diff --git a/.gitattributes b/.gitattributes index f4b031806a3ddbdbb92625f2b882a68bf72bcdb1..e844e6b387aabfb1d51475db6073caa378d06027 100644 --- a/.gitattributes +++ b/.gitattributes @@ -3897,3 +3897,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2023/SpaceEvo_[[:space:]]Hardware-Friendly[[:space:]]Search[[:space:]]Space[[:space:]]Design[[:space:]]for[[:space:]]Efficient[[:space:]]INT8[[:space:]]Inference/7d8d9b36-b64d-45d0-8aa5-d17e8abbee45_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Spacetime[[:space:]]Surface[[:space:]]Regularization[[:space:]]for[[:space:]]Neural[[:space:]]Dynamic[[:space:]]Scene[[:space:]]Reconstruction/11a08db4-fb69-460a-a705-16e05f3f48ac_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Sparse[[:space:]]Instance[[:space:]]Conditioned[[:space:]]Multimodal[[:space:]]Trajectory[[:space:]]Prediction/5a87e2e4-dac0-41f3-9966-85985369c6c6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sparse[[:space:]]Point[[:space:]]Guided[[:space:]]3D[[:space:]]Lane[[:space:]]Detection/6f2348f9-ca1f-4bc7-8757-9bb55f6bb3b9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Sparse[[:space:]]Sampling[[:space:]]Transformer[[:space:]]with[[:space:]]Uncertainty-Driven[[:space:]]Ranking[[:space:]]for[[:space:]]Unified[[:space:]]Removal[[:space:]]of[[:space:]]Raindrops[[:space:]]and[[:space:]]Rain[[:space:]]Streaks/32196ea3-316d-4633-8abd-b15ee0703f11_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SparseBEV_[[:space:]]High-Performance[[:space:]]Sparse[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]from[[:space:]]Multi-Camera[[:space:]]Videos/96a859b1-e12e-4281-b4c4-b8ea44fdbf01_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SparseDet_[[:space:]]Improving[[:space:]]Sparsely[[:space:]]Annotated[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]Pseudo-positive[[:space:]]Mining/5aaa9a89-cdc6-4e0a-94f3-a3b82685ba6a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SparseFusion_[[:space:]]Fusing[[:space:]]Multi-Modal[[:space:]]Sparse[[:space:]]Representations[[:space:]]for[[:space:]]Multi-Sensor[[:space:]]3D[[:space:]]Object[[:space:]]Detection/280aa12d-c539-41fe-b78e-1bcf7c1aa152_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SparseMAE_[[:space:]]Sparse[[:space:]]Training[[:space:]]Meets[[:space:]]Masked[[:space:]]Autoencoders/a542f651-0510-4853-91de-1141d304dc5f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SparseNeRF_[[:space:]]Distilling[[:space:]]Depth[[:space:]]Ranking[[:space:]]for[[:space:]]Few-shot[[:space:]]Novel[[:space:]]View[[:space:]]Synthesis/0d56c181-4428-4dcb-b7bc-71667b8724d5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spatial[[:space:]]Self-Distillation[[:space:]]for[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]Inaccurate[[:space:]]Bounding[[:space:]]Boxes/d223c2f5-fe18-4502-b91f-9060d747649d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spatial-Aware[[:space:]]Token[[:space:]]for[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Object[[:space:]]Localization/f549b4c7-58f5-4874-8f3f-2978b440c9a2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spatially[[:space:]]and[[:space:]]Spectrally[[:space:]]Consistent[[:space:]]Deep[[:space:]]Functional[[:space:]]Maps/38616010-b1fd-470b-8376-180cbf13b5f1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spatially-Adaptive[[:space:]]Feature[[:space:]]Modulation[[:space:]]for[[:space:]]Efficient[[:space:]]Image[[:space:]]Super-Resolution/8203ecb7-a9b9-46de-8112-1277f4eeac63_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spatio-Temporal[[:space:]]Crop[[:space:]]Aggregation[[:space:]]for[[:space:]]Video[[:space:]]Representation[[:space:]]Learning/6732d370-6205-414c-986c-288f16a786f3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spatio-Temporal[[:space:]]Domain[[:space:]]Awareness[[:space:]]for[[:space:]]Multi-Agent[[:space:]]Collaborative[[:space:]]Perception/fe9d96d8-4150-4b6d-a27c-f4b80a133514_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spatio-temporal[[:space:]]Prompting[[:space:]]Network[[:space:]]for[[:space:]]Robust[[:space:]]Video[[:space:]]Feature[[:space:]]Extraction/6b38d58c-2dfe-4bda-8611-478cf7d86d34_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spectral[[:space:]]Graphormer_[[:space:]]Spectral[[:space:]]Graph-Based[[:space:]]Transformer[[:space:]]for[[:space:]]Egocentric[[:space:]]Two-Hand[[:space:]]Reconstruction[[:space:]]using[[:space:]]Multi-View[[:space:]]Color[[:space:]]Images/dbd672e5-b6e6-48ea-861d-97b4378c8ca3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spectrum-guided[[:space:]]Multi-granularity[[:space:]]Referring[[:space:]]Video[[:space:]]Object[[:space:]]Segmentation/dc8d764a-9214-4358-ad24-c2007b58a08c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Speech2Lip_[[:space:]]High-fidelity[[:space:]]Speech[[:space:]]to[[:space:]]Lip[[:space:]]Generation[[:space:]]by[[:space:]]Learning[[:space:]]from[[:space:]]a[[:space:]]Short[[:space:]]Video/b8d7dfde-aec2-4ac3-9ef3-b8789d6534b1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Speech4Mesh_[[:space:]]Speech-Assisted[[:space:]]Monocular[[:space:]]3D[[:space:]]Facial[[:space:]]Reconstruction[[:space:]]for[[:space:]]Speech-Driven[[:space:]]3D[[:space:]]Facial[[:space:]]Animation/236896c7-9d5b-41e0-8f54-b3702ca337d4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spherical[[:space:]]Space[[:space:]]Feature[[:space:]]Decomposition[[:space:]]for[[:space:]]Guided[[:space:]]Depth[[:space:]]Map[[:space:]]Super-Resolution/55278fbf-9750-4827-94bb-81181843441e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SpinCam_[[:space:]]High-Speed[[:space:]]Imaging[[:space:]]via[[:space:]]a[[:space:]]Rotating[[:space:]]Point-Spread[[:space:]]Function/280659a7-98c7-41f5-b75f-aea246ac862a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SportsMOT_[[:space:]]A[[:space:]]Large[[:space:]]Multi-Object[[:space:]]Tracking[[:space:]]Dataset[[:space:]]in[[:space:]]Multiple[[:space:]]Sports[[:space:]]Scenes/4b8d0490-50f4-45bc-9fad-a0ac515c1ca0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Spurious[[:space:]]Features[[:space:]]Everywhere[[:space:]]-[[:space:]]Large-Scale[[:space:]]Detection[[:space:]]of[[:space:]]Harmful[[:space:]]Spurious[[:space:]]Features[[:space:]]in[[:space:]]ImageNet/8d4ee68a-1062-4c0b-81d9-4f01bf36823f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Stabilizing[[:space:]]Visual[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]via[[:space:]]Asymmetric[[:space:]]Interactive[[:space:]]Cooperation/3ce1c749-00ff-4ed7-ac11-5982848d2220_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Stable[[:space:]]Cluster[[:space:]]Discrimination[[:space:]]for[[:space:]]Deep[[:space:]]Clustering/9f881270-2ff5-4f87-a2db-8feef1f47071_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Stable[[:space:]]and[[:space:]]Causal[[:space:]]Inference[[:space:]]for[[:space:]]Discriminative[[:space:]]Self-supervised[[:space:]]Deep[[:space:]]Visual[[:space:]]Representations/fdb72a4a-94ed-4829-ab6f-8cf80d64f7f5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StableVideo_[[:space:]]Text-driven[[:space:]]Consistency-aware[[:space:]]Diffusion[[:space:]]Video[[:space:]]Editing/bcc42e2b-6089-4c27-a5c2-cdec53620ca0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StageInteractor_[[:space:]]Query-based[[:space:]]Object[[:space:]]Detector[[:space:]]with[[:space:]]Cross-stage[[:space:]]Interaction/0e05e637-ac63-484f-9954-22b778c7c120_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Steered[[:space:]]Diffusion_[[:space:]]A[[:space:]]Generalized[[:space:]]Framework[[:space:]]for[[:space:]]Plug-and-Play[[:space:]]Conditional[[:space:]]Image[[:space:]]Synthesis/396c3711-d204-4df2-8e69-9bdf93d4fbff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StegaNeRF_[[:space:]]Embedding[[:space:]]Invisible[[:space:]]Information[[:space:]]within[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields/3c8d0cf6-dc7d-44ba-a804-37f576c9cfcc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Stochastic[[:space:]]Segmentation[[:space:]]with[[:space:]]Conditional[[:space:]]Categorical[[:space:]]Diffusion[[:space:]]Models/36fc9ee6-a13a-40ed-bd76-54644b9d43ea_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Story[[:space:]]Visualization[[:space:]]by[[:space:]]Online[[:space:]]Text[[:space:]]Augmentation[[:space:]]with[[:space:]]Context[[:space:]]Memory/b1180507-958c-45ce-9ffe-139cc0187a6e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Strata-NeRF[[:space:]]_[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]for[[:space:]]Stratified[[:space:]]Scenes/246a74f2-93a7-4d68-8843-6238c9108476_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Strip-MLP_[[:space:]]Efficient[[:space:]]Token[[:space:]]Interaction[[:space:]]for[[:space:]]Vision[[:space:]]MLP/5c4682dc-11cf-43c9-b82f-059171805e73_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Strivec_[[:space:]]Sparse[[:space:]]Tri-Vector[[:space:]]Radiance[[:space:]]Fields/35b59dcb-1ce5-40a4-83d2-07192fb62870_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Structural[[:space:]]Alignment[[:space:]]for[[:space:]]Network[[:space:]]Pruning[[:space:]]through[[:space:]]Partial[[:space:]]Regularization/bbe97c12-8d15-479d-81ec-fff8ed1209f4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Structure[[:space:]]Invariant[[:space:]]Transformation[[:space:]]for[[:space:]]better[[:space:]]Adversarial[[:space:]]Transferability/303ed349-48a8-4393-8e2e-88000a7d20c9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Structure[[:space:]]and[[:space:]]Content-Guided[[:space:]]Video[[:space:]]Synthesis[[:space:]]with[[:space:]]Diffusion[[:space:]]Models/4fea04cc-819a-4990-99d9-0ccc1f6c300d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Structure-Aware[[:space:]]Surface[[:space:]]Reconstruction[[:space:]]via[[:space:]]Primitive[[:space:]]Assembly/4d750dca-9a8b-44f9-b884-62a2d6314ac8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Studying[[:space:]]How[[:space:]]to[[:space:]]Efficiently[[:space:]]and[[:space:]]Effectively[[:space:]]Guide[[:space:]]Models[[:space:]]with[[:space:]]Explanations/88a53c18-99d0-4b4c-b0b2-16a2556a29ae_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleDiffusion_[[:space:]]Controllable[[:space:]]Disentangled[[:space:]]Style[[:space:]]Transfer[[:space:]]via[[:space:]]Diffusion[[:space:]]Models/fea588fd-ec6f-4543-ad58-5dd9ea412b01_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleDomain_[[:space:]]Efficient[[:space:]]and[[:space:]]Lightweight[[:space:]]Parameterizations[[:space:]]of[[:space:]]StyleGAN[[:space:]]for[[:space:]]One-shot[[:space:]]and[[:space:]]Few-shot[[:space:]]Domain[[:space:]]Adaptation/b30ac484-13c8-4c42-8711-cf47c43f4550_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleGANEX_[[:space:]]StyleGAN-Based[[:space:]]Manipulation[[:space:]]Beyond[[:space:]]Cropped[[:space:]]Aligned[[:space:]]Faces/71e23834-fb4f-4c96-b627-797882a6ea6f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleInV_[[:space:]]A[[:space:]]Temporal[[:space:]]Style[[:space:]]Modulated[[:space:]]Inversion[[:space:]]Network[[:space:]]for[[:space:]]Unconditional[[:space:]]Video[[:space:]]Generation/ee18bf56-6e38-4711-9114-1ec06f896ae1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StyleLipSync_[[:space:]]Style-based[[:space:]]Personalized[[:space:]]Lip-sync[[:space:]]Video[[:space:]]Generation/35be9e19-3939-4344-a259-82c8df74d69e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/StylerDALLE_[[:space:]]Language-Guided[[:space:]]Style[[:space:]]Transfer[[:space:]]Using[[:space:]]a[[:space:]]Vector-Quantized[[:space:]]Tokenizer[[:space:]]of[[:space:]]a[[:space:]]Large-Scale[[:space:]]Generative[[:space:]]Model/39699168-4600-4fd3-a205-b0cf85e92a6d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SuS-X_[[:space:]]Training-Free[[:space:]]Name-Only[[:space:]]Transfer[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models/000612a7-f65d-40f9-b28d-4f843782dfff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Subclass-balancing[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Long-tailed[[:space:]]Recognition/bc05a493-9765-4ef8-8b9c-fc9c90911ac5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SupFusion_[[:space:]]Supervised[[:space:]]LiDAR-Camera[[:space:]]Fusion[[:space:]]for[[:space:]]3D[[:space:]]Object[[:space:]]Detection/92e5510f-e140-45f8-8597-b810b193a66f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Supervised[[:space:]]Homography[[:space:]]Learning[[:space:]]with[[:space:]]Realistic[[:space:]]Dataset[[:space:]]Generation/56098eb5-516e-4294-9a4a-ac98eb40dc0f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Surface[[:space:]]Extraction[[:space:]]from[[:space:]]Neural[[:space:]]Unsigned[[:space:]]Distance[[:space:]]Fields/a0829e7a-732d-4761-bffa-ad35624b76ef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Surface[[:space:]]Normal[[:space:]]Clustering[[:space:]]for[[:space:]]Implicit[[:space:]]Representation[[:space:]]of[[:space:]]Manhattan[[:space:]]Scenes/f357e4d1-f9a1-4f72-b3d1-9e59cc154adf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SurfsUP_[[:space:]]Learning[[:space:]]Fluid[[:space:]]Simulation[[:space:]]for[[:space:]]Novel[[:space:]]Surfaces/e47c0301-5f63-4201-a2ee-b44e21fee063_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SurroundOcc_[[:space:]]Multi-camera[[:space:]]3D[[:space:]]Occupancy[[:space:]]Prediction[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/0c733c34-bebb-47b3-8e25-0904ab5a04e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SwiftFormer_[[:space:]]Efficient[[:space:]]Additive[[:space:]]Attention[[:space:]]for[[:space:]]Transformer-based[[:space:]]Real-time[[:space:]]Mobile[[:space:]]Vision[[:space:]]Applications/3fb2491f-940c-4884-86b5-1f7be4430698_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SwinLSTM_[[:space:]]Improving[[:space:]]Spatiotemporal[[:space:]]Prediction[[:space:]]Accuracy[[:space:]]using[[:space:]]Swin[[:space:]]Transformer[[:space:]]and[[:space:]]LSTM/0bea32db-db74-4a9f-85a7-6fe6526e5e78_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/SynBody_[[:space:]]Synthetic[[:space:]]Dataset[[:space:]]with[[:space:]]Layered[[:space:]]Human[[:space:]]Models[[:space:]]for[[:space:]]3D[[:space:]]Human[[:space:]]Perception[[:space:]]and[[:space:]]Modeling/f251ae55-2ca7-4578-ad92-a51119c290a6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Synchronize[[:space:]]Feature[[:space:]]Extracting[[:space:]]and[[:space:]]Matching_[[:space:]]A[[:space:]]Single[[:space:]]Branch[[:space:]]Framework[[:space:]]for[[:space:]]3D[[:space:]]Object[[:space:]]Tracking/d13aa22d-32ff-439b-9d2f-7deb1ed92309_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Synthesizing[[:space:]]Diverse[[:space:]]Human[[:space:]]Motions[[:space:]]in[[:space:]]3D[[:space:]]Indoor[[:space:]]Scenes/39947c2e-9c93-4c2e-8f0c-e7f92b1d1960_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TALL_[[:space:]]Thumbnail[[:space:]]Layout[[:space:]]for[[:space:]]Deepfake[[:space:]]Video[[:space:]]Detection/b759f55b-503f-4164-9793-b1ffe02056fb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TAPIR_[[:space:]]Tracking[[:space:]]Any[[:space:]]Point[[:space:]]with[[:space:]]Per-Frame[[:space:]]Initialization[[:space:]]and[[:space:]]Temporal[[:space:]]Refinement/f0fe9ecb-f819-46e1-ba9a-de70f0257649_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TARGET_[[:space:]]Federated[[:space:]]Class-Continual[[:space:]]Learning[[:space:]]via[[:space:]]Exemplar-Free[[:space:]]Distillation/ff5641a5-17a3-470a-a519-ccac898db108_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TCOVIS_[[:space:]]Temporally[[:space:]]Consistent[[:space:]]Online[[:space:]]Video[[:space:]]Instance[[:space:]]Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TEMPO_[[:space:]]Efficient[[:space:]]Multi-View[[:space:]]Pose[[:space:]]Estimation,[[:space:]]Tracking,[[:space:]]and[[:space:]]Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TF-ICON_[[:space:]]Diffusion-Based[[:space:]]Training-Free[[:space:]]Cross-Domain[[:space:]]Image[[:space:]]Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2023/Sparse Point Guided 3D Lane Detection/6f2348f9-ca1f-4bc7-8757-9bb55f6bb3b9_content_list.json b/2023/Sparse Point Guided 3D Lane Detection/6f2348f9-ca1f-4bc7-8757-9bb55f6bb3b9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..66636438a605704a969b136a0cc2b11b4359f1a3 --- /dev/null +++ b/2023/Sparse Point Guided 3D Lane Detection/6f2348f9-ca1f-4bc7-8757-9bb55f6bb3b9_content_list.json @@ -0,0 +1,1552 @@ +[ + { + "type": "text", + "text": "Sparse Point Guided 3D Lane Detection", + "text_level": 1, + "bbox": [ + 282, + 130, + 687, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengtang Yao $^{1,3}$ , Lidong Yu $^{3}$ , Yuwei Wu $^{1*}$ , Yunde Jia $^{2}$ $^{1}$ Beijing Key Laboratory of Intelligent Information Technology, School of Computer Science & Technology, Beijing Institute of Technology, China \n $^{2}$ Guangdong Laboratory of Machine Perception and Intelligent Computing, Shenzhen MSU-BIT University, China \n $^{3}$ Autonomous Driving Algorithm, DeepRoute", + "bbox": [ + 156, + 179, + 812, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yao.c.t@bit.edu.cn, yvlidong@gmail.com, jiayunde@bit.edu.cn, wuyuwei@bit.edu.cn", + "bbox": [ + 133, + 290, + 834, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D lane detection usually builds a dense correspondence between the front-view space and the BEV space to estimate lane points in the 3D space. 3D lanes only occupy a small ratio of the dense correspondence, while most correspondence belongs to the redundant background. This sparsity phenomenon bottlenecks valuable computation and raises the computation cost of building a high-resolution correspondence for accurate results. In this paper, we propose a sparse point-guided 3D lane detection, focusing on points related to 3D lanes. Our method runs in a coarse-to-fine manner, including coarse-level lane detection and iterative fine-level sparse point refinements. In coarse-level lane detection, we build a dense but efficient correspondence between the front view and BEV space at a very low resolution to compute coarse lanes. Then in fine-level sparse point refinement, we sample sparse points around coarse lanes to extract local features from the high-resolution front-view feature map. The high-resolution local information brought by sparse points refines 3D lanes in the BEV space hierarchically from low resolution to high resolution. The sparse point guides a more effective information flow and greatly promotes the SOTA result by 3 points on the overall F1-score and 6 points on several hard situations while reducing almost half memory cost and speeding up 2 times.", + "bbox": [ + 75, + 371, + 473, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 744, + 209, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D lane detection is an indispensable part of the advanced driver assistance system, supporting functionalities such as automated lane centering and lane departure warning. It relies on building a dense correspondence between the front view space and BEV space to localize lane points in 3D space. Previous methods build a dense correspondence by directly projecting the 2D image or 2D fea", + "bbox": [ + 75, + 768, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b850b6ceb7fd0f7564837946446164bdbe7a844e1d5570f36ca0ea8875a86045.jpg", + "image_caption": [ + "(a) Dense Correspondence", + "Figure 1: The visualization of dense correspondence (a) and sparse correspondence (b) between font view and BEV. The green grids represent areas near lanes, while blue grids are redundant areas far from lanes." + ], + "image_footnote": [], + "bbox": [ + 504, + 338, + 702, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/317220b805f7b838ecdb6615da0c56e7da9430b85430152d2d39a8877fa4a3a1.jpg", + "image_caption": [ + "(b) Sparse Correspondence" + ], + "image_footnote": [], + "bbox": [ + 705, + 338, + 906, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ture map into 3D space using Inverse Perspective Mapping (IPM) [18, 38], but this kind of correspondence is hard to handle the complex road conditions (e.g., uphill and downhill). Recently, some methods [3, 21] build a dense correspondence using a learnable transformer to resolve this problem. The transformer-based correspondence builds a more effective information flow from the front view space to the BEV space, which results in better performance in extreme scenes.", + "bbox": [ + 496, + 578, + 893, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building such dense correspondence is actually redundant for 3D lane detection. As shown in Figure 1, even among a local window near the lane, only $2/5$ correspondences are beneficial for 3D lane detection, while the others are redundant. Among the dense correspondence, 3D lanes only occupy a small ratio, while most correspondences are built for the background. This redundant problem becomes more severe at high resolution, where both the high-resolution front-view and BEV feature maps are necessary for high-quality lane detection in most methods [6, 7, 18, 10, 38, 15]. To this end, we propose to construct a sparse correspondence only focusing on a limited num", + "bbox": [ + 496, + 719, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8363", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ber of points around 3D lanes. The sparse correspondence builds an effective yet efficient information flow from the front-view space to the BEV space. It directly brings the high-resolution lane information from the front-view space, ensuring fine-grained details in high-resolution BEV space. Without redundant correspondence, the information flow learns to focus more on features beneficial for 3D lane results. Meanwhile, it has a naturally efficient performance by only allocating computation to points related to lanes.", + "bbox": [ + 80, + 90, + 467, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a sparse-point-guided 3D lane detection method that decomposes the 3D lane detection into coarse-level lane detection and fine-level sparse point refinement. In coarse-level lane detection, we extract multiscale feature maps from the front-view image. An efficient but dense BEV feature map is built from the front-view feature map at the lowest resolution to detect coarse 3D lanes. Then the fine-level sparse point refinement refines coarse results hierarchically from low resolution to high resolution. In each refinement, we first sample sparse points around coarse 3D lanes within a specific window. And then, we project sampled points onto the front-view plane to extract local features from the high-resolution front-view feature map. The local feature provides fine-grained information to refine the local structure of 3D lanes. At the same time, we compress the front-view feature map into a single vector as the global features to ensure the global smoothness of 3D lanes. We fuse the global feature, the local features, and the sparse point coordinates to predict the location and category of each lane. The fusion of the local and global features refers to the point coordinates, which ensures both the global smoothness and local discrimination of lanes.", + "bbox": [ + 80, + 228, + 467, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate our method on two anchor-based approaches and a widely used segmentation-based approach. The demonstrations are conducted on two real-world datasets (OpenLane [3] and ONCE [18]), including different weathers, lane structures, and road conditions. Our method outperforms the two anchor-based approaches, showing our feasibility to be seamlessly integrated with different prototypical methods to offer consistent improvement. Besides, our method simply outperforms the SOTA anchor-based approach by 2 points on the overall F1-score and more than 6 points on the F1-score in several extreme conditions while reducing half memory cost and speeding up 2 times. As for the segmentation-based approach, our method achieves comparable performance and reduces the memory cost of the 3D lane head by $80\\%$ and speeds up 2 times.", + "bbox": [ + 80, + 561, + 467, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 804, + 215, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Lane Detection", + "text_level": 1, + "bbox": [ + 80, + 830, + 227, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "According to the lane representation, there are mainly three kinds of approaches for lanes detection, including segmentation-based [14, 26, 37, 42], anchor-based [33, 43,", + "bbox": [ + 80, + 854, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "17, 16, 29], and parameter-based methods[5, 35, 32, 19]. In order to estimate the lane representation in the 3D space, some methods first conduct the estimation in the front-view space and then use inverse perspective mapping (IPM) to bridge the front-view space to BEV space for 3D results. The IPM correspondence is based on the planar road assumption, which is hard to handle road undulates, like a hilly road. To tackle this issue, some methods estimate the 3D lanes in the BEV space to resolve complex road conditions [6, 7, 18, 10, 38, 15, 3, 1, 9, 21]. They usually build a dense correspondence by transforming the front-view feature to the BEV feature map. The dense correspondence builds an effective information flow for the BEV feature to improve the quality of the 3D lane.", + "bbox": [ + 503, + 90, + 888, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance the BEV feature, some methods introduce auxiliary tasks for the joint learning[6, 7, 18, 38, 15]. The others focus on designing a more effective transformation [10, 3, 15]. RobustLane [10] uses an attention mechanism to aggregate better global information and thus provides a better global smoothness. RTVLane [15] introduces a geometry consistency between 2D and 3D space to guide the learning of BEV features. Recently, PersFormer [3] uses a transformer to build the dense correspondence between the multi-scale front-view features and BEV feature through cross-attention, which achieves SOTA performance. Building a proper correspondence between the front-view space and the BEV space is the key to improving the 3D lane performance, but a dense correspondence is actually redundant. We observe that lanes only occupy a very small ratio of the dense correspondence. Thus, we build a sparse correspondence that only focuses on the points mostly belonging to a lane. The sparse correspondence efficiently brings the high-resolution lane information from the front-view space, ensuring fine-grained details in high-resolution BEV space. Besides, it also guides the information flow to learn better features beneficial for 3D lane results.", + "bbox": [ + 503, + 303, + 888, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Coarse-to-Fine Methods", + "text_level": 1, + "bbox": [ + 503, + 648, + 718, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The coarse-to-fine design is widely adopted in many methods to hierarchically improve results [43, 42, 13, 39, 30, 24, 28, 40, 11, 22, 2, 36, 25]. The general design consists of a coarse result estimation using low-frequency information and a hierarchical refinement using high-frequency information at high resolution. Due to the sparsity of high-frequency information, many methods [13, 39, 30, 24] propose to only focus on the valuable regions and operate on sparse point sets rather than regular grids to keep the balance between efficiency and accuracy. They mainly focus on common objects and use local features to recover fine-grained structures, like corners and boundaries. Different from common objects, 3D lanes have a uniquely long and thin structure, requiring a globally smooth and locally accurate estimation. In this paper, we fuse the local and global", + "bbox": [ + 503, + 674, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8364", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a682de574565aa88ae80e3ff8deadbb3e6fcaca559687446d822143db9f29bd8.jpg", + "image_caption": [ + "(a) The overall pipeline of our model." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 869, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/92f88d94b97f443e5f2413b294cb85efa0aeb71296b69148d09c05b7a8eef205.jpg", + "image_caption": [ + "(b) The coarse-level lane detection." + ], + "image_footnote": [], + "bbox": [ + 263, + 215, + 707, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0b2ac5337bdf41ff33b82a8918643a070208102da0ab156c3335ff470e166e62.jpg", + "image_caption": [ + "(c) The fine-level sparse point refinement. $\\odot$ represents concatenation.", + "Figure 2: The overall structure of our model is shown in (a). We decompose the 3D lane detection into the coarse-level lane detection shown in (b) and the fine-level sparse point refinement shown in (c)." + ], + "image_footnote": [], + "bbox": [ + 140, + 320, + 828, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "information referring to the coordinates of sparse points and jointly refine the global and local structures of lanes.", + "bbox": [ + 75, + 518, + 468, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 564, + 166, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As 3D lanes only occupy a small ratio among the dense correspondence between the front-view space and BEV space, we propose only focusing on sparse points related to lanes in the high-resolution space. As shown in Figure 2a, we decompose the 3D lane detection into a coarse-level lane detection in the lowest resolution and a series of fine-level sparse point refinement. We extract multi-scale front-view features from a front-view image. The lowest-resolution features are fed into coarse-level lane detection to acquire a rough structure of 3D lanes. The other high-resolution features are used to refine the structure of 3D lanes hierarchically from low resolution to high resolution in fine-level sparse point refinement. As shown in Figure 2b, in coarse-level lane detection, we first build a dense but efficient correspondence between the low-resolution front-view feature and the BEV feature. The low-resolution BEV feature is extracted to estimate the anchor representation of 3D lanes *. As for the fine-level sparse point refinement shown in", + "bbox": [ + 75, + 590, + 470, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2c, candidate points are first sampled around the anchor points of a lane. The candidate point coordinates are then used as indexes to fetch local features from front-view features. We also compress the front-view features into a single vector as global features, which are fused with local features to refine the location and category of lanes.", + "bbox": [ + 504, + 518, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Lane Representation", + "text_level": 1, + "bbox": [ + 500, + 618, + 694, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an image, 3D lane detection aims to detect the location $l$ and category $p$ of lanes $L = \\{l_i,p_i\\}_{i = 0}^{i = N - 1}$ where $N$ is the number of lanes. The location of lane $l_{i}$ is represented by an ordered point sequence:", + "bbox": [ + 500, + 642, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i} = \\left\\{\\left(x _ {j} ^ {i}, y _ {j} ^ {i}, z _ {j} ^ {i}\\right) \\right\\} _ {j = 0} ^ {j = M _ {i} - 1}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 712, + 890, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$M_{i}$ is the number of points.", + "bbox": [ + 500, + 744, + 684, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In anchor-based representation [3, 6, 7], 3D lanes are initialized by a set of anchor lanes with predefined x-coordinates $\\{\\bar{x}^i\\}_{i = 0}^{i = N - 1}$ , where $N$ is the number of lanes. The anchor lanes are further defined by a series of anchor points with predefined y-coordinates $\\{\\bar{y}_j\\}_{j = 0}^{j = M}$ , where $M$ is the anchor point count for each anchor lane. In the anchor-based representation, the location of the 3D lane is formu", + "bbox": [ + 496, + 758, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "*We select the anchor-based approach, i.e., Persformer[3], as the baseline in the main paper. The realization of sparse-point guided lane detec", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "tion for the segmentation-based approach is shown in the Supplementary Material with modification on Persformer backbone.", + "bbox": [ + 500, + 875, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8365", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lated as", + "bbox": [ + 76, + 92, + 130, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i} ^ {\\prime} = \\left\\{\\left(\\Delta_ {x _ {j} ^ {i}}, z _ {j} ^ {i}, v _ {j} ^ {i}\\right) \\right\\} _ {j = 0} ^ {j = M}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 102, + 468, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta_{x_j^i}$ is the offset along the x-axis, $z_j^i$ is the absolute height along the z-axis and $v_{j}^{i}$ is the visibility for each anchor point. The anchor-based representation can be easily converted into the ordered point sequence representation by", + "bbox": [ + 76, + 130, + 468, + 195 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i} = \\left\\{\\left(\\bar {x} ^ {i} + \\Delta_ {x _ {j} ^ {i}}, \\bar {y} _ {j}, z _ {j} ^ {i} \\right\\} _ {j = 0} ^ {j = M - 1}. \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 200, + 468, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Coarse-level Lane Detection", + "text_level": 1, + "bbox": [ + 76, + 229, + 328, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In coarse-level lane detection, we aim to efficiently acquire the coarse location and category of lanes. We first extract the multi-scale features from the front-view image. Then we transform the feature map from the front view to the BEV at the lowest resolution and compute the coarse location and category of 3D lanes, as shown in Figure 2b. Specifically, we build a dense correspondence between the BEV and front view space at the lowest resolution. We then use the front-view points as the index to fetch front-view features for BEV feature extraction. The details of transformation follow the realization of the Persformer [3] and are discussed in Section 4.2. With the dense BEV feature map, we use a standard anchor-based approach [3, 6, 7] to estimate the anchor-based lane representation.", + "bbox": [ + 75, + 253, + 470, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Fine-level Sparse Point Refinement", + "text_level": 1, + "bbox": [ + 76, + 473, + 382, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As aforementioned, 3D lanes occupy very little 3D space, which means a large number of 3D points are redundant. Based on the observation, we design a fine-level sparse point refinement, which only deals with sparse points related to lanes instead of the entire BEV space to improve the coarse results. The refinement is carried out hierarchically from the low resolution to the high resolution. We take coarse 3D lanes estimated from the previous lower resolution as reference lanes and sample candidate sparse points around them, as shown in Figure 2c. The sampled sparse points are then used as indexes to extract the high-resolution front-view features. The fine-grained details from the high-resolution space gradually refine the local structure of coarse results.", + "bbox": [ + 75, + 496, + 468, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Candidate Point Sampling", + "text_level": 1, + "bbox": [ + 76, + 726, + 316, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, points can be sampled along all three dimensions, i.e., $x, y, z$ , but the number of candidate points will become too large to take. In the anchor-based representation, anchor points are predefined and have fixed values on the y-axis. Thus, we sample points only around the anchor points and only on the x-z plane, as shown in Figure 2c. Specifically, we deem coarse points computed from the previous resolution as reference points. The candidate points are uniformly sampled from the neighbors of reference points within a window $\\mathcal{W}$ , which has a predefined", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/94a5065bc1f735bda1b4c079bbf5fe46d5f8d60738078a5a467d0c646078142a.jpg", + "image_caption": [ + "Figure 3: The pipeline of local feature extraction. We project the 3D point $(x,y,z)$ onto image plane $(u,v)$ . A local patch around the point $(u,v)$ is fetched to compute local features using interpolation and multi-layer perception (MLP)." + ], + "image_footnote": [], + "bbox": [ + 506, + 99, + 883, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "size and sampling step. The above sampling process can be indicated as", + "bbox": [ + 498, + 253, + 890, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS \\left(x _ {j} ^ {i}, y _ {j} ^ {i}, z _ {j} ^ {i}\\right) = \\left\\{\\left(x _ {t}, y _ {j} ^ {i}, z _ {t}\\right) \\mid \\left(x _ {t}, z _ {t}\\right) \\in \\mathcal {W} _ {\\left(x _ {j} ^ {i}, z _ {j} ^ {i}\\right)} \\right\\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 294, + 890, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is worth noting that the sampling process has a different requirement for training and inference. During the training stage, we require as much information as possible to supervise the learning of anchors. So, we keep all of the candidate points. In contrast, in the inference stage, we need to reduce as many redundant anchor lanes as possible to save time and memory costs. Thus, we filter redundant anchor lanes in the inference stage before sampling. The filtering is designed by three rules. First, the detected categories of lanes are not backgrounds. Second, there are at least two visible points on a lane. Third, the distance of different lanes is far enough, where we define the distance between two lines as the sum of the distance between anchor points with the same $\\bar{y}_j$ .", + "bbox": [ + 496, + 324, + 892, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Refinement", + "text_level": 1, + "bbox": [ + 500, + 554, + 633, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Anchor-based lane representation consists of two kinds of properties, including the location of each anchor point and the category of each anchor lane.", + "bbox": [ + 496, + 577, + 890, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point Location Refinement Lane has a globally smooth and locally complex structure due to its long and thin shape. To this end, we propose to use both local features of candidate points and global features of the whole image to refine the coarse results. As shown in Figure 3, we project candidate points $(x,y,z)$ onto image plane $(u,v)$ for local feature extraction. This process can be written as", + "bbox": [ + 496, + 623, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left( \\begin{array}{l} u \\\\ v \\\\ 1 \\end{array} \\right) = \\frac {1}{z} \\cdot K \\cdot E \\cdot \\left( \\begin{array}{l} x \\\\ y \\\\ z \\\\ 1 \\end{array} \\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 737, + 890, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K$ is intrinsic matrix, $E$ is extrinsic matrix. As the projected point coordinate is a real value and the feature map only consists of regular grids, we use a local area around the projected point $(u,v)$ to approximate the coordinates. We then use the local area to compute local features $F_{l}$ through bilinear interpolation and multi-layer perception", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8366", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88465d891abf48f72971b568e0cab8c93cba3047e68a1471d7775bdf1ad7d47c.jpg", + "image_caption": [ + "Figure 4: The pipeline of aggregation. Without loss of generality, we take the aggregation on candidate point features as an example to visualize the pipeline of aggregation." + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 848, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MLP). The extracted local feature for sparse points is indicated as $F_{l} \\in \\mathbb{R}^{C \\times N \\times M \\times ||S||}$ , where $C$ is the feature size, $N$ is the number of anchor lanes, $M$ is the number of anchor points and $||S||$ is the sampling size. To guide the fusion of local and global information, we also encode coordinates of sampled sparse points into a high-dimension feature $F_{p} \\in \\mathbb{R}^{C \\times N \\times M \\times ||S||}$ through a single MLP.", + "bbox": [ + 75, + 222, + 470, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The point embedding $F_{p}$ and local features $F_{l}$ are then concatenated as candidate point features $F_{c} \\in \\mathbb{R}^{C \\times N \\times M \\times ||S||}$ . In order to compute anchor point features $F_{a} \\in \\mathbb{R}^{C \\times N \\times M}$ from $F_{c}$ , we propose to aggregate candidate point features belonging to the same anchor point. As shown in Figure 4, we gradually compress the candidate point dimension from $||S||$ to 1. During the compression, we use convolution with the kernel size of $3 \\times 1$ and padding of $(1,0)$ to propagate context information along the anchor point dimension. This operation constrains the context information to only flow among the anchor points in the same lane. Meanwhile, we use convolution with the kernel size of $1 \\times 3$ and padding of $(0,0)$ to compress local information along the candidate point dimension. This operation fuses the local information from the sampled sparse points to the corresponding anchor. These two operations are alternately repeated to propagate and compress information progressively.", + "bbox": [ + 75, + 340, + 472, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As for the global feature, we compress the entire frontview features into a single feature vector $F_{g} \\in \\mathbb{R}^{C}$ through pooling operation, following previous methods [24, 27, 4, 20, 8, 41]. We then concatenate the global feature vector $F_{g}$ to each anchor feature vector $F_{a}$ and fuse them together to balance the global and local information. The fused features are then used to update the anchor point location $\\{(\\Delta_{x_j^i}, z_j^i, v_j^i)\\}_{j=0}^{j=M}$ through 3 convolution layers.", + "bbox": [ + 76, + 628, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lane Category Refinement In order to detect the lane categories, we extract lane features by aggregating the feature of anchor points belonging to the same lane. Similar to the above aggregation operation for the anchor point features, we first concatenate the global feature $F_{g}$ with anchor features $F_{a}$ and then gradually propagate the information along the anchor point dimension. The aggregated lane features $F_{l} \\in \\mathbb{R}^{C \\times N}$ are then used to update the lane category $p_{i}$ through several convolution layers.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss", + "text_level": 1, + "bbox": [ + 500, + 222, + 571, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The predicted anchor-based lane representation consists of category $\\tilde{p}_i$ and location $\\tilde{l}_i = \\{(\\tilde{\\Delta}_{x_j^i},\\tilde{z}_j^i,\\tilde{v}_j^i)\\}_{j = 0}^{j = M}$ . We use the cross entropy to supervise the predicted category $\\tilde{p}$ :", + "bbox": [ + 498, + 244, + 893, + 292 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = - \\frac {1}{N} \\sum_ {i = 0} ^ {i = N - 1} p _ {i} \\log \\left(\\tilde {p} _ {i}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 301, + 890, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As for the location, we supervise the learning of visibility $\\tilde{v}$ by $l_{1}$ loss and use the visibility as weights to compute the $l_{1}$ loss for $\\tilde{\\Delta}_x$ and $\\tilde{z}$ :", + "bbox": [ + 498, + 349, + 892, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\Delta} = \\frac {1}{M \\cdot N} \\sum_ {i = 0} ^ {i = N - 1} \\sum_ {j = 0} ^ {j = M - 1} v _ {j} ^ {i} | | \\Delta_ {x _ {j} ^ {i}} - \\tilde {\\Delta} _ {x _ {j} ^ {i}} | | _ {1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 404, + 890, + 448 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {z} = \\frac {1}{M \\cdot N} \\sum_ {i = 0} ^ {i = N - 1} \\sum_ {j = 0} ^ {j = M - 1} v _ {j} ^ {i} | | z _ {j} ^ {i} - \\tilde {z} _ {j} ^ {i} | | _ {1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 458, + 890, + 501 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v} = \\frac {1}{M \\cdot N} \\sum_ {i = 0} ^ {i = N - 1} \\sum_ {j = 0} ^ {j = M - 1} | | v _ {j} ^ {i} - \\tilde {v} _ {j} ^ {i} | | _ {1}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 507, + 890, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The joint loss can be indicated as", + "bbox": [ + 500, + 554, + 723, + 569 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {p} + \\mathcal {L} _ {\\Delta} + \\mathcal {L} _ {z} + \\mathcal {L} _ {v}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 579, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the hierarchical refinement from the low resolution to the high resolution, we separately compute the above loss for each resolution and sum up the losses as the final loss to train the network.", + "bbox": [ + 498, + 604, + 892, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 679, + 633, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the main paper, we present the sparse-point guided 3D lane detection method on the anchor-based approach. We mainly use Persformer [3] as the baseline, while Gen-LaneNet [7] is taken as an optional baseline to verify the feasibility of our method. We validate our method on the real-world datasets OpenLane [3] and ONCE-3DLanes [18] to demonstrate our priority to handle complex environments and variable scenes. As for our method on the segmentation-based approach, we present the results obtained on the synthetic dataset Apollo 3D Lane [7] in the Supplementary Material. The code is available at https://github.com/YaoChengTang/Sparse-Point-Guided-3D-Lane-Detection", + "bbox": [ + 498, + 704, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8367", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/184b9347a0f344cb0dd3dee72037545709a3c25839c7fe4dbe34548f9511228c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAll ↑Up & Down ↑Curve ↑Extreme Weather ↑Night ↑Intersection ↑Merge & Split ↑
3D-LaneNe [6]44.140.846.547.541.532.141.7
Gen-LaneNet [7]32.325.433.528.118.721.431
PersFormer [3]50.542.455.648.646.64050.7
Ours + PersFormer53.7 (3.2↑)46.2 (3.8↑)59.2 (3.6↑)54.8 (6.2↑)49.8 (3.2↑)41.9 (1.9↑)52.1 (1.4↑)
", + "bbox": [ + 142, + 88, + 834, + 170 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eb895c9e0dc44783d82f900b195e7522e63ee0cb3515fe4094de47b7ccb983b6.jpg", + "table_caption": [ + "Table 1: Performance comparison with state-of-the-art methods on OpenLane benchmark in different scenarios, where the evaluation metric is F1-score. $\\uparrow$ means that the larger the value, the better the result." + ], + "table_footnote": [], + "table_body": "
MethodF-score (%) ↑X error near (m) ↓X error far (m) ↓Z error near (m) ↓Z error far (m) ↓
3D-LaneNe [6]44.10.4790.5720.3670.443
Gen-LaneNet [7]32.30.5910.6840.4110.521
Cond-IPM [3]36.60.5631.0800.4210.892
PersFormer [3]50.50.4850.5530.3640.431
Gen-LaneNet* [7]42.80.4880.6320.3740.481
Ours + Gen-LaneNet*46.60.4750.5770.3710.445
Ours + PersFormer52.30.4680.5140.3710.418
", + "bbox": [ + 140, + 215, + 834, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 382, + 179, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OpneLane Dataset OpneLane [3] is a challenging real-world 3D lane dataset constructed on Waymo Open dataset [31]. It consists of 200K frames with 14 kinds of categories, complex lane structures, and five kinds of weather. $50\\%$ frames have an altitude change of more than $1m$ , and $25\\%$ frames have more than six lanes. Currently, it is the most challenging lane detection dataset, so we choose it as the main dataset to conduct ablation studies and demonstrate the performance promotion of our method.", + "bbox": [ + 75, + 406, + 470, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ONCE-3DLanes Dataset ONCE-3DLanes [18] is a real-world 3D lane dataset constructed on ONCE [23]. It consists of various scenes, such as highways, bridges, tunnels, suburbs, and downtown, with different weather conditions (sunny/rainy) and lighting conditions (day/night).", + "bbox": [ + 75, + 542, + 470, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 633, + 294, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We select the PersFormer [3] as the baseline of the anchor-based approach to leverage the sparse point. EfficientNet [34] is used as the backbone for the front-view feature extraction. Perspective Transformer [3] and 3D-GeoNet [7] are used in coarse-level lane detection to extract the dense BEV features and detect coarse 3D lanes at the lowest resolution. Following PersFormer, we set the highest resolution of BEV space as $208 \\times 128$ with the range of $[-10, 10]$ meters along the x-axis and [3, 101] meters along the y-axis, respectively. We set the count of anchor lanes as $N = 182$ and anchor points for each lane as $M = 10$ , where the fixed $\\{y^i\\}$ is set as $\\{5, 10, 15, 20, 30, 40, 50, 60, 80, 100\\}$ for each lane. In fine-level sparse point refinement, we set the size of sampling window $\\mathcal{W}$ as $3 \\times 3$ with $(1, 0.5)$ sampling step along the x-axis and the z-axis separately.", + "bbox": [ + 75, + 657, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/16bb2c67a45da24dcaafd7f9d1ebccb985d97899760cebd135fdd6c4b182fb01.jpg", + "table_caption": [ + "Table 2: Performance comparison with state-of-the-art methods using different evaluation metrics on the validation set of OpenLane benchmark. * means that we re-implement the method on the OpenLane benchmark. ↑ means that the larger the value, the better the result. ↓ means that the smaller the value, the better the result." + ], + "table_footnote": [], + "table_body": "
MethodTraining \nMemory (G)Inference \nSpeed (FPS)
3D-LaneNe [6]6.6103
Gen-LaneNet [7]4.767
PersFormer [3]19.627
Ours + PersFormer12.545
", + "bbox": [ + 529, + 378, + 867, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3: Training memory cost and inference speed comparison with state-of-the-art approaches. We acquire the memory cost on a single A30 GPU with a batch size of 8. We obtain the inference speed on a single A30 GPU with a batch size of 1.", + "bbox": [ + 496, + 464, + 893, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the training stage, we pre-train the coarse-level lane detection for 100 epochs and then finetune our fine-level sparse point refinement for another 100 epochs. For both pertaining and finetuning, we use Adam optimizer [12] with a base learning rate of $2 \\times 10^{-4}$ and a weight decay of $10^{-3}$ . The cosine learning rate policy is applied to optimization with a maximum iteration of 8 and a minimum learning rate of $10^{-5}$ . We also use gradient norm clipping during the optimization with a maximum norm weight of 35. In the Inference stage, we use the number of valid anchor points in each lane as the threshold to filter anchor lanes that are too close. For more details about training and inference, please refer to the code and our supplemental materials.", + "bbox": [ + 496, + 549, + 893, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 498, + 756, + 684, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following OpenLane, we use bipartite matching to evaluate the predicted and ground-truth lanes. The matching is true positive when the distance error of $75\\%$ lane points is less than $1.5m$ . The percentage of matched ground-truth lanes is deemed as the recall, and the percentage of matched predicted lanes is deemed as the precision. The average of recall and precision is F-score. The lane points are also divided into near points $(3 - 40m)$ and far points $(40 - 101m)$ .", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8368", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afd6c05e45ebccd60776c8079199ce0dc670e7df44f27a6db1ddad4ac81a24f6.jpg", + "image_caption": [ + "Figure 5: The visualization results of PersFormer and our method on the OpneLane dataset." + ], + "image_footnote": [], + "bbox": [ + 181, + 89, + 787, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The point-wise distance error along the x-axis and z-axis is used to evaluate the location accuracy for near and far points. In ONCE, we also use the unilateral Chamfer Distance (CD) metric that computes the chamfer distance between the predicted and ground truth lane.", + "bbox": [ + 76, + 464, + 470, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Benchmark Performance", + "text_level": 1, + "bbox": [ + 76, + 556, + 307, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OpenLane Following the setting in OpenLane, we evaluate the performance in different scenarios. As shown in Table 1, we achieve the best results in all scenarios. It is worth noting that our method achieves more than 6 points promotion in extreme weather scenarios. This promotion shows that our method greatly resists the feature noise and extracts valuable information by only focusing on sparse points around 3D lanes. In Table 2, we further compare our method with state-of-the-art approaches using different evaluation metrics. Our method outperforms all state-of-the-art approaches in almost every evaluation metric. Specifically, we improve the location on the x-axis in both near and far distances and thus improve the overall F1-score. What's more, our method effectively helps Gen-LaneNet outperform both the original result of Gen-LaneNet published by OpenLane and our reproduced Gen-LaneNet* on all metrics. These results prove that our method is flexible and can seamlessly integrate with different prototypical methods to offer consistent improvement.", + "bbox": [ + 76, + 580, + 468, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare our method with state-of-the-art approaches from the training memory cost and the inference", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "speed. As shown in Table 3, our method achieves the lowest memory cost and the fastest inference speed. Compared to the baseline, we greatly reduce the memory cost, which enables the front-view images to have higher resolution. Furthermore, the sparse point guides the network to achieve a better result in most cases with a much faster inference speed. It should be noted that the speed and memory cost reported in this paper is computed based on a pure Python realization of sparse-point sampling and aggregation without any unfair engineering optimization like the custom CUDA kernel that significantly reduces the running time.", + "bbox": [ + 496, + 464, + 890, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides the quantitative analysis, we also compare visualization results of the PersFormer and our method in different scenarios for qualitative analysis. As shown in Figure 5, our method obviously outperforms the PersFormer in cloudy, rainy, night, and different road conditions. This further demonstrates that the sparse point guides the BEV feature to extract more meaningful information from the front-view image and thus improve the recall and accuracy of the detected lanes.", + "bbox": [ + 496, + 637, + 890, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ONCE-3DLanes As illustrated in Table 4, we follow the PersFormer using the validation set of ONCE-3DLanes to evaluate the results of our methods. Considering scenes are easier in the ONCE benchmark, the quantity results are already very high. The sparse point slightly promotes the baseline in the F1-score, Precision, and CD error with a much more efficient performance, which only costs half the memory and inference time.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8369", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/062c8c6471a9bcf40e77be12dc21288035f8df96c9559ce72f21fd7d99f88bdb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodF1-score (%) ↑Precision (%) ↑Recall (%) ↑CD error (m) ↓
3D-LaneNe [6]44.7361.4635.160.127
Gen-LaneNet [7]45.5963.9535.420.121
SALAD [3]64.0775.9055.420.098
PersFormer [3]74.3380.3069.180.074
Ours + PersFormer74.7981.8568.860.070
", + "bbox": [ + 150, + 87, + 823, + 170 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/55048aa91d7ae714f583604cdd56f8aefef50faf55d6cd21ea8f6955ca23520c.jpg", + "table_caption": [ + "Table 4: Performance comparison with state-of-the-art methods on ONCE benchmark. $\\uparrow$ means that the larger the value, the better the result. $\\downarrow$ means that the smaller the value, the better the result." + ], + "table_footnote": [], + "table_body": "
Front-viewFeature ExtractionCoarse levelFine level 1Fine level 2Fine level 3
4.8ms6.1ms3.8ms3.1ms3.4ms
", + "bbox": [ + 107, + 212, + 444, + 257 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ecb1a7fa3c30d81246130d5624e80f533cf978ffa711ba33340ea254e92cae23.jpg", + "table_caption": [ + "Table 5: Inference time cost of each step in our method." + ], + "table_footnote": [], + "table_body": "
Coarse levelFine level 1Fine level 2Fine level 3F1-score (%)X error near (m)
✓51.30.484
✓✓52.10.475
✓✓✓53.60.470
✓✓✓✓53.70.468
", + "bbox": [ + 102, + 281, + 447, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study and Discussion", + "text_level": 1, + "bbox": [ + 76, + 409, + 349, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complexity Analysis We evaluate the time cost of each step in our method to show the efficiency of our sparse refinement. As shown in Table 5, it is more efficient to use sparse points for feature fetching and 3d lane refinement, even compared to the coarse-level lane detection conducted in the lowest resolution.", + "bbox": [ + 75, + 431, + 468, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Refinement In order to validate the effectiveness of our fine-level sparse point refinement, we compare results from the coarse level and from the different fine levels. As shown in Table 6, the fine-level sparse point refinement greatly improves the results of coarse-level lane detection in both F1-score and X error near.", + "bbox": [ + 75, + 523, + 468, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Features To ensure globally smooth and locally discriminative performance, we leverage both the global and local features to refine the lanes. We demonstrate the advantage of balancing the global and local information by comparing three kinds of feature conditions, including only using global features, only using local features, and using both local and global features. As shown in Table 7, the first condition results in a very low F1 score of $46.8\\%$ , and the second one is already comparable to the baseline, while the last one fusing with sparse points gives the best result.", + "bbox": [ + 75, + 613, + 470, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Point Sampling As aforementioned, we sample candidate points around the anchor point on the x-z plane. Here, we analyze the effectiveness of point sampling using different sampling steps.", + "bbox": [ + 75, + 779, + 468, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Table 8, the result gets better when the value of the sampling step is smaller along the $z$ -axis and larger along the $x$ -axis. This phenomenon shows that the local features fetched along the $x$ -axis are more valuable", + "bbox": [ + 75, + 840, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/81b96726f5908e0bf7c18ae722a89de8a61349a759ec180737f977aa6f38f606.jpg", + "table_caption": [ + "Table 6: The performance of coarse-level lane detection and fine-level sparse point refinement." + ], + "table_footnote": [], + "table_body": "
Global FeaturesLocal FeaturesF1-score (%)
✓46.8
✓51.8
✓✓53.7
", + "bbox": [ + 537, + 212, + 857, + 270 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b4c101d59864f58cf02fbf932f9ac886109c4867a9dd892d2502aeddad8f8ae4.jpg", + "table_caption": [ + "Table 7: The performance of our method with/without global features and local features." + ], + "table_footnote": [], + "table_body": "
x-axisz-axisF1-score
1153.61
10.553.65
0.5153.34
0.50.553.46
10.2553.67
", + "bbox": [ + 578, + 306, + 818, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8: The performance of our method with different sampling steps along x-axis and z-axis.", + "bbox": [ + 498, + 393, + 890, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "than the features fetched along the z-axis, where the sampling along the x-axis actually gives us more valid sample points on the front-view image.", + "bbox": [ + 498, + 436, + 890, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Limitations and Discussion", + "text_level": 1, + "bbox": [ + 500, + 491, + 743, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in the above experiments, we have achieved great progress in both accuracy and efficiency. However, our method still has some limitations, e.g., assuming all 3D lane instances could be detected on BEV features in the lowest resolution. This assumption becomes less powerful when meeting extremely complex and crowded road conditions. A slow-fast update strategy for refinement at different scales might give better results, but it requires much more engineering tricks on feature manipulation. In this paper, we mainly focus on proving the priority and possibility of simply using sparse correspondence instead of dense correspondence for 3D lane detection.", + "bbox": [ + 496, + 513, + 893, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 709, + 619, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proved that using sparse points is more efficient and adequate for high-quality 3D lane detection than building a dense correspondence between the HR front-view space and BEV space. We presented a sparse point-guided 3D lane detection method, including coarse-level lane detection and fine-level sparse point refinement. Experiments proved that our method could achieve much better results with less memory and time cost.", + "bbox": [ + 496, + 734, + 890, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by the Natural Science Foundation of China (NSFC) under Grants No.62176021.", + "bbox": [ + 498, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8370", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. arXiv preprint arXiv:2209.07989, 2022.", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022.", + "[3] Li Fei Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proceedings of the European Conference on Computer Vision (ECCV), 2022.", + "[4] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5932-5941, 2019.", + "[5] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17041-17049, 2022.", + "[6] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2921-2930, 2019.", + "[7] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proceedings of the European Conference on Computer Vision (ECCV), 2020.", + "[8] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and Shimin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8928-8937, 2021.", + "[9] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zihan Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. arXiv preprint arXiv:2301.02371, 2023.", + "[10] Yujie Jin, Xiangxuan Ren, Fengxiang Chen, and Weidong Zhang. Robust monocular 3d lane detection with dual attention. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 3348-3352, 2021.", + "[11] Longlong Jing, Yucheng Chen, and Yingli Tian. Coarse-to-fine semantic segmentation from image-level labels. IEEE Transactions on Image Processing (TIP), 29:225-236, 2019.", + "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015.", + "[13] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 9799-9808, 2020." + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Taeyeop Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1965-1973, 2017.", + "[15] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4369-4378, 2022.", + "[16] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Transactions on Intelligent Transportation Systems, 21(1):248-258, 2019.", + "[17] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condilanenet: a top-to-down lane detection framework based on conditional convolution. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 3773-3782, 2021.", + "[18] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Zejian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1765-1772, 2022.", + "[19] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3694-3702, 2021.", + "[20] Shilin Liu, Haoxiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1788-1797, 2021.", + "[21] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022.", + "[22] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 11677-11684, 2020.", + "[23] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Chunjing Xu, et al. One million scenes for autonomous driving: Once dataset. 2021.", + "[24] Lars M. Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4455-4465, 2019.", + "[25] Roozbeh Mottaghi, Yu Xiang, and Silvio Savarese. A coarse-to-fine model for 3d pose estimation and sub-category recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 418-426, 2015." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8371", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaouu Tang. Spatial as deep: Spatial cnn for traffic scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 7276-7283, 2018.", + "[27] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019.", + "[28] Marco Pedersoli, Andrea Vedaldi, Jordi Gonzalez, and Xavier Roca. A coarse-to-fine approach for fast deformable object detection. Pattern Recognition (PR), 48(5):1844-1853, 2015.", + "[29] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 276-291. Springer, 2020.", + "[30] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 81-90, 2020.", + "[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2446-2454, 2020.", + "[32] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021.", + "[33] Lucas Tabelini, Rodrigo Berriel, Thiago Meireles Paixão, Claudine Santos Badue, Alberto Ferreira de Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 294-302, 2021.", + "[34] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 6105-6114. PMLR, 2019.", + "[35] Wouter Van Gansbeke, Bert De Brabandere, Davy Neven, Marc Proesmans, and Luc Van Gool. End-to-end lane detection through differentiable least-squares fitting. In Proceedings of the IEEE International Conference on Computer Vision Workshops (CVPRW), pages 0–0, 2019.", + "[36] Jun Wang, Shiyi Lan, Mingfei Gao, and Larry S Davis. Info-focus: 3d object detection for autonomous driving with dynamic information modeling. In European Conference on Computer Vision, pages 405-420. Springer, 2020.", + "[37] Hang Xu, Shaoju Wang, Xinyue Cai, Wei Zhang, Xiaodan Liang, and Zhenguo Li. Curvelane-nas: Unifying lanesensitive architecture search and adaptive point blending. In Proceedings of the European Conference on Computer Vision (ECCV), pages 689-704. Springer, 2020." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Fan Yan, Ming-Jun Nie, Xinyue Cai, Jianhua Han, Han Xu, Zhengxin Yang, Chao Ye, Yanwei Fu, Michael B. Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17122–17131, 2022.", + "[39] Chengtang Yao, Yunde Jia, Huijun Di, Pengxiang Li, and Yuwei Wu. A decomposition model for stereo matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6091-6100, 2021.", + "[40] Jian Yao, Marko Boben, Sanja Fidler, and Raquel Urtasun. Real-time coarse-to-fine topologically preserving segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2947-2955, 2015.", + "[41] Jianglong Ye, Yuntao Chen, Naiyan Wang, and X. Wang. Gifs: Neural implicit function for general shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12819-12829, 2022.", + "[42] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. Resa: Recurrent feature-shift aggregator for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3547-3554, 2021.", + "[43] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 898-907, 2022." + ], + "bbox": [ + 501, + 92, + 892, + 513 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8372", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Sparse Point Guided 3D Lane Detection/6f2348f9-ca1f-4bc7-8757-9bb55f6bb3b9_model.json b/2023/Sparse Point Guided 3D Lane Detection/6f2348f9-ca1f-4bc7-8757-9bb55f6bb3b9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..25fdba3668c4d5cadb42dec16620e9102f5508a7 --- /dev/null +++ b/2023/Sparse Point Guided 3D Lane Detection/6f2348f9-ca1f-4bc7-8757-9bb55f6bb3b9_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.283, + 0.131, + 0.688, + 0.153 + ], + "angle": 0, + "content": "Sparse Point Guided 3D Lane Detection" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.18, + 0.813, + 0.288 + ], + "angle": 0, + "content": "Chengtang Yao\\(^{1,3}\\), Lidong Yu\\(^{3}\\), Yuwei Wu\\(^{1*}\\), Yunde Jia\\(^{2}\\) \n\\(^{1}\\)Beijing Key Laboratory of Intelligent Information Technology, School of Computer Science & Technology, Beijing Institute of Technology, China \n\\(^{2}\\)Guangdong Laboratory of Machine Perception and Intelligent Computing, Shenzhen MSU-BIT University, China \n\\(^{3}\\)Autonomous Driving Algorithm, DeepRoute" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.291, + 0.835, + 0.305 + ], + "angle": 0, + "content": "yao.c.t@bit.edu.cn, yvlidong@gmail.com, jiayunde@bit.edu.cn, wuyuwei@bit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.474, + 0.735 + ], + "angle": 0, + "content": "3D lane detection usually builds a dense correspondence between the front-view space and the BEV space to estimate lane points in the 3D space. 3D lanes only occupy a small ratio of the dense correspondence, while most correspondence belongs to the redundant background. This sparsity phenomenon bottlenecks valuable computation and raises the computation cost of building a high-resolution correspondence for accurate results. In this paper, we propose a sparse point-guided 3D lane detection, focusing on points related to 3D lanes. Our method runs in a coarse-to-fine manner, including coarse-level lane detection and iterative fine-level sparse point refinements. In coarse-level lane detection, we build a dense but efficient correspondence between the front view and BEV space at a very low resolution to compute coarse lanes. Then in fine-level sparse point refinement, we sample sparse points around coarse lanes to extract local features from the high-resolution front-view feature map. The high-resolution local information brought by sparse points refines 3D lanes in the BEV space hierarchically from low resolution to high resolution. The sparse point guides a more effective information flow and greatly promotes the SOTA result by 3 points on the overall F1-score and 6 points on several hard situations while reducing almost half memory cost and speeding up 2 times." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.745, + 0.21, + 0.76 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.47, + 0.878 + ], + "angle": 0, + "content": "3D lane detection is an indispensable part of the advanced driver assistance system, supporting functionalities such as automated lane centering and lane departure warning. It relies on building a dense correspondence between the front view space and BEV space to localize lane points in 3D space. Previous methods build a dense correspondence by directly projecting the 2D image or 2D fea" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.339, + 0.703, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.481, + 0.669, + 0.493 + ], + "angle": 0, + "content": "(a) Dense Correspondence" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.339, + 0.907, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.481, + 0.872, + 0.493 + ], + "angle": 0, + "content": "(b) Sparse Correspondence" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.505, + 0.895, + 0.566 + ], + "angle": 0, + "content": "Figure 1: The visualization of dense correspondence (a) and sparse correspondence (b) between font view and BEV. The green grids represent areas near lanes, while blue grids are redundant areas far from lanes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.895, + 0.715 + ], + "angle": 0, + "content": "ture map into 3D space using Inverse Perspective Mapping (IPM) [18, 38], but this kind of correspondence is hard to handle the complex road conditions (e.g., uphill and downhill). Recently, some methods [3, 21] build a dense correspondence using a learnable transformer to resolve this problem. The transformer-based correspondence builds a more effective information flow from the front view space to the BEV space, which results in better performance in extreme scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Building such dense correspondence is actually redundant for 3D lane detection. As shown in Figure 1, even among a local window near the lane, only \\(2/5\\) correspondences are beneficial for 3D lane detection, while the others are redundant. Among the dense correspondence, 3D lanes only occupy a small ratio, while most correspondences are built for the background. This redundant problem becomes more severe at high resolution, where both the high-resolution front-view and BEV feature maps are necessary for high-quality lane detection in most methods [6, 7, 18, 10, 38, 15]. To this end, we propose to construct a sparse correspondence only focusing on a limited num" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8363" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.226 + ], + "angle": 0, + "content": "ber of points around 3D lanes. The sparse correspondence builds an effective yet efficient information flow from the front-view space to the BEV space. It directly brings the high-resolution lane information from the front-view space, ensuring fine-grained details in high-resolution BEV space. Without redundant correspondence, the information flow learns to focus more on features beneficial for 3D lane results. Meanwhile, it has a naturally efficient performance by only allocating computation to points related to lanes." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.229, + 0.468, + 0.559 + ], + "angle": 0, + "content": "In this paper, we present a sparse-point-guided 3D lane detection method that decomposes the 3D lane detection into coarse-level lane detection and fine-level sparse point refinement. In coarse-level lane detection, we extract multiscale feature maps from the front-view image. An efficient but dense BEV feature map is built from the front-view feature map at the lowest resolution to detect coarse 3D lanes. Then the fine-level sparse point refinement refines coarse results hierarchically from low resolution to high resolution. In each refinement, we first sample sparse points around coarse 3D lanes within a specific window. And then, we project sampled points onto the front-view plane to extract local features from the high-resolution front-view feature map. The local feature provides fine-grained information to refine the local structure of 3D lanes. At the same time, we compress the front-view feature map into a single vector as the global features to ensure the global smoothness of 3D lanes. We fuse the global feature, the local features, and the sparse point coordinates to predict the location and category of each lane. The fusion of the local and global features refers to the point coordinates, which ensures both the global smoothness and local discrimination of lanes." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.562, + 0.468, + 0.787 + ], + "angle": 0, + "content": "We validate our method on two anchor-based approaches and a widely used segmentation-based approach. The demonstrations are conducted on two real-world datasets (OpenLane [3] and ONCE [18]), including different weathers, lane structures, and road conditions. Our method outperforms the two anchor-based approaches, showing our feasibility to be seamlessly integrated with different prototypical methods to offer consistent improvement. Besides, our method simply outperforms the SOTA anchor-based approach by 2 points on the overall F1-score and more than 6 points on the F1-score in several extreme conditions while reducing half memory cost and speeding up 2 times. As for the segmentation-based approach, our method achieves comparable performance and reduces the memory cost of the 3D lane head by \\(80\\%\\) and speeds up 2 times." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.805, + 0.216, + 0.82 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.831, + 0.228, + 0.845 + ], + "angle": 0, + "content": "2.1. Lane Detection" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "According to the lane representation, there are mainly three kinds of approaches for lanes detection, including segmentation-based [14, 26, 37, 42], anchor-based [33, 43," + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.302 + ], + "angle": 0, + "content": "17, 16, 29], and parameter-based methods[5, 35, 32, 19]. In order to estimate the lane representation in the 3D space, some methods first conduct the estimation in the front-view space and then use inverse perspective mapping (IPM) to bridge the front-view space to BEV space for 3D results. The IPM correspondence is based on the planar road assumption, which is hard to handle road undulates, like a hilly road. To tackle this issue, some methods estimate the 3D lanes in the BEV space to resolve complex road conditions [6, 7, 18, 10, 38, 15, 3, 1, 9, 21]. They usually build a dense correspondence by transforming the front-view feature to the BEV feature map. The dense correspondence builds an effective information flow for the BEV feature to improve the quality of the 3D lane." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.304, + 0.89, + 0.634 + ], + "angle": 0, + "content": "To enhance the BEV feature, some methods introduce auxiliary tasks for the joint learning[6, 7, 18, 38, 15]. The others focus on designing a more effective transformation [10, 3, 15]. RobustLane [10] uses an attention mechanism to aggregate better global information and thus provides a better global smoothness. RTVLane [15] introduces a geometry consistency between 2D and 3D space to guide the learning of BEV features. Recently, PersFormer [3] uses a transformer to build the dense correspondence between the multi-scale front-view features and BEV feature through cross-attention, which achieves SOTA performance. Building a proper correspondence between the front-view space and the BEV space is the key to improving the 3D lane performance, but a dense correspondence is actually redundant. We observe that lanes only occupy a very small ratio of the dense correspondence. Thus, we build a sparse correspondence that only focuses on the points mostly belonging to a lane. The sparse correspondence efficiently brings the high-resolution lane information from the front-view space, ensuring fine-grained details in high-resolution BEV space. Besides, it also guides the information flow to learn better features beneficial for 3D lane results." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.65, + 0.72, + 0.665 + ], + "angle": 0, + "content": "2.2. Coarse-to-Fine Methods" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.675, + 0.89, + 0.9 + ], + "angle": 0, + "content": "The coarse-to-fine design is widely adopted in many methods to hierarchically improve results [43, 42, 13, 39, 30, 24, 28, 40, 11, 22, 2, 36, 25]. The general design consists of a coarse result estimation using low-frequency information and a hierarchical refinement using high-frequency information at high resolution. Due to the sparsity of high-frequency information, many methods [13, 39, 30, 24] propose to only focus on the valuable regions and operate on sparse point sets rather than regular grids to keep the balance between efficiency and accuracy. They mainly focus on common objects and use local features to recover fine-grained structures, like corners and boundaries. Different from common objects, 3D lanes have a uniquely long and thin structure, requiring a globally smooth and locally accurate estimation. In this paper, we fuse the local and global" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8364" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.87, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.199, + 0.596, + 0.212 + ], + "angle": 0, + "content": "(a) The overall pipeline of our model." + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.217, + 0.708, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.3, + 0.59, + 0.313 + ], + "angle": 0, + "content": "(b) The coarse-level lane detection." + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.321, + 0.83, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.451, + 0.695, + 0.466 + ], + "angle": 0, + "content": "(c) The fine-level sparse point refinement. \\(\\odot\\) represents concatenation." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.477, + 0.893, + 0.508 + ], + "angle": 0, + "content": "Figure 2: The overall structure of our model is shown in (a). We decompose the 3D lane detection into the coarse-level lane detection shown in (b) and the fine-level sparse point refinement shown in (c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.549 + ], + "angle": 0, + "content": "information referring to the coordinates of sparse points and jointly refine the global and local structures of lanes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.565, + 0.168, + 0.58 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.471, + 0.864 + ], + "angle": 0, + "content": "As 3D lanes only occupy a small ratio among the dense correspondence between the front-view space and BEV space, we propose only focusing on sparse points related to lanes in the high-resolution space. As shown in Figure 2a, we decompose the 3D lane detection into a coarse-level lane detection in the lowest resolution and a series of fine-level sparse point refinement. We extract multi-scale front-view features from a front-view image. The lowest-resolution features are fed into coarse-level lane detection to acquire a rough structure of 3D lanes. The other high-resolution features are used to refine the structure of 3D lanes hierarchically from low resolution to high resolution in fine-level sparse point refinement. As shown in Figure 2b, in coarse-level lane detection, we first build a dense but efficient correspondence between the low-resolution front-view feature and the BEV feature. The low-resolution BEV feature is extracted to estimate the anchor representation of 3D lanes *. As for the fine-level sparse point refinement shown in" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.519, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Figure 2c, candidate points are first sampled around the anchor points of a lane. The candidate point coordinates are then used as indexes to fetch local features from front-view features. We also compress the front-view features into a single vector as global features, which are fused with local features to refine the location and category of lanes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.696, + 0.635 + ], + "angle": 0, + "content": "3.1. Lane Representation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.643, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Given an image, 3D lane detection aims to detect the location \\(l\\) and category \\(p\\) of lanes \\(L = \\{l_i,p_i\\}_{i = 0}^{i = N - 1}\\) where \\(N\\) is the number of lanes. The location of lane \\(l_{i}\\) is represented by an ordered point sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.713, + 0.892, + 0.735 + ], + "angle": 0, + "content": "\\[\nl _ {i} = \\left\\{\\left(x _ {j} ^ {i}, y _ {j} ^ {i}, z _ {j} ^ {i}\\right) \\right\\} _ {j = 0} ^ {j = M _ {i} - 1}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.745, + 0.685, + 0.759 + ], + "angle": 0, + "content": "\\(M_{i}\\) is the number of points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.892, + 0.866 + ], + "angle": 0, + "content": "In anchor-based representation [3, 6, 7], 3D lanes are initialized by a set of anchor lanes with predefined x-coordinates \\(\\{\\bar{x}^i\\}_{i = 0}^{i = N - 1}\\), where \\(N\\) is the number of lanes. The anchor lanes are further defined by a series of anchor points with predefined y-coordinates \\(\\{\\bar{y}_j\\}_{j = 0}^{j = M}\\), where \\(M\\) is the anchor point count for each anchor lane. In the anchor-based representation, the location of the 3D lane is formu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*We select the anchor-based approach, i.e., Persformer[3], as the baseline in the main paper. The realization of sparse-point guided lane detec" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.899 + ], + "angle": 0, + "content": "tion for the segmentation-based approach is shown in the Supplementary Material with modification on Persformer backbone." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8365" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.132, + 0.105 + ], + "angle": 0, + "content": "lated as" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.103, + 0.469, + 0.127 + ], + "angle": 0, + "content": "\\[\nl _ {i} ^ {\\prime} = \\left\\{\\left(\\Delta_ {x _ {j} ^ {i}}, z _ {j} ^ {i}, v _ {j} ^ {i}\\right) \\right\\} _ {j = 0} ^ {j = M}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.131, + 0.47, + 0.196 + ], + "angle": 0, + "content": "where \\(\\Delta_{x_j^i}\\) is the offset along the x-axis, \\(z_j^i\\) is the absolute height along the z-axis and \\(v_{j}^{i}\\) is the visibility for each anchor point. The anchor-based representation can be easily converted into the ordered point sequence representation by" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.201, + 0.469, + 0.224 + ], + "angle": 0, + "content": "\\[\nl _ {i} = \\left\\{\\left(\\bar {x} ^ {i} + \\Delta_ {x _ {j} ^ {i}}, \\bar {y} _ {j}, z _ {j} ^ {i} \\right\\} _ {j = 0} ^ {j = M - 1}. \\right. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.231, + 0.33, + 0.246 + ], + "angle": 0, + "content": "3.2. Coarse-level Lane Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.255, + 0.471, + 0.466 + ], + "angle": 0, + "content": "In coarse-level lane detection, we aim to efficiently acquire the coarse location and category of lanes. We first extract the multi-scale features from the front-view image. Then we transform the feature map from the front view to the BEV at the lowest resolution and compute the coarse location and category of 3D lanes, as shown in Figure 2b. Specifically, we build a dense correspondence between the BEV and front view space at the lowest resolution. We then use the front-view points as the index to fetch front-view features for BEV feature extraction. The details of transformation follow the realization of the Persformer [3] and are discussed in Section 4.2. With the dense BEV feature map, we use a standard anchor-based approach [3, 6, 7] to estimate the anchor-based lane representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.474, + 0.383, + 0.489 + ], + "angle": 0, + "content": "3.3. Fine-level Sparse Point Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.709 + ], + "angle": 0, + "content": "As aforementioned, 3D lanes occupy very little 3D space, which means a large number of 3D points are redundant. Based on the observation, we design a fine-level sparse point refinement, which only deals with sparse points related to lanes instead of the entire BEV space to improve the coarse results. The refinement is carried out hierarchically from the low resolution to the high resolution. We take coarse 3D lanes estimated from the previous lower resolution as reference lanes and sample candidate sparse points around them, as shown in Figure 2c. The sampled sparse points are then used as indexes to extract the high-resolution front-view features. The fine-grained details from the high-resolution space gradually refine the local structure of coarse results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.317, + 0.743 + ], + "angle": 0, + "content": "3.3.1 Candidate Point Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Intuitively, points can be sampled along all three dimensions, i.e., \\( x, y, z \\), but the number of candidate points will become too large to take. In the anchor-based representation, anchor points are predefined and have fixed values on the y-axis. Thus, we sample points only around the anchor points and only on the x-z plane, as shown in Figure 2c. Specifically, we deem coarse points computed from the previous resolution as reference points. The candidate points are uniformly sampled from the neighbors of reference points within a window \\( \\mathcal{W} \\), which has a predefined" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.1, + 0.885, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.17, + 0.892, + 0.246 + ], + "angle": 0, + "content": "Figure 3: The pipeline of local feature extraction. We project the 3D point \\((x,y,z)\\) onto image plane \\((u,v)\\). A local patch around the point \\((u,v)\\) is fetched to compute local features using interpolation and multi-layer perception (MLP)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.255, + 0.892, + 0.284 + ], + "angle": 0, + "content": "size and sampling step. The above sampling process can be indicated as" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.295, + 0.892, + 0.318 + ], + "angle": 0, + "content": "\\[\nS \\left(x _ {j} ^ {i}, y _ {j} ^ {i}, z _ {j} ^ {i}\\right) = \\left\\{\\left(x _ {t}, y _ {j} ^ {i}, z _ {t}\\right) \\mid \\left(x _ {t}, z _ {t}\\right) \\in \\mathcal {W} _ {\\left(x _ {j} ^ {i}, z _ {j} ^ {i}\\right)} \\right\\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.893, + 0.538 + ], + "angle": 0, + "content": "It is worth noting that the sampling process has a different requirement for training and inference. During the training stage, we require as much information as possible to supervise the learning of anchors. So, we keep all of the candidate points. In contrast, in the inference stage, we need to reduce as many redundant anchor lanes as possible to save time and memory costs. Thus, we filter redundant anchor lanes in the inference stage before sampling. The filtering is designed by three rules. First, the detected categories of lanes are not backgrounds. Second, there are at least two visible points on a lane. Third, the distance of different lanes is far enough, where we define the distance between two lines as the sum of the distance between anchor points with the same \\(\\bar{y}_j\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.635, + 0.569 + ], + "angle": 0, + "content": "3.3.2 Refinement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Anchor-based lane representation consists of two kinds of properties, including the location of each anchor point and the category of each anchor lane." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.73 + ], + "angle": 0, + "content": "Point Location Refinement Lane has a globally smooth and locally complex structure due to its long and thin shape. To this end, we propose to use both local features of candidate points and global features of the whole image to refine the coarse results. As shown in Figure 3, we project candidate points \\((x,y,z)\\) onto image plane \\((u,v)\\) for local feature extraction. This process can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.738, + 0.892, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\left( \\begin{array}{l} u \\\\ v \\\\ 1 \\end{array} \\right) = \\frac {1}{z} \\cdot K \\cdot E \\cdot \\left( \\begin{array}{l} x \\\\ y \\\\ z \\\\ 1 \\end{array} \\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(K\\) is intrinsic matrix, \\(E\\) is extrinsic matrix. As the projected point coordinate is a real value and the feature map only consists of regular grids, we use a local area around the projected point \\((u,v)\\) to approximate the coordinates. We then use the local area to compute local features \\(F_{l}\\) through bilinear interpolation and multi-layer perception" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8366" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.849, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.182, + 0.895, + 0.214 + ], + "angle": 0, + "content": "Figure 4: The pipeline of aggregation. Without loss of generality, we take the aggregation on candidate point features as an example to visualize the pipeline of aggregation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.223, + 0.471, + 0.33 + ], + "angle": 0, + "content": "MLP). The extracted local feature for sparse points is indicated as \\( F_{l} \\in \\mathbb{R}^{C \\times N \\times M \\times ||S||} \\), where \\( C \\) is the feature size, \\( N \\) is the number of anchor lanes, \\( M \\) is the number of anchor points and \\( ||S|| \\) is the sampling size. To guide the fusion of local and global information, we also encode coordinates of sampled sparse points into a high-dimension feature \\( F_{p} \\in \\mathbb{R}^{C \\times N \\times M \\times ||S||} \\) through a single MLP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.473, + 0.617 + ], + "angle": 0, + "content": "The point embedding \\( F_{p} \\) and local features \\( F_{l} \\) are then concatenated as candidate point features \\( F_{c} \\in \\mathbb{R}^{C \\times N \\times M \\times ||S||} \\). In order to compute anchor point features \\( F_{a} \\in \\mathbb{R}^{C \\times N \\times M} \\) from \\( F_{c} \\), we propose to aggregate candidate point features belonging to the same anchor point. As shown in Figure 4, we gradually compress the candidate point dimension from \\( ||S|| \\) to 1. During the compression, we use convolution with the kernel size of \\( 3 \\times 1 \\) and padding of \\( (1,0) \\) to propagate context information along the anchor point dimension. This operation constrains the context information to only flow among the anchor points in the same lane. Meanwhile, we use convolution with the kernel size of \\( 1 \\times 3 \\) and padding of \\( (0,0) \\) to compress local information along the candidate point dimension. This operation fuses the local information from the sampled sparse points to the corresponding anchor. These two operations are alternately repeated to propagate and compress information progressively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.629, + 0.47, + 0.754 + ], + "angle": 0, + "content": "As for the global feature, we compress the entire frontview features into a single feature vector \\( F_{g} \\in \\mathbb{R}^{C} \\) through pooling operation, following previous methods [24, 27, 4, 20, 8, 41]. We then concatenate the global feature vector \\( F_{g} \\) to each anchor feature vector \\( F_{a} \\) and fuse them together to balance the global and local information. The fused features are then used to update the anchor point location \\( \\{(\\Delta_{x_j^i}, z_j^i, v_j^i)\\}_{j=0}^{j=M} \\) through 3 convolution layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Lane Category Refinement In order to detect the lane categories, we extract lane features by aggregating the feature of anchor points belonging to the same lane. Similar to the above aggregation operation for the anchor point features, we first concatenate the global feature \\( F_{g} \\) with anchor features \\( F_{a} \\) and then gradually propagate the information along the anchor point dimension. The aggregated lane features \\( F_{l} \\in \\mathbb{R}^{C \\times N} \\) are then used to update the lane category \\( p_{i} \\) through several convolution layers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.223, + 0.573, + 0.236 + ], + "angle": 0, + "content": "3.4. Loss" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.246, + 0.894, + 0.294 + ], + "angle": 0, + "content": "The predicted anchor-based lane representation consists of category \\(\\tilde{p}_i\\) and location \\(\\tilde{l}_i = \\{(\\tilde{\\Delta}_{x_j^i},\\tilde{z}_j^i,\\tilde{v}_j^i)\\}_{j = 0}^{j = M}\\). We use the cross entropy to supervise the predicted category \\(\\tilde{p}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = - \\frac {1}{N} \\sum_ {i = 0} ^ {i = N - 1} p _ {i} \\log \\left(\\tilde {p} _ {i}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.351, + 0.893, + 0.397 + ], + "angle": 0, + "content": "As for the location, we supervise the learning of visibility \\(\\tilde{v}\\) by \\(l_{1}\\) loss and use the visibility as weights to compute the \\(l_{1}\\) loss for \\(\\tilde{\\Delta}_x\\) and \\(\\tilde{z}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.405, + 0.892, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\Delta} = \\frac {1}{M \\cdot N} \\sum_ {i = 0} ^ {i = N - 1} \\sum_ {j = 0} ^ {j = M - 1} v _ {j} ^ {i} | | \\Delta_ {x _ {j} ^ {i}} - \\tilde {\\Delta} _ {x _ {j} ^ {i}} | | _ {1}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.459, + 0.891, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {z} = \\frac {1}{M \\cdot N} \\sum_ {i = 0} ^ {i = N - 1} \\sum_ {j = 0} ^ {j = M - 1} v _ {j} ^ {i} | | z _ {j} ^ {i} - \\tilde {z} _ {j} ^ {i} | | _ {1}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.508, + 0.892, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v} = \\frac {1}{M \\cdot N} \\sum_ {i = 0} ^ {i = N - 1} \\sum_ {j = 0} ^ {j = M - 1} | | v _ {j} ^ {i} - \\tilde {v} _ {j} ^ {i} | | _ {1}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.555, + 0.724, + 0.57 + ], + "angle": 0, + "content": "The joint loss can be indicated as" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.58, + 0.891, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {p} + \\mathcal {L} _ {\\Delta} + \\mathcal {L} _ {z} + \\mathcal {L} _ {v}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.606, + 0.893, + 0.666 + ], + "angle": 0, + "content": "During the hierarchical refinement from the low resolution to the high resolution, we separately compute the above loss for each resolution and sum up the losses as the final loss to train the network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.634, + 0.697 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In the main paper, we present the sparse-point guided 3D lane detection method on the anchor-based approach. We mainly use Persformer [3] as the baseline, while Gen-LaneNet [7] is taken as an optional baseline to verify the feasibility of our method. We validate our method on the real-world datasets OpenLane [3] and ONCE-3DLanes [18] to demonstrate our priority to handle complex environments and variable scenes. As for our method on the segmentation-based approach, we present the results obtained on the synthetic dataset Apollo 3D Lane [7] in the Supplementary Material. The code is available at https://github.com/YaoChengTang/Sparse-Point-Guided-3D-Lane-Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8367" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.089, + 0.835, + 0.171 + ], + "angle": 0, + "content": "
MethodAll ↑Up & Down ↑Curve ↑Extreme Weather ↑Night ↑Intersection ↑Merge & Split ↑
3D-LaneNe [6]44.140.846.547.541.532.141.7
Gen-LaneNet [7]32.325.433.528.118.721.431
PersFormer [3]50.542.455.648.646.64050.7
Ours + PersFormer53.7 (3.2↑)46.2 (3.8↑)59.2 (3.6↑)54.8 (6.2↑)49.8 (3.2↑)41.9 (1.9↑)52.1 (1.4↑)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.173, + 0.895, + 0.207 + ], + "angle": 0, + "content": "Table 1: Performance comparison with state-of-the-art methods on OpenLane benchmark in different scenarios, where the evaluation metric is F1-score. \\(\\uparrow\\) means that the larger the value, the better the result." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.217, + 0.835, + 0.326 + ], + "angle": 0, + "content": "
MethodF-score (%) ↑X error near (m) ↓X error far (m) ↓Z error near (m) ↓Z error far (m) ↓
3D-LaneNe [6]44.10.4790.5720.3670.443
Gen-LaneNet [7]32.30.5910.6840.4110.521
Cond-IPM [3]36.60.5631.0800.4210.892
PersFormer [3]50.50.4850.5530.3640.431
Gen-LaneNet* [7]42.80.4880.6320.3740.481
Ours + Gen-LaneNet*46.60.4750.5770.3710.445
Ours + PersFormer52.30.4680.5140.3710.418
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.329, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Table 2: Performance comparison with state-of-the-art methods using different evaluation metrics on the validation set of OpenLane benchmark. * means that we re-implement the method on the OpenLane benchmark. ↑ means that the larger the value, the better the result. ↓ means that the smaller the value, the better the result." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.383, + 0.181, + 0.398 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.471, + 0.543 + ], + "angle": 0, + "content": "OpneLane Dataset OpneLane [3] is a challenging real-world 3D lane dataset constructed on Waymo Open dataset [31]. It consists of 200K frames with 14 kinds of categories, complex lane structures, and five kinds of weather. \\(50\\%\\) frames have an altitude change of more than \\(1m\\), and \\(25\\%\\) frames have more than six lanes. Currently, it is the most challenging lane detection dataset, so we choose it as the main dataset to conduct ablation studies and demonstrate the performance promotion of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.472, + 0.622 + ], + "angle": 0, + "content": "ONCE-3DLanes Dataset ONCE-3DLanes [18] is a real-world 3D lane dataset constructed on ONCE [23]. It consists of various scenes, such as highways, bridges, tunnels, suburbs, and downtown, with different weather conditions (sunny/rainy) and lighting conditions (day/night)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.295, + 0.65 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.472, + 0.903 + ], + "angle": 0, + "content": "We select the PersFormer [3] as the baseline of the anchor-based approach to leverage the sparse point. EfficientNet [34] is used as the backbone for the front-view feature extraction. Perspective Transformer [3] and 3D-GeoNet [7] are used in coarse-level lane detection to extract the dense BEV features and detect coarse 3D lanes at the lowest resolution. Following PersFormer, we set the highest resolution of BEV space as \\(208 \\times 128\\) with the range of \\([-10, 10]\\) meters along the x-axis and [3, 101] meters along the y-axis, respectively. We set the count of anchor lanes as \\(N = 182\\) and anchor points for each lane as \\(M = 10\\), where the fixed \\(\\{y^i\\}\\) is set as \\(\\{5, 10, 15, 20, 30, 40, 50, 60, 80, 100\\}\\) for each lane. In fine-level sparse point refinement, we set the size of sampling window \\(\\mathcal{W}\\) as \\(3 \\times 3\\) with \\((1, 0.5)\\) sampling step along the x-axis and the z-axis separately." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.38, + 0.868, + 0.462 + ], + "angle": 0, + "content": "
MethodTraining \nMemory (G)Inference \nSpeed (FPS)
3D-LaneNe [6]6.6103
Gen-LaneNet [7]4.767
PersFormer [3]19.627
Ours + PersFormer12.545
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.465, + 0.895, + 0.541 + ], + "angle": 0, + "content": "Table 3: Training memory cost and inference speed comparison with state-of-the-art approaches. We acquire the memory cost on a single A30 GPU with a batch size of 8. We obtain the inference speed on a single A30 GPU with a batch size of 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.749 + ], + "angle": 0, + "content": "In the training stage, we pre-train the coarse-level lane detection for 100 epochs and then finetune our fine-level sparse point refinement for another 100 epochs. For both pertaining and finetuning, we use Adam optimizer [12] with a base learning rate of \\(2 \\times 10^{-4}\\) and a weight decay of \\(10^{-3}\\). The cosine learning rate policy is applied to optimization with a maximum iteration of 8 and a minimum learning rate of \\(10^{-5}\\). We also use gradient norm clipping during the optimization with a maximum norm weight of 35. In the Inference stage, we use the number of valid anchor points in each lane as the threshold to filter anchor lanes that are too close. For more details about training and inference, please refer to the code and our supplemental materials." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.685, + 0.771 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Following OpenLane, we use bipartite matching to evaluate the predicted and ground-truth lanes. The matching is true positive when the distance error of \\(75\\%\\) lane points is less than \\(1.5m\\). The percentage of matched ground-truth lanes is deemed as the recall, and the percentage of matched predicted lanes is deemed as the precision. The average of recall and precision is F-score. The lane points are also divided into near points \\((3 - 40m)\\) and far points \\((40 - 101m)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8368" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.09, + 0.789, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.44, + 0.785, + 0.457 + ], + "angle": 0, + "content": "Figure 5: The visualization results of PersFormer and our method on the OpneLane dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.465, + 0.471, + 0.541 + ], + "angle": 0, + "content": "The point-wise distance error along the x-axis and z-axis is used to evaluate the location accuracy for near and far points. In ONCE, we also use the unilateral Chamfer Distance (CD) metric that computes the chamfer distance between the predicted and ground truth lane." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.557, + 0.308, + 0.571 + ], + "angle": 0, + "content": "4.4. Benchmark Performance" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.582, + 0.47, + 0.868 + ], + "angle": 0, + "content": "OpenLane Following the setting in OpenLane, we evaluate the performance in different scenarios. As shown in Table 1, we achieve the best results in all scenarios. It is worth noting that our method achieves more than 6 points promotion in extreme weather scenarios. This promotion shows that our method greatly resists the feature noise and extracts valuable information by only focusing on sparse points around 3D lanes. In Table 2, we further compare our method with state-of-the-art approaches using different evaluation metrics. Our method outperforms all state-of-the-art approaches in almost every evaluation metric. Specifically, we improve the location on the x-axis in both near and far distances and thus improve the overall F1-score. What's more, our method effectively helps Gen-LaneNet outperform both the original result of Gen-LaneNet published by OpenLane and our reproduced Gen-LaneNet* on all metrics. These results prove that our method is flexible and can seamlessly integrate with different prototypical methods to offer consistent improvement." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We also compare our method with state-of-the-art approaches from the training memory cost and the inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.892, + 0.631 + ], + "angle": 0, + "content": "speed. As shown in Table 3, our method achieves the lowest memory cost and the fastest inference speed. Compared to the baseline, we greatly reduce the memory cost, which enables the front-view images to have higher resolution. Furthermore, the sparse point guides the network to achieve a better result in most cases with a much faster inference speed. It should be noted that the speed and memory cost reported in this paper is computed based on a pure Python realization of sparse-point sampling and aggregation without any unfair engineering optimization like the custom CUDA kernel that significantly reduces the running time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Besides the quantitative analysis, we also compare visualization results of the PersFormer and our method in different scenarios for qualitative analysis. As shown in Figure 5, our method obviously outperforms the PersFormer in cloudy, rainy, night, and different road conditions. This further demonstrates that the sparse point guides the BEV feature to extract more meaningful information from the front-view image and thus improve the recall and accuracy of the detected lanes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ONCE-3DLanes As illustrated in Table 4, we follow the PersFormer using the validation set of ONCE-3DLanes to evaluate the results of our methods. Considering scenes are easier in the ONCE benchmark, the quantity results are already very high. The sparse point slightly promotes the baseline in the F1-score, Precision, and CD error with a much more efficient performance, which only costs half the memory and inference time." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8369" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.151, + 0.088, + 0.825, + 0.171 + ], + "angle": 0, + "content": "
MethodF1-score (%) ↑Precision (%) ↑Recall (%) ↑CD error (m) ↓
3D-LaneNe [6]44.7361.4635.160.127
Gen-LaneNet [7]45.5963.9535.420.121
SALAD [3]64.0775.9055.420.098
PersFormer [3]74.3380.3069.180.074
Ours + PersFormer74.7981.8568.860.070
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.173, + 0.895, + 0.207 + ], + "angle": 0, + "content": "Table 4: Performance comparison with state-of-the-art methods on ONCE benchmark. \\(\\uparrow\\) means that the larger the value, the better the result. \\(\\downarrow\\) means that the smaller the value, the better the result." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.213, + 0.446, + 0.258 + ], + "angle": 0, + "content": "
Front-viewFeature ExtractionCoarse levelFine level 1Fine level 2Fine level 3
4.8ms6.1ms3.8ms3.1ms3.4ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.26, + 0.457, + 0.275 + ], + "angle": 0, + "content": "Table 5: Inference time cost of each step in our method." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.282, + 0.449, + 0.366 + ], + "angle": 0, + "content": "
Coarse levelFine level 1Fine level 2Fine level 3F1-score (%)X error near (m)
✓51.30.484
✓✓52.10.475
✓✓✓53.60.470
✓✓✓✓53.70.468
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.369, + 0.47, + 0.4 + ], + "angle": 0, + "content": "Table 6: The performance of coarse-level lane detection and fine-level sparse point refinement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.41, + 0.35, + 0.426 + ], + "angle": 0, + "content": "4.5. Ablation Study and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.522 + ], + "angle": 0, + "content": "Complexity Analysis We evaluate the time cost of each step in our method to show the efficiency of our sparse refinement. As shown in Table 5, it is more efficient to use sparse points for feature fetching and 3d lane refinement, even compared to the coarse-level lane detection conducted in the lowest resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.613 + ], + "angle": 0, + "content": "Effectiveness of Refinement In order to validate the effectiveness of our fine-level sparse point refinement, we compare results from the coarse level and from the different fine levels. As shown in Table 6, the fine-level sparse point refinement greatly improves the results of coarse-level lane detection in both F1-score and X error near." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Effectiveness of Features To ensure globally smooth and locally discriminative performance, we leverage both the global and local features to refine the lanes. We demonstrate the advantage of balancing the global and local information by comparing three kinds of feature conditions, including only using global features, only using local features, and using both local and global features. As shown in Table 7, the first condition results in a very low F1 score of \\(46.8\\%\\), and the second one is already comparable to the baseline, while the last one fusing with sparse points gives the best result." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.841 + ], + "angle": 0, + "content": "Effectiveness of Point Sampling As aforementioned, we sample candidate points around the anchor point on the x-z plane. Here, we analyze the effectiveness of point sampling using different sampling steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Table 8, the result gets better when the value of the sampling step is smaller along the \\( z \\)-axis and larger along the \\( x \\)-axis. This phenomenon shows that the local features fetched along the \\( x \\)-axis are more valuable" + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.213, + 0.858, + 0.271 + ], + "angle": 0, + "content": "
Global FeaturesLocal FeaturesF1-score (%)
✓46.8
✓51.8
✓✓53.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.274, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Table 7: The performance of our method with/without global features and local features." + }, + { + "type": "table", + "bbox": [ + 0.579, + 0.308, + 0.819, + 0.391 + ], + "angle": 0, + "content": "
x-axisz-axisF1-score
1153.61
10.553.65
0.5153.34
0.50.553.46
10.2553.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.395, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Table 8: The performance of our method with different sampling steps along x-axis and z-axis." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.438, + 0.892, + 0.485 + ], + "angle": 0, + "content": "than the features fetched along the z-axis, where the sampling along the x-axis actually gives us more valid sample points on the front-view image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.492, + 0.745, + 0.507 + ], + "angle": 0, + "content": "4.6. Limitations and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.895, + 0.697 + ], + "angle": 0, + "content": "As shown in the above experiments, we have achieved great progress in both accuracy and efficiency. However, our method still has some limitations, e.g., assuming all 3D lane instances could be detected on BEV features in the lowest resolution. This assumption becomes less powerful when meeting extremely complex and crowded road conditions. A slow-fast update strategy for refinement at different scales might give better results, but it requires much more engineering tricks on feature manipulation. In this paper, we mainly focus on proving the priority and possibility of simply using sparse correspondence instead of dense correspondence for 3D lane detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.62, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.856 + ], + "angle": 0, + "content": "In this paper, we have proved that using sparse points is more efficient and adequate for high-quality 3D lane detection than building a dense correspondence between the HR front-view space and BEV space. We presented a sparse point-guided 3D lane detection method, including coarse-level lane detection and fine-level sparse point refinement. Experiments proved that our method could achieve much better results with less memory and time cost." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by the Natural Science Foundation of China (NSFC) under Grants No.62176021." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8370" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.17 + ], + "angle": 0, + "content": "[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. arXiv preprint arXiv:2209.07989, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.469, + 0.24 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.469, + 0.324 + ], + "angle": 0, + "content": "[3] Li Fei Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proceedings of the European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.469, + 0.38 + ], + "angle": 0, + "content": "[4] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5932-5941, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.45 + ], + "angle": 0, + "content": "[5] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17041-17049, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[6] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2921-2930, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.469, + 0.577 + ], + "angle": 0, + "content": "[7] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proceedings of the European Conference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[8] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and Shimin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8928-8937, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[9] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zihan Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. arXiv preprint arXiv:2301.02371, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[10] Yujie Jin, Xiangxuan Ren, Fengxiang Chen, and Weidong Zhang. Robust monocular 3d lane detection with dual attention. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 3348-3352, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[11] Longlong Jing, Yucheng Chen, and Yingli Tian. Coarse-to-fine semantic segmentation from image-level labels. IEEE Transactions on Image Processing (TIP), 29:225-236, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 9799-9808, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[14] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Taeyeop Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1965-1973, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[15] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4369-4378, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[16] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Transactions on Intelligent Transportation Systems, 21(1):248-258, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[17] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condilanenet: a top-to-down lane detection framework based on conditional convolution. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 3773-3782, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[18] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Zejian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1765-1772, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[19] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3694-3702, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Shilin Liu, Haoxiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1788-1797, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[21] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[22] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 11677-11684, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[23] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Chunjing Xu, et al. One million scenes for autonomous driving: Once dataset. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[24] Lars M. Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4455-4465, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Roozbeh Mottaghi, Yu Xiang, and Silvio Savarese. A coarse-to-fine model for 3d pose estimation and sub-category recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 418-426, 2015." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8371" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaouu Tang. Spatial as deep: Spatial cnn for traffic scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 7276-7283, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.231 + ], + "angle": 0, + "content": "[27] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[28] Marco Pedersoli, Andrea Vedaldi, Jordi Gonzalez, and Xavier Roca. A coarse-to-fine approach for fast deformable object detection. Pattern Recognition (PR), 48(5):1844-1853, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[29] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 276-291. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[30] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 81-90, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.471, + 0.497 + ], + "angle": 0, + "content": "[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2446-2454, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[32] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.649 + ], + "angle": 0, + "content": "[33] Lucas Tabelini, Rodrigo Berriel, Thiago Meireles Paixão, Claudine Santos Badue, Alberto Ferreira de Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 294-302, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.705 + ], + "angle": 0, + "content": "[34] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 6105-6114. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[35] Wouter Van Gansbeke, Bert De Brabandere, Davy Neven, Marc Proesmans, and Luc Van Gool. End-to-end lane detection through differentiable least-squares fitting. In Proceedings of the IEEE International Conference on Computer Vision Workshops (CVPRW), pages 0–0, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[36] Jun Wang, Shiyi Lan, Mingfei Gao, and Larry S Davis. Info-focus: 3d object detection for autonomous driving with dynamic information modeling. In European Conference on Computer Vision, pages 405-420. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[37] Hang Xu, Shaoju Wang, Xinyue Cai, Wei Zhang, Xiaodan Liang, and Zhenguo Li. Curvelane-nas: Unifying lanesensitive architecture search and adaptive point blending. In Proceedings of the European Conference on Computer Vision (ECCV), pages 689-704. Springer, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.175 + ], + "angle": 0, + "content": "[38] Fan Yan, Ming-Jun Nie, Xinyue Cai, Jianhua Han, Han Xu, Zhengxin Yang, Chao Ye, Yanwei Fu, Michael B. Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17122–17131, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[39] Chengtang Yao, Yunde Jia, Huijun Di, Pengxiang Li, and Yuwei Wu. A decomposition model for stereo matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6091-6100, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.301 + ], + "angle": 0, + "content": "[40] Jian Yao, Marko Boben, Sanja Fidler, and Raquel Urtasun. Real-time coarse-to-fine topologically preserving segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2947-2955, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.893, + 0.372 + ], + "angle": 0, + "content": "[41] Jianglong Ye, Yuntao Chen, Naiyan Wang, and X. Wang. Gifs: Neural implicit function for general shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12819-12829, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.893, + 0.442 + ], + "angle": 0, + "content": "[42] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. Resa: Recurrent feature-shift aggregator for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3547-3554, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.445, + 0.893, + 0.514 + ], + "angle": 0, + "content": "[43] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 898-907, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8372" + } + ] +] \ No newline at end of file diff --git a/2023/Sparse Point Guided 3D Lane Detection/6f2348f9-ca1f-4bc7-8757-9bb55f6bb3b9_origin.pdf b/2023/Sparse Point Guided 3D Lane Detection/6f2348f9-ca1f-4bc7-8757-9bb55f6bb3b9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fea0627cd5919443bb1ca3d6d4f63684916bbfee --- /dev/null +++ b/2023/Sparse Point Guided 3D Lane Detection/6f2348f9-ca1f-4bc7-8757-9bb55f6bb3b9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9d7829bdabc8b48da91eb3d17ad82b56426121814e4a95d28587ea30dc7c60b +size 1403610 diff --git a/2023/Sparse Point Guided 3D Lane Detection/full.md b/2023/Sparse Point Guided 3D Lane Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2febfd3f27e86887b61aefe8cc75e0bffc3ddd52 --- /dev/null +++ b/2023/Sparse Point Guided 3D Lane Detection/full.md @@ -0,0 +1,308 @@ +# Sparse Point Guided 3D Lane Detection + +Chengtang Yao $^{1,3}$ , Lidong Yu $^{3}$ , Yuwei Wu $^{1*}$ , Yunde Jia $^{2}$ $^{1}$ Beijing Key Laboratory of Intelligent Information Technology, School of Computer Science & Technology, Beijing Institute of Technology, China + $^{2}$ Guangdong Laboratory of Machine Perception and Intelligent Computing, Shenzhen MSU-BIT University, China + $^{3}$ Autonomous Driving Algorithm, DeepRoute + +yao.c.t@bit.edu.cn, yvlidong@gmail.com, jiayunde@bit.edu.cn, wuyuwei@bit.edu.cn + +# Abstract + +3D lane detection usually builds a dense correspondence between the front-view space and the BEV space to estimate lane points in the 3D space. 3D lanes only occupy a small ratio of the dense correspondence, while most correspondence belongs to the redundant background. This sparsity phenomenon bottlenecks valuable computation and raises the computation cost of building a high-resolution correspondence for accurate results. In this paper, we propose a sparse point-guided 3D lane detection, focusing on points related to 3D lanes. Our method runs in a coarse-to-fine manner, including coarse-level lane detection and iterative fine-level sparse point refinements. In coarse-level lane detection, we build a dense but efficient correspondence between the front view and BEV space at a very low resolution to compute coarse lanes. Then in fine-level sparse point refinement, we sample sparse points around coarse lanes to extract local features from the high-resolution front-view feature map. The high-resolution local information brought by sparse points refines 3D lanes in the BEV space hierarchically from low resolution to high resolution. The sparse point guides a more effective information flow and greatly promotes the SOTA result by 3 points on the overall F1-score and 6 points on several hard situations while reducing almost half memory cost and speeding up 2 times. + +# 1. Introduction + +3D lane detection is an indispensable part of the advanced driver assistance system, supporting functionalities such as automated lane centering and lane departure warning. It relies on building a dense correspondence between the front view space and BEV space to localize lane points in 3D space. Previous methods build a dense correspondence by directly projecting the 2D image or 2D fea + +![](images/b850b6ceb7fd0f7564837946446164bdbe7a844e1d5570f36ca0ea8875a86045.jpg) +(a) Dense Correspondence +Figure 1: The visualization of dense correspondence (a) and sparse correspondence (b) between font view and BEV. The green grids represent areas near lanes, while blue grids are redundant areas far from lanes. + +![](images/317220b805f7b838ecdb6615da0c56e7da9430b85430152d2d39a8877fa4a3a1.jpg) +(b) Sparse Correspondence + +ture map into 3D space using Inverse Perspective Mapping (IPM) [18, 38], but this kind of correspondence is hard to handle the complex road conditions (e.g., uphill and downhill). Recently, some methods [3, 21] build a dense correspondence using a learnable transformer to resolve this problem. The transformer-based correspondence builds a more effective information flow from the front view space to the BEV space, which results in better performance in extreme scenes. + +Building such dense correspondence is actually redundant for 3D lane detection. As shown in Figure 1, even among a local window near the lane, only $2/5$ correspondences are beneficial for 3D lane detection, while the others are redundant. Among the dense correspondence, 3D lanes only occupy a small ratio, while most correspondences are built for the background. This redundant problem becomes more severe at high resolution, where both the high-resolution front-view and BEV feature maps are necessary for high-quality lane detection in most methods [6, 7, 18, 10, 38, 15]. To this end, we propose to construct a sparse correspondence only focusing on a limited num + +ber of points around 3D lanes. The sparse correspondence builds an effective yet efficient information flow from the front-view space to the BEV space. It directly brings the high-resolution lane information from the front-view space, ensuring fine-grained details in high-resolution BEV space. Without redundant correspondence, the information flow learns to focus more on features beneficial for 3D lane results. Meanwhile, it has a naturally efficient performance by only allocating computation to points related to lanes. + +In this paper, we present a sparse-point-guided 3D lane detection method that decomposes the 3D lane detection into coarse-level lane detection and fine-level sparse point refinement. In coarse-level lane detection, we extract multiscale feature maps from the front-view image. An efficient but dense BEV feature map is built from the front-view feature map at the lowest resolution to detect coarse 3D lanes. Then the fine-level sparse point refinement refines coarse results hierarchically from low resolution to high resolution. In each refinement, we first sample sparse points around coarse 3D lanes within a specific window. And then, we project sampled points onto the front-view plane to extract local features from the high-resolution front-view feature map. The local feature provides fine-grained information to refine the local structure of 3D lanes. At the same time, we compress the front-view feature map into a single vector as the global features to ensure the global smoothness of 3D lanes. We fuse the global feature, the local features, and the sparse point coordinates to predict the location and category of each lane. The fusion of the local and global features refers to the point coordinates, which ensures both the global smoothness and local discrimination of lanes. + +We validate our method on two anchor-based approaches and a widely used segmentation-based approach. The demonstrations are conducted on two real-world datasets (OpenLane [3] and ONCE [18]), including different weathers, lane structures, and road conditions. Our method outperforms the two anchor-based approaches, showing our feasibility to be seamlessly integrated with different prototypical methods to offer consistent improvement. Besides, our method simply outperforms the SOTA anchor-based approach by 2 points on the overall F1-score and more than 6 points on the F1-score in several extreme conditions while reducing half memory cost and speeding up 2 times. As for the segmentation-based approach, our method achieves comparable performance and reduces the memory cost of the 3D lane head by $80\%$ and speeds up 2 times. + +# 2. Related Work + +# 2.1. Lane Detection + +According to the lane representation, there are mainly three kinds of approaches for lanes detection, including segmentation-based [14, 26, 37, 42], anchor-based [33, 43, + +17, 16, 29], and parameter-based methods[5, 35, 32, 19]. In order to estimate the lane representation in the 3D space, some methods first conduct the estimation in the front-view space and then use inverse perspective mapping (IPM) to bridge the front-view space to BEV space for 3D results. The IPM correspondence is based on the planar road assumption, which is hard to handle road undulates, like a hilly road. To tackle this issue, some methods estimate the 3D lanes in the BEV space to resolve complex road conditions [6, 7, 18, 10, 38, 15, 3, 1, 9, 21]. They usually build a dense correspondence by transforming the front-view feature to the BEV feature map. The dense correspondence builds an effective information flow for the BEV feature to improve the quality of the 3D lane. + +To enhance the BEV feature, some methods introduce auxiliary tasks for the joint learning[6, 7, 18, 38, 15]. The others focus on designing a more effective transformation [10, 3, 15]. RobustLane [10] uses an attention mechanism to aggregate better global information and thus provides a better global smoothness. RTVLane [15] introduces a geometry consistency between 2D and 3D space to guide the learning of BEV features. Recently, PersFormer [3] uses a transformer to build the dense correspondence between the multi-scale front-view features and BEV feature through cross-attention, which achieves SOTA performance. Building a proper correspondence between the front-view space and the BEV space is the key to improving the 3D lane performance, but a dense correspondence is actually redundant. We observe that lanes only occupy a very small ratio of the dense correspondence. Thus, we build a sparse correspondence that only focuses on the points mostly belonging to a lane. The sparse correspondence efficiently brings the high-resolution lane information from the front-view space, ensuring fine-grained details in high-resolution BEV space. Besides, it also guides the information flow to learn better features beneficial for 3D lane results. + +# 2.2. Coarse-to-Fine Methods + +The coarse-to-fine design is widely adopted in many methods to hierarchically improve results [43, 42, 13, 39, 30, 24, 28, 40, 11, 22, 2, 36, 25]. The general design consists of a coarse result estimation using low-frequency information and a hierarchical refinement using high-frequency information at high resolution. Due to the sparsity of high-frequency information, many methods [13, 39, 30, 24] propose to only focus on the valuable regions and operate on sparse point sets rather than regular grids to keep the balance between efficiency and accuracy. They mainly focus on common objects and use local features to recover fine-grained structures, like corners and boundaries. Different from common objects, 3D lanes have a uniquely long and thin structure, requiring a globally smooth and locally accurate estimation. In this paper, we fuse the local and global + +![](images/a682de574565aa88ae80e3ff8deadbb3e6fcaca559687446d822143db9f29bd8.jpg) +(a) The overall pipeline of our model. + +![](images/92f88d94b97f443e5f2413b294cb85efa0aeb71296b69148d09c05b7a8eef205.jpg) +(b) The coarse-level lane detection. + +![](images/0b2ac5337bdf41ff33b82a8918643a070208102da0ab156c3335ff470e166e62.jpg) +(c) The fine-level sparse point refinement. $\odot$ represents concatenation. +Figure 2: The overall structure of our model is shown in (a). We decompose the 3D lane detection into the coarse-level lane detection shown in (b) and the fine-level sparse point refinement shown in (c). + +information referring to the coordinates of sparse points and jointly refine the global and local structures of lanes. + +# 3. Method + +As 3D lanes only occupy a small ratio among the dense correspondence between the front-view space and BEV space, we propose only focusing on sparse points related to lanes in the high-resolution space. As shown in Figure 2a, we decompose the 3D lane detection into a coarse-level lane detection in the lowest resolution and a series of fine-level sparse point refinement. We extract multi-scale front-view features from a front-view image. The lowest-resolution features are fed into coarse-level lane detection to acquire a rough structure of 3D lanes. The other high-resolution features are used to refine the structure of 3D lanes hierarchically from low resolution to high resolution in fine-level sparse point refinement. As shown in Figure 2b, in coarse-level lane detection, we first build a dense but efficient correspondence between the low-resolution front-view feature and the BEV feature. The low-resolution BEV feature is extracted to estimate the anchor representation of 3D lanes *. As for the fine-level sparse point refinement shown in + +Figure 2c, candidate points are first sampled around the anchor points of a lane. The candidate point coordinates are then used as indexes to fetch local features from front-view features. We also compress the front-view features into a single vector as global features, which are fused with local features to refine the location and category of lanes. + +# 3.1. Lane Representation + +Given an image, 3D lane detection aims to detect the location $l$ and category $p$ of lanes $L = \{l_i,p_i\}_{i = 0}^{i = N - 1}$ where $N$ is the number of lanes. The location of lane $l_{i}$ is represented by an ordered point sequence: + +$$ +l _ {i} = \left\{\left(x _ {j} ^ {i}, y _ {j} ^ {i}, z _ {j} ^ {i}\right) \right\} _ {j = 0} ^ {j = M _ {i} - 1}. \tag {1} +$$ + +$M_{i}$ is the number of points. + +In anchor-based representation [3, 6, 7], 3D lanes are initialized by a set of anchor lanes with predefined x-coordinates $\{\bar{x}^i\}_{i = 0}^{i = N - 1}$ , where $N$ is the number of lanes. The anchor lanes are further defined by a series of anchor points with predefined y-coordinates $\{\bar{y}_j\}_{j = 0}^{j = M}$ , where $M$ is the anchor point count for each anchor lane. In the anchor-based representation, the location of the 3D lane is formu + +lated as + +$$ +l _ {i} ^ {\prime} = \left\{\left(\Delta_ {x _ {j} ^ {i}}, z _ {j} ^ {i}, v _ {j} ^ {i}\right) \right\} _ {j = 0} ^ {j = M}, \tag {2} +$$ + +where $\Delta_{x_j^i}$ is the offset along the x-axis, $z_j^i$ is the absolute height along the z-axis and $v_{j}^{i}$ is the visibility for each anchor point. The anchor-based representation can be easily converted into the ordered point sequence representation by + +$$ +l _ {i} = \left\{\left(\bar {x} ^ {i} + \Delta_ {x _ {j} ^ {i}}, \bar {y} _ {j}, z _ {j} ^ {i} \right\} _ {j = 0} ^ {j = M - 1}. \right. \tag {3} +$$ + +# 3.2. Coarse-level Lane Detection + +In coarse-level lane detection, we aim to efficiently acquire the coarse location and category of lanes. We first extract the multi-scale features from the front-view image. Then we transform the feature map from the front view to the BEV at the lowest resolution and compute the coarse location and category of 3D lanes, as shown in Figure 2b. Specifically, we build a dense correspondence between the BEV and front view space at the lowest resolution. We then use the front-view points as the index to fetch front-view features for BEV feature extraction. The details of transformation follow the realization of the Persformer [3] and are discussed in Section 4.2. With the dense BEV feature map, we use a standard anchor-based approach [3, 6, 7] to estimate the anchor-based lane representation. + +# 3.3. Fine-level Sparse Point Refinement + +As aforementioned, 3D lanes occupy very little 3D space, which means a large number of 3D points are redundant. Based on the observation, we design a fine-level sparse point refinement, which only deals with sparse points related to lanes instead of the entire BEV space to improve the coarse results. The refinement is carried out hierarchically from the low resolution to the high resolution. We take coarse 3D lanes estimated from the previous lower resolution as reference lanes and sample candidate sparse points around them, as shown in Figure 2c. The sampled sparse points are then used as indexes to extract the high-resolution front-view features. The fine-grained details from the high-resolution space gradually refine the local structure of coarse results. + +# 3.3.1 Candidate Point Sampling + +Intuitively, points can be sampled along all three dimensions, i.e., $x, y, z$ , but the number of candidate points will become too large to take. In the anchor-based representation, anchor points are predefined and have fixed values on the y-axis. Thus, we sample points only around the anchor points and only on the x-z plane, as shown in Figure 2c. Specifically, we deem coarse points computed from the previous resolution as reference points. The candidate points are uniformly sampled from the neighbors of reference points within a window $\mathcal{W}$ , which has a predefined + +![](images/94a5065bc1f735bda1b4c079bbf5fe46d5f8d60738078a5a467d0c646078142a.jpg) +Figure 3: The pipeline of local feature extraction. We project the 3D point $(x,y,z)$ onto image plane $(u,v)$ . A local patch around the point $(u,v)$ is fetched to compute local features using interpolation and multi-layer perception (MLP). + +size and sampling step. The above sampling process can be indicated as + +$$ +S \left(x _ {j} ^ {i}, y _ {j} ^ {i}, z _ {j} ^ {i}\right) = \left\{\left(x _ {t}, y _ {j} ^ {i}, z _ {t}\right) \mid \left(x _ {t}, z _ {t}\right) \in \mathcal {W} _ {\left(x _ {j} ^ {i}, z _ {j} ^ {i}\right)} \right\}. \tag {4} +$$ + +It is worth noting that the sampling process has a different requirement for training and inference. During the training stage, we require as much information as possible to supervise the learning of anchors. So, we keep all of the candidate points. In contrast, in the inference stage, we need to reduce as many redundant anchor lanes as possible to save time and memory costs. Thus, we filter redundant anchor lanes in the inference stage before sampling. The filtering is designed by three rules. First, the detected categories of lanes are not backgrounds. Second, there are at least two visible points on a lane. Third, the distance of different lanes is far enough, where we define the distance between two lines as the sum of the distance between anchor points with the same $\bar{y}_j$ . + +# 3.3.2 Refinement + +Anchor-based lane representation consists of two kinds of properties, including the location of each anchor point and the category of each anchor lane. + +Point Location Refinement Lane has a globally smooth and locally complex structure due to its long and thin shape. To this end, we propose to use both local features of candidate points and global features of the whole image to refine the coarse results. As shown in Figure 3, we project candidate points $(x,y,z)$ onto image plane $(u,v)$ for local feature extraction. This process can be written as + +$$ +\left( \begin{array}{l} u \\ v \\ 1 \end{array} \right) = \frac {1}{z} \cdot K \cdot E \cdot \left( \begin{array}{l} x \\ y \\ z \\ 1 \end{array} \right), \tag {5} +$$ + +where $K$ is intrinsic matrix, $E$ is extrinsic matrix. As the projected point coordinate is a real value and the feature map only consists of regular grids, we use a local area around the projected point $(u,v)$ to approximate the coordinates. We then use the local area to compute local features $F_{l}$ through bilinear interpolation and multi-layer perception + +![](images/88465d891abf48f72971b568e0cab8c93cba3047e68a1471d7775bdf1ad7d47c.jpg) +Figure 4: The pipeline of aggregation. Without loss of generality, we take the aggregation on candidate point features as an example to visualize the pipeline of aggregation. + +MLP). The extracted local feature for sparse points is indicated as $F_{l} \in \mathbb{R}^{C \times N \times M \times ||S||}$ , where $C$ is the feature size, $N$ is the number of anchor lanes, $M$ is the number of anchor points and $||S||$ is the sampling size. To guide the fusion of local and global information, we also encode coordinates of sampled sparse points into a high-dimension feature $F_{p} \in \mathbb{R}^{C \times N \times M \times ||S||}$ through a single MLP. + +The point embedding $F_{p}$ and local features $F_{l}$ are then concatenated as candidate point features $F_{c} \in \mathbb{R}^{C \times N \times M \times ||S||}$ . In order to compute anchor point features $F_{a} \in \mathbb{R}^{C \times N \times M}$ from $F_{c}$ , we propose to aggregate candidate point features belonging to the same anchor point. As shown in Figure 4, we gradually compress the candidate point dimension from $||S||$ to 1. During the compression, we use convolution with the kernel size of $3 \times 1$ and padding of $(1,0)$ to propagate context information along the anchor point dimension. This operation constrains the context information to only flow among the anchor points in the same lane. Meanwhile, we use convolution with the kernel size of $1 \times 3$ and padding of $(0,0)$ to compress local information along the candidate point dimension. This operation fuses the local information from the sampled sparse points to the corresponding anchor. These two operations are alternately repeated to propagate and compress information progressively. + +As for the global feature, we compress the entire frontview features into a single feature vector $F_{g} \in \mathbb{R}^{C}$ through pooling operation, following previous methods [24, 27, 4, 20, 8, 41]. We then concatenate the global feature vector $F_{g}$ to each anchor feature vector $F_{a}$ and fuse them together to balance the global and local information. The fused features are then used to update the anchor point location $\{(\Delta_{x_j^i}, z_j^i, v_j^i)\}_{j=0}^{j=M}$ through 3 convolution layers. + +Lane Category Refinement In order to detect the lane categories, we extract lane features by aggregating the feature of anchor points belonging to the same lane. Similar to the above aggregation operation for the anchor point features, we first concatenate the global feature $F_{g}$ with anchor features $F_{a}$ and then gradually propagate the information along the anchor point dimension. The aggregated lane features $F_{l} \in \mathbb{R}^{C \times N}$ are then used to update the lane category $p_{i}$ through several convolution layers. + +# 3.4. Loss + +The predicted anchor-based lane representation consists of category $\tilde{p}_i$ and location $\tilde{l}_i = \{(\tilde{\Delta}_{x_j^i},\tilde{z}_j^i,\tilde{v}_j^i)\}_{j = 0}^{j = M}$ . We use the cross entropy to supervise the predicted category $\tilde{p}$ : + +$$ +\mathcal {L} _ {p} = - \frac {1}{N} \sum_ {i = 0} ^ {i = N - 1} p _ {i} \log \left(\tilde {p} _ {i}\right). \tag {6} +$$ + +As for the location, we supervise the learning of visibility $\tilde{v}$ by $l_{1}$ loss and use the visibility as weights to compute the $l_{1}$ loss for $\tilde{\Delta}_x$ and $\tilde{z}$ : + +$$ +\mathcal {L} _ {\Delta} = \frac {1}{M \cdot N} \sum_ {i = 0} ^ {i = N - 1} \sum_ {j = 0} ^ {j = M - 1} v _ {j} ^ {i} | | \Delta_ {x _ {j} ^ {i}} - \tilde {\Delta} _ {x _ {j} ^ {i}} | | _ {1}, \tag {7} +$$ + +$$ +\mathcal {L} _ {z} = \frac {1}{M \cdot N} \sum_ {i = 0} ^ {i = N - 1} \sum_ {j = 0} ^ {j = M - 1} v _ {j} ^ {i} | | z _ {j} ^ {i} - \tilde {z} _ {j} ^ {i} | | _ {1}, \tag {8} +$$ + +$$ +\mathcal {L} _ {v} = \frac {1}{M \cdot N} \sum_ {i = 0} ^ {i = N - 1} \sum_ {j = 0} ^ {j = M - 1} | | v _ {j} ^ {i} - \tilde {v} _ {j} ^ {i} | | _ {1}. \tag {9} +$$ + +The joint loss can be indicated as + +$$ +\mathcal {L} = \mathcal {L} _ {p} + \mathcal {L} _ {\Delta} + \mathcal {L} _ {z} + \mathcal {L} _ {v}. \tag {10} +$$ + +During the hierarchical refinement from the low resolution to the high resolution, we separately compute the above loss for each resolution and sum up the losses as the final loss to train the network. + +# 4. Experiments + +In the main paper, we present the sparse-point guided 3D lane detection method on the anchor-based approach. We mainly use Persformer [3] as the baseline, while Gen-LaneNet [7] is taken as an optional baseline to verify the feasibility of our method. We validate our method on the real-world datasets OpenLane [3] and ONCE-3DLanes [18] to demonstrate our priority to handle complex environments and variable scenes. As for our method on the segmentation-based approach, we present the results obtained on the synthetic dataset Apollo 3D Lane [7] in the Supplementary Material. The code is available at https://github.com/YaoChengTang/Sparse-Point-Guided-3D-Lane-Detection + +
MethodAll ↑Up & Down ↑Curve ↑Extreme Weather ↑Night ↑Intersection ↑Merge & Split ↑
3D-LaneNe [6]44.140.846.547.541.532.141.7
Gen-LaneNet [7]32.325.433.528.118.721.431
PersFormer [3]50.542.455.648.646.64050.7
Ours + PersFormer53.7 (3.2↑)46.2 (3.8↑)59.2 (3.6↑)54.8 (6.2↑)49.8 (3.2↑)41.9 (1.9↑)52.1 (1.4↑)
+ +Table 1: Performance comparison with state-of-the-art methods on OpenLane benchmark in different scenarios, where the evaluation metric is F1-score. $\uparrow$ means that the larger the value, the better the result. + +
MethodF-score (%) ↑X error near (m) ↓X error far (m) ↓Z error near (m) ↓Z error far (m) ↓
3D-LaneNe [6]44.10.4790.5720.3670.443
Gen-LaneNet [7]32.30.5910.6840.4110.521
Cond-IPM [3]36.60.5631.0800.4210.892
PersFormer [3]50.50.4850.5530.3640.431
Gen-LaneNet* [7]42.80.4880.6320.3740.481
Ours + Gen-LaneNet*46.60.4750.5770.3710.445
Ours + PersFormer52.30.4680.5140.3710.418
+ +# 4.1. Datasets + +OpneLane Dataset OpneLane [3] is a challenging real-world 3D lane dataset constructed on Waymo Open dataset [31]. It consists of 200K frames with 14 kinds of categories, complex lane structures, and five kinds of weather. $50\%$ frames have an altitude change of more than $1m$ , and $25\%$ frames have more than six lanes. Currently, it is the most challenging lane detection dataset, so we choose it as the main dataset to conduct ablation studies and demonstrate the performance promotion of our method. + +ONCE-3DLanes Dataset ONCE-3DLanes [18] is a real-world 3D lane dataset constructed on ONCE [23]. It consists of various scenes, such as highways, bridges, tunnels, suburbs, and downtown, with different weather conditions (sunny/rainy) and lighting conditions (day/night). + +# 4.2. Implementation Details + +We select the PersFormer [3] as the baseline of the anchor-based approach to leverage the sparse point. EfficientNet [34] is used as the backbone for the front-view feature extraction. Perspective Transformer [3] and 3D-GeoNet [7] are used in coarse-level lane detection to extract the dense BEV features and detect coarse 3D lanes at the lowest resolution. Following PersFormer, we set the highest resolution of BEV space as $208 \times 128$ with the range of $[-10, 10]$ meters along the x-axis and [3, 101] meters along the y-axis, respectively. We set the count of anchor lanes as $N = 182$ and anchor points for each lane as $M = 10$ , where the fixed $\{y^i\}$ is set as $\{5, 10, 15, 20, 30, 40, 50, 60, 80, 100\}$ for each lane. In fine-level sparse point refinement, we set the size of sampling window $\mathcal{W}$ as $3 \times 3$ with $(1, 0.5)$ sampling step along the x-axis and the z-axis separately. + +Table 2: Performance comparison with state-of-the-art methods using different evaluation metrics on the validation set of OpenLane benchmark. * means that we re-implement the method on the OpenLane benchmark. ↑ means that the larger the value, the better the result. ↓ means that the smaller the value, the better the result. + +
MethodTraining +Memory (G)Inference +Speed (FPS)
3D-LaneNe [6]6.6103
Gen-LaneNet [7]4.767
PersFormer [3]19.627
Ours + PersFormer12.545
+ +Table 3: Training memory cost and inference speed comparison with state-of-the-art approaches. We acquire the memory cost on a single A30 GPU with a batch size of 8. We obtain the inference speed on a single A30 GPU with a batch size of 1. + +In the training stage, we pre-train the coarse-level lane detection for 100 epochs and then finetune our fine-level sparse point refinement for another 100 epochs. For both pertaining and finetuning, we use Adam optimizer [12] with a base learning rate of $2 \times 10^{-4}$ and a weight decay of $10^{-3}$ . The cosine learning rate policy is applied to optimization with a maximum iteration of 8 and a minimum learning rate of $10^{-5}$ . We also use gradient norm clipping during the optimization with a maximum norm weight of 35. In the Inference stage, we use the number of valid anchor points in each lane as the threshold to filter anchor lanes that are too close. For more details about training and inference, please refer to the code and our supplemental materials. + +# 4.3. Evaluation Metrics + +Following OpenLane, we use bipartite matching to evaluate the predicted and ground-truth lanes. The matching is true positive when the distance error of $75\%$ lane points is less than $1.5m$ . The percentage of matched ground-truth lanes is deemed as the recall, and the percentage of matched predicted lanes is deemed as the precision. The average of recall and precision is F-score. The lane points are also divided into near points $(3 - 40m)$ and far points $(40 - 101m)$ . + +![](images/afd6c05e45ebccd60776c8079199ce0dc670e7df44f27a6db1ddad4ac81a24f6.jpg) +Figure 5: The visualization results of PersFormer and our method on the OpneLane dataset. + +The point-wise distance error along the x-axis and z-axis is used to evaluate the location accuracy for near and far points. In ONCE, we also use the unilateral Chamfer Distance (CD) metric that computes the chamfer distance between the predicted and ground truth lane. + +# 4.4. Benchmark Performance + +OpenLane Following the setting in OpenLane, we evaluate the performance in different scenarios. As shown in Table 1, we achieve the best results in all scenarios. It is worth noting that our method achieves more than 6 points promotion in extreme weather scenarios. This promotion shows that our method greatly resists the feature noise and extracts valuable information by only focusing on sparse points around 3D lanes. In Table 2, we further compare our method with state-of-the-art approaches using different evaluation metrics. Our method outperforms all state-of-the-art approaches in almost every evaluation metric. Specifically, we improve the location on the x-axis in both near and far distances and thus improve the overall F1-score. What's more, our method effectively helps Gen-LaneNet outperform both the original result of Gen-LaneNet published by OpenLane and our reproduced Gen-LaneNet* on all metrics. These results prove that our method is flexible and can seamlessly integrate with different prototypical methods to offer consistent improvement. + +We also compare our method with state-of-the-art approaches from the training memory cost and the inference + +speed. As shown in Table 3, our method achieves the lowest memory cost and the fastest inference speed. Compared to the baseline, we greatly reduce the memory cost, which enables the front-view images to have higher resolution. Furthermore, the sparse point guides the network to achieve a better result in most cases with a much faster inference speed. It should be noted that the speed and memory cost reported in this paper is computed based on a pure Python realization of sparse-point sampling and aggregation without any unfair engineering optimization like the custom CUDA kernel that significantly reduces the running time. + +Besides the quantitative analysis, we also compare visualization results of the PersFormer and our method in different scenarios for qualitative analysis. As shown in Figure 5, our method obviously outperforms the PersFormer in cloudy, rainy, night, and different road conditions. This further demonstrates that the sparse point guides the BEV feature to extract more meaningful information from the front-view image and thus improve the recall and accuracy of the detected lanes. + +ONCE-3DLanes As illustrated in Table 4, we follow the PersFormer using the validation set of ONCE-3DLanes to evaluate the results of our methods. Considering scenes are easier in the ONCE benchmark, the quantity results are already very high. The sparse point slightly promotes the baseline in the F1-score, Precision, and CD error with a much more efficient performance, which only costs half the memory and inference time. + +
MethodF1-score (%) ↑Precision (%) ↑Recall (%) ↑CD error (m) ↓
3D-LaneNe [6]44.7361.4635.160.127
Gen-LaneNet [7]45.5963.9535.420.121
SALAD [3]64.0775.9055.420.098
PersFormer [3]74.3380.3069.180.074
Ours + PersFormer74.7981.8568.860.070
+ +Table 4: Performance comparison with state-of-the-art methods on ONCE benchmark. $\uparrow$ means that the larger the value, the better the result. $\downarrow$ means that the smaller the value, the better the result. + +
Front-viewFeature ExtractionCoarse levelFine level 1Fine level 2Fine level 3
4.8ms6.1ms3.8ms3.1ms3.4ms
+ +Table 5: Inference time cost of each step in our method. + +
Coarse levelFine level 1Fine level 2Fine level 3F1-score (%)X error near (m)
✓51.30.484
✓✓52.10.475
✓✓✓53.60.470
✓✓✓✓53.70.468
+ +# 4.5. Ablation Study and Discussion + +Complexity Analysis We evaluate the time cost of each step in our method to show the efficiency of our sparse refinement. As shown in Table 5, it is more efficient to use sparse points for feature fetching and 3d lane refinement, even compared to the coarse-level lane detection conducted in the lowest resolution. + +Effectiveness of Refinement In order to validate the effectiveness of our fine-level sparse point refinement, we compare results from the coarse level and from the different fine levels. As shown in Table 6, the fine-level sparse point refinement greatly improves the results of coarse-level lane detection in both F1-score and X error near. + +Effectiveness of Features To ensure globally smooth and locally discriminative performance, we leverage both the global and local features to refine the lanes. We demonstrate the advantage of balancing the global and local information by comparing three kinds of feature conditions, including only using global features, only using local features, and using both local and global features. As shown in Table 7, the first condition results in a very low F1 score of $46.8\%$ , and the second one is already comparable to the baseline, while the last one fusing with sparse points gives the best result. + +Effectiveness of Point Sampling As aforementioned, we sample candidate points around the anchor point on the x-z plane. Here, we analyze the effectiveness of point sampling using different sampling steps. + +As illustrated in Table 8, the result gets better when the value of the sampling step is smaller along the $z$ -axis and larger along the $x$ -axis. This phenomenon shows that the local features fetched along the $x$ -axis are more valuable + +Table 6: The performance of coarse-level lane detection and fine-level sparse point refinement. + +
Global FeaturesLocal FeaturesF1-score (%)
✓46.8
✓51.8
✓✓53.7
+ +Table 7: The performance of our method with/without global features and local features. + +
x-axisz-axisF1-score
1153.61
10.553.65
0.5153.34
0.50.553.46
10.2553.67
+ +Table 8: The performance of our method with different sampling steps along x-axis and z-axis. + +than the features fetched along the z-axis, where the sampling along the x-axis actually gives us more valid sample points on the front-view image. + +# 4.6. Limitations and Discussion + +As shown in the above experiments, we have achieved great progress in both accuracy and efficiency. However, our method still has some limitations, e.g., assuming all 3D lane instances could be detected on BEV features in the lowest resolution. This assumption becomes less powerful when meeting extremely complex and crowded road conditions. A slow-fast update strategy for refinement at different scales might give better results, but it requires much more engineering tricks on feature manipulation. In this paper, we mainly focus on proving the priority and possibility of simply using sparse correspondence instead of dense correspondence for 3D lane detection. + +# 5. Conclusion + +In this paper, we have proved that using sparse points is more efficient and adequate for high-quality 3D lane detection than building a dense correspondence between the HR front-view space and BEV space. We presented a sparse point-guided 3D lane detection method, including coarse-level lane detection and fine-level sparse point refinement. Experiments proved that our method could achieve much better results with less memory and time cost. + +Acknowledgement This work was supported by the Natural Science Foundation of China (NSFC) under Grants No.62176021. + +# References + +[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. arXiv preprint arXiv:2209.07989, 2022. +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022. +[3] Li Fei Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. +[4] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5932-5941, 2019. +[5] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17041-17049, 2022. +[6] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2921-2930, 2019. +[7] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. +[8] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and Shimin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8928-8937, 2021. +[9] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zihan Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. arXiv preprint arXiv:2301.02371, 2023. +[10] Yujie Jin, Xiangxuan Ren, Fengxiang Chen, and Weidong Zhang. Robust monocular 3d lane detection with dual attention. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 3348-3352, 2021. +[11] Longlong Jing, Yucheng Chen, and Yingli Tian. Coarse-to-fine semantic segmentation from image-level labels. IEEE Transactions on Image Processing (TIP), 29:225-236, 2019. +[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015. +[13] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 9799-9808, 2020. + +[14] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Taeyeop Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1965-1973, 2017. +[15] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4369-4378, 2022. +[16] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Transactions on Intelligent Transportation Systems, 21(1):248-258, 2019. +[17] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condilanenet: a top-to-down lane detection framework based on conditional convolution. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 3773-3782, 2021. +[18] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Zejian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1765-1772, 2022. +[19] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3694-3702, 2021. +[20] Shilin Liu, Haoxiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1788-1797, 2021. +[21] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. +[22] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 11677-11684, 2020. +[23] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Chunjing Xu, et al. One million scenes for autonomous driving: Once dataset. 2021. +[24] Lars M. Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4455-4465, 2019. +[25] Roozbeh Mottaghi, Yu Xiang, and Silvio Savarese. A coarse-to-fine model for 3d pose estimation and sub-category recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 418-426, 2015. + +[26] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaouu Tang. Spatial as deep: Spatial cnn for traffic scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 7276-7283, 2018. +[27] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. +[28] Marco Pedersoli, Andrea Vedaldi, Jordi Gonzalez, and Xavier Roca. A coarse-to-fine approach for fast deformable object detection. Pattern Recognition (PR), 48(5):1844-1853, 2015. +[29] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 276-291. Springer, 2020. +[30] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 81-90, 2020. +[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2446-2454, 2020. +[32] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021. +[33] Lucas Tabelini, Rodrigo Berriel, Thiago Meireles Paixão, Claudine Santos Badue, Alberto Ferreira de Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 294-302, 2021. +[34] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 6105-6114. PMLR, 2019. +[35] Wouter Van Gansbeke, Bert De Brabandere, Davy Neven, Marc Proesmans, and Luc Van Gool. End-to-end lane detection through differentiable least-squares fitting. In Proceedings of the IEEE International Conference on Computer Vision Workshops (CVPRW), pages 0–0, 2019. +[36] Jun Wang, Shiyi Lan, Mingfei Gao, and Larry S Davis. Info-focus: 3d object detection for autonomous driving with dynamic information modeling. In European Conference on Computer Vision, pages 405-420. Springer, 2020. +[37] Hang Xu, Shaoju Wang, Xinyue Cai, Wei Zhang, Xiaodan Liang, and Zhenguo Li. Curvelane-nas: Unifying lanesensitive architecture search and adaptive point blending. In Proceedings of the European Conference on Computer Vision (ECCV), pages 689-704. Springer, 2020. + +[38] Fan Yan, Ming-Jun Nie, Xinyue Cai, Jianhua Han, Han Xu, Zhengxin Yang, Chao Ye, Yanwei Fu, Michael B. Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17122–17131, 2022. +[39] Chengtang Yao, Yunde Jia, Huijun Di, Pengxiang Li, and Yuwei Wu. A decomposition model for stereo matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6091-6100, 2021. +[40] Jian Yao, Marko Boben, Sanja Fidler, and Raquel Urtasun. Real-time coarse-to-fine topologically preserving segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2947-2955, 2015. +[41] Jianglong Ye, Yuntao Chen, Naiyan Wang, and X. Wang. Gifs: Neural implicit function for general shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12819-12829, 2022. +[42] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. Resa: Recurrent feature-shift aggregator for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3547-3554, 2021. +[43] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 898-907, 2022. \ No newline at end of file diff --git a/2023/Sparse Point Guided 3D Lane Detection/images.zip b/2023/Sparse Point Guided 3D Lane Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..51d3ebe64566e92ba0d8f1407887684be54d1075 --- /dev/null +++ b/2023/Sparse Point Guided 3D Lane Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75e08f1d8c0829af604945fb8932bb6f3ab1e77daed2f476908962c40fb85eb8 +size 459620 diff --git a/2023/Sparse Point Guided 3D Lane Detection/layout.json b/2023/Sparse Point Guided 3D Lane Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f3ab6b64cfb17b975e6db1624002b40165396aa --- /dev/null +++ b/2023/Sparse Point Guided 3D Lane Detection/layout.json @@ -0,0 +1,8228 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 173, + 103, + 421, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 103, + 421, + 121 + ], + "spans": [ + { + "bbox": [ + 173, + 103, + 421, + 121 + ], + "type": "text", + "content": "Sparse Point Guided 3D Lane Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "spans": [ + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "text", + "content": "Chengtang Yao" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "text", + "content": ", Lidong Yu" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "text", + "content": ", Yuwei Wu" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "text", + "content": ", Yunde Jia" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "text", + "content": "Beijing Key Laboratory of Intelligent Information Technology, School of Computer Science & Technology, Beijing Institute of Technology, China \n" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "text", + "content": "Guangdong Laboratory of Machine Perception and Intelligent Computing, Shenzhen MSU-BIT University, China \n" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 142, + 497, + 228 + ], + "type": "text", + "content": "Autonomous Driving Algorithm, DeepRoute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 230, + 511, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 230, + 511, + 241 + ], + "spans": [ + { + "bbox": [ + 82, + 230, + 511, + 241 + ], + "type": "text", + "content": "yao.c.t@bit.edu.cn, yvlidong@gmail.com, jiayunde@bit.edu.cn, wuyuwei@bit.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 294, + 290, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 290, + 582 + ], + "type": "text", + "content": "3D lane detection usually builds a dense correspondence between the front-view space and the BEV space to estimate lane points in the 3D space. 3D lanes only occupy a small ratio of the dense correspondence, while most correspondence belongs to the redundant background. This sparsity phenomenon bottlenecks valuable computation and raises the computation cost of building a high-resolution correspondence for accurate results. In this paper, we propose a sparse point-guided 3D lane detection, focusing on points related to 3D lanes. Our method runs in a coarse-to-fine manner, including coarse-level lane detection and iterative fine-level sparse point refinements. In coarse-level lane detection, we build a dense but efficient correspondence between the front view and BEV space at a very low resolution to compute coarse lanes. Then in fine-level sparse point refinement, we sample sparse points around coarse lanes to extract local features from the high-resolution front-view feature map. The high-resolution local information brought by sparse points refines 3D lanes in the BEV space hierarchically from low resolution to high resolution. The sparse point guides a more effective information flow and greatly promotes the SOTA result by 3 points on the overall F1-score and 6 points on several hard situations while reducing almost half memory cost and speeding up 2 times." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 590, + 128, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 128, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 128, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 609, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 695 + ], + "type": "text", + "content": "3D lane detection is an indispensable part of the advanced driver assistance system, supporting functionalities such as automated lane centering and lane departure warning. It relies on building a dense correspondence between the front view space and BEV space to localize lane points in 3D space. Previous methods build a dense correspondence by directly projecting the 2D image or 2D fea" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 268, + 430, + 372 + ], + "blocks": [ + { + "bbox": [ + 309, + 268, + 430, + 372 + ], + "lines": [ + { + "bbox": [ + 309, + 268, + 430, + 372 + ], + "spans": [ + { + "bbox": [ + 309, + 268, + 430, + 372 + ], + "type": "image", + "image_path": "b850b6ceb7fd0f7564837946446164bdbe7a844e1d5570f36ca0ea8875a86045.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 380, + 409, + 390 + ], + "lines": [ + { + "bbox": [ + 328, + 380, + 409, + 390 + ], + "spans": [ + { + "bbox": [ + 328, + 380, + 409, + 390 + ], + "type": "text", + "content": "(a) Dense Correspondence" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 399, + 547, + 448 + ], + "lines": [ + { + "bbox": [ + 305, + 399, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 547, + 448 + ], + "type": "text", + "content": "Figure 1: The visualization of dense correspondence (a) and sparse correspondence (b) between font view and BEV. The green grids represent areas near lanes, while blue grids are redundant areas far from lanes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 432, + 268, + 555, + 372 + ], + "blocks": [ + { + "bbox": [ + 432, + 268, + 555, + 372 + ], + "lines": [ + { + "bbox": [ + 432, + 268, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 432, + 268, + 555, + 372 + ], + "type": "image", + "image_path": "317220b805f7b838ecdb6615da0c56e7da9430b85430152d2d39a8877fa4a3a1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 380, + 533, + 390 + ], + "lines": [ + { + "bbox": [ + 451, + 380, + 533, + 390 + ], + "spans": [ + { + "bbox": [ + 451, + 380, + 533, + 390 + ], + "type": "text", + "content": "(b) Sparse Correspondence" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 458, + 547, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 566 + ], + "type": "text", + "content": "ture map into 3D space using Inverse Perspective Mapping (IPM) [18, 38], but this kind of correspondence is hard to handle the complex road conditions (e.g., uphill and downhill). Recently, some methods [3, 21] build a dense correspondence using a learnable transformer to resolve this problem. The transformer-based correspondence builds a more effective information flow from the front view space to the BEV space, which results in better performance in extreme scenes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": "Building such dense correspondence is actually redundant for 3D lane detection. As shown in Figure 1, even among a local window near the lane, only " + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "inline_equation", + "content": "2/5" + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": " correspondences are beneficial for 3D lane detection, while the others are redundant. Among the dense correspondence, 3D lanes only occupy a small ratio, while most correspondences are built for the background. This redundant problem becomes more severe at high resolution, where both the high-resolution front-view and BEV feature maps are necessary for high-quality lane detection in most methods [6, 7, 18, 10, 38, 15]. To this end, we propose to construct a sparse correspondence only focusing on a limited num" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8363" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 178 + ], + "type": "text", + "content": "ber of points around 3D lanes. The sparse correspondence builds an effective yet efficient information flow from the front-view space to the BEV space. It directly brings the high-resolution lane information from the front-view space, ensuring fine-grained details in high-resolution BEV space. Without redundant correspondence, the information flow learns to focus more on features beneficial for 3D lane results. Meanwhile, it has a naturally efficient performance by only allocating computation to points related to lanes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 181, + 286, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 181, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 181, + 286, + 442 + ], + "type": "text", + "content": "In this paper, we present a sparse-point-guided 3D lane detection method that decomposes the 3D lane detection into coarse-level lane detection and fine-level sparse point refinement. In coarse-level lane detection, we extract multiscale feature maps from the front-view image. An efficient but dense BEV feature map is built from the front-view feature map at the lowest resolution to detect coarse 3D lanes. Then the fine-level sparse point refinement refines coarse results hierarchically from low resolution to high resolution. In each refinement, we first sample sparse points around coarse 3D lanes within a specific window. And then, we project sampled points onto the front-view plane to extract local features from the high-resolution front-view feature map. The local feature provides fine-grained information to refine the local structure of 3D lanes. At the same time, we compress the front-view feature map into a single vector as the global features to ensure the global smoothness of 3D lanes. We fuse the global feature, the local features, and the sparse point coordinates to predict the location and category of each lane. The fusion of the local and global features refers to the point coordinates, which ensures both the global smoothness and local discrimination of lanes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 445, + 286, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 286, + 623 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 286, + 623 + ], + "type": "text", + "content": "We validate our method on two anchor-based approaches and a widely used segmentation-based approach. The demonstrations are conducted on two real-world datasets (OpenLane [3] and ONCE [18]), including different weathers, lane structures, and road conditions. Our method outperforms the two anchor-based approaches, showing our feasibility to be seamlessly integrated with different prototypical methods to offer consistent improvement. Besides, our method simply outperforms the SOTA anchor-based approach by 2 points on the overall F1-score and more than 6 points on the F1-score in several extreme conditions while reducing half memory cost and speeding up 2 times. As for the segmentation-based approach, our method achieves comparable performance and reduces the memory cost of the 3D lane head by " + }, + { + "bbox": [ + 49, + 445, + 286, + 623 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 49, + 445, + 286, + 623 + ], + "type": "text", + "content": " and speeds up 2 times." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 637, + 132, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 637, + 132, + 649 + ], + "spans": [ + { + "bbox": [ + 49, + 637, + 132, + 649 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 658, + 139, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 139, + 669 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 139, + 669 + ], + "type": "text", + "content": "2.1. Lane Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 677, + 286, + 712 + ], + "type": "text", + "content": "According to the lane representation, there are mainly three kinds of approaches for lanes detection, including segmentation-based [14, 26, 37, 42], anchor-based [33, 43," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 72, + 544, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 239 + ], + "type": "text", + "content": "17, 16, 29], and parameter-based methods[5, 35, 32, 19]. In order to estimate the lane representation in the 3D space, some methods first conduct the estimation in the front-view space and then use inverse perspective mapping (IPM) to bridge the front-view space to BEV space for 3D results. The IPM correspondence is based on the planar road assumption, which is hard to handle road undulates, like a hilly road. To tackle this issue, some methods estimate the 3D lanes in the BEV space to resolve complex road conditions [6, 7, 18, 10, 38, 15, 3, 1, 9, 21]. They usually build a dense correspondence by transforming the front-view feature to the BEV feature map. The dense correspondence builds an effective information flow for the BEV feature to improve the quality of the 3D lane." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 240, + 544, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 544, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 544, + 502 + ], + "type": "text", + "content": "To enhance the BEV feature, some methods introduce auxiliary tasks for the joint learning[6, 7, 18, 38, 15]. The others focus on designing a more effective transformation [10, 3, 15]. RobustLane [10] uses an attention mechanism to aggregate better global information and thus provides a better global smoothness. RTVLane [15] introduces a geometry consistency between 2D and 3D space to guide the learning of BEV features. Recently, PersFormer [3] uses a transformer to build the dense correspondence between the multi-scale front-view features and BEV feature through cross-attention, which achieves SOTA performance. Building a proper correspondence between the front-view space and the BEV space is the key to improving the 3D lane performance, but a dense correspondence is actually redundant. We observe that lanes only occupy a very small ratio of the dense correspondence. Thus, we build a sparse correspondence that only focuses on the points mostly belonging to a lane. The sparse correspondence efficiently brings the high-resolution lane information from the front-view space, ensuring fine-grained details in high-resolution BEV space. Besides, it also guides the information flow to learn better features beneficial for 3D lane results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 514, + 440, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 440, + 526 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 440, + 526 + ], + "type": "text", + "content": "2.2. Coarse-to-Fine Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 534, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 544, + 712 + ], + "type": "text", + "content": "The coarse-to-fine design is widely adopted in many methods to hierarchically improve results [43, 42, 13, 39, 30, 24, 28, 40, 11, 22, 2, 36, 25]. The general design consists of a coarse result estimation using low-frequency information and a hierarchical refinement using high-frequency information at high resolution. Due to the sparsity of high-frequency information, many methods [13, 39, 30, 24] propose to only focus on the valuable regions and operate on sparse point sets rather than regular grids to keep the balance between efficiency and accuracy. They mainly focus on common objects and use local features to recover fine-grained structures, like corners and boundaries. Different from common objects, 3D lanes have a uniquely long and thin structure, requiring a globally smooth and locally accurate estimation. In this paper, we fuse the local and global" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "8364" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 532, + 155 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 532, + 155 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 532, + 155 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 532, + 155 + ], + "type": "image", + "image_path": "a682de574565aa88ae80e3ff8deadbb3e6fcaca559687446d822143db9f29bd8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 157, + 364, + 167 + ], + "lines": [ + { + "bbox": [ + 228, + 157, + 364, + 167 + ], + "spans": [ + { + "bbox": [ + 228, + 157, + 364, + 167 + ], + "type": "text", + "content": "(a) The overall pipeline of our model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 161, + 171, + 433, + 232 + ], + "blocks": [ + { + "bbox": [ + 161, + 171, + 433, + 232 + ], + "lines": [ + { + "bbox": [ + 161, + 171, + 433, + 232 + ], + "spans": [ + { + "bbox": [ + 161, + 171, + 433, + 232 + ], + "type": "image", + "image_path": "92f88d94b97f443e5f2413b294cb85efa0aeb71296b69148d09c05b7a8eef205.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 237, + 361, + 247 + ], + "lines": [ + { + "bbox": [ + 233, + 237, + 361, + 247 + ], + "spans": [ + { + "bbox": [ + 233, + 237, + 361, + 247 + ], + "type": "text", + "content": "(b) The coarse-level lane detection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 86, + 254, + 507, + 352 + ], + "blocks": [ + { + "bbox": [ + 86, + 254, + 507, + 352 + ], + "lines": [ + { + "bbox": [ + 86, + 254, + 507, + 352 + ], + "spans": [ + { + "bbox": [ + 86, + 254, + 507, + 352 + ], + "type": "image", + "image_path": "0b2ac5337bdf41ff33b82a8918643a070208102da0ab156c3335ff470e166e62.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 357, + 425, + 369 + ], + "lines": [ + { + "bbox": [ + 167, + 357, + 425, + 369 + ], + "spans": [ + { + "bbox": [ + 167, + 357, + 425, + 369 + ], + "type": "text", + "content": "(c) The fine-level sparse point refinement. " + }, + { + "bbox": [ + 167, + 357, + 425, + 369 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 167, + 357, + 425, + 369 + ], + "type": "text", + "content": " represents concatenation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 377, + 546, + 402 + ], + "lines": [ + { + "bbox": [ + 46, + 377, + 546, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 546, + 402 + ], + "type": "text", + "content": "Figure 2: The overall structure of our model is shown in (a). We decompose the 3D lane detection into the coarse-level lane detection shown in (b) and the fine-level sparse point refinement shown in (c)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "type": "text", + "content": "information referring to the coordinates of sparse points and jointly refine the global and local structures of lanes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 447, + 102, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 102, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 102, + 459 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 468, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 288, + 684 + ], + "type": "text", + "content": "As 3D lanes only occupy a small ratio among the dense correspondence between the front-view space and BEV space, we propose only focusing on sparse points related to lanes in the high-resolution space. As shown in Figure 2a, we decompose the 3D lane detection into a coarse-level lane detection in the lowest resolution and a series of fine-level sparse point refinement. We extract multi-scale front-view features from a front-view image. The lowest-resolution features are fed into coarse-level lane detection to acquire a rough structure of 3D lanes. The other high-resolution features are used to refine the structure of 3D lanes hierarchically from low resolution to high resolution in fine-level sparse point refinement. As shown in Figure 2b, in coarse-level lane detection, we first build a dense but efficient correspondence between the low-resolution front-view feature and the BEV feature. The low-resolution BEV feature is extracted to estimate the anchor representation of 3D lanes *. As for the fine-level sparse point refinement shown in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 309, + 411, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 411, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 309, + 411, + 545, + 483 + ], + "type": "text", + "content": "Figure 2c, candidate points are first sampled around the anchor points of a lane. The candidate point coordinates are then used as indexes to fetch local features from front-view features. We also compress the front-view features into a single vector as global features, which are fused with local features to refine the location and category of lanes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 490, + 425, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 425, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 425, + 502 + ], + "type": "text", + "content": "3.1. Lane Representation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "text", + "content": "Given an image, 3D lane detection aims to detect the location " + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "text", + "content": " and category " + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "text", + "content": " of lanes " + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "inline_equation", + "content": "L = \\{l_i,p_i\\}_{i = 0}^{i = N - 1}" + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "text", + "content": " is the number of lanes. The location of lane " + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 306, + 509, + 545, + 557 + ], + "type": "text", + "content": " is represented by an ordered point sequence:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 369, + 564, + 545, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 564, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 369, + 564, + 545, + 582 + ], + "type": "interline_equation", + "content": "l _ {i} = \\left\\{\\left(x _ {j} ^ {i}, y _ {j} ^ {i}, z _ {j} ^ {i}\\right) \\right\\} _ {j = 0} ^ {j = M _ {i} - 1}. \\tag {1}", + "image_path": "41efdee680caa10b3b483481d32d156a998947c47dc62ff40e844940cb843dd0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 590, + 419, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 419, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 419, + 601 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 306, + 590, + 419, + 601 + ], + "type": "text", + "content": " is the number of points." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": "In anchor-based representation [3, 6, 7], 3D lanes are initialized by a set of anchor lanes with predefined x-coordinates " + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\{\\bar{x}^i\\}_{i = 0}^{i = N - 1}" + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": " is the number of lanes. The anchor lanes are further defined by a series of anchor points with predefined y-coordinates " + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\{\\bar{y}_j\\}_{j = 0}^{j = M}" + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": " is the anchor point count for each anchor lane. In the anchor-based representation, the location of the 3D lane is formu" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "*We select the anchor-based approach, i.e., Persformer[3], as the baseline in the main paper. The realization of sparse-point guided lane detec" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": "tion for the segmentation-based approach is shown in the Supplementary Material with modification on Persformer backbone." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8365" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 80, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 80, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 80, + 83 + ], + "type": "text", + "content": "lated as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 81, + 287, + 100 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 81, + 287, + 100 + ], + "spans": [ + { + "bbox": [ + 113, + 81, + 287, + 100 + ], + "type": "interline_equation", + "content": "l _ {i} ^ {\\prime} = \\left\\{\\left(\\Delta_ {x _ {j} ^ {i}}, z _ {j} ^ {i}, v _ {j} ^ {i}\\right) \\right\\} _ {j = 0} ^ {j = M}, \\tag {2}", + "image_path": "0a7f7797e6587693f141bf5e3e0476e53eb1fa7462ba2d540f3acb6c9f7f71fb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 103, + 287, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 287, + 155 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 287, + 155 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 103, + 287, + 155 + ], + "type": "inline_equation", + "content": "\\Delta_{x_j^i}" + }, + { + "bbox": [ + 47, + 103, + 287, + 155 + ], + "type": "text", + "content": " is the offset along the x-axis, " + }, + { + "bbox": [ + 47, + 103, + 287, + 155 + ], + "type": "inline_equation", + "content": "z_j^i" + }, + { + "bbox": [ + 47, + 103, + 287, + 155 + ], + "type": "text", + "content": " is the absolute height along the z-axis and " + }, + { + "bbox": [ + 47, + 103, + 287, + 155 + ], + "type": "inline_equation", + "content": "v_{j}^{i}" + }, + { + "bbox": [ + 47, + 103, + 287, + 155 + ], + "type": "text", + "content": " is the visibility for each anchor point. The anchor-based representation can be easily converted into the ordered point sequence representation by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 159, + 287, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 159, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 99, + 159, + 287, + 177 + ], + "type": "interline_equation", + "content": "l _ {i} = \\left\\{\\left(\\bar {x} ^ {i} + \\Delta_ {x _ {j} ^ {i}}, \\bar {y} _ {j}, z _ {j} ^ {i} \\right\\} _ {j = 0} ^ {j = M - 1}. \\right. \\tag {3}", + "image_path": "4c72fd55221ac0adc39e596571ebf350db41859ff053020587ace0b51fb3a925.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 182, + 201, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 201, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 201, + 194 + ], + "type": "text", + "content": "3.2. Coarse-level Lane Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 201, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 288, + 369 + ], + "type": "text", + "content": "In coarse-level lane detection, we aim to efficiently acquire the coarse location and category of lanes. We first extract the multi-scale features from the front-view image. Then we transform the feature map from the front view to the BEV at the lowest resolution and compute the coarse location and category of 3D lanes, as shown in Figure 2b. Specifically, we build a dense correspondence between the BEV and front view space at the lowest resolution. We then use the front-view points as the index to fetch front-view features for BEV feature extraction. The details of transformation follow the realization of the Persformer [3] and are discussed in Section 4.2. With the dense BEV feature map, we use a standard anchor-based approach [3, 6, 7] to estimate the anchor-based lane representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 375, + 234, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 375, + 234, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 375, + 234, + 387 + ], + "type": "text", + "content": "3.3. Fine-level Sparse Point Refinement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 393, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 561 + ], + "type": "text", + "content": "As aforementioned, 3D lanes occupy very little 3D space, which means a large number of 3D points are redundant. Based on the observation, we design a fine-level sparse point refinement, which only deals with sparse points related to lanes instead of the entire BEV space to improve the coarse results. The refinement is carried out hierarchically from the low resolution to the high resolution. We take coarse 3D lanes estimated from the previous lower resolution as reference lanes and sample candidate sparse points around them, as shown in Figure 2c. The sampled sparse points are then used as indexes to extract the high-resolution front-view features. The fine-grained details from the high-resolution space gradually refine the local structure of coarse results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 575, + 194, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 194, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 194, + 588 + ], + "type": "text", + "content": "3.3.1 Candidate Point Sampling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Intuitively, points can be sampled along all three dimensions, i.e., " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", but the number of candidate points will become too large to take. In the anchor-based representation, anchor points are predefined and have fixed values on the y-axis. Thus, we sample points only around the anchor points and only on the x-z plane, as shown in Figure 2c. Specifically, we deem coarse points computed from the previous resolution as reference points. The candidate points are uniformly sampled from the neighbors of reference points within a window " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", which has a predefined" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 79, + 541, + 128 + ], + "blocks": [ + { + "bbox": [ + 310, + 79, + 541, + 128 + ], + "lines": [ + { + "bbox": [ + 310, + 79, + 541, + 128 + ], + "spans": [ + { + "bbox": [ + 310, + 79, + 541, + 128 + ], + "type": "image", + "image_path": "94a5065bc1f735bda1b4c079bbf5fe46d5f8d60738078a5a467d0c646078142a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 134, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 304, + 134, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 545, + 194 + ], + "type": "text", + "content": "Figure 3: The pipeline of local feature extraction. We project the 3D point " + }, + { + "bbox": [ + 304, + 134, + 545, + 194 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 304, + 134, + 545, + 194 + ], + "type": "text", + "content": " onto image plane " + }, + { + "bbox": [ + 304, + 134, + 545, + 194 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 304, + 134, + 545, + 194 + ], + "type": "text", + "content": ". A local patch around the point " + }, + { + "bbox": [ + 304, + 134, + 545, + 194 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 304, + 134, + 545, + 194 + ], + "type": "text", + "content": " is fetched to compute local features using interpolation and multi-layer perception (MLP)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 224 + ], + "type": "text", + "content": "size and sampling step. The above sampling process can be indicated as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 233, + 545, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 233, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 315, + 233, + 545, + 251 + ], + "type": "interline_equation", + "content": "S \\left(x _ {j} ^ {i}, y _ {j} ^ {i}, z _ {j} ^ {i}\\right) = \\left\\{\\left(x _ {t}, y _ {j} ^ {i}, z _ {t}\\right) \\mid \\left(x _ {t}, z _ {t}\\right) \\in \\mathcal {W} _ {\\left(x _ {j} ^ {i}, z _ {j} ^ {i}\\right)} \\right\\}. \\tag {4}", + "image_path": "14e6911d311160996fe4bcc9aaccf1f6ed50b2783a074c340b6f0c5e7a7be11e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 257, + 546, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 546, + 426 + ], + "type": "text", + "content": "It is worth noting that the sampling process has a different requirement for training and inference. During the training stage, we require as much information as possible to supervise the learning of anchors. So, we keep all of the candidate points. In contrast, in the inference stage, we need to reduce as many redundant anchor lanes as possible to save time and memory costs. Thus, we filter redundant anchor lanes in the inference stage before sampling. The filtering is designed by three rules. First, the detected categories of lanes are not backgrounds. Second, there are at least two visible points on a lane. Third, the distance of different lanes is far enough, where we define the distance between two lines as the sum of the distance between anchor points with the same " + }, + { + "bbox": [ + 304, + 257, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\bar{y}_j" + }, + { + "bbox": [ + 304, + 257, + 546, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 439, + 388, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 388, + 450 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 388, + 450 + ], + "type": "text", + "content": "3.3.2 Refinement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 457, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 493 + ], + "type": "text", + "content": "Anchor-based lane representation consists of two kinds of properties, including the location of each anchor point and the category of each anchor lane." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 494, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 578 + ], + "type": "text", + "content": "Point Location Refinement Lane has a globally smooth and locally complex structure due to its long and thin shape. To this end, we propose to use both local features of candidate points and global features of the whole image to refine the coarse results. As shown in Figure 3, we project candidate points " + }, + { + "bbox": [ + 304, + 494, + 545, + 578 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 304, + 494, + 545, + 578 + ], + "type": "text", + "content": " onto image plane " + }, + { + "bbox": [ + 304, + 494, + 545, + 578 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 304, + 494, + 545, + 578 + ], + "type": "text", + "content": " for local feature extraction. This process can be written as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 584, + 545, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 584, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 365, + 584, + 545, + 635 + ], + "type": "interline_equation", + "content": "\\left( \\begin{array}{l} u \\\\ v \\\\ 1 \\end{array} \\right) = \\frac {1}{z} \\cdot K \\cdot E \\cdot \\left( \\begin{array}{l} x \\\\ y \\\\ z \\\\ 1 \\end{array} \\right), \\tag {5}", + "image_path": "e610f191b29406a1e780db17435203299fe19063c8098feda039e08aae6e8892.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is intrinsic matrix, " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is extrinsic matrix. As the projected point coordinate is a real value and the feature map only consists of regular grids, we use a local area around the projected point " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " to approximate the coordinates. We then use the local area to compute local features " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " through bilinear interpolation and multi-layer perception" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8366" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 519, + 135 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 519, + 135 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 519, + 135 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 519, + 135 + ], + "type": "image", + "image_path": "88465d891abf48f72971b568e0cab8c93cba3047e68a1471d7775bdf1ad7d47c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 144, + 547, + 169 + ], + "lines": [ + { + "bbox": [ + 46, + 144, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 547, + 169 + ], + "type": "text", + "content": "Figure 4: The pipeline of aggregation. Without loss of generality, we take the aggregation on candidate point features as an example to visualize the pipeline of aggregation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "text", + "content": "MLP). The extracted local feature for sparse points is indicated as " + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "inline_equation", + "content": "F_{l} \\in \\mathbb{R}^{C \\times N \\times M \\times ||S||}" + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "text", + "content": " is the feature size, " + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "text", + "content": " is the number of anchor lanes, " + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "text", + "content": " is the number of anchor points and " + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "inline_equation", + "content": "||S||" + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "text", + "content": " is the sampling size. To guide the fusion of local and global information, we also encode coordinates of sampled sparse points into a high-dimension feature " + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "inline_equation", + "content": "F_{p} \\in \\mathbb{R}^{C \\times N \\times M \\times ||S||}" + }, + { + "bbox": [ + 46, + 176, + 288, + 261 + ], + "type": "text", + "content": " through a single MLP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": "The point embedding " + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": " and local features " + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": " are then concatenated as candidate point features " + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "inline_equation", + "content": "F_{c} \\in \\mathbb{R}^{C \\times N \\times M \\times ||S||}" + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": ". In order to compute anchor point features " + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "inline_equation", + "content": "F_{a} \\in \\mathbb{R}^{C \\times N \\times M}" + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": ", we propose to aggregate candidate point features belonging to the same anchor point. As shown in Figure 4, we gradually compress the candidate point dimension from " + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "inline_equation", + "content": "||S||" + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": " to 1. During the compression, we use convolution with the kernel size of " + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "inline_equation", + "content": "3 \\times 1" + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": " and padding of " + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "inline_equation", + "content": "(1,0)" + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": " to propagate context information along the anchor point dimension. This operation constrains the context information to only flow among the anchor points in the same lane. Meanwhile, we use convolution with the kernel size of " + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "inline_equation", + "content": "1 \\times 3" + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": " and padding of " + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 46, + 270, + 289, + 488 + ], + "type": "text", + "content": " to compress local information along the candidate point dimension. This operation fuses the local information from the sampled sparse points to the corresponding anchor. These two operations are alternately repeated to propagate and compress information progressively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "type": "text", + "content": "As for the global feature, we compress the entire frontview features into a single feature vector " + }, + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "type": "inline_equation", + "content": "F_{g} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "type": "text", + "content": " through pooling operation, following previous methods [24, 27, 4, 20, 8, 41]. We then concatenate the global feature vector " + }, + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "type": "inline_equation", + "content": "F_{g}" + }, + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "type": "text", + "content": " to each anchor feature vector " + }, + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "type": "text", + "content": " and fuse them together to balance the global and local information. The fused features are then used to update the anchor point location " + }, + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\{(\\Delta_{x_j^i}, z_j^i, v_j^i)\\}_{j=0}^{j=M}" + }, + { + "bbox": [ + 47, + 498, + 287, + 597 + ], + "type": "text", + "content": " through 3 convolution layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Lane Category Refinement In order to detect the lane categories, we extract lane features by aggregating the feature of anchor points belonging to the same lane. Similar to the above aggregation operation for the anchor point features, we first concatenate the global feature " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{g}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " with anchor features " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " and then gradually propagate the information along the anchor point dimension. The aggregated lane features " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{l} \\in \\mathbb{R}^{C \\times N}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " are then used to update the lane category " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " through several convolution layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 176, + 350, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 176, + 350, + 186 + ], + "spans": [ + { + "bbox": [ + 306, + 176, + 350, + 186 + ], + "type": "text", + "content": "3.4. Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 194, + 547, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 547, + 232 + ], + "type": "text", + "content": "The predicted anchor-based lane representation consists of category " + }, + { + "bbox": [ + 305, + 194, + 547, + 232 + ], + "type": "inline_equation", + "content": "\\tilde{p}_i" + }, + { + "bbox": [ + 305, + 194, + 547, + 232 + ], + "type": "text", + "content": " and location " + }, + { + "bbox": [ + 305, + 194, + 547, + 232 + ], + "type": "inline_equation", + "content": "\\tilde{l}_i = \\{(\\tilde{\\Delta}_{x_j^i},\\tilde{z}_j^i,\\tilde{v}_j^i)\\}_{j = 0}^{j = M}" + }, + { + "bbox": [ + 305, + 194, + 547, + 232 + ], + "type": "text", + "content": ". We use the cross entropy to supervise the predicted category " + }, + { + "bbox": [ + 305, + 194, + 547, + 232 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 305, + 194, + 547, + 232 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 367, + 239, + 545, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 367, + 239, + 545, + 271 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = - \\frac {1}{N} \\sum_ {i = 0} ^ {i = N - 1} p _ {i} \\log \\left(\\tilde {p} _ {i}\\right). \\tag {6}", + "image_path": "eeb180ecd754e1fa3ba1e8f3a5a8ae91d78418040c9ffb97891447817a10ed9e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "text", + "content": "As for the location, we supervise the learning of visibility " + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "inline_equation", + "content": "\\tilde{v}" + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "text", + "content": " loss and use the visibility as weights to compute the " + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "text", + "content": " loss for " + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "inline_equation", + "content": "\\tilde{\\Delta}_x" + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 305, + 277, + 546, + 314 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 329, + 320, + 545, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 320, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 329, + 320, + 545, + 355 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\Delta} = \\frac {1}{M \\cdot N} \\sum_ {i = 0} ^ {i = N - 1} \\sum_ {j = 0} ^ {j = M - 1} v _ {j} ^ {i} | | \\Delta_ {x _ {j} ^ {i}} - \\tilde {\\Delta} _ {x _ {j} ^ {i}} | | _ {1}, \\tag {7}", + "image_path": "8393824b6b5141d5806ff6a4d7f1fd1fe249c99df5962919bc8affee6310da8e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 363, + 545, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 363, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 339, + 363, + 545, + 397 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {z} = \\frac {1}{M \\cdot N} \\sum_ {i = 0} ^ {i = N - 1} \\sum_ {j = 0} ^ {j = M - 1} v _ {j} ^ {i} | | z _ {j} ^ {i} - \\tilde {z} _ {j} ^ {i} | | _ {1}, \\tag {8}", + "image_path": "d1a80919e3cb69eddbcad786f27436a6c6399ae6725e704a5d5626e5eacd3a8e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 402, + 545, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 402, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 342, + 402, + 545, + 435 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v} = \\frac {1}{M \\cdot N} \\sum_ {i = 0} ^ {i = N - 1} \\sum_ {j = 0} ^ {j = M - 1} | | v _ {j} ^ {i} - \\tilde {v} _ {j} ^ {i} | | _ {1}. \\tag {9}", + "image_path": "3e57efd9887dff76993b12ce20ef3da8f17df24cf67854c6351eb2af911beb36.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 439, + 443, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 443, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 443, + 451 + ], + "type": "text", + "content": "The joint loss can be indicated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 459, + 545, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 459, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 370, + 459, + 545, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {p} + \\mathcal {L} _ {\\Delta} + \\mathcal {L} _ {z} + \\mathcal {L} _ {v}. \\tag {10}", + "image_path": "8fc37e26dd975494b8223c8c9a2620e185f7add6b452a7b112f5e9a8b53ca66a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 479, + 546, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 546, + 527 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 546, + 527 + ], + "type": "text", + "content": "During the hierarchical refinement from the low resolution to the high resolution, we separately compute the above loss for each resolution and sum up the losses as the final loss to train the network." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 538, + 388, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 388, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 388, + 552 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 558, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 547, + 714 + ], + "type": "text", + "content": "In the main paper, we present the sparse-point guided 3D lane detection method on the anchor-based approach. We mainly use Persformer [3] as the baseline, while Gen-LaneNet [7] is taken as an optional baseline to verify the feasibility of our method. We validate our method on the real-world datasets OpenLane [3] and ONCE-3DLanes [18] to demonstrate our priority to handle complex environments and variable scenes. As for our method on the segmentation-based approach, we present the results obtained on the synthetic dataset Apollo 3D Lane [7] in the Supplementary Material. The code is available at https://github.com/YaoChengTang/Sparse-Point-Guided-3D-Lane-Detection" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8367" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 70, + 511, + 135 + ], + "blocks": [ + { + "bbox": [ + 87, + 70, + 511, + 135 + ], + "lines": [ + { + "bbox": [ + 87, + 70, + 511, + 135 + ], + "spans": [ + { + "bbox": [ + 87, + 70, + 511, + 135 + ], + "type": "table", + "html": "
MethodAll ↑Up & Down ↑Curve ↑Extreme Weather ↑Night ↑Intersection ↑Merge & Split ↑
3D-LaneNe [6]44.140.846.547.541.532.141.7
Gen-LaneNet [7]32.325.433.528.118.721.431
PersFormer [3]50.542.455.648.646.64050.7
Ours + PersFormer53.7 (3.2↑)46.2 (3.8↑)59.2 (3.6↑)54.8 (6.2↑)49.8 (3.2↑)41.9 (1.9↑)52.1 (1.4↑)
", + "image_path": "184b9347a0f344cb0dd3dee72037545709a3c25839c7fe4dbe34548f9511228c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 86, + 171, + 511, + 258 + ], + "blocks": [ + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "type": "text", + "content": "Table 1: Performance comparison with state-of-the-art methods on OpenLane benchmark in different scenarios, where the evaluation metric is F1-score. " + }, + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "type": "text", + "content": " means that the larger the value, the better the result." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 171, + 511, + 258 + ], + "lines": [ + { + "bbox": [ + 86, + 171, + 511, + 258 + ], + "spans": [ + { + "bbox": [ + 86, + 171, + 511, + 258 + ], + "type": "table", + "html": "
MethodF-score (%) ↑X error near (m) ↓X error far (m) ↓Z error near (m) ↓Z error far (m) ↓
3D-LaneNe [6]44.10.4790.5720.3670.443
Gen-LaneNet [7]32.30.5910.6840.4110.521
Cond-IPM [3]36.60.5631.0800.4210.892
PersFormer [3]50.50.4850.5530.3640.431
Gen-LaneNet* [7]42.80.4880.6320.3740.481
Ours + Gen-LaneNet*46.60.4750.5770.3710.445
Ours + PersFormer52.30.4680.5140.3710.418
", + "image_path": "eb895c9e0dc44783d82f900b195e7522e63ee0cb3515fe4094de47b7ccb983b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 303, + 110, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 110, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 110, + 315 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": "OpneLane Dataset OpneLane [3] is a challenging real-world 3D lane dataset constructed on Waymo Open dataset [31]. It consists of 200K frames with 14 kinds of categories, complex lane structures, and five kinds of weather. " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": " frames have an altitude change of more than " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "1m" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": " frames have more than six lanes. Currently, it is the most challenging lane detection dataset, so we choose it as the main dataset to conduct ablation studies and demonstrate the performance promotion of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 430, + 288, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 288, + 492 + ], + "type": "text", + "content": "ONCE-3DLanes Dataset ONCE-3DLanes [18] is a real-world 3D lane dataset constructed on ONCE [23]. It consists of various scenes, such as highways, bridges, tunnels, suburbs, and downtown, with different weather conditions (sunny/rainy) and lighting conditions (day/night)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 502, + 180, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 180, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 180, + 514 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": "We select the PersFormer [3] as the baseline of the anchor-based approach to leverage the sparse point. EfficientNet [34] is used as the backbone for the front-view feature extraction. Perspective Transformer [3] and 3D-GeoNet [7] are used in coarse-level lane detection to extract the dense BEV features and detect coarse 3D lanes at the lowest resolution. Following PersFormer, we set the highest resolution of BEV space as " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "208 \\times 128" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " with the range of " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "[-10, 10]" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " meters along the x-axis and [3, 101] meters along the y-axis, respectively. We set the count of anchor lanes as " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "N = 182" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " and anchor points for each lane as " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "M = 10" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": ", where the fixed " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{y^i\\}" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " is set as " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{5, 10, 15, 20, 30, 40, 50, 60, 80, 100\\}" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " for each lane. In fine-level sparse point refinement, we set the size of sampling window " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "inline_equation", + "content": "(1, 0.5)" + }, + { + "bbox": [ + 46, + 521, + 288, + 715 + ], + "type": "text", + "content": " sampling step along the x-axis and the z-axis separately." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 324, + 300, + 531, + 365 + ], + "blocks": [ + { + "bbox": [ + 46, + 260, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 547, + 297 + ], + "type": "text", + "content": "Table 2: Performance comparison with state-of-the-art methods using different evaluation metrics on the validation set of OpenLane benchmark. * means that we re-implement the method on the OpenLane benchmark. ↑ means that the larger the value, the better the result. ↓ means that the smaller the value, the better the result." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 300, + 531, + 365 + ], + "lines": [ + { + "bbox": [ + 324, + 300, + 531, + 365 + ], + "spans": [ + { + "bbox": [ + 324, + 300, + 531, + 365 + ], + "type": "table", + "html": "
MethodTraining \nMemory (G)Inference \nSpeed (FPS)
3D-LaneNe [6]6.6103
Gen-LaneNet [7]4.767
PersFormer [3]19.627
Ours + PersFormer12.545
", + "image_path": "16bb2c67a45da24dcaafd7f9d1ebccb985d97899760cebd135fdd6c4b182fb01.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 368, + 547, + 428 + ], + "lines": [ + { + "bbox": [ + 304, + 368, + 547, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 547, + 428 + ], + "type": "text", + "content": "Table 3: Training memory cost and inference speed comparison with state-of-the-art approaches. We acquire the memory cost on a single A30 GPU with a batch size of 8. We obtain the inference speed on a single A30 GPU with a batch size of 1." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "type": "text", + "content": "In the training stage, we pre-train the coarse-level lane detection for 100 epochs and then finetune our fine-level sparse point refinement for another 100 epochs. For both pertaining and finetuning, we use Adam optimizer [12] with a base learning rate of " + }, + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "type": "text", + "content": ". The cosine learning rate policy is applied to optimization with a maximum iteration of 8 and a minimum learning rate of " + }, + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "type": "text", + "content": ". We also use gradient norm clipping during the optimization with a maximum norm weight of 35. In the Inference stage, we use the number of valid anchor points in each lane as the threshold to filter anchor lanes that are too close. For more details about training and inference, please refer to the code and our supplemental materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 599, + 419, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 419, + 610 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 419, + 610 + ], + "type": "text", + "content": "4.3. Evaluation Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Following OpenLane, we use bipartite matching to evaluate the predicted and ground-truth lanes. The matching is true positive when the distance error of " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " lane points is less than " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.5m" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ". The percentage of matched ground-truth lanes is deemed as the recall, and the percentage of matched predicted lanes is deemed as the precision. The average of recall and precision is F-score. The lane points are also divided into near points " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "(3 - 40m)" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and far points " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "(40 - 101m)" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8368" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 71, + 482, + 345 + ], + "blocks": [ + { + "bbox": [ + 111, + 71, + 482, + 345 + ], + "lines": [ + { + "bbox": [ + 111, + 71, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 111, + 71, + 482, + 345 + ], + "type": "image", + "image_path": "afd6c05e45ebccd60776c8079199ce0dc670e7df44f27a6db1ddad4ac81a24f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 348, + 480, + 361 + ], + "lines": [ + { + "bbox": [ + 111, + 348, + 480, + 361 + ], + "spans": [ + { + "bbox": [ + 111, + 348, + 480, + 361 + ], + "type": "text", + "content": "Figure 5: The visualization results of PersFormer and our method on the OpneLane dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 368, + 288, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 288, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 288, + 428 + ], + "type": "text", + "content": "The point-wise distance error along the x-axis and z-axis is used to evaluate the location accuracy for near and far points. In ONCE, we also use the unilateral Chamfer Distance (CD) metric that computes the chamfer distance between the predicted and ground truth lane." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 441, + 188, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 188, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 188, + 452 + ], + "type": "text", + "content": "4.4. Benchmark Performance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 460, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 687 + ], + "type": "text", + "content": "OpenLane Following the setting in OpenLane, we evaluate the performance in different scenarios. As shown in Table 1, we achieve the best results in all scenarios. It is worth noting that our method achieves more than 6 points promotion in extreme weather scenarios. This promotion shows that our method greatly resists the feature noise and extracts valuable information by only focusing on sparse points around 3D lanes. In Table 2, we further compare our method with state-of-the-art approaches using different evaluation metrics. Our method outperforms all state-of-the-art approaches in almost every evaluation metric. Specifically, we improve the location on the x-axis in both near and far distances and thus improve the overall F1-score. What's more, our method effectively helps Gen-LaneNet outperform both the original result of Gen-LaneNet published by OpenLane and our reproduced Gen-LaneNet* on all metrics. These results prove that our method is flexible and can seamlessly integrate with different prototypical methods to offer consistent improvement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We also compare our method with state-of-the-art approaches from the training memory cost and the inference" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 368, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 545, + 499 + ], + "type": "text", + "content": "speed. As shown in Table 3, our method achieves the lowest memory cost and the fastest inference speed. Compared to the baseline, we greatly reduce the memory cost, which enables the front-view images to have higher resolution. Furthermore, the sparse point guides the network to achieve a better result in most cases with a much faster inference speed. It should be noted that the speed and memory cost reported in this paper is computed based on a pure Python realization of sparse-point sampling and aggregation without any unfair engineering optimization like the custom CUDA kernel that significantly reduces the running time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 505, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 611 + ], + "type": "text", + "content": "Besides the quantitative analysis, we also compare visualization results of the PersFormer and our method in different scenarios for qualitative analysis. As shown in Figure 5, our method obviously outperforms the PersFormer in cloudy, rainy, night, and different road conditions. This further demonstrates that the sparse point guides the BEV feature to extract more meaningful information from the front-view image and thus improve the recall and accuracy of the detected lanes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "ONCE-3DLanes As illustrated in Table 4, we follow the PersFormer using the validation set of ONCE-3DLanes to evaluate the results of our methods. Considering scenes are easier in the ONCE benchmark, the quantity results are already very high. The sparse point slightly promotes the baseline in the F1-score, Precision, and CD error with a much more efficient performance, which only costs half the memory and inference time." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8369" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 69, + 504, + 135 + ], + "blocks": [ + { + "bbox": [ + 92, + 69, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 92, + 69, + 504, + 135 + ], + "spans": [ + { + "bbox": [ + 92, + 69, + 504, + 135 + ], + "type": "table", + "html": "
MethodF1-score (%) ↑Precision (%) ↑Recall (%) ↑CD error (m) ↓
3D-LaneNe [6]44.7361.4635.160.127
Gen-LaneNet [7]45.5963.9535.420.121
SALAD [3]64.0775.9055.420.098
PersFormer [3]74.3380.3069.180.074
Ours + PersFormer74.7981.8568.860.070
", + "image_path": "062c8c6471a9bcf40e77be12dc21288035f8df96c9559ce72f21fd7d99f88bdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 168, + 272, + 204 + ], + "blocks": [ + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "type": "text", + "content": "Table 4: Performance comparison with state-of-the-art methods on ONCE benchmark. " + }, + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "type": "text", + "content": " means that the larger the value, the better the result. " + }, + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 137, + 547, + 163 + ], + "type": "text", + "content": " means that the smaller the value, the better the result." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 168, + 272, + 204 + ], + "lines": [ + { + "bbox": [ + 66, + 168, + 272, + 204 + ], + "spans": [ + { + "bbox": [ + 66, + 168, + 272, + 204 + ], + "type": "table", + "html": "
Front-viewFeature ExtractionCoarse levelFine level 1Fine level 2Fine level 3
4.8ms6.1ms3.8ms3.1ms3.4ms
", + "image_path": "55048aa91d7ae714f583604cdd56f8aefef50faf55d6cd21ea8f6955ca23520c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 223, + 274, + 289 + ], + "blocks": [ + { + "bbox": [ + 54, + 205, + 279, + 217 + ], + "lines": [ + { + "bbox": [ + 54, + 205, + 279, + 217 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 279, + 217 + ], + "type": "text", + "content": "Table 5: Inference time cost of each step in our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 223, + 274, + 289 + ], + "lines": [ + { + "bbox": [ + 63, + 223, + 274, + 289 + ], + "spans": [ + { + "bbox": [ + 63, + 223, + 274, + 289 + ], + "type": "table", + "html": "
Coarse levelFine level 1Fine level 2Fine level 3F1-score (%)X error near (m)
✓51.30.484
✓✓52.10.475
✓✓✓53.60.470
✓✓✓✓53.70.468
", + "image_path": "ecb1a7fa3c30d81246130d5624e80f533cf978ffa711ba33340ea254e92cae23.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 324, + 214, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 214, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 214, + 337 + ], + "type": "text", + "content": "4.5. Ablation Study and Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 342, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 413 + ], + "type": "text", + "content": "Complexity Analysis We evaluate the time cost of each step in our method to show the efficiency of our sparse refinement. As shown in Table 5, it is more efficient to use sparse points for feature fetching and 3d lane refinement, even compared to the coarse-level lane detection conducted in the lowest resolution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 415, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 485 + ], + "type": "text", + "content": "Effectiveness of Refinement In order to validate the effectiveness of our fine-level sparse point refinement, we compare results from the coarse level and from the different fine levels. As shown in Table 6, the fine-level sparse point refinement greatly improves the results of coarse-level lane detection in both F1-score and X error near." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 486, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 616 + ], + "type": "text", + "content": "Effectiveness of Features To ensure globally smooth and locally discriminative performance, we leverage both the global and local features to refine the lanes. We demonstrate the advantage of balancing the global and local information by comparing three kinds of feature conditions, including only using global features, only using local features, and using both local and global features. As shown in Table 7, the first condition results in a very low F1 score of " + }, + { + "bbox": [ + 46, + 486, + 288, + 616 + ], + "type": "inline_equation", + "content": "46.8\\%" + }, + { + "bbox": [ + 46, + 486, + 288, + 616 + ], + "type": "text", + "content": ", and the second one is already comparable to the baseline, while the last one fusing with sparse points gives the best result." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": "Effectiveness of Point Sampling As aforementioned, we sample candidate points around the anchor point on the x-z plane. Here, we analyze the effectiveness of point sampling using different sampling steps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": "As illustrated in Table 8, the result gets better when the value of the sampling step is smaller along the " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": "-axis and larger along the " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": "-axis. This phenomenon shows that the local features fetched along the " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": "-axis are more valuable" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 329, + 168, + 525, + 214 + ], + "blocks": [ + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "lines": [ + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "text", + "content": "Table 6: The performance of coarse-level lane detection and fine-level sparse point refinement." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 168, + 525, + 214 + ], + "lines": [ + { + "bbox": [ + 329, + 168, + 525, + 214 + ], + "spans": [ + { + "bbox": [ + 329, + 168, + 525, + 214 + ], + "type": "table", + "html": "
Global FeaturesLocal FeaturesF1-score (%)
✓46.8
✓51.8
✓✓53.7
", + "image_path": "81b96726f5908e0bf7c18ae722a89de8a61349a759ec180737f977aa6f38f606.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 354, + 243, + 501, + 309 + ], + "blocks": [ + { + "bbox": [ + 305, + 217, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 240 + ], + "type": "text", + "content": "Table 7: The performance of our method with/without global features and local features." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 354, + 243, + 501, + 309 + ], + "lines": [ + { + "bbox": [ + 354, + 243, + 501, + 309 + ], + "spans": [ + { + "bbox": [ + 354, + 243, + 501, + 309 + ], + "type": "table", + "html": "
x-axisz-axisF1-score
1153.61
10.553.65
0.5153.34
0.50.553.46
10.2553.67
", + "image_path": "b4c101d59864f58cf02fbf932f9ac886109c4867a9dd892d2502aeddad8f8ae4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 312, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 305, + 312, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 545, + 337 + ], + "type": "text", + "content": "Table 8: The performance of our method with different sampling steps along x-axis and z-axis." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 346, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 384 + ], + "type": "text", + "content": "than the features fetched along the z-axis, where the sampling along the x-axis actually gives us more valid sample points on the front-view image." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 389, + 455, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 389, + 455, + 401 + ], + "spans": [ + { + "bbox": [ + 306, + 389, + 455, + 401 + ], + "type": "text", + "content": "4.6. Limitations and Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 407, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 552 + ], + "type": "text", + "content": "As shown in the above experiments, we have achieved great progress in both accuracy and efficiency. However, our method still has some limitations, e.g., assuming all 3D lane instances could be detected on BEV features in the lowest resolution. This assumption becomes less powerful when meeting extremely complex and crowded road conditions. A slow-fast update strategy for refinement at different scales might give better results, but it requires much more engineering tricks on feature manipulation. In this paper, we mainly focus on proving the priority and possibility of simply using sparse correspondence instead of dense correspondence for 3D lane detection." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 562, + 379, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 379, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 379, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": "In this paper, we have proved that using sparse points is more efficient and adequate for high-quality 3D lane detection than building a dense correspondence between the HR front-view space and BEV space. We presented a sparse point-guided 3D lane detection method, including coarse-level lane detection and fine-level sparse point refinement. Experiments proved that our method could achieve much better results with less memory and time cost." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement This work was supported by the Natural Science Foundation of China (NSFC) under Grants No.62176021." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8370" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 134 + ], + "type": "text", + "content": "[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. arXiv preprint arXiv:2209.07989, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 190 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 256 + ], + "type": "text", + "content": "[3] Li Fei Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, and Junchi Yan. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proceedings of the European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "text", + "content": "[4] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5932-5941, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 356 + ], + "type": "text", + "content": "[5] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17041-17049, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "text", + "content": "[6] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2921-2930, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 456 + ], + "type": "text", + "content": "[7] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proceedings of the European Conference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 511 + ], + "type": "text", + "content": "[8] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and Shimin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8928-8937, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "type": "text", + "content": "[9] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zihan Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. arXiv preprint arXiv:2301.02371, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "type": "text", + "content": "[10] Yujie Jin, Xiangxuan Ren, Fengxiang Chen, and Weidong Zhang. Robust monocular 3d lane detection with dual attention. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 3348-3352, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[11] Longlong Jing, Yucheng Chen, and Yingli Tian. Coarse-to-fine semantic segmentation from image-level labels. IEEE Transactions on Image Processing (TIP), 29:225-236, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 669 + ], + "type": "text", + "content": "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 9799-9808, 2020." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "text", + "content": "[14] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Taeyeop Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1965-1973, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 150, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 205 + ], + "type": "text", + "content": "[15] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4369-4378, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 205, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 545, + 249 + ], + "type": "text", + "content": "[16] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Transactions on Intelligent Transportation Systems, 21(1):248-258, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 250, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 303 + ], + "type": "text", + "content": "[17] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condilanenet: a top-to-down lane detection framework based on conditional convolution. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 3773-3782, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 304, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 360 + ], + "type": "text", + "content": "[18] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Zejian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1765-1772, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "type": "text", + "content": "[19] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3694-3702, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "text", + "content": "[20] Shilin Liu, Haoxiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1788-1797, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 459, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 503 + ], + "type": "text", + "content": "[21] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 503, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 557 + ], + "type": "text", + "content": "[22] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 11677-11684, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 558, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 545, + 602 + ], + "type": "text", + "content": "[23] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Chunjing Xu, et al. One million scenes for autonomous driving: Once dataset. 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 602, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 658 + ], + "type": "text", + "content": "[24] Lars M. Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4455-4465, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Roozbeh Mottaghi, Yu Xiang, and Silvio Savarese. A coarse-to-fine model for 3d pose estimation and sub-category recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 418-426, 2015." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8371" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaouu Tang. Spatial as deep: Spatial cnn for traffic scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 7276-7283, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 182 + ], + "type": "text", + "content": "[27] Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and S. Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 183, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 288, + 227 + ], + "type": "text", + "content": "[28] Marco Pedersoli, Andrea Vedaldi, Jordi Gonzalez, and Xavier Roca. A coarse-to-fine approach for fast deformable object detection. Pattern Recognition (PR), 48(5):1844-1853, 2015." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[29] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 276-291. Springer, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "type": "text", + "content": "[30] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 81-90, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 393 + ], + "type": "text", + "content": "[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2446-2454, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 394, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 288, + 448 + ], + "type": "text", + "content": "[32] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 514 + ], + "type": "text", + "content": "[33] Lucas Tabelini, Rodrigo Berriel, Thiago Meireles Paixão, Claudine Santos Badue, Alberto Ferreira de Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 294-302, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 558 + ], + "type": "text", + "content": "[34] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 6105-6114. PMLR, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "text", + "content": "[35] Wouter Van Gansbeke, Bert De Brabandere, Davy Neven, Marc Proesmans, and Luc Van Gool. End-to-end lane detection through differentiable least-squares fitting. In Proceedings of the IEEE International Conference on Computer Vision Workshops (CVPRW), pages 0–0, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "text", + "content": "[36] Jun Wang, Shiyi Lan, Mingfei Gao, and Larry S Davis. Info-focus: 3d object detection for autonomous driving with dynamic information modeling. In European Conference on Computer Vision, pages 405-420. Springer, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[37] Hang Xu, Shaoju Wang, Xinyue Cai, Wei Zhang, Xiaodan Liang, and Zhenguo Li. Curvelane-nas: Unifying lanesensitive architecture search and adaptive point blending. In Proceedings of the European Conference on Computer Vision (ECCV), pages 689-704. Springer, 2020." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 407 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 138 + ], + "type": "text", + "content": "[38] Fan Yan, Ming-Jun Nie, Xinyue Cai, Jianhua Han, Han Xu, Zhengxin Yang, Chao Ye, Yanwei Fu, Michael B. Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 17122–17131, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "type": "text", + "content": "[39] Chengtang Yao, Yunde Jia, Huijun Di, Pengxiang Li, and Yuwei Wu. A decomposition model for stereo matching. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6091-6100, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 185, + 546, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 238 + ], + "type": "text", + "content": "[40] Jian Yao, Marko Boben, Sanja Fidler, and Raquel Urtasun. Real-time coarse-to-fine topologically preserving segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2947-2955, 2015." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 240, + 546, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 546, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 546, + 294 + ], + "type": "text", + "content": "[41] Jianglong Ye, Yuntao Chen, Naiyan Wang, and X. Wang. Gifs: Neural implicit function for general shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12819-12829, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 297, + 546, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 350 + ], + "type": "text", + "content": "[42] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. Resa: Recurrent feature-shift aggregator for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3547-3554, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 352, + 546, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 352, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 352, + 546, + 407 + ], + "type": "text", + "content": "[43] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 898-907, 2022." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8372" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/32196ea3-316d-4633-8abd-b15ee0703f11_content_list.json b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/32196ea3-316d-4633-8abd-b15ee0703f11_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b0dd5dec349a0fd71b822072f70eec05f77361dd --- /dev/null +++ b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/32196ea3-316d-4633-8abd-b15ee0703f11_content_list.json @@ -0,0 +1,2364 @@ +[ + { + "type": "text", + "text": "Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks", + "text_level": 1, + "bbox": [ + 101, + 128, + 869, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sixiang Chen $^{1,3*}$ Tian Ye $^{1,3*}$ Jinbin Bai $^{2}$ Erkang Chen $^{3}$ Jun Shi $^{4}$ Lei Zhu $^{1,5\\dagger}$", + "bbox": [ + 238, + 200, + 730, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ The Hong Kong University of Science and Technology (Guangzhou) $^{2}$ National University of Singapore $^{3}$ School of Ocean Information Engineering, Jimei University", + "bbox": [ + 171, + 243, + 803, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Xinjiang University $^{5}$ The Hong Kong University of Science and Technology", + "bbox": [ + 248, + 277, + 720, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sixiangchen,owentianye}@hkust-gz.edu.cn", + "bbox": [ + 336, + 296, + 633, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jinbin.bai@u.nus.edu, ekchen@jmu.edu.cn, junshi2022@gmail.com, leizhu@ust.hk", + "bbox": [ + 233, + 313, + 733, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: https://ephemeral182.github.io/UDR_S2Former_deraining", + "bbox": [ + 207, + 330, + 759, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 380, + 313, + 396 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the real world, image degradations caused by rain often exhibit a combination of rain streaks and raindrops, thereby increasing the challenges of recovering the underlying clean image. Note that the rain streaks and raindrops have diverse shapes, sizes, and locations in the captured image, and thus modeling the correlation relationship between irregular degradations caused by rain artifacts is a necessary prerequisite for image deraining. This paper aims to present an efficient and flexible mechanism to learn and model degradation relationships in a global view, thereby achieving a unified removal of intricate rain scenes. To do so, we propose a Sparse Sampling Transformer based on Uncertainty-Driven Ranking, dubbed UDR-S²Former. Compared to previous methods, our UDR-S²Former has three merits. First, it can adaptively sample relevant image degradation information to model underlying degradation relationships. Second, explicit application of the uncertainty-driven ranking strategy can facilitate the network to attend to degradation features and understand the reconstruction process. Finally, experimental results show that our UDR-S²Former clearly outperforms state-of-the-art methods for all benchmarks.", + "bbox": [ + 75, + 412, + 473, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rain is a ubiquitous condition that negatively impacts various computer vision tasks [2, 70]. In real-world rain scenes, raindrops and rain streaks are irregularly superimposed on clean images. Image deraining is employed to restore the clean images from the complex rain degradations. According to previous work [48], the imaging model", + "bbox": [ + 75, + 775, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/410f339b609fb3004addda3fbc86a245d84ee73bd6d09100f23bc0843aad919d.jpg", + "image_caption": [ + "(a) Fixed patch partition breaks the relationship modeling within diverse rain degradations" + ], + "image_footnote": [], + "bbox": [ + 503, + 380, + 627, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd4a186fc2ed84893650b5c348120a6af474faeafcbed90bd80de75d91102006.jpg", + "image_caption": [ + "(b) We utilize the uncertainty-driven sparse sampling transformer to thoroughly explore arbitrary position degradation relationships." + ], + "image_footnote": [], + "bbox": [ + 630, + 378, + 758, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14aa14faaa58ff0bbcb5d23a759f6ec79ab539cb41f466e8781982ef3a781831.jpg", + "image_caption": [ + "(c) Uncertainty Map", + "Figure 1: Illustration of the breakdown of complex rain degradation relationships and the thumbnails of our main ideas. Colored places indicate degradations. Two-way arrows represent modeling between degradations." + ], + "image_footnote": [], + "bbox": [ + 772, + 380, + 898, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of precipitation, inclusive of rain streaks and raindrops, can be expressed as:", + "bbox": [ + 498, + 541, + 892, + 571 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {d s} = \\left(1 - \\mathcal {M} _ {r}\\right) \\odot (\\mathcal {B} + \\mathcal {S}) + \\eta \\mathcal {D}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 588, + 890, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\mathcal{B}$ and $S$ denote the clean background and rain streak map. $\\mathcal{M}_r$ is a binary mask used to judge whether the pixel belongs to the raindrops or the background. $\\mathcal{D}$ is the raindrops and $\\eta$ means global atmospheric lighting coefficient.", + "bbox": [ + 496, + 612, + 892, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As noted by CCN [48], removing rain streaks and raindrops in a unified manner cannot be achieved by simply combining separate methods for removing either. This is due to the complex nature of the physical models involved and the wide array of possible degradation combinations. Previous models developed to address singular forms of degradations [11, 12, 36, 49, 50, 55, 62] face notable obstacles when dealing with irregularly dispersed and diverse rain degradation types.", + "bbox": [ + 496, + 672, + 893, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Specifically, current SOTA methods for image deraining primarily concentrate on using ViTs due to their abilities to model long-range dependencies [27, 40, 57, 58]. Among these methods, window-based self-attention [40, 57, 58] has gained popularity due to its computational efficiency. However, as shown in Fig.1, we argue that utilizing window", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions.", + "bbox": [ + 94, + 875, + 212, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Lei Zhu (leizhu@ust.hk) is the corresponding author.", + "bbox": [ + 94, + 887, + 382, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13106", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/455cfcad4a2f84bf4501ab5bd9f39d3136e940d7c7cf326d729a346ed21a8143.jpg", + "image_caption": [ + "Figure 2: The uncertainty maps (bottom row) correspond to both real and synthetic samples (top row), with more significant uncertainty appearing in areas with severe and complicated degradations. This observation motivates us to use uncertainty explicitly to represent knowledge about degradation and to improve the model's understanding of degradation restoration." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 875, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "based self-attention mechanisms can lead to incomplete degradation coverage, causing the breakdown of degradation relationships for unified rain degradation removal due to fixed window segmentation. This problem can be particularly pronounced when dealing with large raindrops or rain streaks at long distances simultaneously. However, in the case of complicated degradations, it is imperative to model the relationships between related forms of degradations.", + "bbox": [ + 75, + 330, + 470, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, with respect to the dense prediction task of rain removal, the density, shape, position, and size of raindrops and streaks are all uncertain, rendering it arduous for the network to restore clean images from diverse degradations. The CCN [48] requires using expensive and inflexible NAS to select an optimal architecture that effectively handles rain streaks and raindrops precisely. Inspired by uncertainty modeling [34] for image restoration, incorporating uncertainty learning can enhance the performance by reducing the error in model parameters [19], or serve as a regular term constraint to enhance the prediction quality of regions characterized by high uncertainty [44, 53]. Nevertheless, the mentioned design paradigm overlooks the importance of explicitly excavating uncertainty maps in facilitating the network's modeling of degradation features. For intricate rain scenes, we claim that learning uncertainty estimation can in turn affect the network to better focus on the complicated rain degradation areas. As depicted in Fig.2, the uncertainty map of the image exhibits greater concentration within the degraded region. It is our contention that by fully leveraging the properties of this uncertainty map, we can more effectively model the relationships of degradation and drive the network for the understanding of degradation restoration.", + "bbox": [ + 75, + 455, + 470, + 818 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, to address the problems above, we first design the Sparse Sampling Attention to deal with complicated rain scenes. It sparsely learns the relevant degradation relationships from the entire image, thereby alleviating the drawback of large-scale degradation modeling in window", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "based attention. Concurrently, we leverage the uncertainty map to guide feature learning for capturing more discriminative sampling features. To be exact, in order to fully leverage uncertainty information to promote the sampling of degradation features, we propose a novel ranking strategy and present a Constraint Matrix based on it. Such design further restricts the degree of attention to various rain degradations in the sampling process, boosting the modeling of relationships between degradations. Moreover, when restoring partially degraded regions, we consider the internal difference of the uncertainty map and use the ranking of the Correlation Map to strengthen the network to restore the degraded area by leveraging clean cues within the local regions.", + "bbox": [ + 496, + 330, + 893, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contributions can be summarized as follows:", + "bbox": [ + 517, + 541, + 888, + 555 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An uncertainty-driven sparse sampling transformer that fully models the global degradation relationships in an efficient manner is proposed to remove diverse rain streaks and raindrops.", + "- We introduce a ranking strategy in the uncertainty map to enable the model to emphasize various rain degradation features in the sampling process through a constructed constraint matrix.", + "- To enhance local reconstruction, we utilize the internal discrepancies within the uncertainty map to stimulate the network to extract credibly clean information." + ], + "bbox": [ + 517, + 559, + 890, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 738, + 648, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Single Image Deraining", + "text_level": 1, + "bbox": [ + 500, + 760, + 718, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rain streak removal. Image restoration from adverse weather has made significant advancements over the years, owing to its paramount significance [8, 9, 30, 41, 64, 66, 67]. Recently, the field of single image deraining has been predominantly dominated by learning-based methods [28, 36, 37, 50, 55, 58, 59, 61-63, 74, 75]. Zhu et al. [75] proposed a joint optimization algorithm that involves iteratively removing rain streaks from the background layer and non-rain", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13107", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1246007b4512497ade2642e695184ce9421565feb4cf759b160db120c8e0fb57.jpg", + "image_caption": [ + "Figure 3: The overview of our UDR- $S^2$ Former pipeline, which includes (a) our proposed restoration architecture, (b) the Local Reconstruction (LR) in the IRM module, (c) the Sparse Sampling Attention (SSA) constrained by the uncertainty map. For clarity, we only depict one local patch operation. The Feature Extraction and Global Modeling Modules, as well as the simple Refinement Block, are described in supplementary material due to page limitation." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 869, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "details from the rain layer. This is achieved through the incorporation of three essential priors. PreNet [50] offered a recurrent layer to leverage the inter-stage dependencies of deep features, thereby constructing the progressive recurrent network. UMRL [63] proposed the uncertainty map to constrain the rain map in rainscapes and incorporate physical models to estimate the final deraining output. IDT [58] presented a transformer system comprising a complementary window-based transformer and spatial transformer, enabling improved capture of short- and long-range dependencies in rainy scenes.", + "bbox": [ + 75, + 362, + 472, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Raindrop removal. Raindrops are frequently observed in rain scenes, and their diverse shapes and positions present difficulties removing them. Previous attempted to eliminate raindrops utilizing various methods, [14, 47, 49, 65]. The Eigen et al. [14] firstly introduced the learning-based paradigm for raindrop removal. AttenGAN [47] adopted the combination of GAN and attention mechanism to recover the clean image from raindrop degradations.", + "bbox": [ + 75, + 529, + 468, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most of the current design paradigms continue to create specialized networks for removing rain streaks or raindrops. A unified removal network design has still to be developed. While CCN [48] was the first to consider joint removal of rain degradation, their approach still requires expensive design such as NAS to address complex degradation characteristics. Other general-purpose networks [58, 68] disregard these two degradation characteristics, leading to high computation and parameters when improving performance.", + "bbox": [ + 75, + 650, + 470, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Vision Transformers for Image Restoration", + "text_level": 1, + "bbox": [ + 76, + 789, + 444, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ViTs exhibited superior global modeling capabilities, resulting in impressive performance on low-level vision tasks [3, 7, 10, 22, 52, 54, 68, 71] compared with previous CNNs' paradigm [4, 20, 21, 23, 24, 29, 31, 33, 73]. The window-based designs [39, 40, 52, 57, 58] were widely employed to overcome the computational complexity issue of", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathrm{O}(N^2)$ . Additionally, ART [71] utilized a combination of sparse and dense self-attention to manage computational overhead and achieved SOTA outcomes. Restormer [68] incorporated channel-based self-attention to circumvent the square-level complexity problem. Nevertheless, such designs limit the global receptive field of self-attention in spatial dimension and lack flexibility in dealing with intricate and changing degradations. We propose to use sparse sampling to address the aforementioned limitations by adaptively sampling information from the global field to meet the modeling requirements of the local.", + "bbox": [ + 496, + 361, + 893, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Uncertainty in Deep Learning", + "text_level": 1, + "bbox": [ + 500, + 534, + 767, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to Bayesian theory [34], uncertainty in deep learning can be classified into two types: (i) Aleatoric uncertainty, which refers to the inherent noise in the data. (ii) Epistemic uncertainty, which relates to the uncertainty of model parameters. Incorporating modeling uncertainty into the network can enhance its robustness and performance [1, 2]. In low-level fields, leveraging uncertainty can enable the network to prioritize reducing model prediction error [19] or serving as a loss function [32, 44, 53] to improve the reconstruction quality of areas with high uncertainty. Furthermore, uncertainty can assist in the more precise estimation of critical parameters in physical models [63]. Same as [44, 53], we focus on aleatoric uncertainty in this paper. However, these paradigms do not explicitly utilize the distinctive characteristic of the uncertainty map to restrict the network's acquisition of features. This paper proposes leveraging uncertainty-driven ranking to facilitate the network to represent degradations.", + "bbox": [ + 496, + 554, + 890, + 827 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed UDR-S²Former Pipeline", + "text_level": 1, + "bbox": [ + 500, + 832, + 803, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Network architecture. The architecture of the proposed network is illustrated in Fig.3. The network accepts a rain image as input, conducts image processing within the net", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13108", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/62df6801affa3d5489d0566e822b8c7aee6b1e3de9f95453b05e9dfc6ae540c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 85, + 872, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/81ea36e8e73f8046f0293c4b7232ad8436e6ecff48a74542095e7f952587b009.jpg", + "image_caption": [ + "Figure 4: Top row: Constructing a constraint matrix according to the ranking strategy derived from the uncertain map. To simplify the explanation, we only present a one-dimensional ranking strategy. Bottom row: The procedure for utilizing the correlation map and ranking strategy to produce the matrix utilized for modulating the self-attention map. We only present the correlation map and ranking strategy within a patch for easy visualization." + ], + "image_footnote": [], + "bbox": [ + 93, + 213, + 875, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "work, and generates a high-quality restored image as output. Specifically, 1) the degraded image is fed into the Feature Extraction stage to acquire knowledge of features at various scales. This process is accomplished in four stages, each comprising basic convolutional blocks. 2) Upon completion of the feature extraction stage, we employ vanilla transformer to capture deep-level information and ensure the comprehensive utilization of global information1. 3) The Image Reconstruction Module (IRM) is employed to represent degradation relationships in the form of sparse sampling self-attention, which is driven by uncertain learning. Further, it can trigger clean cues excavation to guide the restoration for local reconstruction. Skip connections are utilized in both the feature extraction stage and each stage of the image reconstruction module.", + "bbox": [ + 75, + 409, + 472, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminary for Uncertainty. For image deraining, we model the difference between the image $\\mathcal{B}_{gt}$ and its estimated deraining image $\\hat{\\mathcal{B}}$ as a Laplace distribution. The motivation for this choice is that the Laplace distribution is better suited for characterizing the edges of details in images than the Gaussian distribution. In addition, the L1 loss refers to the Laplace distribution [72]. The likelihood function can be expressed as follows:", + "bbox": [ + 75, + 635, + 472, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathcal {B} _ {g t}, \\sigma ; \\mathcal {R} _ {d s}) = \\frac {1}{2 \\sigma} \\exp \\left(- \\frac {| | \\hat {\\mathcal {B}} - \\mathcal {B} _ {g t} | | _ {1}}{\\sigma}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 767, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{B}_{gt}$ denotes the mean of this distribution. $\\hat{\\mathcal{B}}$ is output calculated by network from rain image $\\mathcal{R}_{ds}$ . $\\sigma$ means the uncertainty (variance) of deraining image $\\hat{\\mathcal{B}}$ . To simplify", + "bbox": [ + 76, + 819, + 470, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the calculation, we transform the likelihood function into its log form and maximize it:", + "bbox": [ + 498, + 409, + 890, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\sigma} {\\arg \\max } \\ln p \\left(\\mathcal {B} _ {g t}, \\sigma ; \\mathcal {R} _ {d s}\\right) = \\left(- \\frac {\\left| \\left| \\hat {\\mathcal {B}} - \\mathcal {B} _ {g t} \\right| \\right| _ {1}}{\\sigma}\\right) - \\ln \\sigma . \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 448, + 890, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the learning process, the aforementioned formula is inverted and utilized as a loss function, which is minimized for optimization. Additionally, we also follow [44] to circumvent the issue of training instability resulting from the presence of uncertainties in the form of zero values. Without being limited to a loss function, we aim to employ the ranking strategy to explicitly utilize estimated uncertainty. It enhances the modeling of complicated degradation relationships and drive the network to restore degraded regions.", + "bbox": [ + 496, + 489, + 890, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Image Reconstruction Module", + "text_level": 1, + "bbox": [ + 500, + 628, + 769, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the image reconstruction module of each stage, we can express it as:", + "bbox": [ + 498, + 648, + 890, + 678 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I R M} (\\mathcal {X}, \\mathcal {U}) = \\left\\{\\operatorname {S S A} (\\mathcal {X}, \\mathcal {U}), \\operatorname {L R} (\\mathcal {X}, \\mathcal {U}) \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 688, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{\\cdot\\}$ indicates that these two modules are alternately formed, with $\\mathcal{X}$ denoting the image feature, and $\\mathcal{U}$ representing the corresponding uncertainty map.", + "bbox": [ + 496, + 714, + 890, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1 Sparse Sampling Attention Constrained by Uncertainty Map", + "text_level": 1, + "bbox": [ + 498, + 762, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The degradation of a large-scale area (raindrops) or long-distance span (rain streaks) can result in the loss of local relationships, particularly when a fixed window division is utilized. Moreover, a window-based design may limit the receptive field to a local area, leading to the loss of global knowledge when dealing with complicated degradations. However, it is crucial to fully exploit the modeling of", + "bbox": [ + 496, + 795, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1Restricted by the number of pages, we introduce the points 1) and 2) of our architecture in the supplementary material.", + "bbox": [ + 76, + 875, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13109", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea18c18415113c0130d0d2b3e1fcde7fdd12d7ab0cc1c830faa04190a2d857a7.jpg", + "image_caption": [ + "Figure 5: Visualization of sparse sampling. The yellow region denotes an $8 \\times 8$ window patch. The red pixels are coordinates obtained from global sparse sampling to interact window patch for modeling the corresponding degradation relationship, which is not limited to the local patch (we draw multi-head sampling coordinates)." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 473, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "degradation relationships for removing diverse rain degradations due to their complicated degraded properties. In this part, we propose an uncertainty-driven sparse sampling approach to learn global coordinates for capturing associated rain degradations adaptively, effectively modeling degradations in diverse regions.", + "bbox": [ + 75, + 314, + 468, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a given image feature $\\mathcal{X} \\in \\mathbb{R}^{C \\times H \\times W}$ , our primary objective is to establish a model that can accurately capture the degradation relationships. For each part of the degradations, the network should be prompted to concurrently consider correlation degradations occurring in other image regions to model it. Specifically, we let the network adaptively learn to match the degraded coordinates of other regions in the whole image, and map them to the same image patch, so as to model the degradation relationship at a low cost. The coordinates are learned as follows:", + "bbox": [ + 75, + 404, + 468, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S}, \\mathcal {B} = \\mathcal {F} \\left(\\mathcal {X} ^ {\\in \\mathbb {R} ^ {C \\times H \\times W}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 565, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}(\\cdot)$ represents the simple convolution and avgpooling operation to estimate scaling factors $S\\in \\mathbb{R}^{C\\times H\\times W\\times 2}$ and biases $\\mathcal{B}\\in \\mathbb{R}^{C\\times H\\times W\\times 2}$ (last dimension denotes x-axis and y-axis coordinates). We use them to transform the original coordinates of the feature map by multiplication and addition:", + "bbox": [ + 75, + 597, + 468, + 689 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C o o r d s} ^ {T} (x, y) = \\mathcal {G} \\left(\\operatorname {C o o r d s} ^ {O} (x, y) \\times \\mathcal {S} + \\mathcal {B}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 704, + 468, + 723 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a39ba2fff3f815cf16deb003d73bb7370b029d38a45d886c13802cb8da037e7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 726, + 419, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scaling factors and biases change the sampling location for each patch. $\\mathrm{Cords}^T (x,y)$ denotes the coordinates of sparse sampling from the global feature, which is leveraged to the local patch $\\mathcal{X}_i^{S\\in \\mathbb{R}^{C\\times \\frac{H}{M}\\times \\frac{H}{M}}}$ ( $i$ denotes the $i$ -th patch of $\\mathcal{M}^2$ patches) after carrying out the function $\\mathcal{G}$ of torch nn. functional grid sample, as shown in Fig.3 (c). We utilize the information obtained through global sparse sampling to match each degraded patches for corresponding degradation modeling and propose a novel attention mech", + "bbox": [ + 75, + 757, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "anism, dubbed Sparse Sampling Attention (SSA):", + "bbox": [ + 500, + 90, + 828, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S S A} = \\mathbf {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {i} ^ {P} \\mathcal {K} _ {i} ^ {S ^ {\\mathrm {T}}}}{\\sqrt {\\mathcal {D}}} + p\\right) \\mathcal {V} _ {i} ^ {S}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 117, + 890, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{Q}_i^P$ denotes the queries projected from original feature of $i$ -th patch. $\\mathcal{K}_i^S$ and $\\mathcal{V}_i^S$ are obtained from global sampling feature. $\\mathcal{D}$ means the dimension number and $p$ is the position embedding like [42].", + "bbox": [ + 496, + 170, + 890, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Uncertainty map driven: Drawing inspiration from the obvious representation of deteriorated regions in the uncertainty map, we aim to incorporate the uncertainty map as an explicit constraint to restrict the level of attention towards the degradations in the sparse sampling procedure, to ensure the modeling of the degradation relationships. The operation is presented in Fig.4.", + "bbox": [ + 496, + 232, + 890, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More specifically, with regard to the uncertain map $\\mathcal{U} \\in \\mathbb{R}^{C \\times H \\times W}$ , we employ a ranking strategy to construct a constraint matrix $\\mathcal{C} \\in \\mathbb{R}^{C \\times H \\times W}$ , which serves as a means of constraining the sampling process:", + "bbox": [ + 496, + 339, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathcal {C} ^ {n} \\left(\\mathcal {U} ^ {n}\\right) \\right] _ {x y} = \\left\\{ \\begin{array}{l l} 1 & \\mathcal {U} _ {x y} ^ {n} \\geq \\mathcal {U} ^ {n} \\text {o f r a n k} \\gamma \\\\ \\beta & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 414, + 890, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ and $\\gamma$ are the constraint factor and a threshold value of $n$ -th dimension. Such a constraint matrix contains crucial cues from the uncertainty map with almost no computational overhead, and we utilize it to regularize our modeling of degradation relationships. The Eq.5 is further optimized as:", + "bbox": [ + 496, + 465, + 890, + 553 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} ^ {U}, \\mathcal {B} ^ {U} = \\mathcal {F} \\left(\\mathcal {X} ^ {\\in \\mathbb {R} ^ {C \\times H \\times W}} \\times \\mathcal {C} _ {x y}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 554, + 890, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "wherein $\\times$ represents the element-wise multiplication. The $\\beta$ of $\\mathcal{C}_{xy}$ reduces the influence of irrelevant background areas in multiplication form. $S^U$ and $\\mathcal{B}^U$ are leveraged to obtain robust coordinates via Eq.6, which promotes the global sparse sampling while mitigating the interference of the background region. Thereby it facilitates the network to concentrate on degradation relationship modeling.", + "bbox": [ + 496, + 582, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our visualization is depicted in Fig.5. As we expected, the sampled red points adaptively have related degradation situations with the target point (see $8 \\times 8$ yellow window). The sparse sampling strategy with uncertainty helps model the relation between local and long-range correlation degradations. Additionally, The most concentration of sampled pixels in a larger area maintains some semantic information and coherence, despite being sampled at a distance.", + "bbox": [ + 496, + 688, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion I: The merits of our sparse sampling compared with previous related work for image restoration. Compared to the latest sparse attention [71], which permitted each token to interact with a limited number of tokens, and with a fixed interval size. However, it still involved a manually designed mechanism. In complex rainy", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13110", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/980449d752487eeff0080a47024774ac06acb5afdddae884766f108a6bf8e14c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 89, + 475, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/03079a5130d4d503a7d48a4e061c5916e78dc994518d7cfe18fe769045c509b9.jpg", + "image_caption": [ + "Figure 6: Top row: Our motivation about needing to strengthen the network to exploit non-degradation regions. Most information is lost due to the complicated rain degradations in the real world, so degradation-free information is needed to reconstruct the clean area. Bottom row: Visualizations on uncertainty map, correlation map and self-attention map. The attention points of the correlation map and the self-attention map differ. We can utilize the correlation map to direct the network's attention towards areas with significant differences, thus enhancing the restoration process through self-attention." + ], + "image_footnote": [], + "bbox": [ + 81, + 166, + 475, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scenes, such an approach cannot flexibly capture degradation information from different locations. Our design aims to enable the network to learn the coordinates that can sample degradation information from any position, leading to an enhanced modeling performance of the degradation relationships and network flexibility.", + "bbox": [ + 75, + 453, + 468, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion II: uncertainty driven compared with prior paradigm and rain mask. Regarding uncertainty, the previous UMRL [63] approach used the uncertainty to optimize the rain streak map based on the physical model of the rain streak. Still, it only focused on optimizing the final map, ignoring the network's learning of the intermediate features. Our method uses uncertainty explicitly to drive the network to provide motivation-driven constraints and optimization for the degradation modeling and local reconstruction processes. Compared to the learned rain mask needed GT [46], i) uncertainty map has a theoretical unsupervised loss for practicing. ii). It outperforms the rain mask in representing the network's focus on intricate rain areas beyond mere location. iii). For large-area raindrops with a complicated physical model, it is difficult to learn a rain mask, but the uncertainty map can solve it well by using the discriminative representation of degradations.", + "bbox": [ + 75, + 542, + 470, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1.2 Local Reconstruction with Correlation Ranking", + "text_level": 1, + "bbox": [ + 76, + 805, + 465, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In rain scenes, information in locally degraded regions is often obscured by rain streaks or raindrops, making it difficult to restore, see Fig.6. To address this issue, it is vital to leverage clean cues in the rain-free area. Building on the discriminative representation of degraded areas and", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "other regions in the uncertainty map, we propose using the uncertainty map to generate a correlation map. This map explicitly adjusts the attention map through a ranking approach, as shown in Fig.4 and Fig.3(b), which promotes the network to better leverage clean cues for reconstruction in the form of Query-Key.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin by partitioning uncertainty map $\\mathcal{U}^{\\in \\mathbb{R}^{C\\times H\\times W}}$ into local patches $\\mathcal{U}^{P\\in \\mathbb{R}^{C\\times \\frac{H}{M}\\times \\frac{H}{M}}}$ that correspond to those patches in the feature map. The presence of discriminative disparities within each patch of the uncertainty map motivates us to calculate the correlation map (CR) between them:", + "bbox": [ + 496, + 180, + 890, + 279 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C R} _ {i} = \\mathcal {U} _ {i} ^ {P} \\times \\mathcal {U} _ {i} ^ {P ^ {\\mathrm {T}}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 279, + 890, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Upon obtaining the correlation map, our purpose is to leverage the degradation-free region to facilitate the restoration process of the degraded areas during image reconstruction. Significant differences (low correlation) between degradations and background in the uncertainty map can be considered a prompt, promoting the network to use clean cues from the background for local reconstruction. To this end, we modulate the self-attention map by selecting the low-correlation regions in the correlation map:", + "bbox": [ + 496, + 306, + 892, + 445 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L R} = \\text {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {i} ^ {P} \\mathcal {K} _ {i} ^ {P T}}{\\sqrt {\\mathcal {D}}} \\times \\underbrace {[ - \\alpha \\boldsymbol {\\Sigma} _ {i} + (1 + \\alpha) ] + p} _ {\\text {M o d u l a t i o n P r o c e s s i n g}}\\right) \\mathcal {V} _ {i} ^ {P}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 455, + 893, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ is a modulation factor. The $\\Sigma$ is acquired via the Top k ranking approach, which can be expressed as follows:", + "bbox": [ + 498, + 535, + 890, + 566 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\boldsymbol {\\Sigma} _ {i} \\left(\\mathcal {C R} _ {i}\\right) \\right] _ {x y} = \\left\\{ \\begin{array}{l l} 1 & \\left(\\mathcal {C R} _ {i}\\right) _ {x y} \\in \\operatorname {T o p k} (\\text {r o w} y) \\\\ 0 & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 577, + 890, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the experiments section, the key parameters of the above used are further studied. About our image reconstruction module, the self-attention FFNs are identical to vanilla vision transformer [13].", + "bbox": [ + 496, + 628, + 890, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Loss Function", + "text_level": 1, + "bbox": [ + 500, + 695, + 643, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our UDR- $S^2$ Former network will predict the derained result and an uncertainty map. Hence, the total loss of our network is defined as follows:", + "bbox": [ + 496, + 715, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {p s n r}} \\left(\\mathcal {B} _ {\\text {p r e}}, \\mathcal {B} _ {\\text {g t}}\\right) + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c e p t u a l}} \\left(\\mathcal {B} _ {\\text {p r e}}, \\mathcal {B} _ {\\text {g t}}\\right) \\\\ + \\lambda_ {3} \\mathcal {L} _ {U D L} (\\mathcal {B} _ {\\text {p r e}}, \\mathcal {U}), \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 772, + 890, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{B}_{pre}$ and $\\mathcal{B}_{gt}$ denote the predicted result of image deraining and the corresponding ground truth. $\\mathcal{U}$ denotes the predicted uncertainty map. $\\mathcal{L}_{psnr}$ and $\\mathcal{L}_{perceptual}$ are the PSNR and perceptual losses, see [6,7]. $\\mathcal{L}_{UDL}$ denotes the uncertainty loss, see [44, 53] for its definition. $\\lambda_{1},\\lambda_{2}$ and $\\lambda_{3}$ are empirically set to 1, 0.2 and 1.", + "bbox": [ + 496, + 821, + 890, + 912 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13111", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/16a136b4cea867bd4e200af1287d95c32431420227d70d60e406a4381e928372.jpg", + "table_caption": [ + "Table 1: Quantitative results compared with SOTA methods on the Rain200H [61] and Rain200L [61] datasets. Underline and bold indicate the first and second best results." + ], + "table_footnote": [], + "table_body": "
MethodRain200H [61]Rain200L [61]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
GMM [38]14.710.43028.990.875
JCAS [18]14.870.47130.050.897
DDN [15]26.360.80334.930.958
NLEDN [35]29.510.89138.560.980
RESCAN [37]27.450.82135.080.959
PreNet [50]29.040.89037.120.976
UMRL [63]28.710.88736.430.973
JORDER-E [60]28.580.87636.900.973
MSPFN [28]29.660.89039.480.984
CCN [48]29.990.91438.260.981
MPRNet [69]30.760.90839.890.985
DGUNet [43]30.850.91140.230.986
Uformer [57]30.800.91140.200.986
Restormer [68]31.390.91640.580.987
IDT [58]32.100.93440.740.988
NAFNet [5]30.980.91240.450.987
UDR-S2Former32.59↑0.490.937↑0.00340.96↑0.220.989↑0.001
", + "bbox": [ + 135, + 117, + 411, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 309, + 209, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Implementaion Details", + "text_level": 1, + "bbox": [ + 76, + 330, + 287, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our UDR- $S^2$ Former model utilizes a 5-level encoder-decoder architecture. Within this architecture, the number of channel dimensions increases to $\\{16,32,64,128,256\\}$ across levels 1 to 5. Additionally, the Feature Extraction stage contains $\\{4,6,7,8\\}$ convolutional blocks, while the transformer block number in the latent layer is 8, with 16 heads. Regarding reconstruction, we employ $\\{3,6,7,8\\}$ Image Reconstruction Modules, composed alternately of Sparse Sampling Attention and Local Reconstruction, with $\\{1,2,4,8\\}$ heads. The window size is fixed at $8\\times 8$ . Finally, we set the constraint factor $\\beta$ , threshold value $\\gamma$ , modulation factor $\\alpha$ , and the k value for the ranking strategy to 0.6, 0.8, 0.2, and 0.8, respectively. For the self-attention mechanism in this paper, we all use multi-head self-attention, which is consistent with vanilla ViT [13]. Additionally, during the last image reconstruction module of each stage, we convert the output feature and corresponding uncertainty map into the final image format, and the $\\mathcal{L}_{UDL}$ is utilized to supervise the learning process of uncertainty map.", + "bbox": [ + 75, + 349, + 468, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During the training phase, we employed the Adam optimizer with initial momentum $\\beta_{1} = 0.9$ and $\\beta_{1} = 0.999$ . We initially set the learning rate to 0.0003 and utilized a cyclic learning rate adjustment strategy, whereby the maximum learning rate is set to 0.00036. We trained our model with a data augmentation strategy that included randomly cropping $256 \\times 256$ patches and using horizontal flipping and random image rotation to a fixed angle. Our training process involved $6 \\times 10^{5}$ steps. To leverage perceptual loss, we used the first and third layers of VGG19 [51]. Our model is implemented using PyTorch [45] and the RTX 3090 GPU.", + "bbox": [ + 75, + 638, + 468, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Evaluation Metrics and Datasets", + "text_level": 1, + "bbox": [ + 76, + 808, + 364, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. In accordance with prior deraining methodologies [48, 58], follow [16, 17], the performance of the model is evaluated using PSNR [25] and SSIM [56]. As suggested in [48, 61], the assessment of PSNR and SSIM is based on the luminance channel, which refers to the Y", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "channel of the YCbCr space.", + "bbox": [ + 498, + 90, + 692, + 106 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rain Streak Datasets. In order to evaluate the effectiveness of our rain streak removal method, we follow the strategy of previous work [48] and select two benchmark datasets, namely Rain200H and Rain200L [61]. Both two datasets have 1800 synthetic images for training and 200 images for testing.", + "bbox": [ + 496, + 107, + 890, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Raindrop Datasets. We also evaluate our method in a raindrop dataset (i.e., AGAN-Data) collected by Qian et al. [47]. AGAN-Data has 861 images for training and 58 images for testing.", + "bbox": [ + 496, + 196, + 890, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Raindrops and Rain Streak Datasets. Our paper aims to develop a methodology that effectively addresses the challenges posed by the presence of large-area raindrops and complex rain streaks. To evaluate the proposed approach, we utilize the RainDS benchmark dataset [48], which includes both real-world and synthetic images, dubbed as RainDS-Real and RainDS-Syn, respectively. These datasets contain images of scenes that include rain streaks only (RS), raindrops only (RD), or both(RDS). RainDS-Syn has 3600 image pairs, with 3000 images used for training and the remaining 600 images for testing. Meanwhile, RainDS-Real comprises of 750 images, with 450 images for training and 300 images for evaluation.", + "bbox": [ + 496, + 258, + 890, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Experimental Evaluation on Benchmarks", + "text_level": 1, + "bbox": [ + 498, + 457, + 854, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared Methods. Regarding removing raindrops and rain streaks in images, we conduct extensive experiments to compare various algorithms that can be used for image rain removal. (i) We compare previous SOTA methods (including GMM [38], JCAS [18], DDN [15], NLEDN [35], RESCAN [37], PreNet [50], UMRL [63], JORDER-E [60], MSPFN [28], CCN [48], IDT [58]). (ii) We also compare our network against universal image restoration methods, including MPRNet [69], DGUNet [43], Uformer [57], Restormer [68], NAFNet [5]). For the specific raindrop removal dataset, we compare against Eigen's model [14], Pix2Pix [26], AttentGAN [47], Quan's network [49], CCN [48], and IDT [58]. In the absence of pre-trained models, we conduct model retraining by utilizing publicly available code and subsequently evaluate the best model performance on test datasets to ensure a fair comparison.", + "bbox": [ + 496, + 477, + 890, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. Tables 1, 2, and 3 reports the quantitative results of our network and state-of-the-art methods on four benchmark datasets, which are Rain200H, Rain200L, RainDS, and AGAN-Data. As demonstrated in these Tables, our approach delivers superior metric results over compared state-of-the-art (SOTA) methods for single rain streaks or raindrop degradations. Moreover, by incorporating uncertainty to steer the network's attention towards challenging degradations, our method outperforms the IDT algorithm by 0.49dB, 0.22dB and 1.01dB on the Rain200H [61], Rain200L [61] datasets and AGAN-Data [47], while significantly outperforming previ", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13112", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4661b3b9060803570aca2d9b60a59e76e27b1fce2c47568332efdbd67f863268.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 173, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e44c919bee67622ed881fccf97ff9eb8250d9a1a91731bdd50aadee962d136c0.jpg", + "image_caption": [ + "(b) CCN" + ], + "image_footnote": [], + "bbox": [ + 174, + 90, + 263, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00d806f974c087b8eb50dd3a85bd604ef500e7c7a5b4381615a2b0d88e183060.jpg", + "image_caption": [ + "(c) MPRNet" + ], + "image_footnote": [], + "bbox": [ + 264, + 90, + 351, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95a0fe0411d66e95ff91de179f0a87375979b33891ac098f48a8c9ccd8c773b1.jpg", + "image_caption": [ + "(d) DGUNet" + ], + "image_footnote": [], + "bbox": [ + 352, + 90, + 439, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91135ad22006f61f0acf3b922ebf4eef6724475bf37937eb8b294df3400e68c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 90, + 527, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae5c647ed1f41523b7c0f1469335909a858512c724ebfffd47a189a69ff4b79d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 90, + 617, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4200f6a7e050e033df00242605ab53e37e65960de2a16e33fee5920b239c25d.jpg", + "image_caption": [ + "(g) NAFNet" + ], + "image_footnote": [], + "bbox": [ + 617, + 90, + 705, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/334e416b92a24680aa163dd7906c8b74ed02ebb9026c7a485398057425833fde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 90, + 794, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/782a0ec52bcc5fe2109dd0da9d7a05d6200e48a5cd80adce08995dd89a1c88dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 90, + 883, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d80353b8f21b50708cc7cec013affaa53c094d0c7475473a5aea5e4c4f7e51d4.jpg", + "image_caption": [ + "(a) Input", + "Table 2: Quantitative comparisons of various SOTA approaches on the RainDS [48] benchmark. Our proposed method outperforms all other approaches on both synthetic and real-world datasets, including all types of precipitation (i.e., rain streaks (RS), raindrops (RD), and a combination of both (RDS)). The results are indicated in bold and underline, representing the first and second-best performances, respectively. $\\uparrow$ means the higher is better. The #GFLOPs are calculated by $256\\times 256$ image resolution for a fair comparison." + ], + "image_footnote": [], + "bbox": [ + 86, + 179, + 341, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc044bc659c8cbc7765ac349d9ef8de1139dbca51891cd1d5198ce6dab256596.jpg", + "image_caption": [ + "(e) Resto", + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 349, + 179, + 473, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d53039f29147929bcffca07ed287a6649f954b48e0f5e2139c548dba336ed9a.jpg", + "image_caption": [ + "(f) IDT", + "(c) CCN" + ], + "image_footnote": [], + "bbox": [ + 483, + 179, + 609, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d69d6ef4c27e24a5ae353d0017ac570cc4040fbe76342db59757c98afe7ba1a.jpg", + "image_caption": [ + "(h) Ours", + "(d) MPRNet" + ], + "image_footnote": [], + "bbox": [ + 619, + 179, + 743, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12853acdb0fc7ae992aeab31ee7366da601a7da7e6a4ad30af1482721275739b.jpg", + "image_caption": [ + "(i) GT", + "(e) DGUNet" + ], + "image_footnote": [], + "bbox": [ + 754, + 179, + 880, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5c9a77c0c56ca1b6fc07ac89e038b7c8e255d5e86fb94bb81b92dd02fca6a5a.jpg", + "image_caption": [ + "(f) Restomer", + "Figure 7: Visual comparisons of removal for raindrops and rain streaks from RainDS-Syn [48] dataset. Zoom it for a better illustration." + ], + "image_footnote": [], + "bbox": [ + 349, + 243, + 473, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f282a81a3ba8d3e62a7df4e37e8cc4894af73180c443580f8c7d09a9bc92bbf.jpg", + "image_caption": [ + "(g) IDT" + ], + "image_footnote": [], + "bbox": [ + 483, + 243, + 609, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c910c365c43ab37ee2b03cad04cc0eb773118dc70f116d111379ae1c9ff4a6c6.jpg", + "image_caption": [ + "(h) NAFNet" + ], + "image_footnote": [], + "bbox": [ + 620, + 243, + 743, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c2d70f53dbff353cd86fa32f1b07f4b92f3e396ca890ad0b2aeab8554fef10f.jpg", + "image_caption": [ + "(i) Ours" + ], + "image_footnote": [], + "bbox": [ + 754, + 243, + 880, + 289 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ab1506730d2addb72eaaaa628571f84a51c369d091725dd322cfa77efef0bb16.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueRainDS-SynRainDS-Real#Param #GFLOPs
RSRDRDSRSRDRDS
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
GMM [38]CVPR'201626.660.78123.040.79321.500.66923.730.56018.600.55421.350.576--
JCAS [18]ICCV'201726.460.78623.150.81120.910.67124.040.55618.180.55521.220.585--
DDN [15]CVPR'201730.410.86927.920.88526.850.79624.850.68323.120.64222.470.606--
NLEDN [35]ACMMM'201836.240.95834.870.95732.130.91727.020.72324.710.67124.060.650--
RESCAN [37]ECCV'201830.990.88729.900.90727.430.81826.700.68324.230.63723.230.5870.15M32.32G
PreNet [50]CVPR'201936.630.96834.580.96432.210.93426.430.72924.420.67923.570.6490.17M66.58G
UMRL [63]CVPR'201935.760.96233.590.95831.570.92925.890.72623.930.67623.010.6470.98M16.50G
JORDER-E [60]TPAMI'201933.650.92533.510.94430.050.87026.560.71324.340.66223.540.6294.17M273.68G
MSPFN [28]CVPR'202038.610.97536.930.97334.080.94726.450.72724.490.68124.110.65121.00M708.44G
CCN [48]CVPR'202139.170.98137.300.97634.790.95727.460.73725.140.70124.930.6793.75M245.85G
MPRNet [69]CVPR'202140.810.98137.030.97234.990.95627.290.73625.260.70124.960.6813.64M148.55G
DGUNet [43]CVPR'202241.090.98337.560.97535.340.95927.520.73725.330.70224.990.68312.18M199.74G
Uformer [57]CVPR'202240.690.97237.080.96634.990.95426.890.73025.310.70124.830.68620.63M43.86G
Restomer [68]CVPR'2022(Oral)41.420.98038.780.97636.080.96127.390.74225.380.70224.920.68526.10M140.99G
IDT [58]TPAMI'202241.610.98339.090.98036.230.96027.510.74325.670.70624.990.68916.00M61.90G
NAFNet [5]ECCV'202240.390.97237.230.97434.990.95727.490.72925.230.70124.640.66340.60M16.19G
UDR-S2Former (Ours)ICCV'202342.39↑0.780.988↑0.00539.78↑0.690.983↑0.00336.91↑0.680.966↑0.00627.90↑0.390.745↑0.00226.01↑0.340.709↑0.00325.52↑0.530.691↑0.0028.53M21.58G
", + "bbox": [ + 89, + 382, + 887, + 571 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8a415f1b0dbbaa8ec5b5824b7cead6a58acf4deed7681fbdea344beb76f8671a.jpg", + "table_caption": [ + "Table 3: Quantitative results compared with SOTA methods on the AGAN-Data [47]. Red and blue indicate the first and second best results." + ], + "table_footnote": [], + "table_body": "
MethodPSNR†SSIM†
(ICCV'2013)Eigen's model [14]21.310.757
(CVPR2017)Pix2Pix [26]27.200.836
(CVPR'2018)AttenGAN [47]31.590.917
(ICCV'2019)Quan's network [49]31.370.918
(CVPR'2021)CCN [48]31.340.929
(TPAMI'2022)IDT [58]31.630.936
UDR-S2Former32.640.943
", + "bbox": [ + 147, + 627, + 401, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ous uncertainty-based [63] methods. Table 2 demonstrates that our approach surpasses all prior designs for jointly removing rain streaks and raindrops. While networks like Restormer [68] and IDT [58] can model long-distance dependencies sufficiently, our method is able to effectively model complex degradation relationships in rain-dominated landscapes (i.e., raindrops and rain streaks both) with uncertainty. Additionally, sparse sampling further facilitates the network's ability to obtain crucial global degradation information at a low cost, see Table 2. Based on the indicators,", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our approach exhibits a significant advantage over the latest designs, including Restormer (36.08PSNR $\\rightarrow$ 36.91PSNR) and IDT (36.23PSNR $\\rightarrow$ 36.91PSNR), in both real-world and synthetic datasets. Additionally, we also present the comparison of speed and memory cost in the supplementary material to further demonstrate our superiority.", + "bbox": [ + 496, + 597, + 890, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Comparison. We present the comparisons of the visual effects in Fig.7-9. As depicted in the figures, we observe that in a scene containing rain streaks and raindrops, the scene information is either occluded by a broad range of raindrops or masked by dense rain streaks of varying scales. Other approaches either struggle to remove complex degradations or significantly compromise image details. Our approach leverages degradation relationship modeling to facilitate image restoration, and utilizes the clean information to recover intricate rain degradations. The resulting output exhibits higher fidelity with the more accurate restoration of great details. In real-world scenarios, our method enables the removal of various forms of degradation while recovering fine details, outperforming previous methods.", + "bbox": [ + 496, + 689, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13113", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d1650df8a459301e2502c0da25dfdebc1fa05f652c967b11e638ac5f7f8965f6.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 343, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2d5f3cf3304e873e3ac7af6f9d440126ad26ef2fbf2f54467209182c99fa0731.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 344, + 89, + 470, + 157 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ede9051f35f56847a0bbea5b66658ae1d14a1ce77742a9711e074e1e305b3f3d.jpg", + "image_caption": [ + "(c) CCN" + ], + "image_footnote": [], + "bbox": [ + 482, + 89, + 609, + 157 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/34e739911121ef46b15cc7ad233e7824b648c540164230269ed7101277082c51.jpg", + "image_caption": [ + "(d) MPRNet" + ], + "image_footnote": [], + "bbox": [ + 619, + 89, + 746, + 157 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e7f0be0f7aa886d0a351f82358ff85cc4438a3ceec842c67f3ab901f031c9616.jpg", + "image_caption": [ + "(e) DGUNet" + ], + "image_footnote": [], + "bbox": [ + 756, + 89, + 883, + 157 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e9882578ce49b3e8a79f0921759750eb2833f77a4c1b8b5e4a11cc5898f271ca.jpg", + "image_caption": [ + "(f) Restormer" + ], + "image_footnote": [], + "bbox": [ + 344, + 181, + 470, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9e2679e241c3d45687e66636ed8409dea6f041bc3cea437448922b5c7e465e63.jpg", + "image_caption": [ + "(g) IDT" + ], + "image_footnote": [], + "bbox": [ + 482, + 183, + 609, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d5d142ce757652236acf13bff8eb361602200c667da6849a62dcd88954a0228a.jpg", + "image_caption": [ + "(h) NAFNet" + ], + "image_footnote": [], + "bbox": [ + 619, + 181, + 746, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a2d3e2b06ea3d6cc0fac92d3ccaa40f344903a7f5742ff5e3cddc2277f7bf948.jpg", + "image_caption": [ + "(i) Ours" + ], + "image_footnote": [], + "bbox": [ + 756, + 181, + 883, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c1650e8e0558982d5749552c994726198a0ca45bdbe24045703b04ff65275441.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 78, + 292, + 176, + 356 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/72012fe37ca959d90e275b379fb72b334d9ac1cece175c83e32cce5052e63334.jpg", + "image_caption": [ + "(b) CCN" + ], + "image_footnote": [], + "bbox": [ + 179, + 292, + 276, + 356 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6a3585fa80c2518b44f44d16cc0ed26a56d78e26e01ca6de5be32964554c81c4.jpg", + "image_caption": [ + "Figure 8: Visual results for real-world sample of RainDS-Real [48] dataset.", + "(c) DGUNet" + ], + "image_footnote": [], + "bbox": [ + 277, + 292, + 375, + 356 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c8e7ffe1bb5c11471f44c577b1faa5f342fa68a10c47b426b2b8facdf1a514e4.jpg", + "image_caption": [ + "(d) Restomer" + ], + "image_footnote": [], + "bbox": [ + 377, + 292, + 475, + 356 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b822c43491e0fe5225eac5c0628d3246ebc7864b362e0cfa590321ecad4576e0.jpg", + "image_caption": [ + "(e) IDT", + "Figure 9: Visual results for real-world sample of RainDS-Real [48] dataset, Please zoom it for a better illustration." + ], + "image_footnote": [], + "bbox": [ + 80, + 369, + 176, + 431 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/82843bbe1e3a8b4031c70e40dab0161769de12f5c4308a4f2ba21c3f371e30e7.jpg", + "image_caption": [ + "(f) NAFNet" + ], + "image_footnote": [], + "bbox": [ + 179, + 369, + 276, + 431 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c9f1f32b47ed79de199abc3da64414af6b9851e4228d6380c3b913b4fd095e08.jpg", + "image_caption": [ + "(g) Ours" + ], + "image_footnote": [], + "bbox": [ + 277, + 369, + 375, + 431 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e35d81cf6ef259b8b35d04d0fb5d7b8e271b6e5131aeb28759b1e60b7bba55ea.jpg", + "image_caption": [ + "(h) GT" + ], + "image_footnote": [], + "bbox": [ + 377, + 369, + 475, + 431 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2d13ca8de42c27a0b28f1f224f5e30ff96401262bfb6d8e72e5240d84d6c2bdd.jpg", + "image_caption": [ + "(a) Ablation of $\\beta$ and $\\gamma$", + "Figure 10: About the ablation experiments of the parameter $\\beta, \\gamma, \\alpha$ and $k$ , the vertical axis is the PSNR metric." + ], + "image_footnote": [], + "bbox": [ + 78, + 486, + 274, + 560 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ecaa1242dbc566f35826dbbf6b7984024361811748464dcfbc8a285a231e6a59.jpg", + "image_caption": [ + "(b) Ablation of $\\alpha$ and k" + ], + "image_footnote": [], + "bbox": [ + 282, + 486, + 475, + 560 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 601, + 240, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To validate the effectiveness of our proposed design, we performed ablation experiments on UDR-S²Former. To this end, we utilized the training dataset of RainDS-Syn [48] for training our model, and evaluated its performance on the RDS scenes of the corresponding testing set. The experimental settings are kept consistent with the aforementioned ones. In the subsequent sections, we analyze the individual contributions of each module toward the performance.", + "bbox": [ + 75, + 623, + 467, + 743 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.1. Effectiveness of Image Reconstruction Module", + "text_level": 1, + "bbox": [ + 76, + 744, + 467, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ablation researches of parameters mentioned in IRM. Our ablation studies involve thorough research of the parameters employed in the method, and we have depicted the results in Fig.10. As demonstrated, the optimal value for each parameter lies within a particular range, which can be attributed to our initial design motivation holding theoretical significance within that range.", + "bbox": [ + 75, + 763, + 467, + 869 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Improvements of Sparse Sampling Attention (SSA) in IRM. We conducted a comparison of our sparse sampling", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "attention with sparse attention in ART [71] (SA), window-based self-attention [58] (WSA), and channel-dimension self-attention [68] (CSA). The results are shown in Table 4. It demonstrates that sparse sampling attention performs significantly better than previous designs, underscoring the importance of incorporating degradation information by adaptive sampling from a global perspective for the unified removal of complex rain scenes.", + "bbox": [ + 496, + 294, + 890, + 415 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f81794038d28ddb7a6e7b07cdfd45072b6c8dbbe9b3972855a0c783c60ee6ff0.jpg", + "table_caption": [ + "Table 4: Ablation studies on Table 5: Ablation studies on SSA (§3.1.1). UDR. (§3.1.1 and 3.1.2)." + ], + "table_footnote": [], + "table_body": "
SettingModelPSNRSSIM
iSA [71]36.390.962
iiWSA [58]36.150.959
iiiCSA [68]36.230.960
ivSSA (Ours)36.910.966
", + "bbox": [ + 511, + 454, + 694, + 512 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a2714fa2c7a88e6d2f67d14e022274e0bd2dc7af38187cf40bf85dbc32cbfe54.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingModelPSNRSSIM
iSSA w/o UD36.250.960
iiSSA w/o RS36.290.961
iiiLR w/o UD36.330.962
ivLR w/o RS36.450.963
vOurs36.910.966
", + "bbox": [ + 712, + 454, + 883, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gains of Uncertainty-Driven Ranking (UDR). In this section, we perform ablation studies of our uncertainty driven (UD) and ranking strategies (RS). The results in Table 5 demonstrate that leveraging uncertainty can effectively constrain the process of sparse sampling and facilitate the network to recover locally degraded regions. However, using uncertainty maps directly without employing a ranking strategy leads to a significant drop in performance, due to the lack of using explicit properties of uncertainty maps.", + "bbox": [ + 496, + 527, + 890, + 665 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 667, + 617, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we present a sparse sampling transformer with the uncertainty-driven ranking that removes rain streaks and raindrops in a unified approach. Our approach employs sparse sampling self-attention to effectively capture global degradation relationships in an adaptive manner. We explicitly leverage the uncertainty map via a ranking strategy to constrain the sampling process and facilitate local reconstruction. Our method achieves SOTA results while requiring minimal computational overhead, showcasing its superiority over existing approaches.", + "bbox": [ + 496, + 689, + 890, + 839 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), and the National Natural Science Foundation of China (Grant No. 61902275).", + "bbox": [ + 496, + 840, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13114", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 39(12):2481-2495, 2017. 3", + "[2] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5710-5719, 2020. 1, 3", + "[3] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 3", + "[4] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 3", + "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 7, 8", + "[6] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 6", + "[7] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Dualformer: Hybrid self-attention transformer for efficient image restoration. arXiv preprint arXiv:2210.01069, 2022. 3, 6", + "[8] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2", + "[9] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Yi Ye, and Erkang Chen. Msp-former: Multi-scale projection transformer for single image desnowing. arXiv preprint arXiv:2207.05621, 2022. 2", + "[10] Sixiang Chen, Tian Ye, Jun Shi, Yun Liu, JingXia Jiang, Erkang Chen, and Peng Chen. Dehrformer: Real-time transformer for depth estimation and haze removal from varicolored haze scenes. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3", + "[11] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 1", + "[12] Sen Deng, Mingqiang Wei, Jun Wang, Yidan Feng, Luming Liang, Haoran Xie, Fu Lee Wang, and Meng Wang. Detail-recovery image deraining via context aggregation networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14560-14569, 2020. 1", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.6.7", + "[14] David Eigen, Dilip Krishnan, and Rob Fergus. Restoring an image taken through a window covered with dirt or rain. In Proceedings of the IEEE international conference on computer vision, pages 633-640, 2013. 3, 7, 8", + "[15] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3855-3863, 2017. 7, 8", + "[16] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 7", + "[17] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 7", + "[18] Shuhang Gu, Deyu Meng, Wangmeng Zuo, and Lei Zhang. Joint convolutional analysis and synthesis sparse representation for single image layer separation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1708-1716, 2017. 7, 8", + "[19] Ming Hong, Jianzhuang Liu, Cuihua Li, and Yanyun Qu. Uncertainty-driven dehazing network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 906–913, 2022. 2, 3", + "[20] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In European Conference on Computer Vision, pages 163-180. Springer, 2022. 3", + "[21] Jie Huang, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Hybrid image enhancement with progressive laplacian enhancing unit. In Proceedings of the 27th ACM International Conference on Multimedia, page 1614-1622, 2019. 3", + "[22] Jie Huang, Feng Zhao, Man Zhou, Jie Xiao, Naishan Zheng, Kaiwen Zheng, and Zhiwei Xiong. Learning sample relationship for exposure correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9904-9913, June 2023. 3", + "[23] Jie Huang, Man Zhou, Yajing Liu, Mingde Yao, Feng Zhao, and Zhiwei Xiong. Exposure-consistency representation learning for exposure correction. In Proceedings of the 30th ACM International Conference on Multimedia, page 6309-6317, 2022. 3", + "[24] Jie Huang, Pengfei Zhu, Mingrui Geng, Jiewen Ran, Xingguang Zhou, Chen Xing, Pengfei Wan, and Xiangyang Ji. Range scaling global u-net for perceptual image enhancement on mobile devices. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, September 2018. 3" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13115", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800–801, 2008. 7", + "[26] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 7, 8", + "[27] Kui Jiang, Zhongyuan Wang, Chen Chen, Zheng Wang, Laizhong Cui, and Chia-Wen Lin. Magic elf: Image deraining meets association learning and transformer. arXiv preprint arXiv:2207.10455, 2022. 1", + "[28] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 2, 7, 8", + "[29] Yeying Jin, Ruoteng Li, Wenhan Yang, and Robby T Tan. Estimating reflectance layer from a single image: Integrating reflectance guidance and shadow/specular aware learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 1069-1077, 2023. 3", + "[30] Yeying Jin, Beibei Lin, Wending Yan, Wei Ye, Yuan Yuan, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution, 2023. 2", + "[31] Yeying Jin, Aashish Sharma, and Robby T Tan. Deshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021. 3", + "[32] Yeying Jin, Wending Yan, Wenhan Yang, and Robby T Tan. Structure representation network and uncertainty feedback learning for dense non-uniform fog removal. In Proceedings of the Asian Conference on Computer Vision, pages 2041–2058, 2022. 3", + "[33] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In European Conference on Computer Vision, pages 404-421. Springer, 2022. 3", + "[34] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2, 3", + "[35] Guanbin Li, Xiang He, Wei Zhang, Huiyou Chang, Le Dong, and Liang Lin. Non-locally enhanced encoder-decoder network for single image de-raining. In Proceedings of the 26th ACM international conference on Multimedia, pages 1056–1064, 2018. 7, 8", + "[36] Ruoteng Li, Loong-Fah Cheong, and Robby T Tan. Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1633-1642, 2019. 1, 2", + "[37] Xia Li, Jianlong Wu, Zhouchen Lin, Hong Liu, and Hongbin Zha. Recurrent squeeze-and-excitation context aggregation net for single image deraining. In Proceedings of the Eu" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ropean conference on computer vision (ECCV), pages 254-269, 2018. 2, 7, 8", + "[38] Yu Li, Robby T Tan, Xiaojie Guo, Jiangbo Lu, and Michael S Brown. Rain streak removal using layer priors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2736-2744, 2016. 7, 8", + "[39] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3", + "[40] Yuanchu Liang, Saeed Anwar, and Yang Liu. Drt: A lightweight single image deraining recursive transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 589-598, 2022. 1, 3", + "[41] Yun Liu, Zhongsheng Yan, Sixiang Chen, Tian Ye, Wenqi Ren, and Erkang Chen. Nighthazeformer: Single nighttime haze removal using prior query transformer. arXiv preprint arXiv:2305.09533, 2023. 2", + "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5", + "[43] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022. 7, 8", + "[44] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 2, 3, 4, 6", + "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 7", + "[46] Kuldeep Purohit, Maitreya Suin, AN Rajagopalan, and Vishnu Naresh Boddeti. Spatially-adaptive image restoration using distortion-guided networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2309-2319, 2021. 6", + "[47] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2482-2491, 2018. 3, 7, 8", + "[48] Ruijie Quan, Xin Yu, Yuanzhi Liang, and Yi Yang. Removing raindrops and rain streaks in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9147-9156, 2021. 1, 2, 3, 7, 8, 9", + "[49] Yuhui Quan, Shijie Deng, Yixin Chen, and Hui Ji. Deep learning for seeing through window with raindrops. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2463-2471, 2019. 1, 3, 7, 8", + "[50] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13116", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "better and simpler baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3937-3946, 2019. 1, 2, 3, 7, 8", + "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7", + "[52] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. arXiv preprint arXiv:2204.03883, 2022. 3", + "[53] Ming Tong, Yongzhen Wang, Peng Cui, Xuefeng Yan, and Mingqiang Wei. Semi-uformer: Semi-supervised uncertainty-aware transformer for image dehazing. arXiv preprint arXiv:2210.16057, 2022. 2, 3, 6", + "[54] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 3", + "[55] Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson WH Lau. Spatial attentive single-image deraining with a high quality real rain dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12270-12279, 2019. 1, 2", + "[56] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[57] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 1, 3, 7, 8", + "[58] Jie Xiao, Xueyang Fu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Image de-raining transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3, 7, 8, 9", + "[59] Ke Xu, Xin Tian, Xin Yang, Baocai Yin, and Rynson WH Lau. Intensity-aware single-image deraining with semantic and color regularization. IEEE Transactions on Image Processing, 30:8497-8509, 2021. 2", + "[60] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 7, 8", + "[61] Wenhan Yang, Robby T Tan, Jiashi Feng, Jiaying Liu, Zongming Guo, and Shuicheng Yan. Deep joint rain detection and removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1357-1366, 2017. 2, 7", + "[62] Wenhan Yang, Robby T Tan, Jiashi Feng, Shiqi Wang, Bin Cheng, and Jiaying Liu. Recurrent multi-frame deraining: Combining physics guidance and adversarial learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8569-8586, 2021. 1, 2", + "[63] Rajeev Yasarla and Vishal M Patel. Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "single image de-raining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8405-8414, 2019. 2, 3, 6, 7, 8", + "[64] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2", + "[65] Shaodi You, Robby T Tan, Rei Kawakami, and Katsushi Ikeuchi. Adherent raindrop detection and removal in video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1035-1042, 2013. 3", + "[66] Hu Yu, Jie Huang, Yajing Liu, Qi Zhu, Man Zhou, and Feng Zhao. Source-free domain adaptation for real-world image dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, page 6645-6654, 2022. 2", + "[67] Hu Yu, Jie Huang, Yajing Liu, Qi Zhu, Man Zhou, and Feng Zhao. Source-free domain adaptation for real-world image dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, page 6645-6654, 2022. 2", + "[68] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 3, 7, 8, 9", + "[69] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14821-14831, 2021. 7, 8", + "[70] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, 2022. 1", + "[71] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. arXiv preprint arXiv:2210.01427, 2022. 3, 5, 9", + "[72] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 4", + "[73] man zhou, Hu Yu, Jie Huang, Feng Zhao, Jinwei Gu, Chen Change Loy, Deyu Meng, and Chongyi Li. Deep fourier up-sampling. In Advances in Neural Information Processing Systems, volume 35, pages 22995-23008. Curran Associates, Inc., 2022. 3", + "[74] Lei Zhu, Zijun Deng, Xiaowei Hu, Haoran Xie, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Learning gated non-local residual for single-image rain streak removal. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2147–2159, 2020. 2", + "[75] Lei Zhu, Chi-Wing Fu, Dani Lischinski, and Pheng-Ann Heng. Joint bi-layer optimization for single-image rain streak removal. In Proceedings of the IEEE international conference on computer vision, pages 2526-2534, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 893 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13117", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/32196ea3-316d-4633-8abd-b15ee0703f11_model.json b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/32196ea3-316d-4633-8abd-b15ee0703f11_model.json new file mode 100644 index 0000000000000000000000000000000000000000..90485ef22b80eecae04daa9a43660411089a7ca5 --- /dev/null +++ b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/32196ea3-316d-4633-8abd-b15ee0703f11_model.json @@ -0,0 +1,3557 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.87, + 0.177 + ], + "angle": 0, + "content": "Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.202, + 0.731, + 0.238 + ], + "angle": 0, + "content": "Sixiang Chen\\(^{1,3*}\\) Tian Ye\\(^{1,3*}\\) Jinbin Bai\\(^{2}\\) Erkang Chen\\(^{3}\\) Jun Shi\\(^{4}\\) Lei Zhu\\(^{1,5\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.244, + 0.804, + 0.277 + ], + "angle": 0, + "content": "\\(^{1}\\)The Hong Kong University of Science and Technology (Guangzhou) \\(^{2}\\)National University of Singapore \\(^{3}\\)School of Ocean Information Engineering, Jimei University" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.278, + 0.721, + 0.295 + ], + "angle": 0, + "content": "\\(^{4}\\)Xinjiang University \\(^{5}\\)The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.297, + 0.634, + 0.312 + ], + "angle": 0, + "content": "{sixiangchen,owentianye}@hkust-gz.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.314, + 0.735, + 0.329 + ], + "angle": 0, + "content": "jinbin.bai@u.nus.edu, ekchen@jmu.edu.cn, junshi2022@gmail.com, leizhu@ust.hk" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.332, + 0.761, + 0.347 + ], + "angle": 0, + "content": "Project page: https://ephemeral182.github.io/UDR_S2Former_deraining" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.381, + 0.314, + 0.397 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.474, + 0.749 + ], + "angle": 0, + "content": "In the real world, image degradations caused by rain often exhibit a combination of rain streaks and raindrops, thereby increasing the challenges of recovering the underlying clean image. Note that the rain streaks and raindrops have diverse shapes, sizes, and locations in the captured image, and thus modeling the correlation relationship between irregular degradations caused by rain artifacts is a necessary prerequisite for image deraining. This paper aims to present an efficient and flexible mechanism to learn and model degradation relationships in a global view, thereby achieving a unified removal of intricate rain scenes. To do so, we propose a Sparse Sampling Transformer based on Uncertainty-Driven Ranking, dubbed UDR-S²Former. Compared to previous methods, our UDR-S²Former has three merits. First, it can adaptively sample relevant image degradation information to model underlying degradation relationships. Second, explicit application of the uncertainty-driven ranking strategy can facilitate the network to attend to degradation features and understand the reconstruction process. Finally, experimental results show that our UDR-S²Former clearly outperforms state-of-the-art methods for all benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.754, + 0.21, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.776, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Rain is a ubiquitous condition that negatively impacts various computer vision tasks [2, 70]. In real-world rain scenes, raindrops and rain streaks are irregularly superimposed on clean images. Image deraining is employed to restore the clean images from the complex rain degradations. According to previous work [48], the imaging model" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.381, + 0.628, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.458, + 0.619, + 0.478 + ], + "angle": 0, + "content": "(a) Fixed patch partition breaks the relationship modeling within diverse rain degradations" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.38, + 0.759, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.458, + 0.766, + 0.478 + ], + "angle": 0, + "content": "(b) We utilize the uncertainty-driven sparse sampling transformer to thoroughly explore arbitrary position degradation relationships." + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.381, + 0.899, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.459, + 0.864, + 0.467 + ], + "angle": 0, + "content": "(c) Uncertainty Map" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.48, + 0.893, + 0.537 + ], + "angle": 0, + "content": "Figure 1: Illustration of the breakdown of complex rain degradation relationships and the thumbnails of our main ideas. Colored places indicate degradations. Two-way arrows represent modeling between degradations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.542, + 0.893, + 0.573 + ], + "angle": 0, + "content": "of precipitation, inclusive of rain streaks and raindrops, can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.589, + 0.892, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {d s} = \\left(1 - \\mathcal {M} _ {r}\\right) \\odot (\\mathcal {B} + \\mathcal {S}) + \\eta \\mathcal {D}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.893, + 0.674 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}\\) and \\(S\\) denote the clean background and rain streak map. \\(\\mathcal{M}_r\\) is a binary mask used to judge whether the pixel belongs to the raindrops or the background. \\(\\mathcal{D}\\) is the raindrops and \\(\\eta\\) means global atmospheric lighting coefficient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.81 + ], + "angle": 0, + "content": "As noted by CCN [48], removing rain streaks and raindrops in a unified manner cannot be achieved by simply combining separate methods for removing either. This is due to the complex nature of the physical models involved and the wide array of possible degradation combinations. Previous models developed to address singular forms of degradations [11, 12, 36, 49, 50, 55, 62] face notable obstacles when dealing with irregularly dispersed and diverse rain degradation types." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Specifically, current SOTA methods for image deraining primarily concentrate on using ViTs due to their abilities to model long-range dependencies [27, 40, 57, 58]. Among these methods, window-based self-attention [40, 57, 58] has gained popularity due to its computational efficiency. However, as shown in Fig.1, we argue that utilizing window" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.213, + 0.888 + ], + "angle": 0, + "content": "*Equal contributions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.383, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) Lei Zhu (leizhu@ust.hk) is the corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.383, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13106" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.877, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.283, + 0.894, + 0.325 + ], + "angle": 0, + "content": "Figure 2: The uncertainty maps (bottom row) correspond to both real and synthetic samples (top row), with more significant uncertainty appearing in areas with severe and complicated degradations. This observation motivates us to use uncertainty explicitly to represent knowledge about degradation and to improve the model's understanding of degradation restoration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.471, + 0.452 + ], + "angle": 0, + "content": "based self-attention mechanisms can lead to incomplete degradation coverage, causing the breakdown of degradation relationships for unified rain degradation removal due to fixed window segmentation. This problem can be particularly pronounced when dealing with large raindrops or rain streaks at long distances simultaneously. However, in the case of complicated degradations, it is imperative to model the relationships between related forms of degradations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.471, + 0.819 + ], + "angle": 0, + "content": "Furthermore, with respect to the dense prediction task of rain removal, the density, shape, position, and size of raindrops and streaks are all uncertain, rendering it arduous for the network to restore clean images from diverse degradations. The CCN [48] requires using expensive and inflexible NAS to select an optimal architecture that effectively handles rain streaks and raindrops precisely. Inspired by uncertainty modeling [34] for image restoration, incorporating uncertainty learning can enhance the performance by reducing the error in model parameters [19], or serve as a regular term constraint to enhance the prediction quality of regions characterized by high uncertainty [44, 53]. Nevertheless, the mentioned design paradigm overlooks the importance of explicitly excavating uncertainty maps in facilitating the network's modeling of degradation features. For intricate rain scenes, we claim that learning uncertainty estimation can in turn affect the network to better focus on the complicated rain degradation areas. As depicted in Fig.2, the uncertainty map of the image exhibits greater concentration within the degraded region. It is our contention that by fully leveraging the properties of this uncertainty map, we can more effectively model the relationships of degradation and drive the network for the understanding of degradation restoration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Specifically, to address the problems above, we first design the Sparse Sampling Attention to deal with complicated rain scenes. It sparsely learns the relevant degradation relationships from the entire image, thereby alleviating the drawback of large-scale degradation modeling in window" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.895, + 0.542 + ], + "angle": 0, + "content": "based attention. Concurrently, we leverage the uncertainty map to guide feature learning for capturing more discriminative sampling features. To be exact, in order to fully leverage uncertainty information to promote the sampling of degradation features, we propose a novel ranking strategy and present a Constraint Matrix based on it. Such design further restricts the degree of attention to various rain degradations in the sampling process, boosting the modeling of relationships between degradations. Moreover, when restoring partially degraded regions, we consider the internal difference of the uncertainty map and use the ranking of the Correlation Map to strengthen the network to restore the degraded area by leveraging clean cues within the local regions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.542, + 0.89, + 0.556 + ], + "angle": 0, + "content": "Overall, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.56, + 0.891, + 0.619 + ], + "angle": 0, + "content": "- An uncertainty-driven sparse sampling transformer that fully models the global degradation relationships in an efficient manner is proposed to remove diverse rain streaks and raindrops." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.623, + 0.89, + 0.681 + ], + "angle": 0, + "content": "- We introduce a ranking strategy in the uncertainty map to enable the model to emphasize various rain degradation features in the sampling process through a constructed constraint matrix." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.686, + 0.89, + 0.73 + ], + "angle": 0, + "content": "- To enhance local reconstruction, we utilize the internal discrepancies within the uncertainty map to stimulate the network to extract credibly clean information." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.56, + 0.891, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.65, + 0.755 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.761, + 0.719, + 0.777 + ], + "angle": 0, + "content": "2.1. Single Image Deraining" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Rain streak removal. Image restoration from adverse weather has made significant advancements over the years, owing to its paramount significance [8, 9, 30, 41, 64, 66, 67]. Recently, the field of single image deraining has been predominantly dominated by learning-based methods [28, 36, 37, 50, 55, 58, 59, 61-63, 74, 75]. Zhu et al. [75] proposed a joint optimization algorithm that involves iteratively removing rain streaks from the background layer and non-rain" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13107" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.87, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.358 + ], + "angle": 0, + "content": "Figure 3: The overview of our UDR-\\(S^2\\)Former pipeline, which includes (a) our proposed restoration architecture, (b) the Local Reconstruction (LR) in the IRM module, (c) the Sparse Sampling Attention (SSA) constrained by the uncertainty map. For clarity, we only depict one local patch operation. The Feature Extraction and Global Modeling Modules, as well as the simple Refinement Block, are described in supplementary material due to page limitation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.473, + 0.528 + ], + "angle": 0, + "content": "details from the rain layer. This is achieved through the incorporation of three essential priors. PreNet [50] offered a recurrent layer to leverage the inter-stage dependencies of deep features, thereby constructing the progressive recurrent network. UMRL [63] proposed the uncertainty map to constrain the rain map in rainscapes and incorporate physical models to estimate the final deraining output. IDT [58] presented a transformer system comprising a complementary window-based transformer and spatial transformer, enabling improved capture of short- and long-range dependencies in rainy scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.65 + ], + "angle": 0, + "content": "Raindrop removal. Raindrops are frequently observed in rain scenes, and their diverse shapes and positions present difficulties removing them. Previous attempted to eliminate raindrops utilizing various methods, [14, 47, 49, 65]. The Eigen et al. [14] firstly introduced the learning-based paradigm for raindrop removal. AttenGAN [47] adopted the combination of GAN and attention mechanism to recover the clean image from raindrop degradations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.471, + 0.788 + ], + "angle": 0, + "content": "Most of the current design paradigms continue to create specialized networks for removing rain streaks or raindrops. A unified removal network design has still to be developed. While CCN [48] was the first to consider joint removal of rain degradation, their approach still requires expensive design such as NAS to address complex degradation characteristics. Other general-purpose networks [58, 68] disregard these two degradation characteristics, leading to high computation and parameters when improving performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.79, + 0.445, + 0.807 + ], + "angle": 0, + "content": "2.2. Vision Transformers for Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "ViTs exhibited superior global modeling capabilities, resulting in impressive performance on low-level vision tasks [3, 7, 10, 22, 52, 54, 68, 71] compared with previous CNNs' paradigm [4, 20, 21, 23, 24, 29, 31, 33, 73]. The window-based designs [39, 40, 52, 57, 58] were widely employed to overcome the computational complexity issue of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.895, + 0.53 + ], + "angle": 0, + "content": "\\(\\mathrm{O}(N^2)\\). Additionally, ART [71] utilized a combination of sparse and dense self-attention to manage computational overhead and achieved SOTA outcomes. Restormer [68] incorporated channel-based self-attention to circumvent the square-level complexity problem. Nevertheless, such designs limit the global receptive field of self-attention in spatial dimension and lack flexibility in dealing with intricate and changing degradations. We propose to use sparse sampling to address the aforementioned limitations by adaptively sampling information from the global field to meet the modeling requirements of the local." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.535, + 0.768, + 0.551 + ], + "angle": 0, + "content": "2.3. Uncertainty in Deep Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.555, + 0.892, + 0.828 + ], + "angle": 0, + "content": "According to Bayesian theory [34], uncertainty in deep learning can be classified into two types: (i) Aleatoric uncertainty, which refers to the inherent noise in the data. (ii) Epistemic uncertainty, which relates to the uncertainty of model parameters. Incorporating modeling uncertainty into the network can enhance its robustness and performance [1, 2]. In low-level fields, leveraging uncertainty can enable the network to prioritize reducing model prediction error [19] or serving as a loss function [32, 44, 53] to improve the reconstruction quality of areas with high uncertainty. Furthermore, uncertainty can assist in the more precise estimation of critical parameters in physical models [63]. Same as [44, 53], we focus on aleatoric uncertainty in this paper. However, these paradigms do not explicitly utilize the distinctive characteristic of the uncertainty map to restrict the network's acquisition of features. This paper proposes leveraging uncertainty-driven ranking to facilitate the network to represent degradations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.805, + 0.851 + ], + "angle": 0, + "content": "3. Proposed UDR-S²Former Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Network architecture. The architecture of the proposed network is illustrated in Fig.3. The network accepts a rain image as input, conducts image processing within the net" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13108" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.086, + 0.874, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.214, + 0.877, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.895, + 0.406 + ], + "angle": 0, + "content": "Figure 4: Top row: Constructing a constraint matrix according to the ranking strategy derived from the uncertain map. To simplify the explanation, we only present a one-dimensional ranking strategy. Bottom row: The procedure for utilizing the correlation map and ranking strategy to produce the matrix utilized for modulating the self-attention map. We only present the correlation map and ranking strategy within a patch for easy visualization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.473, + 0.635 + ], + "angle": 0, + "content": "work, and generates a high-quality restored image as output. Specifically, 1) the degraded image is fed into the Feature Extraction stage to acquire knowledge of features at various scales. This process is accomplished in four stages, each comprising basic convolutional blocks. 2) Upon completion of the feature extraction stage, we employ vanilla transformer to capture deep-level information and ensure the comprehensive utilization of global information1. 3) The Image Reconstruction Module (IRM) is employed to represent degradation relationships in the form of sparse sampling self-attention, which is driven by uncertain learning. Further, it can trigger clean cues excavation to guide the restoration for local reconstruction. Skip connections are utilized in both the feature extraction stage and each stage of the image reconstruction module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.473, + 0.759 + ], + "angle": 0, + "content": "Preliminary for Uncertainty. For image deraining, we model the difference between the image \\(\\mathcal{B}_{gt}\\) and its estimated deraining image \\(\\hat{\\mathcal{B}}\\) as a Laplace distribution. The motivation for this choice is that the Laplace distribution is better suited for characterizing the edges of details in images than the Gaussian distribution. In addition, the L1 loss refers to the Laplace distribution [72]. The likelihood function can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.768, + 0.47, + 0.81 + ], + "angle": 0, + "content": "\\[\np (\\mathcal {B} _ {g t}, \\sigma ; \\mathcal {R} _ {d s}) = \\frac {1}{2 \\sigma} \\exp \\left(- \\frac {| | \\hat {\\mathcal {B}} - \\mathcal {B} _ {g t} | | _ {1}}{\\sigma}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.82, + 0.471, + 0.868 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}_{gt}\\) denotes the mean of this distribution. \\(\\hat{\\mathcal{B}}\\) is output calculated by network from rain image \\(\\mathcal{R}_{ds}\\). \\(\\sigma\\) means the uncertainty (variance) of deraining image \\(\\hat{\\mathcal{B}}\\). To simplify" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.471, + 0.903 + ], + "angle": 0, + "content": "1Restricted by the number of pages, we introduce the points 1) and 2) of our architecture in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.44 + ], + "angle": 0, + "content": "the calculation, we transform the likelihood function into its log form and maximize it:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.449, + 0.892, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\underset {\\sigma} {\\arg \\max } \\ln p \\left(\\mathcal {B} _ {g t}, \\sigma ; \\mathcal {R} _ {d s}\\right) = \\left(- \\frac {\\left| \\left| \\hat {\\mathcal {B}} - \\mathcal {B} _ {g t} \\right| \\right| _ {1}}{\\sigma}\\right) - \\ln \\sigma . \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.627 + ], + "angle": 0, + "content": "During the learning process, the aforementioned formula is inverted and utilized as a loss function, which is minimized for optimization. Additionally, we also follow [44] to circumvent the issue of training instability resulting from the presence of uncertainties in the form of zero values. Without being limited to a loss function, we aim to employ the ranking strategy to explicitly utilize estimated uncertainty. It enhances the modeling of complicated degradation relationships and drive the network to restore degraded regions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.629, + 0.771, + 0.645 + ], + "angle": 0, + "content": "3.1. Image Reconstruction Module" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.649, + 0.892, + 0.679 + ], + "angle": 0, + "content": "For the image reconstruction module of each stage, we can express it as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.689, + 0.892, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I R M} (\\mathcal {X}, \\mathcal {U}) = \\left\\{\\operatorname {S S A} (\\mathcal {X}, \\mathcal {U}), \\operatorname {L R} (\\mathcal {X}, \\mathcal {U}) \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.761 + ], + "angle": 0, + "content": "where \\(\\{\\cdot\\}\\) indicates that these two modules are alternately formed, with \\(\\mathcal{X}\\) denoting the image feature, and \\(\\mathcal{U}\\) representing the corresponding uncertainty map." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.763, + 0.892, + 0.794 + ], + "angle": 0, + "content": "3.1.1 Sparse Sampling Attention Constrained by Uncertainty Map" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The degradation of a large-scale area (raindrops) or long-distance span (rain streaks) can result in the loss of local relationships, particularly when a fixed window division is utilized. Moreover, a window-based design may limit the receptive field to a local area, leading to the loss of global knowledge when dealing with complicated degradations. However, it is crucial to fully exploit the modeling of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13109" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.475, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.226, + 0.47, + 0.309 + ], + "angle": 0, + "content": "Figure 5: Visualization of sparse sampling. The yellow region denotes an \\(8 \\times 8\\) window patch. The red pixels are coordinates obtained from global sparse sampling to interact window patch for modeling the corresponding degradation relationship, which is not limited to the local patch (we draw multi-head sampling coordinates)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.469, + 0.405 + ], + "angle": 0, + "content": "degradation relationships for removing diverse rain degradations due to their complicated degraded properties. In this part, we propose an uncertainty-driven sparse sampling approach to learn global coordinates for capturing associated rain degradations adaptively, effectively modeling degradations in diverse regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.469, + 0.557 + ], + "angle": 0, + "content": "For a given image feature \\(\\mathcal{X} \\in \\mathbb{R}^{C \\times H \\times W}\\), our primary objective is to establish a model that can accurately capture the degradation relationships. For each part of the degradations, the network should be prompted to concurrently consider correlation degradations occurring in other image regions to model it. Specifically, we let the network adaptively learn to match the degraded coordinates of other regions in the whole image, and map them to the same image patch, so as to model the degradation relationship at a low cost. The coordinates are learned as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.566, + 0.469, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S}, \\mathcal {B} = \\mathcal {F} \\left(\\mathcal {X} ^ {\\in \\mathbb {R} ^ {C \\times H \\times W}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.69 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}(\\cdot)\\) represents the simple convolution and avgpooling operation to estimate scaling factors \\(S\\in \\mathbb{R}^{C\\times H\\times W\\times 2}\\) and biases \\(\\mathcal{B}\\in \\mathbb{R}^{C\\times H\\times W\\times 2}\\) (last dimension denotes x-axis and y-axis coordinates). We use them to transform the original coordinates of the feature map by multiplication and addition:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.705, + 0.469, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C o o r d s} ^ {T} (x, y) = \\mathcal {G} \\left(\\operatorname {C o o r d s} ^ {O} (x, y) \\times \\mathcal {S} + \\mathcal {B}\\right), \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.727, + 0.42, + 0.747 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.47, + 0.901 + ], + "angle": 0, + "content": "scaling factors and biases change the sampling location for each patch. \\(\\mathrm{Cords}^T (x,y)\\) denotes the coordinates of sparse sampling from the global feature, which is leveraged to the local patch \\(\\mathcal{X}_i^{S\\in \\mathbb{R}^{C\\times \\frac{H}{M}\\times \\frac{H}{M}}}\\) (\\(i\\) denotes the \\(i\\)-th patch of \\(\\mathcal{M}^2\\) patches) after carrying out the function \\(\\mathcal{G}\\) of torch nn. functional grid sample, as shown in Fig.3 (c). We utilize the information obtained through global sparse sampling to match each degraded patches for corresponding degradation modeling and propose a novel attention mech" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.829, + 0.107 + ], + "angle": 0, + "content": "anism, dubbed Sparse Sampling Attention (SSA):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.118, + 0.892, + 0.16 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S S A} = \\mathbf {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {i} ^ {P} \\mathcal {K} _ {i} ^ {S ^ {\\mathrm {T}}}}{\\sqrt {\\mathcal {D}}} + p\\right) \\mathcal {V} _ {i} ^ {S}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.171, + 0.892, + 0.233 + ], + "angle": 0, + "content": "where \\(\\mathcal{Q}_i^P\\) denotes the queries projected from original feature of \\(i\\)-th patch. \\(\\mathcal{K}_i^S\\) and \\(\\mathcal{V}_i^S\\) are obtained from global sampling feature. \\(\\mathcal{D}\\) means the dimension number and \\(p\\) is the position embedding like [42]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.339 + ], + "angle": 0, + "content": "Uncertainty map driven: Drawing inspiration from the obvious representation of deteriorated regions in the uncertainty map, we aim to incorporate the uncertainty map as an explicit constraint to restrict the level of attention towards the degradations in the sparse sampling procedure, to ensure the modeling of the degradation relationships. The operation is presented in Fig.4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.403 + ], + "angle": 0, + "content": "More specifically, with regard to the uncertain map \\(\\mathcal{U} \\in \\mathbb{R}^{C \\times H \\times W}\\), we employ a ranking strategy to construct a constraint matrix \\(\\mathcal{C} \\in \\mathbb{R}^{C \\times H \\times W}\\), which serves as a means of constraining the sampling process:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.415, + 0.892, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathcal {C} ^ {n} \\left(\\mathcal {U} ^ {n}\\right) \\right] _ {x y} = \\left\\{ \\begin{array}{l l} 1 & \\mathcal {U} _ {x y} ^ {n} \\geq \\mathcal {U} ^ {n} \\text {o f r a n k} \\gamma \\\\ \\beta & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.554 + ], + "angle": 0, + "content": "where \\(\\beta\\) and \\(\\gamma\\) are the constraint factor and a threshold value of \\(n\\)-th dimension. Such a constraint matrix contains crucial cues from the uncertainty map with almost no computational overhead, and we utilize it to regularize our modeling of degradation relationships. The Eq.5 is further optimized as:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.555, + 0.891, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} ^ {U}, \\mathcal {B} ^ {U} = \\mathcal {F} \\left(\\mathcal {X} ^ {\\in \\mathbb {R} ^ {C \\times H \\times W}} \\times \\mathcal {C} _ {x y}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.688 + ], + "angle": 0, + "content": "wherein \\(\\times\\) represents the element-wise multiplication. The \\(\\beta\\) of \\(\\mathcal{C}_{xy}\\) reduces the influence of irrelevant background areas in multiplication form. \\(S^U\\) and \\(\\mathcal{B}^U\\) are leveraged to obtain robust coordinates via Eq.6, which promotes the global sparse sampling while mitigating the interference of the background region. Thereby it facilitates the network to concentrate on degradation relationship modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Our visualization is depicted in Fig.5. As we expected, the sampled red points adaptively have related degradation situations with the target point (see \\(8 \\times 8\\) yellow window). The sparse sampling strategy with uncertainty helps model the relation between local and long-range correlation degradations. Additionally, The most concentration of sampled pixels in a larger area maintains some semantic information and coherence, despite being sampled at a distance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Discussion I: The merits of our sparse sampling compared with previous related work for image restoration. Compared to the latest sparse attention [71], which permitted each token to interact with a limited number of tokens, and with a fixed interval size. However, it still involved a manually designed mechanism. In complex rainy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13110" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.476, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.167, + 0.476, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.306, + 0.47, + 0.446 + ], + "angle": 0, + "content": "Figure 6: Top row: Our motivation about needing to strengthen the network to exploit non-degradation regions. Most information is lost due to the complicated rain degradations in the real world, so degradation-free information is needed to reconstruct the clean area. Bottom row: Visualizations on uncertainty map, correlation map and self-attention map. The attention points of the correlation map and the self-attention map differ. We can utilize the correlation map to direct the network's attention towards areas with significant differences, thus enhancing the restoration process through self-attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.47, + 0.544 + ], + "angle": 0, + "content": "scenes, such an approach cannot flexibly capture degradation information from different locations. Our design aims to enable the network to learn the coordinates that can sample degradation information from any position, leading to an enhanced modeling performance of the degradation relationships and network flexibility." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.471, + 0.802 + ], + "angle": 0, + "content": "Discussion II: uncertainty driven compared with prior paradigm and rain mask. Regarding uncertainty, the previous UMRL [63] approach used the uncertainty to optimize the rain streak map based on the physical model of the rain streak. Still, it only focused on optimizing the final map, ignoring the network's learning of the intermediate features. Our method uses uncertainty explicitly to drive the network to provide motivation-driven constraints and optimization for the degradation modeling and local reconstruction processes. Compared to the learned rain mask needed GT [46], i) uncertainty map has a theoretical unsupervised loss for practicing. ii). It outperforms the rain mask in representing the network's focus on intricate rain areas beyond mere location. iii). For large-area raindrops with a complicated physical model, it is difficult to learn a rain mask, but the uncertainty map can solve it well by using the discriminative representation of degradations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.466, + 0.821 + ], + "angle": 0, + "content": "3.1.2 Local Reconstruction with Correlation Ranking" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In rain scenes, information in locally degraded regions is often obscured by rain streaks or raindrops, making it difficult to restore, see Fig.6. To address this issue, it is vital to leverage clean cues in the rain-free area. Building on the discriminative representation of degraded areas and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "other regions in the uncertainty map, we propose using the uncertainty map to generate a correlation map. This map explicitly adjusts the attention map through a ranking approach, as shown in Fig.4 and Fig.3(b), which promotes the network to better leverage clean cues for reconstruction in the form of Query-Key." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.181, + 0.892, + 0.28 + ], + "angle": 0, + "content": "We begin by partitioning uncertainty map \\(\\mathcal{U}^{\\in \\mathbb{R}^{C\\times H\\times W}}\\) into local patches \\(\\mathcal{U}^{P\\in \\mathbb{R}^{C\\times \\frac{H}{M}\\times \\frac{H}{M}}}\\) that correspond to those patches in the feature map. The presence of discriminative disparities within each patch of the uncertainty map motivates us to calculate the correlation map (CR) between them:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.28, + 0.892, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C R} _ {i} = \\mathcal {U} _ {i} ^ {P} \\times \\mathcal {U} _ {i} ^ {P ^ {\\mathrm {T}}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.893, + 0.446 + ], + "angle": 0, + "content": "Upon obtaining the correlation map, our purpose is to leverage the degradation-free region to facilitate the restoration process of the degraded areas during image reconstruction. Significant differences (low correlation) between degradations and background in the uncertainty map can be considered a prompt, promoting the network to use clean cues from the background for local reconstruction. To this end, we modulate the self-attention map by selecting the low-correlation regions in the correlation map:" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.457, + 0.895, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L R} = \\text {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {i} ^ {P} \\mathcal {K} _ {i} ^ {P T}}{\\sqrt {\\mathcal {D}}} \\times \\underbrace {[ - \\alpha \\boldsymbol {\\Sigma} _ {i} + (1 + \\alpha) ] + p} _ {\\text {M o d u l a t i o n P r o c e s s i n g}}\\right) \\mathcal {V} _ {i} ^ {P}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.536, + 0.892, + 0.567 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a modulation factor. The \\(\\Sigma\\) is acquired via the Top k ranking approach, which can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.578, + 0.892, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\left[ \\boldsymbol {\\Sigma} _ {i} \\left(\\mathcal {C R} _ {i}\\right) \\right] _ {x y} = \\left\\{ \\begin{array}{l l} 1 & \\left(\\mathcal {C R} _ {i}\\right) _ {x y} \\in \\operatorname {T o p k} (\\text {r o w} y) \\\\ 0 & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.692 + ], + "angle": 0, + "content": "In the experiments section, the key parameters of the above used are further studied. About our image reconstruction module, the self-attention FFNs are identical to vanilla vision transformer [13]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.644, + 0.71 + ], + "angle": 0, + "content": "4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Our UDR-\\(S^2\\)Former network will predict the derained result and an uncertainty map. Hence, the total loss of our network is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.773, + 0.891, + 0.821 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {p s n r}} \\left(\\mathcal {B} _ {\\text {p r e}}, \\mathcal {B} _ {\\text {g t}}\\right) + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c e p t u a l}} \\left(\\mathcal {B} _ {\\text {p r e}}, \\mathcal {B} _ {\\text {g t}}\\right) \\\\ + \\lambda_ {3} \\mathcal {L} _ {U D L} (\\mathcal {B} _ {\\text {p r e}}, \\mathcal {U}), \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.822, + 0.892, + 0.913 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}_{pre}\\) and \\(\\mathcal{B}_{gt}\\) denote the predicted result of image deraining and the corresponding ground truth. \\(\\mathcal{U}\\) denotes the predicted uncertainty map. \\(\\mathcal{L}_{psnr}\\) and \\(\\mathcal{L}_{perceptual}\\) are the PSNR and perceptual losses, see [6,7]. \\(\\mathcal{L}_{UDL}\\) denotes the uncertainty loss, see [44, 53] for its definition. \\(\\lambda_{1},\\lambda_{2}\\) and \\(\\lambda_{3}\\) are empirically set to 1, 0.2 and 1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13111" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.076, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 1: Quantitative results compared with SOTA methods on the Rain200H [61] and Rain200L [61] datasets. Underline and bold indicate the first and second best results." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.118, + 0.413, + 0.289 + ], + "angle": 0, + "content": "
MethodRain200H [61]Rain200L [61]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
GMM [38]14.710.43028.990.875
JCAS [18]14.870.47130.050.897
DDN [15]26.360.80334.930.958
NLEDN [35]29.510.89138.560.980
RESCAN [37]27.450.82135.080.959
PreNet [50]29.040.89037.120.976
UMRL [63]28.710.88736.430.973
JORDER-E [60]28.580.87636.900.973
MSPFN [28]29.660.89039.480.984
CCN [48]29.990.91438.260.981
MPRNet [69]30.760.90839.890.985
DGUNet [43]30.850.91140.230.986
Uformer [57]30.800.91140.200.986
Restormer [68]31.390.91640.580.987
IDT [58]32.100.93440.740.988
NAFNet [5]30.980.91240.450.987
UDR-S2Former32.59↑0.490.937↑0.00340.96↑0.220.989↑0.001
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.31, + 0.21, + 0.328 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.331, + 0.288, + 0.348 + ], + "angle": 0, + "content": "5.1. Implementaion Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.639 + ], + "angle": 0, + "content": "Our UDR-\\(S^2\\)Former model utilizes a 5-level encoder-decoder architecture. Within this architecture, the number of channel dimensions increases to \\(\\{16,32,64,128,256\\}\\) across levels 1 to 5. Additionally, the Feature Extraction stage contains \\(\\{4,6,7,8\\}\\) convolutional blocks, while the transformer block number in the latent layer is 8, with 16 heads. Regarding reconstruction, we employ \\(\\{3,6,7,8\\}\\) Image Reconstruction Modules, composed alternately of Sparse Sampling Attention and Local Reconstruction, with \\(\\{1,2,4,8\\}\\) heads. The window size is fixed at \\(8\\times 8\\). Finally, we set the constraint factor \\(\\beta\\), threshold value \\(\\gamma\\), modulation factor \\(\\alpha\\), and the k value for the ranking strategy to 0.6, 0.8, 0.2, and 0.8, respectively. For the self-attention mechanism in this paper, we all use multi-head self-attention, which is consistent with vanilla ViT [13]. Additionally, during the last image reconstruction module of each stage, we convert the output feature and corresponding uncertainty map into the final image format, and the \\(\\mathcal{L}_{UDL}\\) is utilized to supervise the learning process of uncertainty map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.806 + ], + "angle": 0, + "content": "During the training phase, we employed the Adam optimizer with initial momentum \\(\\beta_{1} = 0.9\\) and \\(\\beta_{1} = 0.999\\). We initially set the learning rate to 0.0003 and utilized a cyclic learning rate adjustment strategy, whereby the maximum learning rate is set to 0.00036. We trained our model with a data augmentation strategy that included randomly cropping \\(256 \\times 256\\) patches and using horizontal flipping and random image rotation to a fixed angle. Our training process involved \\(6 \\times 10^{5}\\) steps. To leverage perceptual loss, we used the first and third layers of VGG19 [51]. Our model is implemented using PyTorch [45] and the RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.809, + 0.365, + 0.824 + ], + "angle": 0, + "content": "5.2. Evaluation Metrics and Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Evaluation Metrics. In accordance with prior deraining methodologies [48, 58], follow [16, 17], the performance of the model is evaluated using PSNR [25] and SSIM [56]. As suggested in [48, 61], the assessment of PSNR and SSIM is based on the luminance channel, which refers to the Y" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.694, + 0.107 + ], + "angle": 0, + "content": "channel of the YCbCr space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Rain Streak Datasets. In order to evaluate the effectiveness of our rain streak removal method, we follow the strategy of previous work [48] and select two benchmark datasets, namely Rain200H and Rain200L [61]. Both two datasets have 1800 synthetic images for training and 200 images for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Raindrop Datasets. We also evaluate our method in a raindrop dataset (i.e., AGAN-Data) collected by Qian et al. [47]. AGAN-Data has 861 images for training and 58 images for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.455 + ], + "angle": 0, + "content": "Raindrops and Rain Streak Datasets. Our paper aims to develop a methodology that effectively addresses the challenges posed by the presence of large-area raindrops and complex rain streaks. To evaluate the proposed approach, we utilize the RainDS benchmark dataset [48], which includes both real-world and synthetic images, dubbed as RainDS-Real and RainDS-Syn, respectively. These datasets contain images of scenes that include rain streaks only (RS), raindrops only (RD), or both(RDS). RainDS-Syn has 3600 image pairs, with 3000 images used for training and the remaining 600 images for testing. Meanwhile, RainDS-Real comprises of 750 images, with 450 images for training and 300 images for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.458, + 0.855, + 0.473 + ], + "angle": 0, + "content": "5.3. Experimental Evaluation on Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Compared Methods. Regarding removing raindrops and rain streaks in images, we conduct extensive experiments to compare various algorithms that can be used for image rain removal. (i) We compare previous SOTA methods (including GMM [38], JCAS [18], DDN [15], NLEDN [35], RESCAN [37], PreNet [50], UMRL [63], JORDER-E [60], MSPFN [28], CCN [48], IDT [58]). (ii) We also compare our network against universal image restoration methods, including MPRNet [69], DGUNet [43], Uformer [57], Restormer [68], NAFNet [5]). For the specific raindrop removal dataset, we compare against Eigen's model [14], Pix2Pix [26], AttentGAN [47], Quan's network [49], CCN [48], and IDT [58]. In the absence of pre-trained models, we conduct model retraining by utilizing publicly available code and subsequently evaluate the best model performance on test datasets to ensure a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparison. Tables 1, 2, and 3 reports the quantitative results of our network and state-of-the-art methods on four benchmark datasets, which are Rain200H, Rain200L, RainDS, and AGAN-Data. As demonstrated in these Tables, our approach delivers superior metric results over compared state-of-the-art (SOTA) methods for single rain streaks or raindrop degradations. Moreover, by incorporating uncertainty to steer the network's attention towards challenging degradations, our method outperforms the IDT algorithm by 0.49dB, 0.22dB and 1.01dB on the Rain200H [61], Rain200L [61] datasets and AGAN-Data [47], while significantly outperforming previ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13112" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.174, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.091, + 0.264, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.091, + 0.352, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.091, + 0.441, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.091, + 0.529, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.091, + 0.618, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.091, + 0.707, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.091, + 0.795, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.091, + 0.884, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.167, + 0.155, + 0.178 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.167, + 0.236, + 0.177 + ], + "angle": 0, + "content": "(b) CCN" + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.167, + 0.335, + 0.177 + ], + "angle": 0, + "content": "(c) MPRNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.167, + 0.424, + 0.177 + ], + "angle": 0, + "content": "(d) DGUNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.167, + 0.494, + 0.177 + ], + "angle": 0, + "content": "(e) Resto" + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.167, + 0.591, + 0.177 + ], + "angle": 0, + "content": "(f) IDT" + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.167, + 0.692, + 0.177 + ], + "angle": 0, + "content": "(g) NAFNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.167, + 0.77, + 0.177 + ], + "angle": 0, + "content": "(h) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.167, + 0.855, + 0.177 + ], + "angle": 0, + "content": "(i) GT" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.18, + 0.342, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.294, + 0.239, + 0.305 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.18, + 0.475, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.228, + 0.432, + 0.239 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.18, + 0.61, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.228, + 0.568, + 0.239 + ], + "angle": 0, + "content": "(c) CCN" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.18, + 0.745, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.228, + 0.717, + 0.239 + ], + "angle": 0, + "content": "(d) MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.18, + 0.881, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.228, + 0.845, + 0.239 + ], + "angle": 0, + "content": "(e) DGUNet" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.244, + 0.475, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.293, + 0.452, + 0.304 + ], + "angle": 0, + "content": "(f) Restomer" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.244, + 0.61, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.293, + 0.566, + 0.304 + ], + "angle": 0, + "content": "(g) IDT" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.244, + 0.745, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.293, + 0.715, + 0.304 + ], + "angle": 0, + "content": "(h) NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.244, + 0.881, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.293, + 0.84, + 0.303 + ], + "angle": 0, + "content": "(i) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.306, + 0.885, + 0.32 + ], + "angle": 0, + "content": "Figure 7: Visual comparisons of removal for raindrops and rain streaks from RainDS-Syn [48] dataset. Zoom it for a better illustration." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons of various SOTA approaches on the RainDS [48] benchmark. Our proposed method outperforms all other approaches on both synthetic and real-world datasets, including all types of precipitation (i.e., rain streaks (RS), raindrops (RD), and a combination of both (RDS)). The results are indicated in bold and underline, representing the first and second-best performances, respectively. \\(\\uparrow\\) means the higher is better. The #GFLOPs are calculated by \\(256\\times 256\\) image resolution for a fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.383, + 0.888, + 0.572 + ], + "angle": 0, + "content": "
MethodVenueRainDS-SynRainDS-Real#Param #GFLOPs
RSRDRDSRSRDRDS
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
GMM [38]CVPR'201626.660.78123.040.79321.500.66923.730.56018.600.55421.350.576--
JCAS [18]ICCV'201726.460.78623.150.81120.910.67124.040.55618.180.55521.220.585--
DDN [15]CVPR'201730.410.86927.920.88526.850.79624.850.68323.120.64222.470.606--
NLEDN [35]ACMMM'201836.240.95834.870.95732.130.91727.020.72324.710.67124.060.650--
RESCAN [37]ECCV'201830.990.88729.900.90727.430.81826.700.68324.230.63723.230.5870.15M32.32G
PreNet [50]CVPR'201936.630.96834.580.96432.210.93426.430.72924.420.67923.570.6490.17M66.58G
UMRL [63]CVPR'201935.760.96233.590.95831.570.92925.890.72623.930.67623.010.6470.98M16.50G
JORDER-E [60]TPAMI'201933.650.92533.510.94430.050.87026.560.71324.340.66223.540.6294.17M273.68G
MSPFN [28]CVPR'202038.610.97536.930.97334.080.94726.450.72724.490.68124.110.65121.00M708.44G
CCN [48]CVPR'202139.170.98137.300.97634.790.95727.460.73725.140.70124.930.6793.75M245.85G
MPRNet [69]CVPR'202140.810.98137.030.97234.990.95627.290.73625.260.70124.960.6813.64M148.55G
DGUNet [43]CVPR'202241.090.98337.560.97535.340.95927.520.73725.330.70224.990.68312.18M199.74G
Uformer [57]CVPR'202240.690.97237.080.96634.990.95426.890.73025.310.70124.830.68620.63M43.86G
Restomer [68]CVPR'2022(Oral)41.420.98038.780.97636.080.96127.390.74225.380.70224.920.68526.10M140.99G
IDT [58]TPAMI'202241.610.98339.090.98036.230.96027.510.74325.670.70624.990.68916.00M61.90G
NAFNet [5]ECCV'202240.390.97237.230.97434.990.95727.490.72925.230.70124.640.66340.60M16.19G
UDR-S2Former (Ours)ICCV'202342.39↑0.780.988↑0.00539.78↑0.690.983↑0.00336.91↑0.680.966↑0.00627.90↑0.390.745↑0.00226.01↑0.340.709↑0.00325.52↑0.530.691↑0.0028.53M21.58G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.586, + 0.469, + 0.626 + ], + "angle": 0, + "content": "Table 3: Quantitative results compared with SOTA methods on the AGAN-Data [47]. Red and blue indicate the first and second best results." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.628, + 0.402, + 0.723 + ], + "angle": 0, + "content": "
MethodPSNR†SSIM†
(ICCV'2013)Eigen's model [14]21.310.757
(CVPR2017)Pix2Pix [26]27.200.836
(CVPR'2018)AttenGAN [47]31.590.917
(ICCV'2019)Quan's network [49]31.370.918
(CVPR'2021)CCN [48]31.340.929
(TPAMI'2022)IDT [58]31.630.936
UDR-S2Former32.640.943
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "ous uncertainty-based [63] methods. Table 2 demonstrates that our approach surpasses all prior designs for jointly removing rain streaks and raindrops. While networks like Restormer [68] and IDT [58] can model long-distance dependencies sufficiently, our method is able to effectively model complex degradation relationships in rain-dominated landscapes (i.e., raindrops and rain streaks both) with uncertainty. Additionally, sparse sampling further facilitates the network's ability to obtain crucial global degradation information at a low cost, see Table 2. Based on the indicators," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.688 + ], + "angle": 0, + "content": "our approach exhibits a significant advantage over the latest designs, including Restormer (36.08PSNR \\(\\rightarrow\\) 36.91PSNR) and IDT (36.23PSNR \\(\\rightarrow\\) 36.91PSNR), in both real-world and synthetic datasets. Additionally, we also present the comparison of speed and memory cost in the supplementary material to further demonstrate our superiority." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Visual Comparison. We present the comparisons of the visual effects in Fig.7-9. As depicted in the figures, we observe that in a scene containing rain streaks and raindrops, the scene information is either occluded by a broad range of raindrops or masked by dense rain streaks of varying scales. Other approaches either struggle to remove complex degradations or significantly compromise image details. Our approach leverages degradation relationship modeling to facilitate image restoration, and utilizes the clean information to recover intricate rain degradations. The resulting output exhibits higher fidelity with the more accurate restoration of great details. In real-world scenarios, our method enables the removal of various forms of degradation while recovering fine details, outperforming previous methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13113" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.344, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.255, + 0.24, + 0.266 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.09, + 0.472, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.163, + 0.428, + 0.175 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.09, + 0.61, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.163, + 0.566, + 0.175 + ], + "angle": 0, + "content": "(c) CCN" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.09, + 0.747, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.163, + 0.718, + 0.175 + ], + "angle": 0, + "content": "(d) MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.09, + 0.885, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.163, + 0.856, + 0.175 + ], + "angle": 0, + "content": "(e) DGUNet" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.183, + 0.472, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.255, + 0.448, + 0.267 + ], + "angle": 0, + "content": "(f) Restormer" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.184, + 0.61, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.255, + 0.566, + 0.267 + ], + "angle": 0, + "content": "(g) IDT" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.183, + 0.747, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.256, + 0.716, + 0.267 + ], + "angle": 0, + "content": "(h) NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.183, + 0.885, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.256, + 0.843, + 0.267 + ], + "angle": 0, + "content": "(i) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.269, + 0.71, + 0.283 + ], + "angle": 0, + "content": "Figure 8: Visual results for real-world sample of RainDS-Real [48] dataset." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.293, + 0.178, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.359, + 0.149, + 0.367 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.293, + 0.277, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.359, + 0.245, + 0.366 + ], + "angle": 0, + "content": "(b) CCN" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.293, + 0.376, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.359, + 0.351, + 0.366 + ], + "angle": 0, + "content": "(c) DGUNet" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.293, + 0.476, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.359, + 0.455, + 0.366 + ], + "angle": 0, + "content": "(d) Restomer" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.37, + 0.178, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.434, + 0.143, + 0.442 + ], + "angle": 0, + "content": "(e) IDT" + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.37, + 0.277, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.434, + 0.251, + 0.442 + ], + "angle": 0, + "content": "(f) NAFNet" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.37, + 0.376, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.434, + 0.344, + 0.442 + ], + "angle": 0, + "content": "(g) Ours" + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.37, + 0.476, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.434, + 0.44, + 0.442 + ], + "angle": 0, + "content": "(h) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.445, + 0.468, + 0.472 + ], + "angle": 0, + "content": "Figure 9: Visual results for real-world sample of RainDS-Real [48] dataset, Please zoom it for a better illustration." + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.477, + 0.236, + 0.486 + ], + "angle": 0, + "content": "(a) Ablation of \\(\\beta\\) and \\(\\gamma\\)" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.487, + 0.276, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.477, + 0.435, + 0.486 + ], + "angle": 0, + "content": "(b) Ablation of \\(\\alpha\\) and k" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.487, + 0.476, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.562, + 0.468, + 0.589 + ], + "angle": 0, + "content": "Figure 10: About the ablation experiments of the parameter \\(\\beta, \\gamma, \\alpha\\) and \\(k\\), the vertical axis is the PSNR metric." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.602, + 0.241, + 0.617 + ], + "angle": 0, + "content": "6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.468, + 0.744 + ], + "angle": 0, + "content": "To validate the effectiveness of our proposed design, we performed ablation experiments on UDR-S²Former. To this end, we utilized the training dataset of RainDS-Syn [48] for training our model, and evaluated its performance on the RDS scenes of the corresponding testing set. The experimental settings are kept consistent with the aforementioned ones. In the subsequent sections, we analyze the individual contributions of each module toward the performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.745, + 0.468, + 0.761 + ], + "angle": 0, + "content": "6.1. Effectiveness of Image Reconstruction Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.87 + ], + "angle": 0, + "content": "Ablation researches of parameters mentioned in IRM. Our ablation studies involve thorough research of the parameters employed in the method, and we have depicted the results in Fig.10. As demonstrated, the optimal value for each parameter lies within a particular range, which can be attributed to our initial design motivation holding theoretical significance within that range." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Improvements of Sparse Sampling Attention (SSA) in IRM. We conducted a comparison of our sparse sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.416 + ], + "angle": 0, + "content": "attention with sparse attention in ART [71] (SA), window-based self-attention [58] (WSA), and channel-dimension self-attention [68] (CSA). The results are shown in Table 4. It demonstrates that sparse sampling attention performs significantly better than previous designs, underscoring the importance of incorporating degradation information by adaptive sampling from a global perspective for the unified removal of complex rain scenes." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.427, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Table 4: Ablation studies on Table 5: Ablation studies on SSA (§3.1.1). UDR. (§3.1.1 and 3.1.2)." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.455, + 0.695, + 0.513 + ], + "angle": 0, + "content": "
SettingModelPSNRSSIM
iSA [71]36.390.962
iiWSA [58]36.150.959
iiiCSA [68]36.230.960
ivSSA (Ours)36.910.966
" + }, + { + "type": "table", + "bbox": [ + 0.714, + 0.455, + 0.884, + 0.514 + ], + "angle": 0, + "content": "
SettingModelPSNRSSIM
iSSA w/o UD36.250.960
iiSSA w/o RS36.290.961
iiiLR w/o UD36.330.962
ivLR w/o RS36.450.963
vOurs36.910.966
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.666 + ], + "angle": 0, + "content": "Gains of Uncertainty-Driven Ranking (UDR). In this section, we perform ablation studies of our uncertainty driven (UD) and ranking strategies (RS). The results in Table 5 demonstrate that leveraging uncertainty can effectively constrain the process of sparse sampling and facilitate the network to recover locally degraded regions. However, using uncertainty maps directly without employing a ranking strategy leads to a significant drop in performance, due to the lack of using explicit properties of uncertainty maps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.668, + 0.619, + 0.683 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.84 + ], + "angle": 0, + "content": "In this paper, we present a sparse sampling transformer with the uncertainty-driven ranking that removes rain streaks and raindrops in a unified approach. Our approach employs sparse sampling self-attention to effectively capture global degradation relationships in an adaptive manner. We explicitly leverage the uncertainty map via a ranking strategy to constrain the sampling process and facilitate local reconstruction. Our method achieves SOTA results while requiring minimal computational overhead, showcasing its superiority over existing approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), and the National Natural Science Foundation of China (Grant No. 61902275)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13114" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 39(12):2481-2495, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5710-5719, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.313 + ], + "angle": 0, + "content": "[3] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[6] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[7] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Dualformer: Hybrid self-attention transformer for efficient image restoration. arXiv preprint arXiv:2210.01069, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.586 + ], + "angle": 0, + "content": "[8] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[9] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Yi Ye, and Erkang Chen. Msp-former: Multi-scale projection transformer for single image desnowing. arXiv preprint arXiv:2207.05621, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[10] Sixiang Chen, Tian Ye, Jun Shi, Yun Liu, JingXia Jiang, Erkang Chen, and Peng Chen. Dehrformer: Real-time transformer for depth estimation and haze removal from varicolored haze scenes. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[11] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Sen Deng, Mingqiang Wei, Jun Wang, Yidan Feng, Luming Liang, Haoran Xie, Fu Lee Wang, and Meng Wang. Detail-recovery image deraining via context aggregation networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14560-14569, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.6.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[14] David Eigen, Dilip Krishnan, and Rob Fergus. Restoring an image taken through a window covered with dirt or rain. In Proceedings of the IEEE international conference on computer vision, pages 633-640, 2013. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[15] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3855-3863, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[16] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[17] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[18] Shuhang Gu, Deyu Meng, Wangmeng Zuo, and Lei Zhang. Joint convolutional analysis and synthesis sparse representation for single image layer separation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1708-1716, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[19] Ming Hong, Jianzhuang Liu, Cuihua Li, and Yanyun Qu. Uncertainty-driven dehazing network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 906–913, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[20] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In European Conference on Computer Vision, pages 163-180. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[21] Jie Huang, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Hybrid image enhancement with progressive laplacian enhancing unit. In Proceedings of the 27th ACM International Conference on Multimedia, page 1614-1622, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[22] Jie Huang, Feng Zhao, Man Zhou, Jie Xiao, Naishan Zheng, Kaiwen Zheng, and Zhiwei Xiong. Learning sample relationship for exposure correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9904-9913, June 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[23] Jie Huang, Man Zhou, Yajing Liu, Mingde Yao, Feng Zhao, and Zhiwei Xiong. Exposure-consistency representation learning for exposure correction. In Proceedings of the 30th ACM International Conference on Multimedia, page 6309-6317, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Jie Huang, Pengfei Zhu, Mingrui Geng, Jiewen Ran, Xingguang Zhou, Chen Xing, Pengfei Wan, and Xiangyang Ji. Range scaling global u-net for perceptual image enhancement on mobile devices. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, September 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13115" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[25] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800–801, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[26] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[27] Kui Jiang, Zhongyuan Wang, Chen Chen, Zheng Wang, Laizhong Cui, and Chia-Wen Lin. Magic elf: Image deraining meets association learning and transformer. arXiv preprint arXiv:2207.10455, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[28] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[29] Yeying Jin, Ruoteng Li, Wenhan Yang, and Robby T Tan. Estimating reflectance layer from a single image: Integrating reflectance guidance and shadow/specular aware learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 1069-1077, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[30] Yeying Jin, Beibei Lin, Wending Yan, Wei Ye, Yuan Yuan, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[31] Yeying Jin, Aashish Sharma, and Robby T Tan. Deshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.615 + ], + "angle": 0, + "content": "[32] Yeying Jin, Wending Yan, Wenhan Yang, and Robby T Tan. Structure representation network and uncertainty feedback learning for dense non-uniform fog removal. In Proceedings of the Asian Conference on Computer Vision, pages 2041–2058, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[33] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In European Conference on Computer Vision, pages 404-421. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[34] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[35] Guanbin Li, Xiang He, Wei Zhang, Huiyou Chang, Le Dong, and Liang Lin. Non-locally enhanced encoder-decoder network for single image de-raining. In Proceedings of the 26th ACM international conference on Multimedia, pages 1056–1064, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[36] Ruoteng Li, Loong-Fah Cheong, and Robby T Tan. Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1633-1642, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[37] Xia Li, Jianlong Wu, Zhouchen Lin, Hong Liu, and Hongbin Zha. Recurrent squeeze-and-excitation context aggregation net for single image deraining. In Proceedings of the Eu" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ropean conference on computer vision (ECCV), pages 254-269, 2018. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[38] Yu Li, Robby T Tan, Xiaojie Guo, Jiangbo Lu, and Michael S Brown. Rain streak removal using layer priors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2736-2744, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[39] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[40] Yuanchu Liang, Saeed Anwar, and Yang Liu. Drt: A lightweight single image deraining recursive transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 589-598, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[41] Yun Liu, Zhongsheng Yan, Sixiang Chen, Tian Ye, Wenqi Ren, and Erkang Chen. Nighthazeformer: Single nighttime haze removal using prior query transformer. arXiv preprint arXiv:2305.09533, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[43] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[44] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[46] Kuldeep Purohit, Maitreya Suin, AN Rajagopalan, and Vishnu Naresh Boddeti. Spatially-adaptive image restoration using distortion-guided networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2309-2319, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[47] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2482-2491, 2018. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[48] Ruijie Quan, Xin Yu, Yuanzhi Liang, and Yi Yang. Removing raindrops and rain streaks in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9147-9156, 2021. 1, 2, 3, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[49] Yuhui Quan, Shijie Deng, Yixin Chen, and Hui Ji. Deep learning for seeing through window with raindrops. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2463-2471, 2019. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13116" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "better and simpler baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3937-3946, 2019. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[52] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. arXiv preprint arXiv:2204.03883, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[53] Ming Tong, Yongzhen Wang, Peng Cui, Xuefeng Yan, and Mingqiang Wei. Semi-uformer: Semi-supervised uncertainty-aware transformer for image dehazing. arXiv preprint arXiv:2210.16057, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.279, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[54] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.47, + 0.419 + ], + "angle": 0, + "content": "[55] Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson WH Lau. Spatial attentive single-image deraining with a high quality real rain dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12270-12279, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[56] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[57] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[58] Jie Xiao, Xueyang Fu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Image de-raining transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[59] Ke Xu, Xin Tian, Xin Yang, Baocai Yin, and Rynson WH Lau. Intensity-aware single-image deraining with semantic and color regularization. IEEE Transactions on Image Processing, 30:8497-8509, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[60] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[61] Wenhan Yang, Robby T Tan, Jiashi Feng, Jiaying Liu, Zongming Guo, and Shuicheng Yan. Deep joint rain detection and removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1357-1366, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[62] Wenhan Yang, Robby T Tan, Jiashi Feng, Shiqi Wang, Bin Cheng, and Jiaying Liu. Recurrent multi-frame deraining: Combining physics guidance and adversarial learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8569-8586, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[63] Rajeev Yasarla and Vishal M Patel. Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "single image de-raining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8405-8414, 2019. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[64] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[65] Shaodi You, Robby T Tan, Rei Kawakami, and Katsushi Ikeuchi. Adherent raindrop detection and removal in video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1035-1042, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[66] Hu Yu, Jie Huang, Yajing Liu, Qi Zhu, Man Zhou, and Feng Zhao. Source-free domain adaptation for real-world image dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, page 6645-6654, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[67] Hu Yu, Jie Huang, Yajing Liu, Qi Zhu, Man Zhou, and Feng Zhao. Source-free domain adaptation for real-world image dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, page 6645-6654, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.443 + ], + "angle": 0, + "content": "[68] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 3, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.445, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[69] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14821-14831, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.515, + 0.892, + 0.583 + ], + "angle": 0, + "content": "[70] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[71] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. arXiv preprint arXiv:2210.01427, 2022. 3, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[72] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.698, + 0.892, + 0.766 + ], + "angle": 0, + "content": "[73] man zhou, Hu Yu, Jie Huang, Feng Zhao, Jinwei Gu, Chen Change Loy, Deyu Meng, and Chongyi Li. Deep fourier up-sampling. In Advances in Neural Information Processing Systems, volume 35, pages 22995-23008. Curran Associates, Inc., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.768, + 0.892, + 0.837 + ], + "angle": 0, + "content": "[74] Lei Zhu, Zijun Deng, Xiaowei Hu, Haoran Xie, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Learning gated non-local residual for single-image rain streak removal. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2147–2159, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.839, + 0.892, + 0.895 + ], + "angle": 0, + "content": "[75] Lei Zhu, Chi-Wing Fu, Dani Lischinski, and Pheng-Ann Heng. Joint bi-layer optimization for single-image rain streak removal. In Proceedings of the IEEE international conference on computer vision, pages 2526-2534, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13117" + } + ] +] \ No newline at end of file diff --git a/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/32196ea3-316d-4633-8abd-b15ee0703f11_origin.pdf b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/32196ea3-316d-4633-8abd-b15ee0703f11_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..28cbb4dc66443c0d333a116d011271cd468968f7 --- /dev/null +++ b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/32196ea3-316d-4633-8abd-b15ee0703f11_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:538c573b6919cf071a7b39f05774d2d029ff89ba92bc97585456556027f8802c +size 7690930 diff --git a/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/full.md b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7e97b7fa5398836e41569d2c9f38f55fbab2073e --- /dev/null +++ b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/full.md @@ -0,0 +1,491 @@ +# Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks + +Sixiang Chen $^{1,3*}$ Tian Ye $^{1,3*}$ Jinbin Bai $^{2}$ Erkang Chen $^{3}$ Jun Shi $^{4}$ Lei Zhu $^{1,5\dagger}$ + +$^{1}$ The Hong Kong University of Science and Technology (Guangzhou) $^{2}$ National University of Singapore $^{3}$ School of Ocean Information Engineering, Jimei University + +$^{4}$ Xinjiang University $^{5}$ The Hong Kong University of Science and Technology + +{sixiangchen,owentianye}@hkust-gz.edu.cn + +jinbin.bai@u.nus.edu, ekchen@jmu.edu.cn, junshi2022@gmail.com, leizhu@ust.hk + +Project page: https://ephemeral182.github.io/UDR_S2Former_deraining + +# Abstract + +In the real world, image degradations caused by rain often exhibit a combination of rain streaks and raindrops, thereby increasing the challenges of recovering the underlying clean image. Note that the rain streaks and raindrops have diverse shapes, sizes, and locations in the captured image, and thus modeling the correlation relationship between irregular degradations caused by rain artifacts is a necessary prerequisite for image deraining. This paper aims to present an efficient and flexible mechanism to learn and model degradation relationships in a global view, thereby achieving a unified removal of intricate rain scenes. To do so, we propose a Sparse Sampling Transformer based on Uncertainty-Driven Ranking, dubbed UDR-S²Former. Compared to previous methods, our UDR-S²Former has three merits. First, it can adaptively sample relevant image degradation information to model underlying degradation relationships. Second, explicit application of the uncertainty-driven ranking strategy can facilitate the network to attend to degradation features and understand the reconstruction process. Finally, experimental results show that our UDR-S²Former clearly outperforms state-of-the-art methods for all benchmarks. + +# 1. Introduction + +Rain is a ubiquitous condition that negatively impacts various computer vision tasks [2, 70]. In real-world rain scenes, raindrops and rain streaks are irregularly superimposed on clean images. Image deraining is employed to restore the clean images from the complex rain degradations. According to previous work [48], the imaging model + +![](images/410f339b609fb3004addda3fbc86a245d84ee73bd6d09100f23bc0843aad919d.jpg) +(a) Fixed patch partition breaks the relationship modeling within diverse rain degradations + +![](images/bd4a186fc2ed84893650b5c348120a6af474faeafcbed90bd80de75d91102006.jpg) +(b) We utilize the uncertainty-driven sparse sampling transformer to thoroughly explore arbitrary position degradation relationships. + +![](images/14aa14faaa58ff0bbcb5d23a759f6ec79ab539cb41f466e8781982ef3a781831.jpg) +(c) Uncertainty Map +Figure 1: Illustration of the breakdown of complex rain degradation relationships and the thumbnails of our main ideas. Colored places indicate degradations. Two-way arrows represent modeling between degradations. + +of precipitation, inclusive of rain streaks and raindrops, can be expressed as: + +$$ +\mathcal {R} _ {d s} = \left(1 - \mathcal {M} _ {r}\right) \odot (\mathcal {B} + \mathcal {S}) + \eta \mathcal {D}, \tag {1} +$$ + +where $\mathcal{B}$ and $S$ denote the clean background and rain streak map. $\mathcal{M}_r$ is a binary mask used to judge whether the pixel belongs to the raindrops or the background. $\mathcal{D}$ is the raindrops and $\eta$ means global atmospheric lighting coefficient. + +As noted by CCN [48], removing rain streaks and raindrops in a unified manner cannot be achieved by simply combining separate methods for removing either. This is due to the complex nature of the physical models involved and the wide array of possible degradation combinations. Previous models developed to address singular forms of degradations [11, 12, 36, 49, 50, 55, 62] face notable obstacles when dealing with irregularly dispersed and diverse rain degradation types. + +Specifically, current SOTA methods for image deraining primarily concentrate on using ViTs due to their abilities to model long-range dependencies [27, 40, 57, 58]. Among these methods, window-based self-attention [40, 57, 58] has gained popularity due to its computational efficiency. However, as shown in Fig.1, we argue that utilizing window + +![](images/455cfcad4a2f84bf4501ab5bd9f39d3136e940d7c7cf326d729a346ed21a8143.jpg) +Figure 2: The uncertainty maps (bottom row) correspond to both real and synthetic samples (top row), with more significant uncertainty appearing in areas with severe and complicated degradations. This observation motivates us to use uncertainty explicitly to represent knowledge about degradation and to improve the model's understanding of degradation restoration. + +based self-attention mechanisms can lead to incomplete degradation coverage, causing the breakdown of degradation relationships for unified rain degradation removal due to fixed window segmentation. This problem can be particularly pronounced when dealing with large raindrops or rain streaks at long distances simultaneously. However, in the case of complicated degradations, it is imperative to model the relationships between related forms of degradations. + +Furthermore, with respect to the dense prediction task of rain removal, the density, shape, position, and size of raindrops and streaks are all uncertain, rendering it arduous for the network to restore clean images from diverse degradations. The CCN [48] requires using expensive and inflexible NAS to select an optimal architecture that effectively handles rain streaks and raindrops precisely. Inspired by uncertainty modeling [34] for image restoration, incorporating uncertainty learning can enhance the performance by reducing the error in model parameters [19], or serve as a regular term constraint to enhance the prediction quality of regions characterized by high uncertainty [44, 53]. Nevertheless, the mentioned design paradigm overlooks the importance of explicitly excavating uncertainty maps in facilitating the network's modeling of degradation features. For intricate rain scenes, we claim that learning uncertainty estimation can in turn affect the network to better focus on the complicated rain degradation areas. As depicted in Fig.2, the uncertainty map of the image exhibits greater concentration within the degraded region. It is our contention that by fully leveraging the properties of this uncertainty map, we can more effectively model the relationships of degradation and drive the network for the understanding of degradation restoration. + +Specifically, to address the problems above, we first design the Sparse Sampling Attention to deal with complicated rain scenes. It sparsely learns the relevant degradation relationships from the entire image, thereby alleviating the drawback of large-scale degradation modeling in window + +based attention. Concurrently, we leverage the uncertainty map to guide feature learning for capturing more discriminative sampling features. To be exact, in order to fully leverage uncertainty information to promote the sampling of degradation features, we propose a novel ranking strategy and present a Constraint Matrix based on it. Such design further restricts the degree of attention to various rain degradations in the sampling process, boosting the modeling of relationships between degradations. Moreover, when restoring partially degraded regions, we consider the internal difference of the uncertainty map and use the ranking of the Correlation Map to strengthen the network to restore the degraded area by leveraging clean cues within the local regions. + +Overall, our contributions can be summarized as follows: + +- An uncertainty-driven sparse sampling transformer that fully models the global degradation relationships in an efficient manner is proposed to remove diverse rain streaks and raindrops. +- We introduce a ranking strategy in the uncertainty map to enable the model to emphasize various rain degradation features in the sampling process through a constructed constraint matrix. +- To enhance local reconstruction, we utilize the internal discrepancies within the uncertainty map to stimulate the network to extract credibly clean information. + +# 2. Related Works + +# 2.1. Single Image Deraining + +Rain streak removal. Image restoration from adverse weather has made significant advancements over the years, owing to its paramount significance [8, 9, 30, 41, 64, 66, 67]. Recently, the field of single image deraining has been predominantly dominated by learning-based methods [28, 36, 37, 50, 55, 58, 59, 61-63, 74, 75]. Zhu et al. [75] proposed a joint optimization algorithm that involves iteratively removing rain streaks from the background layer and non-rain + +![](images/1246007b4512497ade2642e695184ce9421565feb4cf759b160db120c8e0fb57.jpg) +Figure 3: The overview of our UDR- $S^2$ Former pipeline, which includes (a) our proposed restoration architecture, (b) the Local Reconstruction (LR) in the IRM module, (c) the Sparse Sampling Attention (SSA) constrained by the uncertainty map. For clarity, we only depict one local patch operation. The Feature Extraction and Global Modeling Modules, as well as the simple Refinement Block, are described in supplementary material due to page limitation. + +details from the rain layer. This is achieved through the incorporation of three essential priors. PreNet [50] offered a recurrent layer to leverage the inter-stage dependencies of deep features, thereby constructing the progressive recurrent network. UMRL [63] proposed the uncertainty map to constrain the rain map in rainscapes and incorporate physical models to estimate the final deraining output. IDT [58] presented a transformer system comprising a complementary window-based transformer and spatial transformer, enabling improved capture of short- and long-range dependencies in rainy scenes. + +Raindrop removal. Raindrops are frequently observed in rain scenes, and their diverse shapes and positions present difficulties removing them. Previous attempted to eliminate raindrops utilizing various methods, [14, 47, 49, 65]. The Eigen et al. [14] firstly introduced the learning-based paradigm for raindrop removal. AttenGAN [47] adopted the combination of GAN and attention mechanism to recover the clean image from raindrop degradations. + +Most of the current design paradigms continue to create specialized networks for removing rain streaks or raindrops. A unified removal network design has still to be developed. While CCN [48] was the first to consider joint removal of rain degradation, their approach still requires expensive design such as NAS to address complex degradation characteristics. Other general-purpose networks [58, 68] disregard these two degradation characteristics, leading to high computation and parameters when improving performance. + +# 2.2. Vision Transformers for Image Restoration + +ViTs exhibited superior global modeling capabilities, resulting in impressive performance on low-level vision tasks [3, 7, 10, 22, 52, 54, 68, 71] compared with previous CNNs' paradigm [4, 20, 21, 23, 24, 29, 31, 33, 73]. The window-based designs [39, 40, 52, 57, 58] were widely employed to overcome the computational complexity issue of + +$\mathrm{O}(N^2)$ . Additionally, ART [71] utilized a combination of sparse and dense self-attention to manage computational overhead and achieved SOTA outcomes. Restormer [68] incorporated channel-based self-attention to circumvent the square-level complexity problem. Nevertheless, such designs limit the global receptive field of self-attention in spatial dimension and lack flexibility in dealing with intricate and changing degradations. We propose to use sparse sampling to address the aforementioned limitations by adaptively sampling information from the global field to meet the modeling requirements of the local. + +# 2.3. Uncertainty in Deep Learning + +According to Bayesian theory [34], uncertainty in deep learning can be classified into two types: (i) Aleatoric uncertainty, which refers to the inherent noise in the data. (ii) Epistemic uncertainty, which relates to the uncertainty of model parameters. Incorporating modeling uncertainty into the network can enhance its robustness and performance [1, 2]. In low-level fields, leveraging uncertainty can enable the network to prioritize reducing model prediction error [19] or serving as a loss function [32, 44, 53] to improve the reconstruction quality of areas with high uncertainty. Furthermore, uncertainty can assist in the more precise estimation of critical parameters in physical models [63]. Same as [44, 53], we focus on aleatoric uncertainty in this paper. However, these paradigms do not explicitly utilize the distinctive characteristic of the uncertainty map to restrict the network's acquisition of features. This paper proposes leveraging uncertainty-driven ranking to facilitate the network to represent degradations. + +# 3. Proposed UDR-S²Former Pipeline + +Network architecture. The architecture of the proposed network is illustrated in Fig.3. The network accepts a rain image as input, conducts image processing within the net + +![](images/62df6801affa3d5489d0566e822b8c7aee6b1e3de9f95453b05e9dfc6ae540c1.jpg) + +![](images/81ea36e8e73f8046f0293c4b7232ad8436e6ecff48a74542095e7f952587b009.jpg) +Figure 4: Top row: Constructing a constraint matrix according to the ranking strategy derived from the uncertain map. To simplify the explanation, we only present a one-dimensional ranking strategy. Bottom row: The procedure for utilizing the correlation map and ranking strategy to produce the matrix utilized for modulating the self-attention map. We only present the correlation map and ranking strategy within a patch for easy visualization. + +work, and generates a high-quality restored image as output. Specifically, 1) the degraded image is fed into the Feature Extraction stage to acquire knowledge of features at various scales. This process is accomplished in four stages, each comprising basic convolutional blocks. 2) Upon completion of the feature extraction stage, we employ vanilla transformer to capture deep-level information and ensure the comprehensive utilization of global information1. 3) The Image Reconstruction Module (IRM) is employed to represent degradation relationships in the form of sparse sampling self-attention, which is driven by uncertain learning. Further, it can trigger clean cues excavation to guide the restoration for local reconstruction. Skip connections are utilized in both the feature extraction stage and each stage of the image reconstruction module. + +Preliminary for Uncertainty. For image deraining, we model the difference between the image $\mathcal{B}_{gt}$ and its estimated deraining image $\hat{\mathcal{B}}$ as a Laplace distribution. The motivation for this choice is that the Laplace distribution is better suited for characterizing the edges of details in images than the Gaussian distribution. In addition, the L1 loss refers to the Laplace distribution [72]. The likelihood function can be expressed as follows: + +$$ +p (\mathcal {B} _ {g t}, \sigma ; \mathcal {R} _ {d s}) = \frac {1}{2 \sigma} \exp \left(- \frac {| | \hat {\mathcal {B}} - \mathcal {B} _ {g t} | | _ {1}}{\sigma}\right), \tag {2} +$$ + +where $\mathcal{B}_{gt}$ denotes the mean of this distribution. $\hat{\mathcal{B}}$ is output calculated by network from rain image $\mathcal{R}_{ds}$ . $\sigma$ means the uncertainty (variance) of deraining image $\hat{\mathcal{B}}$ . To simplify + +the calculation, we transform the likelihood function into its log form and maximize it: + +$$ +\underset {\sigma} {\arg \max } \ln p \left(\mathcal {B} _ {g t}, \sigma ; \mathcal {R} _ {d s}\right) = \left(- \frac {\left| \left| \hat {\mathcal {B}} - \mathcal {B} _ {g t} \right| \right| _ {1}}{\sigma}\right) - \ln \sigma . \tag {3} +$$ + +During the learning process, the aforementioned formula is inverted and utilized as a loss function, which is minimized for optimization. Additionally, we also follow [44] to circumvent the issue of training instability resulting from the presence of uncertainties in the form of zero values. Without being limited to a loss function, we aim to employ the ranking strategy to explicitly utilize estimated uncertainty. It enhances the modeling of complicated degradation relationships and drive the network to restore degraded regions. + +# 3.1. Image Reconstruction Module + +For the image reconstruction module of each stage, we can express it as: + +$$ +\operatorname {I R M} (\mathcal {X}, \mathcal {U}) = \left\{\operatorname {S S A} (\mathcal {X}, \mathcal {U}), \operatorname {L R} (\mathcal {X}, \mathcal {U}) \right\}, \tag {4} +$$ + +where $\{\cdot\}$ indicates that these two modules are alternately formed, with $\mathcal{X}$ denoting the image feature, and $\mathcal{U}$ representing the corresponding uncertainty map. + +# 3.1.1 Sparse Sampling Attention Constrained by Uncertainty Map + +The degradation of a large-scale area (raindrops) or long-distance span (rain streaks) can result in the loss of local relationships, particularly when a fixed window division is utilized. Moreover, a window-based design may limit the receptive field to a local area, leading to the loss of global knowledge when dealing with complicated degradations. However, it is crucial to fully exploit the modeling of + +![](images/ea18c18415113c0130d0d2b3e1fcde7fdd12d7ab0cc1c830faa04190a2d857a7.jpg) +Figure 5: Visualization of sparse sampling. The yellow region denotes an $8 \times 8$ window patch. The red pixels are coordinates obtained from global sparse sampling to interact window patch for modeling the corresponding degradation relationship, which is not limited to the local patch (we draw multi-head sampling coordinates). + +degradation relationships for removing diverse rain degradations due to their complicated degraded properties. In this part, we propose an uncertainty-driven sparse sampling approach to learn global coordinates for capturing associated rain degradations adaptively, effectively modeling degradations in diverse regions. + +For a given image feature $\mathcal{X} \in \mathbb{R}^{C \times H \times W}$ , our primary objective is to establish a model that can accurately capture the degradation relationships. For each part of the degradations, the network should be prompted to concurrently consider correlation degradations occurring in other image regions to model it. Specifically, we let the network adaptively learn to match the degraded coordinates of other regions in the whole image, and map them to the same image patch, so as to model the degradation relationship at a low cost. The coordinates are learned as follows: + +$$ +\mathcal {S}, \mathcal {B} = \mathcal {F} \left(\mathcal {X} ^ {\in \mathbb {R} ^ {C \times H \times W}}\right), \tag {5} +$$ + +where $\mathcal{F}(\cdot)$ represents the simple convolution and avgpooling operation to estimate scaling factors $S\in \mathbb{R}^{C\times H\times W\times 2}$ and biases $\mathcal{B}\in \mathbb{R}^{C\times H\times W\times 2}$ (last dimension denotes x-axis and y-axis coordinates). We use them to transform the original coordinates of the feature map by multiplication and addition: + +$$ +\operatorname {C o o r d s} ^ {T} (x, y) = \mathcal {G} \left(\operatorname {C o o r d s} ^ {O} (x, y) \times \mathcal {S} + \mathcal {B}\right), \tag {6} +$$ + +![](images/a39ba2fff3f815cf16deb003d73bb7370b029d38a45d886c13802cb8da037e7d.jpg) + +scaling factors and biases change the sampling location for each patch. $\mathrm{Cords}^T (x,y)$ denotes the coordinates of sparse sampling from the global feature, which is leveraged to the local patch $\mathcal{X}_i^{S\in \mathbb{R}^{C\times \frac{H}{M}\times \frac{H}{M}}}$ ( $i$ denotes the $i$ -th patch of $\mathcal{M}^2$ patches) after carrying out the function $\mathcal{G}$ of torch nn. functional grid sample, as shown in Fig.3 (c). We utilize the information obtained through global sparse sampling to match each degraded patches for corresponding degradation modeling and propose a novel attention mech + +anism, dubbed Sparse Sampling Attention (SSA): + +$$ +\mathrm {S S A} = \mathbf {S o f t m a x} \left(\frac {\mathcal {Q} _ {i} ^ {P} \mathcal {K} _ {i} ^ {S ^ {\mathrm {T}}}}{\sqrt {\mathcal {D}}} + p\right) \mathcal {V} _ {i} ^ {S}, \tag {7} +$$ + +where $\mathcal{Q}_i^P$ denotes the queries projected from original feature of $i$ -th patch. $\mathcal{K}_i^S$ and $\mathcal{V}_i^S$ are obtained from global sampling feature. $\mathcal{D}$ means the dimension number and $p$ is the position embedding like [42]. + +Uncertainty map driven: Drawing inspiration from the obvious representation of deteriorated regions in the uncertainty map, we aim to incorporate the uncertainty map as an explicit constraint to restrict the level of attention towards the degradations in the sparse sampling procedure, to ensure the modeling of the degradation relationships. The operation is presented in Fig.4. + +More specifically, with regard to the uncertain map $\mathcal{U} \in \mathbb{R}^{C \times H \times W}$ , we employ a ranking strategy to construct a constraint matrix $\mathcal{C} \in \mathbb{R}^{C \times H \times W}$ , which serves as a means of constraining the sampling process: + +$$ +\left[ \mathcal {C} ^ {n} \left(\mathcal {U} ^ {n}\right) \right] _ {x y} = \left\{ \begin{array}{l l} 1 & \mathcal {U} _ {x y} ^ {n} \geq \mathcal {U} ^ {n} \text {o f r a n k} \gamma \\ \beta & \text {o t h e r w i s e} \end{array} , \right. \tag {8} +$$ + +where $\beta$ and $\gamma$ are the constraint factor and a threshold value of $n$ -th dimension. Such a constraint matrix contains crucial cues from the uncertainty map with almost no computational overhead, and we utilize it to regularize our modeling of degradation relationships. The Eq.5 is further optimized as: + +$$ +\mathcal {S} ^ {U}, \mathcal {B} ^ {U} = \mathcal {F} \left(\mathcal {X} ^ {\in \mathbb {R} ^ {C \times H \times W}} \times \mathcal {C} _ {x y}\right), \tag {9} +$$ + +wherein $\times$ represents the element-wise multiplication. The $\beta$ of $\mathcal{C}_{xy}$ reduces the influence of irrelevant background areas in multiplication form. $S^U$ and $\mathcal{B}^U$ are leveraged to obtain robust coordinates via Eq.6, which promotes the global sparse sampling while mitigating the interference of the background region. Thereby it facilitates the network to concentrate on degradation relationship modeling. + +Our visualization is depicted in Fig.5. As we expected, the sampled red points adaptively have related degradation situations with the target point (see $8 \times 8$ yellow window). The sparse sampling strategy with uncertainty helps model the relation between local and long-range correlation degradations. Additionally, The most concentration of sampled pixels in a larger area maintains some semantic information and coherence, despite being sampled at a distance. + +Discussion I: The merits of our sparse sampling compared with previous related work for image restoration. Compared to the latest sparse attention [71], which permitted each token to interact with a limited number of tokens, and with a fixed interval size. However, it still involved a manually designed mechanism. In complex rainy + +![](images/980449d752487eeff0080a47024774ac06acb5afdddae884766f108a6bf8e14c.jpg) + +![](images/03079a5130d4d503a7d48a4e061c5916e78dc994518d7cfe18fe769045c509b9.jpg) +Figure 6: Top row: Our motivation about needing to strengthen the network to exploit non-degradation regions. Most information is lost due to the complicated rain degradations in the real world, so degradation-free information is needed to reconstruct the clean area. Bottom row: Visualizations on uncertainty map, correlation map and self-attention map. The attention points of the correlation map and the self-attention map differ. We can utilize the correlation map to direct the network's attention towards areas with significant differences, thus enhancing the restoration process through self-attention. + +scenes, such an approach cannot flexibly capture degradation information from different locations. Our design aims to enable the network to learn the coordinates that can sample degradation information from any position, leading to an enhanced modeling performance of the degradation relationships and network flexibility. + +Discussion II: uncertainty driven compared with prior paradigm and rain mask. Regarding uncertainty, the previous UMRL [63] approach used the uncertainty to optimize the rain streak map based on the physical model of the rain streak. Still, it only focused on optimizing the final map, ignoring the network's learning of the intermediate features. Our method uses uncertainty explicitly to drive the network to provide motivation-driven constraints and optimization for the degradation modeling and local reconstruction processes. Compared to the learned rain mask needed GT [46], i) uncertainty map has a theoretical unsupervised loss for practicing. ii). It outperforms the rain mask in representing the network's focus on intricate rain areas beyond mere location. iii). For large-area raindrops with a complicated physical model, it is difficult to learn a rain mask, but the uncertainty map can solve it well by using the discriminative representation of degradations. + +# 3.1.2 Local Reconstruction with Correlation Ranking + +In rain scenes, information in locally degraded regions is often obscured by rain streaks or raindrops, making it difficult to restore, see Fig.6. To address this issue, it is vital to leverage clean cues in the rain-free area. Building on the discriminative representation of degraded areas and + +other regions in the uncertainty map, we propose using the uncertainty map to generate a correlation map. This map explicitly adjusts the attention map through a ranking approach, as shown in Fig.4 and Fig.3(b), which promotes the network to better leverage clean cues for reconstruction in the form of Query-Key. + +We begin by partitioning uncertainty map $\mathcal{U}^{\in \mathbb{R}^{C\times H\times W}}$ into local patches $\mathcal{U}^{P\in \mathbb{R}^{C\times \frac{H}{M}\times \frac{H}{M}}}$ that correspond to those patches in the feature map. The presence of discriminative disparities within each patch of the uncertainty map motivates us to calculate the correlation map (CR) between them: + +$$ +\mathcal {C R} _ {i} = \mathcal {U} _ {i} ^ {P} \times \mathcal {U} _ {i} ^ {P ^ {\mathrm {T}}}. \tag {10} +$$ + +Upon obtaining the correlation map, our purpose is to leverage the degradation-free region to facilitate the restoration process of the degraded areas during image reconstruction. Significant differences (low correlation) between degradations and background in the uncertainty map can be considered a prompt, promoting the network to use clean cues from the background for local reconstruction. To this end, we modulate the self-attention map by selecting the low-correlation regions in the correlation map: + +$$ +\mathrm {L R} = \text {S o f t m a x} \left(\frac {\mathcal {Q} _ {i} ^ {P} \mathcal {K} _ {i} ^ {P T}}{\sqrt {\mathcal {D}}} \times \underbrace {[ - \alpha \boldsymbol {\Sigma} _ {i} + (1 + \alpha) ] + p} _ {\text {M o d u l a t i o n P r o c e s s i n g}}\right) \mathcal {V} _ {i} ^ {P}, \tag {11} +$$ + +where $\alpha$ is a modulation factor. The $\Sigma$ is acquired via the Top k ranking approach, which can be expressed as follows: + +$$ +\left[ \boldsymbol {\Sigma} _ {i} \left(\mathcal {C R} _ {i}\right) \right] _ {x y} = \left\{ \begin{array}{l l} 1 & \left(\mathcal {C R} _ {i}\right) _ {x y} \in \operatorname {T o p k} (\text {r o w} y) \\ 0 & \text {o t h e r w i s e} \end{array} . \right. \tag {12} +$$ + +In the experiments section, the key parameters of the above used are further studied. About our image reconstruction module, the self-attention FFNs are identical to vanilla vision transformer [13]. + +# 4. Loss Function + +Our UDR- $S^2$ Former network will predict the derained result and an uncertainty map. Hence, the total loss of our network is defined as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \mathcal {L} _ {\text {p s n r}} \left(\mathcal {B} _ {\text {p r e}}, \mathcal {B} _ {\text {g t}}\right) + \lambda_ {2} \mathcal {L} _ {\text {p e r c e p t u a l}} \left(\mathcal {B} _ {\text {p r e}}, \mathcal {B} _ {\text {g t}}\right) \\ + \lambda_ {3} \mathcal {L} _ {U D L} (\mathcal {B} _ {\text {p r e}}, \mathcal {U}), \tag {13} \\ \end{array} +$$ + +where $\mathcal{B}_{pre}$ and $\mathcal{B}_{gt}$ denote the predicted result of image deraining and the corresponding ground truth. $\mathcal{U}$ denotes the predicted uncertainty map. $\mathcal{L}_{psnr}$ and $\mathcal{L}_{perceptual}$ are the PSNR and perceptual losses, see [6,7]. $\mathcal{L}_{UDL}$ denotes the uncertainty loss, see [44, 53] for its definition. $\lambda_{1},\lambda_{2}$ and $\lambda_{3}$ are empirically set to 1, 0.2 and 1. + +Table 1: Quantitative results compared with SOTA methods on the Rain200H [61] and Rain200L [61] datasets. Underline and bold indicate the first and second best results. + +
MethodRain200H [61]Rain200L [61]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
GMM [38]14.710.43028.990.875
JCAS [18]14.870.47130.050.897
DDN [15]26.360.80334.930.958
NLEDN [35]29.510.89138.560.980
RESCAN [37]27.450.82135.080.959
PreNet [50]29.040.89037.120.976
UMRL [63]28.710.88736.430.973
JORDER-E [60]28.580.87636.900.973
MSPFN [28]29.660.89039.480.984
CCN [48]29.990.91438.260.981
MPRNet [69]30.760.90839.890.985
DGUNet [43]30.850.91140.230.986
Uformer [57]30.800.91140.200.986
Restormer [68]31.390.91640.580.987
IDT [58]32.100.93440.740.988
NAFNet [5]30.980.91240.450.987
UDR-S2Former32.59↑0.490.937↑0.00340.96↑0.220.989↑0.001
+ +# 5. Experiments + +# 5.1. Implementaion Details + +Our UDR- $S^2$ Former model utilizes a 5-level encoder-decoder architecture. Within this architecture, the number of channel dimensions increases to $\{16,32,64,128,256\}$ across levels 1 to 5. Additionally, the Feature Extraction stage contains $\{4,6,7,8\}$ convolutional blocks, while the transformer block number in the latent layer is 8, with 16 heads. Regarding reconstruction, we employ $\{3,6,7,8\}$ Image Reconstruction Modules, composed alternately of Sparse Sampling Attention and Local Reconstruction, with $\{1,2,4,8\}$ heads. The window size is fixed at $8\times 8$ . Finally, we set the constraint factor $\beta$ , threshold value $\gamma$ , modulation factor $\alpha$ , and the k value for the ranking strategy to 0.6, 0.8, 0.2, and 0.8, respectively. For the self-attention mechanism in this paper, we all use multi-head self-attention, which is consistent with vanilla ViT [13]. Additionally, during the last image reconstruction module of each stage, we convert the output feature and corresponding uncertainty map into the final image format, and the $\mathcal{L}_{UDL}$ is utilized to supervise the learning process of uncertainty map. + +During the training phase, we employed the Adam optimizer with initial momentum $\beta_{1} = 0.9$ and $\beta_{1} = 0.999$ . We initially set the learning rate to 0.0003 and utilized a cyclic learning rate adjustment strategy, whereby the maximum learning rate is set to 0.00036. We trained our model with a data augmentation strategy that included randomly cropping $256 \times 256$ patches and using horizontal flipping and random image rotation to a fixed angle. Our training process involved $6 \times 10^{5}$ steps. To leverage perceptual loss, we used the first and third layers of VGG19 [51]. Our model is implemented using PyTorch [45] and the RTX 3090 GPU. + +# 5.2. Evaluation Metrics and Datasets + +Evaluation Metrics. In accordance with prior deraining methodologies [48, 58], follow [16, 17], the performance of the model is evaluated using PSNR [25] and SSIM [56]. As suggested in [48, 61], the assessment of PSNR and SSIM is based on the luminance channel, which refers to the Y + +channel of the YCbCr space. + +Rain Streak Datasets. In order to evaluate the effectiveness of our rain streak removal method, we follow the strategy of previous work [48] and select two benchmark datasets, namely Rain200H and Rain200L [61]. Both two datasets have 1800 synthetic images for training and 200 images for testing. + +Raindrop Datasets. We also evaluate our method in a raindrop dataset (i.e., AGAN-Data) collected by Qian et al. [47]. AGAN-Data has 861 images for training and 58 images for testing. + +Raindrops and Rain Streak Datasets. Our paper aims to develop a methodology that effectively addresses the challenges posed by the presence of large-area raindrops and complex rain streaks. To evaluate the proposed approach, we utilize the RainDS benchmark dataset [48], which includes both real-world and synthetic images, dubbed as RainDS-Real and RainDS-Syn, respectively. These datasets contain images of scenes that include rain streaks only (RS), raindrops only (RD), or both(RDS). RainDS-Syn has 3600 image pairs, with 3000 images used for training and the remaining 600 images for testing. Meanwhile, RainDS-Real comprises of 750 images, with 450 images for training and 300 images for evaluation. + +# 5.3. Experimental Evaluation on Benchmarks + +Compared Methods. Regarding removing raindrops and rain streaks in images, we conduct extensive experiments to compare various algorithms that can be used for image rain removal. (i) We compare previous SOTA methods (including GMM [38], JCAS [18], DDN [15], NLEDN [35], RESCAN [37], PreNet [50], UMRL [63], JORDER-E [60], MSPFN [28], CCN [48], IDT [58]). (ii) We also compare our network against universal image restoration methods, including MPRNet [69], DGUNet [43], Uformer [57], Restormer [68], NAFNet [5]). For the specific raindrop removal dataset, we compare against Eigen's model [14], Pix2Pix [26], AttentGAN [47], Quan's network [49], CCN [48], and IDT [58]. In the absence of pre-trained models, we conduct model retraining by utilizing publicly available code and subsequently evaluate the best model performance on test datasets to ensure a fair comparison. + +Quantitative Comparison. Tables 1, 2, and 3 reports the quantitative results of our network and state-of-the-art methods on four benchmark datasets, which are Rain200H, Rain200L, RainDS, and AGAN-Data. As demonstrated in these Tables, our approach delivers superior metric results over compared state-of-the-art (SOTA) methods for single rain streaks or raindrop degradations. Moreover, by incorporating uncertainty to steer the network's attention towards challenging degradations, our method outperforms the IDT algorithm by 0.49dB, 0.22dB and 1.01dB on the Rain200H [61], Rain200L [61] datasets and AGAN-Data [47], while significantly outperforming previ + +![](images/4661b3b9060803570aca2d9b60a59e76e27b1fce2c47568332efdbd67f863268.jpg) +(a) Input + +![](images/e44c919bee67622ed881fccf97ff9eb8250d9a1a91731bdd50aadee962d136c0.jpg) +(b) CCN + +![](images/00d806f974c087b8eb50dd3a85bd604ef500e7c7a5b4381615a2b0d88e183060.jpg) +(c) MPRNet + +![](images/95a0fe0411d66e95ff91de179f0a87375979b33891ac098f48a8c9ccd8c773b1.jpg) +(d) DGUNet + +![](images/91135ad22006f61f0acf3b922ebf4eef6724475bf37937eb8b294df3400e68c4.jpg) + +![](images/ae5c647ed1f41523b7c0f1469335909a858512c724ebfffd47a189a69ff4b79d.jpg) + +![](images/e4200f6a7e050e033df00242605ab53e37e65960de2a16e33fee5920b239c25d.jpg) +(g) NAFNet + +![](images/334e416b92a24680aa163dd7906c8b74ed02ebb9026c7a485398057425833fde.jpg) + +![](images/782a0ec52bcc5fe2109dd0da9d7a05d6200e48a5cd80adce08995dd89a1c88dd.jpg) + +![](images/d80353b8f21b50708cc7cec013affaa53c094d0c7475473a5aea5e4c4f7e51d4.jpg) +(a) Input +Table 2: Quantitative comparisons of various SOTA approaches on the RainDS [48] benchmark. Our proposed method outperforms all other approaches on both synthetic and real-world datasets, including all types of precipitation (i.e., rain streaks (RS), raindrops (RD), and a combination of both (RDS)). The results are indicated in bold and underline, representing the first and second-best performances, respectively. $\uparrow$ means the higher is better. The #GFLOPs are calculated by $256\times 256$ image resolution for a fair comparison. + +![](images/fc044bc659c8cbc7765ac349d9ef8de1139dbca51891cd1d5198ce6dab256596.jpg) +(e) Resto +(b) GT + +![](images/8d53039f29147929bcffca07ed287a6649f954b48e0f5e2139c548dba336ed9a.jpg) +(f) IDT +(c) CCN + +![](images/1d69d6ef4c27e24a5ae353d0017ac570cc4040fbe76342db59757c98afe7ba1a.jpg) +(h) Ours +(d) MPRNet + +![](images/12853acdb0fc7ae992aeab31ee7366da601a7da7e6a4ad30af1482721275739b.jpg) +(i) GT +(e) DGUNet + +![](images/c5c9a77c0c56ca1b6fc07ac89e038b7c8e255d5e86fb94bb81b92dd02fca6a5a.jpg) +(f) Restomer +Figure 7: Visual comparisons of removal for raindrops and rain streaks from RainDS-Syn [48] dataset. Zoom it for a better illustration. + +![](images/3f282a81a3ba8d3e62a7df4e37e8cc4894af73180c443580f8c7d09a9bc92bbf.jpg) +(g) IDT + +![](images/c910c365c43ab37ee2b03cad04cc0eb773118dc70f116d111379ae1c9ff4a6c6.jpg) +(h) NAFNet + +![](images/8c2d70f53dbff353cd86fa32f1b07f4b92f3e396ca890ad0b2aeab8554fef10f.jpg) +(i) Ours + +
MethodVenueRainDS-SynRainDS-Real#Param #GFLOPs
RSRDRDSRSRDRDS
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
GMM [38]CVPR'201626.660.78123.040.79321.500.66923.730.56018.600.55421.350.576--
JCAS [18]ICCV'201726.460.78623.150.81120.910.67124.040.55618.180.55521.220.585--
DDN [15]CVPR'201730.410.86927.920.88526.850.79624.850.68323.120.64222.470.606--
NLEDN [35]ACMMM'201836.240.95834.870.95732.130.91727.020.72324.710.67124.060.650--
RESCAN [37]ECCV'201830.990.88729.900.90727.430.81826.700.68324.230.63723.230.5870.15M32.32G
PreNet [50]CVPR'201936.630.96834.580.96432.210.93426.430.72924.420.67923.570.6490.17M66.58G
UMRL [63]CVPR'201935.760.96233.590.95831.570.92925.890.72623.930.67623.010.6470.98M16.50G
JORDER-E [60]TPAMI'201933.650.92533.510.94430.050.87026.560.71324.340.66223.540.6294.17M273.68G
MSPFN [28]CVPR'202038.610.97536.930.97334.080.94726.450.72724.490.68124.110.65121.00M708.44G
CCN [48]CVPR'202139.170.98137.300.97634.790.95727.460.73725.140.70124.930.6793.75M245.85G
MPRNet [69]CVPR'202140.810.98137.030.97234.990.95627.290.73625.260.70124.960.6813.64M148.55G
DGUNet [43]CVPR'202241.090.98337.560.97535.340.95927.520.73725.330.70224.990.68312.18M199.74G
Uformer [57]CVPR'202240.690.97237.080.96634.990.95426.890.73025.310.70124.830.68620.63M43.86G
Restomer [68]CVPR'2022(Oral)41.420.98038.780.97636.080.96127.390.74225.380.70224.920.68526.10M140.99G
IDT [58]TPAMI'202241.610.98339.090.98036.230.96027.510.74325.670.70624.990.68916.00M61.90G
NAFNet [5]ECCV'202240.390.97237.230.97434.990.95727.490.72925.230.70124.640.66340.60M16.19G
UDR-S2Former (Ours)ICCV'202342.39↑0.780.988↑0.00539.78↑0.690.983↑0.00336.91↑0.680.966↑0.00627.90↑0.390.745↑0.00226.01↑0.340.709↑0.00325.52↑0.530.691↑0.0028.53M21.58G
+ +Table 3: Quantitative results compared with SOTA methods on the AGAN-Data [47]. Red and blue indicate the first and second best results. + +
MethodPSNR†SSIM†
(ICCV'2013)Eigen's model [14]21.310.757
(CVPR2017)Pix2Pix [26]27.200.836
(CVPR'2018)AttenGAN [47]31.590.917
(ICCV'2019)Quan's network [49]31.370.918
(CVPR'2021)CCN [48]31.340.929
(TPAMI'2022)IDT [58]31.630.936
UDR-S2Former32.640.943
+ +ous uncertainty-based [63] methods. Table 2 demonstrates that our approach surpasses all prior designs for jointly removing rain streaks and raindrops. While networks like Restormer [68] and IDT [58] can model long-distance dependencies sufficiently, our method is able to effectively model complex degradation relationships in rain-dominated landscapes (i.e., raindrops and rain streaks both) with uncertainty. Additionally, sparse sampling further facilitates the network's ability to obtain crucial global degradation information at a low cost, see Table 2. Based on the indicators, + +our approach exhibits a significant advantage over the latest designs, including Restormer (36.08PSNR $\rightarrow$ 36.91PSNR) and IDT (36.23PSNR $\rightarrow$ 36.91PSNR), in both real-world and synthetic datasets. Additionally, we also present the comparison of speed and memory cost in the supplementary material to further demonstrate our superiority. + +Visual Comparison. We present the comparisons of the visual effects in Fig.7-9. As depicted in the figures, we observe that in a scene containing rain streaks and raindrops, the scene information is either occluded by a broad range of raindrops or masked by dense rain streaks of varying scales. Other approaches either struggle to remove complex degradations or significantly compromise image details. Our approach leverages degradation relationship modeling to facilitate image restoration, and utilizes the clean information to recover intricate rain degradations. The resulting output exhibits higher fidelity with the more accurate restoration of great details. In real-world scenarios, our method enables the removal of various forms of degradation while recovering fine details, outperforming previous methods. + +![](images/d1650df8a459301e2502c0da25dfdebc1fa05f652c967b11e638ac5f7f8965f6.jpg) +(a) Input + +![](images/2d5f3cf3304e873e3ac7af6f9d440126ad26ef2fbf2f54467209182c99fa0731.jpg) +(b) GT + +![](images/ede9051f35f56847a0bbea5b66658ae1d14a1ce77742a9711e074e1e305b3f3d.jpg) +(c) CCN + +![](images/34e739911121ef46b15cc7ad233e7824b648c540164230269ed7101277082c51.jpg) +(d) MPRNet + +![](images/e7f0be0f7aa886d0a351f82358ff85cc4438a3ceec842c67f3ab901f031c9616.jpg) +(e) DGUNet + +![](images/e9882578ce49b3e8a79f0921759750eb2833f77a4c1b8b5e4a11cc5898f271ca.jpg) +(f) Restormer + +![](images/9e2679e241c3d45687e66636ed8409dea6f041bc3cea437448922b5c7e465e63.jpg) +(g) IDT + +![](images/d5d142ce757652236acf13bff8eb361602200c667da6849a62dcd88954a0228a.jpg) +(h) NAFNet + +![](images/a2d3e2b06ea3d6cc0fac92d3ccaa40f344903a7f5742ff5e3cddc2277f7bf948.jpg) +(i) Ours + +![](images/c1650e8e0558982d5749552c994726198a0ca45bdbe24045703b04ff65275441.jpg) +(a) Input + +![](images/72012fe37ca959d90e275b379fb72b334d9ac1cece175c83e32cce5052e63334.jpg) +(b) CCN + +![](images/6a3585fa80c2518b44f44d16cc0ed26a56d78e26e01ca6de5be32964554c81c4.jpg) +Figure 8: Visual results for real-world sample of RainDS-Real [48] dataset. +(c) DGUNet + +![](images/c8e7ffe1bb5c11471f44c577b1faa5f342fa68a10c47b426b2b8facdf1a514e4.jpg) +(d) Restomer + +![](images/b822c43491e0fe5225eac5c0628d3246ebc7864b362e0cfa590321ecad4576e0.jpg) +(e) IDT +Figure 9: Visual results for real-world sample of RainDS-Real [48] dataset, Please zoom it for a better illustration. + +![](images/82843bbe1e3a8b4031c70e40dab0161769de12f5c4308a4f2ba21c3f371e30e7.jpg) +(f) NAFNet + +![](images/c9f1f32b47ed79de199abc3da64414af6b9851e4228d6380c3b913b4fd095e08.jpg) +(g) Ours + +![](images/e35d81cf6ef259b8b35d04d0fb5d7b8e271b6e5131aeb28759b1e60b7bba55ea.jpg) +(h) GT + +![](images/2d13ca8de42c27a0b28f1f224f5e30ff96401262bfb6d8e72e5240d84d6c2bdd.jpg) +(a) Ablation of $\beta$ and $\gamma$ +Figure 10: About the ablation experiments of the parameter $\beta, \gamma, \alpha$ and $k$ , the vertical axis is the PSNR metric. + +![](images/ecaa1242dbc566f35826dbbf6b7984024361811748464dcfbc8a285a231e6a59.jpg) +(b) Ablation of $\alpha$ and k + +# 6. Ablation Studies + +To validate the effectiveness of our proposed design, we performed ablation experiments on UDR-S²Former. To this end, we utilized the training dataset of RainDS-Syn [48] for training our model, and evaluated its performance on the RDS scenes of the corresponding testing set. The experimental settings are kept consistent with the aforementioned ones. In the subsequent sections, we analyze the individual contributions of each module toward the performance. + +# 6.1. Effectiveness of Image Reconstruction Module + +Ablation researches of parameters mentioned in IRM. Our ablation studies involve thorough research of the parameters employed in the method, and we have depicted the results in Fig.10. As demonstrated, the optimal value for each parameter lies within a particular range, which can be attributed to our initial design motivation holding theoretical significance within that range. + +Improvements of Sparse Sampling Attention (SSA) in IRM. We conducted a comparison of our sparse sampling + +attention with sparse attention in ART [71] (SA), window-based self-attention [58] (WSA), and channel-dimension self-attention [68] (CSA). The results are shown in Table 4. It demonstrates that sparse sampling attention performs significantly better than previous designs, underscoring the importance of incorporating degradation information by adaptive sampling from a global perspective for the unified removal of complex rain scenes. + +Table 4: Ablation studies on Table 5: Ablation studies on SSA (§3.1.1). UDR. (§3.1.1 and 3.1.2). + +
SettingModelPSNRSSIM
iSA [71]36.390.962
iiWSA [58]36.150.959
iiiCSA [68]36.230.960
ivSSA (Ours)36.910.966
+ +
SettingModelPSNRSSIM
iSSA w/o UD36.250.960
iiSSA w/o RS36.290.961
iiiLR w/o UD36.330.962
ivLR w/o RS36.450.963
vOurs36.910.966
+ +Gains of Uncertainty-Driven Ranking (UDR). In this section, we perform ablation studies of our uncertainty driven (UD) and ranking strategies (RS). The results in Table 5 demonstrate that leveraging uncertainty can effectively constrain the process of sparse sampling and facilitate the network to recover locally degraded regions. However, using uncertainty maps directly without employing a ranking strategy leads to a significant drop in performance, due to the lack of using explicit properties of uncertainty maps. + +# 7. Conclusion + +In this paper, we present a sparse sampling transformer with the uncertainty-driven ranking that removes rain streaks and raindrops in a unified approach. Our approach employs sparse sampling self-attention to effectively capture global degradation relationships in an adaptive manner. We explicitly leverage the uncertainty map via a ranking strategy to constrain the sampling process and facilitate local reconstruction. Our method achieves SOTA results while requiring minimal computational overhead, showcasing its superiority over existing approaches. + +Acknowledgment. This work was supported by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), and the National Natural Science Foundation of China (Grant No. 61902275). + +# References + +[1] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 39(12):2481-2495, 2017. 3 +[2] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5710-5719, 2020. 1, 3 +[3] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 3 +[4] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 3 +[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 7, 8 +[6] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 6 +[7] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Dualformer: Hybrid self-attention transformer for efficient image restoration. arXiv preprint arXiv:2210.01069, 2022. 3, 6 +[8] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2 +[9] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Yi Ye, and Erkang Chen. Msp-former: Multi-scale projection transformer for single image desnowing. arXiv preprint arXiv:2207.05621, 2022. 2 +[10] Sixiang Chen, Tian Ye, Jun Shi, Yun Liu, JingXia Jiang, Erkang Chen, and Peng Chen. Dehrformer: Real-time transformer for depth estimation and haze removal from varicolored haze scenes. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3 +[11] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 1 +[12] Sen Deng, Mingqiang Wei, Jun Wang, Yidan Feng, Luming Liang, Haoran Xie, Fu Lee Wang, and Meng Wang. Detail-recovery image deraining via context aggregation networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14560-14569, 2020. 1 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.6.7 +[14] David Eigen, Dilip Krishnan, and Rob Fergus. Restoring an image taken through a window covered with dirt or rain. In Proceedings of the IEEE international conference on computer vision, pages 633-640, 2013. 3, 7, 8 +[15] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3855-3863, 2017. 7, 8 +[16] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 7 +[17] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 7 +[18] Shuhang Gu, Deyu Meng, Wangmeng Zuo, and Lei Zhang. Joint convolutional analysis and synthesis sparse representation for single image layer separation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1708-1716, 2017. 7, 8 +[19] Ming Hong, Jianzhuang Liu, Cuihua Li, and Yanyun Qu. Uncertainty-driven dehazing network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 906–913, 2022. 2, 3 +[20] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In European Conference on Computer Vision, pages 163-180. Springer, 2022. 3 +[21] Jie Huang, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Hybrid image enhancement with progressive laplacian enhancing unit. In Proceedings of the 27th ACM International Conference on Multimedia, page 1614-1622, 2019. 3 +[22] Jie Huang, Feng Zhao, Man Zhou, Jie Xiao, Naishan Zheng, Kaiwen Zheng, and Zhiwei Xiong. Learning sample relationship for exposure correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9904-9913, June 2023. 3 +[23] Jie Huang, Man Zhou, Yajing Liu, Mingde Yao, Feng Zhao, and Zhiwei Xiong. Exposure-consistency representation learning for exposure correction. In Proceedings of the 30th ACM International Conference on Multimedia, page 6309-6317, 2022. 3 +[24] Jie Huang, Pengfei Zhu, Mingrui Geng, Jiewen Ran, Xingguang Zhou, Chen Xing, Pengfei Wan, and Xiangyang Ji. Range scaling global u-net for perceptual image enhancement on mobile devices. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, September 2018. 3 + +[25] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800–801, 2008. 7 +[26] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 7, 8 +[27] Kui Jiang, Zhongyuan Wang, Chen Chen, Zheng Wang, Laizhong Cui, and Chia-Wen Lin. Magic elf: Image deraining meets association learning and transformer. arXiv preprint arXiv:2207.10455, 2022. 1 +[28] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 2, 7, 8 +[29] Yeying Jin, Ruoteng Li, Wenhan Yang, and Robby T Tan. Estimating reflectance layer from a single image: Integrating reflectance guidance and shadow/specular aware learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 1069-1077, 2023. 3 +[30] Yeying Jin, Beibei Lin, Wending Yan, Wei Ye, Yuan Yuan, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution, 2023. 2 +[31] Yeying Jin, Aashish Sharma, and Robby T Tan. Deshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021. 3 +[32] Yeying Jin, Wending Yan, Wenhan Yang, and Robby T Tan. Structure representation network and uncertainty feedback learning for dense non-uniform fog removal. In Proceedings of the Asian Conference on Computer Vision, pages 2041–2058, 2022. 3 +[33] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In European Conference on Computer Vision, pages 404-421. Springer, 2022. 3 +[34] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2, 3 +[35] Guanbin Li, Xiang He, Wei Zhang, Huiyou Chang, Le Dong, and Liang Lin. Non-locally enhanced encoder-decoder network for single image de-raining. In Proceedings of the 26th ACM international conference on Multimedia, pages 1056–1064, 2018. 7, 8 +[36] Ruoteng Li, Loong-Fah Cheong, and Robby T Tan. Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1633-1642, 2019. 1, 2 +[37] Xia Li, Jianlong Wu, Zhouchen Lin, Hong Liu, and Hongbin Zha. Recurrent squeeze-and-excitation context aggregation net for single image deraining. In Proceedings of the Eu + +ropean conference on computer vision (ECCV), pages 254-269, 2018. 2, 7, 8 +[38] Yu Li, Robby T Tan, Xiaojie Guo, Jiangbo Lu, and Michael S Brown. Rain streak removal using layer priors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2736-2744, 2016. 7, 8 +[39] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3 +[40] Yuanchu Liang, Saeed Anwar, and Yang Liu. Drt: A lightweight single image deraining recursive transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 589-598, 2022. 1, 3 +[41] Yun Liu, Zhongsheng Yan, Sixiang Chen, Tian Ye, Wenqi Ren, and Erkang Chen. Nighthazeformer: Single nighttime haze removal using prior query transformer. arXiv preprint arXiv:2305.09533, 2023. 2 +[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5 +[43] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022. 7, 8 +[44] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 2, 3, 4, 6 +[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 7 +[46] Kuldeep Purohit, Maitreya Suin, AN Rajagopalan, and Vishnu Naresh Boddeti. Spatially-adaptive image restoration using distortion-guided networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2309-2319, 2021. 6 +[47] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2482-2491, 2018. 3, 7, 8 +[48] Ruijie Quan, Xin Yu, Yuanzhi Liang, and Yi Yang. Removing raindrops and rain streaks in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9147-9156, 2021. 1, 2, 3, 7, 8, 9 +[49] Yuhui Quan, Shijie Deng, Yixin Chen, and Hui Ji. Deep learning for seeing through window with raindrops. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2463-2471, 2019. 1, 3, 7, 8 +[50] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A + +better and simpler baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3937-3946, 2019. 1, 2, 3, 7, 8 +[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7 +[52] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. arXiv preprint arXiv:2204.03883, 2022. 3 +[53] Ming Tong, Yongzhen Wang, Peng Cui, Xuefeng Yan, and Mingqiang Wei. Semi-uformer: Semi-supervised uncertainty-aware transformer for image dehazing. arXiv preprint arXiv:2210.16057, 2022. 2, 3, 6 +[54] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 3 +[55] Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson WH Lau. Spatial attentive single-image deraining with a high quality real rain dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12270-12279, 2019. 1, 2 +[56] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[57] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 1, 3, 7, 8 +[58] Jie Xiao, Xueyang Fu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Image de-raining transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3, 7, 8, 9 +[59] Ke Xu, Xin Tian, Xin Yang, Baocai Yin, and Rynson WH Lau. Intensity-aware single-image deraining with semantic and color regularization. IEEE Transactions on Image Processing, 30:8497-8509, 2021. 2 +[60] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 7, 8 +[61] Wenhan Yang, Robby T Tan, Jiashi Feng, Jiaying Liu, Zongming Guo, and Shuicheng Yan. Deep joint rain detection and removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1357-1366, 2017. 2, 7 +[62] Wenhan Yang, Robby T Tan, Jiashi Feng, Shiqi Wang, Bin Cheng, and Jiaying Liu. Recurrent multi-frame deraining: Combining physics guidance and adversarial learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8569-8586, 2021. 1, 2 +[63] Rajeev Yasarla and Vishal M Patel. Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for + +single image de-raining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8405-8414, 2019. 2, 3, 6, 7, 8 +[64] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2 +[65] Shaodi You, Robby T Tan, Rei Kawakami, and Katsushi Ikeuchi. Adherent raindrop detection and removal in video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1035-1042, 2013. 3 +[66] Hu Yu, Jie Huang, Yajing Liu, Qi Zhu, Man Zhou, and Feng Zhao. Source-free domain adaptation for real-world image dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, page 6645-6654, 2022. 2 +[67] Hu Yu, Jie Huang, Yajing Liu, Qi Zhu, Man Zhou, and Feng Zhao. Source-free domain adaptation for real-world image dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, page 6645-6654, 2022. 2 +[68] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 3, 7, 8, 9 +[69] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14821-14831, 2021. 7, 8 +[70] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, 2022. 1 +[71] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. arXiv preprint arXiv:2210.01427, 2022. 3, 5, 9 +[72] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 4 +[73] man zhou, Hu Yu, Jie Huang, Feng Zhao, Jinwei Gu, Chen Change Loy, Deyu Meng, and Chongyi Li. Deep fourier up-sampling. In Advances in Neural Information Processing Systems, volume 35, pages 22995-23008. Curran Associates, Inc., 2022. 3 +[74] Lei Zhu, Zijun Deng, Xiaowei Hu, Haoran Xie, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Learning gated non-local residual for single-image rain streak removal. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2147–2159, 2020. 2 +[75] Lei Zhu, Chi-Wing Fu, Dani Lischinski, and Pheng-Ann Heng. Joint bi-layer optimization for single-image rain streak removal. In Proceedings of the IEEE international conference on computer vision, pages 2526-2534, 2017. 2 \ No newline at end of file diff --git a/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/images.zip b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..60e0f78355c895e179a92bd921597b7bf1a277e7 --- /dev/null +++ b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:167e907e97ea03e147989df75764d80a85f634613487a04c1613085edcba6a67 +size 962195 diff --git a/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/layout.json b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..330430b07d0f255bf40b2b5a5899e6e65a31a19a --- /dev/null +++ b/2023/Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks/layout.json @@ -0,0 +1,13019 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 532, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 532, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 532, + 140 + ], + "type": "text", + "content": "Sparse Sampling Transformer with Uncertainty-Driven Ranking for Unified Removal of Raindrops and Rain Streaks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "spans": [ + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "text", + "content": "Sixiang Chen" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "text", + "content": " Tian Ye" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "text", + "content": " Jinbin Bai" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "text", + "content": " Erkang Chen" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "text", + "content": " Jun Shi" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "text", + "content": " Lei Zhu" + }, + { + "bbox": [ + 146, + 159, + 447, + 188 + ], + "type": "inline_equation", + "content": "^{1,5\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 193, + 492, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 193, + 492, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 492, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 193, + 492, + 219 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology (Guangzhou) " + }, + { + "bbox": [ + 105, + 193, + 492, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 193, + 492, + 219 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 105, + 193, + 492, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 105, + 193, + 492, + 219 + ], + "type": "text", + "content": "School of Ocean Information Engineering, Jimei University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 220, + 441, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 220, + 441, + 233 + ], + "spans": [ + { + "bbox": [ + 152, + 220, + 441, + 233 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 152, + 220, + 441, + 233 + ], + "type": "text", + "content": "Xinjiang University " + }, + { + "bbox": [ + 152, + 220, + 441, + 233 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 152, + 220, + 441, + 233 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 206, + 235, + 388, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 235, + 388, + 247 + ], + "spans": [ + { + "bbox": [ + 206, + 235, + 388, + 247 + ], + "type": "text", + "content": "{sixiangchen,owentianye}@hkust-gz.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 248, + 449, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 248, + 449, + 260 + ], + "spans": [ + { + "bbox": [ + 143, + 248, + 449, + 260 + ], + "type": "text", + "content": "jinbin.bai@u.nus.edu, ekchen@jmu.edu.cn, junshi2022@gmail.com, leizhu@ust.hk" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 262, + 465, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 262, + 465, + 274 + ], + "spans": [ + { + "bbox": [ + 127, + 262, + 465, + 274 + ], + "type": "text", + "content": "Project page: https://ephemeral182.github.io/UDR_S2Former_deraining" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 301, + 192, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 301, + 192, + 314 + ], + "spans": [ + { + "bbox": [ + 143, + 301, + 192, + 314 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 327, + 290, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 290, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 290, + 593 + ], + "type": "text", + "content": "In the real world, image degradations caused by rain often exhibit a combination of rain streaks and raindrops, thereby increasing the challenges of recovering the underlying clean image. Note that the rain streaks and raindrops have diverse shapes, sizes, and locations in the captured image, and thus modeling the correlation relationship between irregular degradations caused by rain artifacts is a necessary prerequisite for image deraining. This paper aims to present an efficient and flexible mechanism to learn and model degradation relationships in a global view, thereby achieving a unified removal of intricate rain scenes. To do so, we propose a Sparse Sampling Transformer based on Uncertainty-Driven Ranking, dubbed UDR-S²Former. Compared to previous methods, our UDR-S²Former has three merits. First, it can adaptively sample relevant image degradation information to model underlying degradation relationships. Second, explicit application of the uncertainty-driven ranking strategy can facilitate the network to attend to degradation features and understand the reconstruction process. Finally, experimental results show that our UDR-S²Former clearly outperforms state-of-the-art methods for all benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 614, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 614, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 614, + 287, + 687 + ], + "type": "text", + "content": "Rain is a ubiquitous condition that negatively impacts various computer vision tasks [2, 70]. In real-world rain scenes, raindrops and rain streaks are irregularly superimposed on clean images. Image deraining is employed to restore the clean images from the complex rain degradations. According to previous work [48], the imaging model" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 308, + 301, + 384, + 361 + ], + "blocks": [ + { + "bbox": [ + 308, + 301, + 384, + 361 + ], + "lines": [ + { + "bbox": [ + 308, + 301, + 384, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 301, + 384, + 361 + ], + "type": "image", + "image_path": "410f339b609fb3004addda3fbc86a245d84ee73bd6d09100f23bc0843aad919d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 362, + 378, + 378 + ], + "lines": [ + { + "bbox": [ + 310, + 362, + 378, + 378 + ], + "spans": [ + { + "bbox": [ + 310, + 362, + 378, + 378 + ], + "type": "text", + "content": "(a) Fixed patch partition breaks the relationship modeling within diverse rain degradations" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 386, + 300, + 464, + 361 + ], + "blocks": [ + { + "bbox": [ + 386, + 300, + 464, + 361 + ], + "lines": [ + { + "bbox": [ + 386, + 300, + 464, + 361 + ], + "spans": [ + { + "bbox": [ + 386, + 300, + 464, + 361 + ], + "type": "image", + "image_path": "bd4a186fc2ed84893650b5c348120a6af474faeafcbed90bd80de75d91102006.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 362, + 468, + 378 + ], + "lines": [ + { + "bbox": [ + 384, + 362, + 468, + 378 + ], + "spans": [ + { + "bbox": [ + 384, + 362, + 468, + 378 + ], + "type": "text", + "content": "(b) We utilize the uncertainty-driven sparse sampling transformer to thoroughly explore arbitrary position degradation relationships." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 473, + 301, + 550, + 361 + ], + "blocks": [ + { + "bbox": [ + 473, + 301, + 550, + 361 + ], + "lines": [ + { + "bbox": [ + 473, + 301, + 550, + 361 + ], + "spans": [ + { + "bbox": [ + 473, + 301, + 550, + 361 + ], + "type": "image", + "image_path": "14aa14faaa58ff0bbcb5d23a759f6ec79ab539cb41f466e8781982ef3a781831.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 363, + 528, + 369 + ], + "lines": [ + { + "bbox": [ + 490, + 363, + 528, + 369 + ], + "spans": [ + { + "bbox": [ + 490, + 363, + 528, + 369 + ], + "type": "text", + "content": "(c) Uncertainty Map" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 380, + 546, + 425 + ], + "lines": [ + { + "bbox": [ + 306, + 380, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 546, + 425 + ], + "type": "text", + "content": "Figure 1: Illustration of the breakdown of complex rain degradation relationships and the thumbnails of our main ideas. Colored places indicate degradations. Two-way arrows represent modeling between degradations." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 429, + 546, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 429, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 546, + 453 + ], + "type": "text", + "content": "of precipitation, inclusive of rain streaks and raindrops, can be expressed as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 351, + 466, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 466, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 351, + 466, + 545, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {d s} = \\left(1 - \\mathcal {M} _ {r}\\right) \\odot (\\mathcal {B} + \\mathcal {S}) + \\eta \\mathcal {D}, \\tag {1}", + "image_path": "9ca2934cb552d40f6ebfeaac45b391d24ac4d27aab29d9e78fab998bc1370a4c.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "text", + "content": " denote the clean background and rain streak map. " + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_r" + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "text", + "content": " is a binary mask used to judge whether the pixel belongs to the raindrops or the background. " + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "text", + "content": " is the raindrops and " + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 485, + 546, + 533 + ], + "type": "text", + "content": " means global atmospheric lighting coefficient." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 533, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 641 + ], + "type": "text", + "content": "As noted by CCN [48], removing rain streaks and raindrops in a unified manner cannot be achieved by simply combining separate methods for removing either. This is due to the complex nature of the physical models involved and the wide array of possible degradation combinations. Previous models developed to address singular forms of degradations [11, 12, 36, 49, 50, 55, 62] face notable obstacles when dealing with irregularly dispersed and diverse rain degradation types." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Specifically, current SOTA methods for image deraining primarily concentrate on using ViTs due to their abilities to model long-range dependencies [27, 40, 57, 58]. Among these methods, window-based self-attention [40, 57, 58] has gained popularity due to its computational efficiency. However, as shown in Fig.1, we argue that utilizing window" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 130, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 130, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 130, + 703 + ], + "type": "text", + "content": "*Equal contributions." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 58, + 703, + 234, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 234, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 234, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 58, + 703, + 234, + 713 + ], + "type": "text", + "content": " Lei Zhu (leizhu@ust.hk) is the corresponding author." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13106" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 536, + 222 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 536, + 222 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 536, + 222 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 536, + 222 + ], + "type": "image", + "image_path": "455cfcad4a2f84bf4501ab5bd9f39d3136e940d7c7cf326d729a346ed21a8143.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 547, + 257 + ], + "type": "text", + "content": "Figure 2: The uncertainty maps (bottom row) correspond to both real and synthetic samples (top row), with more significant uncertainty appearing in areas with severe and complicated degradations. This observation motivates us to use uncertainty explicitly to represent knowledge about degradation and to improve the model's understanding of degradation restoration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 288, + 357 + ], + "type": "text", + "content": "based self-attention mechanisms can lead to incomplete degradation coverage, causing the breakdown of degradation relationships for unified rain degradation removal due to fixed window segmentation. This problem can be particularly pronounced when dealing with large raindrops or rain streaks at long distances simultaneously. However, in the case of complicated degradations, it is imperative to model the relationships between related forms of degradations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 288, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 648 + ], + "type": "text", + "content": "Furthermore, with respect to the dense prediction task of rain removal, the density, shape, position, and size of raindrops and streaks are all uncertain, rendering it arduous for the network to restore clean images from diverse degradations. The CCN [48] requires using expensive and inflexible NAS to select an optimal architecture that effectively handles rain streaks and raindrops precisely. Inspired by uncertainty modeling [34] for image restoration, incorporating uncertainty learning can enhance the performance by reducing the error in model parameters [19], or serve as a regular term constraint to enhance the prediction quality of regions characterized by high uncertainty [44, 53]. Nevertheless, the mentioned design paradigm overlooks the importance of explicitly excavating uncertainty maps in facilitating the network's modeling of degradation features. For intricate rain scenes, we claim that learning uncertainty estimation can in turn affect the network to better focus on the complicated rain degradation areas. As depicted in Fig.2, the uncertainty map of the image exhibits greater concentration within the degraded region. It is our contention that by fully leveraging the properties of this uncertainty map, we can more effectively model the relationships of degradation and drive the network for the understanding of degradation restoration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Specifically, to address the problems above, we first design the Sparse Sampling Attention to deal with complicated rain scenes. It sparsely learns the relevant degradation relationships from the entire image, thereby alleviating the drawback of large-scale degradation modeling in window" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 262, + 547, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 429 + ], + "type": "text", + "content": "based attention. Concurrently, we leverage the uncertainty map to guide feature learning for capturing more discriminative sampling features. To be exact, in order to fully leverage uncertainty information to promote the sampling of degradation features, we propose a novel ranking strategy and present a Constraint Matrix based on it. Such design further restricts the degree of attention to various rain degradations in the sampling process, boosting the modeling of relationships between degradations. Moreover, when restoring partially degraded regions, we consider the internal difference of the uncertainty map and use the ranking of the Correlation Map to strengthen the network to restore the degraded area by leveraging clean cues within the local regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 429, + 544, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 429, + 544, + 440 + ], + "spans": [ + { + "bbox": [ + 317, + 429, + 544, + 440 + ], + "type": "text", + "content": "Overall, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 443, + 545, + 578 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 443, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 443, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 317, + 443, + 545, + 490 + ], + "type": "text", + "content": "- An uncertainty-driven sparse sampling transformer that fully models the global degradation relationships in an efficient manner is proposed to remove diverse rain streaks and raindrops." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 493, + 544, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 544, + 539 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 544, + 539 + ], + "type": "text", + "content": "- We introduce a ranking strategy in the uncertainty map to enable the model to emphasize various rain degradation features in the sampling process through a constructed constraint matrix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 543, + 544, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 543, + 544, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 543, + 544, + 578 + ], + "type": "text", + "content": "- To enhance local reconstruction, we utilize the internal discrepancies within the uncertainty map to stimulate the network to extract credibly clean information." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 585, + 397, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 397, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 397, + 597 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 602, + 440, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 602, + 440, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 602, + 440, + 615 + ], + "type": "text", + "content": "2.1. Single Image Deraining" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Rain streak removal. Image restoration from adverse weather has made significant advancements over the years, owing to its paramount significance [8, 9, 30, 41, 64, 66, 67]. Recently, the field of single image deraining has been predominantly dominated by learning-based methods [28, 36, 37, 50, 55, 58, 59, 61-63, 74, 75]. Zhu et al. [75] proposed a joint optimization algorithm that involves iteratively removing rain streaks from the background layer and non-rain" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13107" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 532, + 236 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 532, + 236 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 532, + 236 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 532, + 236 + ], + "type": "image", + "image_path": "1246007b4512497ade2642e695184ce9421565feb4cf759b160db120c8e0fb57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 283 + ], + "type": "text", + "content": "Figure 3: The overview of our UDR-" + }, + { + "bbox": [ + 46, + 239, + 547, + 283 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 46, + 239, + 547, + 283 + ], + "type": "text", + "content": "Former pipeline, which includes (a) our proposed restoration architecture, (b) the Local Reconstruction (LR) in the IRM module, (c) the Sparse Sampling Attention (SSA) constrained by the uncertainty map. For clarity, we only depict one local patch operation. The Feature Extraction and Global Modeling Modules, as well as the simple Refinement Block, are described in supplementary material due to page limitation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 287, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 289, + 418 + ], + "type": "text", + "content": "details from the rain layer. This is achieved through the incorporation of three essential priors. PreNet [50] offered a recurrent layer to leverage the inter-stage dependencies of deep features, thereby constructing the progressive recurrent network. UMRL [63] proposed the uncertainty map to constrain the rain map in rainscapes and incorporate physical models to estimate the final deraining output. IDT [58] presented a transformer system comprising a complementary window-based transformer and spatial transformer, enabling improved capture of short- and long-range dependencies in rainy scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 514 + ], + "type": "text", + "content": "Raindrop removal. Raindrops are frequently observed in rain scenes, and their diverse shapes and positions present difficulties removing them. Previous attempted to eliminate raindrops utilizing various methods, [14, 47, 49, 65]. The Eigen et al. [14] firstly introduced the learning-based paradigm for raindrop removal. AttenGAN [47] adopted the combination of GAN and attention mechanism to recover the clean image from raindrop degradations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 515, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 288, + 624 + ], + "type": "text", + "content": "Most of the current design paradigms continue to create specialized networks for removing rain streaks or raindrops. A unified removal network design has still to be developed. While CCN [48] was the first to consider joint removal of rain degradation, their approach still requires expensive design such as NAS to address complex degradation characteristics. Other general-purpose networks [58, 68] disregard these two degradation characteristics, leading to high computation and parameters when improving performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 625, + 272, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 272, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 272, + 639 + ], + "type": "text", + "content": "2.2. Vision Transformers for Image Restoration" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "ViTs exhibited superior global modeling capabilities, resulting in impressive performance on low-level vision tasks [3, 7, 10, 22, 52, 54, 68, 71] compared with previous CNNs' paradigm [4, 20, 21, 23, 24, 29, 31, 33, 73]. The window-based designs [39, 40, 52, 57, 58] were widely employed to overcome the computational complexity issue of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 286, + 547, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 547, + 419 + ], + "type": "inline_equation", + "content": "\\mathrm{O}(N^2)" + }, + { + "bbox": [ + 304, + 286, + 547, + 419 + ], + "type": "text", + "content": ". Additionally, ART [71] utilized a combination of sparse and dense self-attention to manage computational overhead and achieved SOTA outcomes. Restormer [68] incorporated channel-based self-attention to circumvent the square-level complexity problem. Nevertheless, such designs limit the global receptive field of self-attention in spatial dimension and lack flexibility in dealing with intricate and changing degradations. We propose to use sparse sampling to address the aforementioned limitations by adaptively sampling information from the global field to meet the modeling requirements of the local." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 423, + 470, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 470, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 470, + 436 + ], + "type": "text", + "content": "2.3. Uncertainty in Deep Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 439, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 439, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 545, + 655 + ], + "type": "text", + "content": "According to Bayesian theory [34], uncertainty in deep learning can be classified into two types: (i) Aleatoric uncertainty, which refers to the inherent noise in the data. (ii) Epistemic uncertainty, which relates to the uncertainty of model parameters. Incorporating modeling uncertainty into the network can enhance its robustness and performance [1, 2]. In low-level fields, leveraging uncertainty can enable the network to prioritize reducing model prediction error [19] or serving as a loss function [32, 44, 53] to improve the reconstruction quality of areas with high uncertainty. Furthermore, uncertainty can assist in the more precise estimation of critical parameters in physical models [63]. Same as [44, 53], we focus on aleatoric uncertainty in this paper. However, these paradigms do not explicitly utilize the distinctive characteristic of the uncertainty map to restrict the network's acquisition of features. This paper proposes leveraging uncertainty-driven ranking to facilitate the network to represent degradations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 659, + 492, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 492, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 492, + 673 + ], + "type": "text", + "content": "3. Proposed UDR-S²Former Pipeline" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "Network architecture. The architecture of the proposed network is illustrated in Fig.3. The network accepts a rain image as input, conducts image processing within the net" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13108" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 68, + 534, + 167 + ], + "blocks": [ + { + "bbox": [ + 107, + 68, + 534, + 167 + ], + "lines": [ + { + "bbox": [ + 107, + 68, + 534, + 167 + ], + "spans": [ + { + "bbox": [ + 107, + 68, + 534, + 167 + ], + "type": "image", + "image_path": "62df6801affa3d5489d0566e822b8c7aee6b1e3de9f95453b05e9dfc6ae540c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 169, + 536, + 276 + ], + "blocks": [ + { + "bbox": [ + 57, + 169, + 536, + 276 + ], + "lines": [ + { + "bbox": [ + 57, + 169, + 536, + 276 + ], + "spans": [ + { + "bbox": [ + 57, + 169, + 536, + 276 + ], + "type": "image", + "image_path": "81ea36e8e73f8046f0293c4b7232ad8436e6ecff48a74542095e7f952587b009.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 275, + 547, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 547, + 321 + ], + "type": "text", + "content": "Figure 4: Top row: Constructing a constraint matrix according to the ranking strategy derived from the uncertain map. To simplify the explanation, we only present a one-dimensional ranking strategy. Bottom row: The procedure for utilizing the correlation map and ranking strategy to produce the matrix utilized for modulating the self-attention map. We only present the correlation map and ranking strategy within a patch for easy visualization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 324, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 289, + 502 + ], + "type": "text", + "content": "work, and generates a high-quality restored image as output. Specifically, 1) the degraded image is fed into the Feature Extraction stage to acquire knowledge of features at various scales. This process is accomplished in four stages, each comprising basic convolutional blocks. 2) Upon completion of the feature extraction stage, we employ vanilla transformer to capture deep-level information and ensure the comprehensive utilization of global information1. 3) The Image Reconstruction Module (IRM) is employed to represent degradation relationships in the form of sparse sampling self-attention, which is driven by uncertain learning. Further, it can trigger clean cues excavation to guide the restoration for local reconstruction. Skip connections are utilized in both the feature extraction stage and each stage of the image reconstruction module." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 503, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 289, + 601 + ], + "type": "text", + "content": "Preliminary for Uncertainty. For image deraining, we model the difference between the image " + }, + { + "bbox": [ + 46, + 503, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{gt}" + }, + { + "bbox": [ + 46, + 503, + 289, + 601 + ], + "type": "text", + "content": " and its estimated deraining image " + }, + { + "bbox": [ + 46, + 503, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}" + }, + { + "bbox": [ + 46, + 503, + 289, + 601 + ], + "type": "text", + "content": " as a Laplace distribution. The motivation for this choice is that the Laplace distribution is better suited for characterizing the edges of details in images than the Gaussian distribution. In addition, the L1 loss refers to the Laplace distribution [72]. The likelihood function can be expressed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 608, + 287, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 608, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 73, + 608, + 287, + 641 + ], + "type": "interline_equation", + "content": "p (\\mathcal {B} _ {g t}, \\sigma ; \\mathcal {R} _ {d s}) = \\frac {1}{2 \\sigma} \\exp \\left(- \\frac {| | \\hat {\\mathcal {B}} - \\mathcal {B} _ {g t} | | _ {1}}{\\sigma}\\right), \\tag {2}", + "image_path": "6d946958e3d182152ed8e92d97622d00befde408e379b8b3607be5647ff20a93.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{gt}" + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "text", + "content": " denotes the mean of this distribution. " + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}" + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "text", + "content": " is output calculated by network from rain image " + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{ds}" + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "text", + "content": " means the uncertainty (variance) of deraining image " + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}" + }, + { + "bbox": [ + 47, + 649, + 288, + 687 + ], + "type": "text", + "content": ". To simplify" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 324, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 348 + ], + "type": "text", + "content": "the calculation, we transform the likelihood function into its log form and maximize it:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 355, + 545, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 355, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 310, + 355, + 545, + 381 + ], + "type": "interline_equation", + "content": "\\underset {\\sigma} {\\arg \\max } \\ln p \\left(\\mathcal {B} _ {g t}, \\sigma ; \\mathcal {R} _ {d s}\\right) = \\left(- \\frac {\\left| \\left| \\hat {\\mathcal {B}} - \\mathcal {B} _ {g t} \\right| \\right| _ {1}}{\\sigma}\\right) - \\ln \\sigma . \\tag {3}", + "image_path": "bba7de1da7705330e28aa1c875b59fa7625f97e2409784b77a729ca535edb10d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 496 + ], + "type": "text", + "content": "During the learning process, the aforementioned formula is inverted and utilized as a loss function, which is minimized for optimization. Additionally, we also follow [44] to circumvent the issue of training instability resulting from the presence of uncertainties in the form of zero values. Without being limited to a loss function, we aim to employ the ranking strategy to explicitly utilize estimated uncertainty. It enhances the modeling of complicated degradation relationships and drive the network to restore degraded regions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 498, + 471, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 471, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 471, + 510 + ], + "type": "text", + "content": "3.1. Image Reconstruction Module" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 514, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 545, + 537 + ], + "type": "text", + "content": "For the image reconstruction module of each stage, we can express it as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 340, + 545, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 545, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 340, + 545, + 545, + 559 + ], + "type": "interline_equation", + "content": "\\operatorname {I R M} (\\mathcal {X}, \\mathcal {U}) = \\left\\{\\operatorname {S S A} (\\mathcal {X}, \\mathcal {U}), \\operatorname {L R} (\\mathcal {X}, \\mathcal {U}) \\right\\}, \\tag {4}", + "image_path": "31a9465488442c9367f148cad67b7abeb9d2245c31cfb7ddff182e5d24608794.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\{\\cdot\\}" + }, + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "text", + "content": " indicates that these two modules are alternately formed, with " + }, + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "text", + "content": " denoting the image feature, and " + }, + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 566, + 545, + 602 + ], + "type": "text", + "content": " representing the corresponding uncertainty map." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 604, + 545, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 604, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 604, + 545, + 628 + ], + "type": "text", + "content": "3.1.1 Sparse Sampling Attention Constrained by Uncertainty Map" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "content": "The degradation of a large-scale area (raindrops) or long-distance span (rain streaks) can result in the loss of local relationships, particularly when a fixed window division is utilized. Moreover, a window-based design may limit the receptive field to a local area, leading to the loss of global knowledge when dealing with complicated degradations. However, it is crucial to fully exploit the modeling of" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 288, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 288, + 715 + ], + "type": "text", + "content": "1Restricted by the number of pages, we introduce the points 1) and 2) of our architecture in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13109" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 290, + 177 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 290, + 177 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 290, + 177 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 290, + 177 + ], + "type": "image", + "image_path": "ea18c18415113c0130d0d2b3e1fcde7fdd12d7ab0cc1c830faa04190a2d857a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 287, + 244 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 287, + 244 + ], + "type": "text", + "content": "Figure 5: Visualization of sparse sampling. The yellow region denotes an " + }, + { + "bbox": [ + 46, + 178, + 287, + 244 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 46, + 178, + 287, + 244 + ], + "type": "text", + "content": " window patch. The red pixels are coordinates obtained from global sparse sampling to interact window patch for modeling the corresponding degradation relationship, which is not limited to the local patch (we draw multi-head sampling coordinates)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 249, + 287, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 287, + 320 + ], + "type": "text", + "content": "degradation relationships for removing diverse rain degradations due to their complicated degraded properties. In this part, we propose an uncertainty-driven sparse sampling approach to learn global coordinates for capturing associated rain degradations adaptively, effectively modeling degradations in diverse regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 320, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 287, + 441 + ], + "type": "text", + "content": "For a given image feature " + }, + { + "bbox": [ + 46, + 320, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 320, + 287, + 441 + ], + "type": "text", + "content": ", our primary objective is to establish a model that can accurately capture the degradation relationships. For each part of the degradations, the network should be prompted to concurrently consider correlation degradations occurring in other image regions to model it. Specifically, we let the network adaptively learn to match the degraded coordinates of other regions in the whole image, and map them to the same image patch, so as to model the degradation relationship at a low cost. The coordinates are learned as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 448, + 287, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 448, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 287, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {S}, \\mathcal {B} = \\mathcal {F} \\left(\\mathcal {X} ^ {\\in \\mathbb {R} ^ {C \\times H \\times W}}\\right), \\tag {5}", + "image_path": "8c10819c118b407ac79639e871b021c7d3b9b453cde59e0f9c8204f04b78a813.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 473, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 473, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 46, + 473, + 287, + 546 + ], + "type": "text", + "content": " represents the simple convolution and avgpooling operation to estimate scaling factors " + }, + { + "bbox": [ + 46, + 473, + 287, + 546 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{R}^{C\\times H\\times W\\times 2}" + }, + { + "bbox": [ + 46, + 473, + 287, + 546 + ], + "type": "text", + "content": " and biases " + }, + { + "bbox": [ + 46, + 473, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{B}\\in \\mathbb{R}^{C\\times H\\times W\\times 2}" + }, + { + "bbox": [ + 46, + 473, + 287, + 546 + ], + "type": "text", + "content": " (last dimension denotes x-axis and y-axis coordinates). We use them to transform the original coordinates of the feature map by multiplication and addition:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 558, + 287, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 558, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 70, + 558, + 287, + 573 + ], + "type": "interline_equation", + "content": "\\operatorname {C o o r d s} ^ {T} (x, y) = \\mathcal {G} \\left(\\operatorname {C o o r d s} ^ {O} (x, y) \\times \\mathcal {S} + \\mathcal {B}\\right), \\tag {6}", + "image_path": "d9afad7d153aef6ca9367a2b919f5394e017c0e7cf258953bf66ef308a2dea56.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 153, + 575, + 257, + 591 + ], + "blocks": [ + { + "bbox": [ + 153, + 575, + 257, + 591 + ], + "lines": [ + { + "bbox": [ + 153, + 575, + 257, + 591 + ], + "spans": [ + { + "bbox": [ + 153, + 575, + 257, + 591 + ], + "type": "image", + "image_path": "a39ba2fff3f815cf16deb003d73bb7370b029d38a45d886c13802cb8da037e7d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "text", + "content": "scaling factors and biases change the sampling location for each patch. " + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Cords}^T (x,y)" + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "text", + "content": " denotes the coordinates of sparse sampling from the global feature, which is leveraged to the local patch " + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i^{S\\in \\mathbb{R}^{C\\times \\frac{H}{M}\\times \\frac{H}{M}}}" + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "text", + "content": "-th patch of " + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^2" + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "text", + "content": " patches) after carrying out the function " + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 600, + 287, + 713 + ], + "type": "text", + "content": " of torch nn. functional grid sample, as shown in Fig.3 (c). We utilize the information obtained through global sparse sampling to match each degraded patches for corresponding degradation modeling and propose a novel attention mech" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 507, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 507, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 507, + 84 + ], + "type": "text", + "content": "anism, dubbed Sparse Sampling Attention (SSA):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 347, + 93, + 545, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 93, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 347, + 93, + 545, + 126 + ], + "type": "interline_equation", + "content": "\\mathrm {S S A} = \\mathbf {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {i} ^ {P} \\mathcal {K} _ {i} ^ {S ^ {\\mathrm {T}}}}{\\sqrt {\\mathcal {D}}} + p\\right) \\mathcal {V} _ {i} ^ {S}, \\tag {7}", + "image_path": "2c3c44595e46b5291846db4e26dd3e5bb874c12fcbc377e7d2015f8846c2625b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_i^P" + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": " denotes the queries projected from original feature of " + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": "-th patch. " + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_i^S" + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i^S" + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": " are obtained from global sampling feature. " + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": " means the dimension number and " + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 135, + 545, + 184 + ], + "type": "text", + "content": " is the position embedding like [42]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 184, + 545, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 545, + 268 + ], + "type": "text", + "content": "Uncertainty map driven: Drawing inspiration from the obvious representation of deteriorated regions in the uncertainty map, we aim to incorporate the uncertainty map as an explicit constraint to restrict the level of attention towards the degradations in the sparse sampling procedure, to ensure the modeling of the degradation relationships. The operation is presented in Fig.4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 269, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 319 + ], + "type": "text", + "content": "More specifically, with regard to the uncertain map " + }, + { + "bbox": [ + 304, + 269, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{U} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 269, + 545, + 319 + ], + "type": "text", + "content": ", we employ a ranking strategy to construct a constraint matrix " + }, + { + "bbox": [ + 304, + 269, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 269, + 545, + 319 + ], + "type": "text", + "content": ", which serves as a means of constraining the sampling process:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 328, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 328, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 339, + 328, + 545, + 360 + ], + "type": "interline_equation", + "content": "\\left[ \\mathcal {C} ^ {n} \\left(\\mathcal {U} ^ {n}\\right) \\right] _ {x y} = \\left\\{ \\begin{array}{l l} 1 & \\mathcal {U} _ {x y} ^ {n} \\geq \\mathcal {U} ^ {n} \\text {o f r a n k} \\gamma \\\\ \\beta & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}", + "image_path": "1a0313fff78fbe57b5425500b24ed895671a789a2262f353d47bd5675f6e4f7c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 369, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 369, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 369, + 545, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 369, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 369, + 545, + 438 + ], + "type": "text", + "content": " are the constraint factor and a threshold value of " + }, + { + "bbox": [ + 304, + 369, + 545, + 438 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 369, + 545, + 438 + ], + "type": "text", + "content": "-th dimension. Such a constraint matrix contains crucial cues from the uncertainty map with almost no computational overhead, and we utilize it to regularize our modeling of degradation relationships. The Eq.5 is further optimized as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 353, + 439, + 545, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 439, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 353, + 439, + 545, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {S} ^ {U}, \\mathcal {B} ^ {U} = \\mathcal {F} \\left(\\mathcal {X} ^ {\\in \\mathbb {R} ^ {C \\times H \\times W}} \\times \\mathcal {C} _ {x y}\\right), \\tag {9}", + "image_path": "cebb5c98a7cdc0b1e6a239e6cfa7ac952ce8f377008225e69d196997e2167d56.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "text", + "content": "wherein " + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "text", + "content": " represents the element-wise multiplication. The " + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{xy}" + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "text", + "content": " reduces the influence of irrelevant background areas in multiplication form. " + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "inline_equation", + "content": "S^U" + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^U" + }, + { + "bbox": [ + 304, + 461, + 545, + 544 + ], + "type": "text", + "content": " are leveraged to obtain robust coordinates via Eq.6, which promotes the global sparse sampling while mitigating the interference of the background region. Thereby it facilitates the network to concentrate on degradation relationship modeling." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "type": "text", + "content": "Our visualization is depicted in Fig.5. As we expected, the sampled red points adaptively have related degradation situations with the target point (see " + }, + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 545, + 545, + 640 + ], + "type": "text", + "content": " yellow window). The sparse sampling strategy with uncertainty helps model the relation between local and long-range correlation degradations. Additionally, The most concentration of sampled pixels in a larger area maintains some semantic information and coherence, despite being sampled at a distance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Discussion I: The merits of our sparse sampling compared with previous related work for image restoration. Compared to the latest sparse attention [71], which permitted each token to interact with a limited number of tokens, and with a fixed interval size. However, it still involved a manually designed mechanism. In complex rainy" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13110" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 291, + 131 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 291, + 131 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 291, + 131 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 291, + 131 + ], + "type": "image", + "image_path": "980449d752487eeff0080a47024774ac06acb5afdddae884766f108a6bf8e14c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 132, + 291, + 242 + ], + "blocks": [ + { + "bbox": [ + 50, + 132, + 291, + 242 + ], + "lines": [ + { + "bbox": [ + 50, + 132, + 291, + 242 + ], + "spans": [ + { + "bbox": [ + 50, + 132, + 291, + 242 + ], + "type": "image", + "image_path": "03079a5130d4d503a7d48a4e061c5916e78dc994518d7cfe18fe769045c509b9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 242, + 287, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 353 + ], + "type": "text", + "content": "Figure 6: Top row: Our motivation about needing to strengthen the network to exploit non-degradation regions. Most information is lost due to the complicated rain degradations in the real world, so degradation-free information is needed to reconstruct the clean area. Bottom row: Visualizations on uncertainty map, correlation map and self-attention map. The attention points of the correlation map and the self-attention map differ. We can utilize the correlation map to direct the network's attention towards areas with significant differences, thus enhancing the restoration process through self-attention." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 359, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 287, + 430 + ], + "type": "text", + "content": "scenes, such an approach cannot flexibly capture degradation information from different locations. Our design aims to enable the network to learn the coordinates that can sample degradation information from any position, leading to an enhanced modeling performance of the degradation relationships and network flexibility." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 430, + 288, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 288, + 635 + ], + "type": "text", + "content": "Discussion II: uncertainty driven compared with prior paradigm and rain mask. Regarding uncertainty, the previous UMRL [63] approach used the uncertainty to optimize the rain streak map based on the physical model of the rain streak. Still, it only focused on optimizing the final map, ignoring the network's learning of the intermediate features. Our method uses uncertainty explicitly to drive the network to provide motivation-driven constraints and optimization for the degradation modeling and local reconstruction processes. Compared to the learned rain mask needed GT [46], i) uncertainty map has a theoretical unsupervised loss for practicing. ii). It outperforms the rain mask in representing the network's focus on intricate rain areas beyond mere location. iii). For large-area raindrops with a complicated physical model, it is difficult to learn a rain mask, but the uncertainty map can solve it well by using the discriminative representation of degradations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 638, + 285, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 285, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 285, + 650 + ], + "type": "text", + "content": "3.1.2 Local Reconstruction with Correlation Ranking" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "In rain scenes, information in locally degraded regions is often obscured by rain streaks or raindrops, making it difficult to restore, see Fig.6. To address this issue, it is vital to leverage clean cues in the rain-free area. Building on the discriminative representation of degraded areas and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "other regions in the uncertainty map, we propose using the uncertainty map to generate a correlation map. This map explicitly adjusts the attention map through a ranking approach, as shown in Fig.4 and Fig.3(b), which promotes the network to better leverage clean cues for reconstruction in the form of Query-Key." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 143, + 545, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 143, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 143, + 545, + 221 + ], + "type": "text", + "content": "We begin by partitioning uncertainty map " + }, + { + "bbox": [ + 304, + 143, + 545, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^{\\in \\mathbb{R}^{C\\times H\\times W}}" + }, + { + "bbox": [ + 304, + 143, + 545, + 221 + ], + "type": "text", + "content": " into local patches " + }, + { + "bbox": [ + 304, + 143, + 545, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{U}^{P\\in \\mathbb{R}^{C\\times \\frac{H}{M}\\times \\frac{H}{M}}}" + }, + { + "bbox": [ + 304, + 143, + 545, + 221 + ], + "type": "text", + "content": " that correspond to those patches in the feature map. The presence of discriminative disparities within each patch of the uncertainty map motivates us to calculate the correlation map (CR) between them:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 384, + 221, + 545, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 221, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 384, + 221, + 545, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {C R} _ {i} = \\mathcal {U} _ {i} ^ {P} \\times \\mathcal {U} _ {i} ^ {P ^ {\\mathrm {T}}}. \\tag {10}", + "image_path": "4504e8fcdb17c915a375b2ee8564c83931b5b098ca2dce434e3d125a8617cfd6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 243, + 546, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 546, + 353 + ], + "type": "text", + "content": "Upon obtaining the correlation map, our purpose is to leverage the degradation-free region to facilitate the restoration process of the degraded areas during image reconstruction. Significant differences (low correlation) between degradations and background in the uncertainty map can be considered a prompt, promoting the network to use clean cues from the background for local reconstruction. To this end, we modulate the self-attention map by selecting the low-correlation regions in the correlation map:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 361, + 547, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 547, + 424 + ], + "type": "interline_equation", + "content": "\\mathrm {L R} = \\text {S o f t m a x} \\left(\\frac {\\mathcal {Q} _ {i} ^ {P} \\mathcal {K} _ {i} ^ {P T}}{\\sqrt {\\mathcal {D}}} \\times \\underbrace {[ - \\alpha \\boldsymbol {\\Sigma} _ {i} + (1 + \\alpha) ] + p} _ {\\text {M o d u l a t i o n P r o c e s s i n g}}\\right) \\mathcal {V} _ {i} ^ {P}, \\tag {11}", + "image_path": "4ab77fc2865d2eb2101786dc1ea75bdd10fa999a8aaaa9c1e09583cbf578c0eb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "text", + "content": " is a modulation factor. The " + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 305, + 424, + 545, + 449 + ], + "type": "text", + "content": " is acquired via the Top k ranking approach, which can be expressed as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 457, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 457, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 457, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\left[ \\boldsymbol {\\Sigma} _ {i} \\left(\\mathcal {C R} _ {i}\\right) \\right] _ {x y} = \\left\\{ \\begin{array}{l l} 1 & \\left(\\mathcal {C R} _ {i}\\right) _ {x y} \\in \\operatorname {T o p k} (\\text {r o w} y) \\\\ 0 & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {12}", + "image_path": "80980c438417ab4da2e843944f4107f3758896d0702aeacf0d8bb5146362c443.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 548 + ], + "type": "text", + "content": "In the experiments section, the key parameters of the above used are further studied. About our image reconstruction module, the self-attention FFNs are identical to vanilla vision transformer [13]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 551, + 394, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 394, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 394, + 562 + ], + "type": "text", + "content": "4. Loss Function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "text", + "content": "Our UDR-" + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 304, + 567, + 545, + 604 + ], + "type": "text", + "content": "Former network will predict the derained result and an uncertainty map. Hence, the total loss of our network is defined as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 309, + 612, + 545, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 612, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 309, + 612, + 545, + 650 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {p s n r}} \\left(\\mathcal {B} _ {\\text {p r e}}, \\mathcal {B} _ {\\text {g t}}\\right) + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c e p t u a l}} \\left(\\mathcal {B} _ {\\text {p r e}}, \\mathcal {B} _ {\\text {g t}}\\right) \\\\ + \\lambda_ {3} \\mathcal {L} _ {U D L} (\\mathcal {B} _ {\\text {p r e}}, \\mathcal {U}), \\tag {13} \\\\ \\end{array}", + "image_path": "3788806c3608b0aef7e96916221b03a6d213f8c4c4060ab824ad0c3ebae0e8d3.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{pre}" + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{gt}" + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "text", + "content": " denote the predicted result of image deraining and the corresponding ground truth. " + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "text", + "content": " denotes the predicted uncertainty map. " + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{psnr}" + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{perceptual}" + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "text", + "content": " are the PSNR and perceptual losses, see [6,7]. " + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{UDL}" + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "text", + "content": " denotes the uncertainty loss, see [44, 53] for its definition. " + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2}" + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 304, + 651, + 545, + 723 + ], + "type": "text", + "content": " are empirically set to 1, 0.2 and 1." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13111" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 93, + 252, + 228 + ], + "blocks": [ + { + "bbox": [ + 47, + 60, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 60, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 60, + 287, + 92 + ], + "type": "text", + "content": "Table 1: Quantitative results compared with SOTA methods on the Rain200H [61] and Rain200L [61] datasets. Underline and bold indicate the first and second best results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 93, + 252, + 228 + ], + "lines": [ + { + "bbox": [ + 83, + 93, + 252, + 228 + ], + "spans": [ + { + "bbox": [ + 83, + 93, + 252, + 228 + ], + "type": "table", + "html": "
MethodRain200H [61]Rain200L [61]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
GMM [38]14.710.43028.990.875
JCAS [18]14.870.47130.050.897
DDN [15]26.360.80334.930.958
NLEDN [35]29.510.89138.560.980
RESCAN [37]27.450.82135.080.959
PreNet [50]29.040.89037.120.976
UMRL [63]28.710.88736.430.973
JORDER-E [60]28.580.87636.900.973
MSPFN [28]29.660.89039.480.984
CCN [48]29.990.91438.260.981
MPRNet [69]30.760.90839.890.985
DGUNet [43]30.850.91140.230.986
Uformer [57]30.800.91140.200.986
Restormer [68]31.390.91640.580.987
IDT [58]32.100.93440.740.988
NAFNet [5]30.980.91240.450.987
UDR-S2Former32.59↑0.490.937↑0.00340.96↑0.220.989↑0.001
", + "image_path": "16a136b4cea867bd4e200af1287d95c32431420227d70d60e406a4381e928372.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 245, + 128, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 245, + 128, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 245, + 128, + 259 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 262, + 176, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 176, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 176, + 275 + ], + "type": "text", + "content": "5.1. Implementaion Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": "Our UDR-" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": "Former model utilizes a 5-level encoder-decoder architecture. Within this architecture, the number of channel dimensions increases to " + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\{16,32,64,128,256\\}" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": " across levels 1 to 5. Additionally, the Feature Extraction stage contains " + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\{4,6,7,8\\}" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": " convolutional blocks, while the transformer block number in the latent layer is 8, with 16 heads. Regarding reconstruction, we employ " + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\{3,6,7,8\\}" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": " Image Reconstruction Modules, composed alternately of Sparse Sampling Attention and Local Reconstruction, with " + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\{1,2,4,8\\}" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": " heads. The window size is fixed at " + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "inline_equation", + "content": "8\\times 8" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": ". Finally, we set the constraint factor " + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": ", threshold value " + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": ", modulation factor " + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": ", and the k value for the ranking strategy to 0.6, 0.8, 0.2, and 0.8, respectively. For the self-attention mechanism in this paper, we all use multi-head self-attention, which is consistent with vanilla ViT [13]. Additionally, during the last image reconstruction module of each stage, we convert the output feature and corresponding uncertainty map into the final image format, and the " + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{UDL}" + }, + { + "bbox": [ + 46, + 277, + 287, + 506 + ], + "type": "text", + "content": " is utilized to supervise the learning process of uncertainty map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "text", + "content": "During the training phase, we employed the Adam optimizer with initial momentum " + }, + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.999" + }, + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "text", + "content": ". We initially set the learning rate to 0.0003 and utilized a cyclic learning rate adjustment strategy, whereby the maximum learning rate is set to 0.00036. We trained our model with a data augmentation strategy that included randomly cropping " + }, + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "text", + "content": " patches and using horizontal flipping and random image rotation to a fixed angle. Our training process involved " + }, + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "inline_equation", + "content": "6 \\times 10^{5}" + }, + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "text", + "content": " steps. To leverage perceptual loss, we used the first and third layers of VGG19 [51]. Our model is implemented using PyTorch [45] and the RTX 3090 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 640, + 223, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 223, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 223, + 652 + ], + "type": "text", + "content": "5.2. Evaluation Metrics and Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "Evaluation Metrics. In accordance with prior deraining methodologies [48, 58], follow [16, 17], the performance of the model is evaluated using PSNR [25] and SSIM [56]. As suggested in [48, 61], the assessment of PSNR and SSIM is based on the luminance channel, which refers to the Y" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 424, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 424, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 424, + 84 + ], + "type": "text", + "content": "channel of the YCbCr space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 85, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 545, + 156 + ], + "type": "text", + "content": "Rain Streak Datasets. In order to evaluate the effectiveness of our rain streak removal method, we follow the strategy of previous work [48] and select two benchmark datasets, namely Rain200H and Rain200L [61]. Both two datasets have 1800 synthetic images for training and 200 images for testing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 156, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 204 + ], + "type": "text", + "content": "Raindrop Datasets. We also evaluate our method in a raindrop dataset (i.e., AGAN-Data) collected by Qian et al. [47]. AGAN-Data has 861 images for training and 58 images for testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 205, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 360 + ], + "type": "text", + "content": "Raindrops and Rain Streak Datasets. Our paper aims to develop a methodology that effectively addresses the challenges posed by the presence of large-area raindrops and complex rain streaks. To evaluate the proposed approach, we utilize the RainDS benchmark dataset [48], which includes both real-world and synthetic images, dubbed as RainDS-Real and RainDS-Syn, respectively. These datasets contain images of scenes that include rain streaks only (RS), raindrops only (RD), or both(RDS). RainDS-Syn has 3600 image pairs, with 3000 images used for training and the remaining 600 images for testing. Meanwhile, RainDS-Real comprises of 750 images, with 450 images for training and 300 images for evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 362, + 523, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 362, + 523, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 523, + 374 + ], + "type": "text", + "content": "5.3. Experimental Evaluation on Benchmarks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 378, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 569 + ], + "type": "text", + "content": "Compared Methods. Regarding removing raindrops and rain streaks in images, we conduct extensive experiments to compare various algorithms that can be used for image rain removal. (i) We compare previous SOTA methods (including GMM [38], JCAS [18], DDN [15], NLEDN [35], RESCAN [37], PreNet [50], UMRL [63], JORDER-E [60], MSPFN [28], CCN [48], IDT [58]). (ii) We also compare our network against universal image restoration methods, including MPRNet [69], DGUNet [43], Uformer [57], Restormer [68], NAFNet [5]). For the specific raindrop removal dataset, we compare against Eigen's model [14], Pix2Pix [26], AttentGAN [47], Quan's network [49], CCN [48], and IDT [58]. In the absence of pre-trained models, we conduct model retraining by utilizing publicly available code and subsequently evaluate the best model performance on test datasets to ensure a fair comparison." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Quantitative Comparison. Tables 1, 2, and 3 reports the quantitative results of our network and state-of-the-art methods on four benchmark datasets, which are Rain200H, Rain200L, RainDS, and AGAN-Data. As demonstrated in these Tables, our approach delivers superior metric results over compared state-of-the-art (SOTA) methods for single rain streaks or raindrop degradations. Moreover, by incorporating uncertainty to steer the network's attention towards challenging degradations, our method outperforms the IDT algorithm by 0.49dB, 0.22dB and 1.01dB on the Rain200H [61], Rain200L [61] datasets and AGAN-Data [47], while significantly outperforming previ" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13112" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 106, + 131 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 106, + 131 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 106, + 131 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 106, + 131 + ], + "type": "image", + "image_path": "4661b3b9060803570aca2d9b60a59e76e27b1fce2c47568332efdbd67f863268.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 132, + 94, + 140 + ], + "lines": [ + { + "bbox": [ + 64, + 132, + 94, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 132, + 94, + 140 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 107, + 72, + 161, + 130 + ], + "blocks": [ + { + "bbox": [ + 107, + 72, + 161, + 130 + ], + "lines": [ + { + "bbox": [ + 107, + 72, + 161, + 130 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 161, + 130 + ], + "type": "image", + "image_path": "e44c919bee67622ed881fccf97ff9eb8250d9a1a91731bdd50aadee962d136c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 132, + 144, + 140 + ], + "lines": [ + { + "bbox": [ + 120, + 132, + 144, + 140 + ], + "spans": [ + { + "bbox": [ + 120, + 132, + 144, + 140 + ], + "type": "text", + "content": "(b) CCN" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 162, + 72, + 215, + 130 + ], + "blocks": [ + { + "bbox": [ + 162, + 72, + 215, + 130 + ], + "lines": [ + { + "bbox": [ + 162, + 72, + 215, + 130 + ], + "spans": [ + { + "bbox": [ + 162, + 72, + 215, + 130 + ], + "type": "image", + "image_path": "00d806f974c087b8eb50dd3a85bd604ef500e7c7a5b4381615a2b0d88e183060.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 132, + 205, + 140 + ], + "lines": [ + { + "bbox": [ + 167, + 132, + 205, + 140 + ], + "spans": [ + { + "bbox": [ + 167, + 132, + 205, + 140 + ], + "type": "text", + "content": "(c) MPRNet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 216, + 72, + 269, + 130 + ], + "blocks": [ + { + "bbox": [ + 216, + 72, + 269, + 130 + ], + "lines": [ + { + "bbox": [ + 216, + 72, + 269, + 130 + ], + "spans": [ + { + "bbox": [ + 216, + 72, + 269, + 130 + ], + "type": "image", + "image_path": "95a0fe0411d66e95ff91de179f0a87375979b33891ac098f48a8c9ccd8c773b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 132, + 259, + 140 + ], + "lines": [ + { + "bbox": [ + 222, + 132, + 259, + 140 + ], + "spans": [ + { + "bbox": [ + 222, + 132, + 259, + 140 + ], + "type": "text", + "content": "(d) DGUNet" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 270, + 72, + 323, + 130 + ], + "blocks": [ + { + "bbox": [ + 270, + 72, + 323, + 130 + ], + "lines": [ + { + "bbox": [ + 270, + 72, + 323, + 130 + ], + "spans": [ + { + "bbox": [ + 270, + 72, + 323, + 130 + ], + "type": "image", + "image_path": "91135ad22006f61f0acf3b922ebf4eef6724475bf37937eb8b294df3400e68c4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 324, + 72, + 378, + 130 + ], + "blocks": [ + { + "bbox": [ + 324, + 72, + 378, + 130 + ], + "lines": [ + { + "bbox": [ + 324, + 72, + 378, + 130 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 378, + 130 + ], + "type": "image", + "image_path": "ae5c647ed1f41523b7c0f1469335909a858512c724ebfffd47a189a69ff4b79d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 378, + 72, + 432, + 130 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 432, + 130 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 432, + 130 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 432, + 130 + ], + "type": "image", + "image_path": "e4200f6a7e050e033df00242605ab53e37e65960de2a16e33fee5920b239c25d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 132, + 423, + 140 + ], + "lines": [ + { + "bbox": [ + 387, + 132, + 423, + 140 + ], + "spans": [ + { + "bbox": [ + 387, + 132, + 423, + 140 + ], + "type": "text", + "content": "(g) NAFNet" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 433, + 72, + 486, + 130 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 486, + 130 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 486, + 130 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 486, + 130 + ], + "type": "image", + "image_path": "334e416b92a24680aa163dd7906c8b74ed02ebb9026c7a485398057425833fde.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 487, + 72, + 541, + 130 + ], + "blocks": [ + { + "bbox": [ + 487, + 72, + 541, + 130 + ], + "lines": [ + { + "bbox": [ + 487, + 72, + 541, + 130 + ], + "spans": [ + { + "bbox": [ + 487, + 72, + 541, + 130 + ], + "type": "image", + "image_path": "782a0ec52bcc5fe2109dd0da9d7a05d6200e48a5cd80adce08995dd89a1c88dd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 53, + 142, + 209, + 232 + ], + "blocks": [ + { + "bbox": [ + 53, + 142, + 209, + 232 + ], + "lines": [ + { + "bbox": [ + 53, + 142, + 209, + 232 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 209, + 232 + ], + "type": "image", + "image_path": "d80353b8f21b50708cc7cec013affaa53c094d0c7475473a5aea5e4c4f7e51d4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 232, + 146, + 241 + ], + "lines": [ + { + "bbox": [ + 116, + 232, + 146, + 241 + ], + "spans": [ + { + "bbox": [ + 116, + 232, + 146, + 241 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 257, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 545, + 300 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons of various SOTA approaches on the RainDS [48] benchmark. Our proposed method outperforms all other approaches on both synthetic and real-world datasets, including all types of precipitation (i.e., rain streaks (RS), raindrops (RD), and a combination of both (RDS)). The results are indicated in bold and underline, representing the first and second-best performances, respectively. " + }, + { + "bbox": [ + 46, + 257, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 257, + 545, + 300 + ], + "type": "text", + "content": " means the higher is better. The #GFLOPs are calculated by " + }, + { + "bbox": [ + 46, + 257, + 545, + 300 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 46, + 257, + 545, + 300 + ], + "type": "text", + "content": " image resolution for a fair comparison." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 214, + 142, + 290, + 178 + ], + "blocks": [ + { + "bbox": [ + 274, + 132, + 302, + 140 + ], + "lines": [ + { + "bbox": [ + 274, + 132, + 302, + 140 + ], + "spans": [ + { + "bbox": [ + 274, + 132, + 302, + 140 + ], + "type": "text", + "content": "(e) Resto" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 214, + 142, + 290, + 178 + ], + "lines": [ + { + "bbox": [ + 214, + 142, + 290, + 178 + ], + "spans": [ + { + "bbox": [ + 214, + 142, + 290, + 178 + ], + "type": "image", + "image_path": "fc044bc659c8cbc7765ac349d9ef8de1139dbca51891cd1d5198ce6dab256596.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 180, + 264, + 189 + ], + "lines": [ + { + "bbox": [ + 242, + 180, + 264, + 189 + ], + "spans": [ + { + "bbox": [ + 242, + 180, + 264, + 189 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 296, + 142, + 373, + 178 + ], + "blocks": [ + { + "bbox": [ + 341, + 132, + 361, + 140 + ], + "lines": [ + { + "bbox": [ + 341, + 132, + 361, + 140 + ], + "spans": [ + { + "bbox": [ + 341, + 132, + 361, + 140 + ], + "type": "text", + "content": "(f) IDT" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 142, + 373, + 178 + ], + "lines": [ + { + "bbox": [ + 296, + 142, + 373, + 178 + ], + "spans": [ + { + "bbox": [ + 296, + 142, + 373, + 178 + ], + "type": "image", + "image_path": "8d53039f29147929bcffca07ed287a6649f954b48e0f5e2139c548dba336ed9a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 180, + 347, + 189 + ], + "lines": [ + { + "bbox": [ + 321, + 180, + 347, + 189 + ], + "spans": [ + { + "bbox": [ + 321, + 180, + 347, + 189 + ], + "type": "text", + "content": "(c) CCN" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 379, + 142, + 455, + 178 + ], + "blocks": [ + { + "bbox": [ + 444, + 132, + 471, + 140 + ], + "lines": [ + { + "bbox": [ + 444, + 132, + 471, + 140 + ], + "spans": [ + { + "bbox": [ + 444, + 132, + 471, + 140 + ], + "type": "text", + "content": "(h) Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 379, + 142, + 455, + 178 + ], + "lines": [ + { + "bbox": [ + 379, + 142, + 455, + 178 + ], + "spans": [ + { + "bbox": [ + 379, + 142, + 455, + 178 + ], + "type": "image", + "image_path": "1d69d6ef4c27e24a5ae353d0017ac570cc4040fbe76342db59757c98afe7ba1a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 180, + 438, + 189 + ], + "lines": [ + { + "bbox": [ + 398, + 180, + 438, + 189 + ], + "spans": [ + { + "bbox": [ + 398, + 180, + 438, + 189 + ], + "type": "text", + "content": "(d) MPRNet" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 462, + 142, + 539, + 178 + ], + "blocks": [ + { + "bbox": [ + 505, + 132, + 523, + 140 + ], + "lines": [ + { + "bbox": [ + 505, + 132, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 505, + 132, + 523, + 140 + ], + "type": "text", + "content": "(i) GT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 142, + 539, + 178 + ], + "lines": [ + { + "bbox": [ + 462, + 142, + 539, + 178 + ], + "spans": [ + { + "bbox": [ + 462, + 142, + 539, + 178 + ], + "type": "image", + "image_path": "12853acdb0fc7ae992aeab31ee7366da601a7da7e6a4ad30af1482721275739b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 180, + 517, + 189 + ], + "lines": [ + { + "bbox": [ + 477, + 180, + 517, + 189 + ], + "spans": [ + { + "bbox": [ + 477, + 180, + 517, + 189 + ], + "type": "text", + "content": "(e) DGUNet" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 214, + 193, + 290, + 230 + ], + "blocks": [ + { + "bbox": [ + 214, + 193, + 290, + 230 + ], + "lines": [ + { + "bbox": [ + 214, + 193, + 290, + 230 + ], + "spans": [ + { + "bbox": [ + 214, + 193, + 290, + 230 + ], + "type": "image", + "image_path": "c5c9a77c0c56ca1b6fc07ac89e038b7c8e255d5e86fb94bb81b92dd02fca6a5a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 232, + 276, + 240 + ], + "lines": [ + { + "bbox": [ + 230, + 232, + 276, + 240 + ], + "spans": [ + { + "bbox": [ + 230, + 232, + 276, + 240 + ], + "type": "text", + "content": "(f) Restomer" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 242, + 541, + 253 + ], + "lines": [ + { + "bbox": [ + 49, + 242, + 541, + 253 + ], + "spans": [ + { + "bbox": [ + 49, + 242, + 541, + 253 + ], + "type": "text", + "content": "Figure 7: Visual comparisons of removal for raindrops and rain streaks from RainDS-Syn [48] dataset. Zoom it for a better illustration." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 296, + 193, + 373, + 229 + ], + "blocks": [ + { + "bbox": [ + 296, + 193, + 373, + 229 + ], + "lines": [ + { + "bbox": [ + 296, + 193, + 373, + 229 + ], + "spans": [ + { + "bbox": [ + 296, + 193, + 373, + 229 + ], + "type": "image", + "image_path": "3f282a81a3ba8d3e62a7df4e37e8cc4894af73180c443580f8c7d09a9bc92bbf.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 232, + 346, + 240 + ], + "lines": [ + { + "bbox": [ + 323, + 232, + 346, + 240 + ], + "spans": [ + { + "bbox": [ + 323, + 232, + 346, + 240 + ], + "type": "text", + "content": "(g) IDT" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 380, + 193, + 455, + 229 + ], + "blocks": [ + { + "bbox": [ + 380, + 193, + 455, + 229 + ], + "lines": [ + { + "bbox": [ + 380, + 193, + 455, + 229 + ], + "spans": [ + { + "bbox": [ + 380, + 193, + 455, + 229 + ], + "type": "image", + "image_path": "c910c365c43ab37ee2b03cad04cc0eb773118dc70f116d111379ae1c9ff4a6c6.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 232, + 437, + 240 + ], + "lines": [ + { + "bbox": [ + 398, + 232, + 437, + 240 + ], + "spans": [ + { + "bbox": [ + 398, + 232, + 437, + 240 + ], + "type": "text", + "content": "(h) NAFNet" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 462, + 193, + 539, + 229 + ], + "blocks": [ + { + "bbox": [ + 462, + 193, + 539, + 229 + ], + "lines": [ + { + "bbox": [ + 462, + 193, + 539, + 229 + ], + "spans": [ + { + "bbox": [ + 462, + 193, + 539, + 229 + ], + "type": "image", + "image_path": "8c2d70f53dbff353cd86fa32f1b07f4b92f3e396ca890ad0b2aeab8554fef10f.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 232, + 514, + 239 + ], + "lines": [ + { + "bbox": [ + 487, + 232, + 514, + 239 + ], + "spans": [ + { + "bbox": [ + 487, + 232, + 514, + 239 + ], + "type": "text", + "content": "(i) Ours" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "table", + "bbox": [ + 55, + 303, + 543, + 453 + ], + "blocks": [ + { + "bbox": [ + 55, + 303, + 543, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 543, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 543, + 453 + ], + "type": "table", + "html": "
MethodVenueRainDS-SynRainDS-Real#Param #GFLOPs
RSRDRDSRSRDRDS
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
GMM [38]CVPR'201626.660.78123.040.79321.500.66923.730.56018.600.55421.350.576--
JCAS [18]ICCV'201726.460.78623.150.81120.910.67124.040.55618.180.55521.220.585--
DDN [15]CVPR'201730.410.86927.920.88526.850.79624.850.68323.120.64222.470.606--
NLEDN [35]ACMMM'201836.240.95834.870.95732.130.91727.020.72324.710.67124.060.650--
RESCAN [37]ECCV'201830.990.88729.900.90727.430.81826.700.68324.230.63723.230.5870.15M32.32G
PreNet [50]CVPR'201936.630.96834.580.96432.210.93426.430.72924.420.67923.570.6490.17M66.58G
UMRL [63]CVPR'201935.760.96233.590.95831.570.92925.890.72623.930.67623.010.6470.98M16.50G
JORDER-E [60]TPAMI'201933.650.92533.510.94430.050.87026.560.71324.340.66223.540.6294.17M273.68G
MSPFN [28]CVPR'202038.610.97536.930.97334.080.94726.450.72724.490.68124.110.65121.00M708.44G
CCN [48]CVPR'202139.170.98137.300.97634.790.95727.460.73725.140.70124.930.6793.75M245.85G
MPRNet [69]CVPR'202140.810.98137.030.97234.990.95627.290.73625.260.70124.960.6813.64M148.55G
DGUNet [43]CVPR'202241.090.98337.560.97535.340.95927.520.73725.330.70224.990.68312.18M199.74G
Uformer [57]CVPR'202240.690.97237.080.96634.990.95426.890.73025.310.70124.830.68620.63M43.86G
Restomer [68]CVPR'2022(Oral)41.420.98038.780.97636.080.96127.390.74225.380.70224.920.68526.10M140.99G
IDT [58]TPAMI'202241.610.98339.090.98036.230.96027.510.74325.670.70624.990.68916.00M61.90G
NAFNet [5]ECCV'202240.390.97237.230.97434.990.95727.490.72925.230.70124.640.66340.60M16.19G
UDR-S2Former (Ours)ICCV'202342.39↑0.780.988↑0.00539.78↑0.690.983↑0.00336.91↑0.680.966↑0.00627.90↑0.390.745↑0.00226.01↑0.340.709↑0.00325.52↑0.530.691↑0.0028.53M21.58G
", + "image_path": "ab1506730d2addb72eaaaa628571f84a51c369d091725dd322cfa77efef0bb16.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_body" + } + ], + "index": 38 + }, + { + "type": "table", + "bbox": [ + 90, + 497, + 246, + 572 + ], + "blocks": [ + { + "bbox": [ + 46, + 464, + 287, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 495 + ], + "type": "text", + "content": "Table 3: Quantitative results compared with SOTA methods on the AGAN-Data [47]. Red and blue indicate the first and second best results." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 90, + 497, + 246, + 572 + ], + "lines": [ + { + "bbox": [ + 90, + 497, + 246, + 572 + ], + "spans": [ + { + "bbox": [ + 90, + 497, + 246, + 572 + ], + "type": "table", + "html": "
MethodPSNR†SSIM†
(ICCV'2013)Eigen's model [14]21.310.757
(CVPR2017)Pix2Pix [26]27.200.836
(CVPR'2018)AttenGAN [47]31.590.917
(ICCV'2019)Quan's network [49]31.370.918
(CVPR'2021)CCN [48]31.340.929
(TPAMI'2022)IDT [58]31.630.936
UDR-S2Former32.640.943
", + "image_path": "8a415f1b0dbbaa8ec5b5824b7cead6a58acf4deed7681fbdea344beb76f8671a.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "table_body" + } + ], + "index": 40 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "ous uncertainty-based [63] methods. Table 2 demonstrates that our approach surpasses all prior designs for jointly removing rain streaks and raindrops. While networks like Restormer [68] and IDT [58] can model long-distance dependencies sufficiently, our method is able to effectively model complex degradation relationships in rain-dominated landscapes (i.e., raindrops and rain streaks both) with uncertainty. Additionally, sparse sampling further facilitates the network's ability to obtain crucial global degradation information at a low cost, see Table 2. Based on the indicators," + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 473, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 544 + ], + "type": "text", + "content": "our approach exhibits a significant advantage over the latest designs, including Restormer (36.08PSNR " + }, + { + "bbox": [ + 304, + 473, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 473, + 545, + 544 + ], + "type": "text", + "content": " 36.91PSNR) and IDT (36.23PSNR " + }, + { + "bbox": [ + 304, + 473, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 473, + 545, + 544 + ], + "type": "text", + "content": " 36.91PSNR), in both real-world and synthetic datasets. Additionally, we also present the comparison of speed and memory cost in the supplementary material to further demonstrate our superiority." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "Visual Comparison. We present the comparisons of the visual effects in Fig.7-9. As depicted in the figures, we observe that in a scene containing rain streaks and raindrops, the scene information is either occluded by a broad range of raindrops or masked by dense rain streaks of varying scales. Other approaches either struggle to remove complex degradations or significantly compromise image details. Our approach leverages degradation relationship modeling to facilitate image restoration, and utilizes the clean information to recover intricate rain degradations. The resulting output exhibits higher fidelity with the more accurate restoration of great details. In real-world scenarios, our method enables the removal of various forms of degradation while recovering fine details, outperforming previous methods." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13113" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 210, + 198 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 210, + 198 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 210, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 210, + 198 + ], + "type": "image", + "image_path": "d1650df8a459301e2502c0da25dfdebc1fa05f652c967b11e638ac5f7f8965f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 201, + 146, + 210 + ], + "lines": [ + { + "bbox": [ + 115, + 201, + 146, + 210 + ], + "spans": [ + { + "bbox": [ + 115, + 201, + 146, + 210 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 71, + 288, + 125 + ], + "blocks": [ + { + "bbox": [ + 211, + 71, + 288, + 125 + ], + "lines": [ + { + "bbox": [ + 211, + 71, + 288, + 125 + ], + "spans": [ + { + "bbox": [ + 211, + 71, + 288, + 125 + ], + "type": "image", + "image_path": "2d5f3cf3304e873e3ac7af6f9d440126ad26ef2fbf2f54467209182c99fa0731.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 129, + 261, + 138 + ], + "lines": [ + { + "bbox": [ + 239, + 129, + 261, + 138 + ], + "spans": [ + { + "bbox": [ + 239, + 129, + 261, + 138 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 295, + 71, + 373, + 125 + ], + "blocks": [ + { + "bbox": [ + 295, + 71, + 373, + 125 + ], + "lines": [ + { + "bbox": [ + 295, + 71, + 373, + 125 + ], + "spans": [ + { + "bbox": [ + 295, + 71, + 373, + 125 + ], + "type": "image", + "image_path": "ede9051f35f56847a0bbea5b66658ae1d14a1ce77742a9711e074e1e305b3f3d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 129, + 346, + 138 + ], + "lines": [ + { + "bbox": [ + 320, + 129, + 346, + 138 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 346, + 138 + ], + "type": "text", + "content": "(c) CCN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 379, + 71, + 457, + 125 + ], + "blocks": [ + { + "bbox": [ + 379, + 71, + 457, + 125 + ], + "lines": [ + { + "bbox": [ + 379, + 71, + 457, + 125 + ], + "spans": [ + { + "bbox": [ + 379, + 71, + 457, + 125 + ], + "type": "image", + "image_path": "34e739911121ef46b15cc7ad233e7824b648c540164230269ed7101277082c51.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 129, + 439, + 138 + ], + "lines": [ + { + "bbox": [ + 397, + 129, + 439, + 138 + ], + "spans": [ + { + "bbox": [ + 397, + 129, + 439, + 138 + ], + "type": "text", + "content": "(d) MPRNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 463, + 71, + 541, + 125 + ], + "blocks": [ + { + "bbox": [ + 463, + 71, + 541, + 125 + ], + "lines": [ + { + "bbox": [ + 463, + 71, + 541, + 125 + ], + "spans": [ + { + "bbox": [ + 463, + 71, + 541, + 125 + ], + "type": "image", + "image_path": "e7f0be0f7aa886d0a351f82358ff85cc4438a3ceec842c67f3ab901f031c9616.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 129, + 523, + 138 + ], + "lines": [ + { + "bbox": [ + 482, + 129, + 523, + 138 + ], + "spans": [ + { + "bbox": [ + 482, + 129, + 523, + 138 + ], + "type": "text", + "content": "(e) DGUNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 211, + 144, + 288, + 198 + ], + "blocks": [ + { + "bbox": [ + 211, + 144, + 288, + 198 + ], + "lines": [ + { + "bbox": [ + 211, + 144, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 211, + 144, + 288, + 198 + ], + "type": "image", + "image_path": "e9882578ce49b3e8a79f0921759750eb2833f77a4c1b8b5e4a11cc5898f271ca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 201, + 274, + 211 + ], + "lines": [ + { + "bbox": [ + 227, + 201, + 274, + 211 + ], + "spans": [ + { + "bbox": [ + 227, + 201, + 274, + 211 + ], + "type": "text", + "content": "(f) Restormer" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 295, + 145, + 373, + 198 + ], + "blocks": [ + { + "bbox": [ + 295, + 145, + 373, + 198 + ], + "lines": [ + { + "bbox": [ + 295, + 145, + 373, + 198 + ], + "spans": [ + { + "bbox": [ + 295, + 145, + 373, + 198 + ], + "type": "image", + "image_path": "9e2679e241c3d45687e66636ed8409dea6f041bc3cea437448922b5c7e465e63.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 201, + 346, + 211 + ], + "lines": [ + { + "bbox": [ + 321, + 201, + 346, + 211 + ], + "spans": [ + { + "bbox": [ + 321, + 201, + 346, + 211 + ], + "type": "text", + "content": "(g) IDT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 379, + 144, + 457, + 198 + ], + "blocks": [ + { + "bbox": [ + 379, + 144, + 457, + 198 + ], + "lines": [ + { + "bbox": [ + 379, + 144, + 457, + 198 + ], + "spans": [ + { + "bbox": [ + 379, + 144, + 457, + 198 + ], + "type": "image", + "image_path": "d5d142ce757652236acf13bff8eb361602200c667da6849a62dcd88954a0228a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 202, + 438, + 211 + ], + "lines": [ + { + "bbox": [ + 398, + 202, + 438, + 211 + ], + "spans": [ + { + "bbox": [ + 398, + 202, + 438, + 211 + ], + "type": "text", + "content": "(h) NAFNet" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 463, + 144, + 541, + 198 + ], + "blocks": [ + { + "bbox": [ + 463, + 144, + 541, + 198 + ], + "lines": [ + { + "bbox": [ + 463, + 144, + 541, + 198 + ], + "spans": [ + { + "bbox": [ + 463, + 144, + 541, + 198 + ], + "type": "image", + "image_path": "a2d3e2b06ea3d6cc0fac92d3ccaa40f344903a7f5742ff5e3cddc2277f7bf948.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 202, + 515, + 211 + ], + "lines": [ + { + "bbox": [ + 488, + 202, + 515, + 211 + ], + "spans": [ + { + "bbox": [ + 488, + 202, + 515, + 211 + ], + "type": "text", + "content": "(i) Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 48, + 232, + 108, + 282 + ], + "blocks": [ + { + "bbox": [ + 48, + 232, + 108, + 282 + ], + "lines": [ + { + "bbox": [ + 48, + 232, + 108, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 108, + 282 + ], + "type": "image", + "image_path": "c1650e8e0558982d5749552c994726198a0ca45bdbe24045703b04ff65275441.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 284, + 91, + 290 + ], + "lines": [ + { + "bbox": [ + 67, + 284, + 91, + 290 + ], + "spans": [ + { + "bbox": [ + 67, + 284, + 91, + 290 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 110, + 232, + 169, + 282 + ], + "blocks": [ + { + "bbox": [ + 110, + 232, + 169, + 282 + ], + "lines": [ + { + "bbox": [ + 110, + 232, + 169, + 282 + ], + "spans": [ + { + "bbox": [ + 110, + 232, + 169, + 282 + ], + "type": "image", + "image_path": "72012fe37ca959d90e275b379fb72b334d9ac1cece175c83e32cce5052e63334.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 284, + 149, + 289 + ], + "lines": [ + { + "bbox": [ + 129, + 284, + 149, + 289 + ], + "spans": [ + { + "bbox": [ + 129, + 284, + 149, + 289 + ], + "type": "text", + "content": "(b) CCN" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 170, + 232, + 230, + 282 + ], + "blocks": [ + { + "bbox": [ + 157, + 213, + 434, + 224 + ], + "lines": [ + { + "bbox": [ + 157, + 213, + 434, + 224 + ], + "spans": [ + { + "bbox": [ + 157, + 213, + 434, + 224 + ], + "type": "text", + "content": "Figure 8: Visual results for real-world sample of RainDS-Real [48] dataset." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 232, + 230, + 282 + ], + "lines": [ + { + "bbox": [ + 170, + 232, + 230, + 282 + ], + "spans": [ + { + "bbox": [ + 170, + 232, + 230, + 282 + ], + "type": "image", + "image_path": "6a3585fa80c2518b44f44d16cc0ed26a56d78e26e01ca6de5be32964554c81c4.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 284, + 214, + 289 + ], + "lines": [ + { + "bbox": [ + 186, + 284, + 214, + 289 + ], + "spans": [ + { + "bbox": [ + 186, + 284, + 214, + 289 + ], + "type": "text", + "content": "(c) DGUNet" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 231, + 232, + 291, + 282 + ], + "blocks": [ + { + "bbox": [ + 231, + 232, + 291, + 282 + ], + "lines": [ + { + "bbox": [ + 231, + 232, + 291, + 282 + ], + "spans": [ + { + "bbox": [ + 231, + 232, + 291, + 282 + ], + "type": "image", + "image_path": "c8e7ffe1bb5c11471f44c577b1faa5f342fa68a10c47b426b2b8facdf1a514e4.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 284, + 278, + 289 + ], + "lines": [ + { + "bbox": [ + 243, + 284, + 278, + 289 + ], + "spans": [ + { + "bbox": [ + 243, + 284, + 278, + 289 + ], + "type": "text", + "content": "(d) Restomer" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 49, + 293, + 108, + 342 + ], + "blocks": [ + { + "bbox": [ + 49, + 293, + 108, + 342 + ], + "lines": [ + { + "bbox": [ + 49, + 293, + 108, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 293, + 108, + 342 + ], + "type": "image", + "image_path": "b822c43491e0fe5225eac5c0628d3246ebc7864b362e0cfa590321ecad4576e0.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 343, + 87, + 350 + ], + "lines": [ + { + "bbox": [ + 70, + 343, + 87, + 350 + ], + "spans": [ + { + "bbox": [ + 70, + 343, + 87, + 350 + ], + "type": "text", + "content": "(e) IDT" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 352, + 286, + 373 + ], + "lines": [ + { + "bbox": [ + 47, + 352, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 286, + 373 + ], + "type": "text", + "content": "Figure 9: Visual results for real-world sample of RainDS-Real [48] dataset, Please zoom it for a better illustration." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 110, + 293, + 169, + 342 + ], + "blocks": [ + { + "bbox": [ + 110, + 293, + 169, + 342 + ], + "lines": [ + { + "bbox": [ + 110, + 293, + 169, + 342 + ], + "spans": [ + { + "bbox": [ + 110, + 293, + 169, + 342 + ], + "type": "image", + "image_path": "82843bbe1e3a8b4031c70e40dab0161769de12f5c4308a4f2ba21c3f371e30e7.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 343, + 153, + 350 + ], + "lines": [ + { + "bbox": [ + 126, + 343, + 153, + 350 + ], + "spans": [ + { + "bbox": [ + 126, + 343, + 153, + 350 + ], + "type": "text", + "content": "(f) NAFNet" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 170, + 293, + 230, + 342 + ], + "blocks": [ + { + "bbox": [ + 170, + 293, + 230, + 342 + ], + "lines": [ + { + "bbox": [ + 170, + 293, + 230, + 342 + ], + "spans": [ + { + "bbox": [ + 170, + 293, + 230, + 342 + ], + "type": "image", + "image_path": "c9f1f32b47ed79de199abc3da64414af6b9851e4228d6380c3b913b4fd095e08.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 343, + 210, + 350 + ], + "lines": [ + { + "bbox": [ + 189, + 343, + 210, + 350 + ], + "spans": [ + { + "bbox": [ + 189, + 343, + 210, + 350 + ], + "type": "text", + "content": "(g) Ours" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 231, + 293, + 291, + 342 + ], + "blocks": [ + { + "bbox": [ + 231, + 293, + 291, + 342 + ], + "lines": [ + { + "bbox": [ + 231, + 293, + 291, + 342 + ], + "spans": [ + { + "bbox": [ + 231, + 293, + 291, + 342 + ], + "type": "image", + "image_path": "e35d81cf6ef259b8b35d04d0fb5d7b8e271b6e5131aeb28759b1e60b7bba55ea.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 343, + 269, + 350 + ], + "lines": [ + { + "bbox": [ + 252, + 343, + 269, + 350 + ], + "spans": [ + { + "bbox": [ + 252, + 343, + 269, + 350 + ], + "type": "text", + "content": "(h) GT" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 48, + 385, + 168, + 444 + ], + "blocks": [ + { + "bbox": [ + 91, + 377, + 144, + 384 + ], + "lines": [ + { + "bbox": [ + 91, + 377, + 144, + 384 + ], + "spans": [ + { + "bbox": [ + 91, + 377, + 144, + 384 + ], + "type": "text", + "content": "(a) Ablation of " + }, + { + "bbox": [ + 91, + 377, + 144, + 384 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 91, + 377, + 144, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 91, + 377, + 144, + 384 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 385, + 168, + 444 + ], + "lines": [ + { + "bbox": [ + 48, + 385, + 168, + 444 + ], + "spans": [ + { + "bbox": [ + 48, + 385, + 168, + 444 + ], + "type": "image", + "image_path": "2d13ca8de42c27a0b28f1f224f5e30ff96401262bfb6d8e72e5240d84d6c2bdd.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 445, + 286, + 466 + ], + "lines": [ + { + "bbox": [ + 47, + 445, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 286, + 466 + ], + "type": "text", + "content": "Figure 10: About the ablation experiments of the parameter " + }, + { + "bbox": [ + 47, + 445, + 286, + 466 + ], + "type": "inline_equation", + "content": "\\beta, \\gamma, \\alpha" + }, + { + "bbox": [ + 47, + 445, + 286, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 445, + 286, + 466 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 445, + 286, + 466 + ], + "type": "text", + "content": ", the vertical axis is the PSNR metric." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 173, + 385, + 291, + 444 + ], + "blocks": [ + { + "bbox": [ + 212, + 377, + 266, + 384 + ], + "lines": [ + { + "bbox": [ + 212, + 377, + 266, + 384 + ], + "spans": [ + { + "bbox": [ + 212, + 377, + 266, + 384 + ], + "type": "text", + "content": "(b) Ablation of " + }, + { + "bbox": [ + 212, + 377, + 266, + 384 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 212, + 377, + 266, + 384 + ], + "type": "text", + "content": " and k" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 385, + 291, + 444 + ], + "lines": [ + { + "bbox": [ + 173, + 385, + 291, + 444 + ], + "spans": [ + { + "bbox": [ + 173, + 385, + 291, + 444 + ], + "type": "image", + "image_path": "ecaa1242dbc566f35826dbbf6b7984024361811748464dcfbc8a285a231e6a59.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 47, + 476, + 147, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 147, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 147, + 488 + ], + "type": "text", + "content": "6. Ablation Studies" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 494, + 286, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 286, + 589 + ], + "type": "text", + "content": "To validate the effectiveness of our proposed design, we performed ablation experiments on UDR-S²Former. To this end, we utilized the training dataset of RainDS-Syn [48] for training our model, and evaluated its performance on the RDS scenes of the corresponding testing set. The experimental settings are kept consistent with the aforementioned ones. In the subsequent sections, we analyze the individual contributions of each module toward the performance." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 47, + 590, + 286, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 286, + 602 + ], + "type": "text", + "content": "6.1. Effectiveness of Image Reconstruction Module" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 46, + 605, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 689 + ], + "type": "text", + "content": "Ablation researches of parameters mentioned in IRM. Our ablation studies involve thorough research of the parameters employed in the method, and we have depicted the results in Fig.10. As demonstrated, the optimal value for each parameter lies within a particular range, which can be attributed to our initial design motivation holding theoretical significance within that range." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Improvements of Sparse Sampling Attention (SSA) in IRM. We conducted a comparison of our sparse sampling" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "content": "attention with sparse attention in ART [71] (SA), window-based self-attention [58] (WSA), and channel-dimension self-attention [68] (CSA). The results are shown in Table 4. It demonstrates that sparse sampling attention performs significantly better than previous designs, underscoring the importance of incorporating degradation information by adaptive sampling from a global perspective for the unified removal of complex rain scenes." + } + ] + } + ], + "index": 46 + }, + { + "type": "table", + "bbox": [ + 313, + 360, + 425, + 406 + ], + "blocks": [ + { + "bbox": [ + 306, + 338, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 306, + 338, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 545, + 358 + ], + "type": "text", + "content": "Table 4: Ablation studies on Table 5: Ablation studies on SSA (§3.1.1). UDR. (§3.1.1 and 3.1.2)." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 360, + 425, + 406 + ], + "lines": [ + { + "bbox": [ + 313, + 360, + 425, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 425, + 406 + ], + "type": "table", + "html": "
SettingModelPSNRSSIM
iSA [71]36.390.962
iiWSA [58]36.150.959
iiiCSA [68]36.230.960
ivSSA (Ours)36.910.966
", + "image_path": "f81794038d28ddb7a6e7b07cdfd45072b6c8dbbe9b3972855a0c783c60ee6ff0.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "table_body" + } + ], + "index": 48 + }, + { + "type": "table", + "bbox": [ + 436, + 360, + 541, + 407 + ], + "blocks": [ + { + "bbox": [ + 436, + 360, + 541, + 407 + ], + "lines": [ + { + "bbox": [ + 436, + 360, + 541, + 407 + ], + "spans": [ + { + "bbox": [ + 436, + 360, + 541, + 407 + ], + "type": "table", + "html": "
SettingModelPSNRSSIM
iSSA w/o UD36.250.960
iiSSA w/o RS36.290.961
iiiLR w/o UD36.330.962
ivLR w/o RS36.450.963
vOurs36.910.966
", + "image_path": "a2714fa2c7a88e6d2f67d14e022274e0bd2dc7af38187cf40bf85dbc32cbfe54.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "table_body" + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 418, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 527 + ], + "type": "text", + "content": "Gains of Uncertainty-Driven Ranking (UDR). In this section, we perform ablation studies of our uncertainty driven (UD) and ranking strategies (RS). The results in Table 5 demonstrate that leveraging uncertainty can effectively constrain the process of sparse sampling and facilitate the network to recover locally degraded regions. However, using uncertainty maps directly without employing a ranking strategy leads to a significant drop in performance, due to the lack of using explicit properties of uncertainty maps." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 306, + 529, + 378, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 378, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 378, + 540 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 546, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 665 + ], + "type": "text", + "content": "In this paper, we present a sparse sampling transformer with the uncertainty-driven ranking that removes rain streaks and raindrops in a unified approach. Our approach employs sparse sampling self-attention to effectively capture global degradation relationships in an adaptive manner. We explicitly leverage the uncertainty map via a ranking strategy to constrain the sampling process and facilitate local reconstruction. Our method achieves SOTA results while requiring minimal computational overhead, showcasing its superiority over existing approaches." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by Guangzhou Municipal Science and Technology Project (Grant No. 2023A03J0671), and the National Natural Science Foundation of China (Grant No. 61902275)." + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13114" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 39(12):2481-2495, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5710-5719, 2020. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 247 + ], + "type": "text", + "content": "[3] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 249, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 287, + 282 + ], + "type": "text", + "content": "[4] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 384 + ], + "type": "text", + "content": "[6] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "type": "text", + "content": "[7] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Dualformer: Hybrid self-attention transformer for efficient image restoration. arXiv preprint arXiv:2210.01069, 2022. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 464 + ], + "type": "text", + "content": "[8] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "type": "text", + "content": "[9] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Yi Ye, and Erkang Chen. Msp-former: Multi-scale projection transformer for single image desnowing. arXiv preprint arXiv:2207.05621, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 577 + ], + "type": "text", + "content": "[10] Sixiang Chen, Tian Ye, Jun Shi, Yun Liu, JingXia Jiang, Erkang Chen, and Peng Chen. Dehrformer: Real-time transformer for depth estimation and haze removal from varicolored haze scenes. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "text", + "content": "[11] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Sen Deng, Mingqiang Wei, Jun Wang, Yidan Feng, Luming Liang, Haoran Xie, Fu Lee Wang, and Meng Wang. Detail-recovery image deraining via context aggregation networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14560-14569, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.6.7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 118, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 163 + ], + "type": "text", + "content": "[14] David Eigen, Dilip Krishnan, and Rob Fergus. Restoring an image taken through a window covered with dirt or rain. In Proceedings of the IEEE international conference on computer vision, pages 633-640, 2013. 3, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 164, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 217 + ], + "type": "text", + "content": "[15] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3855-3863, 2017. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[16] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 265, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 319 + ], + "type": "text", + "content": "[17] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "type": "text", + "content": "[18] Shuhang Gu, Deyu Meng, Wangmeng Zuo, and Lei Zhang. Joint convolutional analysis and synthesis sparse representation for single image layer separation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1708-1716, 2017. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 377, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 420 + ], + "type": "text", + "content": "[19] Ming Hong, Jianzhuang Liu, Cuihua Li, and Yanyun Qu. Uncertainty-driven dehazing network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 906–913, 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 477 + ], + "type": "text", + "content": "[20] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In European Conference on Computer Vision, pages 163-180. Springer, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 479, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 533 + ], + "type": "text", + "content": "[21] Jie Huang, Zhiwei Xiong, Xueyang Fu, Dong Liu, and Zheng-Jun Zha. Hybrid image enhancement with progressive laplacian enhancing unit. In Proceedings of the 27th ACM International Conference on Multimedia, page 1614-1622, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 535, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 590 + ], + "type": "text", + "content": "[22] Jie Huang, Feng Zhao, Man Zhou, Jie Xiao, Naishan Zheng, Kaiwen Zheng, and Zhiwei Xiong. Learning sample relationship for exposure correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9904-9913, June 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 645 + ], + "type": "text", + "content": "[23] Jie Huang, Man Zhou, Yajing Liu, Mingde Yao, Feng Zhao, and Zhiwei Xiong. Exposure-consistency representation learning for exposure correction. In Proceedings of the 30th ACM International Conference on Multimedia, page 6309-6317, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 713 + ], + "type": "text", + "content": "[24] Jie Huang, Pengfei Zhu, Mingrui Geng, Jiewen Ran, Xingguang Zhou, Chen Xing, Pengfei Wan, and Xiangyang Ji. Range scaling global u-net for perceptual image enhancement on mobile devices. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, September 2018. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13115" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[25] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800–801, 2008. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[26] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "text", + "content": "[27] Kui Jiang, Zhongyuan Wang, Chen Chen, Zheng Wang, Laizhong Cui, and Chia-Wen Lin. Magic elf: Image deraining meets association learning and transformer. arXiv preprint arXiv:2207.10455, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 274 + ], + "type": "text", + "content": "[28] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 2, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 288, + 330 + ], + "type": "text", + "content": "[29] Yeying Jin, Ruoteng Li, Wenhan Yang, and Robby T Tan. Estimating reflectance layer from a single image: Integrating reflectance guidance and shadow/specular aware learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 1069-1077, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[30] Yeying Jin, Beibei Lin, Wending Yan, Wei Ye, Yuan Yuan, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 432 + ], + "type": "text", + "content": "[31] Yeying Jin, Aashish Sharma, and Robby T Tan. Deshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 487 + ], + "type": "text", + "content": "[32] Yeying Jin, Wending Yan, Wenhan Yang, and Robby T Tan. Structure representation network and uncertainty feedback learning for dense non-uniform fog removal. In Proceedings of the Asian Conference on Computer Vision, pages 2041–2058, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 488, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 533 + ], + "type": "text", + "content": "[33] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In European Conference on Computer Vision, pages 404-421. Springer, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 567 + ], + "type": "text", + "content": "[34] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "text", + "content": "[35] Guanbin Li, Xiang He, Wei Zhang, Huiyou Chang, Le Dong, and Liang Lin. Non-locally enhanced encoder-decoder network for single image de-raining. In Proceedings of the 26th ACM international conference on Multimedia, pages 1056–1064, 2018. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[36] Ruoteng Li, Loong-Fah Cheong, and Robby T Tan. Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1633-1642, 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[37] Xia Li, Jianlong Wu, Zhouchen Lin, Hong Liu, and Hongbin Zha. Recurrent squeeze-and-excitation context aggregation net for single image deraining. In Proceedings of the Eu" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ropean conference on computer vision (ECCV), pages 254-269, 2018. 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "text", + "content": "[38] Yu Li, Robby T Tan, Xiaojie Guo, Jiangbo Lu, and Michael S Brown. Rain streak removal using layer priors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2736-2744, 2016. 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "text", + "content": "[39] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "text", + "content": "[40] Yuanchu Liang, Saeed Anwar, and Yang Liu. Drt: A lightweight single image deraining recursive transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 589-598, 2022. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[41] Yun Liu, Zhongsheng Yan, Sixiang Chen, Tian Ye, Wenqi Ren, and Erkang Chen. Nighthazeformer: Single nighttime haze removal using prior query transformer. arXiv preprint arXiv:2305.09533, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "type": "text", + "content": "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[43] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17399-17410, 2022. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[44] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "type": "text", + "content": "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 492, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 548 + ], + "type": "text", + "content": "[46] Kuldeep Purohit, Maitreya Suin, AN Rajagopalan, and Vishnu Naresh Boddeti. Spatially-adaptive image restoration using distortion-guided networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2309-2319, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "type": "text", + "content": "[47] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2482-2491, 2018. 3, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "type": "text", + "content": "[48] Ruijie Quan, Xin Yu, Yuanzhi Liang, and Yi Yang. Removing raindrops and rain streaks in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9147-9156, 2021. 1, 2, 3, 7, 8, 9" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 691 + ], + "type": "text", + "content": "[49] Yuhui Quan, Shijie Deng, Yixin Chen, and Hui Ji. Deep learning for seeing through window with raindrops. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2463-2471, 2019. 1, 3, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[50] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13116" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "better and simpler baseline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3937-3946, 2019. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[52] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. arXiv preprint arXiv:2204.03883, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "text", + "content": "[53] Ming Tong, Yongzhen Wang, Peng Cui, Xuefeng Yan, and Mingqiang Wei. Semi-uformer: Semi-supervised uncertainty-aware transformer for image dehazing. arXiv preprint arXiv:2210.16057, 2022. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 275 + ], + "type": "text", + "content": "[54] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 331 + ], + "type": "text", + "content": "[55] Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson WH Lau. Spatial attentive single-image deraining with a high quality real rain dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12270-12279, 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 376 + ], + "type": "text", + "content": "[56] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "type": "text", + "content": "[57] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 1, 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "type": "text", + "content": "[58] Jie Xiao, Xueyang Fu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Image de-raining transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 3, 7, 8, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "text", + "content": "[59] Ke Xu, Xin Tian, Xin Yang, Baocai Yin, and Rynson WH Lau. Intensity-aware single-image deraining with semantic and color regularization. IEEE Transactions on Image Processing, 30:8497-8509, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[60] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 635 + ], + "type": "text", + "content": "[61] Wenhan Yang, Robby T Tan, Jiashi Feng, Jiaying Liu, Zongming Guo, and Shuicheng Yan. Deep joint rain detection and removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1357-1366, 2017. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[62] Wenhan Yang, Robby T Tan, Jiashi Feng, Shiqi Wang, Bin Cheng, and Jiaying Liu. Recurrent multi-frame deraining: Combining physics guidance and adversarial learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8569-8586, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[63] Rajeev Yasarla and Vishal M Patel. Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 708 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "single image de-raining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8405-8414, 2019. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[64] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[65] Shaodi You, Robby T Tan, Rei Kawakami, and Katsushi Ikeuchi. Adherent raindrop detection and removal in video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1035-1042, 2013. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[66] Hu Yu, Jie Huang, Yajing Liu, Qi Zhu, Man Zhou, and Feng Zhao. Source-free domain adaptation for real-world image dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, page 6645-6654, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[67] Hu Yu, Jie Huang, Yajing Liu, Qi Zhu, Man Zhou, and Feng Zhao. Source-free domain adaptation for real-world image dehazing. In Proceedings of the 30th ACM International Conference on Multimedia, page 6645-6654, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 285, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 350 + ], + "type": "text", + "content": "[68] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 3, 7, 8, 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 352, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 352, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 352, + 545, + 407 + ], + "type": "text", + "content": "[69] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14821-14831, 2021. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 407, + 545, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 407, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 407, + 545, + 461 + ], + "type": "text", + "content": "[70] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 463, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 506 + ], + "type": "text", + "content": "[71] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. arXiv preprint arXiv:2210.01427, 2022. 3, 5, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 508, + 545, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 551 + ], + "type": "text", + "content": "[72] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 552, + 545, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 552, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 307, + 552, + 545, + 606 + ], + "type": "text", + "content": "[73] man zhou, Hu Yu, Jie Huang, Feng Zhao, Jinwei Gu, Chen Change Loy, Deyu Meng, and Chongyi Li. Deep fourier up-sampling. In Advances in Neural Information Processing Systems, volume 35, pages 22995-23008. Curran Associates, Inc., 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 608, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 608, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 307, + 608, + 545, + 662 + ], + "type": "text", + "content": "[74] Lei Zhu, Zijun Deng, Xiaowei Hu, Haoran Xie, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Learning gated non-local residual for single-image rain streak removal. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2147–2159, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 664, + 545, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 664, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 307, + 664, + 545, + 708 + ], + "type": "text", + "content": "[75] Lei Zhu, Chi-Wing Fu, Dani Lischinski, and Pheng-Ann Heng. Joint bi-layer optimization for single-image rain streak removal. In Proceedings of the IEEE international conference on computer vision, pages 2526-2534, 2017. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13117" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/96a859b1-e12e-4281-b4c4-b8ea44fdbf01_content_list.json b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/96a859b1-e12e-4281-b4c4-b8ea44fdbf01_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f4c306c31252a98dfff3068f3b8648a794d95e12 --- /dev/null +++ b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/96a859b1-e12e-4281-b4c4-b8ea44fdbf01_content_list.json @@ -0,0 +1,1608 @@ +[ + { + "type": "text", + "text": "SparseBEV: High-Performance Sparse 3D Object Detection from Multi-Camera Videos", + "text_level": 1, + "bbox": [ + 184, + 130, + 785, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haisong Liu $^{1}$ Yao Teng $^{1}$ Tao Lu $^{1}$ Haiguang Wang $^{1}$ Limin Wang $^{1,2,\\boxtimes}$ $^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University $^{2}$ Shanghai AI Lab $\\{\\mathrm{liuhs, yaoteng, taolu, haiguangwang}\\} \\text{@smail.nju.edu.cn, lmwang@nju.edu.cn}$", + "bbox": [ + 104, + 202, + 862, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camera-based 3D object detection in BEV (Bird's Eye View) space has drawn great attention over the past few years. Dense detectors typically follow a two-stage pipeline by first constructing a dense BEV feature and then performing object detection in BEV space, which suffers from complex view transformations and high computation cost. On the other side, sparse detectors follow a query-based paradigm without explicit dense BEV feature construction, but achieve worse performance than the dense counterparts. In this paper, we find that the key to mitigate this performance gap is the adaptability of the detector in both BEV and image space. To achieve this goal, we propose SparseBEV, a fully sparse 3D object detector that outperforms the dense counterparts. SparseBEV contains three key designs, which are (1) scale-adaptive self attention to aggregate features with adaptive receptive field in BEV space, (2) adaptive spatio-temporal sampling to generate sampling locations under the guidance of queries, and (3) adaptive mixing to decode the sampled features with dynamic weights from the queries. On the test split of nuScenes, SparseBEV achieves the state-of-the-art performance of 67.5 NDS. On the val split, SparseBEV achieves 55.8 NDS while maintaining a real-time inference speed of 23.5 FPS. Code is available at https://github.com/MCG-NJU/SparseBEV.", + "bbox": [ + 75, + 323, + 473, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camera-based 3D Object Detection [13, 54, 31, 11, 24, 25, 40] has witnessed great progress over the past few years. Compared with the LiDAR-based counterparts [19, 56, 4, 36], camera-based approaches have lower deployment cost and can detect long-range objects.", + "bbox": [ + 75, + 755, + 468, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous methods can be divided into two paradigms. BEV (Bird's Eye View)-based methods [13, 11, 25, 24, 40] follow a two-stage pipeline by first constructing an explicit", + "bbox": [ + 75, + 830, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/953c39be16319636c67fcc3abb508b6c57cf75b70b13691028eb273f7ed132af.jpg", + "image_caption": [ + "Figure 1: Performance comparison on the val split of nuScenes [1]. All methods use ResNet50 [10] as the image backbone and the input size is set to $704 \\times 256$ . FPS is measured on a single RTX 3090 with the PyTorch fp32 backend. We balance accuracy and speed by reducing the number of decoder layers without re-training." + ], + "image_footnote": [], + "bbox": [ + 513, + 297, + 859, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dense BEV feature from multi-view features and then performing object detection in BEV space. Although achieving remarkable progress, they suffer from high computation cost and rely on complex view transformation operators. Another line of work [54, 31, 32] explores the sparse query-based paradigm by initializing a set of sparse reference points in 3D space. Specifically, DETR3D [54] links the queries to image features using 3D-to-2D projection. It has simpler structure and faster speed, but its performance still lags far behind the dense ones. PETR series [31, 32] uses dense global attention for the interaction between query and image feature, which is computationally expensive and buries the advantage of the sparse paradigm. Thus, a natural question arises whether fully sparse detectors achieve similar accuracy to the dense ones?", + "bbox": [ + 496, + 626, + 892, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we find that the key to obtain high performance in sparse 3D object detection is the adaptability of the detector in both BEV and image space. In BEV space,", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 99, + 886, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18580", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the detector should be able to aggregate multi-scale features adaptively. Dense BEV-based detectors typically use a BEV encoder to encode multi-scale BEV features. It can be a stack of residual blocks with FPN [26] (e.g. BEVDet [26]), or a transformer encoder with multi-scale deformable attention [60] (e.g. BEVFormer [25]). For sparse detectors such as DETR3D, we argue that the multi-head self attention (MHSA) [49] among queries can play the role of the BEV encoder, as queries are defined in BEV space. However, the vanilla MHSA has a global receptive field, lacking an explicit multi-scale design. In image space, the detector should be adaptive to different objects with different sizes and categories. This is because although the objects have similar sizes in 3D space, they might vary greatly in images. However, the single-point sampling in DETR3D has a fixed local receptive field and the sampled feature is processed by static linear layers, hindering its performance.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we present SparseBEV, a fully sparse 3D object detector that matches or even outperforms the dense counterparts. Our SparseBEV detector contains three key designs, which are (1) scale-adaptive self attention to aggregate features with adaptive receptive field in BEV space, (2) adaptive spatio-temporal sampling to generate sampling locations under the guidance of queries, and (3) adaptive mixing to decode the sampled features with dynamic weights from the queries. We also propose to use pillars instead of reference points as the formulation of query, since pillars introduce better spatial priors.", + "bbox": [ + 75, + 349, + 472, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct comprehensive experiments on the nuScenes dataset. As shown in Fig. 1, our SparseBEV achieves the performance of 55.8 NDS and the speed of 23.5 FPS on the val split, surpassing all previous methods in both speed and accuracy. Besides, we can flexibly adjust of the inference speed by reducing the number of decoder layers without re-training. On test split, SparseBEV with V2-99 [20] backbone achieves 63.6 NDS without using future frames or test-time augmentation. By further utilizing future frames, SparseBEV achieves 67.5 NDS, outperforming the previous state-of-the-art BEVFormerV2 [55] by 2.7 NDS.", + "bbox": [ + 75, + 516, + 472, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 699, + 218, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Query-based 2D Object Detection", + "text_level": 1, + "bbox": [ + 76, + 724, + 372, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Transformer [49] with its attention blocks has been widely applied in the computer vision tasks [5, 8, 2]. In object detection, DETR [2] was the first model to predict objects based on learnable queries and treat the detection as a set prediction problem. A lot of works [60, 45, 30, 37, 7, 21, 57] were then proposed to accelerate the convergence of DETR by using the sampled features instead of using the global ones. For example, Deformable-DETR [60] samples image features based on sparse reference points, and then applies the deformable attention on", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the features. Sparse R-CNN [45] uses the ROIAign [9] to obtain local features and then performs the dynamic convolution. AdaMixer [7] combines the advantages of the sampling points and the dynamic convolution to further boost the convergence. DN-DETR [21] devises a denoising mechanism that feeds ground-truth bounding boxes with noises into decoder and trains the model to reconstruct the original boxes. DINO [57] furtherly optimizes DN-DETR [21] and DAB-DETR [30] by proposing a contrastive denoising training strategy and a mixed query selection method. There are also several methods [3, 14, 46] designed to tackle the instability of the training procedure. Our work follows the query-based detection paradigm and extends it to 3D space with temporal information.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Monocular 3D Object Detection", + "text_level": 1, + "bbox": [ + 498, + 313, + 782, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular 3D object detection takes one single image as input and outputs predicted 3D bounding boxes of objects. A significant challenge in monocular 3D object detection is how to transfer 2D features to 3D space. Several works [53, 43, 51, 39] incorporates depth information to deal with this problem. Pseudo LiDAR [53] first estimates the depth of input images and constructs pseudo point clouds. The pseudo point clouds are then sent to a LiDAR-based detection module to predict the 3D boxes of interested objects. CaDDN [43] further proposes a fully differentiable end-to-end network which learns pixel-wise categorical depth distributions to predict appropriate depth intervals in 3D space. Inspired by FCOS [47], FCOS3D [51] projects 3D coordinates onto 2D images and decouples them as 2D attributes (centerness and classification) and 3D attributes (depths, sizes, and orientations).", + "bbox": [ + 496, + 337, + 893, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. 3D Object Detection in BEV", + "text_level": 1, + "bbox": [ + 498, + 589, + 751, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bird's-eye-view (BEV) object detection [38, 42, 13, 54, 31, 58, 25, 55, 40] aims at detecting objects in BEV space given either single-view or multi-view 2D images. Early works [44, 53, 43] typically transformed 2D features to BEV space based on single-view images and conducted monocular 3D object detection. LSS [42] takes six 2D images of different views as input and transforms them into 3D space based on depth estimation. Based on LSS, BEVDet [13] lifts 2D feature to BEV space and uses a BEV encoder with residual blocks and FPN to further encode the BEV features. BEVFormer [25] proposes a spatiotemporal transformer encoder that projects multi-view and multi-timestamp input to BEV representations. To ease the optimization, BEVFormerV2 [55] introduces perspective view supervision and supervises monocular 3D object detection parallel with BEV object detection. SOLOFusion [40] defines the localization potential for depth estimation and fuses short-term, high-resolution and long-term, low-resolution temporal stereo.", + "bbox": [ + 496, + 613, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18581", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/211b6b7907e2bdcd9c9db511ca352701e967087afea83c289a5349f4c1de0b4b.jpg", + "image_caption": [ + "Figure 2: The overall architecture of SparseBEV, a fully-sparse camera-only 3D object detector. Queries are initialized to be a sparse set of pillars in BEV space. The scale-adaptive self attention further encodes the queries with adaptive receptive fields. Next, multi-view and multi-timestamp features are aggregated with adaptive spatio-temporal sampling and decoded by adaptive mixing. The decoder repeats $L$ times to produce final predictions." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 890, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by DETR, another line of works [54, 31, 32] explore the sparse query-based paradigm. DETR3D [54] proposes a top-down framework starting from a learnable sparse set of reference points and refining them iteratively via 3D-to-2D queries. However, such 3D-to-2D projection hinders the receptive field of the query. To handle this, PETR series [31, 32] use global attention for the interaction between queries and image features, and introduce 3D positional embeddings to encode 2D features into 3D representation without explicit projection. Although achieving notable improvements, the expensive dense global attention buries the advantages of the sparse paradigm and makes it difficult to utilize long-term temporal information efficiently. In contrast, we keep the fully sparse design of DETR3D and boost the performance by enhancing the adaptability of the detector.", + "bbox": [ + 75, + 470, + 472, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SparseBEV", + "text_level": 1, + "bbox": [ + 76, + 724, + 199, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, SparseBEV is a query-based one-stage detector with $L$ decoder layers. The input multicamera videos are processed frame-by-frame using image backbone and FPN. Next, a set of sparse pillar queries are initialized in BEV space and aggregated by scale-adaptive self attention. These queries then interact with the image features via adaptive spatio-temporal sampling and adaptive mixing to make 3D object predictions. We also propose a dual-branch version of SparseBEV to further enhance the temporal modeling.", + "bbox": [ + 75, + 750, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Query Formulation", + "text_level": 1, + "bbox": [ + 500, + 470, + 686, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first define a set of learnable queries, where each of them is represented by its translation $[x,y,z]$ , dimension $[w,l,h]$ , rotation $\\theta$ and velocity $[v_x,v_y]$ . The queries are initialized to be pillars in BEV space where $z$ is set to 0 and $h$ is set to $\\sim 4\\mathrm{m}$ . The initial velocity is set to $[v_x,v_y] = 0$ . Other parameters $(x,y,w,l,\\theta)$ are drawn from random gaussian distributions. Following Sparse R-CNN[45], we attach a $D$ -dim query feature to each query box to encode the rich instance characteristics.", + "bbox": [ + 496, + 493, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Scale-adaptive Self Attention", + "text_level": 1, + "bbox": [ + 500, + 637, + 758, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned above, dense BEV-based methods typically use a BEV encoder to encode multi-scale BEV features. However, in our method, since we do not explicitly build a BEV feature, how to aggregate multi-scale features in BEV space is a challenge.", + "bbox": [ + 496, + 660, + 890, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we argue that the self attention can play the role of BEV encoder, since queries are defined in BEV space. The vanilla multi-head self attention has global receptive field and lacks the ability of local multi-scale aggregation. Thus, we propose scale-adaptive self attention (SASA), which learns appropriate receptive fields under the guidance of queries. First, we compute the all-pair distances $D \\in \\mathbb{R}^{N \\times N}$ ( $N$ is the number of queries) between the query centers in BEV space:", + "bbox": [ + 496, + 736, + 890, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i, j} = \\sqrt {\\left(x _ {i} - x _ {j}\\right) ^ {2} + \\left(y _ {i} - y _ {j}\\right) ^ {2}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 878, + 890, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18582", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75d89e7fb1c30aa594196a0bba42b6c0ae657a3d0714724c0909e50b55e74ff6.jpg", + "image_caption": [ + "Figure 3: Visualization of the learned BEV receptive field in SASA. We choose two samples from the validation split and visualize 4 heads for each sample, denoted by $\\tau_{1},\\tau_{2},\\tau_{3}$ and $\\tau_{4}$ . Queries are represented by their centers." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 460, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{i}$ and $y_{i}$ denotes the center of the $i$ -th query. The attention considers not only the similarity between query features, but the distance between them as well. A toy example below shows how it works:", + "bbox": [ + 75, + 342, + 467, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}} - \\tau D\\right) V, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 415, + 468, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q, K, V \\in \\mathbb{R}^{N \\times d}$ is the query itself and $d$ is the channel dimension. $\\tau$ is a scalar to control the receptive field for each query. When $\\tau = 0$ , it degrades to the vanilla self attention with global receptive field. As $\\tau$ grows, the attention weights for distant queries become smaller and the receptive field narrows.", + "bbox": [ + 75, + 462, + 468, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, the receptive field controller $\\tau$ is adaptive to each query and specific to each head. Supposing there are $H$ heads, we use a linear transformation to generate head-specific $\\tau_{1},\\tau_{2},\\dots,\\tau_{H}$ from the give query $\\mathbf{q}\\in \\mathbb{R}^d$ :", + "bbox": [ + 75, + 554, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\tau_ {1}, \\tau_ {2}, \\dots , \\tau_ {H} \\right] = \\operatorname {L i n e a r} _ {d \\rightarrow H} (\\mathbf {q}) \\in \\mathbb {R} ^ {H}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 627, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the weights are shared across different queries.", + "bbox": [ + 76, + 657, + 429, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our experiments, we surprisingly find that the $\\tau$ for each head is learnt to uniformly distribute in a certain range regardless of the initialization. In Fig. 3, we sort the heads according to $\\tau$ and visualize the attention weights for the distance part. As we can see, each head learns a different receptive field from each other, enabling the network to aggregate features in a multi-scale manner like FPN.", + "bbox": [ + 75, + 672, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scale-adaptive self attention (SASA) demonstrates the necessity of FPN, while being more flexible as it learns the scales adaptively from the query. We also find an interesting phenomenon that different categories of queries have different sizes of receptive field. For example, queries representing the bus have larger receptive field than those representing the pedestrians. More details can be found in the ablation studies.", + "bbox": [ + 75, + 779, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adaptive Spatio-temporal Sampling", + "text_level": 1, + "bbox": [ + 500, + 90, + 812, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each frame, we use a linear layer to generate a set of sampling offsets $\\{(\\Delta x_i,\\Delta y_i,\\Delta z_i)\\}$ adaptively from the query feature. These offsets are transformed to 3D sampling points based on the query pillar:", + "bbox": [ + 496, + 114, + 890, + 175 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l} x _ {i} \\\\ y _ {i} \\\\ z _ {i} \\end{array} \\right] = \\left[ \\begin{array}{c c c} \\cos \\theta & - \\sin \\theta & 0 \\\\ \\sin \\theta & \\cos \\theta & 0 \\\\ 0 & 0 & 1 \\end{array} \\right] \\left[ \\begin{array}{l} w \\cdot \\Delta x _ {i} \\\\ l \\cdot \\Delta y _ {i} \\\\ h \\cdot \\Delta z _ {i} \\end{array} \\right] + \\left[ \\begin{array}{l} x \\\\ y \\\\ z \\end{array} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 196, + 890, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with the deformable attention in BEVFormer, our sampling points are adaptive to both query pillar and query feature, thus better covering objects with varying sizes. Besides, these points are not restricted to the query, since we do not limit the range of the sampling offsets.", + "bbox": [ + 496, + 257, + 890, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we perform temporal alignment by warping the sampling points according to motions. In autonomous driving, there are two types of motion: one is ego-motion and the other is object motion. Ego-motion describes the motion of the car from its own perspective as it navigates through the environment, while object motion refers to the movement of other objects in the environment as they move around the self-driving car.", + "bbox": [ + 496, + 334, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dealing with Object Motion. As mentioned above, instantaneous velocity can be equal to average velocity for a short time window in self-driving. Thus, we adaptively warp the sampling points to previous timestamps using the velocity vector $[v_x, v_y]$ from the query:", + "bbox": [ + 496, + 468, + 890, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t, i} = x _ {i} + v _ {x} \\cdot \\left(T _ {t} - T _ {0}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 556, + 890, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t, i} = y _ {i} + v _ {y} \\cdot \\left(T _ {t} - T _ {0}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 575, + 890, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T_{t}$ denotes the timestamp of previous frame $t$ ( $T_{0}$ denotes the current timestamp). Note that $z_{t,i}$ is identical to $z_{i}$ because the velocity vector is defined in BEV plane.", + "bbox": [ + 496, + 604, + 890, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dealing with Ego Motion. Next, we warp the sampling points based on the ego pose provided by the dataset. Points are first transformed to the global coordinate system and then to the local coordinate system of frame $t$ :", + "bbox": [ + 496, + 662, + 890, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l l l l} x _ {t, i} ^ {\\prime} & y _ {t, i} ^ {\\prime} & z _ {t, i} ^ {\\prime} & 1 \\end{array} \\right] ^ {T} = E _ {t} E _ {0} ^ {- 1} \\left[ \\begin{array}{l l l l} x _ {t, i} & y _ {t, i} & z _ {t, i} & 1 \\end{array} \\right] ^ {T}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 734, + 890, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_{t} = [\\mathbf{R}|\\mathbf{t}]$ is the ego pose of frame $t$ .", + "bbox": [ + 500, + 765, + 795, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sampling. For each timestamp, we project the warped sampling points $\\{(x_{t,i}^{\\prime},y_{t,i}^{\\prime},z_{t,i}^{\\prime})\\}$ to each view using the provided camera intrinsics and extrinsics. Since there are overlaps between adjacent views, the projected point may hit one or more views, which are termed as $\\mathcal{V}$ . For each hit view $k$ , we have multi-scale feature maps $\\{F_{k,j}|j\\in$ $\\{1,2,\\dots,N_{\\mathrm{feat}}\\}\\}$ from the image backbone. Features are", + "bbox": [ + 496, + 794, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18583", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e286ebb37855b9bf7976a45342c21af9caa3a227b5a0c3fc29bd3b1a84f23a14.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneInput SizeEpochsNDSmAPmATEmASEmAOEmAVEmAAE
PETRv2 [32]ResNet50704 × 2566045.634.90.7000.2750.5800.4370.187
BEVStereo [22]ResNet50704 × 25690‡50.037.20.5980.2700.4380.3670.190
BEVPoolv2 [12]ResNet50704 × 25690‡52.640.60.5720.2750.4630.2750.188
SOLOFusion [40]ResNet50704 × 25690‡53.442.70.5670.2740.5110.2520.181
Sparse4Dv2 [29]ResNet50704 × 25610053.943.90.5980.2700.4750.2820.179
StreamPETR † [50]ResNet50704 × 2566055.045.00.6130.2670.4130.2650.196
SparseBEVResNet50704 × 2563654.543.20.6060.2740.3870.2510.186
SparseBEV †ResNet50704 × 2563655.844.80.5810.2710.3730.2470.190
DETR3D † [54]ResNet101-DCN1600 × 9002443.434.90.7160.2680.3790.8420.200
BEVFformer † [25]ResNet101-DCN1600 × 9002451.741.60.6730.2740.3720.3940.198
BEVDepth [24]ResNet1011408 × 51290‡53.541.20.5650.2660.3580.3310.190
Sparse4D † [28]ResNet101-DCN1600 × 9004855.044.40.6030.2760.3600.3090.178
SOLOFusion [40]ResNet1011408 × 51290‡58.248.30.5030.2640.3810.2460.207
SparseBEV †ResNet1011408 × 5122459.250.10.5620.2650.3210.2430.195
", + "bbox": [ + 81, + 88, + 885, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Performance comparison on the nuScenes val split. † benefits from perspective pretraining. ‡ indicates methods with CBGS [59] which will elongate 1 epoch into 4.5 epochs.", + "bbox": [ + 75, + 342, + 890, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "first sampled by bilinear interpolation $\\mathbf{B}$ in the image plane and then weighted over the scale axis:", + "bbox": [ + 75, + 398, + 470, + 430 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t, i} = \\frac {1}{| \\mathcal {V} |} \\sum_ {k \\in \\mathcal {V}} \\sum_ {j = 1} ^ {N _ {\\text {f e a t}}} w _ {i j} \\mathbf {B} \\left(F _ {k, j}, \\mathbf {P} _ {k} \\left(x _ {t, i} ^ {\\prime}, y _ {t, i} ^ {\\prime}, z _ {t, i} ^ {\\prime}\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 444, + 468, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{\\mathrm{feat}}$ is the number of the multi-scale feature maps and $\\mathbf{P}_k$ is the projection function for view $k$ . $w_{ij}$ is the weight for the $i$ -th point on the $j$ -th feature map and is generated from the query feature by linear transformation.", + "bbox": [ + 75, + 500, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Adaptive Mixing", + "text_level": 1, + "bbox": [ + 76, + 573, + 243, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the sampled features from different timestamps and locations, the key is how to adaptively decode it under the guidance of queries. Inspired by AdaMixer [7], we introduce a simple but effective approach to decode and aggregate the spatio-temporal features with dynamic convolutions [15] and MLP-Mixer [48]. Supposing there are $T$ frames in total and $S$ sampling points per frame, we first stack them to a total of $P = T \\times S$ points. Thus, the sampled features are organized to $f \\in \\mathbb{R}^{P \\times C}$ .", + "bbox": [ + 75, + 598, + 468, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Channel Mixing. We first perform channel mixing on $f$ to enhance object semantics. The dynamic weights are generated from the query feature $\\mathbf{q}$ :", + "bbox": [ + 75, + 760, + 468, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {c} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {C \\times C} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 818, + 468, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {c} (f) = \\operatorname {R e L U} \\left(\\text {L a y e r N o r m} \\left(f W _ {c}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 839, + 468, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{c}$ is the dynamic weights and is shared across different frames and different sampling points.", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point Mixing. Next, we then transpose the feature and apply the dynamic weights to the point dimension of it:", + "bbox": [ + 498, + 398, + 890, + 430 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {p} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {P \\times P} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 441, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M} _ {p} (f) = \\operatorname {R e L U} \\left(\\operatorname {L a y e r N o r m} \\left(f ^ {T} W _ {p}\\right)\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 463, + 890, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{p}$ is the dynamic weights and is shared across different channels.", + "bbox": [ + 498, + 496, + 890, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After channel mixing and point mixing, the spatiotemporal features are flattened and aggregated by a linear layer. The final regression and classification predictions are computed by two MLPs respectively.", + "bbox": [ + 498, + 527, + 890, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Dual-branch SparseBEV", + "text_level": 1, + "bbox": [ + 500, + 602, + 725, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by SlowFast [6], we further enhance the long-term temporal modeling by dividing the input video into a slow stream and a fast stream, resulting in a dual-branch SparseBEV. The slow stream is designed to capture fine-grained appearance details and it operates at low frame rates and high resolutions. The fast stream is responsible for capturing long-term temporal stereo and it operates at high frame rates and low resolutions. Sampling points are projected to the two streams respectively and the sampled features are stacked before adaptive mixing.", + "bbox": [ + 496, + 627, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, we decouple the learning of static appearance and temporal motion, leading to better performance. Besides, the computation cost is significantly reduced since only a small fraction of frames needs to be processed at high resolution. However, since this dual-branch design makes the framework a little complex, we do not use it unless otherwise stated. The ablations of this part can be found in the supplementary material.", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18584", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e864dec083ec47f6aa6f5c05f7d64c218bcef5a7376265b2499476a8c13ab2f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneEpochsNDSmAPmATEmASEmAOEmAVEmAAE
DETR3D [54]V2-992447.941.20.6410.2550.3940.8450.133
PETR [31]V2-992450.444.10.5930.2490.3830.8080.132
UVTR [23]V2-992455.147.20.5770.2530.3910.5080.123
BEVFormer [25]V2-992456.948.10.5820.2560.3750.3780.126
BEVDet4D [11]Swin-B [33]90‡56.945.10.5110.2410.3860.3010.121
PolarFormer [16]V2-992457.249.30.5560.2560.3640.4400.127
PETRv2 [32]V2-992459.150.80.5430.2410.3600.3670.118
Sparse4D [28]V2-994859.551.10.5330.2630.3690.3170.124
BEVDepth [24]V2-9990‡60.050.30.4450.2450.3780.3200.126
BEVStereo [22]V2-9990‡61.052.50.4310.2460.3580.3570.138
SOLOFusion [40]ConvNeXt-B [34]90‡61.954.00.4530.2570.3760.2760.148
SparseBEVV2-992462.754.30.5020.2440.3240.2510.126
SparseBEV (dual-branch)V2-992463.655.60.4850.2440.3320.2460.117
BEVFormerV2 † [55]InternImage-XL [52]2464.858.00.4480.2620.3420.2380.128
BEVDet-Gamma † [12]Swin-B90‡66.458.60.3750.2430.3770.1740.123
SparseBEV †V2-992467.560.30.4250.2390.3110.1720.116
", + "bbox": [ + 89, + 88, + 877, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Performance comparison on the nuScenes test split. † uses future frames. ‡ indicates methods with CBGS [59] which will elongate 1 epoch into 4.5 epochs.", + "bbox": [ + 75, + 372, + 892, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 428, + 209, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 453, + 294, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement our model using PyTorch [41]. Following previous methods [54, 25, 32, 40], we adopt common image backbones including ResNet [10] and V2-99 [20]. The decoder consists of 6 layers and weights are shared across different layers. By default, we use $T = 8$ frames in total and the interval between adjacent frames is about 0.5s.", + "bbox": [ + 75, + 476, + 470, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, we adopt the Hungarian algorithm [18] for label assignment between ground-truth objects and predictions. Focal loss [27] is used for classification and L1 loss is used for 3D bounding box regression. We use the AdamW [35] optimizer with a global batch size of 8. The initial learning rate is set to $2 \\times 10^{-4}$ and is decayed with cosine annealing policy.", + "bbox": [ + 75, + 566, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recently, we follow the training setting of the very recent work StreamPETR [50] and refresh our results for fair comparison. For the regression loss, we change the loss weight of $x$ and $y$ to 2.0 and leave the others to 1.0. Query denoising [21] is also introduced to stabilize training and speedup convergence.", + "bbox": [ + 75, + 672, + 470, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 76, + 771, + 276, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our model on the nuScenes dataset [1], which consists of large-scale multimodal data collected from 6 surround-view cameras, 1 lidar and 5 radars. The dataset has 1000 videos and is split into 700/150/150 videos for training/validation/testing. Each video has roughly 20s duration and the key samples are annotated every 0.5s.", + "bbox": [ + 75, + 794, + 468, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For 3D object detection, there are up to 1.4M annotated", + "bbox": [ + 94, + 885, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D bounding boxes of 10 classes. The official evaluation metrics include the well-known mean Average Precision (mAP) and five true positive (TP) metrics, including ATE, ASE, AOE, AVE, and AAE for measuring translation, scale, orientation, velocity, and attribute errors respectively. The overall performance is measured by the nuScenes Detection Score (NDS), which is the composite of the above metrics.", + "bbox": [ + 496, + 430, + 890, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with the State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 498, + 547, + 890, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nuScenes val split. In Tab. 1, we compare SparseBEV with previous state-of-the-art methods on the validation split of nuScenes. Unless otherwise stated, the image backbone is pretrained on ImageNet-1k [17] and the number of queries is set to 900. To keep the simplicity of our approach, the dual-branch design is not used here. When adopting ResNet50 as the backbone and $704 \\times 256$ as the input size, SparseBEV outperforms the previous state-of-the-art method SOLOFusion by $1.2\\mathrm{mAP}$ and 1.4 NDS. By further adopting nuImages [1] pretraining and reducing the number queries to 400, SparseBEV sets a new record of 55.8 NDS while maintaining a real-time inference speed of 23.5 FPS (corresponding to Fig. 1). Next, we upgrade the backbone to ResNet101 and scale the input size up to $1408 \\times 512$ . Under this setting, SparseBEV still surpasses SOLOFusion by $1.8\\mathrm{mAP}$ and 1.0 NDS, demonstrating the scalability of our method.", + "bbox": [ + 496, + 571, + 892, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nuScenes test split. We submit our method to the website of nuScenes and report the leaderboard in Tab. 2. Using the V2-99 [20] backbone pretrained by DD3D [39], SparseBEV achieves 62.7 NDS and 54.3 mAP without bells", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18585", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bce5901fbec8654ac5d645b8bd203c9fca179579b05bfbeb493ff2bdec2909ae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Query FormulationNDSmAP
3D Reference Points55.144.0
BEV Pillars55.645.4
", + "bbox": [ + 151, + 88, + 395, + 150 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0e9208010f4be48d2b32363bf22a9e2eea2f856a9057169c72aaed37b32cc78c.jpg", + "table_caption": [ + "Table 3: Ablations on query formulation. Compared with reference points, pillars introduce better spatial priors and lead to better performance." + ], + "table_footnote": [], + "table_body": "
Self AttentionDistance FunctionNDSmAP
MHSA-53.441.4
SASAτD²54.343.8
τD55.645.4
τ√D55.144.3
", + "bbox": [ + 106, + 220, + 441, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4: Comparison between vanilla multi-head self attention (MHSA) and scale-adaptive self attention (SASA). Our SASA achieves significant improvements over the baseline.", + "bbox": [ + 75, + 330, + 468, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and whistles. Built on top of this, the dual-branch design gives us extra bonus on the performance, hitting 63.6 NDS and $55.6\\mathrm{mAP}$ . We also follow the offline setting of BEVFormerV2 [55] which leverages both past and future frames to assist the detection. Remarkably, our method (single branch only) surpasses BEVFormerV2 by $2.8\\mathrm{mAP}$ and 2.2 NDS. In addition, BEVFormerV2 uses InternImage-XL [52] with over 300M parameters as the image backbone, while our V2-99 backbone is much more lightweight with only $\\sim70\\mathrm{M}$ parameters.", + "bbox": [ + 75, + 398, + 470, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 560, + 240, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablations on the validation split of nuScenes. For all experiments, we use ResNet50 pretrained on nuImages [1] as the image backbone. The input contains 8 frames with $704 \\times 256$ resolution. We use 900 queries in the decoder and the model is trained for 24 epochs without CBGS [59].", + "bbox": [ + 75, + 584, + 468, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query Formulation. In Tab. 3, we ablate different formulations of the query. The top row uses a set of reference points distributing uniformly in 3D space (e.g. DETR3D and PETR). By replacing the reference points with pillars in BEV space, we observe performance improvements (+0.5 NDS) over the baseline. This is because pillars introduce better spatial priors than reference points.", + "bbox": [ + 75, + 689, + 468, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scale-adaptive Self Attention. We study the effect of scale-adaptive self attention (SASA) in Tab. 4. Compared with the vanilla multi-head self attention, SASA achieves $+4.0\\mathrm{mAP}$ and $+2.2$ NDS improvements over the baseline. We further ablate different distance functions and find that $L_{2}$ distance works best. Besides, we also observe two in-", + "bbox": [ + 75, + 809, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edf298f2c349c7188f8c124bb404a9adae09c8daf07c17f325f2d504075e0f59.jpg", + "image_caption": [ + "Averaged $\\tau$ for Each Class", + "Figure 4: Averaged $\\tau$ over all queries and all heads for each class in SASA. Larger $\\tau$ indicates smaller receptive fields. We perform statistics on the val split of nuScenes and choose the queries with a confidence score over 0.3." + ], + "image_footnote": [], + "bbox": [ + 516, + 103, + 877, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac449ef0f77a4cbf3f38495d50b2047548ab5771079cd14d96b33667df1190b8.jpg", + "image_caption": [ + "Figure 5: Ablations of adaptive spatio-temporal sampling. The performance continues to increase with the number of frames. For sampling points, we observe that 16 points per frame works best." + ], + "image_footnote": [], + "bbox": [ + 511, + 344, + 694, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1cf1cbe7a3c755477da5d7dd51134a15623a2a17d31aeade1221d24a7381330f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 345, + 880, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "teresting phenomena in SASA. First, different heads learn a different receptive field (see Fig. 3), enabling the model to aggregate multi-scale features in BEV space. Second, queries representing larger objects tend to have larger receptive field. In Fig. 4, we average the receptive field coefficient $\\tau$ over all queries and all heads for each class. As we can see, the receptive field of large objects (such as bus and truck) is larger than small objects (such as pedestrian and traffic cone), demonstrating the effectiveness of our adaptive design.", + "bbox": [ + 496, + 549, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive Spatio-temporal Sampling. In Fig. 5, we ablate the number of frames and sampling points. The performance continues to improve as the number of frames increases, proving that SparseBEV can benefit from the long-term history. Here, we use 8 frames for fair comparison with previous methods. As for the number of sampling points, we observe that dispatching 16 points for each frame leads to the best performance. We also provide the visualization of the sampling points from the last stage of the decoder in Fig. 6. Our sampling scheme has adaptive regions of interest and achieves good temporal alignments for both static and moving objects across different frames.", + "bbox": [ + 496, + 722, + 890, + 904 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18586", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f2cb444e02fd9c7d2d28a4755408ffc9ef72e02d12d8bacdc537fd1fec2f7a0.jpg", + "image_caption": [ + "Figure 6: Visualization of adaptive spatio-temporal sampling over three consecutive frames. Different instances are distinguished by colors. Point size indices the depth: larger points are closer to the camera. Our sampling scheme has an adaptive receptive field and is well aligned across different timestamps." + ], + "image_footnote": [], + "bbox": [ + 80, + 97, + 888, + 239 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0da7e17d6b3bba106a8128736137a69594e846d300ab85c1e3d3ad07b34635e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ego AlignObj. AlignNDSmAPmAVE
--44.434.10.510
✓-54.243.50.281
✓✓55.645.40.243
", + "bbox": [ + 114, + 310, + 433, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal Alignment. In Tab. 5, we validate the necessity of temporal alignment in spatio-temporal sampling. In our method, ego motion is aligned with the provided ego pose, while object motion is aligned with a simple constant velocity motion model. As we can see from the table, both of them contribute to the performance.", + "bbox": [ + 75, + 448, + 468, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive Mixing. In Tab. 6, we validate the design of the adaptive mixing. The top row is a baseline that uses attention weights to aggregate sampled features (as done in DETR3D). Static mixing and adaptive mixing improve the baseline by 2.7 NDS and 6.5 NDS respectively, demonstrating the necessity of the adaptive design. Next, we explore different combinations of channel and point mixing. Channel mixing followed by point mixing leads to better performance, proving that the object semantics should be enhanced before point mixing.", + "bbox": [ + 75, + 550, + 468, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 76, + 710, + 341, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One limitation of SparseBEV is the heavy reliance on ego pose. As we can see from Tab. 5, the performance drops about 10 NDS without ego-based temporal alignment. However, in the real world, the ego pose provided by IMU may be unreliable and inaccurate, seriously affecting the robustness. Another limitation is that the inference latency increases linearly with the number of frames, since we stack the sampled features along the temporal dimension.", + "bbox": [ + 75, + 734, + 468, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the future, we will explore a more elegant and concise way of decoupling spatial appearance and temporal motion. We will also try to extend SparseBEV to other 3D percep", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/df7335a56fade768da81372acf3d5141a84bb01e8b2b5087a84c9271eac4600b.jpg", + "table_caption": [ + "Table 5: Ablations on temporal alignment. Aligning both ego and object motion leads to the best performance." + ], + "table_footnote": [], + "table_body": "
MethodDetailsNDSmAP
W/o Mixing-49.138.6
Static MixingChannel → Point51.841.9
Adaptive MixingChannel Only53.142.7
Point Only53.643.3
Point → Channel55.444.6
Channel → Point55.645.4
", + "bbox": [ + 522, + 310, + 870, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: Ablations on adaptive mixing. Channel mixing followed by point mixing leads to better performance.", + "bbox": [ + 498, + 449, + 890, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion tasks such as BEV segmentation, occupancy prediction and lane detection.", + "bbox": [ + 498, + 497, + 890, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 540, + 617, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a query-based one-stage 3D object detector, named SparseBEV, which can enjoy the benefits of the BEV space without explicitly constructing a dense BEV feature. SparseBEV improves the adaptability of the decoder by three key modules: scale-adaptive self attention, adaptive spatio-temporal sampling and adaptive mixing. We further introduce a dual-branch design to enhance the long-term temporal modeling. Experiments show that SparseBEV achieves the state-of-the-art performance on the dataset of nuScenes for both speed and accuracy. We hope this exciting result will attract more attention to the fully sparse BEV detection paradigm.", + "bbox": [ + 496, + 565, + 890, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization. Besides, the authors would like to thank Ziteng Gao, Zhiqi Li and Ruopeng Gao for their help.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18587", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1, 6, 7", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European conference on computer vision, pages 213-229, 2020. 2", + "[3] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 2", + "[4] Yukang Chen, Jianhui Liu, Xiaojuan Qi, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Scaling up kernels in 3d cnns. arXiv preprint arXiv:2206.10555, 2022. 1", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[6] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5", + "[7] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 2, 5", + "[8] Meng-Hao Guo, Cheng-Ze Lu, Qibin Hou, Zhengning Liu, Ming-Ming Cheng, and Shi-Min Hu. Segnext: Rethinking convolutional attention design for semantic segmentation. arXiv preprint arXiv:2209.08575, 2022. 2", + "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6", + "[11] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 6", + "[12] Junjie Huang and Guan Huang. Bevpoolv2: A cutting-edge implementation of bevdet toward deployment. arXiv preprint arXiv:2211.17111, 2022. 5, 6", + "[13] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 2" + ], + "bbox": [ + 78, + 116, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 2", + "[15] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 5", + "[16] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformers. arXiv preprint arXiv:2206.15398, 2022. 6", + "[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 6", + "[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6", + "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 1", + "[20] Youngwan Lee and Jongyoul Park. Centermask: Real-time anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 13906-13915, 2020. 2, 6", + "[21] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2, 6", + "[22] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo. arXiv preprint arXiv:2209.10248, 2022. 5, 6", + "[23] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022. 6", + "[24] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 1, 5, 6", + "[25] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision (ECCV), pages 1-18. Springer, 2022. 1, 2, 5, 6", + "[26] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 2", + "[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18588", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 5, 6", + "[29] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018, 2023. 5", + "[30] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2", + "[31] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision (ECCV), pages 531-548. Springer, 2022. 1, 2, 3, 6", + "[32] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 1, 3, 5, 6", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6", + "[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6", + "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6", + "[36] Tao Lu, Xiang Ding, Haisong Liu, Gangshan Wu, and Limin Wang. Link: Linear kernel for lidar-based 3d perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1105-1115, 2023. 1", + "[37] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2", + "[38] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters, 5(3):4867-4873, 2020. 2", + "[39] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3142-3152, 2021. 2, 6", + "[40] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 1, 2, 5, 6", + "[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32:8026-8037, 2019. 6", + "[42] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2", + "[43] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2", + "[44] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. arXiv preprint arXiv:1811.08188, 2018. 2", + "[45] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 2, 3", + "[46] Yao Teng, Haisong Liu, Sheng Guo, and Limin Wang. Stageinteractor: Query-based object detector with cross-stage interaction. arXiv preprint arXiv:2304.04978, 2023. 2", + "[47] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 2", + "[48] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261–24272, 2021. 5", + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[50] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926, 2023. 5, 6", + "[51] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 913-922, 2021. 2", + "[52] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, Xiaogang Wang, and Yu Qiao. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022.6.7", + "[53] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18589", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2", + "[54] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 1, 2, 3, 5, 6", + "[55] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. arXiv preprint arXiv:2211.10439, 2022. 2, 6, 7", + "[56] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1", + "[57] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2", + "[58] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. A simple baseline for multi-camera 3d object detection. arXiv preprint arXiv:2208.10035, 2022. 2", + "[59] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019. 5, 6, 7", + "[60] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2" + ], + "bbox": [ + 78, + 90, + 470, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18590", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/96a859b1-e12e-4281-b4c4-b8ea44fdbf01_model.json b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/96a859b1-e12e-4281-b4c4-b8ea44fdbf01_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ba6986759b944f34bdd7730978e6d042d39f8173 --- /dev/null +++ b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/96a859b1-e12e-4281-b4c4-b8ea44fdbf01_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.185, + 0.131, + 0.787, + 0.175 + ], + "angle": 0, + "content": "SparseBEV: High-Performance Sparse 3D Object Detection from Multi-Camera Videos" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.203, + 0.863, + 0.258 + ], + "angle": 0, + "content": "Haisong Liu\\(^{1}\\) Yao Teng\\(^{1}\\) Tao Lu\\(^{1}\\) Haiguang Wang\\(^{1}\\) Limin Wang\\(^{1,2,\\boxtimes}\\) \n\\(^{1}\\)State Key Laboratory for Novel Software Technology, Nanjing University \\(^{2}\\)Shanghai AI Lab \\(\\{\\mathrm{liuhs, yaoteng, taolu, haiguangwang}\\} \\text{@smail.nju.edu.cn, lmwang@nju.edu.cn}\\)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.701 + ], + "angle": 0, + "content": "Camera-based 3D object detection in BEV (Bird's Eye View) space has drawn great attention over the past few years. Dense detectors typically follow a two-stage pipeline by first constructing a dense BEV feature and then performing object detection in BEV space, which suffers from complex view transformations and high computation cost. On the other side, sparse detectors follow a query-based paradigm without explicit dense BEV feature construction, but achieve worse performance than the dense counterparts. In this paper, we find that the key to mitigate this performance gap is the adaptability of the detector in both BEV and image space. To achieve this goal, we propose SparseBEV, a fully sparse 3D object detector that outperforms the dense counterparts. SparseBEV contains three key designs, which are (1) scale-adaptive self attention to aggregate features with adaptive receptive field in BEV space, (2) adaptive spatio-temporal sampling to generate sampling locations under the guidance of queries, and (3) adaptive mixing to decode the sampled features with dynamic weights from the queries. On the test split of nuScenes, SparseBEV achieves the state-of-the-art performance of 67.5 NDS. On the val split, SparseBEV achieves 55.8 NDS while maintaining a real-time inference speed of 23.5 FPS. Code is available at https://github.com/MCG-NJU/SparseBEV." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.731, + 0.21, + 0.746 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.47, + 0.832 + ], + "angle": 0, + "content": "Camera-based 3D Object Detection [13, 54, 31, 11, 24, 25, 40] has witnessed great progress over the past few years. Compared with the LiDAR-based counterparts [19, 56, 4, 36], camera-based approaches have lower deployment cost and can detect long-range objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.832, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Previous methods can be divided into two paradigms. BEV (Bird's Eye View)-based methods [13, 11, 25, 24, 40] follow a two-stage pipeline by first constructing an explicit" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.299, + 0.86, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.518, + 0.895, + 0.61 + ], + "angle": 0, + "content": "Figure 1: Performance comparison on the val split of nuScenes [1]. All methods use ResNet50 [10] as the image backbone and the input size is set to \\(704 \\times 256\\). FPS is measured on a single RTX 3090 with the PyTorch fp32 backend. We balance accuracy and speed by reducing the number of decoder layers without re-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.893, + 0.853 + ], + "angle": 0, + "content": "dense BEV feature from multi-view features and then performing object detection in BEV space. Although achieving remarkable progress, they suffer from high computation cost and rely on complex view transformation operators. Another line of work [54, 31, 32] explores the sparse query-based paradigm by initializing a set of sparse reference points in 3D space. Specifically, DETR3D [54] links the queries to image features using 3D-to-2D projection. It has simpler structure and faster speed, but its performance still lags far behind the dense ones. PETR series [31, 32] uses dense global attention for the interaction between query and image feature, which is computationally expensive and buries the advantage of the sparse paradigm. Thus, a natural question arises whether fully sparse detectors achieve similar accuracy to the dense ones?" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we find that the key to obtain high performance in sparse 3D object detection is the adaptability of the detector in both BEV and image space. In BEV space," + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.244, + 0.901 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18580" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "the detector should be able to aggregate multi-scale features adaptively. Dense BEV-based detectors typically use a BEV encoder to encode multi-scale BEV features. It can be a stack of residual blocks with FPN [26] (e.g. BEVDet [26]), or a transformer encoder with multi-scale deformable attention [60] (e.g. BEVFormer [25]). For sparse detectors such as DETR3D, we argue that the multi-head self attention (MHSA) [49] among queries can play the role of the BEV encoder, as queries are defined in BEV space. However, the vanilla MHSA has a global receptive field, lacking an explicit multi-scale design. In image space, the detector should be adaptive to different objects with different sizes and categories. This is because although the objects have similar sizes in 3D space, they might vary greatly in images. However, the single-point sampling in DETR3D has a fixed local receptive field and the sampled feature is processed by static linear layers, hindering its performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.473, + 0.516 + ], + "angle": 0, + "content": "To this end, we present SparseBEV, a fully sparse 3D object detector that matches or even outperforms the dense counterparts. Our SparseBEV detector contains three key designs, which are (1) scale-adaptive self attention to aggregate features with adaptive receptive field in BEV space, (2) adaptive spatio-temporal sampling to generate sampling locations under the guidance of queries, and (3) adaptive mixing to decode the sampled features with dynamic weights from the queries. We also propose to use pillars instead of reference points as the formulation of query, since pillars introduce better spatial priors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.473, + 0.684 + ], + "angle": 0, + "content": "We conduct comprehensive experiments on the nuScenes dataset. As shown in Fig. 1, our SparseBEV achieves the performance of 55.8 NDS and the speed of 23.5 FPS on the val split, surpassing all previous methods in both speed and accuracy. Besides, we can flexibly adjust of the inference speed by reducing the number of decoder layers without re-training. On test split, SparseBEV with V2-99 [20] backbone achieves 63.6 NDS without using future frames or test-time augmentation. By further utilizing future frames, SparseBEV achieves 67.5 NDS, outperforming the previous state-of-the-art BEVFormerV2 [55] by 2.7 NDS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.7, + 0.22, + 0.715 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.373, + 0.743 + ], + "angle": 0, + "content": "2.1. Query-based 2D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Recently, Transformer [49] with its attention blocks has been widely applied in the computer vision tasks [5, 8, 2]. In object detection, DETR [2] was the first model to predict objects based on learnable queries and treat the detection as a set prediction problem. A lot of works [60, 45, 30, 37, 7, 21, 57] were then proposed to accelerate the convergence of DETR by using the sampled features instead of using the global ones. For example, Deformable-DETR [60] samples image features based on sparse reference points, and then applies the deformable attention on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "the features. Sparse R-CNN [45] uses the ROIAign [9] to obtain local features and then performs the dynamic convolution. AdaMixer [7] combines the advantages of the sampling points and the dynamic convolution to further boost the convergence. DN-DETR [21] devises a denoising mechanism that feeds ground-truth bounding boxes with noises into decoder and trains the model to reconstruct the original boxes. DINO [57] furtherly optimizes DN-DETR [21] and DAB-DETR [30] by proposing a contrastive denoising training strategy and a mixed query selection method. There are also several methods [3, 14, 46] designed to tackle the instability of the training procedure. Our work follows the query-based detection paradigm and extends it to 3D space with temporal information." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.314, + 0.783, + 0.33 + ], + "angle": 0, + "content": "2.2. Monocular 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.895, + 0.581 + ], + "angle": 0, + "content": "Monocular 3D object detection takes one single image as input and outputs predicted 3D bounding boxes of objects. A significant challenge in monocular 3D object detection is how to transfer 2D features to 3D space. Several works [53, 43, 51, 39] incorporates depth information to deal with this problem. Pseudo LiDAR [53] first estimates the depth of input images and constructs pseudo point clouds. The pseudo point clouds are then sent to a LiDAR-based detection module to predict the 3D boxes of interested objects. CaDDN [43] further proposes a fully differentiable end-to-end network which learns pixel-wise categorical depth distributions to predict appropriate depth intervals in 3D space. Inspired by FCOS [47], FCOS3D [51] projects 3D coordinates onto 2D images and decouples them as 2D attributes (centerness and classification) and 3D attributes (depths, sizes, and orientations)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.59, + 0.753, + 0.607 + ], + "angle": 0, + "content": "2.3. 3D Object Detection in BEV" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Bird's-eye-view (BEV) object detection [38, 42, 13, 54, 31, 58, 25, 55, 40] aims at detecting objects in BEV space given either single-view or multi-view 2D images. Early works [44, 53, 43] typically transformed 2D features to BEV space based on single-view images and conducted monocular 3D object detection. LSS [42] takes six 2D images of different views as input and transforms them into 3D space based on depth estimation. Based on LSS, BEVDet [13] lifts 2D feature to BEV space and uses a BEV encoder with residual blocks and FPN to further encode the BEV features. BEVFormer [25] proposes a spatiotemporal transformer encoder that projects multi-view and multi-timestamp input to BEV representations. To ease the optimization, BEVFormerV2 [55] introduces perspective view supervision and supervises monocular 3D object detection parallel with BEV object detection. SOLOFusion [40] defines the localization potential for depth estimation and fuses short-term, high-resolution and long-term, low-resolution temporal stereo." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18581" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.391, + 0.893, + 0.452 + ], + "angle": 0, + "content": "Figure 2: The overall architecture of SparseBEV, a fully-sparse camera-only 3D object detector. Queries are initialized to be a sparse set of pillars in BEV space. The scale-adaptive self attention further encodes the queries with adaptive receptive fields. Next, multi-view and multi-timestamp features are aggregated with adaptive spatio-temporal sampling and decoded by adaptive mixing. The decoder repeats \\( L \\) times to produce final predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.473, + 0.713 + ], + "angle": 0, + "content": "Inspired by DETR, another line of works [54, 31, 32] explore the sparse query-based paradigm. DETR3D [54] proposes a top-down framework starting from a learnable sparse set of reference points and refining them iteratively via 3D-to-2D queries. However, such 3D-to-2D projection hinders the receptive field of the query. To handle this, PETR series [31, 32] use global attention for the interaction between queries and image features, and introduce 3D positional embeddings to encode 2D features into 3D representation without explicit projection. Although achieving notable improvements, the expensive dense global attention buries the advantages of the sparse paradigm and makes it difficult to utilize long-term temporal information efficiently. In contrast, we keep the fully sparse design of DETR3D and boost the performance by enhancing the adaptability of the detector." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.2, + 0.743 + ], + "angle": 0, + "content": "3. SparseBEV" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 2, SparseBEV is a query-based one-stage detector with \\( L \\) decoder layers. The input multicamera videos are processed frame-by-frame using image backbone and FPN. Next, a set of sparse pillar queries are initialized in BEV space and aggregated by scale-adaptive self attention. These queries then interact with the image features via adaptive spatio-temporal sampling and adaptive mixing to make 3D object predictions. We also propose a dual-branch version of SparseBEV to further enhance the temporal modeling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.687, + 0.487 + ], + "angle": 0, + "content": "3.1. Query Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.63 + ], + "angle": 0, + "content": "We first define a set of learnable queries, where each of them is represented by its translation \\([x,y,z]\\), dimension \\([w,l,h]\\), rotation \\(\\theta\\) and velocity \\([v_x,v_y]\\). The queries are initialized to be pillars in BEV space where \\(z\\) is set to 0 and \\(h\\) is set to \\(\\sim 4\\mathrm{m}\\). The initial velocity is set to \\([v_x,v_y] = 0\\). Other parameters \\((x,y,w,l,\\theta)\\) are drawn from random gaussian distributions. Following Sparse R-CNN[45], we attach a \\(D\\)-dim query feature to each query box to encode the rich instance characteristics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.638, + 0.759, + 0.654 + ], + "angle": 0, + "content": "3.2. Scale-adaptive Self Attention" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.892, + 0.737 + ], + "angle": 0, + "content": "As mentioned above, dense BEV-based methods typically use a BEV encoder to encode multi-scale BEV features. However, in our method, since we do not explicitly build a BEV feature, how to aggregate multi-scale features in BEV space is a challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.737, + 0.892, + 0.873 + ], + "angle": 0, + "content": "In this work, we argue that the self attention can play the role of BEV encoder, since queries are defined in BEV space. The vanilla multi-head self attention has global receptive field and lacks the ability of local multi-scale aggregation. Thus, we propose scale-adaptive self attention (SASA), which learns appropriate receptive fields under the guidance of queries. First, we compute the all-pair distances \\( D \\in \\mathbb{R}^{N \\times N} \\) (\\( N \\) is the number of queries) between the query centers in BEV space:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.88, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\nD _ {i, j} = \\sqrt {\\left(x _ {i} - x _ {j}\\right) ^ {2} + \\left(y _ {i} - y _ {j}\\right) ^ {2}}, \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18582" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.462, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.47, + 0.322 + ], + "angle": 0, + "content": "Figure 3: Visualization of the learned BEV receptive field in SASA. We choose two samples from the validation split and visualize 4 heads for each sample, denoted by \\(\\tau_{1},\\tau_{2},\\tau_{3}\\) and \\(\\tau_{4}\\). Queries are represented by their centers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.468, + 0.404 + ], + "angle": 0, + "content": "where \\( x_{i} \\) and \\( y_{i} \\) denotes the center of the \\( i \\)-th query. The attention considers not only the similarity between query features, but the distance between them as well. A toy example below shows how it works:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.416, + 0.469, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}} - \\tau D\\right) V, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.554 + ], + "angle": 0, + "content": "where \\(Q, K, V \\in \\mathbb{R}^{N \\times d}\\) is the query itself and \\(d\\) is the channel dimension. \\(\\tau\\) is a scalar to control the receptive field for each query. When \\(\\tau = 0\\), it degrades to the vanilla self attention with global receptive field. As \\(\\tau\\) grows, the attention weights for distant queries become smaller and the receptive field narrows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.469, + 0.616 + ], + "angle": 0, + "content": "In practice, the receptive field controller \\(\\tau\\) is adaptive to each query and specific to each head. Supposing there are \\(H\\) heads, we use a linear transformation to generate head-specific \\(\\tau_{1},\\tau_{2},\\dots,\\tau_{H}\\) from the give query \\(\\mathbf{q}\\in \\mathbb{R}^d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.628, + 0.469, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\left[ \\tau_ {1}, \\tau_ {2}, \\dots , \\tau_ {H} \\right] = \\operatorname {L i n e a r} _ {d \\rightarrow H} (\\mathbf {q}) \\in \\mathbb {R} ^ {H}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.431, + 0.673 + ], + "angle": 0, + "content": "where the weights are shared across different queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.779 + ], + "angle": 0, + "content": "In our experiments, we surprisingly find that the \\(\\tau\\) for each head is learnt to uniformly distribute in a certain range regardless of the initialization. In Fig. 3, we sort the heads according to \\(\\tau\\) and visualize the attention weights for the distance part. As we can see, each head learns a different receptive field from each other, enabling the network to aggregate features in a multi-scale manner like FPN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Scale-adaptive self attention (SASA) demonstrates the necessity of FPN, while being more flexible as it learns the scales adaptively from the query. We also find an interesting phenomenon that different categories of queries have different sizes of receptive field. For example, queries representing the bus have larger receptive field than those representing the pedestrians. More details can be found in the ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.813, + 0.108 + ], + "angle": 0, + "content": "3.3. Adaptive Spatio-temporal Sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.176 + ], + "angle": 0, + "content": "For each frame, we use a linear layer to generate a set of sampling offsets \\(\\{(\\Delta x_i,\\Delta y_i,\\Delta z_i)\\}\\) adaptively from the query feature. These offsets are transformed to 3D sampling points based on the query pillar:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.198, + 0.892, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l} x _ {i} \\\\ y _ {i} \\\\ z _ {i} \\end{array} \\right] = \\left[ \\begin{array}{c c c} \\cos \\theta & - \\sin \\theta & 0 \\\\ \\sin \\theta & \\cos \\theta & 0 \\\\ 0 & 0 & 1 \\end{array} \\right] \\left[ \\begin{array}{l} w \\cdot \\Delta x _ {i} \\\\ l \\cdot \\Delta y _ {i} \\\\ h \\cdot \\Delta z _ {i} \\end{array} \\right] + \\left[ \\begin{array}{l} x \\\\ y \\\\ z \\end{array} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Compared with the deformable attention in BEVFormer, our sampling points are adaptive to both query pillar and query feature, thus better covering objects with varying sizes. Besides, these points are not restricted to the query, since we do not limit the range of the sampling offsets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Next, we perform temporal alignment by warping the sampling points according to motions. In autonomous driving, there are two types of motion: one is ego-motion and the other is object motion. Ego-motion describes the motion of the car from its own perspective as it navigates through the environment, while object motion refers to the movement of other objects in the environment as they move around the self-driving car." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.546 + ], + "angle": 0, + "content": "Dealing with Object Motion. As mentioned above, instantaneous velocity can be equal to average velocity for a short time window in self-driving. Thus, we adaptively warp the sampling points to previous timestamps using the velocity vector \\([v_x, v_y]\\) from the query:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.558, + 0.891, + 0.573 + ], + "angle": 0, + "content": "\\[\nx _ {t, i} = x _ {i} + v _ {x} \\cdot \\left(T _ {t} - T _ {0}\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.577, + 0.891, + 0.593 + ], + "angle": 0, + "content": "\\[\ny _ {t, i} = y _ {i} + v _ {y} \\cdot \\left(T _ {t} - T _ {0}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.891, + 0.651 + ], + "angle": 0, + "content": "where \\( T_{t} \\) denotes the timestamp of previous frame \\( t \\) (\\( T_{0} \\) denotes the current timestamp). Note that \\( z_{t,i} \\) is identical to \\( z_{i} \\) because the velocity vector is defined in BEV plane." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.725 + ], + "angle": 0, + "content": "Dealing with Ego Motion. Next, we warp the sampling points based on the ego pose provided by the dataset. Points are first transformed to the global coordinate system and then to the local coordinate system of frame \\( t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.736, + 0.892, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l l l l} x _ {t, i} ^ {\\prime} & y _ {t, i} ^ {\\prime} & z _ {t, i} ^ {\\prime} & 1 \\end{array} \\right] ^ {T} = E _ {t} E _ {0} ^ {- 1} \\left[ \\begin{array}{l l l l} x _ {t, i} & y _ {t, i} & z _ {t, i} & 1 \\end{array} \\right] ^ {T}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.766, + 0.797, + 0.782 + ], + "angle": 0, + "content": "where \\( E_{t} = [\\mathbf{R}|\\mathbf{t}] \\) is the ego pose of frame \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Sampling. For each timestamp, we project the warped sampling points \\(\\{(x_{t,i}^{\\prime},y_{t,i}^{\\prime},z_{t,i}^{\\prime})\\}\\) to each view using the provided camera intrinsics and extrinsics. Since there are overlaps between adjacent views, the projected point may hit one or more views, which are termed as \\(\\mathcal{V}\\). For each hit view \\(k\\), we have multi-scale feature maps \\(\\{F_{k,j}|j\\in\\) \\(\\{1,2,\\dots,N_{\\mathrm{feat}}\\}\\}\\) from the image backbone. Features are" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18583" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.887, + 0.338 + ], + "angle": 0, + "content": "
MethodBackboneInput SizeEpochsNDSmAPmATEmASEmAOEmAVEmAAE
PETRv2 [32]ResNet50704 × 2566045.634.90.7000.2750.5800.4370.187
BEVStereo [22]ResNet50704 × 25690‡50.037.20.5980.2700.4380.3670.190
BEVPoolv2 [12]ResNet50704 × 25690‡52.640.60.5720.2750.4630.2750.188
SOLOFusion [40]ResNet50704 × 25690‡53.442.70.5670.2740.5110.2520.181
Sparse4Dv2 [29]ResNet50704 × 25610053.943.90.5980.2700.4750.2820.179
StreamPETR † [50]ResNet50704 × 2566055.045.00.6130.2670.4130.2650.196
SparseBEVResNet50704 × 2563654.543.20.6060.2740.3870.2510.186
SparseBEV †ResNet50704 × 2563655.844.80.5810.2710.3730.2470.190
DETR3D † [54]ResNet101-DCN1600 × 9002443.434.90.7160.2680.3790.8420.200
BEVFformer † [25]ResNet101-DCN1600 × 9002451.741.60.6730.2740.3720.3940.198
BEVDepth [24]ResNet1011408 × 51290‡53.541.20.5650.2660.3580.3310.190
Sparse4D † [28]ResNet101-DCN1600 × 9004855.044.40.6030.2760.3600.3090.178
SOLOFusion [40]ResNet1011408 × 51290‡58.248.30.5030.2640.3810.2460.207
SparseBEV †ResNet1011408 × 5122459.250.10.5620.2650.3210.2430.195
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.343, + 0.892, + 0.375 + ], + "angle": 0, + "content": "Table 1: Performance comparison on the nuScenes val split. † benefits from perspective pretraining. ‡ indicates methods with CBGS [59] which will elongate 1 epoch into 4.5 epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.431 + ], + "angle": 0, + "content": "first sampled by bilinear interpolation \\(\\mathbf{B}\\) in the image plane and then weighted over the scale axis:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.445, + 0.47, + 0.488 + ], + "angle": 0, + "content": "\\[\nf _ {t, i} = \\frac {1}{| \\mathcal {V} |} \\sum_ {k \\in \\mathcal {V}} \\sum_ {j = 1} ^ {N _ {\\text {f e a t}}} w _ {i j} \\mathbf {B} \\left(F _ {k, j}, \\mathbf {P} _ {k} \\left(x _ {t, i} ^ {\\prime}, y _ {t, i} ^ {\\prime}, z _ {t, i} ^ {\\prime}\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.562 + ], + "angle": 0, + "content": "where \\(N_{\\mathrm{feat}}\\) is the number of the multi-scale feature maps and \\(\\mathbf{P}_k\\) is the projection function for view \\(k\\). \\(w_{ij}\\) is the weight for the \\(i\\)-th point on the \\(j\\)-th feature map and is generated from the query feature by linear transformation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.245, + 0.591 + ], + "angle": 0, + "content": "3.4. Adaptive Mixing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.736 + ], + "angle": 0, + "content": "Given the sampled features from different timestamps and locations, the key is how to adaptively decode it under the guidance of queries. Inspired by AdaMixer [7], we introduce a simple but effective approach to decode and aggregate the spatio-temporal features with dynamic convolutions [15] and MLP-Mixer [48]. Supposing there are \\( T \\) frames in total and \\( S \\) sampling points per frame, we first stack them to a total of \\( P = T \\times S \\) points. Thus, the sampled features are organized to \\( f \\in \\mathbb{R}^{P \\times C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Channel Mixing. We first perform channel mixing on \\( f \\) to enhance object semantics. The dynamic weights are generated from the query feature \\( \\mathbf{q} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.819, + 0.469, + 0.837 + ], + "angle": 0, + "content": "\\[\nW _ {c} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {C \\times C} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.84, + 0.469, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {c} (f) = \\operatorname {R e L U} \\left(\\text {L a y e r N o r m} \\left(f W _ {c}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\( W_{c} \\) is the dynamic weights and is shared across different frames and different sampling points." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.4, + 0.892, + 0.431 + ], + "angle": 0, + "content": "Point Mixing. Next, we then transpose the feature and apply the dynamic weights to the point dimension of it:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.443, + 0.892, + 0.462 + ], + "angle": 0, + "content": "\\[\nW _ {p} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {P \\times P} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.464, + 0.892, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M} _ {p} (f) = \\operatorname {R e L U} \\left(\\operatorname {L a y e r N o r m} \\left(f ^ {T} W _ {p}\\right)\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.497, + 0.891, + 0.527 + ], + "angle": 0, + "content": "where \\( W_{p} \\) is the dynamic weights and is shared across different channels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.529, + 0.892, + 0.59 + ], + "angle": 0, + "content": "After channel mixing and point mixing, the spatiotemporal features are flattened and aggregated by a linear layer. The final regression and classification predictions are computed by two MLPs respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.727, + 0.62 + ], + "angle": 0, + "content": "3.5. Dual-branch SparseBEV" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Inspired by SlowFast [6], we further enhance the long-term temporal modeling by dividing the input video into a slow stream and a fast stream, resulting in a dual-branch SparseBEV. The slow stream is designed to capture fine-grained appearance details and it operates at low frame rates and high resolutions. The fast stream is responsible for capturing long-term temporal stereo and it operates at high frame rates and low resolutions. Sampling points are projected to the two streams respectively and the sampled features are stacked before adaptive mixing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this way, we decouple the learning of static appearance and temporal motion, leading to better performance. Besides, the computation cost is significantly reduced since only a small fraction of frames needs to be processed at high resolution. However, since this dual-branch design makes the framework a little complex, we do not use it unless otherwise stated. The ablations of this part can be found in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18584" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.878, + 0.369 + ], + "angle": 0, + "content": "
MethodBackboneEpochsNDSmAPmATEmASEmAOEmAVEmAAE
DETR3D [54]V2-992447.941.20.6410.2550.3940.8450.133
PETR [31]V2-992450.444.10.5930.2490.3830.8080.132
UVTR [23]V2-992455.147.20.5770.2530.3910.5080.123
BEVFormer [25]V2-992456.948.10.5820.2560.3750.3780.126
BEVDet4D [11]Swin-B [33]90‡56.945.10.5110.2410.3860.3010.121
PolarFormer [16]V2-992457.249.30.5560.2560.3640.4400.127
PETRv2 [32]V2-992459.150.80.5430.2410.3600.3670.118
Sparse4D [28]V2-994859.551.10.5330.2630.3690.3170.124
BEVDepth [24]V2-9990‡60.050.30.4450.2450.3780.3200.126
BEVStereo [22]V2-9990‡61.052.50.4310.2460.3580.3570.138
SOLOFusion [40]ConvNeXt-B [34]90‡61.954.00.4530.2570.3760.2760.148
SparseBEVV2-992462.754.30.5020.2440.3240.2510.126
SparseBEV (dual-branch)V2-992463.655.60.4850.2440.3320.2460.117
BEVFormerV2 † [55]InternImage-XL [52]2464.858.00.4480.2620.3420.2380.128
BEVDet-Gamma † [12]Swin-B90‡66.458.60.3750.2430.3770.1740.123
SparseBEV †V2-992467.560.30.4250.2390.3110.1720.116
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.373, + 0.893, + 0.406 + ], + "angle": 0, + "content": "Table 2: Performance comparison on the nuScenes test split. † uses future frames. ‡ indicates methods with CBGS [59] which will elongate 1 epoch into 4.5 epochs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.429, + 0.21, + 0.447 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.454, + 0.295, + 0.47 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.471, + 0.568 + ], + "angle": 0, + "content": "We implement our model using PyTorch [41]. Following previous methods [54, 25, 32, 40], we adopt common image backbones including ResNet [10] and V2-99 [20]. The decoder consists of 6 layers and weights are shared across different layers. By default, we use \\( T = 8 \\) frames in total and the interval between adjacent frames is about 0.5s." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.674 + ], + "angle": 0, + "content": "During training, we adopt the Hungarian algorithm [18] for label assignment between ground-truth objects and predictions. Focal loss [27] is used for classification and L1 loss is used for 3D bounding box regression. We use the AdamW [35] optimizer with a global batch size of 8. The initial learning rate is set to \\(2 \\times 10^{-4}\\) and is decayed with cosine annealing policy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.764 + ], + "angle": 0, + "content": "Recently, we follow the training setting of the very recent work StreamPETR [50] and refresh our results for fair comparison. For the regression loss, we change the loss weight of \\( x \\) and \\( y \\) to 2.0 and leave the others to 1.0. Query denoising [21] is also introduced to stabilize training and speedup convergence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.277, + 0.787 + ], + "angle": 0, + "content": "4.2. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.886 + ], + "angle": 0, + "content": "We evaluate our model on the nuScenes dataset [1], which consists of large-scale multimodal data collected from 6 surround-view cameras, 1 lidar and 5 radars. The dataset has 1000 videos and is split into 700/150/150 videos for training/validation/testing. Each video has roughly 20s duration and the key samples are annotated every 0.5s." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "For 3D object detection, there are up to 1.4M annotated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.537 + ], + "angle": 0, + "content": "3D bounding boxes of 10 classes. The official evaluation metrics include the well-known mean Average Precision (mAP) and five true positive (TP) metrics, including ATE, ASE, AOE, AVE, and AAE for measuring translation, scale, orientation, velocity, and attribute errors respectively. The overall performance is measured by the nuScenes Detection Score (NDS), which is the composite of the above metrics." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.549, + 0.892, + 0.566 + ], + "angle": 0, + "content": "4.3. Comparison with the State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.893, + 0.83 + ], + "angle": 0, + "content": "nuScenes val split. In Tab. 1, we compare SparseBEV with previous state-of-the-art methods on the validation split of nuScenes. Unless otherwise stated, the image backbone is pretrained on ImageNet-1k [17] and the number of queries is set to 900. To keep the simplicity of our approach, the dual-branch design is not used here. When adopting ResNet50 as the backbone and \\(704 \\times 256\\) as the input size, SparseBEV outperforms the previous state-of-the-art method SOLOFusion by \\(1.2\\mathrm{mAP}\\) and 1.4 NDS. By further adopting nuImages [1] pretraining and reducing the number queries to 400, SparseBEV sets a new record of 55.8 NDS while maintaining a real-time inference speed of 23.5 FPS (corresponding to Fig. 1). Next, we upgrade the backbone to ResNet101 and scale the input size up to \\(1408 \\times 512\\). Under this setting, SparseBEV still surpasses SOLOFusion by \\(1.8\\mathrm{mAP}\\) and 1.0 NDS, demonstrating the scalability of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "nuScenes test split. We submit our method to the website of nuScenes and report the leaderboard in Tab. 2. Using the V2-99 [20] backbone pretrained by DD3D [39], SparseBEV achieves 62.7 NDS and 54.3 mAP without bells" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18585" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.152, + 0.089, + 0.396, + 0.151 + ], + "angle": 0, + "content": "
Query FormulationNDSmAP
3D Reference Points55.144.0
BEV Pillars55.645.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.161, + 0.47, + 0.207 + ], + "angle": 0, + "content": "Table 3: Ablations on query formulation. Compared with reference points, pillars introduce better spatial priors and lead to better performance." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.222, + 0.442, + 0.322 + ], + "angle": 0, + "content": "
Self AttentionDistance FunctionNDSmAP
MHSA-53.441.4
SASAτD²54.343.8
τD55.645.4
τ√D55.144.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.379 + ], + "angle": 0, + "content": "Table 4: Comparison between vanilla multi-head self attention (MHSA) and scale-adaptive self attention (SASA). Our SASA achieves significant improvements over the baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.552 + ], + "angle": 0, + "content": "and whistles. Built on top of this, the dual-branch design gives us extra bonus on the performance, hitting 63.6 NDS and \\(55.6\\mathrm{mAP}\\). We also follow the offline setting of BEVFormerV2 [55] which leverages both past and future frames to assist the detection. Remarkably, our method (single branch only) surpasses BEVFormerV2 by \\(2.8\\mathrm{mAP}\\) and 2.2 NDS. In addition, BEVFormerV2 uses InternImage-XL [52] with over 300M parameters as the image backbone, while our V2-99 backbone is much more lightweight with only \\(\\sim70\\mathrm{M}\\) parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.241, + 0.575 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.676 + ], + "angle": 0, + "content": "In this section, we conduct ablations on the validation split of nuScenes. For all experiments, we use ResNet50 pretrained on nuImages [1] as the image backbone. The input contains 8 frames with \\(704 \\times 256\\) resolution. We use 900 queries in the decoder and the model is trained for 24 epochs without CBGS [59]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.469, + 0.796 + ], + "angle": 0, + "content": "Query Formulation. In Tab. 3, we ablate different formulations of the query. The top row uses a set of reference points distributing uniformly in 3D space (e.g. DETR3D and PETR). By replacing the reference points with pillars in BEV space, we observe performance improvements (+0.5 NDS) over the baseline. This is because pillars introduce better spatial priors than reference points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Scale-adaptive Self Attention. We study the effect of scale-adaptive self attention (SASA) in Tab. 4. Compared with the vanilla multi-head self attention, SASA achieves \\(+4.0\\mathrm{mAP}\\) and \\(+2.2\\) NDS improvements over the baseline. We further ablate different distance functions and find that \\(L_{2}\\) distance works best. Besides, we also observe two in-" + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.091, + 0.776, + 0.103 + ], + "angle": 0, + "content": "Averaged \\(\\tau\\) for Each Class" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.104, + 0.878, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.262, + 0.892, + 0.323 + ], + "angle": 0, + "content": "Figure 4: Averaged \\(\\tau\\) over all queries and all heads for each class in SASA. Larger \\(\\tau\\) indicates smaller receptive fields. We perform statistics on the val split of nuScenes and choose the queries with a confidence score over 0.3." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.345, + 0.695, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.346, + 0.882, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Figure 5: Ablations of adaptive spatio-temporal sampling. The performance continues to increase with the number of frames. For sampling points, we observe that 16 points per frame works best." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.701 + ], + "angle": 0, + "content": "teresting phenomena in SASA. First, different heads learn a different receptive field (see Fig. 3), enabling the model to aggregate multi-scale features in BEV space. Second, queries representing larger objects tend to have larger receptive field. In Fig. 4, we average the receptive field coefficient \\(\\tau\\) over all queries and all heads for each class. As we can see, the receptive field of large objects (such as bus and truck) is larger than small objects (such as pedestrian and traffic cone), demonstrating the effectiveness of our adaptive design." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.892, + 0.905 + ], + "angle": 0, + "content": "Adaptive Spatio-temporal Sampling. In Fig. 5, we ablate the number of frames and sampling points. The performance continues to improve as the number of frames increases, proving that SparseBEV can benefit from the long-term history. Here, we use 8 frames for fair comparison with previous methods. As for the number of sampling points, we observe that dispatching 16 points for each frame leads to the best performance. We also provide the visualization of the sampling points from the last stage of the decoder in Fig. 6. Our sampling scheme has adaptive regions of interest and achieves good temporal alignments for both static and moving objects across different frames." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18586" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.098, + 0.89, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.248, + 0.893, + 0.294 + ], + "angle": 0, + "content": "Figure 6: Visualization of adaptive spatio-temporal sampling over three consecutive frames. Different instances are distinguished by colors. Point size indices the depth: larger points are closer to the camera. Our sampling scheme has an adaptive receptive field and is well aligned across different timestamps." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.311, + 0.434, + 0.389 + ], + "angle": 0, + "content": "
Ego AlignObj. AlignNDSmAPmAVE
--44.434.10.510
✓-54.243.50.281
✓✓55.645.40.243
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.399, + 0.47, + 0.43 + ], + "angle": 0, + "content": "Table 5: Ablations on temporal alignment. Aligning both ego and object motion leads to the best performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.54 + ], + "angle": 0, + "content": "Temporal Alignment. In Tab. 5, we validate the necessity of temporal alignment in spatio-temporal sampling. In our method, ego motion is aligned with the provided ego pose, while object motion is aligned with a simple constant velocity motion model. As we can see from the table, both of them contribute to the performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Adaptive Mixing. In Tab. 6, we validate the design of the adaptive mixing. The top row is a baseline that uses attention weights to aggregate sampled features (as done in DETR3D). Static mixing and adaptive mixing improve the baseline by 2.7 NDS and 6.5 NDS respectively, demonstrating the necessity of the adaptive design. Next, we explore different combinations of channel and point mixing. Channel mixing followed by point mixing leads to better performance, proving that the object semantics should be enhanced before point mixing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.342, + 0.726 + ], + "angle": 0, + "content": "4.5. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.856 + ], + "angle": 0, + "content": "One limitation of SparseBEV is the heavy reliance on ego pose. As we can see from Tab. 5, the performance drops about 10 NDS without ego-based temporal alignment. However, in the real world, the ego pose provided by IMU may be unreliable and inaccurate, seriously affecting the robustness. Another limitation is that the inference latency increases linearly with the number of frames, since we stack the sampled features along the temporal dimension." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In the future, we will explore a more elegant and concise way of decoupling spatial appearance and temporal motion. We will also try to extend SparseBEV to other 3D percep" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.311, + 0.871, + 0.439 + ], + "angle": 0, + "content": "
MethodDetailsNDSmAP
W/o Mixing-49.138.6
Static MixingChannel → Point51.841.9
Adaptive MixingChannel Only53.142.7
Point Only53.643.3
Point → Channel55.444.6
Channel → Point55.645.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.45, + 0.892, + 0.481 + ], + "angle": 0, + "content": "Table 6: Ablations on adaptive mixing. Channel mixing followed by point mixing leads to better performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.892, + 0.527 + ], + "angle": 0, + "content": "tion tasks such as BEV segmentation, occupancy prediction and lane detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.619, + 0.557 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.749 + ], + "angle": 0, + "content": "In this paper, we have proposed a query-based one-stage 3D object detector, named SparseBEV, which can enjoy the benefits of the BEV space without explicitly constructing a dense BEV feature. SparseBEV improves the adaptability of the decoder by three key modules: scale-adaptive self attention, adaptive spatio-temporal sampling and adaptive mixing. We further introduce a dual-branch design to enhance the long-term temporal modeling. Experiments show that SparseBEV achieves the state-of-the-art performance on the dataset of nuScenes for both speed and accuracy. We hope this exciting result will attract more attention to the fully sparse BEV detection paradigm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization. Besides, the authors would like to thank Ziteng Gao, Zhiqi Li and Ruopeng Gao for their help." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18587" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.473, + 0.201 + ], + "angle": 0, + "content": "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.472, + 0.272 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European conference on computer vision, pages 213-229, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.276, + 0.472, + 0.33 + ], + "angle": 0, + "content": "[3] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.333, + 0.472, + 0.376 + ], + "angle": 0, + "content": "[4] Yukang Chen, Jianhui Liu, Xiaojuan Qi, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Scaling up kernels in 3d cnns. arXiv preprint arXiv:2206.10555, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.379, + 0.472, + 0.461 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.464, + 0.472, + 0.52 + ], + "angle": 0, + "content": "[6] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.472, + 0.591 + ], + "angle": 0, + "content": "[7] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.595, + 0.472, + 0.651 + ], + "angle": 0, + "content": "[8] Meng-Hao Guo, Cheng-Ze Lu, Qibin Hou, Zhengning Liu, Ming-Ming Cheng, and Shi-Min Hu. Segnext: Rethinking convolutional attention design for semantic segmentation. arXiv preprint arXiv:2209.08575, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.654, + 0.472, + 0.695 + ], + "angle": 0, + "content": "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.698, + 0.472, + 0.754 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.472, + 0.798 + ], + "angle": 0, + "content": "[11] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.472, + 0.842 + ], + "angle": 0, + "content": "[12] Junjie Huang and Guan Huang. Bevpoolv2: A cutting-edge implementation of bevdet toward deployment. arXiv preprint arXiv:2211.17111, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[13] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.473, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[16] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformers. arXiv preprint arXiv:2206.15398, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.331, + 0.894, + 0.399 + ], + "angle": 0, + "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.894, + 0.455 + ], + "angle": 0, + "content": "[20] Youngwan Lee and Jongyoul Park. Centermask: Real-time anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 13906-13915, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[21] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[22] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo. arXiv preprint arXiv:2209.10248, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[23] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[24] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[25] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision (ECCV), pages 1-18. Springer, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18588" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[29] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[30] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[31] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision (ECCV), pages 531-548. Springer, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[32] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[36] Tao Lu, Xiang Ding, Haisong Liu, Gangshan Wu, and Limin Wang. Link: Linear kernel for lidar-based 3d perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1105-1115, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[37] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[38] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters, 5(3):4867-4873, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[39] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3142-3152, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[40] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32:8026-8037, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[42] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[43] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[44] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. arXiv preprint arXiv:1811.08188, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.894, + 0.413 + ], + "angle": 0, + "content": "[45] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[46] Yao Teng, Haisong Liu, Sheng Guo, and Limin Wang. Stageinteractor: Query-based object detector with cross-stage interaction. arXiv preprint arXiv:2304.04978, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[47] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[48] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261–24272, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[50] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[51] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 913-922, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[52] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, Xiaogang Wang, and Yu Qiao. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022.6.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[53] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18589" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[54] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[55] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. arXiv preprint arXiv:2211.10439, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.472, + 0.346 + ], + "angle": 0, + "content": "[56] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[57] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[58] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. A simple baseline for multi-camera 3d object detection. arXiv preprint arXiv:2208.10035, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[59] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[60] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18590" + } + ] +] \ No newline at end of file diff --git a/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/96a859b1-e12e-4281-b4c4-b8ea44fdbf01_origin.pdf b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/96a859b1-e12e-4281-b4c4-b8ea44fdbf01_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..16c747ac5f33460b05993a388dc0156547cbe1f5 --- /dev/null +++ b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/96a859b1-e12e-4281-b4c4-b8ea44fdbf01_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:301d40fc45b9169c04b44ca61b936b29b8c7305ce443b5dce8bef6fa3bf0ba91 +size 2338491 diff --git a/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/full.md b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ed6fdf175d9df1a413d36bf08b677fa0e46baf4b --- /dev/null +++ b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/full.md @@ -0,0 +1,327 @@ +# SparseBEV: High-Performance Sparse 3D Object Detection from Multi-Camera Videos + +Haisong Liu $^{1}$ Yao Teng $^{1}$ Tao Lu $^{1}$ Haiguang Wang $^{1}$ Limin Wang $^{1,2,\boxtimes}$ $^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University $^{2}$ Shanghai AI Lab $\{\mathrm{liuhs, yaoteng, taolu, haiguangwang}\} \text{@smail.nju.edu.cn, lmwang@nju.edu.cn}$ + +# Abstract + +Camera-based 3D object detection in BEV (Bird's Eye View) space has drawn great attention over the past few years. Dense detectors typically follow a two-stage pipeline by first constructing a dense BEV feature and then performing object detection in BEV space, which suffers from complex view transformations and high computation cost. On the other side, sparse detectors follow a query-based paradigm without explicit dense BEV feature construction, but achieve worse performance than the dense counterparts. In this paper, we find that the key to mitigate this performance gap is the adaptability of the detector in both BEV and image space. To achieve this goal, we propose SparseBEV, a fully sparse 3D object detector that outperforms the dense counterparts. SparseBEV contains three key designs, which are (1) scale-adaptive self attention to aggregate features with adaptive receptive field in BEV space, (2) adaptive spatio-temporal sampling to generate sampling locations under the guidance of queries, and (3) adaptive mixing to decode the sampled features with dynamic weights from the queries. On the test split of nuScenes, SparseBEV achieves the state-of-the-art performance of 67.5 NDS. On the val split, SparseBEV achieves 55.8 NDS while maintaining a real-time inference speed of 23.5 FPS. Code is available at https://github.com/MCG-NJU/SparseBEV. + +# 1. Introduction + +Camera-based 3D Object Detection [13, 54, 31, 11, 24, 25, 40] has witnessed great progress over the past few years. Compared with the LiDAR-based counterparts [19, 56, 4, 36], camera-based approaches have lower deployment cost and can detect long-range objects. + +Previous methods can be divided into two paradigms. BEV (Bird's Eye View)-based methods [13, 11, 25, 24, 40] follow a two-stage pipeline by first constructing an explicit + +![](images/953c39be16319636c67fcc3abb508b6c57cf75b70b13691028eb273f7ed132af.jpg) +Figure 1: Performance comparison on the val split of nuScenes [1]. All methods use ResNet50 [10] as the image backbone and the input size is set to $704 \times 256$ . FPS is measured on a single RTX 3090 with the PyTorch fp32 backend. We balance accuracy and speed by reducing the number of decoder layers without re-training. + +dense BEV feature from multi-view features and then performing object detection in BEV space. Although achieving remarkable progress, they suffer from high computation cost and rely on complex view transformation operators. Another line of work [54, 31, 32] explores the sparse query-based paradigm by initializing a set of sparse reference points in 3D space. Specifically, DETR3D [54] links the queries to image features using 3D-to-2D projection. It has simpler structure and faster speed, but its performance still lags far behind the dense ones. PETR series [31, 32] uses dense global attention for the interaction between query and image feature, which is computationally expensive and buries the advantage of the sparse paradigm. Thus, a natural question arises whether fully sparse detectors achieve similar accuracy to the dense ones? + +In this paper, we find that the key to obtain high performance in sparse 3D object detection is the adaptability of the detector in both BEV and image space. In BEV space, + +the detector should be able to aggregate multi-scale features adaptively. Dense BEV-based detectors typically use a BEV encoder to encode multi-scale BEV features. It can be a stack of residual blocks with FPN [26] (e.g. BEVDet [26]), or a transformer encoder with multi-scale deformable attention [60] (e.g. BEVFormer [25]). For sparse detectors such as DETR3D, we argue that the multi-head self attention (MHSA) [49] among queries can play the role of the BEV encoder, as queries are defined in BEV space. However, the vanilla MHSA has a global receptive field, lacking an explicit multi-scale design. In image space, the detector should be adaptive to different objects with different sizes and categories. This is because although the objects have similar sizes in 3D space, they might vary greatly in images. However, the single-point sampling in DETR3D has a fixed local receptive field and the sampled feature is processed by static linear layers, hindering its performance. + +To this end, we present SparseBEV, a fully sparse 3D object detector that matches or even outperforms the dense counterparts. Our SparseBEV detector contains three key designs, which are (1) scale-adaptive self attention to aggregate features with adaptive receptive field in BEV space, (2) adaptive spatio-temporal sampling to generate sampling locations under the guidance of queries, and (3) adaptive mixing to decode the sampled features with dynamic weights from the queries. We also propose to use pillars instead of reference points as the formulation of query, since pillars introduce better spatial priors. + +We conduct comprehensive experiments on the nuScenes dataset. As shown in Fig. 1, our SparseBEV achieves the performance of 55.8 NDS and the speed of 23.5 FPS on the val split, surpassing all previous methods in both speed and accuracy. Besides, we can flexibly adjust of the inference speed by reducing the number of decoder layers without re-training. On test split, SparseBEV with V2-99 [20] backbone achieves 63.6 NDS without using future frames or test-time augmentation. By further utilizing future frames, SparseBEV achieves 67.5 NDS, outperforming the previous state-of-the-art BEVFormerV2 [55] by 2.7 NDS. + +# 2. Related Work + +# 2.1. Query-based 2D Object Detection + +Recently, Transformer [49] with its attention blocks has been widely applied in the computer vision tasks [5, 8, 2]. In object detection, DETR [2] was the first model to predict objects based on learnable queries and treat the detection as a set prediction problem. A lot of works [60, 45, 30, 37, 7, 21, 57] were then proposed to accelerate the convergence of DETR by using the sampled features instead of using the global ones. For example, Deformable-DETR [60] samples image features based on sparse reference points, and then applies the deformable attention on + +the features. Sparse R-CNN [45] uses the ROIAign [9] to obtain local features and then performs the dynamic convolution. AdaMixer [7] combines the advantages of the sampling points and the dynamic convolution to further boost the convergence. DN-DETR [21] devises a denoising mechanism that feeds ground-truth bounding boxes with noises into decoder and trains the model to reconstruct the original boxes. DINO [57] furtherly optimizes DN-DETR [21] and DAB-DETR [30] by proposing a contrastive denoising training strategy and a mixed query selection method. There are also several methods [3, 14, 46] designed to tackle the instability of the training procedure. Our work follows the query-based detection paradigm and extends it to 3D space with temporal information. + +# 2.2. Monocular 3D Object Detection + +Monocular 3D object detection takes one single image as input and outputs predicted 3D bounding boxes of objects. A significant challenge in monocular 3D object detection is how to transfer 2D features to 3D space. Several works [53, 43, 51, 39] incorporates depth information to deal with this problem. Pseudo LiDAR [53] first estimates the depth of input images and constructs pseudo point clouds. The pseudo point clouds are then sent to a LiDAR-based detection module to predict the 3D boxes of interested objects. CaDDN [43] further proposes a fully differentiable end-to-end network which learns pixel-wise categorical depth distributions to predict appropriate depth intervals in 3D space. Inspired by FCOS [47], FCOS3D [51] projects 3D coordinates onto 2D images and decouples them as 2D attributes (centerness and classification) and 3D attributes (depths, sizes, and orientations). + +# 2.3. 3D Object Detection in BEV + +Bird's-eye-view (BEV) object detection [38, 42, 13, 54, 31, 58, 25, 55, 40] aims at detecting objects in BEV space given either single-view or multi-view 2D images. Early works [44, 53, 43] typically transformed 2D features to BEV space based on single-view images and conducted monocular 3D object detection. LSS [42] takes six 2D images of different views as input and transforms them into 3D space based on depth estimation. Based on LSS, BEVDet [13] lifts 2D feature to BEV space and uses a BEV encoder with residual blocks and FPN to further encode the BEV features. BEVFormer [25] proposes a spatiotemporal transformer encoder that projects multi-view and multi-timestamp input to BEV representations. To ease the optimization, BEVFormerV2 [55] introduces perspective view supervision and supervises monocular 3D object detection parallel with BEV object detection. SOLOFusion [40] defines the localization potential for depth estimation and fuses short-term, high-resolution and long-term, low-resolution temporal stereo. + +![](images/211b6b7907e2bdcd9c9db511ca352701e967087afea83c289a5349f4c1de0b4b.jpg) +Figure 2: The overall architecture of SparseBEV, a fully-sparse camera-only 3D object detector. Queries are initialized to be a sparse set of pillars in BEV space. The scale-adaptive self attention further encodes the queries with adaptive receptive fields. Next, multi-view and multi-timestamp features are aggregated with adaptive spatio-temporal sampling and decoded by adaptive mixing. The decoder repeats $L$ times to produce final predictions. + +Inspired by DETR, another line of works [54, 31, 32] explore the sparse query-based paradigm. DETR3D [54] proposes a top-down framework starting from a learnable sparse set of reference points and refining them iteratively via 3D-to-2D queries. However, such 3D-to-2D projection hinders the receptive field of the query. To handle this, PETR series [31, 32] use global attention for the interaction between queries and image features, and introduce 3D positional embeddings to encode 2D features into 3D representation without explicit projection. Although achieving notable improvements, the expensive dense global attention buries the advantages of the sparse paradigm and makes it difficult to utilize long-term temporal information efficiently. In contrast, we keep the fully sparse design of DETR3D and boost the performance by enhancing the adaptability of the detector. + +# 3. SparseBEV + +As shown in Fig. 2, SparseBEV is a query-based one-stage detector with $L$ decoder layers. The input multicamera videos are processed frame-by-frame using image backbone and FPN. Next, a set of sparse pillar queries are initialized in BEV space and aggregated by scale-adaptive self attention. These queries then interact with the image features via adaptive spatio-temporal sampling and adaptive mixing to make 3D object predictions. We also propose a dual-branch version of SparseBEV to further enhance the temporal modeling. + +# 3.1. Query Formulation + +We first define a set of learnable queries, where each of them is represented by its translation $[x,y,z]$ , dimension $[w,l,h]$ , rotation $\theta$ and velocity $[v_x,v_y]$ . The queries are initialized to be pillars in BEV space where $z$ is set to 0 and $h$ is set to $\sim 4\mathrm{m}$ . The initial velocity is set to $[v_x,v_y] = 0$ . Other parameters $(x,y,w,l,\theta)$ are drawn from random gaussian distributions. Following Sparse R-CNN[45], we attach a $D$ -dim query feature to each query box to encode the rich instance characteristics. + +# 3.2. Scale-adaptive Self Attention + +As mentioned above, dense BEV-based methods typically use a BEV encoder to encode multi-scale BEV features. However, in our method, since we do not explicitly build a BEV feature, how to aggregate multi-scale features in BEV space is a challenge. + +In this work, we argue that the self attention can play the role of BEV encoder, since queries are defined in BEV space. The vanilla multi-head self attention has global receptive field and lacks the ability of local multi-scale aggregation. Thus, we propose scale-adaptive self attention (SASA), which learns appropriate receptive fields under the guidance of queries. First, we compute the all-pair distances $D \in \mathbb{R}^{N \times N}$ ( $N$ is the number of queries) between the query centers in BEV space: + +$$ +D _ {i, j} = \sqrt {\left(x _ {i} - x _ {j}\right) ^ {2} + \left(y _ {i} - y _ {j}\right) ^ {2}}, \tag {1} +$$ + +![](images/75d89e7fb1c30aa594196a0bba42b6c0ae657a3d0714724c0909e50b55e74ff6.jpg) +Figure 3: Visualization of the learned BEV receptive field in SASA. We choose two samples from the validation split and visualize 4 heads for each sample, denoted by $\tau_{1},\tau_{2},\tau_{3}$ and $\tau_{4}$ . Queries are represented by their centers. + +where $x_{i}$ and $y_{i}$ denotes the center of the $i$ -th query. The attention considers not only the similarity between query features, but the distance between them as well. A toy example below shows how it works: + +$$ +\operatorname {A t t n} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}} - \tau D\right) V, \tag {2} +$$ + +where $Q, K, V \in \mathbb{R}^{N \times d}$ is the query itself and $d$ is the channel dimension. $\tau$ is a scalar to control the receptive field for each query. When $\tau = 0$ , it degrades to the vanilla self attention with global receptive field. As $\tau$ grows, the attention weights for distant queries become smaller and the receptive field narrows. + +In practice, the receptive field controller $\tau$ is adaptive to each query and specific to each head. Supposing there are $H$ heads, we use a linear transformation to generate head-specific $\tau_{1},\tau_{2},\dots,\tau_{H}$ from the give query $\mathbf{q}\in \mathbb{R}^d$ : + +$$ +\left[ \tau_ {1}, \tau_ {2}, \dots , \tau_ {H} \right] = \operatorname {L i n e a r} _ {d \rightarrow H} (\mathbf {q}) \in \mathbb {R} ^ {H}, \tag {3} +$$ + +where the weights are shared across different queries. + +In our experiments, we surprisingly find that the $\tau$ for each head is learnt to uniformly distribute in a certain range regardless of the initialization. In Fig. 3, we sort the heads according to $\tau$ and visualize the attention weights for the distance part. As we can see, each head learns a different receptive field from each other, enabling the network to aggregate features in a multi-scale manner like FPN. + +Scale-adaptive self attention (SASA) demonstrates the necessity of FPN, while being more flexible as it learns the scales adaptively from the query. We also find an interesting phenomenon that different categories of queries have different sizes of receptive field. For example, queries representing the bus have larger receptive field than those representing the pedestrians. More details can be found in the ablation studies. + +# 3.3. Adaptive Spatio-temporal Sampling + +For each frame, we use a linear layer to generate a set of sampling offsets $\{(\Delta x_i,\Delta y_i,\Delta z_i)\}$ adaptively from the query feature. These offsets are transformed to 3D sampling points based on the query pillar: + +$$ +\left[ \begin{array}{l} x _ {i} \\ y _ {i} \\ z _ {i} \end{array} \right] = \left[ \begin{array}{c c c} \cos \theta & - \sin \theta & 0 \\ \sin \theta & \cos \theta & 0 \\ 0 & 0 & 1 \end{array} \right] \left[ \begin{array}{l} w \cdot \Delta x _ {i} \\ l \cdot \Delta y _ {i} \\ h \cdot \Delta z _ {i} \end{array} \right] + \left[ \begin{array}{l} x \\ y \\ z \end{array} \right]. \tag {4} +$$ + +Compared with the deformable attention in BEVFormer, our sampling points are adaptive to both query pillar and query feature, thus better covering objects with varying sizes. Besides, these points are not restricted to the query, since we do not limit the range of the sampling offsets. + +Next, we perform temporal alignment by warping the sampling points according to motions. In autonomous driving, there are two types of motion: one is ego-motion and the other is object motion. Ego-motion describes the motion of the car from its own perspective as it navigates through the environment, while object motion refers to the movement of other objects in the environment as they move around the self-driving car. + +Dealing with Object Motion. As mentioned above, instantaneous velocity can be equal to average velocity for a short time window in self-driving. Thus, we adaptively warp the sampling points to previous timestamps using the velocity vector $[v_x, v_y]$ from the query: + +$$ +x _ {t, i} = x _ {i} + v _ {x} \cdot \left(T _ {t} - T _ {0}\right) \tag {5} +$$ + +$$ +y _ {t, i} = y _ {i} + v _ {y} \cdot \left(T _ {t} - T _ {0}\right), \tag {6} +$$ + +where $T_{t}$ denotes the timestamp of previous frame $t$ ( $T_{0}$ denotes the current timestamp). Note that $z_{t,i}$ is identical to $z_{i}$ because the velocity vector is defined in BEV plane. + +Dealing with Ego Motion. Next, we warp the sampling points based on the ego pose provided by the dataset. Points are first transformed to the global coordinate system and then to the local coordinate system of frame $t$ : + +$$ +\left[ \begin{array}{l l l l} x _ {t, i} ^ {\prime} & y _ {t, i} ^ {\prime} & z _ {t, i} ^ {\prime} & 1 \end{array} \right] ^ {T} = E _ {t} E _ {0} ^ {- 1} \left[ \begin{array}{l l l l} x _ {t, i} & y _ {t, i} & z _ {t, i} & 1 \end{array} \right] ^ {T}, \tag {7} +$$ + +where $E_{t} = [\mathbf{R}|\mathbf{t}]$ is the ego pose of frame $t$ . + +Sampling. For each timestamp, we project the warped sampling points $\{(x_{t,i}^{\prime},y_{t,i}^{\prime},z_{t,i}^{\prime})\}$ to each view using the provided camera intrinsics and extrinsics. Since there are overlaps between adjacent views, the projected point may hit one or more views, which are termed as $\mathcal{V}$ . For each hit view $k$ , we have multi-scale feature maps $\{F_{k,j}|j\in$ $\{1,2,\dots,N_{\mathrm{feat}}\}\}$ from the image backbone. Features are + +
MethodBackboneInput SizeEpochsNDSmAPmATEmASEmAOEmAVEmAAE
PETRv2 [32]ResNet50704 × 2566045.634.90.7000.2750.5800.4370.187
BEVStereo [22]ResNet50704 × 25690‡50.037.20.5980.2700.4380.3670.190
BEVPoolv2 [12]ResNet50704 × 25690‡52.640.60.5720.2750.4630.2750.188
SOLOFusion [40]ResNet50704 × 25690‡53.442.70.5670.2740.5110.2520.181
Sparse4Dv2 [29]ResNet50704 × 25610053.943.90.5980.2700.4750.2820.179
StreamPETR † [50]ResNet50704 × 2566055.045.00.6130.2670.4130.2650.196
SparseBEVResNet50704 × 2563654.543.20.6060.2740.3870.2510.186
SparseBEV †ResNet50704 × 2563655.844.80.5810.2710.3730.2470.190
DETR3D † [54]ResNet101-DCN1600 × 9002443.434.90.7160.2680.3790.8420.200
BEVFformer † [25]ResNet101-DCN1600 × 9002451.741.60.6730.2740.3720.3940.198
BEVDepth [24]ResNet1011408 × 51290‡53.541.20.5650.2660.3580.3310.190
Sparse4D † [28]ResNet101-DCN1600 × 9004855.044.40.6030.2760.3600.3090.178
SOLOFusion [40]ResNet1011408 × 51290‡58.248.30.5030.2640.3810.2460.207
SparseBEV †ResNet1011408 × 5122459.250.10.5620.2650.3210.2430.195
+ +Table 1: Performance comparison on the nuScenes val split. † benefits from perspective pretraining. ‡ indicates methods with CBGS [59] which will elongate 1 epoch into 4.5 epochs. + +first sampled by bilinear interpolation $\mathbf{B}$ in the image plane and then weighted over the scale axis: + +$$ +f _ {t, i} = \frac {1}{| \mathcal {V} |} \sum_ {k \in \mathcal {V}} \sum_ {j = 1} ^ {N _ {\text {f e a t}}} w _ {i j} \mathbf {B} \left(F _ {k, j}, \mathbf {P} _ {k} \left(x _ {t, i} ^ {\prime}, y _ {t, i} ^ {\prime}, z _ {t, i} ^ {\prime}\right)\right) \tag {8} +$$ + +where $N_{\mathrm{feat}}$ is the number of the multi-scale feature maps and $\mathbf{P}_k$ is the projection function for view $k$ . $w_{ij}$ is the weight for the $i$ -th point on the $j$ -th feature map and is generated from the query feature by linear transformation. + +# 3.4. Adaptive Mixing + +Given the sampled features from different timestamps and locations, the key is how to adaptively decode it under the guidance of queries. Inspired by AdaMixer [7], we introduce a simple but effective approach to decode and aggregate the spatio-temporal features with dynamic convolutions [15] and MLP-Mixer [48]. Supposing there are $T$ frames in total and $S$ sampling points per frame, we first stack them to a total of $P = T \times S$ points. Thus, the sampled features are organized to $f \in \mathbb{R}^{P \times C}$ . + +Channel Mixing. We first perform channel mixing on $f$ to enhance object semantics. The dynamic weights are generated from the query feature $\mathbf{q}$ : + +$$ +W _ {c} = \operatorname {L i n e a r} (\mathbf {q}) \in \mathbb {R} ^ {C \times C} \tag {9} +$$ + +$$ +\mathbf {M} _ {c} (f) = \operatorname {R e L U} \left(\text {L a y e r N o r m} \left(f W _ {c}\right)\right), \tag {10} +$$ + +where $W_{c}$ is the dynamic weights and is shared across different frames and different sampling points. + +Point Mixing. Next, we then transpose the feature and apply the dynamic weights to the point dimension of it: + +$$ +W _ {p} = \operatorname {L i n e a r} (\mathbf {q}) \in \mathbb {R} ^ {P \times P} \tag {11} +$$ + +$$ +\mathrm {M} _ {p} (f) = \operatorname {R e L U} \left(\operatorname {L a y e r N o r m} \left(f ^ {T} W _ {p}\right)\right), \tag {12} +$$ + +where $W_{p}$ is the dynamic weights and is shared across different channels. + +After channel mixing and point mixing, the spatiotemporal features are flattened and aggregated by a linear layer. The final regression and classification predictions are computed by two MLPs respectively. + +# 3.5. Dual-branch SparseBEV + +Inspired by SlowFast [6], we further enhance the long-term temporal modeling by dividing the input video into a slow stream and a fast stream, resulting in a dual-branch SparseBEV. The slow stream is designed to capture fine-grained appearance details and it operates at low frame rates and high resolutions. The fast stream is responsible for capturing long-term temporal stereo and it operates at high frame rates and low resolutions. Sampling points are projected to the two streams respectively and the sampled features are stacked before adaptive mixing. + +In this way, we decouple the learning of static appearance and temporal motion, leading to better performance. Besides, the computation cost is significantly reduced since only a small fraction of frames needs to be processed at high resolution. However, since this dual-branch design makes the framework a little complex, we do not use it unless otherwise stated. The ablations of this part can be found in the supplementary material. + +
MethodBackboneEpochsNDSmAPmATEmASEmAOEmAVEmAAE
DETR3D [54]V2-992447.941.20.6410.2550.3940.8450.133
PETR [31]V2-992450.444.10.5930.2490.3830.8080.132
UVTR [23]V2-992455.147.20.5770.2530.3910.5080.123
BEVFormer [25]V2-992456.948.10.5820.2560.3750.3780.126
BEVDet4D [11]Swin-B [33]90‡56.945.10.5110.2410.3860.3010.121
PolarFormer [16]V2-992457.249.30.5560.2560.3640.4400.127
PETRv2 [32]V2-992459.150.80.5430.2410.3600.3670.118
Sparse4D [28]V2-994859.551.10.5330.2630.3690.3170.124
BEVDepth [24]V2-9990‡60.050.30.4450.2450.3780.3200.126
BEVStereo [22]V2-9990‡61.052.50.4310.2460.3580.3570.138
SOLOFusion [40]ConvNeXt-B [34]90‡61.954.00.4530.2570.3760.2760.148
SparseBEVV2-992462.754.30.5020.2440.3240.2510.126
SparseBEV (dual-branch)V2-992463.655.60.4850.2440.3320.2460.117
BEVFormerV2 † [55]InternImage-XL [52]2464.858.00.4480.2620.3420.2380.128
BEVDet-Gamma † [12]Swin-B90‡66.458.60.3750.2430.3770.1740.123
SparseBEV †V2-992467.560.30.4250.2390.3110.1720.116
+ +Table 2: Performance comparison on the nuScenes test split. † uses future frames. ‡ indicates methods with CBGS [59] which will elongate 1 epoch into 4.5 epochs. + +# 4. Experiments + +# 4.1. Implementation Details + +We implement our model using PyTorch [41]. Following previous methods [54, 25, 32, 40], we adopt common image backbones including ResNet [10] and V2-99 [20]. The decoder consists of 6 layers and weights are shared across different layers. By default, we use $T = 8$ frames in total and the interval between adjacent frames is about 0.5s. + +During training, we adopt the Hungarian algorithm [18] for label assignment between ground-truth objects and predictions. Focal loss [27] is used for classification and L1 loss is used for 3D bounding box regression. We use the AdamW [35] optimizer with a global batch size of 8. The initial learning rate is set to $2 \times 10^{-4}$ and is decayed with cosine annealing policy. + +Recently, we follow the training setting of the very recent work StreamPETR [50] and refresh our results for fair comparison. For the regression loss, we change the loss weight of $x$ and $y$ to 2.0 and leave the others to 1.0. Query denoising [21] is also introduced to stabilize training and speedup convergence. + +# 4.2. Datasets and Metrics + +We evaluate our model on the nuScenes dataset [1], which consists of large-scale multimodal data collected from 6 surround-view cameras, 1 lidar and 5 radars. The dataset has 1000 videos and is split into 700/150/150 videos for training/validation/testing. Each video has roughly 20s duration and the key samples are annotated every 0.5s. + +For 3D object detection, there are up to 1.4M annotated + +3D bounding boxes of 10 classes. The official evaluation metrics include the well-known mean Average Precision (mAP) and five true positive (TP) metrics, including ATE, ASE, AOE, AVE, and AAE for measuring translation, scale, orientation, velocity, and attribute errors respectively. The overall performance is measured by the nuScenes Detection Score (NDS), which is the composite of the above metrics. + +# 4.3. Comparison with the State-of-the-art Methods + +nuScenes val split. In Tab. 1, we compare SparseBEV with previous state-of-the-art methods on the validation split of nuScenes. Unless otherwise stated, the image backbone is pretrained on ImageNet-1k [17] and the number of queries is set to 900. To keep the simplicity of our approach, the dual-branch design is not used here. When adopting ResNet50 as the backbone and $704 \times 256$ as the input size, SparseBEV outperforms the previous state-of-the-art method SOLOFusion by $1.2\mathrm{mAP}$ and 1.4 NDS. By further adopting nuImages [1] pretraining and reducing the number queries to 400, SparseBEV sets a new record of 55.8 NDS while maintaining a real-time inference speed of 23.5 FPS (corresponding to Fig. 1). Next, we upgrade the backbone to ResNet101 and scale the input size up to $1408 \times 512$ . Under this setting, SparseBEV still surpasses SOLOFusion by $1.8\mathrm{mAP}$ and 1.0 NDS, demonstrating the scalability of our method. + +nuScenes test split. We submit our method to the website of nuScenes and report the leaderboard in Tab. 2. Using the V2-99 [20] backbone pretrained by DD3D [39], SparseBEV achieves 62.7 NDS and 54.3 mAP without bells + +
Query FormulationNDSmAP
3D Reference Points55.144.0
BEV Pillars55.645.4
+ +Table 3: Ablations on query formulation. Compared with reference points, pillars introduce better spatial priors and lead to better performance. + +
Self AttentionDistance FunctionNDSmAP
MHSA-53.441.4
SASAτD²54.343.8
τD55.645.4
τ√D55.144.3
+ +Table 4: Comparison between vanilla multi-head self attention (MHSA) and scale-adaptive self attention (SASA). Our SASA achieves significant improvements over the baseline. + +and whistles. Built on top of this, the dual-branch design gives us extra bonus on the performance, hitting 63.6 NDS and $55.6\mathrm{mAP}$ . We also follow the offline setting of BEVFormerV2 [55] which leverages both past and future frames to assist the detection. Remarkably, our method (single branch only) surpasses BEVFormerV2 by $2.8\mathrm{mAP}$ and 2.2 NDS. In addition, BEVFormerV2 uses InternImage-XL [52] with over 300M parameters as the image backbone, while our V2-99 backbone is much more lightweight with only $\sim70\mathrm{M}$ parameters. + +# 4.4. Ablation Studies + +In this section, we conduct ablations on the validation split of nuScenes. For all experiments, we use ResNet50 pretrained on nuImages [1] as the image backbone. The input contains 8 frames with $704 \times 256$ resolution. We use 900 queries in the decoder and the model is trained for 24 epochs without CBGS [59]. + +Query Formulation. In Tab. 3, we ablate different formulations of the query. The top row uses a set of reference points distributing uniformly in 3D space (e.g. DETR3D and PETR). By replacing the reference points with pillars in BEV space, we observe performance improvements (+0.5 NDS) over the baseline. This is because pillars introduce better spatial priors than reference points. + +Scale-adaptive Self Attention. We study the effect of scale-adaptive self attention (SASA) in Tab. 4. Compared with the vanilla multi-head self attention, SASA achieves $+4.0\mathrm{mAP}$ and $+2.2$ NDS improvements over the baseline. We further ablate different distance functions and find that $L_{2}$ distance works best. Besides, we also observe two in- + +![](images/edf298f2c349c7188f8c124bb404a9adae09c8daf07c17f325f2d504075e0f59.jpg) +Averaged $\tau$ for Each Class +Figure 4: Averaged $\tau$ over all queries and all heads for each class in SASA. Larger $\tau$ indicates smaller receptive fields. We perform statistics on the val split of nuScenes and choose the queries with a confidence score over 0.3. + +![](images/ac449ef0f77a4cbf3f38495d50b2047548ab5771079cd14d96b33667df1190b8.jpg) +Figure 5: Ablations of adaptive spatio-temporal sampling. The performance continues to increase with the number of frames. For sampling points, we observe that 16 points per frame works best. + +![](images/1cf1cbe7a3c755477da5d7dd51134a15623a2a17d31aeade1221d24a7381330f.jpg) + +teresting phenomena in SASA. First, different heads learn a different receptive field (see Fig. 3), enabling the model to aggregate multi-scale features in BEV space. Second, queries representing larger objects tend to have larger receptive field. In Fig. 4, we average the receptive field coefficient $\tau$ over all queries and all heads for each class. As we can see, the receptive field of large objects (such as bus and truck) is larger than small objects (such as pedestrian and traffic cone), demonstrating the effectiveness of our adaptive design. + +Adaptive Spatio-temporal Sampling. In Fig. 5, we ablate the number of frames and sampling points. The performance continues to improve as the number of frames increases, proving that SparseBEV can benefit from the long-term history. Here, we use 8 frames for fair comparison with previous methods. As for the number of sampling points, we observe that dispatching 16 points for each frame leads to the best performance. We also provide the visualization of the sampling points from the last stage of the decoder in Fig. 6. Our sampling scheme has adaptive regions of interest and achieves good temporal alignments for both static and moving objects across different frames. + +![](images/9f2cb444e02fd9c7d2d28a4755408ffc9ef72e02d12d8bacdc537fd1fec2f7a0.jpg) +Figure 6: Visualization of adaptive spatio-temporal sampling over three consecutive frames. Different instances are distinguished by colors. Point size indices the depth: larger points are closer to the camera. Our sampling scheme has an adaptive receptive field and is well aligned across different timestamps. + +
Ego AlignObj. AlignNDSmAPmAVE
--44.434.10.510
✓-54.243.50.281
✓✓55.645.40.243
+ +Temporal Alignment. In Tab. 5, we validate the necessity of temporal alignment in spatio-temporal sampling. In our method, ego motion is aligned with the provided ego pose, while object motion is aligned with a simple constant velocity motion model. As we can see from the table, both of them contribute to the performance. + +Adaptive Mixing. In Tab. 6, we validate the design of the adaptive mixing. The top row is a baseline that uses attention weights to aggregate sampled features (as done in DETR3D). Static mixing and adaptive mixing improve the baseline by 2.7 NDS and 6.5 NDS respectively, demonstrating the necessity of the adaptive design. Next, we explore different combinations of channel and point mixing. Channel mixing followed by point mixing leads to better performance, proving that the object semantics should be enhanced before point mixing. + +# 4.5. Limitations and Future Work + +One limitation of SparseBEV is the heavy reliance on ego pose. As we can see from Tab. 5, the performance drops about 10 NDS without ego-based temporal alignment. However, in the real world, the ego pose provided by IMU may be unreliable and inaccurate, seriously affecting the robustness. Another limitation is that the inference latency increases linearly with the number of frames, since we stack the sampled features along the temporal dimension. + +In the future, we will explore a more elegant and concise way of decoupling spatial appearance and temporal motion. We will also try to extend SparseBEV to other 3D percep + +Table 5: Ablations on temporal alignment. Aligning both ego and object motion leads to the best performance. + +
MethodDetailsNDSmAP
W/o Mixing-49.138.6
Static MixingChannel → Point51.841.9
Adaptive MixingChannel Only53.142.7
Point Only53.643.3
Point → Channel55.444.6
Channel → Point55.645.4
+ +Table 6: Ablations on adaptive mixing. Channel mixing followed by point mixing leads to better performance. + +tion tasks such as BEV segmentation, occupancy prediction and lane detection. + +# 5. Conclusion + +In this paper, we have proposed a query-based one-stage 3D object detector, named SparseBEV, which can enjoy the benefits of the BEV space without explicitly constructing a dense BEV feature. SparseBEV improves the adaptability of the decoder by three key modules: scale-adaptive self attention, adaptive spatio-temporal sampling and adaptive mixing. We further introduce a dual-branch design to enhance the long-term temporal modeling. Experiments show that SparseBEV achieves the state-of-the-art performance on the dataset of nuScenes for both speed and accuracy. We hope this exciting result will attract more attention to the fully sparse BEV detection paradigm. + +Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization. Besides, the authors would like to thank Ziteng Gao, Zhiqi Li and Ruopeng Gao for their help. + +# References + +[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1, 6, 7 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European conference on computer vision, pages 213-229, 2020. 2 +[3] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 2 +[4] Yukang Chen, Jianhui Liu, Xiaojuan Qi, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Scaling up kernels in 3d cnns. arXiv preprint arXiv:2206.10555, 2022. 1 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[6] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5 +[7] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 2, 5 +[8] Meng-Hao Guo, Cheng-Ze Lu, Qibin Hou, Zhengning Liu, Ming-Ming Cheng, and Shi-Min Hu. Segnext: Rethinking convolutional attention design for semantic segmentation. arXiv preprint arXiv:2209.08575, 2022. 2 +[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6 +[11] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 6 +[12] Junjie Huang and Guan Huang. Bevpoolv2: A cutting-edge implementation of bevdet toward deployment. arXiv preprint arXiv:2211.17111, 2022. 5, 6 +[13] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 2 + +[14] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 2 +[15] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 5 +[16] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformers. arXiv preprint arXiv:2206.15398, 2022. 6 +[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 6 +[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6 +[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 1 +[20] Youngwan Lee and Jongyoul Park. Centermask: Real-time anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 13906-13915, 2020. 2, 6 +[21] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2, 6 +[22] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo. arXiv preprint arXiv:2209.10248, 2022. 5, 6 +[23] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022. 6 +[24] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 1, 5, 6 +[25] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision (ECCV), pages 1-18. Springer, 2022. 1, 2, 5, 6 +[26] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 2 +[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6 + +[28] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 5, 6 +[29] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018, 2023. 5 +[30] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2 +[31] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision (ECCV), pages 531-548. Springer, 2022. 1, 2, 3, 6 +[32] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 1, 3, 5, 6 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6 +[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6 +[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6 +[36] Tao Lu, Xiang Ding, Haisong Liu, Gangshan Wu, and Limin Wang. Link: Linear kernel for lidar-based 3d perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1105-1115, 2023. 1 +[37] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2 +[38] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters, 5(3):4867-4873, 2020. 2 +[39] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3142-3152, 2021. 2, 6 +[40] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 1, 2, 5, 6 +[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming + +Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32:8026-8037, 2019. 6 +[42] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2 +[43] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2 +[44] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. arXiv preprint arXiv:1811.08188, 2018. 2 +[45] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 2, 3 +[46] Yao Teng, Haisong Liu, Sheng Guo, and Limin Wang. Stageinteractor: Query-based object detector with cross-stage interaction. arXiv preprint arXiv:2304.04978, 2023. 2 +[47] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 2 +[48] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261–24272, 2021. 5 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[50] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926, 2023. 5, 6 +[51] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 913-922, 2021. 2 +[52] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, Xiaogang Wang, and Yu Qiao. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022.6.7 +[53] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo + +lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2 +[54] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 1, 2, 3, 5, 6 +[55] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. arXiv preprint arXiv:2211.10439, 2022. 2, 6, 7 +[56] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1 +[57] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2 +[58] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. A simple baseline for multi-camera 3d object detection. arXiv preprint arXiv:2208.10035, 2022. 2 +[59] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019. 5, 6, 7 +[60] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2 \ No newline at end of file diff --git a/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/images.zip b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4e165223ab799f172c9984a05369c3b6b5c47287 --- /dev/null +++ b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffeae8b94b8978a1345c65822da1314a7f4c93396d9a36790bb2010b75e515b8 +size 741861 diff --git a/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/layout.json b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3eaf0f80d1fd5e8671b85c5311ba37f4cdadd4d9 --- /dev/null +++ b/2023/SparseBEV_ High-Performance Sparse 3D Object Detection from Multi-Camera Videos/layout.json @@ -0,0 +1,8635 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 113, + 103, + 481, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 103, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 113, + 103, + 481, + 138 + ], + "type": "text", + "content": "SparseBEV: High-Performance Sparse 3D Object Detection from Multi-Camera Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "spans": [ + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "text", + "content": "Haisong Liu" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "text", + "content": " Yao Teng" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "text", + "content": " Tao Lu" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "text", + "content": " Haiguang Wang" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "text", + "content": " Limin Wang" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "inline_equation", + "content": "^{1,2,\\boxtimes}" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "text", + "content": "State Key Laboratory for Novel Software Technology, Nanjing University " + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "text", + "content": "Shanghai AI Lab " + }, + { + "bbox": [ + 64, + 160, + 528, + 204 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{liuhs, yaoteng, taolu, haiguangwang}\\} \\text{@smail.nju.edu.cn, lmwang@nju.edu.cn}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 256, + 290, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 555 + ], + "type": "text", + "content": "Camera-based 3D object detection in BEV (Bird's Eye View) space has drawn great attention over the past few years. Dense detectors typically follow a two-stage pipeline by first constructing a dense BEV feature and then performing object detection in BEV space, which suffers from complex view transformations and high computation cost. On the other side, sparse detectors follow a query-based paradigm without explicit dense BEV feature construction, but achieve worse performance than the dense counterparts. In this paper, we find that the key to mitigate this performance gap is the adaptability of the detector in both BEV and image space. To achieve this goal, we propose SparseBEV, a fully sparse 3D object detector that outperforms the dense counterparts. SparseBEV contains three key designs, which are (1) scale-adaptive self attention to aggregate features with adaptive receptive field in BEV space, (2) adaptive spatio-temporal sampling to generate sampling locations under the guidance of queries, and (3) adaptive mixing to decode the sampled features with dynamic weights from the queries. On the test split of nuScenes, SparseBEV achieves the state-of-the-art performance of 67.5 NDS. On the val split, SparseBEV achieves 55.8 NDS while maintaining a real-time inference speed of 23.5 FPS. Code is available at https://github.com/MCG-NJU/SparseBEV." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 598, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 598, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 598, + 287, + 658 + ], + "type": "text", + "content": "Camera-based 3D Object Detection [13, 54, 31, 11, 24, 25, 40] has witnessed great progress over the past few years. Compared with the LiDAR-based counterparts [19, 56, 4, 36], camera-based approaches have lower deployment cost and can detect long-range objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 658, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 287, + 696 + ], + "type": "text", + "content": "Previous methods can be divided into two paradigms. BEV (Bird's Eye View)-based methods [13, 11, 25, 24, 40] follow a two-stage pipeline by first constructing an explicit" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 236, + 526, + 399 + ], + "blocks": [ + { + "bbox": [ + 314, + 236, + 526, + 399 + ], + "lines": [ + { + "bbox": [ + 314, + 236, + 526, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 236, + 526, + 399 + ], + "type": "image", + "image_path": "953c39be16319636c67fcc3abb508b6c57cf75b70b13691028eb273f7ed132af.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 410, + 547, + 483 + ], + "lines": [ + { + "bbox": [ + 304, + 410, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 547, + 483 + ], + "type": "text", + "content": "Figure 1: Performance comparison on the val split of nuScenes [1]. All methods use ResNet50 [10] as the image backbone and the input size is set to " + }, + { + "bbox": [ + 304, + 410, + 547, + 483 + ], + "type": "inline_equation", + "content": "704 \\times 256" + }, + { + "bbox": [ + 304, + 410, + 547, + 483 + ], + "type": "text", + "content": ". FPS is measured on a single RTX 3090 with the PyTorch fp32 backend. We balance accuracy and speed by reducing the number of decoder layers without re-training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 496, + 546, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 546, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 546, + 675 + ], + "type": "text", + "content": "dense BEV feature from multi-view features and then performing object detection in BEV space. Although achieving remarkable progress, they suffer from high computation cost and rely on complex view transformation operators. Another line of work [54, 31, 32] explores the sparse query-based paradigm by initializing a set of sparse reference points in 3D space. Specifically, DETR3D [54] links the queries to image features using 3D-to-2D projection. It has simpler structure and faster speed, but its performance still lags far behind the dense ones. PETR series [31, 32] uses dense global attention for the interaction between query and image feature, which is computationally expensive and buries the advantage of the sparse paradigm. Thus, a natural question arises whether fully sparse detectors achieve similar accuracy to the dense ones?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we find that the key to obtain high performance in sparse 3D object detection is the adaptability of the detector in both BEV and image space. In BEV space," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 149, + 713 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18580" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "the detector should be able to aggregate multi-scale features adaptively. Dense BEV-based detectors typically use a BEV encoder to encode multi-scale BEV features. It can be a stack of residual blocks with FPN [26] (e.g. BEVDet [26]), or a transformer encoder with multi-scale deformable attention [60] (e.g. BEVFormer [25]). For sparse detectors such as DETR3D, we argue that the multi-head self attention (MHSA) [49] among queries can play the role of the BEV encoder, as queries are defined in BEV space. However, the vanilla MHSA has a global receptive field, lacking an explicit multi-scale design. In image space, the detector should be adaptive to different objects with different sizes and categories. This is because although the objects have similar sizes in 3D space, they might vary greatly in images. However, the single-point sampling in DETR3D has a fixed local receptive field and the sampled feature is processed by static linear layers, hindering its performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 408 + ], + "type": "text", + "content": "To this end, we present SparseBEV, a fully sparse 3D object detector that matches or even outperforms the dense counterparts. Our SparseBEV detector contains three key designs, which are (1) scale-adaptive self attention to aggregate features with adaptive receptive field in BEV space, (2) adaptive spatio-temporal sampling to generate sampling locations under the guidance of queries, and (3) adaptive mixing to decode the sampled features with dynamic weights from the queries. We also propose to use pillars instead of reference points as the formulation of query, since pillars introduce better spatial priors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 409, + 289, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 289, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 289, + 541 + ], + "type": "text", + "content": "We conduct comprehensive experiments on the nuScenes dataset. As shown in Fig. 1, our SparseBEV achieves the performance of 55.8 NDS and the speed of 23.5 FPS on the val split, surpassing all previous methods in both speed and accuracy. Besides, we can flexibly adjust of the inference speed by reducing the number of decoder layers without re-training. On test split, SparseBEV with V2-99 [20] backbone achieves 63.6 NDS without using future frames or test-time augmentation. By further utilizing future frames, SparseBEV achieves 67.5 NDS, outperforming the previous state-of-the-art BEVFormerV2 [55] by 2.7 NDS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 554, + 134, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 134, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 134, + 566 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 574, + 228, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 228, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 228, + 588 + ], + "type": "text", + "content": "2.1. Query-based 2D Object Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Recently, Transformer [49] with its attention blocks has been widely applied in the computer vision tasks [5, 8, 2]. In object detection, DETR [2] was the first model to predict objects based on learnable queries and treat the detection as a set prediction problem. A lot of works [60, 45, 30, 37, 7, 21, 57] were then proposed to accelerate the convergence of DETR by using the sampled features instead of using the global ones. For example, Deformable-DETR [60] samples image features based on sparse reference points, and then applies the deformable attention on" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "the features. Sparse R-CNN [45] uses the ROIAign [9] to obtain local features and then performs the dynamic convolution. AdaMixer [7] combines the advantages of the sampling points and the dynamic convolution to further boost the convergence. DN-DETR [21] devises a denoising mechanism that feeds ground-truth bounding boxes with noises into decoder and trains the model to reconstruct the original boxes. DINO [57] furtherly optimizes DN-DETR [21] and DAB-DETR [30] by proposing a contrastive denoising training strategy and a mixed query selection method. There are also several methods [3, 14, 46] designed to tackle the instability of the training procedure. Our work follows the query-based detection paradigm and extends it to 3D space with temporal information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 248, + 479, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 248, + 479, + 261 + ], + "spans": [ + { + "bbox": [ + 305, + 248, + 479, + 261 + ], + "type": "text", + "content": "2.2. Monocular 3D Object Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 267, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 547, + 460 + ], + "type": "text", + "content": "Monocular 3D object detection takes one single image as input and outputs predicted 3D bounding boxes of objects. A significant challenge in monocular 3D object detection is how to transfer 2D features to 3D space. Several works [53, 43, 51, 39] incorporates depth information to deal with this problem. Pseudo LiDAR [53] first estimates the depth of input images and constructs pseudo point clouds. The pseudo point clouds are then sent to a LiDAR-based detection module to predict the 3D boxes of interested objects. CaDDN [43] further proposes a fully differentiable end-to-end network which learns pixel-wise categorical depth distributions to predict appropriate depth intervals in 3D space. Inspired by FCOS [47], FCOS3D [51] projects 3D coordinates onto 2D images and decouples them as 2D attributes (centerness and classification) and 3D attributes (depths, sizes, and orientations)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 467, + 460, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 460, + 480 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 460, + 480 + ], + "type": "text", + "content": "2.3. 3D Object Detection in BEV" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": "Bird's-eye-view (BEV) object detection [38, 42, 13, 54, 31, 58, 25, 55, 40] aims at detecting objects in BEV space given either single-view or multi-view 2D images. Early works [44, 53, 43] typically transformed 2D features to BEV space based on single-view images and conducted monocular 3D object detection. LSS [42] takes six 2D images of different views as input and transforms them into 3D space based on depth estimation. Based on LSS, BEVDet [13] lifts 2D feature to BEV space and uses a BEV encoder with residual blocks and FPN to further encode the BEV features. BEVFormer [25] proposes a spatiotemporal transformer encoder that projects multi-view and multi-timestamp input to BEV representations. To ease the optimization, BEVFormerV2 [55] introduces perspective view supervision and supervises monocular 3D object detection parallel with BEV object detection. SOLOFusion [40] defines the localization potential for depth estimation and fuses short-term, high-resolution and long-term, low-resolution temporal stereo." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18581" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 545, + 297 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 297 + ], + "type": "image", + "image_path": "211b6b7907e2bdcd9c9db511ca352701e967087afea83c289a5349f4c1de0b4b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 309, + 546, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 309, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 546, + 357 + ], + "type": "text", + "content": "Figure 2: The overall architecture of SparseBEV, a fully-sparse camera-only 3D object detector. Queries are initialized to be a sparse set of pillars in BEV space. The scale-adaptive self attention further encodes the queries with adaptive receptive fields. Next, multi-view and multi-timestamp features are aggregated with adaptive spatio-temporal sampling and decoded by adaptive mixing. The decoder repeats " + }, + { + "bbox": [ + 46, + 309, + 546, + 357 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 309, + 546, + 357 + ], + "type": "text", + "content": " times to produce final predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 373, + 289, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 564 + ], + "type": "text", + "content": "Inspired by DETR, another line of works [54, 31, 32] explore the sparse query-based paradigm. DETR3D [54] proposes a top-down framework starting from a learnable sparse set of reference points and refining them iteratively via 3D-to-2D queries. However, such 3D-to-2D projection hinders the receptive field of the query. To handle this, PETR series [31, 32] use global attention for the interaction between queries and image features, and introduce 3D positional embeddings to encode 2D features into 3D representation without explicit projection. Although achieving notable improvements, the expensive dense global attention buries the advantages of the sparse paradigm and makes it difficult to utilize long-term temporal information efficiently. In contrast, we keep the fully sparse design of DETR3D and boost the performance by enhancing the adaptability of the detector." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 574, + 122, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 122, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 122, + 588 + ], + "type": "text", + "content": "3. SparseBEV" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": "As shown in Fig. 2, SparseBEV is a query-based one-stage detector with " + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": " decoder layers. The input multicamera videos are processed frame-by-frame using image backbone and FPN. Next, a set of sparse pillar queries are initialized in BEV space and aggregated by scale-adaptive self attention. These queries then interact with the image features via adaptive spatio-temporal sampling and adaptive mixing to make 3D object predictions. We also propose a dual-branch version of SparseBEV to further enhance the temporal modeling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 373, + 420, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 420, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 420, + 385 + ], + "type": "text", + "content": "3.1. Query Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": "We first define a set of learnable queries, where each of them is represented by its translation " + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "inline_equation", + "content": "[x,y,z]" + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": ", dimension " + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "inline_equation", + "content": "[w,l,h]" + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": " and velocity " + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "inline_equation", + "content": "[v_x,v_y]" + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": ". The queries are initialized to be pillars in BEV space where " + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": " is set to 0 and " + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\sim 4\\mathrm{m}" + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": ". The initial velocity is set to " + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "inline_equation", + "content": "[v_x,v_y] = 0" + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": ". Other parameters " + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "inline_equation", + "content": "(x,y,w,l,\\theta)" + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": " are drawn from random gaussian distributions. Following Sparse R-CNN[45], we attach a " + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 391, + 545, + 498 + ], + "type": "text", + "content": "-dim query feature to each query box to encode the rich instance characteristics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 505, + 464, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 505, + 464, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 505, + 464, + 517 + ], + "type": "text", + "content": "3.2. Scale-adaptive Self Attention" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 523, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 545, + 583 + ], + "type": "text", + "content": "As mentioned above, dense BEV-based methods typically use a BEV encoder to encode multi-scale BEV features. However, in our method, since we do not explicitly build a BEV feature, how to aggregate multi-scale features in BEV space is a challenge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 583, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 583, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 583, + 545, + 691 + ], + "type": "text", + "content": "In this work, we argue that the self attention can play the role of BEV encoder, since queries are defined in BEV space. The vanilla multi-head self attention has global receptive field and lacks the ability of local multi-scale aggregation. Thus, we propose scale-adaptive self attention (SASA), which learns appropriate receptive fields under the guidance of queries. First, we compute the all-pair distances " + }, + { + "bbox": [ + 304, + 583, + 545, + 691 + ], + "type": "inline_equation", + "content": "D \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 304, + 583, + 545, + 691 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 583, + 545, + 691 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 583, + 545, + 691 + ], + "type": "text", + "content": " is the number of queries) between the query centers in BEV space:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 696, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 696, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 353, + 696, + 545, + 716 + ], + "type": "interline_equation", + "content": "D _ {i, j} = \\sqrt {\\left(x _ {i} - x _ {j}\\right) ^ {2} + \\left(y _ {i} - y _ {j}\\right) ^ {2}}, \\tag {1}", + "image_path": "5707f8150c2822c54043dece481d6dfbb6cebcde145d7400aaabf1e0aa3e2d24.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18582" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 282, + 200 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 282, + 200 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 282, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 282, + 200 + ], + "type": "image", + "image_path": "75d89e7fb1c30aa594196a0bba42b6c0ae657a3d0714724c0909e50b55e74ff6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 206, + 287, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 287, + 255 + ], + "type": "text", + "content": "Figure 3: Visualization of the learned BEV receptive field in SASA. We choose two samples from the validation split and visualize 4 heads for each sample, denoted by " + }, + { + "bbox": [ + 46, + 206, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\tau_{1},\\tau_{2},\\tau_{3}" + }, + { + "bbox": [ + 46, + 206, + 287, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 206, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\tau_{4}" + }, + { + "bbox": [ + 46, + 206, + 287, + 255 + ], + "type": "text", + "content": ". Queries are represented by their centers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 271, + 286, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 286, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 271, + 286, + 319 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 271, + 286, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 271, + 286, + 319 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 271, + 286, + 319 + ], + "type": "text", + "content": " denotes the center of the " + }, + { + "bbox": [ + 46, + 271, + 286, + 319 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 271, + 286, + 319 + ], + "type": "text", + "content": "-th query. The attention considers not only the similarity between query features, but the distance between them as well. A toy example below shows how it works:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 329, + 287, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 329, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 78, + 329, + 287, + 357 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}} - \\tau D\\right) V, \\tag {2}", + "image_path": "30345387694b38d9ba129931d92dd9809c332b20a384b61109ea57272686b6fd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "inline_equation", + "content": "Q, K, V \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "text", + "content": " is the query itself and " + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "text", + "content": " is the channel dimension. " + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "text", + "content": " is a scalar to control the receptive field for each query. When " + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "text", + "content": ", it degrades to the vanilla self attention with global receptive field. As " + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 366, + 287, + 438 + ], + "type": "text", + "content": " grows, the attention weights for distant queries become smaller and the receptive field narrows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "content": "In practice, the receptive field controller " + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "content": " is adaptive to each query and specific to each head. Supposing there are " + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "content": " heads, we use a linear transformation to generate head-specific " + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\tau_{1},\\tau_{2},\\dots,\\tau_{H}" + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "content": " from the give query " + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{q}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 497, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 497, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 85, + 497, + 287, + 510 + ], + "type": "interline_equation", + "content": "\\left[ \\tau_ {1}, \\tau_ {2}, \\dots , \\tau_ {H} \\right] = \\operatorname {L i n e a r} _ {d \\rightarrow H} (\\mathbf {q}) \\in \\mathbb {R} ^ {H}, \\tag {3}", + "image_path": "eb309f9043d46bcd5a6b6e771acc7315a7243080a8f07bd8a14dc58146c719ac.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 521, + 263, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 263, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 263, + 533 + ], + "type": "text", + "content": "where the weights are shared across different queries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": "In our experiments, we surprisingly find that the " + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": " for each head is learnt to uniformly distribute in a certain range regardless of the initialization. In Fig. 3, we sort the heads according to " + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": " and visualize the attention weights for the distance part. As we can see, each head learns a different receptive field from each other, enabling the network to aggregate features in a multi-scale manner like FPN." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 712 + ], + "type": "text", + "content": "Scale-adaptive self attention (SASA) demonstrates the necessity of FPN, while being more flexible as it learns the scales adaptively from the query. We also find an interesting phenomenon that different categories of queries have different sizes of receptive field. For example, queries representing the bus have larger receptive field than those representing the pedestrians. More details can be found in the ablation studies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 497, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 497, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 497, + 85 + ], + "type": "text", + "content": "3.3. Adaptive Spatio-temporal Sampling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 91, + 545, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 139 + ], + "type": "text", + "content": "For each frame, we use a linear layer to generate a set of sampling offsets " + }, + { + "bbox": [ + 304, + 91, + 545, + 139 + ], + "type": "inline_equation", + "content": "\\{(\\Delta x_i,\\Delta y_i,\\Delta z_i)\\}" + }, + { + "bbox": [ + 304, + 91, + 545, + 139 + ], + "type": "text", + "content": " adaptively from the query feature. These offsets are transformed to 3D sampling points based on the query pillar:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 156, + 545, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 156, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 318, + 156, + 545, + 196 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l} x _ {i} \\\\ y _ {i} \\\\ z _ {i} \\end{array} \\right] = \\left[ \\begin{array}{c c c} \\cos \\theta & - \\sin \\theta & 0 \\\\ \\sin \\theta & \\cos \\theta & 0 \\\\ 0 & 0 & 1 \\end{array} \\right] \\left[ \\begin{array}{l} w \\cdot \\Delta x _ {i} \\\\ l \\cdot \\Delta y _ {i} \\\\ h \\cdot \\Delta z _ {i} \\end{array} \\right] + \\left[ \\begin{array}{l} x \\\\ y \\\\ z \\end{array} \\right]. \\tag {4}", + "image_path": "f4b9d8ee82ddc9d2c05de333b148b902936368d2f0b7105d7fda2b12acbdfb1a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 204, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 264 + ], + "type": "text", + "content": "Compared with the deformable attention in BEVFormer, our sampling points are adaptive to both query pillar and query feature, thus better covering objects with varying sizes. Besides, these points are not restricted to the query, since we do not limit the range of the sampling offsets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 265, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 361 + ], + "type": "text", + "content": "Next, we perform temporal alignment by warping the sampling points according to motions. In autonomous driving, there are two types of motion: one is ego-motion and the other is object motion. Ego-motion describes the motion of the car from its own perspective as it navigates through the environment, while object motion refers to the movement of other objects in the environment as they move around the self-driving car." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 371, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 432 + ], + "type": "text", + "content": "Dealing with Object Motion. As mentioned above, instantaneous velocity can be equal to average velocity for a short time window in self-driving. Thus, we adaptively warp the sampling points to previous timestamps using the velocity vector " + }, + { + "bbox": [ + 304, + 371, + 545, + 432 + ], + "type": "inline_equation", + "content": "[v_x, v_y]" + }, + { + "bbox": [ + 304, + 371, + 545, + 432 + ], + "type": "text", + "content": " from the query:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 369, + 441, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 441, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 369, + 441, + 545, + 453 + ], + "type": "interline_equation", + "content": "x _ {t, i} = x _ {i} + v _ {x} \\cdot \\left(T _ {t} - T _ {0}\\right) \\tag {5}", + "image_path": "b2385e3813c21886f7a14db018d6b44a061b3724047b0ff053b727f3db147ea9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 371, + 456, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 456, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 371, + 456, + 545, + 469 + ], + "type": "interline_equation", + "content": "y _ {t, i} = y _ {i} + v _ {y} \\cdot \\left(T _ {t} - T _ {0}\\right), \\tag {6}", + "image_path": "1e87f29536281b17cf9d1e0c03e955305ba7d7a0fbe40daf1c7d3c2d04cc55b2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "text", + "content": " denotes the timestamp of previous frame " + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "text", + "content": " denotes the current timestamp). Note that " + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "inline_equation", + "content": "z_{t,i}" + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "text", + "content": " is identical to " + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 479, + 545, + 515 + ], + "type": "text", + "content": " because the velocity vector is defined in BEV plane." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 525, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 574 + ], + "type": "text", + "content": "Dealing with Ego Motion. Next, we warp the sampling points based on the ego pose provided by the dataset. Points are first transformed to the global coordinate system and then to the local coordinate system of frame " + }, + { + "bbox": [ + 304, + 525, + 545, + 574 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 525, + 545, + 574 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 582, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 582, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 318, + 582, + 545, + 597 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l l l l} x _ {t, i} ^ {\\prime} & y _ {t, i} ^ {\\prime} & z _ {t, i} ^ {\\prime} & 1 \\end{array} \\right] ^ {T} = E _ {t} E _ {0} ^ {- 1} \\left[ \\begin{array}{l l l l} x _ {t, i} & y _ {t, i} & z _ {t, i} & 1 \\end{array} \\right] ^ {T}, \\tag {7}", + "image_path": "5af59500452fd43f4db6d47bf38d40069612c2abf7a696b089d9ad91666a670c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 606, + 487, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 606, + 487, + 619 + ], + "spans": [ + { + "bbox": [ + 306, + 606, + 487, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 606, + 487, + 619 + ], + "type": "inline_equation", + "content": "E_{t} = [\\mathbf{R}|\\mathbf{t}]" + }, + { + "bbox": [ + 306, + 606, + 487, + 619 + ], + "type": "text", + "content": " is the ego pose of frame " + }, + { + "bbox": [ + 306, + 606, + 487, + 619 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 606, + 487, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": "Sampling. For each timestamp, we project the warped sampling points " + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\{(x_{t,i}^{\\prime},y_{t,i}^{\\prime},z_{t,i}^{\\prime})\\}" + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": " to each view using the provided camera intrinsics and extrinsics. Since there are overlaps between adjacent views, the projected point may hit one or more views, which are termed as " + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": ". For each hit view " + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": ", we have multi-scale feature maps " + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\{F_{k,j}|j\\in" + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\{1,2,\\dots,N_{\\mathrm{feat}}\\}\\}" + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": " from the image backbone. Features are" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18583" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 542, + 267 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 542, + 267 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 542, + 267 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 542, + 267 + ], + "type": "table", + "html": "
MethodBackboneInput SizeEpochsNDSmAPmATEmASEmAOEmAVEmAAE
PETRv2 [32]ResNet50704 × 2566045.634.90.7000.2750.5800.4370.187
BEVStereo [22]ResNet50704 × 25690‡50.037.20.5980.2700.4380.3670.190
BEVPoolv2 [12]ResNet50704 × 25690‡52.640.60.5720.2750.4630.2750.188
SOLOFusion [40]ResNet50704 × 25690‡53.442.70.5670.2740.5110.2520.181
Sparse4Dv2 [29]ResNet50704 × 25610053.943.90.5980.2700.4750.2820.179
StreamPETR † [50]ResNet50704 × 2566055.045.00.6130.2670.4130.2650.196
SparseBEVResNet50704 × 2563654.543.20.6060.2740.3870.2510.186
SparseBEV †ResNet50704 × 2563655.844.80.5810.2710.3730.2470.190
DETR3D † [54]ResNet101-DCN1600 × 9002443.434.90.7160.2680.3790.8420.200
BEVFformer † [25]ResNet101-DCN1600 × 9002451.741.60.6730.2740.3720.3940.198
BEVDepth [24]ResNet1011408 × 51290‡53.541.20.5650.2660.3580.3310.190
Sparse4D † [28]ResNet101-DCN1600 × 9004855.044.40.6030.2760.3600.3090.178
SOLOFusion [40]ResNet1011408 × 51290‡58.248.30.5030.2640.3810.2460.207
SparseBEV †ResNet1011408 × 5122459.250.10.5620.2650.3210.2430.195
", + "image_path": "e286ebb37855b9bf7976a45342c21af9caa3a227b5a0c3fc29bd3b1a84f23a14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 271, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 545, + 297 + ], + "type": "text", + "content": "Table 1: Performance comparison on the nuScenes val split. † benefits from perspective pretraining. ‡ indicates methods with CBGS [59] which will elongate 1 epoch into 4.5 epochs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 316, + 288, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 341 + ], + "type": "text", + "content": "first sampled by bilinear interpolation " + }, + { + "bbox": [ + 46, + 316, + 288, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 46, + 316, + 288, + 341 + ], + "type": "text", + "content": " in the image plane and then weighted over the scale axis:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 352, + 287, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 352, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 59, + 352, + 287, + 386 + ], + "type": "interline_equation", + "content": "f _ {t, i} = \\frac {1}{| \\mathcal {V} |} \\sum_ {k \\in \\mathcal {V}} \\sum_ {j = 1} ^ {N _ {\\text {f e a t}}} w _ {i j} \\mathbf {B} \\left(F _ {k, j}, \\mathbf {P} _ {k} \\left(x _ {t, i} ^ {\\prime}, y _ {t, i} ^ {\\prime}, z _ {t, i} ^ {\\prime}\\right)\\right) \\tag {8}", + "image_path": "96717d1589298ae038702b0f83a28b526a6d4575cbb84d5c36faed939623add7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{feat}}" + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": " is the number of the multi-scale feature maps and " + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k" + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": " is the projection function for view " + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "inline_equation", + "content": "w_{ij}" + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": " is the weight for the " + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": "-th point on the " + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 396, + 287, + 445 + ], + "type": "text", + "content": "-th feature map and is generated from the query feature by linear transformation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 454, + 149, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 149, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 149, + 468 + ], + "type": "text", + "content": "3.4. Adaptive Mixing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "type": "text", + "content": "Given the sampled features from different timestamps and locations, the key is how to adaptively decode it under the guidance of queries. Inspired by AdaMixer [7], we introduce a simple but effective approach to decode and aggregate the spatio-temporal features with dynamic convolutions [15] and MLP-Mixer [48]. Supposing there are " + }, + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "type": "text", + "content": " frames in total and " + }, + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "type": "text", + "content": " sampling points per frame, we first stack them to a total of " + }, + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "type": "inline_equation", + "content": "P = T \\times S" + }, + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "type": "text", + "content": " points. Thus, the sampled features are organized to " + }, + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{P \\times C}" + }, + { + "bbox": [ + 46, + 474, + 287, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 602, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 639 + ], + "type": "text", + "content": "Channel Mixing. We first perform channel mixing on " + }, + { + "bbox": [ + 46, + 602, + 287, + 639 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 602, + 287, + 639 + ], + "type": "text", + "content": " to enhance object semantics. The dynamic weights are generated from the query feature " + }, + { + "bbox": [ + 46, + 602, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 602, + 287, + 639 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 648, + 287, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 648, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 104, + 648, + 287, + 662 + ], + "type": "interline_equation", + "content": "W _ {c} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {C \\times C} \\tag {9}", + "image_path": "0e5a61bd1596216ab6b33b326b9d7bcafad67a9dbccc760b3a13d5532ff180fc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 665, + 287, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 665, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 91, + 665, + 287, + 678 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {c} (f) = \\operatorname {R e L U} \\left(\\text {L a y e r N o r m} \\left(f W _ {c}\\right)\\right), \\tag {10}", + "image_path": "6452f924d543598f0b54aa9e72fb76f7d9274bf80cba40272a16e0daaf2e8a4b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "W_{c}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " is the dynamic weights and is shared across different frames and different sampling points." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 316, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 341 + ], + "type": "text", + "content": "Point Mixing. Next, we then transpose the feature and apply the dynamic weights to the point dimension of it:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 350, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 350, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 359, + 350, + 545, + 365 + ], + "type": "interline_equation", + "content": "W _ {p} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {P \\times P} \\tag {11}", + "image_path": "793e128fd61d86496b6bf032d9e9ae35269ffe56bb6aaabecd60bff2b6718570.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 346, + 367, + 545, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 367, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 346, + 367, + 545, + 383 + ], + "type": "interline_equation", + "content": "\\mathrm {M} _ {p} (f) = \\operatorname {R e L U} \\left(\\operatorname {L a y e r N o r m} \\left(f ^ {T} W _ {p}\\right)\\right), \\tag {12}", + "image_path": "7fa4e7f2984c7691bfab595ac762ba35026dbd3cbec69523c945051e71ae7284.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 393, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 393, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 393, + 545, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 393, + 545, + 417 + ], + "type": "inline_equation", + "content": "W_{p}" + }, + { + "bbox": [ + 305, + 393, + 545, + 417 + ], + "type": "text", + "content": " is the dynamic weights and is shared across different channels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 418, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 467 + ], + "type": "text", + "content": "After channel mixing and point mixing, the spatiotemporal features are flattened and aggregated by a linear layer. The final regression and classification predictions are computed by two MLPs respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 477, + 444, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 444, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 444, + 491 + ], + "type": "text", + "content": "3.5. Dual-branch SparseBEV" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 616 + ], + "type": "text", + "content": "Inspired by SlowFast [6], we further enhance the long-term temporal modeling by dividing the input video into a slow stream and a fast stream, resulting in a dual-branch SparseBEV. The slow stream is designed to capture fine-grained appearance details and it operates at low frame rates and high resolutions. The fast stream is responsible for capturing long-term temporal stereo and it operates at high frame rates and low resolutions. Sampling points are projected to the two streams respectively and the sampled features are stacked before adaptive mixing." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "In this way, we decouple the learning of static appearance and temporal motion, leading to better performance. Besides, the computation cost is significantly reduced since only a small fraction of frames needs to be processed at high resolution. However, since this dual-branch design makes the framework a little complex, we do not use it unless otherwise stated. The ablations of this part can be found in the supplementary material." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18584" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 537, + 292 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 537, + 292 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 537, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 537, + 292 + ], + "type": "table", + "html": "
MethodBackboneEpochsNDSmAPmATEmASEmAOEmAVEmAAE
DETR3D [54]V2-992447.941.20.6410.2550.3940.8450.133
PETR [31]V2-992450.444.10.5930.2490.3830.8080.132
UVTR [23]V2-992455.147.20.5770.2530.3910.5080.123
BEVFormer [25]V2-992456.948.10.5820.2560.3750.3780.126
BEVDet4D [11]Swin-B [33]90‡56.945.10.5110.2410.3860.3010.121
PolarFormer [16]V2-992457.249.30.5560.2560.3640.4400.127
PETRv2 [32]V2-992459.150.80.5430.2410.3600.3670.118
Sparse4D [28]V2-994859.551.10.5330.2630.3690.3170.124
BEVDepth [24]V2-9990‡60.050.30.4450.2450.3780.3200.126
BEVStereo [22]V2-9990‡61.052.50.4310.2460.3580.3570.138
SOLOFusion [40]ConvNeXt-B [34]90‡61.954.00.4530.2570.3760.2760.148
SparseBEVV2-992462.754.30.5020.2440.3240.2510.126
SparseBEV (dual-branch)V2-992463.655.60.4850.2440.3320.2460.117
BEVFormerV2 † [55]InternImage-XL [52]2464.858.00.4480.2620.3420.2380.128
BEVDet-Gamma † [12]Swin-B90‡66.458.60.3750.2430.3770.1740.123
SparseBEV †V2-992467.560.30.4250.2390.3110.1720.116
", + "image_path": "e864dec083ec47f6aa6f5c05f7d64c218bcef5a7376265b2499476a8c13ab2f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 295, + 546, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 546, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 546, + 321 + ], + "type": "text", + "content": "Table 2: Performance comparison on the nuScenes test split. † uses future frames. ‡ indicates methods with CBGS [59] which will elongate 1 epoch into 4.5 epochs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 339, + 128, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 128, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 128, + 354 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 359, + 180, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 180, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 180, + 372 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 288, + 449 + ], + "type": "text", + "content": "We implement our model using PyTorch [41]. Following previous methods [54, 25, 32, 40], we adopt common image backbones including ResNet [10] and V2-99 [20]. The decoder consists of 6 layers and weights are shared across different layers. By default, we use " + }, + { + "bbox": [ + 46, + 377, + 288, + 449 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 46, + 377, + 288, + 449 + ], + "type": "text", + "content": " frames in total and the interval between adjacent frames is about 0.5s." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "text", + "content": "During training, we adopt the Hungarian algorithm [18] for label assignment between ground-truth objects and predictions. Focal loss [27] is used for classification and L1 loss is used for 3D bounding box regression. We use the AdamW [35] optimizer with a global batch size of 8. The initial learning rate is set to " + }, + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 449, + 287, + 533 + ], + "type": "text", + "content": " and is decayed with cosine annealing policy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 605 + ], + "type": "text", + "content": "Recently, we follow the training setting of the very recent work StreamPETR [50] and refresh our results for fair comparison. For the regression loss, we change the loss weight of " + }, + { + "bbox": [ + 46, + 533, + 288, + 605 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 533, + 288, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 288, + 605 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 533, + 288, + 605 + ], + "type": "text", + "content": " to 2.0 and leave the others to 1.0. Query denoising [21] is also introduced to stabilize training and speedup convergence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 611, + 169, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 169, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 169, + 623 + ], + "type": "text", + "content": "4.2. Datasets and Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 701 + ], + "type": "text", + "content": "We evaluate our model on the nuScenes dataset [1], which consists of large-scale multimodal data collected from 6 surround-view cameras, 1 lidar and 5 radars. The dataset has 1000 videos and is split into 700/150/150 videos for training/validation/testing. Each video has roughly 20s duration and the key samples are annotated every 0.5s." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "type": "text", + "content": "For 3D object detection, there are up to 1.4M annotated" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 341, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 425 + ], + "type": "text", + "content": "3D bounding boxes of 10 classes. The official evaluation metrics include the well-known mean Average Precision (mAP) and five true positive (TP) metrics, including ATE, ASE, AOE, AVE, and AAE for measuring translation, scale, orientation, velocity, and attribute errors respectively. The overall performance is measured by the nuScenes Detection Score (NDS), which is the composite of the above metrics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 434, + 545, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 545, + 448 + ], + "type": "text", + "content": "4.3. Comparison with the State-of-the-art Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "type": "text", + "content": "nuScenes val split. In Tab. 1, we compare SparseBEV with previous state-of-the-art methods on the validation split of nuScenes. Unless otherwise stated, the image backbone is pretrained on ImageNet-1k [17] and the number of queries is set to 900. To keep the simplicity of our approach, the dual-branch design is not used here. When adopting ResNet50 as the backbone and " + }, + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "type": "inline_equation", + "content": "704 \\times 256" + }, + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "type": "text", + "content": " as the input size, SparseBEV outperforms the previous state-of-the-art method SOLOFusion by " + }, + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "type": "inline_equation", + "content": "1.2\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "type": "text", + "content": " and 1.4 NDS. By further adopting nuImages [1] pretraining and reducing the number queries to 400, SparseBEV sets a new record of 55.8 NDS while maintaining a real-time inference speed of 23.5 FPS (corresponding to Fig. 1). Next, we upgrade the backbone to ResNet101 and scale the input size up to " + }, + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "type": "inline_equation", + "content": "1408 \\times 512" + }, + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "type": "text", + "content": ". Under this setting, SparseBEV still surpasses SOLOFusion by " + }, + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "type": "inline_equation", + "content": "1.8\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 453, + 546, + 657 + ], + "type": "text", + "content": " and 1.0 NDS, demonstrating the scalability of our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "nuScenes test split. We submit our method to the website of nuScenes and report the leaderboard in Tab. 2. Using the V2-99 [20] backbone pretrained by DD3D [39], SparseBEV achieves 62.7 NDS and 54.3 mAP without bells" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18585" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 70, + 242, + 119 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 242, + 119 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 242, + 119 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 242, + 119 + ], + "type": "table", + "html": "
Query FormulationNDSmAP
3D Reference Points55.144.0
BEV Pillars55.645.4
", + "image_path": "bce5901fbec8654ac5d645b8bd203c9fca179579b05bfbeb493ff2bdec2909ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 175, + 270, + 255 + ], + "blocks": [ + { + "bbox": [ + 46, + 127, + 287, + 163 + ], + "lines": [ + { + "bbox": [ + 46, + 127, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 127, + 287, + 163 + ], + "type": "text", + "content": "Table 3: Ablations on query formulation. Compared with reference points, pillars introduce better spatial priors and lead to better performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 175, + 270, + 255 + ], + "lines": [ + { + "bbox": [ + 65, + 175, + 270, + 255 + ], + "spans": [ + { + "bbox": [ + 65, + 175, + 270, + 255 + ], + "type": "table", + "html": "
Self AttentionDistance FunctionNDSmAP
MHSA-53.441.4
SASAτD²54.343.8
τD55.645.4
τ√D55.144.3
", + "image_path": "0e9208010f4be48d2b32363bf22a9e2eea2f856a9057169c72aaed37b32cc78c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 262, + 287, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 300 + ], + "type": "text", + "content": "Table 4: Comparison between vanilla multi-head self attention (MHSA) and scale-adaptive self attention (SASA). Our SASA achieves significant improvements over the baseline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 316, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 437 + ], + "type": "text", + "content": "and whistles. Built on top of this, the dual-branch design gives us extra bonus on the performance, hitting 63.6 NDS and " + }, + { + "bbox": [ + 46, + 316, + 288, + 437 + ], + "type": "inline_equation", + "content": "55.6\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 316, + 288, + 437 + ], + "type": "text", + "content": ". We also follow the offline setting of BEVFormerV2 [55] which leverages both past and future frames to assist the detection. Remarkably, our method (single branch only) surpasses BEVFormerV2 by " + }, + { + "bbox": [ + 46, + 316, + 288, + 437 + ], + "type": "inline_equation", + "content": "2.8\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 316, + 288, + 437 + ], + "type": "text", + "content": " and 2.2 NDS. In addition, BEVFormerV2 uses InternImage-XL [52] with over 300M parameters as the image backbone, while our V2-99 backbone is much more lightweight with only " + }, + { + "bbox": [ + 46, + 316, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\sim70\\mathrm{M}" + }, + { + "bbox": [ + 46, + 316, + 288, + 437 + ], + "type": "text", + "content": " parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 444, + 147, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 147, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 147, + 455 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 463, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 535 + ], + "type": "text", + "content": "In this section, we conduct ablations on the validation split of nuScenes. For all experiments, we use ResNet50 pretrained on nuImages [1] as the image backbone. The input contains 8 frames with " + }, + { + "bbox": [ + 46, + 463, + 287, + 535 + ], + "type": "inline_equation", + "content": "704 \\times 256" + }, + { + "bbox": [ + 46, + 463, + 287, + 535 + ], + "type": "text", + "content": " resolution. We use 900 queries in the decoder and the model is trained for 24 epochs without CBGS [59]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 546, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 630 + ], + "type": "text", + "content": "Query Formulation. In Tab. 3, we ablate different formulations of the query. The top row uses a set of reference points distributing uniformly in 3D space (e.g. DETR3D and PETR). By replacing the reference points with pillars in BEV space, we observe performance improvements (+0.5 NDS) over the baseline. This is because pillars introduce better spatial priors than reference points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": "Scale-adaptive Self Attention. We study the effect of scale-adaptive self attention (SASA) in Tab. 4. Compared with the vanilla multi-head self attention, SASA achieves " + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "inline_equation", + "content": "+4.0\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "inline_equation", + "content": "+2.2" + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": " NDS improvements over the baseline. We further ablate different distance functions and find that " + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": " distance works best. Besides, we also observe two in-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 82, + 537, + 202 + ], + "blocks": [ + { + "bbox": [ + 379, + 72, + 474, + 81 + ], + "lines": [ + { + "bbox": [ + 379, + 72, + 474, + 81 + ], + "spans": [ + { + "bbox": [ + 379, + 72, + 474, + 81 + ], + "type": "text", + "content": "Averaged " + }, + { + "bbox": [ + 379, + 72, + 474, + 81 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 379, + 72, + 474, + 81 + ], + "type": "text", + "content": " for Each Class" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 82, + 537, + 202 + ], + "lines": [ + { + "bbox": [ + 316, + 82, + 537, + 202 + ], + "spans": [ + { + "bbox": [ + 316, + 82, + 537, + 202 + ], + "type": "image", + "image_path": "edf298f2c349c7188f8c124bb404a9adae09c8daf07c17f325f2d504075e0f59.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": "Figure 4: Averaged " + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": " over all queries and all heads for each class in SASA. Larger " + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 207, + 545, + 255 + ], + "type": "text", + "content": " indicates smaller receptive fields. We perform statistics on the val split of nuScenes and choose the queries with a confidence score over 0.3." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 273, + 425, + 357 + ], + "blocks": [ + { + "bbox": [ + 313, + 273, + 425, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 273, + 425, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 425, + 357 + ], + "type": "image", + "image_path": "ac449ef0f77a4cbf3f38495d50b2047548ab5771079cd14d96b33667df1190b8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 365, + 545, + 412 + ], + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 412 + ], + "type": "text", + "content": "Figure 5: Ablations of adaptive spatio-temporal sampling. The performance continues to increase with the number of frames. For sampling points, we observe that 16 points per frame works best." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 428, + 274, + 539, + 357 + ], + "blocks": [ + { + "bbox": [ + 428, + 274, + 539, + 357 + ], + "lines": [ + { + "bbox": [ + 428, + 274, + 539, + 357 + ], + "spans": [ + { + "bbox": [ + 428, + 274, + 539, + 357 + ], + "type": "image", + "image_path": "1cf1cbe7a3c755477da5d7dd51134a15623a2a17d31aeade1221d24a7381330f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 435, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 555 + ], + "type": "text", + "content": "teresting phenomena in SASA. First, different heads learn a different receptive field (see Fig. 3), enabling the model to aggregate multi-scale features in BEV space. Second, queries representing larger objects tend to have larger receptive field. In Fig. 4, we average the receptive field coefficient " + }, + { + "bbox": [ + 304, + 435, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 435, + 545, + 555 + ], + "type": "text", + "content": " over all queries and all heads for each class. As we can see, the receptive field of large objects (such as bus and truck) is larger than small objects (such as pedestrian and traffic cone), demonstrating the effectiveness of our adaptive design." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 572, + 545, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 545, + 716 + ], + "type": "text", + "content": "Adaptive Spatio-temporal Sampling. In Fig. 5, we ablate the number of frames and sampling points. The performance continues to improve as the number of frames increases, proving that SparseBEV can benefit from the long-term history. Here, we use 8 frames for fair comparison with previous methods. As for the number of sampling points, we observe that dispatching 16 points for each frame leads to the best performance. We also provide the visualization of the sampling points from the last stage of the decoder in Fig. 6. Our sampling scheme has adaptive regions of interest and achieves good temporal alignments for both static and moving objects across different frames." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18586" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 77, + 544, + 190 + ], + "blocks": [ + { + "bbox": [ + 49, + 77, + 544, + 190 + ], + "lines": [ + { + "bbox": [ + 49, + 77, + 544, + 190 + ], + "spans": [ + { + "bbox": [ + 49, + 77, + 544, + 190 + ], + "type": "image", + "image_path": "9f2cb444e02fd9c7d2d28a4755408ffc9ef72e02d12d8bacdc537fd1fec2f7a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 196, + 546, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 196, + 546, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 546, + 232 + ], + "type": "text", + "content": "Figure 6: Visualization of adaptive spatio-temporal sampling over three consecutive frames. Different instances are distinguished by colors. Point size indices the depth: larger points are closer to the camera. Our sampling scheme has an adaptive receptive field and is well aligned across different timestamps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 246, + 265, + 308 + ], + "blocks": [ + { + "bbox": [ + 70, + 246, + 265, + 308 + ], + "lines": [ + { + "bbox": [ + 70, + 246, + 265, + 308 + ], + "spans": [ + { + "bbox": [ + 70, + 246, + 265, + 308 + ], + "type": "table", + "html": "
Ego AlignObj. AlignNDSmAPmAVE
--44.434.10.510
✓-54.243.50.281
✓✓55.645.40.243
", + "image_path": "0da7e17d6b3bba106a8128736137a69594e846d300ab85c1e3d3ad07b34635e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 355, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 427 + ], + "type": "text", + "content": "Temporal Alignment. In Tab. 5, we validate the necessity of temporal alignment in spatio-temporal sampling. In our method, ego motion is aligned with the provided ego pose, while object motion is aligned with a simple constant velocity motion model. As we can see from the table, both of them contribute to the performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 436, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 556 + ], + "type": "text", + "content": "Adaptive Mixing. In Tab. 6, we validate the design of the adaptive mixing. The top row is a baseline that uses attention weights to aggregate sampled features (as done in DETR3D). Static mixing and adaptive mixing improve the baseline by 2.7 NDS and 6.5 NDS respectively, demonstrating the necessity of the adaptive design. Next, we explore different combinations of channel and point mixing. Channel mixing followed by point mixing leads to better performance, proving that the object semantics should be enhanced before point mixing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 563, + 209, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 209, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 209, + 574 + ], + "type": "text", + "content": "4.5. Limitations and Future Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 677 + ], + "type": "text", + "content": "One limitation of SparseBEV is the heavy reliance on ego pose. As we can see from Tab. 5, the performance drops about 10 NDS without ego-based temporal alignment. However, in the real world, the ego pose provided by IMU may be unreliable and inaccurate, seriously affecting the robustness. Another limitation is that the inference latency increases linearly with the number of frames, since we stack the sampled features along the temporal dimension." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "In the future, we will explore a more elegant and concise way of decoupling spatial appearance and temporal motion. We will also try to extend SparseBEV to other 3D percep" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 320, + 246, + 533, + 347 + ], + "blocks": [ + { + "bbox": [ + 46, + 316, + 287, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 340 + ], + "type": "text", + "content": "Table 5: Ablations on temporal alignment. Aligning both ego and object motion leads to the best performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 246, + 533, + 347 + ], + "lines": [ + { + "bbox": [ + 320, + 246, + 533, + 347 + ], + "spans": [ + { + "bbox": [ + 320, + 246, + 533, + 347 + ], + "type": "table", + "html": "
MethodDetailsNDSmAP
W/o Mixing-49.138.6
Static MixingChannel → Point51.841.9
Adaptive MixingChannel Only53.142.7
Point Only53.643.3
Point → Channel55.444.6
Channel → Point55.645.4
", + "image_path": "df7335a56fade768da81372acf3d5141a84bb01e8b2b5087a84c9271eac4600b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 356, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 305, + 356, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 545, + 380 + ], + "type": "text", + "content": "Table 6: Ablations on adaptive mixing. Channel mixing followed by point mixing leads to better performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 394, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 545, + 417 + ], + "type": "text", + "content": "tion tasks such as BEV segmentation, occupancy prediction and lane detection." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 428, + 378, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 378, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 378, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 448, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 593 + ], + "type": "text", + "content": "In this paper, we have proposed a query-based one-stage 3D object detector, named SparseBEV, which can enjoy the benefits of the BEV space without explicitly constructing a dense BEV feature. SparseBEV improves the adaptability of the decoder by three key modules: scale-adaptive self attention, adaptive spatio-temporal sampling and adaptive mixing. We further introduce a dual-branch design to enhance the long-term temporal modeling. Experiments show that SparseBEV achieves the state-of-the-art performance on the dataset of nuScenes for both speed and accuracy. We hope this exciting result will attract more attention to the fully sparse BEV detection paradigm." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization. Besides, the authors would like to thank Ziteng Gao, Zhiqi Li and Ruopeng Gao for their help." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18587" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 92, + 289, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 289, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 289, + 159 + ], + "type": "text", + "content": "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 1, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 160, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 288, + 215 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European conference on computer vision, pages 213-229, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 218, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 218, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 218, + 288, + 261 + ], + "type": "text", + "content": "[3] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 263, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 263, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 263, + 288, + 297 + ], + "type": "text", + "content": "[4] Yukang Chen, Jianhui Liu, Xiaojuan Qi, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Scaling up kernels in 3d cnns. arXiv preprint arXiv:2206.10555, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 300, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 288, + 365 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 367, + 288, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 367, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 367, + 288, + 411 + ], + "type": "text", + "content": "[6] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "text", + "content": "[7] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 471, + 288, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 471, + 288, + 515 + ], + "spans": [ + { + "bbox": [ + 53, + 471, + 288, + 515 + ], + "type": "text", + "content": "[8] Meng-Hao Guo, Cheng-Ze Lu, Qibin Hou, Zhengning Liu, Ming-Ming Cheng, and Shi-Min Hu. Segnext: Rethinking convolutional attention design for semantic segmentation. arXiv preprint arXiv:2209.08575, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 517, + 288, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 517, + 288, + 550 + ], + "spans": [ + { + "bbox": [ + 53, + 517, + 288, + 550 + ], + "type": "text", + "content": "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 552, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 552, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 552, + 288, + 597 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 599, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 288, + 632 + ], + "type": "text", + "content": "[11] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 634, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 288, + 666 + ], + "type": "text", + "content": "[12] Junjie Huang and Guan Huang. Bevpoolv2: A cutting-edge implementation of bevdet toward deployment. arXiv preprint arXiv:2211.17111, 2022. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[13] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[15] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[16] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformers. arXiv preprint arXiv:2206.15398, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 227 + ], + "type": "text", + "content": "[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "text", + "content": "[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 262, + 547, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 547, + 316 + ], + "type": "text", + "content": "[19] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 317, + 547, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 547, + 360 + ], + "type": "text", + "content": "[20] Youngwan Lee and Jongyoul Park. Centermask: Real-time anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 13906-13915, 2020. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[21] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "type": "text", + "content": "[22] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo. arXiv preprint arXiv:2209.10248, 2022. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[23] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 505, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 548 + ], + "type": "text", + "content": "[24] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 1, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "text", + "content": "[25] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision (ECCV), pages 1-18. Springer, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18588" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "type": "text", + "content": "[29] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. arXiv preprint arXiv:2305.14018, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[30] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "type": "text", + "content": "[31] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. PETR: position embedding transformation for multi-view 3d object detection. In European Conference on Computer Vision (ECCV), pages 531-548. Springer, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 287, + 285 + ], + "type": "text", + "content": "[32] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "type": "text", + "content": "[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 432 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "text", + "content": "[36] Tao Lu, Xiang Ding, Haisong Liu, Gangshan Wu, and Limin Wang. Link: Linear kernel for lidar-based 3d perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1105-1115, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "type": "text", + "content": "[37] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "text", + "content": "[38] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics and Automation Letters, 5(3):4867-4873, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[39] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3142-3152, 2021. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[40] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32:8026-8037, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[42] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "type": "text", + "content": "[43] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "text", + "content": "[44] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. arXiv preprint arXiv:1811.08188, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 547, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 547, + 327 + ], + "type": "text", + "content": "[45] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 369 + ], + "type": "text", + "content": "[46] Yao Teng, Haisong Liu, Sheng Guo, and Limin Wang. Stageinteractor: Query-based object detector with cross-stage interaction. arXiv preprint arXiv:2304.04978, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "text", + "content": "[47] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 415, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 480 + ], + "type": "text", + "content": "[48] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261–24272, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 481, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 545, + 525 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 525, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 568 + ], + "type": "text", + "content": "[50] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. arXiv preprint arXiv:2303.11926, 2023. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 570, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 623 + ], + "type": "text", + "content": "[51] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 913-922, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "text", + "content": "[52] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, Xiaogang Wang, and Yu Qiao. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022.6.7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[53] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18589" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 453 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 116 + ], + "type": "text", + "content": "lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8445-8453, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "text", + "content": "[54] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "text", + "content": "[55] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. arXiv preprint arXiv:2211.10439, 2022. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "type": "text", + "content": "[56] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "text", + "content": "[57] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "type": "text", + "content": "[58] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. A simple baseline for multi-camera 3d object detection. arXiv preprint arXiv:2208.10035, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "type": "text", + "content": "[59] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019. 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "text", + "content": "[60] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18590" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/5aaa9a89-cdc6-4e0a-94f3-a3b82685ba6a_content_list.json b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/5aaa9a89-cdc6-4e0a-94f3-a3b82685ba6a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4822c38b1d36e3c689b29cb531f1c72b59d84680 --- /dev/null +++ b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/5aaa9a89-cdc6-4e0a-94f3-a3b82685ba6a_content_list.json @@ -0,0 +1,2046 @@ +[ + { + "type": "text", + "text": "SparseDet: Improving Sparsely Annotated Object Detection with Pseudo-positive Mining", + "text_level": 1, + "bbox": [ + 156, + 130, + 813, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Saksham Suri\\*1", + "bbox": [ + 127, + 203, + 250, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sakshams@cs.umd.edu", + "bbox": [ + 104, + 224, + 274, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Saketh Rambhatla*1", + "bbox": [ + 313, + 203, + 470, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rssaketh@umd.edu", + "bbox": [ + 320, + 224, + 465, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rama Chellappa2", + "bbox": [ + 511, + 203, + 651, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rchella4@jhu.edu", + "bbox": [ + 511, + 224, + 653, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abhinav Shrivastava", + "bbox": [ + 694, + 203, + 862, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "abhinav@cs.umd.edu", + "bbox": [ + 697, + 224, + 859, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Maryland, College Park1", + "bbox": [ + 205, + 244, + 511, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Johns Hopkins University", + "bbox": [ + 547, + 244, + 761, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 297, + 313, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training with sparse annotations is known to reduce the performance of object detectors. Previous methods have focused on proxies for missing ground truth annotations in the form of pseudo-labels for unlabeled boxes. We observe that existing methods suffer at higher levels of sparsity in the data due to noisy pseudo-labels. To prevent this, we propose an end-to-end system that learns to separate the proposals into labeled and unlabeled regions using Pseudo-positive mining. While the labeled regions are processed as usual, self-supervised learning is used to process the unlabeled regions thereby preventing the negative effects of noisy pseudo-labels. This novel approach has multiple advantages such as improved robustness to higher sparsity when compared to existing methods. We conduct exhaustive experiments on five splits on the PASCAL-VOC and COCO datasets achieving state-of-the-art performance. We also unify various splits used across literature for this task and present a standardized benchmark. On average, we improve by 2.6, 3.9 and 9.6 mAP over previous state-of-the-art methods on three splits of increasing sparsity on COCO. Our project is publicly available at cs.umd.edu/~sakshams/SparseDet.", + "bbox": [ + 75, + 330, + 473, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The performance of object detectors is sensitive to the quality of labeled data [1-3]. Existing object detection methods assume that the training data is pristine and a drop in performance is observed if this assumption fails. Noise in the data used for training object detectors can arise due to incorrect class labels or incorrect/missing bounding boxes. In this work, we deal with the problem of training object detectors with sparse annotations, i.e., missing region or bounding boxes. This problem is of utmost importance, as obtaining crowd-sourced datasets [4, 5] can be expensive", + "bbox": [ + 75, + 722, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0088fc7addd2f7f741af579ba77c33d707eac23ba1e8dbb2f630f0abb57090e.jpg", + "image_caption": [ + "Figure 1: (Top) Most Object Detection datasets have exhaustive annotations for foreground/positives. During training, the unlabeled regions can be safely considered as background/negatives. Sparsely Annotated Object Detection datasets (bottom) have missing annotations. This results in foreground regions (shown in red) being considered as negatives during training, adversely affecting the performance of the classifier." + ], + "image_footnote": [], + "bbox": [ + 506, + 303, + 890, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and laborious. The alternative is to use computer-assisted protocols to collect annotations which have been shown to be noisy and incomplete [6]. This problem of training object detectors with incomplete bounding box annotations is called Sparsely Annotated Object Detection (SAOD).", + "bbox": [ + 496, + 628, + 893, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To understand why training with sparse annotations is detrimental to the performance, consider the example shown in Figure 1 (top). If the annotation were exhaustive, then the negative samples to the classifier contain true background regions. But with sparse annotations, as shown in Figure 1 (bottom), a few positive regions will inevitably be considered as negatives (shown in red), thereby wrongfully penalizing the classifier leading to lower performance. Existing methods[2, 7-10] prevent this by predicting pseudolabels and removing the foreground regions from the negatives to the classifier. However, at higher levels of sparsity, the quality of pseudo-labels is greatly affected, resulting in the same problem noted above.", + "bbox": [ + 496, + 704, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*First two authors contributed equally", + "bbox": [ + 94, + 886, + 297, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6770", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Crowd sourced object detection datasets [5, 11, 12] are ensured to be almost exhaustively labeled. Hence, for SAOD, researchers artificially create sparsely annotated datasets from the original ones. There is no general consensus on the correct way to create the sparsely annotated datasets, a.k.a. splits, and hence each method reports results on one or two different splits. Split can be created by considering the dataset as a whole or per image (i.e. annotations can be removed by considering all the images or only a single image at once). They can also be created by removing annotations in a class-agnostic or class-aware fashion (i.e. remove $p\\%$ , of annotations per category or across all the categories). These variations result in splits with different data distributions making some harder than the others (refer to Table 1). A proper benchmark that analyzes the performance of SAOD methods across these different types of splits is missing. This makes it difficult to compare methods and assess their effectiveness for a specific use case.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the issues discussed above, as our first major contribution, we present SparseDet, a novel SAOD framework that achieves state-of-the-art performance across multiple SAOD benchmarks in practice. SparseDet operates on an image and its augmented counterpart. The combination of features extracted from the two views is used to generate region proposals. Standard detection training methods, consider a region proposal as positive if its intersection over union (IoU) with any ground truth is greater than 0.5 and the rest are treated as negatives. This strategy works when the annotations are exhaustive, which implies that the remaining regions are from background. But due to missing annotations, some of these regions could belong to foreground instances. To prevent considering all region proposals without annotations as negatives, SparseDet partitions all the region proposals into labeled, unlabeled and background. The labeled and background regions are processed as usual. Features extracted from unlabeled regions are then trained with a self-supervised loss. Previous approaches like Co-Mining [10], consider two partitions, labeled and background, and generate pseudo-labels. This is a disadvantage at high sparsity as the generated pseudo-labels can be very noisy. The self-supervised loss in our approach enforces consistency between the features of the two views for the unlabeled regions and prevents penalizing the classifier due to false negatives.", + "bbox": [ + 75, + 367, + 472, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our second major contribution is unifying evaluation. The standard practice is to simulate sparsely annotated training data on COCO [5] and PASCAL-VOC [11] train sets and evaluate them on their corresponding standard validation set. As discussed above, a survey of recent SAOD approaches [2, 7-10, 13-16] reveals that there are at least five different ways to create splits, each differing in the strategy (refer to Section 4.2) used to achieve the desired level of sparsity. However, these splits have not been made public,", + "bbox": [ + 73, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "making it difficult to replicate results for comparison. Additionally, each of these strategies has a different property for simulating sparse data, e.g., different distribution of annotations per class, resulting in a different training set. As a result, methods trained on different splits cannot be compared with one another. To mitigate these issues we standardize the generation of these splits that enables the evaluation of any SAOD methods on all of them for fair comparison. Additionally, we propose a new benchmark that assesses the semi-supervised learning capabilities of SAOD methods, i.e., leveraging unlabeled data to improve performance. We present our approach as a baseline. We will make the data for the new benchmark along with all the SAOD splits public to facilitate future research. We briefly summarize our contributions below.", + "bbox": [ + 496, + 90, + 893, + 318 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel formulation, SparseDet, for SAOD which is an end-to-end approach that identifies labeled, unlabeled and background regions and deals with them in appropriate manner.", + "- We show state-of-the-art performance on sparsely annotated object detection across various splits. On average, we improve by 2.6, 3.9 and $9.6\\mathrm{mAP}$ over previous state-of-the-art methods on three splits of increasing sparsity on COCO.", + "- We standardize the experimental setup for SAOD by evaluating methods on all the splits to facilitate future research. Additionally, we propose a new benchmark that evaluates the semi-supervised learning capabilities of SAOD methods." + ], + "bbox": [ + 517, + 327, + 890, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Section 2, we discuss the related works on SAOD and related fields. We describe our approach in detail in Section 3. We describe our experimental setup and present results in Section 4, and conclude in Section 5.", + "bbox": [ + 496, + 560, + 893, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 633, + 640, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-supervised object detection: Semi-supervised object detection (SSOD) is an active field that also deals with training object detectors with missing annotations. Existing works on semi-supervised object detection, have focused mainly on consistency regularization [17, 18] or pseudolabeling-based approaches [19-22]. The main idea behind these approaches is to perturb the images, or features, and apply a consistency regularization loss to enforce consistency between the predictions using a student teacher framework. However, typical SSOD methods assume a small exhaustively labeled set and a large unlabeled set for training. This is different from Sparsely annotated object detection (SAOD), which assumes a large training set which is sparsely labeled.", + "bbox": [ + 496, + 657, + 893, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sparsely annotated object detection: One of the initial works addressing this problem, by Niitani et al. [7], pro", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6771", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "poses utilizing logical relations between the co-occurrences of objects and pseudo-labeling. Yoon et al. [23] use object tracking to densely label objects across sparsely annotated frames along with single stage detectors to mitigate the negative effects of missing annotations. Wu et al. [2] propose a re-weighting approach where the gradients corresponding to region of interest are weighed as a function of overlap with ground truth instances. Improving upon the previous work, Zhang et al. [8] propose an automatic re-calibration strategy for single stage detectors where the negative branch is changed to take into account low confidence background predictions which might correspond to missing annotations.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, one of the most recent works, Co-mining [10] uses a co-training strategy by using two views of an image and predictions from one view along with the ground truth as supervision for the other view and vice versa. Our method doesn't solely rely on pseudo-labels and leverages a self-supervised loss to prevent propagating negative gradients to the model due to false negatives.", + "bbox": [ + 75, + 276, + 472, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fully supervised object detection: After the success of AlexNet [24] on the image classification challenge (ILSVRC 2012) [25], research on designing deep neural networks for object detection gathered more interest. First among the successful methods are the two-stage Region-based convolutional networks (R-CNN) family of detectors. Two-stage object detectors consists of 1) a region proposal stage which produces a set of candidate object bounding boxes followed by 2) a classification stage which classifies each candidate region as either belonging to a foreground object or \"background\". R-CNN processes a large amount of region proposals by cropping the input image and using a CNN backbone to extract features making it extremely slow. Fast R-CNN [26] was proposed to overcome this limitation. Fast R-CNN computes one convolution feature map for the whole image. RoI Pooling was introduced to pool the feature for each region of interest into a fixed spatial dimension. RoI pooling shares the computation among all the region proposals speeding up training and inference. Fully connected layers are applied on the fixed RoI pooled feature maps which are then passed to two sister heads, for classification and bounding box regression. The whole network is trained end-to-end with a multi-task loss avoiding the multi-stage training in R-CNN [27]. While Fast R-CNN improved the efficiency of its predecessor, test time computation bottleneck is still an issue because of the region proposal method employed. Faster R-CNN [28] proposed a region proposal network (RPN), an elegant solution that trains deep networks to predict region proposals for practically no additional computational overhead. RPN shares convolutional layers with Fast R-CNN [26] and at test time the cost of generating proposals is minimal. Faster R-CNN paved the way for more sophisticated and efficient two stage detection architectures [29-33]. Faster R-CNN", + "bbox": [ + 75, + 386, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "has also been extended to achieve state-of-the-art instance segmentation [34], panoptic segmentation [35, 36], and 3D mesh generation [37] etc. A few limitations of two stage object detectors and anchor boxes has also inspired the family of single-stage [38-40] and anchor-free object detection systems [41, 42].", + "bbox": [ + 496, + 90, + 893, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 498, + 193, + 609, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present SparseDet for Sparsely Annotated Object Detection. Given $N$ images, we denote the set of labeled regions in the dataset as $\\mathcal{B}_{\\mathrm{l}} = \\{b_{i},c_{i}\\}_{i = 1}^{\\mathrm{N}_{\\mathrm{l}}}$ where $(b_{*},c_{*})$ are the bounding box and class labels respectively. The unlabeled regions are denoted as $\\mathcal{B}_{\\mathrm{ul}} = \\{b_k\\}_{k = 1}^{\\mathrm{N}_{\\mathrm{ul}}}$ and the background regions as $\\mathcal{B}_{\\mathrm{bg}} = \\{b_q\\}_{q = 1}^{\\mathrm{N}_{\\mathrm{bg}}}$ . Note that unlike existing SAOD approaches, we do not assume images to contain at least one labeled region in this work. The unlabeled $\\mathcal{B}_{\\mathrm{ul}}$ and background $\\mathcal{B}_{\\mathrm{bg}}$ sets are not known a-priori.", + "bbox": [ + 496, + 218, + 893, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 498, + 363, + 609, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed approach is shown in Figure 2 and consists of a backbone network that extracts features from the original and augmented views of an image. The common RPN (C-RPN), concatenates the features, to generate a set of region proposals. A region proposal $b$ can belong to one of three groups, namely 1) labeled regions $b \\in \\mathcal{B}_1$ , 2) unlabeled foreground regions $b \\in \\mathcal{B}_{\\mathrm{ul}}$ , or 3) background regions $b \\in \\mathcal{B}_{\\mathrm{bg}}$ . For a given set of ground-truth annotations, the first group, i.e. labeled regions can be automatically identified. The problem then reduces to identifying and separating the second group i.e. the unlabeled regions, from the background regions. Given all the region proposals, a pseudo-positive mining (PPM) step identifies the unlabeled regions and segregates them from the background regions. The labeled and unlabeled regions are trained using supervised and self-supervised losses respectively. We describe each stage in detail below.", + "bbox": [ + 496, + 387, + 893, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feature Extraction", + "text_level": 1, + "bbox": [ + 498, + 650, + 683, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an image $I$ , an augmented version of $I$ denoted as $\\mathcal{A}(I)$ is computed. In this work, we use random contrast, brightness, saturation, lighting and bounding box erase in a cascaded fashion to generate $\\mathcal{A}(I)$ . A backbone network is employed to extract two features, $f_{\\mathrm{o}}$ and $f_{\\mathrm{a}}$ , from $I$ and $\\mathcal{A}(I)$ respectively.", + "bbox": [ + 496, + 672, + 893, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Common RPN (C-RPN)", + "text_level": 1, + "bbox": [ + 498, + 771, + 720, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Two stage object detectors [28, 30, 43] use a region proposal network (RPN) [28] to generate regions of interest (RoIs). We propose C-RPN which concatenates $f_{\\mathrm{o}}$ and $f_{\\mathrm{a}}$ to obtain the RoIs. This is different from existing approaches that use $f_{\\mathrm{o}}$ and $f_{\\mathrm{a}}$ to generate two separate sets of RoIs. Operating on two sets increases the difficulty of identifying the labeled, unlabeled and background regions which", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6772", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9725617a6613c754f8f36432dc8cc3994bab2ec28970cf7fe7fb2a7f718810ce.jpg", + "image_caption": [ + "Figure 2: Illustration of SparseDet for sparsely annotated object detection. Following feature extraction from the original and augmented image, a common set of proposals is generated by the common RPN (C-RPN). Using an end to end approach, we identify and mine proposals corresponding to missing annotations using pseudo positive mining (PPM). We train the network end-to-end using a combination of supervised and self-supervised losses. The unlabeled instances (black) are supervised with a self-supervised loss and the labeled instances (green) are supervised with ground truth annotations." + ], + "image_footnote": [], + "bbox": [ + 101, + 95, + 851, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are required for subsequent stages. The sparse ground truth is used as supervision to train the C-RPN with a binary cross entropy loss for foreground classification and smooth L1 [44] loss for bounding box regression as shown below:", + "bbox": [ + 75, + 362, + 470, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r p n}} ^ {\\mathrm {s}} \\left(\\mathbf {c}, \\mathbf {b}, \\mathbf {c} ^ {*}, \\mathbf {b} ^ {*}\\right) = \\mathcal {L} _ {\\mathrm {b c e}} \\left(\\mathbf {c}, \\mathbf {c} ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {r e g}} \\left(\\mathbf {b}, \\mathbf {b} ^ {*}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 446, + 468, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{c}$ , $\\mathbf{b}$ are the objectness scores and bounding boxes, and $\\mathbf{c}^*$ , $\\mathbf{b}^*$ are the corresponding ground truth.", + "bbox": [ + 76, + 474, + 468, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Pseudo Positive Mining (PPM)", + "text_level": 1, + "bbox": [ + 76, + 512, + 348, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the RoIs from C-RPN, the next step is to segregate the unlabeled regions from the labeled and background regions. A standard practice in training detectors is to consider all proposals with an objectness score greater than $\\tau_{\\mathrm{obj}}$ and IoU greater $\\tau_{\\mathrm{fg}}$ with any ground truth as foreground and proposals with IoU less than $\\tau_{\\mathrm{bg}}$ as background. In the presence of missing annotations, detectors are wrongly penalized because the IoU of RoIs corresponding to missing annotations with any ground truth is less than $\\tau_{\\mathrm{bg}}$ . To avoid this, we mine \"potential\" foregrounds and consider them as unlabeled regions $(\\mathcal{B}_{\\mathrm{u}})$ . The proposed PPM module is based on our observation that when trained with sparse annotations, the RPN can reliably distinguish foreground from background regions. We pick all RoIs that have an objectness score greater than $\\tau_{\\mathrm{ppm}}$ and IoU less than $\\tau_{\\mathrm{bg}}$ with any ground truth as the unlabeled regions. The remaining RoIs are assigned as background.", + "bbox": [ + 75, + 536, + 470, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Losses", + "text_level": 1, + "bbox": [ + 76, + 801, + 163, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pseudo positive mining step segregates the RoIs into labeled, unlabeled and background regions. An RoI pooling layer [28] extracts RoI pooled features for the labeled and background regions using the feature $f_{0}$ . The detection head processes the RoI pooled features to predict class-wise", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "probabilities and bounding box adjustments for each region. The ground truth is used to supervise the predictions by applying the cross entropy loss for classification and smooth L1 [44] for bounding box regression as shown below:", + "bbox": [ + 496, + 361, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\det , 0} ^ {s} \\left(\\mathbf {c}, \\mathbf {b}, \\mathbf {c} ^ {*}, \\mathbf {b} ^ {*}\\right) = \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {c}, \\mathbf {c} ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {r e g}} \\left(\\mathbf {b}, \\mathbf {b} ^ {*}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 428, + 890, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{c}$ , $\\mathbf{b}$ are the class and bounding box predictions, and $\\mathbf{c}^*$ , $\\mathbf{b}^*$ are the corresponding ground truth class labels and bounding boxes.", + "bbox": [ + 496, + 450, + 893, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All detections with a score greater than $\\tau_{\\mathrm{m}}$ are combined with the ground truth followed by a class specific NMS to obtain the merged ground truth. It is ensured that no ground truth annotation is discarded during this step. The labeled and background regions along with $f_{\\mathrm{a}}$ are used to extract RoI pooled features which are then fed to the detector head. The predictions of the detector head are supervised with the merged ground truth with the following losses:", + "bbox": [ + 496, + 497, + 890, + 618 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\det , \\mathrm {a}} ^ {\\mathrm {s}} (\\mathbf {c}, \\mathbf {b}, \\mathbf {c} ^ {*}, \\mathbf {b} ^ {*}) = \\mathcal {L} _ {\\mathrm {b c e}} \\left(\\mathbf {c}, \\mathbf {c} _ {\\mathbf {m}} ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {r e g}} (\\mathbf {b}, \\mathbf {b} _ {\\mathbf {m}} ^ {*}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 625, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{c}$ , $\\mathbf{b}$ are the class and bounding box predictions from the detector head, and, $\\mathbf{c}_{\\mathbf{m}}^{*}, \\mathbf{b}_{\\mathbf{m}}^{*}$ are the corresponding merged ground truths.", + "bbox": [ + 496, + 647, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, a class agnostic NMS is performed on the unlabeled regions $\\mathcal{B}_{\\mathrm{u}}$ (obtained after PPM described in Sec. 3.4). The unlabeled regions along with $f_{\\mathrm{o}}$ and $f_{\\mathrm{a}}$ are passed through the ROI pooling layer and the detection head to obtain $f_{\\mathrm{o}}^{\\mathrm{d}}$ and $f_{\\mathrm{a}}^{\\mathrm{d}}$ respectively. A self-supervised loss is applied that enforces the detection head features of the original and augmented regions to be consistent with each other as shown below:", + "bbox": [ + 496, + 693, + 890, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S S L}} ^ {\\mathrm {u}} \\left(f _ {\\mathrm {a}} ^ {\\mathrm {d}}, f _ {\\mathrm {o}} ^ {\\mathrm {d}}\\right) = \\| f _ {\\mathrm {a}} ^ {\\mathrm {d}} - f _ {\\mathrm {o}} ^ {\\mathrm {d}} \\| _ {2} ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 818, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The network is trained end-to-end with both supervised and unsupervised losses as shown below:", + "bbox": [ + 498, + 843, + 890, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = 0. 5 * \\left(\\mathcal {L} _ {\\text {d e t , o}} ^ {\\mathrm {s}} + \\mathcal {L} _ {\\text {d e t , a}} ^ {\\mathrm {s}}\\right) + \\mathcal {L} _ {\\mathrm {r p n}} ^ {\\mathrm {s}} + \\mathcal {L} _ {\\mathrm {S S L}} ^ {\\mathrm {u}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 878, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6773", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/06a9657027fd73d0d4a809763c37d70c7c4327f8df6f6e986754678bf857d4ba.jpg", + "table_caption": [ + "Table 1: Sparsely annotated object detection results on three splits of COCO dataset. \"Oracle\" corresponds to training models using all annotations. Results are reported on the COCO validation set using AP[0.50:0.95]." + ], + "table_footnote": [], + "table_body": "
MethodSplit-1Split-2Split-3100%
30%50%70%30%50%70%30%50%70%
Oracle40.91
Pseudo Label [7]-27.50--------
BRL [8]-32.70--------
Co-mining [10]36.3532.8424.9336.7233.0424.8336.7632.5424.96-
Ours38.2235.9232.6839.7636.9435.3339.5637.1535.48-
", + "bbox": [ + 80, + 130, + 888, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion: In the absence of ground truth annotations, i.e. for completely unlabeled images, the supervised losses $(\\mathcal{L}_{\\mathrm{det,o}}^{\\mathrm{s}},\\mathcal{L}_{\\mathrm{det,a}}^{\\mathrm{s}},\\mathcal{L}_{\\mathrm{rpn}}^{\\mathrm{s}})$ cannot be computed. Our proposed approach leverages self-supervised consistency loss that does not need ground truth. This helps our approach leverage these unlabeled regions unlike contemporary SAOD methods. We claim that this is the first method to use self-supervised losses for SAOD. Even though Co-Mining [10] claims to use self-supervised learning it is technically co-learning. Ours is the first approach to use pseudo-labeling and a self-supervised loss to handle the sparse annotations.", + "bbox": [ + 75, + 266, + 472, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 76, + 446, + 313, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe the experiments to evaluate our proposed approach. In Sections 4.1 and 4.2, we describe data, splits, and metrics. In Section 4.3, we mention the implementation details followed by the baselines in Section 4.4. We compare with contemporary methods in Section 4.5, followed by an ablation study in Section 4.6.", + "bbox": [ + 75, + 470, + 470, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data and Metrics", + "text_level": 1, + "bbox": [ + 76, + 571, + 246, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct all our experiments on the COCO [5] and PASCAL-VOC [11, 12] (2007+2012) datasets. The COCO [5] dataset consists of 118000 and 5000 images for training and validation respectively. Experiments on the PASCAL-VOC07 [11] are conducted on 5011 trainval images and performance is computed on 4952 images of the test set. The PASCAL-VOC12 version consists of 11530 (trainval) images for training and evaluation is done on the PASCAL-VOC07 test set. Following past literature [2, 7-10], we create five different splits (Section 4.2) and report results on them. For splits on the COCO [5] dataset, we use the standard COCO style Average Precision (AP). For splits on the PASCAL-VOC [11] dataset, we use the standard PASCAL-VOC style $\\mathrm{AP}_{50}$ , which is Average Precision computed at an IoU threshold of 0.5.", + "bbox": [ + 75, + 595, + 470, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. SAOD Splits", + "text_level": 1, + "bbox": [ + 76, + 830, + 210, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An extensive review of sparsely annotated object detection methods reveals that there are at least five popular types of splits in use for creating training data. Most methods re", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "port results on a subset of these making it harder to compare across methods.", + "bbox": [ + 496, + 266, + 890, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We standardize the evaluation of SAOD methods by evaluating exhaustively on all the splits facilitating future research. We briefly describe the splits below.", + "bbox": [ + 496, + 296, + 890, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Split-1 [9, 10]: In this split, for each object category, $p\\%$ of annotations are randomly removed from the COCO [5] train set and results are reported on the validation set. This split simulates sparsity at dataset level in a class aware fashion. Note, this split can contains images with no annotations. We experiment with $p = \\{30, 50, 70\\}$ .", + "bbox": [ + 496, + 340, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Split-2: For each image in the COCO [5] train set, all annotations of $p\\%$ of all categories in that image removed and results are reported on the COCO validation set. This split can be considered as image level and class aware. We experiment with $p = \\{30, 50, 70\\}$ . Note, this split might contain images with no annotations.", + "bbox": [ + 496, + 431, + 890, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Split-3: This split, which uses the COCO 2017 [5] train set for training and the validation set for evaluation, deletes $p\\%$ annotations in a class agnostic fashion for each image ensuring at least one annotation. This split is image level but class agnostic. For the experiments, we use $p = \\{30, 50, 70\\}$ .", + "bbox": [ + 496, + 523, + 890, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Split-4 [8, 10]: This split requires evaluating models on three different settings namely easy, hard and extreme ensuring at least one annotation per image constructed using PASCAL-VOC 2007+12 [11, 12] trainval set. Results are reported on the PASCAL-VOC 2007 [11] test set. For each image in the training set, the easy setting randomly removes one annotation while the hard setting randomly removes half of the annotations. The extreme setting retains only one annotation per image. All the sets ensure each image consists of atleast one annotation. This split is a small scale version for an image level class agnostic split.", + "bbox": [ + 496, + 599, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Split-5 [2]: This split uses the PASCAL-VOC 2007 [11] train set for training and the PASCAL-VOC 2007 test set for evaluation and drops $p\\%$ annotations per class. For each image in this construction, instances of randomly selected categories are exhaustively annotated while the remaining categories do not have any annotations. This is a small scale version of Split-1 as annotations are dropped in a class aware manner across the full dataset maintaining at least 1 annotation per image. In this case we use $p = \\{30,40,50\\}$ .", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6774", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3d4716e035cbff5edcbfb6c35def3ce98447fe4abc1c60bb8707fba2f1bbd665.jpg", + "table_caption": [ + "Table 2: Sparsely annotated object detection results on two splits of PASCAL-VOC dataset. \"Oracle\" corresponds to training models using all annotations. Results are reported on VOC 07 test set using $\\mathrm{AP}_{50}$ ." + ], + "table_footnote": [], + "table_body": "
MethodSplit-4Split-5
EasyHardExtreme30%40%50%
Oracle83.0977.47
BRL [8]73.5071.7066.20---
Co-mining [10]79.5978.3869.6074.4273.3069.89
Ours82.1581.5075.5976.8475.8874.35
", + "bbox": [ + 80, + 145, + 566, + 266 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0bad3be2ec215f49d2b8cc30ef557a02ff0918e7f522634c8b3aad84cdbdc0d8.jpg", + "table_caption": [ + "Table 3: Results on the proposed SSL+SAOD setup. VOC12 [12] is used as the unlabeled data and $p$ is the removal percentage" + ], + "table_footnote": [], + "table_body": "
MethodAP
30%40%50%
Co-mining [10]46.5345.9843.21
Ours48.3447.8246.76
", + "bbox": [ + 581, + 146, + 890, + 266 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b0d8db6da3ab120430d6623533b8f8a68f5cfa0e0d86ea080525681d12f77e0.jpg", + "image_caption": [ + "Figure 3: Illustration of the effects of various augmentations used in this work." + ], + "image_footnote": [], + "bbox": [ + 78, + 273, + 194, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/127cdbd84701c76eb453ef5e74577620e145a95d25e394549efc8c2306084aca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 273, + 316, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d877a288dc50ae2b0bdf230522fc51238791e617fa1acf058275223dd40ff44c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 273, + 431, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a4c38e3645337dd3cd2a68f2f44865940f4a5083c98cb91de3af34f37b908f4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 273, + 547, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66cbe7a543613e3cef39c4809b5ff449544eb441c554dc702ce7bb9af25e0a0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 273, + 663, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/83351c1a395a0db47470a9a3cfdf3c99c7854f3a9d255567f8168142c8b5ca5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 273, + 777, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/28bbe280de4a8346d7c59105a89e02c486893a625eb0961f8c6c2e29609189da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 273, + 892, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 433, + 290, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For all our experiments, we use a Faster RCNN [28] architecture with a ResNet-101 [45] FPN backbone [30] implemented using Detector2 [46] framework. For the image augmentations, we apply a For augmentation, a cascade of random contrast, brightness, saturation, lighting and bounding box erase augmentations. The effect of this augmentation is shown in Figure 3. We train all our models with a batch size of 8 for 270000 and 18000 iterations on the COCO and PASCAL-VOC splits respectively with a learning rate of 0.01. The learning rate is decreased ten fold twice at \\{210000, 250000\\} and \\{12000, 15000\\} for COCO and PASCAL-VOC respectively. We adopt a warm-up strategy for 1000 and 100 iterations for the COCO and PASCAL-VOC respectively. Following existing detector implementations we set $\\tau_{\\mathrm{bg}}$ , $\\tau_{\\mathrm{fg}}$ , $\\tau_{\\mathrm{obj}}$ and $\\tau_{\\mathrm{ppm}}$ to 0.2, 0.4, 0.5 and 0.8 respectively. We set $\\tau_{m}$ to 0.9. During inference, we compute the backbone features ( $f_{\\mathrm{o}}$ and $f_{\\mathrm{a}}$ ) for both the original and augmented versions of the input to obtain the RoIs and only $f_{\\mathrm{o}}$ is used for the final detections.", + "bbox": [ + 75, + 459, + 472, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Baselines", + "text_level": 1, + "bbox": [ + 76, + 765, + 186, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method against Co-Mining [10], BRL [8], and Pseudo Label [7].", + "bbox": [ + 76, + 791, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We choose these methods as they outperform previous approaches for the SAOD task. We use the public implementation of these methods and use a ResNet 101 FPN [30] backbone for all the experiments in order to perform a fair comparison.", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Comparison with state-of-the-art", + "text_level": 1, + "bbox": [ + 500, + 433, + 790, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our method with contemporary methods. We evaluate all the models in two different setups. We name the first setup Sparsely annotated setup which evaluates models on SAOD. The second setup, contains labeled and unlabeled images and regions.", + "bbox": [ + 496, + 458, + 892, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sparsely annotated setup: We show results of this setup in Table 1 and Table 2. Splits-2 and 3 ensure at least one annotation per image is retained. On the other hand, Split-1 doesn't ensure this and has significantly more unlabeled data than the other splits. Splits-4 and 5 have no unlabeled images at all. In both the tables, the rows named \"oracle\" refers to the models trained using all annotations.", + "bbox": [ + 496, + 549, + 892, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From Table 1, we observe that our approach outperforms the other baselines and is closer to the oracle performance on the $30\\%$ splits. Co-mining [10], performs competitively on all the splits at $30\\%$ .", + "bbox": [ + 496, + 656, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On Split-1, our method obtains a performance improvement of 1.87, 3.08 and 7.75 on $30\\%$ , $40\\%$ and $50\\%$ sparsity respectively over Co-mining [10].", + "bbox": [ + 496, + 717, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On Split-2, our method obtains an improvement of 3.04, 3.9 and 10.5 at $30\\%$ , $40\\%$ and $50\\%$ sparsity respectively over the previous state of the art Co-mining. On Split-3, an again we see consistent improvement of 2.8, 4.61 and 10.52 at $30\\%$ , $40\\%$ and $50\\%$ sparsity respectively over Co-mining.", + "bbox": [ + 496, + 763, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In particular, on the hardest setting $(70\\%)$ in Split 1-3, we demonstrate that current SAOD methods struggle at higher sparsity in the labeled data. This is because the performance of current SAOD approaches rely on the quality", + "bbox": [ + 496, + 839, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6775", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/564cb0798b5f0563f6e6b9ff43d8ae9a4470bed923ec05e21c15246eb00b2b47.jpg", + "table_caption": [ + "Table 4: Ablation of various components of proposed approach." + ], + "table_footnote": [], + "table_body": "
C-RPNPPM\\( {\\mathcal{L}}_{\\text{SSL }}^{\\mathrm{u}} \\)\\( {\\mathcal{L}}_{\\text{det,a }}^{\\mathrm{s}} \\)AP\\( \\Delta \\)
✘✘✘✘41.77-
✘✘✘✓42.670.90
✓✘✘✓45.303.53
✓✓✘✓45.443.67
✓✓✓✘45.513.74
✘✓✓✓44.863.09
✓✓✓✓46.004.23
", + "bbox": [ + 138, + 119, + 444, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of pseudo labels which degrades at higher sparsity. Our proposed PPM prevents penalizing the classifier irrespective of the quality of pseudo labels and utilizes self-supervised loss to benefit from mined regions. On average, all the methods achieve lower performance on Split-1 compared to the other splits due to the nature of its creation, i.e., class aware dataset level fashion of sparsity.", + "bbox": [ + 75, + 335, + 468, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Table 2, we see a similar trend as above. All the methods perform competitively on the easier settings of both splits. However, the performance gap between the our approach and Co-mining [10] increases at higher sparsity.", + "bbox": [ + 75, + 441, + 468, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On Split-4, we get an improvement of 2.56, 3.12 and $5.99\\mathrm{(AP_{50})}$ percentage points on the easy, hard and extreme settings respectively over Co-mining. On Split-5, an improvement of 2.42, 2.58 and 4.46 points was observed. For details on more experiments refer to supplementary material.", + "bbox": [ + 75, + 503, + 468, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Single Instance Object Detection (SIOD) setup: [47] proposed SIOD where only one instance per category is annotated in every image. While this setup has its benefits [47], note that it is a special case of SAOD. For this setup, we obtain an AP of 32.89 (compared to 31.9 of [47]) which is an improvement of $\\sim 1$ mAP.", + "bbox": [ + 75, + 609, + 468, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$SSL + SAOD$ setup: We propose a semi-supervised learning benchmark for SAOD. This benchmark entails training models on a sparsely annotated labeled and an unlabeled set. As shown in Figure 4, this setup introduces two kinds of sparsity in the data, namely, sparsity in labeled regions (left) and images (right). We believe this is a realistic setup and SAOD methods must be capable of tackling both these kinds of sparsity in the data. For this setup, we use Split-5 (Section 4.2) with increasing sparsity as the labeled set and VOC12 trainval as the unlabeled set. We use the COCO-style AP metric to report results on this setup.", + "bbox": [ + 75, + 717, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 3, we compare against Co-mining [10] and ob", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6400e3da83d042143b2db15d36521d5f28c363d9f99f4a510a52b093a23f2e6d.jpg", + "image_caption": [ + "Sparsity in labeled regions Sparsity in labeled images", + "Figure 4: Types of sparsity in labeled data; sparsity in labeled regions (left) and images (right). Our proposed SSL+SAOD is a realistic setup that presents challenges from both kinds of sparsity." + ], + "image_footnote": [], + "bbox": [ + 498, + 109, + 890, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "serve an improvement of 1.81, 1.84 and $3.55\\mathrm{mAP}$ on the $30\\%$ , $40\\%$ and $50\\%$ sparsity levels respectively. We observe that the gap in performance increases with sparsity, consistent with our observation for Tables 1 and 2. This can be attributed to the inability of methods like Co-mining to handle unlabeled images and high sparsity in the labeled data. We will make this benchmark public and propose this method as a baseline. We encourage researchers to report results on this benchmark in the future.", + "bbox": [ + 496, + 335, + 890, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Experiments", + "text_level": 1, + "bbox": [ + 500, + 482, + 705, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section we conduct ablation experiments to understand the various components. For the ablation experiments, we use a ResNet-101 as the backbone network and train on Split-5 with $p = 50\\%$ . We evaluate on the VOC07 test set and report the COCO style AP metric.", + "bbox": [ + 496, + 507, + 890, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Table 4, a baseline model trained on the ablation set attains 41.77 (row 1). Pseudo-labeling $\\left(\\mathcal{L}_{\\mathrm{det},a}^{s}\\right)$ improves the performance by 0.9 (row 2). Co-mining [10] relies extensively on pseudo-labels. This results in a drop in performance at higher sparsities due to noisy pseudo-labels. Addition of the C-RPN improves our performance by 3.53 points (row 3). C-RPN reduces the overhead of computing two sets of proposals and learns a better notion of objects due to the combined processing of features from the two views.", + "bbox": [ + 496, + 582, + 890, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The combination of C-RPN and PPM improves the performance by 3.67 (row 4). We do not observe major improvements with the introduction of PPM because its task is to identify and segregate the unlabeled regions from the backgrounds. After the segregation, PPM does no further processing to improve performance. The power of PPM can be observed when trained in conjunction with consistency regularization loss $(\\mathcal{L}_{\\mathrm{SSL}}^{\\mathrm{u}})$ which achieves the best performance of 46 (row 7); an improvement of 4.23 points over the baseline. We distinguish ourselves from pseudolabeling approaches like Co-mining in one important as", + "bbox": [ + 496, + 734, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6776", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8e95bf0ee2bedb9f62a17d2367d73ca50a79ac74834c5b64be0b5efc81bc7e10.jpg", + "table_caption": [ + "Table 5: Analysis of the threshold used for PPM." + ], + "table_footnote": [], + "table_body": "
Threshold0.60.70.80.90.95
AP45.9645.9646.0045.7745.30
", + "bbox": [ + 107, + 126, + 433, + 193 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c137d8c98f036771c718b73ab4c316eb9ee1cc2b7508540d1e3cb4a8a69b984.jpg", + "table_caption": [ + "Table 6: Removing test time augmentations." + ], + "table_footnote": [], + "table_body": "
MethodSplit-1Split-2Split-3
30%50%70%30%50%70%30%50%70%
TTA38.2235.9232.6839.7636.9435.3339.5637.1535.48
w/o TTA38.3235.9132.6739.7736.9535.2739.5537.1235.48
", + "bbox": [ + 465, + 116, + 890, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56943a3ba02f47679770eddb80e9fb15362424e65c36d5c6ac1087733f0f1e0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 215, + 264, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44e100c8e228b1bd3e9da7960054013d85c3fd74bb048854d8cb4341632af08a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 215, + 439, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/466bca1565ecfaf5c8d0e149a98af8c30b9580b0f2b855287b34def292c489aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 215, + 612, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4908f32da58fa3d5d9cd90317e1cfdbc5e9aef842392af01a4f3e646107eb5e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 215, + 717, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b421c95f9ca47144aa642efa3bcf377d5e94516a4155002a83d06f4dbb0a39f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 215, + 890, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8e414c0a1ee5cce2a71c8b983479d7b13bbb77761ec61b795c94412bb934734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 306, + 266, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd5c22dd6d8b6849cc3bfc897911f90b5b5f9ac871e0e29d17e74b313f3df2dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 306, + 408, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ffb07e03968fe33caef460caaa04fd813f19297ac479345927f7743c05ff07d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 306, + 612, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1207c722a40a13ae5e31b049fb3f2112de99f7a0fa8ee0184cb142fd6d2fc711.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 306, + 717, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d3cfe20723706485bfb81d4cf1511fe13138de8a36f184af30c21875ac487e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 308, + 890, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9cde6433b013e88104b5cd54b69ca6d65c76cc59cbf56030a841b18b071a8736.jpg", + "image_caption": [ + "Figure 5: Qualitative results showing the unlabeled regions identified by the PPM. The red boxes correspond to the available ground truth. A class agnostic NMS was performed on the regions and the result is shown in white." + ], + "image_footnote": [], + "bbox": [ + 80, + 455, + 253, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0be5711e10edba967aaacfb2d3a05878dc5b057fa58923d6d89e05723d2a1a7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 455, + 372, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f79d2bd3b22a013b842a2c93c7d37aec3e4d53eadc8b0c475106a09f8d6d398.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 455, + 464, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f944d4b57b2e453d1015727317dcc28da07075c7c8a05b422363f8c86382d61f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 455, + 547, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ead32e35cc7a802167f31af89bba5dcfcff9310a723a77b0a6c6fe6e70b11915.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 455, + 723, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf397c441ca6cbf2a46b884debf60b9c692384341352795a76e2ee36752ab9f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 455, + 890, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12dda31c9b2760096dc90be00c9e4ec904ad9d71780bbf4fba533d6d22d766d3.jpg", + "image_caption": [ + "Figure 6: Qualitative results comparing the output of a model trained using available ground truths (top) to a model trained using our approach (bottom). Predictions with a class confidence score greater than 0.9 are shown. Red: Person, Cyan: Dog, Purple: Horse, Yellow: Clock, Green: Stop sign, Blue: Parking meter, Violet: Giraffe, Orange: Potted plant, Black: Surfboard, Dark green: Boat" + ], + "image_footnote": [], + "bbox": [ + 80, + 549, + 253, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/394fb5c3ffa80d65a890f65208469d3c7d0736f1b36c292a00689b8d0a85103a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 547, + 372, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/455ef43f4639b77690f9cecbddaa6d68070c3d0c55efdf9597f600687341aa59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 547, + 464, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c6f45432bfdcd2ee1c69c207e0c45c04b914db15b589ae5c85bd28e2f5f17c2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 547, + 547, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4239b9e8a9242fa7ea905bf76922bc5ea9853025f0230dba4f3079b0a2b9d34d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 547, + 723, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a208970510f4bb326390f91d87805ba0146fd55c5660dc19fd7b10aa0d4a9a54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 549, + 890, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pect. While, Co-mining [10] relies solely on pseudo-labels, we leverage additional components from self-supervised learning along with pseudo-labeling. In row 5, due to C-RPN, PPM and $\\mathcal{L}_{\\mathrm{SSL}}^{\\mathrm{u}}$ , we show an improvement of 3.74 over the baseline. Our proposed components are orthogonal to pseudo-labeling as using them together results in an additional improvement of $\\sim 0.5$ on mAP. At higher sparsity in the labeled dataset, the proposed C-RPN, PPM and $\\mathcal{L}_{\\mathrm{SSL}}^{\\mathrm{u}}$ are less affected than pseudo-labels resulting in the large improvements on these splits. Finally, we show the effect of C-RPN by generating a single set of proposals from the", + "bbox": [ + 75, + 720, + 472, + 888 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "original image and using it for both the branches. We observe a drop in performance (row 6) highlighting the effectiveness of C-RPN.", + "bbox": [ + 498, + 720, + 890, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PPM mines potential positives which can be mistaken for negatives due to missing annotations. We rely on the objectness score of the RPN to identify these regions. In Table 5, we vary the threshold of PPM. For a low threshold, a few hard negatives might also be identified as pseudo-positive leading to a drop in performance. With a high threshold, a few potential positives might not be mined. We observe that a threshold of 0.8 provides a good trade-off and is there-", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6777", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2f725af2a001edb0198a2edb5f0fed260b2563c928be608d34e5c196576e3f38.jpg", + "table_caption": [ + "Table 7: Comparison with a different backbone. Results are reported on the COCO validation set on Splits 1-3 using AP and on VOC 2007 test set on Splits-4,5 using $\\mathrm{AP}_{50}$ ." + ], + "table_footnote": [], + "table_body": "
MethodSplit-1Split-2Split-3Split-4Split-5
30%50%70%30%50%70%30%50%70%EasyHardExtreme30%40%50%
Ours (C4)37.6735.9533.1639.2236.8134.9838.8436.7635.2680.5680.4374.4577.3876.1375.32
Ours (FPN)38.2235.9232.6839.7636.9435.3339.5637.1535.4882.1581.5075.5976.8475.8874.35
", + "bbox": [ + 80, + 130, + 888, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "fore used for all experiments unless stated otherwise. In all our experiments, PPM is performed after an initial warmup of 9000 and 30000 iterations on the PASCAL-VOC and COCO datasets respectively.", + "bbox": [ + 75, + 229, + 470, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6.1 Inference without augmentations", + "text_level": 1, + "bbox": [ + 76, + 311, + 362, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our method requires passing an input with augmentations during inference as well. It should be noted that this is not a test-time augmentation (TTA), a technique that typically involves passing images at a higher resolution. We perform inference by removing the augmentation and extracting the region proposals using C-RPN. We show the results in Table 6 on the three splits of COCO. We do not observe a significant improvement in performance due to the augmentation. For the analysis on the effect of augmentations refer to supplementary material.", + "bbox": [ + 75, + 337, + 470, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6.2 Effect of backbone", + "text_level": 1, + "bbox": [ + 76, + 508, + 261, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We analyze the effect of backbone on our approach and show results for the normal convolution based (C4) and FPN based (FPN) backbones in Table 7. The first row corresponds to C4 while the second row corresponds to FPN. While the gap in performance is lower in most cases, we observe a significant improvement using the FPN on the low sparsity settings of all the split.", + "bbox": [ + 75, + 532, + 468, + 642 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.7. RPN Recall experiments for object discovery", + "text_level": 1, + "bbox": [ + 76, + 650, + 455, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "PPM identifies foreground regions mistakenly assigned as background during training to avoid penalizing the network. To study the effect of PPM on novel [48] classes, we train a network using our approach on randomly chosen 6000 images of the COCO dataset, containing annotations for only 20 classes of PASCAL-VOC. We evaluate the recall@0.5 of the RPN over the remaining 60 classes. A model trained using the standard technique on this dataset achieves a recall@0.5 of $77.46\\%$ and $29.06\\%$ on the known classes (20 categories) and unknown classes (60 categories) respectively. Our proposed approach, with PPM, achieves a recall@0.5 of $78.47\\%$ and $35.20\\%$ respectively. This ability to localize objects not seen during training can be beneficial for object discovery methods like [49-52] which use RPN proposals to learn/discover new categories.", + "bbox": [ + 75, + 672, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.8. Qualitative Results", + "text_level": 1, + "bbox": [ + 498, + 229, + 683, + 246 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Figure 5, we show the pseudo positives mined by PPM. In each figure, the red boxes correspond to the ground truth annotations and the white boxes correspond to the post NMS pseudo positive boxes mined by PPM. We observe that the PPM correctly mines proposals which correspond to missing object annotations. Without PPM, these regions will be used as negatives to the classifier resulting in a reduced class confidence score leading to a drop in mAP. We show the detection results of a model trained on the $50\\%$ Split-1 of our approach in Figure 6. The images on the top corresponds to the model trained using sparse annotations and the bottom image shows the output of our approach.", + "bbox": [ + 496, + 256, + 890, + 439 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 459, + 617, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present SparseDet, a novel end-to-end system for Sparsely Annotated Object Detection (SAOD). We propose a simple yet effective technique for identifying the unlabeled regions using pseudo-positive mining and apply self-supervised loss on them. Through qualitative results we highlight the ability of PPM to mine pseudo-positives. We standardize the evaluation setup and show the effectiveness of our approach with an exhaustive set of experiments on multiple splits of SAOD. While we outperform existing state-of-the-art on all metrics and splits, we observe the gap in performance increases with sparsity demonstrating the short coming of methods that rely solely on pseudolabeling. We propose a new benchmark, that evaluates the semi-supervised learning capabilities of SAOD approaches. We will release the data for the new benchmark along with all the SAOD splits and encourage researchers to evaluate future SAOD methods on these.", + "bbox": [ + 496, + 487, + 890, + 743 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This project was partially funded by DARPA SAIL-ON (W911NF2020009), DARPA SemaFor (HR001119S0085), and IARPA SMART (2021-21040700003) programs. Rama Chellappa was supported by an ONR MURI Grant N00014-20-1-2787. We would like to thank our colleagues from the PI lab for their feedback on this work. The views and conclusions contained herein are those of the authors and do not represent views, policies, and/or endorsements of funding agencies.", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6778", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Simon Chadwick and Paul Newman. Training object detectors with noisy data. 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1319-1325, 2019. 1", + "[2] Zhe Wu, Navaneeth Bodla, Bharat Singh, Mahyar Najibi, Rama Chellappa, and Larry S. Davis. Soft sampling for robust object detection. In BMVC, 2019. 1, 2, 3, 5", + "[3] Junnan Li, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Towards noise-resistant object detection with noisy annotations. ArXiv, abs/2003.01285, 2020. 1", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1", + "[5] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 2, 5", + "[6] Alina Kuznetsova, Hassan Rom, Neil Gordon Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4. International Journal of Computer Vision, 128:1956-1981, 2020. 1", + "[7] Yusuke Niitani, Takuya Akiba, Tommi Kerola, Toru Ogawa, Shotaro Sano, and Shuji Suzuki. Sampling techniques for large-scale object detection from sparsely annotated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6510-6518, 2019. 1, 2, 5, 6", + "[8] Han Zhang, Fangyi Chen, Zhiqiang Shen, Qiqi Hao, Chenchen Zhu, and Marios Savvides. Solving missing-annotation object detection with background recalibration loss. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1888-1892. IEEE, 2020. 3, 5, 6", + "[9] Yuewei Yang, Kevin J Liang, and Lawrence Carin. Object detection as a positive-unlabeled problem. In BMVC, 2020. 5", + "[10] Tiancai Wang, Tong Yang, Jiale Cao, and X. Zhang. Comining: Self-supervised learning for sparsely annotated object detection. In AAAI, 2021. 1, 2, 3, 5, 6, 7, 8", + "[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2007 (VOC2007) Results. http://www.pascalnetwork.org/challenges/VOC/voc2007/workshop/index.html, 2, 5", + "[12] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, 2, 5, 6" + ], + "bbox": [ + 78, + 114, + 472, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Ke Yan, Jinzheng Cai, Adam P. Harrison, Dakai Jin, Jing Xiao, and Le Lu. Universal lesion detection by learning from multiple heterogeneously labeled datasets. ArXiv, abs/2005.13753, 2020. 2", + "[14] Hansheng Li, Xin Han, Yuxin Kang, Xiaoshuang Shi, Mengdi Yan, Zixu Tong, Qirong Bu, Lei Cui, Jun Feng, and Lin Yang. A novel loss calibration strategy for object detection networks training on sparsely annotated pathological datasets. In MICCAI, 2020.", + "[15] Yongqiang Zhang, Mingli Ding, Yancheng Bai, Mengmeng Xu, and Bernard Ghanem. Beyond weakly supervised: Pseudo ground truths mining for missing bounding-boxes object detection. IEEE Transactions on Circuits and Systems for Video Technology, 30:983-997, 2020.", + "[16] Ke Yan, Jinzheng Cai, Youjing Zheng, Adam P. Harrison, Dakai Jin, You-Bao Tang, Yuxing Tang, Lingyun Huang, Jing Xiao, and Le Lu. Learning from multiple datasets with heterogeneous and partial labels for universal lesion detection in ct. IEEE Transactions on Medical Imaging, 40:2759-2770, 2021. 2", + "[17] Jisoo Jeong, Seungeui Lee, Jeesoo Kim, and Nojun Kwak. Consistency-based semi-supervised learning for object detection. Advances in neural information processing systems, 32:10759-10768, 2019. 2", + "[18] Peng Tang, Chetan Ramaiah, Yan Wang, Ran Xu, and Caiming Xiong. Proposal learning for semi-supervised object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2291-2301, 2021. 2", + "[19] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2", + "[20] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. arXiv preprint arXiv:2102.09480, 2021.", + "[21] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. arXiv preprint arXiv:2106.09018, 2021.", + "[22] Qize Yang, Xihan Wei, Biao Wang, Xian-Sheng Hua, and Lei Zhang. Interactive self-training with mean teachers for semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5941-5950, 2021. 2", + "[23] Jihun Yoon, Seungbum Hong, and Min-Kook Choi. Semi-supervised object detection with sparsely annotated dataset. In 2021 IEEE International Conference on Image Processing (ICIP), pages 719-723. IEEE, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6779", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097-1105. Curran Associates, Inc., 2012. 3", + "[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. doi: 10.1007/s11263-015-0816-y. 3", + "[26] Ross Girshick. Fast r-cnn. In Proceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), ICCV '15, pages 1440-1448, Washington, DC, USA, 2015. IEEE Computer Society. ISBN 978-1-4673-8391-2. doi: 10.1109/ICCV.2015.169. URL http://dx.doi.org/10.1109/ICCV.2015.169.3", + "[27] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 3", + "[28] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 3, 4, 6", + "[29] Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. R-fcn: Object detection via region-based fully convolutional networks. In Daniel D. Lee, Masashi Sugiyama, Ulrike von Luxburg, Isabelle Guyon, and Roman Garnett, editors, NIPS, pages 379-387, 2016. URL http://dblp.uni-trier.de/db/conf/nips/nips2016.html#DaiLHS16.3", + "[30] Tsung-Yi Lin, Piotr Dólár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 936-944, 2017. 3, 6", + "[31] Bharat Singh and Larry S Davis. An analysis of scale invariance in object detection-snip. CVPR, 2018.", + "[32] Bharat Singh, Mahyar Najibi, and Larry S Davis. SNIPER: Efficient multi-scale training. NeurIPS, 2018.", + "[33] Mahyar Najibi, Bharat Singh, and Larry S Davis. AutoFocus: Efficient multi-scale inference. ICCV, 2019. 3", + "[34] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. doi: 10.1109/ICCV.2017.322.3", + "[35] Alexander Kirillov, Ross B. Girshick, Kaiming He, and Piotr Dóllár. Panoptic feature pyramid networks. In CVPR, pages 6399-6408. Computer Vision Foundation / IEEE, 2019. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In CVPR, pages 9404-9413. Computer Vision Foundation / IEEE, 2019. 3", + "[37] Justin Johnson Georgia Gkioxari, Jitendra Malik. Mesh r-cnn. In ICCV, 2019. 3", + "[38] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 3", + "[39] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017.", + "[40] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection, 2015. URL http://arxiv.org/abs/1506.02640. cite arxiv:1506.02640.3", + "[41] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 3", + "[42] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 3", + "[43] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:386-397, 2020. 3", + "[44] Ross B. Girshick. Fast R-CNN. 2015 IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 4", + "[45] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. doi: 10.1109/CVPR.2016.90.6", + "[46] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6", + "[47] Hanjun Li, Xingjia Pan, Ke Yan, Fan Tang, and Wei-Shi Zheng. Siod: Single instance annotated per category per image for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14197-14206, 2022. 7" + ], + "bbox": [ + 501, + 90, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6780", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Akshay Raj Dhamija, Manuel Günther, Jonathan Ventura, and Terrance E. Boult. The overlooked elephant of object detection: Open set. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1010-1019, 2020. doi: 10.1109/WACV45572.2020.9093355. 9", + "[49] Huy V. Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 9", + "[50] Huy V. Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 8287-8296, 2019.", + "[51] Sai Saketh Rambhatla, Rama Chellappa, and Abhinav Shrivastava. The pursuit of knowledge: Discovering and localizing novel categories using dual memory. In Proceedings of the International Conference on Computer Vision (ICCV), 2021.", + "[52] Sai Saketh Rambhatla, Ishan Misra, Rama Chellappa, and Abhinav Shrivastava. The pursuit of knowledge: Discovering and localizing novel categories using dual memory. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 9" + ], + "bbox": [ + 78, + 90, + 468, + 477 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6781", + "bbox": [ + 482, + 945, + 511, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/5aaa9a89-cdc6-4e0a-94f3-a3b82685ba6a_model.json b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/5aaa9a89-cdc6-4e0a-94f3-a3b82685ba6a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b245579d39b9ac4bda8490723bd8b15007268897 --- /dev/null +++ b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/5aaa9a89-cdc6-4e0a-94f3-a3b82685ba6a_model.json @@ -0,0 +1,2622 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.815, + 0.177 + ], + "angle": 0, + "content": "SparseDet: Improving Sparsely Annotated Object Detection with Pseudo-positive Mining" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.204, + 0.251, + 0.221 + ], + "angle": 0, + "content": "Saksham Suri\\*1" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.226, + 0.275, + 0.238 + ], + "angle": 0, + "content": "sakshams@cs.umd.edu" + }, + { + "type": "text", + "bbox": [ + 0.314, + 0.204, + 0.471, + 0.221 + ], + "angle": 0, + "content": "Saketh Rambhatla*1" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.226, + 0.466, + 0.238 + ], + "angle": 0, + "content": "rssaketh@umd.edu" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.204, + 0.653, + 0.222 + ], + "angle": 0, + "content": "Rama Chellappa2" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.226, + 0.655, + 0.239 + ], + "angle": 0, + "content": "rchella4@jhu.edu" + }, + { + "type": "text", + "bbox": [ + 0.695, + 0.204, + 0.864, + 0.222 + ], + "angle": 0, + "content": "Abhinav Shrivastava" + }, + { + "type": "text", + "bbox": [ + 0.699, + 0.226, + 0.86, + 0.238 + ], + "angle": 0, + "content": "abhinav@cs.umd.edu" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.246, + 0.513, + 0.265 + ], + "angle": 0, + "content": "University of Maryland, College Park1" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.246, + 0.763, + 0.266 + ], + "angle": 0, + "content": "Johns Hopkins University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.299, + 0.314, + 0.315 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.474, + 0.665 + ], + "angle": 0, + "content": "Training with sparse annotations is known to reduce the performance of object detectors. Previous methods have focused on proxies for missing ground truth annotations in the form of pseudo-labels for unlabeled boxes. We observe that existing methods suffer at higher levels of sparsity in the data due to noisy pseudo-labels. To prevent this, we propose an end-to-end system that learns to separate the proposals into labeled and unlabeled regions using Pseudo-positive mining. While the labeled regions are processed as usual, self-supervised learning is used to process the unlabeled regions thereby preventing the negative effects of noisy pseudo-labels. This novel approach has multiple advantages such as improved robustness to higher sparsity when compared to existing methods. We conduct exhaustive experiments on five splits on the PASCAL-VOC and COCO datasets achieving state-of-the-art performance. We also unify various splits used across literature for this task and present a standardized benchmark. On average, we improve by 2.6, 3.9 and 9.6 mAP over previous state-of-the-art methods on three splits of increasing sparsity on COCO. Our project is publicly available at cs.umd.edu/~sakshams/SparseDet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.697, + 0.21, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.875 + ], + "angle": 0, + "content": "The performance of object detectors is sensitive to the quality of labeled data [1-3]. Existing object detection methods assume that the training data is pristine and a drop in performance is observed if this assumption fails. Noise in the data used for training object detectors can arise due to incorrect class labels or incorrect/missing bounding boxes. In this work, we deal with the problem of training object detectors with sparse annotations, i.e., missing region or bounding boxes. This problem is of utmost importance, as obtaining crowd-sourced datasets [4, 5] can be expensive" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.304, + 0.892, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.481, + 0.895, + 0.602 + ], + "angle": 0, + "content": "Figure 1: (Top) Most Object Detection datasets have exhaustive annotations for foreground/positives. During training, the unlabeled regions can be safely considered as background/negatives. Sparsely Annotated Object Detection datasets (bottom) have missing annotations. This results in foreground regions (shown in red) being considered as negatives during training, adversely affecting the performance of the classifier." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.894, + 0.704 + ], + "angle": 0, + "content": "and laborious. The alternative is to use computer-assisted protocols to collect annotations which have been shown to be noisy and incomplete [6]. This problem of training object detectors with incomplete bounding box annotations is called Sparsely Annotated Object Detection (SAOD)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To understand why training with sparse annotations is detrimental to the performance, consider the example shown in Figure 1 (top). If the annotation were exhaustive, then the negative samples to the classifier contain true background regions. But with sparse annotations, as shown in Figure 1 (bottom), a few positive regions will inevitably be considered as negatives (shown in red), thereby wrongfully penalizing the classifier leading to lower performance. Existing methods[2, 7-10] prevent this by predicting pseudolabels and removing the foreground regions from the negatives to the classifier. However, at higher levels of sparsity, the quality of pseudo-labels is greatly affected, resulting in the same problem noted above." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.298, + 0.901 + ], + "angle": 0, + "content": "*First two authors contributed equally" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6770" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "Crowd sourced object detection datasets [5, 11, 12] are ensured to be almost exhaustively labeled. Hence, for SAOD, researchers artificially create sparsely annotated datasets from the original ones. There is no general consensus on the correct way to create the sparsely annotated datasets, a.k.a. splits, and hence each method reports results on one or two different splits. Split can be created by considering the dataset as a whole or per image (i.e. annotations can be removed by considering all the images or only a single image at once). They can also be created by removing annotations in a class-agnostic or class-aware fashion (i.e. remove \\( p\\% \\), of annotations per category or across all the categories). These variations result in splits with different data distributions making some harder than the others (refer to Table 1). A proper benchmark that analyzes the performance of SAOD methods across these different types of splits is missing. This makes it difficult to compare methods and assess their effectiveness for a specific use case." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.761 + ], + "angle": 0, + "content": "To tackle the issues discussed above, as our first major contribution, we present SparseDet, a novel SAOD framework that achieves state-of-the-art performance across multiple SAOD benchmarks in practice. SparseDet operates on an image and its augmented counterpart. The combination of features extracted from the two views is used to generate region proposals. Standard detection training methods, consider a region proposal as positive if its intersection over union (IoU) with any ground truth is greater than 0.5 and the rest are treated as negatives. This strategy works when the annotations are exhaustive, which implies that the remaining regions are from background. But due to missing annotations, some of these regions could belong to foreground instances. To prevent considering all region proposals without annotations as negatives, SparseDet partitions all the region proposals into labeled, unlabeled and background. The labeled and background regions are processed as usual. Features extracted from unlabeled regions are then trained with a self-supervised loss. Previous approaches like Co-Mining [10], consider two partitions, labeled and background, and generate pseudo-labels. This is a disadvantage at high sparsity as the generated pseudo-labels can be very noisy. The self-supervised loss in our approach enforces consistency between the features of the two views for the unlabeled regions and prevents penalizing the classifier due to false negatives." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.765, + 0.472, + 0.904 + ], + "angle": 0, + "content": "Our second major contribution is unifying evaluation. The standard practice is to simulate sparsely annotated training data on COCO [5] and PASCAL-VOC [11] train sets and evaluate them on their corresponding standard validation set. As discussed above, a survey of recent SAOD approaches [2, 7-10, 13-16] reveals that there are at least five different ways to create splits, each differing in the strategy (refer to Section 4.2) used to achieve the desired level of sparsity. However, these splits have not been made public," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.319 + ], + "angle": 0, + "content": "making it difficult to replicate results for comparison. Additionally, each of these strategies has a different property for simulating sparse data, e.g., different distribution of annotations per class, resulting in a different training set. As a result, methods trained on different splits cannot be compared with one another. To mitigate these issues we standardize the generation of these splits that enables the evaluation of any SAOD methods on all of them for fair comparison. Additionally, we propose a new benchmark that assesses the semi-supervised learning capabilities of SAOD methods, i.e., leveraging unlabeled data to improve performance. We present our approach as a baseline. We will make the data for the new benchmark along with all the SAOD splits public to facilitate future research. We briefly summarize our contributions below." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.328, + 0.892, + 0.389 + ], + "angle": 0, + "content": "- We propose a novel formulation, SparseDet, for SAOD which is an end-to-end approach that identifies labeled, unlabeled and background regions and deals with them in appropriate manner." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.395, + 0.892, + 0.47 + ], + "angle": 0, + "content": "- We show state-of-the-art performance on sparsely annotated object detection across various splits. On average, we improve by 2.6, 3.9 and \\(9.6\\mathrm{mAP}\\) over previous state-of-the-art methods on three splits of increasing sparsity on COCO." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.475, + 0.892, + 0.55 + ], + "angle": 0, + "content": "- We standardize the experimental setup for SAOD by evaluating methods on all the splits to facilitate future research. Additionally, we propose a new benchmark that evaluates the semi-supervised learning capabilities of SAOD methods." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.328, + 0.892, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.895, + 0.622 + ], + "angle": 0, + "content": "In Section 2, we discuss the related works on SAOD and related fields. We describe our approach in detail in Section 3. We describe our experimental setup and present results in Section 4, and conclude in Section 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.642, + 0.65 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Semi-supervised object detection: Semi-supervised object detection (SSOD) is an active field that also deals with training object detectors with missing annotations. Existing works on semi-supervised object detection, have focused mainly on consistency regularization [17, 18] or pseudolabeling-based approaches [19-22]. The main idea behind these approaches is to perturb the images, or features, and apply a consistency regularization loss to enforce consistency between the predictions using a student teacher framework. However, typical SSOD methods assume a small exhaustively labeled set and a large unlabeled set for training. This is different from Sparsely annotated object detection (SAOD), which assumes a large training set which is sparsely labeled." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Sparsely annotated object detection: One of the initial works addressing this problem, by Niitani et al. [7], pro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6771" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "poses utilizing logical relations between the co-occurrences of objects and pseudo-labeling. Yoon et al. [23] use object tracking to densely label objects across sparsely annotated frames along with single stage detectors to mitigate the negative effects of missing annotations. Wu et al. [2] propose a re-weighting approach where the gradients corresponding to region of interest are weighed as a function of overlap with ground truth instances. Improving upon the previous work, Zhang et al. [8] propose an automatic re-calibration strategy for single stage detectors where the negative branch is changed to take into account low confidence background predictions which might correspond to missing annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.473, + 0.385 + ], + "angle": 0, + "content": "Finally, one of the most recent works, Co-mining [10] uses a co-training strategy by using two views of an image and predictions from one view along with the ground truth as supervision for the other view and vice versa. Our method doesn't solely rely on pseudo-labels and leverages a self-supervised loss to prevent propagating negative gradients to the model due to false negatives." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Fully supervised object detection: After the success of AlexNet [24] on the image classification challenge (ILSVRC 2012) [25], research on designing deep neural networks for object detection gathered more interest. First among the successful methods are the two-stage Region-based convolutional networks (R-CNN) family of detectors. Two-stage object detectors consists of 1) a region proposal stage which produces a set of candidate object bounding boxes followed by 2) a classification stage which classifies each candidate region as either belonging to a foreground object or \"background\". R-CNN processes a large amount of region proposals by cropping the input image and using a CNN backbone to extract features making it extremely slow. Fast R-CNN [26] was proposed to overcome this limitation. Fast R-CNN computes one convolution feature map for the whole image. RoI Pooling was introduced to pool the feature for each region of interest into a fixed spatial dimension. RoI pooling shares the computation among all the region proposals speeding up training and inference. Fully connected layers are applied on the fixed RoI pooled feature maps which are then passed to two sister heads, for classification and bounding box regression. The whole network is trained end-to-end with a multi-task loss avoiding the multi-stage training in R-CNN [27]. While Fast R-CNN improved the efficiency of its predecessor, test time computation bottleneck is still an issue because of the region proposal method employed. Faster R-CNN [28] proposed a region proposal network (RPN), an elegant solution that trains deep networks to predict region proposals for practically no additional computational overhead. RPN shares convolutional layers with Fast R-CNN [26] and at test time the cost of generating proposals is minimal. Faster R-CNN paved the way for more sophisticated and efficient two stage detection architectures [29-33]. Faster R-CNN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "has also been extended to achieve state-of-the-art instance segmentation [34], panoptic segmentation [35, 36], and 3D mesh generation [37] etc. A few limitations of two stage object detectors and anchor boxes has also inspired the family of single-stage [38-40] and anchor-free object detection systems [41, 42]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.194, + 0.611, + 0.212 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.895, + 0.359 + ], + "angle": 0, + "content": "We present SparseDet for Sparsely Annotated Object Detection. Given \\(N\\) images, we denote the set of labeled regions in the dataset as \\(\\mathcal{B}_{\\mathrm{l}} = \\{b_{i},c_{i}\\}_{i = 1}^{\\mathrm{N}_{\\mathrm{l}}}\\) where \\((b_{*},c_{*})\\) are the bounding box and class labels respectively. The unlabeled regions are denoted as \\(\\mathcal{B}_{\\mathrm{ul}} = \\{b_k\\}_{k = 1}^{\\mathrm{N}_{\\mathrm{ul}}}\\) and the background regions as \\(\\mathcal{B}_{\\mathrm{bg}} = \\{b_q\\}_{q = 1}^{\\mathrm{N}_{\\mathrm{bg}}}\\). Note that unlike existing SAOD approaches, we do not assume images to contain at least one labeled region in this work. The unlabeled \\(\\mathcal{B}_{\\mathrm{ul}}\\) and background \\(\\mathcal{B}_{\\mathrm{bg}}\\) sets are not known a-priori." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.364, + 0.611, + 0.379 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.645 + ], + "angle": 0, + "content": "The proposed approach is shown in Figure 2 and consists of a backbone network that extracts features from the original and augmented views of an image. The common RPN (C-RPN), concatenates the features, to generate a set of region proposals. A region proposal \\( b \\) can belong to one of three groups, namely 1) labeled regions \\( b \\in \\mathcal{B}_1 \\), 2) unlabeled foreground regions \\( b \\in \\mathcal{B}_{\\mathrm{ul}} \\), or 3) background regions \\( b \\in \\mathcal{B}_{\\mathrm{bg}} \\). For a given set of ground-truth annotations, the first group, i.e. labeled regions can be automatically identified. The problem then reduces to identifying and separating the second group i.e. the unlabeled regions, from the background regions. Given all the region proposals, a pseudo-positive mining (PPM) step identifies the unlabeled regions and segregates them from the background regions. The labeled and unlabeled regions are trained using supervised and self-supervised losses respectively. We describe each stage in detail below." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.651, + 0.684, + 0.667 + ], + "angle": 0, + "content": "3.2. Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.894, + 0.767 + ], + "angle": 0, + "content": "Given an image \\(I\\), an augmented version of \\(I\\) denoted as \\(\\mathcal{A}(I)\\) is computed. In this work, we use random contrast, brightness, saturation, lighting and bounding box erase in a cascaded fashion to generate \\(\\mathcal{A}(I)\\). A backbone network is employed to extract two features, \\(f_{\\mathrm{o}}\\) and \\(f_{\\mathrm{a}}\\), from \\(I\\) and \\(\\mathcal{A}(I)\\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.772, + 0.722, + 0.788 + ], + "angle": 0, + "content": "3.3. Common RPN (C-RPN)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Two stage object detectors [28, 30, 43] use a region proposal network (RPN) [28] to generate regions of interest (RoIs). We propose C-RPN which concatenates \\( f_{\\mathrm{o}} \\) and \\( f_{\\mathrm{a}} \\) to obtain the RoIs. This is different from existing approaches that use \\( f_{\\mathrm{o}} \\) and \\( f_{\\mathrm{a}} \\) to generate two separate sets of RoIs. Operating on two sets increases the difficulty of identifying the labeled, unlabeled and background regions which" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6772" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.096, + 0.852, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Figure 2: Illustration of SparseDet for sparsely annotated object detection. Following feature extraction from the original and augmented image, a common set of proposals is generated by the common RPN (C-RPN). Using an end to end approach, we identify and mine proposals corresponding to missing annotations using pseudo positive mining (PPM). We train the network end-to-end using a combination of supervised and self-supervised losses. The unlabeled instances (black) are supervised with a self-supervised loss and the labeled instances (green) are supervised with ground truth annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.471, + 0.424 + ], + "angle": 0, + "content": "are required for subsequent stages. The sparse ground truth is used as supervision to train the C-RPN with a binary cross entropy loss for foreground classification and smooth L1 [44] loss for bounding box regression as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.448, + 0.47, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r p n}} ^ {\\mathrm {s}} \\left(\\mathbf {c}, \\mathbf {b}, \\mathbf {c} ^ {*}, \\mathbf {b} ^ {*}\\right) = \\mathcal {L} _ {\\mathrm {b c e}} \\left(\\mathbf {c}, \\mathbf {c} ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {r e g}} \\left(\\mathbf {b}, \\mathbf {b} ^ {*}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.475, + 0.47, + 0.506 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}\\), \\(\\mathbf{b}\\) are the objectness scores and bounding boxes, and \\(\\mathbf{c}^*\\), \\(\\mathbf{b}^*\\) are the corresponding ground truth." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.349, + 0.53 + ], + "angle": 0, + "content": "3.4. Pseudo Positive Mining (PPM)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.471, + 0.794 + ], + "angle": 0, + "content": "Given the RoIs from C-RPN, the next step is to segregate the unlabeled regions from the labeled and background regions. A standard practice in training detectors is to consider all proposals with an objectness score greater than \\(\\tau_{\\mathrm{obj}}\\) and IoU greater \\(\\tau_{\\mathrm{fg}}\\) with any ground truth as foreground and proposals with IoU less than \\(\\tau_{\\mathrm{bg}}\\) as background. In the presence of missing annotations, detectors are wrongly penalized because the IoU of RoIs corresponding to missing annotations with any ground truth is less than \\(\\tau_{\\mathrm{bg}}\\). To avoid this, we mine \"potential\" foregrounds and consider them as unlabeled regions \\((\\mathcal{B}_{\\mathrm{u}})\\). The proposed PPM module is based on our observation that when trained with sparse annotations, the RPN can reliably distinguish foreground from background regions. We pick all RoIs that have an objectness score greater than \\(\\tau_{\\mathrm{ppm}}\\) and IoU less than \\(\\tau_{\\mathrm{bg}}\\) with any ground truth as the unlabeled regions. The remaining RoIs are assigned as background." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.165, + 0.817 + ], + "angle": 0, + "content": "3.5. Losses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The pseudo positive mining step segregates the RoIs into labeled, unlabeled and background regions. An RoI pooling layer [28] extracts RoI pooled features for the labeled and background regions using the feature \\( f_{0} \\). The detection head processes the RoI pooled features to predict class-wise" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.424 + ], + "angle": 0, + "content": "probabilities and bounding box adjustments for each region. The ground truth is used to supervise the predictions by applying the cross entropy loss for classification and smooth L1 [44] for bounding box regression as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.429, + 0.892, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\det , 0} ^ {s} \\left(\\mathbf {c}, \\mathbf {b}, \\mathbf {c} ^ {*}, \\mathbf {b} ^ {*}\\right) = \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {c}, \\mathbf {c} ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {r e g}} \\left(\\mathbf {b}, \\mathbf {b} ^ {*}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.894, + 0.497 + ], + "angle": 0, + "content": "where \\( \\mathbf{c} \\), \\( \\mathbf{b} \\) are the class and bounding box predictions, and \\( \\mathbf{c}^* \\), \\( \\mathbf{b}^* \\) are the corresponding ground truth class labels and bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.619 + ], + "angle": 0, + "content": "All detections with a score greater than \\(\\tau_{\\mathrm{m}}\\) are combined with the ground truth followed by a class specific NMS to obtain the merged ground truth. It is ensured that no ground truth annotation is discarded during this step. The labeled and background regions along with \\(f_{\\mathrm{a}}\\) are used to extract RoI pooled features which are then fed to the detector head. The predictions of the detector head are supervised with the merged ground truth with the following losses:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.625, + 0.892, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\det , \\mathrm {a}} ^ {\\mathrm {s}} (\\mathbf {c}, \\mathbf {b}, \\mathbf {c} ^ {*}, \\mathbf {b} ^ {*}) = \\mathcal {L} _ {\\mathrm {b c e}} \\left(\\mathbf {c}, \\mathbf {c} _ {\\mathbf {m}} ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {r e g}} (\\mathbf {b}, \\mathbf {b} _ {\\mathbf {m}} ^ {*}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.891, + 0.693 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}\\), \\(\\mathbf{b}\\) are the class and bounding box predictions from the detector head, and, \\(\\mathbf{c}_{\\mathbf{m}}^{*}, \\mathbf{b}_{\\mathbf{m}}^{*}\\) are the corresponding merged ground truths." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.892, + 0.812 + ], + "angle": 0, + "content": "Finally, a class agnostic NMS is performed on the unlabeled regions \\(\\mathcal{B}_{\\mathrm{u}}\\) (obtained after PPM described in Sec. 3.4). The unlabeled regions along with \\(f_{\\mathrm{o}}\\) and \\(f_{\\mathrm{a}}\\) are passed through the ROI pooling layer and the detection head to obtain \\(f_{\\mathrm{o}}^{\\mathrm{d}}\\) and \\(f_{\\mathrm{a}}^{\\mathrm{d}}\\) respectively. A self-supervised loss is applied that enforces the detection head features of the original and augmented regions to be consistent with each other as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.819, + 0.892, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S S L}} ^ {\\mathrm {u}} \\left(f _ {\\mathrm {a}} ^ {\\mathrm {d}}, f _ {\\mathrm {o}} ^ {\\mathrm {d}}\\right) = \\| f _ {\\mathrm {a}} ^ {\\mathrm {d}} - f _ {\\mathrm {o}} ^ {\\mathrm {d}} \\| _ {2} ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.844, + 0.892, + 0.873 + ], + "angle": 0, + "content": "The network is trained end-to-end with both supervised and unsupervised losses as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.88, + 0.892, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = 0. 5 * \\left(\\mathcal {L} _ {\\text {d e t , o}} ^ {\\mathrm {s}} + \\mathcal {L} _ {\\text {d e t , a}} ^ {\\mathrm {s}}\\right) + \\mathcal {L} _ {\\mathrm {r p n}} ^ {\\mathrm {s}} + \\mathcal {L} _ {\\mathrm {S S L}} ^ {\\mathrm {u}} \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6773" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.121 + ], + "angle": 0, + "content": "Table 1: Sparsely annotated object detection results on three splits of COCO dataset. \"Oracle\" corresponds to training models using all annotations. Results are reported on the COCO validation set using AP[0.50:0.95]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.131, + 0.89, + 0.254 + ], + "angle": 0, + "content": "
MethodSplit-1Split-2Split-3100%
30%50%70%30%50%70%30%50%70%
Oracle40.91
Pseudo Label [7]-27.50--------
BRL [8]-32.70--------
Co-mining [10]36.3532.8424.9336.7233.0424.8336.7632.5424.96-
Ours38.2235.9232.6839.7636.9435.3339.5637.1535.48-
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.473, + 0.435 + ], + "angle": 0, + "content": "Discussion: In the absence of ground truth annotations, i.e. for completely unlabeled images, the supervised losses \\((\\mathcal{L}_{\\mathrm{det,o}}^{\\mathrm{s}},\\mathcal{L}_{\\mathrm{det,a}}^{\\mathrm{s}},\\mathcal{L}_{\\mathrm{rpn}}^{\\mathrm{s}})\\) cannot be computed. Our proposed approach leverages self-supervised consistency loss that does not need ground truth. This helps our approach leverage these unlabeled regions unlike contemporary SAOD methods. We claim that this is the first method to use self-supervised losses for SAOD. Even though Co-Mining [10] claims to use self-supervised learning it is technically co-learning. Ours is the first approach to use pseudo-labeling and a self-supervised loss to handle the sparse annotations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.447, + 0.315, + 0.464 + ], + "angle": 0, + "content": "4. Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.471, + 0.563 + ], + "angle": 0, + "content": "In this section, we describe the experiments to evaluate our proposed approach. In Sections 4.1 and 4.2, we describe data, splits, and metrics. In Section 4.3, we mention the implementation details followed by the baselines in Section 4.4. We compare with contemporary methods in Section 4.5, followed by an ablation study in Section 4.6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.248, + 0.587 + ], + "angle": 0, + "content": "4.1. Data and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.471, + 0.822 + ], + "angle": 0, + "content": "We conduct all our experiments on the COCO [5] and PASCAL-VOC [11, 12] (2007+2012) datasets. The COCO [5] dataset consists of 118000 and 5000 images for training and validation respectively. Experiments on the PASCAL-VOC07 [11] are conducted on 5011 trainval images and performance is computed on 4952 images of the test set. The PASCAL-VOC12 version consists of 11530 (trainval) images for training and evaluation is done on the PASCAL-VOC07 test set. Following past literature [2, 7-10], we create five different splits (Section 4.2) and report results on them. For splits on the COCO [5] dataset, we use the standard COCO style Average Precision (AP). For splits on the PASCAL-VOC [11] dataset, we use the standard PASCAL-VOC style \\(\\mathrm{AP}_{50}\\), which is Average Precision computed at an IoU threshold of 0.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.212, + 0.849 + ], + "angle": 0, + "content": "4.2. SAOD Splits" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "An extensive review of sparsely annotated object detection methods reveals that there are at least five popular types of splits in use for creating training data. Most methods re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.892, + 0.295 + ], + "angle": 0, + "content": "port results on a subset of these making it harder to compare across methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.342 + ], + "angle": 0, + "content": "We standardize the evaluation of SAOD methods by evaluating exhaustively on all the splits facilitating future research. We briefly describe the splits below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Split-1 [9, 10]: In this split, for each object category, \\( p\\% \\) of annotations are randomly removed from the COCO [5] train set and results are reported on the validation set. This split simulates sparsity at dataset level in a class aware fashion. Note, this split can contains images with no annotations. We experiment with \\( p = \\{30, 50, 70\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Split-2: For each image in the COCO [5] train set, all annotations of \\( p\\% \\) of all categories in that image removed and results are reported on the COCO validation set. This split can be considered as image level and class aware. We experiment with \\( p = \\{30, 50, 70\\} \\). Note, this split might contain images with no annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.599 + ], + "angle": 0, + "content": "Split-3: This split, which uses the COCO 2017 [5] train set for training and the validation set for evaluation, deletes \\( p\\% \\) annotations in a class agnostic fashion for each image ensuring at least one annotation. This split is image level but class agnostic. For the experiments, we use \\( p = \\{30, 50, 70\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Split-4 [8, 10]: This split requires evaluating models on three different settings namely easy, hard and extreme ensuring at least one annotation per image constructed using PASCAL-VOC 2007+12 [11, 12] trainval set. Results are reported on the PASCAL-VOC 2007 [11] test set. For each image in the training set, the easy setting randomly removes one annotation while the hard setting randomly removes half of the annotations. The extreme setting retains only one annotation per image. All the sets ensure each image consists of atleast one annotation. This split is a small scale version for an image level class agnostic split." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Split-5 [2]: This split uses the PASCAL-VOC 2007 [11] train set for training and the PASCAL-VOC 2007 test set for evaluation and drops \\( p\\% \\) annotations per class. For each image in this construction, instances of randomly selected categories are exhaustively annotated while the remaining categories do not have any annotations. This is a small scale version of Split-1 as annotations are dropped in a class aware manner across the full dataset maintaining at least 1 annotation per image. In this case we use \\( p = \\{30,40,50\\} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6774" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.571, + 0.136 + ], + "angle": 0, + "content": "Table 2: Sparsely annotated object detection results on two splits of PASCAL-VOC dataset. \"Oracle\" corresponds to training models using all annotations. Results are reported on VOC 07 test set using \\(\\mathrm{AP}_{50}\\)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.146, + 0.568, + 0.267 + ], + "angle": 0, + "content": "
MethodSplit-4Split-5
EasyHardExtreme30%40%50%
Oracle83.0977.47
BRL [8]73.5071.7066.20---
Co-mining [10]79.5978.3869.6074.4273.3069.89
Ours82.1581.5075.5976.8475.8874.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.579, + 0.09, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Table 3: Results on the proposed SSL+SAOD setup. VOC12 [12] is used as the unlabeled data and \\( p \\) is the removal percentage" + }, + { + "type": "table", + "bbox": [ + 0.583, + 0.147, + 0.892, + 0.267 + ], + "angle": 0, + "content": "
MethodAP
30%40%50%
Co-mining [10]46.5345.9843.21
Ours48.3447.8246.76
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.274, + 0.195, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.275, + 0.317, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.275, + 0.432, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.275, + 0.548, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.275, + 0.664, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.275, + 0.778, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.275, + 0.893, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.404, + 0.744, + 0.42 + ], + "angle": 0, + "content": "Figure 3: Illustration of the effects of various augmentations used in this work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.434, + 0.292, + 0.451 + ], + "angle": 0, + "content": "4.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.473, + 0.749 + ], + "angle": 0, + "content": "For all our experiments, we use a Faster RCNN [28] architecture with a ResNet-101 [45] FPN backbone [30] implemented using Detector2 [46] framework. For the image augmentations, we apply a For augmentation, a cascade of random contrast, brightness, saturation, lighting and bounding box erase augmentations. The effect of this augmentation is shown in Figure 3. We train all our models with a batch size of 8 for 270000 and 18000 iterations on the COCO and PASCAL-VOC splits respectively with a learning rate of 0.01. The learning rate is decreased ten fold twice at \\{210000, 250000\\} and \\{12000, 15000\\} for COCO and PASCAL-VOC respectively. We adopt a warm-up strategy for 1000 and 100 iterations for the COCO and PASCAL-VOC respectively. Following existing detector implementations we set \\(\\tau_{\\mathrm{bg}}\\), \\(\\tau_{\\mathrm{fg}}\\), \\(\\tau_{\\mathrm{obj}}\\) and \\(\\tau_{\\mathrm{ppm}}\\) to 0.2, 0.4, 0.5 and 0.8 respectively. We set \\(\\tau_{m}\\) to 0.9. During inference, we compute the backbone features (\\(f_{\\mathrm{o}}\\) and \\(f_{\\mathrm{a}}\\)) for both the original and augmented versions of the input to obtain the RoIs and only \\(f_{\\mathrm{o}}\\) is used for the final detections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.766, + 0.187, + 0.781 + ], + "angle": 0, + "content": "4.4. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.792, + 0.469, + 0.822 + ], + "angle": 0, + "content": "We compare our method against Co-Mining [10], BRL [8], and Pseudo Label [7]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We choose these methods as they outperform previous approaches for the SAOD task. We use the public implementation of these methods and use a ResNet 101 FPN [30] backbone for all the experiments in order to perform a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.434, + 0.791, + 0.451 + ], + "angle": 0, + "content": "4.5. Comparison with state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.893, + 0.535 + ], + "angle": 0, + "content": "In this section, we compare our method with contemporary methods. We evaluate all the models in two different setups. We name the first setup Sparsely annotated setup which evaluates models on SAOD. The second setup, contains labeled and unlabeled images and regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.893, + 0.656 + ], + "angle": 0, + "content": "Sparsely annotated setup: We show results of this setup in Table 1 and Table 2. Splits-2 and 3 ensure at least one annotation per image is retained. On the other hand, Split-1 doesn't ensure this and has significantly more unlabeled data than the other splits. Splits-4 and 5 have no unlabeled images at all. In both the tables, the rows named \"oracle\" refers to the models trained using all annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.717 + ], + "angle": 0, + "content": "From Table 1, we observe that our approach outperforms the other baselines and is closer to the oracle performance on the \\(30\\%\\) splits. Co-mining [10], performs competitively on all the splits at \\(30\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.763 + ], + "angle": 0, + "content": "On Split-1, our method obtains a performance improvement of 1.87, 3.08 and 7.75 on \\(30\\%\\), \\(40\\%\\) and \\(50\\%\\) sparsity respectively over Co-mining [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.84 + ], + "angle": 0, + "content": "On Split-2, our method obtains an improvement of 3.04, 3.9 and 10.5 at \\(30\\%\\), \\(40\\%\\) and \\(50\\%\\) sparsity respectively over the previous state of the art Co-mining. On Split-3, an again we see consistent improvement of 2.8, 4.61 and 10.52 at \\(30\\%\\), \\(40\\%\\) and \\(50\\%\\) sparsity respectively over Co-mining." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In particular, on the hardest setting \\((70\\%)\\) in Split 1-3, we demonstrate that current SAOD methods struggle at higher sparsity in the labeled data. This is because the performance of current SAOD approaches rely on the quality" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6775" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.095, + 0.5, + 0.111 + ], + "angle": 0, + "content": "Table 4: Ablation of various components of proposed approach." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.121, + 0.445, + 0.321 + ], + "angle": 0, + "content": "
C-RPNPPM\\( {\\mathcal{L}}_{\\text{SSL }}^{\\mathrm{u}} \\)\\( {\\mathcal{L}}_{\\text{det,a }}^{\\mathrm{s}} \\)AP\\( \\Delta \\)
✘✘✘✘41.77-
✘✘✘✓42.670.90
✓✘✘✓45.303.53
✓✓✘✓45.443.67
✓✓✓✘45.513.74
✘✓✓✓44.863.09
✓✓✓✓46.004.23
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.442 + ], + "angle": 0, + "content": "of pseudo labels which degrades at higher sparsity. Our proposed PPM prevents penalizing the classifier irrespective of the quality of pseudo labels and utilizes self-supervised loss to benefit from mined regions. On average, all the methods achieve lower performance on Split-1 compared to the other splits due to the nature of its creation, i.e., class aware dataset level fashion of sparsity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.504 + ], + "angle": 0, + "content": "From Table 2, we see a similar trend as above. All the methods perform competitively on the easier settings of both splits. However, the performance gap between the our approach and Co-mining [10] increases at higher sparsity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.595 + ], + "angle": 0, + "content": "On Split-4, we get an improvement of 2.56, 3.12 and \\(5.99\\mathrm{(AP_{50})}\\) percentage points on the easy, hard and extreme settings respectively over Co-mining. On Split-5, an improvement of 2.42, 2.58 and 4.46 points was observed. For details on more experiments refer to supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Single Instance Object Detection (SIOD) setup: [47] proposed SIOD where only one instance per category is annotated in every image. While this setup has its benefits [47], note that it is a special case of SAOD. For this setup, we obtain an AP of 32.89 (compared to 31.9 of [47]) which is an improvement of \\(\\sim 1\\) mAP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.885 + ], + "angle": 0, + "content": "\\( SSL + SAOD \\) setup: We propose a semi-supervised learning benchmark for SAOD. This benchmark entails training models on a sparsely annotated labeled and an unlabeled set. As shown in Figure 4, this setup introduces two kinds of sparsity in the data, namely, sparsity in labeled regions (left) and images (right). We believe this is a realistic setup and SAOD methods must be capable of tackling both these kinds of sparsity in the data. For this setup, we use Split-5 (Section 4.2) with increasing sparsity as the labeled set and VOC12 trainval as the unlabeled set. We use the COCO-style AP metric to report results on this setup." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In Table 3, we compare against Co-mining [10] and ob" + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.094, + 0.851, + 0.108 + ], + "angle": 0, + "content": "Sparsity in labeled regions Sparsity in labeled images" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.11, + 0.892, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.263, + 0.894, + 0.325 + ], + "angle": 0, + "content": "Figure 4: Types of sparsity in labeled data; sparsity in labeled regions (left) and images (right). Our proposed SSL+SAOD is a realistic setup that presents challenges from both kinds of sparsity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.472 + ], + "angle": 0, + "content": "serve an improvement of 1.81, 1.84 and \\(3.55\\mathrm{mAP}\\) on the \\(30\\%\\), \\(40\\%\\) and \\(50\\%\\) sparsity levels respectively. We observe that the gap in performance increases with sparsity, consistent with our observation for Tables 1 and 2. This can be attributed to the inability of methods like Co-mining to handle unlabeled images and high sparsity in the labeled data. We will make this benchmark public and propose this method as a baseline. We encourage researchers to report results on this benchmark in the future." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.706, + 0.5 + ], + "angle": 0, + "content": "4.6. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.582 + ], + "angle": 0, + "content": "In this section we conduct ablation experiments to understand the various components. For the ablation experiments, we use a ResNet-101 as the backbone network and train on Split-5 with \\( p = 50\\% \\). We evaluate on the VOC07 test set and report the COCO style AP metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.734 + ], + "angle": 0, + "content": "From Table 4, a baseline model trained on the ablation set attains 41.77 (row 1). Pseudo-labeling \\(\\left(\\mathcal{L}_{\\mathrm{det},a}^{s}\\right)\\) improves the performance by 0.9 (row 2). Co-mining [10] relies extensively on pseudo-labels. This results in a drop in performance at higher sparsities due to noisy pseudo-labels. Addition of the C-RPN improves our performance by 3.53 points (row 3). C-RPN reduces the overhead of computing two sets of proposals and learns a better notion of objects due to the combined processing of features from the two views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The combination of C-RPN and PPM improves the performance by 3.67 (row 4). We do not observe major improvements with the introduction of PPM because its task is to identify and segregate the unlabeled regions from the backgrounds. After the segregation, PPM does no further processing to improve performance. The power of PPM can be observed when trained in conjunction with consistency regularization loss \\((\\mathcal{L}_{\\mathrm{SSL}}^{\\mathrm{u}})\\) which achieves the best performance of 46 (row 7); an improvement of 4.23 points over the baseline. We distinguish ourselves from pseudolabeling approaches like Co-mining in one important as" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6776" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.101, + 0.432, + 0.115 + ], + "angle": 0, + "content": "Table 5: Analysis of the threshold used for PPM." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.127, + 0.434, + 0.194 + ], + "angle": 0, + "content": "
Threshold0.60.70.80.90.95
AP45.9645.9646.0045.7745.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.09, + 0.824, + 0.105 + ], + "angle": 0, + "content": "Table 6: Removing test time augmentations." + }, + { + "type": "table", + "bbox": [ + 0.466, + 0.117, + 0.892, + 0.205 + ], + "angle": 0, + "content": "
MethodSplit-1Split-2Split-3
30%50%70%30%50%70%30%50%70%
TTA38.2235.9232.6839.7636.9435.3339.5637.1535.48
w/o TTA38.3235.9132.6739.7736.9535.2739.5537.1235.48
" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.216, + 0.266, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.217, + 0.44, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.216, + 0.614, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.216, + 0.718, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.216, + 0.892, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.308, + 0.267, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.308, + 0.41, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.308, + 0.614, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.308, + 0.718, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.309, + 0.892, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.426, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Figure 5: Qualitative results showing the unlabeled regions identified by the PPM. The red boxes correspond to the available ground truth. A class agnostic NMS was performed on the regions and the result is shown in white." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.457, + 0.254, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.457, + 0.374, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.457, + 0.465, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.457, + 0.549, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.457, + 0.724, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.457, + 0.891, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.55, + 0.254, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.549, + 0.374, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.549, + 0.465, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.549, + 0.549, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.549, + 0.724, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.55, + 0.891, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.653, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Figure 6: Qualitative results comparing the output of a model trained using available ground truths (top) to a model trained using our approach (bottom). Predictions with a class confidence score greater than 0.9 are shown. Red: Person, Cyan: Dog, Purple: Horse, Yellow: Clock, Green: Stop sign, Blue: Parking meter, Violet: Giraffe, Orange: Potted plant, Black: Surfboard, Dark green: Boat" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.722, + 0.473, + 0.889 + ], + "angle": 0, + "content": "pect. While, Co-mining [10] relies solely on pseudo-labels, we leverage additional components from self-supervised learning along with pseudo-labeling. In row 5, due to C-RPN, PPM and \\(\\mathcal{L}_{\\mathrm{SSL}}^{\\mathrm{u}}\\), we show an improvement of 3.74 over the baseline. Our proposed components are orthogonal to pseudo-labeling as using them together results in an additional improvement of \\(\\sim 0.5\\) on mAP. At higher sparsity in the labeled dataset, the proposed C-RPN, PPM and \\(\\mathcal{L}_{\\mathrm{SSL}}^{\\mathrm{u}}\\) are less affected than pseudo-labels resulting in the large improvements on these splits. Finally, we show the effect of C-RPN by generating a single set of proposals from the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.722, + 0.892, + 0.766 + ], + "angle": 0, + "content": "original image and using it for both the branches. We observe a drop in performance (row 6) highlighting the effectiveness of C-RPN." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "PPM mines potential positives which can be mistaken for negatives due to missing annotations. We rely on the objectness score of the RPN to identify these regions. In Table 5, we vary the threshold of PPM. For a low threshold, a few hard negatives might also be identified as pseudo-positive leading to a drop in performance. With a high threshold, a few potential positives might not be mined. We observe that a threshold of 0.8 provides a good trade-off and is there-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6777" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.121 + ], + "angle": 0, + "content": "Table 7: Comparison with a different backbone. Results are reported on the COCO validation set on Splits 1-3 using AP and on VOC 2007 test set on Splits-4,5 using \\(\\mathrm{AP}_{50}\\)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.131, + 0.89, + 0.221 + ], + "angle": 0, + "content": "
MethodSplit-1Split-2Split-3Split-4Split-5
30%50%70%30%50%70%30%50%70%EasyHardExtreme30%40%50%
Ours (C4)37.6735.9533.1639.2236.8134.9838.8436.7635.2680.5680.4374.4577.3876.1375.32
Ours (FPN)38.2235.9232.6839.7636.9435.3339.5637.1535.4882.1581.5075.5976.8475.8874.35
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.471, + 0.293 + ], + "angle": 0, + "content": "fore used for all experiments unless stated otherwise. In all our experiments, PPM is performed after an initial warmup of 9000 and 30000 iterations on the PASCAL-VOC and COCO datasets respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.313, + 0.363, + 0.329 + ], + "angle": 0, + "content": "4.6.1 Inference without augmentations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.471, + 0.489 + ], + "angle": 0, + "content": "Our method requires passing an input with augmentations during inference as well. It should be noted that this is not a test-time augmentation (TTA), a technique that typically involves passing images at a higher resolution. We perform inference by removing the augmentation and extracting the region proposals using C-RPN. We show the results in Table 6 on the three splits of COCO. We do not observe a significant improvement in performance due to the augmentation. For the analysis on the effect of augmentations refer to supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.263, + 0.525 + ], + "angle": 0, + "content": "4.6.2 Effect of backbone" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.643 + ], + "angle": 0, + "content": "We analyze the effect of backbone on our approach and show results for the normal convolution based (C4) and FPN based (FPN) backbones in Table 7. The first row corresponds to C4 while the second row corresponds to FPN. While the gap in performance is lower in most cases, we observe a significant improvement using the FPN on the low sparsity settings of all the split." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.456, + 0.667 + ], + "angle": 0, + "content": "4.7. RPN Recall experiments for object discovery" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.903 + ], + "angle": 0, + "content": "PPM identifies foreground regions mistakenly assigned as background during training to avoid penalizing the network. To study the effect of PPM on novel [48] classes, we train a network using our approach on randomly chosen 6000 images of the COCO dataset, containing annotations for only 20 classes of PASCAL-VOC. We evaluate the recall@0.5 of the RPN over the remaining 60 classes. A model trained using the standard technique on this dataset achieves a recall@0.5 of \\(77.46\\%\\) and \\(29.06\\%\\) on the known classes (20 categories) and unknown classes (60 categories) respectively. Our proposed approach, with PPM, achieves a recall@0.5 of \\(78.47\\%\\) and \\(35.20\\%\\) respectively. This ability to localize objects not seen during training can be beneficial for object discovery methods like [49-52] which use RPN proposals to learn/discover new categories." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.231, + 0.684, + 0.247 + ], + "angle": 0, + "content": "4.8. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.257, + 0.892, + 0.44 + ], + "angle": 0, + "content": "In Figure 5, we show the pseudo positives mined by PPM. In each figure, the red boxes correspond to the ground truth annotations and the white boxes correspond to the post NMS pseudo positive boxes mined by PPM. We observe that the PPM correctly mines proposals which correspond to missing object annotations. Without PPM, these regions will be used as negatives to the classifier resulting in a reduced class confidence score leading to a drop in mAP. We show the detection results of a model trained on the \\(50\\%\\) Split-1 of our approach in Figure 6. The images on the top corresponds to the model trained using sparse annotations and the bottom image shows the output of our approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.619, + 0.475 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.744 + ], + "angle": 0, + "content": "We present SparseDet, a novel end-to-end system for Sparsely Annotated Object Detection (SAOD). We propose a simple yet effective technique for identifying the unlabeled regions using pseudo-positive mining and apply self-supervised loss on them. Through qualitative results we highlight the ability of PPM to mine pseudo-positives. We standardize the evaluation setup and show the effectiveness of our approach with an exhaustive set of experiments on multiple splits of SAOD. While we outperform existing state-of-the-art on all metrics and splits, we observe the gap in performance increases with sparsity demonstrating the short coming of methods that rely solely on pseudolabeling. We propose a new benchmark, that evaluates the semi-supervised learning capabilities of SAOD approaches. We will release the data for the new benchmark along with all the SAOD splits and encourage researchers to evaluate future SAOD methods on these." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. This project was partially funded by DARPA SAIL-ON (W911NF2020009), DARPA SemaFor (HR001119S0085), and IARPA SMART (2021-21040700003) programs. Rama Chellappa was supported by an ONR MURI Grant N00014-20-1-2787. We would like to thank our colleagues from the PI lab for their feedback on this work. The views and conclusions contained herein are those of the authors and do not represent views, policies, and/or endorsements of funding agencies." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6778" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Simon Chadwick and Paul Newman. Training object detectors with noisy data. 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1319-1325, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.165, + 0.471, + 0.207 + ], + "angle": 0, + "content": "[2] Zhe Wu, Navaneeth Bodla, Bharat Singh, Mahyar Najibi, Rama Chellappa, and Larry S. Davis. Soft sampling for robust object detection. In BMVC, 2019. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.257 + ], + "angle": 0, + "content": "[3] Junnan Li, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Towards noise-resistant object detection with noisy annotations. ArXiv, abs/2003.01285, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.265, + 0.47, + 0.307 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.47, + 0.371 + ], + "angle": 0, + "content": "[5] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.379, + 0.47, + 0.462 + ], + "angle": 0, + "content": "[6] Alina Kuznetsova, Hassan Rom, Neil Gordon Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4. International Journal of Computer Vision, 128:1956-1981, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.47, + 0.47, + 0.554 + ], + "angle": 0, + "content": "[7] Yusuke Niitani, Takuya Akiba, Tommi Kerola, Toru Ogawa, Shotaro Sano, and Shuji Suzuki. Sampling techniques for large-scale object detection from sparsely annotated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6510-6518, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.562, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[8] Han Zhang, Fangyi Chen, Zhiqiang Shen, Qiqi Hao, Chenchen Zhu, and Marios Savvides. Solving missing-annotation object detection with background recalibration loss. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1888-1892. IEEE, 2020. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.653, + 0.469, + 0.694 + ], + "angle": 0, + "content": "[9] Yuewei Yang, Kevin J Liang, and Lawrence Carin. Object detection as a positive-unlabeled problem. In BMVC, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[10] Tiancai Wang, Tong Yang, Jiale Cao, and X. Zhang. Comining: Self-supervised learning for sparsely annotated object detection. In AAAI, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.753, + 0.472, + 0.823 + ], + "angle": 0, + "content": "[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2007 (VOC2007) Results. http://www.pascalnetwork.org/challenges/VOC/voc2007/workshop/index.html, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.473, + 0.9 + ], + "angle": 0, + "content": "[12] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[13] Ke Yan, Jinzheng Cai, Adam P. Harrison, Dakai Jin, Jing Xiao, and Le Lu. Universal lesion detection by learning from multiple heterogeneously labeled datasets. ArXiv, abs/2005.13753, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.159, + 0.892, + 0.229 + ], + "angle": 0, + "content": "[14] Hansheng Li, Xin Han, Yuxin Kang, Xiaoshuang Shi, Mengdi Yan, Zixu Tong, Qirong Bu, Lei Cui, Jun Feng, and Lin Yang. A novel loss calibration strategy for object detection networks training on sparsely annotated pathological datasets. In MICCAI, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.24, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[15] Yongqiang Zhang, Mingli Ding, Yancheng Bai, Mengmeng Xu, and Bernard Ghanem. Beyond weakly supervised: Pseudo ground truths mining for missing bounding-boxes object detection. IEEE Transactions on Circuits and Systems for Video Technology, 30:983-997, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[16] Ke Yan, Jinzheng Cai, Youjing Zheng, Adam P. Harrison, Dakai Jin, You-Bao Tang, Yuxing Tang, Lingyun Huang, Jing Xiao, and Le Lu. Learning from multiple datasets with heterogeneous and partial labels for universal lesion detection in ct. IEEE Transactions on Medical Imaging, 40:2759-2770, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.416, + 0.892, + 0.47 + ], + "angle": 0, + "content": "[17] Jisoo Jeong, Seungeui Lee, Jeesoo Kim, and Nojun Kwak. Consistency-based semi-supervised learning for object detection. Advances in neural information processing systems, 32:10759-10768, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[18] Peng Tang, Chetan Ramaiah, Yan Wang, Ran Xu, and Caiming Xiong. Proposal learning for semi-supervised object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2291-2301, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[19] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[20] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. arXiv preprint arXiv:2102.09480, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.697, + 0.892, + 0.753 + ], + "angle": 0, + "content": "[21] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. arXiv preprint arXiv:2106.09018, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.765, + 0.892, + 0.834 + ], + "angle": 0, + "content": "[22] Qize Yang, Xihan Wei, Biao Wang, Xian-Sheng Hua, and Lei Zhang. Interactive self-training with mean teachers for semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5941-5950, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Jihun Yoon, Seungbum Hong, and Min-Kook Choi. Semi-supervised object detection with sparsely annotated dataset. In 2021 IEEE International Conference on Image Processing (ICIP), pages 719-723. IEEE, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6779" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097-1105. Curran Associates, Inc., 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.185, + 0.47, + 0.269 + ], + "angle": 0, + "content": "[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. doi: 10.1007/s11263-015-0816-y. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[26] Ross Girshick. Fast r-cnn. In Proceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), ICCV '15, pages 1440-1448, Washington, DC, USA, 2015. IEEE Computer Society. ISBN 978-1-4673-8391-2. doi: 10.1109/ICCV.2015.169. URL http://dx.doi.org/10.1109/ICCV.2015.169.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.369, + 0.469, + 0.439 + ], + "angle": 0, + "content": "[27] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[28] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.596 + ], + "angle": 0, + "content": "[29] Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. R-fcn: Object detection via region-based fully convolutional networks. In Daniel D. Lee, Masashi Sugiyama, Ulrike von Luxburg, Isabelle Guyon, and Roman Garnett, editors, NIPS, pages 379-387, 2016. URL http://dblp.uni-trier.de/db/conf/nips/nips2016.html#DaiLHS16.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Piotr Dólár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 936-944, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.684, + 0.469, + 0.711 + ], + "angle": 0, + "content": "[31] Bharat Singh and Larry S Davis. An analysis of scale invariance in object detection-snip. CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[32] Bharat Singh, Mahyar Najibi, and Larry S Davis. SNIPER: Efficient multi-scale training. NeurIPS, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[33] Mahyar Najibi, Bharat Singh, and Larry S Davis. AutoFocus: Efficient multi-scale inference. ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.794, + 0.469, + 0.849 + ], + "angle": 0, + "content": "[34] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. doi: 10.1109/ICCV.2017.322.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Alexander Kirillov, Ross B. Girshick, Kaiming He, and Piotr Dóllár. Panoptic feature pyramid networks. In CVPR, pages 6399-6408. Computer Vision Foundation / IEEE, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[36] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In CVPR, pages 9404-9413. Computer Vision Foundation / IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.16, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[37] Justin Johnson Georgia Gkioxari, Jitendra Malik. Mesh r-cnn. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.201, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[38] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.297, + 0.892, + 0.354 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.892, + 0.421 + ], + "angle": 0, + "content": "[40] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection, 2015. URL http://arxiv.org/abs/1506.02640. cite arxiv:1506.02640.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[41] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[42] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[43] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:386-397, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.64, + 0.892, + 0.68 + ], + "angle": 0, + "content": "[44] Ross B. Girshick. Fast R-CNN. 2015 IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.694, + 0.892, + 0.763 + ], + "angle": 0, + "content": "[45] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. doi: 10.1109/CVPR.2016.90.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.818 + ], + "angle": 0, + "content": "[46] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[47] Hanjun Li, Xingjia Pan, Ke Yan, Fan Tang, and Wei-Shi Zheng. Siod: Single instance annotated per category per image for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14197-14206, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6780" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[48] Akshay Raj Dhamija, Manuel Günther, Jonathan Ventura, and Terrance E. Boult. The overlooked elephant of object detection: Open set. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1010-1019, 2020. doi: 10.1109/WACV45572.2020.9093355. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.171, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[49] Huy V. Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[50] Huy V. Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 8287-8296, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[51] Sai Saketh Rambhatla, Rama Chellappa, and Abhinav Shrivastava. The pursuit of knowledge: Discovering and localizing novel categories using dual memory. In Proceedings of the International Conference on Computer Vision (ICCV), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.409, + 0.469, + 0.478 + ], + "angle": 0, + "content": "[52] Sai Saketh Rambhatla, Ishan Misra, Rama Chellappa, and Abhinav Shrivastava. The pursuit of knowledge: Discovering and localizing novel categories using dual memory. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 9" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.513, + 0.956 + ], + "angle": 0, + "content": "6781" + } + ] +] \ No newline at end of file diff --git a/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/5aaa9a89-cdc6-4e0a-94f3-a3b82685ba6a_origin.pdf b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/5aaa9a89-cdc6-4e0a-94f3-a3b82685ba6a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..322daad9499c8582a423f4f1fa963a4265a9df31 --- /dev/null +++ b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/5aaa9a89-cdc6-4e0a-94f3-a3b82685ba6a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d25181b634e2b174e36a3955b9826e1a82d55e6f87143b48b7e5b3c99dc90a0c +size 4863869 diff --git a/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/full.md b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..84eaca46f7f5c93201f1777aa482a1e800c3a504 --- /dev/null +++ b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/full.md @@ -0,0 +1,383 @@ +# SparseDet: Improving Sparsely Annotated Object Detection with Pseudo-positive Mining + +Saksham Suri\*1 + +sakshams@cs.umd.edu + +Saketh Rambhatla*1 + +rssaketh@umd.edu + +Rama Chellappa2 + +rchella4@jhu.edu + +Abhinav Shrivastava + +abhinav@cs.umd.edu + +University of Maryland, College Park1 + +Johns Hopkins University + +# Abstract + +Training with sparse annotations is known to reduce the performance of object detectors. Previous methods have focused on proxies for missing ground truth annotations in the form of pseudo-labels for unlabeled boxes. We observe that existing methods suffer at higher levels of sparsity in the data due to noisy pseudo-labels. To prevent this, we propose an end-to-end system that learns to separate the proposals into labeled and unlabeled regions using Pseudo-positive mining. While the labeled regions are processed as usual, self-supervised learning is used to process the unlabeled regions thereby preventing the negative effects of noisy pseudo-labels. This novel approach has multiple advantages such as improved robustness to higher sparsity when compared to existing methods. We conduct exhaustive experiments on five splits on the PASCAL-VOC and COCO datasets achieving state-of-the-art performance. We also unify various splits used across literature for this task and present a standardized benchmark. On average, we improve by 2.6, 3.9 and 9.6 mAP over previous state-of-the-art methods on three splits of increasing sparsity on COCO. Our project is publicly available at cs.umd.edu/~sakshams/SparseDet. + +# 1. Introduction + +The performance of object detectors is sensitive to the quality of labeled data [1-3]. Existing object detection methods assume that the training data is pristine and a drop in performance is observed if this assumption fails. Noise in the data used for training object detectors can arise due to incorrect class labels or incorrect/missing bounding boxes. In this work, we deal with the problem of training object detectors with sparse annotations, i.e., missing region or bounding boxes. This problem is of utmost importance, as obtaining crowd-sourced datasets [4, 5] can be expensive + +![](images/c0088fc7addd2f7f741af579ba77c33d707eac23ba1e8dbb2f630f0abb57090e.jpg) +Figure 1: (Top) Most Object Detection datasets have exhaustive annotations for foreground/positives. During training, the unlabeled regions can be safely considered as background/negatives. Sparsely Annotated Object Detection datasets (bottom) have missing annotations. This results in foreground regions (shown in red) being considered as negatives during training, adversely affecting the performance of the classifier. + +and laborious. The alternative is to use computer-assisted protocols to collect annotations which have been shown to be noisy and incomplete [6]. This problem of training object detectors with incomplete bounding box annotations is called Sparsely Annotated Object Detection (SAOD). + +To understand why training with sparse annotations is detrimental to the performance, consider the example shown in Figure 1 (top). If the annotation were exhaustive, then the negative samples to the classifier contain true background regions. But with sparse annotations, as shown in Figure 1 (bottom), a few positive regions will inevitably be considered as negatives (shown in red), thereby wrongfully penalizing the classifier leading to lower performance. Existing methods[2, 7-10] prevent this by predicting pseudolabels and removing the foreground regions from the negatives to the classifier. However, at higher levels of sparsity, the quality of pseudo-labels is greatly affected, resulting in the same problem noted above. + +Crowd sourced object detection datasets [5, 11, 12] are ensured to be almost exhaustively labeled. Hence, for SAOD, researchers artificially create sparsely annotated datasets from the original ones. There is no general consensus on the correct way to create the sparsely annotated datasets, a.k.a. splits, and hence each method reports results on one or two different splits. Split can be created by considering the dataset as a whole or per image (i.e. annotations can be removed by considering all the images or only a single image at once). They can also be created by removing annotations in a class-agnostic or class-aware fashion (i.e. remove $p\%$ , of annotations per category or across all the categories). These variations result in splits with different data distributions making some harder than the others (refer to Table 1). A proper benchmark that analyzes the performance of SAOD methods across these different types of splits is missing. This makes it difficult to compare methods and assess their effectiveness for a specific use case. + +To tackle the issues discussed above, as our first major contribution, we present SparseDet, a novel SAOD framework that achieves state-of-the-art performance across multiple SAOD benchmarks in practice. SparseDet operates on an image and its augmented counterpart. The combination of features extracted from the two views is used to generate region proposals. Standard detection training methods, consider a region proposal as positive if its intersection over union (IoU) with any ground truth is greater than 0.5 and the rest are treated as negatives. This strategy works when the annotations are exhaustive, which implies that the remaining regions are from background. But due to missing annotations, some of these regions could belong to foreground instances. To prevent considering all region proposals without annotations as negatives, SparseDet partitions all the region proposals into labeled, unlabeled and background. The labeled and background regions are processed as usual. Features extracted from unlabeled regions are then trained with a self-supervised loss. Previous approaches like Co-Mining [10], consider two partitions, labeled and background, and generate pseudo-labels. This is a disadvantage at high sparsity as the generated pseudo-labels can be very noisy. The self-supervised loss in our approach enforces consistency between the features of the two views for the unlabeled regions and prevents penalizing the classifier due to false negatives. + +Our second major contribution is unifying evaluation. The standard practice is to simulate sparsely annotated training data on COCO [5] and PASCAL-VOC [11] train sets and evaluate them on their corresponding standard validation set. As discussed above, a survey of recent SAOD approaches [2, 7-10, 13-16] reveals that there are at least five different ways to create splits, each differing in the strategy (refer to Section 4.2) used to achieve the desired level of sparsity. However, these splits have not been made public, + +making it difficult to replicate results for comparison. Additionally, each of these strategies has a different property for simulating sparse data, e.g., different distribution of annotations per class, resulting in a different training set. As a result, methods trained on different splits cannot be compared with one another. To mitigate these issues we standardize the generation of these splits that enables the evaluation of any SAOD methods on all of them for fair comparison. Additionally, we propose a new benchmark that assesses the semi-supervised learning capabilities of SAOD methods, i.e., leveraging unlabeled data to improve performance. We present our approach as a baseline. We will make the data for the new benchmark along with all the SAOD splits public to facilitate future research. We briefly summarize our contributions below. + +- We propose a novel formulation, SparseDet, for SAOD which is an end-to-end approach that identifies labeled, unlabeled and background regions and deals with them in appropriate manner. +- We show state-of-the-art performance on sparsely annotated object detection across various splits. On average, we improve by 2.6, 3.9 and $9.6\mathrm{mAP}$ over previous state-of-the-art methods on three splits of increasing sparsity on COCO. +- We standardize the experimental setup for SAOD by evaluating methods on all the splits to facilitate future research. Additionally, we propose a new benchmark that evaluates the semi-supervised learning capabilities of SAOD methods. + +In Section 2, we discuss the related works on SAOD and related fields. We describe our approach in detail in Section 3. We describe our experimental setup and present results in Section 4, and conclude in Section 5. + +# 2. Related Work + +Semi-supervised object detection: Semi-supervised object detection (SSOD) is an active field that also deals with training object detectors with missing annotations. Existing works on semi-supervised object detection, have focused mainly on consistency regularization [17, 18] or pseudolabeling-based approaches [19-22]. The main idea behind these approaches is to perturb the images, or features, and apply a consistency regularization loss to enforce consistency between the predictions using a student teacher framework. However, typical SSOD methods assume a small exhaustively labeled set and a large unlabeled set for training. This is different from Sparsely annotated object detection (SAOD), which assumes a large training set which is sparsely labeled. + +Sparsely annotated object detection: One of the initial works addressing this problem, by Niitani et al. [7], pro + +poses utilizing logical relations between the co-occurrences of objects and pseudo-labeling. Yoon et al. [23] use object tracking to densely label objects across sparsely annotated frames along with single stage detectors to mitigate the negative effects of missing annotations. Wu et al. [2] propose a re-weighting approach where the gradients corresponding to region of interest are weighed as a function of overlap with ground truth instances. Improving upon the previous work, Zhang et al. [8] propose an automatic re-calibration strategy for single stage detectors where the negative branch is changed to take into account low confidence background predictions which might correspond to missing annotations. + +Finally, one of the most recent works, Co-mining [10] uses a co-training strategy by using two views of an image and predictions from one view along with the ground truth as supervision for the other view and vice versa. Our method doesn't solely rely on pseudo-labels and leverages a self-supervised loss to prevent propagating negative gradients to the model due to false negatives. + +Fully supervised object detection: After the success of AlexNet [24] on the image classification challenge (ILSVRC 2012) [25], research on designing deep neural networks for object detection gathered more interest. First among the successful methods are the two-stage Region-based convolutional networks (R-CNN) family of detectors. Two-stage object detectors consists of 1) a region proposal stage which produces a set of candidate object bounding boxes followed by 2) a classification stage which classifies each candidate region as either belonging to a foreground object or "background". R-CNN processes a large amount of region proposals by cropping the input image and using a CNN backbone to extract features making it extremely slow. Fast R-CNN [26] was proposed to overcome this limitation. Fast R-CNN computes one convolution feature map for the whole image. RoI Pooling was introduced to pool the feature for each region of interest into a fixed spatial dimension. RoI pooling shares the computation among all the region proposals speeding up training and inference. Fully connected layers are applied on the fixed RoI pooled feature maps which are then passed to two sister heads, for classification and bounding box regression. The whole network is trained end-to-end with a multi-task loss avoiding the multi-stage training in R-CNN [27]. While Fast R-CNN improved the efficiency of its predecessor, test time computation bottleneck is still an issue because of the region proposal method employed. Faster R-CNN [28] proposed a region proposal network (RPN), an elegant solution that trains deep networks to predict region proposals for practically no additional computational overhead. RPN shares convolutional layers with Fast R-CNN [26] and at test time the cost of generating proposals is minimal. Faster R-CNN paved the way for more sophisticated and efficient two stage detection architectures [29-33]. Faster R-CNN + +has also been extended to achieve state-of-the-art instance segmentation [34], panoptic segmentation [35, 36], and 3D mesh generation [37] etc. A few limitations of two stage object detectors and anchor boxes has also inspired the family of single-stage [38-40] and anchor-free object detection systems [41, 42]. + +# 3. Approach + +We present SparseDet for Sparsely Annotated Object Detection. Given $N$ images, we denote the set of labeled regions in the dataset as $\mathcal{B}_{\mathrm{l}} = \{b_{i},c_{i}\}_{i = 1}^{\mathrm{N}_{\mathrm{l}}}$ where $(b_{*},c_{*})$ are the bounding box and class labels respectively. The unlabeled regions are denoted as $\mathcal{B}_{\mathrm{ul}} = \{b_k\}_{k = 1}^{\mathrm{N}_{\mathrm{ul}}}$ and the background regions as $\mathcal{B}_{\mathrm{bg}} = \{b_q\}_{q = 1}^{\mathrm{N}_{\mathrm{bg}}}$ . Note that unlike existing SAOD approaches, we do not assume images to contain at least one labeled region in this work. The unlabeled $\mathcal{B}_{\mathrm{ul}}$ and background $\mathcal{B}_{\mathrm{bg}}$ sets are not known a-priori. + +# 3.1. Overview + +The proposed approach is shown in Figure 2 and consists of a backbone network that extracts features from the original and augmented views of an image. The common RPN (C-RPN), concatenates the features, to generate a set of region proposals. A region proposal $b$ can belong to one of three groups, namely 1) labeled regions $b \in \mathcal{B}_1$ , 2) unlabeled foreground regions $b \in \mathcal{B}_{\mathrm{ul}}$ , or 3) background regions $b \in \mathcal{B}_{\mathrm{bg}}$ . For a given set of ground-truth annotations, the first group, i.e. labeled regions can be automatically identified. The problem then reduces to identifying and separating the second group i.e. the unlabeled regions, from the background regions. Given all the region proposals, a pseudo-positive mining (PPM) step identifies the unlabeled regions and segregates them from the background regions. The labeled and unlabeled regions are trained using supervised and self-supervised losses respectively. We describe each stage in detail below. + +# 3.2. Feature Extraction + +Given an image $I$ , an augmented version of $I$ denoted as $\mathcal{A}(I)$ is computed. In this work, we use random contrast, brightness, saturation, lighting and bounding box erase in a cascaded fashion to generate $\mathcal{A}(I)$ . A backbone network is employed to extract two features, $f_{\mathrm{o}}$ and $f_{\mathrm{a}}$ , from $I$ and $\mathcal{A}(I)$ respectively. + +# 3.3. Common RPN (C-RPN) + +Two stage object detectors [28, 30, 43] use a region proposal network (RPN) [28] to generate regions of interest (RoIs). We propose C-RPN which concatenates $f_{\mathrm{o}}$ and $f_{\mathrm{a}}$ to obtain the RoIs. This is different from existing approaches that use $f_{\mathrm{o}}$ and $f_{\mathrm{a}}$ to generate two separate sets of RoIs. Operating on two sets increases the difficulty of identifying the labeled, unlabeled and background regions which + +![](images/9725617a6613c754f8f36432dc8cc3994bab2ec28970cf7fe7fb2a7f718810ce.jpg) +Figure 2: Illustration of SparseDet for sparsely annotated object detection. Following feature extraction from the original and augmented image, a common set of proposals is generated by the common RPN (C-RPN). Using an end to end approach, we identify and mine proposals corresponding to missing annotations using pseudo positive mining (PPM). We train the network end-to-end using a combination of supervised and self-supervised losses. The unlabeled instances (black) are supervised with a self-supervised loss and the labeled instances (green) are supervised with ground truth annotations. + +are required for subsequent stages. The sparse ground truth is used as supervision to train the C-RPN with a binary cross entropy loss for foreground classification and smooth L1 [44] loss for bounding box regression as shown below: + +$$ +\mathcal {L} _ {\mathrm {r p n}} ^ {\mathrm {s}} \left(\mathbf {c}, \mathbf {b}, \mathbf {c} ^ {*}, \mathbf {b} ^ {*}\right) = \mathcal {L} _ {\mathrm {b c e}} \left(\mathbf {c}, \mathbf {c} ^ {*}\right) + \lambda \mathcal {L} _ {\mathrm {r e g}} \left(\mathbf {b}, \mathbf {b} ^ {*}\right) \tag {1} +$$ + +where $\mathbf{c}$ , $\mathbf{b}$ are the objectness scores and bounding boxes, and $\mathbf{c}^*$ , $\mathbf{b}^*$ are the corresponding ground truth. + +# 3.4. Pseudo Positive Mining (PPM) + +Given the RoIs from C-RPN, the next step is to segregate the unlabeled regions from the labeled and background regions. A standard practice in training detectors is to consider all proposals with an objectness score greater than $\tau_{\mathrm{obj}}$ and IoU greater $\tau_{\mathrm{fg}}$ with any ground truth as foreground and proposals with IoU less than $\tau_{\mathrm{bg}}$ as background. In the presence of missing annotations, detectors are wrongly penalized because the IoU of RoIs corresponding to missing annotations with any ground truth is less than $\tau_{\mathrm{bg}}$ . To avoid this, we mine "potential" foregrounds and consider them as unlabeled regions $(\mathcal{B}_{\mathrm{u}})$ . The proposed PPM module is based on our observation that when trained with sparse annotations, the RPN can reliably distinguish foreground from background regions. We pick all RoIs that have an objectness score greater than $\tau_{\mathrm{ppm}}$ and IoU less than $\tau_{\mathrm{bg}}$ with any ground truth as the unlabeled regions. The remaining RoIs are assigned as background. + +# 3.5. Losses + +The pseudo positive mining step segregates the RoIs into labeled, unlabeled and background regions. An RoI pooling layer [28] extracts RoI pooled features for the labeled and background regions using the feature $f_{0}$ . The detection head processes the RoI pooled features to predict class-wise + +probabilities and bounding box adjustments for each region. The ground truth is used to supervise the predictions by applying the cross entropy loss for classification and smooth L1 [44] for bounding box regression as shown below: + +$$ +\mathcal {L} _ {\det , 0} ^ {s} \left(\mathbf {c}, \mathbf {b}, \mathbf {c} ^ {*}, \mathbf {b} ^ {*}\right) = \mathcal {L} _ {\mathrm {c e}} \left(\mathbf {c}, \mathbf {c} ^ {*}\right) + \lambda \mathcal {L} _ {\mathrm {r e g}} \left(\mathbf {b}, \mathbf {b} ^ {*}\right) \tag {2} +$$ + +where $\mathbf{c}$ , $\mathbf{b}$ are the class and bounding box predictions, and $\mathbf{c}^*$ , $\mathbf{b}^*$ are the corresponding ground truth class labels and bounding boxes. + +All detections with a score greater than $\tau_{\mathrm{m}}$ are combined with the ground truth followed by a class specific NMS to obtain the merged ground truth. It is ensured that no ground truth annotation is discarded during this step. The labeled and background regions along with $f_{\mathrm{a}}$ are used to extract RoI pooled features which are then fed to the detector head. The predictions of the detector head are supervised with the merged ground truth with the following losses: + +$$ +\mathcal {L} _ {\det , \mathrm {a}} ^ {\mathrm {s}} (\mathbf {c}, \mathbf {b}, \mathbf {c} ^ {*}, \mathbf {b} ^ {*}) = \mathcal {L} _ {\mathrm {b c e}} \left(\mathbf {c}, \mathbf {c} _ {\mathbf {m}} ^ {*}\right) + \lambda \mathcal {L} _ {\mathrm {r e g}} (\mathbf {b}, \mathbf {b} _ {\mathbf {m}} ^ {*}) \tag {3} +$$ + +where $\mathbf{c}$ , $\mathbf{b}$ are the class and bounding box predictions from the detector head, and, $\mathbf{c}_{\mathbf{m}}^{*}, \mathbf{b}_{\mathbf{m}}^{*}$ are the corresponding merged ground truths. + +Finally, a class agnostic NMS is performed on the unlabeled regions $\mathcal{B}_{\mathrm{u}}$ (obtained after PPM described in Sec. 3.4). The unlabeled regions along with $f_{\mathrm{o}}$ and $f_{\mathrm{a}}$ are passed through the ROI pooling layer and the detection head to obtain $f_{\mathrm{o}}^{\mathrm{d}}$ and $f_{\mathrm{a}}^{\mathrm{d}}$ respectively. A self-supervised loss is applied that enforces the detection head features of the original and augmented regions to be consistent with each other as shown below: + +$$ +\mathcal {L} _ {\text {S S L}} ^ {\mathrm {u}} \left(f _ {\mathrm {a}} ^ {\mathrm {d}}, f _ {\mathrm {o}} ^ {\mathrm {d}}\right) = \| f _ {\mathrm {a}} ^ {\mathrm {d}} - f _ {\mathrm {o}} ^ {\mathrm {d}} \| _ {2} ^ {2} \tag {4} +$$ + +The network is trained end-to-end with both supervised and unsupervised losses as shown below: + +$$ +\mathcal {L} = 0. 5 * \left(\mathcal {L} _ {\text {d e t , o}} ^ {\mathrm {s}} + \mathcal {L} _ {\text {d e t , a}} ^ {\mathrm {s}}\right) + \mathcal {L} _ {\mathrm {r p n}} ^ {\mathrm {s}} + \mathcal {L} _ {\mathrm {S S L}} ^ {\mathrm {u}} \tag {5} +$$ + +Table 1: Sparsely annotated object detection results on three splits of COCO dataset. "Oracle" corresponds to training models using all annotations. Results are reported on the COCO validation set using AP[0.50:0.95]. + +
MethodSplit-1Split-2Split-3100%
30%50%70%30%50%70%30%50%70%
Oracle40.91
Pseudo Label [7]-27.50--------
BRL [8]-32.70--------
Co-mining [10]36.3532.8424.9336.7233.0424.8336.7632.5424.96-
Ours38.2235.9232.6839.7636.9435.3339.5637.1535.48-
+ +Discussion: In the absence of ground truth annotations, i.e. for completely unlabeled images, the supervised losses $(\mathcal{L}_{\mathrm{det,o}}^{\mathrm{s}},\mathcal{L}_{\mathrm{det,a}}^{\mathrm{s}},\mathcal{L}_{\mathrm{rpn}}^{\mathrm{s}})$ cannot be computed. Our proposed approach leverages self-supervised consistency loss that does not need ground truth. This helps our approach leverage these unlabeled regions unlike contemporary SAOD methods. We claim that this is the first method to use self-supervised losses for SAOD. Even though Co-Mining [10] claims to use self-supervised learning it is technically co-learning. Ours is the first approach to use pseudo-labeling and a self-supervised loss to handle the sparse annotations. + +# 4. Experimental Evaluation + +In this section, we describe the experiments to evaluate our proposed approach. In Sections 4.1 and 4.2, we describe data, splits, and metrics. In Section 4.3, we mention the implementation details followed by the baselines in Section 4.4. We compare with contemporary methods in Section 4.5, followed by an ablation study in Section 4.6. + +# 4.1. Data and Metrics + +We conduct all our experiments on the COCO [5] and PASCAL-VOC [11, 12] (2007+2012) datasets. The COCO [5] dataset consists of 118000 and 5000 images for training and validation respectively. Experiments on the PASCAL-VOC07 [11] are conducted on 5011 trainval images and performance is computed on 4952 images of the test set. The PASCAL-VOC12 version consists of 11530 (trainval) images for training and evaluation is done on the PASCAL-VOC07 test set. Following past literature [2, 7-10], we create five different splits (Section 4.2) and report results on them. For splits on the COCO [5] dataset, we use the standard COCO style Average Precision (AP). For splits on the PASCAL-VOC [11] dataset, we use the standard PASCAL-VOC style $\mathrm{AP}_{50}$ , which is Average Precision computed at an IoU threshold of 0.5. + +# 4.2. SAOD Splits + +An extensive review of sparsely annotated object detection methods reveals that there are at least five popular types of splits in use for creating training data. Most methods re + +port results on a subset of these making it harder to compare across methods. + +We standardize the evaluation of SAOD methods by evaluating exhaustively on all the splits facilitating future research. We briefly describe the splits below. + +Split-1 [9, 10]: In this split, for each object category, $p\%$ of annotations are randomly removed from the COCO [5] train set and results are reported on the validation set. This split simulates sparsity at dataset level in a class aware fashion. Note, this split can contains images with no annotations. We experiment with $p = \{30, 50, 70\}$ . + +Split-2: For each image in the COCO [5] train set, all annotations of $p\%$ of all categories in that image removed and results are reported on the COCO validation set. This split can be considered as image level and class aware. We experiment with $p = \{30, 50, 70\}$ . Note, this split might contain images with no annotations. + +Split-3: This split, which uses the COCO 2017 [5] train set for training and the validation set for evaluation, deletes $p\%$ annotations in a class agnostic fashion for each image ensuring at least one annotation. This split is image level but class agnostic. For the experiments, we use $p = \{30, 50, 70\}$ . + +Split-4 [8, 10]: This split requires evaluating models on three different settings namely easy, hard and extreme ensuring at least one annotation per image constructed using PASCAL-VOC 2007+12 [11, 12] trainval set. Results are reported on the PASCAL-VOC 2007 [11] test set. For each image in the training set, the easy setting randomly removes one annotation while the hard setting randomly removes half of the annotations. The extreme setting retains only one annotation per image. All the sets ensure each image consists of atleast one annotation. This split is a small scale version for an image level class agnostic split. + +Split-5 [2]: This split uses the PASCAL-VOC 2007 [11] train set for training and the PASCAL-VOC 2007 test set for evaluation and drops $p\%$ annotations per class. For each image in this construction, instances of randomly selected categories are exhaustively annotated while the remaining categories do not have any annotations. This is a small scale version of Split-1 as annotations are dropped in a class aware manner across the full dataset maintaining at least 1 annotation per image. In this case we use $p = \{30,40,50\}$ . + +Table 2: Sparsely annotated object detection results on two splits of PASCAL-VOC dataset. "Oracle" corresponds to training models using all annotations. Results are reported on VOC 07 test set using $\mathrm{AP}_{50}$ . + +
MethodSplit-4Split-5
EasyHardExtreme30%40%50%
Oracle83.0977.47
BRL [8]73.5071.7066.20---
Co-mining [10]79.5978.3869.6074.4273.3069.89
Ours82.1581.5075.5976.8475.8874.35
+ +Table 3: Results on the proposed SSL+SAOD setup. VOC12 [12] is used as the unlabeled data and $p$ is the removal percentage + +
MethodAP
30%40%50%
Co-mining [10]46.5345.9843.21
Ours48.3447.8246.76
+ +![](images/4b0d8db6da3ab120430d6623533b8f8a68f5cfa0e0d86ea080525681d12f77e0.jpg) +Figure 3: Illustration of the effects of various augmentations used in this work. + +![](images/127cdbd84701c76eb453ef5e74577620e145a95d25e394549efc8c2306084aca.jpg) + +![](images/d877a288dc50ae2b0bdf230522fc51238791e617fa1acf058275223dd40ff44c.jpg) + +![](images/a4c38e3645337dd3cd2a68f2f44865940f4a5083c98cb91de3af34f37b908f4a.jpg) + +![](images/66cbe7a543613e3cef39c4809b5ff449544eb441c554dc702ce7bb9af25e0a0c.jpg) + +![](images/83351c1a395a0db47470a9a3cfdf3c99c7854f3a9d255567f8168142c8b5ca5c.jpg) + +![](images/28bbe280de4a8346d7c59105a89e02c486893a625eb0961f8c6c2e29609189da.jpg) + +# 4.3. Implementation details + +For all our experiments, we use a Faster RCNN [28] architecture with a ResNet-101 [45] FPN backbone [30] implemented using Detector2 [46] framework. For the image augmentations, we apply a For augmentation, a cascade of random contrast, brightness, saturation, lighting and bounding box erase augmentations. The effect of this augmentation is shown in Figure 3. We train all our models with a batch size of 8 for 270000 and 18000 iterations on the COCO and PASCAL-VOC splits respectively with a learning rate of 0.01. The learning rate is decreased ten fold twice at \{210000, 250000\} and \{12000, 15000\} for COCO and PASCAL-VOC respectively. We adopt a warm-up strategy for 1000 and 100 iterations for the COCO and PASCAL-VOC respectively. Following existing detector implementations we set $\tau_{\mathrm{bg}}$ , $\tau_{\mathrm{fg}}$ , $\tau_{\mathrm{obj}}$ and $\tau_{\mathrm{ppm}}$ to 0.2, 0.4, 0.5 and 0.8 respectively. We set $\tau_{m}$ to 0.9. During inference, we compute the backbone features ( $f_{\mathrm{o}}$ and $f_{\mathrm{a}}$ ) for both the original and augmented versions of the input to obtain the RoIs and only $f_{\mathrm{o}}$ is used for the final detections. + +# 4.4. Baselines + +We compare our method against Co-Mining [10], BRL [8], and Pseudo Label [7]. + +We choose these methods as they outperform previous approaches for the SAOD task. We use the public implementation of these methods and use a ResNet 101 FPN [30] backbone for all the experiments in order to perform a fair comparison. + +# 4.5. Comparison with state-of-the-art + +In this section, we compare our method with contemporary methods. We evaluate all the models in two different setups. We name the first setup Sparsely annotated setup which evaluates models on SAOD. The second setup, contains labeled and unlabeled images and regions. + +Sparsely annotated setup: We show results of this setup in Table 1 and Table 2. Splits-2 and 3 ensure at least one annotation per image is retained. On the other hand, Split-1 doesn't ensure this and has significantly more unlabeled data than the other splits. Splits-4 and 5 have no unlabeled images at all. In both the tables, the rows named "oracle" refers to the models trained using all annotations. + +From Table 1, we observe that our approach outperforms the other baselines and is closer to the oracle performance on the $30\%$ splits. Co-mining [10], performs competitively on all the splits at $30\%$ . + +On Split-1, our method obtains a performance improvement of 1.87, 3.08 and 7.75 on $30\%$ , $40\%$ and $50\%$ sparsity respectively over Co-mining [10]. + +On Split-2, our method obtains an improvement of 3.04, 3.9 and 10.5 at $30\%$ , $40\%$ and $50\%$ sparsity respectively over the previous state of the art Co-mining. On Split-3, an again we see consistent improvement of 2.8, 4.61 and 10.52 at $30\%$ , $40\%$ and $50\%$ sparsity respectively over Co-mining. + +In particular, on the hardest setting $(70\%)$ in Split 1-3, we demonstrate that current SAOD methods struggle at higher sparsity in the labeled data. This is because the performance of current SAOD approaches rely on the quality + +Table 4: Ablation of various components of proposed approach. + +
C-RPNPPM\( {\mathcal{L}}_{\text{SSL }}^{\mathrm{u}} \)\( {\mathcal{L}}_{\text{det,a }}^{\mathrm{s}} \)AP\( \Delta \)
✘✘✘✘41.77-
✘✘✘✓42.670.90
✓✘✘✓45.303.53
✓✓✘✓45.443.67
✓✓✓✘45.513.74
✘✓✓✓44.863.09
✓✓✓✓46.004.23
+ +of pseudo labels which degrades at higher sparsity. Our proposed PPM prevents penalizing the classifier irrespective of the quality of pseudo labels and utilizes self-supervised loss to benefit from mined regions. On average, all the methods achieve lower performance on Split-1 compared to the other splits due to the nature of its creation, i.e., class aware dataset level fashion of sparsity. + +From Table 2, we see a similar trend as above. All the methods perform competitively on the easier settings of both splits. However, the performance gap between the our approach and Co-mining [10] increases at higher sparsity. + +On Split-4, we get an improvement of 2.56, 3.12 and $5.99\mathrm{(AP_{50})}$ percentage points on the easy, hard and extreme settings respectively over Co-mining. On Split-5, an improvement of 2.42, 2.58 and 4.46 points was observed. For details on more experiments refer to supplementary material. + +Single Instance Object Detection (SIOD) setup: [47] proposed SIOD where only one instance per category is annotated in every image. While this setup has its benefits [47], note that it is a special case of SAOD. For this setup, we obtain an AP of 32.89 (compared to 31.9 of [47]) which is an improvement of $\sim 1$ mAP. + +$SSL + SAOD$ setup: We propose a semi-supervised learning benchmark for SAOD. This benchmark entails training models on a sparsely annotated labeled and an unlabeled set. As shown in Figure 4, this setup introduces two kinds of sparsity in the data, namely, sparsity in labeled regions (left) and images (right). We believe this is a realistic setup and SAOD methods must be capable of tackling both these kinds of sparsity in the data. For this setup, we use Split-5 (Section 4.2) with increasing sparsity as the labeled set and VOC12 trainval as the unlabeled set. We use the COCO-style AP metric to report results on this setup. + +In Table 3, we compare against Co-mining [10] and ob + +![](images/6400e3da83d042143b2db15d36521d5f28c363d9f99f4a510a52b093a23f2e6d.jpg) +Sparsity in labeled regions Sparsity in labeled images +Figure 4: Types of sparsity in labeled data; sparsity in labeled regions (left) and images (right). Our proposed SSL+SAOD is a realistic setup that presents challenges from both kinds of sparsity. + +serve an improvement of 1.81, 1.84 and $3.55\mathrm{mAP}$ on the $30\%$ , $40\%$ and $50\%$ sparsity levels respectively. We observe that the gap in performance increases with sparsity, consistent with our observation for Tables 1 and 2. This can be attributed to the inability of methods like Co-mining to handle unlabeled images and high sparsity in the labeled data. We will make this benchmark public and propose this method as a baseline. We encourage researchers to report results on this benchmark in the future. + +# 4.6. Ablation Experiments + +In this section we conduct ablation experiments to understand the various components. For the ablation experiments, we use a ResNet-101 as the backbone network and train on Split-5 with $p = 50\%$ . We evaluate on the VOC07 test set and report the COCO style AP metric. + +From Table 4, a baseline model trained on the ablation set attains 41.77 (row 1). Pseudo-labeling $\left(\mathcal{L}_{\mathrm{det},a}^{s}\right)$ improves the performance by 0.9 (row 2). Co-mining [10] relies extensively on pseudo-labels. This results in a drop in performance at higher sparsities due to noisy pseudo-labels. Addition of the C-RPN improves our performance by 3.53 points (row 3). C-RPN reduces the overhead of computing two sets of proposals and learns a better notion of objects due to the combined processing of features from the two views. + +The combination of C-RPN and PPM improves the performance by 3.67 (row 4). We do not observe major improvements with the introduction of PPM because its task is to identify and segregate the unlabeled regions from the backgrounds. After the segregation, PPM does no further processing to improve performance. The power of PPM can be observed when trained in conjunction with consistency regularization loss $(\mathcal{L}_{\mathrm{SSL}}^{\mathrm{u}})$ which achieves the best performance of 46 (row 7); an improvement of 4.23 points over the baseline. We distinguish ourselves from pseudolabeling approaches like Co-mining in one important as + +Table 5: Analysis of the threshold used for PPM. + +
Threshold0.60.70.80.90.95
AP45.9645.9646.0045.7745.30
+ +Table 6: Removing test time augmentations. + +
MethodSplit-1Split-2Split-3
30%50%70%30%50%70%30%50%70%
TTA38.2235.9232.6839.7636.9435.3339.5637.1535.48
w/o TTA38.3235.9132.6739.7736.9535.2739.5537.1235.48
+ +![](images/56943a3ba02f47679770eddb80e9fb15362424e65c36d5c6ac1087733f0f1e0e.jpg) + +![](images/44e100c8e228b1bd3e9da7960054013d85c3fd74bb048854d8cb4341632af08a.jpg) + +![](images/466bca1565ecfaf5c8d0e149a98af8c30b9580b0f2b855287b34def292c489aa.jpg) + +![](images/4908f32da58fa3d5d9cd90317e1cfdbc5e9aef842392af01a4f3e646107eb5e6.jpg) + +![](images/9b421c95f9ca47144aa642efa3bcf377d5e94516a4155002a83d06f4dbb0a39f.jpg) + +![](images/a8e414c0a1ee5cce2a71c8b983479d7b13bbb77761ec61b795c94412bb934734.jpg) + +![](images/dd5c22dd6d8b6849cc3bfc897911f90b5b5f9ac871e0e29d17e74b313f3df2dd.jpg) + +![](images/ffb07e03968fe33caef460caaa04fd813f19297ac479345927f7743c05ff07d1.jpg) + +![](images/1207c722a40a13ae5e31b049fb3f2112de99f7a0fa8ee0184cb142fd6d2fc711.jpg) + +![](images/5d3cfe20723706485bfb81d4cf1511fe13138de8a36f184af30c21875ac487e2.jpg) + +![](images/9cde6433b013e88104b5cd54b69ca6d65c76cc59cbf56030a841b18b071a8736.jpg) +Figure 5: Qualitative results showing the unlabeled regions identified by the PPM. The red boxes correspond to the available ground truth. A class agnostic NMS was performed on the regions and the result is shown in white. + +![](images/0be5711e10edba967aaacfb2d3a05878dc5b057fa58923d6d89e05723d2a1a7c.jpg) + +![](images/3f79d2bd3b22a013b842a2c93c7d37aec3e4d53eadc8b0c475106a09f8d6d398.jpg) + +![](images/f944d4b57b2e453d1015727317dcc28da07075c7c8a05b422363f8c86382d61f.jpg) + +![](images/ead32e35cc7a802167f31af89bba5dcfcff9310a723a77b0a6c6fe6e70b11915.jpg) + +![](images/cf397c441ca6cbf2a46b884debf60b9c692384341352795a76e2ee36752ab9f4.jpg) + +![](images/12dda31c9b2760096dc90be00c9e4ec904ad9d71780bbf4fba533d6d22d766d3.jpg) +Figure 6: Qualitative results comparing the output of a model trained using available ground truths (top) to a model trained using our approach (bottom). Predictions with a class confidence score greater than 0.9 are shown. Red: Person, Cyan: Dog, Purple: Horse, Yellow: Clock, Green: Stop sign, Blue: Parking meter, Violet: Giraffe, Orange: Potted plant, Black: Surfboard, Dark green: Boat + +![](images/394fb5c3ffa80d65a890f65208469d3c7d0736f1b36c292a00689b8d0a85103a.jpg) + +![](images/455ef43f4639b77690f9cecbddaa6d68070c3d0c55efdf9597f600687341aa59.jpg) + +![](images/c6f45432bfdcd2ee1c69c207e0c45c04b914db15b589ae5c85bd28e2f5f17c2d.jpg) + +![](images/4239b9e8a9242fa7ea905bf76922bc5ea9853025f0230dba4f3079b0a2b9d34d.jpg) + +![](images/a208970510f4bb326390f91d87805ba0146fd55c5660dc19fd7b10aa0d4a9a54.jpg) + +pect. While, Co-mining [10] relies solely on pseudo-labels, we leverage additional components from self-supervised learning along with pseudo-labeling. In row 5, due to C-RPN, PPM and $\mathcal{L}_{\mathrm{SSL}}^{\mathrm{u}}$ , we show an improvement of 3.74 over the baseline. Our proposed components are orthogonal to pseudo-labeling as using them together results in an additional improvement of $\sim 0.5$ on mAP. At higher sparsity in the labeled dataset, the proposed C-RPN, PPM and $\mathcal{L}_{\mathrm{SSL}}^{\mathrm{u}}$ are less affected than pseudo-labels resulting in the large improvements on these splits. Finally, we show the effect of C-RPN by generating a single set of proposals from the + +original image and using it for both the branches. We observe a drop in performance (row 6) highlighting the effectiveness of C-RPN. + +PPM mines potential positives which can be mistaken for negatives due to missing annotations. We rely on the objectness score of the RPN to identify these regions. In Table 5, we vary the threshold of PPM. For a low threshold, a few hard negatives might also be identified as pseudo-positive leading to a drop in performance. With a high threshold, a few potential positives might not be mined. We observe that a threshold of 0.8 provides a good trade-off and is there- + +Table 7: Comparison with a different backbone. Results are reported on the COCO validation set on Splits 1-3 using AP and on VOC 2007 test set on Splits-4,5 using $\mathrm{AP}_{50}$ . + +
MethodSplit-1Split-2Split-3Split-4Split-5
30%50%70%30%50%70%30%50%70%EasyHardExtreme30%40%50%
Ours (C4)37.6735.9533.1639.2236.8134.9838.8436.7635.2680.5680.4374.4577.3876.1375.32
Ours (FPN)38.2235.9232.6839.7636.9435.3339.5637.1535.4882.1581.5075.5976.8475.8874.35
+ +fore used for all experiments unless stated otherwise. In all our experiments, PPM is performed after an initial warmup of 9000 and 30000 iterations on the PASCAL-VOC and COCO datasets respectively. + +# 4.6.1 Inference without augmentations + +Our method requires passing an input with augmentations during inference as well. It should be noted that this is not a test-time augmentation (TTA), a technique that typically involves passing images at a higher resolution. We perform inference by removing the augmentation and extracting the region proposals using C-RPN. We show the results in Table 6 on the three splits of COCO. We do not observe a significant improvement in performance due to the augmentation. For the analysis on the effect of augmentations refer to supplementary material. + +# 4.6.2 Effect of backbone + +We analyze the effect of backbone on our approach and show results for the normal convolution based (C4) and FPN based (FPN) backbones in Table 7. The first row corresponds to C4 while the second row corresponds to FPN. While the gap in performance is lower in most cases, we observe a significant improvement using the FPN on the low sparsity settings of all the split. + +# 4.7. RPN Recall experiments for object discovery + +PPM identifies foreground regions mistakenly assigned as background during training to avoid penalizing the network. To study the effect of PPM on novel [48] classes, we train a network using our approach on randomly chosen 6000 images of the COCO dataset, containing annotations for only 20 classes of PASCAL-VOC. We evaluate the recall@0.5 of the RPN over the remaining 60 classes. A model trained using the standard technique on this dataset achieves a recall@0.5 of $77.46\%$ and $29.06\%$ on the known classes (20 categories) and unknown classes (60 categories) respectively. Our proposed approach, with PPM, achieves a recall@0.5 of $78.47\%$ and $35.20\%$ respectively. This ability to localize objects not seen during training can be beneficial for object discovery methods like [49-52] which use RPN proposals to learn/discover new categories. + +# 4.8. Qualitative Results + +In Figure 5, we show the pseudo positives mined by PPM. In each figure, the red boxes correspond to the ground truth annotations and the white boxes correspond to the post NMS pseudo positive boxes mined by PPM. We observe that the PPM correctly mines proposals which correspond to missing object annotations. Without PPM, these regions will be used as negatives to the classifier resulting in a reduced class confidence score leading to a drop in mAP. We show the detection results of a model trained on the $50\%$ Split-1 of our approach in Figure 6. The images on the top corresponds to the model trained using sparse annotations and the bottom image shows the output of our approach. + +# 5. Conclusion + +We present SparseDet, a novel end-to-end system for Sparsely Annotated Object Detection (SAOD). We propose a simple yet effective technique for identifying the unlabeled regions using pseudo-positive mining and apply self-supervised loss on them. Through qualitative results we highlight the ability of PPM to mine pseudo-positives. We standardize the evaluation setup and show the effectiveness of our approach with an exhaustive set of experiments on multiple splits of SAOD. While we outperform existing state-of-the-art on all metrics and splits, we observe the gap in performance increases with sparsity demonstrating the short coming of methods that rely solely on pseudolabeling. We propose a new benchmark, that evaluates the semi-supervised learning capabilities of SAOD approaches. We will release the data for the new benchmark along with all the SAOD splits and encourage researchers to evaluate future SAOD methods on these. + +Acknowledgements. This project was partially funded by DARPA SAIL-ON (W911NF2020009), DARPA SemaFor (HR001119S0085), and IARPA SMART (2021-21040700003) programs. Rama Chellappa was supported by an ONR MURI Grant N00014-20-1-2787. We would like to thank our colleagues from the PI lab for their feedback on this work. The views and conclusions contained herein are those of the authors and do not represent views, policies, and/or endorsements of funding agencies. + +# References + +[1] Simon Chadwick and Paul Newman. Training object detectors with noisy data. 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1319-1325, 2019. 1 +[2] Zhe Wu, Navaneeth Bodla, Bharat Singh, Mahyar Najibi, Rama Chellappa, and Larry S. Davis. Soft sampling for robust object detection. In BMVC, 2019. 1, 2, 3, 5 +[3] Junnan Li, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Towards noise-resistant object detection with noisy annotations. ArXiv, abs/2003.01285, 2020. 1 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1 +[5] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 2, 5 +[6] Alina Kuznetsova, Hassan Rom, Neil Gordon Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4. International Journal of Computer Vision, 128:1956-1981, 2020. 1 +[7] Yusuke Niitani, Takuya Akiba, Tommi Kerola, Toru Ogawa, Shotaro Sano, and Shuji Suzuki. Sampling techniques for large-scale object detection from sparsely annotated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6510-6518, 2019. 1, 2, 5, 6 +[8] Han Zhang, Fangyi Chen, Zhiqiang Shen, Qiqi Hao, Chenchen Zhu, and Marios Savvides. Solving missing-annotation object detection with background recalibration loss. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1888-1892. IEEE, 2020. 3, 5, 6 +[9] Yuewei Yang, Kevin J Liang, and Lawrence Carin. Object detection as a positive-unlabeled problem. In BMVC, 2020. 5 +[10] Tiancai Wang, Tong Yang, Jiale Cao, and X. Zhang. Comining: Self-supervised learning for sparsely annotated object detection. In AAAI, 2021. 1, 2, 3, 5, 6, 7, 8 +[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2007 (VOC2007) Results. http://www.pascalnetwork.org/challenges/VOC/voc2007/workshop/index.html, 2, 5 +[12] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, 2, 5, 6 + +[13] Ke Yan, Jinzheng Cai, Adam P. Harrison, Dakai Jin, Jing Xiao, and Le Lu. Universal lesion detection by learning from multiple heterogeneously labeled datasets. ArXiv, abs/2005.13753, 2020. 2 +[14] Hansheng Li, Xin Han, Yuxin Kang, Xiaoshuang Shi, Mengdi Yan, Zixu Tong, Qirong Bu, Lei Cui, Jun Feng, and Lin Yang. A novel loss calibration strategy for object detection networks training on sparsely annotated pathological datasets. In MICCAI, 2020. +[15] Yongqiang Zhang, Mingli Ding, Yancheng Bai, Mengmeng Xu, and Bernard Ghanem. Beyond weakly supervised: Pseudo ground truths mining for missing bounding-boxes object detection. IEEE Transactions on Circuits and Systems for Video Technology, 30:983-997, 2020. +[16] Ke Yan, Jinzheng Cai, Youjing Zheng, Adam P. Harrison, Dakai Jin, You-Bao Tang, Yuxing Tang, Lingyun Huang, Jing Xiao, and Le Lu. Learning from multiple datasets with heterogeneous and partial labels for universal lesion detection in ct. IEEE Transactions on Medical Imaging, 40:2759-2770, 2021. 2 +[17] Jisoo Jeong, Seungeui Lee, Jeesoo Kim, and Nojun Kwak. Consistency-based semi-supervised learning for object detection. Advances in neural information processing systems, 32:10759-10768, 2019. 2 +[18] Peng Tang, Chetan Ramaiah, Yan Wang, Ran Xu, and Caiming Xiong. Proposal learning for semi-supervised object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2291-2301, 2021. 2 +[19] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2 +[20] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. arXiv preprint arXiv:2102.09480, 2021. +[21] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. arXiv preprint arXiv:2106.09018, 2021. +[22] Qize Yang, Xihan Wei, Biao Wang, Xian-Sheng Hua, and Lei Zhang. Interactive self-training with mean teachers for semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5941-5950, 2021. 2 +[23] Jihun Yoon, Seungbum Hong, and Min-Kook Choi. Semi-supervised object detection with sparsely annotated dataset. In 2021 IEEE International Conference on Image Processing (ICIP), pages 719-723. IEEE, 2021. 3 + +[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097-1105. Curran Associates, Inc., 2012. 3 +[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. doi: 10.1007/s11263-015-0816-y. 3 +[26] Ross Girshick. Fast r-cnn. In Proceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), ICCV '15, pages 1440-1448, Washington, DC, USA, 2015. IEEE Computer Society. ISBN 978-1-4673-8391-2. doi: 10.1109/ICCV.2015.169. URL http://dx.doi.org/10.1109/ICCV.2015.169.3 +[27] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 3 +[28] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 3, 4, 6 +[29] Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. R-fcn: Object detection via region-based fully convolutional networks. In Daniel D. Lee, Masashi Sugiyama, Ulrike von Luxburg, Isabelle Guyon, and Roman Garnett, editors, NIPS, pages 379-387, 2016. URL http://dblp.uni-trier.de/db/conf/nips/nips2016.html#DaiLHS16.3 +[30] Tsung-Yi Lin, Piotr Dólár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 936-944, 2017. 3, 6 +[31] Bharat Singh and Larry S Davis. An analysis of scale invariance in object detection-snip. CVPR, 2018. +[32] Bharat Singh, Mahyar Najibi, and Larry S Davis. SNIPER: Efficient multi-scale training. NeurIPS, 2018. +[33] Mahyar Najibi, Bharat Singh, and Larry S Davis. AutoFocus: Efficient multi-scale inference. ICCV, 2019. 3 +[34] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. doi: 10.1109/ICCV.2017.322.3 +[35] Alexander Kirillov, Ross B. Girshick, Kaiming He, and Piotr Dóllár. Panoptic feature pyramid networks. In CVPR, pages 6399-6408. Computer Vision Foundation / IEEE, 2019. 3 + +[36] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In CVPR, pages 9404-9413. Computer Vision Foundation / IEEE, 2019. 3 +[37] Justin Johnson Georgia Gkioxari, Jitendra Malik. Mesh r-cnn. In ICCV, 2019. 3 +[38] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 3 +[39] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017. +[40] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection, 2015. URL http://arxiv.org/abs/1506.02640. cite arxiv:1506.02640.3 +[41] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 3 +[42] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 3 +[43] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:386-397, 2020. 3 +[44] Ross B. Girshick. Fast R-CNN. 2015 IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 4 +[45] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. doi: 10.1109/CVPR.2016.90.6 +[46] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6 +[47] Hanjun Li, Xingjia Pan, Ke Yan, Fan Tang, and Wei-Shi Zheng. Siod: Single instance annotated per category per image for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14197-14206, 2022. 7 + +[48] Akshay Raj Dhamija, Manuel Günther, Jonathan Ventura, and Terrance E. Boult. The overlooked elephant of object detection: Open set. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1010-1019, 2020. doi: 10.1109/WACV45572.2020.9093355. 9 +[49] Huy V. Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 9 +[50] Huy V. Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 8287-8296, 2019. +[51] Sai Saketh Rambhatla, Rama Chellappa, and Abhinav Shrivastava. The pursuit of knowledge: Discovering and localizing novel categories using dual memory. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. +[52] Sai Saketh Rambhatla, Ishan Misra, Rama Chellappa, and Abhinav Shrivastava. The pursuit of knowledge: Discovering and localizing novel categories using dual memory. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 9 \ No newline at end of file diff --git a/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/images.zip b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8cac2a63647d40caf41c6ca12d064af61b875006 --- /dev/null +++ b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a1ebbba10b536350fc36a03c56149347908b6b029cd61feb5292c0ddf670206 +size 651659 diff --git a/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/layout.json b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..179fb3567848f6903a14458b3f3aa5c94c0534b1 --- /dev/null +++ b/2023/SparseDet_ Improving Sparsely Annotated Object Detection with Pseudo-positive Mining/layout.json @@ -0,0 +1,10322 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 498, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 498, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 498, + 140 + ], + "type": "text", + "content": "SparseDet: Improving Sparsely Annotated Object Detection with Pseudo-positive Mining" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 161, + 153, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 161, + 153, + 175 + ], + "spans": [ + { + "bbox": [ + 78, + 161, + 153, + 175 + ], + "type": "text", + "content": "Saksham Suri\\*1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 178, + 168, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 178, + 168, + 188 + ], + "spans": [ + { + "bbox": [ + 64, + 178, + 168, + 188 + ], + "type": "text", + "content": "sakshams@cs.umd.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 161, + 288, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 161, + 288, + 175 + ], + "spans": [ + { + "bbox": [ + 192, + 161, + 288, + 175 + ], + "type": "text", + "content": "Saketh Rambhatla*1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 196, + 178, + 285, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 178, + 285, + 188 + ], + "spans": [ + { + "bbox": [ + 196, + 178, + 285, + 188 + ], + "type": "text", + "content": "rssaketh@umd.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 161, + 399, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 399, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 399, + 175 + ], + "type": "text", + "content": "Rama Chellappa2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 178, + 400, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 400, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 400, + 189 + ], + "type": "text", + "content": "rchella4@jhu.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 425, + 161, + 528, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 161, + 528, + 175 + ], + "spans": [ + { + "bbox": [ + 425, + 161, + 528, + 175 + ], + "type": "text", + "content": "Abhinav Shrivastava" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 427, + 178, + 526, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 178, + 526, + 188 + ], + "spans": [ + { + "bbox": [ + 427, + 178, + 526, + 188 + ], + "type": "text", + "content": "abhinav@cs.umd.edu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 194, + 313, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 194, + 313, + 209 + ], + "spans": [ + { + "bbox": [ + 126, + 194, + 313, + 209 + ], + "type": "text", + "content": "University of Maryland, College Park1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 194, + 466, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 194, + 466, + 210 + ], + "spans": [ + { + "bbox": [ + 335, + 194, + 466, + 210 + ], + "type": "text", + "content": "Johns Hopkins University" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 262, + 290, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 290, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 290, + 526 + ], + "type": "text", + "content": "Training with sparse annotations is known to reduce the performance of object detectors. Previous methods have focused on proxies for missing ground truth annotations in the form of pseudo-labels for unlabeled boxes. We observe that existing methods suffer at higher levels of sparsity in the data due to noisy pseudo-labels. To prevent this, we propose an end-to-end system that learns to separate the proposals into labeled and unlabeled regions using Pseudo-positive mining. While the labeled regions are processed as usual, self-supervised learning is used to process the unlabeled regions thereby preventing the negative effects of noisy pseudo-labels. This novel approach has multiple advantages such as improved robustness to higher sparsity when compared to existing methods. We conduct exhaustive experiments on five splits on the PASCAL-VOC and COCO datasets achieving state-of-the-art performance. We also unify various splits used across literature for this task and present a standardized benchmark. On average, we improve by 2.6, 3.9 and 9.6 mAP over previous state-of-the-art methods on three splits of increasing sparsity on COCO. Our project is publicly available at cs.umd.edu/~sakshams/SparseDet." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 693 + ], + "type": "text", + "content": "The performance of object detectors is sensitive to the quality of labeled data [1-3]. Existing object detection methods assume that the training data is pristine and a drop in performance is observed if this assumption fails. Noise in the data used for training object detectors can arise due to incorrect class labels or incorrect/missing bounding boxes. In this work, we deal with the problem of training object detectors with sparse annotations, i.e., missing region or bounding boxes. This problem is of utmost importance, as obtaining crowd-sourced datasets [4, 5] can be expensive" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 310, + 240, + 545, + 373 + ], + "blocks": [ + { + "bbox": [ + 310, + 240, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 310, + 240, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 310, + 240, + 545, + 373 + ], + "type": "image", + "image_path": "c0088fc7addd2f7f741af579ba77c33d707eac23ba1e8dbb2f630f0abb57090e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 380, + 547, + 476 + ], + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 476 + ], + "type": "text", + "content": "Figure 1: (Top) Most Object Detection datasets have exhaustive annotations for foreground/positives. During training, the unlabeled regions can be safely considered as background/negatives. Sparsely Annotated Object Detection datasets (bottom) have missing annotations. This results in foreground regions (shown in red) being considered as negatives during training, adversely affecting the performance of the classifier." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 547, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 557 + ], + "type": "text", + "content": "and laborious. The alternative is to use computer-assisted protocols to collect annotations which have been shown to be noisy and incomplete [6]. This problem of training object detectors with incomplete bounding box annotations is called Sparsely Annotated Object Detection (SAOD)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "To understand why training with sparse annotations is detrimental to the performance, consider the example shown in Figure 1 (top). If the annotation were exhaustive, then the negative samples to the classifier contain true background regions. But with sparse annotations, as shown in Figure 1 (bottom), a few positive regions will inevitably be considered as negatives (shown in red), thereby wrongfully penalizing the classifier leading to lower performance. Existing methods[2, 7-10] prevent this by predicting pseudolabels and removing the foreground regions from the negatives to the classifier. However, at higher levels of sparsity, the quality of pseudo-labels is greatly affected, resulting in the same problem noted above." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 182, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 182, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 182, + 713 + ], + "type": "text", + "content": "*First two authors contributed equally" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6770" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "Crowd sourced object detection datasets [5, 11, 12] are ensured to be almost exhaustively labeled. Hence, for SAOD, researchers artificially create sparsely annotated datasets from the original ones. There is no general consensus on the correct way to create the sparsely annotated datasets, a.k.a. splits, and hence each method reports results on one or two different splits. Split can be created by considering the dataset as a whole or per image (i.e. annotations can be removed by considering all the images or only a single image at once). They can also be created by removing annotations in a class-agnostic or class-aware fashion (i.e. remove " + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": ", of annotations per category or across all the categories). These variations result in splits with different data distributions making some harder than the others (refer to Table 1). A proper benchmark that analyzes the performance of SAOD methods across these different types of splits is missing. This makes it difficult to compare methods and assess their effectiveness for a specific use case." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 291, + 289, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 602 + ], + "type": "text", + "content": "To tackle the issues discussed above, as our first major contribution, we present SparseDet, a novel SAOD framework that achieves state-of-the-art performance across multiple SAOD benchmarks in practice. SparseDet operates on an image and its augmented counterpart. The combination of features extracted from the two views is used to generate region proposals. Standard detection training methods, consider a region proposal as positive if its intersection over union (IoU) with any ground truth is greater than 0.5 and the rest are treated as negatives. This strategy works when the annotations are exhaustive, which implies that the remaining regions are from background. But due to missing annotations, some of these regions could belong to foreground instances. To prevent considering all region proposals without annotations as negatives, SparseDet partitions all the region proposals into labeled, unlabeled and background. The labeled and background regions are processed as usual. Features extracted from unlabeled regions are then trained with a self-supervised loss. Previous approaches like Co-Mining [10], consider two partitions, labeled and background, and generate pseudo-labels. This is a disadvantage at high sparsity as the generated pseudo-labels can be very noisy. The self-supervised loss in our approach enforces consistency between the features of the two views for the unlabeled regions and prevents penalizing the classifier due to false negatives." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 605, + 288, + 715 + ], + "type": "text", + "content": "Our second major contribution is unifying evaluation. The standard practice is to simulate sparsely annotated training data on COCO [5] and PASCAL-VOC [11] train sets and evaluate them on their corresponding standard validation set. As discussed above, a survey of recent SAOD approaches [2, 7-10, 13-16] reveals that there are at least five different ways to create splits, each differing in the strategy (refer to Section 4.2) used to achieve the desired level of sparsity. However, these splits have not been made public," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 252 + ], + "type": "text", + "content": "making it difficult to replicate results for comparison. Additionally, each of these strategies has a different property for simulating sparse data, e.g., different distribution of annotations per class, resulting in a different training set. As a result, methods trained on different splits cannot be compared with one another. To mitigate these issues we standardize the generation of these splits that enables the evaluation of any SAOD methods on all of them for fair comparison. Additionally, we propose a new benchmark that assesses the semi-supervised learning capabilities of SAOD methods, i.e., leveraging unlabeled data to improve performance. We present our approach as a baseline. We will make the data for the new benchmark along with all the SAOD splits public to facilitate future research. We briefly summarize our contributions below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 259, + 545, + 435 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 317, + 259, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 259, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 259, + 545, + 308 + ], + "type": "text", + "content": "- We propose a novel formulation, SparseDet, for SAOD which is an end-to-end approach that identifies labeled, unlabeled and background regions and deals with them in appropriate manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 312, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 312, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 317, + 312, + 545, + 372 + ], + "type": "text", + "content": "- We show state-of-the-art performance on sparsely annotated object detection across various splits. On average, we improve by 2.6, 3.9 and " + }, + { + "bbox": [ + 317, + 312, + 545, + 372 + ], + "type": "inline_equation", + "content": "9.6\\mathrm{mAP}" + }, + { + "bbox": [ + 317, + 312, + 545, + 372 + ], + "type": "text", + "content": " over previous state-of-the-art methods on three splits of increasing sparsity on COCO." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 376, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 545, + 435 + ], + "type": "text", + "content": "- We standardize the experimental setup for SAOD by evaluating methods on all the splits to facilitate future research. Additionally, we propose a new benchmark that evaluates the semi-supervised learning capabilities of SAOD methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 444, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 547, + 492 + ], + "type": "text", + "content": "In Section 2, we discuss the related works on SAOD and related fields. We describe our approach in detail in Section 3. We describe our experimental setup and present results in Section 4, and conclude in Section 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 392, + 514 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "content": "Semi-supervised object detection: Semi-supervised object detection (SSOD) is an active field that also deals with training object detectors with missing annotations. Existing works on semi-supervised object detection, have focused mainly on consistency regularization [17, 18] or pseudolabeling-based approaches [19-22]. The main idea behind these approaches is to perturb the images, or features, and apply a consistency regularization loss to enforce consistency between the predictions using a student teacher framework. However, typical SSOD methods assume a small exhaustively labeled set and a large unlabeled set for training. This is different from Sparsely annotated object detection (SAOD), which assumes a large training set which is sparsely labeled." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Sparsely annotated object detection: One of the initial works addressing this problem, by Niitani et al. [7], pro" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6771" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "poses utilizing logical relations between the co-occurrences of objects and pseudo-labeling. Yoon et al. [23] use object tracking to densely label objects across sparsely annotated frames along with single stage detectors to mitigate the negative effects of missing annotations. Wu et al. [2] propose a re-weighting approach where the gradients corresponding to region of interest are weighed as a function of overlap with ground truth instances. Improving upon the previous work, Zhang et al. [8] propose an automatic re-calibration strategy for single stage detectors where the negative branch is changed to take into account low confidence background predictions which might correspond to missing annotations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 219, + 289, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 289, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 289, + 304 + ], + "type": "text", + "content": "Finally, one of the most recent works, Co-mining [10] uses a co-training strategy by using two views of an image and predictions from one view along with the ground truth as supervision for the other view and vice versa. Our method doesn't solely rely on pseudo-labels and leverages a self-supervised loss to prevent propagating negative gradients to the model due to false negatives." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 306, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 289, + 715 + ], + "type": "text", + "content": "Fully supervised object detection: After the success of AlexNet [24] on the image classification challenge (ILSVRC 2012) [25], research on designing deep neural networks for object detection gathered more interest. First among the successful methods are the two-stage Region-based convolutional networks (R-CNN) family of detectors. Two-stage object detectors consists of 1) a region proposal stage which produces a set of candidate object bounding boxes followed by 2) a classification stage which classifies each candidate region as either belonging to a foreground object or \"background\". R-CNN processes a large amount of region proposals by cropping the input image and using a CNN backbone to extract features making it extremely slow. Fast R-CNN [26] was proposed to overcome this limitation. Fast R-CNN computes one convolution feature map for the whole image. RoI Pooling was introduced to pool the feature for each region of interest into a fixed spatial dimension. RoI pooling shares the computation among all the region proposals speeding up training and inference. Fully connected layers are applied on the fixed RoI pooled feature maps which are then passed to two sister heads, for classification and bounding box regression. The whole network is trained end-to-end with a multi-task loss avoiding the multi-stage training in R-CNN [27]. While Fast R-CNN improved the efficiency of its predecessor, test time computation bottleneck is still an issue because of the region proposal method employed. Faster R-CNN [28] proposed a region proposal network (RPN), an elegant solution that trains deep networks to predict region proposals for practically no additional computational overhead. RPN shares convolutional layers with Fast R-CNN [26] and at test time the cost of generating proposals is minimal. Faster R-CNN paved the way for more sophisticated and efficient two stage detection architectures [29-33]. Faster R-CNN" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": "has also been extended to achieve state-of-the-art instance segmentation [34], panoptic segmentation [35, 36], and 3D mesh generation [37] etc. A few limitations of two stage object detectors and anchor boxes has also inspired the family of single-stage [38-40] and anchor-free object detection systems [41, 42]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 153, + 373, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 153, + 373, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 153, + 373, + 167 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "text", + "content": "We present SparseDet for Sparsely Annotated Object Detection. Given " + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "text", + "content": " images, we denote the set of labeled regions in the dataset as " + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{l}} = \\{b_{i},c_{i}\\}_{i = 1}^{\\mathrm{N}_{\\mathrm{l}}}" + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "inline_equation", + "content": "(b_{*},c_{*})" + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "text", + "content": " are the bounding box and class labels respectively. The unlabeled regions are denoted as " + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{ul}} = \\{b_k\\}_{k = 1}^{\\mathrm{N}_{\\mathrm{ul}}}" + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "text", + "content": " and the background regions as " + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{bg}} = \\{b_q\\}_{q = 1}^{\\mathrm{N}_{\\mathrm{bg}}}" + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "text", + "content": ". Note that unlike existing SAOD approaches, we do not assume images to contain at least one labeled region in this work. The unlabeled " + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{ul}}" + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "text", + "content": " and background " + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{bg}}" + }, + { + "bbox": [ + 304, + 173, + 547, + 284 + ], + "type": "text", + "content": " sets are not known a-priori." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 288, + 373, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 373, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 373, + 300 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "text", + "content": "The proposed approach is shown in Figure 2 and consists of a backbone network that extracts features from the original and augmented views of an image. The common RPN (C-RPN), concatenates the features, to generate a set of region proposals. A region proposal " + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "text", + "content": " can belong to one of three groups, namely 1) labeled regions " + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "inline_equation", + "content": "b \\in \\mathcal{B}_1" + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "text", + "content": ", 2) unlabeled foreground regions " + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "inline_equation", + "content": "b \\in \\mathcal{B}_{\\mathrm{ul}}" + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "text", + "content": ", or 3) background regions " + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "inline_equation", + "content": "b \\in \\mathcal{B}_{\\mathrm{bg}}" + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "text", + "content": ". For a given set of ground-truth annotations, the first group, i.e. labeled regions can be automatically identified. The problem then reduces to identifying and separating the second group i.e. the unlabeled regions, from the background regions. Given all the region proposals, a pseudo-positive mining (PPM) step identifies the unlabeled regions and segregates them from the background regions. The labeled and unlabeled regions are trained using supervised and self-supervised losses respectively. We describe each stage in detail below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 515, + 418, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 515, + 418, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 515, + 418, + 528 + ], + "type": "text", + "content": "3.2. Feature Extraction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "text", + "content": ", an augmented version of " + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "text", + "content": " denoted as " + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(I)" + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "text", + "content": " is computed. In this work, we use random contrast, brightness, saturation, lighting and bounding box erase in a cascaded fashion to generate " + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(I)" + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "text", + "content": ". A backbone network is employed to extract two features, " + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{o}}" + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "text", + "content": ", from " + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(I)" + }, + { + "bbox": [ + 304, + 533, + 547, + 607 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 611, + 441, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 441, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 441, + 624 + ], + "type": "text", + "content": "3.3. Common RPN (C-RPN)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Two stage object detectors [28, 30, 43] use a region proposal network (RPN) [28] to generate regions of interest (RoIs). We propose C-RPN which concatenates " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{o}}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " to obtain the RoIs. This is different from existing approaches that use " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{o}}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " to generate two separate sets of RoIs. Operating on two sets increases the difficulty of identifying the labeled, unlabeled and background regions which" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6772" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 76, + 521, + 206 + ], + "blocks": [ + { + "bbox": [ + 62, + 76, + 521, + 206 + ], + "lines": [ + { + "bbox": [ + 62, + 76, + 521, + 206 + ], + "spans": [ + { + "bbox": [ + 62, + 76, + 521, + 206 + ], + "type": "image", + "image_path": "9725617a6613c754f8f36432dc8cc3994bab2ec28970cf7fe7fb2a7f718810ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 281 + ], + "type": "text", + "content": "Figure 2: Illustration of SparseDet for sparsely annotated object detection. Following feature extraction from the original and augmented image, a common set of proposals is generated by the common RPN (C-RPN). Using an end to end approach, we identify and mine proposals corresponding to missing annotations using pseudo positive mining (PPM). We train the network end-to-end using a combination of supervised and self-supervised losses. The unlabeled instances (black) are supervised with a self-supervised loss and the labeled instances (green) are supervised with ground truth annotations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 287, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 288, + 335 + ], + "type": "text", + "content": "are required for subsequent stages. The sparse ground truth is used as supervision to train the C-RPN with a binary cross entropy loss for foreground classification and smooth L1 [44] loss for bounding box regression as shown below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 354, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r p n}} ^ {\\mathrm {s}} \\left(\\mathbf {c}, \\mathbf {b}, \\mathbf {c} ^ {*}, \\mathbf {b} ^ {*}\\right) = \\mathcal {L} _ {\\mathrm {b c e}} \\left(\\mathbf {c}, \\mathbf {c} ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {r e g}} \\left(\\mathbf {b}, \\mathbf {b} ^ {*}\\right) \\tag {1}", + "image_path": "b859a3a2103f00735f74600370a667573d9e6cff9461efd9aefa4fc4e4b9b692.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "content": " are the objectness scores and bounding boxes, and " + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^*" + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^*" + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "content": " are the corresponding ground truth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 406, + 213, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 213, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 213, + 419 + ], + "type": "text", + "content": "3.4. Pseudo Positive Mining (PPM)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "text", + "content": "Given the RoIs from C-RPN, the next step is to segregate the unlabeled regions from the labeled and background regions. A standard practice in training detectors is to consider all proposals with an objectness score greater than " + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "text", + "content": " and IoU greater " + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{fg}}" + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "text", + "content": " with any ground truth as foreground and proposals with IoU less than " + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "text", + "content": " as background. In the presence of missing annotations, detectors are wrongly penalized because the IoU of RoIs corresponding to missing annotations with any ground truth is less than " + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "text", + "content": ". To avoid this, we mine \"potential\" foregrounds and consider them as unlabeled regions " + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "inline_equation", + "content": "(\\mathcal{B}_{\\mathrm{u}})" + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "text", + "content": ". The proposed PPM module is based on our observation that when trained with sparse annotations, the RPN can reliably distinguish foreground from background regions. We pick all RoIs that have an objectness score greater than " + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{ppm}}" + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "text", + "content": " and IoU less than " + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 425, + 288, + 628 + ], + "type": "text", + "content": " with any ground truth as the unlabeled regions. The remaining RoIs are assigned as background." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 635, + 100, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 100, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 100, + 647 + ], + "type": "text", + "content": "3.5. Losses" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "The pseudo positive mining step segregates the RoIs into labeled, unlabeled and background regions. An RoI pooling layer [28] extracts RoI pooled features for the labeled and background regions using the feature " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{0}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ". The detection head processes the RoI pooled features to predict class-wise" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 286, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 335 + ], + "type": "text", + "content": "probabilities and bounding box adjustments for each region. The ground truth is used to supervise the predictions by applying the cross entropy loss for classification and smooth L1 [44] for bounding box regression as shown below:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 339, + 545, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 339, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 319, + 339, + 545, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\det , 0} ^ {s} \\left(\\mathbf {c}, \\mathbf {b}, \\mathbf {c} ^ {*}, \\mathbf {b} ^ {*}\\right) = \\mathcal {L} _ {\\mathrm {c e}} \\left(\\mathbf {c}, \\mathbf {c} ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {r e g}} \\left(\\mathbf {b}, \\mathbf {b} ^ {*}\\right) \\tag {2}", + "image_path": "70ecd78568c9e2d09f841977c5056e80a6a1701a855723be197bc32cc1c13857.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "type": "text", + "content": " are the class and bounding box predictions, and " + }, + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^*" + }, + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^*" + }, + { + "bbox": [ + 304, + 357, + 547, + 393 + ], + "type": "text", + "content": " are the corresponding ground truth class labels and bounding boxes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "text", + "content": "All detections with a score greater than " + }, + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{m}}" + }, + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "text", + "content": " are combined with the ground truth followed by a class specific NMS to obtain the merged ground truth. It is ensured that no ground truth annotation is discarded during this step. The labeled and background regions along with " + }, + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "text", + "content": " are used to extract RoI pooled features which are then fed to the detector head. The predictions of the detector head are supervised with the merged ground truth with the following losses:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 495, + 545, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 495, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 315, + 495, + 545, + 509 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\det , \\mathrm {a}} ^ {\\mathrm {s}} (\\mathbf {c}, \\mathbf {b}, \\mathbf {c} ^ {*}, \\mathbf {b} ^ {*}) = \\mathcal {L} _ {\\mathrm {b c e}} \\left(\\mathbf {c}, \\mathbf {c} _ {\\mathbf {m}} ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {r e g}} (\\mathbf {b}, \\mathbf {b} _ {\\mathbf {m}} ^ {*}) \\tag {3}", + "image_path": "8543ade36df549a71dacb1a7f9039bdf77dd67fd87e52b2b7db480fc632f4600.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 513, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 513, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 513, + 545, + 548 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 513, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 304, + 513, + 545, + 548 + ], + "type": "text", + "content": " are the class and bounding box predictions from the detector head, and, " + }, + { + "bbox": [ + 304, + 513, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathbf{m}}^{*}, \\mathbf{b}_{\\mathbf{m}}^{*}" + }, + { + "bbox": [ + 304, + 513, + 545, + 548 + ], + "type": "text", + "content": " are the corresponding merged ground truths." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "text", + "content": "Finally, a class agnostic NMS is performed on the unlabeled regions " + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{u}}" + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "text", + "content": " (obtained after PPM described in Sec. 3.4). The unlabeled regions along with " + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{o}}" + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{a}}" + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "text", + "content": " are passed through the ROI pooling layer and the detection head to obtain " + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{o}}^{\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{a}}^{\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 549, + 545, + 643 + ], + "type": "text", + "content": " respectively. A self-supervised loss is applied that enforces the detection head features of the original and augmented regions to be consistent with each other as shown below:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 366, + 648, + 545, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 648, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 366, + 648, + 545, + 664 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S S L}} ^ {\\mathrm {u}} \\left(f _ {\\mathrm {a}} ^ {\\mathrm {d}}, f _ {\\mathrm {o}} ^ {\\mathrm {d}}\\right) = \\| f _ {\\mathrm {a}} ^ {\\mathrm {d}} - f _ {\\mathrm {o}} ^ {\\mathrm {d}} \\| _ {2} ^ {2} \\tag {4}", + "image_path": "a5529f129fcb4fdd7caef46ad90bf83d8d147dd00169a25203181100360d0922.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 668, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 668, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 305, + 668, + 545, + 691 + ], + "type": "text", + "content": "The network is trained end-to-end with both supervised and unsupervised losses as shown below:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 340, + 696, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 696, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 340, + 696, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = 0. 5 * \\left(\\mathcal {L} _ {\\text {d e t , o}} ^ {\\mathrm {s}} + \\mathcal {L} _ {\\text {d e t , a}} ^ {\\mathrm {s}}\\right) + \\mathcal {L} _ {\\mathrm {r p n}} ^ {\\mathrm {s}} + \\mathcal {L} _ {\\mathrm {S S L}} ^ {\\mathrm {u}} \\tag {5}", + "image_path": "90e5c55406f1d002028ec1ba1b3f51c696636cc3544763aa7647762f9a6cbb27.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6773" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 103, + 544, + 201 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 1: Sparsely annotated object detection results on three splits of COCO dataset. \"Oracle\" corresponds to training models using all annotations. Results are reported on the COCO validation set using AP[0.50:0.95]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 103, + 544, + 201 + ], + "lines": [ + { + "bbox": [ + 49, + 103, + 544, + 201 + ], + "spans": [ + { + "bbox": [ + 49, + 103, + 544, + 201 + ], + "type": "table", + "html": "
MethodSplit-1Split-2Split-3100%
30%50%70%30%50%70%30%50%70%
Oracle40.91
Pseudo Label [7]-27.50--------
BRL [8]-32.70--------
Co-mining [10]36.3532.8424.9336.7233.0424.8336.7632.5424.96-
Ours38.2235.9232.6839.7636.9435.3339.5637.1535.48-
", + "image_path": "06a9657027fd73d0d4a809763c37d70c7c4327f8df6f6e986754678bf857d4ba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "text", + "content": "Discussion: In the absence of ground truth annotations, i.e. for completely unlabeled images, the supervised losses " + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{det,o}}^{\\mathrm{s}},\\mathcal{L}_{\\mathrm{det,a}}^{\\mathrm{s}},\\mathcal{L}_{\\mathrm{rpn}}^{\\mathrm{s}})" + }, + { + "bbox": [ + 46, + 211, + 289, + 344 + ], + "type": "text", + "content": " cannot be computed. Our proposed approach leverages self-supervised consistency loss that does not need ground truth. This helps our approach leverage these unlabeled regions unlike contemporary SAOD methods. We claim that this is the first method to use self-supervised losses for SAOD. Even though Co-Mining [10] claims to use self-supervised learning it is technically co-learning. Ours is the first approach to use pseudo-labeling and a self-supervised loss to handle the sparse annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 354, + 192, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 192, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 192, + 367 + ], + "type": "text", + "content": "4. Experimental Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 288, + 445 + ], + "type": "text", + "content": "In this section, we describe the experiments to evaluate our proposed approach. In Sections 4.1 and 4.2, we describe data, splits, and metrics. In Section 4.3, we mention the implementation details followed by the baselines in Section 4.4. We compare with contemporary methods in Section 4.5, followed by an ablation study in Section 4.6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 151, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 151, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 151, + 464 + ], + "type": "text", + "content": "4.1. Data and Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 651 + ], + "type": "text", + "content": "We conduct all our experiments on the COCO [5] and PASCAL-VOC [11, 12] (2007+2012) datasets. The COCO [5] dataset consists of 118000 and 5000 images for training and validation respectively. Experiments on the PASCAL-VOC07 [11] are conducted on 5011 trainval images and performance is computed on 4952 images of the test set. The PASCAL-VOC12 version consists of 11530 (trainval) images for training and evaluation is done on the PASCAL-VOC07 test set. Following past literature [2, 7-10], we create five different splits (Section 4.2) and report results on them. For splits on the COCO [5] dataset, we use the standard COCO style Average Precision (AP). For splits on the PASCAL-VOC [11] dataset, we use the standard PASCAL-VOC style " + }, + { + "bbox": [ + 46, + 472, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 46, + 472, + 288, + 651 + ], + "type": "text", + "content": ", which is Average Precision computed at an IoU threshold of 0.5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 129, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 129, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 129, + 672 + ], + "type": "text", + "content": "4.2. SAOD Splits" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "An extensive review of sparsely annotated object detection methods reveals that there are at least five popular types of splits in use for creating training data. Most methods re" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 211, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 545, + 233 + ], + "type": "text", + "content": "port results on a subset of these making it harder to compare across methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 235, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 270 + ], + "type": "text", + "content": "We standardize the evaluation of SAOD methods by evaluating exhaustively on all the splits facilitating future research. We briefly describe the splits below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "content": "Split-1 [9, 10]: In this split, for each object category, " + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "content": " of annotations are randomly removed from the COCO [5] train set and results are reported on the validation set. This split simulates sparsity at dataset level in a class aware fashion. Note, this split can contains images with no annotations. We experiment with " + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "inline_equation", + "content": "p = \\{30, 50, 70\\}" + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "content": "Split-2: For each image in the COCO [5] train set, all annotations of " + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "content": " of all categories in that image removed and results are reported on the COCO validation set. This split can be considered as image level and class aware. We experiment with " + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "inline_equation", + "content": "p = \\{30, 50, 70\\}" + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "content": ". Note, this split might contain images with no annotations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 415, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 474 + ], + "type": "text", + "content": "Split-3: This split, which uses the COCO 2017 [5] train set for training and the validation set for evaluation, deletes " + }, + { + "bbox": [ + 304, + 415, + 545, + 474 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 304, + 415, + 545, + 474 + ], + "type": "text", + "content": " annotations in a class agnostic fashion for each image ensuring at least one annotation. This split is image level but class agnostic. For the experiments, we use " + }, + { + "bbox": [ + 304, + 415, + 545, + 474 + ], + "type": "inline_equation", + "content": "p = \\{30, 50, 70\\}" + }, + { + "bbox": [ + 304, + 415, + 545, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 475, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 605 + ], + "type": "text", + "content": "Split-4 [8, 10]: This split requires evaluating models on three different settings namely easy, hard and extreme ensuring at least one annotation per image constructed using PASCAL-VOC 2007+12 [11, 12] trainval set. Results are reported on the PASCAL-VOC 2007 [11] test set. For each image in the training set, the easy setting randomly removes one annotation while the hard setting randomly removes half of the annotations. The extreme setting retains only one annotation per image. All the sets ensure each image consists of atleast one annotation. This split is a small scale version for an image level class agnostic split." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Split-5 [2]: This split uses the PASCAL-VOC 2007 [11] train set for training and the PASCAL-VOC 2007 test set for evaluation and drops " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " annotations per class. For each image in this construction, instances of randomly selected categories are exhaustively annotated while the remaining categories do not have any annotations. This is a small scale version of Split-1 as annotations are dropped in a class aware manner across the full dataset maintaining at least 1 annotation per image. In this case we use " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "p = \\{30,40,50\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "6774" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 115, + 347, + 211 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 349, + 107 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 349, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 349, + 107 + ], + "type": "text", + "content": "Table 2: Sparsely annotated object detection results on two splits of PASCAL-VOC dataset. \"Oracle\" corresponds to training models using all annotations. Results are reported on VOC 07 test set using " + }, + { + "bbox": [ + 47, + 70, + 349, + 107 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 47, + 70, + 349, + 107 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 115, + 347, + 211 + ], + "lines": [ + { + "bbox": [ + 49, + 115, + 347, + 211 + ], + "spans": [ + { + "bbox": [ + 49, + 115, + 347, + 211 + ], + "type": "table", + "html": "
MethodSplit-4Split-5
EasyHardExtreme30%40%50%
Oracle83.0977.47
BRL [8]73.5071.7066.20---
Co-mining [10]79.5978.3869.6074.4273.3069.89
Ours82.1581.5075.5976.8475.8874.35
", + "image_path": "3d4716e035cbff5edcbfb6c35def3ce98447fe4abc1c60bb8707fba2f1bbd665.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 356, + 116, + 545, + 211 + ], + "blocks": [ + { + "bbox": [ + 354, + 71, + 545, + 108 + ], + "lines": [ + { + "bbox": [ + 354, + 71, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 354, + 71, + 545, + 108 + ], + "type": "text", + "content": "Table 3: Results on the proposed SSL+SAOD setup. VOC12 [12] is used as the unlabeled data and " + }, + { + "bbox": [ + 354, + 71, + 545, + 108 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 354, + 71, + 545, + 108 + ], + "type": "text", + "content": " is the removal percentage" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 356, + 116, + 545, + 211 + ], + "lines": [ + { + "bbox": [ + 356, + 116, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 356, + 116, + 545, + 211 + ], + "type": "table", + "html": "
MethodAP
30%40%50%
Co-mining [10]46.5345.9843.21
Ours48.3447.8246.76
", + "image_path": "0bad3be2ec215f49d2b8cc30ef557a02ff0918e7f522634c8b3aad84cdbdc0d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 217, + 119, + 312 + ], + "blocks": [ + { + "bbox": [ + 48, + 217, + 119, + 312 + ], + "lines": [ + { + "bbox": [ + 48, + 217, + 119, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 119, + 312 + ], + "type": "image", + "image_path": "4b0d8db6da3ab120430d6623533b8f8a68f5cfa0e0d86ea080525681d12f77e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 319, + 455, + 332 + ], + "lines": [ + { + "bbox": [ + 137, + 319, + 455, + 332 + ], + "spans": [ + { + "bbox": [ + 137, + 319, + 455, + 332 + ], + "type": "text", + "content": "Figure 3: Illustration of the effects of various augmentations used in this work." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 123, + 217, + 194, + 312 + ], + "blocks": [ + { + "bbox": [ + 123, + 217, + 194, + 312 + ], + "lines": [ + { + "bbox": [ + 123, + 217, + 194, + 312 + ], + "spans": [ + { + "bbox": [ + 123, + 217, + 194, + 312 + ], + "type": "image", + "image_path": "127cdbd84701c76eb453ef5e74577620e145a95d25e394549efc8c2306084aca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 194, + 217, + 264, + 312 + ], + "blocks": [ + { + "bbox": [ + 194, + 217, + 264, + 312 + ], + "lines": [ + { + "bbox": [ + 194, + 217, + 264, + 312 + ], + "spans": [ + { + "bbox": [ + 194, + 217, + 264, + 312 + ], + "type": "image", + "image_path": "d877a288dc50ae2b0bdf230522fc51238791e617fa1acf058275223dd40ff44c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 265, + 217, + 335, + 312 + ], + "blocks": [ + { + "bbox": [ + 265, + 217, + 335, + 312 + ], + "lines": [ + { + "bbox": [ + 265, + 217, + 335, + 312 + ], + "spans": [ + { + "bbox": [ + 265, + 217, + 335, + 312 + ], + "type": "image", + "image_path": "a4c38e3645337dd3cd2a68f2f44865940f4a5083c98cb91de3af34f37b908f4a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 336, + 217, + 406, + 312 + ], + "blocks": [ + { + "bbox": [ + 336, + 217, + 406, + 312 + ], + "lines": [ + { + "bbox": [ + 336, + 217, + 406, + 312 + ], + "spans": [ + { + "bbox": [ + 336, + 217, + 406, + 312 + ], + "type": "image", + "image_path": "66cbe7a543613e3cef39c4809b5ff449544eb441c554dc702ce7bb9af25e0a0c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 406, + 217, + 476, + 312 + ], + "blocks": [ + { + "bbox": [ + 406, + 217, + 476, + 312 + ], + "lines": [ + { + "bbox": [ + 406, + 217, + 476, + 312 + ], + "spans": [ + { + "bbox": [ + 406, + 217, + 476, + 312 + ], + "type": "image", + "image_path": "83351c1a395a0db47470a9a3cfdf3c99c7854f3a9d255567f8168142c8b5ca5c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 477, + 217, + 546, + 312 + ], + "blocks": [ + { + "bbox": [ + 477, + 217, + 546, + 312 + ], + "lines": [ + { + "bbox": [ + 477, + 217, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 477, + 217, + 546, + 312 + ], + "type": "image", + "image_path": "28bbe280de4a8346d7c59105a89e02c486893a625eb0961f8c6c2e29609189da.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 343, + 178, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 178, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 178, + 357 + ], + "type": "text", + "content": "4.3. Implementation details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "text", + "content": "For all our experiments, we use a Faster RCNN [28] architecture with a ResNet-101 [45] FPN backbone [30] implemented using Detector2 [46] framework. For the image augmentations, we apply a For augmentation, a cascade of random contrast, brightness, saturation, lighting and bounding box erase augmentations. The effect of this augmentation is shown in Figure 3. We train all our models with a batch size of 8 for 270000 and 18000 iterations on the COCO and PASCAL-VOC splits respectively with a learning rate of 0.01. The learning rate is decreased ten fold twice at \\{210000, 250000\\} and \\{12000, 15000\\} for COCO and PASCAL-VOC respectively. We adopt a warm-up strategy for 1000 and 100 iterations for the COCO and PASCAL-VOC respectively. Following existing detector implementations we set " + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{fg}}" + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{ppm}}" + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "text", + "content": " to 0.2, 0.4, 0.5 and 0.8 respectively. We set " + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "inline_equation", + "content": "\\tau_{m}" + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "text", + "content": " to 0.9. During inference, we compute the backbone features (" + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{o}}" + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{a}}" + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "text", + "content": ") for both the original and augmented versions of the input to obtain the RoIs and only " + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{o}}" + }, + { + "bbox": [ + 46, + 364, + 289, + 593 + ], + "type": "text", + "content": " is used for the final detections." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 606, + 114, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 606, + 114, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 606, + 114, + 618 + ], + "type": "text", + "content": "4.4. Baselines" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 627, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 287, + 651 + ], + "type": "text", + "content": "We compare our method against Co-Mining [10], BRL [8], and Pseudo Label [7]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "We choose these methods as they outperform previous approaches for the SAOD task. We use the public implementation of these methods and use a ResNet 101 FPN [30] backbone for all the experiments in order to perform a fair comparison." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 343, + 484, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 343, + 484, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 343, + 484, + 357 + ], + "type": "text", + "content": "4.5. Comparison with state-of-the-art" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 363, + 546, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 546, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 546, + 423 + ], + "type": "text", + "content": "In this section, we compare our method with contemporary methods. We evaluate all the models in two different setups. We name the first setup Sparsely annotated setup which evaluates models on SAOD. The second setup, contains labeled and unlabeled images and regions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 435, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 546, + 519 + ], + "type": "text", + "content": "Sparsely annotated setup: We show results of this setup in Table 1 and Table 2. Splits-2 and 3 ensure at least one annotation per image is retained. On the other hand, Split-1 doesn't ensure this and has significantly more unlabeled data than the other splits. Splits-4 and 5 have no unlabeled images at all. In both the tables, the rows named \"oracle\" refers to the models trained using all annotations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "text", + "content": "From Table 1, we observe that our approach outperforms the other baselines and is closer to the oracle performance on the " + }, + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "text", + "content": " splits. Co-mining [10], performs competitively on all the splits at " + }, + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 520, + 545, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 568, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 604 + ], + "type": "text", + "content": "On Split-1, our method obtains a performance improvement of 1.87, 3.08 and 7.75 on " + }, + { + "bbox": [ + 304, + 568, + 545, + 604 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 568, + 545, + 604 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 568, + 545, + 604 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 604 + ], + "type": "text", + "content": " sparsity respectively over Co-mining [10]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": "On Split-2, our method obtains an improvement of 3.04, 3.9 and 10.5 at " + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": " sparsity respectively over the previous state of the art Co-mining. On Split-3, an again we see consistent improvement of 2.8, 4.61 and 10.52 at " + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": " sparsity respectively over Co-mining." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "type": "text", + "content": "In particular, on the hardest setting " + }, + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "type": "inline_equation", + "content": "(70\\%)" + }, + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "type": "text", + "content": " in Split 1-3, we demonstrate that current SAOD methods struggle at higher sparsity in the labeled data. This is because the performance of current SAOD approaches rely on the quality" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6775" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 95, + 272, + 254 + ], + "blocks": [ + { + "bbox": [ + 48, + 75, + 306, + 87 + ], + "lines": [ + { + "bbox": [ + 48, + 75, + 306, + 87 + ], + "spans": [ + { + "bbox": [ + 48, + 75, + 306, + 87 + ], + "type": "text", + "content": "Table 4: Ablation of various components of proposed approach." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 95, + 272, + 254 + ], + "lines": [ + { + "bbox": [ + 85, + 95, + 272, + 254 + ], + "spans": [ + { + "bbox": [ + 85, + 95, + 272, + 254 + ], + "type": "table", + "html": "
C-RPNPPM\\( {\\mathcal{L}}_{\\text{SSL }}^{\\mathrm{u}} \\)\\( {\\mathcal{L}}_{\\text{det,a }}^{\\mathrm{s}} \\)AP\\( \\Delta \\)
✘✘✘✘41.77-
✘✘✘✓42.670.90
✓✘✘✓45.303.53
✓✓✘✓45.443.67
✓✓✓✘45.513.74
✘✓✓✓44.863.09
✓✓✓✓46.004.23
", + "image_path": "564cb0798b5f0563f6e6b9ff43d8ae9a4470bed923ec05e21c15246eb00b2b47.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 350 + ], + "type": "text", + "content": "of pseudo labels which degrades at higher sparsity. Our proposed PPM prevents penalizing the classifier irrespective of the quality of pseudo labels and utilizes self-supervised loss to benefit from mined regions. On average, all the methods achieve lower performance on Split-1 compared to the other splits due to the nature of its creation, i.e., class aware dataset level fashion of sparsity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 399 + ], + "type": "text", + "content": "From Table 2, we see a similar trend as above. All the methods perform competitively on the easier settings of both splits. However, the performance gap between the our approach and Co-mining [10] increases at higher sparsity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 399, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 471 + ], + "type": "text", + "content": "On Split-4, we get an improvement of 2.56, 3.12 and " + }, + { + "bbox": [ + 46, + 399, + 287, + 471 + ], + "type": "inline_equation", + "content": "5.99\\mathrm{(AP_{50})}" + }, + { + "bbox": [ + 46, + 399, + 287, + 471 + ], + "type": "text", + "content": " percentage points on the easy, hard and extreme settings respectively over Co-mining. On Split-5, an improvement of 2.42, 2.58 and 4.46 points was observed. For details on more experiments refer to supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 483, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 556 + ], + "type": "text", + "content": "Single Instance Object Detection (SIOD) setup: [47] proposed SIOD where only one instance per category is annotated in every image. While this setup has its benefits [47], note that it is a special case of SAOD. For this setup, we obtain an AP of 32.89 (compared to 31.9 of [47]) which is an improvement of " + }, + { + "bbox": [ + 46, + 483, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\sim 1" + }, + { + "bbox": [ + 46, + 483, + 287, + 556 + ], + "type": "text", + "content": " mAP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "inline_equation", + "content": "SSL + SAOD" + }, + { + "bbox": [ + 46, + 568, + 287, + 700 + ], + "type": "text", + "content": " setup: We propose a semi-supervised learning benchmark for SAOD. This benchmark entails training models on a sparsely annotated labeled and an unlabeled set. As shown in Figure 4, this setup introduces two kinds of sparsity in the data, namely, sparsity in labeled regions (left) and images (right). We believe this is a realistic setup and SAOD methods must be capable of tackling both these kinds of sparsity in the data. For this setup, we use Split-5 (Section 4.2) with increasing sparsity as the labeled set and VOC12 trainval as the unlabeled set. We use the COCO-style AP metric to report results on this setup." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "In Table 3, we compare against Co-mining [10] and ob" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 87, + 545, + 200 + ], + "blocks": [ + { + "bbox": [ + 329, + 74, + 520, + 85 + ], + "lines": [ + { + "bbox": [ + 329, + 74, + 520, + 85 + ], + "spans": [ + { + "bbox": [ + 329, + 74, + 520, + 85 + ], + "type": "text", + "content": "Sparsity in labeled regions Sparsity in labeled images" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 87, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 305, + 87, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 87, + 545, + 200 + ], + "type": "image", + "image_path": "6400e3da83d042143b2db15d36521d5f28c363d9f99f4a510a52b093a23f2e6d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 208, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 304, + 208, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 547, + 257 + ], + "type": "text", + "content": "Figure 4: Types of sparsity in labeled data; sparsity in labeled regions (left) and images (right). Our proposed SSL+SAOD is a realistic setup that presents challenges from both kinds of sparsity." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "text", + "content": "serve an improvement of 1.81, 1.84 and " + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "inline_equation", + "content": "3.55\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 266, + 545, + 373 + ], + "type": "text", + "content": " sparsity levels respectively. We observe that the gap in performance increases with sparsity, consistent with our observation for Tables 1 and 2. This can be attributed to the inability of methods like Co-mining to handle unlabeled images and high sparsity in the labeled data. We will make this benchmark public and propose this method as a baseline. We encourage researchers to report results on this benchmark in the future." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 382, + 432, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 432, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 432, + 396 + ], + "type": "text", + "content": "4.6. Ablation Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 402, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 460 + ], + "type": "text", + "content": "In this section we conduct ablation experiments to understand the various components. For the ablation experiments, we use a ResNet-101 as the backbone network and train on Split-5 with " + }, + { + "bbox": [ + 304, + 402, + 545, + 460 + ], + "type": "inline_equation", + "content": "p = 50\\%" + }, + { + "bbox": [ + 304, + 402, + 545, + 460 + ], + "type": "text", + "content": ". We evaluate on the VOC07 test set and report the COCO style AP metric." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "text", + "content": "From Table 4, a baseline model trained on the ablation set attains 41.77 (row 1). Pseudo-labeling " + }, + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\left(\\mathcal{L}_{\\mathrm{det},a}^{s}\\right)" + }, + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "text", + "content": " improves the performance by 0.9 (row 2). Co-mining [10] relies extensively on pseudo-labels. This results in a drop in performance at higher sparsities due to noisy pseudo-labels. Addition of the C-RPN improves our performance by 3.53 points (row 3). C-RPN reduces the overhead of computing two sets of proposals and learns a better notion of objects due to the combined processing of features from the two views." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": "The combination of C-RPN and PPM improves the performance by 3.67 (row 4). We do not observe major improvements with the introduction of PPM because its task is to identify and segregate the unlabeled regions from the backgrounds. After the segregation, PPM does no further processing to improve performance. The power of PPM can be observed when trained in conjunction with consistency regularization loss " + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{SSL}}^{\\mathrm{u}})" + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": " which achieves the best performance of 46 (row 7); an improvement of 4.23 points over the baseline. We distinguish ourselves from pseudolabeling approaches like Co-mining in one important as" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6776" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 100, + 265, + 153 + ], + "blocks": [ + { + "bbox": [ + 66, + 79, + 264, + 91 + ], + "lines": [ + { + "bbox": [ + 66, + 79, + 264, + 91 + ], + "spans": [ + { + "bbox": [ + 66, + 79, + 264, + 91 + ], + "type": "text", + "content": "Table 5: Analysis of the threshold used for PPM." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 100, + 265, + 153 + ], + "lines": [ + { + "bbox": [ + 66, + 100, + 265, + 153 + ], + "spans": [ + { + "bbox": [ + 66, + 100, + 265, + 153 + ], + "type": "table", + "html": "
Threshold0.60.70.80.90.95
AP45.9645.9646.0045.7745.30
", + "image_path": "8e95bf0ee2bedb9f62a17d2367d73ca50a79ac74834c5b64be0b5efc81bc7e10.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 285, + 92, + 545, + 162 + ], + "blocks": [ + { + "bbox": [ + 325, + 71, + 504, + 83 + ], + "lines": [ + { + "bbox": [ + 325, + 71, + 504, + 83 + ], + "spans": [ + { + "bbox": [ + 325, + 71, + 504, + 83 + ], + "type": "text", + "content": "Table 6: Removing test time augmentations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 285, + 92, + 545, + 162 + ], + "lines": [ + { + "bbox": [ + 285, + 92, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 285, + 92, + 545, + 162 + ], + "type": "table", + "html": "
MethodSplit-1Split-2Split-3
30%50%70%30%50%70%30%50%70%
TTA38.2235.9232.6839.7636.9435.3339.5637.1535.48
w/o TTA38.3235.9132.6739.7736.9535.2739.5537.1235.48
", + "image_path": "0c137d8c98f036771c718b73ab4c316eb9ee1cc2b7508540d1e3cb4a8a69b984.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 171, + 162, + 240 + ], + "blocks": [ + { + "bbox": [ + 48, + 171, + 162, + 240 + ], + "lines": [ + { + "bbox": [ + 48, + 171, + 162, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 171, + 162, + 240 + ], + "type": "image", + "image_path": "56943a3ba02f47679770eddb80e9fb15362424e65c36d5c6ac1087733f0f1e0e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 166, + 171, + 269, + 240 + ], + "blocks": [ + { + "bbox": [ + 166, + 171, + 269, + 240 + ], + "lines": [ + { + "bbox": [ + 166, + 171, + 269, + 240 + ], + "spans": [ + { + "bbox": [ + 166, + 171, + 269, + 240 + ], + "type": "image", + "image_path": "44e100c8e228b1bd3e9da7960054013d85c3fd74bb048854d8cb4341632af08a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 272, + 171, + 375, + 239 + ], + "blocks": [ + { + "bbox": [ + 272, + 171, + 375, + 239 + ], + "lines": [ + { + "bbox": [ + 272, + 171, + 375, + 239 + ], + "spans": [ + { + "bbox": [ + 272, + 171, + 375, + 239 + ], + "type": "image", + "image_path": "466bca1565ecfaf5c8d0e149a98af8c30b9580b0f2b855287b34def292c489aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 378, + 171, + 439, + 240 + ], + "blocks": [ + { + "bbox": [ + 378, + 171, + 439, + 240 + ], + "lines": [ + { + "bbox": [ + 378, + 171, + 439, + 240 + ], + "spans": [ + { + "bbox": [ + 378, + 171, + 439, + 240 + ], + "type": "image", + "image_path": "4908f32da58fa3d5d9cd90317e1cfdbc5e9aef842392af01a4f3e646107eb5e6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 442, + 171, + 545, + 239 + ], + "blocks": [ + { + "bbox": [ + 442, + 171, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 442, + 171, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 442, + 171, + 545, + 239 + ], + "type": "image", + "image_path": "9b421c95f9ca47144aa642efa3bcf377d5e94516a4155002a83d06f4dbb0a39f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 48, + 243, + 163, + 328 + ], + "blocks": [ + { + "bbox": [ + 48, + 243, + 163, + 328 + ], + "lines": [ + { + "bbox": [ + 48, + 243, + 163, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 163, + 328 + ], + "type": "image", + "image_path": "a8e414c0a1ee5cce2a71c8b983479d7b13bbb77761ec61b795c94412bb934734.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 166, + 243, + 250, + 327 + ], + "blocks": [ + { + "bbox": [ + 166, + 243, + 250, + 327 + ], + "lines": [ + { + "bbox": [ + 166, + 243, + 250, + 327 + ], + "spans": [ + { + "bbox": [ + 166, + 243, + 250, + 327 + ], + "type": "image", + "image_path": "dd5c22dd6d8b6849cc3bfc897911f90b5b5f9ac871e0e29d17e74b313f3df2dd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 253, + 243, + 375, + 328 + ], + "blocks": [ + { + "bbox": [ + 253, + 243, + 375, + 328 + ], + "lines": [ + { + "bbox": [ + 253, + 243, + 375, + 328 + ], + "spans": [ + { + "bbox": [ + 253, + 243, + 375, + 328 + ], + "type": "image", + "image_path": "ffb07e03968fe33caef460caaa04fd813f19297ac479345927f7743c05ff07d1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 378, + 243, + 439, + 328 + ], + "blocks": [ + { + "bbox": [ + 378, + 243, + 439, + 328 + ], + "lines": [ + { + "bbox": [ + 378, + 243, + 439, + 328 + ], + "spans": [ + { + "bbox": [ + 378, + 243, + 439, + 328 + ], + "type": "image", + "image_path": "1207c722a40a13ae5e31b049fb3f2112de99f7a0fa8ee0184cb142fd6d2fc711.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 442, + 244, + 545, + 328 + ], + "blocks": [ + { + "bbox": [ + 442, + 244, + 545, + 328 + ], + "lines": [ + { + "bbox": [ + 442, + 244, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 442, + 244, + 545, + 328 + ], + "type": "image", + "image_path": "5d3cfe20723706485bfb81d4cf1511fe13138de8a36f184af30c21875ac487e2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 49, + 361, + 155, + 434 + ], + "blocks": [ + { + "bbox": [ + 46, + 337, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 337, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 545, + 361 + ], + "type": "text", + "content": "Figure 5: Qualitative results showing the unlabeled regions identified by the PPM. The red boxes correspond to the available ground truth. A class agnostic NMS was performed on the regions and the result is shown in white." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 361, + 155, + 434 + ], + "lines": [ + { + "bbox": [ + 49, + 361, + 155, + 434 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 155, + 434 + ], + "type": "image", + "image_path": "9cde6433b013e88104b5cd54b69ca6d65c76cc59cbf56030a841b18b071a8736.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 156, + 361, + 228, + 434 + ], + "blocks": [ + { + "bbox": [ + 156, + 361, + 228, + 434 + ], + "lines": [ + { + "bbox": [ + 156, + 361, + 228, + 434 + ], + "spans": [ + { + "bbox": [ + 156, + 361, + 228, + 434 + ], + "type": "image", + "image_path": "0be5711e10edba967aaacfb2d3a05878dc5b057fa58923d6d89e05723d2a1a7c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 230, + 361, + 284, + 434 + ], + "blocks": [ + { + "bbox": [ + 230, + 361, + 284, + 434 + ], + "lines": [ + { + "bbox": [ + 230, + 361, + 284, + 434 + ], + "spans": [ + { + "bbox": [ + 230, + 361, + 284, + 434 + ], + "type": "image", + "image_path": "3f79d2bd3b22a013b842a2c93c7d37aec3e4d53eadc8b0c475106a09f8d6d398.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 285, + 361, + 335, + 434 + ], + "blocks": [ + { + "bbox": [ + 285, + 361, + 335, + 434 + ], + "lines": [ + { + "bbox": [ + 285, + 361, + 335, + 434 + ], + "spans": [ + { + "bbox": [ + 285, + 361, + 335, + 434 + ], + "type": "image", + "image_path": "f944d4b57b2e453d1015727317dcc28da07075c7c8a05b422363f8c86382d61f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 336, + 361, + 443, + 434 + ], + "blocks": [ + { + "bbox": [ + 336, + 361, + 443, + 434 + ], + "lines": [ + { + "bbox": [ + 336, + 361, + 443, + 434 + ], + "spans": [ + { + "bbox": [ + 336, + 361, + 443, + 434 + ], + "type": "image", + "image_path": "ead32e35cc7a802167f31af89bba5dcfcff9310a723a77b0a6c6fe6e70b11915.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 443, + 361, + 545, + 434 + ], + "blocks": [ + { + "bbox": [ + 443, + 361, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 443, + 361, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 443, + 361, + 545, + 434 + ], + "type": "image", + "image_path": "cf397c441ca6cbf2a46b884debf60b9c692384341352795a76e2ee36752ab9f4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 49, + 435, + 155, + 506 + ], + "blocks": [ + { + "bbox": [ + 49, + 435, + 155, + 506 + ], + "lines": [ + { + "bbox": [ + 49, + 435, + 155, + 506 + ], + "spans": [ + { + "bbox": [ + 49, + 435, + 155, + 506 + ], + "type": "image", + "image_path": "12dda31c9b2760096dc90be00c9e4ec904ad9d71780bbf4fba533d6d22d766d3.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 517, + 545, + 564 + ], + "lines": [ + { + "bbox": [ + 46, + 517, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 545, + 564 + ], + "type": "text", + "content": "Figure 6: Qualitative results comparing the output of a model trained using available ground truths (top) to a model trained using our approach (bottom). Predictions with a class confidence score greater than 0.9 are shown. Red: Person, Cyan: Dog, Purple: Horse, Yellow: Clock, Green: Stop sign, Blue: Parking meter, Violet: Giraffe, Orange: Potted plant, Black: Surfboard, Dark green: Boat" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 156, + 434, + 228, + 506 + ], + "blocks": [ + { + "bbox": [ + 156, + 434, + 228, + 506 + ], + "lines": [ + { + "bbox": [ + 156, + 434, + 228, + 506 + ], + "spans": [ + { + "bbox": [ + 156, + 434, + 228, + 506 + ], + "type": "image", + "image_path": "394fb5c3ffa80d65a890f65208469d3c7d0736f1b36c292a00689b8d0a85103a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 230, + 434, + 284, + 506 + ], + "blocks": [ + { + "bbox": [ + 230, + 434, + 284, + 506 + ], + "lines": [ + { + "bbox": [ + 230, + 434, + 284, + 506 + ], + "spans": [ + { + "bbox": [ + 230, + 434, + 284, + 506 + ], + "type": "image", + "image_path": "455ef43f4639b77690f9cecbddaa6d68070c3d0c55efdf9597f600687341aa59.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 285, + 434, + 335, + 506 + ], + "blocks": [ + { + "bbox": [ + 285, + 434, + 335, + 506 + ], + "lines": [ + { + "bbox": [ + 285, + 434, + 335, + 506 + ], + "spans": [ + { + "bbox": [ + 285, + 434, + 335, + 506 + ], + "type": "image", + "image_path": "c6f45432bfdcd2ee1c69c207e0c45c04b914db15b589ae5c85bd28e2f5f17c2d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 336, + 434, + 443, + 506 + ], + "blocks": [ + { + "bbox": [ + 336, + 434, + 443, + 506 + ], + "lines": [ + { + "bbox": [ + 336, + 434, + 443, + 506 + ], + "spans": [ + { + "bbox": [ + 336, + 434, + 443, + 506 + ], + "type": "image", + "image_path": "4239b9e8a9242fa7ea905bf76922bc5ea9853025f0230dba4f3079b0a2b9d34d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 443, + 435, + 545, + 506 + ], + "blocks": [ + { + "bbox": [ + 443, + 435, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 443, + 435, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 443, + 435, + 545, + 506 + ], + "type": "image", + "image_path": "a208970510f4bb326390f91d87805ba0146fd55c5660dc19fd7b10aa0d4a9a54.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 571, + 289, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 289, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 289, + 704 + ], + "type": "text", + "content": "pect. While, Co-mining [10] relies solely on pseudo-labels, we leverage additional components from self-supervised learning along with pseudo-labeling. In row 5, due to C-RPN, PPM and " + }, + { + "bbox": [ + 46, + 571, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SSL}}^{\\mathrm{u}}" + }, + { + "bbox": [ + 46, + 571, + 289, + 704 + ], + "type": "text", + "content": ", we show an improvement of 3.74 over the baseline. Our proposed components are orthogonal to pseudo-labeling as using them together results in an additional improvement of " + }, + { + "bbox": [ + 46, + 571, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\sim 0.5" + }, + { + "bbox": [ + 46, + 571, + 289, + 704 + ], + "type": "text", + "content": " on mAP. At higher sparsity in the labeled dataset, the proposed C-RPN, PPM and " + }, + { + "bbox": [ + 46, + 571, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SSL}}^{\\mathrm{u}}" + }, + { + "bbox": [ + 46, + 571, + 289, + 704 + ], + "type": "text", + "content": " are less affected than pseudo-labels resulting in the large improvements on these splits. Finally, we show the effect of C-RPN by generating a single set of proposals from the" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 571, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 571, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 305, + 571, + 545, + 606 + ], + "type": "text", + "content": "original image and using it for both the branches. We observe a drop in performance (row 6) highlighting the effectiveness of C-RPN." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "PPM mines potential positives which can be mistaken for negatives due to missing annotations. We rely on the objectness score of the RPN to identify these regions. In Table 5, we vary the threshold of PPM. For a low threshold, a few hard negatives might also be identified as pseudo-positive leading to a drop in performance. With a high threshold, a few potential positives might not be mined. We observe that a threshold of 0.8 provides a good trade-off and is there-" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6777" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 103, + 544, + 175 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 7: Comparison with a different backbone. Results are reported on the COCO validation set on Splits 1-3 using AP and on VOC 2007 test set on Splits-4,5 using " + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 103, + 544, + 175 + ], + "lines": [ + { + "bbox": [ + 49, + 103, + 544, + 175 + ], + "spans": [ + { + "bbox": [ + 49, + 103, + 544, + 175 + ], + "type": "table", + "html": "
MethodSplit-1Split-2Split-3Split-4Split-5
30%50%70%30%50%70%30%50%70%EasyHardExtreme30%40%50%
Ours (C4)37.6735.9533.1639.2236.8134.9838.8436.7635.2680.5680.4374.4577.3876.1375.32
Ours (FPN)38.2235.9232.6839.7636.9435.3339.5637.1535.4882.1581.5075.5976.8475.8874.35
", + "image_path": "2f725af2a001edb0198a2edb5f0fed260b2563c928be608d34e5c196576e3f38.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 288, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 288, + 232 + ], + "type": "text", + "content": "fore used for all experiments unless stated otherwise. In all our experiments, PPM is performed after an initial warmup of 9000 and 30000 iterations on the PASCAL-VOC and COCO datasets respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 247, + 222, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 222, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 222, + 260 + ], + "type": "text", + "content": "4.6.1 Inference without augmentations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 267, + 288, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 288, + 387 + ], + "type": "text", + "content": "Our method requires passing an input with augmentations during inference as well. It should be noted that this is not a test-time augmentation (TTA), a technique that typically involves passing images at a higher resolution. We perform inference by removing the augmentation and extracting the region proposals using C-RPN. We show the results in Table 6 on the three splits of COCO. We do not observe a significant improvement in performance due to the augmentation. For the analysis on the effect of augmentations refer to supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 403, + 160, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 160, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 160, + 415 + ], + "type": "text", + "content": "4.6.2 Effect of backbone" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 422, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 509 + ], + "type": "text", + "content": "We analyze the effect of backbone on our approach and show results for the normal convolution based (C4) and FPN based (FPN) backbones in Table 7. The first row corresponds to C4 while the second row corresponds to FPN. While the gap in performance is lower in most cases, we observe a significant improvement using the FPN on the low sparsity settings of all the split." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 515, + 279, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 279, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 279, + 528 + ], + "type": "text", + "content": "4.7. RPN Recall experiments for object discovery" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": "PPM identifies foreground regions mistakenly assigned as background during training to avoid penalizing the network. To study the effect of PPM on novel [48] classes, we train a network using our approach on randomly chosen 6000 images of the COCO dataset, containing annotations for only 20 classes of PASCAL-VOC. We evaluate the recall@0.5 of the RPN over the remaining 60 classes. A model trained using the standard technique on this dataset achieves a recall@0.5 of " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "77.46\\%" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "29.06\\%" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " on the known classes (20 categories) and unknown classes (60 categories) respectively. Our proposed approach, with PPM, achieves a recall@0.5 of " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "78.47\\%" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "inline_equation", + "content": "35.20\\%" + }, + { + "bbox": [ + 46, + 533, + 288, + 715 + ], + "type": "text", + "content": " respectively. This ability to localize objects not seen during training can be beneficial for object discovery methods like [49-52] which use RPN proposals to learn/discover new categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 182, + 418, + 195 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 418, + 195 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 418, + 195 + ], + "type": "text", + "content": "4.8. Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 203, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 203, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 545, + 348 + ], + "type": "text", + "content": "In Figure 5, we show the pseudo positives mined by PPM. In each figure, the red boxes correspond to the ground truth annotations and the white boxes correspond to the post NMS pseudo positive boxes mined by PPM. We observe that the PPM correctly mines proposals which correspond to missing object annotations. Without PPM, these regions will be used as negatives to the classifier resulting in a reduced class confidence score leading to a drop in mAP. We show the detection results of a model trained on the " + }, + { + "bbox": [ + 304, + 203, + 545, + 348 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 203, + 545, + 348 + ], + "type": "text", + "content": " Split-1 of our approach in Figure 6. The images on the top corresponds to the model trained using sparse annotations and the bottom image shows the output of our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 364, + 378, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 378, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 378, + 376 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 386, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 589 + ], + "type": "text", + "content": "We present SparseDet, a novel end-to-end system for Sparsely Annotated Object Detection (SAOD). We propose a simple yet effective technique for identifying the unlabeled regions using pseudo-positive mining and apply self-supervised loss on them. Through qualitative results we highlight the ability of PPM to mine pseudo-positives. We standardize the evaluation setup and show the effectiveness of our approach with an exhaustive set of experiments on multiple splits of SAOD. While we outperform existing state-of-the-art on all metrics and splits, we observe the gap in performance increases with sparsity demonstrating the short coming of methods that rely solely on pseudolabeling. We propose a new benchmark, that evaluates the semi-supervised learning capabilities of SAOD approaches. We will release the data for the new benchmark along with all the SAOD splits and encourage researchers to evaluate future SAOD methods on these." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgements. This project was partially funded by DARPA SAIL-ON (W911NF2020009), DARPA SemaFor (HR001119S0085), and IARPA SMART (2021-21040700003) programs. Rama Chellappa was supported by an ONR MURI Grant N00014-20-1-2787. We would like to thank our colleagues from the PI lab for their feedback on this work. The views and conclusions contained herein are those of the authors and do not represent views, policies, and/or endorsements of funding agencies." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6778" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Simon Chadwick and Paul Newman. Training object detectors with noisy data. 2019 IEEE Intelligent Vehicles Symposium (IV), pages 1319-1325, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 130, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 130, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 53, + 130, + 288, + 163 + ], + "type": "text", + "content": "[2] Zhe Wu, Navaneeth Bodla, Bharat Singh, Mahyar Najibi, Rama Chellappa, and Larry S. Davis. Soft sampling for robust object detection. In BMVC, 2019. 1, 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "text", + "content": "[3] Junnan Li, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Towards noise-resistant object detection with noisy annotations. ArXiv, abs/2003.01285, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 209, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 287, + 243 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 249, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 287, + 293 + ], + "type": "text", + "content": "[5] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 300, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 287, + 365 + ], + "type": "text", + "content": "[6] Alina Kuznetsova, Hassan Rom, Neil Gordon Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4. International Journal of Computer Vision, 128:1956-1981, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 372, + 287, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 372, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 53, + 372, + 287, + 438 + ], + "type": "text", + "content": "[7] Yusuke Niitani, Takuya Akiba, Tommi Kerola, Toru Ogawa, Shotaro Sano, and Shuji Suzuki. Sampling techniques for large-scale object detection from sparsely annotated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6510-6518, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 445, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 287, + 510 + ], + "type": "text", + "content": "[8] Han Zhang, Fangyi Chen, Zhiqiang Shen, Qiqi Hao, Chenchen Zhu, and Marios Savvides. Solving missing-annotation object detection with background recalibration loss. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1888-1892. IEEE, 2020. 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 517, + 287, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 517, + 287, + 549 + ], + "spans": [ + { + "bbox": [ + 53, + 517, + 287, + 549 + ], + "type": "text", + "content": "[9] Yuewei Yang, Kevin J Liang, and Lawrence Carin. Object detection as a positive-unlabeled problem. In BMVC, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 590 + ], + "type": "text", + "content": "[10] Tiancai Wang, Tong Yang, Jiale Cao, and X. Zhang. Comining: Self-supervised learning for sparsely annotated object detection. In AAAI, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 596, + 288, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 596, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 48, + 596, + 288, + 651 + ], + "type": "text", + "content": "[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2007 (VOC2007) Results. http://www.pascalnetwork.org/challenges/VOC/voc2007/workshop/index.html, 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 289, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 289, + 712 + ], + "type": "text", + "content": "[12] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html, 2, 5, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[13] Ke Yan, Jinzheng Cai, Adam P. Harrison, Dakai Jin, Jing Xiao, and Le Lu. Universal lesion detection by learning from multiple heterogeneously labeled datasets. ArXiv, abs/2005.13753, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 125, + 545, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 125, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 307, + 125, + 545, + 181 + ], + "type": "text", + "content": "[14] Hansheng Li, Xin Han, Yuxin Kang, Xiaoshuang Shi, Mengdi Yan, Zixu Tong, Qirong Bu, Lei Cui, Jun Feng, and Lin Yang. A novel loss calibration strategy for object detection networks training on sparsely annotated pathological datasets. In MICCAI, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 190, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 190, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 190, + 545, + 245 + ], + "type": "text", + "content": "[15] Yongqiang Zhang, Mingli Ding, Yancheng Bai, Mengmeng Xu, and Bernard Ghanem. Beyond weakly supervised: Pseudo ground truths mining for missing bounding-boxes object detection. IEEE Transactions on Circuits and Systems for Video Technology, 30:983-997, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 254, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 319 + ], + "type": "text", + "content": "[16] Ke Yan, Jinzheng Cai, Youjing Zheng, Adam P. Harrison, Dakai Jin, You-Bao Tang, Yuxing Tang, Lingyun Huang, Jing Xiao, and Le Lu. Learning from multiple datasets with heterogeneous and partial labels for universal lesion detection in ct. IEEE Transactions on Medical Imaging, 40:2759-2770, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 329, + 545, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 329, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 307, + 329, + 545, + 372 + ], + "type": "text", + "content": "[17] Jisoo Jeong, Seungeui Lee, Jeesoo Kim, and Nojun Kwak. Consistency-based semi-supervised learning for object detection. Advances in neural information processing systems, 32:10759-10768, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 381, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 437 + ], + "type": "text", + "content": "[18] Peng Tang, Chetan Ramaiah, Yan Wang, Ran Xu, and Caiming Xiong. Proposal learning for semi-supervised object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2291-2301, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 445, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 489 + ], + "type": "text", + "content": "[19] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 498, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 544 + ], + "type": "text", + "content": "[20] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. arXiv preprint arXiv:2102.09480, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 552, + 545, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 552, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 307, + 552, + 545, + 596 + ], + "type": "text", + "content": "[21] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. arXiv preprint arXiv:2106.09018, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 605, + 545, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 605, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 307, + 605, + 545, + 660 + ], + "type": "text", + "content": "[22] Qize Yang, Xihan Wei, Biao Wang, Xian-Sheng Hua, and Lei Zhang. Interactive self-training with mean teachers for semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5941-5950, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[23] Jihun Yoon, Seungbum Hong, and Min-Kook Choi. Semi-supervised object detection with sparsely annotated dataset. In 2021 IEEE International Conference on Image Processing (ICIP), pages 719-723. IEEE, 2021. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6779" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097-1105. Curran Associates, Inc., 2012. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 146, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 146, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 146, + 287, + 213 + ], + "type": "text", + "content": "[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. doi: 10.1007/s11263-015-0816-y. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 219, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 285 + ], + "type": "text", + "content": "[26] Ross Girshick. Fast r-cnn. In Proceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), ICCV '15, pages 1440-1448, Washington, DC, USA, 2015. IEEE Computer Society. ISBN 978-1-4673-8391-2. doi: 10.1109/ICCV.2015.169. URL http://dx.doi.org/10.1109/ICCV.2015.169.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 292, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 292, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 292, + 287, + 347 + ], + "type": "text", + "content": "[27] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 354, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 399 + ], + "type": "text", + "content": "[28] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 405, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 472 + ], + "type": "text", + "content": "[29] Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. R-fcn: Object detection via region-based fully convolutional networks. In Daniel D. Lee, Masashi Sugiyama, Ulrike von Luxburg, Isabelle Guyon, and Roman Garnett, editors, NIPS, pages 379-387, 2016. URL http://dblp.uni-trier.de/db/conf/nips/nips2016.html#DaiLHS16.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Piotr Dólár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 936-944, 2017. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 541, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 541, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 541, + 287, + 563 + ], + "type": "text", + "content": "[31] Bharat Singh and Larry S Davis. An analysis of scale invariance in object detection-snip. CVPR, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 570, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 592 + ], + "type": "text", + "content": "[32] Bharat Singh, Mahyar Najibi, and Larry S Davis. SNIPER: Efficient multi-scale training. NeurIPS, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 599, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 287, + 620 + ], + "type": "text", + "content": "[33] Mahyar Najibi, Bharat Singh, and Larry S Davis. AutoFocus: Efficient multi-scale inference. ICCV, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 628, + 287, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 628, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 48, + 628, + 287, + 672 + ], + "type": "text", + "content": "[34] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. doi: 10.1109/ICCV.2017.322.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[35] Alexander Kirillov, Ross B. Girshick, Kaiming He, and Piotr Dóllár. Panoptic feature pyramid networks. In CVPR, pages 6399-6408. Computer Vision Foundation / IEEE, 2019. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[36] Alexander Kirillov, Kaiming He, Ross B. Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In CVPR, pages 9404-9413. Computer Vision Foundation / IEEE, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 126, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 126, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 126, + 545, + 148 + ], + "type": "text", + "content": "[37] Justin Johnson Georgia Gkioxari, Jitendra Malik. Mesh r-cnn. In ICCV, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 159, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 545, + 225 + ], + "type": "text", + "content": "[38] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 235, + 545, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 235, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 307, + 235, + 545, + 280 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2999-3007, 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 289, + 545, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 333 + ], + "type": "text", + "content": "[40] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection, 2015. URL http://arxiv.org/abs/1506.02640. cite arxiv:1506.02640.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 343, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 388 + ], + "type": "text", + "content": "[41] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 398, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 452 + ], + "type": "text", + "content": "[42] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 496 + ], + "type": "text", + "content": "[43] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:386-397, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 506, + 545, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 506, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 307, + 506, + 545, + 538 + ], + "type": "text", + "content": "[44] Ross B. Girshick. Fast R-CNN. 2015 IEEE International Conference on Computer Vision (ICCV), pages 1440-1448, 2015. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 549, + 545, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 549, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 307, + 549, + 545, + 604 + ], + "type": "text", + "content": "[45] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. doi: 10.1109/CVPR.2016.90.6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 614, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 647 + ], + "type": "text", + "content": "[46] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[47] Hanjun Li, Xingjia Pan, Ke Yan, Fan Tang, and Wei-Shi Zheng. Siod: Single instance annotated per category per image for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14197-14206, 2022. 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6780" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 378 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[48] Akshay Raj Dhamija, Manuel Günther, Jonathan Ventura, and Terrance E. Boult. The overlooked elephant of object detection: Open set. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1010-1019, 2020. doi: 10.1109/WACV45572.2020.9093355. 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 135, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 135, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 135, + 287, + 179 + ], + "type": "text", + "content": "[49] Huy V. Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 253 + ], + "type": "text", + "content": "[50] Huy V. Vo, Francis Bach, Minsu Cho, Kai Han, Yann LeCun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 8287-8296, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "text", + "content": "[51] Sai Saketh Rambhatla, Rama Chellappa, and Abhinav Shrivastava. The pursuit of knowledge: Discovering and localizing novel categories using dual memory. In Proceedings of the International Conference on Computer Vision (ICCV), 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 323, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 323, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 48, + 323, + 287, + 378 + ], + "type": "text", + "content": "[52] Sai Saketh Rambhatla, Ishan Misra, Rama Chellappa, and Abhinav Shrivastava. The pursuit of knowledge: Discovering and localizing novel categories using dual memory. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 9" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "text", + "content": "6781" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/280aa12d-c539-41fe-b78e-1bcf7c1aa152_content_list.json b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/280aa12d-c539-41fe-b78e-1bcf7c1aa152_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..510cde9ec7919fca288ffb664320283dae13f9ff --- /dev/null +++ b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/280aa12d-c539-41fe-b78e-1bcf7c1aa152_content_list.json @@ -0,0 +1,1539 @@ +[ + { + "type": "text", + "text": "SparseFusion: Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection", + "text_level": 1, + "bbox": [ + 184, + 161, + 784, + 201 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yichen Xie $^{1,*}$ , Chenfeng Xu $^{1,*}$ , Marie-Julie Rakotosaona $^{2}$ , Patrick Rim $^{3}$ , Federico Tombari $^{2}$ , Kurt Keutzer $^{1}$ , Masayoshi Tomizuka $^{1}$ , Wei Zhan $^{1}$", + "bbox": [ + 119, + 226, + 850, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of California, Berkeley 2 Google 3 California Institute of Technology", + "bbox": [ + 161, + 260, + 805, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 312, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By identifying four important components of existing LiDAR-camera 3D object detection methods (LiDAR and camera candidates, transformation, and fusion outputs), we observe that all existing methods either find dense candidates or yield dense representations of scenes. However, given that objects occupy only a small part of a scene, finding dense candidates and generating dense representations is noisy and inefficient. We propose SparseFusion, a novel multi-sensor 3D detection method that exclusively uses sparse candidates and sparse representations. Specifically, SparseFusion utilizes the outputs of parallel detectors in the LiDAR and camera modalities as sparse candidates for fusion. We transform the camera candidates into the LiDAR coordinate space by disentangling the object representations. Then, we can fuse the multi-modality candidates in a unified 3D space by a lightweight self-attention module. To mitigate negative transfer between modalities, we propose novel semantic and geometric cross-modality transfer modules that are applied prior to the modality-specific detectors. SparseFusion achieves state-of-the-art performance on the nuScenes benchmark while also running at the fastest speed, even outperforming methods with stronger backbones. We perform extensive experiments to demonstrate the effectiveness and efficiency of our modules and overall method pipeline. Our code will be made publicly available at https://github.com/yichen928/SparseFusion.", + "bbox": [ + 75, + 352, + 473, + 711 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 737, + 208, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving cars rely on multiple sensors, such as LiDAR and cameras, to perceive the surrounding environment. LiDAR sensors provide accurate 3D scene occupancy information through point clouds with points in the xyz coordinate space, and cameras provide rich semantic information through images with pixels in the RGB", + "bbox": [ + 75, + 761, + 468, + 845 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/031d89a9df86fe6b84eb009a319a73d1bff0e654b664a6711427f4fe7fc7f7e1.jpg", + "image_caption": [ + "(a) mAP vs. FPS" + ], + "image_footnote": [], + "bbox": [ + 502, + 308, + 690, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d77b38623dc1b1833de9b991a0d9aaa99ceda8b5b19ab1f59546eb55980df172.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 317, + 890, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1690ec947523fe1b553f01f68c5e7b7c1e3c40c807796ab35ac643998afee96f.jpg", + "image_caption": [ + "(b) Dense-to-dense fusion." + ], + "image_footnote": [ + "(c) Overview of our sparse fusion strategy. We extract instance-level features from the LiDAR and camera modalities separately, and fuse them in a unified 3D space to perform detection.", + "Figure 1: Compared to existing fusion algorithms, Sparse-Fusion achieves state-of-the-art performance as well as the fastest inference speed on nuScenes test set. † : Official code of [5] uses flip as test-time augmentation. ‡ : We use BEVFusion-base results in the official repository of [28] to match the input resolutions of other methods. § : Swin-T [27, 21] is adopted as image backbone." + ], + "bbox": [ + 500, + 431, + 890, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "color space. However, there are often significant discrepancies between representations of the same physical scene acquired by the two sensors, as LiDAR sensors capture point clouds using 360-degree rotation while cameras capture images from a perspective view without a sense of depth. This impedes an effective and efficient fusion of the LiDAR and camera modalities. To tackle this challenge, multi-sensor fusion algorithms were proposed to find correspondences between multi-modality data to transform and fuse them into a unified scene representation space.", + "bbox": [ + 497, + 699, + 892, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dense representations, such as bird-eye-view (BEV), volumetric, and point representations, are commonly used", + "bbox": [ + 499, + 838, + 890, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 85 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 253, + 48, + 789, + 61 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 332, + 62, + 709, + 74 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 304, + 74, + 739, + 87 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* indicates equal contribution.", + "bbox": [ + 99, + 853, + 263, + 865 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17591", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/4b4da354a43dce7df6b1897f6d84c2d2e4ad37fcfd9c6dcd2ca296a40cc622b5.jpg", + "table_caption": [ + "Table 1: For each LiDAR-camera fusion method, we identify the LiDAR candidates and camera candidates that are used, the transformation process used to fuse these candidates into a unified space, and the fusion outputs generated to represent 3D scenes using information from both modalities. Based on these components, we categorize the methods into the following categories: Dense+Sparse $\\rightarrow$ Dense approaches relate the sparse region proposals in images to a dense frustum point cloud and fuse them into a dense point space. Dense+Dense $\\rightarrow$ Dense approaches align each point feature in the point cloud to the corresponding pixel feature in the image and represent the 3D scenes using dense features such as point/BEV features. Sparse+Dense $\\rightarrow$ Sparse approaches generate sparse queries by detecting instance features in the point cloud and then applying cross-attention with dense image features. Dense+Dense $\\rightarrow$ Sparse approaches predict objects using queries that combine dense features from each modality. Sparse+Sparse $\\rightarrow$ Sparse (ours) extracts sparse instance features from each modality and directly fuses them to obtain the final sparse instance features used for detection." + ], + "table_footnote": [], + "table_body": "
CategoryMethodLiDAR CandidateCamera CandidateTransformationFusion Outputs
Dense+Sparse→DenseFrustum PointNets [35]point featuresregion proposalsproj. & concat.point features
Dense+Dense→DensePointPainting [39]point featuressegm. outputproj. & concat.point features
PointAugmenting [40]point featuresimage featuresproj. & concat.point features
PAI3D [25]point featuressegm. outputproj. & concat.point features
BEVFusion [28]BEV featuresimage featuresdepth. est. & proj. & concat.BEV features
AutoAlignV2 [5]voxel featuresimage featuresproj. & attn.voxel features
Sparse+Dense→SparseTransFusion [1]instance featuresimage featuresproj. & attn.instance features
Dense+Dense→SparseFUTR3D [4]voxel featuresimage featuresattn.instance features
UVTR [18]voxel featuresimage featuresdepth. est. & proj. & attn.instance features
DeepInteraction [47]BEV featuresimage featuresproj. & attn.instance features
CMT [45]BEV featuresimage featuresattn.instance features
Sparse+Sparse→SparseSparseFusion (Ours)instance featuresinstance featuresproj. & attn.instance features
", + "bbox": [ + 100, + 271, + 870, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to represent 3D scenes [28, 5, 39, 25, 40, 15]. Most previous works fuse different modalities by aligning low-level data or high-level features to yield dense features that describe the entire 3D scene, e.g., as shown in Fig 1b. However, for the task of 3D object detection, such dense representations are superfluous since we are only interested in instances/objects, which only occupy a small part of the 3D space. Furthermore, noisy backgrounds can be detrimental to object detection performance, and aligning different modalities into the same space is a time-consuming process. For example, generating BEV features from multi-view images takes $500\\mathrm{ms}$ on an RTX 3090 GPU [28].", + "bbox": [ + 75, + 465, + 468, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, sparse representations are more efficient, and methods based on them have achieved state-of-the-art performance in multi-sensor 3D detection [1, 4, 18, 47]. These methods use object queries to represent instances/objects in the scene and interact with the original image and point cloud features. However, most previous works do not take into account the significant domain gap between features from different modalities [44]. The queries may gather information from one modality that has a large distribution shift with respect to another modality, making iterative interaction between modalities with large gaps suboptimal. Recent work [47] mitigates this issue by incorporating modality interaction, i.e. performing cross-attention between features from two different modalities. However, the number of computations performed in this method increases quadratically with the dimensions of features and is thus inefficient. We categorize previous works into four", + "bbox": [ + 75, + 631, + 470, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "groups by identifying four key components, which are outlined in Table 1. Further discussion of the methods in these groups is presented in Sec. 2.", + "bbox": [ + 497, + 465, + 890, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose SparseFusion, a novel method (Fig. 1c) that simultaneously utilizes sparse candidates and yields sparse representations, enabling efficient and effective 3D object detection. SparseFusion is the first LiDAR-camera fusion method, to our knowledge, to perform 3D detection using exclusively sparse candidates and sparse fusion outputs. We highlight a key common ground between the two modalities: an image and a point cloud that represent the same 3D scene will contain mostly the same instances/objects. To leverage this commonality, we perform 3D object detection on the inputs from each modality in two parallel branches. Then, the instance features from each branch are projected into a unified 3D space. Since the instance-level features are sparse representations of the same objects in the same scene, we are able to fuse them with a lightweight attention module [38] in a soft manner. This parallel detection strategy allows the LiDAR and camera branches to take advantage of the unique strengths of the point cloud and image representations, respectively. Nevertheless, the drawbacks of each single-modality detector may result in negative transfer during the fusion phase. For example, the point cloud detector may struggle to distinguish between a standing person and a tree trunk due to a lack of detailed semantic information, while the image detector is hard to localize objects in the 3D space due to a lack of accurate depth information. To mitigate the issue of negative", + "bbox": [ + 497, + 507, + 892, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17592", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "transfer, we introduce a novel cross-modality information transfer method designed to compensate for the deficiencies of each modality. This method is applied to the inputs from both modalities prior to the parallel detection branches.", + "bbox": [ + 75, + 125, + 468, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SparseFusion achieves state-of-the-art results on the competitive nuScenes benchmark [2]. Our instance-level sparse fusion strategy allows for a lighter network and much higher efficiency in comparison with prior work. With the same backbone, SparseFusion outperforms the current state-of-the-art model [47] with 1.8x acceleration. Our contributions are summarized as follows:", + "bbox": [ + 75, + 181, + 470, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We revisit prior LiDAR-camera fusion works and identify four important components that allow us to categorize existing methods into four groups. We propose an entirely new category of methods that exclusively uses sparse candidates and representations.", + "bbox": [ + 95, + 281, + 468, + 350 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose SparseFusion, a novel method for LiDAR-camera 3D object detection that leverages instance-level sparse feature fusion and cross-modality information transfer to take advantage of the strengths of each modality while mitigating their weaknesses.", + "bbox": [ + 95, + 355, + 468, + 423 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We demonstrate that our method achieves state-of-the-art performance in 3D object detection with a lightweight architecture that provides the fastest inference speed.", + "bbox": [ + 95, + 428, + 468, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 77, + 496, + 218, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR-based 3D Object Detection. LiDAR sensors are commonly used for single-modality 3D object detection due to the accurate geometric information provided by point clouds. For detection in outdoor scenes, most existing methods transform unordered point clouds into more structured data formats such as pillars [17], voxels [46, 8], or range views [9, 23]. Features are extracted by standard 2D or 3D convolutional networks, based on which a detection head is used to recognize objects and regress 3D bounding boxes. Mainstream detection heads apply anchor-based [51, 17] or center-based [50] structures. Inspired by the promising performance of transformer-based methods in 2D detection, some recent works explore transformers as feature extractors [30, 36] or as detection heads [1, 31]. Our method is agnostic to the LiDAR-based detector used in the LiDAR branch, and the default setting uses TransFusion-L [1].", + "bbox": [ + 75, + 520, + 468, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera-based 3D Object Detection. Camera-based 3D detection methods are also being studied increasingly. Early work performs monocular 3D object detection by attaching extra 3D bounding box regression heads [48, 42] to 2D detectors. In practice, scenes are often perceived by multiple cameras from different perspective views. Following LSS [34], methods like BEVDet [14] and BEVDepth [19] extract 2D features from multi-view images and project them into the BEV space. Other methods including", + "bbox": [ + 75, + 741, + 468, + 865 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DETR3D [43] and PETR [26] adapt techniques from transformer-based 2D object detection methods [53, 3] to learn correspondences between different perspective views through cross-attention using 3D queries. However, as revealed in [12], there inevitably exists some ambiguity when recovering 3D geometry from 2D images. In response, recent works [13, 20, 32] also explore the positive effects of temporal cues in camera-based 3D detection. In our proposed SparseFusion, we extend deformable-DETR [53] to monocular 3D object detection and explicitly transform the regressed bounding boxes to the LiDAR coordinate space.", + "bbox": [ + 497, + 125, + 890, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-Modality 3D Object Detection. LiDAR and cameras provide complementary information about the surrounding environment, so it is appealing to fuse the multimodality inputs for 3D object detection tasks. As analyzed in Tab. 1, existing fusion methods can be classified into four categories. Early works tend to fuse multi-modality information into a unified dense representation. Frustum Point-Nets [35] utilizes a Dense+Sparse $\\rightarrow$ Dense approach that filters dense point clouds with sparse 2D regions of interest. Subsequent works explore Dense+Dense $\\rightarrow$ Dense approaches by working directly with the dense LiDAR modality and camera modality features instead. Methods such as [39, 40] project point clouds into image perspective views and concatenate the dense image features with point features. BEVFusion [28, 22] significantly improves the performance of this line of methods by projecting dense image features into the LiDAR coordinate space using estimated per-pixel depths. AutoAlignV2 [5] also considers the soft correspondence through cross-modality attention to increase the robustness. However, we point out that dense representations are altogether undesirable for 3D object detection as they are noisy and inefficient.", + "bbox": [ + 497, + 279, + 890, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works have begun to utilize object-centric sparse scene representations. TransFusion [1] adopts a Sparse+Dense $\\rightarrow$ Sparse strategy by extracting sparse instance-level features from the LiDAR modality and refining them using dense image features. Other works [4, 18, 47, 45] utilize a Dense+Dense $\\rightarrow$ Sparse approach where queries are used to extract a sparse instance-level representation from dense BEV and image features. However, it is hard to extract information from multi-modality features with an attention operation given the large cross-modal distribution shift. To this end, UVTR [18] projects image features into the LiDAR coordinate space, CMT [45] encodes modality-specific positional information to its queries, and DeepInteraction [47] proposes cross-modality interaction. However, these methods still need to resolve the large multimodal domain gap by stacking many transformer layers to construct a heavy decoder.", + "bbox": [ + 497, + 587, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to the above methods, our method adopts the previously unexplored Sparse+Sparse $\\rightarrow$ Sparse approach. SparseFusion extracts sparse representations of both modal-", + "bbox": [ + 497, + 825, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17593", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ities and fuses them to generate a more accurate and semantically rich sparse representation that yields great performance while also achieving great efficiency.", + "bbox": [ + 75, + 125, + 468, + 168 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 77, + 182, + 211, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present SparseFusion, an effective and efficient framework for 3D object detection via LiDAR and camera inputs. The overall architecture is illustrated in Fig. 2. We acquire sparse candidates from each modality using modality-specific object detection in the LiDAR and camera branches. The instance-level features generated by the camera branch are transformed into the LiDAR space of the instance-level features generated by the LiDAR branch. They are then fused with a simple self-attention module (Sec. 3.1). To mitigate the negative transfer between modalities, we apply a geometric transfer module and a semantic transfer module prior to the parallel detection branches (Sec. 3.2). Furthermore, we design custom loss functions for each module to ensure stable optimization (Sec. 3.3).", + "bbox": [ + 75, + 205, + 470, + 399 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Sparse Representation Fusion", + "text_level": 1, + "bbox": [ + 77, + 410, + 341, + 423 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Acquiring candidates in two modalities. LiDAR candidates. We follow TransFusion-L [1] and use only one decoder layer for LiDAR modality detection. The LiDAR backbone extracts a BEV feature map from the point cloud inputs. We initialize $N_{L}$ object queries $\\mathbf{Q}_L^0 = \\{\\mathbf{q}_{L,i}^0\\}_{i=1}^{N_L}, \\mathbf{q}_{L,i}^0 \\in \\mathbb{R}^C$ as well as their corresponding reference points $\\mathbf{p}_L^0 = \\{\\mathbf{p}_{L,i}^0\\}_{i=1}^{N_L}, \\mathbf{p}_{L,i}^0 \\in \\mathbb{R}^2$ in the BEV plane. These queries interact with the BEV features through a cross-attention layer to generate the updated object queries $\\mathbf{Q}_L = \\{\\mathbf{q}_{L,i}\\}_{i=1}^{N_L}, \\mathbf{q}_{L,i} \\in \\mathbb{R}^C$ . These updated queries $\\mathbf{Q}_L$ represent the instance-level features of objects in the LiDAR modality, and we use them as the LiDAR candidates in the subsequent multi-modal fusion module. Furthermore, we apply a prediction head to each query to classify the object and regress the bounding box in LiDAR coordinate space.", + "bbox": [ + 75, + 431, + 468, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera candidates. To generate the camera candidates, we utilize a camera-only 3D detector with images from different perspective views as inputs. Specifically, we extend deformable-DETR [53] with 3D box regression heads. We also initialize $N_{C}$ object queries $\\mathbf{Q}_C^0 = \\{\\mathbf{q}_{C,i}^0\\}_{i = 1}^{N_C}$ , $\\mathbf{q}_{C,i}^0 \\in \\mathbb{R}^C$ along with their corresponding reference points $\\mathbf{p}_C^0 = \\{\\mathbf{p}_{C,i}^0\\}_{i = 1}^{N_C}$ , $\\mathbf{p}_{C,i}^0 \\in \\mathbb{R}^2$ on the image. For each perspective view $v$ , queries on its image interact with the corresponding image features using a deformable attention layer [53]. The outputs of all perspective views comprise the updated queries $\\mathbf{Q}_C^P = \\{\\mathbf{q}_{C,i}^P\\}_{i = 1}^{N_C}$ , $\\mathbf{q}_{C,i}^P \\in \\mathbb{R}^C$ . We use these queries as the camera candidates in the subsequent multi-modal fusion module. We provide further details of our architecture, the initialization method, and the prediction heads for the two modalities in our supplementary materials.", + "bbox": [ + 75, + 643, + 468, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformation After acquiring the candidates from each", + "bbox": [ + 77, + 852, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "modality, we aim to transform the candidates from the camera modality to the space of the candidates from the LiDAR modality. Since the candidates from the camera modality are high-dimensional latent features that are distributed differently than the candidates from the LiDAR modality, a naive coordinate transformation between modalities is inapplicable here. To address this issue, we disentangle the representations of the camera candidates. Intrinsically, a camera candidate is an instance feature that is a representation of a specific object's class and 3D bounding box. While an object's class is view-invariant, its 3D bounding box is view-dependent. This motivates us to focus on transforming high-dimensional bounding box representations.", + "bbox": [ + 497, + 125, + 890, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first input the candidate instance features into the prediction head of the camera branch. We label the outputted bounding boxes as $\\mathbf{b}^P$ . Given the extrinsic matrix $[\\mathbf{R}_v|\\mathbf{t}_v]$ and intrinsic matrix $\\mathbf{K}_v$ of the corresponding $v$ -th camera, the bounding boxes can be easily projected into the LiDAR coordinate system. We denote the project bounding boxes as $\\mathbf{b}^L$ . We encode the projected bounding boxes with a multi-layer perceptron (MLP) $g(\\cdot)$ , yielding a high-dimensional box embedding. We also encode the flattened camera parameter with another MLP $m(\\cdot)$ to obtain a camera embedding. The camera embedding is multiplied with the original instance features as done in [19], which are then added to the box embedding, given by", + "bbox": [ + 497, + 306, + 892, + 485 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {C, i} ^ {L} = g \\left(\\mathbf {b} _ {i} ^ {L}\\right) + h \\left(\\mathbf {q} _ {C, i} ^ {P} \\cdot m \\left(\\mathbf {R} _ {v}, \\mathbf {t} _ {v}, \\mathbf {K} _ {v}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 495, + 890, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h(\\cdot)$ is an extra MLP to encode the query features in the perspective view. $h(\\cdot)$ aims to preserve view-agnostic information while discarding view-specific information. Afterward, $\\mathbf{Q}_C^L = \\{\\mathbf{q}_{C,i}^L\\}$ is passed through a self-attention layer to aggregate information from multiple cameras to get the updated queries $\\mathbf{Q}_C$ which represent the image modality instance features in the LiDAR space.", + "bbox": [ + 497, + 523, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sparse candidate fusion. Our parallel modality-specific object detection provides sparse instance candidates $\\mathbf{Q}_L = \\{\\mathbf{q}_{L,i}\\}_{i=1}^{N_L}$ and $\\mathbf{Q}_C = \\{\\mathbf{q}_{C,i}\\}_{i=1}^{N_C}$ from the LiDAR and camera modalities respectively. After the above transformation of the camera candidates into LiDAR space, candidates from both modalities represent bounding boxes in the same LiDAR coordinate space, as well as the view-invariant categories. We now concatenate the candidates together:", + "bbox": [ + 497, + 619, + 890, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {L C} = \\left\\{\\mathbf {q} _ {L C, i} \\right\\} _ {i = 1} ^ {N _ {L} + N _ {C}} = \\left\\{f _ {L} (\\mathbf {q} _ {L, i}) \\right\\} _ {i = 1} ^ {N _ {L}} \\cup \\left\\{f _ {C} (\\mathbf {q} _ {C, i}) \\right\\} _ {i = 1} ^ {N _ {C}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 512, + 742, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{L}(\\cdot), f_{C}(\\cdot)$ are learnable projectors. Afterward, we make novel use of a self-attention module to fuse the two modalities. Despite the simplicity of self-attention, the inherent intuition is novel: the modality-specific detectors encode the advantageous aspects of their respective inputs, and the self-attention module is able to aggregate and preserve the information from both modalities in an efficient", + "bbox": [ + 497, + 769, + 890, + 865 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17594", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b66143a9b0e469d0fda8b48448f1672540f63fae29399532fef1c08e8e6c97ca.jpg", + "image_caption": [ + "Figure 2: Overall framework of SparseFusion. It fuses sparse candidates from LiDAR and camera modalities to obtain a multi-modality instance-level representation in the unified LiDAR space." + ], + "image_footnote": [], + "bbox": [ + 139, + 122, + 831, + 391 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dcecbe289c436eacd957287d6c1024e522935ae42d50c83d1a29e30ee34546bc.jpg", + "image_caption": [ + "Figure 3: Cross-modality information transfer. We transfer geometric information from LiDAR to camera modality, and semantic information from camera to LiDAR modality." + ], + "image_footnote": [], + "bbox": [ + 80, + 453, + 272, + 559 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1bf6d711fbe9c21d28e7b64e12748c757bc2f9d78de6bafcffe9d5599f6e1ab7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 275, + 453, + 467, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "manner. The output of the self-attention module is used for final classification and regression of the bounding boxes.", + "bbox": [ + 75, + 638, + 468, + 668 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Cross-Modality Information Transfer", + "text_level": 1, + "bbox": [ + 75, + 676, + 401, + 692 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although we aim to utilize the advantages of both modalities, we must address that the modalities also have their own disadvantages that can result in negative transfer between modalities. For example, the LiDAR detector struggles to capture rich semantic information, while the camera detector struggles to capture accurate geometric and depth information. To mitigate negative transfer, we propose novel geometric and semantic information transfer modules, as illustrated in Fig. 3, that we apply prior to the modality-specific detectors.", + "bbox": [ + 75, + 699, + 468, + 836 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometric transfer from LiDAR to camera. We project each point in the LiDAR point cloud input to multi-view", + "bbox": [ + 75, + 838, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "images to generate sparse multi-view depth maps. These multi-view depth maps are inputted into a shared encoder to obtain depth features, which are then concatenated with the image features to form depth-aware image features that compensate for the lack of geometric information in camera inputs. The depth-aware image features are used as the input to the camera branch.", + "bbox": [ + 497, + 456, + 890, + 552 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic transfer from camera to LiDAR. We project the points in the LiDAR point cloud input to the image inputs, which yields sparse points on the image features. We perform max-pooling to aggregate the resulting multiscale features, and we combine them with the BEV feature through addition. The concatenated features serve as the queries and interact with the multi-scale image features through deformable-attention [53]. The updated queries replace the original queries in the BEV features, which results in the semantic-aware BEV features, which are used for query initialization.", + "bbox": [ + 497, + 552, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Objective Function", + "text_level": 1, + "bbox": [ + 499, + 712, + 682, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply the Gaussian focal loss [24] to the initialized queries of both modalities, given by", + "bbox": [ + 497, + 733, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n i t}} = \\mathcal {L} _ {G F o c a l} (\\hat {\\mathbf {Y}} _ {L}, \\mathbf {Y} _ {L}) + \\mathcal {L} _ {G F o c a l} (\\hat {\\mathbf {Y}} _ {C}, \\mathbf {Y} _ {C}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 770, + 890, + 788 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{Y}}_L, \\hat{\\mathbf{Y}}_C$ are the dense predictions of category-wise heatmaps of the LiDAR and camera modalities, respectively, and $\\mathbf{Y}_L, \\mathbf{Y}_C$ are the corresponding ground-truths.", + "bbox": [ + 497, + 795, + 890, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we apply the loss function for the detectors of the LiDAR and camera modalities, as well as the view transfor", + "bbox": [ + 499, + 838, + 890, + 865 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17595", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mation of the camera candidates and the candidate fusion stage. Firstly, the predictions of each modality-specific detector are independently matched with the ground-truth using the Hungary algorithm [16]. The object classification is optimized with focal loss [24] and the 3D bounding box regression is optimized with L1 loss. For the camera modality detector, the ground-truth bounding boxes are in separate camera coordinates. For all other detectors, ground-truth bounding boxes are in LiDAR coordinates. The detection loss can be represented as", + "bbox": [ + 75, + 125, + 472, + 264 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e t e c t}} = \\gamma \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {c a m e r a}} + \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {t r a n s}} + \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {L i D A R}} + \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {f u s i o n}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 271, + 468, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our entire network is optimized using $\\mathcal{L} = \\alpha \\mathcal{L}_{init} + \\beta \\mathcal{L}_{detect}$ . In our implementation, we empirically set $\\gamma = 1, \\alpha = 0.1$ , and $\\beta = 1$ to balance different terms.", + "bbox": [ + 75, + 296, + 468, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 75, + 351, + 208, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset and metrics", + "text_level": 1, + "bbox": [ + 75, + 372, + 265, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow previous work [47, 45, 22, 5] to evaluate our method on the nuScenes dataset [2]. It is a challenging dataset for 3D object detection, consisting of 700/150/150 scenes for training/validation/test. It provides point clouds collected using a 32-beam LiDAR and six images from multi-view cameras. There are 1.4 million annotated 3D bounding boxes for objects from 10 different classes. We evaluate performance using the nuScenes detection score (NDS) and mean average precision (mAP) metrics. The final mAP is averaged over distance thresholds $0.5\\mathrm{m}$ , $1\\mathrm{m}$ , $2\\mathrm{m}$ , and $4\\mathrm{m}$ on the BEV across 10 classes. NDS is a weighted average of mAP and other true positive metrics including mATE, mASE, mAOE, mAVE, and mAAE.", + "bbox": [ + 75, + 394, + 472, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 75, + 581, + 294, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our implementation is based on the MMDetection3D framework [6]. For the camera branch, we use ResNet-50 [11] as the backbone and initialize it with the Mask R-CNN [10] instance segmentation network pretrained on nuImage [2]. The input image resolution is $800 \\times 448$ . For the LiDAR branch, we apply VoxelNet [51] with voxel size $(0.075m, 0.075m, 0.2m)$ . The detection range is set as $[-54m, 54m]$ for the XY-axes and $[-5m, 3m]$ for the Z-axis. Our LiDAR and camera modality detectors both include only 1 decoder layer. The query numbers for the LiDAR and camera modalities are set as $N_L = N_C = 200$ , so our fusion stage can detect at most 400 objects per scene.", + "bbox": [ + 75, + 603, + 468, + 769 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since our framework disentangles the camera detector and the LiDAR detector, we can conveniently apply data augmentation separately to the LiDAR inputs and camera inputs. We apply random rotation, scaling, translation, and flipping to the LiDAR inputs, and we apply random scaling and horizontal flipping to the camera inputs. Our training pipeline follows previous works [1, 28, 47]. We first train", + "bbox": [ + 75, + 769, + 470, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TransFusion-L [1] as our LiDAR-only baseline, which is used to initialize our LiDAR backbone and LiDAR modality detector. This LiDAR-only baseline is trained for 20 epochs. Afterward, we freeze the pretrained LiDAR components and train the entire fusion framework for 6 epochs. For both training stages, we use the AdamW optimizer [29] with one-cycle learning rate policy [37]. The initial learning rate is $10^{-4}$ and the weight decay is $10^{-2}$ . The hidden dimensions in the entire model except the backbones are 128. For both training stages, we adopt CBGS [52] to balance the class distribution. We train our method on four NVIDIA A6000 GPUs with batch size 16.", + "bbox": [ + 497, + 125, + 894, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results and Comparison", + "text_level": 1, + "bbox": [ + 497, + 302, + 726, + 317 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report results on the nuScenes validation and test sets, without using any test-time augmentations or model ensembles. As shown in Tab. 2, SparseFusion significantly improves our LiDAR-only baseline, TransFusion-L [1], by $+3.6\\%$ NDS and $+6.3\\%$ mAP on the test set, due to the additional use of camera inputs. More importantly, SparseFusion sets a new state of the art on both the validation set and test set, outperforming prior works including those using stronger backbones. It is noteworthy that SparseFusion demonstrates a $0.4\\%$ NDS and $1.0\\%$ mAP improvement over the most recent state-of-the-art [47] while also achieving a 1.8x speedup (5.6 FPS vs. 3.1 FPS) on A6000 GPU. It can be seen in Fig. 1a that, in addition to superior performance, SparseFusion also provides the fastest inference speed. We also demonstrate the performance of SparseFusion by visualizing some qualitative results in Fig. 4.", + "bbox": [ + 497, + 323, + 894, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Analysis", + "text_level": 1, + "bbox": [ + 497, + 554, + 601, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance Breakdown SparseFusion performs 3D object detection in both the LiDAR and camera branch separately. Tab. 3 shows the detection performance in different parts of SparseFusion including the LiDAR branch, camera branch (before and after view transformation), and the fusion branch. We notice that our LiDAR branch detection results notably surpass the LiDAR-only baseline TransFusion-L [1] since the proposed semantic transfer can compensate for the weakness of point cloud inputs. In comparison with the state-of-the-art single-frame camera detector PETR [26] (six decoder layers), our camera branch achieves much better performance with just one decoder layer, owing to the depth-aware features from the proposed geometric transfer. Besides, our view transformation module not only transforms the instance features from the camera coordinate space into the LiDAR coordinate space, but it also slightly improves the camera branch detection performance by aggregating multi-view information. With this strong performance, the modality-specific detectors in each branch would not cause negative transfer during fusion. We also consider direct Non-Maximum Suppression", + "bbox": [ + 497, + 574, + 894, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17596", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/06eb2d0e51129242144b5cfd0fc6db893ffd3c1c596f4f7a29000258535c6555.jpg", + "table_caption": [ + "Table 2: Comparison with existing methods on nuScenes validation set and test set." + ], + "table_footnote": [], + "table_body": "
MethodsModalityLiDAR BackboneCamera Backbonevalidation settest set
NDSmAPNDSmAP
FCOS3D [42]Camera-ResNet-101 [11]41.534.342.835.8
PETR [26]Camera-ResNet-101 [11]44.237.045.539.1
CenterPoint [48]LiDARVoxelNet [51]-66.859.667.360.3
TransFusion-L [1]LiDARVoxelNet [51]-70.165.170.265.5
PointAugmenting [39]LiDAR+CameraVoxelNet [51]DLA34 [49]--71.066.8
FUTR3D [4]LiDAR+CameraVoxelNet [51]ResNet-101 [11]68.364.5--
UVTR [18]LiDAR+CameraVoxelNet [51]ResNet-101 [11]70.265.471.167.1
TransFusion [1]LiDAR+CameraVoxelNet [51]ResNet-50 [11]71.367.571.668.9
AutoAlignV2 [5]LiDAR+CameraVoxelNet [51]CSPNet [41]71.267.172.468.4
BEVFusion [22]LiDAR+CameraVoxelNet [51]Dual-Swin-T [21]72.169.673.371.3
BEVFusion [28]LiDAR+CameraVoxelNet [51]Swin-T [27]71.468.572.970.2
DeepInteraction [47]LiDAR+CameraVoxelNet [51]ResNet-50 [11]72.669.973.470.8
CMT [45]LiDAR+CameraVoxelNet [51]ResNet-50 [11]70.867.9--
SparseFusion (ours)LiDAR+CameraVoxelNet [51]ResNet-50 [11]72.870.473.872.0
", + "bbox": [ + 87, + 148, + 884, + 403 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93a2f08e907b7768b6021c67114553ff0af23b2938c90c333e07bc4aad988d21.jpg", + "image_caption": [ + "Figure 4: Qualitative results of SparseFusion on nuScenes validation set." + ], + "image_footnote": [], + "bbox": [ + 121, + 415, + 848, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(NMS) to combine the detection results of LiDAR and camera branches, which is even worse than the single LiDAR branch. Therefore, it is also important to make the instance-level features from the two branches interact with each other in the fusion stage.", + "bbox": [ + 75, + 605, + 470, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Strong Image Backbone. We incorporate stronger SwinT [27] backbone into SparseFusion to match some previous work [22, 45]. Tab. 4 compares the performance of different methods on the nuScenes validation set. We do not include any test-time augmentations or model ensembles. Although multi-modality detection relies more on the LiDAR inputs, SparseFusion can still benefit from a stronger image backbone and beat all the counterparts.", + "bbox": [ + 75, + 680, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modality-Specific Object Recall. The parallel detectors in the LiDAR and camera branches enable us to determine which modality recalls each object. Given $N_{L} + N_{C}$ predictions, we know that the first $N_{L}$ instances come from the LiDAR-modality detector, while the last $N_{C}$ are from", + "bbox": [ + 75, + 796, + 470, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Performance breakdown. We show the detection results in different parts of SparseFusion on the validation set. 'L' and 'C' refer to the LiDAR and camera modalities, respectively. 'ST' and 'GT' refer to semantic transfer and geometric transfer, respectively. 'VT' is the view transformation module.", + "bbox": [ + 497, + 603, + 892, + 686 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4cfda4783b3af7890c8e4908afa166e8645ba62f5ba4f79dadd902527b3f9743.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsModalityNDSmAP
TransFusion-L [1]L70.165.1
LiDAR branchL + ST71.868.4
PETR (ResNet50) [26]C38.131.3
Camera branch (before VT)C+GT43.540.6
Camera branch (after VT)C+GT44.341.5
LiDAR/camera branches NMSL+C71.568.2
SparseFusionL+C72.870.4
", + "bbox": [ + 500, + 697, + 912, + 836 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17597", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/50f111b6ff227fcd58c07235ae9949f8035349e4519fa131320261d79230b1bf.jpg", + "table_caption": [ + "Table 4: Results on nuScenes validation set with stronger image backbones. We do not use any test-time augmentations or model ensembles. The inference speed is measured on a single NVIDIA A6000 GPU. $\\dagger$ : CMT [45] adopts flash-attention [7] for transformer acceleration." + ], + "table_footnote": [], + "table_body": "
MethodsImage BackboneInput ResolutionNDSmAPFPS
BEVFusion [22]Dual-Swin-T1600×90072.169.60.8
CMT [45]VoVNet-991600×64072.970.33.8†
SparseFusionResNet-50800×44872.870.45.6
SparseFusionSwin-T800×44873.171.05.3
", + "bbox": [ + 80, + 204, + 465, + 288 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/63be14ae83bf2400e368d6b8b4edac6adbd2f40697cbef7e7152bf986d80f0c5.jpg", + "image_caption": [ + "(a) Objects of different classes. (b) Objects at different distances." + ], + "image_footnote": [], + "bbox": [ + 80, + 304, + 265, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/066a58f9268af594eed9cebf12dc191e3545a0ea764e3657c0750c4a9385e517.jpg", + "image_caption": [ + "Figure 6: Instance-level feature interaction in the fusion stage. Orange boxes are detected after the fusion stage with high confidence in the BEV space. Blue and green dots denote all instances from the LiDAR and camera branches separately. Orange boxes are connected with blue/green dots with blue/green lines. The attention strength is represented by the darkness and thickness of the lines." + ], + "image_footnote": [], + "bbox": [ + 275, + 302, + 460, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the camera modality. An object is recalled if a bounding box with correct classification is predicted within a radius of two meters around its ground-truth position. In Fig. 5, we demonstrate the percentage of objects in the nuScenes validation set recalled by exactly one modality. We observe that each modality can compensate for the weakness of the other to some extent. Although the LiDAR modality is typically more powerful, the camera modality plays an important role in detecting objects from classes such as construction vehicles and barriers. Furthermore, the camera modality is useful for detecting objects at far distances where point clouds are sparse or the LiDAR feature map cannot cover.", + "bbox": [ + 75, + 477, + 468, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-Modality Sparse Representation Interaction. Fig. 6 visualizes the instance feature interaction in the sparse fusion stage. The strength of attention between instance-level features is reflected by the thickness and darkness of lines. We notice that most objects can aggregate multi-modality instance-level features during fusion. Although the strongest interactions exist mainly between neighboring instances, it is interesting that features from the camera modality are also able to deliver strong interactions with the instances at the distant range. This could be a result of the shared semantics among objects in images.", + "bbox": [ + 75, + 644, + 468, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to Sensor Malfunction. With the relatively independent modality-specific detectors, SparseFusion is robust to sensor malfunctions. We evaluate SparseFusion in two different cases: (1) The LiDAR scan range is limited to $(-90^{\\circ}, 90^{\\circ})$ and (2) The most important front camera is", + "bbox": [ + 75, + 796, + 468, + 866 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc8a60f387e60e5d4eb5d09994a5ee653c979f2431056354735e865e4590674c.jpg", + "image_caption": [ + "Figure 5: Percentage of objects recalled by each modality. We do not count those objects recalled by both or neither." + ], + "image_footnote": [], + "bbox": [ + 500, + 122, + 890, + 388 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b476547d04d07f42dfada579fab0e4d7f37102b06598a86b3977333fee697a63.jpg", + "table_caption": [ + "Table 5: Results on nuScenes validation set for LiDAR or camera sensor malfunctions. The models are finetuned on the in-domain data in both cases." + ], + "table_footnote": [], + "table_body": "
MethodsSensor MalfunctionNDSmAP
TransFusion [1]Limited LiDAR Field (-90°, 90°)49.231.1
BEVFusion [22]54.945.5
SparseFusion61.254.3
TransFusion [1]Missing70.165.3
BEVFusion [22]Front Camera70.765.9
SparseFusion72.169.2
", + "bbox": [ + 500, + 562, + 899, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "missing while other cameras are preserved. SparseFusion outperforms previous methods [1, 22] with a large gap under both scenarios, demonstrating great robustness to sensor malfunctions.", + "bbox": [ + 497, + 694, + 890, + 749 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 499, + 760, + 663, + 774 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we study the effect of using alternatives for the different modules in SparseFusion. For our ablation studies, we train on a 1/5 split of nuScenes training set and evaluate on the full nuScenes validation set.", + "bbox": [ + 497, + 782, + 890, + 836 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse Fusion Strategy. We compare our self-attention module for sparse feature fusion with other fusion methods.", + "bbox": [ + 497, + 838, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17598", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/63f0795a5b509d8de1cdb978e3d5ba1cccecfb96233b2d78e267e3eff3ea91ab.jpg", + "table_caption": [ + "Table 6: Ablation study experiments to justify our design of each module.", + "(a) Sparse fusion strategies." + ], + "table_footnote": [], + "table_body": "
Fusion StrategyNDSmAP
MLP67.864.5
Cross-Attention68.665.8
Optimal Transport68.065.6
Self-Attention68.866.4
", + "bbox": [ + 89, + 162, + 337, + 251 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/9fa060d6aec09c954349a8369ee0d26f002f6dec007fad2d6744d50a1c369dc0.jpg", + "table_caption": [ + "(b) Information transfers." + ], + "table_footnote": [], + "table_body": "
GeometricSemanticNDSmAP
✓✗67.764.2
✗✓68.465.7
✓✓68.866.4
", + "bbox": [ + 359, + 168, + 616, + 244 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b0b91e7035c4b307c3db3ec1ae96b6988baf4bb95b97e40bff8be073e595e4a9.jpg", + "table_caption": [ + "(c) Parallel modality-specific detectors." + ], + "table_footnote": [], + "table_body": "
Seq. Pos.Seq. Feat.NDSmAP
✓X68.165.3
✓✓67.964.2
XX68.866.4
", + "bbox": [ + 631, + 168, + 887, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In addition to our self-attention module in Sec. 3.1, three alternatives are considered: 1) Instance-level candidates from the two branches are directly fed into an MLP without any cross-instance aggregation. 2) LiDAR instance candidates are used as queries and camera instance candidates are used as keys/values. A cross-attention module is then used to fuse multi-modality instance features. 3) We make novel use of optimal transport in LiDAR-camera fusion [33]. We propose to learn a distribution transformation from camera candidates to LiDAR candidates through optimal transport. Then, we can directly fuse them by concatenating the candidates of two branches along the channel dimension. More details about this method are provided in appendix.", + "bbox": [ + 75, + 274, + 468, + 454 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The results in Tab. 6a show that only cross-attention achieves competitive performance to self-attention. Yet, we observe that it relies so heavily on the output of the LiDAR branch that the camera branch is not fully utilized to compensate for the weaknesses of the LiDAR branch. The MLP strategy has limited performance as it does not fuse cross-instance and cross-modality information. Despite the impressive progress of optimal transport in other fields, it fails to learn the correspondences between the instance features of the two modalities and thus has limited performance. In contrast, self-attention is simple, efficient, and effective.", + "bbox": [ + 75, + 456, + 468, + 608 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Information Transfer. We then ablate the geometric and semantic transfer between the LiDAR and camera modalities. The results in Tab. 6b show that the fusion performance benefits from both transfers. This also validates that the disadvantages of both modalities result in negative transfer and that our proposed information transfer modules are indeed effective in mitigating this issue. The semantic transfer module especially improves the final performance since it compensates for the LiDAR modality's lack of semantic information, which is critical for 3D detection.", + "bbox": [ + 75, + 611, + 468, + 749 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Despite the importance of the information transfer, SparseFusion trained on the full training set can reach 71.5 NDS on the validation set even if we remove both the semantic and geometric transfers. This still outperforms many other advanced algorithms [1, 28], reflecting the robustness of our fusion stage to the modality-specific weakness.", + "bbox": [ + 75, + 752, + 468, + 836 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "View Transformation. As explained in Sec. 3.1, we transform the sparse representations of both modalities into one", + "bbox": [ + 75, + 838, + 468, + 866 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "unified space. To validate the effectiveness of this approach, we ablate the view transformation of camera candidates into the LiDAR coordinate space. This results in a more straightforward method where we simply obtain the predictions of two modalities and directly fuse them using self-attention. Ablating the view transformation drops performance from $66.4\\%$ mAP and $68.8\\%$ NDS to $65.6\\%$ mAP and $68.3\\%$ NDS, respectively. This demonstrates that the view transformation is indeed helpful to overall performance.", + "bbox": [ + 497, + 274, + 890, + 397 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Parallel Detectors. In addition to the ablation studies on the modules of the proposed pipeline, we also study alternatives to the structure of the pipeline. SparseFusion uses separate 3D object detectors for the LiDAR and camera modalities in parallel for extracting instance-level features to address the cross-modality domain gap. Alternatively, we consider a sequential pipeline where the camera detector runs after the LiDAR detector. The camera-modality detector inherits the output queries from the LiDAR detector. We consider two variants of this inheritance: 1) using the 3D position and instance features for the initial query of the camera modality; 2) using the 3D position but initializing the instance features from the corresponding image features. The camera detector follows the structure of the PETR [26] decoder (one layer). Tab. 6c shows that both sequential structures yield notably inferior performance, justifying our use of parallel modality-specific detectors.", + "bbox": [ + 497, + 399, + 890, + 634 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 648, + 616, + 662 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We revisit previous LiDAR-camera fusion works and propose SparseFusion, a novel 3D object detection method that utilizes the rarely-explored strategy of fusing sparse representations. SparseFusion extracts instance-level features from each modality separately via parallel 3D object detectors and then treats the instance-level features as the modality-specific candidates. Afterward, we transform the candidates into a unified 3D space, and we are able to fuse the candidates with a lightweight attention module. Extensive experiments demonstrate that SparseFusion achieves state-of-the-art performance on the nuScenes benchmark with the fastest inference speed. We hope SparseFusion will serve as a powerful and efficient baseline for further research into this field.", + "bbox": [ + 497, + 672, + 890, + 865 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17599", + "bbox": [ + 465, + 887, + 504, + 897 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 173, + 138 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 2, 3, 4, 6, 7, 8, 9", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3, 6", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3", + "[4] Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. arXiv preprint arXiv:2203.10642, 2022. 2, 3, 7", + "[5] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Autoalignv2: Deformable feature aggregation for dynamic multi-modal 3d object detection. arXiv preprint arXiv:2207.10316, 2022. 1, 2, 3, 6, 7", + "[6] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mm Detection3d, 2020.6", + "[7] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In Advances in Neural Information Processing Systems, 2022. 8", + "[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021. 3", + "[9] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2918-2927, 2021. 3", + "[10] Kaiming He, Georgia Gkioxari, Piotr Dálár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7", + "[12] Chenxi Huang, Tong He, Haidong Ren, Wenxiao Wang, Binbin Lin, and Deng Cai. Obmo: One bounding box multiple objects for monocular 3d object detection. arXiv preprint arXiv:2212.10049, 2022. 3" + ], + "bbox": [ + 78, + 147, + 470, + 865 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 3", + "[14] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3", + "[15] Yang Jiao, Zequn Jie, Shaoxiang Chen, Jingjing Chen, Xiaolin Wei, Lin Ma, and Yu-Gang Jiang. Msmdfusion: Fusing lidar and camera at multiple scales with multi-depth seeds for 3d object detection. arXiv preprint arXiv:2209.03102, 2022. 2", + "[16] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6", + "[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3", + "[18] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022. 2, 3, 7", + "[19] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 3, 4", + "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 1-18. Springer, 2022. 3", + "[21] Tingting Liang, Xiaojie Chu, Yudong Liu, Yongtao Wang, Zhi Tang, Wei Chu, Jingdong Chen, and Haibin Ling. Cbnet: A composite backbone network architecture for object detection. IEEE Transactions on Image Processing, 31:6893-6906, 2022. 1, 7", + "[22] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3, 6, 7, 8", + "[23] Zhidong Liang, Ming Zhang, Zehan Zhang, Xian Zhao, and Shiliang Pu. Rangercnn: Towards fast and accurate 3d object detection with range image representation. arXiv preprint arXiv:2009.00206, 2020. 3", + "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5, 6", + "[25] Hao Liu, Zhuoran Xu, Dan Wang, Baofeng Zhang, Guan Wang, Bo Dong, Xin Wen, and Xinyu Xu. Pai3d: Painting adaptive instance-prior for 3d object detection. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 459-475. Springer, 2023. 2" + ], + "bbox": [ + 502, + 125, + 890, + 864 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17600", + "bbox": [ + 465, + 887, + 505, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 531-548. Springer, 2022. 3, 6, 7, 9", + "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 7", + "[28] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 1, 2, 3, 6, 7, 9", + "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[30] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3164-3173, 2021. 3", + "[31] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3", + "[32] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 3", + "[33] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355-607, 2019. 9", + "[34] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3", + "[35] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 2, 3", + "[36] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2743-2752, 2021. 3", + "[37] Leslie N Smith. Cyclical learning rates for training neural networks. In 2017 IEEE winter conference on applications of computer vision (WACV), pages 464-472. IEEE, 2017. 6", + "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + ], + "bbox": [ + 78, + 125, + 468, + 865 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 2, 3, 7", + "[40] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 2, 3", + "[41] Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh, and I-Hau Yeh. Cspnet: A new backbone that can enhance learning capability of cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 390–391, 2020. 7", + "[42] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 3, 7", + "[43] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 3", + "[44] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In European Conference on Computer Vision, pages 638-656. Springer, 2022. 2", + "[45] Junjie Yan, Yingfei Liu, Jianjian Sun, Fan Jia, Shuai Lin Li, Tiancai Wang, and Xiangyu Zhang. Cross modal transformer via coordinates encoding for 3d object detection. arXiv preprint arXiv:2301.01283, 2023. 2, 3, 6, 7, 8", + "[46] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 3", + "[47] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. arXiv preprint arXiv:2208.11112, 2022. 2, 3, 6, 7", + "[48] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 3, 7", + "[49] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 7", + "[50] Xingyi Zhou, Dequan Wang, and Philipp Krahenbihl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 3", + "[51] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 3, 6, 7", + "[52] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point" + ], + "bbox": [ + 502, + 127, + 890, + 866 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17601", + "bbox": [ + 465, + 887, + 502, + 897 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019.6", + "[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 4, 5" + ], + "bbox": [ + 80, + 127, + 468, + 201 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17602", + "bbox": [ + 467, + 887, + 504, + 897 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/280aa12d-c539-41fe-b78e-1bcf7c1aa152_model.json b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/280aa12d-c539-41fe-b78e-1bcf7c1aa152_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7490f3c37be0f4305de4773a1d963c9a3fa92dd0 --- /dev/null +++ b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/280aa12d-c539-41fe-b78e-1bcf7c1aa152_model.json @@ -0,0 +1,2193 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.086 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.254, + 0.049, + 0.79, + 0.062 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.334, + 0.063, + 0.71, + 0.075 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.305, + 0.076, + 0.74, + 0.088 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.186, + 0.162, + 0.785, + 0.203 + ], + "angle": 0, + "content": "SparseFusion: Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.228, + 0.851, + 0.261 + ], + "angle": 0, + "content": "Yichen Xie\\(^{1,*}\\), Chenfeng Xu\\(^{1,*}\\), Marie-Julie Rakotosaona\\(^{2}\\), Patrick Rim\\(^{3}\\), Federico Tombari\\(^{2}\\), Kurt Keutzer\\(^{1}\\), Masayoshi Tomizuka\\(^{1}\\), Wei Zhan\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.261, + 0.806, + 0.279 + ], + "angle": 0, + "content": "1 University of California, Berkeley 2 Google 3 California Institute of Technology" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.323, + 0.314, + 0.338 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.474, + 0.712 + ], + "angle": 0, + "content": "By identifying four important components of existing LiDAR-camera 3D object detection methods (LiDAR and camera candidates, transformation, and fusion outputs), we observe that all existing methods either find dense candidates or yield dense representations of scenes. However, given that objects occupy only a small part of a scene, finding dense candidates and generating dense representations is noisy and inefficient. We propose SparseFusion, a novel multi-sensor 3D detection method that exclusively uses sparse candidates and sparse representations. Specifically, SparseFusion utilizes the outputs of parallel detectors in the LiDAR and camera modalities as sparse candidates for fusion. We transform the camera candidates into the LiDAR coordinate space by disentangling the object representations. Then, we can fuse the multi-modality candidates in a unified 3D space by a lightweight self-attention module. To mitigate negative transfer between modalities, we propose novel semantic and geometric cross-modality transfer modules that are applied prior to the modality-specific detectors. SparseFusion achieves state-of-the-art performance on the nuScenes benchmark while also running at the fastest speed, even outperforming methods with stronger backbones. We perform extensive experiments to demonstrate the effectiveness and efficiency of our modules and overall method pipeline. Our code will be made publicly available at https://github.com/yichen928/SparseFusion." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.738, + 0.21, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Autonomous driving cars rely on multiple sensors, such as LiDAR and cameras, to perceive the surrounding environment. LiDAR sensors provide accurate 3D scene occupancy information through point clouds with points in the xyz coordinate space, and cameras provide rich semantic information through images with pixels in the RGB" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.309, + 0.691, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.412, + 0.648, + 0.424 + ], + "angle": 0, + "content": "(a) mAP vs. FPS" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.318, + 0.892, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.412, + 0.876, + 0.424 + ], + "angle": 0, + "content": "(b) Dense-to-dense fusion." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.432, + 0.892, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.5, + 0.529, + 0.892, + 0.567 + ], + "angle": 0, + "content": "(c) Overview of our sparse fusion strategy. We extract instance-level features from the LiDAR and camera modalities separately, and fuse them in a unified 3D space to perform detection." + }, + { + "type": "image_footnote", + "bbox": [ + 0.498, + 0.578, + 0.893, + 0.676 + ], + "angle": 0, + "content": "Figure 1: Compared to existing fusion algorithms, Sparse-Fusion achieves state-of-the-art performance as well as the fastest inference speed on nuScenes test set. † : Official code of [5] uses flip as test-time augmentation. ‡ : We use BEVFusion-base results in the official repository of [28] to match the input resolutions of other methods. § : Swin-T [27, 21] is adopted as image backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.893, + 0.838 + ], + "angle": 0, + "content": "color space. However, there are often significant discrepancies between representations of the same physical scene acquired by the two sensors, as LiDAR sensors capture point clouds using 360-degree rotation while cameras capture images from a perspective view without a sense of depth. This impedes an effective and efficient fusion of the LiDAR and camera modalities. To tackle this challenge, multi-sensor fusion algorithms were proposed to find correspondences between multi-modality data to transform and fuse them into a unified scene representation space." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Dense representations, such as bird-eye-view (BEV), volumetric, and point representations, are commonly used" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.854, + 0.264, + 0.866 + ], + "angle": 0, + "content": "* indicates equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "17591" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.124, + 0.895, + 0.265 + ], + "angle": 0, + "content": "Table 1: For each LiDAR-camera fusion method, we identify the LiDAR candidates and camera candidates that are used, the transformation process used to fuse these candidates into a unified space, and the fusion outputs generated to represent 3D scenes using information from both modalities. Based on these components, we categorize the methods into the following categories: Dense+Sparse \\(\\rightarrow\\) Dense approaches relate the sparse region proposals in images to a dense frustum point cloud and fuse them into a dense point space. Dense+Dense \\(\\rightarrow\\) Dense approaches align each point feature in the point cloud to the corresponding pixel feature in the image and represent the 3D scenes using dense features such as point/BEV features. Sparse+Dense \\(\\rightarrow\\) Sparse approaches generate sparse queries by detecting instance features in the point cloud and then applying cross-attention with dense image features. Dense+Dense \\(\\rightarrow\\) Sparse approaches predict objects using queries that combine dense features from each modality. Sparse+Sparse \\(\\rightarrow\\) Sparse (ours) extracts sparse instance features from each modality and directly fuses them to obtain the final sparse instance features used for detection." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.273, + 0.872, + 0.445 + ], + "angle": 0, + "content": "
CategoryMethodLiDAR CandidateCamera CandidateTransformationFusion Outputs
Dense+Sparse→DenseFrustum PointNets [35]point featuresregion proposalsproj. & concat.point features
Dense+Dense→DensePointPainting [39]point featuressegm. outputproj. & concat.point features
PointAugmenting [40]point featuresimage featuresproj. & concat.point features
PAI3D [25]point featuressegm. outputproj. & concat.point features
BEVFusion [28]BEV featuresimage featuresdepth. est. & proj. & concat.BEV features
AutoAlignV2 [5]voxel featuresimage featuresproj. & attn.voxel features
Sparse+Dense→SparseTransFusion [1]instance featuresimage featuresproj. & attn.instance features
Dense+Dense→SparseFUTR3D [4]voxel featuresimage featuresattn.instance features
UVTR [18]voxel featuresimage featuresdepth. est. & proj. & attn.instance features
DeepInteraction [47]BEV featuresimage featuresproj. & attn.instance features
CMT [45]BEV featuresimage featuresattn.instance features
Sparse+Sparse→SparseSparseFusion (Ours)instance featuresinstance featuresproj. & attn.instance features
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.632 + ], + "angle": 0, + "content": "to represent 3D scenes [28, 5, 39, 25, 40, 15]. Most previous works fuse different modalities by aligning low-level data or high-level features to yield dense features that describe the entire 3D scene, e.g., as shown in Fig 1b. However, for the task of 3D object detection, such dense representations are superfluous since we are only interested in instances/objects, which only occupy a small part of the 3D space. Furthermore, noisy backgrounds can be detrimental to object detection performance, and aligning different modalities into the same space is a time-consuming process. For example, generating BEV features from multi-view images takes \\(500\\mathrm{ms}\\) on an RTX 3090 GPU [28]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.471, + 0.868 + ], + "angle": 0, + "content": "In contrast, sparse representations are more efficient, and methods based on them have achieved state-of-the-art performance in multi-sensor 3D detection [1, 4, 18, 47]. These methods use object queries to represent instances/objects in the scene and interact with the original image and point cloud features. However, most previous works do not take into account the significant domain gap between features from different modalities [44]. The queries may gather information from one modality that has a large distribution shift with respect to another modality, making iterative interaction between modalities with large gaps suboptimal. Recent work [47] mitigates this issue by incorporating modality interaction, i.e. performing cross-attention between features from two different modalities. However, the number of computations performed in this method increases quadratically with the dimensions of features and is thus inefficient. We categorize previous works into four" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.508 + ], + "angle": 0, + "content": "groups by identifying four key components, which are outlined in Table 1. Further discussion of the methods in these groups is presented in Sec. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.893, + 0.868 + ], + "angle": 0, + "content": "In this paper, we propose SparseFusion, a novel method (Fig. 1c) that simultaneously utilizes sparse candidates and yields sparse representations, enabling efficient and effective 3D object detection. SparseFusion is the first LiDAR-camera fusion method, to our knowledge, to perform 3D detection using exclusively sparse candidates and sparse fusion outputs. We highlight a key common ground between the two modalities: an image and a point cloud that represent the same 3D scene will contain mostly the same instances/objects. To leverage this commonality, we perform 3D object detection on the inputs from each modality in two parallel branches. Then, the instance features from each branch are projected into a unified 3D space. Since the instance-level features are sparse representations of the same objects in the same scene, we are able to fuse them with a lightweight attention module [38] in a soft manner. This parallel detection strategy allows the LiDAR and camera branches to take advantage of the unique strengths of the point cloud and image representations, respectively. Nevertheless, the drawbacks of each single-modality detector may result in negative transfer during the fusion phase. For example, the point cloud detector may struggle to distinguish between a standing person and a tree trunk due to a lack of detailed semantic information, while the image detector is hard to localize objects in the 3D space due to a lack of accurate depth information. To mitigate the issue of negative" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "17592" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.127, + 0.469, + 0.181 + ], + "angle": 0, + "content": "transfer, we introduce a novel cross-modality information transfer method designed to compensate for the deficiencies of each modality. This method is applied to the inputs from both modalities prior to the parallel detection branches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.182, + 0.471, + 0.278 + ], + "angle": 0, + "content": "SparseFusion achieves state-of-the-art results on the competitive nuScenes benchmark [2]. Our instance-level sparse fusion strategy allows for a lighter network and much higher efficiency in comparison with prior work. With the same backbone, SparseFusion outperforms the current state-of-the-art model [47] with 1.8x acceleration. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.282, + 0.469, + 0.351 + ], + "angle": 0, + "content": "- We revisit prior LiDAR-camera fusion works and identify four important components that allow us to categorize existing methods into four groups. We propose an entirely new category of methods that exclusively uses sparse candidates and representations." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.356, + 0.469, + 0.424 + ], + "angle": 0, + "content": "- We propose SparseFusion, a novel method for LiDAR-camera 3D object detection that leverages instance-level sparse feature fusion and cross-modality information transfer to take advantage of the strengths of each modality while mitigating their weaknesses." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.429, + 0.469, + 0.484 + ], + "angle": 0, + "content": "- We demonstrate that our method achieves state-of-the-art performance in 3D object detection with a lightweight architecture that provides the fastest inference speed." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.497, + 0.219, + 0.511 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.469, + 0.741 + ], + "angle": 0, + "content": "LiDAR-based 3D Object Detection. LiDAR sensors are commonly used for single-modality 3D object detection due to the accurate geometric information provided by point clouds. For detection in outdoor scenes, most existing methods transform unordered point clouds into more structured data formats such as pillars [17], voxels [46, 8], or range views [9, 23]. Features are extracted by standard 2D or 3D convolutional networks, based on which a detection head is used to recognize objects and regress 3D bounding boxes. Mainstream detection heads apply anchor-based [51, 17] or center-based [50] structures. Inspired by the promising performance of transformer-based methods in 2D detection, some recent works explore transformers as feature extractors [30, 36] or as detection heads [1, 31]. Our method is agnostic to the LiDAR-based detector used in the LiDAR branch, and the default setting uses TransFusion-L [1]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.469, + 0.866 + ], + "angle": 0, + "content": "Camera-based 3D Object Detection. Camera-based 3D detection methods are also being studied increasingly. Early work performs monocular 3D object detection by attaching extra 3D bounding box regression heads [48, 42] to 2D detectors. In practice, scenes are often perceived by multiple cameras from different perspective views. Following LSS [34], methods like BEVDet [14] and BEVDepth [19] extract 2D features from multi-view images and project them into the BEV space. Other methods including" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.892, + 0.278 + ], + "angle": 0, + "content": "DETR3D [43] and PETR [26] adapt techniques from transformer-based 2D object detection methods [53, 3] to learn correspondences between different perspective views through cross-attention using 3D queries. However, as revealed in [12], there inevitably exists some ambiguity when recovering 3D geometry from 2D images. In response, recent works [13, 20, 32] also explore the positive effects of temporal cues in camera-based 3D detection. In our proposed SparseFusion, we extend deformable-DETR [53] to monocular 3D object detection and explicitly transform the regressed bounding boxes to the LiDAR coordinate space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Multi-Modality 3D Object Detection. LiDAR and cameras provide complementary information about the surrounding environment, so it is appealing to fuse the multimodality inputs for 3D object detection tasks. As analyzed in Tab. 1, existing fusion methods can be classified into four categories. Early works tend to fuse multi-modality information into a unified dense representation. Frustum Point-Nets [35] utilizes a Dense+Sparse \\(\\rightarrow\\) Dense approach that filters dense point clouds with sparse 2D regions of interest. Subsequent works explore Dense+Dense \\(\\rightarrow\\) Dense approaches by working directly with the dense LiDAR modality and camera modality features instead. Methods such as [39, 40] project point clouds into image perspective views and concatenate the dense image features with point features. BEVFusion [28, 22] significantly improves the performance of this line of methods by projecting dense image features into the LiDAR coordinate space using estimated per-pixel depths. AutoAlignV2 [5] also considers the soft correspondence through cross-modality attention to increase the robustness. However, we point out that dense representations are altogether undesirable for 3D object detection as they are noisy and inefficient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Recent works have begun to utilize object-centric sparse scene representations. TransFusion [1] adopts a Sparse+Dense \\(\\rightarrow\\) Sparse strategy by extracting sparse instance-level features from the LiDAR modality and refining them using dense image features. Other works [4, 18, 47, 45] utilize a Dense+Dense \\(\\rightarrow\\) Sparse approach where queries are used to extract a sparse instance-level representation from dense BEV and image features. However, it is hard to extract information from multi-modality features with an attention operation given the large cross-modal distribution shift. To this end, UVTR [18] projects image features into the LiDAR coordinate space, CMT [45] encodes modality-specific positional information to its queries, and DeepInteraction [47] proposes cross-modality interaction. However, these methods still need to resolve the large multimodal domain gap by stacking many transformer layers to construct a heavy decoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.867 + ], + "angle": 0, + "content": "In contrast to the above methods, our method adopts the previously unexplored Sparse+Sparse \\(\\rightarrow\\) Sparse approach. SparseFusion extracts sparse representations of both modal-" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "17593" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.127, + 0.47, + 0.169 + ], + "angle": 0, + "content": "ities and fuses them to generate a more accurate and semantically rich sparse representation that yields great performance while also achieving great efficiency." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.183, + 0.212, + 0.199 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.206, + 0.471, + 0.4 + ], + "angle": 0, + "content": "We present SparseFusion, an effective and efficient framework for 3D object detection via LiDAR and camera inputs. The overall architecture is illustrated in Fig. 2. We acquire sparse candidates from each modality using modality-specific object detection in the LiDAR and camera branches. The instance-level features generated by the camera branch are transformed into the LiDAR space of the instance-level features generated by the LiDAR branch. They are then fused with a simple self-attention module (Sec. 3.1). To mitigate the negative transfer between modalities, we apply a geometric transfer module and a semantic transfer module prior to the parallel detection branches (Sec. 3.2). Furthermore, we design custom loss functions for each module to ensure stable optimization (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.411, + 0.342, + 0.425 + ], + "angle": 0, + "content": "3.1. Sparse Representation Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Acquiring candidates in two modalities. LiDAR candidates. We follow TransFusion-L [1] and use only one decoder layer for LiDAR modality detection. The LiDAR backbone extracts a BEV feature map from the point cloud inputs. We initialize \\( N_{L} \\) object queries \\( \\mathbf{Q}_L^0 = \\{\\mathbf{q}_{L,i}^0\\}_{i=1}^{N_L}, \\mathbf{q}_{L,i}^0 \\in \\mathbb{R}^C \\) as well as their corresponding reference points \\( \\mathbf{p}_L^0 = \\{\\mathbf{p}_{L,i}^0\\}_{i=1}^{N_L}, \\mathbf{p}_{L,i}^0 \\in \\mathbb{R}^2 \\) in the BEV plane. These queries interact with the BEV features through a cross-attention layer to generate the updated object queries \\( \\mathbf{Q}_L = \\{\\mathbf{q}_{L,i}\\}_{i=1}^{N_L}, \\mathbf{q}_{L,i} \\in \\mathbb{R}^C \\). These updated queries \\( \\mathbf{Q}_L \\) represent the instance-level features of objects in the LiDAR modality, and we use them as the LiDAR candidates in the subsequent multi-modal fusion module. Furthermore, we apply a prediction head to each query to classify the object and regress the bounding box in LiDAR coordinate space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Camera candidates. To generate the camera candidates, we utilize a camera-only 3D detector with images from different perspective views as inputs. Specifically, we extend deformable-DETR [53] with 3D box regression heads. We also initialize \\(N_{C}\\) object queries \\(\\mathbf{Q}_C^0 = \\{\\mathbf{q}_{C,i}^0\\}_{i = 1}^{N_C}\\), \\(\\mathbf{q}_{C,i}^0 \\in \\mathbb{R}^C\\) along with their corresponding reference points \\(\\mathbf{p}_C^0 = \\{\\mathbf{p}_{C,i}^0\\}_{i = 1}^{N_C}\\), \\(\\mathbf{p}_{C,i}^0 \\in \\mathbb{R}^2\\) on the image. For each perspective view \\(v\\), queries on its image interact with the corresponding image features using a deformable attention layer [53]. The outputs of all perspective views comprise the updated queries \\(\\mathbf{Q}_C^P = \\{\\mathbf{q}_{C,i}^P\\}_{i = 1}^{N_C}\\), \\(\\mathbf{q}_{C,i}^P \\in \\mathbb{R}^C\\). We use these queries as the camera candidates in the subsequent multi-modal fusion module. We provide further details of our architecture, the initialization method, and the prediction heads for the two modalities in our supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.853, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Transformation After acquiring the candidates from each" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.892, + 0.306 + ], + "angle": 0, + "content": "modality, we aim to transform the candidates from the camera modality to the space of the candidates from the LiDAR modality. Since the candidates from the camera modality are high-dimensional latent features that are distributed differently than the candidates from the LiDAR modality, a naive coordinate transformation between modalities is inapplicable here. To address this issue, we disentangle the representations of the camera candidates. Intrinsically, a camera candidate is an instance feature that is a representation of a specific object's class and 3D bounding box. While an object's class is view-invariant, its 3D bounding box is view-dependent. This motivates us to focus on transforming high-dimensional bounding box representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.893, + 0.486 + ], + "angle": 0, + "content": "We first input the candidate instance features into the prediction head of the camera branch. We label the outputted bounding boxes as \\(\\mathbf{b}^P\\). Given the extrinsic matrix \\([\\mathbf{R}_v|\\mathbf{t}_v]\\) and intrinsic matrix \\(\\mathbf{K}_v\\) of the corresponding \\(v\\)-th camera, the bounding boxes can be easily projected into the LiDAR coordinate system. We denote the project bounding boxes as \\(\\mathbf{b}^L\\). We encode the projected bounding boxes with a multi-layer perceptron (MLP) \\(g(\\cdot)\\), yielding a high-dimensional box embedding. We also encode the flattened camera parameter with another MLP \\(m(\\cdot)\\) to obtain a camera embedding. The camera embedding is multiplied with the original instance features as done in [19], which are then added to the box embedding, given by" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.496, + 0.891, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {C, i} ^ {L} = g \\left(\\mathbf {b} _ {i} ^ {L}\\right) + h \\left(\\mathbf {q} _ {C, i} ^ {P} \\cdot m \\left(\\mathbf {R} _ {v}, \\mathbf {t} _ {v}, \\mathbf {K} _ {v}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.891, + 0.621 + ], + "angle": 0, + "content": "where \\( h(\\cdot) \\) is an extra MLP to encode the query features in the perspective view. \\( h(\\cdot) \\) aims to preserve view-agnostic information while discarding view-specific information. Afterward, \\( \\mathbf{Q}_C^L = \\{\\mathbf{q}_{C,i}^L\\} \\) is passed through a self-attention layer to aggregate information from multiple cameras to get the updated queries \\( \\mathbf{Q}_C \\) which represent the image modality instance features in the LiDAR space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Sparse candidate fusion. Our parallel modality-specific object detection provides sparse instance candidates \\(\\mathbf{Q}_L = \\{\\mathbf{q}_{L,i}\\}_{i=1}^{N_L}\\) and \\(\\mathbf{Q}_C = \\{\\mathbf{q}_{C,i}\\}_{i=1}^{N_C}\\) from the LiDAR and camera modalities respectively. After the above transformation of the camera candidates into LiDAR space, candidates from both modalities represent bounding boxes in the same LiDAR coordinate space, as well as the view-invariant categories. We now concatenate the candidates together:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.743, + 0.891, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {L C} = \\left\\{\\mathbf {q} _ {L C, i} \\right\\} _ {i = 1} ^ {N _ {L} + N _ {C}} = \\left\\{f _ {L} (\\mathbf {q} _ {L, i}) \\right\\} _ {i = 1} ^ {N _ {L}} \\cup \\left\\{f _ {C} (\\mathbf {q} _ {C, i}) \\right\\} _ {i = 1} ^ {N _ {C}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.892, + 0.866 + ], + "angle": 0, + "content": "where \\( f_{L}(\\cdot), f_{C}(\\cdot) \\) are learnable projectors. Afterward, we make novel use of a self-attention module to fuse the two modalities. Despite the simplicity of self-attention, the inherent intuition is novel: the modality-specific detectors encode the advantageous aspects of their respective inputs, and the self-attention module is able to aggregate and preserve the information from both modalities in an efficient" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "17594" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.123, + 0.832, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.404, + 0.893, + 0.434 + ], + "angle": 0, + "content": "Figure 2: Overall framework of SparseFusion. It fuses sparse candidates from LiDAR and camera modalities to obtain a multi-modality instance-level representation in the unified LiDAR space." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.454, + 0.273, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.454, + 0.468, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.573, + 0.47, + 0.615 + ], + "angle": 0, + "content": "Figure 3: Cross-modality information transfer. We transfer geometric information from LiDAR to camera modality, and semantic information from camera to LiDAR modality." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.64, + 0.47, + 0.669 + ], + "angle": 0, + "content": "manner. The output of the self-attention module is used for final classification and regression of the bounding boxes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.402, + 0.693 + ], + "angle": 0, + "content": "3.2. Cross-Modality Information Transfer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.469, + 0.837 + ], + "angle": 0, + "content": "Although we aim to utilize the advantages of both modalities, we must address that the modalities also have their own disadvantages that can result in negative transfer between modalities. For example, the LiDAR detector struggles to capture rich semantic information, while the camera detector struggles to capture accurate geometric and depth information. To mitigate negative transfer, we propose novel geometric and semantic information transfer modules, as illustrated in Fig. 3, that we apply prior to the modality-specific detectors." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Geometric transfer from LiDAR to camera. We project each point in the LiDAR point cloud input to multi-view" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.553 + ], + "angle": 0, + "content": "images to generate sparse multi-view depth maps. These multi-view depth maps are inputted into a shared encoder to obtain depth features, which are then concatenated with the image features to form depth-aware image features that compensate for the lack of geometric information in camera inputs. The depth-aware image features are used as the input to the camera branch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.706 + ], + "angle": 0, + "content": "Semantic transfer from camera to LiDAR. We project the points in the LiDAR point cloud input to the image inputs, which yields sparse points on the image features. We perform max-pooling to aggregate the resulting multiscale features, and we combine them with the BEV feature through addition. The concatenated features serve as the queries and interact with the multi-scale image features through deformable-attention [53]. The updated queries replace the original queries in the BEV features, which results in the semantic-aware BEV features, which are used for query initialization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.713, + 0.684, + 0.729 + ], + "angle": 0, + "content": "3.3. Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.734, + 0.892, + 0.763 + ], + "angle": 0, + "content": "We apply the Gaussian focal loss [24] to the initialized queries of both modalities, given by" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.771, + 0.891, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n i t}} = \\mathcal {L} _ {G F o c a l} (\\hat {\\mathbf {Y}} _ {L}, \\mathbf {Y} _ {L}) + \\mathcal {L} _ {G F o c a l} (\\hat {\\mathbf {Y}} _ {C}, \\mathbf {Y} _ {C}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.891, + 0.839 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{Y}}_L, \\hat{\\mathbf{Y}}_C\\) are the dense predictions of category-wise heatmaps of the LiDAR and camera modalities, respectively, and \\(\\mathbf{Y}_L, \\mathbf{Y}_C\\) are the corresponding ground-truths." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Then, we apply the loss function for the detectors of the LiDAR and camera modalities, as well as the view transfor" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "17595" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.473, + 0.265 + ], + "angle": 0, + "content": "mation of the camera candidates and the candidate fusion stage. Firstly, the predictions of each modality-specific detector are independently matched with the ground-truth using the Hungary algorithm [16]. The object classification is optimized with focal loss [24] and the 3D bounding box regression is optimized with L1 loss. For the camera modality detector, the ground-truth bounding boxes are in separate camera coordinates. For all other detectors, ground-truth bounding boxes are in LiDAR coordinates. The detection loss can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.273, + 0.47, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e t e c t}} = \\gamma \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {c a m e r a}} + \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {t r a n s}} + \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {L i D A R}} + \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {f u s i o n}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.298, + 0.47, + 0.34 + ], + "angle": 0, + "content": "Our entire network is optimized using \\(\\mathcal{L} = \\alpha \\mathcal{L}_{init} + \\beta \\mathcal{L}_{detect}\\). In our implementation, we empirically set \\(\\gamma = 1, \\alpha = 0.1\\), and \\(\\beta = 1\\) to balance different terms." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.352, + 0.21, + 0.367 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.374, + 0.267, + 0.388 + ], + "angle": 0, + "content": "4.1. Dataset and metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.473, + 0.576 + ], + "angle": 0, + "content": "We follow previous work [47, 45, 22, 5] to evaluate our method on the nuScenes dataset [2]. It is a challenging dataset for 3D object detection, consisting of 700/150/150 scenes for training/validation/test. It provides point clouds collected using a 32-beam LiDAR and six images from multi-view cameras. There are 1.4 million annotated 3D bounding boxes for objects from 10 different classes. We evaluate performance using the nuScenes detection score (NDS) and mean average precision (mAP) metrics. The final mAP is averaged over distance thresholds \\(0.5\\mathrm{m}\\), \\(1\\mathrm{m}\\), \\(2\\mathrm{m}\\), and \\(4\\mathrm{m}\\) on the BEV across 10 classes. NDS is a weighted average of mAP and other true positive metrics including mATE, mASE, mAOE, mAVE, and mAAE." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.583, + 0.295, + 0.598 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.47, + 0.77 + ], + "angle": 0, + "content": "Our implementation is based on the MMDetection3D framework [6]. For the camera branch, we use ResNet-50 [11] as the backbone and initialize it with the Mask R-CNN [10] instance segmentation network pretrained on nuImage [2]. The input image resolution is \\(800 \\times 448\\). For the LiDAR branch, we apply VoxelNet [51] with voxel size \\((0.075m, 0.075m, 0.2m)\\). The detection range is set as \\([-54m, 54m]\\) for the XY-axes and \\([-5m, 3m]\\) for the Z-axis. Our LiDAR and camera modality detectors both include only 1 decoder layer. The query numbers for the LiDAR and camera modalities are set as \\(N_L = N_C = 200\\), so our fusion stage can detect at most 400 objects per scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Since our framework disentangles the camera detector and the LiDAR detector, we can conveniently apply data augmentation separately to the LiDAR inputs and camera inputs. We apply random rotation, scaling, translation, and flipping to the LiDAR inputs, and we apply random scaling and horizontal flipping to the camera inputs. Our training pipeline follows previous works [1, 28, 47]. We first train" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.895, + 0.293 + ], + "angle": 0, + "content": "TransFusion-L [1] as our LiDAR-only baseline, which is used to initialize our LiDAR backbone and LiDAR modality detector. This LiDAR-only baseline is trained for 20 epochs. Afterward, we freeze the pretrained LiDAR components and train the entire fusion framework for 6 epochs. For both training stages, we use the AdamW optimizer [29] with one-cycle learning rate policy [37]. The initial learning rate is \\(10^{-4}\\) and the weight decay is \\(10^{-2}\\). The hidden dimensions in the entire model except the backbones are 128. For both training stages, we adopt CBGS [52] to balance the class distribution. We train our method on four NVIDIA A6000 GPUs with batch size 16." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.303, + 0.727, + 0.318 + ], + "angle": 0, + "content": "4.3. Results and Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.895, + 0.546 + ], + "angle": 0, + "content": "We report results on the nuScenes validation and test sets, without using any test-time augmentations or model ensembles. As shown in Tab. 2, SparseFusion significantly improves our LiDAR-only baseline, TransFusion-L [1], by \\(+3.6\\%\\) NDS and \\(+6.3\\%\\) mAP on the test set, due to the additional use of camera inputs. More importantly, SparseFusion sets a new state of the art on both the validation set and test set, outperforming prior works including those using stronger backbones. It is noteworthy that SparseFusion demonstrates a \\(0.4\\%\\) NDS and \\(1.0\\%\\) mAP improvement over the most recent state-of-the-art [47] while also achieving a 1.8x speedup (5.6 FPS vs. 3.1 FPS) on A6000 GPU. It can be seen in Fig. 1a that, in addition to superior performance, SparseFusion also provides the fastest inference speed. We also demonstrate the performance of SparseFusion by visualizing some qualitative results in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.555, + 0.603, + 0.57 + ], + "angle": 0, + "content": "4.4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.895, + 0.868 + ], + "angle": 0, + "content": "Performance Breakdown SparseFusion performs 3D object detection in both the LiDAR and camera branch separately. Tab. 3 shows the detection performance in different parts of SparseFusion including the LiDAR branch, camera branch (before and after view transformation), and the fusion branch. We notice that our LiDAR branch detection results notably surpass the LiDAR-only baseline TransFusion-L [1] since the proposed semantic transfer can compensate for the weakness of point cloud inputs. In comparison with the state-of-the-art single-frame camera detector PETR [26] (six decoder layers), our camera branch achieves much better performance with just one decoder layer, owing to the depth-aware features from the proposed geometric transfer. Besides, our view transformation module not only transforms the instance features from the camera coordinate space into the LiDAR coordinate space, but it also slightly improves the camera branch detection performance by aggregating multi-view information. With this strong performance, the modality-specific detectors in each branch would not cause negative transfer during fusion. We also consider direct Non-Maximum Suppression" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "17596" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.21, + 0.125, + 0.761, + 0.138 + ], + "angle": 0, + "content": "Table 2: Comparison with existing methods on nuScenes validation set and test set." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.149, + 0.885, + 0.404 + ], + "angle": 0, + "content": "
MethodsModalityLiDAR BackboneCamera Backbonevalidation settest set
NDSmAPNDSmAP
FCOS3D [42]Camera-ResNet-101 [11]41.534.342.835.8
PETR [26]Camera-ResNet-101 [11]44.237.045.539.1
CenterPoint [48]LiDARVoxelNet [51]-66.859.667.360.3
TransFusion-L [1]LiDARVoxelNet [51]-70.165.170.265.5
PointAugmenting [39]LiDAR+CameraVoxelNet [51]DLA34 [49]--71.066.8
FUTR3D [4]LiDAR+CameraVoxelNet [51]ResNet-101 [11]68.364.5--
UVTR [18]LiDAR+CameraVoxelNet [51]ResNet-101 [11]70.265.471.167.1
TransFusion [1]LiDAR+CameraVoxelNet [51]ResNet-50 [11]71.367.571.668.9
AutoAlignV2 [5]LiDAR+CameraVoxelNet [51]CSPNet [41]71.267.172.468.4
BEVFusion [22]LiDAR+CameraVoxelNet [51]Dual-Swin-T [21]72.169.673.371.3
BEVFusion [28]LiDAR+CameraVoxelNet [51]Swin-T [27]71.468.572.970.2
DeepInteraction [47]LiDAR+CameraVoxelNet [51]ResNet-50 [11]72.669.973.470.8
CMT [45]LiDAR+CameraVoxelNet [51]ResNet-50 [11]70.867.9--
SparseFusion (ours)LiDAR+CameraVoxelNet [51]ResNet-50 [11]72.870.473.872.0
" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.416, + 0.85, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.579, + 0.725, + 0.594 + ], + "angle": 0, + "content": "Figure 4: Qualitative results of SparseFusion on nuScenes validation set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.471, + 0.676 + ], + "angle": 0, + "content": "(NMS) to combine the detection results of LiDAR and camera branches, which is even worse than the single LiDAR branch. Therefore, it is also important to make the instance-level features from the two branches interact with each other in the fusion stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Strong Image Backbone. We incorporate stronger SwinT [27] backbone into SparseFusion to match some previous work [22, 45]. Tab. 4 compares the performance of different methods on the nuScenes validation set. We do not include any test-time augmentations or model ensembles. Although multi-modality detection relies more on the LiDAR inputs, SparseFusion can still benefit from a stronger image backbone and beat all the counterparts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Modality-Specific Object Recall. The parallel detectors in the LiDAR and camera branches enable us to determine which modality recalls each object. Given \\( N_{L} + N_{C} \\) predictions, we know that the first \\( N_{L} \\) instances come from the LiDAR-modality detector, while the last \\( N_{C} \\) are from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.894, + 0.687 + ], + "angle": 0, + "content": "Table 3: Performance breakdown. We show the detection results in different parts of SparseFusion on the validation set. 'L' and 'C' refer to the LiDAR and camera modalities, respectively. 'ST' and 'GT' refer to semantic transfer and geometric transfer, respectively. 'VT' is the view transformation module." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.698, + 0.913, + 0.837 + ], + "angle": 0, + "content": "
MethodsModalityNDSmAP
TransFusion-L [1]L70.165.1
LiDAR branchL + ST71.868.4
PETR (ResNet50) [26]C38.131.3
Camera branch (before VT)C+GT43.540.6
Camera branch (after VT)C+GT44.341.5
LiDAR/camera branches NMSL+C71.568.2
SparseFusionL+C72.870.4
" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "17597" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.124, + 0.47, + 0.193 + ], + "angle": 0, + "content": "Table 4: Results on nuScenes validation set with stronger image backbones. We do not use any test-time augmentations or model ensembles. The inference speed is measured on a single NVIDIA A6000 GPU. \\(\\dagger\\) : CMT [45] adopts flash-attention [7] for transformer acceleration." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.205, + 0.467, + 0.289 + ], + "angle": 0, + "content": "
MethodsImage BackboneInput ResolutionNDSmAPFPS
BEVFusion [22]Dual-Swin-T1600×90072.169.60.8
CMT [45]VoVNet-991600×64072.970.33.8†
SparseFusionResNet-50800×44872.870.45.6
SparseFusionSwin-T800×44873.171.05.3
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.305, + 0.266, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.304, + 0.462, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.411, + 0.468, + 0.424 + ], + "angle": 0, + "content": "(a) Objects of different classes. (b) Objects at different distances." + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.434, + 0.468, + 0.463 + ], + "angle": 0, + "content": "Figure 5: Percentage of objects recalled by each modality. We do not count those objects recalled by both or neither." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.469, + 0.643 + ], + "angle": 0, + "content": "the camera modality. An object is recalled if a bounding box with correct classification is predicted within a radius of two meters around its ground-truth position. In Fig. 5, we demonstrate the percentage of objects in the nuScenes validation set recalled by exactly one modality. We observe that each modality can compensate for the weakness of the other to some extent. Although the LiDAR modality is typically more powerful, the camera modality plays an important role in detecting objects from classes such as construction vehicles and barriers. Furthermore, the camera modality is useful for detecting objects at far distances where point clouds are sparse or the LiDAR feature map cannot cover." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.796 + ], + "angle": 0, + "content": "Cross-Modality Sparse Representation Interaction. Fig. 6 visualizes the instance feature interaction in the sparse fusion stage. The strength of attention between instance-level features is reflected by the thickness and darkness of lines. We notice that most objects can aggregate multi-modality instance-level features during fusion. Although the strongest interactions exist mainly between neighboring instances, it is interesting that features from the camera modality are also able to deliver strong interactions with the instances at the distant range. This could be a result of the shared semantics among objects in images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Robustness to Sensor Malfunction. With the relatively independent modality-specific detectors, SparseFusion is robust to sensor malfunctions. We evaluate SparseFusion in two different cases: (1) The LiDAR scan range is limited to \\((-90^{\\circ}, 90^{\\circ})\\) and (2) The most important front camera is" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.123, + 0.892, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.496 + ], + "angle": 0, + "content": "Figure 6: Instance-level feature interaction in the fusion stage. Orange boxes are detected after the fusion stage with high confidence in the BEV space. Blue and green dots denote all instances from the LiDAR and camera branches separately. Orange boxes are connected with blue/green dots with blue/green lines. The attention strength is represented by the darkness and thickness of the lines." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Table 5: Results on nuScenes validation set for LiDAR or camera sensor malfunctions. The models are finetuned on the in-domain data in both cases." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.563, + 0.9, + 0.681 + ], + "angle": 0, + "content": "
MethodsSensor MalfunctionNDSmAP
TransFusion [1]Limited LiDAR Field (-90°, 90°)49.231.1
BEVFusion [22]54.945.5
SparseFusion61.254.3
TransFusion [1]Missing70.165.3
BEVFusion [22]Front Camera70.765.9
SparseFusion72.169.2
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.891, + 0.75 + ], + "angle": 0, + "content": "missing while other cameras are preserved. SparseFusion outperforms previous methods [1, 22] with a large gap under both scenarios, demonstrating great robustness to sensor malfunctions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.761, + 0.664, + 0.775 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.783, + 0.891, + 0.837 + ], + "angle": 0, + "content": "In this section, we study the effect of using alternatives for the different modules in SparseFusion. For our ablation studies, we train on a 1/5 split of nuScenes training set and evaluate on the full nuScenes validation set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.839, + 0.891, + 0.867 + ], + "angle": 0, + "content": "Sparse Fusion Strategy. We compare our self-attention module for sparse feature fusion with other fusion methods." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.899 + ], + "angle": 0, + "content": "17598" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.124, + 0.726, + 0.138 + ], + "angle": 0, + "content": "Table 6: Ablation study experiments to justify our design of each module." + }, + { + "type": "table_caption", + "bbox": [ + 0.132, + 0.151, + 0.295, + 0.163 + ], + "angle": 0, + "content": "(a) Sparse fusion strategies." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.163, + 0.338, + 0.252 + ], + "angle": 0, + "content": "
Fusion StrategyNDSmAP
MLP67.864.5
Cross-Attention68.665.8
Optimal Transport68.065.6
Self-Attention68.866.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.413, + 0.151, + 0.564, + 0.163 + ], + "angle": 0, + "content": "(b) Information transfers." + }, + { + "type": "table", + "bbox": [ + 0.361, + 0.169, + 0.617, + 0.245 + ], + "angle": 0, + "content": "
GeometricSemanticNDSmAP
✓✗67.764.2
✗✓68.465.7
✓✓68.866.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.643, + 0.151, + 0.876, + 0.163 + ], + "angle": 0, + "content": "(c) Parallel modality-specific detectors." + }, + { + "type": "table", + "bbox": [ + 0.633, + 0.169, + 0.888, + 0.245 + ], + "angle": 0, + "content": "
Seq. Pos.Seq. Feat.NDSmAP
✓X68.165.3
✓✓67.964.2
XX68.866.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.469, + 0.455 + ], + "angle": 0, + "content": "In addition to our self-attention module in Sec. 3.1, three alternatives are considered: 1) Instance-level candidates from the two branches are directly fed into an MLP without any cross-instance aggregation. 2) LiDAR instance candidates are used as queries and camera instance candidates are used as keys/values. A cross-attention module is then used to fuse multi-modality instance features. 3) We make novel use of optimal transport in LiDAR-camera fusion [33]. We propose to learn a distribution transformation from camera candidates to LiDAR candidates through optimal transport. Then, we can directly fuse them by concatenating the candidates of two branches along the channel dimension. More details about this method are provided in appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.469, + 0.609 + ], + "angle": 0, + "content": "The results in Tab. 6a show that only cross-attention achieves competitive performance to self-attention. Yet, we observe that it relies so heavily on the output of the LiDAR branch that the camera branch is not fully utilized to compensate for the weaknesses of the LiDAR branch. The MLP strategy has limited performance as it does not fuse cross-instance and cross-modality information. Despite the impressive progress of optimal transport in other fields, it fails to learn the correspondences between the instance features of the two modalities and thus has limited performance. In contrast, self-attention is simple, efficient, and effective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Information Transfer. We then ablate the geometric and semantic transfer between the LiDAR and camera modalities. The results in Tab. 6b show that the fusion performance benefits from both transfers. This also validates that the disadvantages of both modalities result in negative transfer and that our proposed information transfer modules are indeed effective in mitigating this issue. The semantic transfer module especially improves the final performance since it compensates for the LiDAR modality's lack of semantic information, which is critical for 3D detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.469, + 0.837 + ], + "angle": 0, + "content": "Despite the importance of the information transfer, SparseFusion trained on the full training set can reach 71.5 NDS on the validation set even if we remove both the semantic and geometric transfers. This still outperforms many other advanced algorithms [1, 28], reflecting the robustness of our fusion stage to the modality-specific weakness." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.469, + 0.867 + ], + "angle": 0, + "content": "View Transformation. As explained in Sec. 3.1, we transform the sparse representations of both modalities into one" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.399 + ], + "angle": 0, + "content": "unified space. To validate the effectiveness of this approach, we ablate the view transformation of camera candidates into the LiDAR coordinate space. This results in a more straightforward method where we simply obtain the predictions of two modalities and directly fuse them using self-attention. Ablating the view transformation drops performance from \\(66.4\\%\\) mAP and \\(68.8\\%\\) NDS to \\(65.6\\%\\) mAP and \\(68.3\\%\\) NDS, respectively. This demonstrates that the view transformation is indeed helpful to overall performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Parallel Detectors. In addition to the ablation studies on the modules of the proposed pipeline, we also study alternatives to the structure of the pipeline. SparseFusion uses separate 3D object detectors for the LiDAR and camera modalities in parallel for extracting instance-level features to address the cross-modality domain gap. Alternatively, we consider a sequential pipeline where the camera detector runs after the LiDAR detector. The camera-modality detector inherits the output queries from the LiDAR detector. We consider two variants of this inheritance: 1) using the 3D position and instance features for the initial query of the camera modality; 2) using the 3D position but initializing the instance features from the corresponding image features. The camera detector follows the structure of the PETR [26] decoder (one layer). Tab. 6c shows that both sequential structures yield notably inferior performance, justifying our use of parallel modality-specific detectors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.618, + 0.663 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.866 + ], + "angle": 0, + "content": "We revisit previous LiDAR-camera fusion works and propose SparseFusion, a novel 3D object detection method that utilizes the rarely-explored strategy of fusing sparse representations. SparseFusion extracts instance-level features from each modality separately via parallel 3D object detectors and then treats the instance-level features as the modality-specific candidates. Afterward, we transform the candidates into a unified 3D space, and we are able to fuse the candidates with a lightweight attention module. Extensive experiments demonstrate that SparseFusion achieves state-of-the-art performance on the nuScenes benchmark with the fastest inference speed. We hope SparseFusion will serve as a powerful and efficient baseline for further research into this field." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "17599" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.174, + 0.14 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.47, + 0.224 + ], + "angle": 0, + "content": "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 2, 3, 4, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.226, + 0.472, + 0.303 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.471, + 0.381 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.433 + ], + "angle": 0, + "content": "[4] Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. arXiv preprint arXiv:2203.10642, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.435, + 0.47, + 0.487 + ], + "angle": 0, + "content": "[5] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Autoalignv2: Deformable feature aggregation for dynamic multi-modal 3d object detection. arXiv preprint arXiv:2207.10316, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[6] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mm Detection3d, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.54, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[7] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In Advances in Neural Information Processing Systems, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.658, + 0.47, + 0.722 + ], + "angle": 0, + "content": "[9] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2918-2927, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.723, + 0.47, + 0.762 + ], + "angle": 0, + "content": "[10] Kaiming He, Georgia Gkioxari, Piotr Dálár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.866 + ], + "angle": 0, + "content": "[12] Chenxi Huang, Tong He, Haidong Ren, Wenxiao Wang, Binbin Lin, and Deng Cai. Obmo: One bounding box multiple objects for monocular 3d object detection. arXiv preprint arXiv:2212.10049, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.165 + ], + "angle": 0, + "content": "[13] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[14] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.267 + ], + "angle": 0, + "content": "[15] Yang Jiao, Zequn Jie, Shaoxiang Chen, Jingjing Chen, Xiaolin Wei, Lin Ma, and Yu-Gang Jiang. Msmdfusion: Fusing lidar and camera at multiple scales with multi-depth seeds for 3d object detection. arXiv preprint arXiv:2209.03102, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.268, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[16] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.892, + 0.42 + ], + "angle": 0, + "content": "[18] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[19] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.472, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 1-18. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[21] Tingting Liang, Xiaojie Chu, Yudong Liu, Yongtao Wang, Zhi Tang, Wei Chu, Jingdong Chen, and Haibin Ling. Cbnet: A composite backbone network architecture for object detection. IEEE Transactions on Image Processing, 31:6893-6906, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.625, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[22] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[23] Zhidong Liang, Ming Zhang, Zehan Zhang, Xian Zhao, and Shiliang Pu. Rangercnn: Towards fast and accurate 3d object detection with range image representation. arXiv preprint arXiv:2009.00206, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.74, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.865 + ], + "angle": 0, + "content": "[25] Hao Liu, Zhuoran Xu, Dan Wang, Baofeng Zhang, Guan Wang, Bo Dong, Xin Wen, and Xinyu Xu. Pai3d: Painting adaptive instance-prior for 3d object detection. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 459-475. Springer, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.506, + 0.899 + ], + "angle": 0, + "content": "17600" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 531-548. Springer, 2022. 3, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.47, + 0.27 + ], + "angle": 0, + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.272, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[28] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 1, 2, 3, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.337, + 0.469, + 0.363 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.366, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[30] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3164-3173, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.431, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[31] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.485, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[32] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[33] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355-607, 2019. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[34] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.707 + ], + "angle": 0, + "content": "[35] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.71, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[36] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2743-2752, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[37] Leslie N Smith. Cyclical learning rates for training neural networks. In 2017 IEEE winter conference on applications of computer vision (WACV), pages 464-472. IEEE, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[39] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[40] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.257, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[41] Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh, and I-Hau Yeh. Cspnet: A new backbone that can enhance learning capability of cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 390–391, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[42] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.448 + ], + "angle": 0, + "content": "[43] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.451, + 0.892, + 0.526 + ], + "angle": 0, + "content": "[44] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In European Conference on Computer Vision, pages 638-656. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[45] Junjie Yan, Yingfei Liu, Jianjian Sun, Fan Jia, Shuai Lin Li, Tiancai Wang, and Xiangyu Zhang. Cross modal transformer via coordinates encoding for 3d object detection. arXiv preprint arXiv:2301.01283, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.605 + ], + "angle": 0, + "content": "[46] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.607, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[47] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. arXiv preprint arXiv:2208.11112, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.892, + 0.709 + ], + "angle": 0, + "content": "[48] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.711, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[49] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[50] Xingyi Zhou, Dequan Wang, and Philipp Krahenbihl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.84 + ], + "angle": 0, + "content": "[51] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.841, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[52] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.888, + 0.503, + 0.898 + ], + "angle": 0, + "content": "17601" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.128, + 0.468, + 0.152 + ], + "angle": 0, + "content": "cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.155, + 0.469, + 0.203 + ], + "angle": 0, + "content": "[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.128, + 0.469, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.468, + 0.888, + 0.505, + 0.898 + ], + "angle": 0, + "content": "17602" + } + ] +] \ No newline at end of file diff --git a/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/280aa12d-c539-41fe-b78e-1bcf7c1aa152_origin.pdf b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/280aa12d-c539-41fe-b78e-1bcf7c1aa152_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a38a721fb61d076d673e118a66d55949eb84ca6e --- /dev/null +++ b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/280aa12d-c539-41fe-b78e-1bcf7c1aa152_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:133efb748e772b954e7b77a68950e49e36658f72daffe0af3135b310075ed9b4 +size 1637316 diff --git a/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/full.md b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5378ffa6e0102bd655de186a6df767e231f9c1e4 --- /dev/null +++ b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/full.md @@ -0,0 +1,297 @@ +# SparseFusion: Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection + +Yichen Xie $^{1,*}$ , Chenfeng Xu $^{1,*}$ , Marie-Julie Rakotosaona $^{2}$ , Patrick Rim $^{3}$ , Federico Tombari $^{2}$ , Kurt Keutzer $^{1}$ , Masayoshi Tomizuka $^{1}$ , Wei Zhan $^{1}$ + +1 University of California, Berkeley 2 Google 3 California Institute of Technology + +# Abstract + +By identifying four important components of existing LiDAR-camera 3D object detection methods (LiDAR and camera candidates, transformation, and fusion outputs), we observe that all existing methods either find dense candidates or yield dense representations of scenes. However, given that objects occupy only a small part of a scene, finding dense candidates and generating dense representations is noisy and inefficient. We propose SparseFusion, a novel multi-sensor 3D detection method that exclusively uses sparse candidates and sparse representations. Specifically, SparseFusion utilizes the outputs of parallel detectors in the LiDAR and camera modalities as sparse candidates for fusion. We transform the camera candidates into the LiDAR coordinate space by disentangling the object representations. Then, we can fuse the multi-modality candidates in a unified 3D space by a lightweight self-attention module. To mitigate negative transfer between modalities, we propose novel semantic and geometric cross-modality transfer modules that are applied prior to the modality-specific detectors. SparseFusion achieves state-of-the-art performance on the nuScenes benchmark while also running at the fastest speed, even outperforming methods with stronger backbones. We perform extensive experiments to demonstrate the effectiveness and efficiency of our modules and overall method pipeline. Our code will be made publicly available at https://github.com/yichen928/SparseFusion. + +# 1. Introduction + +Autonomous driving cars rely on multiple sensors, such as LiDAR and cameras, to perceive the surrounding environment. LiDAR sensors provide accurate 3D scene occupancy information through point clouds with points in the xyz coordinate space, and cameras provide rich semantic information through images with pixels in the RGB + +![](images/031d89a9df86fe6b84eb009a319a73d1bff0e654b664a6711427f4fe7fc7f7e1.jpg) +(a) mAP vs. FPS + +![](images/d77b38623dc1b1833de9b991a0d9aaa99ceda8b5b19ab1f59546eb55980df172.jpg) + +(b) Dense-to-dense fusion. +![](images/1690ec947523fe1b553f01f68c5e7b7c1e3c40c807796ab35ac643998afee96f.jpg) +(c) Overview of our sparse fusion strategy. We extract instance-level features from the LiDAR and camera modalities separately, and fuse them in a unified 3D space to perform detection. +Figure 1: Compared to existing fusion algorithms, Sparse-Fusion achieves state-of-the-art performance as well as the fastest inference speed on nuScenes test set. † : Official code of [5] uses flip as test-time augmentation. ‡ : We use BEVFusion-base results in the official repository of [28] to match the input resolutions of other methods. § : Swin-T [27, 21] is adopted as image backbone. + +color space. However, there are often significant discrepancies between representations of the same physical scene acquired by the two sensors, as LiDAR sensors capture point clouds using 360-degree rotation while cameras capture images from a perspective view without a sense of depth. This impedes an effective and efficient fusion of the LiDAR and camera modalities. To tackle this challenge, multi-sensor fusion algorithms were proposed to find correspondences between multi-modality data to transform and fuse them into a unified scene representation space. + +Dense representations, such as bird-eye-view (BEV), volumetric, and point representations, are commonly used + +Table 1: For each LiDAR-camera fusion method, we identify the LiDAR candidates and camera candidates that are used, the transformation process used to fuse these candidates into a unified space, and the fusion outputs generated to represent 3D scenes using information from both modalities. Based on these components, we categorize the methods into the following categories: Dense+Sparse $\rightarrow$ Dense approaches relate the sparse region proposals in images to a dense frustum point cloud and fuse them into a dense point space. Dense+Dense $\rightarrow$ Dense approaches align each point feature in the point cloud to the corresponding pixel feature in the image and represent the 3D scenes using dense features such as point/BEV features. Sparse+Dense $\rightarrow$ Sparse approaches generate sparse queries by detecting instance features in the point cloud and then applying cross-attention with dense image features. Dense+Dense $\rightarrow$ Sparse approaches predict objects using queries that combine dense features from each modality. Sparse+Sparse $\rightarrow$ Sparse (ours) extracts sparse instance features from each modality and directly fuses them to obtain the final sparse instance features used for detection. + +
CategoryMethodLiDAR CandidateCamera CandidateTransformationFusion Outputs
Dense+Sparse→DenseFrustum PointNets [35]point featuresregion proposalsproj. & concat.point features
Dense+Dense→DensePointPainting [39]point featuressegm. outputproj. & concat.point features
PointAugmenting [40]point featuresimage featuresproj. & concat.point features
PAI3D [25]point featuressegm. outputproj. & concat.point features
BEVFusion [28]BEV featuresimage featuresdepth. est. & proj. & concat.BEV features
AutoAlignV2 [5]voxel featuresimage featuresproj. & attn.voxel features
Sparse+Dense→SparseTransFusion [1]instance featuresimage featuresproj. & attn.instance features
Dense+Dense→SparseFUTR3D [4]voxel featuresimage featuresattn.instance features
UVTR [18]voxel featuresimage featuresdepth. est. & proj. & attn.instance features
DeepInteraction [47]BEV featuresimage featuresproj. & attn.instance features
CMT [45]BEV featuresimage featuresattn.instance features
Sparse+Sparse→SparseSparseFusion (Ours)instance featuresinstance featuresproj. & attn.instance features
+ +to represent 3D scenes [28, 5, 39, 25, 40, 15]. Most previous works fuse different modalities by aligning low-level data or high-level features to yield dense features that describe the entire 3D scene, e.g., as shown in Fig 1b. However, for the task of 3D object detection, such dense representations are superfluous since we are only interested in instances/objects, which only occupy a small part of the 3D space. Furthermore, noisy backgrounds can be detrimental to object detection performance, and aligning different modalities into the same space is a time-consuming process. For example, generating BEV features from multi-view images takes $500\mathrm{ms}$ on an RTX 3090 GPU [28]. + +In contrast, sparse representations are more efficient, and methods based on them have achieved state-of-the-art performance in multi-sensor 3D detection [1, 4, 18, 47]. These methods use object queries to represent instances/objects in the scene and interact with the original image and point cloud features. However, most previous works do not take into account the significant domain gap between features from different modalities [44]. The queries may gather information from one modality that has a large distribution shift with respect to another modality, making iterative interaction between modalities with large gaps suboptimal. Recent work [47] mitigates this issue by incorporating modality interaction, i.e. performing cross-attention between features from two different modalities. However, the number of computations performed in this method increases quadratically with the dimensions of features and is thus inefficient. We categorize previous works into four + +groups by identifying four key components, which are outlined in Table 1. Further discussion of the methods in these groups is presented in Sec. 2. + +In this paper, we propose SparseFusion, a novel method (Fig. 1c) that simultaneously utilizes sparse candidates and yields sparse representations, enabling efficient and effective 3D object detection. SparseFusion is the first LiDAR-camera fusion method, to our knowledge, to perform 3D detection using exclusively sparse candidates and sparse fusion outputs. We highlight a key common ground between the two modalities: an image and a point cloud that represent the same 3D scene will contain mostly the same instances/objects. To leverage this commonality, we perform 3D object detection on the inputs from each modality in two parallel branches. Then, the instance features from each branch are projected into a unified 3D space. Since the instance-level features are sparse representations of the same objects in the same scene, we are able to fuse them with a lightweight attention module [38] in a soft manner. This parallel detection strategy allows the LiDAR and camera branches to take advantage of the unique strengths of the point cloud and image representations, respectively. Nevertheless, the drawbacks of each single-modality detector may result in negative transfer during the fusion phase. For example, the point cloud detector may struggle to distinguish between a standing person and a tree trunk due to a lack of detailed semantic information, while the image detector is hard to localize objects in the 3D space due to a lack of accurate depth information. To mitigate the issue of negative + +transfer, we introduce a novel cross-modality information transfer method designed to compensate for the deficiencies of each modality. This method is applied to the inputs from both modalities prior to the parallel detection branches. + +SparseFusion achieves state-of-the-art results on the competitive nuScenes benchmark [2]. Our instance-level sparse fusion strategy allows for a lighter network and much higher efficiency in comparison with prior work. With the same backbone, SparseFusion outperforms the current state-of-the-art model [47] with 1.8x acceleration. Our contributions are summarized as follows: + +- We revisit prior LiDAR-camera fusion works and identify four important components that allow us to categorize existing methods into four groups. We propose an entirely new category of methods that exclusively uses sparse candidates and representations. + +- We propose SparseFusion, a novel method for LiDAR-camera 3D object detection that leverages instance-level sparse feature fusion and cross-modality information transfer to take advantage of the strengths of each modality while mitigating their weaknesses. + +- We demonstrate that our method achieves state-of-the-art performance in 3D object detection with a lightweight architecture that provides the fastest inference speed. + +# 2. Related Work + +LiDAR-based 3D Object Detection. LiDAR sensors are commonly used for single-modality 3D object detection due to the accurate geometric information provided by point clouds. For detection in outdoor scenes, most existing methods transform unordered point clouds into more structured data formats such as pillars [17], voxels [46, 8], or range views [9, 23]. Features are extracted by standard 2D or 3D convolutional networks, based on which a detection head is used to recognize objects and regress 3D bounding boxes. Mainstream detection heads apply anchor-based [51, 17] or center-based [50] structures. Inspired by the promising performance of transformer-based methods in 2D detection, some recent works explore transformers as feature extractors [30, 36] or as detection heads [1, 31]. Our method is agnostic to the LiDAR-based detector used in the LiDAR branch, and the default setting uses TransFusion-L [1]. + +Camera-based 3D Object Detection. Camera-based 3D detection methods are also being studied increasingly. Early work performs monocular 3D object detection by attaching extra 3D bounding box regression heads [48, 42] to 2D detectors. In practice, scenes are often perceived by multiple cameras from different perspective views. Following LSS [34], methods like BEVDet [14] and BEVDepth [19] extract 2D features from multi-view images and project them into the BEV space. Other methods including + +DETR3D [43] and PETR [26] adapt techniques from transformer-based 2D object detection methods [53, 3] to learn correspondences between different perspective views through cross-attention using 3D queries. However, as revealed in [12], there inevitably exists some ambiguity when recovering 3D geometry from 2D images. In response, recent works [13, 20, 32] also explore the positive effects of temporal cues in camera-based 3D detection. In our proposed SparseFusion, we extend deformable-DETR [53] to monocular 3D object detection and explicitly transform the regressed bounding boxes to the LiDAR coordinate space. + +Multi-Modality 3D Object Detection. LiDAR and cameras provide complementary information about the surrounding environment, so it is appealing to fuse the multimodality inputs for 3D object detection tasks. As analyzed in Tab. 1, existing fusion methods can be classified into four categories. Early works tend to fuse multi-modality information into a unified dense representation. Frustum Point-Nets [35] utilizes a Dense+Sparse $\rightarrow$ Dense approach that filters dense point clouds with sparse 2D regions of interest. Subsequent works explore Dense+Dense $\rightarrow$ Dense approaches by working directly with the dense LiDAR modality and camera modality features instead. Methods such as [39, 40] project point clouds into image perspective views and concatenate the dense image features with point features. BEVFusion [28, 22] significantly improves the performance of this line of methods by projecting dense image features into the LiDAR coordinate space using estimated per-pixel depths. AutoAlignV2 [5] also considers the soft correspondence through cross-modality attention to increase the robustness. However, we point out that dense representations are altogether undesirable for 3D object detection as they are noisy and inefficient. + +Recent works have begun to utilize object-centric sparse scene representations. TransFusion [1] adopts a Sparse+Dense $\rightarrow$ Sparse strategy by extracting sparse instance-level features from the LiDAR modality and refining them using dense image features. Other works [4, 18, 47, 45] utilize a Dense+Dense $\rightarrow$ Sparse approach where queries are used to extract a sparse instance-level representation from dense BEV and image features. However, it is hard to extract information from multi-modality features with an attention operation given the large cross-modal distribution shift. To this end, UVTR [18] projects image features into the LiDAR coordinate space, CMT [45] encodes modality-specific positional information to its queries, and DeepInteraction [47] proposes cross-modality interaction. However, these methods still need to resolve the large multimodal domain gap by stacking many transformer layers to construct a heavy decoder. + +In contrast to the above methods, our method adopts the previously unexplored Sparse+Sparse $\rightarrow$ Sparse approach. SparseFusion extracts sparse representations of both modal- + +ities and fuses them to generate a more accurate and semantically rich sparse representation that yields great performance while also achieving great efficiency. + +# 3. Methodology + +We present SparseFusion, an effective and efficient framework for 3D object detection via LiDAR and camera inputs. The overall architecture is illustrated in Fig. 2. We acquire sparse candidates from each modality using modality-specific object detection in the LiDAR and camera branches. The instance-level features generated by the camera branch are transformed into the LiDAR space of the instance-level features generated by the LiDAR branch. They are then fused with a simple self-attention module (Sec. 3.1). To mitigate the negative transfer between modalities, we apply a geometric transfer module and a semantic transfer module prior to the parallel detection branches (Sec. 3.2). Furthermore, we design custom loss functions for each module to ensure stable optimization (Sec. 3.3). + +# 3.1. Sparse Representation Fusion + +Acquiring candidates in two modalities. LiDAR candidates. We follow TransFusion-L [1] and use only one decoder layer for LiDAR modality detection. The LiDAR backbone extracts a BEV feature map from the point cloud inputs. We initialize $N_{L}$ object queries $\mathbf{Q}_L^0 = \{\mathbf{q}_{L,i}^0\}_{i=1}^{N_L}, \mathbf{q}_{L,i}^0 \in \mathbb{R}^C$ as well as their corresponding reference points $\mathbf{p}_L^0 = \{\mathbf{p}_{L,i}^0\}_{i=1}^{N_L}, \mathbf{p}_{L,i}^0 \in \mathbb{R}^2$ in the BEV plane. These queries interact with the BEV features through a cross-attention layer to generate the updated object queries $\mathbf{Q}_L = \{\mathbf{q}_{L,i}\}_{i=1}^{N_L}, \mathbf{q}_{L,i} \in \mathbb{R}^C$ . These updated queries $\mathbf{Q}_L$ represent the instance-level features of objects in the LiDAR modality, and we use them as the LiDAR candidates in the subsequent multi-modal fusion module. Furthermore, we apply a prediction head to each query to classify the object and regress the bounding box in LiDAR coordinate space. + +Camera candidates. To generate the camera candidates, we utilize a camera-only 3D detector with images from different perspective views as inputs. Specifically, we extend deformable-DETR [53] with 3D box regression heads. We also initialize $N_{C}$ object queries $\mathbf{Q}_C^0 = \{\mathbf{q}_{C,i}^0\}_{i = 1}^{N_C}$ , $\mathbf{q}_{C,i}^0 \in \mathbb{R}^C$ along with their corresponding reference points $\mathbf{p}_C^0 = \{\mathbf{p}_{C,i}^0\}_{i = 1}^{N_C}$ , $\mathbf{p}_{C,i}^0 \in \mathbb{R}^2$ on the image. For each perspective view $v$ , queries on its image interact with the corresponding image features using a deformable attention layer [53]. The outputs of all perspective views comprise the updated queries $\mathbf{Q}_C^P = \{\mathbf{q}_{C,i}^P\}_{i = 1}^{N_C}$ , $\mathbf{q}_{C,i}^P \in \mathbb{R}^C$ . We use these queries as the camera candidates in the subsequent multi-modal fusion module. We provide further details of our architecture, the initialization method, and the prediction heads for the two modalities in our supplementary materials. + +Transformation After acquiring the candidates from each + +modality, we aim to transform the candidates from the camera modality to the space of the candidates from the LiDAR modality. Since the candidates from the camera modality are high-dimensional latent features that are distributed differently than the candidates from the LiDAR modality, a naive coordinate transformation between modalities is inapplicable here. To address this issue, we disentangle the representations of the camera candidates. Intrinsically, a camera candidate is an instance feature that is a representation of a specific object's class and 3D bounding box. While an object's class is view-invariant, its 3D bounding box is view-dependent. This motivates us to focus on transforming high-dimensional bounding box representations. + +We first input the candidate instance features into the prediction head of the camera branch. We label the outputted bounding boxes as $\mathbf{b}^P$ . Given the extrinsic matrix $[\mathbf{R}_v|\mathbf{t}_v]$ and intrinsic matrix $\mathbf{K}_v$ of the corresponding $v$ -th camera, the bounding boxes can be easily projected into the LiDAR coordinate system. We denote the project bounding boxes as $\mathbf{b}^L$ . We encode the projected bounding boxes with a multi-layer perceptron (MLP) $g(\cdot)$ , yielding a high-dimensional box embedding. We also encode the flattened camera parameter with another MLP $m(\cdot)$ to obtain a camera embedding. The camera embedding is multiplied with the original instance features as done in [19], which are then added to the box embedding, given by + +$$ +\mathbf {q} _ {C, i} ^ {L} = g \left(\mathbf {b} _ {i} ^ {L}\right) + h \left(\mathbf {q} _ {C, i} ^ {P} \cdot m \left(\mathbf {R} _ {v}, \mathbf {t} _ {v}, \mathbf {K} _ {v}\right)\right), \tag {1} +$$ + +where $h(\cdot)$ is an extra MLP to encode the query features in the perspective view. $h(\cdot)$ aims to preserve view-agnostic information while discarding view-specific information. Afterward, $\mathbf{Q}_C^L = \{\mathbf{q}_{C,i}^L\}$ is passed through a self-attention layer to aggregate information from multiple cameras to get the updated queries $\mathbf{Q}_C$ which represent the image modality instance features in the LiDAR space. + +Sparse candidate fusion. Our parallel modality-specific object detection provides sparse instance candidates $\mathbf{Q}_L = \{\mathbf{q}_{L,i}\}_{i=1}^{N_L}$ and $\mathbf{Q}_C = \{\mathbf{q}_{C,i}\}_{i=1}^{N_C}$ from the LiDAR and camera modalities respectively. After the above transformation of the camera candidates into LiDAR space, candidates from both modalities represent bounding boxes in the same LiDAR coordinate space, as well as the view-invariant categories. We now concatenate the candidates together: + +$$ +\mathbf {Q} _ {L C} = \left\{\mathbf {q} _ {L C, i} \right\} _ {i = 1} ^ {N _ {L} + N _ {C}} = \left\{f _ {L} (\mathbf {q} _ {L, i}) \right\} _ {i = 1} ^ {N _ {L}} \cup \left\{f _ {C} (\mathbf {q} _ {C, i}) \right\} _ {i = 1} ^ {N _ {C}} \tag {2} +$$ + +where $f_{L}(\cdot), f_{C}(\cdot)$ are learnable projectors. Afterward, we make novel use of a self-attention module to fuse the two modalities. Despite the simplicity of self-attention, the inherent intuition is novel: the modality-specific detectors encode the advantageous aspects of their respective inputs, and the self-attention module is able to aggregate and preserve the information from both modalities in an efficient + +![](images/b66143a9b0e469d0fda8b48448f1672540f63fae29399532fef1c08e8e6c97ca.jpg) +Figure 2: Overall framework of SparseFusion. It fuses sparse candidates from LiDAR and camera modalities to obtain a multi-modality instance-level representation in the unified LiDAR space. + +![](images/dcecbe289c436eacd957287d6c1024e522935ae42d50c83d1a29e30ee34546bc.jpg) +Figure 3: Cross-modality information transfer. We transfer geometric information from LiDAR to camera modality, and semantic information from camera to LiDAR modality. + +![](images/1bf6d711fbe9c21d28e7b64e12748c757bc2f9d78de6bafcffe9d5599f6e1ab7.jpg) + +manner. The output of the self-attention module is used for final classification and regression of the bounding boxes. + +# 3.2. Cross-Modality Information Transfer + +Although we aim to utilize the advantages of both modalities, we must address that the modalities also have their own disadvantages that can result in negative transfer between modalities. For example, the LiDAR detector struggles to capture rich semantic information, while the camera detector struggles to capture accurate geometric and depth information. To mitigate negative transfer, we propose novel geometric and semantic information transfer modules, as illustrated in Fig. 3, that we apply prior to the modality-specific detectors. + +Geometric transfer from LiDAR to camera. We project each point in the LiDAR point cloud input to multi-view + +images to generate sparse multi-view depth maps. These multi-view depth maps are inputted into a shared encoder to obtain depth features, which are then concatenated with the image features to form depth-aware image features that compensate for the lack of geometric information in camera inputs. The depth-aware image features are used as the input to the camera branch. + +Semantic transfer from camera to LiDAR. We project the points in the LiDAR point cloud input to the image inputs, which yields sparse points on the image features. We perform max-pooling to aggregate the resulting multiscale features, and we combine them with the BEV feature through addition. The concatenated features serve as the queries and interact with the multi-scale image features through deformable-attention [53]. The updated queries replace the original queries in the BEV features, which results in the semantic-aware BEV features, which are used for query initialization. + +# 3.3. Objective Function + +We apply the Gaussian focal loss [24] to the initialized queries of both modalities, given by + +$$ +\mathcal {L} _ {\text {i n i t}} = \mathcal {L} _ {G F o c a l} (\hat {\mathbf {Y}} _ {L}, \mathbf {Y} _ {L}) + \mathcal {L} _ {G F o c a l} (\hat {\mathbf {Y}} _ {C}, \mathbf {Y} _ {C}), \tag {3} +$$ + +where $\hat{\mathbf{Y}}_L, \hat{\mathbf{Y}}_C$ are the dense predictions of category-wise heatmaps of the LiDAR and camera modalities, respectively, and $\mathbf{Y}_L, \mathbf{Y}_C$ are the corresponding ground-truths. + +Then, we apply the loss function for the detectors of the LiDAR and camera modalities, as well as the view transfor + +mation of the camera candidates and the candidate fusion stage. Firstly, the predictions of each modality-specific detector are independently matched with the ground-truth using the Hungary algorithm [16]. The object classification is optimized with focal loss [24] and the 3D bounding box regression is optimized with L1 loss. For the camera modality detector, the ground-truth bounding boxes are in separate camera coordinates. For all other detectors, ground-truth bounding boxes are in LiDAR coordinates. The detection loss can be represented as + +$$ +\mathcal {L} _ {\text {d e t e c t}} = \gamma \mathcal {L} _ {\text {d e t e c t}} ^ {\text {c a m e r a}} + \mathcal {L} _ {\text {d e t e c t}} ^ {\text {t r a n s}} + \mathcal {L} _ {\text {d e t e c t}} ^ {\text {L i D A R}} + \mathcal {L} _ {\text {d e t e c t}} ^ {\text {f u s i o n}}. \tag {4} +$$ + +Our entire network is optimized using $\mathcal{L} = \alpha \mathcal{L}_{init} + \beta \mathcal{L}_{detect}$ . In our implementation, we empirically set $\gamma = 1, \alpha = 0.1$ , and $\beta = 1$ to balance different terms. + +# 4. Experiments + +# 4.1. Dataset and metrics + +We follow previous work [47, 45, 22, 5] to evaluate our method on the nuScenes dataset [2]. It is a challenging dataset for 3D object detection, consisting of 700/150/150 scenes for training/validation/test. It provides point clouds collected using a 32-beam LiDAR and six images from multi-view cameras. There are 1.4 million annotated 3D bounding boxes for objects from 10 different classes. We evaluate performance using the nuScenes detection score (NDS) and mean average precision (mAP) metrics. The final mAP is averaged over distance thresholds $0.5\mathrm{m}$ , $1\mathrm{m}$ , $2\mathrm{m}$ , and $4\mathrm{m}$ on the BEV across 10 classes. NDS is a weighted average of mAP and other true positive metrics including mATE, mASE, mAOE, mAVE, and mAAE. + +# 4.2. Implementation Details + +Our implementation is based on the MMDetection3D framework [6]. For the camera branch, we use ResNet-50 [11] as the backbone and initialize it with the Mask R-CNN [10] instance segmentation network pretrained on nuImage [2]. The input image resolution is $800 \times 448$ . For the LiDAR branch, we apply VoxelNet [51] with voxel size $(0.075m, 0.075m, 0.2m)$ . The detection range is set as $[-54m, 54m]$ for the XY-axes and $[-5m, 3m]$ for the Z-axis. Our LiDAR and camera modality detectors both include only 1 decoder layer. The query numbers for the LiDAR and camera modalities are set as $N_L = N_C = 200$ , so our fusion stage can detect at most 400 objects per scene. + +Since our framework disentangles the camera detector and the LiDAR detector, we can conveniently apply data augmentation separately to the LiDAR inputs and camera inputs. We apply random rotation, scaling, translation, and flipping to the LiDAR inputs, and we apply random scaling and horizontal flipping to the camera inputs. Our training pipeline follows previous works [1, 28, 47]. We first train + +TransFusion-L [1] as our LiDAR-only baseline, which is used to initialize our LiDAR backbone and LiDAR modality detector. This LiDAR-only baseline is trained for 20 epochs. Afterward, we freeze the pretrained LiDAR components and train the entire fusion framework for 6 epochs. For both training stages, we use the AdamW optimizer [29] with one-cycle learning rate policy [37]. The initial learning rate is $10^{-4}$ and the weight decay is $10^{-2}$ . The hidden dimensions in the entire model except the backbones are 128. For both training stages, we adopt CBGS [52] to balance the class distribution. We train our method on four NVIDIA A6000 GPUs with batch size 16. + +# 4.3. Results and Comparison + +We report results on the nuScenes validation and test sets, without using any test-time augmentations or model ensembles. As shown in Tab. 2, SparseFusion significantly improves our LiDAR-only baseline, TransFusion-L [1], by $+3.6\%$ NDS and $+6.3\%$ mAP on the test set, due to the additional use of camera inputs. More importantly, SparseFusion sets a new state of the art on both the validation set and test set, outperforming prior works including those using stronger backbones. It is noteworthy that SparseFusion demonstrates a $0.4\%$ NDS and $1.0\%$ mAP improvement over the most recent state-of-the-art [47] while also achieving a 1.8x speedup (5.6 FPS vs. 3.1 FPS) on A6000 GPU. It can be seen in Fig. 1a that, in addition to superior performance, SparseFusion also provides the fastest inference speed. We also demonstrate the performance of SparseFusion by visualizing some qualitative results in Fig. 4. + +# 4.4. Analysis + +Performance Breakdown SparseFusion performs 3D object detection in both the LiDAR and camera branch separately. Tab. 3 shows the detection performance in different parts of SparseFusion including the LiDAR branch, camera branch (before and after view transformation), and the fusion branch. We notice that our LiDAR branch detection results notably surpass the LiDAR-only baseline TransFusion-L [1] since the proposed semantic transfer can compensate for the weakness of point cloud inputs. In comparison with the state-of-the-art single-frame camera detector PETR [26] (six decoder layers), our camera branch achieves much better performance with just one decoder layer, owing to the depth-aware features from the proposed geometric transfer. Besides, our view transformation module not only transforms the instance features from the camera coordinate space into the LiDAR coordinate space, but it also slightly improves the camera branch detection performance by aggregating multi-view information. With this strong performance, the modality-specific detectors in each branch would not cause negative transfer during fusion. We also consider direct Non-Maximum Suppression + +Table 2: Comparison with existing methods on nuScenes validation set and test set. + +
MethodsModalityLiDAR BackboneCamera Backbonevalidation settest set
NDSmAPNDSmAP
FCOS3D [42]Camera-ResNet-101 [11]41.534.342.835.8
PETR [26]Camera-ResNet-101 [11]44.237.045.539.1
CenterPoint [48]LiDARVoxelNet [51]-66.859.667.360.3
TransFusion-L [1]LiDARVoxelNet [51]-70.165.170.265.5
PointAugmenting [39]LiDAR+CameraVoxelNet [51]DLA34 [49]--71.066.8
FUTR3D [4]LiDAR+CameraVoxelNet [51]ResNet-101 [11]68.364.5--
UVTR [18]LiDAR+CameraVoxelNet [51]ResNet-101 [11]70.265.471.167.1
TransFusion [1]LiDAR+CameraVoxelNet [51]ResNet-50 [11]71.367.571.668.9
AutoAlignV2 [5]LiDAR+CameraVoxelNet [51]CSPNet [41]71.267.172.468.4
BEVFusion [22]LiDAR+CameraVoxelNet [51]Dual-Swin-T [21]72.169.673.371.3
BEVFusion [28]LiDAR+CameraVoxelNet [51]Swin-T [27]71.468.572.970.2
DeepInteraction [47]LiDAR+CameraVoxelNet [51]ResNet-50 [11]72.669.973.470.8
CMT [45]LiDAR+CameraVoxelNet [51]ResNet-50 [11]70.867.9--
SparseFusion (ours)LiDAR+CameraVoxelNet [51]ResNet-50 [11]72.870.473.872.0
+ +![](images/93a2f08e907b7768b6021c67114553ff0af23b2938c90c333e07bc4aad988d21.jpg) +Figure 4: Qualitative results of SparseFusion on nuScenes validation set. + +(NMS) to combine the detection results of LiDAR and camera branches, which is even worse than the single LiDAR branch. Therefore, it is also important to make the instance-level features from the two branches interact with each other in the fusion stage. + +Strong Image Backbone. We incorporate stronger SwinT [27] backbone into SparseFusion to match some previous work [22, 45]. Tab. 4 compares the performance of different methods on the nuScenes validation set. We do not include any test-time augmentations or model ensembles. Although multi-modality detection relies more on the LiDAR inputs, SparseFusion can still benefit from a stronger image backbone and beat all the counterparts. + +Modality-Specific Object Recall. The parallel detectors in the LiDAR and camera branches enable us to determine which modality recalls each object. Given $N_{L} + N_{C}$ predictions, we know that the first $N_{L}$ instances come from the LiDAR-modality detector, while the last $N_{C}$ are from + +Table 3: Performance breakdown. We show the detection results in different parts of SparseFusion on the validation set. 'L' and 'C' refer to the LiDAR and camera modalities, respectively. 'ST' and 'GT' refer to semantic transfer and geometric transfer, respectively. 'VT' is the view transformation module. + +
MethodsModalityNDSmAP
TransFusion-L [1]L70.165.1
LiDAR branchL + ST71.868.4
PETR (ResNet50) [26]C38.131.3
Camera branch (before VT)C+GT43.540.6
Camera branch (after VT)C+GT44.341.5
LiDAR/camera branches NMSL+C71.568.2
SparseFusionL+C72.870.4
+ +Table 4: Results on nuScenes validation set with stronger image backbones. We do not use any test-time augmentations or model ensembles. The inference speed is measured on a single NVIDIA A6000 GPU. $\dagger$ : CMT [45] adopts flash-attention [7] for transformer acceleration. + +
MethodsImage BackboneInput ResolutionNDSmAPFPS
BEVFusion [22]Dual-Swin-T1600×90072.169.60.8
CMT [45]VoVNet-991600×64072.970.33.8†
SparseFusionResNet-50800×44872.870.45.6
SparseFusionSwin-T800×44873.171.05.3
+ +![](images/63be14ae83bf2400e368d6b8b4edac6adbd2f40697cbef7e7152bf986d80f0c5.jpg) +(a) Objects of different classes. (b) Objects at different distances. + +![](images/066a58f9268af594eed9cebf12dc191e3545a0ea764e3657c0750c4a9385e517.jpg) +Figure 6: Instance-level feature interaction in the fusion stage. Orange boxes are detected after the fusion stage with high confidence in the BEV space. Blue and green dots denote all instances from the LiDAR and camera branches separately. Orange boxes are connected with blue/green dots with blue/green lines. The attention strength is represented by the darkness and thickness of the lines. + +the camera modality. An object is recalled if a bounding box with correct classification is predicted within a radius of two meters around its ground-truth position. In Fig. 5, we demonstrate the percentage of objects in the nuScenes validation set recalled by exactly one modality. We observe that each modality can compensate for the weakness of the other to some extent. Although the LiDAR modality is typically more powerful, the camera modality plays an important role in detecting objects from classes such as construction vehicles and barriers. Furthermore, the camera modality is useful for detecting objects at far distances where point clouds are sparse or the LiDAR feature map cannot cover. + +Cross-Modality Sparse Representation Interaction. Fig. 6 visualizes the instance feature interaction in the sparse fusion stage. The strength of attention between instance-level features is reflected by the thickness and darkness of lines. We notice that most objects can aggregate multi-modality instance-level features during fusion. Although the strongest interactions exist mainly between neighboring instances, it is interesting that features from the camera modality are also able to deliver strong interactions with the instances at the distant range. This could be a result of the shared semantics among objects in images. + +Robustness to Sensor Malfunction. With the relatively independent modality-specific detectors, SparseFusion is robust to sensor malfunctions. We evaluate SparseFusion in two different cases: (1) The LiDAR scan range is limited to $(-90^{\circ}, 90^{\circ})$ and (2) The most important front camera is + +![](images/bc8a60f387e60e5d4eb5d09994a5ee653c979f2431056354735e865e4590674c.jpg) +Figure 5: Percentage of objects recalled by each modality. We do not count those objects recalled by both or neither. + +Table 5: Results on nuScenes validation set for LiDAR or camera sensor malfunctions. The models are finetuned on the in-domain data in both cases. + +
MethodsSensor MalfunctionNDSmAP
TransFusion [1]Limited LiDAR Field (-90°, 90°)49.231.1
BEVFusion [22]54.945.5
SparseFusion61.254.3
TransFusion [1]Missing70.165.3
BEVFusion [22]Front Camera70.765.9
SparseFusion72.169.2
+ +missing while other cameras are preserved. SparseFusion outperforms previous methods [1, 22] with a large gap under both scenarios, demonstrating great robustness to sensor malfunctions. + +# 4.5. Ablation Studies + +In this section, we study the effect of using alternatives for the different modules in SparseFusion. For our ablation studies, we train on a 1/5 split of nuScenes training set and evaluate on the full nuScenes validation set. + +Sparse Fusion Strategy. We compare our self-attention module for sparse feature fusion with other fusion methods. + +Table 6: Ablation study experiments to justify our design of each module. +(a) Sparse fusion strategies. + +
Fusion StrategyNDSmAP
MLP67.864.5
Cross-Attention68.665.8
Optimal Transport68.065.6
Self-Attention68.866.4
+ +(b) Information transfers. + +
GeometricSemanticNDSmAP
✓✗67.764.2
✗✓68.465.7
✓✓68.866.4
+ +(c) Parallel modality-specific detectors. + +
Seq. Pos.Seq. Feat.NDSmAP
✓X68.165.3
✓✓67.964.2
XX68.866.4
+ +In addition to our self-attention module in Sec. 3.1, three alternatives are considered: 1) Instance-level candidates from the two branches are directly fed into an MLP without any cross-instance aggregation. 2) LiDAR instance candidates are used as queries and camera instance candidates are used as keys/values. A cross-attention module is then used to fuse multi-modality instance features. 3) We make novel use of optimal transport in LiDAR-camera fusion [33]. We propose to learn a distribution transformation from camera candidates to LiDAR candidates through optimal transport. Then, we can directly fuse them by concatenating the candidates of two branches along the channel dimension. More details about this method are provided in appendix. + +The results in Tab. 6a show that only cross-attention achieves competitive performance to self-attention. Yet, we observe that it relies so heavily on the output of the LiDAR branch that the camera branch is not fully utilized to compensate for the weaknesses of the LiDAR branch. The MLP strategy has limited performance as it does not fuse cross-instance and cross-modality information. Despite the impressive progress of optimal transport in other fields, it fails to learn the correspondences between the instance features of the two modalities and thus has limited performance. In contrast, self-attention is simple, efficient, and effective. + +Information Transfer. We then ablate the geometric and semantic transfer between the LiDAR and camera modalities. The results in Tab. 6b show that the fusion performance benefits from both transfers. This also validates that the disadvantages of both modalities result in negative transfer and that our proposed information transfer modules are indeed effective in mitigating this issue. The semantic transfer module especially improves the final performance since it compensates for the LiDAR modality's lack of semantic information, which is critical for 3D detection. + +Despite the importance of the information transfer, SparseFusion trained on the full training set can reach 71.5 NDS on the validation set even if we remove both the semantic and geometric transfers. This still outperforms many other advanced algorithms [1, 28], reflecting the robustness of our fusion stage to the modality-specific weakness. + +View Transformation. As explained in Sec. 3.1, we transform the sparse representations of both modalities into one + +unified space. To validate the effectiveness of this approach, we ablate the view transformation of camera candidates into the LiDAR coordinate space. This results in a more straightforward method where we simply obtain the predictions of two modalities and directly fuse them using self-attention. Ablating the view transformation drops performance from $66.4\%$ mAP and $68.8\%$ NDS to $65.6\%$ mAP and $68.3\%$ NDS, respectively. This demonstrates that the view transformation is indeed helpful to overall performance. + +Parallel Detectors. In addition to the ablation studies on the modules of the proposed pipeline, we also study alternatives to the structure of the pipeline. SparseFusion uses separate 3D object detectors for the LiDAR and camera modalities in parallel for extracting instance-level features to address the cross-modality domain gap. Alternatively, we consider a sequential pipeline where the camera detector runs after the LiDAR detector. The camera-modality detector inherits the output queries from the LiDAR detector. We consider two variants of this inheritance: 1) using the 3D position and instance features for the initial query of the camera modality; 2) using the 3D position but initializing the instance features from the corresponding image features. The camera detector follows the structure of the PETR [26] decoder (one layer). Tab. 6c shows that both sequential structures yield notably inferior performance, justifying our use of parallel modality-specific detectors. + +# 5. Conclusion + +We revisit previous LiDAR-camera fusion works and propose SparseFusion, a novel 3D object detection method that utilizes the rarely-explored strategy of fusing sparse representations. SparseFusion extracts instance-level features from each modality separately via parallel 3D object detectors and then treats the instance-level features as the modality-specific candidates. Afterward, we transform the candidates into a unified 3D space, and we are able to fuse the candidates with a lightweight attention module. Extensive experiments demonstrate that SparseFusion achieves state-of-the-art performance on the nuScenes benchmark with the fastest inference speed. We hope SparseFusion will serve as a powerful and efficient baseline for further research into this field. + +# References + +[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 2, 3, 4, 6, 7, 8, 9 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3, 6 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3 +[4] Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. arXiv preprint arXiv:2203.10642, 2022. 2, 3, 7 +[5] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Autoalignv2: Deformable feature aggregation for dynamic multi-modal 3d object detection. arXiv preprint arXiv:2207.10316, 2022. 1, 2, 3, 6, 7 +[6] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mm Detection3d, 2020.6 +[7] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In Advances in Neural Information Processing Systems, 2022. 8 +[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021. 3 +[9] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2918-2927, 2021. 3 +[10] Kaiming He, Georgia Gkioxari, Piotr Dálár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7 +[12] Chenxi Huang, Tong He, Haidong Ren, Wenxiao Wang, Binbin Lin, and Deng Cai. Obmo: One bounding box multiple objects for monocular 3d object detection. arXiv preprint arXiv:2212.10049, 2022. 3 + +[13] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 3 +[14] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3 +[15] Yang Jiao, Zequn Jie, Shaoxiang Chen, Jingjing Chen, Xiaolin Wei, Lin Ma, and Yu-Gang Jiang. Msmdfusion: Fusing lidar and camera at multiple scales with multi-depth seeds for 3d object detection. arXiv preprint arXiv:2209.03102, 2022. 2 +[16] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6 +[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3 +[18] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022. 2, 3, 7 +[19] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 3, 4 +[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 1-18. Springer, 2022. 3 +[21] Tingting Liang, Xiaojie Chu, Yudong Liu, Yongtao Wang, Zhi Tang, Wei Chu, Jingdong Chen, and Haibin Ling. Cbnet: A composite backbone network architecture for object detection. IEEE Transactions on Image Processing, 31:6893-6906, 2022. 1, 7 +[22] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3, 6, 7, 8 +[23] Zhidong Liang, Ming Zhang, Zehan Zhang, Xian Zhao, and Shiliang Pu. Rangercnn: Towards fast and accurate 3d object detection with range image representation. arXiv preprint arXiv:2009.00206, 2020. 3 +[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5, 6 +[25] Hao Liu, Zhuoran Xu, Dan Wang, Baofeng Zhang, Guan Wang, Bo Dong, Xin Wen, and Xinyu Xu. Pai3d: Painting adaptive instance-prior for 3d object detection. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 459-475. Springer, 2023. 2 + +[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 531-548. Springer, 2022. 3, 6, 7, 9 +[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 7 +[28] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 1, 2, 3, 6, 7, 9 +[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[30] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3164-3173, 2021. 3 +[31] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3 +[32] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 3 +[33] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355-607, 2019. 9 +[34] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3 +[35] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 2, 3 +[36] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2743-2752, 2021. 3 +[37] Leslie N Smith. Cyclical learning rates for training neural networks. In 2017 IEEE winter conference on applications of computer vision (WACV), pages 464-472. IEEE, 2017. 6 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 + +[39] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 2, 3, 7 +[40] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 2, 3 +[41] Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh, and I-Hau Yeh. Cspnet: A new backbone that can enhance learning capability of cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 390–391, 2020. 7 +[42] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 3, 7 +[43] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 3 +[44] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In European Conference on Computer Vision, pages 638-656. Springer, 2022. 2 +[45] Junjie Yan, Yingfei Liu, Jianjian Sun, Fan Jia, Shuai Lin Li, Tiancai Wang, and Xiangyu Zhang. Cross modal transformer via coordinates encoding for 3d object detection. arXiv preprint arXiv:2301.01283, 2023. 2, 3, 6, 7, 8 +[46] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 3 +[47] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. arXiv preprint arXiv:2208.11112, 2022. 2, 3, 6, 7 +[48] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 3, 7 +[49] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 7 +[50] Xingyi Zhou, Dequan Wang, and Philipp Krahenbihl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 3 +[51] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 3, 6, 7 +[52] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point + +cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019.6 +[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 4, 5 \ No newline at end of file diff --git a/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/images.zip b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..99a2ab0dc579727b69774f5dd7781b45687a9adb --- /dev/null +++ b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12bfb959f2489fb81d7c745664f09a9733152e9d1b49b60426f44d842733caff +size 636469 diff --git a/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/layout.json b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a6e692eb7edded8b901dd134b2eda73f75c2106b --- /dev/null +++ b/2023/SparseFusion_ Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection/layout.json @@ -0,0 +1,8262 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 136, + 467, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 136, + 467, + 170 + ], + "spans": [ + { + "bbox": [ + 110, + 136, + 467, + 170 + ], + "type": "text", + "content": "SparseFusion: Fusing Multi-Modal Sparse Representations for Multi-Sensor 3D Object Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "text", + "content": "Yichen Xie" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "text", + "content": ", Chenfeng Xu" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "text", + "content": ", Marie-Julie Rakotosaona" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "text", + "content": ", Patrick Rim" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "text", + "content": ", Federico Tombari" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "text", + "content": ", Kurt Keutzer" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "text", + "content": ", Masayoshi Tomizuka" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "text", + "content": ", Wei Zhan" + }, + { + "bbox": [ + 71, + 191, + 506, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 219, + 479, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 219, + 479, + 234 + ], + "spans": [ + { + "bbox": [ + 96, + 219, + 479, + 234 + ], + "type": "text", + "content": "1 University of California, Berkeley 2 Google 3 California Institute of Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 139, + 271, + 186, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 271, + 186, + 284 + ], + "spans": [ + { + "bbox": [ + 139, + 271, + 186, + 284 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 297, + 282, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 297, + 282, + 599 + ], + "spans": [ + { + "bbox": [ + 45, + 297, + 282, + 599 + ], + "type": "text", + "content": "By identifying four important components of existing LiDAR-camera 3D object detection methods (LiDAR and camera candidates, transformation, and fusion outputs), we observe that all existing methods either find dense candidates or yield dense representations of scenes. However, given that objects occupy only a small part of a scene, finding dense candidates and generating dense representations is noisy and inefficient. We propose SparseFusion, a novel multi-sensor 3D detection method that exclusively uses sparse candidates and sparse representations. Specifically, SparseFusion utilizes the outputs of parallel detectors in the LiDAR and camera modalities as sparse candidates for fusion. We transform the camera candidates into the LiDAR coordinate space by disentangling the object representations. Then, we can fuse the multi-modality candidates in a unified 3D space by a lightweight self-attention module. To mitigate negative transfer between modalities, we propose novel semantic and geometric cross-modality transfer modules that are applied prior to the modality-specific detectors. SparseFusion achieves state-of-the-art performance on the nuScenes benchmark while also running at the fastest speed, even outperforming methods with stronger backbones. We perform extensive experiments to demonstrate the effectiveness and efficiency of our modules and overall method pipeline. Our code will be made publicly available at https://github.com/yichen928/SparseFusion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 621, + 124, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 124, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 124, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 641, + 279, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 641, + 279, + 712 + ], + "spans": [ + { + "bbox": [ + 45, + 641, + 279, + 712 + ], + "type": "text", + "content": "Autonomous driving cars rely on multiple sensors, such as LiDAR and cameras, to perceive the surrounding environment. LiDAR sensors provide accurate 3D scene occupancy information through point clouds with points in the xyz coordinate space, and cameras provide rich semantic information through images with pixels in the RGB" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 299, + 260, + 411, + 341 + ], + "blocks": [ + { + "bbox": [ + 299, + 260, + 411, + 341 + ], + "lines": [ + { + "bbox": [ + 299, + 260, + 411, + 341 + ], + "spans": [ + { + "bbox": [ + 299, + 260, + 411, + 341 + ], + "type": "image", + "image_path": "031d89a9df86fe6b84eb009a319a73d1bff0e654b664a6711427f4fe7fc7f7e1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 346, + 385, + 357 + ], + "lines": [ + { + "bbox": [ + 324, + 346, + 385, + 357 + ], + "spans": [ + { + "bbox": [ + 324, + 346, + 385, + 357 + ], + "type": "text", + "content": "(a) mAP vs. FPS" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 417, + 267, + 530, + 334 + ], + "blocks": [ + { + "bbox": [ + 417, + 267, + 530, + 334 + ], + "lines": [ + { + "bbox": [ + 417, + 267, + 530, + 334 + ], + "spans": [ + { + "bbox": [ + 417, + 267, + 530, + 334 + ], + "type": "image", + "image_path": "d77b38623dc1b1833de9b991a0d9aaa99ceda8b5b19ab1f59546eb55980df172.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 298, + 363, + 530, + 441 + ], + "blocks": [ + { + "bbox": [ + 427, + 346, + 521, + 357 + ], + "lines": [ + { + "bbox": [ + 427, + 346, + 521, + 357 + ], + "spans": [ + { + "bbox": [ + 427, + 346, + 521, + 357 + ], + "type": "text", + "content": "(b) Dense-to-dense fusion." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 298, + 363, + 530, + 441 + ], + "lines": [ + { + "bbox": [ + 298, + 363, + 530, + 441 + ], + "spans": [ + { + "bbox": [ + 298, + 363, + 530, + 441 + ], + "type": "image", + "image_path": "1690ec947523fe1b553f01f68c5e7b7c1e3c40c807796ab35ac643998afee96f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 445, + 530, + 477 + ], + "lines": [ + { + "bbox": [ + 297, + 445, + 530, + 477 + ], + "spans": [ + { + "bbox": [ + 297, + 445, + 530, + 477 + ], + "type": "text", + "content": "(c) Overview of our sparse fusion strategy. We extract instance-level features from the LiDAR and camera modalities separately, and fuse them in a unified 3D space to perform detection." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 296, + 486, + 531, + 569 + ], + "lines": [ + { + "bbox": [ + 296, + 486, + 531, + 569 + ], + "spans": [ + { + "bbox": [ + 296, + 486, + 531, + 569 + ], + "type": "text", + "content": "Figure 1: Compared to existing fusion algorithms, Sparse-Fusion achieves state-of-the-art performance as well as the fastest inference speed on nuScenes test set. † : Official code of [5] uses flip as test-time augmentation. ‡ : We use BEVFusion-base results in the official repository of [28] to match the input resolutions of other methods. § : Swin-T [27, 21] is adopted as image backbone." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 589, + 531, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 589, + 531, + 705 + ], + "spans": [ + { + "bbox": [ + 296, + 589, + 531, + 705 + ], + "type": "text", + "content": "color space. However, there are often significant discrepancies between representations of the same physical scene acquired by the two sensors, as LiDAR sensors capture point clouds using 360-degree rotation while cameras capture images from a perspective view without a sense of depth. This impedes an effective and efficient fusion of the LiDAR and camera modalities. To tackle this challenge, multi-sensor fusion algorithms were proposed to find correspondences between multi-modality data to transform and fuse them into a unified scene representation space." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 297, + 706, + 530, + 730 + ], + "type": "text", + "content": "Dense representations, such as bird-eye-view (BEV), volumetric, and point representations, are commonly used" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 41, + 470, + 52 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 41, + 470, + 52 + ], + "spans": [ + { + "bbox": [ + 151, + 41, + 470, + 52 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 53, + 422, + 63 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 53, + 422, + 63 + ], + "spans": [ + { + "bbox": [ + 198, + 53, + 422, + 63 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 63, + 440, + 74 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 63, + 440, + 74 + ], + "spans": [ + { + "bbox": [ + 181, + 63, + 440, + 74 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 719, + 157, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 719, + 157, + 729 + ], + "spans": [ + { + "bbox": [ + 59, + 719, + 157, + 729 + ], + "type": "text", + "content": "* indicates equal contribution." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "17591" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 229, + 518, + 374 + ], + "blocks": [ + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "lines": [ + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "spans": [ + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "text", + "content": "Table 1: For each LiDAR-camera fusion method, we identify the LiDAR candidates and camera candidates that are used, the transformation process used to fuse these candidates into a unified space, and the fusion outputs generated to represent 3D scenes using information from both modalities. Based on these components, we categorize the methods into the following categories: Dense+Sparse " + }, + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "text", + "content": " Dense approaches relate the sparse region proposals in images to a dense frustum point cloud and fuse them into a dense point space. Dense+Dense " + }, + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "text", + "content": " Dense approaches align each point feature in the point cloud to the corresponding pixel feature in the image and represent the 3D scenes using dense features such as point/BEV features. Sparse+Dense " + }, + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "text", + "content": " Sparse approaches generate sparse queries by detecting instance features in the point cloud and then applying cross-attention with dense image features. Dense+Dense " + }, + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "text", + "content": " Sparse approaches predict objects using queries that combine dense features from each modality. Sparse+Sparse " + }, + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 45, + 104, + 532, + 223 + ], + "type": "text", + "content": " Sparse (ours) extracts sparse instance features from each modality and directly fuses them to obtain the final sparse instance features used for detection." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 229, + 518, + 374 + ], + "lines": [ + { + "bbox": [ + 60, + 229, + 518, + 374 + ], + "spans": [ + { + "bbox": [ + 60, + 229, + 518, + 374 + ], + "type": "table", + "html": "
CategoryMethodLiDAR CandidateCamera CandidateTransformationFusion Outputs
Dense+Sparse→DenseFrustum PointNets [35]point featuresregion proposalsproj. & concat.point features
Dense+Dense→DensePointPainting [39]point featuressegm. outputproj. & concat.point features
PointAugmenting [40]point featuresimage featuresproj. & concat.point features
PAI3D [25]point featuressegm. outputproj. & concat.point features
BEVFusion [28]BEV featuresimage featuresdepth. est. & proj. & concat.BEV features
AutoAlignV2 [5]voxel featuresimage featuresproj. & attn.voxel features
Sparse+Dense→SparseTransFusion [1]instance featuresimage featuresproj. & attn.instance features
Dense+Dense→SparseFUTR3D [4]voxel featuresimage featuresattn.instance features
UVTR [18]voxel featuresimage featuresdepth. est. & proj. & attn.instance features
DeepInteraction [47]BEV featuresimage featuresproj. & attn.instance features
CMT [45]BEV featuresimage featuresattn.instance features
Sparse+Sparse→SparseSparseFusion (Ours)instance featuresinstance featuresproj. & attn.instance features
", + "image_path": "4b4da354a43dce7df6b1897f6d84c2d2e4ad37fcfd9c6dcd2ca296a40cc622b5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 392, + 279, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 392, + 279, + 532 + ], + "spans": [ + { + "bbox": [ + 45, + 392, + 279, + 532 + ], + "type": "text", + "content": "to represent 3D scenes [28, 5, 39, 25, 40, 15]. Most previous works fuse different modalities by aligning low-level data or high-level features to yield dense features that describe the entire 3D scene, e.g., as shown in Fig 1b. However, for the task of 3D object detection, such dense representations are superfluous since we are only interested in instances/objects, which only occupy a small part of the 3D space. Furthermore, noisy backgrounds can be detrimental to object detection performance, and aligning different modalities into the same space is a time-consuming process. For example, generating BEV features from multi-view images takes " + }, + { + "bbox": [ + 45, + 392, + 279, + 532 + ], + "type": "inline_equation", + "content": "500\\mathrm{ms}" + }, + { + "bbox": [ + 45, + 392, + 279, + 532 + ], + "type": "text", + "content": " on an RTX 3090 GPU [28]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 532, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 532, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 532, + 280, + 730 + ], + "type": "text", + "content": "In contrast, sparse representations are more efficient, and methods based on them have achieved state-of-the-art performance in multi-sensor 3D detection [1, 4, 18, 47]. These methods use object queries to represent instances/objects in the scene and interact with the original image and point cloud features. However, most previous works do not take into account the significant domain gap between features from different modalities [44]. The queries may gather information from one modality that has a large distribution shift with respect to another modality, making iterative interaction between modalities with large gaps suboptimal. Recent work [47] mitigates this issue by incorporating modality interaction, i.e. performing cross-attention between features from two different modalities. However, the number of computations performed in this method increases quadratically with the dimensions of features and is thus inefficient. We categorize previous works into four" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 296, + 392, + 530, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 392, + 530, + 427 + ], + "spans": [ + { + "bbox": [ + 296, + 392, + 530, + 427 + ], + "type": "text", + "content": "groups by identifying four key components, which are outlined in Table 1. Further discussion of the methods in these groups is presented in Sec. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 427, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 427, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 427, + 531, + 730 + ], + "type": "text", + "content": "In this paper, we propose SparseFusion, a novel method (Fig. 1c) that simultaneously utilizes sparse candidates and yields sparse representations, enabling efficient and effective 3D object detection. SparseFusion is the first LiDAR-camera fusion method, to our knowledge, to perform 3D detection using exclusively sparse candidates and sparse fusion outputs. We highlight a key common ground between the two modalities: an image and a point cloud that represent the same 3D scene will contain mostly the same instances/objects. To leverage this commonality, we perform 3D object detection on the inputs from each modality in two parallel branches. Then, the instance features from each branch are projected into a unified 3D space. Since the instance-level features are sparse representations of the same objects in the same scene, we are able to fuse them with a lightweight attention module [38] in a soft manner. This parallel detection strategy allows the LiDAR and camera branches to take advantage of the unique strengths of the point cloud and image representations, respectively. Nevertheless, the drawbacks of each single-modality detector may result in negative transfer during the fusion phase. For example, the point cloud detector may struggle to distinguish between a standing person and a tree trunk due to a lack of detailed semantic information, while the image detector is hard to localize objects in the 3D space due to a lack of accurate depth information. To mitigate the issue of negative" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "17592" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 152 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 152 + ], + "type": "text", + "content": "transfer, we introduce a novel cross-modality information transfer method designed to compensate for the deficiencies of each modality. This method is applied to the inputs from both modalities prior to the parallel detection branches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 153, + 280, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 153, + 280, + 234 + ], + "spans": [ + { + "bbox": [ + 45, + 153, + 280, + 234 + ], + "type": "text", + "content": "SparseFusion achieves state-of-the-art results on the competitive nuScenes benchmark [2]. Our instance-level sparse fusion strategy allows for a lighter network and much higher efficiency in comparison with prior work. With the same backbone, SparseFusion outperforms the current state-of-the-art model [47] with 1.8x acceleration. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 237, + 279, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 237, + 279, + 295 + ], + "spans": [ + { + "bbox": [ + 57, + 237, + 279, + 295 + ], + "type": "text", + "content": "- We revisit prior LiDAR-camera fusion works and identify four important components that allow us to categorize existing methods into four groups. We propose an entirely new category of methods that exclusively uses sparse candidates and representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 299, + 279, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 299, + 279, + 357 + ], + "spans": [ + { + "bbox": [ + 57, + 299, + 279, + 357 + ], + "type": "text", + "content": "- We propose SparseFusion, a novel method for LiDAR-camera 3D object detection that leverages instance-level sparse feature fusion and cross-modality information transfer to take advantage of the strengths of each modality while mitigating their weaknesses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 361, + 279, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 361, + 279, + 407 + ], + "spans": [ + { + "bbox": [ + 57, + 361, + 279, + 407 + ], + "type": "text", + "content": "- We demonstrate that our method achieves state-of-the-art performance in 3D object detection with a lightweight architecture that provides the fastest inference speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 418, + 130, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 130, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 130, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 438, + 279, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 438, + 279, + 623 + ], + "spans": [ + { + "bbox": [ + 45, + 438, + 279, + 623 + ], + "type": "text", + "content": "LiDAR-based 3D Object Detection. LiDAR sensors are commonly used for single-modality 3D object detection due to the accurate geometric information provided by point clouds. For detection in outdoor scenes, most existing methods transform unordered point clouds into more structured data formats such as pillars [17], voxels [46, 8], or range views [9, 23]. Features are extracted by standard 2D or 3D convolutional networks, based on which a detection head is used to recognize objects and regress 3D bounding boxes. Mainstream detection heads apply anchor-based [51, 17] or center-based [50] structures. Inspired by the promising performance of transformer-based methods in 2D detection, some recent works explore transformers as feature extractors [30, 36] or as detection heads [1, 31]. Our method is agnostic to the LiDAR-based detector used in the LiDAR branch, and the default setting uses TransFusion-L [1]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 624, + 279, + 729 + ], + "type": "text", + "content": "Camera-based 3D Object Detection. Camera-based 3D detection methods are also being studied increasingly. Early work performs monocular 3D object detection by attaching extra 3D bounding box regression heads [48, 42] to 2D detectors. In practice, scenes are often perceived by multiple cameras from different perspective views. Following LSS [34], methods like BEVDet [14] and BEVDepth [19] extract 2D features from multi-view images and project them into the BEV space. Other methods including" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 106, + 530, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 234 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 234 + ], + "type": "text", + "content": "DETR3D [43] and PETR [26] adapt techniques from transformer-based 2D object detection methods [53, 3] to learn correspondences between different perspective views through cross-attention using 3D queries. However, as revealed in [12], there inevitably exists some ambiguity when recovering 3D geometry from 2D images. In response, recent works [13, 20, 32] also explore the positive effects of temporal cues in camera-based 3D detection. In our proposed SparseFusion, we extend deformable-DETR [53] to monocular 3D object detection and explicitly transform the regressed bounding boxes to the LiDAR coordinate space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 235, + 530, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 235, + 530, + 491 + ], + "spans": [ + { + "bbox": [ + 296, + 235, + 530, + 491 + ], + "type": "text", + "content": "Multi-Modality 3D Object Detection. LiDAR and cameras provide complementary information about the surrounding environment, so it is appealing to fuse the multimodality inputs for 3D object detection tasks. As analyzed in Tab. 1, existing fusion methods can be classified into four categories. Early works tend to fuse multi-modality information into a unified dense representation. Frustum Point-Nets [35] utilizes a Dense+Sparse " + }, + { + "bbox": [ + 296, + 235, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 296, + 235, + 530, + 491 + ], + "type": "text", + "content": " Dense approach that filters dense point clouds with sparse 2D regions of interest. Subsequent works explore Dense+Dense " + }, + { + "bbox": [ + 296, + 235, + 530, + 491 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 296, + 235, + 530, + 491 + ], + "type": "text", + "content": " Dense approaches by working directly with the dense LiDAR modality and camera modality features instead. Methods such as [39, 40] project point clouds into image perspective views and concatenate the dense image features with point features. BEVFusion [28, 22] significantly improves the performance of this line of methods by projecting dense image features into the LiDAR coordinate space using estimated per-pixel depths. AutoAlignV2 [5] also considers the soft correspondence through cross-modality attention to increase the robustness. However, we point out that dense representations are altogether undesirable for 3D object detection as they are noisy and inefficient." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 495, + 530, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 495, + 530, + 692 + ], + "spans": [ + { + "bbox": [ + 296, + 495, + 530, + 692 + ], + "type": "text", + "content": "Recent works have begun to utilize object-centric sparse scene representations. TransFusion [1] adopts a Sparse+Dense " + }, + { + "bbox": [ + 296, + 495, + 530, + 692 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 296, + 495, + 530, + 692 + ], + "type": "text", + "content": " Sparse strategy by extracting sparse instance-level features from the LiDAR modality and refining them using dense image features. Other works [4, 18, 47, 45] utilize a Dense+Dense " + }, + { + "bbox": [ + 296, + 495, + 530, + 692 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 296, + 495, + 530, + 692 + ], + "type": "text", + "content": " Sparse approach where queries are used to extract a sparse instance-level representation from dense BEV and image features. However, it is hard to extract information from multi-modality features with an attention operation given the large cross-modal distribution shift. To this end, UVTR [18] projects image features into the LiDAR coordinate space, CMT [45] encodes modality-specific positional information to its queries, and DeepInteraction [47] proposes cross-modality interaction. However, these methods still need to resolve the large multimodal domain gap by stacking many transformer layers to construct a heavy decoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "text", + "content": "In contrast to the above methods, our method adopts the previously unexplored Sparse+Sparse " + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 296, + 695, + 530, + 730 + ], + "type": "text", + "content": " Sparse approach. SparseFusion extracts sparse representations of both modal-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "17593" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 142 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 142 + ], + "type": "text", + "content": "ities and fuses them to generate a more accurate and semantically rich sparse representation that yields great performance while also achieving great efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 154, + 126, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 154, + 126, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 126, + 167 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 173, + 280, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 173, + 280, + 336 + ], + "spans": [ + { + "bbox": [ + 45, + 173, + 280, + 336 + ], + "type": "text", + "content": "We present SparseFusion, an effective and efficient framework for 3D object detection via LiDAR and camera inputs. The overall architecture is illustrated in Fig. 2. We acquire sparse candidates from each modality using modality-specific object detection in the LiDAR and camera branches. The instance-level features generated by the camera branch are transformed into the LiDAR space of the instance-level features generated by the LiDAR branch. They are then fused with a simple self-attention module (Sec. 3.1). To mitigate the negative transfer between modalities, we apply a geometric transfer module and a semantic transfer module prior to the parallel detection branches (Sec. 3.2). Furthermore, we design custom loss functions for each module to ensure stable optimization (Sec. 3.3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 203, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 203, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 203, + 357 + ], + "type": "text", + "content": "3.1. Sparse Representation Fusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "spans": [ + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "text", + "content": "Acquiring candidates in two modalities. LiDAR candidates. We follow TransFusion-L [1] and use only one decoder layer for LiDAR modality detection. The LiDAR backbone extracts a BEV feature map from the point cloud inputs. We initialize " + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "text", + "content": " object queries " + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_L^0 = \\{\\mathbf{q}_{L,i}^0\\}_{i=1}^{N_L}, \\mathbf{q}_{L,i}^0 \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "text", + "content": " as well as their corresponding reference points " + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_L^0 = \\{\\mathbf{p}_{L,i}^0\\}_{i=1}^{N_L}, \\mathbf{p}_{L,i}^0 \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "text", + "content": " in the BEV plane. These queries interact with the BEV features through a cross-attention layer to generate the updated object queries " + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_L = \\{\\mathbf{q}_{L,i}\\}_{i=1}^{N_L}, \\mathbf{q}_{L,i} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "text", + "content": ". These updated queries " + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_L" + }, + { + "bbox": [ + 45, + 363, + 279, + 541 + ], + "type": "text", + "content": " represent the instance-level features of objects in the LiDAR modality, and we use them as the LiDAR candidates in the subsequent multi-modal fusion module. Furthermore, we apply a prediction head to each query to classify the object and regress the bounding box in LiDAR coordinate space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "spans": [ + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "text", + "content": "Camera candidates. To generate the camera candidates, we utilize a camera-only 3D detector with images from different perspective views as inputs. Specifically, we extend deformable-DETR [53] with 3D box regression heads. We also initialize " + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "inline_equation", + "content": "N_{C}" + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "text", + "content": " object queries " + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_C^0 = \\{\\mathbf{q}_{C,i}^0\\}_{i = 1}^{N_C}" + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{C,i}^0 \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "text", + "content": " along with their corresponding reference points " + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_C^0 = \\{\\mathbf{p}_{C,i}^0\\}_{i = 1}^{N_C}" + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{C,i}^0 \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "text", + "content": " on the image. For each perspective view " + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "text", + "content": ", queries on its image interact with the corresponding image features using a deformable attention layer [53]. The outputs of all perspective views comprise the updated queries " + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_C^P = \\{\\mathbf{q}_{C,i}^P\\}_{i = 1}^{N_C}" + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{C,i}^P \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 45, + 542, + 279, + 717 + ], + "type": "text", + "content": ". We use these queries as the camera candidates in the subsequent multi-modal fusion module. We provide further details of our architecture, the initialization method, and the prediction heads for the two modalities in our supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 718, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 718, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 46, + 718, + 279, + 730 + ], + "type": "text", + "content": "Transformation After acquiring the candidates from each" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 257 + ], + "type": "text", + "content": "modality, we aim to transform the candidates from the camera modality to the space of the candidates from the LiDAR modality. Since the candidates from the camera modality are high-dimensional latent features that are distributed differently than the candidates from the LiDAR modality, a naive coordinate transformation between modalities is inapplicable here. To address this issue, we disentangle the representations of the camera candidates. Intrinsically, a camera candidate is an instance feature that is a representation of a specific object's class and 3D bounding box. While an object's class is view-invariant, its 3D bounding box is view-dependent. This motivates us to focus on transforming high-dimensional bounding box representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "spans": [ + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "text", + "content": "We first input the candidate instance features into the prediction head of the camera branch. We label the outputted bounding boxes as " + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^P" + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "text", + "content": ". Given the extrinsic matrix " + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "inline_equation", + "content": "[\\mathbf{R}_v|\\mathbf{t}_v]" + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "text", + "content": " and intrinsic matrix " + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_v" + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "text", + "content": " of the corresponding " + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "text", + "content": "-th camera, the bounding boxes can be easily projected into the LiDAR coordinate system. We denote the project bounding boxes as " + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^L" + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "text", + "content": ". We encode the projected bounding boxes with a multi-layer perceptron (MLP) " + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "text", + "content": ", yielding a high-dimensional box embedding. We also encode the flattened camera parameter with another MLP " + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "inline_equation", + "content": "m(\\cdot)" + }, + { + "bbox": [ + 296, + 258, + 531, + 409 + ], + "type": "text", + "content": " to obtain a camera embedding. The camera embedding is multiplied with the original instance features as done in [19], which are then added to the box embedding, given by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 417, + 530, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 417, + 530, + 431 + ], + "spans": [ + { + "bbox": [ + 329, + 417, + 530, + 431 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {C, i} ^ {L} = g \\left(\\mathbf {b} _ {i} ^ {L}\\right) + h \\left(\\mathbf {q} _ {C, i} ^ {P} \\cdot m \\left(\\mathbf {R} _ {v}, \\mathbf {t} _ {v}, \\mathbf {K} _ {v}\\right)\\right), \\tag {1}", + "image_path": "4b41b902d0b49f52c7da8c18088becd788a224430a571d4686d1c69264f73e39.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "spans": [ + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "type": "text", + "content": " is an extra MLP to encode the query features in the perspective view. " + }, + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "type": "text", + "content": " aims to preserve view-agnostic information while discarding view-specific information. Afterward, " + }, + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_C^L = \\{\\mathbf{q}_{C,i}^L\\}" + }, + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "type": "text", + "content": " is passed through a self-attention layer to aggregate information from multiple cameras to get the updated queries " + }, + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_C" + }, + { + "bbox": [ + 296, + 441, + 530, + 522 + ], + "type": "text", + "content": " which represent the image modality instance features in the LiDAR space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 522, + 530, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 522, + 530, + 616 + ], + "spans": [ + { + "bbox": [ + 296, + 522, + 530, + 616 + ], + "type": "text", + "content": "Sparse candidate fusion. Our parallel modality-specific object detection provides sparse instance candidates " + }, + { + "bbox": [ + 296, + 522, + 530, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_L = \\{\\mathbf{q}_{L,i}\\}_{i=1}^{N_L}" + }, + { + "bbox": [ + 296, + 522, + 530, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 522, + 530, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_C = \\{\\mathbf{q}_{C,i}\\}_{i=1}^{N_C}" + }, + { + "bbox": [ + 296, + 522, + 530, + 616 + ], + "type": "text", + "content": " from the LiDAR and camera modalities respectively. After the above transformation of the camera candidates into LiDAR space, candidates from both modalities represent bounding boxes in the same LiDAR coordinate space, as well as the view-invariant categories. We now concatenate the candidates together:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 625, + 530, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 625, + 530, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 625, + 530, + 639 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {L C} = \\left\\{\\mathbf {q} _ {L C, i} \\right\\} _ {i = 1} ^ {N _ {L} + N _ {C}} = \\left\\{f _ {L} (\\mathbf {q} _ {L, i}) \\right\\} _ {i = 1} ^ {N _ {L}} \\cup \\left\\{f _ {C} (\\mathbf {q} _ {C, i}) \\right\\} _ {i = 1} ^ {N _ {C}} \\tag {2}", + "image_path": "e2448e455eb260832fdc4aed970ae12aecf149ff74c4e429ad49b9eab35a0b15.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 648, + 530, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 648, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 296, + 648, + 530, + 729 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 648, + 530, + 729 + ], + "type": "inline_equation", + "content": "f_{L}(\\cdot), f_{C}(\\cdot)" + }, + { + "bbox": [ + 296, + 648, + 530, + 729 + ], + "type": "text", + "content": " are learnable projectors. Afterward, we make novel use of a self-attention module to fuse the two modalities. Despite the simplicity of self-attention, the inherent intuition is novel: the modality-specific detectors encode the advantageous aspects of their respective inputs, and the self-attention module is able to aggregate and preserve the information from both modalities in an efficient" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "17594" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 103, + 495, + 330 + ], + "blocks": [ + { + "bbox": [ + 83, + 103, + 495, + 330 + ], + "lines": [ + { + "bbox": [ + 83, + 103, + 495, + 330 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 495, + 330 + ], + "type": "image", + "image_path": "b66143a9b0e469d0fda8b48448f1672540f63fae29399532fef1c08e8e6c97ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 340, + 531, + 365 + ], + "lines": [ + { + "bbox": [ + 45, + 340, + 531, + 365 + ], + "spans": [ + { + "bbox": [ + 45, + 340, + 531, + 365 + ], + "type": "text", + "content": "Figure 2: Overall framework of SparseFusion. It fuses sparse candidates from LiDAR and camera modalities to obtain a multi-modality instance-level representation in the unified LiDAR space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 382, + 162, + 471 + ], + "blocks": [ + { + "bbox": [ + 48, + 382, + 162, + 471 + ], + "lines": [ + { + "bbox": [ + 48, + 382, + 162, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 162, + 471 + ], + "type": "image", + "image_path": "dcecbe289c436eacd957287d6c1024e522935ae42d50c83d1a29e30ee34546bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 482, + 279, + 517 + ], + "lines": [ + { + "bbox": [ + 45, + 482, + 279, + 517 + ], + "spans": [ + { + "bbox": [ + 45, + 482, + 279, + 517 + ], + "type": "text", + "content": "Figure 3: Cross-modality information transfer. We transfer geometric information from LiDAR to camera modality, and semantic information from camera to LiDAR modality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 164, + 382, + 278, + 471 + ], + "blocks": [ + { + "bbox": [ + 164, + 382, + 278, + 471 + ], + "lines": [ + { + "bbox": [ + 164, + 382, + 278, + 471 + ], + "spans": [ + { + "bbox": [ + 164, + 382, + 278, + 471 + ], + "type": "image", + "image_path": "1bf6d711fbe9c21d28e7b64e12748c757bc2f9d78de6bafcffe9d5599f6e1ab7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 538, + 279, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 538, + 279, + 563 + ], + "spans": [ + { + "bbox": [ + 45, + 538, + 279, + 563 + ], + "type": "text", + "content": "manner. The output of the self-attention module is used for final classification and regression of the bounding boxes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 570, + 239, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 570, + 239, + 583 + ], + "spans": [ + { + "bbox": [ + 45, + 570, + 239, + 583 + ], + "type": "text", + "content": "3.2. Cross-Modality Information Transfer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 589, + 279, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 589, + 279, + 704 + ], + "spans": [ + { + "bbox": [ + 45, + 589, + 279, + 704 + ], + "type": "text", + "content": "Although we aim to utilize the advantages of both modalities, we must address that the modalities also have their own disadvantages that can result in negative transfer between modalities. For example, the LiDAR detector struggles to capture rich semantic information, while the camera detector struggles to capture accurate geometric and depth information. To mitigate negative transfer, we propose novel geometric and semantic information transfer modules, as illustrated in Fig. 3, that we apply prior to the modality-specific detectors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": "Geometric transfer from LiDAR to camera. We project each point in the LiDAR point cloud input to multi-view" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 384, + 530, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 384, + 530, + 465 + ], + "spans": [ + { + "bbox": [ + 296, + 384, + 530, + 465 + ], + "type": "text", + "content": "images to generate sparse multi-view depth maps. These multi-view depth maps are inputted into a shared encoder to obtain depth features, which are then concatenated with the image features to form depth-aware image features that compensate for the lack of geometric information in camera inputs. The depth-aware image features are used as the input to the camera branch." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 465, + 530, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 465, + 530, + 594 + ], + "spans": [ + { + "bbox": [ + 296, + 465, + 530, + 594 + ], + "type": "text", + "content": "Semantic transfer from camera to LiDAR. We project the points in the LiDAR point cloud input to the image inputs, which yields sparse points on the image features. We perform max-pooling to aggregate the resulting multiscale features, and we combine them with the BEV feature through addition. The concatenated features serve as the queries and interact with the multi-scale image features through deformable-attention [53]. The updated queries replace the original queries in the BEV features, which results in the semantic-aware BEV features, which are used for query initialization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 600, + 406, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 600, + 406, + 613 + ], + "spans": [ + { + "bbox": [ + 297, + 600, + 406, + 613 + ], + "type": "text", + "content": "3.3. Objective Function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 618, + 530, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 618, + 530, + 642 + ], + "spans": [ + { + "bbox": [ + 296, + 618, + 530, + 642 + ], + "type": "text", + "content": "We apply the Gaussian focal loss [24] to the initialized queries of both modalities, given by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 649, + 530, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 649, + 530, + 664 + ], + "spans": [ + { + "bbox": [ + 308, + 649, + 530, + 664 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n i t}} = \\mathcal {L} _ {G F o c a l} (\\hat {\\mathbf {Y}} _ {L}, \\mathbf {Y} _ {L}) + \\mathcal {L} _ {G F o c a l} (\\hat {\\mathbf {Y}} _ {C}, \\mathbf {Y} _ {C}), \\tag {3}", + "image_path": "2ee81d3166a093f9f27138b34c5ae625f6ad6feb71bce10260e90bead3422d72.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 670, + 530, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 670, + 530, + 706 + ], + "spans": [ + { + "bbox": [ + 296, + 670, + 530, + 706 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 670, + 530, + 706 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_L, \\hat{\\mathbf{Y}}_C" + }, + { + "bbox": [ + 296, + 670, + 530, + 706 + ], + "type": "text", + "content": " are the dense predictions of category-wise heatmaps of the LiDAR and camera modalities, respectively, and " + }, + { + "bbox": [ + 296, + 670, + 530, + 706 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_L, \\mathbf{Y}_C" + }, + { + "bbox": [ + 296, + 670, + 530, + 706 + ], + "type": "text", + "content": " are the corresponding ground-truths." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 706, + 530, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 706, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 297, + 706, + 530, + 729 + ], + "type": "text", + "content": "Then, we apply the loss function for the detectors of the LiDAR and camera modalities, as well as the view transfor" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "17595" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 281, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 281, + 223 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 281, + 223 + ], + "type": "text", + "content": "mation of the camera candidates and the candidate fusion stage. Firstly, the predictions of each modality-specific detector are independently matched with the ground-truth using the Hungary algorithm [16]. The object classification is optimized with focal loss [24] and the 3D bounding box regression is optimized with L1 loss. For the camera modality detector, the ground-truth bounding boxes are in separate camera coordinates. For all other detectors, ground-truth bounding boxes are in LiDAR coordinates. The detection loss can be represented as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 229, + 279, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 229, + 279, + 245 + ], + "spans": [ + { + "bbox": [ + 50, + 229, + 279, + 245 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e t e c t}} = \\gamma \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {c a m e r a}} + \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {t r a n s}} + \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {L i D A R}} + \\mathcal {L} _ {\\text {d e t e c t}} ^ {\\text {f u s i o n}}. \\tag {4}", + "image_path": "f375a8f4838720a84283581c695ec296137095f4585265d0c033266d64e037fe.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 250, + 279, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 250, + 279, + 286 + ], + "spans": [ + { + "bbox": [ + 45, + 250, + 279, + 286 + ], + "type": "text", + "content": "Our entire network is optimized using " + }, + { + "bbox": [ + 45, + 250, + 279, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\alpha \\mathcal{L}_{init} + \\beta \\mathcal{L}_{detect}" + }, + { + "bbox": [ + 45, + 250, + 279, + 286 + ], + "type": "text", + "content": ". In our implementation, we empirically set " + }, + { + "bbox": [ + 45, + 250, + 279, + 286 + ], + "type": "inline_equation", + "content": "\\gamma = 1, \\alpha = 0.1" + }, + { + "bbox": [ + 45, + 250, + 279, + 286 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 250, + 279, + 286 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 45, + 250, + 279, + 286 + ], + "type": "text", + "content": " to balance different terms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 296, + 124, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 296, + 124, + 309 + ], + "spans": [ + { + "bbox": [ + 45, + 296, + 124, + 309 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 314, + 158, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 314, + 158, + 326 + ], + "spans": [ + { + "bbox": [ + 45, + 314, + 158, + 326 + ], + "type": "text", + "content": "4.1. Dataset and metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "spans": [ + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "type": "text", + "content": "We follow previous work [47, 45, 22, 5] to evaluate our method on the nuScenes dataset [2]. It is a challenging dataset for 3D object detection, consisting of 700/150/150 scenes for training/validation/test. It provides point clouds collected using a 32-beam LiDAR and six images from multi-view cameras. There are 1.4 million annotated 3D bounding boxes for objects from 10 different classes. We evaluate performance using the nuScenes detection score (NDS) and mean average precision (mAP) metrics. The final mAP is averaged over distance thresholds " + }, + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "type": "inline_equation", + "content": "2\\mathrm{m}" + }, + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "type": "inline_equation", + "content": "4\\mathrm{m}" + }, + { + "bbox": [ + 45, + 332, + 281, + 484 + ], + "type": "text", + "content": " on the BEV across 10 classes. NDS is a weighted average of mAP and other true positive metrics including mATE, mASE, mAOE, mAVE, and mAAE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 490, + 175, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 490, + 175, + 503 + ], + "spans": [ + { + "bbox": [ + 45, + 490, + 175, + 503 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "spans": [ + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "text", + "content": "Our implementation is based on the MMDetection3D framework [6]. For the camera branch, we use ResNet-50 [11] as the backbone and initialize it with the Mask R-CNN [10] instance segmentation network pretrained on nuImage [2]. The input image resolution is " + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "inline_equation", + "content": "800 \\times 448" + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "text", + "content": ". For the LiDAR branch, we apply VoxelNet [51] with voxel size " + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "inline_equation", + "content": "(0.075m, 0.075m, 0.2m)" + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "text", + "content": ". The detection range is set as " + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "inline_equation", + "content": "[-54m, 54m]" + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "text", + "content": " for the XY-axes and " + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "inline_equation", + "content": "[-5m, 3m]" + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "text", + "content": " for the Z-axis. Our LiDAR and camera modality detectors both include only 1 decoder layer. The query numbers for the LiDAR and camera modalities are set as " + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "inline_equation", + "content": "N_L = N_C = 200" + }, + { + "bbox": [ + 45, + 508, + 279, + 648 + ], + "type": "text", + "content": ", so our fusion stage can detect at most 400 objects per scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 648, + 280, + 730 + ], + "type": "text", + "content": "Since our framework disentangles the camera detector and the LiDAR detector, we can conveniently apply data augmentation separately to the LiDAR inputs and camera inputs. We apply random rotation, scaling, translation, and flipping to the LiDAR inputs, and we apply random scaling and horizontal flipping to the camera inputs. Our training pipeline follows previous works [1, 28, 47]. We first train" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 106, + 532, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 532, + 246 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 532, + 246 + ], + "type": "text", + "content": "TransFusion-L [1] as our LiDAR-only baseline, which is used to initialize our LiDAR backbone and LiDAR modality detector. This LiDAR-only baseline is trained for 20 epochs. Afterward, we freeze the pretrained LiDAR components and train the entire fusion framework for 6 epochs. For both training stages, we use the AdamW optimizer [29] with one-cycle learning rate policy [37]. The initial learning rate is " + }, + { + "bbox": [ + 296, + 106, + 532, + 246 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 296, + 106, + 532, + 246 + ], + "type": "text", + "content": " and the weight decay is " + }, + { + "bbox": [ + 296, + 106, + 532, + 246 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 296, + 106, + 532, + 246 + ], + "type": "text", + "content": ". The hidden dimensions in the entire model except the backbones are 128. For both training stages, we adopt CBGS [52] to balance the class distribution. We train our method on four NVIDIA A6000 GPUs with batch size 16." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 255, + 432, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 255, + 432, + 267 + ], + "spans": [ + { + "bbox": [ + 296, + 255, + 432, + 267 + ], + "type": "text", + "content": "4.3. Results and Comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "spans": [ + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "type": "text", + "content": "We report results on the nuScenes validation and test sets, without using any test-time augmentations or model ensembles. As shown in Tab. 2, SparseFusion significantly improves our LiDAR-only baseline, TransFusion-L [1], by " + }, + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "type": "inline_equation", + "content": "+3.6\\%" + }, + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "type": "inline_equation", + "content": "+6.3\\%" + }, + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "type": "text", + "content": " mAP on the test set, due to the additional use of camera inputs. More importantly, SparseFusion sets a new state of the art on both the validation set and test set, outperforming prior works including those using stronger backbones. It is noteworthy that SparseFusion demonstrates a " + }, + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 296, + 272, + 532, + 459 + ], + "type": "text", + "content": " mAP improvement over the most recent state-of-the-art [47] while also achieving a 1.8x speedup (5.6 FPS vs. 3.1 FPS) on A6000 GPU. It can be seen in Fig. 1a that, in addition to superior performance, SparseFusion also provides the fastest inference speed. We also demonstrate the performance of SparseFusion by visualizing some qualitative results in Fig. 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 467, + 358, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 467, + 358, + 479 + ], + "spans": [ + { + "bbox": [ + 296, + 467, + 358, + 479 + ], + "type": "text", + "content": "4.4. Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 484, + 532, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 484, + 532, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 484, + 532, + 730 + ], + "type": "text", + "content": "Performance Breakdown SparseFusion performs 3D object detection in both the LiDAR and camera branch separately. Tab. 3 shows the detection performance in different parts of SparseFusion including the LiDAR branch, camera branch (before and after view transformation), and the fusion branch. We notice that our LiDAR branch detection results notably surpass the LiDAR-only baseline TransFusion-L [1] since the proposed semantic transfer can compensate for the weakness of point cloud inputs. In comparison with the state-of-the-art single-frame camera detector PETR [26] (six decoder layers), our camera branch achieves much better performance with just one decoder layer, owing to the depth-aware features from the proposed geometric transfer. Besides, our view transformation module not only transforms the instance features from the camera coordinate space into the LiDAR coordinate space, but it also slightly improves the camera branch detection performance by aggregating multi-view information. With this strong performance, the modality-specific detectors in each branch would not cause negative transfer during fusion. We also consider direct Non-Maximum Suppression" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "17596" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 125, + 526, + 340 + ], + "blocks": [ + { + "bbox": [ + 124, + 105, + 452, + 116 + ], + "lines": [ + { + "bbox": [ + 124, + 105, + 452, + 116 + ], + "spans": [ + { + "bbox": [ + 124, + 105, + 452, + 116 + ], + "type": "text", + "content": "Table 2: Comparison with existing methods on nuScenes validation set and test set." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 125, + 526, + 340 + ], + "lines": [ + { + "bbox": [ + 52, + 125, + 526, + 340 + ], + "spans": [ + { + "bbox": [ + 52, + 125, + 526, + 340 + ], + "type": "table", + "html": "
MethodsModalityLiDAR BackboneCamera Backbonevalidation settest set
NDSmAPNDSmAP
FCOS3D [42]Camera-ResNet-101 [11]41.534.342.835.8
PETR [26]Camera-ResNet-101 [11]44.237.045.539.1
CenterPoint [48]LiDARVoxelNet [51]-66.859.667.360.3
TransFusion-L [1]LiDARVoxelNet [51]-70.165.170.265.5
PointAugmenting [39]LiDAR+CameraVoxelNet [51]DLA34 [49]--71.066.8
FUTR3D [4]LiDAR+CameraVoxelNet [51]ResNet-101 [11]68.364.5--
UVTR [18]LiDAR+CameraVoxelNet [51]ResNet-101 [11]70.265.471.167.1
TransFusion [1]LiDAR+CameraVoxelNet [51]ResNet-50 [11]71.367.571.668.9
AutoAlignV2 [5]LiDAR+CameraVoxelNet [51]CSPNet [41]71.267.172.468.4
BEVFusion [22]LiDAR+CameraVoxelNet [51]Dual-Swin-T [21]72.169.673.371.3
BEVFusion [28]LiDAR+CameraVoxelNet [51]Swin-T [27]71.468.572.970.2
DeepInteraction [47]LiDAR+CameraVoxelNet [51]ResNet-50 [11]72.669.973.470.8
CMT [45]LiDAR+CameraVoxelNet [51]ResNet-50 [11]70.867.9--
SparseFusion (ours)LiDAR+CameraVoxelNet [51]ResNet-50 [11]72.870.473.872.0
", + "image_path": "06eb2d0e51129242144b5cfd0fc6db893ffd3c1c596f4f7a29000258535c6555.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 72, + 350, + 505, + 479 + ], + "blocks": [ + { + "bbox": [ + 72, + 350, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 72, + 350, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 72, + 350, + 505, + 479 + ], + "type": "image", + "image_path": "93a2f08e907b7768b6021c67114553ff0af23b2938c90c333e07bc4aad988d21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 487, + 431, + 500 + ], + "lines": [ + { + "bbox": [ + 144, + 487, + 431, + 500 + ], + "spans": [ + { + "bbox": [ + 144, + 487, + 431, + 500 + ], + "type": "text", + "content": "Figure 4: Qualitative results of SparseFusion on nuScenes validation set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 510, + 280, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 510, + 280, + 569 + ], + "spans": [ + { + "bbox": [ + 45, + 510, + 280, + 569 + ], + "type": "text", + "content": "(NMS) to combine the detection results of LiDAR and camera branches, which is even worse than the single LiDAR branch. Therefore, it is also important to make the instance-level features from the two branches interact with each other in the fusion stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 573, + 279, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 573, + 279, + 667 + ], + "spans": [ + { + "bbox": [ + 45, + 573, + 279, + 667 + ], + "type": "text", + "content": "Strong Image Backbone. We incorporate stronger SwinT [27] backbone into SparseFusion to match some previous work [22, 45]. Tab. 4 compares the performance of different methods on the nuScenes validation set. We do not include any test-time augmentations or model ensembles. Although multi-modality detection relies more on the LiDAR inputs, SparseFusion can still benefit from a stronger image backbone and beat all the counterparts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "text", + "content": "Modality-Specific Object Recall. The parallel detectors in the LiDAR and camera branches enable us to determine which modality recalls each object. Given " + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "inline_equation", + "content": "N_{L} + N_{C}" + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "text", + "content": " predictions, we know that the first " + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "text", + "content": " instances come from the LiDAR-modality detector, while the last " + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "inline_equation", + "content": "N_{C}" + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "text", + "content": " are from" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 508, + 531, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 508, + 531, + 578 + ], + "spans": [ + { + "bbox": [ + 296, + 508, + 531, + 578 + ], + "type": "text", + "content": "Table 3: Performance breakdown. We show the detection results in different parts of SparseFusion on the validation set. 'L' and 'C' refer to the LiDAR and camera modalities, respectively. 'ST' and 'GT' refer to semantic transfer and geometric transfer, respectively. 'VT' is the view transformation module." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 298, + 587, + 543, + 704 + ], + "blocks": [ + { + "bbox": [ + 298, + 587, + 543, + 704 + ], + "lines": [ + { + "bbox": [ + 298, + 587, + 543, + 704 + ], + "spans": [ + { + "bbox": [ + 298, + 587, + 543, + 704 + ], + "type": "table", + "html": "
MethodsModalityNDSmAP
TransFusion-L [1]L70.165.1
LiDAR branchL + ST71.868.4
PETR (ResNet50) [26]C38.131.3
Camera branch (before VT)C+GT43.540.6
Camera branch (after VT)C+GT44.341.5
LiDAR/camera branches NMSL+C71.568.2
SparseFusionL+C72.870.4
", + "image_path": "4cfda4783b3af7890c8e4908afa166e8645ba62f5ba4f79dadd902527b3f9743.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "17597" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 172, + 277, + 243 + ], + "blocks": [ + { + "bbox": [ + 45, + 104, + 279, + 162 + ], + "lines": [ + { + "bbox": [ + 45, + 104, + 279, + 162 + ], + "spans": [ + { + "bbox": [ + 45, + 104, + 279, + 162 + ], + "type": "text", + "content": "Table 4: Results on nuScenes validation set with stronger image backbones. We do not use any test-time augmentations or model ensembles. The inference speed is measured on a single NVIDIA A6000 GPU. " + }, + { + "bbox": [ + 45, + 104, + 279, + 162 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 45, + 104, + 279, + 162 + ], + "type": "text", + "content": " : CMT [45] adopts flash-attention [7] for transformer acceleration." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 172, + 277, + 243 + ], + "lines": [ + { + "bbox": [ + 48, + 172, + 277, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 277, + 243 + ], + "type": "table", + "html": "
MethodsImage BackboneInput ResolutionNDSmAPFPS
BEVFusion [22]Dual-Swin-T1600×90072.169.60.8
CMT [45]VoVNet-991600×64072.970.33.8†
SparseFusionResNet-50800×44872.870.45.6
SparseFusionSwin-T800×44873.171.05.3
", + "image_path": "50f111b6ff227fcd58c07235ae9949f8035349e4519fa131320261d79230b1bf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 256, + 158, + 338 + ], + "blocks": [ + { + "bbox": [ + 48, + 256, + 158, + 338 + ], + "lines": [ + { + "bbox": [ + 48, + 256, + 158, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 158, + 338 + ], + "type": "image", + "image_path": "63be14ae83bf2400e368d6b8b4edac6adbd2f40697cbef7e7152bf986d80f0c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 346, + 278, + 357 + ], + "lines": [ + { + "bbox": [ + 48, + 346, + 278, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 346, + 278, + 357 + ], + "type": "text", + "content": "(a) Objects of different classes. (b) Objects at different distances." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 164, + 255, + 274, + 337 + ], + "blocks": [ + { + "bbox": [ + 164, + 255, + 274, + 337 + ], + "lines": [ + { + "bbox": [ + 164, + 255, + 274, + 337 + ], + "spans": [ + { + "bbox": [ + 164, + 255, + 274, + 337 + ], + "type": "image", + "image_path": "066a58f9268af594eed9cebf12dc191e3545a0ea764e3657c0750c4a9385e517.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 335, + 530, + 417 + ], + "lines": [ + { + "bbox": [ + 296, + 335, + 530, + 417 + ], + "spans": [ + { + "bbox": [ + 296, + 335, + 530, + 417 + ], + "type": "text", + "content": "Figure 6: Instance-level feature interaction in the fusion stage. Orange boxes are detected after the fusion stage with high confidence in the BEV space. Blue and green dots denote all instances from the LiDAR and camera branches separately. Orange boxes are connected with blue/green dots with blue/green lines. The attention strength is represented by the darkness and thickness of the lines." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 402, + 279, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 402, + 279, + 541 + ], + "spans": [ + { + "bbox": [ + 45, + 402, + 279, + 541 + ], + "type": "text", + "content": "the camera modality. An object is recalled if a bounding box with correct classification is predicted within a radius of two meters around its ground-truth position. In Fig. 5, we demonstrate the percentage of objects in the nuScenes validation set recalled by exactly one modality. We observe that each modality can compensate for the weakness of the other to some extent. Although the LiDAR modality is typically more powerful, the camera modality plays an important role in detecting objects from classes such as construction vehicles and barriers. Furthermore, the camera modality is useful for detecting objects at far distances where point clouds are sparse or the LiDAR feature map cannot cover." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 543, + 279, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 543, + 279, + 670 + ], + "spans": [ + { + "bbox": [ + 45, + 543, + 279, + 670 + ], + "type": "text", + "content": "Cross-Modality Sparse Representation Interaction. Fig. 6 visualizes the instance feature interaction in the sparse fusion stage. The strength of attention between instance-level features is reflected by the thickness and darkness of lines. We notice that most objects can aggregate multi-modality instance-level features during fusion. Although the strongest interactions exist mainly between neighboring instances, it is interesting that features from the camera modality are also able to deliver strong interactions with the instances at the distant range. This could be a result of the shared semantics among objects in images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": "Robustness to Sensor Malfunction. With the relatively independent modality-specific detectors, SparseFusion is robust to sensor malfunctions. We evaluate SparseFusion in two different cases: (1) The LiDAR scan range is limited to " + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "inline_equation", + "content": "(-90^{\\circ}, 90^{\\circ})" + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": " and (2) The most important front camera is" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 298, + 103, + 530, + 327 + ], + "blocks": [ + { + "bbox": [ + 45, + 365, + 278, + 389 + ], + "lines": [ + { + "bbox": [ + 45, + 365, + 278, + 389 + ], + "spans": [ + { + "bbox": [ + 45, + 365, + 278, + 389 + ], + "type": "text", + "content": "Figure 5: Percentage of objects recalled by each modality. We do not count those objects recalled by both or neither." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 298, + 103, + 530, + 327 + ], + "lines": [ + { + "bbox": [ + 298, + 103, + 530, + 327 + ], + "spans": [ + { + "bbox": [ + 298, + 103, + 530, + 327 + ], + "type": "image", + "image_path": "bc8a60f387e60e5d4eb5d09994a5ee653c979f2431056354735e865e4590674c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 298, + 474, + 535, + 573 + ], + "blocks": [ + { + "bbox": [ + 296, + 430, + 530, + 464 + ], + "lines": [ + { + "bbox": [ + 296, + 430, + 530, + 464 + ], + "spans": [ + { + "bbox": [ + 296, + 430, + 530, + 464 + ], + "type": "text", + "content": "Table 5: Results on nuScenes validation set for LiDAR or camera sensor malfunctions. The models are finetuned on the in-domain data in both cases." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 474, + 535, + 573 + ], + "lines": [ + { + "bbox": [ + 298, + 474, + 535, + 573 + ], + "spans": [ + { + "bbox": [ + 298, + 474, + 535, + 573 + ], + "type": "table", + "html": "
MethodsSensor MalfunctionNDSmAP
TransFusion [1]Limited LiDAR Field (-90°, 90°)49.231.1
BEVFusion [22]54.945.5
SparseFusion61.254.3
TransFusion [1]Missing70.165.3
BEVFusion [22]Front Camera70.765.9
SparseFusion72.169.2
", + "image_path": "b476547d04d07f42dfada579fab0e4d7f37102b06598a86b3977333fee697a63.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 585, + 530, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 585, + 530, + 631 + ], + "spans": [ + { + "bbox": [ + 296, + 585, + 530, + 631 + ], + "type": "text", + "content": "missing while other cameras are preserved. SparseFusion outperforms previous methods [1, 22] with a large gap under both scenarios, demonstrating great robustness to sensor malfunctions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 640, + 395, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 640, + 395, + 652 + ], + "spans": [ + { + "bbox": [ + 297, + 640, + 395, + 652 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 659, + 530, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 659, + 530, + 704 + ], + "spans": [ + { + "bbox": [ + 296, + 659, + 530, + 704 + ], + "type": "text", + "content": "In this section, we study the effect of using alternatives for the different modules in SparseFusion. For our ablation studies, we train on a 1/5 split of nuScenes training set and evaluate on the full nuScenes validation set." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 706, + 530, + 730 + ], + "type": "text", + "content": "Sparse Fusion Strategy. We compare our self-attention module for sparse feature fusion with other fusion methods." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "17598" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 137, + 201, + 212 + ], + "blocks": [ + { + "bbox": [ + 143, + 104, + 431, + 116 + ], + "lines": [ + { + "bbox": [ + 143, + 104, + 431, + 116 + ], + "spans": [ + { + "bbox": [ + 143, + 104, + 431, + 116 + ], + "type": "text", + "content": "Table 6: Ablation study experiments to justify our design of each module." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 127, + 175, + 137 + ], + "lines": [ + { + "bbox": [ + 78, + 127, + 175, + 137 + ], + "spans": [ + { + "bbox": [ + 78, + 127, + 175, + 137 + ], + "type": "text", + "content": "(a) Sparse fusion strategies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 137, + 201, + 212 + ], + "lines": [ + { + "bbox": [ + 53, + 137, + 201, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 201, + 212 + ], + "type": "table", + "html": "
Fusion StrategyNDSmAP
MLP67.864.5
Cross-Attention68.665.8
Optimal Transport68.065.6
Self-Attention68.866.4
", + "image_path": "63f0795a5b509d8de1cdb978e3d5ba1cccecfb96233b2d78e267e3eff3ea91ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 214, + 142, + 367, + 206 + ], + "blocks": [ + { + "bbox": [ + 245, + 127, + 335, + 137 + ], + "lines": [ + { + "bbox": [ + 245, + 127, + 335, + 137 + ], + "spans": [ + { + "bbox": [ + 245, + 127, + 335, + 137 + ], + "type": "text", + "content": "(b) Information transfers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 214, + 142, + 367, + 206 + ], + "lines": [ + { + "bbox": [ + 214, + 142, + 367, + 206 + ], + "spans": [ + { + "bbox": [ + 214, + 142, + 367, + 206 + ], + "type": "table", + "html": "
GeometricSemanticNDSmAP
✓✗67.764.2
✗✓68.465.7
✓✓68.866.4
", + "image_path": "9fa060d6aec09c954349a8369ee0d26f002f6dec007fad2d6744d50a1c369dc0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 376, + 142, + 528, + 206 + ], + "blocks": [ + { + "bbox": [ + 382, + 127, + 521, + 137 + ], + "lines": [ + { + "bbox": [ + 382, + 127, + 521, + 137 + ], + "spans": [ + { + "bbox": [ + 382, + 127, + 521, + 137 + ], + "type": "text", + "content": "(c) Parallel modality-specific detectors." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 376, + 142, + 528, + 206 + ], + "lines": [ + { + "bbox": [ + 376, + 142, + 528, + 206 + ], + "spans": [ + { + "bbox": [ + 376, + 142, + 528, + 206 + ], + "type": "table", + "html": "
Seq. Pos.Seq. Feat.NDSmAP
✓X68.165.3
✓✓67.964.2
XX68.866.4
", + "image_path": "b0b91e7035c4b307c3db3ec1ae96b6988baf4bb95b97e40bff8be073e595e4a9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 231, + 279, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 231, + 279, + 383 + ], + "spans": [ + { + "bbox": [ + 45, + 231, + 279, + 383 + ], + "type": "text", + "content": "In addition to our self-attention module in Sec. 3.1, three alternatives are considered: 1) Instance-level candidates from the two branches are directly fed into an MLP without any cross-instance aggregation. 2) LiDAR instance candidates are used as queries and camera instance candidates are used as keys/values. A cross-attention module is then used to fuse multi-modality instance features. 3) We make novel use of optimal transport in LiDAR-camera fusion [33]. We propose to learn a distribution transformation from camera candidates to LiDAR candidates through optimal transport. Then, we can directly fuse them by concatenating the candidates of two branches along the channel dimension. More details about this method are provided in appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 384, + 279, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 384, + 279, + 512 + ], + "spans": [ + { + "bbox": [ + 45, + 384, + 279, + 512 + ], + "type": "text", + "content": "The results in Tab. 6a show that only cross-attention achieves competitive performance to self-attention. Yet, we observe that it relies so heavily on the output of the LiDAR branch that the camera branch is not fully utilized to compensate for the weaknesses of the LiDAR branch. The MLP strategy has limited performance as it does not fuse cross-instance and cross-modality information. Despite the impressive progress of optimal transport in other fields, it fails to learn the correspondences between the instance features of the two modalities and thus has limited performance. In contrast, self-attention is simple, efficient, and effective." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 515, + 279, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 515, + 279, + 631 + ], + "spans": [ + { + "bbox": [ + 45, + 515, + 279, + 631 + ], + "type": "text", + "content": "Information Transfer. We then ablate the geometric and semantic transfer between the LiDAR and camera modalities. The results in Tab. 6b show that the fusion performance benefits from both transfers. This also validates that the disadvantages of both modalities result in negative transfer and that our proposed information transfer modules are indeed effective in mitigating this issue. The semantic transfer module especially improves the final performance since it compensates for the LiDAR modality's lack of semantic information, which is critical for 3D detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 634, + 279, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 634, + 279, + 704 + ], + "spans": [ + { + "bbox": [ + 45, + 634, + 279, + 704 + ], + "type": "text", + "content": "Despite the importance of the information transfer, SparseFusion trained on the full training set can reach 71.5 NDS on the validation set even if we remove both the semantic and geometric transfers. This still outperforms many other advanced algorithms [1, 28], reflecting the robustness of our fusion stage to the modality-specific weakness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": "View Transformation. As explained in Sec. 3.1, we transform the sparse representations of both modalities into one" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "spans": [ + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "type": "text", + "content": "unified space. To validate the effectiveness of this approach, we ablate the view transformation of camera candidates into the LiDAR coordinate space. This results in a more straightforward method where we simply obtain the predictions of two modalities and directly fuse them using self-attention. Ablating the view transformation drops performance from " + }, + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "type": "inline_equation", + "content": "66.4\\%" + }, + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "type": "inline_equation", + "content": "68.8\\%" + }, + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "type": "text", + "content": " NDS to " + }, + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "type": "inline_equation", + "content": "65.6\\%" + }, + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "type": "inline_equation", + "content": "68.3\\%" + }, + { + "bbox": [ + 296, + 231, + 530, + 335 + ], + "type": "text", + "content": " NDS, respectively. This demonstrates that the view transformation is indeed helpful to overall performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 336, + 530, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 336, + 530, + 534 + ], + "spans": [ + { + "bbox": [ + 296, + 336, + 530, + 534 + ], + "type": "text", + "content": "Parallel Detectors. In addition to the ablation studies on the modules of the proposed pipeline, we also study alternatives to the structure of the pipeline. SparseFusion uses separate 3D object detectors for the LiDAR and camera modalities in parallel for extracting instance-level features to address the cross-modality domain gap. Alternatively, we consider a sequential pipeline where the camera detector runs after the LiDAR detector. The camera-modality detector inherits the output queries from the LiDAR detector. We consider two variants of this inheritance: 1) using the 3D position and instance features for the initial query of the camera modality; 2) using the 3D position but initializing the instance features from the corresponding image features. The camera detector follows the structure of the PETR [26] decoder (one layer). Tab. 6c shows that both sequential structures yield notably inferior performance, justifying our use of parallel modality-specific detectors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 546, + 367, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 546, + 367, + 558 + ], + "spans": [ + { + "bbox": [ + 297, + 546, + 367, + 558 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 566, + 530, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 566, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 296, + 566, + 530, + 729 + ], + "type": "text", + "content": "We revisit previous LiDAR-camera fusion works and propose SparseFusion, a novel 3D object detection method that utilizes the rarely-explored strategy of fusing sparse representations. SparseFusion extracts instance-level features from each modality separately via parallel 3D object detectors and then treats the instance-level features as the modality-specific candidates. Afterward, we transform the candidates into a unified 3D space, and we are able to fuse the candidates with a lightweight attention module. Extensive experiments demonstrate that SparseFusion achieves state-of-the-art performance on the nuScenes benchmark with the fastest inference speed. We hope SparseFusion will serve as a powerful and efficient baseline for further research into this field." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 300, + 756 + ], + "type": "text", + "content": "17599" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 280, + 729 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 51, + 124, + 279, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 124, + 279, + 188 + ], + "spans": [ + { + "bbox": [ + 51, + 124, + 279, + 188 + ], + "type": "text", + "content": "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1090-1099, 2022. 2, 3, 4, 6, 7, 8, 9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 190, + 280, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 190, + 280, + 255 + ], + "spans": [ + { + "bbox": [ + 51, + 190, + 280, + 255 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 255, + 280, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 255, + 280, + 320 + ], + "spans": [ + { + "bbox": [ + 51, + 255, + 280, + 320 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 321, + 279, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 321, + 279, + 364 + ], + "spans": [ + { + "bbox": [ + 51, + 321, + 279, + 364 + ], + "type": "text", + "content": "[4] Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. arXiv preprint arXiv:2203.10642, 2022. 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 366, + 279, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 366, + 279, + 410 + ], + "spans": [ + { + "bbox": [ + 51, + 366, + 279, + 410 + ], + "type": "text", + "content": "[5] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Autoalignv2: Deformable feature aggregation for dynamic multi-modal 3d object detection. arXiv preprint arXiv:2207.10316, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 410, + 279, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 410, + 279, + 452 + ], + "spans": [ + { + "bbox": [ + 51, + 410, + 279, + 452 + ], + "type": "text", + "content": "[6] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mm Detection3d, 2020.6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 454, + 279, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 454, + 279, + 498 + ], + "spans": [ + { + "bbox": [ + 51, + 454, + 279, + 498 + ], + "type": "text", + "content": "[7] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. In Advances in Neural Information Processing Systems, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 499, + 279, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 499, + 279, + 553 + ], + "spans": [ + { + "bbox": [ + 51, + 499, + 279, + 553 + ], + "type": "text", + "content": "[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 554, + 279, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 554, + 279, + 607 + ], + "spans": [ + { + "bbox": [ + 51, + 554, + 279, + 607 + ], + "type": "text", + "content": "[9] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2918-2927, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 608, + 279, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 279, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 279, + 641 + ], + "type": "text", + "content": "[10] Kaiming He, Georgia Gkioxari, Piotr Dálár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 642, + 279, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 279, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 279, + 685 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "type": "text", + "content": "[12] Chenxi Huang, Tong He, Haidong Ren, Wenxiao Wang, Binbin Lin, and Deng Cai. Obmo: One bounding box multiple objects for monocular 3d object detection. arXiv preprint arXiv:2212.10049, 2022. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 530, + 728 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 299, + 106, + 530, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 106, + 530, + 138 + ], + "spans": [ + { + "bbox": [ + 299, + 106, + 530, + 138 + ], + "type": "text", + "content": "[13] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 139, + 530, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 139, + 530, + 171 + ], + "spans": [ + { + "bbox": [ + 299, + 139, + 530, + 171 + ], + "type": "text", + "content": "[14] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 172, + 530, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 172, + 530, + 224 + ], + "spans": [ + { + "bbox": [ + 299, + 172, + 530, + 224 + ], + "type": "text", + "content": "[15] Yang Jiao, Zequn Jie, Shaoxiang Chen, Jingjing Chen, Xiaolin Wei, Lin Ma, and Yu-Gang Jiang. Msmdfusion: Fusing lidar and camera at multiple scales with multi-depth seeds for 3d object detection. arXiv preprint arXiv:2209.03102, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 225, + 530, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 225, + 530, + 256 + ], + "spans": [ + { + "bbox": [ + 299, + 225, + 530, + 256 + ], + "type": "text", + "content": "[16] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 257, + 530, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 257, + 530, + 311 + ], + "spans": [ + { + "bbox": [ + 299, + 257, + 530, + 311 + ], + "type": "text", + "content": "[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 312, + 530, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 312, + 530, + 353 + ], + "spans": [ + { + "bbox": [ + 299, + 312, + 530, + 353 + ], + "type": "text", + "content": "[18] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 354, + 530, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 354, + 530, + 396 + ], + "spans": [ + { + "bbox": [ + 299, + 354, + 530, + 396 + ], + "type": "text", + "content": "[19] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 397, + 530, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 397, + 530, + 472 + ], + "spans": [ + { + "bbox": [ + 299, + 397, + 530, + 472 + ], + "type": "text", + "content": "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 1-18. Springer, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 472, + 530, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 472, + 530, + 525 + ], + "spans": [ + { + "bbox": [ + 299, + 472, + 530, + 525 + ], + "type": "text", + "content": "[21] Tingting Liang, Xiaojie Chu, Yudong Liu, Yongtao Wang, Zhi Tang, Wei Chu, Jingdong Chen, and Haibin Ling. Cbnet: A composite backbone network architecture for object detection. IEEE Transactions on Image Processing, 31:6893-6906, 2022. 1, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 526, + 530, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 526, + 530, + 578 + ], + "spans": [ + { + "bbox": [ + 299, + 526, + 530, + 578 + ], + "type": "text", + "content": "[22] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 579, + 530, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 579, + 530, + 622 + ], + "spans": [ + { + "bbox": [ + 299, + 579, + 530, + 622 + ], + "type": "text", + "content": "[23] Zhidong Liang, Ming Zhang, Zehan Zhang, Xian Zhao, and Shiliang Pu. Rangercnn: Towards fast and accurate 3d object detection with range image representation. arXiv preprint arXiv:2009.00206, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 623, + 530, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 623, + 530, + 664 + ], + "spans": [ + { + "bbox": [ + 299, + 623, + 530, + 664 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 665, + 530, + 728 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 665, + 530, + 728 + ], + "spans": [ + { + "bbox": [ + 299, + 665, + 530, + 728 + ], + "type": "text", + "content": "[25] Hao Liu, Zhuoran Xu, Dan Wang, Baofeng Zhang, Guan Wang, Bo Dong, Xin Wen, and Xinyu Xu. Pai3d: Painting adaptive instance-prior for 3d object detection. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 459-475. Springer, 2023. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 301, + 756 + ], + "type": "text", + "content": "17600" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 729 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 170 + ], + "type": "text", + "content": "[26] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 531-548. Springer, 2022. 3, 6, 7, 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 173, + 279, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 279, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 279, + 227 + ], + "type": "text", + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 229, + 279, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 279, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 279, + 280 + ], + "type": "text", + "content": "[28] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. 1, 2, 3, 6, 7, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 283, + 279, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 279, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 279, + 305 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 308, + 279, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 279, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 279, + 361 + ], + "type": "text", + "content": "[30] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3164-3173, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 362, + 279, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 279, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 279, + 405 + ], + "type": "text", + "content": "[31] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 408, + 279, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 279, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 279, + 450 + ], + "type": "text", + "content": "[32] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 452, + 279, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 279, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 279, + 493 + ], + "type": "text", + "content": "[33] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355-607, 2019. 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 497, + 279, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 279, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 279, + 551 + ], + "type": "text", + "content": "[34] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 553, + 279, + 595 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 553, + 279, + 595 + ], + "spans": [ + { + "bbox": [ + 47, + 553, + 279, + 595 + ], + "type": "text", + "content": "[35] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 597, + 279, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 279, + 650 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 279, + 650 + ], + "type": "text", + "content": "[36] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2743-2752, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 653, + 279, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 279, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 279, + 685 + ], + "type": "text", + "content": "[37] Leslie N Smith. Cyclical learning rates for training neural networks. In 2017 IEEE winter conference on applications of computer vision (WACV), pages 464-472. IEEE, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "type": "text", + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 730 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 159 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 159 + ], + "type": "text", + "content": "[39] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 2, 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 161, + 530, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 161, + 530, + 214 + ], + "spans": [ + { + "bbox": [ + 299, + 161, + 530, + 214 + ], + "type": "text", + "content": "[40] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11794-11803, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 216, + 530, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 216, + 530, + 278 + ], + "spans": [ + { + "bbox": [ + 299, + 216, + 530, + 278 + ], + "type": "text", + "content": "[41] Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh, and I-Hau Yeh. Cspnet: A new backbone that can enhance learning capability of cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 390–391, 2020. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 281, + 530, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 281, + 530, + 324 + ], + "spans": [ + { + "bbox": [ + 299, + 281, + 530, + 324 + ], + "type": "text", + "content": "[42] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 325, + 530, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 325, + 530, + 377 + ], + "spans": [ + { + "bbox": [ + 299, + 325, + 530, + 377 + ], + "type": "text", + "content": "[43] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 379, + 530, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 379, + 530, + 442 + ], + "spans": [ + { + "bbox": [ + 299, + 379, + 530, + 442 + ], + "type": "text", + "content": "[44] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. In European Conference on Computer Vision, pages 638-656. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 444, + 530, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 444, + 530, + 487 + ], + "spans": [ + { + "bbox": [ + 299, + 444, + 530, + 487 + ], + "type": "text", + "content": "[45] Junjie Yan, Yingfei Liu, Jianjian Sun, Fan Jia, Shuai Lin Li, Tiancai Wang, and Xiangyu Zhang. Cross modal transformer via coordinates encoding for 3d object detection. arXiv preprint arXiv:2301.01283, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 489, + 530, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 489, + 530, + 509 + ], + "spans": [ + { + "bbox": [ + 299, + 489, + 530, + 509 + ], + "type": "text", + "content": "[46] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 511, + 530, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 511, + 530, + 553 + ], + "spans": [ + { + "bbox": [ + 299, + 511, + 530, + 553 + ], + "type": "text", + "content": "[47] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. arXiv preprint arXiv:2208.11112, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 554, + 530, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 554, + 530, + 596 + ], + "spans": [ + { + "bbox": [ + 299, + 554, + 530, + 596 + ], + "type": "text", + "content": "[48] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 598, + 530, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 598, + 530, + 640 + ], + "spans": [ + { + "bbox": [ + 299, + 598, + 530, + 640 + ], + "type": "text", + "content": "[49] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 642, + 530, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 642, + 530, + 663 + ], + "spans": [ + { + "bbox": [ + 299, + 642, + 530, + 663 + ], + "type": "text", + "content": "[50] Xingyi Zhou, Dequan Wang, and Philipp Krahenbihl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 665, + 530, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 665, + 530, + 707 + ], + "spans": [ + { + "bbox": [ + 299, + 665, + 530, + 707 + ], + "type": "text", + "content": "[51] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 708, + 530, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 708, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 708, + 530, + 730 + ], + "type": "text", + "content": "[52] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "spans": [ + { + "bbox": [ + 277, + 747, + 299, + 756 + ], + "type": "text", + "content": "17601" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 107, + 279, + 170 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 65, + 107, + 278, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 107, + 278, + 127 + ], + "spans": [ + { + "bbox": [ + 65, + 107, + 278, + 127 + ], + "type": "text", + "content": "cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019.6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 130, + 279, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 279, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 279, + 170 + ], + "type": "text", + "content": "[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 4, 5" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "spans": [ + { + "bbox": [ + 278, + 747, + 300, + 756 + ], + "type": "text", + "content": "17602" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/a542f651-0510-4853-91de-1141d304dc5f_content_list.json b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/a542f651-0510-4853-91de-1141d304dc5f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c72ffc64576487313b667065912388a9e04e68e3 --- /dev/null +++ b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/a542f651-0510-4853-91de-1141d304dc5f_content_list.json @@ -0,0 +1,1422 @@ +[ + { + "type": "text", + "text": "SparseMAE: Sparse Training Meets Masked Autoencoders", + "text_level": 1, + "bbox": [ + 186, + 130, + 782, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aojun Zhou $^{1*}$ Yang Li $^{2}$ Zipeng Qin $^{1}$ Jianbo Liu $^{1}$ Junting Pan $^{1}$ Renrui Zhang $^{13}$ Rui Zhao $^{2}$ Peng Gao $^{3}$ Hongsheng Li $^{1}$ $^{1}$ The Chinese University of Hong Kong \n $^{2}$ SenseTime Research Shanghai AI Lab", + "bbox": [ + 204, + 179, + 763, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Masked Autoencoders (MAE) and its variants have proven to be effective for pretraining large-scale Vision Transformers (ViTs). However, small-scale models do not benefit from the pretraining mechanisms due to limited capacity. Sparse training is a method of transferring representations from large models to small ones by pruning unimportant parameters. However, naively combining MAE finetuning with sparse training make the network task-specific, resulting in the loss of task-agnostic knowledge, which is crucial for model generalization. In this paper, we aim to reduce model complexity from large vision transformers pretrained by MAE with assistant of sparse training. We summarize various sparse training methods to prune large vision transformers during MAE pretraining and finetuning stages, and discuss their shortcomings. To improve learning both task-agnostic and task-specific knowledge, we propose SparseMAE, a novel two-stage sparse training method that includes sparse pretraining and sparse finetuning. In sparse pretraining, we dynamically prune a small-scale sub-network from a ViT-Base. During finetuning, the sparse sub-network adaptively changes its topology connections under the task-agnostic knowledge of the full model. Extensive experimental results demonstrate the effectiveness of our method and its superiority on small-scale vision transformers. Code will be available at https://github.com/aojunzz/SparseMAE.", + "bbox": [ + 75, + 318, + 473, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 742, + 209, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, several pretrained Masked Image Modeling (MIM) methods, such as MAE [15] and data2vec [1], have achieved great success in various computer vision tasks. They usually involve a two-stage training scheme where the model learns task-agnostic knowledge through a pretraining pretext task, and is subsequently finetuned on downstream tasks to acquire task-specific knowledge. Among", + "bbox": [ + 75, + 768, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "these models, Masked Autoencoders (MAE) demonstrate superior learning capability on large-scale vision transformers (Fig. 1 (a)), thanks to large models' strong capacity to learn powerful general representations in the pretraining phase and versatile transferability to specific vision tasks. Yet, the large deep models are burdensome and difficult to be deployed to computationally restricted real-world scenarios. On the other hand, smaller and more efficient models, such as ViT-Tiny and ViT-Small, fail to perform well after pretrained by MAE or data2vec frameworks. For example, ViT-Tiny's finetuning results are even inferior to fully-supervised training (see Fig. 1 (b)). These empirical findings suggest that, under MIM frameworks, when the model is scaled down, it becomes more difficult to learn well via masked pretraining due to its limited capacity during pretraining and consequently hinders its performance when transferred to downstream tasks. In the realm of MIM pretraining, model capacity becomes a key ingredient to learning task-agnostic knowledge via the pretext tasks.", + "bbox": [ + 500, + 287, + 893, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In our studies, we try to tackle the issues of MIM pretraining for small-scale vision transformers. Particularly, we apply pruning techniques [28, 20, 4, 42, 23] in the MIM frameworks in order to obtain performant small-scale models from larger ones. As a first step in our exploration, we use existing sparse training methods [42, 23] for MAE pretraining and finetuning on unstructured and hardware-friendly N:M sparsity to obtain a model with a similar scale to ViT-Tiny, which is pruned from ViT-Base (see Tab. 5 row 3). It achieves notably improvements compared to directly training ViT-Tiny using MAE on downstream ImageNet finetuning accuracy (77.6% vs. 71.6%) and ADE20K [43] semantic segmentation (39.8 mIoU vs. 37.6 mIoU). The empirical results indicate that on small-scale ViTs sparse training via pruning has promising applications under the MIM pretraining paradigm. However, there is still a large performance gap between the sparse model and the vanilla ViT-Base trained densely under MAE in their ImageNet finetuning accuracies (77.6% vs. 83.2%). The sparsity constraints adversely limit the model capacity and prevent it from learning good task-agnostic knowledge, similar to the predicament faced", + "bbox": [ + 496, + 583, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The first two authors equally contribute to this paper.", + "bbox": [ + 94, + 886, + 380, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16176", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f44911641785043be511cd7ead7c035e0e55ece70627b6f73874a9925ce4ab57.jpg", + "image_caption": [ + "Figure 1: (a) Top-1 accuracy of large-scale models on ImageNet-1K, showing that large-scale vision transformers and self-supervised training can bring significant improvement. (b) On tiny-scale and small-scale, our Sparse-MAE outperforms other self-supervised methods, such as data2vec [1], MAE [15], and training from scratch." + ], + "image_footnote": [], + "bbox": [ + 81, + 104, + 459, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "by ViT-Tiny when it is directly pretrained under MAE.", + "bbox": [ + 75, + 458, + 437, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To obtain more powerful small-scale models, we propose SparseMAE, a simple and unified sparse training method for Masked Autoencoders. Specifically, SparseMAE trains an adaptive sub-network in conjunction with its corresponding large dense model during pretraining. Both the large model and the sparse sub-network predict the same reconstruction objective and are optimized end-to-end. The large full model provides a strong capacity to learn task-agnostic knowledge, while the sparse sub-network can better preserve the knowledge by rewiring the connections in the full model. To better learn task-specific knowledge in downstream tasks, the sparse sub-network inherits the task-agnostic knowledge from the pretraining stage and also dynamically adjusts its connections during the finetuning phase.", + "bbox": [ + 75, + 474, + 468, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on a variety of tasks, including image classification, object detection, segmentation, and robustness evaluation, to thoroughly evaluate the effectiveness of our proposed approach. Notably, our method achieves a top-1 accuracy of $80.5\\%$ with only 6M parameters on ImageNet-1K [8] classification and obtains a 45.2 mIoU on ADE20K segmentation with UperNet [37], significantly outperforming all previous works that have developed small-scale vision transformers using MAE (e.g., MAE-lite, TinyMIM and G2SD distillation).", + "bbox": [ + 75, + 686, + 468, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In all, our contributions can be summarized in three folds:", + "bbox": [ + 96, + 838, + 470, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We thoroughly study the integration of sparse training as a pruning technique into the MAE framework.", + "bbox": [ + 94, + 869, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we design three different strategies to train sparse networks under MAE and analyze their limitations.", + "bbox": [ + 529, + 90, + 893, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Based on the findings from above, we propose a novel sparse training framework, SparseMAE for Masked Autoencoders to improve the acquisition of both the task-agnostic and task-specific knowledge during the pretraining and finetuning stages.", + "- We present extensive experiments to validate the effectiveness of our proposed method and achieve state-of-the-art performance with small-scale vision transformers on ImageNet classification and various downstream tasks." + ], + "bbox": [ + 517, + 154, + 893, + 323 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 500, + 363, + 645, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Masked Image Modeling", + "text_level": 1, + "bbox": [ + 500, + 388, + 725, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised pretraining on large-scale images with Masked Image Modeling (MIM) has shown superior performance on various computer vision tasks. BEiT [2] explores Masked Image Modeling on vision transformers by reconstructing the vision dictionary [30]. MAE [15] further proposes an asymmetric encoder and decoder for scaling up MIM to huge models. Besides, it demonstrates a simple pixel reconstruction loss can learn good visual representations. Due to the simplicity and computational efficiency, MAE is raising to a popular generative pretraining paradigm [40, 39]. As MAE reconstructs low-level signals with an isotropic vision transformer architecture, researchers improve MAE by exploring high-level signals architectures [1, 35, 12], which are more effective than reconstructing low-level signals.", + "bbox": [ + 496, + 415, + 893, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, those methods demonstrate the large capacity model is essential for good representation learning through the MIM pretext tasks. MAE-lite [34] explores Masked Autoencoders to improve the performance of tiny-scale vision transformer models, and introduces knowledge distillation into the pretraining phase. Different from MAE-lite, TinyMIM [31] proposes sequential relation knowledge distillation to improve the performance of small-scale vision transformer models, which utilize the full set of patch tokens and distill the knowledge from a pretrained ViT-Large model. TinyMIM [31] needs the large pretrained vision transformer and the use of sequential models is time-consuming for practical applications. These methods attempt to utilize the task-agnostic knowledge from large models through knowledge distillation. Different from their approaches, we explore sparse training as a pruning technique in order to scale down ViTs from large-scale to small-scale under the Masked Autoencoders framework.", + "bbox": [ + 496, + 628, + 895, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16177", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/77c57227b5868050a5ca693d6c26dd965956994ef58323d46f154debad09561e.jpg", + "image_caption": [ + "(a) General pretraining and finetuning" + ], + "image_footnote": [], + "bbox": [ + 181, + 85, + 442, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1500364156231e2456e525bf77f61489cd891c6da90098c1a2e82b8d5c51f3fc.jpg", + "image_caption": [ + "(b) General pretraining and sparse finetuning" + ], + "image_footnote": [], + "bbox": [ + 524, + 85, + 779, + 198 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8d11eb4c23c8a14ac62c1ab443c30eff3294cf945747cb125935bc9ef9abf80d.jpg", + "image_caption": [ + "(c) Sparse pretraining and sparse finetuning" + ], + "image_footnote": [], + "bbox": [ + 183, + 273, + 439, + 386 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/85315238ec5b5b4751e1e1eecb5802c439d5d9e6afd8486a1604d914ffa065d2.jpg", + "image_caption": [ + "(d) SparseMAE pretraining and finetuning", + "Figure 2: An illustration of four types pretraining and finetuning frameworks. (a) Dense pretraining and finetuning the large full model as done in MAE [15]. $G$ and $D$ are general task-agnostic vision knowledge and task-specific downstream knowledge learnt from pretraining and finetuning respectively. (b) Learning $G$ during dense pretraining with the large full model and acquiring downstream task-specific knowledge $D^{s}$ for the sparse sub-network during sparse finetuning. (c) Learning task-agnostic knowledge $G^{s}$ and task-agnostic knowledge $D^{s}$ with sparse sub-network during sparse pretraining and finetuning. For finetuning, the sparse mask can be kept as fixed or dynamic. (d) Our SparseMAE training framework learns both $G$ and $G^{s}$ during pretraining and subsequently learns better $D^{s}$ during finetuning." + ], + "image_footnote": [], + "bbox": [ + 529, + 243, + 785, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Pruning and Sparse Training", + "text_level": 1, + "bbox": [ + 76, + 568, + 338, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pruning and sparse training are critical methods for reducing the inference costs of deep neural networks [18]. Post-pruning methods [13, 26] typically involve training a dense full network first, followed by one or several rounds of train-prune-retrain rounds. Post-pruning methods are time-consuming due to the iterative manner [11]. Different from post-training methods, sparse training methods [29, 36] aim to adaptively obtain a sparse sub-network while optimizing both model weights and sparse connections from random initialization.", + "bbox": [ + 75, + 594, + 470, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse training is a class of methods stemming from the Sparse Evolutionary Training (SET) algorithm [29]. It starts from a randomly sparsified network and subsequently prunes and grows connections during training. These methods can be applied end-to-end within the network training stage and have achieved promising results on various vision benchmarks with more efficient models by adaptively changing the model topology connections during training.", + "bbox": [ + 75, + 747, + 468, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse training methods have shown promising results in various computer vision benchmarks, achieving state-of-the", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "art performance on different types of sparsity granularities such as unstructured sparsity [4, 23, 27], N:M fine-grained structured sparsity [42, 32, 10, 41], and block sparsity [7]. However, to our knowledge, the studies of pruning techniques for Vision Transformers [4, 20] only focus on the fully-supervised settings. There still lacks a unified method to prune large-scale Vision Transformers under the unsupervised Masked Autoencoders framework.", + "bbox": [ + 496, + 569, + 893, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 500, + 704, + 598, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first revisit the sparse training technique and Masked Autoencoders (MAE) in Sec. 3.1. Then we emphasize the challenges of integrating existing sparse training methods into MAE in Sec. 3.2. Lastly, we introduce our proposed method to effectively obtain a powerful small-scale sub-network from a large-scale model using sparse training in Sec. 3.3", + "bbox": [ + 496, + 729, + 893, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 845, + 640, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Revisiting Sparse Training. The pioneering work of Lottery Ticket Hypothesis (LTH) [11] demonstrates the dense", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16178", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "networks contain sparse matching subnetworks capable of training in isolation to full accuracy. Its findings inspire searching for strong sub-networks within a larger full network. Sparse training starts from randomly initializing the network weights $\\mathbf{W}$ and associating them with a sparse mask $\\mathbf{M}$ , which is a binary tensor with 0 indexing the sparsified entries and 1 indicating the retained ones. $\\mathbf{M}$ is optimized end-to-end and is usually obtained by magnitude pruning [14]. We denote the sparse sub-network as $\\widetilde{\\mathbf{W}} = \\mathbf{W} \\odot \\mathbf{M}$ . Sparse-to-sparse training regime optimizes the sparse weights $\\widetilde{\\mathbf{W}}$ and the corresponding binary mask $\\mathbf{M}$ , which can be updated at each iteration or periodically. Variants of sparse training, such as unstructured sparsity and fine-grained structured sparsity [23, 25, 43, 9] achieve state-of-the-art performance on various convolutional networks with improved efficiency. Masked Autoencoders. Masked Autoencoders are a family of self-supervised methods for pretraining Vision Transformers (ViT). They involve reconstructing masked RGB patches from visible tokens $x_{\\mathrm{vis}}$ with an encoder $f$ and a decoder $g$ . Mathematically, we have", + "bbox": [ + 75, + 90, + 472, + 396 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M A E}} (\\mathbf {W}) = \\mathbb {E} \\left(\\left\\| g _ {\\theta} \\left(f \\left(x _ {\\text {v i s}}, \\mathbf {W}\\right)\\right) - x _ {\\text {m a s k}} \\right\\| ^ {2}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 422, + 470, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, an encoder $f$ maps inputs $x_{\\mathrm{vis}}$ to latent features $z_{\\mathrm{vis}} = f(x_{\\mathrm{vis}})$ , and a decoder $g_{\\theta}$ parameterized by $\\theta$ predicts the masked token $x_{\\mathrm{mask}}$ in the RGB pixel space from the latent feature $z_{\\mathrm{vis}}$ . Encoder weights $\\mathbf{W}$ and decoder weights $\\theta$ are jointly trained. After pretraining, the decoder is discarded and the encoder is finetuned on task-specific downstream datasets. During finetuning, the model acquires downstream specific knowledge $D$ , which is built on general task-agnostic knowledge $G$ learnt in pretraining, as shown in Fig. 2 (a).", + "bbox": [ + 75, + 452, + 472, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Challenges of Sparse Training for MAE", + "text_level": 1, + "bbox": [ + 76, + 614, + 419, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the effectiveness of MAE for large-scale vision transformers, it encounters great difficulties when applied to smaller models (see Fig. 1). The inadequate model capacity prevents small models from learning meaning task-agnostic knowledge and consequently transfers to inferior performance in downstream tasks. To enable the potential of large models' capability for smaller ones to learn powerful representations under MAE framework, we can modify Eq. 1 to combine it with sparse training as follows:", + "bbox": [ + 75, + 637, + 472, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S p a r s e}} (\\mathbf {W}, \\mathbf {M}) = \\mathbb {E} \\left(\\left\\| g _ {\\theta} \\left(f \\left(x _ {\\text {v i s}}, \\mathbf {W} \\odot \\mathbf {M}\\right)\\right) - x _ {\\text {m a s k}} \\right\\| ^ {2}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 796, + 468, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f$ represents the large full encoder with parameters $\\mathbf{W}$ . The encoder binary mask $\\mathbf{M}$ produces a sparse sub-network by magnitude pruning during pretraining, while the decoder is kept dense as it is not used for downstream finetuning. Following the notation in Sec. 3.1, we use $G^{s}$ to represent", + "bbox": [ + 75, + 824, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the general knowledge of the sparse sub-network learnt in pretraining, and we denote $D^{s}$ the specific knowledge of the finetuned sparse sub-network for downstream tasks. Considering both the pretraining and finetuning phases, we outline three intuitive strategies to integrate sparse training into the MAE framework:", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 1 Dense pretraining then sparse finetuning as shown in Fig. 2 (b). Although this strategy allows the sparse finetuning phase to optimize from the large full model's task-agnostic knowledge $G$ , the sub-network does not benefit from the pretraining task and lacks its knowledge $G^{s}$ . This may hinder the model to acquire better downstream knowledge $D^{s}$", + "- 2&3 Sparse pretraining using Eq. (2) then sparse finetuning (see Fig. 2 (c)). During finetuning, the sparsity mask can be kept as fixed or dynamic (dynamic mask is the third strategy). This strategy relies on the architecture and task-agnostic knowledge $G^{s}$ obtained by the sub-network during the pretraining phase, while it ignores the potential benefits brought by incorporating the knowledge $G$ of the large full model." + ], + "bbox": [ + 517, + 195, + 893, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, it is challenging to tailor a sparse training mechanism for MAE framework to sufficiently learn task-agnostic and task-specific knowledge for a small model. In the following section, we introduce our approach to solve this dilemma by fully utilizing the task-agnostic knowledge $G$ of the large full model and adaptively optimizing sub-network's topology connections for various downstream tasks for better task-specific knowledge.", + "bbox": [ + 496, + 446, + 893, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Sparse Masked Autoencoders", + "text_level": 1, + "bbox": [ + 500, + 577, + 764, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To tackle the limitations of the sparse training strategies introduced above, we propose a novel method for pruning ViT models under the MAE framework (see Fig. 2 (d)).", + "bbox": [ + 496, + 601, + 890, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, to ensure a good general task-agnostic knowledge $G$ to be learnt via pretraining, we utilize the dense full model to reconstruct the RGB pixels following Eq. 1, which learns more discriminative representations thanks to its larger capacity. Simultaneously, a sparse sub-network within the full dense network is trained using the same visible tokens as the inputs and regresses the same targets as the large full model. The sparseMAE loss function is therefore extended from Eq. 1 and Eq. 2 as follows:", + "bbox": [ + 496, + 647, + 893, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\underbrace {\\mathcal {L} _ {\\text {S p a r s e}} (\\widetilde {\\mathbf {W}})} _ {\\text {s u b - n e t w o r k}} + \\underbrace {\\alpha \\mathcal {L} _ {\\mathrm {M A E}} (\\mathbf {W})} _ {\\text {f u l l n e t w o r k}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 795, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}$ represents the dense parameters of the large full model, $\\widetilde{\\mathbf{W}} = \\mathbf{W} \\odot \\mathbf{M}$ is the masked sparse parameters of the sub-network, and $\\alpha$ is the scaling hyperparameter balancing the reconstruction losses for the dense full model and the", + "bbox": [ + 496, + 838, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16179", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/22ed134c47fa750236b62ab04a8494baffb2af9d50b552dbe5234ef88e90f38a.jpg", + "image_caption": [ + "Figure 3: Illustration of Sparse Masked Autoencoders. In the forward pass, both the large full model (red solid lines) and the sparse sub-network (red dashed lines) reconstruct the masked patches from the same visible tokens. In the backward pass, the masked weights (green solid lines) and the sparse weights (green dashed lines) are updated separately using the gradients from the corresponding losses." + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 831, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sparse sub-network. In our experiments, we set $\\alpha = 1.0$ by default.", + "bbox": [ + 75, + 503, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Eq. 3, the gradients for the sparse weights $\\widetilde{\\mathbf{W}}$ shall be updated following both the reconstruction losses of the sub-network and the full network, which may lead to gradient imbalance between the masked and unmasked weights [38]. Therefore, in each training iteration we split the weights $\\mathbf{W}$ of the full model into sparse sub-network weights $\\widetilde{\\mathbf{W}}$ and unmasked weights $\\overline{\\mathbf{W}} = \\mathbf{W} \\odot (1 - \\mathbf{M})$ and optimize them separately:", + "bbox": [ + 75, + 534, + 472, + 659 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathbf {W}} _ {t + 1} = \\widetilde {\\mathbf {W}} _ {t} - \\eta (g (\\mathbf {W} _ {t} \\odot \\mathbf {M} _ {t}) \\odot \\mathbf {M} _ {t}), \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 672, + 472, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {W}}} _ {t + 1} = \\overline {{\\mathbf {W}}} _ {t} - \\eta (g (\\mathbf {W} _ {t}) \\odot (1 - \\mathbf {M} _ {t})), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 715, + 470, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g(\\mathbf{W}_t \\odot \\mathbf{M}_t) = \\frac{\\partial \\mathcal{L}_{\\mathrm{Sparse}}(\\mathbf{W}_t \\odot \\mathbf{M}_t)}{\\partial \\mathbf{W}_t}$ are gradients of the sparse sub-network, $g(\\mathbf{W}_t) = \\frac{\\partial \\mathcal{L}_{\\mathrm{MAE}}(\\mathbf{W}_t)}{\\partial \\mathbf{W}_t}$ are the gradients of the full models, $\\eta$ is the learning rate and $t$ indexes the training iteration. The binary mask $\\mathbf{M}_t$ is obtained by on-the-fly magnitude pruning. Here, we only show gradient descent for simplicity and omit other parts (e.g., weight decay). In this way, the large full model can utilize all the model parameters in the forward pass to produce discriminative outputs, while in the backward pass, the sparse sub-network stays unaffected by the gradients from the large full model. During", + "bbox": [ + 75, + 741, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the pretraining phase, the task-agnostic knowledge $G^{s}$ of the sparse sub-network can dynamically select the important sparse connections in the context of the task-agnostic knowledge $G$ of the full dense model, which helps to identify the important sparse topology and is retained for downstream finetuning.", + "bbox": [ + 496, + 503, + 893, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After pretraining, the full dense model with a sparse sub-network is transferred and finetuned on different downstream tasks (e.g., classification, detection, segmentation, etc.). For the finetuning stage, the deep representations from the sparse sub-network are used for the downstream tasks, while its topology connections are optimized end-to-end given the specific downstream objectives. We refrain from fine-tuning dense full models, as the fine-tuning process for both dense and sparse models is costly. Hence, the learning of the task-specific knowledge $D^{s}$ for downstream tasks can benefit from inheriting a pretrained sparse sub-network with knowledge $G^{s}$ , and adaptively selecting the more suitable sparse architecture built on the general knowledge $G$ of the full model.", + "bbox": [ + 496, + 595, + 895, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 827, + 633, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we validate the effectiveness and robustness of our SparseMAE. First, we introduce the implementation details of our SparseMAE in Sec. 4.1. Then, we present", + "bbox": [ + 496, + 854, + 895, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16180", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "our main experimental results in Sec. 4.2, which include classification results on ImageNet-1K [8] and semantic segmentation results on ADE20K [43]. To demonstrate the superiority of our proposed method, we compare SparseMAE with other state-of-the-art ViT pruning methods in Sec. 4.4. In Sec. 4.3, we evaluate the transferring ability of the models trained with our SparseMAE to object detection, instance segmentation tasks on MS COCO [24] and robustness evaluation. Finally, we perform ablative experiments to evaluate the performance of SparseMAE under different settings in Sec. 4.5.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 268, + 294, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sparse Training Methods. We conduct experiments with two different granularities for sparse training, consisting of unstructured sparsity [14, 21] and hardware-friendly N:M sparsity [28, 42]. As we focus on studying how to learn and transfer task-agnostic and task-specific knowledge under the Masked Autoencoders paradigm with sparse training methods, we select two simple and effective sparse training methods, namely DFP [23] for unstructured sparsity and SR-STE [42] for N:M sparsity. We prune all the linear layers within the vision transformers and set the relevant hyperparameters for sparse training the same as [23, 42] by default.", + "bbox": [ + 75, + 292, + 472, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. For all of our main experiments, unless otherwise stated, we select ViT-Base as our large full model and prune it for smaller ones. For hardware-friendly N:M sparsity, we use 2:32 and 4:32 sparse ratios to obtain models with similar scales to ViT-Tiny (5.8M) and ViT-Small (11.3M), the main experiments use N:M sparsity for default, which is applied to all linear layers, including self.qkv, self.proj in the Attention block, and self.fc1 and self.fc2 in the MLP block. For unstructured sparsity, we extend the above N:M ratios and prune ViT-Base with $93.5\\%$ and $87.0\\%$ sparsity. We name our ViT-Base with 2:32/93.5% sparsity as SparseMAE-Tiny (SparseMAE-T) and the one with 4:32/87.0% sparsity as SparseMAE-Small (SparseMAE-S).", + "bbox": [ + 75, + 476, + 472, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretraining. Following MAE [15] pretraining scheme, models start from random initialization, we use ImageNet-1K (IN-1K) [8] for all pretraining experiments. We pretrain SparseMAE for 400 epochs using $224 \\times 224$ images as inputs. The other settings also follow those used in [15]. For example, We use random resized cropping and horizontal flipping for data augmentation, a $75\\%$ mask ratio during pretraining, and a decoder with a single transformer layer and 512 hidden dimensions. Unlike other methods [34, 19] for small and tiny-scale vision transformers pretraining, we do not grid-search the decoder design, possible putting our method at a disadvantage.", + "bbox": [ + 75, + 672, + 472, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finetuning. We evaluate our SparseMAE pretrained models for various downstream tasks, including ImageNet-1K [8] image classification, MS COCO [24] object detection and", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "instance segmentation tasks, ADE20K [43] semantic segmentation.", + "bbox": [ + 496, + 90, + 893, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For image classification, we take a ViT-Base model pretrained using MAE for 400 epochs as the reference model, which is implemented using the officially released codebase [15] and achieves $83.2\\%$ top-1 accuracy. We finetune SparseMAE-Tiny/Small for 200 epochs using the same receipt described in [15] for a fair comparison with other works [31, 19] and set the batch size as 1024 and the base learning rate as $10^{-3}$ .", + "bbox": [ + 496, + 121, + 895, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For semantic segmentation, we follow BEiT [2] to use UperNet[37] as our segmentation framework. We train the model with an input resolution of $512 \\times 512$ for 160k iterations. A batch size of 16 is adopted. For object detection and instance segmentation tasks, we follow ViTDet [22] framework. We train our SparseMAE models with a batch size of 64 for 100 epochs, and the input image resolution is set to $1024 \\times 1024$ . Further details on the hyper-parameters can be found in Appendix.", + "bbox": [ + 496, + 243, + 895, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 498, + 391, + 638, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Classification and Semantic Segmentation. Tab. 1 shows the evaluation of our SparseMAE model against self-supervised methods for ViT-Tiny/Small on ImageNet-1K image classification and ADE20K semantic segmentation. For ImageNet-1K classification, SparseMAE achieves $80.5\\%$ top-1 accuracy. To our knowledge, this is currently the best result achieved by tiny-scale transformer models. Furthermore, SparseMAE outperforms TinyMIM [31]/G2SD [19]/MAE-lite [34] with $+4.7 / 4.2 / 4.0$ accuracy gains on tiny-scale transformers. For semantic segmentation, SparseMAE achieves an mIoU of 45.2, outperforming G2SD/MAE-lite/TinyMIM by $+3.8 / 1.3 / 1.2$ mIoUs. Similar improvements can be observed for the small-scale transformers. SparseMAE achieves a $83.2\\%$ ImageNet-1K classification accuracy and a $48.4\\%$ mIoU on ADE20K semantic segmentation, surpassing TinyMIM [31]/G2SD [19] by $+0.2 / 1.2$ and $+0.0 / 2.2$ on the two tasks respectively. The significant improvements over previous results show the superiority of our method in obtaining performant small models under MAE framework.", + "bbox": [ + 496, + 415, + 895, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Speedups. Unstructured sparsity and N:M structured sparsity is efficient for commercial CPUs and FPGAs respectively. For our SparseMAE with unstructured sparsity, we tested the throughput (img/s) on an Intel RXeon(R) 6238R CPU using the deepsparse engine with a batch size of 64. For our SparseMAE with N:M sparsity, we tested the throughput on Xilinx XCVU13P with a batch size of 16. Our SparseMAE-S with N:M and unstructured sparsity achieved throughput gains of $+71\\%$ and $+9\\%$ over ViT-small with similar amount of parameters on FPGAs and CPUs. Therefore, our approach demonstrate high feasibility in actual scenarios.", + "bbox": [ + 496, + 719, + 895, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16181", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8688c59b47af17d59c84e564172d530120a1d77c71f16c5c8a953654c2483b04.jpg", + "table_caption": [ + "Table 1: Finetuning results on ImageNet-1K and ADE20K. All our models are pre-trained only on ImageNet-1K. Extra pretrained teacher means the compared method cannot trained from scratch. $\\ddagger$ the model adopts the MAE pretrained model before performing the sparse pretraining." + ], + "table_footnote": [], + "table_body": "
MethodParams (M)Pretraining epochsEncoder ratioExtra Pretrained/ TeacherClassification Top-1 Acc (%)Segmentation mIoU (%)
DeiT [33]5.8300-Label72.238.0
Tiny-scale
MAE [15]5.8160025%X71.637.6
MoCo [6]5.8300100%EMA73.339.3
TinyMIM [31]5.8300100%TinyMIM-S75.844.0
G2SD [19]5.8300100%ViT-B76.341.4
MAE-lite [34]5.840025%ViT-B76.543.9
SparseMAE (ours)5.840025%X80.545.2
DeiT [33]22300-Label79.9
Small-scale
MAE [15]22160025%X80.642.8
MoCo [6]22300100%EMA81.443.9
DINO [3]22300100%EMA81.545.3
CAE [5]22300100%DALL-E82.0-
TinyMIM [31]22300100%TinyMIM-ViT-B83.048.4
G2SD [19]2230025%ViT-B82.046.2
SparseMAE (ours)11.330025%ViT-B‡83.248.4
SparseMAE (ours)11.340025%X82.146.7
", + "bbox": [ + 127, + 146, + 839, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation on Other Downstream Tasks", + "text_level": 1, + "bbox": [ + 76, + 516, + 421, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object Detection and Instance Segmentation. We finetune the SparseMAE pretrained models on MS COCO with ViT-Det [22]. As shown in Tab. 2, we report $\\mathrm{AP^{bbox}}$ for object detection and $\\mathrm{AP^{mask}}$ for instance segmentation. We compare SparseMAE-Tiny with MAE-Tiny and its improved variants using distillation methods, such as MAE-lite [34] and G2SD [19]. The results in Tab. 2 show that our Sparse-MAE obtains more than $3.1\\mathrm{AP^{bbox}}$ and $2.4\\mathrm{AP^{mask}}$ gains compared to MAE-lite and G2SD, validating the transferability of the learnt representations in our models.", + "bbox": [ + 75, + 541, + 470, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness Evaluation. In Tab. 3, we evaluate the robustness of our models on different variants of ImageNet validation sets. We use the same models finetuned on original ImageNet (Tab. 1) and only run inference on the different validation sets, such asImagenet-R [17], ImageNet-A [17] and ImageNet-C [16]), without any specialized fine-tuning. Tab. 3 shows that our method has better generalization capability than distillation-based methods, such as G2SD [19] and TinyMIM [31]).", + "bbox": [ + 75, + 694, + 470, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison with State-of-the-art ViT Pruning", + "text_level": 1, + "bbox": [ + 76, + 844, + 470, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To the best of our knowledge, our proposed SparseMAE is the first to study network pruning technique under the", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1eba3410e6487933c65f1487e26c6c6143cfa0827f669bbcee18321ec4cfc1f6.jpg", + "table_caption": [ + "Table 2: Object detection and instance segmentation results on the MS COCO dataset." + ], + "table_footnote": [], + "table_body": "
MethodBackbone Size (M)Det APbboxSeg APmask
Tiny-scale models
DeiT [33]5.840.736.5
MAE [15]5.838.935.1
MoCo v3 [6]5.840.036.0
DINO [3]5.840.236.1
MAE-lite [34]5.842.738.2
G2SD [19]5.844.039.6
SparseMAE (ours)5.847.142.0
", + "bbox": [ + 535, + 556, + 854, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Masked Autoencoders paradigm. To validate the effectiveness of SparseMAE, we compare it against recent state-of-the-art fully-supervised pruning methods for ViTs. We choose SViT [4] and oViT [20] as our compared models, which exploit end-to-end sparse training for ViTs and achieve state-of-the-art results in the fully-supervised setting. SViT [4] adopts a dynamic sparse training method to prune the dense model using random initialization. However, this approach requires training the sparse model for", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16182", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a625b808a20b8bc87828cc770f3d19252c12cafeb1c4918a379a797e2740fdb4.jpg", + "table_caption": [ + "Table 3: Robustness evaluation on ImageNet variants (top-1 accuracy, except for ImageNet-C which evaluates mean corruption error). We test the small-scale and tiny-scale models from Tab. 1 on different ImageNet validation sets, without any fine-tuning." + ], + "table_footnote": [], + "table_body": "
MethodModel SizeIN-A↑IN-R↑IN-C↓
DeiT [33]8.032.754.0
MAE [15]7.036.555.2
TinyMIM [31]Tiny-scale11.039.850.1
G2SD [19]12.939.0-
SparseMAE (ours)18.245.947.2
DeiT [33]18.342.341.4
MAE [15]20.145.640.6
TinyMIM [31]Small-scale27.548.835.8
G2SD [19]29.446.8-
SparseMAE (ours)29.349.235.2
", + "bbox": [ + 78, + 176, + 467, + 366 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69eb1cb78818b3b08d71a22b9ad00a807ab091ef6d216a5aae2caedcb391e605.jpg", + "table_caption": [ + "Table 4: Comparison with state-of-the-art supervised Vision Transformer pruning methods on ImageNet-1K." + ], + "table_footnote": [], + "table_body": "
ModelMethodParams (M)Top-1 Acc (%)
ViT-BaseSViT [4]43.381.5
oViT [20]43.381.6
SViT [4]34.681.3
oViT [20]34.681.5
oViT [20]21.681.1
oViT [20]17.380.8
oViT [20]8.779.7
ViT-BaseSparseMAE-T (ours)5.880.5
SparseMAE-S (ours)11.383.2
", + "bbox": [ + 96, + 422, + 447, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "600 epochs. oViT [20] (post-training pruning) prunes the model initialized from a trained dense model (300 epochs), which requires 300 epochs of finetuning. For our SparseMAE, we only need 400 epochs for pretraining and 200 epochs for fine-tuning. The pretraining of SparseMAE is $3 \\times$ faster than full-token supervised training. Consequently, our SparseMAE approach reduces the training time from 968 GPU hours to 581 GPU hours compared to SOTA SViT and oViT.", + "bbox": [ + 75, + 595, + 468, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 4 demonstrates that SparseMAE can outperform state-of-the-art fully-supervised pruning counterparts significantly for small and tiny-scale transformers. Our SparseMAE-T has a $0.8\\%$ performance gain and a $35.0\\%$ computation complexity reduction compared to oViT [20] for tiny-scale transformers. With our SparseMAE-S, we achieve a $83.2\\%$ ImageNet classification accuracy, outperforming supervised pruning techniques by large margins. These results suggest that our method successfully extends the superiority of MAE for representation learning with large-scale vision transformers to small and tiny-scale models.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: Different sparse pretraining or finetuning strategies with tiny-scale models on ImageNet-1K and ADE20K. \"S\" means sparse pretraining, \"D\" means dense pretraining, and \"D+S\" means sparse pretraining together with a dense full model pretraining. $\\ddagger$ : the model is finetuned with the fixed sparse masks from pretraining. Corr. Strategy.: corresponding sparse training strategies explained in Sec. 4.5.1.", + "bbox": [ + 496, + 88, + 893, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b4d5f52c47841f642a8b272cde1c5c3075ab0a6ff19832a908461fc6ed3d5ad9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPretrain MethodTop-1 Acc (%)Seg mIoU (%)Corr. Strtg.
ViT-BaseD83.248.1-
Unstructured sparsity (93.5% sparse ratio)
SparseMAE-TD77.840.2#1
SparseMAE-TS‡76.737.7#2
SparseMAE-TS77.639.8#3
SparseMAE-TD+S80.445.4-
2:32 sparsity
SparseMAE-TD78.440.7#1
SparseMAE-TS‡76.337.2#2
SparseMAE-TS78.741.3#3
SparseMAE-TD+S80.545.2-
", + "bbox": [ + 517, + 205, + 870, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies and Discussion", + "text_level": 1, + "bbox": [ + 500, + 424, + 782, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.1 Comparison of Sparse Training Strategies.", + "text_level": 1, + "bbox": [ + 500, + 450, + 849, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare three sparse training strategies for MAE framework introduced in Sec. 3.2 under two sparse granularities. Namely, we (1) pretrain a large full model using MAE and finetune it with sparse training, (2) combine sparse training with MAE pretraining and subsequently finetune the model on downstream tasks with the fixed sparse mask from pretraining, or (3) modify (2) to use dynamic sparse mask for finetuning. The results are shown in Tab. 5. Under unstructured sparsity, comparing (1) and (2), we see that the classification performance drops from $77.8\\%$ to $76.7\\%$ , showing that the knowledge $G^{s}$ learnt by the sparse sub-network during pretraining cannot generalize well to the downstream tasks compared to the task-agnostic knowledge $G$ learnt by the large full model. (3) further allows the adaptation of sparse sub-network architecture during finetuning and partially alleviates the problem by better acquiring downstream specific knowledge $D^{s}$ . However, its result still lags significantly behind our proposed method ( $77.6\\%$ vs. $80.4\\%$ ), where the difference is only the absence of a concurrently pretrained large full model. On the other hand, when comparing (1) with our method, the addition of a pretrained sparse sub-network can boost the performance by $2.6\\%$ . Similar observations can also be made under N:M sparsity settings and semantic segmentation results. These empirical results suggest that both the task-agnostic knowledge $G$ of the large full model and $G^{s}$ of the sparse sub-network are necessary for small and tiny-scale vision transformers to effectively benefit from MAE pretraining.", + "bbox": [ + 496, + 477, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16183", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c67e294ceb39929000ef00817ca07e6363b06a37367e134269cb58b3f73a10c6.jpg", + "table_caption": [ + "Table 6: Comparison between different full models for SparseMAE on ImageNet-1K and ADE20K." + ], + "table_footnote": [], + "table_body": "
ModelPatternParams (M)Top-1 Acc (%)Seg mIoU (%)
ViT-Small8:325.877.441.5
ViT-Base2:325.880.545.2
", + "bbox": [ + 91, + 130, + 455, + 209 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5.2 Ablation of Full Model Capacity", + "text_level": 1, + "bbox": [ + 76, + 234, + 357, + 250 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct an ablation study to investigate the performance with different capacities of the full model. We use our Sparse-MAE and set ViT-Small and ViT-Base as the full models, where ViT-Small only has $25\\%$ parameters of ViT-Base. As shown in Tab. 6, the results show that our SparseMAE in ViT-Base achieves $3.1\\%$ top-1 accuracy advantage over ViT-Small, indicating that larger capacity of the full model with SparseMAE can significantly improve its performance by incorporating better pretraining task-agnostic knowledge.", + "bbox": [ + 75, + 258, + 472, + 396 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5.3 Impact of Different Pruning Sparsity Ratios", + "text_level": 1, + "bbox": [ + 76, + 415, + 437, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The main experiments of sparse ratio setting is compared to the ViT-tiny and ViT-small. We provide additional sparse ratios to verify the performance of SparseMAE on ImageNet-1K as in Tab. 7.", + "bbox": [ + 75, + 439, + 470, + 501 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/3fd68e59e3e9a8f9ce70c1a26dec4097d9ef46a6e1a25978b10facf17d9e2ff4.jpg", + "table_caption": [ + "Table 7: Comparison between different sparse ratios for SparseMAE on ImageNet-1K." + ], + "table_footnote": [], + "table_body": "
Sparse Pattern1:322:324:328:32
Top-1 Acc76.880.583.283.7
", + "bbox": [ + 133, + 553, + 410, + 601 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 632, + 194, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we introduce SparseMAE, an innovative method for incorporating sparse training into the Masked Autoencoders (MAE) framework to create efficient and powerful small and tiny-scale transformers. As the first study in this area, we first point out the limitations of sparse training under MAE and design a unified framework to transfer the task-agnostic and task-specific knowledge of large models to lightweight sparse sub-networks. SparseMAE trains the sparse sub-networks by persevering the task-agnostic knowledge of large full models and then adaptively finds the optimal specific sub-networks for downstream tasks. The proposed method outperforms the state-of-the-art small-scale vision transforms in both unsupervised pretraining and fully-supervised settings. We hope this work could shed light on sparse training with the Masked Autoencoders framework in the vision community.", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 89, + 660, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK.", + "bbox": [ + 496, + 114, + 893, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 234, + 596, + 251 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022.", + "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021.", + "[4] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34:19974-19988, 2021.", + "[5] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022.", + "[6] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021.", + "[7] Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. arXiv preprint arXiv:2205.14135, 2022.", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[9] Utku Evci, Trevor Gale, Jacob Menick, Pablo Samuel Castro, and Erich Olsen. Rigging the lottery: Making all tickets winners. In International Conference on Machine Learning, pages 2943-2952. PMLR, 2020.", + "[10] Chao Fang, Aojun Zhou, and Zhongfeng Wang. An algorithm–hardware co-optimized framework for accelerating n: M sparse transformers. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 30(11):1573–1586, 2022.", + "[11] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635, 2018." + ], + "bbox": [ + 501, + 260, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16184", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Peng Gao, Renrui Zhang, Rongyao Fang, Ziyi Lin, Hongyang Li, Hongsheng Li, and Qiao Yu. Mimic before reconstruct: Enhancing masked autoencoders with feature mimicking. IJCV 2023, 2023.", + "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015.", + "[14] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015.", + "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022.", + "[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019.", + "[17] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021.", + "[18] Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks. The Journal of Machine Learning Research, 22(1):10882-11005, 2021.", + "[19] Wei Huang, Zhiliang Peng, Li Dong, Furu Wei, Jianbin Jiao, and Qixiang Ye. Generic-to-specific distillation of masked autoencoders. arXiv preprint arXiv:2302.14771, 2023.", + "[20] Denis Kuznedelev, Eldar Kurtic, Elias Frantar, and Dan Alistarh. ovit: An accurate second-order pruning framework for vision transformers. arXiv preprint arXiv:2210.09223, 2022.", + "[21] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. Advances in neural information processing systems, 2, 1989.", + "[22] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 280-296. Springer, 2022.", + "[23] Tao Lin, Sebastian U Stich, Luis Barba, Daniil Dmitriev, and Martin Jaggi. Dynamic model pruning with feedback. arXiv preprint arXiv:2006.07253, 2020.", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014.", + "[25] Shiwei Liu, Lu Yin, Decebal Constantin Mocanu, and Mykola Pechenizkiy. Do we actually need dense overparameterization? in-time over-parameterization in sparse training. In International Conference on Machine Learning, pages 6989-7000. PMLR, 2021." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018.", + "[27] Zhi-Gang Liu, Paul N Whatmough, and Matthew Mattina. Sparse systolic tensor array for efficient cnn hardware acceleration. arXiv preprint arXiv:2009.02381, 2020.", + "[28] Asit Mishra, Jorge Albericio Latorre, Jeff Pool, Darko Stosic, Dusan Stosic, Ganesh Venkatesh, Chong Yu, and Paulius Miccikevicius. Accelerating sparse deep neural networks. arXiv preprint arXiv:2104.08378, 2021.", + "[29] Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H Nguyen, Madeleine Gibescu, and Antonio Liotta. Scalable training of artificial neural networks with adaptive sparse connectivity inspired by network science. Nature communications, 9(1):2383, 2018.", + "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021.", + "[31] Sucheng Ren, Fangyun Wei, Zheng Zhang, and Han Hu. Tinymim: An empirical study of distilling mim pre-trained models. arXiv preprint arXiv:2301.01296, 2023.", + "[32] Wei Sun, Aojun Zhou, Sander Stuijk, Rob G. J. Wijnhoven, Andrew Nelson, Hongsheng Li, and Henk Corporaal. Dominosearch: Find layer-wise fine-grained n:m sparse schemes from dense neural networks. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021.", + "[33] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021.", + "[34] Shaoru Wang, Jin Gao, Zeming Li, Jian Sun, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. arXiv preprint arXiv:2205.14443, 2022.", + "[35] Chen Wei, Haoqi Fan, Saining Xie, Chao Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. 2021.", + "[36] Mitchell Wortsman, Ali Farhadi, and Mohammad Rastegari. Discovering neural wireings. Advances in Neural Information Processing Systems, 32, 2019.", + "[37] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018.", + "[38] Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. Slimmable neural networks. arXiv preprint arXiv:1812.08928, 2018.", + "[39] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. NeurIPS 2022, 2022.", + "[40] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. CVPR 2023, 2023." + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16185", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yuxin Zhang, Mingbao Lin, Zhihang Lin, Yiting Luo, Ke Li, Fei Chao, Yongjian Wu, and Rongrong Ji. Learning best combination for efficient n:m sparsity. arXiv preprint arXiv:2206.06662, 2022.", + "[42] Aojun Zhou, Yukun Ma, Junnan Zhu, Jianbo Liu, Zhijie Zhang, Kun Yuan, Wenxiu Sun, and Hongsheng Li. Learning n: M fine-grained structured sparse neural networks from scratch. arXiv preprint arXiv:2102.04010, 2021.", + "[43] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + ], + "bbox": [ + 78, + 90, + 470, + 260 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16186", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/a542f651-0510-4853-91de-1141d304dc5f_model.json b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/a542f651-0510-4853-91de-1141d304dc5f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6d2359402d517bed869bbbc489fd7b1169068ea6 --- /dev/null +++ b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/a542f651-0510-4853-91de-1141d304dc5f_model.json @@ -0,0 +1,1960 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.783, + 0.154 + ], + "angle": 0, + "content": "SparseMAE: Sparse Training Meets Masked Autoencoders" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.18, + 0.764, + 0.253 + ], + "angle": 0, + "content": "Aojun Zhou\\(^{1*}\\) Yang Li\\(^{2}\\) Zipeng Qin\\(^{1}\\) Jianbo Liu\\(^{1}\\) Junting Pan\\(^{1}\\) Renrui Zhang\\(^{13}\\) Rui Zhao\\(^{2}\\) Peng Gao\\(^{3}\\) Hongsheng Li\\(^{1}\\) \n\\(^{1}\\)The Chinese University of Hong Kong \n\\(^{2}\\)SenseTime Research Shanghai AI Lab" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.712 + ], + "angle": 0, + "content": "Masked Autoencoders (MAE) and its variants have proven to be effective for pretraining large-scale Vision Transformers (ViTs). However, small-scale models do not benefit from the pretraining mechanisms due to limited capacity. Sparse training is a method of transferring representations from large models to small ones by pruning unimportant parameters. However, naively combining MAE finetuning with sparse training make the network task-specific, resulting in the loss of task-agnostic knowledge, which is crucial for model generalization. In this paper, we aim to reduce model complexity from large vision transformers pretrained by MAE with assistant of sparse training. We summarize various sparse training methods to prune large vision transformers during MAE pretraining and finetuning stages, and discuss their shortcomings. To improve learning both task-agnostic and task-specific knowledge, we propose SparseMAE, a novel two-stage sparse training method that includes sparse pretraining and sparse finetuning. In sparse pretraining, we dynamically prune a small-scale sub-network from a ViT-Base. During finetuning, the sparse sub-network adaptively changes its topology connections under the task-agnostic knowledge of the full model. Extensive experimental results demonstrate the effectiveness of our method and its superiority on small-scale vision transformers. Code will be available at https://github.com/aojunzz/SparseMAE." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.743, + 0.21, + 0.759 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.472, + 0.876 + ], + "angle": 0, + "content": "Recently, several pretrained Masked Image Modeling (MIM) methods, such as MAE [15] and data2vec [1], have achieved great success in various computer vision tasks. They usually involve a two-stage training scheme where the model learns task-agnostic knowledge through a pretraining pretext task, and is subsequently finetuned on downstream tasks to acquire task-specific knowledge. Among" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.895, + 0.576 + ], + "angle": 0, + "content": "these models, Masked Autoencoders (MAE) demonstrate superior learning capability on large-scale vision transformers (Fig. 1 (a)), thanks to large models' strong capacity to learn powerful general representations in the pretraining phase and versatile transferability to specific vision tasks. Yet, the large deep models are burdensome and difficult to be deployed to computationally restricted real-world scenarios. On the other hand, smaller and more efficient models, such as ViT-Tiny and ViT-Small, fail to perform well after pretrained by MAE or data2vec frameworks. For example, ViT-Tiny's finetuning results are even inferior to fully-supervised training (see Fig. 1 (b)). These empirical findings suggest that, under MIM frameworks, when the model is scaled down, it becomes more difficult to learn well via masked pretraining due to its limited capacity during pretraining and consequently hinders its performance when transferred to downstream tasks. In the realm of MIM pretraining, model capacity becomes a key ingredient to learning task-agnostic knowledge via the pretext tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.897, + 0.902 + ], + "angle": 0, + "content": "In our studies, we try to tackle the issues of MIM pretraining for small-scale vision transformers. Particularly, we apply pruning techniques [28, 20, 4, 42, 23] in the MIM frameworks in order to obtain performant small-scale models from larger ones. As a first step in our exploration, we use existing sparse training methods [42, 23] for MAE pretraining and finetuning on unstructured and hardware-friendly N:M sparsity to obtain a model with a similar scale to ViT-Tiny, which is pruned from ViT-Base (see Tab. 5 row 3). It achieves notably improvements compared to directly training ViT-Tiny using MAE on downstream ImageNet finetuning accuracy (77.6% vs. 71.6%) and ADE20K [43] semantic segmentation (39.8 mIoU vs. 37.6 mIoU). The empirical results indicate that on small-scale ViTs sparse training via pruning has promising applications under the MIM pretraining paradigm. However, there is still a large performance gap between the sparse model and the vanilla ViT-Base trained densely under MAE in their ImageNet finetuning accuracies (77.6% vs. 83.2%). The sparsity constraints adversely limit the model capacity and prevent it from learning good task-agnostic knowledge, similar to the predicament faced" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.382, + 0.902 + ], + "angle": 0, + "content": "*The first two authors equally contribute to this paper." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "16176" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.106, + 0.46, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.472, + 0.432 + ], + "angle": 0, + "content": "Figure 1: (a) Top-1 accuracy of large-scale models on ImageNet-1K, showing that large-scale vision transformers and self-supervised training can bring significant improvement. (b) On tiny-scale and small-scale, our Sparse-MAE outperforms other self-supervised methods, such as data2vec [1], MAE [15], and training from scratch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.439, + 0.474 + ], + "angle": 0, + "content": "by ViT-Tiny when it is directly pretrained under MAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.686 + ], + "angle": 0, + "content": "To obtain more powerful small-scale models, we propose SparseMAE, a simple and unified sparse training method for Masked Autoencoders. Specifically, SparseMAE trains an adaptive sub-network in conjunction with its corresponding large dense model during pretraining. Both the large model and the sparse sub-network predict the same reconstruction objective and are optimized end-to-end. The large full model provides a strong capacity to learn task-agnostic knowledge, while the sparse sub-network can better preserve the knowledge by rewiring the connections in the full model. To better learn task-specific knowledge in downstream tasks, the sparse sub-network inherits the task-agnostic knowledge from the pretraining stage and also dynamically adjusts its connections during the finetuning phase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.837 + ], + "angle": 0, + "content": "We conduct extensive experiments on a variety of tasks, including image classification, object detection, segmentation, and robustness evaluation, to thoroughly evaluate the effectiveness of our proposed approach. Notably, our method achieves a top-1 accuracy of \\(80.5\\%\\) with only 6M parameters on ImageNet-1K [8] classification and obtains a 45.2 mIoU on ADE20K segmentation with UperNet [37], significantly outperforming all previous works that have developed small-scale vision transformers using MAE (e.g., MAE-lite, TinyMIM and G2SD distillation)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.839, + 0.471, + 0.854 + ], + "angle": 0, + "content": "In all, our contributions can be summarized in three folds:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.473, + 0.902 + ], + "angle": 0, + "content": "- We thoroughly study the integration of sparse training as a pruning technique into the MAE framework." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.894, + 0.137 + ], + "angle": 0, + "content": "Specifically, we design three different strategies to train sparse networks under MAE and analyze their limitations." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.155, + 0.892, + 0.232 + ], + "angle": 0, + "content": "- Based on the findings from above, we propose a novel sparse training framework, SparseMAE for Masked Autoencoders to improve the acquisition of both the task-agnostic and task-specific knowledge during the pretraining and finetuning stages." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.25, + 0.895, + 0.324 + ], + "angle": 0, + "content": "- We present extensive experiments to validate the effectiveness of our proposed method and achieve state-of-the-art performance with small-scale vision transformers on ImageNet classification and various downstream tasks." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.155, + 0.895, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.364, + 0.646, + 0.379 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.39, + 0.727, + 0.407 + ], + "angle": 0, + "content": "2.1. Masked Image Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.627 + ], + "angle": 0, + "content": "Unsupervised pretraining on large-scale images with Masked Image Modeling (MIM) has shown superior performance on various computer vision tasks. BEiT [2] explores Masked Image Modeling on vision transformers by reconstructing the vision dictionary [30]. MAE [15] further proposes an asymmetric encoder and decoder for scaling up MIM to huge models. Besides, it demonstrates a simple pixel reconstruction loss can learn good visual representations. Due to the simplicity and computational efficiency, MAE is raising to a popular generative pretraining paradigm [40, 39]. As MAE reconstructs low-level signals with an isotropic vision transformer architecture, researchers improve MAE by exploring high-level signals architectures [1, 35, 12], which are more effective than reconstructing low-level signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.897, + 0.901 + ], + "angle": 0, + "content": "However, those methods demonstrate the large capacity model is essential for good representation learning through the MIM pretext tasks. MAE-lite [34] explores Masked Autoencoders to improve the performance of tiny-scale vision transformer models, and introduces knowledge distillation into the pretraining phase. Different from MAE-lite, TinyMIM [31] proposes sequential relation knowledge distillation to improve the performance of small-scale vision transformer models, which utilize the full set of patch tokens and distill the knowledge from a pretrained ViT-Large model. TinyMIM [31] needs the large pretrained vision transformer and the use of sequential models is time-consuming for practical applications. These methods attempt to utilize the task-agnostic knowledge from large models through knowledge distillation. Different from their approaches, we explore sparse training as a pruning technique in order to scale down ViTs from large-scale to small-scale under the Masked Autoencoders framework." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16177" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.087, + 0.443, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.207, + 0.431, + 0.224 + ], + "angle": 0, + "content": "(a) General pretraining and finetuning" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.087, + 0.78, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.203, + 0.752, + 0.234 + ], + "angle": 0, + "content": "(b) General pretraining and sparse finetuning" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.274, + 0.441, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.39, + 0.46, + 0.405 + ], + "angle": 0, + "content": "(c) Sparse pretraining and sparse finetuning" + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.244, + 0.786, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.39, + 0.803, + 0.406 + ], + "angle": 0, + "content": "(d) SparseMAE pretraining and finetuning" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.437, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Figure 2: An illustration of four types pretraining and finetuning frameworks. (a) Dense pretraining and finetuning the large full model as done in MAE [15]. \\( G \\) and \\( D \\) are general task-agnostic vision knowledge and task-specific downstream knowledge learnt from pretraining and finetuning respectively. (b) Learning \\( G \\) during dense pretraining with the large full model and acquiring downstream task-specific knowledge \\( D^{s} \\) for the sparse sub-network during sparse finetuning. (c) Learning task-agnostic knowledge \\( G^{s} \\) and task-agnostic knowledge \\( D^{s} \\) with sparse sub-network during sparse pretraining and finetuning. For finetuning, the sparse mask can be kept as fixed or dynamic. (d) Our SparseMAE training framework learns both \\( G \\) and \\( G^{s} \\) during pretraining and subsequently learns better \\( D^{s} \\) during finetuning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.569, + 0.339, + 0.586 + ], + "angle": 0, + "content": "2.2. Pruning and Sparse Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.471, + 0.744 + ], + "angle": 0, + "content": "Pruning and sparse training are critical methods for reducing the inference costs of deep neural networks [18]. Post-pruning methods [13, 26] typically involve training a dense full network first, followed by one or several rounds of train-prune-retrain rounds. Post-pruning methods are time-consuming due to the iterative manner [11]. Different from post-training methods, sparse training methods [29, 36] aim to adaptively obtain a sparse sub-network while optimizing both model weights and sparse connections from random initialization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Sparse training is a class of methods stemming from the Sparse Evolutionary Training (SET) algorithm [29]. It starts from a randomly sparsified network and subsequently prunes and grows connections during training. These methods can be applied end-to-end within the network training stage and have achieved promising results on various vision benchmarks with more efficient models by adaptively changing the model topology connections during training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Sparse training methods have shown promising results in various computer vision benchmarks, achieving state-of-the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.895, + 0.69 + ], + "angle": 0, + "content": "art performance on different types of sparsity granularities such as unstructured sparsity [4, 23, 27], N:M fine-grained structured sparsity [42, 32, 10, 41], and block sparsity [7]. However, to our knowledge, the studies of pruning techniques for Vision Transformers [4, 20] only focus on the fully-supervised settings. There still lacks a unified method to prune large-scale Vision Transformers under the unsupervised Masked Autoencoders framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.705, + 0.599, + 0.721 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.894, + 0.837 + ], + "angle": 0, + "content": "In this section, we first revisit the sparse training technique and Masked Autoencoders (MAE) in Sec. 3.1. Then we emphasize the challenges of integrating existing sparse training methods into MAE in Sec. 3.2. Lastly, we introduce our proposed method to effectively obtain a powerful small-scale sub-network from a large-scale model using sparse training in Sec. 3.3" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.642, + 0.862 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Revisiting Sparse Training. The pioneering work of Lottery Ticket Hypothesis (LTH) [11] demonstrates the dense" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16178" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.397 + ], + "angle": 0, + "content": "networks contain sparse matching subnetworks capable of training in isolation to full accuracy. Its findings inspire searching for strong sub-networks within a larger full network. Sparse training starts from randomly initializing the network weights \\(\\mathbf{W}\\) and associating them with a sparse mask \\(\\mathbf{M}\\), which is a binary tensor with 0 indexing the sparsified entries and 1 indicating the retained ones. \\(\\mathbf{M}\\) is optimized end-to-end and is usually obtained by magnitude pruning [14]. We denote the sparse sub-network as \\(\\widetilde{\\mathbf{W}} = \\mathbf{W} \\odot \\mathbf{M}\\). Sparse-to-sparse training regime optimizes the sparse weights \\(\\widetilde{\\mathbf{W}}\\) and the corresponding binary mask \\(\\mathbf{M}\\), which can be updated at each iteration or periodically. Variants of sparse training, such as unstructured sparsity and fine-grained structured sparsity [23, 25, 43, 9] achieve state-of-the-art performance on various convolutional networks with improved efficiency. Masked Autoencoders. Masked Autoencoders are a family of self-supervised methods for pretraining Vision Transformers (ViT). They involve reconstructing masked RGB patches from visible tokens \\(x_{\\mathrm{vis}}\\) with an encoder \\(f\\) and a decoder \\(g\\). Mathematically, we have" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.423, + 0.471, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M A E}} (\\mathbf {W}) = \\mathbb {E} \\left(\\left\\| g _ {\\theta} \\left(f \\left(x _ {\\text {v i s}}, \\mathbf {W}\\right)\\right) - x _ {\\text {m a s k}} \\right\\| ^ {2}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.473, + 0.606 + ], + "angle": 0, + "content": "Here, an encoder \\( f \\) maps inputs \\( x_{\\mathrm{vis}} \\) to latent features \\( z_{\\mathrm{vis}} = f(x_{\\mathrm{vis}}) \\), and a decoder \\( g_{\\theta} \\) parameterized by \\( \\theta \\) predicts the masked token \\( x_{\\mathrm{mask}} \\) in the RGB pixel space from the latent feature \\( z_{\\mathrm{vis}} \\). Encoder weights \\( \\mathbf{W} \\) and decoder weights \\( \\theta \\) are jointly trained. After pretraining, the decoder is discarded and the encoder is finetuned on task-specific downstream datasets. During finetuning, the model acquires downstream specific knowledge \\( D \\), which is built on general task-agnostic knowledge \\( G \\) learnt in pretraining, as shown in Fig. 2 (a)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.615, + 0.42, + 0.631 + ], + "angle": 0, + "content": "3.2. Challenges of Sparse Training for MAE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.473, + 0.775 + ], + "angle": 0, + "content": "Despite the effectiveness of MAE for large-scale vision transformers, it encounters great difficulties when applied to smaller models (see Fig. 1). The inadequate model capacity prevents small models from learning meaning task-agnostic knowledge and consequently transfers to inferior performance in downstream tasks. To enable the potential of large models' capability for smaller ones to learn powerful representations under MAE framework, we can modify Eq. 1 to combine it with sparse training as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.797, + 0.47, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S p a r s e}} (\\mathbf {W}, \\mathbf {M}) = \\mathbb {E} \\left(\\left\\| g _ {\\theta} \\left(f \\left(x _ {\\text {v i s}}, \\mathbf {W} \\odot \\mathbf {M}\\right)\\right) - x _ {\\text {m a s k}} \\right\\| ^ {2}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\( f \\) represents the large full encoder with parameters \\( \\mathbf{W} \\). The encoder binary mask \\( \\mathbf{M} \\) produces a sparse sub-network by magnitude pruning during pretraining, while the decoder is kept dense as it is not used for downstream finetuning. Following the notation in Sec. 3.1, we use \\( G^{s} \\) to represent" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.182 + ], + "angle": 0, + "content": "the general knowledge of the sparse sub-network learnt in pretraining, and we denote \\( D^{s} \\) the specific knowledge of the finetuned sparse sub-network for downstream tasks. Considering both the pretraining and finetuning phases, we outline three intuitive strategies to integrate sparse training into the MAE framework:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.196, + 0.895, + 0.302 + ], + "angle": 0, + "content": "- 1 Dense pretraining then sparse finetuning as shown in Fig. 2 (b). Although this strategy allows the sparse finetuning phase to optimize from the large full model's task-agnostic knowledge \\( G \\), the sub-network does not benefit from the pretraining task and lacks its knowledge \\( G^{s} \\). This may hinder the model to acquire better downstream knowledge \\( D^{s} \\)" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.313, + 0.895, + 0.435 + ], + "angle": 0, + "content": "- 2&3 Sparse pretraining using Eq. (2) then sparse finetuning (see Fig. 2 (c)). During finetuning, the sparsity mask can be kept as fixed or dynamic (dynamic mask is the third strategy). This strategy relies on the architecture and task-agnostic knowledge \\( G^{s} \\) obtained by the sub-network during the pretraining phase, while it ignores the potential benefits brought by incorporating the knowledge \\( G \\) of the large full model." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.196, + 0.895, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.568 + ], + "angle": 0, + "content": "In summary, it is challenging to tailor a sparse training mechanism for MAE framework to sufficiently learn task-agnostic and task-specific knowledge for a small model. In the following section, we introduce our approach to solve this dilemma by fully utilizing the task-agnostic knowledge \\( G \\) of the large full model and adaptively optimizing sub-network's topology connections for various downstream tasks for better task-specific knowledge." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.578, + 0.766, + 0.594 + ], + "angle": 0, + "content": "3.3. Sparse Masked Autoencoders" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.647 + ], + "angle": 0, + "content": "To tackle the limitations of the sparse training strategies introduced above, we propose a novel method for pruning ViT models under the MAE framework (see Fig. 2 (d))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.895, + 0.784 + ], + "angle": 0, + "content": "Firstly, to ensure a good general task-agnostic knowledge \\( G \\) to be learnt via pretraining, we utilize the dense full model to reconstruct the RGB pixels following Eq. 1, which learns more discriminative representations thanks to its larger capacity. Simultaneously, a sparse sub-network within the full dense network is trained using the same visible tokens as the inputs and regresses the same targets as the large full model. The sparseMAE loss function is therefore extended from Eq. 1 and Eq. 2 as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.796, + 0.892, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\underbrace {\\mathcal {L} _ {\\text {S p a r s e}} (\\widetilde {\\mathbf {W}})} _ {\\text {s u b - n e t w o r k}} + \\underbrace {\\alpha \\mathcal {L} _ {\\mathrm {M A E}} (\\mathbf {W})} _ {\\text {f u l l n e t w o r k}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.895, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}\\) represents the dense parameters of the large full model, \\(\\widetilde{\\mathbf{W}} = \\mathbf{W} \\odot \\mathbf{M}\\) is the masked sparse parameters of the sub-network, and \\(\\alpha\\) is the scaling hyperparameter balancing the reconstruction losses for the dense full model and the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16179" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.089, + 0.833, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.416, + 0.895, + 0.478 + ], + "angle": 0, + "content": "Figure 3: Illustration of Sparse Masked Autoencoders. In the forward pass, both the large full model (red solid lines) and the sparse sub-network (red dashed lines) reconstruct the masked patches from the same visible tokens. In the backward pass, the masked weights (green solid lines) and the sparse weights (green dashed lines) are updated separately using the gradients from the corresponding losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.533 + ], + "angle": 0, + "content": "sparse sub-network. In our experiments, we set \\(\\alpha = 1.0\\) by default." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.473, + 0.66 + ], + "angle": 0, + "content": "As shown in Eq. 3, the gradients for the sparse weights \\(\\widetilde{\\mathbf{W}}\\) shall be updated following both the reconstruction losses of the sub-network and the full network, which may lead to gradient imbalance between the masked and unmasked weights [38]. Therefore, in each training iteration we split the weights \\(\\mathbf{W}\\) of the full model into sparse sub-network weights \\(\\widetilde{\\mathbf{W}}\\) and unmasked weights \\(\\overline{\\mathbf{W}} = \\mathbf{W} \\odot (1 - \\mathbf{M})\\) and optimize them separately:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.673, + 0.473, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathbf {W}} _ {t + 1} = \\widetilde {\\mathbf {W}} _ {t} - \\eta (g (\\mathbf {W} _ {t} \\odot \\mathbf {M} _ {t}) \\odot \\mathbf {M} _ {t}), \\qquad (4)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.716, + 0.472, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {W}}} _ {t + 1} = \\overline {{\\mathbf {W}}} _ {t} - \\eta (g (\\mathbf {W} _ {t}) \\odot (1 - \\mathbf {M} _ {t})), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\( g(\\mathbf{W}_t \\odot \\mathbf{M}_t) = \\frac{\\partial \\mathcal{L}_{\\mathrm{Sparse}}(\\mathbf{W}_t \\odot \\mathbf{M}_t)}{\\partial \\mathbf{W}_t} \\) are gradients of the sparse sub-network, \\( g(\\mathbf{W}_t) = \\frac{\\partial \\mathcal{L}_{\\mathrm{MAE}}(\\mathbf{W}_t)}{\\partial \\mathbf{W}_t} \\) are the gradients of the full models, \\( \\eta \\) is the learning rate and \\( t \\) indexes the training iteration. The binary mask \\( \\mathbf{M}_t \\) is obtained by on-the-fly magnitude pruning. Here, we only show gradient descent for simplicity and omit other parts (e.g., weight decay). In this way, the large full model can utilize all the model parameters in the forward pass to produce discriminative outputs, while in the backward pass, the sparse sub-network stays unaffected by the gradients from the large full model. During" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.895, + 0.595 + ], + "angle": 0, + "content": "the pretraining phase, the task-agnostic knowledge \\( G^{s} \\) of the sparse sub-network can dynamically select the important sparse connections in the context of the task-agnostic knowledge \\( G \\) of the full dense model, which helps to identify the important sparse topology and is retained for downstream finetuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.896, + 0.807 + ], + "angle": 0, + "content": "After pretraining, the full dense model with a sparse sub-network is transferred and finetuned on different downstream tasks (e.g., classification, detection, segmentation, etc.). For the finetuning stage, the deep representations from the sparse sub-network are used for the downstream tasks, while its topology connections are optimized end-to-end given the specific downstream objectives. We refrain from fine-tuning dense full models, as the fine-tuning process for both dense and sparse models is costly. Hence, the learning of the task-specific knowledge \\( D^{s} \\) for downstream tasks can benefit from inheriting a pretrained sparse sub-network with knowledge \\( G^{s} \\), and adaptively selecting the more suitable sparse architecture built on the general knowledge \\( G \\) of the full model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.828, + 0.634, + 0.846 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.897, + 0.903 + ], + "angle": 0, + "content": "In this section, we validate the effectiveness and robustness of our SparseMAE. First, we introduce the implementation details of our SparseMAE in Sec. 4.1. Then, we present" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16180" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "our main experimental results in Sec. 4.2, which include classification results on ImageNet-1K [8] and semantic segmentation results on ADE20K [43]. To demonstrate the superiority of our proposed method, we compare SparseMAE with other state-of-the-art ViT pruning methods in Sec. 4.4. In Sec. 4.3, we evaluate the transferring ability of the models trained with our SparseMAE to object detection, instance segmentation tasks on MS COCO [24] and robustness evaluation. Finally, we perform ablative experiments to evaluate the performance of SparseMAE under different settings in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.27, + 0.295, + 0.287 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.473, + 0.476 + ], + "angle": 0, + "content": "Sparse Training Methods. We conduct experiments with two different granularities for sparse training, consisting of unstructured sparsity [14, 21] and hardware-friendly N:M sparsity [28, 42]. As we focus on studying how to learn and transfer task-agnostic and task-specific knowledge under the Masked Autoencoders paradigm with sparse training methods, we select two simple and effective sparse training methods, namely DFP [23] for unstructured sparsity and SR-STE [42] for N:M sparsity. We prune all the linear layers within the vision transformers and set the relevant hyperparameters for sparse training the same as [23, 42] by default." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.473, + 0.673 + ], + "angle": 0, + "content": "Models. For all of our main experiments, unless otherwise stated, we select ViT-Base as our large full model and prune it for smaller ones. For hardware-friendly N:M sparsity, we use 2:32 and 4:32 sparse ratios to obtain models with similar scales to ViT-Tiny (5.8M) and ViT-Small (11.3M), the main experiments use N:M sparsity for default, which is applied to all linear layers, including self.qkv, self.proj in the Attention block, and self.fc1 and self.fc2 in the MLP block. For unstructured sparsity, we extend the above N:M ratios and prune ViT-Base with \\(93.5\\%\\) and \\(87.0\\%\\) sparsity. We name our ViT-Base with 2:32/93.5% sparsity as SparseMAE-Tiny (SparseMAE-T) and the one with 4:32/87.0% sparsity as SparseMAE-Small (SparseMAE-S)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.473, + 0.855 + ], + "angle": 0, + "content": "Pretraining. Following MAE [15] pretraining scheme, models start from random initialization, we use ImageNet-1K (IN-1K) [8] for all pretraining experiments. We pretrain SparseMAE for 400 epochs using \\(224 \\times 224\\) images as inputs. The other settings also follow those used in [15]. For example, We use random resized cropping and horizontal flipping for data augmentation, a \\(75\\%\\) mask ratio during pretraining, and a decoder with a single transformer layer and 512 hidden dimensions. Unlike other methods [34, 19] for small and tiny-scale vision transformers pretraining, we do not grid-search the decoder design, possible putting our method at a disadvantage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Finetuning. We evaluate our SparseMAE pretrained models for various downstream tasks, including ImageNet-1K [8] image classification, MS COCO [24] object detection and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.121 + ], + "angle": 0, + "content": "instance segmentation tasks, ADE20K [43] semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.897, + 0.244 + ], + "angle": 0, + "content": "For image classification, we take a ViT-Base model pretrained using MAE for 400 epochs as the reference model, which is implemented using the officially released codebase [15] and achieves \\(83.2\\%\\) top-1 accuracy. We finetune SparseMAE-Tiny/Small for 200 epochs using the same receipt described in [15] for a fair comparison with other works [31, 19] and set the batch size as 1024 and the base learning rate as \\(10^{-3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.897, + 0.382 + ], + "angle": 0, + "content": "For semantic segmentation, we follow BEiT [2] to use UperNet[37] as our segmentation framework. We train the model with an input resolution of \\(512 \\times 512\\) for 160k iterations. A batch size of 16 is adopted. For object detection and instance segmentation tasks, we follow ViTDet [22] framework. We train our SparseMAE models with a batch size of 64 for 100 epochs, and the input image resolution is set to \\(1024 \\times 1024\\). Further details on the hyper-parameters can be found in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.392, + 0.64, + 0.408 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.897, + 0.719 + ], + "angle": 0, + "content": "Classification and Semantic Segmentation. Tab. 1 shows the evaluation of our SparseMAE model against self-supervised methods for ViT-Tiny/Small on ImageNet-1K image classification and ADE20K semantic segmentation. For ImageNet-1K classification, SparseMAE achieves \\(80.5\\%\\) top-1 accuracy. To our knowledge, this is currently the best result achieved by tiny-scale transformer models. Furthermore, SparseMAE outperforms TinyMIM [31]/G2SD [19]/MAE-lite [34] with \\(+4.7 / 4.2 / 4.0\\) accuracy gains on tiny-scale transformers. For semantic segmentation, SparseMAE achieves an mIoU of 45.2, outperforming G2SD/MAE-lite/TinyMIM by \\(+3.8 / 1.3 / 1.2\\) mIoUs. Similar improvements can be observed for the small-scale transformers. SparseMAE achieves a \\(83.2\\%\\) ImageNet-1K classification accuracy and a \\(48.4\\%\\) mIoU on ADE20K semantic segmentation, surpassing TinyMIM [31]/G2SD [19] by \\(+0.2 / 1.2\\) and \\(+0.0 / 2.2\\) on the two tasks respectively. The significant improvements over previous results show the superiority of our method in obtaining performant small models under MAE framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Speedups. Unstructured sparsity and N:M structured sparsity is efficient for commercial CPUs and FPGAs respectively. For our SparseMAE with unstructured sparsity, we tested the throughput (img/s) on an Intel RXeon(R) 6238R CPU using the deepsparse engine with a batch size of 64. For our SparseMAE with N:M sparsity, we tested the throughput on Xilinx XCVU13P with a batch size of 16. Our SparseMAE-S with N:M and unstructured sparsity achieved throughput gains of \\(+71\\%\\) and \\(+9\\%\\) over ViT-small with similar amount of parameters on FPGAs and CPUs. Therefore, our approach demonstrate high feasibility in actual scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16181" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.138 + ], + "angle": 0, + "content": "Table 1: Finetuning results on ImageNet-1K and ADE20K. All our models are pre-trained only on ImageNet-1K. Extra pretrained teacher means the compared method cannot trained from scratch. \\(\\ddagger\\) the model adopts the MAE pretrained model before performing the sparse pretraining." + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.147, + 0.84, + 0.493 + ], + "angle": 0, + "content": "
MethodParams (M)Pretraining epochsEncoder ratioExtra Pretrained/ TeacherClassification Top-1 Acc (%)Segmentation mIoU (%)
DeiT [33]5.8300-Label72.238.0
Tiny-scale
MAE [15]5.8160025%X71.637.6
MoCo [6]5.8300100%EMA73.339.3
TinyMIM [31]5.8300100%TinyMIM-S75.844.0
G2SD [19]5.8300100%ViT-B76.341.4
MAE-lite [34]5.840025%ViT-B76.543.9
SparseMAE (ours)5.840025%X80.545.2
DeiT [33]22300-Label79.9
Small-scale
MAE [15]22160025%X80.642.8
MoCo [6]22300100%EMA81.443.9
DINO [3]22300100%EMA81.545.3
CAE [5]22300100%DALL-E82.0-
TinyMIM [31]22300100%TinyMIM-ViT-B83.048.4
G2SD [19]2230025%ViT-B82.046.2
SparseMAE (ours)11.330025%ViT-B‡83.248.4
SparseMAE (ours)11.340025%X82.146.7
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.517, + 0.423, + 0.534 + ], + "angle": 0, + "content": "4.3. Evaluation on Other Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.471, + 0.693 + ], + "angle": 0, + "content": "Object Detection and Instance Segmentation. We finetune the SparseMAE pretrained models on MS COCO with ViT-Det [22]. As shown in Tab. 2, we report \\(\\mathrm{AP^{bbox}}\\) for object detection and \\(\\mathrm{AP^{mask}}\\) for instance segmentation. We compare SparseMAE-Tiny with MAE-Tiny and its improved variants using distillation methods, such as MAE-lite [34] and G2SD [19]. The results in Tab. 2 show that our Sparse-MAE obtains more than \\(3.1\\mathrm{AP^{bbox}}\\) and \\(2.4\\mathrm{AP^{mask}}\\) gains compared to MAE-lite and G2SD, validating the transferability of the learnt representations in our models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.471, + 0.834 + ], + "angle": 0, + "content": "Robustness Evaluation. In Tab. 3, we evaluate the robustness of our models on different variants of ImageNet validation sets. We use the same models finetuned on original ImageNet (Tab. 1) and only run inference on the different validation sets, such asImagenet-R [17], ImageNet-A [17] and ImageNet-C [16]), without any specialized fine-tuning. Tab. 3 shows that our method has better generalization capability than distillation-based methods, such as G2SD [19] and TinyMIM [31])." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.471, + 0.865 + ], + "angle": 0, + "content": "4.4. Comparison with State-of-the-art ViT Pruning" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To the best of our knowledge, our proposed SparseMAE is the first to study network pruning technique under the" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.546 + ], + "angle": 0, + "content": "Table 2: Object detection and instance segmentation results on the MS COCO dataset." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.557, + 0.855, + 0.727 + ], + "angle": 0, + "content": "
MethodBackbone Size (M)Det APbboxSeg APmask
Tiny-scale models
DeiT [33]5.840.736.5
MAE [15]5.838.935.1
MoCo v3 [6]5.840.036.0
DINO [3]5.840.236.1
MAE-lite [34]5.842.738.2
G2SD [19]5.844.039.6
SparseMAE (ours)5.847.142.0
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Masked Autoencoders paradigm. To validate the effectiveness of SparseMAE, we compare it against recent state-of-the-art fully-supervised pruning methods for ViTs. We choose SViT [4] and oViT [20] as our compared models, which exploit end-to-end sparse training for ViTs and achieve state-of-the-art results in the fully-supervised setting. SViT [4] adopts a dynamic sparse training method to prune the dense model using random initialization. However, this approach requires training the sparse model for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16182" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.472, + 0.166 + ], + "angle": 0, + "content": "Table 3: Robustness evaluation on ImageNet variants (top-1 accuracy, except for ImageNet-C which evaluates mean corruption error). We test the small-scale and tiny-scale models from Tab. 1 on different ImageNet validation sets, without any fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.177, + 0.468, + 0.367 + ], + "angle": 0, + "content": "
MethodModel SizeIN-A↑IN-R↑IN-C↓
DeiT [33]8.032.754.0
MAE [15]7.036.555.2
TinyMIM [31]Tiny-scale11.039.850.1
G2SD [19]12.939.0-
SparseMAE (ours)18.245.947.2
DeiT [33]18.342.341.4
MAE [15]20.145.640.6
TinyMIM [31]Small-scale27.548.835.8
G2SD [19]29.446.8-
SparseMAE (ours)29.349.235.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.414 + ], + "angle": 0, + "content": "Table 4: Comparison with state-of-the-art supervised Vision Transformer pruning methods on ImageNet-1K." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.424, + 0.449, + 0.569 + ], + "angle": 0, + "content": "
ModelMethodParams (M)Top-1 Acc (%)
ViT-BaseSViT [4]43.381.5
oViT [20]43.381.6
SViT [4]34.681.3
oViT [20]34.681.5
oViT [20]21.681.1
oViT [20]17.380.8
oViT [20]8.779.7
ViT-BaseSparseMAE-T (ours)5.880.5
SparseMAE-S (ours)11.383.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.732 + ], + "angle": 0, + "content": "600 epochs. oViT [20] (post-training pruning) prunes the model initialized from a trained dense model (300 epochs), which requires 300 epochs of finetuning. For our SparseMAE, we only need 400 epochs for pretraining and 200 epochs for fine-tuning. The pretraining of SparseMAE is \\(3 \\times\\) faster than full-token supervised training. Consequently, our SparseMAE approach reduces the training time from 968 GPU hours to 581 GPU hours compared to SOTA SViT and oViT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Tab. 4 demonstrates that SparseMAE can outperform state-of-the-art fully-supervised pruning counterparts significantly for small and tiny-scale transformers. Our SparseMAE-T has a \\(0.8\\%\\) performance gain and a \\(35.0\\%\\) computation complexity reduction compared to oViT [20] for tiny-scale transformers. With our SparseMAE-S, we achieve a \\(83.2\\%\\) ImageNet classification accuracy, outperforming supervised pruning techniques by large margins. These results suggest that our method successfully extends the superiority of MAE for representation learning with large-scale vision transformers to small and tiny-scale models." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.089, + 0.895, + 0.198 + ], + "angle": 0, + "content": "Table 5: Different sparse pretraining or finetuning strategies with tiny-scale models on ImageNet-1K and ADE20K. \"S\" means sparse pretraining, \"D\" means dense pretraining, and \"D+S\" means sparse pretraining together with a dense full model pretraining. \\(\\ddagger\\): the model is finetuned with the fixed sparse masks from pretraining. Corr. Strategy.: corresponding sparse training strategies explained in Sec. 4.5.1." + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.207, + 0.872, + 0.396 + ], + "angle": 0, + "content": "
ModelPretrain MethodTop-1 Acc (%)Seg mIoU (%)Corr. Strtg.
ViT-BaseD83.248.1-
Unstructured sparsity (93.5% sparse ratio)
SparseMAE-TD77.840.2#1
SparseMAE-TS‡76.737.7#2
SparseMAE-TS77.639.8#3
SparseMAE-TD+S80.445.4-
2:32 sparsity
SparseMAE-TD78.440.7#1
SparseMAE-TS‡76.337.2#2
SparseMAE-TS78.741.3#3
SparseMAE-TD+S80.545.2-
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.425, + 0.783, + 0.44 + ], + "angle": 0, + "content": "4.5. Ablation Studies and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.451, + 0.851, + 0.466 + ], + "angle": 0, + "content": "4.5.1 Comparison of Sparse Training Strategies." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.478, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We compare three sparse training strategies for MAE framework introduced in Sec. 3.2 under two sparse granularities. Namely, we (1) pretrain a large full model using MAE and finetune it with sparse training, (2) combine sparse training with MAE pretraining and subsequently finetune the model on downstream tasks with the fixed sparse mask from pretraining, or (3) modify (2) to use dynamic sparse mask for finetuning. The results are shown in Tab. 5. Under unstructured sparsity, comparing (1) and (2), we see that the classification performance drops from \\(77.8\\%\\) to \\(76.7\\%\\), showing that the knowledge \\(G^{s}\\) learnt by the sparse sub-network during pretraining cannot generalize well to the downstream tasks compared to the task-agnostic knowledge \\(G\\) learnt by the large full model. (3) further allows the adaptation of sparse sub-network architecture during finetuning and partially alleviates the problem by better acquiring downstream specific knowledge \\(D^{s}\\). However, its result still lags significantly behind our proposed method (\\(77.6\\%\\) vs. \\(80.4\\%\\)), where the difference is only the absence of a concurrently pretrained large full model. On the other hand, when comparing (1) with our method, the addition of a pretrained sparse sub-network can boost the performance by \\(2.6\\%\\). Similar observations can also be made under N:M sparsity settings and semantic segmentation results. These empirical results suggest that both the task-agnostic knowledge \\(G\\) of the large full model and \\(G^{s}\\) of the sparse sub-network are necessary for small and tiny-scale vision transformers to effectively benefit from MAE pretraining." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16183" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.121 + ], + "angle": 0, + "content": "Table 6: Comparison between different full models for SparseMAE on ImageNet-1K and ADE20K." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.131, + 0.457, + 0.21 + ], + "angle": 0, + "content": "
ModelPatternParams (M)Top-1 Acc (%)Seg mIoU (%)
ViT-Small8:325.877.441.5
ViT-Base2:325.880.545.2
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.235, + 0.358, + 0.251 + ], + "angle": 0, + "content": "4.5.2 Ablation of Full Model Capacity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.397 + ], + "angle": 0, + "content": "We conduct an ablation study to investigate the performance with different capacities of the full model. We use our Sparse-MAE and set ViT-Small and ViT-Base as the full models, where ViT-Small only has \\(25\\%\\) parameters of ViT-Base. As shown in Tab. 6, the results show that our SparseMAE in ViT-Base achieves \\(3.1\\%\\) top-1 accuracy advantage over ViT-Small, indicating that larger capacity of the full model with SparseMAE can significantly improve its performance by incorporating better pretraining task-agnostic knowledge." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.416, + 0.439, + 0.433 + ], + "angle": 0, + "content": "4.5.3 Impact of Different Pruning Sparsity Ratios" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.472, + 0.502 + ], + "angle": 0, + "content": "The main experiments of sparse ratio setting is compared to the ViT-tiny and ViT-small. We provide additional sparse ratios to verify the performance of SparseMAE on ImageNet-1K as in Tab. 7." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.511, + 0.472, + 0.543 + ], + "angle": 0, + "content": "Table 7: Comparison between different sparse ratios for SparseMAE on ImageNet-1K." + }, + { + "type": "table", + "bbox": [ + 0.135, + 0.554, + 0.411, + 0.602 + ], + "angle": 0, + "content": "
Sparse Pattern1:322:324:328:32
Top-1 Acc76.880.583.283.7
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.633, + 0.196, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce SparseMAE, an innovative method for incorporating sparse training into the Masked Autoencoders (MAE) framework to create efficient and powerful small and tiny-scale transformers. As the first study in this area, we first point out the limitations of sparse training under MAE and design a unified framework to transfer the task-agnostic and task-specific knowledge of large models to lightweight sparse sub-networks. SparseMAE trains the sparse sub-networks by persevering the task-agnostic knowledge of large full models and then adaptively finds the optimal specific sub-networks for downstream tasks. The proposed method outperforms the state-of-the-art small-scale vision transforms in both unsupervised pretraining and fully-supervised settings. We hope this work could shed light on sparse training with the Masked Autoencoders framework in the vision community." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.661, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.221 + ], + "angle": 0, + "content": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.236, + 0.597, + 0.252 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.261, + 0.895, + 0.331 + ], + "angle": 0, + "content": "[1] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.333, + 0.895, + 0.374 + ], + "angle": 0, + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.375, + 0.895, + 0.445 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.446, + 0.895, + 0.502 + ], + "angle": 0, + "content": "[4] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34:19974-19988, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.504, + 0.895, + 0.572 + ], + "angle": 0, + "content": "[5] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.574, + 0.895, + 0.631 + ], + "angle": 0, + "content": "[6] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.632, + 0.895, + 0.685 + ], + "angle": 0, + "content": "[7] Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. arXiv preprint arXiv:2205.14135, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.688, + 0.895, + 0.744 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.745, + 0.895, + 0.801 + ], + "angle": 0, + "content": "[9] Utku Evci, Trevor Gale, Jacob Menick, Pablo Samuel Castro, and Erich Olsen. Rigging the lottery: Making all tickets winners. In International Conference on Machine Learning, pages 2943-2952. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.802, + 0.895, + 0.858 + ], + "angle": 0, + "content": "[10] Chao Fang, Aojun Zhou, and Zhongfeng Wang. An algorithm–hardware co-optimized framework for accelerating n: M sparse transformers. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 30(11):1573–1586, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.859, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[11] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635, 2018." + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.261, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16184" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[12] Peng Gao, Renrui Zhang, Rongyao Fang, Ziyi Lin, Hongyang Li, Hongsheng Li, and Qiao Yu. Mimic before reconstruct: Enhancing masked autoencoders with feature mimicking. IJCV 2023, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[14] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.472, + 0.318 + ], + "angle": 0, + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.472, + 0.433 + ], + "angle": 0, + "content": "[17] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[18] Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks. The Journal of Machine Learning Research, 22(1):10882-11005, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.472, + 0.546 + ], + "angle": 0, + "content": "[19] Wei Huang, Zhiliang Peng, Li Dong, Furu Wei, Jianbin Jiao, and Qixiang Ye. Generic-to-specific distillation of masked autoencoders. arXiv preprint arXiv:2302.14771, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.472, + 0.589 + ], + "angle": 0, + "content": "[20] Denis Kuznedelev, Eldar Kurtic, Elias Frantar, and Dan Alistarh. ovit: An accurate second-order pruning framework for vision transformers. arXiv preprint arXiv:2210.09223, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.472, + 0.631 + ], + "angle": 0, + "content": "[21] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. Advances in neural information processing systems, 2, 1989." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.472, + 0.702 + ], + "angle": 0, + "content": "[22] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 280-296. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.472, + 0.746 + ], + "angle": 0, + "content": "[23] Tao Lin, Sebastian U Stich, Luis Barba, Daniil Dmitriev, and Martin Jaggi. Dynamic model pruning with feedback. arXiv preprint arXiv:2006.07253, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[25] Shiwei Liu, Lu Yin, Decebal Constantin Mocanu, and Mykola Pechenizkiy. Do we actually need dense overparameterization? in-time over-parameterization in sparse training. In International Conference on Machine Learning, pages 6989-7000. PMLR, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.135 + ], + "angle": 0, + "content": "[26] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[27] Zhi-Gang Liu, Paul N Whatmough, and Matthew Mattina. Sparse systolic tensor array for efficient cnn hardware acceleration. arXiv preprint arXiv:2009.02381, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.894, + 0.235 + ], + "angle": 0, + "content": "[28] Asit Mishra, Jorge Albericio Latorre, Jeff Pool, Darko Stosic, Dusan Stosic, Ganesh Venkatesh, Chong Yu, and Paulius Miccikevicius. Accelerating sparse deep neural networks. arXiv preprint arXiv:2104.08378, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.239, + 0.894, + 0.307 + ], + "angle": 0, + "content": "[29] Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H Nguyen, Madeleine Gibescu, and Antonio Liotta. Scalable training of artificial neural networks with adaptive sparse connectivity inspired by network science. Nature communications, 9(1):2383, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.894, + 0.366 + ], + "angle": 0, + "content": "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.368, + 0.893, + 0.409 + ], + "angle": 0, + "content": "[31] Sucheng Ren, Fangyun Wei, Zheng Zhang, and Han Hu. Tinymim: An empirical study of distilling mim pre-trained models. arXiv preprint arXiv:2301.01296, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.412, + 0.893, + 0.481 + ], + "angle": 0, + "content": "[32] Wei Sun, Aojun Zhou, Sander Stuijk, Rob G. J. Wijnhoven, Andrew Nelson, Hongsheng Li, and Henk Corporaal. Dominosearch: Find layer-wise fine-grained n:m sparse schemes from dense neural networks. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.893, + 0.552 + ], + "angle": 0, + "content": "[33] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.555, + 0.893, + 0.596 + ], + "angle": 0, + "content": "[34] Shaoru Wang, Jin Gao, Zeming Li, Jian Sun, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. arXiv preprint arXiv:2205.14443, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.599, + 0.893, + 0.641 + ], + "angle": 0, + "content": "[35] Chen Wei, Haoqi Fan, Saining Xie, Chao Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.643, + 0.893, + 0.684 + ], + "angle": 0, + "content": "[36] Mitchell Wortsman, Ali Farhadi, and Mohammad Rastegari. Discovering neural wireings. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.893, + 0.741 + ], + "angle": 0, + "content": "[37] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.893, + 0.785 + ], + "angle": 0, + "content": "[38] Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. Slimmable neural networks. arXiv preprint arXiv:1812.08928, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.893, + 0.843 + ], + "angle": 0, + "content": "[39] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. NeurIPS 2022, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.893, + 0.899 + ], + "angle": 0, + "content": "[40] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. CVPR 2023, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16185" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[41] Yuxin Zhang, Mingbao Lin, Zhihang Lin, Yiting Luo, Ke Li, Fei Chao, Yongjian Wu, and Rongrong Ji. Learning best combination for efficient n:m sparsity. arXiv preprint arXiv:2206.06662, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[42] Aojun Zhou, Yukun Ma, Junnan Zhu, Jianbo Liu, Zhijie Zhang, Kun Yuan, Wenxiu Sun, and Hongsheng Li. Learning n: M fine-grained structured sparse neural networks from scratch. arXiv preprint arXiv:2102.04010, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[43] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.471, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16186" + } + ] +] \ No newline at end of file diff --git a/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/a542f651-0510-4853-91de-1141d304dc5f_origin.pdf b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/a542f651-0510-4853-91de-1141d304dc5f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..798a346c47dd50de56c82c23841b0f935b8dc75f --- /dev/null +++ b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/a542f651-0510-4853-91de-1141d304dc5f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a145c866bdabe57dc069b8d849420545b5422f33c39fa79fbd0414cc6dd5a49 +size 759717 diff --git a/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/full.md b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c28d279752f80e8047c4269f3e66f81e3f017812 --- /dev/null +++ b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/full.md @@ -0,0 +1,274 @@ +# SparseMAE: Sparse Training Meets Masked Autoencoders + +Aojun Zhou $^{1*}$ Yang Li $^{2}$ Zipeng Qin $^{1}$ Jianbo Liu $^{1}$ Junting Pan $^{1}$ Renrui Zhang $^{13}$ Rui Zhao $^{2}$ Peng Gao $^{3}$ Hongsheng Li $^{1}$ $^{1}$ The Chinese University of Hong Kong + $^{2}$ SenseTime Research Shanghai AI Lab + +# Abstract + +Masked Autoencoders (MAE) and its variants have proven to be effective for pretraining large-scale Vision Transformers (ViTs). However, small-scale models do not benefit from the pretraining mechanisms due to limited capacity. Sparse training is a method of transferring representations from large models to small ones by pruning unimportant parameters. However, naively combining MAE finetuning with sparse training make the network task-specific, resulting in the loss of task-agnostic knowledge, which is crucial for model generalization. In this paper, we aim to reduce model complexity from large vision transformers pretrained by MAE with assistant of sparse training. We summarize various sparse training methods to prune large vision transformers during MAE pretraining and finetuning stages, and discuss their shortcomings. To improve learning both task-agnostic and task-specific knowledge, we propose SparseMAE, a novel two-stage sparse training method that includes sparse pretraining and sparse finetuning. In sparse pretraining, we dynamically prune a small-scale sub-network from a ViT-Base. During finetuning, the sparse sub-network adaptively changes its topology connections under the task-agnostic knowledge of the full model. Extensive experimental results demonstrate the effectiveness of our method and its superiority on small-scale vision transformers. Code will be available at https://github.com/aojunzz/SparseMAE. + +# 1. Introduction + +Recently, several pretrained Masked Image Modeling (MIM) methods, such as MAE [15] and data2vec [1], have achieved great success in various computer vision tasks. They usually involve a two-stage training scheme where the model learns task-agnostic knowledge through a pretraining pretext task, and is subsequently finetuned on downstream tasks to acquire task-specific knowledge. Among + +these models, Masked Autoencoders (MAE) demonstrate superior learning capability on large-scale vision transformers (Fig. 1 (a)), thanks to large models' strong capacity to learn powerful general representations in the pretraining phase and versatile transferability to specific vision tasks. Yet, the large deep models are burdensome and difficult to be deployed to computationally restricted real-world scenarios. On the other hand, smaller and more efficient models, such as ViT-Tiny and ViT-Small, fail to perform well after pretrained by MAE or data2vec frameworks. For example, ViT-Tiny's finetuning results are even inferior to fully-supervised training (see Fig. 1 (b)). These empirical findings suggest that, under MIM frameworks, when the model is scaled down, it becomes more difficult to learn well via masked pretraining due to its limited capacity during pretraining and consequently hinders its performance when transferred to downstream tasks. In the realm of MIM pretraining, model capacity becomes a key ingredient to learning task-agnostic knowledge via the pretext tasks. + +In our studies, we try to tackle the issues of MIM pretraining for small-scale vision transformers. Particularly, we apply pruning techniques [28, 20, 4, 42, 23] in the MIM frameworks in order to obtain performant small-scale models from larger ones. As a first step in our exploration, we use existing sparse training methods [42, 23] for MAE pretraining and finetuning on unstructured and hardware-friendly N:M sparsity to obtain a model with a similar scale to ViT-Tiny, which is pruned from ViT-Base (see Tab. 5 row 3). It achieves notably improvements compared to directly training ViT-Tiny using MAE on downstream ImageNet finetuning accuracy (77.6% vs. 71.6%) and ADE20K [43] semantic segmentation (39.8 mIoU vs. 37.6 mIoU). The empirical results indicate that on small-scale ViTs sparse training via pruning has promising applications under the MIM pretraining paradigm. However, there is still a large performance gap between the sparse model and the vanilla ViT-Base trained densely under MAE in their ImageNet finetuning accuracies (77.6% vs. 83.2%). The sparsity constraints adversely limit the model capacity and prevent it from learning good task-agnostic knowledge, similar to the predicament faced + +![](images/f44911641785043be511cd7ead7c035e0e55ece70627b6f73874a9925ce4ab57.jpg) +Figure 1: (a) Top-1 accuracy of large-scale models on ImageNet-1K, showing that large-scale vision transformers and self-supervised training can bring significant improvement. (b) On tiny-scale and small-scale, our Sparse-MAE outperforms other self-supervised methods, such as data2vec [1], MAE [15], and training from scratch. + +by ViT-Tiny when it is directly pretrained under MAE. + +To obtain more powerful small-scale models, we propose SparseMAE, a simple and unified sparse training method for Masked Autoencoders. Specifically, SparseMAE trains an adaptive sub-network in conjunction with its corresponding large dense model during pretraining. Both the large model and the sparse sub-network predict the same reconstruction objective and are optimized end-to-end. The large full model provides a strong capacity to learn task-agnostic knowledge, while the sparse sub-network can better preserve the knowledge by rewiring the connections in the full model. To better learn task-specific knowledge in downstream tasks, the sparse sub-network inherits the task-agnostic knowledge from the pretraining stage and also dynamically adjusts its connections during the finetuning phase. + +We conduct extensive experiments on a variety of tasks, including image classification, object detection, segmentation, and robustness evaluation, to thoroughly evaluate the effectiveness of our proposed approach. Notably, our method achieves a top-1 accuracy of $80.5\%$ with only 6M parameters on ImageNet-1K [8] classification and obtains a 45.2 mIoU on ADE20K segmentation with UperNet [37], significantly outperforming all previous works that have developed small-scale vision transformers using MAE (e.g., MAE-lite, TinyMIM and G2SD distillation). + +In all, our contributions can be summarized in three folds: + +- We thoroughly study the integration of sparse training as a pruning technique into the MAE framework. + +Specifically, we design three different strategies to train sparse networks under MAE and analyze their limitations. + +- Based on the findings from above, we propose a novel sparse training framework, SparseMAE for Masked Autoencoders to improve the acquisition of both the task-agnostic and task-specific knowledge during the pretraining and finetuning stages. +- We present extensive experiments to validate the effectiveness of our proposed method and achieve state-of-the-art performance with small-scale vision transformers on ImageNet classification and various downstream tasks. + +# 2. Related works + +# 2.1. Masked Image Modeling + +Unsupervised pretraining on large-scale images with Masked Image Modeling (MIM) has shown superior performance on various computer vision tasks. BEiT [2] explores Masked Image Modeling on vision transformers by reconstructing the vision dictionary [30]. MAE [15] further proposes an asymmetric encoder and decoder for scaling up MIM to huge models. Besides, it demonstrates a simple pixel reconstruction loss can learn good visual representations. Due to the simplicity and computational efficiency, MAE is raising to a popular generative pretraining paradigm [40, 39]. As MAE reconstructs low-level signals with an isotropic vision transformer architecture, researchers improve MAE by exploring high-level signals architectures [1, 35, 12], which are more effective than reconstructing low-level signals. + +However, those methods demonstrate the large capacity model is essential for good representation learning through the MIM pretext tasks. MAE-lite [34] explores Masked Autoencoders to improve the performance of tiny-scale vision transformer models, and introduces knowledge distillation into the pretraining phase. Different from MAE-lite, TinyMIM [31] proposes sequential relation knowledge distillation to improve the performance of small-scale vision transformer models, which utilize the full set of patch tokens and distill the knowledge from a pretrained ViT-Large model. TinyMIM [31] needs the large pretrained vision transformer and the use of sequential models is time-consuming for practical applications. These methods attempt to utilize the task-agnostic knowledge from large models through knowledge distillation. Different from their approaches, we explore sparse training as a pruning technique in order to scale down ViTs from large-scale to small-scale under the Masked Autoencoders framework. + +![](images/77c57227b5868050a5ca693d6c26dd965956994ef58323d46f154debad09561e.jpg) +(a) General pretraining and finetuning + +![](images/1500364156231e2456e525bf77f61489cd891c6da90098c1a2e82b8d5c51f3fc.jpg) +(b) General pretraining and sparse finetuning + +![](images/8d11eb4c23c8a14ac62c1ab443c30eff3294cf945747cb125935bc9ef9abf80d.jpg) +(c) Sparse pretraining and sparse finetuning + +![](images/85315238ec5b5b4751e1e1eecb5802c439d5d9e6afd8486a1604d914ffa065d2.jpg) +(d) SparseMAE pretraining and finetuning +Figure 2: An illustration of four types pretraining and finetuning frameworks. (a) Dense pretraining and finetuning the large full model as done in MAE [15]. $G$ and $D$ are general task-agnostic vision knowledge and task-specific downstream knowledge learnt from pretraining and finetuning respectively. (b) Learning $G$ during dense pretraining with the large full model and acquiring downstream task-specific knowledge $D^{s}$ for the sparse sub-network during sparse finetuning. (c) Learning task-agnostic knowledge $G^{s}$ and task-agnostic knowledge $D^{s}$ with sparse sub-network during sparse pretraining and finetuning. For finetuning, the sparse mask can be kept as fixed or dynamic. (d) Our SparseMAE training framework learns both $G$ and $G^{s}$ during pretraining and subsequently learns better $D^{s}$ during finetuning. + +# 2.2. Pruning and Sparse Training + +Pruning and sparse training are critical methods for reducing the inference costs of deep neural networks [18]. Post-pruning methods [13, 26] typically involve training a dense full network first, followed by one or several rounds of train-prune-retrain rounds. Post-pruning methods are time-consuming due to the iterative manner [11]. Different from post-training methods, sparse training methods [29, 36] aim to adaptively obtain a sparse sub-network while optimizing both model weights and sparse connections from random initialization. + +Sparse training is a class of methods stemming from the Sparse Evolutionary Training (SET) algorithm [29]. It starts from a randomly sparsified network and subsequently prunes and grows connections during training. These methods can be applied end-to-end within the network training stage and have achieved promising results on various vision benchmarks with more efficient models by adaptively changing the model topology connections during training. + +Sparse training methods have shown promising results in various computer vision benchmarks, achieving state-of-the + +art performance on different types of sparsity granularities such as unstructured sparsity [4, 23, 27], N:M fine-grained structured sparsity [42, 32, 10, 41], and block sparsity [7]. However, to our knowledge, the studies of pruning techniques for Vision Transformers [4, 20] only focus on the fully-supervised settings. There still lacks a unified method to prune large-scale Vision Transformers under the unsupervised Masked Autoencoders framework. + +# 3. Methods + +In this section, we first revisit the sparse training technique and Masked Autoencoders (MAE) in Sec. 3.1. Then we emphasize the challenges of integrating existing sparse training methods into MAE in Sec. 3.2. Lastly, we introduce our proposed method to effectively obtain a powerful small-scale sub-network from a large-scale model using sparse training in Sec. 3.3 + +# 3.1. Preliminaries + +Revisiting Sparse Training. The pioneering work of Lottery Ticket Hypothesis (LTH) [11] demonstrates the dense + +networks contain sparse matching subnetworks capable of training in isolation to full accuracy. Its findings inspire searching for strong sub-networks within a larger full network. Sparse training starts from randomly initializing the network weights $\mathbf{W}$ and associating them with a sparse mask $\mathbf{M}$ , which is a binary tensor with 0 indexing the sparsified entries and 1 indicating the retained ones. $\mathbf{M}$ is optimized end-to-end and is usually obtained by magnitude pruning [14]. We denote the sparse sub-network as $\widetilde{\mathbf{W}} = \mathbf{W} \odot \mathbf{M}$ . Sparse-to-sparse training regime optimizes the sparse weights $\widetilde{\mathbf{W}}$ and the corresponding binary mask $\mathbf{M}$ , which can be updated at each iteration or periodically. Variants of sparse training, such as unstructured sparsity and fine-grained structured sparsity [23, 25, 43, 9] achieve state-of-the-art performance on various convolutional networks with improved efficiency. Masked Autoencoders. Masked Autoencoders are a family of self-supervised methods for pretraining Vision Transformers (ViT). They involve reconstructing masked RGB patches from visible tokens $x_{\mathrm{vis}}$ with an encoder $f$ and a decoder $g$ . Mathematically, we have + +$$ +\mathcal {L} _ {\mathrm {M A E}} (\mathbf {W}) = \mathbb {E} \left(\left\| g _ {\theta} \left(f \left(x _ {\text {v i s}}, \mathbf {W}\right)\right) - x _ {\text {m a s k}} \right\| ^ {2}\right). \tag {1} +$$ + +Here, an encoder $f$ maps inputs $x_{\mathrm{vis}}$ to latent features $z_{\mathrm{vis}} = f(x_{\mathrm{vis}})$ , and a decoder $g_{\theta}$ parameterized by $\theta$ predicts the masked token $x_{\mathrm{mask}}$ in the RGB pixel space from the latent feature $z_{\mathrm{vis}}$ . Encoder weights $\mathbf{W}$ and decoder weights $\theta$ are jointly trained. After pretraining, the decoder is discarded and the encoder is finetuned on task-specific downstream datasets. During finetuning, the model acquires downstream specific knowledge $D$ , which is built on general task-agnostic knowledge $G$ learnt in pretraining, as shown in Fig. 2 (a). + +# 3.2. Challenges of Sparse Training for MAE + +Despite the effectiveness of MAE for large-scale vision transformers, it encounters great difficulties when applied to smaller models (see Fig. 1). The inadequate model capacity prevents small models from learning meaning task-agnostic knowledge and consequently transfers to inferior performance in downstream tasks. To enable the potential of large models' capability for smaller ones to learn powerful representations under MAE framework, we can modify Eq. 1 to combine it with sparse training as follows: + +$$ +\mathcal {L} _ {\text {S p a r s e}} (\mathbf {W}, \mathbf {M}) = \mathbb {E} \left(\left\| g _ {\theta} \left(f \left(x _ {\text {v i s}}, \mathbf {W} \odot \mathbf {M}\right)\right) - x _ {\text {m a s k}} \right\| ^ {2}\right), \tag {2} +$$ + +where $f$ represents the large full encoder with parameters $\mathbf{W}$ . The encoder binary mask $\mathbf{M}$ produces a sparse sub-network by magnitude pruning during pretraining, while the decoder is kept dense as it is not used for downstream finetuning. Following the notation in Sec. 3.1, we use $G^{s}$ to represent + +the general knowledge of the sparse sub-network learnt in pretraining, and we denote $D^{s}$ the specific knowledge of the finetuned sparse sub-network for downstream tasks. Considering both the pretraining and finetuning phases, we outline three intuitive strategies to integrate sparse training into the MAE framework: + +- 1 Dense pretraining then sparse finetuning as shown in Fig. 2 (b). Although this strategy allows the sparse finetuning phase to optimize from the large full model's task-agnostic knowledge $G$ , the sub-network does not benefit from the pretraining task and lacks its knowledge $G^{s}$ . This may hinder the model to acquire better downstream knowledge $D^{s}$ +- 2&3 Sparse pretraining using Eq. (2) then sparse finetuning (see Fig. 2 (c)). During finetuning, the sparsity mask can be kept as fixed or dynamic (dynamic mask is the third strategy). This strategy relies on the architecture and task-agnostic knowledge $G^{s}$ obtained by the sub-network during the pretraining phase, while it ignores the potential benefits brought by incorporating the knowledge $G$ of the large full model. + +In summary, it is challenging to tailor a sparse training mechanism for MAE framework to sufficiently learn task-agnostic and task-specific knowledge for a small model. In the following section, we introduce our approach to solve this dilemma by fully utilizing the task-agnostic knowledge $G$ of the large full model and adaptively optimizing sub-network's topology connections for various downstream tasks for better task-specific knowledge. + +# 3.3. Sparse Masked Autoencoders + +To tackle the limitations of the sparse training strategies introduced above, we propose a novel method for pruning ViT models under the MAE framework (see Fig. 2 (d)). + +Firstly, to ensure a good general task-agnostic knowledge $G$ to be learnt via pretraining, we utilize the dense full model to reconstruct the RGB pixels following Eq. 1, which learns more discriminative representations thanks to its larger capacity. Simultaneously, a sparse sub-network within the full dense network is trained using the same visible tokens as the inputs and regresses the same targets as the large full model. The sparseMAE loss function is therefore extended from Eq. 1 and Eq. 2 as follows: + +$$ +\mathcal {L} = \underbrace {\mathcal {L} _ {\text {S p a r s e}} (\widetilde {\mathbf {W}})} _ {\text {s u b - n e t w o r k}} + \underbrace {\alpha \mathcal {L} _ {\mathrm {M A E}} (\mathbf {W})} _ {\text {f u l l n e t w o r k}}, \tag {3} +$$ + +where $\mathbf{W}$ represents the dense parameters of the large full model, $\widetilde{\mathbf{W}} = \mathbf{W} \odot \mathbf{M}$ is the masked sparse parameters of the sub-network, and $\alpha$ is the scaling hyperparameter balancing the reconstruction losses for the dense full model and the + +![](images/22ed134c47fa750236b62ab04a8494baffb2af9d50b552dbe5234ef88e90f38a.jpg) +Figure 3: Illustration of Sparse Masked Autoencoders. In the forward pass, both the large full model (red solid lines) and the sparse sub-network (red dashed lines) reconstruct the masked patches from the same visible tokens. In the backward pass, the masked weights (green solid lines) and the sparse weights (green dashed lines) are updated separately using the gradients from the corresponding losses. + +sparse sub-network. In our experiments, we set $\alpha = 1.0$ by default. + +As shown in Eq. 3, the gradients for the sparse weights $\widetilde{\mathbf{W}}$ shall be updated following both the reconstruction losses of the sub-network and the full network, which may lead to gradient imbalance between the masked and unmasked weights [38]. Therefore, in each training iteration we split the weights $\mathbf{W}$ of the full model into sparse sub-network weights $\widetilde{\mathbf{W}}$ and unmasked weights $\overline{\mathbf{W}} = \mathbf{W} \odot (1 - \mathbf{M})$ and optimize them separately: + +$$ +\widetilde {\mathbf {W}} _ {t + 1} = \widetilde {\mathbf {W}} _ {t} - \eta (g (\mathbf {W} _ {t} \odot \mathbf {M} _ {t}) \odot \mathbf {M} _ {t}), \qquad (4) +$$ + +$$ +\overline {{\mathbf {W}}} _ {t + 1} = \overline {{\mathbf {W}}} _ {t} - \eta (g (\mathbf {W} _ {t}) \odot (1 - \mathbf {M} _ {t})), \tag {5} +$$ + +where $g(\mathbf{W}_t \odot \mathbf{M}_t) = \frac{\partial \mathcal{L}_{\mathrm{Sparse}}(\mathbf{W}_t \odot \mathbf{M}_t)}{\partial \mathbf{W}_t}$ are gradients of the sparse sub-network, $g(\mathbf{W}_t) = \frac{\partial \mathcal{L}_{\mathrm{MAE}}(\mathbf{W}_t)}{\partial \mathbf{W}_t}$ are the gradients of the full models, $\eta$ is the learning rate and $t$ indexes the training iteration. The binary mask $\mathbf{M}_t$ is obtained by on-the-fly magnitude pruning. Here, we only show gradient descent for simplicity and omit other parts (e.g., weight decay). In this way, the large full model can utilize all the model parameters in the forward pass to produce discriminative outputs, while in the backward pass, the sparse sub-network stays unaffected by the gradients from the large full model. During + +the pretraining phase, the task-agnostic knowledge $G^{s}$ of the sparse sub-network can dynamically select the important sparse connections in the context of the task-agnostic knowledge $G$ of the full dense model, which helps to identify the important sparse topology and is retained for downstream finetuning. + +After pretraining, the full dense model with a sparse sub-network is transferred and finetuned on different downstream tasks (e.g., classification, detection, segmentation, etc.). For the finetuning stage, the deep representations from the sparse sub-network are used for the downstream tasks, while its topology connections are optimized end-to-end given the specific downstream objectives. We refrain from fine-tuning dense full models, as the fine-tuning process for both dense and sparse models is costly. Hence, the learning of the task-specific knowledge $D^{s}$ for downstream tasks can benefit from inheriting a pretrained sparse sub-network with knowledge $G^{s}$ , and adaptively selecting the more suitable sparse architecture built on the general knowledge $G$ of the full model. + +# 4. Experiments + +In this section, we validate the effectiveness and robustness of our SparseMAE. First, we introduce the implementation details of our SparseMAE in Sec. 4.1. Then, we present + +our main experimental results in Sec. 4.2, which include classification results on ImageNet-1K [8] and semantic segmentation results on ADE20K [43]. To demonstrate the superiority of our proposed method, we compare SparseMAE with other state-of-the-art ViT pruning methods in Sec. 4.4. In Sec. 4.3, we evaluate the transferring ability of the models trained with our SparseMAE to object detection, instance segmentation tasks on MS COCO [24] and robustness evaluation. Finally, we perform ablative experiments to evaluate the performance of SparseMAE under different settings in Sec. 4.5. + +# 4.1. Implementation Details + +Sparse Training Methods. We conduct experiments with two different granularities for sparse training, consisting of unstructured sparsity [14, 21] and hardware-friendly N:M sparsity [28, 42]. As we focus on studying how to learn and transfer task-agnostic and task-specific knowledge under the Masked Autoencoders paradigm with sparse training methods, we select two simple and effective sparse training methods, namely DFP [23] for unstructured sparsity and SR-STE [42] for N:M sparsity. We prune all the linear layers within the vision transformers and set the relevant hyperparameters for sparse training the same as [23, 42] by default. + +Models. For all of our main experiments, unless otherwise stated, we select ViT-Base as our large full model and prune it for smaller ones. For hardware-friendly N:M sparsity, we use 2:32 and 4:32 sparse ratios to obtain models with similar scales to ViT-Tiny (5.8M) and ViT-Small (11.3M), the main experiments use N:M sparsity for default, which is applied to all linear layers, including self.qkv, self.proj in the Attention block, and self.fc1 and self.fc2 in the MLP block. For unstructured sparsity, we extend the above N:M ratios and prune ViT-Base with $93.5\%$ and $87.0\%$ sparsity. We name our ViT-Base with 2:32/93.5% sparsity as SparseMAE-Tiny (SparseMAE-T) and the one with 4:32/87.0% sparsity as SparseMAE-Small (SparseMAE-S). + +Pretraining. Following MAE [15] pretraining scheme, models start from random initialization, we use ImageNet-1K (IN-1K) [8] for all pretraining experiments. We pretrain SparseMAE for 400 epochs using $224 \times 224$ images as inputs. The other settings also follow those used in [15]. For example, We use random resized cropping and horizontal flipping for data augmentation, a $75\%$ mask ratio during pretraining, and a decoder with a single transformer layer and 512 hidden dimensions. Unlike other methods [34, 19] for small and tiny-scale vision transformers pretraining, we do not grid-search the decoder design, possible putting our method at a disadvantage. + +Finetuning. We evaluate our SparseMAE pretrained models for various downstream tasks, including ImageNet-1K [8] image classification, MS COCO [24] object detection and + +instance segmentation tasks, ADE20K [43] semantic segmentation. + +For image classification, we take a ViT-Base model pretrained using MAE for 400 epochs as the reference model, which is implemented using the officially released codebase [15] and achieves $83.2\%$ top-1 accuracy. We finetune SparseMAE-Tiny/Small for 200 epochs using the same receipt described in [15] for a fair comparison with other works [31, 19] and set the batch size as 1024 and the base learning rate as $10^{-3}$ . + +For semantic segmentation, we follow BEiT [2] to use UperNet[37] as our segmentation framework. We train the model with an input resolution of $512 \times 512$ for 160k iterations. A batch size of 16 is adopted. For object detection and instance segmentation tasks, we follow ViTDet [22] framework. We train our SparseMAE models with a batch size of 64 for 100 epochs, and the input image resolution is set to $1024 \times 1024$ . Further details on the hyper-parameters can be found in Appendix. + +# 4.2. Main Results + +Classification and Semantic Segmentation. Tab. 1 shows the evaluation of our SparseMAE model against self-supervised methods for ViT-Tiny/Small on ImageNet-1K image classification and ADE20K semantic segmentation. For ImageNet-1K classification, SparseMAE achieves $80.5\%$ top-1 accuracy. To our knowledge, this is currently the best result achieved by tiny-scale transformer models. Furthermore, SparseMAE outperforms TinyMIM [31]/G2SD [19]/MAE-lite [34] with $+4.7 / 4.2 / 4.0$ accuracy gains on tiny-scale transformers. For semantic segmentation, SparseMAE achieves an mIoU of 45.2, outperforming G2SD/MAE-lite/TinyMIM by $+3.8 / 1.3 / 1.2$ mIoUs. Similar improvements can be observed for the small-scale transformers. SparseMAE achieves a $83.2\%$ ImageNet-1K classification accuracy and a $48.4\%$ mIoU on ADE20K semantic segmentation, surpassing TinyMIM [31]/G2SD [19] by $+0.2 / 1.2$ and $+0.0 / 2.2$ on the two tasks respectively. The significant improvements over previous results show the superiority of our method in obtaining performant small models under MAE framework. + +Speedups. Unstructured sparsity and N:M structured sparsity is efficient for commercial CPUs and FPGAs respectively. For our SparseMAE with unstructured sparsity, we tested the throughput (img/s) on an Intel RXeon(R) 6238R CPU using the deepsparse engine with a batch size of 64. For our SparseMAE with N:M sparsity, we tested the throughput on Xilinx XCVU13P with a batch size of 16. Our SparseMAE-S with N:M and unstructured sparsity achieved throughput gains of $+71\%$ and $+9\%$ over ViT-small with similar amount of parameters on FPGAs and CPUs. Therefore, our approach demonstrate high feasibility in actual scenarios. + +Table 1: Finetuning results on ImageNet-1K and ADE20K. All our models are pre-trained only on ImageNet-1K. Extra pretrained teacher means the compared method cannot trained from scratch. $\ddagger$ the model adopts the MAE pretrained model before performing the sparse pretraining. + +
MethodParams (M)Pretraining epochsEncoder ratioExtra Pretrained/ TeacherClassification Top-1 Acc (%)Segmentation mIoU (%)
DeiT [33]5.8300-Label72.238.0
Tiny-scale
MAE [15]5.8160025%X71.637.6
MoCo [6]5.8300100%EMA73.339.3
TinyMIM [31]5.8300100%TinyMIM-S75.844.0
G2SD [19]5.8300100%ViT-B76.341.4
MAE-lite [34]5.840025%ViT-B76.543.9
SparseMAE (ours)5.840025%X80.545.2
DeiT [33]22300-Label79.9
Small-scale
MAE [15]22160025%X80.642.8
MoCo [6]22300100%EMA81.443.9
DINO [3]22300100%EMA81.545.3
CAE [5]22300100%DALL-E82.0-
TinyMIM [31]22300100%TinyMIM-ViT-B83.048.4
G2SD [19]2230025%ViT-B82.046.2
SparseMAE (ours)11.330025%ViT-B‡83.248.4
SparseMAE (ours)11.340025%X82.146.7
+ +# 4.3. Evaluation on Other Downstream Tasks + +Object Detection and Instance Segmentation. We finetune the SparseMAE pretrained models on MS COCO with ViT-Det [22]. As shown in Tab. 2, we report $\mathrm{AP^{bbox}}$ for object detection and $\mathrm{AP^{mask}}$ for instance segmentation. We compare SparseMAE-Tiny with MAE-Tiny and its improved variants using distillation methods, such as MAE-lite [34] and G2SD [19]. The results in Tab. 2 show that our Sparse-MAE obtains more than $3.1\mathrm{AP^{bbox}}$ and $2.4\mathrm{AP^{mask}}$ gains compared to MAE-lite and G2SD, validating the transferability of the learnt representations in our models. + +Robustness Evaluation. In Tab. 3, we evaluate the robustness of our models on different variants of ImageNet validation sets. We use the same models finetuned on original ImageNet (Tab. 1) and only run inference on the different validation sets, such asImagenet-R [17], ImageNet-A [17] and ImageNet-C [16]), without any specialized fine-tuning. Tab. 3 shows that our method has better generalization capability than distillation-based methods, such as G2SD [19] and TinyMIM [31]). + +# 4.4. Comparison with State-of-the-art ViT Pruning + +To the best of our knowledge, our proposed SparseMAE is the first to study network pruning technique under the + +Table 2: Object detection and instance segmentation results on the MS COCO dataset. + +
MethodBackbone Size (M)Det APbboxSeg APmask
Tiny-scale models
DeiT [33]5.840.736.5
MAE [15]5.838.935.1
MoCo v3 [6]5.840.036.0
DINO [3]5.840.236.1
MAE-lite [34]5.842.738.2
G2SD [19]5.844.039.6
SparseMAE (ours)5.847.142.0
+ +Masked Autoencoders paradigm. To validate the effectiveness of SparseMAE, we compare it against recent state-of-the-art fully-supervised pruning methods for ViTs. We choose SViT [4] and oViT [20] as our compared models, which exploit end-to-end sparse training for ViTs and achieve state-of-the-art results in the fully-supervised setting. SViT [4] adopts a dynamic sparse training method to prune the dense model using random initialization. However, this approach requires training the sparse model for + +Table 3: Robustness evaluation on ImageNet variants (top-1 accuracy, except for ImageNet-C which evaluates mean corruption error). We test the small-scale and tiny-scale models from Tab. 1 on different ImageNet validation sets, without any fine-tuning. + +
MethodModel SizeIN-A↑IN-R↑IN-C↓
DeiT [33]8.032.754.0
MAE [15]7.036.555.2
TinyMIM [31]Tiny-scale11.039.850.1
G2SD [19]12.939.0-
SparseMAE (ours)18.245.947.2
DeiT [33]18.342.341.4
MAE [15]20.145.640.6
TinyMIM [31]Small-scale27.548.835.8
G2SD [19]29.446.8-
SparseMAE (ours)29.349.235.2
+ +Table 4: Comparison with state-of-the-art supervised Vision Transformer pruning methods on ImageNet-1K. + +
ModelMethodParams (M)Top-1 Acc (%)
ViT-BaseSViT [4]43.381.5
oViT [20]43.381.6
SViT [4]34.681.3
oViT [20]34.681.5
oViT [20]21.681.1
oViT [20]17.380.8
oViT [20]8.779.7
ViT-BaseSparseMAE-T (ours)5.880.5
SparseMAE-S (ours)11.383.2
+ +600 epochs. oViT [20] (post-training pruning) prunes the model initialized from a trained dense model (300 epochs), which requires 300 epochs of finetuning. For our SparseMAE, we only need 400 epochs for pretraining and 200 epochs for fine-tuning. The pretraining of SparseMAE is $3 \times$ faster than full-token supervised training. Consequently, our SparseMAE approach reduces the training time from 968 GPU hours to 581 GPU hours compared to SOTA SViT and oViT. + +Tab. 4 demonstrates that SparseMAE can outperform state-of-the-art fully-supervised pruning counterparts significantly for small and tiny-scale transformers. Our SparseMAE-T has a $0.8\%$ performance gain and a $35.0\%$ computation complexity reduction compared to oViT [20] for tiny-scale transformers. With our SparseMAE-S, we achieve a $83.2\%$ ImageNet classification accuracy, outperforming supervised pruning techniques by large margins. These results suggest that our method successfully extends the superiority of MAE for representation learning with large-scale vision transformers to small and tiny-scale models. + +Table 5: Different sparse pretraining or finetuning strategies with tiny-scale models on ImageNet-1K and ADE20K. "S" means sparse pretraining, "D" means dense pretraining, and "D+S" means sparse pretraining together with a dense full model pretraining. $\ddagger$ : the model is finetuned with the fixed sparse masks from pretraining. Corr. Strategy.: corresponding sparse training strategies explained in Sec. 4.5.1. + +
ModelPretrain MethodTop-1 Acc (%)Seg mIoU (%)Corr. Strtg.
ViT-BaseD83.248.1-
Unstructured sparsity (93.5% sparse ratio)
SparseMAE-TD77.840.2#1
SparseMAE-TS‡76.737.7#2
SparseMAE-TS77.639.8#3
SparseMAE-TD+S80.445.4-
2:32 sparsity
SparseMAE-TD78.440.7#1
SparseMAE-TS‡76.337.2#2
SparseMAE-TS78.741.3#3
SparseMAE-TD+S80.545.2-
+ +# 4.5. Ablation Studies and Discussion + +# 4.5.1 Comparison of Sparse Training Strategies. + +We compare three sparse training strategies for MAE framework introduced in Sec. 3.2 under two sparse granularities. Namely, we (1) pretrain a large full model using MAE and finetune it with sparse training, (2) combine sparse training with MAE pretraining and subsequently finetune the model on downstream tasks with the fixed sparse mask from pretraining, or (3) modify (2) to use dynamic sparse mask for finetuning. The results are shown in Tab. 5. Under unstructured sparsity, comparing (1) and (2), we see that the classification performance drops from $77.8\%$ to $76.7\%$ , showing that the knowledge $G^{s}$ learnt by the sparse sub-network during pretraining cannot generalize well to the downstream tasks compared to the task-agnostic knowledge $G$ learnt by the large full model. (3) further allows the adaptation of sparse sub-network architecture during finetuning and partially alleviates the problem by better acquiring downstream specific knowledge $D^{s}$ . However, its result still lags significantly behind our proposed method ( $77.6\%$ vs. $80.4\%$ ), where the difference is only the absence of a concurrently pretrained large full model. On the other hand, when comparing (1) with our method, the addition of a pretrained sparse sub-network can boost the performance by $2.6\%$ . Similar observations can also be made under N:M sparsity settings and semantic segmentation results. These empirical results suggest that both the task-agnostic knowledge $G$ of the large full model and $G^{s}$ of the sparse sub-network are necessary for small and tiny-scale vision transformers to effectively benefit from MAE pretraining. + +Table 6: Comparison between different full models for SparseMAE on ImageNet-1K and ADE20K. + +
ModelPatternParams (M)Top-1 Acc (%)Seg mIoU (%)
ViT-Small8:325.877.441.5
ViT-Base2:325.880.545.2
+ +# 4.5.2 Ablation of Full Model Capacity + +We conduct an ablation study to investigate the performance with different capacities of the full model. We use our Sparse-MAE and set ViT-Small and ViT-Base as the full models, where ViT-Small only has $25\%$ parameters of ViT-Base. As shown in Tab. 6, the results show that our SparseMAE in ViT-Base achieves $3.1\%$ top-1 accuracy advantage over ViT-Small, indicating that larger capacity of the full model with SparseMAE can significantly improve its performance by incorporating better pretraining task-agnostic knowledge. + +# 4.5.3 Impact of Different Pruning Sparsity Ratios + +The main experiments of sparse ratio setting is compared to the ViT-tiny and ViT-small. We provide additional sparse ratios to verify the performance of SparseMAE on ImageNet-1K as in Tab. 7. + +Table 7: Comparison between different sparse ratios for SparseMAE on ImageNet-1K. + +
Sparse Pattern1:322:324:328:32
Top-1 Acc76.880.583.283.7
+ +# 5. Conclusion + +In this paper, we introduce SparseMAE, an innovative method for incorporating sparse training into the Masked Autoencoders (MAE) framework to create efficient and powerful small and tiny-scale transformers. As the first study in this area, we first point out the limitations of sparse training under MAE and design a unified framework to transfer the task-agnostic and task-specific knowledge of large models to lightweight sparse sub-networks. SparseMAE trains the sparse sub-networks by persevering the task-agnostic knowledge of large full models and then adaptively finds the optimal specific sub-networks for downstream tasks. The proposed method outperforms the state-of-the-art small-scale vision transforms in both unsupervised pretraining and fully-supervised settings. We hope this work could shed light on sparse training with the Masked Autoencoders framework in the vision community. + +# Acknowledgement + +This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. + +# References + +[1] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. +[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. +[4] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34:19974-19988, 2021. +[5] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022. +[6] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. +[7] Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. arXiv preprint arXiv:2205.14135, 2022. +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[9] Utku Evci, Trevor Gale, Jacob Menick, Pablo Samuel Castro, and Erich Olsen. Rigging the lottery: Making all tickets winners. In International Conference on Machine Learning, pages 2943-2952. PMLR, 2020. +[10] Chao Fang, Aojun Zhou, and Zhongfeng Wang. An algorithm–hardware co-optimized framework for accelerating n: M sparse transformers. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 30(11):1573–1586, 2022. +[11] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635, 2018. + +[12] Peng Gao, Renrui Zhang, Rongyao Fang, Ziyi Lin, Hongyang Li, Hongsheng Li, and Qiao Yu. Mimic before reconstruct: Enhancing masked autoencoders with feature mimicking. IJCV 2023, 2023. +[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. +[14] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. +[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. +[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. +[17] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. +[18] Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks. The Journal of Machine Learning Research, 22(1):10882-11005, 2021. +[19] Wei Huang, Zhiliang Peng, Li Dong, Furu Wei, Jianbin Jiao, and Qixiang Ye. Generic-to-specific distillation of masked autoencoders. arXiv preprint arXiv:2302.14771, 2023. +[20] Denis Kuznedelev, Eldar Kurtic, Elias Frantar, and Dan Alistarh. ovit: An accurate second-order pruning framework for vision transformers. arXiv preprint arXiv:2210.09223, 2022. +[21] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. Advances in neural information processing systems, 2, 1989. +[22] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 280-296. Springer, 2022. +[23] Tao Lin, Sebastian U Stich, Luis Barba, Daniil Dmitriev, and Martin Jaggi. Dynamic model pruning with feedback. arXiv preprint arXiv:2006.07253, 2020. +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. +[25] Shiwei Liu, Lu Yin, Decebal Constantin Mocanu, and Mykola Pechenizkiy. Do we actually need dense overparameterization? in-time over-parameterization in sparse training. In International Conference on Machine Learning, pages 6989-7000. PMLR, 2021. + +[26] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018. +[27] Zhi-Gang Liu, Paul N Whatmough, and Matthew Mattina. Sparse systolic tensor array for efficient cnn hardware acceleration. arXiv preprint arXiv:2009.02381, 2020. +[28] Asit Mishra, Jorge Albericio Latorre, Jeff Pool, Darko Stosic, Dusan Stosic, Ganesh Venkatesh, Chong Yu, and Paulius Miccikevicius. Accelerating sparse deep neural networks. arXiv preprint arXiv:2104.08378, 2021. +[29] Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H Nguyen, Madeleine Gibescu, and Antonio Liotta. Scalable training of artificial neural networks with adaptive sparse connectivity inspired by network science. Nature communications, 9(1):2383, 2018. +[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. +[31] Sucheng Ren, Fangyun Wei, Zheng Zhang, and Han Hu. Tinymim: An empirical study of distilling mim pre-trained models. arXiv preprint arXiv:2301.01296, 2023. +[32] Wei Sun, Aojun Zhou, Sander Stuijk, Rob G. J. Wijnhoven, Andrew Nelson, Hongsheng Li, and Henk Corporaal. Dominosearch: Find layer-wise fine-grained n:m sparse schemes from dense neural networks. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. +[33] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. +[34] Shaoru Wang, Jin Gao, Zeming Li, Jian Sun, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. arXiv preprint arXiv:2205.14443, 2022. +[35] Chen Wei, Haoqi Fan, Saining Xie, Chao Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. 2021. +[36] Mitchell Wortsman, Ali Farhadi, and Mohammad Rastegari. Discovering neural wireings. Advances in Neural Information Processing Systems, 32, 2019. +[37] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. +[38] Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. Slimmable neural networks. arXiv preprint arXiv:1812.08928, 2018. +[39] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. NeurIPS 2022, 2022. +[40] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. CVPR 2023, 2023. + +[41] Yuxin Zhang, Mingbao Lin, Zhihang Lin, Yiting Luo, Ke Li, Fei Chao, Yongjian Wu, and Rongrong Ji. Learning best combination for efficient n:m sparsity. arXiv preprint arXiv:2206.06662, 2022. +[42] Aojun Zhou, Yukun Ma, Junnan Zhu, Jianbo Liu, Zhijie Zhang, Kun Yuan, Wenxiu Sun, and Hongsheng Li. Learning n: M fine-grained structured sparse neural networks from scratch. arXiv preprint arXiv:2102.04010, 2021. +[43] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. \ No newline at end of file diff --git a/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/images.zip b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8892f5812850c5a1c355ab3770a10a1c8da65780 --- /dev/null +++ b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cab5e9d771313d618bbdee07df8dcb477ee6e77d03f850a6c8bbdf54399ec64d +size 492036 diff --git a/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/layout.json b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d2a2f27636a8b52c104bf1c202271090d13fb961 --- /dev/null +++ b/2023/SparseMAE_ Sparse Training Meets Masked Autoencoders/layout.json @@ -0,0 +1,8169 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 479, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 479, + 121 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 479, + 121 + ], + "type": "text", + "content": "SparseMAE: Sparse Training Meets Masked Autoencoders" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "spans": [ + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": "Aojun Zhou" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": " Yang Li" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": " Zipeng Qin" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": " Jianbo Liu" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": " Junting Pan" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": " Renrui Zhang" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{13}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": " Rui Zhao" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": " Peng Gao" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": " Hongsheng Li" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": "The Chinese University of Hong Kong \n" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 125, + 142, + 467, + 200 + ], + "type": "text", + "content": "SenseTime Research Shanghai AI Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 252, + 290, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 563 + ], + "type": "text", + "content": "Masked Autoencoders (MAE) and its variants have proven to be effective for pretraining large-scale Vision Transformers (ViTs). However, small-scale models do not benefit from the pretraining mechanisms due to limited capacity. Sparse training is a method of transferring representations from large models to small ones by pruning unimportant parameters. However, naively combining MAE finetuning with sparse training make the network task-specific, resulting in the loss of task-agnostic knowledge, which is crucial for model generalization. In this paper, we aim to reduce model complexity from large vision transformers pretrained by MAE with assistant of sparse training. We summarize various sparse training methods to prune large vision transformers during MAE pretraining and finetuning stages, and discuss their shortcomings. To improve learning both task-agnostic and task-specific knowledge, we propose SparseMAE, a novel two-stage sparse training method that includes sparse pretraining and sparse finetuning. In sparse pretraining, we dynamically prune a small-scale sub-network from a ViT-Base. During finetuning, the sparse sub-network adaptively changes its topology connections under the task-agnostic knowledge of the full model. Extensive experimental results demonstrate the effectiveness of our method and its superiority on small-scale vision transformers. Code will be available at https://github.com/aojunzz/SparseMAE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 588, + 128, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 128, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 128, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 609, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 288, + 693 + ], + "type": "text", + "content": "Recently, several pretrained Masked Image Modeling (MIM) methods, such as MAE [15] and data2vec [1], have achieved great success in various computer vision tasks. They usually involve a two-stage training scheme where the model learns task-agnostic knowledge through a pretraining pretext task, and is subsequently finetuned on downstream tasks to acquire task-specific knowledge. Among" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 228, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 547, + 456 + ], + "type": "text", + "content": "these models, Masked Autoencoders (MAE) demonstrate superior learning capability on large-scale vision transformers (Fig. 1 (a)), thanks to large models' strong capacity to learn powerful general representations in the pretraining phase and versatile transferability to specific vision tasks. Yet, the large deep models are burdensome and difficult to be deployed to computationally restricted real-world scenarios. On the other hand, smaller and more efficient models, such as ViT-Tiny and ViT-Small, fail to perform well after pretrained by MAE or data2vec frameworks. For example, ViT-Tiny's finetuning results are even inferior to fully-supervised training (see Fig. 1 (b)). These empirical findings suggest that, under MIM frameworks, when the model is scaled down, it becomes more difficult to learn well via masked pretraining due to its limited capacity during pretraining and consequently hinders its performance when transferred to downstream tasks. In the realm of MIM pretraining, model capacity becomes a key ingredient to learning task-agnostic knowledge via the pretext tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 548, + 714 + ], + "type": "text", + "content": "In our studies, we try to tackle the issues of MIM pretraining for small-scale vision transformers. Particularly, we apply pruning techniques [28, 20, 4, 42, 23] in the MIM frameworks in order to obtain performant small-scale models from larger ones. As a first step in our exploration, we use existing sparse training methods [42, 23] for MAE pretraining and finetuning on unstructured and hardware-friendly N:M sparsity to obtain a model with a similar scale to ViT-Tiny, which is pruned from ViT-Base (see Tab. 5 row 3). It achieves notably improvements compared to directly training ViT-Tiny using MAE on downstream ImageNet finetuning accuracy (77.6% vs. 71.6%) and ADE20K [43] semantic segmentation (39.8 mIoU vs. 37.6 mIoU). The empirical results indicate that on small-scale ViTs sparse training via pruning has promising applications under the MIM pretraining paradigm. However, there is still a large performance gap between the sparse model and the vanilla ViT-Base trained densely under MAE in their ImageNet finetuning accuracies (77.6% vs. 83.2%). The sparsity constraints adversely limit the model capacity and prevent it from learning good task-agnostic knowledge, similar to the predicament faced" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 233, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 233, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 233, + 714 + ], + "type": "text", + "content": "*The first two authors equally contribute to this paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16176" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 83, + 281, + 255 + ], + "blocks": [ + { + "bbox": [ + 50, + 83, + 281, + 255 + ], + "lines": [ + { + "bbox": [ + 50, + 83, + 281, + 255 + ], + "spans": [ + { + "bbox": [ + 50, + 83, + 281, + 255 + ], + "type": "image", + "image_path": "f44911641785043be511cd7ead7c035e0e55ece70627b6f73874a9925ce4ab57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 268, + 288, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 288, + 342 + ], + "type": "text", + "content": "Figure 1: (a) Top-1 accuracy of large-scale models on ImageNet-1K, showing that large-scale vision transformers and self-supervised training can bring significant improvement. (b) On tiny-scale and small-scale, our Sparse-MAE outperforms other self-supervised methods, such as data2vec [1], MAE [15], and training from scratch." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 363, + 268, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 268, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 268, + 375 + ], + "type": "text", + "content": "by ViT-Tiny when it is directly pretrained under MAE." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 543 + ], + "type": "text", + "content": "To obtain more powerful small-scale models, we propose SparseMAE, a simple and unified sparse training method for Masked Autoencoders. Specifically, SparseMAE trains an adaptive sub-network in conjunction with its corresponding large dense model during pretraining. Both the large model and the sparse sub-network predict the same reconstruction objective and are optimized end-to-end. The large full model provides a strong capacity to learn task-agnostic knowledge, while the sparse sub-network can better preserve the knowledge by rewiring the connections in the full model. To better learn task-specific knowledge in downstream tasks, the sparse sub-network inherits the task-agnostic knowledge from the pretraining stage and also dynamically adjusts its connections during the finetuning phase." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 544, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 662 + ], + "type": "text", + "content": "We conduct extensive experiments on a variety of tasks, including image classification, object detection, segmentation, and robustness evaluation, to thoroughly evaluate the effectiveness of our proposed approach. Notably, our method achieves a top-1 accuracy of " + }, + { + "bbox": [ + 46, + 544, + 287, + 662 + ], + "type": "inline_equation", + "content": "80.5\\%" + }, + { + "bbox": [ + 46, + 544, + 287, + 662 + ], + "type": "text", + "content": " with only 6M parameters on ImageNet-1K [8] classification and obtains a 45.2 mIoU on ADE20K segmentation with UperNet [37], significantly outperforming all previous works that have developed small-scale vision transformers using MAE (e.g., MAE-lite, TinyMIM and G2SD distillation)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 664, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 664, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 59, + 664, + 288, + 676 + ], + "type": "text", + "content": "In all, our contributions can be summarized in three folds:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 689, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 289, + 714 + ], + "type": "text", + "content": "- We thoroughly study the integration of sparse training as a pruning technique into the MAE framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 72, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 547, + 108 + ], + "type": "text", + "content": "Specifically, we design three different strategies to train sparse networks under MAE and analyze their limitations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 122, + 547, + 256 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 122, + 545, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 122, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 122, + 545, + 183 + ], + "type": "text", + "content": "- Based on the findings from above, we propose a novel sparse training framework, SparseMAE for Masked Autoencoders to improve the acquisition of both the task-agnostic and task-specific knowledge during the pretraining and finetuning stages." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 198, + 547, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 547, + 256 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 547, + 256 + ], + "type": "text", + "content": "- We present extensive experiments to validate the effectiveness of our proposed method and achieve state-of-the-art performance with small-scale vision transformers on ImageNet classification and various downstream tasks." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 288, + 395, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 395, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 395, + 300 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 308, + 444, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 444, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 444, + 322 + ], + "type": "text", + "content": "2.1. Masked Image Modeling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 329, + 547, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 496 + ], + "type": "text", + "content": "Unsupervised pretraining on large-scale images with Masked Image Modeling (MIM) has shown superior performance on various computer vision tasks. BEiT [2] explores Masked Image Modeling on vision transformers by reconstructing the vision dictionary [30]. MAE [15] further proposes an asymmetric encoder and decoder for scaling up MIM to huge models. Besides, it demonstrates a simple pixel reconstruction loss can learn good visual representations. Due to the simplicity and computational efficiency, MAE is raising to a popular generative pretraining paradigm [40, 39]. As MAE reconstructs low-level signals with an isotropic vision transformer architecture, researchers improve MAE by exploring high-level signals architectures [1, 35, 12], which are more effective than reconstructing low-level signals." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 548, + 713 + ], + "type": "text", + "content": "However, those methods demonstrate the large capacity model is essential for good representation learning through the MIM pretext tasks. MAE-lite [34] explores Masked Autoencoders to improve the performance of tiny-scale vision transformer models, and introduces knowledge distillation into the pretraining phase. Different from MAE-lite, TinyMIM [31] proposes sequential relation knowledge distillation to improve the performance of small-scale vision transformer models, which utilize the full set of patch tokens and distill the knowledge from a pretrained ViT-Large model. TinyMIM [31] needs the large pretrained vision transformer and the use of sequential models is time-consuming for practical applications. These methods attempt to utilize the task-agnostic knowledge from large models through knowledge distillation. Different from their approaches, we explore sparse training as a pruning technique in order to scale down ViTs from large-scale to small-scale under the Masked Autoencoders framework." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16177" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 68, + 271, + 148 + ], + "blocks": [ + { + "bbox": [ + 111, + 68, + 271, + 148 + ], + "lines": [ + { + "bbox": [ + 111, + 68, + 271, + 148 + ], + "spans": [ + { + "bbox": [ + 111, + 68, + 271, + 148 + ], + "type": "image", + "image_path": "77c57227b5868050a5ca693d6c26dd965956994ef58323d46f154debad09561e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 163, + 263, + 177 + ], + "lines": [ + { + "bbox": [ + 104, + 163, + 263, + 177 + ], + "spans": [ + { + "bbox": [ + 104, + 163, + 263, + 177 + ], + "type": "text", + "content": "(a) General pretraining and finetuning" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 321, + 68, + 477, + 157 + ], + "blocks": [ + { + "bbox": [ + 321, + 68, + 477, + 157 + ], + "lines": [ + { + "bbox": [ + 321, + 68, + 477, + 157 + ], + "spans": [ + { + "bbox": [ + 321, + 68, + 477, + 157 + ], + "type": "image", + "image_path": "1500364156231e2456e525bf77f61489cd891c6da90098c1a2e82b8d5c51f3fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 160, + 460, + 185 + ], + "lines": [ + { + "bbox": [ + 345, + 160, + 460, + 185 + ], + "spans": [ + { + "bbox": [ + 345, + 160, + 460, + 185 + ], + "type": "text", + "content": "(b) General pretraining and sparse finetuning" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 112, + 217, + 269, + 306 + ], + "blocks": [ + { + "bbox": [ + 112, + 217, + 269, + 306 + ], + "lines": [ + { + "bbox": [ + 112, + 217, + 269, + 306 + ], + "spans": [ + { + "bbox": [ + 112, + 217, + 269, + 306 + ], + "type": "image", + "image_path": "8d11eb4c23c8a14ac62c1ab443c30eff3294cf945747cb125935bc9ef9abf80d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 308, + 281, + 320 + ], + "lines": [ + { + "bbox": [ + 100, + 308, + 281, + 320 + ], + "spans": [ + { + "bbox": [ + 100, + 308, + 281, + 320 + ], + "type": "text", + "content": "(c) Sparse pretraining and sparse finetuning" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 324, + 193, + 481, + 306 + ], + "blocks": [ + { + "bbox": [ + 324, + 193, + 481, + 306 + ], + "lines": [ + { + "bbox": [ + 324, + 193, + 481, + 306 + ], + "spans": [ + { + "bbox": [ + 324, + 193, + 481, + 306 + ], + "type": "image", + "image_path": "85315238ec5b5b4751e1e1eecb5802c439d5d9e6afd8486a1604d914ffa065d2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 308, + 491, + 321 + ], + "lines": [ + { + "bbox": [ + 315, + 308, + 491, + 321 + ], + "spans": [ + { + "bbox": [ + 315, + 308, + 491, + 321 + ], + "type": "text", + "content": "(d) SparseMAE pretraining and finetuning" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "text", + "content": "Figure 2: An illustration of four types pretraining and finetuning frameworks. (a) Dense pretraining and finetuning the large full model as done in MAE [15]. " + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "text", + "content": " are general task-agnostic vision knowledge and task-specific downstream knowledge learnt from pretraining and finetuning respectively. (b) Learning " + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "text", + "content": " during dense pretraining with the large full model and acquiring downstream task-specific knowledge " + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "D^{s}" + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "text", + "content": " for the sparse sub-network during sparse finetuning. (c) Learning task-agnostic knowledge " + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "text", + "content": " and task-agnostic knowledge " + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "D^{s}" + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "text", + "content": " with sparse sub-network during sparse pretraining and finetuning. For finetuning, the sparse mask can be kept as fixed or dynamic. (d) Our SparseMAE training framework learns both " + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "text", + "content": " during pretraining and subsequently learns better " + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "D^{s}" + }, + { + "bbox": [ + 46, + 346, + 545, + 430 + ], + "type": "text", + "content": " during finetuning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 450, + 207, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 207, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 207, + 464 + ], + "type": "text", + "content": "2.2. Pruning and Sparse Training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 471, + 288, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 288, + 589 + ], + "type": "text", + "content": "Pruning and sparse training are critical methods for reducing the inference costs of deep neural networks [18]. Post-pruning methods [13, 26] typically involve training a dense full network first, followed by one or several rounds of train-prune-retrain rounds. Post-pruning methods are time-consuming due to the iterative manner [11]. Different from post-training methods, sparse training methods [29, 36] aim to adaptively obtain a sparse sub-network while optimizing both model weights and sparse connections from random initialization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": "Sparse training is a class of methods stemming from the Sparse Evolutionary Training (SET) algorithm [29]. It starts from a randomly sparsified network and subsequently prunes and grows connections during training. These methods can be applied end-to-end within the network training stage and have achieved promising results on various vision benchmarks with more efficient models by adaptively changing the model topology connections during training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "Sparse training methods have shown promising results in various computer vision benchmarks, achieving state-of-the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 451, + 547, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 547, + 546 + ], + "type": "text", + "content": "art performance on different types of sparsity granularities such as unstructured sparsity [4, 23, 27], N:M fine-grained structured sparsity [42, 32, 10, 41], and block sparsity [7]. However, to our knowledge, the studies of pruning techniques for Vision Transformers [4, 20] only focus on the fully-supervised settings. There still lacks a unified method to prune large-scale Vision Transformers under the unsupervised Masked Autoencoders framework." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 558, + 366, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 366, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 366, + 571 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 578, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 662 + ], + "type": "text", + "content": "In this section, we first revisit the sparse training technique and Masked Autoencoders (MAE) in Sec. 3.1. Then we emphasize the challenges of integrating existing sparse training methods into MAE in Sec. 3.2. Lastly, we introduce our proposed method to effectively obtain a powerful small-scale sub-network from a large-scale model using sparse training in Sec. 3.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 670, + 392, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 392, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 392, + 682 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "Revisiting Sparse Training. The pioneering work of Lottery Ticket Hypothesis (LTH) [11] demonstrates the dense" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16178" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "content": "networks contain sparse matching subnetworks capable of training in isolation to full accuracy. Its findings inspire searching for strong sub-networks within a larger full network. Sparse training starts from randomly initializing the network weights " + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "content": " and associating them with a sparse mask " + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "content": ", which is a binary tensor with 0 indexing the sparsified entries and 1 indicating the retained ones. " + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "content": " is optimized end-to-end and is usually obtained by magnitude pruning [14]. We denote the sparse sub-network as " + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{W}} = \\mathbf{W} \\odot \\mathbf{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "content": ". Sparse-to-sparse training regime optimizes the sparse weights " + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{W}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "content": " and the corresponding binary mask " + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "content": ", which can be updated at each iteration or periodically. Variants of sparse training, such as unstructured sparsity and fine-grained structured sparsity [23, 25, 43, 9] achieve state-of-the-art performance on various convolutional networks with improved efficiency. Masked Autoencoders. Masked Autoencoders are a family of self-supervised methods for pretraining Vision Transformers (ViT). They involve reconstructing masked RGB patches from visible tokens " + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{vis}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "content": " with an encoder " + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 72, + 289, + 314 + ], + "type": "text", + "content": ". Mathematically, we have" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 335, + 288, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 335, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 74, + 335, + 288, + 350 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M A E}} (\\mathbf {W}) = \\mathbb {E} \\left(\\left\\| g _ {\\theta} \\left(f \\left(x _ {\\text {v i s}}, \\mathbf {W}\\right)\\right) - x _ {\\text {m a s k}} \\right\\| ^ {2}\\right). \\tag {1}", + "image_path": "33e3415e171c8029a21ae5a9f09c23e13d4fffe8b686abfebf9d5d33be59fe14.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": "Here, an encoder " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " maps inputs " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{vis}}" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " to latent features " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{vis}} = f(x_{\\mathrm{vis}})" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": ", and a decoder " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " predicts the masked token " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{mask}}" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " in the RGB pixel space from the latent feature " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{vis}}" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": ". Encoder weights " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " and decoder weights " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " are jointly trained. After pretraining, the decoder is discarded and the encoder is finetuned on task-specific downstream datasets. During finetuning, the model acquires downstream specific knowledge " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": ", which is built on general task-agnostic knowledge " + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 358, + 289, + 479 + ], + "type": "text", + "content": " learnt in pretraining, as shown in Fig. 2 (a)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 487, + 257, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 257, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 257, + 499 + ], + "type": "text", + "content": "3.2. Challenges of Sparse Training for MAE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 505, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 289, + 613 + ], + "type": "text", + "content": "Despite the effectiveness of MAE for large-scale vision transformers, it encounters great difficulties when applied to smaller models (see Fig. 1). The inadequate model capacity prevents small models from learning meaning task-agnostic knowledge and consequently transfers to inferior performance in downstream tasks. To enable the potential of large models' capability for smaller ones to learn powerful representations under MAE framework, we can modify Eq. 1 to combine it with sparse training as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 631, + 287, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 631, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 59, + 631, + 287, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S p a r s e}} (\\mathbf {W}, \\mathbf {M}) = \\mathbb {E} \\left(\\left\\| g _ {\\theta} \\left(f \\left(x _ {\\text {v i s}}, \\mathbf {W} \\odot \\mathbf {M}\\right)\\right) - x _ {\\text {m a s k}} \\right\\| ^ {2}\\right), \\tag {2}", + "image_path": "8c5ab4d4976a66f5c2ce553d95cc8930b5c53af54fdc66d4a4a2b7c0b6251361.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "text", + "content": " represents the large full encoder with parameters " + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "text", + "content": ". The encoder binary mask " + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "text", + "content": " produces a sparse sub-network by magnitude pruning during pretraining, while the decoder is kept dense as it is not used for downstream finetuning. Following the notation in Sec. 3.1, we use " + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 46, + 653, + 289, + 714 + ], + "type": "text", + "content": " to represent" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "the general knowledge of the sparse sub-network learnt in pretraining, and we denote " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "D^{s}" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " the specific knowledge of the finetuned sparse sub-network for downstream tasks. Considering both the pretraining and finetuning phases, we outline three intuitive strategies to integrate sparse training into the MAE framework:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 155, + 547, + 344 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 155, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 155, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 155, + 547, + 239 + ], + "type": "text", + "content": "- 1 Dense pretraining then sparse finetuning as shown in Fig. 2 (b). Although this strategy allows the sparse finetuning phase to optimize from the large full model's task-agnostic knowledge " + }, + { + "bbox": [ + 317, + 155, + 547, + 239 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 317, + 155, + 547, + 239 + ], + "type": "text", + "content": ", the sub-network does not benefit from the pretraining task and lacks its knowledge " + }, + { + "bbox": [ + 317, + 155, + 547, + 239 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 317, + 155, + 547, + 239 + ], + "type": "text", + "content": ". This may hinder the model to acquire better downstream knowledge " + }, + { + "bbox": [ + 317, + 155, + 547, + 239 + ], + "type": "inline_equation", + "content": "D^{s}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 247, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 247, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 317, + 247, + 547, + 344 + ], + "type": "text", + "content": "- 2&3 Sparse pretraining using Eq. (2) then sparse finetuning (see Fig. 2 (c)). During finetuning, the sparsity mask can be kept as fixed or dynamic (dynamic mask is the third strategy). This strategy relies on the architecture and task-agnostic knowledge " + }, + { + "bbox": [ + 317, + 247, + 547, + 344 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 317, + 247, + 547, + 344 + ], + "type": "text", + "content": " obtained by the sub-network during the pretraining phase, while it ignores the potential benefits brought by incorporating the knowledge " + }, + { + "bbox": [ + 317, + 247, + 547, + 344 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 317, + 247, + 547, + 344 + ], + "type": "text", + "content": " of the large full model." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 354, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 449 + ], + "type": "text", + "content": "In summary, it is challenging to tailor a sparse training mechanism for MAE framework to sufficiently learn task-agnostic and task-specific knowledge for a small model. In the following section, we introduce our approach to solve this dilemma by fully utilizing the task-agnostic knowledge " + }, + { + "bbox": [ + 304, + 354, + 547, + 449 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 354, + 547, + 449 + ], + "type": "text", + "content": " of the large full model and adaptively optimizing sub-network's topology connections for various downstream tasks for better task-specific knowledge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 457, + 468, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 468, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 468, + 470 + ], + "type": "text", + "content": "3.3. Sparse Masked Autoencoders" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 476, + 545, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 512 + ], + "type": "text", + "content": "To tackle the limitations of the sparse training strategies introduced above, we propose a novel method for pruning ViT models under the MAE framework (see Fig. 2 (d))." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 513, + 547, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 547, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 547, + 620 + ], + "type": "text", + "content": "Firstly, to ensure a good general task-agnostic knowledge " + }, + { + "bbox": [ + 304, + 513, + 547, + 620 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 513, + 547, + 620 + ], + "type": "text", + "content": " to be learnt via pretraining, we utilize the dense full model to reconstruct the RGB pixels following Eq. 1, which learns more discriminative representations thanks to its larger capacity. Simultaneously, a sparse sub-network within the full dense network is trained using the same visible tokens as the inputs and regresses the same targets as the large full model. The sparseMAE loss function is therefore extended from Eq. 1 and Eq. 2 as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 355, + 630, + 545, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 630, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 355, + 630, + 545, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\underbrace {\\mathcal {L} _ {\\text {S p a r s e}} (\\widetilde {\\mathbf {W}})} _ {\\text {s u b - n e t w o r k}} + \\underbrace {\\alpha \\mathcal {L} _ {\\mathrm {M A E}} (\\mathbf {W})} _ {\\text {f u l l n e t w o r k}}, \\tag {3}", + "image_path": "bc69130026d1b14e0abbd7fc5c3fc3b76e03ee2db610fd665c38d02ba6791eb2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "text", + "content": " represents the dense parameters of the large full model, " + }, + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{W}} = \\mathbf{W} \\odot \\mathbf{M}" + }, + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "text", + "content": " is the masked sparse parameters of the sub-network, and " + }, + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 664, + 547, + 713 + ], + "type": "text", + "content": " is the scaling hyperparameter balancing the reconstruction losses for the dense full model and the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16179" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 509, + 309 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 509, + 309 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 509, + 309 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 509, + 309 + ], + "type": "image", + "image_path": "22ed134c47fa750236b62ab04a8494baffb2af9d50b552dbe5234ef88e90f38a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 329, + 547, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 329, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 547, + 378 + ], + "type": "text", + "content": "Figure 3: Illustration of Sparse Masked Autoencoders. In the forward pass, both the large full model (red solid lines) and the sparse sub-network (red dashed lines) reconstruct the masked patches from the same visible tokens. In the backward pass, the masked weights (green solid lines) and the sparse weights (green dashed lines) are updated separately using the gradients from the corresponding losses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "type": "text", + "content": "sparse sub-network. In our experiments, we set " + }, + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\alpha = 1.0" + }, + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "type": "text", + "content": "As shown in Eq. 3, the gradients for the sparse weights " + }, + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{W}}" + }, + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "type": "text", + "content": " shall be updated following both the reconstruction losses of the sub-network and the full network, which may lead to gradient imbalance between the masked and unmasked weights [38]. Therefore, in each training iteration we split the weights " + }, + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "type": "text", + "content": " of the full model into sparse sub-network weights " + }, + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{W}}" + }, + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "type": "text", + "content": " and unmasked weights " + }, + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{W}} = \\mathbf{W} \\odot (1 - \\mathbf{M})" + }, + { + "bbox": [ + 46, + 423, + 289, + 522 + ], + "type": "text", + "content": " and optimize them separately:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 533, + 289, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 533, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 83, + 533, + 289, + 548 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathbf {W}} _ {t + 1} = \\widetilde {\\mathbf {W}} _ {t} - \\eta (g (\\mathbf {W} _ {t} \\odot \\mathbf {M} _ {t}) \\odot \\mathbf {M} _ {t}), \\qquad (4)", + "image_path": "8f4e311095d368f12b3ec316025b984b705afa62a1485aaa9db034301da0a1d6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 567, + 288, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 567, + 288, + 582 + ], + "spans": [ + { + "bbox": [ + 84, + 567, + 288, + 582 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {W}}} _ {t + 1} = \\overline {{\\mathbf {W}}} _ {t} - \\eta (g (\\mathbf {W} _ {t}) \\odot (1 - \\mathbf {M} _ {t})), \\tag {5}", + "image_path": "8f209c9176c12e7416039827b5afda6f52b3b90cbe5974b13b7a0937023f8bc2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "inline_equation", + "content": "g(\\mathbf{W}_t \\odot \\mathbf{M}_t) = \\frac{\\partial \\mathcal{L}_{\\mathrm{Sparse}}(\\mathbf{W}_t \\odot \\mathbf{M}_t)}{\\partial \\mathbf{W}_t}" + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "text", + "content": " are gradients of the sparse sub-network, " + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "inline_equation", + "content": "g(\\mathbf{W}_t) = \\frac{\\partial \\mathcal{L}_{\\mathrm{MAE}}(\\mathbf{W}_t)}{\\partial \\mathbf{W}_t}" + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "text", + "content": " are the gradients of the full models, " + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "text", + "content": " is the learning rate and " + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "text", + "content": " indexes the training iteration. The binary mask " + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_t" + }, + { + "bbox": [ + 46, + 587, + 289, + 714 + ], + "type": "text", + "content": " is obtained by on-the-fly magnitude pruning. Here, we only show gradient descent for simplicity and omit other parts (e.g., weight decay). In this way, the large full model can utilize all the model parameters in the forward pass to produce discriminative outputs, while in the backward pass, the sparse sub-network stays unaffected by the gradients from the large full model. During" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 399, + 547, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 547, + 471 + ], + "type": "text", + "content": "the pretraining phase, the task-agnostic knowledge " + }, + { + "bbox": [ + 304, + 399, + 547, + 471 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 304, + 399, + 547, + 471 + ], + "type": "text", + "content": " of the sparse sub-network can dynamically select the important sparse connections in the context of the task-agnostic knowledge " + }, + { + "bbox": [ + 304, + 399, + 547, + 471 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 399, + 547, + 471 + ], + "type": "text", + "content": " of the full dense model, which helps to identify the important sparse topology and is retained for downstream finetuning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 472, + 548, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 548, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 548, + 639 + ], + "type": "text", + "content": "After pretraining, the full dense model with a sparse sub-network is transferred and finetuned on different downstream tasks (e.g., classification, detection, segmentation, etc.). For the finetuning stage, the deep representations from the sparse sub-network are used for the downstream tasks, while its topology connections are optimized end-to-end given the specific downstream objectives. We refrain from fine-tuning dense full models, as the fine-tuning process for both dense and sparse models is costly. Hence, the learning of the task-specific knowledge " + }, + { + "bbox": [ + 304, + 472, + 548, + 639 + ], + "type": "inline_equation", + "content": "D^{s}" + }, + { + "bbox": [ + 304, + 472, + 548, + 639 + ], + "type": "text", + "content": " for downstream tasks can benefit from inheriting a pretrained sparse sub-network with knowledge " + }, + { + "bbox": [ + 304, + 472, + 548, + 639 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 304, + 472, + 548, + 639 + ], + "type": "text", + "content": ", and adaptively selecting the more suitable sparse architecture built on the general knowledge " + }, + { + "bbox": [ + 304, + 472, + 548, + 639 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 472, + 548, + 639 + ], + "type": "text", + "content": " of the full model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 655, + 388, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 655, + 388, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 655, + 388, + 670 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "type": "text", + "content": "In this section, we validate the effectiveness and robustness of our SparseMAE. First, we introduce the implementation details of our SparseMAE in Sec. 4.1. Then, we present" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16180" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "our main experimental results in Sec. 4.2, which include classification results on ImageNet-1K [8] and semantic segmentation results on ADE20K [43]. To demonstrate the superiority of our proposed method, we compare SparseMAE with other state-of-the-art ViT pruning methods in Sec. 4.4. In Sec. 4.3, we evaluate the transferring ability of the models trained with our SparseMAE to object detection, instance segmentation tasks on MS COCO [24] and robustness evaluation. Finally, we perform ablative experiments to evaluate the performance of SparseMAE under different settings in Sec. 4.5." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 180, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 180, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 180, + 227 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 232, + 289, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 289, + 376 + ], + "type": "text", + "content": "Sparse Training Methods. We conduct experiments with two different granularities for sparse training, consisting of unstructured sparsity [14, 21] and hardware-friendly N:M sparsity [28, 42]. As we focus on studying how to learn and transfer task-agnostic and task-specific knowledge under the Masked Autoencoders paradigm with sparse training methods, we select two simple and effective sparse training methods, namely DFP [23] for unstructured sparsity and SR-STE [42] for N:M sparsity. We prune all the linear layers within the vision transformers and set the relevant hyperparameters for sparse training the same as [23, 42] by default." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 377, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 289, + 533 + ], + "type": "text", + "content": "Models. For all of our main experiments, unless otherwise stated, we select ViT-Base as our large full model and prune it for smaller ones. For hardware-friendly N:M sparsity, we use 2:32 and 4:32 sparse ratios to obtain models with similar scales to ViT-Tiny (5.8M) and ViT-Small (11.3M), the main experiments use N:M sparsity for default, which is applied to all linear layers, including self.qkv, self.proj in the Attention block, and self.fc1 and self.fc2 in the MLP block. For unstructured sparsity, we extend the above N:M ratios and prune ViT-Base with " + }, + { + "bbox": [ + 46, + 377, + 289, + 533 + ], + "type": "inline_equation", + "content": "93.5\\%" + }, + { + "bbox": [ + 46, + 377, + 289, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 377, + 289, + 533 + ], + "type": "inline_equation", + "content": "87.0\\%" + }, + { + "bbox": [ + 46, + 377, + 289, + 533 + ], + "type": "text", + "content": " sparsity. We name our ViT-Base with 2:32/93.5% sparsity as SparseMAE-Tiny (SparseMAE-T) and the one with 4:32/87.0% sparsity as SparseMAE-Small (SparseMAE-S)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 289, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 677 + ], + "type": "text", + "content": "Pretraining. Following MAE [15] pretraining scheme, models start from random initialization, we use ImageNet-1K (IN-1K) [8] for all pretraining experiments. We pretrain SparseMAE for 400 epochs using " + }, + { + "bbox": [ + 46, + 533, + 289, + 677 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 533, + 289, + 677 + ], + "type": "text", + "content": " images as inputs. The other settings also follow those used in [15]. For example, We use random resized cropping and horizontal flipping for data augmentation, a " + }, + { + "bbox": [ + 46, + 533, + 289, + 677 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 533, + 289, + 677 + ], + "type": "text", + "content": " mask ratio during pretraining, and a decoder with a single transformer layer and 512 hidden dimensions. Unlike other methods [34, 19] for small and tiny-scale vision transformers pretraining, we do not grid-search the decoder design, possible putting our method at a disadvantage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Finetuning. We evaluate our SparseMAE pretrained models for various downstream tasks, including ImageNet-1K [8] image classification, MS COCO [24] object detection and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 95 + ], + "type": "text", + "content": "instance segmentation tasks, ADE20K [43] semantic segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 96, + 548, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 548, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 548, + 193 + ], + "type": "text", + "content": "For image classification, we take a ViT-Base model pretrained using MAE for 400 epochs as the reference model, which is implemented using the officially released codebase [15] and achieves " + }, + { + "bbox": [ + 304, + 96, + 548, + 193 + ], + "type": "inline_equation", + "content": "83.2\\%" + }, + { + "bbox": [ + 304, + 96, + 548, + 193 + ], + "type": "text", + "content": " top-1 accuracy. We finetune SparseMAE-Tiny/Small for 200 epochs using the same receipt described in [15] for a fair comparison with other works [31, 19] and set the batch size as 1024 and the base learning rate as " + }, + { + "bbox": [ + 304, + 96, + 548, + 193 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 304, + 96, + 548, + 193 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 193, + 548, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 548, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 548, + 302 + ], + "type": "text", + "content": "For semantic segmentation, we follow BEiT [2] to use UperNet[37] as our segmentation framework. We train the model with an input resolution of " + }, + { + "bbox": [ + 304, + 193, + 548, + 302 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 193, + 548, + 302 + ], + "type": "text", + "content": " for 160k iterations. A batch size of 16 is adopted. For object detection and instance segmentation tasks, we follow ViTDet [22] framework. We train our SparseMAE models with a batch size of 64 for 100 epochs, and the input image resolution is set to " + }, + { + "bbox": [ + 304, + 193, + 548, + 302 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 193, + 548, + 302 + ], + "type": "text", + "content": ". Further details on the hyper-parameters can be found in Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 310, + 391, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 391, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 391, + 323 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "text", + "content": "Classification and Semantic Segmentation. Tab. 1 shows the evaluation of our SparseMAE model against self-supervised methods for ViT-Tiny/Small on ImageNet-1K image classification and ADE20K semantic segmentation. For ImageNet-1K classification, SparseMAE achieves " + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "inline_equation", + "content": "80.5\\%" + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "text", + "content": " top-1 accuracy. To our knowledge, this is currently the best result achieved by tiny-scale transformer models. Furthermore, SparseMAE outperforms TinyMIM [31]/G2SD [19]/MAE-lite [34] with " + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "inline_equation", + "content": "+4.7 / 4.2 / 4.0" + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "text", + "content": " accuracy gains on tiny-scale transformers. For semantic segmentation, SparseMAE achieves an mIoU of 45.2, outperforming G2SD/MAE-lite/TinyMIM by " + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "inline_equation", + "content": "+3.8 / 1.3 / 1.2" + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "text", + "content": " mIoUs. Similar improvements can be observed for the small-scale transformers. SparseMAE achieves a " + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "inline_equation", + "content": "83.2\\%" + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "text", + "content": " ImageNet-1K classification accuracy and a " + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "inline_equation", + "content": "48.4\\%" + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "text", + "content": " mIoU on ADE20K semantic segmentation, surpassing TinyMIM [31]/G2SD [19] by " + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "inline_equation", + "content": "+0.2 / 1.2" + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "inline_equation", + "content": "+0.0 / 2.2" + }, + { + "bbox": [ + 304, + 329, + 548, + 569 + ], + "type": "text", + "content": " on the two tasks respectively. The significant improvements over previous results show the superiority of our method in obtaining performant small models under MAE framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": "Speedups. Unstructured sparsity and N:M structured sparsity is efficient for commercial CPUs and FPGAs respectively. For our SparseMAE with unstructured sparsity, we tested the throughput (img/s) on an Intel RXeon(R) 6238R CPU using the deepsparse engine with a batch size of 64. For our SparseMAE with N:M sparsity, we tested the throughput on Xilinx XCVU13P with a batch size of 16. Our SparseMAE-S with N:M and unstructured sparsity achieved throughput gains of " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "+71\\%" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "+9\\%" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " over ViT-small with similar amount of parameters on FPGAs and CPUs. Therefore, our approach demonstrate high feasibility in actual scenarios." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16181" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 116, + 514, + 390 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": "Table 1: Finetuning results on ImageNet-1K and ADE20K. All our models are pre-trained only on ImageNet-1K. Extra pretrained teacher means the compared method cannot trained from scratch. " + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": " the model adopts the MAE pretrained model before performing the sparse pretraining." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 116, + 514, + 390 + ], + "lines": [ + { + "bbox": [ + 78, + 116, + 514, + 390 + ], + "spans": [ + { + "bbox": [ + 78, + 116, + 514, + 390 + ], + "type": "table", + "html": "
MethodParams (M)Pretraining epochsEncoder ratioExtra Pretrained/ TeacherClassification Top-1 Acc (%)Segmentation mIoU (%)
DeiT [33]5.8300-Label72.238.0
Tiny-scale
MAE [15]5.8160025%X71.637.6
MoCo [6]5.8300100%EMA73.339.3
TinyMIM [31]5.8300100%TinyMIM-S75.844.0
G2SD [19]5.8300100%ViT-B76.341.4
MAE-lite [34]5.840025%ViT-B76.543.9
SparseMAE (ours)5.840025%X80.545.2
DeiT [33]22300-Label79.9
Small-scale
MAE [15]22160025%X80.642.8
MoCo [6]22300100%EMA81.443.9
DINO [3]22300100%EMA81.545.3
CAE [5]22300100%DALL-E82.0-
TinyMIM [31]22300100%TinyMIM-ViT-B83.048.4
G2SD [19]2230025%ViT-B82.046.2
SparseMAE (ours)11.330025%ViT-B‡83.248.4
SparseMAE (ours)11.340025%X82.146.7
", + "image_path": "8688c59b47af17d59c84e564172d530120a1d77c71f16c5c8a953654c2483b04.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 409, + 258, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 409, + 258, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 409, + 258, + 422 + ], + "type": "text", + "content": "4.3. Evaluation on Other Downstream Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "type": "text", + "content": "Object Detection and Instance Segmentation. We finetune the SparseMAE pretrained models on MS COCO with ViT-Det [22]. As shown in Tab. 2, we report " + }, + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{bbox}}" + }, + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "type": "text", + "content": " for object detection and " + }, + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{mask}}" + }, + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "type": "text", + "content": " for instance segmentation. We compare SparseMAE-Tiny with MAE-Tiny and its improved variants using distillation methods, such as MAE-lite [34] and G2SD [19]. The results in Tab. 2 show that our Sparse-MAE obtains more than " + }, + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "type": "inline_equation", + "content": "3.1\\mathrm{AP^{bbox}}" + }, + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "type": "inline_equation", + "content": "2.4\\mathrm{AP^{mask}}" + }, + { + "bbox": [ + 46, + 429, + 288, + 548 + ], + "type": "text", + "content": " gains compared to MAE-lite and G2SD, validating the transferability of the learnt representations in our models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 550, + 288, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 288, + 660 + ], + "type": "text", + "content": "Robustness Evaluation. In Tab. 3, we evaluate the robustness of our models on different variants of ImageNet validation sets. We use the same models finetuned on original ImageNet (Tab. 1) and only run inference on the different validation sets, such asImagenet-R [17], ImageNet-A [17] and ImageNet-C [16]), without any specialized fine-tuning. Tab. 3 shows that our method has better generalization capability than distillation-based methods, such as G2SD [19] and TinyMIM [31])." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 288, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 288, + 685 + ], + "type": "text", + "content": "4.4. Comparison with State-of-the-art ViT Pruning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "To the best of our knowledge, our proposed SparseMAE is the first to study network pruning technique under the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 328, + 441, + 523, + 575 + ], + "blocks": [ + { + "bbox": [ + 304, + 408, + 545, + 432 + ], + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 432 + ], + "type": "text", + "content": "Table 2: Object detection and instance segmentation results on the MS COCO dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 441, + 523, + 575 + ], + "lines": [ + { + "bbox": [ + 328, + 441, + 523, + 575 + ], + "spans": [ + { + "bbox": [ + 328, + 441, + 523, + 575 + ], + "type": "table", + "html": "
MethodBackbone Size (M)Det APbboxSeg APmask
Tiny-scale models
DeiT [33]5.840.736.5
MAE [15]5.838.935.1
MoCo v3 [6]5.840.036.0
DINO [3]5.840.236.1
MAE-lite [34]5.842.738.2
G2SD [19]5.844.039.6
SparseMAE (ours)5.847.142.0
", + "image_path": "1eba3410e6487933c65f1487e26c6c6143cfa0827f669bbcee18321ec4cfc1f6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Masked Autoencoders paradigm. To validate the effectiveness of SparseMAE, we compare it against recent state-of-the-art fully-supervised pruning methods for ViTs. We choose SViT [4] and oViT [20] as our compared models, which exploit end-to-end sparse training for ViTs and achieve state-of-the-art results in the fully-supervised setting. SViT [4] adopts a dynamic sparse training method to prune the dense model using random initialization. However, this approach requires training the sparse model for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16182" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 140, + 286, + 290 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 288, + 131 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 288, + 131 + ], + "type": "text", + "content": "Table 3: Robustness evaluation on ImageNet variants (top-1 accuracy, except for ImageNet-C which evaluates mean corruption error). We test the small-scale and tiny-scale models from Tab. 1 on different ImageNet validation sets, without any fine-tuning." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 140, + 286, + 290 + ], + "lines": [ + { + "bbox": [ + 48, + 140, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 286, + 290 + ], + "type": "table", + "html": "
MethodModel SizeIN-A↑IN-R↑IN-C↓
DeiT [33]8.032.754.0
MAE [15]7.036.555.2
TinyMIM [31]Tiny-scale11.039.850.1
G2SD [19]12.939.0-
SparseMAE (ours)18.245.947.2
DeiT [33]18.342.341.4
MAE [15]20.145.640.6
TinyMIM [31]Small-scale27.548.835.8
G2SD [19]29.446.8-
SparseMAE (ours)29.349.235.2
", + "image_path": "a625b808a20b8bc87828cc770f3d19252c12cafeb1c4918a379a797e2740fdb4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 335, + 274, + 450 + ], + "blocks": [ + { + "bbox": [ + 46, + 303, + 287, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 327 + ], + "type": "text", + "content": "Table 4: Comparison with state-of-the-art supervised Vision Transformer pruning methods on ImageNet-1K." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 335, + 274, + 450 + ], + "lines": [ + { + "bbox": [ + 59, + 335, + 274, + 450 + ], + "spans": [ + { + "bbox": [ + 59, + 335, + 274, + 450 + ], + "type": "table", + "html": "
ModelMethodParams (M)Top-1 Acc (%)
ViT-BaseSViT [4]43.381.5
oViT [20]43.381.6
SViT [4]34.681.3
oViT [20]34.681.5
oViT [20]21.681.1
oViT [20]17.380.8
oViT [20]8.779.7
ViT-BaseSparseMAE-T (ours)5.880.5
SparseMAE-S (ours)11.383.2
", + "image_path": "69eb1cb78818b3b08d71a22b9ad00a807ab091ef6d216a5aae2caedcb391e605.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": "600 epochs. oViT [20] (post-training pruning) prunes the model initialized from a trained dense model (300 epochs), which requires 300 epochs of finetuning. For our SparseMAE, we only need 400 epochs for pretraining and 200 epochs for fine-tuning. The pretraining of SparseMAE is " + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": " faster than full-token supervised training. Consequently, our SparseMAE approach reduces the training time from 968 GPU hours to 581 GPU hours compared to SOTA SViT and oViT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Tab. 4 demonstrates that SparseMAE can outperform state-of-the-art fully-supervised pruning counterparts significantly for small and tiny-scale transformers. Our SparseMAE-T has a " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " performance gain and a " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "35.0\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " computation complexity reduction compared to oViT [20] for tiny-scale transformers. With our SparseMAE-S, we achieve a " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "83.2\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " ImageNet classification accuracy, outperforming supervised pruning techniques by large margins. These results suggest that our method successfully extends the superiority of MAE for representation learning with large-scale vision transformers to small and tiny-scale models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 70, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 156 + ], + "type": "text", + "content": "Table 5: Different sparse pretraining or finetuning strategies with tiny-scale models on ImageNet-1K and ADE20K. \"S\" means sparse pretraining, \"D\" means dense pretraining, and \"D+S\" means sparse pretraining together with a dense full model pretraining. " + }, + { + "bbox": [ + 304, + 70, + 547, + 156 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 304, + 70, + 547, + 156 + ], + "type": "text", + "content": ": the model is finetuned with the fixed sparse masks from pretraining. Corr. Strategy.: corresponding sparse training strategies explained in Sec. 4.5.1." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 163, + 533, + 313 + ], + "blocks": [ + { + "bbox": [ + 317, + 163, + 533, + 313 + ], + "lines": [ + { + "bbox": [ + 317, + 163, + 533, + 313 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 533, + 313 + ], + "type": "table", + "html": "
ModelPretrain MethodTop-1 Acc (%)Seg mIoU (%)Corr. Strtg.
ViT-BaseD83.248.1-
Unstructured sparsity (93.5% sparse ratio)
SparseMAE-TD77.840.2#1
SparseMAE-TS‡76.737.7#2
SparseMAE-TS77.639.8#3
SparseMAE-TD+S80.445.4-
2:32 sparsity
SparseMAE-TD78.440.7#1
SparseMAE-TS‡76.337.2#2
SparseMAE-TS78.741.3#3
SparseMAE-TD+S80.545.2-
", + "image_path": "b4d5f52c47841f642a8b272cde1c5c3075ab0a6ff19832a908461fc6ed3d5ad9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 336, + 479, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 336, + 479, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 336, + 479, + 348 + ], + "type": "text", + "content": "4.5. Ablation Studies and Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 357, + 520, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 357, + 520, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 520, + 369 + ], + "type": "text", + "content": "4.5.1 Comparison of Sparse Training Strategies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": "We compare three sparse training strategies for MAE framework introduced in Sec. 3.2 under two sparse granularities. Namely, we (1) pretrain a large full model using MAE and finetune it with sparse training, (2) combine sparse training with MAE pretraining and subsequently finetune the model on downstream tasks with the fixed sparse mask from pretraining, or (3) modify (2) to use dynamic sparse mask for finetuning. The results are shown in Tab. 5. Under unstructured sparsity, comparing (1) and (2), we see that the classification performance drops from " + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "inline_equation", + "content": "77.8\\%" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "inline_equation", + "content": "76.7\\%" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": ", showing that the knowledge " + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": " learnt by the sparse sub-network during pretraining cannot generalize well to the downstream tasks compared to the task-agnostic knowledge " + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": " learnt by the large full model. (3) further allows the adaptation of sparse sub-network architecture during finetuning and partially alleviates the problem by better acquiring downstream specific knowledge " + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "inline_equation", + "content": "D^{s}" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": ". However, its result still lags significantly behind our proposed method (" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "inline_equation", + "content": "77.6\\%" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "inline_equation", + "content": "80.4\\%" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": "), where the difference is only the absence of a concurrently pretrained large full model. On the other hand, when comparing (1) with our method, the addition of a pretrained sparse sub-network can boost the performance by " + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": ". Similar observations can also be made under N:M sparsity settings and semantic segmentation results. These empirical results suggest that both the task-agnostic knowledge " + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": " of the large full model and " + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": " of the sparse sub-network are necessary for small and tiny-scale vision transformers to effectively benefit from MAE pretraining." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16183" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 103, + 279, + 166 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 95 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 95 + ], + "type": "text", + "content": "Table 6: Comparison between different full models for SparseMAE on ImageNet-1K and ADE20K." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 103, + 279, + 166 + ], + "lines": [ + { + "bbox": [ + 56, + 103, + 279, + 166 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 279, + 166 + ], + "type": "table", + "html": "
ModelPatternParams (M)Top-1 Acc (%)Seg mIoU (%)
ViT-Small8:325.877.441.5
ViT-Base2:325.880.545.2
", + "image_path": "c67e294ceb39929000ef00817ca07e6363b06a37367e134269cb58b3f73a10c6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 186, + 219, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 219, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 219, + 198 + ], + "type": "text", + "content": "4.5.2 Ablation of Full Model Capacity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 205, + 289, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 314 + ], + "type": "text", + "content": "We conduct an ablation study to investigate the performance with different capacities of the full model. We use our Sparse-MAE and set ViT-Small and ViT-Base as the full models, where ViT-Small only has " + }, + { + "bbox": [ + 46, + 205, + 289, + 314 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 205, + 289, + 314 + ], + "type": "text", + "content": " parameters of ViT-Base. As shown in Tab. 6, the results show that our SparseMAE in ViT-Base achieves " + }, + { + "bbox": [ + 46, + 205, + 289, + 314 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 46, + 205, + 289, + 314 + ], + "type": "text", + "content": " top-1 accuracy advantage over ViT-Small, indicating that larger capacity of the full model with SparseMAE can significantly improve its performance by incorporating better pretraining task-agnostic knowledge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 329, + 268, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 268, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 268, + 342 + ], + "type": "text", + "content": "4.5.3 Impact of Different Pruning Sparsity Ratios" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 348, + 288, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 288, + 397 + ], + "type": "text", + "content": "The main experiments of sparse ratio setting is compared to the ViT-tiny and ViT-small. We provide additional sparse ratios to verify the performance of SparseMAE on ImageNet-1K as in Tab. 7." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 82, + 438, + 251, + 476 + ], + "blocks": [ + { + "bbox": [ + 46, + 404, + 288, + 430 + ], + "lines": [ + { + "bbox": [ + 46, + 404, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 288, + 430 + ], + "type": "text", + "content": "Table 7: Comparison between different sparse ratios for SparseMAE on ImageNet-1K." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 438, + 251, + 476 + ], + "lines": [ + { + "bbox": [ + 82, + 438, + 251, + 476 + ], + "spans": [ + { + "bbox": [ + 82, + 438, + 251, + 476 + ], + "type": "table", + "html": "
Sparse Pattern1:322:324:328:32
Top-1 Acc76.880.583.283.7
", + "image_path": "3fd68e59e3e9a8f9ce70c1a26dec4097d9ef46a6e1a25978b10facf17d9e2ff4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 501, + 119, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 119, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 119, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "In this paper, we introduce SparseMAE, an innovative method for incorporating sparse training into the Masked Autoencoders (MAE) framework to create efficient and powerful small and tiny-scale transformers. As the first study in this area, we first point out the limitations of sparse training under MAE and design a unified framework to transfer the task-agnostic and task-specific knowledge of large models to lightweight sparse sub-networks. SparseMAE trains the sparse sub-networks by persevering the task-agnostic knowledge of large full models and then adaptively finds the optimal specific sub-networks for downstream tasks. The proposed method outperforms the state-of-the-art small-scale vision transforms in both unsupervised pretraining and fully-supervised settings. We hope this work could shed light on sparse training with the Masked Autoencoders framework in the vision community." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 71, + 404, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 404, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 404, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 186, + 365, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 186, + 365, + 199 + ], + "spans": [ + { + "bbox": [ + 306, + 186, + 365, + 199 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 206, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 310, + 206, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 206, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 310, + 206, + 547, + 262 + ], + "type": "text", + "content": "[1] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 263, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 263, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 311, + 263, + 547, + 296 + ], + "type": "text", + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 297, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 297, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 311, + 297, + 547, + 352 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 353, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 353, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 311, + 353, + 547, + 397 + ], + "type": "text", + "content": "[4] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34:19974-19988, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 399, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 399, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 311, + 399, + 547, + 453 + ], + "type": "text", + "content": "[5] Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 454, + 547, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 454, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 311, + 454, + 547, + 499 + ], + "type": "text", + "content": "[6] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 500, + 547, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 500, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 311, + 500, + 547, + 542 + ], + "type": "text", + "content": "[7] Tri Dao, Daniel Y Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. arXiv preprint arXiv:2205.14135, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 544, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 544, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 311, + 544, + 547, + 589 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 590, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 590, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 311, + 590, + 547, + 634 + ], + "type": "text", + "content": "[9] Utku Evci, Trevor Gale, Jacob Menick, Pablo Samuel Castro, and Erich Olsen. Rigging the lottery: Making all tickets winners. In International Conference on Machine Learning, pages 2943-2952. PMLR, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 635, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 679 + ], + "type": "text", + "content": "[10] Chao Fang, Aojun Zhou, and Zhongfeng Wang. An algorithm–hardware co-optimized framework for accelerating n: M sparse transformers. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 30(11):1573–1586, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "text", + "content": "[11] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635, 2018." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16184" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[12] Peng Gao, Renrui Zhang, Rongyao Fang, Ziyi Lin, Hongyang Li, Hongsheng Li, and Qiao Yu. Mimic before reconstruct: Enhancing masked autoencoders with feature mimicking. IJCV 2023, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "text", + "content": "[14] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 251 + ], + "type": "text", + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "type": "text", + "content": "[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 342 + ], + "type": "text", + "content": "[17] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 288, + 397 + ], + "type": "text", + "content": "[18] Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks. The Journal of Machine Learning Research, 22(1):10882-11005, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 432 + ], + "type": "text", + "content": "[19] Wei Huang, Zhiliang Peng, Li Dong, Furu Wei, Jianbin Jiao, and Qixiang Ye. Generic-to-specific distillation of masked autoencoders. arXiv preprint arXiv:2302.14771, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 288, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 466 + ], + "type": "text", + "content": "[20] Denis Kuznedelev, Eldar Kurtic, Elias Frantar, and Dan Alistarh. ovit: An accurate second-order pruning framework for vision transformers. arXiv preprint arXiv:2210.09223, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 499 + ], + "type": "text", + "content": "[21] Yann LeCun, John Denker, and Sara Solla. Optimal brain damage. Advances in neural information processing systems, 2, 1989." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "text", + "content": "[22] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 280-296. Springer, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 590 + ], + "type": "text", + "content": "[23] Tao Lin, Sebastian U Stich, Luis Barba, Daniil Dmitriev, and Martin Jaggi. Dynamic model pruning with feedback. arXiv preprint arXiv:2006.07253, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 657 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[25] Shiwei Liu, Lu Yin, Decebal Constantin Mocanu, and Mykola Pechenizkiy. Do we actually need dense overparameterization? in-time over-parameterization in sparse training. In International Conference on Machine Learning, pages 6989-7000. PMLR, 2021." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "text", + "content": "[26] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. arXiv preprint arXiv:1810.05270, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "type": "text", + "content": "[27] Zhi-Gang Liu, Paul N Whatmough, and Matthew Mattina. Sparse systolic tensor array for efficient cnn hardware acceleration. arXiv preprint arXiv:2009.02381, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 143, + 547, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 547, + 186 + ], + "type": "text", + "content": "[28] Asit Mishra, Jorge Albericio Latorre, Jeff Pool, Darko Stosic, Dusan Stosic, Ganesh Venkatesh, Chong Yu, and Paulius Miccikevicius. Accelerating sparse deep neural networks. arXiv preprint arXiv:2104.08378, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 189, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 189, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 189, + 547, + 243 + ], + "type": "text", + "content": "[29] Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H Nguyen, Madeleine Gibescu, and Antonio Liotta. Scalable training of artificial neural networks with adaptive sparse connectivity inspired by network science. Nature communications, 9(1):2383, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 245, + 547, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 547, + 289 + ], + "type": "text", + "content": "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 291, + 546, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 291, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 291, + 546, + 323 + ], + "type": "text", + "content": "[31] Sucheng Ren, Fangyun Wei, Zheng Zhang, and Han Hu. Tinymim: An empirical study of distilling mim pre-trained models. arXiv preprint arXiv:2301.01296, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 326, + 546, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 546, + 380 + ], + "type": "text", + "content": "[32] Wei Sun, Aojun Zhou, Sander Stuijk, Rob G. J. Wijnhoven, Andrew Nelson, Hongsheng Li, and Henk Corporaal. Dominosearch: Find layer-wise fine-grained n:m sparse schemes from dense neural networks. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 382, + 546, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 546, + 437 + ], + "type": "text", + "content": "[33] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 439, + 546, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 439, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 308, + 439, + 546, + 472 + ], + "type": "text", + "content": "[34] Shaoru Wang, Jin Gao, Zeming Li, Jian Sun, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. arXiv preprint arXiv:2205.14443, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 474, + 546, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 474, + 546, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 474, + 546, + 507 + ], + "type": "text", + "content": "[35] Chen Wei, Haoqi Fan, Saining Xie, Chao Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 509, + 546, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 509, + 546, + 541 + ], + "spans": [ + { + "bbox": [ + 308, + 509, + 546, + 541 + ], + "type": "text", + "content": "[36] Mitchell Wortsman, Ali Farhadi, and Mohammad Rastegari. Discovering neural wireings. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 544, + 546, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 546, + 586 + ], + "type": "text", + "content": "[37] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 589, + 546, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 546, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 546, + 621 + ], + "type": "text", + "content": "[38] Jiahui Yu, Linjie Yang, Ning Xu, Jianchao Yang, and Thomas Huang. Slimmable neural networks. arXiv preprint arXiv:1812.08928, 2018." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 624, + 546, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 546, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 546, + 667 + ], + "type": "text", + "content": "[39] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. NeurIPS 2022, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 546, + 712 + ], + "type": "text", + "content": "[40] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. CVPR 2023, 2023." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16185" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 206 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[41] Yuxin Zhang, Mingbao Lin, Zhihang Lin, Yiting Luo, Ke Li, Fei Chao, Yongjian Wu, and Rongrong Ji. Learning best combination for efficient n:m sparsity. arXiv preprint arXiv:2206.06662, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[42] Aojun Zhou, Yukun Ma, Junnan Zhu, Jianbo Liu, Zhijie Zhang, Kun Yuan, Wenxiu Sun, and Hongsheng Li. Learning n: M fine-grained structured sparse neural networks from scratch. arXiv preprint arXiv:2102.04010, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 206 + ], + "type": "text", + "content": "[43] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16186" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/0d56c181-4428-4dcb-b7bc-71667b8724d5_content_list.json b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/0d56c181-4428-4dcb-b7bc-71667b8724d5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ebd8d4a5c24ab0c4c5089bdb51ce59c5a921d4c5 --- /dev/null +++ b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/0d56c181-4428-4dcb-b7bc-71667b8724d5_content_list.json @@ -0,0 +1,1815 @@ +[ + { + "type": "text", + "text": "SparseNeRF: Distilling Depth Ranking for Few-shot Novel View Synthesis", + "text_level": 1, + "bbox": [ + 109, + 130, + 857, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guangcong Wang", + "bbox": [ + 166, + 181, + 310, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaoxi Chen", + "bbox": [ + 359, + 181, + 464, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Change Loy", + "bbox": [ + 514, + 181, + 658, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "S-Lab, Nanyang Technological University", + "bbox": [ + 318, + 199, + 651, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziwei Liu", + "bbox": [ + 707, + 181, + 802, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{guangcong.wang,zhaoxi00,ccloy,ziwei.liu}@ntu.edu.sg", + "bbox": [ + 256, + 219, + 712, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25c38c11645eed074b62c9cd251e49a21ca4b5feb2027f7bacbcf93d5ee02a47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 253, + 236, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c000f477d6f0062339011af44cea59e90cc1d9c297bd721e449fa2ee9a2af937.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 324, + 236, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3525d8a67f84551639343e0c7d35420e43791c3fc381f519533a4d755ac8ba24.jpg", + "image_caption": [ + "Novel view (GT)" + ], + "image_footnote": [], + "bbox": [ + 83, + 393, + 236, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ecff797f74986a8d24f98fbc82ca52102ea24122b0fcb1306b7f5be77c5a642.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 253, + 401, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/28f91706afef13fe42b5b4fff2f36c690bddd7e94960c5ed90829a4c159bc04f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 324, + 400, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00557e9e82b0069ff9eaa39ce225d64f476f634688ee1c927bb85482624c8a08.jpg", + "image_caption": [ + "RegNeRF" + ], + "image_footnote": [], + "bbox": [ + 245, + 393, + 400, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64a3f2fecffe8f95e59346a101a75fb57de9e06c4153c47363fce93c651ce5f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 253, + 563, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1841394780b111d42dcab0a5beb53c863612ac147868dd0101632247db5cde5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 324, + 563, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72e5378b92d1d159efc28a474aa139d38d2622e5dcf200a318e81d23c6005967.jpg", + "image_caption": [ + "MonoSDF*", + "Figure 1: Comparing the state-of-the-arts RegNeRF [37], MonoSDF* [65], DSNeRF [17], and our SparseNeRF with three views for training (* denotes a re-implementation for our task). RegNeRF regularizes geometry with sparsity and continuity constraints. MonoSDF, and DSNeRF use scale-invariant depth constraints supervised by coarse depth maps. Our SparseNeRF uses robust depth ranking of coarse depth maps. SparseNeRF can synthesize realistic novel views and coherent geometric depth (Please refer to the supplementary material for rendered videos)." + ], + "image_footnote": [], + "bbox": [ + 408, + 393, + 563, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2533c19f69afba9e908ae86e9e2ae45eb85342aaa2d47b82e371b615bdb4f19d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 253, + 725, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97fe940b37493e134d7c866bb59289c2d1795a559e98eb870ffd908bea07a231.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 324, + 725, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/897fa3094c9469d8eb57d64785200bc4051bd287cbc097e259f59f0239dbf5c9.jpg", + "image_caption": [ + "DSNeRF" + ], + "image_footnote": [], + "bbox": [ + 570, + 393, + 725, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac1b1ff45e41977f79c4b1e24c12922dc34e7d801aae9da99a16a20fde31f722.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 253, + 887, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/129fb981b22218603b81438cf257136b573938aaa5575d6a9a7f763d1feb7ff0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 324, + 887, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2dcadd5a88e5ff3eb09d5d4865a81e4f90813db6baf7442b71da0bd2b74d3a25.jpg", + "image_caption": [ + "SparseNeRF (ours)" + ], + "image_footnote": [], + "bbox": [ + 733, + 393, + 887, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 578, + 313, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Field (NeRF) significantly degrades when only a limited number of views are available. To complement the lack of 3D information, depth-based models, such as DSNeRF and MonoSDF, explicitly assume the availability of accurate depth maps of multiple views. They linearly scale the accurate depth maps as supervision to guide the predicted depth of few-shot NeRFs. However, accurate depth maps are difficult and expensive to capture due to wide-range depth distances in the wild. This work presents a new Sparse-view NeRF (SparseNeRF) framework that exploits depth priors from real-world inaccurate observations. The inaccurate depth observations are either from pre-trained depth models or coarse depth maps of consumer-level depth sensors. Since coarse depth maps are not strictly scaled to the ground-truth depth maps, we propose a simple yet effective constraint, a local depth ranking method, on NeRFs such that the expected depth ranking of the NeRF is consistent with that of the coarse depth maps in local patches. To preserve the spatial continuity of the esti", + "bbox": [ + 75, + 613, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mated depth of NeRF, we further propose a spatial continuity constraint to encourage the consistency of the expected depth continuity of NeRF with coarse depth maps. Surprisingly, with simple depth ranking constraints, SparseNeRF outperforms all state-of-the-art few-shot NeRF methods (including depth-based models) on standard LLFF and DTU datasets. Moreover, we collect a new dataset NVS-RGBD that contains real-world depth maps from Azure Kinect, ZED 2, and iPhone 13 Pro. Extensive experiments on NVS-RGBD dataset also validate the superiority and generalizability of SparseNeRF. Code and dataset are available at https://sparsenerf.github.io/.", + "bbox": [ + 496, + 579, + 893, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 797, + 632, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural radiance fields (NeRFs) [36, 32, 1, 2, 44] have made tremendous progress in generating photo-realistic novel views of scenes by optimizing implicit function representations given a set of 2D input views. However, in a wide range of real-world scenarios, collecting dense views of a", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9065", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scene is often expensive and time-consuming [47]. Therefore, it is necessary to develop few-shot NeRF methods that can be learned from sparse views without significant degradation in performance.", + "bbox": [ + 78, + 90, + 467, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning a NeRF from sparse views is a challenging problem due to under-constrained reconstruction conditions, especially in textureless areas. Directly applying NeRFs to few-shot scenarios suffers from dramatic degradation [37]. Recently, some methods have greatly improved the performance of few-shot NeRF, which can be categorized into three groups. 1) The first group [37, 27, 25] is based on geometry constraints (sparsity and continuity regularizations) and high-level semantics. RegNeRF [37] regularized the geometry and appearance of patches rendered from unobserved viewpoints, and annealed the ray sampling space. InfoNeRF [27] imposed an entropy constraint of the density in each ray and a spatial smoothness constraint into the estimated images. However, since a scene often contains multiple layouts (Figure 1), sparsity and continuity geometric constraints of a few views cannot guarantee the complete 3D geometric reconstruction. 2) The second group [64, 5] resorts to pre-training on similar scenes. For example, PixelNeRF [64] proposed to condition a NeRF on convolutional feature maps to learn high-level semantics from other scenes. 3) The third group [17, 65, 24] exploits depth maps and makes a linearity assumption of depth maps to supervise few-shot NeRFs. For example, DSNeRF [17] exploited sparse 3D points generated by COLMAP [42] or accurate depth maps [13] obtained by high-accuracy depth scanners and the Multi-View Stereo (MVS) algorithm. The depth maps are linearly scaled as supervision to guide the predicted depth of few-shot NeRFs. To use coarse depth maps, MonoSDF [65] uses a local patch-based scale-invariant depth constraint supervised by coarse depth maps instead of global depth maps. However, the scale-invariant depth constraint is strong for real-world coarse depth maps from pre-trained depth models or consumer-level depth sensors due to wide-range depth distances in the wild.", + "bbox": [ + 80, + 152, + 467, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Along the third group, we wish to explore more robust 3D priors from coarse depth maps to complement the under-constrained few-shot NeRF. To address this problem, we present SparseNeRF, a simple yet effective method that distills depth priors from pre-trained depth models [41] or inaccurate depth maps from consumer-level depth sensors (Figure 3), which can be easily obtained from real-world scenes. Deriving useful depth cues from such pre-trained models is non-trivial. In particular, although single-view depth estimation methods have achieved good visual performance, thanks to large-scale monocular depth datasets and large ViT models, they cannot yield accurate 3D depth information due to coarse depth annotations, dataset bias, and ill-posed 2D single-view images. The inaccurate depth information contradicts the density prediction of a NeRF", + "bbox": [ + 80, + 674, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "when reconstructing each pixel of a 3D scene based on volume rendering. Directly scaling the coarse depth maps to a NeRF [17, 65] leads to inconsistent geometry against the expected depth of the NeRF.", + "bbox": [ + 503, + 90, + 888, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of directly supervising a NeRF with coarse depth priors, we relax hard depth constraints [17, 65] and distill robust local depth ranking from the coarse depth maps to a NeRF such that the depth ranking of a NeRF is consistent with that of coarse depth. That is, we supervise a NeRF with relative depth instead of absolute depth [17, 65]. To guarantee the spatial continuity of geometry, we further propose a spatial continuity constraint on depth maps such that the NeRF model imitates the spatial continuity of coarse depth maps. The accurate sparse geometry constraints from a limited number of views, combined with relaxed constraints including depth ranking regularization and continuity regularization, finally achieve promising novel view synthesis (Figure 1). It is noteworthy that SparseNeRF does not increase the running time during inference as it only exploits depth priors from pre-trained depth models or consumer-level sensors during the training stage (Figure 2). In addition, SparseNeRF is a plug-and-play module that can be easily integrated into various NeRFs.", + "bbox": [ + 503, + 152, + 888, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are 1) SparseNeRF, a simple yet effective method that distills local depth ranking priors from pre-trained depth models. With the help of the local depth ranking constraint, SparseNeRF significantly improves the performance of few-shot novel view synthesis over the state-of-the-art models (including depth-based NeRF methods). To preserve the coherent geometry of a scene, we propose a spatial continuity distillation constraint that encourages the spatial continuity of NeRF to be similar to that of the pre-trained depth model. Both depth ranking prior and spatial continuity distillation are new in the literature on NeRF. 2) Apart from SparseNeRF, we also contribute a new dataset, NVS-RGBD, which contains coarse depth maps from Azure Kinect, ZED 2, and iPhone 13 Pro. 3) Extensive experiments on the LLFF, DTU, and NVS-RGBD datasets demonstrate that SparseNeRF achieves a new state-of-the-art performance in few-shot novel view synthesis.", + "bbox": [ + 503, + 438, + 888, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 724, + 638, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields. NeRF [36, 48, 8, 39] has made great success in synthesizing novel views of complex scenes due to good representation of neural networks [20, 52, 50, 53, 18, 28]. Block-NeRF [47], CityNeRF [58], and Mega-NeRF [49] scaled the standard NeRF up to city-scale or urban-scale scenes. NeRF—[56], GNeRF [33], and BARF [31] relaxed the requirements of NeRFs and synthesized novel views without perfect camera poses. Some works attempted to improve NeRFs by considering antialiasing [1], sparse 3D grids with spherical harmonics [19].", + "bbox": [ + 503, + 750, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9066", + "bbox": [ + 483, + 926, + 514, + 935 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some methods [40, 30, 57] extended NeRFs to dynamic scenes. These methods do not focus on generating novel views with a few views. In this paper, we study sparse-view NeRF to reduce dense capture requirements in real-world applications.", + "bbox": [ + 76, + 90, + 470, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot Novel View Synthesis. There are increasing studies on few-shot novel view synthesis. Basically, the existing few-shot NeRF methods can be categorized into three groups. First, some methods exploit continuity constraints on geometry or object semantics. For example, RegNeRF [37] imposed a continuity constraint on geometry and regularized the appearance of patches from unobserved viewpoints with a flow model. InfoNeRF [27] proposed a ray entropy minimization regularization to encourage the density to be as sparse as possible along a ray, and used a ray information gain reduction regularization to constrain the continuous depth of neighbor rays. Second, some methods attempt to pe-train a NeRF on other similar scenes and fine-tune the NeRF on the target scene. For example, PixelNeRF [64] conditioned a NeRF on image inputs in a fully convolutional manner, allowing the model to learn scene priors from other scenes and reduce the requirement of dense views. MVSNeRF [5] leveraged plane-swept cost volumes for geometry-aware scene reasoning and combined it with physically based volume rendering. Similar to PixelNeRF, MVSNeRF was first trained on other real scenes and was finetuned on target scenes to evaluate its effectiveness and generalizability. Third, depth-based models [17, 65] use available depth information to supervise the training of NeRFs. DSNeRF [17] exploited sparse 3D points generated by COLMAP [42] or accurate depth maps [13] obtained by high-accuracy depth scanners. Different from these depth-based models, SparseNeRF distills robust depth ranking from pre-trained depth models or coarse depth maps from consumer-level sensors. As concurrent works, NeRDi [16] and NeuralLift-360 [60] also use ranking-based methods. However, they mainly focus on the single-view setting while SparseNeRF focuses on sparse-view. SparseNeRF uses ranking loss to avoid inconsistent 3D geometry across different views while [16, 60] focuses on a soft geometric regularization on a single view (no cross-view consistency problem during training). Moreover, SparseNeRF introduces a new spatial continuity loss to distill spatial coherence from monocular depth estimators. In addition, some single-view synthesis methods [59] either allow new generative objects from unseen views or focus on specific objects, e.g., face [4] and human [23]. Some NeRF-based methods [22, 11] extended to generation instead of reconstruction. Other AIGC models [21, 51, 54, 67, 10, 9, 63, 29, 62] focus on the 2D generation or category-specific reconstruction [7], which are different.", + "bbox": [ + 76, + 205, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 500, + 89, + 648, + 108 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present SparseNeRF to synthesize novel views given sparse view inputs. Single-view depth estimation is a long-standing computer vision task, aiming to predict a depth map given a single image. In this paper, we are interested in mining the depth priors encapsulated in pre-trained models of single-view depth estimation or coarse depth maps captured by consumer-level depth sensors. However, due to coarse annotations of single-view depth maps, (e.g., user clicks [6], RGB-D [43, 46], and laser/stereo [34]), dataset bias, and imperfect depth estimation models, it is challenging to obtain accurate 3D depth estimation given 2D single-view images. As for consumer-level depth sensors, it still struggles to capture accurate depth maps (Figure 3). Driven by these observations, our goal is to make use of the coarse depth maps and distill useful depth priors to guide the learning of a NeRF.", + "bbox": [ + 496, + 114, + 893, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary and Problem Formulation", + "text_level": 1, + "bbox": [ + 498, + 364, + 834, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields. Let a NeRF [36] be a mapping function $f$ that maps a 3D spatial location $\\mathbf{x}$ and a viewing direction $\\mathbf{d}$ into a volume density $\\sigma$ and a color value $\\mathbf{c}$ . The $f$ is a neural network consisting of eight perceptron (MLP) layers parameterized by $\\theta$ , which is given by $f_{\\theta}: (\\gamma(\\mathbf{x}), \\gamma(\\mathbf{d})) \\mapsto (\\sigma, \\mathbf{c})$ , where $\\gamma$ is a positional encoding. For each expected pixel color $\\hat{C}(\\mathbf{r})$ , it is rendered by casting a ray $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ with near and far bounds $t_n$ and $t_f$ . We evenly partition $[t_n, t_f]$ into $N$ points $(t_1, t_2, \\dots, t_N)$ along a ray $\\mathbf{r}$ and compute expected pixel color $\\hat{C}(\\mathbf{r})$ by $\\hat{C}(\\mathbf{r}) = \\sum_{i=1}^{N} \\mathbf{c}_i^*$ . The weighted color $\\mathbf{c}_i^*$ of a 3D point is computed by $\\mathbf{c}_i^* = w_i\\mathbf{c}_i$ , where $w_i = T_i(1 - \\exp(-\\sigma_i\\delta_i))$ , $T_i = \\exp(-\\sum_{j=1}^{i-1}\\sigma_j\\delta_j)$ and $\\delta_i = t_i - t_{i-1}$ . Therefore, the NeRF reconstruction loss can be formulated as", + "bbox": [ + 496, + 388, + 893, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n e r f}} = \\sum_ {\\mathbf {r} \\in R} \\left| \\left| \\hat {C} (\\mathbf {r}) - C (\\mathbf {r}) \\right| \\right| ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 613, + 890, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{C}(\\mathbf{r})$ are rendered color blended by $N$ samples. $C(\\mathbf{r})$ is the ground-truth pixel color. We use coarse-to-fine sampling as discussed in the vanilla NeRF [36]. Here we omit the fine rendering for simplification.", + "bbox": [ + 496, + 657, + 893, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Vanilla NeRFs aim to learn a mapping function $f_{\\theta}$ with $K_{d}$ dense views by optimizing a color reconstruction loss in Eq. (1). In this work, we aim to study few-shot NeRF when only $K_{s}$ sparse views are available ( $K_{s} \\ll K_{d}$ ). Let $H$ and $W$ denote the height and width of an image, we have $HWK_{s}$ rays as color reconstruction constraints. Without considering view directions $\\mathbf{d}$ and image continuity assumptions (i.e., $\\mathbf{c}_{i}^{*}$ is independent of $\\mathbf{c}_{j}^{*}$ when $i \\neq j$ ), suppose we have a $N_{c} \\times N_{c} \\times N_{c}$ discrete cubic volume to be optimized, which contains $N_{c}^{3}$ weighted color variables $\\mathbf{c}_{i}^{*}$ . We have $N_{c}^{3}$ variables $\\mathbf{c}_{i}^{*}$ and $HWK_{s}$ constraints. Ideally, we are able to solve a discrete cube of edge", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9067", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9bb0f5b71768058838c50b5be5b3168e9fd35622762bea1603009483c95d270f.jpg", + "image_caption": [ + "Figure 2: Framework Overview. SparseNeRF consists of two streams, i.e., NeRF and depth prior distillation. As for NeRF, we use Mip-NeRF as the backbone. We use a NeRF reconstruction loss $\\mathcal{L}_{\\mathrm{nerf}}$ . As for depth prior distillation, we distill depth priors from a pre-trained depth model. Specifically, we propose a local depth ranking regularization and a spatial continuity regularization to distill robust depth priors from coarse depth maps." + ], + "image_footnote": [], + "bbox": [ + 83, + 71, + 890, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "length $N_{c} = \\sqrt[3]{HWK_{s}}$ . That is, we can sample $\\sqrt[3]{HWK_{s}}$ for each edge of a cubic volume. Take $H = W = 512$ and $K_{s} = 3$ as an example, we have $N_{c} \\approx 92$ , which is far from reconstructing a continuous radiance field. To address this under-constrained optimization problem of few-shot NeRF, an intuitive way is to introduce reasonable regularization terms to constrain few-shot NeRFs conditioned on input $\\mathbf{x}$ , $\\mathbf{d}$ , and $C(\\mathbf{r})$ . Considering these constraints, a general formulation is given by", + "bbox": [ + 75, + 508, + 472, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {n e r f}} + \\lambda \\mathcal {R} (\\mathbf {x}, \\mathbf {d}, C (\\mathbf {r})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 672, + 468, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ is a regularization term.", + "bbox": [ + 76, + 712, + 299, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark. RegNeRF [37] and InfoNeRF [27] tackle this problem by introducing regularization terms on continuous depth constraint from unobserved viewpoints, sparsity of density on rays, and patch-based semantic constraint on color appearance. Depth-based models, such as DSNeRF [17] and MonoSDF [65], directly regress depth by leveraging sparse 3D points or deriving ground-truth depth maps with accurate absolute depth maps. Different from them, we propose to use a robust relative depth regularization from coarse depth maps of pre-trained depth models or consumer-level depth sensors.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overview of SparseNeRF", + "text_level": 1, + "bbox": [ + 500, + 508, + 732, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pipeline of SparseNeRF is illustrated in Figure 2. SparseNeRF mainly consists of four components, i.e., a neural radiance field (NeRF), a color reconstruction module, a depth ranking distillation module, and a spatial continuity distillation module. Specifically, we use Mip-NeRF [1] as the backbone and apply an MSE loss for color reconstruction $\\mathcal{L}_{\\mathrm{nerf}}$ . As for depth prior distillation, we either use a pre-trained depth model to estimate depth maps or capture coarse depth maps with consumer-level depth sensors. We use a vision transformer (DPT) [41] that is trained on a large-scale mixed depth dataset (1.4M images with various depth annotations) that covers a wide range of scenes. Therefore, DPT is able to provide general depth priors. Since single-view depth estimation is coarse, we carefully design a local depth ranking regularization and a spatial continuity regularization, which distills robust depth priors from coarse maps to the NeRF.", + "bbox": [ + 496, + 532, + 893, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Local Depth Ranking Distillation", + "text_level": 1, + "bbox": [ + 500, + 801, + 790, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single-view depth estimation is a challenging computer vision task, which aims at predicting a depth map of a scene given a single image as input. Due to dataset bias, coarse depth annotations, and imperfect neural models, it is difficult to achieve accurate depth prediction. Depth maps cap", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9068", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da4c811d6718d8ca818667df2c426261f95e8b65ea0c5ce7bbfe687ea89da00a.jpg", + "image_caption": [ + "Figure 3: Two types of coarse depth maps. Top: predicted by a pre-trained depth model. Bottom: captured by Azure Kinect and ZED 2 depth sensors. Si-MSE is a scale-invariant depth error. Time jittering refers to the difference between two depth maps captured by a static camera within a small time interval." + ], + "image_footnote": [], + "bbox": [ + 80, + 83, + 467, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tured by consumer-level sensors are also inaccurate due to wide-range depth distances. Directly using coarse depth maps to supervise a NeRF leads to ambiguous rendered novel view synthesis. To avoid the error of coarse depth maps, we relax the depth constraint and exploit the robust depth ranking prior. Given a pair of pixels in a single image, the depth ranking regularization only considers which point is nearer or farther. However, when the scene is complex, even depth ranking is not accurate. As shown in Figure 4, it is easy for depth estimation models to compare the depth ranking of white and cyan points, but it is hard to estimate the depth ranking of white and red points. It implies that the depth ranking becomes unreliable as the spatial distance increases.", + "bbox": [ + 75, + 534, + 468, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivated by these observations, we propose a local depth ranking distillation method that distills depth ranking priors from coarse depth maps to a NeRF. On one hand, given a local patch $P$ of an RGB image $I$ with a pose $p$ . We compute the depth $d_{\\mathbf{r}}$ of the rays that trace from $P$ by $d_{\\mathbf{r}} = \\sum_{i=1}^{N} w_{i} t_{i}$ . On the other hand, we use pre-trained depth DPT [41] to estimate the depth of $I$ and crop a local patch $d_{\\mathrm{dpt}}$ with the same spatial location as $d_{\\mathbf{r}}$ . We perform the depth ranking distillation by transferring the depth ranking knowledge from $d_{\\mathrm{dpt}}$ to $d_{\\mathbf{r}}$ . Let $k1$ and $k2$ be the", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6ced21e759b7dfb9e657fc27487da1c300e891a5f38ee58dcfe3b39e49a6530c.jpg", + "image_caption": [ + "Figure 4: Motivation of local depth ranking. It is easy for depth estimation models to tell the white point is nearer than the cyan point, but it is hard to compare while and red points. Best viewed by zooming in." + ], + "image_footnote": [], + "bbox": [ + 501, + 82, + 890, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "indices of 2D pixel coordinate of $d_{\\mathrm{dpt}}$ and $d_{\\mathbf{r}}$ . The depth ranking regularization is given by", + "bbox": [ + 498, + 282, + 890, + 313 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\text {r a n k}} = \\sum_ {d _ {\\mathrm {d p t}} ^ {k 1} \\leq d _ {\\mathrm {d p t}} ^ {k 2}} \\max \\left(d _ {\\mathbf {r}} ^ {k 1} - d _ {\\mathbf {r}} ^ {k 2} + m, 0\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 321, + 890, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m$ is a small margin that allows limited depth ranking errors. In Eq. (3), we randomly sample two depth pixels of $d_{\\mathrm{dpt}}$ , as denoted as $d_{\\mathrm{dpt}}^{k1}$ and $d_{\\mathrm{dpt}}^{k2}$ , $d_{\\mathrm{dpt}}^{k1} \\leq d_{\\mathrm{dpt}}^{k2}$ . If the depth rankings of $d_{\\mathrm{dpt}}$ and $d_{\\mathbf{r}}$ are not consistent, i.e., $d_{\\mathbf{r}}^{k1} > d_{\\mathbf{r}}^{k2}$ and $d_{\\mathrm{dpt}}^{k1} \\leq d_{\\mathrm{dpt}}^{k2}$ , we punish the NeRF. We encourage the corresponding depth estimated by NeRF to satisfy the consistent depth ranking of the pre-trained model. We clip the large depth value and normalize the depth map for depth sensors. For depth models, we use relative inverse depth.", + "bbox": [ + 496, + 364, + 892, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Spatial Continuity Distillation", + "text_level": 1, + "bbox": [ + 500, + 512, + 767, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The local depth ranking constraint guarantees that the predicted depth map estimated by NeRF is consistent with the DPT's depth map. However, it does not constrain the spatial continuity of the depth map. We distill spatial continuity priors from the depth model DPT, which allows large displacement across several depth pixels. If neighbor depth pixels are continuous on the depth map of DPT, we constrain the corresponding depth pixels of NeRF to be continuous. The spatial continuity regularization is given by", + "bbox": [ + 496, + 535, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\text {c o n t i}} = \\sum_ {k 1} \\sum_ {d _ {\\mathrm {d p t}} ^ {k 2} \\in \\mathrm {K N N} \\left(d _ {\\mathrm {d p t}} ^ {k 1}\\right)} \\max \\left(\\left| d _ {\\mathbf {r}} ^ {k 1} - d _ {\\mathbf {r}} ^ {k 2} \\right| - m ^ {\\prime}, 0\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 679, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{KNN}(\\cdot)$ returns k-nearest neighbors measured by depth values within a small region, e.g., $6 \\times 6$ patch. $m'$ is a small margin that allows small depth differences between neighbor pixels.", + "bbox": [ + 496, + 729, + 890, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Full Objective Loss. The full objective loss of SparseNeRF is formulated by", + "bbox": [ + 498, + 790, + 890, + 820 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {n e r f}} + \\lambda \\mathcal {R} _ {\\text {r a n k}} + \\gamma \\mathcal {R} _ {\\text {c o n t i}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 830, + 890, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ and $\\gamma$ control the importance of the regularization terms. In practice, we set $\\lambda = 0.2$ , $\\gamma = 0.02$ , $m = m' = 1.0 \\times 10^{-4}$ .", + "bbox": [ + 498, + 854, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9069", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. NVS-RGBD Dataset", + "text_level": 1, + "bbox": [ + 76, + 89, + 272, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The popular benchmark LLFF [35] on NeRFs is featured with forward-facing and object-centric. Moreover, we focus on consumer-level depth sensors instead of expensive scanners [14]. We study stand-alone depth cameras instead of a complex calibrated structured multiple camera system [15], and collect from real-world scenes instead of synthesized 3D assets [45]. Few RGBD datasets satisfy all of the requirements. Therefore, we collect a new dataset NVS-RGBD that contains real-world depth maps captured by consumer-level depth sensors, including Azure Kinect, ZED 2, and iPhone 13 Pro. We collect 8 scenes for both Azure Kinect and ZED 2, and 4 scenes for iPhone 13 Pro. The depth maps of Azure Kinect often contain noises around object edges. The depth maps of ZED 2 look smooth but unstable and incorrect when observing time jittering (Figure 3). The poses are computed via COLMAP. We collect indoor scenes for ZED 2 and Kinect. For iPhone, we collect both indoor and outdoor scenes. All the scenes are object-centric. Depth maps obtained from sensors have different artifacts coming from the sensor noises. Please refer to the supplementary material for the details.", + "bbox": [ + 75, + 114, + 472, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 449, + 209, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on the LLFF [35], DTU [26], and NVS-RGBD datasets. LLFF contains 8 complex forward-facing scenes. Following RegNeRF [37], we use every 8-th image as the held-out test set, and evenly sample sparse views from the remaining images for training. Different from LLFF, DTU is an object-level dataset. Following PixNeRF [64, 37], we use the same 15 scenes in our experiments. On the DTU dataset, the background of these scenes is a white table or a black background, which has few textures. As mentioned in RegNeRF, we mask the background during inference to avoid background bias. We use the same evaluation protocol for a fair comparison. On the NVS-RGBD dataset, for each scene, we use 3 views for training and the rest views for inference. For LLFF and DTU, we use pre-trained depth maps to implement our method. For NVS-RGBD, we use coarse depth maps captured by depth sensors.", + "bbox": [ + 75, + 474, + 472, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt four evaluation metrics in the experiments, i.e., PSNR, SSIM [55], LPIPS [66], and depth error [17]. Depth error is a scale-invariant MSE derived by minimizing $||wd\\hat{d} + b - d||$ , where $\\hat{d}$ and $d$ denote coarse depth maps and pseudo-ground-truth depth maps predicted by dense-view NeRF. Note that we only use dense views to train a good NeRF to generate pseudo-ground-truth depth maps for evaluation.", + "bbox": [ + 75, + 733, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We implement SparseNeRF based on the official JAX [3]. We use the Adam optimizer for the learning of SparseNeRF. We use an exponen", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tial learning rate, which decays from $2 \\times 10^{-3}$ to $2 \\times 10^{-5}$ . The batch size is set to 4096. For each scene, we use one GPU-v100 with 32G memory for both training and inference. We also implement the proposed SparseNeRF on GPU-v100 with 16G and RTX 2080 Ti with 11G. We reduce the batch size to 2048 and 1024 for these two types of GPUs, respectively. We use the same backbone and sampled points along rays compared with prior arts.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train 90k iterations for each scene given three training views, which takes about 2 hours for each scene. For depth maps captured by Kinect and ZED 2, we mask uncertain regions (black regions in depth maps). More precisely, we sample near-far depth pairs in certain regions in local patches to optimize Eqs. 3 and 4 of the paper.", + "bbox": [ + 496, + 215, + 890, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparisons on LLFF", + "text_level": 1, + "bbox": [ + 498, + 325, + 712, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare SparseNeRF with state-of-the-art methods on the LLFF dataset, including SRF [12], PixelNeRF [64], MVSNeRF [5], Mip-NeRF [1], DietNeRF [25], RegNeRF [37], DSNeRF [17] and MonoSDF[65]. Among them, SRF, PixelNeRF, and MVSNeRF are pre-trained on other similar scenes to exploit high-level semantics. Mip-NeRF is the state-of-the-art NeRF designed for dense-view training. InfoNeRF and RegNeRF mainly constrain sparsity and continuity of geometry or natural appearance semantics. More precisely, the compared methods can be classified into four groups. In the first group, SRF, PixelNeRF, and MVSNeRF are pre-trained in the large-scale DTU datasets (88 scenes) and are directly tested on the LLFF dataset. In the second group, SRF, PixelNeRF, and MVSNeRF are pre-trained on DTU and fine-tuned on LLFF per scene. The third group includes Mip-NeRF, DietNeRF, and RegNeRF. They use geometry continuity and semantic constraints. The fourth group distills knowledge from depth maps or sparse points of COLMAP (DSNeRF). The compared results are shown in Table 1. We can see that the proposed SparseNeRF achieves the best performance in PSNR, SSIM, and LPIPS. Note that the first and second groups have to pre-train on lots of other scenes. SparseNeRF achieves better results than DSNeRF and MonoSDF because the depth ranking is more robust than the absolute scale-invariant constraint.", + "bbox": [ + 496, + 353, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We provide qualitative analysis on LLFF, as shown in Figure 5. It is observed that PixelNeRF and MVSNeRF tend to generate ambiguous pixels. The reason is that they integrate CNN features into NeRFs. CNN features can provide high-level semantics to infer under-constrained textures but suffer pixel-level inference. MipNeRF is designed for dense-view synthesis without considering the under-constrained problem, leading to degraded geometry artifacts. Compared with RegNeRF, SparseNeRF uses robust depth priors from pre-trained depth models, which can handle challenging scene geometry with better performance.", + "bbox": [ + 496, + 734, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9070", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75858878c15072142d88042dca52d770058855cb35cea1341aeae4c6fbac1427.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 71, + 890, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26a7c24ddc19e29e415488dd817b644d64faea2cdd3e15a4b217d933dd618e99.jpg", + "image_caption": [ + "Figure 5: Visual comparisons on the LLFF dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes.", + "Figure 6: Visual comparisons on the DTU dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes." + ], + "image_footnote": [], + "bbox": [ + 81, + 294, + 890, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Quantitative Comparison on LLFF with three views. There are four groups. The first group denotes models that are pre-trained on other scenes and tested on a target scene (P). The second group requires further finetuning on a target scene (P&FT). The third group uses geometry and semantic constraints (geo. & sem.). The fourth group distills depth knowledge (depth KD).", + "bbox": [ + 75, + 518, + 470, + 626 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7a58495f0dacf2a4a4afea6926cd7c02f815805d91f9af962d6126cf54baf99e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingPSNR ↑SSIM ↑LPIPS ↓
SRF [12]P12.340.2500.591
PixelNeRF [64]7.930.2720.682
MVSNeRF [5]17.250.5570.356
SRF ft [12]P&FT17.070.4360.529
PixelNeRF ft [64]16.170.4380.512
MVSNeRF ft [5]17.880.5840.327
Mip-NeRF [1]geo. & sem.14.620.3510.495
DietNeRF [25]14.940.3700.496
RegNeRF [37]19.080.5870.336
MonoSDF*[65]depth KD18.450.5650.388
DSNeRF[17]18.940.5820.362
SparseNeRF (Ours)19.860.6240.328
", + "bbox": [ + 91, + 631, + 457, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparisons on DTU", + "text_level": 1, + "bbox": [ + 76, + 811, + 282, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar to the LLFF dataset, we conduct four-group comparisons on the DTU dataset, as shown in Table 2. In this first group, given a target scene, SRF, PixelNeRF, and MVSNeRF are pre-trained on other DTU scenes and are", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Quantitative Comparison on DTU with three views. The first group is pre-trained on other scenes and tested on a target scene (P). The second group requires further finetuning on a target scene (P&FT). The third group uses geometry and semantic constraints (geo. & sem.). The fourth group distills depth knowledge (depth KD).", + "bbox": [ + 496, + 518, + 890, + 611 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca13b3b973b034d7837782e087dd12dad223952470b4cd7e4bd903b2b185fb54.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingPSNR ↑SSIM ↑LPIPS ↓
SRF [12]P15.320.6710.304
PixelNeRF [64]16.820.6950.270
MVSNeRF [5]18.630.7690.197
SRF ft [12]P&FT15.680.6980.281
PixelNeRF ft [64]18.950.7100.269
MVSNeRF ft [5]18.540.7690.197
Mip-NeRF [1]geo. & sema.8.680.5710.353
DietNeRF [25]11.850.6330.314
RegNeRF [37]18.890.7450.190
MonoSDF* [65]depth KD18.920.7480.237
DSNeRF [17]16.900.5700.450
SparseNeRF(Ours)19.550.7690.201
", + "bbox": [ + 506, + 616, + 888, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tested on the target scene. In the second group, these three methods are further fine-tuned on the target scene. In the third group, we compare Mip-NeRF, DietNeRF, and RegN-eRF. Note that for the pre-trained models, they split DTU as a training set (88 scenes) and a test set (15 scenes). Since", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9071", + "bbox": [ + 480, + 924, + 513, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a8cc286876affb436e5ae62b0bbbd183d09075e72c3231937f127196a7c006f.jpg", + "image_caption": [ + "DSNeRF" + ], + "image_footnote": [], + "bbox": [ + 83, + 104, + 238, + 174 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7fe4728de2a19677aaa89656c081ff72e73158bcc825576d83c99808c0a876b.jpg", + "image_caption": [ + "MonoSDF*" + ], + "image_footnote": [], + "bbox": [ + 246, + 104, + 401, + 174 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92f98aa96d487373d69a8c10c4d586ea005391f268ea9e395baaaf3d158153b7.jpg", + "image_caption": [ + "RegNeRF" + ], + "image_footnote": [], + "bbox": [ + 408, + 104, + 563, + 174 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10731e1d75e132f846ea79fcf79ea2d4fbf94ff214848cb2ed031add2b972680.jpg", + "image_caption": [ + "SparseNeRF (ours)" + ], + "image_footnote": [], + "bbox": [ + 570, + 104, + 725, + 174 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca91274958313cceb7c430db55c5c06db6cbb380cff48b9796002363f91a4366.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 732, + 104, + 887, + 174 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c20aff936b70546755ac1eeabfeda8cd8d180983e3335ea32bcf648c87b074b5.jpg", + "image_caption": [ + "Figure 7: Visual comparisons on the NVS-RGBD dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes." + ], + "image_footnote": [], + "bbox": [ + 83, + 176, + 238, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4d7b76290cfe97e0b20019380f936dc13697917fcb2d43cc6d7836c25427a2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 176, + 401, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/049b3dc06ceb3dac98067a55eb03628ae3315477afa1fb3348d197eafda66a1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 176, + 563, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d75e48e88540eaea2f8ef21b59a760ee1cf17f6e8b0765d5e184fa225f62146b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 176, + 725, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d46d68af4155b95e5f44bb524ec21ef1af188df22ad6ed7fae444e70e5dc4a62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 176, + 887, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8652c6100c9f4481ae8e329de3b0fa81411815f7627378abe711e2e98cac5648.jpg", + "image_caption": [ + "w/o conti. & ranking", + "Figure 8: Ablation study on visual effect. The three columns denote \"w/o continuity and ranking distillation\", \"w/o continuity distillation\", and \"w/ continuity and ranking distillation\" (SparseNeRF, ours), respectively." + ], + "image_footnote": [], + "bbox": [ + 81, + 314, + 176, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f5c61f6a57a90e41dc7ea4c5443bf71e6fd0bb32c0e06afc2d76fcd5cb0152a.jpg", + "image_caption": [ + "w/o ranking" + ], + "image_footnote": [], + "bbox": [ + 178, + 315, + 274, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/888a9648beb336f6626ac7c988d861dc084742a5075cb25eab4e83c486aa8a12.jpg", + "image_caption": [ + "w/o conti." + ], + "image_footnote": [], + "bbox": [ + 276, + 315, + 370, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bac4f1bc85c90cf5283877288d3530b14e8eaa71b23af78ed128d10dab29640d.jpg", + "image_caption": [ + "SparseNeRF(ours)" + ], + "image_footnote": [], + "bbox": [ + 372, + 315, + 467, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the training set and the test set contains similar scenes, a test scene shares a similar distribution (e.g., similar backgrounds) with training scenes. Pre-training on other scenes greatly benefits a target scene. Therefore, SRF, PixelNeRF, and MVSNeRF also greatly achieve promising results. Finally, we compare depth-based DSNeRF [17] and MonoSDF [65]. DSNeRF studies two kinds of depth information, i.e., sparse 3D points generated by COLMAP [42] and ground-truth depth maps obtained by high-accuracy depth scanners and the Multi-View Stereo (MVS) algorithm. The depth maps of NVS-RGBD are captured by consumer-level sensors and contain lots of noise, which is hard to be applied to DSNeRF. Therefore, we adopt sparse 3D points for DSNeRF. Compared with these methods, SparseNeRF still achieves the best performance without pre-training on other scenes. We provide a qualitative analysis, as shown in Figure 6. SparseNeRF achieves better visual results than previous state-of-the-art methods.", + "bbox": [ + 75, + 472, + 468, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Comparisons on NVS-RGBD", + "text_level": 1, + "bbox": [ + 76, + 755, + 336, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then implement three state-of-the-art methods on our new NVS-RGBD dataset, including RegNeRF [37], DSNeRF [17], and MonoSDF [65]. All of them are optimized per scene without pre-training on the other scenes. RegNeRF adopted continuous geometry assumptions from unobserved viewpoints. DSNeRF uses a global scale-invariant depth constraint to supervise NeRFs. MonoSDF mainly focuses on implicit surface reconstruction. It is as-", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3fb45cf9fa821f59b901f984a1e621fc5001ae7276e7417a531b2bd048f6dc50.jpg", + "table_caption": [ + "Table 3: Comparison with State of the Arts on NVS-RGBD. * denotes a new implementation from related tasks." + ], + "table_footnote": [], + "table_body": "
MetricsPSNR↑SSIM↑LPIPS↓depth err↓
KinectRegNeRF [37]25.780.8400.2425.9×10-3
DSNeRF [17]25.910.8380.2384.8×10-3
MonoSDF* [65]25.500.8380.2395.5×10-3
SparseNeRF(ours)26.280.8500.2323.9×10-3
ZED 2RegNeRF [37]24.980.7810.2713.5×10-3
DSNeRF [17]24.970.7890.2623.1×10-3
MonoSDF* [65]24.540.7840.2753.2×10-3
SparseNeRF(ours)26.220.8040.2622.4×10-3
", + "bbox": [ + 504, + 334, + 890, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sumed that local patches of depth maps from monocular cameras satisfy $d' = w d + b$ where $d$ is predicted by the pre-trained model and $d'$ is then computed by the NeRF. In Table 3, we observe that SparseNeRF achieves better performance compared with these prior arts. We provide a qualitative analysis in Figure 7, demonstrating the superiority of SparseNeRF over previous state-of-the-art methods.", + "bbox": [ + 496, + 474, + 890, + 580 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a0ffbc9deb5baf93bc26685bcff098b7343ed6a57d6e0c0de9afa301f1c12003.jpg", + "table_caption": [ + "Table 4: Comparison of Depth Ranking and Depth scaling." + ], + "table_footnote": [], + "table_body": "
LLFFPSNR↑SSIM↑LPIPS↓depth err↓
baseline19.080.5870.33610.1×10-3
w/ depth scaling18.450.5650.3888.9×10-3
w/ depth ranking19.860.6240.3286.3×10-3
", + "bbox": [ + 504, + 617, + 890, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Studies and Further Analyses", + "text_level": 1, + "bbox": [ + 500, + 695, + 833, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Depth Distillation. To validate the effectiveness of the depth distillation, we isolate the depth constraint and keep other modules unchanged. As shown in the first and third rows of Table 4, it is observed that without the local depth ranking distillation, SparseNeRF degrades on PSNR, SSIM, LPIPS, and depth error, which demonstrates the effectiveness of the depth distillation. We further study the depth maps from unobserved views. The ablation study of visual effect is shown in Figure 8. It shows that depth ranking greatly improves performance in geometry. Spatial continuity regularization further improves the detailed coherence of scenes.", + "bbox": [ + 496, + 719, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9072", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5f44f6e1ed381493abb141d7d542d5473cc06c37120f54322dc96229def513b4.jpg", + "table_caption": [ + "Table 5: Impact of different pre-trained depth models." + ], + "table_footnote": [], + "table_body": "
LLFFPSNR↑SSIM↑LPIPS↓depth err↓
Baseline19.080.5870.33610.1×10-3
MiDaS small19.350.5920.3336.9×10-3
DPT Hybrid19.860.6250.3266.3×10-3
DPT Large19.860.6240.3286.3×10-3
", + "bbox": [ + 83, + 109, + 467, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Further Analyses on Depth Ranking and Depth Scaling. To validate the robustness of local depth ranking, we compare depth ranking regularization and depth scaling regularization on LLFF. We use the idea from MonoSDF [65], which mainly focuses on implicit surface reconstruction. With the specific design for surface shapes, MonoSDF slightly degrades the performance of RGB view synthesis. We re-implement it as depth scaling. Since the single-view depth estimation is not accurate, the linear depth scaling constraint is too strong such that depth distillation misleads the learning of the NeRF. Figure 3 shows two cases of scale-invariant errors. Instead, depth ranking relaxes the constraint and only focuses on depth comparison, which is more robust than depth scaling. As shown in Table 4, depth ranking performs better than depth scaling on LLFF.", + "bbox": [ + 75, + 204, + 468, + 430 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Impact of Different Pre-trained Depth Models. We test our method on three depth estimation models, i.e., MiDaS small, DPT Hybrid, and DPT Large. As shown in Table 5, all of the depth models are better than the baseline. DPT Hybrid and DPT Large achieve comparable results, which are better than MiDaS small.", + "bbox": [ + 75, + 431, + 468, + 520 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "FreeNeRF [61] potentially complements SparseNeRF. FreeNeRF and SparseNeRF improve few-shot NeRF in different ways. FreeNeRF introduces a frequency regularization while SparseNeRF distills depth priors from pretrained general depth estimators. We implement SparseNeRF on FreeNeRF and the results significantly improve, as shown in Table 6.", + "bbox": [ + 75, + 521, + 467, + 626 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/467ae9f4dcde4446670317ccf1de668a34cfa26c12f35bc7dcadf005c4c1cf47.jpg", + "table_caption": [ + "Table 6: FreeNeRF potentially complement SparseNeRF." + ], + "table_footnote": [], + "table_body": "
LLFFPSNR↑SSIM↑LPIPS↓
FreeNeRF19.600.6140.302
FreeNeRF+our depth distillation20.260.6460.292
", + "bbox": [ + 81, + 666, + 468, + 714 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "User Study. To show that the proposed spatial continuity regularization improves the 3D consistency in 3D space, we conducted a user study. For each scene, we rendered videos for both with and without spatial continuity regularization terms. Users are asked to select the better-quality video according to 3D consistency/coherence and completeness (good geometry, few floaters). We evaluate 8 scenes on LLFF and there are 27 participants. The result shows $88.9\\%$ of users prefer \"w/ conti\" and $11.1\\%$ prefer \"w/o conti\". To evaluate the advantages of SparseNeRF on 3D consistency and shape completeness, we conduct another user study. Users are asked to evaluate \"baseline+RegNeRF's continu", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ity term” and “baseline+our continuity term”. The 27 users participated in the study and 8 scenes on LLFF are evaluated. The result shows $87.5\\%$ of users prefer “baseline+our continuity term” and $12.5\\%$ prefer “baseline+RegNeRF’s continuity term”. RegNeRF encourages the overall patch to be as smooth as possible without considering the sharp edges of objects, which conflicts with the reconstruction from sparse views. Our continuity regulation distills the neighbor relationship from depth models is more accurate. We show two examples of depth maps in Figure 9.", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e87f7feacc4d162237d85a5348fea56ffb91e41ab9541d1490394fa5f2a2f199.jpg", + "image_caption": [ + "Figure 9: Visual comparison on continuity regulation." + ], + "image_footnote": [], + "bbox": [ + 501, + 255, + 890, + 323 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Discussion. SparseNeRF could have a wide range of applications due to its promising results. We provide two applications on iPhone 13 Pro. Please refer to the project page for more experiments. Like UNISURF [38, 65], we also evaluate the SparseNeRF using Chamfer distance, showing the competitive results with MonoSDF and UNISURF, especially for complex scenes.", + "bbox": [ + 496, + 366, + 890, + 473 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 488, + 617, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a SparseNeRF framework that synthesizes novel views with sparse view inputs. To tackle the under-constrained few-shot NeRF problem, we propose a local depth ranking regularization that distills the depth ranking prior from coarse depth maps. To preserve the spatial continuity of geometry, we propose a spatial continuity regularization that distills the depth continuity priors. With the proposed depth priors, SparseNeRF achieves a new state-of-the-art performance on three datasets.", + "bbox": [ + 496, + 513, + 890, + 648 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitation. SparseNeRF significantly improves the performance of few-shot NeRF, but cannot be generalized to occluded views that are unobserved in the training views.", + "bbox": [ + 496, + 650, + 890, + 695 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Potential Negative Impact. SparseNeRF focuses on scene synthesis. It might be misused to create misleading content. Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is also supported by Singapore MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20221-0011) and NTU NAP Grant.", + "bbox": [ + 496, + 696, + 890, + 830 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 847, + 596, + 862 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srini", + "bbox": [ + 506, + 872, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9073", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vasan. Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In Int. Conf. Comput. Vis., 2021. 1, 2, 4, 6, 7", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1", + "[3] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: Composable transformations of Python+NumPy programs, 2018. 6", + "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3", + "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MvsNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In Int. Conf. Comput. Vis., 2021. 2, 3, 6, 7", + "[6] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Adv. Neural Inform. Process. Syst., 2016. 3", + "[7] Xipeng Chen, Guangrun Wang, Dizhong Zhu, Xiaodan Liang, Philip Torr, and Liang Lin. Structure-preserving 3D garment modeling with neural sewing machines. Adv. Neural Inform. Process. Syst., 2022. 3", + "[8] Zixuan Chen, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. CuNeRF: Cube-based neural radiance field for zero-shot medical image arbitrary-scale super resolution. Int. Conf. Comput. Vis., 2023. 2", + "[9] Zhaoxi Chen and Ziwei Liu. Relighting4D: Neural re- ligable human from videos. In Eur. Conf. Comput. Vis., 2022.3", + "[10] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Trans. Graph., 41(6):1-16, 2022. 3", + "[11] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Scenedreamer: Unbounded 3D scene generation from 2D image collections. In arXiv, 2023. 3", + "[12] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 6, 7", + "[13] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 2, 3", + "[14] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv:1602.02481, 2016. 6", + "[15] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6", + "[16] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "NerDI: Single-view NeRF synthesis with language-guided diffusion as general image priors. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3", + "[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 3, 4, 6, 7, 8", + "[18] Chenhe Dong, Guangrun Wang, Hang Xu, Jiefeng Peng, Xiaozhe Ren, and Xiaodan Liang. Efficientbert: Progressively searching multilayer perceptron via warm-up knowledge distillation. arXiv preprint arXiv:2109.07222, 2021. 2", + "[19] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., June 2022. 2", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst., 2020. 3", + "[22] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. EVA3D: Compositional 3D human generation from 2D image collections. arXiv preprint arXiv:2210.04888, 2022. 3", + "[23] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human NeRF from a single image. arXiv preprint arXiv:2303.12791, 2023. 3", + "[24] Shoukang Hu, Kaichen Zhou, Kaiyu Li, Longhui Yu, Lanqing Hong, Tianyang Hu, Zhenguo Li, Gim Hee Lee, and Ziwei Liu. ConsistentNeRF: Enhancing neural radiance fields with 3D consistency for sparse view synthesis. arXiv preprint arXiv:2305.11031, 2023. 2", + "[25] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a diet: Semantically consistent few-shot view synthesis. In Int. Conf. Comput. Vis., 2021. 2, 6, 7", + "[26] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In IEEE Conf. Comput. Vis. Pattern Recog., 2014. 6", + "[27] Mijeong Kim, Seonguk Seo, and Bohyung Han. InfoNeRF: Ray entropy minimization for few-shot neural volume rendering. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2, 3, 4", + "[28] Changlin Li, Bohan Zhuang, Guangrun Wang, Xiaodan Liang, Xiaojun Chang, and Yi Yang. Automated progressive learning for efficient training of vision transformers. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[29] Xiaoming Li, Wangmeng Zuo, and Chen Change Loy. Learning generative structure prior for blind text image super-resolution. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3", + "[30] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9074", + "bbox": [ + 482, + 925, + 514, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Int. Conf. Comput. Vis., 2021. 2", + "[32] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Adv. Neural Inform. Process. Syst., 2020. 1", + "[33] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. GNeRF: GAN-based neural radiance field without posed camera. In Int. Conf. Comput. Vis., 2021. 2", + "[34] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In IEEE Conf. Comput. Vis. Pattern Recog., 2015. 3", + "[35] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Trans. Graph., 38(4):1-14, 2019. 6", + "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis. Springer, 2020. 1, 2, 3", + "[37] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg-NeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 3, 4, 6, 7, 8", + "[38] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Int. Conf. Comput. Vis., 2021. 9", + "[39] Jiefeng Peng, Jiqi Zhang, Changlin Li, Guangrun Wang, Xiaodan Liang, and Liang Lin. Pi-nas: Improving neural architecture search by reducing supernet training consistency shift. In Int. Conf. Comput. Vis., 2021. 2", + "[40] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3", + "[41] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Int. Conf. Comput. Vis., 2021. 2, 4, 5", + "[42] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 3, 8", + "[43] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Eur. Conf. Comput. Vis. Springer, 2012. 3", + "[44] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Adv. Neural Inform. Process. Syst., 2021. 1", + "[45] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2012. 3", + "[47] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-NeRF: Scalable large scene neural view synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[48] Tang Tao, Longfei Gao, Guangrun Wang, Peng Chen, Dayang Hao, Xiaodan Liang, Mathieu Salzmann, and Kaicheng Yu. LiDAR-NeRF: Novel lidar view synthesis via neural radiance fields. arXiv preprint arXiv:2304.10406, 2023. 2", + "[49] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-NeRF: Scalable construction of large-scale NeRFs for virtual fly-throughs. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[50] Guangrun Wang, Ping Luo, Xinjiang Wang, Liang Lin, et al. Kalman normalization: Normalizing internal representations across network layers. Adv. Neural Inform. Process. Syst., 2018. 2", + "[51] Guangrun Wang and Philip HS Torr. Traditional classification neural networks are good generators: They are competitive with DDPMs and GANs. arXiv preprint arXiv:2211.14794, 2022.3", + "[52] Guangrun Wang, Guangcong Wang, Keze Wang, Xiaodan Liang, and Liang Lin. Grammatically recognizing images with tree convolution. 2020. 2", + "[53] Guangrun Wang, Keze Wang, and Liang Lin. Adaptively connected neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2", + "[54] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In Eur. Conf. Comput. Vis., 2022. 3", + "[55] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 6", + "[56] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2", + "[57] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3", + "[58] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. CityNeRF: Building nerf at city scale. arXiv preprint arXiv:2112.05504, 2021. 2", + "[59] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. SinNeRF: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928, 2022. 3", + "[60] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9075", + "bbox": [ + 482, + 925, + 514, + 935 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2D photo to a 3D object with 360deg views. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3", + "[61] Jiawei Yang, Marco Pavone, and Yue Wang. FreeNeRF: Improving few-shot neural rendering with free frequency regularization. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 9", + "[62] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Vtoonify: Controllable high-resolution portrait video style transfer. ACM Trans. Graph., 41(6):1-15, 2022. 3", + "[63] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Gp-unit: Generative prior for versatile unsupervised image-to-image translation. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 3", + "[64] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 3, 6, 7", + "[65] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 1, 2, 3, 4, 6, 7, 8, 9", + "[66] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6", + "[67] Long Zhuo, Guangcong Wang, Shikai Li, Wanye Wu, and Ziwei Liu. Fast-vid2vid: Spatial-temporal compression for video-to-video synthesis. In Eur. Conf. Comput. Vis., 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 486 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9076", + "bbox": [ + 482, + 924, + 516, + 936 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/0d56c181-4428-4dcb-b7bc-71667b8724d5_model.json b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/0d56c181-4428-4dcb-b7bc-71667b8724d5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..136f1f1cbb6fc51eb8934d3b346b7b2b84083896 --- /dev/null +++ b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/0d56c181-4428-4dcb-b7bc-71667b8724d5_model.json @@ -0,0 +1,2655 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.131, + 0.859, + 0.154 + ], + "angle": 0, + "content": "SparseNeRF: Distilling Depth Ranking for Few-shot Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.182, + 0.312, + 0.2 + ], + "angle": 0, + "content": "Guangcong Wang" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.182, + 0.465, + 0.198 + ], + "angle": 0, + "content": "Zhaoxi Chen" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.182, + 0.66, + 0.199 + ], + "angle": 0, + "content": "Chen Change Loy" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.2, + 0.652, + 0.217 + ], + "angle": 0, + "content": "S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.709, + 0.182, + 0.803, + 0.198 + ], + "angle": 0, + "content": "Ziwei Liu" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.22, + 0.714, + 0.234 + ], + "angle": 0, + "content": "{guangcong.wang,zhaoxi00,ccloy,ziwei.liu}@ntu.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.255, + 0.238, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.325, + 0.238, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.395, + 0.238, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.464, + 0.22, + 0.478 + ], + "angle": 0, + "content": "Novel view (GT)" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.255, + 0.402, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.325, + 0.401, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.394, + 0.401, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.464, + 0.359, + 0.478 + ], + "angle": 0, + "content": "RegNeRF" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.255, + 0.564, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.325, + 0.564, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.394, + 0.564, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.464, + 0.526, + 0.476 + ], + "angle": 0, + "content": "MonoSDF*" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.255, + 0.726, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.325, + 0.726, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.394, + 0.726, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.464, + 0.683, + 0.476 + ], + "angle": 0, + "content": "DSNeRF" + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.255, + 0.888, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.325, + 0.888, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.394, + 0.888, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.464, + 0.872, + 0.478 + ], + "angle": 0, + "content": "SparseNeRF (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.489, + 0.894, + 0.566 + ], + "angle": 0, + "content": "Figure 1: Comparing the state-of-the-arts RegNeRF [37], MonoSDF* [65], DSNeRF [17], and our SparseNeRF with three views for training (* denotes a re-implementation for our task). RegNeRF regularizes geometry with sparsity and continuity constraints. MonoSDF, and DSNeRF use scale-invariant depth constraints supervised by coarse depth maps. Our SparseNeRF uses robust depth ranking of coarse depth maps. SparseNeRF can synthesize realistic novel views and coherent geometric depth (Please refer to the supplementary material for rendered videos)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.579, + 0.314, + 0.594 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Neural Radiance Field (NeRF) significantly degrades when only a limited number of views are available. To complement the lack of 3D information, depth-based models, such as DSNeRF and MonoSDF, explicitly assume the availability of accurate depth maps of multiple views. They linearly scale the accurate depth maps as supervision to guide the predicted depth of few-shot NeRFs. However, accurate depth maps are difficult and expensive to capture due to wide-range depth distances in the wild. This work presents a new Sparse-view NeRF (SparseNeRF) framework that exploits depth priors from real-world inaccurate observations. The inaccurate depth observations are either from pre-trained depth models or coarse depth maps of consumer-level depth sensors. Since coarse depth maps are not strictly scaled to the ground-truth depth maps, we propose a simple yet effective constraint, a local depth ranking method, on NeRFs such that the expected depth ranking of the NeRF is consistent with that of the coarse depth maps in local patches. To preserve the spatial continuity of the esti" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.895, + 0.763 + ], + "angle": 0, + "content": "mated depth of NeRF, we further propose a spatial continuity constraint to encourage the consistency of the expected depth continuity of NeRF with coarse depth maps. Surprisingly, with simple depth ranking constraints, SparseNeRF outperforms all state-of-the-art few-shot NeRF methods (including depth-based models) on standard LLFF and DTU datasets. Moreover, we collect a new dataset NVS-RGBD that contains real-world depth maps from Azure Kinect, ZED 2, and iPhone 13 Pro. Extensive experiments on NVS-RGBD dataset also validate the superiority and generalizability of SparseNeRF. Code and dataset are available at https://sparsenerf.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.798, + 0.633, + 0.813 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Neural radiance fields (NeRFs) [36, 32, 1, 2, 44] have made tremendous progress in generating photo-realistic novel views of scenes by optimizing implicit function representations given a set of 2D input views. However, in a wide range of real-world scenarios, collecting dense views of a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "9065" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.151 + ], + "angle": 0, + "content": "scene is often expensive and time-consuming [47]. Therefore, it is necessary to develop few-shot NeRF methods that can be learned from sparse views without significant degradation in performance." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.154, + 0.468, + 0.67 + ], + "angle": 0, + "content": "Learning a NeRF from sparse views is a challenging problem due to under-constrained reconstruction conditions, especially in textureless areas. Directly applying NeRFs to few-shot scenarios suffers from dramatic degradation [37]. Recently, some methods have greatly improved the performance of few-shot NeRF, which can be categorized into three groups. 1) The first group [37, 27, 25] is based on geometry constraints (sparsity and continuity regularizations) and high-level semantics. RegNeRF [37] regularized the geometry and appearance of patches rendered from unobserved viewpoints, and annealed the ray sampling space. InfoNeRF [27] imposed an entropy constraint of the density in each ray and a spatial smoothness constraint into the estimated images. However, since a scene often contains multiple layouts (Figure 1), sparsity and continuity geometric constraints of a few views cannot guarantee the complete 3D geometric reconstruction. 2) The second group [64, 5] resorts to pre-training on similar scenes. For example, PixelNeRF [64] proposed to condition a NeRF on convolutional feature maps to learn high-level semantics from other scenes. 3) The third group [17, 65, 24] exploits depth maps and makes a linearity assumption of depth maps to supervise few-shot NeRFs. For example, DSNeRF [17] exploited sparse 3D points generated by COLMAP [42] or accurate depth maps [13] obtained by high-accuracy depth scanners and the Multi-View Stereo (MVS) algorithm. The depth maps are linearly scaled as supervision to guide the predicted depth of few-shot NeRFs. To use coarse depth maps, MonoSDF [65] uses a local patch-based scale-invariant depth constraint supervised by coarse depth maps instead of global depth maps. However, the scale-invariant depth constraint is strong for real-world coarse depth maps from pre-trained depth models or consumer-level depth sensors due to wide-range depth distances in the wild." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.675, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Along the third group, we wish to explore more robust 3D priors from coarse depth maps to complement the under-constrained few-shot NeRF. To address this problem, we present SparseNeRF, a simple yet effective method that distills depth priors from pre-trained depth models [41] or inaccurate depth maps from consumer-level depth sensors (Figure 3), which can be easily obtained from real-world scenes. Deriving useful depth cues from such pre-trained models is non-trivial. In particular, although single-view depth estimation methods have achieved good visual performance, thanks to large-scale monocular depth datasets and large ViT models, they cannot yield accurate 3D depth information due to coarse depth annotations, dataset bias, and ill-posed 2D single-view images. The inaccurate depth information contradicts the density prediction of a NeRF" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.151 + ], + "angle": 0, + "content": "when reconstructing each pixel of a 3D scene based on volume rendering. Directly scaling the coarse depth maps to a NeRF [17, 65] leads to inconsistent geometry against the expected depth of the NeRF." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.153, + 0.89, + 0.438 + ], + "angle": 0, + "content": "Instead of directly supervising a NeRF with coarse depth priors, we relax hard depth constraints [17, 65] and distill robust local depth ranking from the coarse depth maps to a NeRF such that the depth ranking of a NeRF is consistent with that of coarse depth. That is, we supervise a NeRF with relative depth instead of absolute depth [17, 65]. To guarantee the spatial continuity of geometry, we further propose a spatial continuity constraint on depth maps such that the NeRF model imitates the spatial continuity of coarse depth maps. The accurate sparse geometry constraints from a limited number of views, combined with relaxed constraints including depth ranking regularization and continuity regularization, finally achieve promising novel view synthesis (Figure 1). It is noteworthy that SparseNeRF does not increase the running time during inference as it only exploits depth priors from pre-trained depth models or consumer-level sensors during the training stage (Figure 2). In addition, SparseNeRF is a plug-and-play module that can be easily integrated into various NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.439, + 0.89, + 0.71 + ], + "angle": 0, + "content": "The main contributions of this paper are 1) SparseNeRF, a simple yet effective method that distills local depth ranking priors from pre-trained depth models. With the help of the local depth ranking constraint, SparseNeRF significantly improves the performance of few-shot novel view synthesis over the state-of-the-art models (including depth-based NeRF methods). To preserve the coherent geometry of a scene, we propose a spatial continuity distillation constraint that encourages the spatial continuity of NeRF to be similar to that of the pre-trained depth model. Both depth ranking prior and spatial continuity distillation are new in the literature on NeRF. 2) Apart from SparseNeRF, we also contribute a new dataset, NVS-RGBD, which contains coarse depth maps from Azure Kinect, ZED 2, and iPhone 13 Pro. 3) Extensive experiments on the LLFF, DTU, and NVS-RGBD datasets demonstrate that SparseNeRF achieves a new state-of-the-art performance in few-shot novel view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.725, + 0.64, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.75, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Neural Radiance Fields. NeRF [36, 48, 8, 39] has made great success in synthesizing novel views of complex scenes due to good representation of neural networks [20, 52, 50, 53, 18, 28]. Block-NeRF [47], CityNeRF [58], and Mega-NeRF [49] scaled the standard NeRF up to city-scale or urban-scale scenes. NeRF—[56], GNeRF [33], and BARF [31] relaxed the requirements of NeRFs and synthesized novel views without perfect camera poses. Some works attempted to improve NeRFs by considering antialiasing [1], sparse 3D grids with spherical harmonics [19]." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.927, + 0.515, + 0.936 + ], + "angle": 0, + "content": "9066" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "Some methods [40, 30, 57] extended NeRFs to dynamic scenes. These methods do not focus on generating novel views with a few views. In this paper, we study sparse-view NeRF to reduce dense capture requirements in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.206, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Few-shot Novel View Synthesis. There are increasing studies on few-shot novel view synthesis. Basically, the existing few-shot NeRF methods can be categorized into three groups. First, some methods exploit continuity constraints on geometry or object semantics. For example, RegNeRF [37] imposed a continuity constraint on geometry and regularized the appearance of patches from unobserved viewpoints with a flow model. InfoNeRF [27] proposed a ray entropy minimization regularization to encourage the density to be as sparse as possible along a ray, and used a ray information gain reduction regularization to constrain the continuous depth of neighbor rays. Second, some methods attempt to pe-train a NeRF on other similar scenes and fine-tune the NeRF on the target scene. For example, PixelNeRF [64] conditioned a NeRF on image inputs in a fully convolutional manner, allowing the model to learn scene priors from other scenes and reduce the requirement of dense views. MVSNeRF [5] leveraged plane-swept cost volumes for geometry-aware scene reasoning and combined it with physically based volume rendering. Similar to PixelNeRF, MVSNeRF was first trained on other real scenes and was finetuned on target scenes to evaluate its effectiveness and generalizability. Third, depth-based models [17, 65] use available depth information to supervise the training of NeRFs. DSNeRF [17] exploited sparse 3D points generated by COLMAP [42] or accurate depth maps [13] obtained by high-accuracy depth scanners. Different from these depth-based models, SparseNeRF distills robust depth ranking from pre-trained depth models or coarse depth maps from consumer-level sensors. As concurrent works, NeRDi [16] and NeuralLift-360 [60] also use ranking-based methods. However, they mainly focus on the single-view setting while SparseNeRF focuses on sparse-view. SparseNeRF uses ranking loss to avoid inconsistent 3D geometry across different views while [16, 60] focuses on a soft geometric regularization on a single view (no cross-view consistency problem during training). Moreover, SparseNeRF introduces a new spatial continuity loss to distill spatial coherence from monocular depth estimators. In addition, some single-view synthesis methods [59] either allow new generative objects from unseen views or focus on specific objects, e.g., face [4] and human [23]. Some NeRF-based methods [22, 11] extended to generation instead of reconstruction. Other AIGC models [21, 51, 54, 67, 10, 9, 63, 29, 62] focus on the 2D generation or category-specific reconstruction [7], which are different." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.65, + 0.109 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.357 + ], + "angle": 0, + "content": "We present SparseNeRF to synthesize novel views given sparse view inputs. Single-view depth estimation is a long-standing computer vision task, aiming to predict a depth map given a single image. In this paper, we are interested in mining the depth priors encapsulated in pre-trained models of single-view depth estimation or coarse depth maps captured by consumer-level depth sensors. However, due to coarse annotations of single-view depth maps, (e.g., user clicks [6], RGB-D [43, 46], and laser/stereo [34]), dataset bias, and imperfect depth estimation models, it is challenging to obtain accurate 3D depth estimation given 2D single-view images. As for consumer-level depth sensors, it still struggles to capture accurate depth maps (Figure 3). Driven by these observations, our goal is to make use of the coarse depth maps and distill useful depth priors to guide the learning of a NeRF." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.366, + 0.835, + 0.383 + ], + "angle": 0, + "content": "3.1. Preliminary and Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.895, + 0.606 + ], + "angle": 0, + "content": "Neural Radiance Fields. Let a NeRF [36] be a mapping function \\( f \\) that maps a 3D spatial location \\( \\mathbf{x} \\) and a viewing direction \\( \\mathbf{d} \\) into a volume density \\( \\sigma \\) and a color value \\( \\mathbf{c} \\). The \\( f \\) is a neural network consisting of eight perceptron (MLP) layers parameterized by \\( \\theta \\), which is given by \\( f_{\\theta}: (\\gamma(\\mathbf{x}), \\gamma(\\mathbf{d})) \\mapsto (\\sigma, \\mathbf{c}) \\), where \\( \\gamma \\) is a positional encoding. For each expected pixel color \\( \\hat{C}(\\mathbf{r}) \\), it is rendered by casting a ray \\( \\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d} \\) with near and far bounds \\( t_n \\) and \\( t_f \\). We evenly partition \\( [t_n, t_f] \\) into \\( N \\) points \\( (t_1, t_2, \\dots, t_N) \\) along a ray \\( \\mathbf{r} \\) and compute expected pixel color \\( \\hat{C}(\\mathbf{r}) \\) by \\( \\hat{C}(\\mathbf{r}) = \\sum_{i=1}^{N} \\mathbf{c}_i^* \\). The weighted color \\( \\mathbf{c}_i^* \\) of a 3D point is computed by \\( \\mathbf{c}_i^* = w_i\\mathbf{c}_i \\), where \\( w_i = T_i(1 - \\exp(-\\sigma_i\\delta_i)) \\), \\( T_i = \\exp(-\\sum_{j=1}^{i-1}\\sigma_j\\delta_j) \\) and \\( \\delta_i = t_i - t_{i-1} \\). Therefore, the NeRF reconstruction loss can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.614, + 0.892, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n e r f}} = \\sum_ {\\mathbf {r} \\in R} \\left| \\left| \\hat {C} (\\mathbf {r}) - C (\\mathbf {r}) \\right| \\right| ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.895, + 0.72 + ], + "angle": 0, + "content": "where \\(\\hat{C}(\\mathbf{r})\\) are rendered color blended by \\(N\\) samples. \\(C(\\mathbf{r})\\) is the ground-truth pixel color. We use coarse-to-fine sampling as discussed in the vanilla NeRF [36]. Here we omit the fine rendering for simplification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Problem Formulation. Vanilla NeRFs aim to learn a mapping function \\( f_{\\theta} \\) with \\( K_{d} \\) dense views by optimizing a color reconstruction loss in Eq. (1). In this work, we aim to study few-shot NeRF when only \\( K_{s} \\) sparse views are available (\\( K_{s} \\ll K_{d} \\)). Let \\( H \\) and \\( W \\) denote the height and width of an image, we have \\( HWK_{s} \\) rays as color reconstruction constraints. Without considering view directions \\( \\mathbf{d} \\) and image continuity assumptions (i.e., \\( \\mathbf{c}_{i}^{*} \\) is independent of \\( \\mathbf{c}_{j}^{*} \\) when \\( i \\neq j \\)), suppose we have a \\( N_{c} \\times N_{c} \\times N_{c} \\) discrete cubic volume to be optimized, which contains \\( N_{c}^{3} \\) weighted color variables \\( \\mathbf{c}_{i}^{*} \\). We have \\( N_{c}^{3} \\) variables \\( \\mathbf{c}_{i}^{*} \\) and \\( HWK_{s} \\) constraints. Ideally, we are able to solve a discrete cube of edge" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "9067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.073, + 0.891, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.436, + 0.895, + 0.498 + ], + "angle": 0, + "content": "Figure 2: Framework Overview. SparseNeRF consists of two streams, i.e., NeRF and depth prior distillation. As for NeRF, we use Mip-NeRF as the backbone. We use a NeRF reconstruction loss \\(\\mathcal{L}_{\\mathrm{nerf}}\\). As for depth prior distillation, we distill depth priors from a pre-trained depth model. Specifically, we propose a local depth ranking regularization and a spatial continuity regularization to distill robust depth priors from coarse depth maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.473, + 0.648 + ], + "angle": 0, + "content": "length \\( N_{c} = \\sqrt[3]{HWK_{s}} \\). That is, we can sample \\( \\sqrt[3]{HWK_{s}} \\) for each edge of a cubic volume. Take \\( H = W = 512 \\) and \\( K_{s} = 3 \\) as an example, we have \\( N_{c} \\approx 92 \\), which is far from reconstructing a continuous radiance field. To address this under-constrained optimization problem of few-shot NeRF, an intuitive way is to introduce reasonable regularization terms to constrain few-shot NeRFs conditioned on input \\( \\mathbf{x} \\), \\( \\mathbf{d} \\), and \\( C(\\mathbf{r}) \\). Considering these constraints, a general formulation is given by" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.673, + 0.47, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {n e r f}} + \\lambda \\mathcal {R} (\\mathbf {x}, \\mathbf {d}, C (\\mathbf {r})), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.713, + 0.3, + 0.729 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) is a regularization term." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Remark. RegNeRF [37] and InfoNeRF [27] tackle this problem by introducing regularization terms on continuous depth constraint from unobserved viewpoints, sparsity of density on rays, and patch-based semantic constraint on color appearance. Depth-based models, such as DSNeRF [17] and MonoSDF [65], directly regress depth by leveraging sparse 3D points or deriving ground-truth depth maps with accurate absolute depth maps. Different from them, we propose to use a robust relative depth regularization from coarse depth maps of pre-trained depth models or consumer-level depth sensors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.51, + 0.733, + 0.527 + ], + "angle": 0, + "content": "3.2. Overview of SparseNeRF" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.895, + 0.791 + ], + "angle": 0, + "content": "The pipeline of SparseNeRF is illustrated in Figure 2. SparseNeRF mainly consists of four components, i.e., a neural radiance field (NeRF), a color reconstruction module, a depth ranking distillation module, and a spatial continuity distillation module. Specifically, we use Mip-NeRF [1] as the backbone and apply an MSE loss for color reconstruction \\(\\mathcal{L}_{\\mathrm{nerf}}\\). As for depth prior distillation, we either use a pre-trained depth model to estimate depth maps or capture coarse depth maps with consumer-level depth sensors. We use a vision transformer (DPT) [41] that is trained on a large-scale mixed depth dataset (1.4M images with various depth annotations) that covers a wide range of scenes. Therefore, DPT is able to provide general depth priors. Since single-view depth estimation is coarse, we carefully design a local depth ranking regularization and a spatial continuity regularization, which distills robust depth priors from coarse maps to the NeRF." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.791, + 0.818 + ], + "angle": 0, + "content": "3.3. Local Depth Ranking Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Single-view depth estimation is a challenging computer vision task, which aims at predicting a depth map of a scene given a single image as input. Due to dataset bias, coarse depth annotations, and imperfect neural models, it is difficult to achieve accurate depth prediction. Depth maps cap" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "9068" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.084, + 0.468, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.418, + 0.472, + 0.507 + ], + "angle": 0, + "content": "Figure 3: Two types of coarse depth maps. Top: predicted by a pre-trained depth model. Bottom: captured by Azure Kinect and ZED 2 depth sensors. Si-MSE is a scale-invariant depth error. Time jittering refers to the difference between two depth maps captured by a static camera within a small time interval." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.744 + ], + "angle": 0, + "content": "tured by consumer-level sensors are also inaccurate due to wide-range depth distances. Directly using coarse depth maps to supervise a NeRF leads to ambiguous rendered novel view synthesis. To avoid the error of coarse depth maps, we relax the depth constraint and exploit the robust depth ranking prior. Given a pair of pixels in a single image, the depth ranking regularization only considers which point is nearer or farther. However, when the scene is complex, even depth ranking is not accurate. As shown in Figure 4, it is easy for depth estimation models to compare the depth ranking of white and cyan points, but it is hard to estimate the depth ranking of white and red points. It implies that the depth ranking becomes unreliable as the spatial distance increases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Motivated by these observations, we propose a local depth ranking distillation method that distills depth ranking priors from coarse depth maps to a NeRF. On one hand, given a local patch \\( P \\) of an RGB image \\( I \\) with a pose \\( p \\). We compute the depth \\( d_{\\mathbf{r}} \\) of the rays that trace from \\( P \\) by \\( d_{\\mathbf{r}} = \\sum_{i=1}^{N} w_{i} t_{i} \\). On the other hand, we use pre-trained depth DPT [41] to estimate the depth of \\( I \\) and crop a local patch \\( d_{\\mathrm{dpt}} \\) with the same spatial location as \\( d_{\\mathbf{r}} \\). We perform the depth ranking distillation by transferring the depth ranking knowledge from \\( d_{\\mathrm{dpt}} \\) to \\( d_{\\mathbf{r}} \\). Let \\( k1 \\) and \\( k2 \\) be the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.083, + 0.892, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.21, + 0.895, + 0.271 + ], + "angle": 0, + "content": "Figure 4: Motivation of local depth ranking. It is easy for depth estimation models to tell the white point is nearer than the cyan point, but it is hard to compare while and red points. Best viewed by zooming in." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.283, + 0.892, + 0.314 + ], + "angle": 0, + "content": "indices of 2D pixel coordinate of \\(d_{\\mathrm{dpt}}\\) and \\(d_{\\mathbf{r}}\\). The depth ranking regularization is given by" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.322, + 0.892, + 0.359 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\text {r a n k}} = \\sum_ {d _ {\\mathrm {d p t}} ^ {k 1} \\leq d _ {\\mathrm {d p t}} ^ {k 2}} \\max \\left(d _ {\\mathbf {r}} ^ {k 1} - d _ {\\mathbf {r}} ^ {k 2} + m, 0\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.893, + 0.505 + ], + "angle": 0, + "content": "where \\(m\\) is a small margin that allows limited depth ranking errors. In Eq. (3), we randomly sample two depth pixels of \\(d_{\\mathrm{dpt}}\\), as denoted as \\(d_{\\mathrm{dpt}}^{k1}\\) and \\(d_{\\mathrm{dpt}}^{k2}\\), \\(d_{\\mathrm{dpt}}^{k1} \\leq d_{\\mathrm{dpt}}^{k2}\\). If the depth rankings of \\(d_{\\mathrm{dpt}}\\) and \\(d_{\\mathbf{r}}\\) are not consistent, i.e., \\(d_{\\mathbf{r}}^{k1} > d_{\\mathbf{r}}^{k2}\\) and \\(d_{\\mathrm{dpt}}^{k1} \\leq d_{\\mathrm{dpt}}^{k2}\\), we punish the NeRF. We encourage the corresponding depth estimated by NeRF to satisfy the consistent depth ranking of the pre-trained model. We clip the large depth value and normalize the depth map for depth sensors. For depth models, we use relative inverse depth." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.768, + 0.529 + ], + "angle": 0, + "content": "3.4. Spatial Continuity Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.673 + ], + "angle": 0, + "content": "The local depth ranking constraint guarantees that the predicted depth map estimated by NeRF is consistent with the DPT's depth map. However, it does not constrain the spatial continuity of the depth map. We distill spatial continuity priors from the depth model DPT, which allows large displacement across several depth pixels. If neighbor depth pixels are continuous on the depth map of DPT, we constrain the corresponding depth pixels of NeRF to be continuous. The spatial continuity regularization is given by" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.68, + 0.892, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\text {c o n t i}} = \\sum_ {k 1} \\sum_ {d _ {\\mathrm {d p t}} ^ {k 2} \\in \\mathrm {K N N} \\left(d _ {\\mathrm {d p t}} ^ {k 1}\\right)} \\max \\left(\\left| d _ {\\mathbf {r}} ^ {k 1} - d _ {\\mathbf {r}} ^ {k 2} \\right| - m ^ {\\prime}, 0\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.79 + ], + "angle": 0, + "content": "where \\(\\mathrm{KNN}(\\cdot)\\) returns k-nearest neighbors measured by depth values within a small region, e.g., \\(6 \\times 6\\) patch. \\(m'\\) is a small margin that allows small depth differences between neighbor pixels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.791, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Full Objective Loss. The full objective loss of SparseNeRF is formulated by" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.831, + 0.892, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {n e r f}} + \\lambda \\mathcal {R} _ {\\text {r a n k}} + \\gamma \\mathcal {R} _ {\\text {c o n t i}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "where \\(\\lambda\\) and \\(\\gamma\\) control the importance of the regularization terms. In practice, we set \\(\\lambda = 0.2\\), \\(\\gamma = 0.02\\), \\(m = m' = 1.0 \\times 10^{-4}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "9069" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.273, + 0.106 + ], + "angle": 0, + "content": "4. NVS-RGBD Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.435 + ], + "angle": 0, + "content": "The popular benchmark LLFF [35] on NeRFs is featured with forward-facing and object-centric. Moreover, we focus on consumer-level depth sensors instead of expensive scanners [14]. We study stand-alone depth cameras instead of a complex calibrated structured multiple camera system [15], and collect from real-world scenes instead of synthesized 3D assets [45]. Few RGBD datasets satisfy all of the requirements. Therefore, we collect a new dataset NVS-RGBD that contains real-world depth maps captured by consumer-level depth sensors, including Azure Kinect, ZED 2, and iPhone 13 Pro. We collect 8 scenes for both Azure Kinect and ZED 2, and 4 scenes for iPhone 13 Pro. The depth maps of Azure Kinect often contain noises around object edges. The depth maps of ZED 2 look smooth but unstable and incorrect when observing time jittering (Figure 3). The poses are computed via COLMAP. We collect indoor scenes for ZED 2 and Kinect. For iPhone, we collect both indoor and outdoor scenes. All the scenes are object-centric. Depth maps obtained from sensors have different artifacts coming from the sensor noises. Please refer to the supplementary material for the details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.21, + 0.468 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.734 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on the LLFF [35], DTU [26], and NVS-RGBD datasets. LLFF contains 8 complex forward-facing scenes. Following RegNeRF [37], we use every 8-th image as the held-out test set, and evenly sample sparse views from the remaining images for training. Different from LLFF, DTU is an object-level dataset. Following PixNeRF [64, 37], we use the same 15 scenes in our experiments. On the DTU dataset, the background of these scenes is a white table or a black background, which has few textures. As mentioned in RegNeRF, we mask the background during inference to avoid background bias. We use the same evaluation protocol for a fair comparison. On the NVS-RGBD dataset, for each scene, we use 3 views for training and the rest views for inference. For LLFF and DTU, we use pre-trained depth maps to implement our method. For NVS-RGBD, we use coarse depth maps captured by depth sensors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt four evaluation metrics in the experiments, i.e., PSNR, SSIM [55], LPIPS [66], and depth error [17]. Depth error is a scale-invariant MSE derived by minimizing \\( ||wd\\hat{d} + b - d|| \\), where \\( \\hat{d} \\) and \\( d \\) denote coarse depth maps and pseudo-ground-truth depth maps predicted by dense-view NeRF. Note that we only use dense views to train a good NeRF to generate pseudo-ground-truth depth maps for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. We implement SparseNeRF based on the official JAX [3]. We use the Adam optimizer for the learning of SparseNeRF. We use an exponen" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.213 + ], + "angle": 0, + "content": "tial learning rate, which decays from \\(2 \\times 10^{-3}\\) to \\(2 \\times 10^{-5}\\). The batch size is set to 4096. For each scene, we use one GPU-v100 with 32G memory for both training and inference. We also implement the proposed SparseNeRF on GPU-v100 with 16G and RTX 2080 Ti with 11G. We reduce the batch size to 2048 and 1024 for these two types of GPUs, respectively. We use the same backbone and sampled points along rays compared with prior arts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.216, + 0.892, + 0.308 + ], + "angle": 0, + "content": "We train 90k iterations for each scene given three training views, which takes about 2 hours for each scene. For depth maps captured by Kinect and ZED 2, we mask uncertain regions (black regions in depth maps). More precisely, we sample near-far depth pairs in certain regions in local patches to optimize Eqs. 3 and 4 of the paper." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.327, + 0.713, + 0.344 + ], + "angle": 0, + "content": "5.1. Comparisons on LLFF" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.732 + ], + "angle": 0, + "content": "We compare SparseNeRF with state-of-the-art methods on the LLFF dataset, including SRF [12], PixelNeRF [64], MVSNeRF [5], Mip-NeRF [1], DietNeRF [25], RegNeRF [37], DSNeRF [17] and MonoSDF[65]. Among them, SRF, PixelNeRF, and MVSNeRF are pre-trained on other similar scenes to exploit high-level semantics. Mip-NeRF is the state-of-the-art NeRF designed for dense-view training. InfoNeRF and RegNeRF mainly constrain sparsity and continuity of geometry or natural appearance semantics. More precisely, the compared methods can be classified into four groups. In the first group, SRF, PixelNeRF, and MVSNeRF are pre-trained in the large-scale DTU datasets (88 scenes) and are directly tested on the LLFF dataset. In the second group, SRF, PixelNeRF, and MVSNeRF are pre-trained on DTU and fine-tuned on LLFF per scene. The third group includes Mip-NeRF, DietNeRF, and RegNeRF. They use geometry continuity and semantic constraints. The fourth group distills knowledge from depth maps or sparse points of COLMAP (DSNeRF). The compared results are shown in Table 1. We can see that the proposed SparseNeRF achieves the best performance in PSNR, SSIM, and LPIPS. Note that the first and second groups have to pre-train on lots of other scenes. SparseNeRF achieves better results than DSNeRF and MonoSDF because the depth ranking is more robust than the absolute scale-invariant constraint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We provide qualitative analysis on LLFF, as shown in Figure 5. It is observed that PixelNeRF and MVSNeRF tend to generate ambiguous pixels. The reason is that they integrate CNN features into NeRFs. CNN features can provide high-level semantics to infer under-constrained textures but suffer pixel-level inference. MipNeRF is designed for dense-view synthesis without considering the under-constrained problem, leading to degraded geometry artifacts. Compared with RegNeRF, SparseNeRF uses robust depth priors from pre-trained depth models, which can handle challenging scene geometry with better performance." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "9070" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.072, + 0.892, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.895, + 0.29 + ], + "angle": 0, + "content": "Figure 5: Visual comparisons on the LLFF dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.295, + 0.891, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.477, + 0.895, + 0.508 + ], + "angle": 0, + "content": "Figure 6: Visual comparisons on the DTU dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.471, + 0.627 + ], + "angle": 0, + "content": "Table 1: Quantitative Comparison on LLFF with three views. There are four groups. The first group denotes models that are pre-trained on other scenes and tested on a target scene (P). The second group requires further finetuning on a target scene (P&FT). The third group uses geometry and semantic constraints (geo. & sem.). The fourth group distills depth knowledge (depth KD)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.632, + 0.458, + 0.796 + ], + "angle": 0, + "content": "
SettingPSNR ↑SSIM ↑LPIPS ↓
SRF [12]P12.340.2500.591
PixelNeRF [64]7.930.2720.682
MVSNeRF [5]17.250.5570.356
SRF ft [12]P&FT17.070.4360.529
PixelNeRF ft [64]16.170.4380.512
MVSNeRF ft [5]17.880.5840.327
Mip-NeRF [1]geo. & sem.14.620.3510.495
DietNeRF [25]14.940.3700.496
RegNeRF [37]19.080.5870.336
MonoSDF*[65]depth KD18.450.5650.388
DSNeRF[17]18.940.5820.362
SparseNeRF (Ours)19.860.6240.328
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.813, + 0.283, + 0.83 + ], + "angle": 0, + "content": "5.2. Comparisons on DTU" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Similar to the LLFF dataset, we conduct four-group comparisons on the DTU dataset, as shown in Table 2. In this first group, given a target scene, SRF, PixelNeRF, and MVSNeRF are pre-trained on other DTU scenes and are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Table 2: Quantitative Comparison on DTU with three views. The first group is pre-trained on other scenes and tested on a target scene (P). The second group requires further finetuning on a target scene (P&FT). The third group uses geometry and semantic constraints (geo. & sem.). The fourth group distills depth knowledge (depth KD)." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.617, + 0.89, + 0.788 + ], + "angle": 0, + "content": "
SettingPSNR ↑SSIM ↑LPIPS ↓
SRF [12]P15.320.6710.304
PixelNeRF [64]16.820.6950.270
MVSNeRF [5]18.630.7690.197
SRF ft [12]P&FT15.680.6980.281
PixelNeRF ft [64]18.950.7100.269
MVSNeRF ft [5]18.540.7690.197
Mip-NeRF [1]geo. & sema.8.680.5710.353
DietNeRF [25]11.850.6330.314
RegNeRF [37]18.890.7450.190
MonoSDF* [65]depth KD18.920.7480.237
DSNeRF [17]16.900.5700.450
SparseNeRF(Ours)19.550.7690.201
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "tested on the target scene. In the second group, these three methods are further fine-tuned on the target scene. In the third group, we compare Mip-NeRF, DietNeRF, and RegN-eRF. Note that for the pre-trained models, they split DTU as a training set (88 scenes) and a test set (15 scenes). Since" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.514, + 0.937 + ], + "angle": 0, + "content": "9071" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.091, + 0.184, + 0.102 + ], + "angle": 0, + "content": "DSNeRF" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.106, + 0.239, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.092, + 0.349, + 0.102 + ], + "angle": 0, + "content": "MonoSDF*" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.106, + 0.402, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.092, + 0.507, + 0.104 + ], + "angle": 0, + "content": "RegNeRF" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.106, + 0.564, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.092, + 0.706, + 0.104 + ], + "angle": 0, + "content": "SparseNeRF (ours)" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.106, + 0.726, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.092, + 0.838, + 0.102 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.106, + 0.888, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.178, + 0.239, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.178, + 0.402, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.178, + 0.564, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.178, + 0.726, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.178, + 0.888, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.892, + 0.293 + ], + "angle": 0, + "content": "Figure 7: Visual comparisons on the NVS-RGBD dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes." + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.304, + 0.173, + 0.315 + ], + "angle": 0, + "content": "w/o conti. & ranking" + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.305, + 0.252, + 0.314 + ], + "angle": 0, + "content": "w/o ranking" + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.305, + 0.341, + 0.313 + ], + "angle": 0, + "content": "w/o conti." + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.305, + 0.454, + 0.314 + ], + "angle": 0, + "content": "SparseNeRF(ours)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.315, + 0.177, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.316, + 0.275, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.316, + 0.371, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.316, + 0.468, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.384, + 0.469, + 0.445 + ], + "angle": 0, + "content": "Figure 8: Ablation study on visual effect. The three columns denote \"w/o continuity and ranking distillation\", \"w/o continuity distillation\", and \"w/ continuity and ranking distillation\" (SparseNeRF, ours), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.744 + ], + "angle": 0, + "content": "the training set and the test set contains similar scenes, a test scene shares a similar distribution (e.g., similar backgrounds) with training scenes. Pre-training on other scenes greatly benefits a target scene. Therefore, SRF, PixelNeRF, and MVSNeRF also greatly achieve promising results. Finally, we compare depth-based DSNeRF [17] and MonoSDF [65]. DSNeRF studies two kinds of depth information, i.e., sparse 3D points generated by COLMAP [42] and ground-truth depth maps obtained by high-accuracy depth scanners and the Multi-View Stereo (MVS) algorithm. The depth maps of NVS-RGBD are captured by consumer-level sensors and contain lots of noise, which is hard to be applied to DSNeRF. Therefore, we adopt sparse 3D points for DSNeRF. Compared with these methods, SparseNeRF still achieves the best performance without pre-training on other scenes. We provide a qualitative analysis, as shown in Figure 6. SparseNeRF achieves better visual results than previous state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.337, + 0.772 + ], + "angle": 0, + "content": "5.3. Comparisons on NVS-RGBD" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We then implement three state-of-the-art methods on our new NVS-RGBD dataset, including RegNeRF [37], DSNeRF [17], and MonoSDF [65]. All of them are optimized per scene without pre-training on the other scenes. RegNeRF adopted continuous geometry assumptions from unobserved viewpoints. DSNeRF uses a global scale-invariant depth constraint to supervise NeRFs. MonoSDF mainly focuses on implicit surface reconstruction. It is as-" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.305, + 0.89, + 0.334 + ], + "angle": 0, + "content": "Table 3: Comparison with State of the Arts on NVS-RGBD. * denotes a new implementation from related tasks." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.335, + 0.892, + 0.455 + ], + "angle": 0, + "content": "
MetricsPSNR↑SSIM↑LPIPS↓depth err↓
KinectRegNeRF [37]25.780.8400.2425.9×10-3
DSNeRF [17]25.910.8380.2384.8×10-3
MonoSDF* [65]25.500.8380.2395.5×10-3
SparseNeRF(ours)26.280.8500.2323.9×10-3
ZED 2RegNeRF [37]24.980.7810.2713.5×10-3
DSNeRF [17]24.970.7890.2623.1×10-3
MonoSDF* [65]24.540.7840.2753.2×10-3
SparseNeRF(ours)26.220.8040.2622.4×10-3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.581 + ], + "angle": 0, + "content": "sumed that local patches of depth maps from monocular cameras satisfy \\( d' = w d + b \\) where \\( d \\) is predicted by the pre-trained model and \\( d' \\) is then computed by the NeRF. In Table 3, we observe that SparseNeRF achieves better performance compared with these prior arts. We provide a qualitative analysis in Figure 7, demonstrating the superiority of SparseNeRF over previous state-of-the-art methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.597, + 0.889, + 0.612 + ], + "angle": 0, + "content": "Table 4: Comparison of Depth Ranking and Depth scaling." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.618, + 0.891, + 0.68 + ], + "angle": 0, + "content": "
LLFFPSNR↑SSIM↑LPIPS↓depth err↓
baseline19.080.5870.33610.1×10-3
w/ depth scaling18.450.5650.3888.9×10-3
w/ depth ranking19.860.6240.3286.3×10-3
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.834, + 0.712 + ], + "angle": 0, + "content": "5.4. Ablation Studies and Further Analyses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Effectiveness of Depth Distillation. To validate the effectiveness of the depth distillation, we isolate the depth constraint and keep other modules unchanged. As shown in the first and third rows of Table 4, it is observed that without the local depth ranking distillation, SparseNeRF degrades on PSNR, SSIM, LPIPS, and depth error, which demonstrates the effectiveness of the depth distillation. We further study the depth maps from unobserved views. The ablation study of visual effect is shown in Figure 8. It shows that depth ranking greatly improves performance in geometry. Spatial continuity regularization further improves the detailed coherence of scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "9072" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.09, + 0.452, + 0.105 + ], + "angle": 0, + "content": "Table 5: Impact of different pre-trained depth models." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.111, + 0.468, + 0.192 + ], + "angle": 0, + "content": "
LLFFPSNR↑SSIM↑LPIPS↓depth err↓
Baseline19.080.5870.33610.1×10-3
MiDaS small19.350.5920.3336.9×10-3
DPT Hybrid19.860.6250.3266.3×10-3
DPT Large19.860.6240.3286.3×10-3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.205, + 0.47, + 0.431 + ], + "angle": 0, + "content": "Further Analyses on Depth Ranking and Depth Scaling. To validate the robustness of local depth ranking, we compare depth ranking regularization and depth scaling regularization on LLFF. We use the idea from MonoSDF [65], which mainly focuses on implicit surface reconstruction. With the specific design for surface shapes, MonoSDF slightly degrades the performance of RGB view synthesis. We re-implement it as depth scaling. Since the single-view depth estimation is not accurate, the linear depth scaling constraint is too strong such that depth distillation misleads the learning of the NeRF. Figure 3 shows two cases of scale-invariant errors. Instead, depth ranking relaxes the constraint and only focuses on depth comparison, which is more robust than depth scaling. As shown in Table 4, depth ranking performs better than depth scaling on LLFF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.521 + ], + "angle": 0, + "content": "Impact of Different Pre-trained Depth Models. We test our method on three depth estimation models, i.e., MiDaS small, DPT Hybrid, and DPT Large. As shown in Table 5, all of the depth models are better than the baseline. DPT Hybrid and DPT Large achieve comparable results, which are better than MiDaS small." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.468, + 0.627 + ], + "angle": 0, + "content": "FreeNeRF [61] potentially complements SparseNeRF. FreeNeRF and SparseNeRF improve few-shot NeRF in different ways. FreeNeRF introduces a frequency regularization while SparseNeRF distills depth priors from pretrained general depth estimators. We implement SparseNeRF on FreeNeRF and the results significantly improve, as shown in Table 6." + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.639, + 0.462, + 0.654 + ], + "angle": 0, + "content": "Table 6: FreeNeRF potentially complement SparseNeRF." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.667, + 0.47, + 0.715 + ], + "angle": 0, + "content": "
LLFFPSNR↑SSIM↑LPIPS↓
FreeNeRF19.600.6140.302
FreeNeRF+our depth distillation20.260.6460.292
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "User Study. To show that the proposed spatial continuity regularization improves the 3D consistency in 3D space, we conducted a user study. For each scene, we rendered videos for both with and without spatial continuity regularization terms. Users are asked to select the better-quality video according to 3D consistency/coherence and completeness (good geometry, few floaters). We evaluate 8 scenes on LLFF and there are 27 participants. The result shows \\(88.9\\%\\) of users prefer \"w/ conti\" and \\(11.1\\%\\) prefer \"w/o conti\". To evaluate the advantages of SparseNeRF on 3D consistency and shape completeness, we conduct another user study. Users are asked to evaluate \"baseline+RegNeRF's continu" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "ity term” and “baseline+our continuity term”. The 27 users participated in the study and 8 scenes on LLFF are evaluated. The result shows \\(87.5\\%\\) of users prefer “baseline+our continuity term” and \\(12.5\\%\\) prefer “baseline+RegNeRF’s continuity term”. RegNeRF encourages the overall patch to be as smooth as possible without considering the sharp edges of objects, which conflicts with the reconstruction from sparse views. Our continuity regulation distills the neighbor relationship from depth models is more accurate. We show two examples of depth maps in Figure 9." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.256, + 0.892, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.335, + 0.874, + 0.35 + ], + "angle": 0, + "content": "Figure 9: Visual comparison on continuity regulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Discussion. SparseNeRF could have a wide range of applications due to its promising results. We provide two applications on iPhone 13 Pro. Please refer to the project page for more experiments. Like UNISURF [38, 65], we also evaluate the SparseNeRF using Chamfer distance, showing the competitive results with MonoSDF and UNISURF, especially for complex scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.489, + 0.618, + 0.504 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.65 + ], + "angle": 0, + "content": "In this paper, we propose a SparseNeRF framework that synthesizes novel views with sparse view inputs. To tackle the under-constrained few-shot NeRF problem, we propose a local depth ranking regularization that distills the depth ranking prior from coarse depth maps. To preserve the spatial continuity of geometry, we propose a spatial continuity regularization that distills the depth continuity priors. With the proposed depth priors, SparseNeRF achieves a new state-of-the-art performance on three datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.696 + ], + "angle": 0, + "content": "Limitation. SparseNeRF significantly improves the performance of few-shot NeRF, but cannot be generalized to occluded views that are unobserved in the training views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.832 + ], + "angle": 0, + "content": "Potential Negative Impact. SparseNeRF focuses on scene synthesis. It might be misused to create misleading content. Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is also supported by Singapore MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20221-0011) and NTU NAP Grant." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.848, + 0.597, + 0.863 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srini" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "9073" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "vasan. Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In Int. Conf. Comput. Vis., 2021. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.192, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[3] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: Composable transformations of Python+NumPy programs, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.263, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MvsNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In Int. Conf. Comput. Vis., 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.39, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[6] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Adv. Neural Inform. Process. Syst., 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.433, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[7] Xipeng Chen, Guangrun Wang, Dizhong Zhu, Xiaodan Liang, Philip Torr, and Liang Lin. Structure-preserving 3D garment modeling with neural sewing machines. Adv. Neural Inform. Process. Syst., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.49, + 0.472, + 0.545 + ], + "angle": 0, + "content": "[8] Zixuan Chen, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. CuNeRF: Cube-based neural radiance field for zero-shot medical image arbitrary-scale super resolution. Int. Conf. Comput. Vis., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.472, + 0.586 + ], + "angle": 0, + "content": "[9] Zhaoxi Chen and Ziwei Liu. Relighting4D: Neural re- ligable human from videos. In Eur. Conf. Comput. Vis., 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[10] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Trans. Graph., 41(6):1-16, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[11] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Scenedreamer: Unbounded 3D scene generation from 2D image collections. In arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[12] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[13] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[14] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv:1602.02481, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[15] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[16] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "NerDI: Single-view NeRF synthesis with language-guided diffusion as general image priors. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[18] Chenhe Dong, Guangrun Wang, Hang Xu, Jiefeng Peng, Xiaozhe Ren, and Xiaodan Liang. Efficientbert: Progressively searching multilayer perceptron via warm-up knowledge distillation. arXiv preprint arXiv:2109.07222, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[19] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[22] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. EVA3D: Compositional 3D human generation from 2D image collections. arXiv preprint arXiv:2210.04888, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[23] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human NeRF from a single image. arXiv preprint arXiv:2303.12791, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[24] Shoukang Hu, Kaichen Zhou, Kaiyu Li, Longhui Yu, Lanqing Hong, Tianyang Hu, Zhenguo Li, Gim Hee Lee, and Ziwei Liu. ConsistentNeRF: Enhancing neural radiance fields with 3D consistency for sparse view synthesis. arXiv preprint arXiv:2305.11031, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[25] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a diet: Semantically consistent few-shot view synthesis. In Int. Conf. Comput. Vis., 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[26] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In IEEE Conf. Comput. Vis. Pattern Recog., 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[27] Mijeong Kim, Seonguk Seo, and Bohyung Han. InfoNeRF: Ray entropy minimization for few-shot neural volume rendering. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[28] Changlin Li, Bohan Zhuang, Guangrun Wang, Xiaodan Liang, Xiaojun Chang, and Yi Yang. Automated progressive learning for efficient training of vision transformers. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[29] Xiaoming Li, Wangmeng Zuo, and Chen Change Loy. Learning generative structure prior for blind text image super-resolution. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.926, + 0.516, + 0.937 + ], + "angle": 0, + "content": "9074" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[31] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Int. Conf. Comput. Vis., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.178 + ], + "angle": 0, + "content": "[32] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Adv. Neural Inform. Process. Syst., 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.181, + 0.469, + 0.235 + ], + "angle": 0, + "content": "[33] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. GNeRF: GAN-based neural radiance field without posed camera. In Int. Conf. Comput. Vis., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.239, + 0.469, + 0.279 + ], + "angle": 0, + "content": "[34] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In IEEE Conf. Comput. Vis. Pattern Recog., 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.282, + 0.469, + 0.35 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Trans. Graph., 38(4):1-14, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.354, + 0.469, + 0.409 + ], + "angle": 0, + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis. Springer, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.412, + 0.469, + 0.48 + ], + "angle": 0, + "content": "[37] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg-NeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.483, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[38] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Int. Conf. Comput. Vis., 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.541, + 0.469, + 0.595 + ], + "angle": 0, + "content": "[39] Jiefeng Peng, Jiqi Zhang, Changlin Li, Guangrun Wang, Xiaodan Liang, and Liang Lin. Pi-nas: Improving neural architecture search by reducing supernet training consistency shift. In Int. Conf. Comput. Vis., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.599, + 0.469, + 0.653 + ], + "angle": 0, + "content": "[40] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.657, + 0.469, + 0.696 + ], + "angle": 0, + "content": "[41] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Int. Conf. Comput. Vis., 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.701, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[42] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.744, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[43] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Eur. Conf. Comput. Vis. Springer, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[44] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Adv. Neural Inform. Process. Syst., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[45] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[46] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[47] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-NeRF: Scalable large scene neural view synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[48] Tang Tao, Longfei Gao, Guangrun Wang, Peng Chen, Dayang Hao, Xiaodan Liang, Mathieu Salzmann, and Kaicheng Yu. LiDAR-NeRF: Novel lidar view synthesis via neural radiance fields. arXiv preprint arXiv:2304.10406, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[49] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-NeRF: Scalable construction of large-scale NeRFs for virtual fly-throughs. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[50] Guangrun Wang, Ping Luo, Xinjiang Wang, Liang Lin, et al. Kalman normalization: Normalizing internal representations across network layers. Adv. Neural Inform. Process. Syst., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[51] Guangrun Wang and Philip HS Torr. Traditional classification neural networks are good generators: They are competitive with DDPMs and GANs. arXiv preprint arXiv:2211.14794, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[52] Guangrun Wang, Guangcong Wang, Keze Wang, Xiaodan Liang, and Liang Lin. Grammatically recognizing images with tree convolution. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[53] Guangrun Wang, Keze Wang, and Liang Lin. Adaptively connected neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[54] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In Eur. Conf. Comput. Vis., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[55] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[56] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[57] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[58] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. CityNeRF: Building nerf at city scale. arXiv preprint arXiv:2112.05504, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[59] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. SinNeRF: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[60] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.926, + 0.515, + 0.936 + ], + "angle": 0, + "content": "9075" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "2D photo to a 3D object with 360deg views. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[61] Jiawei Yang, Marco Pavone, and Yue Wang. FreeNeRF: Improving few-shot neural rendering with free frequency regularization. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[62] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Vtoonify: Controllable high-resolution portrait video style transfer. ACM Trans. Graph., 41(6):1-15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[63] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Gp-unit: Generative prior for versatile unsupervised image-to-image translation. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.468, + 0.319 + ], + "angle": 0, + "content": "[64] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.376 + ], + "angle": 0, + "content": "[65] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 1, 2, 3, 4, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.468, + 0.432 + ], + "angle": 0, + "content": "[66] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[67] Long Zhuo, Guangcong Wang, Shikai Li, Wanye Wu, and Ziwei Liu. Fast-vid2vid: Spatial-temporal compression for video-to-video synthesis. In Eur. Conf. Comput. Vis., 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.517, + 0.937 + ], + "angle": 0, + "content": "9076" + } + ] +] \ No newline at end of file diff --git a/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/0d56c181-4428-4dcb-b7bc-71667b8724d5_origin.pdf b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/0d56c181-4428-4dcb-b7bc-71667b8724d5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b83707f23f34a0a2ccbffef0415a315fae117ac3 --- /dev/null +++ b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/0d56c181-4428-4dcb-b7bc-71667b8724d5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f722b83042f4cda544cf8ab2d2e42ed966173055301cce7c2d4306c495a99d7 +size 5442456 diff --git a/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/full.md b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0ed3ea695b27cbe9dc0c00ae4e7a4f3993953958 --- /dev/null +++ b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/full.md @@ -0,0 +1,362 @@ +# SparseNeRF: Distilling Depth Ranking for Few-shot Novel View Synthesis + +Guangcong Wang + +Zhaoxi Chen + +Chen Change Loy + +S-Lab, Nanyang Technological University + +Ziwei Liu + +{guangcong.wang,zhaoxi00,ccloy,ziwei.liu}@ntu.edu.sg + +![](images/25c38c11645eed074b62c9cd251e49a21ca4b5feb2027f7bacbcf93d5ee02a47.jpg) + +![](images/c000f477d6f0062339011af44cea59e90cc1d9c297bd721e449fa2ee9a2af937.jpg) + +![](images/3525d8a67f84551639343e0c7d35420e43791c3fc381f519533a4d755ac8ba24.jpg) +Novel view (GT) + +![](images/7ecff797f74986a8d24f98fbc82ca52102ea24122b0fcb1306b7f5be77c5a642.jpg) + +![](images/28f91706afef13fe42b5b4fff2f36c690bddd7e94960c5ed90829a4c159bc04f.jpg) + +![](images/00557e9e82b0069ff9eaa39ce225d64f476f634688ee1c927bb85482624c8a08.jpg) +RegNeRF + +![](images/64a3f2fecffe8f95e59346a101a75fb57de9e06c4153c47363fce93c651ce5f9.jpg) + +![](images/b1841394780b111d42dcab0a5beb53c863612ac147868dd0101632247db5cde5.jpg) + +![](images/72e5378b92d1d159efc28a474aa139d38d2622e5dcf200a318e81d23c6005967.jpg) +MonoSDF* +Figure 1: Comparing the state-of-the-arts RegNeRF [37], MonoSDF* [65], DSNeRF [17], and our SparseNeRF with three views for training (* denotes a re-implementation for our task). RegNeRF regularizes geometry with sparsity and continuity constraints. MonoSDF, and DSNeRF use scale-invariant depth constraints supervised by coarse depth maps. Our SparseNeRF uses robust depth ranking of coarse depth maps. SparseNeRF can synthesize realistic novel views and coherent geometric depth (Please refer to the supplementary material for rendered videos). + +![](images/2533c19f69afba9e908ae86e9e2ae45eb85342aaa2d47b82e371b615bdb4f19d.jpg) + +![](images/97fe940b37493e134d7c866bb59289c2d1795a559e98eb870ffd908bea07a231.jpg) + +![](images/897fa3094c9469d8eb57d64785200bc4051bd287cbc097e259f59f0239dbf5c9.jpg) +DSNeRF + +![](images/ac1b1ff45e41977f79c4b1e24c12922dc34e7d801aae9da99a16a20fde31f722.jpg) + +![](images/129fb981b22218603b81438cf257136b573938aaa5575d6a9a7f763d1feb7ff0.jpg) + +![](images/2dcadd5a88e5ff3eb09d5d4865a81e4f90813db6baf7442b71da0bd2b74d3a25.jpg) +SparseNeRF (ours) + +# Abstract + +Neural Radiance Field (NeRF) significantly degrades when only a limited number of views are available. To complement the lack of 3D information, depth-based models, such as DSNeRF and MonoSDF, explicitly assume the availability of accurate depth maps of multiple views. They linearly scale the accurate depth maps as supervision to guide the predicted depth of few-shot NeRFs. However, accurate depth maps are difficult and expensive to capture due to wide-range depth distances in the wild. This work presents a new Sparse-view NeRF (SparseNeRF) framework that exploits depth priors from real-world inaccurate observations. The inaccurate depth observations are either from pre-trained depth models or coarse depth maps of consumer-level depth sensors. Since coarse depth maps are not strictly scaled to the ground-truth depth maps, we propose a simple yet effective constraint, a local depth ranking method, on NeRFs such that the expected depth ranking of the NeRF is consistent with that of the coarse depth maps in local patches. To preserve the spatial continuity of the esti + +mated depth of NeRF, we further propose a spatial continuity constraint to encourage the consistency of the expected depth continuity of NeRF with coarse depth maps. Surprisingly, with simple depth ranking constraints, SparseNeRF outperforms all state-of-the-art few-shot NeRF methods (including depth-based models) on standard LLFF and DTU datasets. Moreover, we collect a new dataset NVS-RGBD that contains real-world depth maps from Azure Kinect, ZED 2, and iPhone 13 Pro. Extensive experiments on NVS-RGBD dataset also validate the superiority and generalizability of SparseNeRF. Code and dataset are available at https://sparsenerf.github.io/. + +# 1. Introduction + +Neural radiance fields (NeRFs) [36, 32, 1, 2, 44] have made tremendous progress in generating photo-realistic novel views of scenes by optimizing implicit function representations given a set of 2D input views. However, in a wide range of real-world scenarios, collecting dense views of a + +scene is often expensive and time-consuming [47]. Therefore, it is necessary to develop few-shot NeRF methods that can be learned from sparse views without significant degradation in performance. + +Learning a NeRF from sparse views is a challenging problem due to under-constrained reconstruction conditions, especially in textureless areas. Directly applying NeRFs to few-shot scenarios suffers from dramatic degradation [37]. Recently, some methods have greatly improved the performance of few-shot NeRF, which can be categorized into three groups. 1) The first group [37, 27, 25] is based on geometry constraints (sparsity and continuity regularizations) and high-level semantics. RegNeRF [37] regularized the geometry and appearance of patches rendered from unobserved viewpoints, and annealed the ray sampling space. InfoNeRF [27] imposed an entropy constraint of the density in each ray and a spatial smoothness constraint into the estimated images. However, since a scene often contains multiple layouts (Figure 1), sparsity and continuity geometric constraints of a few views cannot guarantee the complete 3D geometric reconstruction. 2) The second group [64, 5] resorts to pre-training on similar scenes. For example, PixelNeRF [64] proposed to condition a NeRF on convolutional feature maps to learn high-level semantics from other scenes. 3) The third group [17, 65, 24] exploits depth maps and makes a linearity assumption of depth maps to supervise few-shot NeRFs. For example, DSNeRF [17] exploited sparse 3D points generated by COLMAP [42] or accurate depth maps [13] obtained by high-accuracy depth scanners and the Multi-View Stereo (MVS) algorithm. The depth maps are linearly scaled as supervision to guide the predicted depth of few-shot NeRFs. To use coarse depth maps, MonoSDF [65] uses a local patch-based scale-invariant depth constraint supervised by coarse depth maps instead of global depth maps. However, the scale-invariant depth constraint is strong for real-world coarse depth maps from pre-trained depth models or consumer-level depth sensors due to wide-range depth distances in the wild. + +Along the third group, we wish to explore more robust 3D priors from coarse depth maps to complement the under-constrained few-shot NeRF. To address this problem, we present SparseNeRF, a simple yet effective method that distills depth priors from pre-trained depth models [41] or inaccurate depth maps from consumer-level depth sensors (Figure 3), which can be easily obtained from real-world scenes. Deriving useful depth cues from such pre-trained models is non-trivial. In particular, although single-view depth estimation methods have achieved good visual performance, thanks to large-scale monocular depth datasets and large ViT models, they cannot yield accurate 3D depth information due to coarse depth annotations, dataset bias, and ill-posed 2D single-view images. The inaccurate depth information contradicts the density prediction of a NeRF + +when reconstructing each pixel of a 3D scene based on volume rendering. Directly scaling the coarse depth maps to a NeRF [17, 65] leads to inconsistent geometry against the expected depth of the NeRF. + +Instead of directly supervising a NeRF with coarse depth priors, we relax hard depth constraints [17, 65] and distill robust local depth ranking from the coarse depth maps to a NeRF such that the depth ranking of a NeRF is consistent with that of coarse depth. That is, we supervise a NeRF with relative depth instead of absolute depth [17, 65]. To guarantee the spatial continuity of geometry, we further propose a spatial continuity constraint on depth maps such that the NeRF model imitates the spatial continuity of coarse depth maps. The accurate sparse geometry constraints from a limited number of views, combined with relaxed constraints including depth ranking regularization and continuity regularization, finally achieve promising novel view synthesis (Figure 1). It is noteworthy that SparseNeRF does not increase the running time during inference as it only exploits depth priors from pre-trained depth models or consumer-level sensors during the training stage (Figure 2). In addition, SparseNeRF is a plug-and-play module that can be easily integrated into various NeRFs. + +The main contributions of this paper are 1) SparseNeRF, a simple yet effective method that distills local depth ranking priors from pre-trained depth models. With the help of the local depth ranking constraint, SparseNeRF significantly improves the performance of few-shot novel view synthesis over the state-of-the-art models (including depth-based NeRF methods). To preserve the coherent geometry of a scene, we propose a spatial continuity distillation constraint that encourages the spatial continuity of NeRF to be similar to that of the pre-trained depth model. Both depth ranking prior and spatial continuity distillation are new in the literature on NeRF. 2) Apart from SparseNeRF, we also contribute a new dataset, NVS-RGBD, which contains coarse depth maps from Azure Kinect, ZED 2, and iPhone 13 Pro. 3) Extensive experiments on the LLFF, DTU, and NVS-RGBD datasets demonstrate that SparseNeRF achieves a new state-of-the-art performance in few-shot novel view synthesis. + +# 2. Related Work + +Neural Radiance Fields. NeRF [36, 48, 8, 39] has made great success in synthesizing novel views of complex scenes due to good representation of neural networks [20, 52, 50, 53, 18, 28]. Block-NeRF [47], CityNeRF [58], and Mega-NeRF [49] scaled the standard NeRF up to city-scale or urban-scale scenes. NeRF—[56], GNeRF [33], and BARF [31] relaxed the requirements of NeRFs and synthesized novel views without perfect camera poses. Some works attempted to improve NeRFs by considering antialiasing [1], sparse 3D grids with spherical harmonics [19]. + +Some methods [40, 30, 57] extended NeRFs to dynamic scenes. These methods do not focus on generating novel views with a few views. In this paper, we study sparse-view NeRF to reduce dense capture requirements in real-world applications. + +Few-shot Novel View Synthesis. There are increasing studies on few-shot novel view synthesis. Basically, the existing few-shot NeRF methods can be categorized into three groups. First, some methods exploit continuity constraints on geometry or object semantics. For example, RegNeRF [37] imposed a continuity constraint on geometry and regularized the appearance of patches from unobserved viewpoints with a flow model. InfoNeRF [27] proposed a ray entropy minimization regularization to encourage the density to be as sparse as possible along a ray, and used a ray information gain reduction regularization to constrain the continuous depth of neighbor rays. Second, some methods attempt to pe-train a NeRF on other similar scenes and fine-tune the NeRF on the target scene. For example, PixelNeRF [64] conditioned a NeRF on image inputs in a fully convolutional manner, allowing the model to learn scene priors from other scenes and reduce the requirement of dense views. MVSNeRF [5] leveraged plane-swept cost volumes for geometry-aware scene reasoning and combined it with physically based volume rendering. Similar to PixelNeRF, MVSNeRF was first trained on other real scenes and was finetuned on target scenes to evaluate its effectiveness and generalizability. Third, depth-based models [17, 65] use available depth information to supervise the training of NeRFs. DSNeRF [17] exploited sparse 3D points generated by COLMAP [42] or accurate depth maps [13] obtained by high-accuracy depth scanners. Different from these depth-based models, SparseNeRF distills robust depth ranking from pre-trained depth models or coarse depth maps from consumer-level sensors. As concurrent works, NeRDi [16] and NeuralLift-360 [60] also use ranking-based methods. However, they mainly focus on the single-view setting while SparseNeRF focuses on sparse-view. SparseNeRF uses ranking loss to avoid inconsistent 3D geometry across different views while [16, 60] focuses on a soft geometric regularization on a single view (no cross-view consistency problem during training). Moreover, SparseNeRF introduces a new spatial continuity loss to distill spatial coherence from monocular depth estimators. In addition, some single-view synthesis methods [59] either allow new generative objects from unseen views or focus on specific objects, e.g., face [4] and human [23]. Some NeRF-based methods [22, 11] extended to generation instead of reconstruction. Other AIGC models [21, 51, 54, 67, 10, 9, 63, 29, 62] focus on the 2D generation or category-specific reconstruction [7], which are different. + +# 3. Our Approach + +We present SparseNeRF to synthesize novel views given sparse view inputs. Single-view depth estimation is a long-standing computer vision task, aiming to predict a depth map given a single image. In this paper, we are interested in mining the depth priors encapsulated in pre-trained models of single-view depth estimation or coarse depth maps captured by consumer-level depth sensors. However, due to coarse annotations of single-view depth maps, (e.g., user clicks [6], RGB-D [43, 46], and laser/stereo [34]), dataset bias, and imperfect depth estimation models, it is challenging to obtain accurate 3D depth estimation given 2D single-view images. As for consumer-level depth sensors, it still struggles to capture accurate depth maps (Figure 3). Driven by these observations, our goal is to make use of the coarse depth maps and distill useful depth priors to guide the learning of a NeRF. + +# 3.1. Preliminary and Problem Formulation + +Neural Radiance Fields. Let a NeRF [36] be a mapping function $f$ that maps a 3D spatial location $\mathbf{x}$ and a viewing direction $\mathbf{d}$ into a volume density $\sigma$ and a color value $\mathbf{c}$ . The $f$ is a neural network consisting of eight perceptron (MLP) layers parameterized by $\theta$ , which is given by $f_{\theta}: (\gamma(\mathbf{x}), \gamma(\mathbf{d})) \mapsto (\sigma, \mathbf{c})$ , where $\gamma$ is a positional encoding. For each expected pixel color $\hat{C}(\mathbf{r})$ , it is rendered by casting a ray $\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$ with near and far bounds $t_n$ and $t_f$ . We evenly partition $[t_n, t_f]$ into $N$ points $(t_1, t_2, \dots, t_N)$ along a ray $\mathbf{r}$ and compute expected pixel color $\hat{C}(\mathbf{r})$ by $\hat{C}(\mathbf{r}) = \sum_{i=1}^{N} \mathbf{c}_i^*$ . The weighted color $\mathbf{c}_i^*$ of a 3D point is computed by $\mathbf{c}_i^* = w_i\mathbf{c}_i$ , where $w_i = T_i(1 - \exp(-\sigma_i\delta_i))$ , $T_i = \exp(-\sum_{j=1}^{i-1}\sigma_j\delta_j)$ and $\delta_i = t_i - t_{i-1}$ . Therefore, the NeRF reconstruction loss can be formulated as + +$$ +\mathcal {L} _ {\text {n e r f}} = \sum_ {\mathbf {r} \in R} \left| \left| \hat {C} (\mathbf {r}) - C (\mathbf {r}) \right| \right| ^ {2}, \tag {1} +$$ + +where $\hat{C}(\mathbf{r})$ are rendered color blended by $N$ samples. $C(\mathbf{r})$ is the ground-truth pixel color. We use coarse-to-fine sampling as discussed in the vanilla NeRF [36]. Here we omit the fine rendering for simplification. + +Problem Formulation. Vanilla NeRFs aim to learn a mapping function $f_{\theta}$ with $K_{d}$ dense views by optimizing a color reconstruction loss in Eq. (1). In this work, we aim to study few-shot NeRF when only $K_{s}$ sparse views are available ( $K_{s} \ll K_{d}$ ). Let $H$ and $W$ denote the height and width of an image, we have $HWK_{s}$ rays as color reconstruction constraints. Without considering view directions $\mathbf{d}$ and image continuity assumptions (i.e., $\mathbf{c}_{i}^{*}$ is independent of $\mathbf{c}_{j}^{*}$ when $i \neq j$ ), suppose we have a $N_{c} \times N_{c} \times N_{c}$ discrete cubic volume to be optimized, which contains $N_{c}^{3}$ weighted color variables $\mathbf{c}_{i}^{*}$ . We have $N_{c}^{3}$ variables $\mathbf{c}_{i}^{*}$ and $HWK_{s}$ constraints. Ideally, we are able to solve a discrete cube of edge + +![](images/9bb0f5b71768058838c50b5be5b3168e9fd35622762bea1603009483c95d270f.jpg) +Figure 2: Framework Overview. SparseNeRF consists of two streams, i.e., NeRF and depth prior distillation. As for NeRF, we use Mip-NeRF as the backbone. We use a NeRF reconstruction loss $\mathcal{L}_{\mathrm{nerf}}$ . As for depth prior distillation, we distill depth priors from a pre-trained depth model. Specifically, we propose a local depth ranking regularization and a spatial continuity regularization to distill robust depth priors from coarse depth maps. + +length $N_{c} = \sqrt[3]{HWK_{s}}$ . That is, we can sample $\sqrt[3]{HWK_{s}}$ for each edge of a cubic volume. Take $H = W = 512$ and $K_{s} = 3$ as an example, we have $N_{c} \approx 92$ , which is far from reconstructing a continuous radiance field. To address this under-constrained optimization problem of few-shot NeRF, an intuitive way is to introduce reasonable regularization terms to constrain few-shot NeRFs conditioned on input $\mathbf{x}$ , $\mathbf{d}$ , and $C(\mathbf{r})$ . Considering these constraints, a general formulation is given by + +$$ +\mathcal {L} = \mathcal {L} _ {\text {n e r f}} + \lambda \mathcal {R} (\mathbf {x}, \mathbf {d}, C (\mathbf {r})), \tag {2} +$$ + +where $\mathcal{R}$ is a regularization term. + +Remark. RegNeRF [37] and InfoNeRF [27] tackle this problem by introducing regularization terms on continuous depth constraint from unobserved viewpoints, sparsity of density on rays, and patch-based semantic constraint on color appearance. Depth-based models, such as DSNeRF [17] and MonoSDF [65], directly regress depth by leveraging sparse 3D points or deriving ground-truth depth maps with accurate absolute depth maps. Different from them, we propose to use a robust relative depth regularization from coarse depth maps of pre-trained depth models or consumer-level depth sensors. + +# 3.2. Overview of SparseNeRF + +The pipeline of SparseNeRF is illustrated in Figure 2. SparseNeRF mainly consists of four components, i.e., a neural radiance field (NeRF), a color reconstruction module, a depth ranking distillation module, and a spatial continuity distillation module. Specifically, we use Mip-NeRF [1] as the backbone and apply an MSE loss for color reconstruction $\mathcal{L}_{\mathrm{nerf}}$ . As for depth prior distillation, we either use a pre-trained depth model to estimate depth maps or capture coarse depth maps with consumer-level depth sensors. We use a vision transformer (DPT) [41] that is trained on a large-scale mixed depth dataset (1.4M images with various depth annotations) that covers a wide range of scenes. Therefore, DPT is able to provide general depth priors. Since single-view depth estimation is coarse, we carefully design a local depth ranking regularization and a spatial continuity regularization, which distills robust depth priors from coarse maps to the NeRF. + +# 3.3. Local Depth Ranking Distillation + +Single-view depth estimation is a challenging computer vision task, which aims at predicting a depth map of a scene given a single image as input. Due to dataset bias, coarse depth annotations, and imperfect neural models, it is difficult to achieve accurate depth prediction. Depth maps cap + +![](images/da4c811d6718d8ca818667df2c426261f95e8b65ea0c5ce7bbfe687ea89da00a.jpg) +Figure 3: Two types of coarse depth maps. Top: predicted by a pre-trained depth model. Bottom: captured by Azure Kinect and ZED 2 depth sensors. Si-MSE is a scale-invariant depth error. Time jittering refers to the difference between two depth maps captured by a static camera within a small time interval. + +tured by consumer-level sensors are also inaccurate due to wide-range depth distances. Directly using coarse depth maps to supervise a NeRF leads to ambiguous rendered novel view synthesis. To avoid the error of coarse depth maps, we relax the depth constraint and exploit the robust depth ranking prior. Given a pair of pixels in a single image, the depth ranking regularization only considers which point is nearer or farther. However, when the scene is complex, even depth ranking is not accurate. As shown in Figure 4, it is easy for depth estimation models to compare the depth ranking of white and cyan points, but it is hard to estimate the depth ranking of white and red points. It implies that the depth ranking becomes unreliable as the spatial distance increases. + +Motivated by these observations, we propose a local depth ranking distillation method that distills depth ranking priors from coarse depth maps to a NeRF. On one hand, given a local patch $P$ of an RGB image $I$ with a pose $p$ . We compute the depth $d_{\mathbf{r}}$ of the rays that trace from $P$ by $d_{\mathbf{r}} = \sum_{i=1}^{N} w_{i} t_{i}$ . On the other hand, we use pre-trained depth DPT [41] to estimate the depth of $I$ and crop a local patch $d_{\mathrm{dpt}}$ with the same spatial location as $d_{\mathbf{r}}$ . We perform the depth ranking distillation by transferring the depth ranking knowledge from $d_{\mathrm{dpt}}$ to $d_{\mathbf{r}}$ . Let $k1$ and $k2$ be the + +![](images/6ced21e759b7dfb9e657fc27487da1c300e891a5f38ee58dcfe3b39e49a6530c.jpg) +Figure 4: Motivation of local depth ranking. It is easy for depth estimation models to tell the white point is nearer than the cyan point, but it is hard to compare while and red points. Best viewed by zooming in. + +indices of 2D pixel coordinate of $d_{\mathrm{dpt}}$ and $d_{\mathbf{r}}$ . The depth ranking regularization is given by + +$$ +\mathcal {R} _ {\text {r a n k}} = \sum_ {d _ {\mathrm {d p t}} ^ {k 1} \leq d _ {\mathrm {d p t}} ^ {k 2}} \max \left(d _ {\mathbf {r}} ^ {k 1} - d _ {\mathbf {r}} ^ {k 2} + m, 0\right), \tag {3} +$$ + +where $m$ is a small margin that allows limited depth ranking errors. In Eq. (3), we randomly sample two depth pixels of $d_{\mathrm{dpt}}$ , as denoted as $d_{\mathrm{dpt}}^{k1}$ and $d_{\mathrm{dpt}}^{k2}$ , $d_{\mathrm{dpt}}^{k1} \leq d_{\mathrm{dpt}}^{k2}$ . If the depth rankings of $d_{\mathrm{dpt}}$ and $d_{\mathbf{r}}$ are not consistent, i.e., $d_{\mathbf{r}}^{k1} > d_{\mathbf{r}}^{k2}$ and $d_{\mathrm{dpt}}^{k1} \leq d_{\mathrm{dpt}}^{k2}$ , we punish the NeRF. We encourage the corresponding depth estimated by NeRF to satisfy the consistent depth ranking of the pre-trained model. We clip the large depth value and normalize the depth map for depth sensors. For depth models, we use relative inverse depth. + +# 3.4. Spatial Continuity Distillation + +The local depth ranking constraint guarantees that the predicted depth map estimated by NeRF is consistent with the DPT's depth map. However, it does not constrain the spatial continuity of the depth map. We distill spatial continuity priors from the depth model DPT, which allows large displacement across several depth pixels. If neighbor depth pixels are continuous on the depth map of DPT, we constrain the corresponding depth pixels of NeRF to be continuous. The spatial continuity regularization is given by + +$$ +\mathcal {R} _ {\text {c o n t i}} = \sum_ {k 1} \sum_ {d _ {\mathrm {d p t}} ^ {k 2} \in \mathrm {K N N} \left(d _ {\mathrm {d p t}} ^ {k 1}\right)} \max \left(\left| d _ {\mathbf {r}} ^ {k 1} - d _ {\mathbf {r}} ^ {k 2} \right| - m ^ {\prime}, 0\right), \tag {4} +$$ + +where $\mathrm{KNN}(\cdot)$ returns k-nearest neighbors measured by depth values within a small region, e.g., $6 \times 6$ patch. $m'$ is a small margin that allows small depth differences between neighbor pixels. + +Full Objective Loss. The full objective loss of SparseNeRF is formulated by + +$$ +\mathcal {L} = \mathcal {L} _ {\text {n e r f}} + \lambda \mathcal {R} _ {\text {r a n k}} + \gamma \mathcal {R} _ {\text {c o n t i}}, \tag {5} +$$ + +where $\lambda$ and $\gamma$ control the importance of the regularization terms. In practice, we set $\lambda = 0.2$ , $\gamma = 0.02$ , $m = m' = 1.0 \times 10^{-4}$ . + +# 4. NVS-RGBD Dataset + +The popular benchmark LLFF [35] on NeRFs is featured with forward-facing and object-centric. Moreover, we focus on consumer-level depth sensors instead of expensive scanners [14]. We study stand-alone depth cameras instead of a complex calibrated structured multiple camera system [15], and collect from real-world scenes instead of synthesized 3D assets [45]. Few RGBD datasets satisfy all of the requirements. Therefore, we collect a new dataset NVS-RGBD that contains real-world depth maps captured by consumer-level depth sensors, including Azure Kinect, ZED 2, and iPhone 13 Pro. We collect 8 scenes for both Azure Kinect and ZED 2, and 4 scenes for iPhone 13 Pro. The depth maps of Azure Kinect often contain noises around object edges. The depth maps of ZED 2 look smooth but unstable and incorrect when observing time jittering (Figure 3). The poses are computed via COLMAP. We collect indoor scenes for ZED 2 and Kinect. For iPhone, we collect both indoor and outdoor scenes. All the scenes are object-centric. Depth maps obtained from sensors have different artifacts coming from the sensor noises. Please refer to the supplementary material for the details. + +# 5. Experiments + +Datasets. We conduct experiments on the LLFF [35], DTU [26], and NVS-RGBD datasets. LLFF contains 8 complex forward-facing scenes. Following RegNeRF [37], we use every 8-th image as the held-out test set, and evenly sample sparse views from the remaining images for training. Different from LLFF, DTU is an object-level dataset. Following PixNeRF [64, 37], we use the same 15 scenes in our experiments. On the DTU dataset, the background of these scenes is a white table or a black background, which has few textures. As mentioned in RegNeRF, we mask the background during inference to avoid background bias. We use the same evaluation protocol for a fair comparison. On the NVS-RGBD dataset, for each scene, we use 3 views for training and the rest views for inference. For LLFF and DTU, we use pre-trained depth maps to implement our method. For NVS-RGBD, we use coarse depth maps captured by depth sensors. + +Evaluation Metrics. We adopt four evaluation metrics in the experiments, i.e., PSNR, SSIM [55], LPIPS [66], and depth error [17]. Depth error is a scale-invariant MSE derived by minimizing $||wd\hat{d} + b - d||$ , where $\hat{d}$ and $d$ denote coarse depth maps and pseudo-ground-truth depth maps predicted by dense-view NeRF. Note that we only use dense views to train a good NeRF to generate pseudo-ground-truth depth maps for evaluation. + +Implementation Details. We implement SparseNeRF based on the official JAX [3]. We use the Adam optimizer for the learning of SparseNeRF. We use an exponen + +tial learning rate, which decays from $2 \times 10^{-3}$ to $2 \times 10^{-5}$ . The batch size is set to 4096. For each scene, we use one GPU-v100 with 32G memory for both training and inference. We also implement the proposed SparseNeRF on GPU-v100 with 16G and RTX 2080 Ti with 11G. We reduce the batch size to 2048 and 1024 for these two types of GPUs, respectively. We use the same backbone and sampled points along rays compared with prior arts. + +We train 90k iterations for each scene given three training views, which takes about 2 hours for each scene. For depth maps captured by Kinect and ZED 2, we mask uncertain regions (black regions in depth maps). More precisely, we sample near-far depth pairs in certain regions in local patches to optimize Eqs. 3 and 4 of the paper. + +# 5.1. Comparisons on LLFF + +We compare SparseNeRF with state-of-the-art methods on the LLFF dataset, including SRF [12], PixelNeRF [64], MVSNeRF [5], Mip-NeRF [1], DietNeRF [25], RegNeRF [37], DSNeRF [17] and MonoSDF[65]. Among them, SRF, PixelNeRF, and MVSNeRF are pre-trained on other similar scenes to exploit high-level semantics. Mip-NeRF is the state-of-the-art NeRF designed for dense-view training. InfoNeRF and RegNeRF mainly constrain sparsity and continuity of geometry or natural appearance semantics. More precisely, the compared methods can be classified into four groups. In the first group, SRF, PixelNeRF, and MVSNeRF are pre-trained in the large-scale DTU datasets (88 scenes) and are directly tested on the LLFF dataset. In the second group, SRF, PixelNeRF, and MVSNeRF are pre-trained on DTU and fine-tuned on LLFF per scene. The third group includes Mip-NeRF, DietNeRF, and RegNeRF. They use geometry continuity and semantic constraints. The fourth group distills knowledge from depth maps or sparse points of COLMAP (DSNeRF). The compared results are shown in Table 1. We can see that the proposed SparseNeRF achieves the best performance in PSNR, SSIM, and LPIPS. Note that the first and second groups have to pre-train on lots of other scenes. SparseNeRF achieves better results than DSNeRF and MonoSDF because the depth ranking is more robust than the absolute scale-invariant constraint. + +We provide qualitative analysis on LLFF, as shown in Figure 5. It is observed that PixelNeRF and MVSNeRF tend to generate ambiguous pixels. The reason is that they integrate CNN features into NeRFs. CNN features can provide high-level semantics to infer under-constrained textures but suffer pixel-level inference. MipNeRF is designed for dense-view synthesis without considering the under-constrained problem, leading to degraded geometry artifacts. Compared with RegNeRF, SparseNeRF uses robust depth priors from pre-trained depth models, which can handle challenging scene geometry with better performance. + +![](images/75858878c15072142d88042dca52d770058855cb35cea1341aeae4c6fbac1427.jpg) + +![](images/26a7c24ddc19e29e415488dd817b644d64faea2cdd3e15a4b217d933dd618e99.jpg) +Figure 5: Visual comparisons on the LLFF dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes. +Figure 6: Visual comparisons on the DTU dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes. + +Table 1: Quantitative Comparison on LLFF with three views. There are four groups. The first group denotes models that are pre-trained on other scenes and tested on a target scene (P). The second group requires further finetuning on a target scene (P&FT). The third group uses geometry and semantic constraints (geo. & sem.). The fourth group distills depth knowledge (depth KD). + +
SettingPSNR ↑SSIM ↑LPIPS ↓
SRF [12]P12.340.2500.591
PixelNeRF [64]7.930.2720.682
MVSNeRF [5]17.250.5570.356
SRF ft [12]P&FT17.070.4360.529
PixelNeRF ft [64]16.170.4380.512
MVSNeRF ft [5]17.880.5840.327
Mip-NeRF [1]geo. & sem.14.620.3510.495
DietNeRF [25]14.940.3700.496
RegNeRF [37]19.080.5870.336
MonoSDF*[65]depth KD18.450.5650.388
DSNeRF[17]18.940.5820.362
SparseNeRF (Ours)19.860.6240.328
+ +# 5.2. Comparisons on DTU + +Similar to the LLFF dataset, we conduct four-group comparisons on the DTU dataset, as shown in Table 2. In this first group, given a target scene, SRF, PixelNeRF, and MVSNeRF are pre-trained on other DTU scenes and are + +Table 2: Quantitative Comparison on DTU with three views. The first group is pre-trained on other scenes and tested on a target scene (P). The second group requires further finetuning on a target scene (P&FT). The third group uses geometry and semantic constraints (geo. & sem.). The fourth group distills depth knowledge (depth KD). + +
SettingPSNR ↑SSIM ↑LPIPS ↓
SRF [12]P15.320.6710.304
PixelNeRF [64]16.820.6950.270
MVSNeRF [5]18.630.7690.197
SRF ft [12]P&FT15.680.6980.281
PixelNeRF ft [64]18.950.7100.269
MVSNeRF ft [5]18.540.7690.197
Mip-NeRF [1]geo. & sema.8.680.5710.353
DietNeRF [25]11.850.6330.314
RegNeRF [37]18.890.7450.190
MonoSDF* [65]depth KD18.920.7480.237
DSNeRF [17]16.900.5700.450
SparseNeRF(Ours)19.550.7690.201
+ +tested on the target scene. In the second group, these three methods are further fine-tuned on the target scene. In the third group, we compare Mip-NeRF, DietNeRF, and RegN-eRF. Note that for the pre-trained models, they split DTU as a training set (88 scenes) and a test set (15 scenes). Since + +![](images/3a8cc286876affb436e5ae62b0bbbd183d09075e72c3231937f127196a7c006f.jpg) +DSNeRF + +![](images/d7fe4728de2a19677aaa89656c081ff72e73158bcc825576d83c99808c0a876b.jpg) +MonoSDF* + +![](images/92f98aa96d487373d69a8c10c4d586ea005391f268ea9e395baaaf3d158153b7.jpg) +RegNeRF + +![](images/10731e1d75e132f846ea79fcf79ea2d4fbf94ff214848cb2ed031add2b972680.jpg) +SparseNeRF (ours) + +![](images/ca91274958313cceb7c430db55c5c06db6cbb380cff48b9796002363f91a4366.jpg) +Ground Truth + +![](images/c20aff936b70546755ac1eeabfeda8cd8d180983e3335ea32bcf648c87b074b5.jpg) +Figure 7: Visual comparisons on the NVS-RGBD dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes. + +![](images/d4d7b76290cfe97e0b20019380f936dc13697917fcb2d43cc6d7836c25427a2c.jpg) + +![](images/049b3dc06ceb3dac98067a55eb03628ae3315477afa1fb3348d197eafda66a1f.jpg) + +![](images/d75e48e88540eaea2f8ef21b59a760ee1cf17f6e8b0765d5e184fa225f62146b.jpg) + +![](images/d46d68af4155b95e5f44bb524ec21ef1af188df22ad6ed7fae444e70e5dc4a62.jpg) + +![](images/8652c6100c9f4481ae8e329de3b0fa81411815f7627378abe711e2e98cac5648.jpg) +w/o conti. & ranking +Figure 8: Ablation study on visual effect. The three columns denote "w/o continuity and ranking distillation", "w/o continuity distillation", and "w/ continuity and ranking distillation" (SparseNeRF, ours), respectively. + +![](images/5f5c61f6a57a90e41dc7ea4c5443bf71e6fd0bb32c0e06afc2d76fcd5cb0152a.jpg) +w/o ranking + +![](images/888a9648beb336f6626ac7c988d861dc084742a5075cb25eab4e83c486aa8a12.jpg) +w/o conti. + +![](images/bac4f1bc85c90cf5283877288d3530b14e8eaa71b23af78ed128d10dab29640d.jpg) +SparseNeRF(ours) + +the training set and the test set contains similar scenes, a test scene shares a similar distribution (e.g., similar backgrounds) with training scenes. Pre-training on other scenes greatly benefits a target scene. Therefore, SRF, PixelNeRF, and MVSNeRF also greatly achieve promising results. Finally, we compare depth-based DSNeRF [17] and MonoSDF [65]. DSNeRF studies two kinds of depth information, i.e., sparse 3D points generated by COLMAP [42] and ground-truth depth maps obtained by high-accuracy depth scanners and the Multi-View Stereo (MVS) algorithm. The depth maps of NVS-RGBD are captured by consumer-level sensors and contain lots of noise, which is hard to be applied to DSNeRF. Therefore, we adopt sparse 3D points for DSNeRF. Compared with these methods, SparseNeRF still achieves the best performance without pre-training on other scenes. We provide a qualitative analysis, as shown in Figure 6. SparseNeRF achieves better visual results than previous state-of-the-art methods. + +# 5.3. Comparisons on NVS-RGBD + +We then implement three state-of-the-art methods on our new NVS-RGBD dataset, including RegNeRF [37], DSNeRF [17], and MonoSDF [65]. All of them are optimized per scene without pre-training on the other scenes. RegNeRF adopted continuous geometry assumptions from unobserved viewpoints. DSNeRF uses a global scale-invariant depth constraint to supervise NeRFs. MonoSDF mainly focuses on implicit surface reconstruction. It is as- + +Table 3: Comparison with State of the Arts on NVS-RGBD. * denotes a new implementation from related tasks. + +
MetricsPSNR↑SSIM↑LPIPS↓depth err↓
KinectRegNeRF [37]25.780.8400.2425.9×10-3
DSNeRF [17]25.910.8380.2384.8×10-3
MonoSDF* [65]25.500.8380.2395.5×10-3
SparseNeRF(ours)26.280.8500.2323.9×10-3
ZED 2RegNeRF [37]24.980.7810.2713.5×10-3
DSNeRF [17]24.970.7890.2623.1×10-3
MonoSDF* [65]24.540.7840.2753.2×10-3
SparseNeRF(ours)26.220.8040.2622.4×10-3
+ +sumed that local patches of depth maps from monocular cameras satisfy $d' = w d + b$ where $d$ is predicted by the pre-trained model and $d'$ is then computed by the NeRF. In Table 3, we observe that SparseNeRF achieves better performance compared with these prior arts. We provide a qualitative analysis in Figure 7, demonstrating the superiority of SparseNeRF over previous state-of-the-art methods. + +Table 4: Comparison of Depth Ranking and Depth scaling. + +
LLFFPSNR↑SSIM↑LPIPS↓depth err↓
baseline19.080.5870.33610.1×10-3
w/ depth scaling18.450.5650.3888.9×10-3
w/ depth ranking19.860.6240.3286.3×10-3
+ +# 5.4. Ablation Studies and Further Analyses + +Effectiveness of Depth Distillation. To validate the effectiveness of the depth distillation, we isolate the depth constraint and keep other modules unchanged. As shown in the first and third rows of Table 4, it is observed that without the local depth ranking distillation, SparseNeRF degrades on PSNR, SSIM, LPIPS, and depth error, which demonstrates the effectiveness of the depth distillation. We further study the depth maps from unobserved views. The ablation study of visual effect is shown in Figure 8. It shows that depth ranking greatly improves performance in geometry. Spatial continuity regularization further improves the detailed coherence of scenes. + +Table 5: Impact of different pre-trained depth models. + +
LLFFPSNR↑SSIM↑LPIPS↓depth err↓
Baseline19.080.5870.33610.1×10-3
MiDaS small19.350.5920.3336.9×10-3
DPT Hybrid19.860.6250.3266.3×10-3
DPT Large19.860.6240.3286.3×10-3
+ +Further Analyses on Depth Ranking and Depth Scaling. To validate the robustness of local depth ranking, we compare depth ranking regularization and depth scaling regularization on LLFF. We use the idea from MonoSDF [65], which mainly focuses on implicit surface reconstruction. With the specific design for surface shapes, MonoSDF slightly degrades the performance of RGB view synthesis. We re-implement it as depth scaling. Since the single-view depth estimation is not accurate, the linear depth scaling constraint is too strong such that depth distillation misleads the learning of the NeRF. Figure 3 shows two cases of scale-invariant errors. Instead, depth ranking relaxes the constraint and only focuses on depth comparison, which is more robust than depth scaling. As shown in Table 4, depth ranking performs better than depth scaling on LLFF. + +Impact of Different Pre-trained Depth Models. We test our method on three depth estimation models, i.e., MiDaS small, DPT Hybrid, and DPT Large. As shown in Table 5, all of the depth models are better than the baseline. DPT Hybrid and DPT Large achieve comparable results, which are better than MiDaS small. + +FreeNeRF [61] potentially complements SparseNeRF. FreeNeRF and SparseNeRF improve few-shot NeRF in different ways. FreeNeRF introduces a frequency regularization while SparseNeRF distills depth priors from pretrained general depth estimators. We implement SparseNeRF on FreeNeRF and the results significantly improve, as shown in Table 6. + +Table 6: FreeNeRF potentially complement SparseNeRF. + +
LLFFPSNR↑SSIM↑LPIPS↓
FreeNeRF19.600.6140.302
FreeNeRF+our depth distillation20.260.6460.292
+ +User Study. To show that the proposed spatial continuity regularization improves the 3D consistency in 3D space, we conducted a user study. For each scene, we rendered videos for both with and without spatial continuity regularization terms. Users are asked to select the better-quality video according to 3D consistency/coherence and completeness (good geometry, few floaters). We evaluate 8 scenes on LLFF and there are 27 participants. The result shows $88.9\%$ of users prefer "w/ conti" and $11.1\%$ prefer "w/o conti". To evaluate the advantages of SparseNeRF on 3D consistency and shape completeness, we conduct another user study. Users are asked to evaluate "baseline+RegNeRF's continu + +ity term” and “baseline+our continuity term”. The 27 users participated in the study and 8 scenes on LLFF are evaluated. The result shows $87.5\%$ of users prefer “baseline+our continuity term” and $12.5\%$ prefer “baseline+RegNeRF’s continuity term”. RegNeRF encourages the overall patch to be as smooth as possible without considering the sharp edges of objects, which conflicts with the reconstruction from sparse views. Our continuity regulation distills the neighbor relationship from depth models is more accurate. We show two examples of depth maps in Figure 9. + +![](images/e87f7feacc4d162237d85a5348fea56ffb91e41ab9541d1490394fa5f2a2f199.jpg) +Figure 9: Visual comparison on continuity regulation. + +Discussion. SparseNeRF could have a wide range of applications due to its promising results. We provide two applications on iPhone 13 Pro. Please refer to the project page for more experiments. Like UNISURF [38, 65], we also evaluate the SparseNeRF using Chamfer distance, showing the competitive results with MonoSDF and UNISURF, especially for complex scenes. + +# 6. Conclusion + +In this paper, we propose a SparseNeRF framework that synthesizes novel views with sparse view inputs. To tackle the under-constrained few-shot NeRF problem, we propose a local depth ranking regularization that distills the depth ranking prior from coarse depth maps. To preserve the spatial continuity of geometry, we propose a spatial continuity regularization that distills the depth continuity priors. With the proposed depth priors, SparseNeRF achieves a new state-of-the-art performance on three datasets. + +Limitation. SparseNeRF significantly improves the performance of few-shot NeRF, but cannot be generalized to occluded views that are unobserved in the training views. + +Potential Negative Impact. SparseNeRF focuses on scene synthesis. It might be misused to create misleading content. Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is also supported by Singapore MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20221-0011) and NTU NAP Grant. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srini + +vasan. Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In Int. Conf. Comput. Vis., 2021. 1, 2, 4, 6, 7 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1 +[3] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: Composable transformations of Python+NumPy programs, 2018. 6 +[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3 +[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MvsNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In Int. Conf. Comput. Vis., 2021. 2, 3, 6, 7 +[6] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Adv. Neural Inform. Process. Syst., 2016. 3 +[7] Xipeng Chen, Guangrun Wang, Dizhong Zhu, Xiaodan Liang, Philip Torr, and Liang Lin. Structure-preserving 3D garment modeling with neural sewing machines. Adv. Neural Inform. Process. Syst., 2022. 3 +[8] Zixuan Chen, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. CuNeRF: Cube-based neural radiance field for zero-shot medical image arbitrary-scale super resolution. Int. Conf. Comput. Vis., 2023. 2 +[9] Zhaoxi Chen and Ziwei Liu. Relighting4D: Neural re- ligable human from videos. In Eur. Conf. Comput. Vis., 2022.3 +[10] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Trans. Graph., 41(6):1-16, 2022. 3 +[11] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Scenedreamer: Unbounded 3D scene generation from 2D image collections. In arXiv, 2023. 3 +[12] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 6, 7 +[13] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 2, 3 +[14] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv:1602.02481, 2016. 6 +[15] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6 +[16] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. + +NerDI: Single-view NeRF synthesis with language-guided diffusion as general image priors. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3 +[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 3, 4, 6, 7, 8 +[18] Chenhe Dong, Guangrun Wang, Hang Xu, Jiefeng Peng, Xiaozhe Ren, and Xiaodan Liang. Efficientbert: Progressively searching multilayer perceptron via warm-up knowledge distillation. arXiv preprint arXiv:2109.07222, 2021. 2 +[19] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., June 2022. 2 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst., 2020. 3 +[22] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. EVA3D: Compositional 3D human generation from 2D image collections. arXiv preprint arXiv:2210.04888, 2022. 3 +[23] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human NeRF from a single image. arXiv preprint arXiv:2303.12791, 2023. 3 +[24] Shoukang Hu, Kaichen Zhou, Kaiyu Li, Longhui Yu, Lanqing Hong, Tianyang Hu, Zhenguo Li, Gim Hee Lee, and Ziwei Liu. ConsistentNeRF: Enhancing neural radiance fields with 3D consistency for sparse view synthesis. arXiv preprint arXiv:2305.11031, 2023. 2 +[25] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a diet: Semantically consistent few-shot view synthesis. In Int. Conf. Comput. Vis., 2021. 2, 6, 7 +[26] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In IEEE Conf. Comput. Vis. Pattern Recog., 2014. 6 +[27] Mijeong Kim, Seonguk Seo, and Bohyung Han. InfoNeRF: Ray entropy minimization for few-shot neural volume rendering. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2, 3, 4 +[28] Changlin Li, Bohan Zhuang, Guangrun Wang, Xiaodan Liang, Xiaojun Chang, and Yi Yang. Automated progressive learning for efficient training of vision transformers. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[29] Xiaoming Li, Wangmeng Zuo, and Chen Change Loy. Learning generative structure prior for blind text image super-resolution. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3 +[30] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3 + +[31] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Int. Conf. Comput. Vis., 2021. 2 +[32] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Adv. Neural Inform. Process. Syst., 2020. 1 +[33] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. GNeRF: GAN-based neural radiance field without posed camera. In Int. Conf. Comput. Vis., 2021. 2 +[34] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In IEEE Conf. Comput. Vis. Pattern Recog., 2015. 3 +[35] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Trans. Graph., 38(4):1-14, 2019. 6 +[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis. Springer, 2020. 1, 2, 3 +[37] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg-NeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 3, 4, 6, 7, 8 +[38] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Int. Conf. Comput. Vis., 2021. 9 +[39] Jiefeng Peng, Jiqi Zhang, Changlin Li, Guangrun Wang, Xiaodan Liang, and Liang Lin. Pi-nas: Improving neural architecture search by reducing supernet training consistency shift. In Int. Conf. Comput. Vis., 2021. 2 +[40] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3 +[41] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Int. Conf. Comput. Vis., 2021. 2, 4, 5 +[42] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 3, 8 +[43] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Eur. Conf. Comput. Vis. Springer, 2012. 3 +[44] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Adv. Neural Inform. Process. Syst., 2021. 1 +[45] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6 + +[46] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2012. 3 +[47] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-NeRF: Scalable large scene neural view synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[48] Tang Tao, Longfei Gao, Guangrun Wang, Peng Chen, Dayang Hao, Xiaodan Liang, Mathieu Salzmann, and Kaicheng Yu. LiDAR-NeRF: Novel lidar view synthesis via neural radiance fields. arXiv preprint arXiv:2304.10406, 2023. 2 +[49] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-NeRF: Scalable construction of large-scale NeRFs for virtual fly-throughs. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[50] Guangrun Wang, Ping Luo, Xinjiang Wang, Liang Lin, et al. Kalman normalization: Normalizing internal representations across network layers. Adv. Neural Inform. Process. Syst., 2018. 2 +[51] Guangrun Wang and Philip HS Torr. Traditional classification neural networks are good generators: They are competitive with DDPMs and GANs. arXiv preprint arXiv:2211.14794, 2022.3 +[52] Guangrun Wang, Guangcong Wang, Keze Wang, Xiaodan Liang, and Liang Lin. Grammatically recognizing images with tree convolution. 2020. 2 +[53] Guangrun Wang, Keze Wang, and Liang Lin. Adaptively connected neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2 +[54] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In Eur. Conf. Comput. Vis., 2022. 3 +[55] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 6 +[56] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2 +[57] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3 +[58] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. CityNeRF: Building nerf at city scale. arXiv preprint arXiv:2112.05504, 2021. 2 +[59] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. SinNeRF: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928, 2022. 3 +[60] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild + +2D photo to a 3D object with 360deg views. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3 +[61] Jiawei Yang, Marco Pavone, and Yue Wang. FreeNeRF: Improving few-shot neural rendering with free frequency regularization. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 9 +[62] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Vtoonify: Controllable high-resolution portrait video style transfer. ACM Trans. Graph., 41(6):1-15, 2022. 3 +[63] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Gp-unit: Generative prior for versatile unsupervised image-to-image translation. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 3 +[64] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 3, 6, 7 +[65] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 1, 2, 3, 4, 6, 7, 8, 9 +[66] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6 +[67] Long Zhuo, Guangcong Wang, Shikai Li, Wanye Wu, and Ziwei Liu. Fast-vid2vid: Spatial-temporal compression for video-to-video synthesis. In Eur. Conf. Comput. Vis., 2022. 3 \ No newline at end of file diff --git a/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/images.zip b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2394b74d1f0a5ef2efb7788070a43e0836728d81 --- /dev/null +++ b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ce756cb7522a419dc4335283d77fd13a01850428e87bbf39b08261485b67d7a +size 905575 diff --git a/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/layout.json b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..92d1a2c4bb9a47c192f5a4237c93b9de7468301f --- /dev/null +++ b/2023/SparseNeRF_ Distilling Depth Ranking for Few-shot Novel View Synthesis/layout.json @@ -0,0 +1,10413 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 525, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 525, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 525, + 121 + ], + "type": "text", + "content": "SparseNeRF: Distilling Depth Ranking for Few-shot Novel View Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 144, + 190, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 144, + 190, + 158 + ], + "spans": [ + { + "bbox": [ + 102, + 144, + 190, + 158 + ], + "type": "text", + "content": "Guangcong Wang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 144, + 284, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 144, + 284, + 156 + ], + "spans": [ + { + "bbox": [ + 220, + 144, + 284, + 156 + ], + "type": "text", + "content": "Zhaoxi Chen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 144, + 403, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 144, + 403, + 157 + ], + "spans": [ + { + "bbox": [ + 315, + 144, + 403, + 157 + ], + "type": "text", + "content": "Chen Change Loy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 158, + 399, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 158, + 399, + 171 + ], + "spans": [ + { + "bbox": [ + 195, + 158, + 399, + 171 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 433, + 144, + 491, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 433, + 144, + 491, + 156 + ], + "spans": [ + { + "bbox": [ + 433, + 144, + 491, + 156 + ], + "type": "text", + "content": "Ziwei Liu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 157, + 174, + 436, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 174, + 436, + 185 + ], + "spans": [ + { + "bbox": [ + 157, + 174, + 436, + 185 + ], + "type": "text", + "content": "{guangcong.wang,zhaoxi00,ccloy,ziwei.liu}@ntu.edu.sg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 51, + 201, + 145, + 256 + ], + "blocks": [ + { + "bbox": [ + 51, + 201, + 145, + 256 + ], + "lines": [ + { + "bbox": [ + 51, + 201, + 145, + 256 + ], + "spans": [ + { + "bbox": [ + 51, + 201, + 145, + 256 + ], + "type": "image", + "image_path": "25c38c11645eed074b62c9cd251e49a21ca4b5feb2027f7bacbcf93d5ee02a47.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 51, + 257, + 145, + 311 + ], + "blocks": [ + { + "bbox": [ + 51, + 257, + 145, + 311 + ], + "lines": [ + { + "bbox": [ + 51, + 257, + 145, + 311 + ], + "spans": [ + { + "bbox": [ + 51, + 257, + 145, + 311 + ], + "type": "image", + "image_path": "c000f477d6f0062339011af44cea59e90cc1d9c297bd721e449fa2ee9a2af937.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 51, + 312, + 145, + 365 + ], + "blocks": [ + { + "bbox": [ + 51, + 312, + 145, + 365 + ], + "lines": [ + { + "bbox": [ + 51, + 312, + 145, + 365 + ], + "spans": [ + { + "bbox": [ + 51, + 312, + 145, + 365 + ], + "type": "image", + "image_path": "3525d8a67f84551639343e0c7d35420e43791c3fc381f519533a4d755ac8ba24.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 367, + 134, + 378 + ], + "lines": [ + { + "bbox": [ + 67, + 367, + 134, + 378 + ], + "spans": [ + { + "bbox": [ + 67, + 367, + 134, + 378 + ], + "type": "text", + "content": "Novel view (GT)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 150, + 201, + 246, + 256 + ], + "blocks": [ + { + "bbox": [ + 150, + 201, + 246, + 256 + ], + "lines": [ + { + "bbox": [ + 150, + 201, + 246, + 256 + ], + "spans": [ + { + "bbox": [ + 150, + 201, + 246, + 256 + ], + "type": "image", + "image_path": "7ecff797f74986a8d24f98fbc82ca52102ea24122b0fcb1306b7f5be77c5a642.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 150, + 257, + 245, + 311 + ], + "blocks": [ + { + "bbox": [ + 150, + 257, + 245, + 311 + ], + "lines": [ + { + "bbox": [ + 150, + 257, + 245, + 311 + ], + "spans": [ + { + "bbox": [ + 150, + 257, + 245, + 311 + ], + "type": "image", + "image_path": "28f91706afef13fe42b5b4fff2f36c690bddd7e94960c5ed90829a4c159bc04f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 150, + 312, + 245, + 366 + ], + "blocks": [ + { + "bbox": [ + 150, + 312, + 245, + 366 + ], + "lines": [ + { + "bbox": [ + 150, + 312, + 245, + 366 + ], + "spans": [ + { + "bbox": [ + 150, + 312, + 245, + 366 + ], + "type": "image", + "image_path": "00557e9e82b0069ff9eaa39ce225d64f476f634688ee1c927bb85482624c8a08.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 367, + 219, + 378 + ], + "lines": [ + { + "bbox": [ + 180, + 367, + 219, + 378 + ], + "spans": [ + { + "bbox": [ + 180, + 367, + 219, + 378 + ], + "type": "text", + "content": "RegNeRF" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 250, + 201, + 345, + 256 + ], + "blocks": [ + { + "bbox": [ + 250, + 201, + 345, + 256 + ], + "lines": [ + { + "bbox": [ + 250, + 201, + 345, + 256 + ], + "spans": [ + { + "bbox": [ + 250, + 201, + 345, + 256 + ], + "type": "image", + "image_path": "64a3f2fecffe8f95e59346a101a75fb57de9e06c4153c47363fce93c651ce5f9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 250, + 257, + 345, + 311 + ], + "blocks": [ + { + "bbox": [ + 250, + 257, + 345, + 311 + ], + "lines": [ + { + "bbox": [ + 250, + 257, + 345, + 311 + ], + "spans": [ + { + "bbox": [ + 250, + 257, + 345, + 311 + ], + "type": "image", + "image_path": "b1841394780b111d42dcab0a5beb53c863612ac147868dd0101632247db5cde5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 250, + 312, + 345, + 365 + ], + "blocks": [ + { + "bbox": [ + 250, + 312, + 345, + 365 + ], + "lines": [ + { + "bbox": [ + 250, + 312, + 345, + 365 + ], + "spans": [ + { + "bbox": [ + 250, + 312, + 345, + 365 + ], + "type": "image", + "image_path": "72e5378b92d1d159efc28a474aa139d38d2622e5dcf200a318e81d23c6005967.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 367, + 321, + 376 + ], + "lines": [ + { + "bbox": [ + 276, + 367, + 321, + 376 + ], + "spans": [ + { + "bbox": [ + 276, + 367, + 321, + 376 + ], + "type": "text", + "content": "MonoSDF*" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 387, + 547, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 387, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 547, + 448 + ], + "type": "text", + "content": "Figure 1: Comparing the state-of-the-arts RegNeRF [37], MonoSDF* [65], DSNeRF [17], and our SparseNeRF with three views for training (* denotes a re-implementation for our task). RegNeRF regularizes geometry with sparsity and continuity constraints. MonoSDF, and DSNeRF use scale-invariant depth constraints supervised by coarse depth maps. Our SparseNeRF uses robust depth ranking of coarse depth maps. SparseNeRF can synthesize realistic novel views and coherent geometric depth (Please refer to the supplementary material for rendered videos)." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 349, + 201, + 444, + 256 + ], + "blocks": [ + { + "bbox": [ + 349, + 201, + 444, + 256 + ], + "lines": [ + { + "bbox": [ + 349, + 201, + 444, + 256 + ], + "spans": [ + { + "bbox": [ + 349, + 201, + 444, + 256 + ], + "type": "image", + "image_path": "2533c19f69afba9e908ae86e9e2ae45eb85342aaa2d47b82e371b615bdb4f19d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 349, + 257, + 444, + 311 + ], + "blocks": [ + { + "bbox": [ + 349, + 257, + 444, + 311 + ], + "lines": [ + { + "bbox": [ + 349, + 257, + 444, + 311 + ], + "spans": [ + { + "bbox": [ + 349, + 257, + 444, + 311 + ], + "type": "image", + "image_path": "97fe940b37493e134d7c866bb59289c2d1795a559e98eb870ffd908bea07a231.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 349, + 312, + 444, + 365 + ], + "blocks": [ + { + "bbox": [ + 349, + 312, + 444, + 365 + ], + "lines": [ + { + "bbox": [ + 349, + 312, + 444, + 365 + ], + "spans": [ + { + "bbox": [ + 349, + 312, + 444, + 365 + ], + "type": "image", + "image_path": "897fa3094c9469d8eb57d64785200bc4051bd287cbc097e259f59f0239dbf5c9.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 367, + 417, + 376 + ], + "lines": [ + { + "bbox": [ + 381, + 367, + 417, + 376 + ], + "spans": [ + { + "bbox": [ + 381, + 367, + 417, + 376 + ], + "type": "text", + "content": "DSNeRF" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 449, + 201, + 543, + 256 + ], + "blocks": [ + { + "bbox": [ + 449, + 201, + 543, + 256 + ], + "lines": [ + { + "bbox": [ + 449, + 201, + 543, + 256 + ], + "spans": [ + { + "bbox": [ + 449, + 201, + 543, + 256 + ], + "type": "image", + "image_path": "ac1b1ff45e41977f79c4b1e24c12922dc34e7d801aae9da99a16a20fde31f722.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 449, + 257, + 543, + 311 + ], + "blocks": [ + { + "bbox": [ + 449, + 257, + 543, + 311 + ], + "lines": [ + { + "bbox": [ + 449, + 257, + 543, + 311 + ], + "spans": [ + { + "bbox": [ + 449, + 257, + 543, + 311 + ], + "type": "image", + "image_path": "129fb981b22218603b81438cf257136b573938aaa5575d6a9a7f763d1feb7ff0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 449, + 312, + 543, + 365 + ], + "blocks": [ + { + "bbox": [ + 449, + 312, + 543, + 365 + ], + "lines": [ + { + "bbox": [ + 449, + 312, + 543, + 365 + ], + "spans": [ + { + "bbox": [ + 449, + 312, + 543, + 365 + ], + "type": "image", + "image_path": "2dcadd5a88e5ff3eb09d5d4865a81e4f90813db6baf7442b71da0bd2b74d3a25.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 367, + 533, + 378 + ], + "lines": [ + { + "bbox": [ + 460, + 367, + 533, + 378 + ], + "spans": [ + { + "bbox": [ + 460, + 367, + 533, + 378 + ], + "type": "text", + "content": "SparseNeRF (ours)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 143, + 458, + 192, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 458, + 192, + 470 + ], + "spans": [ + { + "bbox": [ + 143, + 458, + 192, + 470 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "Neural Radiance Field (NeRF) significantly degrades when only a limited number of views are available. To complement the lack of 3D information, depth-based models, such as DSNeRF and MonoSDF, explicitly assume the availability of accurate depth maps of multiple views. They linearly scale the accurate depth maps as supervision to guide the predicted depth of few-shot NeRFs. However, accurate depth maps are difficult and expensive to capture due to wide-range depth distances in the wild. This work presents a new Sparse-view NeRF (SparseNeRF) framework that exploits depth priors from real-world inaccurate observations. The inaccurate depth observations are either from pre-trained depth models or coarse depth maps of consumer-level depth sensors. Since coarse depth maps are not strictly scaled to the ground-truth depth maps, we propose a simple yet effective constraint, a local depth ranking method, on NeRFs such that the expected depth ranking of the NeRF is consistent with that of the coarse depth maps in local patches. To preserve the spatial continuity of the esti" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 459, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 604 + ], + "type": "text", + "content": "mated depth of NeRF, we further propose a spatial continuity constraint to encourage the consistency of the expected depth continuity of NeRF with coarse depth maps. Surprisingly, with simple depth ranking constraints, SparseNeRF outperforms all state-of-the-art few-shot NeRF methods (including depth-based models) on standard LLFF and DTU datasets. Moreover, we collect a new dataset NVS-RGBD that contains real-world depth maps from Azure Kinect, ZED 2, and iPhone 13 Pro. Extensive experiments on NVS-RGBD dataset also validate the superiority and generalizability of SparseNeRF. Code and dataset are available at https://sparsenerf.github.io/." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 632, + 387, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 387, + 643 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 387, + 643 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Neural radiance fields (NeRFs) [36, 32, 1, 2, 44] have made tremendous progress in generating photo-realistic novel views of scenes by optimizing implicit function representations given a set of 2D input views. However, in a wide range of real-world scenarios, collecting dense views of a" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "9065" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "type": "text", + "content": "scene is often expensive and time-consuming [47]. Therefore, it is necessary to develop few-shot NeRF methods that can be learned from sparse views without significant degradation in performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 121, + 286, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 121, + 286, + 530 + ], + "spans": [ + { + "bbox": [ + 49, + 121, + 286, + 530 + ], + "type": "text", + "content": "Learning a NeRF from sparse views is a challenging problem due to under-constrained reconstruction conditions, especially in textureless areas. Directly applying NeRFs to few-shot scenarios suffers from dramatic degradation [37]. Recently, some methods have greatly improved the performance of few-shot NeRF, which can be categorized into three groups. 1) The first group [37, 27, 25] is based on geometry constraints (sparsity and continuity regularizations) and high-level semantics. RegNeRF [37] regularized the geometry and appearance of patches rendered from unobserved viewpoints, and annealed the ray sampling space. InfoNeRF [27] imposed an entropy constraint of the density in each ray and a spatial smoothness constraint into the estimated images. However, since a scene often contains multiple layouts (Figure 1), sparsity and continuity geometric constraints of a few views cannot guarantee the complete 3D geometric reconstruction. 2) The second group [64, 5] resorts to pre-training on similar scenes. For example, PixelNeRF [64] proposed to condition a NeRF on convolutional feature maps to learn high-level semantics from other scenes. 3) The third group [17, 65, 24] exploits depth maps and makes a linearity assumption of depth maps to supervise few-shot NeRFs. For example, DSNeRF [17] exploited sparse 3D points generated by COLMAP [42] or accurate depth maps [13] obtained by high-accuracy depth scanners and the Multi-View Stereo (MVS) algorithm. The depth maps are linearly scaled as supervision to guide the predicted depth of few-shot NeRFs. To use coarse depth maps, MonoSDF [65] uses a local patch-based scale-invariant depth constraint supervised by coarse depth maps instead of global depth maps. However, the scale-invariant depth constraint is strong for real-world coarse depth maps from pre-trained depth models or consumer-level depth sensors due to wide-range depth distances in the wild." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 534, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 534, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 286, + 712 + ], + "type": "text", + "content": "Along the third group, we wish to explore more robust 3D priors from coarse depth maps to complement the under-constrained few-shot NeRF. To address this problem, we present SparseNeRF, a simple yet effective method that distills depth priors from pre-trained depth models [41] or inaccurate depth maps from consumer-level depth sensors (Figure 3), which can be easily obtained from real-world scenes. Deriving useful depth cues from such pre-trained models is non-trivial. In particular, although single-view depth estimation methods have achieved good visual performance, thanks to large-scale monocular depth datasets and large ViT models, they cannot yield accurate 3D depth information due to coarse depth annotations, dataset bias, and ill-posed 2D single-view images. The inaccurate depth information contradicts the density prediction of a NeRF" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 308, + 72, + 544, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 119 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 119 + ], + "type": "text", + "content": "when reconstructing each pixel of a 3D scene based on volume rendering. Directly scaling the coarse depth maps to a NeRF [17, 65] leads to inconsistent geometry against the expected depth of the NeRF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 121, + 544, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 121, + 544, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 121, + 544, + 346 + ], + "type": "text", + "content": "Instead of directly supervising a NeRF with coarse depth priors, we relax hard depth constraints [17, 65] and distill robust local depth ranking from the coarse depth maps to a NeRF such that the depth ranking of a NeRF is consistent with that of coarse depth. That is, we supervise a NeRF with relative depth instead of absolute depth [17, 65]. To guarantee the spatial continuity of geometry, we further propose a spatial continuity constraint on depth maps such that the NeRF model imitates the spatial continuity of coarse depth maps. The accurate sparse geometry constraints from a limited number of views, combined with relaxed constraints including depth ranking regularization and continuity regularization, finally achieve promising novel view synthesis (Figure 1). It is noteworthy that SparseNeRF does not increase the running time during inference as it only exploits depth priors from pre-trained depth models or consumer-level sensors during the training stage (Figure 2). In addition, SparseNeRF is a plug-and-play module that can be easily integrated into various NeRFs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 347, + 544, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 347, + 544, + 562 + ], + "spans": [ + { + "bbox": [ + 308, + 347, + 544, + 562 + ], + "type": "text", + "content": "The main contributions of this paper are 1) SparseNeRF, a simple yet effective method that distills local depth ranking priors from pre-trained depth models. With the help of the local depth ranking constraint, SparseNeRF significantly improves the performance of few-shot novel view synthesis over the state-of-the-art models (including depth-based NeRF methods). To preserve the coherent geometry of a scene, we propose a spatial continuity distillation constraint that encourages the spatial continuity of NeRF to be similar to that of the pre-trained depth model. Both depth ranking prior and spatial continuity distillation are new in the literature on NeRF. 2) Apart from SparseNeRF, we also contribute a new dataset, NVS-RGBD, which contains coarse depth maps from Azure Kinect, ZED 2, and iPhone 13 Pro. 3) Extensive experiments on the LLFF, DTU, and NVS-RGBD datasets demonstrate that SparseNeRF achieves a new state-of-the-art performance in few-shot novel view synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 574, + 391, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 574, + 391, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 574, + 391, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 594, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 594, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 594, + 544, + 712 + ], + "type": "text", + "content": "Neural Radiance Fields. NeRF [36, 48, 8, 39] has made great success in synthesizing novel views of complex scenes due to good representation of neural networks [20, 52, 50, 53, 18, 28]. Block-NeRF [47], CityNeRF [58], and Mega-NeRF [49] scaled the standard NeRF up to city-scale or urban-scale scenes. NeRF—[56], GNeRF [33], and BARF [31] relaxed the requirements of NeRFs and synthesized novel views without perfect camera poses. Some works attempted to improve NeRFs by considering antialiasing [1], sparse 3D grids with spherical harmonics [19]." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 741 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 741 + ], + "type": "text", + "content": "9066" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 133 + ], + "type": "text", + "content": "Some methods [40, 30, 57] extended NeRFs to dynamic scenes. These methods do not focus on generating novel views with a few views. In this paper, we study sparse-view NeRF to reduce dense capture requirements in real-world applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 163, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 289, + 715 + ], + "type": "text", + "content": "Few-shot Novel View Synthesis. There are increasing studies on few-shot novel view synthesis. Basically, the existing few-shot NeRF methods can be categorized into three groups. First, some methods exploit continuity constraints on geometry or object semantics. For example, RegNeRF [37] imposed a continuity constraint on geometry and regularized the appearance of patches from unobserved viewpoints with a flow model. InfoNeRF [27] proposed a ray entropy minimization regularization to encourage the density to be as sparse as possible along a ray, and used a ray information gain reduction regularization to constrain the continuous depth of neighbor rays. Second, some methods attempt to pe-train a NeRF on other similar scenes and fine-tune the NeRF on the target scene. For example, PixelNeRF [64] conditioned a NeRF on image inputs in a fully convolutional manner, allowing the model to learn scene priors from other scenes and reduce the requirement of dense views. MVSNeRF [5] leveraged plane-swept cost volumes for geometry-aware scene reasoning and combined it with physically based volume rendering. Similar to PixelNeRF, MVSNeRF was first trained on other real scenes and was finetuned on target scenes to evaluate its effectiveness and generalizability. Third, depth-based models [17, 65] use available depth information to supervise the training of NeRFs. DSNeRF [17] exploited sparse 3D points generated by COLMAP [42] or accurate depth maps [13] obtained by high-accuracy depth scanners. Different from these depth-based models, SparseNeRF distills robust depth ranking from pre-trained depth models or coarse depth maps from consumer-level sensors. As concurrent works, NeRDi [16] and NeuralLift-360 [60] also use ranking-based methods. However, they mainly focus on the single-view setting while SparseNeRF focuses on sparse-view. SparseNeRF uses ranking loss to avoid inconsistent 3D geometry across different views while [16, 60] focuses on a soft geometric regularization on a single view (no cross-view consistency problem during training). Moreover, SparseNeRF introduces a new spatial continuity loss to distill spatial coherence from monocular depth estimators. In addition, some single-view synthesis methods [59] either allow new generative objects from unseen views or focus on specific objects, e.g., face [4] and human [23]. Some NeRF-based methods [22, 11] extended to generation instead of reconstruction. Other AIGC models [21, 51, 54, 67, 10, 9, 63, 29, 62] focus on the 2D generation or category-specific reconstruction [7], which are different." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 71, + 397, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 397, + 86 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 397, + 86 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 91, + 547, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 282 + ], + "type": "text", + "content": "We present SparseNeRF to synthesize novel views given sparse view inputs. Single-view depth estimation is a long-standing computer vision task, aiming to predict a depth map given a single image. In this paper, we are interested in mining the depth priors encapsulated in pre-trained models of single-view depth estimation or coarse depth maps captured by consumer-level depth sensors. However, due to coarse annotations of single-view depth maps, (e.g., user clicks [6], RGB-D [43, 46], and laser/stereo [34]), dataset bias, and imperfect depth estimation models, it is challenging to obtain accurate 3D depth estimation given 2D single-view images. As for consumer-level depth sensors, it still struggles to capture accurate depth maps (Figure 3). Driven by these observations, our goal is to make use of the coarse depth maps and distill useful depth priors to guide the learning of a NeRF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 289, + 511, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 511, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 511, + 303 + ], + "type": "text", + "content": "3.1. Preliminary and Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": "Neural Radiance Fields. Let a NeRF [36] be a mapping function " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " that maps a 3D spatial location " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " and a viewing direction " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " into a volume density " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " and a color value " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " is a neural network consisting of eight perceptron (MLP) layers parameterized by " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": ", which is given by " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "f_{\\theta}: (\\gamma(\\mathbf{x}), \\gamma(\\mathbf{d})) \\mapsto (\\sigma, \\mathbf{c})" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " is a positional encoding. For each expected pixel color " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r})" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": ", it is rendered by casting a ray " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " with near and far bounds " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": ". We evenly partition " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "[t_n, t_f]" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "(t_1, t_2, \\dots, t_N)" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " along a ray " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " and compute expected pixel color " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r})" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r}) = \\sum_{i=1}^{N} \\mathbf{c}_i^*" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": ". The weighted color " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i^*" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " of a 3D point is computed by " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i^* = w_i\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "w_i = T_i(1 - \\exp(-\\sigma_i\\delta_i))" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "T_i = \\exp(-\\sum_{j=1}^{i-1}\\sigma_j\\delta_j)" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "inline_equation", + "content": "\\delta_i = t_i - t_{i-1}" + }, + { + "bbox": [ + 304, + 308, + 547, + 479 + ], + "type": "text", + "content": ". Therefore, the NeRF reconstruction loss can be formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 362, + 486, + 545, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 486, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 362, + 486, + 545, + 512 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n e r f}} = \\sum_ {\\mathbf {r} \\in R} \\left| \\left| \\hat {C} (\\mathbf {r}) - C (\\mathbf {r}) \\right| \\right| ^ {2}, \\tag {1}", + "image_path": "3d6c0b5ee0af7610106ab6820f835478acdfe7d76d1f163ff4a9510f67b11c8b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 521, + 547, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 521, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\hat{C}(\\mathbf{r})" + }, + { + "bbox": [ + 304, + 521, + 547, + 570 + ], + "type": "text", + "content": " are rendered color blended by " + }, + { + "bbox": [ + 304, + 521, + 547, + 570 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 521, + 547, + 570 + ], + "type": "text", + "content": " samples. " + }, + { + "bbox": [ + 304, + 521, + 547, + 570 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r})" + }, + { + "bbox": [ + 304, + 521, + 547, + 570 + ], + "type": "text", + "content": " is the ground-truth pixel color. We use coarse-to-fine sampling as discussed in the vanilla NeRF [36]. Here we omit the fine rendering for simplification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Problem Formulation. Vanilla NeRFs aim to learn a mapping function " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "K_{d}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " dense views by optimizing a color reconstruction loss in Eq. (1). In this work, we aim to study few-shot NeRF when only " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " sparse views are available (" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "K_{s} \\ll K_{d}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "). Let " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " denote the height and width of an image, we have " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "HWK_{s}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " rays as color reconstruction constraints. Without considering view directions " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " and image continuity assumptions (i.e., " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{i}^{*}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " is independent of " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{j}^{*}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "), suppose we have a " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "N_{c} \\times N_{c} \\times N_{c}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " discrete cubic volume to be optimized, which contains " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "N_{c}^{3}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " weighted color variables " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{i}^{*}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": ". We have " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "N_{c}^{3}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " variables " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{i}^{*}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "HWK_{s}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " constraints. Ideally, we are able to solve a discrete cube of edge" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "9067" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 57, + 545, + 336 + ], + "blocks": [ + { + "bbox": [ + 51, + 57, + 545, + 336 + ], + "lines": [ + { + "bbox": [ + 51, + 57, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 51, + 57, + 545, + 336 + ], + "type": "image", + "image_path": "9bb0f5b71768058838c50b5be5b3168e9fd35622762bea1603009483c95d270f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 345, + 547, + 394 + ], + "lines": [ + { + "bbox": [ + 46, + 345, + 547, + 394 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 547, + 394 + ], + "type": "text", + "content": "Figure 2: Framework Overview. SparseNeRF consists of two streams, i.e., NeRF and depth prior distillation. As for NeRF, we use Mip-NeRF as the backbone. We use a NeRF reconstruction loss " + }, + { + "bbox": [ + 46, + 345, + 547, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nerf}}" + }, + { + "bbox": [ + 46, + 345, + 547, + 394 + ], + "type": "text", + "content": ". As for depth prior distillation, we distill depth priors from a pre-trained depth model. Specifically, we propose a local depth ranking regularization and a spatial continuity regularization to distill robust depth priors from coarse depth maps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "text", + "content": "length " + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "inline_equation", + "content": "N_{c} = \\sqrt[3]{HWK_{s}}" + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "text", + "content": ". That is, we can sample " + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "inline_equation", + "content": "\\sqrt[3]{HWK_{s}}" + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "text", + "content": " for each edge of a cubic volume. Take " + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "inline_equation", + "content": "H = W = 512" + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "inline_equation", + "content": "K_{s} = 3" + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "text", + "content": " as an example, we have " + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "inline_equation", + "content": "N_{c} \\approx 92" + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "text", + "content": ", which is far from reconstructing a continuous radiance field. To address this under-constrained optimization problem of few-shot NeRF, an intuitive way is to introduce reasonable regularization terms to constrain few-shot NeRFs conditioned on input " + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r})" + }, + { + "bbox": [ + 46, + 403, + 289, + 513 + ], + "type": "text", + "content": ". Considering these constraints, a general formulation is given by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 533, + 287, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 533, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 287, + 546 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {n e r f}} + \\lambda \\mathcal {R} (\\mathbf {x}, \\mathbf {d}, C (\\mathbf {r})), \\tag {2}", + "image_path": "87524e6cad11845d23e677ed53718b11246b5acb3f7a3e295652480ce6ec8db1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 564, + 183, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 183, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 183, + 577 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 564, + 183, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 47, + 564, + 183, + 577 + ], + "type": "text", + "content": " is a regularization term." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Remark. RegNeRF [37] and InfoNeRF [27] tackle this problem by introducing regularization terms on continuous depth constraint from unobserved viewpoints, sparsity of density on rays, and patch-based semantic constraint on color appearance. Depth-based models, such as DSNeRF [17] and MonoSDF [65], directly regress depth by leveraging sparse 3D points or deriving ground-truth depth maps with accurate absolute depth maps. Different from them, we propose to use a robust relative depth regularization from coarse depth maps of pre-trained depth models or consumer-level depth sensors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 403, + 448, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 448, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 448, + 417 + ], + "type": "text", + "content": "3.2. Overview of SparseNeRF" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 422, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 626 + ], + "type": "text", + "content": "The pipeline of SparseNeRF is illustrated in Figure 2. SparseNeRF mainly consists of four components, i.e., a neural radiance field (NeRF), a color reconstruction module, a depth ranking distillation module, and a spatial continuity distillation module. Specifically, we use Mip-NeRF [1] as the backbone and apply an MSE loss for color reconstruction " + }, + { + "bbox": [ + 304, + 422, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nerf}}" + }, + { + "bbox": [ + 304, + 422, + 547, + 626 + ], + "type": "text", + "content": ". As for depth prior distillation, we either use a pre-trained depth model to estimate depth maps or capture coarse depth maps with consumer-level depth sensors. We use a vision transformer (DPT) [41] that is trained on a large-scale mixed depth dataset (1.4M images with various depth annotations) that covers a wide range of scenes. Therefore, DPT is able to provide general depth priors. Since single-view depth estimation is coarse, we carefully design a local depth ranking regularization and a spatial continuity regularization, which distills robust depth priors from coarse maps to the NeRF." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 635, + 484, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 484, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 484, + 647 + ], + "type": "text", + "content": "3.3. Local Depth Ranking Distillation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Single-view depth estimation is a challenging computer vision task, which aims at predicting a depth map of a scene given a single image as input. Due to dataset bias, coarse depth annotations, and imperfect neural models, it is difficult to achieve accurate depth prediction. Depth maps cap" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "9068" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 66, + 286, + 322 + ], + "blocks": [ + { + "bbox": [ + 49, + 66, + 286, + 322 + ], + "lines": [ + { + "bbox": [ + 49, + 66, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 49, + 66, + 286, + 322 + ], + "type": "image", + "image_path": "da4c811d6718d8ca818667df2c426261f95e8b65ea0c5ce7bbfe687ea89da00a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 331, + 288, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 288, + 401 + ], + "type": "text", + "content": "Figure 3: Two types of coarse depth maps. Top: predicted by a pre-trained depth model. Bottom: captured by Azure Kinect and ZED 2 depth sensors. Si-MSE is a scale-invariant depth error. Time jittering refers to the difference between two depth maps captured by a static camera within a small time interval." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 423, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 589 + ], + "type": "text", + "content": "tured by consumer-level sensors are also inaccurate due to wide-range depth distances. Directly using coarse depth maps to supervise a NeRF leads to ambiguous rendered novel view synthesis. To avoid the error of coarse depth maps, we relax the depth constraint and exploit the robust depth ranking prior. Given a pair of pixels in a single image, the depth ranking regularization only considers which point is nearer or farther. However, when the scene is complex, even depth ranking is not accurate. As shown in Figure 4, it is easy for depth estimation models to compare the depth ranking of white and cyan points, but it is hard to estimate the depth ranking of white and red points. It implies that the depth ranking becomes unreliable as the spatial distance increases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Motivated by these observations, we propose a local depth ranking distillation method that distills depth ranking priors from coarse depth maps to a NeRF. On one hand, given a local patch " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " of an RGB image " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " with a pose " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". We compute the depth " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{r}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " of the rays that trace from " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{r}} = \\sum_{i=1}^{N} w_{i} t_{i}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". On the other hand, we use pre-trained depth DPT [41] to estimate the depth of " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and crop a local patch " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{dpt}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " with the same spatial location as " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{r}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". We perform the depth ranking distillation by transferring the depth ranking knowledge from " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{dpt}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{r}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "k1" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "k2" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " be the" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 65, + 545, + 155 + ], + "blocks": [ + { + "bbox": [ + 307, + 65, + 545, + 155 + ], + "lines": [ + { + "bbox": [ + 307, + 65, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 307, + 65, + 545, + 155 + ], + "type": "image", + "image_path": "6ced21e759b7dfb9e657fc27487da1c300e891a5f38ee58dcfe3b39e49a6530c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 166, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 304, + 166, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 304, + 166, + 547, + 214 + ], + "type": "text", + "content": "Figure 4: Motivation of local depth ranking. It is easy for depth estimation models to tell the white point is nearer than the cyan point, but it is hard to compare while and red points. Best viewed by zooming in." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 224, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 248 + ], + "type": "text", + "content": "indices of 2D pixel coordinate of " + }, + { + "bbox": [ + 305, + 224, + 545, + 248 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{dpt}}" + }, + { + "bbox": [ + 305, + 224, + 545, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 224, + 545, + 248 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{r}}" + }, + { + "bbox": [ + 305, + 224, + 545, + 248 + ], + "type": "text", + "content": ". The depth ranking regularization is given by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 336, + 255, + 545, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 255, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 336, + 255, + 545, + 284 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\text {r a n k}} = \\sum_ {d _ {\\mathrm {d p t}} ^ {k 1} \\leq d _ {\\mathrm {d p t}} ^ {k 2}} \\max \\left(d _ {\\mathbf {r}} ^ {k 1} - d _ {\\mathbf {r}} ^ {k 2} + m, 0\\right), \\tag {3}", + "image_path": "fad5982cecd755a4f828dcd83b9ce5c0e67590048740e9ed5abf1c8894c70267.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "content": " is a small margin that allows limited depth ranking errors. In Eq. (3), we randomly sample two depth pixels of " + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{dpt}}" + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "content": ", as denoted as " + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{dpt}}^{k1}" + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{dpt}}^{k2}" + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{dpt}}^{k1} \\leq d_{\\mathrm{dpt}}^{k2}" + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "content": ". If the depth rankings of " + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{dpt}}" + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{r}}" + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "content": " are not consistent, i.e., " + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "inline_equation", + "content": "d_{\\mathbf{r}}^{k1} > d_{\\mathbf{r}}^{k2}" + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{dpt}}^{k1} \\leq d_{\\mathrm{dpt}}^{k2}" + }, + { + "bbox": [ + 304, + 289, + 546, + 399 + ], + "type": "text", + "content": ", we punish the NeRF. We encourage the corresponding depth estimated by NeRF to satisfy the consistent depth ranking of the pre-trained model. We clip the large depth value and normalize the depth map for depth sensors. For depth models, we use relative inverse depth." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 406, + 470, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 470, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 470, + 418 + ], + "type": "text", + "content": "3.4. Spatial Continuity Distillation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 424, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 533 + ], + "type": "text", + "content": "The local depth ranking constraint guarantees that the predicted depth map estimated by NeRF is consistent with the DPT's depth map. However, it does not constrain the spatial continuity of the depth map. We distill spatial continuity priors from the depth model DPT, which allows large displacement across several depth pixels. If neighbor depth pixels are continuous on the depth map of DPT, we constrain the corresponding depth pixels of NeRF to be continuous. The spatial continuity regularization is given by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 538, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 538, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 311, + 538, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\text {c o n t i}} = \\sum_ {k 1} \\sum_ {d _ {\\mathrm {d p t}} ^ {k 2} \\in \\mathrm {K N N} \\left(d _ {\\mathrm {d p t}} ^ {k 1}\\right)} \\max \\left(\\left| d _ {\\mathbf {r}} ^ {k 1} - d _ {\\mathbf {r}} ^ {k 2} \\right| - m ^ {\\prime}, 0\\right), \\tag {4}", + "image_path": "0b8f188d7b34dae91c4c7030a15337f6d10d5e0cfc79a90cfb00424b5b82946b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 578, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 578, + 545, + 625 + ], + "type": "inline_equation", + "content": "\\mathrm{KNN}(\\cdot)" + }, + { + "bbox": [ + 304, + 578, + 545, + 625 + ], + "type": "text", + "content": " returns k-nearest neighbors measured by depth values within a small region, e.g., " + }, + { + "bbox": [ + 304, + 578, + 545, + 625 + ], + "type": "inline_equation", + "content": "6 \\times 6" + }, + { + "bbox": [ + 304, + 578, + 545, + 625 + ], + "type": "text", + "content": " patch. " + }, + { + "bbox": [ + 304, + 578, + 545, + 625 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 304, + 578, + 545, + 625 + ], + "type": "text", + "content": " is a small margin that allows small depth differences between neighbor pixels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 626, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 626, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 305, + 626, + 545, + 650 + ], + "type": "text", + "content": "Full Objective Loss. The full objective loss of SparseNeRF is formulated by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 658, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 658, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 358, + 658, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {n e r f}} + \\lambda \\mathcal {R} _ {\\text {r a n k}} + \\gamma \\mathcal {R} _ {\\text {c o n t i}}, \\tag {5}", + "image_path": "15313cdac806d1bdb467e46f31df35129481ccf4051aa3e3342dfda830efcdbd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": " control the importance of the regularization terms. In practice, we set " + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\lambda = 0.2" + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\gamma = 0.02" + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "inline_equation", + "content": "m = m' = 1.0 \\times 10^{-4}" + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "9069" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 167, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 167, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 167, + 83 + ], + "type": "text", + "content": "4. NVS-RGBD Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 344 + ], + "type": "text", + "content": "The popular benchmark LLFF [35] on NeRFs is featured with forward-facing and object-centric. Moreover, we focus on consumer-level depth sensors instead of expensive scanners [14]. We study stand-alone depth cameras instead of a complex calibrated structured multiple camera system [15], and collect from real-world scenes instead of synthesized 3D assets [45]. Few RGBD datasets satisfy all of the requirements. Therefore, we collect a new dataset NVS-RGBD that contains real-world depth maps captured by consumer-level depth sensors, including Azure Kinect, ZED 2, and iPhone 13 Pro. We collect 8 scenes for both Azure Kinect and ZED 2, and 4 scenes for iPhone 13 Pro. The depth maps of Azure Kinect often contain noises around object edges. The depth maps of ZED 2 look smooth but unstable and incorrect when observing time jittering (Figure 3). The poses are computed via COLMAP. We collect indoor scenes for ZED 2 and Kinect. For iPhone, we collect both indoor and outdoor scenes. All the scenes are object-centric. Depth maps obtained from sensors have different artifacts coming from the sensor noises. Please refer to the supplementary material for the details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 356, + 128, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 128, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 128, + 370 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 289, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 581 + ], + "type": "text", + "content": "Datasets. We conduct experiments on the LLFF [35], DTU [26], and NVS-RGBD datasets. LLFF contains 8 complex forward-facing scenes. Following RegNeRF [37], we use every 8-th image as the held-out test set, and evenly sample sparse views from the remaining images for training. Different from LLFF, DTU is an object-level dataset. Following PixNeRF [64, 37], we use the same 15 scenes in our experiments. On the DTU dataset, the background of these scenes is a white table or a black background, which has few textures. As mentioned in RegNeRF, we mask the background during inference to avoid background bias. We use the same evaluation protocol for a fair comparison. On the NVS-RGBD dataset, for each scene, we use 3 views for training and the rest views for inference. For LLFF and DTU, we use pre-trained depth maps to implement our method. For NVS-RGBD, we use coarse depth maps captured by depth sensors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt four evaluation metrics in the experiments, i.e., PSNR, SSIM [55], LPIPS [66], and depth error [17]. Depth error is a scale-invariant MSE derived by minimizing " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "||wd\\hat{d} + b - d||" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\hat{d}" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 581, + 287, + 677 + ], + "type": "text", + "content": " denote coarse depth maps and pseudo-ground-truth depth maps predicted by dense-view NeRF. Note that we only use dense views to train a good NeRF to generate pseudo-ground-truth depth maps for evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Implementation Details. We implement SparseNeRF based on the official JAX [3]. We use the Adam optimizer for the learning of SparseNeRF. We use an exponen" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "tial learning rate, which decays from " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ". The batch size is set to 4096. For each scene, we use one GPU-v100 with 32G memory for both training and inference. We also implement the proposed SparseNeRF on GPU-v100 with 16G and RTX 2080 Ti with 11G. We reduce the batch size to 2048 and 1024 for these two types of GPUs, respectively. We use the same backbone and sampled points along rays compared with prior arts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 171, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 243 + ], + "type": "text", + "content": "We train 90k iterations for each scene given three training views, which takes about 2 hours for each scene. For depth maps captured by Kinect and ZED 2, we mask uncertain regions (black regions in depth maps). More precisely, we sample near-far depth pairs in certain regions in local patches to optimize Eqs. 3 and 4 of the paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 258, + 436, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 436, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 436, + 272 + ], + "type": "text", + "content": "5.1. Comparisons on LLFF" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 280, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 579 + ], + "type": "text", + "content": "We compare SparseNeRF with state-of-the-art methods on the LLFF dataset, including SRF [12], PixelNeRF [64], MVSNeRF [5], Mip-NeRF [1], DietNeRF [25], RegNeRF [37], DSNeRF [17] and MonoSDF[65]. Among them, SRF, PixelNeRF, and MVSNeRF are pre-trained on other similar scenes to exploit high-level semantics. Mip-NeRF is the state-of-the-art NeRF designed for dense-view training. InfoNeRF and RegNeRF mainly constrain sparsity and continuity of geometry or natural appearance semantics. More precisely, the compared methods can be classified into four groups. In the first group, SRF, PixelNeRF, and MVSNeRF are pre-trained in the large-scale DTU datasets (88 scenes) and are directly tested on the LLFF dataset. In the second group, SRF, PixelNeRF, and MVSNeRF are pre-trained on DTU and fine-tuned on LLFF per scene. The third group includes Mip-NeRF, DietNeRF, and RegNeRF. They use geometry continuity and semantic constraints. The fourth group distills knowledge from depth maps or sparse points of COLMAP (DSNeRF). The compared results are shown in Table 1. We can see that the proposed SparseNeRF achieves the best performance in PSNR, SSIM, and LPIPS. Note that the first and second groups have to pre-train on lots of other scenes. SparseNeRF achieves better results than DSNeRF and MonoSDF because the depth ranking is more robust than the absolute scale-invariant constraint." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 715 + ], + "type": "text", + "content": "We provide qualitative analysis on LLFF, as shown in Figure 5. It is observed that PixelNeRF and MVSNeRF tend to generate ambiguous pixels. The reason is that they integrate CNN features into NeRFs. CNN features can provide high-level semantics to infer under-constrained textures but suffer pixel-level inference. MipNeRF is designed for dense-view synthesis without considering the under-constrained problem, leading to degraded geometry artifacts. Compared with RegNeRF, SparseNeRF uses robust depth priors from pre-trained depth models, which can handle challenging scene geometry with better performance." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "9070" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 57, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 49, + 57, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 49, + 57, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 49, + 57, + 545, + 191 + ], + "type": "image", + "image_path": "75858878c15072142d88042dca52d770058855cb35cea1341aeae4c6fbac1427.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 233, + 545, + 365 + ], + "blocks": [ + { + "bbox": [ + 46, + 205, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 547, + 229 + ], + "type": "text", + "content": "Figure 5: Visual comparisons on the LLFF dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 233, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 50, + 233, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 50, + 233, + 545, + 365 + ], + "type": "image", + "image_path": "26a7c24ddc19e29e415488dd817b644d64faea2cdd3e15a4b217d933dd618e99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 377, + 547, + 402 + ], + "lines": [ + { + "bbox": [ + 46, + 377, + 547, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 547, + 402 + ], + "type": "text", + "content": "Figure 6: Visual comparisons on the DTU dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 288, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 496 + ], + "type": "text", + "content": "Table 1: Quantitative Comparison on LLFF with three views. There are four groups. The first group denotes models that are pre-trained on other scenes and tested on a target scene (P). The second group requires further finetuning on a target scene (P&FT). The third group uses geometry and semantic constraints (geo. & sem.). The fourth group distills depth knowledge (depth KD)." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 500, + 280, + 630 + ], + "blocks": [ + { + "bbox": [ + 56, + 500, + 280, + 630 + ], + "lines": [ + { + "bbox": [ + 56, + 500, + 280, + 630 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 280, + 630 + ], + "type": "table", + "html": "
SettingPSNR ↑SSIM ↑LPIPS ↓
SRF [12]P12.340.2500.591
PixelNeRF [64]7.930.2720.682
MVSNeRF [5]17.250.5570.356
SRF ft [12]P&FT17.070.4360.529
PixelNeRF ft [64]16.170.4380.512
MVSNeRF ft [5]17.880.5840.327
Mip-NeRF [1]geo. & sem.14.620.3510.495
DietNeRF [25]14.940.3700.496
RegNeRF [37]19.080.5870.336
MonoSDF*[65]depth KD18.450.5650.388
DSNeRF[17]18.940.5820.362
SparseNeRF (Ours)19.860.6240.328
", + "image_path": "7a58495f0dacf2a4a4afea6926cd7c02f815805d91f9af962d6126cf54baf99e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 643, + 173, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 643, + 173, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 643, + 173, + 657 + ], + "type": "text", + "content": "5.2. Comparisons on DTU" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "Similar to the LLFF dataset, we conduct four-group comparisons on the DTU dataset, as shown in Table 2. In this first group, given a target scene, SRF, PixelNeRF, and MVSNeRF are pre-trained on other DTU scenes and are" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 411, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 484 + ], + "type": "text", + "content": "Table 2: Quantitative Comparison on DTU with three views. The first group is pre-trained on other scenes and tested on a target scene (P). The second group requires further finetuning on a target scene (P&FT). The third group uses geometry and semantic constraints (geo. & sem.). The fourth group distills depth knowledge (depth KD)." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 488, + 544, + 624 + ], + "blocks": [ + { + "bbox": [ + 310, + 488, + 544, + 624 + ], + "lines": [ + { + "bbox": [ + 310, + 488, + 544, + 624 + ], + "spans": [ + { + "bbox": [ + 310, + 488, + 544, + 624 + ], + "type": "table", + "html": "
SettingPSNR ↑SSIM ↑LPIPS ↓
SRF [12]P15.320.6710.304
PixelNeRF [64]16.820.6950.270
MVSNeRF [5]18.630.7690.197
SRF ft [12]P&FT15.680.6980.281
PixelNeRF ft [64]18.950.7100.269
MVSNeRF ft [5]18.540.7690.197
Mip-NeRF [1]geo. & sema.8.680.5710.353
DietNeRF [25]11.850.6330.314
RegNeRF [37]18.890.7450.190
MonoSDF* [65]depth KD18.920.7480.237
DSNeRF [17]16.900.5700.450
SparseNeRF(Ours)19.550.7690.201
", + "image_path": "ca13b3b973b034d7837782e087dd12dad223952470b4cd7e4bd903b2b185fb54.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "tested on the target scene. In the second group, these three methods are further fine-tuned on the target scene. In the third group, we compare Mip-NeRF, DietNeRF, and RegN-eRF. Note that for the pre-trained models, they split DTU as a training set (88 scenes) and a test set (15 scenes). Since" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 314, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 314, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 314, + 742 + ], + "type": "text", + "content": "9071" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 83, + 146, + 138 + ], + "blocks": [ + { + "bbox": [ + 80, + 72, + 112, + 80 + ], + "lines": [ + { + "bbox": [ + 80, + 72, + 112, + 80 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 112, + 80 + ], + "type": "text", + "content": "DSNeRF" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 83, + 146, + 138 + ], + "lines": [ + { + "bbox": [ + 51, + 83, + 146, + 138 + ], + "spans": [ + { + "bbox": [ + 51, + 83, + 146, + 138 + ], + "type": "image", + "image_path": "3a8cc286876affb436e5ae62b0bbbd183d09075e72c3231937f127196a7c006f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 151, + 83, + 246, + 138 + ], + "blocks": [ + { + "bbox": [ + 172, + 72, + 213, + 80 + ], + "lines": [ + { + "bbox": [ + 172, + 72, + 213, + 80 + ], + "spans": [ + { + "bbox": [ + 172, + 72, + 213, + 80 + ], + "type": "text", + "content": "MonoSDF*" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 83, + 246, + 138 + ], + "lines": [ + { + "bbox": [ + 151, + 83, + 246, + 138 + ], + "spans": [ + { + "bbox": [ + 151, + 83, + 246, + 138 + ], + "type": "image", + "image_path": "d7fe4728de2a19677aaa89656c081ff72e73158bcc825576d83c99808c0a876b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 250, + 83, + 345, + 138 + ], + "blocks": [ + { + "bbox": [ + 276, + 72, + 310, + 82 + ], + "lines": [ + { + "bbox": [ + 276, + 72, + 310, + 82 + ], + "spans": [ + { + "bbox": [ + 276, + 72, + 310, + 82 + ], + "type": "text", + "content": "RegNeRF" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 83, + 345, + 138 + ], + "lines": [ + { + "bbox": [ + 250, + 83, + 345, + 138 + ], + "spans": [ + { + "bbox": [ + 250, + 83, + 345, + 138 + ], + "type": "image", + "image_path": "92f98aa96d487373d69a8c10c4d586ea005391f268ea9e395baaaf3d158153b7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 349, + 83, + 444, + 138 + ], + "blocks": [ + { + "bbox": [ + 365, + 72, + 432, + 82 + ], + "lines": [ + { + "bbox": [ + 365, + 72, + 432, + 82 + ], + "spans": [ + { + "bbox": [ + 365, + 72, + 432, + 82 + ], + "type": "text", + "content": "SparseNeRF (ours)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 83, + 444, + 138 + ], + "lines": [ + { + "bbox": [ + 349, + 83, + 444, + 138 + ], + "spans": [ + { + "bbox": [ + 349, + 83, + 444, + 138 + ], + "type": "image", + "image_path": "10731e1d75e132f846ea79fcf79ea2d4fbf94ff214848cb2ed031add2b972680.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 448, + 83, + 543, + 138 + ], + "blocks": [ + { + "bbox": [ + 465, + 72, + 512, + 80 + ], + "lines": [ + { + "bbox": [ + 465, + 72, + 512, + 80 + ], + "spans": [ + { + "bbox": [ + 465, + 72, + 512, + 80 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 448, + 83, + 543, + 138 + ], + "lines": [ + { + "bbox": [ + 448, + 83, + 543, + 138 + ], + "spans": [ + { + "bbox": [ + 448, + 83, + 543, + 138 + ], + "type": "image", + "image_path": "ca91274958313cceb7c430db55c5c06db6cbb380cff48b9796002363f91a4366.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 51, + 140, + 146, + 194 + ], + "blocks": [ + { + "bbox": [ + 51, + 140, + 146, + 194 + ], + "lines": [ + { + "bbox": [ + 51, + 140, + 146, + 194 + ], + "spans": [ + { + "bbox": [ + 51, + 140, + 146, + 194 + ], + "type": "image", + "image_path": "c20aff936b70546755ac1eeabfeda8cd8d180983e3335ea32bcf648c87b074b5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 545, + 232 + ], + "type": "text", + "content": "Figure 7: Visual comparisons on the NVS-RGBD dataset with three views. Red boxes denote compared regions. SparseNeRF achieves consistent improvement in different scenes." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 151, + 140, + 246, + 194 + ], + "blocks": [ + { + "bbox": [ + 151, + 140, + 246, + 194 + ], + "lines": [ + { + "bbox": [ + 151, + 140, + 246, + 194 + ], + "spans": [ + { + "bbox": [ + 151, + 140, + 246, + 194 + ], + "type": "image", + "image_path": "d4d7b76290cfe97e0b20019380f936dc13697917fcb2d43cc6d7836c25427a2c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 250, + 140, + 345, + 194 + ], + "blocks": [ + { + "bbox": [ + 250, + 140, + 345, + 194 + ], + "lines": [ + { + "bbox": [ + 250, + 140, + 345, + 194 + ], + "spans": [ + { + "bbox": [ + 250, + 140, + 345, + 194 + ], + "type": "image", + "image_path": "049b3dc06ceb3dac98067a55eb03628ae3315477afa1fb3348d197eafda66a1f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 349, + 140, + 444, + 194 + ], + "blocks": [ + { + "bbox": [ + 349, + 140, + 444, + 194 + ], + "lines": [ + { + "bbox": [ + 349, + 140, + 444, + 194 + ], + "spans": [ + { + "bbox": [ + 349, + 140, + 444, + 194 + ], + "type": "image", + "image_path": "d75e48e88540eaea2f8ef21b59a760ee1cf17f6e8b0765d5e184fa225f62146b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 448, + 140, + 543, + 194 + ], + "blocks": [ + { + "bbox": [ + 448, + 140, + 543, + 194 + ], + "lines": [ + { + "bbox": [ + 448, + 140, + 543, + 194 + ], + "spans": [ + { + "bbox": [ + 448, + 140, + 543, + 194 + ], + "type": "image", + "image_path": "d46d68af4155b95e5f44bb524ec21ef1af188df22ad6ed7fae444e70e5dc4a62.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 50, + 249, + 108, + 293 + ], + "blocks": [ + { + "bbox": [ + 52, + 240, + 105, + 249 + ], + "lines": [ + { + "bbox": [ + 52, + 240, + 105, + 249 + ], + "spans": [ + { + "bbox": [ + 52, + 240, + 105, + 249 + ], + "type": "text", + "content": "w/o conti. & ranking" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 249, + 108, + 293 + ], + "lines": [ + { + "bbox": [ + 50, + 249, + 108, + 293 + ], + "spans": [ + { + "bbox": [ + 50, + 249, + 108, + 293 + ], + "type": "image", + "image_path": "8652c6100c9f4481ae8e329de3b0fa81411815f7627378abe711e2e98cac5648.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 304, + 287, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 352 + ], + "type": "text", + "content": "Figure 8: Ablation study on visual effect. The three columns denote \"w/o continuity and ranking distillation\", \"w/o continuity distillation\", and \"w/ continuity and ranking distillation\" (SparseNeRF, ours), respectively." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 109, + 250, + 168, + 293 + ], + "blocks": [ + { + "bbox": [ + 123, + 241, + 154, + 248 + ], + "lines": [ + { + "bbox": [ + 123, + 241, + 154, + 248 + ], + "spans": [ + { + "bbox": [ + 123, + 241, + 154, + 248 + ], + "type": "text", + "content": "w/o ranking" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 250, + 168, + 293 + ], + "lines": [ + { + "bbox": [ + 109, + 250, + 168, + 293 + ], + "spans": [ + { + "bbox": [ + 109, + 250, + 168, + 293 + ], + "type": "image", + "image_path": "5f5c61f6a57a90e41dc7ea4c5443bf71e6fd0bb32c0e06afc2d76fcd5cb0152a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 169, + 250, + 227, + 293 + ], + "blocks": [ + { + "bbox": [ + 184, + 241, + 208, + 247 + ], + "lines": [ + { + "bbox": [ + 184, + 241, + 208, + 247 + ], + "spans": [ + { + "bbox": [ + 184, + 241, + 208, + 247 + ], + "type": "text", + "content": "w/o conti." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 250, + 227, + 293 + ], + "lines": [ + { + "bbox": [ + 169, + 250, + 227, + 293 + ], + "spans": [ + { + "bbox": [ + 169, + 250, + 227, + 293 + ], + "type": "image", + "image_path": "888a9648beb336f6626ac7c988d861dc084742a5075cb25eab4e83c486aa8a12.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 228, + 250, + 286, + 293 + ], + "blocks": [ + { + "bbox": [ + 229, + 241, + 277, + 248 + ], + "lines": [ + { + "bbox": [ + 229, + 241, + 277, + 248 + ], + "spans": [ + { + "bbox": [ + 229, + 241, + 277, + 248 + ], + "type": "text", + "content": "SparseNeRF(ours)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 250, + 286, + 293 + ], + "lines": [ + { + "bbox": [ + 228, + 250, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 228, + 250, + 286, + 293 + ], + "type": "image", + "image_path": "bac4f1bc85c90cf5283877288d3530b14e8eaa71b23af78ed128d10dab29640d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 374, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 589 + ], + "type": "text", + "content": "the training set and the test set contains similar scenes, a test scene shares a similar distribution (e.g., similar backgrounds) with training scenes. Pre-training on other scenes greatly benefits a target scene. Therefore, SRF, PixelNeRF, and MVSNeRF also greatly achieve promising results. Finally, we compare depth-based DSNeRF [17] and MonoSDF [65]. DSNeRF studies two kinds of depth information, i.e., sparse 3D points generated by COLMAP [42] and ground-truth depth maps obtained by high-accuracy depth scanners and the Multi-View Stereo (MVS) algorithm. The depth maps of NVS-RGBD are captured by consumer-level sensors and contain lots of noise, which is hard to be applied to DSNeRF. Therefore, we adopt sparse 3D points for DSNeRF. Compared with these methods, SparseNeRF still achieves the best performance without pre-training on other scenes. We provide a qualitative analysis, as shown in Figure 6. SparseNeRF achieves better visual results than previous state-of-the-art methods." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 598, + 206, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 206, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 206, + 611 + ], + "type": "text", + "content": "5.3. Comparisons on NVS-RGBD" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "We then implement three state-of-the-art methods on our new NVS-RGBD dataset, including RegNeRF [37], DSNeRF [17], and MonoSDF [65]. All of them are optimized per scene without pre-training on the other scenes. RegNeRF adopted continuous geometry assumptions from unobserved viewpoints. DSNeRF uses a global scale-invariant depth constraint to supervise NeRFs. MonoSDF mainly focuses on implicit surface reconstruction. It is as-" + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 309, + 265, + 545, + 360 + ], + "blocks": [ + { + "bbox": [ + 306, + 241, + 544, + 264 + ], + "lines": [ + { + "bbox": [ + 306, + 241, + 544, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 544, + 264 + ], + "type": "text", + "content": "Table 3: Comparison with State of the Arts on NVS-RGBD. * denotes a new implementation from related tasks." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 265, + 545, + 360 + ], + "lines": [ + { + "bbox": [ + 309, + 265, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 309, + 265, + 545, + 360 + ], + "type": "table", + "html": "
MetricsPSNR↑SSIM↑LPIPS↓depth err↓
KinectRegNeRF [37]25.780.8400.2425.9×10-3
DSNeRF [17]25.910.8380.2384.8×10-3
MonoSDF* [65]25.500.8380.2395.5×10-3
SparseNeRF(ours)26.280.8500.2323.9×10-3
ZED 2RegNeRF [37]24.980.7810.2713.5×10-3
DSNeRF [17]24.970.7890.2623.1×10-3
MonoSDF* [65]24.540.7840.2753.2×10-3
SparseNeRF(ours)26.220.8040.2622.4×10-3
", + "image_path": "3fb45cf9fa821f59b901f984a1e621fc5001ae7276e7417a531b2bd048f6dc50.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "content": "sumed that local patches of depth maps from monocular cameras satisfy " + }, + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "inline_equation", + "content": "d' = w d + b" + }, + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "content": " is predicted by the pre-trained model and " + }, + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 304, + 376, + 545, + 460 + ], + "type": "text", + "content": " is then computed by the NeRF. In Table 3, we observe that SparseNeRF achieves better performance compared with these prior arts. We provide a qualitative analysis in Figure 7, demonstrating the superiority of SparseNeRF over previous state-of-the-art methods." + } + ] + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 309, + 489, + 545, + 538 + ], + "blocks": [ + { + "bbox": [ + 306, + 472, + 544, + 484 + ], + "lines": [ + { + "bbox": [ + 306, + 472, + 544, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 544, + 484 + ], + "type": "text", + "content": "Table 4: Comparison of Depth Ranking and Depth scaling." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 489, + 545, + 538 + ], + "lines": [ + { + "bbox": [ + 309, + 489, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 309, + 489, + 545, + 538 + ], + "type": "table", + "html": "
LLFFPSNR↑SSIM↑LPIPS↓depth err↓
baseline19.080.5870.33610.1×10-3
w/ depth scaling18.450.5650.3888.9×10-3
w/ depth ranking19.860.6240.3286.3×10-3
", + "image_path": "a0ffbc9deb5baf93bc26685bcff098b7343ed6a57d6e0c0de9afa301f1c12003.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 551, + 510, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 510, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 510, + 563 + ], + "type": "text", + "content": "5.4. Ablation Studies and Further Analyses" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": "Effectiveness of Depth Distillation. To validate the effectiveness of the depth distillation, we isolate the depth constraint and keep other modules unchanged. As shown in the first and third rows of Table 4, it is observed that without the local depth ranking distillation, SparseNeRF degrades on PSNR, SSIM, LPIPS, and depth error, which demonstrates the effectiveness of the depth distillation. We further study the depth maps from unobserved views. The ablation study of visual effect is shown in Figure 8. It shows that depth ranking greatly improves performance in geometry. Spatial continuity regularization further improves the detailed coherence of scenes." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "9072" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 87, + 286, + 152 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 276, + 83 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 276, + 83 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 276, + 83 + ], + "type": "text", + "content": "Table 5: Impact of different pre-trained depth models." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 87, + 286, + 152 + ], + "lines": [ + { + "bbox": [ + 51, + 87, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 51, + 87, + 286, + 152 + ], + "type": "table", + "html": "
LLFFPSNR↑SSIM↑LPIPS↓depth err↓
Baseline19.080.5870.33610.1×10-3
MiDaS small19.350.5920.3336.9×10-3
DPT Hybrid19.860.6250.3266.3×10-3
DPT Large19.860.6240.3286.3×10-3
", + "image_path": "5f44f6e1ed381493abb141d7d542d5473cc06c37120f54322dc96229def513b4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 162, + 287, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 162, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 287, + 341 + ], + "type": "text", + "content": "Further Analyses on Depth Ranking and Depth Scaling. To validate the robustness of local depth ranking, we compare depth ranking regularization and depth scaling regularization on LLFF. We use the idea from MonoSDF [65], which mainly focuses on implicit surface reconstruction. With the specific design for surface shapes, MonoSDF slightly degrades the performance of RGB view synthesis. We re-implement it as depth scaling. Since the single-view depth estimation is not accurate, the linear depth scaling constraint is too strong such that depth distillation misleads the learning of the NeRF. Figure 3 shows two cases of scale-invariant errors. Instead, depth ranking relaxes the constraint and only focuses on depth comparison, which is more robust than depth scaling. As shown in Table 4, depth ranking performs better than depth scaling on LLFF." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 287, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 412 + ], + "type": "text", + "content": "Impact of Different Pre-trained Depth Models. We test our method on three depth estimation models, i.e., MiDaS small, DPT Hybrid, and DPT Large. As shown in Table 5, all of the depth models are better than the baseline. DPT Hybrid and DPT Large achieve comparable results, which are better than MiDaS small." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 413, + 286, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 286, + 496 + ], + "type": "text", + "content": "FreeNeRF [61] potentially complements SparseNeRF. FreeNeRF and SparseNeRF improve few-shot NeRF in different ways. FreeNeRF introduces a frequency regularization while SparseNeRF distills depth priors from pretrained general depth estimators. We implement SparseNeRF on FreeNeRF and the results significantly improve, as shown in Table 6." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 528, + 287, + 566 + ], + "blocks": [ + { + "bbox": [ + 51, + 506, + 282, + 517 + ], + "lines": [ + { + "bbox": [ + 51, + 506, + 282, + 517 + ], + "spans": [ + { + "bbox": [ + 51, + 506, + 282, + 517 + ], + "type": "text", + "content": "Table 6: FreeNeRF potentially complement SparseNeRF." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 528, + 287, + 566 + ], + "lines": [ + { + "bbox": [ + 50, + 528, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 50, + 528, + 287, + 566 + ], + "type": "table", + "html": "
LLFFPSNR↑SSIM↑LPIPS↓
FreeNeRF19.600.6140.302
FreeNeRF+our depth distillation20.260.6460.292
", + "image_path": "467ae9f4dcde4446670317ccf1de668a34cfa26c12f35bc7dcadf005c4c1cf47.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "User Study. To show that the proposed spatial continuity regularization improves the 3D consistency in 3D space, we conducted a user study. For each scene, we rendered videos for both with and without spatial continuity regularization terms. Users are asked to select the better-quality video according to 3D consistency/coherence and completeness (good geometry, few floaters). We evaluate 8 scenes on LLFF and there are 27 participants. The result shows " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "88.9\\%" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " of users prefer \"w/ conti\" and " + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": " prefer \"w/o conti\". To evaluate the advantages of SparseNeRF on 3D consistency and shape completeness, we conduct another user study. Users are asked to evaluate \"baseline+RegNeRF's continu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "ity term” and “baseline+our continuity term”. The 27 users participated in the study and 8 scenes on LLFF are evaluated. The result shows " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "87.5\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " of users prefer “baseline+our continuity term” and " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "12.5\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " prefer “baseline+RegNeRF’s continuity term”. RegNeRF encourages the overall patch to be as smooth as possible without considering the sharp edges of objects, which conflicts with the reconstruction from sparse views. Our continuity regulation distills the neighbor relationship from depth models is more accurate. We show two examples of depth maps in Figure 9." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 202, + 545, + 256 + ], + "blocks": [ + { + "bbox": [ + 307, + 202, + 545, + 256 + ], + "lines": [ + { + "bbox": [ + 307, + 202, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 307, + 202, + 545, + 256 + ], + "type": "image", + "image_path": "e87f7feacc4d162237d85a5348fea56ffb91e41ab9541d1490394fa5f2a2f199.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 265, + 534, + 277 + ], + "lines": [ + { + "bbox": [ + 316, + 265, + 534, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 534, + 277 + ], + "type": "text", + "content": "Figure 9: Visual comparison on continuity regulation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 290, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 375 + ], + "type": "text", + "content": "Discussion. SparseNeRF could have a wide range of applications due to its promising results. We provide two applications on iPhone 13 Pro. Please refer to the project page for more experiments. Like UNISURF [38, 65], we also evaluate the SparseNeRF using Chamfer distance, showing the competitive results with MonoSDF and UNISURF, especially for complex scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 387, + 378, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 387, + 378, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 387, + 378, + 399 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "text", + "content": "In this paper, we propose a SparseNeRF framework that synthesizes novel views with sparse view inputs. To tackle the under-constrained few-shot NeRF problem, we propose a local depth ranking regularization that distills the depth ranking prior from coarse depth maps. To preserve the spatial continuity of geometry, we propose a spatial continuity regularization that distills the depth continuity priors. With the proposed depth priors, SparseNeRF achieves a new state-of-the-art performance on three datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 515, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 551 + ], + "type": "text", + "content": "Limitation. SparseNeRF significantly improves the performance of few-shot NeRF, but cannot be generalized to occluded views that are unobserved in the training views." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 552, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 658 + ], + "type": "text", + "content": "Potential Negative Impact. SparseNeRF focuses on scene synthesis. It might be misused to create misleading content. Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is also supported by Singapore MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20221-0011) and NTU NAP Grant." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 671, + 365, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 365, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 365, + 683 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srini" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "9073" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "vasan. Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In Int. Conf. Comput. Vis., 2021. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 288, + 207 + ], + "type": "text", + "content": "[3] James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: Composable transformations of Python+NumPy programs, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 288, + 262 + ], + "type": "text", + "content": "[4] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 308 + ], + "type": "text", + "content": "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MvsNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In Int. Conf. Comput. Vis., 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 308, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 288, + 342 + ], + "type": "text", + "content": "[6] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Adv. Neural Inform. Process. Syst., 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 387 + ], + "type": "text", + "content": "[7] Xipeng Chen, Guangrun Wang, Dizhong Zhu, Xiaodan Liang, Philip Torr, and Liang Lin. Structure-preserving 3D garment modeling with neural sewing machines. Adv. Neural Inform. Process. Syst., 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 388, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 431 + ], + "type": "text", + "content": "[8] Zixuan Chen, Jianhuang Lai, Lingxiao Yang, and Xiaohua Xie. CuNeRF: Cube-based neural radiance field for zero-shot medical image arbitrary-scale super resolution. Int. Conf. Comput. Vis., 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 433, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 288, + 464 + ], + "type": "text", + "content": "[9] Zhaoxi Chen and Ziwei Liu. Relighting4D: Neural re- ligable human from videos. In Eur. Conf. Comput. Vis., 2022.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "text", + "content": "[10] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Trans. Graph., 41(6):1-16, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[11] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Scenedreamer: Unbounded 3D scene generation from 2D image collections. In arXiv, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "text", + "content": "[12] Julian Chibane, Aayush Bansal, Verica Lazova, and Gerard Pons-Moll. Stereo radiance fields (srf): Learning view synthesis for sparse views of novel scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 611 + ], + "type": "text", + "content": "[13] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "text", + "content": "[14] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv:1602.02481, 2016. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[15] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3D reconstructions of indoor scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[16] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "NerDI: Single-view NeRF synthesis with language-guided diffusion as general image priors. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[18] Chenhe Dong, Guangrun Wang, Hang Xu, Jiefeng Peng, Xiaozhe Ren, and Xiaodan Liang. Efficientbert: Progressively searching multilayer perceptron via warm-up knowledge distillation. arXiv preprint arXiv:2109.07222, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "text", + "content": "[19] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., June 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst., 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 353 + ], + "type": "text", + "content": "[22] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. EVA3D: Compositional 3D human generation from 2D image collections. arXiv preprint arXiv:2210.04888, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 398 + ], + "type": "text", + "content": "[23] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human NeRF from a single image. arXiv preprint arXiv:2303.12791, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 399, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 455 + ], + "type": "text", + "content": "[24] Shoukang Hu, Kaichen Zhou, Kaiyu Li, Longhui Yu, Lanqing Hong, Tianyang Hu, Zhenguo Li, Gim Hee Lee, and Ziwei Liu. ConsistentNeRF: Enhancing neural radiance fields with 3D consistency for sparse view synthesis. arXiv preprint arXiv:2305.11031, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[25] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a diet: Semantically consistent few-shot view synthesis. In Int. Conf. Comput. Vis., 2021. 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 490, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 532 + ], + "type": "text", + "content": "[26] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In IEEE Conf. Comput. Vis. Pattern Recog., 2014. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "type": "text", + "content": "[27] Mijeong Kim, Seonguk Seo, and Bohyung Han. InfoNeRF: Ray entropy minimization for few-shot neural volume rendering. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2, 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "type": "text", + "content": "[28] Changlin Li, Bohan Zhuang, Guangrun Wang, Xiaodan Liang, Xiaojun Chang, and Yi Yang. Automated progressive learning for efficient training of vision transformers. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "text", + "content": "[29] Xiaoming Li, Wangmeng Zuo, and Chen Change Loy. Learning generative structure prior for blind text image super-resolution. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[30] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "text", + "content": "9074" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[31] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Int. Conf. Comput. Vis., 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[32] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Adv. Neural Inform. Process. Syst., 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 143, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 143, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 49, + 143, + 287, + 186 + ], + "type": "text", + "content": "[33] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. GNeRF: GAN-based neural radiance field without posed camera. In Int. Conf. Comput. Vis., 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 189, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 189, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 189, + 287, + 220 + ], + "type": "text", + "content": "[34] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In IEEE Conf. Comput. Vis. Pattern Recog., 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 223, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 223, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 49, + 223, + 287, + 277 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Trans. Graph., 38(4):1-14, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 280, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 280, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 49, + 280, + 287, + 323 + ], + "type": "text", + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Eur. Conf. Comput. Vis. Springer, 2020. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 326, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 326, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 326, + 287, + 380 + ], + "type": "text", + "content": "[37] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg-NeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 382, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 382, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 382, + 287, + 425 + ], + "type": "text", + "content": "[38] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Int. Conf. Comput. Vis., 2021. 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 428, + 287, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 428, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 49, + 428, + 287, + 471 + ], + "type": "text", + "content": "[39] Jiefeng Peng, Jiqi Zhang, Changlin Li, Guangrun Wang, Xiaodan Liang, and Liang Lin. Pi-nas: Improving neural architecture search by reducing supernet training consistency shift. In Int. Conf. Comput. Vis., 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 474, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 474, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 49, + 474, + 287, + 517 + ], + "type": "text", + "content": "[40] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 520, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 520, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 49, + 520, + 287, + 551 + ], + "type": "text", + "content": "[41] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Int. Conf. Comput. Vis., 2021. 2, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 555, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 555, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 49, + 555, + 287, + 586 + ], + "type": "text", + "content": "[42] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 589, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 589, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 589, + 287, + 622 + ], + "type": "text", + "content": "[43] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Eur. Conf. Comput. Vis. Springer, 2012. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 624, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 666 + ], + "type": "text", + "content": "[44] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Adv. Neural Inform. Process. Syst., 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[45] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[46] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2012. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[47] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-NeRF: Scalable large scene neural view synthesis. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 185, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 238 + ], + "type": "text", + "content": "[48] Tang Tao, Longfei Gao, Guangrun Wang, Peng Chen, Dayang Hao, Xiaodan Liang, Mathieu Salzmann, and Kaicheng Yu. LiDAR-NeRF: Novel lidar view synthesis via neural radiance fields. arXiv preprint arXiv:2304.10406, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 241, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 284 + ], + "type": "text", + "content": "[49] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-NeRF: Scalable construction of large-scale NeRFs for virtual fly-throughs. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 328 + ], + "type": "text", + "content": "[50] Guangrun Wang, Ping Luo, Xinjiang Wang, Liang Lin, et al. Kalman normalization: Normalizing internal representations across network layers. Adv. Neural Inform. Process. Syst., 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 373 + ], + "type": "text", + "content": "[51] Guangrun Wang and Philip HS Torr. Traditional classification neural networks are good generators: They are competitive with DDPMs and GANs. arXiv preprint arXiv:2211.14794, 2022.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 376, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 407 + ], + "type": "text", + "content": "[52] Guangrun Wang, Guangcong Wang, Keze Wang, Xiaodan Liang, and Liang Lin. Grammatically recognizing images with tree convolution. 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 410, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 442 + ], + "type": "text", + "content": "[53] Guangrun Wang, Keze Wang, and Liang Lin. Adaptively connected neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 445, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 475 + ], + "type": "text", + "content": "[54] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In Eur. Conf. Comput. Vis., 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 478, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 520 + ], + "type": "text", + "content": "[55] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 523, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 565 + ], + "type": "text", + "content": "[56] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 600 + ], + "type": "text", + "content": "[57] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 602, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 643 + ], + "type": "text", + "content": "[58] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. CityNeRF: Building nerf at city scale. arXiv preprint arXiv:2112.05504, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[59] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. SinNeRF: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[60] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 733, + 315, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 733, + 315, + 741 + ], + "spans": [ + { + "bbox": [ + 295, + 733, + 315, + 741 + ], + "type": "text", + "content": "9075" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 385 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "2D photo to a 3D object with 360deg views. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "text", + "content": "[61] Jiawei Yang, Marco Pavone, and Yue Wang. FreeNeRF: Improving few-shot neural rendering with free frequency regularization. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 174 + ], + "type": "text", + "content": "[62] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Vtoonify: Controllable high-resolution portrait video style transfer. ACM Trans. Graph., 41(6):1-15, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[63] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Gp-unit: Generative prior for versatile unsupervised image-to-image translation. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 286, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 286, + 252 + ], + "type": "text", + "content": "[64] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 286, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 297 + ], + "type": "text", + "content": "[65] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 1, 2, 3, 4, 6, 7, 8, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 286, + 342 + ], + "type": "text", + "content": "[66] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 385 + ], + "type": "text", + "content": "[67] Long Zhuo, Guangcong Wang, Shikai Li, Wanye Wu, and Ziwei Liu. Fast-vid2vid: Spatial-temporal compression for video-to-video synthesis. In Eur. Conf. Comput. Vis., 2022. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 316, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 316, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 316, + 742 + ], + "type": "text", + "content": "9076" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/d223c2f5-fe18-4502-b91f-9060d747649d_content_list.json b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/d223c2f5-fe18-4502-b91f-9060d747649d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b8f1cbd6faef337b2821d8c3051b3aa2a0a198c0 --- /dev/null +++ b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/d223c2f5-fe18-4502-b91f-9060d747649d_content_list.json @@ -0,0 +1,1733 @@ +[ + { + "type": "text", + "text": "Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Di Wu*, Pengfei Chen*, Xuehui Yu*, Guorong Li, Zhenjun Han†, Jianbin Jiao", + "bbox": [ + 179, + 180, + 790, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Chinese Academy of Sciences", + "bbox": [ + 300, + 202, + 666, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection via inaccurate bounding boxes supervision has boosted a broad interest due to the expensive high-quality annotation data or the occasional inevitability of low annotation quality (e.g. tiny objects). The previous works usually utilize multiple instance learning (MIL), which highly depends on category information, to select and refine a low-quality box. Those methods suffer from object drift, group prediction and part domination problems without exploring spatial information. In this paper, we heuristically propose a Spatial Self-Distillation based Object Detector (SSD-Det) to mine spatial information to refine the inaccurate box in a self-distillation fashion. SSD-Det utilizes a Spatial Position Self-Distillation (SPSD) module to exploit spatial information and an interactive structure to combine spatial information and category information, thus constructing a high-quality proposal bag. To further improve the selection procedure, a Spatial Identity Self-Distillation (SISD) module is introduced in SSD-Det to obtain spatial confidence to help select the best proposals. Experiments on MS-COCO and VOC datasets with noisy box annotation verify our method's effectiveness and achieve state-of-the-art performance. The code is available at https://github.com/ucas-vg/PointTinyBenchmark/tree/SSD-Det.", + "bbox": [ + 76, + 282, + 473, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 674, + 209, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection [18, 38, 58, 30, 57] relying on large-scale datasets like MS-COCO[31] has significantly progressed and achieved good performance. However, accurate bounding box annotations are expensive and challenging in natural contexts [12]. Especially in many professional scenarios, it is difficult to label accurate annotations without domain knowledge (e.g., agricultural crop observation and medical image processing) [12, 1]. As shown in Fig. 1a in some complex datasets, the human annotators may also annotate inaccurate bounding boxes due to the inherent ambiguities[20] of objects. In addition, labelling with a de", + "bbox": [ + 75, + 699, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/596443fcd81f05aca705639a72bf222290bab867abf7f86375b54859c6b93eb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 247, + 898, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb1db177af5512fe3557cb3d1d236228818d1612d98f4c281ec3df3f45101fec.jpg", + "image_caption": [ + "(a) Labelling strategies lead to inaccurate annotations (red box).", + "(b) Three problems during previous MIL refining. Because their selections depend solely on classification, (i,ii): the refined box (green box) drifts to another object or makes a group prediction (merging across multiple objects) due to neighbor disturbance. Yellow boxes are proposals in the middle person's MIL bag. (iii): Local part may be more discriminative than the entire object and will be predicted.", + "Figure 1: The sources of inaccurate box annotations and three problems caused by previous refinement methods." + ], + "image_footnote": [], + "bbox": [ + 508, + 443, + 897, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tector or weak signal[11] (e.g., point) is much cheaper but brings more inaccuracy. Therefore learning robust detectors with inaccurate bounding boxes[12, 4, 35, 53, 25] is a practical and meaningful task and has boosted a broad interest.", + "bbox": [ + 496, + 744, + 890, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To use the inaccurate annotations, most related methods [12, 11] refine the inaccurate annotations as Fig. 1a shows, and then train a detector head or re-train a detector with the refined box as the new supervision. There are two main steps during refining: 1) Bag Construction: For each object, obtaining some proposals around the inaccurate an-", + "bbox": [ + 496, + 810, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 94, + 875, + 209, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors. (hanzhj@ucas.ac.cn)", + "bbox": [ + 96, + 887, + 344, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6855", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "notated bounding box to form the object-level proposal bag; 2) Proposal Selection: Selecting the top- $k$ proposals with the highest classification confidence from each bag and then weighting average them to obtain the refined box.", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During the proposal selection, they usually utilize multiple instance learning (MIL) [10] supervised by category information to choose the proposals with high classification confidence from the constructed bags. However, they pay less [12] or no [11] attention to mining spatial information, leading to the following problems as shown in Fig. 1b: (1) Object Drift: For each object, some proposals in the constructed bag do not have a high IoU with the original object but with another nearby object. These proposals are not spatially adjacent to the original object but still have high classification confidence, as the rightmost (the yellow box) proposal of $O_{2}$ shows in the left-bottom corner of Fig. 1b. Only the category confidence is relied on for selecting proposals for $O_{2}$ , and the rightmost proposal will be selected as the refined box. It means the refined box drifts to another object (Fig. 1b (i)), reducing the recall; (2) Group Prediction: Most works [7, 11, 56] select the top- $k$ proposals by classification confidence and then weight average them as the refined box, causing the group prediction problem, as shown in Fig. 1b (ii); (3) Part Domination: The detector often focuses on the object's semantic region, which can statistically represent the category (e.g. the face). As shown in Fig. 1b (iii), the high classification confidence of the animal is in the discriminant part (the face) rather than the entire object as mentioned by [44, 39].", + "bbox": [ + 76, + 155, + 472, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these problems, we propose a Spatial Self-Distillation based detector (SSD-Det) to integrate the spatial cues into the bounding box refinement. SSD-Det has two important components: the Spatial Position Self-Distillation (SPSD) module for the bag construction step and the Spatial Identity Self-Distillation (SISD) module for the proposal selection step. To construct high-quality proposal bags, SPSD utilizes a neighborhood sampler to generate a balanced and flexible initial proposal bag for each object and then trains a regressor with the supervision of the annotated inaccurate bounding boxes. Finally, high-quality proposal bags are constructed with proposals corrected by the regressor. The mechanism behind SPSD is that the network learns the spatial information from the reliable samples, e.g. those low-noise annotations, in the dataset and then guides the noisy samples to produce high-quality proposals, as shown in Fig. 2. In addition, to further combine the category information and the spatial information, an interactive structure is implemented by alternately using SPSD to mine spatial cues and MIL to utilize the category information. With SPSD and the interactive structure, a high-quality proposal bag can be constructed. Experiments on MS-COCO show that SPSD can significantly improve the mean/max IoU between objects and proposals", + "bbox": [ + 76, + 537, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/842a39994e87835a22dd9a60227450ab9bb3ff2e3047497d078fd57f2beee266.jpg", + "image_caption": [ + "Figure 2: The mechanism of Spatial Self-Distillation. By assigning higher weight, the low-noise annotations can be seen as reliable samples to guide the training of proposals' spatial position and identity learning in SPsD and SISD." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(about 18/10 points, Fig. 5) in the constructed bag. Instead of selecting proposals by classification confidence, we have proposed the SISD module in the proposal selection step. We use it to obtain each proposal's spatial confidence by predicting the IoU with the object and combining the IoU with classification confidence to select the top- $k$ proposals. It is worth mentioning that SISD is an object-related IoU predictor, which means that the predicted IoU may be different for the same proposal that appears in different objects' bags. Accordingly, it guarantees that SISD can better handle object drift and group prediction problems. Experiments on MS-COCO and VOC datasets verify the effectiveness of our method and bring state-of-the-art performance. The contributions are as follows:", + "bbox": [ + 496, + 291, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We further investigate the inaccurate-box supervised object detection tasks and propose an end-to-end training SSD-Det that combines the spatial and the category information in an interactive fashion.", + "2) We utilize an SPSD module to generate higher-quality proposals sampling through statistic-guide spatial position distillation, raising the upper bound of the refinement.", + "3) To add spatial cues to classification confidence, we also introduce an SISD module to select a proposal belonging to the object rather than the category.", + "4) The performance of our proposed SSD-Det improves the mean average precision (AP) of the best previous method (e.g. over 10 AP on $40\\%$ noisy MS-COCO) and achieves state-of-the-art under various noise rate box supervision on MS-COCO and VOC datasets." + ], + "bbox": [ + 496, + 503, + 890, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 744, + 640, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Object Detection", + "text_level": 1, + "bbox": [ + 500, + 770, + 666, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classic object detection [15, 38, 37, 32, 30, 3, 43, 57] is supervised by an accurate bounding-box. One-stage detectors utilize anchors as the sliding-window, such asYOLO [37], SSD [32], and RetinaNet [30]. Two-stage detectors mine spatial information to predict proposals (e.g. selective search [47] in Fast R-CNN [15] or RPN in Faster R-CNN [38]) and conduct classification and bounding-box", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6856", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "regression with filtered proposals sparsely. Transformer-based (i.e. DETR [3], Deformable-DETR [62], and SwinTransformer [33]) detectors utilize global information for better representation. Sparse R-CNN [43] combines a transformer's advantages and CNNs for detection.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Weakly-Supervised Object Detection (WSOD)", + "text_level": 1, + "bbox": [ + 76, + 183, + 467, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "WSOD trains object detectors with image tag supervision. Only with the category annotation, the majority of previous methods treat each image as a bag and candidate proposals as instances. They follow the multiple instance learning (MIL) pipeline [1, 44, 46, 7, 48], which highly depends on category information. However, the MIL loss function leads to a non-convex optimization problem; thus, MIL solutions are usually stuck into the local minima. Context information [24, 51], spatial regularization [1, 9, 48], and optimization strategy [44, 48, 46] are proposed to address the problems. SPE [28] introduces Transformer into WSOD and uses attention to generate proposals. SD-LocNet [59] tackles the initialized noisy object locations in WSOD and proposes a self-directed localization network to identify the noisy object instances. [44, 46, 7] use the pseudo label for classification's iterative refinement. However, we use the pseudo box as a better self-distillation teacher. [54, 39] conduct regression to move the proposals, whereas we conduct regression to distill for better bag construction. In this work, we also formulate box correction as a MIL problem.", + "bbox": [ + 75, + 208, + 468, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Semi-Supervised Object Detection (SSOD)", + "text_level": 1, + "bbox": [ + 76, + 527, + 437, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-supervised learning in object detection can be roughly categorized into two groups: consistency based [22, 45] and pseudo label based [27, 36, 41, 49, 64, 52]. [52] presents an end-to-end SSOD approach with two simple techniques named soft teacher and box jittering to facilitate the efficient leverage of the teacher model. Both Soft Teacher [52] and SSD-Det obtain pseudo-labels from candidate boxes, adopting distillation, box jitter and classification scores weighting policy. However, they are different: 1) SSOD selects candidate boxes from the teacher model's detection results, while SSD-Det generates them with a generative approach, due to no any accurate supervised data for ensuring high-quality boxes in the detection results. 2) [52] is based on FixMatch [40] and requires maintaining two networks for teacher-student distillation structure. In contrast, our approach only needs a multi-head detector with a shared backbone for self-distillation. 3) Box Jitter: [52] calculates box variance from jittering for result selection, while we aim to generate candidate boxes that combine SISD-predicted IoU and classification scores, selecting boxes closer to the ground truth. 4) We use classification scores as weights for the next stage's loss, while [52] employs them as a criterion for selecting reliable samples.", + "bbox": [ + 75, + 553, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Learning with Noisy Annotations", + "text_level": 1, + "bbox": [ + 500, + 90, + 790, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training CNNs under noisy labels has been an active research area. Previous research focuses on the classification task, and develops various techniques to deal with noisy labels, such as sample selection [17, 23] for training, label correction [42, 34], and robust loss functions [60, 14] against noisy labels. Recently, many efforts [4, 25, 35, 53, 11, 12] have been devoted to the object detection task. On the one hand, Simon et al. [4] first investigates the impact of different types of label noise on object detection. They propose a per-object co-teaching method to alleviate the effect of noisy labels. On the other hand, [53] proposes a meta-learning framework for noisy annotations consisting of noisy category labels and bounding boxes. [11, 12] utilize object-level MIL to refine the inaccurate box. OAMIL [12] constructs proposal bags through label assignment in a discriminant style. P2BNet [11] originally conducts point-supervised object detection tasks. However, it can be seen as the box correction in its refinement stage. It uses hand-craft anchors to generate proposal bags. Our method inherits the generative style of P2BNet and conducts spatial distillation to mine spatial information.", + "bbox": [ + 496, + 113, + 890, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.5. Knowledge Distillation", + "text_level": 1, + "bbox": [ + 500, + 438, + 710, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge distillation (KD) [21] aims to learn compact and efficient student models guided by excellent teacher networks. It is first applied to object detection in [5], in which hint learning and KD are both used for multi-class object detection. Recently, many efforts [26, 50, 8, 16, 55] aim to mimic the feature. [61] shows that localization knowledge is more important and proposed a localization distillation method. We also transfer the spatial knowledge from reliable labeled instances to correct inaccurate bounding boxes (shown in Fig. 2) in a self-distillation manner.", + "bbox": [ + 496, + 460, + 890, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 623, + 633, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work aims to learn a robust detector with inaccurate bounding boxes. Instead of training a detector with the original inaccurate bounding box, we follow most related works [11, 12] that design a branch to refine the inaccurate bounding box and then train the detector head or detector with the refined bounding box. The most important part is how to design a refining policy. We first design a two-stage basic box refiner (gray region in Fig. 3) as a naive solution that modified from [11]. Then, SPSD and SISD are proposed and added to further mine the spatial cues for box refinement, yielding SSD-Det. Therefore, the overall loss function is formulated as:", + "bbox": [ + 496, + 648, + 890, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {B a s i c}} + \\alpha_ {1} \\cdot \\mathcal {L} _ {\\text {S P S D}} + \\alpha_ {2} \\cdot \\mathcal {L} _ {\\text {S I S D}} + \\alpha_ {3} \\cdot \\mathcal {L} _ {\\text {D e t}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 835, + 890, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{1},\\alpha_{2}$ and $\\alpha_{3}$ are set as 0.25, 0.25 and 4 respectively. $\\mathcal{L}_{Det}$ denotes the loss of detector or detection head. During inference, only the detector or the detection head is used.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6857", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2b0384aa55024bf0fba338b9a639ec4fbe5410b686fbc47463ede4a020def4db.jpg", + "image_caption": [ + "Figure 3: The framework of SSD-Det. It contains basic box refiner, SPSD module, SISD module and a detector head. Neighborhood sampler is adopted around the inaccurate annotation. Then, SPSD module generates better proposal bags which are fed into basic box refiner for MIL training. The selected proposals are average weighted as the refined box and supervise the next SPSD training. Meanwhile, the SISD module predicts the IoU between proposals and the object, and the estimated IoU is multiplied by classification score for better proposal selection to generate the refined box. SPSD shares backbone with the detector." + ], + "image_footnote": [], + "bbox": [ + 78, + 85, + 893, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Basic Box Refiner", + "text_level": 1, + "bbox": [ + 76, + 401, + 250, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by [12] and WSOD [1], we design the basic box refiner (detailed structure figure is in supplementary) that leverages classification confidence to refine the inaccurate box annotation. Then the refine annotation is used to train a detection head or detector. Following [11], we design a two-stream structure as a MIL classifier to select the best proposal for box refinement.", + "bbox": [ + 75, + 441, + 470, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Giving an image with inaccurate box annotation, for each object, $\\mathcal{B}$ is a bag of proposals (bounding boxes) that are generated around its inaccurate annotation by a sampler policy (e.g., selective search[47], edge box[63], neighborhood sampler in Sec. 3.2). Meanwhile a feature map is extracted with a backbone network. And then through $7\\times 7$ RoIAlign [18] and two fully connected (fc) layers, features of proposal in $\\mathcal{B}$ are extracted and denote as $\\mathbf{F}$ . The basic box refiner takes proposal bag $\\mathcal{B} \\in \\mathbb{R}^{P \\times 4}$ and features $\\mathbf{F} \\in \\mathbb{R}^{P \\times D}$ as inputs, where $P$ , $D$ are denoted as the number of proposals in $\\mathcal{B}$ , feature dimension respectively.", + "bbox": [ + 75, + 566, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [11] and [1], as Eq. 2 described, we apply the classification branch $f_{cls}$ to $\\mathbf{F}$ yields $\\mathbf{O}^{cls}$ , which is then passed through the softmax function over classification dimension $K$ to obtain the score $\\mathbf{S}^{cls} \\in \\mathbb{R}^{P \\times K}$ , where $K$ represents the number of instance categories. Likewise, instance selection branch $f_{cls}$ is applied to $\\mathbf{F}$ to yield $\\mathbf{O}^{ins}$ , and instance score $\\mathbf{S}^{ins}$ is obtained through softmax function over $P$ proposals. The proposal score $\\mathbf{S}$ is obtained by computing the Hadamard product of the classification score and the instance score. The bag score $\\widehat{\\mathbf{S}}$ is obtained by the", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "summating of the $P$ proposal boxes' proposal scores.", + "bbox": [ + 500, + 402, + 852, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} ^ {c l s} = f _ {c l s} (\\mathbf {F}) \\in \\mathbb {R} ^ {P \\times K}; [ \\mathbf {S} ^ {c l s} ] _ {p k} = e ^ {[ \\mathbf {O} ^ {c l s} ] _ {p k}} / \\sum_ {k = 1} ^ {K} e ^ {[ \\mathbf {O} ^ {c l s} ] _ {p k}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 419, + 870, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} ^ {i n s} = f _ {i n s} (\\mathbf {F}) \\in \\mathbb {R} ^ {P \\times K}; [ \\mathbf {S} ^ {i n s} ] _ {p k} = e ^ {[ \\mathbf {O} ^ {i n s} ] _ {p k}} / \\sum_ {p = 1} ^ {P} e ^ {[ \\mathbf {O} ^ {i n s} ] _ {p k}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 455, + 883, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\mathbf {S} ^ {c l s} \\odot \\mathbf {S} ^ {i n s} \\in \\mathbb {R} ^ {P \\times K}; \\widehat {\\mathbf {S}} = \\sum_ {p = 1} ^ {P} [ \\mathbf {S} ] _ {p} \\in \\mathbb {R} ^ {K}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 496, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\left[\\cdot \\right]_{pk}$ is the value at row $p$ and column $k$ in the matrix.", + "bbox": [ + 500, + 542, + 890, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The basic box refiner has two similar stages. The loss of stage I(termed $\\mathcal{L}_I$ ) adopt the MIL paradigm with the form of cross-entropy (CE) loss, defined as:", + "bbox": [ + 496, + 558, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I} = C E (\\widehat {\\mathbf {S}}, \\mathbf {c}) = - \\sum_ {k = 1} ^ {K} \\mathbf {c} _ {k} \\log (\\widehat {\\mathbf {S}} _ {k}) + (1 - \\mathbf {c} _ {k}) \\log (1 - \\widehat {\\mathbf {S}} _ {k}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 604, + 890, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{c} \\in \\{0,1\\}^K$ is the one-hot category label. And each object's proposals with the top- $k$ highest proposal score $\\mathbf{S}$ are weighted to obtain the refined box.", + "bbox": [ + 496, + 651, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The stage II takes the refined box of stage I as input and performs fine refining with a similar structure as stage I. Differently, the focal loss is adopted in stage II instead of cross entropy loss. In order to cooperate with focal loss, the classification branch uses the sigmoid $\\sigma(x)$ instead of softmax function and we sample some negative samples $\\mathcal{N}$ to further suppress the background. With the bag score $\\widehat{S}$ and the negative sample scores $\\mathbf{S}_{\\text{neg}}^{\\text{cls}}$ , the loss is:", + "bbox": [ + 496, + 696, + 890, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I I} = \\left\\langle \\mathbf {c} ^ {\\mathrm {T}}, \\widehat {\\mathbf {S}} ^ {*} \\right\\rangle \\cdot \\operatorname {F L} \\left(\\widehat {\\mathbf {S}}, \\mathbf {c}\\right) + \\sum_ {\\mathcal {N}} \\beta \\cdot \\operatorname {F L} \\left(\\mathbf {S} _ {n e g} ^ {c l s}, c _ {n e g}\\right) \\} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 824, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where FL is the focal loss [30], $\\widehat{\\mathbf{S}}_j^*$ represents the bag score predicted by stage I. $\\left\\langle \\mathbf{c}_j^{\\mathrm{T}},\\widehat{\\mathbf{S}}_j^*\\right\\rangle$ represents the inner product", + "bbox": [ + 496, + 861, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6858", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the two vectors, meaning the predicted bag score of the ground-truth category. $\\beta$ is the average of $\\left\\langle \\mathbf{c}_j^{\\mathrm{T}},\\widehat{\\mathbf{S}}_j^*\\right\\rangle$ . They are used to weight each object's FL for stable training. The overall loss function of the basic refiner here is:", + "bbox": [ + 75, + 90, + 468, + 157 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {B a s i c}} = \\mathcal {L} _ {I} + \\alpha_ {I I} \\cdot \\mathcal {L} _ {I I} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 169, + 468, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{II}$ are the loss weights of the two stages.", + "bbox": [ + 76, + 196, + 398, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, the refined box of stage II is used as supervision for a detection head or detector. After training, the basic box refiner will be removed, leaving a well-trained detection head or detector. In this way, we can train a detector under inaccurate annotations.", + "bbox": [ + 75, + 210, + 468, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Spatial Position Self-Distillation (SPSD)", + "text_level": 1, + "bbox": [ + 76, + 297, + 416, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Like most MIL paradigm methods, basic box refiner has two main components: bag construction and proposal selection. And the main idea is to use classification information to guide the refining. In this paper, we add spatial information to improve refining. Specifically, SPSD is proposed to use spatial information to enhance bag construction.", + "bbox": [ + 75, + 321, + 468, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bag construction aims to obtain proposals for each object, while proposal selection is to select the proposals from the object bag. Then, the refined box is averaged over the selected proposals. Therefore, the quality of the proposals in constructed bag determines the upbound of refining. The bag construction can be implemented in a variety of ways. In this paper, the basic box refiner adopts a naive neighborhood sampler for bag construction. basic box refiner adopts a naive neighborhood sampler for bag construction.", + "bbox": [ + 75, + 412, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neighborhood Sampler. Proposals around the inaccurate box are sampled to construct an object bag. For each inaccurate box $b^{*} = (b_{x}^{*}, b_{y}^{*}, b_{w}^{*}, b_{h}^{*})$ , its scale and aspect ratio with $s$ and $v$ are adjusted and its positions $o_{x}, o_{y}$ are jittered to obtain the diverse proposal $b = (b_{x}, b_{y}, b_{w}, b_{h})$ :", + "bbox": [ + 75, + 549, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} b _ {w} = v \\cdot s \\cdot b _ {w} ^ {*}, \\quad b _ {h} = 1 / v \\cdot s \\cdot b _ {h} ^ {*}, \\\\ b _ {x} = b _ {x} ^ {*} + b _ {w} \\cdot o _ {x}, \\quad b _ {y} = b _ {y} ^ {*} + b _ {h} \\cdot o _ {y}. \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 627, + 468, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These proposals $b$ are used to construct the positive proposal bag $\\mathcal{B}$ to train the MIL classifier. Thanks to the hand-craft sampling way, the number of proposals in different objects' proposal bags is controllable and balanced. However, the hand-craft neighborhood sampler strategy is difficult to set hyper-parameters, and the sampling space is discrete. For example, when the jitter region is small, the optimization space of refining is limited, while when it is large, more background will be introduced. Hence, we propose the SPSD module to mine spatial information for higher-quality proposal bag construction.", + "bbox": [ + 75, + 659, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Statistically Guided Adaptive Sampling. Instead of simply using a neighborhood sampler, we adopt a statistically guided adaptive sampling by adding SPSD modules into the basic box refiner. Taking the constructed proposal bag $\\mathcal{B}$ of hand-craft neighborhood sampler as input, the", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RoI features of proposals in $\\mathcal{B}$ are extracted and fed into the two shared fc layers to obtain $\\mathbf{F}$ . Then a regression fc layer $f_{dis}$ , supervised by the inaccurate annotated bounding box $b^{*}$ , is introduced to predict the adaptive proposal bag $\\mathcal{B}^{dis} = f_{dis}(\\mathbf{F}) \\in \\mathbb{R}^{P \\times 4}$ , in which the proposals are closer to the object. Later, $\\mathcal{B}^{dis}$ as the constructed proposal bag is fed into stage I of basic box refiner. In order to combine category and spatial information, we implement an interactive structure by alternately using SPSD to mine the spatial cues and using MIL in basic box refiner to utilize the category information. Specifically, the refined bounding box $\\hat{b}^{*}$ of stage I that selected by the classification confidence is used to supervision of a new SPSD module for stage II. Similar as stage I, The new SPSD takes proposal bag $\\mathcal{B}^{dis}$ of handcraft neighborhood sampler as input. Through the RoI align and the two shared fc layers, the feature $\\hat{\\mathbf{F}}$ is extracted. An extra fc layer $f_{dis}$ is then utilized to conduct further regression. Different with stage I, the obtained $\\mathcal{B}^{dis}$ is supervised by the refined $\\hat{b}^{*}$ . The loss function of the spatial distillation for adaptive sampling can be defined as $\\mathcal{L}_{SPSD}$ in Eq. 7.", + "bbox": [ + 496, + 90, + 890, + 395 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S P S D} = \\frac {1}{P} \\left\\{\\sum_ {p = 1} ^ {P} \\mathbf {L} _ {\\mathbf {1}} \\left(\\left[ \\mathcal {B} ^ {d i s} \\right] _ {p}, b ^ {*}\\right) + \\sum_ {p = 1} ^ {P} \\mathbf {L} _ {\\mathbf {1}} \\left(\\left[ \\mathcal {B} ^ {\\hat {d i s}} \\right] _ {p}, b ^ {*}\\right) \\right\\} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 401, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\mathbf{L}_1$ is the L1 loss function for loose restrictions.", + "bbox": [ + 498, + 444, + 880, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The idea behind SPSD is that the dataset with inaccurate annotation still has many reliable, high-quality boxes and inaccurate boxes. Supervised by the high-quality boxes statistically, the network can guide those proposals sampled around the inaccurate bounding box to regress to the ground truth. With the self-distillation mechanism, SPSD learns the semantic-spatial correspondence knowledge from the reliable samples in the dataset and then propagates the knowledge to produce high-quality proposals.", + "bbox": [ + 496, + 459, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive Negative Sampling. Negative samples are introduced in Stage II to better suppress the background. With the sampled $\\mathcal{B}^{dis}$ , we can adaptively sample the negative samples with a small IoU (set smaller than 0.3 by default) with all positive proposals in all bags, to compose the negative sample set $\\mathcal{N}$ for stage II.", + "bbox": [ + 496, + 595, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Spatial Identity Self-Distillation (SISD)", + "text_level": 1, + "bbox": [ + 500, + 696, + 834, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The basic box refiner selects the proposals only depending on classification confidence during proposal selection. To select the proposal which has high classification confidence and is also spatially close to the object from the bag, we propose a SISD module to predict the IoU between proposals and their corresponding object. Afterwards, through the combination between the IoU and the classification confidence, top- $k$ proposals are selected. In SISD, we design an Object Relevance Enhancement (ORE) module to distinguish different objects' features with the same RoI region. And an identity predictor is designed to predict each proposal's IoU with the object.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6859", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0467d2da66ec35147ce7c8f791b2154e76169d949b8543bd121654d0b11273f4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackbone20% Box Noise Level40% Box Noise Level
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)
Val Set
Clean-FasterRCNN [38]ResNet-5037.958.140.921.641.648.737.958.140.921.641.648.7
Clean-FasterRCNN [38]ResNet-10139.460.143.122.443.751.139.460.143.122.443.751.1
Clean-Retinanet [30]ResNet-5036.756.139.021.640.447.436.756.139.021.640.447.4
Noisy-FasterRCNN [38]ResNet-5030.454.331.417.433.938.710.328.93.35.711.815.1
Noisy-Retinanet [30]ResNet-5030.053.130.817.933.738.213.333.65.78.415.918.0
FreeAnchor[58]ResNet-5028.653.128.516.632.237.010.428.93.35.812.114.9
Co-teaching[17]ResNet-5030.554.930.517.334.039.111.531.44.26.413.116.4
SD-LocNet[59]ResNet-5030.054.530.317.533.638.711.330.34.36.012.716.6
KL loss[20]ResNet-5031.054.332.418.034.939.512.136.73.76.213.017.4
OA-MIL[12]ResNet-5032.155.333.218.135.841.618.642.612.99.219.926.5
SSD-DetResNet-5033.657.335.319.537.243.327.653.926.016.031.034.9
SSD-DetResNet-10134.357.636.719.138.144.328.454.327.216.531.936.4
SSD-Det+FRResNet-5034.457.336.820.038.244.029.354.829.017.132.936.9
SSD-Det+FRResNet-10136.259.139.220.940.247.130.656.730.718.134.539.0
Test Set
Clean-FasterRCNN [38]ResNet-5037.758.740.821.740.646.737.758.740.821.740.646.7
Noisy-FasterRCNN [38]ResNet-5030.754.931.318.033.737.710.429.03.36.011.314.6
OA-MIL[12]ResNet-5032.355.833.718.535.040.218.542.312.89.319.125.1
SSD-DetResNet-5033.557.335.519.136.041.928.054.126.516.530.034.5
SSD-Det+FRResNet-5034.757.937.220.037.742.729.755.629.317.532.436.2
", + "bbox": [ + 81, + 88, + 890, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Relevance Enhancement (ORE). ORE enhances object-relevant features, making SISD an object-relevant IoU predictor. ORE allows the predicted IoU to be different for the same proposal in other objects' bags. In addition, we integrate the feature of the bag's corresponding object into the proposal feature, making the feature of different bags' proposals distinct. That is the so-called ORE. For a proposal bag $\\mathcal{B}$ , the feature $\\mathbf{F}$ is obtained through the RoI align and two fc layers. It is worth mentioning that the two fcs do not share the parameters with those in the refiner since the optimization goals are contradictory. To represent the feature of the bag's corresponding object, $\\mathbf{F}^{+} \\in \\mathbb{R}^{1 \\times D}$ is calculated by averaging features of $P$ proposals in proposal bag $\\mathcal{B}$ . The object feature $\\mathbf{F}^{+}$ is broadcast into $\\mathbb{R}^{P \\times D}$ , and then added to the proposal features to obtain the object-relevant features $\\mathbf{F}^{*} = \\mathbf{F} + \\mathbf{F}^{+}$ .", + "bbox": [ + 75, + 476, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial Identity Prediction. By a following identity fc layer, $U \\in \\mathbb{R}^{P \\times 1}$ is predicted. The pseudo label $T \\in (0,1)$ is IoU between proposals in $\\mathcal{B}$ and the merged box $\\hat{b}^*$ of stage I. For better optimization, the linear normalized $T' = (T - 0.5)/0.5 \\in (-1,1)$ is utilized as supervision. The object function of the identity predictor is identified in:", + "bbox": [ + 76, + 717, + 468, + 811 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S I S D} = \\operatorname {s m o o t h} _ {L 1} \\left(U, T ^ {\\prime}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 816, + 468, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the smooth $L_{1}$ represents the smooth L1 loss. The predicted spatial confidence $U^{\\prime}$ is obtained by normalizing the $U$ . Finally, $S^{*} = U^{\\prime} \\cdot S$ is used to select the top- $k$ proposals for merging as the refined boxes.", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6abbc608d5c12e1e03bd179d7de5d314febc497bc1cc7cb46909642f27fa35c6.jpg", + "table_caption": [ + "Table 1: Performance comparison on COCO. FR is Faster R-CNN. *-FR refers to a retrained Faster R-CNN (R50+FPN) using refined annotations from SSD-Det for improved performance.. Clean-* and Noisy-* means original and noisy annotation." + ], + "table_footnote": [], + "table_body": "
MethodBackboneBox Noise Level
10%20%30%40%
Clean-FasterRCNN [38]ResNet-5077.2 for clean
Clean-RetinaNet [30]ResNet-5073.5 for clean
Noisy-FasterRCNN [38]ResNet-5076.371.260.142.5
Noisy-RetinaNet [30]ResNet-5071.567.657.945.0
KL loss[20]ResNet-5075.872.764.648.6
Co-teaching[17]ResNet-5075.470.660.943.7
SD-LocNet[59]ResNet-5075.771.560.843.9
FreeAnchor[58]ResNet-5073.067.556.241.6
OA-MIL[12]ResNet-5077.474.370.663.8
SSD-DetResNet-5077.174.871.566.9
", + "bbox": [ + 506, + 473, + 888, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Performance comparison on the VOC 2007 test set. The evaluation metric is $AP_{50}$ . The Clean-* and Noisy-* means original annotation and noisy annotation.", + "bbox": [ + 496, + 637, + 890, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 710, + 624, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 737, + 705, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. For experimental comparisons, two publicly available datasets are used for object detection with inaccurate bounding boxes: MS-COCO [31] and PASCAL VOC 2007 [13]. MS-COCO (2017 version) has 118k training and 5k validation images with 80 common object categories. PASCAL VOC 2007 is one of the most popular benchmarks in generic object detection with 20 classes.", + "bbox": [ + 496, + 762, + 890, + 882 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. We use mean average precision", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6860", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/53bc81990f8917d1a72fd51326044c106320f8dabf4201e29bd31b1ef660dbc1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
2-RefSPSDSISDRe-Train20% Box Noise Level40% Box Noise Level
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)\\(AP^{test}\\)\\(AP^{test}_{75}\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)\\(AP^{test}\\)\\(AP^{test}_{75}\\)
30.057.129.016.933.139.8--22.851.116.113.325.030.4--
✓31.256.731.617.834.541.031.432.024.652.020.114.328.231.925.020.5
✓✓33.056.934.818.735.542.233.134.827.253.724.715.930.335.227.625.6
✓✓✓33.657.335.319.537.243.333.535.527.653.926.016.031.034.928.026.5
✓✓31.856.833.118.435.740.832.333.726.554.023.315.730.333.826.823.3
✓✓✓34.157.636.419.037.743.834.336.629.055.127.817.032.536.729.328.4
✓✓✓✓34.457.336.820.038.244.034.737.229.354.829.017.132.936.929.729.3
", + "bbox": [ + 81, + 88, + 893, + 207 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a5285f54cd48f3a17c3f347e76618bbd3cc37e66da65a02663bb47c172a6fc3.jpg", + "table_caption": [ + "Table 3: Modules ablation of SPSD, SISD and Re-Train on MS-COCO validation set (without) and test set (with test). The Re-Train means we generate the pseudo label by SSD-Det and re-train a Faster R-CNN detector." + ], + "table_footnote": [], + "table_body": "
Methods (w/o SISD)APAP50AP75APsAPmAPl
Neighborhood Sampler24.652.020.114.328.231.9
SPSD (II) w/o weighted26.053.322.515.629.433.4
SPSD (II) w/ weighted26.353.422.515.629.333.8
SPSD (I+II) w/ weighted27.253.724.715.930.335.2
", + "bbox": [ + 84, + 271, + 465, + 334 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fe09d1c374a53296449568c1295365f508cec2ad7b5f680698ee64e6c8499c51.jpg", + "table_caption": [ + "Table 4: Different setting of SPSD." + ], + "table_footnote": [], + "table_body": "
ORE Strategies of SISDAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP^s \\)\\( AP^m \\)\\( AP^l \\)
w/o SISD27.253.724.715.930.335.2
SISD w/o ORE27.353.325.716.930.235.0
+ subtract27.253.824.615.730.435.0
+ concatenate27.254.024.516.130.234.8
+ add27.653.926.016.031.034.9
+ add w/ shared fcs23.049.917.612.625.330.4
", + "bbox": [ + 84, + 378, + 460, + 465 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1ab558621bc2291da62176a7fe49804e3afd6e1f3574544c60796aa3e6f1e017.jpg", + "table_caption": [ + "Table 5: Different ORE strategies of SISD." + ], + "table_footnote": [], + "table_body": "
Num.0123
AP/AP5024.6 / 52.026.3 / 53.427.2 / 53.727.0 / 53.1
", + "bbox": [ + 81, + 508, + 467, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6: Number of SPSD module.", + "bbox": [ + 155, + 565, + 387, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mAP@[.5,.95] and (mAP@.5) for MS-COCO and VOC. The $\\{AP, AP_{50}, AP_{75}, AP^{Small}, AP^{Middle}, AP^{Large}\\}$ is reported for MS-COCO and $AP_{50}$ for VOC.", + "bbox": [ + 76, + 609, + 467, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic Noisy Dataset. Following [12], We simulate noisy bounding boxes by perturbing clean boxes from the original annotations. The details are in the appendix. We simulate various box noise levels ranging from $10\\%$ to $40\\%$ for the VOC and $\\{20\\%, 40\\}$ for the MS-COCO.", + "bbox": [ + 76, + 657, + 467, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We implement our method on FasterRCNN [38] with ResNet50-FPN [19, 29] backbone, based on MMDetection [6]. All settings of our method and previous methods employ FPN for fair comparison. Similar to the default setting of object detection on MS-COCO, the stochastic gradient descent [2] algorithm is used to optimize on 1x training schedule. The batch size is two images per GPU on 8 GPUS. For the VOC dataset, the batch size is two images per GPU on 2 GPUS. The performance we report is on a single scale (1333 * 800 for MS-COCO and 1000 * 600 for VOC).", + "bbox": [ + 75, + 734, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art", + "text_level": 1, + "bbox": [ + 500, + 273, + 795, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with several state-of-the-art approaches [20, 17, 59, 58, 12] on MS-COCO and VOC 2007 datasets. We denote Clean-FasterRCNN and Noisy-FasterRCNN as FasterRCNN models trained under clean (original annotations) and noisy annotations with the default setting, respectively.", + "bbox": [ + 498, + 306, + 890, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MS-COCO Dataset. Table 1 shows the comparison results on the MS-COCO. Inaccurate bounding box annotations significantly deteriorate the vanilla Faster R-CNN's detection performance. Co-teaching and SD-LocNet only slightly improve the detection performance, especially under $40\\%$ box noise. That indicates that small-loss sample selection and sample weight assignment can not tackle noisy box annotations well. KL Loss slightly improves the performance under $20\\%$ and $40\\%$ box noise. By treating an object as a bag of instances, OA-MIL is somehow robust to noisy bounding boxes and performs better than other methods. Nevertheless, the previously-mentioned label assignment bag construction limits its ability to handle heavy noise. Our approach is more robust to noisy bounding boxes. It outperforms other methods by a large margin under high box noise levels and significantly boosts the detection performance across all metrics. For example, under $40\\%$ box noise, the end-to-end SSD-Det achieves 27.6 $AP$ and $53.9 AP_{50}$ , 9.0 and attains 11.3 point improvement compared with state-of-the-art method OA-MIL, respectively. Also, through re-training on FasterRCNN, the performance further reaches 29.3 $AP$ and $54.8 AP_{50}$ . With the backbone of ResNet-101, the performance achieves consistent improvement. On MS-COCO test set, our method also achieves state-of-the-art performance.", + "bbox": [ + 496, + 407, + 892, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VOC 2007 Dataset. Table 2 shows the comparison results on the VOC 2007 test set. Co-teaching, SD-LocNet and KL Loss, can not address inaccurate bounding box annotations well. OA-MIL improves the performance on different noisy datasets. Our approach obtains further improvements to 77.10, 74.80, 71.50, 66.90 $AP_{50}$ on $10\\%$ , $20\\%$ , $30\\%$ and $40\\%$ noisy box datasets, respectively.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6861", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b82a78ea638bd787cba4f76d854879baa5c65554b6052c89d801317cd5335c24.jpg", + "image_caption": [ + "Figure 4: Qualitative detection results on COCO validation. Previous methods miss objects and face part prediction problems. Our method misses fewer objects, and the bounding box quality is better, especially for small or overlapped objects." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 367, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c42efc6fd3a318138ab97195e2f6e7c68a77bf843ba4470add0c21737a08ff89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 89, + 490, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e91598751ae71b1b39edb2b75603285d83032dec869312aed2bd3b38ebb1ac5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 90, + 614, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93cb516e89418507075d332120a024065beccae3bea40e96ec6861433fc1725b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 90, + 890, + 282 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a000cf5a9d388a0d3dabc608b3494d3678b7c0ef8e14a2205d3c65bd2b892e93.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsBox Refiner+Re-TrainSSD-DetSSD-Det+Re-Train
AP/AP5029.0 / 54.427.6/53.929.3 / 54.8
", + "bbox": [ + 84, + 349, + 465, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 76, + 431, + 331, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further analyze SSD-Det's effectiveness and robustness, we conduct more experiments on COCO val set if there are no other instructions. Except for Table 3, the noise level of these experiments is $40\\%$ .", + "bbox": [ + 75, + 457, + 468, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation of Modules. Ablation study of each component in our approach is given in Table 3, including: (i) Different stages of our basic box refiner. i.e. training object detector without the stage II (2-Ref), where the pseudo boxes predicted by the stage I are served as the supervision for training a parallel detector. (ii) SPSD, i.e. training without SPSD, where the object-bag is constructed directly by neighborhood sampling around the noisy ground-truth or the predicted pseudo boxes of the stage I. (iii) SISD. (iv) Re-Train with FasterRCNN (Re-Train).", + "bbox": [ + 75, + 518, + 468, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of SPSD. SPSD further improves the detection performance on the MS-COCO, especially under high box noise levels, e.g. under $40\\%$ box noise level, SPSD boosts the performance from 24.6 to 27.2, as shown in Table 3 (row 3). In Table 4, we conduct further ablation on SPSD. With SPSD bag construction only in stage II, the performance increases by 1.4 AP. The performance further improves with the proposal score of stage I as weights. With SPSD in all stages, the AP reaches 27.2. Fig. 5 shows the bag quality. With SPSD, the mean IoU increases from 40.3 to 58.7 and the max and top-10 IoU increase to 78.3 and 75.1, which indicates a better upper bound of proposal selection. More high-quality proposals bring better optimization and easier proposals selection.", + "bbox": [ + 75, + 671, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of SPSD. As shown in Table 6. When adding", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/57db3089771a1f6f47b394dc3abd9b06623b49db9091445d138ad8c19f27a83d.jpg", + "table_caption": [ + "Table 7: Comparisons of end-to-end and Re-Train." + ], + "table_footnote": [], + "table_body": "
DetectorsClean-supervisedNoise-supervised
APAP50APAP50(w/ours)
APAP50
FasterRCNN37.958.110.328.929.354.8
SparseRCNN [43]†45.064.16.020.334.360.2
De-DETR [62]†46.866.35.016.935.260.9
", + "bbox": [ + 504, + 349, + 890, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8: Experiments on advanced detectors. De-DETR is Deformable DETR. † uses multi-scale data augment. ‘ $w/$ ours’ means using our method under noisy supervision.", + "bbox": [ + 498, + 446, + 892, + 493 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3095a5b20676f90b4c2cb6035846f78ae76c3199ff9b55c994e70cb8855d99b9.jpg", + "image_caption": [ + "Figure 5: Bag quality (IoU of proposals with GT) of construction in SSD-Det. B.S. (blue) means neighborhood sampler. SPSD I (orange) denotes single SPSD adopted. SPSD II (yellow) is two SPSD and interactive structure adopted. SPSD II significantly improves the quality." + ], + "image_footnote": [], + "bbox": [ + 501, + 515, + 890, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3 SPSD, performance drops slightly, probably due to the accumulation of errors outweighing the performance gain from extra stages. Hence, 2 SPSD is our default setting.", + "bbox": [ + 498, + 776, + 890, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of SISD. SISD is designed to select object-aware proposals in box selection. Under $40\\%$ and $20\\%$ box noise, the detection performance improves from 27.2 to 27.6 and 33.0 to 33.6, which verifies the effectiveness of the module, shown in Table 3. We also study the", + "bbox": [ + 498, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6862", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ed4b84a6904b6b70f13626687796e53830309a57b6a6369a502ac0a870ee1102.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsDrift Rate (%)Group Rate (%)Part Rate (%)
allsmlallsmlallsml
OA-MIL[12]15.117.817.17.46.72.83.41.42.83.42.72.3
Ours1.51.01.31.41.71.20.50.71.00.51.11.3
", + "bbox": [ + 81, + 89, + 467, + 141 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/44d913ab62a0ab0bacbfcaeaddf08a26f451a6d6c1cc513f9377661d50b2955b.jpg", + "table_caption": [ + "Table 9: Breakdown of different problems during refinement (COCO under $40\\%$ noise level). s, m and 1 mean small, middle and large scale." + ], + "table_footnote": [], + "table_body": "
DatasetourQuality (Average IoU)Frequency (%)
AllPartOversizeShiftReliablePartOversizeShift
Detector as Annotator
Objects-F✓44.324.133.613.340.128.49.821.7
47.030.841.816.849.223.18.019.4
COCO-F✓45.125.433.315.740.027.512.020.5
48.232.140.619.649.522.210.018.3
Point-based Annotator
COCO-P✓55.630.425.029.865.69.523.21.7
65.246.736.440.974.95.918.50.6
", + "bbox": [ + 81, + 227, + 465, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 10: Analysis of noisy annotations types and quality.", + "bbox": [ + 81, + 340, + 460, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "strategies of ORE in SISD (Table 5). The minus or concatenation of object feature $\\mathbf{F}_j^+$ and proposal feature $\\mathbf{F}_j$ do not work. With add strategy, the performance is 27.60. If SISD shares the two fc layers, the performance drops to 22.99 since the optimization goals are contradictory (Identity distinguishes objects in the same category). If we directly use the RoI feature without ORE, the performance drops to 27.32 AP, verifying the effectiveness of the object relevance strategy.", + "bbox": [ + 75, + 386, + 467, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Affect of Re-Train. As most WSOD methods do, we re-run the experiments by training a fully supervised detector for better performance. We find that if the SSD-Det only trains the refiner and uses the pseudo label to train the FasterRCNN, the result is good but lower than re-train after the end-to-end training given in Table 7 (row 1). This is because joint training is beneficial for box refinement.", + "bbox": [ + 75, + 508, + 467, + 614 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Experiments on Advanced Detectors. We re-train recent detectors, e.g. SparseRCNN and Deformable DETR, under the boxes refined by our method. Table 8 verifies that our method achieves consistency improvement.", + "bbox": [ + 75, + 617, + 467, + 678 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The computational cost discussion. Similar to OAMIL, our method adds SSD head to Faster R-CNN for auxiliary training to refine the noisy annotation and the head is not used during inference, the calculation cost during inference is same to standard Faster R-CNN.", + "bbox": [ + 75, + 679, + 467, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4. Visualization and Discussion.", + "text_level": 1, + "bbox": [ + 76, + 768, + 334, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 4 shows that OA-MIL faces missing instances and grouping instances issues for small or overlapped objects (as mentioned in [12]), while our method still works well. For a better intuitive understanding of SISD and SPSD, we visualize the bag construction quality in Fig. 5 Then, we makes noise types breakdown of 'Drift', 'Group' and 'part dominance' issues. We give the definition of IoU, IoG and", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "IoD:", + "bbox": [ + 500, + 90, + 539, + 104 + ], + "page_idx": 8 + }, + { + "type": "equation", + "text": "\n$$\nI o U = \\frac {A (I)}{A (D) + A (G) - A (I)}, I o G = \\frac {A (I)}{A (G)}, I o D = \\frac {A (I)}{A (D)} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 114, + 890, + 141 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "where $\\mathrm{A}(\\ast)$ is area of box *, D and G are refined box and gt box respectively, and I is insertion between D and G. We statistically count the proportion of three noise types of 'bad' refined boxes (having small IoU with gt) in Table 9: (i) Drift: 'bad' refined box has a higher IoU with another nearby object. (ii) Group: 'bad' refined box has high IoG with multiple objects. (iii) Part: 'bad' refined box has a high IoD. Table 9 shows quantitative results for each noise type of baseline and ours. The drift, group, part problems reduce from $15.1\\%$ , $6.7\\%$ , $2.8\\%$ to $1.5\\%$ , $1.7\\%$ , $1.0\\%$ , respectively, demonstrating our improvement.", + "bbox": [ + 496, + 152, + 890, + 319 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Experiments on Real-life Noisy Annotations. Real-life noisy annotations stem from: low-quality data (e.g., occlusion, blur), human annotator errors and automatic machine annotator limitations. Noise from human errors is quite subjective, since differences between annotators. For a more objective analysis, noisy annotations from machine annotator is used for experiments. Without loss of generality, Faster R-CNN well-trained on MS-COCO was applied to Objects365 images, yielding Objects-F dataset, and to COCO-val images, producing COCO-F dataset. P2BNet [11], a point-based annotator, was used on COCO-val images with point annotations, generating COCO-P dataset. SSD-Det effectively improves low-quality boxes. As shown in Table 10, with SSD-Det's refinement, the average IoU increases for Objects-F (from 44.3 to 47.0), COCO-F (from 45.1 to 48.2) and COCO-P (from 55.6 to 65.2) datasets. Further, the proportion of reliable annotations increases, and noise categories' frequency (Part, Oversize, and Shift) decreases for all datasets.", + "bbox": [ + 496, + 319, + 890, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 623, + 617, + 638 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper investigates problems during refinement caused by solely using category information to select proposals. We also propose SSD-Det to mine spatial information in a self-distillation fashion. SSD-Det introduces the SPSD module to learn semantic-spatial correspondence knowledge with neighborhood sampler and an interactive structure to combine spatial information and category information, thus producing a high-quality proposal bag. SISD in SSD-Det is utilized to improve the proposal selection procedure by integrating object-relevant spatial confidence. Complete ablations on multiple datasets verify the effectiveness of SSD-Det.", + "bbox": [ + 496, + 648, + 890, + 829 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 845, + 596, + 861 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[1] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In CVPR, 2016. 1, 3, 4", + "bbox": [ + 506, + 872, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6863", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[2] Léon Bottou. Stochastic gradient descent tricks. In Neural Networks: Tricks of the Trade - Second Edition. Springer, 2012. 7", + "[3] Nicolas Carion, Francisco Massa, and Gabriel Synnaeve et al. End-to-end object detection with transformers. In ECCV, 2020. 2, 3", + "[4] Simon Chadwick and Paul Newman. Training object detectors with noisy data. In IV, 2019. 1, 3", + "[5] Guobin Chen, Wongun Choi, and Xiang Yu et al. Learning efficient object detection models with knowledge distillation. In NeurIPS, 2017. 3", + "[6] Kai Chen, Jiaqi Wang, and Jiangmiao Pang et al. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7", + "[7] Ze Chen, Zhihang Fu, and Rongxin Jiang et al. SLV: spatial likelihood voting for weakly supervised object detection. In CVPR, 2020. 2, 3", + "[8] Xing Dai, Zeren Jiang, and Zhao Wu et al. General instance distillation for object detection. In CVPR, 2021. 3", + "[9] Ali Diba, Vivek Sharma, and Ali Mohammad Pazandeh et al. Weakly supervised cascaded convolutional networks. In CVPR, 2017. 3", + "[10] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the multiple instance problem with axis-parallel rectangles. Artificial Intelligence, 1997. 2", + "[11] Chen, et al. Point-to-box network for accurate object detection via single point supervision. In ECCV, 2022. 1, 2, 3, 4, 9", + "[12] Liu, et al. Robust object detection with inaccurate bounding boxes. In ECCV, 2022. 1, 2, 3, 4, 6, 7, 9", + "[13] Mark Everingham, Luc Van Gool, and Christopher K. I. Williams et al. The pascal visual object classes (VOC) challenge. *IJCV*, 2010. 6", + "[14] Aritra Ghosh, Himanshu Kumar, and P. S. Sastry. Robust loss functions under label noise for deep neural networks. In AAAI, 2017. 3", + "[15] Ross B. Girshick. Fast R-CNN. In ICCV, 2015. 2", + "[16] Jianyuan Guo, Kai Han, and Yunhe Wang et al. Distilling object detectors via decoupled features. In CVPR, 2021. 3", + "[17] Bo Han, Quanming Yao, and Xingrui Yu et al. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, 2018. 3, 6, 7", + "[18] Kaiming He, Georgia Gkioxari, and Piotr Dolkar et al. Mask R-CNN. In ICCV, 2017. 1, 4", + "[19] Kaiming He, Xiangyu Zhang, and Shaoqing Ren et al. Deep residual learning for image recognition. In CVPR, 2016. 7", + "[20] Yihui He, Chenchen Zhu, and Jianren Wang et al. Bounding box regression with uncertainty for accurate object detection. In CVPR, 2019. 1, 6, 7", + "[21] Geoffrey Hinton and Oriol et al. Vinyls. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 3", + "[22] Jisoo Jeong, Seungeui Lee, and Jeesoo Kim et al. Consistency-based semi-supervised learning for object detection. In NeurIPS, 2019. 3", + "[23] Lu Jiang, Zhengyuan Zhou, and Thomas Leung et al. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 3" + ], + "bbox": [ + 78, + 92, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Vadim Kantorov, Maxime Aquab, and Minsu Cho et al. Contextlocnet: Context-aware deep network models for weakly supervised localization. In ECCV, 2016. 3", + "[25] Junnan Li, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Towards noise-resistant object detection with noisy annotations. CoRR, abs/2003.01285, 2020. 1, 3", + "[26] Quanquan Li, Shengying Jin, and Junjie Yan. Mimicking very efficient network for object detection. In CVPR, 2017. 3", + "[27] Yandong Li, Di Huang, and Danfeng Qin et al. Improving object detection with selective self-supervised self-training. In ECCV, 2020. 3", + "[28] Mingxiang Liao, Fang Wan, and Yuan Yao et al. End-to-end weakly supervised object detection with sparse proposal evolution. In ECCV, 2022. 3", + "[29] Tsung-Yi Lin, Piotr Dólar, and Ross B. Girshick et al. Feature pyramid networks for object detection. In CVPR, 2017. 7", + "[30] Tsung-Yi Lin, Priya Goyal, and Ross B. Girshick et al. Focal loss for dense object detection. In ICCV, 2017. 1, 2, 4, 6", + "[31] Tsung-Yi Lin and Michael et al. Maire. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6", + "[32] Wei Liu, Dragomir Anguelov, and Dumitru Erhan et al. SSD: single shot multibox detector. In ECCV, 2016. 2", + "[33] Ze Liu, Yutong Lin, and Yue Cao et al. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3", + "[34] Xingjun Ma, Yisen Wang, and Michael E. Houle et al. Dimensionality-driven learning with noisy labels. In ICML, 2018. 3", + "[35] Jiafeng Mao, Qing Yu, and Kiyoharu Aizawa. Noisy localization annotation refinement for object detection. TIS, 2021. 1, 3", + "[36] Ilija Radosavovic, Piotr Dólar, and Ross B. Girshick et al. Data distillation: Towards omni-supervised learning. In CVPR, 2018. 3", + "[37] Joseph Redmon, Santosh Kumar Divvala, and Ross B. Girshick et al. You only look once: Unified, real-time object detection. In CVPR, 2016. 2", + "[38] Shaoqing Ren, Kaiming He, and Ross B. Girshick et al. Faster R-CNN: towards real-time object detection with region proposal networks. TPAMI, 2017. 1, 2, 6, 7", + "[39] Zhongzheng Ren, Zhiding Yu, and Xiaodong Yang et al. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In CVPR, 2020. 2, 3", + "[40] Kihyuk Sohn, David Berthelot, and Nicholas Carlini et al. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 3", + "[41] Kihyuk Sohn, Zizhao Zhang, and Chun-Liang Li et al. A simple semi-supervised learning framework for object detection. CoRR, 2020. 3", + "[42] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. SELFIE: refurbishing unclean samples for robust deep learning. In ICML, 2019. 3", + "[43] Peize Sun, Rufeng Zhang, and Yi Jiang et al. Sparse R-CNN: end-to-end object detection with learnable proposals. In CVPR, 2021. 2, 3, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6864", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Peng Tang and Xinggang Wang et al. Multiple instance detection network with online instance classifier refinement. In CVPR, 2017. 2, 3", + "[45] Peng Tang, Chetan Ramaiah, and Yan Wang et al. Proposal learning for semi-supervised object detection. In WACV, 2021. 3", + "[46] Peng Tang, Xinggang Wang, and Song Bai et al. PCL: proposal cluster learning for weakly supervised object detection. TPAMI, 2020. 3", + "[47] Koen E. A. van de Sande, Jasper R. R. Uijlings, and Theo Gevers et al. Segmentation as selective search for object recognition. In ICCV, 2011. 2, 4", + "[48] Fang Wan, Pengxu Wei, and Zhenjun Han et al. Min-entropy latent model for weakly supervised object detection. TPAMI, 2019. 3", + "[49] Keze Wang, Xiaopeng Yan, and Dongyu Zhang et al. Towards human-machine cooperation: Self-supervised sample mining for object detection. In CVPR, 2018. 3", + "[50] Tao Wang, Li Yuan, and Xiaopeng Zhang et al. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 3", + "[51] Yunchao Wei, Zhiqiang Shen, and Bowen Cheng et al. TS ^2 2 C: tight box mining with surrounding segmentation context for weakly supervised object detection. In ECCV, 2018. 3", + "[52] Mengde Xu, Zheng Zhang, and Han Hu et al. End-to-end semi-supervised object detection with soft teacher. In ICCV. IEEE, 2021. 3", + "[53] Youjiang Xu, Linchao Zhu, and Yi Yang et al. Training robust object detectors from noisy category labels and imprecise bounding boxes. TIP, 2021. 1, 3", + "[54] Ke Yang, Dongsheng Li, and Yong Dou. Towards precise end-to-end weakly supervised object detection network. In ICCV, 2019. 3", + "[55] Zhendong Yang, Zhe Li, and Xiaohu Jiang et al. Focal and global knowledge distillation for detectors. In CVPR, 2022. 3", + "[56] Xuehui Yu, Pengfei Chen, and Di Wu et al. Object localization under single coarse point supervision. In CVPR, 2022. 2", + "[57] Xuehui Yu, Yuqi Gong, and Nan Jiang et al. Scale match for tiny person detection. In WACV, 2020. 1, 2", + "[58] Xiaosong Zhang, Fang Wan, and Chang Liu et al. Freeanchor: Learning to match anchors for visual object detection. In NeurIPS, 2019. 1, 6, 7", + "[59] Xiaopeng Zhang, Yang Yang, and Jiashi Feng. Learning to localize objects with noisy labeled instances. In AAAI, 2019. 3, 6, 7", + "[60] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In NeurIPS, 2018. 3", + "[61] Zhaohui Zheng, Rongguang Ye, and Ping Wang et al. Localization distillation for dense object detection. In CVPR, 2022. 3", + "[62] Xizhou Zhu, Weijie Su, and Lewei Lu et al. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] C. Lawrence Zitnick and Piotr Dálár. Edge boxes: Locating object proposals from edges. In ECCV, 2014. 4", + "[64] Barret Zoph, Golnaz Ghiasi, and Tsung-Yi Lin et al. Rethinking pre-training and self-training. In NeurIPS, 2020. 3" + ], + "bbox": [ + 501, + 90, + 890, + 148 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6865", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/d223c2f5-fe18-4502-b91f-9060d747649d_model.json b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/d223c2f5-fe18-4502-b91f-9060d747649d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e34b3e1f257e08f250df08c0b7a333e781f3946 --- /dev/null +++ b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/d223c2f5-fe18-4502-b91f-9060d747649d_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.181, + 0.791, + 0.201 + ], + "angle": 0, + "content": "Di Wu*, Pengfei Chen*, Xuehui Yu*, Guorong Li, Zhenjun Han†, Jianbin Jiao" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.203, + 0.668, + 0.217 + ], + "angle": 0, + "content": "University of Chinese Academy of Sciences" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.283, + 0.474, + 0.646 + ], + "angle": 0, + "content": "Object detection via inaccurate bounding boxes supervision has boosted a broad interest due to the expensive high-quality annotation data or the occasional inevitability of low annotation quality (e.g. tiny objects). The previous works usually utilize multiple instance learning (MIL), which highly depends on category information, to select and refine a low-quality box. Those methods suffer from object drift, group prediction and part domination problems without exploring spatial information. In this paper, we heuristically propose a Spatial Self-Distillation based Object Detector (SSD-Det) to mine spatial information to refine the inaccurate box in a self-distillation fashion. SSD-Det utilizes a Spatial Position Self-Distillation (SPSD) module to exploit spatial information and an interactive structure to combine spatial information and category information, thus constructing a high-quality proposal bag. To further improve the selection procedure, a Spatial Identity Self-Distillation (SISD) module is introduced in SSD-Det to obtain spatial confidence to help select the best proposals. Experiments on MS-COCO and VOC datasets with noisy box annotation verify our method's effectiveness and achieve state-of-the-art performance. The code is available at https://github.com/ucas-vg/PointTinyBenchmark/tree/SSD-Det." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.675, + 0.21, + 0.69 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.471, + 0.866 + ], + "angle": 0, + "content": "Object detection [18, 38, 58, 30, 57] relying on large-scale datasets like MS-COCO[31] has significantly progressed and achieved good performance. However, accurate bounding box annotations are expensive and challenging in natural contexts [12]. Especially in many professional scenarios, it is difficult to label accurate annotations without domain knowledge (e.g., agricultural crop observation and medical image processing) [12, 1]. As shown in Fig. 1a in some complex datasets, the human annotators may also annotate inaccurate bounding boxes due to the inherent ambiguities[20] of objects. In addition, labelling with a de" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.248, + 0.9, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.427, + 0.89, + 0.442 + ], + "angle": 0, + "content": "(a) Labelling strategies lead to inaccurate annotations (red box)." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.444, + 0.898, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.586, + 0.9, + 0.671 + ], + "angle": 0, + "content": "(b) Three problems during previous MIL refining. Because their selections depend solely on classification, (i,ii): the refined box (green box) drifts to another object or makes a group prediction (merging across multiple objects) due to neighbor disturbance. Yellow boxes are proposals in the middle person's MIL bag. (iii): Local part may be more discriminative than the entire object and will be predicted." + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.682, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Figure 1: The sources of inaccurate box annotations and three problems caused by previous refinement methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.807 + ], + "angle": 0, + "content": "tector or weak signal[11] (e.g., point) is much cheaper but brings more inaccuracy. Therefore learning robust detectors with inaccurate bounding boxes[12, 4, 35, 53, 25] is a practical and meaningful task and has boosted a broad interest." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.902 + ], + "angle": 0, + "content": "To use the inaccurate annotations, most related methods [12, 11] refine the inaccurate annotations as Fig. 1a shows, and then train a detector head or re-train a detector with the refined box as the new supervision. There are two main steps during refining: 1) Bag Construction: For each object, obtaining some proposals around the inaccurate an-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.21, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.345, + 0.901 + ], + "angle": 0, + "content": "† Corresponding authors. (hanzhj@ucas.ac.cn)" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.345, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6855" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "notated bounding box to form the object-level proposal bag; 2) Proposal Selection: Selecting the top- \\(k\\) proposals with the highest classification confidence from each bag and then weighting average them to obtain the refined box." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.156, + 0.473, + 0.534 + ], + "angle": 0, + "content": "During the proposal selection, they usually utilize multiple instance learning (MIL) [10] supervised by category information to choose the proposals with high classification confidence from the constructed bags. However, they pay less [12] or no [11] attention to mining spatial information, leading to the following problems as shown in Fig. 1b: (1) Object Drift: For each object, some proposals in the constructed bag do not have a high IoU with the original object but with another nearby object. These proposals are not spatially adjacent to the original object but still have high classification confidence, as the rightmost (the yellow box) proposal of \\( O_{2} \\) shows in the left-bottom corner of Fig. 1b. Only the category confidence is relied on for selecting proposals for \\( O_{2} \\), and the rightmost proposal will be selected as the refined box. It means the refined box drifts to another object (Fig. 1b (i)), reducing the recall; (2) Group Prediction: Most works [7, 11, 56] select the top- \\( k \\) proposals by classification confidence and then weight average them as the refined box, causing the group prediction problem, as shown in Fig. 1b (ii); (3) Part Domination: The detector often focuses on the object's semantic region, which can statistically represent the category (e.g. the face). As shown in Fig. 1b (iii), the high classification confidence of the animal is in the discriminant part (the face) rather than the entire object as mentioned by [44, 39]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.539, + 0.473, + 0.904 + ], + "angle": 0, + "content": "To address these problems, we propose a Spatial Self-Distillation based detector (SSD-Det) to integrate the spatial cues into the bounding box refinement. SSD-Det has two important components: the Spatial Position Self-Distillation (SPSD) module for the bag construction step and the Spatial Identity Self-Distillation (SISD) module for the proposal selection step. To construct high-quality proposal bags, SPSD utilizes a neighborhood sampler to generate a balanced and flexible initial proposal bag for each object and then trains a regressor with the supervision of the annotated inaccurate bounding boxes. Finally, high-quality proposal bags are constructed with proposals corrected by the regressor. The mechanism behind SPSD is that the network learns the spatial information from the reliable samples, e.g. those low-noise annotations, in the dataset and then guides the noisy samples to produce high-quality proposals, as shown in Fig. 2. In addition, to further combine the category information and the spatial information, an interactive structure is implemented by alternately using SPSD to mine spatial cues and MIL to utilize the category information. With SPSD and the interactive structure, a high-quality proposal bag can be constructed. Experiments on MS-COCO show that SPSD can significantly improve the mean/max IoU between objects and proposals" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.891, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.203, + 0.894, + 0.265 + ], + "angle": 0, + "content": "Figure 2: The mechanism of Spatial Self-Distillation. By assigning higher weight, the low-noise annotations can be seen as reliable samples to guide the training of proposals' spatial position and identity learning in SPsD and SISD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.501 + ], + "angle": 0, + "content": "(about 18/10 points, Fig. 5) in the constructed bag. Instead of selecting proposals by classification confidence, we have proposed the SISD module in the proposal selection step. We use it to obtain each proposal's spatial confidence by predicting the IoU with the object and combining the IoU with classification confidence to select the top-\\(k\\) proposals. It is worth mentioning that SISD is an object-related IoU predictor, which means that the predicted IoU may be different for the same proposal that appears in different objects' bags. Accordingly, it guarantees that SISD can better handle object drift and group prediction problems. Experiments on MS-COCO and VOC datasets verify the effectiveness of our method and bring state-of-the-art performance. The contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.562 + ], + "angle": 0, + "content": "1) We further investigate the inaccurate-box supervised object detection tasks and propose an end-to-end training SSD-Det that combines the spatial and the category information in an interactive fashion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.609 + ], + "angle": 0, + "content": "2) We utilize an SPSD module to generate higher-quality proposals sampling through statistic-guide spatial position distillation, raising the upper bound of the refinement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.655 + ], + "angle": 0, + "content": "3) To add spatial cues to classification confidence, we also introduce an SISD module to select a proposal belonging to the object rather than the category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.73 + ], + "angle": 0, + "content": "4) The performance of our proposed SSD-Det improves the mean average precision (AP) of the best previous method (e.g. over 10 AP on \\(40\\%\\) noisy MS-COCO) and achieves state-of-the-art under various noise rate box supervision on MS-COCO and VOC datasets." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.746, + 0.642, + 0.763 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.667, + 0.788 + ], + "angle": 0, + "content": "2.1. Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Classic object detection [15, 38, 37, 32, 30, 3, 43, 57] is supervised by an accurate bounding-box. One-stage detectors utilize anchors as the sliding-window, such asYOLO [37], SSD [32], and RetinaNet [30]. Two-stage detectors mine spatial information to predict proposals (e.g. selective search [47] in Fast R-CNN [15] or RPN in Faster R-CNN [38]) and conduct classification and bounding-box" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6856" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "regression with filtered proposals sparsely. Transformer-based (i.e. DETR [3], Deformable-DETR [62], and SwinTransformer [33]) detectors utilize global information for better representation. Sparse R-CNN [43] combines a transformer's advantages and CNNs for detection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.184, + 0.468, + 0.201 + ], + "angle": 0, + "content": "2.2. Weakly-Supervised Object Detection (WSOD)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.209, + 0.47, + 0.512 + ], + "angle": 0, + "content": "WSOD trains object detectors with image tag supervision. Only with the category annotation, the majority of previous methods treat each image as a bag and candidate proposals as instances. They follow the multiple instance learning (MIL) pipeline [1, 44, 46, 7, 48], which highly depends on category information. However, the MIL loss function leads to a non-convex optimization problem; thus, MIL solutions are usually stuck into the local minima. Context information [24, 51], spatial regularization [1, 9, 48], and optimization strategy [44, 48, 46] are proposed to address the problems. SPE [28] introduces Transformer into WSOD and uses attention to generate proposals. SD-LocNet [59] tackles the initialized noisy object locations in WSOD and proposes a self-directed localization network to identify the noisy object instances. [44, 46, 7] use the pseudo label for classification's iterative refinement. However, we use the pseudo box as a better self-distillation teacher. [54, 39] conduct regression to move the proposals, whereas we conduct regression to distill for better bag construction. In this work, we also formulate box correction as a MIL problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.439, + 0.544 + ], + "angle": 0, + "content": "2.3. Semi-Supervised Object Detection (SSOD)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Semi-supervised learning in object detection can be roughly categorized into two groups: consistency based [22, 45] and pseudo label based [27, 36, 41, 49, 64, 52]. [52] presents an end-to-end SSOD approach with two simple techniques named soft teacher and box jittering to facilitate the efficient leverage of the teacher model. Both Soft Teacher [52] and SSD-Det obtain pseudo-labels from candidate boxes, adopting distillation, box jitter and classification scores weighting policy. However, they are different: 1) SSOD selects candidate boxes from the teacher model's detection results, while SSD-Det generates them with a generative approach, due to no any accurate supervised data for ensuring high-quality boxes in the detection results. 2) [52] is based on FixMatch [40] and requires maintaining two networks for teacher-student distillation structure. In contrast, our approach only needs a multi-head detector with a shared backbone for self-distillation. 3) Box Jitter: [52] calculates box variance from jittering for result selection, while we aim to generate candidate boxes that combine SISD-predicted IoU and classification scores, selecting boxes closer to the ground truth. 4) We use classification scores as weights for the next stage's loss, while [52] employs them as a criterion for selecting reliable samples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.791, + 0.108 + ], + "angle": 0, + "content": "2.4. Learning with Noisy Annotations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Training CNNs under noisy labels has been an active research area. Previous research focuses on the classification task, and develops various techniques to deal with noisy labels, such as sample selection [17, 23] for training, label correction [42, 34], and robust loss functions [60, 14] against noisy labels. Recently, many efforts [4, 25, 35, 53, 11, 12] have been devoted to the object detection task. On the one hand, Simon et al. [4] first investigates the impact of different types of label noise on object detection. They propose a per-object co-teaching method to alleviate the effect of noisy labels. On the other hand, [53] proposes a meta-learning framework for noisy annotations consisting of noisy category labels and bounding boxes. [11, 12] utilize object-level MIL to refine the inaccurate box. OAMIL [12] constructs proposal bags through label assignment in a discriminant style. P2BNet [11] originally conducts point-supervised object detection tasks. However, it can be seen as the box correction in its refinement stage. It uses hand-craft anchors to generate proposal bags. Our method inherits the generative style of P2BNet and conducts spatial distillation to mine spatial information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.439, + 0.712, + 0.455 + ], + "angle": 0, + "content": "2.5. Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.613 + ], + "angle": 0, + "content": "Knowledge distillation (KD) [21] aims to learn compact and efficient student models guided by excellent teacher networks. It is first applied to object detection in [5], in which hint learning and KD are both used for multi-class object detection. Recently, many efforts [26, 50, 8, 16, 55] aim to mimic the feature. [61] shows that localization knowledge is more important and proposed a localization distillation method. We also transfer the spatial knowledge from reliable labeled instances to correct inaccurate bounding boxes (shown in Fig. 2) in a self-distillation manner." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.624, + 0.634, + 0.641 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.83 + ], + "angle": 0, + "content": "This work aims to learn a robust detector with inaccurate bounding boxes. Instead of training a detector with the original inaccurate bounding box, we follow most related works [11, 12] that design a branch to refine the inaccurate bounding box and then train the detector head or detector with the refined bounding box. The most important part is how to design a refining policy. We first design a two-stage basic box refiner (gray region in Fig. 3) as a naive solution that modified from [11]. Then, SPSD and SISD are proposed and added to further mine the spatial cues for box refinement, yielding SSD-Det. Therefore, the overall loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.836, + 0.892, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {B a s i c}} + \\alpha_ {1} \\cdot \\mathcal {L} _ {\\text {S P S D}} + \\alpha_ {2} \\cdot \\mathcal {L} _ {\\text {S I S D}} + \\alpha_ {3} \\cdot \\mathcal {L} _ {\\text {D e t}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha_{1},\\alpha_{2}\\) and \\(\\alpha_{3}\\) are set as 0.25, 0.25 and 4 respectively. \\(\\mathcal{L}_{Det}\\) denotes the loss of detector or detection head. During inference, only the detector or the detection head is used." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6857" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.087, + 0.895, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.284, + 0.895, + 0.377 + ], + "angle": 0, + "content": "Figure 3: The framework of SSD-Det. It contains basic box refiner, SPSD module, SISD module and a detector head. Neighborhood sampler is adopted around the inaccurate annotation. Then, SPSD module generates better proposal bags which are fed into basic box refiner for MIL training. The selected proposals are average weighted as the refined box and supervise the next SPSD training. Meanwhile, the SISD module predicts the IoU between proposals and the object, and the estimated IoU is multiplied by classification score for better proposal selection to generate the refined box. SPSD shares backbone with the detector." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.402, + 0.251, + 0.417 + ], + "angle": 0, + "content": "3.1. Basic Box Refiner" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.471, + 0.549 + ], + "angle": 0, + "content": "Motivated by [12] and WSOD [1], we design the basic box refiner (detailed structure figure is in supplementary) that leverages classification confidence to refine the inaccurate box annotation. Then the refine annotation is used to train a detection head or detector. Following [11], we design a two-stream structure as a MIL classifier to select the best proposal for box refinement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Giving an image with inaccurate box annotation, for each object, \\(\\mathcal{B}\\) is a bag of proposals (bounding boxes) that are generated around its inaccurate annotation by a sampler policy (e.g., selective search[47], edge box[63], neighborhood sampler in Sec. 3.2). Meanwhile a feature map is extracted with a backbone network. And then through \\(7\\times 7\\) RoIAlign [18] and two fully connected (fc) layers, features of proposal in \\(\\mathcal{B}\\) are extracted and denote as \\(\\mathbf{F}\\). The basic box refiner takes proposal bag \\(\\mathcal{B} \\in \\mathbb{R}^{P \\times 4}\\) and features \\(\\mathbf{F} \\in \\mathbb{R}^{P \\times D}\\) as inputs, where \\(P\\), \\(D\\) are denoted as the number of proposals in \\(\\mathcal{B}\\), feature dimension respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Following [11] and [1], as Eq. 2 described, we apply the classification branch \\( f_{cls} \\) to \\( \\mathbf{F} \\) yields \\( \\mathbf{O}^{cls} \\), which is then passed through the softmax function over classification dimension \\( K \\) to obtain the score \\( \\mathbf{S}^{cls} \\in \\mathbb{R}^{P \\times K} \\), where \\( K \\) represents the number of instance categories. Likewise, instance selection branch \\( f_{cls} \\) is applied to \\( \\mathbf{F} \\) to yield \\( \\mathbf{O}^{ins} \\), and instance score \\( \\mathbf{S}^{ins} \\) is obtained through softmax function over \\( P \\) proposals. The proposal score \\( \\mathbf{S} \\) is obtained by computing the Hadamard product of the classification score and the instance score. The bag score \\( \\widehat{\\mathbf{S}} \\) is obtained by the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.403, + 0.854, + 0.418 + ], + "angle": 0, + "content": "summating of the \\(P\\) proposal boxes' proposal scores." + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.42, + 0.872, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} ^ {c l s} = f _ {c l s} (\\mathbf {F}) \\in \\mathbb {R} ^ {P \\times K}; [ \\mathbf {S} ^ {c l s} ] _ {p k} = e ^ {[ \\mathbf {O} ^ {c l s} ] _ {p k}} / \\sum_ {k = 1} ^ {K} e ^ {[ \\mathbf {O} ^ {c l s} ] _ {p k}}.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.457, + 0.885, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} ^ {i n s} = f _ {i n s} (\\mathbf {F}) \\in \\mathbb {R} ^ {P \\times K}; [ \\mathbf {S} ^ {i n s} ] _ {p k} = e ^ {[ \\mathbf {O} ^ {i n s} ] _ {p k}} / \\sum_ {p = 1} ^ {P} e ^ {[ \\mathbf {O} ^ {i n s} ] _ {p k}}.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.497, + 0.891, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\mathbf {S} ^ {c l s} \\odot \\mathbf {S} ^ {i n s} \\in \\mathbb {R} ^ {P \\times K}; \\widehat {\\mathbf {S}} = \\sum_ {p = 1} ^ {P} [ \\mathbf {S} ] _ {p} \\in \\mathbb {R} ^ {K}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.543, + 0.891, + 0.558 + ], + "angle": 0, + "content": "where \\(\\left[\\cdot \\right]_{pk}\\) is the value at row \\(p\\) and column \\(k\\) in the matrix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.892, + 0.604 + ], + "angle": 0, + "content": "The basic box refiner has two similar stages. The loss of stage I(termed \\(\\mathcal{L}_I\\)) adopt the MIL paradigm with the form of cross-entropy (CE) loss, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.606, + 0.892, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I} = C E (\\widehat {\\mathbf {S}}, \\mathbf {c}) = - \\sum_ {k = 1} ^ {K} \\mathbf {c} _ {k} \\log (\\widehat {\\mathbf {S}} _ {k}) + (1 - \\mathbf {c} _ {k}) \\log (1 - \\widehat {\\mathbf {S}} _ {k}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.891, + 0.697 + ], + "angle": 0, + "content": "where \\(\\mathbf{c} \\in \\{0,1\\}^K\\) is the one-hot category label. And each object's proposals with the top-\\(k\\) highest proposal score \\(\\mathbf{S}\\) are weighted to obtain the refined box." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.82 + ], + "angle": 0, + "content": "The stage II takes the refined box of stage I as input and performs fine refining with a similar structure as stage I. Differently, the focal loss is adopted in stage II instead of cross entropy loss. In order to cooperate with focal loss, the classification branch uses the sigmoid \\(\\sigma(x)\\) instead of softmax function and we sample some negative samples \\(\\mathcal{N}\\) to further suppress the background. With the bag score \\(\\widehat{S}\\) and the negative sample scores \\(\\mathbf{S}_{\\text{neg}}^{\\text{cls}}\\), the loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.825, + 0.892, + 0.856 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I I} = \\left\\langle \\mathbf {c} ^ {\\mathrm {T}}, \\widehat {\\mathbf {S}} ^ {*} \\right\\rangle \\cdot \\operatorname {F L} \\left(\\widehat {\\mathbf {S}}, \\mathbf {c}\\right) + \\sum_ {\\mathcal {N}} \\beta \\cdot \\operatorname {F L} \\left(\\mathbf {S} _ {n e g} ^ {c l s}, c _ {n e g}\\right) \\} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.862, + 0.892, + 0.904 + ], + "angle": 0, + "content": "where FL is the focal loss [30], \\(\\widehat{\\mathbf{S}}_j^*\\) represents the bag score predicted by stage I. \\(\\left\\langle \\mathbf{c}_j^{\\mathrm{T}},\\widehat{\\mathbf{S}}_j^*\\right\\rangle\\) represents the inner product" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6858" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.158 + ], + "angle": 0, + "content": "of the two vectors, meaning the predicted bag score of the ground-truth category. \\(\\beta\\) is the average of \\(\\left\\langle \\mathbf{c}_j^{\\mathrm{T}},\\widehat{\\mathbf{S}}_j^*\\right\\rangle\\). They are used to weight each object's FL for stable training. The overall loss function of the basic refiner here is:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.17, + 0.469, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {B a s i c}} = \\mathcal {L} _ {I} + \\alpha_ {I I} \\cdot \\mathcal {L} _ {I I} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.197, + 0.4, + 0.211 + ], + "angle": 0, + "content": "where \\(\\alpha_{II}\\) are the loss weights of the two stages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.469, + 0.287 + ], + "angle": 0, + "content": "During training, the refined box of stage II is used as supervision for a detection head or detector. After training, the basic box refiner will be removed, leaving a well-trained detection head or detector. In this way, we can train a detector under inaccurate annotations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.298, + 0.418, + 0.314 + ], + "angle": 0, + "content": "3.2. Spatial Position Self-Distillation (SPSD)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.469, + 0.412 + ], + "angle": 0, + "content": "Like most MIL paradigm methods, basic box refiner has two main components: bag construction and proposal selection. And the main idea is to use classification information to guide the refining. In this paper, we add spatial information to improve refining. Specifically, SPSD is proposed to use spatial information to enhance bag construction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.469, + 0.548 + ], + "angle": 0, + "content": "Bag construction aims to obtain proposals for each object, while proposal selection is to select the proposals from the object bag. Then, the refined box is averaged over the selected proposals. Therefore, the quality of the proposals in constructed bag determines the upbound of refining. The bag construction can be implemented in a variety of ways. In this paper, the basic box refiner adopts a naive neighborhood sampler for bag construction. basic box refiner adopts a naive neighborhood sampler for bag construction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.469, + 0.626 + ], + "angle": 0, + "content": "Neighborhood Sampler. Proposals around the inaccurate box are sampled to construct an object bag. For each inaccurate box \\( b^{*} = (b_{x}^{*}, b_{y}^{*}, b_{w}^{*}, b_{h}^{*}) \\), its scale and aspect ratio with \\( s \\) and \\( v \\) are adjusted and its positions \\( o_{x}, o_{y} \\) are jittered to obtain the diverse proposal \\( b = (b_{x}, b_{y}, b_{w}, b_{h}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.628, + 0.469, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} b _ {w} = v \\cdot s \\cdot b _ {w} ^ {*}, \\quad b _ {h} = 1 / v \\cdot s \\cdot b _ {h} ^ {*}, \\\\ b _ {x} = b _ {x} ^ {*} + b _ {w} \\cdot o _ {x}, \\quad b _ {y} = b _ {y} ^ {*} + b _ {h} \\cdot o _ {y}. \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.824 + ], + "angle": 0, + "content": "These proposals \\( b \\) are used to construct the positive proposal bag \\( \\mathcal{B} \\) to train the MIL classifier. Thanks to the hand-craft sampling way, the number of proposals in different objects' proposal bags is controllable and balanced. However, the hand-craft neighborhood sampler strategy is difficult to set hyper-parameters, and the sampling space is discrete. For example, when the jitter region is small, the optimization space of refining is limited, while when it is large, more background will be introduced. Hence, we propose the SPSD module to mine spatial information for higher-quality proposal bag construction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Statistically Guided Adaptive Sampling. Instead of simply using a neighborhood sampler, we adopt a statistically guided adaptive sampling by adding SPSD modules into the basic box refiner. Taking the constructed proposal bag \\(\\mathcal{B}\\) of hand-craft neighborhood sampler as input, the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.396 + ], + "angle": 0, + "content": "RoI features of proposals in \\(\\mathcal{B}\\) are extracted and fed into the two shared fc layers to obtain \\(\\mathbf{F}\\). Then a regression fc layer \\(f_{dis}\\), supervised by the inaccurate annotated bounding box \\(b^{*}\\), is introduced to predict the adaptive proposal bag \\(\\mathcal{B}^{dis} = f_{dis}(\\mathbf{F}) \\in \\mathbb{R}^{P \\times 4}\\), in which the proposals are closer to the object. Later, \\(\\mathcal{B}^{dis}\\) as the constructed proposal bag is fed into stage I of basic box refiner. In order to combine category and spatial information, we implement an interactive structure by alternately using SPSD to mine the spatial cues and using MIL in basic box refiner to utilize the category information. Specifically, the refined bounding box \\(\\hat{b}^{*}\\) of stage I that selected by the classification confidence is used to supervision of a new SPSD module for stage II. Similar as stage I, The new SPSD takes proposal bag \\(\\mathcal{B}^{dis}\\) of handcraft neighborhood sampler as input. Through the RoI align and the two shared fc layers, the feature \\(\\hat{\\mathbf{F}}\\) is extracted. An extra fc layer \\(f_{dis}\\) is then utilized to conduct further regression. Different with stage I, the obtained \\(\\mathcal{B}^{dis}\\) is supervised by the refined \\(\\hat{b}^{*}\\). The loss function of the spatial distillation for adaptive sampling can be defined as \\(\\mathcal{L}_{SPSD}\\) in Eq. 7." + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.402, + 0.892, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S P S D} = \\frac {1}{P} \\left\\{\\sum_ {p = 1} ^ {P} \\mathbf {L} _ {\\mathbf {1}} \\left(\\left[ \\mathcal {B} ^ {d i s} \\right] _ {p}, b ^ {*}\\right) + \\sum_ {p = 1} ^ {P} \\mathbf {L} _ {\\mathbf {1}} \\left(\\left[ \\mathcal {B} ^ {\\hat {d i s}} \\right] _ {p}, b ^ {*}\\right) \\right\\} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.445, + 0.882, + 0.459 + ], + "angle": 0, + "content": "where the \\(\\mathbf{L}_1\\) is the L1 loss function for loose restrictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.891, + 0.596 + ], + "angle": 0, + "content": "The idea behind SPSD is that the dataset with inaccurate annotation still has many reliable, high-quality boxes and inaccurate boxes. Supervised by the high-quality boxes statistically, the network can guide those proposals sampled around the inaccurate bounding box to regress to the ground truth. With the self-distillation mechanism, SPSD learns the semantic-spatial correspondence knowledge from the reliable samples in the dataset and then propagates the knowledge to produce high-quality proposals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.891, + 0.689 + ], + "angle": 0, + "content": "Adaptive Negative Sampling. Negative samples are introduced in Stage II to better suppress the background. With the sampled \\(\\mathcal{B}^{dis}\\), we can adaptively sample the negative samples with a small IoU (set smaller than 0.3 by default) with all positive proposals in all bags, to compose the negative sample set \\(\\mathcal{N}\\) for stage II." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.836, + 0.713 + ], + "angle": 0, + "content": "3.3. Spatial Identity Self-Distillation (SISD)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The basic box refiner selects the proposals only depending on classification confidence during proposal selection. To select the proposal which has high classification confidence and is also spatially close to the object from the bag, we propose a SISD module to predict the IoU between proposals and their corresponding object. Afterwards, through the combination between the IoU and the classification confidence, top-\\(k\\) proposals are selected. In SISD, we design an Object Relevance Enhancement (ORE) module to distinguish different objects' features with the same RoI region. And an identity predictor is designed to predict each proposal's IoU with the object." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6859" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.892, + 0.41 + ], + "angle": 0, + "content": "
MethodBackbone20% Box Noise Level40% Box Noise Level
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)
Val Set
Clean-FasterRCNN [38]ResNet-5037.958.140.921.641.648.737.958.140.921.641.648.7
Clean-FasterRCNN [38]ResNet-10139.460.143.122.443.751.139.460.143.122.443.751.1
Clean-Retinanet [30]ResNet-5036.756.139.021.640.447.436.756.139.021.640.447.4
Noisy-FasterRCNN [38]ResNet-5030.454.331.417.433.938.710.328.93.35.711.815.1
Noisy-Retinanet [30]ResNet-5030.053.130.817.933.738.213.333.65.78.415.918.0
FreeAnchor[58]ResNet-5028.653.128.516.632.237.010.428.93.35.812.114.9
Co-teaching[17]ResNet-5030.554.930.517.334.039.111.531.44.26.413.116.4
SD-LocNet[59]ResNet-5030.054.530.317.533.638.711.330.34.36.012.716.6
KL loss[20]ResNet-5031.054.332.418.034.939.512.136.73.76.213.017.4
OA-MIL[12]ResNet-5032.155.333.218.135.841.618.642.612.99.219.926.5
SSD-DetResNet-5033.657.335.319.537.243.327.653.926.016.031.034.9
SSD-DetResNet-10134.357.636.719.138.144.328.454.327.216.531.936.4
SSD-Det+FRResNet-5034.457.336.820.038.244.029.354.829.017.132.936.9
SSD-Det+FRResNet-10136.259.139.220.940.247.130.656.730.718.134.539.0
Test Set
Clean-FasterRCNN [38]ResNet-5037.758.740.821.740.646.737.758.740.821.740.646.7
Noisy-FasterRCNN [38]ResNet-5030.754.931.318.033.737.710.429.03.36.011.314.6
OA-MIL[12]ResNet-5032.355.833.718.535.040.218.542.312.89.319.125.1
SSD-DetResNet-5033.557.335.519.136.041.928.054.126.516.530.034.5
SSD-Det+FRResNet-5034.757.937.220.037.742.729.755.629.317.532.436.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.419, + 0.893, + 0.451 + ], + "angle": 0, + "content": "Table 1: Performance comparison on COCO. FR is Faster R-CNN. *-FR refers to a retrained Faster R-CNN (R50+FPN) using refined annotations from SSD-Det for improved performance.. Clean-* and Noisy-* means original and noisy annotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.717 + ], + "angle": 0, + "content": "Object Relevance Enhancement (ORE). ORE enhances object-relevant features, making SISD an object-relevant IoU predictor. ORE allows the predicted IoU to be different for the same proposal in other objects' bags. In addition, we integrate the feature of the bag's corresponding object into the proposal feature, making the feature of different bags' proposals distinct. That is the so-called ORE. For a proposal bag \\(\\mathcal{B}\\), the feature \\(\\mathbf{F}\\) is obtained through the RoI align and two fc layers. It is worth mentioning that the two fcs do not share the parameters with those in the refiner since the optimization goals are contradictory. To represent the feature of the bag's corresponding object, \\(\\mathbf{F}^{+} \\in \\mathbb{R}^{1 \\times D}\\) is calculated by averaging features of \\(P\\) proposals in proposal bag \\(\\mathcal{B}\\). The object feature \\(\\mathbf{F}^{+}\\)is broadcast into \\(\\mathbb{R}^{P \\times D}\\), and then added to the proposal features to obtain the object-relevant features \\(\\mathbf{F}^{*} = \\mathbf{F} + \\mathbf{F}^{+}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.718, + 0.47, + 0.812 + ], + "angle": 0, + "content": "Spatial Identity Prediction. By a following identity fc layer, \\(U \\in \\mathbb{R}^{P \\times 1}\\) is predicted. The pseudo label \\(T \\in (0,1)\\) is IoU between proposals in \\(\\mathcal{B}\\) and the merged box \\(\\hat{b}^*\\) of stage I. For better optimization, the linear normalized \\(T' = (T - 0.5)/0.5 \\in (-1,1)\\) is utilized as supervision. The object function of the identity predictor is identified in:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.817, + 0.47, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S I S D} = \\operatorname {s m o o t h} _ {L 1} \\left(U, T ^ {\\prime}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where the smooth \\(L_{1}\\) represents the smooth L1 loss. The predicted spatial confidence \\(U^{\\prime}\\) is obtained by normalizing the \\(U\\). Finally, \\(S^{*} = U^{\\prime} \\cdot S\\) is used to select the top-\\(k\\) proposals for merging as the refined boxes." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.474, + 0.89, + 0.627 + ], + "angle": 0, + "content": "
MethodBackboneBox Noise Level
10%20%30%40%
Clean-FasterRCNN [38]ResNet-5077.2 for clean
Clean-RetinaNet [30]ResNet-5073.5 for clean
Noisy-FasterRCNN [38]ResNet-5076.371.260.142.5
Noisy-RetinaNet [30]ResNet-5071.567.657.945.0
KL loss[20]ResNet-5075.872.764.648.6
Co-teaching[17]ResNet-5075.470.660.943.7
SD-LocNet[59]ResNet-5075.771.560.843.9
FreeAnchor[58]ResNet-5073.067.556.241.6
OA-MIL[12]ResNet-5077.474.370.663.8
SSD-DetResNet-5077.174.871.566.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Table 2: Performance comparison on the VOC 2007 test set. The evaluation metric is \\( AP_{50} \\). The Clean-* and Noisy-* means original annotation and noisy annotation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.625, + 0.729 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.738, + 0.707, + 0.755 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.883 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. For experimental comparisons, two publicly available datasets are used for object detection with inaccurate bounding boxes: MS-COCO [31] and PASCAL VOC 2007 [13]. MS-COCO (2017 version) has 118k training and 5k validation images with 80 common object categories. PASCAL VOC 2007 is one of the most popular benchmarks in generic object detection with 20 classes." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metric. We use mean average precision" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6860" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.208 + ], + "angle": 0, + "content": "
2-RefSPSDSISDRe-Train20% Box Noise Level40% Box Noise Level
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)\\(AP^{test}\\)\\(AP^{test}_{75}\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)\\(AP^{test}\\)\\(AP^{test}_{75}\\)
30.057.129.016.933.139.8--22.851.116.113.325.030.4--
✓31.256.731.617.834.541.031.432.024.652.020.114.328.231.925.020.5
✓✓33.056.934.818.735.542.233.134.827.253.724.715.930.335.227.625.6
✓✓✓33.657.335.319.537.243.333.535.527.653.926.016.031.034.928.026.5
✓✓31.856.833.118.435.740.832.333.726.554.023.315.730.333.826.823.3
✓✓✓34.157.636.419.037.743.834.336.629.055.127.817.032.536.729.328.4
✓✓✓✓34.457.336.820.038.244.034.737.229.354.829.017.132.936.929.729.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.217, + 0.893, + 0.248 + ], + "angle": 0, + "content": "Table 3: Modules ablation of SPSD, SISD and Re-Train on MS-COCO validation set (without) and test set (with test). The Re-Train means we generate the pseudo label by SSD-Det and re-train a Faster R-CNN detector." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.272, + 0.466, + 0.335 + ], + "angle": 0, + "content": "
Methods (w/o SISD)APAP50AP75APsAPmAPl
Neighborhood Sampler24.652.020.114.328.231.9
SPSD (II) w/o weighted26.053.322.515.629.433.4
SPSD (II) w/ weighted26.353.422.515.629.333.8
SPSD (I+II) w/ weighted27.253.724.715.930.335.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.157, + 0.346, + 0.39, + 0.361 + ], + "angle": 0, + "content": "Table 4: Different setting of SPSD." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.379, + 0.462, + 0.466 + ], + "angle": 0, + "content": "
ORE Strategies of SISDAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP^s \\)\\( AP^m \\)\\( AP^l \\)
w/o SISD27.253.724.715.930.335.2
SISD w/o ORE27.353.325.716.930.235.0
+ subtract27.253.824.615.730.435.0
+ concatenate27.254.024.516.130.234.8
+ add27.653.926.016.031.034.9
+ add w/ shared fcs23.049.917.612.625.330.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.477, + 0.414, + 0.492 + ], + "angle": 0, + "content": "Table 5: Different ORE strategies of SISD." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.51, + 0.468, + 0.543 + ], + "angle": 0, + "content": "
Num.0123
AP/AP5024.6 / 52.026.3 / 53.427.2 / 53.727.0 / 53.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.566, + 0.388, + 0.58 + ], + "angle": 0, + "content": "Table 6: Number of SPSD module." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.611, + 0.468, + 0.656 + ], + "angle": 0, + "content": "mAP@[.5,.95] and (mAP@.5) for MS-COCO and VOC. The \\(\\{AP, AP_{50}, AP_{75}, AP^{Small}, AP^{Middle}, AP^{Large}\\}\\) is reported for MS-COCO and \\(AP_{50}\\) for VOC." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Synthetic Noisy Dataset. Following [12], We simulate noisy bounding boxes by perturbing clean boxes from the original annotations. The details are in the appendix. We simulate various box noise levels ranging from \\(10\\%\\) to \\(40\\%\\) for the VOC and \\(\\{20\\%, 40\\}\\) for the MS-COCO." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Implementation Details. We implement our method on FasterRCNN [38] with ResNet50-FPN [19, 29] backbone, based on MMDetection [6]. All settings of our method and previous methods employ FPN for fair comparison. Similar to the default setting of object detection on MS-COCO, the stochastic gradient descent [2] algorithm is used to optimize on 1x training schedule. The batch size is two images per GPU on 8 GPUS. For the VOC dataset, the batch size is two images per GPU on 2 GPUS. The performance we report is on a single scale (1333 * 800 for MS-COCO and 1000 * 600 for VOC)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.274, + 0.796, + 0.29 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.307, + 0.892, + 0.399 + ], + "angle": 0, + "content": "We compare our method with several state-of-the-art approaches [20, 17, 59, 58, 12] on MS-COCO and VOC 2007 datasets. We denote Clean-FasterRCNN and Noisy-FasterRCNN as FasterRCNN models trained under clean (original annotations) and noisy annotations with the default setting, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.893, + 0.786 + ], + "angle": 0, + "content": "MS-COCO Dataset. Table 1 shows the comparison results on the MS-COCO. Inaccurate bounding box annotations significantly deteriorate the vanilla Faster R-CNN's detection performance. Co-teaching and SD-LocNet only slightly improve the detection performance, especially under \\(40\\%\\) box noise. That indicates that small-loss sample selection and sample weight assignment can not tackle noisy box annotations well. KL Loss slightly improves the performance under \\(20\\%\\) and \\(40\\%\\) box noise. By treating an object as a bag of instances, OA-MIL is somehow robust to noisy bounding boxes and performs better than other methods. Nevertheless, the previously-mentioned label assignment bag construction limits its ability to handle heavy noise. Our approach is more robust to noisy bounding boxes. It outperforms other methods by a large margin under high box noise levels and significantly boosts the detection performance across all metrics. For example, under \\(40\\%\\) box noise, the end-to-end SSD-Det achieves 27.6 \\(AP\\) and \\(53.9 AP_{50}\\), 9.0 and attains 11.3 point improvement compared with state-of-the-art method OA-MIL, respectively. Also, through re-training on FasterRCNN, the performance further reaches 29.3 \\(AP\\) and \\(54.8 AP_{50}\\). With the backbone of ResNet-101, the performance achieves consistent improvement. On MS-COCO test set, our method also achieves state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "VOC 2007 Dataset. Table 2 shows the comparison results on the VOC 2007 test set. Co-teaching, SD-LocNet and KL Loss, can not address inaccurate bounding box annotations well. OA-MIL improves the performance on different noisy datasets. Our approach obtains further improvements to 77.10, 74.80, 71.50, 66.90 \\(AP_{50}\\) on \\(10\\%\\), \\(20\\%\\), \\(30\\%\\) and \\(40\\%\\) noisy box datasets, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6861" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.089, + 0.369, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.09, + 0.491, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.091, + 0.616, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.091, + 0.892, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.296, + 0.89, + 0.327 + ], + "angle": 0, + "content": "Figure 4: Qualitative detection results on COCO validation. Previous methods miss objects and face part prediction problems. Our method misses fewer objects, and the bounding box quality is better, especially for small or overlapped objects." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.35, + 0.466, + 0.378 + ], + "angle": 0, + "content": "
MethodsBox Refiner+Re-TrainSSD-DetSSD-Det+Re-Train
AP/AP5029.0 / 54.427.6/53.929.3 / 54.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.388, + 0.441, + 0.403 + ], + "angle": 0, + "content": "Table 7: Comparisons of end-to-end and Re-Train." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.433, + 0.333, + 0.449 + ], + "angle": 0, + "content": "4.3. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.518 + ], + "angle": 0, + "content": "To further analyze SSD-Det's effectiveness and robustness, we conduct more experiments on COCO val set if there are no other instructions. Except for Table 3, the noise level of these experiments is \\(40\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.469, + 0.671 + ], + "angle": 0, + "content": "Ablation of Modules. Ablation study of each component in our approach is given in Table 3, including: (i) Different stages of our basic box refiner. i.e. training object detector without the stage II (2-Ref), where the pseudo boxes predicted by the stage I are served as the supervision for training a parallel detector. (ii) SPSD, i.e. training without SPSD, where the object-bag is constructed directly by neighborhood sampling around the noisy ground-truth or the predicted pseudo boxes of the stage I. (iii) SISD. (iv) Re-Train with FasterRCNN (Re-Train)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Effectiveness of SPSD. SPSD further improves the detection performance on the MS-COCO, especially under high box noise levels, e.g. under \\(40\\%\\) box noise level, SPSD boosts the performance from 24.6 to 27.2, as shown in Table 3 (row 3). In Table 4, we conduct further ablation on SPSD. With SPSD bag construction only in stage II, the performance increases by 1.4 AP. The performance further improves with the proposal score of stage I as weights. With SPSD in all stages, the AP reaches 27.2. Fig. 5 shows the bag quality. With SPSD, the mean IoU increases from 40.3 to 58.7 and the max and top-10 IoU increase to 78.3 and 75.1, which indicates a better upper bound of proposal selection. More high-quality proposals bring better optimization and easier proposals selection." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Number of SPSD. As shown in Table 6. When adding" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.351, + 0.892, + 0.426 + ], + "angle": 0, + "content": "
DetectorsClean-supervisedNoise-supervised
APAP50APAP50(w/ours)
APAP50
FasterRCNN37.958.110.328.929.354.8
SparseRCNN [43]†45.064.16.020.334.360.2
De-DETR [62]†46.866.35.016.935.260.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.448, + 0.893, + 0.494 + ], + "angle": 0, + "content": "Table 8: Experiments on advanced detectors. De-DETR is Deformable DETR. † uses multi-scale data augment. ‘\\(w/\\)ours’ means using our method under noisy supervision." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.516, + 0.892, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.669, + 0.892, + 0.745 + ], + "angle": 0, + "content": "Figure 5: Bag quality (IoU of proposals with GT) of construction in SSD-Det. B.S. (blue) means neighborhood sampler. SPSD I (orange) denotes single SPSD adopted. SPSD II (yellow) is two SPSD and interactive structure adopted. SPSD II significantly improves the quality." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.777, + 0.891, + 0.824 + ], + "angle": 0, + "content": "3 SPSD, performance drops slightly, probably due to the accumulation of errors outweighing the performance gain from extra stages. Hence, 2 SPSD is our default setting." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of SISD. SISD is designed to select object-aware proposals in box selection. Under \\(40\\%\\) and \\(20\\%\\) box noise, the detection performance improves from 27.2 to 27.6 and 33.0 to 33.6, which verifies the effectiveness of the module, shown in Table 3. We also study the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6862" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.142 + ], + "angle": 0, + "content": "
MethodsDrift Rate (%)Group Rate (%)Part Rate (%)
allsmlallsmlallsml
OA-MIL[12]15.117.817.17.46.72.83.41.42.83.42.72.3
Ours1.51.01.31.41.71.20.50.71.00.51.11.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.165, + 0.468, + 0.21 + ], + "angle": 0, + "content": "Table 9: Breakdown of different problems during refinement (COCO under \\(40\\%\\) noise level). s, m and 1 mean small, middle and large scale." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.228, + 0.467, + 0.332 + ], + "angle": 0, + "content": "
DatasetourQuality (Average IoU)Frequency (%)
AllPartOversizeShiftReliablePartOversizeShift
Detector as Annotator
Objects-F✓44.324.133.613.340.128.49.821.7
47.030.841.816.849.223.18.019.4
COCO-F✓45.125.433.315.740.027.512.020.5
48.232.140.619.649.522.210.018.3
Point-based Annotator
COCO-P✓55.630.425.029.865.69.523.21.7
65.246.736.440.974.95.918.50.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.342, + 0.462, + 0.357 + ], + "angle": 0, + "content": "Table 10: Analysis of noisy annotations types and quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.468, + 0.509 + ], + "angle": 0, + "content": "strategies of ORE in SISD (Table 5). The minus or concatenation of object feature \\(\\mathbf{F}_j^+\\) and proposal feature \\(\\mathbf{F}_j\\) do not work. With add strategy, the performance is 27.60. If SISD shares the two fc layers, the performance drops to 22.99 since the optimization goals are contradictory (Identity distinguishes objects in the same category). If we directly use the RoI feature without ORE, the performance drops to 27.32 AP, verifying the effectiveness of the object relevance strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.468, + 0.616 + ], + "angle": 0, + "content": "Affect of Re-Train. As most WSOD methods do, we re-run the experiments by training a fully supervised detector for better performance. We find that if the SSD-Det only trains the refiner and uses the pseudo label to train the FasterRCNN, the result is good but lower than re-train after the end-to-end training given in Table 7 (row 1). This is because joint training is beneficial for box refinement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.468, + 0.679 + ], + "angle": 0, + "content": "Experiments on Advanced Detectors. We re-train recent detectors, e.g. SparseRCNN and Deformable DETR, under the boxes refined by our method. Table 8 verifies that our method achieves consistency improvement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.468, + 0.755 + ], + "angle": 0, + "content": "The computational cost discussion. Similar to OAMIL, our method adds SSD head to Faster R-CNN for auxiliary training to refine the noisy annotation and the head is not used during inference, the calculation cost during inference is same to standard Faster R-CNN." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.336, + 0.785 + ], + "angle": 0, + "content": "4.4. Visualization and Discussion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Fig. 4 shows that OA-MIL faces missing instances and grouping instances issues for small or overlapped objects (as mentioned in [12]), while our method still works well. For a better intuitive understanding of SISD and SPSD, we visualize the bag construction quality in Fig. 5 Then, we makes noise types breakdown of 'Drift', 'Group' and 'part dominance' issues. We give the definition of IoU, IoG and" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.092, + 0.54, + 0.105 + ], + "angle": 0, + "content": "IoD:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.116, + 0.892, + 0.142 + ], + "angle": 0, + "content": "\\[\nI o U = \\frac {A (I)}{A (D) + A (G) - A (I)}, I o G = \\frac {A (I)}{A (G)}, I o D = \\frac {A (I)}{A (D)} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.32 + ], + "angle": 0, + "content": "where \\( \\mathrm{A}(\\ast) \\) is area of box *, D and G are refined box and gt box respectively, and I is insertion between D and G. We statistically count the proportion of three noise types of 'bad' refined boxes (having small IoU with gt) in Table 9: (i) Drift: 'bad' refined box has a higher IoU with another nearby object. (ii) Group: 'bad' refined box has high IoG with multiple objects. (iii) Part: 'bad' refined box has a high IoD. Table 9 shows quantitative results for each noise type of baseline and ours. The drift, group, part problems reduce from \\( 15.1\\% \\), \\( 6.7\\% \\), \\( 2.8\\% \\) to \\( 1.5\\% \\), \\( 1.7\\% \\), \\( 1.0\\% \\), respectively, demonstrating our improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.607 + ], + "angle": 0, + "content": "Experiments on Real-life Noisy Annotations. Real-life noisy annotations stem from: low-quality data (e.g., occlusion, blur), human annotator errors and automatic machine annotator limitations. Noise from human errors is quite subjective, since differences between annotators. For a more objective analysis, noisy annotations from machine annotator is used for experiments. Without loss of generality, Faster R-CNN well-trained on MS-COCO was applied to Objects365 images, yielding Objects-F dataset, and to COCO-val images, producing COCO-F dataset. P2BNet [11], a point-based annotator, was used on COCO-val images with point annotations, generating COCO-P dataset. SSD-Det effectively improves low-quality boxes. As shown in Table 10, with SSD-Det's refinement, the average IoU increases for Objects-F (from 44.3 to 47.0), COCO-F (from 45.1 to 48.2) and COCO-P (from 55.6 to 65.2) datasets. Further, the proportion of reliable annotations increases, and noise categories' frequency (Part, Oversize, and Shift) decreases for all datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.624, + 0.618, + 0.639 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.83 + ], + "angle": 0, + "content": "This paper investigates problems during refinement caused by solely using category information to select proposals. We also propose SSD-Det to mine spatial information in a self-distillation fashion. SSD-Det introduces the SPSD module to learn semantic-spatial correspondence knowledge with neighborhood sampler and an interactive structure to combine spatial information and category information, thus producing a high-quality proposal bag. SISD in SSD-Det is utilized to improve the proposal selection procedure by integrating object-relevant spatial confidence. Complete ablations on multiple datasets verify the effectiveness of SSD-Det." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.847, + 0.597, + 0.862 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[1] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In CVPR, 2016. 1, 3, 4" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "6863" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[2] Léon Bottou. Stochastic gradient descent tricks. In Neural Networks: Tricks of the Trade - Second Edition. Springer, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.134, + 0.471, + 0.174 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, and Gabriel Synnaeve et al. End-to-end object detection with transformers. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.176, + 0.471, + 0.204 + ], + "angle": 0, + "content": "[4] Simon Chadwick and Paul Newman. Training object detectors with noisy data. In IV, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[5] Guobin Chen, Wongun Choi, and Xiang Yu et al. Learning efficient object detection models with knowledge distillation. In NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[6] Kai Chen, Jiaqi Wang, and Jiangmiao Pang et al. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[7] Ze Chen, Zhihang Fu, and Rongxin Jiang et al. SLV: spatial likelihood voting for weakly supervised object detection. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[8] Xing Dai, Zeren Jiang, and Zhao Wu et al. General instance distillation for object detection. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[9] Ali Diba, Vivek Sharma, and Ali Mohammad Pazandeh et al. Weakly supervised cascaded convolutional networks. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[10] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the multiple instance problem with axis-parallel rectangles. Artificial Intelligence, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.47, + 0.48 + ], + "angle": 0, + "content": "[11] Chen, et al. Point-to-box network for accurate object detection via single point supervision. In ECCV, 2022. 1, 2, 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[12] Liu, et al. Robust object detection with inaccurate bounding boxes. In ECCV, 2022. 1, 2, 3, 4, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.551 + ], + "angle": 0, + "content": "[13] Mark Everingham, Luc Van Gool, and Christopher K. I. Williams et al. The pascal visual object classes (VOC) challenge. *IJCV*, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.47, + 0.593 + ], + "angle": 0, + "content": "[14] Aritra Ghosh, Himanshu Kumar, and P. S. Sastry. Robust loss functions under label noise for deep neural networks. In AAAI, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.411, + 0.607 + ], + "angle": 0, + "content": "[15] Ross B. Girshick. Fast R-CNN. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[16] Jianyuan Guo, Kai Han, and Yunhe Wang et al. Distilling object detectors via decoupled features. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[17] Bo Han, Quanming Yao, and Xingrui Yu et al. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, 2018. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[18] Kaiming He, Georgia Gkioxari, and Piotr Dolkar et al. Mask R-CNN. In ICCV, 2017. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, and Shaoqing Ren et al. Deep residual learning for image recognition. In CVPR, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[20] Yihui He, Chenchen Zhu, and Jianren Wang et al. Bounding box regression with uncertainty for accurate object detection. In CVPR, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[21] Geoffrey Hinton and Oriol et al. Vinyls. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[22] Jisoo Jeong, Seungeui Lee, and Jeesoo Kim et al. Consistency-based semi-supervised learning for object detection. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[23] Lu Jiang, Zhengyuan Zhou, and Thomas Leung et al. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[24] Vadim Kantorov, Maxime Aquab, and Minsu Cho et al. Contextlocnet: Context-aware deep network models for weakly supervised localization. In ECCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[25] Junnan Li, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Towards noise-resistant object detection with noisy annotations. CoRR, abs/2003.01285, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[26] Quanquan Li, Shengying Jin, and Junjie Yan. Mimicking very efficient network for object detection. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[27] Yandong Li, Di Huang, and Danfeng Qin et al. Improving object detection with selective self-supervised self-training. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[28] Mingxiang Liao, Fang Wan, and Yuan Yao et al. End-to-end weakly supervised object detection with sparse proposal evolution. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Piotr Dólar, and Ross B. Girshick et al. Feature pyramid networks for object detection. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Priya Goyal, and Ross B. Girshick et al. Focal loss for dense object detection. In ICCV, 2017. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin and Michael et al. Maire. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[32] Wei Liu, Dragomir Anguelov, and Dumitru Erhan et al. SSD: single shot multibox detector. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, and Yue Cao et al. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[34] Xingjun Ma, Yisen Wang, and Michael E. Houle et al. Dimensionality-driven learning with noisy labels. In ICML, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[35] Jiafeng Mao, Qing Yu, and Kiyoharu Aizawa. Noisy localization annotation refinement for object detection. TIS, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[36] Ilija Radosavovic, Piotr Dólar, and Ross B. Girshick et al. Data distillation: Towards omni-supervised learning. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[37] Joseph Redmon, Santosh Kumar Divvala, and Ross B. Girshick et al. You only look once: Unified, real-time object detection. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[38] Shaoqing Ren, Kaiming He, and Ross B. Girshick et al. Faster R-CNN: towards real-time object detection with region proposal networks. TPAMI, 2017. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[39] Zhongzheng Ren, Zhiding Yu, and Xiaodong Yang et al. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[40] Kihyuk Sohn, David Berthelot, and Nicholas Carlini et al. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[41] Kihyuk Sohn, Zizhao Zhang, and Chun-Liang Li et al. A simple semi-supervised learning framework for object detection. CoRR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[42] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. SELFIE: refurbishing unclean samples for robust deep learning. In ICML, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[43] Peize Sun, Rufeng Zhang, and Yi Jiang et al. Sparse R-CNN: end-to-end object detection with learnable proposals. In CVPR, 2021. 2, 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6864" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[44] Peng Tang and Xinggang Wang et al. Multiple instance detection network with online instance classifier refinement. In CVPR, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[45] Peng Tang, Chetan Ramaiah, and Yan Wang et al. Proposal learning for semi-supervised object detection. In WACV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[46] Peng Tang, Xinggang Wang, and Song Bai et al. PCL: proposal cluster learning for weakly supervised object detection. TPAMI, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.469, + 0.264 + ], + "angle": 0, + "content": "[47] Koen E. A. van de Sande, Jasper R. R. Uijlings, and Theo Gevers et al. Segmentation as selective search for object recognition. In ICCV, 2011. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.267, + 0.469, + 0.306 + ], + "angle": 0, + "content": "[48] Fang Wan, Pengxu Wei, and Zhenjun Han et al. Min-entropy latent model for weakly supervised object detection. TPAMI, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.469, + 0.351 + ], + "angle": 0, + "content": "[49] Keze Wang, Xiaopeng Yan, and Dongyu Zhang et al. Towards human-machine cooperation: Self-supervised sample mining for object detection. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.469, + 0.393 + ], + "angle": 0, + "content": "[50] Tao Wang, Li Yuan, and Xiaopeng Zhang et al. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.396, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[51] Yunchao Wei, Zhiqiang Shen, and Bowen Cheng et al. TS ^2 2 C: tight box mining with surrounding segmentation context for weakly supervised object detection. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.44, + 0.469, + 0.48 + ], + "angle": 0, + "content": "[52] Mengde Xu, Zheng Zhang, and Han Hu et al. End-to-end semi-supervised object detection with soft teacher. In ICCV. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[53] Youjiang Xu, Linchao Zhu, and Yi Yang et al. Training robust object detectors from noisy category labels and imprecise bounding boxes. TIP, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.567 + ], + "angle": 0, + "content": "[54] Ke Yang, Dongsheng Li, and Yong Dou. Towards precise end-to-end weakly supervised object detection network. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.57, + 0.469, + 0.609 + ], + "angle": 0, + "content": "[55] Zhendong Yang, Zhe Li, and Xiaohu Jiang et al. Focal and global knowledge distillation for detectors. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.652 + ], + "angle": 0, + "content": "[56] Xuehui Yu, Pengfei Chen, and Di Wu et al. Object localization under single coarse point supervision. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[57] Xuehui Yu, Yuqi Gong, and Nan Jiang et al. Scale match for tiny person detection. In WACV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[58] Xiaosong Zhang, Fang Wan, and Chang Liu et al. Freeanchor: Learning to match anchors for visual object detection. In NeurIPS, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[59] Xiaopeng Zhang, Yang Yang, and Jiashi Feng. Learning to localize objects with noisy labeled instances. In AAAI, 2019. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[60] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[61] Zhaohui Zheng, Rongguang Ye, and Ping Wang et al. Localization distillation for dense object detection. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[62] Xizhou Zhu, Weijie Su, and Lewei Lu et al. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[63] C. Lawrence Zitnick and Piotr Dálár. Edge boxes: Locating object proposals from edges. In ECCV, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.15 + ], + "angle": 0, + "content": "[64] Barret Zoph, Golnaz Ghiasi, and Tsung-Yi Lin et al. Rethinking pre-training and self-training. In NeurIPS, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6865" + } + ] +] \ No newline at end of file diff --git a/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/d223c2f5-fe18-4502-b91f-9060d747649d_origin.pdf b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/d223c2f5-fe18-4502-b91f-9060d747649d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8ca9ec7778423ad04a697a61fc7fb83e58c5886c --- /dev/null +++ b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/d223c2f5-fe18-4502-b91f-9060d747649d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b18519c535b27eee0e42bfd59e871abb44783171c36b3c21fd5959ce5dd4fc0 +size 4039840 diff --git a/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/full.md b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fbce9dad7aa589bb31bf52597a719873d7aab380 --- /dev/null +++ b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/full.md @@ -0,0 +1,354 @@ +# Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes + +Di Wu*, Pengfei Chen*, Xuehui Yu*, Guorong Li, Zhenjun Han†, Jianbin Jiao + +University of Chinese Academy of Sciences + +# Abstract + +Object detection via inaccurate bounding boxes supervision has boosted a broad interest due to the expensive high-quality annotation data or the occasional inevitability of low annotation quality (e.g. tiny objects). The previous works usually utilize multiple instance learning (MIL), which highly depends on category information, to select and refine a low-quality box. Those methods suffer from object drift, group prediction and part domination problems without exploring spatial information. In this paper, we heuristically propose a Spatial Self-Distillation based Object Detector (SSD-Det) to mine spatial information to refine the inaccurate box in a self-distillation fashion. SSD-Det utilizes a Spatial Position Self-Distillation (SPSD) module to exploit spatial information and an interactive structure to combine spatial information and category information, thus constructing a high-quality proposal bag. To further improve the selection procedure, a Spatial Identity Self-Distillation (SISD) module is introduced in SSD-Det to obtain spatial confidence to help select the best proposals. Experiments on MS-COCO and VOC datasets with noisy box annotation verify our method's effectiveness and achieve state-of-the-art performance. The code is available at https://github.com/ucas-vg/PointTinyBenchmark/tree/SSD-Det. + +# 1. Introduction + +Object detection [18, 38, 58, 30, 57] relying on large-scale datasets like MS-COCO[31] has significantly progressed and achieved good performance. However, accurate bounding box annotations are expensive and challenging in natural contexts [12]. Especially in many professional scenarios, it is difficult to label accurate annotations without domain knowledge (e.g., agricultural crop observation and medical image processing) [12, 1]. As shown in Fig. 1a in some complex datasets, the human annotators may also annotate inaccurate bounding boxes due to the inherent ambiguities[20] of objects. In addition, labelling with a de + +![](images/596443fcd81f05aca705639a72bf222290bab867abf7f86375b54859c6b93eb8.jpg) + +![](images/cb1db177af5512fe3557cb3d1d236228818d1612d98f4c281ec3df3f45101fec.jpg) +(a) Labelling strategies lead to inaccurate annotations (red box). +(b) Three problems during previous MIL refining. Because their selections depend solely on classification, (i,ii): the refined box (green box) drifts to another object or makes a group prediction (merging across multiple objects) due to neighbor disturbance. Yellow boxes are proposals in the middle person's MIL bag. (iii): Local part may be more discriminative than the entire object and will be predicted. +Figure 1: The sources of inaccurate box annotations and three problems caused by previous refinement methods. + +tector or weak signal[11] (e.g., point) is much cheaper but brings more inaccuracy. Therefore learning robust detectors with inaccurate bounding boxes[12, 4, 35, 53, 25] is a practical and meaningful task and has boosted a broad interest. + +To use the inaccurate annotations, most related methods [12, 11] refine the inaccurate annotations as Fig. 1a shows, and then train a detector head or re-train a detector with the refined box as the new supervision. There are two main steps during refining: 1) Bag Construction: For each object, obtaining some proposals around the inaccurate an- + +notated bounding box to form the object-level proposal bag; 2) Proposal Selection: Selecting the top- $k$ proposals with the highest classification confidence from each bag and then weighting average them to obtain the refined box. + +During the proposal selection, they usually utilize multiple instance learning (MIL) [10] supervised by category information to choose the proposals with high classification confidence from the constructed bags. However, they pay less [12] or no [11] attention to mining spatial information, leading to the following problems as shown in Fig. 1b: (1) Object Drift: For each object, some proposals in the constructed bag do not have a high IoU with the original object but with another nearby object. These proposals are not spatially adjacent to the original object but still have high classification confidence, as the rightmost (the yellow box) proposal of $O_{2}$ shows in the left-bottom corner of Fig. 1b. Only the category confidence is relied on for selecting proposals for $O_{2}$ , and the rightmost proposal will be selected as the refined box. It means the refined box drifts to another object (Fig. 1b (i)), reducing the recall; (2) Group Prediction: Most works [7, 11, 56] select the top- $k$ proposals by classification confidence and then weight average them as the refined box, causing the group prediction problem, as shown in Fig. 1b (ii); (3) Part Domination: The detector often focuses on the object's semantic region, which can statistically represent the category (e.g. the face). As shown in Fig. 1b (iii), the high classification confidence of the animal is in the discriminant part (the face) rather than the entire object as mentioned by [44, 39]. + +To address these problems, we propose a Spatial Self-Distillation based detector (SSD-Det) to integrate the spatial cues into the bounding box refinement. SSD-Det has two important components: the Spatial Position Self-Distillation (SPSD) module for the bag construction step and the Spatial Identity Self-Distillation (SISD) module for the proposal selection step. To construct high-quality proposal bags, SPSD utilizes a neighborhood sampler to generate a balanced and flexible initial proposal bag for each object and then trains a regressor with the supervision of the annotated inaccurate bounding boxes. Finally, high-quality proposal bags are constructed with proposals corrected by the regressor. The mechanism behind SPSD is that the network learns the spatial information from the reliable samples, e.g. those low-noise annotations, in the dataset and then guides the noisy samples to produce high-quality proposals, as shown in Fig. 2. In addition, to further combine the category information and the spatial information, an interactive structure is implemented by alternately using SPSD to mine spatial cues and MIL to utilize the category information. With SPSD and the interactive structure, a high-quality proposal bag can be constructed. Experiments on MS-COCO show that SPSD can significantly improve the mean/max IoU between objects and proposals + +![](images/842a39994e87835a22dd9a60227450ab9bb3ff2e3047497d078fd57f2beee266.jpg) +Figure 2: The mechanism of Spatial Self-Distillation. By assigning higher weight, the low-noise annotations can be seen as reliable samples to guide the training of proposals' spatial position and identity learning in SPsD and SISD. + +(about 18/10 points, Fig. 5) in the constructed bag. Instead of selecting proposals by classification confidence, we have proposed the SISD module in the proposal selection step. We use it to obtain each proposal's spatial confidence by predicting the IoU with the object and combining the IoU with classification confidence to select the top- $k$ proposals. It is worth mentioning that SISD is an object-related IoU predictor, which means that the predicted IoU may be different for the same proposal that appears in different objects' bags. Accordingly, it guarantees that SISD can better handle object drift and group prediction problems. Experiments on MS-COCO and VOC datasets verify the effectiveness of our method and bring state-of-the-art performance. The contributions are as follows: + +1) We further investigate the inaccurate-box supervised object detection tasks and propose an end-to-end training SSD-Det that combines the spatial and the category information in an interactive fashion. +2) We utilize an SPSD module to generate higher-quality proposals sampling through statistic-guide spatial position distillation, raising the upper bound of the refinement. +3) To add spatial cues to classification confidence, we also introduce an SISD module to select a proposal belonging to the object rather than the category. +4) The performance of our proposed SSD-Det improves the mean average precision (AP) of the best previous method (e.g. over 10 AP on $40\%$ noisy MS-COCO) and achieves state-of-the-art under various noise rate box supervision on MS-COCO and VOC datasets. + +# 2. Related Work + +# 2.1. Object Detection + +Classic object detection [15, 38, 37, 32, 30, 3, 43, 57] is supervised by an accurate bounding-box. One-stage detectors utilize anchors as the sliding-window, such asYOLO [37], SSD [32], and RetinaNet [30]. Two-stage detectors mine spatial information to predict proposals (e.g. selective search [47] in Fast R-CNN [15] or RPN in Faster R-CNN [38]) and conduct classification and bounding-box + +regression with filtered proposals sparsely. Transformer-based (i.e. DETR [3], Deformable-DETR [62], and SwinTransformer [33]) detectors utilize global information for better representation. Sparse R-CNN [43] combines a transformer's advantages and CNNs for detection. + +# 2.2. Weakly-Supervised Object Detection (WSOD) + +WSOD trains object detectors with image tag supervision. Only with the category annotation, the majority of previous methods treat each image as a bag and candidate proposals as instances. They follow the multiple instance learning (MIL) pipeline [1, 44, 46, 7, 48], which highly depends on category information. However, the MIL loss function leads to a non-convex optimization problem; thus, MIL solutions are usually stuck into the local minima. Context information [24, 51], spatial regularization [1, 9, 48], and optimization strategy [44, 48, 46] are proposed to address the problems. SPE [28] introduces Transformer into WSOD and uses attention to generate proposals. SD-LocNet [59] tackles the initialized noisy object locations in WSOD and proposes a self-directed localization network to identify the noisy object instances. [44, 46, 7] use the pseudo label for classification's iterative refinement. However, we use the pseudo box as a better self-distillation teacher. [54, 39] conduct regression to move the proposals, whereas we conduct regression to distill for better bag construction. In this work, we also formulate box correction as a MIL problem. + +# 2.3. Semi-Supervised Object Detection (SSOD) + +Semi-supervised learning in object detection can be roughly categorized into two groups: consistency based [22, 45] and pseudo label based [27, 36, 41, 49, 64, 52]. [52] presents an end-to-end SSOD approach with two simple techniques named soft teacher and box jittering to facilitate the efficient leverage of the teacher model. Both Soft Teacher [52] and SSD-Det obtain pseudo-labels from candidate boxes, adopting distillation, box jitter and classification scores weighting policy. However, they are different: 1) SSOD selects candidate boxes from the teacher model's detection results, while SSD-Det generates them with a generative approach, due to no any accurate supervised data for ensuring high-quality boxes in the detection results. 2) [52] is based on FixMatch [40] and requires maintaining two networks for teacher-student distillation structure. In contrast, our approach only needs a multi-head detector with a shared backbone for self-distillation. 3) Box Jitter: [52] calculates box variance from jittering for result selection, while we aim to generate candidate boxes that combine SISD-predicted IoU and classification scores, selecting boxes closer to the ground truth. 4) We use classification scores as weights for the next stage's loss, while [52] employs them as a criterion for selecting reliable samples. + +# 2.4. Learning with Noisy Annotations + +Training CNNs under noisy labels has been an active research area. Previous research focuses on the classification task, and develops various techniques to deal with noisy labels, such as sample selection [17, 23] for training, label correction [42, 34], and robust loss functions [60, 14] against noisy labels. Recently, many efforts [4, 25, 35, 53, 11, 12] have been devoted to the object detection task. On the one hand, Simon et al. [4] first investigates the impact of different types of label noise on object detection. They propose a per-object co-teaching method to alleviate the effect of noisy labels. On the other hand, [53] proposes a meta-learning framework for noisy annotations consisting of noisy category labels and bounding boxes. [11, 12] utilize object-level MIL to refine the inaccurate box. OAMIL [12] constructs proposal bags through label assignment in a discriminant style. P2BNet [11] originally conducts point-supervised object detection tasks. However, it can be seen as the box correction in its refinement stage. It uses hand-craft anchors to generate proposal bags. Our method inherits the generative style of P2BNet and conducts spatial distillation to mine spatial information. + +# 2.5. Knowledge Distillation + +Knowledge distillation (KD) [21] aims to learn compact and efficient student models guided by excellent teacher networks. It is first applied to object detection in [5], in which hint learning and KD are both used for multi-class object detection. Recently, many efforts [26, 50, 8, 16, 55] aim to mimic the feature. [61] shows that localization knowledge is more important and proposed a localization distillation method. We also transfer the spatial knowledge from reliable labeled instances to correct inaccurate bounding boxes (shown in Fig. 2) in a self-distillation manner. + +# 3. Methodology + +This work aims to learn a robust detector with inaccurate bounding boxes. Instead of training a detector with the original inaccurate bounding box, we follow most related works [11, 12] that design a branch to refine the inaccurate bounding box and then train the detector head or detector with the refined bounding box. The most important part is how to design a refining policy. We first design a two-stage basic box refiner (gray region in Fig. 3) as a naive solution that modified from [11]. Then, SPSD and SISD are proposed and added to further mine the spatial cues for box refinement, yielding SSD-Det. Therefore, the overall loss function is formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {B a s i c}} + \alpha_ {1} \cdot \mathcal {L} _ {\text {S P S D}} + \alpha_ {2} \cdot \mathcal {L} _ {\text {S I S D}} + \alpha_ {3} \cdot \mathcal {L} _ {\text {D e t}} \tag {1} +$$ + +where $\alpha_{1},\alpha_{2}$ and $\alpha_{3}$ are set as 0.25, 0.25 and 4 respectively. $\mathcal{L}_{Det}$ denotes the loss of detector or detection head. During inference, only the detector or the detection head is used. + +![](images/2b0384aa55024bf0fba338b9a639ec4fbe5410b686fbc47463ede4a020def4db.jpg) +Figure 3: The framework of SSD-Det. It contains basic box refiner, SPSD module, SISD module and a detector head. Neighborhood sampler is adopted around the inaccurate annotation. Then, SPSD module generates better proposal bags which are fed into basic box refiner for MIL training. The selected proposals are average weighted as the refined box and supervise the next SPSD training. Meanwhile, the SISD module predicts the IoU between proposals and the object, and the estimated IoU is multiplied by classification score for better proposal selection to generate the refined box. SPSD shares backbone with the detector. + +# 3.1. Basic Box Refiner + +Motivated by [12] and WSOD [1], we design the basic box refiner (detailed structure figure is in supplementary) that leverages classification confidence to refine the inaccurate box annotation. Then the refine annotation is used to train a detection head or detector. Following [11], we design a two-stream structure as a MIL classifier to select the best proposal for box refinement. + +Giving an image with inaccurate box annotation, for each object, $\mathcal{B}$ is a bag of proposals (bounding boxes) that are generated around its inaccurate annotation by a sampler policy (e.g., selective search[47], edge box[63], neighborhood sampler in Sec. 3.2). Meanwhile a feature map is extracted with a backbone network. And then through $7\times 7$ RoIAlign [18] and two fully connected (fc) layers, features of proposal in $\mathcal{B}$ are extracted and denote as $\mathbf{F}$ . The basic box refiner takes proposal bag $\mathcal{B} \in \mathbb{R}^{P \times 4}$ and features $\mathbf{F} \in \mathbb{R}^{P \times D}$ as inputs, where $P$ , $D$ are denoted as the number of proposals in $\mathcal{B}$ , feature dimension respectively. + +Following [11] and [1], as Eq. 2 described, we apply the classification branch $f_{cls}$ to $\mathbf{F}$ yields $\mathbf{O}^{cls}$ , which is then passed through the softmax function over classification dimension $K$ to obtain the score $\mathbf{S}^{cls} \in \mathbb{R}^{P \times K}$ , where $K$ represents the number of instance categories. Likewise, instance selection branch $f_{cls}$ is applied to $\mathbf{F}$ to yield $\mathbf{O}^{ins}$ , and instance score $\mathbf{S}^{ins}$ is obtained through softmax function over $P$ proposals. The proposal score $\mathbf{S}$ is obtained by computing the Hadamard product of the classification score and the instance score. The bag score $\widehat{\mathbf{S}}$ is obtained by the + +summating of the $P$ proposal boxes' proposal scores. + +$$ +\mathbf {O} ^ {c l s} = f _ {c l s} (\mathbf {F}) \in \mathbb {R} ^ {P \times K}; [ \mathbf {S} ^ {c l s} ] _ {p k} = e ^ {[ \mathbf {O} ^ {c l s} ] _ {p k}} / \sum_ {k = 1} ^ {K} e ^ {[ \mathbf {O} ^ {c l s} ] _ {p k}}. +$$ + +$$ +\mathbf {O} ^ {i n s} = f _ {i n s} (\mathbf {F}) \in \mathbb {R} ^ {P \times K}; [ \mathbf {S} ^ {i n s} ] _ {p k} = e ^ {[ \mathbf {O} ^ {i n s} ] _ {p k}} / \sum_ {p = 1} ^ {P} e ^ {[ \mathbf {O} ^ {i n s} ] _ {p k}}. +$$ + +$$ +\mathbf {S} = \mathbf {S} ^ {c l s} \odot \mathbf {S} ^ {i n s} \in \mathbb {R} ^ {P \times K}; \widehat {\mathbf {S}} = \sum_ {p = 1} ^ {P} [ \mathbf {S} ] _ {p} \in \mathbb {R} ^ {K}. \tag {2} +$$ + +where $\left[\cdot \right]_{pk}$ is the value at row $p$ and column $k$ in the matrix. + +The basic box refiner has two similar stages. The loss of stage I(termed $\mathcal{L}_I$ ) adopt the MIL paradigm with the form of cross-entropy (CE) loss, defined as: + +$$ +\mathcal {L} _ {I} = C E (\widehat {\mathbf {S}}, \mathbf {c}) = - \sum_ {k = 1} ^ {K} \mathbf {c} _ {k} \log (\widehat {\mathbf {S}} _ {k}) + (1 - \mathbf {c} _ {k}) \log (1 - \widehat {\mathbf {S}} _ {k}) \tag {3} +$$ + +where $\mathbf{c} \in \{0,1\}^K$ is the one-hot category label. And each object's proposals with the top- $k$ highest proposal score $\mathbf{S}$ are weighted to obtain the refined box. + +The stage II takes the refined box of stage I as input and performs fine refining with a similar structure as stage I. Differently, the focal loss is adopted in stage II instead of cross entropy loss. In order to cooperate with focal loss, the classification branch uses the sigmoid $\sigma(x)$ instead of softmax function and we sample some negative samples $\mathcal{N}$ to further suppress the background. With the bag score $\widehat{S}$ and the negative sample scores $\mathbf{S}_{\text{neg}}^{\text{cls}}$ , the loss is: + +$$ +\mathcal {L} _ {I I} = \left\langle \mathbf {c} ^ {\mathrm {T}}, \widehat {\mathbf {S}} ^ {*} \right\rangle \cdot \operatorname {F L} \left(\widehat {\mathbf {S}}, \mathbf {c}\right) + \sum_ {\mathcal {N}} \beta \cdot \operatorname {F L} \left(\mathbf {S} _ {n e g} ^ {c l s}, c _ {n e g}\right) \} \tag {4} +$$ + +where FL is the focal loss [30], $\widehat{\mathbf{S}}_j^*$ represents the bag score predicted by stage I. $\left\langle \mathbf{c}_j^{\mathrm{T}},\widehat{\mathbf{S}}_j^*\right\rangle$ represents the inner product + +of the two vectors, meaning the predicted bag score of the ground-truth category. $\beta$ is the average of $\left\langle \mathbf{c}_j^{\mathrm{T}},\widehat{\mathbf{S}}_j^*\right\rangle$ . They are used to weight each object's FL for stable training. The overall loss function of the basic refiner here is: + +$$ +\mathcal {L} _ {\text {B a s i c}} = \mathcal {L} _ {I} + \alpha_ {I I} \cdot \mathcal {L} _ {I I} \tag {5} +$$ + +where $\alpha_{II}$ are the loss weights of the two stages. + +During training, the refined box of stage II is used as supervision for a detection head or detector. After training, the basic box refiner will be removed, leaving a well-trained detection head or detector. In this way, we can train a detector under inaccurate annotations. + +# 3.2. Spatial Position Self-Distillation (SPSD) + +Like most MIL paradigm methods, basic box refiner has two main components: bag construction and proposal selection. And the main idea is to use classification information to guide the refining. In this paper, we add spatial information to improve refining. Specifically, SPSD is proposed to use spatial information to enhance bag construction. + +Bag construction aims to obtain proposals for each object, while proposal selection is to select the proposals from the object bag. Then, the refined box is averaged over the selected proposals. Therefore, the quality of the proposals in constructed bag determines the upbound of refining. The bag construction can be implemented in a variety of ways. In this paper, the basic box refiner adopts a naive neighborhood sampler for bag construction. basic box refiner adopts a naive neighborhood sampler for bag construction. + +Neighborhood Sampler. Proposals around the inaccurate box are sampled to construct an object bag. For each inaccurate box $b^{*} = (b_{x}^{*}, b_{y}^{*}, b_{w}^{*}, b_{h}^{*})$ , its scale and aspect ratio with $s$ and $v$ are adjusted and its positions $o_{x}, o_{y}$ are jittered to obtain the diverse proposal $b = (b_{x}, b_{y}, b_{w}, b_{h})$ : + +$$ +\begin{array}{l} b _ {w} = v \cdot s \cdot b _ {w} ^ {*}, \quad b _ {h} = 1 / v \cdot s \cdot b _ {h} ^ {*}, \\ b _ {x} = b _ {x} ^ {*} + b _ {w} \cdot o _ {x}, \quad b _ {y} = b _ {y} ^ {*} + b _ {h} \cdot o _ {y}. \tag {6} \\ \end{array} +$$ + +These proposals $b$ are used to construct the positive proposal bag $\mathcal{B}$ to train the MIL classifier. Thanks to the hand-craft sampling way, the number of proposals in different objects' proposal bags is controllable and balanced. However, the hand-craft neighborhood sampler strategy is difficult to set hyper-parameters, and the sampling space is discrete. For example, when the jitter region is small, the optimization space of refining is limited, while when it is large, more background will be introduced. Hence, we propose the SPSD module to mine spatial information for higher-quality proposal bag construction. + +Statistically Guided Adaptive Sampling. Instead of simply using a neighborhood sampler, we adopt a statistically guided adaptive sampling by adding SPSD modules into the basic box refiner. Taking the constructed proposal bag $\mathcal{B}$ of hand-craft neighborhood sampler as input, the + +RoI features of proposals in $\mathcal{B}$ are extracted and fed into the two shared fc layers to obtain $\mathbf{F}$ . Then a regression fc layer $f_{dis}$ , supervised by the inaccurate annotated bounding box $b^{*}$ , is introduced to predict the adaptive proposal bag $\mathcal{B}^{dis} = f_{dis}(\mathbf{F}) \in \mathbb{R}^{P \times 4}$ , in which the proposals are closer to the object. Later, $\mathcal{B}^{dis}$ as the constructed proposal bag is fed into stage I of basic box refiner. In order to combine category and spatial information, we implement an interactive structure by alternately using SPSD to mine the spatial cues and using MIL in basic box refiner to utilize the category information. Specifically, the refined bounding box $\hat{b}^{*}$ of stage I that selected by the classification confidence is used to supervision of a new SPSD module for stage II. Similar as stage I, The new SPSD takes proposal bag $\mathcal{B}^{dis}$ of handcraft neighborhood sampler as input. Through the RoI align and the two shared fc layers, the feature $\hat{\mathbf{F}}$ is extracted. An extra fc layer $f_{dis}$ is then utilized to conduct further regression. Different with stage I, the obtained $\mathcal{B}^{dis}$ is supervised by the refined $\hat{b}^{*}$ . The loss function of the spatial distillation for adaptive sampling can be defined as $\mathcal{L}_{SPSD}$ in Eq. 7. + +$$ +\mathcal {L} _ {S P S D} = \frac {1}{P} \left\{\sum_ {p = 1} ^ {P} \mathbf {L} _ {\mathbf {1}} \left(\left[ \mathcal {B} ^ {d i s} \right] _ {p}, b ^ {*}\right) + \sum_ {p = 1} ^ {P} \mathbf {L} _ {\mathbf {1}} \left(\left[ \mathcal {B} ^ {\hat {d i s}} \right] _ {p}, b ^ {*}\right) \right\} \tag {7} +$$ + +where the $\mathbf{L}_1$ is the L1 loss function for loose restrictions. + +The idea behind SPSD is that the dataset with inaccurate annotation still has many reliable, high-quality boxes and inaccurate boxes. Supervised by the high-quality boxes statistically, the network can guide those proposals sampled around the inaccurate bounding box to regress to the ground truth. With the self-distillation mechanism, SPSD learns the semantic-spatial correspondence knowledge from the reliable samples in the dataset and then propagates the knowledge to produce high-quality proposals. + +Adaptive Negative Sampling. Negative samples are introduced in Stage II to better suppress the background. With the sampled $\mathcal{B}^{dis}$ , we can adaptively sample the negative samples with a small IoU (set smaller than 0.3 by default) with all positive proposals in all bags, to compose the negative sample set $\mathcal{N}$ for stage II. + +# 3.3. Spatial Identity Self-Distillation (SISD) + +The basic box refiner selects the proposals only depending on classification confidence during proposal selection. To select the proposal which has high classification confidence and is also spatially close to the object from the bag, we propose a SISD module to predict the IoU between proposals and their corresponding object. Afterwards, through the combination between the IoU and the classification confidence, top- $k$ proposals are selected. In SISD, we design an Object Relevance Enhancement (ORE) module to distinguish different objects' features with the same RoI region. And an identity predictor is designed to predict each proposal's IoU with the object. + +
MethodBackbone20% Box Noise Level40% Box Noise Level
AP\(AP_{50}\)\(AP_{75}\)\(AP^{s}\)\(AP^{m}\)\(AP^{l}\)AP\(AP_{50}\)\(AP_{75}\)\(AP^{s}\)\(AP^{m}\)\(AP^{l}\)
Val Set
Clean-FasterRCNN [38]ResNet-5037.958.140.921.641.648.737.958.140.921.641.648.7
Clean-FasterRCNN [38]ResNet-10139.460.143.122.443.751.139.460.143.122.443.751.1
Clean-Retinanet [30]ResNet-5036.756.139.021.640.447.436.756.139.021.640.447.4
Noisy-FasterRCNN [38]ResNet-5030.454.331.417.433.938.710.328.93.35.711.815.1
Noisy-Retinanet [30]ResNet-5030.053.130.817.933.738.213.333.65.78.415.918.0
FreeAnchor[58]ResNet-5028.653.128.516.632.237.010.428.93.35.812.114.9
Co-teaching[17]ResNet-5030.554.930.517.334.039.111.531.44.26.413.116.4
SD-LocNet[59]ResNet-5030.054.530.317.533.638.711.330.34.36.012.716.6
KL loss[20]ResNet-5031.054.332.418.034.939.512.136.73.76.213.017.4
OA-MIL[12]ResNet-5032.155.333.218.135.841.618.642.612.99.219.926.5
SSD-DetResNet-5033.657.335.319.537.243.327.653.926.016.031.034.9
SSD-DetResNet-10134.357.636.719.138.144.328.454.327.216.531.936.4
SSD-Det+FRResNet-5034.457.336.820.038.244.029.354.829.017.132.936.9
SSD-Det+FRResNet-10136.259.139.220.940.247.130.656.730.718.134.539.0
Test Set
Clean-FasterRCNN [38]ResNet-5037.758.740.821.740.646.737.758.740.821.740.646.7
Noisy-FasterRCNN [38]ResNet-5030.754.931.318.033.737.710.429.03.36.011.314.6
OA-MIL[12]ResNet-5032.355.833.718.535.040.218.542.312.89.319.125.1
SSD-DetResNet-5033.557.335.519.136.041.928.054.126.516.530.034.5
SSD-Det+FRResNet-5034.757.937.220.037.742.729.755.629.317.532.436.2
+ +Object Relevance Enhancement (ORE). ORE enhances object-relevant features, making SISD an object-relevant IoU predictor. ORE allows the predicted IoU to be different for the same proposal in other objects' bags. In addition, we integrate the feature of the bag's corresponding object into the proposal feature, making the feature of different bags' proposals distinct. That is the so-called ORE. For a proposal bag $\mathcal{B}$ , the feature $\mathbf{F}$ is obtained through the RoI align and two fc layers. It is worth mentioning that the two fcs do not share the parameters with those in the refiner since the optimization goals are contradictory. To represent the feature of the bag's corresponding object, $\mathbf{F}^{+} \in \mathbb{R}^{1 \times D}$ is calculated by averaging features of $P$ proposals in proposal bag $\mathcal{B}$ . The object feature $\mathbf{F}^{+}$ is broadcast into $\mathbb{R}^{P \times D}$ , and then added to the proposal features to obtain the object-relevant features $\mathbf{F}^{*} = \mathbf{F} + \mathbf{F}^{+}$ . + +Spatial Identity Prediction. By a following identity fc layer, $U \in \mathbb{R}^{P \times 1}$ is predicted. The pseudo label $T \in (0,1)$ is IoU between proposals in $\mathcal{B}$ and the merged box $\hat{b}^*$ of stage I. For better optimization, the linear normalized $T' = (T - 0.5)/0.5 \in (-1,1)$ is utilized as supervision. The object function of the identity predictor is identified in: + +$$ +\mathcal {L} _ {S I S D} = \operatorname {s m o o t h} _ {L 1} \left(U, T ^ {\prime}\right) \tag {8} +$$ + +where the smooth $L_{1}$ represents the smooth L1 loss. The predicted spatial confidence $U^{\prime}$ is obtained by normalizing the $U$ . Finally, $S^{*} = U^{\prime} \cdot S$ is used to select the top- $k$ proposals for merging as the refined boxes. + +Table 1: Performance comparison on COCO. FR is Faster R-CNN. *-FR refers to a retrained Faster R-CNN (R50+FPN) using refined annotations from SSD-Det for improved performance.. Clean-* and Noisy-* means original and noisy annotation. + +
MethodBackboneBox Noise Level
10%20%30%40%
Clean-FasterRCNN [38]ResNet-5077.2 for clean
Clean-RetinaNet [30]ResNet-5073.5 for clean
Noisy-FasterRCNN [38]ResNet-5076.371.260.142.5
Noisy-RetinaNet [30]ResNet-5071.567.657.945.0
KL loss[20]ResNet-5075.872.764.648.6
Co-teaching[17]ResNet-5075.470.660.943.7
SD-LocNet[59]ResNet-5075.771.560.843.9
FreeAnchor[58]ResNet-5073.067.556.241.6
OA-MIL[12]ResNet-5077.474.370.663.8
SSD-DetResNet-5077.174.871.566.9
+ +Table 2: Performance comparison on the VOC 2007 test set. The evaluation metric is $AP_{50}$ . The Clean-* and Noisy-* means original annotation and noisy annotation. + +# 4. Experiment + +# 4.1. Experimental Settings + +Datasets and Evaluation Metrics. For experimental comparisons, two publicly available datasets are used for object detection with inaccurate bounding boxes: MS-COCO [31] and PASCAL VOC 2007 [13]. MS-COCO (2017 version) has 118k training and 5k validation images with 80 common object categories. PASCAL VOC 2007 is one of the most popular benchmarks in generic object detection with 20 classes. + +Evaluation Metric. We use mean average precision + +
2-RefSPSDSISDRe-Train20% Box Noise Level40% Box Noise Level
AP\(AP_{50}\)\(AP_{75}\)\(AP^{s}\)\(AP^{m}\)\(AP^{l}\)\(AP^{test}\)\(AP^{test}_{75}\)AP\(AP_{50}\)\(AP_{75}\)\(AP^{s}\)\(AP^{m}\)\(AP^{l}\)\(AP^{test}\)\(AP^{test}_{75}\)
30.057.129.016.933.139.8--22.851.116.113.325.030.4--
✓31.256.731.617.834.541.031.432.024.652.020.114.328.231.925.020.5
✓✓33.056.934.818.735.542.233.134.827.253.724.715.930.335.227.625.6
✓✓✓33.657.335.319.537.243.333.535.527.653.926.016.031.034.928.026.5
✓✓31.856.833.118.435.740.832.333.726.554.023.315.730.333.826.823.3
✓✓✓34.157.636.419.037.743.834.336.629.055.127.817.032.536.729.328.4
✓✓✓✓34.457.336.820.038.244.034.737.229.354.829.017.132.936.929.729.3
+ +Table 3: Modules ablation of SPSD, SISD and Re-Train on MS-COCO validation set (without) and test set (with test). The Re-Train means we generate the pseudo label by SSD-Det and re-train a Faster R-CNN detector. + +
Methods (w/o SISD)APAP50AP75APsAPmAPl
Neighborhood Sampler24.652.020.114.328.231.9
SPSD (II) w/o weighted26.053.322.515.629.433.4
SPSD (II) w/ weighted26.353.422.515.629.333.8
SPSD (I+II) w/ weighted27.253.724.715.930.335.2
+ +Table 4: Different setting of SPSD. + +
ORE Strategies of SISDAP\( AP_{50} \)\( AP_{75} \)\( AP^s \)\( AP^m \)\( AP^l \)
w/o SISD27.253.724.715.930.335.2
SISD w/o ORE27.353.325.716.930.235.0
+ subtract27.253.824.615.730.435.0
+ concatenate27.254.024.516.130.234.8
+ add27.653.926.016.031.034.9
+ add w/ shared fcs23.049.917.612.625.330.4
+ +Table 5: Different ORE strategies of SISD. + +
Num.0123
AP/AP5024.6 / 52.026.3 / 53.427.2 / 53.727.0 / 53.1
+ +Table 6: Number of SPSD module. + +mAP@[.5,.95] and (mAP@.5) for MS-COCO and VOC. The $\{AP, AP_{50}, AP_{75}, AP^{Small}, AP^{Middle}, AP^{Large}\}$ is reported for MS-COCO and $AP_{50}$ for VOC. + +Synthetic Noisy Dataset. Following [12], We simulate noisy bounding boxes by perturbing clean boxes from the original annotations. The details are in the appendix. We simulate various box noise levels ranging from $10\%$ to $40\%$ for the VOC and $\{20\%, 40\}$ for the MS-COCO. + +Implementation Details. We implement our method on FasterRCNN [38] with ResNet50-FPN [19, 29] backbone, based on MMDetection [6]. All settings of our method and previous methods employ FPN for fair comparison. Similar to the default setting of object detection on MS-COCO, the stochastic gradient descent [2] algorithm is used to optimize on 1x training schedule. The batch size is two images per GPU on 8 GPUS. For the VOC dataset, the batch size is two images per GPU on 2 GPUS. The performance we report is on a single scale (1333 * 800 for MS-COCO and 1000 * 600 for VOC). + +# 4.2. Comparison with State-of-the-Art + +We compare our method with several state-of-the-art approaches [20, 17, 59, 58, 12] on MS-COCO and VOC 2007 datasets. We denote Clean-FasterRCNN and Noisy-FasterRCNN as FasterRCNN models trained under clean (original annotations) and noisy annotations with the default setting, respectively. + +MS-COCO Dataset. Table 1 shows the comparison results on the MS-COCO. Inaccurate bounding box annotations significantly deteriorate the vanilla Faster R-CNN's detection performance. Co-teaching and SD-LocNet only slightly improve the detection performance, especially under $40\%$ box noise. That indicates that small-loss sample selection and sample weight assignment can not tackle noisy box annotations well. KL Loss slightly improves the performance under $20\%$ and $40\%$ box noise. By treating an object as a bag of instances, OA-MIL is somehow robust to noisy bounding boxes and performs better than other methods. Nevertheless, the previously-mentioned label assignment bag construction limits its ability to handle heavy noise. Our approach is more robust to noisy bounding boxes. It outperforms other methods by a large margin under high box noise levels and significantly boosts the detection performance across all metrics. For example, under $40\%$ box noise, the end-to-end SSD-Det achieves 27.6 $AP$ and $53.9 AP_{50}$ , 9.0 and attains 11.3 point improvement compared with state-of-the-art method OA-MIL, respectively. Also, through re-training on FasterRCNN, the performance further reaches 29.3 $AP$ and $54.8 AP_{50}$ . With the backbone of ResNet-101, the performance achieves consistent improvement. On MS-COCO test set, our method also achieves state-of-the-art performance. + +VOC 2007 Dataset. Table 2 shows the comparison results on the VOC 2007 test set. Co-teaching, SD-LocNet and KL Loss, can not address inaccurate bounding box annotations well. OA-MIL improves the performance on different noisy datasets. Our approach obtains further improvements to 77.10, 74.80, 71.50, 66.90 $AP_{50}$ on $10\%$ , $20\%$ , $30\%$ and $40\%$ noisy box datasets, respectively. + +![](images/b82a78ea638bd787cba4f76d854879baa5c65554b6052c89d801317cd5335c24.jpg) +Figure 4: Qualitative detection results on COCO validation. Previous methods miss objects and face part prediction problems. Our method misses fewer objects, and the bounding box quality is better, especially for small or overlapped objects. + +![](images/c42efc6fd3a318138ab97195e2f6e7c68a77bf843ba4470add0c21737a08ff89.jpg) + +![](images/e91598751ae71b1b39edb2b75603285d83032dec869312aed2bd3b38ebb1ac5b.jpg) + +![](images/93cb516e89418507075d332120a024065beccae3bea40e96ec6861433fc1725b.jpg) + +
MethodsBox Refiner+Re-TrainSSD-DetSSD-Det+Re-Train
AP/AP5029.0 / 54.427.6/53.929.3 / 54.8
+ +# 4.3. Ablation Study and Analysis + +To further analyze SSD-Det's effectiveness and robustness, we conduct more experiments on COCO val set if there are no other instructions. Except for Table 3, the noise level of these experiments is $40\%$ . + +Ablation of Modules. Ablation study of each component in our approach is given in Table 3, including: (i) Different stages of our basic box refiner. i.e. training object detector without the stage II (2-Ref), where the pseudo boxes predicted by the stage I are served as the supervision for training a parallel detector. (ii) SPSD, i.e. training without SPSD, where the object-bag is constructed directly by neighborhood sampling around the noisy ground-truth or the predicted pseudo boxes of the stage I. (iii) SISD. (iv) Re-Train with FasterRCNN (Re-Train). + +Effectiveness of SPSD. SPSD further improves the detection performance on the MS-COCO, especially under high box noise levels, e.g. under $40\%$ box noise level, SPSD boosts the performance from 24.6 to 27.2, as shown in Table 3 (row 3). In Table 4, we conduct further ablation on SPSD. With SPSD bag construction only in stage II, the performance increases by 1.4 AP. The performance further improves with the proposal score of stage I as weights. With SPSD in all stages, the AP reaches 27.2. Fig. 5 shows the bag quality. With SPSD, the mean IoU increases from 40.3 to 58.7 and the max and top-10 IoU increase to 78.3 and 75.1, which indicates a better upper bound of proposal selection. More high-quality proposals bring better optimization and easier proposals selection. + +Number of SPSD. As shown in Table 6. When adding + +Table 7: Comparisons of end-to-end and Re-Train. + +
DetectorsClean-supervisedNoise-supervised
APAP50APAP50(w/ours)
APAP50
FasterRCNN37.958.110.328.929.354.8
SparseRCNN [43]†45.064.16.020.334.360.2
De-DETR [62]†46.866.35.016.935.260.9
+ +Table 8: Experiments on advanced detectors. De-DETR is Deformable DETR. † uses multi-scale data augment. ‘ $w/$ ours’ means using our method under noisy supervision. + +![](images/3095a5b20676f90b4c2cb6035846f78ae76c3199ff9b55c994e70cb8855d99b9.jpg) +Figure 5: Bag quality (IoU of proposals with GT) of construction in SSD-Det. B.S. (blue) means neighborhood sampler. SPSD I (orange) denotes single SPSD adopted. SPSD II (yellow) is two SPSD and interactive structure adopted. SPSD II significantly improves the quality. + +3 SPSD, performance drops slightly, probably due to the accumulation of errors outweighing the performance gain from extra stages. Hence, 2 SPSD is our default setting. + +Effectiveness of SISD. SISD is designed to select object-aware proposals in box selection. Under $40\%$ and $20\%$ box noise, the detection performance improves from 27.2 to 27.6 and 33.0 to 33.6, which verifies the effectiveness of the module, shown in Table 3. We also study the + +
MethodsDrift Rate (%)Group Rate (%)Part Rate (%)
allsmlallsmlallsml
OA-MIL[12]15.117.817.17.46.72.83.41.42.83.42.72.3
Ours1.51.01.31.41.71.20.50.71.00.51.11.3
+ +Table 9: Breakdown of different problems during refinement (COCO under $40\%$ noise level). s, m and 1 mean small, middle and large scale. + +
DatasetourQuality (Average IoU)Frequency (%)
AllPartOversizeShiftReliablePartOversizeShift
Detector as Annotator
Objects-F✓44.324.133.613.340.128.49.821.7
47.030.841.816.849.223.18.019.4
COCO-F✓45.125.433.315.740.027.512.020.5
48.232.140.619.649.522.210.018.3
Point-based Annotator
COCO-P✓55.630.425.029.865.69.523.21.7
65.246.736.440.974.95.918.50.6
+ +Table 10: Analysis of noisy annotations types and quality. + +strategies of ORE in SISD (Table 5). The minus or concatenation of object feature $\mathbf{F}_j^+$ and proposal feature $\mathbf{F}_j$ do not work. With add strategy, the performance is 27.60. If SISD shares the two fc layers, the performance drops to 22.99 since the optimization goals are contradictory (Identity distinguishes objects in the same category). If we directly use the RoI feature without ORE, the performance drops to 27.32 AP, verifying the effectiveness of the object relevance strategy. + +Affect of Re-Train. As most WSOD methods do, we re-run the experiments by training a fully supervised detector for better performance. We find that if the SSD-Det only trains the refiner and uses the pseudo label to train the FasterRCNN, the result is good but lower than re-train after the end-to-end training given in Table 7 (row 1). This is because joint training is beneficial for box refinement. + +Experiments on Advanced Detectors. We re-train recent detectors, e.g. SparseRCNN and Deformable DETR, under the boxes refined by our method. Table 8 verifies that our method achieves consistency improvement. + +The computational cost discussion. Similar to OAMIL, our method adds SSD head to Faster R-CNN for auxiliary training to refine the noisy annotation and the head is not used during inference, the calculation cost during inference is same to standard Faster R-CNN. + +# 4.4. Visualization and Discussion. + +Fig. 4 shows that OA-MIL faces missing instances and grouping instances issues for small or overlapped objects (as mentioned in [12]), while our method still works well. For a better intuitive understanding of SISD and SPSD, we visualize the bag construction quality in Fig. 5 Then, we makes noise types breakdown of 'Drift', 'Group' and 'part dominance' issues. We give the definition of IoU, IoG and + +IoD: + +$$ +I o U = \frac {A (I)}{A (D) + A (G) - A (I)}, I o G = \frac {A (I)}{A (G)}, I o D = \frac {A (I)}{A (D)} \tag {9} +$$ + +where $\mathrm{A}(\ast)$ is area of box *, D and G are refined box and gt box respectively, and I is insertion between D and G. We statistically count the proportion of three noise types of 'bad' refined boxes (having small IoU with gt) in Table 9: (i) Drift: 'bad' refined box has a higher IoU with another nearby object. (ii) Group: 'bad' refined box has high IoG with multiple objects. (iii) Part: 'bad' refined box has a high IoD. Table 9 shows quantitative results for each noise type of baseline and ours. The drift, group, part problems reduce from $15.1\%$ , $6.7\%$ , $2.8\%$ to $1.5\%$ , $1.7\%$ , $1.0\%$ , respectively, demonstrating our improvement. + +Experiments on Real-life Noisy Annotations. Real-life noisy annotations stem from: low-quality data (e.g., occlusion, blur), human annotator errors and automatic machine annotator limitations. Noise from human errors is quite subjective, since differences between annotators. For a more objective analysis, noisy annotations from machine annotator is used for experiments. Without loss of generality, Faster R-CNN well-trained on MS-COCO was applied to Objects365 images, yielding Objects-F dataset, and to COCO-val images, producing COCO-F dataset. P2BNet [11], a point-based annotator, was used on COCO-val images with point annotations, generating COCO-P dataset. SSD-Det effectively improves low-quality boxes. As shown in Table 10, with SSD-Det's refinement, the average IoU increases for Objects-F (from 44.3 to 47.0), COCO-F (from 45.1 to 48.2) and COCO-P (from 55.6 to 65.2) datasets. Further, the proportion of reliable annotations increases, and noise categories' frequency (Part, Oversize, and Shift) decreases for all datasets. + +# 5. Conclusion + +This paper investigates problems during refinement caused by solely using category information to select proposals. We also propose SSD-Det to mine spatial information in a self-distillation fashion. SSD-Det introduces the SPSD module to learn semantic-spatial correspondence knowledge with neighborhood sampler and an interactive structure to combine spatial information and category information, thus producing a high-quality proposal bag. SISD in SSD-Det is utilized to improve the proposal selection procedure by integrating object-relevant spatial confidence. Complete ablations on multiple datasets verify the effectiveness of SSD-Det. + +# References + +[1] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In CVPR, 2016. 1, 3, 4 + +[2] Léon Bottou. Stochastic gradient descent tricks. In Neural Networks: Tricks of the Trade - Second Edition. Springer, 2012. 7 +[3] Nicolas Carion, Francisco Massa, and Gabriel Synnaeve et al. End-to-end object detection with transformers. In ECCV, 2020. 2, 3 +[4] Simon Chadwick and Paul Newman. Training object detectors with noisy data. In IV, 2019. 1, 3 +[5] Guobin Chen, Wongun Choi, and Xiang Yu et al. Learning efficient object detection models with knowledge distillation. In NeurIPS, 2017. 3 +[6] Kai Chen, Jiaqi Wang, and Jiangmiao Pang et al. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7 +[7] Ze Chen, Zhihang Fu, and Rongxin Jiang et al. SLV: spatial likelihood voting for weakly supervised object detection. In CVPR, 2020. 2, 3 +[8] Xing Dai, Zeren Jiang, and Zhao Wu et al. General instance distillation for object detection. In CVPR, 2021. 3 +[9] Ali Diba, Vivek Sharma, and Ali Mohammad Pazandeh et al. Weakly supervised cascaded convolutional networks. In CVPR, 2017. 3 +[10] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the multiple instance problem with axis-parallel rectangles. Artificial Intelligence, 1997. 2 +[11] Chen, et al. Point-to-box network for accurate object detection via single point supervision. In ECCV, 2022. 1, 2, 3, 4, 9 +[12] Liu, et al. Robust object detection with inaccurate bounding boxes. In ECCV, 2022. 1, 2, 3, 4, 6, 7, 9 +[13] Mark Everingham, Luc Van Gool, and Christopher K. I. Williams et al. The pascal visual object classes (VOC) challenge. *IJCV*, 2010. 6 +[14] Aritra Ghosh, Himanshu Kumar, and P. S. Sastry. Robust loss functions under label noise for deep neural networks. In AAAI, 2017. 3 +[15] Ross B. Girshick. Fast R-CNN. In ICCV, 2015. 2 +[16] Jianyuan Guo, Kai Han, and Yunhe Wang et al. Distilling object detectors via decoupled features. In CVPR, 2021. 3 +[17] Bo Han, Quanming Yao, and Xingrui Yu et al. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, 2018. 3, 6, 7 +[18] Kaiming He, Georgia Gkioxari, and Piotr Dolkar et al. Mask R-CNN. In ICCV, 2017. 1, 4 +[19] Kaiming He, Xiangyu Zhang, and Shaoqing Ren et al. Deep residual learning for image recognition. In CVPR, 2016. 7 +[20] Yihui He, Chenchen Zhu, and Jianren Wang et al. Bounding box regression with uncertainty for accurate object detection. In CVPR, 2019. 1, 6, 7 +[21] Geoffrey Hinton and Oriol et al. Vinyls. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 3 +[22] Jisoo Jeong, Seungeui Lee, and Jeesoo Kim et al. Consistency-based semi-supervised learning for object detection. In NeurIPS, 2019. 3 +[23] Lu Jiang, Zhengyuan Zhou, and Thomas Leung et al. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 3 + +[24] Vadim Kantorov, Maxime Aquab, and Minsu Cho et al. Contextlocnet: Context-aware deep network models for weakly supervised localization. In ECCV, 2016. 3 +[25] Junnan Li, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Towards noise-resistant object detection with noisy annotations. CoRR, abs/2003.01285, 2020. 1, 3 +[26] Quanquan Li, Shengying Jin, and Junjie Yan. Mimicking very efficient network for object detection. In CVPR, 2017. 3 +[27] Yandong Li, Di Huang, and Danfeng Qin et al. Improving object detection with selective self-supervised self-training. In ECCV, 2020. 3 +[28] Mingxiang Liao, Fang Wan, and Yuan Yao et al. End-to-end weakly supervised object detection with sparse proposal evolution. In ECCV, 2022. 3 +[29] Tsung-Yi Lin, Piotr Dólar, and Ross B. Girshick et al. Feature pyramid networks for object detection. In CVPR, 2017. 7 +[30] Tsung-Yi Lin, Priya Goyal, and Ross B. Girshick et al. Focal loss for dense object detection. In ICCV, 2017. 1, 2, 4, 6 +[31] Tsung-Yi Lin and Michael et al. Maire. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6 +[32] Wei Liu, Dragomir Anguelov, and Dumitru Erhan et al. SSD: single shot multibox detector. In ECCV, 2016. 2 +[33] Ze Liu, Yutong Lin, and Yue Cao et al. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3 +[34] Xingjun Ma, Yisen Wang, and Michael E. Houle et al. Dimensionality-driven learning with noisy labels. In ICML, 2018. 3 +[35] Jiafeng Mao, Qing Yu, and Kiyoharu Aizawa. Noisy localization annotation refinement for object detection. TIS, 2021. 1, 3 +[36] Ilija Radosavovic, Piotr Dólar, and Ross B. Girshick et al. Data distillation: Towards omni-supervised learning. In CVPR, 2018. 3 +[37] Joseph Redmon, Santosh Kumar Divvala, and Ross B. Girshick et al. You only look once: Unified, real-time object detection. In CVPR, 2016. 2 +[38] Shaoqing Ren, Kaiming He, and Ross B. Girshick et al. Faster R-CNN: towards real-time object detection with region proposal networks. TPAMI, 2017. 1, 2, 6, 7 +[39] Zhongzheng Ren, Zhiding Yu, and Xiaodong Yang et al. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In CVPR, 2020. 2, 3 +[40] Kihyuk Sohn, David Berthelot, and Nicholas Carlini et al. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 3 +[41] Kihyuk Sohn, Zizhao Zhang, and Chun-Liang Li et al. A simple semi-supervised learning framework for object detection. CoRR, 2020. 3 +[42] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. SELFIE: refurbishing unclean samples for robust deep learning. In ICML, 2019. 3 +[43] Peize Sun, Rufeng Zhang, and Yi Jiang et al. Sparse R-CNN: end-to-end object detection with learnable proposals. In CVPR, 2021. 2, 3, 8 + +[44] Peng Tang and Xinggang Wang et al. Multiple instance detection network with online instance classifier refinement. In CVPR, 2017. 2, 3 +[45] Peng Tang, Chetan Ramaiah, and Yan Wang et al. Proposal learning for semi-supervised object detection. In WACV, 2021. 3 +[46] Peng Tang, Xinggang Wang, and Song Bai et al. PCL: proposal cluster learning for weakly supervised object detection. TPAMI, 2020. 3 +[47] Koen E. A. van de Sande, Jasper R. R. Uijlings, and Theo Gevers et al. Segmentation as selective search for object recognition. In ICCV, 2011. 2, 4 +[48] Fang Wan, Pengxu Wei, and Zhenjun Han et al. Min-entropy latent model for weakly supervised object detection. TPAMI, 2019. 3 +[49] Keze Wang, Xiaopeng Yan, and Dongyu Zhang et al. Towards human-machine cooperation: Self-supervised sample mining for object detection. In CVPR, 2018. 3 +[50] Tao Wang, Li Yuan, and Xiaopeng Zhang et al. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 3 +[51] Yunchao Wei, Zhiqiang Shen, and Bowen Cheng et al. TS ^2 2 C: tight box mining with surrounding segmentation context for weakly supervised object detection. In ECCV, 2018. 3 +[52] Mengde Xu, Zheng Zhang, and Han Hu et al. End-to-end semi-supervised object detection with soft teacher. In ICCV. IEEE, 2021. 3 +[53] Youjiang Xu, Linchao Zhu, and Yi Yang et al. Training robust object detectors from noisy category labels and imprecise bounding boxes. TIP, 2021. 1, 3 +[54] Ke Yang, Dongsheng Li, and Yong Dou. Towards precise end-to-end weakly supervised object detection network. In ICCV, 2019. 3 +[55] Zhendong Yang, Zhe Li, and Xiaohu Jiang et al. Focal and global knowledge distillation for detectors. In CVPR, 2022. 3 +[56] Xuehui Yu, Pengfei Chen, and Di Wu et al. Object localization under single coarse point supervision. In CVPR, 2022. 2 +[57] Xuehui Yu, Yuqi Gong, and Nan Jiang et al. Scale match for tiny person detection. In WACV, 2020. 1, 2 +[58] Xiaosong Zhang, Fang Wan, and Chang Liu et al. Freeanchor: Learning to match anchors for visual object detection. In NeurIPS, 2019. 1, 6, 7 +[59] Xiaopeng Zhang, Yang Yang, and Jiashi Feng. Learning to localize objects with noisy labeled instances. In AAAI, 2019. 3, 6, 7 +[60] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In NeurIPS, 2018. 3 +[61] Zhaohui Zheng, Rongguang Ye, and Ping Wang et al. Localization distillation for dense object detection. In CVPR, 2022. 3 +[62] Xizhou Zhu, Weijie Su, and Lewei Lu et al. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3, 8 + +[63] C. Lawrence Zitnick and Piotr Dálár. Edge boxes: Locating object proposals from edges. In ECCV, 2014. 4 +[64] Barret Zoph, Golnaz Ghiasi, and Tsung-Yi Lin et al. Rethinking pre-training and self-training. In NeurIPS, 2020. 3 \ No newline at end of file diff --git a/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/images.zip b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c1ef868a4914be30729ce19fd37a8b58e7a98cc7 --- /dev/null +++ b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b51c2c7f6f1d393f928a9b120e79850e202ca9b36012fa1051260d8b46716d1 +size 891111 diff --git a/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/layout.json b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7d99102886dc4300dbe9b1a5fab97f1d28e798a6 --- /dev/null +++ b/2023/Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes/layout.json @@ -0,0 +1,10074 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "text", + "content": "Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 143, + 484, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 143, + 484, + 159 + ], + "spans": [ + { + "bbox": [ + 110, + 143, + 484, + 159 + ], + "type": "text", + "content": "Di Wu*, Pengfei Chen*, Xuehui Yu*, Guorong Li, Zhenjun Han†, Jianbin Jiao" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 160, + 408, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 160, + 408, + 171 + ], + "spans": [ + { + "bbox": [ + 184, + 160, + 408, + 171 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 224, + 290, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 290, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 290, + 511 + ], + "type": "text", + "content": "Object detection via inaccurate bounding boxes supervision has boosted a broad interest due to the expensive high-quality annotation data or the occasional inevitability of low annotation quality (e.g. tiny objects). The previous works usually utilize multiple instance learning (MIL), which highly depends on category information, to select and refine a low-quality box. Those methods suffer from object drift, group prediction and part domination problems without exploring spatial information. In this paper, we heuristically propose a Spatial Self-Distillation based Object Detector (SSD-Det) to mine spatial information to refine the inaccurate box in a self-distillation fashion. SSD-Det utilizes a Spatial Position Self-Distillation (SPSD) module to exploit spatial information and an interactive structure to combine spatial information and category information, thus constructing a high-quality proposal bag. To further improve the selection procedure, a Spatial Identity Self-Distillation (SISD) module is introduced in SSD-Det to obtain spatial confidence to help select the best proposals. Experiments on MS-COCO and VOC datasets with noisy box annotation verify our method's effectiveness and achieve state-of-the-art performance. The code is available at https://github.com/ucas-vg/PointTinyBenchmark/tree/SSD-Det." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 534, + 128, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 128, + 546 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 128, + 546 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 554, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 288, + 685 + ], + "type": "text", + "content": "Object detection [18, 38, 58, 30, 57] relying on large-scale datasets like MS-COCO[31] has significantly progressed and achieved good performance. However, accurate bounding box annotations are expensive and challenging in natural contexts [12]. Especially in many professional scenarios, it is difficult to label accurate annotations without domain knowledge (e.g., agricultural crop observation and medical image processing) [12, 1]. As shown in Fig. 1a in some complex datasets, the human annotators may also annotate inaccurate bounding boxes due to the inherent ambiguities[20] of objects. In addition, labelling with a de" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 196, + 550, + 335 + ], + "blocks": [ + { + "bbox": [ + 310, + 196, + 550, + 335 + ], + "lines": [ + { + "bbox": [ + 310, + 196, + 550, + 335 + ], + "spans": [ + { + "bbox": [ + 310, + 196, + 550, + 335 + ], + "type": "image", + "image_path": "596443fcd81f05aca705639a72bf222290bab867abf7f86375b54859c6b93eb8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 311, + 351, + 549, + 459 + ], + "blocks": [ + { + "bbox": [ + 312, + 338, + 544, + 350 + ], + "lines": [ + { + "bbox": [ + 312, + 338, + 544, + 350 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 544, + 350 + ], + "type": "text", + "content": "(a) Labelling strategies lead to inaccurate annotations (red box)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 351, + 549, + 459 + ], + "lines": [ + { + "bbox": [ + 311, + 351, + 549, + 459 + ], + "spans": [ + { + "bbox": [ + 311, + 351, + 549, + 459 + ], + "type": "image", + "image_path": "cb1db177af5512fe3557cb3d1d236228818d1612d98f4c281ec3df3f45101fec.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 464, + 550, + 531 + ], + "lines": [ + { + "bbox": [ + 306, + 464, + 550, + 531 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 550, + 531 + ], + "type": "text", + "content": "(b) Three problems during previous MIL refining. Because their selections depend solely on classification, (i,ii): the refined box (green box) drifts to another object or makes a group prediction (merging across multiple objects) due to neighbor disturbance. Yellow boxes are proposals in the middle person's MIL bag. (iii): Local part may be more discriminative than the entire object and will be predicted." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 540, + 545, + 564 + ], + "lines": [ + { + "bbox": [ + 306, + 540, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 545, + 564 + ], + "type": "text", + "content": "Figure 1: The sources of inaccurate box annotations and three problems caused by previous refinement methods." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 590, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 639 + ], + "type": "text", + "content": "tector or weak signal[11] (e.g., point) is much cheaper but brings more inaccuracy. Therefore learning robust detectors with inaccurate bounding boxes[12, 4, 35, 53, 25] is a practical and meaningful task and has boosted a broad interest." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 714 + ], + "type": "text", + "content": "To use the inaccurate annotations, most related methods [12, 11] refine the inaccurate annotations as Fig. 1a shows, and then train a detector head or re-train a detector with the refined box as the new supervision. There are two main steps during refining: 1) Bag Construction: For each object, obtaining some proposals around the inaccurate an-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 128, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 128, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 128, + 703 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 703, + 211, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 211, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 211, + 713 + ], + "type": "text", + "content": "† Corresponding authors. (hanzhj@ucas.ac.cn)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6855" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "notated bounding box to form the object-level proposal bag; 2) Proposal Selection: Selecting the top- " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": " proposals with the highest classification confidence from each bag and then weighting average them to obtain the refined box." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 123, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 123, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 289, + 422 + ], + "type": "text", + "content": "During the proposal selection, they usually utilize multiple instance learning (MIL) [10] supervised by category information to choose the proposals with high classification confidence from the constructed bags. However, they pay less [12] or no [11] attention to mining spatial information, leading to the following problems as shown in Fig. 1b: (1) Object Drift: For each object, some proposals in the constructed bag do not have a high IoU with the original object but with another nearby object. These proposals are not spatially adjacent to the original object but still have high classification confidence, as the rightmost (the yellow box) proposal of " + }, + { + "bbox": [ + 47, + 123, + 289, + 422 + ], + "type": "inline_equation", + "content": "O_{2}" + }, + { + "bbox": [ + 47, + 123, + 289, + 422 + ], + "type": "text", + "content": " shows in the left-bottom corner of Fig. 1b. Only the category confidence is relied on for selecting proposals for " + }, + { + "bbox": [ + 47, + 123, + 289, + 422 + ], + "type": "inline_equation", + "content": "O_{2}" + }, + { + "bbox": [ + 47, + 123, + 289, + 422 + ], + "type": "text", + "content": ", and the rightmost proposal will be selected as the refined box. It means the refined box drifts to another object (Fig. 1b (i)), reducing the recall; (2) Group Prediction: Most works [7, 11, 56] select the top- " + }, + { + "bbox": [ + 47, + 123, + 289, + 422 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 123, + 289, + 422 + ], + "type": "text", + "content": " proposals by classification confidence and then weight average them as the refined box, causing the group prediction problem, as shown in Fig. 1b (ii); (3) Part Domination: The detector often focuses on the object's semantic region, which can statistically represent the category (e.g. the face). As shown in Fig. 1b (iii), the high classification confidence of the animal is in the discriminant part (the face) rather than the entire object as mentioned by [44, 39]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 289, + 715 + ], + "type": "text", + "content": "To address these problems, we propose a Spatial Self-Distillation based detector (SSD-Det) to integrate the spatial cues into the bounding box refinement. SSD-Det has two important components: the Spatial Position Self-Distillation (SPSD) module for the bag construction step and the Spatial Identity Self-Distillation (SISD) module for the proposal selection step. To construct high-quality proposal bags, SPSD utilizes a neighborhood sampler to generate a balanced and flexible initial proposal bag for each object and then trains a regressor with the supervision of the annotated inaccurate bounding boxes. Finally, high-quality proposal bags are constructed with proposals corrected by the regressor. The mechanism behind SPSD is that the network learns the spatial information from the reliable samples, e.g. those low-noise annotations, in the dataset and then guides the noisy samples to produce high-quality proposals, as shown in Fig. 2. In addition, to further combine the category information and the spatial information, an interactive structure is implemented by alternately using SPSD to mine spatial cues and MIL to utilize the category information. With SPSD and the interactive structure, a high-quality proposal bag can be constructed. Experiments on MS-COCO show that SPSD can significantly improve the mean/max IoU between objects and proposals" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 150 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 150 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 150 + ], + "type": "image", + "image_path": "842a39994e87835a22dd9a60227450ab9bb3ff2e3047497d078fd57f2beee266.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 160, + 547, + 209 + ], + "lines": [ + { + "bbox": [ + 304, + 160, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 160, + 547, + 209 + ], + "type": "text", + "content": "Figure 2: The mechanism of Spatial Self-Distillation. By assigning higher weight, the low-noise annotations can be seen as reliable samples to guide the training of proposals' spatial position and identity learning in SPsD and SISD." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 231, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 396 + ], + "type": "text", + "content": "(about 18/10 points, Fig. 5) in the constructed bag. Instead of selecting proposals by classification confidence, we have proposed the SISD module in the proposal selection step. We use it to obtain each proposal's spatial confidence by predicting the IoU with the object and combining the IoU with classification confidence to select the top-" + }, + { + "bbox": [ + 304, + 231, + 545, + 396 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 231, + 545, + 396 + ], + "type": "text", + "content": " proposals. It is worth mentioning that SISD is an object-related IoU predictor, which means that the predicted IoU may be different for the same proposal that appears in different objects' bags. Accordingly, it guarantees that SISD can better handle object drift and group prediction problems. Experiments on MS-COCO and VOC datasets verify the effectiveness of our method and bring state-of-the-art performance. The contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 399, + 545, + 578 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 304, + 399, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 445 + ], + "type": "text", + "content": "1) We further investigate the inaccurate-box supervised object detection tasks and propose an end-to-end training SSD-Det that combines the spatial and the category information in an interactive fashion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 447, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 482 + ], + "type": "text", + "content": "2) We utilize an SPSD module to generate higher-quality proposals sampling through statistic-guide spatial position distillation, raising the upper bound of the refinement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 483, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 518 + ], + "type": "text", + "content": "3) To add spatial cues to classification confidence, we also introduce an SISD module to select a proposal belonging to the object rather than the category." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 519, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 578 + ], + "type": "text", + "content": "4) The performance of our proposed SSD-Det improves the mean average precision (AP) of the best previous method (e.g. over 10 AP on " + }, + { + "bbox": [ + 304, + 519, + 545, + 578 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 519, + 545, + 578 + ], + "type": "text", + "content": " noisy MS-COCO) and achieves state-of-the-art under various noise rate box supervision on MS-COCO and VOC datasets." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 590, + 392, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 392, + 604 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 392, + 604 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 610, + 408, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 408, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 408, + 624 + ], + "type": "text", + "content": "2.1. Object Detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Classic object detection [15, 38, 37, 32, 30, 3, 43, 57] is supervised by an accurate bounding-box. One-stage detectors utilize anchors as the sliding-window, such asYOLO [37], SSD [32], and RetinaNet [30]. Two-stage detectors mine spatial information to predict proposals (e.g. selective search [47] in Fast R-CNN [15] or RPN in Faster R-CNN [38]) and conduct classification and bounding-box" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6856" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "regression with filtered proposals sparsely. Transformer-based (i.e. DETR [3], Deformable-DETR [62], and SwinTransformer [33]) detectors utilize global information for better representation. Sparse R-CNN [43] combines a transformer's advantages and CNNs for detection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 145, + 286, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 145, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 145, + 286, + 159 + ], + "type": "text", + "content": "2.2. Weakly-Supervised Object Detection (WSOD)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 165, + 287, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 165, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 287, + 405 + ], + "type": "text", + "content": "WSOD trains object detectors with image tag supervision. Only with the category annotation, the majority of previous methods treat each image as a bag and candidate proposals as instances. They follow the multiple instance learning (MIL) pipeline [1, 44, 46, 7, 48], which highly depends on category information. However, the MIL loss function leads to a non-convex optimization problem; thus, MIL solutions are usually stuck into the local minima. Context information [24, 51], spatial regularization [1, 9, 48], and optimization strategy [44, 48, 46] are proposed to address the problems. SPE [28] introduces Transformer into WSOD and uses attention to generate proposals. SD-LocNet [59] tackles the initialized noisy object locations in WSOD and proposes a self-directed localization network to identify the noisy object instances. [44, 46, 7] use the pseudo label for classification's iterative refinement. However, we use the pseudo box as a better self-distillation teacher. [54, 39] conduct regression to move the proposals, whereas we conduct regression to distill for better bag construction. In this work, we also formulate box correction as a MIL problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 418, + 268, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 268, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 268, + 430 + ], + "type": "text", + "content": "2.3. Semi-Supervised Object Detection (SSOD)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "text", + "content": "Semi-supervised learning in object detection can be roughly categorized into two groups: consistency based [22, 45] and pseudo label based [27, 36, 41, 49, 64, 52]. [52] presents an end-to-end SSOD approach with two simple techniques named soft teacher and box jittering to facilitate the efficient leverage of the teacher model. Both Soft Teacher [52] and SSD-Det obtain pseudo-labels from candidate boxes, adopting distillation, box jitter and classification scores weighting policy. However, they are different: 1) SSOD selects candidate boxes from the teacher model's detection results, while SSD-Det generates them with a generative approach, due to no any accurate supervised data for ensuring high-quality boxes in the detection results. 2) [52] is based on FixMatch [40] and requires maintaining two networks for teacher-student distillation structure. In contrast, our approach only needs a multi-head detector with a shared backbone for self-distillation. 3) Box Jitter: [52] calculates box variance from jittering for result selection, while we aim to generate candidate boxes that combine SISD-predicted IoU and classification scores, selecting boxes closer to the ground truth. 4) We use classification scores as weights for the next stage's loss, while [52] employs them as a criterion for selecting reliable samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 484, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 484, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 484, + 85 + ], + "type": "text", + "content": "2.4. Learning with Noisy Annotations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 90, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 342 + ], + "type": "text", + "content": "Training CNNs under noisy labels has been an active research area. Previous research focuses on the classification task, and develops various techniques to deal with noisy labels, such as sample selection [17, 23] for training, label correction [42, 34], and robust loss functions [60, 14] against noisy labels. Recently, many efforts [4, 25, 35, 53, 11, 12] have been devoted to the object detection task. On the one hand, Simon et al. [4] first investigates the impact of different types of label noise on object detection. They propose a per-object co-teaching method to alleviate the effect of noisy labels. On the other hand, [53] proposes a meta-learning framework for noisy annotations consisting of noisy category labels and bounding boxes. [11, 12] utilize object-level MIL to refine the inaccurate box. OAMIL [12] constructs proposal bags through label assignment in a discriminant style. P2BNet [11] originally conducts point-supervised object detection tasks. However, it can be seen as the box correction in its refinement stage. It uses hand-craft anchors to generate proposal bags. Our method inherits the generative style of P2BNet and conducts spatial distillation to mine spatial information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 347, + 435, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 435, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 435, + 360 + ], + "type": "text", + "content": "2.5. Knowledge Distillation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 365, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 485 + ], + "type": "text", + "content": "Knowledge distillation (KD) [21] aims to learn compact and efficient student models guided by excellent teacher networks. It is first applied to object detection in [5], in which hint learning and KD are both used for multi-class object detection. Recently, many efforts [26, 50, 8, 16, 55] aim to mimic the feature. [61] shows that localization knowledge is more important and proposed a localization distillation method. We also transfer the spatial knowledge from reliable labeled instances to correct inaccurate bounding boxes (shown in Fig. 2) in a self-distillation manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 494, + 388, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 494, + 388, + 507 + ], + "spans": [ + { + "bbox": [ + 306, + 494, + 388, + 507 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 514, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 657 + ], + "type": "text", + "content": "This work aims to learn a robust detector with inaccurate bounding boxes. Instead of training a detector with the original inaccurate bounding box, we follow most related works [11, 12] that design a branch to refine the inaccurate bounding box and then train the detector head or detector with the refined bounding box. The most important part is how to design a refining policy. We first design a two-stage basic box refiner (gray region in Fig. 3) as a naive solution that modified from [11]. Then, SPSD and SISD are proposed and added to further mine the spatial cues for box refinement, yielding SSD-Det. Therefore, the overall loss function is formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 662, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 662, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 317, + 662, + 545, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {B a s i c}} + \\alpha_ {1} \\cdot \\mathcal {L} _ {\\text {S P S D}} + \\alpha_ {2} \\cdot \\mathcal {L} _ {\\text {S I S D}} + \\alpha_ {3} \\cdot \\mathcal {L} _ {\\text {D e t}} \\tag {1}", + "image_path": "68578ad4100a95f9c95e076fb54213600b5c626208f08403d4c7392f8eb34ad7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{1},\\alpha_{2}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{3}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " are set as 0.25, 0.25 and 4 respectively. " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Det}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " denotes the loss of detector or detection head. During inference, only the detector or the detection head is used." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6857" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 68, + 547, + 216 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 547, + 216 + ], + "type": "image", + "image_path": "2b0384aa55024bf0fba338b9a639ec4fbe5410b686fbc47463ede4a020def4db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 547, + 298 + ], + "type": "text", + "content": "Figure 3: The framework of SSD-Det. It contains basic box refiner, SPSD module, SISD module and a detector head. Neighborhood sampler is adopted around the inaccurate annotation. Then, SPSD module generates better proposal bags which are fed into basic box refiner for MIL training. The selected proposals are average weighted as the refined box and supervise the next SPSD training. Meanwhile, the SISD module predicts the IoU between proposals and the object, and the estimated IoU is multiplied by classification score for better proposal selection to generate the refined box. SPSD shares backbone with the detector." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 318, + 153, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 153, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 153, + 330 + ], + "type": "text", + "content": "3.1. Basic Box Refiner" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 288, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 288, + 434 + ], + "type": "text", + "content": "Motivated by [12] and WSOD [1], we design the basic box refiner (detailed structure figure is in supplementary) that leverages classification confidence to refine the inaccurate box annotation. Then the refine annotation is used to train a detection head or detector. Following [11], we design a two-stream structure as a MIL classifier to select the best proposal for box refinement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": "Giving an image with inaccurate box annotation, for each object, " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": " is a bag of proposals (bounding boxes) that are generated around its inaccurate annotation by a sampler policy (e.g., selective search[47], edge box[63], neighborhood sampler in Sec. 3.2). Meanwhile a feature map is extracted with a backbone network. And then through " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "7\\times 7" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": " RoIAlign [18] and two fully connected (fc) layers, features of proposal in " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": " are extracted and denote as " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": ". The basic box refiner takes proposal bag " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\in \\mathbb{R}^{P \\times 4}" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{P \\times D}" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": " as inputs, where " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": " are denoted as the number of proposals in " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": ", feature dimension respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Following [11] and [1], as Eq. 2 described, we apply the classification branch " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{cls}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " yields " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{cls}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ", which is then passed through the softmax function over classification dimension " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " to obtain the score " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{cls} \\in \\mathbb{R}^{P \\times K}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " represents the number of instance categories. Likewise, instance selection branch " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{cls}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " to yield " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{ins}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ", and instance score " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{ins}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " is obtained through softmax function over " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " proposals. The proposal score " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " is obtained by computing the Hadamard product of the classification score and the instance score. The bag score " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{S}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " is obtained by the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 319, + 522, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 522, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 522, + 331 + ], + "type": "text", + "content": "summating of the " + }, + { + "bbox": [ + 306, + 319, + 522, + 331 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 306, + 319, + 522, + 331 + ], + "type": "text", + "content": " proposal boxes' proposal scores." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 309, + 332, + 533, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 332, + 533, + 360 + ], + "spans": [ + { + "bbox": [ + 309, + 332, + 533, + 360 + ], + "type": "interline_equation", + "content": "\\mathbf {O} ^ {c l s} = f _ {c l s} (\\mathbf {F}) \\in \\mathbb {R} ^ {P \\times K}; [ \\mathbf {S} ^ {c l s} ] _ {p k} = e ^ {[ \\mathbf {O} ^ {c l s} ] _ {p k}} / \\sum_ {k = 1} ^ {K} e ^ {[ \\mathbf {O} ^ {c l s} ] _ {p k}}.", + "image_path": "420ccff4b7a2824300d48a10be7b568aec54bd0c309582ed90352a10ce448ea9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 361, + 541, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 361, + 541, + 390 + ], + "spans": [ + { + "bbox": [ + 309, + 361, + 541, + 390 + ], + "type": "interline_equation", + "content": "\\mathbf {O} ^ {i n s} = f _ {i n s} (\\mathbf {F}) \\in \\mathbb {R} ^ {P \\times K}; [ \\mathbf {S} ^ {i n s} ] _ {p k} = e ^ {[ \\mathbf {O} ^ {i n s} ] _ {p k}} / \\sum_ {p = 1} ^ {P} e ^ {[ \\mathbf {O} ^ {i n s} ] _ {p k}}.", + "image_path": "9e0fcb0aa398900a98676518a6d8247205043cafac49bb8ebb3055b9419f1b55.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 309, + 393, + 545, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 393, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 309, + 393, + 545, + 429 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\mathbf {S} ^ {c l s} \\odot \\mathbf {S} ^ {i n s} \\in \\mathbb {R} ^ {P \\times K}; \\widehat {\\mathbf {S}} = \\sum_ {p = 1} ^ {P} [ \\mathbf {S} ] _ {p} \\in \\mathbb {R} ^ {K}. \\tag {2}", + "image_path": "d9f5db6092ddfafb46571bd79ebc28326e1295800495755b479901feae5a417a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 430, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 545, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 430, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\left[\\cdot \\right]_{pk}" + }, + { + "bbox": [ + 306, + 430, + 545, + 441 + ], + "type": "text", + "content": " is the value at row " + }, + { + "bbox": [ + 306, + 430, + 545, + 441 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 306, + 430, + 545, + 441 + ], + "type": "text", + "content": " and column " + }, + { + "bbox": [ + 306, + 430, + 545, + 441 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 430, + 545, + 441 + ], + "type": "text", + "content": " in the matrix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 442, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 545, + 478 + ], + "type": "text", + "content": "The basic box refiner has two similar stages. The loss of stage I(termed " + }, + { + "bbox": [ + 304, + 442, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_I" + }, + { + "bbox": [ + 304, + 442, + 545, + 478 + ], + "type": "text", + "content": ") adopt the MIL paradigm with the form of cross-entropy (CE) loss, defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 479, + 545, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 545, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I} = C E (\\widehat {\\mathbf {S}}, \\mathbf {c}) = - \\sum_ {k = 1} ^ {K} \\mathbf {c} _ {k} \\log (\\widehat {\\mathbf {S}} _ {k}) + (1 - \\mathbf {c} _ {k}) \\log (1 - \\widehat {\\mathbf {S}} _ {k}) \\tag {3}", + "image_path": "0b9219bedbdea8ab6c1140c95d6198a34986b85f0864bf336fc5d3bf4e014eab.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\{0,1\\}^K" + }, + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "type": "text", + "content": " is the one-hot category label. And each object's proposals with the top-" + }, + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "type": "text", + "content": " highest proposal score " + }, + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 304, + 516, + 545, + 552 + ], + "type": "text", + "content": " are weighted to obtain the refined box." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "type": "text", + "content": "The stage II takes the refined box of stage I as input and performs fine refining with a similar structure as stage I. Differently, the focal loss is adopted in stage II instead of cross entropy loss. In order to cooperate with focal loss, the classification branch uses the sigmoid " + }, + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\sigma(x)" + }, + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "type": "text", + "content": " instead of softmax function and we sample some negative samples " + }, + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "type": "text", + "content": " to further suppress the background. With the bag score " + }, + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\widehat{S}" + }, + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "type": "text", + "content": " and the negative sample scores " + }, + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\text{neg}}^{\\text{cls}}" + }, + { + "bbox": [ + 304, + 552, + 545, + 649 + ], + "type": "text", + "content": ", the loss is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 653, + 545, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 653, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 311, + 653, + 545, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I I} = \\left\\langle \\mathbf {c} ^ {\\mathrm {T}}, \\widehat {\\mathbf {S}} ^ {*} \\right\\rangle \\cdot \\operatorname {F L} \\left(\\widehat {\\mathbf {S}}, \\mathbf {c}\\right) + \\sum_ {\\mathcal {N}} \\beta \\cdot \\operatorname {F L} \\left(\\mathbf {S} _ {n e g} ^ {c l s}, c _ {n e g}\\right) \\} \\tag {4}", + "image_path": "a7203f9f8632ca3b4f413186a1e3eb99f030075f88a2345c33b8234266239856.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 682, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 682, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 682, + 545, + 715 + ], + "type": "text", + "content": "where FL is the focal loss [30], " + }, + { + "bbox": [ + 304, + 682, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{S}}_j^*" + }, + { + "bbox": [ + 304, + 682, + 545, + 715 + ], + "type": "text", + "content": " represents the bag score predicted by stage I. " + }, + { + "bbox": [ + 304, + 682, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\left\\langle \\mathbf{c}_j^{\\mathrm{T}},\\widehat{\\mathbf{S}}_j^*\\right\\rangle" + }, + { + "bbox": [ + 304, + 682, + 545, + 715 + ], + "type": "text", + "content": " represents the inner product" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6858" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 125 + ], + "type": "text", + "content": "of the two vectors, meaning the predicted bag score of the ground-truth category. " + }, + { + "bbox": [ + 46, + 72, + 287, + 125 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 72, + 287, + 125 + ], + "type": "text", + "content": " is the average of " + }, + { + "bbox": [ + 46, + 72, + 287, + 125 + ], + "type": "inline_equation", + "content": "\\left\\langle \\mathbf{c}_j^{\\mathrm{T}},\\widehat{\\mathbf{S}}_j^*\\right\\rangle" + }, + { + "bbox": [ + 46, + 72, + 287, + 125 + ], + "type": "text", + "content": ". They are used to weight each object's FL for stable training. The overall loss function of the basic refiner here is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 134, + 287, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 134, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 117, + 134, + 287, + 146 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {B a s i c}} = \\mathcal {L} _ {I} + \\alpha_ {I I} \\cdot \\mathcal {L} _ {I I} \\tag {5}", + "image_path": "6b52f74a5603fe510a51370b7a45d75ea49453fa18ab46056c0d85ab296ca4ce.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 156, + 244, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 244, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 244, + 167 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 156, + 244, + 167 + ], + "type": "inline_equation", + "content": "\\alpha_{II}" + }, + { + "bbox": [ + 47, + 156, + 244, + 167 + ], + "type": "text", + "content": " are the loss weights of the two stages." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 167, + 287, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 287, + 227 + ], + "type": "text", + "content": "During training, the refined box of stage II is used as supervision for a detection head or detector. After training, the basic box refiner will be removed, leaving a well-trained detection head or detector. In this way, we can train a detector under inaccurate annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 236, + 255, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 255, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 255, + 248 + ], + "type": "text", + "content": "3.2. Spatial Position Self-Distillation (SPSD)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 255, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 326 + ], + "type": "text", + "content": "Like most MIL paradigm methods, basic box refiner has two main components: bag construction and proposal selection. And the main idea is to use classification information to guide the refining. In this paper, we add spatial information to improve refining. Specifically, SPSD is proposed to use spatial information to enhance bag construction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 434 + ], + "type": "text", + "content": "Bag construction aims to obtain proposals for each object, while proposal selection is to select the proposals from the object bag. Then, the refined box is averaged over the selected proposals. Therefore, the quality of the proposals in constructed bag determines the upbound of refining. The bag construction can be implemented in a variety of ways. In this paper, the basic box refiner adopts a naive neighborhood sampler for bag construction. basic box refiner adopts a naive neighborhood sampler for bag construction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "text", + "content": "Neighborhood Sampler. Proposals around the inaccurate box are sampled to construct an object bag. For each inaccurate box " + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "inline_equation", + "content": "b^{*} = (b_{x}^{*}, b_{y}^{*}, b_{w}^{*}, b_{h}^{*})" + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "text", + "content": ", its scale and aspect ratio with " + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "text", + "content": " are adjusted and its positions " + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "inline_equation", + "content": "o_{x}, o_{y}" + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "text", + "content": " are jittered to obtain the diverse proposal " + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "inline_equation", + "content": "b = (b_{x}, b_{y}, b_{w}, b_{h})" + }, + { + "bbox": [ + 46, + 435, + 287, + 495 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 497, + 287, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 497, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 93, + 497, + 287, + 522 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} b _ {w} = v \\cdot s \\cdot b _ {w} ^ {*}, \\quad b _ {h} = 1 / v \\cdot s \\cdot b _ {h} ^ {*}, \\\\ b _ {x} = b _ {x} ^ {*} + b _ {w} \\cdot o _ {x}, \\quad b _ {y} = b _ {y} ^ {*} + b _ {h} \\cdot o _ {y}. \\tag {6} \\\\ \\end{array}", + "image_path": "2cf2c1ae9ac861675dc4963afd911d42cccc311dbc287db2329e49643cc248ad.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 522, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 652 + ], + "type": "text", + "content": "These proposals " + }, + { + "bbox": [ + 46, + 522, + 287, + 652 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 522, + 287, + 652 + ], + "type": "text", + "content": " are used to construct the positive proposal bag " + }, + { + "bbox": [ + 46, + 522, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 522, + 287, + 652 + ], + "type": "text", + "content": " to train the MIL classifier. Thanks to the hand-craft sampling way, the number of proposals in different objects' proposal bags is controllable and balanced. However, the hand-craft neighborhood sampler strategy is difficult to set hyper-parameters, and the sampling space is discrete. For example, when the jitter region is small, the optimization space of refining is limited, while when it is large, more background will be introduced. Hence, we propose the SPSD module to mine spatial information for higher-quality proposal bag construction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Statistically Guided Adaptive Sampling. Instead of simply using a neighborhood sampler, we adopt a statistically guided adaptive sampling by adding SPSD modules into the basic box refiner. Taking the constructed proposal bag " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " of hand-craft neighborhood sampler as input, the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": "RoI features of proposals in " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": " are extracted and fed into the two shared fc layers to obtain " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": ". Then a regression fc layer " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "f_{dis}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": ", supervised by the inaccurate annotated bounding box " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "b^{*}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": ", is introduced to predict the adaptive proposal bag " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{dis} = f_{dis}(\\mathbf{F}) \\in \\mathbb{R}^{P \\times 4}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": ", in which the proposals are closer to the object. Later, " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{dis}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": " as the constructed proposal bag is fed into stage I of basic box refiner. In order to combine category and spatial information, we implement an interactive structure by alternately using SPSD to mine the spatial cues and using MIL in basic box refiner to utilize the category information. Specifically, the refined bounding box " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\hat{b}^{*}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": " of stage I that selected by the classification confidence is used to supervision of a new SPSD module for stage II. Similar as stage I, The new SPSD takes proposal bag " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{dis}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": " of handcraft neighborhood sampler as input. Through the RoI align and the two shared fc layers, the feature " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": " is extracted. An extra fc layer " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "f_{dis}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": " is then utilized to conduct further regression. Different with stage I, the obtained " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{dis}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": " is supervised by the refined " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\hat{b}^{*}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": ". The loss function of the spatial distillation for adaptive sampling can be defined as " + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SPSD}" + }, + { + "bbox": [ + 304, + 72, + 545, + 313 + ], + "type": "text", + "content": " in Eq. 7." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 318, + 545, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 318, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 310, + 318, + 545, + 348 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S P S D} = \\frac {1}{P} \\left\\{\\sum_ {p = 1} ^ {P} \\mathbf {L} _ {\\mathbf {1}} \\left(\\left[ \\mathcal {B} ^ {d i s} \\right] _ {p}, b ^ {*}\\right) + \\sum_ {p = 1} ^ {P} \\mathbf {L} _ {\\mathbf {1}} \\left(\\left[ \\mathcal {B} ^ {\\hat {d i s}} \\right] _ {p}, b ^ {*}\\right) \\right\\} \\tag {7}", + "image_path": "42ff48275ea5bded514e8b1dd721aa4104ea8c6302b41ae818dca44b7748a2d6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 352, + 539, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 352, + 539, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 539, + 363 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 305, + 352, + 539, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1" + }, + { + "bbox": [ + 305, + 352, + 539, + 363 + ], + "type": "text", + "content": " is the L1 loss function for loose restrictions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 364, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 472 + ], + "type": "text", + "content": "The idea behind SPSD is that the dataset with inaccurate annotation still has many reliable, high-quality boxes and inaccurate boxes. Supervised by the high-quality boxes statistically, the network can guide those proposals sampled around the inaccurate bounding box to regress to the ground truth. With the self-distillation mechanism, SPSD learns the semantic-spatial correspondence knowledge from the reliable samples in the dataset and then propagates the knowledge to produce high-quality proposals." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": "Adaptive Negative Sampling. Negative samples are introduced in Stage II to better suppress the background. With the sampled " + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{dis}" + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": ", we can adaptively sample the negative samples with a small IoU (set smaller than 0.3 by default) with all positive proposals in all bags, to compose the negative sample set " + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": " for stage II." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 552, + 511, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 511, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 511, + 564 + ], + "type": "text", + "content": "3.3. Spatial Identity Self-Distillation (SISD)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "The basic box refiner selects the proposals only depending on classification confidence during proposal selection. To select the proposal which has high classification confidence and is also spatially close to the object from the bag, we propose a SISD module to predict the IoU between proposals and their corresponding object. Afterwards, through the combination between the IoU and the classification confidence, top-" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " proposals are selected. In SISD, we design an Object Relevance Enhancement (ORE) module to distinguish different objects' features with the same RoI region. And an identity predictor is designed to predict each proposal's IoU with the object." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6859" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 324 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 324 + ], + "type": "table", + "html": "
MethodBackbone20% Box Noise Level40% Box Noise Level
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)
Val Set
Clean-FasterRCNN [38]ResNet-5037.958.140.921.641.648.737.958.140.921.641.648.7
Clean-FasterRCNN [38]ResNet-10139.460.143.122.443.751.139.460.143.122.443.751.1
Clean-Retinanet [30]ResNet-5036.756.139.021.640.447.436.756.139.021.640.447.4
Noisy-FasterRCNN [38]ResNet-5030.454.331.417.433.938.710.328.93.35.711.815.1
Noisy-Retinanet [30]ResNet-5030.053.130.817.933.738.213.333.65.78.415.918.0
FreeAnchor[58]ResNet-5028.653.128.516.632.237.010.428.93.35.812.114.9
Co-teaching[17]ResNet-5030.554.930.517.334.039.111.531.44.26.413.116.4
SD-LocNet[59]ResNet-5030.054.530.317.533.638.711.330.34.36.012.716.6
KL loss[20]ResNet-5031.054.332.418.034.939.512.136.73.76.213.017.4
OA-MIL[12]ResNet-5032.155.333.218.135.841.618.642.612.99.219.926.5
SSD-DetResNet-5033.657.335.319.537.243.327.653.926.016.031.034.9
SSD-DetResNet-10134.357.636.719.138.144.328.454.327.216.531.936.4
SSD-Det+FRResNet-5034.457.336.820.038.244.029.354.829.017.132.936.9
SSD-Det+FRResNet-10136.259.139.220.940.247.130.656.730.718.134.539.0
Test Set
Clean-FasterRCNN [38]ResNet-5037.758.740.821.740.646.737.758.740.821.740.646.7
Noisy-FasterRCNN [38]ResNet-5030.754.931.318.033.737.710.429.03.36.011.314.6
OA-MIL[12]ResNet-5032.355.833.718.535.040.218.542.312.89.319.125.1
SSD-DetResNet-5033.557.335.519.136.041.928.054.126.516.530.034.5
SSD-Det+FRResNet-5034.757.937.220.037.742.729.755.629.317.532.436.2
", + "image_path": "0467d2da66ec35147ce7c8f791b2154e76169d949b8543bd121654d0b11273f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "text", + "content": "Object Relevance Enhancement (ORE). ORE enhances object-relevant features, making SISD an object-relevant IoU predictor. ORE allows the predicted IoU to be different for the same proposal in other objects' bags. In addition, we integrate the feature of the bag's corresponding object into the proposal feature, making the feature of different bags' proposals distinct. That is the so-called ORE. For a proposal bag " + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "text", + "content": ", the feature " + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "text", + "content": " is obtained through the RoI align and two fc layers. It is worth mentioning that the two fcs do not share the parameters with those in the refiner since the optimization goals are contradictory. To represent the feature of the bag's corresponding object, " + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{+} \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "text", + "content": " is calculated by averaging features of " + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "text", + "content": " proposals in proposal bag " + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "text", + "content": ". The object feature " + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{+}" + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "text", + "content": "is broadcast into " + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{P \\times D}" + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "text", + "content": ", and then added to the proposal features to obtain the object-relevant features " + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{*} = \\mathbf{F} + \\mathbf{F}^{+}" + }, + { + "bbox": [ + 46, + 377, + 287, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "text", + "content": "Spatial Identity Prediction. By a following identity fc layer, " + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{P \\times 1}" + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "text", + "content": " is predicted. The pseudo label " + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "inline_equation", + "content": "T \\in (0,1)" + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "text", + "content": " is IoU between proposals in " + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "text", + "content": " and the merged box " + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\hat{b}^*" + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "text", + "content": " of stage I. For better optimization, the linear normalized " + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "inline_equation", + "content": "T' = (T - 0.5)/0.5 \\in (-1,1)" + }, + { + "bbox": [ + 47, + 568, + 287, + 643 + ], + "type": "text", + "content": " is utilized as supervision. The object function of the identity predictor is identified in:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 647, + 287, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 647, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 113, + 647, + 287, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S I S D} = \\operatorname {s m o o t h} _ {L 1} \\left(U, T ^ {\\prime}\\right) \\tag {8}", + "image_path": "aabfcd92eb69bf91ac1bd945c682d33bd64a7fb3b68eee6f95e4bb2eb5663550.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "where the smooth " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " represents the smooth L1 loss. The predicted spatial confidence " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "U^{\\prime}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " is obtained by normalizing the " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "S^{*} = U^{\\prime} \\cdot S" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " is used to select the top-" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " proposals for merging as the refined boxes." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 375, + 544, + 496 + ], + "blocks": [ + { + "bbox": [ + 46, + 331, + 546, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 546, + 357 + ], + "type": "text", + "content": "Table 1: Performance comparison on COCO. FR is Faster R-CNN. *-FR refers to a retrained Faster R-CNN (R50+FPN) using refined annotations from SSD-Det for improved performance.. Clean-* and Noisy-* means original and noisy annotation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 375, + 544, + 496 + ], + "lines": [ + { + "bbox": [ + 310, + 375, + 544, + 496 + ], + "spans": [ + { + "bbox": [ + 310, + 375, + 544, + 496 + ], + "type": "table", + "html": "
MethodBackboneBox Noise Level
10%20%30%40%
Clean-FasterRCNN [38]ResNet-5077.2 for clean
Clean-RetinaNet [30]ResNet-5073.5 for clean
Noisy-FasterRCNN [38]ResNet-5076.371.260.142.5
Noisy-RetinaNet [30]ResNet-5071.567.657.945.0
KL loss[20]ResNet-5075.872.764.648.6
Co-teaching[17]ResNet-5075.470.660.943.7
SD-LocNet[59]ResNet-5075.771.560.843.9
FreeAnchor[58]ResNet-5073.067.556.241.6
OA-MIL[12]ResNet-5077.474.370.663.8
SSD-DetResNet-5077.174.871.566.9
", + "image_path": "6abbc608d5c12e1e03bd179d7de5d314febc497bc1cc7cb46909642f27fa35c6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 505, + 545, + 541 + ], + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 541 + ], + "type": "text", + "content": "Table 2: Performance comparison on the VOC 2007 test set. The evaluation metric is " + }, + { + "bbox": [ + 304, + 505, + 545, + 541 + ], + "type": "inline_equation", + "content": "AP_{50}" + }, + { + "bbox": [ + 304, + 505, + 545, + 541 + ], + "type": "text", + "content": ". The Clean-* and Noisy-* means original annotation and noisy annotation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 563, + 382, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 382, + 577 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 382, + 577 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 584, + 432, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 584, + 432, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 584, + 432, + 597 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 604, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 699 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. For experimental comparisons, two publicly available datasets are used for object detection with inaccurate bounding boxes: MS-COCO [31] and PASCAL VOC 2007 [13]. MS-COCO (2017 version) has 118k training and 5k validation images with 80 common object categories. PASCAL VOC 2007 is one of the most popular benchmarks in generic object detection with 20 classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Evaluation Metric. We use mean average precision" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6860" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 164 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 164 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 164 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 164 + ], + "type": "table", + "html": "
2-RefSPSDSISDRe-Train20% Box Noise Level40% Box Noise Level
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)\\(AP^{test}\\)\\(AP^{test}_{75}\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AP^{s}\\)\\(AP^{m}\\)\\(AP^{l}\\)\\(AP^{test}\\)\\(AP^{test}_{75}\\)
30.057.129.016.933.139.8--22.851.116.113.325.030.4--
✓31.256.731.617.834.541.031.432.024.652.020.114.328.231.925.020.5
✓✓33.056.934.818.735.542.233.134.827.253.724.715.930.335.227.625.6
✓✓✓33.657.335.319.537.243.333.535.527.653.926.016.031.034.928.026.5
✓✓31.856.833.118.435.740.832.333.726.554.023.315.730.333.826.823.3
✓✓✓34.157.636.419.037.743.834.336.629.055.127.817.032.536.729.328.4
✓✓✓✓34.457.336.820.038.244.034.737.229.354.829.017.132.936.929.729.3
", + "image_path": "53bc81990f8917d1a72fd51326044c106320f8dabf4201e29bd31b1ef660dbc1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 215, + 285, + 265 + ], + "blocks": [ + { + "bbox": [ + 47, + 171, + 546, + 196 + ], + "lines": [ + { + "bbox": [ + 47, + 171, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 546, + 196 + ], + "type": "text", + "content": "Table 3: Modules ablation of SPSD, SISD and Re-Train on MS-COCO validation set (without) and test set (with test). The Re-Train means we generate the pseudo label by SSD-Det and re-train a Faster R-CNN detector." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 215, + 285, + 265 + ], + "lines": [ + { + "bbox": [ + 52, + 215, + 285, + 265 + ], + "spans": [ + { + "bbox": [ + 52, + 215, + 285, + 265 + ], + "type": "table", + "html": "
Methods (w/o SISD)APAP50AP75APsAPmAPl
Neighborhood Sampler24.652.020.114.328.231.9
SPSD (II) w/o weighted26.053.322.515.629.433.4
SPSD (II) w/ weighted26.353.422.515.629.333.8
SPSD (I+II) w/ weighted27.253.724.715.930.335.2
", + "image_path": "8a5285f54cd48f3a17c3f347e76618bbd3cc37e66da65a02663bb47c172a6fc3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 300, + 282, + 369 + ], + "blocks": [ + { + "bbox": [ + 96, + 274, + 238, + 285 + ], + "lines": [ + { + "bbox": [ + 96, + 274, + 238, + 285 + ], + "spans": [ + { + "bbox": [ + 96, + 274, + 238, + 285 + ], + "type": "text", + "content": "Table 4: Different setting of SPSD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 300, + 282, + 369 + ], + "lines": [ + { + "bbox": [ + 52, + 300, + 282, + 369 + ], + "spans": [ + { + "bbox": [ + 52, + 300, + 282, + 369 + ], + "type": "table", + "html": "
ORE Strategies of SISDAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP^s \\)\\( AP^m \\)\\( AP^l \\)
w/o SISD27.253.724.715.930.335.2
SISD w/o ORE27.353.325.716.930.235.0
+ subtract27.253.824.615.730.435.0
+ concatenate27.254.024.516.130.234.8
+ add27.653.926.016.031.034.9
+ add w/ shared fcs23.049.917.612.625.330.4
", + "image_path": "fe09d1c374a53296449568c1295365f508cec2ad7b5f680698ee64e6c8499c51.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 403, + 286, + 430 + ], + "blocks": [ + { + "bbox": [ + 80, + 377, + 253, + 389 + ], + "lines": [ + { + "bbox": [ + 80, + 377, + 253, + 389 + ], + "spans": [ + { + "bbox": [ + 80, + 377, + 253, + 389 + ], + "type": "text", + "content": "Table 5: Different ORE strategies of SISD." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 403, + 286, + 430 + ], + "lines": [ + { + "bbox": [ + 50, + 403, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 50, + 403, + 286, + 430 + ], + "type": "table", + "html": "
Num.0123
AP/AP5024.6 / 52.026.3 / 53.427.2 / 53.727.0 / 53.1
", + "image_path": "1ab558621bc2291da62176a7fe49804e3afd6e1f3574544c60796aa3e6f1e017.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 95, + 448, + 237, + 459 + ], + "lines": [ + { + "bbox": [ + 95, + 448, + 237, + 459 + ], + "spans": [ + { + "bbox": [ + 95, + 448, + 237, + 459 + ], + "type": "text", + "content": "Table 6: Number of SPSD module." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 483, + 286, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 286, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 286, + 519 + ], + "type": "text", + "content": "mAP@[.5,.95] and (mAP@.5) for MS-COCO and VOC. The " + }, + { + "bbox": [ + 47, + 483, + 286, + 519 + ], + "type": "inline_equation", + "content": "\\{AP, AP_{50}, AP_{75}, AP^{Small}, AP^{Middle}, AP^{Large}\\}" + }, + { + "bbox": [ + 47, + 483, + 286, + 519 + ], + "type": "text", + "content": " is reported for MS-COCO and " + }, + { + "bbox": [ + 47, + 483, + 286, + 519 + ], + "type": "inline_equation", + "content": "AP_{50}" + }, + { + "bbox": [ + 47, + 483, + 286, + 519 + ], + "type": "text", + "content": " for VOC." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 521, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 286, + 581 + ], + "type": "text", + "content": "Synthetic Noisy Dataset. Following [12], We simulate noisy bounding boxes by perturbing clean boxes from the original annotations. The details are in the appendix. We simulate various box noise levels ranging from " + }, + { + "bbox": [ + 47, + 521, + 286, + 581 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 521, + 286, + 581 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 521, + 286, + 581 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 47, + 521, + 286, + 581 + ], + "type": "text", + "content": " for the VOC and " + }, + { + "bbox": [ + 47, + 521, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\{20\\%, 40\\}" + }, + { + "bbox": [ + 47, + 521, + 286, + 581 + ], + "type": "text", + "content": " for the MS-COCO." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 712 + ], + "type": "text", + "content": "Implementation Details. We implement our method on FasterRCNN [38] with ResNet50-FPN [19, 29] backbone, based on MMDetection [6]. All settings of our method and previous methods employ FPN for fair comparison. Similar to the default setting of object detection on MS-COCO, the stochastic gradient descent [2] algorithm is used to optimize on 1x training schedule. The batch size is two images per GPU on 8 GPUS. For the VOC dataset, the batch size is two images per GPU on 2 GPUS. The performance we report is on a single scale (1333 * 800 for MS-COCO and 1000 * 600 for VOC)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 217, + 487, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 487, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 487, + 229 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 243, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 316 + ], + "type": "text", + "content": "We compare our method with several state-of-the-art approaches [20, 17, 59, 58, 12] on MS-COCO and VOC 2007 datasets. We denote Clean-FasterRCNN and Noisy-FasterRCNN as FasterRCNN models trained under clean (original annotations) and noisy annotations with the default setting, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "text", + "content": "MS-COCO Dataset. Table 1 shows the comparison results on the MS-COCO. Inaccurate bounding box annotations significantly deteriorate the vanilla Faster R-CNN's detection performance. Co-teaching and SD-LocNet only slightly improve the detection performance, especially under " + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "text", + "content": " box noise. That indicates that small-loss sample selection and sample weight assignment can not tackle noisy box annotations well. KL Loss slightly improves the performance under " + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "text", + "content": " box noise. By treating an object as a bag of instances, OA-MIL is somehow robust to noisy bounding boxes and performs better than other methods. Nevertheless, the previously-mentioned label assignment bag construction limits its ability to handle heavy noise. Our approach is more robust to noisy bounding boxes. It outperforms other methods by a large margin under high box noise levels and significantly boosts the detection performance across all metrics. For example, under " + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "text", + "content": " box noise, the end-to-end SSD-Det achieves 27.6 " + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "inline_equation", + "content": "53.9 AP_{50}" + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "text", + "content": ", 9.0 and attains 11.3 point improvement compared with state-of-the-art method OA-MIL, respectively. Also, through re-training on FasterRCNN, the performance further reaches 29.3 " + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "inline_equation", + "content": "54.8 AP_{50}" + }, + { + "bbox": [ + 304, + 323, + 546, + 622 + ], + "type": "text", + "content": ". With the backbone of ResNet-101, the performance achieves consistent improvement. On MS-COCO test set, our method also achieves state-of-the-art performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "VOC 2007 Dataset. Table 2 shows the comparison results on the VOC 2007 test set. Co-teaching, SD-LocNet and KL Loss, can not address inaccurate bounding box annotations well. OA-MIL improves the performance on different noisy datasets. Our approach obtains further improvements to 77.10, 74.80, 71.50, 66.90 " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "AP_{50}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " noisy box datasets, respectively." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6861" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 225, + 225 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 225, + 225 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 225, + 225 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 225, + 225 + ], + "type": "image", + "image_path": "b82a78ea638bd787cba4f76d854879baa5c65554b6052c89d801317cd5335c24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 234, + 544, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 234, + 544, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 544, + 258 + ], + "type": "text", + "content": "Figure 4: Qualitative detection results on COCO validation. Previous methods miss objects and face part prediction problems. Our method misses fewer objects, and the bounding box quality is better, especially for small or overlapped objects." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 71, + 300, + 224 + ], + "blocks": [ + { + "bbox": [ + 227, + 71, + 300, + 224 + ], + "lines": [ + { + "bbox": [ + 227, + 71, + 300, + 224 + ], + "spans": [ + { + "bbox": [ + 227, + 71, + 300, + 224 + ], + "type": "image", + "image_path": "c42efc6fd3a318138ab97195e2f6e7c68a77bf843ba4470add0c21737a08ff89.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 301, + 72, + 376, + 224 + ], + "blocks": [ + { + "bbox": [ + 301, + 72, + 376, + 224 + ], + "lines": [ + { + "bbox": [ + 301, + 72, + 376, + 224 + ], + "spans": [ + { + "bbox": [ + 301, + 72, + 376, + 224 + ], + "type": "image", + "image_path": "e91598751ae71b1b39edb2b75603285d83032dec869312aed2bd3b38ebb1ac5b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 377, + 72, + 545, + 224 + ], + "blocks": [ + { + "bbox": [ + 377, + 72, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 377, + 72, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 377, + 72, + 545, + 224 + ], + "type": "image", + "image_path": "93cb516e89418507075d332120a024065beccae3bea40e96ec6861433fc1725b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 52, + 277, + 285, + 299 + ], + "blocks": [ + { + "bbox": [ + 52, + 277, + 285, + 299 + ], + "lines": [ + { + "bbox": [ + 52, + 277, + 285, + 299 + ], + "spans": [ + { + "bbox": [ + 52, + 277, + 285, + 299 + ], + "type": "table", + "html": "
MethodsBox Refiner+Re-TrainSSD-DetSSD-Det+Re-Train
AP/AP5029.0 / 54.427.6/53.929.3 / 54.8
", + "image_path": "a000cf5a9d388a0d3dabc608b3494d3678b7c0ef8e14a2205d3c65bd2b892e93.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 342, + 203, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 203, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 203, + 355 + ], + "type": "text", + "content": "4.3. Ablation Study and Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 362, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 410 + ], + "type": "text", + "content": "To further analyze SSD-Det's effectiveness and robustness, we conduct more experiments on COCO val set if there are no other instructions. Except for Table 3, the noise level of these experiments is " + }, + { + "bbox": [ + 46, + 362, + 287, + 410 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 362, + 287, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 411, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 531 + ], + "type": "text", + "content": "Ablation of Modules. Ablation study of each component in our approach is given in Table 3, including: (i) Different stages of our basic box refiner. i.e. training object detector without the stage II (2-Ref), where the pseudo boxes predicted by the stage I are served as the supervision for training a parallel detector. (ii) SPSD, i.e. training without SPSD, where the object-bag is constructed directly by neighborhood sampling around the noisy ground-truth or the predicted pseudo boxes of the stage I. (iii) SISD. (iv) Re-Train with FasterRCNN (Re-Train)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 532, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 700 + ], + "type": "text", + "content": "Effectiveness of SPSD. SPSD further improves the detection performance on the MS-COCO, especially under high box noise levels, e.g. under " + }, + { + "bbox": [ + 46, + 532, + 287, + 700 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 532, + 287, + 700 + ], + "type": "text", + "content": " box noise level, SPSD boosts the performance from 24.6 to 27.2, as shown in Table 3 (row 3). In Table 4, we conduct further ablation on SPSD. With SPSD bag construction only in stage II, the performance increases by 1.4 AP. The performance further improves with the proposal score of stage I as weights. With SPSD in all stages, the AP reaches 27.2. Fig. 5 shows the bag quality. With SPSD, the mean IoU increases from 40.3 to 58.7 and the max and top-10 IoU increase to 78.3 and 75.1, which indicates a better upper bound of proposal selection. More high-quality proposals bring better optimization and easier proposals selection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Number of SPSD. As shown in Table 6. When adding" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 277, + 545, + 337 + ], + "blocks": [ + { + "bbox": [ + 64, + 307, + 269, + 319 + ], + "lines": [ + { + "bbox": [ + 64, + 307, + 269, + 319 + ], + "spans": [ + { + "bbox": [ + 64, + 307, + 269, + 319 + ], + "type": "text", + "content": "Table 7: Comparisons of end-to-end and Re-Train." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 277, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 309, + 277, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 309, + 277, + 545, + 337 + ], + "type": "table", + "html": "
DetectorsClean-supervisedNoise-supervised
APAP50APAP50(w/ours)
APAP50
FasterRCNN37.958.110.328.929.354.8
SparseRCNN [43]†45.064.16.020.334.360.2
De-DETR [62]†46.866.35.016.935.260.9
", + "image_path": "57db3089771a1f6f47b394dc3abd9b06623b49db9091445d138ad8c19f27a83d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 354, + 546, + 391 + ], + "lines": [ + { + "bbox": [ + 305, + 354, + 546, + 391 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 546, + 391 + ], + "type": "text", + "content": "Table 8: Experiments on advanced detectors. De-DETR is Deformable DETR. † uses multi-scale data augment. ‘" + }, + { + "bbox": [ + 305, + 354, + 546, + 391 + ], + "type": "inline_equation", + "content": "w/" + }, + { + "bbox": [ + 305, + 354, + 546, + 391 + ], + "type": "text", + "content": "ours’ means using our method under noisy supervision." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 408, + 545, + 519 + ], + "blocks": [ + { + "bbox": [ + 307, + 408, + 545, + 519 + ], + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 519 + ], + "type": "image", + "image_path": "3095a5b20676f90b4c2cb6035846f78ae76c3199ff9b55c994e70cb8855d99b9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 529, + 545, + 590 + ], + "lines": [ + { + "bbox": [ + 305, + 529, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 305, + 529, + 545, + 590 + ], + "type": "text", + "content": "Figure 5: Bag quality (IoU of proposals with GT) of construction in SSD-Det. B.S. (blue) means neighborhood sampler. SPSD I (orange) denotes single SPSD adopted. SPSD II (yellow) is two SPSD and interactive structure adopted. SPSD II significantly improves the quality." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 615, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 615, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 305, + 615, + 545, + 652 + ], + "type": "text", + "content": "3 SPSD, performance drops slightly, probably due to the accumulation of errors outweighing the performance gain from extra stages. Hence, 2 SPSD is our default setting." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness of SISD. SISD is designed to select object-aware proposals in box selection. Under " + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": " box noise, the detection performance improves from 27.2 to 27.6 and 33.0 to 33.6, which verifies the effectiveness of the module, shown in Table 3. We also study the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6862" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 286, + 112 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 112 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 112 + ], + "type": "table", + "html": "
MethodsDrift Rate (%)Group Rate (%)Part Rate (%)
allsmlallsmlallsml
OA-MIL[12]15.117.817.17.46.72.83.41.42.83.42.72.3
Ours1.51.01.31.41.71.20.50.71.00.51.11.3
", + "image_path": "ed4b84a6904b6b70f13626687796e53830309a57b6a6369a502ac0a870ee1102.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 180, + 285, + 262 + ], + "blocks": [ + { + "bbox": [ + 47, + 130, + 286, + 166 + ], + "lines": [ + { + "bbox": [ + 47, + 130, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 130, + 286, + 166 + ], + "type": "text", + "content": "Table 9: Breakdown of different problems during refinement (COCO under " + }, + { + "bbox": [ + 47, + 130, + 286, + 166 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 47, + 130, + 286, + 166 + ], + "type": "text", + "content": " noise level). s, m and 1 mean small, middle and large scale." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 180, + 285, + 262 + ], + "lines": [ + { + "bbox": [ + 50, + 180, + 285, + 262 + ], + "spans": [ + { + "bbox": [ + 50, + 180, + 285, + 262 + ], + "type": "table", + "html": "
DatasetourQuality (Average IoU)Frequency (%)
AllPartOversizeShiftReliablePartOversizeShift
Detector as Annotator
Objects-F✓44.324.133.613.340.128.49.821.7
47.030.841.816.849.223.18.019.4
COCO-F✓45.125.433.315.740.027.512.020.5
48.232.140.619.649.522.210.018.3
Point-based Annotator
COCO-P✓55.630.425.029.865.69.523.21.7
65.246.736.440.974.95.918.50.6
", + "image_path": "44d913ab62a0ab0bacbfcaeaddf08a26f451a6d6c1cc513f9377661d50b2955b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 50, + 270, + 282, + 282 + ], + "lines": [ + { + "bbox": [ + 50, + 270, + 282, + 282 + ], + "spans": [ + { + "bbox": [ + 50, + 270, + 282, + 282 + ], + "type": "text", + "content": "Table 10: Analysis of noisy annotations types and quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 306, + 286, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 286, + 403 + ], + "type": "text", + "content": "strategies of ORE in SISD (Table 5). The minus or concatenation of object feature " + }, + { + "bbox": [ + 46, + 306, + 286, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j^+" + }, + { + "bbox": [ + 46, + 306, + 286, + 403 + ], + "type": "text", + "content": " and proposal feature " + }, + { + "bbox": [ + 46, + 306, + 286, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_j" + }, + { + "bbox": [ + 46, + 306, + 286, + 403 + ], + "type": "text", + "content": " do not work. With add strategy, the performance is 27.60. If SISD shares the two fc layers, the performance drops to 22.99 since the optimization goals are contradictory (Identity distinguishes objects in the same category). If we directly use the RoI feature without ORE, the performance drops to 27.32 AP, verifying the effectiveness of the object relevance strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 403, + 286, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 286, + 487 + ], + "type": "text", + "content": "Affect of Re-Train. As most WSOD methods do, we re-run the experiments by training a fully supervised detector for better performance. We find that if the SSD-Det only trains the refiner and uses the pseudo label to train the FasterRCNN, the result is good but lower than re-train after the end-to-end training given in Table 7 (row 1). This is because joint training is beneficial for box refinement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 489, + 286, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 286, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 286, + 537 + ], + "type": "text", + "content": "Experiments on Advanced Detectors. We re-train recent detectors, e.g. SparseRCNN and Deformable DETR, under the boxes refined by our method. Table 8 verifies that our method achieves consistency improvement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 538, + 286, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 286, + 597 + ], + "type": "text", + "content": "The computational cost discussion. Similar to OAMIL, our method adds SSD head to Faster R-CNN for auxiliary training to refine the noisy annotation and the head is not used during inference, the calculation cost during inference is same to standard Faster R-CNN." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 609, + 205, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 205, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 205, + 621 + ], + "type": "text", + "content": "4.4. Visualization and Discussion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "Fig. 4 shows that OA-MIL faces missing instances and grouping instances issues for small or overlapped objects (as mentioned in [12]), while our method still works well. For a better intuitive understanding of SISD and SPSD, we visualize the bag construction quality in Fig. 5 Then, we makes noise types breakdown of 'Drift', 'Group' and 'part dominance' issues. We give the definition of IoU, IoG and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 330, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 330, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 330, + 83 + ], + "type": "text", + "content": "IoD:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 91, + 545, + 112 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 91, + 545, + 112 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 545, + 112 + ], + "type": "interline_equation", + "content": "I o U = \\frac {A (I)}{A (D) + A (G) - A (I)}, I o G = \\frac {A (I)}{A (G)}, I o D = \\frac {A (I)}{A (D)} \\tag {9}", + "image_path": "95741281d5d06df098c01ff478ab5dc3970815684d2c035ea0b44a5434d6623b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\mathrm{A}(\\ast)" + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "text", + "content": " is area of box *, D and G are refined box and gt box respectively, and I is insertion between D and G. We statistically count the proportion of three noise types of 'bad' refined boxes (having small IoU with gt) in Table 9: (i) Drift: 'bad' refined box has a higher IoU with another nearby object. (ii) Group: 'bad' refined box has high IoG with multiple objects. (iii) Part: 'bad' refined box has a high IoD. Table 9 shows quantitative results for each noise type of baseline and ours. The drift, group, part problems reduce from " + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "inline_equation", + "content": "15.1\\%" + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 304, + 121, + 545, + 253 + ], + "type": "text", + "content": ", respectively, demonstrating our improvement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 253, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 480 + ], + "type": "text", + "content": "Experiments on Real-life Noisy Annotations. Real-life noisy annotations stem from: low-quality data (e.g., occlusion, blur), human annotator errors and automatic machine annotator limitations. Noise from human errors is quite subjective, since differences between annotators. For a more objective analysis, noisy annotations from machine annotator is used for experiments. Without loss of generality, Faster R-CNN well-trained on MS-COCO was applied to Objects365 images, yielding Objects-F dataset, and to COCO-val images, producing COCO-F dataset. P2BNet [11], a point-based annotator, was used on COCO-val images with point annotations, generating COCO-P dataset. SSD-Det effectively improves low-quality boxes. As shown in Table 10, with SSD-Det's refinement, the average IoU increases for Objects-F (from 44.3 to 47.0), COCO-F (from 45.1 to 48.2) and COCO-P (from 55.6 to 65.2) datasets. Further, the proportion of reliable annotations increases, and noise categories' frequency (Part, Oversize, and Shift) decreases for all datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 494, + 378, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 494, + 378, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 494, + 378, + 506 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 514, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 657 + ], + "type": "text", + "content": "This paper investigates problems during refinement caused by solely using category information to select proposals. We also propose SSD-Det to mine spatial information in a self-distillation fashion. SSD-Det introduces the SPSD module to learn semantic-spatial correspondence knowledge with neighborhood sampler and an interactive structure to combine spatial information and category information, thus producing a high-quality proposal bag. SISD in SSD-Det is utilized to improve the proposal selection procedure by integrating object-relevant spatial confidence. Complete ablations on multiple datasets verify the effectiveness of SSD-Det." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 670, + 365, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 365, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 365, + 682 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 310, + 691, + 545, + 712 + ], + "type": "text", + "content": "[1] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In CVPR, 2016. 1, 3, 4" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6863" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 105 + ], + "type": "text", + "content": "[2] Léon Bottou. Stochastic gradient descent tricks. In Neural Networks: Tricks of the Trade - Second Edition. Springer, 2012. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 288, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 288, + 137 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, and Gabriel Synnaeve et al. End-to-end object detection with transformers. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 139, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 139, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 53, + 139, + 288, + 161 + ], + "type": "text", + "content": "[4] Simon Chadwick and Paul Newman. Training object detectors with noisy data. In IV, 2019. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 162, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 288, + 194 + ], + "type": "text", + "content": "[5] Guobin Chen, Wongun Choi, and Xiang Yu et al. Learning efficient object detection models with knowledge distillation. In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 195, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 288, + 227 + ], + "type": "text", + "content": "[6] Kai Chen, Jiaqi Wang, and Jiangmiao Pang et al. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 259 + ], + "type": "text", + "content": "[7] Ze Chen, Zhihang Fu, and Rongxin Jiang et al. SLV: spatial likelihood voting for weakly supervised object detection. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 261, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 282 + ], + "type": "text", + "content": "[8] Xing Dai, Zeren Jiang, and Zhao Wu et al. General instance distillation for object detection. In CVPR, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 283, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 288, + 315 + ], + "type": "text", + "content": "[9] Ali Diba, Vivek Sharma, and Ali Mohammad Pazandeh et al. Weakly supervised cascaded convolutional networks. In CVPR, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "text", + "content": "[10] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the multiple instance problem with axis-parallel rectangles. Artificial Intelligence, 1997. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 350, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 380 + ], + "type": "text", + "content": "[11] Chen, et al. Point-to-box network for accurate object detection via single point supervision. In ECCV, 2022. 1, 2, 3, 4, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 382, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 403 + ], + "type": "text", + "content": "[12] Liu, et al. Robust object detection with inaccurate bounding boxes. In ECCV, 2022. 1, 2, 3, 4, 6, 7, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 404, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 436 + ], + "type": "text", + "content": "[13] Mark Everingham, Luc Van Gool, and Christopher K. I. Williams et al. The pascal visual object classes (VOC) challenge. *IJCV*, 2010. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 437, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 469 + ], + "type": "text", + "content": "[14] Aritra Ghosh, Himanshu Kumar, and P. S. Sastry. Robust loss functions under label noise for deep neural networks. In AAAI, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 471, + 251, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 251, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 251, + 480 + ], + "type": "text", + "content": "[15] Ross B. Girshick. Fast R-CNN. In ICCV, 2015. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 482, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 503 + ], + "type": "text", + "content": "[16] Jianyuan Guo, Kai Han, and Yunhe Wang et al. Distilling object detectors via decoupled features. In CVPR, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "type": "text", + "content": "[17] Bo Han, Quanming Yao, and Xingrui Yu et al. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In NeurIPS, 2018. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 537, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 557 + ], + "type": "text", + "content": "[18] Kaiming He, Georgia Gkioxari, and Piotr Dolkar et al. Mask R-CNN. In ICCV, 2017. 1, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 559, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 581 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, and Shaoqing Ren et al. Deep residual learning for image recognition. In CVPR, 2016. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "type": "text", + "content": "[20] Yihui He, Chenchen Zhu, and Jianren Wang et al. Bounding box regression with uncertainty for accurate object detection. In CVPR, 2019. 1, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "text", + "content": "[21] Geoffrey Hinton and Oriol et al. Vinyls. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7), 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[22] Jisoo Jeong, Seungeui Lee, and Jeesoo Kim et al. Consistency-based semi-supervised learning for object detection. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[23] Lu Jiang, Zhengyuan Zhou, and Thomas Leung et al. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In ICML, 2018. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 43, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[24] Vadim Kantorov, Maxime Aquab, and Minsu Cho et al. Contextlocnet: Context-aware deep network models for weakly supervised localization. In ECCV, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "text", + "content": "[25] Junnan Li, Caiming Xiong, Richard Socher, and Steven C. H. Hoi. Towards noise-resistant object detection with noisy annotations. CoRR, abs/2003.01285, 2020. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "text", + "content": "[26] Quanquan Li, Shengying Jin, and Junjie Yan. Mimicking very efficient network for object detection. In CVPR, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 175, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 205 + ], + "type": "text", + "content": "[27] Yandong Li, Di Huang, and Danfeng Qin et al. Improving object detection with selective self-supervised self-training. In ECCV, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "type": "text", + "content": "[28] Mingxiang Liao, Fang Wan, and Yuan Yao et al. End-to-end weakly supervised object detection with sparse proposal evolution. In ECCV, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 242, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 273 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Piotr Dólar, and Ross B. Girshick et al. Feature pyramid networks for object detection. In CVPR, 2017. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 275, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 297 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Priya Goyal, and Ross B. Girshick et al. Focal loss for dense object detection. In ICCV, 2017. 1, 2, 4, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 298, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 319 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin and Michael et al. Maire. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 321, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 342 + ], + "type": "text", + "content": "[32] Wei Liu, Dragomir Anguelov, and Dumitru Erhan et al. SSD: single shot multibox detector. In ECCV, 2016. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 344, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 375 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, and Yue Cao et al. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 377, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 408 + ], + "type": "text", + "content": "[34] Xingjun Ma, Yisen Wang, and Michael E. Houle et al. Dimensionality-driven learning with noisy labels. In ICML, 2018. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "type": "text", + "content": "[35] Jiafeng Mao, Qing Yu, and Kiyoharu Aizawa. Noisy localization annotation refinement for object detection. TIS, 2021. 1, 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "type": "text", + "content": "[36] Ilija Radosavovic, Piotr Dólar, and Ross B. Girshick et al. Data distillation: Towards omni-supervised learning. In CVPR, 2018. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "text", + "content": "[37] Joseph Redmon, Santosh Kumar Divvala, and Ross B. Girshick et al. You only look once: Unified, real-time object detection. In CVPR, 2016. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "text", + "content": "[38] Shaoqing Ren, Kaiming He, and Ross B. Girshick et al. Faster R-CNN: towards real-time object detection with region proposal networks. TPAMI, 2017. 1, 2, 6, 7" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[39] Zhongzheng Ren, Zhiding Yu, and Xiaodong Yang et al. Instance-aware, context-focused, and memory-efficient weakly supervised object detection. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 307, + 580, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 611 + ], + "type": "text", + "content": "[40] Kihyuk Sohn, David Berthelot, and Nicholas Carlini et al. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "text", + "content": "[41] Kihyuk Sohn, Zizhao Zhang, and Chun-Liang Li et al. A simple semi-supervised learning framework for object detection. CoRR, 2020. 3" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[42] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. SELFIE: refurbishing unclean samples for robust deep learning. In ICML, 2019. 3" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[43] Peize Sun, Rufeng Zhang, and Yi Jiang et al. Sparse R-CNN: end-to-end object detection with learnable proposals. In CVPR, 2021. 2, 3, 8" + } + ] + } + ], + "index": 42 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6864" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[44] Peng Tang and Xinggang Wang et al. Multiple instance detection network with online instance classifier refinement. In CVPR, 2017. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "text", + "content": "[45] Peng Tang, Chetan Ramaiah, and Yan Wang et al. Proposal learning for semi-supervised object detection. In WACV, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[46] Peng Tang, Xinggang Wang, and Song Bai et al. PCL: proposal cluster learning for weakly supervised object detection. TPAMI, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 176, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 209 + ], + "type": "text", + "content": "[47] Koen E. A. van de Sande, Jasper R. R. Uijlings, and Theo Gevers et al. Segmentation as selective search for object recognition. In ICCV, 2011. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 211, + 287, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 287, + 242 + ], + "type": "text", + "content": "[48] Fang Wan, Pengxu Wei, and Zhenjun Han et al. Min-entropy latent model for weakly supervised object detection. TPAMI, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 277 + ], + "type": "text", + "content": "[49] Keze Wang, Xiaopeng Yan, and Dongyu Zhang et al. Towards human-machine cooperation: Self-supervised sample mining for object detection. In CVPR, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 280, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 311 + ], + "type": "text", + "content": "[50] Tao Wang, Li Yuan, and Xiaopeng Zhang et al. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 313, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 313, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 313, + 287, + 346 + ], + "type": "text", + "content": "[51] Yunchao Wei, Zhiqiang Shen, and Bowen Cheng et al. TS ^2 2 C: tight box mining with surrounding segmentation context for weakly supervised object detection. In ECCV, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 348, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 348, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 348, + 287, + 380 + ], + "type": "text", + "content": "[52] Mengde Xu, Zheng Zhang, and Han Hu et al. End-to-end semi-supervised object detection with soft teacher. In ICCV. IEEE, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 415 + ], + "type": "text", + "content": "[53] Youjiang Xu, Linchao Zhu, and Yi Yang et al. Training robust object detectors from noisy category labels and imprecise bounding boxes. TIP, 2021. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 416, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 449 + ], + "type": "text", + "content": "[54] Ke Yang, Dongsheng Li, and Yong Dou. Towards precise end-to-end weakly supervised object detection network. In ICCV, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 451, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 451, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 451, + 287, + 482 + ], + "type": "text", + "content": "[55] Zhendong Yang, Zhe Li, and Xiaohu Jiang et al. Focal and global knowledge distillation for detectors. In CVPR, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 485, + 287, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 516 + ], + "type": "text", + "content": "[56] Xuehui Yu, Pengfei Chen, and Di Wu et al. Object localization under single coarse point supervision. In CVPR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 519, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 287, + 541 + ], + "type": "text", + "content": "[57] Xuehui Yu, Yuqi Gong, and Nan Jiang et al. Scale match for tiny person detection. In WACV, 2020. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 544, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 574 + ], + "type": "text", + "content": "[58] Xiaosong Zhang, Fang Wan, and Chang Liu et al. Freeanchor: Learning to match anchors for visual object detection. In NeurIPS, 2019. 1, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "type": "text", + "content": "[59] Xiaopeng Zhang, Yang Yang, and Jiashi Feng. Learning to localize objects with noisy labeled instances. In AAAI, 2019. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 643 + ], + "type": "text", + "content": "[60] Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In NeurIPS, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "text", + "content": "[61] Zhaohui Zheng, Rongguang Ye, and Ping Wang et al. Localization distillation for dense object detection. In CVPR, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[62] Xizhou Zhu, Weijie Su, and Lewei Lu et al. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3, 8" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 118 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 95 + ], + "type": "text", + "content": "[63] C. Lawrence Zitnick and Piotr Dálár. Edge boxes: Locating object proposals from edges. In ECCV, 2014. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "text", + "content": "[64] Barret Zoph, Golnaz Ghiasi, and Tsung-Yi Lin et al. Rethinking pre-training and self-training. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6865" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spatial-Aware Token for Weakly Supervised Object Localization/f549b4c7-58f5-4874-8f3f-2978b440c9a2_content_list.json b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/f549b4c7-58f5-4874-8f3f-2978b440c9a2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9dcdd7101ee600f1341ed9fa15a7fec0798db292 --- /dev/null +++ b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/f549b4c7-58f5-4874-8f3f-2978b440c9a2_content_list.json @@ -0,0 +1,1630 @@ +[ + { + "type": "text", + "text": "Spatial-Aware Token for Weakly Supervised Object Localization", + "text_level": 1, + "bbox": [ + 158, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pingyu Wu $^{1}$ , Wei Zhai $^{1,\\dagger}$ , Yang Cao $^{1,3}$ , Jiebo Luo $^{2}$ , Zheng-Jun Zha $^{1}$ $^{1}$ University of Science and Technology of China \n $^{2}$ University of Rochester \n $^{3}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center", + "bbox": [ + 158, + 179, + 810, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wpy364755620@@mail., wzhai056@mail., forrest@}ustc.edu.cn", + "bbox": [ + 228, + 236, + 743, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jluo@cs.ROchester.edu zhazj@ustc.edu.cn", + "bbox": [ + 299, + 253, + 671, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 318, + 313, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weakly supervised object localization (WSOL) is a challenging task aiming to localize objects with only image-level supervision. Recent works apply visual transformer to WSOL and achieve significant success by exploiting the long-range feature dependency in self-attention mechanism. However, existing transformer-based methods synthesize the classification feature maps as the localization map, which leads to optimization conflicts between classification and localization tasks. To address this problem, we propose to learn a task-specific spatial-aware token (SAT) to condition localization in a weakly supervised manner. Specifically, a spatial token is first introduced in the input space to aggregate representations for localization task. Then a spatial aware attention module is constructed, which allows spatial token to generate foreground probabilities of different patches by querying and to extract localization knowledge from the classification task. Besides, for the problem of sparse and unbalanced pixel-level supervision obtained from the image-level label, two spatial constraints, including batch area loss and normalization loss, are designed to compensate and enhance this supervision. Experiments show that the proposed SAT achieves state-of-the-art performance on both CUB-200 and ImageNet, with $98.45\\%$ and $73.13\\%$ GT-known Loc, respectively. Even under the extreme setting of using only 1 image per class from ImageNet for training, SAT already exceeds the SOTA method by $2.1\\%$ GT-known Loc. Code and models are available at https://github.com/wpy1999/SAT.", + "bbox": [ + 75, + 351, + 473, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 804, + 209, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weakly supervised object localization (WSOL) aims to localize objects with only image-level labels available. Since no expensive bounding box or pixel-level annota", + "bbox": [ + 76, + 830, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08b6b50cb1c4fdaed3f540bc09b82ffd93d6f644afd7cd9d4aebf7d540d41ec6.jpg", + "image_caption": [ + "Figure 1. GT-known Loc on ImageNet. The proposed method only needs to fine-tune a small number of parameters or requires a few training images to achieve significant improvement." + ], + "image_footnote": [], + "bbox": [ + 509, + 300, + 699, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6eaee214423d52cf6cbaee399e0e1d7db2238becda7ffbbb3e863d90d5b658de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 300, + 882, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions are required, WSOL significantly reduces the cost of manual annotations [21, 44, 12, 13, 4, 43, 22, 45] and has attracted increasing attention in the research community [40, 42, 41, 39, 10, 27, 29, 28, 14, 38, 23].", + "bbox": [ + 496, + 551, + 892, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a representative work, CAM [48] extracts class activation maps from the classifier as localization maps. However, CAM is usually coarse and focuses on the most discriminative regions, leading to imprecise and incomplete localization results. To solve these problems, many CNN-based methods have been proposed, such as adversarial erasure [22, 45, 6, 16], divergent activation [36, 40, 22], seed region growing [30, 46], regularization [18, 13, 47], feature refining [1, 37, 30], regression-based [35, 43, 9].", + "bbox": [ + 496, + 613, + 893, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, transformer [7, 24] has been introduced to the field of computer vision with great success. Benefiting from the long-range feature dependency, the attention map upon the class token usually can capture the object region relevant to classification in the whole image, thus it is widely used for localization in the transformer-based WSOL methods. The pioneering work TS-CAM [8] accumulates attention maps of each layer and mixes them with semantic-aware maps to achieve localization task. Further, to address the problem that transformer lacks inherent spatial coherence", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 101, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1844", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of the object, LCTR [3] and SCM [2] propose to consider cross-patch information and use activation diffusion to increase the local continuity of localization map, respectively.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing transformer-based methods synthesize the feature maps learned by the classification task, such as attention maps, as the localization map, and try to increase its connectivity and completeness. However, this convenient approach results in optimization conflicts between classification and localization tasks. 1) For classification, making classification feature maps learn more object regions with low discrimination will reduce the classification ability. 2) For localization, the learning of the localization map is limited by the properties of the feature maps. For example, the attention map is generated by the softmax function, which is difficult to produce a balanced and comprehensive response over the object. Therefore, to achieve a more promising localization performance and avoid optimization conflicts, it is necessary to construct task-specific parameters for the generation and learning of localization map.", + "bbox": [ + 75, + 138, + 468, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on transformer architecture, a straightforward idea is to introduce a spatial token in the input space to aggregate global representation for the localization task. And the localization map can be further obtained by interacting the spatial token with each patch in the forward propagation. To this end, we propose a Spatial-Query Attention (SQA) module that takes the spatial token as a query to calculate the similarity with different patches and produces the localization map efficiently. Meanwhile, to make the spatial-aware token obtain localization supervision from image-level labels, the localization map is treated as a visual cue to participate in the calculation of cross attention, thus the generation and learning of localization map can be well adapted to the transformer-based classification model.", + "bbox": [ + 75, + 383, + 468, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon this efficient SQA module, we propose a simple but effective Spatial-Aware Token (SAT) approach, as shown in Fig. 2. Specifically, a task-specific spatial token is introduced in the input space. To achieve the localization task, several SQA modules are applied in the transformer to produce localization maps learned from different layers and aggregate them together. In this way, the final localization map can maximally capture the localization knowledge from the whole classification model. However, the pixel-level supervision generated by image-level labels is sparse and unbalanced. To compensate and strengthen this supervision, we propose two spatial constraints, including batch area loss and normalization loss. Batch area loss aims to provide a sparse area supervision with prior knowledge to compensate for the insufficient supervision. Normalization loss is employed to enhance pixel-level supervision by encouraging the localization map to be more discriminatory.", + "bbox": [ + 75, + 595, + 468, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benefiting from the task-specific spatial token and avoiding optimization conflicts, SAT releases the potential of the model and achieves excellent performance in both classifi", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cation and localization tasks. Besides, unlike synthetic classification feature maps as localization results, the generation and learning of localization map in SAT rely mainly on an extra lightweight token, which brings advantages of data-efficiency and tuning-efficiency. As illustrated in Fig. 1, SAT outperforms SOTA method SCM by $2.1\\%$ GT-known Loc with less than $0.1\\%$ training data on ImageNet.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of this paper include:", + "bbox": [ + 517, + 196, + 861, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We propose to learn a task-specific spatial token to implement the localization task, instead of synthesizing classification feature maps, thus avoiding optimization conflicts between classification and localization tasks.", + "2) We propose a simple but effective spatial-aware token (SAT) pipeline for WSOL, which utilizes a spatial-query attention (SQA) module to generate localization map through a spatial token and supervise it with two spatial constraints, including batch area loss and normalization loss.", + "3) Extensive experiments show that SAT outperforms SOTA methods by a large margin on multiple benchmarks and performs excellently even under extreme settings." + ], + "bbox": [ + 496, + 212, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 407, + 637, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CNN-based methods. WSOL aims to learn the localization of objects with only image-level labels. CAM [48] proposes to synthesize depth feature maps with fully connected weights to obtain class activation maps (CAM). To increase the localization ability of CAM, HaS [22] adopts a random erasure strategy forcing the network to attend to different object regions. ACoL [45] and EIL [16] use adversarial erasing to achieve learning of complementary regions. ADL [4] drops the most discriminative or highlighted regions in the different layers during forward propagation. Besides, SPA [18] designs a post-process approach to extract structure-preserving localization map from the feature maps. SPG [46] and SPOL [30] facilitate learning of localization by selecting and spreading confidence regions as seeds. PSOL [43] and $\\mathrm{C}^2\\mathrm{AM}$ [34] divide WSOL into independent classification task and class-agnostic object localization. I2C [47] and ISIC [31] consider feature similarities across different objects to achieve more complete and robust localization. ORNet [33], FAM [17] and BAS [32] propose to generate a foreground prediction map (FPM) to implement localization. Unlike FPM-based methods that require complex and heavy structures, and the learning of generator relies on a specific feature layer of the classification network. Our method simply and efficiently obtains localization maps from different layers by a task-specific token.", + "bbox": [ + 496, + 431, + 890, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer-based methods. Different from CNNs, transformer [7, 25] can capture global cues with the advantage of long-range dependency in the self-attention mechanism, effectively alleviating the partial activation problem. TS-CAM [8] mixes the attention maps of different layers and combines them with semantic-aware tokens to produce", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1845", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e24218d708a5a464ac22cd185e7c0ecadc2c19c3b02ffed34495bba8e516dd59.jpg", + "image_caption": [ + "Figure 2. Framework of SAT. It includes three spatial aware transformer blocks at the end of the network. Each block generates a localization map $M^l$ using the spatial-query attention module. The final localization map $M$ is obtained by fusing $M^l$ of different layers." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 890, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "semantic-aware localization results. Based on TS-CAM, to address the lack of spatial consistency in transformer, LCTR [3] proposes to enhance the local perception capability among long-range feature dependencies by considering cross-patch information. SCM [2] increases the spatial coherence of attention maps by diffusing the semantic and spatial connections between patch tokens. Notably, they all generate localization map by synthesizing feature maps. Instead, we produce it by a task-specific spatial token.", + "bbox": [ + 75, + 332, + 472, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 488, + 210, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 515, + 187, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall network architecture of SAT is shown in Fig. 2. Given an input image $I \\in \\mathbb{R}^{h \\times w \\times 3}$ , a plain vision transformer splits it into a sequence of non-overlapping patches. Then the divided patches are flattened and transformed into patch tokens $\\{x_{n} \\in \\mathbb{R}^{1 \\times D}, n = 1, 2, \\dots, H \\times W\\}$ by linear projection, where $H = h / P$ , $W = w / P$ , $P$ is the patch size, and $D$ is the number of channels. For simplicity, we omit the description of the batch size $B$ . After grouping the class token and the extra spatial-aware token $x_{spa}$ with patch tokens, this token sequence $\\mathcal{F}^0 \\in \\mathbb{R}^{(N + 2) \\times D}$ is sent into stacked transformer blocks and spatial aware transformer blocks for subsequent representation learning.", + "bbox": [ + 75, + 541, + 468, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed spatial aware transformer block inherits the typical transformer block structure. Thus the entire network is characterized by alternating attention layers and MLP layers. For a given input token subsequent $\\mathcal{F}^{l-1}$ , the overall equations of the $l$ -th block are defined as follows:", + "bbox": [ + 75, + 724, + 468, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} ^ {l} = \\operatorname {L a y e r - N o r m} \\left(\\mathcal {F} ^ {l - 1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 816, + 468, + 842 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Z} ^ {l} = \\mathcal {F} ^ {l - 1} + \\text {A t t e n t i o n} \\left(\\mathcal {X} ^ {l}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 844, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} ^ {l} = \\mathcal {Z} ^ {l} + \\operatorname {M L P} \\left(\\text {L a y e r - N o r m} \\left(\\mathcal {Z} ^ {l}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 872, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Spatial Aware Transformer Block", + "text_level": 1, + "bbox": [ + 500, + 330, + 795, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed spatial aware transformer block mainly contains a spatial-query attention (SQA) module to generate localization map and extract localization knowledge. Besides, two spatial constraints, including batch area loss $\\mathcal{L}_{ba}$ , and normalization loss $\\mathcal{L}_{norm}$ are designed as the complement to the insufficient localization supervision.", + "bbox": [ + 496, + 354, + 890, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial-query Attention. The proposed SQA module is based on self-attention module and aims at converting the spatial-aware token $x_{spa}$ into a localization map $M_{spa}^{l}$ by querying different patches. Meanwhile, the generated $M_{spa}^{l}$ is applied to the cross-attention calculation in the form of dot product, thus obtaining localization supervision from the image-level labels. Specifically, the SQA module first linearly projects the token sequence $\\mathcal{X}^{l - 1}$ to the query matrix $\\mathcal{Q}$ , key matrix $\\mathcal{K}$ , value matrix $\\mathcal{V}$ , where $\\mathcal{Q},\\mathcal{K},\\mathcal{V} \\in \\mathbb{R}^{(N + 2)\\times D}$ . Then the query vector corresponding to the spatial token $\\mathcal{Q}_{spa} \\in \\mathbb{R}^{1\\times D}$ is selected from $\\mathcal{Q}$ and used to query $\\mathcal{K}$ to obtain the query results for each token. The similarity map between the spatial query vector and key matrix in the $l$ -th layer is calculated as follows:", + "bbox": [ + 498, + 446, + 892, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} ^ {l} \\left(\\mathcal {Q} _ {s p a}, \\mathcal {K}\\right) = \\frac {\\mathcal {Q} _ {s p a} \\mathcal {K} ^ {\\mathrm {T}}}{\\sqrt {D}} \\in \\mathbb {R} ^ {1 \\times (N + 2)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 665, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sqrt{D}$ is employed as a scaling factor. After applying sigmoid activation, the similarity map is transformed into foreground probability $M_{pro}^{l}$ with a 0 to 1 interval distribution. This process can be expressed as follows:", + "bbox": [ + 496, + 707, + 890, + 768 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {s p a} ^ {l} = \\operatorname {S i g m o i d} \\left(\\mathcal {S} ^ {l} \\left(\\mathcal {Q} _ {s p a}, \\mathcal {K}\\right)\\right) \\in \\mathbb {R} ^ {1 \\times (N + 2)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 773, + 890, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The generated $M_{spa}^{l}$ is served as a visual cue to point out the object region and combined in the calculation of attention with the form of dot product. In this way, the whole SQA module can be formulated in the following form:", + "bbox": [ + 496, + 801, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S Q A} \\left(\\mathcal {X} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q K} ^ {\\mathrm {T}}}{\\sqrt {D}}\\right) * M _ {s p a} ^ {l} \\mathcal {V}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 869, + 890, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1846", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/86faad96805cfacef49e0f7948004b07a400a3044fe353a45fc7dbb526495375.jpg", + "image_caption": [ + "Figure 3. Visualization comparison. The ground-truth bounding boxes are in red, and the predicted bounding boxes are in green." + ], + "image_footnote": [], + "bbox": [ + 81, + 80, + 893, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $*$ denotes element-wise multiplication.", + "bbox": [ + 76, + 409, + 375, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By cropping and reshaping, the part of the $M_{spa}^{l}$ to the patch tokens can be transformed into a localization map $M^{l} \\in \\mathbb{R}^{H \\times W}$ . After obtaining $M^{l}$ learned from different layers, we take their averages as the final localization map $M$ to increase the completeness and robustness. Although this approach can maximally capture the localization information in the classification network, the pixel-level supervision obtained from image-level labels is sparse and unbalanced. To compensate and strengthen this supervision, we design batch area loss and normalization loss.", + "bbox": [ + 76, + 424, + 468, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Batch Area. To alleviate the insufficient localization supervision caused by the absence of dense labels in WSOL, we propose to use batch area loss as a complement to localization supervision. It aims to constrain the average area of localization maps $\\{M_b\\in \\mathbb{R}^{H\\times W},b = 1,2,\\dots,B\\}$ within a batch $B$ to a hyperparameter $\\lambda$ , as follows:", + "bbox": [ + 76, + 575, + 468, + 665 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b a} = \\left| \\sum_ {b} ^ {B} \\sum_ {i} ^ {H} \\sum_ {j} ^ {W} \\left(\\lambda - \\frac {M _ {b} (i , j)}{B \\times H \\times W}\\right) \\right|, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 670, + 468, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a sparse area supervision with prior knowledge. The $\\lambda$ is set to 0.25 and 0.35 on CUB-200 and ImageNet.", + "bbox": [ + 76, + 719, + 467, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normalization. For unbalanced pixel-level supervision, a normalization loss is applied to enhance the pixel-level supervision based on the pixel response intensity. It encourages pixels to be more distinguishable about whether they are foreground or background, as presented by Eq. 8. Before penalizing the uncertainty value of the localization map $M$ , to further enhance its local consistency, we adopt a gaussian filter with a kernel size of $3 \\times 3$ to establish spatial connections between adjacent patches. The standard deviation of gaussian kernel is set to 6 on all datasets. After", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "applying gaussian filtering, the value of each patch is redistributed according to the adjacent patch values, and $\\mathbf{M}$ is converted to $M^{*}$ . By computing the normalization loss with $M^{*}$ can further encourage $M$ to have local consistency and avoid noise. It can be expressed as follows:", + "bbox": [ + 498, + 409, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m}} = \\frac {1}{H \\times W} \\sum_ {i} ^ {H} \\sum_ {j} ^ {W} M ^ {*} (i, j) (1 - M ^ {*} (i, j)). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 494, + 890, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Total loss", + "text_level": 1, + "bbox": [ + 500, + 547, + 607, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow the design of classification head in the baseline method TS-CAM [8] to obtain the predicted probability $y$ and the cross-entropy classification loss $L_{cls}$ . By jointly optimizing classification loss, batch area loss, and normalization loss, the total loss function is defined as follows:", + "bbox": [ + 498, + 571, + 890, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {b a} + \\mathcal {L} _ {\\text {n o r m}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 660, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where no extra hyperparameter is used to balance each loss.", + "bbox": [ + 500, + 688, + 888, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 715, + 624, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 741, + 689, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We evaluate the proposed SAT on three popular datasets, including CUB-200 [26], ImageNet [20], OpenImages [5]. CUB-200 is a fine-grained dataset with 200 bird species, which consists of 5,994 images for training and 5,794 images for testing. ImageNet contains about 1.2 million training images and 50,000 validation images from 1,000 different classes. OpenImages contains 100 categories, it has 29,819, 2,500, and 5,000 samples in training, val, and test sets, respectively. Apart from the image-level", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1847", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d00f375b66a8e49c8d95c71bcecd9d3b6ea6c050c6fdc015c5dc3e14cdd8c1c9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsVenueBackboneCUB-200 [26]Loc Acc.ImageNet [20]Loc Acc.
Top-1Top-5GT-knownTop-1Top-5GT-known
CAM [48]CVPR16VGG1641.0650.6655.1042.8054.8659.00
ORNet [33]ICCV21VGG1667.7380.7786.2052.0563.9468.27
BAS [32]CVPR22VGG1671.3385.3391.0752.9665.4169.64
Kim et al. [11]CVPR22VGG1670.8388.0793.1749.9463.2568.92
CREAM [35]CVPR22VGG1670.4485.6790.9852.3764.2068.32
GCNet [15]ECCV20InceptionV358.5871.0075.3049.0658.09-
SPA [18]CVPR21InceptionV353.5966.5072.1452.7364.2768.33
FAM [17]ICCV21InceptionV370.67-87.2555.24-68.62
CREAM [35]CVPR22InceptionV371.7686.3790.4356.0766.1969.03
BAS [32]CVPR22InceptionV373.2986.3192.2458.5169.0071.93
BagCAMs [49]ECCV22InceptionV360.07-89.7853.87-71.02
SPOL [30]CVPR21ResNet5080.1293.4496.4659.1467.1569.02
DA-WSOL [50]CVPR22ResNet5066.65-81.8355.84-70.27
BAS [32]CVPR22ResNet5077.2590.0895.1357.1868.4471.77
Kim et al. [11]CVPR22ResNet5073.1686.6891.6053.7665.7569.89
CREAM [35]CVPR22ResNet5076.03-89.8855.66-69.31
BagCAMs [49]ECCV22ResNet5069.67-94.0144.24-72.08
ISIC [31]ECCV22ResNet5080.6894.0897.3259.6167.8470.01
TS-CAM [8]ICCV21Deit-S71.3083.8087.7053.4064.3067.60
LCTR [3]AAAI22Deit-S79.2089.9092.4056.1065.8068.70
SCM [2]ECCV22Deit-S76.4091.6096.6056.1066.4068.80
SAT (ours)This WorkDeit-S80.9694.1398.4560.1570.5273.13
", + "bbox": [ + 114, + 88, + 856, + 429 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a75a80fa1023d22f0fb2b19452797e51f9cc9ea2a9ea841c42865b09856173fe.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art methods. The best results are highlighted in bold, second are underlined." + ], + "table_footnote": [], + "table_body": "
SettingLoc Acc.
LclsLbaLnorm F.Top-1Top-5GT-k.
(a)✓56.8666.6769.29
(b)✓✓58.2168.3671.07
(c)✓✓✓59.8670.2272.78
(d)✓✓✓✓60.1570.5273.13
", + "bbox": [ + 78, + 473, + 344, + 565 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/11bee27d3e8496a83fc8c0c6dd7175be542f69f71cced9cb164379896020fb31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 468, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Ablation study of SAT on ImageNet. $F$ : Gaussian filter.", + "bbox": [ + 76, + 571, + 467, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "labels, both CUB-200 and OpenImages provide pixel-level mask labels, which are only used in the testing phase.", + "bbox": [ + 75, + 612, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. Following [32, 18, 50], for localization, we utilize GT-known localization accuracy (GT-known Loc), Top-1/Top5 localization accuracy (Top-1/Top-5 Loc), and maximal box accuracy (MaxBoxAccV2) [5] as evaluation metrics. GT-known Loc is correct indicating that the intersection over union (IoU) of the predicted bounding box and the ground-truth bounding box is $50\\%$ or more. Top-1/Top-5 Loc is correct when the ground-truth class belongs to Top-1/Top-5 prediction categories and GT-known Loc is correct. For mask, the peak intersection over union (pIoU) [43] and pixel average precision (PxAP) [5] are adopted as metrics.", + "bbox": [ + 75, + 643, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We evaluate our method on the Deit-S [24] pre-trained on ImageNet [20]. On CUB-200 [26], the training process lasts 30 epochs with a batch size of 256. For ImageNet [20], we train 5 epochs and set batch size to 512. In the training phase, the input images are resized to $256 \\times 256$ and then randomly cropped to", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2005113709f5d946aa16ac29d6d23f36103935fda719c1897568f9e0c302385.jpg", + "image_caption": [ + "Figure 4. Hyperparameters. Sensitivity analysis of hyperparameters (a) $\\lambda$ and (b) batch size in $\\mathcal{L}_{ba}$ on ImageNet." + ], + "image_footnote": [], + "bbox": [ + 498, + 468, + 694, + 563 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/556a46dcf2f9f9fad1c2ea261e194bf2374c78cab75032d2483ba4ff262a668c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 468, + 890, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$224 \\times 224$ . In the inference phase, following [32, 9, 43], we adopt ten crop augmentations to obtain classification results and replace random crop with center crop for localization.", + "bbox": [ + 498, + 621, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 678, + 653, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this subsection, we implement a series of ablation experiments with Deit-S [24] as the backbone. All experiments are conducted on ImageNet [20], a universal dataset, to increase the generalizability of the experimental results.", + "bbox": [ + 496, + 703, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ablation studies of SAT components. Table 2 shows the localization accuracy of SAT with different compositions, where we retain the structure of SAT as the baseline. It can be noted that the proposed baseline can effectively activate the object region and already exceeds SOTA method SCM with only $\\mathcal{L}_{cls}$ . Based on the baseline, the addition of $\\mathcal{L}_{ba}$ significantly improves the GT-known Loc with a $1.78\\%$ gain, by limiting the expansion of the localization map and forcing it to focus on the regions relevant to the classifi-", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1848", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d051c84daff14cadd2466fa70dda38c529c81c78f9cd8eb2d28f6b66ee993e95.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueBackboneResolutionCUB-200 [26]ImageNet [20]
δ = 0.3δ = 0.5δ = 0.7Meanδ = 0.3δ = 0.5δ = 0.7Mean
Kim et al. [11]CVPR22ResNet5028×2899.4090.4038.0075.9086.7071.1048.3068.70
BAS [32]CVPR22ResNet5028×2899.4195.1371.9288.8284.8071.7749.2568.61
TS-CAM [8]ICCV21DeiT-S14×1498.8887.7049.9378.8482.1167.6044.7664.82
SCM [2]ECCV22DeiT-S14×1499.6496.6071.7389.3283.6468.8045.3065.91
SAT (ours)This WorkDeiT-S14×1499.88 (+0.24)98.45 (+1.85)79.53 (+7.61)92.62 (+3.30)84.42 (-2.28)73.13 (+1.36)56.83 (+7.58)71.46 (+2.76)
", + "bbox": [ + 81, + 88, + 890, + 207 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4bf90063eb1c7071935a4b480656b2d3a359f2e399d738ddabc93c748fe865fb.jpg", + "image_caption": [ + "Figure 5. Statistical analysis about localization quality." + ], + "image_footnote": [], + "bbox": [ + 78, + 253, + 272, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf14a74d5c0a95193dce7d766148632a5156bcaed8014be42221bbc31a93c9b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 253, + 468, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cation. Then the use of $\\mathcal{L}_{\\text {norm }}$ can effectively increase the distinction between the foreground and background regions, resulting in a remarkable increase of $1.71\\%$ GT-known Loc. Finally, applying gaussian filter can further eliminate noise and increase local connectivity, by considering the consistency of adjacent patches in the calculation of $\\mathcal{L}_{\\text {norm }}$ .", + "bbox": [ + 75, + 415, + 468, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperparameters. Fig. 4 presents the sensitivity of localization quality to the hyperparameters $\\lambda$ and batch size in $\\mathcal{L}_{ba}$ (Eq. 7). $\\lambda$ denotes the constraint for the average area of the localization maps. A larger $\\lambda$ encourages localization map to learn more regions relevant to the classification. Conversely, when $\\lambda$ is small, the region learned by the localization map is small but reliable. Fig. 4 (a) shows that the model behaves insensitive to $\\lambda$ , since GT-known Loc remains stable over a large interval. Batch size affects the calculation of the average area of the localization map within a batch. A sufficiently large batch size can ensure tolerance of area variation between instances. As shown in Fig. 4 (b), when the batch size is larger than 64, the effect of batch on GT-known Loc is limited, with a change of less than $0.5\\%$ .", + "bbox": [ + 75, + 506, + 470, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Performance", + "text_level": 1, + "bbox": [ + 76, + 726, + 212, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main Results. Table 1 compares SAT with state-of-the-art methods on CUB-200 [26] and ImageNet [20]. Experiments show that SAT achieves stable and consistent improvements on both benchmark datasets and significantly outperforms all methods using various backbones. Compared to the CNN counterpart ResNet50-ISIC [31], which uses three separate networks, while we use only one network and surpass ResNet50-ISIC by $1.13\\%$ and $3.12\\%$ GT-known Loc on the CUB-200 and ImageNet, respectively. These quantitative results demonstrate the superi", + "bbox": [ + 75, + 750, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e22ce522dbf1cdb4403c9263c58674a1fc7170b4aaa2ac5de0991f375e35f3bd.jpg", + "table_caption": [ + "Table 3. Localization quality. Comparison of localization accuracy under different IoU thresholds $(\\delta)$ . Best results are highlighted in bold, second are underlined. The last line indicates the improvement (decrease) of our method over the previous best method." + ], + "table_footnote": [], + "table_body": "
MethodVenueBackboneCUB-200OpenImages
pIoUPxAPpIoUPxAP
BagCAMs [49]ECCV22IncepV360.3481.4949.98
DA-WSOL [50]CVPR22ResNet5056.1874.7049.68
BAS [32]CVPR22ResNet5071.0489.2050.72
TS-CAM [8]ICCV21DeiT-S66.5581.4943.65
SCM [2]ECCV22DeiT-S72.0788.15-
SAT (ours)This workDeiT-S77.2189.8759.54
", + "bbox": [ + 500, + 261, + 893, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Segmentation quality of the localization map compared with other SOTA methods. IncepV3: InceptionV3.", + "bbox": [ + 500, + 382, + 890, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ority of the proposed SAT in terms of simplicity and effectiveness. In addition, on CUB-200 [26], SAT achieves a remarkable performance of $80.96\\% / 98.45\\%$ on Top-1/GT-known Loc, exceeding the baseline method TS-CAM [8] by $9.66\\% / 10.75\\%$ . On ImageNet [20], we surpass the SOTA method SCM [2] by a large margin, achieving a $4.33\\%$ boost in GT-known Loc under the same Deit-S [24] backbone. To further demonstrate the effectiveness of SAT, we visualize the localization maps of our method and TS-CAM in Fig. 3. In addition to generating sharper and more complete localization maps than TS-CAM, SAT demonstrates robust localization ability in various complex and challenging scenarios. More accuracy results and visualizations on six datasets are provided in the supplementary materials.", + "bbox": [ + 496, + 443, + 890, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Localization Quality. In Table 3, we compare the localization performance with SOTA methods under the MaxBoxAccV2 [5] criterion. Despite the limitation of localization map resolution $(14\\times 14)$ , we achieve excellent results under different IoU thresholds. Notably, at $\\delta$ of 0.7, a strict criterion for the localization quality, the proposed SAT exceeds the previous best method by a significant margin, with $7.61\\%$ and $7.58\\%$ improvement on CUB-200 [26] and ImageNet [20], respectively. We further plot the IoU distribution of correct bounding boxes in Fig. 5, following DANet [36]. Compared to the latest SCM [2], with the same Deit-S [24] backbone, we obtain a median IoU gain of $2.9\\%$ and $6.1\\%$ on CUB-200 and ImageNet. The IoU distribution map demonstrates that the localization results generated by SAT have higher localization quality, especially on ImageNet [20], a universal and challenging dataset.", + "bbox": [ + 496, + 657, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1849", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae691c9f8a77d3f5f7f0440d1677ac92a471fc3f810d56e39e89f45fcbc46cc8.jpg", + "image_caption": [ + "Figure 6. Density distribution map about IoU and object size on CUB-200, both measured using pixel-level masks." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 269, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac92df289a7ca82909855dcfb79c63d5d949bd5d3eef5fe59dfdfdce5e16b3e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 88, + 467, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/038fdcc0823c5a05e279d7e959d1e4fe346096fae3d1e3994a07bc9c90c917a8.jpg", + "image_caption": [ + "Figure 7. Threshold sensitivity. GT-known Loc-Threshold curves for different methods on CUB-200 and ImageNet." + ], + "image_footnote": [], + "bbox": [ + 76, + 266, + 279, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8c7d6ee241cba9dacb0549531af32aa35ffa9aa6a38b2c9a95cd308b66d8f36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 267, + 465, + 376 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2fcc6aea52b2a539aef920aa7bfe41a45624fdac9b546812ab83a76c22898939.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Localization GenerationCls Acc. Top-1 Top-5Loc Acc. Top-1 GT-k.
Attention map75.8992.7755.9869.74
Spatial token78.4194.4660.1573.13
+Δ+2.52+1.69+4.17+3.39
", + "bbox": [ + 78, + 420, + 328, + 496 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86c419755180c6a119cce9b7893a0507d0f8a8407588a56f719be568b0ae8530.jpg", + "image_caption": [ + "Table 5. Accuracy of different baselines with the same losses. We follow the idea of TS-CAM to extract the attention maps (before softmax and adding a sigmoid function) to synthesize the localization map, and apply the proposed losses to it." + ], + "image_footnote": [], + "bbox": [ + 334, + 420, + 468, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Segmentation Quality. We evaluate the segmentation performance of SAT on CUB-200 [26] and OpenImages [5] since pixel-level labels are available during evaluation. As shown in Table 4, the proposed SAT substantially exceeds all methods in pIoU and PxAP metrics. Especially compared to the baseline method TS-CAM [8], we achieve a remarkable increase of $10.66\\%$ and $15.89\\%$ pIoU on CUB-200 and OpenImages. To further analyze the segmentation quality, we draw the density distribution map about IoU and object size on CUB-200 in Fig. 6. It can be observed the proposed SAT achieves an overall improvement and more stable performance on arbitrary sized objects compared to TS-CAM. Nevertheless, there still exists an imbalance segmentation ability of different sized objects. This is mainly due to the limitation of the localization map resolution, which hinders the acquisition of fine segmentation results.", + "bbox": [ + 75, + 582, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Threshold Sensitivity. Fig. 7 illustrates the curves of GT-known Loc under different thresholds on CUB-200 and ImageNet. Obviously, the curves corresponding to the proposed SAT entirely cover the curves of TS-CAM [8] and SCM [2], indicating that SAT performs stable and achieves", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a7f5ae73763c620ad31f4ef2f2a7f5ac0a7ad7ad11b4af200936610db5166e58.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTop-1 ClsGT-known Loc
TS-CAMSCMOurs (+Δ)TS-CAMSCMOurs (+Δ)
Deit-T72.9-76.2 (+3.3)86.491.895.0 (+3.2)
Deit-S80.377.182.1 (+1.8)87.796.698.5 (+1.9)
Deit-B83.2-85.3 (+2.1)83.393.898.1 (+4.3)
Conf-S81.0-82.4 (+1.4)94.196.197.9 (+1.8)
", + "bbox": [ + 503, + 88, + 888, + 176 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dc3f914981c1d747ff68af0f5e7a730cabe85867ef56f3c20302b9433605db37.jpg", + "table_caption": [ + "Table 6. Generalizability on CUB-200. $\\Delta$ denotes the improvement of SAT over the previous best method. Conf-S: Conformer-S." + ], + "table_footnote": [], + "table_body": "
MethodLoc Acc. Top-1 Top-5 GT-k.
(a)ORNet73.3984.3188.38
(b)BAS49.5960.6863.76
(c)BAS*75.1888.6692.87
(d)Ours80.9694.1398.45
", + "bbox": [ + 501, + 222, + 702, + 314 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/776a76b26daeed25842af12890170a3ff43fe9aa6cbc491fba9a810e7035223e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Image(a)(b)
label(c)(d)
", + "bbox": [ + 714, + 219, + 890, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. SAT vs FPM-based methods on Deit-S backbone. For ORNet, we replace the classification network with Deit-S. For BAS, we insert the generator in the 10-th self-attention module. * denotes replacing the original losses with our proposed losses.", + "bbox": [ + 500, + 316, + 890, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "better results than other methods at arbitrary threshold. Besides, SAT is not sensitive to thresholds and maintains high localization accuracy over a large threshold interval. This phenomenon shows that the localization map generated by SAT has high confidence and excellent visualization.", + "bbox": [ + 496, + 398, + 890, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalizability. To validate the generalizability of the proposed SAT, we implement experiments on different Transformers and various scales. As reported in Table 6, the proposed approach can be well applied to Deit [24] and Conformer [19] on CUB-200 [26]. On Deit-T/Deit-B, we achieve $3.2\\% / 4.3\\%$ GT-known Loc improvement compared to SCM [2], and $3.3\\% / 2.1\\%$ Top-1 classification accuracy gain over TS-CAM [8]. Besides, SAT also achieves state-of-the-art results on Conformer-S. Quantitative experiments show that the proposed SAT achieves excellent performance in both tasks on various transformer backbones.", + "bbox": [ + 496, + 474, + 892, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Discussions", + "text_level": 1, + "bbox": [ + 500, + 651, + 624, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Why spatial token-based structure is critical? To clarify this question, we follow TS-CAM [8] to construct a attention map-based baseline and apply the proposed losses on it for comparison. As shown in Table 5, the attention map-based method substantially reduces the classification and localization accuracy, with $2.52\\%$ and $3.39\\%$ decrease in Top-1 Cls and GT-known Loc, respectively. To explore how this structural difference affects the accuracy, we plot the curves of spatial losses during training in Table 5. It can be noticed that the potential optimization conflicts of attention map-based method will lead to poor convergence of losses and insufficient learning, thus limiting the model to benefit from the losses. Experimental results show that the proposed spatial token-based structure can release the potential of the model by more efficient localization learn", + "bbox": [ + 496, + 672, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1850", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c85d13df1dffc9ca54dee64811002a48211e2aff89207f0e34c4924e6f631eca.jpg", + "image_caption": [ + "Figure 8. Few-shot learning. Box plot of the k-shot localization performance. For each k-shot setting, we repeat the experiment 10 times to randomly select k images as training data, and record the experimental results. The average results are shown in the figure." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 467, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing and avoiding optimization conflicts between two tasks, which are essential for the accuracy improvement.", + "bbox": [ + 75, + 311, + 467, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further verify the effectiveness of the proposed structure in generating localization map, we reproduce two FPM-based methods ORNet [33] and BAS [32] on the Deit-S. Both of them produce localization maps by a generator. As shown in Table 7, first, we note that the original BAS is not suitable for the transformer due to the skip-connection in transformer causing inaccurate evaluation of the masking effect. Second, both FPM-based methods suffer from incomplete localization problem. This may be because they can only mask the input of a specific layer, resulting in overfitting the features that the specific feature layer focuses on. While the proposed structure can extract localization knowledge from different layers in an efficient form, thus improving localization ability and robustness. Quantitative and qualitative results demonstrate the effectiveness of SAT in terms of localization map generation and learning.", + "bbox": [ + 75, + 342, + 467, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data-efficiency and tuning-efficiency. To validate the data-efficiency of the proposed method, we evaluate the localization performance of SAT and TS-CAM [8] under few-shot setting in Fig. 8. Compared to TS-CAM, SAT achieves promising results with only a few training data. Even in the extreme setting of only 1 training image per class, we exceed TS-CAM trained with full data by $3.6\\%$ and $3.3\\%$ GT-known Loc on CUB-200 and ImageNet, respectively. The box plot verifies the effectiveness and efficiency of the proposed SAT that implements localization task by an extra lightweight spatial token and enables high degree of network parameter sharing, thus requiring only a few data to achieve excellent localization results. In contrast, TS-CAM requires large amounts of data to fine-tune the classification feature maps to serve the localization task.", + "bbox": [ + 75, + 584, + 467, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To explore the tuning-efficiency of SAT, we freeze different parts of the network and evaluate the GT-known Loc of the proposed method under few-shot setting. All experiments are performed with ImageNet pre-trained weights. As shown in Table 8, SAT can still obtain competitive results when freezing most parameters of the pre-trained", + "bbox": [ + 75, + 810, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/26f862eae9458d6fff5f49c463e353d5ba759edc5b156a56475b9a33e1158a62.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Fro. rateCUB-200 [26]Fro. rateImageNet [20]
13510Full13510Full
(a)0%91.395.196.197.398.50%70.971.972.672.973.1
(b)24%93.595.996.897.598.121%71.672.372.672.873.1
(c)48%96.496.797.097.498.042%71.872.172.472.772.9
(d)71%93.894.694.995.295.964%71.371.471.771.972.2
(e)91%92.192.592.893.193.481%70.971.171.271.371.4
", + "bbox": [ + 501, + 88, + 890, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Frozen rate w.r.t. few-shot learning. The frozen parts are as follows: (a) None. (b) Attention layer of transformer blocks. (c) MLP layer of transformer blocks. (d) Transformer blocks. (e) Position embedding, projection, transformer blocks, and MLP layer of spatial aware transformer blocks. Fro.: Frozen.", + "bbox": [ + 498, + 195, + 890, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4923b6b770aa9f5021a52caad3ba2b6a0daaa2886edf1a48d652e787339a05b.jpg", + "image_caption": [ + "Figure 9. Localization error analysis on CUB-200." + ], + "image_footnote": [], + "bbox": [ + 501, + 273, + 890, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "classification model. It can be analyzed that the task-specific spatial token can be well adapted to the classification features and plays a major role in the localization learning, which allows efficient implementation of the localization task with fewer parameters to be fine-tuned. Experiments demonstrate both the data-efficiency and the tuning-efficiency of SAT. When freezing $81\\%$ of the parameters and using only 1 image per class, SAT exceeds the SOTA method SCM [2] by $2.1\\%$ GT-known Loc on ImageNet.", + "bbox": [ + 496, + 446, + 890, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 500, + 593, + 625, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 9, we count all localization errors (90 images) on CUB-200 test set (5,794 images) and classify the error causes into six categories, including object occlusion, localization more, water reflection, localization part, multiple instances, and label error. Among them, object occlusion causes the object to be split into two or more parts, resulting in incomplete localization results. Localization more is usually due to the positive effect of co-occurrence context on classification, leading to localizing confounding background regions. Water reflection is an inherent problem of WSOL and difficult to be solved with only image-level labels. Therefore, future work needs to consider more the interaction between object and context to overcome the problems of object occlusion and localization more.", + "bbox": [ + 496, + 617, + 890, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 844, + 617, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes to introduce a spatial-aware token (SAT) specific to the localization task, instead of synthesiz-", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1851", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing the classification feature map, thus avoiding optimization conflicts between classification and localization tasks. To this end, we construct a spatial-query attention (SQA) module to generate localization map and extract localization knowledge from the classification task. Meanwhile, two spatial constraints, including batch area loss and regularization loss, are designed to compensate for the insufficient supervision. Extensive experiments on multiple benchmarks verify the effectiveness and efficiency of the proposed SAT that surpasses previous methods by a large margin.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by National Key R&D Program of China under Grant 2020AAA0105700, National Natural Science Foundation of China (NSFC) under Grants 62225207, U19B2038 and 62121002.", + "bbox": [ + 75, + 242, + 472, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 330, + 174, + 345 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Wonho Bae, Junhyug Noh, and Gunhee Kim. Rethinking class activation mapping for weakly supervised object localization. In European Conference on Computer Vision, pages 618-634. Springer, 2020. 1", + "[2] Haotian Bai, Ruimao Zhang, Jiong Wang, and Xiang Wan. Weakly supervised object localization via transformer with implicit spatial calibration. arXiv preprint arXiv:2207.10447, 2022. 2, 3, 5, 6, 7, 8", + "[3] Zhiwei Chen, Changan Wang, Yabiao Wang, Guannan Jiang, Yunhang Shen, Ying Tai, Chengjie Wang, Wei Zhang, and Liujuan Cao. Lctr: On awakening the local continuity of transformer for weakly supervised object localization. arXiv preprint arXiv:2112.05291, 2021. 2, 3, 5", + "[4] Junsuk Choe, Seungho Lee, and Hyunjung Shim. Attention-based dropout layer for weakly supervised single object localization and semantic segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(12):4256-4271, 2020. 1, 2", + "[5] Junsuk Choe, Seong Joon Oh, Seungho Lee, Sanghyuk Chun, Zeynep Akata, and Hyunjung Shim. Evaluating weakly supervised object localization methods right. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3133-3142, 2020. 4, 5, 6, 7", + "[6] Junsuk Choe and Hyunjung Shim. Attention-based dropout layer for weakly supervised object localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2219-2228, 2019. 1", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2", + "[8] Wei Gao, Fang Wan, Xingjia Pan, Zhiliang Peng, Qi Tian, Zhenjun Han, Bolei Zhou, and Qixiang Ye. Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2886-2895, 2021. 1, 2, 4, 5, 6, 7, 8" + ], + "bbox": [ + 86, + 354, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Guangyu Guo, Junwei Han, Fang Wan, and Dingwen Zhang. Strengthen learning tolerance for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7403-7412, 2021. 1, 5", + "[10] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Amalgamating knowledge from heterogeneous graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15709-15718, 2021. 1", + "[11] Eunji Kim, Siwon Kim, Jungbeom Lee, Hyunwoo Kim, and Sungroh Yoon. Bridging the gap between classification and localization for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14258-14267, 2022. 5, 6", + "[12] Jeesoo Kim, Junsuk Choe, Sangdoo Yun, and Nojun Kwak. Normalization matters in weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3427-3436, 2021. 1", + "[13] Jungbeom Lee, Eunji Kim, Jisoo Mok, and Sungroh Yoon. Anti-adversarily manipulated attributions for weakly supervised semantic segmentation and object localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1", + "[14] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Dechao Meng, and Qingming Huang. Adaptive reconstruction network for weakly supervised referring expression grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2611-2620, 2019. 1", + "[15] Weizeng Lu, Xi Jia, Weicheng Xie, Linlin Shen, Yicong Zhou, and Jinming Duan. Geometry constrained weakly supervised object localization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 481-496. Springer, 2020. 5", + "[16] Jinjie Mai, Meng Yang, and Wenfeng Luo. Erasing integrated learning: A simple yet effective approach for weakly supervised object localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8766-8775, 2020. 1, 2", + "[17] Meng Meng, Tianzhu Zhang, Qi Tian, Yongdong Zhang, and Feng Wu. Foreground activation maps for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3385-3395, 2021. 2, 5", + "[18] Xingjia Pan, Yingguo Gao, Zhiwen Lin, Fan Tang, Weiming Dong, Haolei Yuan, Feiyue Huang, and Changsheng Xu. Unveiling the potential of structure preserving for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11642-11651, 2021. 1, 2, 5", + "[19] Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, and Qixiang Ye. Conformer: Local features coupling global representations for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 367-376, 2021. 7" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1852", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[20] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4, 5, 6, 8", + "[21] Tal Shaharabany, Yoad Tewel, and Lior Wolf. What is where by looking: Weakly-supervised open-world phrase-grounding without text inputs. arXiv preprint arXiv:2206.09358, 2022. 1", + "[22] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In 2017 IEEE international conference on computer vision (ICCV), pages 3544-3553. IEEE, 2017. 1, 2", + "[23] Ganchao Tan, Daqing Liu, Meng Wang, and Zheng-Jun Zha. Learning to discretely compose reasoning module networks for video captioning. arXiv preprint arXiv:2007.09049, 2020.1", + "[24] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 5, 6, 7", + "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[26] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 4, 5, 6, 7, 8", + "[27] Hao Wang, Zheng-Jun Zha, Xuejin Chen, Zhiwei Xiong, and Jiebo Luo. Dual path interaction network for video moment localization. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4116-4124, 2020. 1", + "[28] Hao Wang, Zheng-Jun Zha, Liang Li, Xuejin Chen, and Jiebo Luo. Semantic and relation modulation for audiovisual event localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1", + "[29] Hao Wang, Zheng-Jun Zha, Liang Li, Dong Liu, and Jiebo Luo. Structured multi-level interaction network for video moment localization via language query. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7026-7035, 2021. 1", + "[30] Jun Wei, Qin Wang, Zhen Li, Sheng Wang, S Kevin Zhou, and Shuguang Cui. Shallow feature matters for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5993-6001, 2021. 1, 2, 5", + "[31] Jun Wei, Sheng Wang, S Kevin Zhou, Shuguang Cui, and Zhen Li. Weakly supervised object localization through inter-class feature similarity and intra-class appearance consistency. In European Conference on Computer Vision, pages 195-210. Springer, 2022. 2, 5, 6", + "[32] Pingyu Wu, Wei Zhai, and Yang Cao. Background activation suppression for weakly supervised object localization. arXiv preprint arXiv:2112.00580, 2021. 2, 5, 6, 8", + "[33] Jinheng Xie, Cheng Luo, Xiangping Zhu, Ziqi Jin, Weizeng Lu, and Linlin Shen. Online refinement of low-level feature" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "based activation map for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 132-141, 2021. 2, 5, 8", + "[34] Jinheng Xie, Jianfeng Xiang, Junliang Chen, Xianxu Hou, Xiaodong Zhao, and Linlin Shen. Contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. arXiv preprint arXiv:2203.13505, 2022. 2", + "[35] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Rui-Wei Zhao, Tao Zhang, Xuequan Lu, and Shang Gao. Cream: Weakly supervised object localization via class re-activation mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9437–9446, 2022. 1, 5", + "[36] Haolan Xue, Chang Liu, Fang Wan, Jianbin Jiao, Xi-angyang Ji, and Qixiang Ye. Danet: Divergent activation for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6589–6598, 2019. 1, 6", + "[37] Seunghan Yang, Yoonhyung Kim, Youngeun Kim, and Changick Kim. Combinational class activation maps for weakly supervised object localization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2941-2949, 2020. 1", + "[38] Tianhao Yang, Zheng-Jun Zha, and Hanwang Zhang. Making history matter: History-advantage sequence training for visual dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2561–2569, 2019. 1", + "[39] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. Advances in neural information processing systems, 35:25739-25753, 2022. 1", + "[40] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6023-6032, 2019. 1", + "[41] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 1", + "[42] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10):2472-2500, 2022. 1", + "[43] Chen-Lin Zhang, Yun-Hao Cao, and Jianxin Wu. Rethinking the route towards weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13460-13469, 2020. 1, 2, 5", + "[44] Dingwen Zhang, Junwei Han, Gong Cheng, and Ming-Hsuan Yang. Weakly supervised object localization and detection: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021. 1", + "[45] Xiaolin Zhang, Yunchao Wei, Jiashi Feng, Yi Yang, and Thomas S Huang. Adversarial complementary learning for weakly supervised object localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1325-1334, 2018. 1, 2" + ], + "bbox": [ + 501, + 90, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1853", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Xiaolin Zhang, Yunchao Wei, Guoliang Kang, Yi Yang, and Thomas Huang. Self-produced guidance for weakly-supervised object localization. In Proceedings of the European conference on computer vision (ECCV), pages 597-613, 2018. 1, 2", + "[47] Xiaolin Zhang, Yunchao Wei, and Yi Yang. Inter-image communication for weakly supervised localization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 271-287. Springer, 2020. 1, 2", + "[48] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1, 2, 5", + "[49] Lei Zhu, Qian Chen, Lujia Jin, Yunfei You, and Yanye Lu. Bagging regional classification activation maps for weakly supervised object localization. arXiv preprint arXiv:2207.07818, 2022. 5, 6", + "[50] Lei Zhu, Qi She, Qian Chen, Yunfei You, Boyu Wang, and Yanye Lu. Weakly supervised object localization as domain adaption. arXiv preprint arXiv:2203.01714, 2022. 5, 6" + ], + "bbox": [ + 78, + 90, + 468, + 401 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1854", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Spatial-Aware Token for Weakly Supervised Object Localization/f549b4c7-58f5-4874-8f3f-2978b440c9a2_model.json b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/f549b4c7-58f5-4874-8f3f-2978b440c9a2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7f7d96a4797865defe5ff1fc1c38e4b8487b8b7d --- /dev/null +++ b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/f549b4c7-58f5-4874-8f3f-2978b440c9a2_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.159, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "Spatial-Aware Token for Weakly Supervised Object Localization" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.18, + 0.812, + 0.234 + ], + "angle": 0, + "content": "Pingyu Wu\\(^{1}\\), Wei Zhai\\(^{1,\\dagger}\\), Yang Cao\\(^{1,3}\\), Jiebo Luo\\(^{2}\\), Zheng-Jun Zha\\(^{1}\\) \n\\(^{1}\\) University of Science and Technology of China \n\\(^{2}\\) University of Rochester \n\\(^{3}\\) Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.237, + 0.744, + 0.253 + ], + "angle": 0, + "content": "wpy364755620@@mail., wzhai056@mail., forrest@}ustc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.255, + 0.672, + 0.27 + ], + "angle": 0, + "content": "jluo@cs.ROchester.edu zhazj@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.319, + 0.314, + 0.335 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.474, + 0.776 + ], + "angle": 0, + "content": "Weakly supervised object localization (WSOL) is a challenging task aiming to localize objects with only image-level supervision. Recent works apply visual transformer to WSOL and achieve significant success by exploiting the long-range feature dependency in self-attention mechanism. However, existing transformer-based methods synthesize the classification feature maps as the localization map, which leads to optimization conflicts between classification and localization tasks. To address this problem, we propose to learn a task-specific spatial-aware token (SAT) to condition localization in a weakly supervised manner. Specifically, a spatial token is first introduced in the input space to aggregate representations for localization task. Then a spatial aware attention module is constructed, which allows spatial token to generate foreground probabilities of different patches by querying and to extract localization knowledge from the classification task. Besides, for the problem of sparse and unbalanced pixel-level supervision obtained from the image-level label, two spatial constraints, including batch area loss and normalization loss, are designed to compensate and enhance this supervision. Experiments show that the proposed SAT achieves state-of-the-art performance on both CUB-200 and ImageNet, with \\(98.45\\%\\) and \\(73.13\\%\\) GT-known Loc, respectively. Even under the extreme setting of using only 1 image per class from ImageNet for training, SAT already exceeds the SOTA method by \\(2.1\\%\\) GT-known Loc. Code and models are available at https://github.com/wpy1999/SAT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.805, + 0.21, + 0.82 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.831, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Weakly supervised object localization (WSOL) aims to localize objects with only image-level labels available. Since no expensive bounding box or pixel-level annota" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.301, + 0.7, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.301, + 0.883, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.487, + 0.893, + 0.529 + ], + "angle": 0, + "content": "Figure 1. GT-known Loc on ImageNet. The proposed method only needs to fine-tune a small number of parameters or requires a few training images to achieve significant improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.614 + ], + "angle": 0, + "content": "tions are required, WSOL significantly reduces the cost of manual annotations [21, 44, 12, 13, 4, 43, 22, 45] and has attracted increasing attention in the research community [40, 42, 41, 39, 10, 27, 29, 28, 14, 38, 23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.75 + ], + "angle": 0, + "content": "As a representative work, CAM [48] extracts class activation maps from the classifier as localization maps. However, CAM is usually coarse and focuses on the most discriminative regions, leading to imprecise and incomplete localization results. To solve these problems, many CNN-based methods have been proposed, such as adversarial erasure [22, 45, 6, 16], divergent activation [36, 40, 22], seed region growing [30, 46], regularization [18, 13, 47], feature refining [1, 37, 30], regression-based [35, 43, 9]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Recently, transformer [7, 24] has been introduced to the field of computer vision with great success. Benefiting from the long-range feature dependency, the attention map upon the class token usually can capture the object region relevant to classification in the whole image, thus it is widely used for localization in the transformer-based WSOL methods. The pioneering work TS-CAM [8] accumulates attention maps of each layer and mixes them with semantic-aware maps to achieve localization task. Further, to address the problem that transformer lacks inherent spatial coherence" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1844" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "of the object, LCTR [3] and SCM [2] propose to consider cross-patch information and use activation diffusion to increase the local continuity of localization map, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.47, + 0.382 + ], + "angle": 0, + "content": "Existing transformer-based methods synthesize the feature maps learned by the classification task, such as attention maps, as the localization map, and try to increase its connectivity and completeness. However, this convenient approach results in optimization conflicts between classification and localization tasks. 1) For classification, making classification feature maps learn more object regions with low discrimination will reduce the classification ability. 2) For localization, the learning of the localization map is limited by the properties of the feature maps. For example, the attention map is generated by the softmax function, which is difficult to produce a balanced and comprehensive response over the object. Therefore, to achieve a more promising localization performance and avoid optimization conflicts, it is necessary to construct task-specific parameters for the generation and learning of localization map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Based on transformer architecture, a straightforward idea is to introduce a spatial token in the input space to aggregate global representation for the localization task. And the localization map can be further obtained by interacting the spatial token with each patch in the forward propagation. To this end, we propose a Spatial-Query Attention (SQA) module that takes the spatial token as a query to calculate the similarity with different patches and produces the localization map efficiently. Meanwhile, to make the spatial-aware token obtain localization supervision from image-level labels, the localization map is treated as a visual cue to participate in the calculation of cross attention, thus the generation and learning of localization map can be well adapted to the transformer-based classification model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.853 + ], + "angle": 0, + "content": "Building upon this efficient SQA module, we propose a simple but effective Spatial-Aware Token (SAT) approach, as shown in Fig. 2. Specifically, a task-specific spatial token is introduced in the input space. To achieve the localization task, several SQA modules are applied in the transformer to produce localization maps learned from different layers and aggregate them together. In this way, the final localization map can maximally capture the localization knowledge from the whole classification model. However, the pixel-level supervision generated by image-level labels is sparse and unbalanced. To compensate and strengthen this supervision, we propose two spatial constraints, including batch area loss and normalization loss. Batch area loss aims to provide a sparse area supervision with prior knowledge to compensate for the insufficient supervision. Normalization loss is employed to enhance pixel-level supervision by encouraging the localization map to be more discriminatory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Benefiting from the task-specific spatial token and avoiding optimization conflicts, SAT releases the potential of the model and achieves excellent performance in both classifi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "cation and localization tasks. Besides, unlike synthetic classification feature maps as localization results, the generation and learning of localization map in SAT rely mainly on an extra lightweight token, which brings advantages of data-efficiency and tuning-efficiency. As illustrated in Fig. 1, SAT outperforms SOTA method SCM by \\(2.1\\%\\) GT-known Loc with less than \\(0.1\\%\\) training data on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.198, + 0.862, + 0.212 + ], + "angle": 0, + "content": "In summary, the contributions of this paper include:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.272 + ], + "angle": 0, + "content": "1) We propose to learn a task-specific spatial token to implement the localization task, instead of synthesizing classification feature maps, thus avoiding optimization conflicts between classification and localization tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.348 + ], + "angle": 0, + "content": "2) We propose a simple but effective spatial-aware token (SAT) pipeline for WSOL, which utilizes a spatial-query attention (SQA) module to generate localization map through a spatial token and supervise it with two spatial constraints, including batch area loss and normalization loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.394 + ], + "angle": 0, + "content": "3) Extensive experiments show that SAT outperforms SOTA methods by a large margin on multiple benchmarks and performs excellently even under extreme settings." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.638, + 0.423 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.81 + ], + "angle": 0, + "content": "CNN-based methods. WSOL aims to learn the localization of objects with only image-level labels. CAM [48] proposes to synthesize depth feature maps with fully connected weights to obtain class activation maps (CAM). To increase the localization ability of CAM, HaS [22] adopts a random erasure strategy forcing the network to attend to different object regions. ACoL [45] and EIL [16] use adversarial erasing to achieve learning of complementary regions. ADL [4] drops the most discriminative or highlighted regions in the different layers during forward propagation. Besides, SPA [18] designs a post-process approach to extract structure-preserving localization map from the feature maps. SPG [46] and SPOL [30] facilitate learning of localization by selecting and spreading confidence regions as seeds. PSOL [43] and \\(\\mathrm{C}^2\\mathrm{AM}\\) [34] divide WSOL into independent classification task and class-agnostic object localization. I2C [47] and ISIC [31] consider feature similarities across different objects to achieve more complete and robust localization. ORNet [33], FAM [17] and BAS [32] propose to generate a foreground prediction map (FPM) to implement localization. Unlike FPM-based methods that require complex and heavy structures, and the learning of generator relies on a specific feature layer of the classification network. Our method simply and efficiently obtains localization maps from different layers by a task-specific token." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Transformer-based methods. Different from CNNs, transformer [7, 25] can capture global cues with the advantage of long-range dependency in the self-attention mechanism, effectively alleviating the partial activation problem. TS-CAM [8] mixes the attention maps of different layers and combines them with semantic-aware tokens to produce" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1845" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.086, + 0.891, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.893, + 0.309 + ], + "angle": 0, + "content": "Figure 2. Framework of SAT. It includes three spatial aware transformer blocks at the end of the network. Each block generates a localization map \\( M^l \\) using the spatial-query attention module. The final localization map \\( M \\) is obtained by fusing \\( M^l \\) of different layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.473, + 0.469 + ], + "angle": 0, + "content": "semantic-aware localization results. Based on TS-CAM, to address the lack of spatial consistency in transformer, LCTR [3] proposes to enhance the local perception capability among long-range feature dependencies by considering cross-patch information. SCM [2] increases the spatial coherence of attention maps by diffusing the semantic and spatial connections between patch tokens. Notably, they all generate localization map by synthesizing feature maps. Instead, we produce it by a task-specific spatial token." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.489, + 0.212, + 0.507 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.516, + 0.188, + 0.531 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.724 + ], + "angle": 0, + "content": "The overall network architecture of SAT is shown in Fig. 2. Given an input image \\( I \\in \\mathbb{R}^{h \\times w \\times 3} \\), a plain vision transformer splits it into a sequence of non-overlapping patches. Then the divided patches are flattened and transformed into patch tokens \\( \\{x_{n} \\in \\mathbb{R}^{1 \\times D}, n = 1, 2, \\dots, H \\times W\\} \\) by linear projection, where \\( H = h / P \\), \\( W = w / P \\), \\( P \\) is the patch size, and \\( D \\) is the number of channels. For simplicity, we omit the description of the batch size \\( B \\). After grouping the class token and the extra spatial-aware token \\( x_{spa} \\) with patch tokens, this token sequence \\( \\mathcal{F}^0 \\in \\mathbb{R}^{(N + 2) \\times D} \\) is sent into stacked transformer blocks and spatial aware transformer blocks for subsequent representation learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.47, + 0.803 + ], + "angle": 0, + "content": "The proposed spatial aware transformer block inherits the typical transformer block structure. Thus the entire network is characterized by alternating attention layers and MLP layers. For a given input token subsequent \\(\\mathcal{F}^{l-1}\\), the overall equations of the \\(l\\)-th block are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.818, + 0.469, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} ^ {l} = \\operatorname {L a y e r - N o r m} \\left(\\mathcal {F} ^ {l - 1}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.845, + 0.469, + 0.871 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Z} ^ {l} = \\mathcal {F} ^ {l - 1} + \\text {A t t e n t i o n} \\left(\\mathcal {X} ^ {l}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.873, + 0.469, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} ^ {l} = \\mathcal {Z} ^ {l} + \\operatorname {M L P} \\left(\\text {L a y e r - N o r m} \\left(\\mathcal {Z} ^ {l}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.796, + 0.348 + ], + "angle": 0, + "content": "3.2. Spatial Aware Transformer Block" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.446 + ], + "angle": 0, + "content": "The proposed spatial aware transformer block mainly contains a spatial-query attention (SQA) module to generate localization map and extract localization knowledge. Besides, two spatial constraints, including batch area loss \\(\\mathcal{L}_{ba}\\), and normalization loss \\(\\mathcal{L}_{norm}\\) are designed as the complement to the insufficient localization supervision." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.447, + 0.893, + 0.658 + ], + "angle": 0, + "content": "Spatial-query Attention. The proposed SQA module is based on self-attention module and aims at converting the spatial-aware token \\(x_{spa}\\) into a localization map \\(M_{spa}^{l}\\) by querying different patches. Meanwhile, the generated \\(M_{spa}^{l}\\) is applied to the cross-attention calculation in the form of dot product, thus obtaining localization supervision from the image-level labels. Specifically, the SQA module first linearly projects the token sequence \\(\\mathcal{X}^{l - 1}\\) to the query matrix \\(\\mathcal{Q}\\), key matrix \\(\\mathcal{K}\\), value matrix \\(\\mathcal{V}\\), where \\(\\mathcal{Q},\\mathcal{K},\\mathcal{V} \\in \\mathbb{R}^{(N + 2)\\times D}\\). Then the query vector corresponding to the spatial token \\(\\mathcal{Q}_{spa} \\in \\mathbb{R}^{1\\times D}\\) is selected from \\(\\mathcal{Q}\\) and used to query \\(\\mathcal{K}\\) to obtain the query results for each token. The similarity map between the spatial query vector and key matrix in the \\(l\\)-th layer is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.666, + 0.892, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} ^ {l} \\left(\\mathcal {Q} _ {s p a}, \\mathcal {K}\\right) = \\frac {\\mathcal {Q} _ {s p a} \\mathcal {K} ^ {\\mathrm {T}}}{\\sqrt {D}} \\in \\mathbb {R} ^ {1 \\times (N + 2)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.892, + 0.77 + ], + "angle": 0, + "content": "where \\(\\sqrt{D}\\) is employed as a scaling factor. After applying sigmoid activation, the similarity map is transformed into foreground probability \\(M_{pro}^{l}\\) with a 0 to 1 interval distribution. This process can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.775, + 0.892, + 0.796 + ], + "angle": 0, + "content": "\\[\nM _ {s p a} ^ {l} = \\operatorname {S i g m o i d} \\left(\\mathcal {S} ^ {l} \\left(\\mathcal {Q} _ {s p a}, \\mathcal {K}\\right)\\right) \\in \\mathbb {R} ^ {1 \\times (N + 2)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.892, + 0.865 + ], + "angle": 0, + "content": "The generated \\( M_{spa}^{l} \\) is served as a visual cue to point out the object region and combined in the calculation of attention with the form of dot product. In this way, the whole SQA module can be formulated in the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.87, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S Q A} \\left(\\mathcal {X} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q K} ^ {\\mathrm {T}}}{\\sqrt {D}}\\right) * M _ {s p a} ^ {l} \\mathcal {V}, \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1846" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.082, + 0.895, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.37, + 0.868, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Visualization comparison. The ground-truth bounding boxes are in red, and the predicted bounding boxes are in green." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.41, + 0.377, + 0.424 + ], + "angle": 0, + "content": "where \\(*\\) denotes element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.425, + 0.469, + 0.575 + ], + "angle": 0, + "content": "By cropping and reshaping, the part of the \\( M_{spa}^{l} \\) to the patch tokens can be transformed into a localization map \\( M^{l} \\in \\mathbb{R}^{H \\times W} \\). After obtaining \\( M^{l} \\) learned from different layers, we take their averages as the final localization map \\( M \\) to increase the completeness and robustness. Although this approach can maximally capture the localization information in the classification network, the pixel-level supervision obtained from image-level labels is sparse and unbalanced. To compensate and strengthen this supervision, we design batch area loss and normalization loss." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.576, + 0.469, + 0.666 + ], + "angle": 0, + "content": "Batch Area. To alleviate the insufficient localization supervision caused by the absence of dense labels in WSOL, we propose to use batch area loss as a complement to localization supervision. It aims to constrain the average area of localization maps \\(\\{M_b\\in \\mathbb{R}^{H\\times W},b = 1,2,\\dots,B\\}\\) within a batch \\(B\\) to a hyperparameter \\(\\lambda\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.671, + 0.469, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b a} = \\left| \\sum_ {b} ^ {B} \\sum_ {i} ^ {H} \\sum_ {j} ^ {W} \\left(\\lambda - \\frac {M _ {b} (i , j)}{B \\times H \\times W}\\right) \\right|, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.468, + 0.749 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a sparse area supervision with prior knowledge. The \\(\\lambda\\) is set to 0.25 and 0.35 on CUB-200 and ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Normalization. For unbalanced pixel-level supervision, a normalization loss is applied to enhance the pixel-level supervision based on the pixel response intensity. It encourages pixels to be more distinguishable about whether they are foreground or background, as presented by Eq. 8. Before penalizing the uncertainty value of the localization map \\( M \\), to further enhance its local consistency, we adopt a gaussian filter with a kernel size of \\( 3 \\times 3 \\) to establish spatial connections between adjacent patches. The standard deviation of gaussian kernel is set to 6 on all datasets. After" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.485 + ], + "angle": 0, + "content": "applying gaussian filtering, the value of each patch is redistributed according to the adjacent patch values, and \\(\\mathbf{M}\\) is converted to \\(M^{*}\\). By computing the normalization loss with \\(M^{*}\\) can further encourage \\(M\\) to have local consistency and avoid noise. It can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.496, + 0.892, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m}} = \\frac {1}{H \\times W} \\sum_ {i} ^ {H} \\sum_ {j} ^ {W} M ^ {*} (i, j) (1 - M ^ {*} (i, j)). \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.549, + 0.609, + 0.563 + ], + "angle": 0, + "content": "3.3. Total loss" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.573, + 0.892, + 0.648 + ], + "angle": 0, + "content": "We follow the design of classification head in the baseline method TS-CAM [8] to obtain the predicted probability \\( y \\) and the cross-entropy classification loss \\( L_{cls} \\). By jointly optimizing classification loss, batch area loss, and normalization loss, the total loss function is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.661, + 0.891, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {b a} + \\mathcal {L} _ {\\text {n o r m}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.689, + 0.89, + 0.704 + ], + "angle": 0, + "content": "where no extra hyperparameter is used to balance each loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.625, + 0.734 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.691, + 0.759 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate the proposed SAT on three popular datasets, including CUB-200 [26], ImageNet [20], OpenImages [5]. CUB-200 is a fine-grained dataset with 200 bird species, which consists of 5,994 images for training and 5,794 images for testing. ImageNet contains about 1.2 million training images and 50,000 validation images from 1,000 different classes. OpenImages contains 100 categories, it has 29,819, 2,500, and 5,000 samples in training, val, and test sets, respectively. Apart from the image-level" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1847" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.857, + 0.43 + ], + "angle": 0, + "content": "
MethodsVenueBackboneCUB-200 [26]Loc Acc.ImageNet [20]Loc Acc.
Top-1Top-5GT-knownTop-1Top-5GT-known
CAM [48]CVPR16VGG1641.0650.6655.1042.8054.8659.00
ORNet [33]ICCV21VGG1667.7380.7786.2052.0563.9468.27
BAS [32]CVPR22VGG1671.3385.3391.0752.9665.4169.64
Kim et al. [11]CVPR22VGG1670.8388.0793.1749.9463.2568.92
CREAM [35]CVPR22VGG1670.4485.6790.9852.3764.2068.32
GCNet [15]ECCV20InceptionV358.5871.0075.3049.0658.09-
SPA [18]CVPR21InceptionV353.5966.5072.1452.7364.2768.33
FAM [17]ICCV21InceptionV370.67-87.2555.24-68.62
CREAM [35]CVPR22InceptionV371.7686.3790.4356.0766.1969.03
BAS [32]CVPR22InceptionV373.2986.3192.2458.5169.0071.93
BagCAMs [49]ECCV22InceptionV360.07-89.7853.87-71.02
SPOL [30]CVPR21ResNet5080.1293.4496.4659.1467.1569.02
DA-WSOL [50]CVPR22ResNet5066.65-81.8355.84-70.27
BAS [32]CVPR22ResNet5077.2590.0895.1357.1868.4471.77
Kim et al. [11]CVPR22ResNet5073.1686.6891.6053.7665.7569.89
CREAM [35]CVPR22ResNet5076.03-89.8855.66-69.31
BagCAMs [49]ECCV22ResNet5069.67-94.0144.24-72.08
ISIC [31]ECCV22ResNet5080.6894.0897.3259.6167.8470.01
TS-CAM [8]ICCV21Deit-S71.3083.8087.7053.4064.3067.60
LCTR [3]AAAI22Deit-S79.2089.9092.4056.1065.8068.70
SCM [2]ECCV22Deit-S76.4091.6096.6056.1066.4068.80
SAT (ours)This WorkDeit-S80.9694.1398.4560.1570.5273.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.136, + 0.433, + 0.835, + 0.448 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods. The best results are highlighted in bold, second are underlined." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.474, + 0.345, + 0.566 + ], + "angle": 0, + "content": "
SettingLoc Acc.
LclsLbaLnorm F.Top-1Top-5GT-k.
(a)✓56.8666.6769.29
(b)✓✓58.2168.3671.07
(c)✓✓✓59.8670.2272.78
(d)✓✓✓✓60.1570.5273.13
" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.469, + 0.47, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.572, + 0.468, + 0.586 + ], + "angle": 0, + "content": "Table 2. Ablation study of SAT on ImageNet. \\( F \\) : Gaussian filter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.469, + 0.644 + ], + "angle": 0, + "content": "labels, both CUB-200 and OpenImages provide pixel-level mask labels, which are only used in the testing phase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Metrics. Following [32, 18, 50], for localization, we utilize GT-known localization accuracy (GT-known Loc), Top-1/Top5 localization accuracy (Top-1/Top-5 Loc), and maximal box accuracy (MaxBoxAccV2) [5] as evaluation metrics. GT-known Loc is correct indicating that the intersection over union (IoU) of the predicted bounding box and the ground-truth bounding box is \\(50\\%\\) or more. Top-1/Top-5 Loc is correct when the ground-truth class belongs to Top-1/Top-5 prediction categories and GT-known Loc is correct. For mask, the peak intersection over union (pIoU) [43] and pixel average precision (PxAP) [5] are adopted as metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. We evaluate our method on the Deit-S [24] pre-trained on ImageNet [20]. On CUB-200 [26], the training process lasts 30 epochs with a batch size of 256. For ImageNet [20], we train 5 epochs and set batch size to 512. In the training phase, the input images are resized to \\(256 \\times 256\\) and then randomly cropped to" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.469, + 0.695, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.469, + 0.892, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.567, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Figure 4. Hyperparameters. Sensitivity analysis of hyperparameters (a) \\(\\lambda\\) and (b) batch size in \\(\\mathcal{L}_{ba}\\) on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.622, + 0.892, + 0.668 + ], + "angle": 0, + "content": "\\(224 \\times 224\\). In the inference phase, following [32, 9, 43], we adopt ten crop augmentations to obtain classification results and replace random crop with center crop for localization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.654, + 0.696 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.764 + ], + "angle": 0, + "content": "In this subsection, we implement a series of ablation experiments with Deit-S [24] as the backbone. All experiments are conducted on ImageNet [20], a universal dataset, to increase the generalizability of the experimental results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Ablation studies of SAT components. Table 2 shows the localization accuracy of SAT with different compositions, where we retain the structure of SAT as the baseline. It can be noted that the proposed baseline can effectively activate the object region and already exceeds SOTA method SCM with only \\(\\mathcal{L}_{cls}\\). Based on the baseline, the addition of \\(\\mathcal{L}_{ba}\\) significantly improves the GT-known Loc with a \\(1.78\\%\\) gain, by limiting the expansion of the localization map and forcing it to focus on the regions relevant to the classifi-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1848" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.208 + ], + "angle": 0, + "content": "
MethodVenueBackboneResolutionCUB-200 [26]ImageNet [20]
δ = 0.3δ = 0.5δ = 0.7Meanδ = 0.3δ = 0.5δ = 0.7Mean
Kim et al. [11]CVPR22ResNet5028×2899.4090.4038.0075.9086.7071.1048.3068.70
BAS [32]CVPR22ResNet5028×2899.4195.1371.9288.8284.8071.7749.2568.61
TS-CAM [8]ICCV21DeiT-S14×1498.8887.7049.9378.8482.1167.6044.7664.82
SCM [2]ECCV22DeiT-S14×1499.6496.6071.7389.3283.6468.8045.3065.91
SAT (ours)This WorkDeiT-S14×1499.88 (+0.24)98.45 (+1.85)79.53 (+7.61)92.62 (+3.30)84.42 (-2.28)73.13 (+1.36)56.83 (+7.58)71.46 (+2.76)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.21, + 0.893, + 0.24 + ], + "angle": 0, + "content": "Table 3. Localization quality. Comparison of localization accuracy under different IoU thresholds \\((\\delta)\\). Best results are highlighted in bold, second are underlined. The last line indicates the improvement (decrease) of our method over the previous best method." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.255, + 0.273, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.255, + 0.47, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.377, + 0.441, + 0.392 + ], + "angle": 0, + "content": "Figure 5. Statistical analysis about localization quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.506 + ], + "angle": 0, + "content": "cation. Then the use of \\(\\mathcal{L}_{\\text {norm }}\\) can effectively increase the distinction between the foreground and background regions, resulting in a remarkable increase of \\(1.71\\%\\) GT-known Loc. Finally, applying gaussian filter can further eliminate noise and increase local connectivity, by considering the consistency of adjacent patches in the calculation of \\(\\mathcal{L}_{\\text {norm }}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.471, + 0.719 + ], + "angle": 0, + "content": "Hyperparameters. Fig. 4 presents the sensitivity of localization quality to the hyperparameters \\(\\lambda\\) and batch size in \\(\\mathcal{L}_{ba}\\) (Eq. 7). \\(\\lambda\\) denotes the constraint for the average area of the localization maps. A larger \\(\\lambda\\) encourages localization map to learn more regions relevant to the classification. Conversely, when \\(\\lambda\\) is small, the region learned by the localization map is small but reliable. Fig. 4 (a) shows that the model behaves insensitive to \\(\\lambda\\), since GT-known Loc remains stable over a large interval. Batch size affects the calculation of the average area of the localization map within a batch. A sufficiently large batch size can ensure tolerance of area variation between instances. As shown in Fig. 4 (b), when the batch size is larger than 64, the effect of batch on GT-known Loc is limited, with a change of less than \\(0.5\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.214, + 0.743 + ], + "angle": 0, + "content": "4.3. Performance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Main Results. Table 1 compares SAT with state-of-the-art methods on CUB-200 [26] and ImageNet [20]. Experiments show that SAT achieves stable and consistent improvements on both benchmark datasets and significantly outperforms all methods using various backbones. Compared to the CNN counterpart ResNet50-ISIC [31], which uses three separate networks, while we use only one network and surpass ResNet50-ISIC by \\(1.13\\%\\) and \\(3.12\\%\\) GT-known Loc on the CUB-200 and ImageNet, respectively. These quantitative results demonstrate the superi" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.262, + 0.895, + 0.38 + ], + "angle": 0, + "content": "
MethodVenueBackboneCUB-200OpenImages
pIoUPxAPpIoUPxAP
BagCAMs [49]ECCV22IncepV360.3481.4949.98
DA-WSOL [50]CVPR22ResNet5056.1874.7049.68
BAS [32]CVPR22ResNet5071.0489.2050.72
TS-CAM [8]ICCV21DeiT-S66.5581.4943.65
SCM [2]ECCV22DeiT-S72.0788.15-
SAT (ours)This workDeiT-S77.2189.8759.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.383, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Table 4. Segmentation quality of the localization map compared with other SOTA methods. IncepV3: InceptionV3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.656 + ], + "angle": 0, + "content": "ority of the proposed SAT in terms of simplicity and effectiveness. In addition, on CUB-200 [26], SAT achieves a remarkable performance of \\(80.96\\% / 98.45\\%\\) on Top-1/GT-known Loc, exceeding the baseline method TS-CAM [8] by \\(9.66\\% / 10.75\\%\\). On ImageNet [20], we surpass the SOTA method SCM [2] by a large margin, achieving a \\(4.33\\%\\) boost in GT-known Loc under the same Deit-S [24] backbone. To further demonstrate the effectiveness of SAT, we visualize the localization maps of our method and TS-CAM in Fig. 3. In addition to generating sharper and more complete localization maps than TS-CAM, SAT demonstrates robust localization ability in various complex and challenging scenarios. More accuracy results and visualizations on six datasets are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Localization Quality. In Table 3, we compare the localization performance with SOTA methods under the MaxBoxAccV2 [5] criterion. Despite the limitation of localization map resolution \\((14\\times 14)\\), we achieve excellent results under different IoU thresholds. Notably, at \\(\\delta\\) of 0.7, a strict criterion for the localization quality, the proposed SAT exceeds the previous best method by a significant margin, with \\(7.61\\%\\) and \\(7.58\\%\\) improvement on CUB-200 [26] and ImageNet [20], respectively. We further plot the IoU distribution of correct bounding boxes in Fig. 5, following DANet [36]. Compared to the latest SCM [2], with the same Deit-S [24] backbone, we obtain a median IoU gain of \\(2.9\\%\\) and \\(6.1\\%\\) on CUB-200 and ImageNet. The IoU distribution map demonstrates that the localization results generated by SAT have higher localization quality, especially on ImageNet [20], a universal and challenging dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1849" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.27, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.089, + 0.468, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.228, + 0.47, + 0.256 + ], + "angle": 0, + "content": "Figure 6. Density distribution map about IoU and object size on CUB-200, both measured using pixel-level masks." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.267, + 0.28, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.268, + 0.466, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.379, + 0.469, + 0.407 + ], + "angle": 0, + "content": "Figure 7. Threshold sensitivity. GT-known Loc-Threshold curves for different methods on CUB-200 and ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.421, + 0.33, + 0.497 + ], + "angle": 0, + "content": "
Localization GenerationCls Acc. Top-1 Top-5Loc Acc. Top-1 GT-k.
Attention map75.8992.7755.9869.74
Spatial token78.4194.4660.1573.13
+Δ+2.52+1.69+4.17+3.39
" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.421, + 0.469, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.501, + 0.469, + 0.556 + ], + "angle": 0, + "content": "Table 5. Accuracy of different baselines with the same losses. We follow the idea of TS-CAM to extract the attention maps (before softmax and adding a sigmoid function) to synthesize the localization map, and apply the proposed losses to it." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Segmentation Quality. We evaluate the segmentation performance of SAT on CUB-200 [26] and OpenImages [5] since pixel-level labels are available during evaluation. As shown in Table 4, the proposed SAT substantially exceeds all methods in pIoU and PxAP metrics. Especially compared to the baseline method TS-CAM [8], we achieve a remarkable increase of \\(10.66\\%\\) and \\(15.89\\%\\) pIoU on CUB-200 and OpenImages. To further analyze the segmentation quality, we draw the density distribution map about IoU and object size on CUB-200 in Fig. 6. It can be observed the proposed SAT achieves an overall improvement and more stable performance on arbitrary sized objects compared to TS-CAM. Nevertheless, there still exists an imbalance segmentation ability of different sized objects. This is mainly due to the limitation of the localization map resolution, which hinders the acquisition of fine segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Threshold Sensitivity. Fig. 7 illustrates the curves of GT-known Loc under different thresholds on CUB-200 and ImageNet. Obviously, the curves corresponding to the proposed SAT entirely cover the curves of TS-CAM [8] and SCM [2], indicating that SAT performs stable and achieves" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.089, + 0.89, + 0.178 + ], + "angle": 0, + "content": "
ModelTop-1 ClsGT-known Loc
TS-CAMSCMOurs (+Δ)TS-CAMSCMOurs (+Δ)
Deit-T72.9-76.2 (+3.3)86.491.895.0 (+3.2)
Deit-S80.377.182.1 (+1.8)87.796.698.5 (+1.9)
Deit-B83.2-85.3 (+2.1)83.393.898.1 (+4.3)
Conf-S81.0-82.4 (+1.4)94.196.197.9 (+1.8)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.181, + 0.892, + 0.21 + ], + "angle": 0, + "content": "Table 6. Generalizability on CUB-200. \\( \\Delta \\) denotes the improvement of SAT over the previous best method. Conf-S: Conformer-S." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.223, + 0.703, + 0.315 + ], + "angle": 0, + "content": "
MethodLoc Acc. Top-1 Top-5 GT-k.
(a)ORNet73.3984.3188.38
(b)BAS49.5960.6863.76
(c)BAS*75.1888.6692.87
(d)Ours80.9694.1398.45
" + }, + { + "type": "table", + "bbox": [ + 0.715, + 0.22, + 0.892, + 0.315 + ], + "angle": 0, + "content": "
Image(a)(b)
label(c)(d)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.318, + 0.892, + 0.374 + ], + "angle": 0, + "content": "Table 7. SAT vs FPM-based methods on Deit-S backbone. For ORNet, we replace the classification network with Deit-S. For BAS, we insert the generator in the 10-th self-attention module. * denotes replacing the original losses with our proposed losses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.474 + ], + "angle": 0, + "content": "better results than other methods at arbitrary threshold. Besides, SAT is not sensitive to thresholds and maintains high localization accuracy over a large threshold interval. This phenomenon shows that the localization map generated by SAT has high confidence and excellent visualization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.893, + 0.641 + ], + "angle": 0, + "content": "Generalizability. To validate the generalizability of the proposed SAT, we implement experiments on different Transformers and various scales. As reported in Table 6, the proposed approach can be well applied to Deit [24] and Conformer [19] on CUB-200 [26]. On Deit-T/Deit-B, we achieve \\(3.2\\% / 4.3\\%\\) GT-known Loc improvement compared to SCM [2], and \\(3.3\\% / 2.1\\%\\) Top-1 classification accuracy gain over TS-CAM [8]. Besides, SAT also achieves state-of-the-art results on Conformer-S. Quantitative experiments show that the proposed SAT achieves excellent performance in both tasks on various transformer backbones." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.652, + 0.625, + 0.666 + ], + "angle": 0, + "content": "4.4. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Why spatial token-based structure is critical? To clarify this question, we follow TS-CAM [8] to construct a attention map-based baseline and apply the proposed losses on it for comparison. As shown in Table 5, the attention map-based method substantially reduces the classification and localization accuracy, with \\(2.52\\%\\) and \\(3.39\\%\\) decrease in Top-1 Cls and GT-known Loc, respectively. To explore how this structural difference affects the accuracy, we plot the curves of spatial losses during training in Table 5. It can be noticed that the potential optimization conflicts of attention map-based method will lead to poor convergence of losses and insufficient learning, thus limiting the model to benefit from the losses. Experimental results show that the proposed spatial token-based structure can release the potential of the model by more efficient localization learn" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1850" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.077, + 0.087, + 0.468, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.47, + 0.287 + ], + "angle": 0, + "content": "Figure 8. Few-shot learning. Box plot of the k-shot localization performance. For each k-shot setting, we repeat the experiment 10 times to randomly select k images as training data, and record the experimental results. The average results are shown in the figure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.468, + 0.342 + ], + "angle": 0, + "content": "ing and avoiding optimization conflicts between two tasks, which are essential for the accuracy improvement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.468, + 0.584 + ], + "angle": 0, + "content": "To further verify the effectiveness of the proposed structure in generating localization map, we reproduce two FPM-based methods ORNet [33] and BAS [32] on the Deit-S. Both of them produce localization maps by a generator. As shown in Table 7, first, we note that the original BAS is not suitable for the transformer due to the skip-connection in transformer causing inaccurate evaluation of the masking effect. Second, both FPM-based methods suffer from incomplete localization problem. This may be because they can only mask the input of a specific layer, resulting in overfitting the features that the specific feature layer focuses on. While the proposed structure can extract localization knowledge from different layers in an efficient form, thus improving localization ability and robustness. Quantitative and qualitative results demonstrate the effectiveness of SAT in terms of localization map generation and learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.468, + 0.81 + ], + "angle": 0, + "content": "Data-efficiency and tuning-efficiency. To validate the data-efficiency of the proposed method, we evaluate the localization performance of SAT and TS-CAM [8] under few-shot setting in Fig. 8. Compared to TS-CAM, SAT achieves promising results with only a few training data. Even in the extreme setting of only 1 training image per class, we exceed TS-CAM trained with full data by \\(3.6\\%\\) and \\(3.3\\%\\) GT-known Loc on CUB-200 and ImageNet, respectively. The box plot verifies the effectiveness and efficiency of the proposed SAT that implements localization task by an extra lightweight spatial token and enables high degree of network parameter sharing, thus requiring only a few data to achieve excellent localization results. In contrast, TS-CAM requires large amounts of data to fine-tune the classification feature maps to serve the localization task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To explore the tuning-efficiency of SAT, we freeze different parts of the network and evaluate the GT-known Loc of the proposed method under few-shot setting. All experiments are performed with ImageNet pre-trained weights. As shown in Table 8, SAT can still obtain competitive results when freezing most parameters of the pre-trained" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.193 + ], + "angle": 0, + "content": "
Fro. rateCUB-200 [26]Fro. rateImageNet [20]
13510Full13510Full
(a)0%91.395.196.197.398.50%70.971.972.672.973.1
(b)24%93.595.996.897.598.121%71.672.372.672.873.1
(c)48%96.496.797.097.498.042%71.872.172.472.772.9
(d)71%93.894.694.995.295.964%71.371.471.771.972.2
(e)91%92.192.592.893.193.481%70.971.171.271.371.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.196, + 0.892, + 0.266 + ], + "angle": 0, + "content": "Table 8. Frozen rate w.r.t. few-shot learning. The frozen parts are as follows: (a) None. (b) Attention layer of transformer blocks. (c) MLP layer of transformer blocks. (d) Transformer blocks. (e) Position embedding, projection, transformer blocks, and MLP layer of spatial aware transformer blocks. Fro.: Frozen." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.275, + 0.892, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.408, + 0.851, + 0.421 + ], + "angle": 0, + "content": "Figure 9. Localization error analysis on CUB-200." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.584 + ], + "angle": 0, + "content": "classification model. It can be analyzed that the task-specific spatial token can be well adapted to the classification features and plays a major role in the localization learning, which allows efficient implementation of the localization task with fewer parameters to be fine-tuned. Experiments demonstrate both the data-efficiency and the tuning-efficiency of SAT. When freezing \\(81\\%\\) of the parameters and using only 1 image per class, SAT exceeds the SOTA method SCM [2] by \\(2.1\\%\\) GT-known Loc on ImageNet." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.594, + 0.627, + 0.608 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.83 + ], + "angle": 0, + "content": "As shown in Fig. 9, we count all localization errors (90 images) on CUB-200 test set (5,794 images) and classify the error causes into six categories, including object occlusion, localization more, water reflection, localization part, multiple instances, and label error. Among them, object occlusion causes the object to be split into two or more parts, resulting in incomplete localization results. Localization more is usually due to the positive effect of co-occurrence context on classification, leading to localizing confounding background regions. Water reflection is an inherent problem of WSOL and difficult to be solved with only image-level labels. Therefore, future work needs to consider more the interaction between object and context to overcome the problems of object occlusion and localization more." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.619, + 0.86 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper proposes to introduce a spatial-aware token (SAT) specific to the localization task, instead of synthesiz-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1851" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "ing the classification feature map, thus avoiding optimization conflicts between classification and localization tasks. To this end, we construct a spatial-query attention (SQA) module to generate localization map and extract localization knowledge from the classification task. Meanwhile, two spatial constraints, including batch area loss and regularization loss, are designed to compensate for the insufficient supervision. Extensive experiments on multiple benchmarks verify the effectiveness and efficiency of the proposed SAT that surpasses previous methods by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.318 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by National Key R&D Program of China under Grant 2020AAA0105700, National Natural Science Foundation of China (NSFC) under Grants 62225207, U19B2038 and 62121002." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.331, + 0.175, + 0.347 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.356, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[1] Wonho Bae, Junhyug Noh, and Gunhee Kim. Rethinking class activation mapping for weakly supervised object localization. In European Conference on Computer Vision, pages 618-634. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.413, + 0.472, + 0.467 + ], + "angle": 0, + "content": "[2] Haotian Bai, Ruimao Zhang, Jiong Wang, and Xiang Wan. Weakly supervised object localization via transformer with implicit spatial calibration. arXiv preprint arXiv:2207.10447, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.468, + 0.472, + 0.538 + ], + "angle": 0, + "content": "[3] Zhiwei Chen, Changan Wang, Yabiao Wang, Guannan Jiang, Yunhang Shen, Ying Tai, Chengjie Wang, Wei Zhang, and Liujuan Cao. Lctr: On awakening the local continuity of transformer for weakly supervised object localization. arXiv preprint arXiv:2112.05291, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.539, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[4] Junsuk Choe, Seungho Lee, and Hyunjung Shim. Attention-based dropout layer for weakly supervised single object localization and semantic segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(12):4256-4271, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.608, + 0.472, + 0.678 + ], + "angle": 0, + "content": "[5] Junsuk Choe, Seong Joon Oh, Seungho Lee, Sanghyuk Chun, Zeynep Akata, and Hyunjung Shim. Evaluating weakly supervised object localization methods right. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3133-3142, 2020. 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.679, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[6] Junsuk Choe and Hyunjung Shim. Attention-based dropout layer for weakly supervised object localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2219-2228, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.735, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.818, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[8] Wei Gao, Fang Wan, Xingjia Pan, Zhiliang Peng, Qi Tian, Zhenjun Han, Bolei Zhou, and Qixiang Ye. Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2886-2895, 2021. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.356, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[9] Guangyu Guo, Junwei Han, Fang Wan, and Dingwen Zhang. Strengthen learning tolerance for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7403-7412, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.895, + 0.234 + ], + "angle": 0, + "content": "[10] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Amalgamating knowledge from heterogeneous graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15709-15718, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.895, + 0.317 + ], + "angle": 0, + "content": "[11] Eunji Kim, Siwon Kim, Jungbeom Lee, Hyunwoo Kim, and Sungroh Yoon. Bridging the gap between classification and localization for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14258-14267, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.895, + 0.376 + ], + "angle": 0, + "content": "[12] Jeesoo Kim, Junsuk Choe, Sangdoo Yun, and Nojun Kwak. Normalization matters in weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3427-3436, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.895, + 0.447 + ], + "angle": 0, + "content": "[13] Jungbeom Lee, Eunji Kim, Jisoo Mok, and Sungroh Yoon. Anti-adversarily manipulated attributions for weakly supervised semantic segmentation and object localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.895, + 0.519 + ], + "angle": 0, + "content": "[14] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Dechao Meng, and Qingming Huang. Adaptive reconstruction network for weakly supervised referring expression grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2611-2620, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.895, + 0.603 + ], + "angle": 0, + "content": "[15] Weizeng Lu, Xi Jia, Weicheng Xie, Linlin Shen, Yicong Zhou, and Jinming Duan. Geometry constrained weakly supervised object localization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 481-496. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.895, + 0.675 + ], + "angle": 0, + "content": "[16] Jinjie Mai, Meng Yang, and Wenfeng Luo. Erasing integrated learning: A simple yet effective approach for weakly supervised object localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8766-8775, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.895, + 0.745 + ], + "angle": 0, + "content": "[17] Meng Meng, Tianzhu Zhang, Qi Tian, Yongdong Zhang, and Feng Wu. Foreground activation maps for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3385-3395, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.895, + 0.831 + ], + "angle": 0, + "content": "[18] Xingjia Pan, Yingguo Gao, Zhiwen Lin, Fan Tang, Weiming Dong, Haolei Yuan, Feiyue Huang, and Changsheng Xu. Unveiling the potential of structure preserving for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11642-11651, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[19] Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, and Qixiang Ye. Conformer: Local features coupling global representations for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 367-376, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1852" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[20] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[21] Tal Shaharabany, Yoad Tewel, and Lior Wolf. What is where by looking: Weakly-supervised open-world phrase-grounding without text inputs. arXiv preprint arXiv:2206.09358, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.472, + 0.287 + ], + "angle": 0, + "content": "[22] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In 2017 IEEE international conference on computer vision (ICCV), pages 3544-3553. IEEE, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[23] Ganchao Tan, Daqing Liu, Meng Wang, and Zheng-Jun Zha. Learning to discretely compose reasoning module networks for video captioning. arXiv preprint arXiv:2007.09049, 2020.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[24] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.472, + 0.469 + ], + "angle": 0, + "content": "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[26] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.51, + 0.472, + 0.567 + ], + "angle": 0, + "content": "[27] Hao Wang, Zheng-Jun Zha, Xuejin Chen, Zhiwei Xiong, and Jiebo Luo. Dual path interaction network for video moment localization. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4116-4124, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.472, + 0.622 + ], + "angle": 0, + "content": "[28] Hao Wang, Zheng-Jun Zha, Liang Li, Xuejin Chen, and Jiebo Luo. Semantic and relation modulation for audiovisual event localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.472, + 0.692 + ], + "angle": 0, + "content": "[29] Hao Wang, Zheng-Jun Zha, Liang Li, Dong Liu, and Jiebo Luo. Structured multi-level interaction network for video moment localization via language query. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7026-7035, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.472, + 0.762 + ], + "angle": 0, + "content": "[30] Jun Wei, Qin Wang, Zhen Li, Sheng Wang, S Kevin Zhou, and Shuguang Cui. Shallow feature matters for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5993-6001, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[31] Jun Wei, Sheng Wang, S Kevin Zhou, Shuguang Cui, and Zhen Li. Weakly supervised object localization through inter-class feature similarity and intra-class appearance consistency. In European Conference on Computer Vision, pages 195-210. Springer, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.472, + 0.873 + ], + "angle": 0, + "content": "[32] Pingyu Wu, Wei Zhai, and Yang Cao. Background activation suppression for weakly supervised object localization. arXiv preprint arXiv:2112.00580, 2021. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[33] Jinheng Xie, Cheng Luo, Xiangping Zhu, Ziqi Jin, Weizeng Lu, and Linlin Shen. Online refinement of low-level feature" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "based activation map for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 132-141, 2021. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[34] Jinheng Xie, Jianfeng Xiang, Junliang Chen, Xianxu Hou, Xiaodong Zhao, and Linlin Shen. Contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. arXiv preprint arXiv:2203.13505, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.203, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[35] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Rui-Wei Zhao, Tao Zhang, Xuequan Lu, and Shang Gao. Cream: Weakly supervised object localization via class re-activation mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9437–9446, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.287, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[36] Haolan Xue, Chang Liu, Fang Wan, Jianbin Jiao, Xi-angyang Ji, and Qixiang Ye. Danet: Divergent activation for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6589–6598, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[37] Seunghan Yang, Yoonhyung Kim, Youngeun Kim, and Changick Kim. Combinational class activation maps for weakly supervised object localization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2941-2949, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[38] Tianhao Yang, Zheng-Jun Zha, and Hanwang Zhang. Making history matter: History-advantage sequence training for visual dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2561–2569, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[39] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. Advances in neural information processing systems, 35:25739-25753, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[40] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6023-6032, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[41] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[42] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10):2472-2500, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[43] Chen-Lin Zhang, Yun-Hao Cao, and Jianxin Wu. Rethinking the route towards weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13460-13469, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[44] Dingwen Zhang, Junwei Han, Gong Cheng, and Ming-Hsuan Yang. Weakly supervised object localization and detection: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[45] Xiaolin Zhang, Yunchao Wei, Jiashi Feng, Yi Yang, and Thomas S Huang. Adversarial complementary learning for weakly supervised object localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1325-1334, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1853" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[46] Xiaolin Zhang, Yunchao Wei, Guoliang Kang, Yi Yang, and Thomas Huang. Self-produced guidance for weakly-supervised object localization. In Proceedings of the European conference on computer vision (ECCV), pages 597-613, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[47] Xiaolin Zhang, Yunchao Wei, and Yi Yang. Inter-image communication for weakly supervised localization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 271-287. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[48] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[49] Lei Zhu, Qian Chen, Lujia Jin, Yunfei You, and Yanye Lu. Bagging regional classification activation maps for weakly supervised object localization. arXiv preprint arXiv:2207.07818, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.36, + 0.469, + 0.402 + ], + "angle": 0, + "content": "[50] Lei Zhu, Qi She, Qian Chen, Yunfei You, Boyu Wang, and Yanye Lu. Weakly supervised object localization as domain adaption. arXiv preprint arXiv:2203.01714, 2022. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "1854" + } + ] +] \ No newline at end of file diff --git a/2023/Spatial-Aware Token for Weakly Supervised Object Localization/f549b4c7-58f5-4874-8f3f-2978b440c9a2_origin.pdf b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/f549b4c7-58f5-4874-8f3f-2978b440c9a2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a8a35e065a6ab5e5c5b384a6759fd24b7c54023e --- /dev/null +++ b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/f549b4c7-58f5-4874-8f3f-2978b440c9a2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb33ee2ead5e1492bc3c5b78f00ad4f881e31df9d1a9224545c5b11c3d1c8fc8 +size 2663324 diff --git a/2023/Spatial-Aware Token for Weakly Supervised Object Localization/full.md b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dfaa15d86075c00c95ff6d9673d154c112467cbe --- /dev/null +++ b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/full.md @@ -0,0 +1,319 @@ +# Spatial-Aware Token for Weakly Supervised Object Localization + +Pingyu Wu $^{1}$ , Wei Zhai $^{1,\dagger}$ , Yang Cao $^{1,3}$ , Jiebo Luo $^{2}$ , Zheng-Jun Zha $^{1}$ $^{1}$ University of Science and Technology of China + $^{2}$ University of Rochester + $^{3}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + +wpy364755620@@mail., wzhai056@mail., forrest@}ustc.edu.cn + +jluo@cs.ROchester.edu zhazj@ustc.edu.cn + +# Abstract + +Weakly supervised object localization (WSOL) is a challenging task aiming to localize objects with only image-level supervision. Recent works apply visual transformer to WSOL and achieve significant success by exploiting the long-range feature dependency in self-attention mechanism. However, existing transformer-based methods synthesize the classification feature maps as the localization map, which leads to optimization conflicts between classification and localization tasks. To address this problem, we propose to learn a task-specific spatial-aware token (SAT) to condition localization in a weakly supervised manner. Specifically, a spatial token is first introduced in the input space to aggregate representations for localization task. Then a spatial aware attention module is constructed, which allows spatial token to generate foreground probabilities of different patches by querying and to extract localization knowledge from the classification task. Besides, for the problem of sparse and unbalanced pixel-level supervision obtained from the image-level label, two spatial constraints, including batch area loss and normalization loss, are designed to compensate and enhance this supervision. Experiments show that the proposed SAT achieves state-of-the-art performance on both CUB-200 and ImageNet, with $98.45\%$ and $73.13\%$ GT-known Loc, respectively. Even under the extreme setting of using only 1 image per class from ImageNet for training, SAT already exceeds the SOTA method by $2.1\%$ GT-known Loc. Code and models are available at https://github.com/wpy1999/SAT. + +# 1. Introduction + +Weakly supervised object localization (WSOL) aims to localize objects with only image-level labels available. Since no expensive bounding box or pixel-level annota + +![](images/08b6b50cb1c4fdaed3f540bc09b82ffd93d6f644afd7cd9d4aebf7d540d41ec6.jpg) +Figure 1. GT-known Loc on ImageNet. The proposed method only needs to fine-tune a small number of parameters or requires a few training images to achieve significant improvement. + +![](images/6eaee214423d52cf6cbaee399e0e1d7db2238becda7ffbbb3e863d90d5b658de.jpg) + +tions are required, WSOL significantly reduces the cost of manual annotations [21, 44, 12, 13, 4, 43, 22, 45] and has attracted increasing attention in the research community [40, 42, 41, 39, 10, 27, 29, 28, 14, 38, 23]. + +As a representative work, CAM [48] extracts class activation maps from the classifier as localization maps. However, CAM is usually coarse and focuses on the most discriminative regions, leading to imprecise and incomplete localization results. To solve these problems, many CNN-based methods have been proposed, such as adversarial erasure [22, 45, 6, 16], divergent activation [36, 40, 22], seed region growing [30, 46], regularization [18, 13, 47], feature refining [1, 37, 30], regression-based [35, 43, 9]. + +Recently, transformer [7, 24] has been introduced to the field of computer vision with great success. Benefiting from the long-range feature dependency, the attention map upon the class token usually can capture the object region relevant to classification in the whole image, thus it is widely used for localization in the transformer-based WSOL methods. The pioneering work TS-CAM [8] accumulates attention maps of each layer and mixes them with semantic-aware maps to achieve localization task. Further, to address the problem that transformer lacks inherent spatial coherence + +of the object, LCTR [3] and SCM [2] propose to consider cross-patch information and use activation diffusion to increase the local continuity of localization map, respectively. + +Existing transformer-based methods synthesize the feature maps learned by the classification task, such as attention maps, as the localization map, and try to increase its connectivity and completeness. However, this convenient approach results in optimization conflicts between classification and localization tasks. 1) For classification, making classification feature maps learn more object regions with low discrimination will reduce the classification ability. 2) For localization, the learning of the localization map is limited by the properties of the feature maps. For example, the attention map is generated by the softmax function, which is difficult to produce a balanced and comprehensive response over the object. Therefore, to achieve a more promising localization performance and avoid optimization conflicts, it is necessary to construct task-specific parameters for the generation and learning of localization map. + +Based on transformer architecture, a straightforward idea is to introduce a spatial token in the input space to aggregate global representation for the localization task. And the localization map can be further obtained by interacting the spatial token with each patch in the forward propagation. To this end, we propose a Spatial-Query Attention (SQA) module that takes the spatial token as a query to calculate the similarity with different patches and produces the localization map efficiently. Meanwhile, to make the spatial-aware token obtain localization supervision from image-level labels, the localization map is treated as a visual cue to participate in the calculation of cross attention, thus the generation and learning of localization map can be well adapted to the transformer-based classification model. + +Building upon this efficient SQA module, we propose a simple but effective Spatial-Aware Token (SAT) approach, as shown in Fig. 2. Specifically, a task-specific spatial token is introduced in the input space. To achieve the localization task, several SQA modules are applied in the transformer to produce localization maps learned from different layers and aggregate them together. In this way, the final localization map can maximally capture the localization knowledge from the whole classification model. However, the pixel-level supervision generated by image-level labels is sparse and unbalanced. To compensate and strengthen this supervision, we propose two spatial constraints, including batch area loss and normalization loss. Batch area loss aims to provide a sparse area supervision with prior knowledge to compensate for the insufficient supervision. Normalization loss is employed to enhance pixel-level supervision by encouraging the localization map to be more discriminatory. + +Benefiting from the task-specific spatial token and avoiding optimization conflicts, SAT releases the potential of the model and achieves excellent performance in both classifi + +cation and localization tasks. Besides, unlike synthetic classification feature maps as localization results, the generation and learning of localization map in SAT rely mainly on an extra lightweight token, which brings advantages of data-efficiency and tuning-efficiency. As illustrated in Fig. 1, SAT outperforms SOTA method SCM by $2.1\%$ GT-known Loc with less than $0.1\%$ training data on ImageNet. + +In summary, the contributions of this paper include: + +1) We propose to learn a task-specific spatial token to implement the localization task, instead of synthesizing classification feature maps, thus avoiding optimization conflicts between classification and localization tasks. +2) We propose a simple but effective spatial-aware token (SAT) pipeline for WSOL, which utilizes a spatial-query attention (SQA) module to generate localization map through a spatial token and supervise it with two spatial constraints, including batch area loss and normalization loss. +3) Extensive experiments show that SAT outperforms SOTA methods by a large margin on multiple benchmarks and performs excellently even under extreme settings. + +# 2. Related work + +CNN-based methods. WSOL aims to learn the localization of objects with only image-level labels. CAM [48] proposes to synthesize depth feature maps with fully connected weights to obtain class activation maps (CAM). To increase the localization ability of CAM, HaS [22] adopts a random erasure strategy forcing the network to attend to different object regions. ACoL [45] and EIL [16] use adversarial erasing to achieve learning of complementary regions. ADL [4] drops the most discriminative or highlighted regions in the different layers during forward propagation. Besides, SPA [18] designs a post-process approach to extract structure-preserving localization map from the feature maps. SPG [46] and SPOL [30] facilitate learning of localization by selecting and spreading confidence regions as seeds. PSOL [43] and $\mathrm{C}^2\mathrm{AM}$ [34] divide WSOL into independent classification task and class-agnostic object localization. I2C [47] and ISIC [31] consider feature similarities across different objects to achieve more complete and robust localization. ORNet [33], FAM [17] and BAS [32] propose to generate a foreground prediction map (FPM) to implement localization. Unlike FPM-based methods that require complex and heavy structures, and the learning of generator relies on a specific feature layer of the classification network. Our method simply and efficiently obtains localization maps from different layers by a task-specific token. + +Transformer-based methods. Different from CNNs, transformer [7, 25] can capture global cues with the advantage of long-range dependency in the self-attention mechanism, effectively alleviating the partial activation problem. TS-CAM [8] mixes the attention maps of different layers and combines them with semantic-aware tokens to produce + +![](images/e24218d708a5a464ac22cd185e7c0ecadc2c19c3b02ffed34495bba8e516dd59.jpg) +Figure 2. Framework of SAT. It includes three spatial aware transformer blocks at the end of the network. Each block generates a localization map $M^l$ using the spatial-query attention module. The final localization map $M$ is obtained by fusing $M^l$ of different layers. + +semantic-aware localization results. Based on TS-CAM, to address the lack of spatial consistency in transformer, LCTR [3] proposes to enhance the local perception capability among long-range feature dependencies by considering cross-patch information. SCM [2] increases the spatial coherence of attention maps by diffusing the semantic and spatial connections between patch tokens. Notably, they all generate localization map by synthesizing feature maps. Instead, we produce it by a task-specific spatial token. + +# 3. Methodology + +# 3.1. Overview + +The overall network architecture of SAT is shown in Fig. 2. Given an input image $I \in \mathbb{R}^{h \times w \times 3}$ , a plain vision transformer splits it into a sequence of non-overlapping patches. Then the divided patches are flattened and transformed into patch tokens $\{x_{n} \in \mathbb{R}^{1 \times D}, n = 1, 2, \dots, H \times W\}$ by linear projection, where $H = h / P$ , $W = w / P$ , $P$ is the patch size, and $D$ is the number of channels. For simplicity, we omit the description of the batch size $B$ . After grouping the class token and the extra spatial-aware token $x_{spa}$ with patch tokens, this token sequence $\mathcal{F}^0 \in \mathbb{R}^{(N + 2) \times D}$ is sent into stacked transformer blocks and spatial aware transformer blocks for subsequent representation learning. + +The proposed spatial aware transformer block inherits the typical transformer block structure. Thus the entire network is characterized by alternating attention layers and MLP layers. For a given input token subsequent $\mathcal{F}^{l-1}$ , the overall equations of the $l$ -th block are defined as follows: + +$$ +\mathcal {X} ^ {l} = \operatorname {L a y e r - N o r m} \left(\mathcal {F} ^ {l - 1}\right), \tag {1} +$$ + +$$ +\mathcal {Z} ^ {l} = \mathcal {F} ^ {l - 1} + \text {A t t e n t i o n} \left(\mathcal {X} ^ {l}\right), \tag {2} +$$ + +$$ +\mathcal {F} ^ {l} = \mathcal {Z} ^ {l} + \operatorname {M L P} \left(\text {L a y e r - N o r m} \left(\mathcal {Z} ^ {l}\right)\right). \tag {3} +$$ + +# 3.2. Spatial Aware Transformer Block + +The proposed spatial aware transformer block mainly contains a spatial-query attention (SQA) module to generate localization map and extract localization knowledge. Besides, two spatial constraints, including batch area loss $\mathcal{L}_{ba}$ , and normalization loss $\mathcal{L}_{norm}$ are designed as the complement to the insufficient localization supervision. + +Spatial-query Attention. The proposed SQA module is based on self-attention module and aims at converting the spatial-aware token $x_{spa}$ into a localization map $M_{spa}^{l}$ by querying different patches. Meanwhile, the generated $M_{spa}^{l}$ is applied to the cross-attention calculation in the form of dot product, thus obtaining localization supervision from the image-level labels. Specifically, the SQA module first linearly projects the token sequence $\mathcal{X}^{l - 1}$ to the query matrix $\mathcal{Q}$ , key matrix $\mathcal{K}$ , value matrix $\mathcal{V}$ , where $\mathcal{Q},\mathcal{K},\mathcal{V} \in \mathbb{R}^{(N + 2)\times D}$ . Then the query vector corresponding to the spatial token $\mathcal{Q}_{spa} \in \mathbb{R}^{1\times D}$ is selected from $\mathcal{Q}$ and used to query $\mathcal{K}$ to obtain the query results for each token. The similarity map between the spatial query vector and key matrix in the $l$ -th layer is calculated as follows: + +$$ +\mathcal {S} ^ {l} \left(\mathcal {Q} _ {s p a}, \mathcal {K}\right) = \frac {\mathcal {Q} _ {s p a} \mathcal {K} ^ {\mathrm {T}}}{\sqrt {D}} \in \mathbb {R} ^ {1 \times (N + 2)}, \tag {4} +$$ + +where $\sqrt{D}$ is employed as a scaling factor. After applying sigmoid activation, the similarity map is transformed into foreground probability $M_{pro}^{l}$ with a 0 to 1 interval distribution. This process can be expressed as follows: + +$$ +M _ {s p a} ^ {l} = \operatorname {S i g m o i d} \left(\mathcal {S} ^ {l} \left(\mathcal {Q} _ {s p a}, \mathcal {K}\right)\right) \in \mathbb {R} ^ {1 \times (N + 2)}. \tag {5} +$$ + +The generated $M_{spa}^{l}$ is served as a visual cue to point out the object region and combined in the calculation of attention with the form of dot product. In this way, the whole SQA module can be formulated in the following form: + +$$ +\operatorname {S Q A} \left(\mathcal {X} ^ {l}\right) = \operatorname {S o f t m a x} \left(\frac {\mathcal {Q K} ^ {\mathrm {T}}}{\sqrt {D}}\right) * M _ {s p a} ^ {l} \mathcal {V}, \tag {6} +$$ + +![](images/86faad96805cfacef49e0f7948004b07a400a3044fe353a45fc7dbb526495375.jpg) +Figure 3. Visualization comparison. The ground-truth bounding boxes are in red, and the predicted bounding boxes are in green. + +where $*$ denotes element-wise multiplication. + +By cropping and reshaping, the part of the $M_{spa}^{l}$ to the patch tokens can be transformed into a localization map $M^{l} \in \mathbb{R}^{H \times W}$ . After obtaining $M^{l}$ learned from different layers, we take their averages as the final localization map $M$ to increase the completeness and robustness. Although this approach can maximally capture the localization information in the classification network, the pixel-level supervision obtained from image-level labels is sparse and unbalanced. To compensate and strengthen this supervision, we design batch area loss and normalization loss. + +Batch Area. To alleviate the insufficient localization supervision caused by the absence of dense labels in WSOL, we propose to use batch area loss as a complement to localization supervision. It aims to constrain the average area of localization maps $\{M_b\in \mathbb{R}^{H\times W},b = 1,2,\dots,B\}$ within a batch $B$ to a hyperparameter $\lambda$ , as follows: + +$$ +\mathcal {L} _ {b a} = \left| \sum_ {b} ^ {B} \sum_ {i} ^ {H} \sum_ {j} ^ {W} \left(\lambda - \frac {M _ {b} (i , j)}{B \times H \times W}\right) \right|, \tag {7} +$$ + +where $\lambda$ is a sparse area supervision with prior knowledge. The $\lambda$ is set to 0.25 and 0.35 on CUB-200 and ImageNet. + +Normalization. For unbalanced pixel-level supervision, a normalization loss is applied to enhance the pixel-level supervision based on the pixel response intensity. It encourages pixels to be more distinguishable about whether they are foreground or background, as presented by Eq. 8. Before penalizing the uncertainty value of the localization map $M$ , to further enhance its local consistency, we adopt a gaussian filter with a kernel size of $3 \times 3$ to establish spatial connections between adjacent patches. The standard deviation of gaussian kernel is set to 6 on all datasets. After + +applying gaussian filtering, the value of each patch is redistributed according to the adjacent patch values, and $\mathbf{M}$ is converted to $M^{*}$ . By computing the normalization loss with $M^{*}$ can further encourage $M$ to have local consistency and avoid noise. It can be expressed as follows: + +$$ +\mathcal {L} _ {\text {n o r m}} = \frac {1}{H \times W} \sum_ {i} ^ {H} \sum_ {j} ^ {W} M ^ {*} (i, j) (1 - M ^ {*} (i, j)). \tag {8} +$$ + +# 3.3. Total loss + +We follow the design of classification head in the baseline method TS-CAM [8] to obtain the predicted probability $y$ and the cross-entropy classification loss $L_{cls}$ . By jointly optimizing classification loss, batch area loss, and normalization loss, the total loss function is defined as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {c l s} + \mathcal {L} _ {b a} + \mathcal {L} _ {\text {n o r m}}, \tag {9} +$$ + +where no extra hyperparameter is used to balance each loss. + +# 4. Experiment + +# 4.1. Experimental Setup + +Datasets. We evaluate the proposed SAT on three popular datasets, including CUB-200 [26], ImageNet [20], OpenImages [5]. CUB-200 is a fine-grained dataset with 200 bird species, which consists of 5,994 images for training and 5,794 images for testing. ImageNet contains about 1.2 million training images and 50,000 validation images from 1,000 different classes. OpenImages contains 100 categories, it has 29,819, 2,500, and 5,000 samples in training, val, and test sets, respectively. Apart from the image-level + +
MethodsVenueBackboneCUB-200 [26]Loc Acc.ImageNet [20]Loc Acc.
Top-1Top-5GT-knownTop-1Top-5GT-known
CAM [48]CVPR16VGG1641.0650.6655.1042.8054.8659.00
ORNet [33]ICCV21VGG1667.7380.7786.2052.0563.9468.27
BAS [32]CVPR22VGG1671.3385.3391.0752.9665.4169.64
Kim et al. [11]CVPR22VGG1670.8388.0793.1749.9463.2568.92
CREAM [35]CVPR22VGG1670.4485.6790.9852.3764.2068.32
GCNet [15]ECCV20InceptionV358.5871.0075.3049.0658.09-
SPA [18]CVPR21InceptionV353.5966.5072.1452.7364.2768.33
FAM [17]ICCV21InceptionV370.67-87.2555.24-68.62
CREAM [35]CVPR22InceptionV371.7686.3790.4356.0766.1969.03
BAS [32]CVPR22InceptionV373.2986.3192.2458.5169.0071.93
BagCAMs [49]ECCV22InceptionV360.07-89.7853.87-71.02
SPOL [30]CVPR21ResNet5080.1293.4496.4659.1467.1569.02
DA-WSOL [50]CVPR22ResNet5066.65-81.8355.84-70.27
BAS [32]CVPR22ResNet5077.2590.0895.1357.1868.4471.77
Kim et al. [11]CVPR22ResNet5073.1686.6891.6053.7665.7569.89
CREAM [35]CVPR22ResNet5076.03-89.8855.66-69.31
BagCAMs [49]ECCV22ResNet5069.67-94.0144.24-72.08
ISIC [31]ECCV22ResNet5080.6894.0897.3259.6167.8470.01
TS-CAM [8]ICCV21Deit-S71.3083.8087.7053.4064.3067.60
LCTR [3]AAAI22Deit-S79.2089.9092.4056.1065.8068.70
SCM [2]ECCV22Deit-S76.4091.6096.6056.1066.4068.80
SAT (ours)This WorkDeit-S80.9694.1398.4560.1570.5273.13
+ +Table 1. Comparison with state-of-the-art methods. The best results are highlighted in bold, second are underlined. + +
SettingLoc Acc.
LclsLbaLnorm F.Top-1Top-5GT-k.
(a)✓56.8666.6769.29
(b)✓✓58.2168.3671.07
(c)✓✓✓59.8670.2272.78
(d)✓✓✓✓60.1570.5273.13
+ +![](images/11bee27d3e8496a83fc8c0c6dd7175be542f69f71cced9cb164379896020fb31.jpg) + +Table 2. Ablation study of SAT on ImageNet. $F$ : Gaussian filter. + +labels, both CUB-200 and OpenImages provide pixel-level mask labels, which are only used in the testing phase. + +Metrics. Following [32, 18, 50], for localization, we utilize GT-known localization accuracy (GT-known Loc), Top-1/Top5 localization accuracy (Top-1/Top-5 Loc), and maximal box accuracy (MaxBoxAccV2) [5] as evaluation metrics. GT-known Loc is correct indicating that the intersection over union (IoU) of the predicted bounding box and the ground-truth bounding box is $50\%$ or more. Top-1/Top-5 Loc is correct when the ground-truth class belongs to Top-1/Top-5 prediction categories and GT-known Loc is correct. For mask, the peak intersection over union (pIoU) [43] and pixel average precision (PxAP) [5] are adopted as metrics. + +Implementation Details. We evaluate our method on the Deit-S [24] pre-trained on ImageNet [20]. On CUB-200 [26], the training process lasts 30 epochs with a batch size of 256. For ImageNet [20], we train 5 epochs and set batch size to 512. In the training phase, the input images are resized to $256 \times 256$ and then randomly cropped to + +![](images/a2005113709f5d946aa16ac29d6d23f36103935fda719c1897568f9e0c302385.jpg) +Figure 4. Hyperparameters. Sensitivity analysis of hyperparameters (a) $\lambda$ and (b) batch size in $\mathcal{L}_{ba}$ on ImageNet. + +![](images/556a46dcf2f9f9fad1c2ea261e194bf2374c78cab75032d2483ba4ff262a668c.jpg) + +$224 \times 224$ . In the inference phase, following [32, 9, 43], we adopt ten crop augmentations to obtain classification results and replace random crop with center crop for localization. + +# 4.2. Ablation Study + +In this subsection, we implement a series of ablation experiments with Deit-S [24] as the backbone. All experiments are conducted on ImageNet [20], a universal dataset, to increase the generalizability of the experimental results. + +Ablation studies of SAT components. Table 2 shows the localization accuracy of SAT with different compositions, where we retain the structure of SAT as the baseline. It can be noted that the proposed baseline can effectively activate the object region and already exceeds SOTA method SCM with only $\mathcal{L}_{cls}$ . Based on the baseline, the addition of $\mathcal{L}_{ba}$ significantly improves the GT-known Loc with a $1.78\%$ gain, by limiting the expansion of the localization map and forcing it to focus on the regions relevant to the classifi- + +
MethodVenueBackboneResolutionCUB-200 [26]ImageNet [20]
δ = 0.3δ = 0.5δ = 0.7Meanδ = 0.3δ = 0.5δ = 0.7Mean
Kim et al. [11]CVPR22ResNet5028×2899.4090.4038.0075.9086.7071.1048.3068.70
BAS [32]CVPR22ResNet5028×2899.4195.1371.9288.8284.8071.7749.2568.61
TS-CAM [8]ICCV21DeiT-S14×1498.8887.7049.9378.8482.1167.6044.7664.82
SCM [2]ECCV22DeiT-S14×1499.6496.6071.7389.3283.6468.8045.3065.91
SAT (ours)This WorkDeiT-S14×1499.88 (+0.24)98.45 (+1.85)79.53 (+7.61)92.62 (+3.30)84.42 (-2.28)73.13 (+1.36)56.83 (+7.58)71.46 (+2.76)
+ +![](images/4bf90063eb1c7071935a4b480656b2d3a359f2e399d738ddabc93c748fe865fb.jpg) +Figure 5. Statistical analysis about localization quality. + +![](images/bf14a74d5c0a95193dce7d766148632a5156bcaed8014be42221bbc31a93c9b2.jpg) + +cation. Then the use of $\mathcal{L}_{\text {norm }}$ can effectively increase the distinction between the foreground and background regions, resulting in a remarkable increase of $1.71\%$ GT-known Loc. Finally, applying gaussian filter can further eliminate noise and increase local connectivity, by considering the consistency of adjacent patches in the calculation of $\mathcal{L}_{\text {norm }}$ . + +Hyperparameters. Fig. 4 presents the sensitivity of localization quality to the hyperparameters $\lambda$ and batch size in $\mathcal{L}_{ba}$ (Eq. 7). $\lambda$ denotes the constraint for the average area of the localization maps. A larger $\lambda$ encourages localization map to learn more regions relevant to the classification. Conversely, when $\lambda$ is small, the region learned by the localization map is small but reliable. Fig. 4 (a) shows that the model behaves insensitive to $\lambda$ , since GT-known Loc remains stable over a large interval. Batch size affects the calculation of the average area of the localization map within a batch. A sufficiently large batch size can ensure tolerance of area variation between instances. As shown in Fig. 4 (b), when the batch size is larger than 64, the effect of batch on GT-known Loc is limited, with a change of less than $0.5\%$ . + +# 4.3. Performance + +Main Results. Table 1 compares SAT with state-of-the-art methods on CUB-200 [26] and ImageNet [20]. Experiments show that SAT achieves stable and consistent improvements on both benchmark datasets and significantly outperforms all methods using various backbones. Compared to the CNN counterpart ResNet50-ISIC [31], which uses three separate networks, while we use only one network and surpass ResNet50-ISIC by $1.13\%$ and $3.12\%$ GT-known Loc on the CUB-200 and ImageNet, respectively. These quantitative results demonstrate the superi + +Table 3. Localization quality. Comparison of localization accuracy under different IoU thresholds $(\delta)$ . Best results are highlighted in bold, second are underlined. The last line indicates the improvement (decrease) of our method over the previous best method. + +
MethodVenueBackboneCUB-200OpenImages
pIoUPxAPpIoUPxAP
BagCAMs [49]ECCV22IncepV360.3481.4949.98
DA-WSOL [50]CVPR22ResNet5056.1874.7049.68
BAS [32]CVPR22ResNet5071.0489.2050.72
TS-CAM [8]ICCV21DeiT-S66.5581.4943.65
SCM [2]ECCV22DeiT-S72.0788.15-
SAT (ours)This workDeiT-S77.2189.8759.54
+ +Table 4. Segmentation quality of the localization map compared with other SOTA methods. IncepV3: InceptionV3. + +ority of the proposed SAT in terms of simplicity and effectiveness. In addition, on CUB-200 [26], SAT achieves a remarkable performance of $80.96\% / 98.45\%$ on Top-1/GT-known Loc, exceeding the baseline method TS-CAM [8] by $9.66\% / 10.75\%$ . On ImageNet [20], we surpass the SOTA method SCM [2] by a large margin, achieving a $4.33\%$ boost in GT-known Loc under the same Deit-S [24] backbone. To further demonstrate the effectiveness of SAT, we visualize the localization maps of our method and TS-CAM in Fig. 3. In addition to generating sharper and more complete localization maps than TS-CAM, SAT demonstrates robust localization ability in various complex and challenging scenarios. More accuracy results and visualizations on six datasets are provided in the supplementary materials. + +Localization Quality. In Table 3, we compare the localization performance with SOTA methods under the MaxBoxAccV2 [5] criterion. Despite the limitation of localization map resolution $(14\times 14)$ , we achieve excellent results under different IoU thresholds. Notably, at $\delta$ of 0.7, a strict criterion for the localization quality, the proposed SAT exceeds the previous best method by a significant margin, with $7.61\%$ and $7.58\%$ improvement on CUB-200 [26] and ImageNet [20], respectively. We further plot the IoU distribution of correct bounding boxes in Fig. 5, following DANet [36]. Compared to the latest SCM [2], with the same Deit-S [24] backbone, we obtain a median IoU gain of $2.9\%$ and $6.1\%$ on CUB-200 and ImageNet. The IoU distribution map demonstrates that the localization results generated by SAT have higher localization quality, especially on ImageNet [20], a universal and challenging dataset. + +![](images/ae691c9f8a77d3f5f7f0440d1677ac92a471fc3f810d56e39e89f45fcbc46cc8.jpg) +Figure 6. Density distribution map about IoU and object size on CUB-200, both measured using pixel-level masks. + +![](images/ac92df289a7ca82909855dcfb79c63d5d949bd5d3eef5fe59dfdfdce5e16b3e9.jpg) + +![](images/038fdcc0823c5a05e279d7e959d1e4fe346096fae3d1e3994a07bc9c90c917a8.jpg) +Figure 7. Threshold sensitivity. GT-known Loc-Threshold curves for different methods on CUB-200 and ImageNet. + +![](images/a8c7d6ee241cba9dacb0549531af32aa35ffa9aa6a38b2c9a95cd308b66d8f36.jpg) + +
Localization GenerationCls Acc. Top-1 Top-5Loc Acc. Top-1 GT-k.
Attention map75.8992.7755.9869.74
Spatial token78.4194.4660.1573.13
+Δ+2.52+1.69+4.17+3.39
+ +![](images/86c419755180c6a119cce9b7893a0507d0f8a8407588a56f719be568b0ae8530.jpg) +Table 5. Accuracy of different baselines with the same losses. We follow the idea of TS-CAM to extract the attention maps (before softmax and adding a sigmoid function) to synthesize the localization map, and apply the proposed losses to it. + +Segmentation Quality. We evaluate the segmentation performance of SAT on CUB-200 [26] and OpenImages [5] since pixel-level labels are available during evaluation. As shown in Table 4, the proposed SAT substantially exceeds all methods in pIoU and PxAP metrics. Especially compared to the baseline method TS-CAM [8], we achieve a remarkable increase of $10.66\%$ and $15.89\%$ pIoU on CUB-200 and OpenImages. To further analyze the segmentation quality, we draw the density distribution map about IoU and object size on CUB-200 in Fig. 6. It can be observed the proposed SAT achieves an overall improvement and more stable performance on arbitrary sized objects compared to TS-CAM. Nevertheless, there still exists an imbalance segmentation ability of different sized objects. This is mainly due to the limitation of the localization map resolution, which hinders the acquisition of fine segmentation results. + +Threshold Sensitivity. Fig. 7 illustrates the curves of GT-known Loc under different thresholds on CUB-200 and ImageNet. Obviously, the curves corresponding to the proposed SAT entirely cover the curves of TS-CAM [8] and SCM [2], indicating that SAT performs stable and achieves + +
ModelTop-1 ClsGT-known Loc
TS-CAMSCMOurs (+Δ)TS-CAMSCMOurs (+Δ)
Deit-T72.9-76.2 (+3.3)86.491.895.0 (+3.2)
Deit-S80.377.182.1 (+1.8)87.796.698.5 (+1.9)
Deit-B83.2-85.3 (+2.1)83.393.898.1 (+4.3)
Conf-S81.0-82.4 (+1.4)94.196.197.9 (+1.8)
+ +Table 6. Generalizability on CUB-200. $\Delta$ denotes the improvement of SAT over the previous best method. Conf-S: Conformer-S. + +
MethodLoc Acc. Top-1 Top-5 GT-k.
(a)ORNet73.3984.3188.38
(b)BAS49.5960.6863.76
(c)BAS*75.1888.6692.87
(d)Ours80.9694.1398.45
+ +
Image(a)(b)
label(c)(d)
+ +Table 7. SAT vs FPM-based methods on Deit-S backbone. For ORNet, we replace the classification network with Deit-S. For BAS, we insert the generator in the 10-th self-attention module. * denotes replacing the original losses with our proposed losses. + +better results than other methods at arbitrary threshold. Besides, SAT is not sensitive to thresholds and maintains high localization accuracy over a large threshold interval. This phenomenon shows that the localization map generated by SAT has high confidence and excellent visualization. + +Generalizability. To validate the generalizability of the proposed SAT, we implement experiments on different Transformers and various scales. As reported in Table 6, the proposed approach can be well applied to Deit [24] and Conformer [19] on CUB-200 [26]. On Deit-T/Deit-B, we achieve $3.2\% / 4.3\%$ GT-known Loc improvement compared to SCM [2], and $3.3\% / 2.1\%$ Top-1 classification accuracy gain over TS-CAM [8]. Besides, SAT also achieves state-of-the-art results on Conformer-S. Quantitative experiments show that the proposed SAT achieves excellent performance in both tasks on various transformer backbones. + +# 4.4. Discussions + +Why spatial token-based structure is critical? To clarify this question, we follow TS-CAM [8] to construct a attention map-based baseline and apply the proposed losses on it for comparison. As shown in Table 5, the attention map-based method substantially reduces the classification and localization accuracy, with $2.52\%$ and $3.39\%$ decrease in Top-1 Cls and GT-known Loc, respectively. To explore how this structural difference affects the accuracy, we plot the curves of spatial losses during training in Table 5. It can be noticed that the potential optimization conflicts of attention map-based method will lead to poor convergence of losses and insufficient learning, thus limiting the model to benefit from the losses. Experimental results show that the proposed spatial token-based structure can release the potential of the model by more efficient localization learn + +![](images/c85d13df1dffc9ca54dee64811002a48211e2aff89207f0e34c4924e6f631eca.jpg) +Figure 8. Few-shot learning. Box plot of the k-shot localization performance. For each k-shot setting, we repeat the experiment 10 times to randomly select k images as training data, and record the experimental results. The average results are shown in the figure. + +ing and avoiding optimization conflicts between two tasks, which are essential for the accuracy improvement. + +To further verify the effectiveness of the proposed structure in generating localization map, we reproduce two FPM-based methods ORNet [33] and BAS [32] on the Deit-S. Both of them produce localization maps by a generator. As shown in Table 7, first, we note that the original BAS is not suitable for the transformer due to the skip-connection in transformer causing inaccurate evaluation of the masking effect. Second, both FPM-based methods suffer from incomplete localization problem. This may be because they can only mask the input of a specific layer, resulting in overfitting the features that the specific feature layer focuses on. While the proposed structure can extract localization knowledge from different layers in an efficient form, thus improving localization ability and robustness. Quantitative and qualitative results demonstrate the effectiveness of SAT in terms of localization map generation and learning. + +Data-efficiency and tuning-efficiency. To validate the data-efficiency of the proposed method, we evaluate the localization performance of SAT and TS-CAM [8] under few-shot setting in Fig. 8. Compared to TS-CAM, SAT achieves promising results with only a few training data. Even in the extreme setting of only 1 training image per class, we exceed TS-CAM trained with full data by $3.6\%$ and $3.3\%$ GT-known Loc on CUB-200 and ImageNet, respectively. The box plot verifies the effectiveness and efficiency of the proposed SAT that implements localization task by an extra lightweight spatial token and enables high degree of network parameter sharing, thus requiring only a few data to achieve excellent localization results. In contrast, TS-CAM requires large amounts of data to fine-tune the classification feature maps to serve the localization task. + +To explore the tuning-efficiency of SAT, we freeze different parts of the network and evaluate the GT-known Loc of the proposed method under few-shot setting. All experiments are performed with ImageNet pre-trained weights. As shown in Table 8, SAT can still obtain competitive results when freezing most parameters of the pre-trained + +
Fro. rateCUB-200 [26]Fro. rateImageNet [20]
13510Full13510Full
(a)0%91.395.196.197.398.50%70.971.972.672.973.1
(b)24%93.595.996.897.598.121%71.672.372.672.873.1
(c)48%96.496.797.097.498.042%71.872.172.472.772.9
(d)71%93.894.694.995.295.964%71.371.471.771.972.2
(e)91%92.192.592.893.193.481%70.971.171.271.371.4
+ +Table 8. Frozen rate w.r.t. few-shot learning. The frozen parts are as follows: (a) None. (b) Attention layer of transformer blocks. (c) MLP layer of transformer blocks. (d) Transformer blocks. (e) Position embedding, projection, transformer blocks, and MLP layer of spatial aware transformer blocks. Fro.: Frozen. + +![](images/d4923b6b770aa9f5021a52caad3ba2b6a0daaa2886edf1a48d652e787339a05b.jpg) +Figure 9. Localization error analysis on CUB-200. + +classification model. It can be analyzed that the task-specific spatial token can be well adapted to the classification features and plays a major role in the localization learning, which allows efficient implementation of the localization task with fewer parameters to be fine-tuned. Experiments demonstrate both the data-efficiency and the tuning-efficiency of SAT. When freezing $81\%$ of the parameters and using only 1 image per class, SAT exceeds the SOTA method SCM [2] by $2.1\%$ GT-known Loc on ImageNet. + +# 4.5. Limitations + +As shown in Fig. 9, we count all localization errors (90 images) on CUB-200 test set (5,794 images) and classify the error causes into six categories, including object occlusion, localization more, water reflection, localization part, multiple instances, and label error. Among them, object occlusion causes the object to be split into two or more parts, resulting in incomplete localization results. Localization more is usually due to the positive effect of co-occurrence context on classification, leading to localizing confounding background regions. Water reflection is an inherent problem of WSOL and difficult to be solved with only image-level labels. Therefore, future work needs to consider more the interaction between object and context to overcome the problems of object occlusion and localization more. + +# 5. Conclusion + +This paper proposes to introduce a spatial-aware token (SAT) specific to the localization task, instead of synthesiz- + +ing the classification feature map, thus avoiding optimization conflicts between classification and localization tasks. To this end, we construct a spatial-query attention (SQA) module to generate localization map and extract localization knowledge from the classification task. Meanwhile, two spatial constraints, including batch area loss and regularization loss, are designed to compensate for the insufficient supervision. Extensive experiments on multiple benchmarks verify the effectiveness and efficiency of the proposed SAT that surpasses previous methods by a large margin. + +Acknowledgments. This work is supported by National Key R&D Program of China under Grant 2020AAA0105700, National Natural Science Foundation of China (NSFC) under Grants 62225207, U19B2038 and 62121002. + +# References + +[1] Wonho Bae, Junhyug Noh, and Gunhee Kim. Rethinking class activation mapping for weakly supervised object localization. In European Conference on Computer Vision, pages 618-634. Springer, 2020. 1 +[2] Haotian Bai, Ruimao Zhang, Jiong Wang, and Xiang Wan. Weakly supervised object localization via transformer with implicit spatial calibration. arXiv preprint arXiv:2207.10447, 2022. 2, 3, 5, 6, 7, 8 +[3] Zhiwei Chen, Changan Wang, Yabiao Wang, Guannan Jiang, Yunhang Shen, Ying Tai, Chengjie Wang, Wei Zhang, and Liujuan Cao. Lctr: On awakening the local continuity of transformer for weakly supervised object localization. arXiv preprint arXiv:2112.05291, 2021. 2, 3, 5 +[4] Junsuk Choe, Seungho Lee, and Hyunjung Shim. Attention-based dropout layer for weakly supervised single object localization and semantic segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(12):4256-4271, 2020. 1, 2 +[5] Junsuk Choe, Seong Joon Oh, Seungho Lee, Sanghyuk Chun, Zeynep Akata, and Hyunjung Shim. Evaluating weakly supervised object localization methods right. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3133-3142, 2020. 4, 5, 6, 7 +[6] Junsuk Choe and Hyunjung Shim. Attention-based dropout layer for weakly supervised object localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2219-2228, 2019. 1 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2 +[8] Wei Gao, Fang Wan, Xingjia Pan, Zhiliang Peng, Qi Tian, Zhenjun Han, Bolei Zhou, and Qixiang Ye. Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2886-2895, 2021. 1, 2, 4, 5, 6, 7, 8 + +[9] Guangyu Guo, Junwei Han, Fang Wan, and Dingwen Zhang. Strengthen learning tolerance for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7403-7412, 2021. 1, 5 +[10] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Amalgamating knowledge from heterogeneous graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15709-15718, 2021. 1 +[11] Eunji Kim, Siwon Kim, Jungbeom Lee, Hyunwoo Kim, and Sungroh Yoon. Bridging the gap between classification and localization for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14258-14267, 2022. 5, 6 +[12] Jeesoo Kim, Junsuk Choe, Sangdoo Yun, and Nojun Kwak. Normalization matters in weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3427-3436, 2021. 1 +[13] Jungbeom Lee, Eunji Kim, Jisoo Mok, and Sungroh Yoon. Anti-adversarily manipulated attributions for weakly supervised semantic segmentation and object localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1 +[14] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Dechao Meng, and Qingming Huang. Adaptive reconstruction network for weakly supervised referring expression grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2611-2620, 2019. 1 +[15] Weizeng Lu, Xi Jia, Weicheng Xie, Linlin Shen, Yicong Zhou, and Jinming Duan. Geometry constrained weakly supervised object localization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 481-496. Springer, 2020. 5 +[16] Jinjie Mai, Meng Yang, and Wenfeng Luo. Erasing integrated learning: A simple yet effective approach for weakly supervised object localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8766-8775, 2020. 1, 2 +[17] Meng Meng, Tianzhu Zhang, Qi Tian, Yongdong Zhang, and Feng Wu. Foreground activation maps for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3385-3395, 2021. 2, 5 +[18] Xingjia Pan, Yingguo Gao, Zhiwen Lin, Fan Tang, Weiming Dong, Haolei Yuan, Feiyue Huang, and Changsheng Xu. Unveiling the potential of structure preserving for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11642-11651, 2021. 1, 2, 5 +[19] Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, and Qixiang Ye. Conformer: Local features coupling global representations for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 367-376, 2021. 7 + +[20] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4, 5, 6, 8 +[21] Tal Shaharabany, Yoad Tewel, and Lior Wolf. What is where by looking: Weakly-supervised open-world phrase-grounding without text inputs. arXiv preprint arXiv:2206.09358, 2022. 1 +[22] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In 2017 IEEE international conference on computer vision (ICCV), pages 3544-3553. IEEE, 2017. 1, 2 +[23] Ganchao Tan, Daqing Liu, Meng Wang, and Zheng-Jun Zha. Learning to discretely compose reasoning module networks for video captioning. arXiv preprint arXiv:2007.09049, 2020.1 +[24] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 5, 6, 7 +[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[26] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 4, 5, 6, 7, 8 +[27] Hao Wang, Zheng-Jun Zha, Xuejin Chen, Zhiwei Xiong, and Jiebo Luo. Dual path interaction network for video moment localization. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4116-4124, 2020. 1 +[28] Hao Wang, Zheng-Jun Zha, Liang Li, Xuejin Chen, and Jiebo Luo. Semantic and relation modulation for audiovisual event localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1 +[29] Hao Wang, Zheng-Jun Zha, Liang Li, Dong Liu, and Jiebo Luo. Structured multi-level interaction network for video moment localization via language query. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7026-7035, 2021. 1 +[30] Jun Wei, Qin Wang, Zhen Li, Sheng Wang, S Kevin Zhou, and Shuguang Cui. Shallow feature matters for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5993-6001, 2021. 1, 2, 5 +[31] Jun Wei, Sheng Wang, S Kevin Zhou, Shuguang Cui, and Zhen Li. Weakly supervised object localization through inter-class feature similarity and intra-class appearance consistency. In European Conference on Computer Vision, pages 195-210. Springer, 2022. 2, 5, 6 +[32] Pingyu Wu, Wei Zhai, and Yang Cao. Background activation suppression for weakly supervised object localization. arXiv preprint arXiv:2112.00580, 2021. 2, 5, 6, 8 +[33] Jinheng Xie, Cheng Luo, Xiangping Zhu, Ziqi Jin, Weizeng Lu, and Linlin Shen. Online refinement of low-level feature + +based activation map for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 132-141, 2021. 2, 5, 8 +[34] Jinheng Xie, Jianfeng Xiang, Junliang Chen, Xianxu Hou, Xiaodong Zhao, and Linlin Shen. Contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. arXiv preprint arXiv:2203.13505, 2022. 2 +[35] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Rui-Wei Zhao, Tao Zhang, Xuequan Lu, and Shang Gao. Cream: Weakly supervised object localization via class re-activation mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9437–9446, 2022. 1, 5 +[36] Haolan Xue, Chang Liu, Fang Wan, Jianbin Jiao, Xi-angyang Ji, and Qixiang Ye. Danet: Divergent activation for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6589–6598, 2019. 1, 6 +[37] Seunghan Yang, Yoonhyung Kim, Youngeun Kim, and Changick Kim. Combinational class activation maps for weakly supervised object localization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2941-2949, 2020. 1 +[38] Tianhao Yang, Zheng-Jun Zha, and Hanwang Zhang. Making history matter: History-advantage sequence training for visual dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2561–2569, 2019. 1 +[39] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. Advances in neural information processing systems, 35:25739-25753, 2022. 1 +[40] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6023-6032, 2019. 1 +[41] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 1 +[42] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10):2472-2500, 2022. 1 +[43] Chen-Lin Zhang, Yun-Hao Cao, and Jianxin Wu. Rethinking the route towards weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13460-13469, 2020. 1, 2, 5 +[44] Dingwen Zhang, Junwei Han, Gong Cheng, and Ming-Hsuan Yang. Weakly supervised object localization and detection: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021. 1 +[45] Xiaolin Zhang, Yunchao Wei, Jiashi Feng, Yi Yang, and Thomas S Huang. Adversarial complementary learning for weakly supervised object localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1325-1334, 2018. 1, 2 + +[46] Xiaolin Zhang, Yunchao Wei, Guoliang Kang, Yi Yang, and Thomas Huang. Self-produced guidance for weakly-supervised object localization. In Proceedings of the European conference on computer vision (ECCV), pages 597-613, 2018. 1, 2 +[47] Xiaolin Zhang, Yunchao Wei, and Yi Yang. Inter-image communication for weakly supervised localization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 271-287. Springer, 2020. 1, 2 +[48] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1, 2, 5 +[49] Lei Zhu, Qian Chen, Lujia Jin, Yunfei You, and Yanye Lu. Bagging regional classification activation maps for weakly supervised object localization. arXiv preprint arXiv:2207.07818, 2022. 5, 6 +[50] Lei Zhu, Qi She, Qian Chen, Yunfei You, Boyu Wang, and Yanye Lu. Weakly supervised object localization as domain adaption. arXiv preprint arXiv:2203.01714, 2022. 5, 6 \ No newline at end of file diff --git a/2023/Spatial-Aware Token for Weakly Supervised Object Localization/images.zip b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..70bd4702ec5b0e16c0fb7d51c231d72985d862a4 --- /dev/null +++ b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47503eb1df6493a1f2aeb21025e80416756aade979faf54a8d3221cf2516d62d +size 1066456 diff --git a/2023/Spatial-Aware Token for Weakly Supervised Object Localization/layout.json b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f881f3560231428364dc7597c66faafefd27bbe0 --- /dev/null +++ b/2023/Spatial-Aware Token for Weakly Supervised Object Localization/layout.json @@ -0,0 +1,8978 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 496, + 121 + ], + "type": "text", + "content": "Spatial-Aware Token for Weakly Supervised Object Localization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "text", + "content": "Pingyu Wu" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "text", + "content": ", Wei Zhai" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "text", + "content": ", Yang Cao" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "text", + "content": ", Jiebo Luo" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "text", + "content": ", Zheng-Jun Zha" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "text", + "content": " University of Science and Technology of China \n" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "text", + "content": " University of Rochester \n" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 142, + 496, + 185 + ], + "type": "text", + "content": " Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 187, + 455, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 187, + 455, + 200 + ], + "spans": [ + { + "bbox": [ + 140, + 187, + 455, + 200 + ], + "type": "text", + "content": "wpy364755620@@mail., wzhai056@mail., forrest@}ustc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 183, + 201, + 411, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 201, + 411, + 213 + ], + "spans": [ + { + "bbox": [ + 183, + 201, + 411, + 213 + ], + "type": "text", + "content": "jluo@cs.ROchester.edu zhazj@ustc.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "spans": [ + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 278, + 290, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 290, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 290, + 614 + ], + "type": "text", + "content": "Weakly supervised object localization (WSOL) is a challenging task aiming to localize objects with only image-level supervision. Recent works apply visual transformer to WSOL and achieve significant success by exploiting the long-range feature dependency in self-attention mechanism. However, existing transformer-based methods synthesize the classification feature maps as the localization map, which leads to optimization conflicts between classification and localization tasks. To address this problem, we propose to learn a task-specific spatial-aware token (SAT) to condition localization in a weakly supervised manner. Specifically, a spatial token is first introduced in the input space to aggregate representations for localization task. Then a spatial aware attention module is constructed, which allows spatial token to generate foreground probabilities of different patches by querying and to extract localization knowledge from the classification task. Besides, for the problem of sparse and unbalanced pixel-level supervision obtained from the image-level label, two spatial constraints, including batch area loss and normalization loss, are designed to compensate and enhance this supervision. Experiments show that the proposed SAT achieves state-of-the-art performance on both CUB-200 and ImageNet, with " + }, + { + "bbox": [ + 46, + 278, + 290, + 614 + ], + "type": "inline_equation", + "content": "98.45\\%" + }, + { + "bbox": [ + 46, + 278, + 290, + 614 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 278, + 290, + 614 + ], + "type": "inline_equation", + "content": "73.13\\%" + }, + { + "bbox": [ + 46, + 278, + 290, + 614 + ], + "type": "text", + "content": " GT-known Loc, respectively. Even under the extreme setting of using only 1 image per class from ImageNet for training, SAT already exceeds the SOTA method by " + }, + { + "bbox": [ + 46, + 278, + 290, + 614 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 46, + 278, + 290, + 614 + ], + "type": "text", + "content": " GT-known Loc. Code and models are available at https://github.com/wpy1999/SAT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "text", + "content": "Weakly supervised object localization (WSOL) aims to localize objects with only image-level labels available. Since no expensive bounding box or pixel-level annota" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 238, + 428, + 383 + ], + "blocks": [ + { + "bbox": [ + 312, + 238, + 428, + 383 + ], + "lines": [ + { + "bbox": [ + 312, + 238, + 428, + 383 + ], + "spans": [ + { + "bbox": [ + 312, + 238, + 428, + 383 + ], + "type": "image", + "image_path": "08b6b50cb1c4fdaed3f540bc09b82ffd93d6f644afd7cd9d4aebf7d540d41ec6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 385, + 546, + 418 + ], + "lines": [ + { + "bbox": [ + 305, + 385, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 546, + 418 + ], + "type": "text", + "content": "Figure 1. GT-known Loc on ImageNet. The proposed method only needs to fine-tune a small number of parameters or requires a few training images to achieve significant improvement." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 430, + 238, + 540, + 384 + ], + "blocks": [ + { + "bbox": [ + 430, + 238, + 540, + 384 + ], + "lines": [ + { + "bbox": [ + 430, + 238, + 540, + 384 + ], + "spans": [ + { + "bbox": [ + 430, + 238, + 540, + 384 + ], + "type": "image", + "image_path": "6eaee214423d52cf6cbaee399e0e1d7db2238becda7ffbbb3e863d90d5b658de.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 486 + ], + "type": "text", + "content": "tions are required, WSOL significantly reduces the cost of manual annotations [21, 44, 12, 13, 4, 43, 22, 45] and has attracted increasing attention in the research community [40, 42, 41, 39, 10, 27, 29, 28, 14, 38, 23]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 486, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 594 + ], + "type": "text", + "content": "As a representative work, CAM [48] extracts class activation maps from the classifier as localization maps. However, CAM is usually coarse and focuses on the most discriminative regions, leading to imprecise and incomplete localization results. To solve these problems, many CNN-based methods have been proposed, such as adversarial erasure [22, 45, 6, 16], divergent activation [36, 40, 22], seed region growing [30, 46], regularization [18, 13, 47], feature refining [1, 37, 30], regression-based [35, 43, 9]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "Recently, transformer [7, 24] has been introduced to the field of computer vision with great success. Benefiting from the long-range feature dependency, the attention map upon the class token usually can capture the object region relevant to classification in the whole image, thus it is widely used for localization in the transformer-based WSOL methods. The pioneering work TS-CAM [8] accumulates attention maps of each layer and mixes them with semantic-aware maps to achieve localization task. Further, to address the problem that transformer lacks inherent spatial coherence" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1844" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "of the object, LCTR [3] and SCM [2] propose to consider cross-patch information and use activation diffusion to increase the local continuity of localization map, respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 302 + ], + "type": "text", + "content": "Existing transformer-based methods synthesize the feature maps learned by the classification task, such as attention maps, as the localization map, and try to increase its connectivity and completeness. However, this convenient approach results in optimization conflicts between classification and localization tasks. 1) For classification, making classification feature maps learn more object regions with low discrimination will reduce the classification ability. 2) For localization, the learning of the localization map is limited by the properties of the feature maps. For example, the attention map is generated by the softmax function, which is difficult to produce a balanced and comprehensive response over the object. Therefore, to achieve a more promising localization performance and avoid optimization conflicts, it is necessary to construct task-specific parameters for the generation and learning of localization map." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 471 + ], + "type": "text", + "content": "Based on transformer architecture, a straightforward idea is to introduce a spatial token in the input space to aggregate global representation for the localization task. And the localization map can be further obtained by interacting the spatial token with each patch in the forward propagation. To this end, we propose a Spatial-Query Attention (SQA) module that takes the spatial token as a query to calculate the similarity with different patches and produces the localization map efficiently. Meanwhile, to make the spatial-aware token obtain localization supervision from image-level labels, the localization map is treated as a visual cue to participate in the calculation of cross attention, thus the generation and learning of localization map can be well adapted to the transformer-based classification model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 675 + ], + "type": "text", + "content": "Building upon this efficient SQA module, we propose a simple but effective Spatial-Aware Token (SAT) approach, as shown in Fig. 2. Specifically, a task-specific spatial token is introduced in the input space. To achieve the localization task, several SQA modules are applied in the transformer to produce localization maps learned from different layers and aggregate them together. In this way, the final localization map can maximally capture the localization knowledge from the whole classification model. However, the pixel-level supervision generated by image-level labels is sparse and unbalanced. To compensate and strengthen this supervision, we propose two spatial constraints, including batch area loss and normalization loss. Batch area loss aims to provide a sparse area supervision with prior knowledge to compensate for the insufficient supervision. Normalization loss is employed to enhance pixel-level supervision by encouraging the localization map to be more discriminatory." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Benefiting from the task-specific spatial token and avoiding optimization conflicts, SAT releases the potential of the model and achieves excellent performance in both classifi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "cation and localization tasks. Besides, unlike synthetic classification feature maps as localization results, the generation and learning of localization map in SAT rely mainly on an extra lightweight token, which brings advantages of data-efficiency and tuning-efficiency. As illustrated in Fig. 1, SAT outperforms SOTA method SCM by " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " GT-known Loc with less than " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " training data on ImageNet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 156, + 527, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 156, + 527, + 167 + ], + "spans": [ + { + "bbox": [ + 317, + 156, + 527, + 167 + ], + "type": "text", + "content": "In summary, the contributions of this paper include:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 168, + 545, + 312 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 304, + 168, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 215 + ], + "type": "text", + "content": "1) We propose to learn a task-specific spatial token to implement the localization task, instead of synthesizing classification feature maps, thus avoiding optimization conflicts between classification and localization tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 216, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 275 + ], + "type": "text", + "content": "2) We propose a simple but effective spatial-aware token (SAT) pipeline for WSOL, which utilizes a spatial-query attention (SQA) module to generate localization map through a spatial token and supervise it with two spatial constraints, including batch area loss and normalization loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 275, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 312 + ], + "type": "text", + "content": "3) Extensive experiments show that SAT outperforms SOTA methods by a large margin on multiple benchmarks and performs excellently even under extreme settings." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 323, + 390, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 390, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 390, + 335 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 342, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 641 + ], + "type": "text", + "content": "CNN-based methods. WSOL aims to learn the localization of objects with only image-level labels. CAM [48] proposes to synthesize depth feature maps with fully connected weights to obtain class activation maps (CAM). To increase the localization ability of CAM, HaS [22] adopts a random erasure strategy forcing the network to attend to different object regions. ACoL [45] and EIL [16] use adversarial erasing to achieve learning of complementary regions. ADL [4] drops the most discriminative or highlighted regions in the different layers during forward propagation. Besides, SPA [18] designs a post-process approach to extract structure-preserving localization map from the feature maps. SPG [46] and SPOL [30] facilitate learning of localization by selecting and spreading confidence regions as seeds. PSOL [43] and " + }, + { + "bbox": [ + 304, + 342, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{AM}" + }, + { + "bbox": [ + 304, + 342, + 545, + 641 + ], + "type": "text", + "content": " [34] divide WSOL into independent classification task and class-agnostic object localization. I2C [47] and ISIC [31] consider feature similarities across different objects to achieve more complete and robust localization. ORNet [33], FAM [17] and BAS [32] propose to generate a foreground prediction map (FPM) to implement localization. Unlike FPM-based methods that require complex and heavy structures, and the learning of generator relies on a specific feature layer of the classification network. Our method simply and efficiently obtains localization maps from different layers by a task-specific token." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Transformer-based methods. Different from CNNs, transformer [7, 25] can capture global cues with the advantage of long-range dependency in the self-attention mechanism, effectively alleviating the partial activation problem. TS-CAM [8] mixes the attention maps of different layers and combines them with semantic-aware tokens to produce" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1845" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 545, + 220 + ], + "type": "image", + "image_path": "e24218d708a5a464ac22cd185e7c0ecadc2c19c3b02ffed34495bba8e516dd59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 221, + 546, + 244 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 546, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 546, + 244 + ], + "type": "text", + "content": "Figure 2. Framework of SAT. It includes three spatial aware transformer blocks at the end of the network. Each block generates a localization map " + }, + { + "bbox": [ + 46, + 221, + 546, + 244 + ], + "type": "inline_equation", + "content": "M^l" + }, + { + "bbox": [ + 46, + 221, + 546, + 244 + ], + "type": "text", + "content": " using the spatial-query attention module. The final localization map " + }, + { + "bbox": [ + 46, + 221, + 546, + 244 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 221, + 546, + 244 + ], + "type": "text", + "content": " is obtained by fusing " + }, + { + "bbox": [ + 46, + 221, + 546, + 244 + ], + "type": "inline_equation", + "content": "M^l" + }, + { + "bbox": [ + 46, + 221, + 546, + 244 + ], + "type": "text", + "content": " of different layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 289, + 371 + ], + "type": "text", + "content": "semantic-aware localization results. Based on TS-CAM, to address the lack of spatial consistency in transformer, LCTR [3] proposes to enhance the local perception capability among long-range feature dependencies by considering cross-patch information. SCM [2] increases the spatial coherence of attention maps by diffusing the semantic and spatial connections between patch tokens. Notably, they all generate localization map by synthesizing feature maps. Instead, we produce it by a task-specific spatial token." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 387, + 129, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 129, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 129, + 401 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 408, + 115, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 115, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 115, + 420 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "content": "The overall network architecture of SAT is shown in Fig. 2. Given an input image " + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "content": ", a plain vision transformer splits it into a sequence of non-overlapping patches. Then the divided patches are flattened and transformed into patch tokens " + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\{x_{n} \\in \\mathbb{R}^{1 \\times D}, n = 1, 2, \\dots, H \\times W\\}" + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "content": " by linear projection, where " + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "inline_equation", + "content": "H = h / P" + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "inline_equation", + "content": "W = w / P" + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "content": " is the patch size, and " + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "content": " is the number of channels. For simplicity, we omit the description of the batch size " + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "content": ". After grouping the class token and the extra spatial-aware token " + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "inline_equation", + "content": "x_{spa}" + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "content": " with patch tokens, this token sequence " + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^0 \\in \\mathbb{R}^{(N + 2) \\times D}" + }, + { + "bbox": [ + 46, + 429, + 287, + 573 + ], + "type": "text", + "content": " is sent into stacked transformer blocks and spatial aware transformer blocks for subsequent representation learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 574, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 635 + ], + "type": "text", + "content": "The proposed spatial aware transformer block inherits the typical transformer block structure. Thus the entire network is characterized by alternating attention layers and MLP layers. For a given input token subsequent " + }, + { + "bbox": [ + 46, + 574, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{l-1}" + }, + { + "bbox": [ + 46, + 574, + 287, + 635 + ], + "type": "text", + "content": ", the overall equations of the " + }, + { + "bbox": [ + 46, + 574, + 287, + 635 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 574, + 287, + 635 + ], + "type": "text", + "content": "-th block are defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 647, + 287, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 647, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 87, + 647, + 287, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {X} ^ {l} = \\operatorname {L a y e r - N o r m} \\left(\\mathcal {F} ^ {l - 1}\\right), \\tag {1}", + "image_path": "6cb0a93c8400cfd9ebf237f18e4480c2a3d476932e11860e0c735c21da59a5bf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 87, + 669, + 287, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 669, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 87, + 669, + 287, + 689 + ], + "type": "interline_equation", + "content": "\\mathcal {Z} ^ {l} = \\mathcal {F} ^ {l - 1} + \\text {A t t e n t i o n} \\left(\\mathcal {X} ^ {l}\\right), \\tag {2}", + "image_path": "a3d240b059f9b24c376488a6c14129a6542ec826c0a2f3cd7f6c6263c89369f2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 87, + 691, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 87, + 691, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {F} ^ {l} = \\mathcal {Z} ^ {l} + \\operatorname {M L P} \\left(\\text {L a y e r - N o r m} \\left(\\mathcal {Z} ^ {l}\\right)\\right). \\tag {3}", + "image_path": "7e756142d954ff08e5552b1972b7f82650d68975ce960dcbe998d22fd0ba1693.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 262, + 487, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 487, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 487, + 275 + ], + "type": "text", + "content": "3.2. Spatial Aware Transformer Block" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "type": "text", + "content": "The proposed spatial aware transformer block mainly contains a spatial-query attention (SQA) module to generate localization map and extract localization knowledge. Besides, two spatial constraints, including batch area loss " + }, + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ba}" + }, + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "type": "text", + "content": ", and normalization loss " + }, + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{norm}" + }, + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "type": "text", + "content": " are designed as the complement to the insufficient localization supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": "Spatial-query Attention. The proposed SQA module is based on self-attention module and aims at converting the spatial-aware token " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "x_{spa}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": " into a localization map " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "M_{spa}^{l}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": " by querying different patches. Meanwhile, the generated " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "M_{spa}^{l}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": " is applied to the cross-attention calculation in the form of dot product, thus obtaining localization supervision from the image-level labels. Specifically, the SQA module first linearly projects the token sequence " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^{l - 1}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": " to the query matrix " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": ", key matrix " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": ", value matrix " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{Q},\\mathcal{K},\\mathcal{V} \\in \\mathbb{R}^{(N + 2)\\times D}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": ". Then the query vector corresponding to the spatial token " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{spa} \\in \\mathbb{R}^{1\\times D}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": " is selected from " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": " and used to query " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": " to obtain the query results for each token. The similarity map between the spatial query vector and key matrix in the " + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 354, + 546, + 521 + ], + "type": "text", + "content": "-th layer is calculated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 345, + 527, + 545, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 527, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 345, + 527, + 545, + 554 + ], + "type": "interline_equation", + "content": "\\mathcal {S} ^ {l} \\left(\\mathcal {Q} _ {s p a}, \\mathcal {K}\\right) = \\frac {\\mathcal {Q} _ {s p a} \\mathcal {K} ^ {\\mathrm {T}}}{\\sqrt {D}} \\in \\mathbb {R} ^ {1 \\times (N + 2)}, \\tag {4}", + "image_path": "b6ca4a54882ef03fa8c45ae3ba08ac71195e39f42e522d4bd75b0ad9daf8d9fb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 560, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 545, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 560, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\sqrt{D}" + }, + { + "bbox": [ + 304, + 560, + 545, + 609 + ], + "type": "text", + "content": " is employed as a scaling factor. After applying sigmoid activation, the similarity map is transformed into foreground probability " + }, + { + "bbox": [ + 304, + 560, + 545, + 609 + ], + "type": "inline_equation", + "content": "M_{pro}^{l}" + }, + { + "bbox": [ + 304, + 560, + 545, + 609 + ], + "type": "text", + "content": " with a 0 to 1 interval distribution. This process can be expressed as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 331, + 613, + 545, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 613, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 331, + 613, + 545, + 630 + ], + "type": "interline_equation", + "content": "M _ {s p a} ^ {l} = \\operatorname {S i g m o i d} \\left(\\mathcal {S} ^ {l} \\left(\\mathcal {Q} _ {s p a}, \\mathcal {K}\\right)\\right) \\in \\mathbb {R} ^ {1 \\times (N + 2)}. \\tag {5}", + "image_path": "0f5e6be2059371b405132f0c8074118dc9fe6b74a8e9fffd61b26ca8a0b7f611.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 635, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 545, + 685 + ], + "type": "text", + "content": "The generated " + }, + { + "bbox": [ + 304, + 635, + 545, + 685 + ], + "type": "inline_equation", + "content": "M_{spa}^{l}" + }, + { + "bbox": [ + 304, + 635, + 545, + 685 + ], + "type": "text", + "content": " is served as a visual cue to point out the object region and combined in the calculation of attention with the form of dot product. In this way, the whole SQA module can be formulated in the following form:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 342, + 689, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 689, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 342, + 689, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\operatorname {S Q A} \\left(\\mathcal {X} ^ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {Q K} ^ {\\mathrm {T}}}{\\sqrt {D}}\\right) * M _ {s p a} ^ {l} \\mathcal {V}, \\tag {6}", + "image_path": "90750d5a6e6689c29d5c9e65c1bea9f5f71f194237dcf4902963023659c0fdc7.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1846" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 64, + 547, + 290 + ], + "blocks": [ + { + "bbox": [ + 50, + 64, + 547, + 290 + ], + "lines": [ + { + "bbox": [ + 50, + 64, + 547, + 290 + ], + "spans": [ + { + "bbox": [ + 50, + 64, + 547, + 290 + ], + "type": "image", + "image_path": "86faad96805cfacef49e0f7948004b07a400a3044fe353a45fc7dbb526495375.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 293, + 531, + 304 + ], + "lines": [ + { + "bbox": [ + 61, + 293, + 531, + 304 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 531, + 304 + ], + "type": "text", + "content": "Figure 3. Visualization comparison. The ground-truth bounding boxes are in red, and the predicted bounding boxes are in green." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 324, + 230, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 230, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 230, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 324, + 230, + 335 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 47, + 324, + 230, + 335 + ], + "type": "text", + "content": " denotes element-wise multiplication." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "type": "text", + "content": "By cropping and reshaping, the part of the " + }, + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "type": "inline_equation", + "content": "M_{spa}^{l}" + }, + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "type": "text", + "content": " to the patch tokens can be transformed into a localization map " + }, + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "type": "inline_equation", + "content": "M^{l} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "type": "text", + "content": ". After obtaining " + }, + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "type": "inline_equation", + "content": "M^{l}" + }, + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "type": "text", + "content": " learned from different layers, we take their averages as the final localization map " + }, + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 336, + 287, + 455 + ], + "type": "text", + "content": " to increase the completeness and robustness. Although this approach can maximally capture the localization information in the classification network, the pixel-level supervision obtained from image-level labels is sparse and unbalanced. To compensate and strengthen this supervision, we design batch area loss and normalization loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 456, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 527 + ], + "type": "text", + "content": "Batch Area. To alleviate the insufficient localization supervision caused by the absence of dense labels in WSOL, we propose to use batch area loss as a complement to localization supervision. It aims to constrain the average area of localization maps " + }, + { + "bbox": [ + 47, + 456, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\{M_b\\in \\mathbb{R}^{H\\times W},b = 1,2,\\dots,B\\}" + }, + { + "bbox": [ + 47, + 456, + 287, + 527 + ], + "type": "text", + "content": " within a batch " + }, + { + "bbox": [ + 47, + 456, + 287, + 527 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 456, + 287, + 527 + ], + "type": "text", + "content": " to a hyperparameter " + }, + { + "bbox": [ + 47, + 456, + 287, + 527 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 456, + 287, + 527 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 531, + 287, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 531, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 82, + 531, + 287, + 566 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b a} = \\left| \\sum_ {b} ^ {B} \\sum_ {i} ^ {H} \\sum_ {j} ^ {W} \\left(\\lambda - \\frac {M _ {b} (i , j)}{B \\times H \\times W}\\right) \\right|, \\tag {7}", + "image_path": "cd861f82723ab593c3b91c27ffda665b4baaf3a87d0ebdc2c8e0793adf67945d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 570, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 286, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 570, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 570, + 286, + 593 + ], + "type": "text", + "content": " is a sparse area supervision with prior knowledge. The " + }, + { + "bbox": [ + 47, + 570, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 570, + 286, + 593 + ], + "type": "text", + "content": " is set to 0.25 and 0.35 on CUB-200 and ImageNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Normalization. For unbalanced pixel-level supervision, a normalization loss is applied to enhance the pixel-level supervision based on the pixel response intensity. It encourages pixels to be more distinguishable about whether they are foreground or background, as presented by Eq. 8. Before penalizing the uncertainty value of the localization map " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", to further enhance its local consistency, we adopt a gaussian filter with a kernel size of " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " to establish spatial connections between adjacent patches. The standard deviation of gaussian kernel is set to 6 on all datasets. After" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "type": "text", + "content": "applying gaussian filtering, the value of each patch is redistributed according to the adjacent patch values, and " + }, + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "type": "text", + "content": " is converted to " + }, + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "type": "inline_equation", + "content": "M^{*}" + }, + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "type": "text", + "content": ". By computing the normalization loss with " + }, + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "type": "inline_equation", + "content": "M^{*}" + }, + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "type": "text", + "content": " can further encourage " + }, + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 324, + 545, + 384 + ], + "type": "text", + "content": " to have local consistency and avoid noise. It can be expressed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 392, + 545, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 392, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 392, + 545, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m}} = \\frac {1}{H \\times W} \\sum_ {i} ^ {H} \\sum_ {j} ^ {W} M ^ {*} (i, j) (1 - M ^ {*} (i, j)). \\tag {8}", + "image_path": "77951bef1be0f6bd030c4eff564e0b9d2bceb9871cd6a4b61121602e0165782a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 434, + 372, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 372, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 372, + 445 + ], + "type": "text", + "content": "3.3. Total loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 453, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 513 + ], + "type": "text", + "content": "We follow the design of classification head in the baseline method TS-CAM [8] to obtain the predicted probability " + }, + { + "bbox": [ + 305, + 453, + 545, + 513 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 453, + 545, + 513 + ], + "type": "text", + "content": " and the cross-entropy classification loss " + }, + { + "bbox": [ + 305, + 453, + 545, + 513 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 305, + 453, + 545, + 513 + ], + "type": "text", + "content": ". By jointly optimizing classification loss, batch area loss, and normalization loss, the total loss function is defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 370, + 523, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 523, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 370, + 523, + 545, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {b a} + \\mathcal {L} _ {\\text {n o r m}}, \\tag {9}", + "image_path": "2bbb45553d745c19c31d9af92bdb1ecda0f57f152969e0fc3306b00422a5307f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 545, + 544, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 544, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 544, + 557 + ], + "type": "text", + "content": "where no extra hyperparameter is used to balance each loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 567, + 382, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 382, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 382, + 581 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 587, + 422, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 422, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 422, + 601 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate the proposed SAT on three popular datasets, including CUB-200 [26], ImageNet [20], OpenImages [5]. CUB-200 is a fine-grained dataset with 200 bird species, which consists of 5,994 images for training and 5,794 images for testing. ImageNet contains about 1.2 million training images and 50,000 validation images from 1,000 different classes. OpenImages contains 100 categories, it has 29,819, 2,500, and 5,000 samples in training, val, and test sets, respectively. Apart from the image-level" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1847" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 524, + 340 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 524, + 340 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 524, + 340 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 524, + 340 + ], + "type": "table", + "html": "
MethodsVenueBackboneCUB-200 [26]Loc Acc.ImageNet [20]Loc Acc.
Top-1Top-5GT-knownTop-1Top-5GT-known
CAM [48]CVPR16VGG1641.0650.6655.1042.8054.8659.00
ORNet [33]ICCV21VGG1667.7380.7786.2052.0563.9468.27
BAS [32]CVPR22VGG1671.3385.3391.0752.9665.4169.64
Kim et al. [11]CVPR22VGG1670.8388.0793.1749.9463.2568.92
CREAM [35]CVPR22VGG1670.4485.6790.9852.3764.2068.32
GCNet [15]ECCV20InceptionV358.5871.0075.3049.0658.09-
SPA [18]CVPR21InceptionV353.5966.5072.1452.7364.2768.33
FAM [17]ICCV21InceptionV370.67-87.2555.24-68.62
CREAM [35]CVPR22InceptionV371.7686.3790.4356.0766.1969.03
BAS [32]CVPR22InceptionV373.2986.3192.2458.5169.0071.93
BagCAMs [49]ECCV22InceptionV360.07-89.7853.87-71.02
SPOL [30]CVPR21ResNet5080.1293.4496.4659.1467.1569.02
DA-WSOL [50]CVPR22ResNet5066.65-81.8355.84-70.27
BAS [32]CVPR22ResNet5077.2590.0895.1357.1868.4471.77
Kim et al. [11]CVPR22ResNet5073.1686.6891.6053.7665.7569.89
CREAM [35]CVPR22ResNet5076.03-89.8855.66-69.31
BagCAMs [49]ECCV22ResNet5069.67-94.0144.24-72.08
ISIC [31]ECCV22ResNet5080.6894.0897.3259.6167.8470.01
TS-CAM [8]ICCV21Deit-S71.3083.8087.7053.4064.3067.60
LCTR [3]AAAI22Deit-S79.2089.9092.4056.1065.8068.70
SCM [2]ECCV22Deit-S76.4091.6096.6056.1066.4068.80
SAT (ours)This WorkDeit-S80.9694.1398.4560.1570.5273.13
", + "image_path": "d00f375b66a8e49c8d95c71bcecd9d3b6ea6c050c6fdc015c5dc3e14cdd8c1c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 375, + 211, + 448 + ], + "blocks": [ + { + "bbox": [ + 83, + 342, + 511, + 354 + ], + "lines": [ + { + "bbox": [ + 83, + 342, + 511, + 354 + ], + "spans": [ + { + "bbox": [ + 83, + 342, + 511, + 354 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods. The best results are highlighted in bold, second are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 375, + 211, + 448 + ], + "lines": [ + { + "bbox": [ + 48, + 375, + 211, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 211, + 448 + ], + "type": "table", + "html": "
SettingLoc Acc.
LclsLbaLnorm F.Top-1Top-5GT-k.
(a)✓56.8666.6769.29
(b)✓✓58.2168.3671.07
(c)✓✓✓59.8670.2272.78
(d)✓✓✓✓60.1570.5273.13
", + "image_path": "a75a80fa1023d22f0fb2b19452797e51f9cc9ea2a9ea841c42865b09856173fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 212, + 371, + 287, + 449 + ], + "blocks": [ + { + "bbox": [ + 212, + 371, + 287, + 449 + ], + "lines": [ + { + "bbox": [ + 212, + 371, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 212, + 371, + 287, + 449 + ], + "type": "image", + "image_path": "11bee27d3e8496a83fc8c0c6dd7175be542f69f71cced9cb164379896020fb31.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 453, + 286, + 464 + ], + "lines": [ + { + "bbox": [ + 47, + 453, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 286, + 464 + ], + "type": "text", + "content": "Table 2. Ablation study of SAT on ImageNet. " + }, + { + "bbox": [ + 47, + 453, + 286, + 464 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 453, + 286, + 464 + ], + "type": "text", + "content": " : Gaussian filter." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "type": "text", + "content": "labels, both CUB-200 and OpenImages provide pixel-level mask labels, which are only used in the testing phase." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": "Metrics. Following [32, 18, 50], for localization, we utilize GT-known localization accuracy (GT-known Loc), Top-1/Top5 localization accuracy (Top-1/Top-5 Loc), and maximal box accuracy (MaxBoxAccV2) [5] as evaluation metrics. GT-known Loc is correct indicating that the intersection over union (IoU) of the predicted bounding box and the ground-truth bounding box is " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": " or more. Top-1/Top-5 Loc is correct when the ground-truth class belongs to Top-1/Top-5 prediction categories and GT-known Loc is correct. For mask, the peak intersection over union (pIoU) [43] and pixel average precision (PxAP) [5] are adopted as metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Implementation Details. We evaluate our method on the Deit-S [24] pre-trained on ImageNet [20]. On CUB-200 [26], the training process lasts 30 epochs with a batch size of 256. For ImageNet [20], we train 5 epochs and set batch size to 512. In the training phase, the input images are resized to " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " and then randomly cropped to" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 371, + 425, + 446 + ], + "blocks": [ + { + "bbox": [ + 305, + 371, + 425, + 446 + ], + "lines": [ + { + "bbox": [ + 305, + 371, + 425, + 446 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 425, + 446 + ], + "type": "image", + "image_path": "a2005113709f5d946aa16ac29d6d23f36103935fda719c1897568f9e0c302385.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 449, + 545, + 471 + ], + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 471 + ], + "type": "text", + "content": "Figure 4. Hyperparameters. Sensitivity analysis of hyperparameters (a) " + }, + { + "bbox": [ + 305, + 449, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 449, + 545, + 471 + ], + "type": "text", + "content": " and (b) batch size in " + }, + { + "bbox": [ + 305, + 449, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ba}" + }, + { + "bbox": [ + 305, + 449, + 545, + 471 + ], + "type": "text", + "content": " on ImageNet." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 371, + 545, + 446 + ], + "blocks": [ + { + "bbox": [ + 425, + 371, + 545, + 446 + ], + "lines": [ + { + "bbox": [ + 425, + 371, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 425, + 371, + 545, + 446 + ], + "type": "image", + "image_path": "556a46dcf2f9f9fad1c2ea261e194bf2374c78cab75032d2483ba4ff262a668c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 492, + 545, + 529 + ], + "type": "text", + "content": ". In the inference phase, following [32, 9, 43], we adopt ten crop augmentations to obtain classification results and replace random crop with center crop for localization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 537, + 400, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 400, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 400, + 551 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "content": "In this subsection, we implement a series of ablation experiments with Deit-S [24] as the backbone. All experiments are conducted on ImageNet [20], a universal dataset, to increase the generalizability of the experimental results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "Ablation studies of SAT components. Table 2 shows the localization accuracy of SAT with different compositions, where we retain the structure of SAT as the baseline. It can be noted that the proposed baseline can effectively activate the object region and already exceeds SOTA method SCM with only " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": ". Based on the baseline, the addition of " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ba}" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " significantly improves the GT-known Loc with a " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "1.78\\%" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " gain, by limiting the expansion of the localization map and forcing it to focus on the regions relevant to the classifi-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1848" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 164 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 164 + ], + "type": "table", + "html": "
MethodVenueBackboneResolutionCUB-200 [26]ImageNet [20]
δ = 0.3δ = 0.5δ = 0.7Meanδ = 0.3δ = 0.5δ = 0.7Mean
Kim et al. [11]CVPR22ResNet5028×2899.4090.4038.0075.9086.7071.1048.3068.70
BAS [32]CVPR22ResNet5028×2899.4195.1371.9288.8284.8071.7749.2568.61
TS-CAM [8]ICCV21DeiT-S14×1498.8887.7049.9378.8482.1167.6044.7664.82
SCM [2]ECCV22DeiT-S14×1499.6496.6071.7389.3283.6468.8045.3065.91
SAT (ours)This WorkDeiT-S14×1499.88 (+0.24)98.45 (+1.85)79.53 (+7.61)92.62 (+3.30)84.42 (-2.28)73.13 (+1.36)56.83 (+7.58)71.46 (+2.76)
", + "image_path": "d051c84daff14cadd2466fa70dda38c529c81c78f9cd8eb2d28f6b66ee993e95.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 201, + 167, + 297 + ], + "blocks": [ + { + "bbox": [ + 48, + 201, + 167, + 297 + ], + "lines": [ + { + "bbox": [ + 48, + 201, + 167, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 167, + 297 + ], + "type": "image", + "image_path": "4bf90063eb1c7071935a4b480656b2d3a359f2e399d738ddabc93c748fe865fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 298, + 269, + 310 + ], + "lines": [ + { + "bbox": [ + 65, + 298, + 269, + 310 + ], + "spans": [ + { + "bbox": [ + 65, + 298, + 269, + 310 + ], + "type": "text", + "content": "Figure 5. Statistical analysis about localization quality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 201, + 287, + 297 + ], + "blocks": [ + { + "bbox": [ + 167, + 201, + 287, + 297 + ], + "lines": [ + { + "bbox": [ + 167, + 201, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 167, + 201, + 287, + 297 + ], + "type": "image", + "image_path": "bf14a74d5c0a95193dce7d766148632a5156bcaed8014be42221bbc31a93c9b2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "content": "cation. Then the use of " + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {norm }}" + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "content": " can effectively increase the distinction between the foreground and background regions, resulting in a remarkable increase of " + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "inline_equation", + "content": "1.71\\%" + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "content": " GT-known Loc. Finally, applying gaussian filter can further eliminate noise and increase local connectivity, by considering the consistency of adjacent patches in the calculation of " + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {norm }}" + }, + { + "bbox": [ + 46, + 329, + 287, + 400 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "text", + "content": "Hyperparameters. Fig. 4 presents the sensitivity of localization quality to the hyperparameters " + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "text", + "content": " and batch size in " + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ba}" + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "text", + "content": " (Eq. 7). " + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "text", + "content": " denotes the constraint for the average area of the localization maps. A larger " + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "text", + "content": " encourages localization map to learn more regions relevant to the classification. Conversely, when " + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "text", + "content": " is small, the region learned by the localization map is small but reliable. Fig. 4 (a) shows that the model behaves insensitive to " + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "text", + "content": ", since GT-known Loc remains stable over a large interval. Batch size affects the calculation of the average area of the localization map within a batch. A sufficiently large batch size can ensure tolerance of area variation between instances. As shown in Fig. 4 (b), when the batch size is larger than 64, the effect of batch on GT-known Loc is limited, with a change of less than " + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 401, + 288, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 575, + 130, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 130, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 130, + 588 + ], + "type": "text", + "content": "4.3. Performance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": "Main Results. Table 1 compares SAT with state-of-the-art methods on CUB-200 [26] and ImageNet [20]. Experiments show that SAT achieves stable and consistent improvements on both benchmark datasets and significantly outperforms all methods using various backbones. Compared to the CNN counterpart ResNet50-ISIC [31], which uses three separate networks, while we use only one network and surpass ResNet50-ISIC by " + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "inline_equation", + "content": "1.13\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "inline_equation", + "content": "3.12\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": " GT-known Loc on the CUB-200 and ImageNet, respectively. These quantitative results demonstrate the superi" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 207, + 547, + 300 + ], + "blocks": [ + { + "bbox": [ + 46, + 166, + 546, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 546, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 546, + 190 + ], + "type": "text", + "content": "Table 3. Localization quality. Comparison of localization accuracy under different IoU thresholds " + }, + { + "bbox": [ + 46, + 166, + 546, + 190 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 46, + 166, + 546, + 190 + ], + "type": "text", + "content": ". Best results are highlighted in bold, second are underlined. The last line indicates the improvement (decrease) of our method over the previous best method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 207, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 306, + 207, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 547, + 300 + ], + "type": "table", + "html": "
MethodVenueBackboneCUB-200OpenImages
pIoUPxAPpIoUPxAP
BagCAMs [49]ECCV22IncepV360.3481.4949.98
DA-WSOL [50]CVPR22ResNet5056.1874.7049.68
BAS [32]CVPR22ResNet5071.0489.2050.72
TS-CAM [8]ICCV21DeiT-S66.5581.4943.65
SCM [2]ECCV22DeiT-S72.0788.15-
SAT (ours)This workDeiT-S77.2189.8759.54
", + "image_path": "e22ce522dbf1cdb4403c9263c58674a1fc7170b4aaa2ac5de0991f375e35f3bd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 303, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 306, + 303, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 545, + 326 + ], + "type": "text", + "content": "Table 4. Segmentation quality of the localization map compared with other SOTA methods. IncepV3: InceptionV3." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 351, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 519 + ], + "type": "text", + "content": "ority of the proposed SAT in terms of simplicity and effectiveness. In addition, on CUB-200 [26], SAT achieves a remarkable performance of " + }, + { + "bbox": [ + 304, + 351, + 545, + 519 + ], + "type": "inline_equation", + "content": "80.96\\% / 98.45\\%" + }, + { + "bbox": [ + 304, + 351, + 545, + 519 + ], + "type": "text", + "content": " on Top-1/GT-known Loc, exceeding the baseline method TS-CAM [8] by " + }, + { + "bbox": [ + 304, + 351, + 545, + 519 + ], + "type": "inline_equation", + "content": "9.66\\% / 10.75\\%" + }, + { + "bbox": [ + 304, + 351, + 545, + 519 + ], + "type": "text", + "content": ". On ImageNet [20], we surpass the SOTA method SCM [2] by a large margin, achieving a " + }, + { + "bbox": [ + 304, + 351, + 545, + 519 + ], + "type": "inline_equation", + "content": "4.33\\%" + }, + { + "bbox": [ + 304, + 351, + 545, + 519 + ], + "type": "text", + "content": " boost in GT-known Loc under the same Deit-S [24] backbone. To further demonstrate the effectiveness of SAT, we visualize the localization maps of our method and TS-CAM in Fig. 3. In addition to generating sharper and more complete localization maps than TS-CAM, SAT demonstrates robust localization ability in various complex and challenging scenarios. More accuracy results and visualizations on six datasets are provided in the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": "Localization Quality. In Table 3, we compare the localization performance with SOTA methods under the MaxBoxAccV2 [5] criterion. Despite the limitation of localization map resolution " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "(14\\times 14)" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": ", we achieve excellent results under different IoU thresholds. Notably, at " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " of 0.7, a strict criterion for the localization quality, the proposed SAT exceeds the previous best method by a significant margin, with " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "7.61\\%" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "7.58\\%" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " improvement on CUB-200 [26] and ImageNet [20], respectively. We further plot the IoU distribution of correct bounding boxes in Fig. 5, following DANet [36]. Compared to the latest SCM [2], with the same Deit-S [24] backbone, we obtain a median IoU gain of " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": " on CUB-200 and ImageNet. The IoU distribution map demonstrates that the localization results generated by SAT have higher localization quality, especially on ImageNet [20], a universal and challenging dataset." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1849" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 165, + 178 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 165, + 178 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 165, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 165, + 178 + ], + "type": "image", + "image_path": "ae691c9f8a77d3f5f7f0440d1677ac92a471fc3f810d56e39e89f45fcbc46cc8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 180, + 287, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 180, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 287, + 202 + ], + "type": "text", + "content": "Figure 6. Density distribution map about IoU and object size on CUB-200, both measured using pixel-level masks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 286, + 177 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 286, + 177 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 286, + 177 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 286, + 177 + ], + "type": "image", + "image_path": "ac92df289a7ca82909855dcfb79c63d5d949bd5d3eef5fe59dfdfdce5e16b3e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 47, + 211, + 171, + 298 + ], + "blocks": [ + { + "bbox": [ + 47, + 211, + 171, + 298 + ], + "lines": [ + { + "bbox": [ + 47, + 211, + 171, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 171, + 298 + ], + "type": "image", + "image_path": "038fdcc0823c5a05e279d7e959d1e4fe346096fae3d1e3994a07bc9c90c917a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 300, + 287, + 322 + ], + "lines": [ + { + "bbox": [ + 47, + 300, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 287, + 322 + ], + "type": "text", + "content": "Figure 7. Threshold sensitivity. GT-known Loc-Threshold curves for different methods on CUB-200 and ImageNet." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 171, + 212, + 285, + 298 + ], + "blocks": [ + { + "bbox": [ + 171, + 212, + 285, + 298 + ], + "lines": [ + { + "bbox": [ + 171, + 212, + 285, + 298 + ], + "spans": [ + { + "bbox": [ + 171, + 212, + 285, + 298 + ], + "type": "image", + "image_path": "a8c7d6ee241cba9dacb0549531af32aa35ffa9aa6a38b2c9a95cd308b66d8f36.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 333, + 201, + 393 + ], + "blocks": [ + { + "bbox": [ + 48, + 333, + 201, + 393 + ], + "lines": [ + { + "bbox": [ + 48, + 333, + 201, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 201, + 393 + ], + "type": "table", + "html": "
Localization GenerationCls Acc. Top-1 Top-5Loc Acc. Top-1 GT-k.
Attention map75.8992.7755.9869.74
Spatial token78.4194.4660.1573.13
+Δ+2.52+1.69+4.17+3.39
", + "image_path": "2fcc6aea52b2a539aef920aa7bfe41a45624fdac9b546812ab83a76c22898939.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 205, + 333, + 287, + 394 + ], + "blocks": [ + { + "bbox": [ + 205, + 333, + 287, + 394 + ], + "lines": [ + { + "bbox": [ + 205, + 333, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 205, + 333, + 287, + 394 + ], + "type": "image", + "image_path": "86c419755180c6a119cce9b7893a0507d0f8a8407588a56f719be568b0ae8530.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 396, + 287, + 440 + ], + "lines": [ + { + "bbox": [ + 47, + 396, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 287, + 440 + ], + "type": "text", + "content": "Table 5. Accuracy of different baselines with the same losses. We follow the idea of TS-CAM to extract the attention maps (before softmax and adding a sigmoid function) to synthesize the localization map, and apply the proposed losses to it." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "text", + "content": "Segmentation Quality. We evaluate the segmentation performance of SAT on CUB-200 [26] and OpenImages [5] since pixel-level labels are available during evaluation. As shown in Table 4, the proposed SAT substantially exceeds all methods in pIoU and PxAP metrics. Especially compared to the baseline method TS-CAM [8], we achieve a remarkable increase of " + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "inline_equation", + "content": "10.66\\%" + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "inline_equation", + "content": "15.89\\%" + }, + { + "bbox": [ + 46, + 461, + 287, + 652 + ], + "type": "text", + "content": " pIoU on CUB-200 and OpenImages. To further analyze the segmentation quality, we draw the density distribution map about IoU and object size on CUB-200 in Fig. 6. It can be observed the proposed SAT achieves an overall improvement and more stable performance on arbitrary sized objects compared to TS-CAM. Nevertheless, there still exists an imbalance segmentation ability of different sized objects. This is mainly due to the limitation of the localization map resolution, which hinders the acquisition of fine segmentation results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Threshold Sensitivity. Fig. 7 illustrates the curves of GT-known Loc under different thresholds on CUB-200 and ImageNet. Obviously, the curves corresponding to the proposed SAT entirely cover the curves of TS-CAM [8] and SCM [2], indicating that SAT performs stable and achieves" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 70, + 544, + 140 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 544, + 140 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 544, + 140 + ], + "type": "table", + "html": "
ModelTop-1 ClsGT-known Loc
TS-CAMSCMOurs (+Δ)TS-CAMSCMOurs (+Δ)
Deit-T72.9-76.2 (+3.3)86.491.895.0 (+3.2)
Deit-S80.377.182.1 (+1.8)87.796.698.5 (+1.9)
Deit-B83.2-85.3 (+2.1)83.393.898.1 (+4.3)
Conf-S81.0-82.4 (+1.4)94.196.197.9 (+1.8)
", + "image_path": "a7f5ae73763c620ad31f4ef2f2a7f5ac0a7ad7ad11b4af200936610db5166e58.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 176, + 430, + 249 + ], + "blocks": [ + { + "bbox": [ + 306, + 143, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 306, + 143, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 545, + 166 + ], + "type": "text", + "content": "Table 6. Generalizability on CUB-200. " + }, + { + "bbox": [ + 306, + 143, + 545, + 166 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 306, + 143, + 545, + 166 + ], + "type": "text", + "content": " denotes the improvement of SAT over the previous best method. Conf-S: Conformer-S." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 176, + 430, + 249 + ], + "lines": [ + { + "bbox": [ + 307, + 176, + 430, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 430, + 249 + ], + "type": "table", + "html": "
MethodLoc Acc. Top-1 Top-5 GT-k.
(a)ORNet73.3984.3188.38
(b)BAS49.5960.6863.76
(c)BAS*75.1888.6692.87
(d)Ours80.9694.1398.45
", + "image_path": "dc3f914981c1d747ff68af0f5e7a730cabe85867ef56f3c20302b9433605db37.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 437, + 174, + 545, + 249 + ], + "blocks": [ + { + "bbox": [ + 437, + 174, + 545, + 249 + ], + "lines": [ + { + "bbox": [ + 437, + 174, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 437, + 174, + 545, + 249 + ], + "type": "table", + "html": "
Image(a)(b)
label(c)(d)
", + "image_path": "776a76b26daeed25842af12890170a3ff43fe9aa6cbc491fba9a810e7035223e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 251, + 545, + 296 + ], + "lines": [ + { + "bbox": [ + 306, + 251, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 545, + 296 + ], + "type": "text", + "content": "Table 7. SAT vs FPM-based methods on Deit-S backbone. For ORNet, we replace the classification network with Deit-S. For BAS, we insert the generator in the 10-th self-attention module. * denotes replacing the original losses with our proposed losses." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 316, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 375 + ], + "type": "text", + "content": "better results than other methods at arbitrary threshold. Besides, SAT is not sensitive to thresholds and maintains high localization accuracy over a large threshold interval. This phenomenon shows that the localization map generated by SAT has high confidence and excellent visualization." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 376, + 546, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 507 + ], + "type": "text", + "content": "Generalizability. To validate the generalizability of the proposed SAT, we implement experiments on different Transformers and various scales. As reported in Table 6, the proposed approach can be well applied to Deit [24] and Conformer [19] on CUB-200 [26]. On Deit-T/Deit-B, we achieve " + }, + { + "bbox": [ + 304, + 376, + 546, + 507 + ], + "type": "inline_equation", + "content": "3.2\\% / 4.3\\%" + }, + { + "bbox": [ + 304, + 376, + 546, + 507 + ], + "type": "text", + "content": " GT-known Loc improvement compared to SCM [2], and " + }, + { + "bbox": [ + 304, + 376, + 546, + 507 + ], + "type": "inline_equation", + "content": "3.3\\% / 2.1\\%" + }, + { + "bbox": [ + 304, + 376, + 546, + 507 + ], + "type": "text", + "content": " Top-1 classification accuracy gain over TS-CAM [8]. Besides, SAT also achieves state-of-the-art results on Conformer-S. Quantitative experiments show that the proposed SAT achieves excellent performance in both tasks on various transformer backbones." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 516, + 382, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 516, + 382, + 527 + ], + "spans": [ + { + "bbox": [ + 306, + 516, + 382, + 527 + ], + "type": "text", + "content": "4.4. Discussions" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "text", + "content": "Why spatial token-based structure is critical? To clarify this question, we follow TS-CAM [8] to construct a attention map-based baseline and apply the proposed losses on it for comparison. As shown in Table 5, the attention map-based method substantially reduces the classification and localization accuracy, with " + }, + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "inline_equation", + "content": "2.52\\%" + }, + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "inline_equation", + "content": "3.39\\%" + }, + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "text", + "content": " decrease in Top-1 Cls and GT-known Loc, respectively. To explore how this structural difference affects the accuracy, we plot the curves of spatial losses during training in Table 5. It can be noticed that the potential optimization conflicts of attention map-based method will lead to poor convergence of losses and insufficient learning, thus limiting the model to benefit from the losses. Experimental results show that the proposed spatial token-based structure can release the potential of the model by more efficient localization learn" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1850" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 286, + 180 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 286, + 180 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 286, + 180 + ], + "type": "image", + "image_path": "c85d13df1dffc9ca54dee64811002a48211e2aff89207f0e34c4924e6f631eca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 287, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 227 + ], + "type": "text", + "content": "Figure 8. Few-shot learning. Box plot of the k-shot localization performance. For each k-shot setting, we repeat the experiment 10 times to randomly select k images as training data, and record the experimental results. The average results are shown in the figure." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 247, + 286, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 286, + 270 + ], + "type": "text", + "content": "ing and avoiding optimization conflicts between two tasks, which are essential for the accuracy improvement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 271, + 286, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 286, + 462 + ], + "type": "text", + "content": "To further verify the effectiveness of the proposed structure in generating localization map, we reproduce two FPM-based methods ORNet [33] and BAS [32] on the Deit-S. Both of them produce localization maps by a generator. As shown in Table 7, first, we note that the original BAS is not suitable for the transformer due to the skip-connection in transformer causing inaccurate evaluation of the masking effect. Second, both FPM-based methods suffer from incomplete localization problem. This may be because they can only mask the input of a specific layer, resulting in overfitting the features that the specific feature layer focuses on. While the proposed structure can extract localization knowledge from different layers in an efficient form, thus improving localization ability and robustness. Quantitative and qualitative results demonstrate the effectiveness of SAT in terms of localization map generation and learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 463, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 286, + 641 + ], + "type": "text", + "content": "Data-efficiency and tuning-efficiency. To validate the data-efficiency of the proposed method, we evaluate the localization performance of SAT and TS-CAM [8] under few-shot setting in Fig. 8. Compared to TS-CAM, SAT achieves promising results with only a few training data. Even in the extreme setting of only 1 training image per class, we exceed TS-CAM trained with full data by " + }, + { + "bbox": [ + 46, + 463, + 286, + 641 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 46, + 463, + 286, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 463, + 286, + 641 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 46, + 463, + 286, + 641 + ], + "type": "text", + "content": " GT-known Loc on CUB-200 and ImageNet, respectively. The box plot verifies the effectiveness and efficiency of the proposed SAT that implements localization task by an extra lightweight spatial token and enables high degree of network parameter sharing, thus requiring only a few data to achieve excellent localization results. In contrast, TS-CAM requires large amounts of data to fine-tune the classification feature maps to serve the localization task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": "To explore the tuning-efficiency of SAT, we freeze different parts of the network and evaluate the GT-known Loc of the proposed method under few-shot setting. All experiments are performed with ImageNet pre-trained weights. As shown in Table 8, SAT can still obtain competitive results when freezing most parameters of the pre-trained" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 152 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 152 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 152 + ], + "type": "table", + "html": "
Fro. rateCUB-200 [26]Fro. rateImageNet [20]
13510Full13510Full
(a)0%91.395.196.197.398.50%70.971.972.672.973.1
(b)24%93.595.996.897.598.121%71.672.372.672.873.1
(c)48%96.496.797.097.498.042%71.872.172.472.772.9
(d)71%93.894.694.995.295.964%71.371.471.771.972.2
(e)91%92.192.592.893.193.481%70.971.171.271.371.4
", + "image_path": "26f862eae9458d6fff5f49c463e353d5ba759edc5b156a56475b9a33e1158a62.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 545, + 210 + ], + "type": "text", + "content": "Table 8. Frozen rate w.r.t. few-shot learning. The frozen parts are as follows: (a) None. (b) Attention layer of transformer blocks. (c) MLP layer of transformer blocks. (d) Transformer blocks. (e) Position embedding, projection, transformer blocks, and MLP layer of spatial aware transformer blocks. Fro.: Frozen." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 217, + 545, + 320 + ], + "blocks": [ + { + "bbox": [ + 307, + 217, + 545, + 320 + ], + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 320 + ], + "type": "image", + "image_path": "d4923b6b770aa9f5021a52caad3ba2b6a0daaa2886edf1a48d652e787339a05b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 323, + 520, + 333 + ], + "lines": [ + { + "bbox": [ + 330, + 323, + 520, + 333 + ], + "spans": [ + { + "bbox": [ + 330, + 323, + 520, + 333 + ], + "type": "text", + "content": "Figure 9. Localization error analysis on CUB-200." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 462 + ], + "type": "text", + "content": "classification model. It can be analyzed that the task-specific spatial token can be well adapted to the classification features and plays a major role in the localization learning, which allows efficient implementation of the localization task with fewer parameters to be fine-tuned. Experiments demonstrate both the data-efficiency and the tuning-efficiency of SAT. When freezing " + }, + { + "bbox": [ + 304, + 354, + 545, + 462 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 304, + 354, + 545, + 462 + ], + "type": "text", + "content": " of the parameters and using only 1 image per class, SAT exceeds the SOTA method SCM [2] by " + }, + { + "bbox": [ + 304, + 354, + 545, + 462 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 354, + 545, + 462 + ], + "type": "text", + "content": " GT-known Loc on ImageNet." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 470, + 383, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 383, + 481 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 383, + 481 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 489, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 657 + ], + "type": "text", + "content": "As shown in Fig. 9, we count all localization errors (90 images) on CUB-200 test set (5,794 images) and classify the error causes into six categories, including object occlusion, localization more, water reflection, localization part, multiple instances, and label error. Among them, object occlusion causes the object to be split into two or more parts, resulting in incomplete localization results. Localization more is usually due to the positive effect of co-occurrence context on classification, leading to localizing confounding background regions. Water reflection is an inherent problem of WSOL and difficult to be solved with only image-level labels. Therefore, future work needs to consider more the interaction between object and context to overcome the problems of object occlusion and localization more." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 669, + 378, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 378, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 378, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "This paper proposes to introduce a spatial-aware token (SAT) specific to the localization task, instead of synthesiz-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1851" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "ing the classification feature map, thus avoiding optimization conflicts between classification and localization tasks. To this end, we construct a spatial-query attention (SQA) module to generate localization map and extract localization knowledge from the classification task. Meanwhile, two spatial constraints, including batch area loss and regularization loss, are designed to compensate for the insufficient supervision. Extensive experiments on multiple benchmarks verify the effectiveness and efficiency of the proposed SAT that surpasses previous methods by a large margin." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 251 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by National Key R&D Program of China under Grant 2020AAA0105700, National Natural Science Foundation of China (NSFC) under Grants 62225207, U19B2038 and 62121002." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 262, + 107, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 262, + 107, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 107, + 274 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 281, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 281, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 326 + ], + "type": "text", + "content": "[1] Wonho Bae, Junhyug Noh, and Gunhee Kim. Rethinking class activation mapping for weakly supervised object localization. In European Conference on Computer Vision, pages 618-634. Springer, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 327, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 288, + 369 + ], + "type": "text", + "content": "[2] Haotian Bai, Ruimao Zhang, Jiong Wang, and Xiang Wan. Weakly supervised object localization via transformer with implicit spatial calibration. arXiv preprint arXiv:2207.10447, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 370, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 288, + 426 + ], + "type": "text", + "content": "[3] Zhiwei Chen, Changan Wang, Yabiao Wang, Guannan Jiang, Yunhang Shen, Ying Tai, Chengjie Wang, Wei Zhang, and Liujuan Cao. Lctr: On awakening the local continuity of transformer for weakly supervised object localization. arXiv preprint arXiv:2112.05291, 2021. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 426, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 480 + ], + "type": "text", + "content": "[4] Junsuk Choe, Seungho Lee, and Hyunjung Shim. Attention-based dropout layer for weakly supervised single object localization and semantic segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(12):4256-4271, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 481, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 536 + ], + "type": "text", + "content": "[5] Junsuk Choe, Seong Joon Oh, Seungho Lee, Sanghyuk Chun, Zeynep Akata, and Hyunjung Shim. Evaluating weakly supervised object localization methods right. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3133-3142, 2020. 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 537, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 537, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 53, + 537, + 288, + 581 + ], + "type": "text", + "content": "[6] Junsuk Choe and Hyunjung Shim. Attention-based dropout layer for weakly supervised object localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2219-2228, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 582, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 582, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 53, + 582, + 288, + 647 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 712 + ], + "type": "text", + "content": "[8] Wei Gao, Fang Wan, Xingjia Pan, Zhiliang Peng, Qi Tian, Zhenjun Han, Bolei Zhou, and Qixiang Ye. Ts-cam: Token semantic coupled attention map for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2886-2895, 2021. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 128 + ], + "type": "text", + "content": "[9] Guangyu Guo, Junwei Han, Fang Wan, and Dingwen Zhang. Strengthen learning tolerance for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7403-7412, 2021. 1, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 129, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 185 + ], + "type": "text", + "content": "[10] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Amalgamating knowledge from heterogeneous graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15709-15718, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 186, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 251 + ], + "type": "text", + "content": "[11] Eunji Kim, Siwon Kim, Jungbeom Lee, Hyunwoo Kim, and Sungroh Yoon. Bridging the gap between classification and localization for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14258-14267, 2022. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 253, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 547, + 297 + ], + "type": "text", + "content": "[12] Jeesoo Kim, Junsuk Choe, Sangdoo Yun, and Nojun Kwak. Normalization matters in weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3427-3436, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 298, + 547, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 547, + 354 + ], + "type": "text", + "content": "[13] Jungbeom Lee, Eunji Kim, Jisoo Mok, and Sungroh Yoon. Anti-adversarily manipulated attributions for weakly supervised semantic segmentation and object localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 355, + 547, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 547, + 411 + ], + "type": "text", + "content": "[14] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Dechao Meng, and Qingming Huang. Adaptive reconstruction network for weakly supervised referring expression grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2611-2620, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 411, + 547, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 547, + 477 + ], + "type": "text", + "content": "[15] Weizeng Lu, Xi Jia, Weicheng Xie, Linlin Shen, Yicong Zhou, and Jinming Duan. Geometry constrained weakly supervised object localization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 481-496. Springer, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 478, + 547, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 547, + 534 + ], + "type": "text", + "content": "[16] Jinjie Mai, Meng Yang, and Wenfeng Luo. Erasing integrated learning: A simple yet effective approach for weakly supervised object localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8766-8775, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 535, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 547, + 590 + ], + "type": "text", + "content": "[17] Meng Meng, Tianzhu Zhang, Qi Tian, Yongdong Zhang, and Feng Wu. Foreground activation maps for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3385-3395, 2021. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 590, + 547, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 547, + 658 + ], + "type": "text", + "content": "[18] Xingjia Pan, Yingguo Gao, Zhiwen Lin, Fan Tang, Weiming Dong, Haolei Yuan, Feiyue Huang, and Changsheng Xu. Unveiling the potential of structure preserving for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11642-11651, 2021. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 658, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 714 + ], + "type": "text", + "content": "[19] Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, and Qixiang Ye. Conformer: Local features coupling global representations for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 367-376, 2021. 7" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1852" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[20] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4, 5, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 172 + ], + "type": "text", + "content": "[21] Tal Shaharabany, Yoad Tewel, and Lior Wolf. What is where by looking: Weakly-supervised open-world phrase-grounding without text inputs. arXiv preprint arXiv:2206.09358, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "type": "text", + "content": "[22] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous for weakly-supervised object and action localization. In 2017 IEEE international conference on computer vision (ICCV), pages 3544-3553. IEEE, 2017. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 270 + ], + "type": "text", + "content": "[23] Ganchao Tan, Daqing Liu, Meng Wang, and Zheng-Jun Zha. Learning to discretely compose reasoning module networks for video captioning. arXiv preprint arXiv:2007.09049, 2020.1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 271, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 288, + 327 + ], + "type": "text", + "content": "[24] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 1, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "type": "text", + "content": "[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 403 + ], + "type": "text", + "content": "[26] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 403, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 288, + 449 + ], + "type": "text", + "content": "[27] Hao Wang, Zheng-Jun Zha, Xuejin Chen, Zhiwei Xiong, and Jiebo Luo. Dual path interaction network for video moment localization. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4116-4124, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 492 + ], + "type": "text", + "content": "[28] Hao Wang, Zheng-Jun Zha, Liang Li, Xuejin Chen, and Jiebo Luo. Semantic and relation modulation for audiovisual event localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 548 + ], + "type": "text", + "content": "[29] Hao Wang, Zheng-Jun Zha, Liang Li, Dong Liu, and Jiebo Luo. Structured multi-level interaction network for video moment localization via language query. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7026-7035, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 288, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 603 + ], + "type": "text", + "content": "[30] Jun Wei, Qin Wang, Zhen Li, Sheng Wang, S Kevin Zhou, and Shuguang Cui. Shallow feature matters for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5993-6001, 2021. 1, 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "type": "text", + "content": "[31] Jun Wei, Sheng Wang, S Kevin Zhou, Shuguang Cui, and Zhen Li. Weakly supervised object localization through inter-class feature similarity and intra-class appearance consistency. In European Conference on Computer Vision, pages 195-210. Springer, 2022. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "text", + "content": "[32] Pingyu Wu, Wei Zhai, and Yang Cao. Background activation suppression for weakly supervised object localization. arXiv preprint arXiv:2112.00580, 2021. 2, 5, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[33] Jinheng Xie, Cheng Luo, Xiangping Zhu, Ziqi Jin, Weizeng Lu, and Linlin Shen. Online refinement of low-level feature" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 106 + ], + "type": "text", + "content": "based activation map for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 132-141, 2021. 2, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 106, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 160 + ], + "type": "text", + "content": "[34] Jinheng Xie, Jianfeng Xiang, Junliang Chen, Xianxu Hou, Xiaodong Zhao, and Linlin Shen. Contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. arXiv preprint arXiv:2203.13505, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 160, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 160, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 160, + 545, + 227 + ], + "type": "text", + "content": "[35] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Rui-Wei Zhao, Tao Zhang, Xuequan Lu, and Shang Gao. Cream: Weakly supervised object localization via class re-activation mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9437–9446, 2022. 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 227, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 227, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 545, + 282 + ], + "type": "text", + "content": "[36] Haolan Xue, Chang Liu, Fang Wan, Jianbin Jiao, Xi-angyang Ji, and Qixiang Ye. Danet: Divergent activation for weakly supervised object localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6589–6598, 2019. 1, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 282, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 545, + 338 + ], + "type": "text", + "content": "[37] Seunghan Yang, Yoonhyung Kim, Youngeun Kim, and Changick Kim. Combinational class activation maps for weakly supervised object localization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2941-2949, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 338, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 382 + ], + "type": "text", + "content": "[38] Tianhao Yang, Zheng-Jun Zha, and Hanwang Zhang. Making history matter: History-advantage sequence training for visual dialog. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2561–2569, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "text", + "content": "[39] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. Advances in neural information processing systems, 35:25739-25753, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 415, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 471 + ], + "type": "text", + "content": "[40] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6023-6032, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "text", + "content": "[41] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Exploring figure-ground assignment mechanism in perceptual organization. Advances in Neural Information Processing Systems, 35:17030-17042, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 558 + ], + "type": "text", + "content": "[42] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10):2472-2500, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 558, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 545, + 613 + ], + "type": "text", + "content": "[43] Chen-Lin Zhang, Yun-Hao Cao, and Jianxin Wu. Rethinking the route towards weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13460-13469, 2020. 1, 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "type": "text", + "content": "[44] Dingwen Zhang, Junwei Han, Gong Cheng, and Ming-Hsuan Yang. Weakly supervised object localization and detection: A survey. IEEE transactions on pattern analysis and machine intelligence, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[45] Xiaolin Zhang, Yunchao Wei, Jiashi Feng, Yi Yang, and Thomas S Huang. Adversarial complementary learning for weakly supervised object localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1325-1334, 2018. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1853" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 318 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[46] Xiaolin Zhang, Yunchao Wei, Guoliang Kang, Yi Yang, and Thomas Huang. Self-produced guidance for weakly-supervised object localization. In Proceedings of the European conference on computer vision (ECCV), pages 597-613, 2018. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 185 + ], + "type": "text", + "content": "[47] Xiaolin Zhang, Yunchao Wei, and Yi Yang. Inter-image communication for weakly supervised localization. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 271-287. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "text", + "content": "[48] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 283 + ], + "type": "text", + "content": "[49] Lei Zhu, Qian Chen, Lujia Jin, Yunfei You, and Yanye Lu. Bagging regional classification activation maps for weakly supervised object localization. arXiv preprint arXiv:2207.07818, 2022. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 318 + ], + "type": "text", + "content": "[50] Lei Zhu, Qi She, Qian Chen, Yunfei You, Boyu Wang, and Yanye Lu. Weakly supervised object localization as domain adaption. arXiv preprint arXiv:2203.01714, 2022. 5, 6" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1854" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spatially and Spectrally Consistent Deep Functional Maps/38616010-b1fd-470b-8376-180cbf13b5f1_content_list.json b/2023/Spatially and Spectrally Consistent Deep Functional Maps/38616010-b1fd-470b-8376-180cbf13b5f1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6593411d534459bbc7ce413b3c601d254c4aeca6 --- /dev/null +++ b/2023/Spatially and Spectrally Consistent Deep Functional Maps/38616010-b1fd-470b-8376-180cbf13b5f1_content_list.json @@ -0,0 +1,1769 @@ +[ + { + "type": "text", + "text": "Spatially and Spectrally Consistent Deep Functional Maps", + "text_level": 1, + "bbox": [ + 189, + 130, + 779, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingze Sun1 Shiwei Mao1 Puhua Jiang1,2 Maks Ovsjanikov3 Ruqi Huang1* \n1Tsinghua Shenzhen International Graduate School, China \n2Peng Cheng Laboratory, China \n3LIX, École polytechnique, IP Paris, France", + "bbox": [ + 148, + 178, + 825, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 287, + 313, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cycle consistency has long been exploited as a powerful prior for jointly optimizing maps within a collection of shapes. In this paper, we investigate its utility in the approaches of Deep Functional Maps, which are considered state-of-the-art in non-rigid shape matching. We first justify that under certain conditions, the learned maps, when represented in the spectral domain, are already cycle consistent. Furthermore, we identify the discrepancy that spectrally consistent maps are not necessarily spatially, or point-wise, consistent. In light of this, we present a novel design of unsupervised Deep Functional Maps, which effectively enforces the harmony of learned maps under the spectral and the point-wise representation. By taking advantage of cycle consistency, our framework produces state-of-the-art results in mapping shapes even under significant distortions. Beyond that, by independently estimating maps in both spectral and spatial domains, our method naturally alleviates over-fitting in network training, yielding superior generalization performance and accuracy within an array of challenging tests for both near-isometric and non-isometric datasets. Codes are available at https://github.com/rqhuang88/Spatially-and-Spectrally-Consistent-Deep-Functional-Maps.", + "bbox": [ + 75, + 319, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Non-rigid shape matching is a fundamental task in shape analysis, playing a pivotal role in a wide spectrum of applications including texture transfer [1], volume parameterization [31], statistical shape analysis [2, 6], to name a few. In order to establish correspondences between shapes undergoing non-rigid deformations, early approaches [46, 4, 47] focus on designing hand-crafted features by exploiting geometric invariance. Nowadays, data-driven techniques have been widely adopted to learn features in a more flexible way, leading to significant improvements regarding accu", + "bbox": [ + 75, + 723, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb7ef1846ad012ddb4f92b9a3fb2f9b9ff9b2cc6e8a344dda77a6cd6675d878f.jpg", + "image_caption": [ + "Figure 1. We train unsupervised non-rigid shape matching pipelines on the challenging non-isometric dataset DT4D-H, which involves 8 categories of humanoid shapes undergoing significant distortions. Correspondences are visualized by texture transfer. Our method (bottom row) outperforms the state-of-the-art, AttentiveFMaps [28] (middle row) and UDMSM [9] (top row), by a large margin." + ], + "image_footnote": [], + "bbox": [ + 516, + 286, + 924, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "racy, efficiency, and robustness [10].", + "bbox": [ + 500, + 534, + 743, + 550 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A noticeable trend among the learning-based shape matching approaches is based on the formalism of Deep Functional Maps (DFM), pioneered by the FMNet [30]. Functional maps [37], as a spectral map representation, allows to encode maps into compact matrices and to express desirable map priors (e.g., area-preservation, isometry, bijectivity) in simple algebraic forms. In particular, by learning features that optimize functional map priors, several unsupervised DFM frameworks [19, 43, 18, 13, 9, 28] have been proposed and, remarkably, achieve even comparable results with respect to the supervised counterparts. On the other hand, while the majority of DFM works focus on exploring local, pairwise-level map priors, there is relatively less attention (with the exception of [9]) paid to the problem of injecting global map priors into DFM pipelines.", + "bbox": [ + 496, + 551, + 892, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a global map prior, cycle consistency has long been used as a powerful regularizer for jointly optimizing maps among shapes, both under point-wise [22] and functional [21, 24] map representation. In this paper, we exploit the utility of cycle consistency within DFM frameworks and propose a novel two-branch unsupervised DFM framework, which promotes cycle consistency in both spectral and spatial domains. We first re-examine the generic DFM frame", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: ruqihuang@sz.tsinghua.edu.cn", + "bbox": [ + 94, + 886, + 388, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14497", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "work and outline a theoretical condition, based on the residuals of a linear solver used in functional map estimation, that guarantees spectrally cycle consistency in DFM over training shapes. Then we identify the discrepancy that spectrally consistent maps are not necessarily spatially, or pointwise, consistent. In light of this, we leverage our theoretical insight to construct a latent shape in DFM and employ this construction to formulate a novel two-branch design of unsupervised DFM. We align each shape's spectral embedding, represented as the eigenbasis of the Laplace-Beltrami operator, to a canonical space given by the constructed universal latent shape. We then compute point-wise maps in this canonical embedding domain, which promotes consistency across computed point-wise maps. In the end, the point-wise maps are converted to functional maps, which are expected to agree with the ones from the original spectral branch. Overall, our two-branch DFM takes advantage of the justified spectrally cycle consistency and further enhances it to spatially cycle consistent.", + "bbox": [ + 76, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on several non-rigid shape matching benchmarks, and demonstrate that our method achieves superior performance over existing methods, especially in the presence of significant distortions among shapes (see, e.g., Fig. 1). We also observe that our method, by enforcing the harmony of spatial and spectral map representations, reduces over-fitting during training, which leads to remarkable generalization performance within an array of challenging tests. Finally, our two-branch design can be easily plugged into any existing DFM framework, and we show evidence that it achieves significant performance gains upon both recent and early DFM approaches [43] with marginal computational burden.", + "bbox": [ + 76, + 378, + 472, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are as follows: (1) We perform theoretical analysis on spectrally cycle consistency of DFM frameworks; (2) We formulate a simple yet effective two-branch design of unsupervised DFM based on our theoretical justification, which introduces spatially cycle consistency. (3) We prove the effectiveness of our method through an array of challenging non-rigid shape matching tasks, demonstrating improvements upon existing methods in terms of accuracy, consistency, and generalization performance.", + "bbox": [ + 76, + 574, + 472, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 739, + 218, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Functional Maps Our method is built upon functional map representation, introduced in [37] and then significantly extended in follow-up works (see, e.g., [38]). The key idea is to encode shape correspondences as transformations between the respective spectral embeddings, which are represented by compact matrices by using reduced eigenbasis. The functional maps pipeline has been further improved in accuracy, efficiency, and robustness by many recent works including [26, 21, 8, 42, 36].", + "bbox": [ + 76, + 763, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Functional Maps In contrast to axiomatic approaches that use hand-crafted features [46, 4], the deep functional maps approach, pioneered by FMNet [30], aims to learn the optimal features from data. FMNet contributes several key designs of DFM: (1) it leverages Siamese network to conduct learning in a set of shape pairs; (2) it advocates refining the input descriptors with non-linear transformations. FMNet is then supervised by labeled maps to learn optimal features.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of learning from labeled maps, unsupervised approaches [19, 43] demonstrate that it is sufficient to learn from geometric map priors. More recently, with the development of robust mesh feature extractors [44], more frameworks [9, 28, 11, 3] are proposed to learn directly from geometry, yielding state-of-the-art performance.", + "bbox": [ + 496, + 229, + 893, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cycle Consistency Cycle consistency has long been used as a strong prior for joint map optimization among a collection of shapes. Axiomatic approaches detect and eliminate inconsistent cycles using consistency constraints [25, 52, 35, 51, 18], as well as associate cycle consistency with low-rank properties of matrices encoding map networks [22, 50, 27]. Related to the latter, the matrix nature of the functional maps enables convenient access to map composition, which naturally bridges the functional map framework and consistent map refinement techniques [49, 21, 23, 45]. Some recent learning-based approach [20] also incorporate cycle consistency in the pipeline. It is worth noting, though, the above works all utilize cycle consistency as a prior for test-time optimization, which depends on test shape collection and initial maps. Contrastingly, our approach exploits cycle consistency during training to boost feature learning and poses no constraint on a test.", + "bbox": [ + 496, + 321, + 895, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From this viewpoint, UDMSM [9] is perhaps the most relevant work to ours, as both construct a latent shape during training to inject cycle consistency. The key difference between their work and ours, though, is how the latent shape is constructed. In [9], the authors propose to construct a universal shape in the learned feature space and establish pointwise maps from real shapes to the universal one by training a classifier. As shown in Sec.5, though this construction leads to great performance on matching near-isometric shapes, it suffers from large shape variability in mapping non-isometric ones. On the other hand, we leverage spectral information in estimating point-wise maps. As a result, our approach benefits from more direct usage of intrinsic geometric information encoded in the spectral embedding, yielding better generalization performance.", + "bbox": [ + 496, + 580, + 895, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dual Map Representations Thanks to the inherent connection and efficient conversion between point-wise and functional maps, it has long been observed that jointly estimating both map representations can improve the mapping quality. For instance, in the original work [37], the authors have already proposed to apply an ICP-like technique on", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14498", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "functional maps. More recent advances take advantage of the multi-scale properties of the eigenbasis of the Laplace-Beltrami operator. In the works [34, 24, 13, 15], conversions are done between spatial domain and a series of spectral domains spanned by eigenfunctions of increasing dimensions. This idea has also been incorporated into DFM. In AttentiveFMaps [28], the authors propose to fuse functional maps of different dimensions by converting them into the common spatial domain. This fusion technique in turn allows training an attention network to dynamically choose the optimal spectral resolution.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above methods enforce consistency across map representations by iterative projections, which is computationally heavy. By exploiting our theoretical insight on the spectral cycle consistency of DFM, we only introduce a marginal computational overhead to a standard DFM trained on a single spectral resolution. Our method is much lighter and simpler but also shows superior performance in accuracy and generalization.", + "bbox": [ + 75, + 257, + 472, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Cycle Consistency of Deep Functional Maps", + "text_level": 1, + "bbox": [ + 76, + 392, + 468, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first briefly review the general deep functional maps pipeline. Then we present a theoretical analysis of its cycle consistency.", + "bbox": [ + 75, + 417, + 468, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Deep Functional Maps", + "text_level": 1, + "bbox": [ + 76, + 472, + 287, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We assume to be given a pair of deformable shapes $S_{1}$ and $S_{2}$ , which are discretized as triangular meshes of $n_{1}$ and $n_{2}$ vertices, respectively. The generic deep functional map pipeline introduced in [30] learns a map that spectrally aligns the shapes through the following four steps:", + "bbox": [ + 75, + 496, + 468, + 571 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Compute the leading $k$ eigenfunctions of the Laplace-Beltrami operator on each shape, which can be treated as a high-dimensional spectral embedding of the respective shape. The eigenfunctions are stored as matrices $\\Phi_{i}\\in \\mathbb{R}^{n_{i}\\times k},i = 1,2$", + "2. Instantiate a feature extractor network, $\\mathcal{F}_{\\Theta}$ , where $\\Theta$ denotes the set of learnable parameters. By feeding forward shapes through $\\mathcal{F}_{\\Theta}$ , descriptors are obtained and expected to be approximately preserved by the underlying map. In general, we denote $\\mathbf{G}_i = \\mathcal{F}_{\\Theta}(S_i) \\in \\mathbb{R}^{n_i \\times d}$ , $i = 1, 2$ , where $d$ is the predefined number of descriptors. They are then projected onto the eigenbasis above, resulting in a couple of coefficient matrices $\\mathbf{A}_i = \\boldsymbol{\\Phi}_i^\\dagger \\mathbf{G}_i \\in \\mathbb{R}^{k \\times d}$ , $i = 1, 2$ .", + "3. Estimate the optimal functional map, $\\mathbf{C}^*$ , by solving the following linear system:" + ], + "bbox": [ + 88, + 584, + 468, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} ^ {*} = \\arg \\min _ {\\mathbf {C}} E _ {\\mathrm {d e s c}} (\\mathbf {C}) + E _ {\\mathrm {r e g}} (\\mathbf {C}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 835, + 468, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{\\mathrm{desc}}(\\mathbf{C}) = \\| \\mathbf{CA}_1 - \\mathbf{A}_2\\|^2$ , and $E_{\\mathrm{reg}}(\\mathbf{C})$ is the regularization term promoting structural properties", + "bbox": [ + 107, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/143874733320dcf517fd437577aa58b5339c02e0b3042929de4bb47ed9008857.jpg", + "image_caption": [ + "Figure 2. Overview of our two-branch DFM network. The part colored in black corresponds to a standard DFM, we introduce a novel branch, colored in red, that estimates maps from spatial perspective, our loss function is defined as Eqn.6. $\\Phi_i,\\Phi_i^\\dagger$ indicate projection of the regarding features into the spatial and spectral domain, respectively. See more details in Section 4.2." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 890, + 202 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "of $\\mathbf{C}$ , e.g., enforcing $\\mathbf{C}$ to be commutative with the Laplace-Beltrami operators and to be orthogonal. In particular, we let $E_{\\mathrm{total}}(\\mathbf{C}) = E_{\\mathrm{desc}}(\\mathbf{C}) + E_{\\mathrm{reg}}(\\mathbf{C})$", + "4. Convert the estimated functional map $\\mathbf{C}$ to a pointwise map by conducting nearest neighbor search between the rows of $\\Phi_1\\mathbf{C}$ and that of $\\Phi_{2}$ ." + ], + "bbox": [ + 511, + 295, + 890, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The feature extractor is then trained on a set of training shape pairs, which typically enumerates all pairs from some given training set. The sub-optimization problem in Step (3) is generally decomposed into two parts to circumvent using an iterative solver. Namely, a closed-form solution is obtained either by a least-square estimation w.r.t $E_{\\mathrm{desc}}(\\mathbf{C})$ [30], or by solving a sequence of linear systems involving Laplacian commutativity as well [11]. Both correspond to the FMreg layer shown in Fig. 2. The rest of $E_{\\mathrm{reg}}(\\mathbf{C})$ is then set as the structural loss, corresponding to $\\mathcal{L}(\\mathbf{C})$ in Fig. 2.", + "bbox": [ + 496, + 398, + 892, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cycle Consistency", + "text_level": 1, + "bbox": [ + 500, + 574, + 674, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by giving a formal definition of cycle consistency. We assume to be given a collection of shapes $S = \\{S_i\\}_{i=1}^n$ and a set of maps $\\mathcal{T} = \\{T_{ij}\\}_{i,j \\in [1..n]}$ . We call $\\mathcal{T}$ cycle consistent if for any shape $S_i$ , and any closed path $(i, i_1, i_2, \\dots, i_k, i)$ , the map composition along this path $T_{ii} = T_{i_k i} \\circ \\dots \\circ T_{i_1 i_2} \\circ T_{ii_1}$ is an identity map on $S_i$ . We can similarly define cycle consistency on functional maps, the only difference is that map composition is given by matrix multiplication and we require the final functional map $C_{ii}$ to be an identity matrix.", + "bbox": [ + 496, + 597, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We suppose that deep functional maps are trained on $S$ with respect to all possible pairs. Then the global energy is given by $E_{\\mathrm{total}}(\\mathcal{C}) = \\sum_{i,j} \\| \\mathbf{C}_{ij} \\mathbf{A}_i - \\mathbf{A}_j \\|^2 + \\sum_{i,j} E_{\\mathrm{reg}}(\\mathbf{C}_{ij})$ , where $\\mathcal{C} = \\{\\mathbf{C}_{ij}\\}_{i,j \\in [1..n]}$ is the set of functional maps among training shapes.", + "bbox": [ + 496, + 750, + 893, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1 If $E_{\\text{total}}(\\mathcal{C}) = 0$ , then for any shape $S_i$ , and any path $(i, i_1, i_2, \\dots, i_k, i)$ , the map composition $\\mathbf{C}_{ii}$ is cycle consistent within the functional space spanned by the columns of $\\mathbf{A}_i$ , i.e., $\\mathbf{C}_{ii}\\mathbf{A}_i = \\mathbf{A}_i$ .", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14499", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a consequence of Prop. 1, when $\\mathbf{A}_i\\in \\mathbb{R}^{k\\times d}$ is of full row rank, then $\\mathcal{C} = \\{\\mathbf{C}_{ij}\\}_{i,j\\in [1\\dots n]}$ is cycle consistent. In practice, we generally set $d > k$ . Moreover, during network training, $\\mathbf{C}_{ij}$ is computed via pseudo-inverse, which implicitly assumes the full-rankness of $\\mathbf{A}_i,\\mathbf{A}_j$ . We refer readers to the empirical validation in Section 4.2.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We defer the proof of Prop. 1 to Supp. Material. In fact, a similar claim has been formulated and proven in [49] (see Sec.3.1 therein), but in the context of map refinement via promoting cycle consistency.", + "bbox": [ + 75, + 181, + 468, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Though being technically similar, the theoretical argument of [49] and that of ours have fundamentally different implications. More specifically, the former justifies a test-time optimization algorithm, which is used to promote cycle consistency of maps among a fixed test shape collection. While the latter suggests that spectrally cycle consistency has been ensured and further leveraged to enhance the universal feature extractor (independent of the test data) during training in any DFM framework following the generic pipeline presented in Sec. 3.1.", + "bbox": [ + 75, + 242, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Two-branch Deep Functional Maps", + "text_level": 1, + "bbox": [ + 76, + 407, + 397, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It has long been recognized, both theoretically and empirically, that optimizing purely in the spectral domain is not sufficient. As a toy example, a trivial solution attaining global optima can be constructed as follows: Suppose that we have learned a feature extractor $\\mathcal{F}_{\\Theta}$ , which returns the respective eigenbasis transformed by a universal $\\mathbf{A}_0$ . That is, $\\mathbf{G}_i = \\boldsymbol{\\Phi}_i\\mathbf{A}_0, \\forall i$ , which implies $\\mathbf{A}_i = \\boldsymbol{\\Phi}_i^\\dagger\\mathbf{G}_i = \\mathbf{A}_0, \\forall i$ . Then we have $\\mathbf{C}_{ij} \\equiv I_k$ , which exactly satisfies $E_{\\mathrm{desc}}(\\mathbf{C}_{ij}) = E_{\\mathrm{reg}}(\\mathbf{C}_{ij}) = 0, \\forall i, j$ . However, it probably induces poor point-wise maps.", + "bbox": [ + 75, + 433, + 468, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In fact, in [37] the authors have proposed to use an ICP-like technique to encourage the estimated functional maps to be induced by some point-wise maps. In [34], the authors propose a spectral upsampling method for map refinement, which essentially converts maps back and forth between spectral and spatial domains. Moreover, the following lemma from [34] sheds light on the necessity of taking both spectral and spatial representations into consideration.", + "bbox": [ + 75, + 585, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lemma 1 Given a pair of shapes $S_{1}, S_{2}$ each having nonrepeating Laplacian eigenvalues, which are the same. A point-wise map $T: S_{1} \\to S_{2}$ is an isometry if and only if the corresponding functional map $\\mathbf{C}$ in the complete Laplacian basis is both diagonal and orthonormal.", + "bbox": [ + 75, + 722, + 468, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above lemma suggests that apart from promoting the structural properties of functional maps, it is also critical to enforce them to be associated with certain point-wise maps, or termed as properness of functional maps in [40].", + "bbox": [ + 75, + 809, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we remark that some recent DFM advances also promote the properness of the resulting spectral maps.", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For instance, AttentiveFMaps [28] follows the spirit of ZoomOut [34] and explicitly performs a conversion between spectral and spatial map representations across different dimensions of eigenbasis; UDMSM [9] constructs explicitly a universal shape in the feature space, and enforce the spectral map estimation to be consistent with the spatial maps induced via the universal shape.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Two-branch Map Estimation", + "text_level": 1, + "bbox": [ + 498, + 205, + 758, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this part, we leverage our observation made in Prop. 1 and propose a novel, simple yet effective design of unsupervised deep functional maps, which introduces a new branch that independently estimates maps from spatial perspective.", + "bbox": [ + 496, + 229, + 890, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our key insight is that, once cycle consistency is valid and $\\mathbf{A}_i$ is of full row rank, $\\mathbf{A}_i$ can be seen as a functional map from a universal latent shape, $S_0$ , to $S_i$ . This perspective has been explored in several prior works [49, 23, 24], we provide the following details to be self-contained. The above assumption implies $\\mathbf{C}_{ij} = \\mathbf{A}_j\\mathbf{A}_i^\\dagger$ . Then $\\mathbf{C}_{ij}$ can be interpreted as a functional map composition from $S_i$ to $S_0$ , followed by a map from $S_0$ to $S_j$ . On the other hand, one can align the spectral embeddings of $S_j$ to that of $S_i$ by simply transforming the former by $\\mathbf{C}_{ij}$ . Indeed, we convert $\\mathbf{C}_{ij}$ into the point-wise map by the nearest neighbor searching between the rows of $\\Phi_j\\mathbf{C}_{ij}$ and that of $\\Phi_i$ . From this point of view, denoting the virtual spectral embedding of the latent shape by $\\Phi_0$ , $\\Phi_i\\mathbf{A}_i$ can be then treated as the spectral embedding of $S_i$ aligned to that of $S_0$ . Therefore, given a pair of shapes $S_i, S_j$ , since we have aligned their eigenbasis to the canonical frame defined by the virtual spectral embedding $\\Phi_0$ , we can align the spectral embedding of $S_i$ to $\\Phi_0$ by computing $\\Phi_i\\mathbf{A}_i$ . Once all the spectral embeddings are aligned to the canonical embedding domain regarding $\\Phi_0$ , we can compute the soft point-wise map between $S_i$ and $S_j$ by nearest neighbor searching between the rows of $\\Phi_i\\mathbf{A}_i$ and those of $\\Phi_j\\mathbf{A}_j$ .", + "bbox": [ + 496, + 290, + 892, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above derivation, given the learned features projected in the spectral domain, $\\mathbf{A}_i,\\mathbf{A}_j$ , and a pair of indices $p\\in [1..n_i],q\\in [1..n_j]$ , we can compute point-wise maps. Firstly we compute residual:", + "bbox": [ + 496, + 638, + 890, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {q p} = \\left\\| \\Phi_ {i} [ p ] \\mathbf {A} _ {i} - \\Phi_ {j} [ q ] \\mathbf {A} _ {j} \\right\\| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 710, + 890, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Phi_i[p]$ denotes the $p$ -th row of $\\Phi_i$ , and similarly we define $\\Phi_j[q]$ . The soft point-wise map $\\Pi \\in \\mathbb{R}^{n_j \\times n_i}$ is then given by:", + "bbox": [ + 496, + 738, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Pi (q, p) = \\frac {\\exp (- \\alpha \\delta_ {q p})}{\\sum_ {p ^ {\\prime}} \\exp (- \\alpha \\delta_ {q p ^ {\\prime}})}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 795, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that by construction, each row of $\\Pi$ is non-negative and sums up to 1, forming a probability distribution. The parameter $\\alpha$ controls the entropy of each distribution - the smaller/larger $\\alpha$ is, the fuzzier/sharper the distribution is.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14500", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/95ad605773bb07df2bf392fdc7ed80286db371f9990e4c239a2561d6072d540c.jpg", + "image_caption": [ + "Figure 3. We train our two-branch DFM and a vanilla single-branch version on DT4D-H and monitor spectrally and spatially cycle consistency along the training." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 449, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of manually tuning the optimal $\\alpha$ , we propose a learning scheme that dynamically controls $\\alpha$ over training, which is inspired by curriculum learning [5]. We defer the respective details to Section 4.3.", + "bbox": [ + 75, + 303, + 468, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We convert the soft point-wise map to a functional map by", + "bbox": [ + 75, + 364, + 468, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {2} = \\boldsymbol {\\Phi} _ {j} ^ {\\dagger} \\Pi \\boldsymbol {\\Phi} _ {i}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 392, + 468, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the end, we enforce $\\mathbf{C}_2$ to be consistent with $\\mathbf{C}_1$ , the intermediate output from the FMreg layer.", + "bbox": [ + 75, + 415, + 468, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To summarize, thanks to the spectrally cycle consistency we identify in Sec. 3.2, we are allowed to construct a spectral latent shape and induce spatial maps via it. By enforcing the spatial estimations to be consistent with the spectral ones, we obtain a spatially and spectrally consistent deep functional maps framework.", + "bbox": [ + 75, + 446, + 468, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Network Design", + "text_level": 1, + "bbox": [ + 76, + 544, + 236, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our two-branch network is built upon a standard DFM framework. In the following, we denote by $\\mathbf{C}_1$ and $\\mathbf{C}_2$ the estimated functional maps from the original FMreg layer and our novel branch, respectively.", + "bbox": [ + 75, + 568, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we use DiffusionNet [44] as our feature extractor. And WKS [4] descriptors are fed into it as initialization of learned features. We borrow the FMreg layer from [12]. It takes both $E_{\\mathrm{desc}}(\\mathbf{C})$ and commutativity with the Laplace-Beltrami operator into consideration, where the latter is given as:", + "bbox": [ + 75, + 628, + 468, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {l a p}} = \\left\\| \\mathbf {C} _ {1} \\Lambda_ {1} - \\Lambda_ {2} \\mathbf {C} _ {1} \\right\\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 727, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Lambda_{1}$ and $\\Lambda_{2}$ are diagonal matrices of the Laplace-Beltrami eigenvalues on the two shapes.", + "bbox": [ + 75, + 758, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The estimation of $\\mathbf{C}_2$ has been described in detail in Section 4.1. In the end, we formulate the training loss as:", + "bbox": [ + 75, + 789, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\mathbf {C} _ {1}, \\mathbf {C} _ {2}\\right) = \\left\\| \\mathbf {C} _ {1} ^ {T} \\mathbf {C} _ {1} - \\mathbf {I} \\right\\| ^ {2} + \\left\\| \\mathbf {C} _ {1} - \\mathbf {C} _ {2} \\right\\| ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 827, + 468, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the first term promotes the orthogonality of $\\mathbf{C}_1$ , while the second term promotes the consistency between the functional maps estimated from different branches.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we remark that by combining the FMreg layer and $\\mathcal{L}(\\mathbf{C}_1,\\mathbf{C}_2)$ , we have incorporated every factor in Lemma 1 into our design.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conceptual Validation In this part, we train a network on DT4D-H dataset (see Section 5.1 for details) with our two-branch network, and a single-branch variant without our spatial map estimation branch. We monitor and plot the following quantities along training: (1) Average spectral cycle consistency over sampled triplets, i.e., $\\frac{1}{M}\\sum_{(i,j,k)}\\| \\mathbf{C}_{ki}\\mathbf{C}_{jk}\\mathbf{C}_{ij} - I\\| ^2 /\\| I\\| ^2$ ; (2) Average spatial cycle consistency over sampled triplets, i.e., the mean Euclidean deviation from composed maps $T_{ki}\\circ T_{jk}\\circ T_{ij}$ to identity map on $S_{i}$ . Here $n = 80$ is the number of training shapes, and $M = 1000$ is the number of sampled triplets. The behavior of the blue curves after 4500 iterations verifies our argument that spectrally cycle consistency does not imply spatially cycle consistency. On the other hand, by introducing our two-branch design, the discrepancy is well compensated and evidently better cycle consistencies in both spatial and spectral domains are achieved.", + "bbox": [ + 498, + 155, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Updating Scheme of $\\alpha$ in Eqn. (3)", + "text_level": 1, + "bbox": [ + 500, + 421, + 792, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The soft point-map conversion (Eqn. 3) has been applied in several prior works [30, 28], which all set $\\alpha$ to be a manually selected constant. Ideally, we expect $\\Pi$ in Eqn. (3) to be close to a permutation matrix, i.e., each row forms a binary vector. This seems to suggest a preference for a large $\\alpha$ . Unfortunately, it would severely hinder network training, since the learned features and maps are of low quality in the early stage. On the other hand, a small $\\alpha$ can alleviate such difficulty but falls short of fully pushing functional maps to be proper. As demonstrated in Section 5.5, neither small nor large $\\alpha$ produces satisfying results.", + "bbox": [ + 498, + 444, + 890, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based upon the above analysis, we propose a novel updating scheme, which is inspired by curriculum learning [5]. Namely, we initiate a small $\\alpha$ at the beginning of network training and increase it by a constant step size for every fixed number of epochs. As shown in Table 1, 2, 3, our scheme does not rely on hyper-parameter tuning but also achieves state-of-the-art results.", + "bbox": [ + 498, + 611, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 726, + 715, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement our network with PyTorch [39]. We use four DiffusionNet blocks [44] as feature backbone and borrow the functional map block with Laplacian regularizer from [12]. The dimension of the Laplace-Beltrami eigenbasis is set to 50. WKS [4] descriptors are used as the input signal to our network. The dimensions of the input and the output descriptors are both set to 128. During training, the value of the learning rate is set to 2e-4 with ADAM optimizer. In all experiments, we train our method for 10,000 iterations with a batch size of 1. Following the learning", + "bbox": [ + 498, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14501", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "strategy in Section 4.3, we initialize $\\alpha$ to 1 and increase it by 5 per epoch. As indicated in Eqn. (6), We weigh equally the orthogonality loss with respect to $\\mathbf{C}_1$ and the residual between $\\mathbf{C}_1$ and $\\mathbf{C}_2$ . More implementation details are provided in the Supp. Material.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 188, + 282, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct an extensive set of experiments of non-rigid shape matching on various datasets including humanoids and animals. We test on both near-isometric and non-isometric shape pairs. Our method is compared to a set of competitive baselines including axiomatic, supervised, weakly-supervised, and unsupervised learning methods. We emphasize that in this section, all the maps from the learning-based pipelines are directly inferred from the trained models, without any post-processing procedure. We evaluate the matching results in terms of a mean geodesic error on shapes normalized to unit area. Finally, our point-wise maps are all inferred by converting the output functional maps, as all the other DFM frameworks.", + "bbox": [ + 75, + 218, + 470, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 431, + 181, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FAUST_r: The remeshed version [41] of FAUST dataset[6] contains 100 human shapes. Following [43], it is split into 80/20 for training and testing.", + "bbox": [ + 75, + 458, + 468, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SCAPE_r: The remeshed version [41] of SCAPE dataset[2] contains 71 human shapes. Following [43], it is split into 51/20 for training and testing.", + "bbox": [ + 75, + 507, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SHREC19_r: The remeshed version of SHREC19 dataset [33] collects 44 human shapes from 11 independent datasets with distinctive poses and styles. We abandon shape 40 due to its partiality, we test on 407 pairs among the rest 43 shapes, which come with ground-truth.", + "bbox": [ + 75, + 555, + 468, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DT4D-H [32]: The remeshed subset of the large scale animation dataset DeformingThings4D [29]. In particular, DT4D-H includes 10 categories of humanoid shapes undergoing significant pose and style variances, forming a challenging benchmark.", + "bbox": [ + 75, + 633, + 468, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SMAL_r: The remeshed SMAL dataset [53] contains 49 animal shapes with 8 species. We follow the setting from [28], which splits 29 (5 species) and 20 (3 species) shapes for training and testing.", + "bbox": [ + 75, + 712, + 468, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TOSCA_r: The remeshed TOSCA dataset [7] contains multiple shape categories. We choose 4 animal categories, including cat, dog, horse, and wolf to verify the generalization performance of networks trained on SMAL_r. Note that we only infer the intra-category maps, due to the absence of ground-truth inter-category maps.", + "bbox": [ + 75, + 776, + 468, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We refer readers to Supp. Material for visualizations illustrating the variability of the above datasets.", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Near-isometric Shape Matching", + "text_level": 1, + "bbox": [ + 500, + 90, + 781, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this part, we perform comparisons with an array of non-rigid shape matching methods: (1) Axiomatic methods including ZoomOut [34], BCICP [41], IsoMuSh [17], Smooth Shells [13], CZO [24]; (2) Supervised learning methods including TransMatch [48], GeomFMaps [12], and supervised version of AttentiveFMaps [28]; (3) Unsupervised learning methods including NeuroMorph [14], SyNoRiM [16], Deep Shell [15], AttentiveFMaps [28], UDMSM [9], DUO-FM [11].", + "bbox": [ + 496, + 114, + 890, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For all learning-based methods, we train models on FAUST_r and SCAPE_r respectively. Table 1 reports results on both standard tests and more challenging generalizations. We observe a trade-off between the two tasks, methods that perform the best in the former (e.g., supervised AttentiveFMaps and UDMSM) tend to overfit, and therefore suffer poor generalization (especially to SREHC19_r). Meanwhile, our default setting, denoted by Ours, achieves reasonable performance in the standard tests but also outperforms the external baselines in 3 out of 4 generalization tests. Especially, in generalizing to SHREC19_r, Ours outperforms the external baselines by a large margin, resulting in $41\\%$ (3.8 vs. 6.4) and $46\\%$ (4.5 vs. 8.4) error reduction upon the second best.", + "bbox": [ + 496, + 251, + 890, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We highlight that post-processing with cycle consistency generally depends on the initialized map quality and the size of the test set (e.g., $\\geq 3$ shapes). In contrast, we leverage cycle consistency to improve the feature extractor during training. We also report the results from post-processing techniques based on cycle consistency [17, 24, 16] in Table 1. They are significantly outperformed by our method, which is inferred per-pair and without any post-processing.", + "bbox": [ + 496, + 463, + 890, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further augment the dimension of functional maps in network training to 80 (same as UDMSM), which is beneficial to near-isometric matching [28]. It is evident that Ours (80 dim) achieves on-par performance with the regarding state-of-the-art methods in standard tests. On the other hand, due to the significant variability between SHREC19_r and the training sets (see Supp. Material), augmenting dimension leads to worse generalization than before (5.5 vs. 3.8, 5.8 vs. 4.5). Nevertheless, even in this case, our method outperforms the external baselines in all generalization tests by a notable margin.", + "bbox": [ + 496, + 585, + 890, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Non-isometric Shape Matching", + "text_level": 1, + "bbox": [ + 500, + 763, + 774, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also train our network on non-isometric datasets, SMAL_r and DT4D-H, and compare it with the state-of-the-art baselines including DeepShells [15], AttentiveFMaps [28], UDMSM [9] and DUO-FM [11].", + "bbox": [ + 496, + 787, + 890, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SMAL_r: We follow the split and input descriptors from [28] (more details are provided in the Supp. Material). Table 2 reports results on SMAL_r, our method", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14502", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/debab315f179157f4fdf5153877d5de3c77d29771d6dbc46dff585efc62c28af.jpg", + "table_caption": [ + "Table 1. Mean geodesic errors (×100) on FAUST_r, SCAPE_r, and SHREC19_r. The best and the second best are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodTrain TestFAUST_rSCAPE_r
FAUST_rSCAPE_rSHREC19_rSCAPE_rFAUST_rSHREC19_r
ZM[34]6.1\\\\7.5\\\\
BCICP[41]6.4\\\\11.0\\\\
IsoMuSh[17]4.4\\\\5.6\\\\
Smooth Shell[13]2.5\\\\4.7\\\\
CZO[24]2.2\\\\2.5\\\\
TransMatch[48]2.733.621.018.318.638.8
GeomFMaps[12]supervised2.63.39.93.03.012.2
AttentiveFMaps[28]1.42.29.41.71.812.2
NeuroMorph[14]8.528.526.329.918.227.6
SyNoRiM[16]7.921.725.59.524.626.8
Deep Shell[15]1.75.427.42.52.723.4
AttentiveFMaps[28]unsupervised1.92.66.42.22.29.9
UDMSM[9]1.57.321.52.08.630.7
DUO-FM[11]2.54.26.42.72.88.4
Ours2.32.63.82.42.54.5
Ours (80 dim)1.72.65.52.22.05.8
", + "bbox": [ + 81, + 101, + 893, + 340 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/27f93f5c5bb9217e3ee51f9dd99d68cd46707bcfb296e8dd056473ce295d81f3.jpg", + "table_caption": [ + "Table 2. Mean geodesic errors $(\\times 100)$ on SMAL_r. The best and the second best are highlighted correspondingly." + ], + "table_footnote": [], + "table_body": "
MethodTrain TestSMAL_r
SMAL_rTOSCA_r
DeepShell[15]29.38.7
GeomFMaps[12]7.624.5
AttentiveFMaps[28]5.420.9
UDMSM[9]24.621.7
DUO-FM[11]32.815.3
Ours5.47.9
", + "bbox": [ + 86, + 388, + 459, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves the best performance, which is on-par with AttentiveFMaps [28]. To evaluate generalization performance, we use the trained models to directly infer intra-category maps within TOSCA-r. It turns out that AttentiveFMaps and GeomFMaps both suffer from significant performance drops ( $\\times 3.8$ and $\\times 3.2$ larger geodesic error). It is also worth noting that DeepShells achieves the second-best generalization score in the relatively simpler task. However, it fails dramatically regarding the base task. In contrast, our method achieves the best balance between learning in difficult non-isometric pairs and generalizing to relatively easy near-isometric pairs.", + "bbox": [ + 75, + 544, + 468, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DT4D-H: We follow the train/test (198/95) split of [28], but ignore the categories mousey and ortiz in both train and test, due to the lack of inter-category map labels regarding them, resulting a split of 168/80. We emphasize that we conduct training and test in a category-agnostic manner, i.e., no class label is used, and the training pairs can consist of shapes from arbitrary two categories. This is significantly different from [28], in which training pairs are selected according to clustering information. Obviously, our setting is more practical, but also more challenging. For completeness, we report results under the setting of [28] in Supp. Material", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d41dc492e8de2981b3f434e3f45cafc90393bd9a3fa7c3dec84657091f86aaa5.jpg", + "image_caption": [ + "$T_{11}^{(8)} = T_{81}\\circ T_{78}\\circ \\dots \\circ T_{23}\\circ T_{12}$" + ], + "image_footnote": [], + "bbox": [ + 516, + 364, + 653, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64ccf4fd1979f7cea35e717ec983ffe0a1bc56f37a4e7455c25af9d29ed047e1.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 669, + 364, + 818, + 444 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e3df59357d281208aec89702037854f0565f8c2c96010fc7d04560cabb6c515.jpg", + "image_caption": [ + "UDMSM" + ], + "image_footnote": [], + "bbox": [ + 818, + 364, + 895, + 444 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ea98edba3fcd026d767278c8b76c4d8fe67e02aa930fe28629e6dab1125d9c1.jpg", + "image_caption": [ + "DUO-FM", + "Deep Shell" + ], + "image_footnote": [], + "bbox": [ + 669, + 455, + 818, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 738, + 455, + 818, + 534 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/047059c4388edffeebf8516da0a316d882ca06a1329d31fa27fc53f8f487a78b.jpg", + "image_caption": [ + "BCICP", + "Figure 4. Qualitative evaluation of spatial cycle consistency of different methods. Even composed along a path of 8 highly deformed shapes, our resulting map remains close to identity, while all the baselines fail significantly." + ], + "image_footnote": [], + "bbox": [ + 820, + 455, + 895, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and our method outperforms the baselines in both intra- and inter-category evaluation by a notable margin.", + "bbox": [ + 496, + 621, + 890, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 reports results on DT4D-H, in which we preserve 80 shapes for test and train networks with 168 and 80 shapes, respectively. Note that we report mean geodesic errors over all possible test shape pairs, which may undergo significant distortions (see, e.g., Fig. 1). Our method obtains a $67.2\\%$ (7.7 vs. 22.4) geodesic error reduction with respect to the second-best baseline. On top of that, we also test the generalization of the trained model on near-isometric benchmarks – our method also generalizes the best in generalization to FAUST_r and SCAPE_r. The same pattern is observed when the training set is reduced by more than half. Remarkably, our network trained on the reduced set still outperforms all the baselines trained on the full set.", + "bbox": [ + 496, + 655, + 890, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, we attribute our performance on matching challenging non-isometric shapes (Table 3) and generalizing to unseen shapes (Table 1) to our effort to promote both spec", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14503", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3004ac49aba51731e73092f2fa7d6726e09d9547e53dcb9a01335d50acd8fd62.jpg", + "table_caption": [ + "Table 3. Mean geodesic errors $\\left( {\\times {100}}\\right)$ on DT4D-H. The best and the second best are highlighted correspondingly." + ], + "table_footnote": [], + "table_body": "
MethodTrain TestDT4D-H (168)DT4D-H (80)
DT4D-HFAUST_rSCAPE_rDT4D-HFAUST_rSCAPE_r
DeepShell[15]27.04.96.529.34.77.0
AttentiveFMaps[28]25.73.46.428.92.76.3
UDMSM[9]46.843.347.949.742.540.0
DUO-FM[11]22.410.012.224.78.09.2
Ours7.73.16.19.02.66.2
", + "bbox": [ + 81, + 99, + 893, + 198 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5e30d750640885c0515730b9c0e6cc7b6949bf42b1f0bbed88df5b2fd768af20.jpg", + "table_caption": [ + "Table 4. Mean geodesic errors $\\left( {\\times {100}}\\right)$ of SURFMNet and our variant trained on 4 datasets" + ], + "table_footnote": [], + "table_body": "
MethodSURFMNetSURFMNet + Ours
FAUST_r6.03.5
SCAPE_r6.83.4
SMAL_r20.413.3
DT4D-H18.315.0
", + "bbox": [ + 107, + 238, + 437, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trally and spatially cycle consistency. Especially, the isometry assumption is likely violated in the former case, thus cycle consistency, as a generic prior, plays an important role of regularizing maps.", + "bbox": [ + 75, + 344, + 467, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As an illustration, we present a qualitative evaluation on the point-wise cycle consistency in Fig. 4. We sample 8 shapes from the test set of DT4D-H (one from each category) and compose the maps along the path $(S_{1} \\to S_{2} \\to \\dots \\to S_{8} \\to S_{1})$ with respect to different approaches. It is evident that due to the significant distortion undergoing among the shapes, all but our method fails to preserve cycle consistency in this demanding test, while our composing map approximates the identity map on $S_{1}$ . It also aligns nicely with the quantitative results reported in Table 3.", + "bbox": [ + 75, + 405, + 467, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Integration with SURFMNet [43]", + "text_level": 1, + "bbox": [ + 76, + 566, + 369, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our two-branch design can be easily incorporated into any existing DFM framework following the general design outlined in Sec. 3.1. To demonstrate this, we modify the SURFMNet [43], one of the earliest approaches of unsupervised DFM, by adding our new branch. Table 4 shows the matching accuracy on the four main benchmarks. It is evident that in every case, incorporating our design leads to significant error reduction ranging from $18\\%$ to $50\\%$ . Especially, in the near-isometric cases, we obtain $41.6\\%$ and $50\\%$ error reduction respectively. Note that the absolute scores, 0.035, 0.034, are reasonable even compared to the state-of-the-art results reported in Table 1.", + "bbox": [ + 75, + 592, + 467, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 785, + 228, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we present a set of ablation studies consisting of two parts. The first part verifies the rationality of our method, and the second part demonstrates the robustness of our method. We conduct all experiments on SMAL_r dataset [53].", + "bbox": [ + 75, + 809, + 467, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First of all, instead of using the updating scheme in Sec-", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion 4.3, we test the performance of our pipeline using two fixed values of $\\alpha$ in Eqn. (3): $\\alpha = 1$ and $\\alpha = 50$ . Compared to our proposed model, the two variants yield a noticeable performance drop. Especially in the case $\\alpha = 50$ , the network fails to deliver reasonable matching results. We believe it is because a large $\\alpha$ amplifies the noise of maps learned at the early training stage.", + "bbox": [ + 496, + 217, + 890, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then we justify our two-branch network design. Removing spatial branch amounts to training a standard single-branch DFM. To remove the spectral branch, we remove the FMreg layer and instead use our new branch to compute point-wise maps and convert them to functional maps. In the end, we modify the training loss so that it covers descriptor preservation, commutativity with the Laplace-Beltrami operator, and orthogonality (the latter two compensate the removed FMreg layer). The accuracy drop reported in the third and fourth row of Table 5 clearly suggests the necessity of our two-branch design.", + "bbox": [ + 496, + 324, + 890, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our experiments, we always use the full resolution meshes ( $\\sim 5k$ vertices) and compute in Eqn. (2) with all of the 128 descriptors. We anticipate that efficiency can become an issue when the input mesh resolution is high, and/or we would like to increase the size of learned descriptors. Therefore, we test the robustness of our pipeline with respect to down-sampling, which is commonly used in functional maps-based frameworks [34, 28]: 1) We down-sample 3000 vertices on each shape via furthest point sampling; 2) In order to down-sample the feature dimension, we operate as the following during training: given a $\\mathbf{A}_1$ , $\\mathbf{A}_2$ , we perform SVD on $\\mathbf{A}_1$ , i.e., $\\mathbf{A}_1 = U_1\\Sigma_1V_1^T$ , then we set $\\hat{\\mathbf{A}}_1 = \\mathbf{A}_1\\hat{V}_1$ , and set $\\hat{\\mathbf{A}}_2 = \\mathbf{A}_2\\hat{V}_1$ , where $\\hat{V}_1$ is the first $m$ columns of $V_{1}$ . We set $m = 30$ by replacing $\\mathbf{A}_i$ with $\\hat{\\mathbf{A}}_i$ in Eqn. (2). The results in the bottom two rows show that the above operation has a relatively minor effect on the performance, proving the robustness of our method.", + "bbox": [ + 496, + 492, + 890, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 767, + 617, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To conclude, we provide a theoretical justification for spectral cycle consistency in DFM. To compensate for the discrepancy of purely spectrally cycle consistency, we formulate a spectral latent shape that allows the alignment of the spectral embeddings of each shape to a canonical embedding domain. Based on this construction, we introduce a two-branch architecture for estimating maps in both spec", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14504", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1157f33293cfad9fa526bcecf3106a394fa4fb72aa46aa4cbcf01c8a22328a08.jpg", + "table_caption": [ + "Table 5. Mean geodesic errors $(\\times 100)$ on different ablated settings, the models are all train on SMAL_r." + ], + "table_footnote": [], + "table_body": "
Ours5.4
α = 16.6
α = 5035.2
remove spatial branch33.4
remove spectral branch14.3
Vertex downsampling5.7
Feature downsampling5.9
", + "bbox": [ + 161, + 112, + 387, + 223 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tral and spatial domains within DFM. The resulting network is simple, computationally efficient, and compatible with most existing DFM frameworks. We demonstrate the effectiveness of our framework through a comprehensive set of experiments, showing significant improvements upon state-of-the-art approaches in terms of accuracy, consistency, and generalization performance.", + "bbox": [ + 75, + 247, + 468, + 352 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "It is worth noting, though, our method is developed upon clean and complete meshes. It would be interesting for future work to investigate the utility of our approach in more general tasks, involving partial meshes, noisy point clouds, and other representations.", + "bbox": [ + 75, + 353, + 468, + 428 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported in part by the National Natural Science Foundation of China under contract No. 62171256 and Shenzhen Key Laboratory of next-generation interactive media innovative technology (No. ZDSYS20210623092001004), and in part by the ERC Starting Grant No. 758800 (EXPROTEA) and the ANR AI Chair AIGRETTE.", + "bbox": [ + 75, + 446, + 468, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 565, + 171, + 580 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Noam Aigerman, Roi Poranne, and Yaron Lipman. Seamless Surface Mappings. ACM Transactions on Graphics (TOG), 34(4):72, 2015. 1", + "[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape Completion and Animation of People. In ACM Transactions on Graphics (TOG), volume 24, pages 408-416. ACM, 2005. 1, 6", + "[3] Souhaib Attaiki, Gautam Pai, and Maks Ovsjanikov. Dpfm: Deep partial functional maps. In 2021 International Conference on 3D Vision (3DV), pages 175-185. IEEE, 2021. 2", + "[4] Mathieu Aubry, Ulrich Schlickewei, and Daniel Cremers. The Wave Kernel Signature: A Quantum Mechanical Approach to Shape Analysis. In Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on, pages 1626-1633. IEEE, 2011. 1, 2, 5", + "[5] Yoshua Bengio, Jerome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In International Conference on Machine Learning, 2009. 5", + "[6] Federica Bogo, Javier Romero, Matthew Loper, and Michael J. Black. FAUST: Dataset and evaluation for 3D mesh registration. In Proceedings IEEE Conf. on Com" + ], + "bbox": [ + 86, + 590, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "puter Vision and Pattern Recognition (CVPR), Piscataway, NJ, USA, June 2014. IEEE. 1, 6", + "[7] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Numerical geometry of non-rigid shapes. Springer Science & Business Media, 2008. 6", + "[8] Oliver Burghard, Alexander Dieckmann, and Reinhard Klein. Embedding shapes with Green's functions for global shape matching. Computers & Graphics, 68:1-10, 2017. 2", + "[9] D. Cao and F. Bernard. Unsupervised deep multi-shape matching. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 4, 6, 7, 8", + "[10] Bailing Deng, Yuxin Yao, Roberto M. Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. 2022. 1", + "[11] Nicolas Donati, Etienne Corman, and Maks Ovsjanikov. Deep orientation-aware functional maps: Tackling symmetry issues in shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-751, 2022. 2, 3, 6, 7, 8", + "[12] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In CVPR, 2020. 5, 6, 7", + "[13] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In CVPR, 2020. 1, 3, 6, 7", + "[14] Marvin Eisenberger, David Novotny, Gael Kerchenbaum, Patrick Labatut, Natalia Neverova, Daniel Cremers, and Andrea Vedaldi. Neuromorph: Unsupervised shape interpolation and correspondence in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7473-7483, 2021. 6, 7", + "[15] Marvin Eisenberger, Aysim Token, Laura Leal-Taixe, and Daniel Cremers. Deep shells: Unsupervised shape correspondence with optimal transport. Advances in Neural Information Processing Systems, 33:10491-10502, 2020. 3, 6, 7, 8", + "[16] Jiahui Huang et. al. Multiway non-rigid point cloud registration via learned functional map synchronization, 2022. 6, 7", + "[17] Maolin Gao, Zorah Lahner, Johan Thunberg, Daniel Cremers, and Florian Bernard. Isometric multi-shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14183-14193, 2021. 6, 7", + "[18] Dvir Ginzburg and D. Raviv. Cyclic functional mapping: Self-supervised correspondence between non-isometric deformable shapes. In ECCV, 2020. 1, 2", + "[19] Oshri Halimi, Or Litany, Emanuele Rodol'a, Alex Bronstein, and Ron Kimmel. Unsupervised learning of dense shape correspondence. In CVPR, 2019. 1, 2", + "[20] Jiahui Huang, Tolga Birdal, Zan Gojcic, Leonidas J. Guibas, and Shi-Min Hu. Multiway Non-rigid Point Cloud Registration via Learned Functional Map Synchronization. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2", + "[21] Qixing Huang, Fan Wang, and Leonidas Guibas. Functional map networks for analyzing and exploring large shape collections. ACM TOG, 33(4), 2014. 1, 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14505", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Qi-Xing Huang and Leonidas Guibas. Consistent shape maps via semidefinite programming. In Computer Graphics Forum, volume 32, pages 177-186. Wiley Online Library, 2013. 1, 2", + "[23] Ruqi Huang, Panos Achlioptas, Leonidas Guibas, and Maks Ovsjanikov. Limit shapes – a tool for understanding shape differences and variability in 3d model collections. Computer Graphics Forum, 38(5):187–202, 2019. 2, 4", + "[24] Ruqi Huang, Jing Ren, Peter Wonka, and Maks Ovsjanikov. Consistent zoomout: Efficient spectral map synchronization. In Computer Graphics Forum, volume 39, pages 265-278. Wiley Online Library, 2020. 1, 3, 4, 6, 7", + "[25] Daniel F Huber. Automatic three-dimensional modeling from reality. PhD thesis, Citeseer, 2002. 2", + "[26] Artiom Kovnatsky, Michael M Bronstein, Alexander M Bronstein, Klaus Glashoff, and Ron Kimmel. Coupled quasiharmonic bases. In Computer Graphics Forum, volume 32, pages 439-448, 2013. 2", + "[27] Spyridon Leonardos, Xiaowei Zhou, and Kostas Daniilidis. Distributed consistent data association via permutation synchronization. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 2645-2652. IEEE, 2017. 2", + "[28] Lei Li, Nicolas Donati, and Maks Ovsjanikov. Learning multi-resolution functional maps with spectral attention for robust shape matching. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[29] Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12706-12716, 2021. 6", + "[30] Or Litany, Tal Remez, Emanuele Rodola, Alexander M. Bronstein, and Michael M. Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5660-5668, 2017. 1, 2, 3, 5", + "[31] Rabinovich M., Poranne R., Panozzo D., and Sorkine-Hornung O. Scalable locally injective mappings. In ACM Transactions on Graphics, 2017. 1", + "[32] Robin Magnet, Jing Ren, Olga Sorkine-Hornung, and Maks Ovsjanikov. Smooth non-rigid shape matching via effective dirichlet energy optimization. In International Conference on 3D Vision, 2022. 6", + "[33] Simone Melzi, Riccardo Marin, Emanuele Rodolà, Umberto Castellani, Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Shrec 2019: Matching humans with different connectivity. In Eurographics Workshop on 3D Object Retrieval, volume 7, page 3, 2019. 6", + "[34] Simone Melzi, Jing Ren, Emanuele Rodola, Peter Wonka, and Maks Ovsjanikov. Zoomout: Spectral upsampling for efficient shape correspondence. Proc. SIGGRAPH Asia, 2019, 3, 4, 6, 7, 8", + "[35] Andy Nguyen, Mirela Ben-Chen, Katarzyna Welnicka, Yinyu Ye, and Leonidas Guibas. An optimization approach to improving collections of shape maps. In Computer Graphics Forum, volume 30, pages 1481-1491. Wiley Online Library, 2011. 2" + ], + "bbox": [ + 78, + 90, + 468, + 904 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Dorian Nogneng and Maks Ovsjanikov. Informative descriptor preservation via commutativity for shape matching. Computer Graphics Forum, 36(2):259-267, 2017. 2", + "[37] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional Maps: A Flexible Representation of Maps Between Shapes. ACM Transactions on Graphics (TOG), 31(4):30, 2012. 1, 2, 4", + "[38] Maks Ovsjanikov, Etienne Corman, Michael Bronstein, Emanuele Rodola, Mirela Ben-Chen, Leonidas Guibas, Frederic Chazal, and Alex Bronstein. Computing and processing correspondences with functional maps. In ACM SIGGRAPH 2017 Courses, page 5. ACM, 2017. 2", + "[39] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Linand N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. Pytorch: An imperative style, high-performance deep learning library, 2019. 5", + "[40] Jing Ren, Simone Melzi, Peter Wonka, and Maks Ovsjanikov. Discrete optimization for shape matching, 2021. 4", + "[41] Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Continuous and orientation-preserving correspondences via functional maps. ACM Trans. Graph., 37(6):248:1-248:16, Dec. 2018. 6, 7", + "[42] Emanuele Rodolà, Luca Cosmo, Michael M Bronstein, Andrea Torsello, and Daniel Cremers. Partial functional correspondence. In Computer Graphics Forum, volume 36, pages 222–236. Wiley Online Library, 2017. 2", + "[43] Jean-Michel Roufosse, Abhishek Sharma, and Maks Ovsjanikov. Unsupervised deep learning for structured shape matching. ICCV, 2019. 1, 2, 6, 8", + "[44] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM Transactions on Graphics, 2022. 2, 5", + "[45] Meged Shoham, Amir Vaxman, and Mirela Ben-Chen. Hierarchical functional maps between subdivision surfaces. In Computer Graphics Forum, volume 38, pages 55-73. Wiley Online Library, 2019. 2", + "[46] Jian Sun, Maks Ovsjanikov, and Leonidas Guibas. A Concise and Provably Informative Multi-Scale Signature Based on Heat Diffusion. In Computer graphics forum, volume 28, pages 1383-1392, 2009. 1, 2", + "[47] Federico Tombari, Samuele Salti, and Luigi Di Stefano. Unique signatures of histograms for local surface description. In International Conference on Computer Vision (ICCV), pages 356-369, 2010. 1", + "[48] Giovanni Trappolini, Luca Cosmo, Luca Moschella, Ricardo Marin, Simone Melzi, and Emanuele Rodola. Shape registration in the time of transformers. Advances in Neural Information Processing Systems, 34:5731-5744, 2021. 6, 7", + "[49] Fan Wang, Qixing Huang, and Leonidas J. Guibas. Image co-segmentation via consistent functional maps. In Proceedings of the IEEE International Conference on Computer Vision, pages 849-856, 2013. 2, 4", + "[50] Lanhui Wang and Amit Singer. Exact and stable recovery of rotations for robust synchronization. Information and Inference: A Journal of the IMA, 2(2):145-193, 2013. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14506", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Lei Yang, Wenxi Liu, Zhiming Cui, Nenglun Chen, and Wenping Wang. Mapping in a cycle: Sinkhorn regularized unsupervised learning for point cloud shapes. In ECCV, 2020. 2", + "[52] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In 2010 IEEE Computer Society Conference on Computer" + ], + "bbox": [ + 78, + 90, + 468, + 190 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition, pages 1426-1433. IEEE, 2010. 2", + "[53] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 6, 8" + ], + "bbox": [ + 503, + 90, + 890, + 189 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14507", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Spatially and Spectrally Consistent Deep Functional Maps/38616010-b1fd-470b-8376-180cbf13b5f1_model.json b/2023/Spatially and Spectrally Consistent Deep Functional Maps/38616010-b1fd-470b-8376-180cbf13b5f1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7763f4fa353f26ff7dd805b9800772e0d0eafdf --- /dev/null +++ b/2023/Spatially and Spectrally Consistent Deep Functional Maps/38616010-b1fd-470b-8376-180cbf13b5f1_model.json @@ -0,0 +1,2455 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.131, + 0.78, + 0.154 + ], + "angle": 0, + "content": "Spatially and Spectrally Consistent Deep Functional Maps" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.179, + 0.826, + 0.254 + ], + "angle": 0, + "content": "Mingze Sun1 Shiwei Mao1 Puhua Jiang1,2 Maks Ovsjanikov3 Ruqi Huang1* \n1Tsinghua Shenzhen International Graduate School, China \n2Peng Cheng Laboratory, China \n3LIX, École polytechnique, IP Paris, France" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.288, + 0.314, + 0.304 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.474, + 0.669 + ], + "angle": 0, + "content": "Cycle consistency has long been exploited as a powerful prior for jointly optimizing maps within a collection of shapes. In this paper, we investigate its utility in the approaches of Deep Functional Maps, which are considered state-of-the-art in non-rigid shape matching. We first justify that under certain conditions, the learned maps, when represented in the spectral domain, are already cycle consistent. Furthermore, we identify the discrepancy that spectrally consistent maps are not necessarily spatially, or point-wise, consistent. In light of this, we present a novel design of unsupervised Deep Functional Maps, which effectively enforces the harmony of learned maps under the spectral and the point-wise representation. By taking advantage of cycle consistency, our framework produces state-of-the-art results in mapping shapes even under significant distortions. Beyond that, by independently estimating maps in both spectral and spatial domains, our method naturally alleviates over-fitting in network training, yielding superior generalization performance and accuracy within an array of challenging tests for both near-isometric and non-isometric datasets. Codes are available at https://github.com/rqhuang88/Spatially-and-Spectrally-Consistent-Deep-Functional-Maps." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.698, + 0.21, + 0.714 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Non-rigid shape matching is a fundamental task in shape analysis, playing a pivotal role in a wide spectrum of applications including texture transfer [1], volume parameterization [31], statistical shape analysis [2, 6], to name a few. In order to establish correspondences between shapes undergoing non-rigid deformations, early approaches [46, 4, 47] focus on designing hand-crafted features by exploiting geometric invariance. Nowadays, data-driven techniques have been widely adopted to learn features in a more flexible way, leading to significant improvements regarding accu" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.287, + 0.925, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.428, + 0.895, + 0.527 + ], + "angle": 0, + "content": "Figure 1. We train unsupervised non-rigid shape matching pipelines on the challenging non-isometric dataset DT4D-H, which involves 8 categories of humanoid shapes undergoing significant distortions. Correspondences are visualized by texture transfer. Our method (bottom row) outperforms the state-of-the-art, AttentiveFMaps [28] (middle row) and UDMSM [9] (top row), by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.535, + 0.744, + 0.551 + ], + "angle": 0, + "content": "racy, efficiency, and robustness [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.893, + 0.78 + ], + "angle": 0, + "content": "A noticeable trend among the learning-based shape matching approaches is based on the formalism of Deep Functional Maps (DFM), pioneered by the FMNet [30]. Functional maps [37], as a spectral map representation, allows to encode maps into compact matrices and to express desirable map priors (e.g., area-preservation, isometry, bijectivity) in simple algebraic forms. In particular, by learning features that optimize functional map priors, several unsupervised DFM frameworks [19, 43, 18, 13, 9, 28] have been proposed and, remarkably, achieve even comparable results with respect to the supervised counterparts. On the other hand, while the majority of DFM works focus on exploring local, pairwise-level map priors, there is relatively less attention (with the exception of [9]) paid to the problem of injecting global map priors into DFM pipelines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As a global map prior, cycle consistency has long been used as a powerful regularizer for jointly optimizing maps among shapes, both under point-wise [22] and functional [21, 24] map representation. In this paper, we exploit the utility of cycle consistency within DFM frameworks and propose a novel two-branch unsupervised DFM framework, which promotes cycle consistency in both spectral and spatial domains. We first re-examine the generic DFM frame" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.39, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: ruqihuang@sz.tsinghua.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14497" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.379 + ], + "angle": 0, + "content": "work and outline a theoretical condition, based on the residuals of a linear solver used in functional map estimation, that guarantees spectrally cycle consistency in DFM over training shapes. Then we identify the discrepancy that spectrally consistent maps are not necessarily spatially, or pointwise, consistent. In light of this, we leverage our theoretical insight to construct a latent shape in DFM and employ this construction to formulate a novel two-branch design of unsupervised DFM. We align each shape's spectral embedding, represented as the eigenbasis of the Laplace-Beltrami operator, to a canonical space given by the constructed universal latent shape. We then compute point-wise maps in this canonical embedding domain, which promotes consistency across computed point-wise maps. In the end, the point-wise maps are converted to functional maps, which are expected to agree with the ones from the original spectral branch. Overall, our two-branch DFM takes advantage of the justified spectrally cycle consistency and further enhances it to spatially cycle consistent." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.379, + 0.473, + 0.575 + ], + "angle": 0, + "content": "We conduct extensive experiments on several non-rigid shape matching benchmarks, and demonstrate that our method achieves superior performance over existing methods, especially in the presence of significant distortions among shapes (see, e.g., Fig. 1). We also observe that our method, by enforcing the harmony of spatial and spectral map representations, reduces over-fitting during training, which leads to remarkable generalization performance within an array of challenging tests. Finally, our two-branch design can be easily plugged into any existing DFM framework, and we show evidence that it achieves significant performance gains upon both recent and early DFM approaches [43] with marginal computational burden." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.575, + 0.473, + 0.727 + ], + "angle": 0, + "content": "To summarize, our main contributions are as follows: (1) We perform theoretical analysis on spectrally cycle consistency of DFM frameworks; (2) We formulate a simple yet effective two-branch design of unsupervised DFM based on our theoretical justification, which introduces spatially cycle consistency. (3) We prove the effectiveness of our method through an array of challenging non-rigid shape matching tasks, demonstrating improvements upon existing methods in terms of accuracy, consistency, and generalization performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.22, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Functional Maps Our method is built upon functional map representation, introduced in [37] and then significantly extended in follow-up works (see, e.g., [38]). The key idea is to encode shape correspondences as transformations between the respective spectral embeddings, which are represented by compact matrices by using reduced eigenbasis. The functional maps pipeline has been further improved in accuracy, efficiency, and robustness by many recent works including [26, 21, 8, 42, 36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "Deep Functional Maps In contrast to axiomatic approaches that use hand-crafted features [46, 4], the deep functional maps approach, pioneered by FMNet [30], aims to learn the optimal features from data. FMNet contributes several key designs of DFM: (1) it leverages Siamese network to conduct learning in a set of shape pairs; (2) it advocates refining the input descriptors with non-linear transformations. FMNet is then supervised by labeled maps to learn optimal features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Instead of learning from labeled maps, unsupervised approaches [19, 43] demonstrate that it is sufficient to learn from geometric map priors. More recently, with the development of robust mesh feature extractors [44], more frameworks [9, 28, 11, 3] are proposed to learn directly from geometry, yielding state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.896, + 0.58 + ], + "angle": 0, + "content": "Cycle Consistency Cycle consistency has long been used as a strong prior for joint map optimization among a collection of shapes. Axiomatic approaches detect and eliminate inconsistent cycles using consistency constraints [25, 52, 35, 51, 18], as well as associate cycle consistency with low-rank properties of matrices encoding map networks [22, 50, 27]. Related to the latter, the matrix nature of the functional maps enables convenient access to map composition, which naturally bridges the functional map framework and consistent map refinement techniques [49, 21, 23, 45]. Some recent learning-based approach [20] also incorporate cycle consistency in the pipeline. It is worth noting, though, the above works all utilize cycle consistency as a prior for test-time optimization, which depends on test shape collection and initial maps. Contrastingly, our approach exploits cycle consistency during training to boost feature learning and poses no constraint on a test." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.896, + 0.809 + ], + "angle": 0, + "content": "From this viewpoint, UDMSM [9] is perhaps the most relevant work to ours, as both construct a latent shape during training to inject cycle consistency. The key difference between their work and ours, though, is how the latent shape is constructed. In [9], the authors propose to construct a universal shape in the learned feature space and establish pointwise maps from real shapes to the universal one by training a classifier. As shown in Sec.5, though this construction leads to great performance on matching near-isometric shapes, it suffers from large shape variability in mapping non-isometric ones. On the other hand, we leverage spectral information in estimating point-wise maps. As a result, our approach benefits from more direct usage of intrinsic geometric information encoded in the spectral embedding, yielding better generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Dual Map Representations Thanks to the inherent connection and efficient conversion between point-wise and functional maps, it has long been observed that jointly estimating both map representations can improve the mapping quality. For instance, in the original work [37], the authors have already proposed to apply an ICP-like technique on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14498" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "functional maps. More recent advances take advantage of the multi-scale properties of the eigenbasis of the Laplace-Beltrami operator. In the works [34, 24, 13, 15], conversions are done between spatial domain and a series of spectral domains spanned by eigenfunctions of increasing dimensions. This idea has also been incorporated into DFM. In AttentiveFMaps [28], the authors propose to fuse functional maps of different dimensions by converting them into the common spatial domain. This fusion technique in turn allows training an attention network to dynamically choose the optimal spectral resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.381 + ], + "angle": 0, + "content": "The above methods enforce consistency across map representations by iterative projections, which is computationally heavy. By exploiting our theoretical insight on the spectral cycle consistency of DFM, we only introduce a marginal computational overhead to a standard DFM trained on a single spectral resolution. Our method is much lighter and simpler but also shows superior performance in accuracy and generalization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.47, + 0.412 + ], + "angle": 0, + "content": "3. Cycle Consistency of Deep Functional Maps" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.464 + ], + "angle": 0, + "content": "In this section, we first briefly review the general deep functional maps pipeline. Then we present a theoretical analysis of its cycle consistency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.473, + 0.289, + 0.49 + ], + "angle": 0, + "content": "3.1. Deep Functional Maps" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.469, + 0.573 + ], + "angle": 0, + "content": "We assume to be given a pair of deformable shapes \\( S_{1} \\) and \\( S_{2} \\), which are discretized as triangular meshes of \\( n_{1} \\) and \\( n_{2} \\) vertices, respectively. The generic deep functional map pipeline introduced in [30] learns a map that spectrally aligns the shapes through the following four steps:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.47, + 0.659 + ], + "angle": 0, + "content": "1. Compute the leading \\(k\\) eigenfunctions of the Laplace-Beltrami operator on each shape, which can be treated as a high-dimensional spectral embedding of the respective shape. The eigenfunctions are stored as matrices \\(\\Phi_{i}\\in \\mathbb{R}^{n_{i}\\times k},i = 1,2\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.47, + 0.795 + ], + "angle": 0, + "content": "2. Instantiate a feature extractor network, \\(\\mathcal{F}_{\\Theta}\\), where \\(\\Theta\\) denotes the set of learnable parameters. By feeding forward shapes through \\(\\mathcal{F}_{\\Theta}\\), descriptors are obtained and expected to be approximately preserved by the underlying map. In general, we denote \\(\\mathbf{G}_i = \\mathcal{F}_{\\Theta}(S_i) \\in \\mathbb{R}^{n_i \\times d}\\), \\(i = 1, 2\\), where \\(d\\) is the predefined number of descriptors. They are then projected onto the eigenbasis above, resulting in a couple of coefficient matrices \\(\\mathbf{A}_i = \\boldsymbol{\\Phi}_i^\\dagger \\mathbf{G}_i \\in \\mathbb{R}^{k \\times d}\\), \\(i = 1, 2\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.469, + 0.825 + ], + "angle": 0, + "content": "3. Estimate the optimal functional map, \\(\\mathbf{C}^*\\), by solving the following linear system:" + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.585, + 0.47, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.837, + 0.47, + 0.856 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} ^ {*} = \\arg \\min _ {\\mathbf {C}} E _ {\\mathrm {d e s c}} (\\mathbf {C}) + E _ {\\mathrm {r e g}} (\\mathbf {C}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.87, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(E_{\\mathrm{desc}}(\\mathbf{C}) = \\| \\mathbf{CA}_1 - \\mathbf{A}_2\\|^2\\), and \\(E_{\\mathrm{reg}}(\\mathbf{C})\\) is the regularization term promoting structural properties" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.891, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.206, + 0.894, + 0.29 + ], + "angle": 0, + "content": "Figure 2. Overview of our two-branch DFM network. The part colored in black corresponds to a standard DFM, we introduce a novel branch, colored in red, that estimates maps from spatial perspective, our loss function is defined as Eqn.6. \\(\\Phi_i,\\Phi_i^\\dagger\\) indicate projection of the regarding features into the spatial and spectral domain, respectively. See more details in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.892, + 0.342 + ], + "angle": 0, + "content": "of \\(\\mathbf{C}\\) , e.g., enforcing \\(\\mathbf{C}\\) to be commutative with the Laplace-Beltrami operators and to be orthogonal. In particular, we let \\(E_{\\mathrm{total}}(\\mathbf{C}) = E_{\\mathrm{desc}}(\\mathbf{C}) + E_{\\mathrm{reg}}(\\mathbf{C})\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.892, + 0.386 + ], + "angle": 0, + "content": "4. Convert the estimated functional map \\(\\mathbf{C}\\) to a pointwise map by conducting nearest neighbor search between the rows of \\(\\Phi_1\\mathbf{C}\\) and that of \\(\\Phi_{2}\\)." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.296, + 0.892, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.893, + 0.566 + ], + "angle": 0, + "content": "The feature extractor is then trained on a set of training shape pairs, which typically enumerates all pairs from some given training set. The sub-optimization problem in Step (3) is generally decomposed into two parts to circumvent using an iterative solver. Namely, a closed-form solution is obtained either by a least-square estimation w.r.t \\( E_{\\mathrm{desc}}(\\mathbf{C}) \\) [30], or by solving a sequence of linear systems involving Laplacian commutativity as well [11]. Both correspond to the FMreg layer shown in Fig. 2. The rest of \\( E_{\\mathrm{reg}}(\\mathbf{C}) \\) is then set as the structural loss, corresponding to \\( \\mathcal{L}(\\mathbf{C}) \\) in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.676, + 0.591 + ], + "angle": 0, + "content": "3.2. Cycle Consistency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.749 + ], + "angle": 0, + "content": "We start by giving a formal definition of cycle consistency. We assume to be given a collection of shapes \\( S = \\{S_i\\}_{i=1}^n \\) and a set of maps \\( \\mathcal{T} = \\{T_{ij}\\}_{i,j \\in [1..n]} \\). We call \\( \\mathcal{T} \\) cycle consistent if for any shape \\( S_i \\), and any closed path \\( (i, i_1, i_2, \\dots, i_k, i) \\), the map composition along this path \\( T_{ii} = T_{i_k i} \\circ \\dots \\circ T_{i_1 i_2} \\circ T_{ii_1} \\) is an identity map on \\( S_i \\). We can similarly define cycle consistency on functional maps, the only difference is that map composition is given by matrix multiplication and we require the final functional map \\( C_{ii} \\) to be an identity matrix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.826 + ], + "angle": 0, + "content": "We suppose that deep functional maps are trained on \\(S\\) with respect to all possible pairs. Then the global energy is given by \\(E_{\\mathrm{total}}(\\mathcal{C}) = \\sum_{i,j} \\| \\mathbf{C}_{ij} \\mathbf{A}_i - \\mathbf{A}_j \\|^2 + \\sum_{i,j} E_{\\mathrm{reg}}(\\mathbf{C}_{ij})\\), where \\(\\mathcal{C} = \\{\\mathbf{C}_{ij}\\}_{i,j \\in [1..n]}\\) is the set of functional maps among training shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Proposition 1 If \\( E_{\\text{total}}(\\mathcal{C}) = 0 \\), then for any shape \\( S_i \\), and any path \\( (i, i_1, i_2, \\dots, i_k, i) \\), the map composition \\( \\mathbf{C}_{ii} \\) is cycle consistent within the functional space spanned by the columns of \\( \\mathbf{A}_i \\), i.e., \\( \\mathbf{C}_{ii}\\mathbf{A}_i = \\mathbf{A}_i \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14499" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.182 + ], + "angle": 0, + "content": "As a consequence of Prop. 1, when \\(\\mathbf{A}_i\\in \\mathbb{R}^{k\\times d}\\) is of full row rank, then \\(\\mathcal{C} = \\{\\mathbf{C}_{ij}\\}_{i,j\\in [1\\dots n]}\\) is cycle consistent. In practice, we generally set \\(d > k\\). Moreover, during network training, \\(\\mathbf{C}_{ij}\\) is computed via pseudo-inverse, which implicitly assumes the full-rankness of \\(\\mathbf{A}_i,\\mathbf{A}_j\\). We refer readers to the empirical validation in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.469, + 0.243 + ], + "angle": 0, + "content": "We defer the proof of Prop. 1 to Supp. Material. In fact, a similar claim has been formulated and proven in [49] (see Sec.3.1 therein), but in the context of map refinement via promoting cycle consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.469, + 0.394 + ], + "angle": 0, + "content": "Though being technically similar, the theoretical argument of [49] and that of ours have fundamentally different implications. More specifically, the former justifies a test-time optimization algorithm, which is used to promote cycle consistency of maps among a fixed test shape collection. While the latter suggests that spectrally cycle consistency has been ensured and further leveraged to enhance the universal feature extractor (independent of the test data) during training in any DFM framework following the generic pipeline presented in Sec. 3.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.408, + 0.398, + 0.426 + ], + "angle": 0, + "content": "4. Two-branch Deep Functional Maps" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.469, + 0.585 + ], + "angle": 0, + "content": "It has long been recognized, both theoretically and empirically, that optimizing purely in the spectral domain is not sufficient. As a toy example, a trivial solution attaining global optima can be constructed as follows: Suppose that we have learned a feature extractor \\(\\mathcal{F}_{\\Theta}\\), which returns the respective eigenbasis transformed by a universal \\(\\mathbf{A}_0\\). That is, \\(\\mathbf{G}_i = \\boldsymbol{\\Phi}_i\\mathbf{A}_0, \\forall i\\), which implies \\(\\mathbf{A}_i = \\boldsymbol{\\Phi}_i^\\dagger\\mathbf{G}_i = \\mathbf{A}_0, \\forall i\\). Then we have \\(\\mathbf{C}_{ij} \\equiv I_k\\), which exactly satisfies \\(E_{\\mathrm{desc}}(\\mathbf{C}_{ij}) = E_{\\mathrm{reg}}(\\mathbf{C}_{ij}) = 0, \\forall i, j\\). However, it probably induces poor point-wise maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.469, + 0.706 + ], + "angle": 0, + "content": "In fact, in [37] the authors have proposed to use an ICP-like technique to encourage the estimated functional maps to be induced by some point-wise maps. In [34], the authors propose a spectral upsampling method for map refinement, which essentially converts maps back and forth between spectral and spatial domains. Moreover, the following lemma from [34] sheds light on the necessity of taking both spectral and spatial representations into consideration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.469, + 0.799 + ], + "angle": 0, + "content": "Lemma 1 Given a pair of shapes \\( S_{1}, S_{2} \\) each having nonrepeating Laplacian eigenvalues, which are the same. A point-wise map \\( T: S_{1} \\to S_{2} \\) is an isometry if and only if the corresponding functional map \\( \\mathbf{C} \\) in the complete Laplacian basis is both diagonal and orthonormal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.87 + ], + "angle": 0, + "content": "The above lemma suggests that apart from promoting the structural properties of functional maps, it is also critical to enforce them to be associated with certain point-wise maps, or termed as properness of functional maps in [40]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Finally, we remark that some recent DFM advances also promote the properness of the resulting spectral maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "For instance, AttentiveFMaps [28] follows the spirit of ZoomOut [34] and explicitly performs a conversion between spectral and spatial map representations across different dimensions of eigenbasis; UDMSM [9] constructs explicitly a universal shape in the feature space, and enforce the spectral map estimation to be consistent with the spatial maps induced via the universal shape." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.207, + 0.759, + 0.223 + ], + "angle": 0, + "content": "4.1. Two-branch Map Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.291 + ], + "angle": 0, + "content": "In this part, we leverage our observation made in Prop. 1 and propose a novel, simple yet effective design of unsupervised deep functional maps, which introduces a new branch that independently estimates maps from spatial perspective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.893, + 0.638 + ], + "angle": 0, + "content": "Our key insight is that, once cycle consistency is valid and \\(\\mathbf{A}_i\\) is of full row rank, \\(\\mathbf{A}_i\\) can be seen as a functional map from a universal latent shape, \\(S_0\\), to \\(S_i\\). This perspective has been explored in several prior works [49, 23, 24], we provide the following details to be self-contained. The above assumption implies \\(\\mathbf{C}_{ij} = \\mathbf{A}_j\\mathbf{A}_i^\\dagger\\). Then \\(\\mathbf{C}_{ij}\\) can be interpreted as a functional map composition from \\(S_i\\) to \\(S_0\\), followed by a map from \\(S_0\\) to \\(S_j\\). On the other hand, one can align the spectral embeddings of \\(S_j\\) to that of \\(S_i\\) by simply transforming the former by \\(\\mathbf{C}_{ij}\\). Indeed, we convert \\(\\mathbf{C}_{ij}\\) into the point-wise map by the nearest neighbor searching between the rows of \\(\\Phi_j\\mathbf{C}_{ij}\\) and that of \\(\\Phi_i\\). From this point of view, denoting the virtual spectral embedding of the latent shape by \\(\\Phi_0\\), \\(\\Phi_i\\mathbf{A}_i\\) can be then treated as the spectral embedding of \\(S_i\\) aligned to that of \\(S_0\\). Therefore, given a pair of shapes \\(S_i, S_j\\), since we have aligned their eigenbasis to the canonical frame defined by the virtual spectral embedding \\(\\Phi_0\\), we can align the spectral embedding of \\(S_i\\) to \\(\\Phi_0\\) by computing \\(\\Phi_i\\mathbf{A}_i\\). Once all the spectral embeddings are aligned to the canonical embedding domain regarding \\(\\Phi_0\\), we can compute the soft point-wise map between \\(S_i\\) and \\(S_j\\) by nearest neighbor searching between the rows of \\(\\Phi_i\\mathbf{A}_i\\) and those of \\(\\Phi_j\\mathbf{A}_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Based on the above derivation, given the learned features projected in the spectral domain, \\(\\mathbf{A}_i,\\mathbf{A}_j\\), and a pair of indices \\(p\\in [1..n_i],q\\in [1..n_j]\\), we can compute point-wise maps. Firstly we compute residual:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.711, + 0.891, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\delta_ {q p} = \\left\\| \\Phi_ {i} [ p ] \\mathbf {A} _ {i} - \\Phi_ {j} [ q ] \\mathbf {A} _ {j} \\right\\| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.739, + 0.892, + 0.784 + ], + "angle": 0, + "content": "where \\(\\Phi_i[p]\\) denotes the \\(p\\)-th row of \\(\\Phi_i\\), and similarly we define \\(\\Phi_j[q]\\). The soft point-wise map \\(\\Pi \\in \\mathbb{R}^{n_j \\times n_i}\\) is then given by:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.796, + 0.891, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\Pi (q, p) = \\frac {\\exp (- \\alpha \\delta_ {q p})}{\\sum_ {p ^ {\\prime}} \\exp (- \\alpha \\delta_ {q p ^ {\\prime}})}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Note that by construction, each row of \\(\\Pi\\) is non-negative and sums up to 1, forming a probability distribution. The parameter \\(\\alpha\\) controls the entropy of each distribution - the smaller/larger \\(\\alpha\\) is, the fuzzier/sharper the distribution is." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14500" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.45, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.47, + 0.3 + ], + "angle": 0, + "content": "Figure 3. We train our two-branch DFM and a vanilla single-branch version on DT4D-H and monitor spectrally and spatially cycle consistency along the training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.469, + 0.364 + ], + "angle": 0, + "content": "Instead of manually tuning the optimal \\(\\alpha\\), we propose a learning scheme that dynamically controls \\(\\alpha\\) over training, which is inspired by curriculum learning [5]. We defer the respective details to Section 4.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.469, + 0.395 + ], + "angle": 0, + "content": "We convert the soft point-wise map to a functional map by" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.393, + 0.469, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {2} = \\boldsymbol {\\Phi} _ {j} ^ {\\dagger} \\Pi \\boldsymbol {\\Phi} _ {i}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.469, + 0.446 + ], + "angle": 0, + "content": "In the end, we enforce \\(\\mathbf{C}_2\\) to be consistent with \\(\\mathbf{C}_1\\), the intermediate output from the FMreg layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.469, + 0.537 + ], + "angle": 0, + "content": "To summarize, thanks to the spectrally cycle consistency we identify in Sec. 3.2, we are allowed to construct a spectral latent shape and induce spatial maps via it. By enforcing the spatial estimations to be consistent with the spectral ones, we obtain a spatially and spectrally consistent deep functional maps framework." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.238, + 0.561 + ], + "angle": 0, + "content": "4.2. Network Design" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.629 + ], + "angle": 0, + "content": "As shown in Fig. 2, our two-branch network is built upon a standard DFM framework. In the following, we denote by \\(\\mathbf{C}_1\\) and \\(\\mathbf{C}_2\\) the estimated functional maps from the original FMreg layer and our novel branch, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.72 + ], + "angle": 0, + "content": "Specifically, we use DiffusionNet [44] as our feature extractor. And WKS [4] descriptors are fed into it as initialization of learned features. We borrow the FMreg layer from [12]. It takes both \\( E_{\\mathrm{desc}}(\\mathbf{C}) \\) and commutativity with the Laplace-Beltrami operator into consideration, where the latter is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.728, + 0.469, + 0.751 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {l a p}} = \\left\\| \\mathbf {C} _ {1} \\Lambda_ {1} - \\Lambda_ {2} \\mathbf {C} _ {1} \\right\\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.469, + 0.789 + ], + "angle": 0, + "content": "where \\(\\Lambda_{1}\\) and \\(\\Lambda_{2}\\) are diagonal matrices of the Laplace-Beltrami eigenvalues on the two shapes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.469, + 0.82 + ], + "angle": 0, + "content": "The estimation of \\(\\mathbf{C}_2\\) has been described in detail in Section 4.1. In the end, we formulate the training loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.828, + 0.469, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\mathbf {C} _ {1}, \\mathbf {C} _ {2}\\right) = \\left\\| \\mathbf {C} _ {1} ^ {T} \\mathbf {C} _ {1} - \\mathbf {I} \\right\\| ^ {2} + \\left\\| \\mathbf {C} _ {1} - \\mathbf {C} _ {2} \\right\\| ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where the first term promotes the orthogonality of \\(\\mathbf{C}_1\\), while the second term promotes the consistency between the functional maps estimated from different branches." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Finally, we remark that by combining the FMreg layer and \\(\\mathcal{L}(\\mathbf{C}_1,\\mathbf{C}_2)\\), we have incorporated every factor in Lemma 1 into our design." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.156, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Conceptual Validation In this part, we train a network on DT4D-H dataset (see Section 5.1 for details) with our two-branch network, and a single-branch variant without our spatial map estimation branch. We monitor and plot the following quantities along training: (1) Average spectral cycle consistency over sampled triplets, i.e., \\(\\frac{1}{M}\\sum_{(i,j,k)}\\| \\mathbf{C}_{ki}\\mathbf{C}_{jk}\\mathbf{C}_{ij} - I\\| ^2 /\\| I\\| ^2\\) ; (2) Average spatial cycle consistency over sampled triplets, i.e., the mean Euclidean deviation from composed maps \\(T_{ki}\\circ T_{jk}\\circ T_{ij}\\) to identity map on \\(S_{i}\\). Here \\(n = 80\\) is the number of training shapes, and \\(M = 1000\\) is the number of sampled triplets. The behavior of the blue curves after 4500 iterations verifies our argument that spectrally cycle consistency does not imply spatially cycle consistency. On the other hand, by introducing our two-branch design, the discrepancy is well compensated and evidently better cycle consistencies in both spatial and spectral domains are achieved." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.422, + 0.794, + 0.438 + ], + "angle": 0, + "content": "4.3. Updating Scheme of \\(\\alpha\\) in Eqn. (3)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.445, + 0.892, + 0.611 + ], + "angle": 0, + "content": "The soft point-map conversion (Eqn. 3) has been applied in several prior works [30, 28], which all set \\(\\alpha\\) to be a manually selected constant. Ideally, we expect \\(\\Pi\\) in Eqn. (3) to be close to a permutation matrix, i.e., each row forms a binary vector. This seems to suggest a preference for a large \\(\\alpha\\). Unfortunately, it would severely hinder network training, since the learned features and maps are of low quality in the early stage. On the other hand, a small \\(\\alpha\\) can alleviate such difficulty but falls short of fully pushing functional maps to be proper. As demonstrated in Section 5.5, neither small nor large \\(\\alpha\\) produces satisfying results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.612, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Based upon the above analysis, we propose a novel updating scheme, which is inspired by curriculum learning [5]. Namely, we initiate a small \\(\\alpha\\) at the beginning of network training and increase it by a constant step size for every fixed number of epochs. As shown in Table 1, 2, 3, our scheme does not rely on hyper-parameter tuning but also achieves state-of-the-art results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.717, + 0.742 + ], + "angle": 0, + "content": "4.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We implement our network with PyTorch [39]. We use four DiffusionNet blocks [44] as feature backbone and borrow the functional map block with Laplacian regularizer from [12]. The dimension of the Laplace-Beltrami eigenbasis is set to 50. WKS [4] descriptors are used as the input signal to our network. The dimensions of the input and the output descriptors are both set to 128. During training, the value of the learning rate is set to 2e-4 with ADAM optimizer. In all experiments, we train our method for 10,000 iterations with a batch size of 1. Following the learning" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14501" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "strategy in Section 4.3, we initialize \\(\\alpha\\) to 1 and increase it by 5 per epoch. As indicated in Eqn. (6), We weigh equally the orthogonality loss with respect to \\(\\mathbf{C}_1\\) and the residual between \\(\\mathbf{C}_1\\) and \\(\\mathbf{C}_2\\). More implementation details are provided in the Supp. Material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.189, + 0.283, + 0.208 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.219, + 0.471, + 0.415 + ], + "angle": 0, + "content": "In this section, we conduct an extensive set of experiments of non-rigid shape matching on various datasets including humanoids and animals. We test on both near-isometric and non-isometric shape pairs. Our method is compared to a set of competitive baselines including axiomatic, supervised, weakly-supervised, and unsupervised learning methods. We emphasize that in this section, all the maps from the learning-based pipelines are directly inferred from the trained models, without any post-processing procedure. We evaluate the matching results in terms of a mean geodesic error on shapes normalized to unit area. Finally, our point-wise maps are all inferred by converting the output functional maps, as all the other DFM frameworks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.433, + 0.182, + 0.449 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.469, + 0.506 + ], + "angle": 0, + "content": "FAUST_r: The remeshed version [41] of FAUST dataset[6] contains 100 human shapes. Following [43], it is split into 80/20 for training and testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.553 + ], + "angle": 0, + "content": "SCAPE_r: The remeshed version [41] of SCAPE dataset[2] contains 71 human shapes. Following [43], it is split into 51/20 for training and testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.469, + 0.632 + ], + "angle": 0, + "content": "SHREC19_r: The remeshed version of SHREC19 dataset [33] collects 44 human shapes from 11 independent datasets with distinctive poses and styles. We abandon shape 40 due to its partiality, we test on 407 pairs among the rest 43 shapes, which come with ground-truth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.469, + 0.71 + ], + "angle": 0, + "content": "DT4D-H [32]: The remeshed subset of the large scale animation dataset DeformingThings4D [29]. In particular, DT4D-H includes 10 categories of humanoid shapes undergoing significant pose and style variances, forming a challenging benchmark." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.469, + 0.774 + ], + "angle": 0, + "content": "SMAL_r: The remeshed SMAL dataset [53] contains 49 animal shapes with 8 species. We follow the setting from [28], which splits 29 (5 species) and 20 (3 species) shapes for training and testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.469, + 0.868 + ], + "angle": 0, + "content": "TOSCA_r: The remeshed TOSCA dataset [7] contains multiple shape categories. We choose 4 animal categories, including cat, dog, horse, and wolf to verify the generalization performance of networks trained on SMAL_r. Note that we only infer the intra-category maps, due to the absence of ground-truth inter-category maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We refer readers to Supp. Material for visualizations illustrating the variability of the above datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.782, + 0.108 + ], + "angle": 0, + "content": "5.2. Near-isometric Shape Matching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.251 + ], + "angle": 0, + "content": "In this part, we perform comparisons with an array of non-rigid shape matching methods: (1) Axiomatic methods including ZoomOut [34], BCICP [41], IsoMuSh [17], Smooth Shells [13], CZO [24]; (2) Supervised learning methods including TransMatch [48], GeomFMaps [12], and supervised version of AttentiveFMaps [28]; (3) Unsupervised learning methods including NeuroMorph [14], SyNoRiM [16], Deep Shell [15], AttentiveFMaps [28], UDMSM [9], DUO-FM [11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.252, + 0.892, + 0.463 + ], + "angle": 0, + "content": "For all learning-based methods, we train models on FAUST_r and SCAPE_r respectively. Table 1 reports results on both standard tests and more challenging generalizations. We observe a trade-off between the two tasks, methods that perform the best in the former (e.g., supervised AttentiveFMaps and UDMSM) tend to overfit, and therefore suffer poor generalization (especially to SREHC19_r). Meanwhile, our default setting, denoted by Ours, achieves reasonable performance in the standard tests but also outperforms the external baselines in 3 out of 4 generalization tests. Especially, in generalizing to SHREC19_r, Ours outperforms the external baselines by a large margin, resulting in \\(41\\%\\) (3.8 vs. 6.4) and \\(46\\%\\) (4.5 vs. 8.4) error reduction upon the second best." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.892, + 0.586 + ], + "angle": 0, + "content": "We highlight that post-processing with cycle consistency generally depends on the initialized map quality and the size of the test set (e.g., \\(\\geq 3\\) shapes). In contrast, we leverage cycle consistency to improve the feature extractor during training. We also report the results from post-processing techniques based on cycle consistency [17, 24, 16] in Table 1. They are significantly outperformed by our method, which is inferred per-pair and without any post-processing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.753 + ], + "angle": 0, + "content": "We further augment the dimension of functional maps in network training to 80 (same as UDMSM), which is beneficial to near-isometric matching [28]. It is evident that Ours (80 dim) achieves on-par performance with the regarding state-of-the-art methods in standard tests. On the other hand, due to the significant variability between SHREC19_r and the training sets (see Supp. Material), augmenting dimension leads to worse generalization than before (5.5 vs. 3.8, 5.8 vs. 4.5). Nevertheless, even in this case, our method outperforms the external baselines in all generalization tests by a notable margin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.764, + 0.776, + 0.781 + ], + "angle": 0, + "content": "5.3. Non-isometric Shape Matching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.788, + 0.892, + 0.849 + ], + "angle": 0, + "content": "We also train our network on non-isometric datasets, SMAL_r and DT4D-H, and compare it with the state-of-the-art baselines including DeepShells [15], AttentiveFMaps [28], UDMSM [9] and DUO-FM [11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SMAL_r: We follow the split and input descriptors from [28] (more details are provided in the Supp. Material). Table 2 reports results on SMAL_r, our method" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14502" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.088, + 0.854, + 0.101 + ], + "angle": 0, + "content": "Table 1. Mean geodesic errors (×100) on FAUST_r, SCAPE_r, and SHREC19_r. The best and the second best are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.102, + 0.894, + 0.342 + ], + "angle": 0, + "content": "
MethodTrain TestFAUST_rSCAPE_r
FAUST_rSCAPE_rSHREC19_rSCAPE_rFAUST_rSHREC19_r
ZM[34]6.1\\\\7.5\\\\
BCICP[41]6.4\\\\11.0\\\\
IsoMuSh[17]4.4\\\\5.6\\\\
Smooth Shell[13]2.5\\\\4.7\\\\
CZO[24]2.2\\\\2.5\\\\
TransMatch[48]2.733.621.018.318.638.8
GeomFMaps[12]supervised2.63.39.93.03.012.2
AttentiveFMaps[28]1.42.29.41.71.812.2
NeuroMorph[14]8.528.526.329.918.227.6
SyNoRiM[16]7.921.725.59.524.626.8
Deep Shell[15]1.75.427.42.52.723.4
AttentiveFMaps[28]unsupervised1.92.66.42.22.29.9
UDMSM[9]1.57.321.52.08.630.7
DUO-FM[11]2.54.26.42.72.88.4
Ours2.32.63.82.42.54.5
Ours (80 dim)1.72.65.52.22.05.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.362, + 0.468, + 0.389 + ], + "angle": 0, + "content": "Table 2. Mean geodesic errors \\((\\times 100)\\) on SMAL_r. The best and the second best are highlighted correspondingly." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.389, + 0.46, + 0.514 + ], + "angle": 0, + "content": "
MethodTrain TestSMAL_r
SMAL_rTOSCA_r
DeepShell[15]29.38.7
GeomFMaps[12]7.624.5
AttentiveFMaps[28]5.420.9
UDMSM[9]24.621.7
DUO-FM[11]32.815.3
Ours5.47.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.47, + 0.726 + ], + "angle": 0, + "content": "achieves the best performance, which is on-par with AttentiveFMaps [28]. To evaluate generalization performance, we use the trained models to directly infer intra-category maps within TOSCA-r. It turns out that AttentiveFMaps and GeomFMaps both suffer from significant performance drops (\\( \\times 3.8 \\) and \\( \\times 3.2 \\) larger geodesic error). It is also worth noting that DeepShells achieves the second-best generalization score in the relatively simpler task. However, it fails dramatically regarding the base task. In contrast, our method achieves the best balance between learning in difficult non-isometric pairs and generalizing to relatively easy near-isometric pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "DT4D-H: We follow the train/test (198/95) split of [28], but ignore the categories mousey and ortiz in both train and test, due to the lack of inter-category map labels regarding them, resulting a split of 168/80. We emphasize that we conduct training and test in a category-agnostic manner, i.e., no class label is used, and the training pairs can consist of shapes from arbitrary two categories. This is significantly different from [28], in which training pairs are selected according to clustering information. Obviously, our setting is more practical, but also more challenging. For completeness, we report results under the setting of [28] in Supp. Material" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.366, + 0.655, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.521, + 0.645, + 0.535 + ], + "angle": 0, + "content": "\\(T_{11}^{(8)} = T_{81}\\circ T_{78}\\circ \\dots \\circ T_{23}\\circ T_{12}\\)" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.365, + 0.819, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.447, + 0.715, + 0.455 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.366, + 0.896, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.447, + 0.867, + 0.455 + ], + "angle": 0, + "content": "UDMSM" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.457, + 0.82, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.537, + 0.727, + 0.546 + ], + "angle": 0, + "content": "DUO-FM" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.457, + 0.82, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.537, + 0.802, + 0.546 + ], + "angle": 0, + "content": "Deep Shell" + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.457, + 0.896, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.835, + 0.537, + 0.866, + 0.546 + ], + "angle": 0, + "content": "BCICP" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Figure 4. Qualitative evaluation of spatial cycle consistency of different methods. Even composed along a path of 8 highly deformed shapes, our resulting map remains close to identity, while all the baselines fail significantly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.653 + ], + "angle": 0, + "content": "and our method outperforms the baselines in both intra- and inter-category evaluation by a notable margin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Table 3 reports results on DT4D-H, in which we preserve 80 shapes for test and train networks with 168 and 80 shapes, respectively. Note that we report mean geodesic errors over all possible test shape pairs, which may undergo significant distortions (see, e.g., Fig. 1). Our method obtains a \\(67.2\\%\\) (7.7 vs. 22.4) geodesic error reduction with respect to the second-best baseline. On top of that, we also test the generalization of the trained model on near-isometric benchmarks – our method also generalizes the best in generalization to FAUST_r and SCAPE_r. The same pattern is observed when the training set is reduced by more than half. Remarkably, our network trained on the reduced set still outperforms all the baselines trained on the full set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Overall, we attribute our performance on matching challenging non-isometric shapes (Table 3) and generalizing to unseen shapes (Table 1) to our effort to promote both spec" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14503" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.088, + 0.819, + 0.101 + ], + "angle": 0, + "content": "Table 3. Mean geodesic errors \\( \\left( {\\times {100}}\\right) \\) on DT4D-H. The best and the second best are highlighted correspondingly." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.101, + 0.894, + 0.199 + ], + "angle": 0, + "content": "
MethodTrain TestDT4D-H (168)DT4D-H (80)
DT4D-HFAUST_rSCAPE_rDT4D-HFAUST_rSCAPE_r
DeepShell[15]27.04.96.529.34.77.0
AttentiveFMaps[28]25.73.46.428.92.76.3
UDMSM[9]46.843.347.949.742.540.0
DUO-FM[11]22.410.012.224.78.09.2
Ours7.73.16.19.02.66.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.214, + 0.468, + 0.24 + ], + "angle": 0, + "content": "Table 4. Mean geodesic errors \\( \\left( {\\times {100}}\\right) \\) of SURFMNet and our variant trained on 4 datasets" + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.239, + 0.438, + 0.318 + ], + "angle": 0, + "content": "
MethodSURFMNetSURFMNet + Ours
FAUST_r6.03.5
SCAPE_r6.83.4
SMAL_r20.413.3
DT4D-H18.315.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.468, + 0.405 + ], + "angle": 0, + "content": "trally and spatially cycle consistency. Especially, the isometry assumption is likely violated in the former case, thus cycle consistency, as a generic prior, plays an important role of regularizing maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.468, + 0.557 + ], + "angle": 0, + "content": "As an illustration, we present a qualitative evaluation on the point-wise cycle consistency in Fig. 4. We sample 8 shapes from the test set of DT4D-H (one from each category) and compose the maps along the path \\((S_{1} \\to S_{2} \\to \\dots \\to S_{8} \\to S_{1})\\) with respect to different approaches. It is evident that due to the significant distortion undergoing among the shapes, all but our method fails to preserve cycle consistency in this demanding test, while our composing map approximates the identity map on \\(S_{1}\\). It also aligns nicely with the quantitative results reported in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.37, + 0.584 + ], + "angle": 0, + "content": "5.4. Integration with SURFMNet [43]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.468, + 0.773 + ], + "angle": 0, + "content": "Our two-branch design can be easily incorporated into any existing DFM framework following the general design outlined in Sec. 3.1. To demonstrate this, we modify the SURFMNet [43], one of the earliest approaches of unsupervised DFM, by adding our new branch. Table 4 shows the matching accuracy on the four main benchmarks. It is evident that in every case, incorporating our design leads to significant error reduction ranging from \\(18\\%\\) to \\(50\\%\\). Especially, in the near-isometric cases, we obtain \\(41.6\\%\\) and \\(50\\%\\) error reduction respectively. Note that the absolute scores, 0.035, 0.034, are reasonable even compared to the state-of-the-art results reported in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.229, + 0.802 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.884 + ], + "angle": 0, + "content": "In this section, we present a set of ablation studies consisting of two parts. The first part verifies the rationality of our method, and the second part demonstrates the robustness of our method. We conduct all experiments on SMAL_r dataset [53]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "First of all, instead of using the updating scheme in Sec-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.218, + 0.892, + 0.324 + ], + "angle": 0, + "content": "tion 4.3, we test the performance of our pipeline using two fixed values of \\(\\alpha\\) in Eqn. (3): \\(\\alpha = 1\\) and \\(\\alpha = 50\\). Compared to our proposed model, the two variants yield a noticeable performance drop. Especially in the case \\(\\alpha = 50\\), the network fails to deliver reasonable matching results. We believe it is because a large \\(\\alpha\\) amplifies the noise of maps learned at the early training stage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Then we justify our two-branch network design. Removing spatial branch amounts to training a standard single-branch DFM. To remove the spectral branch, we remove the FMreg layer and instead use our new branch to compute point-wise maps and convert them to functional maps. In the end, we modify the training loss so that it covers descriptor preservation, commutativity with the Laplace-Beltrami operator, and orthogonality (the latter two compensate the removed FMreg layer). The accuracy drop reported in the third and fourth row of Table 5 clearly suggests the necessity of our two-branch design." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.75 + ], + "angle": 0, + "content": "In our experiments, we always use the full resolution meshes (\\(\\sim 5k\\) vertices) and compute in Eqn. (2) with all of the 128 descriptors. We anticipate that efficiency can become an issue when the input mesh resolution is high, and/or we would like to increase the size of learned descriptors. Therefore, we test the robustness of our pipeline with respect to down-sampling, which is commonly used in functional maps-based frameworks [34, 28]: 1) We down-sample 3000 vertices on each shape via furthest point sampling; 2) In order to down-sample the feature dimension, we operate as the following during training: given a \\(\\mathbf{A}_1\\), \\(\\mathbf{A}_2\\), we perform SVD on \\(\\mathbf{A}_1\\), i.e., \\(\\mathbf{A}_1 = U_1\\Sigma_1V_1^T\\), then we set \\(\\hat{\\mathbf{A}}_1 = \\mathbf{A}_1\\hat{V}_1\\), and set \\(\\hat{\\mathbf{A}}_2 = \\mathbf{A}_2\\hat{V}_1\\), where \\(\\hat{V}_1\\) is the first \\(m\\) columns of \\(V_{1}\\). We set \\(m = 30\\) by replacing \\(\\mathbf{A}_i\\) with \\(\\hat{\\mathbf{A}}_i\\) in Eqn. (2). The results in the bottom two rows show that the above operation has a relatively minor effect on the performance, proving the robustness of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.768, + 0.619, + 0.784 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To conclude, we provide a theoretical justification for spectral cycle consistency in DFM. To compensate for the discrepancy of purely spectrally cycle consistency, we formulate a spectral latent shape that allows the alignment of the spectral embeddings of each shape to a canonical embedding domain. Based on this construction, we introduce a two-branch architecture for estimating maps in both spec" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14504" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.113 + ], + "angle": 0, + "content": "Table 5. Mean geodesic errors \\((\\times 100)\\) on different ablated settings, the models are all train on SMAL_r." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.113, + 0.388, + 0.224 + ], + "angle": 0, + "content": "
Ours5.4
α = 16.6
α = 5035.2
remove spatial branch33.4
remove spectral branch14.3
Vertex downsampling5.7
Feature downsampling5.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.47, + 0.353 + ], + "angle": 0, + "content": "tral and spatial domains within DFM. The resulting network is simple, computationally efficient, and compatible with most existing DFM frameworks. We demonstrate the effectiveness of our framework through a comprehensive set of experiments, showing significant improvements upon state-of-the-art approaches in terms of accuracy, consistency, and generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.469, + 0.429 + ], + "angle": 0, + "content": "It is worth noting, though, our method is developed upon clean and complete meshes. It would be interesting for future work to investigate the utility of our approach in more general tasks, involving partial meshes, noisy point clouds, and other representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Acknowledgement This work was supported in part by the National Natural Science Foundation of China under contract No. 62171256 and Shenzhen Key Laboratory of next-generation interactive media innovative technology (No. ZDSYS20210623092001004), and in part by the ERC Starting Grant No. 758800 (EXPROTEA) and the ANR AI Chair AIGRETTE." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.566, + 0.173, + 0.581 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.469, + 0.632 + ], + "angle": 0, + "content": "[1] Noam Aigerman, Roi Poranne, and Yaron Lipman. Seamless Surface Mappings. ACM Transactions on Graphics (TOG), 34(4):72, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape Completion and Animation of People. In ACM Transactions on Graphics (TOG), volume 24, pages 408-416. ACM, 2005. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.705, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[3] Souhaib Attaiki, Gautam Pai, and Maks Ovsjanikov. Dpfm: Deep partial functional maps. In 2021 International Conference on 3D Vision (3DV), pages 175-185. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.748, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[4] Mathieu Aubry, Ulrich Schlickewei, and Daniel Cremers. The Wave Kernel Signature: A Quantum Mechanical Approach to Shape Analysis. In Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on, pages 1626-1633. IEEE, 2011. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.818, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[5] Yoshua Bengio, Jerome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In International Conference on Machine Learning, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.861, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[6] Federica Bogo, Javier Romero, Matthew Loper, and Michael J. Black. FAUST: Dataset and evaluation for 3D mesh registration. In Proceedings IEEE Conf. on Com" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.591, + 0.469, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "puter Vision and Pattern Recognition (CVPR), Piscataway, NJ, USA, June 2014. IEEE. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[7] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Numerical geometry of non-rigid shapes. Springer Science & Business Media, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[8] Oliver Burghard, Alexander Dieckmann, and Reinhard Klein. Embedding shapes with Green's functions for global shape matching. Computers & Graphics, 68:1-10, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[9] D. Cao and F. Bernard. Unsupervised deep multi-shape matching. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[10] Bailing Deng, Yuxin Yao, Roberto M. Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[11] Nicolas Donati, Etienne Corman, and Maks Ovsjanikov. Deep orientation-aware functional maps: Tackling symmetry issues in shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-751, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[12] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In CVPR, 2020. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[13] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In CVPR, 2020. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.894, + 0.518 + ], + "angle": 0, + "content": "[14] Marvin Eisenberger, David Novotny, Gael Kerchenbaum, Patrick Labatut, Natalia Neverova, Daniel Cremers, and Andrea Vedaldi. Neuromorph: Unsupervised shape interpolation and correspondence in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7473-7483, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[15] Marvin Eisenberger, Aysim Token, Laura Leal-Taixe, and Daniel Cremers. Deep shells: Unsupervised shape correspondence with optimal transport. Advances in Neural Information Processing Systems, 33:10491-10502, 2020. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[16] Jiahui Huang et. al. Multiway non-rigid point cloud registration via learned functional map synchronization, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[17] Maolin Gao, Zorah Lahner, Johan Thunberg, Daniel Cremers, and Florian Bernard. Isometric multi-shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14183-14193, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[18] Dvir Ginzburg and D. Raviv. Cyclic functional mapping: Self-supervised correspondence between non-isometric deformable shapes. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[19] Oshri Halimi, Or Litany, Emanuele Rodol'a, Alex Bronstein, and Ron Kimmel. Unsupervised learning of dense shape correspondence. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[20] Jiahui Huang, Tolga Birdal, Zan Gojcic, Leonidas J. Guibas, and Shi-Min Hu. Multiway Non-rigid Point Cloud Registration via Learned Functional Map Synchronization. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[21] Qixing Huang, Fan Wang, and Leonidas Guibas. Functional map networks for analyzing and exploring large shape collections. ACM TOG, 33(4), 2014. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14505" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[22] Qi-Xing Huang and Leonidas Guibas. Consistent shape maps via semidefinite programming. In Computer Graphics Forum, volume 32, pages 177-186. Wiley Online Library, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[23] Ruqi Huang, Panos Achlioptas, Leonidas Guibas, and Maks Ovsjanikov. Limit shapes – a tool for understanding shape differences and variability in 3d model collections. Computer Graphics Forum, 38(5):187–202, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[24] Ruqi Huang, Jing Ren, Peter Wonka, and Maks Ovsjanikov. Consistent zoomout: Efficient spectral map synchronization. In Computer Graphics Forum, volume 39, pages 265-278. Wiley Online Library, 2020. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.288 + ], + "angle": 0, + "content": "[25] Daniel F Huber. Automatic three-dimensional modeling from reality. PhD thesis, Citeseer, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[26] Artiom Kovnatsky, Michael M Bronstein, Alexander M Bronstein, Klaus Glashoff, and Ron Kimmel. Coupled quasiharmonic bases. In Computer Graphics Forum, volume 32, pages 439-448, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.346, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[27] Spyridon Leonardos, Xiaowei Zhou, and Kostas Daniilidis. Distributed consistent data association via permutation synchronization. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 2645-2652. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.415, + 0.469, + 0.47 + ], + "angle": 0, + "content": "[28] Lei Li, Nicolas Donati, and Maks Ovsjanikov. Learning multi-resolution functional maps with spectral attention for robust shape matching. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.47, + 0.54 + ], + "angle": 0, + "content": "[29] Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12706-12716, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.542, + 0.469, + 0.61 + ], + "angle": 0, + "content": "[30] Or Litany, Tal Remez, Emanuele Rodola, Alexander M. Bronstein, and Michael M. Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5660-5668, 2017. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.611, + 0.469, + 0.652 + ], + "angle": 0, + "content": "[31] Rabinovich M., Poranne R., Panozzo D., and Sorkine-Hornung O. Scalable locally injective mappings. In ACM Transactions on Graphics, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.653, + 0.469, + 0.707 + ], + "angle": 0, + "content": "[32] Robin Magnet, Jing Ren, Olga Sorkine-Hornung, and Maks Ovsjanikov. Smooth non-rigid shape matching via effective dirichlet energy optimization. In International Conference on 3D Vision, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.709, + 0.469, + 0.779 + ], + "angle": 0, + "content": "[33] Simone Melzi, Riccardo Marin, Emanuele Rodolà, Umberto Castellani, Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Shrec 2019: Matching humans with different connectivity. In Eurographics Workshop on 3D Object Retrieval, volume 7, page 3, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.78, + 0.469, + 0.834 + ], + "angle": 0, + "content": "[34] Simone Melzi, Jing Ren, Emanuele Rodola, Peter Wonka, and Maks Ovsjanikov. Zoomout: Spectral upsampling for efficient shape correspondence. Proc. SIGGRAPH Asia, 2019, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.836, + 0.469, + 0.905 + ], + "angle": 0, + "content": "[35] Andy Nguyen, Mirela Ben-Chen, Katarzyna Welnicka, Yinyu Ye, and Leonidas Guibas. An optimization approach to improving collections of shape maps. In Computer Graphics Forum, volume 30, pages 1481-1491. Wiley Online Library, 2011. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.905 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[36] Dorian Nogneng and Maks Ovsjanikov. Informative descriptor preservation via commutativity for shape matching. Computer Graphics Forum, 36(2):259-267, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[37] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional Maps: A Flexible Representation of Maps Between Shapes. ACM Transactions on Graphics (TOG), 31(4):30, 2012. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[38] Maks Ovsjanikov, Etienne Corman, Michael Bronstein, Emanuele Rodola, Mirela Ben-Chen, Leonidas Guibas, Frederic Chazal, and Alex Bronstein. Computing and processing correspondences with functional maps. In ACM SIGGRAPH 2017 Courses, page 5. ACM, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[39] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Linand N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. Pytorch: An imperative style, high-performance deep learning library, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[40] Jing Ren, Simone Melzi, Peter Wonka, and Maks Ovsjanikov. Discrete optimization for shape matching, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[41] Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Continuous and orientation-preserving correspondences via functional maps. ACM Trans. Graph., 37(6):248:1-248:16, Dec. 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[42] Emanuele Rodolà, Luca Cosmo, Michael M Bronstein, Andrea Torsello, and Daniel Cremers. Partial functional correspondence. In Computer Graphics Forum, volume 36, pages 222–236. Wiley Online Library, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[43] Jean-Michel Roufosse, Abhishek Sharma, and Maks Ovsjanikov. Unsupervised deep learning for structured shape matching. ICCV, 2019. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[44] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM Transactions on Graphics, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[45] Meged Shoham, Amir Vaxman, and Mirela Ben-Chen. Hierarchical functional maps between subdivision surfaces. In Computer Graphics Forum, volume 38, pages 55-73. Wiley Online Library, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.631, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[46] Jian Sun, Maks Ovsjanikov, and Leonidas Guibas. A Concise and Provably Informative Multi-Scale Signature Based on Heat Diffusion. In Computer graphics forum, volume 28, pages 1383-1392, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[47] Federico Tombari, Samuele Salti, and Luigi Di Stefano. Unique signatures of histograms for local surface description. In International Conference on Computer Vision (ICCV), pages 356-369, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[48] Giovanni Trappolini, Luca Cosmo, Luca Moschella, Ricardo Marin, Simone Melzi, and Emanuele Rodola. Shape registration in the time of transformers. Advances in Neural Information Processing Systems, 34:5731-5744, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[49] Fan Wang, Qixing Huang, and Leonidas J. Guibas. Image co-segmentation via consistent functional maps. In Proceedings of the IEEE International Conference on Computer Vision, pages 849-856, 2013. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[50] Lanhui Wang and Amit Singer. Exact and stable recovery of rotations for robust synchronization. Information and Inference: A Journal of the IMA, 2(2):145-193, 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14506" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[51] Lei Yang, Wenxi Liu, Zhiming Cui, Nenglun Chen, and Wenping Wang. Mapping in a cycle: Sinkhorn regularized unsupervised learning for point cloud shapes. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[52] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In 2010 IEEE Computer Society Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition, pages 1426-1433. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[53] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.092, + 0.892, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14507" + } + ] +] \ No newline at end of file diff --git a/2023/Spatially and Spectrally Consistent Deep Functional Maps/38616010-b1fd-470b-8376-180cbf13b5f1_origin.pdf b/2023/Spatially and Spectrally Consistent Deep Functional Maps/38616010-b1fd-470b-8376-180cbf13b5f1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..58d2197bdf365613ce2ae57c8f47e6233da04375 --- /dev/null +++ b/2023/Spatially and Spectrally Consistent Deep Functional Maps/38616010-b1fd-470b-8376-180cbf13b5f1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:475a4dfabd2eb54243ded9d798df52b657d2cf957e0e7ed0b32e37bfcad216b4 +size 3841653 diff --git a/2023/Spatially and Spectrally Consistent Deep Functional Maps/full.md b/2023/Spatially and Spectrally Consistent Deep Functional Maps/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4bd7bc3947739744567e5ecc9552db24c1a0b989 --- /dev/null +++ b/2023/Spatially and Spectrally Consistent Deep Functional Maps/full.md @@ -0,0 +1,348 @@ +# Spatially and Spectrally Consistent Deep Functional Maps + +Mingze Sun1 Shiwei Mao1 Puhua Jiang1,2 Maks Ovsjanikov3 Ruqi Huang1* +1Tsinghua Shenzhen International Graduate School, China +2Peng Cheng Laboratory, China +3LIX, École polytechnique, IP Paris, France + +# Abstract + +Cycle consistency has long been exploited as a powerful prior for jointly optimizing maps within a collection of shapes. In this paper, we investigate its utility in the approaches of Deep Functional Maps, which are considered state-of-the-art in non-rigid shape matching. We first justify that under certain conditions, the learned maps, when represented in the spectral domain, are already cycle consistent. Furthermore, we identify the discrepancy that spectrally consistent maps are not necessarily spatially, or point-wise, consistent. In light of this, we present a novel design of unsupervised Deep Functional Maps, which effectively enforces the harmony of learned maps under the spectral and the point-wise representation. By taking advantage of cycle consistency, our framework produces state-of-the-art results in mapping shapes even under significant distortions. Beyond that, by independently estimating maps in both spectral and spatial domains, our method naturally alleviates over-fitting in network training, yielding superior generalization performance and accuracy within an array of challenging tests for both near-isometric and non-isometric datasets. Codes are available at https://github.com/rqhuang88/Spatially-and-Spectrally-Consistent-Deep-Functional-Maps. + +# 1. Introduction + +Non-rigid shape matching is a fundamental task in shape analysis, playing a pivotal role in a wide spectrum of applications including texture transfer [1], volume parameterization [31], statistical shape analysis [2, 6], to name a few. In order to establish correspondences between shapes undergoing non-rigid deformations, early approaches [46, 4, 47] focus on designing hand-crafted features by exploiting geometric invariance. Nowadays, data-driven techniques have been widely adopted to learn features in a more flexible way, leading to significant improvements regarding accu + +![](images/cb7ef1846ad012ddb4f92b9a3fb2f9b9ff9b2cc6e8a344dda77a6cd6675d878f.jpg) +Figure 1. We train unsupervised non-rigid shape matching pipelines on the challenging non-isometric dataset DT4D-H, which involves 8 categories of humanoid shapes undergoing significant distortions. Correspondences are visualized by texture transfer. Our method (bottom row) outperforms the state-of-the-art, AttentiveFMaps [28] (middle row) and UDMSM [9] (top row), by a large margin. + +racy, efficiency, and robustness [10]. + +A noticeable trend among the learning-based shape matching approaches is based on the formalism of Deep Functional Maps (DFM), pioneered by the FMNet [30]. Functional maps [37], as a spectral map representation, allows to encode maps into compact matrices and to express desirable map priors (e.g., area-preservation, isometry, bijectivity) in simple algebraic forms. In particular, by learning features that optimize functional map priors, several unsupervised DFM frameworks [19, 43, 18, 13, 9, 28] have been proposed and, remarkably, achieve even comparable results with respect to the supervised counterparts. On the other hand, while the majority of DFM works focus on exploring local, pairwise-level map priors, there is relatively less attention (with the exception of [9]) paid to the problem of injecting global map priors into DFM pipelines. + +As a global map prior, cycle consistency has long been used as a powerful regularizer for jointly optimizing maps among shapes, both under point-wise [22] and functional [21, 24] map representation. In this paper, we exploit the utility of cycle consistency within DFM frameworks and propose a novel two-branch unsupervised DFM framework, which promotes cycle consistency in both spectral and spatial domains. We first re-examine the generic DFM frame + +work and outline a theoretical condition, based on the residuals of a linear solver used in functional map estimation, that guarantees spectrally cycle consistency in DFM over training shapes. Then we identify the discrepancy that spectrally consistent maps are not necessarily spatially, or pointwise, consistent. In light of this, we leverage our theoretical insight to construct a latent shape in DFM and employ this construction to formulate a novel two-branch design of unsupervised DFM. We align each shape's spectral embedding, represented as the eigenbasis of the Laplace-Beltrami operator, to a canonical space given by the constructed universal latent shape. We then compute point-wise maps in this canonical embedding domain, which promotes consistency across computed point-wise maps. In the end, the point-wise maps are converted to functional maps, which are expected to agree with the ones from the original spectral branch. Overall, our two-branch DFM takes advantage of the justified spectrally cycle consistency and further enhances it to spatially cycle consistent. + +We conduct extensive experiments on several non-rigid shape matching benchmarks, and demonstrate that our method achieves superior performance over existing methods, especially in the presence of significant distortions among shapes (see, e.g., Fig. 1). We also observe that our method, by enforcing the harmony of spatial and spectral map representations, reduces over-fitting during training, which leads to remarkable generalization performance within an array of challenging tests. Finally, our two-branch design can be easily plugged into any existing DFM framework, and we show evidence that it achieves significant performance gains upon both recent and early DFM approaches [43] with marginal computational burden. + +To summarize, our main contributions are as follows: (1) We perform theoretical analysis on spectrally cycle consistency of DFM frameworks; (2) We formulate a simple yet effective two-branch design of unsupervised DFM based on our theoretical justification, which introduces spatially cycle consistency. (3) We prove the effectiveness of our method through an array of challenging non-rigid shape matching tasks, demonstrating improvements upon existing methods in terms of accuracy, consistency, and generalization performance. + +# 2. Related Work + +Functional Maps Our method is built upon functional map representation, introduced in [37] and then significantly extended in follow-up works (see, e.g., [38]). The key idea is to encode shape correspondences as transformations between the respective spectral embeddings, which are represented by compact matrices by using reduced eigenbasis. The functional maps pipeline has been further improved in accuracy, efficiency, and robustness by many recent works including [26, 21, 8, 42, 36]. + +Deep Functional Maps In contrast to axiomatic approaches that use hand-crafted features [46, 4], the deep functional maps approach, pioneered by FMNet [30], aims to learn the optimal features from data. FMNet contributes several key designs of DFM: (1) it leverages Siamese network to conduct learning in a set of shape pairs; (2) it advocates refining the input descriptors with non-linear transformations. FMNet is then supervised by labeled maps to learn optimal features. + +Instead of learning from labeled maps, unsupervised approaches [19, 43] demonstrate that it is sufficient to learn from geometric map priors. More recently, with the development of robust mesh feature extractors [44], more frameworks [9, 28, 11, 3] are proposed to learn directly from geometry, yielding state-of-the-art performance. + +Cycle Consistency Cycle consistency has long been used as a strong prior for joint map optimization among a collection of shapes. Axiomatic approaches detect and eliminate inconsistent cycles using consistency constraints [25, 52, 35, 51, 18], as well as associate cycle consistency with low-rank properties of matrices encoding map networks [22, 50, 27]. Related to the latter, the matrix nature of the functional maps enables convenient access to map composition, which naturally bridges the functional map framework and consistent map refinement techniques [49, 21, 23, 45]. Some recent learning-based approach [20] also incorporate cycle consistency in the pipeline. It is worth noting, though, the above works all utilize cycle consistency as a prior for test-time optimization, which depends on test shape collection and initial maps. Contrastingly, our approach exploits cycle consistency during training to boost feature learning and poses no constraint on a test. + +From this viewpoint, UDMSM [9] is perhaps the most relevant work to ours, as both construct a latent shape during training to inject cycle consistency. The key difference between their work and ours, though, is how the latent shape is constructed. In [9], the authors propose to construct a universal shape in the learned feature space and establish pointwise maps from real shapes to the universal one by training a classifier. As shown in Sec.5, though this construction leads to great performance on matching near-isometric shapes, it suffers from large shape variability in mapping non-isometric ones. On the other hand, we leverage spectral information in estimating point-wise maps. As a result, our approach benefits from more direct usage of intrinsic geometric information encoded in the spectral embedding, yielding better generalization performance. + +Dual Map Representations Thanks to the inherent connection and efficient conversion between point-wise and functional maps, it has long been observed that jointly estimating both map representations can improve the mapping quality. For instance, in the original work [37], the authors have already proposed to apply an ICP-like technique on + +functional maps. More recent advances take advantage of the multi-scale properties of the eigenbasis of the Laplace-Beltrami operator. In the works [34, 24, 13, 15], conversions are done between spatial domain and a series of spectral domains spanned by eigenfunctions of increasing dimensions. This idea has also been incorporated into DFM. In AttentiveFMaps [28], the authors propose to fuse functional maps of different dimensions by converting them into the common spatial domain. This fusion technique in turn allows training an attention network to dynamically choose the optimal spectral resolution. + +The above methods enforce consistency across map representations by iterative projections, which is computationally heavy. By exploiting our theoretical insight on the spectral cycle consistency of DFM, we only introduce a marginal computational overhead to a standard DFM trained on a single spectral resolution. Our method is much lighter and simpler but also shows superior performance in accuracy and generalization. + +# 3. Cycle Consistency of Deep Functional Maps + +In this section, we first briefly review the general deep functional maps pipeline. Then we present a theoretical analysis of its cycle consistency. + +# 3.1. Deep Functional Maps + +We assume to be given a pair of deformable shapes $S_{1}$ and $S_{2}$ , which are discretized as triangular meshes of $n_{1}$ and $n_{2}$ vertices, respectively. The generic deep functional map pipeline introduced in [30] learns a map that spectrally aligns the shapes through the following four steps: + +1. Compute the leading $k$ eigenfunctions of the Laplace-Beltrami operator on each shape, which can be treated as a high-dimensional spectral embedding of the respective shape. The eigenfunctions are stored as matrices $\Phi_{i}\in \mathbb{R}^{n_{i}\times k},i = 1,2$ +2. Instantiate a feature extractor network, $\mathcal{F}_{\Theta}$ , where $\Theta$ denotes the set of learnable parameters. By feeding forward shapes through $\mathcal{F}_{\Theta}$ , descriptors are obtained and expected to be approximately preserved by the underlying map. In general, we denote $\mathbf{G}_i = \mathcal{F}_{\Theta}(S_i) \in \mathbb{R}^{n_i \times d}$ , $i = 1, 2$ , where $d$ is the predefined number of descriptors. They are then projected onto the eigenbasis above, resulting in a couple of coefficient matrices $\mathbf{A}_i = \boldsymbol{\Phi}_i^\dagger \mathbf{G}_i \in \mathbb{R}^{k \times d}$ , $i = 1, 2$ . +3. Estimate the optimal functional map, $\mathbf{C}^*$ , by solving the following linear system: + +$$ +\mathbf {C} ^ {*} = \arg \min _ {\mathbf {C}} E _ {\mathrm {d e s c}} (\mathbf {C}) + E _ {\mathrm {r e g}} (\mathbf {C}), \tag {1} +$$ + +where $E_{\mathrm{desc}}(\mathbf{C}) = \| \mathbf{CA}_1 - \mathbf{A}_2\|^2$ , and $E_{\mathrm{reg}}(\mathbf{C})$ is the regularization term promoting structural properties + +![](images/143874733320dcf517fd437577aa58b5339c02e0b3042929de4bb47ed9008857.jpg) +Figure 2. Overview of our two-branch DFM network. The part colored in black corresponds to a standard DFM, we introduce a novel branch, colored in red, that estimates maps from spatial perspective, our loss function is defined as Eqn.6. $\Phi_i,\Phi_i^\dagger$ indicate projection of the regarding features into the spatial and spectral domain, respectively. See more details in Section 4.2. + +of $\mathbf{C}$ , e.g., enforcing $\mathbf{C}$ to be commutative with the Laplace-Beltrami operators and to be orthogonal. In particular, we let $E_{\mathrm{total}}(\mathbf{C}) = E_{\mathrm{desc}}(\mathbf{C}) + E_{\mathrm{reg}}(\mathbf{C})$ +4. Convert the estimated functional map $\mathbf{C}$ to a pointwise map by conducting nearest neighbor search between the rows of $\Phi_1\mathbf{C}$ and that of $\Phi_{2}$ . + +The feature extractor is then trained on a set of training shape pairs, which typically enumerates all pairs from some given training set. The sub-optimization problem in Step (3) is generally decomposed into two parts to circumvent using an iterative solver. Namely, a closed-form solution is obtained either by a least-square estimation w.r.t $E_{\mathrm{desc}}(\mathbf{C})$ [30], or by solving a sequence of linear systems involving Laplacian commutativity as well [11]. Both correspond to the FMreg layer shown in Fig. 2. The rest of $E_{\mathrm{reg}}(\mathbf{C})$ is then set as the structural loss, corresponding to $\mathcal{L}(\mathbf{C})$ in Fig. 2. + +# 3.2. Cycle Consistency + +We start by giving a formal definition of cycle consistency. We assume to be given a collection of shapes $S = \{S_i\}_{i=1}^n$ and a set of maps $\mathcal{T} = \{T_{ij}\}_{i,j \in [1..n]}$ . We call $\mathcal{T}$ cycle consistent if for any shape $S_i$ , and any closed path $(i, i_1, i_2, \dots, i_k, i)$ , the map composition along this path $T_{ii} = T_{i_k i} \circ \dots \circ T_{i_1 i_2} \circ T_{ii_1}$ is an identity map on $S_i$ . We can similarly define cycle consistency on functional maps, the only difference is that map composition is given by matrix multiplication and we require the final functional map $C_{ii}$ to be an identity matrix. + +We suppose that deep functional maps are trained on $S$ with respect to all possible pairs. Then the global energy is given by $E_{\mathrm{total}}(\mathcal{C}) = \sum_{i,j} \| \mathbf{C}_{ij} \mathbf{A}_i - \mathbf{A}_j \|^2 + \sum_{i,j} E_{\mathrm{reg}}(\mathbf{C}_{ij})$ , where $\mathcal{C} = \{\mathbf{C}_{ij}\}_{i,j \in [1..n]}$ is the set of functional maps among training shapes. + +Proposition 1 If $E_{\text{total}}(\mathcal{C}) = 0$ , then for any shape $S_i$ , and any path $(i, i_1, i_2, \dots, i_k, i)$ , the map composition $\mathbf{C}_{ii}$ is cycle consistent within the functional space spanned by the columns of $\mathbf{A}_i$ , i.e., $\mathbf{C}_{ii}\mathbf{A}_i = \mathbf{A}_i$ . + +As a consequence of Prop. 1, when $\mathbf{A}_i\in \mathbb{R}^{k\times d}$ is of full row rank, then $\mathcal{C} = \{\mathbf{C}_{ij}\}_{i,j\in [1\dots n]}$ is cycle consistent. In practice, we generally set $d > k$ . Moreover, during network training, $\mathbf{C}_{ij}$ is computed via pseudo-inverse, which implicitly assumes the full-rankness of $\mathbf{A}_i,\mathbf{A}_j$ . We refer readers to the empirical validation in Section 4.2. + +We defer the proof of Prop. 1 to Supp. Material. In fact, a similar claim has been formulated and proven in [49] (see Sec.3.1 therein), but in the context of map refinement via promoting cycle consistency. + +Though being technically similar, the theoretical argument of [49] and that of ours have fundamentally different implications. More specifically, the former justifies a test-time optimization algorithm, which is used to promote cycle consistency of maps among a fixed test shape collection. While the latter suggests that spectrally cycle consistency has been ensured and further leveraged to enhance the universal feature extractor (independent of the test data) during training in any DFM framework following the generic pipeline presented in Sec. 3.1. + +# 4. Two-branch Deep Functional Maps + +It has long been recognized, both theoretically and empirically, that optimizing purely in the spectral domain is not sufficient. As a toy example, a trivial solution attaining global optima can be constructed as follows: Suppose that we have learned a feature extractor $\mathcal{F}_{\Theta}$ , which returns the respective eigenbasis transformed by a universal $\mathbf{A}_0$ . That is, $\mathbf{G}_i = \boldsymbol{\Phi}_i\mathbf{A}_0, \forall i$ , which implies $\mathbf{A}_i = \boldsymbol{\Phi}_i^\dagger\mathbf{G}_i = \mathbf{A}_0, \forall i$ . Then we have $\mathbf{C}_{ij} \equiv I_k$ , which exactly satisfies $E_{\mathrm{desc}}(\mathbf{C}_{ij}) = E_{\mathrm{reg}}(\mathbf{C}_{ij}) = 0, \forall i, j$ . However, it probably induces poor point-wise maps. + +In fact, in [37] the authors have proposed to use an ICP-like technique to encourage the estimated functional maps to be induced by some point-wise maps. In [34], the authors propose a spectral upsampling method for map refinement, which essentially converts maps back and forth between spectral and spatial domains. Moreover, the following lemma from [34] sheds light on the necessity of taking both spectral and spatial representations into consideration. + +Lemma 1 Given a pair of shapes $S_{1}, S_{2}$ each having nonrepeating Laplacian eigenvalues, which are the same. A point-wise map $T: S_{1} \to S_{2}$ is an isometry if and only if the corresponding functional map $\mathbf{C}$ in the complete Laplacian basis is both diagonal and orthonormal. + +The above lemma suggests that apart from promoting the structural properties of functional maps, it is also critical to enforce them to be associated with certain point-wise maps, or termed as properness of functional maps in [40]. + +Finally, we remark that some recent DFM advances also promote the properness of the resulting spectral maps. + +For instance, AttentiveFMaps [28] follows the spirit of ZoomOut [34] and explicitly performs a conversion between spectral and spatial map representations across different dimensions of eigenbasis; UDMSM [9] constructs explicitly a universal shape in the feature space, and enforce the spectral map estimation to be consistent with the spatial maps induced via the universal shape. + +# 4.1. Two-branch Map Estimation + +In this part, we leverage our observation made in Prop. 1 and propose a novel, simple yet effective design of unsupervised deep functional maps, which introduces a new branch that independently estimates maps from spatial perspective. + +Our key insight is that, once cycle consistency is valid and $\mathbf{A}_i$ is of full row rank, $\mathbf{A}_i$ can be seen as a functional map from a universal latent shape, $S_0$ , to $S_i$ . This perspective has been explored in several prior works [49, 23, 24], we provide the following details to be self-contained. The above assumption implies $\mathbf{C}_{ij} = \mathbf{A}_j\mathbf{A}_i^\dagger$ . Then $\mathbf{C}_{ij}$ can be interpreted as a functional map composition from $S_i$ to $S_0$ , followed by a map from $S_0$ to $S_j$ . On the other hand, one can align the spectral embeddings of $S_j$ to that of $S_i$ by simply transforming the former by $\mathbf{C}_{ij}$ . Indeed, we convert $\mathbf{C}_{ij}$ into the point-wise map by the nearest neighbor searching between the rows of $\Phi_j\mathbf{C}_{ij}$ and that of $\Phi_i$ . From this point of view, denoting the virtual spectral embedding of the latent shape by $\Phi_0$ , $\Phi_i\mathbf{A}_i$ can be then treated as the spectral embedding of $S_i$ aligned to that of $S_0$ . Therefore, given a pair of shapes $S_i, S_j$ , since we have aligned their eigenbasis to the canonical frame defined by the virtual spectral embedding $\Phi_0$ , we can align the spectral embedding of $S_i$ to $\Phi_0$ by computing $\Phi_i\mathbf{A}_i$ . Once all the spectral embeddings are aligned to the canonical embedding domain regarding $\Phi_0$ , we can compute the soft point-wise map between $S_i$ and $S_j$ by nearest neighbor searching between the rows of $\Phi_i\mathbf{A}_i$ and those of $\Phi_j\mathbf{A}_j$ . + +Based on the above derivation, given the learned features projected in the spectral domain, $\mathbf{A}_i,\mathbf{A}_j$ , and a pair of indices $p\in [1..n_i],q\in [1..n_j]$ , we can compute point-wise maps. Firstly we compute residual: + +$$ +\delta_ {q p} = \left\| \Phi_ {i} [ p ] \mathbf {A} _ {i} - \Phi_ {j} [ q ] \mathbf {A} _ {j} \right\| _ {2}, \tag {2} +$$ + +where $\Phi_i[p]$ denotes the $p$ -th row of $\Phi_i$ , and similarly we define $\Phi_j[q]$ . The soft point-wise map $\Pi \in \mathbb{R}^{n_j \times n_i}$ is then given by: + +$$ +\Pi (q, p) = \frac {\exp (- \alpha \delta_ {q p})}{\sum_ {p ^ {\prime}} \exp (- \alpha \delta_ {q p ^ {\prime}})}. \tag {3} +$$ + +Note that by construction, each row of $\Pi$ is non-negative and sums up to 1, forming a probability distribution. The parameter $\alpha$ controls the entropy of each distribution - the smaller/larger $\alpha$ is, the fuzzier/sharper the distribution is. + +![](images/95ad605773bb07df2bf392fdc7ed80286db371f9990e4c239a2561d6072d540c.jpg) +Figure 3. We train our two-branch DFM and a vanilla single-branch version on DT4D-H and monitor spectrally and spatially cycle consistency along the training. + +Instead of manually tuning the optimal $\alpha$ , we propose a learning scheme that dynamically controls $\alpha$ over training, which is inspired by curriculum learning [5]. We defer the respective details to Section 4.3. + +We convert the soft point-wise map to a functional map by + +$$ +\mathbf {C} _ {2} = \boldsymbol {\Phi} _ {j} ^ {\dagger} \Pi \boldsymbol {\Phi} _ {i}. \tag {4} +$$ + +In the end, we enforce $\mathbf{C}_2$ to be consistent with $\mathbf{C}_1$ , the intermediate output from the FMreg layer. + +To summarize, thanks to the spectrally cycle consistency we identify in Sec. 3.2, we are allowed to construct a spectral latent shape and induce spatial maps via it. By enforcing the spatial estimations to be consistent with the spectral ones, we obtain a spatially and spectrally consistent deep functional maps framework. + +# 4.2. Network Design + +As shown in Fig. 2, our two-branch network is built upon a standard DFM framework. In the following, we denote by $\mathbf{C}_1$ and $\mathbf{C}_2$ the estimated functional maps from the original FMreg layer and our novel branch, respectively. + +Specifically, we use DiffusionNet [44] as our feature extractor. And WKS [4] descriptors are fed into it as initialization of learned features. We borrow the FMreg layer from [12]. It takes both $E_{\mathrm{desc}}(\mathbf{C})$ and commutativity with the Laplace-Beltrami operator into consideration, where the latter is given as: + +$$ +L _ {\text {l a p}} = \left\| \mathbf {C} _ {1} \Lambda_ {1} - \Lambda_ {2} \mathbf {C} _ {1} \right\| ^ {2}, \tag {5} +$$ + +where $\Lambda_{1}$ and $\Lambda_{2}$ are diagonal matrices of the Laplace-Beltrami eigenvalues on the two shapes. + +The estimation of $\mathbf{C}_2$ has been described in detail in Section 4.1. In the end, we formulate the training loss as: + +$$ +\mathcal {L} \left(\mathbf {C} _ {1}, \mathbf {C} _ {2}\right) = \left\| \mathbf {C} _ {1} ^ {T} \mathbf {C} _ {1} - \mathbf {I} \right\| ^ {2} + \left\| \mathbf {C} _ {1} - \mathbf {C} _ {2} \right\| ^ {2}, \tag {6} +$$ + +where the first term promotes the orthogonality of $\mathbf{C}_1$ , while the second term promotes the consistency between the functional maps estimated from different branches. + +Finally, we remark that by combining the FMreg layer and $\mathcal{L}(\mathbf{C}_1,\mathbf{C}_2)$ , we have incorporated every factor in Lemma 1 into our design. + +Conceptual Validation In this part, we train a network on DT4D-H dataset (see Section 5.1 for details) with our two-branch network, and a single-branch variant without our spatial map estimation branch. We monitor and plot the following quantities along training: (1) Average spectral cycle consistency over sampled triplets, i.e., $\frac{1}{M}\sum_{(i,j,k)}\| \mathbf{C}_{ki}\mathbf{C}_{jk}\mathbf{C}_{ij} - I\| ^2 /\| I\| ^2$ ; (2) Average spatial cycle consistency over sampled triplets, i.e., the mean Euclidean deviation from composed maps $T_{ki}\circ T_{jk}\circ T_{ij}$ to identity map on $S_{i}$ . Here $n = 80$ is the number of training shapes, and $M = 1000$ is the number of sampled triplets. The behavior of the blue curves after 4500 iterations verifies our argument that spectrally cycle consistency does not imply spatially cycle consistency. On the other hand, by introducing our two-branch design, the discrepancy is well compensated and evidently better cycle consistencies in both spatial and spectral domains are achieved. + +# 4.3. Updating Scheme of $\alpha$ in Eqn. (3) + +The soft point-map conversion (Eqn. 3) has been applied in several prior works [30, 28], which all set $\alpha$ to be a manually selected constant. Ideally, we expect $\Pi$ in Eqn. (3) to be close to a permutation matrix, i.e., each row forms a binary vector. This seems to suggest a preference for a large $\alpha$ . Unfortunately, it would severely hinder network training, since the learned features and maps are of low quality in the early stage. On the other hand, a small $\alpha$ can alleviate such difficulty but falls short of fully pushing functional maps to be proper. As demonstrated in Section 5.5, neither small nor large $\alpha$ produces satisfying results. + +Based upon the above analysis, we propose a novel updating scheme, which is inspired by curriculum learning [5]. Namely, we initiate a small $\alpha$ at the beginning of network training and increase it by a constant step size for every fixed number of epochs. As shown in Table 1, 2, 3, our scheme does not rely on hyper-parameter tuning but also achieves state-of-the-art results. + +# 4.4. Implementation Details + +We implement our network with PyTorch [39]. We use four DiffusionNet blocks [44] as feature backbone and borrow the functional map block with Laplacian regularizer from [12]. The dimension of the Laplace-Beltrami eigenbasis is set to 50. WKS [4] descriptors are used as the input signal to our network. The dimensions of the input and the output descriptors are both set to 128. During training, the value of the learning rate is set to 2e-4 with ADAM optimizer. In all experiments, we train our method for 10,000 iterations with a batch size of 1. Following the learning + +strategy in Section 4.3, we initialize $\alpha$ to 1 and increase it by 5 per epoch. As indicated in Eqn. (6), We weigh equally the orthogonality loss with respect to $\mathbf{C}_1$ and the residual between $\mathbf{C}_1$ and $\mathbf{C}_2$ . More implementation details are provided in the Supp. Material. + +# 5. Experimental Results + +In this section, we conduct an extensive set of experiments of non-rigid shape matching on various datasets including humanoids and animals. We test on both near-isometric and non-isometric shape pairs. Our method is compared to a set of competitive baselines including axiomatic, supervised, weakly-supervised, and unsupervised learning methods. We emphasize that in this section, all the maps from the learning-based pipelines are directly inferred from the trained models, without any post-processing procedure. We evaluate the matching results in terms of a mean geodesic error on shapes normalized to unit area. Finally, our point-wise maps are all inferred by converting the output functional maps, as all the other DFM frameworks. + +# 5.1. Datasets + +FAUST_r: The remeshed version [41] of FAUST dataset[6] contains 100 human shapes. Following [43], it is split into 80/20 for training and testing. + +SCAPE_r: The remeshed version [41] of SCAPE dataset[2] contains 71 human shapes. Following [43], it is split into 51/20 for training and testing. + +SHREC19_r: The remeshed version of SHREC19 dataset [33] collects 44 human shapes from 11 independent datasets with distinctive poses and styles. We abandon shape 40 due to its partiality, we test on 407 pairs among the rest 43 shapes, which come with ground-truth. + +DT4D-H [32]: The remeshed subset of the large scale animation dataset DeformingThings4D [29]. In particular, DT4D-H includes 10 categories of humanoid shapes undergoing significant pose and style variances, forming a challenging benchmark. + +SMAL_r: The remeshed SMAL dataset [53] contains 49 animal shapes with 8 species. We follow the setting from [28], which splits 29 (5 species) and 20 (3 species) shapes for training and testing. + +TOSCA_r: The remeshed TOSCA dataset [7] contains multiple shape categories. We choose 4 animal categories, including cat, dog, horse, and wolf to verify the generalization performance of networks trained on SMAL_r. Note that we only infer the intra-category maps, due to the absence of ground-truth inter-category maps. + +We refer readers to Supp. Material for visualizations illustrating the variability of the above datasets. + +# 5.2. Near-isometric Shape Matching + +In this part, we perform comparisons with an array of non-rigid shape matching methods: (1) Axiomatic methods including ZoomOut [34], BCICP [41], IsoMuSh [17], Smooth Shells [13], CZO [24]; (2) Supervised learning methods including TransMatch [48], GeomFMaps [12], and supervised version of AttentiveFMaps [28]; (3) Unsupervised learning methods including NeuroMorph [14], SyNoRiM [16], Deep Shell [15], AttentiveFMaps [28], UDMSM [9], DUO-FM [11]. + +For all learning-based methods, we train models on FAUST_r and SCAPE_r respectively. Table 1 reports results on both standard tests and more challenging generalizations. We observe a trade-off between the two tasks, methods that perform the best in the former (e.g., supervised AttentiveFMaps and UDMSM) tend to overfit, and therefore suffer poor generalization (especially to SREHC19_r). Meanwhile, our default setting, denoted by Ours, achieves reasonable performance in the standard tests but also outperforms the external baselines in 3 out of 4 generalization tests. Especially, in generalizing to SHREC19_r, Ours outperforms the external baselines by a large margin, resulting in $41\%$ (3.8 vs. 6.4) and $46\%$ (4.5 vs. 8.4) error reduction upon the second best. + +We highlight that post-processing with cycle consistency generally depends on the initialized map quality and the size of the test set (e.g., $\geq 3$ shapes). In contrast, we leverage cycle consistency to improve the feature extractor during training. We also report the results from post-processing techniques based on cycle consistency [17, 24, 16] in Table 1. They are significantly outperformed by our method, which is inferred per-pair and without any post-processing. + +We further augment the dimension of functional maps in network training to 80 (same as UDMSM), which is beneficial to near-isometric matching [28]. It is evident that Ours (80 dim) achieves on-par performance with the regarding state-of-the-art methods in standard tests. On the other hand, due to the significant variability between SHREC19_r and the training sets (see Supp. Material), augmenting dimension leads to worse generalization than before (5.5 vs. 3.8, 5.8 vs. 4.5). Nevertheless, even in this case, our method outperforms the external baselines in all generalization tests by a notable margin. + +# 5.3. Non-isometric Shape Matching + +We also train our network on non-isometric datasets, SMAL_r and DT4D-H, and compare it with the state-of-the-art baselines including DeepShells [15], AttentiveFMaps [28], UDMSM [9] and DUO-FM [11]. + +SMAL_r: We follow the split and input descriptors from [28] (more details are provided in the Supp. Material). Table 2 reports results on SMAL_r, our method + +Table 1. Mean geodesic errors (×100) on FAUST_r, SCAPE_r, and SHREC19_r. The best and the second best are highlighted. + +
MethodTrain TestFAUST_rSCAPE_r
FAUST_rSCAPE_rSHREC19_rSCAPE_rFAUST_rSHREC19_r
ZM[34]6.1\\7.5\\
BCICP[41]6.4\\11.0\\
IsoMuSh[17]4.4\\5.6\\
Smooth Shell[13]2.5\\4.7\\
CZO[24]2.2\\2.5\\
TransMatch[48]2.733.621.018.318.638.8
GeomFMaps[12]supervised2.63.39.93.03.012.2
AttentiveFMaps[28]1.42.29.41.71.812.2
NeuroMorph[14]8.528.526.329.918.227.6
SyNoRiM[16]7.921.725.59.524.626.8
Deep Shell[15]1.75.427.42.52.723.4
AttentiveFMaps[28]unsupervised1.92.66.42.22.29.9
UDMSM[9]1.57.321.52.08.630.7
DUO-FM[11]2.54.26.42.72.88.4
Ours2.32.63.82.42.54.5
Ours (80 dim)1.72.65.52.22.05.8
+ +Table 2. Mean geodesic errors $(\times 100)$ on SMAL_r. The best and the second best are highlighted correspondingly. + +
MethodTrain TestSMAL_r
SMAL_rTOSCA_r
DeepShell[15]29.38.7
GeomFMaps[12]7.624.5
AttentiveFMaps[28]5.420.9
UDMSM[9]24.621.7
DUO-FM[11]32.815.3
Ours5.47.9
+ +achieves the best performance, which is on-par with AttentiveFMaps [28]. To evaluate generalization performance, we use the trained models to directly infer intra-category maps within TOSCA-r. It turns out that AttentiveFMaps and GeomFMaps both suffer from significant performance drops ( $\times 3.8$ and $\times 3.2$ larger geodesic error). It is also worth noting that DeepShells achieves the second-best generalization score in the relatively simpler task. However, it fails dramatically regarding the base task. In contrast, our method achieves the best balance between learning in difficult non-isometric pairs and generalizing to relatively easy near-isometric pairs. + +DT4D-H: We follow the train/test (198/95) split of [28], but ignore the categories mousey and ortiz in both train and test, due to the lack of inter-category map labels regarding them, resulting a split of 168/80. We emphasize that we conduct training and test in a category-agnostic manner, i.e., no class label is used, and the training pairs can consist of shapes from arbitrary two categories. This is significantly different from [28], in which training pairs are selected according to clustering information. Obviously, our setting is more practical, but also more challenging. For completeness, we report results under the setting of [28] in Supp. Material + +![](images/d41dc492e8de2981b3f434e3f45cafc90393bd9a3fa7c3dec84657091f86aaa5.jpg) +$T_{11}^{(8)} = T_{81}\circ T_{78}\circ \dots \circ T_{23}\circ T_{12}$ + +![](images/64ccf4fd1979f7cea35e717ec983ffe0a1bc56f37a4e7455c25af9d29ed047e1.jpg) +Ours + +![](images/2e3df59357d281208aec89702037854f0565f8c2c96010fc7d04560cabb6c515.jpg) +UDMSM + +![](images/4ea98edba3fcd026d767278c8b76c4d8fe67e02aa930fe28629e6dab1125d9c1.jpg) +DUO-FM +Deep Shell + +![](images/047059c4388edffeebf8516da0a316d882ca06a1329d31fa27fc53f8f487a78b.jpg) +BCICP +Figure 4. Qualitative evaluation of spatial cycle consistency of different methods. Even composed along a path of 8 highly deformed shapes, our resulting map remains close to identity, while all the baselines fail significantly. + +and our method outperforms the baselines in both intra- and inter-category evaluation by a notable margin. + +Table 3 reports results on DT4D-H, in which we preserve 80 shapes for test and train networks with 168 and 80 shapes, respectively. Note that we report mean geodesic errors over all possible test shape pairs, which may undergo significant distortions (see, e.g., Fig. 1). Our method obtains a $67.2\%$ (7.7 vs. 22.4) geodesic error reduction with respect to the second-best baseline. On top of that, we also test the generalization of the trained model on near-isometric benchmarks – our method also generalizes the best in generalization to FAUST_r and SCAPE_r. The same pattern is observed when the training set is reduced by more than half. Remarkably, our network trained on the reduced set still outperforms all the baselines trained on the full set. + +Overall, we attribute our performance on matching challenging non-isometric shapes (Table 3) and generalizing to unseen shapes (Table 1) to our effort to promote both spec + +Table 3. Mean geodesic errors $\left( {\times {100}}\right)$ on DT4D-H. The best and the second best are highlighted correspondingly. + +
MethodTrain TestDT4D-H (168)DT4D-H (80)
DT4D-HFAUST_rSCAPE_rDT4D-HFAUST_rSCAPE_r
DeepShell[15]27.04.96.529.34.77.0
AttentiveFMaps[28]25.73.46.428.92.76.3
UDMSM[9]46.843.347.949.742.540.0
DUO-FM[11]22.410.012.224.78.09.2
Ours7.73.16.19.02.66.2
+ +Table 4. Mean geodesic errors $\left( {\times {100}}\right)$ of SURFMNet and our variant trained on 4 datasets + +
MethodSURFMNetSURFMNet + Ours
FAUST_r6.03.5
SCAPE_r6.83.4
SMAL_r20.413.3
DT4D-H18.315.0
+ +trally and spatially cycle consistency. Especially, the isometry assumption is likely violated in the former case, thus cycle consistency, as a generic prior, plays an important role of regularizing maps. + +As an illustration, we present a qualitative evaluation on the point-wise cycle consistency in Fig. 4. We sample 8 shapes from the test set of DT4D-H (one from each category) and compose the maps along the path $(S_{1} \to S_{2} \to \dots \to S_{8} \to S_{1})$ with respect to different approaches. It is evident that due to the significant distortion undergoing among the shapes, all but our method fails to preserve cycle consistency in this demanding test, while our composing map approximates the identity map on $S_{1}$ . It also aligns nicely with the quantitative results reported in Table 3. + +# 5.4. Integration with SURFMNet [43] + +Our two-branch design can be easily incorporated into any existing DFM framework following the general design outlined in Sec. 3.1. To demonstrate this, we modify the SURFMNet [43], one of the earliest approaches of unsupervised DFM, by adding our new branch. Table 4 shows the matching accuracy on the four main benchmarks. It is evident that in every case, incorporating our design leads to significant error reduction ranging from $18\%$ to $50\%$ . Especially, in the near-isometric cases, we obtain $41.6\%$ and $50\%$ error reduction respectively. Note that the absolute scores, 0.035, 0.034, are reasonable even compared to the state-of-the-art results reported in Table 1. + +# 5.5. Ablation Study + +In this section, we present a set of ablation studies consisting of two parts. The first part verifies the rationality of our method, and the second part demonstrates the robustness of our method. We conduct all experiments on SMAL_r dataset [53]. + +First of all, instead of using the updating scheme in Sec- + +tion 4.3, we test the performance of our pipeline using two fixed values of $\alpha$ in Eqn. (3): $\alpha = 1$ and $\alpha = 50$ . Compared to our proposed model, the two variants yield a noticeable performance drop. Especially in the case $\alpha = 50$ , the network fails to deliver reasonable matching results. We believe it is because a large $\alpha$ amplifies the noise of maps learned at the early training stage. + +Then we justify our two-branch network design. Removing spatial branch amounts to training a standard single-branch DFM. To remove the spectral branch, we remove the FMreg layer and instead use our new branch to compute point-wise maps and convert them to functional maps. In the end, we modify the training loss so that it covers descriptor preservation, commutativity with the Laplace-Beltrami operator, and orthogonality (the latter two compensate the removed FMreg layer). The accuracy drop reported in the third and fourth row of Table 5 clearly suggests the necessity of our two-branch design. + +In our experiments, we always use the full resolution meshes ( $\sim 5k$ vertices) and compute in Eqn. (2) with all of the 128 descriptors. We anticipate that efficiency can become an issue when the input mesh resolution is high, and/or we would like to increase the size of learned descriptors. Therefore, we test the robustness of our pipeline with respect to down-sampling, which is commonly used in functional maps-based frameworks [34, 28]: 1) We down-sample 3000 vertices on each shape via furthest point sampling; 2) In order to down-sample the feature dimension, we operate as the following during training: given a $\mathbf{A}_1$ , $\mathbf{A}_2$ , we perform SVD on $\mathbf{A}_1$ , i.e., $\mathbf{A}_1 = U_1\Sigma_1V_1^T$ , then we set $\hat{\mathbf{A}}_1 = \mathbf{A}_1\hat{V}_1$ , and set $\hat{\mathbf{A}}_2 = \mathbf{A}_2\hat{V}_1$ , where $\hat{V}_1$ is the first $m$ columns of $V_{1}$ . We set $m = 30$ by replacing $\mathbf{A}_i$ with $\hat{\mathbf{A}}_i$ in Eqn. (2). The results in the bottom two rows show that the above operation has a relatively minor effect on the performance, proving the robustness of our method. + +# 6. Conclusion + +To conclude, we provide a theoretical justification for spectral cycle consistency in DFM. To compensate for the discrepancy of purely spectrally cycle consistency, we formulate a spectral latent shape that allows the alignment of the spectral embeddings of each shape to a canonical embedding domain. Based on this construction, we introduce a two-branch architecture for estimating maps in both spec + +Table 5. Mean geodesic errors $(\times 100)$ on different ablated settings, the models are all train on SMAL_r. + +
Ours5.4
α = 16.6
α = 5035.2
remove spatial branch33.4
remove spectral branch14.3
Vertex downsampling5.7
Feature downsampling5.9
+ +tral and spatial domains within DFM. The resulting network is simple, computationally efficient, and compatible with most existing DFM frameworks. We demonstrate the effectiveness of our framework through a comprehensive set of experiments, showing significant improvements upon state-of-the-art approaches in terms of accuracy, consistency, and generalization performance. + +It is worth noting, though, our method is developed upon clean and complete meshes. It would be interesting for future work to investigate the utility of our approach in more general tasks, involving partial meshes, noisy point clouds, and other representations. + +Acknowledgement This work was supported in part by the National Natural Science Foundation of China under contract No. 62171256 and Shenzhen Key Laboratory of next-generation interactive media innovative technology (No. ZDSYS20210623092001004), and in part by the ERC Starting Grant No. 758800 (EXPROTEA) and the ANR AI Chair AIGRETTE. + +# References + +[1] Noam Aigerman, Roi Poranne, and Yaron Lipman. Seamless Surface Mappings. ACM Transactions on Graphics (TOG), 34(4):72, 2015. 1 +[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape Completion and Animation of People. In ACM Transactions on Graphics (TOG), volume 24, pages 408-416. ACM, 2005. 1, 6 +[3] Souhaib Attaiki, Gautam Pai, and Maks Ovsjanikov. Dpfm: Deep partial functional maps. In 2021 International Conference on 3D Vision (3DV), pages 175-185. IEEE, 2021. 2 +[4] Mathieu Aubry, Ulrich Schlickewei, and Daniel Cremers. The Wave Kernel Signature: A Quantum Mechanical Approach to Shape Analysis. In Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on, pages 1626-1633. IEEE, 2011. 1, 2, 5 +[5] Yoshua Bengio, Jerome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In International Conference on Machine Learning, 2009. 5 +[6] Federica Bogo, Javier Romero, Matthew Loper, and Michael J. Black. FAUST: Dataset and evaluation for 3D mesh registration. In Proceedings IEEE Conf. on Com + +puter Vision and Pattern Recognition (CVPR), Piscataway, NJ, USA, June 2014. IEEE. 1, 6 +[7] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Numerical geometry of non-rigid shapes. Springer Science & Business Media, 2008. 6 +[8] Oliver Burghard, Alexander Dieckmann, and Reinhard Klein. Embedding shapes with Green's functions for global shape matching. Computers & Graphics, 68:1-10, 2017. 2 +[9] D. Cao and F. Bernard. Unsupervised deep multi-shape matching. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 4, 6, 7, 8 +[10] Bailing Deng, Yuxin Yao, Roberto M. Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. 2022. 1 +[11] Nicolas Donati, Etienne Corman, and Maks Ovsjanikov. Deep orientation-aware functional maps: Tackling symmetry issues in shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-751, 2022. 2, 3, 6, 7, 8 +[12] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In CVPR, 2020. 5, 6, 7 +[13] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In CVPR, 2020. 1, 3, 6, 7 +[14] Marvin Eisenberger, David Novotny, Gael Kerchenbaum, Patrick Labatut, Natalia Neverova, Daniel Cremers, and Andrea Vedaldi. Neuromorph: Unsupervised shape interpolation and correspondence in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7473-7483, 2021. 6, 7 +[15] Marvin Eisenberger, Aysim Token, Laura Leal-Taixe, and Daniel Cremers. Deep shells: Unsupervised shape correspondence with optimal transport. Advances in Neural Information Processing Systems, 33:10491-10502, 2020. 3, 6, 7, 8 +[16] Jiahui Huang et. al. Multiway non-rigid point cloud registration via learned functional map synchronization, 2022. 6, 7 +[17] Maolin Gao, Zorah Lahner, Johan Thunberg, Daniel Cremers, and Florian Bernard. Isometric multi-shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14183-14193, 2021. 6, 7 +[18] Dvir Ginzburg and D. Raviv. Cyclic functional mapping: Self-supervised correspondence between non-isometric deformable shapes. In ECCV, 2020. 1, 2 +[19] Oshri Halimi, Or Litany, Emanuele Rodol'a, Alex Bronstein, and Ron Kimmel. Unsupervised learning of dense shape correspondence. In CVPR, 2019. 1, 2 +[20] Jiahui Huang, Tolga Birdal, Zan Gojcic, Leonidas J. Guibas, and Shi-Min Hu. Multiway Non-rigid Point Cloud Registration via Learned Functional Map Synchronization. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2 +[21] Qixing Huang, Fan Wang, and Leonidas Guibas. Functional map networks for analyzing and exploring large shape collections. ACM TOG, 33(4), 2014. 1, 2 + +[22] Qi-Xing Huang and Leonidas Guibas. Consistent shape maps via semidefinite programming. In Computer Graphics Forum, volume 32, pages 177-186. Wiley Online Library, 2013. 1, 2 +[23] Ruqi Huang, Panos Achlioptas, Leonidas Guibas, and Maks Ovsjanikov. Limit shapes – a tool for understanding shape differences and variability in 3d model collections. Computer Graphics Forum, 38(5):187–202, 2019. 2, 4 +[24] Ruqi Huang, Jing Ren, Peter Wonka, and Maks Ovsjanikov. Consistent zoomout: Efficient spectral map synchronization. In Computer Graphics Forum, volume 39, pages 265-278. Wiley Online Library, 2020. 1, 3, 4, 6, 7 +[25] Daniel F Huber. Automatic three-dimensional modeling from reality. PhD thesis, Citeseer, 2002. 2 +[26] Artiom Kovnatsky, Michael M Bronstein, Alexander M Bronstein, Klaus Glashoff, and Ron Kimmel. Coupled quasiharmonic bases. In Computer Graphics Forum, volume 32, pages 439-448, 2013. 2 +[27] Spyridon Leonardos, Xiaowei Zhou, and Kostas Daniilidis. Distributed consistent data association via permutation synchronization. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 2645-2652. IEEE, 2017. 2 +[28] Lei Li, Nicolas Donati, and Maks Ovsjanikov. Learning multi-resolution functional maps with spectral attention for robust shape matching. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[29] Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12706-12716, 2021. 6 +[30] Or Litany, Tal Remez, Emanuele Rodola, Alexander M. Bronstein, and Michael M. Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5660-5668, 2017. 1, 2, 3, 5 +[31] Rabinovich M., Poranne R., Panozzo D., and Sorkine-Hornung O. Scalable locally injective mappings. In ACM Transactions on Graphics, 2017. 1 +[32] Robin Magnet, Jing Ren, Olga Sorkine-Hornung, and Maks Ovsjanikov. Smooth non-rigid shape matching via effective dirichlet energy optimization. In International Conference on 3D Vision, 2022. 6 +[33] Simone Melzi, Riccardo Marin, Emanuele Rodolà, Umberto Castellani, Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Shrec 2019: Matching humans with different connectivity. In Eurographics Workshop on 3D Object Retrieval, volume 7, page 3, 2019. 6 +[34] Simone Melzi, Jing Ren, Emanuele Rodola, Peter Wonka, and Maks Ovsjanikov. Zoomout: Spectral upsampling for efficient shape correspondence. Proc. SIGGRAPH Asia, 2019, 3, 4, 6, 7, 8 +[35] Andy Nguyen, Mirela Ben-Chen, Katarzyna Welnicka, Yinyu Ye, and Leonidas Guibas. An optimization approach to improving collections of shape maps. In Computer Graphics Forum, volume 30, pages 1481-1491. Wiley Online Library, 2011. 2 + +[36] Dorian Nogneng and Maks Ovsjanikov. Informative descriptor preservation via commutativity for shape matching. Computer Graphics Forum, 36(2):259-267, 2017. 2 +[37] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional Maps: A Flexible Representation of Maps Between Shapes. ACM Transactions on Graphics (TOG), 31(4):30, 2012. 1, 2, 4 +[38] Maks Ovsjanikov, Etienne Corman, Michael Bronstein, Emanuele Rodola, Mirela Ben-Chen, Leonidas Guibas, Frederic Chazal, and Alex Bronstein. Computing and processing correspondences with functional maps. In ACM SIGGRAPH 2017 Courses, page 5. ACM, 2017. 2 +[39] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Linand N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. Pytorch: An imperative style, high-performance deep learning library, 2019. 5 +[40] Jing Ren, Simone Melzi, Peter Wonka, and Maks Ovsjanikov. Discrete optimization for shape matching, 2021. 4 +[41] Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Continuous and orientation-preserving correspondences via functional maps. ACM Trans. Graph., 37(6):248:1-248:16, Dec. 2018. 6, 7 +[42] Emanuele Rodolà, Luca Cosmo, Michael M Bronstein, Andrea Torsello, and Daniel Cremers. Partial functional correspondence. In Computer Graphics Forum, volume 36, pages 222–236. Wiley Online Library, 2017. 2 +[43] Jean-Michel Roufosse, Abhishek Sharma, and Maks Ovsjanikov. Unsupervised deep learning for structured shape matching. ICCV, 2019. 1, 2, 6, 8 +[44] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM Transactions on Graphics, 2022. 2, 5 +[45] Meged Shoham, Amir Vaxman, and Mirela Ben-Chen. Hierarchical functional maps between subdivision surfaces. In Computer Graphics Forum, volume 38, pages 55-73. Wiley Online Library, 2019. 2 +[46] Jian Sun, Maks Ovsjanikov, and Leonidas Guibas. A Concise and Provably Informative Multi-Scale Signature Based on Heat Diffusion. In Computer graphics forum, volume 28, pages 1383-1392, 2009. 1, 2 +[47] Federico Tombari, Samuele Salti, and Luigi Di Stefano. Unique signatures of histograms for local surface description. In International Conference on Computer Vision (ICCV), pages 356-369, 2010. 1 +[48] Giovanni Trappolini, Luca Cosmo, Luca Moschella, Ricardo Marin, Simone Melzi, and Emanuele Rodola. Shape registration in the time of transformers. Advances in Neural Information Processing Systems, 34:5731-5744, 2021. 6, 7 +[49] Fan Wang, Qixing Huang, and Leonidas J. Guibas. Image co-segmentation via consistent functional maps. In Proceedings of the IEEE International Conference on Computer Vision, pages 849-856, 2013. 2, 4 +[50] Lanhui Wang and Amit Singer. Exact and stable recovery of rotations for robust synchronization. Information and Inference: A Journal of the IMA, 2(2):145-193, 2013. 2 + +[51] Lei Yang, Wenxi Liu, Zhiming Cui, Nenglun Chen, and Wenping Wang. Mapping in a cycle: Sinkhorn regularized unsupervised learning for point cloud shapes. In ECCV, 2020. 2 +[52] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In 2010 IEEE Computer Society Conference on Computer + +Vision and Pattern Recognition, pages 1426-1433. IEEE, 2010. 2 +[53] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 6, 8 \ No newline at end of file diff --git a/2023/Spatially and Spectrally Consistent Deep Functional Maps/images.zip b/2023/Spatially and Spectrally Consistent Deep Functional Maps/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d94be4259ccc6fead5dc3e9f69e0661c95ad0086 --- /dev/null +++ b/2023/Spatially and Spectrally Consistent Deep Functional Maps/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64719cfc6f3683f30819ac33c074ec75c32d647f186364a45d2437a2d4806285 +size 298647 diff --git a/2023/Spatially and Spectrally Consistent Deep Functional Maps/layout.json b/2023/Spatially and Spectrally Consistent Deep Functional Maps/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..667065c878bb13ff64395398440a3e42d41a72c3 --- /dev/null +++ b/2023/Spatially and Spectrally Consistent Deep Functional Maps/layout.json @@ -0,0 +1,10615 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 477, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 477, + 121 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 477, + 121 + ], + "type": "text", + "content": "Spatially and Spectrally Consistent Deep Functional Maps" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 141, + 505, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 141, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 91, + 141, + 505, + 201 + ], + "type": "text", + "content": "Mingze Sun1 Shiwei Mao1 Puhua Jiang1,2 Maks Ovsjanikov3 Ruqi Huang1* \n1Tsinghua Shenzhen International Graduate School, China \n2Peng Cheng Laboratory, China \n3LIX, École polytechnique, IP Paris, France" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 228, + 192, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 228, + 192, + 240 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 192, + 240 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 253, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 290, + 529 + ], + "type": "text", + "content": "Cycle consistency has long been exploited as a powerful prior for jointly optimizing maps within a collection of shapes. In this paper, we investigate its utility in the approaches of Deep Functional Maps, which are considered state-of-the-art in non-rigid shape matching. We first justify that under certain conditions, the learned maps, when represented in the spectral domain, are already cycle consistent. Furthermore, we identify the discrepancy that spectrally consistent maps are not necessarily spatially, or point-wise, consistent. In light of this, we present a novel design of unsupervised Deep Functional Maps, which effectively enforces the harmony of learned maps under the spectral and the point-wise representation. By taking advantage of cycle consistency, our framework produces state-of-the-art results in mapping shapes even under significant distortions. Beyond that, by independently estimating maps in both spectral and spatial domains, our method naturally alleviates over-fitting in network training, yielding superior generalization performance and accuracy within an array of challenging tests for both near-isometric and non-isometric datasets. Codes are available at https://github.com/rqhuang88/Spatially-and-Spectrally-Consistent-Deep-Functional-Maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 573, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 694 + ], + "type": "text", + "content": "Non-rigid shape matching is a fundamental task in shape analysis, playing a pivotal role in a wide spectrum of applications including texture transfer [1], volume parameterization [31], statistical shape analysis [2, 6], to name a few. In order to establish correspondences between shapes undergoing non-rigid deformations, early approaches [46, 4, 47] focus on designing hand-crafted features by exploiting geometric invariance. Nowadays, data-driven techniques have been widely adopted to learn features in a more flexible way, leading to significant improvements regarding accu" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 227, + 566, + 337 + ], + "blocks": [ + { + "bbox": [ + 316, + 227, + 566, + 337 + ], + "lines": [ + { + "bbox": [ + 316, + 227, + 566, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 566, + 337 + ], + "type": "image", + "image_path": "cb7ef1846ad012ddb4f92b9a3fb2f9b9ff9b2cc6e8a344dda77a6cd6675d878f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 338, + 547, + 417 + ], + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 417 + ], + "type": "text", + "content": "Figure 1. We train unsupervised non-rigid shape matching pipelines on the challenging non-isometric dataset DT4D-H, which involves 8 categories of humanoid shapes undergoing significant distortions. Correspondences are visualized by texture transfer. Our method (bottom row) outperforms the state-of-the-art, AttentiveFMaps [28] (middle row) and UDMSM [9] (top row), by a large margin." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 423, + 455, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 455, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 455, + 436 + ], + "type": "text", + "content": "racy, efficiency, and robustness [10]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 617 + ], + "type": "text", + "content": "A noticeable trend among the learning-based shape matching approaches is based on the formalism of Deep Functional Maps (DFM), pioneered by the FMNet [30]. Functional maps [37], as a spectral map representation, allows to encode maps into compact matrices and to express desirable map priors (e.g., area-preservation, isometry, bijectivity) in simple algebraic forms. In particular, by learning features that optimize functional map priors, several unsupervised DFM frameworks [19, 43, 18, 13, 9, 28] have been proposed and, remarkably, achieve even comparable results with respect to the supervised counterparts. On the other hand, while the majority of DFM works focus on exploring local, pairwise-level map priors, there is relatively less attention (with the exception of [9]) paid to the problem of injecting global map priors into DFM pipelines." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "As a global map prior, cycle consistency has long been used as a powerful regularizer for jointly optimizing maps among shapes, both under point-wise [22] and functional [21, 24] map representation. In this paper, we exploit the utility of cycle consistency within DFM frameworks and propose a novel two-branch unsupervised DFM framework, which promotes cycle consistency in both spectral and spatial domains. We first re-examine the generic DFM frame" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 238, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 238, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 238, + 713 + ], + "type": "text", + "content": "*Corresponding author: ruqihuang@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14497" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": "work and outline a theoretical condition, based on the residuals of a linear solver used in functional map estimation, that guarantees spectrally cycle consistency in DFM over training shapes. Then we identify the discrepancy that spectrally consistent maps are not necessarily spatially, or pointwise, consistent. In light of this, we leverage our theoretical insight to construct a latent shape in DFM and employ this construction to formulate a novel two-branch design of unsupervised DFM. We align each shape's spectral embedding, represented as the eigenbasis of the Laplace-Beltrami operator, to a canonical space given by the constructed universal latent shape. We then compute point-wise maps in this canonical embedding domain, which promotes consistency across computed point-wise maps. In the end, the point-wise maps are converted to functional maps, which are expected to agree with the ones from the original spectral branch. Overall, our two-branch DFM takes advantage of the justified spectrally cycle consistency and further enhances it to spatially cycle consistent." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 300, + 289, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 289, + 455 + ], + "type": "text", + "content": "We conduct extensive experiments on several non-rigid shape matching benchmarks, and demonstrate that our method achieves superior performance over existing methods, especially in the presence of significant distortions among shapes (see, e.g., Fig. 1). We also observe that our method, by enforcing the harmony of spatial and spectral map representations, reduces over-fitting during training, which leads to remarkable generalization performance within an array of challenging tests. Finally, our two-branch design can be easily plugged into any existing DFM framework, and we show evidence that it achieves significant performance gains upon both recent and early DFM approaches [43] with marginal computational burden." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 455, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 289, + 575 + ], + "type": "text", + "content": "To summarize, our main contributions are as follows: (1) We perform theoretical analysis on spectrally cycle consistency of DFM frameworks; (2) We formulate a simple yet effective two-branch design of unsupervised DFM based on our theoretical justification, which introduces spatially cycle consistency. (3) We prove the effectiveness of our method through an array of challenging non-rigid shape matching tasks, demonstrating improvements upon existing methods in terms of accuracy, consistency, and generalization performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 586, + 134, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 134, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 134, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 289, + 715 + ], + "type": "text", + "content": "Functional Maps Our method is built upon functional map representation, introduced in [37] and then significantly extended in follow-up works (see, e.g., [38]). The key idea is to encode shape correspondences as transformations between the respective spectral embeddings, which are represented by compact matrices by using reduced eigenbasis. The functional maps pipeline has been further improved in accuracy, efficiency, and robustness by many recent works including [26, 21, 8, 42, 36]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "Deep Functional Maps In contrast to axiomatic approaches that use hand-crafted features [46, 4], the deep functional maps approach, pioneered by FMNet [30], aims to learn the optimal features from data. FMNet contributes several key designs of DFM: (1) it leverages Siamese network to conduct learning in a set of shape pairs; (2) it advocates refining the input descriptors with non-linear transformations. FMNet is then supervised by labeled maps to learn optimal features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 255 + ], + "type": "text", + "content": "Instead of learning from labeled maps, unsupervised approaches [19, 43] demonstrate that it is sufficient to learn from geometric map priors. More recently, with the development of robust mesh feature extractors [44], more frameworks [9, 28, 11, 3] are proposed to learn directly from geometry, yielding state-of-the-art performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 255, + 548, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 548, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 548, + 459 + ], + "type": "text", + "content": "Cycle Consistency Cycle consistency has long been used as a strong prior for joint map optimization among a collection of shapes. Axiomatic approaches detect and eliminate inconsistent cycles using consistency constraints [25, 52, 35, 51, 18], as well as associate cycle consistency with low-rank properties of matrices encoding map networks [22, 50, 27]. Related to the latter, the matrix nature of the functional maps enables convenient access to map composition, which naturally bridges the functional map framework and consistent map refinement techniques [49, 21, 23, 45]. Some recent learning-based approach [20] also incorporate cycle consistency in the pipeline. It is worth noting, though, the above works all utilize cycle consistency as a prior for test-time optimization, which depends on test shape collection and initial maps. Contrastingly, our approach exploits cycle consistency during training to boost feature learning and poses no constraint on a test." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 460, + 548, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 548, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 548, + 640 + ], + "type": "text", + "content": "From this viewpoint, UDMSM [9] is perhaps the most relevant work to ours, as both construct a latent shape during training to inject cycle consistency. The key difference between their work and ours, though, is how the latent shape is constructed. In [9], the authors propose to construct a universal shape in the learned feature space and establish pointwise maps from real shapes to the universal one by training a classifier. As shown in Sec.5, though this construction leads to great performance on matching near-isometric shapes, it suffers from large shape variability in mapping non-isometric ones. On the other hand, we leverage spectral information in estimating point-wise maps. As a result, our approach benefits from more direct usage of intrinsic geometric information encoded in the spectral embedding, yielding better generalization performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Dual Map Representations Thanks to the inherent connection and efficient conversion between point-wise and functional maps, it has long been observed that jointly estimating both map representations can improve the mapping quality. For instance, in the original work [37], the authors have already proposed to apply an ICP-like technique on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14498" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "functional maps. More recent advances take advantage of the multi-scale properties of the eigenbasis of the Laplace-Beltrami operator. In the works [34, 24, 13, 15], conversions are done between spatial domain and a series of spectral domains spanned by eigenfunctions of increasing dimensions. This idea has also been incorporated into DFM. In AttentiveFMaps [28], the authors propose to fuse functional maps of different dimensions by converting them into the common spatial domain. This fusion technique in turn allows training an attention network to dynamically choose the optimal spectral resolution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 301 + ], + "type": "text", + "content": "The above methods enforce consistency across map representations by iterative projections, which is computationally heavy. By exploiting our theoretical insight on the spectral cycle consistency of DFM, we only introduce a marginal computational overhead to a standard DFM trained on a single spectral resolution. Our method is much lighter and simpler but also shows superior performance in accuracy and generalization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 311, + 287, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 287, + 326 + ], + "type": "text", + "content": "3. Cycle Consistency of Deep Functional Maps" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "content": "In this section, we first briefly review the general deep functional maps pipeline. Then we present a theoretical analysis of its cycle consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 374, + 176, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 176, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 176, + 388 + ], + "type": "text", + "content": "3.1. Deep Functional Maps" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": "We assume to be given a pair of deformable shapes " + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": ", which are discretized as triangular meshes of " + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "n_{1}" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "n_{2}" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": " vertices, respectively. The generic deep functional map pipeline introduced in [30] learns a map that spectrally aligns the shapes through the following four steps:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 463, + 287, + 653 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 463, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 287, + 521 + ], + "type": "text", + "content": "1. Compute the leading " + }, + { + "bbox": [ + 55, + 463, + 287, + 521 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 463, + 287, + 521 + ], + "type": "text", + "content": " eigenfunctions of the Laplace-Beltrami operator on each shape, which can be treated as a high-dimensional spectral embedding of the respective shape. The eigenfunctions are stored as matrices " + }, + { + "bbox": [ + 55, + 463, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\Phi_{i}\\in \\mathbb{R}^{n_{i}\\times k},i = 1,2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "text", + "content": "2. Instantiate a feature extractor network, " + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\Theta}" + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "text", + "content": " denotes the set of learnable parameters. By feeding forward shapes through " + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\Theta}" + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "text", + "content": ", descriptors are obtained and expected to be approximately preserved by the underlying map. In general, we denote " + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_i = \\mathcal{F}_{\\Theta}(S_i) \\in \\mathbb{R}^{n_i \\times d}" + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "inline_equation", + "content": "i = 1, 2" + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "text", + "content": " is the predefined number of descriptors. They are then projected onto the eigenbasis above, resulting in a couple of coefficient matrices " + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i = \\boldsymbol{\\Phi}_i^\\dagger \\mathbf{G}_i \\in \\mathbb{R}^{k \\times d}" + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "inline_equation", + "content": "i = 1, 2" + }, + { + "bbox": [ + 54, + 522, + 287, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 629, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 287, + 653 + ], + "type": "text", + "content": "3. Estimate the optimal functional map, " + }, + { + "bbox": [ + 54, + 629, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^*" + }, + { + "bbox": [ + 54, + 629, + 287, + 653 + ], + "type": "text", + "content": ", by solving the following linear system:" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 96, + 662, + 287, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 662, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 96, + 662, + 287, + 677 + ], + "type": "interline_equation", + "content": "\\mathbf {C} ^ {*} = \\arg \\min _ {\\mathbf {C}} E _ {\\mathrm {d e s c}} (\\mathbf {C}) + E _ {\\mathrm {r e g}} (\\mathbf {C}), \\tag {1}", + "image_path": "4d0e76213c07aa7b29b2784bc38d2b15375d5177dfe3e87c4b4b026ab9a513c4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 66, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{desc}}(\\mathbf{C}) = \\| \\mathbf{CA}_1 - \\mathbf{A}_2\\|^2" + }, + { + "bbox": [ + 66, + 689, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 66, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{reg}}(\\mathbf{C})" + }, + { + "bbox": [ + 66, + 689, + 287, + 714 + ], + "type": "text", + "content": " is the regularization term promoting structural properties" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 545, + 160 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 545, + 160 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 545, + 160 + ], + "type": "image", + "image_path": "143874733320dcf517fd437577aa58b5339c02e0b3042929de4bb47ed9008857.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 163, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 304, + 163, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 547, + 229 + ], + "type": "text", + "content": "Figure 2. Overview of our two-branch DFM network. The part colored in black corresponds to a standard DFM, we introduce a novel branch, colored in red, that estimates maps from spatial perspective, our loss function is defined as Eqn.6. " + }, + { + "bbox": [ + 304, + 163, + 547, + 229 + ], + "type": "inline_equation", + "content": "\\Phi_i,\\Phi_i^\\dagger" + }, + { + "bbox": [ + 304, + 163, + 547, + 229 + ], + "type": "text", + "content": " indicate projection of the regarding features into the spatial and spectral domain, respectively. See more details in Section 4.2." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 234, + 545, + 305 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 234, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 545, + 270 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 313, + 234, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 234, + 545, + 270 + ], + "type": "text", + "content": " , e.g., enforcing " + }, + { + "bbox": [ + 313, + 234, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 234, + 545, + 270 + ], + "type": "text", + "content": " to be commutative with the Laplace-Beltrami operators and to be orthogonal. In particular, we let " + }, + { + "bbox": [ + 313, + 234, + 545, + 270 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{total}}(\\mathbf{C}) = E_{\\mathrm{desc}}(\\mathbf{C}) + E_{\\mathrm{reg}}(\\mathbf{C})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 270, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 545, + 305 + ], + "type": "text", + "content": "4. Convert the estimated functional map " + }, + { + "bbox": [ + 313, + 270, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 270, + 545, + 305 + ], + "type": "text", + "content": " to a pointwise map by conducting nearest neighbor search between the rows of " + }, + { + "bbox": [ + 313, + 270, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\Phi_1\\mathbf{C}" + }, + { + "bbox": [ + 313, + 270, + 545, + 305 + ], + "type": "text", + "content": " and that of " + }, + { + "bbox": [ + 313, + 270, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\Phi_{2}" + }, + { + "bbox": [ + 313, + 270, + 545, + 305 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 316, + 546, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 546, + 448 + ], + "type": "text", + "content": "The feature extractor is then trained on a set of training shape pairs, which typically enumerates all pairs from some given training set. The sub-optimization problem in Step (3) is generally decomposed into two parts to circumvent using an iterative solver. Namely, a closed-form solution is obtained either by a least-square estimation w.r.t " + }, + { + "bbox": [ + 304, + 316, + 546, + 448 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{desc}}(\\mathbf{C})" + }, + { + "bbox": [ + 304, + 316, + 546, + 448 + ], + "type": "text", + "content": " [30], or by solving a sequence of linear systems involving Laplacian commutativity as well [11]. Both correspond to the FMreg layer shown in Fig. 2. The rest of " + }, + { + "bbox": [ + 304, + 316, + 546, + 448 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{reg}}(\\mathbf{C})" + }, + { + "bbox": [ + 304, + 316, + 546, + 448 + ], + "type": "text", + "content": " is then set as the structural loss, corresponding to " + }, + { + "bbox": [ + 304, + 316, + 546, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{C})" + }, + { + "bbox": [ + 304, + 316, + 546, + 448 + ], + "type": "text", + "content": " in Fig. 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 455, + 413, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 413, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 413, + 468 + ], + "type": "text", + "content": "3.2. Cycle Consistency" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": "We start by giving a formal definition of cycle consistency. We assume to be given a collection of shapes " + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "inline_equation", + "content": "S = \\{S_i\\}_{i=1}^n" + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": " and a set of maps " + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{T_{ij}\\}_{i,j \\in [1..n]}" + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": ". We call " + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": " cycle consistent if for any shape " + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": ", and any closed path " + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "inline_equation", + "content": "(i, i_1, i_2, \\dots, i_k, i)" + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": ", the map composition along this path " + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "inline_equation", + "content": "T_{ii} = T_{i_k i} \\circ \\dots \\circ T_{i_1 i_2} \\circ T_{ii_1}" + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": " is an identity map on " + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": ". We can similarly define cycle consistency on functional maps, the only difference is that map composition is given by matrix multiplication and we require the final functional map " + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "inline_equation", + "content": "C_{ii}" + }, + { + "bbox": [ + 304, + 473, + 545, + 593 + ], + "type": "text", + "content": " to be an identity matrix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "text", + "content": "We suppose that deep functional maps are trained on " + }, + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "text", + "content": " with respect to all possible pairs. Then the global energy is given by " + }, + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{total}}(\\mathcal{C}) = \\sum_{i,j} \\| \\mathbf{C}_{ij} \\mathbf{A}_i - \\mathbf{A}_j \\|^2 + \\sum_{i,j} E_{\\mathrm{reg}}(\\mathbf{C}_{ij})" + }, + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\mathbf{C}_{ij}\\}_{i,j \\in [1..n]}" + }, + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "text", + "content": " is the set of functional maps among training shapes." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Proposition 1 If " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "E_{\\text{total}}(\\mathcal{C}) = 0" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": ", then for any shape " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": ", and any path " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "(i, i_1, i_2, \\dots, i_k, i)" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": ", the map composition " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ii}" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " is cycle consistent within the functional space spanned by the columns of " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ii}\\mathbf{A}_i = \\mathbf{A}_i" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14499" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "As a consequence of Prop. 1, when " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i\\in \\mathbb{R}^{k\\times d}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " is of full row rank, then " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\mathbf{C}_{ij}\\}_{i,j\\in [1\\dots n]}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " is cycle consistent. In practice, we generally set " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "d > k" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": ". Moreover, during network training, " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " is computed via pseudo-inverse, which implicitly assumes the full-rankness of " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i,\\mathbf{A}_j" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": ". We refer readers to the empirical validation in Section 4.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 192 + ], + "type": "text", + "content": "We defer the proof of Prop. 1 to Supp. Material. In fact, a similar claim has been formulated and proven in [49] (see Sec.3.1 therein), but in the context of map refinement via promoting cycle consistency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 192, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 287, + 312 + ], + "type": "text", + "content": "Though being technically similar, the theoretical argument of [49] and that of ours have fundamentally different implications. More specifically, the former justifies a test-time optimization algorithm, which is used to promote cycle consistency of maps among a fixed test shape collection. While the latter suggests that spectrally cycle consistency has been ensured and further leveraged to enhance the universal feature extractor (independent of the test data) during training in any DFM framework following the generic pipeline presented in Sec. 3.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 323, + 243, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 243, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 243, + 337 + ], + "type": "text", + "content": "4. Two-branch Deep Functional Maps" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "text", + "content": "It has long been recognized, both theoretically and empirically, that optimizing purely in the spectral domain is not sufficient. As a toy example, a trivial solution attaining global optima can be constructed as follows: Suppose that we have learned a feature extractor " + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\Theta}" + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "text", + "content": ", which returns the respective eigenbasis transformed by a universal " + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_0" + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "text", + "content": ". That is, " + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_i = \\boldsymbol{\\Phi}_i\\mathbf{A}_0, \\forall i" + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "text", + "content": ", which implies " + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i = \\boldsymbol{\\Phi}_i^\\dagger\\mathbf{G}_i = \\mathbf{A}_0, \\forall i" + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "text", + "content": ". Then we have " + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij} \\equiv I_k" + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "text", + "content": ", which exactly satisfies " + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{desc}}(\\mathbf{C}_{ij}) = E_{\\mathrm{reg}}(\\mathbf{C}_{ij}) = 0, \\forall i, j" + }, + { + "bbox": [ + 46, + 343, + 287, + 463 + ], + "type": "text", + "content": ". However, it probably induces poor point-wise maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 464, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 559 + ], + "type": "text", + "content": "In fact, in [37] the authors have proposed to use an ICP-like technique to encourage the estimated functional maps to be induced by some point-wise maps. In [34], the authors propose a spectral upsampling method for map refinement, which essentially converts maps back and forth between spectral and spatial domains. Moreover, the following lemma from [34] sheds light on the necessity of taking both spectral and spatial representations into consideration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 572, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 632 + ], + "type": "text", + "content": "Lemma 1 Given a pair of shapes " + }, + { + "bbox": [ + 46, + 572, + 287, + 632 + ], + "type": "inline_equation", + "content": "S_{1}, S_{2}" + }, + { + "bbox": [ + 46, + 572, + 287, + 632 + ], + "type": "text", + "content": " each having nonrepeating Laplacian eigenvalues, which are the same. A point-wise map " + }, + { + "bbox": [ + 46, + 572, + 287, + 632 + ], + "type": "inline_equation", + "content": "T: S_{1} \\to S_{2}" + }, + { + "bbox": [ + 46, + 572, + 287, + 632 + ], + "type": "text", + "content": " is an isometry if and only if the corresponding functional map " + }, + { + "bbox": [ + 46, + 572, + 287, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 46, + 572, + 287, + 632 + ], + "type": "text", + "content": " in the complete Laplacian basis is both diagonal and orthonormal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": "The above lemma suggests that apart from promoting the structural properties of functional maps, it is also critical to enforce them to be associated with certain point-wise maps, or termed as properness of functional maps in [40]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Finally, we remark that some recent DFM advances also promote the properness of the resulting spectral maps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "For instance, AttentiveFMaps [28] follows the spirit of ZoomOut [34] and explicitly performs a conversion between spectral and spatial map representations across different dimensions of eigenbasis; UDMSM [9] constructs explicitly a universal shape in the feature space, and enforce the spectral map estimation to be consistent with the spatial maps induced via the universal shape." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 163, + 464, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 163, + 464, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 464, + 176 + ], + "type": "text", + "content": "4.1. Two-branch Map Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 182, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 230 + ], + "type": "text", + "content": "In this part, we leverage our observation made in Prop. 1 and propose a novel, simple yet effective design of unsupervised deep functional maps, which introduces a new branch that independently estimates maps from spatial perspective." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": "Our key insight is that, once cycle consistency is valid and " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " is of full row rank, " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " can be seen as a functional map from a universal latent shape, " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ", to " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ". This perspective has been explored in several prior works [49, 23, 24], we provide the following details to be self-contained. The above assumption implies " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij} = \\mathbf{A}_j\\mathbf{A}_i^\\dagger" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij}" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " can be interpreted as a functional map composition from " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ", followed by a map from " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_j" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ". On the other hand, one can align the spectral embeddings of " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_j" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " to that of " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " by simply transforming the former by " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij}" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ". Indeed, we convert " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{ij}" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " into the point-wise map by the nearest neighbor searching between the rows of " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\Phi_j\\mathbf{C}_{ij}" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " and that of " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\Phi_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ". From this point of view, denoting the virtual spectral embedding of the latent shape by " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\Phi_0" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\Phi_i\\mathbf{A}_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " can be then treated as the spectral embedding of " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " aligned to that of " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ". Therefore, given a pair of shapes " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_i, S_j" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ", since we have aligned their eigenbasis to the canonical frame defined by the virtual spectral embedding " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\Phi_0" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ", we can align the spectral embedding of " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\Phi_0" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " by computing " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\Phi_i\\mathbf{A}_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ". Once all the spectral embeddings are aligned to the canonical embedding domain regarding " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\Phi_0" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": ", we can compute the soft point-wise map between " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "S_j" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " by nearest neighbor searching between the rows of " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\Phi_i\\mathbf{A}_i" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": " and those of " + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "inline_equation", + "content": "\\Phi_j\\mathbf{A}_j" + }, + { + "bbox": [ + 304, + 230, + 546, + 505 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "type": "text", + "content": "Based on the above derivation, given the learned features projected in the spectral domain, " + }, + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i,\\mathbf{A}_j" + }, + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "type": "text", + "content": ", and a pair of indices " + }, + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "type": "inline_equation", + "content": "p\\in [1..n_i],q\\in [1..n_j]" + }, + { + "bbox": [ + 304, + 506, + 545, + 553 + ], + "type": "text", + "content": ", we can compute point-wise maps. Firstly we compute residual:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 563, + 545, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 563, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 361, + 563, + 545, + 576 + ], + "type": "interline_equation", + "content": "\\delta_ {q p} = \\left\\| \\Phi_ {i} [ p ] \\mathbf {A} _ {i} - \\Phi_ {j} [ q ] \\mathbf {A} _ {j} \\right\\| _ {2}, \\tag {2}", + "image_path": "5afc36e2177ce3188b614bd6fcaaa4bf274163da167ae31e9959a2c1a327a183.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\Phi_i[p]" + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "text", + "content": "-th row of " + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\Phi_i" + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "text", + "content": ", and similarly we define " + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\Phi_j[q]" + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "text", + "content": ". The soft point-wise map " + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "inline_equation", + "content": "\\Pi \\in \\mathbb{R}^{n_j \\times n_i}" + }, + { + "bbox": [ + 304, + 585, + 545, + 620 + ], + "type": "text", + "content": " is then given by:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 363, + 630, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 630, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 363, + 630, + 545, + 658 + ], + "type": "interline_equation", + "content": "\\Pi (q, p) = \\frac {\\exp (- \\alpha \\delta_ {q p})}{\\sum_ {p ^ {\\prime}} \\exp (- \\alpha \\delta_ {q p ^ {\\prime}})}. \\tag {3}", + "image_path": "7c0f9ab1af880208bf9944361e11ef5d9b4fbea88282e866e9bb1c2057d20712.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Note that by construction, each row of " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is non-negative and sums up to 1, forming a probability distribution. The parameter " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " controls the entropy of each distribution - the smaller/larger " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is, the fuzzier/sharper the distribution is." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14500" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 275, + 202 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 275, + 202 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 275, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 275, + 202 + ], + "type": "image", + "image_path": "95ad605773bb07df2bf392fdc7ed80286db371f9990e4c239a2561d6072d540c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 287, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 287, + 237 + ], + "type": "text", + "content": "Figure 3. We train our two-branch DFM and a vanilla single-branch version on DT4D-H and monitor spectrally and spatially cycle consistency along the training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 240, + 287, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 288 + ], + "type": "text", + "content": "Instead of manually tuning the optimal " + }, + { + "bbox": [ + 46, + 240, + 287, + 288 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 240, + 287, + 288 + ], + "type": "text", + "content": ", we propose a learning scheme that dynamically controls " + }, + { + "bbox": [ + 46, + 240, + 287, + 288 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 240, + 287, + 288 + ], + "type": "text", + "content": " over training, which is inspired by curriculum learning [5]. We defer the respective details to Section 4.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 312 + ], + "type": "text", + "content": "We convert the soft point-wise map to a functional map by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 311, + 287, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 311, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 135, + 311, + 287, + 327 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {2} = \\boldsymbol {\\Phi} _ {j} ^ {\\dagger} \\Pi \\boldsymbol {\\Phi} _ {i}. \\tag {4}", + "image_path": "85fa4712329a5336a061b78a9597a5c2e5bab1a04298c07eea19fc93011eebe3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 329, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 353 + ], + "type": "text", + "content": "In the end, we enforce " + }, + { + "bbox": [ + 46, + 329, + 287, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2" + }, + { + "bbox": [ + 46, + 329, + 287, + 353 + ], + "type": "text", + "content": " to be consistent with " + }, + { + "bbox": [ + 46, + 329, + 287, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1" + }, + { + "bbox": [ + 46, + 329, + 287, + 353 + ], + "type": "text", + "content": ", the intermediate output from the FMreg layer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 425 + ], + "type": "text", + "content": "To summarize, thanks to the spectrally cycle consistency we identify in Sec. 3.2, we are allowed to construct a spectral latent shape and induce spatial maps via it. By enforcing the spatial estimations to be consistent with the spectral ones, we obtain a spatially and spectrally consistent deep functional maps framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 431, + 145, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 145, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 145, + 444 + ], + "type": "text", + "content": "4.2. Network Design" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 450, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 498 + ], + "type": "text", + "content": "As shown in Fig. 2, our two-branch network is built upon a standard DFM framework. In the following, we denote by " + }, + { + "bbox": [ + 46, + 450, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1" + }, + { + "bbox": [ + 46, + 450, + 287, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2" + }, + { + "bbox": [ + 46, + 450, + 287, + 498 + ], + "type": "text", + "content": " the estimated functional maps from the original FMreg layer and our novel branch, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 498, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 570 + ], + "type": "text", + "content": "Specifically, we use DiffusionNet [44] as our feature extractor. And WKS [4] descriptors are fed into it as initialization of learned features. We borrow the FMreg layer from [12]. It takes both " + }, + { + "bbox": [ + 46, + 498, + 287, + 570 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{desc}}(\\mathbf{C})" + }, + { + "bbox": [ + 46, + 498, + 287, + 570 + ], + "type": "text", + "content": " and commutativity with the Laplace-Beltrami operator into consideration, where the latter is given as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 110, + 576, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 576, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 110, + 576, + 287, + 594 + ], + "type": "interline_equation", + "content": "L _ {\\text {l a p}} = \\left\\| \\mathbf {C} _ {1} \\Lambda_ {1} - \\Lambda_ {2} \\mathbf {C} _ {1} \\right\\| ^ {2}, \\tag {5}", + "image_path": "3043b5ff40fc3da3f932bce81146c0cdad0ac90bf71611019dba3965d821d64b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 601, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 601, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\Lambda_{1}" + }, + { + "bbox": [ + 46, + 601, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 601, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\Lambda_{2}" + }, + { + "bbox": [ + 46, + 601, + 287, + 624 + ], + "type": "text", + "content": " are diagonal matrices of the Laplace-Beltrami eigenvalues on the two shapes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 625, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 287, + 649 + ], + "type": "text", + "content": "The estimation of " + }, + { + "bbox": [ + 46, + 625, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2" + }, + { + "bbox": [ + 46, + 625, + 287, + 649 + ], + "type": "text", + "content": " has been described in detail in Section 4.1. In the end, we formulate the training loss as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 73, + 655, + 287, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 655, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 73, + 655, + 287, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\mathbf {C} _ {1}, \\mathbf {C} _ {2}\\right) = \\left\\| \\mathbf {C} _ {1} ^ {T} \\mathbf {C} _ {1} - \\mathbf {I} \\right\\| ^ {2} + \\left\\| \\mathbf {C} _ {1} - \\mathbf {C} _ {2} \\right\\| ^ {2}, \\tag {6}", + "image_path": "4617468fab1d20181f6ddc923e04070ca490db575c2345f4eea396448b26c421.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "where the first term promotes the orthogonality of " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": ", while the second term promotes the consistency between the functional maps estimated from different branches." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "Finally, we remark that by combining the FMreg layer and " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{C}_1,\\mathbf{C}_2)" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": ", we have incorporated every factor in Lemma 1 into our design." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "text", + "content": "Conceptual Validation In this part, we train a network on DT4D-H dataset (see Section 5.1 for details) with our two-branch network, and a single-branch variant without our spatial map estimation branch. We monitor and plot the following quantities along training: (1) Average spectral cycle consistency over sampled triplets, i.e., " + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\frac{1}{M}\\sum_{(i,j,k)}\\| \\mathbf{C}_{ki}\\mathbf{C}_{jk}\\mathbf{C}_{ij} - I\\| ^2 /\\| I\\| ^2" + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "text", + "content": " ; (2) Average spatial cycle consistency over sampled triplets, i.e., the mean Euclidean deviation from composed maps " + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "inline_equation", + "content": "T_{ki}\\circ T_{jk}\\circ T_{ij}" + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "text", + "content": " to identity map on " + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "inline_equation", + "content": "n = 80" + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "text", + "content": " is the number of training shapes, and " + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "inline_equation", + "content": "M = 1000" + }, + { + "bbox": [ + 305, + 123, + 545, + 327 + ], + "type": "text", + "content": " is the number of sampled triplets. The behavior of the blue curves after 4500 iterations verifies our argument that spectrally cycle consistency does not imply spatially cycle consistency. On the other hand, by introducing our two-branch design, the discrepancy is well compensated and evidently better cycle consistencies in both spatial and spectral domains are achieved." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 334, + 485, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 485, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 485, + 346 + ], + "type": "text", + "content": "4.3. Updating Scheme of " + }, + { + "bbox": [ + 306, + 334, + 485, + 346 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 334, + 485, + 346 + ], + "type": "text", + "content": " in Eqn. (3)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "text", + "content": "The soft point-map conversion (Eqn. 3) has been applied in several prior works [30, 28], which all set " + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "text", + "content": " to be a manually selected constant. Ideally, we expect " + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "text", + "content": " in Eqn. (3) to be close to a permutation matrix, i.e., each row forms a binary vector. This seems to suggest a preference for a large " + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "text", + "content": ". Unfortunately, it would severely hinder network training, since the learned features and maps are of low quality in the early stage. On the other hand, a small " + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "text", + "content": " can alleviate such difficulty but falls short of fully pushing functional maps to be proper. As demonstrated in Section 5.5, neither small nor large " + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 352, + 545, + 483 + ], + "type": "text", + "content": " produces satisfying results." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 484, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 484, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 484, + 545, + 567 + ], + "type": "text", + "content": "Based upon the above analysis, we propose a novel updating scheme, which is inspired by curriculum learning [5]. Namely, we initiate a small " + }, + { + "bbox": [ + 305, + 484, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 484, + 545, + 567 + ], + "type": "text", + "content": " at the beginning of network training and increase it by a constant step size for every fixed number of epochs. As shown in Table 1, 2, 3, our scheme does not rely on hyper-parameter tuning but also achieves state-of-the-art results." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 575, + 438, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 438, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 438, + 587 + ], + "type": "text", + "content": "4.4. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "type": "text", + "content": "We implement our network with PyTorch [39]. We use four DiffusionNet blocks [44] as feature backbone and borrow the functional map block with Laplacian regularizer from [12]. The dimension of the Laplace-Beltrami eigenbasis is set to 50. WKS [4] descriptors are used as the input signal to our network. The dimensions of the input and the output descriptors are both set to 128. During training, the value of the learning rate is set to 2e-4 with ADAM optimizer. In all experiments, we train our method for 10,000 iterations with a batch size of 1. Following the learning" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14501" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "strategy in Section 4.3, we initialize " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " to 1 and increase it by 5 per epoch. As indicated in Eqn. (6), We weigh equally the orthogonality loss with respect to " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " and the residual between " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": ". More implementation details are provided in the Supp. Material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 149, + 173, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 149, + 173, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 149, + 173, + 164 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 173, + 288, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 173, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 288, + 328 + ], + "type": "text", + "content": "In this section, we conduct an extensive set of experiments of non-rigid shape matching on various datasets including humanoids and animals. We test on both near-isometric and non-isometric shape pairs. Our method is compared to a set of competitive baselines including axiomatic, supervised, weakly-supervised, and unsupervised learning methods. We emphasize that in this section, all the maps from the learning-based pipelines are directly inferred from the trained models, without any post-processing procedure. We evaluate the matching results in terms of a mean geodesic error on shapes normalized to unit area. Finally, our point-wise maps are all inferred by converting the output functional maps, as all the other DFM frameworks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 342, + 111, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 111, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 111, + 355 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 363, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 400 + ], + "type": "text", + "content": "FAUST_r: The remeshed version [41] of FAUST dataset[6] contains 100 human shapes. Following [43], it is split into 80/20 for training and testing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 437 + ], + "type": "text", + "content": "SCAPE_r: The remeshed version [41] of SCAPE dataset[2] contains 71 human shapes. Following [43], it is split into 51/20 for training and testing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 440, + 287, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 500 + ], + "type": "text", + "content": "SHREC19_r: The remeshed version of SHREC19 dataset [33] collects 44 human shapes from 11 independent datasets with distinctive poses and styles. We abandon shape 40 due to its partiality, we test on 407 pairs among the rest 43 shapes, which come with ground-truth." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 502, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 562 + ], + "type": "text", + "content": "DT4D-H [32]: The remeshed subset of the large scale animation dataset DeformingThings4D [29]. In particular, DT4D-H includes 10 categories of humanoid shapes undergoing significant pose and style variances, forming a challenging benchmark." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 564, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 613 + ], + "type": "text", + "content": "SMAL_r: The remeshed SMAL dataset [53] contains 49 animal shapes with 8 species. We follow the setting from [28], which splits 29 (5 species) and 20 (3 species) shapes for training and testing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 615, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 287, + 687 + ], + "type": "text", + "content": "TOSCA_r: The remeshed TOSCA dataset [7] contains multiple shape categories. We choose 4 animal categories, including cat, dog, horse, and wolf to verify the generalization performance of networks trained on SMAL_r. Note that we only infer the intra-category maps, due to the absence of ground-truth inter-category maps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "We refer readers to Supp. Material for visualizations illustrating the variability of the above datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 478, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 478, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 478, + 85 + ], + "type": "text", + "content": "5.2. Near-isometric Shape Matching" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 91, + 545, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 198 + ], + "type": "text", + "content": "In this part, we perform comparisons with an array of non-rigid shape matching methods: (1) Axiomatic methods including ZoomOut [34], BCICP [41], IsoMuSh [17], Smooth Shells [13], CZO [24]; (2) Supervised learning methods including TransMatch [48], GeomFMaps [12], and supervised version of AttentiveFMaps [28]; (3) Unsupervised learning methods including NeuroMorph [14], SyNoRiM [16], Deep Shell [15], AttentiveFMaps [28], UDMSM [9], DUO-FM [11]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 199, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 199, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 199, + 545, + 366 + ], + "type": "text", + "content": "For all learning-based methods, we train models on FAUST_r and SCAPE_r respectively. Table 1 reports results on both standard tests and more challenging generalizations. We observe a trade-off between the two tasks, methods that perform the best in the former (e.g., supervised AttentiveFMaps and UDMSM) tend to overfit, and therefore suffer poor generalization (especially to SREHC19_r). Meanwhile, our default setting, denoted by Ours, achieves reasonable performance in the standard tests but also outperforms the external baselines in 3 out of 4 generalization tests. Especially, in generalizing to SHREC19_r, Ours outperforms the external baselines by a large margin, resulting in " + }, + { + "bbox": [ + 304, + 199, + 545, + 366 + ], + "type": "inline_equation", + "content": "41\\%" + }, + { + "bbox": [ + 304, + 199, + 545, + 366 + ], + "type": "text", + "content": " (3.8 vs. 6.4) and " + }, + { + "bbox": [ + 304, + 199, + 545, + 366 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 304, + 199, + 545, + 366 + ], + "type": "text", + "content": " (4.5 vs. 8.4) error reduction upon the second best." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 367, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 545, + 464 + ], + "type": "text", + "content": "We highlight that post-processing with cycle consistency generally depends on the initialized map quality and the size of the test set (e.g., " + }, + { + "bbox": [ + 304, + 367, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\geq 3" + }, + { + "bbox": [ + 304, + 367, + 545, + 464 + ], + "type": "text", + "content": " shapes). In contrast, we leverage cycle consistency to improve the feature extractor during training. We also report the results from post-processing techniques based on cycle consistency [17, 24, 16] in Table 1. They are significantly outperformed by our method, which is inferred per-pair and without any post-processing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 464, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 596 + ], + "type": "text", + "content": "We further augment the dimension of functional maps in network training to 80 (same as UDMSM), which is beneficial to near-isometric matching [28]. It is evident that Ours (80 dim) achieves on-par performance with the regarding state-of-the-art methods in standard tests. On the other hand, due to the significant variability between SHREC19_r and the training sets (see Supp. Material), augmenting dimension leads to worse generalization than before (5.5 vs. 3.8, 5.8 vs. 4.5). Nevertheless, even in this case, our method outperforms the external baselines in all generalization tests by a notable margin." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 605, + 474, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 474, + 618 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 474, + 618 + ], + "type": "text", + "content": "5.3. Non-isometric Shape Matching" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 624, + 545, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 672 + ], + "type": "text", + "content": "We also train our network on non-isometric datasets, SMAL_r and DT4D-H, and compare it with the state-of-the-art baselines including DeepShells [15], AttentiveFMaps [28], UDMSM [9] and DUO-FM [11]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "SMAL_r: We follow the split and input descriptors from [28] (more details are provided in the Supp. Material). Table 2 reports results on SMAL_r, our method" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14502" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 80, + 547, + 270 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 522, + 79 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 522, + 79 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 522, + 79 + ], + "type": "text", + "content": "Table 1. Mean geodesic errors (×100) on FAUST_r, SCAPE_r, and SHREC19_r. The best and the second best are highlighted." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 80, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 50, + 80, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 50, + 80, + 547, + 270 + ], + "type": "table", + "html": "
MethodTrain TestFAUST_rSCAPE_r
FAUST_rSCAPE_rSHREC19_rSCAPE_rFAUST_rSHREC19_r
ZM[34]6.1\\\\7.5\\\\
BCICP[41]6.4\\\\11.0\\\\
IsoMuSh[17]4.4\\\\5.6\\\\
Smooth Shell[13]2.5\\\\4.7\\\\
CZO[24]2.2\\\\2.5\\\\
TransMatch[48]2.733.621.018.318.638.8
GeomFMaps[12]supervised2.63.39.93.03.012.2
AttentiveFMaps[28]1.42.29.41.71.812.2
NeuroMorph[14]8.528.526.329.918.227.6
SyNoRiM[16]7.921.725.59.524.626.8
Deep Shell[15]1.75.427.42.52.723.4
AttentiveFMaps[28]unsupervised1.92.66.42.22.29.9
UDMSM[9]1.57.321.52.08.630.7
DUO-FM[11]2.54.26.42.72.88.4
Ours2.32.63.82.42.54.5
Ours (80 dim)1.72.65.52.22.05.8
", + "image_path": "debab315f179157f4fdf5153877d5de3c77d29771d6dbc46dff585efc62c28af.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 53, + 308, + 281, + 407 + ], + "blocks": [ + { + "bbox": [ + 47, + 286, + 286, + 308 + ], + "lines": [ + { + "bbox": [ + 47, + 286, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 286, + 308 + ], + "type": "text", + "content": "Table 2. Mean geodesic errors " + }, + { + "bbox": [ + 47, + 286, + 286, + 308 + ], + "type": "inline_equation", + "content": "(\\times 100)" + }, + { + "bbox": [ + 47, + 286, + 286, + 308 + ], + "type": "text", + "content": " on SMAL_r. The best and the second best are highlighted correspondingly." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 308, + 281, + 407 + ], + "lines": [ + { + "bbox": [ + 53, + 308, + 281, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 281, + 407 + ], + "type": "table", + "html": "
MethodTrain TestSMAL_r
SMAL_rTOSCA_r
DeepShell[15]29.38.7
GeomFMaps[12]7.624.5
AttentiveFMaps[28]5.420.9
UDMSM[9]24.621.7
DUO-FM[11]32.815.3
Ours5.47.9
", + "image_path": "27f93f5c5bb9217e3ee51f9dd99d68cd46707bcfb296e8dd056473ce295d81f3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 431, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 574 + ], + "type": "text", + "content": "achieves the best performance, which is on-par with AttentiveFMaps [28]. To evaluate generalization performance, we use the trained models to directly infer intra-category maps within TOSCA-r. It turns out that AttentiveFMaps and GeomFMaps both suffer from significant performance drops (" + }, + { + "bbox": [ + 46, + 431, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\times 3.8" + }, + { + "bbox": [ + 46, + 431, + 287, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 431, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\times 3.2" + }, + { + "bbox": [ + 46, + 431, + 287, + 574 + ], + "type": "text", + "content": " larger geodesic error). It is also worth noting that DeepShells achieves the second-best generalization score in the relatively simpler task. However, it fails dramatically regarding the base task. In contrast, our method achieves the best balance between learning in difficult non-isometric pairs and generalizing to relatively easy near-isometric pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "DT4D-H: We follow the train/test (198/95) split of [28], but ignore the categories mousey and ortiz in both train and test, due to the lack of inter-category map labels regarding them, resulting a split of 168/80. We emphasize that we conduct training and test in a category-agnostic manner, i.e., no class label is used, and the training pairs can consist of shapes from arbitrary two categories. This is significantly different from [28], in which training pairs are selected according to clustering information. Obviously, our setting is more practical, but also more challenging. For completeness, we report results under the setting of [28] in Supp. Material" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 289, + 400, + 411 + ], + "blocks": [ + { + "bbox": [ + 316, + 289, + 400, + 411 + ], + "lines": [ + { + "bbox": [ + 316, + 289, + 400, + 411 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 400, + 411 + ], + "type": "image", + "image_path": "d41dc492e8de2981b3f434e3f45cafc90393bd9a3fa7c3dec84657091f86aaa5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 412, + 394, + 423 + ], + "lines": [ + { + "bbox": [ + 326, + 412, + 394, + 423 + ], + "spans": [ + { + "bbox": [ + 326, + 412, + 394, + 423 + ], + "type": "inline_equation", + "content": "T_{11}^{(8)} = T_{81}\\circ T_{78}\\circ \\dots \\circ T_{23}\\circ T_{12}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 410, + 289, + 501, + 352 + ], + "blocks": [ + { + "bbox": [ + 410, + 289, + 501, + 352 + ], + "lines": [ + { + "bbox": [ + 410, + 289, + 501, + 352 + ], + "spans": [ + { + "bbox": [ + 410, + 289, + 501, + 352 + ], + "type": "image", + "image_path": "64ccf4fd1979f7cea35e717ec983ffe0a1bc56f37a4e7455c25af9d29ed047e1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 354, + 437, + 360 + ], + "lines": [ + { + "bbox": [ + 423, + 354, + 437, + 360 + ], + "spans": [ + { + "bbox": [ + 423, + 354, + 437, + 360 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 501, + 289, + 548, + 352 + ], + "blocks": [ + { + "bbox": [ + 501, + 289, + 548, + 352 + ], + "lines": [ + { + "bbox": [ + 501, + 289, + 548, + 352 + ], + "spans": [ + { + "bbox": [ + 501, + 289, + 548, + 352 + ], + "type": "image", + "image_path": "2e3df59357d281208aec89702037854f0565f8c2c96010fc7d04560cabb6c515.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 354, + 530, + 360 + ], + "lines": [ + { + "bbox": [ + 507, + 354, + 530, + 360 + ], + "spans": [ + { + "bbox": [ + 507, + 354, + 530, + 360 + ], + "type": "text", + "content": "UDMSM" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 410, + 361, + 501, + 423 + ], + "blocks": [ + { + "bbox": [ + 410, + 361, + 501, + 423 + ], + "lines": [ + { + "bbox": [ + 410, + 361, + 501, + 423 + ], + "spans": [ + { + "bbox": [ + 410, + 361, + 501, + 423 + ], + "type": "image", + "image_path": "4ea98edba3fcd026d767278c8b76c4d8fe67e02aa930fe28629e6dab1125d9c1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 425, + 444, + 432 + ], + "lines": [ + { + "bbox": [ + 420, + 425, + 444, + 432 + ], + "spans": [ + { + "bbox": [ + 420, + 425, + 444, + 432 + ], + "type": "text", + "content": "DUO-FM" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 425, + 490, + 432 + ], + "lines": [ + { + "bbox": [ + 463, + 425, + 490, + 432 + ], + "spans": [ + { + "bbox": [ + 463, + 425, + 490, + 432 + ], + "type": "text", + "content": "Deep Shell" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 452, + 361, + 501, + 423 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 452, + 361, + 501, + 423 + ], + "spans": [ + { + "bbox": [ + 452, + 361, + 501, + 423 + ], + "type": "image", + "image_path": "cbb85b510f8610adec33f582d03fca7ed6028aac0d266a364cee39dcbf0eff7b.jpg" + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 502, + 361, + 548, + 423 + ], + "blocks": [ + { + "bbox": [ + 502, + 361, + 548, + 423 + ], + "lines": [ + { + "bbox": [ + 502, + 361, + 548, + 423 + ], + "spans": [ + { + "bbox": [ + 502, + 361, + 548, + 423 + ], + "type": "image", + "image_path": "047059c4388edffeebf8516da0a316d882ca06a1329d31fa27fc53f8f487a78b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 425, + 529, + 432 + ], + "lines": [ + { + "bbox": [ + 511, + 425, + 529, + 432 + ], + "spans": [ + { + "bbox": [ + 511, + 425, + 529, + 432 + ], + "type": "text", + "content": "BCICP" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 437, + 545, + 481 + ], + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 481 + ], + "type": "text", + "content": "Figure 4. Qualitative evaluation of spatial cycle consistency of different methods. Even composed along a path of 8 highly deformed shapes, our resulting map remains close to identity, while all the baselines fail significantly." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 492, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 517 + ], + "type": "text", + "content": "and our method outperforms the baselines in both intra- and inter-category evaluation by a notable margin." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 519, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 674 + ], + "type": "text", + "content": "Table 3 reports results on DT4D-H, in which we preserve 80 shapes for test and train networks with 168 and 80 shapes, respectively. Note that we report mean geodesic errors over all possible test shape pairs, which may undergo significant distortions (see, e.g., Fig. 1). Our method obtains a " + }, + { + "bbox": [ + 304, + 519, + 545, + 674 + ], + "type": "inline_equation", + "content": "67.2\\%" + }, + { + "bbox": [ + 304, + 519, + 545, + 674 + ], + "type": "text", + "content": " (7.7 vs. 22.4) geodesic error reduction with respect to the second-best baseline. On top of that, we also test the generalization of the trained model on near-isometric benchmarks – our method also generalizes the best in generalization to FAUST_r and SCAPE_r. The same pattern is observed when the training set is reduced by more than half. Remarkably, our network trained on the reduced set still outperforms all the baselines trained on the full set." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Overall, we attribute our performance on matching challenging non-isometric shapes (Table 3) and generalizing to unseen shapes (Table 1) to our effort to promote both spec" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14503" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 79, + 547, + 157 + ], + "blocks": [ + { + "bbox": [ + 88, + 69, + 501, + 79 + ], + "lines": [ + { + "bbox": [ + 88, + 69, + 501, + 79 + ], + "spans": [ + { + "bbox": [ + 88, + 69, + 501, + 79 + ], + "type": "text", + "content": "Table 3. Mean geodesic errors " + }, + { + "bbox": [ + 88, + 69, + 501, + 79 + ], + "type": "inline_equation", + "content": "\\left( {\\times {100}}\\right)" + }, + { + "bbox": [ + 88, + 69, + 501, + 79 + ], + "type": "text", + "content": " on DT4D-H. The best and the second best are highlighted correspondingly." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 79, + 547, + 157 + ], + "lines": [ + { + "bbox": [ + 50, + 79, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 50, + 79, + 547, + 157 + ], + "type": "table", + "html": "
MethodTrain TestDT4D-H (168)DT4D-H (80)
DT4D-HFAUST_rSCAPE_rDT4D-HFAUST_rSCAPE_r
DeepShell[15]27.04.96.529.34.77.0
AttentiveFMaps[28]25.73.46.428.92.76.3
UDMSM[9]46.843.347.949.742.540.0
DUO-FM[11]22.410.012.224.78.09.2
Ours7.73.16.19.02.66.2
", + "image_path": "3004ac49aba51731e73092f2fa7d6726e09d9547e53dcb9a01335d50acd8fd62.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 189, + 268, + 251 + ], + "blocks": [ + { + "bbox": [ + 47, + 169, + 286, + 190 + ], + "lines": [ + { + "bbox": [ + 47, + 169, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 169, + 286, + 190 + ], + "type": "text", + "content": "Table 4. Mean geodesic errors " + }, + { + "bbox": [ + 47, + 169, + 286, + 190 + ], + "type": "inline_equation", + "content": "\\left( {\\times {100}}\\right)" + }, + { + "bbox": [ + 47, + 169, + 286, + 190 + ], + "type": "text", + "content": " of SURFMNet and our variant trained on 4 datasets" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 189, + 268, + 251 + ], + "lines": [ + { + "bbox": [ + 66, + 189, + 268, + 251 + ], + "spans": [ + { + "bbox": [ + 66, + 189, + 268, + 251 + ], + "type": "table", + "html": "
MethodSURFMNetSURFMNet + Ours
FAUST_r6.03.5
SCAPE_r6.83.4
SMAL_r20.413.3
DT4D-H18.315.0
", + "image_path": "5e30d750640885c0515730b9c0e6cc7b6949bf42b1f0bbed88df5b2fd768af20.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 273, + 286, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 286, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 286, + 320 + ], + "type": "text", + "content": "trally and spatially cycle consistency. Especially, the isometry assumption is likely violated in the former case, thus cycle consistency, as a generic prior, plays an important role of regularizing maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 321, + 286, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 286, + 441 + ], + "type": "text", + "content": "As an illustration, we present a qualitative evaluation on the point-wise cycle consistency in Fig. 4. We sample 8 shapes from the test set of DT4D-H (one from each category) and compose the maps along the path " + }, + { + "bbox": [ + 46, + 321, + 286, + 441 + ], + "type": "inline_equation", + "content": "(S_{1} \\to S_{2} \\to \\dots \\to S_{8} \\to S_{1})" + }, + { + "bbox": [ + 46, + 321, + 286, + 441 + ], + "type": "text", + "content": " with respect to different approaches. It is evident that due to the significant distortion undergoing among the shapes, all but our method fails to preserve cycle consistency in this demanding test, while our composing map approximates the identity map on " + }, + { + "bbox": [ + 46, + 321, + 286, + 441 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 46, + 321, + 286, + 441 + ], + "type": "text", + "content": ". It also aligns nicely with the quantitative results reported in Table 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 449, + 226, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 226, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 226, + 462 + ], + "type": "text", + "content": "5.4. Integration with SURFMNet [43]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "type": "text", + "content": "Our two-branch design can be easily incorporated into any existing DFM framework following the general design outlined in Sec. 3.1. To demonstrate this, we modify the SURFMNet [43], one of the earliest approaches of unsupervised DFM, by adding our new branch. Table 4 shows the matching accuracy on the four main benchmarks. It is evident that in every case, incorporating our design leads to significant error reduction ranging from " + }, + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "type": "text", + "content": ". Especially, in the near-isometric cases, we obtain " + }, + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "type": "inline_equation", + "content": "41.6\\%" + }, + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 469, + 286, + 612 + ], + "type": "text", + "content": " error reduction respectively. Note that the absolute scores, 0.035, 0.034, are reasonable even compared to the state-of-the-art results reported in Table 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 622, + 140, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 140, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 140, + 635 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 700 + ], + "type": "text", + "content": "In this section, we present a set of ablation studies consisting of two parts. The first part verifies the rationality of our method, and the second part demonstrates the robustness of our method. We conduct all experiments on SMAL_r dataset [53]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "First of all, instead of using the updating scheme in Sec-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "text", + "content": "tion 4.3, we test the performance of our pipeline using two fixed values of " + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "text", + "content": " in Eqn. (3): " + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\alpha = 50" + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "text", + "content": ". Compared to our proposed model, the two variants yield a noticeable performance drop. Especially in the case " + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\alpha = 50" + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "text", + "content": ", the network fails to deliver reasonable matching results. We believe it is because a large " + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 172, + 545, + 256 + ], + "type": "text", + "content": " amplifies the noise of maps learned at the early training stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 257, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 389 + ], + "type": "text", + "content": "Then we justify our two-branch network design. Removing spatial branch amounts to training a standard single-branch DFM. To remove the spectral branch, we remove the FMreg layer and instead use our new branch to compute point-wise maps and convert them to functional maps. In the end, we modify the training loss so that it covers descriptor preservation, commutativity with the Laplace-Beltrami operator, and orthogonality (the latter two compensate the removed FMreg layer). The accuracy drop reported in the third and fourth row of Table 5 clearly suggests the necessity of our two-branch design." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": "In our experiments, we always use the full resolution meshes (" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\sim 5k" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": " vertices) and compute in Eqn. (2) with all of the 128 descriptors. We anticipate that efficiency can become an issue when the input mesh resolution is high, and/or we would like to increase the size of learned descriptors. Therefore, we test the robustness of our pipeline with respect to down-sampling, which is commonly used in functional maps-based frameworks [34, 28]: 1) We down-sample 3000 vertices on each shape via furthest point sampling; 2) In order to down-sample the feature dimension, we operate as the following during training: given a " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_1" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_2" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": ", we perform SVD on " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_1" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_1 = U_1\\Sigma_1V_1^T" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": ", then we set " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}_1 = \\mathbf{A}_1\\hat{V}_1" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": ", and set " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}_2 = \\mathbf{A}_2\\hat{V}_1" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\hat{V}_1" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": " is the first " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": " columns of " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "m = 30" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": " by replacing " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}_i" + }, + { + "bbox": [ + 304, + 390, + 545, + 594 + ], + "type": "text", + "content": " in Eqn. (2). The results in the bottom two rows show that the above operation has a relatively minor effect on the performance, proving the robustness of our method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 608, + 378, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 378, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 378, + 620 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "To conclude, we provide a theoretical justification for spectral cycle consistency in DFM. To compensate for the discrepancy of purely spectrally cycle consistency, we formulate a spectral latent shape that allows the alignment of the spectral embeddings of each shape to a canonical embedding domain. Based on this construction, we introduce a two-branch architecture for estimating maps in both spec" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14504" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 89, + 237, + 177 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 89 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 89 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 89 + ], + "type": "text", + "content": "Table 5. Mean geodesic errors " + }, + { + "bbox": [ + 47, + 69, + 287, + 89 + ], + "type": "inline_equation", + "content": "(\\times 100)" + }, + { + "bbox": [ + 47, + 69, + 287, + 89 + ], + "type": "text", + "content": " on different ablated settings, the models are all train on SMAL_r." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 89, + 237, + 177 + ], + "lines": [ + { + "bbox": [ + 99, + 89, + 237, + 177 + ], + "spans": [ + { + "bbox": [ + 99, + 89, + 237, + 177 + ], + "type": "table", + "html": "
Ours5.4
α = 16.6
α = 5035.2
remove spatial branch33.4
remove spectral branch14.3
Vertex downsampling5.7
Feature downsampling5.9
", + "image_path": "1157f33293cfad9fa526bcecf3106a394fa4fb72aa46aa4cbcf01c8a22328a08.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 196, + 287, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 287, + 279 + ], + "type": "text", + "content": "tral and spatial domains within DFM. The resulting network is simple, computationally efficient, and compatible with most existing DFM frameworks. We demonstrate the effectiveness of our framework through a comprehensive set of experiments, showing significant improvements upon state-of-the-art approaches in terms of accuracy, consistency, and generalization performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 339 + ], + "type": "text", + "content": "It is worth noting, though, our method is developed upon clean and complete meshes. It would be interesting for future work to investigate the utility of our approach in more general tasks, involving partial meshes, noisy point clouds, and other representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "text", + "content": "Acknowledgement This work was supported in part by the National Natural Science Foundation of China under contract No. 62171256 and Shenzhen Key Laboratory of next-generation interactive media innovative technology (No. ZDSYS20210623092001004), and in part by the ERC Starting Grant No. 758800 (EXPROTEA) and the ANR AI Chair AIGRETTE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 448, + 105, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 105, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 105, + 460 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 468, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 468, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 500 + ], + "type": "text", + "content": "[1] Noam Aigerman, Roi Poranne, and Yaron Lipman. Seamless Surface Mappings. ACM Transactions on Graphics (TOG), 34(4):72, 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 502, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 555 + ], + "type": "text", + "content": "[2] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. SCAPE: Shape Completion and Animation of People. In ACM Transactions on Graphics (TOG), volume 24, pages 408-416. ACM, 2005. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 558, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 287, + 590 + ], + "type": "text", + "content": "[3] Souhaib Attaiki, Gautam Pai, and Maks Ovsjanikov. Dpfm: Deep partial functional maps. In 2021 International Conference on 3D Vision (3DV), pages 175-185. IEEE, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 592, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 592, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 53, + 592, + 287, + 646 + ], + "type": "text", + "content": "[4] Mathieu Aubry, Ulrich Schlickewei, and Daniel Cremers. The Wave Kernel Signature: A Quantum Mechanical Approach to Shape Analysis. In Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on, pages 1626-1633. IEEE, 2011. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 287, + 680 + ], + "type": "text", + "content": "[5] Yoshua Bengio, Jerome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In International Conference on Machine Learning, 2009. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 681, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 287, + 714 + ], + "type": "text", + "content": "[6] Federica Bogo, Javier Romero, Matthew Loper, and Michael J. Black. FAUST: Dataset and evaluation for 3D mesh registration. In Proceedings IEEE Conf. on Com" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "puter Vision and Pattern Recognition (CVPR), Piscataway, NJ, USA, June 2014. IEEE. 1, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 128 + ], + "type": "text", + "content": "[7] Alexander M Bronstein, Michael M Bronstein, and Ron Kimmel. Numerical geometry of non-rigid shapes. Springer Science & Business Media, 2008. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 312, + 130, + 545, + 163 + ], + "type": "text", + "content": "[8] Oliver Burghard, Alexander Dieckmann, and Reinhard Klein. Embedding shapes with Green's functions for global shape matching. Computers & Graphics, 68:1-10, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 164, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 164, + 545, + 196 + ], + "type": "text", + "content": "[9] D. Cao and F. Bernard. Unsupervised deep multi-shape matching. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 220 + ], + "type": "text", + "content": "[10] Bailing Deng, Yuxin Yao, Roberto M. Dyke, and Juyong Zhang. A survey of non-rigid 3d registration. 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 275 + ], + "type": "text", + "content": "[11] Nicolas Donati, Etienne Corman, and Maks Ovsjanikov. Deep orientation-aware functional maps: Tackling symmetry issues in shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-751, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 277, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 308 + ], + "type": "text", + "content": "[12] Nicolas Donati, Abhishek Sharma, and Maks Ovsjanikov. Deep geometric functional maps: Robust feature learning for shape correspondence. In CVPR, 2020. 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 310, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 342 + ], + "type": "text", + "content": "[13] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In CVPR, 2020. 1, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 344, + 547, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 547, + 410 + ], + "type": "text", + "content": "[14] Marvin Eisenberger, David Novotny, Gael Kerchenbaum, Patrick Labatut, Natalia Neverova, Daniel Cremers, and Andrea Vedaldi. Neuromorph: Unsupervised shape interpolation and correspondence in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7473-7483, 2021. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 464 + ], + "type": "text", + "content": "[15] Marvin Eisenberger, Aysim Token, Laura Leal-Taixe, and Daniel Cremers. Deep shells: Unsupervised shape correspondence with optimal transport. Advances in Neural Information Processing Systems, 33:10491-10502, 2020. 3, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 467, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 498 + ], + "type": "text", + "content": "[16] Jiahui Huang et. al. Multiway non-rigid point cloud registration via learned functional map synchronization, 2022. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 500, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 555 + ], + "type": "text", + "content": "[17] Maolin Gao, Zorah Lahner, Johan Thunberg, Daniel Cremers, and Florian Bernard. Isometric multi-shape matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14183-14193, 2021. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 589 + ], + "type": "text", + "content": "[18] Dvir Ginzburg and D. Raviv. Cyclic functional mapping: Self-supervised correspondence between non-isometric deformable shapes. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 623 + ], + "type": "text", + "content": "[19] Oshri Halimi, Or Litany, Emanuele Rodol'a, Alex Bronstein, and Ron Kimmel. Unsupervised learning of dense shape correspondence. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "text", + "content": "[20] Jiahui Huang, Tolga Birdal, Zan Gojcic, Leonidas J. Guibas, and Shi-Min Hu. Multiway Non-rigid Point Cloud Registration via Learned Functional Map Synchronization. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[21] Qixing Huang, Fan Wang, and Leonidas Guibas. Functional map networks for analyzing and exploring large shape collections. ACM TOG, 33(4), 2014. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14505" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 716 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[22] Qi-Xing Huang and Leonidas Guibas. Consistent shape maps via semidefinite programming. In Computer Graphics Forum, volume 32, pages 177-186. Wiley Online Library, 2013. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[23] Ruqi Huang, Panos Achlioptas, Leonidas Guibas, and Maks Ovsjanikov. Limit shapes – a tool for understanding shape differences and variability in 3d model collections. Computer Graphics Forum, 38(5):187–202, 2019. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "type": "text", + "content": "[24] Ruqi Huang, Jing Ren, Peter Wonka, and Maks Ovsjanikov. Consistent zoomout: Efficient spectral map synchronization. In Computer Graphics Forum, volume 39, pages 265-278. Wiley Online Library, 2020. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 228 + ], + "type": "text", + "content": "[25] Daniel F Huber. Automatic three-dimensional modeling from reality. PhD thesis, Citeseer, 2002. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 272 + ], + "type": "text", + "content": "[26] Artiom Kovnatsky, Michael M Bronstein, Alexander M Bronstein, Klaus Glashoff, and Ron Kimmel. Coupled quasiharmonic bases. In Computer Graphics Forum, volume 32, pages 439-448, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 274, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 327 + ], + "type": "text", + "content": "[27] Spyridon Leonardos, Xiaowei Zhou, and Kostas Daniilidis. Distributed consistent data association via permutation synchronization. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 2645-2652. IEEE, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 328, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 328, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 48, + 328, + 287, + 372 + ], + "type": "text", + "content": "[28] Lei Li, Nicolas Donati, and Maks Ovsjanikov. Learning multi-resolution functional maps with spectral attention for robust shape matching. In Advances in Neural Information Processing Systems, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 373, + 287, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 427 + ], + "type": "text", + "content": "[29] Yang Li, Hikari Takehara, Takafumi Taketomi, Bo Zheng, and Matthias Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12706-12716, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 429, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 429, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 429, + 287, + 483 + ], + "type": "text", + "content": "[30] Or Litany, Tal Remez, Emanuele Rodola, Alexander M. Bronstein, and Michael M. Bronstein. Deep functional maps: Structured prediction for dense shape correspondence. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5660-5668, 2017. 1, 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 483, + 287, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 287, + 516 + ], + "type": "text", + "content": "[31] Rabinovich M., Poranne R., Panozzo D., and Sorkine-Hornung O. Scalable locally injective mappings. In ACM Transactions on Graphics, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 517, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 517, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 517, + 287, + 559 + ], + "type": "text", + "content": "[32] Robin Magnet, Jing Ren, Olga Sorkine-Hornung, and Maks Ovsjanikov. Smooth non-rigid shape matching via effective dirichlet energy optimization. In International Conference on 3D Vision, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 561, + 287, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 561, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 48, + 561, + 287, + 616 + ], + "type": "text", + "content": "[33] Simone Melzi, Riccardo Marin, Emanuele Rodolà, Umberto Castellani, Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Shrec 2019: Matching humans with different connectivity. In Eurographics Workshop on 3D Object Retrieval, volume 7, page 3, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 617, + 287, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 617, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 48, + 617, + 287, + 660 + ], + "type": "text", + "content": "[34] Simone Melzi, Jing Ren, Emanuele Rodola, Peter Wonka, and Maks Ovsjanikov. Zoomout: Spectral upsampling for efficient shape correspondence. Proc. SIGGRAPH Asia, 2019, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 662, + 287, + 716 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 662, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 48, + 662, + 287, + 716 + ], + "type": "text", + "content": "[35] Andy Nguyen, Mirela Ben-Chen, Katarzyna Welnicka, Yinyu Ye, and Leonidas Guibas. An optimization approach to improving collections of shape maps. In Computer Graphics Forum, volume 30, pages 1481-1491. Wiley Online Library, 2011. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[36] Dorian Nogneng and Maks Ovsjanikov. Informative descriptor preservation via commutativity for shape matching. Computer Graphics Forum, 36(2):259-267, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[37] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional Maps: A Flexible Representation of Maps Between Shapes. ACM Transactions on Graphics (TOG), 31(4):30, 2012. 1, 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[38] Maks Ovsjanikov, Etienne Corman, Michael Bronstein, Emanuele Rodola, Mirela Ben-Chen, Leonidas Guibas, Frederic Chazal, and Alex Bronstein. Computing and processing correspondences with functional maps. In ACM SIGGRAPH 2017 Courses, page 5. ACM, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 273 + ], + "type": "text", + "content": "[39] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Linand N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. Pytorch: An imperative style, high-performance deep learning library, 2019. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 274, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 297 + ], + "type": "text", + "content": "[40] Jing Ren, Simone Melzi, Peter Wonka, and Maks Ovsjanikov. Discrete optimization for shape matching, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "text", + "content": "[41] Jing Ren, Adrien Poulenard, Peter Wonka, and Maks Ovsjanikov. Continuous and orientation-preserving correspondences via functional maps. ACM Trans. Graph., 37(6):248:1-248:16, Dec. 2018. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "text", + "content": "[42] Emanuele Rodolà, Luca Cosmo, Michael M Bronstein, Andrea Torsello, and Daniel Cremers. Partial functional correspondence. In Computer Graphics Forum, volume 36, pages 222–236. Wiley Online Library, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 387, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 420 + ], + "type": "text", + "content": "[43] Jean-Michel Roufosse, Abhishek Sharma, and Maks Ovsjanikov. Unsupervised deep learning for structured shape matching. ICCV, 2019. 1, 2, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "text", + "content": "[44] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM Transactions on Graphics, 2022. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "text", + "content": "[45] Meged Shoham, Amir Vaxman, and Mirela Ben-Chen. Hierarchical functional maps between subdivision surfaces. In Computer Graphics Forum, volume 38, pages 55-73. Wiley Online Library, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 499, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 499, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 499, + 545, + 543 + ], + "type": "text", + "content": "[46] Jian Sun, Maks Ovsjanikov, and Leonidas Guibas. A Concise and Provably Informative Multi-Scale Signature Based on Heat Diffusion. In Computer graphics forum, volume 28, pages 1383-1392, 2009. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "type": "text", + "content": "[47] Federico Tombari, Samuele Salti, and Luigi Di Stefano. Unique signatures of histograms for local surface description. In International Conference on Computer Vision (ICCV), pages 356-369, 2010. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 589, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 633 + ], + "type": "text", + "content": "[48] Giovanni Trappolini, Luca Cosmo, Luca Moschella, Ricardo Marin, Simone Melzi, and Emanuele Rodola. Shape registration in the time of transformers. Advances in Neural Information Processing Systems, 34:5731-5744, 2021. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 634, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 678 + ], + "type": "text", + "content": "[49] Fan Wang, Qixing Huang, and Leonidas J. Guibas. Image co-segmentation via consistent functional maps. In Proceedings of the IEEE International Conference on Computer Vision, pages 849-856, 2013. 2, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 679, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 545, + 712 + ], + "type": "text", + "content": "[50] Lanhui Wang and Amit Singer. Exact and stable recovery of rotations for robust synchronization. Information and Inference: A Journal of the IMA, 2(2):145-193, 2013. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14506" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 151 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[51] Lei Yang, Wenxi Liu, Zhiming Cui, Nenglun Chen, and Wenping Wang. Mapping in a cycle: Sinkhorn regularized unsupervised learning for point cloud shapes. In ECCV, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "type": "text", + "content": "[52] Christopher Zach, Manfred Klopschitz, and Marc Pollefeys. Disambiguating visual relations using loop constraints. In 2010 IEEE Computer Society Conference on Computer" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 72, + 545, + 150 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 326, + 72, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition, pages 1426-1433. IEEE, 2010. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[53] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 6, 8" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14507" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/8203ecb7-a9b9-46de-8112-1277f4eeac63_content_list.json b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/8203ecb7-a9b9-46de-8112-1277f4eeac63_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f4508d5e4249bfe7025422ab910622a18f20bd09 --- /dev/null +++ b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/8203ecb7-a9b9-46de-8112-1277f4eeac63_content_list.json @@ -0,0 +1,1877 @@ +[ + { + "type": "text", + "text": "Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution", + "text_level": 1, + "bbox": [ + 98, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long Sun, Jiangxin Dong, Jinhui Tang, and Jinshan Pan*", + "bbox": [ + 261, + 181, + 712, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computer Science and Engineering, Nanjing University of Science and Technology", + "bbox": [ + 114, + 198, + 854, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although deep learning-based solutions have achieved impressive reconstruction performance in image superresolution (SR), these models are generally large, with complex architectures, making them incompatible with low-power devices with many computational and memory constraints. To overcome these challenges, we propose a spatially-adaptive feature modulation (SAFM) mechanism for efficient SR design. In detail, the SAFM layer uses independent computations to learn multi-scale feature representations and aggregates these features for dynamic spatial modulation. As the SAFM prioritizes exploiting non-local feature dependencies, we further introduce a convolutional channel mixer (CCM) to encode local contextual information and mix channels simultaneously. Extensive experimental results show that the proposed method is $3 \\times$ smaller than state-of-the-art efficient SR methods, e.g., IMDN, and yields comparable performance with much less memory usage. Our source codes and pre-trained models are available at: https://github.com/sunny2109/SAFMN.", + "bbox": [ + 73, + 282, + 473, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 580, + 209, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single image super-resolution (SISR) aims to restore a high-resolution (HR) image from its low-resolution (LR) counterpart by recovering lost details. This longstanding and challenging task has recently attracted much attention due to the rapid development of streaming media or high-definition devices. As these scenarios are usually resource-constrained, it is of great interest to develop an efficient and effective SR method to estimate HR images for better visual display on these platforms or products.", + "bbox": [ + 75, + 604, + 468, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning-based SR methods have achieved significant performance improvements with the great evolution of hardware technologies, as we can use large amounts of data to train much larger or deeper neural networks for image SR [32, 53, 31, 5]. For example, RCAN [53] is a representative CNN-based image SR network with 15.59M parameters and reaching a depth of over 400 layers. One of the most significant drawbacks of these large models is that they require high computational costs, which makes them", + "bbox": [ + 75, + 741, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f5f94fe3b6904c2563a45c10789fd0ac832f8a950da2e113df2b59caeb481bb.jpg", + "image_caption": [ + "Figure 1. Model complexity and performance comparison between our proposed SAFMN model and other lightweight methods on Set5 [4] for $\\times 2$ SR. Circle sizes indicate the number of parameters. The proposed method achieves a better trade-off between model complexity and reconstruction performance." + ], + "image_footnote": [], + "bbox": [ + 504, + 253, + 885, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "challenging to deploy. Moreover, recent visual transformers (ViTs) [12, 31, 5] outperform convolutional neural networks (CNNs) in low-level vision tasks, and their results demonstrate that exploring non-local feature interactions is essential for high-quality reconstruction. But existing self-attention mechanisms are computationally expensive and unfriendly to efficient SR design. This, therefore, motivates us to develop a lightweight yet effective model for real-world applications of image super-resolution by integrating the principles of convolution and self-attention.", + "bbox": [ + 496, + 549, + 892, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To reduce the heavy computational burden, various methods, including efficient module design [11, 41, 1, 20, 26, 33, 44, 45, 54, 30], knowledge distillation [16], neural architecture search [7], and structural re-parameterization [52], are trying to improve the efficiency of SR algorithms. Among these efficient SR models, one direction is to reduce model parameters or complexity (FLOPs). Lightweight strategies like recursive manner [23, 46], parameter sharing [1], and spare convolutions [1, 45, 30] are adopted. Although these approaches certainly reduce the model size, they usually compensate for the performance drop caused by shared recursive modules or sparse convolutions by increasing the depth or width", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13190", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of the model, which affects the inference efficiency when performing SR reconstruction.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another direction is to accelerate the inference time. The post-upsampling [11, 44] is an important replacement for the pre-defined input [10, 24], which significantly speeds up the runtime. Model quantization [21] effectively accelerates latency and reduces energy consumption, particularly when deploying algorithms in edge-devices. Structural re-parameterization [52, 8] improves the speed of a well-trained model in the inference stage. These methods enjoy fast running time but poor reconstruction performance. Consequently, there is still room for a better trade-off between model efficiency and reconstruction performance.", + "bbox": [ + 75, + 121, + 467, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above-mentioned issues, we design a simple yet effective model by developing a spatially-adaptive feature modulation, namely SAFMN, to realize a favorable trade-off between performance and efficiency. Specifically, we first exploit non-local feature relations to dynamically select representative features (see Figure 2 and 5) by implementing a feature modulation mechanism based on multiscale representations. As the modulation mechanism processes input features from a non-local perspective, there is a requirement to complement local contextual information. To this end, we present a convolutional channel mixer based on the FMBCov [47] to encode local features and mix channels. Taken together, we find that the SAFMN network is able to achieve a better trade-off between SR performance and model complexity, as shown in Figure 1.", + "bbox": [ + 75, + 287, + 467, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are summarized as follows:", + "bbox": [ + 76, + 513, + 467, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop an efficient feature modulation mechanism to learn feature dependencies, which absorbs CNN-like efficiency and transformer-like adaptability.", + "- We present a compact convolutional channel mixer that simultaneously encodes local contextual information and performs channel mixing.", + "- We evaluate the proposed method quantitatively and qualitatively on benchmark datasets, and the results show that our SAFMN achieves a favorable trade-off between accuracy and model complexity." + ], + "bbox": [ + 94, + 550, + 467, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 739, + 218, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Learning-based Image Super-Resolution. Classical interpolation algorithms, such as linear or bicubic upsampling, create high-resolution images by inserting zeros between adjacent pixels in the low resolution and then using a low-pass filter to preserve the content information of the input image [41]. Unlike these interpolation-based upsamplers, deep learning-based approaches learn a nonlinear mapping between the input image and the target output in an end-to-end training fashion. SRCNN [10] is the first attempt", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8a6be1bfafab898fefc3d04f472d9d70902d0229d999d1a0a4e310f1d54ec6a7.jpg", + "image_caption": [ + "DI", + "(a)", + "Figure 2. Comparison of local attribution maps (LAMs) [13] and diffusion indices (DIs) [13] between our SAFMN and other efficient SR models. (a) is the LR input, (b)-(e) are the results of CARN [1], EDSR-baseline [32], PAN [54] and SAFMN, respectively. The LAM results denote the importance of each pixel in the input LR image when super-resolving the patch marked with a red box. The DI value reflects the range of involved pixels. A larger DI value means a wider range of attention. The proposed method can exploit more feature information." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 586, + 146 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30acbc6abd5027152737258acc386d49e5d935ca744b4b69563848a6b88583c8.jpg", + "image_caption": [ + "3.2180", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 588, + 90, + 658, + 146 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c13192d0602280aaff706dd38fe8e389fce80cc98dd7e5214ca12d8aa7b863e7.jpg", + "image_caption": [ + "5.1936", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 663, + 90, + 733, + 146 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e67f2fbc1dd450fa7088d7a56e0dd2173494720f44016025753f5788df065d9.jpg", + "image_caption": [ + "4.0629", + "(d)" + ], + "image_footnote": [], + "bbox": [ + 738, + 90, + 808, + 146 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d4f6e8dd8f63302bb3ef3dd889913f61010d21db5dec58e2aa426494d618784e.jpg", + "image_caption": [ + "10.4295", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 815, + 90, + 887, + 146 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to use a convolutional neural network to tackle the image SR problem and achieves a considerable performance gain compared with conventional methods. Since then, many improvements have been proposed. VDSR [24] uses global residual learning [15] to solve the problem of difficulty in training an SR model with deep layers. DRRN [46] integrates the local residual learning and global residual connection to ease the training difficulty and enhance high-frequency details. EDSR [32] further increases the model footprint to 43M, achieving a significant breakthrough in reconstruction performance and showing that the BatchNorm (BN) [22] layer is not necessary for the SR task. RCAN [53] builds a more than 400 layers model based on channel attention and dense connections for accurate SR. NLSA [40] utilizes a non-local sparse attention to enhance the feature modeling capability and reaches state-of-the-art performance. With the successful application of ViT in various high-level vision tasks, image SR also follows this ViT [12] framework, achieving higher reconstruction performance than CNN-based models on extensive public benchmarks. For instance, SwinIR [31] acts as a strong baseline for image restoration tasks based on the Swin Transformer [34]. While these approaches achieve impressive reconstruction performance, the required high computational costs make them challenging to deploy in real-world applications on resource-constrained devices.", + "bbox": [ + 496, + 311, + 890, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient Image Super-Resolution. To improve the model efficiency, many CNN-based SR works try to alleviate this issue. FSRCNN [11] and ESPCN [44] utilize a postupsampling manner to reduce the computational burden from the pre-defined inputs significantly. CARN [1] uses group convolutions and a cascading mechanism upon residual networks to improve efficiency. IMDN [20] adopts feature splitting and concatenation operations to progressively aggregate features, and its improved variants [33, 26] won the AIM2020 and NTIRE2022 Efficient SR challenge. ShuffleMixer [45] introduces a large kernel convolution for lightweight SR design. BSRN [30] employs blueprint-separable convolutions to reduce model complexity. Mean", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13191", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/393c8fd6b666674a4e818e2c5fbfd0168d7f4e5b4b09f2957d58231ad3184e63.jpg", + "image_caption": [ + "Figure 3. An overview of the proposed SAFMN. SAFMN first transforms the input LR image into the feature space using a convolutional layer, performs feature extraction using a series of feature mixing modules (FMMs), and then reconstructs these extracted features by an upsampler module. The FMM block is implemented by a spatially-adaptive feature modulation (SAFM) layer, a convolutional channel mixer (CCM) and two skip-connections." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 893, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while, an increasingly popular direction is to compress or accelerate a well-trained deep model through model quantization [21], structural re-parameterization [52] or knowledge distillation [16]. Neural architecture search (NAS) is also commonly used to search a well-constrained architecture for image super-resolution [7]. Note that the efficiency of a deep neural network could be measured in different metrics, including the number of parameters, FLOPs, activations, memory consumption and inference running time [29]. Although the above approaches have been improved in different efficiency aspects, there is still room for a favorable trade-off between reconstruction performance and model efficiency.", + "bbox": [ + 75, + 309, + 472, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 521, + 250, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the core components of our proposed model for efficient SISR. As shown in Figure 3, the network consists of the following parts: a shallow convolution, a stacking of feature mixing modules (FMMs), and an upsampler layer. Specifically, we first apply a $3 \\times 3$ convolution layer to transform the input LR image to feature space and generate the shallow feature $\\mathbf{F}_0$ . Then, the multiple stacked FMMs are used to generate finer deep features from $\\mathbf{F}_0$ for HR image reconstruction, where an FMM layer has a spatially-adaptive feature modulation (SAFM) layer and a convolutional channel mixer (CCM). To recover the HR target image, we introduce a global residual connection to learn high-frequency details and employ a lightweight upsampling layer for fast reconstruction, which only contains a $3 \\times 3$ convolution and a pixel-shuffle layer [44].", + "bbox": [ + 75, + 547, + 468, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Spatially-adaptive Feature Modulation", + "text_level": 1, + "bbox": [ + 76, + 785, + 410, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies [9, 35, 14, 39, 31] suggest that the notable performance of ViTs across diverse tasks stems from their implementation of the key multi-head self-attention (MHSA) mechanism. This mechanism enables the model with the capability of long-range feature interaction and dynamic spatial weighting, both of which contribute to pro", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6320c5b4e1c0681f68965f91caa103296b32cdac5aa5978562909a3ea02eec1b.jpg", + "image_caption": [ + "Figure 4. Network architecture of the proposed SAFM module. This module performs spatially-adaptive modulation on the input features. Please see the Section 3.1 for more details." + ], + "image_footnote": [], + "bbox": [ + 514, + 309, + 879, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ducing promising results. But existing self-attention variants are computationally expensive and unfriendly to efficient SR design. In contrast, standard convolution is an efficient operation, but its performance is limited by the static weights and locality. This observation inspires us to borrow the idea of MHSA to enhance the representation capability of convolution.", + "bbox": [ + 496, + 462, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To introduce the ability of long-range interaction and dynamic modeling into convolution, we follow the multi-head paradigm, using parallel and independent computations that allow each head to process different scale information of the input, and then aggregate these features to generate an attention map for spatially-adaptive feature modulation.", + "bbox": [ + 496, + 569, + 892, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The detailed architecture of our proposed SAFM is depicted in Figure 4. Specifically, we first split the normalized input features into four-group components, and feed them into a multi-scale feature generation unit (MFGU), where a $3 \\times 3$ depth-wise convolution processes the first one, and the rest parts are sampled individually by pooling operations. As we desire to select discriminative features towards learning non-local feature interactions, adaptive max pooling operators are applied over the input features to gather information. Given the input feature $X$ , this procedure can be formulated as:", + "bbox": [ + 496, + 662, + 893, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ X _ {0}, X _ {1}, X _ {2}, X _ {3} \\right] = \\operatorname {S p l i t} (X),\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 840, + 743, + 858 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {0} = \\operatorname {D W - C o n v} _ {3 \\times 3} \\left(X _ {0}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 861, + 890, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {i} = \\uparrow_ {p} \\left(\\operatorname {D W - C o n v} _ {3 \\times 3} \\left(\\downarrow_ {\\frac {p}{2 ^ {i}}} \\left(X _ {i}\\right)\\right)\\right), 1 \\leq i \\leq 3,\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 883, + 851, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13192", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\text{Split}(\\cdot)$ corresponds to the channel split operation, $\\text{DW-Conv}_{3\\times 3}(\\cdot)$ is a $3\\times 3$ depth-wise convolution, $\\uparrow_p(\\cdot)$ represents upsampling features at a specific level to the original resolution $p$ via nearest interpolation for fast implementation, and $\\downarrow_{\\frac{p}{2^i}}$ denotes pooling the input features to the size of $\\frac{p}{2^i}$ . Figure 5(b)-(e) show that using such a design, we can obtain varying features from different representation subspaces.", + "bbox": [ + 76, + 90, + 467, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterward, we aggregate these extracted short- or long-range features by concatenating them on the channel dimension and performing a $1 \\times 1$ convolution. It can be formulated as:", + "bbox": [ + 76, + 214, + 467, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} = \\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left([ \\hat {X} _ {0}, \\hat {X} _ {1}, \\hat {X} _ {2}, \\hat {X} _ {3} ]\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 282, + 468, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Concat}(\\cdot)$ denotes the concatenation operation, and $\\operatorname{Conv}_{1\\times 1}(\\cdot)$ is the $1\\times 1$ convolution. After obtaining the aggregated representation $\\hat{X}$ (see Figure 5(f)), we normalize it through a GELU non-linearity [17] to estimate the attention map and adaptively modulate the input $X$ according to the estimated attention via element-wise product. This process can be written as:", + "bbox": [ + 76, + 319, + 467, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {X} = \\phi (\\hat {X}) \\odot X, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 436, + 468, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi(\\cdot)$ represents the GELU function and $\\odot$ is the element-wise product. The benefits of those operators are compared and results are shown in Table 3.", + "bbox": [ + 76, + 465, + 467, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Feature Mixing Module", + "text_level": 1, + "bbox": [ + 76, + 521, + 294, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned previously, the SAFM module prioritizes exploiting non-local feature interactions. To further incorporate local contextual information and perform channel mixing simultaneously, we introduce a compact convolutional channel mixer (CCM) based on the FMBCnv [47]. The CCM comprises a $3 \\times 3$ convolution and a $1 \\times 1$ convolution. Within this, the first $3 \\times 3$ convolution encodes the spatially local contexts and doubles the number of channels of the input features for mixing channels; the later $1 \\times 1$ convolution reduces the channels back to the original input dimension. A GELU [17] function is applied to the hidden layer for non-linear mapping.", + "bbox": [ + 76, + 544, + 467, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We formulate the proposed SAFM and CCM into a unified feature mixing module (FMM) to select representative features. The FMM can be written as:", + "bbox": [ + 76, + 724, + 467, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Y = \\operatorname {S A F M} (\\operatorname {L N} (X)) + X, \\\\ Z = \\operatorname {C C M} (\\operatorname {L N} (Y)) + Y, \\end{array} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 779, + 467, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{LN}(\\cdot)$ is the LayerNorm [3] layer, $X$ , $Y$ , and $Z$ are the intermediate features. The additional residual learning is employed to stabilize training processes and learn high frequency details to facilitate high-quality image reconstruction.", + "bbox": [ + 76, + 824, + 467, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b45186fcca63868e4a60edf820afb6e75a97b394336a82cb8d6e3e295e00402d.jpg", + "image_caption": [ + "(a) LR input", + "Figure 5. Illustration of learned deep features from the SAFM module. (a) is the LR input. (b)-(e) are features learned by the multi-scale feature generation unit. (f) is the results of aggregating features (b)-(e). (g) denotes the attended output of the SAFM." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 653, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da54769aa655ccede61a63ec8b80831fc3be69faa912f41969987d152a73dc9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 90, + 730, + 138 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2daa8534748e0c099c21c63eaefee922dbe6cbd0a131557e011047cad6bd5b25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 90, + 805, + 138 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f986e3064fc4a1f5a8ba0ffea38aa718d3edd04fd8814d17a25c53e33473dc0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 90, + 883, + 138 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ed4d6368d75c596548d09b452c59bffa36d273764cc35a515a9891acee9cc800.jpg", + "image_caption": [ + "(b) Scale 1", + "(e) Scale 8" + ], + "image_footnote": [], + "bbox": [ + 663, + 151, + 730, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/be287c37315a90c066b9eeeb172012e46548abba9841e76ac836fc2d6ec71b94.jpg", + "image_caption": [ + "(c) Scale 2", + "(f) Aggregated" + ], + "image_footnote": [], + "bbox": [ + 740, + 151, + 805, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/33d97f925d1edad43e51fd429aeacc579f0298e50710996fca4b079542ceb7dd.jpg", + "image_caption": [ + "(d) Scale 4", + "(g) Attended" + ], + "image_footnote": [], + "bbox": [ + 818, + 151, + 883, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 281, + 702, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we perform quantitative and qualitative evaluations with state-of-the-art algorithms to demonstrate the effectiveness of the proposed method.", + "bbox": [ + 498, + 306, + 890, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Dataset and Implementation", + "text_level": 1, + "bbox": [ + 500, + 361, + 754, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. Following previous works [28, 31, 45], we use DIV2K [48] and Flickr2K [32] as the training data and generate LR images by applying the bicubic downscaling to reference HR images. We use five commonly used benchmark datasets including Set5 [4], Set14 [51], B100 [2], Urban100 [19] and Manga109 [38] as test data. We use the peak signal to noise ratio (PSNR) and the structural similarity index (SSIM) to evaluate the quality of the restored images. All PSNR and SSIM values are calculated on the Y channel of images transformed to YCbCr color space.", + "bbox": [ + 498, + 383, + 890, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation details. During the training, the data argumentation is performed on the input patches with random horizontal flips and rotations. In addition, we randomly crop 64 patches of size $64 \\times 64$ pixels from LR images as the basic training inputs. The number of FMM and feature channels is set to 8 and 36, respectively. We use the Adam [25] optimizer with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ to solve the proposed model. The number of iterations is set to 500,000. We set the initial learning rate to $1 \\times 10^{-3}$ and the minimum one to $1 \\times 10^{-5}$ , which is updated by the Cosine Annealing scheme [36]. We use a combination of mean absolute error (MAE) loss and an FFT-based frequency loss function to constrain the model training, which is the same as [45]. All experiments are conducted with the PyTorch framework on an NVIDIA GeForce RTX 3090 GPU. Due to the page limit, we include more comparison results in the supplemental material.", + "bbox": [ + 498, + 536, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 500, + 801, + 874, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quantitative comparisons. To evaluate the performance of the proposed approach, we compare it with state-of-the-art lightweight SR methods on different scaling factors, including SRCNN [10], FSRCNN [11], ESPCN [44], VDSR [24], LapSRN [27], CARN [1], EDSR-baseline [32], IMDN [20],", + "bbox": [ + 498, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13193", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/b3c1623dfa780fe64348dbbcf3b1f053f6014fe44744bc8cc5af525b43ffffca.jpg", + "table_caption": [ + "Table 1. Comparisons of efficient SR networks on the commonly used benchmark datasets. All PSNR/SSIM results are calculated on the Y-channel. #Acts represents all elements of the output of convolutional layers. #FLOPs and #Acts are measured corresponding to an HR image of the size ${1280} \\times {720}$ pixels. Red color denotes the best performance. Blanked entries indicate results not reported or not available from previous work." + ], + "table_footnote": [], + "table_body": "
MethodsScale#Params [K]#FLOPs [G]#Acts [M]Set5Set14B100Urban100Manga109
Bicubic×2---33.66/0.929930.24/0.868829.56/0.843126.88/0.840330.80/0.9339
SRCNN [10]57538936.66/0.954232.42/0.906331.36/0.887929.50/0.894635.74/0.9661
FSRCNN [11]1264137.00/0.955832.63/0.908831.53/0.892029.88/0.902036.67/0.9694
ESPCN [44]2152336.83/0.956432.40/0.909631.29/0.891729.48/0.8975-
VDSR [24]6656131,12037.53/0.958733.03/0.912431.90/0.896030.76/0.914037.22/0.9729
LapSRN [27]8133022337.52/0.959033.08/0.913031.80/0.895030.41/0.910037.27/0.9740
CARN-M [1]4159165537.53/0.958333.26/0.914131.92/0.896031.23/0.9193-
CARN [1]1,59222352237.76/0.959033.52/0.916632.09/0.897831.92/0.9256-
EDSR-baseline [32]1,37031656337.99/0.960433.57/0.917532.16/0.899431.98/0.927238.54/0.9769
IMDN [20]69415942338.00/0.960533.63/0.917732.19/0.899632.17/0.928338.88/0.9774
PAN [54]2617167738.00/0.960533.59/0.918132.18/0.899732.01/0.927338.70/0.9773
LAPAR-A [28]54817165638.01/0.960533.62/0.918332.19/0.899932.10/0.928338.67/0.9772
ECBSR-M16C64 [52]59613725237.90/0.961533.34/0.917832.10/0.901831.71/0.9250-
SMSR [49]985132-38.00/0.960133.64/0.917932.17/0.899032.19/0.928438.76/0.9771
ShuffleMixer [45]3949183238.01/0.960633.63/0.918032.17/0.899531.89/0.925738.83/0.9774
SAFMN (Ours)2285229938.00/0.960533.54/0.917732.16/0.899531.84/0.925638.71/0.9771
Bicubic×3---30.39/0.868227.55/0.774227.21/0.738524.46/0.734926.95/0.8556
SRCNN [10]57538932.75/0.909029.28/0.820928.41/0.786326.24/0.798930.59/0.9107
FSRCNN [11]1251933.16/0.914029.43/0.824228.53/0.791026.43/0.808030.98/0.9212
VDSR [24]6656131,12033.66/0.921329.77/0.831428.82/0.797627.14/0.827932.01/0.9310
CARN-M [1]4154632733.99/0.923630.08/0.836728.91/0.800027.55/0.8385-
CARN [1]1,59211926834.29/0.925530.29/0.840729.06/0.803428.06/0.8493-
EDSR-baseline [32]1,55516028534.37/0.927030.28/0.841729.09/0.805228.15/0.852733.45/0.9439
IMDN [20]7037219034.36/0.927030.32/0.841729.09/0.804628.17/0.851933.61/0.9445
PAN [54]2613934034.40/0.927130.36/0.842329.11/0.805028.11/0.851133.61/0.9448
LAPAR-A [28]59411450534.36/0.926730.34/0.842129.11/0.805428.15/0.852333.51/0.9441
SMSR [49]99368-34.40/0.927030.33/0.841229.10/0.805028.25/0.853633.68/0.9445
ShuffleMixer [45]4154340434.40/0.927230.37/0.842329.12/0.805128.08/0.849833.69/0.9448
SAFMN (Ours)2332313434.34/0.926730.33/0.841829.08/0.804827.95/0.847433.52/0.9437
Bicubic×4---28.42/0.810426.00/0.702725.96/0.667523.14/0.657724.89/0.7866
SRCNN [10]57538930.48/0.862827.49/0.750326.90/0.710124.52/0.722127.66/0.8505
FSRCNN [11]1251130.71/0.865727.59/0.753526.98/0.715024.62/0.728027.90/0.8517
ESPCN [44]251630.52/0.869727.42/0.760626.87/0.721624.39/0.7241-
VDSR [24]6656131,12031.35/0.883828.01/0.767427.29/0.725125.18/0.752428.83/0.8809
LapSRN [27]81314926431.54/0.885028.19/0.772027.32/0.728025.21/0.756029.09/0.8845
CARN-M [1]4153322731.92/0.890328.42/0.776227.44/0.730425.62/0.7694-
CARN [1]1,5929119432.13/0.893728.60/0.780627.58/0.734926.07/0.7837-
EDSR-baseline [32]1,51811420232.09/0.893828.58/0.781327.57/0.735726.04/0.784930.35/0.9067
IMDN [20]7154110832.21/0.894828.58/0.781127.56/0.735326.04/0.783830.45/0.9075
PAN [54]2722823832.13/0.894828.61/0.782227.59/0.736326.11/0.785430.51/0.9095
LAPAR-A [28]6599445232.15/0.894428.61/0.781827.61/0.736626.14/0.787130.42/0.9074
ECBSR-M16C64 [52]603356431.92/0.894628.34/0.781727.48/0.739325.81/0.7773-
SMSR [49]100642-32.12/0.893228.55/0.780827.55/0.735126.11/0.786830.54/0.9085
ShuffleMixer [45]4112826932.21/0.895328.66/0.782727.61/0.736626.08/0.783530.65/0.9093
SAFMN (Ours)240147732.18/0.894828.60/0.781327.58/0.735925.97/0.780930.43/0.9063
", + "bbox": [ + 89, + 152, + 883, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PAN [54], LAPAR [28], ECBSR [52], SMSR [49], and ShuffleMixer [45].", + "bbox": [ + 75, + 699, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The quantitative comparisons on benchmark datasets for the upscaling factors of $\\times 2$ , $\\times 3$ , and $\\times 4$ are reported in Table 1. In addition to PSNR/SSIM metrics, we list the number of parameters (#Params), FLOPs (#FLOPs) and activations (#Acts). We calculate the number of FLOPs and activations with the fvcore1 library under a setting of super-resolving an LR image to $1280 \\times 720$ pixels. Among these metrics, #Params and #Acts are linked to memory consumption, and #FLOPs is related to energy usage. In particular,", + "bbox": [ + 75, + 729, + 470, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acts is a better metric for measuring the efficiency of a model than the number of parameters and FLOPs as suggested in recent works [42, 52, 29].", + "bbox": [ + 500, + 699, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benefiting from the simple yet efficient structure, the proposed SAFMN obtains comparable performance with significantly fewer parameters and memory consumption. For instance, the $\\times 4$ SR results on the B100 dataset reveal that our SAFMN has parameters about $85\\%$ less than the CARN [1], $66\\%$ less than the IMDN [20], and $42\\%$ less than the ShuffleMixer [45]. As for activations, we have $60\\%$ , $29\\%$ and $71\\%$ fewer than them, respectively. While our model has a smaller footprint, we achieve similar performance among these methods. Moreover, we compare", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1We use the fvcore.nn.flop_count_str command to calculate the number of parameters, FLOPs and activations.", + "bbox": [ + 75, + 875, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13194", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f95ca739a69a4cdd8c4742c439b180c6e154276564b0f83236be35c2f9653b4b.jpg", + "image_caption": [ + "img024 from Urban100" + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 377, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f08c0aad803c7c0e0f51133f12e5ab352dc5c33c9205208e4768847194cf5fa.jpg", + "image_caption": [ + "(a) HR patch" + ], + "image_footnote": [], + "bbox": [ + 379, + 88, + 506, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db4592ed12cb81241769af5a9a5d57bb3ed6b21fb98a4af844d09d7f342c54f7.jpg", + "image_caption": [ + "(b) Bicubic" + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 637, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/10d9983022c13b989d58fdb4f73a23d2af3c7e30e1620f9385c84323ae25a849.jpg", + "image_caption": [ + "(c) VDSR [24]" + ], + "image_footnote": [], + "bbox": [ + 640, + 88, + 767, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8bad97d5cf8b0a62f617e00ed6a09482c4f3f0cd18dec9190e3a9f26ee12c9e.jpg", + "image_caption": [ + "(d) ShuffleMixer [45]" + ], + "image_footnote": [], + "bbox": [ + 771, + 88, + 897, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38b79e2d65202dd949a67dd38b8d5f5c749ece62ba1cc2efbbb4e123c4fc758b.jpg", + "image_caption": [ + "(e) LapSRN [27]" + ], + "image_footnote": [], + "bbox": [ + 380, + 170, + 506, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/61fda92f8777e8bc55c4c5e3d5ed4061cfcfc40a46e56dd137576598147b05f5.jpg", + "image_caption": [ + "(f) CARN [1]" + ], + "image_footnote": [], + "bbox": [ + 509, + 170, + 637, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/07e881585ecd68c4795f71db766598c00e10ed81f8162b5299c6b104296572d4.jpg", + "image_caption": [ + "(g) IMDN [20]" + ], + "image_footnote": [], + "bbox": [ + 640, + 170, + 767, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/caf2e5c4e65179ea6cd14438872d6100b484045b74b75eee5f0da56f3c0b7f4e.jpg", + "image_caption": [ + "(h) SAFMN" + ], + "image_footnote": [], + "bbox": [ + 771, + 170, + 897, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6f7c923d0205866c8c827499c5d78cb70f1963b9d5fa0258b78f8a0e3b762da5.jpg", + "image_caption": [ + "img062 from Urban100" + ], + "image_footnote": [], + "bbox": [ + 86, + 261, + 375, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/622f350b20b08c08ca2a81decc4d8fa343a730e114883df4e5ed492fbbfc5e46.jpg", + "image_caption": [ + "(a) HR patch" + ], + "image_footnote": [], + "bbox": [ + 379, + 260, + 506, + 329 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d0e7cb23b4086db9d47a77daec5596f4c7f7727454712d6e44257ff17d49eb1.jpg", + "image_caption": [ + "(b) Bicubic" + ], + "image_footnote": [], + "bbox": [ + 509, + 261, + 637, + 329 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71e5d281c226a473a50dee59c44baab8cf287c14f7834c54a3574e144061bd74.jpg", + "image_caption": [ + "(c) VDSR [24]" + ], + "image_footnote": [], + "bbox": [ + 640, + 260, + 767, + 329 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1176d288c6656b23b12f713ee7c43bb1e3da66961ba56ca3367783ba9e10710.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 260, + 897, + 329 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8820d156b0c82fae48dce421a21ce6c3b903332c6ed019aba3c797762d567a2b.jpg", + "image_caption": [ + "(e) LapSRN [27]" + ], + "image_footnote": [], + "bbox": [ + 379, + 342, + 506, + 409 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/274b0fa28ba294b772474b88dd130cf5ac21656b8ccc4c84d4d7b0697d5cb9a0.jpg", + "image_caption": [ + "(f) CARN [1]" + ], + "image_footnote": [], + "bbox": [ + 509, + 342, + 637, + 409 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42eaf61c08a0c8515e656227235af5d0c0aae50f5290e2b5d01e5327c19f0d4d.jpg", + "image_caption": [ + "(g) IMDN [20]" + ], + "image_footnote": [], + "bbox": [ + 640, + 342, + 767, + 409 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e6bf4e654fd4d339ce58f0610f926a38a2a5efc1c0f7320057a366f9bf720bd.jpg", + "image_caption": [ + "(d) ShuffleMixer [45]", + "(h) SAFMN" + ], + "image_footnote": [], + "bbox": [ + 769, + 342, + 897, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1adef27a5c285f442aee31378aa9e76a80203eb26cb485480b46fdcaf3e8174.jpg", + "image_caption": [ + "img098 from Urban100" + ], + "image_footnote": [], + "bbox": [ + 86, + 433, + 375, + 579 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6cd63aeac1fe8ca0610913d9ff06eef59ca503f0fd4a8dc38ff024de7ef7ac4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 433, + 506, + 501 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29d9f88e24fcfa660b6e81d81fe1a59f1c1e0a15cf7b33f1206d3b9c148b024b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 433, + 637, + 501 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/102ec09a11e96e2616ea61933e51d6e081ef910c421dfc88171f190ab35c7e78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 433, + 767, + 501 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/695a38c4790475a345b79fe11b9817bf3733219f2d4758432aec88d49de4d9d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 433, + 897, + 501 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2b6dca1ce0e7511b68bd3f0dfdbd518a73e55e4cd26423a7344461ed0411231.jpg", + "image_caption": [ + "(a) HR patch", + "(e) LapSRN [27]", + "Figure 6. Visual comparisons for $\\times 3$ SR on the Urban100 dataset. The proposed method recovers the image with clearer structures." + ], + "image_footnote": [], + "bbox": [ + 380, + 513, + 506, + 580 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/220e43d6dd4197ddb818451e4b9b2d69eb8d39156832abf7b85deee4f024ca71.jpg", + "image_caption": [ + "(b) Bicubic", + "(f) CARN [1]" + ], + "image_footnote": [], + "bbox": [ + 509, + 513, + 637, + 580 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef2a68857cb59fd7ab26ce8c84aa8002da778f481a48fc8113fe9cc22b435fdb.jpg", + "image_caption": [ + "(c) VDSR [24]", + "(g) IMDN [20]" + ], + "image_footnote": [], + "bbox": [ + 640, + 513, + 767, + 580 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6ea036299f59c8afbec5ef5d4e7807642ce93d1b1048ade568b17cd1dfc3c03.jpg", + "image_caption": [ + "(d) ShuffleMixer [45]", + "(h) SAFMN" + ], + "image_footnote": [], + "bbox": [ + 769, + 513, + 897, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the reconstruction accuracy, FLOPs and parameters on the $\\times 2$ Set5 dataset in Figure 1. The proposed SAFMN model achieves a favorable trade-off between model complexity and reconstruction performance.", + "bbox": [ + 75, + 633, + 468, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparisons. In addition to the quantitative evaluations, we provide qualitative comparisons of the proposed method. Figure 6 shows the visual comparisons on the Urban100 dataset for $\\times 3$ SR. Our approach generates parallel straight lines and grid patterns more accurately than the listed methods. These results also demonstrate the effectiveness of our method for adaptive feature modulation by exploiting non-local feature interactions.", + "bbox": [ + 75, + 699, + 468, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Memory and running time comparisons. To fully examine the efficiency of our proposed method, we further evaluate our method against five representative ones, including CARN-M [1], CARN [1], EDSR-baseline [32], IMDN [20], and LAPAR-A [28] on $\\times 4$ SR in terms of", + "bbox": [ + 75, + 825, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the GPU memory consumption (#GPU Mem.) and running time (#Avg. Time). The maximum GPU memory consumption is recorded during the inference. The running time is averaged on 50 test images with a resolution of $320 \\times 180$ pixels. We show the memory and running time comparisons in Table 2, our method achieves a clear improvement over other state-of-the-art methods. By using the spatially-adaptive feature modulation layer and the compact channel mixer, the GPU consumption of our SAFMN is only $10\\%$ of the CARN series and $4\\%$ of LAPAR-A; the running time is nearly twice as fast as other evaluated methods, except for IMDN. Compared to IMDN [20], our method has a similar running time speed while significantly reducing memory usage. Tables 1 and 2 show that the proposed model achieves a favorable trade-off in terms of inference speed, model complexity and reconstruction performance against state-of-the-art methods.", + "bbox": [ + 496, + 633, + 893, + 890 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13195", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4841e3cdda00c78a2ceca854042bccb24c4b56be35aa3c1476f9193439e4600e.jpg", + "table_caption": [ + "Table 2. Memory and running time comparisons on $\\times 4$ SR. #GPU Mem. denotes the maximum GPU memory consumption during the inference phase, derived with the Pytorch torch.cuda.max_memory_allocated() function. #Avg. Time is the average running time on 50 LR images of $320 \\times 180$ pixels." + ], + "table_footnote": [], + "table_body": "
Methods#GPU Mem. [M]#Avg.Time [ms]
CARN-M [1]680.8417.85
CARN [1]689.8318.90
EDSR-baseline [32]486.5819.81
IMDN [20]203.4410.22
LAPAR-A [28]1811.4724.91
SAFMN65.2610.71
", + "bbox": [ + 83, + 162, + 460, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 76, + 273, + 302, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further conduct extensive ablation studies to better understand and evaluate each component in the proposed SAFMN. For fair comparisons with the designed baselines, we implement all experiments based on $\\times 4$ SAFMN and train them with the same setting. The experimental results in Table 3 are measured on DIV2K-val [48] and Manga109 [38] datasets.", + "bbox": [ + 75, + 297, + 467, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the SAFM. To demonstrate the effect of the spatially-adaptive feature modulation, we first remove this module for comparison. Without it, the PSNR values will drop by 0.17dB (30.26 vs. 30.43) and 0.34dB (30.09 vs. 30.43) on the DIV2K-val and Manga109 datasets, respectively. These results show the importance of the SAFM. Therefore, we further investigate this module to figure out why it works.", + "bbox": [ + 75, + 404, + 467, + 523 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Feature modulation. The feature modulation mechanism is introduced to enable the network with adaptive properties. Without this operation, the baseline model is lowered by 0.11dB on the Manga109 dataset.", + "- Multi-scale representation. Here, \"w/o MR\" in Table 3 denotes that we directly use a depth-wise convolution with a kernel size of $3 \\times 3$ pixels to extract spatial information and do not employ multi-scale features. A noticeable performance drop of 0.13dB on the Manga109 dataset is observed without these multiscale features. Moreover, we apply the adaptive max pooling over the input features in this module to build feature pyramids. Compared to using adaptive average pooling or nearest interpolation, adaptive max pooling allows the model to detect discriminative features, resulting in better reconstruction results.", + "- Feature aggregation. We use a $1 \\times 1$ convolution to aggregate the multi-scale features on the channel dimension. Combined with the modulation mechanism, it brings a PSNR improvement of 0.14dB on the Manga109 dataset, which proves the necessity of aggregating the multi-scale features." + ], + "bbox": [ + 96, + 527, + 467, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without all three components mentioned before, it represents that only a $3 \\times 3$ depth-wise convolution is", + "bbox": [ + 107, + 869, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "used to encode the spatial information and leads to a PSNR reduction of 0.12dB and 0.2dB on the DIV2Kval and Manga109 datasets, respectively. This performance drop suggests that the spatially-adaptive modulation based on the multi-scale feature representation effectively boosts SR reconstruction performance.", + "bbox": [ + 529, + 90, + 890, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- GELU function. Here, we use the GELU [17] function to normalize the modulation map. Table 3 shows that better experimental results can be achieved with GELU than with Sigmoid or without GELU, mainly because it weights the input features by their percentile and allows better-activated features.", + "bbox": [ + 517, + 184, + 890, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we use the local attribution map (LAM) [13] and diffusion index (DI) [13] to illustrate the feature interaction range. Figure 2 clearly shows that the proposed model can utilise more feature information than previous CNN-based efficient SR methods. The more information used, the better the reconstruction performance in general.", + "bbox": [ + 496, + 281, + 890, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above analysis shows that benefiting from the multiscale feature representation, our proposed SAFM can effectively exploit non-local feature interactions with small memory and computational costs.", + "bbox": [ + 496, + 371, + 890, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the CCM. Compared with the original FMBCConv [47], the main change made by CCM is removing the SE [18] block. As shown in Table 3, using SE blocks, its performance is almost the same as without them. This result is mainly caused by the SAFM block allows adaptability in channel and spatial dimensions. Hence, we do not use the SE blocks for saving parameters. We next conduct a series of ablations to verify that CCM can effectively encode local contextual information and perform channel mixing. Without it, the model only achieves the accuracy of 29.69dB and 28.49dB on DIV2K-val and Manga109 datasets, proving the indispensability and locality modelling ability of this part. We then change CCM to other channel mixers commonly used in ViT architectures: channel MLP [12] or inverted residual block [43]. When the channel MLP is adopted for the channel mixer, there is a significant performance drop of 0.63dB (29.80 vs. 30.43) on the Manga109 dataset. This performance reduction results from a lack of capacity for modeling local contextual information. For the inverted residual block, the performance is highly similar to that of CCM, but the corresponding #Acts is increased by nearly 33M, which means more memory consumption and slower inference time. Thus, we use the CCM as the default channel mixer.", + "bbox": [ + 496, + 431, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of the LayerNorm layer. Since we use the elementwise product in the SAFM module, it will result in abnormal gradient values and unstable model training, as shown in Figure 7(a). It is, therefore, necessary to normalize the input features. Here, we perform normalization with the LayerNorm [3] layer. To verify this assumption, we first remove the LayerNorm layer. Table 3 and Figure 7(a) show", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13196", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation for SAFMN on DIV2K-val and Manga109 datasets. SAFMN with a scaling factor of $\\times 4$ is utilized as the baseline for ablation studies. The PSNR/SSIM values on benchmarks are reported. “ $\\mathrm{A} \\rightarrow \\mathrm{B}$ ” is to replace A with B. “None” means to remove the operation. “Ir” denotes the learning rate. “FBN” is the abbreviation of the Frozen BatchNorm. “ $\\mathrm{L}_2$ normalization” represents that the inputs are normalized by $\\mathrm{L}_2$ -norm over the channel dimension. “FM”, “MR”, and “FA” are the abbreviations corresponding to feature modulation, multi-scale representation, and feature aggregation. * indicates that the corresponding results are obtained before the training collapse. The number of parameters, FLOPs and Acts are calculated in the same way as in Table 1.", + "bbox": [ + 75, + 85, + 893, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b237df714f76606e6cca04b26ef5f99c204623ba60354e3aaf0dc35f15d3b3ec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AblationVariant#Params [K]#FLOPs [G]#Acts [M]DIV2K_valManga109
Baseline-239.5213.5676.7030.43/0.837230.43/0.9063
Main moduleSAFM → None225.4112.9054.6130.26/0.833030.09/0.9018
CCM → None30.721.6126.9329.69/0.819328.49/0.8756
SAFM(a): w/o FM239.5213.5676.7030.36/0.835730.32/0.9048
(b): w/o MR239.5213.6487.7830.34/0.835030.30/0.9047
(c): w/o FA228.8612.9660.1130.36/0.835530.29/0.9049
(a) + (b)239.5213.6487.7830.32/0.834530.24/0.9038
(a) + (c)228.8612.9660.1130.34/0.835130.28/0.9043
(a) + (b) + (c)228.8613.0571.1930.31/0.834430.23/0.9036
AdaptiveMaxPool → AdaptiveAvgPool239.5213.5676.7030.40/0.836430.41/0.9061
AdaptiveMaxPool → Nearest interpolate239.5213.5676.7030.36/0.835430.31/0.9048
GELU → None239.5213.5676.7030.40/0.836630.37/0.9058
GELU → Sigmoid239.5213.5676.7030.36/0.835530.29/0.9044
CCMw/ SE260.9813.5976.7030.39/0.836030.46/0.9067
CCM → Channel MLP73.634.0076.7030.17/0.831329.80/0.8980
CCM → Inverted residual block245.2813.85110.0030.43/0.837330.43/0.9064
NormalizationLN → None, lr=1 × 10-3*238.3713.5576.7030.29/0.834030.04/0.9014
LN → None, lr=1 × 10-4238.3713.5576.7030.15/0.830629.74/0.8970
LN → BN239.5213.7276.7030.28/0.835430.05/0.9029
LN → FBN*238.3713.5576.7030.30/0.834330.15/0.9028
LN → L2 normalization238.3713.5576.7030.37/0.835830.31/0.9049
", + "bbox": [ + 86, + 174, + 883, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c3078570a3532d142635edd2c3b35e19e0f056ff7d1e04e9402978aa7aa07a5.jpg", + "image_caption": [ + "(a) Effect of LayerNorm", + "(b) Comparison of different normalizations" + ], + "image_footnote": [], + "bbox": [ + 78, + 492, + 267, + 599 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44de5793af16a9eb02ab82b433a0a47f0f1e7db89e3a7e16b681ced063f6ee55.jpg", + "image_caption": [ + "Figure 7. Smoothed training loss curves with or without normalization. Figure (a) shows LayerNorm stabilizes model training and converges better than without it. Figure (b) compares the effects of different normalization approaches on model training." + ], + "image_footnote": [], + "bbox": [ + 272, + 492, + 457, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that without it, the model suffers a training crash at a large learning rate and does not converge well at small ones with its PSNR of only 30.15dB and 29.74dB on the DIV2K-val and Manga109 datasets, respectively. Next, we compare the LayerNorm with three representative normalization methods, including BatchNorm [22], Frozen BatchNorm (FBN) [22], and $\\mathrm{L}_2$ normalization in Figure 7(b). The results in Table 3 and Figure 7(b) demonstrate that the BN family decrease the PSNR/SSIM values by a large margin, and the FBN even does not guarantee a stable model training process. Although the $\\mathrm{L}_2$ normalization allows the model to be trained successfully, its performance is not as good as that of using LayerNorm. Thus, the LayerNorm layer is set as default for the proposed SAFMN.", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7a66f8b5173635210b952c85f5925aa2cd715040f0b5cac9381bbef4cd12e31e.jpg", + "table_caption": [ + "Table 4. Effect of the number of FMMs on performance." + ], + "table_footnote": [], + "table_body": "
#N#Params [K]#FLOPs [G]#Acts [M]DIV2K_valManga109
4128.067.2540.7730.21/0.831929.86/0.8988
6183.7910.4058.6430.32/0.834730.18/0.9033
8239.5213.5676.7030.43/0.837230.43/0.9063
12350.9719.86113.0030.51/0.838930.71/0.9096
", + "bbox": [ + 504, + 498, + 885, + 553 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7fd168afea0491e1d2db3be921c13548ec89d4ff62646b063865ee38e9fc1285.jpg", + "table_caption": [ + "Table 5. GPU runtime of meta-operations in the SAFM layer." + ], + "table_footnote": [], + "table_body": "
Meta-operationsLayerNormChannel SplittingAdaptive MaxPoolNearest InterpolationElement-wise ProductElement-wise Addition
Runtime [us]4651338524156193
", + "bbox": [ + 504, + 573, + 885, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the number of FMMs. The proposed network contains several FMMs. We evaluate the effect of the number of FMMs by setting the number #N from 4 to 12. Table 4 shows that the performance gradually improves as #N increases. To achieve a favorable trade-off between model complexity and reconstruction performance, we choose 8 blocks as the default setting.", + "bbox": [ + 496, + 598, + 890, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Relations to lightweight ViT models. To alleviate the quadratic complexity of vanilla ViT [12], MobileViT [39], Mobile-Former [6], and EdgeNeXt [37] adopt a hybrid CNN-Transformer design, which replaces the shallow transformer blocks with convolutional layers to avoid using self-attention on high-resolution representations, and efficiently combines the transformer blocks with the modified mobilenet blocks to perform feature interactions in the later stages. Another alternative [50, 14] is to incorporate concepts from the transformer into convolutional modules. For example, MetaFormer [50] and VAN [14] use pooling operation or large kernel attention to replace the heavy-weight self-attention for contextual information aggregation. Com", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13197", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pared to these approaches, we develop an efficient feature modulation mechanism upon a grouped multi-scale representation to learn feature dependencies. Benefiting from the multi-head paradigm, the proposed model can exploit multiscale contexts across different heads for adaptive feature modulation while maintaining a low computational cost.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations. Although our SAFMN has better overall performance than the evaluated methods [1, 20, 49, 28, 45], its running time is relatively slow. As shown in Table 5, we count the inference time of different commonly used tensor operations with the torch.profiler function, and the results show that the utilized LayerNorm layers and residual connections lead to increased runtime. The running time of SAFMN can be further accelerated by optimizing the implementation of these components. We will investigate these in the future work. Another limitation is that the PSNR performance of SAFMN is suboptimal, especially on the Urban100 dataset, which is mainly due to the fact that we perform the feature modulation without considering the integration of local features.", + "bbox": [ + 75, + 181, + 472, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 411, + 194, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a simple yet efficient CNN model to solve the efficient image super-resolution problem. Our SAFMN exploits non-local feature interactions upon a multi-scale feature representation. To complement the local contextual information, we further develop a compact convolutional channel mixer to encode spatially local context and conduct channel mixing simultaneously. We both qualitatively and quantitatively evaluate the proposed method on commonly used benchmarks. Experimental results show that the proposed SAFMN model is more efficient than state-of-the-art methods while achieving competitive performance.", + "bbox": [ + 75, + 436, + 472, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work has been partly supported by the National Key R&D Program of China (No. 2018AAA0102001), the National Natural Science Foundation of China (Nos. U22B2049, 62272233, 61922043, 61925204), and the Fundamental Research Funds for the Central Universities (Nos. 30922010910, 30920041109).", + "bbox": [ + 75, + 637, + 468, + 728 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 744, + 173, + 760 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, 2018. 1, 2, 4, 5, 6, 7, 9", + "[2] Pablo Arbeláez, Michael Maire, Charless C. Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. PAMI, 33(5):898-916, 2011. 4", + "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4, 7" + ], + "bbox": [ + 84, + 770, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 1, 4", + "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 1", + "[6] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In CVPR, 2022.8", + "[7] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2021. 1, 3", + "[8] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In CVPR, 2021. 2", + "[9] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 3", + "[10] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. PAMI, 38(2):295-307, 2016. 2, 4, 5", + "[11] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In ECCV, 2016. 1, 2, 4, 5", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 7, 8", + "[13] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In CVPR, 2021. 2, 7", + "[14] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 3, 8", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2", + "[16] Zibin He, Tao Dai, Jian Lu, Yong Jiang, and Shu-Tao Xia. Faked: Feature-affinity based knowledge distillation for efficient image super-resolution. In ICIP, 2020. 1, 3", + "[17] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units. arXiv preprint arXiv:1606.08415, 2016. 4, 7", + "[18] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 7", + "[19] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 4", + "[20] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multistillation network. In ACM MM, 2019. 1, 2, 4, 5, 6, 7, 9" + ], + "bbox": [ + 501, + 92, + 893, + 897 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13198", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Andrey Ignatov, Radu Timofte, Maurizio Denna, and et al. Efficient and accurate quantized image super-resolution on Mobile NPUs, Mobile AI & AIM 2022 challenge: Report. In ECCV Workshops, 2022. 2, 3", + "[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2, 8", + "[23] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, 2016. 1", + "[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2, 4, 5, 6", + "[25] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 4", + "[26] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPR Workshops, 2022. 1, 2", + "[27] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017. 4, 5, 6", + "[28] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. LAPAR: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. In NeurIPS, 2020. 4, 5, 6, 7, 9", + "[29] Yawei Li, Kai Zhang, Luc Van Gool, Radu Timofte, et al. NTIRE 2022 challenge on efficient super-resolution: Methods and results. In CVPR Workshops, 2022. 3, 5", + "[30] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In CVPR Workshops, 2022. 1, 2", + "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In ICCV Workshops, 2021. 1, 2, 3, 4", + "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPR Workshops, 2017. 1, 2, 4, 5, 6, 7", + "[33] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV Workshops, 2020. 1, 2", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2", + "[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In CVPR, 2022. 3", + "[36] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 4", + "[37] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: Efficiently amalgamated cnn-transformer architecture for mobile vision applications. In ECCV Workshops, 2022. 8" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yusuke Matsui, Kota Ito, Yuji Aramaki, Toshihiko Yasasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. arXiv preprint arXiv:1510.04389, 2015. 4, 7", + "[39] Sachin Mehta and Mohammad Rastegari. MobileViT: Lightweight, general-purpose, and mobile-friendly vision transformer. In ICLR, 2022. 3, 8", + "[40] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In CVPR, 2021. 2", + "[41] Pablo Navarrete Michelini, Yunhua Lu, and Xingqun Jiang. edge-SR: Super-resolution for the masses. In WACV, 2022. 1, 2", + "[42] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. In CVPR, 2020. 5", + "[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 7", + "[44] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 1, 2, 3, 4, 5", + "[45] Long Sun, Jinshan Pan, and Jinhui Tang. ShuffleMixer: An efficient convnet for image super-resolution. In NeurIPS, 2022. 1, 2, 4, 5, 6, 9", + "[46] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. CVPR, 2017. 1, 2", + "[47] Mingxing Tan and Quoc Le. EfficientNetV2: Smaller models and faster training. In ICML, 2021. 2, 4, 7", + "[48] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, Lei Zhang, et al. NTIRE 2017 challenge on single image super-resolution: Methods and results. In CVPR Workshops, 2017. 4, 7", + "[49] Longguang Wang, Xiaoyu Dong, Yingqian Wang, Xinyi Ying, Zaiping Lin, Wei An, and Yulan Guo. Exploring sparsity in image super-resolution for efficient inference. In CVPR, 2021. 5, 9", + "[50] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 8", + "[51] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces*, 2012. 4", + "[52] Xindong Zhang, Hui Zeng, and Lei Zhang. Edge-oriented convolution block for real-time super resolution on mobile devices. In ACM MM, 2021. 1, 2, 3, 5", + "[53] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 2", + "[54] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCV Workshops, 2020. 1, 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 873 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13199", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/8203ecb7-a9b9-46de-8112-1277f4eeac63_model.json b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/8203ecb7-a9b9-46de-8112-1277f4eeac63_model.json new file mode 100644 index 0000000000000000000000000000000000000000..15de3629d2bce72b3db9617a4d2e3e4a04809326 --- /dev/null +++ b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/8203ecb7-a9b9-46de-8112-1277f4eeac63_model.json @@ -0,0 +1,2893 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.182, + 0.713, + 0.199 + ], + "angle": 0, + "content": "Long Sun, Jiangxin Dong, Jinhui Tang, and Jinshan Pan*" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.199, + 0.856, + 0.218 + ], + "angle": 0, + "content": "School of Computer Science and Engineering, Nanjing University of Science and Technology" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.283, + 0.474, + 0.571 + ], + "angle": 0, + "content": "Although deep learning-based solutions have achieved impressive reconstruction performance in image superresolution (SR), these models are generally large, with complex architectures, making them incompatible with low-power devices with many computational and memory constraints. To overcome these challenges, we propose a spatially-adaptive feature modulation (SAFM) mechanism for efficient SR design. In detail, the SAFM layer uses independent computations to learn multi-scale feature representations and aggregates these features for dynamic spatial modulation. As the SAFM prioritizes exploiting non-local feature dependencies, we further introduce a convolutional channel mixer (CCM) to encode local contextual information and mix channels simultaneously. Extensive experimental results show that the proposed method is \\(3 \\times\\) smaller than state-of-the-art efficient SR methods, e.g., IMDN, and yields comparable performance with much less memory usage. Our source codes and pre-trained models are available at: https://github.com/sunny2109/SAFMN." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.581, + 0.21, + 0.597 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Single image super-resolution (SISR) aims to restore a high-resolution (HR) image from its low-resolution (LR) counterpart by recovering lost details. This longstanding and challenging task has recently attracted much attention due to the rapid development of streaming media or high-definition devices. As these scenarios are usually resource-constrained, it is of great interest to develop an efficient and effective SR method to estimate HR images for better visual display on these platforms or products." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Deep learning-based SR methods have achieved significant performance improvements with the great evolution of hardware technologies, as we can use large amounts of data to train much larger or deeper neural networks for image SR [32, 53, 31, 5]. For example, RCAN [53] is a representative CNN-based image SR network with 15.59M parameters and reaching a depth of over 400 layers. One of the most significant drawbacks of these large models is that they require high computational costs, which makes them" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.254, + 0.887, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.467, + 0.894, + 0.537 + ], + "angle": 0, + "content": "Figure 1. Model complexity and performance comparison between our proposed SAFMN model and other lightweight methods on Set5 [4] for \\(\\times 2\\) SR. Circle sizes indicate the number of parameters. The proposed method achieves a better trade-off between model complexity and reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.893, + 0.702 + ], + "angle": 0, + "content": "challenging to deploy. Moreover, recent visual transformers (ViTs) [12, 31, 5] outperform convolutional neural networks (CNNs) in low-level vision tasks, and their results demonstrate that exploring non-local feature interactions is essential for high-quality reconstruction. But existing self-attention mechanisms are computationally expensive and unfriendly to efficient SR design. This, therefore, motivates us to develop a lightweight yet effective model for real-world applications of image super-resolution by integrating the principles of convolution and self-attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To reduce the heavy computational burden, various methods, including efficient module design [11, 41, 1, 20, 26, 33, 44, 45, 54, 30], knowledge distillation [16], neural architecture search [7], and structural re-parameterization [52], are trying to improve the efficiency of SR algorithms. Among these efficient SR models, one direction is to reduce model parameters or complexity (FLOPs). Lightweight strategies like recursive manner [23, 46], parameter sharing [1], and spare convolutions [1, 45, 30] are adopted. Although these approaches certainly reduce the model size, they usually compensate for the performance drop caused by shared recursive modules or sparse convolutions by increasing the depth or width" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13190" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "of the model, which affects the inference efficiency when performing SR reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.287 + ], + "angle": 0, + "content": "Another direction is to accelerate the inference time. The post-upsampling [11, 44] is an important replacement for the pre-defined input [10, 24], which significantly speeds up the runtime. Model quantization [21] effectively accelerates latency and reduces energy consumption, particularly when deploying algorithms in edge-devices. Structural re-parameterization [52, 8] improves the speed of a well-trained model in the inference stage. These methods enjoy fast running time but poor reconstruction performance. Consequently, there is still room for a better trade-off between model efficiency and reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.468, + 0.514 + ], + "angle": 0, + "content": "To address the above-mentioned issues, we design a simple yet effective model by developing a spatially-adaptive feature modulation, namely SAFMN, to realize a favorable trade-off between performance and efficiency. Specifically, we first exploit non-local feature relations to dynamically select representative features (see Figure 2 and 5) by implementing a feature modulation mechanism based on multiscale representations. As the modulation mechanism processes input features from a non-local perspective, there is a requirement to complement local contextual information. To this end, we present a convolutional channel mixer based on the FMBCov [47] to encode local features and mix channels. Taken together, we find that the SAFMN network is able to achieve a better trade-off between SR performance and model complexity, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.515, + 0.468, + 0.543 + ], + "angle": 0, + "content": "The main contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.551, + 0.468, + 0.597 + ], + "angle": 0, + "content": "- We develop an efficient feature modulation mechanism to learn feature dependencies, which absorbs CNN-like efficiency and transformer-like adaptability." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.603, + 0.468, + 0.648 + ], + "angle": 0, + "content": "- We present a compact convolutional channel mixer that simultaneously encodes local contextual information and performs channel mixing." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.655, + 0.468, + 0.715 + ], + "angle": 0, + "content": "- We evaluate the proposed method quantitatively and qualitatively on benchmark datasets, and the results show that our SAFMN achieves a favorable trade-off between accuracy and model complexity." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.551, + 0.468, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.219, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Deep Learning-based Image Super-Resolution. Classical interpolation algorithms, such as linear or bicubic upsampling, create high-resolution images by inserting zeros between adjacent pixels in the low resolution and then using a low-pass filter to preserve the content information of the input image [41]. Unlike these interpolation-based upsamplers, deep learning-based approaches learn a nonlinear mapping between the input image and the target output in an end-to-end training fashion. SRCNN [10] is the first attempt" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.092, + 0.587, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.15, + 0.558, + 0.157 + ], + "angle": 0, + "content": "DI" + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.161, + 0.559, + 0.171 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.092, + 0.66, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.149, + 0.641, + 0.157 + ], + "angle": 0, + "content": "3.2180" + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.161, + 0.634, + 0.171 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.092, + 0.735, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.149, + 0.715, + 0.157 + ], + "angle": 0, + "content": "5.1936" + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.161, + 0.708, + 0.171 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.092, + 0.81, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.149, + 0.792, + 0.157 + ], + "angle": 0, + "content": "4.0629" + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.161, + 0.784, + 0.171 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.092, + 0.888, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.834, + 0.149, + 0.868, + 0.157 + ], + "angle": 0, + "content": "10.4295" + }, + { + "type": "image_caption", + "bbox": [ + 0.845, + 0.161, + 0.859, + 0.17 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.177, + 0.892, + 0.301 + ], + "angle": 0, + "content": "Figure 2. Comparison of local attribution maps (LAMs) [13] and diffusion indices (DIs) [13] between our SAFMN and other efficient SR models. (a) is the LR input, (b)-(e) are the results of CARN [1], EDSR-baseline [32], PAN [54] and SAFMN, respectively. The LAM results denote the importance of each pixel in the input LR image when super-resolving the patch marked with a red box. The DI value reflects the range of involved pixels. A larger DI value means a wider range of attention. The proposed method can exploit more feature information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.892, + 0.702 + ], + "angle": 0, + "content": "to use a convolutional neural network to tackle the image SR problem and achieves a considerable performance gain compared with conventional methods. Since then, many improvements have been proposed. VDSR [24] uses global residual learning [15] to solve the problem of difficulty in training an SR model with deep layers. DRRN [46] integrates the local residual learning and global residual connection to ease the training difficulty and enhance high-frequency details. EDSR [32] further increases the model footprint to 43M, achieving a significant breakthrough in reconstruction performance and showing that the BatchNorm (BN) [22] layer is not necessary for the SR task. RCAN [53] builds a more than 400 layers model based on channel attention and dense connections for accurate SR. NLSA [40] utilizes a non-local sparse attention to enhance the feature modeling capability and reaches state-of-the-art performance. With the successful application of ViT in various high-level vision tasks, image SR also follows this ViT [12] framework, achieving higher reconstruction performance than CNN-based models on extensive public benchmarks. For instance, SwinIR [31] acts as a strong baseline for image restoration tasks based on the Swin Transformer [34]. While these approaches achieve impressive reconstruction performance, the required high computational costs make them challenging to deploy in real-world applications on resource-constrained devices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Efficient Image Super-Resolution. To improve the model efficiency, many CNN-based SR works try to alleviate this issue. FSRCNN [11] and ESPCN [44] utilize a postupsampling manner to reduce the computational burden from the pre-defined inputs significantly. CARN [1] uses group convolutions and a cascading mechanism upon residual networks to improve efficiency. IMDN [20] adopts feature splitting and concatenation operations to progressively aggregate features, and its improved variants [33, 26] won the AIM2020 and NTIRE2022 Efficient SR challenge. ShuffleMixer [45] introduces a large kernel convolution for lightweight SR design. BSRN [30] employs blueprint-separable convolutions to reduce model complexity. Mean" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13191" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.895, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.895, + 0.298 + ], + "angle": 0, + "content": "Figure 3. An overview of the proposed SAFMN. SAFMN first transforms the input LR image into the feature space using a convolutional layer, performs feature extraction using a series of feature mixing modules (FMMs), and then reconstructs these extracted features by an upsampler module. The FMM block is implemented by a spatially-adaptive feature modulation (SAFM) layer, a convolutional channel mixer (CCM) and two skip-connections." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.473, + 0.507 + ], + "angle": 0, + "content": "while, an increasingly popular direction is to compress or accelerate a well-trained deep model through model quantization [21], structural re-parameterization [52] or knowledge distillation [16]. Neural architecture search (NAS) is also commonly used to search a well-constrained architecture for image super-resolution [7]. Note that the efficiency of a deep neural network could be measured in different metrics, including the number of parameters, FLOPs, activations, memory consumption and inference running time [29]. Although the above approaches have been improved in different efficiency aspects, there is still room for a favorable trade-off between reconstruction performance and model efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.522, + 0.251, + 0.539 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.777 + ], + "angle": 0, + "content": "In this section, we present the core components of our proposed model for efficient SISR. As shown in Figure 3, the network consists of the following parts: a shallow convolution, a stacking of feature mixing modules (FMMs), and an upsampler layer. Specifically, we first apply a \\(3 \\times 3\\) convolution layer to transform the input LR image to feature space and generate the shallow feature \\(\\mathbf{F}_0\\). Then, the multiple stacked FMMs are used to generate finer deep features from \\(\\mathbf{F}_0\\) for HR image reconstruction, where an FMM layer has a spatially-adaptive feature modulation (SAFM) layer and a convolutional channel mixer (CCM). To recover the HR target image, we introduce a global residual connection to learn high-frequency details and employ a lightweight upsampling layer for fast reconstruction, which only contains a \\(3 \\times 3\\) convolution and a pixel-shuffle layer [44]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.411, + 0.803 + ], + "angle": 0, + "content": "3.1. Spatially-adaptive Feature Modulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Recent studies [9, 35, 14, 39, 31] suggest that the notable performance of ViTs across diverse tasks stems from their implementation of the key multi-head self-attention (MHSA) mechanism. This mechanism enables the model with the capability of long-range feature interaction and dynamic spatial weighting, both of which contribute to pro" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.31, + 0.88, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.413, + 0.893, + 0.455 + ], + "angle": 0, + "content": "Figure 4. Network architecture of the proposed SAFM module. This module performs spatially-adaptive modulation on the input features. Please see the Section 3.1 for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.568 + ], + "angle": 0, + "content": "ducing promising results. But existing self-attention variants are computationally expensive and unfriendly to efficient SR design. In contrast, standard convolution is an efficient operation, but its performance is limited by the static weights and locality. This observation inspires us to borrow the idea of MHSA to enhance the representation capability of convolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.893, + 0.662 + ], + "angle": 0, + "content": "To introduce the ability of long-range interaction and dynamic modeling into convolution, we follow the multi-head paradigm, using parallel and independent computations that allow each head to process different scale information of the input, and then aggregate these features to generate an attention map for spatially-adaptive feature modulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.895, + 0.83 + ], + "angle": 0, + "content": "The detailed architecture of our proposed SAFM is depicted in Figure 4. Specifically, we first split the normalized input features into four-group components, and feed them into a multi-scale feature generation unit (MFGU), where a \\(3 \\times 3\\) depth-wise convolution processes the first one, and the rest parts are sampled individually by pooling operations. As we desire to select discriminative features towards learning non-local feature interactions, adaptive max pooling operators are applied over the input features to gather information. Given the input feature \\(X\\), this procedure can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.841, + 0.744, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\left[ X _ {0}, X _ {1}, X _ {2}, X _ {3} \\right] = \\operatorname {S p l i t} (X),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.862, + 0.892, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {0} = \\operatorname {D W - C o n v} _ {3 \\times 3} \\left(X _ {0}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.884, + 0.852, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {i} = \\uparrow_ {p} \\left(\\operatorname {D W - C o n v} _ {3 \\times 3} \\left(\\downarrow_ {\\frac {p}{2 ^ {i}}} \\left(X _ {i}\\right)\\right)\\right), 1 \\leq i \\leq 3,\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13192" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.214 + ], + "angle": 0, + "content": "where \\( \\text{Split}(\\cdot) \\) corresponds to the channel split operation, \\( \\text{DW-Conv}_{3\\times 3}(\\cdot) \\) is a \\( 3\\times 3 \\) depth-wise convolution, \\( \\uparrow_p(\\cdot) \\) represents upsampling features at a specific level to the original resolution \\( p \\) via nearest interpolation for fast implementation, and \\( \\downarrow_{\\frac{p}{2^i}} \\) denotes pooling the input features to the size of \\( \\frac{p}{2^i} \\). Figure 5(b)-(e) show that using such a design, we can obtain varying features from different representation subspaces." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.215, + 0.468, + 0.273 + ], + "angle": 0, + "content": "Afterward, we aggregate these extracted short- or long-range features by concatenating them on the channel dimension and performing a \\(1 \\times 1\\) convolution. It can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.284, + 0.469, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\hat {X} = \\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left([ \\hat {X} _ {0}, \\hat {X} _ {1}, \\hat {X} _ {2}, \\hat {X} _ {3} ]\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.468, + 0.424 + ], + "angle": 0, + "content": "where \\(\\operatorname{Concat}(\\cdot)\\) denotes the concatenation operation, and \\(\\operatorname{Conv}_{1\\times 1}(\\cdot)\\) is the \\(1\\times 1\\) convolution. After obtaining the aggregated representation \\(\\hat{X}\\) (see Figure 5(f)), we normalize it through a GELU non-linearity [17] to estimate the attention map and adaptively modulate the input \\(X\\) according to the estimated attention via element-wise product. This process can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.437, + 0.469, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\bar {X} = \\phi (\\hat {X}) \\odot X, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.467, + 0.468, + 0.512 + ], + "angle": 0, + "content": "where \\(\\phi(\\cdot)\\) represents the GELU function and \\(\\odot\\) is the element-wise product. The benefits of those operators are compared and results are shown in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.522, + 0.295, + 0.538 + ], + "angle": 0, + "content": "3.2. Feature Mixing Module" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.545, + 0.468, + 0.726 + ], + "angle": 0, + "content": "As mentioned previously, the SAFM module prioritizes exploiting non-local feature interactions. To further incorporate local contextual information and perform channel mixing simultaneously, we introduce a compact convolutional channel mixer (CCM) based on the FMBCnv [47]. The CCM comprises a \\(3 \\times 3\\) convolution and a \\(1 \\times 1\\) convolution. Within this, the first \\(3 \\times 3\\) convolution encodes the spatially local contexts and doubles the number of channels of the input features for mixing channels; the later \\(1 \\times 1\\) convolution reduces the channels back to the original input dimension. A GELU [17] function is applied to the hidden layer for non-linear mapping." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.726, + 0.468, + 0.77 + ], + "angle": 0, + "content": "We formulate the proposed SAFM and CCM into a unified feature mixing module (FMM) to select representative features. The FMM can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.78, + 0.468, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Y = \\operatorname {S A F M} (\\operatorname {L N} (X)) + X, \\\\ Z = \\operatorname {C C M} (\\operatorname {L N} (Y)) + Y, \\end{array} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.468, + 0.899 + ], + "angle": 0, + "content": "where \\(\\mathrm{LN}(\\cdot)\\) is the LayerNorm [3] layer, \\(X\\), \\(Y\\), and \\(Z\\) are the intermediate features. The additional residual learning is employed to stabilize training processes and learn high frequency details to facilitate high-quality image reconstruction." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.091, + 0.655, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.201, + 0.614, + 0.212 + ], + "angle": 0, + "content": "(a) LR input" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.091, + 0.731, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.142, + 0.723, + 0.151 + ], + "angle": 0, + "content": "(b) Scale 1" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.091, + 0.806, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.142, + 0.8, + 0.151 + ], + "angle": 0, + "content": "(c) Scale 2" + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.091, + 0.884, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.142, + 0.878, + 0.151 + ], + "angle": 0, + "content": "(d) Scale 4" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.152, + 0.731, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.201, + 0.723, + 0.211 + ], + "angle": 0, + "content": "(e) Scale 8" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.152, + 0.806, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.201, + 0.808, + 0.212 + ], + "angle": 0, + "content": "(f) Aggregated" + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.152, + 0.884, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.201, + 0.882, + 0.211 + ], + "angle": 0, + "content": "(g) Attended" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.215, + 0.892, + 0.27 + ], + "angle": 0, + "content": "Figure 5. Illustration of learned deep features from the SAFM module. (a) is the LR input. (b)-(e) are features learned by the multi-scale feature generation unit. (f) is the results of aggregating features (b)-(e). (g) denotes the attended output of the SAFM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.282, + 0.704, + 0.299 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.308, + 0.891, + 0.353 + ], + "angle": 0, + "content": "In this section, we perform quantitative and qualitative evaluations with state-of-the-art algorithms to demonstrate the effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.362, + 0.755, + 0.378 + ], + "angle": 0, + "content": "4.1. Dataset and Implementation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.891, + 0.536 + ], + "angle": 0, + "content": "Datasets. Following previous works [28, 31, 45], we use DIV2K [48] and Flickr2K [32] as the training data and generate LR images by applying the bicubic downscaling to reference HR images. We use five commonly used benchmark datasets including Set5 [4], Set14 [51], B100 [2], Urban100 [19] and Manga109 [38] as test data. We use the peak signal to noise ratio (PSNR) and the structural similarity index (SSIM) to evaluate the quality of the restored images. All PSNR and SSIM values are calculated on the Y channel of images transformed to YCbCr color space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.537, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Implementation details. During the training, the data argumentation is performed on the input patches with random horizontal flips and rotations. In addition, we randomly crop 64 patches of size \\(64 \\times 64\\) pixels from LR images as the basic training inputs. The number of FMM and feature channels is set to 8 and 36, respectively. We use the Adam [25] optimizer with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.99\\) to solve the proposed model. The number of iterations is set to 500,000. We set the initial learning rate to \\(1 \\times 10^{-3}\\) and the minimum one to \\(1 \\times 10^{-5}\\), which is updated by the Cosine Annealing scheme [36]. We use a combination of mean absolute error (MAE) loss and an FFT-based frequency loss function to constrain the model training, which is the same as [45]. All experiments are conducted with the PyTorch framework on an NVIDIA GeForce RTX 3090 GPU. Due to the page limit, we include more comparison results in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.875, + 0.817 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparisons. To evaluate the performance of the proposed approach, we compare it with state-of-the-art lightweight SR methods on different scaling factors, including SRCNN [10], FSRCNN [11], ESPCN [44], VDSR [24], LapSRN [27], CARN [1], EDSR-baseline [32], IMDN [20]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13193" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.087, + 0.895, + 0.144 + ], + "angle": 0, + "content": "Table 1. Comparisons of efficient SR networks on the commonly used benchmark datasets. All PSNR/SSIM results are calculated on the Y-channel. #Acts represents all elements of the output of convolutional layers. #FLOPs and #Acts are measured corresponding to an HR image of the size \\( {1280} \\times {720} \\) pixels. Red color denotes the best performance. Blanked entries indicate results not reported or not available from previous work." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.154, + 0.885, + 0.688 + ], + "angle": 0, + "content": "
MethodsScale#Params [K]#FLOPs [G]#Acts [M]Set5Set14B100Urban100Manga109
Bicubic×2---33.66/0.929930.24/0.868829.56/0.843126.88/0.840330.80/0.9339
SRCNN [10]57538936.66/0.954232.42/0.906331.36/0.887929.50/0.894635.74/0.9661
FSRCNN [11]1264137.00/0.955832.63/0.908831.53/0.892029.88/0.902036.67/0.9694
ESPCN [44]2152336.83/0.956432.40/0.909631.29/0.891729.48/0.8975-
VDSR [24]6656131,12037.53/0.958733.03/0.912431.90/0.896030.76/0.914037.22/0.9729
LapSRN [27]8133022337.52/0.959033.08/0.913031.80/0.895030.41/0.910037.27/0.9740
CARN-M [1]4159165537.53/0.958333.26/0.914131.92/0.896031.23/0.9193-
CARN [1]1,59222352237.76/0.959033.52/0.916632.09/0.897831.92/0.9256-
EDSR-baseline [32]1,37031656337.99/0.960433.57/0.917532.16/0.899431.98/0.927238.54/0.9769
IMDN [20]69415942338.00/0.960533.63/0.917732.19/0.899632.17/0.928338.88/0.9774
PAN [54]2617167738.00/0.960533.59/0.918132.18/0.899732.01/0.927338.70/0.9773
LAPAR-A [28]54817165638.01/0.960533.62/0.918332.19/0.899932.10/0.928338.67/0.9772
ECBSR-M16C64 [52]59613725237.90/0.961533.34/0.917832.10/0.901831.71/0.9250-
SMSR [49]985132-38.00/0.960133.64/0.917932.17/0.899032.19/0.928438.76/0.9771
ShuffleMixer [45]3949183238.01/0.960633.63/0.918032.17/0.899531.89/0.925738.83/0.9774
SAFMN (Ours)2285229938.00/0.960533.54/0.917732.16/0.899531.84/0.925638.71/0.9771
Bicubic×3---30.39/0.868227.55/0.774227.21/0.738524.46/0.734926.95/0.8556
SRCNN [10]57538932.75/0.909029.28/0.820928.41/0.786326.24/0.798930.59/0.9107
FSRCNN [11]1251933.16/0.914029.43/0.824228.53/0.791026.43/0.808030.98/0.9212
VDSR [24]6656131,12033.66/0.921329.77/0.831428.82/0.797627.14/0.827932.01/0.9310
CARN-M [1]4154632733.99/0.923630.08/0.836728.91/0.800027.55/0.8385-
CARN [1]1,59211926834.29/0.925530.29/0.840729.06/0.803428.06/0.8493-
EDSR-baseline [32]1,55516028534.37/0.927030.28/0.841729.09/0.805228.15/0.852733.45/0.9439
IMDN [20]7037219034.36/0.927030.32/0.841729.09/0.804628.17/0.851933.61/0.9445
PAN [54]2613934034.40/0.927130.36/0.842329.11/0.805028.11/0.851133.61/0.9448
LAPAR-A [28]59411450534.36/0.926730.34/0.842129.11/0.805428.15/0.852333.51/0.9441
SMSR [49]99368-34.40/0.927030.33/0.841229.10/0.805028.25/0.853633.68/0.9445
ShuffleMixer [45]4154340434.40/0.927230.37/0.842329.12/0.805128.08/0.849833.69/0.9448
SAFMN (Ours)2332313434.34/0.926730.33/0.841829.08/0.804827.95/0.847433.52/0.9437
Bicubic×4---28.42/0.810426.00/0.702725.96/0.667523.14/0.657724.89/0.7866
SRCNN [10]57538930.48/0.862827.49/0.750326.90/0.710124.52/0.722127.66/0.8505
FSRCNN [11]1251130.71/0.865727.59/0.753526.98/0.715024.62/0.728027.90/0.8517
ESPCN [44]251630.52/0.869727.42/0.760626.87/0.721624.39/0.7241-
VDSR [24]6656131,12031.35/0.883828.01/0.767427.29/0.725125.18/0.752428.83/0.8809
LapSRN [27]81314926431.54/0.885028.19/0.772027.32/0.728025.21/0.756029.09/0.8845
CARN-M [1]4153322731.92/0.890328.42/0.776227.44/0.730425.62/0.7694-
CARN [1]1,5929119432.13/0.893728.60/0.780627.58/0.734926.07/0.7837-
EDSR-baseline [32]1,51811420232.09/0.893828.58/0.781327.57/0.735726.04/0.784930.35/0.9067
IMDN [20]7154110832.21/0.894828.58/0.781127.56/0.735326.04/0.783830.45/0.9075
PAN [54]2722823832.13/0.894828.61/0.782227.59/0.736326.11/0.785430.51/0.9095
LAPAR-A [28]6599445232.15/0.894428.61/0.781827.61/0.736626.14/0.787130.42/0.9074
ECBSR-M16C64 [52]603356431.92/0.894628.34/0.781727.48/0.739325.81/0.7773-
SMSR [49]100642-32.12/0.893228.55/0.780827.55/0.735126.11/0.786830.54/0.9085
ShuffleMixer [45]4112826932.21/0.895328.66/0.782727.61/0.736626.08/0.783530.65/0.9093
SAFMN (Ours)240147732.18/0.894828.60/0.781327.58/0.735925.97/0.780930.43/0.9063
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.73 + ], + "angle": 0, + "content": "PAN [54], LAPAR [28], ECBSR [52], SMSR [49], and ShuffleMixer [45]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.471, + 0.868 + ], + "angle": 0, + "content": "The quantitative comparisons on benchmark datasets for the upscaling factors of \\(\\times 2\\), \\(\\times 3\\), and \\(\\times 4\\) are reported in Table 1. In addition to PSNR/SSIM metrics, we list the number of parameters (#Params), FLOPs (#FLOPs) and activations (#Acts). We calculate the number of FLOPs and activations with the fvcore1 library under a setting of super-resolving an LR image to \\(1280 \\times 720\\) pixels. Among these metrics, #Params and #Acts are linked to memory consumption, and #FLOPs is related to energy usage. In particular," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.7, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Acts is a better metric for measuring the efficiency of a model than the number of parameters and FLOPs as suggested in recent works [42, 52, 29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Benefiting from the simple yet efficient structure, the proposed SAFMN obtains comparable performance with significantly fewer parameters and memory consumption. For instance, the \\(\\times 4\\) SR results on the B100 dataset reveal that our SAFMN has parameters about \\(85\\%\\) less than the CARN [1], \\(66\\%\\) less than the IMDN [20], and \\(42\\%\\) less than the ShuffleMixer [45]. As for activations, we have \\(60\\%\\), \\(29\\%\\) and \\(71\\%\\) fewer than them, respectively. While our model has a smaller footprint, we achieve similar performance among these methods. Moreover, we compare" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.471, + 0.901 + ], + "angle": 0, + "content": "1We use the fvcore.nn.flop_count_str command to calculate the number of parameters, FLOPs and activations." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "13194" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.379, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.239, + 0.296, + 0.251 + ], + "angle": 0, + "content": "img024 from Urban100" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.089, + 0.508, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.159, + 0.478, + 0.171 + ], + "angle": 0, + "content": "(a) HR patch" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.09, + 0.638, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.159, + 0.604, + 0.17 + ], + "angle": 0, + "content": "(b) Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.089, + 0.768, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.159, + 0.744, + 0.171 + ], + "angle": 0, + "content": "(c) VDSR [24]" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.089, + 0.898, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.159, + 0.89, + 0.171 + ], + "angle": 0, + "content": "(d) ShuffleMixer [45]" + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.171, + 0.508, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.239, + 0.489, + 0.251 + ], + "angle": 0, + "content": "(e) LapSRN [27]" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.171, + 0.638, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.239, + 0.611, + 0.251 + ], + "angle": 0, + "content": "(f) CARN [1]" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.171, + 0.768, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.239, + 0.744, + 0.251 + ], + "angle": 0, + "content": "(g) IMDN [20]" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.171, + 0.898, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.239, + 0.865, + 0.25 + ], + "angle": 0, + "content": "(h) SAFMN" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.262, + 0.376, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.411, + 0.296, + 0.423 + ], + "angle": 0, + "content": "img062 from Urban100" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.261, + 0.507, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.332, + 0.478, + 0.342 + ], + "angle": 0, + "content": "(a) HR patch" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.262, + 0.638, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.332, + 0.604, + 0.342 + ], + "angle": 0, + "content": "(b) Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.261, + 0.768, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.332, + 0.743, + 0.342 + ], + "angle": 0, + "content": "(c) VDSR [24]" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.261, + 0.898, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.332, + 0.891, + 0.343 + ], + "angle": 0, + "content": "(d) ShuffleMixer [45]" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.343, + 0.507, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.411, + 0.489, + 0.423 + ], + "angle": 0, + "content": "(e) LapSRN [27]" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.343, + 0.638, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.411, + 0.61, + 0.423 + ], + "angle": 0, + "content": "(f) CARN [1]" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.343, + 0.768, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.411, + 0.744, + 0.423 + ], + "angle": 0, + "content": "(g) IMDN [20]" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.343, + 0.898, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.411, + 0.866, + 0.422 + ], + "angle": 0, + "content": "(h) SAFMN" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.434, + 0.377, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.583, + 0.296, + 0.594 + ], + "angle": 0, + "content": "img098 from Urban100" + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.434, + 0.508, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.503, + 0.478, + 0.514 + ], + "angle": 0, + "content": "(a) HR patch" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.434, + 0.638, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.503, + 0.604, + 0.514 + ], + "angle": 0, + "content": "(b) Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.434, + 0.768, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.503, + 0.743, + 0.514 + ], + "angle": 0, + "content": "(c) VDSR [24]" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.434, + 0.898, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.503, + 0.89, + 0.514 + ], + "angle": 0, + "content": "(d) ShuffleMixer [45]" + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.515, + 0.508, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.583, + 0.488, + 0.595 + ], + "angle": 0, + "content": "(e) LapSRN [27]" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.515, + 0.638, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.583, + 0.61, + 0.595 + ], + "angle": 0, + "content": "(f) CARN [1]" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.515, + 0.768, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.583, + 0.744, + 0.595 + ], + "angle": 0, + "content": "(g) IMDN [20]" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.515, + 0.898, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.583, + 0.867, + 0.593 + ], + "angle": 0, + "content": "(h) SAFMN" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.607, + 0.873, + 0.621 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons for \\(\\times 3\\) SR on the Urban100 dataset. The proposed method recovers the image with clearer structures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.696 + ], + "angle": 0, + "content": "the reconstruction accuracy, FLOPs and parameters on the \\(\\times 2\\) Set5 dataset in Figure 1. The proposed SAFMN model achieves a favorable trade-off between model complexity and reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.821 + ], + "angle": 0, + "content": "Qualitative comparisons. In addition to the quantitative evaluations, we provide qualitative comparisons of the proposed method. Figure 6 shows the visual comparisons on the Urban100 dataset for \\(\\times 3\\) SR. Our approach generates parallel straight lines and grid patterns more accurately than the listed methods. These results also demonstrate the effectiveness of our method for adaptive feature modulation by exploiting non-local feature interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Memory and running time comparisons. To fully examine the efficiency of our proposed method, we further evaluate our method against five representative ones, including CARN-M [1], CARN [1], EDSR-baseline [32], IMDN [20], and LAPAR-A [28] on \\(\\times 4\\) SR in terms of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.895, + 0.891 + ], + "angle": 0, + "content": "the GPU memory consumption (#GPU Mem.) and running time (#Avg. Time). The maximum GPU memory consumption is recorded during the inference. The running time is averaged on 50 test images with a resolution of \\(320 \\times 180\\) pixels. We show the memory and running time comparisons in Table 2, our method achieves a clear improvement over other state-of-the-art methods. By using the spatially-adaptive feature modulation layer and the compact channel mixer, the GPU consumption of our SAFMN is only \\(10\\%\\) of the CARN series and \\(4\\%\\) of LAPAR-A; the running time is nearly twice as fast as other evaluated methods, except for IMDN. Compared to IMDN [20], our method has a similar running time speed while significantly reducing memory usage. Tables 1 and 2 show that the proposed model achieves a favorable trade-off in terms of inference speed, model complexity and reconstruction performance against state-of-the-art methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13195" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.158 + ], + "angle": 0, + "content": "Table 2. Memory and running time comparisons on \\(\\times 4\\) SR. #GPU Mem. denotes the maximum GPU memory consumption during the inference phase, derived with the Pytorch torch.cuda.max_memory_allocated() function. #Avg. Time is the average running time on 50 LR images of \\(320 \\times 180\\) pixels." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.163, + 0.462, + 0.263 + ], + "angle": 0, + "content": "
Methods#GPU Mem. [M]#Avg.Time [ms]
CARN-M [1]680.8417.85
CARN [1]689.8318.90
EDSR-baseline [32]486.5819.81
IMDN [20]203.4410.22
LAPAR-A [28]1811.4724.91
SAFMN65.2610.71
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.274, + 0.303, + 0.291 + ], + "angle": 0, + "content": "5. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.468, + 0.404 + ], + "angle": 0, + "content": "We further conduct extensive ablation studies to better understand and evaluate each component in the proposed SAFMN. For fair comparisons with the designed baselines, we implement all experiments based on \\(\\times 4\\) SAFMN and train them with the same setting. The experimental results in Table 3 are measured on DIV2K-val [48] and Manga109 [38] datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.468, + 0.525 + ], + "angle": 0, + "content": "Effectiveness of the SAFM. To demonstrate the effect of the spatially-adaptive feature modulation, we first remove this module for comparison. Without it, the PSNR values will drop by 0.17dB (30.26 vs. 30.43) and 0.34dB (30.09 vs. 30.43) on the DIV2K-val and Manga109 datasets, respectively. These results show the importance of the SAFM. Therefore, we further investigate this module to figure out why it works." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.528, + 0.468, + 0.587 + ], + "angle": 0, + "content": "- Feature modulation. The feature modulation mechanism is introduced to enable the network with adaptive properties. Without this operation, the baseline model is lowered by 0.11dB on the Manga109 dataset." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.591, + 0.468, + 0.771 + ], + "angle": 0, + "content": "- Multi-scale representation. Here, \"w/o MR\" in Table 3 denotes that we directly use a depth-wise convolution with a kernel size of \\(3 \\times 3\\) pixels to extract spatial information and do not employ multi-scale features. A noticeable performance drop of 0.13dB on the Manga109 dataset is observed without these multiscale features. Moreover, we apply the adaptive max pooling over the input features in this module to build feature pyramids. Compared to using adaptive average pooling or nearest interpolation, adaptive max pooling allows the model to detect discriminative features, resulting in better reconstruction results." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.776, + 0.468, + 0.866 + ], + "angle": 0, + "content": "- Feature aggregation. We use a \\(1 \\times 1\\) convolution to aggregate the multi-scale features on the channel dimension. Combined with the modulation mechanism, it brings a PSNR improvement of 0.14dB on the Manga109 dataset, which proves the necessity of aggregating the multi-scale features." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.528, + 0.468, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Without all three components mentioned before, it represents that only a \\(3 \\times 3\\) depth-wise convolution is" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "used to encode the spatial information and leads to a PSNR reduction of 0.12dB and 0.2dB on the DIV2Kval and Manga109 datasets, respectively. This performance drop suggests that the spatially-adaptive modulation based on the multi-scale feature representation effectively boosts SR reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.185, + 0.892, + 0.274 + ], + "angle": 0, + "content": "- GELU function. Here, we use the GELU [17] function to normalize the modulation map. Table 3 shows that better experimental results can be achieved with GELU than with Sigmoid or without GELU, mainly because it weights the input features by their percentile and allows better-activated features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.372 + ], + "angle": 0, + "content": "In addition, we use the local attribution map (LAM) [13] and diffusion index (DI) [13] to illustrate the feature interaction range. Figure 2 clearly shows that the proposed model can utilise more feature information than previous CNN-based efficient SR methods. The more information used, the better the reconstruction performance in general." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.433 + ], + "angle": 0, + "content": "The above analysis shows that benefiting from the multiscale feature representation, our proposed SAFM can effectively exploit non-local feature interactions with small memory and computational costs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Effectiveness of the CCM. Compared with the original FMBCConv [47], the main change made by CCM is removing the SE [18] block. As shown in Table 3, using SE blocks, its performance is almost the same as without them. This result is mainly caused by the SAFM block allows adaptability in channel and spatial dimensions. Hence, we do not use the SE blocks for saving parameters. We next conduct a series of ablations to verify that CCM can effectively encode local contextual information and perform channel mixing. Without it, the model only achieves the accuracy of 29.69dB and 28.49dB on DIV2K-val and Manga109 datasets, proving the indispensability and locality modelling ability of this part. We then change CCM to other channel mixers commonly used in ViT architectures: channel MLP [12] or inverted residual block [43]. When the channel MLP is adopted for the channel mixer, there is a significant performance drop of 0.63dB (29.80 vs. 30.43) on the Manga109 dataset. This performance reduction results from a lack of capacity for modeling local contextual information. For the inverted residual block, the performance is highly similar to that of CCM, but the corresponding #Acts is increased by nearly 33M, which means more memory consumption and slower inference time. Thus, we use the CCM as the default channel mixer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of the LayerNorm layer. Since we use the elementwise product in the SAFM module, it will result in abnormal gradient values and unstable model training, as shown in Figure 7(a). It is, therefore, necessary to normalize the input features. Here, we perform normalization with the LayerNorm [3] layer. To verify this assumption, we first remove the LayerNorm layer. Table 3 and Figure 7(a) show" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13196" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.087, + 0.895, + 0.173 + ], + "angle": 0, + "content": "Table 3. Ablation for SAFMN on DIV2K-val and Manga109 datasets. SAFMN with a scaling factor of \\(\\times 4\\) is utilized as the baseline for ablation studies. The PSNR/SSIM values on benchmarks are reported. “\\(\\mathrm{A} \\rightarrow \\mathrm{B}\\)” is to replace A with B. “None” means to remove the operation. “Ir” denotes the learning rate. “FBN” is the abbreviation of the Frozen BatchNorm. “\\(\\mathrm{L}_2\\) normalization” represents that the inputs are normalized by \\(\\mathrm{L}_2\\)-norm over the channel dimension. “FM”, “MR”, and “FA” are the abbreviations corresponding to feature modulation, multi-scale representation, and feature aggregation. * indicates that the corresponding results are obtained before the training collapse. The number of parameters, FLOPs and Acts are calculated in the same way as in Table 1." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.175, + 0.885, + 0.475 + ], + "angle": 0, + "content": "
AblationVariant#Params [K]#FLOPs [G]#Acts [M]DIV2K_valManga109
Baseline-239.5213.5676.7030.43/0.837230.43/0.9063
Main moduleSAFM → None225.4112.9054.6130.26/0.833030.09/0.9018
CCM → None30.721.6126.9329.69/0.819328.49/0.8756
SAFM(a): w/o FM239.5213.5676.7030.36/0.835730.32/0.9048
(b): w/o MR239.5213.6487.7830.34/0.835030.30/0.9047
(c): w/o FA228.8612.9660.1130.36/0.835530.29/0.9049
(a) + (b)239.5213.6487.7830.32/0.834530.24/0.9038
(a) + (c)228.8612.9660.1130.34/0.835130.28/0.9043
(a) + (b) + (c)228.8613.0571.1930.31/0.834430.23/0.9036
AdaptiveMaxPool → AdaptiveAvgPool239.5213.5676.7030.40/0.836430.41/0.9061
AdaptiveMaxPool → Nearest interpolate239.5213.5676.7030.36/0.835430.31/0.9048
GELU → None239.5213.5676.7030.40/0.836630.37/0.9058
GELU → Sigmoid239.5213.5676.7030.36/0.835530.29/0.9044
CCMw/ SE260.9813.5976.7030.39/0.836030.46/0.9067
CCM → Channel MLP73.634.0076.7030.17/0.831329.80/0.8980
CCM → Inverted residual block245.2813.85110.0030.43/0.837330.43/0.9064
NormalizationLN → None, lr=1 × 10-3*238.3713.5576.7030.29/0.834030.04/0.9014
LN → None, lr=1 × 10-4238.3713.5576.7030.15/0.830629.74/0.8970
LN → BN239.5213.7276.7030.28/0.835430.05/0.9029
LN → FBN*238.3713.5576.7030.30/0.834330.15/0.9028
LN → L2 normalization238.3713.5576.7030.37/0.835830.31/0.9049
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.493, + 0.268, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.605, + 0.241, + 0.618 + ], + "angle": 0, + "content": "(a) Effect of LayerNorm" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.493, + 0.459, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.605, + 0.483, + 0.619 + ], + "angle": 0, + "content": "(b) Comparison of different normalizations" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.626, + 0.472, + 0.682 + ], + "angle": 0, + "content": "Figure 7. Smoothed training loss curves with or without normalization. Figure (a) shows LayerNorm stabilizes model training and converges better than without it. Figure (b) compares the effects of different normalization approaches on model training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "that without it, the model suffers a training crash at a large learning rate and does not converge well at small ones with its PSNR of only 30.15dB and 29.74dB on the DIV2K-val and Manga109 datasets, respectively. Next, we compare the LayerNorm with three representative normalization methods, including BatchNorm [22], Frozen BatchNorm (FBN) [22], and \\(\\mathrm{L}_2\\) normalization in Figure 7(b). The results in Table 3 and Figure 7(b) demonstrate that the BN family decrease the PSNR/SSIM values by a large margin, and the FBN even does not guarantee a stable model training process. Although the \\(\\mathrm{L}_2\\) normalization allows the model to be trained successfully, its performance is not as good as that of using LayerNorm. Thus, the LayerNorm layer is set as default for the proposed SAFMN." + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.487, + 0.854, + 0.499 + ], + "angle": 0, + "content": "Table 4. Effect of the number of FMMs on performance." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.499, + 0.887, + 0.554 + ], + "angle": 0, + "content": "
#N#Params [K]#FLOPs [G]#Acts [M]DIV2K_valManga109
4128.067.2540.7730.21/0.831929.86/0.8988
6183.7910.4058.6430.32/0.834730.18/0.9033
8239.5213.5676.7030.43/0.837230.43/0.9063
12350.9719.86113.0030.51/0.838930.71/0.9096
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.559, + 0.884, + 0.572 + ], + "angle": 0, + "content": "Table 5. GPU runtime of meta-operations in the SAFM layer." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.574, + 0.887, + 0.591 + ], + "angle": 0, + "content": "
Meta-operationsLayerNormChannel SplittingAdaptive MaxPoolNearest InterpolationElement-wise ProductElement-wise Addition
Runtime [us]4651338524156193
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Effect of the number of FMMs. The proposed network contains several FMMs. We evaluate the effect of the number of FMMs by setting the number #N from 4 to 12. Table 4 shows that the performance gradually improves as #N increases. To achieve a favorable trade-off between model complexity and reconstruction performance, we choose 8 blocks as the default setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Relations to lightweight ViT models. To alleviate the quadratic complexity of vanilla ViT [12], MobileViT [39], Mobile-Former [6], and EdgeNeXt [37] adopt a hybrid CNN-Transformer design, which replaces the shallow transformer blocks with convolutional layers to avoid using self-attention on high-resolution representations, and efficiently combines the transformer blocks with the modified mobilenet blocks to perform feature interactions in the later stages. Another alternative [50, 14] is to incorporate concepts from the transformer into convolutional modules. For example, MetaFormer [50] and VAN [14] use pooling operation or large kernel attention to replace the heavy-weight self-attention for contextual information aggregation. Com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13197" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "pared to these approaches, we develop an efficient feature modulation mechanism upon a grouped multi-scale representation to learn feature dependencies. Benefiting from the multi-head paradigm, the proposed model can exploit multiscale contexts across different heads for adaptive feature modulation while maintaining a low computational cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.473, + 0.395 + ], + "angle": 0, + "content": "Limitations. Although our SAFMN has better overall performance than the evaluated methods [1, 20, 49, 28, 45], its running time is relatively slow. As shown in Table 5, we count the inference time of different commonly used tensor operations with the torch.profiler function, and the results show that the utilized LayerNorm layers and residual connections lead to increased runtime. The running time of SAFMN can be further accelerated by optimizing the implementation of these components. We will investigate these in the future work. Another limitation is that the PSNR performance of SAFMN is suboptimal, especially on the Urban100 dataset, which is mainly due to the fact that we perform the feature modulation without considering the integration of local features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.412, + 0.196, + 0.428 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.473, + 0.62 + ], + "angle": 0, + "content": "In this paper, we propose a simple yet efficient CNN model to solve the efficient image super-resolution problem. Our SAFMN exploits non-local feature interactions upon a multi-scale feature representation. To complement the local contextual information, we further develop a compact convolutional channel mixer to encode spatially local context and conduct channel mixing simultaneously. We both qualitatively and quantitatively evaluate the proposed method on commonly used benchmarks. Experimental results show that the proposed SAFMN model is more efficient than state-of-the-art methods while achieving competitive performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.729 + ], + "angle": 0, + "content": "Acknowledgements. This work has been partly supported by the National Key R&D Program of China (No. 2018AAA0102001), the National Natural Science Foundation of China (Nos. U22B2049, 62272233, 61922043, 61925204), and the Fundamental Research Funds for the Central Universities (Nos. 30922010910, 30920041109)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.745, + 0.174, + 0.761 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.771, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, 2018. 1, 2, 4, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.816, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[2] Pablo Arbeláez, Michael Maire, Charless C. Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. PAMI, 33(5):898-916, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.771, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.15, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.208, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[6] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In CVPR, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.265, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[7] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.309, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[8] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.352, + 0.892, + 0.406 + ], + "angle": 0, + "content": "[9] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.41, + 0.892, + 0.451 + ], + "angle": 0, + "content": "[10] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. PAMI, 38(2):295-307, 2016. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.453, + 0.892, + 0.494 + ], + "angle": 0, + "content": "[11] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In ECCV, 2016. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[13] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In CVPR, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.612, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[14] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.655, + 0.892, + 0.695 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.699, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[16] Zibin He, Tao Dai, Jian Lu, Yong Jiang, and Shu-Tao Xia. Faked: Feature-affinity based knowledge distillation for efficient image super-resolution. In ICIP, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.743, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[17] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units. arXiv preprint arXiv:1606.08415, 2016. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[18] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[19] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.898 + ], + "angle": 0, + "content": "[20] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multistillation network. In ACM MM, 2019. 1, 2, 4, 5, 6, 7, 9" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13198" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[21] Andrey Ignatov, Radu Timofte, Maurizio Denna, and et al. Efficient and accurate quantized image super-resolution on Mobile NPUs, Mobile AI & AIM 2022 challenge: Report. In ECCV Workshops, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[23] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[25] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[26] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPR Workshops, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[27] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[28] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. LAPAR: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. In NeurIPS, 2020. 4, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[29] Yawei Li, Kai Zhang, Luc Van Gool, Radu Timofte, et al. NTIRE 2022 challenge on efficient super-resolution: Methods and results. In CVPR Workshops, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[30] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In CVPR Workshops, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In ICCV Workshops, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPR Workshops, 2017. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[33] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV Workshops, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[36] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: Efficiently amalgamated cnn-transformer architecture for mobile vision applications. In ECCV Workshops, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[38] Yusuke Matsui, Kota Ito, Yuji Aramaki, Toshihiko Yasasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. arXiv preprint arXiv:1510.04389, 2015. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[39] Sachin Mehta and Mohammad Rastegari. MobileViT: Lightweight, general-purpose, and mobile-friendly vision transformer. In ICLR, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[40] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[41] Pablo Navarrete Michelini, Yunhua Lu, and Xingqun Jiang. edge-SR: Super-resolution for the masses. In WACV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[42] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[44] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[45] Long Sun, Jinshan Pan, and Jinhui Tang. ShuffleMixer: An efficient convnet for image super-resolution. In NeurIPS, 2022. 1, 2, 4, 5, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[46] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. CVPR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[47] Mingxing Tan and Quoc Le. EfficientNetV2: Smaller models and faster training. In ICML, 2021. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[48] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, Lei Zhang, et al. NTIRE 2017 challenge on single image super-resolution: Methods and results. In CVPR Workshops, 2017. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[49] Longguang Wang, Xiaoyu Dong, Yingqian Wang, Xinyi Ying, Zaiping Lin, Wei An, and Yulan Guo. Exploring sparsity in image super-resolution for efficient inference. In CVPR, 2021. 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[50] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[51] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces*, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[52] Xindong Zhang, Hui Zeng, and Lei Zhang. Edge-oriented convolution block for real-time super resolution on mobile devices. In ACM MM, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[53] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.874 + ], + "angle": 0, + "content": "[54] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCV Workshops, 2020. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.874 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13199" + } + ] +] \ No newline at end of file diff --git a/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/8203ecb7-a9b9-46de-8112-1277f4eeac63_origin.pdf b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/8203ecb7-a9b9-46de-8112-1277f4eeac63_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..643e90a8fd80366e93dcf7bfb368445bdcde2347 --- /dev/null +++ b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/8203ecb7-a9b9-46de-8112-1277f4eeac63_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:531457c98127b833d5205027c4e5e5374e315c9a0c29f2715de4e5af3662544d +size 5941430 diff --git a/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/full.md b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bae05f93a731744e69d48a4c708cc2d0354276e6 --- /dev/null +++ b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/full.md @@ -0,0 +1,386 @@ +# Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution + +Long Sun, Jiangxin Dong, Jinhui Tang, and Jinshan Pan* + +School of Computer Science and Engineering, Nanjing University of Science and Technology + +# Abstract + +Although deep learning-based solutions have achieved impressive reconstruction performance in image superresolution (SR), these models are generally large, with complex architectures, making them incompatible with low-power devices with many computational and memory constraints. To overcome these challenges, we propose a spatially-adaptive feature modulation (SAFM) mechanism for efficient SR design. In detail, the SAFM layer uses independent computations to learn multi-scale feature representations and aggregates these features for dynamic spatial modulation. As the SAFM prioritizes exploiting non-local feature dependencies, we further introduce a convolutional channel mixer (CCM) to encode local contextual information and mix channels simultaneously. Extensive experimental results show that the proposed method is $3 \times$ smaller than state-of-the-art efficient SR methods, e.g., IMDN, and yields comparable performance with much less memory usage. Our source codes and pre-trained models are available at: https://github.com/sunny2109/SAFMN. + +# 1. Introduction + +Single image super-resolution (SISR) aims to restore a high-resolution (HR) image from its low-resolution (LR) counterpart by recovering lost details. This longstanding and challenging task has recently attracted much attention due to the rapid development of streaming media or high-definition devices. As these scenarios are usually resource-constrained, it is of great interest to develop an efficient and effective SR method to estimate HR images for better visual display on these platforms or products. + +Deep learning-based SR methods have achieved significant performance improvements with the great evolution of hardware technologies, as we can use large amounts of data to train much larger or deeper neural networks for image SR [32, 53, 31, 5]. For example, RCAN [53] is a representative CNN-based image SR network with 15.59M parameters and reaching a depth of over 400 layers. One of the most significant drawbacks of these large models is that they require high computational costs, which makes them + +![](images/7f5f94fe3b6904c2563a45c10789fd0ac832f8a950da2e113df2b59caeb481bb.jpg) +Figure 1. Model complexity and performance comparison between our proposed SAFMN model and other lightweight methods on Set5 [4] for $\times 2$ SR. Circle sizes indicate the number of parameters. The proposed method achieves a better trade-off between model complexity and reconstruction performance. + +challenging to deploy. Moreover, recent visual transformers (ViTs) [12, 31, 5] outperform convolutional neural networks (CNNs) in low-level vision tasks, and their results demonstrate that exploring non-local feature interactions is essential for high-quality reconstruction. But existing self-attention mechanisms are computationally expensive and unfriendly to efficient SR design. This, therefore, motivates us to develop a lightweight yet effective model for real-world applications of image super-resolution by integrating the principles of convolution and self-attention. + +To reduce the heavy computational burden, various methods, including efficient module design [11, 41, 1, 20, 26, 33, 44, 45, 54, 30], knowledge distillation [16], neural architecture search [7], and structural re-parameterization [52], are trying to improve the efficiency of SR algorithms. Among these efficient SR models, one direction is to reduce model parameters or complexity (FLOPs). Lightweight strategies like recursive manner [23, 46], parameter sharing [1], and spare convolutions [1, 45, 30] are adopted. Although these approaches certainly reduce the model size, they usually compensate for the performance drop caused by shared recursive modules or sparse convolutions by increasing the depth or width + +of the model, which affects the inference efficiency when performing SR reconstruction. + +Another direction is to accelerate the inference time. The post-upsampling [11, 44] is an important replacement for the pre-defined input [10, 24], which significantly speeds up the runtime. Model quantization [21] effectively accelerates latency and reduces energy consumption, particularly when deploying algorithms in edge-devices. Structural re-parameterization [52, 8] improves the speed of a well-trained model in the inference stage. These methods enjoy fast running time but poor reconstruction performance. Consequently, there is still room for a better trade-off between model efficiency and reconstruction performance. + +To address the above-mentioned issues, we design a simple yet effective model by developing a spatially-adaptive feature modulation, namely SAFMN, to realize a favorable trade-off between performance and efficiency. Specifically, we first exploit non-local feature relations to dynamically select representative features (see Figure 2 and 5) by implementing a feature modulation mechanism based on multiscale representations. As the modulation mechanism processes input features from a non-local perspective, there is a requirement to complement local contextual information. To this end, we present a convolutional channel mixer based on the FMBCov [47] to encode local features and mix channels. Taken together, we find that the SAFMN network is able to achieve a better trade-off between SR performance and model complexity, as shown in Figure 1. + +The main contributions of this paper are summarized as follows: + +- We develop an efficient feature modulation mechanism to learn feature dependencies, which absorbs CNN-like efficiency and transformer-like adaptability. +- We present a compact convolutional channel mixer that simultaneously encodes local contextual information and performs channel mixing. +- We evaluate the proposed method quantitatively and qualitatively on benchmark datasets, and the results show that our SAFMN achieves a favorable trade-off between accuracy and model complexity. + +# 2. Related Work + +Deep Learning-based Image Super-Resolution. Classical interpolation algorithms, such as linear or bicubic upsampling, create high-resolution images by inserting zeros between adjacent pixels in the low resolution and then using a low-pass filter to preserve the content information of the input image [41]. Unlike these interpolation-based upsamplers, deep learning-based approaches learn a nonlinear mapping between the input image and the target output in an end-to-end training fashion. SRCNN [10] is the first attempt + +![](images/8a6be1bfafab898fefc3d04f472d9d70902d0229d999d1a0a4e310f1d54ec6a7.jpg) +DI +(a) +Figure 2. Comparison of local attribution maps (LAMs) [13] and diffusion indices (DIs) [13] between our SAFMN and other efficient SR models. (a) is the LR input, (b)-(e) are the results of CARN [1], EDSR-baseline [32], PAN [54] and SAFMN, respectively. The LAM results denote the importance of each pixel in the input LR image when super-resolving the patch marked with a red box. The DI value reflects the range of involved pixels. A larger DI value means a wider range of attention. The proposed method can exploit more feature information. + +![](images/30acbc6abd5027152737258acc386d49e5d935ca744b4b69563848a6b88583c8.jpg) +3.2180 +(b) + +![](images/c13192d0602280aaff706dd38fe8e389fce80cc98dd7e5214ca12d8aa7b863e7.jpg) +5.1936 +(c) + +![](images/2e67f2fbc1dd450fa7088d7a56e0dd2173494720f44016025753f5788df065d9.jpg) +4.0629 +(d) + +![](images/d4f6e8dd8f63302bb3ef3dd889913f61010d21db5dec58e2aa426494d618784e.jpg) +10.4295 +(e) + +to use a convolutional neural network to tackle the image SR problem and achieves a considerable performance gain compared with conventional methods. Since then, many improvements have been proposed. VDSR [24] uses global residual learning [15] to solve the problem of difficulty in training an SR model with deep layers. DRRN [46] integrates the local residual learning and global residual connection to ease the training difficulty and enhance high-frequency details. EDSR [32] further increases the model footprint to 43M, achieving a significant breakthrough in reconstruction performance and showing that the BatchNorm (BN) [22] layer is not necessary for the SR task. RCAN [53] builds a more than 400 layers model based on channel attention and dense connections for accurate SR. NLSA [40] utilizes a non-local sparse attention to enhance the feature modeling capability and reaches state-of-the-art performance. With the successful application of ViT in various high-level vision tasks, image SR also follows this ViT [12] framework, achieving higher reconstruction performance than CNN-based models on extensive public benchmarks. For instance, SwinIR [31] acts as a strong baseline for image restoration tasks based on the Swin Transformer [34]. While these approaches achieve impressive reconstruction performance, the required high computational costs make them challenging to deploy in real-world applications on resource-constrained devices. + +Efficient Image Super-Resolution. To improve the model efficiency, many CNN-based SR works try to alleviate this issue. FSRCNN [11] and ESPCN [44] utilize a postupsampling manner to reduce the computational burden from the pre-defined inputs significantly. CARN [1] uses group convolutions and a cascading mechanism upon residual networks to improve efficiency. IMDN [20] adopts feature splitting and concatenation operations to progressively aggregate features, and its improved variants [33, 26] won the AIM2020 and NTIRE2022 Efficient SR challenge. ShuffleMixer [45] introduces a large kernel convolution for lightweight SR design. BSRN [30] employs blueprint-separable convolutions to reduce model complexity. Mean + +![](images/393c8fd6b666674a4e818e2c5fbfd0168d7f4e5b4b09f2957d58231ad3184e63.jpg) +Figure 3. An overview of the proposed SAFMN. SAFMN first transforms the input LR image into the feature space using a convolutional layer, performs feature extraction using a series of feature mixing modules (FMMs), and then reconstructs these extracted features by an upsampler module. The FMM block is implemented by a spatially-adaptive feature modulation (SAFM) layer, a convolutional channel mixer (CCM) and two skip-connections. + +while, an increasingly popular direction is to compress or accelerate a well-trained deep model through model quantization [21], structural re-parameterization [52] or knowledge distillation [16]. Neural architecture search (NAS) is also commonly used to search a well-constrained architecture for image super-resolution [7]. Note that the efficiency of a deep neural network could be measured in different metrics, including the number of parameters, FLOPs, activations, memory consumption and inference running time [29]. Although the above approaches have been improved in different efficiency aspects, there is still room for a favorable trade-off between reconstruction performance and model efficiency. + +# 3. Proposed Method + +In this section, we present the core components of our proposed model for efficient SISR. As shown in Figure 3, the network consists of the following parts: a shallow convolution, a stacking of feature mixing modules (FMMs), and an upsampler layer. Specifically, we first apply a $3 \times 3$ convolution layer to transform the input LR image to feature space and generate the shallow feature $\mathbf{F}_0$ . Then, the multiple stacked FMMs are used to generate finer deep features from $\mathbf{F}_0$ for HR image reconstruction, where an FMM layer has a spatially-adaptive feature modulation (SAFM) layer and a convolutional channel mixer (CCM). To recover the HR target image, we introduce a global residual connection to learn high-frequency details and employ a lightweight upsampling layer for fast reconstruction, which only contains a $3 \times 3$ convolution and a pixel-shuffle layer [44]. + +# 3.1. Spatially-adaptive Feature Modulation + +Recent studies [9, 35, 14, 39, 31] suggest that the notable performance of ViTs across diverse tasks stems from their implementation of the key multi-head self-attention (MHSA) mechanism. This mechanism enables the model with the capability of long-range feature interaction and dynamic spatial weighting, both of which contribute to pro + +![](images/6320c5b4e1c0681f68965f91caa103296b32cdac5aa5978562909a3ea02eec1b.jpg) +Figure 4. Network architecture of the proposed SAFM module. This module performs spatially-adaptive modulation on the input features. Please see the Section 3.1 for more details. + +ducing promising results. But existing self-attention variants are computationally expensive and unfriendly to efficient SR design. In contrast, standard convolution is an efficient operation, but its performance is limited by the static weights and locality. This observation inspires us to borrow the idea of MHSA to enhance the representation capability of convolution. + +To introduce the ability of long-range interaction and dynamic modeling into convolution, we follow the multi-head paradigm, using parallel and independent computations that allow each head to process different scale information of the input, and then aggregate these features to generate an attention map for spatially-adaptive feature modulation. + +The detailed architecture of our proposed SAFM is depicted in Figure 4. Specifically, we first split the normalized input features into four-group components, and feed them into a multi-scale feature generation unit (MFGU), where a $3 \times 3$ depth-wise convolution processes the first one, and the rest parts are sampled individually by pooling operations. As we desire to select discriminative features towards learning non-local feature interactions, adaptive max pooling operators are applied over the input features to gather information. Given the input feature $X$ , this procedure can be formulated as: + +$$ +\left[ X _ {0}, X _ {1}, X _ {2}, X _ {3} \right] = \operatorname {S p l i t} (X), +$$ + +$$ +\hat {X} _ {0} = \operatorname {D W - C o n v} _ {3 \times 3} \left(X _ {0}\right), \tag {1} +$$ + +$$ +\hat {X} _ {i} = \uparrow_ {p} \left(\operatorname {D W - C o n v} _ {3 \times 3} \left(\downarrow_ {\frac {p}{2 ^ {i}}} \left(X _ {i}\right)\right)\right), 1 \leq i \leq 3, +$$ + +where $\text{Split}(\cdot)$ corresponds to the channel split operation, $\text{DW-Conv}_{3\times 3}(\cdot)$ is a $3\times 3$ depth-wise convolution, $\uparrow_p(\cdot)$ represents upsampling features at a specific level to the original resolution $p$ via nearest interpolation for fast implementation, and $\downarrow_{\frac{p}{2^i}}$ denotes pooling the input features to the size of $\frac{p}{2^i}$ . Figure 5(b)-(e) show that using such a design, we can obtain varying features from different representation subspaces. + +Afterward, we aggregate these extracted short- or long-range features by concatenating them on the channel dimension and performing a $1 \times 1$ convolution. It can be formulated as: + +$$ +\hat {X} = \operatorname {C o n v} _ {1 \times 1} \left(\operatorname {C o n c a t} \left([ \hat {X} _ {0}, \hat {X} _ {1}, \hat {X} _ {2}, \hat {X} _ {3} ]\right)\right), \tag {2} +$$ + +where $\operatorname{Concat}(\cdot)$ denotes the concatenation operation, and $\operatorname{Conv}_{1\times 1}(\cdot)$ is the $1\times 1$ convolution. After obtaining the aggregated representation $\hat{X}$ (see Figure 5(f)), we normalize it through a GELU non-linearity [17] to estimate the attention map and adaptively modulate the input $X$ according to the estimated attention via element-wise product. This process can be written as: + +$$ +\bar {X} = \phi (\hat {X}) \odot X, \tag {3} +$$ + +where $\phi(\cdot)$ represents the GELU function and $\odot$ is the element-wise product. The benefits of those operators are compared and results are shown in Table 3. + +# 3.2. Feature Mixing Module + +As mentioned previously, the SAFM module prioritizes exploiting non-local feature interactions. To further incorporate local contextual information and perform channel mixing simultaneously, we introduce a compact convolutional channel mixer (CCM) based on the FMBCnv [47]. The CCM comprises a $3 \times 3$ convolution and a $1 \times 1$ convolution. Within this, the first $3 \times 3$ convolution encodes the spatially local contexts and doubles the number of channels of the input features for mixing channels; the later $1 \times 1$ convolution reduces the channels back to the original input dimension. A GELU [17] function is applied to the hidden layer for non-linear mapping. + +We formulate the proposed SAFM and CCM into a unified feature mixing module (FMM) to select representative features. The FMM can be written as: + +$$ +\begin{array}{l} Y = \operatorname {S A F M} (\operatorname {L N} (X)) + X, \\ Z = \operatorname {C C M} (\operatorname {L N} (Y)) + Y, \end{array} \tag {4} +$$ + +where $\mathrm{LN}(\cdot)$ is the LayerNorm [3] layer, $X$ , $Y$ , and $Z$ are the intermediate features. The additional residual learning is employed to stabilize training processes and learn high frequency details to facilitate high-quality image reconstruction. + +![](images/b45186fcca63868e4a60edf820afb6e75a97b394336a82cb8d6e3e295e00402d.jpg) +(a) LR input +Figure 5. Illustration of learned deep features from the SAFM module. (a) is the LR input. (b)-(e) are features learned by the multi-scale feature generation unit. (f) is the results of aggregating features (b)-(e). (g) denotes the attended output of the SAFM. + +![](images/da54769aa655ccede61a63ec8b80831fc3be69faa912f41969987d152a73dc9c.jpg) + +![](images/2daa8534748e0c099c21c63eaefee922dbe6cbd0a131557e011047cad6bd5b25.jpg) + +![](images/f986e3064fc4a1f5a8ba0ffea38aa718d3edd04fd8814d17a25c53e33473dc0d.jpg) + +![](images/ed4d6368d75c596548d09b452c59bffa36d273764cc35a515a9891acee9cc800.jpg) +(b) Scale 1 +(e) Scale 8 + +![](images/be287c37315a90c066b9eeeb172012e46548abba9841e76ac836fc2d6ec71b94.jpg) +(c) Scale 2 +(f) Aggregated + +![](images/33d97f925d1edad43e51fd429aeacc579f0298e50710996fca4b079542ceb7dd.jpg) +(d) Scale 4 +(g) Attended + +# 4. Experimental Results + +In this section, we perform quantitative and qualitative evaluations with state-of-the-art algorithms to demonstrate the effectiveness of the proposed method. + +# 4.1. Dataset and Implementation + +Datasets. Following previous works [28, 31, 45], we use DIV2K [48] and Flickr2K [32] as the training data and generate LR images by applying the bicubic downscaling to reference HR images. We use five commonly used benchmark datasets including Set5 [4], Set14 [51], B100 [2], Urban100 [19] and Manga109 [38] as test data. We use the peak signal to noise ratio (PSNR) and the structural similarity index (SSIM) to evaluate the quality of the restored images. All PSNR and SSIM values are calculated on the Y channel of images transformed to YCbCr color space. + +Implementation details. During the training, the data argumentation is performed on the input patches with random horizontal flips and rotations. In addition, we randomly crop 64 patches of size $64 \times 64$ pixels from LR images as the basic training inputs. The number of FMM and feature channels is set to 8 and 36, respectively. We use the Adam [25] optimizer with $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ to solve the proposed model. The number of iterations is set to 500,000. We set the initial learning rate to $1 \times 10^{-3}$ and the minimum one to $1 \times 10^{-5}$ , which is updated by the Cosine Annealing scheme [36]. We use a combination of mean absolute error (MAE) loss and an FFT-based frequency loss function to constrain the model training, which is the same as [45]. All experiments are conducted with the PyTorch framework on an NVIDIA GeForce RTX 3090 GPU. Due to the page limit, we include more comparison results in the supplemental material. + +# 4.2. Comparisons with State-of-the-Art Methods + +Quantitative comparisons. To evaluate the performance of the proposed approach, we compare it with state-of-the-art lightweight SR methods on different scaling factors, including SRCNN [10], FSRCNN [11], ESPCN [44], VDSR [24], LapSRN [27], CARN [1], EDSR-baseline [32], IMDN [20], + +Table 1. Comparisons of efficient SR networks on the commonly used benchmark datasets. All PSNR/SSIM results are calculated on the Y-channel. #Acts represents all elements of the output of convolutional layers. #FLOPs and #Acts are measured corresponding to an HR image of the size ${1280} \times {720}$ pixels. Red color denotes the best performance. Blanked entries indicate results not reported or not available from previous work. + +
MethodsScale#Params [K]#FLOPs [G]#Acts [M]Set5Set14B100Urban100Manga109
Bicubic×2---33.66/0.929930.24/0.868829.56/0.843126.88/0.840330.80/0.9339
SRCNN [10]57538936.66/0.954232.42/0.906331.36/0.887929.50/0.894635.74/0.9661
FSRCNN [11]1264137.00/0.955832.63/0.908831.53/0.892029.88/0.902036.67/0.9694
ESPCN [44]2152336.83/0.956432.40/0.909631.29/0.891729.48/0.8975-
VDSR [24]6656131,12037.53/0.958733.03/0.912431.90/0.896030.76/0.914037.22/0.9729
LapSRN [27]8133022337.52/0.959033.08/0.913031.80/0.895030.41/0.910037.27/0.9740
CARN-M [1]4159165537.53/0.958333.26/0.914131.92/0.896031.23/0.9193-
CARN [1]1,59222352237.76/0.959033.52/0.916632.09/0.897831.92/0.9256-
EDSR-baseline [32]1,37031656337.99/0.960433.57/0.917532.16/0.899431.98/0.927238.54/0.9769
IMDN [20]69415942338.00/0.960533.63/0.917732.19/0.899632.17/0.928338.88/0.9774
PAN [54]2617167738.00/0.960533.59/0.918132.18/0.899732.01/0.927338.70/0.9773
LAPAR-A [28]54817165638.01/0.960533.62/0.918332.19/0.899932.10/0.928338.67/0.9772
ECBSR-M16C64 [52]59613725237.90/0.961533.34/0.917832.10/0.901831.71/0.9250-
SMSR [49]985132-38.00/0.960133.64/0.917932.17/0.899032.19/0.928438.76/0.9771
ShuffleMixer [45]3949183238.01/0.960633.63/0.918032.17/0.899531.89/0.925738.83/0.9774
SAFMN (Ours)2285229938.00/0.960533.54/0.917732.16/0.899531.84/0.925638.71/0.9771
Bicubic×3---30.39/0.868227.55/0.774227.21/0.738524.46/0.734926.95/0.8556
SRCNN [10]57538932.75/0.909029.28/0.820928.41/0.786326.24/0.798930.59/0.9107
FSRCNN [11]1251933.16/0.914029.43/0.824228.53/0.791026.43/0.808030.98/0.9212
VDSR [24]6656131,12033.66/0.921329.77/0.831428.82/0.797627.14/0.827932.01/0.9310
CARN-M [1]4154632733.99/0.923630.08/0.836728.91/0.800027.55/0.8385-
CARN [1]1,59211926834.29/0.925530.29/0.840729.06/0.803428.06/0.8493-
EDSR-baseline [32]1,55516028534.37/0.927030.28/0.841729.09/0.805228.15/0.852733.45/0.9439
IMDN [20]7037219034.36/0.927030.32/0.841729.09/0.804628.17/0.851933.61/0.9445
PAN [54]2613934034.40/0.927130.36/0.842329.11/0.805028.11/0.851133.61/0.9448
LAPAR-A [28]59411450534.36/0.926730.34/0.842129.11/0.805428.15/0.852333.51/0.9441
SMSR [49]99368-34.40/0.927030.33/0.841229.10/0.805028.25/0.853633.68/0.9445
ShuffleMixer [45]4154340434.40/0.927230.37/0.842329.12/0.805128.08/0.849833.69/0.9448
SAFMN (Ours)2332313434.34/0.926730.33/0.841829.08/0.804827.95/0.847433.52/0.9437
Bicubic×4---28.42/0.810426.00/0.702725.96/0.667523.14/0.657724.89/0.7866
SRCNN [10]57538930.48/0.862827.49/0.750326.90/0.710124.52/0.722127.66/0.8505
FSRCNN [11]1251130.71/0.865727.59/0.753526.98/0.715024.62/0.728027.90/0.8517
ESPCN [44]251630.52/0.869727.42/0.760626.87/0.721624.39/0.7241-
VDSR [24]6656131,12031.35/0.883828.01/0.767427.29/0.725125.18/0.752428.83/0.8809
LapSRN [27]81314926431.54/0.885028.19/0.772027.32/0.728025.21/0.756029.09/0.8845
CARN-M [1]4153322731.92/0.890328.42/0.776227.44/0.730425.62/0.7694-
CARN [1]1,5929119432.13/0.893728.60/0.780627.58/0.734926.07/0.7837-
EDSR-baseline [32]1,51811420232.09/0.893828.58/0.781327.57/0.735726.04/0.784930.35/0.9067
IMDN [20]7154110832.21/0.894828.58/0.781127.56/0.735326.04/0.783830.45/0.9075
PAN [54]2722823832.13/0.894828.61/0.782227.59/0.736326.11/0.785430.51/0.9095
LAPAR-A [28]6599445232.15/0.894428.61/0.781827.61/0.736626.14/0.787130.42/0.9074
ECBSR-M16C64 [52]603356431.92/0.894628.34/0.781727.48/0.739325.81/0.7773-
SMSR [49]100642-32.12/0.893228.55/0.780827.55/0.735126.11/0.786830.54/0.9085
ShuffleMixer [45]4112826932.21/0.895328.66/0.782727.61/0.736626.08/0.783530.65/0.9093
SAFMN (Ours)240147732.18/0.894828.60/0.781327.58/0.735925.97/0.780930.43/0.9063
+ +PAN [54], LAPAR [28], ECBSR [52], SMSR [49], and ShuffleMixer [45]. + +The quantitative comparisons on benchmark datasets for the upscaling factors of $\times 2$ , $\times 3$ , and $\times 4$ are reported in Table 1. In addition to PSNR/SSIM metrics, we list the number of parameters (#Params), FLOPs (#FLOPs) and activations (#Acts). We calculate the number of FLOPs and activations with the fvcore1 library under a setting of super-resolving an LR image to $1280 \times 720$ pixels. Among these metrics, #Params and #Acts are linked to memory consumption, and #FLOPs is related to energy usage. In particular, + +Acts is a better metric for measuring the efficiency of a model than the number of parameters and FLOPs as suggested in recent works [42, 52, 29]. + +Benefiting from the simple yet efficient structure, the proposed SAFMN obtains comparable performance with significantly fewer parameters and memory consumption. For instance, the $\times 4$ SR results on the B100 dataset reveal that our SAFMN has parameters about $85\%$ less than the CARN [1], $66\%$ less than the IMDN [20], and $42\%$ less than the ShuffleMixer [45]. As for activations, we have $60\%$ , $29\%$ and $71\%$ fewer than them, respectively. While our model has a smaller footprint, we achieve similar performance among these methods. Moreover, we compare + +![](images/f95ca739a69a4cdd8c4742c439b180c6e154276564b0f83236be35c2f9653b4b.jpg) +img024 from Urban100 + +![](images/5f08c0aad803c7c0e0f51133f12e5ab352dc5c33c9205208e4768847194cf5fa.jpg) +(a) HR patch + +![](images/db4592ed12cb81241769af5a9a5d57bb3ed6b21fb98a4af844d09d7f342c54f7.jpg) +(b) Bicubic + +![](images/10d9983022c13b989d58fdb4f73a23d2af3c7e30e1620f9385c84323ae25a849.jpg) +(c) VDSR [24] + +![](images/e8bad97d5cf8b0a62f617e00ed6a09482c4f3f0cd18dec9190e3a9f26ee12c9e.jpg) +(d) ShuffleMixer [45] + +![](images/38b79e2d65202dd949a67dd38b8d5f5c749ece62ba1cc2efbbb4e123c4fc758b.jpg) +(e) LapSRN [27] + +![](images/61fda92f8777e8bc55c4c5e3d5ed4061cfcfc40a46e56dd137576598147b05f5.jpg) +(f) CARN [1] + +![](images/07e881585ecd68c4795f71db766598c00e10ed81f8162b5299c6b104296572d4.jpg) +(g) IMDN [20] + +![](images/caf2e5c4e65179ea6cd14438872d6100b484045b74b75eee5f0da56f3c0b7f4e.jpg) +(h) SAFMN + +![](images/6f7c923d0205866c8c827499c5d78cb70f1963b9d5fa0258b78f8a0e3b762da5.jpg) +img062 from Urban100 + +![](images/622f350b20b08c08ca2a81decc4d8fa343a730e114883df4e5ed492fbbfc5e46.jpg) +(a) HR patch + +![](images/5d0e7cb23b4086db9d47a77daec5596f4c7f7727454712d6e44257ff17d49eb1.jpg) +(b) Bicubic + +![](images/71e5d281c226a473a50dee59c44baab8cf287c14f7834c54a3574e144061bd74.jpg) +(c) VDSR [24] + +![](images/c1176d288c6656b23b12f713ee7c43bb1e3da66961ba56ca3367783ba9e10710.jpg) + +![](images/8820d156b0c82fae48dce421a21ce6c3b903332c6ed019aba3c797762d567a2b.jpg) +(e) LapSRN [27] + +![](images/274b0fa28ba294b772474b88dd130cf5ac21656b8ccc4c84d4d7b0697d5cb9a0.jpg) +(f) CARN [1] + +![](images/42eaf61c08a0c8515e656227235af5d0c0aae50f5290e2b5d01e5327c19f0d4d.jpg) +(g) IMDN [20] + +![](images/3e6bf4e654fd4d339ce58f0610f926a38a2a5efc1c0f7320057a366f9bf720bd.jpg) +(d) ShuffleMixer [45] +(h) SAFMN + +![](images/b1adef27a5c285f442aee31378aa9e76a80203eb26cb485480b46fdcaf3e8174.jpg) +img098 from Urban100 + +![](images/6cd63aeac1fe8ca0610913d9ff06eef59ca503f0fd4a8dc38ff024de7ef7ac4c.jpg) + +![](images/29d9f88e24fcfa660b6e81d81fe1a59f1c1e0a15cf7b33f1206d3b9c148b024b.jpg) + +![](images/102ec09a11e96e2616ea61933e51d6e081ef910c421dfc88171f190ab35c7e78.jpg) + +![](images/695a38c4790475a345b79fe11b9817bf3733219f2d4758432aec88d49de4d9d8.jpg) + +![](images/b2b6dca1ce0e7511b68bd3f0dfdbd518a73e55e4cd26423a7344461ed0411231.jpg) +(a) HR patch +(e) LapSRN [27] +Figure 6. Visual comparisons for $\times 3$ SR on the Urban100 dataset. The proposed method recovers the image with clearer structures. + +![](images/220e43d6dd4197ddb818451e4b9b2d69eb8d39156832abf7b85deee4f024ca71.jpg) +(b) Bicubic +(f) CARN [1] + +![](images/ef2a68857cb59fd7ab26ce8c84aa8002da778f481a48fc8113fe9cc22b435fdb.jpg) +(c) VDSR [24] +(g) IMDN [20] + +![](images/a6ea036299f59c8afbec5ef5d4e7807642ce93d1b1048ade568b17cd1dfc3c03.jpg) +(d) ShuffleMixer [45] +(h) SAFMN + +the reconstruction accuracy, FLOPs and parameters on the $\times 2$ Set5 dataset in Figure 1. The proposed SAFMN model achieves a favorable trade-off between model complexity and reconstruction performance. + +Qualitative comparisons. In addition to the quantitative evaluations, we provide qualitative comparisons of the proposed method. Figure 6 shows the visual comparisons on the Urban100 dataset for $\times 3$ SR. Our approach generates parallel straight lines and grid patterns more accurately than the listed methods. These results also demonstrate the effectiveness of our method for adaptive feature modulation by exploiting non-local feature interactions. + +Memory and running time comparisons. To fully examine the efficiency of our proposed method, we further evaluate our method against five representative ones, including CARN-M [1], CARN [1], EDSR-baseline [32], IMDN [20], and LAPAR-A [28] on $\times 4$ SR in terms of + +the GPU memory consumption (#GPU Mem.) and running time (#Avg. Time). The maximum GPU memory consumption is recorded during the inference. The running time is averaged on 50 test images with a resolution of $320 \times 180$ pixels. We show the memory and running time comparisons in Table 2, our method achieves a clear improvement over other state-of-the-art methods. By using the spatially-adaptive feature modulation layer and the compact channel mixer, the GPU consumption of our SAFMN is only $10\%$ of the CARN series and $4\%$ of LAPAR-A; the running time is nearly twice as fast as other evaluated methods, except for IMDN. Compared to IMDN [20], our method has a similar running time speed while significantly reducing memory usage. Tables 1 and 2 show that the proposed model achieves a favorable trade-off in terms of inference speed, model complexity and reconstruction performance against state-of-the-art methods. + +Table 2. Memory and running time comparisons on $\times 4$ SR. #GPU Mem. denotes the maximum GPU memory consumption during the inference phase, derived with the Pytorch torch.cuda.max_memory_allocated() function. #Avg. Time is the average running time on 50 LR images of $320 \times 180$ pixels. + +
Methods#GPU Mem. [M]#Avg.Time [ms]
CARN-M [1]680.8417.85
CARN [1]689.8318.90
EDSR-baseline [32]486.5819.81
IMDN [20]203.4410.22
LAPAR-A [28]1811.4724.91
SAFMN65.2610.71
+ +# 5. Analysis and Discussion + +We further conduct extensive ablation studies to better understand and evaluate each component in the proposed SAFMN. For fair comparisons with the designed baselines, we implement all experiments based on $\times 4$ SAFMN and train them with the same setting. The experimental results in Table 3 are measured on DIV2K-val [48] and Manga109 [38] datasets. + +Effectiveness of the SAFM. To demonstrate the effect of the spatially-adaptive feature modulation, we first remove this module for comparison. Without it, the PSNR values will drop by 0.17dB (30.26 vs. 30.43) and 0.34dB (30.09 vs. 30.43) on the DIV2K-val and Manga109 datasets, respectively. These results show the importance of the SAFM. Therefore, we further investigate this module to figure out why it works. + +- Feature modulation. The feature modulation mechanism is introduced to enable the network with adaptive properties. Without this operation, the baseline model is lowered by 0.11dB on the Manga109 dataset. +- Multi-scale representation. Here, "w/o MR" in Table 3 denotes that we directly use a depth-wise convolution with a kernel size of $3 \times 3$ pixels to extract spatial information and do not employ multi-scale features. A noticeable performance drop of 0.13dB on the Manga109 dataset is observed without these multiscale features. Moreover, we apply the adaptive max pooling over the input features in this module to build feature pyramids. Compared to using adaptive average pooling or nearest interpolation, adaptive max pooling allows the model to detect discriminative features, resulting in better reconstruction results. +- Feature aggregation. We use a $1 \times 1$ convolution to aggregate the multi-scale features on the channel dimension. Combined with the modulation mechanism, it brings a PSNR improvement of 0.14dB on the Manga109 dataset, which proves the necessity of aggregating the multi-scale features. + +Without all three components mentioned before, it represents that only a $3 \times 3$ depth-wise convolution is + +used to encode the spatial information and leads to a PSNR reduction of 0.12dB and 0.2dB on the DIV2Kval and Manga109 datasets, respectively. This performance drop suggests that the spatially-adaptive modulation based on the multi-scale feature representation effectively boosts SR reconstruction performance. + +- GELU function. Here, we use the GELU [17] function to normalize the modulation map. Table 3 shows that better experimental results can be achieved with GELU than with Sigmoid or without GELU, mainly because it weights the input features by their percentile and allows better-activated features. + +In addition, we use the local attribution map (LAM) [13] and diffusion index (DI) [13] to illustrate the feature interaction range. Figure 2 clearly shows that the proposed model can utilise more feature information than previous CNN-based efficient SR methods. The more information used, the better the reconstruction performance in general. + +The above analysis shows that benefiting from the multiscale feature representation, our proposed SAFM can effectively exploit non-local feature interactions with small memory and computational costs. + +Effectiveness of the CCM. Compared with the original FMBCConv [47], the main change made by CCM is removing the SE [18] block. As shown in Table 3, using SE blocks, its performance is almost the same as without them. This result is mainly caused by the SAFM block allows adaptability in channel and spatial dimensions. Hence, we do not use the SE blocks for saving parameters. We next conduct a series of ablations to verify that CCM can effectively encode local contextual information and perform channel mixing. Without it, the model only achieves the accuracy of 29.69dB and 28.49dB on DIV2K-val and Manga109 datasets, proving the indispensability and locality modelling ability of this part. We then change CCM to other channel mixers commonly used in ViT architectures: channel MLP [12] or inverted residual block [43]. When the channel MLP is adopted for the channel mixer, there is a significant performance drop of 0.63dB (29.80 vs. 30.43) on the Manga109 dataset. This performance reduction results from a lack of capacity for modeling local contextual information. For the inverted residual block, the performance is highly similar to that of CCM, but the corresponding #Acts is increased by nearly 33M, which means more memory consumption and slower inference time. Thus, we use the CCM as the default channel mixer. + +Effect of the LayerNorm layer. Since we use the elementwise product in the SAFM module, it will result in abnormal gradient values and unstable model training, as shown in Figure 7(a). It is, therefore, necessary to normalize the input features. Here, we perform normalization with the LayerNorm [3] layer. To verify this assumption, we first remove the LayerNorm layer. Table 3 and Figure 7(a) show + +Table 3. Ablation for SAFMN on DIV2K-val and Manga109 datasets. SAFMN with a scaling factor of $\times 4$ is utilized as the baseline for ablation studies. The PSNR/SSIM values on benchmarks are reported. “ $\mathrm{A} \rightarrow \mathrm{B}$ ” is to replace A with B. “None” means to remove the operation. “Ir” denotes the learning rate. “FBN” is the abbreviation of the Frozen BatchNorm. “ $\mathrm{L}_2$ normalization” represents that the inputs are normalized by $\mathrm{L}_2$ -norm over the channel dimension. “FM”, “MR”, and “FA” are the abbreviations corresponding to feature modulation, multi-scale representation, and feature aggregation. * indicates that the corresponding results are obtained before the training collapse. The number of parameters, FLOPs and Acts are calculated in the same way as in Table 1. + +
AblationVariant#Params [K]#FLOPs [G]#Acts [M]DIV2K_valManga109
Baseline-239.5213.5676.7030.43/0.837230.43/0.9063
Main moduleSAFM → None225.4112.9054.6130.26/0.833030.09/0.9018
CCM → None30.721.6126.9329.69/0.819328.49/0.8756
SAFM(a): w/o FM239.5213.5676.7030.36/0.835730.32/0.9048
(b): w/o MR239.5213.6487.7830.34/0.835030.30/0.9047
(c): w/o FA228.8612.9660.1130.36/0.835530.29/0.9049
(a) + (b)239.5213.6487.7830.32/0.834530.24/0.9038
(a) + (c)228.8612.9660.1130.34/0.835130.28/0.9043
(a) + (b) + (c)228.8613.0571.1930.31/0.834430.23/0.9036
AdaptiveMaxPool → AdaptiveAvgPool239.5213.5676.7030.40/0.836430.41/0.9061
AdaptiveMaxPool → Nearest interpolate239.5213.5676.7030.36/0.835430.31/0.9048
GELU → None239.5213.5676.7030.40/0.836630.37/0.9058
GELU → Sigmoid239.5213.5676.7030.36/0.835530.29/0.9044
CCMw/ SE260.9813.5976.7030.39/0.836030.46/0.9067
CCM → Channel MLP73.634.0076.7030.17/0.831329.80/0.8980
CCM → Inverted residual block245.2813.85110.0030.43/0.837330.43/0.9064
NormalizationLN → None, lr=1 × 10-3*238.3713.5576.7030.29/0.834030.04/0.9014
LN → None, lr=1 × 10-4238.3713.5576.7030.15/0.830629.74/0.8970
LN → BN239.5213.7276.7030.28/0.835430.05/0.9029
LN → FBN*238.3713.5576.7030.30/0.834330.15/0.9028
LN → L2 normalization238.3713.5576.7030.37/0.835830.31/0.9049
+ +![](images/7c3078570a3532d142635edd2c3b35e19e0f056ff7d1e04e9402978aa7aa07a5.jpg) +(a) Effect of LayerNorm +(b) Comparison of different normalizations + +![](images/44de5793af16a9eb02ab82b433a0a47f0f1e7db89e3a7e16b681ced063f6ee55.jpg) +Figure 7. Smoothed training loss curves with or without normalization. Figure (a) shows LayerNorm stabilizes model training and converges better than without it. Figure (b) compares the effects of different normalization approaches on model training. + +that without it, the model suffers a training crash at a large learning rate and does not converge well at small ones with its PSNR of only 30.15dB and 29.74dB on the DIV2K-val and Manga109 datasets, respectively. Next, we compare the LayerNorm with three representative normalization methods, including BatchNorm [22], Frozen BatchNorm (FBN) [22], and $\mathrm{L}_2$ normalization in Figure 7(b). The results in Table 3 and Figure 7(b) demonstrate that the BN family decrease the PSNR/SSIM values by a large margin, and the FBN even does not guarantee a stable model training process. Although the $\mathrm{L}_2$ normalization allows the model to be trained successfully, its performance is not as good as that of using LayerNorm. Thus, the LayerNorm layer is set as default for the proposed SAFMN. + +Table 4. Effect of the number of FMMs on performance. + +
#N#Params [K]#FLOPs [G]#Acts [M]DIV2K_valManga109
4128.067.2540.7730.21/0.831929.86/0.8988
6183.7910.4058.6430.32/0.834730.18/0.9033
8239.5213.5676.7030.43/0.837230.43/0.9063
12350.9719.86113.0030.51/0.838930.71/0.9096
+ +Table 5. GPU runtime of meta-operations in the SAFM layer. + +
Meta-operationsLayerNormChannel SplittingAdaptive MaxPoolNearest InterpolationElement-wise ProductElement-wise Addition
Runtime [us]4651338524156193
+ +Effect of the number of FMMs. The proposed network contains several FMMs. We evaluate the effect of the number of FMMs by setting the number #N from 4 to 12. Table 4 shows that the performance gradually improves as #N increases. To achieve a favorable trade-off between model complexity and reconstruction performance, we choose 8 blocks as the default setting. + +Relations to lightweight ViT models. To alleviate the quadratic complexity of vanilla ViT [12], MobileViT [39], Mobile-Former [6], and EdgeNeXt [37] adopt a hybrid CNN-Transformer design, which replaces the shallow transformer blocks with convolutional layers to avoid using self-attention on high-resolution representations, and efficiently combines the transformer blocks with the modified mobilenet blocks to perform feature interactions in the later stages. Another alternative [50, 14] is to incorporate concepts from the transformer into convolutional modules. For example, MetaFormer [50] and VAN [14] use pooling operation or large kernel attention to replace the heavy-weight self-attention for contextual information aggregation. Com + +pared to these approaches, we develop an efficient feature modulation mechanism upon a grouped multi-scale representation to learn feature dependencies. Benefiting from the multi-head paradigm, the proposed model can exploit multiscale contexts across different heads for adaptive feature modulation while maintaining a low computational cost. + +Limitations. Although our SAFMN has better overall performance than the evaluated methods [1, 20, 49, 28, 45], its running time is relatively slow. As shown in Table 5, we count the inference time of different commonly used tensor operations with the torch.profiler function, and the results show that the utilized LayerNorm layers and residual connections lead to increased runtime. The running time of SAFMN can be further accelerated by optimizing the implementation of these components. We will investigate these in the future work. Another limitation is that the PSNR performance of SAFMN is suboptimal, especially on the Urban100 dataset, which is mainly due to the fact that we perform the feature modulation without considering the integration of local features. + +# 6. Conclusion + +In this paper, we propose a simple yet efficient CNN model to solve the efficient image super-resolution problem. Our SAFMN exploits non-local feature interactions upon a multi-scale feature representation. To complement the local contextual information, we further develop a compact convolutional channel mixer to encode spatially local context and conduct channel mixing simultaneously. We both qualitatively and quantitatively evaluate the proposed method on commonly used benchmarks. Experimental results show that the proposed SAFMN model is more efficient than state-of-the-art methods while achieving competitive performance. + +Acknowledgements. This work has been partly supported by the National Key R&D Program of China (No. 2018AAA0102001), the National Natural Science Foundation of China (Nos. U22B2049, 62272233, 61922043, 61925204), and the Fundamental Research Funds for the Central Universities (Nos. 30922010910, 30920041109). + +# References + +[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, 2018. 1, 2, 4, 5, 6, 7, 9 +[2] Pablo Arbeláez, Michael Maire, Charless C. Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. PAMI, 33(5):898-916, 2011. 4 +[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4, 7 + +[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 1, 4 +[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 1 +[6] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In CVPR, 2022.8 +[7] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2021. 1, 3 +[8] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In CVPR, 2021. 2 +[9] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 3 +[10] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. PAMI, 38(2):295-307, 2016. 2, 4, 5 +[11] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In ECCV, 2016. 1, 2, 4, 5 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 7, 8 +[13] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In CVPR, 2021. 2, 7 +[14] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 3, 8 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2 +[16] Zibin He, Tao Dai, Jian Lu, Yong Jiang, and Shu-Tao Xia. Faked: Feature-affinity based knowledge distillation for efficient image super-resolution. In ICIP, 2020. 1, 3 +[17] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units. arXiv preprint arXiv:1606.08415, 2016. 4, 7 +[18] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 7 +[19] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 4 +[20] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multistillation network. In ACM MM, 2019. 1, 2, 4, 5, 6, 7, 9 + +[21] Andrey Ignatov, Radu Timofte, Maurizio Denna, and et al. Efficient and accurate quantized image super-resolution on Mobile NPUs, Mobile AI & AIM 2022 challenge: Report. In ECCV Workshops, 2022. 2, 3 +[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2, 8 +[23] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, 2016. 1 +[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2, 4, 5, 6 +[25] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 4 +[26] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPR Workshops, 2022. 1, 2 +[27] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017. 4, 5, 6 +[28] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. LAPAR: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. In NeurIPS, 2020. 4, 5, 6, 7, 9 +[29] Yawei Li, Kai Zhang, Luc Van Gool, Radu Timofte, et al. NTIRE 2022 challenge on efficient super-resolution: Methods and results. In CVPR Workshops, 2022. 3, 5 +[30] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In CVPR Workshops, 2022. 1, 2 +[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In ICCV Workshops, 2021. 1, 2, 3, 4 +[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPR Workshops, 2017. 1, 2, 4, 5, 6, 7 +[33] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV Workshops, 2020. 1, 2 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2 +[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In CVPR, 2022. 3 +[36] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 4 +[37] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: Efficiently amalgamated cnn-transformer architecture for mobile vision applications. In ECCV Workshops, 2022. 8 + +[38] Yusuke Matsui, Kota Ito, Yuji Aramaki, Toshihiko Yasasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. arXiv preprint arXiv:1510.04389, 2015. 4, 7 +[39] Sachin Mehta and Mohammad Rastegari. MobileViT: Lightweight, general-purpose, and mobile-friendly vision transformer. In ICLR, 2022. 3, 8 +[40] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In CVPR, 2021. 2 +[41] Pablo Navarrete Michelini, Yunhua Lu, and Xingqun Jiang. edge-SR: Super-resolution for the masses. In WACV, 2022. 1, 2 +[42] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. In CVPR, 2020. 5 +[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 7 +[44] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 1, 2, 3, 4, 5 +[45] Long Sun, Jinshan Pan, and Jinhui Tang. ShuffleMixer: An efficient convnet for image super-resolution. In NeurIPS, 2022. 1, 2, 4, 5, 6, 9 +[46] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. CVPR, 2017. 1, 2 +[47] Mingxing Tan and Quoc Le. EfficientNetV2: Smaller models and faster training. In ICML, 2021. 2, 4, 7 +[48] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, Lei Zhang, et al. NTIRE 2017 challenge on single image super-resolution: Methods and results. In CVPR Workshops, 2017. 4, 7 +[49] Longguang Wang, Xiaoyu Dong, Yingqian Wang, Xinyi Ying, Zaiping Lin, Wei An, and Yulan Guo. Exploring sparsity in image super-resolution for efficient inference. In CVPR, 2021. 5, 9 +[50] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 8 +[51] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces*, 2012. 4 +[52] Xindong Zhang, Hui Zeng, and Lei Zhang. Edge-oriented convolution block for real-time super resolution on mobile devices. In ACM MM, 2021. 1, 2, 3, 5 +[53] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 2 +[54] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCV Workshops, 2020. 1, 2, 5 \ No newline at end of file diff --git a/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/images.zip b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4c7ea26f6feec30cb8bed67183dbaa8086d1beb2 --- /dev/null +++ b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e4008ef9700d09e90f6ed35fb7e7b4d519b26ea9f37026176a6af27a8669f5c +size 1076065 diff --git a/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/layout.json b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bbb46ca261c923036c6f46867db0234102c1ca06 --- /dev/null +++ b/2023/Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution/layout.json @@ -0,0 +1,10554 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "text", + "content": "Spatially-Adaptive Feature Modulation for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 144, + 436, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 144, + 436, + 157 + ], + "spans": [ + { + "bbox": [ + 160, + 144, + 436, + 157 + ], + "type": "text", + "content": "Long Sun, Jiangxin Dong, Jinhui Tang, and Jinshan Pan*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 157, + 523, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 157, + 523, + 172 + ], + "spans": [ + { + "bbox": [ + 70, + 157, + 523, + 172 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Nanjing University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 224, + 290, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 224, + 290, + 452 + ], + "spans": [ + { + "bbox": [ + 45, + 224, + 290, + 452 + ], + "type": "text", + "content": "Although deep learning-based solutions have achieved impressive reconstruction performance in image superresolution (SR), these models are generally large, with complex architectures, making them incompatible with low-power devices with many computational and memory constraints. To overcome these challenges, we propose a spatially-adaptive feature modulation (SAFM) mechanism for efficient SR design. In detail, the SAFM layer uses independent computations to learn multi-scale feature representations and aggregates these features for dynamic spatial modulation. As the SAFM prioritizes exploiting non-local feature dependencies, we further introduce a convolutional channel mixer (CCM) to encode local contextual information and mix channels simultaneously. Extensive experimental results show that the proposed method is " + }, + { + "bbox": [ + 45, + 224, + 290, + 452 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 45, + 224, + 290, + 452 + ], + "type": "text", + "content": " smaller than state-of-the-art efficient SR methods, e.g., IMDN, and yields comparable performance with much less memory usage. Our source codes and pre-trained models are available at: https://github.com/sunny2109/SAFMN." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 460, + 128, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 128, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 128, + 472 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 479, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 586 + ], + "type": "text", + "content": "Single image super-resolution (SISR) aims to restore a high-resolution (HR) image from its low-resolution (LR) counterpart by recovering lost details. This longstanding and challenging task has recently attracted much attention due to the rapid development of streaming media or high-definition devices. As these scenarios are usually resource-constrained, it is of great interest to develop an efficient and effective SR method to estimate HR images for better visual display on these platforms or products." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 587, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 288, + 696 + ], + "type": "text", + "content": "Deep learning-based SR methods have achieved significant performance improvements with the great evolution of hardware technologies, as we can use large amounts of data to train much larger or deeper neural networks for image SR [32, 53, 31, 5]. For example, RCAN [53] is a representative CNN-based image SR network with 15.59M parameters and reaching a depth of over 400 layers. One of the most significant drawbacks of these large models is that they require high computational costs, which makes them" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 201, + 542, + 363 + ], + "blocks": [ + { + "bbox": [ + 309, + 201, + 542, + 363 + ], + "lines": [ + { + "bbox": [ + 309, + 201, + 542, + 363 + ], + "spans": [ + { + "bbox": [ + 309, + 201, + 542, + 363 + ], + "type": "image", + "image_path": "7f5f94fe3b6904c2563a45c10789fd0ac832f8a950da2e113df2b59caeb481bb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 369, + 547, + 425 + ], + "lines": [ + { + "bbox": [ + 305, + 369, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 547, + 425 + ], + "type": "text", + "content": "Figure 1. Model complexity and performance comparison between our proposed SAFMN model and other lightweight methods on Set5 [4] for " + }, + { + "bbox": [ + 305, + 369, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 305, + 369, + 547, + 425 + ], + "type": "text", + "content": " SR. Circle sizes indicate the number of parameters. The proposed method achieves a better trade-off between model complexity and reconstruction performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 435, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 546, + 555 + ], + "type": "text", + "content": "challenging to deploy. Moreover, recent visual transformers (ViTs) [12, 31, 5] outperform convolutional neural networks (CNNs) in low-level vision tasks, and their results demonstrate that exploring non-local feature interactions is essential for high-quality reconstruction. But existing self-attention mechanisms are computationally expensive and unfriendly to efficient SR design. This, therefore, motivates us to develop a lightweight yet effective model for real-world applications of image super-resolution by integrating the principles of convolution and self-attention." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "To reduce the heavy computational burden, various methods, including efficient module design [11, 41, 1, 20, 26, 33, 44, 45, 54, 30], knowledge distillation [16], neural architecture search [7], and structural re-parameterization [52], are trying to improve the efficiency of SR algorithms. Among these efficient SR models, one direction is to reduce model parameters or complexity (FLOPs). Lightweight strategies like recursive manner [23, 46], parameter sharing [1], and spare convolutions [1, 45, 30] are adopted. Although these approaches certainly reduce the model size, they usually compensate for the performance drop caused by shared recursive modules or sparse convolutions by increasing the depth or width" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13190" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "of the model, which affects the inference efficiency when performing SR reconstruction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 227 + ], + "type": "text", + "content": "Another direction is to accelerate the inference time. The post-upsampling [11, 44] is an important replacement for the pre-defined input [10, 24], which significantly speeds up the runtime. Model quantization [21] effectively accelerates latency and reduces energy consumption, particularly when deploying algorithms in edge-devices. Structural re-parameterization [52, 8] improves the speed of a well-trained model in the inference stage. These methods enjoy fast running time but poor reconstruction performance. Consequently, there is still room for a better trade-off between model efficiency and reconstruction performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 286, + 407 + ], + "type": "text", + "content": "To address the above-mentioned issues, we design a simple yet effective model by developing a spatially-adaptive feature modulation, namely SAFMN, to realize a favorable trade-off between performance and efficiency. Specifically, we first exploit non-local feature relations to dynamically select representative features (see Figure 2 and 5) by implementing a feature modulation mechanism based on multiscale representations. As the modulation mechanism processes input features from a non-local perspective, there is a requirement to complement local contextual information. To this end, we present a convolutional channel mixer based on the FMBCov [47] to encode local features and mix channels. Taken together, we find that the SAFMN network is able to achieve a better trade-off between SR performance and model complexity, as shown in Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 286, + 430 + ], + "type": "text", + "content": "The main contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 436, + 286, + 566 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 436, + 286, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 436, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 58, + 436, + 286, + 472 + ], + "type": "text", + "content": "- We develop an efficient feature modulation mechanism to learn feature dependencies, which absorbs CNN-like efficiency and transformer-like adaptability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 477, + 286, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 477, + 286, + 513 + ], + "spans": [ + { + "bbox": [ + 58, + 477, + 286, + 513 + ], + "type": "text", + "content": "- We present a compact convolutional channel mixer that simultaneously encodes local contextual information and performs channel mixing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 518, + 286, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 518, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 58, + 518, + 286, + 566 + ], + "type": "text", + "content": "- We evaluate the proposed method quantitatively and qualitatively on benchmark datasets, and the results show that our SAFMN achieves a favorable trade-off between accuracy and model complexity." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 586, + 134, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 134, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 134, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Deep Learning-based Image Super-Resolution. Classical interpolation algorithms, such as linear or bicubic upsampling, create high-resolution images by inserting zeros between adjacent pixels in the low resolution and then using a low-pass filter to preserve the content information of the input image [41]. Unlike these interpolation-based upsamplers, deep learning-based approaches learn a nonlinear mapping between the input image and the target output in an end-to-end training fashion. SRCNN [10] is the first attempt" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 359, + 116 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 359, + 116 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 359, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 359, + 116 + ], + "type": "image", + "image_path": "8a6be1bfafab898fefc3d04f472d9d70902d0229d999d1a0a4e310f1d54ec6a7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 118, + 341, + 124 + ], + "lines": [ + { + "bbox": [ + 333, + 118, + 341, + 124 + ], + "spans": [ + { + "bbox": [ + 333, + 118, + 341, + 124 + ], + "type": "text", + "content": "DI" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 127, + 342, + 135 + ], + "lines": [ + { + "bbox": [ + 332, + 127, + 342, + 135 + ], + "spans": [ + { + "bbox": [ + 332, + 127, + 342, + 135 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 140, + 545, + 238 + ], + "lines": [ + { + "bbox": [ + 304, + 140, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 545, + 238 + ], + "type": "text", + "content": "Figure 2. Comparison of local attribution maps (LAMs) [13] and diffusion indices (DIs) [13] between our SAFMN and other efficient SR models. (a) is the LR input, (b)-(e) are the results of CARN [1], EDSR-baseline [32], PAN [54] and SAFMN, respectively. The LAM results denote the importance of each pixel in the input LR image when super-resolving the patch marked with a red box. The DI value reflects the range of involved pixels. A larger DI value means a wider range of attention. The proposed method can exploit more feature information." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 360, + 72, + 403, + 116 + ], + "blocks": [ + { + "bbox": [ + 360, + 72, + 403, + 116 + ], + "lines": [ + { + "bbox": [ + 360, + 72, + 403, + 116 + ], + "spans": [ + { + "bbox": [ + 360, + 72, + 403, + 116 + ], + "type": "image", + "image_path": "30acbc6abd5027152737258acc386d49e5d935ca744b4b69563848a6b88583c8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 118, + 392, + 124 + ], + "lines": [ + { + "bbox": [ + 372, + 118, + 392, + 124 + ], + "spans": [ + { + "bbox": [ + 372, + 118, + 392, + 124 + ], + "type": "text", + "content": "3.2180" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 127, + 388, + 135 + ], + "lines": [ + { + "bbox": [ + 378, + 127, + 388, + 135 + ], + "spans": [ + { + "bbox": [ + 378, + 127, + 388, + 135 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 406, + 72, + 449, + 116 + ], + "blocks": [ + { + "bbox": [ + 406, + 72, + 449, + 116 + ], + "lines": [ + { + "bbox": [ + 406, + 72, + 449, + 116 + ], + "spans": [ + { + "bbox": [ + 406, + 72, + 449, + 116 + ], + "type": "image", + "image_path": "c13192d0602280aaff706dd38fe8e389fce80cc98dd7e5214ca12d8aa7b863e7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 118, + 437, + 124 + ], + "lines": [ + { + "bbox": [ + 419, + 118, + 437, + 124 + ], + "spans": [ + { + "bbox": [ + 419, + 118, + 437, + 124 + ], + "type": "text", + "content": "5.1936" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 127, + 433, + 135 + ], + "lines": [ + { + "bbox": [ + 424, + 127, + 433, + 135 + ], + "spans": [ + { + "bbox": [ + 424, + 127, + 433, + 135 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 452, + 72, + 495, + 116 + ], + "blocks": [ + { + "bbox": [ + 452, + 72, + 495, + 116 + ], + "lines": [ + { + "bbox": [ + 452, + 72, + 495, + 116 + ], + "spans": [ + { + "bbox": [ + 452, + 72, + 495, + 116 + ], + "type": "image", + "image_path": "2e67f2fbc1dd450fa7088d7a56e0dd2173494720f44016025753f5788df065d9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 118, + 484, + 124 + ], + "lines": [ + { + "bbox": [ + 465, + 118, + 484, + 124 + ], + "spans": [ + { + "bbox": [ + 465, + 118, + 484, + 124 + ], + "type": "text", + "content": "4.0629" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 127, + 479, + 135 + ], + "lines": [ + { + "bbox": [ + 470, + 127, + 479, + 135 + ], + "spans": [ + { + "bbox": [ + 470, + 127, + 479, + 135 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 499, + 72, + 543, + 116 + ], + "blocks": [ + { + "bbox": [ + 499, + 72, + 543, + 116 + ], + "lines": [ + { + "bbox": [ + 499, + 72, + 543, + 116 + ], + "spans": [ + { + "bbox": [ + 499, + 72, + 543, + 116 + ], + "type": "image", + "image_path": "d4f6e8dd8f63302bb3ef3dd889913f61010d21db5dec58e2aa426494d618784e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 510, + 118, + 531, + 124 + ], + "lines": [ + { + "bbox": [ + 510, + 118, + 531, + 124 + ], + "spans": [ + { + "bbox": [ + 510, + 118, + 531, + 124 + ], + "type": "text", + "content": "10.4295" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 517, + 127, + 525, + 134 + ], + "lines": [ + { + "bbox": [ + 517, + 127, + 525, + 134 + ], + "spans": [ + { + "bbox": [ + 517, + 127, + 525, + 134 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 247, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 555 + ], + "type": "text", + "content": "to use a convolutional neural network to tackle the image SR problem and achieves a considerable performance gain compared with conventional methods. Since then, many improvements have been proposed. VDSR [24] uses global residual learning [15] to solve the problem of difficulty in training an SR model with deep layers. DRRN [46] integrates the local residual learning and global residual connection to ease the training difficulty and enhance high-frequency details. EDSR [32] further increases the model footprint to 43M, achieving a significant breakthrough in reconstruction performance and showing that the BatchNorm (BN) [22] layer is not necessary for the SR task. RCAN [53] builds a more than 400 layers model based on channel attention and dense connections for accurate SR. NLSA [40] utilizes a non-local sparse attention to enhance the feature modeling capability and reaches state-of-the-art performance. With the successful application of ViT in various high-level vision tasks, image SR also follows this ViT [12] framework, achieving higher reconstruction performance than CNN-based models on extensive public benchmarks. For instance, SwinIR [31] acts as a strong baseline for image restoration tasks based on the Swin Transformer [34]. While these approaches achieve impressive reconstruction performance, the required high computational costs make them challenging to deploy in real-world applications on resource-constrained devices." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Efficient Image Super-Resolution. To improve the model efficiency, many CNN-based SR works try to alleviate this issue. FSRCNN [11] and ESPCN [44] utilize a postupsampling manner to reduce the computational burden from the pre-defined inputs significantly. CARN [1] uses group convolutions and a cascading mechanism upon residual networks to improve efficiency. IMDN [20] adopts feature splitting and concatenation operations to progressively aggregate features, and its improved variants [33, 26] won the AIM2020 and NTIRE2022 Efficient SR challenge. ShuffleMixer [45] introduces a large kernel convolution for lightweight SR design. BSRN [30] employs blueprint-separable convolutions to reduce model complexity. Mean" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13191" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 547, + 178 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 547, + 178 + ], + "type": "image", + "image_path": "393c8fd6b666674a4e818e2c5fbfd0168d7f4e5b4b09f2957d58231ad3184e63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 236 + ], + "type": "text", + "content": "Figure 3. An overview of the proposed SAFMN. SAFMN first transforms the input LR image into the feature space using a convolutional layer, performs feature extraction using a series of feature mixing modules (FMMs), and then reconstructs these extracted features by an upsampler module. The FMM block is implemented by a spatially-adaptive feature modulation (SAFM) layer, a convolutional channel mixer (CCM) and two skip-connections." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 245, + 289, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 289, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 289, + 401 + ], + "type": "text", + "content": "while, an increasingly popular direction is to compress or accelerate a well-trained deep model through model quantization [21], structural re-parameterization [52] or knowledge distillation [16]. Neural architecture search (NAS) is also commonly used to search a well-constrained architecture for image super-resolution [7]. Note that the efficiency of a deep neural network could be measured in different metrics, including the number of parameters, FLOPs, activations, memory consumption and inference running time [29]. Although the above approaches have been improved in different efficiency aspects, there is still room for a favorable trade-off between reconstruction performance and model efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 413, + 153, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 153, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 153, + 426 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "type": "text", + "content": "In this section, we present the core components of our proposed model for efficient SISR. As shown in Figure 3, the network consists of the following parts: a shallow convolution, a stacking of feature mixing modules (FMMs), and an upsampler layer. Specifically, we first apply a " + }, + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "type": "text", + "content": " convolution layer to transform the input LR image to feature space and generate the shallow feature " + }, + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0" + }, + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "type": "text", + "content": ". Then, the multiple stacked FMMs are used to generate finer deep features from " + }, + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0" + }, + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "type": "text", + "content": " for HR image reconstruction, where an FMM layer has a spatially-adaptive feature modulation (SAFM) layer and a convolutional channel mixer (CCM). To recover the HR target image, we introduce a global residual connection to learn high-frequency details and employ a lightweight upsampling layer for fast reconstruction, which only contains a " + }, + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 434, + 287, + 615 + ], + "type": "text", + "content": " convolution and a pixel-shuffle layer [44]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 622, + 251, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 251, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 251, + 635 + ], + "type": "text", + "content": "3.1. Spatially-adaptive Feature Modulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Recent studies [9, 35, 14, 39, 31] suggest that the notable performance of ViTs across diverse tasks stems from their implementation of the key multi-head self-attention (MHSA) mechanism. This mechanism enables the model with the capability of long-range feature interaction and dynamic spatial weighting, both of which contribute to pro" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 245, + 538, + 325 + ], + "blocks": [ + { + "bbox": [ + 315, + 245, + 538, + 325 + ], + "lines": [ + { + "bbox": [ + 315, + 245, + 538, + 325 + ], + "spans": [ + { + "bbox": [ + 315, + 245, + 538, + 325 + ], + "type": "image", + "image_path": "6320c5b4e1c0681f68965f91caa103296b32cdac5aa5978562909a3ea02eec1b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 327, + 546, + 360 + ], + "lines": [ + { + "bbox": [ + 305, + 327, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 546, + 360 + ], + "type": "text", + "content": "Figure 4. Network architecture of the proposed SAFM module. This module performs spatially-adaptive modulation on the input features. Please see the Section 3.1 for more details." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 366, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 449 + ], + "type": "text", + "content": "ducing promising results. But existing self-attention variants are computationally expensive and unfriendly to efficient SR design. In contrast, standard convolution is an efficient operation, but its performance is limited by the static weights and locality. This observation inspires us to borrow the idea of MHSA to enhance the representation capability of convolution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 451, + 546, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 546, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 546, + 524 + ], + "type": "text", + "content": "To introduce the ability of long-range interaction and dynamic modeling into convolution, we follow the multi-head paradigm, using parallel and independent computations that allow each head to process different scale information of the input, and then aggregate these features to generate an attention map for spatially-adaptive feature modulation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 525, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 547, + 657 + ], + "type": "text", + "content": "The detailed architecture of our proposed SAFM is depicted in Figure 4. Specifically, we first split the normalized input features into four-group components, and feed them into a multi-scale feature generation unit (MFGU), where a " + }, + { + "bbox": [ + 304, + 525, + 547, + 657 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 525, + 547, + 657 + ], + "type": "text", + "content": " depth-wise convolution processes the first one, and the rest parts are sampled individually by pooling operations. As we desire to select discriminative features towards learning non-local feature interactions, adaptive max pooling operators are applied over the input features to gather information. Given the input feature " + }, + { + "bbox": [ + 304, + 525, + 547, + 657 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 525, + 547, + 657 + ], + "type": "text", + "content": ", this procedure can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 330, + 666, + 455, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 666, + 455, + 680 + ], + "spans": [ + { + "bbox": [ + 330, + 666, + 455, + 680 + ], + "type": "interline_equation", + "content": "\\left[ X _ {0}, X _ {1}, X _ {2}, X _ {3} \\right] = \\operatorname {S p l i t} (X),", + "image_path": "74909055bac72f679b245f61a8d83746646a339504821e9c88d1cdd550eef580.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 330, + 682, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 682, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 330, + 682, + 545, + 696 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {0} = \\operatorname {D W - C o n v} _ {3 \\times 3} \\left(X _ {0}\\right), \\tag {1}", + "image_path": "e66bc6c08025f028f1a98ffdff0dc01950db63ab6d2ffd879b2344d290c44675.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 700, + 521, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 700, + 521, + 717 + ], + "spans": [ + { + "bbox": [ + 330, + 700, + 521, + 717 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {i} = \\uparrow_ {p} \\left(\\operatorname {D W - C o n v} _ {3 \\times 3} \\left(\\downarrow_ {\\frac {p}{2 ^ {i}}} \\left(X _ {i}\\right)\\right)\\right), 1 \\leq i \\leq 3,", + "image_path": "196397f2f3b00b79b0ab67d2fbdda0166f7744065e77672f82b99ea3fca1760b.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "13192" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "inline_equation", + "content": "\\text{Split}(\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "text", + "content": " corresponds to the channel split operation, " + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "inline_equation", + "content": "\\text{DW-Conv}_{3\\times 3}(\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "text", + "content": " depth-wise convolution, " + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "inline_equation", + "content": "\\uparrow_p(\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "text", + "content": " represents upsampling features at a specific level to the original resolution " + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "text", + "content": " via nearest interpolation for fast implementation, and " + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "inline_equation", + "content": "\\downarrow_{\\frac{p}{2^i}}" + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "text", + "content": " denotes pooling the input features to the size of " + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "inline_equation", + "content": "\\frac{p}{2^i}" + }, + { + "bbox": [ + 47, + 72, + 286, + 169 + ], + "type": "text", + "content": ". Figure 5(b)-(e) show that using such a design, we can obtain varying features from different representation subspaces." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 170, + 286, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 170, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 286, + 216 + ], + "type": "text", + "content": "Afterward, we aggregate these extracted short- or long-range features by concatenating them on the channel dimension and performing a " + }, + { + "bbox": [ + 47, + 170, + 286, + 216 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 170, + 286, + 216 + ], + "type": "text", + "content": " convolution. It can be formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 224, + 287, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 224, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 74, + 224, + 287, + 245 + ], + "type": "interline_equation", + "content": "\\hat {X} = \\operatorname {C o n v} _ {1 \\times 1} \\left(\\operatorname {C o n c a t} \\left([ \\hat {X} _ {0}, \\hat {X} _ {1}, \\hat {X} _ {2}, \\hat {X} _ {3} ]\\right)\\right), \\tag {2}", + "image_path": "e0e9c33820d898a856e2c8d9afbd069b776c992de1dc35f9b2524e12a50f1c70.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "inline_equation", + "content": "\\operatorname{Concat}(\\cdot)" + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "text", + "content": " denotes the concatenation operation, and " + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "inline_equation", + "content": "\\operatorname{Conv}_{1\\times 1}(\\cdot)" + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "text", + "content": " convolution. After obtaining the aggregated representation " + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "text", + "content": " (see Figure 5(f)), we normalize it through a GELU non-linearity [17] to estimate the attention map and adaptively modulate the input " + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 253, + 286, + 335 + ], + "type": "text", + "content": " according to the estimated attention via element-wise product. This process can be written as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 346, + 287, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 346, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 131, + 346, + 287, + 360 + ], + "type": "interline_equation", + "content": "\\bar {X} = \\phi (\\hat {X}) \\odot X, \\tag {3}", + "image_path": "58a9dfcc6fee71e66bb081edb3dd521e34b09854826c53b2cc20b531f231a028.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 369, + 286, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 286, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 286, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 369, + 286, + 405 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 47, + 369, + 286, + 405 + ], + "type": "text", + "content": " represents the GELU function and " + }, + { + "bbox": [ + 47, + 369, + 286, + 405 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 369, + 286, + 405 + ], + "type": "text", + "content": " is the element-wise product. The benefits of those operators are compared and results are shown in Table 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 413, + 180, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 180, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 180, + 426 + ], + "type": "text", + "content": "3.2. Feature Mixing Module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "type": "text", + "content": "As mentioned previously, the SAFM module prioritizes exploiting non-local feature interactions. To further incorporate local contextual information and perform channel mixing simultaneously, we introduce a compact convolutional channel mixer (CCM) based on the FMBCnv [47]. The CCM comprises a " + }, + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "type": "text", + "content": " convolution and a " + }, + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "type": "text", + "content": " convolution. Within this, the first " + }, + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "type": "text", + "content": " convolution encodes the spatially local contexts and doubles the number of channels of the input features for mixing channels; the later " + }, + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 431, + 286, + 574 + ], + "type": "text", + "content": " convolution reduces the channels back to the original input dimension. A GELU [17] function is applied to the hidden layer for non-linear mapping." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 574, + 286, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 286, + 609 + ], + "type": "text", + "content": "We formulate the proposed SAFM and CCM into a unified feature mixing module (FMM) to select representative features. The FMM can be written as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 617, + 286, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 617, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 110, + 617, + 286, + 644 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Y = \\operatorname {S A F M} (\\operatorname {L N} (X)) + X, \\\\ Z = \\operatorname {C C M} (\\operatorname {L N} (Y)) + Y, \\end{array} \\tag {4}", + "image_path": "46f044823cf3a3516e077b6520dc78df248054c3f8356098ea10e2d6cb95d6d8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "type": "text", + "content": " is the LayerNorm [3] layer, " + }, + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 47, + 653, + 286, + 712 + ], + "type": "text", + "content": " are the intermediate features. The additional residual learning is employed to stabilize training processes and learn high frequency details to facilitate high-quality image reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 400, + 159 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 400, + 159 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 400, + 159 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 400, + 159 + ], + "type": "image", + "image_path": "b45186fcca63868e4a60edf820afb6e75a97b394336a82cb8d6e3e295e00402d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 159, + 375, + 167 + ], + "lines": [ + { + "bbox": [ + 339, + 159, + 375, + 167 + ], + "spans": [ + { + "bbox": [ + 339, + 159, + 375, + 167 + ], + "type": "text", + "content": "(a) LR input" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 170, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 305, + 170, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 545, + 213 + ], + "type": "text", + "content": "Figure 5. Illustration of learned deep features from the SAFM module. (a) is the LR input. (b)-(e) are features learned by the multi-scale feature generation unit. (f) is the results of aggregating features (b)-(e). (g) denotes the attended output of the SAFM." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 406, + 72, + 447, + 110 + ], + "blocks": [ + { + "bbox": [ + 406, + 72, + 447, + 110 + ], + "lines": [ + { + "bbox": [ + 406, + 72, + 447, + 110 + ], + "spans": [ + { + "bbox": [ + 406, + 72, + 447, + 110 + ], + "type": "image", + "image_path": "da54769aa655ccede61a63ec8b80831fc3be69faa912f41969987d152a73dc9c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 453, + 72, + 493, + 110 + ], + "blocks": [ + { + "bbox": [ + 453, + 72, + 493, + 110 + ], + "lines": [ + { + "bbox": [ + 453, + 72, + 493, + 110 + ], + "spans": [ + { + "bbox": [ + 453, + 72, + 493, + 110 + ], + "type": "image", + "image_path": "2daa8534748e0c099c21c63eaefee922dbe6cbd0a131557e011047cad6bd5b25.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 501, + 72, + 541, + 110 + ], + "blocks": [ + { + "bbox": [ + 501, + 72, + 541, + 110 + ], + "lines": [ + { + "bbox": [ + 501, + 72, + 541, + 110 + ], + "spans": [ + { + "bbox": [ + 501, + 72, + 541, + 110 + ], + "type": "image", + "image_path": "f986e3064fc4a1f5a8ba0ffea38aa718d3edd04fd8814d17a25c53e33473dc0d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 406, + 120, + 447, + 159 + ], + "blocks": [ + { + "bbox": [ + 410, + 112, + 442, + 119 + ], + "lines": [ + { + "bbox": [ + 410, + 112, + 442, + 119 + ], + "spans": [ + { + "bbox": [ + 410, + 112, + 442, + 119 + ], + "type": "text", + "content": "(b) Scale 1" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 120, + 447, + 159 + ], + "lines": [ + { + "bbox": [ + 406, + 120, + 447, + 159 + ], + "spans": [ + { + "bbox": [ + 406, + 120, + 447, + 159 + ], + "type": "image", + "image_path": "ed4d6368d75c596548d09b452c59bffa36d273764cc35a515a9891acee9cc800.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 159, + 442, + 167 + ], + "lines": [ + { + "bbox": [ + 411, + 159, + 442, + 167 + ], + "spans": [ + { + "bbox": [ + 411, + 159, + 442, + 167 + ], + "type": "text", + "content": "(e) Scale 8" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 453, + 120, + 493, + 159 + ], + "blocks": [ + { + "bbox": [ + 458, + 112, + 489, + 119 + ], + "lines": [ + { + "bbox": [ + 458, + 112, + 489, + 119 + ], + "spans": [ + { + "bbox": [ + 458, + 112, + 489, + 119 + ], + "type": "text", + "content": "(c) Scale 2" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 453, + 120, + 493, + 159 + ], + "lines": [ + { + "bbox": [ + 453, + 120, + 493, + 159 + ], + "spans": [ + { + "bbox": [ + 453, + 120, + 493, + 159 + ], + "type": "image", + "image_path": "be287c37315a90c066b9eeeb172012e46548abba9841e76ac836fc2d6ec71b94.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 159, + 494, + 167 + ], + "lines": [ + { + "bbox": [ + 452, + 159, + 494, + 167 + ], + "spans": [ + { + "bbox": [ + 452, + 159, + 494, + 167 + ], + "type": "text", + "content": "(f) Aggregated" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 501, + 120, + 541, + 159 + ], + "blocks": [ + { + "bbox": [ + 504, + 112, + 537, + 119 + ], + "lines": [ + { + "bbox": [ + 504, + 112, + 537, + 119 + ], + "spans": [ + { + "bbox": [ + 504, + 112, + 537, + 119 + ], + "type": "text", + "content": "(d) Scale 4" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 501, + 120, + 541, + 159 + ], + "lines": [ + { + "bbox": [ + 501, + 120, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 501, + 120, + 541, + 159 + ], + "type": "image", + "image_path": "33d97f925d1edad43e51fd429aeacc579f0298e50710996fca4b079542ceb7dd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 159, + 539, + 167 + ], + "lines": [ + { + "bbox": [ + 503, + 159, + 539, + 167 + ], + "spans": [ + { + "bbox": [ + 503, + 159, + 539, + 167 + ], + "type": "text", + "content": "(g) Attended" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 223, + 430, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 223, + 430, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 223, + 430, + 236 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 243, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 279 + ], + "type": "text", + "content": "In this section, we perform quantitative and qualitative evaluations with state-of-the-art algorithms to demonstrate the effectiveness of the proposed method." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 286, + 462, + 299 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 462, + 299 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 462, + 299 + ], + "type": "text", + "content": "4.1. Dataset and Implementation" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 304, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 424 + ], + "type": "text", + "content": "Datasets. Following previous works [28, 31, 45], we use DIV2K [48] and Flickr2K [32] as the training data and generate LR images by applying the bicubic downscaling to reference HR images. We use five commonly used benchmark datasets including Set5 [4], Set14 [51], B100 [2], Urban100 [19] and Manga109 [38] as test data. We use the peak signal to noise ratio (PSNR) and the structural similarity index (SSIM) to evaluate the quality of the restored images. All PSNR and SSIM values are calculated on the Y channel of images transformed to YCbCr color space." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "text", + "content": "Implementation details. During the training, the data argumentation is performed on the input patches with random horizontal flips and rotations. In addition, we randomly crop 64 patches of size " + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "text", + "content": " pixels from LR images as the basic training inputs. The number of FMM and feature channels is set to 8 and 36, respectively. We use the Adam [25] optimizer with " + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "text", + "content": " to solve the proposed model. The number of iterations is set to 500,000. We set the initial learning rate to " + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "text", + "content": " and the minimum one to " + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 305, + 425, + 545, + 628 + ], + "type": "text", + "content": ", which is updated by the Cosine Annealing scheme [36]. We use a combination of mean absolute error (MAE) loss and an FFT-based frequency loss function to constrain the model training, which is the same as [45]. All experiments are conducted with the PyTorch framework on an NVIDIA GeForce RTX 3090 GPU. Due to the page limit, we include more comparison results in the supplemental material." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 635, + 535, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 535, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 535, + 647 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": "Quantitative comparisons. To evaluate the performance of the proposed approach, we compare it with state-of-the-art lightweight SR methods on different scaling factors, including SRCNN [10], FSRCNN [11], ESPCN [44], VDSR [24], LapSRN [27], CARN [1], EDSR-baseline [32], IMDN [20]," + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13193" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 121, + 541, + 544 + ], + "blocks": [ + { + "bbox": [ + 46, + 68, + 547, + 114 + ], + "lines": [ + { + "bbox": [ + 46, + 68, + 547, + 114 + ], + "spans": [ + { + "bbox": [ + 46, + 68, + 547, + 114 + ], + "type": "text", + "content": "Table 1. Comparisons of efficient SR networks on the commonly used benchmark datasets. All PSNR/SSIM results are calculated on the Y-channel. #Acts represents all elements of the output of convolutional layers. #FLOPs and #Acts are measured corresponding to an HR image of the size " + }, + { + "bbox": [ + 46, + 68, + 547, + 114 + ], + "type": "inline_equation", + "content": "{1280} \\times {720}" + }, + { + "bbox": [ + 46, + 68, + 547, + 114 + ], + "type": "text", + "content": " pixels. Red color denotes the best performance. Blanked entries indicate results not reported or not available from previous work." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 121, + 541, + 544 + ], + "lines": [ + { + "bbox": [ + 55, + 121, + 541, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 541, + 544 + ], + "type": "table", + "html": "
MethodsScale#Params [K]#FLOPs [G]#Acts [M]Set5Set14B100Urban100Manga109
Bicubic×2---33.66/0.929930.24/0.868829.56/0.843126.88/0.840330.80/0.9339
SRCNN [10]57538936.66/0.954232.42/0.906331.36/0.887929.50/0.894635.74/0.9661
FSRCNN [11]1264137.00/0.955832.63/0.908831.53/0.892029.88/0.902036.67/0.9694
ESPCN [44]2152336.83/0.956432.40/0.909631.29/0.891729.48/0.8975-
VDSR [24]6656131,12037.53/0.958733.03/0.912431.90/0.896030.76/0.914037.22/0.9729
LapSRN [27]8133022337.52/0.959033.08/0.913031.80/0.895030.41/0.910037.27/0.9740
CARN-M [1]4159165537.53/0.958333.26/0.914131.92/0.896031.23/0.9193-
CARN [1]1,59222352237.76/0.959033.52/0.916632.09/0.897831.92/0.9256-
EDSR-baseline [32]1,37031656337.99/0.960433.57/0.917532.16/0.899431.98/0.927238.54/0.9769
IMDN [20]69415942338.00/0.960533.63/0.917732.19/0.899632.17/0.928338.88/0.9774
PAN [54]2617167738.00/0.960533.59/0.918132.18/0.899732.01/0.927338.70/0.9773
LAPAR-A [28]54817165638.01/0.960533.62/0.918332.19/0.899932.10/0.928338.67/0.9772
ECBSR-M16C64 [52]59613725237.90/0.961533.34/0.917832.10/0.901831.71/0.9250-
SMSR [49]985132-38.00/0.960133.64/0.917932.17/0.899032.19/0.928438.76/0.9771
ShuffleMixer [45]3949183238.01/0.960633.63/0.918032.17/0.899531.89/0.925738.83/0.9774
SAFMN (Ours)2285229938.00/0.960533.54/0.917732.16/0.899531.84/0.925638.71/0.9771
Bicubic×3---30.39/0.868227.55/0.774227.21/0.738524.46/0.734926.95/0.8556
SRCNN [10]57538932.75/0.909029.28/0.820928.41/0.786326.24/0.798930.59/0.9107
FSRCNN [11]1251933.16/0.914029.43/0.824228.53/0.791026.43/0.808030.98/0.9212
VDSR [24]6656131,12033.66/0.921329.77/0.831428.82/0.797627.14/0.827932.01/0.9310
CARN-M [1]4154632733.99/0.923630.08/0.836728.91/0.800027.55/0.8385-
CARN [1]1,59211926834.29/0.925530.29/0.840729.06/0.803428.06/0.8493-
EDSR-baseline [32]1,55516028534.37/0.927030.28/0.841729.09/0.805228.15/0.852733.45/0.9439
IMDN [20]7037219034.36/0.927030.32/0.841729.09/0.804628.17/0.851933.61/0.9445
PAN [54]2613934034.40/0.927130.36/0.842329.11/0.805028.11/0.851133.61/0.9448
LAPAR-A [28]59411450534.36/0.926730.34/0.842129.11/0.805428.15/0.852333.51/0.9441
SMSR [49]99368-34.40/0.927030.33/0.841229.10/0.805028.25/0.853633.68/0.9445
ShuffleMixer [45]4154340434.40/0.927230.37/0.842329.12/0.805128.08/0.849833.69/0.9448
SAFMN (Ours)2332313434.34/0.926730.33/0.841829.08/0.804827.95/0.847433.52/0.9437
Bicubic×4---28.42/0.810426.00/0.702725.96/0.667523.14/0.657724.89/0.7866
SRCNN [10]57538930.48/0.862827.49/0.750326.90/0.710124.52/0.722127.66/0.8505
FSRCNN [11]1251130.71/0.865727.59/0.753526.98/0.715024.62/0.728027.90/0.8517
ESPCN [44]251630.52/0.869727.42/0.760626.87/0.721624.39/0.7241-
VDSR [24]6656131,12031.35/0.883828.01/0.767427.29/0.725125.18/0.752428.83/0.8809
LapSRN [27]81314926431.54/0.885028.19/0.772027.32/0.728025.21/0.756029.09/0.8845
CARN-M [1]4153322731.92/0.890328.42/0.776227.44/0.730425.62/0.7694-
CARN [1]1,5929119432.13/0.893728.60/0.780627.58/0.734926.07/0.7837-
EDSR-baseline [32]1,51811420232.09/0.893828.58/0.781327.57/0.735726.04/0.784930.35/0.9067
IMDN [20]7154110832.21/0.894828.58/0.781127.56/0.735326.04/0.783830.45/0.9075
PAN [54]2722823832.13/0.894828.61/0.782227.59/0.736326.11/0.785430.51/0.9095
LAPAR-A [28]6599445232.15/0.894428.61/0.781827.61/0.736626.14/0.787130.42/0.9074
ECBSR-M16C64 [52]603356431.92/0.894628.34/0.781727.48/0.739325.81/0.7773-
SMSR [49]100642-32.12/0.893228.55/0.780827.55/0.735126.11/0.786830.54/0.9085
ShuffleMixer [45]4112826932.21/0.895328.66/0.782727.61/0.736626.08/0.783530.65/0.9093
SAFMN (Ours)240147732.18/0.894828.60/0.781327.58/0.735925.97/0.780930.43/0.9063
", + "image_path": "b3c1623dfa780fe64348dbbcf3b1f053f6014fe44744bc8cc5af525b43ffffca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 554, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 578 + ], + "type": "text", + "content": "PAN [54], LAPAR [28], ECBSR [52], SMSR [49], and ShuffleMixer [45]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "type": "text", + "content": "The quantitative comparisons on benchmark datasets for the upscaling factors of " + }, + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "type": "text", + "content": " are reported in Table 1. In addition to PSNR/SSIM metrics, we list the number of parameters (#Params), FLOPs (#FLOPs) and activations (#Acts). We calculate the number of FLOPs and activations with the fvcore1 library under a setting of super-resolving an LR image to " + }, + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 46, + 578, + 288, + 687 + ], + "type": "text", + "content": " pixels. Among these metrics, #Params and #Acts are linked to memory consumption, and #FLOPs is related to energy usage. In particular," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 554, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 554, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 554, + 545, + 590 + ], + "type": "text", + "content": "Acts is a better metric for measuring the efficiency of a model than the number of parameters and FLOPs as suggested in recent works [42, 52, 29]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Benefiting from the simple yet efficient structure, the proposed SAFMN obtains comparable performance with significantly fewer parameters and memory consumption. For instance, the " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " SR results on the B100 dataset reveal that our SAFMN has parameters about " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " less than the CARN [1], " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " less than the IMDN [20], and " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " less than the ShuffleMixer [45]. As for activations, we have " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " fewer than them, respectively. While our model has a smaller footprint, we achieve similar performance among these methods. Moreover, we compare" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "text", + "content": "1We use the fvcore.nn.flop_count_str command to calculate the number of parameters, FLOPs and activations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "13194" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 231, + 187 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 231, + 187 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 231, + 187 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 231, + 187 + ], + "type": "image", + "image_path": "f95ca739a69a4cdd8c4742c439b180c6e154276564b0f83236be35c2f9653b4b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 189, + 181, + 198 + ], + "lines": [ + { + "bbox": [ + 103, + 189, + 181, + 198 + ], + "spans": [ + { + "bbox": [ + 103, + 189, + 181, + 198 + ], + "type": "text", + "content": "img024 from Urban100" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 232, + 70, + 310, + 125 + ], + "blocks": [ + { + "bbox": [ + 232, + 70, + 310, + 125 + ], + "lines": [ + { + "bbox": [ + 232, + 70, + 310, + 125 + ], + "spans": [ + { + "bbox": [ + 232, + 70, + 310, + 125 + ], + "type": "image", + "image_path": "5f08c0aad803c7c0e0f51133f12e5ab352dc5c33c9205208e4768847194cf5fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 125, + 292, + 135 + ], + "lines": [ + { + "bbox": [ + 250, + 125, + 292, + 135 + ], + "spans": [ + { + "bbox": [ + 250, + 125, + 292, + 135 + ], + "type": "text", + "content": "(a) HR patch" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 390, + 125 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 390, + 125 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 390, + 125 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 390, + 125 + ], + "type": "image", + "image_path": "db4592ed12cb81241769af5a9a5d57bb3ed6b21fb98a4af844d09d7f342c54f7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 125, + 369, + 134 + ], + "lines": [ + { + "bbox": [ + 332, + 125, + 369, + 134 + ], + "spans": [ + { + "bbox": [ + 332, + 125, + 369, + 134 + ], + "type": "text", + "content": "(b) Bicubic" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 392, + 70, + 470, + 125 + ], + "blocks": [ + { + "bbox": [ + 392, + 70, + 470, + 125 + ], + "lines": [ + { + "bbox": [ + 392, + 70, + 470, + 125 + ], + "spans": [ + { + "bbox": [ + 392, + 70, + 470, + 125 + ], + "type": "image", + "image_path": "10d9983022c13b989d58fdb4f73a23d2af3c7e30e1620f9385c84323ae25a849.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 125, + 455, + 135 + ], + "lines": [ + { + "bbox": [ + 406, + 125, + 455, + 135 + ], + "spans": [ + { + "bbox": [ + 406, + 125, + 455, + 135 + ], + "type": "text", + "content": "(c) VDSR [24]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 472, + 70, + 549, + 125 + ], + "blocks": [ + { + "bbox": [ + 472, + 70, + 549, + 125 + ], + "lines": [ + { + "bbox": [ + 472, + 70, + 549, + 125 + ], + "spans": [ + { + "bbox": [ + 472, + 70, + 549, + 125 + ], + "type": "image", + "image_path": "e8bad97d5cf8b0a62f617e00ed6a09482c4f3f0cd18dec9190e3a9f26ee12c9e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 125, + 544, + 135 + ], + "lines": [ + { + "bbox": [ + 476, + 125, + 544, + 135 + ], + "spans": [ + { + "bbox": [ + 476, + 125, + 544, + 135 + ], + "type": "text", + "content": "(d) ShuffleMixer [45]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 233, + 135, + 310, + 187 + ], + "blocks": [ + { + "bbox": [ + 233, + 135, + 310, + 187 + ], + "lines": [ + { + "bbox": [ + 233, + 135, + 310, + 187 + ], + "spans": [ + { + "bbox": [ + 233, + 135, + 310, + 187 + ], + "type": "image", + "image_path": "38b79e2d65202dd949a67dd38b8d5f5c749ece62ba1cc2efbbb4e123c4fc758b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 189, + 299, + 198 + ], + "lines": [ + { + "bbox": [ + 244, + 189, + 299, + 198 + ], + "spans": [ + { + "bbox": [ + 244, + 189, + 299, + 198 + ], + "type": "text", + "content": "(e) LapSRN [27]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 135, + 390, + 187 + ], + "blocks": [ + { + "bbox": [ + 312, + 135, + 390, + 187 + ], + "lines": [ + { + "bbox": [ + 312, + 135, + 390, + 187 + ], + "spans": [ + { + "bbox": [ + 312, + 135, + 390, + 187 + ], + "type": "image", + "image_path": "61fda92f8777e8bc55c4c5e3d5ed4061cfcfc40a46e56dd137576598147b05f5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 189, + 373, + 198 + ], + "lines": [ + { + "bbox": [ + 329, + 189, + 373, + 198 + ], + "spans": [ + { + "bbox": [ + 329, + 189, + 373, + 198 + ], + "type": "text", + "content": "(f) CARN [1]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 392, + 135, + 470, + 187 + ], + "blocks": [ + { + "bbox": [ + 392, + 135, + 470, + 187 + ], + "lines": [ + { + "bbox": [ + 392, + 135, + 470, + 187 + ], + "spans": [ + { + "bbox": [ + 392, + 135, + 470, + 187 + ], + "type": "image", + "image_path": "07e881585ecd68c4795f71db766598c00e10ed81f8162b5299c6b104296572d4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 189, + 455, + 198 + ], + "lines": [ + { + "bbox": [ + 406, + 189, + 455, + 198 + ], + "spans": [ + { + "bbox": [ + 406, + 189, + 455, + 198 + ], + "type": "text", + "content": "(g) IMDN [20]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 472, + 135, + 549, + 187 + ], + "blocks": [ + { + "bbox": [ + 472, + 135, + 549, + 187 + ], + "lines": [ + { + "bbox": [ + 472, + 135, + 549, + 187 + ], + "spans": [ + { + "bbox": [ + 472, + 135, + 549, + 187 + ], + "type": "image", + "image_path": "caf2e5c4e65179ea6cd14438872d6100b484045b74b75eee5f0da56f3c0b7f4e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 189, + 529, + 198 + ], + "lines": [ + { + "bbox": [ + 489, + 189, + 529, + 198 + ], + "spans": [ + { + "bbox": [ + 489, + 189, + 529, + 198 + ], + "type": "text", + "content": "(h) SAFMN" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 53, + 207, + 230, + 323 + ], + "blocks": [ + { + "bbox": [ + 53, + 207, + 230, + 323 + ], + "lines": [ + { + "bbox": [ + 53, + 207, + 230, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 207, + 230, + 323 + ], + "type": "image", + "image_path": "6f7c923d0205866c8c827499c5d78cb70f1963b9d5fa0258b78f8a0e3b762da5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 325, + 181, + 335 + ], + "lines": [ + { + "bbox": [ + 103, + 325, + 181, + 335 + ], + "spans": [ + { + "bbox": [ + 103, + 325, + 181, + 335 + ], + "type": "text", + "content": "img062 from Urban100" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 232, + 206, + 310, + 261 + ], + "blocks": [ + { + "bbox": [ + 232, + 206, + 310, + 261 + ], + "lines": [ + { + "bbox": [ + 232, + 206, + 310, + 261 + ], + "spans": [ + { + "bbox": [ + 232, + 206, + 310, + 261 + ], + "type": "image", + "image_path": "622f350b20b08c08ca2a81decc4d8fa343a730e114883df4e5ed492fbbfc5e46.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 262, + 292, + 270 + ], + "lines": [ + { + "bbox": [ + 250, + 262, + 292, + 270 + ], + "spans": [ + { + "bbox": [ + 250, + 262, + 292, + 270 + ], + "type": "text", + "content": "(a) HR patch" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 312, + 207, + 390, + 261 + ], + "blocks": [ + { + "bbox": [ + 312, + 207, + 390, + 261 + ], + "lines": [ + { + "bbox": [ + 312, + 207, + 390, + 261 + ], + "spans": [ + { + "bbox": [ + 312, + 207, + 390, + 261 + ], + "type": "image", + "image_path": "5d0e7cb23b4086db9d47a77daec5596f4c7f7727454712d6e44257ff17d49eb1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 262, + 369, + 270 + ], + "lines": [ + { + "bbox": [ + 331, + 262, + 369, + 270 + ], + "spans": [ + { + "bbox": [ + 331, + 262, + 369, + 270 + ], + "type": "text", + "content": "(b) Bicubic" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 392, + 206, + 470, + 261 + ], + "blocks": [ + { + "bbox": [ + 392, + 206, + 470, + 261 + ], + "lines": [ + { + "bbox": [ + 392, + 206, + 470, + 261 + ], + "spans": [ + { + "bbox": [ + 392, + 206, + 470, + 261 + ], + "type": "image", + "image_path": "71e5d281c226a473a50dee59c44baab8cf287c14f7834c54a3574e144061bd74.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 262, + 454, + 270 + ], + "lines": [ + { + "bbox": [ + 406, + 262, + 454, + 270 + ], + "spans": [ + { + "bbox": [ + 406, + 262, + 454, + 270 + ], + "type": "text", + "content": "(c) VDSR [24]" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 471, + 206, + 549, + 261 + ], + "blocks": [ + { + "bbox": [ + 471, + 206, + 549, + 261 + ], + "lines": [ + { + "bbox": [ + 471, + 206, + 549, + 261 + ], + "spans": [ + { + "bbox": [ + 471, + 206, + 549, + 261 + ], + "type": "image", + "image_path": "c1176d288c6656b23b12f713ee7c43bb1e3da66961ba56ca3367783ba9e10710.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 232, + 271, + 310, + 324 + ], + "blocks": [ + { + "bbox": [ + 232, + 271, + 310, + 324 + ], + "lines": [ + { + "bbox": [ + 232, + 271, + 310, + 324 + ], + "spans": [ + { + "bbox": [ + 232, + 271, + 310, + 324 + ], + "type": "image", + "image_path": "8820d156b0c82fae48dce421a21ce6c3b903332c6ed019aba3c797762d567a2b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 325, + 299, + 335 + ], + "lines": [ + { + "bbox": [ + 244, + 325, + 299, + 335 + ], + "spans": [ + { + "bbox": [ + 244, + 325, + 299, + 335 + ], + "type": "text", + "content": "(e) LapSRN [27]" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 312, + 271, + 390, + 324 + ], + "blocks": [ + { + "bbox": [ + 312, + 271, + 390, + 324 + ], + "lines": [ + { + "bbox": [ + 312, + 271, + 390, + 324 + ], + "spans": [ + { + "bbox": [ + 312, + 271, + 390, + 324 + ], + "type": "image", + "image_path": "274b0fa28ba294b772474b88dd130cf5ac21656b8ccc4c84d4d7b0697d5cb9a0.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 325, + 373, + 335 + ], + "lines": [ + { + "bbox": [ + 329, + 325, + 373, + 335 + ], + "spans": [ + { + "bbox": [ + 329, + 325, + 373, + 335 + ], + "type": "text", + "content": "(f) CARN [1]" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 392, + 271, + 470, + 324 + ], + "blocks": [ + { + "bbox": [ + 392, + 271, + 470, + 324 + ], + "lines": [ + { + "bbox": [ + 392, + 271, + 470, + 324 + ], + "spans": [ + { + "bbox": [ + 392, + 271, + 470, + 324 + ], + "type": "image", + "image_path": "42eaf61c08a0c8515e656227235af5d0c0aae50f5290e2b5d01e5327c19f0d4d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 325, + 455, + 335 + ], + "lines": [ + { + "bbox": [ + 406, + 325, + 455, + 335 + ], + "spans": [ + { + "bbox": [ + 406, + 325, + 455, + 335 + ], + "type": "text", + "content": "(g) IMDN [20]" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 471, + 271, + 549, + 323 + ], + "blocks": [ + { + "bbox": [ + 475, + 262, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 475, + 262, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 475, + 262, + 545, + 271 + ], + "type": "text", + "content": "(d) ShuffleMixer [45]" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 271, + 549, + 323 + ], + "lines": [ + { + "bbox": [ + 471, + 271, + 549, + 323 + ], + "spans": [ + { + "bbox": [ + 471, + 271, + 549, + 323 + ], + "type": "image", + "image_path": "3e6bf4e654fd4d339ce58f0610f926a38a2a5efc1c0f7320057a366f9bf720bd.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 325, + 529, + 334 + ], + "lines": [ + { + "bbox": [ + 490, + 325, + 529, + 334 + ], + "spans": [ + { + "bbox": [ + 490, + 325, + 529, + 334 + ], + "type": "text", + "content": "(h) SAFMN" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 53, + 343, + 230, + 459 + ], + "blocks": [ + { + "bbox": [ + 53, + 343, + 230, + 459 + ], + "lines": [ + { + "bbox": [ + 53, + 343, + 230, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 343, + 230, + 459 + ], + "type": "image", + "image_path": "b1adef27a5c285f442aee31378aa9e76a80203eb26cb485480b46fdcaf3e8174.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 461, + 181, + 470 + ], + "lines": [ + { + "bbox": [ + 103, + 461, + 181, + 470 + ], + "spans": [ + { + "bbox": [ + 103, + 461, + 181, + 470 + ], + "type": "text", + "content": "img098 from Urban100" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 233, + 343, + 310, + 397 + ], + "blocks": [ + { + "bbox": [ + 233, + 343, + 310, + 397 + ], + "lines": [ + { + "bbox": [ + 233, + 343, + 310, + 397 + ], + "spans": [ + { + "bbox": [ + 233, + 343, + 310, + 397 + ], + "type": "image", + "image_path": "6cd63aeac1fe8ca0610913d9ff06eef59ca503f0fd4a8dc38ff024de7ef7ac4c.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 312, + 343, + 390, + 397 + ], + "blocks": [ + { + "bbox": [ + 312, + 343, + 390, + 397 + ], + "lines": [ + { + "bbox": [ + 312, + 343, + 390, + 397 + ], + "spans": [ + { + "bbox": [ + 312, + 343, + 390, + 397 + ], + "type": "image", + "image_path": "29d9f88e24fcfa660b6e81d81fe1a59f1c1e0a15cf7b33f1206d3b9c148b024b.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 392, + 343, + 470, + 397 + ], + "blocks": [ + { + "bbox": [ + 392, + 343, + 470, + 397 + ], + "lines": [ + { + "bbox": [ + 392, + 343, + 470, + 397 + ], + "spans": [ + { + "bbox": [ + 392, + 343, + 470, + 397 + ], + "type": "image", + "image_path": "102ec09a11e96e2616ea61933e51d6e081ef910c421dfc88171f190ab35c7e78.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 471, + 343, + 549, + 397 + ], + "blocks": [ + { + "bbox": [ + 471, + 343, + 549, + 397 + ], + "lines": [ + { + "bbox": [ + 471, + 343, + 549, + 397 + ], + "spans": [ + { + "bbox": [ + 471, + 343, + 549, + 397 + ], + "type": "image", + "image_path": "695a38c4790475a345b79fe11b9817bf3733219f2d4758432aec88d49de4d9d8.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 233, + 407, + 310, + 460 + ], + "blocks": [ + { + "bbox": [ + 250, + 398, + 292, + 407 + ], + "lines": [ + { + "bbox": [ + 250, + 398, + 292, + 407 + ], + "spans": [ + { + "bbox": [ + 250, + 398, + 292, + 407 + ], + "type": "text", + "content": "(a) HR patch" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 407, + 310, + 460 + ], + "lines": [ + { + "bbox": [ + 233, + 407, + 310, + 460 + ], + "spans": [ + { + "bbox": [ + 233, + 407, + 310, + 460 + ], + "type": "image", + "image_path": "b2b6dca1ce0e7511b68bd3f0dfdbd518a73e55e4cd26423a7344461ed0411231.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 461, + 298, + 471 + ], + "lines": [ + { + "bbox": [ + 244, + 461, + 298, + 471 + ], + "spans": [ + { + "bbox": [ + 244, + 461, + 298, + 471 + ], + "type": "text", + "content": "(e) LapSRN [27]" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 480, + 534, + 491 + ], + "lines": [ + { + "bbox": [ + 47, + 480, + 534, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 534, + 491 + ], + "type": "text", + "content": "Figure 6. Visual comparisons for " + }, + { + "bbox": [ + 47, + 480, + 534, + 491 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 47, + 480, + 534, + 491 + ], + "type": "text", + "content": " SR on the Urban100 dataset. The proposed method recovers the image with clearer structures." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 312, + 407, + 390, + 460 + ], + "blocks": [ + { + "bbox": [ + 332, + 398, + 369, + 407 + ], + "lines": [ + { + "bbox": [ + 332, + 398, + 369, + 407 + ], + "spans": [ + { + "bbox": [ + 332, + 398, + 369, + 407 + ], + "type": "text", + "content": "(b) Bicubic" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 407, + 390, + 460 + ], + "lines": [ + { + "bbox": [ + 312, + 407, + 390, + 460 + ], + "spans": [ + { + "bbox": [ + 312, + 407, + 390, + 460 + ], + "type": "image", + "image_path": "220e43d6dd4197ddb818451e4b9b2d69eb8d39156832abf7b85deee4f024ca71.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 461, + 373, + 471 + ], + "lines": [ + { + "bbox": [ + 328, + 461, + 373, + 471 + ], + "spans": [ + { + "bbox": [ + 328, + 461, + 373, + 471 + ], + "type": "text", + "content": "(f) CARN [1]" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 392, + 407, + 470, + 460 + ], + "blocks": [ + { + "bbox": [ + 406, + 398, + 454, + 407 + ], + "lines": [ + { + "bbox": [ + 406, + 398, + 454, + 407 + ], + "spans": [ + { + "bbox": [ + 406, + 398, + 454, + 407 + ], + "type": "text", + "content": "(c) VDSR [24]" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 407, + 470, + 460 + ], + "lines": [ + { + "bbox": [ + 392, + 407, + 470, + 460 + ], + "spans": [ + { + "bbox": [ + 392, + 407, + 470, + 460 + ], + "type": "image", + "image_path": "ef2a68857cb59fd7ab26ce8c84aa8002da778f481a48fc8113fe9cc22b435fdb.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 461, + 455, + 471 + ], + "lines": [ + { + "bbox": [ + 406, + 461, + 455, + 471 + ], + "spans": [ + { + "bbox": [ + 406, + 461, + 455, + 471 + ], + "type": "text", + "content": "(g) IMDN [20]" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 471, + 407, + 549, + 460 + ], + "blocks": [ + { + "bbox": [ + 475, + 398, + 544, + 407 + ], + "lines": [ + { + "bbox": [ + 475, + 398, + 544, + 407 + ], + "spans": [ + { + "bbox": [ + 475, + 398, + 544, + 407 + ], + "type": "text", + "content": "(d) ShuffleMixer [45]" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 407, + 549, + 460 + ], + "lines": [ + { + "bbox": [ + 471, + 407, + 549, + 460 + ], + "spans": [ + { + "bbox": [ + 471, + 407, + 549, + 460 + ], + "type": "image", + "image_path": "a6ea036299f59c8afbec5ef5d4e7807642ce93d1b1048ade568b17cd1dfc3c03.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 461, + 530, + 469 + ], + "lines": [ + { + "bbox": [ + 490, + 461, + 530, + 469 + ], + "spans": [ + { + "bbox": [ + 490, + 461, + 530, + 469 + ], + "type": "text", + "content": "(h) SAFMN" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "bbox": [ + 46, + 502, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 551 + ], + "type": "text", + "content": "the reconstruction accuracy, FLOPs and parameters on the " + }, + { + "bbox": [ + 46, + 502, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 46, + 502, + 287, + 551 + ], + "type": "text", + "content": " Set5 dataset in Figure 1. The proposed SAFMN model achieves a favorable trade-off between model complexity and reconstruction performance." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "type": "text", + "content": "Qualitative comparisons. In addition to the quantitative evaluations, we provide qualitative comparisons of the proposed method. Figure 6 shows the visual comparisons on the Urban100 dataset for " + }, + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 46, + 554, + 287, + 650 + ], + "type": "text", + "content": " SR. Our approach generates parallel straight lines and grid patterns more accurately than the listed methods. These results also demonstrate the effectiveness of our method for adaptive feature modulation by exploiting non-local feature interactions." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "content": "Memory and running time comparisons. To fully examine the efficiency of our proposed method, we further evaluate our method against five representative ones, including CARN-M [1], CARN [1], EDSR-baseline [32], IMDN [20], and LAPAR-A [28] on " + }, + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "content": " SR in terms of" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 502, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 705 + ], + "type": "text", + "content": "the GPU memory consumption (#GPU Mem.) and running time (#Avg. Time). The maximum GPU memory consumption is recorded during the inference. The running time is averaged on 50 test images with a resolution of " + }, + { + "bbox": [ + 304, + 502, + 547, + 705 + ], + "type": "inline_equation", + "content": "320 \\times 180" + }, + { + "bbox": [ + 304, + 502, + 547, + 705 + ], + "type": "text", + "content": " pixels. We show the memory and running time comparisons in Table 2, our method achieves a clear improvement over other state-of-the-art methods. By using the spatially-adaptive feature modulation layer and the compact channel mixer, the GPU consumption of our SAFMN is only " + }, + { + "bbox": [ + 304, + 502, + 547, + 705 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 502, + 547, + 705 + ], + "type": "text", + "content": " of the CARN series and " + }, + { + "bbox": [ + 304, + 502, + 547, + 705 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 502, + 547, + 705 + ], + "type": "text", + "content": " of LAPAR-A; the running time is nearly twice as fast as other evaluated methods, except for IMDN. Compared to IMDN [20], our method has a similar running time speed while significantly reducing memory usage. Tables 1 and 2 show that the proposed model achieves a favorable trade-off in terms of inference speed, model complexity and reconstruction performance against state-of-the-art methods." + } + ] + } + ], + "index": 58 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13195" + } + ] + } + ], + "index": 59 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 129, + 282, + 208 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "type": "text", + "content": "Table 2. Memory and running time comparisons on " + }, + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "type": "text", + "content": " SR. #GPU Mem. denotes the maximum GPU memory consumption during the inference phase, derived with the Pytorch torch.cuda.max_memory_allocated() function. #Avg. Time is the average running time on 50 LR images of " + }, + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "type": "inline_equation", + "content": "320 \\times 180" + }, + { + "bbox": [ + 47, + 69, + 287, + 125 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 129, + 282, + 208 + ], + "lines": [ + { + "bbox": [ + 51, + 129, + 282, + 208 + ], + "spans": [ + { + "bbox": [ + 51, + 129, + 282, + 208 + ], + "type": "table", + "html": "
Methods#GPU Mem. [M]#Avg.Time [ms]
CARN-M [1]680.8417.85
CARN [1]689.8318.90
EDSR-baseline [32]486.5819.81
IMDN [20]203.4410.22
LAPAR-A [28]1811.4724.91
SAFMN65.2610.71
", + "image_path": "4841e3cdda00c78a2ceca854042bccb24c4b56be35aa3c1476f9193439e4600e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 217, + 185, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 185, + 230 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 185, + 230 + ], + "type": "text", + "content": "5. Analysis and Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 286, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 286, + 319 + ], + "type": "text", + "content": "We further conduct extensive ablation studies to better understand and evaluate each component in the proposed SAFMN. For fair comparisons with the designed baselines, we implement all experiments based on " + }, + { + "bbox": [ + 46, + 236, + 286, + 319 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 236, + 286, + 319 + ], + "type": "text", + "content": " SAFMN and train them with the same setting. The experimental results in Table 3 are measured on DIV2K-val [48] and Manga109 [38] datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 320, + 286, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 286, + 415 + ], + "type": "text", + "content": "Effectiveness of the SAFM. To demonstrate the effect of the spatially-adaptive feature modulation, we first remove this module for comparison. Without it, the PSNR values will drop by 0.17dB (30.26 vs. 30.43) and 0.34dB (30.09 vs. 30.43) on the DIV2K-val and Manga109 datasets, respectively. These results show the importance of the SAFM. Therefore, we further investigate this module to figure out why it works." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 418, + 286, + 685 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 59, + 418, + 286, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 418, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 59, + 418, + 286, + 464 + ], + "type": "text", + "content": "- Feature modulation. The feature modulation mechanism is introduced to enable the network with adaptive properties. Without this operation, the baseline model is lowered by 0.11dB on the Manga109 dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 468, + 286, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 468, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 59, + 468, + 286, + 610 + ], + "type": "text", + "content": "- Multi-scale representation. Here, \"w/o MR\" in Table 3 denotes that we directly use a depth-wise convolution with a kernel size of " + }, + { + "bbox": [ + 59, + 468, + 286, + 610 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 59, + 468, + 286, + 610 + ], + "type": "text", + "content": " pixels to extract spatial information and do not employ multi-scale features. A noticeable performance drop of 0.13dB on the Manga109 dataset is observed without these multiscale features. Moreover, we apply the adaptive max pooling over the input features in this module to build feature pyramids. Compared to using adaptive average pooling or nearest interpolation, adaptive max pooling allows the model to detect discriminative features, resulting in better reconstruction results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 614, + 286, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 614, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 59, + 614, + 286, + 685 + ], + "type": "text", + "content": "- Feature aggregation. We use a " + }, + { + "bbox": [ + 59, + 614, + 286, + 685 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 59, + 614, + 286, + 685 + ], + "type": "text", + "content": " convolution to aggregate the multi-scale features on the channel dimension. Combined with the modulation mechanism, it brings a PSNR improvement of 0.14dB on the Manga109 dataset, which proves the necessity of aggregating the multi-scale features." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 66, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 689, + 286, + 713 + ], + "type": "text", + "content": "Without all three components mentioned before, it represents that only a " + }, + { + "bbox": [ + 66, + 689, + 286, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 66, + 689, + 286, + 713 + ], + "type": "text", + "content": " depth-wise convolution is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 144 + ], + "type": "text", + "content": "used to encode the spatial information and leads to a PSNR reduction of 0.12dB and 0.2dB on the DIV2Kval and Manga109 datasets, respectively. This performance drop suggests that the spatially-adaptive modulation based on the multi-scale feature representation effectively boosts SR reconstruction performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 146, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 146, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 146, + 545, + 217 + ], + "type": "text", + "content": "- GELU function. Here, we use the GELU [17] function to normalize the modulation map. Table 3 shows that better experimental results can be achieved with GELU than with Sigmoid or without GELU, mainly because it weights the input features by their percentile and allows better-activated features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 223, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 294 + ], + "type": "text", + "content": "In addition, we use the local attribution map (LAM) [13] and diffusion index (DI) [13] to illustrate the feature interaction range. Figure 2 clearly shows that the proposed model can utilise more feature information than previous CNN-based efficient SR methods. The more information used, the better the reconstruction performance in general." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "text", + "content": "The above analysis shows that benefiting from the multiscale feature representation, our proposed SAFM can effectively exploit non-local feature interactions with small memory and computational costs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 342, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 628 + ], + "type": "text", + "content": "Effectiveness of the CCM. Compared with the original FMBCConv [47], the main change made by CCM is removing the SE [18] block. As shown in Table 3, using SE blocks, its performance is almost the same as without them. This result is mainly caused by the SAFM block allows adaptability in channel and spatial dimensions. Hence, we do not use the SE blocks for saving parameters. We next conduct a series of ablations to verify that CCM can effectively encode local contextual information and perform channel mixing. Without it, the model only achieves the accuracy of 29.69dB and 28.49dB on DIV2K-val and Manga109 datasets, proving the indispensability and locality modelling ability of this part. We then change CCM to other channel mixers commonly used in ViT architectures: channel MLP [12] or inverted residual block [43]. When the channel MLP is adopted for the channel mixer, there is a significant performance drop of 0.63dB (29.80 vs. 30.43) on the Manga109 dataset. This performance reduction results from a lack of capacity for modeling local contextual information. For the inverted residual block, the performance is highly similar to that of CCM, but the corresponding #Acts is increased by nearly 33M, which means more memory consumption and slower inference time. Thus, we use the CCM as the default channel mixer." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Effect of the LayerNorm layer. Since we use the elementwise product in the SAFM module, it will result in abnormal gradient values and unstable model training, as shown in Figure 7(a). It is, therefore, necessary to normalize the input features. Here, we perform normalization with the LayerNorm [3] layer. To verify this assumption, we first remove the LayerNorm layer. Table 3 and Figure 7(a) show" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13196" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "type": "text", + "content": "Table 3. Ablation for SAFMN on DIV2K-val and Manga109 datasets. SAFMN with a scaling factor of " + }, + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "type": "text", + "content": " is utilized as the baseline for ablation studies. The PSNR/SSIM values on benchmarks are reported. “" + }, + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "type": "inline_equation", + "content": "\\mathrm{A} \\rightarrow \\mathrm{B}" + }, + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "type": "text", + "content": "” is to replace A with B. “None” means to remove the operation. “Ir” denotes the learning rate. “FBN” is the abbreviation of the Frozen BatchNorm. “" + }, + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2" + }, + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "type": "text", + "content": " normalization” represents that the inputs are normalized by " + }, + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2" + }, + { + "bbox": [ + 46, + 68, + 547, + 137 + ], + "type": "text", + "content": "-norm over the channel dimension. “FM”, “MR”, and “FA” are the abbreviations corresponding to feature modulation, multi-scale representation, and feature aggregation. * indicates that the corresponding results are obtained before the training collapse. The number of parameters, FLOPs and Acts are calculated in the same way as in Table 1." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 138, + 541, + 376 + ], + "blocks": [ + { + "bbox": [ + 53, + 138, + 541, + 376 + ], + "lines": [ + { + "bbox": [ + 53, + 138, + 541, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 138, + 541, + 376 + ], + "type": "table", + "html": "
AblationVariant#Params [K]#FLOPs [G]#Acts [M]DIV2K_valManga109
Baseline-239.5213.5676.7030.43/0.837230.43/0.9063
Main moduleSAFM → None225.4112.9054.6130.26/0.833030.09/0.9018
CCM → None30.721.6126.9329.69/0.819328.49/0.8756
SAFM(a): w/o FM239.5213.5676.7030.36/0.835730.32/0.9048
(b): w/o MR239.5213.6487.7830.34/0.835030.30/0.9047
(c): w/o FA228.8612.9660.1130.36/0.835530.29/0.9049
(a) + (b)239.5213.6487.7830.32/0.834530.24/0.9038
(a) + (c)228.8612.9660.1130.34/0.835130.28/0.9043
(a) + (b) + (c)228.8613.0571.1930.31/0.834430.23/0.9036
AdaptiveMaxPool → AdaptiveAvgPool239.5213.5676.7030.40/0.836430.41/0.9061
AdaptiveMaxPool → Nearest interpolate239.5213.5676.7030.36/0.835430.31/0.9048
GELU → None239.5213.5676.7030.40/0.836630.37/0.9058
GELU → Sigmoid239.5213.5676.7030.36/0.835530.29/0.9044
CCMw/ SE260.9813.5976.7030.39/0.836030.46/0.9067
CCM → Channel MLP73.634.0076.7030.17/0.831329.80/0.8980
CCM → Inverted residual block245.2813.85110.0030.43/0.837330.43/0.9064
NormalizationLN → None, lr=1 × 10-3*238.3713.5576.7030.29/0.834030.04/0.9014
LN → None, lr=1 × 10-4238.3713.5576.7030.15/0.830629.74/0.8970
LN → BN239.5213.7276.7030.28/0.835430.05/0.9029
LN → FBN*238.3713.5576.7030.30/0.834330.15/0.9028
LN → L2 normalization238.3713.5576.7030.37/0.835830.31/0.9049
", + "image_path": "b237df714f76606e6cca04b26ef5f99c204623ba60354e3aaf0dc35f15d3b3ec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 390, + 164, + 475 + ], + "blocks": [ + { + "bbox": [ + 48, + 390, + 164, + 475 + ], + "lines": [ + { + "bbox": [ + 48, + 390, + 164, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 390, + 164, + 475 + ], + "type": "image", + "image_path": "7c3078570a3532d142635edd2c3b35e19e0f056ff7d1e04e9402978aa7aa07a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 479, + 147, + 489 + ], + "lines": [ + { + "bbox": [ + 66, + 479, + 147, + 489 + ], + "spans": [ + { + "bbox": [ + 66, + 479, + 147, + 489 + ], + "type": "text", + "content": "(a) Effect of LayerNorm" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 479, + 295, + 490 + ], + "lines": [ + { + "bbox": [ + 156, + 479, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 156, + 479, + 295, + 490 + ], + "type": "text", + "content": "(b) Comparison of different normalizations" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 390, + 280, + 474 + ], + "blocks": [ + { + "bbox": [ + 167, + 390, + 280, + 474 + ], + "lines": [ + { + "bbox": [ + 167, + 390, + 280, + 474 + ], + "spans": [ + { + "bbox": [ + 167, + 390, + 280, + 474 + ], + "type": "image", + "image_path": "44de5793af16a9eb02ab82b433a0a47f0f1e7db89e3a7e16b681ced063f6ee55.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 495, + 288, + 540 + ], + "lines": [ + { + "bbox": [ + 46, + 495, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 288, + 540 + ], + "type": "text", + "content": "Figure 7. Smoothed training loss curves with or without normalization. Figure (a) shows LayerNorm stabilizes model training and converges better than without it. Figure (b) compares the effects of different normalization approaches on model training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "that without it, the model suffers a training crash at a large learning rate and does not converge well at small ones with its PSNR of only 30.15dB and 29.74dB on the DIV2K-val and Manga109 datasets, respectively. Next, we compare the LayerNorm with three representative normalization methods, including BatchNorm [22], Frozen BatchNorm (FBN) [22], and " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " normalization in Figure 7(b). The results in Table 3 and Figure 7(b) demonstrate that the BN family decrease the PSNR/SSIM values by a large margin, and the FBN even does not guarantee a stable model training process. Although the " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_2" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": " normalization allows the model to be trained successfully, its performance is not as good as that of using LayerNorm. Thus, the LayerNorm layer is set as default for the proposed SAFMN." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 395, + 542, + 438 + ], + "blocks": [ + { + "bbox": [ + 306, + 385, + 522, + 395 + ], + "lines": [ + { + "bbox": [ + 306, + 385, + 522, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 385, + 522, + 395 + ], + "type": "text", + "content": "Table 4. Effect of the number of FMMs on performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 395, + 542, + 438 + ], + "lines": [ + { + "bbox": [ + 309, + 395, + 542, + 438 + ], + "spans": [ + { + "bbox": [ + 309, + 395, + 542, + 438 + ], + "type": "table", + "html": "
#N#Params [K]#FLOPs [G]#Acts [M]DIV2K_valManga109
4128.067.2540.7730.21/0.831929.86/0.8988
6183.7910.4058.6430.32/0.834730.18/0.9033
8239.5213.5676.7030.43/0.837230.43/0.9063
12350.9719.86113.0030.51/0.838930.71/0.9096
", + "image_path": "7a66f8b5173635210b952c85f5925aa2cd715040f0b5cac9381bbef4cd12e31e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 454, + 542, + 468 + ], + "blocks": [ + { + "bbox": [ + 306, + 442, + 541, + 453 + ], + "lines": [ + { + "bbox": [ + 306, + 442, + 541, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 541, + 453 + ], + "type": "text", + "content": "Table 5. GPU runtime of meta-operations in the SAFM layer." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 454, + 542, + 468 + ], + "lines": [ + { + "bbox": [ + 309, + 454, + 542, + 468 + ], + "spans": [ + { + "bbox": [ + 309, + 454, + 542, + 468 + ], + "type": "table", + "html": "
Meta-operationsLayerNormChannel SplittingAdaptive MaxPoolNearest InterpolationElement-wise ProductElement-wise Addition
Runtime [us]4651338524156193
", + "image_path": "7fd168afea0491e1d2db3be921c13548ec89d4ff62646b063865ee38e9fc1285.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "content": "Effect of the number of FMMs. The proposed network contains several FMMs. We evaluate the effect of the number of FMMs by setting the number #N from 4 to 12. Table 4 shows that the performance gradually improves as #N increases. To achieve a favorable trade-off between model complexity and reconstruction performance, we choose 8 blocks as the default setting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Relations to lightweight ViT models. To alleviate the quadratic complexity of vanilla ViT [12], MobileViT [39], Mobile-Former [6], and EdgeNeXt [37] adopt a hybrid CNN-Transformer design, which replaces the shallow transformer blocks with convolutional layers to avoid using self-attention on high-resolution representations, and efficiently combines the transformer blocks with the modified mobilenet blocks to perform feature interactions in the later stages. Another alternative [50, 14] is to incorporate concepts from the transformer into convolutional modules. For example, MetaFormer [50] and VAN [14] use pooling operation or large kernel attention to replace the heavy-weight self-attention for contextual information aggregation. Com" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13197" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "pared to these approaches, we develop an efficient feature modulation mechanism upon a grouped multi-scale representation to learn feature dependencies. Benefiting from the multi-head paradigm, the proposed model can exploit multiscale contexts across different heads for adaptive feature modulation while maintaining a low computational cost." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 289, + 312 + ], + "type": "text", + "content": "Limitations. Although our SAFMN has better overall performance than the evaluated methods [1, 20, 49, 28, 45], its running time is relatively slow. As shown in Table 5, we count the inference time of different commonly used tensor operations with the torch.profiler function, and the results show that the utilized LayerNorm layers and residual connections lead to increased runtime. The running time of SAFMN can be further accelerated by optimizing the implementation of these components. We will investigate these in the future work. Another limitation is that the PSNR performance of SAFMN is suboptimal, especially on the Urban100 dataset, which is mainly due to the fact that we perform the feature modulation without considering the integration of local features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 326, + 119, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 119, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 119, + 338 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 289, + 491 + ], + "type": "text", + "content": "In this paper, we propose a simple yet efficient CNN model to solve the efficient image super-resolution problem. Our SAFMN exploits non-local feature interactions upon a multi-scale feature representation. To complement the local contextual information, we further develop a compact convolutional channel mixer to encode spatially local context and conduct channel mixing simultaneously. We both qualitatively and quantitatively evaluate the proposed method on commonly used benchmarks. Experimental results show that the proposed SAFMN model is more efficient than state-of-the-art methods while achieving competitive performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 505, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 577 + ], + "type": "text", + "content": "Acknowledgements. This work has been partly supported by the National Key R&D Program of China (No. 2018AAA0102001), the National Natural Science Foundation of China (Nos. U22B2049, 62272233, 61922043, 61925204), and the Fundamental Research Funds for the Central Universities (Nos. 30922010910, 30920041109)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 590, + 106, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 106, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 106, + 602 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 610, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 52, + 610, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 610, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 52, + 610, + 287, + 643 + ], + "type": "text", + "content": "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, 2018. 1, 2, 4, 5, 6, 7, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 646, + 287, + 678 + ], + "type": "text", + "content": "[2] Pablo Arbeláez, Michael Maire, Charless C. Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. PAMI, 33(5):898-916, 2011. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 287, + 712 + ], + "type": "text", + "content": "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4, 7" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 711 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 311, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 311, + 73, + 545, + 116 + ], + "type": "text", + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie line Alberi Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 1, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 118, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 547, + 162 + ], + "type": "text", + "content": "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 164, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 164, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 312, + 164, + 545, + 206 + ], + "type": "text", + "content": "[6] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In CVPR, 2022.8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 209, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 209, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 545, + 243 + ], + "type": "text", + "content": "[7] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2021. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 244, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 244, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 312, + 244, + 545, + 277 + ], + "type": "text", + "content": "[8] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 278, + 545, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 278, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 312, + 278, + 545, + 321 + ], + "type": "text", + "content": "[9] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 324, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 324, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 324, + 545, + 357 + ], + "type": "text", + "content": "[10] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. PAMI, 38(2):295-307, 2016. 2, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 358, + 545, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 358, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 308, + 358, + 545, + 391 + ], + "type": "text", + "content": "[11] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In ECCV, 2016. 1, 2, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 393, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 458 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 460, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 482 + ], + "type": "text", + "content": "[13] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In CVPR, 2021. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 484, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 484, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 484, + 545, + 517 + ], + "type": "text", + "content": "[14] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 518, + 545, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 518, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 308, + 518, + 545, + 550 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 553, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 553, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 553, + 545, + 586 + ], + "type": "text", + "content": "[16] Zibin He, Tao Dai, Jian Lu, Yong Jiang, and Shu-Tao Xia. Faked: Feature-affinity based knowledge distillation for efficient image super-resolution. In ICIP, 2020. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 588, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 588, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 588, + 545, + 609 + ], + "type": "text", + "content": "[17] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units. arXiv preprint arXiv:1606.08415, 2016. 4, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 611, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 545, + 632 + ], + "type": "text", + "content": "[18] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 666 + ], + "type": "text", + "content": "[19] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 545, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 711 + ], + "type": "text", + "content": "[20] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multistillation network. In ACM MM, 2019. 1, 2, 4, 5, 6, 7, 9" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13198" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[21] Andrey Ignatov, Radu Timofte, Maurizio Denna, and et al. Efficient and accurate quantized image super-resolution on Mobile NPUs, Mobile AI & AIM 2022 challenge: Report. In ECCV Workshops, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[23] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "type": "text", + "content": "[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2, 4, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 240 + ], + "type": "text", + "content": "[25] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "text", + "content": "[26] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPR Workshops, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "text", + "content": "[27] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, 2017. 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 321, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 364 + ], + "type": "text", + "content": "[28] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. LAPAR: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. In NeurIPS, 2020. 4, 5, 6, 7, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "type": "text", + "content": "[29] Yawei Li, Kai Zhang, Luc Van Gool, Radu Timofte, et al. NTIRE 2022 challenge on efficient super-resolution: Methods and results. In CVPR Workshops, 2022. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "text", + "content": "[30] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In CVPR Workshops, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 477 + ], + "type": "text", + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In ICCV Workshops, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "text", + "content": "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPR Workshops, 2017. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "text", + "content": "[33] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV Workshops, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 634 + ], + "type": "text", + "content": "[35] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In CVPR, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 657 + ], + "type": "text", + "content": "[36] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[37] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: Efficiently amalgamated cnn-transformer architecture for mobile vision applications. In ECCV Workshops, 2022. 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 692 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[38] Yusuke Matsui, Kota Ito, Yuji Aramaki, Toshihiko Yasasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. arXiv preprint arXiv:1510.04389, 2015. 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[39] Sachin Mehta and Mohammad Rastegari. MobileViT: Lightweight, general-purpose, and mobile-friendly vision transformer. In ICLR, 2022. 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "text", + "content": "[40] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "text", + "content": "[41] Pablo Navarrete Michelini, Yunhua Lu, and Xingqun Jiang. edge-SR: Super-resolution for the masses. In WACV, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[42] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. In CVPR, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "text", + "content": "[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "type": "text", + "content": "[44] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "text", + "content": "[45] Long Sun, Jinshan Pan, and Jinhui Tang. ShuffleMixer: An efficient convnet for image super-resolution. In NeurIPS, 2022. 1, 2, 4, 5, 6, 9" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "text", + "content": "[46] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. CVPR, 2017. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 411, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 432 + ], + "type": "text", + "content": "[47] Mingxing Tan and Quoc Le. EfficientNetV2: Smaller models and faster training. In ICML, 2021. 2, 4, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "type": "text", + "content": "[48] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, Lei Zhang, et al. NTIRE 2017 challenge on single image super-resolution: Methods and results. In CVPR Workshops, 2017. 4, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "text", + "content": "[49] Longguang Wang, Xiaoyu Dong, Yingqian Wang, Xinyi Ying, Zaiping Lin, Wei An, and Yulan Guo. Exploring sparsity in image super-resolution for efficient inference. In CVPR, 2021. 5, 9" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[50] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 557, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 590 + ], + "type": "text", + "content": "[51] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces*, 2012. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 591, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 624 + ], + "type": "text", + "content": "[52] Xindong Zhang, Hui Zeng, and Lei Zhang. Edge-oriented convolution block for real-time super resolution on mobile devices. In ACM MM, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 625, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 657 + ], + "type": "text", + "content": "[53] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 658, + 545, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 692 + ], + "type": "text", + "content": "[54] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCV Workshops, 2020. 1, 2, 5" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13199" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/6732d370-6205-414c-986c-288f16a786f3_content_list.json b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/6732d370-6205-414c-986c-288f16a786f3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f64f211759a0c45d16b04e18b6e84635673a3973 --- /dev/null +++ b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/6732d370-6205-414c-986c-288f16a786f3_content_list.json @@ -0,0 +1,1566 @@ +[ + { + "type": "text", + "text": "Spatio-Temporal Crop Aggregation for Video Representation Learning", + "text_level": 1, + "bbox": [ + 125, + 130, + 843, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sepehr Sameni", + "bbox": [ + 212, + 181, + 334, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Bern", + "bbox": [ + 197, + 199, + 346, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sepehr.sameni@unibe.ch", + "bbox": [ + 174, + 220, + 370, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simon Jenni", + "bbox": [ + 437, + 181, + 539, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adobe Research", + "bbox": [ + 423, + 198, + 555, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jenni@adobe.com", + "bbox": [ + 423, + 219, + 557, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Paolo Favaro", + "bbox": [ + 647, + 181, + 754, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Bern", + "bbox": [ + 625, + 198, + 774, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "paolo.favaro@unibe.ch", + "bbox": [ + 606, + 219, + 795, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose Spatio-temporal Crop Aggregation for video representation Learning (SCALE), a novel method that enjoys high scalability at both training and inference time. Our model builds long-range video features by learning from sets of video clip-level features extracted with a pretrained backbone. To train the model, we propose a self-supervised objective consisting of masked clip feature predictions. We apply sparsity to both the input, by extracting a random set of video clips, and to the loss function, by only reconstructing the sparse inputs. Moreover, we use dimensionality reduction by working in the latent space of a pre-trained backbone applied to single video clips. These techniques make our method not only extremely efficient to train but also highly effective in transfer learning. We demonstrate that our video representation yields state-of-the-art performance with linear, nonlinear, and $k$ -NN probing on common action classification and video understanding datasets.", + "bbox": [ + 73, + 300, + 473, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 602, + 209, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Videos provide rich and detailed information about objects and their activities. Their analysis is, however, made challenging not only by the difference in the information provided across space and time but also by the high dimensionality of the data [45, 29, 70]. While computational resources are expected to scale over time, so is the demand for higher data resolution (both in space and time) and also the need for processing data with even more dimensions, such as videos of volumetric data [10]. Therefore, it is of paramount importance to explore methods that drastically reduce the computational requirements to process videos. Moreover, the annotation of videos is an extremely costly and time-consuming burden that makes the use of models pre-trained in a self-supervised manner essential [44].", + "bbox": [ + 73, + 628, + 468, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) is a very popular technique to reduce the need for annotation because it can build useful representations from unlabeled data through an artificial goal, also called pseudo- or pretext-task. These repre", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sentations can either be evaluated through K-nearest neighbor or linear probing [71, 72] or through fine-tuning (i.e., as the initialization parameters of the trained model) [25, 54, 35] on a downstream task, where only a small labeled dataset is available. More remarkably, SSL pre-trained models can outperform models that were pre-trained on an annotated dataset [53, 4].", + "bbox": [ + 496, + 268, + 893, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SSL methods for video representation learning present fundamental scalability challenges [54, 45, 29, 65]. A first major challenge is that training models from scratch for any new pseudo-task is not sustainable. A more viable setting is one where the processing of videos is broken down into multiple steps. First, one pre-processes video clips (spatiotemporal crops) only once via some pre-trained general-purpose model (e.g., trained via self-supervised learning [18, 23, 26, 44, 7, 37, 41]). In general, the idea is that this (compressed) local representation is stored and used later for other training purposes or retrieval. Then, one can train more efficiently other super-features that aggregate information from these local representations. Within this setting, we address two questions: 1) Can one further improve the performance of pre-computed video features by training a model on top of them? 2) Can such training be made computationally scalable? In this paper, we show that both questions have a positive answer. The answer exploits four fronts:", + "bbox": [ + 496, + 376, + 893, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Input Sparsity: Sparsity in the input to the model [54, 25, 3, 19, 1] drastically reduces computational load and memory requirements, leveraging information redundancy in images and videos [16]. Inspired by prior work, we extract a sparse set of clips from a video (these are spatio-temporal video crops and not image patches or video tubelets); each clip is fed separately to a neural network so that we obtain a set of video clip representations.", + "bbox": [ + 513, + 675, + 893, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Output Sparsity: Another way to reduce the computational cost is to use a sparse reconstruction output instead of a dense one [71, 51, 59, 5]. This is not just a reduction of the number of terms in the loss function, but also in the number of actual outputs of", + "bbox": [ + 511, + 824, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5664", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the model, and thus also a reduction of the computations needed to obtain them. This is in contrast to MAE-based SSL methods for vision, where the proposed pseudo-tasks are based on the reconstruction of the whole input [25, 54, 19, 16], and even if the loss uses a subset of the tokens (the masked ones) for the loss calculation, the remaining tokens are still part of the decoder's output and computational graph.", + "bbox": [ + 107, + 90, + 468, + 212 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3. Dimensionality Reduction: Inspired by prior work [71, 41, 13], instead of directly processing the raw input data, we work in the latent space. This allows us to further reduce the dimensionality of both input and output data.", + "4. Use of a Pre-trained Backbone: To reduce training time and further speed up the processing per iteration, we exploit SSL pre-trained models. These pre-trained backbones already learn very strong short-term spatiotemporal features, and our approach is a way to extract longer-term features (see LVU results 2) by building a video representation on top of a set of pre-trained features (one for each video clip)." + ], + "bbox": [ + 88, + 220, + 468, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To integrate all these components, we propose a novel SSL method that we call Spatio-temporal Crop Aggregation for video representation LEarning (SCALE). Given an input video, SCALE extracts a random set of clips and produces an embedding for each clip through a pre-trained backbone, which is kept frozen. These initial embeddings are then augmented in two ways: 1) each one is refined into a more discriminative feature, and 2) the set of all embeddings is summarized in a global feature. These global features can learn long-term correlations in the whole video by aggregating the short-term information in each clip embedding. The combination of the initial embeddings with their refinement and the global feature is then used in an ensemble for applications on new downstream tasks.", + "bbox": [ + 76, + 431, + 467, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To train SCALE, we introduce two novel pseudo-tasks, which aim to improve the discriminability of the embeddings of each video clip as well as obtain a global representation of the video. One task, which we call Masked Clip Modeling (MCM), is the reconstruction of a video clip embedding as in masked autoencoders [54]. Masked embeddings are combined with positional encodings so that the model can (spatio-temporally) relate the missing input embeddings to the other available embeddings (similarly to BERT in Natural Language Processing [12]). A second task is to train the model to output a global feature token (which we refer to as CLS, just for consistency with previous works [7, 12, 14, 55, 9]) for a set of clips via contrastive learning, so that the global feature is invariant to the chosen set of clips from the same video, but can discriminate summary features of clips from other videos. Both tasks are trained via contrastive losses.", + "bbox": [ + 76, + 643, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SCALE consistently improves upon pretrained features, with more evident gains considering the computational cost of other methods. For instance, SCALE outperforms VideoMAE [54] by $0.3\\%$ (1600 epochs checkpoint – fine-tuning) and $\\rho$ BYOL [18] by $0.6\\%$ (800 epochs checkpoint – linear probe) on Kinetics400 with 64 V100 GPUs in about 5 mins. In contrast, VideoMAE takes 27.7 hours to gain $0.5\\%$ (800 to 1600 epochs checkpoint), and $\\rho$ BYOL needs 48 hours to gain $0.4\\%$ (200 to 400 epochs checkpoint) with the same resources.", + "bbox": [ + 496, + 90, + 890, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize our contributions, we propose SCALE, a novel and highly scalable video representation method that", + "bbox": [ + 498, + 243, + 890, + 273 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- is trained via novel pseudo-tasks on sets of video clips (in contrast to existing methods that work only with pairs of clips at a time [44, 18, 45]);", + "- results in video feature representations with a significant performance improvement in $k$ -NN (retrieval), linear, and nonlinear probing across a wide range of datasets for action classification (UCF [48], HMDB [34], SSv2 [22], Kinetics400 [33]) and long-form video understanding (LVU [63]);", + "- achieves consistent transfer learning performance improvement across diverse SotA pre-trained backbones (architectures, scale, and pre-training tasks). For example, our nonlinear probed model even outperforms fully fine-tuned SVT [44] on HMDB [34]." + ], + "bbox": [ + 517, + 285, + 890, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 540, + 640, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach relates to many prior works on (self-supervised) representation learning. In particular, SCALE relates to SSL approaches on videos, methods that rely on multiple views of the data, and predictive methods, where part of the data is predicted from another part.", + "bbox": [ + 496, + 566, + 890, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Video Representation Learning. Early SSL approaches on video were based on pseudo tasks, e.g., the recognition of transformations of video frame sequences [38, 6, 62]. More recently, popular methods developed on images have been successfully translated to video, e.g., contrastive methods [18, 23, 26], clustering-based methods [44, 7], or predictive approaches [37, 41]. Often, these approaches are tailored to video data by including additional learning signals, e.g., by combining contrastive methods with temporal constraints [11] and pretext tasks [32], or by incorporating audio [39], or optical flow [24]. These approaches typically learn representations with limited temporal extent, which can serve as backbones for our approach.", + "bbox": [ + 496, + 643, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning from Multiple Views. Several methods have exploited multiple views (e.g., generated through space-time cropping) to represent and learn from videos. For example, many SSL approaches rely on multiple views of the", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5665", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "data, e.g., in contrastive formulations [18, 44, 42], or predictive learning [45], where invariance to views is the goal. Other approaches aim to learn from the relation between two views, e.g., by predicting overlap [69], the relative distance [49], or cross-view feature prediction [68, 52] and reconstruction [40]. Besides exploiting multiple views for SSL, some works also propose general multi-view video models, e.g., by capturing and fusing features at different spatio-temporal resolutions [67], by aggregating information over longer time-spans [64, 47, 65, 57], or by selecting important frames [21]. These approaches, however, are learned end-to-end. In contrast, we propose a more scalable approach by learning a global video representation of pretrained features extracted over multiple crops using self-supervision.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Predictive Learning. One of our proposed SSL objectives is a prediction of space-time crop features given other video crop features. This approach is similar to masked token prediction as in BERT [12] and relates to several other methods in the literature. Masked input reconstruction methods have recently become popular on images [25] and successfully translated to video [54, 16]. Other approaches formulate masked prediction tasks in the learned feature space [71, 52, 13] or some fixed latent space [51]. Another line of work considers directional predictions (e.g., into the future) often formulated via contrastive predictive coding [41] applied to video [37, 36, 50, 66, 20]. These masked prediction tasks are typically formulated on a fixed grid (e.g., at the level of patches or frames). In contrast, our formulation is continuous, predicting features of arbitrarily sampled space-time crops.", + "bbox": [ + 75, + 321, + 472, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Scalable Video Representation Learning", + "text_level": 1, + "bbox": [ + 76, + 590, + 439, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To describe SCALE, we first define some basic notation and functions, including a general contrastive loss notation that we use for all training losses.", + "bbox": [ + 75, + 622, + 468, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notation", + "text_level": 1, + "bbox": [ + 76, + 691, + 181, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use lower-case letters (e.g., $z$ ) to denote generic vectors and capital letters (e.g., $Z$ ) to denote their sets. The expression $a \\circ b$ denotes the concatenation of $a$ and $b$ . We also avoid indicating the parameters of networks (usually denoted by $\\theta$ ) if their presence and role are clear from the context. Throughout the description of the method we refer to the training of neural networks and therefore at each iteration of the training we sample a minibatch of videos. All the equations in the next sections are written for a single video in the minibatch. Although we do not explicitly indicate it, all the contrastive losses also use the other videos in the minibatch as negatives.", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Contrastive Loss", + "text_level": 1, + "bbox": [ + 500, + 90, + 666, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "InfoNCE is a powerful method for representation learning [41] that can be used to maximize the mutual information between two variables. Because we use this loss between different variables throughout our method, we introduce here a unified notation. Let the paired sets $A$ and $B$ have $N$ elements each, $A = \\{a^i\\}_{i=1}^N$ and $B = \\{b^i\\}_{i=1}^N$ , where $a^i$ are vectors of dimension $d_A$ and $b^i$ are vectors of dimension $d_B$ . We also introduce two Multi Layer Perceptrons (MLP), parameterized with $\\theta_A$ and $\\theta_B$ , to project these vectors onto a common space of dimension $d$ . After feeding the elements $a^i$ and $b^i$ to the MLPs and normalizing them, we obtain", + "bbox": [ + 496, + 113, + 893, + 295 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {a} ^ {i} = \\frac {\\operatorname {M L P} _ {\\theta_ {A}} \\left(a ^ {i}\\right)}{\\| \\operatorname {M L P} _ {\\theta_ {A}} \\left(a ^ {i}\\right) \\|} \\quad \\text {a n d} \\quad \\tilde {b} ^ {i} = \\frac {\\operatorname {M L P} _ {\\theta_ {B}} \\left(b ^ {i}\\right)}{\\| \\operatorname {M L P} _ {\\theta_ {B}} \\left(b ^ {i}\\right) \\|}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 303, + 892, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|$ denotes the $L_{2}$ norm. We define the per-element loss based on the relative similarity of $\\tilde{a}^i$ and $\\tilde{b}^i$ , and by using a temperature $\\tau$", + "bbox": [ + 496, + 347, + 893, + 393 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\ell} ^ {i} (A, B, \\theta_ {A}, \\theta_ {B}) = - \\log \\frac {\\exp \\left(\\frac {\\tilde {a} ^ {i} \\cdot \\tilde {b} ^ {i}}{\\tau}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\tilde {a} ^ {i} \\cdot \\tilde {b} ^ {j}}{\\tau}\\right)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 401, + 890, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then make the loss symmetric [43] by using", + "bbox": [ + 500, + 479, + 816, + 496 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\ell^ {i} (A, B, \\theta_ {A}, \\theta_ {B}) = \\tilde {\\ell} ^ {i} (A, B) + \\tilde {\\ell} ^ {i} (B, A). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 506, + 890, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we define the contrastive loss $\\mathcal{L}_{\\mathrm{cntr}}(A,B,\\theta_A,\\theta_B)$ as the mean of $\\ell^i$", + "bbox": [ + 500, + 535, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c n t r}} (A, B, \\theta_ {A}, \\theta_ {B}) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\ell^ {i} (A, B, \\theta_ {A}, \\theta_ {B}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 575, + 890, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned earlier on, for simplicity, in the rest of the paper we will not indicate the parameters of the MLPs, and simply write $\\mathcal{L}_{\\mathrm{cntr}}(A,B)$ or $\\ell^i (A,B)$ .", + "bbox": [ + 496, + 626, + 893, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Training SCALE", + "text_level": 1, + "bbox": [ + 500, + 680, + 668, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our method, we integrate 4 principles to drastically reduce the computational complexity of learning a video representation: Input sparsity, output sparsity, dimensionality reduction, and use of a pre-trained backbone. Moreover, we introduce two pseudo-tasks to train the model. One task is based on the (contrastive) reconstruction of a masked video clip given some context video clips. The second task is to build a global representation that is (contrastively) invariant to the set of input sampled video clips. The overall training scheme of SCALE is illustrated in Figure 1.", + "bbox": [ + 496, + 704, + 893, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input sparsity. As a first step, rather than processing a whole video, we collect a sparse set of short video clips from the same video. Given a video $V \\in \\mathbb{R}^{H \\times W \\times T \\times 3}$ ,", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5666", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ad3701cf1f5887dcfc3526852ced09c74df589e1fcf58ac08b692ec715456196.jpg", + "image_caption": [ + "Figure 1. Video Representation Learning with SCALE. For each video, SCALE extracts two sets of video clips $V_{1}^{1}, \\ldots, V_{K}^{1}$ and $V_{1}^{2}, \\ldots, V_{K}^{2}$ . Each video clip is processed separately through a frozen backbone $\\mathrm{E}_{\\text{基}}$ and results in encoded video clips $C_{1}^{1}, \\ldots, C_{K}^{1}$ and $C_{1}^{2}, \\ldots, C_{K}^{2}$ . Then, a random set of encodings in each set is masked and reconstructed at the output of the predictor network (a transformer) $(\\ell^{i})$ . The predictor network is also fed a class token CLS. The corresponding output token encodes a summary $\\mathrm{CLS}^{m}$ of the $m$ -th set of video clips. The objective for these summary tokens is to be similar only when encoding video clips from the same video $(\\mathcal{L}_{\\mathrm{SET}})$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 870, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H$ , $W$ and $T$ are the height, width, and duration (in frames) of the video, we sample $2 \\times K$ clips. We divide the clips into two sets randomly. Each clip in the first set $V_{i}^{1} \\in \\mathbb{R}^{H_{i}^{1} \\times W_{i}^{1} \\times T_{i}^{1} \\times 3}$ , with $i = 1, \\ldots, K$ , is obtained at the spatio-temporal location $X_{i}^{1}, Y_{i}^{1}, Q_{i}^{1}$ with different data augmentations and dimensions $H_{i}^{1}$ , $W_{i}^{1}$ and $T_{i}^{1}$ . Similarly, we denote the second set of clips $V_{i}^{2}$ , for $i = 1, \\ldots, K$ . We also normalize their coordinates relative to the video dimensions and embed them onto a feature vector $P_{i}^{j}$ by feeding them to a learnable MLP. We denote these embeddings", + "bbox": [ + 75, + 446, + 472, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} ^ {j} = \\operatorname {M L P} \\left(\\left[ \\frac {X _ {i} ^ {j}}{H}, \\frac {Y _ {i} ^ {j}}{W}, \\frac {Q _ {i} ^ {j}}{T}, \\frac {X _ {i} ^ {j} + H _ {i} ^ {j}}{H}, \\frac {Y _ {i} ^ {j} + W _ {i} ^ {j}}{W}, \\frac {Q _ {i} ^ {j} + T _ {i} ^ {j}}{T} \\right] ^ {\\top}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 611, + 468, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $j = 1,2$ and $i = 1,\\dots ,K$", + "bbox": [ + 76, + 672, + 302, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dimensionality reduction and pre-trained backbone. To reduce the dimensionality of each video clip, we feed them independently to a frozen encoder $\\mathrm{E}_{\\ast}$ , to obtain the encodings $C_{i}^{j} = \\mathrm{E}_{\\ast}(V_{i}^{j})$ , where $j = 1,2$ and $i = 1,\\dots ,K$ . Our framework is encoder-agnostic and thus can work with encoders obtained through different training schemes (supervised, contrastive, or autoencoder). In addition to reducing the dimensionality, we make the training even more scalable by using pre-trained and frozen encoders. Note, however, that if performance is the main goal, it is possible to also train a sparse backbone end to end with multiple clips. Thanks to token dropping, one can drop up to $95\\%$ of the tokens [19] and still build a good representation.", + "bbox": [ + 75, + 688, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output sparsity. As a self-supervised signal for our video", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "representation learning, we use a (contrastive) reconstruction loss. To reduce the computational cost, instead of predicting the features for the whole video [59, 19, 54, 16] (asymmetric decoding), we only reconstruct a sparse set of masked clips. Our reconstruction objective is based on the observation that video signals carry a lot of redundancy. Hence, we introduce a model, the predictor network, to predict masked video clip embeddings given the other video clip embeddings (the context). We follow the general approach of BERT [12] but implement the predictor network as a masked autoencoder, where the reconstruction is based on a contrastive loss. The loss is applied only to a sparse set $M^1 \\subset \\{1, \\dots, K\\}$ of masked video clips. These clips are replaced by a learned MSK token. All embeddings $C_i^1$ , including the masked ones, are added to their corresponding position encoding $P_i^1$ and are then fed to the predictor network. We also include an additional learnable CLS token as input for the predictor network, which will be used for tasks with multiple video clips. We denote the outputs of the predictor network as $\\hat{C}_i^1$ for the tokens corresponding to $C_i^1$ , and as $\\mathrm{CLS}^1$ for the token corresponding to CLS. Similarly, we feed as inputs separately from the previous set all the video clips $C_i^2$ with their corresponding positional embeddings $P_i^2$ , and the same CLS token, and obtain $\\hat{C}_i^2$ and $\\mathrm{CLS}^2$ respectively (see Figure 1 for a visual depiction of these processing steps).", + "bbox": [ + 496, + 446, + 893, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive reconstruction. Modeling all the details of a masked clip, even in the latent space and even given the redundancy in videos, is a demanding task. Rather than", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5667", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "increasing the capacity of our model, since we are aiming for scalability, we keep our predictor network a (relatively) shallow network and use contrastive learning [41]. With contrastive learning, the predicted representation of the masked tokens should only be closer to the original un-masked clip representation (after an MLP projection) than from all other clips from the same video and the rest of the minibatch. Note that the rest of the clips in the same video act as hard negatives in contrastive learning [46]. Also, since we are using a frozen backbone, we can afford to use large minibatch sizes, which is known to be beneficial for contrastive learning [9]. We call this contrastive reconstruction loss the Masked Clip Modeling (MCM) loss", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M C M}} = \\sum_ {i \\in M ^ {1}} \\ell^ {i} \\left(\\hat {C} ^ {1}, C ^ {1}\\right) + \\sum_ {j \\in M ^ {2}} \\ell^ {j} \\left(\\hat {C} ^ {2}, C ^ {2}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 295, + 468, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multiple video clips loss. The predictor network outputs features for each video clip that are highly discriminative. This task is similar to that of a masked autoencoder [25] and gives you an enhanced per-clip representation. For many video tasks, we need a global representation for the whole video; for that, we introduce an additional pseudo-task that captures a more global representation of a set of video clips. Our task takes inspiration from contrastive learning methods used in SSL, which yield representations that perform well with linear probing [8]. The loss aims to make the $\\mathrm{CLS}^1$ and $\\mathrm{CLS}^2$ tokens returned from the predictor network more similar (recall that these two tokens are obtained from two separate groups of video clips extracted from the same video) than to other class tokens from other videos within the minibatch", + "bbox": [ + 75, + 335, + 472, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S E T}} = \\mathcal {L} _ {\\mathrm {c n t r}} \\left(\\mathrm {C L S} ^ {1}, \\mathrm {C L S} ^ {2}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 570, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to our contrastive loss (InfoNCE), one can use clustering [7] or regression [23] losses. We choose InfoNCE for simplicity and for better compatibility between the losses. As the overall loss, we use the sum of both loss terms (without any weights)", + "bbox": [ + 75, + 597, + 468, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M C M}} + \\mathcal {L} _ {\\mathrm {S E T}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 681, + 468, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 708, + 209, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate SCALE on several commonly used action classification datasets for video representation learning. As our performance metric, we primarily use linear probing and nonlinear probing [27]. For the smaller datasets, we also use $k$ -NN classifiers (which are training-free) and demonstrate that the proposed method improves upon both unsupervised and supervised backbones.", + "bbox": [ + 75, + 732, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup and Protocols", + "text_level": 1, + "bbox": [ + 76, + 847, + 377, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computational Efficiency: In Table 1, we show estimates of the maximum batch sizes and the training throughput for", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/749833dfbe6aededec5c035b4915f9efa8de72ebe4c987fedd9cf9c37c6b2747.jpg", + "table_caption": [ + "Table 1. Computational Resources. GPU VRAM usage (GB) and max training speed (using one 3090 GPU) of $\\mathrm{ViT_B}$ with varying batch sizes and SSL tasks for videos. $\\mathrm{MoCo_{Sparse}^{V3}}$ is akin to MSN [3]. OOM indicates Out Of Memory error." + ], + "table_footnote": [], + "table_body": "
Batch Size819572048Samples/s
MoCoV323.47OOMOOMOOM8.66
VMAE11.4222.80OOMOOM33.54
MoCoV3Sparse5.499.2723.62OOM28.83
SCALE1.211.231.5819.529224.65
", + "bbox": [ + 500, + 142, + 893, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "different methods trained with the same computational and memory resources. As can be seen, SCALE is orders of magnitude more efficient than other SotA methods. Also during multi-crop inference, our method only results in an FLOP increase of approximately $0.001\\%$ .", + "bbox": [ + 496, + 262, + 893, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: Following prior work [44, 18, 45] we use Kinetics-400 [33], UCF-101 [48] (split 1), HMDB-51 [34] (split 1), and Something-Something v2 (SSv2) [22] to train and evaluate our models. We also use the LVU benchmark [63] to showcase our long-form video understanding capabilities. Note that almost $35\\%$ of LVU videos are not available to download from YouTube anymore; thus, our results are not directly comparable with prior methods.", + "bbox": [ + 496, + 338, + 893, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretrained backbones: We use the pretrained checkpoints of $\\rho$ BYOL [18], SVT [44], and three variants of Video-MAE [54] (base(B), large(L), and fine-tuned base(FT)). We choose $\\rho$ BYOL for their excellent linear performance, SVT for the usage of ViT [14], and VMAE for showing 1) the applicability of our proposed method to MAE models, 2) the scalability of our method to larger models, and 3) possibility of using supervisedly fine-tuned models as our backbone. All the models are self-supervisedly pretrained on Kinetics-400, except the fine-tuned VMAE base that was also supervisedly finetuned on Kinetics-400. We also used a backbone pretrained and fine-tuned on SSv2 (VMAE $_{SSv2}$ ) for the SSv2 experiment to show the universality of SCALE with respect to the pretraining dataset.", + "bbox": [ + 496, + 460, + 893, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-supervised Training: For training data, we extract 16 clips of 16 frames from each video per dataset and save their feature encodings to disk. We use PySlowFast's common data augmentations for that [15]. For evaluation, we follow the $5 \\times 3$ scheme [17] (uniformly sampling 5 clips from a video along its temporal axis and then taking 3 spatial crops) and extract 15 clips from each video (except for SSv2, where we extract $2 \\times 3$ clips [58]). As the architecture for the predictor network, we use an encoder-only Transformer [56] and a three-layer MLP (without batch normalization [30]) for the contrastive heads. Unless stated otherwise, we train our models for 500 epochs (for example, training with $\\mathrm{VMAE}^{\\mathrm{B}}$ on SSV2 takes 137 minutes with one 3090 GPU) with a batch size of 512 and use all 16 clips.", + "bbox": [ + 496, + 672, + 893, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation: Since our focus is on efficient and scalable", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5668", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/92065676fb9af73983265bbd397d796bc3142b315d3b4a4458a4c7da6fa15804.jpg", + "table_caption": [ + "Table 2. Long-Form Video Understanding Results. Linear and nonlinear probing accuracies on LVU [63] classification tasks. SCALE shows significant performance improvement compared to baselines, indicating its ability to capture long-form video features. The first two rows are greyed out as direct comparison is not possible due to the unavailability of the full dataset for download." + ], + "table_footnote": [], + "table_body": "
RelationSpeakSceneDirectorGenreWriterYear
SlowFast+NL ViS4mer[17, 60]52.4035.8054.7044.9053.0036.3052.50
[31]57.1440.7967.4462.6154.7148.8044.75
SVTLinear64.7035.7762.3337.2754.3552.1228.46
SCALE linear73.52 (+8.82)40.65 (+4.88)68.83 (+6.50)46.36 (+9.09)57.94 (+3.59)56.38 (+4.26)39.23 (+10.77)
MLP67.6439.0266.2345.4556.9257.4436.15
Transformer70.5840.6568.8347.2757.1758.5136.92
SCALE ft76.47 (+5.89)42.27 (+1.62)74.02 (+5.19)49.09 (+1.82)58.97 (+1.80)62.76 (+4.25)39.23 (+2.31)
ρBYOLLinear52.9438.2153.2436.3651.7957.4433.84
SCALE linear67.64 (+14.70)43.08 (+4.87)66.23 (+12.99)44.54 (+8.18)53.33 (+1.54)60.63 (+3.19)40.00 (+6.16)
MLP62.3541.4662.4247.2752.5659.5740.00
Transformer65.0944.7166.2350.9053.0761.7043.84
SCALE ft67.64 (+2.55)45.52 (+0.81)71.42 (+5.19)51.81 (+0.91)55.72 (+2.65)65.95 (+4.25)46.92 (+3.08)
", + "bbox": [ + 76, + 128, + 903, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "video classification, we always freeze the backbones in our evaluation (as in our self-supervised pretraining) and either train a linear classifier [44, 18] or fine-tune the predictor network (the transformer) with an additional linear head. Therefore, when we refer to fine-tuning (ft), we only adapt the nonlinear head (e.g., predictor network) but not the backbone. We apply a grid search for the hyper-parameters of the heads covering learning rate, weight decay, batch size, and optimizer type. Similar to MAE [25], we found that applying a batch normalization layer [30] without affine transformations is beneficial for VideoMAE models. As the linear baseline, we consider the well-established ensembling approach, i.e., we average the softmax predictions of the 15 clips (6 for SSv2) to obtain the final prediction. For models that process multiple clips at once (like ours), we likewise apply a linear softmax head on the concatenation of the individual clip features and the [SET] token before averaging to obtain the final prediction. For the smaller datasets, we also use $k$ -NN classification, where, similar to DINO [7], we always use $k = 20$ and work with 12 normalized representations.", + "bbox": [ + 75, + 330, + 472, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nonlinear baselines: As SCALE is a nonlinear model, we consider an MLP on top of the frozen backbone as a nonlinear baseline. As a further baseline and to illustrate the effect of our self-supervised pre-training, we consider a Transformer trained on all the clip representations. This Transformer uses the exact same architecture as SCALE, and only differs in the initialization: in the case of SCALE we start from our proposed SSL pre-trained weights instead of random initialization.", + "bbox": [ + 75, + 650, + 468, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 76, + 800, + 171, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LVU: One of the benefits of SCALE is that it can be used to process long videos, even though the backbones were trained on short videos only. To demonstrate the ability to capture long-form video features, we evaluated SCALE on LVU [63], a benchmark that involves seven classifica", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/af92a4b99b5208c81c76fcc4c5b1a2f6510ea88a0a855dc4cbd3c4392eea3915.jpg", + "table_caption": [ + "Table 3. SSv2 Results. Linear and nonlinear probing accuracies on SSv2 [22]. We see that both SCALE linear and SCALE ft outperform other methods and improve the classification accuracies by a large margin. We also see that SCALE ft, with its better initialization, always outperforms the Transformer. $\\mathrm{VMAE}_{\\mathrm{SSv2}}^{\\mathrm{B}}$ was pretrained and fine-tuned on SSv2." + ], + "table_footnote": [], + "table_body": "
SVTρBYOLVMAEBVMAELVMAEftVMAEBSSv2
Linear20.3025.3018.3127.9428.9070.53
SCALE linear25.2627.1621.2430.1833.2570.63
MLP21.4326.4619.4227.9629.8370.52
Transformer29.2430.9924.2634.3935.6070.57
SCALE ft29.6831.8325.2536.3437.3870.69
", + "bbox": [ + 500, + 410, + 890, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion (and two regression) tasks on minute-long videos. Past studies [31, 60] have established that increasing the input's time span enhances accuracy in this challenging dataset. As shown in Table 2, our experiments indicate that SCALE can improve the baseline model's performance by a considerable margin. Moreover, we found that fine-tuning SCALE can lead to further enhancements.", + "bbox": [ + 496, + 534, + 890, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SSv2: Multiple classes in SSv2 share similar backgrounds and only differ in motion [28], suggesting that high performance on this dataset demonstrates that the model has captured strong motion-related contextual cues [44]. Results in Table 3 show that we outperform the state-of-the-art. On this dataset, we see a large performance gap between models that process single clips at a time (Linear and MLP) and the models that work with multiple clips (SCALE and Transformer). We can see SCALE linear is also outperforming the MLP, showing that SCALE is able to capture motion and long-form temporal features of the video. We even improve the supervised model trained on SSv2 $\\mathrm{(VMAE_{SSv2}^B)}$", + "bbox": [ + 496, + 641, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UCF-101 & HMDB-51: For these smaller datasets, besides linear and nonlinear probing, we also use $k$ -NN probing (see Table 4 and Table 5). With SCALE $k$ -NN, we see a consistent improvement over the baseline and find that pre-trained MAE-based models greatly benefit from our", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5669", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ada5b02e549ae549e59cce9d81c1df747e671d32a5a93b5497ef58c04e74a85a.jpg", + "table_caption": [ + "Table 4. UCF Results. Linear and nonlinear probing accuracies on UCF-101 [48]. SCALE fit outperforms all the other models and, in the case of $\\rho$ BYOL, even gets performance close to a fully fine-tuned model. Also, in most cases, SCALE linear outperforms the fine-tuned Transformer and achieves state-of-the-art results in linear probing (previous SotA using RGB frames was 92.6 [45]). We further see a significant accuracy improvement in $k$ -NN probing, especially for pre-trained MAE-based models. As a point of reference, the current best fully fine-tuned accuracy (which is not comparable with our setting) is $96.8\\%$ [57]." + ], + "table_footnote": [], + "table_body": "
SVTρBYOLVMAEBVMAELVMAEft
k-NN87.2085.1935.0549.1496.82
SCALE k-NN89.0083.4765.6376.0297.38
Linear91.2789.5566.5384.5397.91
SCALE linear92.6591.4374.4686.7898.14
MLP91.1793.6071.9787.0498.04
Transformer92.2094.3468.2286.3098.04
SCALE ft92.9495.0076.0789.9298.46
FTreported93.795.496.1--
", + "bbox": [ + 76, + 224, + 470, + 367 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/49d5ea81ff2d75043b9e1ed22652893e29ebf36c1bf7574ebf91e801ac1a0921.jpg", + "table_caption": [ + "Table 5. HMDB Results. Linear and nonlinear probing accuracies on HMDB-51 [34]. Despite the small size of the dataset, we see that $\\mathrm{SCALE}_{\\mathrm{ft}}$ is outperforming all the other methods, and in the case of SVT, it even outperforms the fully fine-tuned model. We also see that $\\mathrm{SCALE}_{\\mathrm{linear}}$ outperforms the Transformer in most cases with only a single linear layer (the best linear accuracy in the literature is $66.7\\%$ [45]). Similar to UCF results, we see a considerable increase in the performance of $k$ -NN classifiers for pre-trained MAE-based models." + ], + "table_footnote": [], + "table_body": "
SVTρBYOLVMAEBVMAELVMAEft
k-NN51.8349.6721.9629.2172.81
SCALE k-NN56.0151.5637.1851.3071.83
Linear63.0761.1745.2260.2676.33
SCALE linear66.3363.9252.1562.3578.36
MLP63.0064.7749.0162.6177.45
Transformer63.9866.1647.3261.5076.86
SCALE ft68.1066.7951.8964.8379.34
FTreported67.273.673.3--
", + "bbox": [ + 76, + 505, + 470, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training. This can be explained by the additional invariance properties introduced through the SET loss term in SCALE training. Across the board, we also see that in the case of linear probing, not only does SCALE linear outperform Linear, but it also outperforms Transformer, which leverages many more parameters. In the case of SVT, our SCALE linear also outperforms the best reported linear accuracy on UCF101 (92.7% vs. 92.6% [45]). Finally, $\\mathrm{SCALE}_{\\mathrm{ft}}$ achieves better results than all the nonlinear baselines and even outperforms the fully fine-tuned SVT (68.1% vs. 67.2%).", + "bbox": [ + 75, + 672, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Kinetics-400: We present our main results on Kinetics-400 [33] in Table 6. Our SCALE linear with SVT backbone beats the previous state of the art (71.8% vs. 71.5% [18]) and SCALE ft can even improve the accuracy of $\\mathrm{VMAE}_{\\mathrm{ft}}^{\\mathrm{B}}$", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8baf8edbbb85a91b209ff1c612b1149ac8c9156cbe7ffd1a5848194b0da0f1fa.jpg", + "table_caption": [ + "Table 6. Kinetics-400 Results. Linear and nonlinear probing accuracies on Kinetics-400 [33] without any extra data and using RGB frames only. While SCALE linear is on par with Linear, we observe clear improvements for nonlinear probing in the case of SCALE ft. Note that the best linear accuracy on this dataset (without any extra data) is 71.5 [18] and the best full fine-tuning accuracy is 86.7 [61]." + ], + "table_footnote": [], + "table_body": "
SVTρBYOLVMAEBVMAELVMAEftb
Linear71.7168.8243.5060.7381.52
SCALE linear71.7868.3843.9660.6681.44
MLP71.1969.4245.4861.6481.27
Transformer72.1869.2844.8562.1581.70
SCALE ft72.3869.6346.1562.6781.84
", + "bbox": [ + 500, + 184, + 890, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5646ec0ffda4941dd3c8120f6e525249126bd9765e5f93aa64a1992a781f3e32.jpg", + "table_caption": [ + "Table 7. Kinetics-400 Low-shot Results. Linear and nonlinear probing accuracies on $10\\%$ of Kinetics-400 [33]. SCALE is more robust to the size of the labeled dataset. $\\mathrm{SCALE}_{\\mathrm{ft}}$ does not overfit like the other nonlinear probes (MLP and Transformer) and outperforms the baselines. $\\mathrm{VMAE}_{\\mathrm{ft}}^{\\mathrm{B}}$ is greyed out since it was already fine-tuned on the whole labeled dataset and is only reported here for consistency with the other tables." + ], + "table_footnote": [], + "table_body": "
SVTρBYOLVMAEBVMAELVMAEft
Linear66.4356.4331.2548.4279.79
SCALE linear65.9657.7434.0349.2179.94
MLP65.4458.6830.4748.2779.37
Transformer64.9758.9529.8948.2779.47
SCALE ft67.0159.5233.9250.3680.47
", + "bbox": [ + 500, + 386, + 890, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which is a strong supervised model, from $81.5\\%$ to $81.84\\%$ .", + "bbox": [ + 496, + 505, + 888, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the evaluation setup of self-supervised image representations [8, 7, 71], we also introduce low-shot K400 video classifications by sampling 10 percent of the videos (in a class-balanced way) and training the probes only on those. We still test on the whole evaluation set of K400. This low-shot setting is more aligned with the typical use-case of self-supervised models in which there is abundant unlabeled data for training via self-supervision and a small set of labeled data for fine-tuning. Results in Table 7 show that our method is particularly effective in this low-shot setting. While most other nonlinear probes overfit and perform worse than the linear probes, our SCALE ft does not overfit and clearly outperforms the baselines.", + "bbox": [ + 496, + 521, + 890, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 500, + 726, + 609, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we start from a baseline setup consisting of a two-layer transformer with a hidden size of 256, $20\\%$ chance of masking clips, trained with a batch size of 512 for 200 epochs, and using two sets of 8 views for representation learning. Using SCALE ft, we explore different loss functions, masking ratios, number of layers, and finally, the number of views during training and testing. All experiments are performed on UCF and HMDB.", + "bbox": [ + 496, + 750, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss Function: As explained in the method section, we have two loss terms, and each of them can be enabled or", + "bbox": [ + 496, + 869, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5670", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd4ed6b28e9cf28cc3ec44236b723b3a830767dad5c18ed354c9689bd28564f2.jpg", + "table_caption": [ + "Table 8. Loss Function. SCALE fit accuracy with different loss function combinations (the masking ratio here is $20\\%$ ). We can see that having MCM is always beneficial, and the SET loss is almost always helpful. We use both loss terms for our final model." + ], + "table_footnote": [], + "table_body": "
SETMCMUCF-101HMDB-51
SVTρBYOLSVTρBYOL
✓✗91.8092.2064.5063.59
✗✓92.0193.8162.8164.05
✓✓93.2092.9964.5765.61
", + "bbox": [ + 76, + 152, + 468, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6db47fa8f338658d26cd59683fb5897a9b9aeb7271a8efb202dc1bec125f964b.jpg", + "table_caption": [ + "Table 9. Masking Ratio. SCALEft accuracy with different masking ratios. We observe best results around 25% similar to NLP models [12] (15%), and different from low-level video models like VideoMAE [54] (90%)." + ], + "table_footnote": [], + "table_body": "
Masking RatioUCF-101HMDB-51
SVTρBYOLSVTρBYOL
0.1593.1893.2564.3764.83
0.2593.2093.8165.4965.62
0.3593.1593.0664.1865.22
0.4592.9693.0263.3964.96
", + "bbox": [ + 76, + 295, + 468, + 402 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/49d39c211f932285f84af6ee1a1994958231c389e7acc6d64fca7c3fa19881f6.jpg", + "table_caption": [ + "Table 10. Transformer Capacity. SCALEft accuracy with different model capacities. Having more than one transformer layer and not too few hidden channels is necessary for the best performance." + ], + "table_footnote": [], + "table_body": "
Hidden \nDimNum \nLayersUCF-101HMDB-51
SVTρBYOLSVTρBYOL
641-92.62-63.16
1281-92.83-63.68
2561-92.86-64.39
128292.7893.5263.2665.55
256293.2093.8165.4965.62
512292.5793.6665.6864.77
128392.3393.2564.8365.55
256392.7593.5265.4965.16
512392.8692.9365.4964.84
", + "bbox": [ + 76, + 465, + 470, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "disabled for the pretraining. In Table 8 we show that having both loss terms is better than the individual loss terms.", + "bbox": [ + 75, + 672, + 467, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Masking Ratio: Masking ratio is an important hyperparameter and depends on the data modality, for example, BERT [12] uses $15\\%$ , MSN [3] uses $30\\%$ (for ViT-Base), MAE [25] uses $75\\%$ , and VideoMAE [54] uses 90 to $95\\%$ masking. Since our clip representations are somewhat abstract representations of the video, we expect the optimal masking ratio to be close to NLP models rather than video MAEs. We have observed a steady decrease in the pretraining task's performance with higher masking ratios, so we only tested low masking ratios in Table 9 and found out that $25\\%$ is the optimal masking ratio.", + "bbox": [ + 75, + 703, + 467, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transformer Capacity: We also explore the number of", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/55940f158f1307601fe757b30501859598002bec76735fcf3611c12f2734b4dc.jpg", + "table_caption": [ + "Table 11. Number of Views. SCALE ft accuracy with different numbers of clips and batch sizes. More views lead to consistent improvement, and large batch sizes are not necessary because of the hard negative samples." + ], + "table_footnote": [], + "table_body": "
Num ViewsBatch SizeUCF-101HMDB-51
SVTρBYOLSVTρBYOL
4 × 225692.6592.8364.3564.24
6 × 225692.7093.0764.5764.37
8 × 225692.8093.4964.6464.63
4 × 251292.6793.4964.8564.50
6 × 251293.1893.6864.9065.35
8 × 251293.2093.8165.4965.62
4 × 2102492.7593.3664.7765.15
6 × 2102492.9693.5764.9665.48
8 × 2102493.07OOM65.29OOM
", + "bbox": [ + 500, + 142, + 893, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "transformer layers and their hidden size in Table 10. We can see that having more than one transformer layer is necessary for good results and too few hidden channels can hurt performance. However, there is a trade-off, and deeper transformers can lead to worse performance.", + "bbox": [ + 496, + 357, + 890, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of Views: Finally, we studied the model performance as we changed the number of views and batch size fed to the model. As can be seen in Table 11, having more views has a large and consistent impact on the performance, and since we have hard negatives for contrastive loss within the video, we are not too reliant on large batch sizes.", + "bbox": [ + 496, + 436, + 890, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 553, + 619, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced SCALE, a framework for video representation learning that aggregates the information from multiple clips from the same video. We combine contrastive learning and masked modeling with intuitions from predictive coding to obtain improved global and local representations of clips starting from frozen backbones. We evaluated these features using a wide array of backbones on different action classification and video understanding datasets and achieved strong or state-of-the-art results. The computational efficiency of our method is extremely useful for videos and opens the possibility to a wider group of researchers (with limited computational resources) to work on video representation learning than previously possible. We also believe that working with a set of clips is an interesting direction for representation learning. Finally, as a surprising and maybe alarming observation, even contrastive representations that were trained to be invariant to data augmentations and spatio-temporal crops can be used for contrastive masked modeling. This might be due to benign memorization [2], and understanding why this phenomenon happens might lead to a better understanding of contrastive learning.", + "bbox": [ + 496, + 583, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5671", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 76, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by grant 200020_188690 of the Swiss National Science Foundation (SNSF) and an Adobe award.", + "bbox": [ + 76, + 114, + 468, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 172, + 174, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hassan Akbari, Linagzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. ArXiv, abs/2104.11178, 2021. 1", + "[2] Sotiris Anagnostidis, Gregor Bachmann, Lorenzo Noci, and Thomas Hofmann. The curious case of benign memorization. ArXiv, abs/2210.14019, 2022. 8", + "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. 1, 5, 8", + "[4] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zach Beaver, Jana von Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, Vivek Natarajan, and Mohammad Norouzi. Big self-supervised models advance medical image classification. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3458-3468, 2021. 1", + "[5] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. ArXiv, abs/2106.08254, 2022. 1", + "[6] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9922-9931, 2020. 2", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1, 2, 5, 6, 7", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 5, 7", + "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2, 5", + "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 1", + "[11] Ishan Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TCLR: Temporal contrastive learning for video representation. Computer Vision and Image Understanding, 219:103406, 2022. 2", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + ], + "bbox": [ + 78, + 198, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3, 4, 8", + "[13] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. ArXiv, abs/2207.07116, 2022. 2, 3", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 5", + "[15] Haoqi Fan, Yanghao Li, Bo Xiong, Wan-Yen Lo, and Christoph Feichtenhofer. Pyslowfast. https://github.com/facebookresearch/slowfast, 2020.5", + "[16] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022. 1, 2, 3, 4", + "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6201-6210, 2019. 5, 6", + "[18] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021. 1, 2, 3, 5, 6, 7", + "[19] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimae: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 1, 2, 4", + "[20] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13505-13515, 2021. 3", + "[21] Shreyank N Gowda, Marcus Rohrbach, and Laura Sevilla-Lara. Smart frame selection for action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1451-1459, 2021. 3", + "[22] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 5, 6", + "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 5", + "[24] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in Neural Information Processing Systems, 33:5679-5690, 2020. 2", + "[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5672", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1, 2, 3, 5, 6, 8", + "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2", + "[27] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. ArXiv, abs/1808.06670, 2019. 5", + "[28] Kaiqin Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7929, 2021. 6", + "[29] Deng Huang, Wenhao Wu, Weiwen Hu, Xu Liu, Dongliang He, Zhihua Wu, Xiangmiao Wu, Mingkui Tan, and Errui Ding. Ascnet: Self-supervised video representation learning with appearance-speed consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8096-8105, 2021. 1", + "[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 5, 6", + "[31] Md. Mohaiminul Islam and Gedas Bertasius. Long movie clip classification with state-space video models. ArXiv, abs/2204.01692, 2022. 6", + "[32] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9970-9980, 2021. 2", + "[33] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 5, 7", + "[34] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 2, 5, 7", + "[35] Xingbin Liu, Jinghao Zhou, Tao Kong, Xianming Lin, and Rongrong Ji. Exploring target representations for masked autoencoders. arXiv preprint arXiv:2209.03917, 2022. 1", + "[36] Yue Liu, Junqi Ma, Yufei Xie, Xuefeng Yang, Xingzhen Tao, Lin Peng, and Wei Gao. Contrastive predictive coding with transformer for video representation learning. Neurocomputing, 482:154-162, 2022. 3", + "[37] Guillaume Lorre, Jaxonary Rabarisoa, Astrid Orcesi, Samia Ainouz, and Stephane Canu. Temporal contrastive pretraining for video action recognition. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 662-670, 2020. 1, 2, 3", + "[38] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "verification. In European conference on computer vision, pages 527-544. Springer, 2016. 2", + "[39] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2", + "[40] Charlie Nash, João Carreira, Jacob Walker, Iain Barr, Andrew Jaegle, Mateusz Malinowski, and Peter Battaglia. Transframer: Arbitrary frame prediction with generative models. arXiv preprint arXiv:2203.09494, 2022. 3", + "[41] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 2, 3, 5", + "[42] Mandela Patrick, Yuki M. Asano, Bernie Huang, Ishan Misra, Florian Metze, João F. Henriques, and Andrea Vedaldi. Space-time crop & attend: Improving cross-modal video representation learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10540-10552, 2021. 3", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[44] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2874-2884, 2022. 1, 2, 3, 5, 6", + "[45] Adria Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Pătrăucean, Florent Altché, Michal Valko, et al. Broaden your views for self-supervised video learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1255-1265, 2021. 1, 2, 3, 5, 7", + "[46] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. ArXiv, abs/2010.04592, 2021. 5", + "[47] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In European Conference on Computer Vision, pages 154-171. Springer, 2020. 3", + "[48] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 5, 7", + "[49] Chen Sun, Arsha Nagrani, Yonglong Tian, and Cordelia Schmid. Composable augmentation encoding for video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8834-8844, 2021. 3", + "[50] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12607-12617, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5673", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 1, 3", + "[52] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. arXiv preprint arXiv:2206.01204, 2022. 3", + "[53] Nenad Tomasev, Ioana Bica, Brian McWilliams, Lars Holger Buesing, Razvan Pascanu, Charles Blundell, and Jovana Mitrovic. Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels on imageNet? In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 1", + "[54] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. 1, 2, 3, 4, 5, 8", + "[55] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2", + "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[57] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14010-14020, 2022. 3, 7", + "[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:2740-2755, 2019. 5", + "[59] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14713-14723, 2022. 1, 4", + "[60] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 6", + "[61] Chen Wei, Haoqi Fan, Saining Xie, Chaoxia Wu, Alan Loddon Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14648-14658, 2022. 7", + "[62] Donglai Wei, Joseph J Lim, Andrew Zisserman, and William T Freeman. Learning and using the arrow of time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8052-8060, 2018. 2", + "[63] Chaoxia Wu and Philipp Krahenbuhl. Towards long-form video understanding. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1884-1894, 2021. 2, 5, 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 3", + "[65] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 1, 3", + "[66] Yu Wu, Linchao Zhu, Xiaohan Wang, Yi Yang, and Fei Wu. Learning to anticipate egocentric actions by imagination. IEEE Transactions on Image Processing, 30:1143-1152, 2020. 3", + "[67] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3333-3343, 2022. 3", + "[68] Liangzhe Yuan, Rui Qian, Yin Cui, Boqing Gong, Florian Schroff, Ming-Hsuan Yang, Hartwig Adam, and Ting Liu. Contextualized spatio-temporal contrastive learning with self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13977-13986, 2022. 3", + "[69] Yujia Zhang, Lai-Man Po, Xuyuan Xu, Mengyang Liu, Yexin Wang, Weifeng Ou, Yuzhi Zhao, and Wing-Yin Yu. Contrastive spatio-temporal pretext learning for self-supervised video representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3380-3389, 2022. 3", + "[70] Zhang Zhang and Dacheng Tao. Slow feature analysis for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 34(3):436-450, 2012. 1", + "[71] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 7", + "[72] Pan Zhou, Yichen Zhou, Chenyang Si, Weihao Yu, Teck Khim Ng, and Shuicheng Yan. Mugs: A multi-granular self-supervised learning framework. arXiv preprint arXiv:2203.14415, 2022. 1" + ], + "bbox": [ + 501, + 92, + 893, + 696 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5674", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/6732d370-6205-414c-986c-288f16a786f3_model.json b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/6732d370-6205-414c-986c-288f16a786f3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..77fd13af7540a189028a84be6ed28c716b331cc9 --- /dev/null +++ b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/6732d370-6205-414c-986c-288f16a786f3_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.131, + 0.844, + 0.154 + ], + "angle": 0, + "content": "Spatio-Temporal Crop Aggregation for Video Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.182, + 0.335, + 0.199 + ], + "angle": 0, + "content": "Sepehr Sameni" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.2, + 0.348, + 0.216 + ], + "angle": 0, + "content": "University of Bern" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.221, + 0.372, + 0.233 + ], + "angle": 0, + "content": "sepehr.sameni@unibe.ch" + }, + { + "type": "text", + "bbox": [ + 0.439, + 0.182, + 0.54, + 0.198 + ], + "angle": 0, + "content": "Simon Jenni" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.199, + 0.556, + 0.216 + ], + "angle": 0, + "content": "Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.22, + 0.558, + 0.234 + ], + "angle": 0, + "content": "jenni@adobe.com" + }, + { + "type": "text", + "bbox": [ + 0.648, + 0.182, + 0.756, + 0.198 + ], + "angle": 0, + "content": "Paolo Favaro" + }, + { + "type": "text", + "bbox": [ + 0.627, + 0.199, + 0.776, + 0.217 + ], + "angle": 0, + "content": "University of Bern" + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.22, + 0.796, + 0.234 + ], + "angle": 0, + "content": "paolo.favaro@unibe.ch" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.301, + 0.474, + 0.573 + ], + "angle": 0, + "content": "We propose Spatio-temporal Crop Aggregation for video representation Learning (SCALE), a novel method that enjoys high scalability at both training and inference time. Our model builds long-range video features by learning from sets of video clip-level features extracted with a pretrained backbone. To train the model, we propose a self-supervised objective consisting of masked clip feature predictions. We apply sparsity to both the input, by extracting a random set of video clips, and to the loss function, by only reconstructing the sparse inputs. Moreover, we use dimensionality reduction by working in the latent space of a pre-trained backbone applied to single video clips. These techniques make our method not only extremely efficient to train but also highly effective in transfer learning. We demonstrate that our video representation yields state-of-the-art performance with linear, nonlinear, and \\(k\\)-NN probing on common action classification and video understanding datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.603, + 0.21, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Videos provide rich and detailed information about objects and their activities. Their analysis is, however, made challenging not only by the difference in the information provided across space and time but also by the high dimensionality of the data [45, 29, 70]. While computational resources are expected to scale over time, so is the demand for higher data resolution (both in space and time) and also the need for processing data with even more dimensions, such as videos of volumetric data [10]. Therefore, it is of paramount importance to explore methods that drastically reduce the computational requirements to process videos. Moreover, the annotation of videos is an extremely costly and time-consuming burden that makes the use of models pre-trained in a self-supervised manner essential [44]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) is a very popular technique to reduce the need for annotation because it can build useful representations from unlabeled data through an artificial goal, also called pseudo- or pretext-task. These repre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.376 + ], + "angle": 0, + "content": "sentations can either be evaluated through K-nearest neighbor or linear probing [71, 72] or through fine-tuning (i.e., as the initialization parameters of the trained model) [25, 54, 35] on a downstream task, where only a small labeled dataset is available. More remarkably, SSL pre-trained models can outperform models that were pre-trained on an annotated dataset [53, 4]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.895, + 0.663 + ], + "angle": 0, + "content": "SSL methods for video representation learning present fundamental scalability challenges [54, 45, 29, 65]. A first major challenge is that training models from scratch for any new pseudo-task is not sustainable. A more viable setting is one where the processing of videos is broken down into multiple steps. First, one pre-processes video clips (spatiotemporal crops) only once via some pre-trained general-purpose model (e.g., trained via self-supervised learning [18, 23, 26, 44, 7, 37, 41]). In general, the idea is that this (compressed) local representation is stored and used later for other training purposes or retrieval. Then, one can train more efficiently other super-features that aggregate information from these local representations. Within this setting, we address two questions: 1) Can one further improve the performance of pre-computed video features by training a model on top of them? 2) Can such training be made computationally scalable? In this paper, we show that both questions have a positive answer. The answer exploits four fronts:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.676, + 0.895, + 0.813 + ], + "angle": 0, + "content": "1. Input Sparsity: Sparsity in the input to the model [54, 25, 3, 19, 1] drastically reduces computational load and memory requirements, leveraging information redundancy in images and videos [16]. Inspired by prior work, we extract a sparse set of clips from a video (these are spatio-temporal video crops and not image patches or video tubelets); each clip is fed separately to a neural network so that we obtain a set of video clip representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "2. Output Sparsity: Another way to reduce the computational cost is to use a sparse reconstruction output instead of a dense one [71, 51, 59, 5]. This is not just a reduction of the number of terms in the loss function, but also in the number of actual outputs of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5664" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.213 + ], + "angle": 0, + "content": "the model, and thus also a reduction of the computations needed to obtain them. This is in contrast to MAE-based SSL methods for vision, where the proposed pseudo-tasks are based on the reconstruction of the whole input [25, 54, 19, 16], and even if the loss uses a subset of the tokens (the masked ones) for the loss calculation, the remaining tokens are still part of the decoder's output and computational graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.47, + 0.297 + ], + "angle": 0, + "content": "3. Dimensionality Reduction: Inspired by prior work [71, 41, 13], instead of directly processing the raw input data, we work in the latent space. This allows us to further reduce the dimensionality of both input and output data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.305, + 0.468, + 0.425 + ], + "angle": 0, + "content": "4. Use of a Pre-trained Backbone: To reduce training time and further speed up the processing per iteration, we exploit SSL pre-trained models. These pre-trained backbones already learn very strong short-term spatiotemporal features, and our approach is a way to extract longer-term features (see LVU results 2) by building a video representation on top of a set of pre-trained features (one for each video clip)." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.221, + 0.47, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.433, + 0.468, + 0.643 + ], + "angle": 0, + "content": "To integrate all these components, we propose a novel SSL method that we call Spatio-temporal Crop Aggregation for video representation LEarning (SCALE). Given an input video, SCALE extracts a random set of clips and produces an embedding for each clip through a pre-trained backbone, which is kept frozen. These initial embeddings are then augmented in two ways: 1) each one is refined into a more discriminative feature, and 2) the set of all embeddings is summarized in a global feature. These global features can learn long-term correlations in the whole video by aggregating the short-term information in each clip embedding. The combination of the initial embeddings with their refinement and the global feature is then used in an ensemble for applications on new downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.645, + 0.468, + 0.9 + ], + "angle": 0, + "content": "To train SCALE, we introduce two novel pseudo-tasks, which aim to improve the discriminability of the embeddings of each video clip as well as obtain a global representation of the video. One task, which we call Masked Clip Modeling (MCM), is the reconstruction of a video clip embedding as in masked autoencoders [54]. Masked embeddings are combined with positional encodings so that the model can (spatio-temporally) relate the missing input embeddings to the other available embeddings (similarly to BERT in Natural Language Processing [12]). A second task is to train the model to output a global feature token (which we refer to as CLS, just for consistency with previous works [7, 12, 14, 55, 9]) for a set of clips via contrastive learning, so that the global feature is invariant to the chosen set of clips from the same video, but can discriminate summary features of clips from other videos. Both tasks are trained via contrastive losses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.241 + ], + "angle": 0, + "content": "SCALE consistently improves upon pretrained features, with more evident gains considering the computational cost of other methods. For instance, SCALE outperforms VideoMAE [54] by \\(0.3\\%\\) (1600 epochs checkpoint – fine-tuning) and \\(\\rho\\) BYOL [18] by \\(0.6\\%\\) (800 epochs checkpoint – linear probe) on Kinetics400 with 64 V100 GPUs in about 5 mins. In contrast, VideoMAE takes 27.7 hours to gain \\(0.5\\%\\) (800 to 1600 epochs checkpoint), and \\(\\rho\\) BYOL needs 48 hours to gain \\(0.4\\%\\) (200 to 400 epochs checkpoint) with the same resources." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.244, + 0.892, + 0.274 + ], + "angle": 0, + "content": "To summarize our contributions, we propose SCALE, a novel and highly scalable video representation method that" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.286, + 0.892, + 0.332 + ], + "angle": 0, + "content": "- is trained via novel pseudo-tasks on sets of video clips (in contrast to existing methods that work only with pairs of clips at a time [44, 18, 45]);" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.345, + 0.892, + 0.435 + ], + "angle": 0, + "content": "- results in video feature representations with a significant performance improvement in \\( k \\)-NN (retrieval), linear, and nonlinear probing across a wide range of datasets for action classification (UCF [48], HMDB [34], SSv2 [22], Kinetics400 [33]) and long-form video understanding (LVU [63]);" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.449, + 0.892, + 0.525 + ], + "angle": 0, + "content": "- achieves consistent transfer learning performance improvement across diverse SotA pre-trained backbones (architectures, scale, and pre-training tasks). For example, our nonlinear probed model even outperforms fully fine-tuned SVT [44] on HMDB [34]." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.286, + 0.892, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.642, + 0.557 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Our approach relates to many prior works on (self-supervised) representation learning. In particular, SCALE relates to SSL approaches on videos, methods that rely on multiple views of the data, and predictive methods, where part of the data is predicted from another part." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Self-Supervised Video Representation Learning. Early SSL approaches on video were based on pseudo tasks, e.g., the recognition of transformations of video frame sequences [38, 6, 62]. More recently, popular methods developed on images have been successfully translated to video, e.g., contrastive methods [18, 23, 26], clustering-based methods [44, 7], or predictive approaches [37, 41]. Often, these approaches are tailored to video data by including additional learning signals, e.g., by combining contrastive methods with temporal constraints [11] and pretext tasks [32], or by incorporating audio [39], or optical flow [24]. These approaches typically learn representations with limited temporal extent, which can serve as backbones for our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Learning from Multiple Views. Several methods have exploited multiple views (e.g., generated through space-time cropping) to represent and learn from videos. For example, many SSL approaches rely on multiple views of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5665" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "data, e.g., in contrastive formulations [18, 44, 42], or predictive learning [45], where invariance to views is the goal. Other approaches aim to learn from the relation between two views, e.g., by predicting overlap [69], the relative distance [49], or cross-view feature prediction [68, 52] and reconstruction [40]. Besides exploiting multiple views for SSL, some works also propose general multi-view video models, e.g., by capturing and fusing features at different spatio-temporal resolutions [67], by aggregating information over longer time-spans [64, 47, 65, 57], or by selecting important frames [21]. These approaches, however, are learned end-to-end. In contrast, we propose a more scalable approach by learning a global video representation of pretrained features extracted over multiple crops using self-supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.473, + 0.566 + ], + "angle": 0, + "content": "Predictive Learning. One of our proposed SSL objectives is a prediction of space-time crop features given other video crop features. This approach is similar to masked token prediction as in BERT [12] and relates to several other methods in the literature. Masked input reconstruction methods have recently become popular on images [25] and successfully translated to video [54, 16]. Other approaches formulate masked prediction tasks in the learned feature space [71, 52, 13] or some fixed latent space [51]. Another line of work considers directional predictions (e.g., into the future) often formulated via contrastive predictive coding [41] applied to video [37, 36, 50, 66, 20]. These masked prediction tasks are typically formulated on a fixed grid (e.g., at the level of patches or frames). In contrast, our formulation is continuous, predicting features of arbitrarily sampled space-time crops." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.592, + 0.441, + 0.612 + ], + "angle": 0, + "content": "3. Scalable Video Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.47, + 0.669 + ], + "angle": 0, + "content": "To describe SCALE, we first define some basic notation and functions, including a general contrastive loss notation that we use for all training losses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.692, + 0.182, + 0.707 + ], + "angle": 0, + "content": "3.1. Notation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We use lower-case letters (e.g., \\( z \\)) to denote generic vectors and capital letters (e.g., \\( Z \\)) to denote their sets. The expression \\( a \\circ b \\) denotes the concatenation of \\( a \\) and \\( b \\). We also avoid indicating the parameters of networks (usually denoted by \\( \\theta \\)) if their presence and role are clear from the context. Throughout the description of the method we refer to the training of neural networks and therefore at each iteration of the training we sample a minibatch of videos. All the equations in the next sections are written for a single video in the minibatch. Although we do not explicitly indicate it, all the contrastive losses also use the other videos in the minibatch as negatives." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.667, + 0.106 + ], + "angle": 0, + "content": "3.2. Contrastive Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.296 + ], + "angle": 0, + "content": "InfoNCE is a powerful method for representation learning [41] that can be used to maximize the mutual information between two variables. Because we use this loss between different variables throughout our method, we introduce here a unified notation. Let the paired sets \\( A \\) and \\( B \\) have \\( N \\) elements each, \\( A = \\{a^i\\}_{i=1}^N \\) and \\( B = \\{b^i\\}_{i=1}^N \\), where \\( a^i \\) are vectors of dimension \\( d_A \\) and \\( b^i \\) are vectors of dimension \\( d_B \\). We also introduce two Multi Layer Perceptrons (MLP), parameterized with \\( \\theta_A \\) and \\( \\theta_B \\), to project these vectors onto a common space of dimension \\( d \\). After feeding the elements \\( a^i \\) and \\( b^i \\) to the MLPs and normalizing them, we obtain" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.304, + 0.893, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\tilde {a} ^ {i} = \\frac {\\operatorname {M L P} _ {\\theta_ {A}} \\left(a ^ {i}\\right)}{\\| \\operatorname {M L P} _ {\\theta_ {A}} \\left(a ^ {i}\\right) \\|} \\quad \\text {a n d} \\quad \\tilde {b} ^ {i} = \\frac {\\operatorname {M L P} _ {\\theta_ {B}} \\left(b ^ {i}\\right)}{\\| \\operatorname {M L P} _ {\\theta_ {B}} \\left(b ^ {i}\\right) \\|}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.895, + 0.394 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\| \\) denotes the \\( L_{2} \\) norm. We define the per-element loss based on the relative similarity of \\( \\tilde{a}^i \\) and \\( \\tilde{b}^i \\), and by using a temperature \\( \\tau \\)" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.402, + 0.892, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\ell} ^ {i} (A, B, \\theta_ {A}, \\theta_ {B}) = - \\log \\frac {\\exp \\left(\\frac {\\tilde {a} ^ {i} \\cdot \\tilde {b} ^ {i}}{\\tau}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\tilde {a} ^ {i} \\cdot \\tilde {b} ^ {j}}{\\tau}\\right)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.481, + 0.818, + 0.497 + ], + "angle": 0, + "content": "We then make the loss symmetric [43] by using" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.507, + 0.892, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\ell^ {i} (A, B, \\theta_ {A}, \\theta_ {B}) = \\tilde {\\ell} ^ {i} (A, B) + \\tilde {\\ell} ^ {i} (B, A). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.536, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Finally, we define the contrastive loss \\(\\mathcal{L}_{\\mathrm{cntr}}(A,B,\\theta_A,\\theta_B)\\) as the mean of \\(\\ell^i\\)" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.577, + 0.892, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c n t r}} (A, B, \\theta_ {A}, \\theta_ {B}) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\ell^ {i} (A, B, \\theta_ {A}, \\theta_ {B}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.894, + 0.673 + ], + "angle": 0, + "content": "As mentioned earlier on, for simplicity, in the rest of the paper we will not indicate the parameters of the MLPs, and simply write \\(\\mathcal{L}_{\\mathrm{cntr}}(A,B)\\) or \\(\\ell^i (A,B)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.669, + 0.698 + ], + "angle": 0, + "content": "3.3. Training SCALE" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.856 + ], + "angle": 0, + "content": "In our method, we integrate 4 principles to drastically reduce the computational complexity of learning a video representation: Input sparsity, output sparsity, dimensionality reduction, and use of a pre-trained backbone. Moreover, we introduce two pseudo-tasks to train the model. One task is based on the (contrastive) reconstruction of a masked video clip given some context video clips. The second task is to build a global representation that is (contrastively) invariant to the set of input sampled video clips. The overall training scheme of SCALE is illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Input sparsity. As a first step, rather than processing a whole video, we collect a sparse set of short video clips from the same video. Given a video \\(V \\in \\mathbb{R}^{H \\times W \\times T \\times 3}\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5666" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.871, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.896, + 0.427 + ], + "angle": 0, + "content": "Figure 1. Video Representation Learning with SCALE. For each video, SCALE extracts two sets of video clips \\( V_{1}^{1}, \\ldots, V_{K}^{1} \\) and \\( V_{1}^{2}, \\ldots, V_{K}^{2} \\). Each video clip is processed separately through a frozen backbone \\( \\mathrm{E}_{\\text{基}} \\) and results in encoded video clips \\( C_{1}^{1}, \\ldots, C_{K}^{1} \\) and \\( C_{1}^{2}, \\ldots, C_{K}^{2} \\). Then, a random set of encodings in each set is masked and reconstructed at the output of the predictor network (a transformer) \\( (\\ell^{i}) \\). The predictor network is also fed a class token CLS. The corresponding output token encodes a summary \\( \\mathrm{CLS}^{m} \\) of the \\( m \\)-th set of video clips. The objective for these summary tokens is to be similar only when encoding video clips from the same video \\( (\\mathcal{L}_{\\mathrm{SET}}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.473, + 0.603 + ], + "angle": 0, + "content": "where \\(H\\), \\(W\\) and \\(T\\) are the height, width, and duration (in frames) of the video, we sample \\(2 \\times K\\) clips. We divide the clips into two sets randomly. Each clip in the first set \\(V_{i}^{1} \\in \\mathbb{R}^{H_{i}^{1} \\times W_{i}^{1} \\times T_{i}^{1} \\times 3}\\), with \\(i = 1, \\ldots, K\\), is obtained at the spatio-temporal location \\(X_{i}^{1}, Y_{i}^{1}, Q_{i}^{1}\\) with different data augmentations and dimensions \\(H_{i}^{1}\\), \\(W_{i}^{1}\\) and \\(T_{i}^{1}\\). Similarly, we denote the second set of clips \\(V_{i}^{2}\\), for \\(i = 1, \\ldots, K\\). We also normalize their coordinates relative to the video dimensions and embed them onto a feature vector \\(P_{i}^{j}\\) by feeding them to a learnable MLP. We denote these embeddings" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.612, + 0.47, + 0.661 + ], + "angle": 0, + "content": "\\[\nP _ {i} ^ {j} = \\operatorname {M L P} \\left(\\left[ \\frac {X _ {i} ^ {j}}{H}, \\frac {Y _ {i} ^ {j}}{W}, \\frac {Q _ {i} ^ {j}}{T}, \\frac {X _ {i} ^ {j} + H _ {i} ^ {j}}{H}, \\frac {Y _ {i} ^ {j} + W _ {i} ^ {j}}{W}, \\frac {Q _ {i} ^ {j} + T _ {i} ^ {j}}{T} \\right] ^ {\\top}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.673, + 0.303, + 0.688 + ], + "angle": 0, + "content": "where \\(j = 1,2\\) and \\(i = 1,\\dots ,K\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Dimensionality reduction and pre-trained backbone. To reduce the dimensionality of each video clip, we feed them independently to a frozen encoder \\(\\mathrm{E}_{\\ast}\\), to obtain the encodings \\(C_{i}^{j} = \\mathrm{E}_{\\ast}(V_{i}^{j})\\), where \\(j = 1,2\\) and \\(i = 1,\\dots ,K\\). Our framework is encoder-agnostic and thus can work with encoders obtained through different training schemes (supervised, contrastive, or autoencoder). In addition to reducing the dimensionality, we make the training even more scalable by using pre-trained and frozen encoders. Note, however, that if performance is the main goal, it is possible to also train a sparse backbone end to end with multiple clips. Thanks to token dropping, one can drop up to \\(95\\%\\) of the tokens [19] and still build a good representation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Output sparsity. As a self-supervised signal for our video" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.844 + ], + "angle": 0, + "content": "representation learning, we use a (contrastive) reconstruction loss. To reduce the computational cost, instead of predicting the features for the whole video [59, 19, 54, 16] (asymmetric decoding), we only reconstruct a sparse set of masked clips. Our reconstruction objective is based on the observation that video signals carry a lot of redundancy. Hence, we introduce a model, the predictor network, to predict masked video clip embeddings given the other video clip embeddings (the context). We follow the general approach of BERT [12] but implement the predictor network as a masked autoencoder, where the reconstruction is based on a contrastive loss. The loss is applied only to a sparse set \\( M^1 \\subset \\{1, \\dots, K\\} \\) of masked video clips. These clips are replaced by a learned MSK token. All embeddings \\( C_i^1 \\), including the masked ones, are added to their corresponding position encoding \\( P_i^1 \\) and are then fed to the predictor network. We also include an additional learnable CLS token as input for the predictor network, which will be used for tasks with multiple video clips. We denote the outputs of the predictor network as \\( \\hat{C}_i^1 \\) for the tokens corresponding to \\( C_i^1 \\), and as \\( \\mathrm{CLS}^1 \\) for the token corresponding to CLS. Similarly, we feed as inputs separately from the previous set all the video clips \\( C_i^2 \\) with their corresponding positional embeddings \\( P_i^2 \\), and the same CLS token, and obtain \\( \\hat{C}_i^2 \\) and \\( \\mathrm{CLS}^2 \\) respectively (see Figure 1 for a visual depiction of these processing steps)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Contrastive reconstruction. Modeling all the details of a masked clip, even in the latent space and even given the redundancy in videos, is a demanding task. Rather than" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5667" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "increasing the capacity of our model, since we are aiming for scalability, we keep our predictor network a (relatively) shallow network and use contrastive learning [41]. With contrastive learning, the predicted representation of the masked tokens should only be closer to the original un-masked clip representation (after an MLP projection) than from all other clips from the same video and the rest of the minibatch. Note that the rest of the clips in the same video act as hard negatives in contrastive learning [46]. Also, since we are using a frozen backbone, we can afford to use large minibatch sizes, which is known to be beneficial for contrastive learning [9]. We call this contrastive reconstruction loss the Masked Clip Modeling (MCM) loss" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.296, + 0.47, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M C M}} = \\sum_ {i \\in M ^ {1}} \\ell^ {i} \\left(\\hat {C} ^ {1}, C ^ {1}\\right) + \\sum_ {j \\in M ^ {2}} \\ell^ {j} \\left(\\hat {C} ^ {2}, C ^ {2}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.563 + ], + "angle": 0, + "content": "Multiple video clips loss. The predictor network outputs features for each video clip that are highly discriminative. This task is similar to that of a masked autoencoder [25] and gives you an enhanced per-clip representation. For many video tasks, we need a global representation for the whole video; for that, we introduce an additional pseudo-task that captures a more global representation of a set of video clips. Our task takes inspiration from contrastive learning methods used in SSL, which yield representations that perform well with linear probing [8]. The loss aims to make the \\(\\mathrm{CLS}^1\\) and \\(\\mathrm{CLS}^2\\) tokens returned from the predictor network more similar (recall that these two tokens are obtained from two separate groups of video clips extracted from the same video) than to other class tokens from other videos within the minibatch" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.571, + 0.47, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S E T}} = \\mathcal {L} _ {\\mathrm {c n t r}} \\left(\\mathrm {C L S} ^ {1}, \\mathrm {C L S} ^ {2}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.675 + ], + "angle": 0, + "content": "In addition to our contrastive loss (InfoNCE), one can use clustering [7] or regression [23] losses. We choose InfoNCE for simplicity and for better compatibility between the losses. As the overall loss, we use the sum of both loss terms (without any weights)" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.683, + 0.47, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M C M}} + \\mathcal {L} _ {\\mathrm {S E T}}, \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.709, + 0.21, + 0.725 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.84 + ], + "angle": 0, + "content": "We evaluate SCALE on several commonly used action classification datasets for video representation learning. As our performance metric, we primarily use linear probing and nonlinear probing [27]. For the smaller datasets, we also use \\( k \\)-NN classifiers (which are training-free) and demonstrate that the proposed method improves upon both unsupervised and supervised backbones." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.379, + 0.865 + ], + "angle": 0, + "content": "4.1. Experimental Setup and Protocols" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Computational Efficiency: In Table 1, we show estimates of the maximum batch sizes and the training throughput for" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.087, + 0.892, + 0.143 + ], + "angle": 0, + "content": "Table 1. Computational Resources. GPU VRAM usage (GB) and max training speed (using one 3090 GPU) of \\(\\mathrm{ViT_B}\\) with varying batch sizes and SSL tasks for videos. \\(\\mathrm{MoCo_{Sparse}^{V3}}\\) is akin to MSN [3]. OOM indicates Out Of Memory error." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.143, + 0.894, + 0.235 + ], + "angle": 0, + "content": "
Batch Size819572048Samples/s
MoCoV323.47OOMOOMOOM8.66
VMAE11.4222.80OOMOOM33.54
MoCoV3Sparse5.499.2723.62OOM28.83
SCALE1.211.231.5819.529224.65
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.894, + 0.338 + ], + "angle": 0, + "content": "different methods trained with the same computational and memory resources. As can be seen, SCALE is orders of magnitude more efficient than other SotA methods. Also during multi-crop inference, our method only results in an FLOP increase of approximately \\(0.001\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.894, + 0.46 + ], + "angle": 0, + "content": "Datasets: Following prior work [44, 18, 45] we use Kinetics-400 [33], UCF-101 [48] (split 1), HMDB-51 [34] (split 1), and Something-Something v2 (SSv2) [22] to train and evaluate our models. We also use the LVU benchmark [63] to showcase our long-form video understanding capabilities. Note that almost \\(35\\%\\) of LVU videos are not available to download from YouTube anymore; thus, our results are not directly comparable with prior methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Pretrained backbones: We use the pretrained checkpoints of \\(\\rho\\) BYOL [18], SVT [44], and three variants of Video-MAE [54] (base(B), large(L), and fine-tuned base(FT)). We choose \\(\\rho\\) BYOL for their excellent linear performance, SVT for the usage of ViT [14], and VMAE for showing 1) the applicability of our proposed method to MAE models, 2) the scalability of our method to larger models, and 3) possibility of using supervisedly fine-tuned models as our backbone. All the models are self-supervisedly pretrained on Kinetics-400, except the fine-tuned VMAE base that was also supervisedly finetuned on Kinetics-400. We also used a backbone pretrained and fine-tuned on SSv2 (VMAE\\(_{SSv2}\\)) for the SSv2 experiment to show the universality of SCALE with respect to the pretraining dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.884 + ], + "angle": 0, + "content": "Self-supervised Training: For training data, we extract 16 clips of 16 frames from each video per dataset and save their feature encodings to disk. We use PySlowFast's common data augmentations for that [15]. For evaluation, we follow the \\(5 \\times 3\\) scheme [17] (uniformly sampling 5 clips from a video along its temporal axis and then taking 3 spatial crops) and extract 15 clips from each video (except for SSv2, where we extract \\(2 \\times 3\\) clips [58]). As the architecture for the predictor network, we use an encoder-only Transformer [56] and a three-layer MLP (without batch normalization [30]) for the contrastive heads. Unless stated otherwise, we train our models for 500 epochs (for example, training with \\(\\mathrm{VMAE}^{\\mathrm{B}}\\) on SSV2 takes 137 minutes with one 3090 GPU) with a batch size of 512 and use all 16 clips." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation: Since our focus is on efficient and scalable" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5668" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.892, + 0.129 + ], + "angle": 0, + "content": "Table 2. Long-Form Video Understanding Results. Linear and nonlinear probing accuracies on LVU [63] classification tasks. SCALE shows significant performance improvement compared to baselines, indicating its ability to capture long-form video features. The first two rows are greyed out as direct comparison is not possible due to the unavailability of the full dataset for download." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.13, + 0.904, + 0.307 + ], + "angle": 0, + "content": "
RelationSpeakSceneDirectorGenreWriterYear
SlowFast+NL ViS4mer[17, 60]52.4035.8054.7044.9053.0036.3052.50
[31]57.1440.7967.4462.6154.7148.8044.75
SVTLinear64.7035.7762.3337.2754.3552.1228.46
SCALE linear73.52 (+8.82)40.65 (+4.88)68.83 (+6.50)46.36 (+9.09)57.94 (+3.59)56.38 (+4.26)39.23 (+10.77)
MLP67.6439.0266.2345.4556.9257.4436.15
Transformer70.5840.6568.8347.2757.1758.5136.92
SCALE ft76.47 (+5.89)42.27 (+1.62)74.02 (+5.19)49.09 (+1.82)58.97 (+1.80)62.76 (+4.25)39.23 (+2.31)
ρBYOLLinear52.9438.2153.2436.3651.7957.4433.84
SCALE linear67.64 (+14.70)43.08 (+4.87)66.23 (+12.99)44.54 (+8.18)53.33 (+1.54)60.63 (+3.19)40.00 (+6.16)
MLP62.3541.4662.4247.2752.5659.5740.00
Transformer65.0944.7166.2350.9053.0761.7043.84
SCALE ft67.64 (+2.55)45.52 (+0.81)71.42 (+5.19)51.81 (+0.91)55.72 (+2.65)65.95 (+4.25)46.92 (+3.08)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.473, + 0.649 + ], + "angle": 0, + "content": "video classification, we always freeze the backbones in our evaluation (as in our self-supervised pretraining) and either train a linear classifier [44, 18] or fine-tune the predictor network (the transformer) with an additional linear head. Therefore, when we refer to fine-tuning (ft), we only adapt the nonlinear head (e.g., predictor network) but not the backbone. We apply a grid search for the hyper-parameters of the heads covering learning rate, weight decay, batch size, and optimizer type. Similar to MAE [25], we found that applying a batch normalization layer [30] without affine transformations is beneficial for VideoMAE models. As the linear baseline, we consider the well-established ensembling approach, i.e., we average the softmax predictions of the 15 clips (6 for SSv2) to obtain the final prediction. For models that process multiple clips at once (like ours), we likewise apply a linear softmax head on the concatenation of the individual clip features and the [SET] token before averaging to obtain the final prediction. For the smaller datasets, we also use \\( k \\)-NN classification, where, similar to DINO [7], we always use \\( k = 20 \\) and work with 12 normalized representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.786 + ], + "angle": 0, + "content": "Nonlinear baselines: As SCALE is a nonlinear model, we consider an MLP on top of the frozen backbone as a nonlinear baseline. As a further baseline and to illustrate the effect of our self-supervised pre-training, we consider a Transformer trained on all the clip representations. This Transformer uses the exact same architecture as SCALE, and only differs in the initialization: in the case of SCALE we start from our proposed SSL pre-trained weights instead of random initialization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.172, + 0.815 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "LVU: One of the benefits of SCALE is that it can be used to process long videos, even though the backbones were trained on short videos only. To demonstrate the ability to capture long-form video features, we evaluated SCALE on LVU [63], a benchmark that involves seven classifica" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.328, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Table 3. SSv2 Results. Linear and nonlinear probing accuracies on SSv2 [22]. We see that both SCALE linear and SCALE ft outperform other methods and improve the classification accuracies by a large margin. We also see that SCALE ft, with its better initialization, always outperforms the Transformer. \\(\\mathrm{VMAE}_{\\mathrm{SSv2}}^{\\mathrm{B}}\\) was pretrained and fine-tuned on SSv2." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.411, + 0.892, + 0.505 + ], + "angle": 0, + "content": "
SVTρBYOLVMAEBVMAELVMAEftVMAEBSSv2
Linear20.3025.3018.3127.9428.9070.53
SCALE linear25.2627.1621.2430.1833.2570.63
MLP21.4326.4619.4227.9629.8370.52
Transformer29.2430.9924.2634.3935.6070.57
SCALE ft29.6831.8325.2536.3437.3870.69
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.64 + ], + "angle": 0, + "content": "tion (and two regression) tasks on minute-long videos. Past studies [31, 60] have established that increasing the input's time span enhances accuracy in this challenging dataset. As shown in Table 2, our experiments indicate that SCALE can improve the baseline model's performance by a considerable margin. Moreover, we found that fine-tuning SCALE can lead to further enhancements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.825 + ], + "angle": 0, + "content": "SSv2: Multiple classes in SSv2 share similar backgrounds and only differ in motion [28], suggesting that high performance on this dataset demonstrates that the model has captured strong motion-related contextual cues [44]. Results in Table 3 show that we outperform the state-of-the-art. On this dataset, we see a large performance gap between models that process single clips at a time (Linear and MLP) and the models that work with multiple clips (SCALE and Transformer). We can see SCALE linear is also outperforming the MLP, showing that SCALE is able to capture motion and long-form temporal features of the video. We even improve the supervised model trained on SSv2 \\(\\mathrm{(VMAE_{SSv2}^B)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "UCF-101 & HMDB-51: For these smaller datasets, besides linear and nonlinear probing, we also use \\( k \\)-NN probing (see Table 4 and Table 5). With SCALE \\( k \\)-NN, we see a consistent improvement over the baseline and find that pre-trained MAE-based models greatly benefit from our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5669" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.087, + 0.473, + 0.226 + ], + "angle": 0, + "content": "Table 4. UCF Results. Linear and nonlinear probing accuracies on UCF-101 [48]. SCALE fit outperforms all the other models and, in the case of \\(\\rho\\) BYOL, even gets performance close to a fully fine-tuned model. Also, in most cases, SCALE linear outperforms the fine-tuned Transformer and achieves state-of-the-art results in linear probing (previous SotA using RGB frames was 92.6 [45]). We further see a significant accuracy improvement in \\(k\\)-NN probing, especially for pre-trained MAE-based models. As a point of reference, the current best fully fine-tuned accuracy (which is not comparable with our setting) is \\(96.8\\%\\) [57]." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.226, + 0.472, + 0.368 + ], + "angle": 0, + "content": "
SVTρBYOLVMAEBVMAELVMAEft
k-NN87.2085.1935.0549.1496.82
SCALE k-NN89.0083.4765.6376.0297.38
Linear91.2789.5566.5384.5397.91
SCALE linear92.6591.4374.4686.7898.14
MLP91.1793.6071.9787.0498.04
Transformer92.2094.3468.2286.3098.04
SCALE ft92.9495.0076.0789.9298.46
FTreported93.795.496.1--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.381, + 0.472, + 0.505 + ], + "angle": 0, + "content": "Table 5. HMDB Results. Linear and nonlinear probing accuracies on HMDB-51 [34]. Despite the small size of the dataset, we see that \\(\\mathrm{SCALE}_{\\mathrm{ft}}\\) is outperforming all the other methods, and in the case of SVT, it even outperforms the fully fine-tuned model. We also see that \\(\\mathrm{SCALE}_{\\mathrm{linear}}\\) outperforms the Transformer in most cases with only a single linear layer (the best linear accuracy in the literature is \\(66.7\\%\\) [45]). Similar to UCF results, we see a considerable increase in the performance of \\(k\\)-NN classifiers for pre-trained MAE-based models." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.506, + 0.472, + 0.648 + ], + "angle": 0, + "content": "
SVTρBYOLVMAEBVMAELVMAEft
k-NN51.8349.6721.9629.2172.81
SCALE k-NN56.0151.5637.1851.3071.83
Linear63.0761.1745.2260.2676.33
SCALE linear66.3363.9252.1562.3578.36
MLP63.0064.7749.0162.6177.45
Transformer63.9866.1647.3261.5076.86
SCALE ft68.1066.7951.8964.8379.34
FTreported67.273.673.3--
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.839 + ], + "angle": 0, + "content": "training. This can be explained by the additional invariance properties introduced through the SET loss term in SCALE training. Across the board, we also see that in the case of linear probing, not only does SCALE linear outperform Linear, but it also outperforms Transformer, which leverages many more parameters. In the case of SVT, our SCALE linear also outperforms the best reported linear accuracy on UCF101 (92.7% vs. 92.6% [45]). Finally, \\(\\mathrm{SCALE}_{\\mathrm{ft}}\\) achieves better results than all the nonlinear baselines and even outperforms the fully fine-tuned SVT (68.1% vs. 67.2%)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Kinetics-400: We present our main results on Kinetics-400 [33] in Table 6. Our SCALE linear with SVT backbone beats the previous state of the art (71.8% vs. 71.5% [18]) and SCALE ft can even improve the accuracy of \\(\\mathrm{VMAE}_{\\mathrm{ft}}^{\\mathrm{B}}\\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.087, + 0.892, + 0.184 + ], + "angle": 0, + "content": "Table 6. Kinetics-400 Results. Linear and nonlinear probing accuracies on Kinetics-400 [33] without any extra data and using RGB frames only. While SCALE linear is on par with Linear, we observe clear improvements for nonlinear probing in the case of SCALE ft. Note that the best linear accuracy on this dataset (without any extra data) is 71.5 [18] and the best full fine-tuning accuracy is 86.7 [61]." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.185, + 0.892, + 0.279 + ], + "angle": 0, + "content": "
SVTρBYOLVMAEBVMAELVMAEftb
Linear71.7168.8243.5060.7381.52
SCALE linear71.7868.3843.9660.6681.44
MLP71.1969.4245.4861.6481.27
Transformer72.1869.2844.8562.1581.70
SCALE ft72.3869.6346.1562.6781.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Table 7. Kinetics-400 Low-shot Results. Linear and nonlinear probing accuracies on \\(10\\%\\) of Kinetics-400 [33]. SCALE is more robust to the size of the labeled dataset. \\(\\mathrm{SCALE}_{\\mathrm{ft}}\\) does not overfit like the other nonlinear probes (MLP and Transformer) and outperforms the baselines. \\(\\mathrm{VMAE}_{\\mathrm{ft}}^{\\mathrm{B}}\\) is greyed out since it was already fine-tuned on the whole labeled dataset and is only reported here for consistency with the other tables." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.387, + 0.892, + 0.481 + ], + "angle": 0, + "content": "
SVTρBYOLVMAEBVMAELVMAEft
Linear66.4356.4331.2548.4279.79
SCALE linear65.9657.7434.0349.2179.94
MLP65.4458.6830.4748.2779.37
Transformer64.9758.9529.8948.2779.47
SCALE ft67.0159.5233.9250.3680.47
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.89, + 0.521 + ], + "angle": 0, + "content": "which is a strong supervised model, from \\(81.5\\%\\) to \\(81.84\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Following the evaluation setup of self-supervised image representations [8, 7, 71], we also introduce low-shot K400 video classifications by sampling 10 percent of the videos (in a class-balanced way) and training the probes only on those. We still test on the whole evaluation set of K400. This low-shot setting is more aligned with the typical use-case of self-supervised models in which there is abundant unlabeled data for training via self-supervision and a small set of labeled data for fine-tuning. Results in Table 7 show that our method is particularly effective in this low-shot setting. While most other nonlinear probes overfit and perform worse than the linear probes, our SCALE ft does not overfit and clearly outperforms the baselines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.611, + 0.741 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.87 + ], + "angle": 0, + "content": "In this section, we start from a baseline setup consisting of a two-layer transformer with a hidden size of 256, \\(20\\%\\) chance of masking clips, trained with a batch size of 512 for 200 epochs, and using two sets of 8 views for representation learning. Using SCALE ft, we explore different loss functions, masking ratios, number of layers, and finally, the number of views during training and testing. All experiments are performed on UCF and HMDB." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Loss Function: As explained in the method section, we have two loss terms, and each of them can be enabled or" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5670" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.472, + 0.144 + ], + "angle": 0, + "content": "Table 8. Loss Function. SCALE fit accuracy with different loss function combinations (the masking ratio here is \\(20\\%\\)). We can see that having MCM is always beneficial, and the SET loss is almost always helpful. We use both loss terms for our final model." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.153, + 0.47, + 0.23 + ], + "angle": 0, + "content": "
SETMCMUCF-101HMDB-51
SVTρBYOLSVTρBYOL
✓✗91.8092.2064.5063.59
✗✓92.0193.8162.8164.05
✓✓93.2092.9964.5765.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.241, + 0.47, + 0.297 + ], + "angle": 0, + "content": "Table 9. Masking Ratio. SCALEft accuracy with different masking ratios. We observe best results around 25% similar to NLP models [12] (15%), and different from low-level video models like VideoMAE [54] (90%)." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.296, + 0.47, + 0.403 + ], + "angle": 0, + "content": "
Masking RatioUCF-101HMDB-51
SVTρBYOLSVTρBYOL
0.1593.1893.2564.3764.83
0.2593.2093.8165.4965.62
0.3593.1593.0664.1865.22
0.4592.9693.0263.3964.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.414, + 0.47, + 0.456 + ], + "angle": 0, + "content": "Table 10. Transformer Capacity. SCALEft accuracy with different model capacities. Having more than one transformer layer and not too few hidden channels is necessary for the best performance." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.466, + 0.472, + 0.649 + ], + "angle": 0, + "content": "
Hidden \nDimNum \nLayersUCF-101HMDB-51
SVTρBYOLSVTρBYOL
641-92.62-63.16
1281-92.83-63.68
2561-92.86-64.39
128292.7893.5263.2665.55
256293.2093.8165.4965.62
512292.5793.6665.6864.77
128392.3393.2564.8365.55
256392.7593.5265.4965.16
512392.8692.9365.4964.84
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.468, + 0.703 + ], + "angle": 0, + "content": "disabled for the pretraining. In Table 8 we show that having both loss terms is better than the individual loss terms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.87 + ], + "angle": 0, + "content": "Masking Ratio: Masking ratio is an important hyperparameter and depends on the data modality, for example, BERT [12] uses \\(15\\%\\), MSN [3] uses \\(30\\%\\) (for ViT-Base), MAE [25] uses \\(75\\%\\), and VideoMAE [54] uses 90 to \\(95\\%\\) masking. Since our clip representations are somewhat abstract representations of the video, we expect the optimal masking ratio to be close to NLP models rather than video MAEs. We have observed a steady decrease in the pretraining task's performance with higher masking ratios, so we only tested low masking ratios in Table 9 and found out that \\(25\\%\\) is the optimal masking ratio." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Transformer Capacity: We also explore the number of" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.087, + 0.892, + 0.142 + ], + "angle": 0, + "content": "Table 11. Number of Views. SCALE ft accuracy with different numbers of clips and batch sizes. More views lead to consistent improvement, and large batch sizes are not necessary because of the hard negative samples." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.143, + 0.894, + 0.325 + ], + "angle": 0, + "content": "
Num ViewsBatch SizeUCF-101HMDB-51
SVTρBYOLSVTρBYOL
4 × 225692.6592.8364.3564.24
6 × 225692.7093.0764.5764.37
8 × 225692.8093.4964.6464.63
4 × 251292.6793.4964.8564.50
6 × 251293.1893.6864.9065.35
8 × 251293.2093.8165.4965.62
4 × 2102492.7593.3664.7765.15
6 × 2102492.9693.5764.9665.48
8 × 2102493.07OOM65.29OOM
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.435 + ], + "angle": 0, + "content": "transformer layers and their hidden size in Table 10. We can see that having more than one transformer layer is necessary for good results and too few hidden channels can hurt performance. However, there is a trade-off, and deeper transformers can lead to worse performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.529 + ], + "angle": 0, + "content": "Number of Views: Finally, we studied the model performance as we changed the number of views and batch size fed to the model. As can be seen in Table 11, having more views has a large and consistent impact on the performance, and since we have hard negatives for contrastive loss within the video, we are not too reliant on large batch sizes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.554, + 0.62, + 0.57 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduced SCALE, a framework for video representation learning that aggregates the information from multiple clips from the same video. We combine contrastive learning and masked modeling with intuitions from predictive coding to obtain improved global and local representations of clips starting from frozen backbones. We evaluated these features using a wide array of backbones on different action classification and video understanding datasets and achieved strong or state-of-the-art results. The computational efficiency of our method is extremely useful for videos and opens the possibility to a wider group of researchers (with limited computational resources) to work on video representation learning than previously possible. We also believe that working with a set of clips is an interesting direction for representation learning. Finally, as a surprising and maybe alarming observation, even contrastive representations that were trained to be invariant to data augmentations and spatio-temporal crops can be used for contrastive masked modeling. This might be due to benign memorization [2], and understanding why this phenomenon happens might lead to a better understanding of contrastive learning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5671" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.16 + ], + "angle": 0, + "content": "This work was supported by grant 200020_188690 of the Swiss National Science Foundation (SNSF) and an Adobe award." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.173, + 0.175, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.199, + 0.47, + 0.254 + ], + "angle": 0, + "content": "[1] Hassan Akbari, Linagzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. ArXiv, abs/2104.11178, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[2] Sotiris Anagnostidis, Gregor Bachmann, Lorenzo Noci, and Thomas Hofmann. The curious case of benign memorization. ArXiv, abs/2210.14019, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[4] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zach Beaver, Jana von Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, Vivek Natarajan, and Mohammad Norouzi. Big self-supervised models advance medical image classification. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3458-3468, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[5] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. ArXiv, abs/2106.08254, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.563 + ], + "angle": 0, + "content": "[6] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9922-9931, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.633 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.47, + 0.689 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.691, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[11] Ishan Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TCLR: Temporal contrastive learning for video representation. Computer Vision and Image Understanding, 219:103406, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.199, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. ArXiv, abs/2207.07116, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[15] Haoqi Fan, Yanghao Li, Bo Xiong, Wan-Yen Lo, and Christoph Feichtenhofer. Pyslowfast. https://github.com/facebookresearch/slowfast, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6201-6210, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[18] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[19] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimae: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[20] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13505-13515, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[21] Shreyank N Gowda, Marcus Rohrbach, and Laura Sevilla-Lara. Smart frame selection for action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1451-1459, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[22] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in Neural Information Processing Systems, 33:5679-5690, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5672" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[27] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. ArXiv, abs/1808.06670, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[28] Kaiqin Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7929, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[29] Deng Huang, Wenhao Wu, Weiwen Hu, Xu Liu, Dongliang He, Zhihua Wu, Xiangmiao Wu, Mingkui Tan, and Errui Ding. Ascnet: Self-supervised video representation learning with appearance-speed consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8096-8105, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[31] Md. Mohaiminul Islam and Gedas Bertasius. Long movie clip classification with state-space video models. ArXiv, abs/2204.01692, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[32] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9970-9980, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[33] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[34] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[35] Xingbin Liu, Jinghao Zhou, Tao Kong, Xianming Lin, and Rongrong Ji. Exploring target representations for masked autoencoders. arXiv preprint arXiv:2209.03917, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[36] Yue Liu, Junqi Ma, Yufei Xie, Xuefeng Yang, Xingzhen Tao, Lin Peng, and Wei Gao. Contrastive predictive coding with transformer for video representation learning. Neurocomputing, 482:154-162, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[37] Guillaume Lorre, Jaxonary Rabarisoa, Astrid Orcesi, Samia Ainouz, and Stephane Canu. Temporal contrastive pretraining for video action recognition. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 662-670, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "verification. In European conference on computer vision, pages 527-544. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.124, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[39] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.195, + 0.892, + 0.25 + ], + "angle": 0, + "content": "[40] Charlie Nash, João Carreira, Jacob Walker, Iain Barr, Andrew Jaegle, Mateusz Malinowski, and Peter Battaglia. Transframer: Arbitrary frame prediction with generative models. arXiv preprint arXiv:2203.09494, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.253, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[41] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.297, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[42] Mandela Patrick, Yuki M. Asano, Bernie Huang, Ishan Misra, Florian Metze, João F. Henriques, and Andrea Vedaldi. Space-time crop & attend: Improving cross-modal video representation learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10540-10552, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.464 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[44] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2874-2884, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[45] Adria Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Pătrăucean, Florent Altché, Michal Valko, et al. Broaden your views for self-supervised video learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1255-1265, 2021. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.627, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[46] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. ArXiv, abs/2010.04592, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[47] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In European Conference on Computer Vision, pages 154-171. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[48] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[49] Chen Sun, Arsha Nagrani, Yonglong Tian, and Cordelia Schmid. Composable augmentation encoding for video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8834-8844, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12607-12617, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5673" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[51] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[52] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. arXiv preprint arXiv:2206.01204, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[53] Nenad Tomasev, Ioana Bica, Brian McWilliams, Lars Holger Buesing, Razvan Pascanu, Charles Blundell, and Jovana Mitrovic. Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels on imageNet? In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[54] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. 1, 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[55] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[57] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14010-14020, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:2740-2755, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[59] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14713-14723, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[60] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[61] Chen Wei, Haoqi Fan, Saining Xie, Chaoxia Wu, Alan Loddon Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14648-14658, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[62] Donglai Wei, Joseph J Lim, Andrew Zisserman, and William T Freeman. Learning and using the arrow of time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8052-8060, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[63] Chaoxia Wu and Philipp Krahenbuhl. Towards long-form video understanding. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1884-1894, 2021. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[64] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[65] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.893, + 0.302 + ], + "angle": 0, + "content": "[66] Yu Wu, Linchao Zhu, Xiaohan Wang, Yi Yang, and Fei Wu. Learning to anticipate egocentric actions by imagination. IEEE Transactions on Image Processing, 30:1143-1152, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.893, + 0.373 + ], + "angle": 0, + "content": "[67] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3333-3343, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.893, + 0.457 + ], + "angle": 0, + "content": "[68] Liangzhe Yuan, Rui Qian, Yin Cui, Boqing Gong, Florian Schroff, Ming-Hsuan Yang, Hartwig Adam, and Ting Liu. Contextualized spatio-temporal contrastive learning with self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13977-13986, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.893, + 0.541 + ], + "angle": 0, + "content": "[69] Yujia Zhang, Lai-Man Po, Xuyuan Xu, Mengyang Liu, Yexin Wang, Weifeng Ou, Yuzhi Zhao, and Wing-Yin Yu. Contrastive spatio-temporal pretext learning for self-supervised video representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3380-3389, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.543, + 0.893, + 0.585 + ], + "angle": 0, + "content": "[70] Zhang Zhang and Dacheng Tao. Slow feature analysis for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 34(3):436-450, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.893, + 0.64 + ], + "angle": 0, + "content": "[71] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.893, + 0.697 + ], + "angle": 0, + "content": "[72] Pan Zhou, Yichen Zhou, Chenyang Si, Weihao Yu, Teck Khim Ng, and Shuicheng Yan. Mugs: A multi-granular self-supervised learning framework. arXiv preprint arXiv:2203.14415, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5674" + } + ] +] \ No newline at end of file diff --git a/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/6732d370-6205-414c-986c-288f16a786f3_origin.pdf b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/6732d370-6205-414c-986c-288f16a786f3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8626e405558da1df9bd997de376856a5f157357c --- /dev/null +++ b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/6732d370-6205-414c-986c-288f16a786f3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26b2b210f70b9899eac377519e39da4181081295e74888bf2c74427e9ed9471f +size 1143690 diff --git a/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/full.md b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..924328bd34624a6920fc9f6b5ee9b91cdf89f238 --- /dev/null +++ b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/full.md @@ -0,0 +1,338 @@ +# Spatio-Temporal Crop Aggregation for Video Representation Learning + +Sepehr Sameni + +University of Bern + +sepehr.sameni@unibe.ch + +Simon Jenni + +Adobe Research + +jenni@adobe.com + +Paolo Favaro + +University of Bern + +paolo.favaro@unibe.ch + +# Abstract + +We propose Spatio-temporal Crop Aggregation for video representation Learning (SCALE), a novel method that enjoys high scalability at both training and inference time. Our model builds long-range video features by learning from sets of video clip-level features extracted with a pretrained backbone. To train the model, we propose a self-supervised objective consisting of masked clip feature predictions. We apply sparsity to both the input, by extracting a random set of video clips, and to the loss function, by only reconstructing the sparse inputs. Moreover, we use dimensionality reduction by working in the latent space of a pre-trained backbone applied to single video clips. These techniques make our method not only extremely efficient to train but also highly effective in transfer learning. We demonstrate that our video representation yields state-of-the-art performance with linear, nonlinear, and $k$ -NN probing on common action classification and video understanding datasets. + +# 1. Introduction + +Videos provide rich and detailed information about objects and their activities. Their analysis is, however, made challenging not only by the difference in the information provided across space and time but also by the high dimensionality of the data [45, 29, 70]. While computational resources are expected to scale over time, so is the demand for higher data resolution (both in space and time) and also the need for processing data with even more dimensions, such as videos of volumetric data [10]. Therefore, it is of paramount importance to explore methods that drastically reduce the computational requirements to process videos. Moreover, the annotation of videos is an extremely costly and time-consuming burden that makes the use of models pre-trained in a self-supervised manner essential [44]. + +Self-supervised learning (SSL) is a very popular technique to reduce the need for annotation because it can build useful representations from unlabeled data through an artificial goal, also called pseudo- or pretext-task. These repre + +sentations can either be evaluated through K-nearest neighbor or linear probing [71, 72] or through fine-tuning (i.e., as the initialization parameters of the trained model) [25, 54, 35] on a downstream task, where only a small labeled dataset is available. More remarkably, SSL pre-trained models can outperform models that were pre-trained on an annotated dataset [53, 4]. + +SSL methods for video representation learning present fundamental scalability challenges [54, 45, 29, 65]. A first major challenge is that training models from scratch for any new pseudo-task is not sustainable. A more viable setting is one where the processing of videos is broken down into multiple steps. First, one pre-processes video clips (spatiotemporal crops) only once via some pre-trained general-purpose model (e.g., trained via self-supervised learning [18, 23, 26, 44, 7, 37, 41]). In general, the idea is that this (compressed) local representation is stored and used later for other training purposes or retrieval. Then, one can train more efficiently other super-features that aggregate information from these local representations. Within this setting, we address two questions: 1) Can one further improve the performance of pre-computed video features by training a model on top of them? 2) Can such training be made computationally scalable? In this paper, we show that both questions have a positive answer. The answer exploits four fronts: + +1. Input Sparsity: Sparsity in the input to the model [54, 25, 3, 19, 1] drastically reduces computational load and memory requirements, leveraging information redundancy in images and videos [16]. Inspired by prior work, we extract a sparse set of clips from a video (these are spatio-temporal video crops and not image patches or video tubelets); each clip is fed separately to a neural network so that we obtain a set of video clip representations. + +2. Output Sparsity: Another way to reduce the computational cost is to use a sparse reconstruction output instead of a dense one [71, 51, 59, 5]. This is not just a reduction of the number of terms in the loss function, but also in the number of actual outputs of + +the model, and thus also a reduction of the computations needed to obtain them. This is in contrast to MAE-based SSL methods for vision, where the proposed pseudo-tasks are based on the reconstruction of the whole input [25, 54, 19, 16], and even if the loss uses a subset of the tokens (the masked ones) for the loss calculation, the remaining tokens are still part of the decoder's output and computational graph. + +3. Dimensionality Reduction: Inspired by prior work [71, 41, 13], instead of directly processing the raw input data, we work in the latent space. This allows us to further reduce the dimensionality of both input and output data. +4. Use of a Pre-trained Backbone: To reduce training time and further speed up the processing per iteration, we exploit SSL pre-trained models. These pre-trained backbones already learn very strong short-term spatiotemporal features, and our approach is a way to extract longer-term features (see LVU results 2) by building a video representation on top of a set of pre-trained features (one for each video clip). + +To integrate all these components, we propose a novel SSL method that we call Spatio-temporal Crop Aggregation for video representation LEarning (SCALE). Given an input video, SCALE extracts a random set of clips and produces an embedding for each clip through a pre-trained backbone, which is kept frozen. These initial embeddings are then augmented in two ways: 1) each one is refined into a more discriminative feature, and 2) the set of all embeddings is summarized in a global feature. These global features can learn long-term correlations in the whole video by aggregating the short-term information in each clip embedding. The combination of the initial embeddings with their refinement and the global feature is then used in an ensemble for applications on new downstream tasks. + +To train SCALE, we introduce two novel pseudo-tasks, which aim to improve the discriminability of the embeddings of each video clip as well as obtain a global representation of the video. One task, which we call Masked Clip Modeling (MCM), is the reconstruction of a video clip embedding as in masked autoencoders [54]. Masked embeddings are combined with positional encodings so that the model can (spatio-temporally) relate the missing input embeddings to the other available embeddings (similarly to BERT in Natural Language Processing [12]). A second task is to train the model to output a global feature token (which we refer to as CLS, just for consistency with previous works [7, 12, 14, 55, 9]) for a set of clips via contrastive learning, so that the global feature is invariant to the chosen set of clips from the same video, but can discriminate summary features of clips from other videos. Both tasks are trained via contrastive losses. + +SCALE consistently improves upon pretrained features, with more evident gains considering the computational cost of other methods. For instance, SCALE outperforms VideoMAE [54] by $0.3\%$ (1600 epochs checkpoint – fine-tuning) and $\rho$ BYOL [18] by $0.6\%$ (800 epochs checkpoint – linear probe) on Kinetics400 with 64 V100 GPUs in about 5 mins. In contrast, VideoMAE takes 27.7 hours to gain $0.5\%$ (800 to 1600 epochs checkpoint), and $\rho$ BYOL needs 48 hours to gain $0.4\%$ (200 to 400 epochs checkpoint) with the same resources. + +To summarize our contributions, we propose SCALE, a novel and highly scalable video representation method that + +- is trained via novel pseudo-tasks on sets of video clips (in contrast to existing methods that work only with pairs of clips at a time [44, 18, 45]); +- results in video feature representations with a significant performance improvement in $k$ -NN (retrieval), linear, and nonlinear probing across a wide range of datasets for action classification (UCF [48], HMDB [34], SSv2 [22], Kinetics400 [33]) and long-form video understanding (LVU [63]); +- achieves consistent transfer learning performance improvement across diverse SotA pre-trained backbones (architectures, scale, and pre-training tasks). For example, our nonlinear probed model even outperforms fully fine-tuned SVT [44] on HMDB [34]. + +# 2. Related Work + +Our approach relates to many prior works on (self-supervised) representation learning. In particular, SCALE relates to SSL approaches on videos, methods that rely on multiple views of the data, and predictive methods, where part of the data is predicted from another part. + +Self-Supervised Video Representation Learning. Early SSL approaches on video were based on pseudo tasks, e.g., the recognition of transformations of video frame sequences [38, 6, 62]. More recently, popular methods developed on images have been successfully translated to video, e.g., contrastive methods [18, 23, 26], clustering-based methods [44, 7], or predictive approaches [37, 41]. Often, these approaches are tailored to video data by including additional learning signals, e.g., by combining contrastive methods with temporal constraints [11] and pretext tasks [32], or by incorporating audio [39], or optical flow [24]. These approaches typically learn representations with limited temporal extent, which can serve as backbones for our approach. + +Learning from Multiple Views. Several methods have exploited multiple views (e.g., generated through space-time cropping) to represent and learn from videos. For example, many SSL approaches rely on multiple views of the + +data, e.g., in contrastive formulations [18, 44, 42], or predictive learning [45], where invariance to views is the goal. Other approaches aim to learn from the relation between two views, e.g., by predicting overlap [69], the relative distance [49], or cross-view feature prediction [68, 52] and reconstruction [40]. Besides exploiting multiple views for SSL, some works also propose general multi-view video models, e.g., by capturing and fusing features at different spatio-temporal resolutions [67], by aggregating information over longer time-spans [64, 47, 65, 57], or by selecting important frames [21]. These approaches, however, are learned end-to-end. In contrast, we propose a more scalable approach by learning a global video representation of pretrained features extracted over multiple crops using self-supervision. + +Predictive Learning. One of our proposed SSL objectives is a prediction of space-time crop features given other video crop features. This approach is similar to masked token prediction as in BERT [12] and relates to several other methods in the literature. Masked input reconstruction methods have recently become popular on images [25] and successfully translated to video [54, 16]. Other approaches formulate masked prediction tasks in the learned feature space [71, 52, 13] or some fixed latent space [51]. Another line of work considers directional predictions (e.g., into the future) often formulated via contrastive predictive coding [41] applied to video [37, 36, 50, 66, 20]. These masked prediction tasks are typically formulated on a fixed grid (e.g., at the level of patches or frames). In contrast, our formulation is continuous, predicting features of arbitrarily sampled space-time crops. + +# 3. Scalable Video Representation Learning + +To describe SCALE, we first define some basic notation and functions, including a general contrastive loss notation that we use for all training losses. + +# 3.1. Notation + +We use lower-case letters (e.g., $z$ ) to denote generic vectors and capital letters (e.g., $Z$ ) to denote their sets. The expression $a \circ b$ denotes the concatenation of $a$ and $b$ . We also avoid indicating the parameters of networks (usually denoted by $\theta$ ) if their presence and role are clear from the context. Throughout the description of the method we refer to the training of neural networks and therefore at each iteration of the training we sample a minibatch of videos. All the equations in the next sections are written for a single video in the minibatch. Although we do not explicitly indicate it, all the contrastive losses also use the other videos in the minibatch as negatives. + +# 3.2. Contrastive Loss + +InfoNCE is a powerful method for representation learning [41] that can be used to maximize the mutual information between two variables. Because we use this loss between different variables throughout our method, we introduce here a unified notation. Let the paired sets $A$ and $B$ have $N$ elements each, $A = \{a^i\}_{i=1}^N$ and $B = \{b^i\}_{i=1}^N$ , where $a^i$ are vectors of dimension $d_A$ and $b^i$ are vectors of dimension $d_B$ . We also introduce two Multi Layer Perceptrons (MLP), parameterized with $\theta_A$ and $\theta_B$ , to project these vectors onto a common space of dimension $d$ . After feeding the elements $a^i$ and $b^i$ to the MLPs and normalizing them, we obtain + +$$ +\tilde {a} ^ {i} = \frac {\operatorname {M L P} _ {\theta_ {A}} \left(a ^ {i}\right)}{\| \operatorname {M L P} _ {\theta_ {A}} \left(a ^ {i}\right) \|} \quad \text {a n d} \quad \tilde {b} ^ {i} = \frac {\operatorname {M L P} _ {\theta_ {B}} \left(b ^ {i}\right)}{\| \operatorname {M L P} _ {\theta_ {B}} \left(b ^ {i}\right) \|}, \tag {1} +$$ + +where $\| \cdot \|$ denotes the $L_{2}$ norm. We define the per-element loss based on the relative similarity of $\tilde{a}^i$ and $\tilde{b}^i$ , and by using a temperature $\tau$ + +$$ +\tilde {\ell} ^ {i} (A, B, \theta_ {A}, \theta_ {B}) = - \log \frac {\exp \left(\frac {\tilde {a} ^ {i} \cdot \tilde {b} ^ {i}}{\tau}\right)}{\sum_ {j = 1} ^ {N} \exp \left(\frac {\tilde {a} ^ {i} \cdot \tilde {b} ^ {j}}{\tau}\right)}. \tag {2} +$$ + +We then make the loss symmetric [43] by using + +$$ +\ell^ {i} (A, B, \theta_ {A}, \theta_ {B}) = \tilde {\ell} ^ {i} (A, B) + \tilde {\ell} ^ {i} (B, A). \tag {3} +$$ + +Finally, we define the contrastive loss $\mathcal{L}_{\mathrm{cntr}}(A,B,\theta_A,\theta_B)$ as the mean of $\ell^i$ + +$$ +\mathcal {L} _ {\mathrm {c n t r}} (A, B, \theta_ {A}, \theta_ {B}) = \frac {1}{N} \sum_ {i = 1} ^ {N} \ell^ {i} (A, B, \theta_ {A}, \theta_ {B}). \tag {4} +$$ + +As mentioned earlier on, for simplicity, in the rest of the paper we will not indicate the parameters of the MLPs, and simply write $\mathcal{L}_{\mathrm{cntr}}(A,B)$ or $\ell^i (A,B)$ . + +# 3.3. Training SCALE + +In our method, we integrate 4 principles to drastically reduce the computational complexity of learning a video representation: Input sparsity, output sparsity, dimensionality reduction, and use of a pre-trained backbone. Moreover, we introduce two pseudo-tasks to train the model. One task is based on the (contrastive) reconstruction of a masked video clip given some context video clips. The second task is to build a global representation that is (contrastively) invariant to the set of input sampled video clips. The overall training scheme of SCALE is illustrated in Figure 1. + +Input sparsity. As a first step, rather than processing a whole video, we collect a sparse set of short video clips from the same video. Given a video $V \in \mathbb{R}^{H \times W \times T \times 3}$ , + +![](images/ad3701cf1f5887dcfc3526852ced09c74df589e1fcf58ac08b692ec715456196.jpg) +Figure 1. Video Representation Learning with SCALE. For each video, SCALE extracts two sets of video clips $V_{1}^{1}, \ldots, V_{K}^{1}$ and $V_{1}^{2}, \ldots, V_{K}^{2}$ . Each video clip is processed separately through a frozen backbone $\mathrm{E}_{\text{基}}$ and results in encoded video clips $C_{1}^{1}, \ldots, C_{K}^{1}$ and $C_{1}^{2}, \ldots, C_{K}^{2}$ . Then, a random set of encodings in each set is masked and reconstructed at the output of the predictor network (a transformer) $(\ell^{i})$ . The predictor network is also fed a class token CLS. The corresponding output token encodes a summary $\mathrm{CLS}^{m}$ of the $m$ -th set of video clips. The objective for these summary tokens is to be similar only when encoding video clips from the same video $(\mathcal{L}_{\mathrm{SET}})$ . + +where $H$ , $W$ and $T$ are the height, width, and duration (in frames) of the video, we sample $2 \times K$ clips. We divide the clips into two sets randomly. Each clip in the first set $V_{i}^{1} \in \mathbb{R}^{H_{i}^{1} \times W_{i}^{1} \times T_{i}^{1} \times 3}$ , with $i = 1, \ldots, K$ , is obtained at the spatio-temporal location $X_{i}^{1}, Y_{i}^{1}, Q_{i}^{1}$ with different data augmentations and dimensions $H_{i}^{1}$ , $W_{i}^{1}$ and $T_{i}^{1}$ . Similarly, we denote the second set of clips $V_{i}^{2}$ , for $i = 1, \ldots, K$ . We also normalize their coordinates relative to the video dimensions and embed them onto a feature vector $P_{i}^{j}$ by feeding them to a learnable MLP. We denote these embeddings + +$$ +P _ {i} ^ {j} = \operatorname {M L P} \left(\left[ \frac {X _ {i} ^ {j}}{H}, \frac {Y _ {i} ^ {j}}{W}, \frac {Q _ {i} ^ {j}}{T}, \frac {X _ {i} ^ {j} + H _ {i} ^ {j}}{H}, \frac {Y _ {i} ^ {j} + W _ {i} ^ {j}}{W}, \frac {Q _ {i} ^ {j} + T _ {i} ^ {j}}{T} \right] ^ {\top}\right), \tag {5} +$$ + +where $j = 1,2$ and $i = 1,\dots ,K$ + +Dimensionality reduction and pre-trained backbone. To reduce the dimensionality of each video clip, we feed them independently to a frozen encoder $\mathrm{E}_{\ast}$ , to obtain the encodings $C_{i}^{j} = \mathrm{E}_{\ast}(V_{i}^{j})$ , where $j = 1,2$ and $i = 1,\dots ,K$ . Our framework is encoder-agnostic and thus can work with encoders obtained through different training schemes (supervised, contrastive, or autoencoder). In addition to reducing the dimensionality, we make the training even more scalable by using pre-trained and frozen encoders. Note, however, that if performance is the main goal, it is possible to also train a sparse backbone end to end with multiple clips. Thanks to token dropping, one can drop up to $95\%$ of the tokens [19] and still build a good representation. + +Output sparsity. As a self-supervised signal for our video + +representation learning, we use a (contrastive) reconstruction loss. To reduce the computational cost, instead of predicting the features for the whole video [59, 19, 54, 16] (asymmetric decoding), we only reconstruct a sparse set of masked clips. Our reconstruction objective is based on the observation that video signals carry a lot of redundancy. Hence, we introduce a model, the predictor network, to predict masked video clip embeddings given the other video clip embeddings (the context). We follow the general approach of BERT [12] but implement the predictor network as a masked autoencoder, where the reconstruction is based on a contrastive loss. The loss is applied only to a sparse set $M^1 \subset \{1, \dots, K\}$ of masked video clips. These clips are replaced by a learned MSK token. All embeddings $C_i^1$ , including the masked ones, are added to their corresponding position encoding $P_i^1$ and are then fed to the predictor network. We also include an additional learnable CLS token as input for the predictor network, which will be used for tasks with multiple video clips. We denote the outputs of the predictor network as $\hat{C}_i^1$ for the tokens corresponding to $C_i^1$ , and as $\mathrm{CLS}^1$ for the token corresponding to CLS. Similarly, we feed as inputs separately from the previous set all the video clips $C_i^2$ with their corresponding positional embeddings $P_i^2$ , and the same CLS token, and obtain $\hat{C}_i^2$ and $\mathrm{CLS}^2$ respectively (see Figure 1 for a visual depiction of these processing steps). + +Contrastive reconstruction. Modeling all the details of a masked clip, even in the latent space and even given the redundancy in videos, is a demanding task. Rather than + +increasing the capacity of our model, since we are aiming for scalability, we keep our predictor network a (relatively) shallow network and use contrastive learning [41]. With contrastive learning, the predicted representation of the masked tokens should only be closer to the original un-masked clip representation (after an MLP projection) than from all other clips from the same video and the rest of the minibatch. Note that the rest of the clips in the same video act as hard negatives in contrastive learning [46]. Also, since we are using a frozen backbone, we can afford to use large minibatch sizes, which is known to be beneficial for contrastive learning [9]. We call this contrastive reconstruction loss the Masked Clip Modeling (MCM) loss + +$$ +\mathcal {L} _ {\mathrm {M C M}} = \sum_ {i \in M ^ {1}} \ell^ {i} \left(\hat {C} ^ {1}, C ^ {1}\right) + \sum_ {j \in M ^ {2}} \ell^ {j} \left(\hat {C} ^ {2}, C ^ {2}\right). \tag {6} +$$ + +Multiple video clips loss. The predictor network outputs features for each video clip that are highly discriminative. This task is similar to that of a masked autoencoder [25] and gives you an enhanced per-clip representation. For many video tasks, we need a global representation for the whole video; for that, we introduce an additional pseudo-task that captures a more global representation of a set of video clips. Our task takes inspiration from contrastive learning methods used in SSL, which yield representations that perform well with linear probing [8]. The loss aims to make the $\mathrm{CLS}^1$ and $\mathrm{CLS}^2$ tokens returned from the predictor network more similar (recall that these two tokens are obtained from two separate groups of video clips extracted from the same video) than to other class tokens from other videos within the minibatch + +$$ +\mathcal {L} _ {\mathrm {S E T}} = \mathcal {L} _ {\mathrm {c n t r}} \left(\mathrm {C L S} ^ {1}, \mathrm {C L S} ^ {2}\right). \tag {7} +$$ + +In addition to our contrastive loss (InfoNCE), one can use clustering [7] or regression [23] losses. We choose InfoNCE for simplicity and for better compatibility between the losses. As the overall loss, we use the sum of both loss terms (without any weights) + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {M C M}} + \mathcal {L} _ {\mathrm {S E T}}, \tag {8} +$$ + +# 4. Experiments + +We evaluate SCALE on several commonly used action classification datasets for video representation learning. As our performance metric, we primarily use linear probing and nonlinear probing [27]. For the smaller datasets, we also use $k$ -NN classifiers (which are training-free) and demonstrate that the proposed method improves upon both unsupervised and supervised backbones. + +# 4.1. Experimental Setup and Protocols + +Computational Efficiency: In Table 1, we show estimates of the maximum batch sizes and the training throughput for + +Table 1. Computational Resources. GPU VRAM usage (GB) and max training speed (using one 3090 GPU) of $\mathrm{ViT_B}$ with varying batch sizes and SSL tasks for videos. $\mathrm{MoCo_{Sparse}^{V3}}$ is akin to MSN [3]. OOM indicates Out Of Memory error. + +
Batch Size819572048Samples/s
MoCoV323.47OOMOOMOOM8.66
VMAE11.4222.80OOMOOM33.54
MoCoV3Sparse5.499.2723.62OOM28.83
SCALE1.211.231.5819.529224.65
+ +different methods trained with the same computational and memory resources. As can be seen, SCALE is orders of magnitude more efficient than other SotA methods. Also during multi-crop inference, our method only results in an FLOP increase of approximately $0.001\%$ . + +Datasets: Following prior work [44, 18, 45] we use Kinetics-400 [33], UCF-101 [48] (split 1), HMDB-51 [34] (split 1), and Something-Something v2 (SSv2) [22] to train and evaluate our models. We also use the LVU benchmark [63] to showcase our long-form video understanding capabilities. Note that almost $35\%$ of LVU videos are not available to download from YouTube anymore; thus, our results are not directly comparable with prior methods. + +Pretrained backbones: We use the pretrained checkpoints of $\rho$ BYOL [18], SVT [44], and three variants of Video-MAE [54] (base(B), large(L), and fine-tuned base(FT)). We choose $\rho$ BYOL for their excellent linear performance, SVT for the usage of ViT [14], and VMAE for showing 1) the applicability of our proposed method to MAE models, 2) the scalability of our method to larger models, and 3) possibility of using supervisedly fine-tuned models as our backbone. All the models are self-supervisedly pretrained on Kinetics-400, except the fine-tuned VMAE base that was also supervisedly finetuned on Kinetics-400. We also used a backbone pretrained and fine-tuned on SSv2 (VMAE $_{SSv2}$ ) for the SSv2 experiment to show the universality of SCALE with respect to the pretraining dataset. + +Self-supervised Training: For training data, we extract 16 clips of 16 frames from each video per dataset and save their feature encodings to disk. We use PySlowFast's common data augmentations for that [15]. For evaluation, we follow the $5 \times 3$ scheme [17] (uniformly sampling 5 clips from a video along its temporal axis and then taking 3 spatial crops) and extract 15 clips from each video (except for SSv2, where we extract $2 \times 3$ clips [58]). As the architecture for the predictor network, we use an encoder-only Transformer [56] and a three-layer MLP (without batch normalization [30]) for the contrastive heads. Unless stated otherwise, we train our models for 500 epochs (for example, training with $\mathrm{VMAE}^{\mathrm{B}}$ on SSV2 takes 137 minutes with one 3090 GPU) with a batch size of 512 and use all 16 clips. + +Evaluation: Since our focus is on efficient and scalable + +Table 2. Long-Form Video Understanding Results. Linear and nonlinear probing accuracies on LVU [63] classification tasks. SCALE shows significant performance improvement compared to baselines, indicating its ability to capture long-form video features. The first two rows are greyed out as direct comparison is not possible due to the unavailability of the full dataset for download. + +
RelationSpeakSceneDirectorGenreWriterYear
SlowFast+NL ViS4mer[17, 60]52.4035.8054.7044.9053.0036.3052.50
[31]57.1440.7967.4462.6154.7148.8044.75
SVTLinear64.7035.7762.3337.2754.3552.1228.46
SCALE linear73.52 (+8.82)40.65 (+4.88)68.83 (+6.50)46.36 (+9.09)57.94 (+3.59)56.38 (+4.26)39.23 (+10.77)
MLP67.6439.0266.2345.4556.9257.4436.15
Transformer70.5840.6568.8347.2757.1758.5136.92
SCALE ft76.47 (+5.89)42.27 (+1.62)74.02 (+5.19)49.09 (+1.82)58.97 (+1.80)62.76 (+4.25)39.23 (+2.31)
ρBYOLLinear52.9438.2153.2436.3651.7957.4433.84
SCALE linear67.64 (+14.70)43.08 (+4.87)66.23 (+12.99)44.54 (+8.18)53.33 (+1.54)60.63 (+3.19)40.00 (+6.16)
MLP62.3541.4662.4247.2752.5659.5740.00
Transformer65.0944.7166.2350.9053.0761.7043.84
SCALE ft67.64 (+2.55)45.52 (+0.81)71.42 (+5.19)51.81 (+0.91)55.72 (+2.65)65.95 (+4.25)46.92 (+3.08)
+ +video classification, we always freeze the backbones in our evaluation (as in our self-supervised pretraining) and either train a linear classifier [44, 18] or fine-tune the predictor network (the transformer) with an additional linear head. Therefore, when we refer to fine-tuning (ft), we only adapt the nonlinear head (e.g., predictor network) but not the backbone. We apply a grid search for the hyper-parameters of the heads covering learning rate, weight decay, batch size, and optimizer type. Similar to MAE [25], we found that applying a batch normalization layer [30] without affine transformations is beneficial for VideoMAE models. As the linear baseline, we consider the well-established ensembling approach, i.e., we average the softmax predictions of the 15 clips (6 for SSv2) to obtain the final prediction. For models that process multiple clips at once (like ours), we likewise apply a linear softmax head on the concatenation of the individual clip features and the [SET] token before averaging to obtain the final prediction. For the smaller datasets, we also use $k$ -NN classification, where, similar to DINO [7], we always use $k = 20$ and work with 12 normalized representations. + +Nonlinear baselines: As SCALE is a nonlinear model, we consider an MLP on top of the frozen backbone as a nonlinear baseline. As a further baseline and to illustrate the effect of our self-supervised pre-training, we consider a Transformer trained on all the clip representations. This Transformer uses the exact same architecture as SCALE, and only differs in the initialization: in the case of SCALE we start from our proposed SSL pre-trained weights instead of random initialization. + +# 4.2. Results + +LVU: One of the benefits of SCALE is that it can be used to process long videos, even though the backbones were trained on short videos only. To demonstrate the ability to capture long-form video features, we evaluated SCALE on LVU [63], a benchmark that involves seven classifica + +Table 3. SSv2 Results. Linear and nonlinear probing accuracies on SSv2 [22]. We see that both SCALE linear and SCALE ft outperform other methods and improve the classification accuracies by a large margin. We also see that SCALE ft, with its better initialization, always outperforms the Transformer. $\mathrm{VMAE}_{\mathrm{SSv2}}^{\mathrm{B}}$ was pretrained and fine-tuned on SSv2. + +
SVTρBYOLVMAEBVMAELVMAEftVMAEBSSv2
Linear20.3025.3018.3127.9428.9070.53
SCALE linear25.2627.1621.2430.1833.2570.63
MLP21.4326.4619.4227.9629.8370.52
Transformer29.2430.9924.2634.3935.6070.57
SCALE ft29.6831.8325.2536.3437.3870.69
+ +tion (and two regression) tasks on minute-long videos. Past studies [31, 60] have established that increasing the input's time span enhances accuracy in this challenging dataset. As shown in Table 2, our experiments indicate that SCALE can improve the baseline model's performance by a considerable margin. Moreover, we found that fine-tuning SCALE can lead to further enhancements. + +SSv2: Multiple classes in SSv2 share similar backgrounds and only differ in motion [28], suggesting that high performance on this dataset demonstrates that the model has captured strong motion-related contextual cues [44]. Results in Table 3 show that we outperform the state-of-the-art. On this dataset, we see a large performance gap between models that process single clips at a time (Linear and MLP) and the models that work with multiple clips (SCALE and Transformer). We can see SCALE linear is also outperforming the MLP, showing that SCALE is able to capture motion and long-form temporal features of the video. We even improve the supervised model trained on SSv2 $\mathrm{(VMAE_{SSv2}^B)}$ + +UCF-101 & HMDB-51: For these smaller datasets, besides linear and nonlinear probing, we also use $k$ -NN probing (see Table 4 and Table 5). With SCALE $k$ -NN, we see a consistent improvement over the baseline and find that pre-trained MAE-based models greatly benefit from our + +Table 4. UCF Results. Linear and nonlinear probing accuracies on UCF-101 [48]. SCALE fit outperforms all the other models and, in the case of $\rho$ BYOL, even gets performance close to a fully fine-tuned model. Also, in most cases, SCALE linear outperforms the fine-tuned Transformer and achieves state-of-the-art results in linear probing (previous SotA using RGB frames was 92.6 [45]). We further see a significant accuracy improvement in $k$ -NN probing, especially for pre-trained MAE-based models. As a point of reference, the current best fully fine-tuned accuracy (which is not comparable with our setting) is $96.8\%$ [57]. + +
SVTρBYOLVMAEBVMAELVMAEft
k-NN87.2085.1935.0549.1496.82
SCALE k-NN89.0083.4765.6376.0297.38
Linear91.2789.5566.5384.5397.91
SCALE linear92.6591.4374.4686.7898.14
MLP91.1793.6071.9787.0498.04
Transformer92.2094.3468.2286.3098.04
SCALE ft92.9495.0076.0789.9298.46
FTreported93.795.496.1--
+ +Table 5. HMDB Results. Linear and nonlinear probing accuracies on HMDB-51 [34]. Despite the small size of the dataset, we see that $\mathrm{SCALE}_{\mathrm{ft}}$ is outperforming all the other methods, and in the case of SVT, it even outperforms the fully fine-tuned model. We also see that $\mathrm{SCALE}_{\mathrm{linear}}$ outperforms the Transformer in most cases with only a single linear layer (the best linear accuracy in the literature is $66.7\%$ [45]). Similar to UCF results, we see a considerable increase in the performance of $k$ -NN classifiers for pre-trained MAE-based models. + +
SVTρBYOLVMAEBVMAELVMAEft
k-NN51.8349.6721.9629.2172.81
SCALE k-NN56.0151.5637.1851.3071.83
Linear63.0761.1745.2260.2676.33
SCALE linear66.3363.9252.1562.3578.36
MLP63.0064.7749.0162.6177.45
Transformer63.9866.1647.3261.5076.86
SCALE ft68.1066.7951.8964.8379.34
FTreported67.273.673.3--
+ +training. This can be explained by the additional invariance properties introduced through the SET loss term in SCALE training. Across the board, we also see that in the case of linear probing, not only does SCALE linear outperform Linear, but it also outperforms Transformer, which leverages many more parameters. In the case of SVT, our SCALE linear also outperforms the best reported linear accuracy on UCF101 (92.7% vs. 92.6% [45]). Finally, $\mathrm{SCALE}_{\mathrm{ft}}$ achieves better results than all the nonlinear baselines and even outperforms the fully fine-tuned SVT (68.1% vs. 67.2%). + +Kinetics-400: We present our main results on Kinetics-400 [33] in Table 6. Our SCALE linear with SVT backbone beats the previous state of the art (71.8% vs. 71.5% [18]) and SCALE ft can even improve the accuracy of $\mathrm{VMAE}_{\mathrm{ft}}^{\mathrm{B}}$ + +Table 6. Kinetics-400 Results. Linear and nonlinear probing accuracies on Kinetics-400 [33] without any extra data and using RGB frames only. While SCALE linear is on par with Linear, we observe clear improvements for nonlinear probing in the case of SCALE ft. Note that the best linear accuracy on this dataset (without any extra data) is 71.5 [18] and the best full fine-tuning accuracy is 86.7 [61]. + +
SVTρBYOLVMAEBVMAELVMAEftb
Linear71.7168.8243.5060.7381.52
SCALE linear71.7868.3843.9660.6681.44
MLP71.1969.4245.4861.6481.27
Transformer72.1869.2844.8562.1581.70
SCALE ft72.3869.6346.1562.6781.84
+ +Table 7. Kinetics-400 Low-shot Results. Linear and nonlinear probing accuracies on $10\%$ of Kinetics-400 [33]. SCALE is more robust to the size of the labeled dataset. $\mathrm{SCALE}_{\mathrm{ft}}$ does not overfit like the other nonlinear probes (MLP and Transformer) and outperforms the baselines. $\mathrm{VMAE}_{\mathrm{ft}}^{\mathrm{B}}$ is greyed out since it was already fine-tuned on the whole labeled dataset and is only reported here for consistency with the other tables. + +
SVTρBYOLVMAEBVMAELVMAEft
Linear66.4356.4331.2548.4279.79
SCALE linear65.9657.7434.0349.2179.94
MLP65.4458.6830.4748.2779.37
Transformer64.9758.9529.8948.2779.47
SCALE ft67.0159.5233.9250.3680.47
+ +which is a strong supervised model, from $81.5\%$ to $81.84\%$ . + +Following the evaluation setup of self-supervised image representations [8, 7, 71], we also introduce low-shot K400 video classifications by sampling 10 percent of the videos (in a class-balanced way) and training the probes only on those. We still test on the whole evaluation set of K400. This low-shot setting is more aligned with the typical use-case of self-supervised models in which there is abundant unlabeled data for training via self-supervision and a small set of labeled data for fine-tuning. Results in Table 7 show that our method is particularly effective in this low-shot setting. While most other nonlinear probes overfit and perform worse than the linear probes, our SCALE ft does not overfit and clearly outperforms the baselines. + +# 4.3. Ablations + +In this section, we start from a baseline setup consisting of a two-layer transformer with a hidden size of 256, $20\%$ chance of masking clips, trained with a batch size of 512 for 200 epochs, and using two sets of 8 views for representation learning. Using SCALE ft, we explore different loss functions, masking ratios, number of layers, and finally, the number of views during training and testing. All experiments are performed on UCF and HMDB. + +Loss Function: As explained in the method section, we have two loss terms, and each of them can be enabled or + +Table 8. Loss Function. SCALE fit accuracy with different loss function combinations (the masking ratio here is $20\%$ ). We can see that having MCM is always beneficial, and the SET loss is almost always helpful. We use both loss terms for our final model. + +
SETMCMUCF-101HMDB-51
SVTρBYOLSVTρBYOL
✓✗91.8092.2064.5063.59
✗✓92.0193.8162.8164.05
✓✓93.2092.9964.5765.61
+ +Table 9. Masking Ratio. SCALEft accuracy with different masking ratios. We observe best results around 25% similar to NLP models [12] (15%), and different from low-level video models like VideoMAE [54] (90%). + +
Masking RatioUCF-101HMDB-51
SVTρBYOLSVTρBYOL
0.1593.1893.2564.3764.83
0.2593.2093.8165.4965.62
0.3593.1593.0664.1865.22
0.4592.9693.0263.3964.96
+ +Table 10. Transformer Capacity. SCALEft accuracy with different model capacities. Having more than one transformer layer and not too few hidden channels is necessary for the best performance. + +
Hidden +DimNum +LayersUCF-101HMDB-51
SVTρBYOLSVTρBYOL
641-92.62-63.16
1281-92.83-63.68
2561-92.86-64.39
128292.7893.5263.2665.55
256293.2093.8165.4965.62
512292.5793.6665.6864.77
128392.3393.2564.8365.55
256392.7593.5265.4965.16
512392.8692.9365.4964.84
+ +disabled for the pretraining. In Table 8 we show that having both loss terms is better than the individual loss terms. + +Masking Ratio: Masking ratio is an important hyperparameter and depends on the data modality, for example, BERT [12] uses $15\%$ , MSN [3] uses $30\%$ (for ViT-Base), MAE [25] uses $75\%$ , and VideoMAE [54] uses 90 to $95\%$ masking. Since our clip representations are somewhat abstract representations of the video, we expect the optimal masking ratio to be close to NLP models rather than video MAEs. We have observed a steady decrease in the pretraining task's performance with higher masking ratios, so we only tested low masking ratios in Table 9 and found out that $25\%$ is the optimal masking ratio. + +Transformer Capacity: We also explore the number of + +Table 11. Number of Views. SCALE ft accuracy with different numbers of clips and batch sizes. More views lead to consistent improvement, and large batch sizes are not necessary because of the hard negative samples. + +
Num ViewsBatch SizeUCF-101HMDB-51
SVTρBYOLSVTρBYOL
4 × 225692.6592.8364.3564.24
6 × 225692.7093.0764.5764.37
8 × 225692.8093.4964.6464.63
4 × 251292.6793.4964.8564.50
6 × 251293.1893.6864.9065.35
8 × 251293.2093.8165.4965.62
4 × 2102492.7593.3664.7765.15
6 × 2102492.9693.5764.9665.48
8 × 2102493.07OOM65.29OOM
+ +transformer layers and their hidden size in Table 10. We can see that having more than one transformer layer is necessary for good results and too few hidden channels can hurt performance. However, there is a trade-off, and deeper transformers can lead to worse performance. + +Number of Views: Finally, we studied the model performance as we changed the number of views and batch size fed to the model. As can be seen in Table 11, having more views has a large and consistent impact on the performance, and since we have hard negatives for contrastive loss within the video, we are not too reliant on large batch sizes. + +# 5. Conclusion + +In this paper, we introduced SCALE, a framework for video representation learning that aggregates the information from multiple clips from the same video. We combine contrastive learning and masked modeling with intuitions from predictive coding to obtain improved global and local representations of clips starting from frozen backbones. We evaluated these features using a wide array of backbones on different action classification and video understanding datasets and achieved strong or state-of-the-art results. The computational efficiency of our method is extremely useful for videos and opens the possibility to a wider group of researchers (with limited computational resources) to work on video representation learning than previously possible. We also believe that working with a set of clips is an interesting direction for representation learning. Finally, as a surprising and maybe alarming observation, even contrastive representations that were trained to be invariant to data augmentations and spatio-temporal crops can be used for contrastive masked modeling. This might be due to benign memorization [2], and understanding why this phenomenon happens might lead to a better understanding of contrastive learning. + +# 6. Acknowledgments + +This work was supported by grant 200020_188690 of the Swiss National Science Foundation (SNSF) and an Adobe award. + +# References + +[1] Hassan Akbari, Linagzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. ArXiv, abs/2104.11178, 2021. 1 +[2] Sotiris Anagnostidis, Gregor Bachmann, Lorenzo Noci, and Thomas Hofmann. The curious case of benign memorization. ArXiv, abs/2210.14019, 2022. 8 +[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. 1, 5, 8 +[4] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zach Beaver, Jana von Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, Vivek Natarajan, and Mohammad Norouzi. Big self-supervised models advance medical image classification. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3458-3468, 2021. 1 +[5] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. ArXiv, abs/2106.08254, 2022. 1 +[6] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9922-9931, 2020. 2 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1, 2, 5, 6, 7 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 5, 7 +[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2, 5 +[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 1 +[11] Ishan Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TCLR: Temporal contrastive learning for video representation. Computer Vision and Image Understanding, 219:103406, 2022. 2 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional + +transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3, 4, 8 +[13] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. ArXiv, abs/2207.07116, 2022. 2, 3 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 5 +[15] Haoqi Fan, Yanghao Li, Bo Xiong, Wan-Yen Lo, and Christoph Feichtenhofer. Pyslowfast. https://github.com/facebookresearch/slowfast, 2020.5 +[16] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022. 1, 2, 3, 4 +[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6201-6210, 2019. 5, 6 +[18] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021. 1, 2, 3, 5, 6, 7 +[19] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimae: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 1, 2, 4 +[20] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13505-13515, 2021. 3 +[21] Shreyank N Gowda, Marcus Rohrbach, and Laura Sevilla-Lara. Smart frame selection for action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1451-1459, 2021. 3 +[22] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 5, 6 +[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 5 +[24] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in Neural Information Processing Systems, 33:5679-5690, 2020. 2 +[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable + +vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1, 2, 3, 5, 6, 8 +[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2 +[27] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. ArXiv, abs/1808.06670, 2019. 5 +[28] Kaiqin Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7929, 2021. 6 +[29] Deng Huang, Wenhao Wu, Weiwen Hu, Xu Liu, Dongliang He, Zhihua Wu, Xiangmiao Wu, Mingkui Tan, and Errui Ding. Ascnet: Self-supervised video representation learning with appearance-speed consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8096-8105, 2021. 1 +[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 5, 6 +[31] Md. Mohaiminul Islam and Gedas Bertasius. Long movie clip classification with state-space video models. ArXiv, abs/2204.01692, 2022. 6 +[32] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9970-9980, 2021. 2 +[33] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 5, 7 +[34] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 2, 5, 7 +[35] Xingbin Liu, Jinghao Zhou, Tao Kong, Xianming Lin, and Rongrong Ji. Exploring target representations for masked autoencoders. arXiv preprint arXiv:2209.03917, 2022. 1 +[36] Yue Liu, Junqi Ma, Yufei Xie, Xuefeng Yang, Xingzhen Tao, Lin Peng, and Wei Gao. Contrastive predictive coding with transformer for video representation learning. Neurocomputing, 482:154-162, 2022. 3 +[37] Guillaume Lorre, Jaxonary Rabarisoa, Astrid Orcesi, Samia Ainouz, and Stephane Canu. Temporal contrastive pretraining for video action recognition. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 662-670, 2020. 1, 2, 3 +[38] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order + +verification. In European conference on computer vision, pages 527-544. Springer, 2016. 2 +[39] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2 +[40] Charlie Nash, João Carreira, Jacob Walker, Iain Barr, Andrew Jaegle, Mateusz Malinowski, and Peter Battaglia. Transframer: Arbitrary frame prediction with generative models. arXiv preprint arXiv:2203.09494, 2022. 3 +[41] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 2, 3, 5 +[42] Mandela Patrick, Yuki M. Asano, Bernie Huang, Ishan Misra, Florian Metze, João F. Henriques, and Andrea Vedaldi. Space-time crop & attend: Improving cross-modal video representation learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10540-10552, 2021. 3 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[44] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2874-2884, 2022. 1, 2, 3, 5, 6 +[45] Adria Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Pătrăucean, Florent Altché, Michal Valko, et al. Broaden your views for self-supervised video learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1255-1265, 2021. 1, 2, 3, 5, 7 +[46] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. ArXiv, abs/2010.04592, 2021. 5 +[47] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In European Conference on Computer Vision, pages 154-171. Springer, 2020. 3 +[48] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 5, 7 +[49] Chen Sun, Arsha Nagrani, Yonglong Tian, and Cordelia Schmid. Composable augmentation encoding for video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8834-8844, 2021. 3 +[50] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12607-12617, 2021. 3 + +[51] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 1, 3 +[52] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. arXiv preprint arXiv:2206.01204, 2022. 3 +[53] Nenad Tomasev, Ioana Bica, Brian McWilliams, Lars Holger Buesing, Razvan Pascanu, Charles Blundell, and Jovana Mitrovic. Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels on imageNet? In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 1 +[54] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. 1, 2, 3, 4, 5, 8 +[55] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2 +[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[57] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14010-14020, 2022. 3, 7 +[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:2740-2755, 2019. 5 +[59] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14713-14723, 2022. 1, 4 +[60] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 6 +[61] Chen Wei, Haoqi Fan, Saining Xie, Chaoxia Wu, Alan Loddon Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14648-14658, 2022. 7 +[62] Donglai Wei, Joseph J Lim, Andrew Zisserman, and William T Freeman. Learning and using the arrow of time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8052-8060, 2018. 2 +[63] Chaoxia Wu and Philipp Krahenbuhl. Towards long-form video understanding. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1884-1894, 2021. 2, 5, 6 + +[64] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 3 +[65] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 1, 3 +[66] Yu Wu, Linchao Zhu, Xiaohan Wang, Yi Yang, and Fei Wu. Learning to anticipate egocentric actions by imagination. IEEE Transactions on Image Processing, 30:1143-1152, 2020. 3 +[67] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3333-3343, 2022. 3 +[68] Liangzhe Yuan, Rui Qian, Yin Cui, Boqing Gong, Florian Schroff, Ming-Hsuan Yang, Hartwig Adam, and Ting Liu. Contextualized spatio-temporal contrastive learning with self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13977-13986, 2022. 3 +[69] Yujia Zhang, Lai-Man Po, Xuyuan Xu, Mengyang Liu, Yexin Wang, Weifeng Ou, Yuzhi Zhao, and Wing-Yin Yu. Contrastive spatio-temporal pretext learning for self-supervised video representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3380-3389, 2022. 3 +[70] Zhang Zhang and Dacheng Tao. Slow feature analysis for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 34(3):436-450, 2012. 1 +[71] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 7 +[72] Pan Zhou, Yichen Zhou, Chenyang Si, Weihao Yu, Teck Khim Ng, and Shuicheng Yan. Mugs: A multi-granular self-supervised learning framework. arXiv preprint arXiv:2203.14415, 2022. 1 \ No newline at end of file diff --git a/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/images.zip b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..09f6ebbcf8d08b64b18120a6d1bbbb7a8bf9e82f --- /dev/null +++ b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61ca9456c3433733735f8ca395ee25ff59ad122caccaee313ee3850eeccca0eb +size 486802 diff --git a/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/layout.json b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..06a52a6de1f221bc824babb0e74ac857f40e3f96 --- /dev/null +++ b/2023/Spatio-Temporal Crop Aggregation for Video Representation Learning/layout.json @@ -0,0 +1,9622 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 516, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 516, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 516, + 121 + ], + "type": "text", + "content": "Spatio-Temporal Crop Aggregation for Video Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 144, + 205, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 144, + 205, + 157 + ], + "spans": [ + { + "bbox": [ + 130, + 144, + 205, + 157 + ], + "type": "text", + "content": "Sepehr Sameni" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 158, + 212, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 158, + 212, + 171 + ], + "spans": [ + { + "bbox": [ + 121, + 158, + 212, + 171 + ], + "type": "text", + "content": "University of Bern" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 175, + 227, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 175, + 227, + 184 + ], + "spans": [ + { + "bbox": [ + 107, + 175, + 227, + 184 + ], + "type": "text", + "content": "sepehr.sameni@unibe.ch" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 268, + 144, + 330, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 144, + 330, + 156 + ], + "spans": [ + { + "bbox": [ + 268, + 144, + 330, + 156 + ], + "type": "text", + "content": "Simon Jenni" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 259, + 157, + 340, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 157, + 340, + 171 + ], + "spans": [ + { + "bbox": [ + 259, + 157, + 340, + 171 + ], + "type": "text", + "content": "Adobe Research" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 259, + 174, + 341, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 174, + 341, + 185 + ], + "spans": [ + { + "bbox": [ + 259, + 174, + 341, + 185 + ], + "type": "text", + "content": "jenni@adobe.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 396, + 144, + 462, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 144, + 462, + 156 + ], + "spans": [ + { + "bbox": [ + 396, + 144, + 462, + 156 + ], + "type": "text", + "content": "Paolo Favaro" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 383, + 157, + 474, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 157, + 474, + 171 + ], + "spans": [ + { + "bbox": [ + 383, + 157, + 474, + 171 + ], + "type": "text", + "content": "University of Bern" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 371, + 174, + 487, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 174, + 487, + 185 + ], + "spans": [ + { + "bbox": [ + 371, + 174, + 487, + 185 + ], + "type": "text", + "content": "paolo.favaro@unibe.ch" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 238, + 290, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 238, + 290, + 453 + ], + "spans": [ + { + "bbox": [ + 45, + 238, + 290, + 453 + ], + "type": "text", + "content": "We propose Spatio-temporal Crop Aggregation for video representation Learning (SCALE), a novel method that enjoys high scalability at both training and inference time. Our model builds long-range video features by learning from sets of video clip-level features extracted with a pretrained backbone. To train the model, we propose a self-supervised objective consisting of masked clip feature predictions. We apply sparsity to both the input, by extracting a random set of video clips, and to the loss function, by only reconstructing the sparse inputs. Moreover, we use dimensionality reduction by working in the latent space of a pre-trained backbone applied to single video clips. These techniques make our method not only extremely efficient to train but also highly effective in transfer learning. We demonstrate that our video representation yields state-of-the-art performance with linear, nonlinear, and " + }, + { + "bbox": [ + 45, + 238, + 290, + 453 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 45, + 238, + 290, + 453 + ], + "type": "text", + "content": "-NN probing on common action classification and video understanding datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 477, + 128, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 128, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 128, + 490 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 498, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 287, + 665 + ], + "type": "text", + "content": "Videos provide rich and detailed information about objects and their activities. Their analysis is, however, made challenging not only by the difference in the information provided across space and time but also by the high dimensionality of the data [45, 29, 70]. While computational resources are expected to scale over time, so is the demand for higher data resolution (both in space and time) and also the need for processing data with even more dimensions, such as videos of volumetric data [10]. Therefore, it is of paramount importance to explore methods that drastically reduce the computational requirements to process videos. Moreover, the annotation of videos is an extremely costly and time-consuming burden that makes the use of models pre-trained in a self-supervised manner essential [44]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Self-supervised learning (SSL) is a very popular technique to reduce the need for annotation because it can build useful representations from unlabeled data through an artificial goal, also called pseudo- or pretext-task. These repre" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 213, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 297 + ], + "type": "text", + "content": "sentations can either be evaluated through K-nearest neighbor or linear probing [71, 72] or through fine-tuning (i.e., as the initialization parameters of the trained model) [25, 54, 35] on a downstream task, where only a small labeled dataset is available. More remarkably, SSL pre-trained models can outperform models that were pre-trained on an annotated dataset [53, 4]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 298, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 547, + 525 + ], + "type": "text", + "content": "SSL methods for video representation learning present fundamental scalability challenges [54, 45, 29, 65]. A first major challenge is that training models from scratch for any new pseudo-task is not sustainable. A more viable setting is one where the processing of videos is broken down into multiple steps. First, one pre-processes video clips (spatiotemporal crops) only once via some pre-trained general-purpose model (e.g., trained via self-supervised learning [18, 23, 26, 44, 7, 37, 41]). In general, the idea is that this (compressed) local representation is stored and used later for other training purposes or retrieval. Then, one can train more efficiently other super-features that aggregate information from these local representations. Within this setting, we address two questions: 1) Can one further improve the performance of pre-computed video features by training a model on top of them? 2) Can such training be made computationally scalable? In this paper, we show that both questions have a positive answer. The answer exploits four fronts:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 535, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 535, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 314, + 535, + 547, + 643 + ], + "type": "text", + "content": "1. Input Sparsity: Sparsity in the input to the model [54, 25, 3, 19, 1] drastically reduces computational load and memory requirements, leveraging information redundancy in images and videos [16]. Inspired by prior work, we extract a sparse set of clips from a video (these are spatio-temporal video crops and not image patches or video tubelets); each clip is fed separately to a neural network so that we obtain a set of video clip representations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 547, + 713 + ], + "type": "text", + "content": "2. Output Sparsity: Another way to reduce the computational cost is to use a sparse reconstruction output instead of a dense one [71, 51, 59, 5]. This is not just a reduction of the number of terms in the loss function, but also in the number of actual outputs of" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5664" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 168 + ], + "type": "text", + "content": "the model, and thus also a reduction of the computations needed to obtain them. This is in contrast to MAE-based SSL methods for vision, where the proposed pseudo-tasks are based on the reconstruction of the whole input [25, 54, 19, 16], and even if the loss uses a subset of the tokens (the masked ones) for the loss calculation, the remaining tokens are still part of the decoder's output and computational graph." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 175, + 287, + 336 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 175, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 287, + 235 + ], + "type": "text", + "content": "3. Dimensionality Reduction: Inspired by prior work [71, 41, 13], instead of directly processing the raw input data, we work in the latent space. This allows us to further reduce the dimensionality of both input and output data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 241, + 286, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 241, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 54, + 241, + 286, + 336 + ], + "type": "text", + "content": "4. Use of a Pre-trained Backbone: To reduce training time and further speed up the processing per iteration, we exploit SSL pre-trained models. These pre-trained backbones already learn very strong short-term spatiotemporal features, and our approach is a way to extract longer-term features (see LVU results 2) by building a video representation on top of a set of pre-trained features (one for each video clip)." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 342, + 286, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 286, + 509 + ], + "type": "text", + "content": "To integrate all these components, we propose a novel SSL method that we call Spatio-temporal Crop Aggregation for video representation LEarning (SCALE). Given an input video, SCALE extracts a random set of clips and produces an embedding for each clip through a pre-trained backbone, which is kept frozen. These initial embeddings are then augmented in two ways: 1) each one is refined into a more discriminative feature, and 2) the set of all embeddings is summarized in a global feature. These global features can learn long-term correlations in the whole video by aggregating the short-term information in each clip embedding. The combination of the initial embeddings with their refinement and the global feature is then used in an ensemble for applications on new downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 510, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 286, + 712 + ], + "type": "text", + "content": "To train SCALE, we introduce two novel pseudo-tasks, which aim to improve the discriminability of the embeddings of each video clip as well as obtain a global representation of the video. One task, which we call Masked Clip Modeling (MCM), is the reconstruction of a video clip embedding as in masked autoencoders [54]. Masked embeddings are combined with positional encodings so that the model can (spatio-temporally) relate the missing input embeddings to the other available embeddings (similarly to BERT in Natural Language Processing [12]). A second task is to train the model to output a global feature token (which we refer to as CLS, just for consistency with previous works [7, 12, 14, 55, 9]) for a set of clips via contrastive learning, so that the global feature is invariant to the chosen set of clips from the same video, but can discriminate summary features of clips from other videos. Both tasks are trained via contrastive losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "text", + "content": "SCALE consistently improves upon pretrained features, with more evident gains considering the computational cost of other methods. For instance, SCALE outperforms VideoMAE [54] by " + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "text", + "content": " (1600 epochs checkpoint – fine-tuning) and " + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "text", + "content": " BYOL [18] by " + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "text", + "content": " (800 epochs checkpoint – linear probe) on Kinetics400 with 64 V100 GPUs in about 5 mins. In contrast, VideoMAE takes 27.7 hours to gain " + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "text", + "content": " (800 to 1600 epochs checkpoint), and " + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "text", + "content": " BYOL needs 48 hours to gain " + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 190 + ], + "type": "text", + "content": " (200 to 400 epochs checkpoint) with the same resources." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 193, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 193, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 193, + 545, + 217 + ], + "type": "text", + "content": "To summarize our contributions, we propose SCALE, a novel and highly scalable video representation method that" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 226, + 545, + 415 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 317, + 226, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 226, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 226, + 545, + 262 + ], + "type": "text", + "content": "- is trained via novel pseudo-tasks on sets of video clips (in contrast to existing methods that work only with pairs of clips at a time [44, 18, 45]);" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 273, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 545, + 344 + ], + "type": "text", + "content": "- results in video feature representations with a significant performance improvement in " + }, + { + "bbox": [ + 317, + 273, + 545, + 344 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 317, + 273, + 545, + 344 + ], + "type": "text", + "content": "-NN (retrieval), linear, and nonlinear probing across a wide range of datasets for action classification (UCF [48], HMDB [34], SSv2 [22], Kinetics400 [33]) and long-form video understanding (LVU [63]);" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 355, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 545, + 415 + ], + "type": "text", + "content": "- achieves consistent transfer learning performance improvement across diverse SotA pre-trained backbones (architectures, scale, and pre-training tasks). For example, our nonlinear probed model even outperforms fully fine-tuned SVT [44] on HMDB [34]." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 428, + 392, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 392, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 392, + 441 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 509 + ], + "type": "text", + "content": "Our approach relates to many prior works on (self-supervised) representation learning. In particular, SCALE relates to SSL approaches on videos, methods that rely on multiple views of the data, and predictive methods, where part of the data is predicted from another part." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 665 + ], + "type": "text", + "content": "Self-Supervised Video Representation Learning. Early SSL approaches on video were based on pseudo tasks, e.g., the recognition of transformations of video frame sequences [38, 6, 62]. More recently, popular methods developed on images have been successfully translated to video, e.g., contrastive methods [18, 23, 26], clustering-based methods [44, 7], or predictive approaches [37, 41]. Often, these approaches are tailored to video data by including additional learning signals, e.g., by combining contrastive methods with temporal constraints [11] and pretext tasks [32], or by incorporating audio [39], or optical flow [24]. These approaches typically learn representations with limited temporal extent, which can serve as backbones for our approach." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Learning from Multiple Views. Several methods have exploited multiple views (e.g., generated through space-time cropping) to represent and learn from videos. For example, many SSL approaches rely on multiple views of the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5665" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "data, e.g., in contrastive formulations [18, 44, 42], or predictive learning [45], where invariance to views is the goal. Other approaches aim to learn from the relation between two views, e.g., by predicting overlap [69], the relative distance [49], or cross-view feature prediction [68, 52] and reconstruction [40]. Besides exploiting multiple views for SSL, some works also propose general multi-view video models, e.g., by capturing and fusing features at different spatio-temporal resolutions [67], by aggregating information over longer time-spans [64, 47, 65, 57], or by selecting important frames [21]. These approaches, however, are learned end-to-end. In contrast, we propose a more scalable approach by learning a global video representation of pretrained features extracted over multiple crops using self-supervision." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 289, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 448 + ], + "type": "text", + "content": "Predictive Learning. One of our proposed SSL objectives is a prediction of space-time crop features given other video crop features. This approach is similar to masked token prediction as in BERT [12] and relates to several other methods in the literature. Masked input reconstruction methods have recently become popular on images [25] and successfully translated to video [54, 16]. Other approaches formulate masked prediction tasks in the learned feature space [71, 52, 13] or some fixed latent space [51]. Another line of work considers directional predictions (e.g., into the future) often formulated via contrastive predictive coding [41] applied to video [37, 36, 50, 66, 20]. These masked prediction tasks are typically formulated on a fixed grid (e.g., at the level of patches or frames). In contrast, our formulation is continuous, predicting features of arbitrarily sampled space-time crops." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 468, + 269, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 269, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 269, + 484 + ], + "type": "text", + "content": "3. Scalable Video Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 493, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 529 + ], + "type": "text", + "content": "To describe SCALE, we first define some basic notation and functions, including a general contrastive loss notation that we use for all training losses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 548, + 111, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 111, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 111, + 559 + ], + "type": "text", + "content": "3.1. Notation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "We use lower-case letters (e.g., " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ") to denote generic vectors and capital letters (e.g., " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ") to denote their sets. The expression " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "a \\circ b" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " denotes the concatenation of " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ". We also avoid indicating the parameters of networks (usually denoted by " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ") if their presence and role are clear from the context. Throughout the description of the method we refer to the training of neural networks and therefore at each iteration of the training we sample a minibatch of videos. All the equations in the next sections are written for a single video in the minibatch. Although we do not explicitly indicate it, all the contrastive losses also use the other videos in the minibatch as negatives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 408, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 408, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 408, + 83 + ], + "type": "text", + "content": "3.2. Contrastive Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": "InfoNCE is a powerful method for representation learning [41] that can be used to maximize the mutual information between two variables. Because we use this loss between different variables throughout our method, we introduce here a unified notation. Let the paired sets " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": " have " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": " elements each, " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "A = \\{a^i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "B = \\{b^i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "a^i" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": " are vectors of dimension " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "d_A" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "b^i" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": " are vectors of dimension " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "d_B" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": ". We also introduce two Multi Layer Perceptrons (MLP), parameterized with " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "\\theta_A" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "\\theta_B" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": ", to project these vectors onto a common space of dimension " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": ". After feeding the elements " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "a^i" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "inline_equation", + "content": "b^i" + }, + { + "bbox": [ + 304, + 90, + 547, + 234 + ], + "type": "text", + "content": " to the MLPs and normalizing them, we obtain" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 240, + 546, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 240, + 546, + 269 + ], + "spans": [ + { + "bbox": [ + 318, + 240, + 546, + 269 + ], + "type": "interline_equation", + "content": "\\tilde {a} ^ {i} = \\frac {\\operatorname {M L P} _ {\\theta_ {A}} \\left(a ^ {i}\\right)}{\\| \\operatorname {M L P} _ {\\theta_ {A}} \\left(a ^ {i}\\right) \\|} \\quad \\text {a n d} \\quad \\tilde {b} ^ {i} = \\frac {\\operatorname {M L P} _ {\\theta_ {B}} \\left(b ^ {i}\\right)}{\\| \\operatorname {M L P} _ {\\theta_ {B}} \\left(b ^ {i}\\right) \\|}, \\tag {1}", + "image_path": "bf5d93032a5d726942eb97a467b4fc9036e7367f196c03a4654ad8d0b6540879.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "text", + "content": " norm. We define the per-element loss based on the relative similarity of " + }, + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\tilde{a}^i" + }, + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\tilde{b}^i" + }, + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "text", + "content": ", and by using a temperature " + }, + { + "bbox": [ + 304, + 275, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\tau" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 337, + 318, + 545, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 318, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 337, + 318, + 545, + 373 + ], + "type": "interline_equation", + "content": "\\tilde {\\ell} ^ {i} (A, B, \\theta_ {A}, \\theta_ {B}) = - \\log \\frac {\\exp \\left(\\frac {\\tilde {a} ^ {i} \\cdot \\tilde {b} ^ {i}}{\\tau}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\tilde {a} ^ {i} \\cdot \\tilde {b} ^ {j}}{\\tau}\\right)}. \\tag {2}", + "image_path": "f6caff58da49cd1b2da57ab9b927c56730f309f59ca9231eb6c463a8d3566254.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 380, + 500, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 500, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 500, + 393 + ], + "type": "text", + "content": "We then make the loss symmetric [43] by using" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 401, + 545, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 401, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 341, + 401, + 545, + 417 + ], + "type": "interline_equation", + "content": "\\ell^ {i} (A, B, \\theta_ {A}, \\theta_ {B}) = \\tilde {\\ell} ^ {i} (A, B) + \\tilde {\\ell} ^ {i} (B, A). \\tag {3}", + "image_path": "c969086c559b08e7c12794945dae5c56ed7a4015e319058fe21a7a3a43167af7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "text", + "content": "Finally, we define the contrastive loss " + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cntr}}(A,B,\\theta_A,\\theta_B)" + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "text", + "content": " as the mean of " + }, + { + "bbox": [ + 306, + 424, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\ell^i" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 329, + 456, + 545, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 456, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 329, + 456, + 545, + 489 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c n t r}} (A, B, \\theta_ {A}, \\theta_ {B}) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\ell^ {i} (A, B, \\theta_ {A}, \\theta_ {B}). \\tag {4}", + "image_path": "7e76a4e91c8f39b386b9eb3a10e59663b2727951f9922fa26e4ed3e6eb4f97ed.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 496, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 533 + ], + "type": "text", + "content": "As mentioned earlier on, for simplicity, in the rest of the paper we will not indicate the parameters of the MLPs, and simply write " + }, + { + "bbox": [ + 304, + 496, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cntr}}(A,B)" + }, + { + "bbox": [ + 304, + 496, + 547, + 533 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 496, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\ell^i (A,B)" + }, + { + "bbox": [ + 304, + 496, + 547, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 539, + 409, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 409, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 409, + 552 + ], + "type": "text", + "content": "3.3. Training SCALE" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 558, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 677 + ], + "type": "text", + "content": "In our method, we integrate 4 principles to drastically reduce the computational complexity of learning a video representation: Input sparsity, output sparsity, dimensionality reduction, and use of a pre-trained backbone. Moreover, we introduce two pseudo-tasks to train the model. One task is based on the (contrastive) reconstruction of a masked video clip given some context video clips. The second task is to build a global representation that is (contrastively) invariant to the set of input sampled video clips. The overall training scheme of SCALE is illustrated in Figure 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Input sparsity. As a first step, rather than processing a whole video, we collect a sparse set of short video clips from the same video. Given a video " + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{H \\times W \\times T \\times 3}" + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5666" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 533, + 270 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 533, + 270 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 533, + 270 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 533, + 270 + ], + "type": "image", + "image_path": "ad3701cf1f5887dcfc3526852ced09c74df589e1fcf58ac08b692ec715456196.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "text", + "content": "Figure 1. Video Representation Learning with SCALE. For each video, SCALE extracts two sets of video clips " + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "inline_equation", + "content": "V_{1}^{1}, \\ldots, V_{K}^{1}" + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "inline_equation", + "content": "V_{1}^{2}, \\ldots, V_{K}^{2}" + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "text", + "content": ". Each video clip is processed separately through a frozen backbone " + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{\\text{基}}" + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "text", + "content": " and results in encoded video clips " + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "inline_equation", + "content": "C_{1}^{1}, \\ldots, C_{K}^{1}" + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "inline_equation", + "content": "C_{1}^{2}, \\ldots, C_{K}^{2}" + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "text", + "content": ". Then, a random set of encodings in each set is masked and reconstructed at the output of the predictor network (a transformer) " + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "inline_equation", + "content": "(\\ell^{i})" + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "text", + "content": ". The predictor network is also fed a class token CLS. The corresponding output token encodes a summary " + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\mathrm{CLS}^{m}" + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "text", + "content": "-th set of video clips. The objective for these summary tokens is to be similar only when encoding video clips from the same video " + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{SET}})" + }, + { + "bbox": [ + 46, + 278, + 548, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": " are the height, width, and duration (in frames) of the video, we sample " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "2 \\times K" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": " clips. We divide the clips into two sets randomly. Each clip in the first set " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "V_{i}^{1} \\in \\mathbb{R}^{H_{i}^{1} \\times W_{i}^{1} \\times T_{i}^{1} \\times 3}" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "i = 1, \\ldots, K" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": ", is obtained at the spatio-temporal location " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "X_{i}^{1}, Y_{i}^{1}, Q_{i}^{1}" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": " with different data augmentations and dimensions " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "H_{i}^{1}" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "W_{i}^{1}" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "T_{i}^{1}" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": ". Similarly, we denote the second set of clips " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "V_{i}^{2}" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "i = 1, \\ldots, K" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": ". We also normalize their coordinates relative to the video dimensions and embed them onto a feature vector " + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "inline_equation", + "content": "P_{i}^{j}" + }, + { + "bbox": [ + 46, + 354, + 289, + 477 + ], + "type": "text", + "content": " by feeding them to a learnable MLP. We denote these embeddings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 484, + 287, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 484, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 484, + 287, + 523 + ], + "type": "interline_equation", + "content": "P _ {i} ^ {j} = \\operatorname {M L P} \\left(\\left[ \\frac {X _ {i} ^ {j}}{H}, \\frac {Y _ {i} ^ {j}}{W}, \\frac {Q _ {i} ^ {j}}{T}, \\frac {X _ {i} ^ {j} + H _ {i} ^ {j}}{H}, \\frac {Y _ {i} ^ {j} + W _ {i} ^ {j}}{W}, \\frac {Q _ {i} ^ {j} + T _ {i} ^ {j}}{T} \\right] ^ {\\top}\\right), \\tag {5}", + "image_path": "3364285879076788484b4d11df884dbc8eee5f928165768014b769281eb50cd6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 533, + 185, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 185, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 185, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 533, + 185, + 544 + ], + "type": "inline_equation", + "content": "j = 1,2" + }, + { + "bbox": [ + 47, + 533, + 185, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 533, + 185, + 544 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,K" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": "Dimensionality reduction and pre-trained backbone. To reduce the dimensionality of each video clip, we feed them independently to a frozen encoder " + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{\\ast}" + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": ", to obtain the encodings " + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "inline_equation", + "content": "C_{i}^{j} = \\mathrm{E}_{\\ast}(V_{i}^{j})" + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "inline_equation", + "content": "j = 1,2" + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,K" + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": ". Our framework is encoder-agnostic and thus can work with encoders obtained through different training schemes (supervised, contrastive, or autoencoder). In addition to reducing the dimensionality, we make the training even more scalable by using pre-trained and frozen encoders. Note, however, that if performance is the main goal, it is possible to also train a sparse backbone end to end with multiple clips. Thanks to token dropping, one can drop up to " + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": " of the tokens [19] and still build a good representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Output sparsity. As a self-supervised signal for our video" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": "representation learning, we use a (contrastive) reconstruction loss. To reduce the computational cost, instead of predicting the features for the whole video [59, 19, 54, 16] (asymmetric decoding), we only reconstruct a sparse set of masked clips. Our reconstruction objective is based on the observation that video signals carry a lot of redundancy. Hence, we introduce a model, the predictor network, to predict masked video clip embeddings given the other video clip embeddings (the context). We follow the general approach of BERT [12] but implement the predictor network as a masked autoencoder, where the reconstruction is based on a contrastive loss. The loss is applied only to a sparse set " + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "inline_equation", + "content": "M^1 \\subset \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": " of masked video clips. These clips are replaced by a learned MSK token. All embeddings " + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "inline_equation", + "content": "C_i^1" + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": ", including the masked ones, are added to their corresponding position encoding " + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "inline_equation", + "content": "P_i^1" + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": " and are then fed to the predictor network. We also include an additional learnable CLS token as input for the predictor network, which will be used for tasks with multiple video clips. We denote the outputs of the predictor network as " + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "inline_equation", + "content": "\\hat{C}_i^1" + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": " for the tokens corresponding to " + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "inline_equation", + "content": "C_i^1" + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": ", and as " + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{CLS}^1" + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": " for the token corresponding to CLS. Similarly, we feed as inputs separately from the previous set all the video clips " + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "inline_equation", + "content": "C_i^2" + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": " with their corresponding positional embeddings " + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "inline_equation", + "content": "P_i^2" + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": ", and the same CLS token, and obtain " + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "inline_equation", + "content": "\\hat{C}_i^2" + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{CLS}^2" + }, + { + "bbox": [ + 304, + 354, + 547, + 668 + ], + "type": "text", + "content": " respectively (see Figure 1 for a visual depiction of these processing steps)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Contrastive reconstruction. Modeling all the details of a masked clip, even in the latent space and even given the redundancy in videos, is a demanding task. Rather than" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5667" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "increasing the capacity of our model, since we are aiming for scalability, we keep our predictor network a (relatively) shallow network and use contrastive learning [41]. With contrastive learning, the predicted representation of the masked tokens should only be closer to the original un-masked clip representation (after an MLP projection) than from all other clips from the same video and the rest of the minibatch. Note that the rest of the clips in the same video act as hard negatives in contrastive learning [46]. Also, since we are using a frozen backbone, we can afford to use large minibatch sizes, which is known to be beneficial for contrastive learning [9]. We call this contrastive reconstruction loss the Masked Clip Modeling (MCM) loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 234, + 287, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 234, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 71, + 234, + 287, + 261 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M C M}} = \\sum_ {i \\in M ^ {1}} \\ell^ {i} \\left(\\hat {C} ^ {1}, C ^ {1}\\right) + \\sum_ {j \\in M ^ {2}} \\ell^ {j} \\left(\\hat {C} ^ {2}, C ^ {2}\\right). \\tag {6}", + "image_path": "9d386163745e955c058a091fb1b45c8a3700e5ced87da9e768d26f3e7508ff4a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 445 + ], + "type": "text", + "content": "Multiple video clips loss. The predictor network outputs features for each video clip that are highly discriminative. This task is similar to that of a masked autoencoder [25] and gives you an enhanced per-clip representation. For many video tasks, we need a global representation for the whole video; for that, we introduce an additional pseudo-task that captures a more global representation of a set of video clips. Our task takes inspiration from contrastive learning methods used in SSL, which yield representations that perform well with linear probing [8]. The loss aims to make the " + }, + { + "bbox": [ + 46, + 266, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{CLS}^1" + }, + { + "bbox": [ + 46, + 266, + 289, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\mathrm{CLS}^2" + }, + { + "bbox": [ + 46, + 266, + 289, + 445 + ], + "type": "text", + "content": " tokens returned from the predictor network more similar (recall that these two tokens are obtained from two separate groups of video clips extracted from the same video) than to other class tokens from other videos within the minibatch" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 452, + 287, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 452, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 107, + 452, + 287, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S E T}} = \\mathcal {L} _ {\\mathrm {c n t r}} \\left(\\mathrm {C L S} ^ {1}, \\mathrm {C L S} ^ {2}\\right). \\tag {7}", + "image_path": "b7503219e837b8b3c4f420cee077ea6d657eda29905eda396c1af8709264c84e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 534 + ], + "type": "text", + "content": "In addition to our contrastive loss (InfoNCE), one can use clustering [7] or regression [23] losses. We choose InfoNCE for simplicity and for better compatibility between the losses. As the overall loss, we use the sum of both loss terms (without any weights)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 540, + 287, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 540, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 125, + 540, + 287, + 553 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M C M}} + \\mathcal {L} _ {\\mathrm {S E T}}, \\tag {8}", + "image_path": "b4bf2a22961268e3a2d7e31d3bc5a459fc1d9564a33fef41814d098000ad9469.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 580, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 665 + ], + "type": "text", + "content": "We evaluate SCALE on several commonly used action classification datasets for video representation learning. As our performance metric, we primarily use linear probing and nonlinear probing [27]. For the smaller datasets, we also use " + }, + { + "bbox": [ + 46, + 580, + 287, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 580, + 287, + 665 + ], + "type": "text", + "content": "-NN classifiers (which are training-free) and demonstrate that the proposed method improves upon both unsupervised and supervised backbones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 231, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 231, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 231, + 685 + ], + "type": "text", + "content": "4.1. Experimental Setup and Protocols" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Computational Efficiency: In Table 1, we show estimates of the maximum batch sizes and the training throughput for" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 113, + 547, + 186 + ], + "blocks": [ + { + "bbox": [ + 305, + 68, + 545, + 113 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 545, + 113 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 545, + 113 + ], + "type": "text", + "content": "Table 1. Computational Resources. GPU VRAM usage (GB) and max training speed (using one 3090 GPU) of " + }, + { + "bbox": [ + 305, + 68, + 545, + 113 + ], + "type": "inline_equation", + "content": "\\mathrm{ViT_B}" + }, + { + "bbox": [ + 305, + 68, + 545, + 113 + ], + "type": "text", + "content": " with varying batch sizes and SSL tasks for videos. " + }, + { + "bbox": [ + 305, + 68, + 545, + 113 + ], + "type": "inline_equation", + "content": "\\mathrm{MoCo_{Sparse}^{V3}}" + }, + { + "bbox": [ + 305, + 68, + 545, + 113 + ], + "type": "text", + "content": " is akin to MSN [3]. OOM indicates Out Of Memory error." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 113, + 547, + 186 + ], + "lines": [ + { + "bbox": [ + 306, + 113, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 306, + 113, + 547, + 186 + ], + "type": "table", + "html": "
Batch Size819572048Samples/s
MoCoV323.47OOMOOMOOM8.66
VMAE11.4222.80OOMOOM33.54
MoCoV3Sparse5.499.2723.62OOM28.83
SCALE1.211.231.5819.529224.65
", + "image_path": "749833dfbe6aededec5c035b4915f9efa8de72ebe4c987fedd9cf9c37c6b2747.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 208, + 547, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 547, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 547, + 267 + ], + "type": "text", + "content": "different methods trained with the same computational and memory resources. As can be seen, SCALE is orders of magnitude more efficient than other SotA methods. Also during multi-crop inference, our method only results in an FLOP increase of approximately " + }, + { + "bbox": [ + 304, + 208, + 547, + 267 + ], + "type": "inline_equation", + "content": "0.001\\%" + }, + { + "bbox": [ + 304, + 208, + 547, + 267 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 268, + 547, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 364 + ], + "type": "text", + "content": "Datasets: Following prior work [44, 18, 45] we use Kinetics-400 [33], UCF-101 [48] (split 1), HMDB-51 [34] (split 1), and Something-Something v2 (SSv2) [22] to train and evaluate our models. We also use the LVU benchmark [63] to showcase our long-form video understanding capabilities. Note that almost " + }, + { + "bbox": [ + 304, + 268, + 547, + 364 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 304, + 268, + 547, + 364 + ], + "type": "text", + "content": " of LVU videos are not available to download from YouTube anymore; thus, our results are not directly comparable with prior methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 365, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 533 + ], + "type": "text", + "content": "Pretrained backbones: We use the pretrained checkpoints of " + }, + { + "bbox": [ + 304, + 365, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 365, + 547, + 533 + ], + "type": "text", + "content": " BYOL [18], SVT [44], and three variants of Video-MAE [54] (base(B), large(L), and fine-tuned base(FT)). We choose " + }, + { + "bbox": [ + 304, + 365, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 365, + 547, + 533 + ], + "type": "text", + "content": " BYOL for their excellent linear performance, SVT for the usage of ViT [14], and VMAE for showing 1) the applicability of our proposed method to MAE models, 2) the scalability of our method to larger models, and 3) possibility of using supervisedly fine-tuned models as our backbone. All the models are self-supervisedly pretrained on Kinetics-400, except the fine-tuned VMAE base that was also supervisedly finetuned on Kinetics-400. We also used a backbone pretrained and fine-tuned on SSv2 (VMAE" + }, + { + "bbox": [ + 304, + 365, + 547, + 533 + ], + "type": "inline_equation", + "content": "_{SSv2}" + }, + { + "bbox": [ + 304, + 365, + 547, + 533 + ], + "type": "text", + "content": ") for the SSv2 experiment to show the universality of SCALE with respect to the pretraining dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": "Self-supervised Training: For training data, we extract 16 clips of 16 frames from each video per dataset and save their feature encodings to disk. We use PySlowFast's common data augmentations for that [15]. For evaluation, we follow the " + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "inline_equation", + "content": "5 \\times 3" + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": " scheme [17] (uniformly sampling 5 clips from a video along its temporal axis and then taking 3 spatial crops) and extract 15 clips from each video (except for SSv2, where we extract " + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "inline_equation", + "content": "2 \\times 3" + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": " clips [58]). As the architecture for the predictor network, we use an encoder-only Transformer [56] and a three-layer MLP (without batch normalization [30]) for the contrastive heads. Unless stated otherwise, we train our models for 500 epochs (for example, training with " + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{VMAE}^{\\mathrm{B}}" + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": " on SSV2 takes 137 minutes with one 3090 GPU) with a batch size of 512 and use all 16 clips." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Evaluation: Since our focus is on efficient and scalable" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5668" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 102, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 102 + ], + "type": "text", + "content": "Table 2. Long-Form Video Understanding Results. Linear and nonlinear probing accuracies on LVU [63] classification tasks. SCALE shows significant performance improvement compared to baselines, indicating its ability to capture long-form video features. The first two rows are greyed out as direct comparison is not possible due to the unavailability of the full dataset for download." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 102, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 47, + 102, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 553, + 243 + ], + "type": "table", + "html": "
RelationSpeakSceneDirectorGenreWriterYear
SlowFast+NL ViS4mer[17, 60]52.4035.8054.7044.9053.0036.3052.50
[31]57.1440.7967.4462.6154.7148.8044.75
SVTLinear64.7035.7762.3337.2754.3552.1228.46
SCALE linear73.52 (+8.82)40.65 (+4.88)68.83 (+6.50)46.36 (+9.09)57.94 (+3.59)56.38 (+4.26)39.23 (+10.77)
MLP67.6439.0266.2345.4556.9257.4436.15
Transformer70.5840.6568.8347.2757.1758.5136.92
SCALE ft76.47 (+5.89)42.27 (+1.62)74.02 (+5.19)49.09 (+1.82)58.97 (+1.80)62.76 (+4.25)39.23 (+2.31)
ρBYOLLinear52.9438.2153.2436.3651.7957.4433.84
SCALE linear67.64 (+14.70)43.08 (+4.87)66.23 (+12.99)44.54 (+8.18)53.33 (+1.54)60.63 (+3.19)40.00 (+6.16)
MLP62.3541.4662.4247.2752.5659.5740.00
Transformer65.0944.7166.2350.9053.0761.7043.84
SCALE ft67.64 (+2.55)45.52 (+0.81)71.42 (+5.19)51.81 (+0.91)55.72 (+2.65)65.95 (+4.25)46.92 (+3.08)
", + "image_path": "92065676fb9af73983265bbd397d796bc3142b315d3b4a4458a4c7da6fa15804.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 262, + 289, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 289, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 289, + 514 + ], + "type": "text", + "content": "video classification, we always freeze the backbones in our evaluation (as in our self-supervised pretraining) and either train a linear classifier [44, 18] or fine-tune the predictor network (the transformer) with an additional linear head. Therefore, when we refer to fine-tuning (ft), we only adapt the nonlinear head (e.g., predictor network) but not the backbone. We apply a grid search for the hyper-parameters of the heads covering learning rate, weight decay, batch size, and optimizer type. Similar to MAE [25], we found that applying a batch normalization layer [30] without affine transformations is beneficial for VideoMAE models. As the linear baseline, we consider the well-established ensembling approach, i.e., we average the softmax predictions of the 15 clips (6 for SSv2) to obtain the final prediction. For models that process multiple clips at once (like ours), we likewise apply a linear softmax head on the concatenation of the individual clip features and the [SET] token before averaging to obtain the final prediction. For the smaller datasets, we also use " + }, + { + "bbox": [ + 46, + 262, + 289, + 514 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 262, + 289, + 514 + ], + "type": "text", + "content": "-NN classification, where, similar to DINO [7], we always use " + }, + { + "bbox": [ + 46, + 262, + 289, + 514 + ], + "type": "inline_equation", + "content": "k = 20" + }, + { + "bbox": [ + 46, + 262, + 289, + 514 + ], + "type": "text", + "content": " and work with 12 normalized representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "type": "text", + "content": "Nonlinear baselines: As SCALE is a nonlinear model, we consider an MLP on top of the frozen backbone as a nonlinear baseline. As a further baseline and to illustrate the effect of our self-supervised pre-training, we consider a Transformer trained on all the clip representations. This Transformer uses the exact same architecture as SCALE, and only differs in the initialization: in the case of SCALE we start from our proposed SSL pre-trained weights instead of random initialization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 634, + 105, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 105, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 105, + 645 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "LVU: One of the benefits of SCALE is that it can be used to process long videos, even though the backbones were trained on short videos only. To demonstrate the ability to capture long-form video features, we evaluated SCALE on LVU [63], a benchmark that involves seven classifica" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 325, + 545, + 399 + ], + "blocks": [ + { + "bbox": [ + 304, + 259, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 304, + 259, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 545, + 326 + ], + "type": "text", + "content": "Table 3. SSv2 Results. Linear and nonlinear probing accuracies on SSv2 [22]. We see that both SCALE linear and SCALE ft outperform other methods and improve the classification accuracies by a large margin. We also see that SCALE ft, with its better initialization, always outperforms the Transformer. " + }, + { + "bbox": [ + 304, + 259, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathrm{VMAE}_{\\mathrm{SSv2}}^{\\mathrm{B}}" + }, + { + "bbox": [ + 304, + 259, + 545, + 326 + ], + "type": "text", + "content": " was pretrained and fine-tuned on SSv2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 325, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 306, + 325, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 325, + 545, + 399 + ], + "type": "table", + "html": "
SVTρBYOLVMAEBVMAELVMAEftVMAEBSSv2
Linear20.3025.3018.3127.9428.9070.53
SCALE linear25.2627.1621.2430.1833.2570.63
MLP21.4326.4619.4227.9629.8370.52
Transformer29.2430.9924.2634.3935.6070.57
SCALE ft29.6831.8325.2536.3437.3870.69
", + "image_path": "af92a4b99b5208c81c76fcc4c5b1a2f6510ea88a0a855dc4cbd3c4392eea3915.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 423, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 506 + ], + "type": "text", + "content": "tion (and two regression) tasks on minute-long videos. Past studies [31, 60] have established that increasing the input's time span enhances accuracy in this challenging dataset. As shown in Table 2, our experiments indicate that SCALE can improve the baseline model's performance by a considerable margin. Moreover, we found that fine-tuning SCALE can lead to further enhancements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 508, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 653 + ], + "type": "text", + "content": "SSv2: Multiple classes in SSv2 share similar backgrounds and only differ in motion [28], suggesting that high performance on this dataset demonstrates that the model has captured strong motion-related contextual cues [44]. Results in Table 3 show that we outperform the state-of-the-art. On this dataset, we see a large performance gap between models that process single clips at a time (Linear and MLP) and the models that work with multiple clips (SCALE and Transformer). We can see SCALE linear is also outperforming the MLP, showing that SCALE is able to capture motion and long-form temporal features of the video. We even improve the supervised model trained on SSv2 " + }, + { + "bbox": [ + 304, + 508, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{(VMAE_{SSv2}^B)}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "UCF-101 & HMDB-51: For these smaller datasets, besides linear and nonlinear probing, we also use " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "-NN probing (see Table 4 and Table 5). With SCALE " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "-NN, we see a consistent improvement over the baseline and find that pre-trained MAE-based models greatly benefit from our" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5669" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 178, + 288, + 291 + ], + "blocks": [ + { + "bbox": [ + 46, + 68, + 289, + 178 + ], + "lines": [ + { + "bbox": [ + 46, + 68, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 68, + 289, + 178 + ], + "type": "text", + "content": "Table 4. UCF Results. Linear and nonlinear probing accuracies on UCF-101 [48]. SCALE fit outperforms all the other models and, in the case of " + }, + { + "bbox": [ + 46, + 68, + 289, + 178 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 68, + 289, + 178 + ], + "type": "text", + "content": " BYOL, even gets performance close to a fully fine-tuned model. Also, in most cases, SCALE linear outperforms the fine-tuned Transformer and achieves state-of-the-art results in linear probing (previous SotA using RGB frames was 92.6 [45]). We further see a significant accuracy improvement in " + }, + { + "bbox": [ + 46, + 68, + 289, + 178 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 68, + 289, + 178 + ], + "type": "text", + "content": "-NN probing, especially for pre-trained MAE-based models. As a point of reference, the current best fully fine-tuned accuracy (which is not comparable with our setting) is " + }, + { + "bbox": [ + 46, + 68, + 289, + 178 + ], + "type": "inline_equation", + "content": "96.8\\%" + }, + { + "bbox": [ + 46, + 68, + 289, + 178 + ], + "type": "text", + "content": " [57]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 178, + 288, + 291 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 288, + 291 + ], + "type": "table", + "html": "
SVTρBYOLVMAEBVMAELVMAEft
k-NN87.2085.1935.0549.1496.82
SCALE k-NN89.0083.4765.6376.0297.38
Linear91.2789.5566.5384.5397.91
SCALE linear92.6591.4374.4686.7898.14
MLP91.1793.6071.9787.0498.04
Transformer92.2094.3468.2286.3098.04
SCALE ft92.9495.0076.0789.9298.46
FTreported93.795.496.1--
", + "image_path": "ada5b02e549ae549e59cce9d81c1df747e671d32a5a93b5497ef58c04e74a85a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 400, + 288, + 513 + ], + "blocks": [ + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "type": "text", + "content": "Table 5. HMDB Results. Linear and nonlinear probing accuracies on HMDB-51 [34]. Despite the small size of the dataset, we see that " + }, + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{SCALE}_{\\mathrm{ft}}" + }, + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "type": "text", + "content": " is outperforming all the other methods, and in the case of SVT, it even outperforms the fully fine-tuned model. We also see that " + }, + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{SCALE}_{\\mathrm{linear}}" + }, + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "type": "text", + "content": " outperforms the Transformer in most cases with only a single linear layer (the best linear accuracy in the literature is " + }, + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "type": "inline_equation", + "content": "66.7\\%" + }, + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "type": "text", + "content": " [45]). Similar to UCF results, we see a considerable increase in the performance of " + }, + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 301, + 288, + 399 + ], + "type": "text", + "content": "-NN classifiers for pre-trained MAE-based models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 400, + 288, + 513 + ], + "lines": [ + { + "bbox": [ + 47, + 400, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 288, + 513 + ], + "type": "table", + "html": "
SVTρBYOLVMAEBVMAELVMAEft
k-NN51.8349.6721.9629.2172.81
SCALE k-NN56.0151.5637.1851.3071.83
Linear63.0761.1745.2260.2676.33
SCALE linear66.3363.9252.1562.3578.36
MLP63.0064.7749.0162.6177.45
Transformer63.9866.1647.3261.5076.86
SCALE ft68.1066.7951.8964.8379.34
FTreported67.273.673.3--
", + "image_path": "49d5ea81ff2d75043b9e1ed22652893e29ebf36c1bf7574ebf91e801ac1a0921.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": "training. This can be explained by the additional invariance properties introduced through the SET loss term in SCALE training. Across the board, we also see that in the case of linear probing, not only does SCALE linear outperform Linear, but it also outperforms Transformer, which leverages many more parameters. In the case of SVT, our SCALE linear also outperforms the best reported linear accuracy on UCF101 (92.7% vs. 92.6% [45]). Finally, " + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{SCALE}_{\\mathrm{ft}}" + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": " achieves better results than all the nonlinear baselines and even outperforms the fully fine-tuned SVT (68.1% vs. 67.2%)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Kinetics-400: We present our main results on Kinetics-400 [33] in Table 6. Our SCALE linear with SVT backbone beats the previous state of the art (71.8% vs. 71.5% [18]) and SCALE ft can even improve the accuracy of " + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{VMAE}_{\\mathrm{ft}}^{\\mathrm{B}}" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 306, + 146, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 545, + 145 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 545, + 145 + ], + "type": "text", + "content": "Table 6. Kinetics-400 Results. Linear and nonlinear probing accuracies on Kinetics-400 [33] without any extra data and using RGB frames only. While SCALE linear is on par with Linear, we observe clear improvements for nonlinear probing in the case of SCALE ft. Note that the best linear accuracy on this dataset (without any extra data) is 71.5 [18] and the best full fine-tuning accuracy is 86.7 [61]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 146, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 306, + 146, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 146, + 545, + 220 + ], + "type": "table", + "html": "
SVTρBYOLVMAEBVMAELVMAEftb
Linear71.7168.8243.5060.7381.52
SCALE linear71.7868.3843.9660.6681.44
MLP71.1969.4245.4861.6481.27
Transformer72.1869.2844.8562.1581.70
SCALE ft72.3869.6346.1562.6781.84
", + "image_path": "8baf8edbbb85a91b209ff1c612b1149ac8c9156cbe7ffd1a5848194b0da0f1fa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 306, + 545, + 380 + ], + "blocks": [ + { + "bbox": [ + 304, + 229, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 305 + ], + "type": "text", + "content": "Table 7. Kinetics-400 Low-shot Results. Linear and nonlinear probing accuracies on " + }, + { + "bbox": [ + 304, + 229, + 545, + 305 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 229, + 545, + 305 + ], + "type": "text", + "content": " of Kinetics-400 [33]. SCALE is more robust to the size of the labeled dataset. " + }, + { + "bbox": [ + 304, + 229, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\mathrm{SCALE}_{\\mathrm{ft}}" + }, + { + "bbox": [ + 304, + 229, + 545, + 305 + ], + "type": "text", + "content": " does not overfit like the other nonlinear probes (MLP and Transformer) and outperforms the baselines. " + }, + { + "bbox": [ + 304, + 229, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\mathrm{VMAE}_{\\mathrm{ft}}^{\\mathrm{B}}" + }, + { + "bbox": [ + 304, + 229, + 545, + 305 + ], + "type": "text", + "content": " is greyed out since it was already fine-tuned on the whole labeled dataset and is only reported here for consistency with the other tables." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 306, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 306, + 306, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 545, + 380 + ], + "type": "table", + "html": "
SVTρBYOLVMAEBVMAELVMAEft
Linear66.4356.4331.2548.4279.79
SCALE linear65.9657.7434.0349.2179.94
MLP65.4458.6830.4748.2779.37
Transformer64.9758.9529.8948.2779.47
SCALE ft67.0159.5233.9250.3680.47
", + "image_path": "5646ec0ffda4941dd3c8120f6e525249126bd9765e5f93aa64a1992a781f3e32.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 400, + 544, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 544, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 544, + 412 + ], + "type": "text", + "content": "which is a strong supervised model, from " + }, + { + "bbox": [ + 304, + 400, + 544, + 412 + ], + "type": "inline_equation", + "content": "81.5\\%" + }, + { + "bbox": [ + 304, + 400, + 544, + 412 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 400, + 544, + 412 + ], + "type": "inline_equation", + "content": "81.84\\%" + }, + { + "bbox": [ + 304, + 400, + 544, + 412 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 567 + ], + "type": "text", + "content": "Following the evaluation setup of self-supervised image representations [8, 7, 71], we also introduce low-shot K400 video classifications by sampling 10 percent of the videos (in a class-balanced way) and training the probes only on those. We still test on the whole evaluation set of K400. This low-shot setting is more aligned with the typical use-case of self-supervised models in which there is abundant unlabeled data for training via self-supervision and a small set of labeled data for fine-tuning. Results in Table 7 show that our method is particularly effective in this low-shot setting. While most other nonlinear probes overfit and perform worse than the linear probes, our SCALE ft does not overfit and clearly outperforms the baselines." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 575, + 373, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 373, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 373, + 586 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "In this section, we start from a baseline setup consisting of a two-layer transformer with a hidden size of 256, " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " chance of masking clips, trained with a batch size of 512 for 200 epochs, and using two sets of 8 views for representation learning. Using SCALE ft, we explore different loss functions, masking ratios, number of layers, and finally, the number of views during training and testing. All experiments are performed on UCF and HMDB." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "content": "Loss Function: As explained in the method section, we have two loss terms, and each of them can be enabled or" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5670" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 121, + 287, + 182 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 288, + 114 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 288, + 114 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 288, + 114 + ], + "type": "text", + "content": "Table 8. Loss Function. SCALE fit accuracy with different loss function combinations (the masking ratio here is " + }, + { + "bbox": [ + 47, + 69, + 288, + 114 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 69, + 288, + 114 + ], + "type": "text", + "content": "). We can see that having MCM is always beneficial, and the SET loss is almost always helpful. We use both loss terms for our final model." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 121, + 287, + 182 + ], + "lines": [ + { + "bbox": [ + 47, + 121, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 287, + 182 + ], + "type": "table", + "html": "
SETMCMUCF-101HMDB-51
SVTρBYOLSVTρBYOL
✓✗91.8092.2064.5063.59
✗✓92.0193.8162.8164.05
✓✓93.2092.9964.5765.61
", + "image_path": "bd4ed6b28e9cf28cc3ec44236b723b3a830767dad5c18ed354c9689bd28564f2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 234, + 287, + 319 + ], + "blocks": [ + { + "bbox": [ + 47, + 190, + 287, + 235 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 287, + 235 + ], + "type": "text", + "content": "Table 9. Masking Ratio. SCALEft accuracy with different masking ratios. We observe best results around 25% similar to NLP models [12] (15%), and different from low-level video models like VideoMAE [54] (90%)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 234, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 47, + 234, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 287, + 319 + ], + "type": "table", + "html": "
Masking RatioUCF-101HMDB-51
SVTρBYOLSVTρBYOL
0.1593.1893.2564.3764.83
0.2593.2093.8165.4965.62
0.3593.1593.0664.1865.22
0.4592.9693.0263.3964.96
", + "image_path": "6db47fa8f338658d26cd59683fb5897a9b9aeb7271a8efb202dc1bec125f964b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 47, + 369, + 288, + 514 + ], + "blocks": [ + { + "bbox": [ + 47, + 327, + 287, + 361 + ], + "lines": [ + { + "bbox": [ + 47, + 327, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 287, + 361 + ], + "type": "text", + "content": "Table 10. Transformer Capacity. SCALEft accuracy with different model capacities. Having more than one transformer layer and not too few hidden channels is necessary for the best performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 369, + 288, + 514 + ], + "lines": [ + { + "bbox": [ + 47, + 369, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 288, + 514 + ], + "type": "table", + "html": "
Hidden \nDimNum \nLayersUCF-101HMDB-51
SVTρBYOLSVTρBYOL
641-92.62-63.16
1281-92.83-63.68
2561-92.86-64.39
128292.7893.5263.2665.55
256293.2093.8165.4965.62
512292.5793.6665.6864.77
128392.3393.2564.8365.55
256392.7593.5265.4965.16
512392.8692.9365.4964.84
", + "image_path": "49d39c211f932285f84af6ee1a1994958231c389e7acc6d64fca7c3fa19881f6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 286, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 556 + ], + "type": "text", + "content": "disabled for the pretraining. In Table 8 we show that having both loss terms is better than the individual loss terms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "text", + "content": "Masking Ratio: Masking ratio is an important hyperparameter and depends on the data modality, for example, BERT [12] uses " + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "text", + "content": ", MSN [3] uses " + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "text", + "content": " (for ViT-Base), MAE [25] uses " + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "text", + "content": ", and VideoMAE [54] uses 90 to " + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "text", + "content": " masking. Since our clip representations are somewhat abstract representations of the video, we expect the optimal masking ratio to be close to NLP models rather than video MAEs. We have observed a steady decrease in the pretraining task's performance with higher masking ratios, so we only tested low masking ratios in Table 9 and found out that " + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 557, + 286, + 689 + ], + "type": "text", + "content": " is the optimal masking ratio." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Transformer Capacity: We also explore the number of" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 113, + 547, + 257 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 545, + 112 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 545, + 112 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 545, + 112 + ], + "type": "text", + "content": "Table 11. Number of Views. SCALE ft accuracy with different numbers of clips and batch sizes. More views lead to consistent improvement, and large batch sizes are not necessary because of the hard negative samples." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 113, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 306, + 113, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 306, + 113, + 547, + 257 + ], + "type": "table", + "html": "
Num ViewsBatch SizeUCF-101HMDB-51
SVTρBYOLSVTρBYOL
4 × 225692.6592.8364.3564.24
6 × 225692.7093.0764.5764.37
8 × 225692.8093.4964.6464.63
4 × 251292.6793.4964.8564.50
6 × 251293.1893.6864.9065.35
8 × 251293.2093.8165.4965.62
4 × 2102492.7593.3664.7765.15
6 × 2102492.9693.5764.9665.48
8 × 2102493.07OOM65.29OOM
", + "image_path": "55940f158f1307601fe757b30501859598002bec76735fcf3611c12f2734b4dc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 283, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 344 + ], + "type": "text", + "content": "transformer layers and their hidden size in Table 10. We can see that having more than one transformer layer is necessary for good results and too few hidden channels can hurt performance. However, there is a trade-off, and deeper transformers can lead to worse performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 346, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 418 + ], + "type": "text", + "content": "Number of Views: Finally, we studied the model performance as we changed the number of views and batch size fed to the model. As can be seen in Table 11, having more views has a large and consistent impact on the performance, and since we have hard negatives for contrastive loss within the video, we are not too reliant on large batch sizes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 438, + 379, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 438, + 379, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 438, + 379, + 451 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 715 + ], + "type": "text", + "content": "In this paper, we introduced SCALE, a framework for video representation learning that aggregates the information from multiple clips from the same video. We combine contrastive learning and masked modeling with intuitions from predictive coding to obtain improved global and local representations of clips starting from frozen backbones. We evaluated these features using a wide array of backbones on different action classification and video understanding datasets and achieved strong or state-of-the-art results. The computational efficiency of our method is extremely useful for videos and opens the possibility to a wider group of researchers (with limited computational resources) to work on video representation learning than previously possible. We also believe that working with a set of clips is an interesting direction for representation learning. Finally, as a surprising and maybe alarming observation, even contrastive representations that were trained to be invariant to data augmentations and spatio-temporal crops can be used for contrastive masked modeling. This might be due to benign memorization [2], and understanding why this phenomenon happens might lead to a better understanding of contrastive learning." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5671" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 126 + ], + "type": "text", + "content": "This work was supported by grant 200020_188690 of the Swiss National Science Foundation (SNSF) and an Adobe award." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 107, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 137, + 107, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 107, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 157, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 157, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 157, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 157, + 287, + 201 + ], + "type": "text", + "content": "[1] Hassan Akbari, Linagzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. ArXiv, abs/2104.11178, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 235 + ], + "type": "text", + "content": "[2] Sotiris Anagnostidis, Gregor Bachmann, Lorenzo Noci, and Thomas Hofmann. The curious case of benign memorization. ArXiv, abs/2210.14019, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 290 + ], + "type": "text", + "content": "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In European Conference on Computer Vision, pages 456-473. Springer, 2022. 1, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 367 + ], + "type": "text", + "content": "[4] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zach Beaver, Jana von Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, Vivek Natarajan, and Mohammad Norouzi. Big self-supervised models advance medical image classification. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3458-3468, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 390 + ], + "type": "text", + "content": "[5] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. ArXiv, abs/2106.08254, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 445 + ], + "type": "text", + "content": "[6] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9922-9931, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 545 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 547, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 547, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 547, + 287, + 590 + ], + "type": "text", + "content": "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "text", + "content": "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3075-3084, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[11] Ishan Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TCLR: Temporal contrastive learning for video representation. Computer Vision and Image Understanding, 219:103406, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3, 4, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 139 + ], + "type": "text", + "content": "[13] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. ArXiv, abs/2207.07116, 2022. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 140, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 205 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "type": "text", + "content": "[15] Haoqi Fan, Yanghao Li, Bo Xiong, Wan-Yen Lo, and Christoph Feichtenhofer. Pyslowfast. https://github.com/facebookresearch/slowfast, 2020.5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "text", + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "text", + "content": "[17] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6201-6210, 2019. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "type": "text", + "content": "[18] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "text", + "content": "[19] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimae: Single model masked pretraining on images and videos. arXiv preprint arXiv:2206.08356, 2022. 1, 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 458 + ], + "type": "text", + "content": "[20] Rohit Girdhar and Kristen Grauman. Anticipative video transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13505-13515, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[21] Shreyank N Gowda, Marcus Rohrbach, and Laura Sevilla-Lara. Smart frame selection for action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1451-1459, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 504, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 580 + ], + "type": "text", + "content": "[22] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 581, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 647 + ], + "type": "text", + "content": "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[24] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in Neural Information Processing Systems, 33:5679-5690, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5672" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "text", + "content": "[26] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 205 + ], + "type": "text", + "content": "[27] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. ArXiv, abs/1808.06670, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 259 + ], + "type": "text", + "content": "[28] Kaiqin Hu, Jie Shao, Yuan Liu, Bhiksha Raj, Marios Savvides, and Zhiqiang Shen. Contrast and order representations for video self-supervised learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7919-7929, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 327 + ], + "type": "text", + "content": "[29] Deng Huang, Wenhao Wu, Weiwen Hu, Xu Liu, Dongliang He, Zhihua Wu, Xiangmiao Wu, Mingkui Tan, and Errui Ding. Ascnet: Self-supervised video representation learning with appearance-speed consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8096-8105, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "text", + "content": "[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. PMLR, 2015. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "text", + "content": "[31] Md. Mohaiminul Islam and Gedas Bertasius. Long movie clip classification with state-space video models. ArXiv, abs/2204.01692, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "text", + "content": "[32] Simon Jenni and Hailin Jin. Time-equivariant contrastive video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9970-9980, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[33] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "text", + "content": "[34] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 2, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "type": "text", + "content": "[35] Xingbin Liu, Jinghao Zhou, Tao Kong, Xianming Lin, and Rongrong Ji. Exploring target representations for masked autoencoders. arXiv preprint arXiv:2209.03917, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[36] Yue Liu, Junqi Ma, Yufei Xie, Xuefeng Yang, Xingzhen Tao, Lin Peng, and Wei Gao. Contrastive predictive coding with transformer for video representation learning. Neurocomputing, 482:154-162, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 690 + ], + "type": "text", + "content": "[37] Guillaume Lorre, Jaxonary Rabarisoa, Astrid Orcesi, Samia Ainouz, and Stephane Canu. Temporal contrastive pretraining for video action recognition. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 662-670, 2020. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[38] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "verification. In European conference on computer vision, pages 527-544. Springer, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 98, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 98, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 98, + 545, + 150 + ], + "type": "text", + "content": "[39] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12475-12486, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 154, + 545, + 198 + ], + "type": "text", + "content": "[40] Charlie Nash, João Carreira, Jacob Walker, Iain Barr, Andrew Jaegle, Mateusz Malinowski, and Peter Battaglia. Transframer: Arbitrary frame prediction with generative models. arXiv preprint arXiv:2203.09494, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 200, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 200, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 200, + 545, + 233 + ], + "type": "text", + "content": "[41] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 235, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 235, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 235, + 545, + 300 + ], + "type": "text", + "content": "[42] Mandela Patrick, Yuki M. Asano, Bernie Huang, Ishan Misra, Florian Metze, João F. Henriques, and Andrea Vedaldi. Space-time crop & attend: Improving cross-modal video representation learning. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10540-10552, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 303, + 545, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 367 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 371, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 425 + ], + "type": "text", + "content": "[44] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2874-2884, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 428, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 428, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 428, + 545, + 494 + ], + "type": "text", + "content": "[45] Adria Recasens, Pauline Luc, Jean-Baptiste Alayrac, Luyu Wang, Florian Strub, Corentin Tallec, Mateusz Malinowski, Viorica Pătrăucean, Florent Altché, Michal Valko, et al. Broaden your views for self-supervised video learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1255-1265, 2021. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 496, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 496, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 496, + 545, + 529 + ], + "type": "text", + "content": "[46] Joshua Robinson, Ching-Yao Chuang, Suvrit Sra, and Stefanie Jegelka. Contrastive learning with hard negative samples. ArXiv, abs/2010.04592, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 531, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 575 + ], + "type": "text", + "content": "[47] Fadime Sener, Dipika Singhania, and Angela Yao. Temporal aggregate representations for long-range video understanding. In European Conference on Computer Vision, pages 154-171. Springer, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 610 + ], + "type": "text", + "content": "[48] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 2, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "type": "text", + "content": "[49] Chen Sun, Arsha Nagrani, Yonglong Tian, and Cordelia Schmid. Composable augmentation encoding for video representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8834-8844, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[50] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12607-12617, 2021. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5673" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[51] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. Vimpac: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 160 + ], + "type": "text", + "content": "[52] Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. arXiv preprint arXiv:2206.01204, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 227 + ], + "type": "text", + "content": "[53] Nenad Tomasev, Ioana Bica, Brian McWilliams, Lars Holger Buesing, Razvan Pascanu, Charles Blundell, and Jovana Mitrovic. Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels on imageNet? In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 270 + ], + "type": "text", + "content": "[54] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. 1, 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 315 + ], + "type": "text", + "content": "[55] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "text", + "content": "[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "type": "text", + "content": "[57] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14010-14020, 2022. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 468 + ], + "type": "text", + "content": "[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:2740-2755, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "text", + "content": "[59] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14713-14723, 2022. 1, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 526, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 569 + ], + "type": "text", + "content": "[60] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "text", + "content": "[61] Chen Wei, Haoqi Fan, Saining Xie, Chaoxia Wu, Alan Loddon Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14648-14658, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[62] Donglai Wei, Joseph J Lim, Andrew Zisserman, and William T Freeman. Learning and using the arrow of time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8052-8060, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[63] Chaoxia Wu and Philipp Krahenbuhl. Towards long-form video understanding. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1884-1894, 2021. 2, 5, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 552 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[64] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "type": "text", + "content": "[65] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 196, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 546, + 239 + ], + "type": "text", + "content": "[66] Yu Wu, Linchao Zhu, Xiaohan Wang, Yi Yang, and Fei Wu. Learning to anticipate egocentric actions by imagination. IEEE Transactions on Image Processing, 30:1143-1152, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 240, + 546, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 546, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 546, + 295 + ], + "type": "text", + "content": "[67] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3333-3343, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 297, + 546, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 361 + ], + "type": "text", + "content": "[68] Liangzhe Yuan, Rui Qian, Yin Cui, Boqing Gong, Florian Schroff, Ming-Hsuan Yang, Hartwig Adam, and Ting Liu. Contextualized spatio-temporal contrastive learning with self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13977-13986, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 363, + 546, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 546, + 428 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 546, + 428 + ], + "type": "text", + "content": "[69] Yujia Zhang, Lai-Man Po, Xuyuan Xu, Mengyang Liu, Yexin Wang, Weifeng Ou, Yuzhi Zhao, and Wing-Yin Yu. Contrastive spatio-temporal pretext learning for self-supervised video representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 3380-3389, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 430, + 546, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 546, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 546, + 463 + ], + "type": "text", + "content": "[70] Zhang Zhang and Dacheng Tao. Slow feature analysis for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 34(3):436-450, 2012. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 464, + 546, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 546, + 506 + ], + "type": "text", + "content": "[71] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 508, + 546, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 546, + 552 + ], + "type": "text", + "content": "[72] Pan Zhou, Yichen Zhou, Chenyang Si, Weihao Yu, Teck Khim Ng, and Shuicheng Yan. Mugs: A multi-granular self-supervised learning framework. arXiv preprint arXiv:2203.14415, 2022. 1" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5674" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/fe9d96d8-4150-4b6d-a27c-f4b80a133514_content_list.json b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/fe9d96d8-4150-4b6d-a27c-f4b80a133514_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6eb054a162eeb0ece9bcb280316367b6b1b1b501 --- /dev/null +++ b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/fe9d96d8-4150-4b6d-a27c-f4b80a133514_content_list.json @@ -0,0 +1,1537 @@ +[ + { + "type": "text", + "text": "Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception", + "text_level": 1, + "bbox": [ + 88, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kun Yang $^{1\\dagger}$ Dingkang Yang $^{1\\dagger}$ Jingyu Zhang $^{1}$ Mingcheng Li $^{1}$ Yang Liu $^{1}$ Jing Liu $^{1}$ Hanqi Wang $^{1}$ Peng Sun $^{2}$ Liang Song $^{1*}$ $^{1}$ Academy for Engineering and Technology, Fudan University $^{2}$ Duke Kunshan University {kunyang20,dkyang20,songl}@fudan.edu.cn", + "bbox": [ + 120, + 178, + 848, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-agent collaborative perception as a potential application for vehicle-to-everything communication could significantly improve the perception performance of autonomous vehicles over single-agent perception. However, several challenges remain in achieving pragmatic information sharing in this emerging research. In this paper, we propose SCOPE, a novel collaborative perception framework that aggregates the spatio-temporal awareness characteristics across on-road agents in an end-to-end manner. Specifically, SCOPE has three distinct strengths: i) it considers effective semantic cues of the temporal context to enhance current representations of the target agent; ii) it aggregates perceptually critical spatial information from heterogeneous agents and overcomes localization errors via multi-scale feature interactions; iii) it integrates multi-source representations of the target agent based on their complementary contributions by an adaptive fusion paradigm. To thoroughly evaluate SCOPE, we consider both real-world and simulated scenarios of collaborative 3D object detection tasks on three datasets. Extensive experiments show the superiority of our approach and the necessity of the proposed components. The project link is https://ydk122024.github.io/SCOPE/.", + "bbox": [ + 75, + 316, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 695, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Perceiving complex driving environments is essential for ensuring road safety [23] and traffic efficiency [21] in autonomous driving. Despite the rapid development in the perception ability of intelligent vehicles through deep learning technologies [5, 6, 14, 41, 43], the conventional single-agent perception paradigm remains challenging. Single-agent perception system is generally vulnerable and defective due to the inevitable practical difficulties on the road,", + "bbox": [ + 75, + 720, + 468, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such as restricted detection range and occlusion [47, 48]. To alleviate the inadequate observation constraints of isolated agents, multi-agent collaborative perception has emerged as a promising solution in recent years [17, 29]. With Vehicle-to-Everything (V2X) oriented communication applications, collaborative perception effectively promotes perspective complementation and information sharing among agents.", + "bbox": [ + 500, + 287, + 892, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Based on recent LiDAR-based 3D collaborative perception datasets [35, 36, 37, 46], several impressive works have achieved a trade-off between collaborative perception performance and communication bandwidth of Autonomous Vehicles (AVs) via well-designed mechanisms. These remarkable efforts include knowledge distillation-based intermediate collaboration [17], handshake communication mechanism [20], V2X-visual transformer [36], and spatial information selection strategy [10]. However, there is still considerable room for progress in achieving a pragmatic and effective collaborative perception. Concretely, existing methods invariably follow a single-frame static perception pattern, suffering from the data sparsity shortcoming of 3D point clouds [47] and ignoring meaningful semantic cues in the temporal context [49]. The insufficient representation of the moving objects in the current frame potentially restricts the perception performance of target vehicles.", + "bbox": [ + 496, + 397, + 893, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Moreover, spatial information aggregation in collaborative perception systems has exposed several problems. For the collaborator-shared message fusion, the per-agent/location-based fusion strategies of previous attempts [16, 17, 20, 28, 36] fail to deal with feature map misalignment from heterogeneous agents due to localization errors. Consequently, misleading features from collaborators could lead to object position misjudgment of the ego vehicle (i.e., receiver) and harm its detection precision. For the ego vehicle message refinement, existing approaches [10, 17, 20, 21, 28, 33] rely on fused representations with collaborator information to implement detection, abandoning the natural perception advantages of the ego vehicle and introducing the potential noises. The ego-centered characteristics may contain locally critical spatial information that is not disturbed by noise from collaborative agents", + "bbox": [ + 496, + 657, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ Equal contributions in no particular order. The work was done when Dingkang Yang was at the Institute of Meta-Medical, IPASS. The two first authors thank Runsheng Xu for providing constructive suggestions. *Corresponding author.", + "bbox": [ + 75, + 849, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23383", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/367b6d237d71ba4996473067053e82e0dd9db99ac2d68dbea701532133eceebe.jpg", + "image_caption": [ + "Figure 1. The overall architecture of the proposed SCOPE. The framework consists of five parts: metadata conversion and feature extraction, context-aware information aggregation, confidence-aware cross-agent collaboration, importance-aware adaptive fusion, and detection decoders. The details of each individual component are illustrated in Section 3." + ], + "image_footnote": [], + "bbox": [ + 125, + 87, + 851, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with asynchronous sensor measurements. To this end, how to effectively break the aforementioned limitations becomes a priority for achieving robust collaborative perception.", + "bbox": [ + 75, + 420, + 470, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above observations, we present SCOPE, a Spatio-temporal domain awareness multi-agent CollabOrative PErception approach to tackle existing challenges jointly. From Figure 1, SCOPE achieves effective information collaboration among agents and precise perception of the ego agent via three proposed core components. Specifically, (i) for the data sparsity challenge in single-frame point clouds, the context-aware information aggregation is proposed to aggregate the ego agent's context information from preceding frames. We employ selective information filtering and spatio-temporal feature integration modules to capture informative temporal cues and historical context semantics. (ii) For the collaborator-shared message fusion challenge, the confidence-aware cross-agent collaboration is introduced to guarantee that the ego agent aggregates complementary information from heterogeneous agents. Based on the confidence-aware multi-scale feature interaction, we facilitate holistic communication and mitigate feature map misalignment due to localization errors of collaborators. (iii) For the multi-feature integration challenge of the ego agent, the importance-aware adaptive fusion is designed to flexibly incorporate the advantages of diverse representations based on distinctive perception contributions. We evaluate the superiority of SCOPE on various collaborative 3D object detection datasets, including DAIR-V2X [46], OPV2V [37], and V2XSet [36]. Extensive experiments in real-world and simulated scenarios provide favorable evidence that our approach is competitive in", + "bbox": [ + 75, + 478, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multi-agent collaborative perception tasks.", + "bbox": [ + 500, + 420, + 782, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions can be summarized as follows:", + "bbox": [ + 517, + 435, + 880, + 450 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present SCOPE, a novel framework for multi-agent collaborative perception. The framework facilitates information collaboration and feature fusion among agents, achieving a reasonable performance-bandwidth trade-off. Comprehensive experiments on collaborative detection tasks show that SCOPE outperforms previous state-of-the-art methods.", + "- To our best knowledge, we are the first to consider the temporal context of the ego agent in collaborative perception systems. Based on the proposed context-aware component, the target frame of point clouds efficiently integrates historical cues from preceding frames to capture valuable temporal information.", + "- We introduce two spatial information aggregation components to address the challenges of collaboration heterogeneity and fused representation unicity. These tailored components effectively perform multigrained information interactions among agents and multi-source feature refinement of the ego agent." + ], + "bbox": [ + 517, + 459, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 777, + 640, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-Agent Communication", + "text_level": 1, + "bbox": [ + 500, + 801, + 759, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benefiting from rapid advances in learning-based technologies [18, 38, 39, 40, 42, 44, 45], communication has become an essential component in constructing robust multiagent systems. Previous works [7, 25, 26, 27] typically focus on centralized paradigms or predefined protocols to", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23384", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "explore information-sharing mechanisms among multiple agents. However, these efforts are potentially domain-constrained and scenario-isolated. Recent methods [3, 9, 12] achieve effective communication strategies in diverse scenarios. For instance, Vain [9] determines which agents would share information by modeling the locality of interactions. ATOC [12] introduces an attentional communication strategy to learn when communicating and integrating useful information to execute collaborative decisions. In comparison, we focus on the more challenging 3D perception tasks in complex driving scenarios. Based on the proposed confidence-aware multi-scale information interaction, SCOPE achieves an information-critical and holistic communication pattern among heterogeneous agents.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Collaborative Perception", + "text_level": 1, + "bbox": [ + 76, + 310, + 305, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Collaborative perception investigates how to aggregate complementary perception semantics across connected agents to improve overall system performance. Current works are mainly categorized as early [2], intermediate [1, 10, 17, 20, 28, 34, 36, 37], and late [24, 31] fusion, where intermediate fusion is more favoured due to its good trade-off between performance and transmission bandwidth. In this case, several remarkable works explore efficient collaboration mechanisms. In DiscoNet [17], the teacher-student network is trained to aggregate the shared information. V2X-ViT [36] proposes heterogeneous multi-agent self-attention to handle device heterogeneity, and Xu et al. [34] design an axial attention module to capture sparse spatial semantics. Furthermore, Where2comm [10] filters irrelevant semantics by quantifying the spatial importance through confidence maps. Unlike these single-frame methods, we propose an information aggregation component to extract rich context information from local historical observations. Moreover, our cross-agent collaboration mechanism outperforms the previous per-agent/location message fusion models, significantly mitigating the impact of localization errors.", + "bbox": [ + 75, + 333, + 472, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 662, + 212, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we focus on developing a robust collaborative perception system to improve the ego vehicle's perception capability. Figure 1 illustrates the overall architecture of the proposed framework. It consists of five sequential parts, where the system input accommodates different types of connected agents (i.e., infrastructures and AVs).", + "bbox": [ + 75, + 688, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Metadata Conversion and Feature Extraction", + "text_level": 1, + "bbox": [ + 76, + 786, + 464, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the collaboration, one of the connected agents is identified as the ego agent $(i)$ to build a communication graph, where other agents within the communication range serve as collaborators. Upon receiving the metadata $(e.g.,$ poses and extrinsics) broadcast by the ego agent, the collaborators project their local LiDAR point clouds", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/39b6fb7b98a911884f35aee76d423898913923fe38a44a77a2bfd7a535440964.jpg", + "image_caption": [ + "Figure 2. (a) The architecture of the proposed CIA component, including selective information filtering (left) and spatio-temporal feature integration (middle). (b) The multi-scale convolutional structure in the pyramid LSTM." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 764, + 226 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9c047eff5af653d2c89c4ebdc503f8f9658027d248b9a5076e7837d6b585fae4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 90, + 888, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to the coordinate system of the ego agent. Similarly, the ego agent's previous point cloud frames are synchronized to the current frame. Each agent $k \\in \\{1, \\dots, K\\}$ encodes the projected point clouds into Bird's Eye View (BEV) features to extract local visual representations. Given the point cloud $\\mathbf{X}_k^{(t)}$ of $k$ -th agent at timestamp $t$ , the extracted feature is $\\mathbf{F}_k^{(t)} = f_{enc}(\\mathbf{X}_k^{(t)}) \\in \\mathbb{R}^{C \\times H \\times W}$ , where $f_{enc}(\\cdot)$ is the PointPillar [14] encoder shared among all agents, and $C, H, W$ stand for the channel, height, and width.", + "bbox": [ + 496, + 297, + 890, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Context-aware Information Aggregation", + "text_level": 1, + "bbox": [ + 498, + 446, + 846, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conventional 3D detection tasks [47, 49] have shown the potential of temporal information in improving perception performance. Unfortunately, current multi-agent collaborative efforts [10, 17, 20, 21, 28] invariably focus on exploring the current frame while ignoring context cues from previous frames. The single-frame solution fails to effectively detect fast-moving objects (e.g., surrounding vehicles) due to the sparsity and insufficiency of point clouds. To this end, we present a Context-aware Information Aggregation (CIA) component to capture spatio-temporal representations of the ego agent's previous frames to fuse valuable semantics. As shown in Figure 2(a), CIA consists of two core phases.", + "bbox": [ + 496, + 469, + 890, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Selective Information Filtering. This phase aims to extract refined information by filtering redundant features from the target $\\pmb{F}_i^{(t)} \\in \\mathbb{R}^{C \\times H \\times W}$ and distilling meaningful characteristics from the previous $\\pmb{F}_i^{(t - \\tau)} \\in \\mathbb{R}^{\\tau \\times C \\times H \\times W}$ , where $\\tau$ is the time shift. Concretely, the channel-wise average and max pooling operations $\\Psi_{a / m}(\\cdot)$ are first utilized to aggregate rich channel semantics, denoted as $\\mathcal{M}_{a / m}^{(t)} = \\Psi_{a / m}(\\pmb{F}_i^{(t)})$ and $\\mathcal{M}_{a / m}^{(\\tau)} = \\Psi_{a / m}(\\pmb{F}_i^{(\\tau)})$ , where $\\pmb{F}_i^{(\\tau)} = \\sum_{n=1}^{\\tau} \\pmb{F}_i^{(t-n)}$ . These refined feature maps are combined to obtain a well-chosen spatial selection map. The procedure is expressed as follows:", + "bbox": [ + 496, + 651, + 893, + 837 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {U} = \\sigma \\cdot \\aleph (\\mathcal {M} _ {a} ^ {(t)} \\| \\mathcal {M} _ {m} ^ {(t)} + \\mathcal {M} _ {a} ^ {(\\tau)} \\| \\mathcal {M} _ {m} ^ {(\\tau)}) \\in \\mathbb {R} ^ {H \\times W}, \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 843, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma$ is the sigmoid activation, $\\aleph$ denotes a convolutional layer for fusion, and $\\parallel$ denotes the concatenation op", + "bbox": [ + 496, + 869, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23385", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/63e56030fa655eb97aab668dd2b6ed5c5b81ea5e717f094b6119abf2b0feae59.jpg", + "image_caption": [ + "Figure 3. (a) The architecture of the proposed CCC component. (b) CCC consists of two phases at each scale: reference point proposal (middle) and deformable cross-attention module (right)." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 467, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "eration. Immediately, the information filtering strategy is implemented via pixel-wise attention weighting to enhance the salient historical features of $H_{i}^{(\\tau)} \\in \\mathbb{R}^{\\tau \\times C \\times H \\times W}$ :", + "bbox": [ + 75, + 316, + 470, + 367 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {H} _ {i} ^ {(\\tau)} = (1 - \\mathcal {U}) \\odot \\tanh \\left(\\boldsymbol {F} _ {i} ^ {(t)}\\right) + \\mathcal {U} \\odot \\boldsymbol {F} _ {i} ^ {(t - \\tau)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 376, + 468, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatio-temporal Feature Integration. In order to integrate the historical prototype to improve the perception ability of the current representation, we introduce a pyramid LSTM [15] to learn the contextual dependencies of interframe features $\\pmb{F}_i^{(t)}\\| \\pmb{H}_i^{(\\tau)}$ . The difference between the pyramid LSTM and vanilla LSTM [8] is that the initial matrix multiplication is replaced with a multi-scale convolutional structure, as shown in Figure 2(b). Note that the intrinsic long short-term memory mechanism in the LSTM is adopted, and please refer to [8] for details. In practice, multi-scale spatial features are extracted by two successive 2D convolutions among different scales, followed by batch normalization and ReLU activation. To achieve multilevel spatial semantic fusion, the downsampled features are progressively interpolated to deliver to the upsampled layers via lateral connections. The overall design facilitates learning spatio-temporal correlations in historical contexts, which benefits our multi-frame perception paradigm. Ultimately, the final hidden state in the pyramid LSTM is used as the refined context-aware feature $\\pmb{H}_i^{(t)}\\in \\mathbb{R}^{C\\times H\\times W}$", + "bbox": [ + 75, + 406, + 468, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Confidence-aware Cross-agent Collaboration", + "text_level": 1, + "bbox": [ + 76, + 720, + 459, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cross-agent collaboration targets to enhance the ego agent's visual representations by aggregating the complementary semantics from the collaborator-shared features. Existing works [10, 11, 36, 37] propose attention-based perlocation feature fusion methods, which are vulnerable to localization errors and ignore the sparsity of point clouds. To tackle these issues, we implement a novel Confidence-aware Cross-agent Collaboration (CCC) component. We first obtain the confidence map $S_{k}^{(t)}$ and binary selection matrix $\\mathcal{C}_k^{(t)}$ of feature $F_{k}^{(t)}, k\\in \\{1,\\dots,K\\}$ following [10].", + "bbox": [ + 75, + 744, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The filtered feature shared by $k$ -th agent to ego agent is $\\pmb{F}_k^{(t)} = \\pmb{F}_k^{(t)} \\odot \\mathcal{C}_k^{(t)}$ . As Figure 3(a) shows, the CCC component encodes the features and confidence maps into three scales and performs feature fusion at each scale. $\\pmb{F}_{k,l}^{(t)}$ and $S_{k,l}^{(t)}$ denote the feature and confidence map at $l$ -th scale. We take two agents as an example to elaborate on the feature fusion in Figure 3(b), including the following two phases.", + "bbox": [ + 498, + 90, + 890, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reference Point Proposal. To predict the spatial locations containing meaningful object information, we obtain the element-wise summation of all confidence maps as $S_{sum,l}^{(t)} = \\sum_{k=1}^{K} S_{k,l}^{(t)}$ . Since the confidence map reflects the spatially critical levels, $S_{sum,l}^{(t)}$ presents the potential positions of the targets within the detection range, called the reference points. Accordingly, we apply a threshold-based selection function $f_{sel}(\\cdot)$ to extract the reference points $S_{re,l}^{(t)}$ . This design could actively guide the subsequent fusion network to concentrate on the essential spatial areas.", + "bbox": [ + 498, + 205, + 892, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deformable Cross-attention Module. We extract the ego agent's feature $\\pmb{F}_{i,l}^{(t)}$ at the reference points as the initial query embedding and apply a linear layer to encode the locations of reference points into a position embedding. To resolve the misalignment of feature maps and acquire more robust representations against localization errors, the Deformable Cross-attention Module (DCM) aggregates the information from sampled keypoints through a deformable cross-attention layer [30]. Concretely, a linear layer is leveraged to learn the 2D spatial offset $\\Delta q$ of the reference point $q$ for sampling attending keypoints at $q + \\Delta q$ . We simultaneously select $M$ keypoints and extract their features as the attending features. The output of DCM at position $q$ is:", + "bbox": [ + 498, + 367, + 892, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {D C M} (q) = \\sum_ {a = 1} ^ {A} \\boldsymbol {W} _ {a} \\left[ \\sum_ {k = 1} ^ {K} \\sum_ {m = 1} ^ {M} \\phi \\left(\\boldsymbol {W} _ {b} \\boldsymbol {F} _ {i, l} ^ {(t)} (q)\\right) \\boldsymbol {F} _ {k, l} ^ {(t)} \\left(q + \\Delta q _ {m}\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 571, + 890, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $a$ indexes the attention head, $\\mathbf{W}_{a / b}$ are the learnable weights, and $\\phi (\\cdot)$ is the softmax function. Ultimately, we derive a filling operation that fills $\\mathrm{DCM}(q)$ into the ego agent's feature $\\pmb{F}_{i,l}^{(t)}$ based on the initial position $q$ and outputs $\\pmb{Z}_{i,l}^{(t)}$ . The output-enhanced features under the three scales are encoded into the same size and concatenated in the channel dimension. We utilize a $1\\times 1$ convolutional layer to fuse information of the three scales and get the final collaboration feature $\\pmb{Z}_i^{(t)}\\in \\mathbb{R}^{C\\times H\\times W}$ .", + "bbox": [ + 498, + 614, + 890, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Importance-aware Adaptive Fusion", + "text_level": 1, + "bbox": [ + 500, + 770, + 808, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although previous works [17, 20, 21, 28, 36] have achieved impressive performance through ego features that aggregate collaborative information, they potentially suffer from noise interference introduced by collaborators with asynchronous measurements. A promising solution is to consider purely ego-centered characteristics, which contain the natural perception advantages of the target agent. To", + "bbox": [ + 498, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23386", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26dbb363aa27207230655e7bc0098c7122196a976e6b1a158ff9d7e8cf002baa.jpg", + "image_caption": [ + "Figure 4. The architecture of the proposed IAF component." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 464, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "this end, we propose an Importance-aware Adaptive Fusion (IAF) component to fuse multi-source features based on their complementary contributions. These informative features include $\\pmb{H}_i^{(t)}, \\pmb{Z}_i^{(t)}$ , and $\\pmb{F}_i^{(t)}$ . From Figure 4, the importance generator $f_{gen}(\\cdot)$ is used to generate the spatial importance maps, which processes as follows:", + "bbox": [ + 75, + 296, + 468, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {i} ^ {(t)}, \\mathcal {Z} _ {i} ^ {(t)}, \\mathcal {S} _ {i} ^ {(t)} = \\sigma \\cdot \\Psi_ {m} (f _ {g e n} (\\boldsymbol {H} _ {i} ^ {(t)}, \\boldsymbol {Z} _ {i} ^ {(t)}, \\boldsymbol {F} _ {i} ^ {(t)})), \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 393, + 468, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{H}_i^{(t)},\\mathcal{Z}_i^{(t)},S_i^{(t)}\\in [0,1]^{H\\times W}$ . Then, these importance maps are concatenated to get the normalized attention maps across distinct features by the softmax function $\\phi (\\cdot)$ :", + "bbox": [ + 76, + 420, + 468, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\mathcal {H}} ^ {(t)} = \\phi \\left(\\mathcal {H} _ {i} ^ {(t)}\\right) = \\frac {\\exp \\left(\\mathcal {H} _ {i} ^ {(t)}\\right)}{\\exp \\left(\\mathcal {H} _ {i} ^ {(t)}\\right) + \\exp \\left(\\mathcal {Z} _ {i} ^ {(t)}\\right) + \\exp \\left(\\mathcal {S} _ {i} ^ {(t)}\\right)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 474, + 468, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A larger value in $\\mathcal{E}_{\\mathcal{H}}^{(t)} \\in \\mathbb{R}^{H \\times W}$ implies that the corresponding feature position is more significant. Similarly, $\\mathcal{E}_{\\mathcal{Z}}^{(t)} = \\phi(\\mathcal{Z}_i^{(t)})$ and $\\mathcal{E}_S^{(t)} = \\phi(S_i^{(t)})$ . Finally, we adaptively activate each feature map's perceptually critical spatial information by a weighted summation. The fused feature map $\\mathcal{F}_i^{(t)} \\in \\mathbb{R}^{C \\times H \\times W}$ is obtained as follows:", + "bbox": [ + 76, + 515, + 468, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {i} ^ {(t)} = \\mathcal {E} _ {\\mathcal {H}} ^ {(t)} \\odot \\boldsymbol {H} _ {i} ^ {(t)} + \\mathcal {E} _ {\\mathcal {Z}} ^ {(t)} \\odot \\boldsymbol {Z} _ {i} ^ {(t)} + \\mathcal {E} _ {\\mathcal {S}} ^ {(t)} \\odot \\boldsymbol {F} _ {i} ^ {(t)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 618, + 468, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Detection Decoder and Objective Optimization", + "text_level": 1, + "bbox": [ + 76, + 643, + 468, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We leverage two detection decoders to decode the final fused feature $\\mathcal{F}_i^{(t)}$ into the predicted outputs. The regression output is $\\mathcal{Y}_{i,r}^{(t)} = f_{dec}^r (\\mathcal{F}_i^{(t)})\\in \\mathbb{R}^{7\\times H\\times W}$ , representing the position, size, and yaw angle of the bounding box at each location. The classification output reveals the confidence value of each predefined box to be a target or background, which is $\\mathcal{Y}_{i,c}^{(t)} = f_{dec}^c (\\mathcal{F}_i^{(t)})\\in \\mathbb{R}^{2\\times H\\times W}$ . Following existing work [14], we adopt the smooth $L_{1}$ loss for regressing the bounding boxes and a focal loss [19] for classification.", + "bbox": [ + 75, + 667, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 821, + 209, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 847, + 362, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct comprehensive experiments on three benchmark datasets for collaborative perception tasks, in", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "including DAIR-V2X [46], V2XSet [36], and OPV2V [37]. These datasets provide annotated sensor data from multiple connected agents for each sample. Concretely, DAIR-V2X is a large-scale real-world dataset supporting collaborative 3D object detection, containing 71,254 samples and dividing the training/validation/testing sets as 5:2:3. Each sample contains the LiDAR point clouds from a vehicle and an infrastructure. During feature extraction, we transform the point cloud into the BEV feature map of the size (200, 504, 64). V2XSet is a public simulation dataset for V2X collaborative perception, including 73 representative scenarios and 11,447 frames of labeled point clouds generated by Carla [4]. The training/validation/testing sets are 6,694, 1,920, and 2,833 frames, respectively. The size of each extracted feature map is (100, 384, 64). OPV2V is a vehicle-to-vehicle collaborative perception dataset collected by Carla and OpenCDA [32]. It includes 11,464 frames of point clouds and RGB images with 3D annotation, splitting into training/validation/testing sets of 6,764, 1,981, and 2,719 frames, respectively. The 3D point cloud is extracted as a feature map of the size (100, 384, 64).", + "bbox": [ + 496, + 90, + 890, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following [36], we leverage the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the 3D object detection performance. The communication results count the message size by byte in the log scale with base 2 [10].", + "bbox": [ + 496, + 407, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details.", + "text_level": 1, + "bbox": [ + 500, + 494, + 720, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We build the proposed models on the Pytorch toolbox [22] and train them on four Nvidia Tesla V100 GPUs using the Adam optimizer [13]. The training settings of the DAIR-V2X, V2XSet, and OPV2V datasets follow: the batch sizes are $\\{5, 3, 3\\}$ and the epochs are $\\{30, 40, 40\\}$ . The initial learning rate is $2\\mathrm{e} - 3$ and decays every 15 epochs with a factor of 0.1. Our CIA module uses the previous 2 frames, the attention head in the CCC component is 8, and the number of keypoints is 15. To simulate the default localization and heading errors, we add Gaussian noise with a standard deviation of $0.2\\mathrm{m}$ and $0.2^{\\circ}$ to the initial position and orientation of the agents. The voxel resolution of the observation encoder $f_{enc}(\\cdot)$ is set to $0.4\\mathrm{m}$ in height and width. The generator $f_{gen}(\\cdot)$ follows a detection decoder from [14], and the summed confidence map $S_{sum,l}^{(t)}$ is filtered using a threshold of 0.1 in the function $f_{sel}(\\cdot)$ . The decoders $\\{f_{dec}^{r}(\\cdot), f_{dec}^{c}(\\cdot)\\}$ are derived with two $1 \\times 1$ convolutional layers. We provide results on the validation set of DAIR-V2X and the testing set of V2XSet and OPV2V.", + "bbox": [ + 496, + 516, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 816, + 720, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detection Performance Comparison. Table 1 shows the 3D detection performance comparison results on the three datasets. We compare the following methods. No Fusion as a baseline considers only the point cloud data of ego vehi", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23387", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a89cbb5d8c3de522e8f7b24f4d4ba2745219562dbc5ecd2f1603891c0af0fc05.jpg", + "image_caption": [ + "Figure 5. Collaborative perception performance comparison of SCOPE and Where2comm [10] on the DAIR-V2X, V2XSet and OPV2V datasets with varying communication volume." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 344, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e451d18b2479306ccbff34efb6869ab379bac8a0f5854bd7c596591249867bed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 90, + 614, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/10bcef53aa60ef03160c33255cafd189b73cf887ba1c6f361c150f14cc4ef2cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 90, + 887, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/866dc6a87604fdd1680e42da86f0d9b393df137582b6a5173ebef319e79a059c.jpg", + "image_caption": [ + "Figure 6. Robustness to the localization error on the DAIR-V2X, V2XSet, and OPV2V datasets." + ], + "image_footnote": [], + "bbox": [ + 81, + 266, + 356, + 396 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9734047217852d965ebab3c7a3f58b121229c521afc331558d263ff75920525.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 266, + 620, + 396 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6dc8fe174cb7352ae503e9020ba8ea9ab9d6addd7ee2ca13507dd1ddcd8a79ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 267, + 888, + 396 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3cdd9dc15dcbb4aa96dbd083e4d9f5499f0665dd3c6af91c0bd05c044a199182.jpg", + "table_caption": [ + "Table 1. Performance comparison on the DAIR-V2X, V2XSet, and OPV2V datasets. The results are reported in AP@0.5/0.7." + ], + "table_footnote": [], + "table_body": "
ModelDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
No Fusion50.03/43.5760.60/40.2068.71/48.66
Late Fusion53.12/37.8866.79/50.9582.24/65.78
Early Fusion61.74/46.5377.39/50.4581.30/68.69
When2com [20]51.12/36.1770.16/53.7277.85/62.40
V2VNet [28]56.01/42.2581.80/61.3582.79/70.31
AttFuse [37]53.79/42.6176.27/57.9383.21/70.09
V2X-ViT [36]54.26/43.3585.13/68.6786.72/74.94
DiscoNet [17]54.29/44.8882.18/63.7387.38/73.19
CoBEVT [34]54.82/43.9583.01/62.6787.40/74.35
Where2comm [10]63.71/48.9385.78/72.4288.07/75.06
SCOPE (Ours)65.18/49.8987.52/75.0589.71/80.62
", + "bbox": [ + 78, + 462, + 465, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cles without collaboration. Late Fusion allows agents to exchange detected results and utilizes non-maximum suppression to produce the final outputs. Early Fusion means the raw LiDAR point clouds are shared to complete the perception view. Meanwhile, the existing state-of-the-art (SOTA) models are fully considered, including When2com [20], V2VNet [28], AttFuse [37], V2X-ViT [36], DiscoNet [17], CoBEVT [34], and Where2comm [10]. Intuitively, the proposed SCOPE outperforms No Fusion, Late Fusion, and Early Fusion by large margins across all datasets, demonstrating the superiority of our perception paradigm. Moreover, our method significantly outperforms previous SOTA models in simulated and real-world (DAIR-V2X) scenarios. In particular, the SOTA performance is improved by $3.63\\%$ and $7.41\\%$ on the V2XSet and OPV2V datasets separately", + "bbox": [ + 75, + 672, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in AP@0.7. Compared to single-frame and collaboration-only based efforts [10, 17, 20, 28, 36], SCOPE proves the effectiveness and rationality of simultaneously considering the temporal context, collaboration, and ego characteristics.", + "bbox": [ + 496, + 438, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of Communication Volume. Evaluating performance at varying bandwidths plays an essential role in achieving pragmatic perception. To this purpose, Figure 5 shows the perception performance of our SCOPE and Where2comm [10] with various communication volumes. Note that we do not consider communication methods that use any extra model/data/feature compression for a fair comparison. We have the following observations: (i) SCOPE achieves better performance than previous models [17, 20, 28, 36] on all datasets with less communication volume. (ii) Compared to Where2comm [10], our method obtains a superior perception-communication tradeoff across all the communication bandwidth choices.", + "bbox": [ + 496, + 503, + 892, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robustness to Localization and Heading Errors. Following the noise settings from [36], we perform experiments to evaluate the robustness against natural localization noises of existing methods. The noises are sampled from a Gaussian distribution with standard deviation $\\sigma_{xyz} \\in [0, 0.5]$ m, $\\sigma_{\\mathrm{heading}} \\in [0^{\\circ}, 1.0^{\\circ}]$ . From Figures 6 and 7, the overall collaborative perception performance consistently deteriorates as errors continue to increase. For instance, When2com and V2VNet fail even worse than No Fusion on the DAIR-V2X when the localization and heading errors are over 0.2 m and $0.4^{\\circ}$ , respectively. Noticeably, our method consistently surpasses the previous SOTA models under all noise levels, clearly displaying the robustness of SCOPE against", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23388", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f11b5a7e055828d46820c0eb0fa0c641e2d2f092501d0fb7aa62a64c3997cc33.jpg", + "image_caption": [ + "Figure 7. Robustness to the heading error on the DAIR-V2X, V2XSet, and OPV2V datasets." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 346, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/452945dfc3e6d64ec3cf1d46f6a6ec8b89db3b7e2f6835a2e7e528f945bc7e0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 93, + 616, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f757639bab00616afa0b5fd38075ec66444ad6ff0aecf46d3f6585e478e1333.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 93, + 887, + 218 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7697102e806923bf7ec5e1ca2eb4d7ec318762de074aa76fb57815898ca5425d.jpg", + "table_caption": [ + "Table 2. Ablation study results of the proposed core components on all the datasets. PL: pyramid LSTM; SIF: selective information filtering; DCM: deformable cross-attention module; RPP: reference point proposal; IAF: importance-aware adaptive fusion." + ], + "table_footnote": [], + "table_body": "
PLSIFDCMRPPIAFDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
60.31/40.3576.04/62.4177.22/65.51
✓61.75/44.5880.36/65.7481.19/69.02
✓✓62.48/45.6281.88/67.4282.34/71.23
✓✓✓64.06/47.8785.47/72.2986.96/76.84
✓✓✓✓64.74/48.4586.33/73.2788.01/78.57
✓✓✓✓✓65.18/49.8987.52/75.0589.71/80.62
", + "bbox": [ + 76, + 311, + 467, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pose errors. The reasonable explanations are: (i) the deformable cross-attention module facilitates capturing perceptually critical information across agents; (ii) the proposed multi-scale feature interaction alleviates the feature map misalignment about collaborative agents due to noises.", + "bbox": [ + 75, + 430, + 468, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 513, + 241, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform thorough ablation studies on all datasets to understand the necessity of the different designs. Tables 2 and 3 show the following observations.", + "bbox": [ + 75, + 537, + 468, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of Core Components. We investigate the contribution of each core component in SCOPE. A base version (first row of Table 2) is provided as the gain reference. In the base model, we use the same fusion pattern to aggregate context information, collaborator-shared features, and multi-source features, where features are concatenated on the channel dimension and fused by a $1 \\times 1$ convolutional network. As Table 2 demonstrates, we progressively add (1) Pyramid LSTM, (2) SIF, (3) DCM, (4) RPP, and (5) IAF and present the corresponding detection precision. The consistently increased detection results over the three datasets reveal the effectiveness of each proposed component, which all provide valuable performance gains.", + "bbox": [ + 75, + 584, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Historical Frame Number. As shown in the top half of Table 3, we evaluate the effect of integrating different degrees of the temporal context from the ego agent. The interesting finding is that choosing a suitable historical frame count (i.e., 2 frames) is beneficial in achieving better performance gains. The potential reason is that adding excessive historical frames may lead to accumulated errors under the noise setting and degrade the performance.", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/db7cb919adf01446054ede96d72f03311c5d074fe00aa2a663603b0db4117580.jpg", + "table_caption": [ + "Table 3. Ablation study results of candidate designs and strategies on all the datasets. \"w/f\" and \"w/o\" mean the with and without." + ], + "table_footnote": [], + "table_body": "
Designs/StrategiesDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
Full Model65.18/49.8987.52/75.0589.71/80.62
Effect of Historical Frame Number
1 Frame64.28/49.1686.16/74.2188.85/78.49
2 Frames (Default)65.18/49.8987.52/75.0589.71/80.62
3 Frames65.03/49.2486.99/74.6589.67/80.45
Effect of Keypoint Number
9 Keypoints64.36/49.2286.34/74.0788.32/77.94
15 Keypoints (Default)65.18/49.8987.52/75.0589.71/80.62
21 Keypoints64.72/49.4087.12/74.6389.08/78.38
Importance of Distinct Features
w/o Ego-agent Featue64.96/49.3486.47/74.2488.63/79.38
w/o Context Feature64.14/48.8185.93/73.8888.26/77.18
w/o Collaboration Feature52.35/44.4663.91/45.3471.36/50.94
Effect of Fusion Strategies
w/ Summation Fusion63.65/48.9185.81/72.6488.05/76.13
w/ Maximum Fusion64.92/49.2786.28/73.6187.95/78.17
w/ Average Fusion64.85/48.9686.55/73.4688.24/78.73
", + "bbox": [ + 501, + 280, + 888, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Keypoint Number. We investigate the rationality of the keypoint number in the deformable crossattention. Noticeably, the DCM component with 15 keypoints achieves better detection performance than other settings. This result shows that SCOPE mitigates the impact of localization errors and fully extracts essential spatial semantics by setting a reasonable keypoint number.", + "bbox": [ + 496, + 561, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of Distinct Features. We explore the importance of distinct representations by removing the ego-agent, context, and collaboration features (i.e., $\\pmb{F}_i^{(t)}, \\pmb{H}_i^{(t)}, \\pmb{Z}_i^{(t)}$ ). There is a significant drop in the model's performance when the collaboration feature is removed, indicating that multiagent information sharing dominates various collaborative perception tasks. Furthermore, the consistently decreased performance suggests that ego-centered characteristics (i.e., pure ego-agent feature and corresponding context information) provide an indispensable contribution.", + "bbox": [ + 496, + 669, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Fusion Strategies. Finally, we demonstrate the effectiveness of the proposed adaptive fusion strategy by providing candidate fusion mechanisms. Summation Fusion is a direct pixel-wise addition operation. Maximum and Average Fusion select elements at corresponding po", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23389", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04fd1d07398a19438ad2d4faa76ed8e950b1d39da7f8222e34097cb4b8737187.jpg", + "image_caption": [ + "(a) DiscoNet" + ], + "image_footnote": [], + "bbox": [ + 124, + 90, + 302, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4758aec46258a60fe5f4d4d55718824214a96fac70a6c7e93d98f1857c50d22.jpg", + "image_caption": [ + "(b) V2X-ViT" + ], + "image_footnote": [], + "bbox": [ + 305, + 90, + 483, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfb1871e46055a4ed03979d1c9c674a52be3855f7632bf2897b7cbd5b599653d.jpg", + "image_caption": [ + "(c) Where2comm" + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 665, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c335617087060a539efecf0a6fd025d4ab0ea6dcf9db90ef39a018a93fd71b1f.jpg", + "image_caption": [ + "(d) SCOPE (ours)" + ], + "image_footnote": [], + "bbox": [ + 668, + 90, + 846, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0191c4efa9eb134c2cf71b8fd9aaf00c156435b469c5f96577c97aff3f497807.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison results in real-world scenarios from the DAIR-V2X dataset. Green and red boxes denote ground truths and detection results, respectively. Compared to the previous SOTA models, our method achieves more accurate detection results.", + "GT Bounding Boxes", + "Figure 9. Visualization of learned multi-scale deformable cross-attention. Green cross marker denotes the proposed reference point. The sampled keypoints are represented by the brown dots whose colors reflect the corresponding attention weights." + ], + "image_footnote": [], + "bbox": [ + 120, + 318, + 302, + 424 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5e140de898e6e8c59eeb909dd2f4560ef1f4a11e71e8c6c5e0f89aacc99a82d.jpg", + "image_caption": [ + "Scale One" + ], + "image_footnote": [], + "bbox": [ + 305, + 318, + 486, + 424 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dac8ab962b3d0b89d684f4a857bf9e413d1cbbd42d5848e650b748dd1442c607.jpg", + "image_caption": [ + "Scale Two" + ], + "image_footnote": [], + "bbox": [ + 486, + 318, + 663, + 424 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c984beec5d3048727ba5502cd28d9c5988311d57e50b10b45f1c31e7a3471238.jpg", + "image_caption": [ + "Scale Three" + ], + "image_footnote": [], + "bbox": [ + 666, + 318, + 848, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sitions among features to retain the maximum or average, respectively, and then reshape them into a fused representation. The poor results in Table 3 show that the above mechanisms fail to tackle the complex multi-source feature fusion challenge. In contrast, the superiority of our IAF module is reflected in considering the complementary contributions of each feature at different positions, adaptively assigning higher importance to spatially beneficial semantics.", + "bbox": [ + 75, + 487, + 468, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 619, + 287, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Detection Results. To intuitively illustrate the perception capability of different models, we provide a detection visualization comparison in Figure 8 under two challenging scenarios. SCOPE achieves more accurate and robust detection results than previous SOTA methods, where the predicted bounding boxes are well aligned to ground truths. In particular, DiscoNet [17], V2X-ViT [36], and Where2comm [10] fail to identify fast-moving objects (i.e., more ground truth bounding boxes are not matched). In contrast, our method achieves holistic perception across heterogeneous agents by effectively aggregating the salient information in the spatio-temporal domains.", + "bbox": [ + 75, + 642, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Attention Map. We visualize the multiscale deformable cross-attention in the CCC component. As Figure 9 shows, our tailored module can dynamically aggregate the semantics of multiple sampled keypoints at different scales to enhance the visual representation of the", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reference point. Concretely, the deformable cross-attention learns to extract object-related valuable perception cues at lower scales by sampling keypoints inside or around targets. In comparison, it reasonably samples long-range features at higher scales to capture spatially critical information.", + "bbox": [ + 496, + 487, + 892, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 574, + 617, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present SCOPE, a learning-based framework to address existing multi-agent collaborative perception challenges in an end-to-end manner. Our approach is the first time considering the temporal semantics from ego agents to capture valuable context cues through the context-aware component. Moreover, two spatial information aggregation components are introduced to achieve comprehensive information interaction among agents and refined feature fusion of ego agents. Extensive experiments on diverse 3D detection datasets prove the superiority of SCOPE and the effectiveness of our components. The future work focuses on extending SCOPE into more complex perception scenarios.", + "bbox": [ + 496, + 601, + 892, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 814, + 666, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the Shanghai Key Research Laboratory of NSAI, the National Natural Science Foundation of China (Grant No. 62250410368), and the Nanjing First Automobile Works Grant.", + "bbox": [ + 496, + 839, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23390", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3D point clouds. In Proceedings of ACM/IEEE Symposium on Edge Computing, pages 88-100, Arlington Virginia, Nov. 2019. 3", + "[2] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524, 2019. 3", + "[3] Abhishek Das, Theophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. In International Conference on Machine Learning (ICML), pages 1538-1546, 2019. 3", + "[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on Robot Learning (CoRL), pages 1-16, 2017. 5", + "[5] Yangtao Du, Dingkang Yang, Peng Zhai, Mingchen Li, and Lihua Zhang. Learning associative representation for facial expression recognition. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 889-893, 2021. 1", + "[6] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Yogamani. Rgb and lidar fusion based 3d semantic segmentation for autonomous driving. In Proceedings of the IEEE Intelligent Transportation Systems Conference (ITSC), pages 7-12, 2019. 1", + "[7] Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. Advances in Neural Information Processing Systems (NIPS), 29, 2016. 2", + "[8] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735-1780, 1997. 4", + "[9] Yedid Hoshen. Vain: Attentional multi-agent predictive modeling. Advances in Neural Information Processing Systems (NIPS), 30, 2017. 3", + "[10] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Advances in Neural Information Processing Systems (NIPS), 2022. 1, 3, 4, 5, 6, 8", + "[11] Yue Hu, Yifan Lu, Runsheng Xu, Weidi Xie, Siheng Chen, and Yanfeng Wang. Collaboration helps camera overtake lidar in 3d detection. arXiv preprint arXiv:2303.13560, 2023. 4", + "[12] Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent cooperation. Advances in Neural Information Processing Systems (NIPS), 31, 2018. 3", + "[13] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5", + "[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12697-12705, 2019. 1, 3, 5", + "[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Proceedings of the European Conference on Computer Vision (ECCV), pages 316-332, 2022. 4", + "[16] Jinlong Li, Runsheng Xu, Xinyu Liu, Jin Ma, Zicheng Chi, Jiaqi Ma, and Hongkai Yu. Learning for vehicle-to-vehicle cooperative perception under lossy communication. IEEE Transactions on Intelligent Vehicles, 2023. 1", + "[17] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems (NIPS), 34:29541-29552, 2021. 1, 3, 4, 6, 8", + "[18] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 641-656, 2018. 2", + "[19] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 5", + "[20] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4106-4115, 2020. 1, 3, 4, 6", + "[21] Guiyang Luo, Hui Zhang, Quan Yuan, and Jinglin Li. Complementarity-enhanced and redundancy-minimized collaboration network for multi-agent perception. In Proceedings of the ACM International Conference on Multimedia (ACMM), pages 3578-3586, 2022. 1, 3, 4", + "[22] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NIPS), 32, 2019. 5", + "[23] Tika Ram and Khem Chand. Effect of drivers' risk perception and perception of driving tasks on road safety attitude. Transportation Research Part F: Traffic Psychology and Behaviour, 42:162-176, 2016. 1", + "[24] Andreas Rauch, Felix Klanner, Ralph Rasshofer, and Klaus Dietmayer. Car2x-based perception in a high-level fusion architecture for cooperative perception systems. In IEEE Intelligent Vehicles Symposium, pages 270-275, 2012. 3", + "[25] Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in multiagent cooperative and competitive tasks. arXiv preprint arXiv:1812.09755, 2018. 2", + "[26] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation. Advances in Neural Information Processing Systems (NIPS), 29, 2016. 2", + "[27] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In International Conference on Machine Learning (ICML), pages 330-337, 1993. 2" + ], + "bbox": [ + 503, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23391", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 605–621, 2020. 1, 3, 4, 6", + "[29] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. In IEEE International Conference on Robotics and Automation (ICRA), pages 3584-3591. IEEE, 2023. 1", + "[30] Zhu Xizhou, Su Weijie, Lu Lewei, Li Bin, Wang Xiaogang, and Dai Jifeng. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2021. 4", + "[31] Runsheng Xu, Weizhe Chen, Hao Xiang, Xin Xia, Lantao Liu, and Jiaqi Ma. Model-agnostic multi-agent perception framework. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1471–1478. IEEE, 2023. 3", + "[32] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 5", + "[33] Runsheng Xu, Jinlong Li, Xiaoyu Dong, Hongkai Yu, and Jiaqi Ma. Bridging the domain gap for multi-agent perception. arXiv preprint arXiv:2210.08451, 2022. 1", + "[34] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. In Conference on Robot Learning (CoRL), 2022. 3, 6", + "[35] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13712–13722, 2023. 1", + "[36] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6, 8", + "[37] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In Proceedings of the International Conference on Robotics and Automation (ICRA), pages 2583–2589, 2022. 1, 2, 3, 4, 5, 6", + "[38] Dingkang Yang, Zhaoyu Chen, Yuzheng Wang, Shunli Wang, Mingcheng Li, Siao Liu, Xiao Zhao, Shuai Huang, Zhiyan Dong, Peng Zhai, and Lihua Zhang. Context deconfounded emotion recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19005-19015, June 2023. 2", + "[39] Dingkang Yang, Shuai Huang, Haopeng Kuang, Yangtao Du, and Lihua Zhang. Disentangled representation learning for multimodal emotion recognition. In Proceedings of the 30th" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ACM International Conference on Multimedia (ACM MM), page 1642-1651, 2022. 2", + "[40] Dingkang Yang, Shuai Huang, Yang Liu, and Lihua Zhang. Contextual and cross-modal interaction for multi-modal speech emotion recognition. IEEE Signal Processing Letters, 29:2093-2097, 2022. 2", + "[41] Dingkang Yang, Shuai Huang, Shunli Wang, Yang Liu, Peng Zhai, Liuzhen Su, Mingcheng Li, and Lihua Zhang. Emotion recognition for multiple context awareness. In Proceedings of the European Conference on Computer Vision (ECCV), volume 13697, pages 144-162. Springer, 2022. 1", + "[42] Dingkang Yang, Haopeng Kuang, Shuai Huang, and Lihua Zhang. Learning modality-specific and -agnostic representations for asynchronous multimodal language sequences. In Proceedings of the 30th ACM International Conference on Multimedia (ACMM), page 1708-1717, 2022. 2", + "[43] Dingkang Yang, Yang Liu, Can Huang, Mingcheng Li, Xiao Zhao, Yuzheng Wang, Kun Yang, Yan Wang, Peng Zhai, and Lihua Zhang. Target and source modality co-reinforcement for emotion understanding from asynchronous multimodal sequences. Knowledge-Based Systems, page 110370, 2023. 1", + "[44] Kun Yang, Jing Liu, Dingkang Yang, Hanqi Wang, Peng Sun, Yanni Zhang, Yan Liu, and Liang Song. A novel efficient multi-view traffic-related object detection framework. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2", + "[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In IEEE International Conference on Distributed Computing Systems (ICDCS), pages 972-982, 2022. 2", + "[46] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21361-21370, 2022. 1, 2, 5", + "[47] Zhenxun Yuan, Xiao Song, Lei Bai, Zhe Wang, and Wanli Ouyang. Temporal-channel transformer for 3d lidar-based video object detection for autonomous driving. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2068-2078, 2021. 1, 3", + "[48] Zixu Zhang and Jaime F Fisac. Safe occlusion-aware autonomous driving via game-theoretic active perception. arXiv preprint arXiv:2105.08169, 2021. 1", + "[49] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 496-513, 2022. 1, 3" + ], + "bbox": [ + 501, + 92, + 890, + 823 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23392", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/fe9d96d8-4150-4b6d-a27c-f4b80a133514_model.json b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/fe9d96d8-4150-4b6d-a27c-f4b80a133514_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa1cc3bc3024294c5580e6430ca5eeb239ff9a1d --- /dev/null +++ b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/fe9d96d8-4150-4b6d-a27c-f4b80a133514_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.089, + 0.131, + 0.882, + 0.154 + ], + "angle": 0, + "content": "Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.179, + 0.849, + 0.253 + ], + "angle": 0, + "content": "Kun Yang\\(^{1\\dagger}\\) Dingkang Yang\\(^{1\\dagger}\\) Jingyu Zhang\\(^{1}\\) Mingcheng Li\\(^{1}\\) Yang Liu\\(^{1}\\) Jing Liu\\(^{1}\\) Hanqi Wang\\(^{1}\\) Peng Sun\\(^{2}\\) Liang Song\\(^{1*}\\) \n\\(^{1}\\)Academy for Engineering and Technology, Fudan University \\(^{2}\\)Duke Kunshan University {kunyang20,dkyang20,songl}@fudan.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.474, + 0.668 + ], + "angle": 0, + "content": "Multi-agent collaborative perception as a potential application for vehicle-to-everything communication could significantly improve the perception performance of autonomous vehicles over single-agent perception. However, several challenges remain in achieving pragmatic information sharing in this emerging research. In this paper, we propose SCOPE, a novel collaborative perception framework that aggregates the spatio-temporal awareness characteristics across on-road agents in an end-to-end manner. Specifically, SCOPE has three distinct strengths: i) it considers effective semantic cues of the temporal context to enhance current representations of the target agent; ii) it aggregates perceptually critical spatial information from heterogeneous agents and overcomes localization errors via multi-scale feature interactions; iii) it integrates multi-source representations of the target agent based on their complementary contributions by an adaptive fusion paradigm. To thoroughly evaluate SCOPE, we consider both real-world and simulated scenarios of collaborative 3D object detection tasks on three datasets. Extensive experiments show the superiority of our approach and the necessity of the proposed components. The project link is https://ydk122024.github.io/SCOPE/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.696, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.721, + 0.47, + 0.842 + ], + "angle": 0, + "content": "Perceiving complex driving environments is essential for ensuring road safety [23] and traffic efficiency [21] in autonomous driving. Despite the rapid development in the perception ability of intelligent vehicles through deep learning technologies [5, 6, 14, 41, 43], the conventional single-agent perception paradigm remains challenging. Single-agent perception system is generally vulnerable and defective due to the inevitable practical difficulties on the road," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.893, + 0.395 + ], + "angle": 0, + "content": "such as restricted detection range and occlusion [47, 48]. To alleviate the inadequate observation constraints of isolated agents, multi-agent collaborative perception has emerged as a promising solution in recent years [17, 29]. With Vehicle-to-Everything (V2X) oriented communication applications, collaborative perception effectively promotes perspective complementation and information sharing among agents." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Based on recent LiDAR-based 3D collaborative perception datasets [35, 36, 37, 46], several impressive works have achieved a trade-off between collaborative perception performance and communication bandwidth of Autonomous Vehicles (AVs) via well-designed mechanisms. These remarkable efforts include knowledge distillation-based intermediate collaboration [17], handshake communication mechanism [20], V2X-visual transformer [36], and spatial information selection strategy [10]. However, there is still considerable room for progress in achieving a pragmatic and effective collaborative perception. Concretely, existing methods invariably follow a single-frame static perception pattern, suffering from the data sparsity shortcoming of 3D point clouds [47] and ignoring meaningful semantic cues in the temporal context [49]. The insufficient representation of the moving objects in the current frame potentially restricts the perception performance of target vehicles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Moreover, spatial information aggregation in collaborative perception systems has exposed several problems. For the collaborator-shared message fusion, the per-agent/location-based fusion strategies of previous attempts [16, 17, 20, 28, 36] fail to deal with feature map misalignment from heterogeneous agents due to localization errors. Consequently, misleading features from collaborators could lead to object position misjudgment of the ego vehicle (i.e., receiver) and harm its detection precision. For the ego vehicle message refinement, existing approaches [10, 17, 20, 21, 28, 33] rely on fused representations with collaborator information to implement detection, abandoning the natural perception advantages of the ego vehicle and introducing the potential noises. The ego-centered characteristics may contain locally critical spatial information that is not disturbed by noise from collaborative agents" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.47, + 0.901 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Equal contributions in no particular order. The work was done when Dingkang Yang was at the Institute of Meta-Medical, IPASS. The two first authors thank Runsheng Xu for providing constructive suggestions. *Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23383" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.088, + 0.852, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.366, + 0.895, + 0.41 + ], + "angle": 0, + "content": "Figure 1. The overall architecture of the proposed SCOPE. The framework consists of five parts: metadata conversion and feature extraction, context-aware information aggregation, confidence-aware cross-agent collaboration, importance-aware adaptive fusion, and detection decoders. The details of each individual component are illustrated in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.471, + 0.468 + ], + "angle": 0, + "content": "with asynchronous sensor measurements. To this end, how to effectively break the aforementioned limitations becomes a priority for achieving robust collaborative perception." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Motivated by the above observations, we present SCOPE, a Spatio-temporal domain awareness multi-agent CollabOrative PErception approach to tackle existing challenges jointly. From Figure 1, SCOPE achieves effective information collaboration among agents and precise perception of the ego agent via three proposed core components. Specifically, (i) for the data sparsity challenge in single-frame point clouds, the context-aware information aggregation is proposed to aggregate the ego agent's context information from preceding frames. We employ selective information filtering and spatio-temporal feature integration modules to capture informative temporal cues and historical context semantics. (ii) For the collaborator-shared message fusion challenge, the confidence-aware cross-agent collaboration is introduced to guarantee that the ego agent aggregates complementary information from heterogeneous agents. Based on the confidence-aware multi-scale feature interaction, we facilitate holistic communication and mitigate feature map misalignment due to localization errors of collaborators. (iii) For the multi-feature integration challenge of the ego agent, the importance-aware adaptive fusion is designed to flexibly incorporate the advantages of diverse representations based on distinctive perception contributions. We evaluate the superiority of SCOPE on various collaborative 3D object detection datasets, including DAIR-V2X [46], OPV2V [37], and V2XSet [36]. Extensive experiments in real-world and simulated scenarios provide favorable evidence that our approach is competitive in" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.421, + 0.783, + 0.435 + ], + "angle": 0, + "content": "multi-agent collaborative perception tasks." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.436, + 0.882, + 0.451 + ], + "angle": 0, + "content": "The main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.46, + 0.892, + 0.565 + ], + "angle": 0, + "content": "- We present SCOPE, a novel framework for multi-agent collaborative perception. The framework facilitates information collaboration and feature fusion among agents, achieving a reasonable performance-bandwidth trade-off. Comprehensive experiments on collaborative detection tasks show that SCOPE outperforms previous state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.574, + 0.892, + 0.664 + ], + "angle": 0, + "content": "- To our best knowledge, we are the first to consider the temporal context of the ego agent in collaborative perception systems. Based on the proposed context-aware component, the target frame of point clouds efficiently integrates historical cues from preceding frames to capture valuable temporal information." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.674, + 0.892, + 0.765 + ], + "angle": 0, + "content": "- We introduce two spatial information aggregation components to address the challenges of collaboration heterogeneity and fused representation unicity. These tailored components effectively perform multigrained information interactions among agents and multi-source feature refinement of the ego agent." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.46, + 0.892, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.642, + 0.793 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.76, + 0.818 + ], + "angle": 0, + "content": "2.1. Multi-Agent Communication" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Benefiting from rapid advances in learning-based technologies [18, 38, 39, 40, 42, 44, 45], communication has become an essential component in constructing robust multiagent systems. Previous works [7, 25, 26, 27] typically focus on centralized paradigms or predefined protocols to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23384" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "explore information-sharing mechanisms among multiple agents. However, these efforts are potentially domain-constrained and scenario-isolated. Recent methods [3, 9, 12] achieve effective communication strategies in diverse scenarios. For instance, Vain [9] determines which agents would share information by modeling the locality of interactions. ATOC [12] introduces an attentional communication strategy to learn when communicating and integrating useful information to execute collaborative decisions. In comparison, we focus on the more challenging 3D perception tasks in complex driving scenarios. Based on the proposed confidence-aware multi-scale information interaction, SCOPE achieves an information-critical and holistic communication pattern among heterogeneous agents." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.311, + 0.307, + 0.327 + ], + "angle": 0, + "content": "2.2. Collaborative Perception" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.652 + ], + "angle": 0, + "content": "Collaborative perception investigates how to aggregate complementary perception semantics across connected agents to improve overall system performance. Current works are mainly categorized as early [2], intermediate [1, 10, 17, 20, 28, 34, 36, 37], and late [24, 31] fusion, where intermediate fusion is more favoured due to its good trade-off between performance and transmission bandwidth. In this case, several remarkable works explore efficient collaboration mechanisms. In DiscoNet [17], the teacher-student network is trained to aggregate the shared information. V2X-ViT [36] proposes heterogeneous multi-agent self-attention to handle device heterogeneity, and Xu et al. [34] design an axial attention module to capture sparse spatial semantics. Furthermore, Where2comm [10] filters irrelevant semantics by quantifying the spatial importance through confidence maps. Unlike these single-frame methods, we propose an information aggregation component to extract rich context information from local historical observations. Moreover, our cross-agent collaboration mechanism outperforms the previous per-agent/location message fusion models, significantly mitigating the impact of localization errors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.663, + 0.214, + 0.681 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.78 + ], + "angle": 0, + "content": "In this paper, we focus on developing a robust collaborative perception system to improve the ego vehicle's perception capability. Figure 1 illustrates the overall architecture of the proposed framework. It consists of five sequential parts, where the system input accommodates different types of connected agents (i.e., infrastructures and AVs)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.465, + 0.803 + ], + "angle": 0, + "content": "3.1. Metadata Conversion and Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "During the collaboration, one of the connected agents is identified as the ego agent \\((i)\\) to build a communication graph, where other agents within the communication range serve as collaborators. Upon receiving the metadata \\((e.g.,\\) poses and extrinsics) broadcast by the ego agent, the collaborators project their local LiDAR point clouds" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.766, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.091, + 0.889, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.231, + 0.894, + 0.288 + ], + "angle": 0, + "content": "Figure 2. (a) The architecture of the proposed CIA component, including selective information filtering (left) and spatio-temporal feature integration (middle). (b) The multi-scale convolutional structure in the pyramid LSTM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.892, + 0.44 + ], + "angle": 0, + "content": "to the coordinate system of the ego agent. Similarly, the ego agent's previous point cloud frames are synchronized to the current frame. Each agent \\( k \\in \\{1, \\dots, K\\} \\) encodes the projected point clouds into Bird's Eye View (BEV) features to extract local visual representations. Given the point cloud \\( \\mathbf{X}_k^{(t)} \\) of \\( k \\)-th agent at timestamp \\( t \\), the extracted feature is \\( \\mathbf{F}_k^{(t)} = f_{enc}(\\mathbf{X}_k^{(t)}) \\in \\mathbb{R}^{C \\times H \\times W} \\), where \\( f_{enc}(\\cdot) \\) is the PointPillar [14] encoder shared among all agents, and \\( C, H, W \\) stand for the channel, height, and width." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.448, + 0.847, + 0.464 + ], + "angle": 0, + "content": "3.2. Context-aware Information Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.652 + ], + "angle": 0, + "content": "Conventional 3D detection tasks [47, 49] have shown the potential of temporal information in improving perception performance. Unfortunately, current multi-agent collaborative efforts [10, 17, 20, 21, 28] invariably focus on exploring the current frame while ignoring context cues from previous frames. The single-frame solution fails to effectively detect fast-moving objects (e.g., surrounding vehicles) due to the sparsity and insufficiency of point clouds. To this end, we present a Context-aware Information Aggregation (CIA) component to capture spatio-temporal representations of the ego agent's previous frames to fuse valuable semantics. As shown in Figure 2(a), CIA consists of two core phases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.895, + 0.838 + ], + "angle": 0, + "content": "Selective Information Filtering. This phase aims to extract refined information by filtering redundant features from the target \\( \\pmb{F}_i^{(t)} \\in \\mathbb{R}^{C \\times H \\times W} \\) and distilling meaningful characteristics from the previous \\( \\pmb{F}_i^{(t - \\tau)} \\in \\mathbb{R}^{\\tau \\times C \\times H \\times W} \\), where \\( \\tau \\) is the time shift. Concretely, the channel-wise average and max pooling operations \\( \\Psi_{a / m}(\\cdot) \\) are first utilized to aggregate rich channel semantics, denoted as \\( \\mathcal{M}_{a / m}^{(t)} = \\Psi_{a / m}(\\pmb{F}_i^{(t)}) \\) and \\( \\mathcal{M}_{a / m}^{(\\tau)} = \\Psi_{a / m}(\\pmb{F}_i^{(\\tau)}) \\), where \\( \\pmb{F}_i^{(\\tau)} = \\sum_{n=1}^{\\tau} \\pmb{F}_i^{(t-n)} \\). These refined feature maps are combined to obtain a well-chosen spatial selection map. The procedure is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.844, + 0.892, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {U} = \\sigma \\cdot \\aleph (\\mathcal {M} _ {a} ^ {(t)} \\| \\mathcal {M} _ {m} ^ {(t)} + \\mathcal {M} _ {a} ^ {(\\tau)} \\| \\mathcal {M} _ {m} ^ {(\\tau)}) \\in \\mathbb {R} ^ {H \\times W}, \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the sigmoid activation, \\(\\aleph\\) denotes a convolutional layer for fusion, and \\(\\parallel\\) denotes the concatenation op" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23385" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.092, + 0.468, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.47, + 0.306 + ], + "angle": 0, + "content": "Figure 3. (a) The architecture of the proposed CCC component. (b) CCC consists of two phases at each scale: reference point proposal (middle) and deformable cross-attention module (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.368 + ], + "angle": 0, + "content": "eration. Immediately, the information filtering strategy is implemented via pixel-wise attention weighting to enhance the salient historical features of \\( H_{i}^{(\\tau)} \\in \\mathbb{R}^{\\tau \\times C \\times H \\times W} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.377, + 0.469, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {H} _ {i} ^ {(\\tau)} = (1 - \\mathcal {U}) \\odot \\tanh \\left(\\boldsymbol {F} _ {i} ^ {(t)}\\right) + \\mathcal {U} \\odot \\boldsymbol {F} _ {i} ^ {(t - \\tau)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.47, + 0.715 + ], + "angle": 0, + "content": "Spatio-temporal Feature Integration. In order to integrate the historical prototype to improve the perception ability of the current representation, we introduce a pyramid LSTM [15] to learn the contextual dependencies of interframe features \\( \\pmb{F}_i^{(t)}\\| \\pmb{H}_i^{(\\tau)} \\). The difference between the pyramid LSTM and vanilla LSTM [8] is that the initial matrix multiplication is replaced with a multi-scale convolutional structure, as shown in Figure 2(b). Note that the intrinsic long short-term memory mechanism in the LSTM is adopted, and please refer to [8] for details. In practice, multi-scale spatial features are extracted by two successive 2D convolutions among different scales, followed by batch normalization and ReLU activation. To achieve multilevel spatial semantic fusion, the downsampled features are progressively interpolated to deliver to the upsampled layers via lateral connections. The overall design facilitates learning spatio-temporal correlations in historical contexts, which benefits our multi-frame perception paradigm. Ultimately, the final hidden state in the pyramid LSTM is used as the refined context-aware feature \\( \\pmb{H}_i^{(t)}\\in \\mathbb{R}^{C\\times H\\times W} \\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.722, + 0.46, + 0.739 + ], + "angle": 0, + "content": "3.3. Confidence-aware Cross-agent Collaboration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.47, + 0.904 + ], + "angle": 0, + "content": "The cross-agent collaboration targets to enhance the ego agent's visual representations by aggregating the complementary semantics from the collaborator-shared features. Existing works [10, 11, 36, 37] propose attention-based perlocation feature fusion methods, which are vulnerable to localization errors and ignore the sparsity of point clouds. To tackle these issues, we implement a novel Confidence-aware Cross-agent Collaboration (CCC) component. We first obtain the confidence map \\( S_{k}^{(t)} \\) and binary selection matrix \\( \\mathcal{C}_k^{(t)} \\) of feature \\( F_{k}^{(t)}, k\\in \\{1,\\dots,K\\} \\) following [10]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.206 + ], + "angle": 0, + "content": "The filtered feature shared by \\( k \\)-th agent to ego agent is \\( \\pmb{F}_k^{(t)} = \\pmb{F}_k^{(t)} \\odot \\mathcal{C}_k^{(t)} \\). As Figure 3(a) shows, the CCC component encodes the features and confidence maps into three scales and performs feature fusion at each scale. \\( \\pmb{F}_{k,l}^{(t)} \\) and \\( S_{k,l}^{(t)} \\) denote the feature and confidence map at \\( l \\)-th scale. We take two agents as an example to elaborate on the feature fusion in Figure 3(b), including the following two phases." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.207, + 0.893, + 0.367 + ], + "angle": 0, + "content": "Reference Point Proposal. To predict the spatial locations containing meaningful object information, we obtain the element-wise summation of all confidence maps as \\( S_{sum,l}^{(t)} = \\sum_{k=1}^{K} S_{k,l}^{(t)} \\). Since the confidence map reflects the spatially critical levels, \\( S_{sum,l}^{(t)} \\) presents the potential positions of the targets within the detection range, called the reference points. Accordingly, we apply a threshold-based selection function \\( f_{sel}(\\cdot) \\) to extract the reference points \\( S_{re,l}^{(t)} \\). This design could actively guide the subsequent fusion network to concentrate on the essential spatial areas." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.368, + 0.893, + 0.567 + ], + "angle": 0, + "content": "Deformable Cross-attention Module. We extract the ego agent's feature \\( \\pmb{F}_{i,l}^{(t)} \\) at the reference points as the initial query embedding and apply a linear layer to encode the locations of reference points into a position embedding. To resolve the misalignment of feature maps and acquire more robust representations against localization errors, the Deformable Cross-attention Module (DCM) aggregates the information from sampled keypoints through a deformable cross-attention layer [30]. Concretely, a linear layer is leveraged to learn the 2D spatial offset \\( \\Delta q \\) of the reference point \\( q \\) for sampling attending keypoints at \\( q + \\Delta q \\). We simultaneously select \\( M \\) keypoints and extract their features as the attending features. The output of DCM at position \\( q \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.573, + 0.892, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\operatorname {D C M} (q) = \\sum_ {a = 1} ^ {A} \\boldsymbol {W} _ {a} \\left[ \\sum_ {k = 1} ^ {K} \\sum_ {m = 1} ^ {M} \\phi \\left(\\boldsymbol {W} _ {b} \\boldsymbol {F} _ {i, l} ^ {(t)} (q)\\right) \\boldsymbol {F} _ {k, l} ^ {(t)} \\left(q + \\Delta q _ {m}\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.615, + 0.892, + 0.761 + ], + "angle": 0, + "content": "where \\(a\\) indexes the attention head, \\(\\mathbf{W}_{a / b}\\) are the learnable weights, and \\(\\phi (\\cdot)\\) is the softmax function. Ultimately, we derive a filling operation that fills \\(\\mathrm{DCM}(q)\\) into the ego agent's feature \\(\\pmb{F}_{i,l}^{(t)}\\) based on the initial position \\(q\\) and outputs \\(\\pmb{Z}_{i,l}^{(t)}\\). The output-enhanced features under the three scales are encoded into the same size and concatenated in the channel dimension. We utilize a \\(1\\times 1\\) convolutional layer to fuse information of the three scales and get the final collaboration feature \\(\\pmb{Z}_i^{(t)}\\in \\mathbb{R}^{C\\times H\\times W}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.809, + 0.788 + ], + "angle": 0, + "content": "3.4. Importance-aware Adaptive Fusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although previous works [17, 20, 21, 28, 36] have achieved impressive performance through ego features that aggregate collaborative information, they potentially suffer from noise interference introduced by collaborators with asynchronous measurements. A promising solution is to consider purely ego-centered characteristics, which contain the natural perception advantages of the target agent. To" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23386" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.465, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.272, + 0.43, + 0.286 + ], + "angle": 0, + "content": "Figure 4. The architecture of the proposed IAF component." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.47, + 0.39 + ], + "angle": 0, + "content": "this end, we propose an Importance-aware Adaptive Fusion (IAF) component to fuse multi-source features based on their complementary contributions. These informative features include \\( \\pmb{H}_i^{(t)}, \\pmb{Z}_i^{(t)} \\), and \\( \\pmb{F}_i^{(t)} \\). From Figure 4, the importance generator \\( f_{gen}(\\cdot) \\) is used to generate the spatial importance maps, which processes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.395, + 0.469, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {i} ^ {(t)}, \\mathcal {Z} _ {i} ^ {(t)}, \\mathcal {S} _ {i} ^ {(t)} = \\sigma \\cdot \\Psi_ {m} (f _ {g e n} (\\boldsymbol {H} _ {i} ^ {(t)}, \\boldsymbol {Z} _ {i} ^ {(t)}, \\boldsymbol {F} _ {i} ^ {(t)})), \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.421, + 0.469, + 0.471 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}_i^{(t)},\\mathcal{Z}_i^{(t)},S_i^{(t)}\\in [0,1]^{H\\times W}\\). Then, these importance maps are concatenated to get the normalized attention maps across distinct features by the softmax function \\(\\phi (\\cdot)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.476, + 0.469, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\mathcal {H}} ^ {(t)} = \\phi \\left(\\mathcal {H} _ {i} ^ {(t)}\\right) = \\frac {\\exp \\left(\\mathcal {H} _ {i} ^ {(t)}\\right)}{\\exp \\left(\\mathcal {H} _ {i} ^ {(t)}\\right) + \\exp \\left(\\mathcal {Z} _ {i} ^ {(t)}\\right) + \\exp \\left(\\mathcal {S} _ {i} ^ {(t)}\\right)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.516, + 0.469, + 0.613 + ], + "angle": 0, + "content": "A larger value in \\(\\mathcal{E}_{\\mathcal{H}}^{(t)} \\in \\mathbb{R}^{H \\times W}\\) implies that the corresponding feature position is more significant. Similarly, \\(\\mathcal{E}_{\\mathcal{Z}}^{(t)} = \\phi(\\mathcal{Z}_i^{(t)})\\) and \\(\\mathcal{E}_S^{(t)} = \\phi(S_i^{(t)})\\). Finally, we adaptively activate each feature map's perceptually critical spatial information by a weighted summation. The fused feature map \\(\\mathcal{F}_i^{(t)} \\in \\mathbb{R}^{C \\times H \\times W}\\) is obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.619, + 0.469, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {i} ^ {(t)} = \\mathcal {E} _ {\\mathcal {H}} ^ {(t)} \\odot \\boldsymbol {H} _ {i} ^ {(t)} + \\mathcal {E} _ {\\mathcal {Z}} ^ {(t)} \\odot \\boldsymbol {Z} _ {i} ^ {(t)} + \\mathcal {E} _ {\\mathcal {S}} ^ {(t)} \\odot \\boldsymbol {F} _ {i} ^ {(t)}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.645, + 0.469, + 0.661 + ], + "angle": 0, + "content": "3.5. Detection Decoder and Objective Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.469, + 0.812 + ], + "angle": 0, + "content": "We leverage two detection decoders to decode the final fused feature \\(\\mathcal{F}_i^{(t)}\\) into the predicted outputs. The regression output is \\(\\mathcal{Y}_{i,r}^{(t)} = f_{dec}^r (\\mathcal{F}_i^{(t)})\\in \\mathbb{R}^{7\\times H\\times W}\\), representing the position, size, and yaw angle of the bounding box at each location. The classification output reveals the confidence value of each predefined box to be a target or background, which is \\(\\mathcal{Y}_{i,c}^{(t)} = f_{dec}^c (\\mathcal{F}_i^{(t)})\\in \\mathbb{R}^{2\\times H\\times W}\\). Following existing work [14], we adopt the smooth \\(L_{1}\\) loss for regressing the bounding boxes and a focal loss [19] for classification." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.21, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.364, + 0.862 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets. We conduct comprehensive experiments on three benchmark datasets for collaborative perception tasks, in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.408 + ], + "angle": 0, + "content": "including DAIR-V2X [46], V2XSet [36], and OPV2V [37]. These datasets provide annotated sensor data from multiple connected agents for each sample. Concretely, DAIR-V2X is a large-scale real-world dataset supporting collaborative 3D object detection, containing 71,254 samples and dividing the training/validation/testing sets as 5:2:3. Each sample contains the LiDAR point clouds from a vehicle and an infrastructure. During feature extraction, we transform the point cloud into the BEV feature map of the size (200, 504, 64). V2XSet is a public simulation dataset for V2X collaborative perception, including 73 representative scenarios and 11,447 frames of labeled point clouds generated by Carla [4]. The training/validation/testing sets are 6,694, 1,920, and 2,833 frames, respectively. The size of each extracted feature map is (100, 384, 64). OPV2V is a vehicle-to-vehicle collaborative perception dataset collected by Carla and OpenCDA [32]. It includes 11,464 frames of point clouds and RGB images with 3D annotation, splitting into training/validation/testing sets of 6,764, 1,981, and 2,719 frames, respectively. The 3D point cloud is extracted as a feature map of the size (100, 384, 64)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Evaluation Metrics. Following [36], we leverage the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the 3D object detection performance. The communication results count the message size by byte in the log scale with base 2 [10]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.722, + 0.51 + ], + "angle": 0, + "content": "4.2. Implementation Details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.807 + ], + "angle": 0, + "content": "We build the proposed models on the Pytorch toolbox [22] and train them on four Nvidia Tesla V100 GPUs using the Adam optimizer [13]. The training settings of the DAIR-V2X, V2XSet, and OPV2V datasets follow: the batch sizes are \\(\\{5, 3, 3\\}\\) and the epochs are \\(\\{30, 40, 40\\}\\). The initial learning rate is \\(2\\mathrm{e} - 3\\) and decays every 15 epochs with a factor of 0.1. Our CIA module uses the previous 2 frames, the attention head in the CCC component is 8, and the number of keypoints is 15. To simulate the default localization and heading errors, we add Gaussian noise with a standard deviation of \\(0.2\\mathrm{m}\\) and \\(0.2^{\\circ}\\) to the initial position and orientation of the agents. The voxel resolution of the observation encoder \\(f_{enc}(\\cdot)\\) is set to \\(0.4\\mathrm{m}\\) in height and width. The generator \\(f_{gen}(\\cdot)\\) follows a detection decoder from [14], and the summed confidence map \\(S_{sum,l}^{(t)}\\) is filtered using a threshold of 0.1 in the function \\(f_{sel}(\\cdot)\\). The decoders \\(\\{f_{dec}^{r}(\\cdot), f_{dec}^{c}(\\cdot)\\}\\) are derived with two \\(1 \\times 1\\) convolutional layers. We provide results on the validation set of DAIR-V2X and the testing set of V2XSet and OPV2V." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.722, + 0.833 + ], + "angle": 0, + "content": "4.3.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Detection Performance Comparison. Table 1 shows the 3D detection performance comparison results on the three datasets. We compare the following methods. No Fusion as a baseline considers only the point cloud data of ego vehi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23387" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.345, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.092, + 0.616, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.092, + 0.888, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.231, + 0.893, + 0.26 + ], + "angle": 0, + "content": "Figure 5. Collaborative perception performance comparison of SCOPE and Where2comm [10] on the DAIR-V2X, V2XSet and OPV2V datasets with varying communication volume." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.267, + 0.357, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.267, + 0.622, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.268, + 0.889, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.407, + 0.768, + 0.421 + ], + "angle": 0, + "content": "Figure 6. Robustness to the localization error on the DAIR-V2X, V2XSet, and OPV2V datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.435, + 0.469, + 0.462 + ], + "angle": 0, + "content": "Table 1. Performance comparison on the DAIR-V2X, V2XSet, and OPV2V datasets. The results are reported in AP@0.5/0.7." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.463, + 0.467, + 0.644 + ], + "angle": 0, + "content": "
ModelDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
No Fusion50.03/43.5760.60/40.2068.71/48.66
Late Fusion53.12/37.8866.79/50.9582.24/65.78
Early Fusion61.74/46.5377.39/50.4581.30/68.69
When2com [20]51.12/36.1770.16/53.7277.85/62.40
V2VNet [28]56.01/42.2581.80/61.3582.79/70.31
AttFuse [37]53.79/42.6176.27/57.9383.21/70.09
V2X-ViT [36]54.26/43.3585.13/68.6786.72/74.94
DiscoNet [17]54.29/44.8882.18/63.7387.38/73.19
CoBEVT [34]54.82/43.9583.01/62.6787.40/74.35
Where2comm [10]63.71/48.9385.78/72.4288.07/75.06
SCOPE (Ours)65.18/49.8987.52/75.0589.71/80.62
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.901 + ], + "angle": 0, + "content": "cles without collaboration. Late Fusion allows agents to exchange detected results and utilizes non-maximum suppression to produce the final outputs. Early Fusion means the raw LiDAR point clouds are shared to complete the perception view. Meanwhile, the existing state-of-the-art (SOTA) models are fully considered, including When2com [20], V2VNet [28], AttFuse [37], V2X-ViT [36], DiscoNet [17], CoBEVT [34], and Where2comm [10]. Intuitively, the proposed SCOPE outperforms No Fusion, Late Fusion, and Early Fusion by large margins across all datasets, demonstrating the superiority of our perception paradigm. Moreover, our method significantly outperforms previous SOTA models in simulated and real-world (DAIR-V2X) scenarios. In particular, the SOTA performance is improved by \\(3.63\\%\\) and \\(7.41\\%\\) on the V2XSet and OPV2V datasets separately" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.501 + ], + "angle": 0, + "content": "in AP@0.7. Compared to single-frame and collaboration-only based efforts [10, 17, 20, 28, 36], SCOPE proves the effectiveness and rationality of simultaneously considering the temporal context, collaboration, and ego characteristics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.893, + 0.7 + ], + "angle": 0, + "content": "Comparison of Communication Volume. Evaluating performance at varying bandwidths plays an essential role in achieving pragmatic perception. To this purpose, Figure 5 shows the perception performance of our SCOPE and Where2comm [10] with various communication volumes. Note that we do not consider communication methods that use any extra model/data/feature compression for a fair comparison. We have the following observations: (i) SCOPE achieves better performance than previous models [17, 20, 28, 36] on all datasets with less communication volume. (ii) Compared to Where2comm [10], our method obtains a superior perception-communication tradeoff across all the communication bandwidth choices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Robustness to Localization and Heading Errors. Following the noise settings from [36], we perform experiments to evaluate the robustness against natural localization noises of existing methods. The noises are sampled from a Gaussian distribution with standard deviation \\(\\sigma_{xyz} \\in [0, 0.5]\\) m, \\(\\sigma_{\\mathrm{heading}} \\in [0^{\\circ}, 1.0^{\\circ}]\\). From Figures 6 and 7, the overall collaborative perception performance consistently deteriorates as errors continue to increase. For instance, When2com and V2VNet fail even worse than No Fusion on the DAIR-V2X when the localization and heading errors are over 0.2 m and \\(0.4^{\\circ}\\), respectively. Noticeably, our method consistently surpasses the previous SOTA models under all noise levels, clearly displaying the robustness of SCOPE against" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23388" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.348, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.094, + 0.617, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.094, + 0.888, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.227, + 0.759, + 0.242 + ], + "angle": 0, + "content": "Figure 7. Robustness to the heading error on the DAIR-V2X, V2XSet, and OPV2V datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.256, + 0.47, + 0.311 + ], + "angle": 0, + "content": "Table 2. Ablation study results of the proposed core components on all the datasets. PL: pyramid LSTM; SIF: selective information filtering; DCM: deformable cross-attention module; RPP: reference point proposal; IAF: importance-aware adaptive fusion." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.312, + 0.468, + 0.417 + ], + "angle": 0, + "content": "
PLSIFDCMRPPIAFDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
60.31/40.3576.04/62.4177.22/65.51
✓61.75/44.5880.36/65.7481.19/69.02
✓✓62.48/45.6281.88/67.4282.34/71.23
✓✓✓64.06/47.8785.47/72.2986.96/76.84
✓✓✓✓64.74/48.4586.33/73.2788.01/78.57
✓✓✓✓✓65.18/49.8987.52/75.0589.71/80.62
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.507 + ], + "angle": 0, + "content": "pose errors. The reasonable explanations are: (i) the deformable cross-attention module facilitates capturing perceptually critical information across agents; (ii) the proposed multi-scale feature interaction alleviates the feature map misalignment about collaborative agents due to noises." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.242, + 0.53 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.584 + ], + "angle": 0, + "content": "We perform thorough ablation studies on all datasets to understand the necessity of the different designs. Tables 2 and 3 show the following observations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Importance of Core Components. We investigate the contribution of each core component in SCOPE. A base version (first row of Table 2) is provided as the gain reference. In the base model, we use the same fusion pattern to aggregate context information, collaborator-shared features, and multi-source features, where features are concatenated on the channel dimension and fused by a \\(1 \\times 1\\) convolutional network. As Table 2 demonstrates, we progressively add (1) Pyramid LSTM, (2) SIF, (3) DCM, (4) RPP, and (5) IAF and present the corresponding detection precision. The consistently increased detection results over the three datasets reveal the effectiveness of each proposed component, which all provide valuable performance gains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effect of Historical Frame Number. As shown in the top half of Table 3, we evaluate the effect of integrating different degrees of the temporal context from the ego agent. The interesting finding is that choosing a suitable historical frame count (i.e., 2 frames) is beneficial in achieving better performance gains. The potential reason is that adding excessive historical frames may lead to accumulated errors under the noise setting and degrade the performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.281 + ], + "angle": 0, + "content": "Table 3. Ablation study results of candidate designs and strategies on all the datasets. \"w/f\" and \"w/o\" mean the with and without." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.281, + 0.89, + 0.543 + ], + "angle": 0, + "content": "
Designs/StrategiesDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
Full Model65.18/49.8987.52/75.0589.71/80.62
Effect of Historical Frame Number
1 Frame64.28/49.1686.16/74.2188.85/78.49
2 Frames (Default)65.18/49.8987.52/75.0589.71/80.62
3 Frames65.03/49.2486.99/74.6589.67/80.45
Effect of Keypoint Number
9 Keypoints64.36/49.2286.34/74.0788.32/77.94
15 Keypoints (Default)65.18/49.8987.52/75.0589.71/80.62
21 Keypoints64.72/49.4087.12/74.6389.08/78.38
Importance of Distinct Features
w/o Ego-agent Featue64.96/49.3486.47/74.2488.63/79.38
w/o Context Feature64.14/48.8185.93/73.8888.26/77.18
w/o Collaboration Feature52.35/44.4663.91/45.3471.36/50.94
Effect of Fusion Strategies
w/ Summation Fusion63.65/48.9185.81/72.6488.05/76.13
w/ Maximum Fusion64.92/49.2786.28/73.6187.95/78.17
w/ Average Fusion64.85/48.9686.55/73.4688.24/78.73
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.669 + ], + "angle": 0, + "content": "Effect of Keypoint Number. We investigate the rationality of the keypoint number in the deformable crossattention. Noticeably, the DCM component with 15 keypoints achieves better detection performance than other settings. This result shows that SCOPE mitigates the impact of localization errors and fully extracts essential spatial semantics by setting a reasonable keypoint number." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Importance of Distinct Features. We explore the importance of distinct representations by removing the ego-agent, context, and collaboration features (i.e., \\( \\pmb{F}_i^{(t)}, \\pmb{H}_i^{(t)}, \\pmb{Z}_i^{(t)} \\)). There is a significant drop in the model's performance when the collaboration feature is removed, indicating that multiagent information sharing dominates various collaborative perception tasks. Furthermore, the consistently decreased performance suggests that ego-centered characteristics (i.e., pure ego-agent feature and corresponding context information) provide an indispensable contribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Effect of Fusion Strategies. Finally, we demonstrate the effectiveness of the proposed adaptive fusion strategy by providing candidate fusion mechanisms. Summation Fusion is a direct pixel-wise addition operation. Maximum and Average Fusion select elements at corresponding po" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23389" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.091, + 0.303, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.266, + 0.243, + 0.277 + ], + "angle": 0, + "content": "(a) DiscoNet" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.092, + 0.484, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.266, + 0.428, + 0.277 + ], + "angle": 0, + "content": "(b) V2X-ViT" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.092, + 0.666, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.266, + 0.616, + 0.277 + ], + "angle": 0, + "content": "(c) Where2comm" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.092, + 0.848, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.266, + 0.808, + 0.277 + ], + "angle": 0, + "content": "(d) SCOPE (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.281, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison results in real-world scenarios from the DAIR-V2X dataset. Green and red boxes denote ground truths and detection results, respectively. Compared to the previous SOTA models, our method achieves more accurate detection results." + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.319, + 0.303, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.426, + 0.26, + 0.437 + ], + "angle": 0, + "content": "GT Bounding Boxes" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.319, + 0.488, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.426, + 0.417, + 0.436 + ], + "angle": 0, + "content": "Scale One" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.319, + 0.665, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.426, + 0.602, + 0.436 + ], + "angle": 0, + "content": "Scale Two" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.319, + 0.849, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.426, + 0.788, + 0.436 + ], + "angle": 0, + "content": "Scale Three" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.441, + 0.893, + 0.469 + ], + "angle": 0, + "content": "Figure 9. Visualization of learned multi-scale deformable cross-attention. Green cross marker denotes the proposed reference point. The sampled keypoints are represented by the brown dots whose colors reflect the corresponding attention weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.609 + ], + "angle": 0, + "content": "sitions among features to retain the maximum or average, respectively, and then reshape them into a fused representation. The poor results in Table 3 show that the above mechanisms fail to tackle the complex multi-source feature fusion challenge. In contrast, the superiority of our IAF module is reflected in considering the complementary contributions of each feature at different positions, adaptively assigning higher importance to spatially beneficial semantics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.288, + 0.635 + ], + "angle": 0, + "content": "4.5. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Visualization of Detection Results. To intuitively illustrate the perception capability of different models, we provide a detection visualization comparison in Figure 8 under two challenging scenarios. SCOPE achieves more accurate and robust detection results than previous SOTA methods, where the predicted bounding boxes are well aligned to ground truths. In particular, DiscoNet [17], V2X-ViT [36], and Where2comm [10] fail to identify fast-moving objects (i.e., more ground truth bounding boxes are not matched). In contrast, our method achieves holistic perception across heterogeneous agents by effectively aggregating the salient information in the spatio-temporal domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Visualization of Attention Map. We visualize the multiscale deformable cross-attention in the CCC component. As Figure 9 shows, our tailored module can dynamically aggregate the semantics of multiple sampled keypoints at different scales to enhance the visual representation of the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.893, + 0.563 + ], + "angle": 0, + "content": "reference point. Concretely, the deformable cross-attention learns to extract object-related valuable perception cues at lower scales by sampling keypoints inside or around targets. In comparison, it reasonably samples long-range features at higher scales to capture spatially critical information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.619, + 0.591 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.893, + 0.799 + ], + "angle": 0, + "content": "In this paper, we present SCOPE, a learning-based framework to address existing multi-agent collaborative perception challenges in an end-to-end manner. Our approach is the first time considering the temporal semantics from ego agents to capture valuable context cues through the context-aware component. Moreover, two spatial information aggregation components are introduced to achieve comprehensive information interaction among agents and refined feature fusion of ego agents. Extensive experiments on diverse 3D detection datasets prove the superiority of SCOPE and the effectiveness of our components. The future work focuses on extending SCOPE into more complex perception scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.667, + 0.831 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.9 + ], + "angle": 0, + "content": "This work is supported by the Shanghai Key Research Laboratory of NSAI, the National Natural Science Foundation of China (Grant No. 62250410368), and the Nanjing First Automobile Works Grant." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23390" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3D point clouds. In Proceedings of ACM/IEEE Symposium on Edge Computing, pages 88-100, Arlington Virginia, Nov. 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.323 + ], + "angle": 0, + "content": "[3] Abhishek Das, Theophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. In International Conference on Machine Learning (ICML), pages 1538-1546, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.379 + ], + "angle": 0, + "content": "[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on Robot Learning (CoRL), pages 1-16, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.449 + ], + "angle": 0, + "content": "[5] Yangtao Du, Dingkang Yang, Peng Zhai, Mingchen Li, and Lihua Zhang. Learning associative representation for facial expression recognition. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 889-893, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.521 + ], + "angle": 0, + "content": "[6] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Yogamani. Rgb and lidar fusion based 3d semantic segmentation for autonomous driving. In Proceedings of the IEEE Intelligent Transportation Systems Conference (ITSC), pages 7-12, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[7] Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. Advances in Neural Information Processing Systems (NIPS), 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.47, + 0.605 + ], + "angle": 0, + "content": "[8] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735-1780, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.647 + ], + "angle": 0, + "content": "[9] Yedid Hoshen. Vain: Attentional multi-agent predictive modeling. Advances in Neural Information Processing Systems (NIPS), 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[10] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Advances in Neural Information Processing Systems (NIPS), 2022. 1, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[11] Yue Hu, Yifan Lu, Runsheng Xu, Weidi Xie, Siheng Chen, and Yanfeng Wang. Collaboration helps camera overtake lidar in 3d detection. arXiv preprint arXiv:2303.13560, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[12] Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent cooperation. Advances in Neural Information Processing Systems (NIPS), 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[13] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12697-12705, 2019. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Proceedings of the European Conference on Computer Vision (ECCV), pages 316-332, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[16] Jinlong Li, Runsheng Xu, Xinyu Liu, Jin Ma, Zicheng Chi, Jiaqi Ma, and Hongkai Yu. Learning for vehicle-to-vehicle cooperative perception under lossy communication. IEEE Transactions on Intelligent Vehicles, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[17] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems (NIPS), 34:29541-29552, 2021. 1, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[18] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 641-656, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[19] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[20] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4106-4115, 2020. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[21] Guiyang Luo, Hui Zhang, Quan Yuan, and Jinglin Li. Complementarity-enhanced and redundancy-minimized collaboration network for multi-agent perception. In Proceedings of the ACM International Conference on Multimedia (ACMM), pages 3578-3586, 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[22] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NIPS), 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[23] Tika Ram and Khem Chand. Effect of drivers' risk perception and perception of driving tasks on road safety attitude. Transportation Research Part F: Traffic Psychology and Behaviour, 42:162-176, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[24] Andreas Rauch, Felix Klanner, Ralph Rasshofer, and Klaus Dietmayer. Car2x-based perception in a high-level fusion architecture for cooperative perception systems. In IEEE Intelligent Vehicles Symposium, pages 270-275, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[25] Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in multiagent cooperative and competitive tasks. arXiv preprint arXiv:1812.09755, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[26] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation. Advances in Neural Information Processing Systems (NIPS), 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In International Conference on Machine Learning (ICML), pages 330-337, 1993. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23391" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[28] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 605–621, 2020. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[29] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. In IEEE International Conference on Robotics and Automation (ICRA), pages 3584-3591. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[30] Zhu Xizhou, Su Weijie, Lu Lewei, Li Bin, Wang Xiaogang, and Dai Jifeng. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.47, + 0.36 + ], + "angle": 0, + "content": "[31] Runsheng Xu, Weizhe Chen, Hao Xiang, Xin Xia, Lantao Liu, and Jiaqi Ma. Model-agnostic multi-agent perception framework. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1471–1478. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[32] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.476 + ], + "angle": 0, + "content": "[33] Runsheng Xu, Jinlong Li, Xiaoyu Dong, Hongkai Yu, and Jiaqi Ma. Bridging the domain gap for multi-agent perception. arXiv preprint arXiv:2210.08451, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.533 + ], + "angle": 0, + "content": "[34] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. In Conference on Robot Learning (CoRL), 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.47, + 0.618 + ], + "angle": 0, + "content": "[35] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13712–13722, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.47, + 0.688 + ], + "angle": 0, + "content": "[36] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[37] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In Proceedings of the International Conference on Robotics and Automation (ICRA), pages 2583–2589, 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[38] Dingkang Yang, Zhaoyu Chen, Yuzheng Wang, Shunli Wang, Mingcheng Li, Siao Liu, Xiao Zhao, Shuai Huang, Zhiyan Dong, Peng Zhai, and Lihua Zhang. Context deconfounded emotion recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19005-19015, June 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[39] Dingkang Yang, Shuai Huang, Haopeng Kuang, Yangtao Du, and Lihua Zhang. Disentangled representation learning for multimodal emotion recognition. In Proceedings of the 30th" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ACM International Conference on Multimedia (ACM MM), page 1642-1651, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[40] Dingkang Yang, Shuai Huang, Yang Liu, and Lihua Zhang. Contextual and cross-modal interaction for multi-modal speech emotion recognition. IEEE Signal Processing Letters, 29:2093-2097, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[41] Dingkang Yang, Shuai Huang, Shunli Wang, Yang Liu, Peng Zhai, Liuzhen Su, Mingcheng Li, and Lihua Zhang. Emotion recognition for multiple context awareness. In Proceedings of the European Conference on Computer Vision (ECCV), volume 13697, pages 144-162. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[42] Dingkang Yang, Haopeng Kuang, Shuai Huang, and Lihua Zhang. Learning modality-specific and -agnostic representations for asynchronous multimodal language sequences. In Proceedings of the 30th ACM International Conference on Multimedia (ACMM), page 1708-1717, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[43] Dingkang Yang, Yang Liu, Can Huang, Mingcheng Li, Xiao Zhao, Yuzheng Wang, Kun Yang, Yan Wang, Peng Zhai, and Lihua Zhang. Target and source modality co-reinforcement for emotion understanding from asynchronous multimodal sequences. Knowledge-Based Systems, page 110370, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[44] Kun Yang, Jing Liu, Dingkang Yang, Hanqi Wang, Peng Sun, Yanni Zhang, Yan Liu, and Liang Song. A novel efficient multi-view traffic-related object detection framework. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In IEEE International Conference on Distributed Computing Systems (ICDCS), pages 972-982, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[46] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21361-21370, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.642, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[47] Zhenxun Yuan, Xiao Song, Lei Bai, Zhe Wang, and Wanli Ouyang. Temporal-channel transformer for 3d lidar-based video object detection for autonomous driving. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2068-2078, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.713, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[48] Zixu Zhang and Jaime F Fisac. Safe occlusion-aware autonomous driving via game-theoretic active perception. arXiv preprint arXiv:2105.08169, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.756, + 0.892, + 0.824 + ], + "angle": 0, + "content": "[49] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 496-513, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23392" + } + ] +] \ No newline at end of file diff --git a/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/fe9d96d8-4150-4b6d-a27c-f4b80a133514_origin.pdf b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/fe9d96d8-4150-4b6d-a27c-f4b80a133514_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4120f0f6553a56fa8c3705a8ad99db6e26239354 --- /dev/null +++ b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/fe9d96d8-4150-4b6d-a27c-f4b80a133514_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68327429778c0d47962041c4633fd4af5dca5baebdc1299cae876ed2e9c62178 +size 2372793 diff --git a/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/full.md b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/full.md new file mode 100644 index 0000000000000000000000000000000000000000..88c53079286af129fada410db4f91df13cfeb971 --- /dev/null +++ b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/full.md @@ -0,0 +1,303 @@ +# Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception + +Kun Yang $^{1\dagger}$ Dingkang Yang $^{1\dagger}$ Jingyu Zhang $^{1}$ Mingcheng Li $^{1}$ Yang Liu $^{1}$ Jing Liu $^{1}$ Hanqi Wang $^{1}$ Peng Sun $^{2}$ Liang Song $^{1*}$ $^{1}$ Academy for Engineering and Technology, Fudan University $^{2}$ Duke Kunshan University {kunyang20,dkyang20,songl}@fudan.edu.cn + +# Abstract + +Multi-agent collaborative perception as a potential application for vehicle-to-everything communication could significantly improve the perception performance of autonomous vehicles over single-agent perception. However, several challenges remain in achieving pragmatic information sharing in this emerging research. In this paper, we propose SCOPE, a novel collaborative perception framework that aggregates the spatio-temporal awareness characteristics across on-road agents in an end-to-end manner. Specifically, SCOPE has three distinct strengths: i) it considers effective semantic cues of the temporal context to enhance current representations of the target agent; ii) it aggregates perceptually critical spatial information from heterogeneous agents and overcomes localization errors via multi-scale feature interactions; iii) it integrates multi-source representations of the target agent based on their complementary contributions by an adaptive fusion paradigm. To thoroughly evaluate SCOPE, we consider both real-world and simulated scenarios of collaborative 3D object detection tasks on three datasets. Extensive experiments show the superiority of our approach and the necessity of the proposed components. The project link is https://ydk122024.github.io/SCOPE/. + +# 1. Introduction + +Perceiving complex driving environments is essential for ensuring road safety [23] and traffic efficiency [21] in autonomous driving. Despite the rapid development in the perception ability of intelligent vehicles through deep learning technologies [5, 6, 14, 41, 43], the conventional single-agent perception paradigm remains challenging. Single-agent perception system is generally vulnerable and defective due to the inevitable practical difficulties on the road, + +such as restricted detection range and occlusion [47, 48]. To alleviate the inadequate observation constraints of isolated agents, multi-agent collaborative perception has emerged as a promising solution in recent years [17, 29]. With Vehicle-to-Everything (V2X) oriented communication applications, collaborative perception effectively promotes perspective complementation and information sharing among agents. + +Based on recent LiDAR-based 3D collaborative perception datasets [35, 36, 37, 46], several impressive works have achieved a trade-off between collaborative perception performance and communication bandwidth of Autonomous Vehicles (AVs) via well-designed mechanisms. These remarkable efforts include knowledge distillation-based intermediate collaboration [17], handshake communication mechanism [20], V2X-visual transformer [36], and spatial information selection strategy [10]. However, there is still considerable room for progress in achieving a pragmatic and effective collaborative perception. Concretely, existing methods invariably follow a single-frame static perception pattern, suffering from the data sparsity shortcoming of 3D point clouds [47] and ignoring meaningful semantic cues in the temporal context [49]. The insufficient representation of the moving objects in the current frame potentially restricts the perception performance of target vehicles. + +Moreover, spatial information aggregation in collaborative perception systems has exposed several problems. For the collaborator-shared message fusion, the per-agent/location-based fusion strategies of previous attempts [16, 17, 20, 28, 36] fail to deal with feature map misalignment from heterogeneous agents due to localization errors. Consequently, misleading features from collaborators could lead to object position misjudgment of the ego vehicle (i.e., receiver) and harm its detection precision. For the ego vehicle message refinement, existing approaches [10, 17, 20, 21, 28, 33] rely on fused representations with collaborator information to implement detection, abandoning the natural perception advantages of the ego vehicle and introducing the potential noises. The ego-centered characteristics may contain locally critical spatial information that is not disturbed by noise from collaborative agents + +![](images/367b6d237d71ba4996473067053e82e0dd9db99ac2d68dbea701532133eceebe.jpg) +Figure 1. The overall architecture of the proposed SCOPE. The framework consists of five parts: metadata conversion and feature extraction, context-aware information aggregation, confidence-aware cross-agent collaboration, importance-aware adaptive fusion, and detection decoders. The details of each individual component are illustrated in Section 3. + +with asynchronous sensor measurements. To this end, how to effectively break the aforementioned limitations becomes a priority for achieving robust collaborative perception. + +Motivated by the above observations, we present SCOPE, a Spatio-temporal domain awareness multi-agent CollabOrative PErception approach to tackle existing challenges jointly. From Figure 1, SCOPE achieves effective information collaboration among agents and precise perception of the ego agent via three proposed core components. Specifically, (i) for the data sparsity challenge in single-frame point clouds, the context-aware information aggregation is proposed to aggregate the ego agent's context information from preceding frames. We employ selective information filtering and spatio-temporal feature integration modules to capture informative temporal cues and historical context semantics. (ii) For the collaborator-shared message fusion challenge, the confidence-aware cross-agent collaboration is introduced to guarantee that the ego agent aggregates complementary information from heterogeneous agents. Based on the confidence-aware multi-scale feature interaction, we facilitate holistic communication and mitigate feature map misalignment due to localization errors of collaborators. (iii) For the multi-feature integration challenge of the ego agent, the importance-aware adaptive fusion is designed to flexibly incorporate the advantages of diverse representations based on distinctive perception contributions. We evaluate the superiority of SCOPE on various collaborative 3D object detection datasets, including DAIR-V2X [46], OPV2V [37], and V2XSet [36]. Extensive experiments in real-world and simulated scenarios provide favorable evidence that our approach is competitive in + +multi-agent collaborative perception tasks. + +The main contributions can be summarized as follows: + +- We present SCOPE, a novel framework for multi-agent collaborative perception. The framework facilitates information collaboration and feature fusion among agents, achieving a reasonable performance-bandwidth trade-off. Comprehensive experiments on collaborative detection tasks show that SCOPE outperforms previous state-of-the-art methods. +- To our best knowledge, we are the first to consider the temporal context of the ego agent in collaborative perception systems. Based on the proposed context-aware component, the target frame of point clouds efficiently integrates historical cues from preceding frames to capture valuable temporal information. +- We introduce two spatial information aggregation components to address the challenges of collaboration heterogeneity and fused representation unicity. These tailored components effectively perform multigrained information interactions among agents and multi-source feature refinement of the ego agent. + +# 2. Related Work + +# 2.1. Multi-Agent Communication + +Benefiting from rapid advances in learning-based technologies [18, 38, 39, 40, 42, 44, 45], communication has become an essential component in constructing robust multiagent systems. Previous works [7, 25, 26, 27] typically focus on centralized paradigms or predefined protocols to + +explore information-sharing mechanisms among multiple agents. However, these efforts are potentially domain-constrained and scenario-isolated. Recent methods [3, 9, 12] achieve effective communication strategies in diverse scenarios. For instance, Vain [9] determines which agents would share information by modeling the locality of interactions. ATOC [12] introduces an attentional communication strategy to learn when communicating and integrating useful information to execute collaborative decisions. In comparison, we focus on the more challenging 3D perception tasks in complex driving scenarios. Based on the proposed confidence-aware multi-scale information interaction, SCOPE achieves an information-critical and holistic communication pattern among heterogeneous agents. + +# 2.2. Collaborative Perception + +Collaborative perception investigates how to aggregate complementary perception semantics across connected agents to improve overall system performance. Current works are mainly categorized as early [2], intermediate [1, 10, 17, 20, 28, 34, 36, 37], and late [24, 31] fusion, where intermediate fusion is more favoured due to its good trade-off between performance and transmission bandwidth. In this case, several remarkable works explore efficient collaboration mechanisms. In DiscoNet [17], the teacher-student network is trained to aggregate the shared information. V2X-ViT [36] proposes heterogeneous multi-agent self-attention to handle device heterogeneity, and Xu et al. [34] design an axial attention module to capture sparse spatial semantics. Furthermore, Where2comm [10] filters irrelevant semantics by quantifying the spatial importance through confidence maps. Unlike these single-frame methods, we propose an information aggregation component to extract rich context information from local historical observations. Moreover, our cross-agent collaboration mechanism outperforms the previous per-agent/location message fusion models, significantly mitigating the impact of localization errors. + +# 3. Methodology + +In this paper, we focus on developing a robust collaborative perception system to improve the ego vehicle's perception capability. Figure 1 illustrates the overall architecture of the proposed framework. It consists of five sequential parts, where the system input accommodates different types of connected agents (i.e., infrastructures and AVs). + +# 3.1. Metadata Conversion and Feature Extraction + +During the collaboration, one of the connected agents is identified as the ego agent $(i)$ to build a communication graph, where other agents within the communication range serve as collaborators. Upon receiving the metadata $(e.g.,$ poses and extrinsics) broadcast by the ego agent, the collaborators project their local LiDAR point clouds + +![](images/39b6fb7b98a911884f35aee76d423898913923fe38a44a77a2bfd7a535440964.jpg) +Figure 2. (a) The architecture of the proposed CIA component, including selective information filtering (left) and spatio-temporal feature integration (middle). (b) The multi-scale convolutional structure in the pyramid LSTM. + +![](images/9c047eff5af653d2c89c4ebdc503f8f9658027d248b9a5076e7837d6b585fae4.jpg) + +to the coordinate system of the ego agent. Similarly, the ego agent's previous point cloud frames are synchronized to the current frame. Each agent $k \in \{1, \dots, K\}$ encodes the projected point clouds into Bird's Eye View (BEV) features to extract local visual representations. Given the point cloud $\mathbf{X}_k^{(t)}$ of $k$ -th agent at timestamp $t$ , the extracted feature is $\mathbf{F}_k^{(t)} = f_{enc}(\mathbf{X}_k^{(t)}) \in \mathbb{R}^{C \times H \times W}$ , where $f_{enc}(\cdot)$ is the PointPillar [14] encoder shared among all agents, and $C, H, W$ stand for the channel, height, and width. + +# 3.2. Context-aware Information Aggregation + +Conventional 3D detection tasks [47, 49] have shown the potential of temporal information in improving perception performance. Unfortunately, current multi-agent collaborative efforts [10, 17, 20, 21, 28] invariably focus on exploring the current frame while ignoring context cues from previous frames. The single-frame solution fails to effectively detect fast-moving objects (e.g., surrounding vehicles) due to the sparsity and insufficiency of point clouds. To this end, we present a Context-aware Information Aggregation (CIA) component to capture spatio-temporal representations of the ego agent's previous frames to fuse valuable semantics. As shown in Figure 2(a), CIA consists of two core phases. + +Selective Information Filtering. This phase aims to extract refined information by filtering redundant features from the target $\pmb{F}_i^{(t)} \in \mathbb{R}^{C \times H \times W}$ and distilling meaningful characteristics from the previous $\pmb{F}_i^{(t - \tau)} \in \mathbb{R}^{\tau \times C \times H \times W}$ , where $\tau$ is the time shift. Concretely, the channel-wise average and max pooling operations $\Psi_{a / m}(\cdot)$ are first utilized to aggregate rich channel semantics, denoted as $\mathcal{M}_{a / m}^{(t)} = \Psi_{a / m}(\pmb{F}_i^{(t)})$ and $\mathcal{M}_{a / m}^{(\tau)} = \Psi_{a / m}(\pmb{F}_i^{(\tau)})$ , where $\pmb{F}_i^{(\tau)} = \sum_{n=1}^{\tau} \pmb{F}_i^{(t-n)}$ . These refined feature maps are combined to obtain a well-chosen spatial selection map. The procedure is expressed as follows: + +$$ +\mathcal {U} = \sigma \cdot \aleph (\mathcal {M} _ {a} ^ {(t)} \| \mathcal {M} _ {m} ^ {(t)} + \mathcal {M} _ {a} ^ {(\tau)} \| \mathcal {M} _ {m} ^ {(\tau)}) \in \mathbb {R} ^ {H \times W}, \quad (1) +$$ + +where $\sigma$ is the sigmoid activation, $\aleph$ denotes a convolutional layer for fusion, and $\parallel$ denotes the concatenation op + +![](images/63e56030fa655eb97aab668dd2b6ed5c5b81ea5e717f094b6119abf2b0feae59.jpg) +Figure 3. (a) The architecture of the proposed CCC component. (b) CCC consists of two phases at each scale: reference point proposal (middle) and deformable cross-attention module (right). + +eration. Immediately, the information filtering strategy is implemented via pixel-wise attention weighting to enhance the salient historical features of $H_{i}^{(\tau)} \in \mathbb{R}^{\tau \times C \times H \times W}$ : + +$$ +\boldsymbol {H} _ {i} ^ {(\tau)} = (1 - \mathcal {U}) \odot \tanh \left(\boldsymbol {F} _ {i} ^ {(t)}\right) + \mathcal {U} \odot \boldsymbol {F} _ {i} ^ {(t - \tau)}. \tag {2} +$$ + +Spatio-temporal Feature Integration. In order to integrate the historical prototype to improve the perception ability of the current representation, we introduce a pyramid LSTM [15] to learn the contextual dependencies of interframe features $\pmb{F}_i^{(t)}\| \pmb{H}_i^{(\tau)}$ . The difference between the pyramid LSTM and vanilla LSTM [8] is that the initial matrix multiplication is replaced with a multi-scale convolutional structure, as shown in Figure 2(b). Note that the intrinsic long short-term memory mechanism in the LSTM is adopted, and please refer to [8] for details. In practice, multi-scale spatial features are extracted by two successive 2D convolutions among different scales, followed by batch normalization and ReLU activation. To achieve multilevel spatial semantic fusion, the downsampled features are progressively interpolated to deliver to the upsampled layers via lateral connections. The overall design facilitates learning spatio-temporal correlations in historical contexts, which benefits our multi-frame perception paradigm. Ultimately, the final hidden state in the pyramid LSTM is used as the refined context-aware feature $\pmb{H}_i^{(t)}\in \mathbb{R}^{C\times H\times W}$ + +# 3.3. Confidence-aware Cross-agent Collaboration + +The cross-agent collaboration targets to enhance the ego agent's visual representations by aggregating the complementary semantics from the collaborator-shared features. Existing works [10, 11, 36, 37] propose attention-based perlocation feature fusion methods, which are vulnerable to localization errors and ignore the sparsity of point clouds. To tackle these issues, we implement a novel Confidence-aware Cross-agent Collaboration (CCC) component. We first obtain the confidence map $S_{k}^{(t)}$ and binary selection matrix $\mathcal{C}_k^{(t)}$ of feature $F_{k}^{(t)}, k\in \{1,\dots,K\}$ following [10]. + +The filtered feature shared by $k$ -th agent to ego agent is $\pmb{F}_k^{(t)} = \pmb{F}_k^{(t)} \odot \mathcal{C}_k^{(t)}$ . As Figure 3(a) shows, the CCC component encodes the features and confidence maps into three scales and performs feature fusion at each scale. $\pmb{F}_{k,l}^{(t)}$ and $S_{k,l}^{(t)}$ denote the feature and confidence map at $l$ -th scale. We take two agents as an example to elaborate on the feature fusion in Figure 3(b), including the following two phases. + +Reference Point Proposal. To predict the spatial locations containing meaningful object information, we obtain the element-wise summation of all confidence maps as $S_{sum,l}^{(t)} = \sum_{k=1}^{K} S_{k,l}^{(t)}$ . Since the confidence map reflects the spatially critical levels, $S_{sum,l}^{(t)}$ presents the potential positions of the targets within the detection range, called the reference points. Accordingly, we apply a threshold-based selection function $f_{sel}(\cdot)$ to extract the reference points $S_{re,l}^{(t)}$ . This design could actively guide the subsequent fusion network to concentrate on the essential spatial areas. + +Deformable Cross-attention Module. We extract the ego agent's feature $\pmb{F}_{i,l}^{(t)}$ at the reference points as the initial query embedding and apply a linear layer to encode the locations of reference points into a position embedding. To resolve the misalignment of feature maps and acquire more robust representations against localization errors, the Deformable Cross-attention Module (DCM) aggregates the information from sampled keypoints through a deformable cross-attention layer [30]. Concretely, a linear layer is leveraged to learn the 2D spatial offset $\Delta q$ of the reference point $q$ for sampling attending keypoints at $q + \Delta q$ . We simultaneously select $M$ keypoints and extract their features as the attending features. The output of DCM at position $q$ is: + +$$ +\operatorname {D C M} (q) = \sum_ {a = 1} ^ {A} \boldsymbol {W} _ {a} \left[ \sum_ {k = 1} ^ {K} \sum_ {m = 1} ^ {M} \phi \left(\boldsymbol {W} _ {b} \boldsymbol {F} _ {i, l} ^ {(t)} (q)\right) \boldsymbol {F} _ {k, l} ^ {(t)} \left(q + \Delta q _ {m}\right) \right], \tag {3} +$$ + +where $a$ indexes the attention head, $\mathbf{W}_{a / b}$ are the learnable weights, and $\phi (\cdot)$ is the softmax function. Ultimately, we derive a filling operation that fills $\mathrm{DCM}(q)$ into the ego agent's feature $\pmb{F}_{i,l}^{(t)}$ based on the initial position $q$ and outputs $\pmb{Z}_{i,l}^{(t)}$ . The output-enhanced features under the three scales are encoded into the same size and concatenated in the channel dimension. We utilize a $1\times 1$ convolutional layer to fuse information of the three scales and get the final collaboration feature $\pmb{Z}_i^{(t)}\in \mathbb{R}^{C\times H\times W}$ . + +# 3.4. Importance-aware Adaptive Fusion + +Although previous works [17, 20, 21, 28, 36] have achieved impressive performance through ego features that aggregate collaborative information, they potentially suffer from noise interference introduced by collaborators with asynchronous measurements. A promising solution is to consider purely ego-centered characteristics, which contain the natural perception advantages of the target agent. To + +![](images/26dbb363aa27207230655e7bc0098c7122196a976e6b1a158ff9d7e8cf002baa.jpg) +Figure 4. The architecture of the proposed IAF component. + +this end, we propose an Importance-aware Adaptive Fusion (IAF) component to fuse multi-source features based on their complementary contributions. These informative features include $\pmb{H}_i^{(t)}, \pmb{Z}_i^{(t)}$ , and $\pmb{F}_i^{(t)}$ . From Figure 4, the importance generator $f_{gen}(\cdot)$ is used to generate the spatial importance maps, which processes as follows: + +$$ +\mathcal {H} _ {i} ^ {(t)}, \mathcal {Z} _ {i} ^ {(t)}, \mathcal {S} _ {i} ^ {(t)} = \sigma \cdot \Psi_ {m} (f _ {g e n} (\boldsymbol {H} _ {i} ^ {(t)}, \boldsymbol {Z} _ {i} ^ {(t)}, \boldsymbol {F} _ {i} ^ {(t)})), \quad (4) +$$ + +where $\mathcal{H}_i^{(t)},\mathcal{Z}_i^{(t)},S_i^{(t)}\in [0,1]^{H\times W}$ . Then, these importance maps are concatenated to get the normalized attention maps across distinct features by the softmax function $\phi (\cdot)$ : + +$$ +\mathcal {E} _ {\mathcal {H}} ^ {(t)} = \phi \left(\mathcal {H} _ {i} ^ {(t)}\right) = \frac {\exp \left(\mathcal {H} _ {i} ^ {(t)}\right)}{\exp \left(\mathcal {H} _ {i} ^ {(t)}\right) + \exp \left(\mathcal {Z} _ {i} ^ {(t)}\right) + \exp \left(\mathcal {S} _ {i} ^ {(t)}\right)}. \tag {5} +$$ + +A larger value in $\mathcal{E}_{\mathcal{H}}^{(t)} \in \mathbb{R}^{H \times W}$ implies that the corresponding feature position is more significant. Similarly, $\mathcal{E}_{\mathcal{Z}}^{(t)} = \phi(\mathcal{Z}_i^{(t)})$ and $\mathcal{E}_S^{(t)} = \phi(S_i^{(t)})$ . Finally, we adaptively activate each feature map's perceptually critical spatial information by a weighted summation. The fused feature map $\mathcal{F}_i^{(t)} \in \mathbb{R}^{C \times H \times W}$ is obtained as follows: + +$$ +\mathcal {F} _ {i} ^ {(t)} = \mathcal {E} _ {\mathcal {H}} ^ {(t)} \odot \boldsymbol {H} _ {i} ^ {(t)} + \mathcal {E} _ {\mathcal {Z}} ^ {(t)} \odot \boldsymbol {Z} _ {i} ^ {(t)} + \mathcal {E} _ {\mathcal {S}} ^ {(t)} \odot \boldsymbol {F} _ {i} ^ {(t)}. \tag {6} +$$ + +# 3.5. Detection Decoder and Objective Optimization + +We leverage two detection decoders to decode the final fused feature $\mathcal{F}_i^{(t)}$ into the predicted outputs. The regression output is $\mathcal{Y}_{i,r}^{(t)} = f_{dec}^r (\mathcal{F}_i^{(t)})\in \mathbb{R}^{7\times H\times W}$ , representing the position, size, and yaw angle of the bounding box at each location. The classification output reveals the confidence value of each predefined box to be a target or background, which is $\mathcal{Y}_{i,c}^{(t)} = f_{dec}^c (\mathcal{F}_i^{(t)})\in \mathbb{R}^{2\times H\times W}$ . Following existing work [14], we adopt the smooth $L_{1}$ loss for regressing the bounding boxes and a focal loss [19] for classification. + +# 4. Experiments + +# 4.1. Datasets and Evaluation Metrics + +Datasets. We conduct comprehensive experiments on three benchmark datasets for collaborative perception tasks, in + +including DAIR-V2X [46], V2XSet [36], and OPV2V [37]. These datasets provide annotated sensor data from multiple connected agents for each sample. Concretely, DAIR-V2X is a large-scale real-world dataset supporting collaborative 3D object detection, containing 71,254 samples and dividing the training/validation/testing sets as 5:2:3. Each sample contains the LiDAR point clouds from a vehicle and an infrastructure. During feature extraction, we transform the point cloud into the BEV feature map of the size (200, 504, 64). V2XSet is a public simulation dataset for V2X collaborative perception, including 73 representative scenarios and 11,447 frames of labeled point clouds generated by Carla [4]. The training/validation/testing sets are 6,694, 1,920, and 2,833 frames, respectively. The size of each extracted feature map is (100, 384, 64). OPV2V is a vehicle-to-vehicle collaborative perception dataset collected by Carla and OpenCDA [32]. It includes 11,464 frames of point clouds and RGB images with 3D annotation, splitting into training/validation/testing sets of 6,764, 1,981, and 2,719 frames, respectively. The 3D point cloud is extracted as a feature map of the size (100, 384, 64). + +Evaluation Metrics. Following [36], we leverage the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the 3D object detection performance. The communication results count the message size by byte in the log scale with base 2 [10]. + +# 4.2. Implementation Details. + +We build the proposed models on the Pytorch toolbox [22] and train them on four Nvidia Tesla V100 GPUs using the Adam optimizer [13]. The training settings of the DAIR-V2X, V2XSet, and OPV2V datasets follow: the batch sizes are $\{5, 3, 3\}$ and the epochs are $\{30, 40, 40\}$ . The initial learning rate is $2\mathrm{e} - 3$ and decays every 15 epochs with a factor of 0.1. Our CIA module uses the previous 2 frames, the attention head in the CCC component is 8, and the number of keypoints is 15. To simulate the default localization and heading errors, we add Gaussian noise with a standard deviation of $0.2\mathrm{m}$ and $0.2^{\circ}$ to the initial position and orientation of the agents. The voxel resolution of the observation encoder $f_{enc}(\cdot)$ is set to $0.4\mathrm{m}$ in height and width. The generator $f_{gen}(\cdot)$ follows a detection decoder from [14], and the summed confidence map $S_{sum,l}^{(t)}$ is filtered using a threshold of 0.1 in the function $f_{sel}(\cdot)$ . The decoders $\{f_{dec}^{r}(\cdot), f_{dec}^{c}(\cdot)\}$ are derived with two $1 \times 1$ convolutional layers. We provide results on the validation set of DAIR-V2X and the testing set of V2XSet and OPV2V. + +# 4.3.Quantitative Evaluation + +Detection Performance Comparison. Table 1 shows the 3D detection performance comparison results on the three datasets. We compare the following methods. No Fusion as a baseline considers only the point cloud data of ego vehi + +![](images/a89cbb5d8c3de522e8f7b24f4d4ba2745219562dbc5ecd2f1603891c0af0fc05.jpg) +Figure 5. Collaborative perception performance comparison of SCOPE and Where2comm [10] on the DAIR-V2X, V2XSet and OPV2V datasets with varying communication volume. + +![](images/e451d18b2479306ccbff34efb6869ab379bac8a0f5854bd7c596591249867bed.jpg) + +![](images/10bcef53aa60ef03160c33255cafd189b73cf887ba1c6f361c150f14cc4ef2cc.jpg) + +![](images/866dc6a87604fdd1680e42da86f0d9b393df137582b6a5173ebef319e79a059c.jpg) +Figure 6. Robustness to the localization error on the DAIR-V2X, V2XSet, and OPV2V datasets. + +![](images/c9734047217852d965ebab3c7a3f58b121229c521afc331558d263ff75920525.jpg) + +![](images/6dc8fe174cb7352ae503e9020ba8ea9ab9d6addd7ee2ca13507dd1ddcd8a79ac.jpg) + +Table 1. Performance comparison on the DAIR-V2X, V2XSet, and OPV2V datasets. The results are reported in AP@0.5/0.7. + +
ModelDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
No Fusion50.03/43.5760.60/40.2068.71/48.66
Late Fusion53.12/37.8866.79/50.9582.24/65.78
Early Fusion61.74/46.5377.39/50.4581.30/68.69
When2com [20]51.12/36.1770.16/53.7277.85/62.40
V2VNet [28]56.01/42.2581.80/61.3582.79/70.31
AttFuse [37]53.79/42.6176.27/57.9383.21/70.09
V2X-ViT [36]54.26/43.3585.13/68.6786.72/74.94
DiscoNet [17]54.29/44.8882.18/63.7387.38/73.19
CoBEVT [34]54.82/43.9583.01/62.6787.40/74.35
Where2comm [10]63.71/48.9385.78/72.4288.07/75.06
SCOPE (Ours)65.18/49.8987.52/75.0589.71/80.62
+ +cles without collaboration. Late Fusion allows agents to exchange detected results and utilizes non-maximum suppression to produce the final outputs. Early Fusion means the raw LiDAR point clouds are shared to complete the perception view. Meanwhile, the existing state-of-the-art (SOTA) models are fully considered, including When2com [20], V2VNet [28], AttFuse [37], V2X-ViT [36], DiscoNet [17], CoBEVT [34], and Where2comm [10]. Intuitively, the proposed SCOPE outperforms No Fusion, Late Fusion, and Early Fusion by large margins across all datasets, demonstrating the superiority of our perception paradigm. Moreover, our method significantly outperforms previous SOTA models in simulated and real-world (DAIR-V2X) scenarios. In particular, the SOTA performance is improved by $3.63\%$ and $7.41\%$ on the V2XSet and OPV2V datasets separately + +in AP@0.7. Compared to single-frame and collaboration-only based efforts [10, 17, 20, 28, 36], SCOPE proves the effectiveness and rationality of simultaneously considering the temporal context, collaboration, and ego characteristics. + +Comparison of Communication Volume. Evaluating performance at varying bandwidths plays an essential role in achieving pragmatic perception. To this purpose, Figure 5 shows the perception performance of our SCOPE and Where2comm [10] with various communication volumes. Note that we do not consider communication methods that use any extra model/data/feature compression for a fair comparison. We have the following observations: (i) SCOPE achieves better performance than previous models [17, 20, 28, 36] on all datasets with less communication volume. (ii) Compared to Where2comm [10], our method obtains a superior perception-communication tradeoff across all the communication bandwidth choices. + +Robustness to Localization and Heading Errors. Following the noise settings from [36], we perform experiments to evaluate the robustness against natural localization noises of existing methods. The noises are sampled from a Gaussian distribution with standard deviation $\sigma_{xyz} \in [0, 0.5]$ m, $\sigma_{\mathrm{heading}} \in [0^{\circ}, 1.0^{\circ}]$ . From Figures 6 and 7, the overall collaborative perception performance consistently deteriorates as errors continue to increase. For instance, When2com and V2VNet fail even worse than No Fusion on the DAIR-V2X when the localization and heading errors are over 0.2 m and $0.4^{\circ}$ , respectively. Noticeably, our method consistently surpasses the previous SOTA models under all noise levels, clearly displaying the robustness of SCOPE against + +![](images/f11b5a7e055828d46820c0eb0fa0c641e2d2f092501d0fb7aa62a64c3997cc33.jpg) +Figure 7. Robustness to the heading error on the DAIR-V2X, V2XSet, and OPV2V datasets. + +![](images/452945dfc3e6d64ec3cf1d46f6a6ec8b89db3b7e2f6835a2e7e528f945bc7e0c.jpg) + +![](images/7f757639bab00616afa0b5fd38075ec66444ad6ff0aecf46d3f6585e478e1333.jpg) + +Table 2. Ablation study results of the proposed core components on all the datasets. PL: pyramid LSTM; SIF: selective information filtering; DCM: deformable cross-attention module; RPP: reference point proposal; IAF: importance-aware adaptive fusion. + +
PLSIFDCMRPPIAFDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
60.31/40.3576.04/62.4177.22/65.51
✓61.75/44.5880.36/65.7481.19/69.02
✓✓62.48/45.6281.88/67.4282.34/71.23
✓✓✓64.06/47.8785.47/72.2986.96/76.84
✓✓✓✓64.74/48.4586.33/73.2788.01/78.57
✓✓✓✓✓65.18/49.8987.52/75.0589.71/80.62
+ +pose errors. The reasonable explanations are: (i) the deformable cross-attention module facilitates capturing perceptually critical information across agents; (ii) the proposed multi-scale feature interaction alleviates the feature map misalignment about collaborative agents due to noises. + +# 4.4. Ablation Studies + +We perform thorough ablation studies on all datasets to understand the necessity of the different designs. Tables 2 and 3 show the following observations. + +Importance of Core Components. We investigate the contribution of each core component in SCOPE. A base version (first row of Table 2) is provided as the gain reference. In the base model, we use the same fusion pattern to aggregate context information, collaborator-shared features, and multi-source features, where features are concatenated on the channel dimension and fused by a $1 \times 1$ convolutional network. As Table 2 demonstrates, we progressively add (1) Pyramid LSTM, (2) SIF, (3) DCM, (4) RPP, and (5) IAF and present the corresponding detection precision. The consistently increased detection results over the three datasets reveal the effectiveness of each proposed component, which all provide valuable performance gains. + +Effect of Historical Frame Number. As shown in the top half of Table 3, we evaluate the effect of integrating different degrees of the temporal context from the ego agent. The interesting finding is that choosing a suitable historical frame count (i.e., 2 frames) is beneficial in achieving better performance gains. The potential reason is that adding excessive historical frames may lead to accumulated errors under the noise setting and degrade the performance. + +Table 3. Ablation study results of candidate designs and strategies on all the datasets. "w/f" and "w/o" mean the with and without. + +
Designs/StrategiesDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
Full Model65.18/49.8987.52/75.0589.71/80.62
Effect of Historical Frame Number
1 Frame64.28/49.1686.16/74.2188.85/78.49
2 Frames (Default)65.18/49.8987.52/75.0589.71/80.62
3 Frames65.03/49.2486.99/74.6589.67/80.45
Effect of Keypoint Number
9 Keypoints64.36/49.2286.34/74.0788.32/77.94
15 Keypoints (Default)65.18/49.8987.52/75.0589.71/80.62
21 Keypoints64.72/49.4087.12/74.6389.08/78.38
Importance of Distinct Features
w/o Ego-agent Featue64.96/49.3486.47/74.2488.63/79.38
w/o Context Feature64.14/48.8185.93/73.8888.26/77.18
w/o Collaboration Feature52.35/44.4663.91/45.3471.36/50.94
Effect of Fusion Strategies
w/ Summation Fusion63.65/48.9185.81/72.6488.05/76.13
w/ Maximum Fusion64.92/49.2786.28/73.6187.95/78.17
w/ Average Fusion64.85/48.9686.55/73.4688.24/78.73
+ +Effect of Keypoint Number. We investigate the rationality of the keypoint number in the deformable crossattention. Noticeably, the DCM component with 15 keypoints achieves better detection performance than other settings. This result shows that SCOPE mitigates the impact of localization errors and fully extracts essential spatial semantics by setting a reasonable keypoint number. + +Importance of Distinct Features. We explore the importance of distinct representations by removing the ego-agent, context, and collaboration features (i.e., $\pmb{F}_i^{(t)}, \pmb{H}_i^{(t)}, \pmb{Z}_i^{(t)}$ ). There is a significant drop in the model's performance when the collaboration feature is removed, indicating that multiagent information sharing dominates various collaborative perception tasks. Furthermore, the consistently decreased performance suggests that ego-centered characteristics (i.e., pure ego-agent feature and corresponding context information) provide an indispensable contribution. + +Effect of Fusion Strategies. Finally, we demonstrate the effectiveness of the proposed adaptive fusion strategy by providing candidate fusion mechanisms. Summation Fusion is a direct pixel-wise addition operation. Maximum and Average Fusion select elements at corresponding po + +![](images/04fd1d07398a19438ad2d4faa76ed8e950b1d39da7f8222e34097cb4b8737187.jpg) +(a) DiscoNet + +![](images/a4758aec46258a60fe5f4d4d55718824214a96fac70a6c7e93d98f1857c50d22.jpg) +(b) V2X-ViT + +![](images/bfb1871e46055a4ed03979d1c9c674a52be3855f7632bf2897b7cbd5b599653d.jpg) +(c) Where2comm + +![](images/c335617087060a539efecf0a6fd025d4ab0ea6dcf9db90ef39a018a93fd71b1f.jpg) +(d) SCOPE (ours) + +![](images/0191c4efa9eb134c2cf71b8fd9aaf00c156435b469c5f96577c97aff3f497807.jpg) +Figure 8. Qualitative comparison results in real-world scenarios from the DAIR-V2X dataset. Green and red boxes denote ground truths and detection results, respectively. Compared to the previous SOTA models, our method achieves more accurate detection results. +GT Bounding Boxes +Figure 9. Visualization of learned multi-scale deformable cross-attention. Green cross marker denotes the proposed reference point. The sampled keypoints are represented by the brown dots whose colors reflect the corresponding attention weights. + +![](images/c5e140de898e6e8c59eeb909dd2f4560ef1f4a11e71e8c6c5e0f89aacc99a82d.jpg) +Scale One + +![](images/dac8ab962b3d0b89d684f4a857bf9e413d1cbbd42d5848e650b748dd1442c607.jpg) +Scale Two + +![](images/c984beec5d3048727ba5502cd28d9c5988311d57e50b10b45f1c31e7a3471238.jpg) +Scale Three + +sitions among features to retain the maximum or average, respectively, and then reshape them into a fused representation. The poor results in Table 3 show that the above mechanisms fail to tackle the complex multi-source feature fusion challenge. In contrast, the superiority of our IAF module is reflected in considering the complementary contributions of each feature at different positions, adaptively assigning higher importance to spatially beneficial semantics. + +# 4.5. Qualitative Evaluation + +Visualization of Detection Results. To intuitively illustrate the perception capability of different models, we provide a detection visualization comparison in Figure 8 under two challenging scenarios. SCOPE achieves more accurate and robust detection results than previous SOTA methods, where the predicted bounding boxes are well aligned to ground truths. In particular, DiscoNet [17], V2X-ViT [36], and Where2comm [10] fail to identify fast-moving objects (i.e., more ground truth bounding boxes are not matched). In contrast, our method achieves holistic perception across heterogeneous agents by effectively aggregating the salient information in the spatio-temporal domains. + +Visualization of Attention Map. We visualize the multiscale deformable cross-attention in the CCC component. As Figure 9 shows, our tailored module can dynamically aggregate the semantics of multiple sampled keypoints at different scales to enhance the visual representation of the + +reference point. Concretely, the deformable cross-attention learns to extract object-related valuable perception cues at lower scales by sampling keypoints inside or around targets. In comparison, it reasonably samples long-range features at higher scales to capture spatially critical information. + +# 5. Conclusion + +In this paper, we present SCOPE, a learning-based framework to address existing multi-agent collaborative perception challenges in an end-to-end manner. Our approach is the first time considering the temporal semantics from ego agents to capture valuable context cues through the context-aware component. Moreover, two spatial information aggregation components are introduced to achieve comprehensive information interaction among agents and refined feature fusion of ego agents. Extensive experiments on diverse 3D detection datasets prove the superiority of SCOPE and the effectiveness of our components. The future work focuses on extending SCOPE into more complex perception scenarios. + +# Acknowledgements + +This work is supported by the Shanghai Key Research Laboratory of NSAI, the National Natural Science Foundation of China (Grant No. 62250410368), and the Nanjing First Automobile Works Grant. + +# References + +[1] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3D point clouds. In Proceedings of ACM/IEEE Symposium on Edge Computing, pages 88-100, Arlington Virginia, Nov. 2019. 3 +[2] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524, 2019. 3 +[3] Abhishek Das, Theophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. In International Conference on Machine Learning (ICML), pages 1538-1546, 2019. 3 +[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on Robot Learning (CoRL), pages 1-16, 2017. 5 +[5] Yangtao Du, Dingkang Yang, Peng Zhai, Mingchen Li, and Lihua Zhang. Learning associative representation for facial expression recognition. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 889-893, 2021. 1 +[6] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Yogamani. Rgb and lidar fusion based 3d semantic segmentation for autonomous driving. In Proceedings of the IEEE Intelligent Transportation Systems Conference (ITSC), pages 7-12, 2019. 1 +[7] Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. Advances in Neural Information Processing Systems (NIPS), 29, 2016. 2 +[8] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735-1780, 1997. 4 +[9] Yedid Hoshen. Vain: Attentional multi-agent predictive modeling. Advances in Neural Information Processing Systems (NIPS), 30, 2017. 3 +[10] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Advances in Neural Information Processing Systems (NIPS), 2022. 1, 3, 4, 5, 6, 8 +[11] Yue Hu, Yifan Lu, Runsheng Xu, Weidi Xie, Siheng Chen, and Yanfeng Wang. Collaboration helps camera overtake lidar in 3d detection. arXiv preprint arXiv:2303.13560, 2023. 4 +[12] Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent cooperation. Advances in Neural Information Processing Systems (NIPS), 31, 2018. 3 +[13] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5 +[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12697-12705, 2019. 1, 3, 5 +[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Proceedings of the European Conference on Computer Vision (ECCV), pages 316-332, 2022. 4 +[16] Jinlong Li, Runsheng Xu, Xinyu Liu, Jin Ma, Zicheng Chi, Jiaqi Ma, and Hongkai Yu. Learning for vehicle-to-vehicle cooperative perception under lossy communication. IEEE Transactions on Intelligent Vehicles, 2023. 1 +[17] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems (NIPS), 34:29541-29552, 2021. 1, 3, 4, 6, 8 +[18] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 641-656, 2018. 2 +[19] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 5 +[20] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4106-4115, 2020. 1, 3, 4, 6 +[21] Guiyang Luo, Hui Zhang, Quan Yuan, and Jinglin Li. Complementarity-enhanced and redundancy-minimized collaboration network for multi-agent perception. In Proceedings of the ACM International Conference on Multimedia (ACMM), pages 3578-3586, 2022. 1, 3, 4 +[22] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NIPS), 32, 2019. 5 +[23] Tika Ram and Khem Chand. Effect of drivers' risk perception and perception of driving tasks on road safety attitude. Transportation Research Part F: Traffic Psychology and Behaviour, 42:162-176, 2016. 1 +[24] Andreas Rauch, Felix Klanner, Ralph Rasshofer, and Klaus Dietmayer. Car2x-based perception in a high-level fusion architecture for cooperative perception systems. In IEEE Intelligent Vehicles Symposium, pages 270-275, 2012. 3 +[25] Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in multiagent cooperative and competitive tasks. arXiv preprint arXiv:1812.09755, 2018. 2 +[26] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation. Advances in Neural Information Processing Systems (NIPS), 29, 2016. 2 +[27] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In International Conference on Machine Learning (ICML), pages 330-337, 1993. 2 + +[28] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 605–621, 2020. 1, 3, 4, 6 +[29] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. In IEEE International Conference on Robotics and Automation (ICRA), pages 3584-3591. IEEE, 2023. 1 +[30] Zhu Xizhou, Su Weijie, Lu Lewei, Li Bin, Wang Xiaogang, and Dai Jifeng. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2021. 4 +[31] Runsheng Xu, Weizhe Chen, Hao Xiang, Xin Xia, Lantao Liu, and Jiaqi Ma. Model-agnostic multi-agent perception framework. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1471–1478. IEEE, 2023. 3 +[32] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 5 +[33] Runsheng Xu, Jinlong Li, Xiaoyu Dong, Hongkai Yu, and Jiaqi Ma. Bridging the domain gap for multi-agent perception. arXiv preprint arXiv:2210.08451, 2022. 1 +[34] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. In Conference on Robot Learning (CoRL), 2022. 3, 6 +[35] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13712–13722, 2023. 1 +[36] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6, 8 +[37] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In Proceedings of the International Conference on Robotics and Automation (ICRA), pages 2583–2589, 2022. 1, 2, 3, 4, 5, 6 +[38] Dingkang Yang, Zhaoyu Chen, Yuzheng Wang, Shunli Wang, Mingcheng Li, Siao Liu, Xiao Zhao, Shuai Huang, Zhiyan Dong, Peng Zhai, and Lihua Zhang. Context deconfounded emotion recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19005-19015, June 2023. 2 +[39] Dingkang Yang, Shuai Huang, Haopeng Kuang, Yangtao Du, and Lihua Zhang. Disentangled representation learning for multimodal emotion recognition. In Proceedings of the 30th + +ACM International Conference on Multimedia (ACM MM), page 1642-1651, 2022. 2 +[40] Dingkang Yang, Shuai Huang, Yang Liu, and Lihua Zhang. Contextual and cross-modal interaction for multi-modal speech emotion recognition. IEEE Signal Processing Letters, 29:2093-2097, 2022. 2 +[41] Dingkang Yang, Shuai Huang, Shunli Wang, Yang Liu, Peng Zhai, Liuzhen Su, Mingcheng Li, and Lihua Zhang. Emotion recognition for multiple context awareness. In Proceedings of the European Conference on Computer Vision (ECCV), volume 13697, pages 144-162. Springer, 2022. 1 +[42] Dingkang Yang, Haopeng Kuang, Shuai Huang, and Lihua Zhang. Learning modality-specific and -agnostic representations for asynchronous multimodal language sequences. In Proceedings of the 30th ACM International Conference on Multimedia (ACMM), page 1708-1717, 2022. 2 +[43] Dingkang Yang, Yang Liu, Can Huang, Mingcheng Li, Xiao Zhao, Yuzheng Wang, Kun Yang, Yan Wang, Peng Zhai, and Lihua Zhang. Target and source modality co-reinforcement for emotion understanding from asynchronous multimodal sequences. Knowledge-Based Systems, page 110370, 2023. 1 +[44] Kun Yang, Jing Liu, Dingkang Yang, Hanqi Wang, Peng Sun, Yanni Zhang, Yan Liu, and Liang Song. A novel efficient multi-view traffic-related object detection framework. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2 +[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In IEEE International Conference on Distributed Computing Systems (ICDCS), pages 972-982, 2022. 2 +[46] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21361-21370, 2022. 1, 2, 5 +[47] Zhenxun Yuan, Xiao Song, Lei Bai, Zhe Wang, and Wanli Ouyang. Temporal-channel transformer for 3d lidar-based video object detection for autonomous driving. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2068-2078, 2021. 1, 3 +[48] Zixu Zhang and Jaime F Fisac. Safe occlusion-aware autonomous driving via game-theoretic active perception. arXiv preprint arXiv:2105.08169, 2021. 1 +[49] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 496-513, 2022. 1, 3 \ No newline at end of file diff --git a/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/images.zip b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..715b1be3acb5f5cf69a390fd91ece51bbd8a510a --- /dev/null +++ b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:685d29e3ad6ec62bdd935547b21c12df70681854297582e1c4f424d07ba13a88 +size 800008 diff --git a/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/layout.json b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..54ac9a1a29badc14b458de32982bef0564a9796f --- /dev/null +++ b/2023/Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception/layout.json @@ -0,0 +1,8661 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 54, + 103, + 539, + 121 + ], + "type": "text", + "content": "Spatio-Temporal Domain Awareness for Multi-Agent Collaborative Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "spans": [ + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": "Kun Yang" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": " Dingkang Yang" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": " Jingyu Zhang" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": " Mingcheng Li" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": " Yang Liu" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": " Jing Liu" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": " Hanqi Wang" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": " Peng Sun" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": " Liang Song" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": "Academy for Engineering and Technology, Fudan University " + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 141, + 519, + 200 + ], + "type": "text", + "content": "Duke Kunshan University {kunyang20,dkyang20,songl}@fudan.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 251, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 529 + ], + "type": "text", + "content": "Multi-agent collaborative perception as a potential application for vehicle-to-everything communication could significantly improve the perception performance of autonomous vehicles over single-agent perception. However, several challenges remain in achieving pragmatic information sharing in this emerging research. In this paper, we propose SCOPE, a novel collaborative perception framework that aggregates the spatio-temporal awareness characteristics across on-road agents in an end-to-end manner. Specifically, SCOPE has three distinct strengths: i) it considers effective semantic cues of the temporal context to enhance current representations of the target agent; ii) it aggregates perceptually critical spatial information from heterogeneous agents and overcomes localization errors via multi-scale feature interactions; iii) it integrates multi-source representations of the target agent based on their complementary contributions by an adaptive fusion paradigm. To thoroughly evaluate SCOPE, we consider both real-world and simulated scenarios of collaborative 3D object detection tasks on three datasets. Extensive experiments show the superiority of our approach and the necessity of the proposed components. The project link is https://ydk122024.github.io/SCOPE/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 551, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "text", + "content": "Perceiving complex driving environments is essential for ensuring road safety [23] and traffic efficiency [21] in autonomous driving. Despite the rapid development in the perception ability of intelligent vehicles through deep learning technologies [5, 6, 14, 41, 43], the conventional single-agent perception paradigm remains challenging. Single-agent perception system is generally vulnerable and defective due to the inevitable practical difficulties on the road," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 228, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 546, + 312 + ], + "type": "text", + "content": "such as restricted detection range and occlusion [47, 48]. To alleviate the inadequate observation constraints of isolated agents, multi-agent collaborative perception has emerged as a promising solution in recent years [17, 29]. With Vehicle-to-Everything (V2X) oriented communication applications, collaborative perception effectively promotes perspective complementation and information sharing among agents." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 315, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 547, + 519 + ], + "type": "text", + "content": "Based on recent LiDAR-based 3D collaborative perception datasets [35, 36, 37, 46], several impressive works have achieved a trade-off between collaborative perception performance and communication bandwidth of Autonomous Vehicles (AVs) via well-designed mechanisms. These remarkable efforts include knowledge distillation-based intermediate collaboration [17], handshake communication mechanism [20], V2X-visual transformer [36], and spatial information selection strategy [10]. However, there is still considerable room for progress in achieving a pragmatic and effective collaborative perception. Concretely, existing methods invariably follow a single-frame static perception pattern, suffering from the data sparsity shortcoming of 3D point clouds [47] and ignoring meaningful semantic cues in the temporal context [49]. The insufficient representation of the moving objects in the current frame potentially restricts the perception performance of target vehicles." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 714 + ], + "type": "text", + "content": "Moreover, spatial information aggregation in collaborative perception systems has exposed several problems. For the collaborator-shared message fusion, the per-agent/location-based fusion strategies of previous attempts [16, 17, 20, 28, 36] fail to deal with feature map misalignment from heterogeneous agents due to localization errors. Consequently, misleading features from collaborators could lead to object position misjudgment of the ego vehicle (i.e., receiver) and harm its detection precision. For the ego vehicle message refinement, existing approaches [10, 17, 20, 21, 28, 33] rely on fused representations with collaborator information to implement detection, abandoning the natural perception advantages of the ego vehicle and introducing the potential noises. The ego-centered characteristics may contain locally critical spatial information that is not disturbed by noise from collaborative agents" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "text", + "content": "Equal contributions in no particular order. The work was done when Dingkang Yang was at the Institute of Meta-Medical, IPASS. The two first authors thank Runsheng Xu for providing constructive suggestions. *Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23383" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 69, + 521, + 286 + ], + "blocks": [ + { + "bbox": [ + 77, + 69, + 521, + 286 + ], + "lines": [ + { + "bbox": [ + 77, + 69, + 521, + 286 + ], + "spans": [ + { + "bbox": [ + 77, + 69, + 521, + 286 + ], + "type": "image", + "image_path": "367b6d237d71ba4996473067053e82e0dd9db99ac2d68dbea701532133eceebe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 289, + 547, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 547, + 324 + ], + "type": "text", + "content": "Figure 1. The overall architecture of the proposed SCOPE. The framework consists of five parts: metadata conversion and feature extraction, context-aware information aggregation, confidence-aware cross-agent collaboration, importance-aware adaptive fusion, and detection decoders. The details of each individual component are illustrated in Section 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 333, + 288, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 288, + 370 + ], + "type": "text", + "content": "with asynchronous sensor measurements. To this end, how to effectively break the aforementioned limitations becomes a priority for achieving robust collaborative perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 379, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 289, + 715 + ], + "type": "text", + "content": "Motivated by the above observations, we present SCOPE, a Spatio-temporal domain awareness multi-agent CollabOrative PErception approach to tackle existing challenges jointly. From Figure 1, SCOPE achieves effective information collaboration among agents and precise perception of the ego agent via three proposed core components. Specifically, (i) for the data sparsity challenge in single-frame point clouds, the context-aware information aggregation is proposed to aggregate the ego agent's context information from preceding frames. We employ selective information filtering and spatio-temporal feature integration modules to capture informative temporal cues and historical context semantics. (ii) For the collaborator-shared message fusion challenge, the confidence-aware cross-agent collaboration is introduced to guarantee that the ego agent aggregates complementary information from heterogeneous agents. Based on the confidence-aware multi-scale feature interaction, we facilitate holistic communication and mitigate feature map misalignment due to localization errors of collaborators. (iii) For the multi-feature integration challenge of the ego agent, the importance-aware adaptive fusion is designed to flexibly incorporate the advantages of diverse representations based on distinctive perception contributions. We evaluate the superiority of SCOPE on various collaborative 3D object detection datasets, including DAIR-V2X [46], OPV2V [37], and V2XSet [36]. Extensive experiments in real-world and simulated scenarios provide favorable evidence that our approach is competitive in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 333, + 479, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 479, + 344 + ], + "type": "text", + "content": "multi-agent collaborative perception tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 345, + 539, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 345, + 539, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 345, + 539, + 357 + ], + "type": "text", + "content": "The main contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 364, + 545, + 605 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 364, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 364, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 317, + 364, + 545, + 447 + ], + "type": "text", + "content": "- We present SCOPE, a novel framework for multi-agent collaborative perception. The framework facilitates information collaboration and feature fusion among agents, achieving a reasonable performance-bandwidth trade-off. Comprehensive experiments on collaborative detection tasks show that SCOPE outperforms previous state-of-the-art methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 454, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 454, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 454, + 545, + 525 + ], + "type": "text", + "content": "- To our best knowledge, we are the first to consider the temporal context of the ego agent in collaborative perception systems. Based on the proposed context-aware component, the target frame of point clouds efficiently integrates historical cues from preceding frames to capture valuable temporal information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 533, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 545, + 605 + ], + "type": "text", + "content": "- We introduce two spatial information aggregation components to address the challenges of collaboration heterogeneity and fused representation unicity. These tailored components effectively perform multigrained information interactions among agents and multi-source feature refinement of the ego agent." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 616, + 392, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 392, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 392, + 628 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 635, + 465, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 465, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 465, + 647 + ], + "type": "text", + "content": "2.1. Multi-Agent Communication" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Benefiting from rapid advances in learning-based technologies [18, 38, 39, 40, 42, 44, 45], communication has become an essential component in constructing robust multiagent systems. Previous works [7, 25, 26, 27] typically focus on centralized paradigms or predefined protocols to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23384" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "explore information-sharing mechanisms among multiple agents. However, these efforts are potentially domain-constrained and scenario-isolated. Recent methods [3, 9, 12] achieve effective communication strategies in diverse scenarios. For instance, Vain [9] determines which agents would share information by modeling the locality of interactions. ATOC [12] introduces an attentional communication strategy to learn when communicating and integrating useful information to execute collaborative decisions. In comparison, we focus on the more challenging 3D perception tasks in complex driving scenarios. Based on the proposed confidence-aware multi-scale information interaction, SCOPE achieves an information-critical and holistic communication pattern among heterogeneous agents." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 246, + 187, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 246, + 187, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 187, + 258 + ], + "type": "text", + "content": "2.2. Collaborative Perception" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 264, + 289, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 516 + ], + "type": "text", + "content": "Collaborative perception investigates how to aggregate complementary perception semantics across connected agents to improve overall system performance. Current works are mainly categorized as early [2], intermediate [1, 10, 17, 20, 28, 34, 36, 37], and late [24, 31] fusion, where intermediate fusion is more favoured due to its good trade-off between performance and transmission bandwidth. In this case, several remarkable works explore efficient collaboration mechanisms. In DiscoNet [17], the teacher-student network is trained to aggregate the shared information. V2X-ViT [36] proposes heterogeneous multi-agent self-attention to handle device heterogeneity, and Xu et al. [34] design an axial attention module to capture sparse spatial semantics. Furthermore, Where2comm [10] filters irrelevant semantics by quantifying the spatial importance through confidence maps. Unlike these single-frame methods, we propose an information aggregation component to extract rich context information from local historical observations. Moreover, our cross-agent collaboration mechanism outperforms the previous per-agent/location message fusion models, significantly mitigating the impact of localization errors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 525, + 130, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 130, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 130, + 539 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 617 + ], + "type": "text", + "content": "In this paper, we focus on developing a robust collaborative perception system to improve the ego vehicle's perception capability. Figure 1 illustrates the overall architecture of the proposed framework. It consists of five sequential parts, where the system input accommodates different types of connected agents (i.e., infrastructures and AVs)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 623, + 284, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 284, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 284, + 635 + ], + "type": "text", + "content": "3.1. Metadata Conversion and Feature Extraction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "During the collaboration, one of the connected agents is identified as the ego agent " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " to build a communication graph, where other agents within the communication range serve as collaborators. Upon receiving the metadata " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "(e.g.," + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " poses and extrinsics) broadcast by the ego agent, the collaborators project their local LiDAR point clouds" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 468, + 179 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 468, + 179 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 468, + 179 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 468, + 179 + ], + "type": "image", + "image_path": "39b6fb7b98a911884f35aee76d423898913923fe38a44a77a2bfd7a535440964.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 182, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 228 + ], + "type": "text", + "content": "Figure 2. (a) The architecture of the proposed CIA component, including selective information filtering (left) and spatio-temporal feature integration (middle). (b) The multi-scale convolutional structure in the pyramid LSTM." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 470, + 72, + 544, + 178 + ], + "blocks": [ + { + "bbox": [ + 470, + 72, + 544, + 178 + ], + "lines": [ + { + "bbox": [ + 470, + 72, + 544, + 178 + ], + "spans": [ + { + "bbox": [ + 470, + 72, + 544, + 178 + ], + "type": "image", + "image_path": "9c047eff5af653d2c89c4ebdc503f8f9658027d248b9a5076e7837d6b585fae4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "text", + "content": "to the coordinate system of the ego agent. Similarly, the ego agent's previous point cloud frames are synchronized to the current frame. Each agent " + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "text", + "content": " encodes the projected point clouds into Bird's Eye View (BEV) features to extract local visual representations. Given the point cloud " + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_k^{(t)}" + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "text", + "content": "-th agent at timestamp " + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "text", + "content": ", the extracted feature is " + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_k^{(t)} = f_{enc}(\\mathbf{X}_k^{(t)}) \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "inline_equation", + "content": "f_{enc}(\\cdot)" + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "text", + "content": " is the PointPillar [14] encoder shared among all agents, and " + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "inline_equation", + "content": "C, H, W" + }, + { + "bbox": [ + 304, + 236, + 545, + 348 + ], + "type": "text", + "content": " stand for the channel, height, and width." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 354, + 518, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 518, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 518, + 367 + ], + "type": "text", + "content": "3.2. Context-aware Information Aggregation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 372, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 516 + ], + "type": "text", + "content": "Conventional 3D detection tasks [47, 49] have shown the potential of temporal information in improving perception performance. Unfortunately, current multi-agent collaborative efforts [10, 17, 20, 21, 28] invariably focus on exploring the current frame while ignoring context cues from previous frames. The single-frame solution fails to effectively detect fast-moving objects (e.g., surrounding vehicles) due to the sparsity and insufficiency of point clouds. To this end, we present a Context-aware Information Aggregation (CIA) component to capture spatio-temporal representations of the ego agent's previous frames to fuse valuable semantics. As shown in Figure 2(a), CIA consists of two core phases." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "text", + "content": "Selective Information Filtering. This phase aims to extract refined information by filtering redundant features from the target " + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i^{(t)} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "text", + "content": " and distilling meaningful characteristics from the previous " + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i^{(t - \\tau)} \\in \\mathbb{R}^{\\tau \\times C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "text", + "content": " is the time shift. Concretely, the channel-wise average and max pooling operations " + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\Psi_{a / m}(\\cdot)" + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "text", + "content": " are first utilized to aggregate rich channel semantics, denoted as " + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{a / m}^{(t)} = \\Psi_{a / m}(\\pmb{F}_i^{(t)})" + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{a / m}^{(\\tau)} = \\Psi_{a / m}(\\pmb{F}_i^{(\\tau)})" + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i^{(\\tau)} = \\sum_{n=1}^{\\tau} \\pmb{F}_i^{(t-n)}" + }, + { + "bbox": [ + 304, + 516, + 547, + 663 + ], + "type": "text", + "content": ". These refined feature maps are combined to obtain a well-chosen spatial selection map. The procedure is expressed as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 668, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 668, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 315, + 668, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {U} = \\sigma \\cdot \\aleph (\\mathcal {M} _ {a} ^ {(t)} \\| \\mathcal {M} _ {m} ^ {(t)} + \\mathcal {M} _ {a} ^ {(\\tau)} \\| \\mathcal {M} _ {m} ^ {(\\tau)}) \\in \\mathbb {R} ^ {H \\times W}, \\quad (1)", + "image_path": "892854ad4a8b88912ee4f685ce0c6710d01fd9bd693c3ac99239988ac2491971.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " is the sigmoid activation, " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\aleph" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " denotes a convolutional layer for fusion, and " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\parallel" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " denotes the concatenation op" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23385" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 286, + 203 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 203 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 203 + ], + "type": "image", + "image_path": "63e56030fa655eb97aab668dd2b6ed5c5b81ea5e717f094b6119abf2b0feae59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 287, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 287, + 242 + ], + "type": "text", + "content": "Figure 3. (a) The architecture of the proposed CCC component. (b) CCC consists of two phases at each scale: reference point proposal (middle) and deformable cross-attention module (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 288, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 291 + ], + "type": "text", + "content": "eration. Immediately, the information filtering strategy is implemented via pixel-wise attention weighting to enhance the salient historical features of " + }, + { + "bbox": [ + 46, + 251, + 288, + 291 + ], + "type": "inline_equation", + "content": "H_{i}^{(\\tau)} \\in \\mathbb{R}^{\\tau \\times C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 251, + 288, + 291 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 298, + 287, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 298, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 70, + 298, + 287, + 314 + ], + "type": "interline_equation", + "content": "\\boldsymbol {H} _ {i} ^ {(\\tau)} = (1 - \\mathcal {U}) \\odot \\tanh \\left(\\boldsymbol {F} _ {i} ^ {(t)}\\right) + \\mathcal {U} \\odot \\boldsymbol {F} _ {i} ^ {(t - \\tau)}. \\tag {2}", + "image_path": "6c01bf842c311c83bb188208f62706e20a4f7d97269861c196d264f4be5dd265.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 322, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 566 + ], + "type": "text", + "content": "Spatio-temporal Feature Integration. In order to integrate the historical prototype to improve the perception ability of the current representation, we introduce a pyramid LSTM [15] to learn the contextual dependencies of interframe features " + }, + { + "bbox": [ + 46, + 322, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i^{(t)}\\| \\pmb{H}_i^{(\\tau)}" + }, + { + "bbox": [ + 46, + 322, + 287, + 566 + ], + "type": "text", + "content": ". The difference between the pyramid LSTM and vanilla LSTM [8] is that the initial matrix multiplication is replaced with a multi-scale convolutional structure, as shown in Figure 2(b). Note that the intrinsic long short-term memory mechanism in the LSTM is adopted, and please refer to [8] for details. In practice, multi-scale spatial features are extracted by two successive 2D convolutions among different scales, followed by batch normalization and ReLU activation. To achieve multilevel spatial semantic fusion, the downsampled features are progressively interpolated to deliver to the upsampled layers via lateral connections. The overall design facilitates learning spatio-temporal correlations in historical contexts, which benefits our multi-frame perception paradigm. Ultimately, the final hidden state in the pyramid LSTM is used as the refined context-aware feature " + }, + { + "bbox": [ + 46, + 322, + 287, + 566 + ], + "type": "inline_equation", + "content": "\\pmb{H}_i^{(t)}\\in \\mathbb{R}^{C\\times H\\times W}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 571, + 281, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 281, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 281, + 585 + ], + "type": "text", + "content": "3.3. Confidence-aware Cross-agent Collaboration" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 590, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 715 + ], + "type": "text", + "content": "The cross-agent collaboration targets to enhance the ego agent's visual representations by aggregating the complementary semantics from the collaborator-shared features. Existing works [10, 11, 36, 37] propose attention-based perlocation feature fusion methods, which are vulnerable to localization errors and ignore the sparsity of point clouds. To tackle these issues, we implement a novel Confidence-aware Cross-agent Collaboration (CCC) component. We first obtain the confidence map " + }, + { + "bbox": [ + 46, + 590, + 287, + 715 + ], + "type": "inline_equation", + "content": "S_{k}^{(t)}" + }, + { + "bbox": [ + 46, + 590, + 287, + 715 + ], + "type": "text", + "content": " and binary selection matrix " + }, + { + "bbox": [ + 46, + 590, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_k^{(t)}" + }, + { + "bbox": [ + 46, + 590, + 287, + 715 + ], + "type": "text", + "content": " of feature " + }, + { + "bbox": [ + 46, + 590, + 287, + 715 + ], + "type": "inline_equation", + "content": "F_{k}^{(t)}, k\\in \\{1,\\dots,K\\}" + }, + { + "bbox": [ + 46, + 590, + 287, + 715 + ], + "type": "text", + "content": " following [10]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "text", + "content": "The filtered feature shared by " + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "text", + "content": "-th agent to ego agent is " + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{F}_k^{(t)} = \\pmb{F}_k^{(t)} \\odot \\mathcal{C}_k^{(t)}" + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "text", + "content": ". As Figure 3(a) shows, the CCC component encodes the features and confidence maps into three scales and performs feature fusion at each scale. " + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{k,l}^{(t)}" + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "S_{k,l}^{(t)}" + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "text", + "content": " denote the feature and confidence map at " + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 72, + 545, + 163 + ], + "type": "text", + "content": "-th scale. We take two agents as an example to elaborate on the feature fusion in Figure 3(b), including the following two phases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "type": "text", + "content": "Reference Point Proposal. To predict the spatial locations containing meaningful object information, we obtain the element-wise summation of all confidence maps as " + }, + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "type": "inline_equation", + "content": "S_{sum,l}^{(t)} = \\sum_{k=1}^{K} S_{k,l}^{(t)}" + }, + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "type": "text", + "content": ". Since the confidence map reflects the spatially critical levels, " + }, + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "type": "inline_equation", + "content": "S_{sum,l}^{(t)}" + }, + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "type": "text", + "content": " presents the potential positions of the targets within the detection range, called the reference points. Accordingly, we apply a threshold-based selection function " + }, + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "type": "inline_equation", + "content": "f_{sel}(\\cdot)" + }, + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "type": "text", + "content": " to extract the reference points " + }, + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "type": "inline_equation", + "content": "S_{re,l}^{(t)}" + }, + { + "bbox": [ + 305, + 163, + 546, + 290 + ], + "type": "text", + "content": ". This design could actively guide the subsequent fusion network to concentrate on the essential spatial areas." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "text", + "content": "Deformable Cross-attention Module. We extract the ego agent's feature " + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{i,l}^{(t)}" + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "text", + "content": " at the reference points as the initial query embedding and apply a linear layer to encode the locations of reference points into a position embedding. To resolve the misalignment of feature maps and acquire more robust representations against localization errors, the Deformable Cross-attention Module (DCM) aggregates the information from sampled keypoints through a deformable cross-attention layer [30]. Concretely, a linear layer is leveraged to learn the 2D spatial offset " + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "inline_equation", + "content": "\\Delta q" + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "text", + "content": " of the reference point " + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "text", + "content": " for sampling attending keypoints at " + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "inline_equation", + "content": "q + \\Delta q" + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "text", + "content": ". We simultaneously select " + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "text", + "content": " keypoints and extract their features as the attending features. The output of DCM at position " + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 291, + 546, + 449 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 453, + 545, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 453, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 312, + 453, + 545, + 482 + ], + "type": "interline_equation", + "content": "\\operatorname {D C M} (q) = \\sum_ {a = 1} ^ {A} \\boldsymbol {W} _ {a} \\left[ \\sum_ {k = 1} ^ {K} \\sum_ {m = 1} ^ {M} \\phi \\left(\\boldsymbol {W} _ {b} \\boldsymbol {F} _ {i, l} ^ {(t)} (q)\\right) \\boldsymbol {F} _ {k, l} ^ {(t)} \\left(q + \\Delta q _ {m}\\right) \\right], \\tag {3}", + "image_path": "61d3155c4a82f7cd9ad506a1584405050796592e8cb2afe325e898a0d3d167cc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "content": " indexes the attention head, " + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{a / b}" + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "content": " are the learnable weights, and " + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "content": " is the softmax function. Ultimately, we derive a filling operation that fills " + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathrm{DCM}(q)" + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "content": " into the ego agent's feature " + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{i,l}^{(t)}" + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "content": " based on the initial position " + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "content": " and outputs " + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\pmb{Z}_{i,l}^{(t)}" + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "content": ". The output-enhanced features under the three scales are encoded into the same size and concatenated in the channel dimension. We utilize a " + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "content": " convolutional layer to fuse information of the three scales and get the final collaboration feature " + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\pmb{Z}_i^{(t)}\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 305, + 487, + 545, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 610, + 495, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 495, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 495, + 624 + ], + "type": "text", + "content": "3.4. Importance-aware Adaptive Fusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": "Although previous works [17, 20, 21, 28, 36] have achieved impressive performance through ego features that aggregate collaborative information, they potentially suffer from noise interference introduced by collaborators with asynchronous measurements. A promising solution is to consider purely ego-centered characteristics, which contain the natural perception advantages of the target agent. To" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23386" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 284, + 211 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 284, + 211 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 284, + 211 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 284, + 211 + ], + "type": "image", + "image_path": "26dbb363aa27207230655e7bc0098c7122196a976e6b1a158ff9d7e8cf002baa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 215, + 263, + 226 + ], + "lines": [ + { + "bbox": [ + 47, + 215, + 263, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 263, + 226 + ], + "type": "text", + "content": "Figure 4. The architecture of the proposed IAF component." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 235, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 287, + 308 + ], + "type": "text", + "content": "this end, we propose an Importance-aware Adaptive Fusion (IAF) component to fuse multi-source features based on their complementary contributions. These informative features include " + }, + { + "bbox": [ + 46, + 235, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{H}_i^{(t)}, \\pmb{Z}_i^{(t)}" + }, + { + "bbox": [ + 46, + 235, + 287, + 308 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 235, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i^{(t)}" + }, + { + "bbox": [ + 46, + 235, + 287, + 308 + ], + "type": "text", + "content": ". From Figure 4, the importance generator " + }, + { + "bbox": [ + 46, + 235, + 287, + 308 + ], + "type": "inline_equation", + "content": "f_{gen}(\\cdot)" + }, + { + "bbox": [ + 46, + 235, + 287, + 308 + ], + "type": "text", + "content": " is used to generate the spatial importance maps, which processes as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 312, + 287, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 312, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 287, + 330 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {i} ^ {(t)}, \\mathcal {Z} _ {i} ^ {(t)}, \\mathcal {S} _ {i} ^ {(t)} = \\sigma \\cdot \\Psi_ {m} (f _ {g e n} (\\boldsymbol {H} _ {i} ^ {(t)}, \\boldsymbol {Z} _ {i} ^ {(t)}, \\boldsymbol {F} _ {i} ^ {(t)})), \\quad (4)", + "image_path": "3898769634ff4c104f09c09960bb191a8493ae397f2e5bf4b9ed18d1f5eb8fea.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 333, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 333, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 287, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 333, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i^{(t)},\\mathcal{Z}_i^{(t)},S_i^{(t)}\\in [0,1]^{H\\times W}" + }, + { + "bbox": [ + 47, + 333, + 287, + 373 + ], + "type": "text", + "content": ". Then, these importance maps are concatenated to get the normalized attention maps across distinct features by the softmax function " + }, + { + "bbox": [ + 47, + 333, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 47, + 333, + 287, + 373 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 376, + 287, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 287, + 404 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\mathcal {H}} ^ {(t)} = \\phi \\left(\\mathcal {H} _ {i} ^ {(t)}\\right) = \\frac {\\exp \\left(\\mathcal {H} _ {i} ^ {(t)}\\right)}{\\exp \\left(\\mathcal {H} _ {i} ^ {(t)}\\right) + \\exp \\left(\\mathcal {Z} _ {i} ^ {(t)}\\right) + \\exp \\left(\\mathcal {S} _ {i} ^ {(t)}\\right)}. \\tag {5}", + "image_path": "966d95be0a36bd8d4308aa66d3efe827999da7ba723ca8bf4602e20901129212.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "type": "text", + "content": "A larger value in " + }, + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{H}}^{(t)} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "type": "text", + "content": " implies that the corresponding feature position is more significant. Similarly, " + }, + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{Z}}^{(t)} = \\phi(\\mathcal{Z}_i^{(t)})" + }, + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_S^{(t)} = \\phi(S_i^{(t)})" + }, + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "type": "text", + "content": ". Finally, we adaptively activate each feature map's perceptually critical spatial information by a weighted summation. The fused feature map " + }, + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^{(t)} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 408, + 287, + 485 + ], + "type": "text", + "content": " is obtained as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 490, + 287, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 490, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 59, + 490, + 287, + 507 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {i} ^ {(t)} = \\mathcal {E} _ {\\mathcal {H}} ^ {(t)} \\odot \\boldsymbol {H} _ {i} ^ {(t)} + \\mathcal {E} _ {\\mathcal {Z}} ^ {(t)} \\odot \\boldsymbol {Z} _ {i} ^ {(t)} + \\mathcal {E} _ {\\mathcal {S}} ^ {(t)} \\odot \\boldsymbol {F} _ {i} ^ {(t)}. \\tag {6}", + "image_path": "c1e27232646c96cb51a94d53f688a7d2d55ec7a5ef068cda9beecc8aa0078b75.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 510, + 287, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 287, + 523 + ], + "type": "text", + "content": "3.5. Detection Decoder and Objective Optimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "type": "text", + "content": "We leverage two detection decoders to decode the final fused feature " + }, + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i^{(t)}" + }, + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "type": "text", + "content": " into the predicted outputs. The regression output is " + }, + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{i,r}^{(t)} = f_{dec}^r (\\mathcal{F}_i^{(t)})\\in \\mathbb{R}^{7\\times H\\times W}" + }, + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "type": "text", + "content": ", representing the position, size, and yaw angle of the bounding box at each location. The classification output reveals the confidence value of each predefined box to be a target or background, which is " + }, + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{i,c}^{(t)} = f_{dec}^c (\\mathcal{F}_i^{(t)})\\in \\mathbb{R}^{2\\times H\\times W}" + }, + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "type": "text", + "content": ". Following existing work [14], we adopt the smooth " + }, + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 529, + 287, + 643 + ], + "type": "text", + "content": " loss for regressing the bounding boxes and a focal loss [19] for classification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 671, + 222, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 222, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 222, + 682 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Datasets. We conduct comprehensive experiments on three benchmark datasets for collaborative perception tasks, in" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": "including DAIR-V2X [46], V2XSet [36], and OPV2V [37]. These datasets provide annotated sensor data from multiple connected agents for each sample. Concretely, DAIR-V2X is a large-scale real-world dataset supporting collaborative 3D object detection, containing 71,254 samples and dividing the training/validation/testing sets as 5:2:3. Each sample contains the LiDAR point clouds from a vehicle and an infrastructure. During feature extraction, we transform the point cloud into the BEV feature map of the size (200, 504, 64). V2XSet is a public simulation dataset for V2X collaborative perception, including 73 representative scenarios and 11,447 frames of labeled point clouds generated by Carla [4]. The training/validation/testing sets are 6,694, 1,920, and 2,833 frames, respectively. The size of each extracted feature map is (100, 384, 64). OPV2V is a vehicle-to-vehicle collaborative perception dataset collected by Carla and OpenCDA [32]. It includes 11,464 frames of point clouds and RGB images with 3D annotation, splitting into training/validation/testing sets of 6,764, 1,981, and 2,719 frames, respectively. The 3D point cloud is extracted as a feature map of the size (100, 384, 64)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 323, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 384 + ], + "type": "text", + "content": "Evaluation Metrics. Following [36], we leverage the Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 and 0.7 to evaluate the 3D object detection performance. The communication results count the message size by byte in the log scale with base 2 [10]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 392, + 441, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 441, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 441, + 403 + ], + "type": "text", + "content": "4.2. Implementation Details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": "We build the proposed models on the Pytorch toolbox [22] and train them on four Nvidia Tesla V100 GPUs using the Adam optimizer [13]. The training settings of the DAIR-V2X, V2XSet, and OPV2V datasets follow: the batch sizes are " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\{5, 3, 3\\}" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": " and the epochs are " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\{30, 40, 40\\}" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": ". The initial learning rate is " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "2\\mathrm{e} - 3" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": " and decays every 15 epochs with a factor of 0.1. Our CIA module uses the previous 2 frames, the attention head in the CCC component is 8, and the number of keypoints is 15. To simulate the default localization and heading errors, we add Gaussian noise with a standard deviation of " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{m}" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "0.2^{\\circ}" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": " to the initial position and orientation of the agents. The voxel resolution of the observation encoder " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "f_{enc}(\\cdot)" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{m}" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": " in height and width. The generator " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "f_{gen}(\\cdot)" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": " follows a detection decoder from [14], and the summed confidence map " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "S_{sum,l}^{(t)}" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": " is filtered using a threshold of 0.1 in the function " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "f_{sel}(\\cdot)" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": ". The decoders " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\{f_{dec}^{r}(\\cdot), f_{dec}^{c}(\\cdot)\\}" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": " are derived with two " + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 409, + 545, + 639 + ], + "type": "text", + "content": " convolutional layers. We provide results on the validation set of DAIR-V2X and the testing set of V2XSet and OPV2V." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 647, + 441, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 441, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 441, + 659 + ], + "type": "text", + "content": "4.3.Quantitative Evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Detection Performance Comparison. Table 1 shows the 3D detection performance comparison results on the three datasets. We compare the following methods. No Fusion as a baseline considers only the point cloud data of ego vehi" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23387" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 211, + 177 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 211, + 177 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 211, + 177 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 211, + 177 + ], + "type": "image", + "image_path": "a89cbb5d8c3de522e8f7b24f4d4ba2745219562dbc5ecd2f1603891c0af0fc05.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 546, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 546, + 205 + ], + "type": "text", + "content": "Figure 5. Collaborative perception performance comparison of SCOPE and Where2comm [10] on the DAIR-V2X, V2XSet and OPV2V datasets with varying communication volume." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 217, + 72, + 376, + 176 + ], + "blocks": [ + { + "bbox": [ + 217, + 72, + 376, + 176 + ], + "lines": [ + { + "bbox": [ + 217, + 72, + 376, + 176 + ], + "spans": [ + { + "bbox": [ + 217, + 72, + 376, + 176 + ], + "type": "image", + "image_path": "e451d18b2479306ccbff34efb6869ab379bac8a0f5854bd7c596591249867bed.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 72, + 543, + 176 + ], + "blocks": [ + { + "bbox": [ + 379, + 72, + 543, + 176 + ], + "lines": [ + { + "bbox": [ + 379, + 72, + 543, + 176 + ], + "spans": [ + { + "bbox": [ + 379, + 72, + 543, + 176 + ], + "type": "image", + "image_path": "10bcef53aa60ef03160c33255cafd189b73cf887ba1c6f361c150f14cc4ef2cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 211, + 218, + 314 + ], + "blocks": [ + { + "bbox": [ + 50, + 211, + 218, + 314 + ], + "lines": [ + { + "bbox": [ + 50, + 211, + 218, + 314 + ], + "spans": [ + { + "bbox": [ + 50, + 211, + 218, + 314 + ], + "type": "image", + "image_path": "866dc6a87604fdd1680e42da86f0d9b393df137582b6a5173ebef319e79a059c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 322, + 470, + 333 + ], + "lines": [ + { + "bbox": [ + 122, + 322, + 470, + 333 + ], + "spans": [ + { + "bbox": [ + 122, + 322, + 470, + 333 + ], + "type": "text", + "content": "Figure 6. Robustness to the localization error on the DAIR-V2X, V2XSet, and OPV2V datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 220, + 211, + 380, + 314 + ], + "blocks": [ + { + "bbox": [ + 220, + 211, + 380, + 314 + ], + "lines": [ + { + "bbox": [ + 220, + 211, + 380, + 314 + ], + "spans": [ + { + "bbox": [ + 220, + 211, + 380, + 314 + ], + "type": "image", + "image_path": "c9734047217852d965ebab3c7a3f58b121229c521afc331558d263ff75920525.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 382, + 212, + 544, + 314 + ], + "blocks": [ + { + "bbox": [ + 382, + 212, + 544, + 314 + ], + "lines": [ + { + "bbox": [ + 382, + 212, + 544, + 314 + ], + "spans": [ + { + "bbox": [ + 382, + 212, + 544, + 314 + ], + "type": "image", + "image_path": "6dc8fe174cb7352ae503e9020ba8ea9ab9d6addd7ee2ca13507dd1ddcd8a79ac.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 48, + 366, + 285, + 510 + ], + "blocks": [ + { + "bbox": [ + 47, + 344, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 47, + 344, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 344, + 287, + 365 + ], + "type": "text", + "content": "Table 1. Performance comparison on the DAIR-V2X, V2XSet, and OPV2V datasets. The results are reported in AP@0.5/0.7." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 366, + 285, + 510 + ], + "lines": [ + { + "bbox": [ + 48, + 366, + 285, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 285, + 510 + ], + "type": "table", + "html": "
ModelDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
No Fusion50.03/43.5760.60/40.2068.71/48.66
Late Fusion53.12/37.8866.79/50.9582.24/65.78
Early Fusion61.74/46.5377.39/50.4581.30/68.69
When2com [20]51.12/36.1770.16/53.7277.85/62.40
V2VNet [28]56.01/42.2581.80/61.3582.79/70.31
AttFuse [37]53.79/42.6176.27/57.9383.21/70.09
V2X-ViT [36]54.26/43.3585.13/68.6786.72/74.94
DiscoNet [17]54.29/44.8882.18/63.7387.38/73.19
CoBEVT [34]54.82/43.9583.01/62.6787.40/74.35
Where2comm [10]63.71/48.9385.78/72.4288.07/75.06
SCOPE (Ours)65.18/49.8987.52/75.0589.71/80.62
", + "image_path": "3cdd9dc15dcbb4aa96dbd083e4d9f5499f0665dd3c6af91c0bd05c044a199182.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "content": "cles without collaboration. Late Fusion allows agents to exchange detected results and utilizes non-maximum suppression to produce the final outputs. Early Fusion means the raw LiDAR point clouds are shared to complete the perception view. Meanwhile, the existing state-of-the-art (SOTA) models are fully considered, including When2com [20], V2VNet [28], AttFuse [37], V2X-ViT [36], DiscoNet [17], CoBEVT [34], and Where2comm [10]. Intuitively, the proposed SCOPE outperforms No Fusion, Late Fusion, and Early Fusion by large margins across all datasets, demonstrating the superiority of our perception paradigm. Moreover, our method significantly outperforms previous SOTA models in simulated and real-world (DAIR-V2X) scenarios. In particular, the SOTA performance is improved by " + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "inline_equation", + "content": "3.63\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "inline_equation", + "content": "7.41\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 713 + ], + "type": "text", + "content": " on the V2XSet and OPV2V datasets separately" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "text", + "content": "in AP@0.7. Compared to single-frame and collaboration-only based efforts [10, 17, 20, 28, 36], SCOPE proves the effectiveness and rationality of simultaneously considering the temporal context, collaboration, and ego characteristics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 399, + 546, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 554 + ], + "type": "text", + "content": "Comparison of Communication Volume. Evaluating performance at varying bandwidths plays an essential role in achieving pragmatic perception. To this purpose, Figure 5 shows the perception performance of our SCOPE and Where2comm [10] with various communication volumes. Note that we do not consider communication methods that use any extra model/data/feature compression for a fair comparison. We have the following observations: (i) SCOPE achieves better performance than previous models [17, 20, 28, 36] on all datasets with less communication volume. (ii) Compared to Where2comm [10], our method obtains a superior perception-communication tradeoff across all the communication bandwidth choices." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "Robustness to Localization and Heading Errors. Following the noise settings from [36], we perform experiments to evaluate the robustness against natural localization noises of existing methods. The noises are sampled from a Gaussian distribution with standard deviation " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{xyz} \\in [0, 0.5]" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " m, " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{heading}} \\in [0^{\\circ}, 1.0^{\\circ}]" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ". From Figures 6 and 7, the overall collaborative perception performance consistently deteriorates as errors continue to increase. For instance, When2com and V2VNet fail even worse than No Fusion on the DAIR-V2X when the localization and heading errors are over 0.2 m and " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "0.4^{\\circ}" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": ", respectively. Noticeably, our method consistently surpasses the previous SOTA models under all noise levels, clearly displaying the robustness of SCOPE against" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23388" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 212, + 173 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 212, + 173 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 212, + 173 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 212, + 173 + ], + "type": "image", + "image_path": "f11b5a7e055828d46820c0eb0fa0c641e2d2f092501d0fb7aa62a64c3997cc33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 179, + 464, + 191 + ], + "lines": [ + { + "bbox": [ + 128, + 179, + 464, + 191 + ], + "spans": [ + { + "bbox": [ + 128, + 179, + 464, + 191 + ], + "type": "text", + "content": "Figure 7. Robustness to the heading error on the DAIR-V2X, V2XSet, and OPV2V datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 74, + 377, + 173 + ], + "blocks": [ + { + "bbox": [ + 215, + 74, + 377, + 173 + ], + "lines": [ + { + "bbox": [ + 215, + 74, + 377, + 173 + ], + "spans": [ + { + "bbox": [ + 215, + 74, + 377, + 173 + ], + "type": "image", + "image_path": "452945dfc3e6d64ec3cf1d46f6a6ec8b89db3b7e2f6835a2e7e528f945bc7e0c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 74, + 543, + 173 + ], + "blocks": [ + { + "bbox": [ + 379, + 74, + 543, + 173 + ], + "lines": [ + { + "bbox": [ + 379, + 74, + 543, + 173 + ], + "spans": [ + { + "bbox": [ + 379, + 74, + 543, + 173 + ], + "type": "image", + "image_path": "7f757639bab00616afa0b5fd38075ec66444ad6ff0aecf46d3f6585e478e1333.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 47, + 247, + 286, + 330 + ], + "blocks": [ + { + "bbox": [ + 46, + 202, + 287, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 287, + 246 + ], + "type": "text", + "content": "Table 2. Ablation study results of the proposed core components on all the datasets. PL: pyramid LSTM; SIF: selective information filtering; DCM: deformable cross-attention module; RPP: reference point proposal; IAF: importance-aware adaptive fusion." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 247, + 286, + 330 + ], + "lines": [ + { + "bbox": [ + 47, + 247, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 286, + 330 + ], + "type": "table", + "html": "
PLSIFDCMRPPIAFDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
60.31/40.3576.04/62.4177.22/65.51
✓61.75/44.5880.36/65.7481.19/69.02
✓✓62.48/45.6281.88/67.4282.34/71.23
✓✓✓64.06/47.8785.47/72.2986.96/76.84
✓✓✓✓64.74/48.4586.33/73.2788.01/78.57
✓✓✓✓✓65.18/49.8987.52/75.0589.71/80.62
", + "image_path": "7697102e806923bf7ec5e1ca2eb4d7ec318762de074aa76fb57815898ca5425d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 341, + 287, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 401 + ], + "type": "text", + "content": "pose errors. The reasonable explanations are: (i) the deformable cross-attention module facilitates capturing perceptually critical information across agents; (ii) the proposed multi-scale feature interaction alleviates the feature map misalignment about collaborative agents due to noises." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 407, + 148, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 148, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 148, + 419 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 426, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 462 + ], + "type": "text", + "content": "We perform thorough ablation studies on all datasets to understand the necessity of the different designs. Tables 2 and 3 show the following observations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "type": "text", + "content": "Importance of Core Components. We investigate the contribution of each core component in SCOPE. A base version (first row of Table 2) is provided as the gain reference. In the base model, we use the same fusion pattern to aggregate context information, collaborator-shared features, and multi-source features, where features are concatenated on the channel dimension and fused by a " + }, + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 463, + 287, + 617 + ], + "type": "text", + "content": " convolutional network. As Table 2 demonstrates, we progressively add (1) Pyramid LSTM, (2) SIF, (3) DCM, (4) RPP, and (5) IAF and present the corresponding detection precision. The consistently increased detection results over the three datasets reveal the effectiveness of each proposed component, which all provide valuable performance gains." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Effect of Historical Frame Number. As shown in the top half of Table 3, we evaluate the effect of integrating different degrees of the temporal context from the ego agent. The interesting finding is that choosing a suitable historical frame count (i.e., 2 frames) is beneficial in achieving better performance gains. The potential reason is that adding excessive historical frames may lead to accumulated errors under the noise setting and degrade the performance." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 222, + 544, + 430 + ], + "blocks": [ + { + "bbox": [ + 305, + 202, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 222 + ], + "type": "text", + "content": "Table 3. Ablation study results of candidate designs and strategies on all the datasets. \"w/f\" and \"w/o\" mean the with and without." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 222, + 544, + 430 + ], + "lines": [ + { + "bbox": [ + 307, + 222, + 544, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 544, + 430 + ], + "type": "table", + "html": "
Designs/StrategiesDAIR-V2XV2XSetOPV2V
AP@0.5/0.7AP@0.5/0.7AP@0.5/0.7
Full Model65.18/49.8987.52/75.0589.71/80.62
Effect of Historical Frame Number
1 Frame64.28/49.1686.16/74.2188.85/78.49
2 Frames (Default)65.18/49.8987.52/75.0589.71/80.62
3 Frames65.03/49.2486.99/74.6589.67/80.45
Effect of Keypoint Number
9 Keypoints64.36/49.2286.34/74.0788.32/77.94
15 Keypoints (Default)65.18/49.8987.52/75.0589.71/80.62
21 Keypoints64.72/49.4087.12/74.6389.08/78.38
Importance of Distinct Features
w/o Ego-agent Featue64.96/49.3486.47/74.2488.63/79.38
w/o Context Feature64.14/48.8185.93/73.8888.26/77.18
w/o Collaboration Feature52.35/44.4663.91/45.3471.36/50.94
Effect of Fusion Strategies
w/ Summation Fusion63.65/48.9185.81/72.6488.05/76.13
w/ Maximum Fusion64.92/49.2786.28/73.6187.95/78.17
w/ Average Fusion64.85/48.9686.55/73.4688.24/78.73
", + "image_path": "db7cb919adf01446054ede96d72f03311c5d074fe00aa2a663603b0db4117580.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "type": "text", + "content": "Effect of Keypoint Number. We investigate the rationality of the keypoint number in the deformable crossattention. Noticeably, the DCM component with 15 keypoints achieves better detection performance than other settings. This result shows that SCOPE mitigates the impact of localization errors and fully extracts essential spatial semantics by setting a reasonable keypoint number." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 530, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 652 + ], + "type": "text", + "content": "Importance of Distinct Features. We explore the importance of distinct representations by removing the ego-agent, context, and collaboration features (i.e., " + }, + { + "bbox": [ + 304, + 530, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i^{(t)}, \\pmb{H}_i^{(t)}, \\pmb{Z}_i^{(t)}" + }, + { + "bbox": [ + 304, + 530, + 545, + 652 + ], + "type": "text", + "content": "). There is a significant drop in the model's performance when the collaboration feature is removed, indicating that multiagent information sharing dominates various collaborative perception tasks. Furthermore, the consistently decreased performance suggests that ego-centered characteristics (i.e., pure ego-agent feature and corresponding context information) provide an indispensable contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "Effect of Fusion Strategies. Finally, we demonstrate the effectiveness of the proposed adaptive fusion strategy by providing candidate fusion mechanisms. Summation Fusion is a direct pixel-wise addition operation. Maximum and Average Fusion select elements at corresponding po" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23389" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 72, + 185, + 208 + ], + "blocks": [ + { + "bbox": [ + 76, + 72, + 185, + 208 + ], + "lines": [ + { + "bbox": [ + 76, + 72, + 185, + 208 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 185, + 208 + ], + "type": "image", + "image_path": "04fd1d07398a19438ad2d4faa76ed8e950b1d39da7f8222e34097cb4b8737187.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 210, + 148, + 219 + ], + "lines": [ + { + "bbox": [ + 108, + 210, + 148, + 219 + ], + "spans": [ + { + "bbox": [ + 108, + 210, + 148, + 219 + ], + "type": "text", + "content": "(a) DiscoNet" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 72, + 296, + 208 + ], + "blocks": [ + { + "bbox": [ + 187, + 72, + 296, + 208 + ], + "lines": [ + { + "bbox": [ + 187, + 72, + 296, + 208 + ], + "spans": [ + { + "bbox": [ + 187, + 72, + 296, + 208 + ], + "type": "image", + "image_path": "a4758aec46258a60fe5f4d4d55718824214a96fac70a6c7e93d98f1857c50d22.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 210, + 261, + 219 + ], + "lines": [ + { + "bbox": [ + 221, + 210, + 261, + 219 + ], + "spans": [ + { + "bbox": [ + 221, + 210, + 261, + 219 + ], + "type": "text", + "content": "(b) V2X-ViT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 407, + 208 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 407, + 208 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 407, + 208 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 407, + 208 + ], + "type": "image", + "image_path": "bfb1871e46055a4ed03979d1c9c674a52be3855f7632bf2897b7cbd5b599653d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 210, + 376, + 219 + ], + "lines": [ + { + "bbox": [ + 323, + 210, + 376, + 219 + ], + "spans": [ + { + "bbox": [ + 323, + 210, + 376, + 219 + ], + "type": "text", + "content": "(c) Where2comm" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 409, + 72, + 518, + 208 + ], + "blocks": [ + { + "bbox": [ + 409, + 72, + 518, + 208 + ], + "lines": [ + { + "bbox": [ + 409, + 72, + 518, + 208 + ], + "spans": [ + { + "bbox": [ + 409, + 72, + 518, + 208 + ], + "type": "image", + "image_path": "c335617087060a539efecf0a6fd025d4ab0ea6dcf9db90ef39a018a93fd71b1f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 210, + 494, + 219 + ], + "lines": [ + { + "bbox": [ + 438, + 210, + 494, + 219 + ], + "spans": [ + { + "bbox": [ + 438, + 210, + 494, + 219 + ], + "type": "text", + "content": "(d) SCOPE (ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 74, + 252, + 185, + 336 + ], + "blocks": [ + { + "bbox": [ + 46, + 222, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 545, + 245 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison results in real-world scenarios from the DAIR-V2X dataset. Green and red boxes denote ground truths and detection results, respectively. Compared to the previous SOTA models, our method achieves more accurate detection results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 252, + 185, + 336 + ], + "lines": [ + { + "bbox": [ + 74, + 252, + 185, + 336 + ], + "spans": [ + { + "bbox": [ + 74, + 252, + 185, + 336 + ], + "type": "image", + "image_path": "0191c4efa9eb134c2cf71b8fd9aaf00c156435b469c5f96577c97aff3f497807.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 337, + 159, + 346 + ], + "lines": [ + { + "bbox": [ + 96, + 337, + 159, + 346 + ], + "spans": [ + { + "bbox": [ + 96, + 337, + 159, + 346 + ], + "type": "text", + "content": "GT Bounding Boxes" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 349, + 546, + 371 + ], + "lines": [ + { + "bbox": [ + 46, + 349, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 546, + 371 + ], + "type": "text", + "content": "Figure 9. Visualization of learned multi-scale deformable cross-attention. Green cross marker denotes the proposed reference point. The sampled keypoints are represented by the brown dots whose colors reflect the corresponding attention weights." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 187, + 252, + 298, + 336 + ], + "blocks": [ + { + "bbox": [ + 187, + 252, + 298, + 336 + ], + "lines": [ + { + "bbox": [ + 187, + 252, + 298, + 336 + ], + "spans": [ + { + "bbox": [ + 187, + 252, + 298, + 336 + ], + "type": "image", + "image_path": "c5e140de898e6e8c59eeb909dd2f4560ef1f4a11e71e8c6c5e0f89aacc99a82d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 337, + 255, + 345 + ], + "lines": [ + { + "bbox": [ + 222, + 337, + 255, + 345 + ], + "spans": [ + { + "bbox": [ + 222, + 337, + 255, + 345 + ], + "type": "text", + "content": "Scale One" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 298, + 252, + 406, + 336 + ], + "blocks": [ + { + "bbox": [ + 298, + 252, + 406, + 336 + ], + "lines": [ + { + "bbox": [ + 298, + 252, + 406, + 336 + ], + "spans": [ + { + "bbox": [ + 298, + 252, + 406, + 336 + ], + "type": "image", + "image_path": "dac8ab962b3d0b89d684f4a857bf9e413d1cbbd42d5848e650b748dd1442c607.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 337, + 368, + 345 + ], + "lines": [ + { + "bbox": [ + 336, + 337, + 368, + 345 + ], + "spans": [ + { + "bbox": [ + 336, + 337, + 368, + 345 + ], + "type": "text", + "content": "Scale Two" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 408, + 252, + 519, + 336 + ], + "blocks": [ + { + "bbox": [ + 408, + 252, + 519, + 336 + ], + "lines": [ + { + "bbox": [ + 408, + 252, + 519, + 336 + ], + "spans": [ + { + "bbox": [ + 408, + 252, + 519, + 336 + ], + "type": "image", + "image_path": "c984beec5d3048727ba5502cd28d9c5988311d57e50b10b45f1c31e7a3471238.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 337, + 482, + 345 + ], + "lines": [ + { + "bbox": [ + 445, + 337, + 482, + 345 + ], + "spans": [ + { + "bbox": [ + 445, + 337, + 482, + 345 + ], + "type": "text", + "content": "Scale Three" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 386, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 482 + ], + "type": "text", + "content": "sitions among features to retain the maximum or average, respectively, and then reshape them into a fused representation. The poor results in Table 3 show that the above mechanisms fail to tackle the complex multi-source feature fusion challenge. In contrast, the superiority of our IAF module is reflected in considering the complementary contributions of each feature at different positions, adaptively assigning higher importance to spatially beneficial semantics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 491, + 176, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 176, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 176, + 502 + ], + "type": "text", + "content": "4.5. Qualitative Evaluation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 509, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 652 + ], + "type": "text", + "content": "Visualization of Detection Results. To intuitively illustrate the perception capability of different models, we provide a detection visualization comparison in Figure 8 under two challenging scenarios. SCOPE achieves more accurate and robust detection results than previous SOTA methods, where the predicted bounding boxes are well aligned to ground truths. In particular, DiscoNet [17], V2X-ViT [36], and Where2comm [10] fail to identify fast-moving objects (i.e., more ground truth bounding boxes are not matched). In contrast, our method achieves holistic perception across heterogeneous agents by effectively aggregating the salient information in the spatio-temporal domains." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "Visualization of Attention Map. We visualize the multiscale deformable cross-attention in the CCC component. As Figure 9 shows, our tailored module can dynamically aggregate the semantics of multiple sampled keypoints at different scales to enhance the visual representation of the" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 386, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 546, + 445 + ], + "type": "text", + "content": "reference point. Concretely, the deformable cross-attention learns to extract object-related valuable perception cues at lower scales by sampling keypoints inside or around targets. In comparison, it reasonably samples long-range features at higher scales to capture spatially critical information." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 455, + 378, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 378, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 378, + 468 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 476, + 546, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 546, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 546, + 632 + ], + "type": "text", + "content": "In this paper, we present SCOPE, a learning-based framework to address existing multi-agent collaborative perception challenges in an end-to-end manner. Our approach is the first time considering the temporal semantics from ego agents to capture valuable context cues through the context-aware component. Moreover, two spatial information aggregation components are introduced to achieve comprehensive information interaction among agents and refined feature fusion of ego agents. Extensive experiments on diverse 3D detection datasets prove the superiority of SCOPE and the effectiveness of our components. The future work focuses on extending SCOPE into more complex perception scenarios." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 645, + 408, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 408, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 408, + 658 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "type": "text", + "content": "This work is supported by the Shanghai Key Research Laboratory of NSAI, the National Natural Science Foundation of China (Grant No. 62250410368), and the Nanjing First Automobile Works Grant." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23390" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3D point clouds. In Proceedings of ACM/IEEE Symposium on Edge Computing, pages 88-100, Arlington Virginia, Nov. 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "type": "text", + "content": "[2] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 255 + ], + "type": "text", + "content": "[3] Abhishek Das, Theophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and Joelle Pineau. Tarmac: Targeted multi-agent communication. In International Conference on Machine Learning (ICML), pages 1538-1546, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 300 + ], + "type": "text", + "content": "[4] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on Robot Learning (CoRL), pages 1-16, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 355 + ], + "type": "text", + "content": "[5] Yangtao Du, Dingkang Yang, Peng Zhai, Mingchen Li, and Lihua Zhang. Learning associative representation for facial expression recognition. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 889-893, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 412 + ], + "type": "text", + "content": "[6] Khaled El Madawi, Hazem Rashed, Ahmad El Sallab, Omar Nasr, Hanan Kamel, and Senthil Yogamani. Rgb and lidar fusion based 3d semantic segmentation for autonomous driving. In Proceedings of the IEEE Intelligent Transportation Systems Conference (ITSC), pages 7-12, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 456 + ], + "type": "text", + "content": "[7] Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson. Learning to communicate with deep multi-agent reinforcement learning. Advances in Neural Information Processing Systems (NIPS), 29, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 287, + 479 + ], + "type": "text", + "content": "[8] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735-1780, 1997. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 512 + ], + "type": "text", + "content": "[9] Yedid Hoshen. Vain: Attentional multi-agent predictive modeling. Advances in Neural Information Processing Systems (NIPS), 30, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "text", + "content": "[10] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Advances in Neural Information Processing Systems (NIPS), 2022. 1, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 611 + ], + "type": "text", + "content": "[11] Yue Hu, Yifan Lu, Runsheng Xu, Weidi Xie, Siheng Chen, and Yanfeng Wang. Collaboration helps camera overtake lidar in 3d detection. arXiv preprint arXiv:2303.13560, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "text", + "content": "[12] Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent cooperation. Advances in Neural Information Processing Systems (NIPS), 31, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[13] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12697-12705, 2019. 1, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 140 + ], + "type": "text", + "content": "[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Proceedings of the European Conference on Computer Vision (ECCV), pages 316-332, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 185 + ], + "type": "text", + "content": "[16] Jinlong Li, Runsheng Xu, Xinyu Liu, Jin Ma, Zicheng Chi, Jiaqi Ma, and Hongkai Yu. Learning for vehicle-to-vehicle cooperative perception under lossy communication. IEEE Transactions on Intelligent Vehicles, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "text", + "content": "[17] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems (NIPS), 34:29541-29552, 2021. 1, 3, 4, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "text", + "content": "[18] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 641-656, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "text", + "content": "[19] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 332, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 387 + ], + "type": "text", + "content": "[20] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4106-4115, 2020. 1, 3, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 443 + ], + "type": "text", + "content": "[21] Guiyang Luo, Hui Zhang, Quan Yuan, and Jinglin Li. Complementarity-enhanced and redundancy-minimized collaboration network for multi-agent perception. In Proceedings of the ACM International Conference on Multimedia (ACMM), pages 3578-3586, 2022. 1, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 445, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 509 + ], + "type": "text", + "content": "[22] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems (NIPS), 32, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "text", + "content": "[23] Tika Ram and Khem Chand. Effect of drivers' risk perception and perception of driving tasks on road safety attitude. Transportation Research Part F: Traffic Psychology and Behaviour, 42:162-176, 2016. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 556, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 600 + ], + "type": "text", + "content": "[24] Andreas Rauch, Felix Klanner, Ralph Rasshofer, and Klaus Dietmayer. Car2x-based perception in a high-level fusion architecture for cooperative perception systems. In IEEE Intelligent Vehicles Symposium, pages 270-275, 2012. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "type": "text", + "content": "[25] Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in multiagent cooperative and competitive tasks. arXiv preprint arXiv:1812.09755, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[26] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation. Advances in Neural Information Processing Systems (NIPS), 29, 2016. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[27] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In International Conference on Machine Learning (ICML), pages 330-337, 1993. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23391" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "text", + "content": "[28] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 605–621, 2020. 1, 3, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 185 + ], + "type": "text", + "content": "[29] Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2xp-asg: Generating adversarial scenes for vehicle-to-everything perception. In IEEE International Conference on Robotics and Automation (ICRA), pages 3584-3591. IEEE, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "text", + "content": "[30] Zhu Xizhou, Su Weijie, Lu Lewei, Li Bin, Wang Xiaogang, and Dai Jifeng. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[31] Runsheng Xu, Weizhe Chen, Hao Xiang, Xin Xia, Lantao Liu, and Jiaqi Ma. Model-agnostic multi-agent perception framework. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1471–1478. IEEE, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "type": "text", + "content": "[32] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "type": "text", + "content": "[33] Runsheng Xu, Jinlong Li, Xiaoyu Dong, Hongkai Yu, and Jiaqi Ma. Bridging the domain gap for multi-agent perception. arXiv preprint arXiv:2210.08451, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 422 + ], + "type": "text", + "content": "[34] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. In Conference on Robot Learning (CoRL), 2022. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 489 + ], + "type": "text", + "content": "[35] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13712–13722, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[36] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 612 + ], + "type": "text", + "content": "[37] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In Proceedings of the International Conference on Robotics and Automation (ICRA), pages 2583–2589, 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 679 + ], + "type": "text", + "content": "[38] Dingkang Yang, Zhaoyu Chen, Yuzheng Wang, Shunli Wang, Mingcheng Li, Siao Liu, Xiao Zhao, Shuai Huang, Zhiyan Dong, Peng Zhai, and Lihua Zhang. Context deconfounded emotion recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19005-19015, June 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[39] Dingkang Yang, Shuai Huang, Haopeng Kuang, Yangtao Du, and Lihua Zhang. Disentangled representation learning for multimodal emotion recognition. In Proceedings of the 30th" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 652 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ACM International Conference on Multimedia (ACM MM), page 1642-1651, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[40] Dingkang Yang, Shuai Huang, Yang Liu, and Lihua Zhang. Contextual and cross-modal interaction for multi-modal speech emotion recognition. IEEE Signal Processing Letters, 29:2093-2097, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "text", + "content": "[41] Dingkang Yang, Shuai Huang, Shunli Wang, Yang Liu, Peng Zhai, Liuzhen Su, Mingcheng Li, and Lihua Zhang. Emotion recognition for multiple context awareness. In Proceedings of the European Conference on Computer Vision (ECCV), volume 13697, pages 144-162. Springer, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 197, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 252 + ], + "type": "text", + "content": "[42] Dingkang Yang, Haopeng Kuang, Shuai Huang, and Lihua Zhang. Learning modality-specific and -agnostic representations for asynchronous multimodal language sequences. In Proceedings of the 30th ACM International Conference on Multimedia (ACMM), page 1708-1717, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 253, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 317 + ], + "type": "text", + "content": "[43] Dingkang Yang, Yang Liu, Can Huang, Mingcheng Li, Xiao Zhao, Yuzheng Wang, Kun Yang, Yan Wang, Peng Zhai, and Lihua Zhang. Target and source modality co-reinforcement for emotion understanding from asynchronous multimodal sequences. Knowledge-Based Systems, page 110370, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 319, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 374 + ], + "type": "text", + "content": "[44] Kun Yang, Jing Liu, Dingkang Yang, Hanqi Wang, Peng Sun, Yanni Zhang, Yan Liu, and Liang Song. A novel efficient multi-view traffic-related object detection framework. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 430 + ], + "type": "text", + "content": "[45] Kun Yang, Peng Sun, Jieyu Lin, Azzedine Boukerche, and Liang Song. A novel distributed task scheduling framework for supporting vehicular edge intelligence. In IEEE International Conference on Distributed Computing Systems (ICDCS), pages 972-982, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 431, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 431, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 431, + 545, + 506 + ], + "type": "text", + "content": "[46] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21361-21370, 2022. 1, 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 508, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 508, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 308, + 508, + 545, + 563 + ], + "type": "text", + "content": "[47] Zhenxun Yuan, Xiao Song, Lei Bai, Zhe Wang, and Wanli Ouyang. Temporal-channel transformer for 3d lidar-based video object detection for autonomous driving. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2068-2078, 2021. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 564, + 545, + 597 + ], + "type": "text", + "content": "[48] Zixu Zhang and Jaime F Fisac. Safe occlusion-aware autonomous driving via game-theoretic active perception. arXiv preprint arXiv:2105.08169, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 598, + 545, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 598, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 308, + 598, + 545, + 652 + ], + "type": "text", + "content": "[49] Zixiang Zhou, Xiangchen Zhao, Yu Wang, Panqu Wang, and Hassan Foroosh. Centerformer: Center-based transformer for 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 496-513, 2022. 1, 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23392" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/6b38d58c-2dfe-4bda-8611-478cf7d86d34_content_list.json b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/6b38d58c-2dfe-4bda-8611-478cf7d86d34_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef7f5e0146437eadc361e0544df945cea706d2c3 --- /dev/null +++ b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/6b38d58c-2dfe-4bda-8611-478cf7d86d34_content_list.json @@ -0,0 +1,1570 @@ +[ + { + "type": "text", + "text": "Spatio-temporal Prompting Network for Robust Video Feature Extraction", + "text_level": 1, + "bbox": [ + 109, + 130, + 857, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guanxiong Sun $^{1,2}$ , Chi Wang $^{1}$ , Zhaoyu Zhang $^{1}$ , Jiankang Deng $^{2,3}$ , Stefanos Zafeiriou $^{3}$ , Yang Hua $^{1}$", + "bbox": [ + 84, + 191, + 880, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Queen's University Belfast $^{2}$ Huawei UKRD $^{3}$ Imperial College London", + "bbox": [ + 158, + 222, + 808, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gsun02, cwang38, zhang55,Y.Hua} @qub.ac.uk, {j.deng16,s.zafeiriou} @imperial.ac.uk", + "bbox": [ + 133, + 241, + 836, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 305, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Frame quality deterioration is one of the main challenges in the field of video understanding. To compensate for the information loss caused by deteriorated frames, recent approaches exploit transformer-based integration modules to obtain spatio-temporal information. However, these integration modules are heavy and complex. Furthermore, each integration module is specifically tailored for its target task, making it difficult to generalise to multiple tasks. In this paper, we present a neat and unified framework, called Spatio-Temporal Prompt Network (STPN). It can efficiently extract robust and accurate video features by dynamically adjusting the input features in the backbone network. Specifically, STPN predicts several video prompts containing spatio-temporal information of neighbour frames. Then, these video prompts are pretended to the patch embeddings of the current frame as the updated input for video feature extraction. Moreover, STPN is easy to generalise to various video tasks because it does not contain task-specific modules. Without bells and whistles, STPN achieves state-of-the-art performance on three widely-used datasets for different video understanding tasks, i.e., ImageNetVID for video object detection, YouTubeVIS for video instance segmentation, and GOT-10k for visual object tracking. Codes are available at https://github.com/guanxiongsun/STPN", + "bbox": [ + 75, + 339, + 473, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 752, + 209, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video understanding is a fundamental research direction in the field of computer vision. It plays an important role in many real-world applications, such as autonomous driving [64, 32], video surveillance [34, 5], and sports analysis [56, 25]. However, a significant challenge in this field is the deterioration of video frames due to motion blur, occlusion, and deformation, which makes it difficult to extract relevant information.", + "bbox": [ + 75, + 779, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c51f56c56ea88b261e25340276a14265d27c959fa6d11c2f69d511fe7150034.jpg", + "image_caption": [ + "(a) Current pipeline" + ], + "image_footnote": [], + "bbox": [ + 501, + 309, + 699, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a951f0b22ba2fe2e8fa03193987df5e5509e7638d5ad9b1bdeecef2a83eaf4db.jpg", + "image_caption": [ + "(b) STPN", + "Figure 1. Comparisons between pipelines of (a) existing methods and (b) the proposed Spatio-temporal Prompting Network (STPN). Existing methods introduce complex and task-specific integration modules (□ □ □) after backbone networks (□). In contrast, STPN is a unified framework for multiple tasks. A lightweight dynamic video prompt (DVP) predictor (□) generates a set of DVPs to adjust input before backbone networks (□). Best viewed in colour." + ], + "image_footnote": [], + "bbox": [ + 709, + 318, + 883, + 438 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome this challenge, inspired by the great success of transformers in many computer vision tasks [49, 54, 27, 19, 4], researchers explore various transformer-based integration modules to alleviate the information loss on the deteriorated video frames. For example, in video object detection, transformer-based feature aggregation methods [13, 7, 57, 43] are investigated to enhance the feature of proposals [40] in the detection head. In video instance segmentation, deteriorated frames usually cause wrong instance associations in a sequence, so 3D mask decoders [55, 22, 8] are introduced to learn associations of instance masks in an end-to-end manner. For visual object tracking, traditional correlation filters [26, 47] meet their limitations in severely degraded frames. Therefore, transformer-based integration modules [11, 62, 6, 10] are proposed to better capture complicated correlations between the template region and the search region. To conclude, the summarised pipeline of recent methods is shown in Figure 1 (a). A backbone network extracts spatially-only features from the current frame and", + "bbox": [ + 496, + 613, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13587", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the support frames in the same video. Then, different integration modules integrate spatial-only features on multiple frames to obtain spatio-temporal features. Finally, the spatio-temporal features are used for the targeted video understanding task.", + "bbox": [ + 75, + 90, + 467, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the current pipeline achieves good performance, there are still two major limitations of transformer-based integration modules. Firstly, these integration modules are involved as an extra component after the backbone network, leading to increased complexity and additional computational costs. Secondly, each integration module is tailored specifically for the target task and thus cannot be generalised to multiple video tasks. Hence, we put forward the following question: can we remove the complex integration modules and directly obtain spatio-temporal information in backbone networks?", + "bbox": [ + 75, + 167, + 467, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer this question, inspired by recent prompting techniques [24, 23, 28], in this paper, we present a neat and unified framework, called Spatio-Temporal Prompting Network (STPN). Instead of using complex integration modules, STPN simplifies the current pipeline for video understanding by introducing spatio-temporal information into the backbone network, as shown in Figure 1 (b). Specifically, given a video frame, we propose a dynamic video prompt (DVP) predictor to generate several video prompts according to support frames. Then, the predicted DVPs are prepended to the patch embeddings of the current frame as the updated input. Finally, a vision transformer backbone network extracts video features using the updated input for future video understanding tasks. It is worth noting that the DVP predictor is a lightweight structure and introduces only a small number of extra parameters, e.g., 0.11M. Moreover, STPN can be easily adapted to different video understanding tasks, since it does not contain task-specific modules, and all modifications happen before the backbone network.", + "bbox": [ + 75, + 332, + 467, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are: (1) We present the Spatio-Temporal Prompting Network (STPN) that can extract robust video features on deteriorated video frames. STPN simplifies the current pipeline for video understanding and is easy to generalise to different video understanding tasks. (2) To the best of our knowledge, we are the first to explore promoting techniques for robust video feature extraction on the task of video object detection (VOD), video instance segmentation (VIS), and visual object tracking (VOT). (3) Without bells and whistles, STPN achieves state-of-the-art performance on three widely-used video understanding benchmarks, i.e., ImageNet-VID [41] for VOD, YouTube-VIS [63] for VIS, and GOT-10k [21] for VOT.", + "bbox": [ + 75, + 619, + 467, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 829, + 217, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first review the state-of-the-art (SOTA) methods for related video understanding tasks. Then, we introduce the visual prompting techniques.", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Object Detection (VOD). SOTA VOD methods use attention modules to aggregate features of support frames onto the current frame and thus enhance the feature quality of the current frame. For example, RDN [13] and SELSA [57] utilise relation modules to aggregate proposal features from local frames and global frames, respectively. These methods use sample memory architectures to store features on support frames, e.g., a sliding window or a memory queue, which limit the \"diversity\" of support features. Therefore, MAMBA [43] proposes a memory bank architecture that can partially update and sample support features and thus achieve a good speed-accuracy trade-off. More recent approaches TransVOD [17] and TDViT [45] introduce transformers into the field and further improve the performance of VOD.", + "bbox": [ + 496, + 108, + 890, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Instance Segmentation (VIS). The goal of VIS is to simultaneously segment and track all object instances in the video sequence. There are two categories of approaches: Firstly, per-frame approaches conduct instance segmentation on each frame independently and then associate instance masks by post-processing. MaskTrackRCNN [63] adds a tracking prediction head on MaskRCNN [16] for instance association. More recently, some approaches [60, 20, 18] use transformers [49, 4] to reason the relationships between instances. Secondly, per-clip approaches [55, 59, 8, 22] directly predict 3D masks for each instance in a video. VisTR [55] proposes a transformer encoder-decoder structure to obtain spatio-temporal features, which are then used in an instance sequence matching module to get object tracks. IFC [22] uses memory tokens inside of the transformer encoder module to transfer temporal information between input frames.", + "bbox": [ + 496, + 352, + 890, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Object Tracking (VOT). Object tracking is one of the most fundamental research topics in computer vision and has a long history. Siamese-based methods [2, 30, 29, 53, 67] consider tracking as a template-matching problem by extracting correlation feature maps of the target and the search region via two-head siamese networks. Recently, transformers have been introduced to perform attention-based correspondence modelling and these methods [6, 51, 10, 62] achieve SOTA performance. Among these trackers, MixFormer [11] is a unified framework established solely on transformers for feature extraction, correspondence modelling, and result generation.", + "bbox": [ + 496, + 626, + 890, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompting. The concept of prompting is originally proposed in natural language processing (NLP). Prompting refers to designing instructions and preponding them to the input so that the pre-trained language models [38, 14, 3] can \"understand\" the task. Recently, prompting techniques", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13588", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/fa96e9674be6aa0356ecce6f27806d54b0ec77b578d07407c49314c8cad22559.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MeaningSymbol
The i-th transformer layerLi
Collection of the output embeddings from LiEi
Number of embeddings in Eini
Dimension of embeddings in Eidi
Collection of dynamic video promptsP
Number of dynamic video promptsNP
", + "bbox": [ + 78, + 88, + 467, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. List of notations.", + "bbox": [ + 194, + 219, + 349, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "have been explored in many computer vision tasks. A major trend [35, 36, 24] is to transfer pre-trained vision-language models [15, 37, 39] to the task of video recognition, e.g., video classification and action recognition. Another popular research direction is parameter-efficient transfer learning via prompt tuning. For example, VPT [23] and UPT [65] tune a small number of parameters in the input space of a model and can achieve good performance comparable to the full fine-tuning scheme on a variety of recognition tasks.", + "bbox": [ + 75, + 255, + 472, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 404, + 215, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we define notations necessary for subsequent demonstrations. $I_{t}$ denotes the current frame and $t$ is the current time step in the video. $I_{sup}$ denotes the support frames used to predict video prompts and aid in the feature extraction process. Specifically, $I_{sup} = \\{I_{\\tau}\\}_{t - S(K / 2)}^{t + S(K / 2)}$ , where $K$ denotes the number of support frames (assumed to be an even number) and $S$ denotes the frame interval in the time dimension between support frames.", + "bbox": [ + 75, + 429, + 468, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformers encoders [27, 33, 58] are used to extract feature embeddings from video frames. A transformer encoder contains a patch embedding layer and $L$ transformer layers. The input frame is first divided into $n$ non-overlapping patches $\\{\\mathbf{x}_j\\in \\mathbb{R}^{3\\times h\\times w}|j\\in \\mathbb{N},1\\leq j\\leq n\\}$ where $h$ and $w$ are the height and width of the patches. Each patch is then embedded into patch embeddings, denoted as $\\mathbf{E}\\in \\mathbb{R}^{n_0\\times d_0}$ . Other notations are summarised in Table 1.", + "bbox": [ + 75, + 551, + 468, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 76, + 686, + 210, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 710, + 187, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of our framework is shown as in Figure 2 (a). There are two stages in the framework: the predicting stage and the prompting stage. The goal of the predicting stage is to generate a set of prompts that vary according to spatio-temporal information. We first select $K$ support frames around $I_{t}$ and then pass the support frames $I_{sup}$ to the patch embedding layer and the transformer encoder to extract image embeddings, called support embeddings. The support embeddings are then passed to the dynamic video prompt (DVP) predictor network as the input. The DVP predictor outputs $N_{P}$ dynamic video prompts, denoted as", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{P}$ , which contain the spatio-temporal information of the current video. Details of the DVP predictor network are illustrated in $\\S 4.2$ .", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second stage is the prompting stage whose goal is to extract spatio-temporal features of the current frame $I_{t}$ using $\\mathbf{P}$ . Specifically, $\\mathbf{P}$ is prepended to the patch embeddings of the current frame $I_{t}$ . Then, the mixed embeddings are passed into the transformer encoder to produce the spatiotemporal embeddings. Finally, the spatio-temporal embeddings are used for different video understanding tasks, e.g., video object detection, video instance segmentation and visual object tracking. More details of the prompting stage are presented in §4.3.", + "bbox": [ + 496, + 137, + 892, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Predicting Stage", + "text_level": 1, + "bbox": [ + 498, + 299, + 663, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the current frame $I_{t}$ , we first sample $K$ support frames around $I_{t}$ as support frames. The support frames are then passed into the patch embedding layer and the transformer encoder:", + "bbox": [ + 496, + 323, + 890, + 383 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} _ {\\tau} ^ {i} = \\mathbf {L} ^ {i} \\left(\\mathbf {E} _ {\\tau} ^ {i - 1}\\right) \\quad \\mathbf {E} _ {\\tau} ^ {i} \\in \\mathbb {R} ^ {n _ {i} \\times d _ {i}}, i = 1, 2, \\dots , L. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 396, + 890, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, the collection of support embeddings is denoted as: $\\mathbf{E}_{sup} = \\{\\mathbf{E}_{\\tau}^{\\mathbf{L}} \\in \\mathbb{R}^{n_L \\times d_L} | \\tau \\in \\mathcal{T}\\}$ , where $\\mathcal{T}$ is the collection of sampled time steps around the current frame, $\\mathcal{T} = [t - S(K/2), \\dots, t - S, t + S, \\dots, t + S(K/2)]$ . For example, if $K = 6$ and $S = 2$ , three support frames are sampled from the past of the current time step $t$ , and three are sampled from the future, with a time interval of 2 between each sampled frame in both directions. Given $\\mathbf{E}_{sup}$ , we design two predicting networks, i.e., transformer-based predictor and mixer-based predictor, to obtain dynamic video prompts $P \\in \\mathbb{R}^{N_P \\times d}$ .", + "bbox": [ + 496, + 426, + 892, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer-based Predictor. This design is inspired by the object decoder in DETR [4] where a set of randomly initialised object queries are used to retrieve object features on the image embeddings. As depicted in Figure 2 (b), we introduce a set of learnable prompt queries $Q \\in \\mathbb{R}^{N_P \\times d_L}$ to gather spatio-temporal information from the support embeddings. Specifically, a multi-head attention (MHA) [49] module takes the support embeddings and the prompt queries as input, and produces the aggregated embeddings. The process is formulated as follows:", + "bbox": [ + 496, + 609, + 890, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nK = \\operatorname {C o n c a t} \\left(\\mathbf {E} _ {\\sup }\\right) \\quad K \\in \\mathbb {R} ^ {K \\cdot n _ {L} \\times d _ {L}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 771, + 890, + 791 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Q} = \\operatorname {M H A} (\\ln (Q), \\ln (K)) + Q \\quad \\hat {Q} \\in \\mathbb {R} ^ {N _ {P} \\times d _ {L}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 792, + 890, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{Concat}(\\cdot)$ indicates the concatenation operation and LN denotes the LayerNorm layer. The output embeddings are then added back to the prompt queries and passed to the LayerNorm layer. The normalised output embeddings are further transformed by a simple feed-forward network", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13589", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ead93793a658ca120a22f9feca0ba40bbdeb68013e09e969ae7b4cc6824c71c0.jpg", + "image_caption": [ + "(a) Overview" + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 549, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f240f4df6c29134140ab3c4a679019ffbffa4a8bb8b150fd65afaefaaa835021.jpg", + "image_caption": [ + "(b) Transformer-based predictor", + "(c) Mixer-based predictor", + "Figure 2. (a) An overview of our approach. In the predicting stage, a set of dynamic visual prompts (DVPs) $\\mathbf{P}$ is generated by the DVP predictor that takes support embeddings $\\mathbf{E}_{sup}$ on support frames $I_{sup}$ as the input. Then, in the prompting stage, predicted DVPs are prepended with the patch embeddings of the current frame to extract spatio-temporal embeddings via a transformer encoder which contains $L$ transformer layers. Finally, different task heads take the spatio-temporal embeddings and output final results for various general video tasks, e.g., video object detection, video instance segmentation, and visual object tracking. (b) Details of the transformer-based predictor. (c) Details of the Mixer-based predictor." + ], + "image_footnote": [], + "bbox": [ + 550, + 95, + 846, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(FFN) with a residual connection that makes the dynamic video prompts:", + "bbox": [ + 75, + 525, + 468, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\operatorname {F F N} (\\operatorname {L N} (\\hat {Q})) + \\operatorname {L N} (\\hat {Q}) \\quad \\mathbf {P} \\in \\mathbb {R} ^ {N _ {P} \\times d}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 559, + 468, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mixer-based Predictor. Inspired by the MLP-Mixer [46] (shortened as Mixer), we design a Mixer-based predictor to transform support embedding into dynamic video prompts. The details of a Mixer-based predictor are shown in Figure 2 (c). Unlike the transformer-based predictor whose first step is to concatenate all support embeddings, we conduct average pooling in the temporal dimension to obtain the average support embeddings $\\overline{\\mathbf{E}}_{sup}$ as follows:", + "bbox": [ + 75, + 583, + 468, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {E}}} _ {s u p} = \\operatorname {A v g P o o l} \\left(\\mathbf {E} _ {s u p}\\right) \\quad \\overline {{\\mathbf {E}}} _ {s u p} \\in \\mathbb {R} ^ {n _ {L} \\times d _ {L}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 709, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where AvgPool denotes the average pooling across the time dimension. Then, the average support embedding is normalised by a LayerNorm layer and further transposed and passed into a feed-forward network (FFN) to obtain the hidden feature $h$ with $N_P$ channels. Finally, the hidden feature $h$ is used to generate $N_P$ dynamic video prompts with $d$ channels. These two steps can be formulated as follows:", + "bbox": [ + 75, + 732, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh = \\operatorname {F F N} \\left(\\left(\\operatorname {L N} \\left(\\overline {{\\mathbf {E}}} _ {s u p}\\right)\\right) ^ {T}\\right) \\quad h \\in \\mathbb {R} ^ {d _ {L} \\times N _ {P}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 857, + 468, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\operatorname {F F N} \\left(\\left(\\operatorname {L N} (h) ^ {T}\\right) \\quad \\mathbf {P} \\in \\mathbb {R} ^ {N _ {P} \\times d}, \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 878, + 468, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where LN and $(\\cdot)^T$ denote the LayerNorm layer [1] and the transpose layer, respectively.", + "bbox": [ + 500, + 523, + 888, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Prompting Stage", + "text_level": 1, + "bbox": [ + 500, + 561, + 666, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the predicted dynamic video prompts $\\mathbf{P}$ , we concatenate $\\mathbf{P}$ together with the patch embeddings of the current frame $\\mathbf{E}_t$ as the input of the transformer encoder. The details of this process are shown in the prompting stage (right part) of Figure 2 (a). More concretely, the process of applying $\\mathbf{P}$ and extracting the spatio-temporal feature of the current frame $I_t$ can be formulated as:", + "bbox": [ + 496, + 585, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {Z} _ {t} ^ {1}, \\mathbf {E} _ {t} ^ {1} \\right] = \\mathbf {L} ^ {1} (\\operatorname {C o n c a t} (\\mathbf {P}, \\mathbf {E} _ {t})) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 696, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {Z} _ {t} ^ {i}, \\mathbf {E} _ {t} ^ {i} \\right] = \\mathbf {L} ^ {i} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {Z} ^ {i - 1}, \\mathbf {E} _ {t} ^ {i - 1}\\right)\\right) \\quad i = 2, 3, \\dots , L \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 718, + 890, + 738 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = \\operatorname {H e a d} \\left(\\mathbf {E} _ {t} ^ {L}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 739, + 890, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Concat}(\\cdot)$ indicates the concatenation operation and $\\mathbf{Z}_t^i\\in \\mathbb{R}^{N_P\\times d_i}$ represents the extra embeddings produced by adding $\\mathbf{P}$ in the input embeddings. The weights of the transformer encoder are shared in the predicting and prompting stages. Finally, the output embeddings $\\mathbf{E}_t^L$ are passed into the head network for different video tasks.", + "bbox": [ + 496, + 763, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "STPN Shallow v.s. STPN Deep. Following the protocols in VPT [23], we design two STPN variants, i.e., STPN shal", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13590", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cdbc7086953477a806bd7a791abc73e5ba842f6f738957c74fbb450ea128d27a.jpg", + "image_caption": [ + "(a) SPTN shallow", + "Figure 3. A comparison between the (a) STPN shallow and the (b) STPN deep. The green rounded rectangles denote predicted DVPs. In (a), one set of DVPs is prepended to the patch embeddings before the first layer of the transformer encoder. In (b), $L$ sets of DVPs are predicted and then prepended to the input embeddings of all $L$ transformer layers in the transformer encoder." + ], + "image_footnote": [], + "bbox": [ + 109, + 87, + 274, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4077239bead0aa142b0b0ca06cc2ad4721a0f54e4cd63a2508fd3bf891ee8615.jpg", + "image_caption": [ + "(b) SPTN deep" + ], + "image_footnote": [], + "bbox": [ + 282, + 88, + 450, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "low $v.s.$ STPN deep, according to how many sets of dynamic video prompts (DVPs) are generated to adjust the transformer encoder. For example, in the shallow version, one set of DVPs is predicted and prepended only before the first transformer layer. In the deep version, $L$ sets of DVPs are predicted via $L$ separate FC layers, and each set of DVPs is prepended to the corresponding transformer layer. Formally, the DVPs of the $i$ -th layer are denoted as $\\mathbf{P}^i \\in \\mathbb{R}^{N_P \\times d_i}$ and the feature extraction process Equation (9) is changed to:", + "bbox": [ + 75, + 356, + 468, + 507 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\dots , \\mathbf {E} _ {t} ^ {i} \\right] = \\mathbf {L} ^ {i} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {P} ^ {i - 1}, \\mathbf {E} _ {t} ^ {i - 1}\\right)\\right) \\quad i = 2, 3, \\dots , L. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 516, + 468, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The difference between the shallow and deep variants is shown in Figure 3.", + "bbox": [ + 75, + 542, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 585, + 207, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first conduct experiments and compare SPTN with other state-of-the-art methods on three general video understanding tasks, i.e., video object detection (VOD), video instance segmentation (VIS) and visual object tracking (VOT). Then, in §5.3, we conduct ablation studies on how different design choices affect the overall performance, which helps to gain a deeper understanding of SPTN.", + "bbox": [ + 75, + 611, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Settings", + "text_level": 1, + "bbox": [ + 76, + 726, + 174, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. For VOD, the ImageNetVID dataset [41] is used. It contains 3,862 training and 555 validation videos. Following previous approaches [43, 57, 66], we add overlapped 30 classes of the ImageNet DET dataset into the train set. Specifically, we sample 15 frames from each video in the VID train and at most 2,000 images per class from the DET train. For VIS, we train our models and report results on the YouTube VIS 2019 dataset [63]. The dataset contains 2,238 training, 302 validation, and 343 test videos, which covers 40 object categories. For VOT, we", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "choose the GOT10k [21] dataset, because it is a large-scale dataset (10,000 training videos and 180 test videos) with more than 1.5 million manually labelled bounding boxes, enabling stable training and evaluation of trackers.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. For video object detection and video instance segmentation, we report detailed results in the COCO [31] evaluation format. Specifically, we report the average precision (AP) metric which computes the average precision over ten IoU thresholds $[0.5:0.05:0.95]$ for all categories. Meanwhile, the COCO evaluation contains other important metrics, e.g., $\\mathrm{AP}_{50}$ and $\\mathrm{AP}_{75}$ that are calculated at IoU thresholds of 0.50 and 0.75, and $\\mathrm{AP}_S$ , $\\mathrm{AP}_M$ , $\\mathrm{AP}_L$ that are calculated on different object scales, i.e., small, medium and large. The metrics for detection and segmentation are noted with box and mask, respectively. For visual object tracking, we report the widely used average overlap (AO) and the success rate (SR). AO denotes the average overlaps between all ground truth and estimated bounding boxes, while SR measures the percentage of successfully tracked frames where the overlaps exceed a threshold, e.g., $\\mathrm{SR}_{0.5}$ and $\\mathrm{SR}_{0.75}$ .", + "bbox": [ + 496, + 167, + 890, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer Encoders. Our method is compatible with different transformer backbones. For demonstration, we conduct experiments on two well-known transformer backbones, i.e., Swin [33] and CvT [58]. Swin is used for video object detection and video instance segmentation. CvT is used for visual object tracking. The details of the backbones are illustrated in the supplementary material.", + "bbox": [ + 496, + 440, + 890, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training and Inference. All reported results are obtained with Python 3.7 and PyTorch 1.8.1 on Tesla V100 GPUs. We follow the protocols in TransVOD [17], Mask2Former [9], and MixFormer [11] to set the hyper-parameters, e.g., learning rate, data augmentations, optimiser, etc., for three tasks, respectively. More details about the hyper-parameter settings are described in the supplementary material.", + "bbox": [ + 496, + 561, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Comparisons with SOTA Methods", + "text_level": 1, + "bbox": [ + 498, + 680, + 799, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Object Detection. Table 2 shows the results of STPN and state-of-the-art methods. Firstly, in the AP50 metric, using a simple FasterRCNN [40] base detector with the Swin-T [33] backbone, STPN achieves $85.0\\%$ of AP50, which makes a $1.3\\%$ improvement over the best competitor TransVOD [17]. Furthermore, when changing the FasterRCNN base detector to a multi-frame aggregation method SELSA [57], STPN increases further to $86.5\\%$ of AP50. When using a more strict AP metric, STPN with FasterRCNN and SELSA achieve remarkable improvements of $+6.5 / 6.9\\%$ over the best competitor EOVOD [44], respectively. Finally, when changing to stronger backbones, STPN with the Swin-B backbone achieves further improved", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13591", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/748f95824273036ff409bb9df00f87852de42152fd59399b1f0d39370c128694.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBase DetectorBackboneAPAP50AP75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
CHP [61]CenterNetR101-76.7----
EOVOD [44]FCOSR10154.179.859.510.528.360.1
RDN [13]FasterRCNNR10153.481.860.18.527.459.6
TransVOD [17]Deformable DETRR101-82.0----
MEGA [7]FasterRCNNR10153.282.959.29.129.459.1
TDViT [45]FasterRCNNSwinT50.979.955.79.126.957.2
SELSA* [57]FasterRCNNSwinT52.382.358.211.428.857.9
TransVOD [17]Deformable DETRSwinT-83.7----
STPNFasterRCNNSwinT60.685.068.812.133.766.8
STPNSELSASwinT60.986.568.613.334.867.1
MEGA [7]FasterRCNNResNeXt101-84.1----
MAMBA [43]FasterRCNNResNeXt101-86.7----
TransVOD [17]Deformable DETRSwinB-90.0----
STPNFasterRCNNSwinB63.686.171.612.133.870.3
STPNSELSASwinB65.290.673.514.137.771.8
", + "bbox": [ + 135, + 85, + 831, + 358 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8684c4596a2efdf564538e408df8c299a67f660d123f6cf0368d3e0ad4b63ca5.jpg", + "table_caption": [ + "Table 2. Comparison with state-of-the-art methods on ImageNet VID for video object detection. * represents our re-implementation." + ], + "table_footnote": [], + "table_body": "
ModelBackboneAPAP50AP75FPS
VisTR [55]R10140.164.045.057.7
CrossVISR10136.657.339.735.6
MaskPropR10142.5-45.6-
SeqFormerR10149.071.155.764.6
IFC [22]R10144.669.249.589.4
IDOL [60]R10150.173.156.126.0
VITA [18]R10151.975.457.0-
M2F-VIS [8]SwinT51.575.056.5-
MinVIS [20]SwinT51.975.158.227.1
STPNMinVISSwinT54.878.761.026.5
", + "bbox": [ + 78, + 406, + 465, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performance of $65.2\\%$ AP. Compared to TransVOD, STPN with SELSA makes a $0.6\\%$ improvement in AP50.", + "bbox": [ + 76, + 652, + 467, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video Instance Segmentation. We compare STPN with state-of-the-art methods in Table 3. Using the M2F-VIS [8] as the baseline segmentation method, STPN improves performance by large margins with $2.9 / 3.6\\%$ of AP/AP50, respectively. Compared with SOTA methods with similar backbone complexities, for example, Swin-T or ResNet-101, STPN achieves the best accuracy. Specifically, STPN improves the performance of the best competitor VITA [18] by $2.9 / 3.3\\%$ of AP/AP50, respectively.", + "bbox": [ + 76, + 700, + 467, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Object Tracking. Table 4 compares STPN with state-of-the-art methods. We use MixFormer [11] as our baseline. Specifically, when CVT-22k and CVT-1k are used", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0aac90a83b307dc136326284c7ad39783a695972cf7438d529adc71e330b8ce5.jpg", + "table_caption": [ + "Table 3. Comparison with state-of-the-art methods for video instance segmentation on the YouTube VIS 2019 dataset." + ], + "table_footnote": [], + "table_body": "
ModelBackboneAOSR0.5SR0.75
SiamRPN++ [29]R10151.761.632.5
SiamR-CNN [50]R10164.972.859.7
PrDiMP [12]R10163.473.854.3
TrDiMP [52]R10167.177.758.3
TREG [10]R10166.877.857.2
TransT [6]R10167.176.860.9
STARK [62]R10168.878.164.1
MixFormer[11]CVT-22K70.780.067.8
MixFormer [11]CVT-1K71.279.965.8
STPNMixFormerCVT-22K71.881.669.0
STPNMixFormerCVT-1K72.781.967.7
", + "bbox": [ + 501, + 406, + 890, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Comparison with state-of-the-art tracking methods on the GOT-10k dataset.", + "bbox": [ + 500, + 614, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as the backbone, STPN improves the AO of MixFormer by $+1.1\\%$ and $+1.5\\%$ , respectively. STPN achieves the best performance compared to other SOTA methods on the GOT-10K dataset.", + "bbox": [ + 500, + 667, + 890, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These experimental results demonstrate that STPN is easy to generalise to various video understanding tasks and achieves remarkable performance across all three different video tasks.", + "bbox": [ + 500, + 729, + 890, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 800, + 651, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We investigate the effect of different designs of our approach on the overall performance. All experiments in §5.3 are conducted on the ImageNet VID dataset for video object detection.", + "bbox": [ + 500, + 824, + 890, + 885 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13592", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1742272ceb30c4fa5cacb8a90d74ec157f4b54d88e0bcae83df0fe71bdcd75d1.jpg", + "image_caption": [ + "(a) STPN with Swin-T backbone" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 344, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d35e8a0c5af68a5f91bb35cfefab442cae421062b4d9d594962028686558e1c2.jpg", + "image_caption": [ + "(b) STPN with Swin-B backbone" + ], + "image_footnote": [], + "bbox": [ + 349, + 90, + 620, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8528e5a397210b5681701b2b60a2805689c6295a707676118c9df7053ee1fd4d.jpg", + "image_caption": [ + "(c) Comparison with STOA" + ], + "image_footnote": [], + "bbox": [ + 624, + 94, + 887, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d02caf1694f3ee160a8f9cfa7c658e7b62745af656d84cbee610f8bee087dc76.jpg", + "image_caption": [ + "Figure 4. Comparisons between the transformer-based and the Mixer-based DVP predictors.", + "Figure 5. Comparisons of STPN deep and STPN shallow on transformer encoders with different scales." + ], + "image_footnote": [], + "bbox": [ + 83, + 306, + 464, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DVP Predictor Designs. We compare the effects of DVP predictor designs on different baselines. Figure 4 (a) shows the results on the Swin-T backbone. The FasterRCNN baseline achieves $80.7\\%$ AP50, runs at 23.12 FPS, and has 44.90M parameters. Using the transformer-based predictor and the Mixer-based predictor, STPN achieves $85.0 / 84.5\\%$ AP50 and runs at 22.27/22.41 FPS and has 45.01/44.96M parameters, respectively. These results show that the transformer-based predictor can achieve better accuracy, while the Mixer-based predictor has less complexity and faster speed. Experiments on the Swin-B backbone are shown in Figure 4 (b) and the experimental results indicate the same conclusion.", + "bbox": [ + 75, + 518, + 468, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency of STPN. Compared with other state-of-the-art methods, STPN with both predictor designs actually achieves a significantly improved speed-accuracy trade-off, as shown in Figure 4 (c). These results demonstrate that the pipeline of STPN is very efficient and is not sensitive to different predictor designs. We choose the transformer-based predictor as our default setting since it achieves a slightly higher accuracy with a comparable fast speed.", + "bbox": [ + 75, + 731, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "STPN Shallow v.s. STPN Deep. As shown in the Figure 5, we conduct experiments on Swin-T/S/B/L variants. The", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b4b28ef148b8349a632c9c31aab96145aae58dddd56000d85815369fde5b6c9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(a)S12481632
AP58.759.660.160.660.560.5
(b)K1357911
AP52.653.558.960.660.660.7
FPS22.422.422.322.321.720.5
(c)NP35791113
AP60.260.360.660.660.560.6
FPS22.922.722.321.621.019.9
", + "bbox": [ + 504, + 303, + 888, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Effect of hyper-parameters in STPN: the temporal stride $S$ , the number of support frames $K$ , and the number of dynamic video prompts $N_P$ . Different choices of hyper-parameters are listed in rows highlighted in grey.", + "bbox": [ + 498, + 458, + 890, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "first finding is that both STPN shallow and STPN deep can continuously improve performance on all Swin variants, indicating that STPN has good compatibility with backbone scales. Another finding which contradicts our intuition is that surprisingly STPN shallow achieves better performance than STPN deep on three Swin variants. We believe the reason is that STPN shallow has fewer parameters and thus is easier to optimise in the training process. In contrast, the Swin-L variant is much more complex and thus needs more adjustments in the intermediate transformer layers. By default, we use STPN shallow.", + "bbox": [ + 496, + 547, + 890, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Support Frame Sampling. We use two hyper-parameters $S$ and $K$ to control the sampling of support frames. Table 5 (a) and Table 5 (b) show the effects of choosing different values of $S$ and $K$ , respectively. Firstly, we fix the value of $K$ as 7, so 15 surrounding support frames are used in total. Then, we vary the temporal stride $S$ from 1 to 32. As shown in Table 5 (a), $S = 8$ achieves the best performance. Afterwards, with $S$ fixed as 8, we verify the effect of different $K$ with respect to both performance (AP) and speed (FPS). Specifically, when $K = 7$ , STPN achieves a good speed-accuracy trade-off with $60.6\\%$ AP and 22.3", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13593", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ada8eb29dcbe02f5dbc9a30def17471d47b3707032bad0dde499e02bbcdd2f7d.jpg", + "image_caption": [ + "Figure 6. Bounding boxes and Grad-CAM visualisations of Faster-RCNN and STPN. Each column shows a video frame from the ImageNet VID validation set." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 450, + 295 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66c89b35a9f5accb099d5241eb3615b1061a0dc6ba4ddb9912809dc388f7732a.jpg", + "image_caption": [ + "Figure 7. Instance mask visualisations of MinVIS [20] and STPN. Each column shows a video frame from the YouTube VIS dataset. MinVIS fails to generate accurate masks in low-quality frames." + ], + "image_footnote": [], + "bbox": [ + 99, + 364, + 450, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FPS. Therefore, we choose $S = 8$ and $K = 7$ as default.", + "bbox": [ + 76, + 564, + 441, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of dynamic video prompts. To explore the effect of the number of dynamic video prompts $N_P$ , we show the performance and speed results by varying this number from 3 to 13 in Table 5 (c). The best speed-accuracy trade-off is obtained when $N_P$ is 7. In particular, once $N_P$ is greater than 7, AP is less affected by the change of $N_P$ . Thus, we use 7 as the default value of $N_P$ .", + "bbox": [ + 75, + 594, + 468, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 710, + 261, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Grad-CAM [42] is used to visualise the \"concentration\" areas of the FasterRCNN baseline and STPN in some challenging scenarios, for example, motion blur, occlusion, and both. As shown in Figure 6, FasterRCNN either fails to detect the blurry object or generates false positive detections, while STPN can produce accurate detections. In the first column, the squirrel is very blurry because of its fast movement. As a result, FasterRCNN cannot detect the squirrel, and thus the Grad-CAM does not show much attention to the region of the squirrel. On the contrary, STPN successfully detects the squirrel, and Grad-CAM", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9263a3183aed34777fc3a75a304c487996835dbbd2957f0f3045f510fadc61d6.jpg", + "image_caption": [ + "(a) Fast-speed objects", + "Figure 8. t-SNE visualisation of the object features produced by FasterRCNN and STPN on (a) fast speed and (b) medium speed objects in the ImageNet VID dataset." + ], + "image_footnote": [], + "bbox": [ + 521, + 88, + 694, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/173ffe4c3326581a32b8b833d1a268085b38a726b0cdb71cc6dc263eba87451d.jpg", + "image_caption": [ + "(b) Medium-speed objects" + ], + "image_footnote": [], + "bbox": [ + 697, + 89, + 870, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shows a good concentration map in the squirrel region.", + "bbox": [ + 500, + 256, + 861, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instance Mask Visualisation. Figure 7 shows the mask segmentation results of the state-of-the-art method MinVIS [20] and the STPN in deteriorated frames. For example, in the first and the second columns, MinVIS fails to generate instance masks for the surfboard and the fox because of motion blur but our STPN can generate accurate instance masks in these frames. The third column shows a case of occlusion. The left deer is occluded by the right one, so the MinVIS fails to generate a mask in the head area of the left dear. Besides, it cannot distinguish the two deer and generate one instance mask. In contrast, STPN works well for generating the mask of the occluded deer and distinguishing the two instances of deer.", + "bbox": [ + 498, + 286, + 892, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "t-SNE [48] is a statistical method for visualising high-dimensional data in 2D maps and is usually used to demonstrate the discriminative abilities of different models in classification tasks. We visualise the object features extracted by FasterRCNN and STPN using t-SNE. Figure 8 (a) and (b) show the results on the objects with the fast and medium moving speeds, respectively. We can see that STPN has better clustering results than the FasterRCNN baseline, demonstrating that STPN can improve the discriminative ability of features. Details of how to define different moving speeds and how to generate object features are illustrated in the supplementary material.", + "bbox": [ + 498, + 498, + 890, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 694, + 617, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To the best of our knowledge, we are the first to explore promoting techniques for robust video feature extraction. We present a neat and unified framework, named Sptio-Temporal Prompting Network (STPN). STPN simplifies the current pipeline for video understanding tasks. Moreover, STPN is easy to generalise to various video tasks. We conduct extensive experiments and report STPN's superior performance on three widely-used video benchmarks: ImageNet VID for video object detection, YouTube for video instance segmentation, and GOT-10K for visual object tracking. We hope our work can inform future research on robust video understanding.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13594", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4", + "[2] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolitional siamese networks for object tracking. In ECCV, 2016. 2", + "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 3", + "[5] Jianguo Chen, Kenli Li, Qingying Deng, Keqin Li, and S Yu Philip. Distributed deep learning model for intelligent video surveillance systems with edge computing. Industrial Informatics, 2019. 1", + "[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Hutchuan Lu. Transformer tracking. In CVPR, 2021. 1, 2, 6", + "[7] Yihong Chen, Yue Cao, Han Hu, and Liwei Wang. Memory enhanced global-local aggregation for video object detection. In CVPR, 2020. 1, 6", + "[8] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 2, 6", + "[9] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 5", + "[10] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Target transformed regression for accurate tracking. arXiv preprint arXiv:2104.00403, 2021. 1, 2, 6", + "[11] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 1, 2, 5, 6", + "[12] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 6", + "[13] Jiajun Deng, Yingwei Pan, Ting Yao, Wengang Zhou, Houqiang Li, and Tao Mei. Relation distillation networks for video object detection. In ICCV, 2019. 1, 2, 6", + "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 2", + "[15] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. NeurIPS, 2020. 3", + "[16] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2", + "[17] Lu He, Qianyu Zhou, Xiangtai Li, Li Niu, Guangliang Cheng, Xiao Li, Wenxuan Liu, Yunhai Tong, Lizhuang Ma, and Liqing Zhang. End-to-end video object detection with spatial-temporal transformers. In ACMMM, 2021. 2, 5, 6" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 6", + "[19] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, 2018. 1", + "[20] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6, 8", + "[21] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE transactions on pattern analysis and machine intelligence, 43(5):1562-1577, 2019. 2, 5", + "[22] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 2021. 1, 2, 6", + "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 2, 3, 4", + "[24] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 2, 3", + "[25] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 1", + "[26] Hamed Kiani Galoogahi, Ashton Fagg, and Simon Lucey. Learning background-aware correlation filters for visual tracking. In ICCV, 2017. 1", + "[27] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 3", + "[28] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2", + "[29] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2, 6", + "[30] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 2", + "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5", + "[32] Liangkai Liu, Sidi Lu, Ren Zhong, Baofu Wu, Yongtao Yao, Qingyang Zhang, and Weisong Shi. Computing systems for autonomous driving: State of the art and challenges. IEEE Internet of Things Journal, 8(8):6469-6486, 2020. 1", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3, 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Rashmika Nawaratne, Damminda Alahakoon, Daswin De Silva, and Xinghuo Yu. Spatiotemporal anomaly detection using deep learning for real-time video surveillance. Industrial Informatics, 16(1):393-402, 2019. 1", + "[35] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, 2022. 3", + "[36] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. In NeurIPS, 2022. 3", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2", + "[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3", + "[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 5", + "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5", + "[42] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 8", + "[43] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Mamba: Multi-level aggregation via memory bank for video object detection. In AAAI, 2021. 1, 2, 5, 6", + "[44] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Efficient one-stage video object detection by exploiting temporal consistency. In ECCV, 2022. 5, 6", + "[45] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Tdvit: Temporal dilated video transformer for dense video tasks. In ECCV, 2022. 2, 6", + "[46] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. In NeurIPS, 2021. 4", + "[47] Jack Valmadre, Luca Bertinetto, João Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In CVPR, 2017. 1", + "[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2, 3", + "[50] Paul Voigtlaender, Jonathon Luiten, Philip H. S. Torr, and Bastian Leibe. Siam R-CNN: visual tracking by re-detection. In CVPR, 2020. 6", + "[51] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 2", + "[52] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 6", + "[53] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation: A unifying approach. In CVPR, 2019. 2", + "[54] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 1", + "[55] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 1, 2, 6", + "[56] Fei Wu, Qingzhong Wang, Jiang Bian, Ning Ding, Feixiang Lu, Jun Cheng, Dejing Dou, and Haoyi Xiong. A survey on video action recognition in sports: Datasets, methods and applications. Multimedia, 2022. 1", + "[57] Haiping Wu, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Sequence level semantics aggregation for video object detection. In ICCV, 2019. 1, 2, 5, 6", + "[58] Haiping Wu, Bin Xiao, Noel C. F. Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In ICCV, 2021. 3, 5", + "[59] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 2", + "[60] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 2, 6", + "[61] Zhujun Xu, Emir Hrustic, and Damien Vivet. Centernet heatmap propagation for real-time video object detection. In ECCV, 2020. 6", + "[62] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 1, 2, 6", + "[63] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 5", + "[64] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1", + "[65] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Unified vision and language prompt learning. arXiv preprint arXiv:2210.07225, 2022. 3", + "[66] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In ICCV, 2017. 5" + ], + "bbox": [ + 501, + 90, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13596", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[67] Zheng Zhu, Qiang Wang, Bo Li, Wei Wu, Junjie Yan, and Weiming Hu. Distractor-aware siamese networks for visual object tracking. In ECCV, 2018. 2", + "bbox": [ + 78, + 90, + 470, + 135 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13597", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/6b38d58c-2dfe-4bda-8611-478cf7d86d34_model.json b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/6b38d58c-2dfe-4bda-8611-478cf7d86d34_model.json new file mode 100644 index 0000000000000000000000000000000000000000..47f5034c8c57c9b2ecf5b9b1892def5b1dea16f3 --- /dev/null +++ b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/6b38d58c-2dfe-4bda-8611-478cf7d86d34_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.131, + 0.859, + 0.154 + ], + "angle": 0, + "content": "Spatio-temporal Prompting Network for Robust Video Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.193, + 0.882, + 0.213 + ], + "angle": 0, + "content": "Guanxiong Sun\\(^{1,2}\\), Chi Wang\\(^{1}\\), Zhaoyu Zhang\\(^{1}\\), Jiankang Deng\\(^{2,3}\\), Stefanos Zafeiriou\\(^{3}\\), Yang Hua\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.223, + 0.81, + 0.242 + ], + "angle": 0, + "content": "\\(^{1}\\)Queen's University Belfast \\(^{2}\\)Huawei UKRD \\(^{3}\\)Imperial College London" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.242, + 0.838, + 0.261 + ], + "angle": 0, + "content": "{gsun02, cwang38, zhang55,Y.Hua} @qub.ac.uk, {j.deng16,s.zafeiriou} @imperial.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.306, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.474, + 0.718 + ], + "angle": 0, + "content": "Frame quality deterioration is one of the main challenges in the field of video understanding. To compensate for the information loss caused by deteriorated frames, recent approaches exploit transformer-based integration modules to obtain spatio-temporal information. However, these integration modules are heavy and complex. Furthermore, each integration module is specifically tailored for its target task, making it difficult to generalise to multiple tasks. In this paper, we present a neat and unified framework, called Spatio-Temporal Prompt Network (STPN). It can efficiently extract robust and accurate video features by dynamically adjusting the input features in the backbone network. Specifically, STPN predicts several video prompts containing spatio-temporal information of neighbour frames. Then, these video prompts are pretended to the patch embeddings of the current frame as the updated input for video feature extraction. Moreover, STPN is easy to generalise to various video tasks because it does not contain task-specific modules. Without bells and whistles, STPN achieves state-of-the-art performance on three widely-used datasets for different video understanding tasks, i.e., ImageNetVID for video object detection, YouTubeVIS for video instance segmentation, and GOT-10k for visual object tracking. Codes are available at https://github.com/guanxiongsun/STPN" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.753, + 0.21, + 0.768 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Video understanding is a fundamental research direction in the field of computer vision. It plays an important role in many real-world applications, such as autonomous driving [64, 32], video surveillance [34, 5], and sports analysis [56, 25]. However, a significant challenge in this field is the deterioration of video frames due to motion blur, occlusion, and deformation, which makes it difficult to extract relevant information." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.31, + 0.7, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.442, + 0.657, + 0.454 + ], + "angle": 0, + "content": "(a) Current pipeline" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.319, + 0.885, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.441, + 0.819, + 0.454 + ], + "angle": 0, + "content": "(b) STPN" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Figure 1. Comparisons between pipelines of (a) existing methods and (b) the proposed Spatio-temporal Prompting Network (STPN). Existing methods introduce complex and task-specific integration modules (□ □ □) after backbone networks (□). In contrast, STPN is a unified framework for multiple tasks. A lightweight dynamic video prompt (DVP) predictor (□) generates a set of DVPs to adjust input before backbone networks (□). Best viewed in colour." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To overcome this challenge, inspired by the great success of transformers in many computer vision tasks [49, 54, 27, 19, 4], researchers explore various transformer-based integration modules to alleviate the information loss on the deteriorated video frames. For example, in video object detection, transformer-based feature aggregation methods [13, 7, 57, 43] are investigated to enhance the feature of proposals [40] in the detection head. In video instance segmentation, deteriorated frames usually cause wrong instance associations in a sequence, so 3D mask decoders [55, 22, 8] are introduced to learn associations of instance masks in an end-to-end manner. For visual object tracking, traditional correlation filters [26, 47] meet their limitations in severely degraded frames. Therefore, transformer-based integration modules [11, 62, 6, 10] are proposed to better capture complicated correlations between the template region and the search region. To conclude, the summarised pipeline of recent methods is shown in Figure 1 (a). A backbone network extracts spatially-only features from the current frame and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13587" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "the support frames in the same video. Then, different integration modules integrate spatial-only features on multiple frames to obtain spatio-temporal features. Finally, the spatio-temporal features are used for the targeted video understanding task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.332 + ], + "angle": 0, + "content": "While the current pipeline achieves good performance, there are still two major limitations of transformer-based integration modules. Firstly, these integration modules are involved as an extra component after the backbone network, leading to increased complexity and additional computational costs. Secondly, each integration module is tailored specifically for the target task and thus cannot be generalised to multiple video tasks. Hence, we put forward the following question: can we remove the complex integration modules and directly obtain spatio-temporal information in backbone networks?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.468, + 0.62 + ], + "angle": 0, + "content": "To answer this question, inspired by recent prompting techniques [24, 23, 28], in this paper, we present a neat and unified framework, called Spatio-Temporal Prompting Network (STPN). Instead of using complex integration modules, STPN simplifies the current pipeline for video understanding by introducing spatio-temporal information into the backbone network, as shown in Figure 1 (b). Specifically, given a video frame, we propose a dynamic video prompt (DVP) predictor to generate several video prompts according to support frames. Then, the predicted DVPs are prepended to the patch embeddings of the current frame as the updated input. Finally, a vision transformer backbone network extracts video features using the updated input for future video understanding tasks. It is worth noting that the DVP predictor is a lightweight structure and introduces only a small number of extra parameters, e.g., 0.11M. Moreover, STPN can be easily adapted to different video understanding tasks, since it does not contain task-specific modules, and all modifications happen before the backbone network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.468, + 0.816 + ], + "angle": 0, + "content": "In summary, our key contributions are: (1) We present the Spatio-Temporal Prompting Network (STPN) that can extract robust video features on deteriorated video frames. STPN simplifies the current pipeline for video understanding and is easy to generalise to different video understanding tasks. (2) To the best of our knowledge, we are the first to explore promoting techniques for robust video feature extraction on the task of video object detection (VOD), video instance segmentation (VIS), and visual object tracking (VOT). (3) Without bells and whistles, STPN achieves state-of-the-art performance on three widely-used video understanding benchmarks, i.e., ImageNet-VID [41] for VOD, YouTube-VIS [63] for VIS, and GOT-10k [21] for VOT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.83, + 0.218, + 0.846 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this section, we first review the state-of-the-art (SOTA) methods for related video understanding tasks. Then, we introduce the visual prompting techniques." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.109, + 0.892, + 0.335 + ], + "angle": 0, + "content": "Video Object Detection (VOD). SOTA VOD methods use attention modules to aggregate features of support frames onto the current frame and thus enhance the feature quality of the current frame. For example, RDN [13] and SELSA [57] utilise relation modules to aggregate proposal features from local frames and global frames, respectively. These methods use sample memory architectures to store features on support frames, e.g., a sliding window or a memory queue, which limit the \"diversity\" of support features. Therefore, MAMBA [43] proposes a memory bank architecture that can partially update and sample support features and thus achieve a good speed-accuracy trade-off. More recent approaches TransVOD [17] and TDViT [45] introduce transformers into the field and further improve the performance of VOD." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Video Instance Segmentation (VIS). The goal of VIS is to simultaneously segment and track all object instances in the video sequence. There are two categories of approaches: Firstly, per-frame approaches conduct instance segmentation on each frame independently and then associate instance masks by post-processing. MaskTrackRCNN [63] adds a tracking prediction head on MaskRCNN [16] for instance association. More recently, some approaches [60, 20, 18] use transformers [49, 4] to reason the relationships between instances. Secondly, per-clip approaches [55, 59, 8, 22] directly predict 3D masks for each instance in a video. VisTR [55] proposes a transformer encoder-decoder structure to obtain spatio-temporal features, which are then used in an instance sequence matching module to get object tracks. IFC [22] uses memory tokens inside of the transformer encoder module to transfer temporal information between input frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Visual Object Tracking (VOT). Object tracking is one of the most fundamental research topics in computer vision and has a long history. Siamese-based methods [2, 30, 29, 53, 67] consider tracking as a template-matching problem by extracting correlation feature maps of the target and the search region via two-head siamese networks. Recently, transformers have been introduced to perform attention-based correspondence modelling and these methods [6, 51, 10, 62] achieve SOTA performance. Among these trackers, MixFormer [11] is a unified framework established solely on transformers for feature extraction, correspondence modelling, and result generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Prompting. The concept of prompting is originally proposed in natural language processing (NLP). Prompting refers to designing instructions and preponding them to the input so that the pre-trained language models [38, 14, 3] can \"understand\" the task. Recently, prompting techniques" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "13588" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.212 + ], + "angle": 0, + "content": "
MeaningSymbol
The i-th transformer layerLi
Collection of the output embeddings from LiEi
Number of embeddings in Eini
Dimension of embeddings in Eidi
Collection of dynamic video promptsP
Number of dynamic video promptsNP
" + }, + { + "type": "table_caption", + "bbox": [ + 0.195, + 0.22, + 0.351, + 0.234 + ], + "angle": 0, + "content": "Table 1. List of notations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.256, + 0.473, + 0.392 + ], + "angle": 0, + "content": "have been explored in many computer vision tasks. A major trend [35, 36, 24] is to transfer pre-trained vision-language models [15, 37, 39] to the task of video recognition, e.g., video classification and action recognition. Another popular research direction is parameter-efficient transfer learning via prompt tuning. For example, VPT [23] and UPT [65] tune a small number of parameters in the input space of a model and can achieve good performance comparable to the full fine-tuning scheme on a variety of recognition tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.405, + 0.216, + 0.421 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.553 + ], + "angle": 0, + "content": "In this section, we define notations necessary for subsequent demonstrations. \\( I_{t} \\) denotes the current frame and \\( t \\) is the current time step in the video. \\( I_{sup} \\) denotes the support frames used to predict video prompts and aid in the feature extraction process. Specifically, \\( I_{sup} = \\{I_{\\tau}\\}_{t - S(K / 2)}^{t + S(K / 2)} \\), where \\( K \\) denotes the number of support frames (assumed to be an even number) and \\( S \\) denotes the frame interval in the time dimension between support frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.675 + ], + "angle": 0, + "content": "Transformers encoders [27, 33, 58] are used to extract feature embeddings from video frames. A transformer encoder contains a patch embedding layer and \\(L\\) transformer layers. The input frame is first divided into \\(n\\) non-overlapping patches \\(\\{\\mathbf{x}_j\\in \\mathbb{R}^{3\\times h\\times w}|j\\in \\mathbb{N},1\\leq j\\leq n\\}\\) where \\(h\\) and \\(w\\) are the height and width of the patches. Each patch is then embedded into patch embeddings, denoted as \\(\\mathbf{E}\\in \\mathbb{R}^{n_0\\times d_0}\\). Other notations are summarised in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.687, + 0.212, + 0.705 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.188, + 0.726 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The overview of our framework is shown as in Figure 2 (a). There are two stages in the framework: the predicting stage and the prompting stage. The goal of the predicting stage is to generate a set of prompts that vary according to spatio-temporal information. We first select \\( K \\) support frames around \\( I_{t} \\) and then pass the support frames \\( I_{sup} \\) to the patch embedding layer and the transformer encoder to extract image embeddings, called support embeddings. The support embeddings are then passed to the dynamic video prompt (DVP) predictor network as the input. The DVP predictor outputs \\( N_{P} \\) dynamic video prompts, denoted as" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "\\(\\mathbf{P}\\), which contain the spatio-temporal information of the current video. Details of the DVP predictor network are illustrated in \\(\\S 4.2\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.893, + 0.289 + ], + "angle": 0, + "content": "The second stage is the prompting stage whose goal is to extract spatio-temporal features of the current frame \\( I_{t} \\) using \\( \\mathbf{P} \\). Specifically, \\( \\mathbf{P} \\) is prepended to the patch embeddings of the current frame \\( I_{t} \\). Then, the mixed embeddings are passed into the transformer encoder to produce the spatiotemporal embeddings. Finally, the spatio-temporal embeddings are used for different video understanding tasks, e.g., video object detection, video instance segmentation and visual object tracking. More details of the prompting stage are presented in §4.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.3, + 0.664, + 0.317 + ], + "angle": 0, + "content": "4.2. Predicting Stage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Given the current frame \\( I_{t} \\), we first sample \\( K \\) support frames around \\( I_{t} \\) as support frames. The support frames are then passed into the patch embedding layer and the transformer encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.397, + 0.892, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} _ {\\tau} ^ {i} = \\mathbf {L} ^ {i} \\left(\\mathbf {E} _ {\\tau} ^ {i - 1}\\right) \\quad \\mathbf {E} _ {\\tau} ^ {i} \\in \\mathbb {R} ^ {n _ {i} \\times d _ {i}}, i = 1, 2, \\dots , L. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.893, + 0.595 + ], + "angle": 0, + "content": "Therefore, the collection of support embeddings is denoted as: \\(\\mathbf{E}_{sup} = \\{\\mathbf{E}_{\\tau}^{\\mathbf{L}} \\in \\mathbb{R}^{n_L \\times d_L} | \\tau \\in \\mathcal{T}\\}\\), where \\(\\mathcal{T}\\) is the collection of sampled time steps around the current frame, \\(\\mathcal{T} = [t - S(K/2), \\dots, t - S, t + S, \\dots, t + S(K/2)]\\). For example, if \\(K = 6\\) and \\(S = 2\\), three support frames are sampled from the past of the current time step \\(t\\), and three are sampled from the future, with a time interval of 2 between each sampled frame in both directions. Given \\(\\mathbf{E}_{sup}\\), we design two predicting networks, i.e., transformer-based predictor and mixer-based predictor, to obtain dynamic video prompts \\(P \\in \\mathbb{R}^{N_P \\times d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Transformer-based Predictor. This design is inspired by the object decoder in DETR [4] where a set of randomly initialised object queries are used to retrieve object features on the image embeddings. As depicted in Figure 2 (b), we introduce a set of learnable prompt queries \\( Q \\in \\mathbb{R}^{N_P \\times d_L} \\) to gather spatio-temporal information from the support embeddings. Specifically, a multi-head attention (MHA) [49] module takes the support embeddings and the prompt queries as input, and produces the aggregated embeddings. The process is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.772, + 0.891, + 0.792 + ], + "angle": 0, + "content": "\\[\nK = \\operatorname {C o n c a t} \\left(\\mathbf {E} _ {\\sup }\\right) \\quad K \\in \\mathbb {R} ^ {K \\cdot n _ {L} \\times d _ {L}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.794, + 0.891, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\hat {Q} = \\operatorname {M H A} (\\ln (Q), \\ln (K)) + Q \\quad \\hat {Q} \\in \\mathbb {R} ^ {N _ {P} \\times d _ {L}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathrm{Concat}(\\cdot)\\) indicates the concatenation operation and LN denotes the LayerNorm layer. The output embeddings are then added back to the prompt queries and passed to the LayerNorm layer. The normalised output embeddings are further transformed by a simple feed-forward network" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13589" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.55, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.382, + 0.322, + 0.394 + ], + "angle": 0, + "content": "(a) Overview" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.096, + 0.847, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.374, + 0.691, + 0.399 + ], + "angle": 0, + "content": "(b) Transformer-based predictor" + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.375, + 0.824, + 0.398 + ], + "angle": 0, + "content": "(c) Mixer-based predictor" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.417, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Figure 2. (a) An overview of our approach. In the predicting stage, a set of dynamic visual prompts (DVPs) \\(\\mathbf{P}\\) is generated by the DVP predictor that takes support embeddings \\(\\mathbf{E}_{sup}\\) on support frames \\(I_{sup}\\) as the input. Then, in the prompting stage, predicted DVPs are prepended with the patch embeddings of the current frame to extract spatio-temporal embeddings via a transformer encoder which contains \\(L\\) transformer layers. Finally, different task heads take the spatio-temporal embeddings and output final results for various general video tasks, e.g., video object detection, video instance segmentation, and visual object tracking. (b) Details of the transformer-based predictor. (c) Details of the Mixer-based predictor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.556 + ], + "angle": 0, + "content": "(FFN) with a residual connection that makes the dynamic video prompts:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.56, + 0.47, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\operatorname {F F N} (\\operatorname {L N} (\\hat {Q})) + \\operatorname {L N} (\\hat {Q}) \\quad \\mathbf {P} \\in \\mathbb {R} ^ {N _ {P} \\times d}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.706 + ], + "angle": 0, + "content": "Mixer-based Predictor. Inspired by the MLP-Mixer [46] (shortened as Mixer), we design a Mixer-based predictor to transform support embedding into dynamic video prompts. The details of a Mixer-based predictor are shown in Figure 2 (c). Unlike the transformer-based predictor whose first step is to concatenate all support embeddings, we conduct average pooling in the temporal dimension to obtain the average support embeddings \\(\\overline{\\mathbf{E}}_{sup}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.71, + 0.469, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {E}}} _ {s u p} = \\operatorname {A v g P o o l} \\left(\\mathbf {E} _ {s u p}\\right) \\quad \\overline {{\\mathbf {E}}} _ {s u p} \\in \\mathbb {R} ^ {n _ {L} \\times d _ {L}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.839 + ], + "angle": 0, + "content": "where AvgPool denotes the average pooling across the time dimension. Then, the average support embedding is normalised by a LayerNorm layer and further transposed and passed into a feed-forward network (FFN) to obtain the hidden feature \\( h \\) with \\( N_P \\) channels. Finally, the hidden feature \\( h \\) is used to generate \\( N_P \\) dynamic video prompts with \\( d \\) channels. These two steps can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.858, + 0.469, + 0.877 + ], + "angle": 0, + "content": "\\[\nh = \\operatorname {F F N} \\left(\\left(\\operatorname {L N} \\left(\\overline {{\\mathbf {E}}} _ {s u p}\\right)\\right) ^ {T}\\right) \\quad h \\in \\mathbb {R} ^ {d _ {L} \\times N _ {P}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.879, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\operatorname {F F N} \\left(\\left(\\operatorname {L N} (h) ^ {T}\\right) \\quad \\mathbf {P} \\in \\mathbb {R} ^ {N _ {P} \\times d}, \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.525, + 0.89, + 0.557 + ], + "angle": 0, + "content": "where LN and \\((\\cdot)^T\\) denote the LayerNorm layer [1] and the transpose layer, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.563, + 0.667, + 0.58 + ], + "angle": 0, + "content": "4.3. Prompting Stage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.692 + ], + "angle": 0, + "content": "Given the predicted dynamic video prompts \\(\\mathbf{P}\\), we concatenate \\(\\mathbf{P}\\) together with the patch embeddings of the current frame \\(\\mathbf{E}_t\\) as the input of the transformer encoder. The details of this process are shown in the prompting stage (right part) of Figure 2 (a). More concretely, the process of applying \\(\\mathbf{P}\\) and extracting the spatio-temporal feature of the current frame \\(I_t\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.698, + 0.891, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {Z} _ {t} ^ {1}, \\mathbf {E} _ {t} ^ {1} \\right] = \\mathbf {L} ^ {1} (\\operatorname {C o n c a t} (\\mathbf {P}, \\mathbf {E} _ {t})) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.719, + 0.891, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {Z} _ {t} ^ {i}, \\mathbf {E} _ {t} ^ {i} \\right] = \\mathbf {L} ^ {i} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {Z} ^ {i - 1}, \\mathbf {E} _ {t} ^ {i - 1}\\right)\\right) \\quad i = 2, 3, \\dots , L \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.741, + 0.891, + 0.758 + ], + "angle": 0, + "content": "\\[\ny _ {t} = \\operatorname {H e a d} \\left(\\mathbf {E} _ {t} ^ {L}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.856 + ], + "angle": 0, + "content": "where \\(\\mathrm{Concat}(\\cdot)\\) indicates the concatenation operation and \\(\\mathbf{Z}_t^i\\in \\mathbb{R}^{N_P\\times d_i}\\) represents the extra embeddings produced by adding \\(\\mathbf{P}\\) in the input embeddings. The weights of the transformer encoder are shared in the predicting and prompting stages. Finally, the output embeddings \\(\\mathbf{E}_t^L\\) are passed into the head network for different video tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "STPN Shallow v.s. STPN Deep. Following the protocols in VPT [23], we design two STPN variants, i.e., STPN shal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13590" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.088, + 0.275, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.218, + 0.242, + 0.23 + ], + "angle": 0, + "content": "(a) SPTN shallow" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.451, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.219, + 0.409, + 0.23 + ], + "angle": 0, + "content": "(b) SPTN deep" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.333 + ], + "angle": 0, + "content": "Figure 3. A comparison between the (a) STPN shallow and the (b) STPN deep. The green rounded rectangles denote predicted DVPs. In (a), one set of DVPs is prepended to the patch embeddings before the first layer of the transformer encoder. In (b), \\( L \\) sets of DVPs are predicted and then prepended to the input embeddings of all \\( L \\) transformer layers in the transformer encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.508 + ], + "angle": 0, + "content": "low \\( v.s. \\) STPN deep, according to how many sets of dynamic video prompts (DVPs) are generated to adjust the transformer encoder. For example, in the shallow version, one set of DVPs is predicted and prepended only before the first transformer layer. In the deep version, \\( L \\) sets of DVPs are predicted via \\( L \\) separate FC layers, and each set of DVPs is prepended to the corresponding transformer layer. Formally, the DVPs of the \\( i \\)-th layer are denoted as \\( \\mathbf{P}^i \\in \\mathbb{R}^{N_P \\times d_i} \\) and the feature extraction process Equation (9) is changed to:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.517, + 0.469, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\left[ \\dots , \\mathbf {E} _ {t} ^ {i} \\right] = \\mathbf {L} ^ {i} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {P} ^ {i - 1}, \\mathbf {E} _ {t} ^ {i - 1}\\right)\\right) \\quad i = 2, 3, \\dots , L. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.469, + 0.575 + ], + "angle": 0, + "content": "The difference between the shallow and deep variants is shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.587, + 0.208, + 0.604 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.469, + 0.718 + ], + "angle": 0, + "content": "We first conduct experiments and compare SPTN with other state-of-the-art methods on three general video understanding tasks, i.e., video object detection (VOD), video instance segmentation (VIS) and visual object tracking (VOT). Then, in §5.3, we conduct ablation studies on how different design choices affect the overall performance, which helps to gain a deeper understanding of SPTN." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.175, + 0.743 + ], + "angle": 0, + "content": "5.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Datasets. For VOD, the ImageNetVID dataset [41] is used. It contains 3,862 training and 555 validation videos. Following previous approaches [43, 57, 66], we add overlapped 30 classes of the ImageNet DET dataset into the train set. Specifically, we sample 15 frames from each video in the VID train and at most 2,000 images per class from the DET train. For VIS, we train our models and report results on the YouTube VIS 2019 dataset [63]. The dataset contains 2,238 training, 302 validation, and 343 test videos, which covers 40 object categories. For VOT, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "choose the GOT10k [21] dataset, because it is a large-scale dataset (10,000 training videos and 180 test videos) with more than 1.5 million manually labelled bounding boxes, enabling stable training and evaluation of trackers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Evaluation Metrics. For video object detection and video instance segmentation, we report detailed results in the COCO [31] evaluation format. Specifically, we report the average precision (AP) metric which computes the average precision over ten IoU thresholds \\([0.5:0.05:0.95]\\) for all categories. Meanwhile, the COCO evaluation contains other important metrics, e.g., \\(\\mathrm{AP}_{50}\\) and \\(\\mathrm{AP}_{75}\\) that are calculated at IoU thresholds of 0.50 and 0.75, and \\(\\mathrm{AP}_S\\), \\(\\mathrm{AP}_M\\), \\(\\mathrm{AP}_L\\) that are calculated on different object scales, i.e., small, medium and large. The metrics for detection and segmentation are noted with box and mask, respectively. For visual object tracking, we report the widely used average overlap (AO) and the success rate (SR). AO denotes the average overlaps between all ground truth and estimated bounding boxes, while SR measures the percentage of successfully tracked frames where the overlaps exceed a threshold, e.g., \\(\\mathrm{SR}_{0.5}\\) and \\(\\mathrm{SR}_{0.75}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.547 + ], + "angle": 0, + "content": "Transformer Encoders. Our method is compatible with different transformer backbones. For demonstration, we conduct experiments on two well-known transformer backbones, i.e., Swin [33] and CvT [58]. Swin is used for video object detection and video instance segmentation. CvT is used for visual object tracking. The details of the backbones are illustrated in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.669 + ], + "angle": 0, + "content": "Training and Inference. All reported results are obtained with Python 3.7 and PyTorch 1.8.1 on Tesla V100 GPUs. We follow the protocols in TransVOD [17], Mask2Former [9], and MixFormer [11] to set the hyper-parameters, e.g., learning rate, data augmentations, optimiser, etc., for three tasks, respectively. More details about the hyper-parameter settings are described in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.681, + 0.8, + 0.696 + ], + "angle": 0, + "content": "5.2. Comparisons with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video Object Detection. Table 2 shows the results of STPN and state-of-the-art methods. Firstly, in the AP50 metric, using a simple FasterRCNN [40] base detector with the Swin-T [33] backbone, STPN achieves \\(85.0\\%\\) of AP50, which makes a \\(1.3\\%\\) improvement over the best competitor TransVOD [17]. Furthermore, when changing the FasterRCNN base detector to a multi-frame aggregation method SELSA [57], STPN increases further to \\(86.5\\%\\) of AP50. When using a more strict AP metric, STPN with FasterRCNN and SELSA achieve remarkable improvements of \\(+6.5 / 6.9\\%\\) over the best competitor EOVOD [44], respectively. Finally, when changing to stronger backbones, STPN with the Swin-B backbone achieves further improved" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13591" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.087, + 0.833, + 0.359 + ], + "angle": 0, + "content": "
ModelBase DetectorBackboneAPAP50AP75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
CHP [61]CenterNetR101-76.7----
EOVOD [44]FCOSR10154.179.859.510.528.360.1
RDN [13]FasterRCNNR10153.481.860.18.527.459.6
TransVOD [17]Deformable DETRR101-82.0----
MEGA [7]FasterRCNNR10153.282.959.29.129.459.1
TDViT [45]FasterRCNNSwinT50.979.955.79.126.957.2
SELSA* [57]FasterRCNNSwinT52.382.358.211.428.857.9
TransVOD [17]Deformable DETRSwinT-83.7----
STPNFasterRCNNSwinT60.685.068.812.133.766.8
STPNSELSASwinT60.986.568.613.334.867.1
MEGA [7]FasterRCNNResNeXt101-84.1----
MAMBA [43]FasterRCNNResNeXt101-86.7----
TransVOD [17]Deformable DETRSwinB-90.0----
STPNFasterRCNNSwinB63.686.171.612.133.870.3
STPNSELSASwinB65.290.673.514.137.771.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.369, + 0.859, + 0.384 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art methods on ImageNet VID for video object detection. * represents our re-implementation." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.407, + 0.467, + 0.585 + ], + "angle": 0, + "content": "
ModelBackboneAPAP50AP75FPS
VisTR [55]R10140.164.045.057.7
CrossVISR10136.657.339.735.6
MaskPropR10142.5-45.6-
SeqFormerR10149.071.155.764.6
IFC [22]R10144.669.249.589.4
IDOL [60]R10150.173.156.126.0
VITA [18]R10151.975.457.0-
M2F-VIS [8]SwinT51.575.056.5-
MinVIS [20]SwinT51.975.158.227.1
STPNMinVISSwinT54.878.761.026.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.595, + 0.468, + 0.623 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-art methods for video instance segmentation on the YouTube VIS 2019 dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.654, + 0.468, + 0.684 + ], + "angle": 0, + "content": "performance of \\(65.2\\%\\) AP. Compared to TransVOD, STPN with SELSA makes a \\(0.6\\%\\) improvement in AP50." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.702, + 0.468, + 0.838 + ], + "angle": 0, + "content": "Video Instance Segmentation. We compare STPN with state-of-the-art methods in Table 3. Using the M2F-VIS [8] as the baseline segmentation method, STPN improves performance by large margins with \\(2.9 / 3.6\\%\\) of AP/AP50, respectively. Compared with SOTA methods with similar backbone complexities, for example, Swin-T or ResNet-101, STPN achieves the best accuracy. Specifically, STPN improves the performance of the best competitor VITA [18] by \\(2.9 / 3.3\\%\\) of AP/AP50, respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Visual Object Tracking. Table 4 compares STPN with state-of-the-art methods. We use MixFormer [11] as our baseline. Specifically, when CVT-22k and CVT-1k are used" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.606 + ], + "angle": 0, + "content": "
ModelBackboneAOSR0.5SR0.75
SiamRPN++ [29]R10151.761.632.5
SiamR-CNN [50]R10164.972.859.7
PrDiMP [12]R10163.473.854.3
TrDiMP [52]R10167.177.758.3
TREG [10]R10166.877.857.2
TransT [6]R10167.176.860.9
STARK [62]R10168.878.164.1
MixFormer[11]CVT-22K70.780.067.8
MixFormer [11]CVT-1K71.279.965.8
STPNMixFormerCVT-22K71.881.669.0
STPNMixFormerCVT-1K72.781.967.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.616, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Table 4. Comparison with state-of-the-art tracking methods on the GOT-10k dataset." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.668, + 0.892, + 0.728 + ], + "angle": 0, + "content": "as the backbone, STPN improves the AO of MixFormer by \\(+1.1\\%\\) and \\(+1.5\\%\\), respectively. STPN achieves the best performance compared to other SOTA methods on the GOT-10K dataset." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.73, + 0.892, + 0.789 + ], + "angle": 0, + "content": "These experimental results demonstrate that STPN is easy to generalise to various video understanding tasks and achieves remarkable performance across all three different video tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.653, + 0.818 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.825, + 0.892, + 0.886 + ], + "angle": 0, + "content": "We investigate the effect of different designs of our approach on the overall performance. All experiments in §5.3 are conducted on the ImageNet VID dataset for video object detection." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13592" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.346, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.233, + 0.316, + 0.247 + ], + "angle": 0, + "content": "(a) STPN with Swin-T backbone" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.092, + 0.621, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.233, + 0.591, + 0.247 + ], + "angle": 0, + "content": "(b) STPN with Swin-B backbone" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.095, + 0.888, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.234, + 0.85, + 0.247 + ], + "angle": 0, + "content": "(c) Comparison with STOA" + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.266, + 0.757, + 0.281 + ], + "angle": 0, + "content": "Figure 4. Comparisons between the transformer-based and the Mixer-based DVP predictors." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.307, + 0.465, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.464, + 0.469, + 0.491 + ], + "angle": 0, + "content": "Figure 5. Comparisons of STPN deep and STPN shallow on transformer encoders with different scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.716 + ], + "angle": 0, + "content": "DVP Predictor Designs. We compare the effects of DVP predictor designs on different baselines. Figure 4 (a) shows the results on the Swin-T backbone. The FasterRCNN baseline achieves \\(80.7\\%\\) AP50, runs at 23.12 FPS, and has 44.90M parameters. Using the transformer-based predictor and the Mixer-based predictor, STPN achieves \\(85.0 / 84.5\\%\\) AP50 and runs at 22.27/22.41 FPS and has 45.01/44.96M parameters, respectively. These results show that the transformer-based predictor can achieve better accuracy, while the Mixer-based predictor has less complexity and faster speed. Experiments on the Swin-B backbone are shown in Figure 4 (b) and the experimental results indicate the same conclusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Efficiency of STPN. Compared with other state-of-the-art methods, STPN with both predictor designs actually achieves a significantly improved speed-accuracy trade-off, as shown in Figure 4 (c). These results demonstrate that the pipeline of STPN is very efficient and is not sensitive to different predictor designs. We choose the transformer-based predictor as our default setting since it achieves a slightly higher accuracy with a comparable fast speed." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "STPN Shallow v.s. STPN Deep. As shown in the Figure 5, we conduct experiments on Swin-T/S/B/L variants. The" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.304, + 0.89, + 0.45 + ], + "angle": 0, + "content": "
(a)S12481632
AP58.759.660.160.660.560.5
(b)K1357911
AP52.653.558.960.660.660.7
FPS22.422.422.322.321.720.5
(c)NP35791113
AP60.260.360.660.660.560.6
FPS22.922.722.321.621.019.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Table 5. Effect of hyper-parameters in STPN: the temporal stride \\( S \\), the number of support frames \\( K \\), and the number of dynamic video prompts \\( N_P \\). Different choices of hyper-parameters are listed in rows highlighted in grey." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.716 + ], + "angle": 0, + "content": "first finding is that both STPN shallow and STPN deep can continuously improve performance on all Swin variants, indicating that STPN has good compatibility with backbone scales. Another finding which contradicts our intuition is that surprisingly STPN shallow achieves better performance than STPN deep on three Swin variants. We believe the reason is that STPN shallow has fewer parameters and thus is easier to optimise in the training process. In contrast, the Swin-L variant is much more complex and thus needs more adjustments in the intermediate transformer layers. By default, we use STPN shallow." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Support Frame Sampling. We use two hyper-parameters \\( S \\) and \\( K \\) to control the sampling of support frames. Table 5 (a) and Table 5 (b) show the effects of choosing different values of \\( S \\) and \\( K \\), respectively. Firstly, we fix the value of \\( K \\) as 7, so 15 surrounding support frames are used in total. Then, we vary the temporal stride \\( S \\) from 1 to 32. As shown in Table 5 (a), \\( S = 8 \\) achieves the best performance. Afterwards, with \\( S \\) fixed as 8, we verify the effect of different \\( K \\) with respect to both performance (AP) and speed (FPS). Specifically, when \\( K = 7 \\), STPN achieves a good speed-accuracy trade-off with \\( 60.6\\% \\) AP and 22.3" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13593" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.451, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.311, + 0.47, + 0.354 + ], + "angle": 0, + "content": "Figure 6. Bounding boxes and Grad-CAM visualisations of Faster-RCNN and STPN. Each column shows a video frame from the ImageNet VID validation set." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.366, + 0.451, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.497, + 0.468, + 0.54 + ], + "angle": 0, + "content": "Figure 7. Instance mask visualisations of MinVIS [20] and STPN. Each column shows a video frame from the YouTube VIS dataset. MinVIS fails to generate accurate masks in low-quality frames." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.565, + 0.442, + 0.58 + ], + "angle": 0, + "content": "FPS. Therefore, we choose \\( S = 8 \\) and \\( K = 7 \\) as default." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Number of dynamic video prompts. To explore the effect of the number of dynamic video prompts \\( N_P \\), we show the performance and speed results by varying this number from 3 to 13 in Table 5 (c). The best speed-accuracy trade-off is obtained when \\( N_P \\) is 7. In particular, once \\( N_P \\) is greater than 7, AP is less affected by the change of \\( N_P \\). Thus, we use 7 as the default value of \\( N_P \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.262, + 0.727 + ], + "angle": 0, + "content": "5.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Grad-CAM [42] is used to visualise the \"concentration\" areas of the FasterRCNN baseline and STPN in some challenging scenarios, for example, motion blur, occlusion, and both. As shown in Figure 6, FasterRCNN either fails to detect the blurry object or generates false positive detections, while STPN can produce accurate detections. In the first column, the squirrel is very blurry because of its fast movement. As a result, FasterRCNN cannot detect the squirrel, and thus the Grad-CAM does not show much attention to the region of the squirrel. On the contrary, STPN successfully detects the squirrel, and Grad-CAM" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.089, + 0.695, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.164, + 0.656, + 0.173 + ], + "angle": 0, + "content": "(a) Fast-speed objects" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.09, + 0.871, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.164, + 0.845, + 0.173 + ], + "angle": 0, + "content": "(b) Medium-speed objects" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.189, + 0.892, + 0.232 + ], + "angle": 0, + "content": "Figure 8. t-SNE visualisation of the object features produced by FasterRCNN and STPN on (a) fast speed and (b) medium speed objects in the ImageNet VID dataset." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.257, + 0.862, + 0.273 + ], + "angle": 0, + "content": "shows a good concentration map in the squirrel region." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.287, + 0.893, + 0.484 + ], + "angle": 0, + "content": "Instance Mask Visualisation. Figure 7 shows the mask segmentation results of the state-of-the-art method MinVIS [20] and the STPN in deteriorated frames. For example, in the first and the second columns, MinVIS fails to generate instance masks for the surfboard and the fox because of motion blur but our STPN can generate accurate instance masks in these frames. The third column shows a case of occlusion. The left deer is occluded by the right one, so the MinVIS fails to generate a mask in the head area of the left dear. Besides, it cannot distinguish the two deer and generate one instance mask. In contrast, STPN works well for generating the mask of the occluded deer and distinguishing the two instances of deer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.499, + 0.892, + 0.68 + ], + "angle": 0, + "content": "t-SNE [48] is a statistical method for visualising high-dimensional data in 2D maps and is usually used to demonstrate the discriminative abilities of different models in classification tasks. We visualise the object features extracted by FasterRCNN and STPN using t-SNE. Figure 8 (a) and (b) show the results on the objects with the fast and medium moving speeds, respectively. We can see that STPN has better clustering results than the FasterRCNN baseline, demonstrating that STPN can improve the discriminative ability of features. Details of how to define different moving speeds and how to generate object features are illustrated in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.619, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To the best of our knowledge, we are the first to explore promoting techniques for robust video feature extraction. We present a neat and unified framework, named Sptio-Temporal Prompting Network (STPN). STPN simplifies the current pipeline for video understanding tasks. Moreover, STPN is easy to generalise to various video tasks. We conduct extensive experiments and report STPN's superior performance on three widely-used video benchmarks: ImageNet VID for video object detection, YouTube for video instance segmentation, and GOT-10K for visual object tracking. We hope our work can inform future research on robust video understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13594" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.155 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.199 + ], + "angle": 0, + "content": "[2] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolitional siamese networks for object tracking. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.47, + 0.255 + ], + "angle": 0, + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.469, + 0.367 + ], + "angle": 0, + "content": "[5] Jianguo Chen, Kenli Li, Qingying Deng, Keqin Li, and S Yu Philip. Distributed deep learning model for intelligent video surveillance systems with edge computing. Industrial Informatics, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.469, + 0.408 + ], + "angle": 0, + "content": "[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Hutchuan Lu. Transformer tracking. In CVPR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.469, + 0.45 + ], + "angle": 0, + "content": "[7] Yihong Chen, Yue Cao, Han Hu, and Liwei Wang. Memory enhanced global-local aggregation for video object detection. In CVPR, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.469, + 0.562 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[10] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Target transformed regression for accurate tracking. arXiv preprint arXiv:2104.00403, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[11] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[12] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[13] Jiajun Deng, Yingwei Pan, Ting Yao, Wengang Zhou, Houqiang Li, and Tao Mei. Relation distillation networks for video object detection. In ICCV, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[15] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[16] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[17] Lu He, Qianyu Zhou, Xiangtai Li, Li Niu, Guangliang Cheng, Xiao Li, Wenxuan Liu, Yunhai Tong, Lizhuang Ma, and Liqing Zhang. End-to-end video object detection with spatial-temporal transformers. In ACMMM, 2021. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[18] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[19] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[20] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[21] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE transactions on pattern analysis and machine intelligence, 43(5):1562-1577, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[22] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[24] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[25] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[26] Hamed Kiani Galoogahi, Ashton Fagg, and Simon Lucey. Learning background-aware correlation filters for visual tracking. In ICCV, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[27] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[28] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[29] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[30] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[32] Liangkai Liu, Sidi Lu, Ren Zhong, Baofu Wu, Yongtao Yao, Qingyang Zhang, and Weisong Shi. Computing systems for autonomous driving: State of the art and challenges. IEEE Internet of Things Journal, 8(8):6469-6486, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13595" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[34] Rashmika Nawaratne, Damminda Alahakoon, Daswin De Silva, and Xinghuo Yu. Spatiotemporal anomaly detection using deep learning for real-time video surveillance. Industrial Informatics, 16(1):393-402, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[35] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.47, + 0.263 + ], + "angle": 0, + "content": "[36] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.47, + 0.335 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.337, + 0.47, + 0.38 + ], + "angle": 0, + "content": "[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.382, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.425, + 0.47, + 0.467 + ], + "angle": 0, + "content": "[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.539 + ], + "angle": 0, + "content": "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.469, + 0.596 + ], + "angle": 0, + "content": "[42] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[43] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Mamba: Multi-level aggregation via memory bank for video object detection. In AAAI, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[44] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Efficient one-stage video object detection by exploiting temporal consistency. In ECCV, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[45] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Tdvit: Temporal dilated video transformer for dense video tasks. In ECCV, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.798 + ], + "angle": 0, + "content": "[46] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. In NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[47] Jack Valmadre, Luca Bertinetto, João Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[50] Paul Voigtlaender, Jonathon Luiten, Philip H. S. Torr, and Bastian Leibe. Siam R-CNN: visual tracking by re-detection. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[51] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.237, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[52] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[53] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation: A unifying approach. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[54] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.353, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[55] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[56] Fei Wu, Qingzhong Wang, Jiang Bian, Ning Ding, Feixiang Lu, Jun Cheng, Dejing Dou, and Haoyi Xiong. A survey on video action recognition in sports: Datasets, methods and applications. Multimedia, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.468, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[57] Haiping Wu, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Sequence level semantics aggregation for video object detection. In ICCV, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[58] Haiping Wu, Bin Xiao, Noel C. F. Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In ICCV, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.555, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[59] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[60] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[61] Zhujun Xu, Emir Hrustic, and Damien Vivet. Centernet heatmap propagation for real-time video object detection. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[62] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[63] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[64] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[65] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Unified vision and language prompt learning. arXiv preprint arXiv:2210.07225, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[66] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In ICCV, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13596" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.136 + ], + "angle": 0, + "content": "[67] Zheng Zhu, Qiang Wang, Bo Li, Wei Wu, Junjie Yan, and Weiming Hu. Distractor-aware siamese networks for visual object tracking. In ECCV, 2018. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13597" + } + ] +] \ No newline at end of file diff --git a/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/6b38d58c-2dfe-4bda-8611-478cf7d86d34_origin.pdf b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/6b38d58c-2dfe-4bda-8611-478cf7d86d34_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..475cb68c8b914e491b6f92c4371a8e525de976e7 --- /dev/null +++ b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/6b38d58c-2dfe-4bda-8611-478cf7d86d34_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2703fbf7d5d12720714fb91a4c0a2897624c11aeec888cf10f34ea324cfc067 +size 8624662 diff --git a/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/full.md b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2953fbd5ebe987c8d6e7710d76512e17aeb51eae --- /dev/null +++ b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/full.md @@ -0,0 +1,333 @@ +# Spatio-temporal Prompting Network for Robust Video Feature Extraction + +Guanxiong Sun $^{1,2}$ , Chi Wang $^{1}$ , Zhaoyu Zhang $^{1}$ , Jiankang Deng $^{2,3}$ , Stefanos Zafeiriou $^{3}$ , Yang Hua $^{1}$ + +$^{1}$ Queen's University Belfast $^{2}$ Huawei UKRD $^{3}$ Imperial College London + +{gsun02, cwang38, zhang55,Y.Hua} @qub.ac.uk, {j.deng16,s.zafeiriou} @imperial.ac.uk + +# Abstract + +Frame quality deterioration is one of the main challenges in the field of video understanding. To compensate for the information loss caused by deteriorated frames, recent approaches exploit transformer-based integration modules to obtain spatio-temporal information. However, these integration modules are heavy and complex. Furthermore, each integration module is specifically tailored for its target task, making it difficult to generalise to multiple tasks. In this paper, we present a neat and unified framework, called Spatio-Temporal Prompt Network (STPN). It can efficiently extract robust and accurate video features by dynamically adjusting the input features in the backbone network. Specifically, STPN predicts several video prompts containing spatio-temporal information of neighbour frames. Then, these video prompts are pretended to the patch embeddings of the current frame as the updated input for video feature extraction. Moreover, STPN is easy to generalise to various video tasks because it does not contain task-specific modules. Without bells and whistles, STPN achieves state-of-the-art performance on three widely-used datasets for different video understanding tasks, i.e., ImageNetVID for video object detection, YouTubeVIS for video instance segmentation, and GOT-10k for visual object tracking. Codes are available at https://github.com/guanxiongsun/STPN + +# 1. Introduction + +Video understanding is a fundamental research direction in the field of computer vision. It plays an important role in many real-world applications, such as autonomous driving [64, 32], video surveillance [34, 5], and sports analysis [56, 25]. However, a significant challenge in this field is the deterioration of video frames due to motion blur, occlusion, and deformation, which makes it difficult to extract relevant information. + +![](images/7c51f56c56ea88b261e25340276a14265d27c959fa6d11c2f69d511fe7150034.jpg) +(a) Current pipeline + +![](images/a951f0b22ba2fe2e8fa03193987df5e5509e7638d5ad9b1bdeecef2a83eaf4db.jpg) +(b) STPN +Figure 1. Comparisons between pipelines of (a) existing methods and (b) the proposed Spatio-temporal Prompting Network (STPN). Existing methods introduce complex and task-specific integration modules (□ □ □) after backbone networks (□). In contrast, STPN is a unified framework for multiple tasks. A lightweight dynamic video prompt (DVP) predictor (□) generates a set of DVPs to adjust input before backbone networks (□). Best viewed in colour. + +To overcome this challenge, inspired by the great success of transformers in many computer vision tasks [49, 54, 27, 19, 4], researchers explore various transformer-based integration modules to alleviate the information loss on the deteriorated video frames. For example, in video object detection, transformer-based feature aggregation methods [13, 7, 57, 43] are investigated to enhance the feature of proposals [40] in the detection head. In video instance segmentation, deteriorated frames usually cause wrong instance associations in a sequence, so 3D mask decoders [55, 22, 8] are introduced to learn associations of instance masks in an end-to-end manner. For visual object tracking, traditional correlation filters [26, 47] meet their limitations in severely degraded frames. Therefore, transformer-based integration modules [11, 62, 6, 10] are proposed to better capture complicated correlations between the template region and the search region. To conclude, the summarised pipeline of recent methods is shown in Figure 1 (a). A backbone network extracts spatially-only features from the current frame and + +the support frames in the same video. Then, different integration modules integrate spatial-only features on multiple frames to obtain spatio-temporal features. Finally, the spatio-temporal features are used for the targeted video understanding task. + +While the current pipeline achieves good performance, there are still two major limitations of transformer-based integration modules. Firstly, these integration modules are involved as an extra component after the backbone network, leading to increased complexity and additional computational costs. Secondly, each integration module is tailored specifically for the target task and thus cannot be generalised to multiple video tasks. Hence, we put forward the following question: can we remove the complex integration modules and directly obtain spatio-temporal information in backbone networks? + +To answer this question, inspired by recent prompting techniques [24, 23, 28], in this paper, we present a neat and unified framework, called Spatio-Temporal Prompting Network (STPN). Instead of using complex integration modules, STPN simplifies the current pipeline for video understanding by introducing spatio-temporal information into the backbone network, as shown in Figure 1 (b). Specifically, given a video frame, we propose a dynamic video prompt (DVP) predictor to generate several video prompts according to support frames. Then, the predicted DVPs are prepended to the patch embeddings of the current frame as the updated input. Finally, a vision transformer backbone network extracts video features using the updated input for future video understanding tasks. It is worth noting that the DVP predictor is a lightweight structure and introduces only a small number of extra parameters, e.g., 0.11M. Moreover, STPN can be easily adapted to different video understanding tasks, since it does not contain task-specific modules, and all modifications happen before the backbone network. + +In summary, our key contributions are: (1) We present the Spatio-Temporal Prompting Network (STPN) that can extract robust video features on deteriorated video frames. STPN simplifies the current pipeline for video understanding and is easy to generalise to different video understanding tasks. (2) To the best of our knowledge, we are the first to explore promoting techniques for robust video feature extraction on the task of video object detection (VOD), video instance segmentation (VIS), and visual object tracking (VOT). (3) Without bells and whistles, STPN achieves state-of-the-art performance on three widely-used video understanding benchmarks, i.e., ImageNet-VID [41] for VOD, YouTube-VIS [63] for VIS, and GOT-10k [21] for VOT. + +# 2. Related Work + +In this section, we first review the state-of-the-art (SOTA) methods for related video understanding tasks. Then, we introduce the visual prompting techniques. + +Video Object Detection (VOD). SOTA VOD methods use attention modules to aggregate features of support frames onto the current frame and thus enhance the feature quality of the current frame. For example, RDN [13] and SELSA [57] utilise relation modules to aggregate proposal features from local frames and global frames, respectively. These methods use sample memory architectures to store features on support frames, e.g., a sliding window or a memory queue, which limit the "diversity" of support features. Therefore, MAMBA [43] proposes a memory bank architecture that can partially update and sample support features and thus achieve a good speed-accuracy trade-off. More recent approaches TransVOD [17] and TDViT [45] introduce transformers into the field and further improve the performance of VOD. + +Video Instance Segmentation (VIS). The goal of VIS is to simultaneously segment and track all object instances in the video sequence. There are two categories of approaches: Firstly, per-frame approaches conduct instance segmentation on each frame independently and then associate instance masks by post-processing. MaskTrackRCNN [63] adds a tracking prediction head on MaskRCNN [16] for instance association. More recently, some approaches [60, 20, 18] use transformers [49, 4] to reason the relationships between instances. Secondly, per-clip approaches [55, 59, 8, 22] directly predict 3D masks for each instance in a video. VisTR [55] proposes a transformer encoder-decoder structure to obtain spatio-temporal features, which are then used in an instance sequence matching module to get object tracks. IFC [22] uses memory tokens inside of the transformer encoder module to transfer temporal information between input frames. + +Visual Object Tracking (VOT). Object tracking is one of the most fundamental research topics in computer vision and has a long history. Siamese-based methods [2, 30, 29, 53, 67] consider tracking as a template-matching problem by extracting correlation feature maps of the target and the search region via two-head siamese networks. Recently, transformers have been introduced to perform attention-based correspondence modelling and these methods [6, 51, 10, 62] achieve SOTA performance. Among these trackers, MixFormer [11] is a unified framework established solely on transformers for feature extraction, correspondence modelling, and result generation. + +Prompting. The concept of prompting is originally proposed in natural language processing (NLP). Prompting refers to designing instructions and preponding them to the input so that the pre-trained language models [38, 14, 3] can "understand" the task. Recently, prompting techniques + +
MeaningSymbol
The i-th transformer layerLi
Collection of the output embeddings from LiEi
Number of embeddings in Eini
Dimension of embeddings in Eidi
Collection of dynamic video promptsP
Number of dynamic video promptsNP
+ +Table 1. List of notations. + +have been explored in many computer vision tasks. A major trend [35, 36, 24] is to transfer pre-trained vision-language models [15, 37, 39] to the task of video recognition, e.g., video classification and action recognition. Another popular research direction is parameter-efficient transfer learning via prompt tuning. For example, VPT [23] and UPT [65] tune a small number of parameters in the input space of a model and can achieve good performance comparable to the full fine-tuning scheme on a variety of recognition tasks. + +# 3. Preliminaries + +In this section, we define notations necessary for subsequent demonstrations. $I_{t}$ denotes the current frame and $t$ is the current time step in the video. $I_{sup}$ denotes the support frames used to predict video prompts and aid in the feature extraction process. Specifically, $I_{sup} = \{I_{\tau}\}_{t - S(K / 2)}^{t + S(K / 2)}$ , where $K$ denotes the number of support frames (assumed to be an even number) and $S$ denotes the frame interval in the time dimension between support frames. + +Transformers encoders [27, 33, 58] are used to extract feature embeddings from video frames. A transformer encoder contains a patch embedding layer and $L$ transformer layers. The input frame is first divided into $n$ non-overlapping patches $\{\mathbf{x}_j\in \mathbb{R}^{3\times h\times w}|j\in \mathbb{N},1\leq j\leq n\}$ where $h$ and $w$ are the height and width of the patches. Each patch is then embedded into patch embeddings, denoted as $\mathbf{E}\in \mathbb{R}^{n_0\times d_0}$ . Other notations are summarised in Table 1. + +# 4. Methodology + +# 4.1. Overview + +The overview of our framework is shown as in Figure 2 (a). There are two stages in the framework: the predicting stage and the prompting stage. The goal of the predicting stage is to generate a set of prompts that vary according to spatio-temporal information. We first select $K$ support frames around $I_{t}$ and then pass the support frames $I_{sup}$ to the patch embedding layer and the transformer encoder to extract image embeddings, called support embeddings. The support embeddings are then passed to the dynamic video prompt (DVP) predictor network as the input. The DVP predictor outputs $N_{P}$ dynamic video prompts, denoted as + +$\mathbf{P}$ , which contain the spatio-temporal information of the current video. Details of the DVP predictor network are illustrated in $\S 4.2$ . + +The second stage is the prompting stage whose goal is to extract spatio-temporal features of the current frame $I_{t}$ using $\mathbf{P}$ . Specifically, $\mathbf{P}$ is prepended to the patch embeddings of the current frame $I_{t}$ . Then, the mixed embeddings are passed into the transformer encoder to produce the spatiotemporal embeddings. Finally, the spatio-temporal embeddings are used for different video understanding tasks, e.g., video object detection, video instance segmentation and visual object tracking. More details of the prompting stage are presented in §4.3. + +# 4.2. Predicting Stage + +Given the current frame $I_{t}$ , we first sample $K$ support frames around $I_{t}$ as support frames. The support frames are then passed into the patch embedding layer and the transformer encoder: + +$$ +\mathbf {E} _ {\tau} ^ {i} = \mathbf {L} ^ {i} \left(\mathbf {E} _ {\tau} ^ {i - 1}\right) \quad \mathbf {E} _ {\tau} ^ {i} \in \mathbb {R} ^ {n _ {i} \times d _ {i}}, i = 1, 2, \dots , L. \tag {1} +$$ + +Therefore, the collection of support embeddings is denoted as: $\mathbf{E}_{sup} = \{\mathbf{E}_{\tau}^{\mathbf{L}} \in \mathbb{R}^{n_L \times d_L} | \tau \in \mathcal{T}\}$ , where $\mathcal{T}$ is the collection of sampled time steps around the current frame, $\mathcal{T} = [t - S(K/2), \dots, t - S, t + S, \dots, t + S(K/2)]$ . For example, if $K = 6$ and $S = 2$ , three support frames are sampled from the past of the current time step $t$ , and three are sampled from the future, with a time interval of 2 between each sampled frame in both directions. Given $\mathbf{E}_{sup}$ , we design two predicting networks, i.e., transformer-based predictor and mixer-based predictor, to obtain dynamic video prompts $P \in \mathbb{R}^{N_P \times d}$ . + +Transformer-based Predictor. This design is inspired by the object decoder in DETR [4] where a set of randomly initialised object queries are used to retrieve object features on the image embeddings. As depicted in Figure 2 (b), we introduce a set of learnable prompt queries $Q \in \mathbb{R}^{N_P \times d_L}$ to gather spatio-temporal information from the support embeddings. Specifically, a multi-head attention (MHA) [49] module takes the support embeddings and the prompt queries as input, and produces the aggregated embeddings. The process is formulated as follows: + +$$ +K = \operatorname {C o n c a t} \left(\mathbf {E} _ {\sup }\right) \quad K \in \mathbb {R} ^ {K \cdot n _ {L} \times d _ {L}}, \tag {2} +$$ + +$$ +\hat {Q} = \operatorname {M H A} (\ln (Q), \ln (K)) + Q \quad \hat {Q} \in \mathbb {R} ^ {N _ {P} \times d _ {L}}, \tag {3} +$$ + +where $\mathrm{Concat}(\cdot)$ indicates the concatenation operation and LN denotes the LayerNorm layer. The output embeddings are then added back to the prompt queries and passed to the LayerNorm layer. The normalised output embeddings are further transformed by a simple feed-forward network + +![](images/ead93793a658ca120a22f9feca0ba40bbdeb68013e09e969ae7b4cc6824c71c0.jpg) +(a) Overview + +![](images/f240f4df6c29134140ab3c4a679019ffbffa4a8bb8b150fd65afaefaaa835021.jpg) +(b) Transformer-based predictor +(c) Mixer-based predictor +Figure 2. (a) An overview of our approach. In the predicting stage, a set of dynamic visual prompts (DVPs) $\mathbf{P}$ is generated by the DVP predictor that takes support embeddings $\mathbf{E}_{sup}$ on support frames $I_{sup}$ as the input. Then, in the prompting stage, predicted DVPs are prepended with the patch embeddings of the current frame to extract spatio-temporal embeddings via a transformer encoder which contains $L$ transformer layers. Finally, different task heads take the spatio-temporal embeddings and output final results for various general video tasks, e.g., video object detection, video instance segmentation, and visual object tracking. (b) Details of the transformer-based predictor. (c) Details of the Mixer-based predictor. + +(FFN) with a residual connection that makes the dynamic video prompts: + +$$ +\mathbf {P} = \operatorname {F F N} (\operatorname {L N} (\hat {Q})) + \operatorname {L N} (\hat {Q}) \quad \mathbf {P} \in \mathbb {R} ^ {N _ {P} \times d}. \tag {4} +$$ + +Mixer-based Predictor. Inspired by the MLP-Mixer [46] (shortened as Mixer), we design a Mixer-based predictor to transform support embedding into dynamic video prompts. The details of a Mixer-based predictor are shown in Figure 2 (c). Unlike the transformer-based predictor whose first step is to concatenate all support embeddings, we conduct average pooling in the temporal dimension to obtain the average support embeddings $\overline{\mathbf{E}}_{sup}$ as follows: + +$$ +\overline {{\mathbf {E}}} _ {s u p} = \operatorname {A v g P o o l} \left(\mathbf {E} _ {s u p}\right) \quad \overline {{\mathbf {E}}} _ {s u p} \in \mathbb {R} ^ {n _ {L} \times d _ {L}}, \tag {5} +$$ + +where AvgPool denotes the average pooling across the time dimension. Then, the average support embedding is normalised by a LayerNorm layer and further transposed and passed into a feed-forward network (FFN) to obtain the hidden feature $h$ with $N_P$ channels. Finally, the hidden feature $h$ is used to generate $N_P$ dynamic video prompts with $d$ channels. These two steps can be formulated as follows: + +$$ +h = \operatorname {F F N} \left(\left(\operatorname {L N} \left(\overline {{\mathbf {E}}} _ {s u p}\right)\right) ^ {T}\right) \quad h \in \mathbb {R} ^ {d _ {L} \times N _ {P}}, \tag {6} +$$ + +$$ +\mathbf {P} = \operatorname {F F N} \left(\left(\operatorname {L N} (h) ^ {T}\right) \quad \mathbf {P} \in \mathbb {R} ^ {N _ {P} \times d}, \right. \tag {7} +$$ + +where LN and $(\cdot)^T$ denote the LayerNorm layer [1] and the transpose layer, respectively. + +# 4.3. Prompting Stage + +Given the predicted dynamic video prompts $\mathbf{P}$ , we concatenate $\mathbf{P}$ together with the patch embeddings of the current frame $\mathbf{E}_t$ as the input of the transformer encoder. The details of this process are shown in the prompting stage (right part) of Figure 2 (a). More concretely, the process of applying $\mathbf{P}$ and extracting the spatio-temporal feature of the current frame $I_t$ can be formulated as: + +$$ +\left[ \mathbf {Z} _ {t} ^ {1}, \mathbf {E} _ {t} ^ {1} \right] = \mathbf {L} ^ {1} (\operatorname {C o n c a t} (\mathbf {P}, \mathbf {E} _ {t})) \tag {8} +$$ + +$$ +\left[ \mathbf {Z} _ {t} ^ {i}, \mathbf {E} _ {t} ^ {i} \right] = \mathbf {L} ^ {i} \left(\operatorname {C o n c a t} \left(\mathbf {Z} ^ {i - 1}, \mathbf {E} _ {t} ^ {i - 1}\right)\right) \quad i = 2, 3, \dots , L \tag {9} +$$ + +$$ +y _ {t} = \operatorname {H e a d} \left(\mathbf {E} _ {t} ^ {L}\right), \tag {10} +$$ + +where $\mathrm{Concat}(\cdot)$ indicates the concatenation operation and $\mathbf{Z}_t^i\in \mathbb{R}^{N_P\times d_i}$ represents the extra embeddings produced by adding $\mathbf{P}$ in the input embeddings. The weights of the transformer encoder are shared in the predicting and prompting stages. Finally, the output embeddings $\mathbf{E}_t^L$ are passed into the head network for different video tasks. + +STPN Shallow v.s. STPN Deep. Following the protocols in VPT [23], we design two STPN variants, i.e., STPN shal + +![](images/cdbc7086953477a806bd7a791abc73e5ba842f6f738957c74fbb450ea128d27a.jpg) +(a) SPTN shallow +Figure 3. A comparison between the (a) STPN shallow and the (b) STPN deep. The green rounded rectangles denote predicted DVPs. In (a), one set of DVPs is prepended to the patch embeddings before the first layer of the transformer encoder. In (b), $L$ sets of DVPs are predicted and then prepended to the input embeddings of all $L$ transformer layers in the transformer encoder. + +![](images/4077239bead0aa142b0b0ca06cc2ad4721a0f54e4cd63a2508fd3bf891ee8615.jpg) +(b) SPTN deep + +low $v.s.$ STPN deep, according to how many sets of dynamic video prompts (DVPs) are generated to adjust the transformer encoder. For example, in the shallow version, one set of DVPs is predicted and prepended only before the first transformer layer. In the deep version, $L$ sets of DVPs are predicted via $L$ separate FC layers, and each set of DVPs is prepended to the corresponding transformer layer. Formally, the DVPs of the $i$ -th layer are denoted as $\mathbf{P}^i \in \mathbb{R}^{N_P \times d_i}$ and the feature extraction process Equation (9) is changed to: + +$$ +\left[ \dots , \mathbf {E} _ {t} ^ {i} \right] = \mathbf {L} ^ {i} \left(\operatorname {C o n c a t} \left(\mathbf {P} ^ {i - 1}, \mathbf {E} _ {t} ^ {i - 1}\right)\right) \quad i = 2, 3, \dots , L. \tag {11} +$$ + +The difference between the shallow and deep variants is shown in Figure 3. + +# 5. Experiments + +We first conduct experiments and compare SPTN with other state-of-the-art methods on three general video understanding tasks, i.e., video object detection (VOD), video instance segmentation (VIS) and visual object tracking (VOT). Then, in §5.3, we conduct ablation studies on how different design choices affect the overall performance, which helps to gain a deeper understanding of SPTN. + +# 5.1. Settings + +Datasets. For VOD, the ImageNetVID dataset [41] is used. It contains 3,862 training and 555 validation videos. Following previous approaches [43, 57, 66], we add overlapped 30 classes of the ImageNet DET dataset into the train set. Specifically, we sample 15 frames from each video in the VID train and at most 2,000 images per class from the DET train. For VIS, we train our models and report results on the YouTube VIS 2019 dataset [63]. The dataset contains 2,238 training, 302 validation, and 343 test videos, which covers 40 object categories. For VOT, we + +choose the GOT10k [21] dataset, because it is a large-scale dataset (10,000 training videos and 180 test videos) with more than 1.5 million manually labelled bounding boxes, enabling stable training and evaluation of trackers. + +Evaluation Metrics. For video object detection and video instance segmentation, we report detailed results in the COCO [31] evaluation format. Specifically, we report the average precision (AP) metric which computes the average precision over ten IoU thresholds $[0.5:0.05:0.95]$ for all categories. Meanwhile, the COCO evaluation contains other important metrics, e.g., $\mathrm{AP}_{50}$ and $\mathrm{AP}_{75}$ that are calculated at IoU thresholds of 0.50 and 0.75, and $\mathrm{AP}_S$ , $\mathrm{AP}_M$ , $\mathrm{AP}_L$ that are calculated on different object scales, i.e., small, medium and large. The metrics for detection and segmentation are noted with box and mask, respectively. For visual object tracking, we report the widely used average overlap (AO) and the success rate (SR). AO denotes the average overlaps between all ground truth and estimated bounding boxes, while SR measures the percentage of successfully tracked frames where the overlaps exceed a threshold, e.g., $\mathrm{SR}_{0.5}$ and $\mathrm{SR}_{0.75}$ . + +Transformer Encoders. Our method is compatible with different transformer backbones. For demonstration, we conduct experiments on two well-known transformer backbones, i.e., Swin [33] and CvT [58]. Swin is used for video object detection and video instance segmentation. CvT is used for visual object tracking. The details of the backbones are illustrated in the supplementary material. + +Training and Inference. All reported results are obtained with Python 3.7 and PyTorch 1.8.1 on Tesla V100 GPUs. We follow the protocols in TransVOD [17], Mask2Former [9], and MixFormer [11] to set the hyper-parameters, e.g., learning rate, data augmentations, optimiser, etc., for three tasks, respectively. More details about the hyper-parameter settings are described in the supplementary material. + +# 5.2. Comparisons with SOTA Methods + +Video Object Detection. Table 2 shows the results of STPN and state-of-the-art methods. Firstly, in the AP50 metric, using a simple FasterRCNN [40] base detector with the Swin-T [33] backbone, STPN achieves $85.0\%$ of AP50, which makes a $1.3\%$ improvement over the best competitor TransVOD [17]. Furthermore, when changing the FasterRCNN base detector to a multi-frame aggregation method SELSA [57], STPN increases further to $86.5\%$ of AP50. When using a more strict AP metric, STPN with FasterRCNN and SELSA achieve remarkable improvements of $+6.5 / 6.9\%$ over the best competitor EOVOD [44], respectively. Finally, when changing to stronger backbones, STPN with the Swin-B backbone achieves further improved + +
ModelBase DetectorBackboneAPAP50AP75\( AP_S \)\( AP_M \)\( AP_L \)
CHP [61]CenterNetR101-76.7----
EOVOD [44]FCOSR10154.179.859.510.528.360.1
RDN [13]FasterRCNNR10153.481.860.18.527.459.6
TransVOD [17]Deformable DETRR101-82.0----
MEGA [7]FasterRCNNR10153.282.959.29.129.459.1
TDViT [45]FasterRCNNSwinT50.979.955.79.126.957.2
SELSA* [57]FasterRCNNSwinT52.382.358.211.428.857.9
TransVOD [17]Deformable DETRSwinT-83.7----
STPNFasterRCNNSwinT60.685.068.812.133.766.8
STPNSELSASwinT60.986.568.613.334.867.1
MEGA [7]FasterRCNNResNeXt101-84.1----
MAMBA [43]FasterRCNNResNeXt101-86.7----
TransVOD [17]Deformable DETRSwinB-90.0----
STPNFasterRCNNSwinB63.686.171.612.133.870.3
STPNSELSASwinB65.290.673.514.137.771.8
+ +Table 2. Comparison with state-of-the-art methods on ImageNet VID for video object detection. * represents our re-implementation. + +
ModelBackboneAPAP50AP75FPS
VisTR [55]R10140.164.045.057.7
CrossVISR10136.657.339.735.6
MaskPropR10142.5-45.6-
SeqFormerR10149.071.155.764.6
IFC [22]R10144.669.249.589.4
IDOL [60]R10150.173.156.126.0
VITA [18]R10151.975.457.0-
M2F-VIS [8]SwinT51.575.056.5-
MinVIS [20]SwinT51.975.158.227.1
STPNMinVISSwinT54.878.761.026.5
+ +performance of $65.2\%$ AP. Compared to TransVOD, STPN with SELSA makes a $0.6\%$ improvement in AP50. + +Video Instance Segmentation. We compare STPN with state-of-the-art methods in Table 3. Using the M2F-VIS [8] as the baseline segmentation method, STPN improves performance by large margins with $2.9 / 3.6\%$ of AP/AP50, respectively. Compared with SOTA methods with similar backbone complexities, for example, Swin-T or ResNet-101, STPN achieves the best accuracy. Specifically, STPN improves the performance of the best competitor VITA [18] by $2.9 / 3.3\%$ of AP/AP50, respectively. + +Visual Object Tracking. Table 4 compares STPN with state-of-the-art methods. We use MixFormer [11] as our baseline. Specifically, when CVT-22k and CVT-1k are used + +Table 3. Comparison with state-of-the-art methods for video instance segmentation on the YouTube VIS 2019 dataset. + +
ModelBackboneAOSR0.5SR0.75
SiamRPN++ [29]R10151.761.632.5
SiamR-CNN [50]R10164.972.859.7
PrDiMP [12]R10163.473.854.3
TrDiMP [52]R10167.177.758.3
TREG [10]R10166.877.857.2
TransT [6]R10167.176.860.9
STARK [62]R10168.878.164.1
MixFormer[11]CVT-22K70.780.067.8
MixFormer [11]CVT-1K71.279.965.8
STPNMixFormerCVT-22K71.881.669.0
STPNMixFormerCVT-1K72.781.967.7
+ +Table 4. Comparison with state-of-the-art tracking methods on the GOT-10k dataset. + +as the backbone, STPN improves the AO of MixFormer by $+1.1\%$ and $+1.5\%$ , respectively. STPN achieves the best performance compared to other SOTA methods on the GOT-10K dataset. + +These experimental results demonstrate that STPN is easy to generalise to various video understanding tasks and achieves remarkable performance across all three different video tasks. + +# 5.3. Ablation Study + +We investigate the effect of different designs of our approach on the overall performance. All experiments in §5.3 are conducted on the ImageNet VID dataset for video object detection. + +![](images/1742272ceb30c4fa5cacb8a90d74ec157f4b54d88e0bcae83df0fe71bdcd75d1.jpg) +(a) STPN with Swin-T backbone + +![](images/d35e8a0c5af68a5f91bb35cfefab442cae421062b4d9d594962028686558e1c2.jpg) +(b) STPN with Swin-B backbone + +![](images/8528e5a397210b5681701b2b60a2805689c6295a707676118c9df7053ee1fd4d.jpg) +(c) Comparison with STOA + +![](images/d02caf1694f3ee160a8f9cfa7c658e7b62745af656d84cbee610f8bee087dc76.jpg) +Figure 4. Comparisons between the transformer-based and the Mixer-based DVP predictors. +Figure 5. Comparisons of STPN deep and STPN shallow on transformer encoders with different scales. + +DVP Predictor Designs. We compare the effects of DVP predictor designs on different baselines. Figure 4 (a) shows the results on the Swin-T backbone. The FasterRCNN baseline achieves $80.7\%$ AP50, runs at 23.12 FPS, and has 44.90M parameters. Using the transformer-based predictor and the Mixer-based predictor, STPN achieves $85.0 / 84.5\%$ AP50 and runs at 22.27/22.41 FPS and has 45.01/44.96M parameters, respectively. These results show that the transformer-based predictor can achieve better accuracy, while the Mixer-based predictor has less complexity and faster speed. Experiments on the Swin-B backbone are shown in Figure 4 (b) and the experimental results indicate the same conclusion. + +Efficiency of STPN. Compared with other state-of-the-art methods, STPN with both predictor designs actually achieves a significantly improved speed-accuracy trade-off, as shown in Figure 4 (c). These results demonstrate that the pipeline of STPN is very efficient and is not sensitive to different predictor designs. We choose the transformer-based predictor as our default setting since it achieves a slightly higher accuracy with a comparable fast speed. + +STPN Shallow v.s. STPN Deep. As shown in the Figure 5, we conduct experiments on Swin-T/S/B/L variants. The + +
(a)S12481632
AP58.759.660.160.660.560.5
(b)K1357911
AP52.653.558.960.660.660.7
FPS22.422.422.322.321.720.5
(c)NP35791113
AP60.260.360.660.660.560.6
FPS22.922.722.321.621.019.9
+ +Table 5. Effect of hyper-parameters in STPN: the temporal stride $S$ , the number of support frames $K$ , and the number of dynamic video prompts $N_P$ . Different choices of hyper-parameters are listed in rows highlighted in grey. + +first finding is that both STPN shallow and STPN deep can continuously improve performance on all Swin variants, indicating that STPN has good compatibility with backbone scales. Another finding which contradicts our intuition is that surprisingly STPN shallow achieves better performance than STPN deep on three Swin variants. We believe the reason is that STPN shallow has fewer parameters and thus is easier to optimise in the training process. In contrast, the Swin-L variant is much more complex and thus needs more adjustments in the intermediate transformer layers. By default, we use STPN shallow. + +Support Frame Sampling. We use two hyper-parameters $S$ and $K$ to control the sampling of support frames. Table 5 (a) and Table 5 (b) show the effects of choosing different values of $S$ and $K$ , respectively. Firstly, we fix the value of $K$ as 7, so 15 surrounding support frames are used in total. Then, we vary the temporal stride $S$ from 1 to 32. As shown in Table 5 (a), $S = 8$ achieves the best performance. Afterwards, with $S$ fixed as 8, we verify the effect of different $K$ with respect to both performance (AP) and speed (FPS). Specifically, when $K = 7$ , STPN achieves a good speed-accuracy trade-off with $60.6\%$ AP and 22.3 + +![](images/ada8eb29dcbe02f5dbc9a30def17471d47b3707032bad0dde499e02bbcdd2f7d.jpg) +Figure 6. Bounding boxes and Grad-CAM visualisations of Faster-RCNN and STPN. Each column shows a video frame from the ImageNet VID validation set. + +![](images/66c89b35a9f5accb099d5241eb3615b1061a0dc6ba4ddb9912809dc388f7732a.jpg) +Figure 7. Instance mask visualisations of MinVIS [20] and STPN. Each column shows a video frame from the YouTube VIS dataset. MinVIS fails to generate accurate masks in low-quality frames. + +FPS. Therefore, we choose $S = 8$ and $K = 7$ as default. + +Number of dynamic video prompts. To explore the effect of the number of dynamic video prompts $N_P$ , we show the performance and speed results by varying this number from 3 to 13 in Table 5 (c). The best speed-accuracy trade-off is obtained when $N_P$ is 7. In particular, once $N_P$ is greater than 7, AP is less affected by the change of $N_P$ . Thus, we use 7 as the default value of $N_P$ . + +# 5.4. Qualitative Results + +Grad-CAM [42] is used to visualise the "concentration" areas of the FasterRCNN baseline and STPN in some challenging scenarios, for example, motion blur, occlusion, and both. As shown in Figure 6, FasterRCNN either fails to detect the blurry object or generates false positive detections, while STPN can produce accurate detections. In the first column, the squirrel is very blurry because of its fast movement. As a result, FasterRCNN cannot detect the squirrel, and thus the Grad-CAM does not show much attention to the region of the squirrel. On the contrary, STPN successfully detects the squirrel, and Grad-CAM + +![](images/9263a3183aed34777fc3a75a304c487996835dbbd2957f0f3045f510fadc61d6.jpg) +(a) Fast-speed objects +Figure 8. t-SNE visualisation of the object features produced by FasterRCNN and STPN on (a) fast speed and (b) medium speed objects in the ImageNet VID dataset. + +![](images/173ffe4c3326581a32b8b833d1a268085b38a726b0cdb71cc6dc263eba87451d.jpg) +(b) Medium-speed objects + +shows a good concentration map in the squirrel region. + +Instance Mask Visualisation. Figure 7 shows the mask segmentation results of the state-of-the-art method MinVIS [20] and the STPN in deteriorated frames. For example, in the first and the second columns, MinVIS fails to generate instance masks for the surfboard and the fox because of motion blur but our STPN can generate accurate instance masks in these frames. The third column shows a case of occlusion. The left deer is occluded by the right one, so the MinVIS fails to generate a mask in the head area of the left dear. Besides, it cannot distinguish the two deer and generate one instance mask. In contrast, STPN works well for generating the mask of the occluded deer and distinguishing the two instances of deer. + +t-SNE [48] is a statistical method for visualising high-dimensional data in 2D maps and is usually used to demonstrate the discriminative abilities of different models in classification tasks. We visualise the object features extracted by FasterRCNN and STPN using t-SNE. Figure 8 (a) and (b) show the results on the objects with the fast and medium moving speeds, respectively. We can see that STPN has better clustering results than the FasterRCNN baseline, demonstrating that STPN can improve the discriminative ability of features. Details of how to define different moving speeds and how to generate object features are illustrated in the supplementary material. + +# 6. Conclusion + +To the best of our knowledge, we are the first to explore promoting techniques for robust video feature extraction. We present a neat and unified framework, named Sptio-Temporal Prompting Network (STPN). STPN simplifies the current pipeline for video understanding tasks. Moreover, STPN is easy to generalise to various video tasks. We conduct extensive experiments and report STPN's superior performance on three widely-used video benchmarks: ImageNet VID for video object detection, YouTube for video instance segmentation, and GOT-10K for visual object tracking. We hope our work can inform future research on robust video understanding. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4 +[2] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolitional siamese networks for object tracking. In ECCV, 2016. 2 +[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 3 +[5] Jianguo Chen, Kenli Li, Qingying Deng, Keqin Li, and S Yu Philip. Distributed deep learning model for intelligent video surveillance systems with edge computing. Industrial Informatics, 2019. 1 +[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Hutchuan Lu. Transformer tracking. In CVPR, 2021. 1, 2, 6 +[7] Yihong Chen, Yue Cao, Han Hu, and Liwei Wang. Memory enhanced global-local aggregation for video object detection. In CVPR, 2020. 1, 6 +[8] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 2, 6 +[9] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 5 +[10] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Target transformed regression for accurate tracking. arXiv preprint arXiv:2104.00403, 2021. 1, 2, 6 +[11] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 1, 2, 5, 6 +[12] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 6 +[13] Jiajun Deng, Yingwei Pan, Ting Yao, Wengang Zhou, Houqiang Li, and Tao Mei. Relation distillation networks for video object detection. In ICCV, 2019. 1, 2, 6 +[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 2 +[15] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. NeurIPS, 2020. 3 +[16] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2 +[17] Lu He, Qianyu Zhou, Xiangtai Li, Li Niu, Guangliang Cheng, Xiao Li, Wenxuan Liu, Yunhai Tong, Lizhuang Ma, and Liqing Zhang. End-to-end video object detection with spatial-temporal transformers. In ACMMM, 2021. 2, 5, 6 + +[18] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 6 +[19] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, 2018. 1 +[20] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6, 8 +[21] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE transactions on pattern analysis and machine intelligence, 43(5):1562-1577, 2019. 2, 5 +[22] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 2021. 1, 2, 6 +[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 2, 3, 4 +[24] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 2, 3 +[25] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 1 +[26] Hamed Kiani Galoogahi, Ashton Fagg, and Simon Lucey. Learning background-aware correlation filters for visual tracking. In ICCV, 2017. 1 +[27] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 3 +[28] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2 +[29] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2, 6 +[30] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 2 +[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5 +[32] Liangkai Liu, Sidi Lu, Ren Zhong, Baofu Wu, Yongtao Yao, Qingyang Zhang, and Weisong Shi. Computing systems for autonomous driving: State of the art and challenges. IEEE Internet of Things Journal, 8(8):6469-6486, 2020. 1 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3, 5 + +[34] Rashmika Nawaratne, Damminda Alahakoon, Daswin De Silva, and Xinghuo Yu. Spatiotemporal anomaly detection using deep learning for real-time video surveillance. Industrial Informatics, 16(1):393-402, 2019. 1 +[35] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, 2022. 3 +[36] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. In NeurIPS, 2022. 3 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2 +[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3 +[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 5 +[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5 +[42] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 8 +[43] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Mamba: Multi-level aggregation via memory bank for video object detection. In AAAI, 2021. 1, 2, 5, 6 +[44] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Efficient one-stage video object detection by exploiting temporal consistency. In ECCV, 2022. 5, 6 +[45] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Tdvit: Temporal dilated video transformer for dense video tasks. In ECCV, 2022. 2, 6 +[46] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. In NeurIPS, 2021. 4 +[47] Jack Valmadre, Luca Bertinetto, João Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In CVPR, 2017. 1 +[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8 + +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2, 3 +[50] Paul Voigtlaender, Jonathon Luiten, Philip H. S. Torr, and Bastian Leibe. Siam R-CNN: visual tracking by re-detection. In CVPR, 2020. 6 +[51] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 2 +[52] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 6 +[53] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation: A unifying approach. In CVPR, 2019. 2 +[54] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 1 +[55] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 1, 2, 6 +[56] Fei Wu, Qingzhong Wang, Jiang Bian, Ning Ding, Feixiang Lu, Jun Cheng, Dejing Dou, and Haoyi Xiong. A survey on video action recognition in sports: Datasets, methods and applications. Multimedia, 2022. 1 +[57] Haiping Wu, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Sequence level semantics aggregation for video object detection. In ICCV, 2019. 1, 2, 5, 6 +[58] Haiping Wu, Bin Xiao, Noel C. F. Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In ICCV, 2021. 3, 5 +[59] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 2 +[60] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 2, 6 +[61] Zhujun Xu, Emir Hrustic, and Damien Vivet. Centernet heatmap propagation for real-time video object detection. In ECCV, 2020. 6 +[62] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 1, 2, 6 +[63] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 5 +[64] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1 +[65] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Unified vision and language prompt learning. arXiv preprint arXiv:2210.07225, 2022. 3 +[66] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In ICCV, 2017. 5 + +[67] Zheng Zhu, Qiang Wang, Bo Li, Wei Wu, Junjie Yan, and Weiming Hu. Distractor-aware siamese networks for visual object tracking. In ECCV, 2018. 2 \ No newline at end of file diff --git a/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/images.zip b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..08783be056ef4590ffcc1af40440989895b10950 --- /dev/null +++ b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21bbc02fe03d6b1c981afdbf38879f7e7381d3dd0af8dc38eea35487d4cbd8bf +size 684201 diff --git a/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/layout.json b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6253575cdc15d116320651701ee449234e3c92cb --- /dev/null +++ b/2023/Spatio-temporal Prompting Network for Robust Video Feature Extraction/layout.json @@ -0,0 +1,9567 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 525, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 525, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 525, + 121 + ], + "type": "text", + "content": "Spatio-temporal Prompting Network for Robust Video Feature Extraction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "spans": [ + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "text", + "content": "Guanxiong Sun" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "text", + "content": ", Chi Wang" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "text", + "content": ", Zhaoyu Zhang" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "text", + "content": ", Jiankang Deng" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "text", + "content": ", Stefanos Zafeiriou" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "text", + "content": ", Yang Hua" + }, + { + "bbox": [ + 52, + 152, + 539, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 176, + 495, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 176, + 495, + 191 + ], + "spans": [ + { + "bbox": [ + 97, + 176, + 495, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 176, + 495, + 191 + ], + "type": "text", + "content": "Queen's University Belfast " + }, + { + "bbox": [ + 97, + 176, + 495, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 176, + 495, + 191 + ], + "type": "text", + "content": "Huawei UKRD " + }, + { + "bbox": [ + 97, + 176, + 495, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 176, + 495, + 191 + ], + "type": "text", + "content": "Imperial College London" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 191, + 512, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 191, + 512, + 206 + ], + "spans": [ + { + "bbox": [ + 82, + 191, + 512, + 206 + ], + "type": "text", + "content": "{gsun02, cwang38, zhang55,Y.Hua} @qub.ac.uk, {j.deng16,s.zafeiriou} @imperial.ac.uk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 242, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 242, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 242, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 269, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 290, + 568 + ], + "type": "text", + "content": "Frame quality deterioration is one of the main challenges in the field of video understanding. To compensate for the information loss caused by deteriorated frames, recent approaches exploit transformer-based integration modules to obtain spatio-temporal information. However, these integration modules are heavy and complex. Furthermore, each integration module is specifically tailored for its target task, making it difficult to generalise to multiple tasks. In this paper, we present a neat and unified framework, called Spatio-Temporal Prompt Network (STPN). It can efficiently extract robust and accurate video features by dynamically adjusting the input features in the backbone network. Specifically, STPN predicts several video prompts containing spatio-temporal information of neighbour frames. Then, these video prompts are pretended to the patch embeddings of the current frame as the updated input for video feature extraction. Moreover, STPN is easy to generalise to various video tasks because it does not contain task-specific modules. Without bells and whistles, STPN achieves state-of-the-art performance on three widely-used datasets for different video understanding tasks, i.e., ImageNetVID for video object detection, YouTubeVIS for video instance segmentation, and GOT-10k for visual object tracking. Codes are available at https://github.com/guanxiongsun/STPN" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 128, + 608 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 712 + ], + "type": "text", + "content": "Video understanding is a fundamental research direction in the field of computer vision. It plays an important role in many real-world applications, such as autonomous driving [64, 32], video surveillance [34, 5], and sports analysis [56, 25]. However, a significant challenge in this field is the deterioration of video frames due to motion blur, occlusion, and deformation, which makes it difficult to extract relevant information." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 245, + 428, + 346 + ], + "blocks": [ + { + "bbox": [ + 307, + 245, + 428, + 346 + ], + "lines": [ + { + "bbox": [ + 307, + 245, + 428, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 428, + 346 + ], + "type": "image", + "image_path": "7c51f56c56ea88b261e25340276a14265d27c959fa6d11c2f69d511fe7150034.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 350, + 402, + 359 + ], + "lines": [ + { + "bbox": [ + 337, + 350, + 402, + 359 + ], + "spans": [ + { + "bbox": [ + 337, + 350, + 402, + 359 + ], + "type": "text", + "content": "(a) Current pipeline" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 434, + 252, + 541, + 347 + ], + "blocks": [ + { + "bbox": [ + 434, + 252, + 541, + 347 + ], + "lines": [ + { + "bbox": [ + 434, + 252, + 541, + 347 + ], + "spans": [ + { + "bbox": [ + 434, + 252, + 541, + 347 + ], + "type": "image", + "image_path": "a951f0b22ba2fe2e8fa03193987df5e5509e7638d5ad9b1bdeecef2a83eaf4db.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 349, + 501, + 359 + ], + "lines": [ + { + "bbox": [ + 468, + 349, + 501, + 359 + ], + "spans": [ + { + "bbox": [ + 468, + 349, + 501, + 359 + ], + "type": "text", + "content": "(b) STPN" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 374, + 547, + 462 + ], + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 462 + ], + "type": "text", + "content": "Figure 1. Comparisons between pipelines of (a) existing methods and (b) the proposed Spatio-temporal Prompting Network (STPN). Existing methods introduce complex and task-specific integration modules (□ □ □) after backbone networks (□). In contrast, STPN is a unified framework for multiple tasks. A lightweight dynamic video prompt (DVP) predictor (□) generates a set of DVPs to adjust input before backbone networks (□). Best viewed in colour." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": "To overcome this challenge, inspired by the great success of transformers in many computer vision tasks [49, 54, 27, 19, 4], researchers explore various transformer-based integration modules to alleviate the information loss on the deteriorated video frames. For example, in video object detection, transformer-based feature aggregation methods [13, 7, 57, 43] are investigated to enhance the feature of proposals [40] in the detection head. In video instance segmentation, deteriorated frames usually cause wrong instance associations in a sequence, so 3D mask decoders [55, 22, 8] are introduced to learn associations of instance masks in an end-to-end manner. For visual object tracking, traditional correlation filters [26, 47] meet their limitations in severely degraded frames. Therefore, transformer-based integration modules [11, 62, 6, 10] are proposed to better capture complicated correlations between the template region and the search region. To conclude, the summarised pipeline of recent methods is shown in Figure 1 (a). A backbone network extracts spatially-only features from the current frame and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13587" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": "the support frames in the same video. Then, different integration modules integrate spatial-only features on multiple frames to obtain spatio-temporal features. Finally, the spatio-temporal features are used for the targeted video understanding task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 262 + ], + "type": "text", + "content": "While the current pipeline achieves good performance, there are still two major limitations of transformer-based integration modules. Firstly, these integration modules are involved as an extra component after the backbone network, leading to increased complexity and additional computational costs. Secondly, each integration module is tailored specifically for the target task and thus cannot be generalised to multiple video tasks. Hence, we put forward the following question: can we remove the complex integration modules and directly obtain spatio-temporal information in backbone networks?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 263, + 286, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 286, + 491 + ], + "type": "text", + "content": "To answer this question, inspired by recent prompting techniques [24, 23, 28], in this paper, we present a neat and unified framework, called Spatio-Temporal Prompting Network (STPN). Instead of using complex integration modules, STPN simplifies the current pipeline for video understanding by introducing spatio-temporal information into the backbone network, as shown in Figure 1 (b). Specifically, given a video frame, we propose a dynamic video prompt (DVP) predictor to generate several video prompts according to support frames. Then, the predicted DVPs are prepended to the patch embeddings of the current frame as the updated input. Finally, a vision transformer backbone network extracts video features using the updated input for future video understanding tasks. It is worth noting that the DVP predictor is a lightweight structure and introduces only a small number of extra parameters, e.g., 0.11M. Moreover, STPN can be easily adapted to different video understanding tasks, since it does not contain task-specific modules, and all modifications happen before the backbone network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 491, + 286, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 286, + 646 + ], + "type": "text", + "content": "In summary, our key contributions are: (1) We present the Spatio-Temporal Prompting Network (STPN) that can extract robust video features on deteriorated video frames. STPN simplifies the current pipeline for video understanding and is easy to generalise to different video understanding tasks. (2) To the best of our knowledge, we are the first to explore promoting techniques for robust video feature extraction on the task of video object detection (VOD), video instance segmentation (VIS), and visual object tracking (VOT). (3) Without bells and whistles, STPN achieves state-of-the-art performance on three widely-used video understanding benchmarks, i.e., ImageNet-VID [41] for VOD, YouTube-VIS [63] for VIS, and GOT-10k [21] for VOT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 657, + 133, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 133, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 133, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "In this section, we first review the state-of-the-art (SOTA) methods for related video understanding tasks. Then, we introduce the visual prompting techniques." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 86, + 545, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 86, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 86, + 545, + 265 + ], + "type": "text", + "content": "Video Object Detection (VOD). SOTA VOD methods use attention modules to aggregate features of support frames onto the current frame and thus enhance the feature quality of the current frame. For example, RDN [13] and SELSA [57] utilise relation modules to aggregate proposal features from local frames and global frames, respectively. These methods use sample memory architectures to store features on support frames, e.g., a sliding window or a memory queue, which limit the \"diversity\" of support features. Therefore, MAMBA [43] proposes a memory bank architecture that can partially update and sample support features and thus achieve a good speed-accuracy trade-off. More recent approaches TransVOD [17] and TDViT [45] introduce transformers into the field and further improve the performance of VOD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 279, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 483 + ], + "type": "text", + "content": "Video Instance Segmentation (VIS). The goal of VIS is to simultaneously segment and track all object instances in the video sequence. There are two categories of approaches: Firstly, per-frame approaches conduct instance segmentation on each frame independently and then associate instance masks by post-processing. MaskTrackRCNN [63] adds a tracking prediction head on MaskRCNN [16] for instance association. More recently, some approaches [60, 20, 18] use transformers [49, 4] to reason the relationships between instances. Secondly, per-clip approaches [55, 59, 8, 22] directly predict 3D masks for each instance in a video. VisTR [55] proposes a transformer encoder-decoder structure to obtain spatio-temporal features, which are then used in an instance sequence matching module to get object tracks. IFC [22] uses memory tokens inside of the transformer encoder module to transfer temporal information between input frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 496, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 640 + ], + "type": "text", + "content": "Visual Object Tracking (VOT). Object tracking is one of the most fundamental research topics in computer vision and has a long history. Siamese-based methods [2, 30, 29, 53, 67] consider tracking as a template-matching problem by extracting correlation feature maps of the target and the search region via two-head siamese networks. Recently, transformers have been introduced to perform attention-based correspondence modelling and these methods [6, 51, 10, 62] achieve SOTA performance. Among these trackers, MixFormer [11] is a unified framework established solely on transformers for feature extraction, correspondence modelling, and result generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Prompting. The concept of prompting is originally proposed in natural language processing (NLP). Prompting refers to designing instructions and preponding them to the input so that the pre-trained language models [38, 14, 3] can \"understand\" the task. Recently, prompting techniques" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13588" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 167 + ], + "type": "table", + "html": "
MeaningSymbol
The i-th transformer layerLi
Collection of the output embeddings from LiEi
Number of embeddings in Eini
Dimension of embeddings in Eidi
Collection of dynamic video promptsP
Number of dynamic video promptsNP
", + "image_path": "fa96e9674be6aa0356ecce6f27806d54b0ec77b578d07407c49314c8cad22559.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 174, + 214, + 185 + ], + "lines": [ + { + "bbox": [ + 119, + 174, + 214, + 185 + ], + "spans": [ + { + "bbox": [ + 119, + 174, + 214, + 185 + ], + "type": "text", + "content": "Table 1. List of notations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 202, + 289, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 202, + 289, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 289, + 310 + ], + "type": "text", + "content": "have been explored in many computer vision tasks. A major trend [35, 36, 24] is to transfer pre-trained vision-language models [15, 37, 39] to the task of video recognition, e.g., video classification and action recognition. Another popular research direction is parameter-efficient transfer learning via prompt tuning. For example, VPT [23] and UPT [65] tune a small number of parameters in the input space of a model and can achieve good performance comparable to the full fine-tuning scheme on a variety of recognition tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 320, + 132, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 320, + 132, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 320, + 132, + 333 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "text", + "content": "In this section, we define notations necessary for subsequent demonstrations. " + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "text", + "content": " denotes the current frame and " + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "text", + "content": " is the current time step in the video. " + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "inline_equation", + "content": "I_{sup}" + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "text", + "content": " denotes the support frames used to predict video prompts and aid in the feature extraction process. Specifically, " + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "inline_equation", + "content": "I_{sup} = \\{I_{\\tau}\\}_{t - S(K / 2)}^{t + S(K / 2)}" + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "text", + "content": " denotes the number of support frames (assumed to be an even number) and " + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 340, + 287, + 437 + ], + "type": "text", + "content": " denotes the frame interval in the time dimension between support frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": "Transformers encoders [27, 33, 58] are used to extract feature embeddings from video frames. A transformer encoder contains a patch embedding layer and " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": " transformer layers. The input frame is first divided into " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": " non-overlapping patches " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_j\\in \\mathbb{R}^{3\\times h\\times w}|j\\in \\mathbb{N},1\\leq j\\leq n\\}" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": " are the height and width of the patches. Each patch is then embedded into patch embeddings, denoted as " + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{E}\\in \\mathbb{R}^{n_0\\times d_0}" + }, + { + "bbox": [ + 46, + 437, + 287, + 534 + ], + "type": "text", + "content": ". Other notations are summarised in Table 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 544, + 129, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 129, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 129, + 558 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 115, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 115, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 115, + 574 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "The overview of our framework is shown as in Figure 2 (a). There are two stages in the framework: the predicting stage and the prompting stage. The goal of the predicting stage is to generate a set of prompts that vary according to spatio-temporal information. We first select " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " support frames around " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and then pass the support frames " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_{sup}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " to the patch embedding layer and the transformer encoder to extract image embeddings, called support embeddings. The support embeddings are then passed to the dynamic video prompt (DVP) predictor network as the input. The DVP predictor outputs " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{P}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " dynamic video prompts, denoted as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": ", which contain the spatio-temporal information of the current video. Details of the DVP predictor network are illustrated in " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\S 4.2" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "type": "text", + "content": "The second stage is the prompting stage whose goal is to extract spatio-temporal features of the current frame " + }, + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "type": "text", + "content": " is prepended to the patch embeddings of the current frame " + }, + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 109, + 546, + 228 + ], + "type": "text", + "content": ". Then, the mixed embeddings are passed into the transformer encoder to produce the spatiotemporal embeddings. Finally, the spatio-temporal embeddings are used for different video understanding tasks, e.g., video object detection, video instance segmentation and visual object tracking. More details of the prompting stage are presented in §4.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 237, + 406, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 237, + 406, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 237, + 406, + 251 + ], + "type": "text", + "content": "4.2. Predicting Stage" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": "Given the current frame " + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": ", we first sample " + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": " support frames around " + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": " as support frames. The support frames are then passed into the patch embedding layer and the transformer encoder:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 314, + 545, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 314, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 319, + 314, + 545, + 329 + ], + "type": "interline_equation", + "content": "\\mathbf {E} _ {\\tau} ^ {i} = \\mathbf {L} ^ {i} \\left(\\mathbf {E} _ {\\tau} ^ {i - 1}\\right) \\quad \\mathbf {E} _ {\\tau} ^ {i} \\in \\mathbb {R} ^ {n _ {i} \\times d _ {i}}, i = 1, 2, \\dots , L. \\tag {1}", + "image_path": "061704ad3036f9bb7feaee42b79486e331fbd11017e82e62b68c12153dea5551.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "text", + "content": "Therefore, the collection of support embeddings is denoted as: " + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{sup} = \\{\\mathbf{E}_{\\tau}^{\\mathbf{L}} \\in \\mathbb{R}^{n_L \\times d_L} | \\tau \\in \\mathcal{T}\\}" + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "text", + "content": " is the collection of sampled time steps around the current frame, " + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = [t - S(K/2), \\dots, t - S, t + S, \\dots, t + S(K/2)]" + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "text", + "content": ". For example, if " + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "inline_equation", + "content": "K = 6" + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "inline_equation", + "content": "S = 2" + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "text", + "content": ", three support frames are sampled from the past of the current time step " + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "text", + "content": ", and three are sampled from the future, with a time interval of 2 between each sampled frame in both directions. Given " + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{sup}" + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "text", + "content": ", we design two predicting networks, i.e., transformer-based predictor and mixer-based predictor, to obtain dynamic video prompts " + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N_P \\times d}" + }, + { + "bbox": [ + 304, + 338, + 546, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 483, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 602 + ], + "type": "text", + "content": "Transformer-based Predictor. This design is inspired by the object decoder in DETR [4] where a set of randomly initialised object queries are used to retrieve object features on the image embeddings. As depicted in Figure 2 (b), we introduce a set of learnable prompt queries " + }, + { + "bbox": [ + 304, + 483, + 545, + 602 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{N_P \\times d_L}" + }, + { + "bbox": [ + 304, + 483, + 545, + 602 + ], + "type": "text", + "content": " to gather spatio-temporal information from the support embeddings. Specifically, a multi-head attention (MHA) [49] module takes the support embeddings and the prompt queries as input, and produces the aggregated embeddings. The process is formulated as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 611, + 545, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 611, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 315, + 611, + 545, + 627 + ], + "type": "interline_equation", + "content": "K = \\operatorname {C o n c a t} \\left(\\mathbf {E} _ {\\sup }\\right) \\quad K \\in \\mathbb {R} ^ {K \\cdot n _ {L} \\times d _ {L}}, \\tag {2}", + "image_path": "93b0750296d6358c3eceb120466a2052b1ba4374cb7f1b68b401615f842638a9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 628, + 545, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 628, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 628, + 545, + 643 + ], + "type": "interline_equation", + "content": "\\hat {Q} = \\operatorname {M H A} (\\ln (Q), \\ln (K)) + Q \\quad \\hat {Q} \\in \\mathbb {R} ^ {N _ {P} \\times d _ {L}}, \\tag {3}", + "image_path": "02a489e7e2134a401a0d38042417feaab882db270b7f11eacc37881be915fd9c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}(\\cdot)" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " indicates the concatenation operation and LN denotes the LayerNorm layer. The output embeddings are then added back to the prompt queries and passed to the LayerNorm layer. The normalised output embeddings are further transformed by a simple feed-forward network" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13589" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 336, + 300 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 336, + 300 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 336, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 336, + 300 + ], + "type": "image", + "image_path": "ead93793a658ca120a22f9feca0ba40bbdeb68013e09e969ae7b4cc6824c71c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 302, + 197, + 312 + ], + "lines": [ + { + "bbox": [ + 151, + 302, + 197, + 312 + ], + "spans": [ + { + "bbox": [ + 151, + 302, + 197, + 312 + ], + "type": "text", + "content": "(a) Overview" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 337, + 76, + 518, + 293 + ], + "blocks": [ + { + "bbox": [ + 337, + 76, + 518, + 293 + ], + "lines": [ + { + "bbox": [ + 337, + 76, + 518, + 293 + ], + "spans": [ + { + "bbox": [ + 337, + 76, + 518, + 293 + ], + "type": "image", + "image_path": "f240f4df6c29134140ab3c4a679019ffbffa4a8bb8b150fd65afaefaaa835021.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 296, + 422, + 316 + ], + "lines": [ + { + "bbox": [ + 345, + 296, + 422, + 316 + ], + "spans": [ + { + "bbox": [ + 345, + 296, + 422, + 316 + ], + "type": "text", + "content": "(b) Transformer-based predictor" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 448, + 297, + 504, + 315 + ], + "lines": [ + { + "bbox": [ + 448, + 297, + 504, + 315 + ], + "spans": [ + { + "bbox": [ + 448, + 297, + 504, + 315 + ], + "type": "text", + "content": "(c) Mixer-based predictor" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "type": "text", + "content": "Figure 2. (a) An overview of our approach. In the predicting stage, a set of dynamic visual prompts (DVPs) " + }, + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "type": "text", + "content": " is generated by the DVP predictor that takes support embeddings " + }, + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{sup}" + }, + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "type": "text", + "content": " on support frames " + }, + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "type": "inline_equation", + "content": "I_{sup}" + }, + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "type": "text", + "content": " as the input. Then, in the prompting stage, predicted DVPs are prepended with the patch embeddings of the current frame to extract spatio-temporal embeddings via a transformer encoder which contains " + }, + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 330, + 545, + 396 + ], + "type": "text", + "content": " transformer layers. Finally, different task heads take the spatio-temporal embeddings and output final results for various general video tasks, e.g., video object detection, video instance segmentation, and visual object tracking. (b) Details of the transformer-based predictor. (c) Details of the Mixer-based predictor." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 416, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 440 + ], + "type": "text", + "content": "(FFN) with a residual connection that makes the dynamic video prompts:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 443, + 287, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 443, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 75, + 443, + 287, + 459 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\operatorname {F F N} (\\operatorname {L N} (\\hat {Q})) + \\operatorname {L N} (\\hat {Q}) \\quad \\mathbf {P} \\in \\mathbb {R} ^ {N _ {P} \\times d}. \\tag {4}", + "image_path": "2c36595179b4e0330734dcc5d1214d2f86cdd909b7fafee099bf5e28f36088ba.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 462, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 559 + ], + "type": "text", + "content": "Mixer-based Predictor. Inspired by the MLP-Mixer [46] (shortened as Mixer), we design a Mixer-based predictor to transform support embedding into dynamic video prompts. The details of a Mixer-based predictor are shown in Figure 2 (c). Unlike the transformer-based predictor whose first step is to concatenate all support embeddings, we conduct average pooling in the temporal dimension to obtain the average support embeddings " + }, + { + "bbox": [ + 46, + 462, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{E}}_{sup}" + }, + { + "bbox": [ + 46, + 462, + 287, + 559 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 562, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 562, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 77, + 562, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {E}}} _ {s u p} = \\operatorname {A v g P o o l} \\left(\\mathbf {E} _ {s u p}\\right) \\quad \\overline {{\\mathbf {E}}} _ {s u p} \\in \\mathbb {R} ^ {n _ {L} \\times d _ {L}}, \\tag {5}", + "image_path": "256d35e80a27f125f0600bb96e92d46e579708807c362a0af2254a02dfdcd69a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "text", + "content": "where AvgPool denotes the average pooling across the time dimension. Then, the average support embedding is normalised by a LayerNorm layer and further transposed and passed into a feed-forward network (FFN) to obtain the hidden feature " + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "inline_equation", + "content": "N_P" + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "text", + "content": " channels. Finally, the hidden feature " + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "text", + "content": " is used to generate " + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "inline_equation", + "content": "N_P" + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "text", + "content": " dynamic video prompts with " + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 580, + 287, + 664 + ], + "type": "text", + "content": " channels. These two steps can be formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 82, + 679, + 287, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 679, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 82, + 679, + 287, + 694 + ], + "type": "interline_equation", + "content": "h = \\operatorname {F F N} \\left(\\left(\\operatorname {L N} \\left(\\overline {{\\mathbf {E}}} _ {s u p}\\right)\\right) ^ {T}\\right) \\quad h \\in \\mathbb {R} ^ {d _ {L} \\times N _ {P}}, \\tag {6}", + "image_path": "66cdf2fd6f9049d2501ce6fd1e5df2a121f8cbca666ff34d9a18f9e2af1973b9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 82, + 696, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 696, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 82, + 696, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\operatorname {F F N} \\left(\\left(\\operatorname {L N} (h) ^ {T}\\right) \\quad \\mathbf {P} \\in \\mathbb {R} ^ {N _ {P} \\times d}, \\right. \\tag {7}", + "image_path": "85be03bb0f9895e9cde374f8d810707efd7e42ffd6148ca5d67ed7dd95eee136.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 415, + 544, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 544, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 544, + 441 + ], + "type": "text", + "content": "where LN and " + }, + { + "bbox": [ + 306, + 415, + 544, + 441 + ], + "type": "inline_equation", + "content": "(\\cdot)^T" + }, + { + "bbox": [ + 306, + 415, + 544, + 441 + ], + "type": "text", + "content": " denote the LayerNorm layer [1] and the transpose layer, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 445, + 408, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 408, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 408, + 459 + ], + "type": "text", + "content": "4.3. Prompting Stage" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "text", + "content": "Given the predicted dynamic video prompts " + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "text", + "content": ", we concatenate " + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "text", + "content": " together with the patch embeddings of the current frame " + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t" + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "text", + "content": " as the input of the transformer encoder. The details of this process are shown in the prompting stage (right part) of Figure 2 (a). More concretely, the process of applying " + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "text", + "content": " and extracting the spatio-temporal feature of the current frame " + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 304, + 464, + 545, + 548 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 552, + 545, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 552, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 552, + 545, + 568 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {Z} _ {t} ^ {1}, \\mathbf {E} _ {t} ^ {1} \\right] = \\mathbf {L} ^ {1} (\\operatorname {C o n c a t} (\\mathbf {P}, \\mathbf {E} _ {t})) \\tag {8}", + "image_path": "1596623aa6dfbe8e2557463285ef9437b40f2341c7580f847ad94d06ebc6a744.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 569, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 569, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 315, + 569, + 545, + 585 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {Z} _ {t} ^ {i}, \\mathbf {E} _ {t} ^ {i} \\right] = \\mathbf {L} ^ {i} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {Z} ^ {i - 1}, \\mathbf {E} _ {t} ^ {i - 1}\\right)\\right) \\quad i = 2, 3, \\dots , L \\tag {9}", + "image_path": "ddd7309d141e4949688114a10c2dbf31e4996d2ef600e0c071af03b7ff1b50b7.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 339, + 586, + 545, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 586, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 339, + 586, + 545, + 600 + ], + "type": "interline_equation", + "content": "y _ {t} = \\operatorname {H e a d} \\left(\\mathbf {E} _ {t} ^ {L}\\right), \\tag {10}", + "image_path": "93147d8c748e1b1c3cb23e6eb636bcd526aa00c3e4716f4f8e3ba393f58ba995.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": " indicates the concatenation operation and " + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_t^i\\in \\mathbb{R}^{N_P\\times d_i}" + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": " represents the extra embeddings produced by adding " + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": " in the input embeddings. The weights of the transformer encoder are shared in the predicting and prompting stages. Finally, the output embeddings " + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t^L" + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": " are passed into the head network for different video tasks." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "STPN Shallow v.s. STPN Deep. Following the protocols in VPT [23], we design two STPN variants, i.e., STPN shal" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13590" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 69, + 168, + 171 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 168, + 171 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 168, + 171 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 168, + 171 + ], + "type": "image", + "image_path": "cdbc7086953477a806bd7a791abc73e5ba842f6f738957c74fbb450ea128d27a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 172, + 148, + 182 + ], + "lines": [ + { + "bbox": [ + 88, + 172, + 148, + 182 + ], + "spans": [ + { + "bbox": [ + 88, + 172, + 148, + 182 + ], + "type": "text", + "content": "(a) SPTN shallow" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 198, + 287, + 263 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 263 + ], + "type": "text", + "content": "Figure 3. A comparison between the (a) STPN shallow and the (b) STPN deep. The green rounded rectangles denote predicted DVPs. In (a), one set of DVPs is prepended to the patch embeddings before the first layer of the transformer encoder. In (b), " + }, + { + "bbox": [ + 46, + 198, + 287, + 263 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 198, + 287, + 263 + ], + "type": "text", + "content": " sets of DVPs are predicted and then prepended to the input embeddings of all " + }, + { + "bbox": [ + 46, + 198, + 287, + 263 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 198, + 287, + 263 + ], + "type": "text", + "content": " transformer layers in the transformer encoder." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 276, + 171 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 276, + 171 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 276, + 171 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 276, + 171 + ], + "type": "image", + "image_path": "4077239bead0aa142b0b0ca06cc2ad4721a0f54e4cd63a2508fd3bf891ee8615.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 173, + 250, + 182 + ], + "lines": [ + { + "bbox": [ + 199, + 173, + 250, + 182 + ], + "spans": [ + { + "bbox": [ + 199, + 173, + 250, + 182 + ], + "type": "text", + "content": "(b) SPTN deep" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "text", + "content": "low " + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "inline_equation", + "content": "v.s." + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "text", + "content": " STPN deep, according to how many sets of dynamic video prompts (DVPs) are generated to adjust the transformer encoder. For example, in the shallow version, one set of DVPs is predicted and prepended only before the first transformer layer. In the deep version, " + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "text", + "content": " sets of DVPs are predicted via " + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "text", + "content": " separate FC layers, and each set of DVPs is prepended to the corresponding transformer layer. Formally, the DVPs of the " + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "text", + "content": "-th layer are denoted as " + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^i \\in \\mathbb{R}^{N_P \\times d_i}" + }, + { + "bbox": [ + 46, + 282, + 287, + 402 + ], + "type": "text", + "content": " and the feature extraction process Equation (9) is changed to:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 409, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 423 + ], + "type": "interline_equation", + "content": "\\left[ \\dots , \\mathbf {E} _ {t} ^ {i} \\right] = \\mathbf {L} ^ {i} \\left(\\operatorname {C o n c a t} \\left(\\mathbf {P} ^ {i - 1}, \\mathbf {E} _ {t} ^ {i - 1}\\right)\\right) \\quad i = 2, 3, \\dots , L. \\tag {11}", + "image_path": "6ceb30d520beaa65dc67fd9ca29c7e22faef0be372c6ec2467424339318c9101.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 430, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 455 + ], + "type": "text", + "content": "The difference between the shallow and deep variants is shown in Figure 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 464, + 127, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 127, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 127, + 478 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 484, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 568 + ], + "type": "text", + "content": "We first conduct experiments and compare SPTN with other state-of-the-art methods on three general video understanding tasks, i.e., video object detection (VOD), video instance segmentation (VIS) and visual object tracking (VOT). Then, in §5.3, we conduct ablation studies on how different design choices affect the overall performance, which helps to gain a deeper understanding of SPTN." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 575, + 107, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 107, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 107, + 588 + ], + "type": "text", + "content": "5.1. Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Datasets. For VOD, the ImageNetVID dataset [41] is used. It contains 3,862 training and 555 validation videos. Following previous approaches [43, 57, 66], we add overlapped 30 classes of the ImageNet DET dataset into the train set. Specifically, we sample 15 frames from each video in the VID train and at most 2,000 images per class from the DET train. For VIS, we train our models and report results on the YouTube VIS 2019 dataset [63]. The dataset contains 2,238 training, 302 validation, and 343 test videos, which covers 40 object categories. For VOT, we" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "choose the GOT10k [21] dataset, because it is a large-scale dataset (10,000 training videos and 180 test videos) with more than 1.5 million manually labelled bounding boxes, enabling stable training and evaluation of trackers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": "Evaluation Metrics. For video object detection and video instance segmentation, we report detailed results in the COCO [31] evaluation format. Specifically, we report the average precision (AP) metric which computes the average precision over ten IoU thresholds " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "[0.5:0.05:0.95]" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": " for all categories. Meanwhile, the COCO evaluation contains other important metrics, e.g., " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": " that are calculated at IoU thresholds of 0.50 and 0.75, and " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_S" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_M" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_L" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": " that are calculated on different object scales, i.e., small, medium and large. The metrics for detection and segmentation are noted with box and mask, respectively. For visual object tracking, we report the widely used average overlap (AO) and the success rate (SR). AO denotes the average overlaps between all ground truth and estimated bounding boxes, while SR measures the percentage of successfully tracked frames where the overlaps exceed a threshold, e.g., " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{SR}_{0.5}" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{SR}_{0.75}" + }, + { + "bbox": [ + 304, + 133, + 545, + 337 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 349, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 433 + ], + "type": "text", + "content": "Transformer Encoders. Our method is compatible with different transformer backbones. For demonstration, we conduct experiments on two well-known transformer backbones, i.e., Swin [33] and CvT [58]. Swin is used for video object detection and video instance segmentation. CvT is used for visual object tracking. The details of the backbones are illustrated in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "type": "text", + "content": "Training and Inference. All reported results are obtained with Python 3.7 and PyTorch 1.8.1 on Tesla V100 GPUs. We follow the protocols in TransVOD [17], Mask2Former [9], and MixFormer [11] to set the hyper-parameters, e.g., learning rate, data augmentations, optimiser, etc., for three tasks, respectively. More details about the hyper-parameter settings are described in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 539, + 489, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 539, + 489, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 539, + 489, + 551 + ], + "type": "text", + "content": "5.2. Comparisons with SOTA Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "Video Object Detection. Table 2 shows the results of STPN and state-of-the-art methods. Firstly, in the AP50 metric, using a simple FasterRCNN [40] base detector with the Swin-T [33] backbone, STPN achieves " + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "inline_equation", + "content": "85.0\\%" + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": " of AP50, which makes a " + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": " improvement over the best competitor TransVOD [17]. Furthermore, when changing the FasterRCNN base detector to a multi-frame aggregation method SELSA [57], STPN increases further to " + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "inline_equation", + "content": "86.5\\%" + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": " of AP50. When using a more strict AP metric, STPN with FasterRCNN and SELSA achieve remarkable improvements of " + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "inline_equation", + "content": "+6.5 / 6.9\\%" + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": " over the best competitor EOVOD [44], respectively. Finally, when changing to stronger backbones, STPN with the Swin-B backbone achieves further improved" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13591" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 68, + 509, + 284 + ], + "blocks": [ + { + "bbox": [ + 83, + 68, + 509, + 284 + ], + "lines": [ + { + "bbox": [ + 83, + 68, + 509, + 284 + ], + "spans": [ + { + "bbox": [ + 83, + 68, + 509, + 284 + ], + "type": "table", + "html": "
ModelBase DetectorBackboneAPAP50AP75\\( AP_S \\)\\( AP_M \\)\\( AP_L \\)
CHP [61]CenterNetR101-76.7----
EOVOD [44]FCOSR10154.179.859.510.528.360.1
RDN [13]FasterRCNNR10153.481.860.18.527.459.6
TransVOD [17]Deformable DETRR101-82.0----
MEGA [7]FasterRCNNR10153.282.959.29.129.459.1
TDViT [45]FasterRCNNSwinT50.979.955.79.126.957.2
SELSA* [57]FasterRCNNSwinT52.382.358.211.428.857.9
TransVOD [17]Deformable DETRSwinT-83.7----
STPNFasterRCNNSwinT60.685.068.812.133.766.8
STPNSELSASwinT60.986.568.613.334.867.1
MEGA [7]FasterRCNNResNeXt101-84.1----
MAMBA [43]FasterRCNNResNeXt101-86.7----
TransVOD [17]Deformable DETRSwinB-90.0----
STPNFasterRCNNSwinB63.686.171.612.133.870.3
STPNSELSASwinB65.290.673.514.137.771.8
", + "image_path": "748f95824273036ff409bb9df00f87852de42152fd59399b1f0d39370c128694.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 322, + 285, + 463 + ], + "blocks": [ + { + "bbox": [ + 47, + 292, + 525, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 292, + 525, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 525, + 304 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art methods on ImageNet VID for video object detection. * represents our re-implementation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 322, + 285, + 463 + ], + "lines": [ + { + "bbox": [ + 48, + 322, + 285, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 285, + 463 + ], + "type": "table", + "html": "
ModelBackboneAPAP50AP75FPS
VisTR [55]R10140.164.045.057.7
CrossVISR10136.657.339.735.6
MaskPropR10142.5-45.6-
SeqFormerR10149.071.155.764.6
IFC [22]R10144.669.249.589.4
IDOL [60]R10150.173.156.126.0
VITA [18]R10151.975.457.0-
M2F-VIS [8]SwinT51.575.056.5-
MinVIS [20]SwinT51.975.158.227.1
STPNMinVISSwinT54.878.761.026.5
", + "image_path": "8684c4596a2efdf564538e408df8c299a67f660d123f6cf0368d3e0ad4b63ca5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 517, + 286, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 286, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 286, + 541 + ], + "type": "text", + "content": "performance of " + }, + { + "bbox": [ + 47, + 517, + 286, + 541 + ], + "type": "inline_equation", + "content": "65.2\\%" + }, + { + "bbox": [ + 47, + 517, + 286, + 541 + ], + "type": "text", + "content": " AP. Compared to TransVOD, STPN with SELSA makes a " + }, + { + "bbox": [ + 47, + 517, + 286, + 541 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 47, + 517, + 286, + 541 + ], + "type": "text", + "content": " improvement in AP50." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 555, + 286, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 286, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 286, + 663 + ], + "type": "text", + "content": "Video Instance Segmentation. We compare STPN with state-of-the-art methods in Table 3. Using the M2F-VIS [8] as the baseline segmentation method, STPN improves performance by large margins with " + }, + { + "bbox": [ + 47, + 555, + 286, + 663 + ], + "type": "inline_equation", + "content": "2.9 / 3.6\\%" + }, + { + "bbox": [ + 47, + 555, + 286, + 663 + ], + "type": "text", + "content": " of AP/AP50, respectively. Compared with SOTA methods with similar backbone complexities, for example, Swin-T or ResNet-101, STPN achieves the best accuracy. Specifically, STPN improves the performance of the best competitor VITA [18] by " + }, + { + "bbox": [ + 47, + 555, + 286, + 663 + ], + "type": "inline_equation", + "content": "2.9 / 3.3\\%" + }, + { + "bbox": [ + 47, + 555, + 286, + 663 + ], + "type": "text", + "content": " of AP/AP50, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Visual Object Tracking. Table 4 compares STPN with state-of-the-art methods. We use MixFormer [11] as our baseline. Specifically, when CVT-22k and CVT-1k are used" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 322, + 545, + 479 + ], + "blocks": [ + { + "bbox": [ + 47, + 471, + 286, + 493 + ], + "lines": [ + { + "bbox": [ + 47, + 471, + 286, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 286, + 493 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-art methods for video instance segmentation on the YouTube VIS 2019 dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 322, + 545, + 479 + ], + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 479 + ], + "type": "table", + "html": "
ModelBackboneAOSR0.5SR0.75
SiamRPN++ [29]R10151.761.632.5
SiamR-CNN [50]R10164.972.859.7
PrDiMP [12]R10163.473.854.3
TrDiMP [52]R10167.177.758.3
TREG [10]R10166.877.857.2
TransT [6]R10167.176.860.9
STARK [62]R10168.878.164.1
MixFormer[11]CVT-22K70.780.067.8
MixFormer [11]CVT-1K71.279.965.8
STPNMixFormerCVT-22K71.881.669.0
STPNMixFormerCVT-1K72.781.967.7
", + "image_path": "0aac90a83b307dc136326284c7ad39783a695972cf7438d529adc71e330b8ce5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 487, + 545, + 510 + ], + "lines": [ + { + "bbox": [ + 306, + 487, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 545, + 510 + ], + "type": "text", + "content": "Table 4. Comparison with state-of-the-art tracking methods on the GOT-10k dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 529, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 545, + 576 + ], + "type": "text", + "content": "as the backbone, STPN improves the AO of MixFormer by " + }, + { + "bbox": [ + 306, + 529, + 545, + 576 + ], + "type": "inline_equation", + "content": "+1.1\\%" + }, + { + "bbox": [ + 306, + 529, + 545, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 529, + 545, + 576 + ], + "type": "inline_equation", + "content": "+1.5\\%" + }, + { + "bbox": [ + 306, + 529, + 545, + 576 + ], + "type": "text", + "content": ", respectively. STPN achieves the best performance compared to other SOTA methods on the GOT-10K dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 578, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 545, + 624 + ], + "type": "text", + "content": "These experimental results demonstrate that STPN is easy to generalise to various video understanding tasks and achieves remarkable performance across all three different video tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 634, + 399, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 399, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 399, + 647 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 653, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 545, + 701 + ], + "type": "text", + "content": "We investigate the effect of different designs of our approach on the overall performance. All experiments in §5.3 are conducted on the ImageNet VID dataset for video object detection." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13592" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 211, + 182 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 211, + 182 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 211, + 182 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 211, + 182 + ], + "type": "image", + "image_path": "1742272ceb30c4fa5cacb8a90d74ec157f4b54d88e0bcae83df0fe71bdcd75d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 184, + 193, + 195 + ], + "lines": [ + { + "bbox": [ + 78, + 184, + 193, + 195 + ], + "spans": [ + { + "bbox": [ + 78, + 184, + 193, + 195 + ], + "type": "text", + "content": "(a) STPN with Swin-T backbone" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 214, + 72, + 380, + 182 + ], + "blocks": [ + { + "bbox": [ + 214, + 72, + 380, + 182 + ], + "lines": [ + { + "bbox": [ + 214, + 72, + 380, + 182 + ], + "spans": [ + { + "bbox": [ + 214, + 72, + 380, + 182 + ], + "type": "image", + "image_path": "d35e8a0c5af68a5f91bb35cfefab442cae421062b4d9d594962028686558e1c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 184, + 361, + 195 + ], + "lines": [ + { + "bbox": [ + 247, + 184, + 361, + 195 + ], + "spans": [ + { + "bbox": [ + 247, + 184, + 361, + 195 + ], + "type": "text", + "content": "(b) STPN with Swin-B backbone" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 382, + 75, + 543, + 184 + ], + "blocks": [ + { + "bbox": [ + 382, + 75, + 543, + 184 + ], + "lines": [ + { + "bbox": [ + 382, + 75, + 543, + 184 + ], + "spans": [ + { + "bbox": [ + 382, + 75, + 543, + 184 + ], + "type": "image", + "image_path": "8528e5a397210b5681701b2b60a2805689c6295a707676118c9df7053ee1fd4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 185, + 520, + 195 + ], + "lines": [ + { + "bbox": [ + 422, + 185, + 520, + 195 + ], + "spans": [ + { + "bbox": [ + 422, + 185, + 520, + 195 + ], + "type": "text", + "content": "(c) Comparison with STOA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 243, + 284, + 350 + ], + "blocks": [ + { + "bbox": [ + 129, + 210, + 463, + 222 + ], + "lines": [ + { + "bbox": [ + 129, + 210, + 463, + 222 + ], + "spans": [ + { + "bbox": [ + 129, + 210, + 463, + 222 + ], + "type": "text", + "content": "Figure 4. Comparisons between the transformer-based and the Mixer-based DVP predictors." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 243, + 284, + 350 + ], + "lines": [ + { + "bbox": [ + 51, + 243, + 284, + 350 + ], + "spans": [ + { + "bbox": [ + 51, + 243, + 284, + 350 + ], + "type": "image", + "image_path": "d02caf1694f3ee160a8f9cfa7c658e7b62745af656d84cbee610f8bee087dc76.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 367, + 287, + 388 + ], + "lines": [ + { + "bbox": [ + 47, + 367, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 287, + 388 + ], + "type": "text", + "content": "Figure 5. Comparisons of STPN deep and STPN shallow on transformer encoders with different scales." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": "DVP Predictor Designs. We compare the effects of DVP predictor designs on different baselines. Figure 4 (a) shows the results on the Swin-T backbone. The FasterRCNN baseline achieves " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "80.7\\%" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " AP50, runs at 23.12 FPS, and has 44.90M parameters. Using the transformer-based predictor and the Mixer-based predictor, STPN achieves " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "85.0 / 84.5\\%" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " AP50 and runs at 22.27/22.41 FPS and has 45.01/44.96M parameters, respectively. These results show that the transformer-based predictor can achieve better accuracy, while the Mixer-based predictor has less complexity and faster speed. Experiments on the Swin-B backbone are shown in Figure 4 (b) and the experimental results indicate the same conclusion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 579, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 676 + ], + "type": "text", + "content": "Efficiency of STPN. Compared with other state-of-the-art methods, STPN with both predictor designs actually achieves a significantly improved speed-accuracy trade-off, as shown in Figure 4 (c). These results demonstrate that the pipeline of STPN is very efficient and is not sensitive to different predictor designs. We choose the transformer-based predictor as our default setting since it achieves a slightly higher accuracy with a comparable fast speed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "STPN Shallow v.s. STPN Deep. As shown in the Figure 5, we conduct experiments on Swin-T/S/B/L variants. The" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 240, + 544, + 356 + ], + "blocks": [ + { + "bbox": [ + 309, + 240, + 544, + 356 + ], + "lines": [ + { + "bbox": [ + 309, + 240, + 544, + 356 + ], + "spans": [ + { + "bbox": [ + 309, + 240, + 544, + 356 + ], + "type": "table", + "html": "
(a)S12481632
AP58.759.660.160.660.560.5
(b)K1357911
AP52.653.558.960.660.660.7
FPS22.422.422.322.321.720.5
(c)NP35791113
AP60.260.360.660.660.560.6
FPS22.922.722.321.621.019.9
", + "image_path": "6b4b28ef148b8349a632c9c31aab96145aae58dddd56000d85815369fde5b6c9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 363, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 407 + ], + "type": "text", + "content": "Table 5. Effect of hyper-parameters in STPN: the temporal stride " + }, + { + "bbox": [ + 305, + 363, + 545, + 407 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 363, + 545, + 407 + ], + "type": "text", + "content": ", the number of support frames " + }, + { + "bbox": [ + 305, + 363, + 545, + 407 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 363, + 545, + 407 + ], + "type": "text", + "content": ", and the number of dynamic video prompts " + }, + { + "bbox": [ + 305, + 363, + 545, + 407 + ], + "type": "inline_equation", + "content": "N_P" + }, + { + "bbox": [ + 305, + 363, + 545, + 407 + ], + "type": "text", + "content": ". Different choices of hyper-parameters are listed in rows highlighted in grey." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": "first finding is that both STPN shallow and STPN deep can continuously improve performance on all Swin variants, indicating that STPN has good compatibility with backbone scales. Another finding which contradicts our intuition is that surprisingly STPN shallow achieves better performance than STPN deep on three Swin variants. We believe the reason is that STPN shallow has fewer parameters and thus is easier to optimise in the training process. In contrast, the Swin-L variant is much more complex and thus needs more adjustments in the intermediate transformer layers. By default, we use STPN shallow." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "Support Frame Sampling. We use two hyper-parameters " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " to control the sampling of support frames. Table 5 (a) and Table 5 (b) show the effects of choosing different values of " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": ", respectively. Firstly, we fix the value of " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " as 7, so 15 surrounding support frames are used in total. Then, we vary the temporal stride " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " from 1 to 32. As shown in Table 5 (a), " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "S = 8" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " achieves the best performance. Afterwards, with " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " fixed as 8, we verify the effect of different " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " with respect to both performance (AP) and speed (FPS). Specifically, when " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "K = 7" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": ", STPN achieves a good speed-accuracy trade-off with " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "60.6\\%" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " AP and 22.3" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13593" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 276, + 234 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 276, + 234 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 276, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 276, + 234 + ], + "type": "image", + "image_path": "ada8eb29dcbe02f5dbc9a30def17471d47b3707032bad0dde499e02bbcdd2f7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 246, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 47, + 246, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 287, + 280 + ], + "type": "text", + "content": "Figure 6. Bounding boxes and Grad-CAM visualisations of Faster-RCNN and STPN. Each column shows a video frame from the ImageNet VID validation set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 289, + 276, + 381 + ], + "blocks": [ + { + "bbox": [ + 61, + 289, + 276, + 381 + ], + "lines": [ + { + "bbox": [ + 61, + 289, + 276, + 381 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 276, + 381 + ], + "type": "image", + "image_path": "66c89b35a9f5accb099d5241eb3615b1061a0dc6ba4ddb9912809dc388f7732a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 393, + 286, + 427 + ], + "lines": [ + { + "bbox": [ + 47, + 393, + 286, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 286, + 427 + ], + "type": "text", + "content": "Figure 7. Instance mask visualisations of MinVIS [20] and STPN. Each column shows a video frame from the YouTube VIS dataset. MinVIS fails to generate accurate masks in low-quality frames." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 447, + 270, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 270, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 270, + 459 + ], + "type": "text", + "content": "FPS. Therefore, we choose " + }, + { + "bbox": [ + 47, + 447, + 270, + 459 + ], + "type": "inline_equation", + "content": "S = 8" + }, + { + "bbox": [ + 47, + 447, + 270, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 447, + 270, + 459 + ], + "type": "inline_equation", + "content": "K = 7" + }, + { + "bbox": [ + 47, + 447, + 270, + 459 + ], + "type": "text", + "content": " as default." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "text", + "content": "Number of dynamic video prompts. To explore the effect of the number of dynamic video prompts " + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "inline_equation", + "content": "N_P" + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "text", + "content": ", we show the performance and speed results by varying this number from 3 to 13 in Table 5 (c). The best speed-accuracy trade-off is obtained when " + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "inline_equation", + "content": "N_P" + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "text", + "content": " is 7. In particular, once " + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "inline_equation", + "content": "N_P" + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "text", + "content": " is greater than 7, AP is less affected by the change of " + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "inline_equation", + "content": "N_P" + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "text", + "content": ". Thus, we use 7 as the default value of " + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "inline_equation", + "content": "N_P" + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 563, + 160, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 160, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 160, + 575 + ], + "type": "text", + "content": "5.4. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Grad-CAM [42] is used to visualise the \"concentration\" areas of the FasterRCNN baseline and STPN in some challenging scenarios, for example, motion blur, occlusion, and both. As shown in Figure 6, FasterRCNN either fails to detect the blurry object or generates false positive detections, while STPN can produce accurate detections. In the first column, the squirrel is very blurry because of its fast movement. As a result, FasterRCNN cannot detect the squirrel, and thus the Grad-CAM does not show much attention to the region of the squirrel. On the contrary, STPN successfully detects the squirrel, and Grad-CAM" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 70, + 425, + 129 + ], + "blocks": [ + { + "bbox": [ + 319, + 70, + 425, + 129 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 425, + 129 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 425, + 129 + ], + "type": "image", + "image_path": "9263a3183aed34777fc3a75a304c487996835dbbd2957f0f3045f510fadc61d6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 129, + 401, + 137 + ], + "lines": [ + { + "bbox": [ + 341, + 129, + 401, + 137 + ], + "spans": [ + { + "bbox": [ + 341, + 129, + 401, + 137 + ], + "type": "text", + "content": "(a) Fast-speed objects" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 149, + 545, + 183 + ], + "lines": [ + { + "bbox": [ + 306, + 149, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 306, + 149, + 545, + 183 + ], + "type": "text", + "content": "Figure 8. t-SNE visualisation of the object features produced by FasterRCNN and STPN on (a) fast speed and (b) medium speed objects in the ImageNet VID dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 533, + 129 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 533, + 129 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 533, + 129 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 533, + 129 + ], + "type": "image", + "image_path": "173ffe4c3326581a32b8b833d1a268085b38a726b0cdb71cc6dc263eba87451d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 129, + 517, + 137 + ], + "lines": [ + { + "bbox": [ + 447, + 129, + 517, + 137 + ], + "spans": [ + { + "bbox": [ + 447, + 129, + 517, + 137 + ], + "type": "text", + "content": "(b) Medium-speed objects" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 203, + 527, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 527, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 527, + 216 + ], + "type": "text", + "content": "shows a good concentration map in the squirrel region." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 227, + 546, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 227, + 546, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 227, + 546, + 383 + ], + "type": "text", + "content": "Instance Mask Visualisation. Figure 7 shows the mask segmentation results of the state-of-the-art method MinVIS [20] and the STPN in deteriorated frames. For example, in the first and the second columns, MinVIS fails to generate instance masks for the surfboard and the fox because of motion blur but our STPN can generate accurate instance masks in these frames. The third column shows a case of occlusion. The left deer is occluded by the right one, so the MinVIS fails to generate a mask in the head area of the left dear. Besides, it cannot distinguish the two deer and generate one instance mask. In contrast, STPN works well for generating the mask of the occluded deer and distinguishing the two instances of deer." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 395, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 395, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 545, + 538 + ], + "type": "text", + "content": "t-SNE [48] is a statistical method for visualising high-dimensional data in 2D maps and is usually used to demonstrate the discriminative abilities of different models in classification tasks. We visualise the object features extracted by FasterRCNN and STPN using t-SNE. Figure 8 (a) and (b) show the results on the objects with the fast and medium moving speeds, respectively. We can see that STPN has better clustering results than the FasterRCNN baseline, demonstrating that STPN can improve the discriminative ability of features. Details of how to define different moving speeds and how to generate object features are illustrated in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "To the best of our knowledge, we are the first to explore promoting techniques for robust video feature extraction. We present a neat and unified framework, named Sptio-Temporal Prompting Network (STPN). STPN simplifies the current pipeline for video understanding tasks. Moreover, STPN is easy to generalise to various video tasks. We conduct extensive experiments and report STPN's superior performance on three widely-used video benchmarks: ImageNet VID for video object detection, YouTube for video instance segmentation, and GOT-10K for visual object tracking. We hope our work can inform future research on robust video understanding." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13594" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "text", + "content": "[2] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolitional siamese networks for object tracking. In ECCV, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 287, + 201 + ], + "type": "text", + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 245 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 290 + ], + "type": "text", + "content": "[5] Jianguo Chen, Kenli Li, Qingying Deng, Keqin Li, and S Yu Philip. Distributed deep learning model for intelligent video surveillance systems with edge computing. Industrial Informatics, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "type": "text", + "content": "[6] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Hutchuan Lu. Transformer tracking. In CVPR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 324, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 356 + ], + "type": "text", + "content": "[7] Yihong Chen, Yue Cao, Han Hu, and Liwei Wang. Memory enhanced global-local aggregation for video object detection. In CVPR, 2020. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 357, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 402 + ], + "type": "text", + "content": "[8] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 445 + ], + "type": "text", + "content": "[9] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "text", + "content": "[10] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Target transformed regression for accurate tracking. arXiv preprint arXiv:2104.00403, 2021. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "text", + "content": "[11] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In CVPR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 535 + ], + "type": "text", + "content": "[12] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 567 + ], + "type": "text", + "content": "[13] Jiajun Deng, Yingwei Pan, Ting Yao, Wengang Zhou, Houqiang Li, and Tao Mei. Relation distillation networks for video object detection. In ICCV, 2019. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 612 + ], + "type": "text", + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 646 + ], + "type": "text", + "content": "[15] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. NeurIPS, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 668 + ], + "type": "text", + "content": "[16] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[17] Lu He, Qianyu Zhou, Xiangtai Li, Li Niu, Guangliang Cheng, Xiao Li, Wenxuan Liu, Yunhai Tong, Lizhuang Ma, and Liqing Zhang. End-to-end video object detection with spatial-temporal transformers. In ACMMM, 2021. 2, 5, 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. Vita: Video instance segmentation via object token association. In NeurIPS, 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[19] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In CVPR, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[20] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 2, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[21] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE transactions on pattern analysis and machine intelligence, 43(5):1562-1577, 2019. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "text", + "content": "[22] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 2021. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 254, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 296 + ], + "type": "text", + "content": "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "text", + "content": "[24] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "text", + "content": "[25] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "text", + "content": "[26] Hamed Kiani Galoogahi, Ashton Fagg, and Simon Lucey. Learning background-aware correlation filters for visual tracking. In ICCV, 2017. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 411, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 476 + ], + "type": "text", + "content": "[27] Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "text", + "content": "[28] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "text", + "content": "[29] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In CVPR, 2019. 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[30] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In CVPR, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 579, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 622 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[32] Liangkai Liu, Sidi Lu, Ren Zhong, Baofu Wu, Yongtao Yao, Qingyang Zhang, and Weisong Shi. Computing systems for autonomous driving: State of the art and challenges. IEEE Internet of Things Journal, 8(8):6469-6486, 2020. 1" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 3, 5" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13595" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[34] Rashmika Nawaratne, Damminda Alahakoon, Daswin De Silva, and Xinghuo Yu. Spatiotemporal anomaly detection using deep learning for real-time video surveillance. Industrial Informatics, 16(1):393-402, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "type": "text", + "content": "[35] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 287, + 208 + ], + "type": "text", + "content": "[36] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning for action recognition. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 287, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 265 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 266, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 287, + 300 + ], + "type": "text", + "content": "[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 302, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 302, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 287, + 334 + ], + "type": "text", + "content": "[39] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 336, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 336, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 336, + 287, + 369 + ], + "type": "text", + "content": "[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "text", + "content": "[41] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 428, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 287, + 472 + ], + "type": "text", + "content": "[42] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "type": "text", + "content": "[43] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Mamba: Multi-level aggregation via memory bank for video object detection. In AAAI, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 509, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 541 + ], + "type": "text", + "content": "[44] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Efficient one-stage video object detection by exploiting temporal consistency. In ECCV, 2022. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "type": "text", + "content": "[45] Guanxiong Sun, Yang Hua, Guosheng Hu, and Neil Robertson. Tdvit: Temporal dilated video transformer for dense video tasks. In ECCV, 2022. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "text", + "content": "[46] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. In NeurIPS, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 677 + ], + "type": "text", + "content": "[47] Jack Valmadre, Luca Bertinetto, João Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In CVPR, 2017. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[48] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[50] Paul Voigtlaender, Jonathon Luiten, Philip H. S. Torr, and Bastian Leibe. Siam R-CNN: visual tracking by re-detection. In CVPR, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 186 + ], + "type": "text", + "content": "[51] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 187, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 187, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 187, + 545, + 220 + ], + "type": "text", + "content": "[52] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In CVPR, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "text", + "content": "[53] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip HS Torr. Fast online object tracking and segmentation: A unifying approach. In CVPR, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 256, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 277 + ], + "type": "text", + "content": "[54] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 279, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 279, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 279, + 545, + 323 + ], + "type": "text", + "content": "[55] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 369 + ], + "type": "text", + "content": "[56] Fei Wu, Qingzhong Wang, Jiang Bian, Ning Ding, Feixiang Lu, Jun Cheng, Dejing Dou, and Haoyi Xiong. A survey on video action recognition in sports: Datasets, methods and applications. Multimedia, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 370, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 370, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 545, + 403 + ], + "type": "text", + "content": "[57] Haiping Wu, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Sequence level semantics aggregation for video object detection. In ICCV, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "text", + "content": "[58] Haiping Wu, Bin Xiao, Noel C. F. Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In ICCV, 2021. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 439, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 439, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 439, + 545, + 472 + ], + "type": "text", + "content": "[59] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "type": "text", + "content": "[60] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, 2022. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 508, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 540 + ], + "type": "text", + "content": "[61] Zhujun Xu, Emir Hrustic, and Damien Vivet. Centernet heatmap propagation for real-time video object detection. In ECCV, 2020. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 543, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 575 + ], + "type": "text", + "content": "[62] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 1, 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 577, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 598 + ], + "type": "text", + "content": "[63] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 643 + ], + "type": "text", + "content": "[64] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[65] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Unified vision and language prompt learning. arXiv preprint arXiv:2210.07225, 2022. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[66] Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object detection. In ICCV, 2017. 5" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13596" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 107 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 107 + ], + "type": "text", + "content": "[67] Zheng Zhu, Qiang Wang, Bo Li, Wei Wu, Junjie Yan, and Weiming Hu. Distractor-aware siamese networks for visual object tracking. In ECCV, 2018. 2" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13597" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/dbd672e5-b6e6-48ea-861d-97b4378c8ca3_content_list.json b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/dbd672e5-b6e6-48ea-861d-97b4378c8ca3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc64cbbacb8b5ad28d74c9653426e2f9b1a7f170 --- /dev/null +++ b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/dbd672e5-b6e6-48ea-861d-97b4378c8ca3_content_list.json @@ -0,0 +1,1647 @@ +[ + { + "type": "text", + "text": "Spectral Graphormer: Spectral Graph-based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images", + "text_level": 1, + "bbox": [ + 114, + 130, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tze Ho Elden Tse $^{1,2*}$ Franziska Mueller $^{1}$ Zhengyang Shen $^{1}$ Danhang Tang $^{1}$ Thabo Beeler $^{1}$ Mingsong Dou $^{1}$ Yinda Zhang $^{1}$ Sasa Petrovic $^{1}$ Hyung Jin Chang $^{2}$ Jonathan Taylor $^{1}$ Bardia Doosti $^{1}$ $^{1}$ Google $^{2}$ University of Birmingham", + "bbox": [ + 84, + 202, + 903, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel transformer-based framework that reconstructs two high fidelity hands from multi-view RGB images. Unlike existing hand pose estimation methods, where one typically trains a deep network to regress hand model parameters from single RGB image, we consider a more challenging problem setting where we directly regress the absolute root poses of two-hands with extended forearm at high resolution from egocentric view. As existing datasets are either infeasible for egocentric viewpoints or lack background variations, we create a large-scale synthetic dataset with diverse scenarios and collect a real dataset from multi-calibrated camera setup to verify our proposed multi-view image feature fusion strategy. To make the reconstruction physically plausible, we propose two strategies: (i) a coarse-to-fine spectral graph convolution decoder to smoothen the meshes during upsampling and (ii) an optimisation-based refinement stage at inference to prevent self-penetrations. Through extensive quantitative and qualitative evaluations, we show that our framework is able to produce realistic two-hand reconstructions and demonstrate the generalisation of synthetic-trained models to real data, as well as real-time AR/VR applications.", + "bbox": [ + 75, + 323, + 473, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 684, + 209, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D hand pose estimation is a fundamental problem in various downstream applications including augmented and virtual reality (AR/VR) [19, 20, 25, 26, 47, 58, 63]. Research efforts in designing 3D hand or hand-object reconstruction networks have shown great effectiveness for single-hand pose estimation [15, 22, 27, 59, 60, 66]. However, two-hand pose estimation has received relatively little attention. In this paper, as illustrated in Fig. 1, we focus on high fidelity two-hand reconstruction from multi-view RGB images.", + "bbox": [ + 75, + 710, + 468, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the availability of the InterHand2.6M dataset [46], various methods have been recently proposed. More recent", + "bbox": [ + 76, + 845, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9e1a2ba69ca581c8b2f6bfcbf10cc1e94b1a459c3acf9f157c7b0b419663fdb.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 501, + 301, + 633, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0579d14bbeee00972e7d57d7879038422fec69acb90e0d9c6b86dd095ac41e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 404, + 633, + 505 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc98eeb63b063af0c96f1f8761486c78d7ee7b75178cd75cfe6472be1737e082.jpg", + "image_caption": [ + "Front view" + ], + "image_footnote": [], + "bbox": [ + 638, + 301, + 759, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/99ec366cdc10acb3392148a5249e44be9efc4941542eae2061db8dead4202421.jpg", + "image_caption": [ + "Rotated view" + ], + "image_footnote": [], + "bbox": [ + 638, + 405, + 761, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e653ccdf381ab53d91c7e758fe3a13819461dcb55b3706447a42bb2856be0c27.jpg", + "image_caption": [ + "Zoomed view" + ], + "image_footnote": [], + "bbox": [ + 764, + 301, + 885, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1920d8931781f6238e0dd968166ee32da01ec6039b1ed5953d4f6736f8c5c95d.jpg", + "image_caption": [ + "Figure 1: Our method jointly reconstructs high fidelity two-hand meshes from multi-view RGB image. The input images shown here are from our synthetic dataset, which contains challenging video sequences of two hands rendered into egocentric views." + ], + "image_footnote": [], + "bbox": [ + 764, + 388, + 883, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods focused on alleviating the self-similarity problem between interacting hands by exploiting hand part segmentation probability [18], joint visibility [33], cascaded refinement modules [65] or keypoints using Transformer [24]. Although they perform well on complex configurations, their dependence on root joint alignment does not provide absolute root pose recovery when applied to multi-view scenarios, which is especially crucial for interactions in VR.", + "bbox": [ + 496, + 611, + 893, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Moreover, many immersive AR/VR applications require accurate estimation of two hands including extended forearms as it allows for a more realistic and accurate representation of hand movements and gestures. By including the forearm, the orientation and movement of the hand in relation to the arm can provide important contextual information for the user's actions in the virtual environment. Additionally, it can reduce errors and improve the stability of the tracking system, which is important for maintaining immersion and avoiding disorientation in AR/VR applications. However, as in many other areas of computer vision, there", + "bbox": [ + 496, + 734, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during an internship at Google.", + "bbox": [ + 94, + 886, + 377, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14666", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is currently no suitable dataset for a typically supervised deep learning setup. Specifically, existing datasets are either infeasible for egocentric views [21] or lack variations in background and lighting conditions [46].", + "bbox": [ + 78, + 90, + 467, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a spectral graph-based transformer architecture that can reconstruct high resolution two-hand meshes with extended forearms from multi-view RGB images. However, directly extending state-of-the-art methods [39, 40] from single-view to multi-view setting is nontrivial as they consume a substantial number of parameters and are computationally expensive. As pointed out by Cho et al. [8], the current encoder-based transformers [39, 40] have overlooked the importance of efficient token design. In order to retain spatial information in image features and avoid passing redundant global image features to the transformer encoder, we propose a soft-attention-based multiview image feature fusion strategy to obtain region-specific features. Further, as a hand mesh in essence is graph-structured data, we incorporate properties of graph Laplacian from spectral graph theory [10] into the design of our transformer-based network architecture. To obtain realistic hand reconstructions, we leverage a hierarchical graph decoder and propose an optimisation-based refinement stage during inference to prevent self-penetrations. We show that each technical component above contributes meaningfully in our ablation study.", + "bbox": [ + 80, + 152, + 467, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides the expensive computational cost, an additional challenge is the lack of high fidelity 3D mesh ground truth for training such a model. As manually annotating 3D hand meshes on real data is extremely laborious and expensive, we create a large-scale synthetic multi-view dataset containing realistic hand motions from egocentric camera viewpoints rendered under a large variation of background and illumination. To generalise and evaluate on real-world data, we further collected real data from a precisely calibrated multi-view studio with 18 number of high-resolution cameras. We use an automatic approach for registering meshes and obtain ground truth hand poses. Our proposed method trained on both datasets shows robust performance on challenging scenarios and can serve as a strong baseline.", + "bbox": [ + 80, + 484, + 467, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are the following:", + "bbox": [ + 98, + 696, + 334, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We propose a novel end-to-end trainable spectral graph-based transformer for high fidelity two-hand reconstruction from multi-view RGB image.", + "bbox": [ + 89, + 719, + 467, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. We design an efficient soft attention-based multi-view image feature fusion in which the resulting image features are region-specific to segmented hand mesh. We further demonstrate a minimal reduction of $35\\%$ in the model size with this approach.", + "bbox": [ + 89, + 768, + 467, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. We introduce an optimisation-based method to refine physically-implausible meshes at inference.", + "bbox": [ + 89, + 849, + 467, + 880 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4. We create a large-scale synthetic multi-view dataset", + "bbox": [ + 88, + 885, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with high resolution 3D hand meshes and collect real dataset to verify our proposed method.", + "bbox": [ + 535, + 90, + 888, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 145, + 638, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work tackles the problem of two-hand reconstruction from multi-view images. We first review the literature on Hand reconstruction from RGB images. Then, we focus on the line that leverages Transformer on human body/hand reconstruction. Finally, we provide a brief review on existing Hand datasets.", + "bbox": [ + 503, + 170, + 888, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-Hand Reconstruction. Many approaches have already been proposed for hand pose estimation from either RGB images or depth maps. Here we focus mainly on works that reconstruct full hand meshes. As most earlier methods do not have access to suitable datasets which contain dense 3D ground truths, the parametric hand model MANO [52] provides an alternative as its input parameters can be directly regressed and it contains prior 3D information about hand geometry. Hence, discriminative approaches [2, 5, 27] which directly predict input hand pose and shape parameters from single images have been a popular approach. [45] learns the pose and shape corrective parameters to reconstruct high-fidelity hand meshes with multi-view depth images. Concurrently, non-parametric methods [22, 36] leveraging graph convolutions demonstrate the ability to directly regress 3D hand meshes. Recently, there is an increasing interest in representing object shapes by learning an implicit function [1, 7, 42]. By extending NASA [13], [31] is a keypoint-driven implicit hand representation and [11] disentangles shape, color and pose representations.", + "bbox": [ + 503, + 263, + 888, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to the above mentioned approaches, we propose in this work a non-parametric method that can be extended to multi-view settings and reconstruct high-fidelity hand meshes with extended forearms. Further, our approach is not restricted to a single subject, which is the primary constraint of [45]. Recent work on multi-view fusion leverage epipolar geometry [29, 50] to aggregate estimated 2D heatmaps from different views or Feature Transform Layers (FTL) [26, 51] to learn camera geometry-aware latent features. However, these methods require known camera parameters and the feature matching process based on feature similarity is sensitive to occlusion. Our work introduces a soft-attention mechanism coupled with transformer tokenisation which is robust to occlusion and aggregates multiview image features efficiently.", + "bbox": [ + 503, + 566, + 888, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Two-Hand Pose Estimation. Pose estimation of interacting hands can be broadly classified as discriminative and generative (or hybrid) approaches. Early methods typically follow hybrid approaches by leveraging visual cues detected by discriminative methods, followed by model fitting. [3, 61] apply collision optimisation terms and physical modelling on detected fingertips. [25, 26, 47, 63] ex", + "bbox": [ + 503, + 795, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14667", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/29fcfb6ac8d7b1a52209c85878503dce4d6d8eb9bcaf2f870b84fd5c9f5f7d65.jpg", + "image_caption": [ + "Figure 2: A schematic illustration of our framework. Given multi-view RGB images, we extract volumetric features $\\mathbf{f}$ with a shared CNN backbone. The soft-attention fusion block generates the attention mask $\\mathcal{M}$ and finer image features $\\mathbf{f}'$ through multiple upsampling and convolution blocks. Region-specific features $\\mathbf{f}_r$ are computed by first aggregating $\\mathbf{f}'$ along the feature channel dimension via the attention mask $\\mathcal{M}$ , followed by a max-pooling operation across multi-view images to focus on useful features. Then, we apply mesh segmentation via spectral clustering on template hand meshes and uniformly subsample them to obtain coarse meshes. We perform position encoding by concatenating coarse template meshes to the corresponding region-specific features $\\mathbf{f}_r$ , i.e. matching colored features to mesh segments. Finally, our multi-layer transformer encoder takes the resulting features as input and outputs a coarse mesh representation $\\mathbf{f}_c$ which is then decoded by a spectral graph decoder to produce the final two-hand meshes at target resolution. Here, each hand contains 4023 vertices." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 872, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tract image features or keypoints from RGB or depth images and fit hand models with physical constraints. Recent largescale datasets of interacting hands [46] have led to the emergence of fully discriminative methods [24,33,38,53,65] that jointly estimate the 3D joint locations or hand model parameters from a single RGB image. In this work, we target to reconstruct two hands at high resolution while keeping the model complexity low.", + "bbox": [ + 75, + 415, + 470, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer in 3D Vision. Transformer-based architectures have been gaining increasing popularity in the vision community. Here we focus on methods that reconstruct the human body or hands from RGB images and refer readers to [32] for a detailed survey. In a closely related work, [39] uses cascaded Transformer encoders to reconstruct the human body and hands from a single RGB image and achieves state-of-the-art performance. [40] extends [39] with graph convolutions along the Transformer encoder. [8] improves the efficiency of [39] by disentangling image encoding and mesh estimation via an encoder-decoder architecture. [24] extends \"Detection Transformer\" [6] with hand-object pose estimations. While these works are targeted at a single image and their extension to the multi-view setting is nontrivial due to large number of learnable parameters, our architecture is designed to efficiently reconstruct two hands at high resolution.", + "bbox": [ + 75, + 544, + 468, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hand Pose Datasets. The success of discriminative methods depend on the availability and variability of hands interaction datasets. [46] introduces the large-scale dataset InterHand2.6M with closely interacting hand motions using a semi-automatic annotation process with multiple high-resolution RGB cameras. However, it contains minimal", + "bbox": [ + 75, + 809, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "background and lighting variation. FreiHAND [67] includes wider background variation but the dataset is limited to a single hand and third-person views. Therefore, the above datasets are not suitable for the egocentric two-hand reconstruction task. The available egocentric datasets are limited by either visible markers [21] or constrained lab environments [37]. This motivates us to create a large-scale egocentric synthetic dataset with improved environment and lighting variations. To validate our proposed multi-view fusion strategy, we further collect a real dataset with more challenging camera viewing angles. Altogether, both synthetic and real datasets contain diverse egocentric multi-view and -frame data points with multiple subjects. Our architecture demonstrates promising performance on both datasets and constitutes a strong baseline.", + "bbox": [ + 496, + 415, + 893, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 661, + 591, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our architecture first passes $N$ number of multi-view RGB input images $\\mathbf{x} \\in \\mathbb{R}^{N \\times 224 \\times 224 \\times 3}$ to a shared CNN backbone to extract volumetric features $\\mathbf{f} \\in \\mathbb{R}^{N \\times 7 \\times 7 \\times 2048}$ , i.e. features before the global average pooling layer for ResNet [28]. The volumetric features are then fed into a soft attention-based multi-view feature encoder and output $K$ region-specific features $\\mathbf{f}_r \\in \\mathbb{R}^{K \\times C}$ where $C$ refers to the feature channel size. The Transformer encoder takes $\\mathbf{f}_r$ together with template hand meshes $\\mathbf{m}' \\in \\mathbb{R}^{V' \\times 3}$ and outputs a coarse mesh representation $\\mathbf{f}_c \\in \\mathbb{R}^{V' \\times F}$ . Finally, the spectral graph decoder generates hand meshes $\\mathbf{m} \\in \\mathbb{R}^{V \\times 3}$ by upsampling on $\\mathbf{f}_c$ where $V \\gg V'$ . With slight abuse of notation, $\\mathbf{m}$ can either be two-hand or single-hand depending on the application.", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14668", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we detail the multi-view image feature encoder in Section 3.2, the spectral graph convolution decoder in Section 3.3 and the loss used for training in Section 3.4. We also present an optimisation-based refinement procedure at inference time in Section 3.5.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Notations", + "text_level": 1, + "bbox": [ + 76, + 178, + 187, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A 3D mesh can be represented as an undirected graph $\\mathcal{G} = (\\mathcal{V},\\mathcal{E},\\mathbf{A})$ , where $\\mathcal{V}$ is a node set and $\\mathcal{E}$ is an edge set. An adjacency matrix $\\mathbf{A} \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}$ encodes information of pairwise relations between nodes. A degree matrix $\\mathbf{D} \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}$ is a diagonal matrix whose diagonal element $\\mathbf{D}_{ii} = \\sum_{j} \\mathbf{A}_{ij}$ refers to the degree value of each node. An essential operator in spectral graph theory [10] is the graph Laplacian $\\mathbf{L}$ , whose definition is $\\mathbf{L} = \\mathbf{D} - \\mathbf{A}$ , and $\\mathbf{L} = \\mathbf{U} \\boldsymbol{\\Lambda} \\mathbf{U}^T$ where the graph Laplacian can be diagonalised by the Fourier basis $\\mathbf{U} = [\\mathbf{u}_1, \\dots, \\mathbf{u}_{|\\mathcal{V}|}] \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}$ and $\\boldsymbol{\\Lambda} = \\text{diag}([\\lambda_1, \\dots, \\lambda_{|\\mathcal{V}|}]) \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}$ where $\\{\\mathbf{u}_i\\}_{i=1}^{|\\mathcal{V}|}$ are the eigenvectors and $\\{\\lambda_i\\}_{i=1}^{|\\mathcal{V}|}$ are the non-negative eigenvalues of graph Laplacian ( $0 = \\lambda_1 \\leq \\dots \\leq \\lambda_{|\\mathcal{V}|}$ ).", + "bbox": [ + 75, + 200, + 470, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multi-View Image Feature Encoder", + "text_level": 1, + "bbox": [ + 76, + 417, + 385, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extend transformer-based architecture to multi-view settings and prevent concatenating global image features in an overly-duplicated way as in [39], we propose a simple soft-attention fusion strategy to better aggregate features across multiple views and attend to different hand parts via mesh segmentation. The resulting $K$ region-specific features $\\mathbf{f}_r$ defined by spectral clustering are fed to our Transformer encoder to obtain a coarse mesh representation $\\mathbf{f}_c$ .", + "bbox": [ + 73, + 440, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Soft-Attention Fusion. Given volumetric features $\\mathbf{f}$ , we do not apply any pooling operations to avoid losing spatial information as shown by our experiments (see Table 3). Instead, we propose to aggregate multi-view features with a soft-attention mask. We first obtain a finer representation of $\\mathbf{f}$ , denoted as $\\mathbf{f}' \\in \\mathbb{R}^{N \\times (H \\times W) \\times C}$ , by feeding it through two blocks each comprised of 2D upsampling with bilinear interpolation, $3 \\times 3$ convolution layers, batch-normalisation [30] and ReLU. The soft-attention mask $\\mathcal{M} \\in \\mathbb{R}^{N \\times (H \\times W) \\times K}$ is obtained by applying 1) $K$ $1 \\times 1$ convolutional filters to reduce the feature channel of $\\mathbf{f}'$ to $K$ and 2) spatial soft arg-max which determines the image-space point of maximal activation in each $C$ . At this stage, we can compute per-frame features $\\mathbf{f}'' \\in \\mathbb{R}^{N \\times K \\times C}$ by $\\mathbf{f}'' = \\mathcal{M}^{\\mathrm{T}}\\mathbf{f}'$ . We finally obtain region-specific features $\\mathbf{f}_r \\in \\mathbb{R}^{K \\times C}$ by max-pooling along the second dimension of $\\mathbf{f}''$ . There are two intuitions to this design: 1) the corresponding feature with higher attention weight contributes more to the final feature representation spatially and 2) max-pooling allows efficient feature selection across multiple views and it does not overfit to any multi-camera configurations as CNN filters are shared across all views.", + "bbox": [ + 75, + 568, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mesh Segmentation via Spectral Clustering. Inspired by [41], we perform spectral clustering to obtain a 3D mesh segmentation. Different from [41], we apply the eigendecomposition of the graph Laplacian $\\mathbf{L}$ instead of the affinity matrix (which encodes pairwise point affinities with exponential kernel), followed by $k$ -means clustering into $K$ clusters. With this approach, one can efficiently segment any template model mesh without manual effort. In addition, as shown in Fig. 3, the segmented mesh does not exhibit clear-cut boundaries, and certain clusters are scattered throughout the entire surface. This is in spirit similar to mask vertex modeling occlusions in [39] by encouraging the transformer to consider other relevant vertex queries.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer Encoder. We now concatenate the $C$ -dimensional region-specific image features to the corresponding $K$ clusters of the segmented template hand mesh. The resulting features $\\mathbf{f}_t \\in \\mathbb{R}^{V' \\times (C + 3)^1}$ are fed into a multi-layer transformer encoder with progressive dimensionality reduction as described in [39]. The output is a coarse mesh representation $\\mathbf{f}_c \\in \\mathbb{R}^{V' \\times F}$ , where $F = (C + 3)/2^n$ with $n$ layers of transformer encoder.", + "bbox": [ + 496, + 294, + 890, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretical Motivations. Since there do not exist explicit coordinate systems as in grid graphs, aligning nodes of highly irregular graphs in nature is a non-trivial problem. Recent studies [16, 17, 35] attempted to encode positional information by leveraging the spectral domain in a way that nearby nodes have similar values and distant nodes have different values. This can be achieved as the eigenvectors of the graph Laplacian can be interpreted as the generalised concepts of sinusoidal functions of positional encoding in Transformers [62]. For instance with any given graph, nodes close to each other can be assigned values of similar positional features as the smaller the eigenvalue of the graph Laplacian $\\lambda_{i}$ (closer to 0) the smoother the coordinates of the corresponding eigenvector $\\mathbf{u}_{i}$ .", + "bbox": [ + 496, + 422, + 892, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Spectral Graph Decoder", + "text_level": 1, + "bbox": [ + 500, + 646, + 723, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We find that simply relying on fully-connected layers to upsample meshes to target resolution is insufficient as this process introduces instability and disruption to the mesh, even with mesh regularisation loss terms [27] (see Fig. 4). Therefore, we propose to couple the fully-connected layers for upsampling with spectral filtering as meshes can be treated as graph signals $\\mathbf{f}_c = (f_1,\\dots ,f_{V'})\\in \\mathbb{R}^{V'\\times F}$ , i.e. $V^{\\prime}$ vertices with $F$ -dimensional features for batch size of 1.", + "bbox": [ + 496, + 670, + 890, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spectral Filtering. As spectral graph convolution can be defined via point-wise products in the transformed Fourier space, graph signals $\\mathbf{f}_c$ filtered by $g_{\\theta}$ can then be expressed as $\\mathbf{U}g_{\\theta}(\\boldsymbol {\\Lambda})\\mathbf{U}^{T}\\mathbf{f}_{c}$ where $g_{\\theta}(\\boldsymbol {\\Lambda})$ is a diagonal matrix with Fourier coefficients. To ensure that the spectral filter corre", + "bbox": [ + 496, + 797, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "${}^{1} + 3$ refers to the 3D positions of mesh template.", + "bbox": [ + 517, + 886, + 777, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14669", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d4fe74583be375d47f5db27e8e4eca094a30c92d8beea57350fd175229e2782.jpg", + "image_caption": [ + "Figure 3: Illustration of mesh segmentation via spectral clustering. Here, we have chosen $K = 7$ clusters and each has its own color for a single right-handed template mesh. As shown, there are no rigid boundaries across the mesh. To prepare inputs $\\mathbf{f}_t$ to transformer, we decrease the size of this mesh (similarly for left hand) uniformly by a factor of 10 through subsampling, i.e. $V' = V / 10$ and concatenate with the corresponding region-specific features $\\mathbf{f}_r$ ." + ], + "image_footnote": [], + "bbox": [ + 122, + 84, + 433, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sponds to a meaningful convolution on the graph, a natural solution is to parameterise based on the eigenvalues of the Laplacian. Here we have used the Chebyshev polynomial parametrisation of $g_{\\theta}(\\mathbf{L})$ for fast computation and readers are referred to [12] for more details.", + "bbox": [ + 75, + 369, + 470, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. Similar to [9, 22], our decoder is based on a hierarchical architecture where the mesh is recovered by using fully-connected and graph convolution layers for upsampling. We followed [12, 22] to pre-computed coarse graphs using [14] and used the third-order polynomial in the Laplacian. Given the coarse mesh representation $\\mathbf{f}_c\\in \\mathbb{R}^{V'\\times F}$ , the key idea here is to smooth out the $F$ -dimensional values after upsampling on $V^{\\prime}$ . Note that the fundamental difference between this approach and the one in [9, 22] is that they have a hierarchical architecture on both $V^{\\prime}$ and $F$ dimensions, we apply only spectral filtering and keep $F = 3$ constant which massively reduces the model size while maintaining performance (see Table 1).", + "bbox": [ + 75, + 452, + 472, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Discussions.** Recall that in Section 3.2, we leverage the properties of Laplacian eigenvectors to perform spectral clustering. Here we can interpret from a signal processing perspective where the Laplacian eigenvectors define signals that vary smoothly across the graph, with the smoothest signals indicating the coarse community structure of the mesh.", + "bbox": [ + 75, + 654, + 470, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 76, + 753, + 181, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [39, 40], our model can be trained end-to-end with L1 losses on 3D mesh vertices and 2D re-projection using the predicted camera parameters to improve image-mesh alignment. In addition, we apply an edge length regularisation $\\mathcal{L}_{edge}$ [27] to encourage smoothness of the mesh:", + "bbox": [ + 75, + 776, + 470, + 853 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e d g e}} (\\mathbf {m}) = \\frac {1}{| \\mathcal {E} _ {L} |} \\sum_ {l \\in \\mathcal {E} _ {L}} \\left| l ^ {2} - \\mu \\left(\\mathcal {E} _ {L} ^ {2}\\right) \\right|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 859, + 468, + 897 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{E}_L$ refers to the set of edge lengths, defined as the L2 norms of all edges and $\\mu (\\mathcal{E}_L^2)$ is the average of the squared edge lengths.", + "bbox": [ + 498, + 90, + 893, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Mesh Refinement at Inference", + "text_level": 1, + "bbox": [ + 500, + 143, + 767, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are two main approaches in the literature for producing realistic mesh reconstruction: learning-based and optimisation-based methods. Learning-based methods [27, 45] approach the problem by proposing various repulsive losses that penalise penetration during training. However, their generalisation ability to other meshes is limited due to the need for mesh-specific pre-computation. In particular, they require manual selection of areas of interest (i.e. fingertips and palm) which does not prevent other forms of self-penetrations, such as finger-finger. More importantly, the presence of interpenetration at test time shows that the model is not able to learn the physical rule implicitly [27, 59]. On the other hand, recent optimisation-based methods [23, 60] leverage contact maps to refine meshes at inference. However, they rely heavily on accurate contact map estimations and are sensitive to initialisation as contact optimisation is local.", + "bbox": [ + 496, + 167, + 893, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposed Method. To avoid penetrations, we extend the repulsion loss from [27] into an optimisation-based strategy which does not require any form of pre-computation and is more generalisable to other meshes. To identify hand vertices that contribute to collision, we follow [27,43] and cast rays from each vertex and count the number of surface intersections. If the number is odd, it indicates penetration. After obtaining the collision mask $\\mathcal{M}_C$ , we compute the nearest point in the source mesh $\\mathbf{m}$ with respect to the set of collision points. If the point and its nearest corresponding point have a different normal, we compute their distance as loss. We minimise this collision loss $\\mathcal{L}_{collision}$ with as-rigid-as-possible (ARAP) [56] regularisation on the mesh shape:", + "bbox": [ + 496, + 430, + 893, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o l l i s i o n}} (\\mathbf {m}, \\mathcal {M} _ {C}) = \\sum_ {v \\in \\mathcal {V}} \\mathcal {M} _ {C} \\cdot d (v, \\mathcal {V}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 635, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{M}_C = \\mathbb{1}_{v\\in \\mathrm{Int}(\\mathcal{V})}$ indicates which vertices belong to the interior of the mesh and $d(v,\\mathcal{V}) = \\min_{v'\\in \\mathcal{V}}\\| v - v'\\| _2$ denotes distances from point $v^{\\prime}$ to set $\\mathcal{V}$ . We show that our method is able to remove self-penetration in Fig. 5 and 9.", + "bbox": [ + 496, + 675, + 893, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 748, + 633, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We implement our method in TensorFlow. We train all parts of the network simultaneously with the Adam optimiser [34] using a learning rate of $10^{-4}$ when training on the synthetic dataset and $10^{-5}$ when fine-tuning on the real dataset. We use ResNet [28] pre-trained on ImageNet [54] for our CNN backbone. For computing region-specific features $\\mathbf{f}_r$ , we set $K = 7$ and $C = 256$ .", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14670", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2a1860447dab3f115e4823d25edd80ff279df4fcbc0119f8b545c8aebde37774.jpg", + "image_caption": [ + "Figure 4: Qualitative comparison with METRO [39]. We show that relying on fully-connected layers to upsample meshes is inadequate for high-resolution mesh reconstruction. In contrast, our spectral graph decoder can accurately capture intricate surface features like nails." + ], + "image_footnote": [], + "bbox": [ + 112, + 84, + 450, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/acfbc6247fac97551120a7a16e353e48a5b2e870a763664886c74b713191ad72.jpg", + "image_caption": [ + "Figure 5: Qualitative example of mesh refinement at inference. More examples are available in Fig. 9." + ], + "image_footnote": [], + "bbox": [ + 99, + 393, + 450, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic Dataset Creation. A large-scale multi-view egocentric dataset with challenging interacting hand motion is required to train our pipeline. However, existing datasets are either infeasible for egocentric views or lack variations in terms of background and lighting conditions, which are crucial for AR/VR applications. In addition, the ground-truth mesh in [46] contains $5mm$ fitting error which is sufficient to cause inaccuracy in hand tracking or misalignment when interacting with virtual objects. Therefore, it is not suitable to validate our approach. To this end, we create a new large-scale synthetic egocentric dataset with two high-fidelity 3D hand meshes. In particular, we purchase commercial 3D hand models for both left and right hands. Each of them contains 4023 vertices and 4008 quad faces. By dividing each quad into two triangles, they can be decomposed into 8016 triangular faces. The size and textures of the hand", + "bbox": [ + 75, + 659, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/69bb32fe51f1cfc2398bfb003159c040f7a2dd46a8064fdceb339e33bcd5c9e3.jpg", + "image_caption": [ + "Figure 6: Qualitative examples of our synthetic dataset. The scenes are rendered with large variations in lighting, texture and background." + ], + "image_footnote": [], + "bbox": [ + 503, + 89, + 890, + 203 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models can vary, and it can also be rigged with 3D joint locations. We apply photorealistic textures as well as natural lighting using High-Dynamic-Range (HDR) images. To create realistic hand motions, we generate 100 poses and interpolate between pairs of these poses randomly over 1000 frame sequences, i.e. pose pair is swapped every 10 frames. Then, we render hands onto 480 4K backgrounds using the Cycles renderer [49]. Our synthetic dataset comprises 1M data points. Each data point consists of 3D annotations for two hands, rendered in two egocentric views to simulate a customised camera headset scenario. We divide the entire dataset into 4 groups and use $25\\%$ randomly for testing and provide additional image examples in Fig. 6.", + "bbox": [ + 496, + 267, + 893, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real Dataset Collection. Our real hand data is captured from a multi-view stereo system with 18 synchronised Z-Cams. We build a NeRF-based [4, 48] reconstruction pipeline that simultaneously reconstructs and disentangles the foreground and the background, from which a meshing module extracts hand meshes as ground-truth target mesh. We then apply a mesh and tetrahedral registration approach [55] that registers a template hand model [64] to get a well-registered mesh for each reconstruction.", + "bbox": [ + 496, + 470, + 893, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare with the state-of-the-art transformer-based architecture METRO [39] that is explicitly designed for human body mesh reconstruction. We follow their official implementation and use 3 transformer encoder layers. We attempted to include other strong baselines to compare with. However, given our challenging multi-view settings on reconstructing high resolution two-hand meshes, existing methods are constrained by either 1) relying on intermediate 3D pose supervision [9,44] or 2) directly regressing a low resolution parametric hand model [2,5,27,59,67]. In addition, the direct extension of METRO [39], Mesh Graphormer [40], is infeasible to extend to multi-view settings as they tokenised volumetric features which increases the computational complexity of each transformer layer quadratically [8]. Therefore, by considering that the performance gain is minor, we pick METRO [39] as a strong baseline and study in-depth. We provide more details about the baselines in the supplementary materials.", + "bbox": [ + 496, + 613, + 895, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14671", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/29679ff5c7576c2c51a093ad0d78c522043e4529dd55b2c86b9ec5d5103bcdfc.jpg", + "table_caption": [ + "Table 1: Error rates on our synthetic dataset. Parametric baseline is trained and tested on right hand only. Our proposed method achieves strong performance with less than half the METRO model size." + ], + "table_footnote": [], + "table_body": "
Hand error# Params
Parametric baseline24.540.9M
METRO [39]7.09116.1M
Ours (w/o graph decoder)3.7275.2M
Ours1.3858.3M
", + "bbox": [ + 98, + 161, + 446, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fa04e2f26bb2514e2ef1caf8f199faee728903f8e002eb1c9366b45a1dbf7b07.jpg", + "table_caption": [ + "Table 2: Error rates (in mm) on FreiHAND." + ], + "table_footnote": [], + "table_body": "
METRO [39]Graphormer [40]Ours
PA-MPVPE ↓6.75.95.5
PA-MPJPE ↓6.86.05.6
", + "bbox": [ + 107, + 287, + 444, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metric. We report the Mean-Per-Vertex-Error (MPVE) in $mm$ to evaluate the hand reconstruction error. MPVE measures the mean Euclidean distances between the ground-truth vertices and the predicted vertices.", + "bbox": [ + 75, + 358, + 468, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. We perform a quantitative comparison for the two-hand reconstruction task on our synthetic dataset. As previously mentioned, comparing with other methods is not straightforward since the majority of existing methods focus on single-view settings. We extend METRO [39] to the multi-view setting by applying max-pool to image features and concatenating them with vertex queries. For fair comparisons, we share the same hyperparameter setting for the multi-layer transformer encoder and report the results in Table 1. In this experiment, we downsample the template hand mesh by 10 times. To provide more context regarding the difficulty of our synthetic dataset, we also experiment on a parametric baseline. We create a parametric hand model which uses the 200-dimensional PCA (principal component analysis) subspace from $25\\%$ of the training data. We use ResNet-50 as the backbone for all models in Table 1 and the parametric baseline has an MLP head to predict the input parameters to recover the hand mesh. Our method significantly outperforms both baseline methods with less than half the model size of METRO. We show that our soft-attention feature fusion strategy coupled with mesh segmentation using spectral clustering can effectively reduce the feature channel size from 2048 to 256 without performance drop. In addition, the performance of the parametric baseline is in-line with existing single-hand benchmarks. Lastly, we report results on single-view benchmark FreiHAND [67] in Table 2.", + "bbox": [ + 75, + 425, + 468, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. To motivate our design choices, we present a quantitative evaluation of our method with various components disabled. We validate that each of our proposed technical component contributes meaningfully.", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/58e98dd5f446fab194a315333154e6b6d05311659318b9340a3b8e5a86d7df5e.jpg", + "table_caption": [ + "Table 3: Performance of different multi-view fusion strategies. We report hand error for both settings. $K$ refers to the number of clusters for template hand mesh. Note that we do not include spectral filtering in the graph decoder here." + ], + "table_footnote": [], + "table_body": "
Single-viewMulti-view
METRO [39]10.87-
METRO [39] + avg. pool-8.71
METRO [39] + max pool-7.09
Ours (K=1)-6.59
Ours (K=4)-4.79
Ours (K=7)-3.72
", + "bbox": [ + 519, + 161, + 870, + 276 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6da4621b85caa801ee02d6d0d763ee95e88bdeb6c5fb15c0a76f8277e51f9fa6.jpg", + "table_caption": [ + "Table 4: Ablations on multi-view feature fusion. We report hand error on 3 experimental settings ((a)-(c)) with different feature fusion strategies. We keep the same transformer and spectral graph decoder for all fusion strategies." + ], + "table_footnote": [], + "table_body": "
(a)(b)(c)
Direct concatenation11.313.914.7
Max pool9.311.212.5
Soft-attention + max pool6.77.27.4
", + "bbox": [ + 511, + 364, + 877, + 436 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/675b4d13a61b34fc061d505876cf4926afc2c93e66f2ef579c5e896f9954e9dd.jpg", + "table_caption": [ + "Table 5: Ablations of different spectral filters." + ], + "table_footnote": [], + "table_body": "
ggauglapgcheb3gcheb4gcheb5gcheb6
Hand error1.561.451.381.471.381.39
", + "bbox": [ + 509, + 479, + 877, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Multi-View Feature Fusion. Table 3 shows the results of varying number of spectral clusters $K$ (full results in supp.). The combination of soft-attention fusion and mesh segmentation consistently improves the performance. As our synthetic dataset contains only two views, we further verify our multi-view fusion strategy on more challenging viewing angles captured in our real dataset. We also demonstrate that our method does not overfit to camera setup when tested on unseen camera views in Table 4. In these experiments, we divide the 18 camera views into 2 groups, i.e. the first group contains the first 15 camera views and the second group contains the last 3 unseen camera views for evaluation. Note that the evaluation group contains the only egocentric view. We consider 3 experimental settings by varying number of camera views present in the training data: (a) 15 views, (b) 6 views and (c) 3 views in Table 4.", + "bbox": [ + 496, + 551, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Spectral Filters. We experiment with three commonly-used spectral filters: Gaussian $g_{\\text{gau}}$ , Laplacian $g_{\\text{lap}}$ and Chebyshev filters $g_{\\text{cheb}}$ with varying order of polynomials in Table 5. We provide implementation details and additional qualitative examples in the supplementary. We choose $g_{\\text{cheb}}$ for efficiency and do not find increasing order of polynomials improves performance further.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14672", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cac81a46ce0877043d538abf13499afa6d8eab91dbf7c91d1e98d0bf09b38ae.jpg", + "image_caption": [ + "Figure 7: Qualitative examples on our synthetic dataset. Our model shows robustness to various extreme self-occlusions (left), lighting conditions (middle) and scenes (right). Fine mesh reconstruction details are well-preserved in all scenarios." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 184, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dfaf1ac16563ea1b5ac874d3ffc7df376f2e245ea5aa854e257f4724562748b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 186, + 90, + 299, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3e2e913e07322fa1584d7505a4ba7730d23218db64582dca0ce4db0312ad52d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 89, + 415, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73f8481e5c0ba1ec12e61631c89f092bdaecc93662fa24d2d8c292ae8d96c6bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 89, + 539, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/205fd080caea2fe4bdf645b47b8c68f9ddd29a230e944aa435a48d9c4e916299.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 90, + 617, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6164d9ad3781182432cb66835e5bee09a676acc1d7640f65b0d994509f69f285.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 90, + 720, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2dc6ec83cfec752ac3d06b59c9bed8cca42a4c67b742c9e023509deb2037624d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 90, + 890, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e3b017d84d011132dec8f5d8f059a5cf232961209f7e17b062c9b9d35a359c9d.jpg", + "table_caption": [ + "Table 6: Ablations of different backbones and hyperparameters. We denote $P_{cnn}$ and $P_{total}$ to be the number of parameters for CNN backbone and total model, respectively." + ], + "table_footnote": [], + "table_body": "
BackbonePcnnV'CErrorPtotal
ResNet-5023.5M8042561.3858.3M
EfficientNet-B05.3M1602564.1242.8M
EfficientNet-B05.3M1601284.9637.8M
EfficientNet-B05.3M80646.8934.2M
", + "bbox": [ + 98, + 356, + 446, + 439 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7e5b193a1879bb49dd8641d67ac64821c6d025895197bd980a3b0446e6261a5.jpg", + "image_caption": [ + "Figure 8: Training examples of our real dataset (top) and 3D hand mesh estimation results on in-the-wild image (bottom). The resulting mesh contains 3745 vertices." + ], + "image_footnote": [], + "bbox": [ + 125, + 452, + 410, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Synthetic to Real Transfer. Large-scale synthetic dataset can be used to pre-train models in the absence of suitable real datasets. In the following, we demonstrate a use case of deploying our pre-trained model which was trained on synthetic data to general real images. First, we train on a subset of our real dataset and freeze all part of the pre-trained model except CNN backbone. As the registered meshes for our real dataset have a different mesh topology, we drop the pink cluster of our original template mesh (shown in Fig. 3) before training. This approach is analogous to", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MANO-based 3D hand mesh estimation methods [2, 5] and has been demonstrated in [45]. We show that our model can generalise to in-the-wild-images in Fig. 8.", + "bbox": [ + 496, + 301, + 890, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model Compression. We are interested in finding the minimal size to which model can be compressed while maintaining the METRO baseline performance in Table 1. First, we perform computational complexity analysis on the self-attention layer in the Transformer encoder. There are two key steps to compute self-attention: 1) linear projection to $C$ -dimensional query, key and value matrices from $V'$ vertex queries requires $O(V'C^2)$ and softmax operation for layer output requires another $O(V'^2C)$ . The total computational complexity of each transformer layer is therefore quadratic no matter whether $V'$ or $C$ dominates. Recall that as we uniformly subsample template hand meshes by 10 times to obtain $V' = 804$ and $C$ is empirically set to 256 in earlier sections, $V'$ contributes more to overall complexity as it is dominantly larger. In the following experiments, we gradually decrease $V'$ before $C$ while using a variant of EfficientNet [57]. We present main results in Table 6, while full results are provided in the supplementary.", + "bbox": [ + 496, + 351, + 892, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 631, + 619, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a novel spectral graph-based transformer framework which reconstructs high-fidelity two-hand meshes from egocentric views. The main idea behind this study was to demonstrate that the fundamental properties of the graph Laplacian from the spectral graph theory can be applied to a Transformer. We have shown that our multi-view feature fusion strategy is most effective when coupled with mesh segmentation using spectral clustering. We further present spectral filtering and an optimisation-based refinement to reconstruct more physically plausible meshes. Our framework is general and can be extended to other multi-view reconstruction tasks.", + "bbox": [ + 496, + 652, + 890, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 838, + 638, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2023-2020-0-01789), supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation).", + "bbox": [ + 496, + 851, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14673", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6a2d4ee2a648be09e691072b01186cb4819a94e6ed1f51430e66b916b0ffd82.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 96, + 141, + 336, + 213 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c927a5734fc3d4c695e7a461e0149987c19b319896fa85a6a5cb98cc7951b2a0.jpg", + "image_caption": [ + "Input zoomed view" + ], + "image_footnote": [], + "bbox": [ + 352, + 107, + 612, + 242 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f0a972495511209885c5ab509a8e691d8eb8edd515df133adac06bc3810ca69d.jpg", + "image_caption": [ + "Output zoomed view" + ], + "image_footnote": [], + "bbox": [ + 611, + 114, + 875, + 237 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1fbbfb1b0a1679f18106bf1badfdbd7cafa471d4c0f0933018f816b8c0e29fe6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 311, + 316, + 387 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/76135f1c83a2964e87c1c9f1ad0085202a0cbb0bfcbbe284d99b4fbf4eca06ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 277, + 604, + 404 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a03828bee0b822a8a0bca4201e12f487a1e038c9aa69e1e8bf7b30b7a470b5ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 282, + 875, + 404 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6cd0a39094f2b955da7a35926a8500a70e7077ec04ad0b6e75d3ff9f98f3e39f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 508, + 313, + 580 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a7050f79ee12ee5154d19cf508f26cff30ea525d205a4b9438524e15b2ed2f60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 448, + 593, + 627 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/20f63ca0a3e2a35fc96c231faac23ebf9d0a06be8f645516dd86725a15834904.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 452, + 870, + 636 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/926d9c1c67882e578a833fcf078db310478d41f0fb0015799ff573f7164f30b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 691, + 307, + 751 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/591a7b4c4bf9fc63203bff2f47b2b499e1bd490bdfdc15b4407859a91204eef0.jpg", + "image_caption": [ + "Figure 9: Additional qualitative examples of mesh refinement at inference. Our optimisation-based strategy shows robustness to various hand poses. We present failure cases when self-penetration is highly complex in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 336, + 646, + 591, + 810 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/27c88ef3c9574cec71e64838d3732986439f61980183ca32d1db8176fb284e71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 643, + 870, + 815 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14674", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 2", + "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for RGB-based dense 3D hand pose estimation via neural rendering. In CVPR, 2019. 2, 6, 8", + "[3] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 2", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 6", + "[5] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3D hand shape and pose from images in the wild. In CVPR, 2019. 2, 6, 8", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3", + "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2", + "[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3D human mesh recovery with transformers. In ECCV, 2022. 2, 3, 6", + "[9] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2Mesh: Graph convolutional network for 3D human pose and mesh recovery from a 2D human pose. In ECCV, 2020. 5, 6", + "[10] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 2, 4", + "[11] Enric Corona, Tomas Hodan, Minh Vo, Francesc Moreno-Noguer, Chris Sweeney, Richard Newcombe, and Lingni Ma. LISA: Learning implicit shape and appearance of hands. In CVPR, 2022. 2", + "[12] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016. 5", + "[13] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. NASA neural articulated shape approximation. In ECCV, 2020. 2", + "[14] Inderjit S Dhillon, Yuqiang Guan, and Brian Kulis. Weighted graph cuts without eigenvectors a multilevel approach. TPAMI, 2007. 5", + "[15] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. HOPE-Net: A graph-based model for hand-object pose estimation. In CVPR, 2020. 1", + "[16] Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. AAAI Workshop on Deep Learning on Graphs: Methods and Applications, 2021. 4", + "[17] Vijay Prakash Dwivedi, Chaitanya K Joshi, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. JMLR, 2022. 4", + "[18] Zicong Fan, Adrian Spurr, Muhammed Kocabas, Siyu Tang, Michael J Black, and Otmar Hilliges. Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In 3DV, 2021. 1" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Runyang Feng, Yixing Gao, Xueqing Ma, Tze Ho Elden Tse, and Hyung Jin Chang. Mutual information-based temporal difference learning for human pose estimation in video. In CVPR, 2023. 1", + "[20] Runyang Feng, Yixing Gao, Tze Ho Elden Tse, Xueqing Ma, and Hyung Jin Chang. DiffPose: SpatioTemporal diffusion model for video-based human pose estimation. arXiv preprint arXiv:2307.16687, 2023. 1", + "[21] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with RGB-D videos and 3D hand pose annotations. In CVPR, 2018. 2, 3", + "[22] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3D hand shape and pose estimation from a single RGB image. In CVPR, 2019. 1, 2, 5", + "[23] Patrick Grady, Chengcheng Tang, Christopher D Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C Kemp. ContactOpt: Optimizing contact to improve grasps. In CVPR, 2021. 5", + "[24] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint Transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In CVPR, 2022. 1, 3", + "[25] Shangchen Han, Beibei Liu, Randi Cabezas, Christopher D Twigg, Peizhao Zhang, Jeff Petkau, Tsz-Ho Yu, Chun-Jung Tai, Muzaffer Akbay, Zheng Wang, et al. MEgATrack: Monochrome egocentric articulated hand-tracking for virtual reality. ACM Transactions on Graphics (ToG), 2020. 1, 2", + "[26] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, et al. UmeTrack: Unified multiview end-to-end hand tracking for vr. In SIGGRAPH, 2022. 1, 2", + "[27] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In CVPR, 2019. 1, 2, 4, 5, 6", + "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 5", + "[29] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In CVPR, 2020. 2", + "[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 4", + "[31] Korrawe Karunratanakul, Adrian Spurr, Zicong Fan, Otmar Hilliges, and Siyu Tang. A skeleton-driven neural occupancy representation for articulated hands. In 3DV, 2021. 2", + "[32] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 2022. 3", + "[33] Dong Uk Kim, Kwang In Kim, and Seungryul Baek. End-to-end detection and pose estimation of two interacting hands. In ICCV, 2021. 1, 3", + "[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14675", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Devin Kreuzer, Dominique Beaini, Will Hamilton, Vincent Létourneau, and Prudencio Tossou. Rethinking graph transformers with spectral attention. NeurIPS, 2021. 4", + "[36] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 2", + "[37] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2O: Two hands manipulating objects for first person interaction recognition. In ICCV, 2021. 3", + "[38] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In CVPR, 2022. 3", + "[39] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 2, 3, 4, 5, 6, 7", + "[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, 2021. 2, 3, 5, 6, 7", + "[41] Rong Liu and Hao Zhang. Segmentation of 3D meshes through spectral clustering. In Pacific Graphics, 2004. 4", + "[42] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, 2019. 2", + "[43] Tomas Möller and Ben Trumbore. Fast, minimum storage ray-triangle intersection. Journal of graphics tools, 1997. 5", + "[44] Gyeongsik Moon and Kyoung Mu Lee. I2L-MeshNet: Image-to-lixel prediction network for accurate 3D human pose and mesh estimation from a single RGB image. In ECCV, 2020. 6", + "[45] Gyeongsik Moon, Takaaki Shiratori, and Kyoung Mu Lee. DeepHandMesh: A weakly-supervised deep encoder-decoder framework for high-fidelity hand mesh modeling. In ECCV, 2020. 2, 5, 8", + "[46] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. InterHand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single RGB image. In ECCV, 2020. 1, 2, 3, 6", + "[47] Franziska Mueller, Micah Davis, Florian Bernard, Oleksandr Sotnychenko, Micekal Verschooor, Miguel A Otaduy, Dan Casas, and Christian Theobalt. Real-time pose and shape reconstruction of two interacting hands with a single depth camera. In SIGGRAPH, 2019. 1, 2", + "[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 6", + "[49] Blender project. Cycles. https://www.cycles-renderer.org/.6", + "[50] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3D human pose estimation. In ICCV, 2019. 2", + "[51] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3D pose estimation through camera-disentangled representation. In CVPR, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 36(6):245, 2017. 2", + "[53] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3D reconstruction of interacting hands via collision-aware factorized refinements. In 3DV, 2021. 3", + "[54] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large scale visual recognition challenge. IJCV, 2015. 5", + "[55] Breannan Smith, Chenglei Wu, He Wen, Patrick Peluse, Yaser Sheikh, Jessica K Hodgins, and Takaaki Shiratori. Constraining dense hand surface tracking with elasticity. ACM Transactions on Graphics (ToG), 2020. 6", + "[56] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, 2007. 5", + "[57] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 8", + "[58] Tze Ho Elden Tse, Daniele De Martini, and Letizia Marchegiani. No need to scream: Robust sound-based speaker localisation in challenging scenarios. In ICSR 2019, 2019. 1", + "[59] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In CVPR, 2022. 1, 5, 6", + "[60] Tze Ho Elden Tse, Zhongqun Zhang, Kwang In Kim, Ales Leonardis, Feng Zheng, and Hyung Jin Chang. $S^2$ Contact: Graph-based network for 3D hand-object contact estimation with semi-supervised learning. In ECCV, 2022. 1, 5", + "[61] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. IJCV, 2016. 2", + "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4", + "[63] Jiayi Wang, Franziska Mueller, Florian Bernard, Suzanne Sorli, Oleksandr Sotnychenko, Neng Qian, Miguel A Otabuy, Dan Casas, and Christian Theobalt. RGB2Hands: real-time tracking of 3D hand interactions from monocular rgb video. In SIGGRAPH, 2020. 1, 2", + "[64] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3D human shape and articulated pose models. In CVPR, 2020. 6", + "[65] Baowen Zhang, Yangang Wang, Xiaoming Deng, Yinda Zhang, Ping Tan, Cuixia Ma, and Hongan Wang. Interacting two-hand 3D pose and shape reconstruction from single color image. In ICCV, 2021. 1, 3", + "[66] Linfang Zheng, Ales Leonardis, Tze Ho Elden Tse, Nora Horanyi, Hua Chen, Wei Zhang, and Hyung Jin Chang. TP-AE: Temporally primed 6D object pose tracking with autoencoders. In ICRA, 2022. 1", + "[67] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. FreiHAND: A" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14676", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "header", + "text": "dataset for markerless capture of hand pose and shape from single RGB images. In ICCV, 2019. 3, 6, 7", + "bbox": [ + 109, + 92, + 468, + 119 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14677", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/dbd672e5-b6e6-48ea-861d-97b4378c8ca3_model.json b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/dbd672e5-b6e6-48ea-861d-97b4378c8ca3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bd063f672ece7d3c4f87676d8f4f7231f70886a4 --- /dev/null +++ b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/dbd672e5-b6e6-48ea-861d-97b4378c8ca3_model.json @@ -0,0 +1,2435 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Spectral Graphormer: Spectral Graph-based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.203, + 0.905, + 0.259 + ], + "angle": 0, + "content": "Tze Ho Elden Tse\\(^{1,2*}\\) Franziska Mueller\\(^{1}\\) Zhengyang Shen\\(^{1}\\) Danhang Tang\\(^{1}\\) Thabo Beeler\\(^{1}\\) Mingsong Dou\\(^{1}\\) Yinda Zhang\\(^{1}\\) Sasa Petrovic\\(^{1}\\) Hyung Jin Chang\\(^{2}\\) Jonathan Taylor\\(^{1}\\) Bardia Doosti\\(^{1}\\) \\(^{1}\\)Google \\(^{2}\\)University of Birmingham" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.657 + ], + "angle": 0, + "content": "We propose a novel transformer-based framework that reconstructs two high fidelity hands from multi-view RGB images. Unlike existing hand pose estimation methods, where one typically trains a deep network to regress hand model parameters from single RGB image, we consider a more challenging problem setting where we directly regress the absolute root poses of two-hands with extended forearm at high resolution from egocentric view. As existing datasets are either infeasible for egocentric viewpoints or lack background variations, we create a large-scale synthetic dataset with diverse scenarios and collect a real dataset from multi-calibrated camera setup to verify our proposed multi-view image feature fusion strategy. To make the reconstruction physically plausible, we propose two strategies: (i) a coarse-to-fine spectral graph convolution decoder to smoothen the meshes during upsampling and (ii) an optimisation-based refinement stage at inference to prevent self-penetrations. Through extensive quantitative and qualitative evaluations, we show that our framework is able to produce realistic two-hand reconstructions and demonstrate the generalisation of synthetic-trained models to real data, as well as real-time AR/VR applications." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.685, + 0.21, + 0.701 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.47, + 0.846 + ], + "angle": 0, + "content": "3D hand pose estimation is a fundamental problem in various downstream applications including augmented and virtual reality (AR/VR) [19, 20, 25, 26, 47, 58, 63]. Research efforts in designing 3D hand or hand-object reconstruction networks have shown great effectiveness for single-hand pose estimation [15, 22, 27, 59, 60, 66]. However, two-hand pose estimation has received relatively little attention. In this paper, as illustrated in Fig. 1, we focus on high fidelity two-hand reconstruction from multi-view RGB images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.847, + 0.47, + 0.878 + ], + "angle": 0, + "content": "With the availability of the InterHand2.6M dataset [46], various methods have been recently proposed. More recent" + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.289, + 0.588, + 0.302 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.302, + 0.635, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.405, + 0.635, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.289, + 0.747, + 0.302 + ], + "angle": 0, + "content": "Front view" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.302, + 0.76, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.393, + 0.753, + 0.406 + ], + "angle": 0, + "content": "Rotated view" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.406, + 0.762, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.289, + 0.877, + 0.302 + ], + "angle": 0, + "content": "Zoomed view" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.303, + 0.887, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.389, + 0.884, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.519, + 0.895, + 0.596 + ], + "angle": 0, + "content": "Figure 1: Our method jointly reconstructs high fidelity two-hand meshes from multi-view RGB image. The input images shown here are from our synthetic dataset, which contains challenging video sequences of two hands rendered into egocentric views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.895, + 0.734 + ], + "angle": 0, + "content": "methods focused on alleviating the self-similarity problem between interacting hands by exploiting hand part segmentation probability [18], joint visibility [33], cascaded refinement modules [65] or keypoints using Transformer [24]. Although they perform well on complex configurations, their dependence on root joint alignment does not provide absolute root pose recovery when applied to multi-view scenarios, which is especially crucial for interactions in VR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Moreover, many immersive AR/VR applications require accurate estimation of two hands including extended forearms as it allows for a more realistic and accurate representation of hand movements and gestures. By including the forearm, the orientation and movement of the hand in relation to the arm can provide important contextual information for the user's actions in the virtual environment. Additionally, it can reduce errors and improve the stability of the tracking system, which is important for maintaining immersion and avoiding disorientation in AR/VR applications. However, as in many other areas of computer vision, there" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.378, + 0.902 + ], + "angle": 0, + "content": "*This work was done during an internship at Google." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14666" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.151 + ], + "angle": 0, + "content": "is currently no suitable dataset for a typically supervised deep learning setup. Specifically, existing datasets are either infeasible for egocentric views [21] or lack variations in background and lighting conditions [46]." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.153, + 0.468, + 0.483 + ], + "angle": 0, + "content": "In this paper, we propose a spectral graph-based transformer architecture that can reconstruct high resolution two-hand meshes with extended forearms from multi-view RGB images. However, directly extending state-of-the-art methods [39, 40] from single-view to multi-view setting is nontrivial as they consume a substantial number of parameters and are computationally expensive. As pointed out by Cho et al. [8], the current encoder-based transformers [39, 40] have overlooked the importance of efficient token design. In order to retain spatial information in image features and avoid passing redundant global image features to the transformer encoder, we propose a soft-attention-based multiview image feature fusion strategy to obtain region-specific features. Further, as a hand mesh in essence is graph-structured data, we incorporate properties of graph Laplacian from spectral graph theory [10] into the design of our transformer-based network architecture. To obtain realistic hand reconstructions, we leverage a hierarchical graph decoder and propose an optimisation-based refinement stage during inference to prevent self-penetrations. We show that each technical component above contributes meaningfully in our ablation study." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.485, + 0.468, + 0.695 + ], + "angle": 0, + "content": "Besides the expensive computational cost, an additional challenge is the lack of high fidelity 3D mesh ground truth for training such a model. As manually annotating 3D hand meshes on real data is extremely laborious and expensive, we create a large-scale synthetic multi-view dataset containing realistic hand motions from egocentric camera viewpoints rendered under a large variation of background and illumination. To generalise and evaluate on real-world data, we further collected real data from a precisely calibrated multi-view studio with 18 number of high-resolution cameras. We use an automatic approach for registering meshes and obtain ground truth hand poses. Our proposed method trained on both datasets shows robust performance on challenging scenarios and can serve as a strong baseline." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.697, + 0.336, + 0.71 + ], + "angle": 0, + "content": "Our contributions are the following:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.72, + 0.468, + 0.765 + ], + "angle": 0, + "content": "1. We propose a novel end-to-end trainable spectral graph-based transformer for high fidelity two-hand reconstruction from multi-view RGB image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.468, + 0.845 + ], + "angle": 0, + "content": "2. We design an efficient soft attention-based multi-view image feature fusion in which the resulting image features are region-specific to segmented hand mesh. We further demonstrate a minimal reduction of \\(35\\%\\) in the model size with this approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.851, + 0.468, + 0.881 + ], + "angle": 0, + "content": "3. We introduce an optimisation-based method to refine physically-implausible meshes at inference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.886, + 0.468, + 0.9 + ], + "angle": 0, + "content": "4. We create a large-scale synthetic multi-view dataset" + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "with high resolution 3D hand meshes and collect real dataset to verify our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.146, + 0.64, + 0.161 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.171, + 0.89, + 0.261 + ], + "angle": 0, + "content": "Our work tackles the problem of two-hand reconstruction from multi-view images. We first review the literature on Hand reconstruction from RGB images. Then, we focus on the line that leverages Transformer on human body/hand reconstruction. Finally, we provide a brief review on existing Hand datasets." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.264, + 0.89, + 0.565 + ], + "angle": 0, + "content": "Single-Hand Reconstruction. Many approaches have already been proposed for hand pose estimation from either RGB images or depth maps. Here we focus mainly on works that reconstruct full hand meshes. As most earlier methods do not have access to suitable datasets which contain dense 3D ground truths, the parametric hand model MANO [52] provides an alternative as its input parameters can be directly regressed and it contains prior 3D information about hand geometry. Hence, discriminative approaches [2, 5, 27] which directly predict input hand pose and shape parameters from single images have been a popular approach. [45] learns the pose and shape corrective parameters to reconstruct high-fidelity hand meshes with multi-view depth images. Concurrently, non-parametric methods [22, 36] leveraging graph convolutions demonstrate the ability to directly regress 3D hand meshes. Recently, there is an increasing interest in representing object shapes by learning an implicit function [1, 7, 42]. By extending NASA [13], [31] is a keypoint-driven implicit hand representation and [11] disentangles shape, color and pose representations." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.568, + 0.89, + 0.793 + ], + "angle": 0, + "content": "In contrast to the above mentioned approaches, we propose in this work a non-parametric method that can be extended to multi-view settings and reconstruct high-fidelity hand meshes with extended forearms. Further, our approach is not restricted to a single subject, which is the primary constraint of [45]. Recent work on multi-view fusion leverage epipolar geometry [29, 50] to aggregate estimated 2D heatmaps from different views or Feature Transform Layers (FTL) [26, 51] to learn camera geometry-aware latent features. However, these methods require known camera parameters and the feature matching process based on feature similarity is sensitive to occlusion. Our work introduces a soft-attention mechanism coupled with transformer tokenisation which is robust to occlusion and aggregates multiview image features efficiently." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.796, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Two-Hand Pose Estimation. Pose estimation of interacting hands can be broadly classified as discriminative and generative (or hybrid) approaches. Early methods typically follow hybrid approaches by leveraging visual cues detected by discriminative methods, followed by model fitting. [3, 61] apply collision optimisation terms and physical modelling on detected fingertips. [25, 26, 47, 63] ex" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14667" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.092, + 0.873, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Figure 2: A schematic illustration of our framework. Given multi-view RGB images, we extract volumetric features \\(\\mathbf{f}\\) with a shared CNN backbone. The soft-attention fusion block generates the attention mask \\(\\mathcal{M}\\) and finer image features \\(\\mathbf{f}'\\) through multiple upsampling and convolution blocks. Region-specific features \\(\\mathbf{f}_r\\) are computed by first aggregating \\(\\mathbf{f}'\\) along the feature channel dimension via the attention mask \\(\\mathcal{M}\\), followed by a max-pooling operation across multi-view images to focus on useful features. Then, we apply mesh segmentation via spectral clustering on template hand meshes and uniformly subsample them to obtain coarse meshes. We perform position encoding by concatenating coarse template meshes to the corresponding region-specific features \\(\\mathbf{f}_r\\), i.e. matching colored features to mesh segments. Finally, our multi-layer transformer encoder takes the resulting features as input and outputs a coarse mesh representation \\(\\mathbf{f}_c\\) which is then decoded by a spectral graph decoder to produce the final two-hand meshes at target resolution. Here, each hand contains 4023 vertices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.471, + 0.538 + ], + "angle": 0, + "content": "tract image features or keypoints from RGB or depth images and fit hand models with physical constraints. Recent largescale datasets of interacting hands [46] have led to the emergence of fully discriminative methods [24,33,38,53,65] that jointly estimate the 3D joint locations or hand model parameters from a single RGB image. In this work, we target to reconstruct two hands at high resolution while keeping the model complexity low." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.47, + 0.803 + ], + "angle": 0, + "content": "Transformer in 3D Vision. Transformer-based architectures have been gaining increasing popularity in the vision community. Here we focus on methods that reconstruct the human body or hands from RGB images and refer readers to [32] for a detailed survey. In a closely related work, [39] uses cascaded Transformer encoders to reconstruct the human body and hands from a single RGB image and achieves state-of-the-art performance. [40] extends [39] with graph convolutions along the Transformer encoder. [8] improves the efficiency of [39] by disentangling image encoding and mesh estimation via an encoder-decoder architecture. [24] extends \"Detection Transformer\" [6] with hand-object pose estimations. While these works are targeted at a single image and their extension to the multi-view setting is nontrivial due to large number of learnable parameters, our architecture is designed to efficiently reconstruct two hands at high resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Hand Pose Datasets. The success of discriminative methods depend on the availability and variability of hands interaction datasets. [46] introduces the large-scale dataset InterHand2.6M with closely interacting hand motions using a semi-automatic annotation process with multiple high-resolution RGB cameras. However, it contains minimal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.644 + ], + "angle": 0, + "content": "background and lighting variation. FreiHAND [67] includes wider background variation but the dataset is limited to a single hand and third-person views. Therefore, the above datasets are not suitable for the egocentric two-hand reconstruction task. The available egocentric datasets are limited by either visible markers [21] or constrained lab environments [37]. This motivates us to create a large-scale egocentric synthetic dataset with improved environment and lighting variations. To validate our proposed multi-view fusion strategy, we further collect a real dataset with more challenging camera viewing angles. Altogether, both synthetic and real datasets contain diverse egocentric multi-view and -frame data points with multiple subjects. Our architecture demonstrates promising performance on both datasets and constitutes a strong baseline." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.662, + 0.593, + 0.679 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 2, our architecture first passes \\(N\\) number of multi-view RGB input images \\(\\mathbf{x} \\in \\mathbb{R}^{N \\times 224 \\times 224 \\times 3}\\) to a shared CNN backbone to extract volumetric features \\(\\mathbf{f} \\in \\mathbb{R}^{N \\times 7 \\times 7 \\times 2048}\\), i.e. features before the global average pooling layer for ResNet [28]. The volumetric features are then fed into a soft attention-based multi-view feature encoder and output \\(K\\) region-specific features \\(\\mathbf{f}_r \\in \\mathbb{R}^{K \\times C}\\) where \\(C\\) refers to the feature channel size. The Transformer encoder takes \\(\\mathbf{f}_r\\) together with template hand meshes \\(\\mathbf{m}' \\in \\mathbb{R}^{V' \\times 3}\\) and outputs a coarse mesh representation \\(\\mathbf{f}_c \\in \\mathbb{R}^{V' \\times F}\\). Finally, the spectral graph decoder generates hand meshes \\(\\mathbf{m} \\in \\mathbb{R}^{V \\times 3}\\) by upsampling on \\(\\mathbf{f}_c\\) where \\(V \\gg V'\\). With slight abuse of notation, \\(\\mathbf{m}\\) can either be two-hand or single-hand depending on the application." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14668" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "In the following, we detail the multi-view image feature encoder in Section 3.2, the spectral graph convolution decoder in Section 3.3 and the loss used for training in Section 3.4. We also present an optimisation-based refinement procedure at inference time in Section 3.5." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.179, + 0.188, + 0.193 + ], + "angle": 0, + "content": "3.1. Notations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.471, + 0.408 + ], + "angle": 0, + "content": "A 3D mesh can be represented as an undirected graph \\(\\mathcal{G} = (\\mathcal{V},\\mathcal{E},\\mathbf{A})\\), where \\(\\mathcal{V}\\) is a node set and \\(\\mathcal{E}\\) is an edge set. An adjacency matrix \\(\\mathbf{A} \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}\\) encodes information of pairwise relations between nodes. A degree matrix \\(\\mathbf{D} \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}\\) is a diagonal matrix whose diagonal element \\(\\mathbf{D}_{ii} = \\sum_{j} \\mathbf{A}_{ij}\\) refers to the degree value of each node. An essential operator in spectral graph theory [10] is the graph Laplacian \\(\\mathbf{L}\\), whose definition is \\(\\mathbf{L} = \\mathbf{D} - \\mathbf{A}\\), and \\(\\mathbf{L} = \\mathbf{U} \\boldsymbol{\\Lambda} \\mathbf{U}^T\\) where the graph Laplacian can be diagonalised by the Fourier basis \\(\\mathbf{U} = [\\mathbf{u}_1, \\dots, \\mathbf{u}_{|\\mathcal{V}|}] \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}\\) and \\(\\boldsymbol{\\Lambda} = \\text{diag}([\\lambda_1, \\dots, \\lambda_{|\\mathcal{V}|}]) \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}\\) where \\(\\{\\mathbf{u}_i\\}_{i=1}^{|\\mathcal{V}|}\\) are the eigenvectors and \\(\\{\\lambda_i\\}_{i=1}^{|\\mathcal{V}|}\\) are the non-negative eigenvalues of graph Laplacian (\\(0 = \\lambda_1 \\leq \\dots \\leq \\lambda_{|\\mathcal{V}|}\\))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.418, + 0.386, + 0.433 + ], + "angle": 0, + "content": "3.2. Multi-View Image Feature Encoder" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.441, + 0.469, + 0.562 + ], + "angle": 0, + "content": "To extend transformer-based architecture to multi-view settings and prevent concatenating global image features in an overly-duplicated way as in [39], we propose a simple soft-attention fusion strategy to better aggregate features across multiple views and attend to different hand parts via mesh segmentation. The resulting \\(K\\) region-specific features \\(\\mathbf{f}_r\\) defined by spectral clustering are fed to our Transformer encoder to obtain a coarse mesh representation \\(\\mathbf{f}_c\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Soft-Attention Fusion. Given volumetric features \\(\\mathbf{f}\\), we do not apply any pooling operations to avoid losing spatial information as shown by our experiments (see Table 3). Instead, we propose to aggregate multi-view features with a soft-attention mask. We first obtain a finer representation of \\(\\mathbf{f}\\), denoted as \\(\\mathbf{f}' \\in \\mathbb{R}^{N \\times (H \\times W) \\times C}\\), by feeding it through two blocks each comprised of 2D upsampling with bilinear interpolation, \\(3 \\times 3\\) convolution layers, batch-normalisation [30] and ReLU. The soft-attention mask \\(\\mathcal{M} \\in \\mathbb{R}^{N \\times (H \\times W) \\times K}\\) is obtained by applying 1) \\(K\\) \\(1 \\times 1\\) convolutional filters to reduce the feature channel of \\(\\mathbf{f}'\\) to \\(K\\) and 2) spatial soft arg-max which determines the image-space point of maximal activation in each \\(C\\). At this stage, we can compute per-frame features \\(\\mathbf{f}'' \\in \\mathbb{R}^{N \\times K \\times C}\\) by \\(\\mathbf{f}'' = \\mathcal{M}^{\\mathrm{T}}\\mathbf{f}'\\). We finally obtain region-specific features \\(\\mathbf{f}_r \\in \\mathbb{R}^{K \\times C}\\) by max-pooling along the second dimension of \\(\\mathbf{f}''\\). There are two intuitions to this design: 1) the corresponding feature with higher attention weight contributes more to the final feature representation spatially and 2) max-pooling allows efficient feature selection across multiple views and it does not overfit to any multi-camera configurations as CNN filters are shared across all views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Mesh Segmentation via Spectral Clustering. Inspired by [41], we perform spectral clustering to obtain a 3D mesh segmentation. Different from [41], we apply the eigendecomposition of the graph Laplacian \\(\\mathbf{L}\\) instead of the affinity matrix (which encodes pairwise point affinities with exponential kernel), followed by \\(k\\)-means clustering into \\(K\\) clusters. With this approach, one can efficiently segment any template model mesh without manual effort. In addition, as shown in Fig. 3, the segmented mesh does not exhibit clear-cut boundaries, and certain clusters are scattered throughout the entire surface. This is in spirit similar to mask vertex modeling occlusions in [39] by encouraging the transformer to consider other relevant vertex queries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Transformer Encoder. We now concatenate the \\(C\\)-dimensional region-specific image features to the corresponding \\(K\\) clusters of the segmented template hand mesh. The resulting features \\(\\mathbf{f}_t \\in \\mathbb{R}^{V' \\times (C + 3)^1}\\) are fed into a multi-layer transformer encoder with progressive dimensionality reduction as described in [39]. The output is a coarse mesh representation \\(\\mathbf{f}_c \\in \\mathbb{R}^{V' \\times F}\\), where \\(F = (C + 3)/2^n\\) with \\(n\\) layers of transformer encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.893, + 0.635 + ], + "angle": 0, + "content": "Theoretical Motivations. Since there do not exist explicit coordinate systems as in grid graphs, aligning nodes of highly irregular graphs in nature is a non-trivial problem. Recent studies [16, 17, 35] attempted to encode positional information by leveraging the spectral domain in a way that nearby nodes have similar values and distant nodes have different values. This can be achieved as the eigenvectors of the graph Laplacian can be interpreted as the generalised concepts of sinusoidal functions of positional encoding in Transformers [62]. For instance with any given graph, nodes close to each other can be assigned values of similar positional features as the smaller the eigenvalue of the graph Laplacian \\(\\lambda_{i}\\) (closer to 0) the smoother the coordinates of the corresponding eigenvector \\(\\mathbf{u}_{i}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.647, + 0.724, + 0.663 + ], + "angle": 0, + "content": "3.3. Spectral Graph Decoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.792 + ], + "angle": 0, + "content": "We find that simply relying on fully-connected layers to upsample meshes to target resolution is insufficient as this process introduces instability and disruption to the mesh, even with mesh regularisation loss terms [27] (see Fig. 4). Therefore, we propose to couple the fully-connected layers for upsampling with spectral filtering as meshes can be treated as graph signals \\(\\mathbf{f}_c = (f_1,\\dots ,f_{V'})\\in \\mathbb{R}^{V'\\times F}\\), i.e. \\(V^{\\prime}\\) vertices with \\(F\\)-dimensional features for batch size of 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.799, + 0.892, + 0.875 + ], + "angle": 0, + "content": "Spectral Filtering. As spectral graph convolution can be defined via point-wise products in the transformed Fourier space, graph signals \\(\\mathbf{f}_c\\) filtered by \\(g_{\\theta}\\) can then be expressed as \\(\\mathbf{U}g_{\\theta}(\\boldsymbol {\\Lambda})\\mathbf{U}^{T}\\mathbf{f}_{c}\\) where \\(g_{\\theta}(\\boldsymbol {\\Lambda})\\) is a diagonal matrix with Fourier coefficients. To ensure that the spectral filter corre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.778, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{1} + 3 \\) refers to the 3D positions of mesh template." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14669" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.085, + 0.434, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.473, + 0.364 + ], + "angle": 0, + "content": "Figure 3: Illustration of mesh segmentation via spectral clustering. Here, we have chosen \\( K = 7 \\) clusters and each has its own color for a single right-handed template mesh. As shown, there are no rigid boundaries across the mesh. To prepare inputs \\( \\mathbf{f}_t \\) to transformer, we decrease the size of this mesh (similarly for left hand) uniformly by a factor of 10 through subsampling, i.e. \\( V' = V / 10 \\) and concatenate with the corresponding region-specific features \\( \\mathbf{f}_r \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.471, + 0.448 + ], + "angle": 0, + "content": "sponds to a meaningful convolution on the graph, a natural solution is to parameterise based on the eigenvalues of the Laplacian. Here we have used the Chebyshev polynomial parametrisation of \\( g_{\\theta}(\\mathbf{L}) \\) for fast computation and readers are referred to [12] for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.473, + 0.651 + ], + "angle": 0, + "content": "Architecture. Similar to [9, 22], our decoder is based on a hierarchical architecture where the mesh is recovered by using fully-connected and graph convolution layers for upsampling. We followed [12, 22] to pre-computed coarse graphs using [14] and used the third-order polynomial in the Laplacian. Given the coarse mesh representation \\(\\mathbf{f}_c\\in \\mathbb{R}^{V'\\times F}\\), the key idea here is to smooth out the \\(F\\)-dimensional values after upsampling on \\(V^{\\prime}\\). Note that the fundamental difference between this approach and the one in [9, 22] is that they have a hierarchical architecture on both \\(V^{\\prime}\\) and \\(F\\) dimensions, we apply only spectral filtering and keep \\(F = 3\\) constant which massively reduces the model size while maintaining performance (see Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.471, + 0.748 + ], + "angle": 0, + "content": "**Discussions.** Recall that in Section 3.2, we leverage the properties of Laplacian eigenvectors to perform spectral clustering. Here we can interpret from a signal processing perspective where the Laplacian eigenvectors define signals that vary smoothly across the graph, with the smoothest signals indicating the coarse community structure of the mesh." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.754, + 0.183, + 0.77 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.471, + 0.854 + ], + "angle": 0, + "content": "Following [39, 40], our model can be trained end-to-end with L1 losses on 3D mesh vertices and 2D re-projection using the predicted camera parameters to improve image-mesh alignment. In addition, we apply an edge length regularisation \\(\\mathcal{L}_{edge}\\) [27] to encourage smoothness of the mesh:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.86, + 0.47, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e d g e}} (\\mathbf {m}) = \\frac {1}{| \\mathcal {E} _ {L} |} \\sum_ {l \\in \\mathcal {E} _ {L}} \\left| l ^ {2} - \\mu \\left(\\mathcal {E} _ {L} ^ {2}\\right) \\right|, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.138 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}_L\\) refers to the set of edge lengths, defined as the L2 norms of all edges and \\(\\mu (\\mathcal{E}_L^2)\\) is the average of the squared edge lengths." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.145, + 0.768, + 0.16 + ], + "angle": 0, + "content": "3.5. Mesh Refinement at Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.895, + 0.426 + ], + "angle": 0, + "content": "There are two main approaches in the literature for producing realistic mesh reconstruction: learning-based and optimisation-based methods. Learning-based methods [27, 45] approach the problem by proposing various repulsive losses that penalise penetration during training. However, their generalisation ability to other meshes is limited due to the need for mesh-specific pre-computation. In particular, they require manual selection of areas of interest (i.e. fingertips and palm) which does not prevent other forms of self-penetrations, such as finger-finger. More importantly, the presence of interpenetration at test time shows that the model is not able to learn the physical rule implicitly [27, 59]. On the other hand, recent optimisation-based methods [23, 60] leverage contact maps to refine meshes at inference. However, they rely heavily on accurate contact map estimations and are sensitive to initialisation as contact optimisation is local." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.895, + 0.629 + ], + "angle": 0, + "content": "Proposed Method. To avoid penetrations, we extend the repulsion loss from [27] into an optimisation-based strategy which does not require any form of pre-computation and is more generalisable to other meshes. To identify hand vertices that contribute to collision, we follow [27,43] and cast rays from each vertex and count the number of surface intersections. If the number is odd, it indicates penetration. After obtaining the collision mask \\(\\mathcal{M}_C\\), we compute the nearest point in the source mesh \\(\\mathbf{m}\\) with respect to the set of collision points. If the point and its nearest corresponding point have a different normal, we compute their distance as loss. We minimise this collision loss \\(\\mathcal{L}_{collision}\\) with as-rigid-as-possible (ARAP) [56] regularisation on the mesh shape:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.636, + 0.892, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o l l i s i o n}} (\\mathbf {m}, \\mathcal {M} _ {C}) = \\sum_ {v \\in \\mathcal {V}} \\mathcal {M} _ {C} \\cdot d (v, \\mathcal {V}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.676, + 0.894, + 0.738 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}_C = \\mathbb{1}_{v\\in \\mathrm{Int}(\\mathcal{V})}\\) indicates which vertices belong to the interior of the mesh and \\(d(v,\\mathcal{V}) = \\min_{v'\\in \\mathcal{V}}\\| v - v'\\| _2\\) denotes distances from point \\(v^{\\prime}\\) to set \\(\\mathcal{V}\\). We show that our method is able to remove self-penetration in Fig. 5 and 9." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.749, + 0.634, + 0.767 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We implement our method in TensorFlow. We train all parts of the network simultaneously with the Adam optimiser [34] using a learning rate of \\(10^{-4}\\) when training on the synthetic dataset and \\(10^{-5}\\) when fine-tuning on the real dataset. We use ResNet [28] pre-trained on ImageNet [54] for our CNN backbone. For computing region-specific features \\(\\mathbf{f}_r\\), we set \\(K = 7\\) and \\(C = 256\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14670" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.085, + 0.452, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.473, + 0.385 + ], + "angle": 0, + "content": "Figure 4: Qualitative comparison with METRO [39]. We show that relying on fully-connected layers to upsample meshes is inadequate for high-resolution mesh reconstruction. In contrast, our spectral graph decoder can accurately capture intricate surface features like nails." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.395, + 0.452, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.601, + 0.47, + 0.633 + ], + "angle": 0, + "content": "Figure 5: Qualitative example of mesh refinement at inference. More examples are available in Fig. 9." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Synthetic Dataset Creation. A large-scale multi-view egocentric dataset with challenging interacting hand motion is required to train our pipeline. However, existing datasets are either infeasible for egocentric views or lack variations in terms of background and lighting conditions, which are crucial for AR/VR applications. In addition, the ground-truth mesh in [46] contains \\(5mm\\) fitting error which is sufficient to cause inaccuracy in hand tracking or misalignment when interacting with virtual objects. Therefore, it is not suitable to validate our approach. To this end, we create a new large-scale synthetic egocentric dataset with two high-fidelity 3D hand meshes. In particular, we purchase commercial 3D hand models for both left and right hands. Each of them contains 4023 vertices and 4008 quad faces. By dividing each quad into two triangles, they can be decomposed into 8016 triangular faces. The size and textures of the hand" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.09, + 0.892, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.215, + 0.895, + 0.261 + ], + "angle": 0, + "content": "Figure 6: Qualitative examples of our synthetic dataset. The scenes are rendered with large variations in lighting, texture and background." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.268, + 0.895, + 0.466 + ], + "angle": 0, + "content": "models can vary, and it can also be rigged with 3D joint locations. We apply photorealistic textures as well as natural lighting using High-Dynamic-Range (HDR) images. To create realistic hand motions, we generate 100 poses and interpolate between pairs of these poses randomly over 1000 frame sequences, i.e. pose pair is swapped every 10 frames. Then, we render hands onto 480 4K backgrounds using the Cycles renderer [49]. Our synthetic dataset comprises 1M data points. Each data point consists of 3D annotations for two hands, rendered in two egocentric views to simulate a customised camera headset scenario. We divide the entire dataset into 4 groups and use \\(25\\%\\) randomly for testing and provide additional image examples in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.895, + 0.608 + ], + "angle": 0, + "content": "Real Dataset Collection. Our real hand data is captured from a multi-view stereo system with 18 synchronised Z-Cams. We build a NeRF-based [4, 48] reconstruction pipeline that simultaneously reconstructs and disentangles the foreground and the background, from which a meshing module extracts hand meshes as ground-truth target mesh. We then apply a mesh and tetrahedral registration approach [55] that registers a template hand model [64] to get a well-registered mesh for each reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Baselines. We compare with the state-of-the-art transformer-based architecture METRO [39] that is explicitly designed for human body mesh reconstruction. We follow their official implementation and use 3 transformer encoder layers. We attempted to include other strong baselines to compare with. However, given our challenging multi-view settings on reconstructing high resolution two-hand meshes, existing methods are constrained by either 1) relying on intermediate 3D pose supervision [9,44] or 2) directly regressing a low resolution parametric hand model [2,5,27,59,67]. In addition, the direct extension of METRO [39], Mesh Graphormer [40], is infeasible to extend to multi-view settings as they tokenised volumetric features which increases the computational complexity of each transformer layer quadratically [8]. Therefore, by considering that the performance gain is minor, we pick METRO [39] as a strong baseline and study in-depth. We provide more details about the baselines in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14671" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.472, + 0.15 + ], + "angle": 0, + "content": "Table 1: Error rates on our synthetic dataset. Parametric baseline is trained and tested on right hand only. Our proposed method achieves strong performance with less than half the METRO model size." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.162, + 0.447, + 0.25 + ], + "angle": 0, + "content": "
Hand error# Params
Parametric baseline24.540.9M
METRO [39]7.09116.1M
Ours (w/o graph decoder)3.7275.2M
Ours1.3858.3M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.262, + 0.419, + 0.277 + ], + "angle": 0, + "content": "Table 2: Error rates (in mm) on FreiHAND." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.289, + 0.446, + 0.342 + ], + "angle": 0, + "content": "
METRO [39]Graphormer [40]Ours
PA-MPVPE ↓6.75.95.5
PA-MPJPE ↓6.86.05.6
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.421 + ], + "angle": 0, + "content": "Evaluation Metric. We report the Mean-Per-Vertex-Error (MPVE) in \\(mm\\) to evaluate the hand reconstruction error. MPVE measures the mean Euclidean distances between the ground-truth vertices and the predicted vertices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Quantitative Comparison. We perform a quantitative comparison for the two-hand reconstruction task on our synthetic dataset. As previously mentioned, comparing with other methods is not straightforward since the majority of existing methods focus on single-view settings. We extend METRO [39] to the multi-view setting by applying max-pool to image features and concatenating them with vertex queries. For fair comparisons, we share the same hyperparameter setting for the multi-layer transformer encoder and report the results in Table 1. In this experiment, we downsample the template hand mesh by 10 times. To provide more context regarding the difficulty of our synthetic dataset, we also experiment on a parametric baseline. We create a parametric hand model which uses the 200-dimensional PCA (principal component analysis) subspace from \\(25\\%\\) of the training data. We use ResNet-50 as the backbone for all models in Table 1 and the parametric baseline has an MLP head to predict the input parameters to recover the hand mesh. Our method significantly outperforms both baseline methods with less than half the model size of METRO. We show that our soft-attention feature fusion strategy coupled with mesh segmentation using spectral clustering can effectively reduce the feature channel size from 2048 to 256 without performance drop. In addition, the performance of the parametric baseline is in-line with existing single-hand benchmarks. Lastly, we report results on single-view benchmark FreiHAND [67] in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Ablation Study. To motivate our design choices, we present a quantitative evaluation of our method with various components disabled. We validate that each of our proposed technical component contributes meaningfully." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.151 + ], + "angle": 0, + "content": "Table 3: Performance of different multi-view fusion strategies. We report hand error for both settings. \\( K \\) refers to the number of clusters for template hand mesh. Note that we do not include spectral filtering in the graph decoder here." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.162, + 0.871, + 0.277 + ], + "angle": 0, + "content": "
Single-viewMulti-view
METRO [39]10.87-
METRO [39] + avg. pool-8.71
METRO [39] + max pool-7.09
Ours (K=1)-6.59
Ours (K=4)-4.79
Ours (K=7)-3.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.354 + ], + "angle": 0, + "content": "Table 4: Ablations on multi-view feature fusion. We report hand error on 3 experimental settings ((a)-(c)) with different feature fusion strategies. We keep the same transformer and spectral graph decoder for all fusion strategies." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.365, + 0.879, + 0.438 + ], + "angle": 0, + "content": "
(a)(b)(c)
Direct concatenation11.313.914.7
Max pool9.311.212.5
Soft-attention + max pool6.77.27.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.545, + 0.453, + 0.847, + 0.468 + ], + "angle": 0, + "content": "Table 5: Ablations of different spectral filters." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.48, + 0.879, + 0.524 + ], + "angle": 0, + "content": "
ggauglapgcheb3gcheb4gcheb5gcheb6
Hand error1.561.451.381.471.381.39
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Effects of Multi-View Feature Fusion. Table 3 shows the results of varying number of spectral clusters \\( K \\) (full results in supp.). The combination of soft-attention fusion and mesh segmentation consistently improves the performance. As our synthetic dataset contains only two views, we further verify our multi-view fusion strategy on more challenging viewing angles captured in our real dataset. We also demonstrate that our method does not overfit to camera setup when tested on unseen camera views in Table 4. In these experiments, we divide the 18 camera views into 2 groups, i.e. the first group contains the first 15 camera views and the second group contains the last 3 unseen camera views for evaluation. Note that the evaluation group contains the only egocentric view. We consider 3 experimental settings by varying number of camera views present in the training data: (a) 15 views, (b) 6 views and (c) 3 views in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effects of Spectral Filters. We experiment with three commonly-used spectral filters: Gaussian \\( g_{\\text{gau}} \\), Laplacian \\( g_{\\text{lap}} \\) and Chebyshev filters \\( g_{\\text{cheb}} \\) with varying order of polynomials in Table 5. We provide implementation details and additional qualitative examples in the supplementary. We choose \\( g_{\\text{cheb}} \\) for efficiency and do not find increasing order of polynomials improves performance further." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14672" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.185, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.091, + 0.3, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.09, + 0.416, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.09, + 0.54, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.091, + 0.618, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.091, + 0.721, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.091, + 0.892, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.892, + 0.292 + ], + "angle": 0, + "content": "Figure 7: Qualitative examples on our synthetic dataset. Our model shows robustness to various extreme self-occlusions (left), lighting conditions (middle) and scenes (right). Fine mesh reconstruction details are well-preserved in all scenarios." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.3, + 0.471, + 0.348 + ], + "angle": 0, + "content": "Table 6: Ablations of different backbones and hyperparameters. We denote \\( P_{cnn} \\) and \\( P_{total} \\) to be the number of parameters for CNN backbone and total model, respectively." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.357, + 0.447, + 0.44 + ], + "angle": 0, + "content": "
BackbonePcnnV'CErrorPtotal
ResNet-5023.5M8042561.3858.3M
EfficientNet-B05.3M1602564.1242.8M
EfficientNet-B05.3M1601284.9637.8M
EfficientNet-B05.3M80646.8934.2M
" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.453, + 0.411, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.69, + 0.471, + 0.737 + ], + "angle": 0, + "content": "Figure 8: Training examples of our real dataset (top) and 3D hand mesh estimation results on in-the-wild image (bottom). The resulting mesh contains 3745 vertices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Synthetic to Real Transfer. Large-scale synthetic dataset can be used to pre-train models in the absence of suitable real datasets. In the following, we demonstrate a use case of deploying our pre-trained model which was trained on synthetic data to general real images. First, we train on a subset of our real dataset and freeze all part of the pre-trained model except CNN backbone. As the registered meshes for our real dataset have a different mesh topology, we drop the pink cluster of our original template mesh (shown in Fig. 3) before training. This approach is analogous to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.348 + ], + "angle": 0, + "content": "MANO-based 3D hand mesh estimation methods [2, 5] and has been demonstrated in [45]. We show that our model can generalise to in-the-wild-images in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.893, + 0.626 + ], + "angle": 0, + "content": "Model Compression. We are interested in finding the minimal size to which model can be compressed while maintaining the METRO baseline performance in Table 1. First, we perform computational complexity analysis on the self-attention layer in the Transformer encoder. There are two key steps to compute self-attention: 1) linear projection to \\( C \\)-dimensional query, key and value matrices from \\( V' \\) vertex queries requires \\( O(V'C^2) \\) and softmax operation for layer output requires another \\( O(V'^2C) \\). The total computational complexity of each transformer layer is therefore quadratic no matter whether \\( V' \\) or \\( C \\) dominates. Recall that as we uniformly subsample template hand meshes by 10 times to obtain \\( V' = 804 \\) and \\( C \\) is empirically set to 256 in earlier sections, \\( V' \\) contributes more to overall complexity as it is dominantly larger. In the following experiments, we gradually decrease \\( V' \\) before \\( C \\) while using a variant of EfficientNet [57]. We present main results in Table 6, while full results are provided in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.632, + 0.62, + 0.647 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.835 + ], + "angle": 0, + "content": "In this paper, we have proposed a novel spectral graph-based transformer framework which reconstructs high-fidelity two-hand meshes from egocentric views. The main idea behind this study was to demonstrate that the fundamental properties of the graph Laplacian from the spectral graph theory can be applied to a Transformer. We have shown that our multi-view feature fusion strategy is most effective when coupled with mesh segmentation using spectral clustering. We further present spectral filtering and an optimisation-based refinement to reconstruct more physically plausible meshes. Our framework is general and can be extended to other multi-view reconstruction tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.839, + 0.64, + 0.852 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.852, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2023-2020-0-01789), supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14673" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.088, + 0.251, + 0.102 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.088, + 0.55, + 0.102 + ], + "angle": 0, + "content": "Input zoomed view" + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.088, + 0.816, + 0.102 + ], + "angle": 0, + "content": "Output zoomed view" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.142, + 0.337, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.108, + 0.613, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.116, + 0.877, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.313, + 0.317, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.279, + 0.605, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.284, + 0.877, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.51, + 0.315, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.449, + 0.594, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.453, + 0.872, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.693, + 0.308, + 0.752 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.647, + 0.592, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.645, + 0.872, + 0.816 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.831, + 0.892, + 0.862 + ], + "angle": 0, + "content": "Figure 9: Additional qualitative examples of mesh refinement at inference. Our optimisation-based strategy shows robustness to various hand poses. We present failure cases when self-penetration is highly complex in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14674" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.469, + 0.186 + ], + "angle": 0, + "content": "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for RGB-based dense 3D hand pose estimation via neural rendering. In CVPR, 2019. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.469, + 0.228 + ], + "angle": 0, + "content": "[3] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[5] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3D hand shape and pose from images in the wild. In CVPR, 2019. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.469, + 0.358 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.469, + 0.387 + ], + "angle": 0, + "content": "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.389, + 0.469, + 0.43 + ], + "angle": 0, + "content": "[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3D human mesh recovery with transformers. In ECCV, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[9] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2Mesh: Graph convolutional network for 3D human pose and mesh recovery from a 2D human pose. In ECCV, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[10] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[11] Enric Corona, Tomas Hodan, Minh Vo, Francesc Moreno-Noguer, Chris Sweeney, Richard Newcombe, and Lingni Ma. LISA: Learning implicit shape and appearance of hands. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[12] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[13] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. NASA neural articulated shape approximation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[14] Inderjit S Dhillon, Yuqiang Guan, and Brian Kulis. Weighted graph cuts without eigenvectors a multilevel approach. TPAMI, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[15] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. HOPE-Net: A graph-based model for hand-object pose estimation. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[16] Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. AAAI Workshop on Deep Learning on Graphs: Methods and Applications, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[17] Vijay Prakash Dwivedi, Chaitanya K Joshi, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. JMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[18] Zicong Fan, Adrian Spurr, Muhammed Kocabas, Siyu Tang, Michael J Black, and Otmar Hilliges. Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In 3DV, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[19] Runyang Feng, Yixing Gao, Xueqing Ma, Tze Ho Elden Tse, and Hyung Jin Chang. Mutual information-based temporal difference learning for human pose estimation in video. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[20] Runyang Feng, Yixing Gao, Tze Ho Elden Tse, Xueqing Ma, and Hyung Jin Chang. DiffPose: SpatioTemporal diffusion model for video-based human pose estimation. arXiv preprint arXiv:2307.16687, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[21] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with RGB-D videos and 3D hand pose annotations. In CVPR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[22] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3D hand shape and pose estimation from a single RGB image. In CVPR, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.369 + ], + "angle": 0, + "content": "[23] Patrick Grady, Chengcheng Tang, Christopher D Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C Kemp. ContactOpt: Optimizing contact to improve grasps. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[24] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint Transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[25] Shangchen Han, Beibei Liu, Randi Cabezas, Christopher D Twigg, Peizhao Zhang, Jeff Petkau, Tsz-Ho Yu, Chun-Jung Tai, Muzaffer Akbay, Zheng Wang, et al. MEgATrack: Monochrome egocentric articulated hand-tracking for virtual reality. ACM Transactions on Graphics (ToG), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[26] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, et al. UmeTrack: Unified multiview end-to-end hand tracking for vr. In SIGGRAPH, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[27] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In CVPR, 2019. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[29] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.693, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[31] Korrawe Karunratanakul, Adrian Spurr, Zicong Fan, Otmar Hilliges, and Siyu Tang. A skeleton-driven neural occupancy representation for articulated hands. In 3DV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[32] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[33] Dong Uk Kim, Kwang In Kim, and Seungryul Baek. End-to-end detection and pose estimation of two interacting hands. In ICCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14675" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[35] Devin Kreuzer, Dominique Beaini, Will Hamilton, Vincent Létourneau, and Prudencio Tossou. Rethinking graph transformers with spectral attention. NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[36] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[37] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2O: Two hands manipulating objects for first person interaction recognition. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[38] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[39] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.362 + ], + "angle": 0, + "content": "[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.47, + 0.391 + ], + "angle": 0, + "content": "[41] Rong Liu and Hao Zhang. Segmentation of 3D meshes through spectral clustering. In Pacific Graphics, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[42] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.476 + ], + "angle": 0, + "content": "[43] Tomas Möller and Ben Trumbore. Fast, minimum storage ray-triangle intersection. Journal of graphics tools, 1997. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[44] Gyeongsik Moon and Kyoung Mu Lee. I2L-MeshNet: Image-to-lixel prediction network for accurate 3D human pose and mesh estimation from a single RGB image. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[45] Gyeongsik Moon, Takaaki Shiratori, and Kyoung Mu Lee. DeepHandMesh: A weakly-supervised deep encoder-decoder framework for high-fidelity hand mesh modeling. In ECCV, 2020. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[46] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. InterHand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single RGB image. In ECCV, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[47] Franziska Mueller, Micah Davis, Florian Bernard, Oleksandr Sotnychenko, Micekal Verschooor, Miguel A Otaduy, Dan Casas, and Christian Theobalt. Real-time pose and shape reconstruction of two interacting hands with a single depth camera. In SIGGRAPH, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[49] Blender project. Cycles. https://www.cycles-renderer.org/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[50] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3D human pose estimation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[51] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3D pose estimation through camera-disentangled representation. In CVPR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[52] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 36(6):245, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[53] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3D reconstruction of interacting hands via collision-aware factorized refinements. In 3DV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[54] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large scale visual recognition challenge. IJCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[55] Breannan Smith, Chenglei Wu, He Wen, Patrick Peluse, Yaser Sheikh, Jessica K Hodgins, and Takaaki Shiratori. Constraining dense hand surface tracking with elasticity. ACM Transactions on Graphics (ToG), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[56] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[57] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[58] Tze Ho Elden Tse, Daniele De Martini, and Letizia Marchegiani. No need to scream: Robust sound-based speaker localisation in challenging scenarios. In ICSR 2019, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[59] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In CVPR, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[60] Tze Ho Elden Tse, Zhongqun Zhang, Kwang In Kim, Ales Leonardis, Feng Zheng, and Hyung Jin Chang. \\(S^2\\) Contact: Graph-based network for 3D hand-object contact estimation with semi-supervised learning. In ECCV, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[61] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. IJCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[63] Jiayi Wang, Franziska Mueller, Florian Bernard, Suzanne Sorli, Oleksandr Sotnychenko, Neng Qian, Miguel A Otabuy, Dan Casas, and Christian Theobalt. RGB2Hands: real-time tracking of 3D hand interactions from monocular rgb video. In SIGGRAPH, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[64] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3D human shape and articulated pose models. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[65] Baowen Zhang, Yangang Wang, Xiaoming Deng, Yinda Zhang, Ping Tan, Cuixia Ma, and Hongan Wang. Interacting two-hand 3D pose and shape reconstruction from single color image. In ICCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[66] Linfang Zheng, Ales Leonardis, Tze Ho Elden Tse, Nora Horanyi, Hua Chen, Wei Zhang, and Hyung Jin Chang. TP-AE: Temporally primed 6D object pose tracking with autoencoders. In ICRA, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[67] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. FreiHAND: A" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14676" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.121 + ], + "angle": 0, + "content": "dataset for markerless capture of hand pose and shape from single RGB images. In ICCV, 2019. 3, 6, 7" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14677" + } + ] +] \ No newline at end of file diff --git a/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/dbd672e5-b6e6-48ea-861d-97b4378c8ca3_origin.pdf b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/dbd672e5-b6e6-48ea-861d-97b4378c8ca3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d09461114266e7bf429c07a27dc460fc021787e6 --- /dev/null +++ b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/dbd672e5-b6e6-48ea-861d-97b4378c8ca3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:391bae2d055613a9a63adc34a859af8bc2b57a39f199f088a3413b5633c8f12e +size 6158880 diff --git a/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/full.md b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ced002e7aaa34a28651c64ab6f0ee3092d312525 --- /dev/null +++ b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/full.md @@ -0,0 +1,320 @@ +# Spectral Graphormer: Spectral Graph-based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images + +Tze Ho Elden Tse $^{1,2*}$ Franziska Mueller $^{1}$ Zhengyang Shen $^{1}$ Danhang Tang $^{1}$ Thabo Beeler $^{1}$ Mingsong Dou $^{1}$ Yinda Zhang $^{1}$ Sasa Petrovic $^{1}$ Hyung Jin Chang $^{2}$ Jonathan Taylor $^{1}$ Bardia Doosti $^{1}$ $^{1}$ Google $^{2}$ University of Birmingham + +# Abstract + +We propose a novel transformer-based framework that reconstructs two high fidelity hands from multi-view RGB images. Unlike existing hand pose estimation methods, where one typically trains a deep network to regress hand model parameters from single RGB image, we consider a more challenging problem setting where we directly regress the absolute root poses of two-hands with extended forearm at high resolution from egocentric view. As existing datasets are either infeasible for egocentric viewpoints or lack background variations, we create a large-scale synthetic dataset with diverse scenarios and collect a real dataset from multi-calibrated camera setup to verify our proposed multi-view image feature fusion strategy. To make the reconstruction physically plausible, we propose two strategies: (i) a coarse-to-fine spectral graph convolution decoder to smoothen the meshes during upsampling and (ii) an optimisation-based refinement stage at inference to prevent self-penetrations. Through extensive quantitative and qualitative evaluations, we show that our framework is able to produce realistic two-hand reconstructions and demonstrate the generalisation of synthetic-trained models to real data, as well as real-time AR/VR applications. + +# 1. Introduction + +3D hand pose estimation is a fundamental problem in various downstream applications including augmented and virtual reality (AR/VR) [19, 20, 25, 26, 47, 58, 63]. Research efforts in designing 3D hand or hand-object reconstruction networks have shown great effectiveness for single-hand pose estimation [15, 22, 27, 59, 60, 66]. However, two-hand pose estimation has received relatively little attention. In this paper, as illustrated in Fig. 1, we focus on high fidelity two-hand reconstruction from multi-view RGB images. + +With the availability of the InterHand2.6M dataset [46], various methods have been recently proposed. More recent + +![](images/f9e1a2ba69ca581c8b2f6bfcbf10cc1e94b1a459c3acf9f157c7b0b419663fdb.jpg) +Input + +![](images/b0579d14bbeee00972e7d57d7879038422fec69acb90e0d9c6b86dd095ac41e6.jpg) + +![](images/cc98eeb63b063af0c96f1f8761486c78d7ee7b75178cd75cfe6472be1737e082.jpg) +Front view + +![](images/99ec366cdc10acb3392148a5249e44be9efc4941542eae2061db8dead4202421.jpg) +Rotated view + +![](images/e653ccdf381ab53d91c7e758fe3a13819461dcb55b3706447a42bb2856be0c27.jpg) +Zoomed view + +![](images/1920d8931781f6238e0dd968166ee32da01ec6039b1ed5953d4f6736f8c5c95d.jpg) +Figure 1: Our method jointly reconstructs high fidelity two-hand meshes from multi-view RGB image. The input images shown here are from our synthetic dataset, which contains challenging video sequences of two hands rendered into egocentric views. + +methods focused on alleviating the self-similarity problem between interacting hands by exploiting hand part segmentation probability [18], joint visibility [33], cascaded refinement modules [65] or keypoints using Transformer [24]. Although they perform well on complex configurations, their dependence on root joint alignment does not provide absolute root pose recovery when applied to multi-view scenarios, which is especially crucial for interactions in VR. + +Moreover, many immersive AR/VR applications require accurate estimation of two hands including extended forearms as it allows for a more realistic and accurate representation of hand movements and gestures. By including the forearm, the orientation and movement of the hand in relation to the arm can provide important contextual information for the user's actions in the virtual environment. Additionally, it can reduce errors and improve the stability of the tracking system, which is important for maintaining immersion and avoiding disorientation in AR/VR applications. However, as in many other areas of computer vision, there + +is currently no suitable dataset for a typically supervised deep learning setup. Specifically, existing datasets are either infeasible for egocentric views [21] or lack variations in background and lighting conditions [46]. + +In this paper, we propose a spectral graph-based transformer architecture that can reconstruct high resolution two-hand meshes with extended forearms from multi-view RGB images. However, directly extending state-of-the-art methods [39, 40] from single-view to multi-view setting is nontrivial as they consume a substantial number of parameters and are computationally expensive. As pointed out by Cho et al. [8], the current encoder-based transformers [39, 40] have overlooked the importance of efficient token design. In order to retain spatial information in image features and avoid passing redundant global image features to the transformer encoder, we propose a soft-attention-based multiview image feature fusion strategy to obtain region-specific features. Further, as a hand mesh in essence is graph-structured data, we incorporate properties of graph Laplacian from spectral graph theory [10] into the design of our transformer-based network architecture. To obtain realistic hand reconstructions, we leverage a hierarchical graph decoder and propose an optimisation-based refinement stage during inference to prevent self-penetrations. We show that each technical component above contributes meaningfully in our ablation study. + +Besides the expensive computational cost, an additional challenge is the lack of high fidelity 3D mesh ground truth for training such a model. As manually annotating 3D hand meshes on real data is extremely laborious and expensive, we create a large-scale synthetic multi-view dataset containing realistic hand motions from egocentric camera viewpoints rendered under a large variation of background and illumination. To generalise and evaluate on real-world data, we further collected real data from a precisely calibrated multi-view studio with 18 number of high-resolution cameras. We use an automatic approach for registering meshes and obtain ground truth hand poses. Our proposed method trained on both datasets shows robust performance on challenging scenarios and can serve as a strong baseline. + +Our contributions are the following: + +1. We propose a novel end-to-end trainable spectral graph-based transformer for high fidelity two-hand reconstruction from multi-view RGB image. + +2. We design an efficient soft attention-based multi-view image feature fusion in which the resulting image features are region-specific to segmented hand mesh. We further demonstrate a minimal reduction of $35\%$ in the model size with this approach. + +3. We introduce an optimisation-based method to refine physically-implausible meshes at inference. + +4. We create a large-scale synthetic multi-view dataset + +with high resolution 3D hand meshes and collect real dataset to verify our proposed method. + +# 2. Related Work + +Our work tackles the problem of two-hand reconstruction from multi-view images. We first review the literature on Hand reconstruction from RGB images. Then, we focus on the line that leverages Transformer on human body/hand reconstruction. Finally, we provide a brief review on existing Hand datasets. + +Single-Hand Reconstruction. Many approaches have already been proposed for hand pose estimation from either RGB images or depth maps. Here we focus mainly on works that reconstruct full hand meshes. As most earlier methods do not have access to suitable datasets which contain dense 3D ground truths, the parametric hand model MANO [52] provides an alternative as its input parameters can be directly regressed and it contains prior 3D information about hand geometry. Hence, discriminative approaches [2, 5, 27] which directly predict input hand pose and shape parameters from single images have been a popular approach. [45] learns the pose and shape corrective parameters to reconstruct high-fidelity hand meshes with multi-view depth images. Concurrently, non-parametric methods [22, 36] leveraging graph convolutions demonstrate the ability to directly regress 3D hand meshes. Recently, there is an increasing interest in representing object shapes by learning an implicit function [1, 7, 42]. By extending NASA [13], [31] is a keypoint-driven implicit hand representation and [11] disentangles shape, color and pose representations. + +In contrast to the above mentioned approaches, we propose in this work a non-parametric method that can be extended to multi-view settings and reconstruct high-fidelity hand meshes with extended forearms. Further, our approach is not restricted to a single subject, which is the primary constraint of [45]. Recent work on multi-view fusion leverage epipolar geometry [29, 50] to aggregate estimated 2D heatmaps from different views or Feature Transform Layers (FTL) [26, 51] to learn camera geometry-aware latent features. However, these methods require known camera parameters and the feature matching process based on feature similarity is sensitive to occlusion. Our work introduces a soft-attention mechanism coupled with transformer tokenisation which is robust to occlusion and aggregates multiview image features efficiently. + +Two-Hand Pose Estimation. Pose estimation of interacting hands can be broadly classified as discriminative and generative (or hybrid) approaches. Early methods typically follow hybrid approaches by leveraging visual cues detected by discriminative methods, followed by model fitting. [3, 61] apply collision optimisation terms and physical modelling on detected fingertips. [25, 26, 47, 63] ex + +![](images/29fcfb6ac8d7b1a52209c85878503dce4d6d8eb9bcaf2f870b84fd5c9f5f7d65.jpg) +Figure 2: A schematic illustration of our framework. Given multi-view RGB images, we extract volumetric features $\mathbf{f}$ with a shared CNN backbone. The soft-attention fusion block generates the attention mask $\mathcal{M}$ and finer image features $\mathbf{f}'$ through multiple upsampling and convolution blocks. Region-specific features $\mathbf{f}_r$ are computed by first aggregating $\mathbf{f}'$ along the feature channel dimension via the attention mask $\mathcal{M}$ , followed by a max-pooling operation across multi-view images to focus on useful features. Then, we apply mesh segmentation via spectral clustering on template hand meshes and uniformly subsample them to obtain coarse meshes. We perform position encoding by concatenating coarse template meshes to the corresponding region-specific features $\mathbf{f}_r$ , i.e. matching colored features to mesh segments. Finally, our multi-layer transformer encoder takes the resulting features as input and outputs a coarse mesh representation $\mathbf{f}_c$ which is then decoded by a spectral graph decoder to produce the final two-hand meshes at target resolution. Here, each hand contains 4023 vertices. + +tract image features or keypoints from RGB or depth images and fit hand models with physical constraints. Recent largescale datasets of interacting hands [46] have led to the emergence of fully discriminative methods [24,33,38,53,65] that jointly estimate the 3D joint locations or hand model parameters from a single RGB image. In this work, we target to reconstruct two hands at high resolution while keeping the model complexity low. + +Transformer in 3D Vision. Transformer-based architectures have been gaining increasing popularity in the vision community. Here we focus on methods that reconstruct the human body or hands from RGB images and refer readers to [32] for a detailed survey. In a closely related work, [39] uses cascaded Transformer encoders to reconstruct the human body and hands from a single RGB image and achieves state-of-the-art performance. [40] extends [39] with graph convolutions along the Transformer encoder. [8] improves the efficiency of [39] by disentangling image encoding and mesh estimation via an encoder-decoder architecture. [24] extends "Detection Transformer" [6] with hand-object pose estimations. While these works are targeted at a single image and their extension to the multi-view setting is nontrivial due to large number of learnable parameters, our architecture is designed to efficiently reconstruct two hands at high resolution. + +Hand Pose Datasets. The success of discriminative methods depend on the availability and variability of hands interaction datasets. [46] introduces the large-scale dataset InterHand2.6M with closely interacting hand motions using a semi-automatic annotation process with multiple high-resolution RGB cameras. However, it contains minimal + +background and lighting variation. FreiHAND [67] includes wider background variation but the dataset is limited to a single hand and third-person views. Therefore, the above datasets are not suitable for the egocentric two-hand reconstruction task. The available egocentric datasets are limited by either visible markers [21] or constrained lab environments [37]. This motivates us to create a large-scale egocentric synthetic dataset with improved environment and lighting variations. To validate our proposed multi-view fusion strategy, we further collect a real dataset with more challenging camera viewing angles. Altogether, both synthetic and real datasets contain diverse egocentric multi-view and -frame data points with multiple subjects. Our architecture demonstrates promising performance on both datasets and constitutes a strong baseline. + +# 3. Method + +As shown in Fig. 2, our architecture first passes $N$ number of multi-view RGB input images $\mathbf{x} \in \mathbb{R}^{N \times 224 \times 224 \times 3}$ to a shared CNN backbone to extract volumetric features $\mathbf{f} \in \mathbb{R}^{N \times 7 \times 7 \times 2048}$ , i.e. features before the global average pooling layer for ResNet [28]. The volumetric features are then fed into a soft attention-based multi-view feature encoder and output $K$ region-specific features $\mathbf{f}_r \in \mathbb{R}^{K \times C}$ where $C$ refers to the feature channel size. The Transformer encoder takes $\mathbf{f}_r$ together with template hand meshes $\mathbf{m}' \in \mathbb{R}^{V' \times 3}$ and outputs a coarse mesh representation $\mathbf{f}_c \in \mathbb{R}^{V' \times F}$ . Finally, the spectral graph decoder generates hand meshes $\mathbf{m} \in \mathbb{R}^{V \times 3}$ by upsampling on $\mathbf{f}_c$ where $V \gg V'$ . With slight abuse of notation, $\mathbf{m}$ can either be two-hand or single-hand depending on the application. + +In the following, we detail the multi-view image feature encoder in Section 3.2, the spectral graph convolution decoder in Section 3.3 and the loss used for training in Section 3.4. We also present an optimisation-based refinement procedure at inference time in Section 3.5. + +# 3.1. Notations + +A 3D mesh can be represented as an undirected graph $\mathcal{G} = (\mathcal{V},\mathcal{E},\mathbf{A})$ , where $\mathcal{V}$ is a node set and $\mathcal{E}$ is an edge set. An adjacency matrix $\mathbf{A} \in \mathbb{R}^{|\mathcal{V}| \times |\mathcal{V}|}$ encodes information of pairwise relations between nodes. A degree matrix $\mathbf{D} \in \mathbb{R}^{|\mathcal{V}| \times |\mathcal{V}|}$ is a diagonal matrix whose diagonal element $\mathbf{D}_{ii} = \sum_{j} \mathbf{A}_{ij}$ refers to the degree value of each node. An essential operator in spectral graph theory [10] is the graph Laplacian $\mathbf{L}$ , whose definition is $\mathbf{L} = \mathbf{D} - \mathbf{A}$ , and $\mathbf{L} = \mathbf{U} \boldsymbol{\Lambda} \mathbf{U}^T$ where the graph Laplacian can be diagonalised by the Fourier basis $\mathbf{U} = [\mathbf{u}_1, \dots, \mathbf{u}_{|\mathcal{V}|}] \in \mathbb{R}^{|\mathcal{V}| \times |\mathcal{V}|}$ and $\boldsymbol{\Lambda} = \text{diag}([\lambda_1, \dots, \lambda_{|\mathcal{V}|}]) \in \mathbb{R}^{|\mathcal{V}| \times |\mathcal{V}|}$ where $\{\mathbf{u}_i\}_{i=1}^{|\mathcal{V}|}$ are the eigenvectors and $\{\lambda_i\}_{i=1}^{|\mathcal{V}|}$ are the non-negative eigenvalues of graph Laplacian ( $0 = \lambda_1 \leq \dots \leq \lambda_{|\mathcal{V}|}$ ). + +# 3.2. Multi-View Image Feature Encoder + +To extend transformer-based architecture to multi-view settings and prevent concatenating global image features in an overly-duplicated way as in [39], we propose a simple soft-attention fusion strategy to better aggregate features across multiple views and attend to different hand parts via mesh segmentation. The resulting $K$ region-specific features $\mathbf{f}_r$ defined by spectral clustering are fed to our Transformer encoder to obtain a coarse mesh representation $\mathbf{f}_c$ . + +Soft-Attention Fusion. Given volumetric features $\mathbf{f}$ , we do not apply any pooling operations to avoid losing spatial information as shown by our experiments (see Table 3). Instead, we propose to aggregate multi-view features with a soft-attention mask. We first obtain a finer representation of $\mathbf{f}$ , denoted as $\mathbf{f}' \in \mathbb{R}^{N \times (H \times W) \times C}$ , by feeding it through two blocks each comprised of 2D upsampling with bilinear interpolation, $3 \times 3$ convolution layers, batch-normalisation [30] and ReLU. The soft-attention mask $\mathcal{M} \in \mathbb{R}^{N \times (H \times W) \times K}$ is obtained by applying 1) $K$ $1 \times 1$ convolutional filters to reduce the feature channel of $\mathbf{f}'$ to $K$ and 2) spatial soft arg-max which determines the image-space point of maximal activation in each $C$ . At this stage, we can compute per-frame features $\mathbf{f}'' \in \mathbb{R}^{N \times K \times C}$ by $\mathbf{f}'' = \mathcal{M}^{\mathrm{T}}\mathbf{f}'$ . We finally obtain region-specific features $\mathbf{f}_r \in \mathbb{R}^{K \times C}$ by max-pooling along the second dimension of $\mathbf{f}''$ . There are two intuitions to this design: 1) the corresponding feature with higher attention weight contributes more to the final feature representation spatially and 2) max-pooling allows efficient feature selection across multiple views and it does not overfit to any multi-camera configurations as CNN filters are shared across all views. + +Mesh Segmentation via Spectral Clustering. Inspired by [41], we perform spectral clustering to obtain a 3D mesh segmentation. Different from [41], we apply the eigendecomposition of the graph Laplacian $\mathbf{L}$ instead of the affinity matrix (which encodes pairwise point affinities with exponential kernel), followed by $k$ -means clustering into $K$ clusters. With this approach, one can efficiently segment any template model mesh without manual effort. In addition, as shown in Fig. 3, the segmented mesh does not exhibit clear-cut boundaries, and certain clusters are scattered throughout the entire surface. This is in spirit similar to mask vertex modeling occlusions in [39] by encouraging the transformer to consider other relevant vertex queries. + +Transformer Encoder. We now concatenate the $C$ -dimensional region-specific image features to the corresponding $K$ clusters of the segmented template hand mesh. The resulting features $\mathbf{f}_t \in \mathbb{R}^{V' \times (C + 3)^1}$ are fed into a multi-layer transformer encoder with progressive dimensionality reduction as described in [39]. The output is a coarse mesh representation $\mathbf{f}_c \in \mathbb{R}^{V' \times F}$ , where $F = (C + 3)/2^n$ with $n$ layers of transformer encoder. + +Theoretical Motivations. Since there do not exist explicit coordinate systems as in grid graphs, aligning nodes of highly irregular graphs in nature is a non-trivial problem. Recent studies [16, 17, 35] attempted to encode positional information by leveraging the spectral domain in a way that nearby nodes have similar values and distant nodes have different values. This can be achieved as the eigenvectors of the graph Laplacian can be interpreted as the generalised concepts of sinusoidal functions of positional encoding in Transformers [62]. For instance with any given graph, nodes close to each other can be assigned values of similar positional features as the smaller the eigenvalue of the graph Laplacian $\lambda_{i}$ (closer to 0) the smoother the coordinates of the corresponding eigenvector $\mathbf{u}_{i}$ . + +# 3.3. Spectral Graph Decoder + +We find that simply relying on fully-connected layers to upsample meshes to target resolution is insufficient as this process introduces instability and disruption to the mesh, even with mesh regularisation loss terms [27] (see Fig. 4). Therefore, we propose to couple the fully-connected layers for upsampling with spectral filtering as meshes can be treated as graph signals $\mathbf{f}_c = (f_1,\dots ,f_{V'})\in \mathbb{R}^{V'\times F}$ , i.e. $V^{\prime}$ vertices with $F$ -dimensional features for batch size of 1. + +Spectral Filtering. As spectral graph convolution can be defined via point-wise products in the transformed Fourier space, graph signals $\mathbf{f}_c$ filtered by $g_{\theta}$ can then be expressed as $\mathbf{U}g_{\theta}(\boldsymbol {\Lambda})\mathbf{U}^{T}\mathbf{f}_{c}$ where $g_{\theta}(\boldsymbol {\Lambda})$ is a diagonal matrix with Fourier coefficients. To ensure that the spectral filter corre + +![](images/5d4fe74583be375d47f5db27e8e4eca094a30c92d8beea57350fd175229e2782.jpg) +Figure 3: Illustration of mesh segmentation via spectral clustering. Here, we have chosen $K = 7$ clusters and each has its own color for a single right-handed template mesh. As shown, there are no rigid boundaries across the mesh. To prepare inputs $\mathbf{f}_t$ to transformer, we decrease the size of this mesh (similarly for left hand) uniformly by a factor of 10 through subsampling, i.e. $V' = V / 10$ and concatenate with the corresponding region-specific features $\mathbf{f}_r$ . + +sponds to a meaningful convolution on the graph, a natural solution is to parameterise based on the eigenvalues of the Laplacian. Here we have used the Chebyshev polynomial parametrisation of $g_{\theta}(\mathbf{L})$ for fast computation and readers are referred to [12] for more details. + +Architecture. Similar to [9, 22], our decoder is based on a hierarchical architecture where the mesh is recovered by using fully-connected and graph convolution layers for upsampling. We followed [12, 22] to pre-computed coarse graphs using [14] and used the third-order polynomial in the Laplacian. Given the coarse mesh representation $\mathbf{f}_c\in \mathbb{R}^{V'\times F}$ , the key idea here is to smooth out the $F$ -dimensional values after upsampling on $V^{\prime}$ . Note that the fundamental difference between this approach and the one in [9, 22] is that they have a hierarchical architecture on both $V^{\prime}$ and $F$ dimensions, we apply only spectral filtering and keep $F = 3$ constant which massively reduces the model size while maintaining performance (see Table 1). + +**Discussions.** Recall that in Section 3.2, we leverage the properties of Laplacian eigenvectors to perform spectral clustering. Here we can interpret from a signal processing perspective where the Laplacian eigenvectors define signals that vary smoothly across the graph, with the smoothest signals indicating the coarse community structure of the mesh. + +# 3.4. Training + +Following [39, 40], our model can be trained end-to-end with L1 losses on 3D mesh vertices and 2D re-projection using the predicted camera parameters to improve image-mesh alignment. In addition, we apply an edge length regularisation $\mathcal{L}_{edge}$ [27] to encourage smoothness of the mesh: + +$$ +\mathcal {L} _ {\text {e d g e}} (\mathbf {m}) = \frac {1}{| \mathcal {E} _ {L} |} \sum_ {l \in \mathcal {E} _ {L}} \left| l ^ {2} - \mu \left(\mathcal {E} _ {L} ^ {2}\right) \right|, \tag {1} +$$ + +where $\mathcal{E}_L$ refers to the set of edge lengths, defined as the L2 norms of all edges and $\mu (\mathcal{E}_L^2)$ is the average of the squared edge lengths. + +# 3.5. Mesh Refinement at Inference + +There are two main approaches in the literature for producing realistic mesh reconstruction: learning-based and optimisation-based methods. Learning-based methods [27, 45] approach the problem by proposing various repulsive losses that penalise penetration during training. However, their generalisation ability to other meshes is limited due to the need for mesh-specific pre-computation. In particular, they require manual selection of areas of interest (i.e. fingertips and palm) which does not prevent other forms of self-penetrations, such as finger-finger. More importantly, the presence of interpenetration at test time shows that the model is not able to learn the physical rule implicitly [27, 59]. On the other hand, recent optimisation-based methods [23, 60] leverage contact maps to refine meshes at inference. However, they rely heavily on accurate contact map estimations and are sensitive to initialisation as contact optimisation is local. + +Proposed Method. To avoid penetrations, we extend the repulsion loss from [27] into an optimisation-based strategy which does not require any form of pre-computation and is more generalisable to other meshes. To identify hand vertices that contribute to collision, we follow [27,43] and cast rays from each vertex and count the number of surface intersections. If the number is odd, it indicates penetration. After obtaining the collision mask $\mathcal{M}_C$ , we compute the nearest point in the source mesh $\mathbf{m}$ with respect to the set of collision points. If the point and its nearest corresponding point have a different normal, we compute their distance as loss. We minimise this collision loss $\mathcal{L}_{collision}$ with as-rigid-as-possible (ARAP) [56] regularisation on the mesh shape: + +$$ +\mathcal {L} _ {\text {c o l l i s i o n}} (\mathbf {m}, \mathcal {M} _ {C}) = \sum_ {v \in \mathcal {V}} \mathcal {M} _ {C} \cdot d (v, \mathcal {V}), \tag {2} +$$ + +where $\mathcal{M}_C = \mathbb{1}_{v\in \mathrm{Int}(\mathcal{V})}$ indicates which vertices belong to the interior of the mesh and $d(v,\mathcal{V}) = \min_{v'\in \mathcal{V}}\| v - v'\| _2$ denotes distances from point $v^{\prime}$ to set $\mathcal{V}$ . We show that our method is able to remove self-penetration in Fig. 5 and 9. + +# 4. Experiments + +Implementation Details. We implement our method in TensorFlow. We train all parts of the network simultaneously with the Adam optimiser [34] using a learning rate of $10^{-4}$ when training on the synthetic dataset and $10^{-5}$ when fine-tuning on the real dataset. We use ResNet [28] pre-trained on ImageNet [54] for our CNN backbone. For computing region-specific features $\mathbf{f}_r$ , we set $K = 7$ and $C = 256$ . + +![](images/2a1860447dab3f115e4823d25edd80ff279df4fcbc0119f8b545c8aebde37774.jpg) +Figure 4: Qualitative comparison with METRO [39]. We show that relying on fully-connected layers to upsample meshes is inadequate for high-resolution mesh reconstruction. In contrast, our spectral graph decoder can accurately capture intricate surface features like nails. + +![](images/acfbc6247fac97551120a7a16e353e48a5b2e870a763664886c74b713191ad72.jpg) +Figure 5: Qualitative example of mesh refinement at inference. More examples are available in Fig. 9. + +Synthetic Dataset Creation. A large-scale multi-view egocentric dataset with challenging interacting hand motion is required to train our pipeline. However, existing datasets are either infeasible for egocentric views or lack variations in terms of background and lighting conditions, which are crucial for AR/VR applications. In addition, the ground-truth mesh in [46] contains $5mm$ fitting error which is sufficient to cause inaccuracy in hand tracking or misalignment when interacting with virtual objects. Therefore, it is not suitable to validate our approach. To this end, we create a new large-scale synthetic egocentric dataset with two high-fidelity 3D hand meshes. In particular, we purchase commercial 3D hand models for both left and right hands. Each of them contains 4023 vertices and 4008 quad faces. By dividing each quad into two triangles, they can be decomposed into 8016 triangular faces. The size and textures of the hand + +![](images/69bb32fe51f1cfc2398bfb003159c040f7a2dd46a8064fdceb339e33bcd5c9e3.jpg) +Figure 6: Qualitative examples of our synthetic dataset. The scenes are rendered with large variations in lighting, texture and background. + +models can vary, and it can also be rigged with 3D joint locations. We apply photorealistic textures as well as natural lighting using High-Dynamic-Range (HDR) images. To create realistic hand motions, we generate 100 poses and interpolate between pairs of these poses randomly over 1000 frame sequences, i.e. pose pair is swapped every 10 frames. Then, we render hands onto 480 4K backgrounds using the Cycles renderer [49]. Our synthetic dataset comprises 1M data points. Each data point consists of 3D annotations for two hands, rendered in two egocentric views to simulate a customised camera headset scenario. We divide the entire dataset into 4 groups and use $25\%$ randomly for testing and provide additional image examples in Fig. 6. + +Real Dataset Collection. Our real hand data is captured from a multi-view stereo system with 18 synchronised Z-Cams. We build a NeRF-based [4, 48] reconstruction pipeline that simultaneously reconstructs and disentangles the foreground and the background, from which a meshing module extracts hand meshes as ground-truth target mesh. We then apply a mesh and tetrahedral registration approach [55] that registers a template hand model [64] to get a well-registered mesh for each reconstruction. + +Baselines. We compare with the state-of-the-art transformer-based architecture METRO [39] that is explicitly designed for human body mesh reconstruction. We follow their official implementation and use 3 transformer encoder layers. We attempted to include other strong baselines to compare with. However, given our challenging multi-view settings on reconstructing high resolution two-hand meshes, existing methods are constrained by either 1) relying on intermediate 3D pose supervision [9,44] or 2) directly regressing a low resolution parametric hand model [2,5,27,59,67]. In addition, the direct extension of METRO [39], Mesh Graphormer [40], is infeasible to extend to multi-view settings as they tokenised volumetric features which increases the computational complexity of each transformer layer quadratically [8]. Therefore, by considering that the performance gain is minor, we pick METRO [39] as a strong baseline and study in-depth. We provide more details about the baselines in the supplementary materials. + +Table 1: Error rates on our synthetic dataset. Parametric baseline is trained and tested on right hand only. Our proposed method achieves strong performance with less than half the METRO model size. + +
Hand error# Params
Parametric baseline24.540.9M
METRO [39]7.09116.1M
Ours (w/o graph decoder)3.7275.2M
Ours1.3858.3M
+ +Table 2: Error rates (in mm) on FreiHAND. + +
METRO [39]Graphormer [40]Ours
PA-MPVPE ↓6.75.95.5
PA-MPJPE ↓6.86.05.6
+ +Evaluation Metric. We report the Mean-Per-Vertex-Error (MPVE) in $mm$ to evaluate the hand reconstruction error. MPVE measures the mean Euclidean distances between the ground-truth vertices and the predicted vertices. + +Quantitative Comparison. We perform a quantitative comparison for the two-hand reconstruction task on our synthetic dataset. As previously mentioned, comparing with other methods is not straightforward since the majority of existing methods focus on single-view settings. We extend METRO [39] to the multi-view setting by applying max-pool to image features and concatenating them with vertex queries. For fair comparisons, we share the same hyperparameter setting for the multi-layer transformer encoder and report the results in Table 1. In this experiment, we downsample the template hand mesh by 10 times. To provide more context regarding the difficulty of our synthetic dataset, we also experiment on a parametric baseline. We create a parametric hand model which uses the 200-dimensional PCA (principal component analysis) subspace from $25\%$ of the training data. We use ResNet-50 as the backbone for all models in Table 1 and the parametric baseline has an MLP head to predict the input parameters to recover the hand mesh. Our method significantly outperforms both baseline methods with less than half the model size of METRO. We show that our soft-attention feature fusion strategy coupled with mesh segmentation using spectral clustering can effectively reduce the feature channel size from 2048 to 256 without performance drop. In addition, the performance of the parametric baseline is in-line with existing single-hand benchmarks. Lastly, we report results on single-view benchmark FreiHAND [67] in Table 2. + +Ablation Study. To motivate our design choices, we present a quantitative evaluation of our method with various components disabled. We validate that each of our proposed technical component contributes meaningfully. + +Table 3: Performance of different multi-view fusion strategies. We report hand error for both settings. $K$ refers to the number of clusters for template hand mesh. Note that we do not include spectral filtering in the graph decoder here. + +
Single-viewMulti-view
METRO [39]10.87-
METRO [39] + avg. pool-8.71
METRO [39] + max pool-7.09
Ours (K=1)-6.59
Ours (K=4)-4.79
Ours (K=7)-3.72
+ +Table 4: Ablations on multi-view feature fusion. We report hand error on 3 experimental settings ((a)-(c)) with different feature fusion strategies. We keep the same transformer and spectral graph decoder for all fusion strategies. + +
(a)(b)(c)
Direct concatenation11.313.914.7
Max pool9.311.212.5
Soft-attention + max pool6.77.27.4
+ +Table 5: Ablations of different spectral filters. + +
ggauglapgcheb3gcheb4gcheb5gcheb6
Hand error1.561.451.381.471.381.39
+ +Effects of Multi-View Feature Fusion. Table 3 shows the results of varying number of spectral clusters $K$ (full results in supp.). The combination of soft-attention fusion and mesh segmentation consistently improves the performance. As our synthetic dataset contains only two views, we further verify our multi-view fusion strategy on more challenging viewing angles captured in our real dataset. We also demonstrate that our method does not overfit to camera setup when tested on unseen camera views in Table 4. In these experiments, we divide the 18 camera views into 2 groups, i.e. the first group contains the first 15 camera views and the second group contains the last 3 unseen camera views for evaluation. Note that the evaluation group contains the only egocentric view. We consider 3 experimental settings by varying number of camera views present in the training data: (a) 15 views, (b) 6 views and (c) 3 views in Table 4. + +Effects of Spectral Filters. We experiment with three commonly-used spectral filters: Gaussian $g_{\text{gau}}$ , Laplacian $g_{\text{lap}}$ and Chebyshev filters $g_{\text{cheb}}$ with varying order of polynomials in Table 5. We provide implementation details and additional qualitative examples in the supplementary. We choose $g_{\text{cheb}}$ for efficiency and do not find increasing order of polynomials improves performance further. + +![](images/7cac81a46ce0877043d538abf13499afa6d8eab91dbf7c91d1e98d0bf09b38ae.jpg) +Figure 7: Qualitative examples on our synthetic dataset. Our model shows robustness to various extreme self-occlusions (left), lighting conditions (middle) and scenes (right). Fine mesh reconstruction details are well-preserved in all scenarios. + +![](images/dfaf1ac16563ea1b5ac874d3ffc7df376f2e245ea5aa854e257f4724562748b8.jpg) + +![](images/c3e2e913e07322fa1584d7505a4ba7730d23218db64582dca0ce4db0312ad52d.jpg) + +![](images/73f8481e5c0ba1ec12e61631c89f092bdaecc93662fa24d2d8c292ae8d96c6bc.jpg) + +![](images/205fd080caea2fe4bdf645b47b8c68f9ddd29a230e944aa435a48d9c4e916299.jpg) + +![](images/6164d9ad3781182432cb66835e5bee09a676acc1d7640f65b0d994509f69f285.jpg) + +![](images/2dc6ec83cfec752ac3d06b59c9bed8cca42a4c67b742c9e023509deb2037624d.jpg) + +Table 6: Ablations of different backbones and hyperparameters. We denote $P_{cnn}$ and $P_{total}$ to be the number of parameters for CNN backbone and total model, respectively. + +
BackbonePcnnV'CErrorPtotal
ResNet-5023.5M8042561.3858.3M
EfficientNet-B05.3M1602564.1242.8M
EfficientNet-B05.3M1601284.9637.8M
EfficientNet-B05.3M80646.8934.2M
+ +![](images/f7e5b193a1879bb49dd8641d67ac64821c6d025895197bd980a3b0446e6261a5.jpg) +Figure 8: Training examples of our real dataset (top) and 3D hand mesh estimation results on in-the-wild image (bottom). The resulting mesh contains 3745 vertices. + +Synthetic to Real Transfer. Large-scale synthetic dataset can be used to pre-train models in the absence of suitable real datasets. In the following, we demonstrate a use case of deploying our pre-trained model which was trained on synthetic data to general real images. First, we train on a subset of our real dataset and freeze all part of the pre-trained model except CNN backbone. As the registered meshes for our real dataset have a different mesh topology, we drop the pink cluster of our original template mesh (shown in Fig. 3) before training. This approach is analogous to + +MANO-based 3D hand mesh estimation methods [2, 5] and has been demonstrated in [45]. We show that our model can generalise to in-the-wild-images in Fig. 8. + +Model Compression. We are interested in finding the minimal size to which model can be compressed while maintaining the METRO baseline performance in Table 1. First, we perform computational complexity analysis on the self-attention layer in the Transformer encoder. There are two key steps to compute self-attention: 1) linear projection to $C$ -dimensional query, key and value matrices from $V'$ vertex queries requires $O(V'C^2)$ and softmax operation for layer output requires another $O(V'^2C)$ . The total computational complexity of each transformer layer is therefore quadratic no matter whether $V'$ or $C$ dominates. Recall that as we uniformly subsample template hand meshes by 10 times to obtain $V' = 804$ and $C$ is empirically set to 256 in earlier sections, $V'$ contributes more to overall complexity as it is dominantly larger. In the following experiments, we gradually decrease $V'$ before $C$ while using a variant of EfficientNet [57]. We present main results in Table 6, while full results are provided in the supplementary. + +# 5. Conclusion + +In this paper, we have proposed a novel spectral graph-based transformer framework which reconstructs high-fidelity two-hand meshes from egocentric views. The main idea behind this study was to demonstrate that the fundamental properties of the graph Laplacian from the spectral graph theory can be applied to a Transformer. We have shown that our multi-view feature fusion strategy is most effective when coupled with mesh segmentation using spectral clustering. We further present spectral filtering and an optimisation-based refinement to reconstruct more physically plausible meshes. Our framework is general and can be extended to other multi-view reconstruction tasks. + +# Acknowledgements + +This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2023-2020-0-01789), supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation). + +![](images/f6a2d4ee2a648be09e691072b01186cb4819a94e6ed1f51430e66b916b0ffd82.jpg) +Input + +![](images/c927a5734fc3d4c695e7a461e0149987c19b319896fa85a6a5cb98cc7951b2a0.jpg) +Input zoomed view + +![](images/f0a972495511209885c5ab509a8e691d8eb8edd515df133adac06bc3810ca69d.jpg) +Output zoomed view + +![](images/1fbbfb1b0a1679f18106bf1badfdbd7cafa471d4c0f0933018f816b8c0e29fe6.jpg) + +![](images/76135f1c83a2964e87c1c9f1ad0085202a0cbb0bfcbbe284d99b4fbf4eca06ff.jpg) + +![](images/a03828bee0b822a8a0bca4201e12f487a1e038c9aa69e1e8bf7b30b7a470b5ee.jpg) + +![](images/6cd0a39094f2b955da7a35926a8500a70e7077ec04ad0b6e75d3ff9f98f3e39f.jpg) + +![](images/a7050f79ee12ee5154d19cf508f26cff30ea525d205a4b9438524e15b2ed2f60.jpg) + +![](images/20f63ca0a3e2a35fc96c231faac23ebf9d0a06be8f645516dd86725a15834904.jpg) + +![](images/926d9c1c67882e578a833fcf078db310478d41f0fb0015799ff573f7164f30b5.jpg) + +![](images/591a7b4c4bf9fc63203bff2f47b2b499e1bd490bdfdc15b4407859a91204eef0.jpg) +Figure 9: Additional qualitative examples of mesh refinement at inference. Our optimisation-based strategy shows robustness to various hand poses. We present failure cases when self-penetration is highly complex in the supplementary materials. + +![](images/27c88ef3c9574cec71e64838d3732986439f61980183ca32d1db8176fb284e71.jpg) + +# References + +[1] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 2 +[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for RGB-based dense 3D hand pose estimation via neural rendering. In CVPR, 2019. 2, 6, 8 +[3] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 2 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 6 +[5] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3D hand shape and pose from images in the wild. In CVPR, 2019. 2, 6, 8 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3 +[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2 +[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3D human mesh recovery with transformers. In ECCV, 2022. 2, 3, 6 +[9] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2Mesh: Graph convolutional network for 3D human pose and mesh recovery from a 2D human pose. In ECCV, 2020. 5, 6 +[10] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 2, 4 +[11] Enric Corona, Tomas Hodan, Minh Vo, Francesc Moreno-Noguer, Chris Sweeney, Richard Newcombe, and Lingni Ma. LISA: Learning implicit shape and appearance of hands. In CVPR, 2022. 2 +[12] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016. 5 +[13] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. NASA neural articulated shape approximation. In ECCV, 2020. 2 +[14] Inderjit S Dhillon, Yuqiang Guan, and Brian Kulis. Weighted graph cuts without eigenvectors a multilevel approach. TPAMI, 2007. 5 +[15] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. HOPE-Net: A graph-based model for hand-object pose estimation. In CVPR, 2020. 1 +[16] Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. AAAI Workshop on Deep Learning on Graphs: Methods and Applications, 2021. 4 +[17] Vijay Prakash Dwivedi, Chaitanya K Joshi, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. JMLR, 2022. 4 +[18] Zicong Fan, Adrian Spurr, Muhammed Kocabas, Siyu Tang, Michael J Black, and Otmar Hilliges. Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In 3DV, 2021. 1 + +[19] Runyang Feng, Yixing Gao, Xueqing Ma, Tze Ho Elden Tse, and Hyung Jin Chang. Mutual information-based temporal difference learning for human pose estimation in video. In CVPR, 2023. 1 +[20] Runyang Feng, Yixing Gao, Tze Ho Elden Tse, Xueqing Ma, and Hyung Jin Chang. DiffPose: SpatioTemporal diffusion model for video-based human pose estimation. arXiv preprint arXiv:2307.16687, 2023. 1 +[21] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with RGB-D videos and 3D hand pose annotations. In CVPR, 2018. 2, 3 +[22] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3D hand shape and pose estimation from a single RGB image. In CVPR, 2019. 1, 2, 5 +[23] Patrick Grady, Chengcheng Tang, Christopher D Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C Kemp. ContactOpt: Optimizing contact to improve grasps. In CVPR, 2021. 5 +[24] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint Transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In CVPR, 2022. 1, 3 +[25] Shangchen Han, Beibei Liu, Randi Cabezas, Christopher D Twigg, Peizhao Zhang, Jeff Petkau, Tsz-Ho Yu, Chun-Jung Tai, Muzaffer Akbay, Zheng Wang, et al. MEgATrack: Monochrome egocentric articulated hand-tracking for virtual reality. ACM Transactions on Graphics (ToG), 2020. 1, 2 +[26] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, et al. UmeTrack: Unified multiview end-to-end hand tracking for vr. In SIGGRAPH, 2022. 1, 2 +[27] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In CVPR, 2019. 1, 2, 4, 5, 6 +[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 5 +[29] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In CVPR, 2020. 2 +[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 4 +[31] Korrawe Karunratanakul, Adrian Spurr, Zicong Fan, Otmar Hilliges, and Siyu Tang. A skeleton-driven neural occupancy representation for articulated hands. In 3DV, 2021. 2 +[32] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 2022. 3 +[33] Dong Uk Kim, Kwang In Kim, and Seungryul Baek. End-to-end detection and pose estimation of two interacting hands. In ICCV, 2021. 1, 3 +[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 + +[35] Devin Kreuzer, Dominique Beaini, Will Hamilton, Vincent Létourneau, and Prudencio Tossou. Rethinking graph transformers with spectral attention. NeurIPS, 2021. 4 +[36] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 2 +[37] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2O: Two hands manipulating objects for first person interaction recognition. In ICCV, 2021. 3 +[38] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In CVPR, 2022. 3 +[39] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 2, 3, 4, 5, 6, 7 +[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, 2021. 2, 3, 5, 6, 7 +[41] Rong Liu and Hao Zhang. Segmentation of 3D meshes through spectral clustering. In Pacific Graphics, 2004. 4 +[42] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, 2019. 2 +[43] Tomas Möller and Ben Trumbore. Fast, minimum storage ray-triangle intersection. Journal of graphics tools, 1997. 5 +[44] Gyeongsik Moon and Kyoung Mu Lee. I2L-MeshNet: Image-to-lixel prediction network for accurate 3D human pose and mesh estimation from a single RGB image. In ECCV, 2020. 6 +[45] Gyeongsik Moon, Takaaki Shiratori, and Kyoung Mu Lee. DeepHandMesh: A weakly-supervised deep encoder-decoder framework for high-fidelity hand mesh modeling. In ECCV, 2020. 2, 5, 8 +[46] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. InterHand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single RGB image. In ECCV, 2020. 1, 2, 3, 6 +[47] Franziska Mueller, Micah Davis, Florian Bernard, Oleksandr Sotnychenko, Micekal Verschooor, Miguel A Otaduy, Dan Casas, and Christian Theobalt. Real-time pose and shape reconstruction of two interacting hands with a single depth camera. In SIGGRAPH, 2019. 1, 2 +[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 6 +[49] Blender project. Cycles. https://www.cycles-renderer.org/.6 +[50] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3D human pose estimation. In ICCV, 2019. 2 +[51] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3D pose estimation through camera-disentangled representation. In CVPR, 2020. 2 + +[52] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 36(6):245, 2017. 2 +[53] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3D reconstruction of interacting hands via collision-aware factorized refinements. In 3DV, 2021. 3 +[54] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large scale visual recognition challenge. IJCV, 2015. 5 +[55] Breannan Smith, Chenglei Wu, He Wen, Patrick Peluse, Yaser Sheikh, Jessica K Hodgins, and Takaaki Shiratori. Constraining dense hand surface tracking with elasticity. ACM Transactions on Graphics (ToG), 2020. 6 +[56] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, 2007. 5 +[57] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 8 +[58] Tze Ho Elden Tse, Daniele De Martini, and Letizia Marchegiani. No need to scream: Robust sound-based speaker localisation in challenging scenarios. In ICSR 2019, 2019. 1 +[59] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In CVPR, 2022. 1, 5, 6 +[60] Tze Ho Elden Tse, Zhongqun Zhang, Kwang In Kim, Ales Leonardis, Feng Zheng, and Hyung Jin Chang. $S^2$ Contact: Graph-based network for 3D hand-object contact estimation with semi-supervised learning. In ECCV, 2022. 1, 5 +[61] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. IJCV, 2016. 2 +[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4 +[63] Jiayi Wang, Franziska Mueller, Florian Bernard, Suzanne Sorli, Oleksandr Sotnychenko, Neng Qian, Miguel A Otabuy, Dan Casas, and Christian Theobalt. RGB2Hands: real-time tracking of 3D hand interactions from monocular rgb video. In SIGGRAPH, 2020. 1, 2 +[64] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3D human shape and articulated pose models. In CVPR, 2020. 6 +[65] Baowen Zhang, Yangang Wang, Xiaoming Deng, Yinda Zhang, Ping Tan, Cuixia Ma, and Hongan Wang. Interacting two-hand 3D pose and shape reconstruction from single color image. In ICCV, 2021. 1, 3 +[66] Linfang Zheng, Ales Leonardis, Tze Ho Elden Tse, Nora Horanyi, Hua Chen, Wei Zhang, and Hyung Jin Chang. TP-AE: Temporally primed 6D object pose tracking with autoencoders. In ICRA, 2022. 1 +[67] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. FreiHAND: A \ No newline at end of file diff --git a/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/images.zip b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d3bb0becc7694e2bc2cecfe8647070b2f51527c9 --- /dev/null +++ b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32c68370fd08f9116214b299627edfd25b7e45873b8b778ee6a680cc68373828 +size 574980 diff --git a/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/layout.json b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4610f39b901db6ec8da5d9dce45b5b23c9eb5c1f --- /dev/null +++ b/2023/Spectral Graphormer_ Spectral Graph-Based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images/layout.json @@ -0,0 +1,10410 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 523, + 140 + ], + "type": "text", + "content": "Spectral Graphormer: Spectral Graph-based Transformer for Egocentric Two-Hand Reconstruction using Multi-View Color Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": "Tze Ho Elden Tse" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": " Franziska Mueller" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": " Zhengyang Shen" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": " Danhang Tang" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": " Thabo Beeler" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": " Mingsong Dou" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": " Yinda Zhang" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": " Sasa Petrovic" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": " Hyung Jin Chang" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": " Jonathan Taylor" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": " Bardia Doosti" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": "Google " + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 160, + 553, + 205 + ], + "type": "text", + "content": "University of Birmingham" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 256, + 290, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 520 + ], + "type": "text", + "content": "We propose a novel transformer-based framework that reconstructs two high fidelity hands from multi-view RGB images. Unlike existing hand pose estimation methods, where one typically trains a deep network to regress hand model parameters from single RGB image, we consider a more challenging problem setting where we directly regress the absolute root poses of two-hands with extended forearm at high resolution from egocentric view. As existing datasets are either infeasible for egocentric viewpoints or lack background variations, we create a large-scale synthetic dataset with diverse scenarios and collect a real dataset from multi-calibrated camera setup to verify our proposed multi-view image feature fusion strategy. To make the reconstruction physically plausible, we propose two strategies: (i) a coarse-to-fine spectral graph convolution decoder to smoothen the meshes during upsampling and (ii) an optimisation-based refinement stage at inference to prevent self-penetrations. Through extensive quantitative and qualitative evaluations, we show that our framework is able to produce realistic two-hand reconstructions and demonstrate the generalisation of synthetic-trained models to real data, as well as real-time AR/VR applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 563, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 670 + ], + "type": "text", + "content": "3D hand pose estimation is a fundamental problem in various downstream applications including augmented and virtual reality (AR/VR) [19, 20, 25, 26, 47, 58, 63]. Research efforts in designing 3D hand or hand-object reconstruction networks have shown great effectiveness for single-hand pose estimation [15, 22, 27, 59, 60, 66]. However, two-hand pose estimation has received relatively little attention. In this paper, as illustrated in Fig. 1, we focus on high fidelity two-hand reconstruction from multi-view RGB images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 695 + ], + "type": "text", + "content": "With the availability of the InterHand2.6M dataset [46], various methods have been recently proposed. More recent" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 239, + 388, + 319 + ], + "blocks": [ + { + "bbox": [ + 333, + 228, + 359, + 239 + ], + "lines": [ + { + "bbox": [ + 333, + 228, + 359, + 239 + ], + "spans": [ + { + "bbox": [ + 333, + 228, + 359, + 239 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 239, + 388, + 319 + ], + "lines": [ + { + "bbox": [ + 307, + 239, + 388, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 388, + 319 + ], + "type": "image", + "image_path": "f9e1a2ba69ca581c8b2f6bfcbf10cc1e94b1a459c3acf9f157c7b0b419663fdb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 320, + 388, + 400 + ], + "blocks": [ + { + "bbox": [ + 307, + 320, + 388, + 400 + ], + "lines": [ + { + "bbox": [ + 307, + 320, + 388, + 400 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 388, + 400 + ], + "type": "image", + "image_path": "b0579d14bbeee00972e7d57d7879038422fec69acb90e0d9c6b86dd095ac41e6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 391, + 239, + 465, + 309 + ], + "blocks": [ + { + "bbox": [ + 408, + 228, + 457, + 239 + ], + "lines": [ + { + "bbox": [ + 408, + 228, + 457, + 239 + ], + "spans": [ + { + "bbox": [ + 408, + 228, + 457, + 239 + ], + "type": "text", + "content": "Front view" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 239, + 465, + 309 + ], + "lines": [ + { + "bbox": [ + 391, + 239, + 465, + 309 + ], + "spans": [ + { + "bbox": [ + 391, + 239, + 465, + 309 + ], + "type": "image", + "image_path": "cc98eeb63b063af0c96f1f8761486c78d7ee7b75178cd75cfe6472be1737e082.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 391, + 321, + 466, + 399 + ], + "blocks": [ + { + "bbox": [ + 403, + 311, + 460, + 321 + ], + "lines": [ + { + "bbox": [ + 403, + 311, + 460, + 321 + ], + "spans": [ + { + "bbox": [ + 403, + 311, + 460, + 321 + ], + "type": "text", + "content": "Rotated view" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 321, + 466, + 399 + ], + "lines": [ + { + "bbox": [ + 391, + 321, + 466, + 399 + ], + "spans": [ + { + "bbox": [ + 391, + 321, + 466, + 399 + ], + "type": "image", + "image_path": "99ec366cdc10acb3392148a5249e44be9efc4941542eae2061db8dead4202421.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 468, + 239, + 542, + 305 + ], + "blocks": [ + { + "bbox": [ + 477, + 228, + 536, + 239 + ], + "lines": [ + { + "bbox": [ + 477, + 228, + 536, + 239 + ], + "spans": [ + { + "bbox": [ + 477, + 228, + 536, + 239 + ], + "type": "text", + "content": "Zoomed view" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 239, + 542, + 305 + ], + "lines": [ + { + "bbox": [ + 468, + 239, + 542, + 305 + ], + "spans": [ + { + "bbox": [ + 468, + 239, + 542, + 305 + ], + "type": "image", + "image_path": "e653ccdf381ab53d91c7e758fe3a13819461dcb55b3706447a42bb2856be0c27.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 468, + 308, + 541, + 401 + ], + "blocks": [ + { + "bbox": [ + 468, + 308, + 541, + 401 + ], + "lines": [ + { + "bbox": [ + 468, + 308, + 541, + 401 + ], + "spans": [ + { + "bbox": [ + 468, + 308, + 541, + 401 + ], + "type": "image", + "image_path": "1920d8931781f6238e0dd968166ee32da01ec6039b1ed5953d4f6736f8c5c95d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 411, + 547, + 472 + ], + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 472 + ], + "type": "text", + "content": "Figure 1: Our method jointly reconstructs high fidelity two-hand meshes from multi-view RGB image. The input images shown here are from our synthetic dataset, which contains challenging video sequences of two hands rendered into egocentric views." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 484, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 581 + ], + "type": "text", + "content": "methods focused on alleviating the self-similarity problem between interacting hands by exploiting hand part segmentation probability [18], joint visibility [33], cascaded refinement modules [65] or keypoints using Transformer [24]. Although they perform well on complex configurations, their dependence on root joint alignment does not provide absolute root pose recovery when applied to multi-view scenarios, which is especially crucial for interactions in VR." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 714 + ], + "type": "text", + "content": "Moreover, many immersive AR/VR applications require accurate estimation of two hands including extended forearms as it allows for a more realistic and accurate representation of hand movements and gestures. By including the forearm, the orientation and movement of the hand in relation to the arm can provide important contextual information for the user's actions in the virtual environment. Additionally, it can reduce errors and improve the stability of the tracking system, which is important for maintaining immersion and avoiding disorientation in AR/VR applications. However, as in many other areas of computer vision, there" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 231, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 231, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 231, + 714 + ], + "type": "text", + "content": "*This work was done during an internship at Google." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14666" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "type": "text", + "content": "is currently no suitable dataset for a typically supervised deep learning setup. Specifically, existing datasets are either infeasible for egocentric views [21] or lack variations in background and lighting conditions [46]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 121, + 286, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 121, + 286, + 382 + ], + "spans": [ + { + "bbox": [ + 49, + 121, + 286, + 382 + ], + "type": "text", + "content": "In this paper, we propose a spectral graph-based transformer architecture that can reconstruct high resolution two-hand meshes with extended forearms from multi-view RGB images. However, directly extending state-of-the-art methods [39, 40] from single-view to multi-view setting is nontrivial as they consume a substantial number of parameters and are computationally expensive. As pointed out by Cho et al. [8], the current encoder-based transformers [39, 40] have overlooked the importance of efficient token design. In order to retain spatial information in image features and avoid passing redundant global image features to the transformer encoder, we propose a soft-attention-based multiview image feature fusion strategy to obtain region-specific features. Further, as a hand mesh in essence is graph-structured data, we incorporate properties of graph Laplacian from spectral graph theory [10] into the design of our transformer-based network architecture. To obtain realistic hand reconstructions, we leverage a hierarchical graph decoder and propose an optimisation-based refinement stage during inference to prevent self-penetrations. We show that each technical component above contributes meaningfully in our ablation study." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 384, + 286, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 384, + 286, + 550 + ], + "spans": [ + { + "bbox": [ + 49, + 384, + 286, + 550 + ], + "type": "text", + "content": "Besides the expensive computational cost, an additional challenge is the lack of high fidelity 3D mesh ground truth for training such a model. As manually annotating 3D hand meshes on real data is extremely laborious and expensive, we create a large-scale synthetic multi-view dataset containing realistic hand motions from egocentric camera viewpoints rendered under a large variation of background and illumination. To generalise and evaluate on real-world data, we further collected real data from a precisely calibrated multi-view studio with 18 number of high-resolution cameras. We use an automatic approach for registering meshes and obtain ground truth hand poses. Our proposed method trained on both datasets shows robust performance on challenging scenarios and can serve as a strong baseline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 552, + 205, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 552, + 205, + 562 + ], + "spans": [ + { + "bbox": [ + 60, + 552, + 205, + 562 + ], + "type": "text", + "content": "Our contributions are the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 286, + 605 + ], + "type": "text", + "content": "1. We propose a novel end-to-end trainable spectral graph-based transformer for high fidelity two-hand reconstruction from multi-view RGB image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 609, + 286, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 286, + 669 + ], + "type": "text", + "content": "2. We design an efficient soft attention-based multi-view image feature fusion in which the resulting image features are region-specific to segmented hand mesh. We further demonstrate a minimal reduction of " + }, + { + "bbox": [ + 55, + 609, + 286, + 669 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 55, + 609, + 286, + 669 + ], + "type": "text", + "content": " in the model size with this approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 673, + 286, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 286, + 697 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 286, + 697 + ], + "type": "text", + "content": "3. We introduce an optimisation-based method to refine physically-implausible meshes at inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 701, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 701, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 701, + 286, + 712 + ], + "type": "text", + "content": "4. We create a large-scale synthetic multi-view dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 328, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 544, + 95 + ], + "type": "text", + "content": "with high resolution 3D hand meshes and collect real dataset to verify our proposed method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 115, + 391, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 115, + 391, + 127 + ], + "spans": [ + { + "bbox": [ + 308, + 115, + 391, + 127 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 135, + 544, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 135, + 544, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 135, + 544, + 206 + ], + "type": "text", + "content": "Our work tackles the problem of two-hand reconstruction from multi-view images. We first review the literature on Hand reconstruction from RGB images. Then, we focus on the line that leverages Transformer on human body/hand reconstruction. Finally, we provide a brief review on existing Hand datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 209, + 544, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 544, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 544, + 447 + ], + "type": "text", + "content": "Single-Hand Reconstruction. Many approaches have already been proposed for hand pose estimation from either RGB images or depth maps. Here we focus mainly on works that reconstruct full hand meshes. As most earlier methods do not have access to suitable datasets which contain dense 3D ground truths, the parametric hand model MANO [52] provides an alternative as its input parameters can be directly regressed and it contains prior 3D information about hand geometry. Hence, discriminative approaches [2, 5, 27] which directly predict input hand pose and shape parameters from single images have been a popular approach. [45] learns the pose and shape corrective parameters to reconstruct high-fidelity hand meshes with multi-view depth images. Concurrently, non-parametric methods [22, 36] leveraging graph convolutions demonstrate the ability to directly regress 3D hand meshes. Recently, there is an increasing interest in representing object shapes by learning an implicit function [1, 7, 42]. By extending NASA [13], [31] is a keypoint-driven implicit hand representation and [11] disentangles shape, color and pose representations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 449, + 544, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 544, + 628 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 544, + 628 + ], + "type": "text", + "content": "In contrast to the above mentioned approaches, we propose in this work a non-parametric method that can be extended to multi-view settings and reconstruct high-fidelity hand meshes with extended forearms. Further, our approach is not restricted to a single subject, which is the primary constraint of [45]. Recent work on multi-view fusion leverage epipolar geometry [29, 50] to aggregate estimated 2D heatmaps from different views or Feature Transform Layers (FTL) [26, 51] to learn camera geometry-aware latent features. However, these methods require known camera parameters and the feature matching process based on feature similarity is sensitive to occlusion. Our work introduces a soft-attention mechanism coupled with transformer tokenisation which is robust to occlusion and aggregates multiview image features efficiently." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 630, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 630, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 630, + 544, + 712 + ], + "type": "text", + "content": "Two-Hand Pose Estimation. Pose estimation of interacting hands can be broadly classified as discriminative and generative (or hybrid) approaches. Early methods typically follow hybrid approaches by leveraging visual cues detected by discriminative methods, followed by model fitting. [3, 61] apply collision optimisation terms and physical modelling on detected fingertips. [25, 26, 47, 63] ex" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "14667" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 534, + 208 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 534, + 208 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 534, + 208 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 534, + 208 + ], + "type": "image", + "image_path": "29fcfb6ac8d7b1a52209c85878503dce4d6d8eb9bcaf2f870b84fd5c9f5f7d65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "text", + "content": "Figure 2: A schematic illustration of our framework. Given multi-view RGB images, we extract volumetric features " + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "text", + "content": " with a shared CNN backbone. The soft-attention fusion block generates the attention mask " + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "text", + "content": " and finer image features " + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{f}'" + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "text", + "content": " through multiple upsampling and convolution blocks. Region-specific features " + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r" + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "text", + "content": " are computed by first aggregating " + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{f}'" + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "text", + "content": " along the feature channel dimension via the attention mask " + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "text", + "content": ", followed by a max-pooling operation across multi-view images to focus on useful features. Then, we apply mesh segmentation via spectral clustering on template hand meshes and uniformly subsample them to obtain coarse meshes. We perform position encoding by concatenating coarse template meshes to the corresponding region-specific features " + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r" + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "text", + "content": ", i.e. matching colored features to mesh segments. Finally, our multi-layer transformer encoder takes the resulting features as input and outputs a coarse mesh representation " + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c" + }, + { + "bbox": [ + 46, + 211, + 547, + 319 + ], + "type": "text", + "content": " which is then decoded by a spectral graph decoder to produce the final two-hand meshes at target resolution. Here, each hand contains 4023 vertices." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 329, + 288, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 288, + 426 + ], + "type": "text", + "content": "tract image features or keypoints from RGB or depth images and fit hand models with physical constraints. Recent largescale datasets of interacting hands [46] have led to the emergence of fully discriminative methods [24,33,38,53,65] that jointly estimate the 3D joint locations or hand model parameters from a single RGB image. In this work, we target to reconstruct two hands at high resolution while keeping the model complexity low." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 431, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 635 + ], + "type": "text", + "content": "Transformer in 3D Vision. Transformer-based architectures have been gaining increasing popularity in the vision community. Here we focus on methods that reconstruct the human body or hands from RGB images and refer readers to [32] for a detailed survey. In a closely related work, [39] uses cascaded Transformer encoders to reconstruct the human body and hands from a single RGB image and achieves state-of-the-art performance. [40] extends [39] with graph convolutions along the Transformer encoder. [8] improves the efficiency of [39] by disentangling image encoding and mesh estimation via an encoder-decoder architecture. [24] extends \"Detection Transformer\" [6] with hand-object pose estimations. While these works are targeted at a single image and their extension to the multi-view setting is nontrivial due to large number of learnable parameters, our architecture is designed to efficiently reconstruct two hands at high resolution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": "Hand Pose Datasets. The success of discriminative methods depend on the availability and variability of hands interaction datasets. [46] introduces the large-scale dataset InterHand2.6M with closely interacting hand motions using a semi-automatic annotation process with multiple high-resolution RGB cameras. However, it contains minimal" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 329, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 510 + ], + "type": "text", + "content": "background and lighting variation. FreiHAND [67] includes wider background variation but the dataset is limited to a single hand and third-person views. Therefore, the above datasets are not suitable for the egocentric two-hand reconstruction task. The available egocentric datasets are limited by either visible markers [21] or constrained lab environments [37]. This motivates us to create a large-scale egocentric synthetic dataset with improved environment and lighting variations. To validate our proposed multi-view fusion strategy, we further collect a real dataset with more challenging camera viewing angles. Altogether, both synthetic and real datasets contain diverse egocentric multi-view and -frame data points with multiple subjects. Our architecture demonstrates promising performance on both datasets and constitutes a strong baseline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 524, + 362, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 524, + 362, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 524, + 362, + 537 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "As shown in Fig. 2, our architecture first passes " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " number of multi-view RGB input images " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{N \\times 224 \\times 224 \\times 3}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " to a shared CNN backbone to extract volumetric features " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbb{R}^{N \\times 7 \\times 7 \\times 2048}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ", i.e. features before the global average pooling layer for ResNet [28]. The volumetric features are then fed into a soft attention-based multi-view feature encoder and output " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " region-specific features " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r \\in \\mathbb{R}^{K \\times C}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " refers to the feature channel size. The Transformer encoder takes " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " together with template hand meshes " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m}' \\in \\mathbb{R}^{V' \\times 3}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " and outputs a coarse mesh representation " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c \\in \\mathbb{R}^{V' \\times F}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ". Finally, the spectral graph decoder generates hand meshes " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\mathbb{R}^{V \\times 3}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " by upsampling on " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "V \\gg V'" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ". With slight abuse of notation, " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " can either be two-hand or single-hand depending on the application." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14668" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "In the following, we detail the multi-view image feature encoder in Section 3.2, the spectral graph convolution decoder in Section 3.3 and the loss used for training in Section 3.4. We also present an optimisation-based refinement procedure at inference time in Section 3.5." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 141, + 115, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 141, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 115, + 152 + ], + "type": "text", + "content": "3.1. Notations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": "A 3D mesh can be represented as an undirected graph " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V},\\mathcal{E},\\mathbf{A})" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": " is a node set and " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": " is an edge set. An adjacency matrix " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": " encodes information of pairwise relations between nodes. A degree matrix " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{D} \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": " is a diagonal matrix whose diagonal element " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{ii} = \\sum_{j} \\mathbf{A}_{ij}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": " refers to the degree value of each node. An essential operator in spectral graph theory [10] is the graph Laplacian " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": ", whose definition is " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = \\mathbf{D} - \\mathbf{A}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = \\mathbf{U} \\boldsymbol{\\Lambda} \\mathbf{U}^T" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": " where the graph Laplacian can be diagonalised by the Fourier basis " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{U} = [\\mathbf{u}_1, \\dots, \\mathbf{u}_{|\\mathcal{V}|}] \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Lambda} = \\text{diag}([\\lambda_1, \\dots, \\lambda_{|\\mathcal{V}|}]) \\in \\mathbb{R}^{|\\mathcal{V}| \\times |\\mathcal{V}|}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{u}_i\\}_{i=1}^{|\\mathcal{V}|}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": " are the eigenvectors and " + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "\\{\\lambda_i\\}_{i=1}^{|\\mathcal{V}|}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": " are the non-negative eigenvalues of graph Laplacian (" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "inline_equation", + "content": "0 = \\lambda_1 \\leq \\dots \\leq \\lambda_{|\\mathcal{V}|}" + }, + { + "bbox": [ + 46, + 159, + 288, + 323 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 331, + 236, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 236, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 236, + 342 + ], + "type": "text", + "content": "3.2. Multi-View Image Feature Encoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 349, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 349, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 45, + 349, + 287, + 445 + ], + "type": "text", + "content": "To extend transformer-based architecture to multi-view settings and prevent concatenating global image features in an overly-duplicated way as in [39], we propose a simple soft-attention fusion strategy to better aggregate features across multiple views and attend to different hand parts via mesh segmentation. The resulting " + }, + { + "bbox": [ + 45, + 349, + 287, + 445 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 349, + 287, + 445 + ], + "type": "text", + "content": " region-specific features " + }, + { + "bbox": [ + 45, + 349, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r" + }, + { + "bbox": [ + 45, + 349, + 287, + 445 + ], + "type": "text", + "content": " defined by spectral clustering are fed to our Transformer encoder to obtain a coarse mesh representation " + }, + { + "bbox": [ + 45, + 349, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c" + }, + { + "bbox": [ + 45, + 349, + 287, + 445 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": "Soft-Attention Fusion. Given volumetric features " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": ", we do not apply any pooling operations to avoid losing spatial information as shown by our experiments (see Table 3). Instead, we propose to aggregate multi-view features with a soft-attention mask. We first obtain a finer representation of " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}' \\in \\mathbb{R}^{N \\times (H \\times W) \\times C}" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": ", by feeding it through two blocks each comprised of 2D upsampling with bilinear interpolation, " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": " convolution layers, batch-normalisation [30] and ReLU. The soft-attention mask " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{N \\times (H \\times W) \\times K}" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": " is obtained by applying 1) " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": " convolutional filters to reduce the feature channel of " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}'" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": " and 2) spatial soft arg-max which determines the image-space point of maximal activation in each " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": ". At this stage, we can compute per-frame features " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}'' \\in \\mathbb{R}^{N \\times K \\times C}" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}'' = \\mathcal{M}^{\\mathrm{T}}\\mathbf{f}'" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": ". We finally obtain region-specific features " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r \\in \\mathbb{R}^{K \\times C}" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": " by max-pooling along the second dimension of " + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}''" + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": ". There are two intuitions to this design: 1) the corresponding feature with higher attention weight contributes more to the final feature representation spatially and 2) max-pooling allows efficient feature selection across multiple views and it does not overfit to any multi-camera configurations as CNN filters are shared across all views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "Mesh Segmentation via Spectral Clustering. Inspired by [41], we perform spectral clustering to obtain a 3D mesh segmentation. Different from [41], we apply the eigendecomposition of the graph Laplacian " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " instead of the affinity matrix (which encodes pairwise point affinities with exponential kernel), followed by " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "-means clustering into " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " clusters. With this approach, one can efficiently segment any template model mesh without manual effort. In addition, as shown in Fig. 3, the segmented mesh does not exhibit clear-cut boundaries, and certain clusters are scattered throughout the entire surface. This is in spirit similar to mask vertex modeling occlusions in [39] by encouraging the transformer to consider other relevant vertex queries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "content": "Transformer Encoder. We now concatenate the " + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "content": "-dimensional region-specific image features to the corresponding " + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "content": " clusters of the segmented template hand mesh. The resulting features " + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t \\in \\mathbb{R}^{V' \\times (C + 3)^1}" + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "content": " are fed into a multi-layer transformer encoder with progressive dimensionality reduction as described in [39]. The output is a coarse mesh representation " + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c \\in \\mathbb{R}^{V' \\times F}" + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "inline_equation", + "content": "F = (C + 3)/2^n" + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 233, + 545, + 329 + ], + "type": "text", + "content": " layers of transformer encoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 335, + 546, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 546, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 546, + 502 + ], + "type": "text", + "content": "Theoretical Motivations. Since there do not exist explicit coordinate systems as in grid graphs, aligning nodes of highly irregular graphs in nature is a non-trivial problem. Recent studies [16, 17, 35] attempted to encode positional information by leveraging the spectral domain in a way that nearby nodes have similar values and distant nodes have different values. This can be achieved as the eigenvectors of the graph Laplacian can be interpreted as the generalised concepts of sinusoidal functions of positional encoding in Transformers [62]. For instance with any given graph, nodes close to each other can be assigned values of similar positional features as the smaller the eigenvalue of the graph Laplacian " + }, + { + "bbox": [ + 304, + 335, + 546, + 502 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 304, + 335, + 546, + 502 + ], + "type": "text", + "content": " (closer to 0) the smoother the coordinates of the corresponding eigenvector " + }, + { + "bbox": [ + 304, + 335, + 546, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i}" + }, + { + "bbox": [ + 304, + 335, + 546, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 512, + 443, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 512, + 443, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 512, + 443, + 525 + ], + "type": "text", + "content": "3.3. Spectral Graph Decoder" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 531, + 545, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 627 + ], + "type": "text", + "content": "We find that simply relying on fully-connected layers to upsample meshes to target resolution is insufficient as this process introduces instability and disruption to the mesh, even with mesh regularisation loss terms [27] (see Fig. 4). Therefore, we propose to couple the fully-connected layers for upsampling with spectral filtering as meshes can be treated as graph signals " + }, + { + "bbox": [ + 304, + 531, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c = (f_1,\\dots ,f_{V'})\\in \\mathbb{R}^{V'\\times F}" + }, + { + "bbox": [ + 304, + 531, + 545, + 627 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 304, + 531, + 545, + 627 + ], + "type": "inline_equation", + "content": "V^{\\prime}" + }, + { + "bbox": [ + 304, + 531, + 545, + 627 + ], + "type": "text", + "content": " vertices with " + }, + { + "bbox": [ + 304, + 531, + 545, + 627 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 531, + 545, + 627 + ], + "type": "text", + "content": "-dimensional features for batch size of 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "type": "text", + "content": "Spectral Filtering. As spectral graph convolution can be defined via point-wise products in the transformed Fourier space, graph signals " + }, + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c" + }, + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "type": "text", + "content": " filtered by " + }, + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "type": "text", + "content": " can then be expressed as " + }, + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{U}g_{\\theta}(\\boldsymbol {\\Lambda})\\mathbf{U}^{T}\\mathbf{f}_{c}" + }, + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "type": "inline_equation", + "content": "g_{\\theta}(\\boldsymbol {\\Lambda})" + }, + { + "bbox": [ + 304, + 632, + 545, + 693 + ], + "type": "text", + "content": " is a diagonal matrix with Fourier coefficients. To ensure that the spectral filter corre" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 476, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 476, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 476, + 713 + ], + "type": "inline_equation", + "content": "{}^{1} + 3" + }, + { + "bbox": [ + 317, + 702, + 476, + 713 + ], + "type": "text", + "content": " refers to the 3D positions of mesh template." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14669" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 67, + 265, + 180 + ], + "blocks": [ + { + "bbox": [ + 75, + 67, + 265, + 180 + ], + "lines": [ + { + "bbox": [ + 75, + 67, + 265, + 180 + ], + "spans": [ + { + "bbox": [ + 75, + 67, + 265, + 180 + ], + "type": "image", + "image_path": "5d4fe74583be375d47f5db27e8e4eca094a30c92d8beea57350fd175229e2782.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "type": "text", + "content": "Figure 3: Illustration of mesh segmentation via spectral clustering. Here, we have chosen " + }, + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "type": "inline_equation", + "content": "K = 7" + }, + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "type": "text", + "content": " clusters and each has its own color for a single right-handed template mesh. As shown, there are no rigid boundaries across the mesh. To prepare inputs " + }, + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "type": "text", + "content": " to transformer, we decrease the size of this mesh (similarly for left hand) uniformly by a factor of 10 through subsampling, i.e. " + }, + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "type": "inline_equation", + "content": "V' = V / 10" + }, + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "type": "text", + "content": " and concatenate with the corresponding region-specific features " + }, + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r" + }, + { + "bbox": [ + 46, + 190, + 289, + 288 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 288, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 288, + 354 + ], + "type": "text", + "content": "sponds to a meaningful convolution on the graph, a natural solution is to parameterise based on the eigenvalues of the Laplacian. Here we have used the Chebyshev polynomial parametrisation of " + }, + { + "bbox": [ + 46, + 293, + 288, + 354 + ], + "type": "inline_equation", + "content": "g_{\\theta}(\\mathbf{L})" + }, + { + "bbox": [ + 46, + 293, + 288, + 354 + ], + "type": "text", + "content": " for fast computation and readers are referred to [12] for more details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "text", + "content": "Architecture. Similar to [9, 22], our decoder is based on a hierarchical architecture where the mesh is recovered by using fully-connected and graph convolution layers for upsampling. We followed [12, 22] to pre-computed coarse graphs using [14] and used the third-order polynomial in the Laplacian. Given the coarse mesh representation " + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_c\\in \\mathbb{R}^{V'\\times F}" + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "text", + "content": ", the key idea here is to smooth out the " + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "text", + "content": "-dimensional values after upsampling on " + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "inline_equation", + "content": "V^{\\prime}" + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "text", + "content": ". Note that the fundamental difference between this approach and the one in [9, 22] is that they have a hierarchical architecture on both " + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "inline_equation", + "content": "V^{\\prime}" + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "text", + "content": " dimensions, we apply only spectral filtering and keep " + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "inline_equation", + "content": "F = 3" + }, + { + "bbox": [ + 46, + 358, + 289, + 515 + ], + "type": "text", + "content": " constant which massively reduces the model size while maintaining performance (see Table 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 518, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 288, + 592 + ], + "type": "text", + "content": "**Discussions.** Recall that in Section 3.2, we leverage the properties of Laplacian eigenvectors to perform spectral clustering. Here we can interpret from a signal processing perspective where the Laplacian eigenvectors define signals that vary smoothly across the graph, with the smoothest signals indicating the coarse community structure of the mesh." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 597, + 111, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 111, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 111, + 609 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 615, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 288, + 676 + ], + "type": "text", + "content": "Following [39, 40], our model can be trained end-to-end with L1 losses on 3D mesh vertices and 2D re-projection using the predicted camera parameters to improve image-mesh alignment. In addition, we apply an edge length regularisation " + }, + { + "bbox": [ + 46, + 615, + 288, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{edge}" + }, + { + "bbox": [ + 46, + 615, + 288, + 676 + ], + "type": "text", + "content": " [27] to encourage smoothness of the mesh:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 92, + 681, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 681, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 92, + 681, + 287, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e d g e}} (\\mathbf {m}) = \\frac {1}{| \\mathcal {E} _ {L} |} \\sum_ {l \\in \\mathcal {E} _ {L}} \\left| l ^ {2} - \\mu \\left(\\mathcal {E} _ {L} ^ {2}\\right) \\right|, \\tag {1}", + "image_path": "5be14038e8e008e0861abbaba1a9d6d9c739a800692a1726f7a16f0df437460a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 547, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 547, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_L" + }, + { + "bbox": [ + 305, + 72, + 547, + 109 + ], + "type": "text", + "content": " refers to the set of edge lengths, defined as the L2 norms of all edges and " + }, + { + "bbox": [ + 305, + 72, + 547, + 109 + ], + "type": "inline_equation", + "content": "\\mu (\\mathcal{E}_L^2)" + }, + { + "bbox": [ + 305, + 72, + 547, + 109 + ], + "type": "text", + "content": " is the average of the squared edge lengths." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 114, + 470, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 114, + 470, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 470, + 126 + ], + "type": "text", + "content": "3.5. Mesh Refinement at Inference" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 133, + 547, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 337 + ], + "type": "text", + "content": "There are two main approaches in the literature for producing realistic mesh reconstruction: learning-based and optimisation-based methods. Learning-based methods [27, 45] approach the problem by proposing various repulsive losses that penalise penetration during training. However, their generalisation ability to other meshes is limited due to the need for mesh-specific pre-computation. In particular, they require manual selection of areas of interest (i.e. fingertips and palm) which does not prevent other forms of self-penetrations, such as finger-finger. More importantly, the presence of interpenetration at test time shows that the model is not able to learn the physical rule implicitly [27, 59]. On the other hand, recent optimisation-based methods [23, 60] leverage contact maps to refine meshes at inference. However, they rely heavily on accurate contact map estimations and are sensitive to initialisation as contact optimisation is local." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 341, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 547, + 498 + ], + "type": "text", + "content": "Proposed Method. To avoid penetrations, we extend the repulsion loss from [27] into an optimisation-based strategy which does not require any form of pre-computation and is more generalisable to other meshes. To identify hand vertices that contribute to collision, we follow [27,43] and cast rays from each vertex and count the number of surface intersections. If the number is odd, it indicates penetration. After obtaining the collision mask " + }, + { + "bbox": [ + 304, + 341, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_C" + }, + { + "bbox": [ + 304, + 341, + 547, + 498 + ], + "type": "text", + "content": ", we compute the nearest point in the source mesh " + }, + { + "bbox": [ + 304, + 341, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 304, + 341, + 547, + 498 + ], + "type": "text", + "content": " with respect to the set of collision points. If the point and its nearest corresponding point have a different normal, we compute their distance as loss. We minimise this collision loss " + }, + { + "bbox": [ + 304, + 341, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{collision}" + }, + { + "bbox": [ + 304, + 341, + 547, + 498 + ], + "type": "text", + "content": " with as-rigid-as-possible (ARAP) [56] regularisation on the mesh shape:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 340, + 503, + 545, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 503, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 340, + 503, + 545, + 529 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o l l i s i o n}} (\\mathbf {m}, \\mathcal {M} _ {C}) = \\sum_ {v \\in \\mathcal {V}} \\mathcal {M} _ {C} \\cdot d (v, \\mathcal {V}), \\tag {2}", + "image_path": "99841cd39d68348567696c57ff88599238e9adc43747bc1250aadafafafd98a8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_C = \\mathbb{1}_{v\\in \\mathrm{Int}(\\mathcal{V})}" + }, + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "type": "text", + "content": " indicates which vertices belong to the interior of the mesh and " + }, + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "type": "inline_equation", + "content": "d(v,\\mathcal{V}) = \\min_{v'\\in \\mathcal{V}}\\| v - v'\\| _2" + }, + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "type": "text", + "content": " denotes distances from point " + }, + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "type": "inline_equation", + "content": "v^{\\prime}" + }, + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "type": "text", + "content": " to set " + }, + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 535, + 547, + 584 + ], + "type": "text", + "content": ". We show that our method is able to remove self-penetration in Fig. 5 and 9." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 593, + 388, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 593, + 388, + 607 + ], + "spans": [ + { + "bbox": [ + 306, + 593, + 388, + 607 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "Implementation Details. We implement our method in TensorFlow. We train all parts of the network simultaneously with the Adam optimiser [34] using a learning rate of " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": " when training on the synthetic dataset and " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": " when fine-tuning on the real dataset. We use ResNet [28] pre-trained on ImageNet [54] for our CNN backbone. For computing region-specific features " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "K = 7" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "C = 256" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14670" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 67, + 276, + 236 + ], + "blocks": [ + { + "bbox": [ + 69, + 67, + 276, + 236 + ], + "lines": [ + { + "bbox": [ + 69, + 67, + 276, + 236 + ], + "spans": [ + { + "bbox": [ + 69, + 67, + 276, + 236 + ], + "type": "image", + "image_path": "2a1860447dab3f115e4823d25edd80ff279df4fcbc0119f8b545c8aebde37774.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 289, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 289, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 289, + 304 + ], + "type": "text", + "content": "Figure 4: Qualitative comparison with METRO [39]. We show that relying on fully-connected layers to upsample meshes is inadequate for high-resolution mesh reconstruction. In contrast, our spectral graph decoder can accurately capture intricate surface features like nails." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 312, + 276, + 468 + ], + "blocks": [ + { + "bbox": [ + 61, + 312, + 276, + 468 + ], + "lines": [ + { + "bbox": [ + 61, + 312, + 276, + 468 + ], + "spans": [ + { + "bbox": [ + 61, + 312, + 276, + 468 + ], + "type": "image", + "image_path": "acfbc6247fac97551120a7a16e353e48a5b2e870a763664886c74b713191ad72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 475, + 287, + 501 + ], + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 501 + ], + "type": "text", + "content": "Figure 5: Qualitative example of mesh refinement at inference. More examples are available in Fig. 9." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": "Synthetic Dataset Creation. A large-scale multi-view egocentric dataset with challenging interacting hand motion is required to train our pipeline. However, existing datasets are either infeasible for egocentric views or lack variations in terms of background and lighting conditions, which are crucial for AR/VR applications. In addition, the ground-truth mesh in [46] contains " + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "inline_equation", + "content": "5mm" + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": " fitting error which is sufficient to cause inaccuracy in hand tracking or misalignment when interacting with virtual objects. Therefore, it is not suitable to validate our approach. To this end, we create a new large-scale synthetic egocentric dataset with two high-fidelity 3D hand meshes. In particular, we purchase commercial 3D hand models for both left and right hands. Each of them contains 4023 vertices and 4008 quad faces. By dividing each quad into two triangles, they can be decomposed into 8016 triangular faces. The size and textures of the hand" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 71, + 545, + 161 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 545, + 161 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 545, + 161 + ], + "type": "image", + "image_path": "69bb32fe51f1cfc2398bfb003159c040f7a2dd46a8064fdceb339e33bcd5c9e3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 170, + 547, + 206 + ], + "lines": [ + { + "bbox": [ + 305, + 170, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 547, + 206 + ], + "type": "text", + "content": "Figure 6: Qualitative examples of our synthetic dataset. The scenes are rendered with large variations in lighting, texture and background." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 212, + 547, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 212, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 547, + 369 + ], + "type": "text", + "content": "models can vary, and it can also be rigged with 3D joint locations. We apply photorealistic textures as well as natural lighting using High-Dynamic-Range (HDR) images. To create realistic hand motions, we generate 100 poses and interpolate between pairs of these poses randomly over 1000 frame sequences, i.e. pose pair is swapped every 10 frames. Then, we render hands onto 480 4K backgrounds using the Cycles renderer [49]. Our synthetic dataset comprises 1M data points. Each data point consists of 3D annotations for two hands, rendered in two egocentric views to simulate a customised camera headset scenario. We divide the entire dataset into 4 groups and use " + }, + { + "bbox": [ + 304, + 212, + 547, + 369 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 212, + 547, + 369 + ], + "type": "text", + "content": " randomly for testing and provide additional image examples in Fig. 6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 481 + ], + "type": "text", + "content": "Real Dataset Collection. Our real hand data is captured from a multi-view stereo system with 18 synchronised Z-Cams. We build a NeRF-based [4, 48] reconstruction pipeline that simultaneously reconstructs and disentangles the foreground and the background, from which a meshing module extracts hand meshes as ground-truth target mesh. We then apply a mesh and tetrahedral registration approach [55] that registers a template hand model [64] to get a well-registered mesh for each reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 548, + 714 + ], + "type": "text", + "content": "Baselines. We compare with the state-of-the-art transformer-based architecture METRO [39] that is explicitly designed for human body mesh reconstruction. We follow their official implementation and use 3 transformer encoder layers. We attempted to include other strong baselines to compare with. However, given our challenging multi-view settings on reconstructing high resolution two-hand meshes, existing methods are constrained by either 1) relying on intermediate 3D pose supervision [9,44] or 2) directly regressing a low resolution parametric hand model [2,5,27,59,67]. In addition, the direct extension of METRO [39], Mesh Graphormer [40], is infeasible to extend to multi-view settings as they tokenised volumetric features which increases the computational complexity of each transformer layer quadratically [8]. Therefore, by considering that the performance gain is minor, we pick METRO [39] as a strong baseline and study in-depth. We provide more details about the baselines in the supplementary materials." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14671" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 128, + 273, + 198 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 288, + 118 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 288, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 288, + 118 + ], + "type": "text", + "content": "Table 1: Error rates on our synthetic dataset. Parametric baseline is trained and tested on right hand only. Our proposed method achieves strong performance with less than half the METRO model size." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 128, + 273, + 198 + ], + "lines": [ + { + "bbox": [ + 60, + 128, + 273, + 198 + ], + "spans": [ + { + "bbox": [ + 60, + 128, + 273, + 198 + ], + "type": "table", + "html": "
Hand error# Params
Parametric baseline24.540.9M
METRO [39]7.09116.1M
Ours (w/o graph decoder)3.7275.2M
Ours1.3858.3M
", + "image_path": "29679ff5c7576c2c51a093ad0d78c522043e4529dd55b2c86b9ec5d5103bcdfc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 228, + 272, + 270 + ], + "blocks": [ + { + "bbox": [ + 78, + 207, + 256, + 219 + ], + "lines": [ + { + "bbox": [ + 78, + 207, + 256, + 219 + ], + "spans": [ + { + "bbox": [ + 78, + 207, + 256, + 219 + ], + "type": "text", + "content": "Table 2: Error rates (in mm) on FreiHAND." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 228, + 272, + 270 + ], + "lines": [ + { + "bbox": [ + 66, + 228, + 272, + 270 + ], + "spans": [ + { + "bbox": [ + 66, + 228, + 272, + 270 + ], + "type": "table", + "html": "
METRO [39]Graphormer [40]Ours
PA-MPVPE ↓6.75.95.5
PA-MPJPE ↓6.86.05.6
", + "image_path": "fa04e2f26bb2514e2ef1caf8f199faee728903f8e002eb1c9366b45a1dbf7b07.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 284, + 287, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 333 + ], + "type": "text", + "content": "Evaluation Metric. We report the Mean-Per-Vertex-Error (MPVE) in " + }, + { + "bbox": [ + 46, + 284, + 287, + 333 + ], + "type": "inline_equation", + "content": "mm" + }, + { + "bbox": [ + 46, + 284, + 287, + 333 + ], + "type": "text", + "content": " to evaluate the hand reconstruction error. MPVE measures the mean Euclidean distances between the ground-truth vertices and the predicted vertices." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 337, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 661 + ], + "type": "text", + "content": "Quantitative Comparison. We perform a quantitative comparison for the two-hand reconstruction task on our synthetic dataset. As previously mentioned, comparing with other methods is not straightforward since the majority of existing methods focus on single-view settings. We extend METRO [39] to the multi-view setting by applying max-pool to image features and concatenating them with vertex queries. For fair comparisons, we share the same hyperparameter setting for the multi-layer transformer encoder and report the results in Table 1. In this experiment, we downsample the template hand mesh by 10 times. To provide more context regarding the difficulty of our synthetic dataset, we also experiment on a parametric baseline. We create a parametric hand model which uses the 200-dimensional PCA (principal component analysis) subspace from " + }, + { + "bbox": [ + 46, + 337, + 287, + 661 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 337, + 287, + 661 + ], + "type": "text", + "content": " of the training data. We use ResNet-50 as the backbone for all models in Table 1 and the parametric baseline has an MLP head to predict the input parameters to recover the hand mesh. Our method significantly outperforms both baseline methods with less than half the model size of METRO. We show that our soft-attention feature fusion strategy coupled with mesh segmentation using spectral clustering can effectively reduce the feature channel size from 2048 to 256 without performance drop. In addition, the performance of the parametric baseline is in-line with existing single-hand benchmarks. Lastly, we report results on single-view benchmark FreiHAND [67] in Table 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Ablation Study. To motivate our design choices, we present a quantitative evaluation of our method with various components disabled. We validate that each of our proposed technical component contributes meaningfully." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 128, + 533, + 219 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 119 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 119 + ], + "type": "text", + "content": "Table 3: Performance of different multi-view fusion strategies. We report hand error for both settings. " + }, + { + "bbox": [ + 304, + 70, + 545, + 119 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 70, + 545, + 119 + ], + "type": "text", + "content": " refers to the number of clusters for template hand mesh. Note that we do not include spectral filtering in the graph decoder here." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 128, + 533, + 219 + ], + "lines": [ + { + "bbox": [ + 318, + 128, + 533, + 219 + ], + "spans": [ + { + "bbox": [ + 318, + 128, + 533, + 219 + ], + "type": "table", + "html": "
Single-viewMulti-view
METRO [39]10.87-
METRO [39] + avg. pool-8.71
METRO [39] + max pool-7.09
Ours (K=1)-6.59
Ours (K=4)-4.79
Ours (K=7)-3.72
", + "image_path": "58e98dd5f446fab194a315333154e6b6d05311659318b9340a3b8e5a86d7df5e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 313, + 289, + 537, + 346 + ], + "blocks": [ + { + "bbox": [ + 304, + 231, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 280 + ], + "type": "text", + "content": "Table 4: Ablations on multi-view feature fusion. We report hand error on 3 experimental settings ((a)-(c)) with different feature fusion strategies. We keep the same transformer and spectral graph decoder for all fusion strategies." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 289, + 537, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 537, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 537, + 346 + ], + "type": "table", + "html": "
(a)(b)(c)
Direct concatenation11.313.914.7
Max pool9.311.212.5
Soft-attention + max pool6.77.27.4
", + "image_path": "6da4621b85caa801ee02d6d0d763ee95e88bdeb6c5fb15c0a76f8277e51f9fa6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 312, + 380, + 537, + 415 + ], + "blocks": [ + { + "bbox": [ + 333, + 358, + 518, + 370 + ], + "lines": [ + { + "bbox": [ + 333, + 358, + 518, + 370 + ], + "spans": [ + { + "bbox": [ + 333, + 358, + 518, + 370 + ], + "type": "text", + "content": "Table 5: Ablations of different spectral filters." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 380, + 537, + 415 + ], + "lines": [ + { + "bbox": [ + 312, + 380, + 537, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 380, + 537, + 415 + ], + "type": "table", + "html": "
ggauglapgcheb3gcheb4gcheb5gcheb6
Hand error1.561.451.381.471.381.39
", + "image_path": "675b4d13a61b34fc061d505876cf4926afc2c93e66f2ef579c5e896f9954e9dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "content": "Effects of Multi-View Feature Fusion. Table 3 shows the results of varying number of spectral clusters " + }, + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "content": " (full results in supp.). The combination of soft-attention fusion and mesh segmentation consistently improves the performance. As our synthetic dataset contains only two views, we further verify our multi-view fusion strategy on more challenging viewing angles captured in our real dataset. We also demonstrate that our method does not overfit to camera setup when tested on unseen camera views in Table 4. In these experiments, we divide the 18 camera views into 2 groups, i.e. the first group contains the first 15 camera views and the second group contains the last 3 unseen camera views for evaluation. Note that the evaluation group contains the only egocentric view. We consider 3 experimental settings by varying number of camera views present in the training data: (a) 15 views, (b) 6 views and (c) 3 views in Table 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Effects of Spectral Filters. We experiment with three commonly-used spectral filters: Gaussian " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{\\text{gau}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", Laplacian " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{\\text{lap}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and Chebyshev filters " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{\\text{cheb}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " with varying order of polynomials in Table 5. We provide implementation details and additional qualitative examples in the supplementary. We choose " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{\\text{cheb}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " for efficiency and do not find increasing order of polynomials improves performance further." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14672" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 113, + 196 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 113, + 196 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 113, + 196 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 113, + 196 + ], + "type": "image", + "image_path": "7cac81a46ce0877043d538abf13499afa6d8eab91dbf7c91d1e98d0bf09b38ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 545, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 545, + 231 + ], + "type": "text", + "content": "Figure 7: Qualitative examples on our synthetic dataset. Our model shows robustness to various extreme self-occlusions (left), lighting conditions (middle) and scenes (right). Fine mesh reconstruction details are well-preserved in all scenarios." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 114, + 72, + 183, + 196 + ], + "blocks": [ + { + "bbox": [ + 114, + 72, + 183, + 196 + ], + "lines": [ + { + "bbox": [ + 114, + 72, + 183, + 196 + ], + "spans": [ + { + "bbox": [ + 114, + 72, + 183, + 196 + ], + "type": "image", + "image_path": "dfaf1ac16563ea1b5ac874d3ffc7df376f2e245ea5aa854e257f4724562748b8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 192, + 71, + 254, + 195 + ], + "blocks": [ + { + "bbox": [ + 192, + 71, + 254, + 195 + ], + "lines": [ + { + "bbox": [ + 192, + 71, + 254, + 195 + ], + "spans": [ + { + "bbox": [ + 192, + 71, + 254, + 195 + ], + "type": "image", + "image_path": "c3e2e913e07322fa1584d7505a4ba7730d23218db64582dca0ce4db0312ad52d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 255, + 71, + 330, + 195 + ], + "blocks": [ + { + "bbox": [ + 255, + 71, + 330, + 195 + ], + "lines": [ + { + "bbox": [ + 255, + 71, + 330, + 195 + ], + "spans": [ + { + "bbox": [ + 255, + 71, + 330, + 195 + ], + "type": "image", + "image_path": "73f8481e5c0ba1ec12e61631c89f092bdaecc93662fa24d2d8c292ae8d96c6bc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 378, + 195 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 378, + 195 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 378, + 195 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 378, + 195 + ], + "type": "image", + "image_path": "205fd080caea2fe4bdf645b47b8c68f9ddd29a230e944aa435a48d9c4e916299.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 379, + 72, + 441, + 195 + ], + "blocks": [ + { + "bbox": [ + 379, + 72, + 441, + 195 + ], + "lines": [ + { + "bbox": [ + 379, + 72, + 441, + 195 + ], + "spans": [ + { + "bbox": [ + 379, + 72, + 441, + 195 + ], + "type": "image", + "image_path": "6164d9ad3781182432cb66835e5bee09a676acc1d7640f65b0d994509f69f285.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 443, + 72, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 443, + 72, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 443, + 72, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 443, + 72, + 545, + 194 + ], + "type": "image", + "image_path": "2dc6ec83cfec752ac3d06b59c9bed8cca42a4c67b742c9e023509deb2037624d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 60, + 282, + 273, + 348 + ], + "blocks": [ + { + "bbox": [ + 46, + 237, + 288, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 288, + 275 + ], + "type": "text", + "content": "Table 6: Ablations of different backbones and hyperparameters. We denote " + }, + { + "bbox": [ + 46, + 237, + 288, + 275 + ], + "type": "inline_equation", + "content": "P_{cnn}" + }, + { + "bbox": [ + 46, + 237, + 288, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 237, + 288, + 275 + ], + "type": "inline_equation", + "content": "P_{total}" + }, + { + "bbox": [ + 46, + 237, + 288, + 275 + ], + "type": "text", + "content": " to be the number of parameters for CNN backbone and total model, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 282, + 273, + 348 + ], + "lines": [ + { + "bbox": [ + 60, + 282, + 273, + 348 + ], + "spans": [ + { + "bbox": [ + 60, + 282, + 273, + 348 + ], + "type": "table", + "html": "
BackbonePcnnV'CErrorPtotal
ResNet-5023.5M8042561.3858.3M
EfficientNet-B05.3M1602564.1242.8M
EfficientNet-B05.3M1601284.9637.8M
EfficientNet-B05.3M80646.8934.2M
", + "image_path": "e3b017d84d011132dec8f5d8f059a5cf232961209f7e17b062c9b9d35a359c9d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 77, + 358, + 251, + 537 + ], + "blocks": [ + { + "bbox": [ + 77, + 358, + 251, + 537 + ], + "lines": [ + { + "bbox": [ + 77, + 358, + 251, + 537 + ], + "spans": [ + { + "bbox": [ + 77, + 358, + 251, + 537 + ], + "type": "image", + "image_path": "f7e5b193a1879bb49dd8641d67ac64821c6d025895197bd980a3b0446e6261a5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 546, + 288, + 583 + ], + "lines": [ + { + "bbox": [ + 46, + 546, + 288, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 288, + 583 + ], + "type": "text", + "content": "Figure 8: Training examples of our real dataset (top) and 3D hand mesh estimation results on in-the-wild image (bottom). The resulting mesh contains 3745 vertices." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Synthetic to Real Transfer. Large-scale synthetic dataset can be used to pre-train models in the absence of suitable real datasets. In the following, we demonstrate a use case of deploying our pre-trained model which was trained on synthetic data to general real images. First, we train on a subset of our real dataset and freeze all part of the pre-trained model except CNN backbone. As the registered meshes for our real dataset have a different mesh topology, we drop the pink cluster of our original template mesh (shown in Fig. 3) before training. This approach is analogous to" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 275 + ], + "type": "text", + "content": "MANO-based 3D hand mesh estimation methods [2, 5] and has been demonstrated in [45]. We show that our model can generalise to in-the-wild-images in Fig. 8." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": "Model Compression. We are interested in finding the minimal size to which model can be compressed while maintaining the METRO baseline performance in Table 1. First, we perform computational complexity analysis on the self-attention layer in the Transformer encoder. There are two key steps to compute self-attention: 1) linear projection to " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": "-dimensional query, key and value matrices from " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": " vertex queries requires " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "O(V'C^2)" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": " and softmax operation for layer output requires another " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "O(V'^2C)" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": ". The total computational complexity of each transformer layer is therefore quadratic no matter whether " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": " dominates. Recall that as we uniformly subsample template hand meshes by 10 times to obtain " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "V' = 804" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": " is empirically set to 256 in earlier sections, " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": " contributes more to overall complexity as it is dominantly larger. In the following experiments, we gradually decrease " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": " before " + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 278, + 546, + 495 + ], + "type": "text", + "content": " while using a variant of EfficientNet [57]. We present main results in Table 6, while full results are provided in the supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 500, + 379, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 379, + 512 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 379, + 512 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 517, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 661 + ], + "type": "text", + "content": "In this paper, we have proposed a novel spectral graph-based transformer framework which reconstructs high-fidelity two-hand meshes from egocentric views. The main idea behind this study was to demonstrate that the fundamental properties of the graph Laplacian from the spectral graph theory can be applied to a Transformer. We have shown that our multi-view feature fusion strategy is most effective when coupled with mesh segmentation using spectral clustering. We further present spectral filtering and an optimisation-based refinement to reconstruct more physically plausible meshes. Our framework is general and can be extended to other multi-view reconstruction tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 664, + 391, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 664, + 391, + 674 + ], + "spans": [ + { + "bbox": [ + 306, + 664, + 391, + 674 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 674, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 547, + 713 + ], + "type": "text", + "content": "This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2023-2020-0-01789), supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14673" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 112, + 206, + 169 + ], + "blocks": [ + { + "bbox": [ + 127, + 69, + 153, + 80 + ], + "lines": [ + { + "bbox": [ + 127, + 69, + 153, + 80 + ], + "spans": [ + { + "bbox": [ + 127, + 69, + 153, + 80 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 112, + 206, + 169 + ], + "lines": [ + { + "bbox": [ + 59, + 112, + 206, + 169 + ], + "spans": [ + { + "bbox": [ + 59, + 112, + 206, + 169 + ], + "type": "image", + "image_path": "f6a2d4ee2a648be09e691072b01186cb4819a94e6ed1f51430e66b916b0ffd82.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 216, + 85, + 375, + 192 + ], + "blocks": [ + { + "bbox": [ + 254, + 69, + 336, + 80 + ], + "lines": [ + { + "bbox": [ + 254, + 69, + 336, + 80 + ], + "spans": [ + { + "bbox": [ + 254, + 69, + 336, + 80 + ], + "type": "text", + "content": "Input zoomed view" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 216, + 85, + 375, + 192 + ], + "lines": [ + { + "bbox": [ + 216, + 85, + 375, + 192 + ], + "spans": [ + { + "bbox": [ + 216, + 85, + 375, + 192 + ], + "type": "image", + "image_path": "c927a5734fc3d4c695e7a461e0149987c19b319896fa85a6a5cb98cc7951b2a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 374, + 91, + 536, + 188 + ], + "blocks": [ + { + "bbox": [ + 410, + 69, + 499, + 80 + ], + "lines": [ + { + "bbox": [ + 410, + 69, + 499, + 80 + ], + "spans": [ + { + "bbox": [ + 410, + 69, + 499, + 80 + ], + "type": "text", + "content": "Output zoomed view" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 374, + 91, + 536, + 188 + ], + "lines": [ + { + "bbox": [ + 374, + 91, + 536, + 188 + ], + "spans": [ + { + "bbox": [ + 374, + 91, + 536, + 188 + ], + "type": "image", + "image_path": "f0a972495511209885c5ab509a8e691d8eb8edd515df133adac06bc3810ca69d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 51, + 247, + 194, + 307 + ], + "blocks": [ + { + "bbox": [ + 51, + 247, + 194, + 307 + ], + "lines": [ + { + "bbox": [ + 51, + 247, + 194, + 307 + ], + "spans": [ + { + "bbox": [ + 51, + 247, + 194, + 307 + ], + "type": "image", + "image_path": "1fbbfb1b0a1679f18106bf1badfdbd7cafa471d4c0f0933018f816b8c0e29fe6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 197, + 220, + 370, + 320 + ], + "blocks": [ + { + "bbox": [ + 197, + 220, + 370, + 320 + ], + "lines": [ + { + "bbox": [ + 197, + 220, + 370, + 320 + ], + "spans": [ + { + "bbox": [ + 197, + 220, + 370, + 320 + ], + "type": "image", + "image_path": "76135f1c83a2964e87c1c9f1ad0085202a0cbb0bfcbbe284d99b4fbf4eca06ff.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 370, + 224, + 536, + 320 + ], + "blocks": [ + { + "bbox": [ + 370, + 224, + 536, + 320 + ], + "lines": [ + { + "bbox": [ + 370, + 224, + 536, + 320 + ], + "spans": [ + { + "bbox": [ + 370, + 224, + 536, + 320 + ], + "type": "image", + "image_path": "a03828bee0b822a8a0bca4201e12f487a1e038c9aa69e1e8bf7b30b7a470b5ee.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 55, + 403, + 192, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 403, + 192, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 192, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 192, + 460 + ], + "type": "image", + "image_path": "6cd0a39094f2b955da7a35926a8500a70e7077ec04ad0b6e75d3ff9f98f3e39f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 210, + 355, + 363, + 497 + ], + "blocks": [ + { + "bbox": [ + 210, + 355, + 363, + 497 + ], + "lines": [ + { + "bbox": [ + 210, + 355, + 363, + 497 + ], + "spans": [ + { + "bbox": [ + 210, + 355, + 363, + 497 + ], + "type": "image", + "image_path": "a7050f79ee12ee5154d19cf508f26cff30ea525d205a4b9438524e15b2ed2f60.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 378, + 358, + 533, + 504 + ], + "blocks": [ + { + "bbox": [ + 378, + 358, + 533, + 504 + ], + "lines": [ + { + "bbox": [ + 378, + 358, + 533, + 504 + ], + "spans": [ + { + "bbox": [ + 378, + 358, + 533, + 504 + ], + "type": "image", + "image_path": "20f63ca0a3e2a35fc96c231faac23ebf9d0a06be8f645516dd86725a15834904.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 58, + 548, + 188, + 595 + ], + "blocks": [ + { + "bbox": [ + 58, + 548, + 188, + 595 + ], + "lines": [ + { + "bbox": [ + 58, + 548, + 188, + 595 + ], + "spans": [ + { + "bbox": [ + 58, + 548, + 188, + 595 + ], + "type": "image", + "image_path": "926d9c1c67882e578a833fcf078db310478d41f0fb0015799ff573f7164f30b5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 206, + 512, + 362, + 642 + ], + "blocks": [ + { + "bbox": [ + 206, + 512, + 362, + 642 + ], + "lines": [ + { + "bbox": [ + 206, + 512, + 362, + 642 + ], + "spans": [ + { + "bbox": [ + 206, + 512, + 362, + 642 + ], + "type": "image", + "image_path": "591a7b4c4bf9fc63203bff2f47b2b499e1bd490bdfdc15b4407859a91204eef0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 658, + 545, + 682 + ], + "lines": [ + { + "bbox": [ + 46, + 658, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 545, + 682 + ], + "type": "text", + "content": "Figure 9: Additional qualitative examples of mesh refinement at inference. Our optimisation-based strategy shows robustness to various hand poses. We present failure cases when self-penetration is highly complex in the supplementary materials." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 388, + 510, + 533, + 646 + ], + "blocks": [ + { + "bbox": [ + 388, + 510, + 533, + 646 + ], + "lines": [ + { + "bbox": [ + 388, + 510, + 533, + 646 + ], + "spans": [ + { + "bbox": [ + 388, + 510, + 533, + 646 + ], + "type": "image", + "image_path": "27c88ef3c9574cec71e64838d3732986439f61980183ca32d1db8176fb284e71.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14674" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "text", + "content": "[2] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for RGB-based dense 3D hand pose estimation via neural rendering. In CVPR, 2019. 2, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 180 + ], + "type": "text", + "content": "[3] Luca Ballan, Aparna Taneja, Jürgen Gall, Luc Van Gool, and Marc Pollefeys. Motion capture of hands in action using discriminative salient points. In ECCV, 2012. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 217, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 248 + ], + "type": "text", + "content": "[5] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3D hand shape and pose from images in the wild. In CVPR, 2019. 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 251, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 283 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 285, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 306 + ], + "type": "text", + "content": "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 308, + 287, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 287, + 340 + ], + "type": "text", + "content": "[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3D human mesh recovery with transformers. In ECCV, 2022. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 342, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 384 + ], + "type": "text", + "content": "[9] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2Mesh: Graph convolutional network for 3D human pose and mesh recovery from a 2D human pose. In ECCV, 2020. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 387, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 407 + ], + "type": "text", + "content": "[10] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "type": "text", + "content": "[11] Enric Corona, Tomas Hodan, Minh Vo, Francesc Moreno-Noguer, Chris Sweeney, Richard Newcombe, and Lingni Ma. LISA: Learning implicit shape and appearance of hands. In CVPR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 487 + ], + "type": "text", + "content": "[12] Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. In NeurIPS, 2016. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 488, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 531 + ], + "type": "text", + "content": "[13] Boyang Deng, John P Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. NASA neural articulated shape approximation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 533, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 565 + ], + "type": "text", + "content": "[14] Inderjit S Dhillon, Yuqiang Guan, and Brian Kulis. Weighted graph cuts without eigenvectors a multilevel approach. TPAMI, 2007. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "text", + "content": "[15] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. HOPE-Net: A graph-based model for hand-object pose estimation. In CVPR, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "text", + "content": "[16] Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. AAAI Workshop on Deep Learning on Graphs: Methods and Applications, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[17] Vijay Prakash Dwivedi, Chaitanya K Joshi, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. JMLR, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[18] Zicong Fan, Adrian Spurr, Muhammed Kocabas, Siyu Tang, Michael J Black, and Otmar Hilliges. Learning to disambiguate strongly interacting hands via probabilistic per-pixel part segmentation. In 3DV, 2021. 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[19] Runyang Feng, Yixing Gao, Xueqing Ma, Tze Ho Elden Tse, and Hyung Jin Chang. Mutual information-based temporal difference learning for human pose estimation in video. In CVPR, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[20] Runyang Feng, Yixing Gao, Tze Ho Elden Tse, Xueqing Ma, and Hyung Jin Chang. DiffPose: SpatioTemporal diffusion model for video-based human pose estimation. arXiv preprint arXiv:2307.16687, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 162, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 204 + ], + "type": "text", + "content": "[21] Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with RGB-D videos and 3D hand pose annotations. In CVPR, 2018. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 205, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 545, + 248 + ], + "type": "text", + "content": "[22] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3D hand shape and pose estimation from a single RGB image. In CVPR, 2019. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 250, + 545, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 292 + ], + "type": "text", + "content": "[23] Patrick Grady, Chengcheng Tang, Christopher D Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C Kemp. ContactOpt: Optimizing contact to improve grasps. In CVPR, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "text", + "content": "[24] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint Transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "text", + "content": "[25] Shangchen Han, Beibei Liu, Randi Cabezas, Christopher D Twigg, Peizhao Zhang, Jeff Petkau, Tsz-Ho Yu, Chun-Jung Tai, Muzaffer Akbay, Zheng Wang, et al. MEgATrack: Monochrome egocentric articulated hand-tracking for virtual reality. ACM Transactions on Graphics (ToG), 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 394, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 446 + ], + "type": "text", + "content": "[26] Shangchen Han, Po-chen Wu, Yubo Zhang, Beibei Liu, Linguuang Zhang, Zheng Wang, Weiguang Si, Peizhao Zhang, Yujun Cai, Tomas Hodan, et al. UmeTrack: Unified multiview end-to-end hand tracking for vr. In SIGGRAPH, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "text", + "content": "[27] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In CVPR, 2019. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 493, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 525 + ], + "type": "text", + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 526, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 547 + ], + "type": "text", + "content": "[29] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In CVPR, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 548, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 579 + ], + "type": "text", + "content": "[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 581, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 545, + 613 + ], + "type": "text", + "content": "[31] Korrawe Karunratanakul, Adrian Spurr, Zicong Fan, Otmar Hilliges, and Siyu Tang. A skeleton-driven neural occupancy representation for articulated hands. In 3DV, 2021. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 657 + ], + "type": "text", + "content": "[32] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM computing surveys (CSUR), 2022. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[33] Dong Uk Kim, Kwang In Kim, and Seungryul Baek. End-to-end detection and pose estimation of two interacting hands. In ICCV, 2021. 1, 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14675" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[35] Devin Kreuzer, Dominique Beaini, Will Hamilton, Vincent Létourneau, and Prudencio Tossou. Rethinking graph transformers with spectral attention. NeurIPS, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[36] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[37] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2O: Two hands manipulating objects for first person interaction recognition. In ICCV, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 228 + ], + "type": "text", + "content": "[38] Mengcheng Li, Liang An, Hongwen Zhang, Lianpeng Wu, Feng Chen, Tao Yu, and Yebin Liu. Interacting attention graph for single image two-hand reconstruction. In CVPR, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "text", + "content": "[39] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 286 + ], + "type": "text", + "content": "[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 287, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 309 + ], + "type": "text", + "content": "[41] Rong Liu and Hao Zhang. Segmentation of 3D meshes through spectral clustering. In Pacific Graphics, 2004. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "text", + "content": "[42] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 376 + ], + "type": "text", + "content": "[43] Tomas Möller and Ben Trumbore. Fast, minimum storage ray-triangle intersection. Journal of graphics tools, 1997. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "type": "text", + "content": "[44] Gyeongsik Moon and Kyoung Mu Lee. I2L-MeshNet: Image-to-lixel prediction network for accurate 3D human pose and mesh estimation from a single RGB image. In ECCV, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[45] Gyeongsik Moon, Takaaki Shiratori, and Kyoung Mu Lee. DeepHandMesh: A weakly-supervised deep encoder-decoder framework for high-fidelity hand mesh modeling. In ECCV, 2020. 2, 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 467, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 511 + ], + "type": "text", + "content": "[46] Gyeongsik Moon, Shouu-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. InterHand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single RGB image. In ECCV, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[47] Franziska Mueller, Micah Davis, Florian Bernard, Oleksandr Sotnychenko, Micekal Verschooor, Miguel A Otaduy, Dan Casas, and Christian Theobalt. Real-time pose and shape reconstruction of two interacting hands with a single depth camera. In SIGGRAPH, 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "text", + "content": "[48] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 2022. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 634 + ], + "type": "text", + "content": "[49] Blender project. Cycles. https://www.cycles-renderer.org/.6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[50] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3D human pose estimation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[51] Edoardo Remelli, Shangchen Han, Sina Honari, Pascal Fua, and Robert Wang. Lightweight multi-view 3D pose estimation through camera-disentangled representation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[52] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 36(6):245, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[53] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3D reconstruction of interacting hands via collision-aware factorized refinements. In 3DV, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[54] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large scale visual recognition challenge. IJCV, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[55] Breannan Smith, Chenglei Wu, He Wen, Patrick Peluse, Yaser Sheikh, Jessica K Hodgins, and Takaaki Shiratori. Constraining dense hand surface tracking with elasticity. ACM Transactions on Graphics (ToG), 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 241, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 263 + ], + "type": "text", + "content": "[56] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, 2007. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 286 + ], + "type": "text", + "content": "[57] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "type": "text", + "content": "[58] Tze Ho Elden Tse, Daniele De Martini, and Letizia Marchegiani. No need to scream: Robust sound-based speaker localisation in challenging scenarios. In ICSR 2019, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 375 + ], + "type": "text", + "content": "[59] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In CVPR, 2022. 1, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "text", + "content": "[60] Tze Ho Elden Tse, Zhongqun Zhang, Kwang In Kim, Ales Leonardis, Feng Zheng, and Hyung Jin Chang. " + }, + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "text", + "content": " Contact: Graph-based network for 3D hand-object contact estimation with semi-supervised learning. In ECCV, 2022. 1, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "text", + "content": "[61] Dimitrios Tzionas, Luca Ballan, Abhilash Srikantha, Pablo Aponte, Marc Pollefeys, and Juergen Gall. Capturing hands in action using discriminative salient points and physics simulation. IJCV, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 466, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 499 + ], + "type": "text", + "content": "[62] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "text", + "content": "[63] Jiayi Wang, Franziska Mueller, Florian Bernard, Suzanne Sorli, Oleksandr Sotnychenko, Neng Qian, Miguel A Otabuy, Dan Casas, and Christian Theobalt. RGB2Hands: real-time tracking of 3D hand interactions from monocular rgb video. In SIGGRAPH, 2020. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "text", + "content": "[64] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3D human shape and articulated pose models. In CVPR, 2020. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[65] Baowen Zhang, Yangang Wang, Xiaoming Deng, Yinda Zhang, Ping Tan, Cuixia Ma, and Hongan Wang. Interacting two-hand 3D pose and shape reconstruction from single color image. In ICCV, 2021. 1, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[66] Linfang Zheng, Ales Leonardis, Tze Ho Elden Tse, Nora Horanyi, Hua Chen, Wei Zhang, and Hyung Jin Chang. TP-AE: Temporally primed 6D object pose tracking with autoencoders. In ICRA, 2022. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 690, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 712 + ], + "type": "text", + "content": "[67] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. FreiHAND: A" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14676" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [], + "discarded_blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "dataset for markerless capture of hand pose and shape from single RGB images. In ICCV, 2019. 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14677" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/dc8d764a-9214-4358-ad24-c2007b58a08c_content_list.json b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/dc8d764a-9214-4358-ad24-c2007b58a08c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..edf120fdc9fdd5bb3c7da4b7b550628d5a5ef5a1 --- /dev/null +++ b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/dc8d764a-9214-4358-ad24-c2007b58a08c_content_list.json @@ -0,0 +1,1574 @@ +[ + { + "type": "text", + "text": "Spectrum-guided Multi-granularity Referring Video Object Segmentation", + "text_level": 1, + "bbox": [ + 110, + 161, + 855, + 182 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bo Miao $^{1}$ , Mohammed Bennamoun $^{1}$ , Yongsheng Gao $^{2}$ , Ajmal Mian $^{1}$ $^{1}$ The University of Western Australia \n $^{2}$ Griffith University", + "bbox": [ + 194, + 206, + 771, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 290, + 312, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current referring video object segmentation (R-VOS) techniques extract conditional kernels from encoded (low-resolution) vision-language features to segment the decoded high-resolution features. We discovered that this causes significant feature drift, which the segmentation kernels struggle to perceive during the forward computation. This negatively affects the ability of segmentation kernels. To address the drift problem, we propose a Spectrum-guided Multigranularity (SgMg) approach, which performs direct segmentation on the encoded features and employs visual details to further optimize the masks. In addition, we propose Spectrum-guided Cross-modal Fusion (SCF) to perform intra-frame global interactions in the spectral domain for effective multimodal representation. Finally, we extend SgMg to perform multi-object R-VOS, a new paradigm that enables simultaneous segmentation of multiple referred objects in a video. This not only makes R-VOS faster, but also more practical. Extensive experiments show that SgMg achieves state-of-the-art performance on four video benchmark datasets, outperforming the nearest competitor by $2.8\\%$ points on Ref-YouTube-VOS. Our extended SgMg enables multi-object R-VOS, runs about $3\\times$ faster while maintaining satisfactory performance. Code is available at https://github.com/bo-miao/SgMg.", + "bbox": [ + 75, + 318, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 676, + 208, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring video object segmentation (R-VOS) aims at segmenting objects in a video, referred to by linguistic descriptions. R-VOS is an emerging task for multimodal reasoning and promotes a wide range of applications, including language-guided video editing and human-machine interaction. Different from conventional semi-supervised video object segmentation [44, 7, 41], where the mask annotation for the first frame is provided for reference, R-VOS is more challenging due to the need for cross-modal understanding between vision and free-form language expressions.", + "bbox": [ + 75, + 700, + 468, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early R-VOS techniques [2, 21, 61] perform feature encoding, cross-modal interaction, and language grounding", + "bbox": [ + 75, + 838, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/89d6e208c163b4cbcd2b4ac342aeb8c5568808f60e50da5bd790db27323de29d.jpg", + "image_caption": [ + "(b) SgMg (Ours): segment-and-optimize w/ encoded features $F_{vl}$" + ], + "image_footnote": [], + "bbox": [ + 504, + 290, + 890, + 423 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc603f36ec204257f55403249444807ca1ada13606a4021792a967903df7776c.jpg", + "image_caption": [ + "Conditional Segmentation Fusion: Cross-modal Fusion $F_{\\mu l}$ : Encoded Vision-Language Features $F_{\\nu l}^{d}$ : Decoded Features Q: Instance Query $K_{c}$ : Conditional Kernel $K_{cp}$ : Conditional Patch Kernel $M_{Q}$ : Predicted Mask", + "Figure 1. (a) Previous methods [4, 57] apply segmentation kernels $\\mathcal{K}_c$ [50], extracted from encoded features $\\mathcal{F}_{vl}$ , to segment the decoded high-resolution features $\\mathcal{F}_{vl}^d$ . (b) We use segmentation kernels $\\mathcal{K}_{cp}$ , extracted from encoded features $\\mathcal{F}_{vl}$ , to segment the encoded features $\\mathcal{F}_{vl}$ directly, and propose multi-granularity optimization to recover visual details and produce fine-grained masks." + ], + "image_footnote": [], + "bbox": [ + 502, + 451, + 884, + 514 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "using convolutional neural networks (CNNs). However, the limited ability of CNNs to capture long-range dependencies and handle free-form features constrains the model performance. With the advancement of attention mechanisms [52, 43, 15, 16, 58], recent methods achieved significant improvement on R-VOS using cross-attention [49, 23, 27] for multimodal understanding and transformers [6, 56] for spatio-temporal representation. Based on transformers, conditional kernel [50] is then introduced to separate foreground from semantic features given its high adaptability to different instances [4, 57]. As illustrated in Fig. 1(a), these methods attend to encoded vision-language features $\\mathcal{F}_{vl}$ using instance queries $Q$ to predict conditional kernels $\\kappa_{c}$ , and employ $\\kappa_{c}$ as the segmentation head to segment decoded features $\\mathcal{F}_{vl}^{d}$ . Despite the promising performance,", + "bbox": [ + 497, + 657, + 892, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 85 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 253, + 48, + 789, + 86 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "920", + "bbox": [ + 472, + 885, + 497, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "this paradigm still has some limitations. Firstly, as shown in the t-SNE [51] visualization in Fig. 1(a), although the nonlinear decoding process introduces visual details, this is accompanied by a significant feature drift, which increases the difficulty of segmentation since $\\mathcal{K}_c$ is predicted before feature decoding. Secondly, bilinear upsampling of the predicted masks $\\mathcal{M}_Q$ to increase resolution impedes the segmentation performance. Thirdly, these methods only support single expression-based segmentation, making R-VOS inefficient when multiple referred objects exist in a video.", + "bbox": [ + 75, + 125, + 468, + 264 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a Spectrum-guided Multi-granularity (SgMg) approach that follows a segment-and-optimize pipeline to address the above problems. As depicted in Fig. 1(b), SgMg introduces Conditional Patch Kernel (CPK) $\\mathcal{K}_{cp}$ to directly segment its fully perceived encoded features $\\mathcal{F}_{vl}$ , avoiding the feature drift and its adverse effects. The segmentation is then refined using our proposed Multi-granularity Segmentation Optimizer (MSO), which employs low-level visual details to produce full-resolution masks. Within the SgMg framework, we further develop Spectrum-guided Cross-modal Fusion (SCF) that performs intra-frame global interactions in the spectral domain to facilitate multimodal understanding. Finally, we introduce a new paradigm called multi-object R-VOS to simultaneously segment multiple referred objects in a video. To achieve this, we extend SgMg by devising multi-instance fusion and decoupling. Our main contributions are summarized as follows:", + "bbox": [ + 75, + 264, + 470, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explain how existing R-VOS methods suffer from the feature drift problem. To address this problem, we propose $\\mathrm{SgMg}$ that follows a segment-and-optimize pipeline and achieves top-ranked overall performance on multiple benchmark datasets.", + "- We propose Spectrum-guided Cross-modal Fusion to encourage intra-frame global interactions in the spectral domain.", + "- We extend $\\mathrm{SgMg}$ to perform multi-object R-VOS, a new paradigm that enables simultaneous segmentation of multiple referred objects in a video. Our multi-object variant is more practical and runs $3\\times$ faster." + ], + "bbox": [ + 92, + 530, + 468, + 725 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on multiple benchmark datasets, including Ref-YouTube-VOS [49], RefDAVIS17 [21], A2D-Sentences [17], and JHMDB-Sentences [20], and achieve state-of-the-art performance on all four. On the largest validation set Ref-YouTube-VOS, $\\mathrm{SgMg}$ achieves $65.7\\mathcal{J}\\& \\mathcal{F}$ which is $2.8\\%$ points higher than that of the closest competitor ReferFormer [57]. On the A2D-Sentences, $\\mathrm{SgMg}$ achieves $58.5\\mathrm{mAP}$ which is $3.5\\%$ points higher than that of ReferFormer.", + "bbox": [ + 75, + 742, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 499, + 124, + 648, + 138 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Object Segmentation techniques fall into two categories: unsupervised and semi-supervised. Unsupervised approaches segment the most salient instances in each video without user interactions [38, 47]. They often employ two-stream networks to fuse motion and appearance cues for segmentation. Semi-supervised approaches track the given first frame object mask by performing online learning [5] or spatial-temporal association [44, 7, 66, 42, 53]. Unlike conventional semi-supervised video object segmentation, R-VOS takes a free-form linguistic expression as guidance to detect and segment referred objects in videos.", + "bbox": [ + 497, + 147, + 890, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Referring Video Object Segmentation. R-VOS methods mainly use deep neural networks with vision-and-language interaction to empower visual features with corresponding linguistic information for pixel-level segmentation. For example, [49] employs a unified R-VOS framework that performs iterative segmentation guided by both language and temporal features. [32, 29] adopt progressive segmentation by perceiving potential objects and discriminating the best match. [70] fuses visual and motion features for segmentation under the guidance of linguistic cues. [30] models object relations to form tracklets and performs tracklet-language grounding. To enhance multi-modal interactions, [61, 14, 11, 10] perform hierarchical vision-language fusion on multiple feature layers.", + "bbox": [ + 497, + 299, + 890, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their promising performance, the complex multistage pipelines and use of multiple networks make R-VOS burdensome. To address these problems, MTTR [4] proposes an end-to-end transformer-based network with conditional kernels [50] to segment target objects. ReferFormer [57] further introduces language-guided instance queries to predict instance-aware conditional kernels and an auxiliary detection task to aid localization. These methods follow a decode-and-segment pipeline, which adopts conditional kernels to segment decoded high-resolution features to achieve promising performance. However, the nonlinear decoding process leads to significant feature drift that negatively affects the conditional kernels. In contrast to previous works, our approach follows a segment-and-optimize pipeline to avoid the adverse drift effects and to predict full-resolution masks in an efficient manner.", + "bbox": [ + 497, + 492, + 890, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision and Language Representation Learning aims to learn vision-language semantics and alignment for multimodal reasoning tasks. It has achieved significant success in various tasks [37, 67, 68], including video question answering [71], video captioning [1], video-text retrieval [12, 69], zero-shot classification [46], referring image/video segmentation [49], etc. Some approaches [46, 24] rely on contrastive pre-training using large-scale datasets to project different modalities into unified embedding space. Others [36, 13] develop cross-modal interaction layers for multimodal feature fusion and understanding. Recent deep", + "bbox": [ + 497, + 714, + 890, + 866 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "921", + "bbox": [ + 470, + 885, + 495, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a094d27b35f2ca47ca5c6634e7b32f40cd1830c5bc4be6a1bb9f6c78fe79b680.jpg", + "image_caption": [ + "Figure 2. The overall framework of $\\mathrm{SgMg}$ . Taking a video sequence $\\mathcal{V} = \\{I_i\\}_{i=1}^T$ and a language expression $\\mathcal{L} = \\{S_i\\}_{i=1}^N$ as input, $\\mathrm{SgMg}$ predicts the masks of referred object $\\mathcal{O}_{\\mathcal{L}}$ in each frame. SCF projects visual features $\\mathcal{F}_v$ to vision-language features $\\mathcal{F}_{vl}$ , instance-aware CPK predicts patch masks by segmenting encoded $\\mathcal{F}_{vl}$ , and MSO optimizes patch masks to get fine-grained results." + ], + "image_footnote": [], + "bbox": [ + 97, + 125, + 867, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "learning methods in spectral domain [18, 40, 8, 34] have raised widespread awareness because of their ability to perform global interactions. We take inspiration from these spectral-based methods and employ spectrum guidance in the field of vision-language representation to encourage multimodal global interactions.", + "bbox": [ + 75, + 428, + 470, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. $\\mathrm{SgMg}$ : Spectrum-guided Multi-granularity Referring Video Object Segmentation", + "text_level": 1, + "bbox": [ + 75, + 530, + 468, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video sequence $\\mathcal{V} = \\{I_i\\}_{i=1}^T$ with $T$ frames and a language query $\\mathcal{L} = \\{S_i\\}_{i=1}^N$ with $N$ words. The goal of R-VOS is to segment the referred object $\\mathcal{O}_{\\mathcal{L}}$ in $\\mathcal{V}$ at pixel-level. To this end, we introduce a new approach termed $\\mathrm{SgMg}$ . Different from previous R-VOS methods [4, 57], our approach follows a segment-and-optimize pipeline.", + "bbox": [ + 75, + 572, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of $\\mathrm{SgMg}$ is shown in Fig. 2. Video Swin Transformer [35] is adopted to extract visual feature $\\mathcal{F}_v$ and RoBERTa [33] is adopted to extract sentence $\\mathcal{F}_s$ and word $\\mathcal{F}_w$ features. The channel dimension of all features is projected to 256. Spectrum-guided Cross-modal Fusion (SCF) cross attends $\\mathcal{F}_v$ with $\\mathcal{F}_w$ to compute vision-language features $\\mathcal{F}_{vl}$ . Deformable Transformer [72] encoder is used to encode $\\mathcal{F}_{vl}$ and the decoder associates instance queries created based on $\\mathcal{F}_s$ to predict instance embeddings and the corresponding Conditional Patch Kernels (CPKs). Finally, the CPKs are employed to segment $\\mathcal{F}_{vl}$ and predict patch masks that are further optimized with visual details through Multi-granularity Segmentation Optimizer (MSO). The choice of the encoder and transformer follows previous works to avoid distractions [4, 57].", + "bbox": [ + 75, + 659, + 470, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature Drift Analysis", + "text_level": 1, + "bbox": [ + 499, + 427, + 707, + 442 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing R-VOS methods [57, 4] follow a decode-and-segment pipeline where conditional kernels $\\mathcal{K}_c$ [50] are extracted from encoded features $\\mathcal{F}_{vl}$ and used to segment the decoded features $\\mathcal{F}_{vl}^d$ . However, the decoding process leads to feature drift, which is evident in the t-SNE visualization depicted in Fig. 1(a). This drift is difficult for the kernels $\\mathcal{K}_c$ to perceive during the forward computation since $\\mathcal{K}_c$ is predicted before the feature decoding. Therefore, we argue that even though the feature decoding enhances visual details, it also causes the drift problem that negatively affects the segmentation kernels. This makes the existing decode-and-segment pipeline sub-optimal.", + "bbox": [ + 497, + 448, + 890, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome the adverse effects of feature drift while recovering visual details, we present $\\mathrm{SgMg}$ , a novel approach that follows a segment-and-optimize pipeline. In a nutshell, $\\mathrm{SgMg}$ performs Spectrum-guided Cross-modal Fusion to compute $\\mathcal{F}_{vl}$ , leverages Conditional Patch Kernels to segment encoded features $\\mathcal{F}_{vl}$ to avoid the drift effects, and recovers visual details with Multi-granularity Segmentation Optimizer to generate fine-grained masks.", + "bbox": [ + 497, + 614, + 892, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Spectrum-guided Cross-modal Fusion", + "text_level": 1, + "bbox": [ + 499, + 733, + 826, + 749 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The two-dimensional discrete Fourier transform converts spatial data into the spectral domain. Based on the spectral convolution theorem [3], point-wise update of signals in the spectral domain globally affects all inputs in the spatial domain, which gives the insight to design spectrum-based modules so as to efficiently facilitate global interactions, which is critical for multimodal understanding. In addition, Low-frequency components in the spectral domain usually", + "bbox": [ + 497, + 755, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "922", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2759da62b255c26bbafd9000234dd434c8a8d9e64545b5286f5ee30137e530ce.jpg", + "image_caption": [ + "(a) Spectrum-Guided Cross-Modal Fusion", + "(b) Spectrum Augmentation" + ], + "image_footnote": [ + "Flatten: reshape $\\mathcal{M}_P$ from $\\mathbb{R}^{\\frac{H}{i}\\times \\frac{W}{i}\\times p^2}$ to $\\mathbb{R}^{\\frac{Hp}{i}\\times \\frac{Wp}{i}}$ for visualization. $\\times 2$ : upsampling operation." + ], + "bbox": [ + 78, + 122, + 460, + 301 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e07e40cf9af570d6f85d2835a23d862802b7928425ca500af0c62a6251d34559.jpg", + "image_caption": [ + "Figure 3. Spectrum-guided Cross-modal Fusion. Imag.: Imaginary. Pre-spectrum augmentation and post-spectrum augmentation share an identical structure." + ], + "image_footnote": [], + "bbox": [ + 78, + 315, + 467, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "correspond to the general semantic information according to previous theoretical studies [59, 60, 64].", + "bbox": [ + 75, + 396, + 468, + 423 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the above observations, we conjecture that low-frequency components can benefit higher dimensional semantic features and propose Spectrum-guided Cross-modal Fusion (SCF). As shown in Fig. 3, SCF performs pre-spectrum augmentation to enhance visual features before cross-modal fusion and post-spectrum augmentation to facilitate global vision-language interactions after the fusion process. Let $\\mathcal{F} \\in \\mathbb{R}^{C \\times H \\times W}$ denotes the input features, the spectrum augmentation (SA) is computed as:", + "bbox": [ + 75, + 423, + 468, + 538 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A} (\\mathcal {F}, K) = \\mathcal {F} + \\Theta_ {I F F T} (\\operatorname {C o n v} (\\sigma (K, \\mathcal {F}) \\odot \\Theta_ {F F T} (\\mathcal {F}))) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 546, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes low-pass filtering with adaptive Gaussian smoothed filters $\\sigma(K, \\mathcal{F})$ , which has the same spatial size as $\\mathcal{F}$ , and $K$ is the bandwidth. To make $\\sigma(K, \\mathcal{F})$ input-aware, we create an initial 2D Gaussian map based on $K$ , and apply pooling and linear layers on $\\mathcal{F}$ to predict a scale parameter to update the Gaussian map. Thanks to the spectral convolution theorem, the efficient point-wise spectral convolution globally updates $\\mathcal{F}$ . We treat the spectral-operated features as residuals and add them to the original input features for enhancement. Overall, SCF, which takes visual features $\\mathcal{F}_v$ and word-level text features $\\mathcal{F}_w$ as input, is computed as:", + "bbox": [ + 75, + 568, + 468, + 707 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S C F} \\left(\\mathcal {F} _ {w}, \\mathcal {F} _ {v}\\right) = \\operatorname {S A} \\left(\\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right) \\otimes \\operatorname {A t t} \\left(\\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right), \\mathcal {F} _ {w}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 110, + 716, + 468, + 730 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Conditional Patch Segmentation", + "text_level": 1, + "bbox": [ + 75, + 748, + 363, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We devise Conditional Patch Kernel (CPK) as the segmentation head to predict patch masks from the encoded vision-language features $\\mathcal{F}_{vl}$ that are fully perceived by CPK. Unlike previous works [4, 57], CPK predicts a sequence of labels for each token rather than a single label, efficiently improving segmentation resolution along the channel dimension.", + "bbox": [ + 75, + 769, + 468, + 865 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5352f0d2490d1e1f8f1404aadde77b7151607047992d91378dfdada585618560.jpg", + "image_caption": [ + "Figure 4. Multi-granularity Segmentation Optimizer, which predicts residual maps to optimize patch masks $\\mathcal{M}_P$ progressively." + ], + "image_footnote": [], + "bbox": [ + 500, + 124, + 890, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first use sentence-level text features $\\mathcal{F}_s$ and multiple learnable embeddings to generate instance queries $Q\\in \\mathbb{R}^{N\\times C}$ . Next, $Q$ is projected into instance embeddings $\\mathcal{E}\\in \\mathbb{R}^{N\\times C}$ using the transformer decoder and $\\mathcal{E}$ is leveraged to predict CPK for each instance query:", + "bbox": [ + 497, + 312, + 890, + 382 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C P K} (Q, \\mathcal {F} _ {v l}) = \\Theta (\\operatorname {F C} (\\operatorname {A t t} (Q, \\mathcal {F} _ {v l}))) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 564, + 391, + 890, + 408 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Theta$ denotes the parameterization operation that reshapes CPK to form two point-wise convolutions with the output channel number of 16, which is similar to [50]. Since $Q$ changes dynamically according to different linguistic expressions, CPK becomes instance-aware and can separate objects of interest from $F_{vl}$ . Finally, we apply the parameterized CPK (dynamic point-wise convolutions) on $F_{vl}$ to predict patch masks $\\mathcal{M}_P\\in \\mathbb{R}^{\\frac{H}{i}\\times \\frac{W}{i}\\times p^2}$ , where $\\frac{H}{i}\\times \\frac{W}{i}$ denotes the spatial resolution of $F_{vl}$ and $p^2$ denotes the increased segmentation resolution on the channel dimension.", + "bbox": [ + 497, + 416, + 890, + 557 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During inference, we can reshape patch masks to $\\mathcal{M}_P\\in$ $\\mathbb{R}\\frac{Hp}{i}\\times \\frac{Wp}{i}$ to efficiently generate fine-grained segmentation from low-resolution $F_{vl}$ . The resolution of prediction will be consistent with the input when $p$ equals to $i$ . We found that this efficient CPK can achieve competitive performance compared to methods that use heavy decoders.", + "bbox": [ + 497, + 557, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Multi-granularity Segmentation Optimizer", + "text_level": 1, + "bbox": [ + 499, + 650, + 865, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Segmenting encoded features $\\mathcal{F}_{vl}$ with CPK avoids the detrimental drift effect on the segmentation head. However, visual details are required to produce accurate fine-grained masks. We propose Multi-granularity Segmentation Optimizer (MSO) to achieve this goal.", + "bbox": [ + 497, + 671, + 890, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An overview of MSO is shown in Fig. 4. It takes the predicted patch masks $\\mathcal{M}_P$ as object priors and reuses visual features $\\mathcal{F}_v$ with spatial strides of $\\{4,8\\}$ to gradually recover visual details and refine the priors. Specifically, MSO first concatenates $\\mathcal{M}_P$ and $\\mathcal{F}_v$ and projects them to low dimensional bases. Next, residual masks predicted by performing another convolution on these bases are used to correct $\\mathcal{M}_P$ . Finally, the optimized patch masks achieve the input resolution by reshaping from $\\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times 4^2}$ to $\\mathbb{R}^{H \\times W}$ . Since", + "bbox": [ + 497, + 739, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "923", + "bbox": [ + 472, + 887, + 495, + 897 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/02f7359109a6fa52d753e26b6c274c9a713998aa69959a5b8775d86504c6a7a3.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparing different methods for their ability to segment single or multiple frames or multiple objects simultaneously." + ], + "table_body": "
MethodsSingle-frameMulti-framesMulti-objects
[49, 63, 11] et al.✓
[4, 57, 56, 23] et al.✓✓
Fast SgMg (Ours)✓✓✓
", + "bbox": [ + 82, + 122, + 460, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MSO does not include heavy computations, the segment-and-optimize pipeline makes our approach perform better with efficient inference time.", + "bbox": [ + 75, + 225, + 468, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Multi-Object R-VOS", + "text_level": 1, + "bbox": [ + 77, + 275, + 273, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing R-VOS methods perform single-frame (frameworkwise) segmentation [49, 63, 11] or multi-frame (clip-wise) segmentation [4, 57, 56] for an individual referred object at a time. However, to the best of our knowledge, no existing work explores the simultaneous segmentation of multiple referred objects in video using a common GPU, which is important for real-world scenarios. To fill this gap, we present a new paradigm called multi-object R-VOS.", + "bbox": [ + 75, + 296, + 468, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key to multi-object R-VOS is designing a network that shares computationally intensive features for multiple objects, and enables different instance features to be decoupled before segmentation. To achieve this, we extend $\\mathrm{SgMg}$ for multi-object R-VOS by introducing multi-instance fusion and decoupling. As shown in Table 1, our method, dubbed Fast $\\mathrm{SgMg}$ , can simultaneously segment multiple objects (in multiple frames) using a single 24GB GPU.", + "bbox": [ + 75, + 408, + 468, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fast $\\mathrm{SgMg}$ shares visual features as well as vision-language features for all referred objects to make the network efficient, and decouples the shared features to make them instance-specific before the segmentation stage. Firstly, visual features $(\\mathcal{F}_v)$ and language features $(\\mathcal{F}_w$ and $\\mathcal{F}_s)$ are extracted. Next, we associate $\\mathcal{F}_v$ and $\\mathcal{F}_w$ using multi-instance fusion rather than the previous SCF. Multi-instance fusion is built on the foundation of SCF, which is depicted in Fig. 3. The difference is that multi-instance fusion includes semantic fusion, which performs an elementwise add operation, after cross-attention to merge vision-language features of different expressions. The features after semantic fusion perform Hadamard product with $\\mathcal{F}_v$ to generate the vision-language features $\\mathcal{F}_{vl}$ for all objects:", + "bbox": [ + 75, + 517, + 468, + 695 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S F} \\left(\\mathcal {F} _ {w}, \\mathcal {F} _ {v}\\right) = \\sum_ {i = 1} ^ {N} \\operatorname {A t t} \\left(\\mathcal {F} _ {w} ^ {i}, \\mathcal {F} _ {v}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 701, + 468, + 735 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M I F} \\left(\\mathcal {F} _ {w}, \\mathcal {F} _ {v}\\right) = \\operatorname {S A} \\left(\\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right) \\otimes \\operatorname {S F} \\left(\\mathcal {F} _ {w}, \\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right)\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 758, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes Hadamard product and $N$ denotes the number of expressions. After vision-language fusion, we encode $\\mathcal{F}_{vl}$ using the transformer encoder to enrich its semantic information, and plug multi-instance decoupling to decouple features for each instance. Multi-instance decoupling employs $\\mathcal{F}_w$ and cross-attention to decouple $\\mathcal{F}_{vl}$ to", + "bbox": [ + 75, + 783, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "predict instance embeddings $\\mathcal{E}$ for different referred objects. These embeddings are then projected to CPKs to predict the patch masks. Thus, FAST SgMg shares features, which account for most of the computational overhead, for different expressions, making it efficient for referring segmentation.", + "bbox": [ + 497, + 125, + 890, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Instance Matching and Loss Functions", + "text_level": 1, + "bbox": [ + 499, + 204, + 831, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [4, 57], we perform instance matching with $N = 5$ learnable instance queries to improve fault tolerance. These queries are projected to CPKs to predict $N$ potential patch masks $\\mathcal{M}_P$ for each expression. The Hungarian algorithm [22] is then adopted to select the best match based on the matching loss for training. During inference, we directly employ the predicted confidence scores $S$ to measure the instance queries and select the results.", + "bbox": [ + 497, + 226, + 890, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt the same training losses and weights as used in [57, 72] for a fair comparison. Specifically, we use Dice loss [26] and Focal loss [31] for patch mask $\\mathcal{M}_P$ and optimized mask $\\mathcal{M}_O$ , Focal loss [31] for confidence scores $S$ , and L1 and GIoU [48] loss for bounding boxes $\\mathcal{B}$ . The final training loss functions are:", + "bbox": [ + 497, + 336, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a i n}} = \\lambda_ {\\mathcal {M} _ {P}} \\mathcal {L} _ {\\mathcal {M} _ {P}} + \\lambda_ {\\mathcal {M} _ {O}} \\mathcal {L} _ {\\mathcal {M} _ {O}} + \\lambda_ {\\mathcal {B}} \\mathcal {L} _ {\\mathcal {B}} + \\lambda_ {\\mathcal {S}} \\mathcal {L} _ {\\mathcal {S}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 422, + 890, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ and $\\lambda$ are the loss term and weight, respectively.", + "bbox": [ + 502, + 446, + 880, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 499, + 472, + 631, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 499, + 496, + 697, + 509 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate $\\mathrm{SgMg}$ on four video benchmarks: Ref-YouTube-VOS [49], Ref-DAVIS17 [21], A2D-Sentences [17], and JHMDB-Sentences [20]. RefYouTube-VOS is currently the largest dataset for R-VOS, containing 3,978 videos with about $13\\mathrm{K}$ expressions. RefDAVIS17 is an extension of DAVIS17 [45] by including the language expressions of different objects and contains 90 videos. A2D-Sentences is a general actor and action segmentation dataset with over 3.7K videos and 6.6K action descriptions. JHMDB-Sentences includes 928 videos and 928 descriptions covering 21 different action classes.", + "bbox": [ + 497, + 517, + 890, + 668 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt the standard metrics to evaluate our models: region similarity $\\mathcal{I}$ (average IoU), contour accuracy $\\mathcal{F}$ (average boundary similarity), and their mean value $\\mathcal{I} \\& \\mathcal{F}$ . All results are evaluated using the official code or server. On A2D-Sentences and JHMDB-Sentences, we adopt mAP, overall IoU, and mean IoU for evaluation.", + "bbox": [ + 497, + 669, + 890, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 499, + 762, + 715, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [4, 6, 57], we train our models on the training set of Ref-YouTube-VOS, and directly evaluate them on the validation split of Ref-YouTube-VOS and Ref-DAVIS17 without any additional techniques, e.g., model ensemble, joint training, and mask propagation, since they are not the focus of this paper. Additionally, we present results for our", + "bbox": [ + 497, + 783, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "924", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/03b0aa0f0b4461880ac4be5591110fa1d089e948417be1ed8e301ada90f11e57.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodYearBackboneRef-YouTube-VOSRef-DAVIS17
J&FJFFPSJ&FJF
CMSA [63]2019ResNet-5036.434.838.1-40.236.943.5
URVOS [49]2020ResNet-5047.245.349.2-51.547.356.0
CMPC-V [32]2021I3D47.545.649.3----
PMINet [11]2021ResNeSt-10153.051.554.5----
YOFO [23]2022ResNet-5048.647.549.71053.348.857.8
LBDT [10]2022ResNet-5049.448.250.6-54.3--
MLRL [56]2022ResNet-5049.748.451.0-52.850.055.4
MTTR [4]2022Video-Swin-T55.354.056.6----
MANet [6]2022Video-Swin-T55.654.856.5----
ReferFormer [57]2022Video-Swin-T56.054.857.350---
SgMg (Ours)2023Video-Swin-T58.957.760.06556.753.360.0
", + "bbox": [ + 139, + 123, + 828, + 301 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dc3f028f9e0cd45daf190f6ec9ff502797e6f61ff60845ec66307fa5c105045e.jpg", + "table_caption": [ + "Pre-training with RefCOCO/+/g & larger backbone" + ], + "table_footnote": [], + "table_body": "
ReferFormer [57]2022Video-Swin-T59.458.060.95059.656.562.7
SgMg (Ours)2023Video-Swin-T62.060.463.56561.959.064.8
ReferFormer [57]2022Video-Swin-B62.961.364.63361.158.164.1
SgMg (Ours)2023Video-Swin-B65.763.967.44063.360.666.0
", + "bbox": [ + 141, + 318, + 828, + 375 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ba07cbfefe8f697b403953944b94e99ba3625f1163306b85755c2c0feee15f43.jpg", + "table_caption": [ + "Table 2. Quantitative comparison to state-of-the-art methods on the validation split of Ref-YouTube-VOS and Ref-DAVIS17." + ], + "table_footnote": [], + "table_body": "
MethodBackboneA2D-SentencesJHMDB-Sentences
mAPOverall IoUMean IoUmAPOverall IoUMean IoU
Hu et al. [19]VGG-1613.247.435.017.854.652.8
Gavrilyuk et al. [17]I3D19.853.642.123.354.154.2
ACAN [54]I3D27.460.149.028.957.658.4
CMPC-V [32]I3D40.465.357.334.261.661.7
ClawCraneNet [29]ResNet-50/101-63.159.9-64.465.6
MTTR [4]Video-Swin-T46.172.064.039.270.169.8
ReferFormer [57]Video-Swin-T52.877.669.642.271.971.0
SgMg (Ours)Video-Swin-T56.178.070.444.472.871.7
ReferFormer [57]Video-Swin-B55.078.670.343.773.071.8
SgMg (Ours)Video-Swin-B58.579.972.045.073.772.5
", + "bbox": [ + 139, + 399, + 831, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison to state-of-the-art R-VOS methods on A2D-Sentences and JHMDB-Sentences.", + "bbox": [ + 151, + 566, + 815, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models first pre-trained on RefCOCO+/g [39, 65] and then fine-tuned on Ref-YouTube-VOS. Similar to [57, 72], we set the coefficients for different losses $\\lambda_{dice}$ , $\\lambda_{focal}$ , $\\lambda_{L1}$ , $\\lambda_{giou}$ to 5, 2, 5, 2, respectively. The models are trained using 2 RTX 3090 GPUs with 5 frames per clip for 9 epochs. All frames are resized to have the longest side of 640 pixels. Further implementation details are in the supplementary material.", + "bbox": [ + 75, + 600, + 468, + 711 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 75, + 720, + 270, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ref-YouTube-VOS and Ref-DAVIS17. We compare $\\mathrm{SgMg}$ with recently published works in Table 2. Our approach surpasses present solutions on the two datasets across all metrics. On Ref-YouTube-VOS, $\\mathrm{SgMg}$ with the Video Swin Tiny backbone achieves $58.9\\mathcal{J}\\&\\mathcal{F}$ at 65 FPS, which is $2.9\\%$ higher and $1.3\\times$ faster than the previous state-of-the-art ReferFormer [57]. Our approach runs faster due to the use of the segment-and-optimize pipeline, which avoids the need for heavy feature decoders. When pre", + "bbox": [ + 75, + 741, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "training with RefCOCO/+/g and using a larger backbone, i.e., Video Swin Base, the performance of $\\mathrm{SgMg}$ further boosts to 65.7 $\\mathcal{J}\\&\\mathcal{F}$ , consistently leading all other solutions by more than $2.8\\%$ . On Ref-DAVIS17, $\\mathrm{SgMg}$ achieves 63.3 $\\mathcal{J}\\&\\mathcal{F}$ , outperforming state-of-the-art by $2.2\\%$ and demonstrating the generality of our approach.", + "bbox": [ + 497, + 600, + 890, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A2D-Sentences and JHMDB-Sentences. We further evaluate $\\mathrm{SgMg}$ on A2D-Sentences and JHMDB-Sentences in Table 3. Following [57], the models are first pre-trained on RefCOCO+/g and then fine-tuned on A2D-Sentences. JHMDB-Sentences is used only for evaluation. As shown in Table 3, $\\mathrm{SgMg}$ achieves superior performance compared to other state-of-the-art R-VOS methods and surpasses the nearest competitor Referformer [57] by $3.5 / 1.3\\%$ mAP on A2D-Sentences and JHMDB-Sentences, respectively.", + "bbox": [ + 497, + 685, + 890, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-object R-VOS. We extend $\\mathrm{SgMg}$ to perform multi-object R-VOS, which is more practical and efficient for deployment. Fast $\\mathrm{SgMg}$ is trained on Ref-YouTube-VOS without pre-training or postprocessing techniques. We", + "bbox": [ + 497, + 809, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "925", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e0c167ebe36658d6be154ac5022a6ac627ec2e510bbd1b774c6ad8f4a3677b38.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRef-DAVIS17Ref-YouTube-VOS
J&FJFJ&FJFFPS
ReferFormer [57]54.551.058.056.054.857.350
Fast SgMg (Ours)54.251.157.354.253.155.3185
", + "bbox": [ + 78, + 123, + 470, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f7841f11983ce39fe5d13c2a03de7d9150de07f0859b010f5cb900a698aa2b78.jpg", + "table_caption": [ + "Table 4. Evaluation of Fast SgMg on Ref-DAVIS17 and RefYouTube-VOS. Video-Swin-T is adopted as the backbone." + ], + "table_footnote": [], + "table_body": "
ComponentsPerformance
CPKMSOSCFJ&FJFFPS
54.452.756.270
✓55.854.557.170
✓57.756.359.169
✓55.854.357.466
✓✓57.956.759.169
✓✓✓58.957.760.065
", + "bbox": [ + 102, + 218, + 443, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "benchmark Fast SgMg on Ref-YouTube-VOS and RefDAVIS17 using the commonly used Video Swin Tiny, and compare the results with the state-of-the-art R-VOS method, which performs single-object segmentation.", + "bbox": [ + 75, + 366, + 468, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 4, Fast $\\mathrm{SgMg}$ achieves reasonable performance and runs about $3.7\\times$ faster (185 vs 50 FPS) compared to ReferFormer [57]. It should be noted that each object in the above datasets contains multiple expressions. On Ref-DAVIS17, we group expressions to have only one expression per object within each group and segment all expressions in each group simultaneously since the object identity is given. On Ref-YouTube-VOS, all expressions in a video are segmented simultaneously due to the lack of object identity, making it more challenging.", + "bbox": [ + 75, + 422, + 468, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study for Different Components", + "text_level": 1, + "bbox": [ + 75, + 568, + 431, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments to evaluate the effectiveness of different components in $\\mathrm{SgMg}$ . The components are added to the baseline model step-by-step.", + "bbox": [ + 75, + 590, + 468, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conditional Patch Kernel. As shown in Table 5, CPK boosts the performance by $1.4\\%$ compared with the recent instance-aware conditional kernels [57]. The sequential labels of each token predicted by CPK contain more fine-grained information, making the prediction more accurate.", + "bbox": [ + 75, + 631, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-granularity Segmentation Optimizer. We devise MSO to optimize the predicted patch masks. As shown in Table 5, MSO improves the performance by $3.3\\%$ , indicating the importance of fine-grained visual details in R-VOS. Spectrum-guided Cross-modal Fusion. We present SCF to perform global interactions by operating in the spectral domain. In Table 5, using SCF to replace the traditional cross-attention in [57, 52] improves the $\\mathcal{J}\\&\\mathcal{F}$ by $1.4\\%$ . We consider SCF extracts important low-frequency features and facilitates multimodal understanding globally, which is suitable for R-VOS since locating referred objects requires understanding the global context and token relations.", + "bbox": [ + 75, + 700, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e789f9b848a06b92e7859a139cc944de2de6a5d7ebe0aca650919001541463d.jpg", + "table_caption": [ + "Table 5. Ablation of different components on Ref-YouTube-VOS." + ], + "table_footnote": [], + "table_body": "
SettingsDriftPipelineJ&FFPS
Baseline + Decoder✓decode-and-segment56.050
Baseline + Decoder + MSO✓decode-and-segment56.449
Baseline + MSO×segment-and-optimize57.769
", + "bbox": [ + 500, + 123, + 890, + 179 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cce2e01bab5d357ae545be79411791a1336fc3e6319b6a1d1a33c0b56865542e.jpg", + "table_caption": [ + "Table 6. Feature drift analysis using ReferFormer [57] (Baseline + Decoder) and $\\mathrm{SgMg}$ w/o CPK & SCF (Baseline + MSO). Significant improvement is achieved by addressing the drift issue (last row). Adding MSO on top of ReferFormer to recover visual details (for a second time) still performs worse than our basic pipeline." + ], + "table_footnote": [], + "table_body": "
MethodRefCOCORefCOCO+RefCOCOg
MaIL [28]70.162.262.5
CRIS [55]70.562.359.9
RefTR [25]70.6--
LAVT [62]72.762.161.2
VLT [9]73.063.563.5
SgMg (Ours)76.366.470.0
", + "bbox": [ + 514, + 255, + 878, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Quantitative evaluation on the validation split of Ref-COCO/+/g. Overall IoU is adopted as the evaluation metric.", + "bbox": [ + 499, + 357, + 890, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study for Feature Drift", + "text_level": 1, + "bbox": [ + 499, + 406, + 784, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation study in Table 6 to demonstrate the feature drift problem. Our segment-and-optimize pipeline addresses the adverse drift effect discussed in Section 3.1 to significantly outperform ReferFormer [57] by $1.7\\%$ points and runs $1.4 \\times$ faster. Furthermore, adding MSO on top of ReferFormer still performs worse due to the negative drift impact caused by the decode-and-segment pipeline. These results demonstrate the efficacy of our proposed segment-and-optimize pipeline.", + "bbox": [ + 497, + 428, + 890, + 552 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Referring Image Segmentation Results", + "text_level": 1, + "bbox": [ + 499, + 562, + 831, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We apply $\\mathrm{SgMg}$ to referring image (expression) segmentation without any architectural modifications, and compare against the current state-of-the-art methods on RefCOCO $+ / g$ [39, 65]. A single $\\mathrm{SgMg}$ model is trained on RefCOCO $+ / g$ without large-scale pre-training. As shown in Table 7, $\\mathrm{SgMg}$ achieves advanced performance on all three benchmarks. These results demonstrate the efficacy of $\\mathrm{SgMg}$ in referring image segmentation.", + "bbox": [ + 497, + 585, + 890, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Inference Time Analysis of Multi-Object RVOS", + "text_level": 1, + "bbox": [ + 499, + 706, + 890, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the efficiency of the proposed multi-object R-VOS paradigm by comparing the FPS of Fast $\\mathrm{SgMg}$ and $\\mathrm{SgMg}$ on videos with different numbers of expressions. As illustrated in Fig. 6, Fast $\\mathrm{SgMg}$ performs about $2\\times$ faster than $\\mathrm{SgMg}$ when there are two expressions per video on average. As the number of expressions increases, Fast $\\mathrm{SgMg}$ achieves faster reasoning time per object per frame due to its utilization of the multi-object R-VOS paradigm. When there are ten expressions in each video, Fast $\\mathrm{SgMg}$ performs at nearly 300 FPS, which is about $5\\times$ faster than $\\mathrm{SgMg}$ .", + "bbox": [ + 497, + 726, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "926", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cbcb32da13ed3ec608c9c4aada756dbe130df296a2ba08aa0c6443ae4e1029e8.jpg", + "image_caption": [ + "Expression: a turtle to the bottom left.", + "Figure 5. Qualitative comparison of our method with others." + ], + "image_footnote": [], + "bbox": [ + 97, + 138, + 867, + 377 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/55fe67df66eb4d6cc68ac6b33fcae467955bcb6523e69cd99e50525433d96014.jpg", + "image_caption": [ + "Figure 6. Efficiency analysis of $\\mathrm{SgMg}$ and Fast $\\mathrm{SgMg}$ for videos with different numbers of expressions on Ref-YouTube-VOS." + ], + "image_footnote": [], + "bbox": [ + 82, + 408, + 465, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.8. Qualitative Results", + "text_level": 1, + "bbox": [ + 77, + 597, + 260, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 5, we show qualitative comparison with ReferFormer [57] and MTTR [4]. $\\mathrm{SgMg}$ can handle different objects of the same category or with the same behavior.", + "bbox": [ + 75, + 618, + 468, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.9. Feature Visualization of SCF", + "text_level": 1, + "bbox": [ + 77, + 669, + 336, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 7, we visualize the vision-language features extracted by our SCF in comparison to the cross-attention used in [57]. The features extracted by SCF exhibit superior grounding ability in locating target objects, resulting in better performance for $\\mathrm{SgMg}$ .", + "bbox": [ + 75, + 691, + 470, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 77, + 773, + 196, + 788 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We discovered the feature drift issue in current referring video object segmentation (R-VOS) methods, which negatively affects the segmentation kernels. We presented $\\mathrm{SgMg}$ , a novel segment-and-optimize approach for R-VOS that avoids the drift issue and optimizes masks with visual", + "bbox": [ + 75, + 796, + 470, + 866 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e457b08af0144c867c4fde393fa6bb8ccf2eada1d54b3835be531cac4aa387d0.jpg", + "image_caption": [ + "Figure 7. Visualization of the vision-language features extracted w/o and w/ our SCF." + ], + "image_footnote": [], + "bbox": [ + 505, + 408, + 890, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "details. We also provided a new perspective to encourage vision-language global interactions in the spectral domain with Spectrum-guided Cross-modal Fusion. Additionally, we proposed the multi-object R-VOS paradigm by extending $\\mathrm{SgMg}$ with multi-instance fusion and decoupling. Finally, we evaluated our models on four video benchmarks and demonstrated that our approach achieves state-of-the-art performance on all four datasets.", + "bbox": [ + 497, + 659, + 890, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This research was supported by the Australian Research Council Industrial Transformation Research Hub IH180100002. Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government.", + "bbox": [ + 497, + 776, + 892, + 859 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "927", + "bbox": [ + 472, + 885, + 497, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 124, + 173, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nayyer Aafaq, Naveed Akhtar, Wei Liu, Syed Zulqarnain Gilani, and Ajmal Mian. Spatio-temporal dynamics and semantic attribute enriched visual encoding for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12487-12496, 2019. 2", + "[2] Miriam Bellver, Carles Ventura, Carina Silberer, Ioannis Kazakos, Jordi Torres, and Xavier Giro-i Nieto. A closer look at referring expressions for video object segmentation. Multimedia Tools and Applications, pages 1-20, 2022. 1", + "[3] Glenn D Bergland. A guided tour of the fast fourier transform. IEEE spectrum, 6(7):41-52, 1969. 3", + "[4] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In CVPR, pages 4985-4995, 2022. 1, 2, 3, 4, 5, 6, 8", + "[5] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 221-230, 2017. 2", + "[6] Weidong Chen, Dexiang Hong, Yuankai Qi, Zhenjun Han, Shuhui Wang, Laiyun Qing, Qingming Huang, and Guorong Li. Multi-attention network for compressed video referring object segmentation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4416–4425, 2022. 1, 5, 6", + "[7] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Proc. Adv. Neural Inf. Process. Syst. (NIPS), 34:11781-11794, 2021. 1, 2", + "[8] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3", + "[9] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 7", + "[10] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 2, 6", + "[11] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, page 7, 2021. 2, 5, 6", + "[12] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 2", + "[13] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 2" + ], + "bbox": [ + 78, + 147, + 470, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Guang Feng, Lihe Zhang, Zhiwei Hu, and Hutchuan Lu. Deeply interleaved two-stream encoder for referring video segmentation. arXiv preprint arXiv:2203.15969, 2022. 2", + "[15] Mingtao Feng, Haoran Hou, Liang Zhang, Yulan Guo, Hongshan Yu, Yaonan Wang, and Ajmal Mian. Exploring hierarchical spatial layout cues for 3d point cloud based scene graph prediction. IEEE Transactions on Multimedia, 2023. 1", + "[16] Zhiheng Fu, Siyu Hong, Mengyi Liu, Hamid Laga, Mohammed Bennamoun, Farid Boussaid, and Yulan Guo. Multi-stage information diffusion for joint depth and surface normal estimation. Pattern Recognition, 141:109660, 2023. 1", + "[17] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 2, 5, 6", + "[18] Lorenzo Giambagli, Lorenzo Buffoni, Timoteo Carletti, Walter Nocentini, and Duccio Fanelli. Machine learning in spectral domain. Nature communications, 12(1):1-9, 2021. 3", + "[19] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 6", + "[20] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 2, 5", + "[21] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Asian Conference on Computer Vision, pages 123-141. Springer, 2018. 1, 2, 5", + "[22] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5", + "[23] Dezhuang Li, Ruoqi Li, Lijun Wang, Yifan Wang, Jinqing Qi, Lu Zhang, Ting Liu, Qingquan Xu, and Huchuan Lu. You only infer once: Cross-modal meta-transfer for referring video object segmentation. In AAAI Conference on Artificial Intelligence, 2022. 1, 5, 6", + "[24] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2", + "[25] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in neural information processing systems, 34:19652-19664, 2021.7", + "[26] Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855, 2019. 5", + "[27] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Yan Lu, and Bhiksha Raj. R^2vos: Robust referring video object segmentation via relational multimodal cycle consistency. arXiv preprint arXiv:2207.01203, 2022. 1" + ], + "bbox": [ + 502, + 125, + 890, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "928", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Zizhang Li, Mengmeng Wang, Jianbiao Mei, and Yong Liu. Mail: A unified mask-image-language trimodal network for referring image segmentation. arXiv preprint arXiv:2111.10747, 2021. 7", + "[29] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawcranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 2, 6", + "[30] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 2", + "[31] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5", + "[32] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 6", + "[33] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3", + "[34] Yong Liu, Ran Yu, Jiahao Wang, Xinyuan Zhao, Yitong Wang, Yansong Tang, and Yujiu Yang. Global spectral filter memory network for video object segmentation. In European Conference on Computer Vision, pages 648-665. Springer, 2022. 3", + "[35] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 3", + "[36] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2", + "[37] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10437–10446, 2020. 2", + "[38] Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsupervised video object segmentation with co-attention siamese networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3623–3632, 2019. 2", + "[39] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6, 7", + "[40] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and" + ], + "bbox": [ + 78, + 125, + 468, + 866 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3", + "[41] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Region aware video object segmentation with deep motion modeling. arXiv preprint arXiv:2207.10258, 2022.1", + "[42] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Self-supervised video object segmentation by motion-aware mask propagation. In Proceedings of the IEEE International Conference on Multimedia and Expo, 2022. 2", + "[43] Bo Miao, Liguang Zhou, Ajmal Mian, Tin Lun Lam, and Yangsheng Xu. Object-to-scene: Learning to transfer object knowledge to indoor scene recognition. In Proc. IEEE Int. Conf. Intell. Robots Syst. (IROS). IEEE, 2021. 1", + "[44] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proc. Int. Conf. Comput. Vis. (ICCV), pages 9226–9235, 2019. 1, 2", + "[45] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 5", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[47] Sucheng Ren, Wenxi Liu, Yongtuo Liu, Haoxin Chen, Guoqiang Han, and Shengfeng He. Reciprocal transformations for unsupervised video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15455-15464, 2021. 2", + "[48] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5", + "[49] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223. Springer, 2020. 1, 2, 5, 6", + "[50] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In European conference on computer vision, pages 282-298. Springer, 2020. 1, 2, 3, 4", + "[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 2", + "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 7", + "[53] Paul Voiglaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast" + ], + "bbox": [ + 502, + 125, + 890, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "929", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "end-to-end embedding learning for video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9481-9490, 2019. 2", + "[54] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 6", + "[55] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 7", + "[56] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 1, 5, 6", + "[57] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[58] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. arXiv preprint arXiv:2308.02874, 2023.1", + "[59] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019.4", + "[60] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In International Conference on Neural Information Processing, pages 264-274. Springer, 2019. 4", + "[61] Zhao Yang, Yansong Tang, Luca Bertinetto, Hengshuang Zhao, and Philip HS Torr. Hierarchical interaction network for video object segmentation from referring expressions. In BMVC, 2021. 1, 2", + "[62] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Ladv: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 7", + "[63] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 5, 6", + "[64] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 4", + "[65] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 6, 7" + ], + "bbox": [ + 78, + 125, + 470, + 865 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Z. Yang and Y, Wei and Y. Yang. Associating objects with transformers for video object segmentation. Proc. Adv. Neural Inf. Process. Syst. (NIPS), 34:2491-2502, 2021. 2", + "[67] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 2", + "[68] Shuai Zhao, Xiaohan Wang, Linchao Zhu, and Yi Yang. Test-time adaptation with clip reward for zero-shot generalization in vision-language models. arXiv preprint arXiv:2305.18010, 2023. 2", + "[69] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. arXiv preprint arXiv:2205.00823, 2022. 2", + "[70] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 2", + "[71] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13041-13049, 2020. 2", + "[72] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 5, 6" + ], + "bbox": [ + 502, + 127, + 890, + 511 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "930", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/dc8d764a-9214-4358-ad24-c2007b58a08c_model.json b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/dc8d764a-9214-4358-ad24-c2007b58a08c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb104ca437a833ab8e65874a56b7ea14445b81cf --- /dev/null +++ b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/dc8d764a-9214-4358-ad24-c2007b58a08c_model.json @@ -0,0 +1,2455 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.086 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.254, + 0.049, + 0.79, + 0.087 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.162, + 0.857, + 0.183 + ], + "angle": 0, + "content": "Spectrum-guided Multi-granularity Referring Video Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.207, + 0.772, + 0.248 + ], + "angle": 0, + "content": "Bo Miao\\(^{1}\\), Mohammed Bennamoun\\(^{1}\\), Yongsheng Gao\\(^{2}\\), Ajmal Mian\\(^{1}\\) \n\\(^{1}\\)The University of Western Australia \n\\(^{2}\\)Griffith University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.291, + 0.314, + 0.305 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.652 + ], + "angle": 0, + "content": "Current referring video object segmentation (R-VOS) techniques extract conditional kernels from encoded (low-resolution) vision-language features to segment the decoded high-resolution features. We discovered that this causes significant feature drift, which the segmentation kernels struggle to perceive during the forward computation. This negatively affects the ability of segmentation kernels. To address the drift problem, we propose a Spectrum-guided Multigranularity (SgMg) approach, which performs direct segmentation on the encoded features and employs visual details to further optimize the masks. In addition, we propose Spectrum-guided Cross-modal Fusion (SCF) to perform intra-frame global interactions in the spectral domain for effective multimodal representation. Finally, we extend SgMg to perform multi-object R-VOS, a new paradigm that enables simultaneous segmentation of multiple referred objects in a video. This not only makes R-VOS faster, but also more practical. Extensive experiments show that SgMg achieves state-of-the-art performance on four video benchmark datasets, outperforming the nearest competitor by \\(2.8\\%\\) points on Ref-YouTube-VOS. Our extended SgMg enables multi-object R-VOS, runs about \\(3\\times\\) faster while maintaining satisfactory performance. Code is available at https://github.com/bo-miao/SgMg." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.677, + 0.21, + 0.691 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Referring video object segmentation (R-VOS) aims at segmenting objects in a video, referred to by linguistic descriptions. R-VOS is an emerging task for multimodal reasoning and promotes a wide range of applications, including language-guided video editing and human-machine interaction. Different from conventional semi-supervised video object segmentation [44, 7, 41], where the mask annotation for the first frame is provided for reference, R-VOS is more challenging due to the need for cross-modal understanding between vision and free-form language expressions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Early R-VOS techniques [2, 21, 61] perform feature encoding, cross-modal interaction, and language grounding" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.291, + 0.891, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.435, + 0.752, + 0.444 + ], + "angle": 0, + "content": "(b) SgMg (Ours): segment-and-optimize w/ encoded features \\(F_{vl}\\)" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.452, + 0.885, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.517, + 0.88, + 0.549 + ], + "angle": 0, + "content": "Conditional Segmentation Fusion: Cross-modal Fusion \\(F_{\\mu l}\\) : Encoded Vision-Language Features \\(F_{\\nu l}^{d}\\) : Decoded Features Q: Instance Query \\(K_{c}\\) : Conditional Kernel \\(K_{cp}\\) : Conditional Patch Kernel \\(M_{Q}\\) : Predicted Mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.552, + 0.892, + 0.63 + ], + "angle": 0, + "content": "Figure 1. (a) Previous methods [4, 57] apply segmentation kernels \\(\\mathcal{K}_c\\) [50], extracted from encoded features \\(\\mathcal{F}_{vl}\\), to segment the decoded high-resolution features \\(\\mathcal{F}_{vl}^d\\). (b) We use segmentation kernels \\(\\mathcal{K}_{cp}\\), extracted from encoded features \\(\\mathcal{F}_{vl}\\), to segment the encoded features \\(\\mathcal{F}_{vl}\\) directly, and propose multi-granularity optimization to recover visual details and produce fine-grained masks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.868 + ], + "angle": 0, + "content": "using convolutional neural networks (CNNs). However, the limited ability of CNNs to capture long-range dependencies and handle free-form features constrains the model performance. With the advancement of attention mechanisms [52, 43, 15, 16, 58], recent methods achieved significant improvement on R-VOS using cross-attention [49, 23, 27] for multimodal understanding and transformers [6, 56] for spatio-temporal representation. Based on transformers, conditional kernel [50] is then introduced to separate foreground from semantic features given its high adaptability to different instances [4, 57]. As illustrated in Fig. 1(a), these methods attend to encoded vision-language features \\(\\mathcal{F}_{vl}\\) using instance queries \\(Q\\) to predict conditional kernels \\(\\kappa_{c}\\), and employ \\(\\kappa_{c}\\) as the segmentation head to segment decoded features \\(\\mathcal{F}_{vl}^{d}\\). Despite the promising performance," + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.887, + 0.498, + 0.898 + ], + "angle": 0, + "content": "920" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.127, + 0.469, + 0.265 + ], + "angle": 0, + "content": "this paradigm still has some limitations. Firstly, as shown in the t-SNE [51] visualization in Fig. 1(a), although the nonlinear decoding process introduces visual details, this is accompanied by a significant feature drift, which increases the difficulty of segmentation since \\(\\mathcal{K}_c\\) is predicted before feature decoding. Secondly, bilinear upsampling of the predicted masks \\(\\mathcal{M}_Q\\) to increase resolution impedes the segmentation performance. Thirdly, these methods only support single expression-based segmentation, making R-VOS inefficient when multiple referred objects exist in a video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.471, + 0.514 + ], + "angle": 0, + "content": "In this work, we propose a Spectrum-guided Multi-granularity (SgMg) approach that follows a segment-and-optimize pipeline to address the above problems. As depicted in Fig. 1(b), SgMg introduces Conditional Patch Kernel (CPK) \\(\\mathcal{K}_{cp}\\) to directly segment its fully perceived encoded features \\(\\mathcal{F}_{vl}\\), avoiding the feature drift and its adverse effects. The segmentation is then refined using our proposed Multi-granularity Segmentation Optimizer (MSO), which employs low-level visual details to produce full-resolution masks. Within the SgMg framework, we further develop Spectrum-guided Cross-modal Fusion (SCF) that performs intra-frame global interactions in the spectral domain to facilitate multimodal understanding. Finally, we introduce a new paradigm called multi-object R-VOS to simultaneously segment multiple referred objects in a video. To achieve this, we extend SgMg by devising multi-instance fusion and decoupling. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.531, + 0.47, + 0.601 + ], + "angle": 0, + "content": "- We explain how existing R-VOS methods suffer from the feature drift problem. To address this problem, we propose \\(\\mathrm{SgMg}\\) that follows a segment-and-optimize pipeline and achieves top-ranked overall performance on multiple benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.615, + 0.469, + 0.655 + ], + "angle": 0, + "content": "- We propose Spectrum-guided Cross-modal Fusion to encourage intra-frame global interactions in the spectral domain." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.67, + 0.469, + 0.726 + ], + "angle": 0, + "content": "- We extend \\(\\mathrm{SgMg}\\) to perform multi-object R-VOS, a new paradigm that enables simultaneous segmentation of multiple referred objects in a video. Our multi-object variant is more practical and runs \\(3\\times\\) faster." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.531, + 0.47, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.469, + 0.867 + ], + "angle": 0, + "content": "We conduct extensive experiments on multiple benchmark datasets, including Ref-YouTube-VOS [49], RefDAVIS17 [21], A2D-Sentences [17], and JHMDB-Sentences [20], and achieve state-of-the-art performance on all four. On the largest validation set Ref-YouTube-VOS, \\(\\mathrm{SgMg}\\) achieves \\(65.7\\mathcal{J}\\& \\mathcal{F}\\) which is \\(2.8\\%\\) points higher than that of the closest competitor ReferFormer [57]. On the A2D-Sentences, \\(\\mathrm{SgMg}\\) achieves \\(58.5\\mathrm{mAP}\\) which is \\(3.5\\%\\) points higher than that of ReferFormer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.125, + 0.65, + 0.14 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.148, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Video Object Segmentation techniques fall into two categories: unsupervised and semi-supervised. Unsupervised approaches segment the most salient instances in each video without user interactions [38, 47]. They often employ two-stream networks to fuse motion and appearance cues for segmentation. Semi-supervised approaches track the given first frame object mask by performing online learning [5] or spatial-temporal association [44, 7, 66, 42, 53]. Unlike conventional semi-supervised video object segmentation, R-VOS takes a free-form linguistic expression as guidance to detect and segment referred objects in videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Referring Video Object Segmentation. R-VOS methods mainly use deep neural networks with vision-and-language interaction to empower visual features with corresponding linguistic information for pixel-level segmentation. For example, [49] employs a unified R-VOS framework that performs iterative segmentation guided by both language and temporal features. [32, 29] adopt progressive segmentation by perceiving potential objects and discriminating the best match. [70] fuses visual and motion features for segmentation under the guidance of linguistic cues. [30] models object relations to form tracklets and performs tracklet-language grounding. To enhance multi-modal interactions, [61, 14, 11, 10] perform hierarchical vision-language fusion on multiple feature layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Despite their promising performance, the complex multistage pipelines and use of multiple networks make R-VOS burdensome. To address these problems, MTTR [4] proposes an end-to-end transformer-based network with conditional kernels [50] to segment target objects. ReferFormer [57] further introduces language-guided instance queries to predict instance-aware conditional kernels and an auxiliary detection task to aid localization. These methods follow a decode-and-segment pipeline, which adopts conditional kernels to segment decoded high-resolution features to achieve promising performance. However, the nonlinear decoding process leads to significant feature drift that negatively affects the conditional kernels. In contrast to previous works, our approach follows a segment-and-optimize pipeline to avoid the adverse drift effects and to predict full-resolution masks in an efficient manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Vision and Language Representation Learning aims to learn vision-language semantics and alignment for multimodal reasoning tasks. It has achieved significant success in various tasks [37, 67, 68], including video question answering [71], video captioning [1], video-text retrieval [12, 69], zero-shot classification [46], referring image/video segmentation [49], etc. Some approaches [46, 24] rely on contrastive pre-training using large-scale datasets to project different modalities into unified embedding space. Others [36, 13] develop cross-modal interaction layers for multimodal feature fusion and understanding. Recent deep" + }, + { + "type": "page_number", + "bbox": [ + 0.472, + 0.887, + 0.496, + 0.898 + ], + "angle": 0, + "content": "921" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.127, + 0.868, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.367, + 0.893, + 0.407 + ], + "angle": 0, + "content": "Figure 2. The overall framework of \\(\\mathrm{SgMg}\\). Taking a video sequence \\(\\mathcal{V} = \\{I_i\\}_{i=1}^T\\) and a language expression \\(\\mathcal{L} = \\{S_i\\}_{i=1}^N\\) as input, \\(\\mathrm{SgMg}\\) predicts the masks of referred object \\(\\mathcal{O}_{\\mathcal{L}}\\) in each frame. SCF projects visual features \\(\\mathcal{F}_v\\) to vision-language features \\(\\mathcal{F}_{vl}\\), instance-aware CPK predicts patch masks by segmenting encoded \\(\\mathcal{F}_{vl}\\), and MSO optimizes patch masks to get fine-grained results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.471, + 0.512 + ], + "angle": 0, + "content": "learning methods in spectral domain [18, 40, 8, 34] have raised widespread awareness because of their ability to perform global interactions. We take inspiration from these spectral-based methods and employ spectrum guidance in the field of vision-language representation to encourage multimodal global interactions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.532, + 0.47, + 0.566 + ], + "angle": 0, + "content": "3. \\(\\mathrm{SgMg}\\): Spectrum-guided Multi-granularity Referring Video Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Given a video sequence \\(\\mathcal{V} = \\{I_i\\}_{i=1}^T\\) with \\(T\\) frames and a language query \\(\\mathcal{L} = \\{S_i\\}_{i=1}^N\\) with \\(N\\) words. The goal of R-VOS is to segment the referred object \\(\\mathcal{O}_{\\mathcal{L}}\\) in \\(\\mathcal{V}\\) at pixel-level. To this end, we introduce a new approach termed \\(\\mathrm{SgMg}\\). Different from previous R-VOS methods [4, 57], our approach follows a segment-and-optimize pipeline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.867 + ], + "angle": 0, + "content": "An overview of \\(\\mathrm{SgMg}\\) is shown in Fig. 2. Video Swin Transformer [35] is adopted to extract visual feature \\(\\mathcal{F}_v\\) and RoBERTa [33] is adopted to extract sentence \\(\\mathcal{F}_s\\) and word \\(\\mathcal{F}_w\\) features. The channel dimension of all features is projected to 256. Spectrum-guided Cross-modal Fusion (SCF) cross attends \\(\\mathcal{F}_v\\) with \\(\\mathcal{F}_w\\) to compute vision-language features \\(\\mathcal{F}_{vl}\\). Deformable Transformer [72] encoder is used to encode \\(\\mathcal{F}_{vl}\\) and the decoder associates instance queries created based on \\(\\mathcal{F}_s\\) to predict instance embeddings and the corresponding Conditional Patch Kernels (CPKs). Finally, the CPKs are employed to segment \\(\\mathcal{F}_{vl}\\) and predict patch masks that are further optimized with visual details through Multi-granularity Segmentation Optimizer (MSO). The choice of the encoder and transformer follows previous works to avoid distractions [4, 57]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.428, + 0.708, + 0.443 + ], + "angle": 0, + "content": "3.1. Feature Drift Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.617 + ], + "angle": 0, + "content": "Existing R-VOS methods [57, 4] follow a decode-and-segment pipeline where conditional kernels \\(\\mathcal{K}_c\\) [50] are extracted from encoded features \\(\\mathcal{F}_{vl}\\) and used to segment the decoded features \\(\\mathcal{F}_{vl}^d\\). However, the decoding process leads to feature drift, which is evident in the t-SNE visualization depicted in Fig. 1(a). This drift is difficult for the kernels \\(\\mathcal{K}_c\\) to perceive during the forward computation since \\(\\mathcal{K}_c\\) is predicted before the feature decoding. Therefore, we argue that even though the feature decoding enhances visual details, it also causes the drift problem that negatively affects the segmentation kernels. This makes the existing decode-and-segment pipeline sub-optimal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.893, + 0.727 + ], + "angle": 0, + "content": "To overcome the adverse effects of feature drift while recovering visual details, we present \\(\\mathrm{SgMg}\\), a novel approach that follows a segment-and-optimize pipeline. In a nutshell, \\(\\mathrm{SgMg}\\) performs Spectrum-guided Cross-modal Fusion to compute \\(\\mathcal{F}_{vl}\\), leverages Conditional Patch Kernels to segment encoded features \\(\\mathcal{F}_{vl}\\) to avoid the drift effects, and recovers visual details with Multi-granularity Segmentation Optimizer to generate fine-grained masks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.735, + 0.828, + 0.75 + ], + "angle": 0, + "content": "3.2. Spectrum-guided Cross-modal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.892, + 0.868 + ], + "angle": 0, + "content": "The two-dimensional discrete Fourier transform converts spatial data into the spectral domain. Based on the spectral convolution theorem [3], point-wise update of signals in the spectral domain globally affects all inputs in the spatial domain, which gives the insight to design spectrum-based modules so as to efficiently facilitate global interactions, which is critical for multimodal understanding. In addition, Low-frequency components in the spectral domain usually" + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "922" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.123, + 0.462, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.083, + 0.304, + 0.284, + 0.313 + ], + "angle": 0, + "content": "(a) Spectrum-Guided Cross-Modal Fusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.303, + 0.436, + 0.313 + ], + "angle": 0, + "content": "(b) Spectrum Augmentation" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.316, + 0.468, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.339, + 0.469, + 0.376 + ], + "angle": 0, + "content": "Figure 3. Spectrum-guided Cross-modal Fusion. Imag.: Imaginary. Pre-spectrum augmentation and post-spectrum augmentation share an identical structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.469, + 0.424 + ], + "angle": 0, + "content": "correspond to the general semantic information according to previous theoretical studies [59, 60, 64]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.469, + 0.539 + ], + "angle": 0, + "content": "Inspired by the above observations, we conjecture that low-frequency components can benefit higher dimensional semantic features and propose Spectrum-guided Cross-modal Fusion (SCF). As shown in Fig. 3, SCF performs pre-spectrum augmentation to enhance visual features before cross-modal fusion and post-spectrum augmentation to facilitate global vision-language interactions after the fusion process. Let \\(\\mathcal{F} \\in \\mathbb{R}^{C \\times H \\times W}\\) denotes the input features, the spectrum augmentation (SA) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.547, + 0.469, + 0.562 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A} (\\mathcal {F}, K) = \\mathcal {F} + \\Theta_ {I F F T} (\\operatorname {C o n v} (\\sigma (K, \\mathcal {F}) \\odot \\Theta_ {F F T} (\\mathcal {F}))) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.709 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes low-pass filtering with adaptive Gaussian smoothed filters \\(\\sigma(K, \\mathcal{F})\\), which has the same spatial size as \\(\\mathcal{F}\\), and \\(K\\) is the bandwidth. To make \\(\\sigma(K, \\mathcal{F})\\) input-aware, we create an initial 2D Gaussian map based on \\(K\\), and apply pooling and linear layers on \\(\\mathcal{F}\\) to predict a scale parameter to update the Gaussian map. Thanks to the spectral convolution theorem, the efficient point-wise spectral convolution globally updates \\(\\mathcal{F}\\). We treat the spectral-operated features as residuals and add them to the original input features for enhancement. Overall, SCF, which takes visual features \\(\\mathcal{F}_v\\) and word-level text features \\(\\mathcal{F}_w\\) as input, is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.717, + 0.469, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S C F} \\left(\\mathcal {F} _ {w}, \\mathcal {F} _ {v}\\right) = \\operatorname {S A} \\left(\\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right) \\otimes \\operatorname {A t t} \\left(\\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right), \\mathcal {F} _ {w}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.749, + 0.364, + 0.763 + ], + "angle": 0, + "content": "3.3. Conditional Patch Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.469, + 0.866 + ], + "angle": 0, + "content": "We devise Conditional Patch Kernel (CPK) as the segmentation head to predict patch masks from the encoded vision-language features \\(\\mathcal{F}_{vl}\\) that are fully perceived by CPK. Unlike previous works [4, 57], CPK predicts a sequence of labels for each token rather than a single label, efficiently improving segmentation resolution along the channel dimension." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.125, + 0.891, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.237, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Figure 4. Multi-granularity Segmentation Optimizer, which predicts residual maps to optimize patch masks \\(\\mathcal{M}_P\\) progressively." + }, + { + "type": "image_footnote", + "bbox": [ + 0.5, + 0.263, + 0.892, + 0.291 + ], + "angle": 0, + "content": "Flatten: reshape \\(\\mathcal{M}_P\\) from \\(\\mathbb{R}^{\\frac{H}{i}\\times \\frac{W}{i}\\times p^2}\\) to \\(\\mathbb{R}^{\\frac{Hp}{i}\\times \\frac{Wp}{i}}\\) for visualization. \\(\\times 2\\): upsampling operation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.383 + ], + "angle": 0, + "content": "Specifically, we first use sentence-level text features \\(\\mathcal{F}_s\\) and multiple learnable embeddings to generate instance queries \\(Q\\in \\mathbb{R}^{N\\times C}\\). Next, \\(Q\\) is projected into instance embeddings \\(\\mathcal{E}\\in \\mathbb{R}^{N\\times C}\\) using the transformer decoder and \\(\\mathcal{E}\\) is leveraged to predict CPK for each instance query:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.393, + 0.891, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C P K} (Q, \\mathcal {F} _ {v l}) = \\Theta (\\operatorname {F C} (\\operatorname {A t t} (Q, \\mathcal {F} _ {v l}))) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.558 + ], + "angle": 0, + "content": "where \\(\\Theta\\) denotes the parameterization operation that reshapes CPK to form two point-wise convolutions with the output channel number of 16, which is similar to [50]. Since \\(Q\\) changes dynamically according to different linguistic expressions, CPK becomes instance-aware and can separate objects of interest from \\(F_{vl}\\). Finally, we apply the parameterized CPK (dynamic point-wise convolutions) on \\(F_{vl}\\) to predict patch masks \\(\\mathcal{M}_P\\in \\mathbb{R}^{\\frac{H}{i}\\times \\frac{W}{i}\\times p^2}\\), where \\(\\frac{H}{i}\\times \\frac{W}{i}\\) denotes the spatial resolution of \\(F_{vl}\\) and \\(p^2\\) denotes the increased segmentation resolution on the channel dimension." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.558, + 0.892, + 0.643 + ], + "angle": 0, + "content": "During inference, we can reshape patch masks to \\(\\mathcal{M}_P\\in\\) \\(\\mathbb{R}\\frac{Hp}{i}\\times \\frac{Wp}{i}\\) to efficiently generate fine-grained segmentation from low-resolution \\(F_{vl}\\). The resolution of prediction will be consistent with the input when \\(p\\) equals to \\(i\\). We found that this efficient CPK can achieve competitive performance compared to methods that use heavy decoders." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.866, + 0.667 + ], + "angle": 0, + "content": "3.4. Multi-granularity Segmentation Optimizer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.891, + 0.741 + ], + "angle": 0, + "content": "Segmenting encoded features \\(\\mathcal{F}_{vl}\\) with CPK avoids the detrimental drift effect on the segmentation head. However, visual details are required to produce accurate fine-grained masks. We propose Multi-granularity Segmentation Optimizer (MSO) to achieve this goal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.741, + 0.892, + 0.867 + ], + "angle": 0, + "content": "An overview of MSO is shown in Fig. 4. It takes the predicted patch masks \\(\\mathcal{M}_P\\) as object priors and reuses visual features \\(\\mathcal{F}_v\\) with spatial strides of \\(\\{4,8\\}\\) to gradually recover visual details and refine the priors. Specifically, MSO first concatenates \\(\\mathcal{M}_P\\) and \\(\\mathcal{F}_v\\) and projects them to low dimensional bases. Next, residual masks predicted by performing another convolution on these bases are used to correct \\(\\mathcal{M}_P\\). Finally, the optimized patch masks achieve the input resolution by reshaping from \\(\\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times 4^2}\\) to \\(\\mathbb{R}^{H \\times W}\\). Since" + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.497, + 0.898 + ], + "angle": 0, + "content": "923" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.123, + 0.462, + 0.181 + ], + "angle": 0, + "content": "
MethodsSingle-frameMulti-framesMulti-objects
[49, 63, 11] et al.✓
[4, 57, 56, 23] et al.✓✓
Fast SgMg (Ours)✓✓✓
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.078, + 0.182, + 0.468, + 0.207 + ], + "angle": 0, + "content": "Table 1. Comparing different methods for their ability to segment single or multiple frames or multiple objects simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.226, + 0.469, + 0.267 + ], + "angle": 0, + "content": "MSO does not include heavy computations, the segment-and-optimize pipeline makes our approach perform better with efficient inference time." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.276, + 0.274, + 0.291 + ], + "angle": 0, + "content": "3.5. Multi-Object R-VOS" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.469, + 0.408 + ], + "angle": 0, + "content": "Existing R-VOS methods perform single-frame (frameworkwise) segmentation [49, 63, 11] or multi-frame (clip-wise) segmentation [4, 57, 56] for an individual referred object at a time. However, to the best of our knowledge, no existing work explores the simultaneous segmentation of multiple referred objects in video using a common GPU, which is important for real-world scenarios. To fill this gap, we present a new paradigm called multi-object R-VOS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.469, + 0.518 + ], + "angle": 0, + "content": "The key to multi-object R-VOS is designing a network that shares computationally intensive features for multiple objects, and enables different instance features to be decoupled before segmentation. To achieve this, we extend \\(\\mathrm{SgMg}\\) for multi-object R-VOS by introducing multi-instance fusion and decoupling. As shown in Table 1, our method, dubbed Fast \\(\\mathrm{SgMg}\\), can simultaneously segment multiple objects (in multiple frames) using a single 24GB GPU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.469, + 0.697 + ], + "angle": 0, + "content": "Fast \\(\\mathrm{SgMg}\\) shares visual features as well as vision-language features for all referred objects to make the network efficient, and decouples the shared features to make them instance-specific before the segmentation stage. Firstly, visual features \\((\\mathcal{F}_v)\\) and language features \\((\\mathcal{F}_w\\) and \\(\\mathcal{F}_s)\\) are extracted. Next, we associate \\(\\mathcal{F}_v\\) and \\(\\mathcal{F}_w\\) using multi-instance fusion rather than the previous SCF. Multi-instance fusion is built on the foundation of SCF, which is depicted in Fig. 3. The difference is that multi-instance fusion includes semantic fusion, which performs an elementwise add operation, after cross-attention to merge vision-language features of different expressions. The features after semantic fusion perform Hadamard product with \\(\\mathcal{F}_v\\) to generate the vision-language features \\(\\mathcal{F}_{vl}\\) for all objects:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.703, + 0.469, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S F} \\left(\\mathcal {F} _ {w}, \\mathcal {F} _ {v}\\right) = \\sum_ {i = 1} ^ {N} \\operatorname {A t t} \\left(\\mathcal {F} _ {w} ^ {i}, \\mathcal {F} _ {v}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.759, + 0.469, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M I F} \\left(\\mathcal {F} _ {w}, \\mathcal {F} _ {v}\\right) = \\operatorname {S A} \\left(\\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right) \\otimes \\operatorname {S F} \\left(\\mathcal {F} _ {w}, \\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right)\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.469, + 0.867 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes Hadamard product and \\(N\\) denotes the number of expressions. After vision-language fusion, we encode \\(\\mathcal{F}_{vl}\\) using the transformer encoder to enrich its semantic information, and plug multi-instance decoupling to decouple features for each instance. Multi-instance decoupling employs \\(\\mathcal{F}_w\\) and cross-attention to decouple \\(\\mathcal{F}_{vl}\\) to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.127, + 0.892, + 0.196 + ], + "angle": 0, + "content": "predict instance embeddings \\(\\mathcal{E}\\) for different referred objects. These embeddings are then projected to CPKs to predict the patch masks. Thus, FAST SgMg shares features, which account for most of the computational overhead, for different expressions, making it efficient for referring segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.205, + 0.833, + 0.219 + ], + "angle": 0, + "content": "3.6. Instance Matching and Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.227, + 0.892, + 0.336 + ], + "angle": 0, + "content": "Following [4, 57], we perform instance matching with \\( N = 5 \\) learnable instance queries to improve fault tolerance. These queries are projected to CPKs to predict \\( N \\) potential patch masks \\( \\mathcal{M}_P \\) for each expression. The Hungarian algorithm [22] is then adopted to select the best match based on the matching loss for training. During inference, we directly employ the predicted confidence scores \\( S \\) to measure the instance queries and select the results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.337, + 0.892, + 0.413 + ], + "angle": 0, + "content": "We adopt the same training losses and weights as used in [57, 72] for a fair comparison. Specifically, we use Dice loss [26] and Focal loss [31] for patch mask \\(\\mathcal{M}_P\\) and optimized mask \\(\\mathcal{M}_O\\), Focal loss [31] for confidence scores \\(S\\), and L1 and GIoU [48] loss for bounding boxes \\(\\mathcal{B}\\). The final training loss functions are:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.423, + 0.891, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a i n}} = \\lambda_ {\\mathcal {M} _ {P}} \\mathcal {L} _ {\\mathcal {M} _ {P}} + \\lambda_ {\\mathcal {M} _ {O}} \\mathcal {L} _ {\\mathcal {M} _ {O}} + \\lambda_ {\\mathcal {B}} \\mathcal {L} _ {\\mathcal {B}} + \\lambda_ {\\mathcal {S}} \\mathcal {L} _ {\\mathcal {S}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.447, + 0.882, + 0.46 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) and \\(\\lambda\\) are the loss term and weight, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.473, + 0.632, + 0.489 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.698, + 0.51 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.669 + ], + "angle": 0, + "content": "Datasets. We evaluate \\(\\mathrm{SgMg}\\) on four video benchmarks: Ref-YouTube-VOS [49], Ref-DAVIS17 [21], A2D-Sentences [17], and JHMDB-Sentences [20]. RefYouTube-VOS is currently the largest dataset for R-VOS, containing 3,978 videos with about \\(13\\mathrm{K}\\) expressions. RefDAVIS17 is an extension of DAVIS17 [45] by including the language expressions of different objects and contains 90 videos. A2D-Sentences is a general actor and action segmentation dataset with over 3.7K videos and 6.6K action descriptions. JHMDB-Sentences includes 928 videos and 928 descriptions covering 21 different action classes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.67, + 0.892, + 0.753 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt the standard metrics to evaluate our models: region similarity \\(\\mathcal{I}\\) (average IoU), contour accuracy \\(\\mathcal{F}\\) (average boundary similarity), and their mean value \\(\\mathcal{I} \\& \\mathcal{F}\\). All results are evaluated using the official code or server. On A2D-Sentences and JHMDB-Sentences, we adopt mAP, overall IoU, and mean IoU for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.716, + 0.777 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.784, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Following [4, 6, 57], we train our models on the training set of Ref-YouTube-VOS, and directly evaluate them on the validation split of Ref-YouTube-VOS and Ref-DAVIS17 without any additional techniques, e.g., model ensemble, joint training, and mask propagation, since they are not the focus of this paper. Additionally, we present results for our" + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "924" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.141, + 0.124, + 0.829, + 0.302 + ], + "angle": 0, + "content": "
MethodYearBackboneRef-YouTube-VOSRef-DAVIS17
J&FJFFPSJ&FJF
CMSA [63]2019ResNet-5036.434.838.1-40.236.943.5
URVOS [49]2020ResNet-5047.245.349.2-51.547.356.0
CMPC-V [32]2021I3D47.545.649.3----
PMINet [11]2021ResNeSt-10153.051.554.5----
YOFO [23]2022ResNet-5048.647.549.71053.348.857.8
LBDT [10]2022ResNet-5049.448.250.6-54.3--
MLRL [56]2022ResNet-5049.748.451.0-52.850.055.4
MTTR [4]2022Video-Swin-T55.354.056.6----
MANet [6]2022Video-Swin-T55.654.856.5----
ReferFormer [57]2022Video-Swin-T56.054.857.350---
SgMg (Ours)2023Video-Swin-T58.957.760.06556.753.360.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.305, + 0.461, + 0.317 + ], + "angle": 0, + "content": "Pre-training with RefCOCO/+/g & larger backbone" + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.319, + 0.829, + 0.376 + ], + "angle": 0, + "content": "
ReferFormer [57]2022Video-Swin-T59.458.060.95059.656.562.7
SgMg (Ours)2023Video-Swin-T62.060.463.56561.959.064.8
ReferFormer [57]2022Video-Swin-B62.961.364.63361.158.164.1
SgMg (Ours)2023Video-Swin-B65.763.967.44063.360.666.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.377, + 0.85, + 0.389 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison to state-of-the-art methods on the validation split of Ref-YouTube-VOS and Ref-DAVIS17." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.4, + 0.832, + 0.566 + ], + "angle": 0, + "content": "
MethodBackboneA2D-SentencesJHMDB-Sentences
mAPOverall IoUMean IoUmAPOverall IoUMean IoU
Hu et al. [19]VGG-1613.247.435.017.854.652.8
Gavrilyuk et al. [17]I3D19.853.642.123.354.154.2
ACAN [54]I3D27.460.149.028.957.658.4
CMPC-V [32]I3D40.465.357.334.261.661.7
ClawCraneNet [29]ResNet-50/101-63.159.9-64.465.6
MTTR [4]Video-Swin-T46.172.064.039.270.169.8
ReferFormer [57]Video-Swin-T52.877.669.642.271.971.0
SgMg (Ours)Video-Swin-T56.178.070.444.472.871.7
ReferFormer [57]Video-Swin-B55.078.670.343.773.071.8
SgMg (Ours)Video-Swin-B58.579.972.045.073.772.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.152, + 0.567, + 0.816, + 0.578 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison to state-of-the-art R-VOS methods on A2D-Sentences and JHMDB-Sentences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.47, + 0.712 + ], + "angle": 0, + "content": "models first pre-trained on RefCOCO+/g [39, 65] and then fine-tuned on Ref-YouTube-VOS. Similar to [57, 72], we set the coefficients for different losses \\(\\lambda_{dice}\\), \\(\\lambda_{focal}\\), \\(\\lambda_{L1}\\), \\(\\lambda_{giou}\\) to 5, 2, 5, 2, respectively. The models are trained using 2 RTX 3090 GPUs with 5 frames per clip for 9 epochs. All frames are resized to have the longest side of 640 pixels. Further implementation details are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.721, + 0.272, + 0.735 + ], + "angle": 0, + "content": "4.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Ref-YouTube-VOS and Ref-DAVIS17. We compare \\(\\mathrm{SgMg}\\) with recently published works in Table 2. Our approach surpasses present solutions on the two datasets across all metrics. On Ref-YouTube-VOS, \\(\\mathrm{SgMg}\\) with the Video Swin Tiny backbone achieves \\(58.9\\mathcal{J}\\&\\mathcal{F}\\) at 65 FPS, which is \\(2.9\\%\\) higher and \\(1.3\\times\\) faster than the previous state-of-the-art ReferFormer [57]. Our approach runs faster due to the use of the segment-and-optimize pipeline, which avoids the need for heavy feature decoders. When pre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.892, + 0.685 + ], + "angle": 0, + "content": "training with RefCOCO/+/g and using a larger backbone, i.e., Video Swin Base, the performance of \\(\\mathrm{SgMg}\\) further boosts to 65.7 \\(\\mathcal{J}\\&\\mathcal{F}\\), consistently leading all other solutions by more than \\(2.8\\%\\). On Ref-DAVIS17, \\(\\mathrm{SgMg}\\) achieves 63.3 \\(\\mathcal{J}\\&\\mathcal{F}\\), outperforming state-of-the-art by \\(2.2\\%\\) and demonstrating the generality of our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.81 + ], + "angle": 0, + "content": "A2D-Sentences and JHMDB-Sentences. We further evaluate \\(\\mathrm{SgMg}\\) on A2D-Sentences and JHMDB-Sentences in Table 3. Following [57], the models are first pre-trained on RefCOCO+/g and then fine-tuned on A2D-Sentences. JHMDB-Sentences is used only for evaluation. As shown in Table 3, \\(\\mathrm{SgMg}\\) achieves superior performance compared to other state-of-the-art R-VOS methods and surpasses the nearest competitor Referformer [57] by \\(3.5 / 1.3\\%\\) mAP on A2D-Sentences and JHMDB-Sentences, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Multi-object R-VOS. We extend \\(\\mathrm{SgMg}\\) to perform multi-object R-VOS, which is more practical and efficient for deployment. Fast \\(\\mathrm{SgMg}\\) is trained on Ref-YouTube-VOS without pre-training or postprocessing techniques. We" + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "925" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.124, + 0.472, + 0.187 + ], + "angle": 0, + "content": "
MethodRef-DAVIS17Ref-YouTube-VOS
J&FJFJ&FJFFPS
ReferFormer [57]54.551.058.056.054.857.350
Fast SgMg (Ours)54.251.157.354.253.155.3185
" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.187, + 0.471, + 0.212 + ], + "angle": 0, + "content": "Table 4. Evaluation of Fast SgMg on Ref-DAVIS17 and RefYouTube-VOS. Video-Swin-T is adopted as the backbone." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.219, + 0.445, + 0.339 + ], + "angle": 0, + "content": "
ComponentsPerformance
CPKMSOSCFJ&FJFFPS
54.452.756.270
✓55.854.557.170
✓57.756.359.169
✓55.854.357.466
✓✓57.956.759.169
✓✓✓58.957.760.065
" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.34, + 0.466, + 0.352 + ], + "angle": 0, + "content": "Table 5. Ablation of different components on Ref-YouTube-VOS." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.367, + 0.469, + 0.422 + ], + "angle": 0, + "content": "benchmark Fast SgMg on Ref-YouTube-VOS and RefDAVIS17 using the commonly used Video Swin Tiny, and compare the results with the state-of-the-art R-VOS method, which performs single-object segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.469, + 0.561 + ], + "angle": 0, + "content": "As shown in Table 4, Fast \\(\\mathrm{SgMg}\\) achieves reasonable performance and runs about \\(3.7\\times\\) faster (185 vs 50 FPS) compared to ReferFormer [57]. It should be noted that each object in the above datasets contains multiple expressions. On Ref-DAVIS17, we group expressions to have only one expression per object within each group and segment all expressions in each group simultaneously since the object identity is given. On Ref-YouTube-VOS, all expressions in a video are segmented simultaneously due to the lack of object identity, making it more challenging." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.569, + 0.433, + 0.584 + ], + "angle": 0, + "content": "4.4. Ablation Study for Different Components" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.469, + 0.632 + ], + "angle": 0, + "content": "We conduct ablation experiments to evaluate the effectiveness of different components in \\(\\mathrm{SgMg}\\). The components are added to the baseline model step-by-step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.469, + 0.701 + ], + "angle": 0, + "content": "Conditional Patch Kernel. As shown in Table 5, CPK boosts the performance by \\(1.4\\%\\) compared with the recent instance-aware conditional kernels [57]. The sequential labels of each token predicted by CPK contain more fine-grained information, making the prediction more accurate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Multi-granularity Segmentation Optimizer. We devise MSO to optimize the predicted patch masks. As shown in Table 5, MSO improves the performance by \\(3.3\\%\\), indicating the importance of fine-grained visual details in R-VOS. Spectrum-guided Cross-modal Fusion. We present SCF to perform global interactions by operating in the spectral domain. In Table 5, using SCF to replace the traditional cross-attention in [57, 52] improves the \\(\\mathcal{J}\\&\\mathcal{F}\\) by \\(1.4\\%\\). We consider SCF extracts important low-frequency features and facilitates multimodal understanding globally, which is suitable for R-VOS since locating referred objects requires understanding the global context and token relations." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.124, + 0.892, + 0.18 + ], + "angle": 0, + "content": "
SettingsDriftPipelineJ&FFPS
Baseline + Decoder✓decode-and-segment56.050
Baseline + Decoder + MSO✓decode-and-segment56.449
Baseline + MSO×segment-and-optimize57.769
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.181, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Table 6. Feature drift analysis using ReferFormer [57] (Baseline + Decoder) and \\(\\mathrm{SgMg}\\) w/o CPK & SCF (Baseline + MSO). Significant improvement is achieved by addressing the drift issue (last row). Adding MSO on top of ReferFormer to recover visual details (for a second time) still performs worse than our basic pipeline." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.256, + 0.879, + 0.357 + ], + "angle": 0, + "content": "
MethodRefCOCORefCOCO+RefCOCOg
MaIL [28]70.162.262.5
CRIS [55]70.562.359.9
RefTR [25]70.6--
LAVT [62]72.762.161.2
VLT [9]73.063.563.5
SgMg (Ours)76.366.470.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.358, + 0.891, + 0.384 + ], + "angle": 0, + "content": "Table 7. Quantitative evaluation on the validation split of Ref-COCO/+/g. Overall IoU is adopted as the evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.407, + 0.785, + 0.421 + ], + "angle": 0, + "content": "4.5. Ablation Study for Feature Drift" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.553 + ], + "angle": 0, + "content": "We conduct ablation study in Table 6 to demonstrate the feature drift problem. Our segment-and-optimize pipeline addresses the adverse drift effect discussed in Section 3.1 to significantly outperform ReferFormer [57] by \\(1.7\\%\\) points and runs \\(1.4 \\times\\) faster. Furthermore, adding MSO on top of ReferFormer still performs worse due to the negative drift impact caused by the decode-and-segment pipeline. These results demonstrate the efficacy of our proposed segment-and-optimize pipeline." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.564, + 0.833, + 0.579 + ], + "angle": 0, + "content": "4.6. Referring Image Segmentation Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.891, + 0.697 + ], + "angle": 0, + "content": "We apply \\(\\mathrm{SgMg}\\) to referring image (expression) segmentation without any architectural modifications, and compare against the current state-of-the-art methods on RefCOCO \\(+ / g\\) [39, 65]. A single \\(\\mathrm{SgMg}\\) model is trained on RefCOCO \\(+ / g\\) without large-scale pre-training. As shown in Table 7, \\(\\mathrm{SgMg}\\) achieves advanced performance on all three benchmarks. These results demonstrate the efficacy of \\(\\mathrm{SgMg}\\) in referring image segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.707, + 0.891, + 0.722 + ], + "angle": 0, + "content": "4.7. Inference Time Analysis of Multi-Object RVOS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.892, + 0.867 + ], + "angle": 0, + "content": "We analyze the efficiency of the proposed multi-object R-VOS paradigm by comparing the FPS of Fast \\(\\mathrm{SgMg}\\) and \\(\\mathrm{SgMg}\\) on videos with different numbers of expressions. As illustrated in Fig. 6, Fast \\(\\mathrm{SgMg}\\) performs about \\(2\\times\\) faster than \\(\\mathrm{SgMg}\\) when there are two expressions per video on average. As the number of expressions increases, Fast \\(\\mathrm{SgMg}\\) achieves faster reasoning time per object per frame due to its utilization of the multi-object R-VOS paradigm. When there are ten expressions in each video, Fast \\(\\mathrm{SgMg}\\) performs at nearly 300 FPS, which is about \\(5\\times\\) faster than \\(\\mathrm{SgMg}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "926" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.125, + 0.388, + 0.138 + ], + "angle": 0, + "content": "Expression: a turtle to the bottom left." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.14, + 0.868, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.38, + 0.663, + 0.393 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of our method with others." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.409, + 0.466, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.55, + 0.47, + 0.577 + ], + "angle": 0, + "content": "Figure 6. Efficiency analysis of \\(\\mathrm{SgMg}\\) and Fast \\(\\mathrm{SgMg}\\) for videos with different numbers of expressions on Ref-YouTube-VOS." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.598, + 0.262, + 0.613 + ], + "angle": 0, + "content": "4.8. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.47, + 0.662 + ], + "angle": 0, + "content": "In Fig. 5, we show qualitative comparison with ReferFormer [57] and MTTR [4]. \\(\\mathrm{SgMg}\\) can handle different objects of the same category or with the same behavior." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.67, + 0.337, + 0.685 + ], + "angle": 0, + "content": "4.9. Feature Visualization of SCF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.471, + 0.762 + ], + "angle": 0, + "content": "In Fig. 7, we visualize the vision-language features extracted by our SCF in comparison to the cross-attention used in [57]. The features extracted by SCF exhibit superior grounding ability in locating target objects, resulting in better performance for \\(\\mathrm{SgMg}\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.774, + 0.197, + 0.789 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.797, + 0.471, + 0.868 + ], + "angle": 0, + "content": "We discovered the feature drift issue in current referring video object segmentation (R-VOS) methods, which negatively affects the segmentation kernels. We presented \\(\\mathrm{SgMg}\\), a novel segment-and-optimize approach for R-VOS that avoids the drift issue and optimizes masks with visual" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.409, + 0.892, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.617, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Figure 7. Visualization of the vision-language features extracted w/o and w/ our SCF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.772 + ], + "angle": 0, + "content": "details. We also provided a new perspective to encourage vision-language global interactions in the spectral domain with Spectrum-guided Cross-modal Fusion. Additionally, we proposed the multi-object R-VOS paradigm by extending \\(\\mathrm{SgMg}\\) with multi-instance fusion and decoupling. Finally, we evaluated our models on four video benchmarks and demonstrated that our approach achieves state-of-the-art performance on all four datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.893, + 0.86 + ], + "angle": 0, + "content": "Acknowledgment. This research was supported by the Australian Research Council Industrial Transformation Research Hub IH180100002. Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government." + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.887, + 0.498, + 0.898 + ], + "angle": 0, + "content": "927" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.174, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.47, + 0.223 + ], + "angle": 0, + "content": "[1] Nayyer Aafaq, Naveed Akhtar, Wei Liu, Syed Zulqarnain Gilani, and Ajmal Mian. Spatio-temporal dynamics and semantic attribute enriched visual encoding for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12487-12496, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.225, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[2] Miriam Bellver, Carles Ventura, Carina Silberer, Ioannis Kazakos, Jordi Torres, and Xavier Giro-i Nieto. A closer look at referring expressions for video object segmentation. Multimedia Tools and Applications, pages 1-20, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.277, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[3] Glenn D Bergland. A guided tour of the fast fourier transform. IEEE spectrum, 6(7):41-52, 1969. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.471, + 0.352 + ], + "angle": 0, + "content": "[4] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In CVPR, pages 4985-4995, 2022. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[5] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 221-230, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Weidong Chen, Dexiang Hong, Yuankai Qi, Zhenjun Han, Shuhui Wang, Laiyun Qing, Qingming Huang, and Guorong Li. Multi-attention network for compressed video referring object segmentation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4416–4425, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[7] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Proc. Adv. Neural Inf. Process. Syst. (NIPS), 34:11781-11794, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.47, + 0.583 + ], + "angle": 0, + "content": "[8] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.585, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[9] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[10] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.47, + 0.763 + ], + "angle": 0, + "content": "[11] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, page 7, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.765, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[12] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.867 + ], + "angle": 0, + "content": "[13] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.165 + ], + "angle": 0, + "content": "[14] Guang Feng, Lihe Zhang, Zhiwei Hu, and Hutchuan Lu. Deeply interleaved two-stream encoder for referring video segmentation. arXiv preprint arXiv:2203.15969, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.228 + ], + "angle": 0, + "content": "[15] Mingtao Feng, Haoran Hou, Liang Zhang, Yulan Guo, Hongshan Yu, Yaonan Wang, and Ajmal Mian. Exploring hierarchical spatial layout cues for 3d point cloud based scene graph prediction. IEEE Transactions on Multimedia, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.23, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[16] Zhiheng Fu, Siyu Hong, Mengyi Liu, Hamid Laga, Mohammed Bennamoun, Farid Boussaid, and Yulan Guo. Multi-stage information diffusion for joint depth and surface normal estimation. Pattern Recognition, 141:109660, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[17] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[18] Lorenzo Giambagli, Lorenzo Buffoni, Timoteo Carletti, Walter Nocentini, and Duccio Fanelli. Machine learning in spectral domain. Nature communications, 12(1):1-9, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.395, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[19] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[20] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.51, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[21] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Asian Conference on Computer Vision, pages 123-141. Springer, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[22] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.599, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[23] Dezhuang Li, Ruoqi Li, Lijun Wang, Yifan Wang, Jinqing Qi, Lu Zhang, Ting Liu, Qingquan Xu, and Huchuan Lu. You only infer once: Cross-modal meta-transfer for referring video object segmentation. In AAAI Conference on Artificial Intelligence, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.663, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[24] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.727, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[25] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in neural information processing systems, 34:19652-19664, 2021.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.778, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[26] Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[27] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Yan Lu, and Bhiksha Raj. R^2vos: Robust referring video object segmentation via relational multimodal cycle consistency. arXiv preprint arXiv:2207.01203, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "928" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[28] Zizhang Li, Mengmeng Wang, Jianbiao Mei, and Yong Liu. Mail: A unified mask-image-language trimodal network for referring image segmentation. arXiv preprint arXiv:2111.10747, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.18, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[29] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawcranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.469, + 0.282 + ], + "angle": 0, + "content": "[30] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.284, + 0.469, + 0.334 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.336, + 0.469, + 0.386 + ], + "angle": 0, + "content": "[32] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.388, + 0.469, + 0.45 + ], + "angle": 0, + "content": "[33] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[34] Yong Liu, Ran Yu, Jiahao Wang, Xinyuan Zhao, Yitong Wang, Yansong Tang, and Yujiu Yang. Global spectral filter memory network for video object segmentation. In European Conference on Computer Vision, pages 648-665. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[35] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[36] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[37] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10437–10446, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[38] Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsupervised video object segmentation with co-attention siamese networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3623–3632, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.469, + 0.827 + ], + "angle": 0, + "content": "[39] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.828, + 0.469, + 0.867 + ], + "angle": 0, + "content": "[40] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.127, + 0.892, + 0.165 + ], + "angle": 0, + "content": "localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[41] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Region aware video object segmentation with deep motion modeling. arXiv preprint arXiv:2207.10258, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.269 + ], + "angle": 0, + "content": "[42] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Self-supervised video object segmentation by motion-aware mask propagation. In Proceedings of the IEEE International Conference on Multimedia and Expo, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.271, + 0.892, + 0.321 + ], + "angle": 0, + "content": "[43] Bo Miao, Liguang Zhou, Ajmal Mian, Tin Lun Lam, and Yangsheng Xu. Object-to-scene: Learning to transfer object knowledge to indoor scene recognition. In Proc. IEEE Int. Conf. Intell. Robots Syst. (IROS). IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.323, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[44] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proc. Int. Conf. Comput. Vis. (ICCV), pages 9226–9235, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[45] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[47] Sucheng Ren, Wenxi Liu, Yongtuo Liu, Haoxin Chen, Guoqiang Han, and Shengfeng He. Reciprocal transformations for unsupervised video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15455-15464, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.569, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[48] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[49] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223. Springer, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.698, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[50] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In European conference on computer vision, pages 282-298. Springer, 2020. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.84 + ], + "angle": 0, + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.841, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[53] Paul Voiglaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "929" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.127, + 0.47, + 0.177 + ], + "angle": 0, + "content": "end-to-end embedding learning for video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9481-9490, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.18, + 0.47, + 0.243 + ], + "angle": 0, + "content": "[54] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.245, + 0.471, + 0.307 + ], + "angle": 0, + "content": "[55] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.47, + 0.373 + ], + "angle": 0, + "content": "[56] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.374, + 0.469, + 0.423 + ], + "angle": 0, + "content": "[57] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[58] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. arXiv preprint arXiv:2308.02874, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[59] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.53, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[60] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In International Conference on Neural Information Processing, pages 264-274. Springer, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.582, + 0.469, + 0.632 + ], + "angle": 0, + "content": "[61] Zhao Yang, Yansong Tang, Luca Bertinetto, Hengshuang Zhao, and Philip HS Torr. Hierarchical interaction network for video object segmentation from referring expressions. In BMVC, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[62] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Ladv: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[63] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[64] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[65] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.471, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.165 + ], + "angle": 0, + "content": "[66] Z. Yang and Y, Wei and Y. Yang. Associating objects with transformers for video object segmentation. Proc. Adv. Neural Inf. Process. Syst. (NIPS), 34:2491-2502, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[67] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.244, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[68] Shuai Zhao, Xiaohan Wang, Linchao Zhu, and Yi Yang. Test-time adaptation with clip reward for zero-shot generalization in vision-language models. arXiv preprint arXiv:2305.18010, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[69] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. arXiv preprint arXiv:2205.00823, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[70] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[71] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13041-13049, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.464, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[72] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "930" + } + ] +] \ No newline at end of file diff --git a/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/dc8d764a-9214-4358-ad24-c2007b58a08c_origin.pdf b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/dc8d764a-9214-4358-ad24-c2007b58a08c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..35d6b4c180aa34144196870ec8ab141ed80cd97a --- /dev/null +++ b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/dc8d764a-9214-4358-ad24-c2007b58a08c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cee35a8745b3a69a6ab2e8248c297c34a1c23f4e1799604a9d7d8355a8ed77f +size 1864430 diff --git a/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/full.md b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4edd081f492eddf2cb094f1fb87646df05d8a9ec --- /dev/null +++ b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/full.md @@ -0,0 +1,331 @@ +# Spectrum-guided Multi-granularity Referring Video Object Segmentation + +Bo Miao $^{1}$ , Mohammed Bennamoun $^{1}$ , Yongsheng Gao $^{2}$ , Ajmal Mian $^{1}$ $^{1}$ The University of Western Australia + $^{2}$ Griffith University + +# Abstract + +Current referring video object segmentation (R-VOS) techniques extract conditional kernels from encoded (low-resolution) vision-language features to segment the decoded high-resolution features. We discovered that this causes significant feature drift, which the segmentation kernels struggle to perceive during the forward computation. This negatively affects the ability of segmentation kernels. To address the drift problem, we propose a Spectrum-guided Multigranularity (SgMg) approach, which performs direct segmentation on the encoded features and employs visual details to further optimize the masks. In addition, we propose Spectrum-guided Cross-modal Fusion (SCF) to perform intra-frame global interactions in the spectral domain for effective multimodal representation. Finally, we extend SgMg to perform multi-object R-VOS, a new paradigm that enables simultaneous segmentation of multiple referred objects in a video. This not only makes R-VOS faster, but also more practical. Extensive experiments show that SgMg achieves state-of-the-art performance on four video benchmark datasets, outperforming the nearest competitor by $2.8\%$ points on Ref-YouTube-VOS. Our extended SgMg enables multi-object R-VOS, runs about $3\times$ faster while maintaining satisfactory performance. Code is available at https://github.com/bo-miao/SgMg. + +# 1. Introduction + +Referring video object segmentation (R-VOS) aims at segmenting objects in a video, referred to by linguistic descriptions. R-VOS is an emerging task for multimodal reasoning and promotes a wide range of applications, including language-guided video editing and human-machine interaction. Different from conventional semi-supervised video object segmentation [44, 7, 41], where the mask annotation for the first frame is provided for reference, R-VOS is more challenging due to the need for cross-modal understanding between vision and free-form language expressions. + +Early R-VOS techniques [2, 21, 61] perform feature encoding, cross-modal interaction, and language grounding + +![](images/89d6e208c163b4cbcd2b4ac342aeb8c5568808f60e50da5bd790db27323de29d.jpg) +(b) SgMg (Ours): segment-and-optimize w/ encoded features $F_{vl}$ + +![](images/fc603f36ec204257f55403249444807ca1ada13606a4021792a967903df7776c.jpg) +Conditional Segmentation Fusion: Cross-modal Fusion $F_{\mu l}$ : Encoded Vision-Language Features $F_{\nu l}^{d}$ : Decoded Features Q: Instance Query $K_{c}$ : Conditional Kernel $K_{cp}$ : Conditional Patch Kernel $M_{Q}$ : Predicted Mask +Figure 1. (a) Previous methods [4, 57] apply segmentation kernels $\mathcal{K}_c$ [50], extracted from encoded features $\mathcal{F}_{vl}$ , to segment the decoded high-resolution features $\mathcal{F}_{vl}^d$ . (b) We use segmentation kernels $\mathcal{K}_{cp}$ , extracted from encoded features $\mathcal{F}_{vl}$ , to segment the encoded features $\mathcal{F}_{vl}$ directly, and propose multi-granularity optimization to recover visual details and produce fine-grained masks. + +using convolutional neural networks (CNNs). However, the limited ability of CNNs to capture long-range dependencies and handle free-form features constrains the model performance. With the advancement of attention mechanisms [52, 43, 15, 16, 58], recent methods achieved significant improvement on R-VOS using cross-attention [49, 23, 27] for multimodal understanding and transformers [6, 56] for spatio-temporal representation. Based on transformers, conditional kernel [50] is then introduced to separate foreground from semantic features given its high adaptability to different instances [4, 57]. As illustrated in Fig. 1(a), these methods attend to encoded vision-language features $\mathcal{F}_{vl}$ using instance queries $Q$ to predict conditional kernels $\kappa_{c}$ , and employ $\kappa_{c}$ as the segmentation head to segment decoded features $\mathcal{F}_{vl}^{d}$ . Despite the promising performance, + +this paradigm still has some limitations. Firstly, as shown in the t-SNE [51] visualization in Fig. 1(a), although the nonlinear decoding process introduces visual details, this is accompanied by a significant feature drift, which increases the difficulty of segmentation since $\mathcal{K}_c$ is predicted before feature decoding. Secondly, bilinear upsampling of the predicted masks $\mathcal{M}_Q$ to increase resolution impedes the segmentation performance. Thirdly, these methods only support single expression-based segmentation, making R-VOS inefficient when multiple referred objects exist in a video. + +In this work, we propose a Spectrum-guided Multi-granularity (SgMg) approach that follows a segment-and-optimize pipeline to address the above problems. As depicted in Fig. 1(b), SgMg introduces Conditional Patch Kernel (CPK) $\mathcal{K}_{cp}$ to directly segment its fully perceived encoded features $\mathcal{F}_{vl}$ , avoiding the feature drift and its adverse effects. The segmentation is then refined using our proposed Multi-granularity Segmentation Optimizer (MSO), which employs low-level visual details to produce full-resolution masks. Within the SgMg framework, we further develop Spectrum-guided Cross-modal Fusion (SCF) that performs intra-frame global interactions in the spectral domain to facilitate multimodal understanding. Finally, we introduce a new paradigm called multi-object R-VOS to simultaneously segment multiple referred objects in a video. To achieve this, we extend SgMg by devising multi-instance fusion and decoupling. Our main contributions are summarized as follows: + +- We explain how existing R-VOS methods suffer from the feature drift problem. To address this problem, we propose $\mathrm{SgMg}$ that follows a segment-and-optimize pipeline and achieves top-ranked overall performance on multiple benchmark datasets. +- We propose Spectrum-guided Cross-modal Fusion to encourage intra-frame global interactions in the spectral domain. +- We extend $\mathrm{SgMg}$ to perform multi-object R-VOS, a new paradigm that enables simultaneous segmentation of multiple referred objects in a video. Our multi-object variant is more practical and runs $3\times$ faster. + +We conduct extensive experiments on multiple benchmark datasets, including Ref-YouTube-VOS [49], RefDAVIS17 [21], A2D-Sentences [17], and JHMDB-Sentences [20], and achieve state-of-the-art performance on all four. On the largest validation set Ref-YouTube-VOS, $\mathrm{SgMg}$ achieves $65.7\mathcal{J}\& \mathcal{F}$ which is $2.8\%$ points higher than that of the closest competitor ReferFormer [57]. On the A2D-Sentences, $\mathrm{SgMg}$ achieves $58.5\mathrm{mAP}$ which is $3.5\%$ points higher than that of ReferFormer. + +# 2. Related Works + +Video Object Segmentation techniques fall into two categories: unsupervised and semi-supervised. Unsupervised approaches segment the most salient instances in each video without user interactions [38, 47]. They often employ two-stream networks to fuse motion and appearance cues for segmentation. Semi-supervised approaches track the given first frame object mask by performing online learning [5] or spatial-temporal association [44, 7, 66, 42, 53]. Unlike conventional semi-supervised video object segmentation, R-VOS takes a free-form linguistic expression as guidance to detect and segment referred objects in videos. + +Referring Video Object Segmentation. R-VOS methods mainly use deep neural networks with vision-and-language interaction to empower visual features with corresponding linguistic information for pixel-level segmentation. For example, [49] employs a unified R-VOS framework that performs iterative segmentation guided by both language and temporal features. [32, 29] adopt progressive segmentation by perceiving potential objects and discriminating the best match. [70] fuses visual and motion features for segmentation under the guidance of linguistic cues. [30] models object relations to form tracklets and performs tracklet-language grounding. To enhance multi-modal interactions, [61, 14, 11, 10] perform hierarchical vision-language fusion on multiple feature layers. + +Despite their promising performance, the complex multistage pipelines and use of multiple networks make R-VOS burdensome. To address these problems, MTTR [4] proposes an end-to-end transformer-based network with conditional kernels [50] to segment target objects. ReferFormer [57] further introduces language-guided instance queries to predict instance-aware conditional kernels and an auxiliary detection task to aid localization. These methods follow a decode-and-segment pipeline, which adopts conditional kernels to segment decoded high-resolution features to achieve promising performance. However, the nonlinear decoding process leads to significant feature drift that negatively affects the conditional kernels. In contrast to previous works, our approach follows a segment-and-optimize pipeline to avoid the adverse drift effects and to predict full-resolution masks in an efficient manner. + +Vision and Language Representation Learning aims to learn vision-language semantics and alignment for multimodal reasoning tasks. It has achieved significant success in various tasks [37, 67, 68], including video question answering [71], video captioning [1], video-text retrieval [12, 69], zero-shot classification [46], referring image/video segmentation [49], etc. Some approaches [46, 24] rely on contrastive pre-training using large-scale datasets to project different modalities into unified embedding space. Others [36, 13] develop cross-modal interaction layers for multimodal feature fusion and understanding. Recent deep + +![](images/a094d27b35f2ca47ca5c6634e7b32f40cd1830c5bc4be6a1bb9f6c78fe79b680.jpg) +Figure 2. The overall framework of $\mathrm{SgMg}$ . Taking a video sequence $\mathcal{V} = \{I_i\}_{i=1}^T$ and a language expression $\mathcal{L} = \{S_i\}_{i=1}^N$ as input, $\mathrm{SgMg}$ predicts the masks of referred object $\mathcal{O}_{\mathcal{L}}$ in each frame. SCF projects visual features $\mathcal{F}_v$ to vision-language features $\mathcal{F}_{vl}$ , instance-aware CPK predicts patch masks by segmenting encoded $\mathcal{F}_{vl}$ , and MSO optimizes patch masks to get fine-grained results. + +learning methods in spectral domain [18, 40, 8, 34] have raised widespread awareness because of their ability to perform global interactions. We take inspiration from these spectral-based methods and employ spectrum guidance in the field of vision-language representation to encourage multimodal global interactions. + +# 3. $\mathrm{SgMg}$ : Spectrum-guided Multi-granularity Referring Video Object Segmentation + +Given a video sequence $\mathcal{V} = \{I_i\}_{i=1}^T$ with $T$ frames and a language query $\mathcal{L} = \{S_i\}_{i=1}^N$ with $N$ words. The goal of R-VOS is to segment the referred object $\mathcal{O}_{\mathcal{L}}$ in $\mathcal{V}$ at pixel-level. To this end, we introduce a new approach termed $\mathrm{SgMg}$ . Different from previous R-VOS methods [4, 57], our approach follows a segment-and-optimize pipeline. + +An overview of $\mathrm{SgMg}$ is shown in Fig. 2. Video Swin Transformer [35] is adopted to extract visual feature $\mathcal{F}_v$ and RoBERTa [33] is adopted to extract sentence $\mathcal{F}_s$ and word $\mathcal{F}_w$ features. The channel dimension of all features is projected to 256. Spectrum-guided Cross-modal Fusion (SCF) cross attends $\mathcal{F}_v$ with $\mathcal{F}_w$ to compute vision-language features $\mathcal{F}_{vl}$ . Deformable Transformer [72] encoder is used to encode $\mathcal{F}_{vl}$ and the decoder associates instance queries created based on $\mathcal{F}_s$ to predict instance embeddings and the corresponding Conditional Patch Kernels (CPKs). Finally, the CPKs are employed to segment $\mathcal{F}_{vl}$ and predict patch masks that are further optimized with visual details through Multi-granularity Segmentation Optimizer (MSO). The choice of the encoder and transformer follows previous works to avoid distractions [4, 57]. + +# 3.1. Feature Drift Analysis + +Existing R-VOS methods [57, 4] follow a decode-and-segment pipeline where conditional kernels $\mathcal{K}_c$ [50] are extracted from encoded features $\mathcal{F}_{vl}$ and used to segment the decoded features $\mathcal{F}_{vl}^d$ . However, the decoding process leads to feature drift, which is evident in the t-SNE visualization depicted in Fig. 1(a). This drift is difficult for the kernels $\mathcal{K}_c$ to perceive during the forward computation since $\mathcal{K}_c$ is predicted before the feature decoding. Therefore, we argue that even though the feature decoding enhances visual details, it also causes the drift problem that negatively affects the segmentation kernels. This makes the existing decode-and-segment pipeline sub-optimal. + +To overcome the adverse effects of feature drift while recovering visual details, we present $\mathrm{SgMg}$ , a novel approach that follows a segment-and-optimize pipeline. In a nutshell, $\mathrm{SgMg}$ performs Spectrum-guided Cross-modal Fusion to compute $\mathcal{F}_{vl}$ , leverages Conditional Patch Kernels to segment encoded features $\mathcal{F}_{vl}$ to avoid the drift effects, and recovers visual details with Multi-granularity Segmentation Optimizer to generate fine-grained masks. + +# 3.2. Spectrum-guided Cross-modal Fusion + +The two-dimensional discrete Fourier transform converts spatial data into the spectral domain. Based on the spectral convolution theorem [3], point-wise update of signals in the spectral domain globally affects all inputs in the spatial domain, which gives the insight to design spectrum-based modules so as to efficiently facilitate global interactions, which is critical for multimodal understanding. In addition, Low-frequency components in the spectral domain usually + +(a) Spectrum-Guided Cross-Modal Fusion +(b) Spectrum Augmentation +![](images/2759da62b255c26bbafd9000234dd434c8a8d9e64545b5286f5ee30137e530ce.jpg) +Flatten: reshape $\mathcal{M}_P$ from $\mathbb{R}^{\frac{H}{i}\times \frac{W}{i}\times p^2}$ to $\mathbb{R}^{\frac{Hp}{i}\times \frac{Wp}{i}}$ for visualization. $\times 2$ : upsampling operation. + +![](images/e07e40cf9af570d6f85d2835a23d862802b7928425ca500af0c62a6251d34559.jpg) +Figure 3. Spectrum-guided Cross-modal Fusion. Imag.: Imaginary. Pre-spectrum augmentation and post-spectrum augmentation share an identical structure. + +correspond to the general semantic information according to previous theoretical studies [59, 60, 64]. + +Inspired by the above observations, we conjecture that low-frequency components can benefit higher dimensional semantic features and propose Spectrum-guided Cross-modal Fusion (SCF). As shown in Fig. 3, SCF performs pre-spectrum augmentation to enhance visual features before cross-modal fusion and post-spectrum augmentation to facilitate global vision-language interactions after the fusion process. Let $\mathcal{F} \in \mathbb{R}^{C \times H \times W}$ denotes the input features, the spectrum augmentation (SA) is computed as: + +$$ +\operatorname {S A} (\mathcal {F}, K) = \mathcal {F} + \Theta_ {I F F T} (\operatorname {C o n v} (\sigma (K, \mathcal {F}) \odot \Theta_ {F F T} (\mathcal {F}))) \tag {1} +$$ + +where $\odot$ denotes low-pass filtering with adaptive Gaussian smoothed filters $\sigma(K, \mathcal{F})$ , which has the same spatial size as $\mathcal{F}$ , and $K$ is the bandwidth. To make $\sigma(K, \mathcal{F})$ input-aware, we create an initial 2D Gaussian map based on $K$ , and apply pooling and linear layers on $\mathcal{F}$ to predict a scale parameter to update the Gaussian map. Thanks to the spectral convolution theorem, the efficient point-wise spectral convolution globally updates $\mathcal{F}$ . We treat the spectral-operated features as residuals and add them to the original input features for enhancement. Overall, SCF, which takes visual features $\mathcal{F}_v$ and word-level text features $\mathcal{F}_w$ as input, is computed as: + +$$ +\operatorname {S C F} \left(\mathcal {F} _ {w}, \mathcal {F} _ {v}\right) = \operatorname {S A} \left(\operatorname {S A} \left(\mathcal {F} _ {v}\right) \otimes \operatorname {A t t} \left(\operatorname {S A} \left(\mathcal {F} _ {v}\right), \mathcal {F} _ {w}\right)\right) \tag {2} +$$ + +# 3.3. Conditional Patch Segmentation + +We devise Conditional Patch Kernel (CPK) as the segmentation head to predict patch masks from the encoded vision-language features $\mathcal{F}_{vl}$ that are fully perceived by CPK. Unlike previous works [4, 57], CPK predicts a sequence of labels for each token rather than a single label, efficiently improving segmentation resolution along the channel dimension. + +![](images/5352f0d2490d1e1f8f1404aadde77b7151607047992d91378dfdada585618560.jpg) +Figure 4. Multi-granularity Segmentation Optimizer, which predicts residual maps to optimize patch masks $\mathcal{M}_P$ progressively. + +Specifically, we first use sentence-level text features $\mathcal{F}_s$ and multiple learnable embeddings to generate instance queries $Q\in \mathbb{R}^{N\times C}$ . Next, $Q$ is projected into instance embeddings $\mathcal{E}\in \mathbb{R}^{N\times C}$ using the transformer decoder and $\mathcal{E}$ is leveraged to predict CPK for each instance query: + +$$ +\operatorname {C P K} (Q, \mathcal {F} _ {v l}) = \Theta (\operatorname {F C} (\operatorname {A t t} (Q, \mathcal {F} _ {v l}))) \tag {3} +$$ + +where $\Theta$ denotes the parameterization operation that reshapes CPK to form two point-wise convolutions with the output channel number of 16, which is similar to [50]. Since $Q$ changes dynamically according to different linguistic expressions, CPK becomes instance-aware and can separate objects of interest from $F_{vl}$ . Finally, we apply the parameterized CPK (dynamic point-wise convolutions) on $F_{vl}$ to predict patch masks $\mathcal{M}_P\in \mathbb{R}^{\frac{H}{i}\times \frac{W}{i}\times p^2}$ , where $\frac{H}{i}\times \frac{W}{i}$ denotes the spatial resolution of $F_{vl}$ and $p^2$ denotes the increased segmentation resolution on the channel dimension. + +During inference, we can reshape patch masks to $\mathcal{M}_P\in$ $\mathbb{R}\frac{Hp}{i}\times \frac{Wp}{i}$ to efficiently generate fine-grained segmentation from low-resolution $F_{vl}$ . The resolution of prediction will be consistent with the input when $p$ equals to $i$ . We found that this efficient CPK can achieve competitive performance compared to methods that use heavy decoders. + +# 3.4. Multi-granularity Segmentation Optimizer + +Segmenting encoded features $\mathcal{F}_{vl}$ with CPK avoids the detrimental drift effect on the segmentation head. However, visual details are required to produce accurate fine-grained masks. We propose Multi-granularity Segmentation Optimizer (MSO) to achieve this goal. + +An overview of MSO is shown in Fig. 4. It takes the predicted patch masks $\mathcal{M}_P$ as object priors and reuses visual features $\mathcal{F}_v$ with spatial strides of $\{4,8\}$ to gradually recover visual details and refine the priors. Specifically, MSO first concatenates $\mathcal{M}_P$ and $\mathcal{F}_v$ and projects them to low dimensional bases. Next, residual masks predicted by performing another convolution on these bases are used to correct $\mathcal{M}_P$ . Finally, the optimized patch masks achieve the input resolution by reshaping from $\mathbb{R}^{\frac{H}{4} \times \frac{W}{4} \times 4^2}$ to $\mathbb{R}^{H \times W}$ . Since + +
MethodsSingle-frameMulti-framesMulti-objects
[49, 63, 11] et al.✓
[4, 57, 56, 23] et al.✓✓
Fast SgMg (Ours)✓✓✓
+ +Table 1. Comparing different methods for their ability to segment single or multiple frames or multiple objects simultaneously. + +MSO does not include heavy computations, the segment-and-optimize pipeline makes our approach perform better with efficient inference time. + +# 3.5. Multi-Object R-VOS + +Existing R-VOS methods perform single-frame (frameworkwise) segmentation [49, 63, 11] or multi-frame (clip-wise) segmentation [4, 57, 56] for an individual referred object at a time. However, to the best of our knowledge, no existing work explores the simultaneous segmentation of multiple referred objects in video using a common GPU, which is important for real-world scenarios. To fill this gap, we present a new paradigm called multi-object R-VOS. + +The key to multi-object R-VOS is designing a network that shares computationally intensive features for multiple objects, and enables different instance features to be decoupled before segmentation. To achieve this, we extend $\mathrm{SgMg}$ for multi-object R-VOS by introducing multi-instance fusion and decoupling. As shown in Table 1, our method, dubbed Fast $\mathrm{SgMg}$ , can simultaneously segment multiple objects (in multiple frames) using a single 24GB GPU. + +Fast $\mathrm{SgMg}$ shares visual features as well as vision-language features for all referred objects to make the network efficient, and decouples the shared features to make them instance-specific before the segmentation stage. Firstly, visual features $(\mathcal{F}_v)$ and language features $(\mathcal{F}_w$ and $\mathcal{F}_s)$ are extracted. Next, we associate $\mathcal{F}_v$ and $\mathcal{F}_w$ using multi-instance fusion rather than the previous SCF. Multi-instance fusion is built on the foundation of SCF, which is depicted in Fig. 3. The difference is that multi-instance fusion includes semantic fusion, which performs an elementwise add operation, after cross-attention to merge vision-language features of different expressions. The features after semantic fusion perform Hadamard product with $\mathcal{F}_v$ to generate the vision-language features $\mathcal{F}_{vl}$ for all objects: + +$$ +\operatorname {S F} \left(\mathcal {F} _ {w}, \mathcal {F} _ {v}\right) = \sum_ {i = 1} ^ {N} \operatorname {A t t} \left(\mathcal {F} _ {w} ^ {i}, \mathcal {F} _ {v}\right) \tag {4} +$$ + +$$ +\operatorname {M I F} \left(\mathcal {F} _ {w}, \mathcal {F} _ {v}\right) = \operatorname {S A} \left(\operatorname {S A} \left(\mathcal {F} _ {v}\right) \otimes \operatorname {S F} \left(\mathcal {F} _ {w}, \operatorname {S A} \left(\mathcal {F} _ {v}\right)\right)\right) \tag {5} +$$ + +where $\otimes$ denotes Hadamard product and $N$ denotes the number of expressions. After vision-language fusion, we encode $\mathcal{F}_{vl}$ using the transformer encoder to enrich its semantic information, and plug multi-instance decoupling to decouple features for each instance. Multi-instance decoupling employs $\mathcal{F}_w$ and cross-attention to decouple $\mathcal{F}_{vl}$ to + +predict instance embeddings $\mathcal{E}$ for different referred objects. These embeddings are then projected to CPKs to predict the patch masks. Thus, FAST SgMg shares features, which account for most of the computational overhead, for different expressions, making it efficient for referring segmentation. + +# 3.6. Instance Matching and Loss Functions + +Following [4, 57], we perform instance matching with $N = 5$ learnable instance queries to improve fault tolerance. These queries are projected to CPKs to predict $N$ potential patch masks $\mathcal{M}_P$ for each expression. The Hungarian algorithm [22] is then adopted to select the best match based on the matching loss for training. During inference, we directly employ the predicted confidence scores $S$ to measure the instance queries and select the results. + +We adopt the same training losses and weights as used in [57, 72] for a fair comparison. Specifically, we use Dice loss [26] and Focal loss [31] for patch mask $\mathcal{M}_P$ and optimized mask $\mathcal{M}_O$ , Focal loss [31] for confidence scores $S$ , and L1 and GIoU [48] loss for bounding boxes $\mathcal{B}$ . The final training loss functions are: + +$$ +\mathcal {L} _ {\text {t r a i n}} = \lambda_ {\mathcal {M} _ {P}} \mathcal {L} _ {\mathcal {M} _ {P}} + \lambda_ {\mathcal {M} _ {O}} \mathcal {L} _ {\mathcal {M} _ {O}} + \lambda_ {\mathcal {B}} \mathcal {L} _ {\mathcal {B}} + \lambda_ {\mathcal {S}} \mathcal {L} _ {\mathcal {S}} \tag {6} +$$ + +where $\mathcal{L}$ and $\lambda$ are the loss term and weight, respectively. + +# 4. Experiments + +# 4.1. Datasets and Metrics + +Datasets. We evaluate $\mathrm{SgMg}$ on four video benchmarks: Ref-YouTube-VOS [49], Ref-DAVIS17 [21], A2D-Sentences [17], and JHMDB-Sentences [20]. RefYouTube-VOS is currently the largest dataset for R-VOS, containing 3,978 videos with about $13\mathrm{K}$ expressions. RefDAVIS17 is an extension of DAVIS17 [45] by including the language expressions of different objects and contains 90 videos. A2D-Sentences is a general actor and action segmentation dataset with over 3.7K videos and 6.6K action descriptions. JHMDB-Sentences includes 928 videos and 928 descriptions covering 21 different action classes. + +Evaluation Metrics. We adopt the standard metrics to evaluate our models: region similarity $\mathcal{I}$ (average IoU), contour accuracy $\mathcal{F}$ (average boundary similarity), and their mean value $\mathcal{I} \& \mathcal{F}$ . All results are evaluated using the official code or server. On A2D-Sentences and JHMDB-Sentences, we adopt mAP, overall IoU, and mean IoU for evaluation. + +# 4.2. Implementation Details + +Following [4, 6, 57], we train our models on the training set of Ref-YouTube-VOS, and directly evaluate them on the validation split of Ref-YouTube-VOS and Ref-DAVIS17 without any additional techniques, e.g., model ensemble, joint training, and mask propagation, since they are not the focus of this paper. Additionally, we present results for our + +
MethodYearBackboneRef-YouTube-VOSRef-DAVIS17
J&FJFFPSJ&FJF
CMSA [63]2019ResNet-5036.434.838.1-40.236.943.5
URVOS [49]2020ResNet-5047.245.349.2-51.547.356.0
CMPC-V [32]2021I3D47.545.649.3----
PMINet [11]2021ResNeSt-10153.051.554.5----
YOFO [23]2022ResNet-5048.647.549.71053.348.857.8
LBDT [10]2022ResNet-5049.448.250.6-54.3--
MLRL [56]2022ResNet-5049.748.451.0-52.850.055.4
MTTR [4]2022Video-Swin-T55.354.056.6----
MANet [6]2022Video-Swin-T55.654.856.5----
ReferFormer [57]2022Video-Swin-T56.054.857.350---
SgMg (Ours)2023Video-Swin-T58.957.760.06556.753.360.0
+ +Pre-training with RefCOCO/+/g & larger backbone + +
ReferFormer [57]2022Video-Swin-T59.458.060.95059.656.562.7
SgMg (Ours)2023Video-Swin-T62.060.463.56561.959.064.8
ReferFormer [57]2022Video-Swin-B62.961.364.63361.158.164.1
SgMg (Ours)2023Video-Swin-B65.763.967.44063.360.666.0
+ +Table 2. Quantitative comparison to state-of-the-art methods on the validation split of Ref-YouTube-VOS and Ref-DAVIS17. + +
MethodBackboneA2D-SentencesJHMDB-Sentences
mAPOverall IoUMean IoUmAPOverall IoUMean IoU
Hu et al. [19]VGG-1613.247.435.017.854.652.8
Gavrilyuk et al. [17]I3D19.853.642.123.354.154.2
ACAN [54]I3D27.460.149.028.957.658.4
CMPC-V [32]I3D40.465.357.334.261.661.7
ClawCraneNet [29]ResNet-50/101-63.159.9-64.465.6
MTTR [4]Video-Swin-T46.172.064.039.270.169.8
ReferFormer [57]Video-Swin-T52.877.669.642.271.971.0
SgMg (Ours)Video-Swin-T56.178.070.444.472.871.7
ReferFormer [57]Video-Swin-B55.078.670.343.773.071.8
SgMg (Ours)Video-Swin-B58.579.972.045.073.772.5
+ +Table 3. Quantitative comparison to state-of-the-art R-VOS methods on A2D-Sentences and JHMDB-Sentences. + +models first pre-trained on RefCOCO+/g [39, 65] and then fine-tuned on Ref-YouTube-VOS. Similar to [57, 72], we set the coefficients for different losses $\lambda_{dice}$ , $\lambda_{focal}$ , $\lambda_{L1}$ , $\lambda_{giou}$ to 5, 2, 5, 2, respectively. The models are trained using 2 RTX 3090 GPUs with 5 frames per clip for 9 epochs. All frames are resized to have the longest side of 640 pixels. Further implementation details are in the supplementary material. + +# 4.3.Quantitative Results + +Ref-YouTube-VOS and Ref-DAVIS17. We compare $\mathrm{SgMg}$ with recently published works in Table 2. Our approach surpasses present solutions on the two datasets across all metrics. On Ref-YouTube-VOS, $\mathrm{SgMg}$ with the Video Swin Tiny backbone achieves $58.9\mathcal{J}\&\mathcal{F}$ at 65 FPS, which is $2.9\%$ higher and $1.3\times$ faster than the previous state-of-the-art ReferFormer [57]. Our approach runs faster due to the use of the segment-and-optimize pipeline, which avoids the need for heavy feature decoders. When pre + +training with RefCOCO/+/g and using a larger backbone, i.e., Video Swin Base, the performance of $\mathrm{SgMg}$ further boosts to 65.7 $\mathcal{J}\&\mathcal{F}$ , consistently leading all other solutions by more than $2.8\%$ . On Ref-DAVIS17, $\mathrm{SgMg}$ achieves 63.3 $\mathcal{J}\&\mathcal{F}$ , outperforming state-of-the-art by $2.2\%$ and demonstrating the generality of our approach. + +A2D-Sentences and JHMDB-Sentences. We further evaluate $\mathrm{SgMg}$ on A2D-Sentences and JHMDB-Sentences in Table 3. Following [57], the models are first pre-trained on RefCOCO+/g and then fine-tuned on A2D-Sentences. JHMDB-Sentences is used only for evaluation. As shown in Table 3, $\mathrm{SgMg}$ achieves superior performance compared to other state-of-the-art R-VOS methods and surpasses the nearest competitor Referformer [57] by $3.5 / 1.3\%$ mAP on A2D-Sentences and JHMDB-Sentences, respectively. + +Multi-object R-VOS. We extend $\mathrm{SgMg}$ to perform multi-object R-VOS, which is more practical and efficient for deployment. Fast $\mathrm{SgMg}$ is trained on Ref-YouTube-VOS without pre-training or postprocessing techniques. We + +
MethodRef-DAVIS17Ref-YouTube-VOS
J&FJFJ&FJFFPS
ReferFormer [57]54.551.058.056.054.857.350
Fast SgMg (Ours)54.251.157.354.253.155.3185
+ +Table 4. Evaluation of Fast SgMg on Ref-DAVIS17 and RefYouTube-VOS. Video-Swin-T is adopted as the backbone. + +
ComponentsPerformance
CPKMSOSCFJ&FJFFPS
54.452.756.270
✓55.854.557.170
✓57.756.359.169
✓55.854.357.466
✓✓57.956.759.169
✓✓✓58.957.760.065
+ +benchmark Fast SgMg on Ref-YouTube-VOS and RefDAVIS17 using the commonly used Video Swin Tiny, and compare the results with the state-of-the-art R-VOS method, which performs single-object segmentation. + +As shown in Table 4, Fast $\mathrm{SgMg}$ achieves reasonable performance and runs about $3.7\times$ faster (185 vs 50 FPS) compared to ReferFormer [57]. It should be noted that each object in the above datasets contains multiple expressions. On Ref-DAVIS17, we group expressions to have only one expression per object within each group and segment all expressions in each group simultaneously since the object identity is given. On Ref-YouTube-VOS, all expressions in a video are segmented simultaneously due to the lack of object identity, making it more challenging. + +# 4.4. Ablation Study for Different Components + +We conduct ablation experiments to evaluate the effectiveness of different components in $\mathrm{SgMg}$ . The components are added to the baseline model step-by-step. + +Conditional Patch Kernel. As shown in Table 5, CPK boosts the performance by $1.4\%$ compared with the recent instance-aware conditional kernels [57]. The sequential labels of each token predicted by CPK contain more fine-grained information, making the prediction more accurate. + +Multi-granularity Segmentation Optimizer. We devise MSO to optimize the predicted patch masks. As shown in Table 5, MSO improves the performance by $3.3\%$ , indicating the importance of fine-grained visual details in R-VOS. Spectrum-guided Cross-modal Fusion. We present SCF to perform global interactions by operating in the spectral domain. In Table 5, using SCF to replace the traditional cross-attention in [57, 52] improves the $\mathcal{J}\&\mathcal{F}$ by $1.4\%$ . We consider SCF extracts important low-frequency features and facilitates multimodal understanding globally, which is suitable for R-VOS since locating referred objects requires understanding the global context and token relations. + +Table 5. Ablation of different components on Ref-YouTube-VOS. + +
SettingsDriftPipelineJ&FFPS
Baseline + Decoder✓decode-and-segment56.050
Baseline + Decoder + MSO✓decode-and-segment56.449
Baseline + MSO×segment-and-optimize57.769
+ +Table 6. Feature drift analysis using ReferFormer [57] (Baseline + Decoder) and $\mathrm{SgMg}$ w/o CPK & SCF (Baseline + MSO). Significant improvement is achieved by addressing the drift issue (last row). Adding MSO on top of ReferFormer to recover visual details (for a second time) still performs worse than our basic pipeline. + +
MethodRefCOCORefCOCO+RefCOCOg
MaIL [28]70.162.262.5
CRIS [55]70.562.359.9
RefTR [25]70.6--
LAVT [62]72.762.161.2
VLT [9]73.063.563.5
SgMg (Ours)76.366.470.0
+ +Table 7. Quantitative evaluation on the validation split of Ref-COCO/+/g. Overall IoU is adopted as the evaluation metric. + +# 4.5. Ablation Study for Feature Drift + +We conduct ablation study in Table 6 to demonstrate the feature drift problem. Our segment-and-optimize pipeline addresses the adverse drift effect discussed in Section 3.1 to significantly outperform ReferFormer [57] by $1.7\%$ points and runs $1.4 \times$ faster. Furthermore, adding MSO on top of ReferFormer still performs worse due to the negative drift impact caused by the decode-and-segment pipeline. These results demonstrate the efficacy of our proposed segment-and-optimize pipeline. + +# 4.6. Referring Image Segmentation Results + +We apply $\mathrm{SgMg}$ to referring image (expression) segmentation without any architectural modifications, and compare against the current state-of-the-art methods on RefCOCO $+ / g$ [39, 65]. A single $\mathrm{SgMg}$ model is trained on RefCOCO $+ / g$ without large-scale pre-training. As shown in Table 7, $\mathrm{SgMg}$ achieves advanced performance on all three benchmarks. These results demonstrate the efficacy of $\mathrm{SgMg}$ in referring image segmentation. + +# 4.7. Inference Time Analysis of Multi-Object RVOS + +We analyze the efficiency of the proposed multi-object R-VOS paradigm by comparing the FPS of Fast $\mathrm{SgMg}$ and $\mathrm{SgMg}$ on videos with different numbers of expressions. As illustrated in Fig. 6, Fast $\mathrm{SgMg}$ performs about $2\times$ faster than $\mathrm{SgMg}$ when there are two expressions per video on average. As the number of expressions increases, Fast $\mathrm{SgMg}$ achieves faster reasoning time per object per frame due to its utilization of the multi-object R-VOS paradigm. When there are ten expressions in each video, Fast $\mathrm{SgMg}$ performs at nearly 300 FPS, which is about $5\times$ faster than $\mathrm{SgMg}$ . + +![](images/cbcb32da13ed3ec608c9c4aada756dbe130df296a2ba08aa0c6443ae4e1029e8.jpg) +Expression: a turtle to the bottom left. +Figure 5. Qualitative comparison of our method with others. + +![](images/55fe67df66eb4d6cc68ac6b33fcae467955bcb6523e69cd99e50525433d96014.jpg) +Figure 6. Efficiency analysis of $\mathrm{SgMg}$ and Fast $\mathrm{SgMg}$ for videos with different numbers of expressions on Ref-YouTube-VOS. + +# 4.8. Qualitative Results + +In Fig. 5, we show qualitative comparison with ReferFormer [57] and MTTR [4]. $\mathrm{SgMg}$ can handle different objects of the same category or with the same behavior. + +# 4.9. Feature Visualization of SCF + +In Fig. 7, we visualize the vision-language features extracted by our SCF in comparison to the cross-attention used in [57]. The features extracted by SCF exhibit superior grounding ability in locating target objects, resulting in better performance for $\mathrm{SgMg}$ . + +# 5. Conclusion + +We discovered the feature drift issue in current referring video object segmentation (R-VOS) methods, which negatively affects the segmentation kernels. We presented $\mathrm{SgMg}$ , a novel segment-and-optimize approach for R-VOS that avoids the drift issue and optimizes masks with visual + +![](images/e457b08af0144c867c4fde393fa6bb8ccf2eada1d54b3835be531cac4aa387d0.jpg) +Figure 7. Visualization of the vision-language features extracted w/o and w/ our SCF. + +details. We also provided a new perspective to encourage vision-language global interactions in the spectral domain with Spectrum-guided Cross-modal Fusion. Additionally, we proposed the multi-object R-VOS paradigm by extending $\mathrm{SgMg}$ with multi-instance fusion and decoupling. Finally, we evaluated our models on four video benchmarks and demonstrated that our approach achieves state-of-the-art performance on all four datasets. + +Acknowledgment. This research was supported by the Australian Research Council Industrial Transformation Research Hub IH180100002. Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government. + +# References + +[1] Nayyer Aafaq, Naveed Akhtar, Wei Liu, Syed Zulqarnain Gilani, and Ajmal Mian. Spatio-temporal dynamics and semantic attribute enriched visual encoding for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12487-12496, 2019. 2 +[2] Miriam Bellver, Carles Ventura, Carina Silberer, Ioannis Kazakos, Jordi Torres, and Xavier Giro-i Nieto. A closer look at referring expressions for video object segmentation. Multimedia Tools and Applications, pages 1-20, 2022. 1 +[3] Glenn D Bergland. A guided tour of the fast fourier transform. IEEE spectrum, 6(7):41-52, 1969. 3 +[4] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In CVPR, pages 4985-4995, 2022. 1, 2, 3, 4, 5, 6, 8 +[5] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 221-230, 2017. 2 +[6] Weidong Chen, Dexiang Hong, Yuankai Qi, Zhenjun Han, Shuhui Wang, Laiyun Qing, Qingming Huang, and Guorong Li. Multi-attention network for compressed video referring object segmentation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4416–4425, 2022. 1, 5, 6 +[7] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Proc. Adv. Neural Inf. Process. Syst. (NIPS), 34:11781-11794, 2021. 1, 2 +[8] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3 +[9] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 7 +[10] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 2, 6 +[11] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, page 7, 2021. 2, 5, 6 +[12] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 2 +[13] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 2 + +[14] Guang Feng, Lihe Zhang, Zhiwei Hu, and Hutchuan Lu. Deeply interleaved two-stream encoder for referring video segmentation. arXiv preprint arXiv:2203.15969, 2022. 2 +[15] Mingtao Feng, Haoran Hou, Liang Zhang, Yulan Guo, Hongshan Yu, Yaonan Wang, and Ajmal Mian. Exploring hierarchical spatial layout cues for 3d point cloud based scene graph prediction. IEEE Transactions on Multimedia, 2023. 1 +[16] Zhiheng Fu, Siyu Hong, Mengyi Liu, Hamid Laga, Mohammed Bennamoun, Farid Boussaid, and Yulan Guo. Multi-stage information diffusion for joint depth and surface normal estimation. Pattern Recognition, 141:109660, 2023. 1 +[17] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 2, 5, 6 +[18] Lorenzo Giambagli, Lorenzo Buffoni, Timoteo Carletti, Walter Nocentini, and Duccio Fanelli. Machine learning in spectral domain. Nature communications, 12(1):1-9, 2021. 3 +[19] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 6 +[20] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 2, 5 +[21] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Asian Conference on Computer Vision, pages 123-141. Springer, 2018. 1, 2, 5 +[22] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5 +[23] Dezhuang Li, Ruoqi Li, Lijun Wang, Yifan Wang, Jinqing Qi, Lu Zhang, Ting Liu, Qingquan Xu, and Huchuan Lu. You only infer once: Cross-modal meta-transfer for referring video object segmentation. In AAAI Conference on Artificial Intelligence, 2022. 1, 5, 6 +[24] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2 +[25] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in neural information processing systems, 34:19652-19664, 2021.7 +[26] Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855, 2019. 5 +[27] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Yan Lu, and Bhiksha Raj. R^2vos: Robust referring video object segmentation via relational multimodal cycle consistency. arXiv preprint arXiv:2207.01203, 2022. 1 + +[28] Zizhang Li, Mengmeng Wang, Jianbiao Mei, and Yong Liu. Mail: A unified mask-image-language trimodal network for referring image segmentation. arXiv preprint arXiv:2111.10747, 2021. 7 +[29] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawcranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 2, 6 +[30] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 2 +[31] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[32] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 6 +[33] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3 +[34] Yong Liu, Ran Yu, Jiahao Wang, Xinyuan Zhao, Yitong Wang, Yansong Tang, and Yujiu Yang. Global spectral filter memory network for video object segmentation. In European Conference on Computer Vision, pages 648-665. Springer, 2022. 3 +[35] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 3 +[36] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2 +[37] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10437–10446, 2020. 2 +[38] Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsupervised video object segmentation with co-attention siamese networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3623–3632, 2019. 2 +[39] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6, 7 +[40] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and + +localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3 +[41] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Region aware video object segmentation with deep motion modeling. arXiv preprint arXiv:2207.10258, 2022.1 +[42] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Self-supervised video object segmentation by motion-aware mask propagation. In Proceedings of the IEEE International Conference on Multimedia and Expo, 2022. 2 +[43] Bo Miao, Liguang Zhou, Ajmal Mian, Tin Lun Lam, and Yangsheng Xu. Object-to-scene: Learning to transfer object knowledge to indoor scene recognition. In Proc. IEEE Int. Conf. Intell. Robots Syst. (IROS). IEEE, 2021. 1 +[44] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proc. Int. Conf. Comput. Vis. (ICCV), pages 9226–9235, 2019. 1, 2 +[45] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 5 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[47] Sucheng Ren, Wenxi Liu, Yongtuo Liu, Haoxin Chen, Guoqiang Han, and Shengfeng He. Reciprocal transformations for unsupervised video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15455-15464, 2021. 2 +[48] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5 +[49] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223. Springer, 2020. 1, 2, 5, 6 +[50] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In European conference on computer vision, pages 282-298. Springer, 2020. 1, 2, 3, 4 +[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 2 +[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 7 +[53] Paul Voiglaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast + +end-to-end embedding learning for video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9481-9490, 2019. 2 +[54] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 6 +[55] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 7 +[56] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 1, 5, 6 +[57] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[58] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. arXiv preprint arXiv:2308.02874, 2023.1 +[59] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019.4 +[60] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In International Conference on Neural Information Processing, pages 264-274. Springer, 2019. 4 +[61] Zhao Yang, Yansong Tang, Luca Bertinetto, Hengshuang Zhao, and Philip HS Torr. Hierarchical interaction network for video object segmentation from referring expressions. In BMVC, 2021. 1, 2 +[62] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Ladv: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 7 +[63] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 5, 6 +[64] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 4 +[65] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 6, 7 + +[66] Z. Yang and Y, Wei and Y. Yang. Associating objects with transformers for video object segmentation. Proc. Adv. Neural Inf. Process. Syst. (NIPS), 34:2491-2502, 2021. 2 +[67] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 2 +[68] Shuai Zhao, Xiaohan Wang, Linchao Zhu, and Yi Yang. Test-time adaptation with clip reward for zero-shot generalization in vision-language models. arXiv preprint arXiv:2305.18010, 2023. 2 +[69] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. arXiv preprint arXiv:2205.00823, 2022. 2 +[70] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 2 +[71] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13041-13049, 2020. 2 +[72] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 5, 6 \ No newline at end of file diff --git a/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/images.zip b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2a54f587226344d0c78dcf5656fab20e405f415e --- /dev/null +++ b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0511f6e73bf14a5585a1614460a6122ea3d450a9c895938db0f2794a5c6b374 +size 685680 diff --git a/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/layout.json b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..85c00cf977747b813fb1ac4e83d76978b66e7e55 --- /dev/null +++ b/2023/Spectrum-guided Multi-granularity Referring Video Object Segmentation/layout.json @@ -0,0 +1,11455 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 136, + 509, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 136, + 509, + 154 + ], + "spans": [ + { + "bbox": [ + 66, + 136, + 509, + 154 + ], + "type": "text", + "content": "Spectrum-guided Multi-granularity Referring Video Object Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "spans": [ + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "text", + "content": "Bo Miao" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "text", + "content": ", Mohammed Bennamoun" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "text", + "content": ", Yongsheng Gao" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "text", + "content": ", Ajmal Mian" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "text", + "content": "The University of Western Australia \n" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 174, + 459, + 208 + ], + "type": "text", + "content": "Griffith University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 245, + 186, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 245, + 186, + 256 + ], + "spans": [ + { + "bbox": [ + 139, + 245, + 186, + 256 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 268, + 282, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 268, + 282, + 548 + ], + "spans": [ + { + "bbox": [ + 45, + 268, + 282, + 548 + ], + "type": "text", + "content": "Current referring video object segmentation (R-VOS) techniques extract conditional kernels from encoded (low-resolution) vision-language features to segment the decoded high-resolution features. We discovered that this causes significant feature drift, which the segmentation kernels struggle to perceive during the forward computation. This negatively affects the ability of segmentation kernels. To address the drift problem, we propose a Spectrum-guided Multigranularity (SgMg) approach, which performs direct segmentation on the encoded features and employs visual details to further optimize the masks. In addition, we propose Spectrum-guided Cross-modal Fusion (SCF) to perform intra-frame global interactions in the spectral domain for effective multimodal representation. Finally, we extend SgMg to perform multi-object R-VOS, a new paradigm that enables simultaneous segmentation of multiple referred objects in a video. This not only makes R-VOS faster, but also more practical. Extensive experiments show that SgMg achieves state-of-the-art performance on four video benchmark datasets, outperforming the nearest competitor by " + }, + { + "bbox": [ + 45, + 268, + 282, + 548 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 45, + 268, + 282, + 548 + ], + "type": "text", + "content": " points on Ref-YouTube-VOS. Our extended SgMg enables multi-object R-VOS, runs about " + }, + { + "bbox": [ + 45, + 268, + 282, + 548 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 45, + 268, + 282, + 548 + ], + "type": "text", + "content": " faster while maintaining satisfactory performance. Code is available at https://github.com/bo-miao/SgMg." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 124, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 124, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 124, + 581 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 590, + 279, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 590, + 279, + 706 + ], + "spans": [ + { + "bbox": [ + 45, + 590, + 279, + 706 + ], + "type": "text", + "content": "Referring video object segmentation (R-VOS) aims at segmenting objects in a video, referred to by linguistic descriptions. R-VOS is an emerging task for multimodal reasoning and promotes a wide range of applications, including language-guided video editing and human-machine interaction. Different from conventional semi-supervised video object segmentation [44, 7, 41], where the mask annotation for the first frame is provided for reference, R-VOS is more challenging due to the need for cross-modal understanding between vision and free-form language expressions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 706, + 279, + 730 + ], + "type": "text", + "content": "Early R-VOS techniques [2, 21, 61] perform feature encoding, cross-modal interaction, and language grounding" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 300, + 245, + 530, + 357 + ], + "blocks": [ + { + "bbox": [ + 300, + 245, + 530, + 357 + ], + "lines": [ + { + "bbox": [ + 300, + 245, + 530, + 357 + ], + "spans": [ + { + "bbox": [ + 300, + 245, + 530, + 357 + ], + "type": "image", + "image_path": "89d6e208c163b4cbcd2b4ac342aeb8c5568808f60e50da5bd790db27323de29d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 366, + 447, + 373 + ], + "lines": [ + { + "bbox": [ + 301, + 366, + 447, + 373 + ], + "spans": [ + { + "bbox": [ + 301, + 366, + 447, + 373 + ], + "type": "text", + "content": "(b) SgMg (Ours): segment-and-optimize w/ encoded features " + }, + { + "bbox": [ + 301, + 366, + 447, + 373 + ], + "type": "inline_equation", + "content": "F_{vl}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 299, + 380, + 526, + 433 + ], + "blocks": [ + { + "bbox": [ + 299, + 380, + 526, + 433 + ], + "lines": [ + { + "bbox": [ + 299, + 380, + 526, + 433 + ], + "spans": [ + { + "bbox": [ + 299, + 380, + 526, + 433 + ], + "type": "image", + "image_path": "fc603f36ec204257f55403249444807ca1ada13606a4021792a967903df7776c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "lines": [ + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "spans": [ + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "text", + "content": "Conditional Segmentation Fusion: Cross-modal Fusion " + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "inline_equation", + "content": "F_{\\mu l}" + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "text", + "content": " : Encoded Vision-Language Features " + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "inline_equation", + "content": "F_{\\nu l}^{d}" + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "text", + "content": " : Decoded Features Q: Instance Query " + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "inline_equation", + "content": "K_{c}" + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "text", + "content": " : Conditional Kernel " + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "inline_equation", + "content": "K_{cp}" + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "text", + "content": " : Conditional Patch Kernel " + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "inline_equation", + "content": "M_{Q}" + }, + { + "bbox": [ + 301, + 435, + 523, + 462 + ], + "type": "text", + "content": " : Predicted Mask" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "lines": [ + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "spans": [ + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "text", + "content": "Figure 1. (a) Previous methods [4, 57] apply segmentation kernels " + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_c" + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "text", + "content": " [50], extracted from encoded features " + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "text", + "content": ", to segment the decoded high-resolution features " + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}^d" + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "text", + "content": ". (b) We use segmentation kernels " + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{cp}" + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "text", + "content": ", extracted from encoded features " + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "text", + "content": ", to segment the encoded features " + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 296, + 464, + 530, + 530 + ], + "type": "text", + "content": " directly, and propose multi-granularity optimization to recover visual details and produce fine-grained masks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "text", + "content": "using convolutional neural networks (CNNs). However, the limited ability of CNNs to capture long-range dependencies and handle free-form features constrains the model performance. With the advancement of attention mechanisms [52, 43, 15, 16, 58], recent methods achieved significant improvement on R-VOS using cross-attention [49, 23, 27] for multimodal understanding and transformers [6, 56] for spatio-temporal representation. Based on transformers, conditional kernel [50] is then introduced to separate foreground from semantic features given its high adaptability to different instances [4, 57]. As illustrated in Fig. 1(a), these methods attend to encoded vision-language features " + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "text", + "content": " using instance queries " + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "text", + "content": " to predict conditional kernels " + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\kappa_{c}" + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "text", + "content": ", and employ " + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\kappa_{c}" + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "text", + "content": " as the segmentation head to segment decoded features " + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}^{d}" + }, + { + "bbox": [ + 296, + 554, + 531, + 730 + ], + "type": "text", + "content": ". Despite the promising performance," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 72 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 41, + 470, + 73 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 41, + 470, + 73 + ], + "spans": [ + { + "bbox": [ + 151, + 41, + 470, + 73 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 281, + 746, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 746, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 746, + 296, + 756 + ], + "type": "text", + "content": "920" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 223 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 223 + ], + "type": "text", + "content": "this paradigm still has some limitations. Firstly, as shown in the t-SNE [51] visualization in Fig. 1(a), although the nonlinear decoding process introduces visual details, this is accompanied by a significant feature drift, which increases the difficulty of segmentation since " + }, + { + "bbox": [ + 45, + 106, + 279, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_c" + }, + { + "bbox": [ + 45, + 106, + 279, + 223 + ], + "type": "text", + "content": " is predicted before feature decoding. Secondly, bilinear upsampling of the predicted masks " + }, + { + "bbox": [ + 45, + 106, + 279, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_Q" + }, + { + "bbox": [ + 45, + 106, + 279, + 223 + ], + "type": "text", + "content": " to increase resolution impedes the segmentation performance. Thirdly, these methods only support single expression-based segmentation, making R-VOS inefficient when multiple referred objects exist in a video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 223, + 280, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 223, + 280, + 432 + ], + "spans": [ + { + "bbox": [ + 45, + 223, + 280, + 432 + ], + "type": "text", + "content": "In this work, we propose a Spectrum-guided Multi-granularity (SgMg) approach that follows a segment-and-optimize pipeline to address the above problems. As depicted in Fig. 1(b), SgMg introduces Conditional Patch Kernel (CPK) " + }, + { + "bbox": [ + 45, + 223, + 280, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_{cp}" + }, + { + "bbox": [ + 45, + 223, + 280, + 432 + ], + "type": "text", + "content": " to directly segment its fully perceived encoded features " + }, + { + "bbox": [ + 45, + 223, + 280, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 45, + 223, + 280, + 432 + ], + "type": "text", + "content": ", avoiding the feature drift and its adverse effects. The segmentation is then refined using our proposed Multi-granularity Segmentation Optimizer (MSO), which employs low-level visual details to produce full-resolution masks. Within the SgMg framework, we further develop Spectrum-guided Cross-modal Fusion (SCF) that performs intra-frame global interactions in the spectral domain to facilitate multimodal understanding. Finally, we introduce a new paradigm called multi-object R-VOS to simultaneously segment multiple referred objects in a video. To achieve this, we extend SgMg by devising multi-instance fusion and decoupling. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 447, + 279, + 611 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 447, + 279, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 279, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 279, + 506 + ], + "type": "text", + "content": "- We explain how existing R-VOS methods suffer from the feature drift problem. To address this problem, we propose " + }, + { + "bbox": [ + 55, + 447, + 279, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 55, + 447, + 279, + 506 + ], + "type": "text", + "content": " that follows a segment-and-optimize pipeline and achieves top-ranked overall performance on multiple benchmark datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 279, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 279, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 279, + 551 + ], + "type": "text", + "content": "- We propose Spectrum-guided Cross-modal Fusion to encourage intra-frame global interactions in the spectral domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 564, + 279, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 279, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 279, + 611 + ], + "type": "text", + "content": "- We extend " + }, + { + "bbox": [ + 55, + 564, + 279, + 611 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 55, + 564, + 279, + 611 + ], + "type": "text", + "content": " to perform multi-object R-VOS, a new paradigm that enables simultaneous segmentation of multiple referred objects in a video. Our multi-object variant is more practical and runs " + }, + { + "bbox": [ + 55, + 564, + 279, + 611 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 55, + 564, + 279, + 611 + ], + "type": "text", + "content": " faster." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "text", + "content": "We conduct extensive experiments on multiple benchmark datasets, including Ref-YouTube-VOS [49], RefDAVIS17 [21], A2D-Sentences [17], and JHMDB-Sentences [20], and achieve state-of-the-art performance on all four. On the largest validation set Ref-YouTube-VOS, " + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "text", + "content": " achieves " + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "inline_equation", + "content": "65.7\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "text", + "content": " which is " + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "text", + "content": " points higher than that of the closest competitor ReferFormer [57]. On the A2D-Sentences, " + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "text", + "content": " achieves " + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "inline_equation", + "content": "58.5\\mathrm{mAP}" + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "text", + "content": " which is " + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 45, + 625, + 279, + 730 + ], + "type": "text", + "content": " points higher than that of ReferFormer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 297, + 105, + 386, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 105, + 386, + 117 + ], + "spans": [ + { + "bbox": [ + 297, + 105, + 386, + 117 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 124, + 530, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 124, + 530, + 252 + ], + "spans": [ + { + "bbox": [ + 296, + 124, + 530, + 252 + ], + "type": "text", + "content": "Video Object Segmentation techniques fall into two categories: unsupervised and semi-supervised. Unsupervised approaches segment the most salient instances in each video without user interactions [38, 47]. They often employ two-stream networks to fuse motion and appearance cues for segmentation. Semi-supervised approaches track the given first frame object mask by performing online learning [5] or spatial-temporal association [44, 7, 66, 42, 53]. Unlike conventional semi-supervised video object segmentation, R-VOS takes a free-form linguistic expression as guidance to detect and segment referred objects in videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 252, + 530, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 252, + 530, + 415 + ], + "spans": [ + { + "bbox": [ + 296, + 252, + 530, + 415 + ], + "type": "text", + "content": "Referring Video Object Segmentation. R-VOS methods mainly use deep neural networks with vision-and-language interaction to empower visual features with corresponding linguistic information for pixel-level segmentation. For example, [49] employs a unified R-VOS framework that performs iterative segmentation guided by both language and temporal features. [32, 29] adopt progressive segmentation by perceiving potential objects and discriminating the best match. [70] fuses visual and motion features for segmentation under the guidance of linguistic cues. [30] models object relations to form tracklets and performs tracklet-language grounding. To enhance multi-modal interactions, [61, 14, 11, 10] perform hierarchical vision-language fusion on multiple feature layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 415, + 530, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 415, + 530, + 600 + ], + "spans": [ + { + "bbox": [ + 296, + 415, + 530, + 600 + ], + "type": "text", + "content": "Despite their promising performance, the complex multistage pipelines and use of multiple networks make R-VOS burdensome. To address these problems, MTTR [4] proposes an end-to-end transformer-based network with conditional kernels [50] to segment target objects. ReferFormer [57] further introduces language-guided instance queries to predict instance-aware conditional kernels and an auxiliary detection task to aid localization. These methods follow a decode-and-segment pipeline, which adopts conditional kernels to segment decoded high-resolution features to achieve promising performance. However, the nonlinear decoding process leads to significant feature drift that negatively affects the conditional kernels. In contrast to previous works, our approach follows a segment-and-optimize pipeline to avoid the adverse drift effects and to predict full-resolution masks in an efficient manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 602, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 602, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 602, + 530, + 730 + ], + "type": "text", + "content": "Vision and Language Representation Learning aims to learn vision-language semantics and alignment for multimodal reasoning tasks. It has achieved significant success in various tasks [37, 67, 68], including video question answering [71], video captioning [1], video-text retrieval [12, 69], zero-shot classification [46], referring image/video segmentation [49], etc. Some approaches [46, 24] rely on contrastive pre-training using large-scale datasets to project different modalities into unified embedding space. Others [36, 13] develop cross-modal interaction layers for multimodal feature fusion and understanding. Recent deep" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 280, + 746, + 295, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 746, + 295, + 756 + ], + "spans": [ + { + "bbox": [ + 280, + 746, + 295, + 756 + ], + "type": "text", + "content": "921" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 106, + 516, + 304 + ], + "blocks": [ + { + "bbox": [ + 58, + 106, + 516, + 304 + ], + "lines": [ + { + "bbox": [ + 58, + 106, + 516, + 304 + ], + "spans": [ + { + "bbox": [ + 58, + 106, + 516, + 304 + ], + "type": "image", + "image_path": "a094d27b35f2ca47ca5c6634e7b32f40cd1830c5bc4be6a1bb9f6c78fe79b680.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "lines": [ + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "spans": [ + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "text", + "content": "Figure 2. The overall framework of " + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "text", + "content": ". Taking a video sequence " + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{I_i\\}_{i=1}^T" + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "text", + "content": " and a language expression " + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\{S_i\\}_{i=1}^N" + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "text", + "content": " as input, " + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "text", + "content": " predicts the masks of referred object " + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathcal{L}}" + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "text", + "content": " in each frame. SCF projects visual features " + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v" + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "text", + "content": " to vision-language features " + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "text", + "content": ", instance-aware CPK predicts patch masks by segmenting encoded " + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 45, + 309, + 531, + 342 + ], + "type": "text", + "content": ", and MSO optimizes patch masks to get fine-grained results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 361, + 280, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 361, + 280, + 431 + ], + "spans": [ + { + "bbox": [ + 45, + 361, + 280, + 431 + ], + "type": "text", + "content": "learning methods in spectral domain [18, 40, 8, 34] have raised widespread awareness because of their ability to perform global interactions. We take inspiration from these spectral-based methods and employ spectrum guidance in the field of vision-language representation to encourage multimodal global interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 447, + 279, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 447, + 279, + 476 + ], + "spans": [ + { + "bbox": [ + 45, + 447, + 279, + 476 + ], + "type": "text", + "content": "3. " + }, + { + "bbox": [ + 45, + 447, + 279, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 447, + 279, + 476 + ], + "type": "text", + "content": ": Spectrum-guided Multi-granularity Referring Video Object Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "spans": [ + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "text", + "content": "Given a video sequence " + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{I_i\\}_{i=1}^T" + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "text", + "content": " frames and a language query " + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\{S_i\\}_{i=1}^N" + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "text", + "content": " words. The goal of R-VOS is to segment the referred object " + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathcal{L}}" + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "text", + "content": " at pixel-level. To this end, we introduce a new approach termed " + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 482, + 279, + 554 + ], + "type": "text", + "content": ". Different from previous R-VOS methods [4, 57], our approach follows a segment-and-optimize pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": "An overview of " + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": " is shown in Fig. 2. Video Swin Transformer [35] is adopted to extract visual feature " + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v" + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": " and RoBERTa [33] is adopted to extract sentence " + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": " and word " + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w" + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": " features. The channel dimension of all features is projected to 256. Spectrum-guided Cross-modal Fusion (SCF) cross attends " + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v" + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w" + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": " to compute vision-language features " + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": ". Deformable Transformer [72] encoder is used to encode " + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": " and the decoder associates instance queries created based on " + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": " to predict instance embeddings and the corresponding Conditional Patch Kernels (CPKs). Finally, the CPKs are employed to segment " + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 45, + 555, + 280, + 730 + ], + "type": "text", + "content": " and predict patch masks that are further optimized with visual details through Multi-granularity Segmentation Optimizer (MSO). The choice of the encoder and transformer follows previous works to avoid distractions [4, 57]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 360, + 421, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 360, + 421, + 373 + ], + "spans": [ + { + "bbox": [ + 297, + 360, + 421, + 373 + ], + "type": "text", + "content": "3.1. Feature Drift Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "spans": [ + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "text", + "content": "Existing R-VOS methods [57, 4] follow a decode-and-segment pipeline where conditional kernels " + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_c" + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "text", + "content": " [50] are extracted from encoded features " + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "text", + "content": " and used to segment the decoded features " + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}^d" + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "text", + "content": ". However, the decoding process leads to feature drift, which is evident in the t-SNE visualization depicted in Fig. 1(a). This drift is difficult for the kernels " + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_c" + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "text", + "content": " to perceive during the forward computation since " + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_c" + }, + { + "bbox": [ + 296, + 378, + 530, + 519 + ], + "type": "text", + "content": " is predicted before the feature decoding. Therefore, we argue that even though the feature decoding enhances visual details, it also causes the drift problem that negatively affects the segmentation kernels. This makes the existing decode-and-segment pipeline sub-optimal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "spans": [ + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "type": "text", + "content": "To overcome the adverse effects of feature drift while recovering visual details, we present " + }, + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "type": "text", + "content": ", a novel approach that follows a segment-and-optimize pipeline. In a nutshell, " + }, + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "type": "text", + "content": " performs Spectrum-guided Cross-modal Fusion to compute " + }, + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "type": "text", + "content": ", leverages Conditional Patch Kernels to segment encoded features " + }, + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 296, + 517, + 531, + 612 + ], + "type": "text", + "content": " to avoid the drift effects, and recovers visual details with Multi-granularity Segmentation Optimizer to generate fine-grained masks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 297, + 618, + 492, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 618, + 492, + 631 + ], + "spans": [ + { + "bbox": [ + 297, + 618, + 492, + 631 + ], + "type": "text", + "content": "3.2. Spectrum-guided Cross-modal Fusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 530, + 730 + ], + "type": "text", + "content": "The two-dimensional discrete Fourier transform converts spatial data into the spectral domain. Based on the spectral convolution theorem [3], point-wise update of signals in the spectral domain globally affects all inputs in the spatial domain, which gives the insight to design spectrum-based modules so as to efficiently facilitate global interactions, which is critical for multimodal understanding. In addition, Low-frequency components in the spectral domain usually" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "922" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 103, + 274, + 254 + ], + "blocks": [ + { + "bbox": [ + 47, + 103, + 274, + 254 + ], + "lines": [ + { + "bbox": [ + 47, + 103, + 274, + 254 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 274, + 254 + ], + "type": "image", + "image_path": "2759da62b255c26bbafd9000234dd434c8a8d9e64545b5286f5ee30137e530ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 255, + 168, + 263 + ], + "lines": [ + { + "bbox": [ + 49, + 255, + 168, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 255, + 168, + 263 + ], + "type": "text", + "content": "(a) Spectrum-Guided Cross-Modal Fusion" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 255, + 259, + 263 + ], + "lines": [ + { + "bbox": [ + 177, + 255, + 259, + 263 + ], + "spans": [ + { + "bbox": [ + 177, + 255, + 259, + 263 + ], + "type": "text", + "content": "(b) Spectrum Augmentation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "lines": [ + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "spans": [ + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "type": "text", + "content": "Flatten: reshape " + }, + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_P" + }, + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\frac{H}{i}\\times \\frac{W}{i}\\times p^2}" + }, + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\frac{Hp}{i}\\times \\frac{Wp}{i}}" + }, + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "type": "text", + "content": " for visualization. " + }, + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 297, + 221, + 530, + 245 + ], + "type": "text", + "content": ": upsampling operation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 266, + 278, + 281 + ], + "blocks": [ + { + "bbox": [ + 47, + 266, + 278, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 266, + 278, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 278, + 281 + ], + "type": "image", + "image_path": "e07e40cf9af570d6f85d2835a23d862802b7928425ca500af0c62a6251d34559.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 285, + 279, + 316 + ], + "lines": [ + { + "bbox": [ + 45, + 285, + 279, + 316 + ], + "spans": [ + { + "bbox": [ + 45, + 285, + 279, + 316 + ], + "type": "text", + "content": "Figure 3. Spectrum-guided Cross-modal Fusion. Imag.: Imaginary. Pre-spectrum augmentation and post-spectrum augmentation share an identical structure." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 334, + 279, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 334, + 279, + 357 + ], + "spans": [ + { + "bbox": [ + 45, + 334, + 279, + 357 + ], + "type": "text", + "content": "correspond to the general semantic information according to previous theoretical studies [59, 60, 64]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 357, + 279, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 357, + 279, + 453 + ], + "spans": [ + { + "bbox": [ + 45, + 357, + 279, + 453 + ], + "type": "text", + "content": "Inspired by the above observations, we conjecture that low-frequency components can benefit higher dimensional semantic features and propose Spectrum-guided Cross-modal Fusion (SCF). As shown in Fig. 3, SCF performs pre-spectrum augmentation to enhance visual features before cross-modal fusion and post-spectrum augmentation to facilitate global vision-language interactions after the fusion process. Let " + }, + { + "bbox": [ + 45, + 357, + 279, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 45, + 357, + 279, + 453 + ], + "type": "text", + "content": " denotes the input features, the spectrum augmentation (SA) is computed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 50, + 460, + 279, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 460, + 279, + 473 + ], + "spans": [ + { + "bbox": [ + 50, + 460, + 279, + 473 + ], + "type": "interline_equation", + "content": "\\operatorname {S A} (\\mathcal {F}, K) = \\mathcal {F} + \\Theta_ {I F F T} (\\operatorname {C o n v} (\\sigma (K, \\mathcal {F}) \\odot \\Theta_ {F F T} (\\mathcal {F}))) \\tag {1}", + "image_path": "d47b2bcaa7d39cd2988e507fcb2abd6080d805207f6c7f9b74baacd0999d59b3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "spans": [ + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": " denotes low-pass filtering with adaptive Gaussian smoothed filters " + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "inline_equation", + "content": "\\sigma(K, \\mathcal{F})" + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": ", which has the same spatial size as " + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": " is the bandwidth. To make " + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "inline_equation", + "content": "\\sigma(K, \\mathcal{F})" + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": " input-aware, we create an initial 2D Gaussian map based on " + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": ", and apply pooling and linear layers on " + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": " to predict a scale parameter to update the Gaussian map. Thanks to the spectral convolution theorem, the efficient point-wise spectral convolution globally updates " + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": ". We treat the spectral-operated features as residuals and add them to the original input features for enhancement. Overall, SCF, which takes visual features " + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v" + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": " and word-level text features " + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w" + }, + { + "bbox": [ + 45, + 479, + 279, + 596 + ], + "type": "text", + "content": " as input, is computed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 603, + 279, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 603, + 279, + 615 + ], + "spans": [ + { + "bbox": [ + 66, + 603, + 279, + 615 + ], + "type": "interline_equation", + "content": "\\operatorname {S C F} \\left(\\mathcal {F} _ {w}, \\mathcal {F} _ {v}\\right) = \\operatorname {S A} \\left(\\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right) \\otimes \\operatorname {A t t} \\left(\\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right), \\mathcal {F} _ {w}\\right)\\right) \\tag {2}", + "image_path": "7be7e81aef15e5543a089b3e9bd968d7c5c2008de687bd80a607792e3a37055c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 630, + 216, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 630, + 216, + 642 + ], + "spans": [ + { + "bbox": [ + 45, + 630, + 216, + 642 + ], + "type": "text", + "content": "3.3. Conditional Patch Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 648, + 279, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 648, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 45, + 648, + 279, + 729 + ], + "type": "text", + "content": "We devise Conditional Patch Kernel (CPK) as the segmentation head to predict patch masks from the encoded vision-language features " + }, + { + "bbox": [ + 45, + 648, + 279, + 729 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 45, + 648, + 279, + 729 + ], + "type": "text", + "content": " that are fully perceived by CPK. Unlike previous works [4, 57], CPK predicts a sequence of labels for each token rather than a single label, efficiently improving segmentation resolution along the channel dimension." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 298, + 105, + 530, + 196 + ], + "blocks": [ + { + "bbox": [ + 298, + 105, + 530, + 196 + ], + "lines": [ + { + "bbox": [ + 298, + 105, + 530, + 196 + ], + "spans": [ + { + "bbox": [ + 298, + 105, + 530, + 196 + ], + "type": "image", + "image_path": "5352f0d2490d1e1f8f1404aadde77b7151607047992d91378dfdada585618560.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 199, + 530, + 221 + ], + "lines": [ + { + "bbox": [ + 296, + 199, + 530, + 221 + ], + "spans": [ + { + "bbox": [ + 296, + 199, + 530, + 221 + ], + "type": "text", + "content": "Figure 4. Multi-granularity Segmentation Optimizer, which predicts residual maps to optimize patch masks " + }, + { + "bbox": [ + 296, + 199, + 530, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_P" + }, + { + "bbox": [ + 296, + 199, + 530, + 221 + ], + "type": "text", + "content": " progressively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "spans": [ + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "text", + "content": "Specifically, we first use sentence-level text features " + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "text", + "content": " and multiple learnable embeddings to generate instance queries " + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "inline_equation", + "content": "Q\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "text", + "content": ". Next, " + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "text", + "content": " is projected into instance embeddings " + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "text", + "content": " using the transformer decoder and " + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 296, + 263, + 530, + 322 + ], + "type": "text", + "content": " is leveraged to predict CPK for each instance query:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 330, + 530, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 330, + 530, + 344 + ], + "spans": [ + { + "bbox": [ + 336, + 330, + 530, + 344 + ], + "type": "interline_equation", + "content": "\\operatorname {C P K} (Q, \\mathcal {F} _ {v l}) = \\Theta (\\operatorname {F C} (\\operatorname {A t t} (Q, \\mathcal {F} _ {v l}))) \\tag {3}", + "image_path": "1f70224931acdef8603dc104c49aca5ce1d088ab33f6c74bec9bbf0d2aefad37.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "spans": [ + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "text", + "content": " denotes the parameterization operation that reshapes CPK to form two point-wise convolutions with the output channel number of 16, which is similar to [50]. Since " + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "text", + "content": " changes dynamically according to different linguistic expressions, CPK becomes instance-aware and can separate objects of interest from " + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "text", + "content": ". Finally, we apply the parameterized CPK (dynamic point-wise convolutions) on " + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "text", + "content": " to predict patch masks " + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_P\\in \\mathbb{R}^{\\frac{H}{i}\\times \\frac{W}{i}\\times p^2}" + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "inline_equation", + "content": "\\frac{H}{i}\\times \\frac{W}{i}" + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "text", + "content": " denotes the spatial resolution of " + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "inline_equation", + "content": "p^2" + }, + { + "bbox": [ + 296, + 351, + 530, + 469 + ], + "type": "text", + "content": " denotes the increased segmentation resolution on the channel dimension." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "spans": [ + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "text", + "content": "During inference, we can reshape patch masks to " + }, + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_P\\in" + }, + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "inline_equation", + "content": "\\mathbb{R}\\frac{Hp}{i}\\times \\frac{Wp}{i}" + }, + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "text", + "content": " to efficiently generate fine-grained segmentation from low-resolution " + }, + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "text", + "content": ". The resolution of prediction will be consistent with the input when " + }, + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "text", + "content": " equals to " + }, + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 296, + 469, + 530, + 541 + ], + "type": "text", + "content": ". We found that this efficient CPK can achieve competitive performance compared to methods that use heavy decoders." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 297, + 548, + 515, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 548, + 515, + 561 + ], + "spans": [ + { + "bbox": [ + 297, + 548, + 515, + 561 + ], + "type": "text", + "content": "3.4. Multi-granularity Segmentation Optimizer" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 296, + 565, + 530, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 565, + 530, + 623 + ], + "spans": [ + { + "bbox": [ + 296, + 565, + 530, + 623 + ], + "type": "text", + "content": "Segmenting encoded features " + }, + { + "bbox": [ + 296, + 565, + 530, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 296, + 565, + 530, + 623 + ], + "type": "text", + "content": " with CPK avoids the detrimental drift effect on the segmentation head. However, visual details are required to produce accurate fine-grained masks. We propose Multi-granularity Segmentation Optimizer (MSO) to achieve this goal." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "text", + "content": "An overview of MSO is shown in Fig. 4. It takes the predicted patch masks " + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_P" + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "text", + "content": " as object priors and reuses visual features " + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v" + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "text", + "content": " with spatial strides of " + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\{4,8\\}" + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "text", + "content": " to gradually recover visual details and refine the priors. Specifically, MSO first concatenates " + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_P" + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v" + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "text", + "content": " and projects them to low dimensional bases. Next, residual masks predicted by performing another convolution on these bases are used to correct " + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_P" + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "text", + "content": ". Finally, the optimized patch masks achieve the input resolution by reshaping from " + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times 4^2}" + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 296, + 623, + 530, + 730 + ], + "type": "text", + "content": ". Since" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 295, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 295, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 295, + 756 + ], + "type": "text", + "content": "923" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 103, + 274, + 152 + ], + "blocks": [ + { + "bbox": [ + 49, + 103, + 274, + 152 + ], + "lines": [ + { + "bbox": [ + 49, + 103, + 274, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 103, + 274, + 152 + ], + "type": "table", + "html": "
MethodsSingle-frameMulti-framesMulti-objects
[49, 63, 11] et al.✓
[4, 57, 56, 23] et al.✓✓
Fast SgMg (Ours)✓✓✓
", + "image_path": "02f7359109a6fa52d753e26b6c274c9a713998aa69959a5b8775d86504c6a7a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 153, + 278, + 174 + ], + "lines": [ + { + "bbox": [ + 46, + 153, + 278, + 174 + ], + "spans": [ + { + "bbox": [ + 46, + 153, + 278, + 174 + ], + "type": "text", + "content": "Table 1. Comparing different methods for their ability to segment single or multiple frames or multiple objects simultaneously." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 190, + 279, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 190, + 279, + 224 + ], + "spans": [ + { + "bbox": [ + 45, + 190, + 279, + 224 + ], + "type": "text", + "content": "MSO does not include heavy computations, the segment-and-optimize pipeline makes our approach perform better with efficient inference time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 232, + 163, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 163, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 163, + 245 + ], + "type": "text", + "content": "3.5. Multi-Object R-VOS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 250, + 279, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 250, + 279, + 343 + ], + "spans": [ + { + "bbox": [ + 45, + 250, + 279, + 343 + ], + "type": "text", + "content": "Existing R-VOS methods perform single-frame (frameworkwise) segmentation [49, 63, 11] or multi-frame (clip-wise) segmentation [4, 57, 56] for an individual referred object at a time. However, to the best of our knowledge, no existing work explores the simultaneous segmentation of multiple referred objects in video using a common GPU, which is important for real-world scenarios. To fill this gap, we present a new paradigm called multi-object R-VOS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 344, + 279, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 344, + 279, + 436 + ], + "spans": [ + { + "bbox": [ + 45, + 344, + 279, + 436 + ], + "type": "text", + "content": "The key to multi-object R-VOS is designing a network that shares computationally intensive features for multiple objects, and enables different instance features to be decoupled before segmentation. To achieve this, we extend " + }, + { + "bbox": [ + 45, + 344, + 279, + 436 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 344, + 279, + 436 + ], + "type": "text", + "content": " for multi-object R-VOS by introducing multi-instance fusion and decoupling. As shown in Table 1, our method, dubbed Fast " + }, + { + "bbox": [ + 45, + 344, + 279, + 436 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 344, + 279, + 436 + ], + "type": "text", + "content": ", can simultaneously segment multiple objects (in multiple frames) using a single 24GB GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "spans": [ + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "text", + "content": "Fast " + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "text", + "content": " shares visual features as well as vision-language features for all referred objects to make the network efficient, and decouples the shared features to make them instance-specific before the segmentation stage. Firstly, visual features " + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_v)" + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "text", + "content": " and language features " + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_w" + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s)" + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "text", + "content": " are extracted. Next, we associate " + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v" + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w" + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "text", + "content": " using multi-instance fusion rather than the previous SCF. Multi-instance fusion is built on the foundation of SCF, which is depicted in Fig. 3. The difference is that multi-instance fusion includes semantic fusion, which performs an elementwise add operation, after cross-attention to merge vision-language features of different expressions. The features after semantic fusion perform Hadamard product with " + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v" + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "text", + "content": " to generate the vision-language features " + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 45, + 436, + 279, + 586 + ], + "type": "text", + "content": " for all objects:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 101, + 591, + 279, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 591, + 279, + 619 + ], + "spans": [ + { + "bbox": [ + 101, + 591, + 279, + 619 + ], + "type": "interline_equation", + "content": "\\operatorname {S F} \\left(\\mathcal {F} _ {w}, \\mathcal {F} _ {v}\\right) = \\sum_ {i = 1} ^ {N} \\operatorname {A t t} \\left(\\mathcal {F} _ {w} ^ {i}, \\mathcal {F} _ {v}\\right) \\tag {4}", + "image_path": "ec415cf9e91fde94707bf83405d9c2d7a963e57893aa286510ced0b37a1ba3ac.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 639, + 279, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 639, + 279, + 651 + ], + "spans": [ + { + "bbox": [ + 67, + 639, + 279, + 651 + ], + "type": "interline_equation", + "content": "\\operatorname {M I F} \\left(\\mathcal {F} _ {w}, \\mathcal {F} _ {v}\\right) = \\operatorname {S A} \\left(\\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right) \\otimes \\operatorname {S F} \\left(\\mathcal {F} _ {w}, \\operatorname {S A} \\left(\\mathcal {F} _ {v}\\right)\\right)\\right) \\tag {5}", + "image_path": "c1b210c3ce8874ea61f16b8433c8d15bd7cfe0ec43d2646e091877cedae3844c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " denotes Hadamard product and " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " denotes the number of expressions. After vision-language fusion, we encode " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " using the transformer encoder to enrich its semantic information, and plug multi-instance decoupling to decouple features for each instance. Multi-instance decoupling employs " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " and cross-attention to decouple " + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{vl}" + }, + { + "bbox": [ + 45, + 660, + 279, + 730 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "spans": [ + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "text", + "content": "predict instance embeddings " + }, + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 296, + 106, + 530, + 165 + ], + "type": "text", + "content": " for different referred objects. These embeddings are then projected to CPKs to predict the patch masks. Thus, FAST SgMg shares features, which account for most of the computational overhead, for different expressions, making it efficient for referring segmentation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 172, + 495, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 172, + 495, + 184 + ], + "spans": [ + { + "bbox": [ + 297, + 172, + 495, + 184 + ], + "type": "text", + "content": "3.6. Instance Matching and Loss Functions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "spans": [ + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "type": "text", + "content": "Following [4, 57], we perform instance matching with " + }, + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "type": "text", + "content": " learnable instance queries to improve fault tolerance. These queries are projected to CPKs to predict " + }, + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "type": "text", + "content": " potential patch masks " + }, + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_P" + }, + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "type": "text", + "content": " for each expression. The Hungarian algorithm [22] is then adopted to select the best match based on the matching loss for training. During inference, we directly employ the predicted confidence scores " + }, + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 296, + 191, + 530, + 282 + ], + "type": "text", + "content": " to measure the instance queries and select the results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "spans": [ + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "type": "text", + "content": "We adopt the same training losses and weights as used in [57, 72] for a fair comparison. Specifically, we use Dice loss [26] and Focal loss [31] for patch mask " + }, + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_P" + }, + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "type": "text", + "content": " and optimized mask " + }, + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_O" + }, + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "type": "text", + "content": ", Focal loss [31] for confidence scores " + }, + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "type": "text", + "content": ", and L1 and GIoU [48] loss for bounding boxes " + }, + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 296, + 283, + 530, + 347 + ], + "type": "text", + "content": ". The final training loss functions are:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 356, + 530, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 356, + 530, + 367 + ], + "spans": [ + { + "bbox": [ + 310, + 356, + 530, + 367 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a i n}} = \\lambda_ {\\mathcal {M} _ {P}} \\mathcal {L} _ {\\mathcal {M} _ {P}} + \\lambda_ {\\mathcal {M} _ {O}} \\mathcal {L} _ {\\mathcal {M} _ {O}} + \\lambda_ {\\mathcal {B}} \\mathcal {L} _ {\\mathcal {B}} + \\lambda_ {\\mathcal {S}} \\mathcal {L} _ {\\mathcal {S}} \\tag {6}", + "image_path": "b29adcc416103e14b522aa5ec38d4370bf7a36e27d063d69b1c568eaf16104a6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 376, + 524, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 376, + 524, + 387 + ], + "spans": [ + { + "bbox": [ + 299, + 376, + 524, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 299, + 376, + 524, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 299, + 376, + 524, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 299, + 376, + 524, + 387 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 299, + 376, + 524, + 387 + ], + "type": "text", + "content": " are the loss term and weight, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 398, + 376, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 398, + 376, + 411 + ], + "spans": [ + { + "bbox": [ + 297, + 398, + 376, + 411 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 418, + 415, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 418, + 415, + 429 + ], + "spans": [ + { + "bbox": [ + 297, + 418, + 415, + 429 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 436, + 530, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 436, + 530, + 563 + ], + "spans": [ + { + "bbox": [ + 296, + 436, + 530, + 563 + ], + "type": "text", + "content": "Datasets. We evaluate " + }, + { + "bbox": [ + 296, + 436, + 530, + 563 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 436, + 530, + 563 + ], + "type": "text", + "content": " on four video benchmarks: Ref-YouTube-VOS [49], Ref-DAVIS17 [21], A2D-Sentences [17], and JHMDB-Sentences [20]. RefYouTube-VOS is currently the largest dataset for R-VOS, containing 3,978 videos with about " + }, + { + "bbox": [ + 296, + 436, + 530, + 563 + ], + "type": "inline_equation", + "content": "13\\mathrm{K}" + }, + { + "bbox": [ + 296, + 436, + 530, + 563 + ], + "type": "text", + "content": " expressions. RefDAVIS17 is an extension of DAVIS17 [45] by including the language expressions of different objects and contains 90 videos. A2D-Sentences is a general actor and action segmentation dataset with over 3.7K videos and 6.6K action descriptions. JHMDB-Sentences includes 928 videos and 928 descriptions covering 21 different action classes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 564, + 530, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 564, + 530, + 634 + ], + "spans": [ + { + "bbox": [ + 296, + 564, + 530, + 634 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt the standard metrics to evaluate our models: region similarity " + }, + { + "bbox": [ + 296, + 564, + 530, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 296, + 564, + 530, + 634 + ], + "type": "text", + "content": " (average IoU), contour accuracy " + }, + { + "bbox": [ + 296, + 564, + 530, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 296, + 564, + 530, + 634 + ], + "type": "text", + "content": " (average boundary similarity), and their mean value " + }, + { + "bbox": [ + 296, + 564, + 530, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\& \\mathcal{F}" + }, + { + "bbox": [ + 296, + 564, + 530, + 634 + ], + "type": "text", + "content": ". All results are evaluated using the official code or server. On A2D-Sentences and JHMDB-Sentences, we adopt mAP, overall IoU, and mean IoU for evaluation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 297, + 642, + 426, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 642, + 426, + 654 + ], + "spans": [ + { + "bbox": [ + 297, + 642, + 426, + 654 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "content": "Following [4, 6, 57], we train our models on the training set of Ref-YouTube-VOS, and directly evaluate them on the validation split of Ref-YouTube-VOS and Ref-DAVIS17 without any additional techniques, e.g., model ensemble, joint training, and mask propagation, since they are not the focus of this paper. Additionally, we present results for our" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "924" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 104, + 493, + 254 + ], + "blocks": [ + { + "bbox": [ + 83, + 104, + 493, + 254 + ], + "lines": [ + { + "bbox": [ + 83, + 104, + 493, + 254 + ], + "spans": [ + { + "bbox": [ + 83, + 104, + 493, + 254 + ], + "type": "table", + "html": "
MethodYearBackboneRef-YouTube-VOSRef-DAVIS17
J&FJFFPSJ&FJF
CMSA [63]2019ResNet-5036.434.838.1-40.236.943.5
URVOS [49]2020ResNet-5047.245.349.2-51.547.356.0
CMPC-V [32]2021I3D47.545.649.3----
PMINet [11]2021ResNeSt-10153.051.554.5----
YOFO [23]2022ResNet-5048.647.549.71053.348.857.8
LBDT [10]2022ResNet-5049.448.250.6-54.3--
MLRL [56]2022ResNet-5049.748.451.0-52.850.055.4
MTTR [4]2022Video-Swin-T55.354.056.6----
MANet [6]2022Video-Swin-T55.654.856.5----
ReferFormer [57]2022Video-Swin-T56.054.857.350---
SgMg (Ours)2023Video-Swin-T58.957.760.06556.753.360.0
", + "image_path": "03b0aa0f0b4461880ac4be5591110fa1d089e948417be1ed8e301ada90f11e57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 84, + 268, + 493, + 316 + ], + "blocks": [ + { + "bbox": [ + 92, + 256, + 274, + 266 + ], + "lines": [ + { + "bbox": [ + 92, + 256, + 274, + 266 + ], + "spans": [ + { + "bbox": [ + 92, + 256, + 274, + 266 + ], + "type": "text", + "content": "Pre-training with RefCOCO/+/g & larger backbone" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 84, + 268, + 493, + 316 + ], + "lines": [ + { + "bbox": [ + 84, + 268, + 493, + 316 + ], + "spans": [ + { + "bbox": [ + 84, + 268, + 493, + 316 + ], + "type": "table", + "html": "
ReferFormer [57]2022Video-Swin-T59.458.060.95059.656.562.7
SgMg (Ours)2023Video-Swin-T62.060.463.56561.959.064.8
ReferFormer [57]2022Video-Swin-B62.961.364.63361.158.164.1
SgMg (Ours)2023Video-Swin-B65.763.967.44063.360.666.0
", + "image_path": "dc3f028f9e0cd45daf190f6ec9ff502797e6f61ff60845ec66307fa5c105045e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 83, + 336, + 495, + 476 + ], + "blocks": [ + { + "bbox": [ + 68, + 317, + 505, + 327 + ], + "lines": [ + { + "bbox": [ + 68, + 317, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 68, + 317, + 505, + 327 + ], + "type": "text", + "content": "Table 2. Quantitative comparison to state-of-the-art methods on the validation split of Ref-YouTube-VOS and Ref-DAVIS17." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 336, + 495, + 476 + ], + "lines": [ + { + "bbox": [ + 83, + 336, + 495, + 476 + ], + "spans": [ + { + "bbox": [ + 83, + 336, + 495, + 476 + ], + "type": "table", + "html": "
MethodBackboneA2D-SentencesJHMDB-Sentences
mAPOverall IoUMean IoUmAPOverall IoUMean IoU
Hu et al. [19]VGG-1613.247.435.017.854.652.8
Gavrilyuk et al. [17]I3D19.853.642.123.354.154.2
ACAN [54]I3D27.460.149.028.957.658.4
CMPC-V [32]I3D40.465.357.334.261.661.7
ClawCraneNet [29]ResNet-50/101-63.159.9-64.465.6
MTTR [4]Video-Swin-T46.172.064.039.270.169.8
ReferFormer [57]Video-Swin-T52.877.669.642.271.971.0
SgMg (Ours)Video-Swin-T56.178.070.444.472.871.7
ReferFormer [57]Video-Swin-B55.078.670.343.773.071.8
SgMg (Ours)Video-Swin-B58.579.972.045.073.772.5
", + "image_path": "ba07cbfefe8f697b403953944b94e99ba3625f1163306b85755c2c0feee15f43.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 477, + 485, + 486 + ], + "lines": [ + { + "bbox": [ + 90, + 477, + 485, + 486 + ], + "spans": [ + { + "bbox": [ + 90, + 477, + 485, + 486 + ], + "type": "text", + "content": "Table 3. Quantitative comparison to state-of-the-art R-VOS methods on A2D-Sentences and JHMDB-Sentences." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "spans": [ + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "text", + "content": "models first pre-trained on RefCOCO+/g [39, 65] and then fine-tuned on Ref-YouTube-VOS. Similar to [57, 72], we set the coefficients for different losses " + }, + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "inline_equation", + "content": "\\lambda_{focal}" + }, + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "inline_equation", + "content": "\\lambda_{L1}" + }, + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "inline_equation", + "content": "\\lambda_{giou}" + }, + { + "bbox": [ + 45, + 506, + 279, + 599 + ], + "type": "text", + "content": " to 5, 2, 5, 2, respectively. The models are trained using 2 RTX 3090 GPUs with 5 frames per clip for 9 epochs. All frames are resized to have the longest side of 640 pixels. Further implementation details are in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 607, + 161, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 607, + 161, + 618 + ], + "spans": [ + { + "bbox": [ + 45, + 607, + 161, + 618 + ], + "type": "text", + "content": "4.3.Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": "Ref-YouTube-VOS and Ref-DAVIS17. We compare " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " with recently published works in Table 2. Our approach surpasses present solutions on the two datasets across all metrics. On Ref-YouTube-VOS, " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " with the Video Swin Tiny backbone achieves " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "58.9\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " at 65 FPS, which is " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " higher and " + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "inline_equation", + "content": "1.3\\times" + }, + { + "bbox": [ + 45, + 624, + 279, + 730 + ], + "type": "text", + "content": " faster than the previous state-of-the-art ReferFormer [57]. Our approach runs faster due to the use of the segment-and-optimize pipeline, which avoids the need for heavy feature decoders. When pre" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "spans": [ + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "text", + "content": "training with RefCOCO/+/g and using a larger backbone, i.e., Video Swin Base, the performance of " + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "text", + "content": " further boosts to 65.7 " + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "text", + "content": ", consistently leading all other solutions by more than " + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "text", + "content": ". On Ref-DAVIS17, " + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "text", + "content": " achieves 63.3 " + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "text", + "content": ", outperforming state-of-the-art by " + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 296, + 506, + 530, + 576 + ], + "type": "text", + "content": " and demonstrating the generality of our approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "spans": [ + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "type": "text", + "content": "A2D-Sentences and JHMDB-Sentences. We further evaluate " + }, + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "type": "text", + "content": " on A2D-Sentences and JHMDB-Sentences in Table 3. Following [57], the models are first pre-trained on RefCOCO+/g and then fine-tuned on A2D-Sentences. JHMDB-Sentences is used only for evaluation. As shown in Table 3, " + }, + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "type": "text", + "content": " achieves superior performance compared to other state-of-the-art R-VOS methods and surpasses the nearest competitor Referformer [57] by " + }, + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "type": "inline_equation", + "content": "3.5 / 1.3\\%" + }, + { + "bbox": [ + 296, + 577, + 530, + 682 + ], + "type": "text", + "content": " mAP on A2D-Sentences and JHMDB-Sentences, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": "Multi-object R-VOS. We extend " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " to perform multi-object R-VOS, which is more practical and efficient for deployment. Fast " + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 682, + 530, + 730 + ], + "type": "text", + "content": " is trained on Ref-YouTube-VOS without pre-training or postprocessing techniques. We" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "925" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 104, + 280, + 157 + ], + "blocks": [ + { + "bbox": [ + 47, + 104, + 280, + 157 + ], + "lines": [ + { + "bbox": [ + 47, + 104, + 280, + 157 + ], + "spans": [ + { + "bbox": [ + 47, + 104, + 280, + 157 + ], + "type": "table", + "html": "
MethodRef-DAVIS17Ref-YouTube-VOS
J&FJFJ&FJFFPS
ReferFormer [57]54.551.058.056.054.857.350
Fast SgMg (Ours)54.251.157.354.253.155.3185
", + "image_path": "e0c167ebe36658d6be154ac5022a6ac627ec2e510bbd1b774c6ad8f4a3677b38.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 184, + 264, + 285 + ], + "blocks": [ + { + "bbox": [ + 47, + 157, + 280, + 178 + ], + "lines": [ + { + "bbox": [ + 47, + 157, + 280, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 157, + 280, + 178 + ], + "type": "text", + "content": "Table 4. Evaluation of Fast SgMg on Ref-DAVIS17 and RefYouTube-VOS. Video-Swin-T is adopted as the backbone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 184, + 264, + 285 + ], + "lines": [ + { + "bbox": [ + 61, + 184, + 264, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 264, + 285 + ], + "type": "table", + "html": "
ComponentsPerformance
CPKMSOSCFJ&FJFFPS
54.452.756.270
✓55.854.557.170
✓57.756.359.169
✓55.854.357.466
✓✓57.956.759.169
✓✓✓58.957.760.065
", + "image_path": "f7841f11983ce39fe5d13c2a03de7d9150de07f0859b010f5cb900a698aa2b78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 309, + 279, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 309, + 279, + 355 + ], + "spans": [ + { + "bbox": [ + 45, + 309, + 279, + 355 + ], + "type": "text", + "content": "benchmark Fast SgMg on Ref-YouTube-VOS and RefDAVIS17 using the commonly used Video Swin Tiny, and compare the results with the state-of-the-art R-VOS method, which performs single-object segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 356, + 279, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 356, + 279, + 472 + ], + "spans": [ + { + "bbox": [ + 45, + 356, + 279, + 472 + ], + "type": "text", + "content": "As shown in Table 4, Fast " + }, + { + "bbox": [ + 45, + 356, + 279, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 356, + 279, + 472 + ], + "type": "text", + "content": " achieves reasonable performance and runs about " + }, + { + "bbox": [ + 45, + 356, + 279, + 472 + ], + "type": "inline_equation", + "content": "3.7\\times" + }, + { + "bbox": [ + 45, + 356, + 279, + 472 + ], + "type": "text", + "content": " faster (185 vs 50 FPS) compared to ReferFormer [57]. It should be noted that each object in the above datasets contains multiple expressions. On Ref-DAVIS17, we group expressions to have only one expression per object within each group and segment all expressions in each group simultaneously since the object identity is given. On Ref-YouTube-VOS, all expressions in a video are segmented simultaneously due to the lack of object identity, making it more challenging." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 479, + 257, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 479, + 257, + 491 + ], + "spans": [ + { + "bbox": [ + 45, + 479, + 257, + 491 + ], + "type": "text", + "content": "4.4. Ablation Study for Different Components" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 497, + 279, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 497, + 279, + 532 + ], + "spans": [ + { + "bbox": [ + 45, + 497, + 279, + 532 + ], + "type": "text", + "content": "We conduct ablation experiments to evaluate the effectiveness of different components in " + }, + { + "bbox": [ + 45, + 497, + 279, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 497, + 279, + 532 + ], + "type": "text", + "content": ". The components are added to the baseline model step-by-step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 532, + 279, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 532, + 279, + 590 + ], + "spans": [ + { + "bbox": [ + 45, + 532, + 279, + 590 + ], + "type": "text", + "content": "Conditional Patch Kernel. As shown in Table 5, CPK boosts the performance by " + }, + { + "bbox": [ + 45, + 532, + 279, + 590 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 45, + 532, + 279, + 590 + ], + "type": "text", + "content": " compared with the recent instance-aware conditional kernels [57]. The sequential labels of each token predicted by CPK contain more fine-grained information, making the prediction more accurate." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": "Multi-granularity Segmentation Optimizer. We devise MSO to optimize the predicted patch masks. As shown in Table 5, MSO improves the performance by " + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": ", indicating the importance of fine-grained visual details in R-VOS. Spectrum-guided Cross-modal Fusion. We present SCF to perform global interactions by operating in the spectral domain. In Table 5, using SCF to replace the traditional cross-attention in [57, 52] improves the " + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 45, + 590, + 279, + 730 + ], + "type": "text", + "content": ". We consider SCF extracts important low-frequency features and facilitates multimodal understanding globally, which is suitable for R-VOS since locating referred objects requires understanding the global context and token relations." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 298, + 104, + 530, + 151 + ], + "blocks": [ + { + "bbox": [ + 47, + 286, + 277, + 296 + ], + "lines": [ + { + "bbox": [ + 47, + 286, + 277, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 277, + 296 + ], + "type": "text", + "content": "Table 5. Ablation of different components on Ref-YouTube-VOS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 298, + 104, + 530, + 151 + ], + "lines": [ + { + "bbox": [ + 298, + 104, + 530, + 151 + ], + "spans": [ + { + "bbox": [ + 298, + 104, + 530, + 151 + ], + "type": "table", + "html": "
SettingsDriftPipelineJ&FFPS
Baseline + Decoder✓decode-and-segment56.050
Baseline + Decoder + MSO✓decode-and-segment56.449
Baseline + MSO×segment-and-optimize57.769
", + "image_path": "6e789f9b848a06b92e7859a139cc944de2de6a5d7ebe0aca650919001541463d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 306, + 215, + 523, + 300 + ], + "blocks": [ + { + "bbox": [ + 297, + 152, + 530, + 205 + ], + "lines": [ + { + "bbox": [ + 297, + 152, + 530, + 205 + ], + "spans": [ + { + "bbox": [ + 297, + 152, + 530, + 205 + ], + "type": "text", + "content": "Table 6. Feature drift analysis using ReferFormer [57] (Baseline + Decoder) and " + }, + { + "bbox": [ + 297, + 152, + 530, + 205 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 297, + 152, + 530, + 205 + ], + "type": "text", + "content": " w/o CPK & SCF (Baseline + MSO). Significant improvement is achieved by addressing the drift issue (last row). Adding MSO on top of ReferFormer to recover visual details (for a second time) still performs worse than our basic pipeline." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 215, + 523, + 300 + ], + "lines": [ + { + "bbox": [ + 306, + 215, + 523, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 215, + 523, + 300 + ], + "type": "table", + "html": "
MethodRefCOCORefCOCO+RefCOCOg
MaIL [28]70.162.262.5
CRIS [55]70.562.359.9
RefTR [25]70.6--
LAVT [62]72.762.161.2
VLT [9]73.063.563.5
SgMg (Ours)76.366.470.0
", + "image_path": "cce2e01bab5d357ae545be79411791a1336fc3e6319b6a1d1a33c0b56865542e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 301, + 530, + 323 + ], + "lines": [ + { + "bbox": [ + 297, + 301, + 530, + 323 + ], + "spans": [ + { + "bbox": [ + 297, + 301, + 530, + 323 + ], + "type": "text", + "content": "Table 7. Quantitative evaluation on the validation split of Ref-COCO/+/g. Overall IoU is adopted as the evaluation metric." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 297, + 342, + 467, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 342, + 467, + 354 + ], + "spans": [ + { + "bbox": [ + 297, + 342, + 467, + 354 + ], + "type": "text", + "content": "4.5. Ablation Study for Feature Drift" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 361, + 530, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 361, + 530, + 465 + ], + "spans": [ + { + "bbox": [ + 296, + 361, + 530, + 465 + ], + "type": "text", + "content": "We conduct ablation study in Table 6 to demonstrate the feature drift problem. Our segment-and-optimize pipeline addresses the adverse drift effect discussed in Section 3.1 to significantly outperform ReferFormer [57] by " + }, + { + "bbox": [ + 296, + 361, + 530, + 465 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 296, + 361, + 530, + 465 + ], + "type": "text", + "content": " points and runs " + }, + { + "bbox": [ + 296, + 361, + 530, + 465 + ], + "type": "inline_equation", + "content": "1.4 \\times" + }, + { + "bbox": [ + 296, + 361, + 530, + 465 + ], + "type": "text", + "content": " faster. Furthermore, adding MSO on top of ReferFormer still performs worse due to the negative drift impact caused by the decode-and-segment pipeline. These results demonstrate the efficacy of our proposed segment-and-optimize pipeline." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 474, + 495, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 474, + 495, + 487 + ], + "spans": [ + { + "bbox": [ + 297, + 474, + 495, + 487 + ], + "type": "text", + "content": "4.6. Referring Image Segmentation Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "spans": [ + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "text", + "content": "We apply " + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "text", + "content": " to referring image (expression) segmentation without any architectural modifications, and compare against the current state-of-the-art methods on RefCOCO " + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "inline_equation", + "content": "+ / g" + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "text", + "content": " [39, 65]. A single " + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "text", + "content": " model is trained on RefCOCO " + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "inline_equation", + "content": "+ / g" + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "text", + "content": " without large-scale pre-training. As shown in Table 7, " + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "text", + "content": " achieves advanced performance on all three benchmarks. These results demonstrate the efficacy of " + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 493, + 530, + 586 + ], + "type": "text", + "content": " in referring image segmentation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 595, + 530, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 595, + 530, + 607 + ], + "spans": [ + { + "bbox": [ + 297, + 595, + 530, + 607 + ], + "type": "text", + "content": "4.7. Inference Time Analysis of Multi-Object RVOS" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": "We analyze the efficiency of the proposed multi-object R-VOS paradigm by comparing the FPS of Fast " + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": " on videos with different numbers of expressions. As illustrated in Fig. 6, Fast " + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": " performs about " + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": " faster than " + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": " when there are two expressions per video on average. As the number of expressions increases, Fast " + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": " achieves faster reasoning time per object per frame due to its utilization of the multi-object R-VOS paradigm. When there are ten expressions in each video, Fast " + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": " performs at nearly 300 FPS, which is about " + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "inline_equation", + "content": "5\\times" + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": " faster than " + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 612, + 530, + 730 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "926" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 117, + 516, + 318 + ], + "blocks": [ + { + "bbox": [ + 74, + 105, + 230, + 116 + ], + "lines": [ + { + "bbox": [ + 74, + 105, + 230, + 116 + ], + "spans": [ + { + "bbox": [ + 74, + 105, + 230, + 116 + ], + "type": "text", + "content": "Expression: a turtle to the bottom left." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 117, + 516, + 318 + ], + "lines": [ + { + "bbox": [ + 58, + 117, + 516, + 318 + ], + "spans": [ + { + "bbox": [ + 58, + 117, + 516, + 318 + ], + "type": "image", + "image_path": "cbcb32da13ed3ec608c9c4aada756dbe130df296a2ba08aa0c6443ae4e1029e8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 319, + 394, + 330 + ], + "lines": [ + { + "bbox": [ + 181, + 319, + 394, + 330 + ], + "spans": [ + { + "bbox": [ + 181, + 319, + 394, + 330 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of our method with others." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 344, + 277, + 459 + ], + "blocks": [ + { + "bbox": [ + 49, + 344, + 277, + 459 + ], + "lines": [ + { + "bbox": [ + 49, + 344, + 277, + 459 + ], + "spans": [ + { + "bbox": [ + 49, + 344, + 277, + 459 + ], + "type": "image", + "image_path": "55fe67df66eb4d6cc68ac6b33fcae467955bcb6523e69cd99e50525433d96014.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 463, + 279, + 485 + ], + "lines": [ + { + "bbox": [ + 45, + 463, + 279, + 485 + ], + "spans": [ + { + "bbox": [ + 45, + 463, + 279, + 485 + ], + "type": "text", + "content": "Figure 6. Efficiency analysis of " + }, + { + "bbox": [ + 45, + 463, + 279, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 463, + 279, + 485 + ], + "type": "text", + "content": " and Fast " + }, + { + "bbox": [ + 45, + 463, + 279, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 463, + 279, + 485 + ], + "type": "text", + "content": " for videos with different numbers of expressions on Ref-YouTube-VOS." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 503, + 155, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 155, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 155, + 516 + ], + "type": "text", + "content": "4.8. Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 521, + 279, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 521, + 279, + 557 + ], + "spans": [ + { + "bbox": [ + 45, + 521, + 279, + 557 + ], + "type": "text", + "content": "In Fig. 5, we show qualitative comparison with ReferFormer [57] and MTTR [4]. " + }, + { + "bbox": [ + 45, + 521, + 279, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 521, + 279, + 557 + ], + "type": "text", + "content": " can handle different objects of the same category or with the same behavior." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 564, + 200, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 200, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 200, + 576 + ], + "type": "text", + "content": "4.9. Feature Visualization of SCF" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 582, + 280, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 582, + 280, + 641 + ], + "spans": [ + { + "bbox": [ + 45, + 582, + 280, + 641 + ], + "type": "text", + "content": "In Fig. 7, we visualize the vision-language features extracted by our SCF in comparison to the cross-attention used in [57]. The features extracted by SCF exhibit superior grounding ability in locating target objects, resulting in better performance for " + }, + { + "bbox": [ + 45, + 582, + 280, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 582, + 280, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 651, + 117, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 117, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 117, + 664 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "text", + "content": "We discovered the feature drift issue in current referring video object segmentation (R-VOS) methods, which negatively affects the segmentation kernels. We presented " + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "text", + "content": ", a novel segment-and-optimize approach for R-VOS that avoids the drift issue and optimizes masks with visual" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 301, + 344, + 530, + 516 + ], + "blocks": [ + { + "bbox": [ + 301, + 344, + 530, + 516 + ], + "lines": [ + { + "bbox": [ + 301, + 344, + 530, + 516 + ], + "spans": [ + { + "bbox": [ + 301, + 344, + 530, + 516 + ], + "type": "image", + "image_path": "e457b08af0144c867c4fde393fa6bb8ccf2eada1d54b3835be531cac4aa387d0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 519, + 530, + 541 + ], + "lines": [ + { + "bbox": [ + 297, + 519, + 530, + 541 + ], + "spans": [ + { + "bbox": [ + 297, + 519, + 530, + 541 + ], + "type": "text", + "content": "Figure 7. Visualization of the vision-language features extracted w/o and w/ our SCF." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 555, + 530, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 555, + 530, + 650 + ], + "spans": [ + { + "bbox": [ + 296, + 555, + 530, + 650 + ], + "type": "text", + "content": "details. We also provided a new perspective to encourage vision-language global interactions in the spectral domain with Spectrum-guided Cross-modal Fusion. Additionally, we proposed the multi-object R-VOS paradigm by extending " + }, + { + "bbox": [ + 296, + 555, + 530, + 650 + ], + "type": "inline_equation", + "content": "\\mathrm{SgMg}" + }, + { + "bbox": [ + 296, + 555, + 530, + 650 + ], + "type": "text", + "content": " with multi-instance fusion and decoupling. Finally, we evaluated our models on four video benchmarks and demonstrated that our approach achieves state-of-the-art performance on all four datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 654, + 531, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 654, + 531, + 724 + ], + "spans": [ + { + "bbox": [ + 296, + 654, + 531, + 724 + ], + "type": "text", + "content": "Acknowledgment. This research was supported by the Australian Research Council Industrial Transformation Research Hub IH180100002. Professor Ajmal Mian is the recipient of an Australian Research Council Future Fellowship Award (project number FT210100268) funded by the Australian Government." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 746, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 746, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 746, + 296, + 756 + ], + "type": "text", + "content": "927" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 103, + 117 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 280, + 730 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 51, + 124, + 279, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 124, + 279, + 187 + ], + "spans": [ + { + "bbox": [ + 51, + 124, + 279, + 187 + ], + "type": "text", + "content": "[1] Nayyer Aafaq, Naveed Akhtar, Wei Liu, Syed Zulqarnain Gilani, and Ajmal Mian. Spatio-temporal dynamics and semantic attribute enriched visual encoding for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12487-12496, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 189, + 280, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 189, + 280, + 231 + ], + "spans": [ + { + "bbox": [ + 51, + 189, + 280, + 231 + ], + "type": "text", + "content": "[2] Miriam Bellver, Carles Ventura, Carina Silberer, Ioannis Kazakos, Jordi Torres, and Xavier Giro-i Nieto. A closer look at referring expressions for video object segmentation. Multimedia Tools and Applications, pages 1-20, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 233, + 280, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 233, + 280, + 254 + ], + "spans": [ + { + "bbox": [ + 51, + 233, + 280, + 254 + ], + "type": "text", + "content": "[3] Glenn D Bergland. A guided tour of the fast fourier transform. IEEE spectrum, 6(7):41-52, 1969. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 255, + 280, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 255, + 280, + 296 + ], + "spans": [ + { + "bbox": [ + 51, + 255, + 280, + 296 + ], + "type": "text", + "content": "[4] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In CVPR, pages 4985-4995, 2022. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 298, + 279, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 298, + 279, + 351 + ], + "spans": [ + { + "bbox": [ + 51, + 298, + 279, + 351 + ], + "type": "text", + "content": "[5] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 221-230, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 351, + 280, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 351, + 280, + 415 + ], + "spans": [ + { + "bbox": [ + 51, + 351, + 280, + 415 + ], + "type": "text", + "content": "[6] Weidong Chen, Dexiang Hong, Yuankai Qi, Zhenjun Han, Shuhui Wang, Laiyun Qing, Qingming Huang, and Guorong Li. Multi-attention network for compressed video referring object segmentation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4416–4425, 2022. 1, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 416, + 280, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 416, + 280, + 458 + ], + "spans": [ + { + "bbox": [ + 51, + 416, + 280, + 458 + ], + "type": "text", + "content": "[7] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. Proc. Adv. Neural Inf. Process. Syst. (NIPS), 34:11781-11794, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 460, + 279, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 460, + 279, + 490 + ], + "spans": [ + { + "bbox": [ + 51, + 460, + 279, + 490 + ], + "type": "text", + "content": "[8] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 492, + 279, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 492, + 279, + 534 + ], + "spans": [ + { + "bbox": [ + 51, + 492, + 279, + 534 + ], + "type": "text", + "content": "[9] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 536, + 279, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 279, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 279, + 588 + ], + "type": "text", + "content": "[10] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 590, + 279, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 279, + 642 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 279, + 642 + ], + "type": "text", + "content": "[11] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, page 7, 2021. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 644, + 279, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 279, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 279, + 675 + ], + "type": "text", + "content": "[12] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 676, + 279, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 676, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 676, + 279, + 730 + ], + "type": "text", + "content": "[13] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 530, + 730 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 299, + 106, + 530, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 106, + 530, + 138 + ], + "spans": [ + { + "bbox": [ + 299, + 106, + 530, + 138 + ], + "type": "text", + "content": "[14] Guang Feng, Lihe Zhang, Zhiwei Hu, and Hutchuan Lu. Deeply interleaved two-stream encoder for referring video segmentation. arXiv preprint arXiv:2203.15969, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 139, + 530, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 139, + 530, + 191 + ], + "spans": [ + { + "bbox": [ + 299, + 139, + 530, + 191 + ], + "type": "text", + "content": "[15] Mingtao Feng, Haoran Hou, Liang Zhang, Yulan Guo, Hongshan Yu, Yaonan Wang, and Ajmal Mian. Exploring hierarchical spatial layout cues for 3d point cloud based scene graph prediction. IEEE Transactions on Multimedia, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 193, + 530, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 193, + 530, + 245 + ], + "spans": [ + { + "bbox": [ + 299, + 193, + 530, + 245 + ], + "type": "text", + "content": "[16] Zhiheng Fu, Siyu Hong, Mengyi Liu, Hamid Laga, Mohammed Bennamoun, Farid Boussaid, and Yulan Guo. Multi-stage information diffusion for joint depth and surface normal estimation. Pattern Recognition, 141:109660, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 246, + 530, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 246, + 530, + 289 + ], + "spans": [ + { + "bbox": [ + 299, + 246, + 530, + 289 + ], + "type": "text", + "content": "[17] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 290, + 530, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 290, + 530, + 330 + ], + "spans": [ + { + "bbox": [ + 299, + 290, + 530, + 330 + ], + "type": "text", + "content": "[18] Lorenzo Giambagli, Lorenzo Buffoni, Timoteo Carletti, Walter Nocentini, and Duccio Fanelli. Machine learning in spectral domain. Nature communications, 12(1):1-9, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 332, + 530, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 332, + 530, + 373 + ], + "spans": [ + { + "bbox": [ + 299, + 332, + 530, + 373 + ], + "type": "text", + "content": "[19] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In European Conference on Computer Vision, pages 108-124. Springer, 2016. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 375, + 530, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 375, + 530, + 427 + ], + "spans": [ + { + "bbox": [ + 299, + 375, + 530, + 427 + ], + "type": "text", + "content": "[20] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 429, + 530, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 429, + 530, + 471 + ], + "spans": [ + { + "bbox": [ + 299, + 429, + 530, + 471 + ], + "type": "text", + "content": "[21] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Asian Conference on Computer Vision, pages 123-141. Springer, 2018. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 472, + 530, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 472, + 530, + 502 + ], + "spans": [ + { + "bbox": [ + 299, + 472, + 530, + 502 + ], + "type": "text", + "content": "[22] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 504, + 530, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 504, + 530, + 556 + ], + "spans": [ + { + "bbox": [ + 299, + 504, + 530, + 556 + ], + "type": "text", + "content": "[23] Dezhuang Li, Ruoqi Li, Lijun Wang, Yifan Wang, Jinqing Qi, Lu Zhang, Ting Liu, Qingquan Xu, and Huchuan Lu. You only infer once: Cross-modal meta-transfer for referring video object segmentation. In AAAI Conference on Artificial Intelligence, 2022. 1, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 558, + 530, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 558, + 530, + 611 + ], + "spans": [ + { + "bbox": [ + 299, + 558, + 530, + 611 + ], + "type": "text", + "content": "[24] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 612, + 530, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 612, + 530, + 652 + ], + "spans": [ + { + "bbox": [ + 299, + 612, + 530, + 652 + ], + "type": "text", + "content": "[25] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. Advances in neural information processing systems, 34:19652-19664, 2021.7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 655, + 530, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 655, + 530, + 686 + ], + "spans": [ + { + "bbox": [ + 299, + 655, + 530, + 686 + ], + "type": "text", + "content": "[26] Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855, 2019. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 687, + 530, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 687, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 687, + 530, + 730 + ], + "type": "text", + "content": "[27] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Yan Lu, and Bhiksha Raj. R^2vos: Robust referring video object segmentation via relational multimodal cycle consistency. arXiv preprint arXiv:2207.01203, 2022. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "928" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 730 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 149 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 149 + ], + "type": "text", + "content": "[28] Zizhang Li, Mengmeng Wang, Jianbiao Mei, and Yong Liu. Mail: A unified mask-image-language trimodal network for referring image segmentation. arXiv preprint arXiv:2111.10747, 2021. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 151, + 279, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 151, + 279, + 183 + ], + "spans": [ + { + "bbox": [ + 47, + 151, + 279, + 183 + ], + "type": "text", + "content": "[29] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawcranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 184, + 279, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 184, + 279, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 279, + 237 + ], + "type": "text", + "content": "[30] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 239, + 279, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 279, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 279, + 281 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 282, + 279, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 279, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 279, + 325 + ], + "type": "text", + "content": "[32] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 326, + 279, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 279, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 279, + 378 + ], + "type": "text", + "content": "[33] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 381, + 279, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 279, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 279, + 433 + ], + "type": "text", + "content": "[34] Yong Liu, Ran Yu, Jiahao Wang, Xinyuan Zhao, Yitong Wang, Yansong Tang, and Yujiu Yang. Global spectral filter memory network for video object segmentation. In European Conference on Computer Vision, pages 648-665. Springer, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 435, + 279, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 279, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 279, + 478 + ], + "type": "text", + "content": "[35] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 479, + 279, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 279, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 279, + 522 + ], + "type": "text", + "content": "[36] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 523, + 279, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 279, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 279, + 576 + ], + "type": "text", + "content": "[37] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10437–10446, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 577, + 279, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 577, + 279, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 577, + 279, + 640 + ], + "type": "text", + "content": "[38] Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsupervised video object segmentation with co-attention siamese networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3623–3632, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 642, + 279, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 279, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 279, + 696 + ], + "type": "text", + "content": "[39] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 697, + 279, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 697, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 697, + 279, + 730 + ], + "type": "text", + "content": "[40] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 530, + 730 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 317, + 106, + 530, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 530, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 530, + 138 + ], + "type": "text", + "content": "localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 140, + 530, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 140, + 530, + 181 + ], + "spans": [ + { + "bbox": [ + 299, + 140, + 530, + 181 + ], + "type": "text", + "content": "[41] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Region aware video object segmentation with deep motion modeling. arXiv preprint arXiv:2207.10258, 2022.1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 184, + 530, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 184, + 530, + 226 + ], + "spans": [ + { + "bbox": [ + 299, + 184, + 530, + 226 + ], + "type": "text", + "content": "[42] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Self-supervised video object segmentation by motion-aware mask propagation. In Proceedings of the IEEE International Conference on Multimedia and Expo, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 228, + 530, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 228, + 530, + 270 + ], + "spans": [ + { + "bbox": [ + 299, + 228, + 530, + 270 + ], + "type": "text", + "content": "[43] Bo Miao, Liguang Zhou, Ajmal Mian, Tin Lun Lam, and Yangsheng Xu. Object-to-scene: Learning to transfer object knowledge to indoor scene recognition. In Proc. IEEE Int. Conf. Intell. Robots Syst. (IROS). IEEE, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 271, + 530, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 271, + 530, + 314 + ], + "spans": [ + { + "bbox": [ + 299, + 271, + 530, + 314 + ], + "type": "text", + "content": "[44] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proc. Int. Conf. Comput. Vis. (ICCV), pages 9226–9235, 2019. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 315, + 530, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 315, + 530, + 357 + ], + "spans": [ + { + "bbox": [ + 299, + 315, + 530, + 357 + ], + "type": "text", + "content": "[45] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 358, + 530, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 358, + 530, + 423 + ], + "spans": [ + { + "bbox": [ + 299, + 358, + 530, + 423 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 424, + 530, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 424, + 530, + 478 + ], + "spans": [ + { + "bbox": [ + 299, + 424, + 530, + 478 + ], + "type": "text", + "content": "[47] Sucheng Ren, Wenxi Liu, Yongtuo Liu, Haoxin Chen, Guoqiang Han, and Shengfeng He. Reciprocal transformations for unsupervised video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15455-15464, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 479, + 530, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 479, + 530, + 541 + ], + "spans": [ + { + "bbox": [ + 299, + 479, + 530, + 541 + ], + "type": "text", + "content": "[48] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 543, + 530, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 543, + 530, + 586 + ], + "spans": [ + { + "bbox": [ + 299, + 543, + 530, + 586 + ], + "type": "text", + "content": "[49] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In ECCV, pages 208–223. Springer, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 587, + 530, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 587, + 530, + 628 + ], + "spans": [ + { + "bbox": [ + 299, + 587, + 530, + 628 + ], + "type": "text", + "content": "[50] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In European conference on computer vision, pages 282-298. Springer, 2020. 1, 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 630, + 530, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 630, + 530, + 662 + ], + "spans": [ + { + "bbox": [ + 299, + 630, + 530, + 662 + ], + "type": "text", + "content": "[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 664, + 530, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 664, + 530, + 707 + ], + "spans": [ + { + "bbox": [ + 299, + 664, + 530, + 707 + ], + "type": "text", + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 708, + 530, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 708, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 299, + 708, + 530, + 730 + ], + "type": "text", + "content": "[53] Paul Voiglaender, Yuning Chai, Florian Schroff, Hartwig Adam, Bastian Leibe, and Liang-Chieh Chen. Feelvos: Fast" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "929" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 280, + 729 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 64, + 106, + 279, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 106, + 279, + 149 + ], + "spans": [ + { + "bbox": [ + 64, + 106, + 279, + 149 + ], + "type": "text", + "content": "end-to-end embedding learning for video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9481-9490, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 151, + 279, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 151, + 279, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 151, + 279, + 204 + ], + "type": "text", + "content": "[54] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 206, + 280, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 280, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 280, + 258 + ], + "type": "text", + "content": "[55] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 260, + 279, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 279, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 279, + 314 + ], + "type": "text", + "content": "[56] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 1, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 314, + 279, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 279, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 279, + 356 + ], + "type": "text", + "content": "[57] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4974-4984, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 358, + 279, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 279, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 279, + 399 + ], + "type": "text", + "content": "[58] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. arXiv preprint arXiv:2308.02874, 2023.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 402, + 279, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 279, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 279, + 444 + ], + "type": "text", + "content": "[59] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019.4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 446, + 279, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 446, + 279, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 446, + 279, + 488 + ], + "type": "text", + "content": "[60] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In International Conference on Neural Information Processing, pages 264-274. Springer, 2019. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 490, + 279, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 279, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 279, + 532 + ], + "type": "text", + "content": "[61] Zhao Yang, Yansong Tang, Luca Bertinetto, Hengshuang Zhao, and Philip HS Torr. Hierarchical interaction network for video object segmentation from referring expressions. In BMVC, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 533, + 279, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 279, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 279, + 586 + ], + "type": "text", + "content": "[62] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Ladv: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 588, + 279, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 279, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 279, + 640 + ], + "type": "text", + "content": "[63] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 642, + 279, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 279, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 279, + 685 + ], + "type": "text", + "content": "[64] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "type": "text", + "content": "[65] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer, 2016. 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 431 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "spans": [ + { + "bbox": [ + 299, + 107, + 530, + 138 + ], + "type": "text", + "content": "[66] Z. Yang and Y, Wei and Y. Yang. Associating objects with transformers for video object segmentation. Proc. Adv. Neural Inf. Process. Syst. (NIPS), 34:2491-2502, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 140, + 530, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 140, + 530, + 203 + ], + "spans": [ + { + "bbox": [ + 299, + 140, + 530, + 203 + ], + "type": "text", + "content": "[67] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 205, + 530, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 205, + 530, + 246 + ], + "spans": [ + { + "bbox": [ + 299, + 205, + 530, + 246 + ], + "type": "text", + "content": "[68] Shuai Zhao, Xiaohan Wang, Linchao Zhu, and Yi Yang. Test-time adaptation with clip reward for zero-shot generalization in vision-language models. arXiv preprint arXiv:2305.18010, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 249, + 530, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 249, + 530, + 280 + ], + "spans": [ + { + "bbox": [ + 299, + 249, + 530, + 280 + ], + "type": "text", + "content": "[69] Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. Centerclip: Token clustering for efficient text-video retrieval. arXiv preprint arXiv:2205.00823, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 282, + 530, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 282, + 530, + 335 + ], + "spans": [ + { + "bbox": [ + 299, + 282, + 530, + 335 + ], + "type": "text", + "content": "[70] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 335, + 530, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 335, + 530, + 389 + ], + "spans": [ + { + "bbox": [ + 299, + 335, + 530, + 389 + ], + "type": "text", + "content": "[71] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and vqa. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13041-13049, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 390, + 530, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 390, + 530, + 431 + ], + "spans": [ + { + "bbox": [ + 299, + 390, + 530, + 431 + ], + "type": "text", + "content": "[72] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 5, 6" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "930" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/b8d7dfde-aec2-4ac3-9ef3-b8789d6534b1_content_list.json b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/b8d7dfde-aec2-4ac3-9ef3-b8789d6534b1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4160b78fd366c3cd320f712a8efd0741d8798869 --- /dev/null +++ b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/b8d7dfde-aec2-4ac3-9ef3-b8789d6534b1_content_list.json @@ -0,0 +1,1491 @@ +[ + { + "type": "text", + "text": "Speech2Lip: High-fidelity Speech to Lip Generation by Learning from a Short Video", + "text_level": 1, + "bbox": [ + 91, + 128, + 880, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiuzhe $\\mathsf{W u}^{1}$ , Pengfei $\\mathsf{H u}^{2}$ , Yang $\\mathsf{W u}^{3,4}$ , Xiaoyang $\\mathsf{L y u}^{1}$ , Yan-Pei $\\mathsf{C a o}^{3}$ , Ying $\\mathsf{S h a n}^{3}$ , Wenming $\\mathsf{Y a n g}^{2}$ , Zhongqian $\\mathsf{S u n}^{4}$ , Xiaojuan $\\mathsf{Q i}^{1}$ ,", + "bbox": [ + 99, + 214, + 870, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong, 2Tsinghua University, 3ARC Lab, Tencent PCG, 4Tencent AI Lab", + "bbox": [ + 91, + 273, + 877, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synthesizing realistic videos according to a given speech is still an open challenge. Previous works have been plagued by issues such as inaccurate lip shape generation and poor image quality. The key reason is that only motions and appearances on limited facial areas (e.g., lip area) are mainly driven by the input speech. Therefore, directly learning a mapping function from speech to the entire head image is prone to ambiguity, particularly when using a short video for training. We thus propose a decomposition-synthesis-composition framework named Speech to Lip (Speech2Lip) that disentangles speech-sensitive and speech-insensitive motion/appearance to facilitate effective learning from limited training data, resulting in the generation of natural-looking videos. First, given a fixed head pose (i.e., canonical space), we present a speech-driven implicit model for lip image generation which concentrates on learning speech-sensitive motion and appearance. Next, to model the major speech-insensitive motion (i.e., head movement), we introduce a geometry-aware mutual explicit mapping (GAMEM) module that establishes geometric mappings between different head poses. This allows us to paste generated lip images at the canonical space onto head images with arbitrary poses and synthesize talking videos with natural head movements. In addition, a Blend-Net and a contrastive sync loss are introduced to enhance the overall synthesis performance. Quantitative and qualitative results on three benchmarks demonstrate that our model can be trained by a video of just a few minutes in length and achieve state-of-the-art performance in both visual quality and speech-visual synchronization. Code: https://github.com/CVMI-Lab/Speech2Lip.", + "bbox": [ + 75, + 415, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a6576cc4cc0f72ccf3597bdaafdedc96a8ea8445204707a4a2b246f1c96cfb8.jpg", + "image_caption": [ + "Figure 1. Given a speech as input, our model generates high-quality talking-head videos and supports pose-controllable synthesis. The decomposition and synthesis modules make learning from a short video more effective and the composition module enables us to synthesize high-fidelity videos." + ], + "image_footnote": [], + "bbox": [ + 501, + 338, + 893, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 583, + 632, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning from a short video to generate personalized audio-synchronized talking videos driven by a speech is of great importance to various applications, for instance, digital human animation, video dubbing, and UGC video creation. However, synthesizing high-fidelity videos from speech for a desired speaker remains a challenging task.", + "bbox": [ + 496, + 611, + 893, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The first challenge arises from complicated motion patterns. Although speech majorly influences lip areas (i.e., speech-sensitive areas), lip movements are often accompanied by other motions, such as global head movements, which greatly impact lip shapes and appearances. Thus, directly synthesizing a whole image from speech often leads to inaccurate lip synthesis. Second, existing methods still struggle to satisfy appearance fidelity requirements, which include both identity preservation (speaker-specific) and high-quality detail generation, such as clear details of teeth, tongue, and eye-blinking [9, 42, 26, 44, 7]. Third, it's difficult to acquire videos longer than 10 hours for a speaker which is yet required by conventional methods [17, 31].", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22168", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle the aforementioned challenges, existing attempts can be coarsely categorized into two lines of research: speaker-independent and speaker-specific methods. The first line often exploits GANs [15] that need to be trained on large-scale multi-person datasets. However, GAN-based models [4, 7, 9, 26, 33, 38, 42, 43] usually synthesize low-resolution images and unnatural motions (i.e., background movements). They thus hardly meet the appearance fidelity requirement in terms of sharpness and fine appearance details. Moreover, preserving the identity of the speaker remains a challenging task [7, 43].", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the consideration of high-fidelity and identity preservation, another strategy focuses on learning from a specific speaker. Although attaining high-fidelity results, early works [17, 31] often require several hours of video footage from a speaker for training, hindering their practical applicability. Recently, NeRF [20] has emerged as a promising approach for generating high-fidelity videos, which succeeds in learning from a short video of just a few minutes and having the potential to generate high-fidelity results [16, 18, 28]. Nevertheless, the models still struggle with appearance and motion ambiguity issues because they model speech-sensitive motions/appearances together with other facial areas less correlated to the given speech. This issue becomes more severe when training data is limited since no extra information can be exploited to avoid interference from signals that are not correlated with the speech. As a result, they tend to generate lip sequences that do not synchronize well with the speech and produce blurry images (see Figure 5 and Table 1). Therefore, reducing the complexity of modeling motions is critical to enable effective learning from a short video and synthesizing high-quality videos for a specific speaker.", + "bbox": [ + 75, + 261, + 472, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We thus design a preliminary experiment to identify that motion and appearance of lip areas have a strong correlation with speech, while head motion and other facial areas are less related to speech (Figure ??). Motivated by the observation, we propose decomposing speech-insensitive motion/appearance from speech-sensitive one, and synthesizing them separately before composing them into a new talking video that aligns with the given speech (Figure 1). Toward this goal, we present a decomposition-synthesis-composition framework named Speech to Lip (Speech2Lip). In the decomposition stage, we introduce a speech-driven implicit model that generates high-fidelity synced lip sequences in a fixed view (i.e., canonical view). To model 3D head motion effectively, we design a Geometry-Aware Mutual Explicit Mapping (GAMEM) module that estimates explicit geometric mappings between an arbitrary observed view and the canonical view. GAMEM also includes a jointly optimized canonical-view full-head depth map, which enables the model to be 3D-aware and supports controllable synthesis driven by new", + "bbox": [ + 75, + 598, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "head poses. In the composition stage, GAMEM allows us to flexibly paste the synthesized canonical-view lips onto an arbitrary observed view to obtain natural synchronized talking videos. To improve the synthesis and synchronization qualities after composition, we incorporate a blending network (i.e., Blend-Net) to refine the results and a contrastive sync loss to facilitate learning from a short video for generating synchronized talking videos.", + "bbox": [ + 496, + 90, + 893, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our major contributions are summarized as follows:", + "bbox": [ + 517, + 212, + 864, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We introduce a novel framework that disentangles speech-sensitive and speech-insensitive motion/appearance in high-fidelity video synthesis. By separating these components, the framework can effectively learn from limited training data.", + "2) The proposed speech-driven implicit model synthesizes speech-sensitive contents and the GAMEM flexibly combines them with given speech-insensitive areas to generate synchronized talking heads with natural movements and support pose-controllable synthesis.", + "3) Both qualitative and quantitative experimental results demonstrate the superiority of our method over the state-of-the-art speaker-specific methods." + ], + "bbox": [ + 509, + 232, + 890, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 460, + 640, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Speech-Driven Talking Face Synthesis. Video synthesis from speech is a long-standing problem. Recent works can be generally divided into two categories: speaker-specific [17, 31, 19, 16, 39, 18, 28] and speaker-independent [9, 42, 26, 44, 7, 5, 41, 35, 36]. In the first track, earlier works have succeeded in obtaining realistic visual results for a target person [17, 31] but required hours of video belonging to one specific speaker. Therefore, many efforts [19, 16, 39, 18, 28] have been made to train the model with a shorter talking video (3-5 minutes). However, they utilize speech to forecast overall motions including speech-insensitive ones, failing to learn accurate lip shapes and appearances. The other line (i.e., speaker-independent method) aims to build a universal model for all identities [9, 42, 26, 44, 7]. The end-to-end pipeline [9, 42] is usually developed on GAN [15]. To boost performance, Richard et al. [27] only synthesize the mesh of limited facial areas, Prajwal et al. [26] present an extra lip-sync discriminator, Zhou et al. [43] provide additional head poses as inputs, and some other models [7, 14, 44, 32, 30] leverage the intermediate representation (e.g., 2D facial landmarks [7, 14, 44] or expression parameters [32, 30]). Nevertheless, they often suffer from low video quality, difficult identity preservation, and abnormal head motion generation.", + "bbox": [ + 496, + 489, + 893, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit Representation based Talking Head Methods. Recently, implicit representations have shown high model", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22169", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing capabilities in multiple tasks [25, 20, 24, 34]. Among them, NeRF [20] obtains extraordinary performance in novel view synthesis by training on only hundreds of images. Guo et al. [16] first apply it in the speaker-specific talking head synthesis task to enable learning from a short video. However, they utilize two NeRFs to model the head and torso, resulting in the head-torso separation phenomenon. Liu et al. [18] solve it by leveraging one unified NeRF with two semantic-aware modules. Shen et al. [28] incorporate 2D image features as additional inputs to further reduce training data requirements (i.e., videos with only 10-15 seconds). Regardless, these methods only ensure visual quality, and simply use speech to drive all the complex motions, leading to the out-of-sync problem.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Empirical Study and Motivations", + "text_level": 1, + "bbox": [ + 76, + 314, + 382, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach is motivated by the observation that only limited facial areas are highly correlated to speech. To verify this, we conduct a preliminary experiment to determine which areas are most speech-sensitive. Specifically, we apply warping to all captured images with varying head poses (refer to observed views) using the 3D Morphable Model (3DMM) [2] and bring them to a fixed head pose (known as the canonical view). We then compute a motion heatmap in the canonical view to determine the areas most responsive to speech. As depicted in Figure 2, the elimination of head motion results in most of the motions around the lip regions displaying high sensitivity to the input speech. Additional examples can be found in the supplementary file. For more details on the warping process, please refer to Sec. 4.3.", + "bbox": [ + 75, + 339, + 473, + 551 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/af6ab167b3373f3bbf76d9a7778a725e22eae26d77ad4220fe57fab0f27e251a.jpg", + "image_caption": [ + "Figure 2. Speech-sensitive Motion Heatmap." + ], + "image_footnote": [], + "bbox": [ + 114, + 561, + 429, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 656, + 168, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 681, + 187, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 3 provides an overview of our model. We define each frame captured in the observed view with its own head motion as the \"observed space\". To disentangle speech-sensitive and speech-insensitive motions (i.e., head motions), we randomly select one observed space to serve as the \"canonical space\". By doing so, we are able to align all observed spaces with the canonical space, thereby eliminating the effects of head motions. Note that once selected, the canonical space is fixed during training and inference. To model speech-sensitive motions, we propose a speech-driven implicit model to generate lip images in the canonical space without head motion effect (Sec. 4.2). Thus, the generated lip image is canonical-view. A Geometry-Aware", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mutual Explicit Mapping (GAMEM) module is further proposed to model the speech-insensitive head motions without speech effect. Next, we project the lip images in canonical space generated by the above implicit model to the observed space based on GAMEM so that the synthesized lip image can be aligned and composed with any arbitrary observed frame, giving the model the flexibility for diversified composition (Sec. 4.3). Finally, a blending network (i.e., Blend-Net) and a contrastive sync loss are presented to improve the quality of final synthetic images $\\hat{I}_o$ (Sec. 4.4).", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Disentangled and Synced Implicit Modeling", + "text_level": 1, + "bbox": [ + 498, + 252, + 872, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The central part of Speech2Lip is the disentangled and syncsed speech-driven implicit generator (Figure 3(a)), which focuses on generating 2D lip appearances that are synchronized with cross-modal speech. By employing motion and appearance disentanglement, the generator only retains speech-sensitive motions and appearances. A lip-syncing contrastive loss is utilized (Figure 3(c)) so that it can generate syncsed high-fidelity mouth appearance by learning from very short video data. All these together make the implicit modeling disentangled and syncsed.", + "bbox": [ + 496, + 275, + 890, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Speech-driven Implicit Model To achieve high-fidelity canonical lip image generation, we utilize a speech-driven implicit model. An implicit model is defined as a function that maps the coordinate signal to another signal [29], e.g., color. The input coordinates of the implicit model are defined in a continuous space (i.e., the real field) so that it helps exploit the inherent relationship between spatially adjacent locations. Also, it can be further mapped into the frequency domain and benefit the learning of high-frequency information [20]. Besides, by learning a dedicated model for a specific scenario, implicit models also gain a strong ability to generate high-quality desired outputs. In our model, color information $\\hat{c}_{c,n} = (r,g,b)$ is regarded as our output signal (appearance color), and the canonical implicit function $f_{\\theta}$ can be defined as", + "bbox": [ + 496, + 431, + 892, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} _ {c, n} = \\boldsymbol {f} _ {\\theta} \\left(\\boldsymbol {x} _ {c, n}, \\boldsymbol {a}, \\boldsymbol {t} _ {s}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 671, + 890, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{c,n}$ is the continuous 2D pixel coordinate vector (u,v) in the canonical space. $\\pmb{a} \\in \\mathbb{R}^{64}$ is the feature vector of the input speech at the concerned moment and $t_s$ represents the timestamp information, which is utilized to enhance the temporal consistency between adjacent frames.", + "bbox": [ + 496, + 699, + 893, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continuous Sampling The next question that arises is how to obtain the corresponding supervision signal to train the implicit model, which poses a challenge as the coordinates of the pixels are always integers, while we define the coordinates in a continuous space for more expressive features. To overcome this challenge, we adopt the approach developed in [8], which enables us to sample and generate corresponding supervisions in the continuous coordinate space.", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22170", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cd3a30426c3a701cd31a2d098b0f095e111dfda4f46823ca215b7bb432dd11bb.jpg", + "image_caption": [ + "Figure 3. The overall framework of Speech2Lip. Our framework decomposes speech-sensitive and speech-insensitive motions/appearances first, models them individually (the major part of (a)), and composes the ultimate output image (b). Besides, the synchronization performance enhancement module and the GAMEM are illustrated in (c) and (d), respectively. The inputs include continuous pixel coordinates, speech audio signals, and timestamps. The speech-driven implicit model will generate speech-sensitive canonical-view lip images, which will be further transformed into observed space to composite the eventual output image. A full-head depth map is learned along with the training process, supporting pose-controllable synthesis." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 571, + 311 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/747bdadaa66b16a42672747da5613f67fc2d593df9ca61fa30bbbc045bea7914.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 90, + 888, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Particularly, for each pixel $x_{c}$ , a rectangle with an arbitrary shape is randomly sampled, the four corner points (i.e., $x_{c,n}, n \\in \\{00, 01, 10, 11\\}$ ) of which are regarded as our inputs. A weighted averaged value $\\hat{c}_{c}$ is then calculated as", + "bbox": [ + 76, + 426, + 470, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {c}} _ {c} = \\sum_ {n \\in \\{0 0, 0 1, 1 0, 1 1 \\}} \\frac {S _ {n}}{S} \\cdot \\hat {\\boldsymbol {c}} _ {c, n}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 496, + 468, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{c}_{c,n}$ are the output appearance values of the sampled points. As illustrated in Figure 4, the blue-dashed-line rectangle represents the area $S$ , which is divided into four sub-rectangles by gray lines, each with an area $S_{n}$ . By leveraging this strategy, we can generate supervision for points with continuous positions, which enables us to fully exploit the advantages of the implicit model and achieve high-fidelity visual results.", + "bbox": [ + 75, + 546, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Geometry-Aware Mutual Explicit Mapping", + "text_level": 1, + "bbox": [ + 76, + 676, + 446, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once we have obtained the speech-sensitive lip sequences in the canonical view, we can construct the eventually synthesized image $\\hat{I}_o$ by assigning the synthesized pixels to a location in the observed space. However, since the assigning process is speech-insensitive and only relies on the geometry information, we introduce the GAMEM module, illustrated in Figure 3, to explicitly model it. The GAMEM module comprises Canonical-to-Observed (C2O), Observed-to-Canonical (O2C) Mapping, and a learnable full-head depth map, which supports various applications, including pose-controllable image synthesis.", + "bbox": [ + 75, + 699, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "C2O/O2C Mapping These mappings represent how the images transform between two spaces (i.e., canonical space", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b29d85034d561383250a604e9080fffbb758e88a3fee026ae3de0a125f4169c6.jpg", + "image_caption": [ + "Figure 4. Continuous Sampling strategy. This strategy can generate supervision signals for sampled positions at different resolutions in the training time." + ], + "image_footnote": [], + "bbox": [ + 584, + 428, + 810, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and observed space). As depicted in Figure 3, O2C Mapping aims at creating supervisions $I_{o \\to c}^{m}$ for training the lip generator model $f_{\\theta}$ and C2O Mapping is utilized to warp the generated lip image into the observed view to produce $\\hat{I}_{c \\to o}^{m}$ for further composition. As they follow similar principles, we take the O2C Mapping as an example to illustrate the process. The input of O2C Mapping contains $I_{o}$ and $I_{c}$ , and the output are pixel correspondences that map pixel coordinates in observed space to the canonical space. To achieve it, we use the 3D Morphable Model (3DMM) [2] to calculate the overall camera intrinsic matrix $K$ and rough face geometry $G$ . The head poses $T_{o} \\in \\mathbb{R}^{4 \\times 4}$ and $T_{c} \\in \\mathbb{R}^{4 \\times 4}$ , which include a rotation matrix and a translation vector, are also estimated. And the relative head pose $T_{c \\to o} \\in \\mathbb{R}^{4 \\times 4}$ between two spaces can be computed as", + "bbox": [ + 496, + 571, + 890, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {c \\rightarrow o} = T _ {o} \\times T _ {c} ^ {- 1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 814, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, face depth maps $D_{o}$ and $D_{c}$ for $I_{o}$ and $I_{c}$ obtained from 3DMM can be further interpolated based on $G$ , and the corresponding head poses. Noticing that 3DMM can only be aware of the face area instead of the head area,", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22171", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "but the information about the rest part of the head (e.g., forehead, ears, eyes, mouth) is lacking. Hence we optimize the $D_{c}$ along with our model to complete the missing depth region. If $p_{o}$ and $p_{c}$ are the corresponding 2D homogeneous pixel grid coordinates on $I_{o}$ and $I_{c}$ , the geometric relationships can be explicitly formulated as", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {o} p _ {o} = K T _ {c \\rightarrow o} D _ {c} K ^ {- 1} p _ {c}, \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 193, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "so the position mapping from canonical space to observed space can be easily obtained as", + "bbox": [ + 75, + 222, + 468, + 253 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {\\boldsymbol {c} \\rightarrow \\boldsymbol {o}} \\left(T _ {c \\rightarrow o}, D _ {c}, p _ {c}\\right) = p _ {o}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 265, + 468, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we can warp the observed ground-truth lip images $I_{o}^{m}$ into the canonical space by backward warping, denoted by $I_{o\\rightarrow c}^{m}$ , to guide the learning of the lip generation model. The C2O Mapping can be similarly defined by $T_{o\\rightarrow c}$ and $D_{o}$ . With it, we can warp the generated canonical lip image $\\hat{I}_c^m$ into the observed space by backward warping, denoted by $\\hat{I}_{c\\to o}^{m}$ , to compose with the given observed image $I_{o}$ for synthesizing a talking face.", + "bbox": [ + 76, + 294, + 470, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Controllable Synthesis Thanks to the simultaneously learned complete depth map $D_{c}$ , pose-controllable full-head image synthesis according to users' requirements is also supported in our model. Specifically, we employ $D_{c}$ to determine location correspondences based on Eq. 3 and Eq. 5 by altering $T_{o}$ . There is a slight difference in this situation, which is to use the forward warping strategy as the ground-truth depths at new given poses are missing, which will produce black holes. To mitigate this issue, we propose a data augmentation method in training by randomly adding black holes with a ratio of $50\\%$ to the image to let the Blend-Net learn to fill these areas.", + "bbox": [ + 75, + 419, + 472, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Overall Refinement", + "text_level": 1, + "bbox": [ + 76, + 611, + 264, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Blending With $\\hat{I}_{c\\rightarrow o}^{m}$ , we can integrate it by directly pasting it to the original observed frame based on 2D lip landmarks [3]. However, the paste operation often results in mismatching artifacts in the boundary area, and the image artifacts after data augmentation should also be modified. Therefore, we propose a Blend-Net for blending. The Blend-Net takes the pasted image after the paste operation as input to synthesize harmonized final output $\\hat{I}_o$ . Since the target is fusion and amending instead of generation, we predict the pixel residual, which is further added back to the input image to composite the final output image. Detailed network structure is illustrated in the supplementary file.", + "bbox": [ + 75, + 638, + 468, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synchronization Enhancement. To further improve the synchronization performance, we introduce a pre-trained sync expert network to boost the model's performance in synchronization inspiring by [26]. The sync expert network consists of two pre-trained encoders, which extract features", + "bbox": [ + 75, + 825, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of image and speech audio within a sliding window, denoted by $i$ and $a$ respectively. Different from [26], we only have a short video for training. Therefore, the unsynced speech-image pairs are also exploited to construct the contrastive loss which helps avoid falling into a trivial solution. Its effectiveness is also verified in our experiments. The distance of syncsed speech-image pairs is desired to be closer while that of the unsynced speech-image pairs should be farther. Thus, we define a contrastive sync loss $\\mathcal{L}_s$ based on the metric which is widely used in contrastive learning:", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = y \\cdot (1 - \\cos (\\boldsymbol {i}, \\boldsymbol {a})) + (1 - y) \\cdot \\max (0, \\cos (\\boldsymbol {i}, \\boldsymbol {a})), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 252, + 890, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cos$ is the cosine similarity metric. $y = 1$ and $y = 0$ represent positive and negative speech-image pairs, respectively. For the input speech, the positive images are the matched images at the same timestamp while the negative images are randomly chosen at some timestamps else.", + "bbox": [ + 496, + 279, + 890, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Training Objectives", + "text_level": 1, + "bbox": [ + 498, + 363, + 689, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our whole pipeline is trained in a self-supervised manner using the observed frames to provide the supervisory signals. The overall loss includes a canonical mouth image reconstruction loss $\\mathcal{L}_m$ , a depth loss $\\mathcal{L}_d$ , a sync loss $\\mathcal{L}_s$ , and a whole observed image reconstruction loss $\\mathcal{L}_w$ .", + "bbox": [ + 496, + 387, + 890, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_m$ measures the reconstruction error between the predicted lip image $\\hat{I}_c^m$ and ground-truth lip image $I_{o\\rightarrow c}^{m}$ as", + "bbox": [ + 498, + 463, + 890, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = \\mathcal {L} _ {p} \\left(\\hat {I} _ {c} ^ {m}, I _ {o \\rightarrow c} ^ {m}\\right) + \\| \\hat {I} _ {c} ^ {m} - I _ {o \\rightarrow c} ^ {m} \\| _ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 501, + 890, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{o\\rightarrow c}^{m}$ is the warped ground truth lip image using O2C space mapping (see the supplementary file for more details) and the widely-used perceptual loss $\\mathcal{L}_p$ is defined in [40]. Similarly, overall reconstruction loss $\\mathcal{L}_w$ is computed as", + "bbox": [ + 496, + 530, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {w} = \\mathcal {L} _ {p} \\left(\\hat {I} _ {o}, I _ {o}\\right) + \\| \\hat {I} _ {o} - I _ {o} \\| _ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 599, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The canonical head depth map is initialized by the incomplete depth map computed by 3DMM and is trained with the help of the photometric loss as", + "bbox": [ + 496, + 628, + 890, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\left\\| \\hat {I} _ {o \\rightarrow c} - I _ {c} \\right\\| _ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 683, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{I}_{o\\rightarrow c}$ is the warped predicted image. With the contrastive sync loss (Sec. 4.4), the overall loss function is", + "bbox": [ + 496, + 710, + 890, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\omega_ {m} \\cdot \\mathcal {L} _ {m} + \\omega_ {w} \\cdot \\mathcal {L} _ {w} + \\omega_ {d} \\cdot \\mathcal {L} _ {d} + \\omega_ {s} \\cdot \\mathcal {L} _ {s}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 752, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details are shown in the supplementary file.", + "bbox": [ + 500, + 779, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 806, + 633, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets and metrics", + "text_level": 1, + "bbox": [ + 500, + 832, + 696, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our algorithm and other methods on three datasets belonging to three specific speakers (Testset I, Testset II and Testset III) as recent speaker-specific", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22172", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f3792334849146e9664a2484666ae7010ff0ea591429c1a2e0bd226d3760f65b.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative results compared with the SOTA methods. Image quality assessment metrics (i.e., PSNR, SSIM, and CPBD) are computed within mouth region. Algorithms are categorized into two classes based on the training datasets for a fair comparison. The first ones are trained in large public datasets while the other ones are trained using only 3-5min videos. The best results of each class are in bold and overall best results are with underlines." + ], + "table_body": "
MethodTrained with large extra dataTestset ITestset IITestset III Sync↑
PSNR↑SSIM↑CPBD↑LMD↓Sync↑PSNR↑SSIM↑CPBD↑LMD↓Sync↑
Ground TruthN/AN/A1.0000.1860.0009.102N/A1.0000.1210.0008.6884.877
ATVG [7]Yes28.4520.8180.0195.4235.81328.0510.6680.0034.2036.2243.571
MakeitTalk [44]Yes29.6920.9060.0504.2154.11528.9960.8130.0614.4635.5592.320
Wav2Lip [26]Yes31.5570.9800.1153.05310.03131.7930.9560.0653.4159.9365.809
PC-AVS [43]Yes29.0720.8800.0404.5959.25828.3590.7340.0464.3058.5865.206
LSP [19]No29.5150.9000.0983.1745.37728.8950.7760.1174.9726.8113.046
AD-NeRF [16]No32.2230.9540.0512.9896.04230.8850.9090.0553.2105.9103.285
DFRF [28]No33.2920.9740.0943.0795.25231.4190.9440.1243.1395.5522.879
Speech2LipNo34.8150.9870.2242.9767.77133.1970.9620.1253.0827.3704.379
", + "bbox": [ + 89, + 88, + 880, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods [16, 19] usually do. Training videos are collected from [19] as they are publicly available. All video sequences are resampled to 25 FPS and split into training part and test part with a ratio of $90\\% / 10\\%$ . The training splits of Testset I $(500 \\times 500)$ and Testset II $(624 \\times 624)$ are employed to train the models and then same-identity evaluations are conducted on the test splits. Testset III is utilized for cross-identity, -gender, and -language tests. We present more evaluation results in the supplementary file.", + "bbox": [ + 75, + 334, + 472, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Since both face generation and lip generation models are contained in our quantitative evaluations, we conduct the image quality assessment just around the mouth region for a fair comparison. To evaluate the appearance quality, we first utilize image quality metrics Peak Signal-to-Noise Ratio (PSNR) and SSIM [37]. A no-reference objective image sharpness metric based on Cumulative Probability of Blur Detection (CPBD) [23, 21, 22] is further utilized to measure the overall perceptual image quality. For synchronization evaluation, Landmarks Distance (LMD) around the mouth region is exploited to compute the accuracy of lip shape following [6]. The confidence score of lip synchronization computed by a pretrained SyncNet [12] is also applied to measure the synchronization performance, labeled by Sync.", + "bbox": [ + 75, + 477, + 472, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparisons with State of the Arts", + "text_level": 1, + "bbox": [ + 76, + 719, + 383, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Same-identity Evaluation. For speaker-independent models, we conduct experiments using their officially-released pre-trained models. Since AD-NeRF [16] and DFRF [28] are speaker-specific models, we retrain them using our training data for a fair comparison. DFRF aims at few-shot learning but its lip synchronization performance witnesses a significant drop when the number of training images decreases from 5000 to 15. Hence, we still train DFRF on thousands of images. Also, we directly use LSP's [19] pretrained models on Testset I and Testset II to conduct experi", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ments (subject May for Testset I and Testset III, and subject Obama2 for Testset II). We cannot provide a quantitative comparison with SSP-NeRF [18] here, as neither codes nor pre-trained weights are available. Thus, for comparison, we extract the speech from their released demo and then present qualitative comparisons in the supplementary file.", + "bbox": [ + 496, + 334, + 893, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results are shown in Table 1. We divide all the models into two types based on the scale of the training dataset. The upper four algorithms (speaker-independent models) are trained using public video datasets with large amounts of identities and speech-image pairs (e.g., LRS2 [1, 11, 13] and VoxCeleb2 [10]) while the latter four methods (speaker-specific models) only use 5-minute videos of a specific identity for training. Speaker-independent models tend to perform well in synchronization because they can learn lip motions well from various speech-visual pairs in training time. However, the cost is sacrificing the visual quality. The metrics of image quality (e.g., PSNR, SSIM, and CPBD) are much lower than that of specific models. Wav2lip [26] is an exception because it only generates lip images, but the details still can not be produced well (Figure 5). In contrast, speaker-specific models usually have much better image qualities, which is favorable in practice. Our model outperforms all the SOTA algorithms in image quality and achieves the best synchronization performance among all the speaker-specific models, and the Sync score is also competitive when compared to those speaker-independent models. It is also worth mentioning that the Sync scores of Wav2lip [26] and PC-AVS [43] are higher than that of the ground truth. The reason might be that the Sync score is sensitive to synchronization quality only when synchronization quality is within a range and a higher score above a threshold may not necessarily imply a higher synchronization quality. We design a toy experiment to verify it in the supplementary file and our user study can also demonstrate it.", + "bbox": [ + 496, + 426, + 893, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-identity Evaluation. Our model also has the abil", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22173", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54512ee06f077ac6673d5030f7d533444c5e53634b0733abd8ebb7ddfaa7bfba.jpg", + "image_caption": [ + "Figure 5. Qualitative results compared with SOTA methods. The Lip area is cropped based on the detected 2D landmarks for a clear comparison. Speaker-independent and speaker-specific models are remarked by different colors." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 166, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85cffeeca06aad9ce7bb1d34f3ef9f6984bfc5548b2ef9e4c70d9d69a3bbec2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 90, + 527, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4885db8e54c2ee536f28575f4512b8534e61893a95201e143fb9b23c9bf49076.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 90, + 888, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ity to generate cross-identity, -gender, and -language results. Since there is no ground truth image, we provide the sync performance comparison on generating a video for the model trained on Testset I (female, British) using another speaker's speech (male, French). This setting is defined as Testset III, and results are shown in the last column of Table 1. Our model still has significant superiority and outperforms other speaker-specific models by a large margin. We also test the effect of unsynced speech-image pairs used in Eq. 6 on Testset III. Without negative pairs, the Sync score decreases from 4.379 to 3.830, showing the effectiveness of our contrastive sync loss.", + "bbox": [ + 75, + 375, + 472, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Qualitative comparisons of different algorithms are presented in Figure 5. We can see that speaker-independent methods all suffer from low image quality. Besides, ATVG and MakeiTalk have unnatural speech-insensitive movements (e.g., head motion and background motion) and PC-AVS has an extra issue of struggling with identity preservation. Speaker-specific approaches have much better visual quality but most of them can hardly estimate precise lip details when there exist considerable significant lip motions. And fine details like teeth and tongue are not well modeled either. Besides, AD-NeRF sometimes causes head-torso separation due to the adoption of two separate NeRFs. In contrast, our method Speech2Lip performs well in all these situations, as the generated shapes are more accurate and the synthetic images are more clear and realistic, especially the lip area. More results can be found in the supplementary videos.", + "bbox": [ + 75, + 561, + 468, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Complexity Comparisons. To further demonstrate the superiority of our model, complexity and computational cost comparisons with speaker-specific models are conducted on Testset I. In Table 2, our model gains the best results with much less complexity and lower computational cost.", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/18299d8cc8024099a3c46fa770d4c38cf0726b6c04435afcdc12d4a231e8babf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLSPAD-NeRFDFRFOurs
Model size (MB) ↓500302030
Train time (hour) ↓38.5806030
Test speed (FPS) ↑350.060.0418
PSNR (dB) ↑29.5232.2233.2934.82
Sync ↑5.386.045.257.77
", + "bbox": [ + 532, + 371, + 859, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Speaker-specific model comparisons. The best results are in bold and the second best results are with underlines.", + "bbox": [ + 500, + 465, + 890, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. To verify the perceptual quality, user studies are conducted on 16 generated video clips covering both Testset I and Testset II. Videos are generated by our proposed Speech2Lip and other SOTA algorithms, with algorithm names hidden and video order randomized. 17 participants independently evaluated all the videos. For each video, each participant gave a Mean Opinion Score (MOS) from 1-5 for each of three aspects: qualities of speech-visual synchronization, image fidelity, and image realness. Higher scores represent better quality. The overall results are shown in Table 3. Speech2Lip not only achieves the highest MOS in image quality measurement (e.g., 4.618 and 4.382) but also outperforms all the other algorithms in lip synchronization (e.g., 4.265), exhibiting Speech2Lip has the most satisfactory overall video quality.", + "bbox": [ + 496, + 518, + 890, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 756, + 653, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contributions of loss functions. We first conduct an ablation study on loss functions. The evaluation metrics contain the PSNR around the lip area and the sync score. The PSNR of the whole image is also retained since practical applications attach more attention to the overall quality of the talking portrait scene. Results are shown in Table 4, demonstrating the loss functions all contribute to the increased performance on both visual quality and synchronization.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22174", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ea9073f00849656e86fa7d14b2e794a154435e9f3a75f642b1aaa20d5a46d840.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsATVGMakeitTalkWav2LipPC-AVSLSPAD-NeRFDFRFSpeech2Lip
Lip Synchronization1.5001.8534.1913.0593.1183.0882.8534.265
Image Fidelity1.1472.7943.2652.5293.8823.2353.2654.618
Image Realness1.1182.0883.7652.2653.0002.3242.9714.382
", + "bbox": [ + 158, + 89, + 812, + 155 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1f21ac99ed47689a7c34ef27d8f042b0f4b6b487ee6783ea33d8f266f6c07ac1.jpg", + "table_caption": [ + "Table 3. Detailed user study results compared with SOTA methods. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodsPSNRlip↑PSNRimg↑Sync↑
Lw34.10436.9476.313
Lw + Ld34.42936.9946.359
Lw + Ld + Lm34.69437.1937.194
Lw + Ld + Lm + Ls34.81537.2457.771
", + "bbox": [ + 84, + 191, + 460, + 271 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e6ee61cdeb8040da8e3e73ac9ffe08cab13133aac007d4d2a10fe68ac0f3a063.jpg", + "table_caption": [ + "Table 4. Ablation study results about loss function on Testset I." + ], + "table_footnote": [], + "table_body": "
MethodsPSNRlip↑PSNRimg↑Sync↑
w/o continuous sampling34.70237.2147.553
w/o implicit model34.25037.0586.790
w/o Blend-Net33.73636.8816.065
w/o time34.51237.1217.146
Speech2Lip34.81537.2457.771
", + "bbox": [ + 84, + 297, + 460, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study results about network design on Testset I.", + "bbox": [ + 76, + 392, + 465, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contributions of individual components. We also explore the benefits of our key components by removing each component to see how the performance changes in Table 5. \"w/o implicit model\" represents we employ an explicit lip generation model instead, and \"w/o time\" indicates times-tamp is deleted from inputs. From the results, it can be concluded that each design plays an important role. Among them, the Blend-Net significantly improves video quality, and the design of implicit modeling schema is also critical.", + "bbox": [ + 75, + 429, + 468, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Controllable Synthesis Results", + "text_level": 1, + "bbox": [ + 76, + 575, + 346, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Thanks to the joint optimization and completion of a full-head depth map $D_{c}$ in GAMEM, our model supports synthesizing full-head images driven by novel head poses. Specifically, $D_{c}$ can be projected into any observed view to obtain $D_{o}$ . $D_{c}$ and $T_{o}$ in Eq. (3) are further replaced by $D_{o}$ and the new target head pose, respectively. Then, we adopt Eq. (5) to calculate the correspondence. The correspondence is further exploited to warp the synthesized image in the observed space into a new target space. It is noted that depth maps are used for coarse mapping between spaces while facial expression is more taken charge of by fine-grained contents from observed views. The pose-controllable novel view synthesis results are shown in Figure 7. When testing with moderate pose changes, our model achieves comparable results as AD-NeRF.", + "bbox": [ + 75, + 598, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We note that existing methods with 3D modeling such as AD-NeRF also cannot generate good synthesis results if the evaluated poses deviate too much from the training data. The reason is that the training data widely used in speaker-specific models [19, 16, 28, 18] only show the front view", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe3e1b2368d9dd5d8ac162519478c0aa7315a28e3ff6b00e8346834a3fda973c.jpg", + "image_caption": [ + "Figure 6. 3D Mesh of AD-NeRF on Testset I." + ], + "image_footnote": [], + "bbox": [ + 555, + 181, + 683, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b65d7199e3c382b4c1ed709256cb5fdf92e81f522b603c15fe6c98f4cc085c28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 181, + 834, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9adf590cf715a8fd37d35fa7d3fa95b7f0d64a6670283ff5e075ba4d322f5e5c.jpg", + "image_caption": [ + "(a) AD-NeRF" + ], + "image_footnote": [], + "bbox": [ + 537, + 316, + 849, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fcdee3442b9ba562bb89e6e2979c44d8d7aefbea96911ed622dd9973c9654cc7.jpg", + "image_caption": [ + "(b) Speech2Lip", + "Figure 7. Novel view synthesis capability of our model. We rotate the head region with a random angle." + ], + "image_footnote": [], + "bbox": [ + 537, + 391, + 849, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with limited pose variations, being naturally hard for 3D modeling. The 3D mesh is incomplete, and the surface is rough (Figure 6). In addition, the computational costs of 3D methods are extremely high (Table 2).", + "bbox": [ + 498, + 503, + 893, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 584, + 617, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel decomposition-synthesis-composition framework called Speech2Lip for high-fidelity talking head video synthesis, which disentangles speech-sensitive and speech-insensitive motions/appearances. By presenting a syncsed speech-driven implicit model, a GAMEM module, a Blend-Net, and a contrastive sync loss, we can achieve satisfactory results with only a few minutes of training video. Our framework also supports pose-controllable synthesis. In the future, we plan to study generating realistic expressions driven by speech and explore combining our insights with advanced general image generation methods such as diffusion-based models for better generalizability. We hope that our work inspires more future research in this field and encourages the development of positive applications. However, we also urge caution to prevent any potential abuses. More discussions about limitations are described in the supplementary file.", + "bbox": [ + 496, + 611, + 890, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work has been supported by the Research Fund from Tencent ARC lab.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22175", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] T. Afouras, J. S. Chung, A. Senior, O. Vinyals, and A. Zisserman. Deep audio-visual speech recognition. In arXiv:1809.02108, 2018. 6", + "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3, 4", + "[3] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 5", + "[4] Lele Chen, Guofeng Cui, Ziyi Kou, Haitian Zheng, and Chenliang Xu. What comprises a good talking-head video generation? In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 2", + "[5] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX, pages 35-51. Springer, 2020. 2", + "[6] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 6", + "[7] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. 1, 2, 6", + "[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 3", + "[9] Joon Son Chung, Amir Jamaludin, and Andrew Zisserman. You said that? arXiv preprint arXiv:1705.02966, 2017. 1, 2", + "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 6", + "[11] J. S. Chung, A. Senior, O. Vinyals, and A. Zisserman. Lip reading sentences in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 6", + "[12] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian conference on computer vision, pages 251-263. Springer, 2016. 6", + "[13] J. S. Chung and A. Zisserman. Lip reading in profile. In British Machine Vision Conference, 2017. 6", + "[14] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision, pages 408-424. Springer, 2020. 2", + "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 2, 3, 6, 8", + "[17] Rithesh Kumar, Jose Sotelo, Kundan Kumar, Alexandre de Brébisson, and Yoshua Bengio. Obamanet: Photo-realistic lip-sync from text. arXiv preprint arXiv:1801.01442, 2017. 1, 2", + "[18] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. arXiv preprint arXiv:2201.07786, 2022. 2, 3, 6, 8", + "[19] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2, 6, 8", + "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3", + "[21] Niranjan Narvekar and Lina J Karam. An improved no-reference sharpness metric based on the probability of blur detection. In Workshop on Video Processing and Quality Metrics, 2010. 6", + "[22] Niranjan D Narvekar and Lina J Karam. A no-reference perceptual image sharpness metric based on a cumulative probability of blur detection. In 2009 International Workshop on Quality of Multimedia Experience, pages 87-91. IEEE, 2009. 6", + "[23] Niranjan D Narvekar and Lina J Karam. A no-reference image blur metric based on the cumulative probability of blur detection (cpbd). IEEE Transactions on Image Processing, 20(9):2678-2683, 2011. 6", + "[24] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 3", + "[25] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3", + "[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1, 2, 5, 6", + "[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 2", + "[28] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European conference on computer vision, 2022. 2, 3, 6, 8" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22176", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 3", + "[30] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 2022. 2", + "[31] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 1, 2", + "[32] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European conference on computer vision, pages 716-731. Springer, 2020. 2", + "[33] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 128(5):1398-1413, 2020. 2", + "[34] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3", + "[35] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2", + "[36] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2531–2539, 2022. 2", + "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[38] Olivia Wiles, A Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Proceedings of the European conference on computer vision (ECCV), pages 670-686, 2018. 2", + "[39] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 2", + "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[41] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2", + "[42] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9299–9306, 2019. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 2, 6", + "[44] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 1, 2, 6" + ], + "bbox": [ + 501, + 90, + 892, + 231 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22177", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/b8d7dfde-aec2-4ac3-9ef3-b8789d6534b1_model.json b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/b8d7dfde-aec2-4ac3-9ef3-b8789d6534b1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d3954d500365c11ff48655ea15fb536806d3dc0 --- /dev/null +++ b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/b8d7dfde-aec2-4ac3-9ef3-b8789d6534b1_model.json @@ -0,0 +1,2013 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.13, + 0.882, + 0.175 + ], + "angle": 0, + "content": "Speech2Lip: High-fidelity Speech to Lip Generation by Learning from a Short Video" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.215, + 0.872, + 0.253 + ], + "angle": 0, + "content": "Xiuzhe \\(\\mathsf{W u}^{1}\\), Pengfei \\(\\mathsf{H u}^{2}\\), Yang \\(\\mathsf{W u}^{3,4}\\), Xiaoyang \\(\\mathsf{L y u}^{1}\\), Yan-Pei \\(\\mathsf{C a o}^{3}\\), Ying \\(\\mathsf{S h a n}^{3}\\), Wenming \\(\\mathsf{Y a n g}^{2}\\), Zhongqian \\(\\mathsf{S u n}^{4}\\), Xiaojuan \\(\\mathsf{Q i}^{1}\\)," + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.275, + 0.879, + 0.295 + ], + "angle": 0, + "content": "1The University of Hong Kong, 2Tsinghua University, 3ARC Lab, Tencent PCG, 4Tencent AI Lab" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.473, + 0.884 + ], + "angle": 0, + "content": "Synthesizing realistic videos according to a given speech is still an open challenge. Previous works have been plagued by issues such as inaccurate lip shape generation and poor image quality. The key reason is that only motions and appearances on limited facial areas (e.g., lip area) are mainly driven by the input speech. Therefore, directly learning a mapping function from speech to the entire head image is prone to ambiguity, particularly when using a short video for training. We thus propose a decomposition-synthesis-composition framework named Speech to Lip (Speech2Lip) that disentangles speech-sensitive and speech-insensitive motion/appearance to facilitate effective learning from limited training data, resulting in the generation of natural-looking videos. First, given a fixed head pose (i.e., canonical space), we present a speech-driven implicit model for lip image generation which concentrates on learning speech-sensitive motion and appearance. Next, to model the major speech-insensitive motion (i.e., head movement), we introduce a geometry-aware mutual explicit mapping (GAMEM) module that establishes geometric mappings between different head poses. This allows us to paste generated lip images at the canonical space onto head images with arbitrary poses and synthesize talking videos with natural head movements. In addition, a Blend-Net and a contrastive sync loss are introduced to enhance the overall synthesis performance. Quantitative and qualitative results on three benchmarks demonstrate that our model can be trained by a video of just a few minutes in length and achieve state-of-the-art performance in both visual quality and speech-visual synchronization. Code: https://github.com/CVMI-Lab/Speech2Lip." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.339, + 0.895, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.486, + 0.895, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Given a speech as input, our model generates high-quality talking-head videos and supports pose-controllable synthesis. The decomposition and synthesis modules make learning from a short video more effective and the composition module enables us to synthesize high-fidelity videos." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.584, + 0.633, + 0.6 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.894, + 0.703 + ], + "angle": 0, + "content": "Learning from a short video to generate personalized audio-synchronized talking videos driven by a speech is of great importance to various applications, for instance, digital human animation, video dubbing, and UGC video creation. However, synthesizing high-fidelity videos from speech for a desired speaker remains a challenging task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The first challenge arises from complicated motion patterns. Although speech majorly influences lip areas (i.e., speech-sensitive areas), lip movements are often accompanied by other motions, such as global head movements, which greatly impact lip shapes and appearances. Thus, directly synthesizing a whole image from speech often leads to inaccurate lip synthesis. Second, existing methods still struggle to satisfy appearance fidelity requirements, which include both identity preservation (speaker-specific) and high-quality detail generation, such as clear details of teeth, tongue, and eye-blinking [9, 42, 26, 44, 7]. Third, it's difficult to acquire videos longer than 10 hours for a speaker which is yet required by conventional methods [17, 31]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22168" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "To tackle the aforementioned challenges, existing attempts can be coarsely categorized into two lines of research: speaker-independent and speaker-specific methods. The first line often exploits GANs [15] that need to be trained on large-scale multi-person datasets. However, GAN-based models [4, 7, 9, 26, 33, 38, 42, 43] usually synthesize low-resolution images and unnatural motions (i.e., background movements). They thus hardly meet the appearance fidelity requirement in terms of sharpness and fine appearance details. Moreover, preserving the identity of the speaker remains a challenging task [7, 43]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.596 + ], + "angle": 0, + "content": "For the consideration of high-fidelity and identity preservation, another strategy focuses on learning from a specific speaker. Although attaining high-fidelity results, early works [17, 31] often require several hours of video footage from a speaker for training, hindering their practical applicability. Recently, NeRF [20] has emerged as a promising approach for generating high-fidelity videos, which succeeds in learning from a short video of just a few minutes and having the potential to generate high-fidelity results [16, 18, 28]. Nevertheless, the models still struggle with appearance and motion ambiguity issues because they model speech-sensitive motions/appearances together with other facial areas less correlated to the given speech. This issue becomes more severe when training data is limited since no extra information can be exploited to avoid interference from signals that are not correlated with the speech. As a result, they tend to generate lip sequences that do not synchronize well with the speech and produce blurry images (see Figure 5 and Table 1). Therefore, reducing the complexity of modeling motions is critical to enable effective learning from a short video and synthesizing high-quality videos for a specific speaker." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We thus design a preliminary experiment to identify that motion and appearance of lip areas have a strong correlation with speech, while head motion and other facial areas are less related to speech (Figure ??). Motivated by the observation, we propose decomposing speech-insensitive motion/appearance from speech-sensitive one, and synthesizing them separately before composing them into a new talking video that aligns with the given speech (Figure 1). Toward this goal, we present a decomposition-synthesis-composition framework named Speech to Lip (Speech2Lip). In the decomposition stage, we introduce a speech-driven implicit model that generates high-fidelity synced lip sequences in a fixed view (i.e., canonical view). To model 3D head motion effectively, we design a Geometry-Aware Mutual Explicit Mapping (GAMEM) module that estimates explicit geometric mappings between an arbitrary observed view and the canonical view. GAMEM also includes a jointly optimized canonical-view full-head depth map, which enables the model to be 3D-aware and supports controllable synthesis driven by new" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.212 + ], + "angle": 0, + "content": "head poses. In the composition stage, GAMEM allows us to flexibly paste the synthesized canonical-view lips onto an arbitrary observed view to obtain natural synchronized talking videos. To improve the synthesis and synchronization qualities after composition, we incorporate a blending network (i.e., Blend-Net) to refine the results and a contrastive sync loss to facilitate learning from a short video for generating synchronized talking videos." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.213, + 0.866, + 0.227 + ], + "angle": 0, + "content": "Our major contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.892, + 0.308 + ], + "angle": 0, + "content": "1) We introduce a novel framework that disentangles speech-sensitive and speech-insensitive motion/appearance in high-fidelity video synthesis. By separating these components, the framework can effectively learn from limited training data." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.311, + 0.892, + 0.4 + ], + "angle": 0, + "content": "2) The proposed speech-driven implicit model synthesizes speech-sensitive contents and the GAMEM flexibly combines them with given speech-insensitive areas to generate synchronized talking heads with natural movements and support pose-controllable synthesis." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.405, + 0.892, + 0.45 + ], + "angle": 0, + "content": "3) Both qualitative and quantitative experimental results demonstrate the superiority of our method over the state-of-the-art speaker-specific methods." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.233, + 0.892, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.461, + 0.642, + 0.477 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.855 + ], + "angle": 0, + "content": "Speech-Driven Talking Face Synthesis. Video synthesis from speech is a long-standing problem. Recent works can be generally divided into two categories: speaker-specific [17, 31, 19, 16, 39, 18, 28] and speaker-independent [9, 42, 26, 44, 7, 5, 41, 35, 36]. In the first track, earlier works have succeeded in obtaining realistic visual results for a target person [17, 31] but required hours of video belonging to one specific speaker. Therefore, many efforts [19, 16, 39, 18, 28] have been made to train the model with a shorter talking video (3-5 minutes). However, they utilize speech to forecast overall motions including speech-insensitive ones, failing to learn accurate lip shapes and appearances. The other line (i.e., speaker-independent method) aims to build a universal model for all identities [9, 42, 26, 44, 7]. The end-to-end pipeline [9, 42] is usually developed on GAN [15]. To boost performance, Richard et al. [27] only synthesize the mesh of limited facial areas, Prajwal et al. [26] present an extra lip-sync discriminator, Zhou et al. [43] provide additional head poses as inputs, and some other models [7, 14, 44, 32, 30] leverage the intermediate representation (e.g., 2D facial landmarks [7, 14, 44] or expression parameters [32, 30]). Nevertheless, they often suffer from low video quality, difficult identity preservation, and abnormal head motion generation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Implicit Representation based Talking Head Methods. Recently, implicit representations have shown high model" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22169" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "ing capabilities in multiple tasks [25, 20, 24, 34]. Among them, NeRF [20] obtains extraordinary performance in novel view synthesis by training on only hundreds of images. Guo et al. [16] first apply it in the speaker-specific talking head synthesis task to enable learning from a short video. However, they utilize two NeRFs to model the head and torso, resulting in the head-torso separation phenomenon. Liu et al. [18] solve it by leveraging one unified NeRF with two semantic-aware modules. Shen et al. [28] incorporate 2D image features as additional inputs to further reduce training data requirements (i.e., videos with only 10-15 seconds). Regardless, these methods only ensure visual quality, and simply use speech to drive all the complex motions, leading to the out-of-sync problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.315, + 0.383, + 0.333 + ], + "angle": 0, + "content": "3. Empirical Study and Motivations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.474, + 0.552 + ], + "angle": 0, + "content": "Our approach is motivated by the observation that only limited facial areas are highly correlated to speech. To verify this, we conduct a preliminary experiment to determine which areas are most speech-sensitive. Specifically, we apply warping to all captured images with varying head poses (refer to observed views) using the 3D Morphable Model (3DMM) [2] and bring them to a fixed head pose (known as the canonical view). We then compute a motion heatmap in the canonical view to determine the areas most responsive to speech. As depicted in Figure 2, the elimination of head motion results in most of the motions around the lip regions displaying high sensitivity to the input speech. Additional examples can be found in the supplementary file. For more details on the warping process, please refer to Sec. 4.3." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.563, + 0.43, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.643, + 0.406, + 0.656 + ], + "angle": 0, + "content": "Figure 2. Speech-sensitive Motion Heatmap." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.657, + 0.169, + 0.672 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.188, + 0.696 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Figure 3 provides an overview of our model. We define each frame captured in the observed view with its own head motion as the \"observed space\". To disentangle speech-sensitive and speech-insensitive motions (i.e., head motions), we randomly select one observed space to serve as the \"canonical space\". By doing so, we are able to align all observed spaces with the canonical space, thereby eliminating the effects of head motions. Note that once selected, the canonical space is fixed during training and inference. To model speech-sensitive motions, we propose a speech-driven implicit model to generate lip images in the canonical space without head motion effect (Sec. 4.2). Thus, the generated lip image is canonical-view. A Geometry-Aware" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "Mutual Explicit Mapping (GAMEM) module is further proposed to model the speech-insensitive head motions without speech effect. Next, we project the lip images in canonical space generated by the above implicit model to the observed space based on GAMEM so that the synthesized lip image can be aligned and composed with any arbitrary observed frame, giving the model the flexibility for diversified composition (Sec. 4.3). Finally, a blending network (i.e., Blend-Net) and a contrastive sync loss are presented to improve the quality of final synthetic images \\(\\hat{I}_o\\) (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.253, + 0.874, + 0.27 + ], + "angle": 0, + "content": "4.2. Disentangled and Synced Implicit Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.429 + ], + "angle": 0, + "content": "The central part of Speech2Lip is the disentangled and syncsed speech-driven implicit generator (Figure 3(a)), which focuses on generating 2D lip appearances that are synchronized with cross-modal speech. By employing motion and appearance disentanglement, the generator only retains speech-sensitive motions and appearances. A lip-syncing contrastive loss is utilized (Figure 3(c)) so that it can generate syncsed high-fidelity mouth appearance by learning from very short video data. All these together make the implicit modeling disentangled and syncsed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.893, + 0.659 + ], + "angle": 0, + "content": "Speech-driven Implicit Model To achieve high-fidelity canonical lip image generation, we utilize a speech-driven implicit model. An implicit model is defined as a function that maps the coordinate signal to another signal [29], e.g., color. The input coordinates of the implicit model are defined in a continuous space (i.e., the real field) so that it helps exploit the inherent relationship between spatially adjacent locations. Also, it can be further mapped into the frequency domain and benefit the learning of high-frequency information [20]. Besides, by learning a dedicated model for a specific scenario, implicit models also gain a strong ability to generate high-quality desired outputs. In our model, color information \\(\\hat{c}_{c,n} = (r,g,b)\\) is regarded as our output signal (appearance color), and the canonical implicit function \\(f_{\\theta}\\) can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.672, + 0.892, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\hat {c} _ {c, n} = \\boldsymbol {f} _ {\\theta} \\left(\\boldsymbol {x} _ {c, n}, \\boldsymbol {a}, \\boldsymbol {t} _ {s}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.895, + 0.776 + ], + "angle": 0, + "content": "where \\( x_{c,n} \\) is the continuous 2D pixel coordinate vector (u,v) in the canonical space. \\( \\pmb{a} \\in \\mathbb{R}^{64} \\) is the feature vector of the input speech at the concerned moment and \\( t_s \\) represents the timestamp information, which is utilized to enhance the temporal consistency between adjacent frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Continuous Sampling The next question that arises is how to obtain the corresponding supervision signal to train the implicit model, which poses a challenge as the coordinates of the pixels are always integers, while we define the coordinates in a continuous space for more expressive features. To overcome this challenge, we adopt the approach developed in [8], which enables us to sample and generate corresponding supervisions in the continuous coordinate space." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22170" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.092, + 0.572, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.092, + 0.889, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.896, + 0.404 + ], + "angle": 0, + "content": "Figure 3. The overall framework of Speech2Lip. Our framework decomposes speech-sensitive and speech-insensitive motions/appearances first, models them individually (the major part of (a)), and composes the ultimate output image (b). Besides, the synchronization performance enhancement module and the GAMEM are illustrated in (c) and (d), respectively. The inputs include continuous pixel coordinates, speech audio signals, and timestamps. The speech-driven implicit model will generate speech-sensitive canonical-view lip images, which will be further transformed into observed space to composite the eventual output image. A full-head depth map is learned along with the training process, supporting pose-controllable synthesis." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.427, + 0.471, + 0.489 + ], + "angle": 0, + "content": "Particularly, for each pixel \\( x_{c} \\), a rectangle with an arbitrary shape is randomly sampled, the four corner points (i.e., \\( x_{c,n}, n \\in \\{00, 01, 10, 11\\} \\)) of which are regarded as our inputs. A weighted averaged value \\( \\hat{c}_{c} \\) is then calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.497, + 0.47, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {c}} _ {c} = \\sum_ {n \\in \\{0 0, 0 1, 1 0, 1 1 \\}} \\frac {S _ {n}}{S} \\cdot \\hat {\\boldsymbol {c}} _ {c, n}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.669 + ], + "angle": 0, + "content": "where \\(\\hat{c}_{c,n}\\) are the output appearance values of the sampled points. As illustrated in Figure 4, the blue-dashed-line rectangle represents the area \\(S\\), which is divided into four sub-rectangles by gray lines, each with an area \\(S_{n}\\). By leveraging this strategy, we can generate supervision for points with continuous positions, which enables us to fully exploit the advantages of the implicit model and achieve high-fidelity visual results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.677, + 0.447, + 0.695 + ], + "angle": 0, + "content": "4.3. Geometry-Aware Mutual Explicit Mapping" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Once we have obtained the speech-sensitive lip sequences in the canonical view, we can construct the eventually synthesized image \\(\\hat{I}_o\\) by assigning the synthesized pixels to a location in the observed space. However, since the assigning process is speech-insensitive and only relies on the geometry information, we introduce the GAMEM module, illustrated in Figure 3, to explicitly model it. The GAMEM module comprises Canonical-to-Observed (C2O), Observed-to-Canonical (O2C) Mapping, and a learnable full-head depth map, which supports various applications, including pose-controllable image synthesis." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "C2O/O2C Mapping These mappings represent how the images transform between two spaces (i.e., canonical space" + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.429, + 0.811, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.524, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Figure 4. Continuous Sampling strategy. This strategy can generate supervision signals for sampled positions at different resolutions in the training time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.892, + 0.814 + ], + "angle": 0, + "content": "and observed space). As depicted in Figure 3, O2C Mapping aims at creating supervisions \\( I_{o \\to c}^{m} \\) for training the lip generator model \\( f_{\\theta} \\) and C2O Mapping is utilized to warp the generated lip image into the observed view to produce \\( \\hat{I}_{c \\to o}^{m} \\) for further composition. As they follow similar principles, we take the O2C Mapping as an example to illustrate the process. The input of O2C Mapping contains \\( I_{o} \\) and \\( I_{c} \\), and the output are pixel correspondences that map pixel coordinates in observed space to the canonical space. To achieve it, we use the 3D Morphable Model (3DMM) [2] to calculate the overall camera intrinsic matrix \\( K \\) and rough face geometry \\( G \\). The head poses \\( T_{o} \\in \\mathbb{R}^{4 \\times 4} \\) and \\( T_{c} \\in \\mathbb{R}^{4 \\times 4} \\), which include a rotation matrix and a translation vector, are also estimated. And the relative head pose \\( T_{c \\to o} \\in \\mathbb{R}^{4 \\times 4} \\) between two spaces can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.815, + 0.891, + 0.832 + ], + "angle": 0, + "content": "\\[\nT _ {c \\rightarrow o} = T _ {o} \\times T _ {c} ^ {- 1}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Furthermore, face depth maps \\( D_{o} \\) and \\( D_{c} \\) for \\( I_{o} \\) and \\( I_{c} \\) obtained from 3DMM can be further interpolated based on \\( G \\), and the corresponding head poses. Noticing that 3DMM can only be aware of the face area instead of the head area," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22171" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "but the information about the rest part of the head (e.g., forehead, ears, eyes, mouth) is lacking. Hence we optimize the \\(D_{c}\\) along with our model to complete the missing depth region. If \\(p_{o}\\) and \\(p_{c}\\) are the corresponding 2D homogeneous pixel grid coordinates on \\(I_{o}\\) and \\(I_{c}\\), the geometric relationships can be explicitly formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.194, + 0.47, + 0.213 + ], + "angle": 0, + "content": "\\[\nD _ {o} p _ {o} = K T _ {c \\rightarrow o} D _ {c} K ^ {- 1} p _ {c}, \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.223, + 0.47, + 0.254 + ], + "angle": 0, + "content": "so the position mapping from canonical space to observed space can be easily obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.266, + 0.47, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {\\boldsymbol {c} \\rightarrow \\boldsymbol {o}} \\left(T _ {c \\rightarrow o}, D _ {c}, p _ {c}\\right) = p _ {o}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.295, + 0.471, + 0.417 + ], + "angle": 0, + "content": "Then, we can warp the observed ground-truth lip images \\( I_{o}^{m} \\) into the canonical space by backward warping, denoted by \\( I_{o\\rightarrow c}^{m} \\), to guide the learning of the lip generation model. The C2O Mapping can be similarly defined by \\( T_{o\\rightarrow c} \\) and \\( D_{o} \\). With it, we can warp the generated canonical lip image \\( \\hat{I}_c^m \\) into the observed space by backward warping, denoted by \\( \\hat{I}_{c\\to o}^{m} \\), to compose with the given observed image \\( I_{o} \\) for synthesizing a talking face." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.473, + 0.603 + ], + "angle": 0, + "content": "Pose Controllable Synthesis Thanks to the simultaneously learned complete depth map \\( D_{c} \\), pose-controllable full-head image synthesis according to users' requirements is also supported in our model. Specifically, we employ \\( D_{c} \\) to determine location correspondences based on Eq. 3 and Eq. 5 by altering \\( T_{o} \\). There is a slight difference in this situation, which is to use the forward warping strategy as the ground-truth depths at new given poses are missing, which will produce black holes. To mitigate this issue, we propose a data augmentation method in training by randomly adding black holes with a ratio of \\( 50\\% \\) to the image to let the Blend-Net learn to fill these areas." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.612, + 0.266, + 0.627 + ], + "angle": 0, + "content": "4.4. Overall Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.47, + 0.821 + ], + "angle": 0, + "content": "Image Blending With \\(\\hat{I}_{c\\rightarrow o}^{m}\\), we can integrate it by directly pasting it to the original observed frame based on 2D lip landmarks [3]. However, the paste operation often results in mismatching artifacts in the boundary area, and the image artifacts after data augmentation should also be modified. Therefore, we propose a Blend-Net for blending. The Blend-Net takes the pasted image after the paste operation as input to synthesize harmonized final output \\(\\hat{I}_o\\). Since the target is fusion and amending instead of generation, we predict the pixel residual, which is further added back to the input image to composite the final output image. Detailed network structure is illustrated in the supplementary file." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Synchronization Enhancement. To further improve the synchronization performance, we introduce a pre-trained sync expert network to boost the model's performance in synchronization inspiring by [26]. The sync expert network consists of two pre-trained encoders, which extract features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "of image and speech audio within a sliding window, denoted by \\( i \\) and \\( a \\) respectively. Different from [26], we only have a short video for training. Therefore, the unsynced speech-image pairs are also exploited to construct the contrastive loss which helps avoid falling into a trivial solution. Its effectiveness is also verified in our experiments. The distance of syncsed speech-image pairs is desired to be closer while that of the unsynced speech-image pairs should be farther. Thus, we define a contrastive sync loss \\( \\mathcal{L}_s \\) based on the metric which is widely used in contrastive learning:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.253, + 0.892, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = y \\cdot (1 - \\cos (\\boldsymbol {i}, \\boldsymbol {a})) + (1 - y) \\cdot \\max (0, \\cos (\\boldsymbol {i}, \\boldsymbol {a})), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.356 + ], + "angle": 0, + "content": "where \\( \\cos \\) is the cosine similarity metric. \\( y = 1 \\) and \\( y = 0 \\) represent positive and negative speech-image pairs, respectively. For the input speech, the positive images are the matched images at the same timestamp while the negative images are randomly chosen at some timestamps else." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.364, + 0.69, + 0.381 + ], + "angle": 0, + "content": "4.5. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.463 + ], + "angle": 0, + "content": "Our whole pipeline is trained in a self-supervised manner using the observed frames to provide the supervisory signals. The overall loss includes a canonical mouth image reconstruction loss \\(\\mathcal{L}_m\\), a depth loss \\(\\mathcal{L}_d\\), a sync loss \\(\\mathcal{L}_s\\), and a whole observed image reconstruction loss \\(\\mathcal{L}_w\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.464, + 0.892, + 0.495 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_m\\) measures the reconstruction error between the predicted lip image \\(\\hat{I}_c^m\\) and ground-truth lip image \\(I_{o\\rightarrow c}^{m}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.502, + 0.892, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = \\mathcal {L} _ {p} \\left(\\hat {I} _ {c} ^ {m}, I _ {o \\rightarrow c} ^ {m}\\right) + \\| \\hat {I} _ {c} ^ {m} - I _ {o \\rightarrow c} ^ {m} \\| _ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.592 + ], + "angle": 0, + "content": "where \\( I_{o\\rightarrow c}^{m} \\) is the warped ground truth lip image using O2C space mapping (see the supplementary file for more details) and the widely-used perceptual loss \\( \\mathcal{L}_p \\) is defined in [40]. Similarly, overall reconstruction loss \\( \\mathcal{L}_w \\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.601, + 0.892, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {w} = \\mathcal {L} _ {p} \\left(\\hat {I} _ {o}, I _ {o}\\right) + \\| \\hat {I} _ {o} - I _ {o} \\| _ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.675 + ], + "angle": 0, + "content": "The canonical head depth map is initialized by the incomplete depth map computed by 3DMM and is trained with the help of the photometric loss as" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.684, + 0.892, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\left\\| \\hat {I} _ {o \\rightarrow c} - I _ {c} \\right\\| _ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.892, + 0.743 + ], + "angle": 0, + "content": "where \\(\\hat{I}_{o\\rightarrow c}\\) is the warped predicted image. With the contrastive sync loss (Sec. 4.4), the overall loss function is" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.753, + 0.892, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\omega_ {m} \\cdot \\mathcal {L} _ {m} + \\omega_ {w} \\cdot \\mathcal {L} _ {w} + \\omega_ {d} \\cdot \\mathcal {L} _ {d} + \\omega_ {s} \\cdot \\mathcal {L} _ {s}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.78, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Implementation details are shown in the supplementary file." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.634, + 0.825 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.697, + 0.848 + ], + "angle": 0, + "content": "5.1. Datasets and metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate our algorithm and other methods on three datasets belonging to three specific speakers (Testset I, Testset II and Testset III) as recent speaker-specific" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22172" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.882, + 0.256 + ], + "angle": 0, + "content": "
MethodTrained with large extra dataTestset ITestset IITestset III Sync↑
PSNR↑SSIM↑CPBD↑LMD↓Sync↑PSNR↑SSIM↑CPBD↑LMD↓Sync↑
Ground TruthN/AN/A1.0000.1860.0009.102N/A1.0000.1210.0008.6884.877
ATVG [7]Yes28.4520.8180.0195.4235.81328.0510.6680.0034.2036.2243.571
MakeitTalk [44]Yes29.6920.9060.0504.2154.11528.9960.8130.0614.4635.5592.320
Wav2Lip [26]Yes31.5570.9800.1153.05310.03131.7930.9560.0653.4159.9365.809
PC-AVS [43]Yes29.0720.8800.0404.5959.25828.3590.7340.0464.3058.5865.206
LSP [19]No29.5150.9000.0983.1745.37728.8950.7760.1174.9726.8113.046
AD-NeRF [16]No32.2230.9540.0512.9896.04230.8850.9090.0553.2105.9103.285
DFRF [28]No33.2920.9740.0943.0795.25231.4190.9440.1243.1395.5522.879
Speech2LipNo34.8150.9870.2242.9767.77133.1970.9620.1253.0827.3704.379
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.256, + 0.894, + 0.311 + ], + "angle": 0, + "content": "Table 1. Quantitative results compared with the SOTA methods. Image quality assessment metrics (i.e., PSNR, SSIM, and CPBD) are computed within mouth region. Algorithms are categorized into two classes based on the training datasets for a fair comparison. The first ones are trained in large public datasets while the other ones are trained using only 3-5min videos. The best results of each class are in bold and overall best results are with underlines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.473, + 0.472 + ], + "angle": 0, + "content": "methods [16, 19] usually do. Training videos are collected from [19] as they are publicly available. All video sequences are resampled to 25 FPS and split into training part and test part with a ratio of \\(90\\% / 10\\%\\). The training splits of Testset I \\((500 \\times 500)\\) and Testset II \\((624 \\times 624)\\) are employed to train the models and then same-identity evaluations are conducted on the test splits. Testset III is utilized for cross-identity, -gender, and -language tests. We present more evaluation results in the supplementary file." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.473, + 0.706 + ], + "angle": 0, + "content": "Evaluation Metrics. Since both face generation and lip generation models are contained in our quantitative evaluations, we conduct the image quality assessment just around the mouth region for a fair comparison. To evaluate the appearance quality, we first utilize image quality metrics Peak Signal-to-Noise Ratio (PSNR) and SSIM [37]. A no-reference objective image sharpness metric based on Cumulative Probability of Blur Detection (CPBD) [23, 21, 22] is further utilized to measure the overall perceptual image quality. For synchronization evaluation, Landmarks Distance (LMD) around the mouth region is exploited to compute the accuracy of lip shape following [6]. The confidence score of lip synchronization computed by a pretrained SyncNet [12] is also applied to measure the synchronization performance, labeled by Sync." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.72, + 0.385, + 0.737 + ], + "angle": 0, + "content": "5.2. Comparisons with State of the Arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Same-identity Evaluation. For speaker-independent models, we conduct experiments using their officially-released pre-trained models. Since AD-NeRF [16] and DFRF [28] are speaker-specific models, we retrain them using our training data for a fair comparison. DFRF aims at few-shot learning but its lip synchronization performance witnesses a significant drop when the number of training images decreases from 5000 to 15. Hence, we still train DFRF on thousands of images. Also, we directly use LSP's [19] pretrained models on Testset I and Testset II to conduct experi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.895, + 0.427 + ], + "angle": 0, + "content": "ments (subject May for Testset I and Testset III, and subject Obama2 for Testset II). We cannot provide a quantitative comparison with SSP-NeRF [18] here, as neither codes nor pre-trained weights are available. Thus, for comparison, we extract the speech from their released demo and then present qualitative comparisons in the supplementary file." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.895, + 0.882 + ], + "angle": 0, + "content": "Quantitative results are shown in Table 1. We divide all the models into two types based on the scale of the training dataset. The upper four algorithms (speaker-independent models) are trained using public video datasets with large amounts of identities and speech-image pairs (e.g., LRS2 [1, 11, 13] and VoxCeleb2 [10]) while the latter four methods (speaker-specific models) only use 5-minute videos of a specific identity for training. Speaker-independent models tend to perform well in synchronization because they can learn lip motions well from various speech-visual pairs in training time. However, the cost is sacrificing the visual quality. The metrics of image quality (e.g., PSNR, SSIM, and CPBD) are much lower than that of specific models. Wav2lip [26] is an exception because it only generates lip images, but the details still can not be produced well (Figure 5). In contrast, speaker-specific models usually have much better image qualities, which is favorable in practice. Our model outperforms all the SOTA algorithms in image quality and achieves the best synchronization performance among all the speaker-specific models, and the Sync score is also competitive when compared to those speaker-independent models. It is also worth mentioning that the Sync scores of Wav2lip [26] and PC-AVS [43] are higher than that of the ground truth. The reason might be that the Sync score is sensitive to synchronization quality only when synchronization quality is within a range and a higher score above a threshold may not necessarily imply a higher synchronization quality. We design a toy experiment to verify it in the supplementary file and our user study can also demonstrate it." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Cross-identity Evaluation. Our model also has the abil" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22173" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.168, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.091, + 0.529, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.091, + 0.89, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.893, + 0.358 + ], + "angle": 0, + "content": "Figure 5. Qualitative results compared with SOTA methods. The Lip area is cropped based on the detected 2D landmarks for a clear comparison. Speaker-independent and speaker-specific models are remarked by different colors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.473, + 0.555 + ], + "angle": 0, + "content": "ity to generate cross-identity, -gender, and -language results. Since there is no ground truth image, we provide the sync performance comparison on generating a video for the model trained on Testset I (female, British) using another speaker's speech (male, French). This setting is defined as Testset III, and results are shown in the last column of Table 1. Our model still has significant superiority and outperforms other speaker-specific models by a large margin. We also test the effect of unsynced speech-image pairs used in Eq. 6 on Testset III. Without negative pairs, the Sync score decreases from 4.379 to 3.830, showing the effectiveness of our contrastive sync loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.819 + ], + "angle": 0, + "content": "Qualitative Results. Qualitative comparisons of different algorithms are presented in Figure 5. We can see that speaker-independent methods all suffer from low image quality. Besides, ATVG and MakeiTalk have unnatural speech-insensitive movements (e.g., head motion and background motion) and PC-AVS has an extra issue of struggling with identity preservation. Speaker-specific approaches have much better visual quality but most of them can hardly estimate precise lip details when there exist considerable significant lip motions. And fine details like teeth and tongue are not well modeled either. Besides, AD-NeRF sometimes causes head-torso separation due to the adoption of two separate NeRFs. In contrast, our method Speech2Lip performs well in all these situations, as the generated shapes are more accurate and the synthetic images are more clear and realistic, especially the lip area. More results can be found in the supplementary videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Complexity Comparisons. To further demonstrate the superiority of our model, complexity and computational cost comparisons with speaker-specific models are conducted on Testset I. In Table 2, our model gains the best results with much less complexity and lower computational cost." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.372, + 0.86, + 0.466 + ], + "angle": 0, + "content": "
MethodsLSPAD-NeRFDFRFOurs
Model size (MB) ↓500302030
Train time (hour) ↓38.5806030
Test speed (FPS) ↑350.060.0418
PSNR (dB) ↑29.5232.2233.2934.82
Sync ↑5.386.045.257.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.466, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Table 2. Speaker-specific model comparisons. The best results are in bold and the second best results are with underlines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.746 + ], + "angle": 0, + "content": "User Study. To verify the perceptual quality, user studies are conducted on 16 generated video clips covering both Testset I and Testset II. Videos are generated by our proposed Speech2Lip and other SOTA algorithms, with algorithm names hidden and video order randomized. 17 participants independently evaluated all the videos. For each video, each participant gave a Mean Opinion Score (MOS) from 1-5 for each of three aspects: qualities of speech-visual synchronization, image fidelity, and image realness. Higher scores represent better quality. The overall results are shown in Table 3. Speech2Lip not only achieves the highest MOS in image quality measurement (e.g., 4.618 and 4.382) but also outperforms all the other algorithms in lip synchronization (e.g., 4.265), exhibiting Speech2Lip has the most satisfactory overall video quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.655, + 0.773 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Contributions of loss functions. We first conduct an ablation study on loss functions. The evaluation metrics contain the PSNR around the lip area and the sync score. The PSNR of the whole image is also retained since practical applications attach more attention to the overall quality of the talking portrait scene. Results are shown in Table 4, demonstrating the loss functions all contribute to the increased performance on both visual quality and synchronization." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22174" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.159, + 0.09, + 0.813, + 0.156 + ], + "angle": 0, + "content": "
MethodsATVGMakeitTalkWav2LipPC-AVSLSPAD-NeRFDFRFSpeech2Lip
Lip Synchronization1.5001.8534.1913.0593.1183.0882.8534.265
Image Fidelity1.1472.7943.2652.5293.8823.2353.2654.618
Image Realness1.1182.0883.7652.2653.0002.3242.9714.382
" + }, + { + "type": "table_caption", + "bbox": [ + 0.201, + 0.156, + 0.768, + 0.168 + ], + "angle": 0, + "content": "Table 3. Detailed user study results compared with SOTA methods. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.192, + 0.462, + 0.272 + ], + "angle": 0, + "content": "
MethodsPSNRlip↑PSNRimg↑Sync↑
Lw34.10436.9476.313
Lw + Ld34.42936.9946.359
Lw + Ld + Lm34.69437.1937.194
Lw + Ld + Lm + Ls34.81537.2457.771
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.272, + 0.453, + 0.285 + ], + "angle": 0, + "content": "Table 4. Ablation study results about loss function on Testset I." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.298, + 0.462, + 0.392 + ], + "angle": 0, + "content": "
MethodsPSNRlip↑PSNRimg↑Sync↑
w/o continuous sampling34.70237.2147.553
w/o implicit model34.25037.0586.790
w/o Blend-Net33.73636.8816.065
w/o time34.51237.1217.146
Speech2Lip34.81537.2457.771
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.393, + 0.466, + 0.406 + ], + "angle": 0, + "content": "Table 5. Ablation study results about network design on Testset I." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.567 + ], + "angle": 0, + "content": "Contributions of individual components. We also explore the benefits of our key components by removing each component to see how the performance changes in Table 5. \"w/o implicit model\" represents we employ an explicit lip generation model instead, and \"w/o time\" indicates times-tamp is deleted from inputs. From the results, it can be concluded that each design plays an important role. Among them, the Blend-Net significantly improves video quality, and the design of implicit modeling schema is also critical." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.347, + 0.591 + ], + "angle": 0, + "content": "5.4. Controllable Synthesis Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Thanks to the joint optimization and completion of a full-head depth map \\( D_{c} \\) in GAMEM, our model supports synthesizing full-head images driven by novel head poses. Specifically, \\( D_{c} \\) can be projected into any observed view to obtain \\( D_{o} \\). \\( D_{c} \\) and \\( T_{o} \\) in Eq. (3) are further replaced by \\( D_{o} \\) and the new target head pose, respectively. Then, we adopt Eq. (5) to calculate the correspondence. The correspondence is further exploited to warp the synthesized image in the observed space into a new target space. It is noted that depth maps are used for coarse mapping between spaces while facial expression is more taken charge of by fine-grained contents from observed views. The pose-controllable novel view synthesis results are shown in Figure 7. When testing with moderate pose changes, our model achieves comparable results as AD-NeRF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We note that existing methods with 3D modeling such as AD-NeRF also cannot generate good synthesis results if the evaluated poses deviate too much from the training data. The reason is that the training data widely used in speaker-specific models [19, 16, 28, 18] only show the front view" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.183, + 0.684, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.183, + 0.835, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.296, + 0.832, + 0.31 + ], + "angle": 0, + "content": "Figure 6. 3D Mesh of AD-NeRF on Testset I." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.318, + 0.851, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.381, + 0.718, + 0.39 + ], + "angle": 0, + "content": "(a) AD-NeRF" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.392, + 0.851, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.455, + 0.724, + 0.464 + ], + "angle": 0, + "content": "(b) Speech2Lip" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.462, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Figure 7. Novel view synthesis capability of our model. We rotate the head region with a random angle." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.504, + 0.894, + 0.565 + ], + "angle": 0, + "content": "with limited pose variations, being naturally hard for 3D modeling. The 3D mesh is incomplete, and the surface is rough (Figure 6). In addition, the computational costs of 3D methods are extremely high (Table 2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.585, + 0.619, + 0.6 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.869 + ], + "angle": 0, + "content": "In this paper, we propose a novel decomposition-synthesis-composition framework called Speech2Lip for high-fidelity talking head video synthesis, which disentangles speech-sensitive and speech-insensitive motions/appearances. By presenting a syncsed speech-driven implicit model, a GAMEM module, a Blend-Net, and a contrastive sync loss, we can achieve satisfactory results with only a few minutes of training video. Our framework also supports pose-controllable synthesis. In the future, we plan to study generating realistic expressions driven by speech and explore combining our insights with advanced general image generation methods such as diffusion-based models for better generalizability. We hope that our work inspires more future research in this field and encourages the development of positive applications. However, we also urge caution to prevent any potential abuses. More discussions about limitations are described in the supplementary file." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement This work has been supported by the Research Fund from Tencent ARC lab." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22175" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] T. Afouras, J. S. Chung, A. Senior, O. Vinyals, and A. Zisserman. Deep audio-visual speech recognition. In arXiv:1809.02108, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[3] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[4] Lele Chen, Guofeng Cui, Ziyi Kou, Haitian Zheng, and Chenliang Xu. What comprises a good talking-head video generation? In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.396 + ], + "angle": 0, + "content": "[5] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX, pages 35-51. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.47, + 0.452 + ], + "angle": 0, + "content": "[6] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[7] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.47, + 0.59 + ], + "angle": 0, + "content": "[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[9] Joon Son Chung, Amir Jamaludin, and Andrew Zisserman. You said that? arXiv preprint arXiv:1705.02966, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.469, + 0.661 + ], + "angle": 0, + "content": "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[11] J. S. Chung, A. Senior, O. Vinyals, and A. Zisserman. Lip reading sentences in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[12] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian conference on computer vision, pages 251-263. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[13] J. S. Chung and A. Zisserman. Lip reading in profile. In British Machine Vision Conference, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[14] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision, pages 408-424. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[16] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.162, + 0.893, + 0.217 + ], + "angle": 0, + "content": "[17] Rithesh Kumar, Jose Sotelo, Kundan Kumar, Alexandre de Brébisson, and Yoshua Bengio. Obamanet: Photo-realistic lip-sync from text. arXiv preprint arXiv:1801.01442, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.893, + 0.273 + ], + "angle": 0, + "content": "[18] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. arXiv preprint arXiv:2201.07786, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.273, + 0.893, + 0.315 + ], + "angle": 0, + "content": "[19] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.893, + 0.386 + ], + "angle": 0, + "content": "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.893, + 0.44 + ], + "angle": 0, + "content": "[21] Niranjan Narvekar and Lina J Karam. An improved no-reference sharpness metric based on the probability of blur detection. In Workshop on Video Processing and Quality Metrics, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.893, + 0.509 + ], + "angle": 0, + "content": "[22] Niranjan D Narvekar and Lina J Karam. A no-reference perceptual image sharpness metric based on a cumulative probability of blur detection. In 2009 International Workshop on Quality of Multimedia Experience, pages 87-91. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.51, + 0.893, + 0.566 + ], + "angle": 0, + "content": "[23] Niranjan D Narvekar and Lina J Karam. A no-reference image blur metric based on the cumulative probability of blur detection (cpbd). IEEE Transactions on Image Processing, 20(9):2678-2683, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.566, + 0.893, + 0.636 + ], + "angle": 0, + "content": "[24] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.893, + 0.705 + ], + "angle": 0, + "content": "[25] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.705, + 0.893, + 0.775 + ], + "angle": 0, + "content": "[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.893, + 0.845 + ], + "angle": 0, + "content": "[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[28] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European conference on computer vision, 2022. 2, 3, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22176" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[30] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[31] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[32] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European conference on computer vision, pages 716-731. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[33] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 128(5):1398-1413, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[34] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.488 + ], + "angle": 0, + "content": "[35] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[36] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2531–2539, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[38] Olivia Wiles, A Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Proceedings of the European conference on computer vision (ECCV), pages 670-686, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[39] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[41] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[42] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9299–9306, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.175 + ], + "angle": 0, + "content": "[43] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[44] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22177" + } + ] +] \ No newline at end of file diff --git a/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/b8d7dfde-aec2-4ac3-9ef3-b8789d6534b1_origin.pdf b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/b8d7dfde-aec2-4ac3-9ef3-b8789d6534b1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6670572c905c0e92f97808f64c3b8bdce7ab30ff --- /dev/null +++ b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/b8d7dfde-aec2-4ac3-9ef3-b8789d6534b1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aea8f865d165519f1df6e986edb77511b5c1a41d612c530336a479183bac684b +size 3479160 diff --git a/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/full.md b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..24bec7cdf2f5a33ec4da18d131dd33ace340ab3c --- /dev/null +++ b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/full.md @@ -0,0 +1,296 @@ +# Speech2Lip: High-fidelity Speech to Lip Generation by Learning from a Short Video + +Xiuzhe $\mathsf{W u}^{1}$ , Pengfei $\mathsf{H u}^{2}$ , Yang $\mathsf{W u}^{3,4}$ , Xiaoyang $\mathsf{L y u}^{1}$ , Yan-Pei $\mathsf{C a o}^{3}$ , Ying $\mathsf{S h a n}^{3}$ , Wenming $\mathsf{Y a n g}^{2}$ , Zhongqian $\mathsf{S u n}^{4}$ , Xiaojuan $\mathsf{Q i}^{1}$ , + +1The University of Hong Kong, 2Tsinghua University, 3ARC Lab, Tencent PCG, 4Tencent AI Lab + +# Abstract + +Synthesizing realistic videos according to a given speech is still an open challenge. Previous works have been plagued by issues such as inaccurate lip shape generation and poor image quality. The key reason is that only motions and appearances on limited facial areas (e.g., lip area) are mainly driven by the input speech. Therefore, directly learning a mapping function from speech to the entire head image is prone to ambiguity, particularly when using a short video for training. We thus propose a decomposition-synthesis-composition framework named Speech to Lip (Speech2Lip) that disentangles speech-sensitive and speech-insensitive motion/appearance to facilitate effective learning from limited training data, resulting in the generation of natural-looking videos. First, given a fixed head pose (i.e., canonical space), we present a speech-driven implicit model for lip image generation which concentrates on learning speech-sensitive motion and appearance. Next, to model the major speech-insensitive motion (i.e., head movement), we introduce a geometry-aware mutual explicit mapping (GAMEM) module that establishes geometric mappings between different head poses. This allows us to paste generated lip images at the canonical space onto head images with arbitrary poses and synthesize talking videos with natural head movements. In addition, a Blend-Net and a contrastive sync loss are introduced to enhance the overall synthesis performance. Quantitative and qualitative results on three benchmarks demonstrate that our model can be trained by a video of just a few minutes in length and achieve state-of-the-art performance in both visual quality and speech-visual synchronization. Code: https://github.com/CVMI-Lab/Speech2Lip. + +![](images/4a6576cc4cc0f72ccf3597bdaafdedc96a8ea8445204707a4a2b246f1c96cfb8.jpg) +Figure 1. Given a speech as input, our model generates high-quality talking-head videos and supports pose-controllable synthesis. The decomposition and synthesis modules make learning from a short video more effective and the composition module enables us to synthesize high-fidelity videos. + +# 1. Introduction + +Learning from a short video to generate personalized audio-synchronized talking videos driven by a speech is of great importance to various applications, for instance, digital human animation, video dubbing, and UGC video creation. However, synthesizing high-fidelity videos from speech for a desired speaker remains a challenging task. + +The first challenge arises from complicated motion patterns. Although speech majorly influences lip areas (i.e., speech-sensitive areas), lip movements are often accompanied by other motions, such as global head movements, which greatly impact lip shapes and appearances. Thus, directly synthesizing a whole image from speech often leads to inaccurate lip synthesis. Second, existing methods still struggle to satisfy appearance fidelity requirements, which include both identity preservation (speaker-specific) and high-quality detail generation, such as clear details of teeth, tongue, and eye-blinking [9, 42, 26, 44, 7]. Third, it's difficult to acquire videos longer than 10 hours for a speaker which is yet required by conventional methods [17, 31]. + +To tackle the aforementioned challenges, existing attempts can be coarsely categorized into two lines of research: speaker-independent and speaker-specific methods. The first line often exploits GANs [15] that need to be trained on large-scale multi-person datasets. However, GAN-based models [4, 7, 9, 26, 33, 38, 42, 43] usually synthesize low-resolution images and unnatural motions (i.e., background movements). They thus hardly meet the appearance fidelity requirement in terms of sharpness and fine appearance details. Moreover, preserving the identity of the speaker remains a challenging task [7, 43]. + +For the consideration of high-fidelity and identity preservation, another strategy focuses on learning from a specific speaker. Although attaining high-fidelity results, early works [17, 31] often require several hours of video footage from a speaker for training, hindering their practical applicability. Recently, NeRF [20] has emerged as a promising approach for generating high-fidelity videos, which succeeds in learning from a short video of just a few minutes and having the potential to generate high-fidelity results [16, 18, 28]. Nevertheless, the models still struggle with appearance and motion ambiguity issues because they model speech-sensitive motions/appearances together with other facial areas less correlated to the given speech. This issue becomes more severe when training data is limited since no extra information can be exploited to avoid interference from signals that are not correlated with the speech. As a result, they tend to generate lip sequences that do not synchronize well with the speech and produce blurry images (see Figure 5 and Table 1). Therefore, reducing the complexity of modeling motions is critical to enable effective learning from a short video and synthesizing high-quality videos for a specific speaker. + +We thus design a preliminary experiment to identify that motion and appearance of lip areas have a strong correlation with speech, while head motion and other facial areas are less related to speech (Figure ??). Motivated by the observation, we propose decomposing speech-insensitive motion/appearance from speech-sensitive one, and synthesizing them separately before composing them into a new talking video that aligns with the given speech (Figure 1). Toward this goal, we present a decomposition-synthesis-composition framework named Speech to Lip (Speech2Lip). In the decomposition stage, we introduce a speech-driven implicit model that generates high-fidelity synced lip sequences in a fixed view (i.e., canonical view). To model 3D head motion effectively, we design a Geometry-Aware Mutual Explicit Mapping (GAMEM) module that estimates explicit geometric mappings between an arbitrary observed view and the canonical view. GAMEM also includes a jointly optimized canonical-view full-head depth map, which enables the model to be 3D-aware and supports controllable synthesis driven by new + +head poses. In the composition stage, GAMEM allows us to flexibly paste the synthesized canonical-view lips onto an arbitrary observed view to obtain natural synchronized talking videos. To improve the synthesis and synchronization qualities after composition, we incorporate a blending network (i.e., Blend-Net) to refine the results and a contrastive sync loss to facilitate learning from a short video for generating synchronized talking videos. + +Our major contributions are summarized as follows: + +1) We introduce a novel framework that disentangles speech-sensitive and speech-insensitive motion/appearance in high-fidelity video synthesis. By separating these components, the framework can effectively learn from limited training data. +2) The proposed speech-driven implicit model synthesizes speech-sensitive contents and the GAMEM flexibly combines them with given speech-insensitive areas to generate synchronized talking heads with natural movements and support pose-controllable synthesis. +3) Both qualitative and quantitative experimental results demonstrate the superiority of our method over the state-of-the-art speaker-specific methods. + +# 2. Related Work + +Speech-Driven Talking Face Synthesis. Video synthesis from speech is a long-standing problem. Recent works can be generally divided into two categories: speaker-specific [17, 31, 19, 16, 39, 18, 28] and speaker-independent [9, 42, 26, 44, 7, 5, 41, 35, 36]. In the first track, earlier works have succeeded in obtaining realistic visual results for a target person [17, 31] but required hours of video belonging to one specific speaker. Therefore, many efforts [19, 16, 39, 18, 28] have been made to train the model with a shorter talking video (3-5 minutes). However, they utilize speech to forecast overall motions including speech-insensitive ones, failing to learn accurate lip shapes and appearances. The other line (i.e., speaker-independent method) aims to build a universal model for all identities [9, 42, 26, 44, 7]. The end-to-end pipeline [9, 42] is usually developed on GAN [15]. To boost performance, Richard et al. [27] only synthesize the mesh of limited facial areas, Prajwal et al. [26] present an extra lip-sync discriminator, Zhou et al. [43] provide additional head poses as inputs, and some other models [7, 14, 44, 32, 30] leverage the intermediate representation (e.g., 2D facial landmarks [7, 14, 44] or expression parameters [32, 30]). Nevertheless, they often suffer from low video quality, difficult identity preservation, and abnormal head motion generation. + +Implicit Representation based Talking Head Methods. Recently, implicit representations have shown high model + +ing capabilities in multiple tasks [25, 20, 24, 34]. Among them, NeRF [20] obtains extraordinary performance in novel view synthesis by training on only hundreds of images. Guo et al. [16] first apply it in the speaker-specific talking head synthesis task to enable learning from a short video. However, they utilize two NeRFs to model the head and torso, resulting in the head-torso separation phenomenon. Liu et al. [18] solve it by leveraging one unified NeRF with two semantic-aware modules. Shen et al. [28] incorporate 2D image features as additional inputs to further reduce training data requirements (i.e., videos with only 10-15 seconds). Regardless, these methods only ensure visual quality, and simply use speech to drive all the complex motions, leading to the out-of-sync problem. + +# 3. Empirical Study and Motivations + +Our approach is motivated by the observation that only limited facial areas are highly correlated to speech. To verify this, we conduct a preliminary experiment to determine which areas are most speech-sensitive. Specifically, we apply warping to all captured images with varying head poses (refer to observed views) using the 3D Morphable Model (3DMM) [2] and bring them to a fixed head pose (known as the canonical view). We then compute a motion heatmap in the canonical view to determine the areas most responsive to speech. As depicted in Figure 2, the elimination of head motion results in most of the motions around the lip regions displaying high sensitivity to the input speech. Additional examples can be found in the supplementary file. For more details on the warping process, please refer to Sec. 4.3. + +![](images/af6ab167b3373f3bbf76d9a7778a725e22eae26d77ad4220fe57fab0f27e251a.jpg) +Figure 2. Speech-sensitive Motion Heatmap. + +# 4. Method + +# 4.1. Overview + +Figure 3 provides an overview of our model. We define each frame captured in the observed view with its own head motion as the "observed space". To disentangle speech-sensitive and speech-insensitive motions (i.e., head motions), we randomly select one observed space to serve as the "canonical space". By doing so, we are able to align all observed spaces with the canonical space, thereby eliminating the effects of head motions. Note that once selected, the canonical space is fixed during training and inference. To model speech-sensitive motions, we propose a speech-driven implicit model to generate lip images in the canonical space without head motion effect (Sec. 4.2). Thus, the generated lip image is canonical-view. A Geometry-Aware + +Mutual Explicit Mapping (GAMEM) module is further proposed to model the speech-insensitive head motions without speech effect. Next, we project the lip images in canonical space generated by the above implicit model to the observed space based on GAMEM so that the synthesized lip image can be aligned and composed with any arbitrary observed frame, giving the model the flexibility for diversified composition (Sec. 4.3). Finally, a blending network (i.e., Blend-Net) and a contrastive sync loss are presented to improve the quality of final synthetic images $\hat{I}_o$ (Sec. 4.4). + +# 4.2. Disentangled and Synced Implicit Modeling + +The central part of Speech2Lip is the disentangled and syncsed speech-driven implicit generator (Figure 3(a)), which focuses on generating 2D lip appearances that are synchronized with cross-modal speech. By employing motion and appearance disentanglement, the generator only retains speech-sensitive motions and appearances. A lip-syncing contrastive loss is utilized (Figure 3(c)) so that it can generate syncsed high-fidelity mouth appearance by learning from very short video data. All these together make the implicit modeling disentangled and syncsed. + +Speech-driven Implicit Model To achieve high-fidelity canonical lip image generation, we utilize a speech-driven implicit model. An implicit model is defined as a function that maps the coordinate signal to another signal [29], e.g., color. The input coordinates of the implicit model are defined in a continuous space (i.e., the real field) so that it helps exploit the inherent relationship between spatially adjacent locations. Also, it can be further mapped into the frequency domain and benefit the learning of high-frequency information [20]. Besides, by learning a dedicated model for a specific scenario, implicit models also gain a strong ability to generate high-quality desired outputs. In our model, color information $\hat{c}_{c,n} = (r,g,b)$ is regarded as our output signal (appearance color), and the canonical implicit function $f_{\theta}$ can be defined as + +$$ +\hat {c} _ {c, n} = \boldsymbol {f} _ {\theta} \left(\boldsymbol {x} _ {c, n}, \boldsymbol {a}, \boldsymbol {t} _ {s}\right), \tag {1} +$$ + +where $x_{c,n}$ is the continuous 2D pixel coordinate vector (u,v) in the canonical space. $\pmb{a} \in \mathbb{R}^{64}$ is the feature vector of the input speech at the concerned moment and $t_s$ represents the timestamp information, which is utilized to enhance the temporal consistency between adjacent frames. + +Continuous Sampling The next question that arises is how to obtain the corresponding supervision signal to train the implicit model, which poses a challenge as the coordinates of the pixels are always integers, while we define the coordinates in a continuous space for more expressive features. To overcome this challenge, we adopt the approach developed in [8], which enables us to sample and generate corresponding supervisions in the continuous coordinate space. + +![](images/cd3a30426c3a701cd31a2d098b0f095e111dfda4f46823ca215b7bb432dd11bb.jpg) +Figure 3. The overall framework of Speech2Lip. Our framework decomposes speech-sensitive and speech-insensitive motions/appearances first, models them individually (the major part of (a)), and composes the ultimate output image (b). Besides, the synchronization performance enhancement module and the GAMEM are illustrated in (c) and (d), respectively. The inputs include continuous pixel coordinates, speech audio signals, and timestamps. The speech-driven implicit model will generate speech-sensitive canonical-view lip images, which will be further transformed into observed space to composite the eventual output image. A full-head depth map is learned along with the training process, supporting pose-controllable synthesis. + +![](images/747bdadaa66b16a42672747da5613f67fc2d593df9ca61fa30bbbc045bea7914.jpg) + +Particularly, for each pixel $x_{c}$ , a rectangle with an arbitrary shape is randomly sampled, the four corner points (i.e., $x_{c,n}, n \in \{00, 01, 10, 11\}$ ) of which are regarded as our inputs. A weighted averaged value $\hat{c}_{c}$ is then calculated as + +$$ +\hat {\boldsymbol {c}} _ {c} = \sum_ {n \in \{0 0, 0 1, 1 0, 1 1 \}} \frac {S _ {n}}{S} \cdot \hat {\boldsymbol {c}} _ {c, n}, \tag {2} +$$ + +where $\hat{c}_{c,n}$ are the output appearance values of the sampled points. As illustrated in Figure 4, the blue-dashed-line rectangle represents the area $S$ , which is divided into four sub-rectangles by gray lines, each with an area $S_{n}$ . By leveraging this strategy, we can generate supervision for points with continuous positions, which enables us to fully exploit the advantages of the implicit model and achieve high-fidelity visual results. + +# 4.3. Geometry-Aware Mutual Explicit Mapping + +Once we have obtained the speech-sensitive lip sequences in the canonical view, we can construct the eventually synthesized image $\hat{I}_o$ by assigning the synthesized pixels to a location in the observed space. However, since the assigning process is speech-insensitive and only relies on the geometry information, we introduce the GAMEM module, illustrated in Figure 3, to explicitly model it. The GAMEM module comprises Canonical-to-Observed (C2O), Observed-to-Canonical (O2C) Mapping, and a learnable full-head depth map, which supports various applications, including pose-controllable image synthesis. + +C2O/O2C Mapping These mappings represent how the images transform between two spaces (i.e., canonical space + +![](images/b29d85034d561383250a604e9080fffbb758e88a3fee026ae3de0a125f4169c6.jpg) +Figure 4. Continuous Sampling strategy. This strategy can generate supervision signals for sampled positions at different resolutions in the training time. + +and observed space). As depicted in Figure 3, O2C Mapping aims at creating supervisions $I_{o \to c}^{m}$ for training the lip generator model $f_{\theta}$ and C2O Mapping is utilized to warp the generated lip image into the observed view to produce $\hat{I}_{c \to o}^{m}$ for further composition. As they follow similar principles, we take the O2C Mapping as an example to illustrate the process. The input of O2C Mapping contains $I_{o}$ and $I_{c}$ , and the output are pixel correspondences that map pixel coordinates in observed space to the canonical space. To achieve it, we use the 3D Morphable Model (3DMM) [2] to calculate the overall camera intrinsic matrix $K$ and rough face geometry $G$ . The head poses $T_{o} \in \mathbb{R}^{4 \times 4}$ and $T_{c} \in \mathbb{R}^{4 \times 4}$ , which include a rotation matrix and a translation vector, are also estimated. And the relative head pose $T_{c \to o} \in \mathbb{R}^{4 \times 4}$ between two spaces can be computed as + +$$ +T _ {c \rightarrow o} = T _ {o} \times T _ {c} ^ {- 1}. \tag {3} +$$ + +Furthermore, face depth maps $D_{o}$ and $D_{c}$ for $I_{o}$ and $I_{c}$ obtained from 3DMM can be further interpolated based on $G$ , and the corresponding head poses. Noticing that 3DMM can only be aware of the face area instead of the head area, + +but the information about the rest part of the head (e.g., forehead, ears, eyes, mouth) is lacking. Hence we optimize the $D_{c}$ along with our model to complete the missing depth region. If $p_{o}$ and $p_{c}$ are the corresponding 2D homogeneous pixel grid coordinates on $I_{o}$ and $I_{c}$ , the geometric relationships can be explicitly formulated as + +$$ +D _ {o} p _ {o} = K T _ {c \rightarrow o} D _ {c} K ^ {- 1} p _ {c}, \qquad (4) +$$ + +so the position mapping from canonical space to observed space can be easily obtained as + +$$ +\boldsymbol {F} _ {\boldsymbol {c} \rightarrow \boldsymbol {o}} \left(T _ {c \rightarrow o}, D _ {c}, p _ {c}\right) = p _ {o}. \tag {5} +$$ + +Then, we can warp the observed ground-truth lip images $I_{o}^{m}$ into the canonical space by backward warping, denoted by $I_{o\rightarrow c}^{m}$ , to guide the learning of the lip generation model. The C2O Mapping can be similarly defined by $T_{o\rightarrow c}$ and $D_{o}$ . With it, we can warp the generated canonical lip image $\hat{I}_c^m$ into the observed space by backward warping, denoted by $\hat{I}_{c\to o}^{m}$ , to compose with the given observed image $I_{o}$ for synthesizing a talking face. + +Pose Controllable Synthesis Thanks to the simultaneously learned complete depth map $D_{c}$ , pose-controllable full-head image synthesis according to users' requirements is also supported in our model. Specifically, we employ $D_{c}$ to determine location correspondences based on Eq. 3 and Eq. 5 by altering $T_{o}$ . There is a slight difference in this situation, which is to use the forward warping strategy as the ground-truth depths at new given poses are missing, which will produce black holes. To mitigate this issue, we propose a data augmentation method in training by randomly adding black holes with a ratio of $50\%$ to the image to let the Blend-Net learn to fill these areas. + +# 4.4. Overall Refinement + +Image Blending With $\hat{I}_{c\rightarrow o}^{m}$ , we can integrate it by directly pasting it to the original observed frame based on 2D lip landmarks [3]. However, the paste operation often results in mismatching artifacts in the boundary area, and the image artifacts after data augmentation should also be modified. Therefore, we propose a Blend-Net for blending. The Blend-Net takes the pasted image after the paste operation as input to synthesize harmonized final output $\hat{I}_o$ . Since the target is fusion and amending instead of generation, we predict the pixel residual, which is further added back to the input image to composite the final output image. Detailed network structure is illustrated in the supplementary file. + +Synchronization Enhancement. To further improve the synchronization performance, we introduce a pre-trained sync expert network to boost the model's performance in synchronization inspiring by [26]. The sync expert network consists of two pre-trained encoders, which extract features + +of image and speech audio within a sliding window, denoted by $i$ and $a$ respectively. Different from [26], we only have a short video for training. Therefore, the unsynced speech-image pairs are also exploited to construct the contrastive loss which helps avoid falling into a trivial solution. Its effectiveness is also verified in our experiments. The distance of syncsed speech-image pairs is desired to be closer while that of the unsynced speech-image pairs should be farther. Thus, we define a contrastive sync loss $\mathcal{L}_s$ based on the metric which is widely used in contrastive learning: + +$$ +\mathcal {L} _ {s} = y \cdot (1 - \cos (\boldsymbol {i}, \boldsymbol {a})) + (1 - y) \cdot \max (0, \cos (\boldsymbol {i}, \boldsymbol {a})), \tag {6} +$$ + +where $\cos$ is the cosine similarity metric. $y = 1$ and $y = 0$ represent positive and negative speech-image pairs, respectively. For the input speech, the positive images are the matched images at the same timestamp while the negative images are randomly chosen at some timestamps else. + +# 4.5. Training Objectives + +Our whole pipeline is trained in a self-supervised manner using the observed frames to provide the supervisory signals. The overall loss includes a canonical mouth image reconstruction loss $\mathcal{L}_m$ , a depth loss $\mathcal{L}_d$ , a sync loss $\mathcal{L}_s$ , and a whole observed image reconstruction loss $\mathcal{L}_w$ . + +$\mathcal{L}_m$ measures the reconstruction error between the predicted lip image $\hat{I}_c^m$ and ground-truth lip image $I_{o\rightarrow c}^{m}$ as + +$$ +\mathcal {L} _ {m} = \mathcal {L} _ {p} \left(\hat {I} _ {c} ^ {m}, I _ {o \rightarrow c} ^ {m}\right) + \| \hat {I} _ {c} ^ {m} - I _ {o \rightarrow c} ^ {m} \| _ {2}, \tag {7} +$$ + +where $I_{o\rightarrow c}^{m}$ is the warped ground truth lip image using O2C space mapping (see the supplementary file for more details) and the widely-used perceptual loss $\mathcal{L}_p$ is defined in [40]. Similarly, overall reconstruction loss $\mathcal{L}_w$ is computed as + +$$ +\mathcal {L} _ {w} = \mathcal {L} _ {p} \left(\hat {I} _ {o}, I _ {o}\right) + \| \hat {I} _ {o} - I _ {o} \| _ {2}. \tag {8} +$$ + +The canonical head depth map is initialized by the incomplete depth map computed by 3DMM and is trained with the help of the photometric loss as + +$$ +\mathcal {L} _ {d} = \left\| \hat {I} _ {o \rightarrow c} - I _ {c} \right\| _ {2}, \tag {9} +$$ + +where $\hat{I}_{o\rightarrow c}$ is the warped predicted image. With the contrastive sync loss (Sec. 4.4), the overall loss function is + +$$ +\mathcal {L} = \omega_ {m} \cdot \mathcal {L} _ {m} + \omega_ {w} \cdot \mathcal {L} _ {w} + \omega_ {d} \cdot \mathcal {L} _ {d} + \omega_ {s} \cdot \mathcal {L} _ {s}. \tag {10} +$$ + +Implementation details are shown in the supplementary file. + +# 5. Experiments + +# 5.1. Datasets and metrics + +Datasets. We evaluate our algorithm and other methods on three datasets belonging to three specific speakers (Testset I, Testset II and Testset III) as recent speaker-specific + +
MethodTrained with large extra dataTestset ITestset IITestset III Sync↑
PSNR↑SSIM↑CPBD↑LMD↓Sync↑PSNR↑SSIM↑CPBD↑LMD↓Sync↑
Ground TruthN/AN/A1.0000.1860.0009.102N/A1.0000.1210.0008.6884.877
ATVG [7]Yes28.4520.8180.0195.4235.81328.0510.6680.0034.2036.2243.571
MakeitTalk [44]Yes29.6920.9060.0504.2154.11528.9960.8130.0614.4635.5592.320
Wav2Lip [26]Yes31.5570.9800.1153.05310.03131.7930.9560.0653.4159.9365.809
PC-AVS [43]Yes29.0720.8800.0404.5959.25828.3590.7340.0464.3058.5865.206
LSP [19]No29.5150.9000.0983.1745.37728.8950.7760.1174.9726.8113.046
AD-NeRF [16]No32.2230.9540.0512.9896.04230.8850.9090.0553.2105.9103.285
DFRF [28]No33.2920.9740.0943.0795.25231.4190.9440.1243.1395.5522.879
Speech2LipNo34.8150.9870.2242.9767.77133.1970.9620.1253.0827.3704.379
+ +Table 1. Quantitative results compared with the SOTA methods. Image quality assessment metrics (i.e., PSNR, SSIM, and CPBD) are computed within mouth region. Algorithms are categorized into two classes based on the training datasets for a fair comparison. The first ones are trained in large public datasets while the other ones are trained using only 3-5min videos. The best results of each class are in bold and overall best results are with underlines. + +methods [16, 19] usually do. Training videos are collected from [19] as they are publicly available. All video sequences are resampled to 25 FPS and split into training part and test part with a ratio of $90\% / 10\%$ . The training splits of Testset I $(500 \times 500)$ and Testset II $(624 \times 624)$ are employed to train the models and then same-identity evaluations are conducted on the test splits. Testset III is utilized for cross-identity, -gender, and -language tests. We present more evaluation results in the supplementary file. + +Evaluation Metrics. Since both face generation and lip generation models are contained in our quantitative evaluations, we conduct the image quality assessment just around the mouth region for a fair comparison. To evaluate the appearance quality, we first utilize image quality metrics Peak Signal-to-Noise Ratio (PSNR) and SSIM [37]. A no-reference objective image sharpness metric based on Cumulative Probability of Blur Detection (CPBD) [23, 21, 22] is further utilized to measure the overall perceptual image quality. For synchronization evaluation, Landmarks Distance (LMD) around the mouth region is exploited to compute the accuracy of lip shape following [6]. The confidence score of lip synchronization computed by a pretrained SyncNet [12] is also applied to measure the synchronization performance, labeled by Sync. + +# 5.2. Comparisons with State of the Arts + +Same-identity Evaluation. For speaker-independent models, we conduct experiments using their officially-released pre-trained models. Since AD-NeRF [16] and DFRF [28] are speaker-specific models, we retrain them using our training data for a fair comparison. DFRF aims at few-shot learning but its lip synchronization performance witnesses a significant drop when the number of training images decreases from 5000 to 15. Hence, we still train DFRF on thousands of images. Also, we directly use LSP's [19] pretrained models on Testset I and Testset II to conduct experi + +ments (subject May for Testset I and Testset III, and subject Obama2 for Testset II). We cannot provide a quantitative comparison with SSP-NeRF [18] here, as neither codes nor pre-trained weights are available. Thus, for comparison, we extract the speech from their released demo and then present qualitative comparisons in the supplementary file. + +Quantitative results are shown in Table 1. We divide all the models into two types based on the scale of the training dataset. The upper four algorithms (speaker-independent models) are trained using public video datasets with large amounts of identities and speech-image pairs (e.g., LRS2 [1, 11, 13] and VoxCeleb2 [10]) while the latter four methods (speaker-specific models) only use 5-minute videos of a specific identity for training. Speaker-independent models tend to perform well in synchronization because they can learn lip motions well from various speech-visual pairs in training time. However, the cost is sacrificing the visual quality. The metrics of image quality (e.g., PSNR, SSIM, and CPBD) are much lower than that of specific models. Wav2lip [26] is an exception because it only generates lip images, but the details still can not be produced well (Figure 5). In contrast, speaker-specific models usually have much better image qualities, which is favorable in practice. Our model outperforms all the SOTA algorithms in image quality and achieves the best synchronization performance among all the speaker-specific models, and the Sync score is also competitive when compared to those speaker-independent models. It is also worth mentioning that the Sync scores of Wav2lip [26] and PC-AVS [43] are higher than that of the ground truth. The reason might be that the Sync score is sensitive to synchronization quality only when synchronization quality is within a range and a higher score above a threshold may not necessarily imply a higher synchronization quality. We design a toy experiment to verify it in the supplementary file and our user study can also demonstrate it. + +Cross-identity Evaluation. Our model also has the abil + +![](images/54512ee06f077ac6673d5030f7d533444c5e53634b0733abd8ebb7ddfaa7bfba.jpg) +Figure 5. Qualitative results compared with SOTA methods. The Lip area is cropped based on the detected 2D landmarks for a clear comparison. Speaker-independent and speaker-specific models are remarked by different colors. + +![](images/85cffeeca06aad9ce7bb1d34f3ef9f6984bfc5548b2ef9e4c70d9d69a3bbec2e.jpg) + +![](images/4885db8e54c2ee536f28575f4512b8534e61893a95201e143fb9b23c9bf49076.jpg) + +ity to generate cross-identity, -gender, and -language results. Since there is no ground truth image, we provide the sync performance comparison on generating a video for the model trained on Testset I (female, British) using another speaker's speech (male, French). This setting is defined as Testset III, and results are shown in the last column of Table 1. Our model still has significant superiority and outperforms other speaker-specific models by a large margin. We also test the effect of unsynced speech-image pairs used in Eq. 6 on Testset III. Without negative pairs, the Sync score decreases from 4.379 to 3.830, showing the effectiveness of our contrastive sync loss. + +Qualitative Results. Qualitative comparisons of different algorithms are presented in Figure 5. We can see that speaker-independent methods all suffer from low image quality. Besides, ATVG and MakeiTalk have unnatural speech-insensitive movements (e.g., head motion and background motion) and PC-AVS has an extra issue of struggling with identity preservation. Speaker-specific approaches have much better visual quality but most of them can hardly estimate precise lip details when there exist considerable significant lip motions. And fine details like teeth and tongue are not well modeled either. Besides, AD-NeRF sometimes causes head-torso separation due to the adoption of two separate NeRFs. In contrast, our method Speech2Lip performs well in all these situations, as the generated shapes are more accurate and the synthetic images are more clear and realistic, especially the lip area. More results can be found in the supplementary videos. + +Complexity Comparisons. To further demonstrate the superiority of our model, complexity and computational cost comparisons with speaker-specific models are conducted on Testset I. In Table 2, our model gains the best results with much less complexity and lower computational cost. + +
MethodsLSPAD-NeRFDFRFOurs
Model size (MB) ↓500302030
Train time (hour) ↓38.5806030
Test speed (FPS) ↑350.060.0418
PSNR (dB) ↑29.5232.2233.2934.82
Sync ↑5.386.045.257.77
+ +Table 2. Speaker-specific model comparisons. The best results are in bold and the second best results are with underlines. + +User Study. To verify the perceptual quality, user studies are conducted on 16 generated video clips covering both Testset I and Testset II. Videos are generated by our proposed Speech2Lip and other SOTA algorithms, with algorithm names hidden and video order randomized. 17 participants independently evaluated all the videos. For each video, each participant gave a Mean Opinion Score (MOS) from 1-5 for each of three aspects: qualities of speech-visual synchronization, image fidelity, and image realness. Higher scores represent better quality. The overall results are shown in Table 3. Speech2Lip not only achieves the highest MOS in image quality measurement (e.g., 4.618 and 4.382) but also outperforms all the other algorithms in lip synchronization (e.g., 4.265), exhibiting Speech2Lip has the most satisfactory overall video quality. + +# 5.3. Ablation Study + +Contributions of loss functions. We first conduct an ablation study on loss functions. The evaluation metrics contain the PSNR around the lip area and the sync score. The PSNR of the whole image is also retained since practical applications attach more attention to the overall quality of the talking portrait scene. Results are shown in Table 4, demonstrating the loss functions all contribute to the increased performance on both visual quality and synchronization. + +
MethodsATVGMakeitTalkWav2LipPC-AVSLSPAD-NeRFDFRFSpeech2Lip
Lip Synchronization1.5001.8534.1913.0593.1183.0882.8534.265
Image Fidelity1.1472.7943.2652.5293.8823.2353.2654.618
Image Realness1.1182.0883.7652.2653.0002.3242.9714.382
+ +Table 3. Detailed user study results compared with SOTA methods. The best results are in bold. + +
MethodsPSNRlip↑PSNRimg↑Sync↑
Lw34.10436.9476.313
Lw + Ld34.42936.9946.359
Lw + Ld + Lm34.69437.1937.194
Lw + Ld + Lm + Ls34.81537.2457.771
+ +Table 4. Ablation study results about loss function on Testset I. + +
MethodsPSNRlip↑PSNRimg↑Sync↑
w/o continuous sampling34.70237.2147.553
w/o implicit model34.25037.0586.790
w/o Blend-Net33.73636.8816.065
w/o time34.51237.1217.146
Speech2Lip34.81537.2457.771
+ +Table 5. Ablation study results about network design on Testset I. + +Contributions of individual components. We also explore the benefits of our key components by removing each component to see how the performance changes in Table 5. "w/o implicit model" represents we employ an explicit lip generation model instead, and "w/o time" indicates times-tamp is deleted from inputs. From the results, it can be concluded that each design plays an important role. Among them, the Blend-Net significantly improves video quality, and the design of implicit modeling schema is also critical. + +# 5.4. Controllable Synthesis Results + +Thanks to the joint optimization and completion of a full-head depth map $D_{c}$ in GAMEM, our model supports synthesizing full-head images driven by novel head poses. Specifically, $D_{c}$ can be projected into any observed view to obtain $D_{o}$ . $D_{c}$ and $T_{o}$ in Eq. (3) are further replaced by $D_{o}$ and the new target head pose, respectively. Then, we adopt Eq. (5) to calculate the correspondence. The correspondence is further exploited to warp the synthesized image in the observed space into a new target space. It is noted that depth maps are used for coarse mapping between spaces while facial expression is more taken charge of by fine-grained contents from observed views. The pose-controllable novel view synthesis results are shown in Figure 7. When testing with moderate pose changes, our model achieves comparable results as AD-NeRF. + +We note that existing methods with 3D modeling such as AD-NeRF also cannot generate good synthesis results if the evaluated poses deviate too much from the training data. The reason is that the training data widely used in speaker-specific models [19, 16, 28, 18] only show the front view + +![](images/fe3e1b2368d9dd5d8ac162519478c0aa7315a28e3ff6b00e8346834a3fda973c.jpg) +Figure 6. 3D Mesh of AD-NeRF on Testset I. + +![](images/b65d7199e3c382b4c1ed709256cb5fdf92e81f522b603c15fe6c98f4cc085c28.jpg) + +![](images/9adf590cf715a8fd37d35fa7d3fa95b7f0d64a6670283ff5e075ba4d322f5e5c.jpg) +(a) AD-NeRF + +![](images/fcdee3442b9ba562bb89e6e2979c44d8d7aefbea96911ed622dd9973c9654cc7.jpg) +(b) Speech2Lip +Figure 7. Novel view synthesis capability of our model. We rotate the head region with a random angle. + +with limited pose variations, being naturally hard for 3D modeling. The 3D mesh is incomplete, and the surface is rough (Figure 6). In addition, the computational costs of 3D methods are extremely high (Table 2). + +# 6. Conclusion + +In this paper, we propose a novel decomposition-synthesis-composition framework called Speech2Lip for high-fidelity talking head video synthesis, which disentangles speech-sensitive and speech-insensitive motions/appearances. By presenting a syncsed speech-driven implicit model, a GAMEM module, a Blend-Net, and a contrastive sync loss, we can achieve satisfactory results with only a few minutes of training video. Our framework also supports pose-controllable synthesis. In the future, we plan to study generating realistic expressions driven by speech and explore combining our insights with advanced general image generation methods such as diffusion-based models for better generalizability. We hope that our work inspires more future research in this field and encourages the development of positive applications. However, we also urge caution to prevent any potential abuses. More discussions about limitations are described in the supplementary file. + +Acknowledgement This work has been supported by the Research Fund from Tencent ARC lab. + +# References + +[1] T. Afouras, J. S. Chung, A. Senior, O. Vinyals, and A. Zisserman. Deep audio-visual speech recognition. In arXiv:1809.02108, 2018. 6 +[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3, 4 +[3] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 5 +[4] Lele Chen, Guofeng Cui, Ziyi Kou, Haitian Zheng, and Chenliang Xu. What comprises a good talking-head video generation? In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 2 +[5] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX, pages 35-51. Springer, 2020. 2 +[6] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 6 +[7] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. 1, 2, 6 +[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 3 +[9] Joon Son Chung, Amir Jamaludin, and Andrew Zisserman. You said that? arXiv preprint arXiv:1705.02966, 2017. 1, 2 +[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 6 +[11] J. S. Chung, A. Senior, O. Vinyals, and A. Zisserman. Lip reading sentences in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 6 +[12] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian conference on computer vision, pages 251-263. Springer, 2016. 6 +[13] J. S. Chung and A. Zisserman. Lip reading in profile. In British Machine Vision Conference, 2017. 6 +[14] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision, pages 408-424. Springer, 2020. 2 +[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 + +[16] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 2, 3, 6, 8 +[17] Rithesh Kumar, Jose Sotelo, Kundan Kumar, Alexandre de Brébisson, and Yoshua Bengio. Obamanet: Photo-realistic lip-sync from text. arXiv preprint arXiv:1801.01442, 2017. 1, 2 +[18] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. arXiv preprint arXiv:2201.07786, 2022. 2, 3, 6, 8 +[19] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2, 6, 8 +[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3 +[21] Niranjan Narvekar and Lina J Karam. An improved no-reference sharpness metric based on the probability of blur detection. In Workshop on Video Processing and Quality Metrics, 2010. 6 +[22] Niranjan D Narvekar and Lina J Karam. A no-reference perceptual image sharpness metric based on a cumulative probability of blur detection. In 2009 International Workshop on Quality of Multimedia Experience, pages 87-91. IEEE, 2009. 6 +[23] Niranjan D Narvekar and Lina J Karam. A no-reference image blur metric based on the cumulative probability of blur detection (cpbd). IEEE Transactions on Image Processing, 20(9):2678-2683, 2011. 6 +[24] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 3 +[25] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3 +[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1, 2, 5, 6 +[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 2 +[28] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European conference on computer vision, 2022. 2, 3, 6, 8 + +[29] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 3 +[30] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 2022. 2 +[31] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 1, 2 +[32] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European conference on computer vision, pages 716-731. Springer, 2020. 2 +[33] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 128(5):1398-1413, 2020. 2 +[34] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3 +[35] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2 +[36] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2531–2539, 2022. 2 +[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[38] Olivia Wiles, A Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Proceedings of the European conference on computer vision (ECCV), pages 670-686, 2018. 2 +[39] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 2 +[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[41] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2 +[42] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9299–9306, 2019. 1, 2 + +[43] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 2, 6 +[44] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 1, 2, 6 \ No newline at end of file diff --git a/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/images.zip b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..af26386941b6c1e07b89ccff3ff74b70ef92af22 --- /dev/null +++ b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93ff7cfd27e2f136a83f7f638dac889bfaf680d2fe059c39471d2a0eb438c30b +size 529727 diff --git a/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/layout.json b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6705623caa40f1d5d486c9f1ad93df9ae74d5bfd --- /dev/null +++ b/2023/Speech2Lip_ High-fidelity Speech to Lip Generation by Learning from a Short Video/layout.json @@ -0,0 +1,7823 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 539, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 539, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 539, + 138 + ], + "type": "text", + "content": "Speech2Lip: High-fidelity Speech to Lip Generation by Learning from a Short Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "content": "Xiuzhe " + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "inline_equation", + "content": "\\mathsf{W u}^{1}" + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "content": ", Pengfei " + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "inline_equation", + "content": "\\mathsf{H u}^{2}" + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "content": ", Yang " + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "inline_equation", + "content": "\\mathsf{W u}^{3,4}" + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "content": ", Xiaoyang " + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "inline_equation", + "content": "\\mathsf{L y u}^{1}" + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "content": ", Yan-Pei " + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "inline_equation", + "content": "\\mathsf{C a o}^{3}" + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "content": ", Ying " + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "inline_equation", + "content": "\\mathsf{S h a n}^{3}" + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "content": ", Wenming " + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "inline_equation", + "content": "\\mathsf{Y a n g}^{2}" + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "content": ", Zhongqian " + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "inline_equation", + "content": "\\mathsf{S u n}^{4}" + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "content": ", Xiaojuan " + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "inline_equation", + "content": "\\mathsf{Q i}^{1}" + }, + { + "bbox": [ + 61, + 170, + 533, + 200 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 537, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 537, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 537, + 233 + ], + "type": "text", + "content": "1The University of Hong Kong, 2Tsinghua University, 3ARC Lab, Tencent PCG, 4Tencent AI Lab" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 329, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 289, + 700 + ], + "type": "text", + "content": "Synthesizing realistic videos according to a given speech is still an open challenge. Previous works have been plagued by issues such as inaccurate lip shape generation and poor image quality. The key reason is that only motions and appearances on limited facial areas (e.g., lip area) are mainly driven by the input speech. Therefore, directly learning a mapping function from speech to the entire head image is prone to ambiguity, particularly when using a short video for training. We thus propose a decomposition-synthesis-composition framework named Speech to Lip (Speech2Lip) that disentangles speech-sensitive and speech-insensitive motion/appearance to facilitate effective learning from limited training data, resulting in the generation of natural-looking videos. First, given a fixed head pose (i.e., canonical space), we present a speech-driven implicit model for lip image generation which concentrates on learning speech-sensitive motion and appearance. Next, to model the major speech-insensitive motion (i.e., head movement), we introduce a geometry-aware mutual explicit mapping (GAMEM) module that establishes geometric mappings between different head poses. This allows us to paste generated lip images at the canonical space onto head images with arbitrary poses and synthesize talking videos with natural head movements. In addition, a Blend-Net and a contrastive sync loss are introduced to enhance the overall synthesis performance. Quantitative and qualitative results on three benchmarks demonstrate that our model can be trained by a video of just a few minutes in length and achieve state-of-the-art performance in both visual quality and speech-visual synchronization. Code: https://github.com/CVMI-Lab/Speech2Lip." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 268, + 547, + 383 + ], + "blocks": [ + { + "bbox": [ + 307, + 268, + 547, + 383 + ], + "lines": [ + { + "bbox": [ + 307, + 268, + 547, + 383 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 547, + 383 + ], + "type": "image", + "image_path": "4a6576cc4cc0f72ccf3597bdaafdedc96a8ea8445204707a4a2b246f1c96cfb8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 384, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 305, + 384, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 547, + 440 + ], + "type": "text", + "content": "Figure 1. Given a speech as input, our model generates high-quality talking-head videos and supports pose-controllable synthesis. The decomposition and synthesis modules make learning from a short video more effective and the composition module enables us to synthesize high-fidelity videos." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 462, + 387, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 462, + 387, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 462, + 387, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 484, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 556 + ], + "type": "text", + "content": "Learning from a short video to generate personalized audio-synchronized talking videos driven by a speech is of great importance to various applications, for instance, digital human animation, video dubbing, and UGC video creation. However, synthesizing high-fidelity videos from speech for a desired speaker remains a challenging task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "The first challenge arises from complicated motion patterns. Although speech majorly influences lip areas (i.e., speech-sensitive areas), lip movements are often accompanied by other motions, such as global head movements, which greatly impact lip shapes and appearances. Thus, directly synthesizing a whole image from speech often leads to inaccurate lip synthesis. Second, existing methods still struggle to satisfy appearance fidelity requirements, which include both identity preservation (speaker-specific) and high-quality detail generation, such as clear details of teeth, tongue, and eye-blinking [9, 42, 26, 44, 7]. Third, it's difficult to acquire videos longer than 10 hours for a speaker which is yet required by conventional methods [17, 31]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22168" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "To tackle the aforementioned challenges, existing attempts can be coarsely categorized into two lines of research: speaker-independent and speaker-specific methods. The first line often exploits GANs [15] that need to be trained on large-scale multi-person datasets. However, GAN-based models [4, 7, 9, 26, 33, 38, 42, 43] usually synthesize low-resolution images and unnatural motions (i.e., background movements). They thus hardly meet the appearance fidelity requirement in terms of sharpness and fine appearance details. Moreover, preserving the identity of the speaker remains a challenging task [7, 43]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 472 + ], + "type": "text", + "content": "For the consideration of high-fidelity and identity preservation, another strategy focuses on learning from a specific speaker. Although attaining high-fidelity results, early works [17, 31] often require several hours of video footage from a speaker for training, hindering their practical applicability. Recently, NeRF [20] has emerged as a promising approach for generating high-fidelity videos, which succeeds in learning from a short video of just a few minutes and having the potential to generate high-fidelity results [16, 18, 28]. Nevertheless, the models still struggle with appearance and motion ambiguity issues because they model speech-sensitive motions/appearances together with other facial areas less correlated to the given speech. This issue becomes more severe when training data is limited since no extra information can be exploited to avoid interference from signals that are not correlated with the speech. As a result, they tend to generate lip sequences that do not synchronize well with the speech and produce blurry images (see Figure 5 and Table 1). Therefore, reducing the complexity of modeling motions is critical to enable effective learning from a short video and synthesizing high-quality videos for a specific speaker." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": "We thus design a preliminary experiment to identify that motion and appearance of lip areas have a strong correlation with speech, while head motion and other facial areas are less related to speech (Figure ??). Motivated by the observation, we propose decomposing speech-insensitive motion/appearance from speech-sensitive one, and synthesizing them separately before composing them into a new talking video that aligns with the given speech (Figure 1). Toward this goal, we present a decomposition-synthesis-composition framework named Speech to Lip (Speech2Lip). In the decomposition stage, we introduce a speech-driven implicit model that generates high-fidelity synced lip sequences in a fixed view (i.e., canonical view). To model 3D head motion effectively, we design a Geometry-Aware Mutual Explicit Mapping (GAMEM) module that estimates explicit geometric mappings between an arbitrary observed view and the canonical view. GAMEM also includes a jointly optimized canonical-view full-head depth map, which enables the model to be 3D-aware and supports controllable synthesis driven by new" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": "head poses. In the composition stage, GAMEM allows us to flexibly paste the synthesized canonical-view lips onto an arbitrary observed view to obtain natural synchronized talking videos. To improve the synthesis and synchronization qualities after composition, we incorporate a blending network (i.e., Blend-Net) to refine the results and a contrastive sync loss to facilitate learning from a short video for generating synchronized talking videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 168, + 529, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 168, + 529, + 179 + ], + "spans": [ + { + "bbox": [ + 317, + 168, + 529, + 179 + ], + "type": "text", + "content": "Our major contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 312, + 184, + 545, + 356 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 184, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 545, + 243 + ], + "type": "text", + "content": "1) We introduce a novel framework that disentangles speech-sensitive and speech-insensitive motion/appearance in high-fidelity video synthesis. By separating these components, the framework can effectively learn from limited training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 246, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 246, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 246, + 545, + 316 + ], + "type": "text", + "content": "2) The proposed speech-driven implicit model synthesizes speech-sensitive contents and the GAMEM flexibly combines them with given speech-insensitive areas to generate synchronized talking heads with natural movements and support pose-controllable synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 320, + 545, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 320, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 312, + 320, + 545, + 356 + ], + "type": "text", + "content": "3) Both qualitative and quantitative experimental results demonstrate the superiority of our method over the state-of-the-art speaker-specific methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 365, + 392, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 392, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 392, + 377 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 388, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 677 + ], + "type": "text", + "content": "Speech-Driven Talking Face Synthesis. Video synthesis from speech is a long-standing problem. Recent works can be generally divided into two categories: speaker-specific [17, 31, 19, 16, 39, 18, 28] and speaker-independent [9, 42, 26, 44, 7, 5, 41, 35, 36]. In the first track, earlier works have succeeded in obtaining realistic visual results for a target person [17, 31] but required hours of video belonging to one specific speaker. Therefore, many efforts [19, 16, 39, 18, 28] have been made to train the model with a shorter talking video (3-5 minutes). However, they utilize speech to forecast overall motions including speech-insensitive ones, failing to learn accurate lip shapes and appearances. The other line (i.e., speaker-independent method) aims to build a universal model for all identities [9, 42, 26, 44, 7]. The end-to-end pipeline [9, 42] is usually developed on GAN [15]. To boost performance, Richard et al. [27] only synthesize the mesh of limited facial areas, Prajwal et al. [26] present an extra lip-sync discriminator, Zhou et al. [43] provide additional head poses as inputs, and some other models [7, 14, 44, 32, 30] leverage the intermediate representation (e.g., 2D facial landmarks [7, 14, 44] or expression parameters [32, 30]). Nevertheless, they often suffer from low video quality, difficult identity preservation, and abnormal head motion generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Implicit Representation based Talking Head Methods. Recently, implicit representations have shown high model" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22169" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "ing capabilities in multiple tasks [25, 20, 24, 34]. Among them, NeRF [20] obtains extraordinary performance in novel view synthesis by training on only hundreds of images. Guo et al. [16] first apply it in the speaker-specific talking head synthesis task to enable learning from a short video. However, they utilize two NeRFs to model the head and torso, resulting in the head-torso separation phenomenon. Liu et al. [18] solve it by leveraging one unified NeRF with two semantic-aware modules. Shen et al. [28] incorporate 2D image features as additional inputs to further reduce training data requirements (i.e., videos with only 10-15 seconds). Regardless, these methods only ensure visual quality, and simply use speech to drive all the complex motions, leading to the out-of-sync problem." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 249, + 234, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 249, + 234, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 234, + 263 + ], + "type": "text", + "content": "3. Empirical Study and Motivations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 269, + 290, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 290, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 290, + 437 + ], + "type": "text", + "content": "Our approach is motivated by the observation that only limited facial areas are highly correlated to speech. To verify this, we conduct a preliminary experiment to determine which areas are most speech-sensitive. Specifically, we apply warping to all captured images with varying head poses (refer to observed views) using the 3D Morphable Model (3DMM) [2] and bring them to a fixed head pose (known as the canonical view). We then compute a motion heatmap in the canonical view to determine the areas most responsive to speech. As depicted in Figure 2, the elimination of head motion results in most of the motions around the lip regions displaying high sensitivity to the input speech. Additional examples can be found in the supplementary file. For more details on the warping process, please refer to Sec. 4.3." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 445, + 263, + 510 + ], + "blocks": [ + { + "bbox": [ + 70, + 445, + 263, + 510 + ], + "lines": [ + { + "bbox": [ + 70, + 445, + 263, + 510 + ], + "spans": [ + { + "bbox": [ + 70, + 445, + 263, + 510 + ], + "type": "image", + "image_path": "af6ab167b3373f3bbf76d9a7778a725e22eae26d77ad4220fe57fab0f27e251a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 509, + 248, + 519 + ], + "lines": [ + { + "bbox": [ + 85, + 509, + 248, + 519 + ], + "spans": [ + { + "bbox": [ + 85, + 509, + 248, + 519 + ], + "type": "text", + "content": "Figure 2. Speech-sensitive Motion Heatmap." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 520, + 103, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 103, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 103, + 532 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 540, + 115, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 115, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 115, + 551 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "Figure 3 provides an overview of our model. We define each frame captured in the observed view with its own head motion as the \"observed space\". To disentangle speech-sensitive and speech-insensitive motions (i.e., head motions), we randomly select one observed space to serve as the \"canonical space\". By doing so, we are able to align all observed spaces with the canonical space, thereby eliminating the effects of head motions. Note that once selected, the canonical space is fixed during training and inference. To model speech-sensitive motions, we propose a speech-driven implicit model to generate lip images in the canonical space without head motion effect (Sec. 4.2). Thus, the generated lip image is canonical-view. A Geometry-Aware" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "Mutual Explicit Mapping (GAMEM) module is further proposed to model the speech-insensitive head motions without speech effect. Next, we project the lip images in canonical space generated by the above implicit model to the observed space based on GAMEM so that the synthesized lip image can be aligned and composed with any arbitrary observed frame, giving the model the flexibility for diversified composition (Sec. 4.3). Finally, a blending network (i.e., Blend-Net) and a contrastive sync loss are presented to improve the quality of final synthetic images " + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "inline_equation", + "content": "\\hat{I}_o" + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": " (Sec. 4.4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 200, + 534, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 200, + 534, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 200, + 534, + 213 + ], + "type": "text", + "content": "4.2. Disentangled and Synced Implicit Modeling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 218, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 339 + ], + "type": "text", + "content": "The central part of Speech2Lip is the disentangled and syncsed speech-driven implicit generator (Figure 3(a)), which focuses on generating 2D lip appearances that are synchronized with cross-modal speech. By employing motion and appearance disentanglement, the generator only retains speech-sensitive motions and appearances. A lip-syncing contrastive loss is utilized (Figure 3(c)) so that it can generate syncsed high-fidelity mouth appearance by learning from very short video data. All these together make the implicit modeling disentangled and syncsed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "text", + "content": "Speech-driven Implicit Model To achieve high-fidelity canonical lip image generation, we utilize a speech-driven implicit model. An implicit model is defined as a function that maps the coordinate signal to another signal [29], e.g., color. The input coordinates of the implicit model are defined in a continuous space (i.e., the real field) so that it helps exploit the inherent relationship between spatially adjacent locations. Also, it can be further mapped into the frequency domain and benefit the learning of high-frequency information [20]. Besides, by learning a dedicated model for a specific scenario, implicit models also gain a strong ability to generate high-quality desired outputs. In our model, color information " + }, + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\hat{c}_{c,n} = (r,g,b)" + }, + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "text", + "content": " is regarded as our output signal (appearance color), and the canonical implicit function " + }, + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 342, + 546, + 521 + ], + "type": "text", + "content": " can be defined as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 376, + 532, + 545, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 532, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 376, + 532, + 545, + 545 + ], + "type": "interline_equation", + "content": "\\hat {c} _ {c, n} = \\boldsymbol {f} _ {\\theta} \\left(\\boldsymbol {x} _ {c, n}, \\boldsymbol {a}, \\boldsymbol {t} _ {s}\\right), \\tag {1}", + "image_path": "85e5156d7aefd0d9038d1698b8ba6d41d96ef44ccec12af50cba668589b02f6b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "inline_equation", + "content": "x_{c,n}" + }, + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "text", + "content": " is the continuous 2D pixel coordinate vector (u,v) in the canonical space. " + }, + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "inline_equation", + "content": "\\pmb{a} \\in \\mathbb{R}^{64}" + }, + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "text", + "content": " is the feature vector of the input speech at the concerned moment and " + }, + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "text", + "content": " represents the timestamp information, which is utilized to enhance the temporal consistency between adjacent frames." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Continuous Sampling The next question that arises is how to obtain the corresponding supervision signal to train the implicit model, which poses a challenge as the coordinates of the pixels are always integers, while we define the coordinates in a continuous space for more expressive features. To overcome this challenge, we adopt the approach developed in [8], which enables us to sample and generate corresponding supervisions in the continuous coordinate space." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22170" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 350, + 247 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 350, + 247 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 350, + 247 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 350, + 247 + ], + "type": "image", + "image_path": "cd3a30426c3a701cd31a2d098b0f095e111dfda4f46823ca215b7bb432dd11bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 548, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 548, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 548, + 319 + ], + "type": "text", + "content": "Figure 3. The overall framework of Speech2Lip. Our framework decomposes speech-sensitive and speech-insensitive motions/appearances first, models them individually (the major part of (a)), and composes the ultimate output image (b). Besides, the synchronization performance enhancement module and the GAMEM are illustrated in (c) and (d), respectively. The inputs include continuous pixel coordinates, speech audio signals, and timestamps. The speech-driven implicit model will generate speech-sensitive canonical-view lip images, which will be further transformed into observed space to composite the eventual output image. A full-head depth map is learned along with the training process, supporting pose-controllable synthesis." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 350, + 72, + 544, + 247 + ], + "blocks": [ + { + "bbox": [ + 350, + 72, + 544, + 247 + ], + "lines": [ + { + "bbox": [ + 350, + 72, + 544, + 247 + ], + "spans": [ + { + "bbox": [ + 350, + 72, + 544, + 247 + ], + "type": "image", + "image_path": "747bdadaa66b16a42672747da5613f67fc2d593df9ca61fa30bbbc045bea7914.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 338, + 288, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 288, + 387 + ], + "type": "text", + "content": "Particularly, for each pixel " + }, + { + "bbox": [ + 47, + 338, + 288, + 387 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 47, + 338, + 288, + 387 + ], + "type": "text", + "content": ", a rectangle with an arbitrary shape is randomly sampled, the four corner points (i.e., " + }, + { + "bbox": [ + 47, + 338, + 288, + 387 + ], + "type": "inline_equation", + "content": "x_{c,n}, n \\in \\{00, 01, 10, 11\\}" + }, + { + "bbox": [ + 47, + 338, + 288, + 387 + ], + "type": "text", + "content": ") of which are regarded as our inputs. A weighted averaged value " + }, + { + "bbox": [ + 47, + 338, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\hat{c}_{c}" + }, + { + "bbox": [ + 47, + 338, + 288, + 387 + ], + "type": "text", + "content": " is then calculated as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 393, + 287, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 393, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 102, + 393, + 287, + 425 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {c}} _ {c} = \\sum_ {n \\in \\{0 0, 0 1, 1 0, 1 1 \\}} \\frac {S _ {n}}{S} \\cdot \\hat {\\boldsymbol {c}} _ {c, n}, \\tag {2}", + "image_path": "5dd90119dcce17239f8b24b2a675feecfc5988e8f4bcc7d835b5bd14c1715e22.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\hat{c}_{c,n}" + }, + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "type": "text", + "content": " are the output appearance values of the sampled points. As illustrated in Figure 4, the blue-dashed-line rectangle represents the area " + }, + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "type": "text", + "content": ", which is divided into four sub-rectangles by gray lines, each with an area " + }, + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "type": "text", + "content": ". By leveraging this strategy, we can generate supervision for points with continuous positions, which enables us to fully exploit the advantages of the implicit model and achieve high-fidelity visual results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 536, + 273, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 273, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 273, + 550 + ], + "type": "text", + "content": "4.3. Geometry-Aware Mutual Explicit Mapping" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 554, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 686 + ], + "type": "text", + "content": "Once we have obtained the speech-sensitive lip sequences in the canonical view, we can construct the eventually synthesized image " + }, + { + "bbox": [ + 46, + 554, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\hat{I}_o" + }, + { + "bbox": [ + 46, + 554, + 287, + 686 + ], + "type": "text", + "content": " by assigning the synthesized pixels to a location in the observed space. However, since the assigning process is speech-insensitive and only relies on the geometry information, we introduce the GAMEM module, illustrated in Figure 3, to explicitly model it. The GAMEM module comprises Canonical-to-Observed (C2O), Observed-to-Canonical (O2C) Mapping, and a learnable full-head depth map, which supports various applications, including pose-controllable image synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "C2O/O2C Mapping These mappings represent how the images transform between two spaces (i.e., canonical space" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 358, + 339, + 496, + 408 + ], + "blocks": [ + { + "bbox": [ + 358, + 339, + 496, + 408 + ], + "lines": [ + { + "bbox": [ + 358, + 339, + 496, + 408 + ], + "spans": [ + { + "bbox": [ + 358, + 339, + 496, + 408 + ], + "type": "image", + "image_path": "b29d85034d561383250a604e9080fffbb758e88a3fee026ae3de0a125f4169c6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 415, + 545, + 448 + ], + "lines": [ + { + "bbox": [ + 305, + 415, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 545, + 448 + ], + "type": "text", + "content": "Figure 4. Continuous Sampling strategy. This strategy can generate supervision signals for sampled positions at different resolutions in the training time." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": "and observed space). As depicted in Figure 3, O2C Mapping aims at creating supervisions " + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "inline_equation", + "content": "I_{o \\to c}^{m}" + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": " for training the lip generator model " + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": " and C2O Mapping is utilized to warp the generated lip image into the observed view to produce " + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\hat{I}_{c \\to o}^{m}" + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": " for further composition. As they follow similar principles, we take the O2C Mapping as an example to illustrate the process. The input of O2C Mapping contains " + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": ", and the output are pixel correspondences that map pixel coordinates in observed space to the canonical space. To achieve it, we use the 3D Morphable Model (3DMM) [2] to calculate the overall camera intrinsic matrix " + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": " and rough face geometry " + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": ". The head poses " + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "inline_equation", + "content": "T_{o} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "inline_equation", + "content": "T_{c} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": ", which include a rotation matrix and a translation vector, are also estimated. And the relative head pose " + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "inline_equation", + "content": "T_{c \\to o} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 304, + 453, + 545, + 644 + ], + "type": "text", + "content": " between two spaces can be computed as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 384, + 645, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 645, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 384, + 645, + 545, + 658 + ], + "type": "interline_equation", + "content": "T _ {c \\rightarrow o} = T _ {o} \\times T _ {c} ^ {- 1}. \\tag {3}", + "image_path": "423e9dcd1e98228ead963700a59ac1aaaedd1c828d67b74b940103db9b476844.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "Furthermore, face depth maps " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " obtained from 3DMM can be further interpolated based on " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": ", and the corresponding head poses. Noticing that 3DMM can only be aware of the face area instead of the head area," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22171" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "but the information about the rest part of the head (e.g., forehead, ears, eyes, mouth) is lacking. Hence we optimize the " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " along with our model to complete the missing depth region. If " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "p_{o}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "p_{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " are the corresponding 2D homogeneous pixel grid coordinates on " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": ", the geometric relationships can be explicitly formulated as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 153, + 287, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 153, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 110, + 153, + 287, + 168 + ], + "type": "interline_equation", + "content": "D _ {o} p _ {o} = K T _ {c \\rightarrow o} D _ {c} K ^ {- 1} p _ {c}, \\qquad (4)", + "image_path": "9b1190c679e669ad171270bf9ee4a35bf2d17d8c8a69d3c9a10fbb723cc53f3f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 176, + 287, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 176, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 287, + 201 + ], + "type": "text", + "content": "so the position mapping from canonical space to observed space can be easily obtained as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 210, + 287, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 210, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 110, + 210, + 287, + 224 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {\\boldsymbol {c} \\rightarrow \\boldsymbol {o}} \\left(T _ {c \\rightarrow o}, D _ {c}, p _ {c}\\right) = p _ {o}. \\tag {5}", + "image_path": "6e974a059dfae188cfc40e84b2140fcde30a9e95d163dbccbb5012aa30b5b4dc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "text", + "content": "Then, we can warp the observed ground-truth lip images " + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "inline_equation", + "content": "I_{o}^{m}" + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "text", + "content": " into the canonical space by backward warping, denoted by " + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "inline_equation", + "content": "I_{o\\rightarrow c}^{m}" + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "text", + "content": ", to guide the learning of the lip generation model. The C2O Mapping can be similarly defined by " + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "inline_equation", + "content": "T_{o\\rightarrow c}" + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "text", + "content": ". With it, we can warp the generated canonical lip image " + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\hat{I}_c^m" + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "text", + "content": " into the observed space by backward warping, denoted by " + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\hat{I}_{c\\to o}^{m}" + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "text", + "content": ", to compose with the given observed image " + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 47, + 233, + 288, + 330 + ], + "type": "text", + "content": " for synthesizing a talking face." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "type": "text", + "content": "Pose Controllable Synthesis Thanks to the simultaneously learned complete depth map " + }, + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "type": "text", + "content": ", pose-controllable full-head image synthesis according to users' requirements is also supported in our model. Specifically, we employ " + }, + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "type": "text", + "content": " to determine location correspondences based on Eq. 3 and Eq. 5 by altering " + }, + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "type": "inline_equation", + "content": "T_{o}" + }, + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "type": "text", + "content": ". There is a slight difference in this situation, which is to use the forward warping strategy as the ground-truth depths at new given poses are missing, which will produce black holes. To mitigate this issue, we propose a data augmentation method in training by randomly adding black holes with a ratio of " + }, + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 332, + 289, + 477 + ], + "type": "text", + "content": " to the image to let the Blend-Net learn to fill these areas." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 484, + 162, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 162, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 162, + 496 + ], + "type": "text", + "content": "4.4. Overall Refinement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 506, + 287, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 650 + ], + "type": "text", + "content": "Image Blending With " + }, + { + "bbox": [ + 46, + 506, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\hat{I}_{c\\rightarrow o}^{m}" + }, + { + "bbox": [ + 46, + 506, + 287, + 650 + ], + "type": "text", + "content": ", we can integrate it by directly pasting it to the original observed frame based on 2D lip landmarks [3]. However, the paste operation often results in mismatching artifacts in the boundary area, and the image artifacts after data augmentation should also be modified. Therefore, we propose a Blend-Net for blending. The Blend-Net takes the pasted image after the paste operation as input to synthesize harmonized final output " + }, + { + "bbox": [ + 46, + 506, + 287, + 650 + ], + "type": "inline_equation", + "content": "\\hat{I}_o" + }, + { + "bbox": [ + 46, + 506, + 287, + 650 + ], + "type": "text", + "content": ". Since the target is fusion and amending instead of generation, we predict the pixel residual, which is further added back to the input image to composite the final output image. Detailed network structure is illustrated in the supplementary file." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 713 + ], + "type": "text", + "content": "Synchronization Enhancement. To further improve the synchronization performance, we introduce a pre-trained sync expert network to boost the model's performance in synchronization inspiring by [26]. The sync expert network consists of two pre-trained encoders, which extract features" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "of image and speech audio within a sliding window, denoted by " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " respectively. Different from [26], we only have a short video for training. Therefore, the unsynced speech-image pairs are also exploited to construct the contrastive loss which helps avoid falling into a trivial solution. Its effectiveness is also verified in our experiments. The distance of syncsed speech-image pairs is desired to be closer while that of the unsynced speech-image pairs should be farther. Thus, we define a contrastive sync loss " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " based on the metric which is widely used in contrastive learning:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 200, + 545, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 200, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 310, + 200, + 545, + 214 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = y \\cdot (1 - \\cos (\\boldsymbol {i}, \\boldsymbol {a})) + (1 - y) \\cdot \\max (0, \\cos (\\boldsymbol {i}, \\boldsymbol {a})), \\tag {6}", + "image_path": "f395d497380ffe9b746a51cca059c72141c3cffb3acb14f88b62730e56d519b2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 221, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 221, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\cos" + }, + { + "bbox": [ + 304, + 221, + 545, + 281 + ], + "type": "text", + "content": " is the cosine similarity metric. " + }, + { + "bbox": [ + 304, + 221, + 545, + 281 + ], + "type": "inline_equation", + "content": "y = 1" + }, + { + "bbox": [ + 304, + 221, + 545, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 221, + 545, + 281 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 304, + 221, + 545, + 281 + ], + "type": "text", + "content": " represent positive and negative speech-image pairs, respectively. For the input speech, the positive images are the matched images at the same timestamp while the negative images are randomly chosen at some timestamps else." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 288, + 422, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 422, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 422, + 301 + ], + "type": "text", + "content": "4.5. Training Objectives" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "text", + "content": "Our whole pipeline is trained in a self-supervised manner using the observed frames to provide the supervisory signals. The overall loss includes a canonical mouth image reconstruction loss " + }, + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "text", + "content": ", a depth loss " + }, + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "text", + "content": ", a sync loss " + }, + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "text", + "content": ", and a whole observed image reconstruction loss " + }, + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_w" + }, + { + "bbox": [ + 304, + 307, + 545, + 366 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 367, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 367, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 305, + 367, + 545, + 392 + ], + "type": "text", + "content": " measures the reconstruction error between the predicted lip image " + }, + { + "bbox": [ + 305, + 367, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\hat{I}_c^m" + }, + { + "bbox": [ + 305, + 367, + 545, + 392 + ], + "type": "text", + "content": " and ground-truth lip image " + }, + { + "bbox": [ + 305, + 367, + 545, + 392 + ], + "type": "inline_equation", + "content": "I_{o\\rightarrow c}^{m}" + }, + { + "bbox": [ + 305, + 367, + 545, + 392 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 343, + 397, + 545, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 397, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 343, + 397, + 545, + 413 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = \\mathcal {L} _ {p} \\left(\\hat {I} _ {c} ^ {m}, I _ {o \\rightarrow c} ^ {m}\\right) + \\| \\hat {I} _ {c} ^ {m} - I _ {o \\rightarrow c} ^ {m} \\| _ {2}, \\tag {7}", + "image_path": "fcb470ae01d96fc6123a809f6af9d3da542cff8051835b80d4028fb897eb53f9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 420, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 420, + 545, + 468 + ], + "type": "inline_equation", + "content": "I_{o\\rightarrow c}^{m}" + }, + { + "bbox": [ + 304, + 420, + 545, + 468 + ], + "type": "text", + "content": " is the warped ground truth lip image using O2C space mapping (see the supplementary file for more details) and the widely-used perceptual loss " + }, + { + "bbox": [ + 304, + 420, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 304, + 420, + 545, + 468 + ], + "type": "text", + "content": " is defined in [40]. Similarly, overall reconstruction loss " + }, + { + "bbox": [ + 304, + 420, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_w" + }, + { + "bbox": [ + 304, + 420, + 545, + 468 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 360, + 475, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 475, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 360, + 475, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {w} = \\mathcal {L} _ {p} \\left(\\hat {I} _ {o}, I _ {o}\\right) + \\| \\hat {I} _ {o} - I _ {o} \\| _ {2}. \\tag {8}", + "image_path": "18ef43f3a062c8400aa99dcd0cb10250c64c6ffefb04133ff5c0c434f417b5cf.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 534 + ], + "type": "text", + "content": "The canonical head depth map is initialized by the incomplete depth map computed by 3DMM and is trained with the help of the photometric loss as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 381, + 541, + 545, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 541, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 381, + 541, + 545, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\left\\| \\hat {I} _ {o \\rightarrow c} - I _ {c} \\right\\| _ {2}, \\tag {9}", + "image_path": "eed8ea39d6da92d46bcec7b0012f94b7ca34212ae5778072df2c9ff0c12406e0.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\hat{I}_{o\\rightarrow c}" + }, + { + "bbox": [ + 304, + 563, + 545, + 588 + ], + "type": "text", + "content": " is the warped predicted image. With the contrastive sync loss (Sec. 4.4), the overall loss function is" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 322, + 596, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 596, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 322, + 596, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\omega_ {m} \\cdot \\mathcal {L} _ {m} + \\omega_ {w} \\cdot \\mathcal {L} _ {w} + \\omega_ {d} \\cdot \\mathcal {L} _ {d} + \\omega_ {s} \\cdot \\mathcal {L} _ {s}. \\tag {10}", + "image_path": "51468afffc44487a578fffa4a81ccefdc591d425720f8888fac92fc79affb13d.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 617, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 617, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 545, + 629 + ], + "type": "text", + "content": "Implementation details are shown in the supplementary file." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 659, + 426, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 426, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 426, + 671 + ], + "type": "text", + "content": "5.1. Datasets and metrics" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate our algorithm and other methods on three datasets belonging to three specific speakers (Testset I, Testset II and Testset III) as recent speaker-specific" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22172" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 539, + 202 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 539, + 202 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 539, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 539, + 202 + ], + "type": "table", + "html": "
MethodTrained with large extra dataTestset ITestset IITestset III Sync↑
PSNR↑SSIM↑CPBD↑LMD↓Sync↑PSNR↑SSIM↑CPBD↑LMD↓Sync↑
Ground TruthN/AN/A1.0000.1860.0009.102N/A1.0000.1210.0008.6884.877
ATVG [7]Yes28.4520.8180.0195.4235.81328.0510.6680.0034.2036.2243.571
MakeitTalk [44]Yes29.6920.9060.0504.2154.11528.9960.8130.0614.4635.5592.320
Wav2Lip [26]Yes31.5570.9800.1153.05310.03131.7930.9560.0653.4159.9365.809
PC-AVS [43]Yes29.0720.8800.0404.5959.25828.3590.7340.0464.3058.5865.206
LSP [19]No29.5150.9000.0983.1745.37728.8950.7760.1174.9726.8113.046
AD-NeRF [16]No32.2230.9540.0512.9896.04230.8850.9090.0553.2105.9103.285
DFRF [28]No33.2920.9740.0943.0795.25231.4190.9440.1243.1395.5522.879
Speech2LipNo34.8150.9870.2242.9767.77133.1970.9620.1253.0827.3704.379
", + "image_path": "f3792334849146e9664a2484666ae7010ff0ea591429c1a2e0bd226d3760f65b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 202, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 547, + 246 + ], + "type": "text", + "content": "Table 1. Quantitative results compared with the SOTA methods. Image quality assessment metrics (i.e., PSNR, SSIM, and CPBD) are computed within mouth region. Algorithms are categorized into two classes based on the training datasets for a fair comparison. The first ones are trained in large public datasets while the other ones are trained using only 3-5min videos. The best results of each class are in bold and overall best results are with underlines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 265, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 289, + 373 + ], + "type": "text", + "content": "methods [16, 19] usually do. Training videos are collected from [19] as they are publicly available. All video sequences are resampled to 25 FPS and split into training part and test part with a ratio of " + }, + { + "bbox": [ + 46, + 265, + 289, + 373 + ], + "type": "inline_equation", + "content": "90\\% / 10\\%" + }, + { + "bbox": [ + 46, + 265, + 289, + 373 + ], + "type": "text", + "content": ". The training splits of Testset I " + }, + { + "bbox": [ + 46, + 265, + 289, + 373 + ], + "type": "inline_equation", + "content": "(500 \\times 500)" + }, + { + "bbox": [ + 46, + 265, + 289, + 373 + ], + "type": "text", + "content": " and Testset II " + }, + { + "bbox": [ + 46, + 265, + 289, + 373 + ], + "type": "inline_equation", + "content": "(624 \\times 624)" + }, + { + "bbox": [ + 46, + 265, + 289, + 373 + ], + "type": "text", + "content": " are employed to train the models and then same-identity evaluations are conducted on the test splits. Testset III is utilized for cross-identity, -gender, and -language tests. We present more evaluation results in the supplementary file." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 378, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 289, + 559 + ], + "type": "text", + "content": "Evaluation Metrics. Since both face generation and lip generation models are contained in our quantitative evaluations, we conduct the image quality assessment just around the mouth region for a fair comparison. To evaluate the appearance quality, we first utilize image quality metrics Peak Signal-to-Noise Ratio (PSNR) and SSIM [37]. A no-reference objective image sharpness metric based on Cumulative Probability of Blur Detection (CPBD) [23, 21, 22] is further utilized to measure the overall perceptual image quality. For synchronization evaluation, Landmarks Distance (LMD) around the mouth region is exploited to compute the accuracy of lip shape following [6]. The confidence score of lip synchronization computed by a pretrained SyncNet [12] is also applied to measure the synchronization performance, labeled by Sync." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 570, + 235, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 235, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 235, + 583 + ], + "type": "text", + "content": "5.2. Comparisons with State of the Arts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Same-identity Evaluation. For speaker-independent models, we conduct experiments using their officially-released pre-trained models. Since AD-NeRF [16] and DFRF [28] are speaker-specific models, we retrain them using our training data for a fair comparison. DFRF aims at few-shot learning but its lip synchronization performance witnesses a significant drop when the number of training images decreases from 5000 to 15. Hence, we still train DFRF on thousands of images. Also, we directly use LSP's [19] pretrained models on Testset I and Testset II to conduct experi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 265, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 547, + 338 + ], + "type": "text", + "content": "ments (subject May for Testset I and Testset III, and subject Obama2 for Testset II). We cannot provide a quantitative comparison with SSP-NeRF [18] here, as neither codes nor pre-trained weights are available. Thus, for comparison, we extract the speech from their released demo and then present qualitative comparisons in the supplementary file." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 338, + 547, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 698 + ], + "type": "text", + "content": "Quantitative results are shown in Table 1. We divide all the models into two types based on the scale of the training dataset. The upper four algorithms (speaker-independent models) are trained using public video datasets with large amounts of identities and speech-image pairs (e.g., LRS2 [1, 11, 13] and VoxCeleb2 [10]) while the latter four methods (speaker-specific models) only use 5-minute videos of a specific identity for training. Speaker-independent models tend to perform well in synchronization because they can learn lip motions well from various speech-visual pairs in training time. However, the cost is sacrificing the visual quality. The metrics of image quality (e.g., PSNR, SSIM, and CPBD) are much lower than that of specific models. Wav2lip [26] is an exception because it only generates lip images, but the details still can not be produced well (Figure 5). In contrast, speaker-specific models usually have much better image qualities, which is favorable in practice. Our model outperforms all the SOTA algorithms in image quality and achieves the best synchronization performance among all the speaker-specific models, and the Sync score is also competitive when compared to those speaker-independent models. It is also worth mentioning that the Sync scores of Wav2lip [26] and PC-AVS [43] are higher than that of the ground truth. The reason might be that the Sync score is sensitive to synchronization quality only when synchronization quality is within a range and a higher score above a threshold may not necessarily imply a higher synchronization quality. We design a toy experiment to verify it in the supplementary file and our user study can also demonstrate it." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Cross-identity Evaluation. Our model also has the abil" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22173" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 102, + 258 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 102, + 258 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 102, + 258 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 102, + 258 + ], + "type": "image", + "image_path": "54512ee06f077ac6673d5030f7d533444c5e53634b0733abd8ebb7ddfaa7bfba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 261, + 546, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 546, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 546, + 283 + ], + "type": "text", + "content": "Figure 5. Qualitative results compared with SOTA methods. The Lip area is cropped based on the detected 2D landmarks for a clear comparison. Speaker-independent and speaker-specific models are remarked by different colors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 72, + 323, + 258 + ], + "blocks": [ + { + "bbox": [ + 108, + 72, + 323, + 258 + ], + "lines": [ + { + "bbox": [ + 108, + 72, + 323, + 258 + ], + "spans": [ + { + "bbox": [ + 108, + 72, + 323, + 258 + ], + "type": "image", + "image_path": "85cffeeca06aad9ce7bb1d34f3ef9f6984bfc5548b2ef9e4c70d9d69a3bbec2e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 544, + 258 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 544, + 258 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 544, + 258 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 544, + 258 + ], + "type": "image", + "image_path": "4885db8e54c2ee536f28575f4512b8534e61893a95201e143fb9b23c9bf49076.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 297, + 289, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 289, + 439 + ], + "type": "text", + "content": "ity to generate cross-identity, -gender, and -language results. Since there is no ground truth image, we provide the sync performance comparison on generating a video for the model trained on Testset I (female, British) using another speaker's speech (male, French). This setting is defined as Testset III, and results are shown in the last column of Table 1. Our model still has significant superiority and outperforms other speaker-specific models by a large margin. We also test the effect of unsynced speech-image pairs used in Eq. 6 on Testset III. Without negative pairs, the Sync score decreases from 4.379 to 3.830, showing the effectiveness of our contrastive sync loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 648 + ], + "type": "text", + "content": "Qualitative Results. Qualitative comparisons of different algorithms are presented in Figure 5. We can see that speaker-independent methods all suffer from low image quality. Besides, ATVG and MakeiTalk have unnatural speech-insensitive movements (e.g., head motion and background motion) and PC-AVS has an extra issue of struggling with identity preservation. Speaker-specific approaches have much better visual quality but most of them can hardly estimate precise lip details when there exist considerable significant lip motions. And fine details like teeth and tongue are not well modeled either. Besides, AD-NeRF sometimes causes head-torso separation due to the adoption of two separate NeRFs. In contrast, our method Speech2Lip performs well in all these situations, as the generated shapes are more accurate and the synthetic images are more clear and realistic, especially the lip area. More results can be found in the supplementary videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Complexity Comparisons. To further demonstrate the superiority of our model, complexity and computational cost comparisons with speaker-specific models are conducted on Testset I. In Table 2, our model gains the best results with much less complexity and lower computational cost." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 294, + 526, + 369 + ], + "blocks": [ + { + "bbox": [ + 326, + 294, + 526, + 369 + ], + "lines": [ + { + "bbox": [ + 326, + 294, + 526, + 369 + ], + "spans": [ + { + "bbox": [ + 326, + 294, + 526, + 369 + ], + "type": "table", + "html": "
MethodsLSPAD-NeRFDFRFOurs
Model size (MB) ↓500302030
Train time (hour) ↓38.5806030
Test speed (FPS) ↑350.060.0418
PSNR (dB) ↑29.5232.2233.2934.82
Sync ↑5.386.045.257.77
", + "image_path": "18299d8cc8024099a3c46fa770d4c38cf0726b6c04435afcdc12d4a231e8babf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 369, + 545, + 390 + ], + "lines": [ + { + "bbox": [ + 306, + 369, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 545, + 390 + ], + "type": "text", + "content": "Table 2. Speaker-specific model comparisons. The best results are in bold and the second best results are with underlines." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 411, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 590 + ], + "type": "text", + "content": "User Study. To verify the perceptual quality, user studies are conducted on 16 generated video clips covering both Testset I and Testset II. Videos are generated by our proposed Speech2Lip and other SOTA algorithms, with algorithm names hidden and video order randomized. 17 participants independently evaluated all the videos. For each video, each participant gave a Mean Opinion Score (MOS) from 1-5 for each of three aspects: qualities of speech-visual synchronization, image fidelity, and image realness. Higher scores represent better quality. The overall results are shown in Table 3. Speech2Lip not only achieves the highest MOS in image quality measurement (e.g., 4.618 and 4.382) but also outperforms all the other algorithms in lip synchronization (e.g., 4.265), exhibiting Speech2Lip has the most satisfactory overall video quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 599, + 400, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 400, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 400, + 612 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Contributions of loss functions. We first conduct an ablation study on loss functions. The evaluation metrics contain the PSNR around the lip area and the sync score. The PSNR of the whole image is also retained since practical applications attach more attention to the overall quality of the talking portrait scene. Results are shown in Table 4, demonstrating the loss functions all contribute to the increased performance on both visual quality and synchronization." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22174" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 71, + 497, + 123 + ], + "blocks": [ + { + "bbox": [ + 97, + 71, + 497, + 123 + ], + "lines": [ + { + "bbox": [ + 97, + 71, + 497, + 123 + ], + "spans": [ + { + "bbox": [ + 97, + 71, + 497, + 123 + ], + "type": "table", + "html": "
MethodsATVGMakeitTalkWav2LipPC-AVSLSPAD-NeRFDFRFSpeech2Lip
Lip Synchronization1.5001.8534.1913.0593.1183.0882.8534.265
Image Fidelity1.1472.7943.2652.5293.8823.2353.2654.618
Image Realness1.1182.0883.7652.2653.0002.3242.9714.382
", + "image_path": "ea9073f00849656e86fa7d14b2e794a154435e9f3a75f642b1aaa20d5a46d840.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 152, + 282, + 215 + ], + "blocks": [ + { + "bbox": [ + 123, + 123, + 470, + 133 + ], + "lines": [ + { + "bbox": [ + 123, + 123, + 470, + 133 + ], + "spans": [ + { + "bbox": [ + 123, + 123, + 470, + 133 + ], + "type": "text", + "content": "Table 3. Detailed user study results compared with SOTA methods. The best results are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 152, + 282, + 215 + ], + "lines": [ + { + "bbox": [ + 52, + 152, + 282, + 215 + ], + "spans": [ + { + "bbox": [ + 52, + 152, + 282, + 215 + ], + "type": "table", + "html": "
MethodsPSNRlip↑PSNRimg↑Sync↑
Lw34.10436.9476.313
Lw + Ld34.42936.9946.359
Lw + Ld + Lm34.69437.1937.194
Lw + Ld + Lm + Ls34.81537.2457.771
", + "image_path": "1f21ac99ed47689a7c34ef27d8f042b0f4b6b487ee6783ea33d8f266f6c07ac1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 236, + 282, + 310 + ], + "blocks": [ + { + "bbox": [ + 48, + 215, + 277, + 225 + ], + "lines": [ + { + "bbox": [ + 48, + 215, + 277, + 225 + ], + "spans": [ + { + "bbox": [ + 48, + 215, + 277, + 225 + ], + "type": "text", + "content": "Table 4. Ablation study results about loss function on Testset I." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 236, + 282, + 310 + ], + "lines": [ + { + "bbox": [ + 52, + 236, + 282, + 310 + ], + "spans": [ + { + "bbox": [ + 52, + 236, + 282, + 310 + ], + "type": "table", + "html": "
MethodsPSNRlip↑PSNRimg↑Sync↑
w/o continuous sampling34.70237.2147.553
w/o implicit model34.25037.0586.790
w/o Blend-Net33.73636.8816.065
w/o time34.51237.1217.146
Speech2Lip34.81537.2457.771
", + "image_path": "e6ee61cdeb8040da8e3e73ac9ffe08cab13133aac007d4d2a10fe68ac0f3a063.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 311, + 285, + 321 + ], + "lines": [ + { + "bbox": [ + 47, + 311, + 285, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 285, + 321 + ], + "type": "text", + "content": "Table 5. Ablation study results about network design on Testset I." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 340, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 449 + ], + "type": "text", + "content": "Contributions of individual components. We also explore the benefits of our key components by removing each component to see how the performance changes in Table 5. \"w/o implicit model\" represents we employ an explicit lip generation model instead, and \"w/o time\" indicates times-tamp is deleted from inputs. From the results, it can be concluded that each design plays an important role. Among them, the Blend-Net significantly improves video quality, and the design of implicit modeling schema is also critical." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 456, + 212, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 212, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 212, + 468 + ], + "type": "text", + "content": "5.4. Controllable Synthesis Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "text", + "content": "Thanks to the joint optimization and completion of a full-head depth map " + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "text", + "content": " in GAMEM, our model supports synthesizing full-head images driven by novel head poses. Specifically, " + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "text", + "content": " can be projected into any observed view to obtain " + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "inline_equation", + "content": "T_{o}" + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "text", + "content": " in Eq. (3) are further replaced by " + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 46, + 474, + 287, + 652 + ], + "type": "text", + "content": " and the new target head pose, respectively. Then, we adopt Eq. (5) to calculate the correspondence. The correspondence is further exploited to warp the synthesized image in the observed space into a new target space. It is noted that depth maps are used for coarse mapping between spaces while facial expression is more taken charge of by fine-grained contents from observed views. The pose-controllable novel view synthesis results are shown in Figure 7. When testing with moderate pose changes, our model achieves comparable results as AD-NeRF." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "We note that existing methods with 3D modeling such as AD-NeRF also cannot generate good synthesis results if the evaluated poses deviate too much from the training data. The reason is that the training data widely used in speaker-specific models [19, 16, 28, 18] only show the front view" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 340, + 144, + 418, + 233 + ], + "blocks": [ + { + "bbox": [ + 340, + 144, + 418, + 233 + ], + "lines": [ + { + "bbox": [ + 340, + 144, + 418, + 233 + ], + "spans": [ + { + "bbox": [ + 340, + 144, + 418, + 233 + ], + "type": "image", + "image_path": "fe3e1b2368d9dd5d8ac162519478c0aa7315a28e3ff6b00e8346834a3fda973c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 234, + 509, + 245 + ], + "lines": [ + { + "bbox": [ + 343, + 234, + 509, + 245 + ], + "spans": [ + { + "bbox": [ + 343, + 234, + 509, + 245 + ], + "type": "text", + "content": "Figure 6. 3D Mesh of AD-NeRF on Testset I." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 435, + 144, + 511, + 233 + ], + "blocks": [ + { + "bbox": [ + 435, + 144, + 511, + 233 + ], + "lines": [ + { + "bbox": [ + 435, + 144, + 511, + 233 + ], + "spans": [ + { + "bbox": [ + 435, + 144, + 511, + 233 + ], + "type": "image", + "image_path": "b65d7199e3c382b4c1ed709256cb5fdf92e81f522b603c15fe6c98f4cc085c28.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 329, + 251, + 520, + 300 + ], + "blocks": [ + { + "bbox": [ + 329, + 251, + 520, + 300 + ], + "lines": [ + { + "bbox": [ + 329, + 251, + 520, + 300 + ], + "spans": [ + { + "bbox": [ + 329, + 251, + 520, + 300 + ], + "type": "image", + "image_path": "9adf590cf715a8fd37d35fa7d3fa95b7f0d64a6670283ff5e075ba4d322f5e5c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 301, + 439, + 308 + ], + "lines": [ + { + "bbox": [ + 406, + 301, + 439, + 308 + ], + "spans": [ + { + "bbox": [ + 406, + 301, + 439, + 308 + ], + "type": "text", + "content": "(a) AD-NeRF" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 329, + 310, + 520, + 360 + ], + "blocks": [ + { + "bbox": [ + 329, + 310, + 520, + 360 + ], + "lines": [ + { + "bbox": [ + 329, + 310, + 520, + 360 + ], + "spans": [ + { + "bbox": [ + 329, + 310, + 520, + 360 + ], + "type": "image", + "image_path": "fcdee3442b9ba562bb89e6e2979c44d8d7aefbea96911ed622dd9973c9654cc7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 360, + 443, + 367 + ], + "lines": [ + { + "bbox": [ + 406, + 360, + 443, + 367 + ], + "spans": [ + { + "bbox": [ + 406, + 360, + 443, + 367 + ], + "type": "text", + "content": "(b) Speech2Lip" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 365, + 545, + 387 + ], + "lines": [ + { + "bbox": [ + 306, + 365, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 545, + 387 + ], + "type": "text", + "content": "Figure 7. Novel view synthesis capability of our model. We rotate the head region with a random angle." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 399, + 547, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 547, + 447 + ], + "type": "text", + "content": "with limited pose variations, being naturally hard for 3D modeling. The 3D mesh is incomplete, and the surface is rough (Figure 6). In addition, the computational costs of 3D methods are extremely high (Table 2)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 463, + 378, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 463, + 378, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 463, + 378, + 475 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 484, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 688 + ], + "type": "text", + "content": "In this paper, we propose a novel decomposition-synthesis-composition framework called Speech2Lip for high-fidelity talking head video synthesis, which disentangles speech-sensitive and speech-insensitive motions/appearances. By presenting a syncsed speech-driven implicit model, a GAMEM module, a Blend-Net, and a contrastive sync loss, we can achieve satisfactory results with only a few minutes of training video. Our framework also supports pose-controllable synthesis. In the future, we plan to study generating realistic expressions driven by speech and explore combining our insights with advanced general image generation methods such as diffusion-based models for better generalizability. We hope that our work inspires more future research in this field and encourages the development of positive applications. However, we also urge caution to prevent any potential abuses. More discussions about limitations are described in the supplementary file." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement This work has been supported by the Research Fund from Tencent ARC lab." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22175" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] T. Afouras, J. S. Chung, A. Senior, O. Vinyals, and A. Zisserman. Deep audio-visual speech recognition. In arXiv:1809.02108, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "text", + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "text", + "content": "[3] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 257 + ], + "type": "text", + "content": "[4] Lele Chen, Guofeng Cui, Ziyi Kou, Haitian Zheng, and Chenliang Xu. What comprises a good talking-head video generation? In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 313 + ], + "type": "text", + "content": "[5] Lele Chen, Guofeng Cui, Celong Liu, Zhong Li, Ziyi Kou, Yi Xu, and Chenliang Xu. Talking-head generation with rhythmic head motion. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX, pages 35-51. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 287, + 357 + ], + "type": "text", + "content": "[6] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 413 + ], + "type": "text", + "content": "[7] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 413, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 467 + ], + "type": "text", + "content": "[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 468, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 491 + ], + "type": "text", + "content": "[9] Joon Son Chung, Amir Jamaludin, and Andrew Zisserman. You said that? arXiv preprint arXiv:1705.02966, 2017. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 523 + ], + "type": "text", + "content": "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 557 + ], + "type": "text", + "content": "[11] J. S. Chung, A. Senior, O. Vinyals, and A. Zisserman. Lip reading sentences in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "text", + "content": "[12] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian conference on computer vision, pages 251-263. Springer, 2016. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 613 + ], + "type": "text", + "content": "[13] J. S. Chung and A. Zisserman. Lip reading in profile. In British Machine Vision Conference, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[14] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision, pages 408-424. Springer, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[16] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. 2, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 128, + 546, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 128, + 546, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 546, + 171 + ], + "type": "text", + "content": "[17] Rithesh Kumar, Jose Sotelo, Kundan Kumar, Alexandre de Brébisson, and Yoshua Bengio. Obamanet: Photo-realistic lip-sync from text. arXiv preprint arXiv:1801.01442, 2017. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 172, + 546, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 546, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 546, + 216 + ], + "type": "text", + "content": "[18] Xian Liu, Yinghao Xu, Qianyi Wu, Hang Zhou, Wayne Wu, and Bolei Zhou. Semantic-aware implicit neural audio-driven video portrait generation. arXiv preprint arXiv:2201.07786, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 216, + 546, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 216, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 216, + 546, + 249 + ], + "type": "text", + "content": "[19] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 250, + 546, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 546, + 305 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 546, + 305 + ], + "type": "text", + "content": "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 305, + 546, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 546, + 348 + ], + "type": "text", + "content": "[21] Niranjan Narvekar and Lina J Karam. An improved no-reference sharpness metric based on the probability of blur detection. In Workshop on Video Processing and Quality Metrics, 2010. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 349, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 546, + 403 + ], + "type": "text", + "content": "[22] Niranjan D Narvekar and Lina J Karam. A no-reference perceptual image sharpness metric based on a cumulative probability of blur detection. In 2009 International Workshop on Quality of Multimedia Experience, pages 87-91. IEEE, 2009. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 403, + 546, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 403, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 403, + 546, + 448 + ], + "type": "text", + "content": "[23] Niranjan D Narvekar and Lina J Karam. A no-reference image blur metric based on the cumulative probability of blur detection (cpbd). IEEE Transactions on Image Processing, 20(9):2678-2683, 2011. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 448, + 546, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 448, + 546, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 448, + 546, + 503 + ], + "type": "text", + "content": "[24] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 503, + 546, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 546, + 558 + ], + "type": "text", + "content": "[25] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 558, + 546, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 558, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 558, + 546, + 613 + ], + "type": "text", + "content": "[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 614, + 546, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 546, + 669 + ], + "type": "text", + "content": "[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "type": "text", + "content": "[28] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European conference on computer vision, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22176" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[30] Linsen Song, Wayne Wu, Chen Qian, Ran He, and Chen Change Loy. Everybody's talkin': Let me talk as you want. IEEE Transactions on Information Forensics and Security, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[31] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[32] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European conference on computer vision, pages 716-731. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 296 + ], + "type": "text", + "content": "[33] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. International Journal of Computer Vision, 128(5):1398-1413, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "text", + "content": "[34] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "type": "text", + "content": "[35] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "text", + "content": "[36] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2531–2539, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "type": "text", + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "text", + "content": "[38] Olivia Wiles, A Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Proceedings of the European conference on computer vision (ECCV), pages 670-686, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[39] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 601 + ], + "type": "text", + "content": "[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "text", + "content": "[41] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[42] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9299–9306, 2019. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 183 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 138 + ], + "type": "text", + "content": "[43] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "type": "text", + "content": "[44] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 1, 2, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22177" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/236896c7-9d5b-41e0-8f54-b3702ca337d4_content_list.json b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/236896c7-9d5b-41e0-8f54-b3702ca337d4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f3262363bf5a04c1f7ff14a462c30d152d8792de --- /dev/null +++ b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/236896c7-9d5b-41e0-8f54-b3702ca337d4_content_list.json @@ -0,0 +1,1744 @@ +[ + { + "type": "text", + "text": "Speech4Mesh: Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation", + "text_level": 1, + "bbox": [ + 120, + 128, + 849, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shan He $^{1,2}$ , Haonan He $^{2}$ , Shuo Yang $^{2}$ , Xiaoyan Wu $^{2}$ , Pengcheng Xia $^{2}$ , Bing Yin $^{2}$ , Cong Liu $^{2}$ , Lirong Dai $^{1}$ , Chang Xu $^{3}$", + "bbox": [ + 205, + 202, + 763, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China, 2iFLYTEK Research, 3University of Sydney {shanhe2, hnhe, shuoyang7, xywu10, pcxia, bingyin, congliu2}@iflytek.com lrdai@ustc.edu.cn, c.xu@sydney.edu.au", + "bbox": [ + 114, + 238, + 854, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent audio2mesh-based methods have shown promising prospects for speech-driven 3D facial animation tasks. However, some intractable challenges are urgent to be settled. For example, the data-scarcity problem is intrinsically inevitable due to the difficulty of 4D data collection. Besides, current methods generally lack controllability on the animated face. To this end, we propose a novel framework named Speech4Mesh to consecutively generate 4D talking head data and train the audio2mesh network with the reconstructed meshes. In our framework, we first reconstruct the 4D talking head sequence based on the monocular videos. For precise capture of the talking-related variation on the face, we exploit the audio-visual alignment information from the video by employing a contrastive learning scheme. We next can train the audio2mesh network (e.g., FaceFormer) based on the generated 4D data. To get control of the animated talking face, we encode the speaking-unrelated factors (e.g., emotion, etc.) into an emotion embedding for manipulation. Finally, a differentiable renderer guarantees more accurate photometric details of the reconstruction and animation results. Empirical experiments demonstrate that the Speech4Mesh framework can not only outperform state-of-the-art reconstruction methods, especially on the lower-face part but also achieve better animation performance both perceptually and objectively after pre-trained on the synthesized data. Besides, we also verify that the proposed framework is able to explicitly control the emotion of the animated talking face.", + "bbox": [ + 76, + 359, + 473, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 813, + 209, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Speech-driven facial animation is aimed at guiding the face model (either 2D image or 3D mesh) to have perceptually rational motion (especially for lip sync) only according to the input audio signal. It is drawing increasingly more", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db64bfb2d479e1ede49ea6bdb631142f53d4050c80a55e001b3bd360a71aa2b6.jpg", + "image_caption": [ + "Figure 1: Overview of the proposed Speech4Mesh framework. In this framework, the speech signal is leveraged for both the reconstruction module and the audio2mesh module. An emotion code is embedded into the audio2mesh module for emotional manipulation. Differentiable rendering is employed for recovering the details of the face." + ], + "image_footnote": [], + "bbox": [ + 511, + 325, + 880, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "attention with the development of virtual reality, game and film production, etc. There are currently two mainstreams to deal with this task, either from 2D or 3D perspectives. By learning from video data [9, 41, 55, 71], 2D solutions try to generate the talking head images frame by frame. Although sufficient data is available to be used, the synthesized images cannot directly be applied to 3D scenes. Thus, in this paper, we mainly focus on 3D facial animation driven by speech. In recent years, a handful of 3D speech-driven facial animation methods [20, 47, 39, 13, 7, 34] have shown promising performance by directly learning a mapping from audio to 3D avatar mesh deformation with deep neural networks (hereinafter called audio2mesh methods). In this way, some subtle variations on the face mesh can be captured. Generally speaking, for the audio2mesh methods, data quality is one of the most important factors to be considered [20] since they generally demand high-precision", + "bbox": [ + 496, + 636, + 893, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14192", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4D talking head data to learn the correspondence from audio embedding to mesh movement in a supervised manner. However, in practice, real 4D data is extremely hard to collect. It is generally reconstructed and registered from synchronized multi-view images or 3D scan frame by frame, which is extraordinarily costly [13, 21]. Therefore, how to deal with the data scarcity problem becomes an urgent agenda to be discussed.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, the accessibility of monocular videos is much higher. The scarcity of 4D data may be mitigated through the use of 3D reconstruction techniques applied to 2D videos. The majority of current monocular 3D face reconstruction methods (e.g., 3DDFA-V2 [27], DECA [22]) are based on the self-supervised training scheme, whose essential guidances are landmark loss and photometric loss. Nonetheless, these losses can only provide 2D information from the image plane, which is relatively weak to capture the complex nonrigid deformation of the lower face region [51, 62]. Therefore, the reconstructed results always demonstrate obvious artifacts and ambiguities between utterance and visual perception, such as mouth funnel, pucker, and rounded vowels, which may considerably limit the precision of the audio2mesh methods [20].", + "bbox": [ + 75, + 215, + 472, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the problems mentioned above, we propose a novel framework named Speech4Mesh to consecutively synthesize 4D talking head data and train the audio2mesh network with the reconstructed mesh sequences. An overview of the Speech4Mesh framework is illustrated in Fig. 1. Firstly, to obtain better synthesis results, we propose a novel speech-assisted monocular 3D face reconstruction module to exploit multi-modal correlation from the video data rather than merely from the RGB image, since the mouth geometry and expression are naturally correlated to the speech. Specifically, this module is built based on a backbone reconstruction model DECA [22], one of the state-of-the-art monocular 3D face reconstruction methods. By employing a speech-visual contrastive loss, it can learn the correspondence between audio and visemes to make the reconstructed 3D face to be more reliable and perceptually reasonable.", + "bbox": [ + 75, + 444, + 472, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, once sufficient 4D data has been acquired, we can train an audio2mesh network to model the deformation of a face mesh based on input speech. For our work, we employed FaceFormer [20] as the backbone for the audio2mesh network. Compared with the vanilla training scheme of FaceFormer, the primary benefit of Speech4Mesh is data sufficiency. In fact, Speech4Mesh provides a low-cost self-supervised approach to generate \"infinite\" 4D training data from infinite RGB videos. Besides, owing to the diversity of the talking head videos, the reconstructed data empowers our model to further disentangle other auxiliary properties of the face (e.g., emotion, etc.) for manipulation. To be specific, we first pre-train Face-", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Former on the reconstructed emotional 4D dataset with the corresponding emotion codes. At the fine-tuning stage, different emotion codes can be embedded into the FaceFormer backbone to realize the emotion control of the animated talking head.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our work can be summarized as follows:", + "bbox": [ + 498, + 167, + 893, + 196 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Speech4Mesh, a novel framework that tackles the data scarcity issue in speech-driven 3D facial animation by consecutively reconstructing pseudo-4D data and training an audio2mesh network. This framework also provides a referable scheme to alleviate the data scarcity problem of similar 4D tasks.", + "- This framework enables emotional control in the absence of high-quality expressive 4D data by pretraining with the emotional pseudo-4D talking head data reconstructed from 2D videos. And the synthesized emotional 4D talking head dataset can be found in https://github.com/haonanhe/MEAD-3D/tree/main.", + "- To the best of our knowledge, we are the first to utilize speech information as complement for monocular 3D face reconstruction. Empirical experiments show that the exploitation of multi-modal information yields more natural and precise results." + ], + "bbox": [ + 517, + 207, + 890, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 512, + 640, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Speech-driven 3D Facial Animation: Speech-driven facial animation has been widely explored over the years. Existing methods could be categorized into 2D-based approaches which attempt to operate in 2D pixel space [72, 70, 45, 15, 11, 8], and 3D-based approaches that we focus on. Some works animate 3D models based on visemes through complex rules of coarticulation [19, 73, 68, 57, 12]. Though they could capture accurate lip movements and achieve explicit control over the entire process, such methods usually involve plenty of manual adjustments. Data-driven approaches have been proposed to automatically learn a map from acoustic features to 3D models. Most of them were trained in a supervised manner [20, 13, 47, 6, 34] with multi-view motion capture data [4, 6] or high-resolution 4D scans [34, 13]. [6] searches the matching mouth animation for phonemes based on an Anime Graph structure. [34] proposes an end-to-end convolutional network and uses a latent code to control the emotional state of animated mesh. VOCA [13] achieves speaker-independent facial animation by leveraging a subject-specific latent code. MeshTalk [47] learns a categorical latent space to disentangle audio-correlated and audio-uncorrelated information, which enables it to animate the entire face. FaceFormer [20] first applies Transformer to 3D facial animation to make an autore", + "bbox": [ + 496, + 537, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14193", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pressive prediction of the animated 3D face mesh sequence, which claims to be able to capture accurate lip movements.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, 4D scans are usually too expensive to access, which leads to the data scarcity issue. There are methods trying to address this problem by using 4D data generated from videos [56, 65, 44]. [44] trains an LSTM-RNN model to regress the shape parameter of the parametric face model tracked from input video frames. [56] uses a sliding window predictor to learn mappings from phoneme sequences to lower facial shape and appearance that an Active Appearance Model (AAM) tracks from video frames. [65] aims to synthesize 3D talking-head with rich emotion. However, their 3D faces tracked from videos are only optimized by 2D landmarks loss, which may lead to inaccurate reconstruction and animation results. In contrast, our method leverages speech features to capture 3D faces with more accurate facial movements. In addition, pre-training on emotion datasets enables our reconstructed 3D face meshes to have emotion states, which leads to 3D facial animation with rich emotion.", + "bbox": [ + 75, + 125, + 470, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monocular 3D face reconstruction: Monocular 3D face reconstruction aims to track and recover the 3D shape and texture of the human face from monocular 2D data. Early methods often perform an optimization procedure in an analysis-by-synthesis framework, predicting landmarks[75, 62, 30, 5] or local features[29, 50] to regress parameters of 3D Morphable Models (3DMM) [2, 43, 3, 35]. Although optimization-based approaches could recover high-quality 3D faces, they usually involve costly optimization processes. In recent years, learning-based methods were proposed for more efficient reconstruction. Some of them predict 3DMM coefficients [32, 74, 48, 37, 49] directly. To alleviate constraints caused by models' limited geometry, there are also model-free methods directly regressing 3D meshes [16, 18, 23, 33, 53], voxels [31] or Signed Distance Function (SDF) [42].", + "bbox": [ + 75, + 398, + 470, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the greatest limitation of learning-based methods is the lack of paired training data. Some methods address this problem by generating synthetic data or using image-model pairs fitted by traditional optimization-based methods [63, 66, 28, 38, 74, 48, 53]. Others have tried to solve it in an unsupervised manner. Most of unsupervised methods try to regress 3DMM parameters by predicting 2D landmarks [17, 22, 40, 52, 54, 58, 59, 60, 69], photometric constraints [17, 22, 25, 54, 59, 60, 69] or multi-image constraints[17, 22, 25, 52, 58]. Perceptual constraints such as face recognition features [17, 22, 25, 54] or emotion features [14] are also utilized to constrain models' learning objects. While unsupervised methods provide a large amount of training data, reconstructing 3D faces from merely 2D information is still underconstrained. To capture more accurate expressions from monocular videos, our framework leverages speech information as another source", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of constraints, which has a strong relationship with facial movements, especially the mouth area.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 136, + 638, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current monocular 3D face reconstruction methods generally leverage the self-supervised scheme for training. In such a way, a parametric face model and a differentiable renderer are required to recover the 3D face mesh and render the 3D face back to the image for training, respectively. In this section, we review the basic knowledge of the parametric face model and differentiable renderer for reference.", + "bbox": [ + 498, + 162, + 890, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. FLAME", + "text_level": 1, + "bbox": [ + 500, + 273, + 602, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we used FLAME [36] as the prior parametric face model for reconstruction, which is a generic 3D head model using standard vertex-based linear blend skinning (LBS) with corrective blendshapes. Specifically, it consists of $N = 5023$ vertices and could be defined as:", + "bbox": [ + 498, + 297, + 890, + 372 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM (\\boldsymbol {\\beta}, \\boldsymbol {\\theta}, \\boldsymbol {\\psi}) \\in \\mathbb {R} ^ {3 N}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 383, + 890, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which contains shape parameters $\\beta \\in \\mathbb{R}^{|\\beta|}$ accounting for shape variation, expression $\\psi \\in \\mathbb{R}^{|\\psi|}$ to capture facial expressions, and pose parameters $\\theta \\in \\mathbb{R}^{3K + 3}$ to rotate vertices around $K = 4$ joints (neck, jaw, and eyeballs) as well as correcting pose deformations. FLAME provides a geometric constraint and a low-dimensional latent space, making our reconstruction process much easier. Besides, highly disentangled parameters allow us to control the face model's expression or pose independently, which benefits us in both the reconstruction and animation phases.", + "bbox": [ + 496, + 414, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We obtain FLAME's texture map through an appearance model $A(\\alpha) \\in \\mathbb{R}^{d \\times d \\times 3}$ with albedo parameters $\\alpha \\in \\mathbb{R}^{|\\alpha|}$ as input. This is achieved by converting Basel Face Model's linear albedo space to FLAME's UV layout. More details can be found in [35].", + "bbox": [ + 498, + 566, + 890, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Differentiable Rendering", + "text_level": 1, + "bbox": [ + 500, + 650, + 728, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D rendering is a function that maps the 3D object to a 2D image. Differentiable rendering, as the name implies, provides a differentiable rendering function where the gradient of the function with respect to the 3D object parameters can be effectively computed. In this work, we use the differentiable rasterizer from Pytorch3D [46] to render parametric FLAME model $M(\\beta, \\theta, \\psi)$ to 2D images $I_R$ as", + "bbox": [ + 498, + 674, + 890, + 780 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR (M (\\boldsymbol {\\beta}, \\boldsymbol {\\theta}, \\psi), \\boldsymbol {\\alpha}, \\mathbf {l}, \\mathbf {c}) \\rightarrow I _ {R}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 792, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with albedo parameters $\\alpha$ , Spherical Harmonics (SH) lighting parameters $\\mathbf{l} = \\left[\\mathbf{l}_1^T, \\dots, \\mathbf{l}_9^T\\right]^T$ , $\\mathbf{l}_k \\in \\mathbb{R}^3$ to create shaded textures and camera parameters $\\mathbf{c} \\in \\mathbb{R}^3$ which is the concatenation of isotropic scale $s \\in \\mathbb{R}$ and 2D translation $t \\in \\mathbb{R}^2$ .", + "bbox": [ + 498, + 821, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14194", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa7ddd1d4745c95c0e35e7d58047ae79b6071804af14b24a9967df591f9a8f43.jpg", + "image_caption": [ + "Figure 2: Speech-Assisted Monocular 3D Face Reconstruction. The block in grey color are fixed models or networks, and the Expression Encoder $E_{e}$ in blue color is the learnable network." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 459, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Speech4Mesh", + "text_level": 1, + "bbox": [ + 76, + 369, + 217, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Speech-Assisted Monocular 3D Face Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 395, + 468, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the proposed Speech4Mesh framework, 3D face sequences are first reconstructed from RGB videos. The detailed structure of our reconstruction module is illustrated in Fig. 2. The overall structure is built on DECA [22], the state-of-the-art method in monocular 3D face reconstruction. DECA learns animatable detailed 3D face models from static RGB images $I$ in a two-stage scheme. The coarse stage first predicts parameters to reconstruct a rough face shape based on the FLAME model. The detail stage then generates a UV displacement map conditioned on expression and jaw pose parameters to achieve fine-grained details. Since the detail stage may bring noise to the reconstructed mesh, we only borrow the coarse part as our backbone. In DECA's coarse part, an encoder (ResNet-50) learns to regress 3DMM coefficients merely under visual constraints, including landmark-related losses, photometric loss, perceptual-level loss, and shape consistency loss. We will introduce some of the losses we have used later in the text.", + "bbox": [ + 75, + 433, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Expression encoder: Since DECA already can capture the overall facial features well, we fix the pre-trained coarse encoder $E_{c}$ in DECA for shape, pose (except jaw), camera, albedo, and light parameters. We then train an additional expression encoder $E_{e}$ to regress expression parameters $\\psi_{e}$ and jaw parameters $\\theta_{e}^{jaw}$ as follows:", + "bbox": [ + 75, + 720, + 467, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {e} ^ {j a w}, \\psi_ {e} = E _ {e} (I). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 816, + 468, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that, for computational efficiency, we use MobileNetV2 instead of ResNet-50 as the backbone network.", + "bbox": [ + 75, + 840, + 467, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Only training $E_{e}$ simplifies the learning goal and enables us to remove losses that are unrelated to lip motion such as", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "identity loss $L_{id}$ and shape consistency loss $L_{sc}$ in DECA, which leads to faster training speed and fewer training resources. Also, speech content would mainly affect mouth-related movements, which are most relevant to expression and jaw pose. The rendering output of the reconstruction part is then denoted as:", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR \\left(M \\left(\\boldsymbol {\\beta}, [ \\boldsymbol {\\theta}, \\boldsymbol {\\theta} _ {e} ^ {j a w} ], \\psi_ {e}\\right), \\boldsymbol {\\alpha}, 1, \\mathbf {c}\\right)\\rightarrow I _ {R e}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 191, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which $[\\theta, \\theta_{e}^{jaw}]$ means we replace the original vectors representing jaw pose in $\\theta$ with $\\theta_{e}^{jaw}$ .", + "bbox": [ + 496, + 218, + 890, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Speech-visual contrastive loss: As discussed in previous sections, the majority of current works only utilize the visual information from utterance videos. While these methods have achieved acceptable reconstruction results, it is still an underconstrained problem to retrieve 3D faces merely from 2D planes. Ignoring information contained in audio can lead to a considerable loss in reconstruction. Thus, we propose to leverage speech constraints here to achieve more precise 3D face reconstruction in a contrastive manner. We calculate the contrastive loss using a minibatch of N speech-visual representation pairs $(\\epsilon_s,\\epsilon_v)$ , where $\\epsilon_v$ is the hidden state before the last 4-layer MLP in expression encoder, which projects $\\epsilon_v$ to the target 3DMM space, and $\\epsilon_s$ is the audio embedding. In specific, $\\epsilon_s$ is obtained by the pre-trained wav2vec 2.0 [1] model as follows:", + "bbox": [ + 496, + 250, + 892, + 476 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {s} = E _ {s} (S), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 486, + 890, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which can extract powerful representations from speech audio. Since we interpolate the frequency of audio features to be two times the video frame rate, we then adopt a linear combination of audio features from time steps $(t - \\lfloor T / 2\\rfloor)\\cdot 2$ to $(t + \\lfloor T / 2\\rfloor)\\cdot 2$ for frame $t$ as the speech-visual positive pair, where $T$ is set to 5 in our experiment. Minimizing the contrastive loss leads to encoders that maximally preserve the mutual information between positive speech-visual pairs. The contrastive loss consists of two terms, in which the first term calculates the speech-to-visual contrastive cost for the i-th pair:", + "bbox": [ + 496, + 513, + 890, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {i} ^ {(s \\rightarrow v)} = - \\log \\frac {\\exp \\left(\\left\\langle \\epsilon_ {\\mathbf {i}} ^ {\\mathbf {s}}, \\epsilon_ {\\mathbf {i}} ^ {\\mathbf {v}} \\right\\rangle / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(\\left\\langle \\epsilon_ {\\mathbf {i}} ^ {\\mathbf {s}}, \\epsilon_ {\\mathbf {k}} ^ {\\mathbf {v}} \\right\\rangle / \\tau\\right)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 688, + 890, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\langle \\epsilon_{\\mathbf{i}}^{\\mathbf{s}},\\epsilon_{\\mathbf{i}}^{\\mathbf{v}}\\rangle$ calculates cosine similarity between two vectors, i.e., $\\langle \\mathbf{v},\\mathbf{u}\\rangle = \\mathbf{v}^\\top \\mathbf{u} / \\| \\mathbf{v}\\| \\| \\mathbf{u}\\|$ . $\\tau \\in \\mathbb{R}^{+}$ is the temperature parameter, which is fixed to 0.07 in our experiment. While the second term is similar to the first term, it calculates cosine similarity as $\\langle \\epsilon_{\\mathbf{i}}^{\\mathbf{v}},\\epsilon_{\\mathbf{i}}^{\\mathbf{s}}\\rangle$ . The final contrastive loss is the weighted sum of the two terms:", + "bbox": [ + 496, + 734, + 890, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o n} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\lambda \\ell_ {i} ^ {(s \\rightarrow v)} + (1 - \\lambda) \\ell_ {i} ^ {(v \\rightarrow s)}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 834, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we set $\\lambda$ to 0.5 in our experiment as a scalar weight.", + "bbox": [ + 500, + 885, + 885, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14195", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weighted landmark loss: We use weighted landmark loss to minimize the $L_{1}$ distance between ground-truth 2D key points and corresponding landmarks projected into the image plane from mesh.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {l m k} = \\sum_ {i = 1} ^ {6 8} \\lambda_ {i} \\| \\mathbf {k} _ {i} - s \\Pi (M _ {i}) + t \\| _ {1} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 156, + 468, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in which $\\lambda_{i}$ is the weight of the $i$ th landmark, $\\Pi(M_{i})$ represents projecting 3D meshes $M_{i}$ to the 2D pixel space. We set weights of landmarks in the mouth, jaw, and nose area higher than others to capture more accurate geometry representations.", + "bbox": [ + 75, + 202, + 468, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eye closure loss: We calculate the eye closure loss", + "bbox": [ + 96, + 277, + 434, + 292 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e y e} = \\sum_ {(i, j) \\in E} \\| \\mathbf {k} _ {i} - \\mathbf {k} _ {j} - s \\Pi (M _ {i} - M _ {j}) \\| _ {1} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 297, + 468, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to force the relative offset of vertices $M_{i}$ and $M_{j}$ on the upper and lower eyelid to be close to the offset of ground-truth eyelid landmarks pairs $\\mathbf{k}_i$ and $\\mathbf{k}_j$ . In this way, the reconstructed mesh could capture the blink of eye.", + "bbox": [ + 75, + 335, + 468, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lip closure loss: The lip closure loss $L_{lp}$ is similar to the eye closure loss. It measures the relative distance between the upper/lower lip and left/right lip corners keypoint pairs.", + "bbox": [ + 75, + 398, + 468, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Photometric loss: The photometric loss", + "bbox": [ + 96, + 443, + 366, + 457 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p h o} = \\left\\| V _ {I} \\odot (I - I _ {R e}) \\right\\| _ {1, 1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 463, + 468, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "measures the pixel-to-pixel differences between ground-truth training images and rendered images. $V_{I}$ is the face skin mask with a value 1 in the face skin area and $\\odot$ denotes the Hadamard product.", + "bbox": [ + 75, + 486, + 468, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Emotion consistency loss: To capture more vivid expressions, we borrow the idea from EMOCA [14] that introduces an emotion consistency loss with a fixed emotion recognition network. We compute the $L_{2}$ distance between emotion features of input images and output rendered images as follows:", + "bbox": [ + 75, + 546, + 468, + 636 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e m o} = \\left\\| \\epsilon_ {I} ^ {e m o}, \\epsilon_ {R e} ^ {e m o} \\right\\| _ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 642, + 468, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_I^{emo}$ and $\\epsilon_{Re}^{emo}$ are produced by a fixed emotion recognition network.", + "bbox": [ + 75, + 664, + 468, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regularization term: We constrain the expression parameters $\\psi_{e}$ and jaw parameters $\\theta_{e}^{jaw}$ produced by our expression encoder to be close to the original parameters produced by pre-trained DECA by minimizing their $L_{2}$ distance as performed in [24]: $\\| \\psi_{e} - \\psi \\|_{2}$ and $\\left\\| \\theta_e^{jaw} - \\theta^{jaw}\\right\\| _2$", + "bbox": [ + 75, + 696, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss function: To sum up, our optimization goal is:", + "bbox": [ + 96, + 786, + 441, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {r e c o n}} = \\lambda_ {c o n} L _ {c o n} + \\lambda_ {l m k} L _ {l m k} + \\lambda_ {e y e} L _ {e y e} + \\lambda_ {l p} L _ {l p} \\\\ \\lambda_ {p h o} L _ {p h o} + \\lambda_ {e m o} L _ {e m o} + \\lambda_ {r e g} L _ {r e g}, \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 806, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ with different subscripts are hyper-parameters to balance the corresponding losses. The detailed values of $\\lambda$ s can be found in the supplemental material.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Speech-Driven Emotional 3D Facial Animation", + "text_level": 1, + "bbox": [ + 500, + 90, + 890, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After acquiring sufficient synthesized 4D training data, we can train the audio2mesh module for animation. We use FaceFormer [20], a Transformer-based model to drive 3D face vertices with speech embeddings. FaceFormer [20] is the first work to apply Transformer to 3D facial animation. It has achieved better performances in both realism and lip synchronization than prior methods owing to the Transformer's ability to learn long-term audio dependencies. Specifically, FaceFormer predicts the vertices' movement in an autoregressive manner as follows:", + "bbox": [ + 496, + 113, + 890, + 265 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} _ {\\mathbf {t}} = \\operatorname {F a c e F o r m e r} \\left(\\hat {\\mathbf {y}} _ {< \\mathbf {t}}, \\mathbf {i} _ {\\mathbf {n}}, \\mathbf {S}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 277, + 890, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{y}}_{\\mathrm{t}}$ represents the predicted vertices offset at time step $t$ , and $\\mathbf{i}_{\\mathbf{n}}$ represents the identity code. Although FaceFormer has achieved remarkable animation results, its performance highly depends on the data quality, since it is trained in a fully supervised scheme where the only supervision is the MSE loss between predicted and ground-truth meshes:", + "bbox": [ + 496, + 306, + 890, + 397 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S E}} = \\sum_ {\\mathbf {t} = 1} ^ {\\mathbf {T}} \\sum_ {\\mathbf {v} = 1} ^ {\\mathbf {V}} \\left\\| \\hat {\\mathbf {y}} _ {\\mathbf {t}, \\mathbf {v}} - \\mathbf {y} _ {\\mathbf {t}, \\mathbf {v}} \\right\\| ^ {2}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 409, + 890, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In such a way, it's hard to capture some facial details due to the registration loss of the scanned 4D training data. To this end, during pre-training with the synthesized data, we further introduce a pixel-level constraint by computing the photometric loss between the rendered images and corresponding real images as follow:", + "bbox": [ + 496, + 460, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR \\left(M _ {T} + \\hat {\\mathbf {y}} _ {\\mathbf {t}}, \\boldsymbol {\\theta} _ {\\mathbf {t}}, \\boldsymbol {\\alpha} _ {\\mathbf {t}}, \\mathbf {l} _ {\\mathbf {t}}, \\mathbf {c} _ {\\mathbf {t}}\\right)\\rightarrow I _ {R e} ^ {t}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 564, + 890, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_T$ is the template face mesh (the mean of all faces in the training set as in FaceFormer). Note that, when performing differentiable rendering, since the output meshes of the FaceFormer network are all aligned, i.e., without any rotation or camera transform, we borrow $\\theta_t^1$ , $\\alpha_t$ , $l_t$ , $c_t$ estimated by our 3D reconstruction encoders to render the vertices to 2D images. Except for the pixel-to-pixel loss, we also calculate landmark-related losses as in the reconstruction part.", + "bbox": [ + 496, + 593, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, due to the insufficiency of emotional 4D data, current audio2mesh methods generally lack emotion controllability. To make our meshes more lifelike, we introduce a one-hot emotion code to explicitly control the emotional state. Specifically, during pre-training on the emotional 2D video dataset, we embed the emotion category into the FaceFormer network as follows:", + "bbox": [ + 496, + 729, + 890, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} _ {\\mathbf {t}} = \\operatorname {F a c e F o r m e r} \\left(\\hat {\\mathbf {y}} _ {< \\mathbf {t}}, \\mathbf {i} _ {\\mathbf {n}}, \\mathbf {e} _ {\\mathbf {m}}, \\mathbf {S}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 848, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1With slight abuse of notation, the $\\theta_t$ here does not include the jaw rotation.", + "bbox": [ + 500, + 875, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14196", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/45bd049caaf13a367ce7d68d980459b57a782213f60b66bcf43945c592b8857a.jpg", + "image_caption": [ + "Figure 3: Comparison of the reconstruction results on MEAD (left) and VoxCeleb2 (right) datasets. The first column on the left is the original RGB images. The reconstructed faces from the second to the sixth column are from 3DDFA-V2, Deep3DFace (Pytorch), DECA (coarse), EMOCA (coarse), and ours, respectively. More qualitative results could be found in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 869, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{e}_{\\mathbf{m}}$ is the one-hot vector projected to a higher dimension space. In this way, during the inference stage, we can achieve speaker-independent emotion-controllable 3D facial animation by embedding different emotion vectors into FaceFormer.", + "bbox": [ + 75, + 382, + 470, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To sum up, by employing the self-supervised differentiable rendering scheme and emotion code, our modified FaceFormer model can generate not only more realistic but also emotion-controllable animated face meshes.", + "bbox": [ + 75, + 458, + 470, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 534, + 209, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 559, + 171, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, we leveraged RGB video data and 4D mesh data for training the reconstruction module and fine-tuning the audio2mesh module, respectively. A brief introduction of the datasets we used is as follows.", + "bbox": [ + 75, + 583, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2D Training Data: MEAD [64], VoxCeleb2 [10]. The MEAD dataset is a talking head corpus with different intensities of emotions. It has a total of 281,400 video clips collected from 60 English-speaking actors and actresses under 8 common emotions with 3 levels of intensity. Note that although the dataset contains videos from different views, we only took the front-face data for training to keep more visible details. VoxCeleb2 is a large-scale talking head video dataset collected from open-source media which covers a huge range of identities, races, languages, and head poses. In practice, we randomly sampled 2,000 identities for training. The images were extracted under 30 and 25 fps for MEAD and VoxCeleb2 videos, respectively. We used wav2vec 2.0 [1] to exact the audio embeddings, in which the output frequency is $49\\mathrm{Hz}$ . To ensure temporal alignment of the image and audio, we interpolate the audio embeddings to $50\\mathrm{Hz}$ for the VoxCeleb2 dataset and to $60\\mathrm{Hz}$", + "bbox": [ + 75, + 643, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the MEAD dataset.", + "bbox": [ + 498, + 382, + 653, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4D Fine-Tuning Data: VOCASET [13]. VOCASET is a 4D dataset composed of consecutive scanned 3D heads which are all registered to a 5023 vertices mesh. It contains a total of 480 talking motion sequences from 12 subjects under the frame rate of 60 fps. The training, validation, and testing splits are directly adopted from [20].", + "bbox": [ + 498, + 397, + 890, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Training Setup", + "text_level": 1, + "bbox": [ + 500, + 498, + 651, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The reconstruction part was trained for a maximum of 20 epochs, using the Adam optimizer and a learning rate of $5e - 5$ . In one iteration, we used 4 mini-batches of data, each of which contains 16 frames randomly sampled from a single video. We further implied distributed data-parallel technique to collect data from 4 GPUs to enlarge the pool of training pairs for contrastive learning.", + "bbox": [ + 498, + 522, + 890, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The audio2mesh network was trained using the Adam optimizer with a learning rate of $1e - 4$ for a total of 100 epochs. The parameters of the encoder were initialized with the pre-trained wav2vec 2.0 weights and the parameters of TCN were fixed during training. The period $p$ for the biased causal multi-head self-attention was set to 30.", + "bbox": [ + 496, + 628, + 890, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We pre-trained our audio2mesh network on 4D data produced by our monocular 3D reconstruction network from the MEAD dataset. While it contains 3 levels of emotional intensity, we only use the mildest first level to train. We created 8 one-hot emotion codes (angry, contempt, disgusted, fear, happy, sad, surprised, or neutral) and 42 one-hot speaker identity codes. For each pair of image and its 3D reconstruction in the MEAD dataset, we embedded its corresponding emotion code and speaker ID code to a higher dimension and then concatenated and projected it to the audio2mesh network.", + "bbox": [ + 496, + 719, + 890, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We fine-tuned the audio2mesh network on VOCASET", + "bbox": [ + 519, + 885, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14197", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which contains high-resolution 4D scans. During training, we randomly chose 8 kinds of speaker ID codes defined in the pre-train phase and set the emotion code to neutral.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Analysis of 3D Face Reconstruction Results", + "text_level": 1, + "bbox": [ + 76, + 148, + 446, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subjective Evaluation: Visualization of the reconstruction results is demonstrated in Fig. 3. We randomly sampled several frames from MEAD and VoxCeleb2 datasets for comparison. For clear observation of the expression details, we did not apply texture on the face. The sampled frames contain a variety of talking heads with various lip motions, extreme emotions, different genders, head poses, ages, and races. From Fig. 3, we can find that our method apparently outperforms the 3DDFA-V2 and DECA regarding the lip motion as well as the overall appearance. The curly lips and emotional expressions cannot be recovered by these two baselines. EMOCA is an expert in capturing extreme emotions. However, some mild expressions will be significantly exaggerated by EMOCA, which causes unnatural artifacts and inaccurate lip shapes in the reconstruction results. More comparison results with EMOCA could be found in supplementary materials. Note that, we used the Pytorch version of Deep3DFace for comparison, which is claimed to be better than the Tensorflow implementation. As shown in the third column in Fig. 3, the recovered faces by Deep3DFace reveal impressive identity consistency compared with the original RGB image. This may attribute to the identity aggregation mechanism and a different parametric face model (i.e., Basel Face Model [43]). However, the lower face expressiveness is still lower than our proposed method.", + "bbox": [ + 75, + 172, + 472, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further conducted a user study to make a perceptual evaluation between our and the competitors' reconstruction results. Specifically, 150 participants were asked to judge a side-by-side image composed of the original image, our reconstruction result, and the competitor's result. The participants can select their favorable one or choose the equal option if they cannot distinguish the difference. 50 samples were randomly selected from the MEAD and Voxceleb2 datasets. The result of the study is summarized in Tab. 1.", + "bbox": [ + 75, + 551, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As we can see from Tab. 1, the result demonstrates an obvious inclination to support our results. It is worth mentioning that about 80 percent of our reconstructed samples are deemed better than or equal with 3DDFA-V2 and DECA methods regarding both full and lower face. EMOCA improves the perception of the results due to its emotion prior. Deep3DFace also achieves better performance as we discussed above. However, owing to the audio-visual alignment, our method achieves better performance at both full-face and the low-face region according to the user study.", + "bbox": [ + 75, + 688, + 468, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Objective Evaluation: Our reconstruction model's ability to capture lip shapes is measured using the lip vertex error as the objective metric, which has been extensively leveraged to evaluate the precision of lip synchronization", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4cb19f1070bd50aeaffb05dede2642387ca0a7363120434843da6e1bd5ca961c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ours vs. CompetitorFace regionFavorability(%) ↑Ours better or equal
EqualComp.Ours
Ours vs. 3DDFA-V2Full5.419.475.280.6
Lower5.421.373.378.7
Ours vs. DECAFull11.418.869.881.2
Lower9.818.571.781.5
Ours vs. EMOCAFull8.438.353.358.9
Lower8.337.254.562.8
Ours vs. Deep3DFaceFull5.941.153.158.9
Lower8.734.257.065.8
", + "bbox": [ + 514, + 88, + 879, + 198 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c381aa38cc2fc73446f2887157cd44d4b041542aaf1ea67aa082d9dfad6b71d1.jpg", + "table_caption": [ + "Table 1: A user study of the reconstruction results. We compare the participants' favorability of our reconstructions with the competitors' w.r.t. different regions of the face." + ], + "table_footnote": [], + "table_body": "
MetricOursEMOCA
Lip Vertex Error ↓0.995e-21.359e-2
", + "bbox": [ + 571, + 252, + 818, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Reconstruction error on VOCASET.", + "bbox": [ + 545, + 290, + 844, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[20, 47]. To calculate this error, we determine the maximal $L_{2}$ distance between the lip vertices of the reconstructed meshes and their corresponding ground truth in the VO-CASET. We compared Speech4Mesh with EMOCA, which performs the best in the subjective evaluation. As presented in Table 2, our method outperforms EMOCA in terms of lower lip vertex error, demonstrating our superior ability to capture accurate lip shapes. On the contrary, EMOCA performs poorly on this metric due to its exaggerated facial expression with contorted mouth movements.", + "bbox": [ + 496, + 306, + 890, + 457 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Analysis of Speech-Driven Facial Animation Results", + "text_level": 1, + "bbox": [ + 498, + 468, + 890, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subjective Evaluation: We also performed a user study on the speech-driven animation results. Similar to the reconstruction user study, we recruited 150 participants to judge the animation quality of our method and the vanilla FaceFormer from the realism and lip synchronization dimensions. The speeches for evaluation are from FaceFormer's test data (the demo video of FaceFormer) and the VoxCeleb2 dataset. For each dataset, we sampled 56 speeches (7 ids, 8 sentences for each id). Each participant evaluates 8 videos selected at random. The statistic of the user study is shown in Tab. 3. From Tab. 3 we can find that the favorability rating of our Speech4Mesh is consistently higher than that of FaceFormer, especially on the more challenging VoxCeleb2 speeches. We also visualize several representative syllables and corresponding animations in Fig. 4 for demonstration. In Fig. 4, our animated faces are more expressive and perceptually reasonable. We would like to note that MeshTalk cannot be well-trained on VOCASET which in a different topology, as mentioned in [61]. We have refrained from presenting any quantitative or qualitative comparison results here. However, we do incorporate a straightforward comparison in the supplementary videos. It is highly recommended to see those videos for more details.", + "bbox": [ + 496, + 507, + 892, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Objective Evaluation: We evaluate the lip vertex error of our proposed method and the vanilla FaceFormer on the test set of VOCASET. For simplicity, the scale we used for eval", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14198", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e9116db4c11dd58686f94a168a900af0c1ebb2f2585e838bd367959e299b63ca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Test Speech DataCriterionFavorability(%) ↑Ours better or equal
EqualFaceFormerSpeech4Mesh
Speech from FaceFormer DemoRealism24.831.843.468.2
Lip Sync21.936.441.863.6
Speech from VoxCeleb2Realism16.326.157.773.9
Lip Sync16.730.752.569.3
", + "bbox": [ + 80, + 88, + 468, + 185 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/77ec48d62bf1a7affe8a387848a75f25b960203dc8736f042f35166339c17ea3.jpg", + "table_caption": [ + "Table 3: A user study of the speech-driven results. We compare the participants' favorability of our results with the FaceFormer with respect to realism and lip synchronization." + ], + "table_footnote": [], + "table_body": "
MethodLip Vertex Error ↓
VOCA6.428e-3
FaceFormer5.355e-3
Speech4Mesh4.863e-3
", + "bbox": [ + 148, + 242, + 393, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Comparison of the lip-sync error between VOCA, the vanilla FaceFormer, and Speech4Mesh.", + "bbox": [ + 75, + 316, + 468, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "uation is the same as the FLAME model.", + "bbox": [ + 76, + 349, + 349, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 4 displays the lip vertex errors of the vanilla Face-Former method and our Speech4Mesh method. It can be seen that more than a $9\\%$ improvement in lip synchronization is brought by our training framework, which restates the importance of data sufficiency for 4D tasks.", + "bbox": [ + 75, + 367, + 468, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Emotion Control", + "text_level": 1, + "bbox": [ + 76, + 450, + 243, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As introduced in previous sections, our audio2mesh module can be trained on the reconstructed meshes with different emotions. By employing an emotion code, we can manipulate the originally neural 4D talking heads to have different emotions. As shown in Fig. 5, different emotions (e.g., happy, surprised, sad) are embedded into the talking face driven by the same speech. The result reveals the potentiality of more diverse applications by exploiting the 2D videos under the Speech4Mesh framework. More emotion control results could be found in supplementary videos.", + "bbox": [ + 75, + 474, + 468, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 633, + 241, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To understand the impact of the contrastive loss and the photometric loss on speech-driven animation, we conducted quantitative experiments as shown in Tab. 5. More ablation studies could be found in supplementary materials. For the ablation study of contrastive loss, we pre-trained our audio2mesh module using 4D meshes reconstructed from the MEAD dataset, without incorporating contrastive loss during training. For the ablation study of photometric loss, we simply removed it in the pre-training stage. The quantitative results in Tab. 5. show that pre-training without contrastive loss or photometric loss results in a higher lip vertex error. These results underscore the importance of high-quality reconstructed 4D meshes in our method. They also support the notion that the photometric loss can encourage the audio2mesh network to learn detailed information which can not be captured solely from 4D meshes.", + "bbox": [ + 75, + 657, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19755bcc633b5ef5f2330fec81ab25631352b285c8cb374bf567f94378903fd7.jpg", + "image_caption": [ + "Figure 4: Representative syllables and corresponding animations. The animation results of our Speech4Mesh and FaceFormer are displayed in the first and second columns." + ], + "image_footnote": [], + "bbox": [ + 540, + 88, + 854, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8753daa8ceb5377c021443de597afbfdc507cd40bf4bda7e7106ba3fcd768100.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 279, + 805, + 441 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8f5aa289bb75f7ecfc25db3378732960ed731dbe015ffe65e1f5388db345350f.jpg", + "table_caption": [ + "Figure 5: Speech4Mesh samples with different emotions." + ], + "table_footnote": [], + "table_body": "
MetricFullwo. Pho.wo. Con.FaceFormer
Lip Vertex Error ↓4.863e-35.233e-35.271e-35.355e-3
", + "bbox": [ + 511, + 467, + 877, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: Ablation study on the photometric loss and contrastive loss w.r.t. the lip-sync error.", + "bbox": [ + 498, + 503, + 888, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions and Future Work", + "text_level": 1, + "bbox": [ + 500, + 540, + 777, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel framework named Speech4Mesh to train the audio2mesh network for speech-driven 3D facial animation. In Speech4Mesh, the speech-assisted monocular 3D face reconstruction module can provide significantly more accurate 4D synthesized data for pre-training by leveraging the audio-visual contrastive alignment. By employing the differentiable rendering loss and an emotion code, the modified audio2mesh network can achieve more realistic animation performance and unprecedentedly realize emotion manipulation. Empirical experiments verify the effectiveness and superiority of the Speech4Mesh framework both quantitatively and perceptually. We hope that the proposed Speech4Mesh framework can bring inspiration for similar 4D tasks to address the data scarcity problem.", + "bbox": [ + 496, + 566, + 890, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the future, we plan to replace the image-based reconstruction backbone DECA with a video-based reconstruction backbone such as MICA [76], neural head avatar [26] or [67]. We will also leverage more high-quality 2D video data for better animation performance and devise a more effective emotion modeling scheme to improve the expressiveness of expressions.", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14199", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems, 33:12449-12460, 2020. 4, 6", + "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3", + "[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5543-5552, 2016. 3", + "[4] Carlos Busso, Zhigang Deng, Michael Grimm, Ulrich Neumann, and Shrikanth Narayanan. Rigid head motion in expressive speech animation: Analysis and synthesis. IEEE transactions on audio, speech, and language processing, 15(3):1075-1086, 2007. 2", + "[5] Chen Cao, Qiming Hou, and Kun Zhou. Displaced dynamic expression regression for real-time facial tracking and animation. ACM Trans. Graph, 33(4):1-10, 2014. 3", + "[6] Yong Cao, Wen C Tien, Petros Faloutsos, and Frédéric Pighin. Expressive speech-driven facial animation. ACM Transactions on Graphics (TOG), 24(4):1283-1302, 2005. 2", + "[7] Yujin Chai, Yanlin Weng, Lvdi Wang, and Kun Zhou. Speech-driven facial animation with spectral gathering and temporal attention. Frontiers of Computer Science, 16(3):1-10, 2022. 1", + "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 2", + "[9] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 1", + "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. Proc. Interspeech 2018, pages 1086-1090, 2018. 6", + "[11] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian conference on computer vision, pages 251-263. Springer, 2016. 2", + "[12] Michael M Cohen, Rashid Clark, and Dominic W Massaro. Animated speech: Research progress and applications. In AVSP 2001-International Conference on Auditory-Visual Speech Processing, 2001. 2", + "[13] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 1, 2, 6", + "[14] Radek Daneček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In CVPR, pages 20311-20322, 2022. 3, 5" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European conference on computer vision, pages 408-424. Springer, 2020. 2", + "[16] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 3", + "[17] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 3", + "[18] Pengfei Dou, Shishir K Shah, and Ioannis A Kakadiaris. End-to-end 3d face reconstruction with deep neural networks. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 5908-5917, 2017. 3", + "[19] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph, 35(4):1-11, 2016. 2", + "[20] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 1, 2, 5, 6, 7", + "[21] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2", + "[22] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Trans. Graph., 40(4):1-13, 2021. 2, 3, 4", + "[23] Yao Feng, Fan Wu, Xiaohu Shao, Yanfeng Wang, and Xi Zhou. Joint 3d face reconstruction and dense alignment with position map regression network. In Proceedings of the European conference on computer vision, pages 534-551, 2018. 3", + "[24] Panagiotis P Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Visual speech-aware perceptual 3d facial expression reconstruction from videos. arXiv preprint arXiv:2207.11094, 2022.5", + "[25] Kyle Genova, Forrester Cole, Aaron Maschinot, Aaron Sarna, Daniel Vlasic, and William T Freeman. Unsupervised training for 3d morphable model regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8377-8386, 2018. 3", + "[26] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 8" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14200", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Jianzhu Guo, Xiangyu Zhu, Yang Yang, Fan Yang, Zhen Lei, and Stan Z Li. Towards fast, accurate and stable 3d dense face alignment. In European Conference on Computer Vision, pages 152-168. Springer, 2020. 2", + "[28] Yudong Guo, Jianfei Cai, Boyi Jiang, Jianmin Zheng, et al. CNN-based real-time dense face reconstruction with inverse-rendered photo-realistic face images. IEEE transactions on pattern analysis and machine intelligence, 41(6):1294-1307, 2018. 3", + "[29] Patrik Huber, Zhen-Hua Feng, William Christmas, Josef Kittler, and Matthias Ratsch. Fitting 3d morphable face models using local features. In 2015 IEEE international conference on image processing, pages 1195-1199. IEEE, 2015. 3", + "[30] Patrik Huber, Guosheng Hu, Rafael Tena, Pouria Mortazavian, P Koppen, William J Christmas, Matthias Ratsch, and Josef Kittler. A multiresolution 3d morphable face model and fitting framework. In Proceedings of the 11th international joint conference on computer vision, imaging and computer graphics theory and applications, 2016. 3", + "[31] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In Proceedings of the IEEE international conference on computer vision, pages 1031-1039, 2017. 3", + "[32] Amin Jourabloo and Xiaoming Liu. Large-pose face alignment via cnn-based dense 3d model fitting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4188–4196, 2016. 3", + "[33] Harim Jung, Myeong-Seok Oh, and Seong-Whan Lee. Learning free-form deformation for 3d face reconstruction from in-the-wild images. In 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 2737–2742. IEEE, 2021. 3", + "[34] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Trans. Graph., 36(4):1-12, 2017. 1, 2", + "[35] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3", + "[36] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3", + "[37] Feng Liu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Joint face alignment and 3d face reconstruction. In European Conference on Computer Vision, pages 545-560. Springer, 2016. 3", + "[38] Feng Liu, Ronghang Zhu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Disentangling features in 3d face shapes for joint face reconstruction and recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5216-5225, 2018. 3", + "[39] Jingying Liu, Binyuan Hui, Kun Li, Yunke Liu, Yu-Kun Lai, Yuxiang Zhang, Yebin Liu, and Jingyu Yang. Geometry-guided dense perspective network for speech-driven facial animation. IEEE Transactions on Visualization and Computer Graphics, 2021. 1" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Yaojie Liu, Amin Jourabloo, William Ren, and Xiaoming Liu. Dense face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1619-1628, 2017. 3", + "[41] Gaurav Mittal and Baoyuan Wang. Animating face using disentangled audio representations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3290–3298, 2020. 1", + "[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3", + "[43] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 3, 7", + "[44] Hai X Pham, Samuel Cheung, and Vladimir Pavlovic. Speech-driven 3d facial animation with implicit emotional awareness: a deep learning approach. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 80-88, 2017. 3", + "[45] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2", + "[46] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 3", + "[47] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 1, 2, 7", + "[48] Elad Richardson, Matan Sela, and Ron Kimmel. 3d face reconstruction by learning from synthetic data. In 2016 fourth international conference on 3D vision (3DV), pages 460-469. IEEE, 2016. 3", + "[49] Elad Richardson, Matan Sela, Roy Or-El, and Ron Kimmel. Learning detailed face reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1259–1268, 2017. 3", + "[50] Sami Romdhani and Thomas Vetter. Estimating 3d shape and texture using pixel intensity, edges, specular highlights, texture constraints and a prior. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 2, pages 986-993. IEEE, 2005. 3", + "[51] Shunsuke Saito, Tianye Li, and Hao Li. Real-time facial segmentation and performance capture from rgb input. In European conference on computer vision, pages 244–261. Springer, 2016. 2", + "[52] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14201", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7763-7772, 2019. 3", + "[53] Matan Sela, Elad Richardson, and Ron Kimmel. Unrestricted facial geometry reconstruction using image-to-image translation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1576-1585, 2017. 3", + "[54] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Mingmin Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In European Conference on Computer Vision, pages 53-70. Springer, 2020. 3", + "[55] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Trans. Graph., 36(4):1-13, 2017. 1", + "[56] Sarah Taylor, Taehwan Kim, Yisong Yue, Moshe Mahler, James Krahe, Anastasio Garcia Rodriguez, Jessica Hodgins, and Iain Matthews. A deep learning approach for generalized speech animation. ACM Transactions on Graphics (TOG), 36(4):1-11, 2017. 3", + "[57] Sarah L Taylor, Moshe Mahler, Barry-John Theobald, and Iain Matthews. Dynamic units of visual speech. In Proceedings of the 11th ACM SIGGRAPH/Eurographics conference on Computer Animation, pages 275-284, 2012. 2", + "[58] Ayush Tewari, Florian Bernard, Pablo Garrido, Gaurav Bharaj, Mohamed Elgharib, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Fml: Face model learning from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10812–10822, 2019. 3", + "[59] Ayush Tewari, Michael Zollhöfer, Pablo Garrido, Florian Bernard, Hyeongwoo Kim, Patrick Pérez, and Christian Theobalt. Self-supervised multi-level face model learning for monocular reconstruction at over $250\\mathrm{~hz}$ . In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2549-2559, 2018. 3", + "[60] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1274-1283, 2017. 3", + "[61] Balamurugan Thambiraja, Ikhsanul Habibie, Sadegh Aliakbarian, Darren Cosker, Christian Theobalt, and Justus Thies. Imitator: Personalized speech-driven 3d facial animation. arXiv preprint arXiv:2301.00023, 2022. 7", + "[62] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3", + "[63] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5163-5172, 2017. 3", + "[64] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700-717. Springer, 2020. 6", + "[65] Qianyun Wang, Zhenfeng Fan, and Shihong Xia. 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051, 2021. 3", + "[66] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 3", + "[67] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljkovic, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 8", + "[68] Yuyu Xu, Andrew W Feng, Stacy Marsella, and Ari Shapiro. A practical and configurable lip sync method for games. In Proceedings of Motion on Games, pages 131-140, 2013. 2", + "[69] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 3", + "[70] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2", + "[71] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, pages 9299–9306, 2019. 1", + "[72] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2", + "[73] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2", + "[74] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 3", + "[75] Xiangyu Zhu, Zhen Lei, Junjie Yan, Dong Yi, and Stan Z Li. High-fidelity pose and expression normalization for face recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 787-796, 2015. 3", + "[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision, pages 250–269. Springer, 2022. 8" + ], + "bbox": [ + 503, + 92, + 890, + 890 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14202", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/236896c7-9d5b-41e0-8f54-b3702ca337d4_model.json b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/236896c7-9d5b-41e0-8f54-b3702ca337d4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bc9c50c0ca6f8950df46003cc1f2ae7607d11a80 --- /dev/null +++ b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/236896c7-9d5b-41e0-8f54-b3702ca337d4_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.13, + 0.85, + 0.175 + ], + "angle": 0, + "content": "Speech4Mesh: Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.203, + 0.764, + 0.24 + ], + "angle": 0, + "content": "Shan He\\(^{1,2}\\), Haonan He\\(^{2}\\), Shuo Yang\\(^{2}\\), Xiaoyan Wu\\(^{2}\\), Pengcheng Xia\\(^{2}\\), Bing Yin\\(^{2}\\), Cong Liu\\(^{2}\\), Lirong Dai\\(^{1}\\), Chang Xu\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.239, + 0.856, + 0.293 + ], + "angle": 0, + "content": "1University of Science and Technology of China, 2iFLYTEK Research, 3University of Sydney {shanhe2, hnhe, shuoyang7, xywu10, pcxia, bingyin, congliu2}@iflytek.com lrdai@ustc.edu.cn, c.xu@sydney.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.36, + 0.474, + 0.785 + ], + "angle": 0, + "content": "Recent audio2mesh-based methods have shown promising prospects for speech-driven 3D facial animation tasks. However, some intractable challenges are urgent to be settled. For example, the data-scarcity problem is intrinsically inevitable due to the difficulty of 4D data collection. Besides, current methods generally lack controllability on the animated face. To this end, we propose a novel framework named Speech4Mesh to consecutively generate 4D talking head data and train the audio2mesh network with the reconstructed meshes. In our framework, we first reconstruct the 4D talking head sequence based on the monocular videos. For precise capture of the talking-related variation on the face, we exploit the audio-visual alignment information from the video by employing a contrastive learning scheme. We next can train the audio2mesh network (e.g., FaceFormer) based on the generated 4D data. To get control of the animated talking face, we encode the speaking-unrelated factors (e.g., emotion, etc.) into an emotion embedding for manipulation. Finally, a differentiable renderer guarantees more accurate photometric details of the reconstruction and animation results. Empirical experiments demonstrate that the Speech4Mesh framework can not only outperform state-of-the-art reconstruction methods, especially on the lower-face part but also achieve better animation performance both perceptually and objectively after pre-trained on the synthesized data. Besides, we also verify that the proposed framework is able to explicitly control the emotion of the animated talking face." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.814, + 0.21, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Speech-driven facial animation is aimed at guiding the face model (either 2D image or 3D mesh) to have perceptually rational motion (especially for lip sync) only according to the input audio signal. It is drawing increasingly more" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.327, + 0.882, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.631 + ], + "angle": 0, + "content": "Figure 1: Overview of the proposed Speech4Mesh framework. In this framework, the speech signal is leveraged for both the reconstruction module and the audio2mesh module. An emotion code is embedded into the audio2mesh module for emotional manipulation. Differentiable rendering is employed for recovering the details of the face." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.895, + 0.895 + ], + "angle": 0, + "content": "attention with the development of virtual reality, game and film production, etc. There are currently two mainstreams to deal with this task, either from 2D or 3D perspectives. By learning from video data [9, 41, 55, 71], 2D solutions try to generate the talking head images frame by frame. Although sufficient data is available to be used, the synthesized images cannot directly be applied to 3D scenes. Thus, in this paper, we mainly focus on 3D facial animation driven by speech. In recent years, a handful of 3D speech-driven facial animation methods [20, 47, 39, 13, 7, 34] have shown promising performance by directly learning a mapping from audio to 3D avatar mesh deformation with deep neural networks (hereinafter called audio2mesh methods). In this way, some subtle variations on the face mesh can be captured. Generally speaking, for the audio2mesh methods, data quality is one of the most important factors to be considered [20] since they generally demand high-precision" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14192" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "4D talking head data to learn the correspondence from audio embedding to mesh movement in a supervised manner. However, in practice, real 4D data is extremely hard to collect. It is generally reconstructed and registered from synchronized multi-view images or 3D scan frame by frame, which is extraordinarily costly [13, 21]. Therefore, how to deal with the data scarcity problem becomes an urgent agenda to be discussed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.216, + 0.473, + 0.442 + ], + "angle": 0, + "content": "In contrast, the accessibility of monocular videos is much higher. The scarcity of 4D data may be mitigated through the use of 3D reconstruction techniques applied to 2D videos. The majority of current monocular 3D face reconstruction methods (e.g., 3DDFA-V2 [27], DECA [22]) are based on the self-supervised training scheme, whose essential guidances are landmark loss and photometric loss. Nonetheless, these losses can only provide 2D information from the image plane, which is relatively weak to capture the complex nonrigid deformation of the lower face region [51, 62]. Therefore, the reconstructed results always demonstrate obvious artifacts and ambiguities between utterance and visual perception, such as mouth funnel, pucker, and rounded vowels, which may considerably limit the precision of the audio2mesh methods [20]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.702 + ], + "angle": 0, + "content": "To address the problems mentioned above, we propose a novel framework named Speech4Mesh to consecutively synthesize 4D talking head data and train the audio2mesh network with the reconstructed mesh sequences. An overview of the Speech4Mesh framework is illustrated in Fig. 1. Firstly, to obtain better synthesis results, we propose a novel speech-assisted monocular 3D face reconstruction module to exploit multi-modal correlation from the video data rather than merely from the RGB image, since the mouth geometry and expression are naturally correlated to the speech. Specifically, this module is built based on a backbone reconstruction model DECA [22], one of the state-of-the-art monocular 3D face reconstruction methods. By employing a speech-visual contrastive loss, it can learn the correspondence between audio and visemes to make the reconstructed 3D face to be more reliable and perceptually reasonable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Secondly, once sufficient 4D data has been acquired, we can train an audio2mesh network to model the deformation of a face mesh based on input speech. For our work, we employed FaceFormer [20] as the backbone for the audio2mesh network. Compared with the vanilla training scheme of FaceFormer, the primary benefit of Speech4Mesh is data sufficiency. In fact, Speech4Mesh provides a low-cost self-supervised approach to generate \"infinite\" 4D training data from infinite RGB videos. Besides, owing to the diversity of the talking head videos, the reconstructed data empowers our model to further disentangle other auxiliary properties of the face (e.g., emotion, etc.) for manipulation. To be specific, we first pre-train Face-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "Former on the reconstructed emotional 4D dataset with the corresponding emotion codes. At the fine-tuning stage, different emotion codes can be embedded into the FaceFormer backbone to realize the emotion control of the animated talking head." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.168, + 0.894, + 0.197 + ], + "angle": 0, + "content": "The main contributions of our work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.208, + 0.892, + 0.298 + ], + "angle": 0, + "content": "- We propose Speech4Mesh, a novel framework that tackles the data scarcity issue in speech-driven 3D facial animation by consecutively reconstructing pseudo-4D data and training an audio2mesh network. This framework also provides a referable scheme to alleviate the data scarcity problem of similar 4D tasks." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.308, + 0.892, + 0.413 + ], + "angle": 0, + "content": "- This framework enables emotional control in the absence of high-quality expressive 4D data by pretraining with the emotional pseudo-4D talking head data reconstructed from 2D videos. And the synthesized emotional 4D talking head dataset can be found in https://github.com/haonanhe/MEAD-3D/tree/main." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.424, + 0.892, + 0.5 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to utilize speech information as complement for monocular 3D face reconstruction. Empirical experiments show that the exploitation of multi-modal information yields more natural and precise results." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.208, + 0.892, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.642, + 0.529 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Speech-driven 3D Facial Animation: Speech-driven facial animation has been widely explored over the years. Existing methods could be categorized into 2D-based approaches which attempt to operate in 2D pixel space [72, 70, 45, 15, 11, 8], and 3D-based approaches that we focus on. Some works animate 3D models based on visemes through complex rules of coarticulation [19, 73, 68, 57, 12]. Though they could capture accurate lip movements and achieve explicit control over the entire process, such methods usually involve plenty of manual adjustments. Data-driven approaches have been proposed to automatically learn a map from acoustic features to 3D models. Most of them were trained in a supervised manner [20, 13, 47, 6, 34] with multi-view motion capture data [4, 6] or high-resolution 4D scans [34, 13]. [6] searches the matching mouth animation for phonemes based on an Anime Graph structure. [34] proposes an end-to-end convolutional network and uses a latent code to control the emotional state of animated mesh. VOCA [13] achieves speaker-independent facial animation by leveraging a subject-specific latent code. MeshTalk [47] learns a categorical latent space to disentangle audio-correlated and audio-uncorrelated information, which enables it to animate the entire face. FaceFormer [20] first applies Transformer to 3D facial animation to make an autore" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14193" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "pressive prediction of the animated 3D face mesh sequence, which claims to be able to capture accurate lip movements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.125, + 0.471, + 0.395 + ], + "angle": 0, + "content": "However, 4D scans are usually too expensive to access, which leads to the data scarcity issue. There are methods trying to address this problem by using 4D data generated from videos [56, 65, 44]. [44] trains an LSTM-RNN model to regress the shape parameter of the parametric face model tracked from input video frames. [56] uses a sliding window predictor to learn mappings from phoneme sequences to lower facial shape and appearance that an Active Appearance Model (AAM) tracks from video frames. [65] aims to synthesize 3D talking-head with rich emotion. However, their 3D faces tracked from videos are only optimized by 2D landmarks loss, which may lead to inaccurate reconstruction and animation results. In contrast, our method leverages speech features to capture 3D faces with more accurate facial movements. In addition, pre-training on emotion datasets enables our reconstructed 3D face meshes to have emotion states, which leads to 3D facial animation with rich emotion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.471, + 0.642 + ], + "angle": 0, + "content": "Monocular 3D face reconstruction: Monocular 3D face reconstruction aims to track and recover the 3D shape and texture of the human face from monocular 2D data. Early methods often perform an optimization procedure in an analysis-by-synthesis framework, predicting landmarks[75, 62, 30, 5] or local features[29, 50] to regress parameters of 3D Morphable Models (3DMM) [2, 43, 3, 35]. Although optimization-based approaches could recover high-quality 3D faces, they usually involve costly optimization processes. In recent years, learning-based methods were proposed for more efficient reconstruction. Some of them predict 3DMM coefficients [32, 74, 48, 37, 49] directly. To alleviate constraints caused by models' limited geometry, there are also model-free methods directly regressing 3D meshes [16, 18, 23, 33, 53], voxels [31] or Signed Distance Function (SDF) [42]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.902 + ], + "angle": 0, + "content": "However, the greatest limitation of learning-based methods is the lack of paired training data. Some methods address this problem by generating synthetic data or using image-model pairs fitted by traditional optimization-based methods [63, 66, 28, 38, 74, 48, 53]. Others have tried to solve it in an unsupervised manner. Most of unsupervised methods try to regress 3DMM parameters by predicting 2D landmarks [17, 22, 40, 52, 54, 58, 59, 60, 69], photometric constraints [17, 22, 25, 54, 59, 60, 69] or multi-image constraints[17, 22, 25, 52, 58]. Perceptual constraints such as face recognition features [17, 22, 25, 54] or emotion features [14] are also utilized to constrain models' learning objects. While unsupervised methods provide a large amount of training data, reconstructing 3D faces from merely 2D information is still underconstrained. To capture more accurate expressions from monocular videos, our framework leverages speech information as another source" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "of constraints, which has a strong relationship with facial movements, especially the mouth area." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.137, + 0.64, + 0.154 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.163, + 0.892, + 0.269 + ], + "angle": 0, + "content": "Current monocular 3D face reconstruction methods generally leverage the self-supervised scheme for training. In such a way, a parametric face model and a differentiable renderer are required to recover the 3D face mesh and render the 3D face back to the image for training, respectively. In this section, we review the basic knowledge of the parametric face model and differentiable renderer for reference." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.274, + 0.603, + 0.289 + ], + "angle": 0, + "content": "3.1. FLAME" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.298, + 0.892, + 0.373 + ], + "angle": 0, + "content": "In this work, we used FLAME [36] as the prior parametric face model for reconstruction, which is a generic 3D head model using standard vertex-based linear blend skinning (LBS) with corrective blendshapes. Specifically, it consists of \\(N = 5023\\) vertices and could be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.385, + 0.892, + 0.404 + ], + "angle": 0, + "content": "\\[\nM (\\boldsymbol {\\beta}, \\boldsymbol {\\theta}, \\boldsymbol {\\psi}) \\in \\mathbb {R} ^ {3 N}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.567 + ], + "angle": 0, + "content": "which contains shape parameters \\(\\beta \\in \\mathbb{R}^{|\\beta|}\\) accounting for shape variation, expression \\(\\psi \\in \\mathbb{R}^{|\\psi|}\\) to capture facial expressions, and pose parameters \\(\\theta \\in \\mathbb{R}^{3K + 3}\\) to rotate vertices around \\(K = 4\\) joints (neck, jaw, and eyeballs) as well as correcting pose deformations. FLAME provides a geometric constraint and a low-dimensional latent space, making our reconstruction process much easier. Besides, highly disentangled parameters allow us to control the face model's expression or pose independently, which benefits us in both the reconstruction and animation phases." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.568, + 0.892, + 0.643 + ], + "angle": 0, + "content": "We obtain FLAME's texture map through an appearance model \\( A(\\alpha) \\in \\mathbb{R}^{d \\times d \\times 3} \\) with albedo parameters \\( \\alpha \\in \\mathbb{R}^{|\\alpha|} \\) as input. This is achieved by converting Basel Face Model's linear albedo space to FLAME's UV layout. More details can be found in [35]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.729, + 0.667 + ], + "angle": 0, + "content": "3.2. Differentiable Rendering" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.675, + 0.892, + 0.781 + ], + "angle": 0, + "content": "3D rendering is a function that maps the 3D object to a 2D image. Differentiable rendering, as the name implies, provides a differentiable rendering function where the gradient of the function with respect to the 3D object parameters can be effectively computed. In this work, we use the differentiable rasterizer from Pytorch3D [46] to render parametric FLAME model \\( M(\\beta, \\theta, \\psi) \\) to 2D images \\( I_R \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.794, + 0.892, + 0.81 + ], + "angle": 0, + "content": "\\[\nR (M (\\boldsymbol {\\beta}, \\boldsymbol {\\theta}, \\psi), \\boldsymbol {\\alpha}, \\mathbf {l}, \\mathbf {c}) \\rightarrow I _ {R}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.823, + 0.892, + 0.9 + ], + "angle": 0, + "content": "with albedo parameters \\(\\alpha\\), Spherical Harmonics (SH) lighting parameters \\(\\mathbf{l} = \\left[\\mathbf{l}_1^T, \\dots, \\mathbf{l}_9^T\\right]^T\\), \\(\\mathbf{l}_k \\in \\mathbb{R}^3\\) to create shaded textures and camera parameters \\(\\mathbf{c} \\in \\mathbb{R}^3\\) which is the concatenation of isotropic scale \\(s \\in \\mathbb{R}\\) and 2D translation \\(t \\in \\mathbb{R}^2\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14194" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.09, + 0.46, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.471, + 0.365 + ], + "angle": 0, + "content": "Figure 2: Speech-Assisted Monocular 3D Face Reconstruction. The block in grey color are fixed models or networks, and the Expression Encoder \\( E_{e} \\) in blue color is the learnable network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.371, + 0.218, + 0.387 + ], + "angle": 0, + "content": "4. Speech4Mesh" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.396, + 0.47, + 0.426 + ], + "angle": 0, + "content": "4.1. Speech-Assisted Monocular 3D Face Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.47, + 0.72 + ], + "angle": 0, + "content": "In the proposed Speech4Mesh framework, 3D face sequences are first reconstructed from RGB videos. The detailed structure of our reconstruction module is illustrated in Fig. 2. The overall structure is built on DECA [22], the state-of-the-art method in monocular 3D face reconstruction. DECA learns animatable detailed 3D face models from static RGB images \\(I\\) in a two-stage scheme. The coarse stage first predicts parameters to reconstruct a rough face shape based on the FLAME model. The detail stage then generates a UV displacement map conditioned on expression and jaw pose parameters to achieve fine-grained details. Since the detail stage may bring noise to the reconstructed mesh, we only borrow the coarse part as our backbone. In DECA's coarse part, an encoder (ResNet-50) learns to regress 3DMM coefficients merely under visual constraints, including landmark-related losses, photometric loss, perceptual-level loss, and shape consistency loss. We will introduce some of the losses we have used later in the text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.722, + 0.468, + 0.814 + ], + "angle": 0, + "content": "Expression encoder: Since DECA already can capture the overall facial features well, we fix the pre-trained coarse encoder \\( E_{c} \\) in DECA for shape, pose (except jaw), camera, albedo, and light parameters. We then train an additional expression encoder \\( E_{e} \\) to regress expression parameters \\( \\psi_{e} \\) and jaw parameters \\( \\theta_{e}^{jaw} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.817, + 0.469, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {e} ^ {j a w}, \\psi_ {e} = E _ {e} (I). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.468, + 0.871 + ], + "angle": 0, + "content": "Note that, for computational efficiency, we use MobileNetV2 instead of ResNet-50 as the backbone network." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Only training \\( E_{e} \\) simplifies the learning goal and enables us to remove losses that are unrelated to lip motion such as" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "identity loss \\( L_{id} \\) and shape consistency loss \\( L_{sc} \\) in DECA, which leads to faster training speed and fewer training resources. Also, speech content would mainly affect mouth-related movements, which are most relevant to expression and jaw pose. The rendering output of the reconstruction part is then denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.192, + 0.892, + 0.211 + ], + "angle": 0, + "content": "\\[\nR \\left(M \\left(\\boldsymbol {\\beta}, [ \\boldsymbol {\\theta}, \\boldsymbol {\\theta} _ {e} ^ {j a w} ], \\psi_ {e}\\right), \\boldsymbol {\\alpha}, 1, \\mathbf {c}\\right)\\rightarrow I _ {R e}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.25 + ], + "angle": 0, + "content": "in which \\([\\theta, \\theta_{e}^{jaw}]\\) means we replace the original vectors representing jaw pose in \\(\\theta\\) with \\(\\theta_{e}^{jaw}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.893, + 0.477 + ], + "angle": 0, + "content": "Speech-visual contrastive loss: As discussed in previous sections, the majority of current works only utilize the visual information from utterance videos. While these methods have achieved acceptable reconstruction results, it is still an underconstrained problem to retrieve 3D faces merely from 2D planes. Ignoring information contained in audio can lead to a considerable loss in reconstruction. Thus, we propose to leverage speech constraints here to achieve more precise 3D face reconstruction in a contrastive manner. We calculate the contrastive loss using a minibatch of N speech-visual representation pairs \\((\\epsilon_s,\\epsilon_v)\\), where \\(\\epsilon_v\\) is the hidden state before the last 4-layer MLP in expression encoder, which projects \\(\\epsilon_v\\) to the target 3DMM space, and \\(\\epsilon_s\\) is the audio embedding. In specific, \\(\\epsilon_s\\) is obtained by the pre-trained wav2vec 2.0 [1] model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.487, + 0.891, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {s} = E _ {s} (S), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.892, + 0.68 + ], + "angle": 0, + "content": "which can extract powerful representations from speech audio. Since we interpolate the frequency of audio features to be two times the video frame rate, we then adopt a linear combination of audio features from time steps \\((t - \\lfloor T / 2\\rfloor)\\cdot 2\\) to \\((t + \\lfloor T / 2\\rfloor)\\cdot 2\\) for frame \\(t\\) as the speech-visual positive pair, where \\(T\\) is set to 5 in our experiment. Minimizing the contrastive loss leads to encoders that maximally preserve the mutual information between positive speech-visual pairs. The contrastive loss consists of two terms, in which the first term calculates the speech-to-visual contrastive cost for the i-th pair:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.689, + 0.891, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\ell_ {i} ^ {(s \\rightarrow v)} = - \\log \\frac {\\exp \\left(\\left\\langle \\epsilon_ {\\mathbf {i}} ^ {\\mathbf {s}}, \\epsilon_ {\\mathbf {i}} ^ {\\mathbf {v}} \\right\\rangle / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(\\left\\langle \\epsilon_ {\\mathbf {i}} ^ {\\mathbf {s}}, \\epsilon_ {\\mathbf {k}} ^ {\\mathbf {v}} \\right\\rangle / \\tau\\right)} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.826 + ], + "angle": 0, + "content": "where \\(\\langle \\epsilon_{\\mathbf{i}}^{\\mathbf{s}},\\epsilon_{\\mathbf{i}}^{\\mathbf{v}}\\rangle\\) calculates cosine similarity between two vectors, i.e., \\(\\langle \\mathbf{v},\\mathbf{u}\\rangle = \\mathbf{v}^\\top \\mathbf{u} / \\| \\mathbf{v}\\| \\| \\mathbf{u}\\|\\). \\(\\tau \\in \\mathbb{R}^{+}\\) is the temperature parameter, which is fixed to 0.07 in our experiment. While the second term is similar to the first term, it calculates cosine similarity as \\(\\langle \\epsilon_{\\mathbf{i}}^{\\mathbf{v}},\\epsilon_{\\mathbf{i}}^{\\mathbf{s}}\\rangle\\). The final contrastive loss is the weighted sum of the two terms:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.835, + 0.891, + 0.876 + ], + "angle": 0, + "content": "\\[\nL _ {c o n} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\lambda \\ell_ {i} ^ {(s \\rightarrow v)} + (1 - \\lambda) \\ell_ {i} ^ {(v \\rightarrow s)}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.887, + 0.902 + ], + "angle": 0, + "content": "where we set \\(\\lambda\\) to 0.5 in our experiment as a scalar weight." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14195" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "Weighted landmark loss: We use weighted landmark loss to minimize the \\( L_{1} \\) distance between ground-truth 2D key points and corresponding landmarks projected into the image plane from mesh." + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.157, + 0.47, + 0.198 + ], + "angle": 0, + "content": "\\[\nL _ {l m k} = \\sum_ {i = 1} ^ {6 8} \\lambda_ {i} \\| \\mathbf {k} _ {i} - s \\Pi (M _ {i}) + t \\| _ {1} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.203, + 0.469, + 0.277 + ], + "angle": 0, + "content": "in which \\(\\lambda_{i}\\) is the weight of the \\(i\\)th landmark, \\(\\Pi(M_{i})\\) represents projecting 3D meshes \\(M_{i}\\) to the 2D pixel space. We set weights of landmarks in the mouth, jaw, and nose area higher than others to capture more accurate geometry representations." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.279, + 0.436, + 0.293 + ], + "angle": 0, + "content": "Eye closure loss: We calculate the eye closure loss" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.299, + 0.469, + 0.333 + ], + "angle": 0, + "content": "\\[\nL _ {e y e} = \\sum_ {(i, j) \\in E} \\| \\mathbf {k} _ {i} - \\mathbf {k} _ {j} - s \\Pi (M _ {i} - M _ {j}) \\| _ {1} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.469, + 0.398 + ], + "angle": 0, + "content": "to force the relative offset of vertices \\( M_{i} \\) and \\( M_{j} \\) on the upper and lower eyelid to be close to the offset of ground-truth eyelid landmarks pairs \\( \\mathbf{k}_i \\) and \\( \\mathbf{k}_j \\). In this way, the reconstructed mesh could capture the blink of eye." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.469, + 0.444 + ], + "angle": 0, + "content": "Lip closure loss: The lip closure loss \\( L_{lp} \\) is similar to the eye closure loss. It measures the relative distance between the upper/lower lip and left/right lip corners keypoint pairs." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.444, + 0.367, + 0.458 + ], + "angle": 0, + "content": "Photometric loss: The photometric loss" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.464, + 0.469, + 0.482 + ], + "angle": 0, + "content": "\\[\nL _ {p h o} = \\left\\| V _ {I} \\odot (I - I _ {R e}) \\right\\| _ {1, 1} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.469, + 0.546 + ], + "angle": 0, + "content": "measures the pixel-to-pixel differences between ground-truth training images and rendered images. \\(V_{I}\\) is the face skin mask with a value 1 in the face skin area and \\(\\odot\\) denotes the Hadamard product." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.469, + 0.637 + ], + "angle": 0, + "content": "Emotion consistency loss: To capture more vivid expressions, we borrow the idea from EMOCA [14] that introduces an emotion consistency loss with a fixed emotion recognition network. We compute the \\( L_{2} \\) distance between emotion features of input images and output rendered images as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.643, + 0.469, + 0.661 + ], + "angle": 0, + "content": "\\[\nL _ {e m o} = \\left\\| \\epsilon_ {I} ^ {e m o}, \\epsilon_ {R e} ^ {e m o} \\right\\| _ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.469, + 0.695 + ], + "angle": 0, + "content": "where \\(\\epsilon_I^{emo}\\) and \\(\\epsilon_{Re}^{emo}\\) are produced by a fixed emotion recognition network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.469, + 0.788 + ], + "angle": 0, + "content": "Regularization term: We constrain the expression parameters \\(\\psi_{e}\\) and jaw parameters \\(\\theta_{e}^{jaw}\\) produced by our expression encoder to be close to the original parameters produced by pre-trained DECA by minimizing their \\(L_{2}\\) distance as performed in [24]: \\(\\| \\psi_{e} - \\psi \\|_{2}\\) and \\(\\left\\| \\theta_e^{jaw} - \\theta^{jaw}\\right\\| _2\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.787, + 0.442, + 0.802 + ], + "angle": 0, + "content": "Loss function: To sum up, our optimization goal is:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.807, + 0.469, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {r e c o n}} = \\lambda_ {c o n} L _ {c o n} + \\lambda_ {l m k} L _ {l m k} + \\lambda_ {e y e} L _ {e y e} + \\lambda_ {l p} L _ {l p} \\\\ \\lambda_ {p h o} L _ {p h o} + \\lambda_ {e m o} L _ {e m o} + \\lambda_ {r e g} L _ {r e g}, \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda\\) with different subscripts are hyper-parameters to balance the corresponding losses. The detailed values of \\(\\lambda\\)s can be found in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.107 + ], + "angle": 0, + "content": "4.2. Speech-Driven Emotional 3D Facial Animation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.266 + ], + "angle": 0, + "content": "After acquiring sufficient synthesized 4D training data, we can train the audio2mesh module for animation. We use FaceFormer [20], a Transformer-based model to drive 3D face vertices with speech embeddings. FaceFormer [20] is the first work to apply Transformer to 3D facial animation. It has achieved better performances in both realism and lip synchronization than prior methods owing to the Transformer's ability to learn long-term audio dependencies. Specifically, FaceFormer predicts the vertices' movement in an autoregressive manner as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.279, + 0.892, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} _ {\\mathbf {t}} = \\operatorname {F a c e F o r m e r} \\left(\\hat {\\mathbf {y}} _ {< \\mathbf {t}}, \\mathbf {i} _ {\\mathbf {n}}, \\mathbf {S}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.398 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{y}}_{\\mathrm{t}}\\) represents the predicted vertices offset at time step \\(t\\), and \\(\\mathbf{i}_{\\mathbf{n}}\\) represents the identity code. Although FaceFormer has achieved remarkable animation results, its performance highly depends on the data quality, since it is trained in a fully supervised scheme where the only supervision is the MSE loss between predicted and ground-truth meshes:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.41, + 0.891, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S E}} = \\sum_ {\\mathbf {t} = 1} ^ {\\mathbf {T}} \\sum_ {\\mathbf {v} = 1} ^ {\\mathbf {V}} \\left\\| \\hat {\\mathbf {y}} _ {\\mathbf {t}, \\mathbf {v}} - \\mathbf {y} _ {\\mathbf {t}, \\mathbf {v}} \\right\\| ^ {2}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.553 + ], + "angle": 0, + "content": "In such a way, it's hard to capture some facial details due to the registration loss of the scanned 4D training data. To this end, during pre-training with the synthesized data, we further introduce a pixel-level constraint by computing the photometric loss between the rendered images and corresponding real images as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.565, + 0.891, + 0.582 + ], + "angle": 0, + "content": "\\[\nR \\left(M _ {T} + \\hat {\\mathbf {y}} _ {\\mathbf {t}}, \\boldsymbol {\\theta} _ {\\mathbf {t}}, \\boldsymbol {\\alpha} _ {\\mathbf {t}}, \\mathbf {l} _ {\\mathbf {t}}, \\mathbf {c} _ {\\mathbf {t}}\\right)\\rightarrow I _ {R e} ^ {t}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.73 + ], + "angle": 0, + "content": "where \\( M_T \\) is the template face mesh (the mean of all faces in the training set as in FaceFormer). Note that, when performing differentiable rendering, since the output meshes of the FaceFormer network are all aligned, i.e., without any rotation or camera transform, we borrow \\( \\theta_t^1 \\), \\( \\alpha_t \\), \\( l_t \\), \\( c_t \\) estimated by our 3D reconstruction encoders to render the vertices to 2D images. Except for the pixel-to-pixel loss, we also calculate landmark-related losses as in the reconstruction part." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Furthermore, due to the insufficiency of emotional 4D data, current audio2mesh methods generally lack emotion controllability. To make our meshes more lifelike, we introduce a one-hot emotion code to explicitly control the emotional state. Specifically, during pre-training on the emotional 2D video dataset, we embed the emotion category into the FaceFormer network as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.849, + 0.891, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} _ {\\mathbf {t}} = \\operatorname {F a c e F o r m e r} \\left(\\hat {\\mathbf {y}} _ {< \\mathbf {t}}, \\mathbf {i} _ {\\mathbf {n}}, \\mathbf {e} _ {\\mathbf {m}}, \\mathbf {S}\\right), \\tag {16}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.9 + ], + "angle": 0, + "content": "1With slight abuse of notation, the \\( \\theta_t \\) here does not include the jaw rotation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14196" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.091, + 0.87, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Figure 3: Comparison of the reconstruction results on MEAD (left) and VoxCeleb2 (right) datasets. The first column on the left is the original RGB images. The reconstructed faces from the second to the sixth column are from 3DDFA-V2, Deep3DFace (Pytorch), DECA (coarse), EMOCA (coarse), and ours, respectively. More qualitative results could be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.471, + 0.457 + ], + "angle": 0, + "content": "where \\(\\mathbf{e}_{\\mathbf{m}}\\) is the one-hot vector projected to a higher dimension space. In this way, during the inference stage, we can achieve speaker-independent emotion-controllable 3D facial animation by embedding different emotion vectors into FaceFormer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.471, + 0.519 + ], + "angle": 0, + "content": "To sum up, by employing the self-supervised differentiable rendering scheme and emotion code, our modified FaceFormer model can generate not only more realistic but also emotion-controllable animated face meshes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.21, + 0.551 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.173, + 0.574 + ], + "angle": 0, + "content": "5.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.643 + ], + "angle": 0, + "content": "In this paper, we leveraged RGB video data and 4D mesh data for training the reconstruction module and fine-tuning the audio2mesh module, respectively. A brief introduction of the datasets we used is as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.901 + ], + "angle": 0, + "content": "2D Training Data: MEAD [64], VoxCeleb2 [10]. The MEAD dataset is a talking head corpus with different intensities of emotions. It has a total of 281,400 video clips collected from 60 English-speaking actors and actresses under 8 common emotions with 3 levels of intensity. Note that although the dataset contains videos from different views, we only took the front-face data for training to keep more visible details. VoxCeleb2 is a large-scale talking head video dataset collected from open-source media which covers a huge range of identities, races, languages, and head poses. In practice, we randomly sampled 2,000 identities for training. The images were extracted under 30 and 25 fps for MEAD and VoxCeleb2 videos, respectively. We used wav2vec 2.0 [1] to exact the audio embeddings, in which the output frequency is \\(49\\mathrm{Hz}\\). To ensure temporal alignment of the image and audio, we interpolate the audio embeddings to \\(50\\mathrm{Hz}\\) for the VoxCeleb2 dataset and to \\(60\\mathrm{Hz}\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.383, + 0.655, + 0.396 + ], + "angle": 0, + "content": "for the MEAD dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.398, + 0.892, + 0.489 + ], + "angle": 0, + "content": "4D Fine-Tuning Data: VOCASET [13]. VOCASET is a 4D dataset composed of consecutive scanned 3D heads which are all registered to a 5023 vertices mesh. It contains a total of 480 talking motion sequences from 12 subjects under the frame rate of 60 fps. The training, validation, and testing splits are directly adopted from [20]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.652, + 0.515 + ], + "angle": 0, + "content": "5.2. Training Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.523, + 0.892, + 0.629 + ], + "angle": 0, + "content": "The reconstruction part was trained for a maximum of 20 epochs, using the Adam optimizer and a learning rate of \\(5e - 5\\). In one iteration, we used 4 mini-batches of data, each of which contains 16 frames randomly sampled from a single video. We further implied distributed data-parallel technique to collect data from 4 GPUs to enlarge the pool of training pairs for contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.718 + ], + "angle": 0, + "content": "The audio2mesh network was trained using the Adam optimizer with a learning rate of \\( 1e - 4 \\) for a total of 100 epochs. The parameters of the encoder were initialized with the pre-trained wav2vec 2.0 weights and the parameters of TCN were fixed during training. The period \\( p \\) for the biased causal multi-head self-attention was set to 30." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.884 + ], + "angle": 0, + "content": "We pre-trained our audio2mesh network on 4D data produced by our monocular 3D reconstruction network from the MEAD dataset. While it contains 3 levels of emotional intensity, we only use the mildest first level to train. We created 8 one-hot emotion codes (angry, contempt, disgusted, fear, happy, sad, surprised, or neutral) and 42 one-hot speaker identity codes. For each pair of image and its 3D reconstruction in the MEAD dataset, we embedded its corresponding emotion code and speaker ID code to a higher dimension and then concatenated and projected it to the audio2mesh network." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We fine-tuned the audio2mesh network on VOCASET" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14197" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "which contains high-resolution 4D scans. During training, we randomly chose 8 kinds of speaker ID codes defined in the pre-train phase and set the emotion code to neutral." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.149, + 0.447, + 0.166 + ], + "angle": 0, + "content": "5.3. Analysis of 3D Face Reconstruction Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.173, + 0.473, + 0.55 + ], + "angle": 0, + "content": "Subjective Evaluation: Visualization of the reconstruction results is demonstrated in Fig. 3. We randomly sampled several frames from MEAD and VoxCeleb2 datasets for comparison. For clear observation of the expression details, we did not apply texture on the face. The sampled frames contain a variety of talking heads with various lip motions, extreme emotions, different genders, head poses, ages, and races. From Fig. 3, we can find that our method apparently outperforms the 3DDFA-V2 and DECA regarding the lip motion as well as the overall appearance. The curly lips and emotional expressions cannot be recovered by these two baselines. EMOCA is an expert in capturing extreme emotions. However, some mild expressions will be significantly exaggerated by EMOCA, which causes unnatural artifacts and inaccurate lip shapes in the reconstruction results. More comparison results with EMOCA could be found in supplementary materials. Note that, we used the Pytorch version of Deep3DFace for comparison, which is claimed to be better than the Tensorflow implementation. As shown in the third column in Fig. 3, the recovered faces by Deep3DFace reveal impressive identity consistency compared with the original RGB image. This may attribute to the identity aggregation mechanism and a different parametric face model (i.e., Basel Face Model [43]). However, the lower face expressiveness is still lower than our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.688 + ], + "angle": 0, + "content": "We further conducted a user study to make a perceptual evaluation between our and the competitors' reconstruction results. Specifically, 150 participants were asked to judge a side-by-side image composed of the original image, our reconstruction result, and the competitor's result. The participants can select their favorable one or choose the equal option if they cannot distinguish the difference. 50 samples were randomly selected from the MEAD and Voxceleb2 datasets. The result of the study is summarized in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.84 + ], + "angle": 0, + "content": "As we can see from Tab. 1, the result demonstrates an obvious inclination to support our results. It is worth mentioning that about 80 percent of our reconstructed samples are deemed better than or equal with 3DDFA-V2 and DECA methods regarding both full and lower face. EMOCA improves the perception of the results due to its emotion prior. Deep3DFace also achieves better performance as we discussed above. However, owing to the audio-visual alignment, our method achieves better performance at both full-face and the low-face region according to the user study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Objective Evaluation: Our reconstruction model's ability to capture lip shapes is measured using the lip vertex error as the objective metric, which has been extensively leveraged to evaluate the precision of lip synchronization" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.88, + 0.199 + ], + "angle": 0, + "content": "
Ours vs. CompetitorFace regionFavorability(%) ↑Ours better or equal
EqualComp.Ours
Ours vs. 3DDFA-V2Full5.419.475.280.6
Lower5.421.373.378.7
Ours vs. DECAFull11.418.869.881.2
Lower9.818.571.781.5
Ours vs. EMOCAFull8.438.353.358.9
Lower8.337.254.562.8
Ours vs. Deep3DFaceFull5.941.153.158.9
Lower8.734.257.065.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.203, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Table 1: A user study of the reconstruction results. We compare the participants' favorability of our reconstructions with the competitors' w.r.t. different regions of the face." + }, + { + "type": "table", + "bbox": [ + 0.572, + 0.253, + 0.819, + 0.287 + ], + "angle": 0, + "content": "
MetricOursEMOCA
Lip Vertex Error ↓0.995e-21.359e-2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.546, + 0.291, + 0.845, + 0.305 + ], + "angle": 0, + "content": "Table 2: Reconstruction error on VOCASET." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.458 + ], + "angle": 0, + "content": "[20, 47]. To calculate this error, we determine the maximal \\( L_{2} \\) distance between the lip vertices of the reconstructed meshes and their corresponding ground truth in the VO-CASET. We compared Speech4Mesh with EMOCA, which performs the best in the subjective evaluation. As presented in Table 2, our method outperforms EMOCA in terms of lower lip vertex error, demonstrating our superior ability to capture accurate lip shapes. On the contrary, EMOCA performs poorly on this metric due to its exaggerated facial expression with contorted mouth movements." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.469, + 0.892, + 0.499 + ], + "angle": 0, + "content": "5.4. Analysis of Speech-Driven Facial Animation Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.893, + 0.854 + ], + "angle": 0, + "content": "Subjective Evaluation: We also performed a user study on the speech-driven animation results. Similar to the reconstruction user study, we recruited 150 participants to judge the animation quality of our method and the vanilla FaceFormer from the realism and lip synchronization dimensions. The speeches for evaluation are from FaceFormer's test data (the demo video of FaceFormer) and the VoxCeleb2 dataset. For each dataset, we sampled 56 speeches (7 ids, 8 sentences for each id). Each participant evaluates 8 videos selected at random. The statistic of the user study is shown in Tab. 3. From Tab. 3 we can find that the favorability rating of our Speech4Mesh is consistently higher than that of FaceFormer, especially on the more challenging VoxCeleb2 speeches. We also visualize several representative syllables and corresponding animations in Fig. 4 for demonstration. In Fig. 4, our animated faces are more expressive and perceptually reasonable. We would like to note that MeshTalk cannot be well-trained on VOCASET which in a different topology, as mentioned in [61]. We have refrained from presenting any quantitative or qualitative comparison results here. However, we do incorporate a straightforward comparison in the supplementary videos. It is highly recommended to see those videos for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Objective Evaluation: We evaluate the lip vertex error of our proposed method and the vanilla FaceFormer on the test set of VOCASET. For simplicity, the scale we used for eval" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14198" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.47, + 0.186 + ], + "angle": 0, + "content": "
Test Speech DataCriterionFavorability(%) ↑Ours better or equal
EqualFaceFormerSpeech4Mesh
Speech from FaceFormer DemoRealism24.831.843.468.2
Lip Sync21.936.441.863.6
Speech from VoxCeleb2Realism16.326.157.773.9
Lip Sync16.730.752.569.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.189, + 0.471, + 0.237 + ], + "angle": 0, + "content": "Table 3: A user study of the speech-driven results. We compare the participants' favorability of our results with the FaceFormer with respect to realism and lip synchronization." + }, + { + "type": "table", + "bbox": [ + 0.15, + 0.243, + 0.395, + 0.313 + ], + "angle": 0, + "content": "
MethodLip Vertex Error ↓
VOCA6.428e-3
FaceFormer5.355e-3
Speech4Mesh4.863e-3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.317, + 0.469, + 0.348 + ], + "angle": 0, + "content": "Table 4: Comparison of the lip-sync error between VOCA, the vanilla FaceFormer, and Speech4Mesh." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.351, + 0.351, + 0.366 + ], + "angle": 0, + "content": "uation is the same as the FLAME model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.444 + ], + "angle": 0, + "content": "Tab. 4 displays the lip vertex errors of the vanilla Face-Former method and our Speech4Mesh method. It can be seen that more than a \\(9\\%\\) improvement in lip synchronization is brought by our training framework, which restates the importance of data sufficiency for 4D tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.451, + 0.245, + 0.467 + ], + "angle": 0, + "content": "5.5. Emotion Control" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.628 + ], + "angle": 0, + "content": "As introduced in previous sections, our audio2mesh module can be trained on the reconstructed meshes with different emotions. By employing an emotion code, we can manipulate the originally neural 4D talking heads to have different emotions. As shown in Fig. 5, different emotions (e.g., happy, surprised, sad) are embedded into the talking face driven by the same speech. The result reveals the potentiality of more diverse applications by exploiting the 2D videos under the Speech4Mesh framework. More emotion control results could be found in supplementary videos." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.242, + 0.65 + ], + "angle": 0, + "content": "5.6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To understand the impact of the contrastive loss and the photometric loss on speech-driven animation, we conducted quantitative experiments as shown in Tab. 5. More ablation studies could be found in supplementary materials. For the ablation study of contrastive loss, we pre-trained our audio2mesh module using 4D meshes reconstructed from the MEAD dataset, without incorporating contrastive loss during training. For the ablation study of photometric loss, we simply removed it in the pre-training stage. The quantitative results in Tab. 5. show that pre-training without contrastive loss or photometric loss results in a higher lip vertex error. These results underscore the importance of high-quality reconstructed 4D meshes in our method. They also support the notion that the photometric loss can encourage the audio2mesh network to learn detailed information which can not be captured solely from 4D meshes." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.089, + 0.856, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.227, + 0.894, + 0.272 + ], + "angle": 0, + "content": "Figure 4: Representative syllables and corresponding animations. The animation results of our Speech4Mesh and FaceFormer are displayed in the first and second columns." + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.28, + 0.807, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.506, + 0.443, + 0.884, + 0.459 + ], + "angle": 0, + "content": "Figure 5: Speech4Mesh samples with different emotions." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.468, + 0.879, + 0.5 + ], + "angle": 0, + "content": "
MetricFullwo. Pho.wo. Con.FaceFormer
Lip Vertex Error ↓4.863e-35.233e-35.271e-35.355e-3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.504, + 0.89, + 0.535 + ], + "angle": 0, + "content": "Table 5: Ablation study on the photometric loss and contrastive loss w.r.t. the lip-sync error." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.779, + 0.558 + ], + "angle": 0, + "content": "6. Conclusions and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.795 + ], + "angle": 0, + "content": "In this paper, we propose a novel framework named Speech4Mesh to train the audio2mesh network for speech-driven 3D facial animation. In Speech4Mesh, the speech-assisted monocular 3D face reconstruction module can provide significantly more accurate 4D synthesized data for pre-training by leveraging the audio-visual contrastive alignment. By employing the differentiable rendering loss and an emotion code, the modified audio2mesh network can achieve more realistic animation performance and unprecedentedly realize emotion manipulation. Empirical experiments verify the effectiveness and superiority of the Speech4Mesh framework both quantitatively and perceptually. We hope that the proposed Speech4Mesh framework can bring inspiration for similar 4D tasks to address the data scarcity problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In the future, we plan to replace the image-based reconstruction backbone DECA with a video-based reconstruction backbone such as MICA [76], neural head avatar [26] or [67]. We will also leverage more high-quality 2D video data for better animation performance and devise a more effective emotion modeling scheme to improve the expressiveness of expressions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14199" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems, 33:12449-12460, 2020. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5543-5552, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[4] Carlos Busso, Zhigang Deng, Michael Grimm, Ulrich Neumann, and Shrikanth Narayanan. Rigid head motion in expressive speech animation: Analysis and synthesis. IEEE transactions on audio, speech, and language processing, 15(3):1075-1086, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[5] Chen Cao, Qiming Hou, and Kun Zhou. Displaced dynamic expression regression for real-time facial tracking and animation. ACM Trans. Graph, 33(4):1-10, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[6] Yong Cao, Wen C Tien, Petros Faloutsos, and Frédéric Pighin. Expressive speech-driven facial animation. ACM Transactions on Graphics (TOG), 24(4):1283-1302, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[7] Yujin Chai, Yanlin Weng, Lvdi Wang, and Kun Zhou. Speech-driven facial animation with spectral gathering and temporal attention. Frontiers of Computer Science, 16(3):1-10, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.578 + ], + "angle": 0, + "content": "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[9] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. Proc. Interspeech 2018, pages 1086-1090, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[11] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian conference on computer vision, pages 251-263. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[12] Michael M Cohen, Rashid Clark, and Dominic W Massaro. Animated speech: Research progress and applications. In AVSP 2001-International Conference on Auditory-Visual Speech Processing, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[13] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Radek Daneček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In CVPR, pages 20311-20322, 2022. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[15] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European conference on computer vision, pages 408-424. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[16] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.893, + 0.316 + ], + "angle": 0, + "content": "[17] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.893, + 0.375 + ], + "angle": 0, + "content": "[18] Pengfei Dou, Shishir K Shah, and Ioannis A Kakadiaris. End-to-end 3d face reconstruction with deep neural networks. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 5908-5917, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.893, + 0.431 + ], + "angle": 0, + "content": "[19] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph, 35(4):1-11, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.893, + 0.503 + ], + "angle": 0, + "content": "[20] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.893, + 0.56 + ], + "angle": 0, + "content": "[21] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.893, + 0.616 + ], + "angle": 0, + "content": "[22] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Trans. Graph., 40(4):1-13, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.893, + 0.687 + ], + "angle": 0, + "content": "[23] Yao Feng, Fan Wu, Xiaohu Shao, Yanfeng Wang, and Xi Zhou. Joint 3d face reconstruction and dense alignment with position map regression network. In Proceedings of the European conference on computer vision, pages 534-551, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.893, + 0.759 + ], + "angle": 0, + "content": "[24] Panagiotis P Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Visual speech-aware perceptual 3d facial expression reconstruction from videos. arXiv preprint arXiv:2207.11094, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.893, + 0.831 + ], + "angle": 0, + "content": "[25] Kyle Genova, Forrester Cole, Aaron Maschinot, Aaron Sarna, Daniel Vlasic, and William T Freeman. Unsupervised training for 3d morphable model regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8377-8386, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[26] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14200" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Jianzhu Guo, Xiangyu Zhu, Yang Yang, Fan Yang, Zhen Lei, and Stan Z Li. Towards fast, accurate and stable 3d dense face alignment. In European Conference on Computer Vision, pages 152-168. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.216 + ], + "angle": 0, + "content": "[28] Yudong Guo, Jianfei Cai, Boyi Jiang, Jianmin Zheng, et al. CNN-based real-time dense face reconstruction with inverse-rendered photo-realistic face images. IEEE transactions on pattern analysis and machine intelligence, 41(6):1294-1307, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[29] Patrik Huber, Zhen-Hua Feng, William Christmas, Josef Kittler, and Matthias Ratsch. Fitting 3d morphable face models using local features. In 2015 IEEE international conference on image processing, pages 1195-1199. IEEE, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[30] Patrik Huber, Guosheng Hu, Rafael Tena, Pouria Mortazavian, P Koppen, William J Christmas, Matthias Ratsch, and Josef Kittler. A multiresolution 3d morphable face model and fitting framework. In Proceedings of the 11th international joint conference on computer vision, imaging and computer graphics theory and applications, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[31] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In Proceedings of the IEEE international conference on computer vision, pages 1031-1039, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[32] Amin Jourabloo and Xiaoming Liu. Large-pose face alignment via cnn-based dense 3d model fitting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4188–4196, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[33] Harim Jung, Myeong-Seok Oh, and Seong-Whan Lee. Learning free-form deformation for 3d face reconstruction from in-the-wild images. In 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 2737–2742. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[34] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Trans. Graph., 36(4):1-12, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[35] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[36] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[37] Feng Liu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Joint face alignment and 3d face reconstruction. In European Conference on Computer Vision, pages 545-560. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[38] Feng Liu, Ronghang Zhu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Disentangling features in 3d face shapes for joint face reconstruction and recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5216-5225, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[39] Jingying Liu, Binyuan Hui, Kun Li, Yunke Liu, Yu-Kun Lai, Yuxiang Zhang, Yebin Liu, and Jingyu Yang. Geometry-guided dense perspective network for speech-driven facial animation. IEEE Transactions on Visualization and Computer Graphics, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[40] Yaojie Liu, Amin Jourabloo, William Ren, and Xiaoming Liu. Dense face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1619-1628, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[41] Gaurav Mittal and Baoyuan Wang. Animating face using disentangled audio representations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3290–3298, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[43] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[44] Hai X Pham, Samuel Cheung, and Vladimir Pavlovic. Speech-driven 3d facial animation with implicit emotional awareness: a deep learning approach. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 80-88, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[45] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[46] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[47] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[48] Elad Richardson, Matan Sela, and Ron Kimmel. 3d face reconstruction by learning from synthetic data. In 2016 fourth international conference on 3D vision (3DV), pages 460-469. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[49] Elad Richardson, Matan Sela, Roy Or-El, and Ron Kimmel. Learning detailed face reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1259–1268, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[50] Sami Romdhani and Thomas Vetter. Estimating 3d shape and texture using pixel intensity, edges, specular highlights, texture constraints and a prior. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 2, pages 986-993. IEEE, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[51] Shunsuke Saito, Tianye Li, and Hao Li. Real-time facial segmentation and performance capture from rgb input. In European conference on computer vision, pages 244–261. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[52] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14201" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7763-7772, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[53] Matan Sela, Elad Richardson, and Ron Kimmel. Unrestricted facial geometry reconstruction using image-to-image translation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1576-1585, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.178, + 0.469, + 0.246 + ], + "angle": 0, + "content": "[54] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Mingmin Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In European Conference on Computer Vision, pages 53-70. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.247, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[55] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Trans. Graph., 36(4):1-13, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.302, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[56] Sarah Taylor, Taehwan Kim, Yisong Yue, Moshe Mahler, James Krahe, Anastasio Garcia Rodriguez, Jessica Hodgins, and Iain Matthews. A deep learning approach for generalized speech animation. ACM Transactions on Graphics (TOG), 36(4):1-11, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.373, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[57] Sarah L Taylor, Moshe Mahler, Barry-John Theobald, and Iain Matthews. Dynamic units of visual speech. In Proceedings of the 11th ACM SIGGRAPH/Eurographics conference on Computer Animation, pages 275-284, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.428, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[58] Ayush Tewari, Florian Bernard, Pablo Garrido, Gaurav Bharaj, Mohamed Elgharib, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Fml: Face model learning from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10812–10822, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.512, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[59] Ayush Tewari, Michael Zollhöfer, Pablo Garrido, Florian Bernard, Hyeongwoo Kim, Patrick Pérez, and Christian Theobalt. Self-supervised multi-level face model learning for monocular reconstruction at over \\(250\\mathrm{~hz}\\). In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2549-2559, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.595, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[60] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1274-1283, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[61] Balamurugan Thambiraja, Ikhsanul Habibie, Sadegh Aliakbarian, Darren Cosker, Christian Theobalt, and Justus Thies. Imitator: Personalized speech-driven 3d facial animation. arXiv preprint arXiv:2301.00023, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.735, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[62] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.805, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[63] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5163-5172, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[64] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700-717. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[65] Qianyun Wang, Zhenfeng Fan, and Shihong Xia. 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[66] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[67] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljkovic, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.332, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[68] Yuyu Xu, Andrew W Feng, Stacy Marsella, and Ari Shapiro. A practical and configurable lip sync method for games. In Proceedings of Motion on Games, pages 131-140, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.375, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[69] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.459, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[70] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.515, + 0.892, + 0.571 + ], + "angle": 0, + "content": "[71] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, pages 9299–9306, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.572, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[72] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.656, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[73] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.712, + 0.892, + 0.767 + ], + "angle": 0, + "content": "[74] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.769, + 0.892, + 0.836 + ], + "angle": 0, + "content": "[75] Xiangyu Zhu, Zhen Lei, Junjie Yan, Dong Yi, and Stan Z Li. High-fidelity pose and expression normalization for face recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 787-796, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.839, + 0.892, + 0.891 + ], + "angle": 0, + "content": "[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision, pages 250–269. Springer, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14202" + } + ] +] \ No newline at end of file diff --git a/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/236896c7-9d5b-41e0-8f54-b3702ca337d4_origin.pdf b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/236896c7-9d5b-41e0-8f54-b3702ca337d4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8d2979b0ebc26896dc5fc335839e36b2cd75d68a --- /dev/null +++ b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/236896c7-9d5b-41e0-8f54-b3702ca337d4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6e74088d0fe5d0848d92ec1eddde860fdf9ca5b203c5cb211226afae977c8dd +size 3902375 diff --git a/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/full.md b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f7eb65e9a66024fedf67371c52a637bcf52fc0fa --- /dev/null +++ b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/full.md @@ -0,0 +1,373 @@ +# Speech4Mesh: Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation + +Shan He $^{1,2}$ , Haonan He $^{2}$ , Shuo Yang $^{2}$ , Xiaoyan Wu $^{2}$ , Pengcheng Xia $^{2}$ , Bing Yin $^{2}$ , Cong Liu $^{2}$ , Lirong Dai $^{1}$ , Chang Xu $^{3}$ + +1University of Science and Technology of China, 2iFLYTEK Research, 3University of Sydney {shanhe2, hnhe, shuoyang7, xywu10, pcxia, bingyin, congliu2}@iflytek.com lrdai@ustc.edu.cn, c.xu@sydney.edu.au + +# Abstract + +Recent audio2mesh-based methods have shown promising prospects for speech-driven 3D facial animation tasks. However, some intractable challenges are urgent to be settled. For example, the data-scarcity problem is intrinsically inevitable due to the difficulty of 4D data collection. Besides, current methods generally lack controllability on the animated face. To this end, we propose a novel framework named Speech4Mesh to consecutively generate 4D talking head data and train the audio2mesh network with the reconstructed meshes. In our framework, we first reconstruct the 4D talking head sequence based on the monocular videos. For precise capture of the talking-related variation on the face, we exploit the audio-visual alignment information from the video by employing a contrastive learning scheme. We next can train the audio2mesh network (e.g., FaceFormer) based on the generated 4D data. To get control of the animated talking face, we encode the speaking-unrelated factors (e.g., emotion, etc.) into an emotion embedding for manipulation. Finally, a differentiable renderer guarantees more accurate photometric details of the reconstruction and animation results. Empirical experiments demonstrate that the Speech4Mesh framework can not only outperform state-of-the-art reconstruction methods, especially on the lower-face part but also achieve better animation performance both perceptually and objectively after pre-trained on the synthesized data. Besides, we also verify that the proposed framework is able to explicitly control the emotion of the animated talking face. + +# 1. Introduction + +Speech-driven facial animation is aimed at guiding the face model (either 2D image or 3D mesh) to have perceptually rational motion (especially for lip sync) only according to the input audio signal. It is drawing increasingly more + +![](images/db64bfb2d479e1ede49ea6bdb631142f53d4050c80a55e001b3bd360a71aa2b6.jpg) +Figure 1: Overview of the proposed Speech4Mesh framework. In this framework, the speech signal is leveraged for both the reconstruction module and the audio2mesh module. An emotion code is embedded into the audio2mesh module for emotional manipulation. Differentiable rendering is employed for recovering the details of the face. + +attention with the development of virtual reality, game and film production, etc. There are currently two mainstreams to deal with this task, either from 2D or 3D perspectives. By learning from video data [9, 41, 55, 71], 2D solutions try to generate the talking head images frame by frame. Although sufficient data is available to be used, the synthesized images cannot directly be applied to 3D scenes. Thus, in this paper, we mainly focus on 3D facial animation driven by speech. In recent years, a handful of 3D speech-driven facial animation methods [20, 47, 39, 13, 7, 34] have shown promising performance by directly learning a mapping from audio to 3D avatar mesh deformation with deep neural networks (hereinafter called audio2mesh methods). In this way, some subtle variations on the face mesh can be captured. Generally speaking, for the audio2mesh methods, data quality is one of the most important factors to be considered [20] since they generally demand high-precision + +4D talking head data to learn the correspondence from audio embedding to mesh movement in a supervised manner. However, in practice, real 4D data is extremely hard to collect. It is generally reconstructed and registered from synchronized multi-view images or 3D scan frame by frame, which is extraordinarily costly [13, 21]. Therefore, how to deal with the data scarcity problem becomes an urgent agenda to be discussed. + +In contrast, the accessibility of monocular videos is much higher. The scarcity of 4D data may be mitigated through the use of 3D reconstruction techniques applied to 2D videos. The majority of current monocular 3D face reconstruction methods (e.g., 3DDFA-V2 [27], DECA [22]) are based on the self-supervised training scheme, whose essential guidances are landmark loss and photometric loss. Nonetheless, these losses can only provide 2D information from the image plane, which is relatively weak to capture the complex nonrigid deformation of the lower face region [51, 62]. Therefore, the reconstructed results always demonstrate obvious artifacts and ambiguities between utterance and visual perception, such as mouth funnel, pucker, and rounded vowels, which may considerably limit the precision of the audio2mesh methods [20]. + +To address the problems mentioned above, we propose a novel framework named Speech4Mesh to consecutively synthesize 4D talking head data and train the audio2mesh network with the reconstructed mesh sequences. An overview of the Speech4Mesh framework is illustrated in Fig. 1. Firstly, to obtain better synthesis results, we propose a novel speech-assisted monocular 3D face reconstruction module to exploit multi-modal correlation from the video data rather than merely from the RGB image, since the mouth geometry and expression are naturally correlated to the speech. Specifically, this module is built based on a backbone reconstruction model DECA [22], one of the state-of-the-art monocular 3D face reconstruction methods. By employing a speech-visual contrastive loss, it can learn the correspondence between audio and visemes to make the reconstructed 3D face to be more reliable and perceptually reasonable. + +Secondly, once sufficient 4D data has been acquired, we can train an audio2mesh network to model the deformation of a face mesh based on input speech. For our work, we employed FaceFormer [20] as the backbone for the audio2mesh network. Compared with the vanilla training scheme of FaceFormer, the primary benefit of Speech4Mesh is data sufficiency. In fact, Speech4Mesh provides a low-cost self-supervised approach to generate "infinite" 4D training data from infinite RGB videos. Besides, owing to the diversity of the talking head videos, the reconstructed data empowers our model to further disentangle other auxiliary properties of the face (e.g., emotion, etc.) for manipulation. To be specific, we first pre-train Face- + +Former on the reconstructed emotional 4D dataset with the corresponding emotion codes. At the fine-tuning stage, different emotion codes can be embedded into the FaceFormer backbone to realize the emotion control of the animated talking head. + +The main contributions of our work can be summarized as follows: + +- We propose Speech4Mesh, a novel framework that tackles the data scarcity issue in speech-driven 3D facial animation by consecutively reconstructing pseudo-4D data and training an audio2mesh network. This framework also provides a referable scheme to alleviate the data scarcity problem of similar 4D tasks. +- This framework enables emotional control in the absence of high-quality expressive 4D data by pretraining with the emotional pseudo-4D talking head data reconstructed from 2D videos. And the synthesized emotional 4D talking head dataset can be found in https://github.com/haonanhe/MEAD-3D/tree/main. +- To the best of our knowledge, we are the first to utilize speech information as complement for monocular 3D face reconstruction. Empirical experiments show that the exploitation of multi-modal information yields more natural and precise results. + +# 2. Related Work + +Speech-driven 3D Facial Animation: Speech-driven facial animation has been widely explored over the years. Existing methods could be categorized into 2D-based approaches which attempt to operate in 2D pixel space [72, 70, 45, 15, 11, 8], and 3D-based approaches that we focus on. Some works animate 3D models based on visemes through complex rules of coarticulation [19, 73, 68, 57, 12]. Though they could capture accurate lip movements and achieve explicit control over the entire process, such methods usually involve plenty of manual adjustments. Data-driven approaches have been proposed to automatically learn a map from acoustic features to 3D models. Most of them were trained in a supervised manner [20, 13, 47, 6, 34] with multi-view motion capture data [4, 6] or high-resolution 4D scans [34, 13]. [6] searches the matching mouth animation for phonemes based on an Anime Graph structure. [34] proposes an end-to-end convolutional network and uses a latent code to control the emotional state of animated mesh. VOCA [13] achieves speaker-independent facial animation by leveraging a subject-specific latent code. MeshTalk [47] learns a categorical latent space to disentangle audio-correlated and audio-uncorrelated information, which enables it to animate the entire face. FaceFormer [20] first applies Transformer to 3D facial animation to make an autore + +pressive prediction of the animated 3D face mesh sequence, which claims to be able to capture accurate lip movements. + +However, 4D scans are usually too expensive to access, which leads to the data scarcity issue. There are methods trying to address this problem by using 4D data generated from videos [56, 65, 44]. [44] trains an LSTM-RNN model to regress the shape parameter of the parametric face model tracked from input video frames. [56] uses a sliding window predictor to learn mappings from phoneme sequences to lower facial shape and appearance that an Active Appearance Model (AAM) tracks from video frames. [65] aims to synthesize 3D talking-head with rich emotion. However, their 3D faces tracked from videos are only optimized by 2D landmarks loss, which may lead to inaccurate reconstruction and animation results. In contrast, our method leverages speech features to capture 3D faces with more accurate facial movements. In addition, pre-training on emotion datasets enables our reconstructed 3D face meshes to have emotion states, which leads to 3D facial animation with rich emotion. + +Monocular 3D face reconstruction: Monocular 3D face reconstruction aims to track and recover the 3D shape and texture of the human face from monocular 2D data. Early methods often perform an optimization procedure in an analysis-by-synthesis framework, predicting landmarks[75, 62, 30, 5] or local features[29, 50] to regress parameters of 3D Morphable Models (3DMM) [2, 43, 3, 35]. Although optimization-based approaches could recover high-quality 3D faces, they usually involve costly optimization processes. In recent years, learning-based methods were proposed for more efficient reconstruction. Some of them predict 3DMM coefficients [32, 74, 48, 37, 49] directly. To alleviate constraints caused by models' limited geometry, there are also model-free methods directly regressing 3D meshes [16, 18, 23, 33, 53], voxels [31] or Signed Distance Function (SDF) [42]. + +However, the greatest limitation of learning-based methods is the lack of paired training data. Some methods address this problem by generating synthetic data or using image-model pairs fitted by traditional optimization-based methods [63, 66, 28, 38, 74, 48, 53]. Others have tried to solve it in an unsupervised manner. Most of unsupervised methods try to regress 3DMM parameters by predicting 2D landmarks [17, 22, 40, 52, 54, 58, 59, 60, 69], photometric constraints [17, 22, 25, 54, 59, 60, 69] or multi-image constraints[17, 22, 25, 52, 58]. Perceptual constraints such as face recognition features [17, 22, 25, 54] or emotion features [14] are also utilized to constrain models' learning objects. While unsupervised methods provide a large amount of training data, reconstructing 3D faces from merely 2D information is still underconstrained. To capture more accurate expressions from monocular videos, our framework leverages speech information as another source + +of constraints, which has a strong relationship with facial movements, especially the mouth area. + +# 3. Preliminaries + +Current monocular 3D face reconstruction methods generally leverage the self-supervised scheme for training. In such a way, a parametric face model and a differentiable renderer are required to recover the 3D face mesh and render the 3D face back to the image for training, respectively. In this section, we review the basic knowledge of the parametric face model and differentiable renderer for reference. + +# 3.1. FLAME + +In this work, we used FLAME [36] as the prior parametric face model for reconstruction, which is a generic 3D head model using standard vertex-based linear blend skinning (LBS) with corrective blendshapes. Specifically, it consists of $N = 5023$ vertices and could be defined as: + +$$ +M (\boldsymbol {\beta}, \boldsymbol {\theta}, \boldsymbol {\psi}) \in \mathbb {R} ^ {3 N}, \tag {1} +$$ + +which contains shape parameters $\beta \in \mathbb{R}^{|\beta|}$ accounting for shape variation, expression $\psi \in \mathbb{R}^{|\psi|}$ to capture facial expressions, and pose parameters $\theta \in \mathbb{R}^{3K + 3}$ to rotate vertices around $K = 4$ joints (neck, jaw, and eyeballs) as well as correcting pose deformations. FLAME provides a geometric constraint and a low-dimensional latent space, making our reconstruction process much easier. Besides, highly disentangled parameters allow us to control the face model's expression or pose independently, which benefits us in both the reconstruction and animation phases. + +We obtain FLAME's texture map through an appearance model $A(\alpha) \in \mathbb{R}^{d \times d \times 3}$ with albedo parameters $\alpha \in \mathbb{R}^{|\alpha|}$ as input. This is achieved by converting Basel Face Model's linear albedo space to FLAME's UV layout. More details can be found in [35]. + +# 3.2. Differentiable Rendering + +3D rendering is a function that maps the 3D object to a 2D image. Differentiable rendering, as the name implies, provides a differentiable rendering function where the gradient of the function with respect to the 3D object parameters can be effectively computed. In this work, we use the differentiable rasterizer from Pytorch3D [46] to render parametric FLAME model $M(\beta, \theta, \psi)$ to 2D images $I_R$ as + +$$ +R (M (\boldsymbol {\beta}, \boldsymbol {\theta}, \psi), \boldsymbol {\alpha}, \mathbf {l}, \mathbf {c}) \rightarrow I _ {R}, \tag {2} +$$ + +with albedo parameters $\alpha$ , Spherical Harmonics (SH) lighting parameters $\mathbf{l} = \left[\mathbf{l}_1^T, \dots, \mathbf{l}_9^T\right]^T$ , $\mathbf{l}_k \in \mathbb{R}^3$ to create shaded textures and camera parameters $\mathbf{c} \in \mathbb{R}^3$ which is the concatenation of isotropic scale $s \in \mathbb{R}$ and 2D translation $t \in \mathbb{R}^2$ . + +![](images/fa7ddd1d4745c95c0e35e7d58047ae79b6071804af14b24a9967df591f9a8f43.jpg) +Figure 2: Speech-Assisted Monocular 3D Face Reconstruction. The block in grey color are fixed models or networks, and the Expression Encoder $E_{e}$ in blue color is the learnable network. + +# 4. Speech4Mesh + +# 4.1. Speech-Assisted Monocular 3D Face Reconstruction + +In the proposed Speech4Mesh framework, 3D face sequences are first reconstructed from RGB videos. The detailed structure of our reconstruction module is illustrated in Fig. 2. The overall structure is built on DECA [22], the state-of-the-art method in monocular 3D face reconstruction. DECA learns animatable detailed 3D face models from static RGB images $I$ in a two-stage scheme. The coarse stage first predicts parameters to reconstruct a rough face shape based on the FLAME model. The detail stage then generates a UV displacement map conditioned on expression and jaw pose parameters to achieve fine-grained details. Since the detail stage may bring noise to the reconstructed mesh, we only borrow the coarse part as our backbone. In DECA's coarse part, an encoder (ResNet-50) learns to regress 3DMM coefficients merely under visual constraints, including landmark-related losses, photometric loss, perceptual-level loss, and shape consistency loss. We will introduce some of the losses we have used later in the text. + +Expression encoder: Since DECA already can capture the overall facial features well, we fix the pre-trained coarse encoder $E_{c}$ in DECA for shape, pose (except jaw), camera, albedo, and light parameters. We then train an additional expression encoder $E_{e}$ to regress expression parameters $\psi_{e}$ and jaw parameters $\theta_{e}^{jaw}$ as follows: + +$$ +\boldsymbol {\theta} _ {e} ^ {j a w}, \psi_ {e} = E _ {e} (I). \tag {3} +$$ + +Note that, for computational efficiency, we use MobileNetV2 instead of ResNet-50 as the backbone network. + +Only training $E_{e}$ simplifies the learning goal and enables us to remove losses that are unrelated to lip motion such as + +identity loss $L_{id}$ and shape consistency loss $L_{sc}$ in DECA, which leads to faster training speed and fewer training resources. Also, speech content would mainly affect mouth-related movements, which are most relevant to expression and jaw pose. The rendering output of the reconstruction part is then denoted as: + +$$ +R \left(M \left(\boldsymbol {\beta}, [ \boldsymbol {\theta}, \boldsymbol {\theta} _ {e} ^ {j a w} ], \psi_ {e}\right), \boldsymbol {\alpha}, 1, \mathbf {c}\right)\rightarrow I _ {R e}, \tag {4} +$$ + +in which $[\theta, \theta_{e}^{jaw}]$ means we replace the original vectors representing jaw pose in $\theta$ with $\theta_{e}^{jaw}$ . + +Speech-visual contrastive loss: As discussed in previous sections, the majority of current works only utilize the visual information from utterance videos. While these methods have achieved acceptable reconstruction results, it is still an underconstrained problem to retrieve 3D faces merely from 2D planes. Ignoring information contained in audio can lead to a considerable loss in reconstruction. Thus, we propose to leverage speech constraints here to achieve more precise 3D face reconstruction in a contrastive manner. We calculate the contrastive loss using a minibatch of N speech-visual representation pairs $(\epsilon_s,\epsilon_v)$ , where $\epsilon_v$ is the hidden state before the last 4-layer MLP in expression encoder, which projects $\epsilon_v$ to the target 3DMM space, and $\epsilon_s$ is the audio embedding. In specific, $\epsilon_s$ is obtained by the pre-trained wav2vec 2.0 [1] model as follows: + +$$ +\epsilon_ {s} = E _ {s} (S), \tag {5} +$$ + +which can extract powerful representations from speech audio. Since we interpolate the frequency of audio features to be two times the video frame rate, we then adopt a linear combination of audio features from time steps $(t - \lfloor T / 2\rfloor)\cdot 2$ to $(t + \lfloor T / 2\rfloor)\cdot 2$ for frame $t$ as the speech-visual positive pair, where $T$ is set to 5 in our experiment. Minimizing the contrastive loss leads to encoders that maximally preserve the mutual information between positive speech-visual pairs. The contrastive loss consists of two terms, in which the first term calculates the speech-to-visual contrastive cost for the i-th pair: + +$$ +\ell_ {i} ^ {(s \rightarrow v)} = - \log \frac {\exp \left(\left\langle \epsilon_ {\mathbf {i}} ^ {\mathbf {s}}, \epsilon_ {\mathbf {i}} ^ {\mathbf {v}} \right\rangle / \tau\right)}{\sum_ {k = 1} ^ {N} \exp \left(\left\langle \epsilon_ {\mathbf {i}} ^ {\mathbf {s}}, \epsilon_ {\mathbf {k}} ^ {\mathbf {v}} \right\rangle / \tau\right)} \tag {6} +$$ + +where $\langle \epsilon_{\mathbf{i}}^{\mathbf{s}},\epsilon_{\mathbf{i}}^{\mathbf{v}}\rangle$ calculates cosine similarity between two vectors, i.e., $\langle \mathbf{v},\mathbf{u}\rangle = \mathbf{v}^\top \mathbf{u} / \| \mathbf{v}\| \| \mathbf{u}\|$ . $\tau \in \mathbb{R}^{+}$ is the temperature parameter, which is fixed to 0.07 in our experiment. While the second term is similar to the first term, it calculates cosine similarity as $\langle \epsilon_{\mathbf{i}}^{\mathbf{v}},\epsilon_{\mathbf{i}}^{\mathbf{s}}\rangle$ . The final contrastive loss is the weighted sum of the two terms: + +$$ +L _ {c o n} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\lambda \ell_ {i} ^ {(s \rightarrow v)} + (1 - \lambda) \ell_ {i} ^ {(v \rightarrow s)}\right), \tag {7} +$$ + +where we set $\lambda$ to 0.5 in our experiment as a scalar weight. + +Weighted landmark loss: We use weighted landmark loss to minimize the $L_{1}$ distance between ground-truth 2D key points and corresponding landmarks projected into the image plane from mesh. + +$$ +L _ {l m k} = \sum_ {i = 1} ^ {6 8} \lambda_ {i} \| \mathbf {k} _ {i} - s \Pi (M _ {i}) + t \| _ {1} \tag {8} +$$ + +in which $\lambda_{i}$ is the weight of the $i$ th landmark, $\Pi(M_{i})$ represents projecting 3D meshes $M_{i}$ to the 2D pixel space. We set weights of landmarks in the mouth, jaw, and nose area higher than others to capture more accurate geometry representations. + +Eye closure loss: We calculate the eye closure loss + +$$ +L _ {e y e} = \sum_ {(i, j) \in E} \| \mathbf {k} _ {i} - \mathbf {k} _ {j} - s \Pi (M _ {i} - M _ {j}) \| _ {1} \tag {9} +$$ + +to force the relative offset of vertices $M_{i}$ and $M_{j}$ on the upper and lower eyelid to be close to the offset of ground-truth eyelid landmarks pairs $\mathbf{k}_i$ and $\mathbf{k}_j$ . In this way, the reconstructed mesh could capture the blink of eye. + +Lip closure loss: The lip closure loss $L_{lp}$ is similar to the eye closure loss. It measures the relative distance between the upper/lower lip and left/right lip corners keypoint pairs. + +Photometric loss: The photometric loss + +$$ +L _ {p h o} = \left\| V _ {I} \odot (I - I _ {R e}) \right\| _ {1, 1} \tag {10} +$$ + +measures the pixel-to-pixel differences between ground-truth training images and rendered images. $V_{I}$ is the face skin mask with a value 1 in the face skin area and $\odot$ denotes the Hadamard product. + +Emotion consistency loss: To capture more vivid expressions, we borrow the idea from EMOCA [14] that introduces an emotion consistency loss with a fixed emotion recognition network. We compute the $L_{2}$ distance between emotion features of input images and output rendered images as follows: + +$$ +L _ {e m o} = \left\| \epsilon_ {I} ^ {e m o}, \epsilon_ {R e} ^ {e m o} \right\| _ {2}, \tag {11} +$$ + +where $\epsilon_I^{emo}$ and $\epsilon_{Re}^{emo}$ are produced by a fixed emotion recognition network. + +Regularization term: We constrain the expression parameters $\psi_{e}$ and jaw parameters $\theta_{e}^{jaw}$ produced by our expression encoder to be close to the original parameters produced by pre-trained DECA by minimizing their $L_{2}$ distance as performed in [24]: $\| \psi_{e} - \psi \|_{2}$ and $\left\| \theta_e^{jaw} - \theta^{jaw}\right\| _2$ + +Loss function: To sum up, our optimization goal is: + +$$ +\begin{array}{l} L _ {\text {r e c o n}} = \lambda_ {c o n} L _ {c o n} + \lambda_ {l m k} L _ {l m k} + \lambda_ {e y e} L _ {e y e} + \lambda_ {l p} L _ {l p} \\ \lambda_ {p h o} L _ {p h o} + \lambda_ {e m o} L _ {e m o} + \lambda_ {r e g} L _ {r e g}, \tag {12} \\ \end{array} +$$ + +where $\lambda$ with different subscripts are hyper-parameters to balance the corresponding losses. The detailed values of $\lambda$ s can be found in the supplemental material. + +# 4.2. Speech-Driven Emotional 3D Facial Animation + +After acquiring sufficient synthesized 4D training data, we can train the audio2mesh module for animation. We use FaceFormer [20], a Transformer-based model to drive 3D face vertices with speech embeddings. FaceFormer [20] is the first work to apply Transformer to 3D facial animation. It has achieved better performances in both realism and lip synchronization than prior methods owing to the Transformer's ability to learn long-term audio dependencies. Specifically, FaceFormer predicts the vertices' movement in an autoregressive manner as follows: + +$$ +\hat {\mathbf {y}} _ {\mathbf {t}} = \operatorname {F a c e F o r m e r} \left(\hat {\mathbf {y}} _ {< \mathbf {t}}, \mathbf {i} _ {\mathbf {n}}, \mathbf {S}\right), \tag {13} +$$ + +where $\hat{\mathbf{y}}_{\mathrm{t}}$ represents the predicted vertices offset at time step $t$ , and $\mathbf{i}_{\mathbf{n}}$ represents the identity code. Although FaceFormer has achieved remarkable animation results, its performance highly depends on the data quality, since it is trained in a fully supervised scheme where the only supervision is the MSE loss between predicted and ground-truth meshes: + +$$ +\mathcal {L} _ {\mathrm {M S E}} = \sum_ {\mathbf {t} = 1} ^ {\mathbf {T}} \sum_ {\mathbf {v} = 1} ^ {\mathbf {V}} \left\| \hat {\mathbf {y}} _ {\mathbf {t}, \mathbf {v}} - \mathbf {y} _ {\mathbf {t}, \mathbf {v}} \right\| ^ {2}. \tag {14} +$$ + +In such a way, it's hard to capture some facial details due to the registration loss of the scanned 4D training data. To this end, during pre-training with the synthesized data, we further introduce a pixel-level constraint by computing the photometric loss between the rendered images and corresponding real images as follow: + +$$ +R \left(M _ {T} + \hat {\mathbf {y}} _ {\mathbf {t}}, \boldsymbol {\theta} _ {\mathbf {t}}, \boldsymbol {\alpha} _ {\mathbf {t}}, \mathbf {l} _ {\mathbf {t}}, \mathbf {c} _ {\mathbf {t}}\right)\rightarrow I _ {R e} ^ {t}, \tag {15} +$$ + +where $M_T$ is the template face mesh (the mean of all faces in the training set as in FaceFormer). Note that, when performing differentiable rendering, since the output meshes of the FaceFormer network are all aligned, i.e., without any rotation or camera transform, we borrow $\theta_t^1$ , $\alpha_t$ , $l_t$ , $c_t$ estimated by our 3D reconstruction encoders to render the vertices to 2D images. Except for the pixel-to-pixel loss, we also calculate landmark-related losses as in the reconstruction part. + +Furthermore, due to the insufficiency of emotional 4D data, current audio2mesh methods generally lack emotion controllability. To make our meshes more lifelike, we introduce a one-hot emotion code to explicitly control the emotional state. Specifically, during pre-training on the emotional 2D video dataset, we embed the emotion category into the FaceFormer network as follows: + +$$ +\hat {\mathbf {y}} _ {\mathbf {t}} = \operatorname {F a c e F o r m e r} \left(\hat {\mathbf {y}} _ {< \mathbf {t}}, \mathbf {i} _ {\mathbf {n}}, \mathbf {e} _ {\mathbf {m}}, \mathbf {S}\right), \tag {16} +$$ + +![](images/45bd049caaf13a367ce7d68d980459b57a782213f60b66bcf43945c592b8857a.jpg) +Figure 3: Comparison of the reconstruction results on MEAD (left) and VoxCeleb2 (right) datasets. The first column on the left is the original RGB images. The reconstructed faces from the second to the sixth column are from 3DDFA-V2, Deep3DFace (Pytorch), DECA (coarse), EMOCA (coarse), and ours, respectively. More qualitative results could be found in the supplementary materials. + +where $\mathbf{e}_{\mathbf{m}}$ is the one-hot vector projected to a higher dimension space. In this way, during the inference stage, we can achieve speaker-independent emotion-controllable 3D facial animation by embedding different emotion vectors into FaceFormer. + +To sum up, by employing the self-supervised differentiable rendering scheme and emotion code, our modified FaceFormer model can generate not only more realistic but also emotion-controllable animated face meshes. + +# 5. Experiments + +# 5.1. Dataset + +In this paper, we leveraged RGB video data and 4D mesh data for training the reconstruction module and fine-tuning the audio2mesh module, respectively. A brief introduction of the datasets we used is as follows. + +2D Training Data: MEAD [64], VoxCeleb2 [10]. The MEAD dataset is a talking head corpus with different intensities of emotions. It has a total of 281,400 video clips collected from 60 English-speaking actors and actresses under 8 common emotions with 3 levels of intensity. Note that although the dataset contains videos from different views, we only took the front-face data for training to keep more visible details. VoxCeleb2 is a large-scale talking head video dataset collected from open-source media which covers a huge range of identities, races, languages, and head poses. In practice, we randomly sampled 2,000 identities for training. The images were extracted under 30 and 25 fps for MEAD and VoxCeleb2 videos, respectively. We used wav2vec 2.0 [1] to exact the audio embeddings, in which the output frequency is $49\mathrm{Hz}$ . To ensure temporal alignment of the image and audio, we interpolate the audio embeddings to $50\mathrm{Hz}$ for the VoxCeleb2 dataset and to $60\mathrm{Hz}$ + +for the MEAD dataset. + +4D Fine-Tuning Data: VOCASET [13]. VOCASET is a 4D dataset composed of consecutive scanned 3D heads which are all registered to a 5023 vertices mesh. It contains a total of 480 talking motion sequences from 12 subjects under the frame rate of 60 fps. The training, validation, and testing splits are directly adopted from [20]. + +# 5.2. Training Setup + +The reconstruction part was trained for a maximum of 20 epochs, using the Adam optimizer and a learning rate of $5e - 5$ . In one iteration, we used 4 mini-batches of data, each of which contains 16 frames randomly sampled from a single video. We further implied distributed data-parallel technique to collect data from 4 GPUs to enlarge the pool of training pairs for contrastive learning. + +The audio2mesh network was trained using the Adam optimizer with a learning rate of $1e - 4$ for a total of 100 epochs. The parameters of the encoder were initialized with the pre-trained wav2vec 2.0 weights and the parameters of TCN were fixed during training. The period $p$ for the biased causal multi-head self-attention was set to 30. + +We pre-trained our audio2mesh network on 4D data produced by our monocular 3D reconstruction network from the MEAD dataset. While it contains 3 levels of emotional intensity, we only use the mildest first level to train. We created 8 one-hot emotion codes (angry, contempt, disgusted, fear, happy, sad, surprised, or neutral) and 42 one-hot speaker identity codes. For each pair of image and its 3D reconstruction in the MEAD dataset, we embedded its corresponding emotion code and speaker ID code to a higher dimension and then concatenated and projected it to the audio2mesh network. + +We fine-tuned the audio2mesh network on VOCASET + +which contains high-resolution 4D scans. During training, we randomly chose 8 kinds of speaker ID codes defined in the pre-train phase and set the emotion code to neutral. + +# 5.3. Analysis of 3D Face Reconstruction Results + +Subjective Evaluation: Visualization of the reconstruction results is demonstrated in Fig. 3. We randomly sampled several frames from MEAD and VoxCeleb2 datasets for comparison. For clear observation of the expression details, we did not apply texture on the face. The sampled frames contain a variety of talking heads with various lip motions, extreme emotions, different genders, head poses, ages, and races. From Fig. 3, we can find that our method apparently outperforms the 3DDFA-V2 and DECA regarding the lip motion as well as the overall appearance. The curly lips and emotional expressions cannot be recovered by these two baselines. EMOCA is an expert in capturing extreme emotions. However, some mild expressions will be significantly exaggerated by EMOCA, which causes unnatural artifacts and inaccurate lip shapes in the reconstruction results. More comparison results with EMOCA could be found in supplementary materials. Note that, we used the Pytorch version of Deep3DFace for comparison, which is claimed to be better than the Tensorflow implementation. As shown in the third column in Fig. 3, the recovered faces by Deep3DFace reveal impressive identity consistency compared with the original RGB image. This may attribute to the identity aggregation mechanism and a different parametric face model (i.e., Basel Face Model [43]). However, the lower face expressiveness is still lower than our proposed method. + +We further conducted a user study to make a perceptual evaluation between our and the competitors' reconstruction results. Specifically, 150 participants were asked to judge a side-by-side image composed of the original image, our reconstruction result, and the competitor's result. The participants can select their favorable one or choose the equal option if they cannot distinguish the difference. 50 samples were randomly selected from the MEAD and Voxceleb2 datasets. The result of the study is summarized in Tab. 1. + +As we can see from Tab. 1, the result demonstrates an obvious inclination to support our results. It is worth mentioning that about 80 percent of our reconstructed samples are deemed better than or equal with 3DDFA-V2 and DECA methods regarding both full and lower face. EMOCA improves the perception of the results due to its emotion prior. Deep3DFace also achieves better performance as we discussed above. However, owing to the audio-visual alignment, our method achieves better performance at both full-face and the low-face region according to the user study. + +Objective Evaluation: Our reconstruction model's ability to capture lip shapes is measured using the lip vertex error as the objective metric, which has been extensively leveraged to evaluate the precision of lip synchronization + +
Ours vs. CompetitorFace regionFavorability(%) ↑Ours better or equal
EqualComp.Ours
Ours vs. 3DDFA-V2Full5.419.475.280.6
Lower5.421.373.378.7
Ours vs. DECAFull11.418.869.881.2
Lower9.818.571.781.5
Ours vs. EMOCAFull8.438.353.358.9
Lower8.337.254.562.8
Ours vs. Deep3DFaceFull5.941.153.158.9
Lower8.734.257.065.8
+ +Table 1: A user study of the reconstruction results. We compare the participants' favorability of our reconstructions with the competitors' w.r.t. different regions of the face. + +
MetricOursEMOCA
Lip Vertex Error ↓0.995e-21.359e-2
+ +Table 2: Reconstruction error on VOCASET. + +[20, 47]. To calculate this error, we determine the maximal $L_{2}$ distance between the lip vertices of the reconstructed meshes and their corresponding ground truth in the VO-CASET. We compared Speech4Mesh with EMOCA, which performs the best in the subjective evaluation. As presented in Table 2, our method outperforms EMOCA in terms of lower lip vertex error, demonstrating our superior ability to capture accurate lip shapes. On the contrary, EMOCA performs poorly on this metric due to its exaggerated facial expression with contorted mouth movements. + +# 5.4. Analysis of Speech-Driven Facial Animation Results + +Subjective Evaluation: We also performed a user study on the speech-driven animation results. Similar to the reconstruction user study, we recruited 150 participants to judge the animation quality of our method and the vanilla FaceFormer from the realism and lip synchronization dimensions. The speeches for evaluation are from FaceFormer's test data (the demo video of FaceFormer) and the VoxCeleb2 dataset. For each dataset, we sampled 56 speeches (7 ids, 8 sentences for each id). Each participant evaluates 8 videos selected at random. The statistic of the user study is shown in Tab. 3. From Tab. 3 we can find that the favorability rating of our Speech4Mesh is consistently higher than that of FaceFormer, especially on the more challenging VoxCeleb2 speeches. We also visualize several representative syllables and corresponding animations in Fig. 4 for demonstration. In Fig. 4, our animated faces are more expressive and perceptually reasonable. We would like to note that MeshTalk cannot be well-trained on VOCASET which in a different topology, as mentioned in [61]. We have refrained from presenting any quantitative or qualitative comparison results here. However, we do incorporate a straightforward comparison in the supplementary videos. It is highly recommended to see those videos for more details. + +Objective Evaluation: We evaluate the lip vertex error of our proposed method and the vanilla FaceFormer on the test set of VOCASET. For simplicity, the scale we used for eval + +
Test Speech DataCriterionFavorability(%) ↑Ours better or equal
EqualFaceFormerSpeech4Mesh
Speech from FaceFormer DemoRealism24.831.843.468.2
Lip Sync21.936.441.863.6
Speech from VoxCeleb2Realism16.326.157.773.9
Lip Sync16.730.752.569.3
+ +Table 3: A user study of the speech-driven results. We compare the participants' favorability of our results with the FaceFormer with respect to realism and lip synchronization. + +
MethodLip Vertex Error ↓
VOCA6.428e-3
FaceFormer5.355e-3
Speech4Mesh4.863e-3
+ +Table 4: Comparison of the lip-sync error between VOCA, the vanilla FaceFormer, and Speech4Mesh. + +uation is the same as the FLAME model. + +Tab. 4 displays the lip vertex errors of the vanilla Face-Former method and our Speech4Mesh method. It can be seen that more than a $9\%$ improvement in lip synchronization is brought by our training framework, which restates the importance of data sufficiency for 4D tasks. + +# 5.5. Emotion Control + +As introduced in previous sections, our audio2mesh module can be trained on the reconstructed meshes with different emotions. By employing an emotion code, we can manipulate the originally neural 4D talking heads to have different emotions. As shown in Fig. 5, different emotions (e.g., happy, surprised, sad) are embedded into the talking face driven by the same speech. The result reveals the potentiality of more diverse applications by exploiting the 2D videos under the Speech4Mesh framework. More emotion control results could be found in supplementary videos. + +# 5.6. Ablation Studies + +To understand the impact of the contrastive loss and the photometric loss on speech-driven animation, we conducted quantitative experiments as shown in Tab. 5. More ablation studies could be found in supplementary materials. For the ablation study of contrastive loss, we pre-trained our audio2mesh module using 4D meshes reconstructed from the MEAD dataset, without incorporating contrastive loss during training. For the ablation study of photometric loss, we simply removed it in the pre-training stage. The quantitative results in Tab. 5. show that pre-training without contrastive loss or photometric loss results in a higher lip vertex error. These results underscore the importance of high-quality reconstructed 4D meshes in our method. They also support the notion that the photometric loss can encourage the audio2mesh network to learn detailed information which can not be captured solely from 4D meshes. + +![](images/19755bcc633b5ef5f2330fec81ab25631352b285c8cb374bf567f94378903fd7.jpg) +Figure 4: Representative syllables and corresponding animations. The animation results of our Speech4Mesh and FaceFormer are displayed in the first and second columns. + +![](images/8753daa8ceb5377c021443de597afbfdc507cd40bf4bda7e7106ba3fcd768100.jpg) + +Figure 5: Speech4Mesh samples with different emotions. + +
MetricFullwo. Pho.wo. Con.FaceFormer
Lip Vertex Error ↓4.863e-35.233e-35.271e-35.355e-3
+ +Table 5: Ablation study on the photometric loss and contrastive loss w.r.t. the lip-sync error. + +# 6. Conclusions and Future Work + +In this paper, we propose a novel framework named Speech4Mesh to train the audio2mesh network for speech-driven 3D facial animation. In Speech4Mesh, the speech-assisted monocular 3D face reconstruction module can provide significantly more accurate 4D synthesized data for pre-training by leveraging the audio-visual contrastive alignment. By employing the differentiable rendering loss and an emotion code, the modified audio2mesh network can achieve more realistic animation performance and unprecedentedly realize emotion manipulation. Empirical experiments verify the effectiveness and superiority of the Speech4Mesh framework both quantitatively and perceptually. We hope that the proposed Speech4Mesh framework can bring inspiration for similar 4D tasks to address the data scarcity problem. + +In the future, we plan to replace the image-based reconstruction backbone DECA with a video-based reconstruction backbone such as MICA [76], neural head avatar [26] or [67]. We will also leverage more high-quality 2D video data for better animation performance and devise a more effective emotion modeling scheme to improve the expressiveness of expressions. + +# References + +[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems, 33:12449-12460, 2020. 4, 6 +[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3 +[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5543-5552, 2016. 3 +[4] Carlos Busso, Zhigang Deng, Michael Grimm, Ulrich Neumann, and Shrikanth Narayanan. Rigid head motion in expressive speech animation: Analysis and synthesis. IEEE transactions on audio, speech, and language processing, 15(3):1075-1086, 2007. 2 +[5] Chen Cao, Qiming Hou, and Kun Zhou. Displaced dynamic expression regression for real-time facial tracking and animation. ACM Trans. Graph, 33(4):1-10, 2014. 3 +[6] Yong Cao, Wen C Tien, Petros Faloutsos, and Frédéric Pighin. Expressive speech-driven facial animation. ACM Transactions on Graphics (TOG), 24(4):1283-1302, 2005. 2 +[7] Yujin Chai, Yanlin Weng, Lvdi Wang, and Kun Zhou. Speech-driven facial animation with spectral gathering and temporal attention. Frontiers of Computer Science, 16(3):1-10, 2022. 1 +[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 2 +[9] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 1 +[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. Proc. Interspeech 2018, pages 1086-1090, 2018. 6 +[11] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian conference on computer vision, pages 251-263. Springer, 2016. 2 +[12] Michael M Cohen, Rashid Clark, and Dominic W Massaro. Animated speech: Research progress and applications. In AVSP 2001-International Conference on Auditory-Visual Speech Processing, 2001. 2 +[13] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 1, 2, 6 +[14] Radek Daneček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In CVPR, pages 20311-20322, 2022. 3, 5 + +[15] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European conference on computer vision, pages 408-424. Springer, 2020. 2 +[16] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 3 +[17] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 3 +[18] Pengfei Dou, Shishir K Shah, and Ioannis A Kakadiaris. End-to-end 3d face reconstruction with deep neural networks. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 5908-5917, 2017. 3 +[19] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph, 35(4):1-11, 2016. 2 +[20] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 1, 2, 5, 6, 7 +[21] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2 +[22] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Trans. Graph., 40(4):1-13, 2021. 2, 3, 4 +[23] Yao Feng, Fan Wu, Xiaohu Shao, Yanfeng Wang, and Xi Zhou. Joint 3d face reconstruction and dense alignment with position map regression network. In Proceedings of the European conference on computer vision, pages 534-551, 2018. 3 +[24] Panagiotis P Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Visual speech-aware perceptual 3d facial expression reconstruction from videos. arXiv preprint arXiv:2207.11094, 2022.5 +[25] Kyle Genova, Forrester Cole, Aaron Maschinot, Aaron Sarna, Daniel Vlasic, and William T Freeman. Unsupervised training for 3d morphable model regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8377-8386, 2018. 3 +[26] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 8 + +[27] Jianzhu Guo, Xiangyu Zhu, Yang Yang, Fan Yang, Zhen Lei, and Stan Z Li. Towards fast, accurate and stable 3d dense face alignment. In European Conference on Computer Vision, pages 152-168. Springer, 2020. 2 +[28] Yudong Guo, Jianfei Cai, Boyi Jiang, Jianmin Zheng, et al. CNN-based real-time dense face reconstruction with inverse-rendered photo-realistic face images. IEEE transactions on pattern analysis and machine intelligence, 41(6):1294-1307, 2018. 3 +[29] Patrik Huber, Zhen-Hua Feng, William Christmas, Josef Kittler, and Matthias Ratsch. Fitting 3d morphable face models using local features. In 2015 IEEE international conference on image processing, pages 1195-1199. IEEE, 2015. 3 +[30] Patrik Huber, Guosheng Hu, Rafael Tena, Pouria Mortazavian, P Koppen, William J Christmas, Matthias Ratsch, and Josef Kittler. A multiresolution 3d morphable face model and fitting framework. In Proceedings of the 11th international joint conference on computer vision, imaging and computer graphics theory and applications, 2016. 3 +[31] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In Proceedings of the IEEE international conference on computer vision, pages 1031-1039, 2017. 3 +[32] Amin Jourabloo and Xiaoming Liu. Large-pose face alignment via cnn-based dense 3d model fitting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4188–4196, 2016. 3 +[33] Harim Jung, Myeong-Seok Oh, and Seong-Whan Lee. Learning free-form deformation for 3d face reconstruction from in-the-wild images. In 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 2737–2742. IEEE, 2021. 3 +[34] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Trans. Graph., 36(4):1-12, 2017. 1, 2 +[35] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3 +[36] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3 +[37] Feng Liu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Joint face alignment and 3d face reconstruction. In European Conference on Computer Vision, pages 545-560. Springer, 2016. 3 +[38] Feng Liu, Ronghang Zhu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Disentangling features in 3d face shapes for joint face reconstruction and recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5216-5225, 2018. 3 +[39] Jingying Liu, Binyuan Hui, Kun Li, Yunke Liu, Yu-Kun Lai, Yuxiang Zhang, Yebin Liu, and Jingyu Yang. Geometry-guided dense perspective network for speech-driven facial animation. IEEE Transactions on Visualization and Computer Graphics, 2021. 1 + +[40] Yaojie Liu, Amin Jourabloo, William Ren, and Xiaoming Liu. Dense face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1619-1628, 2017. 3 +[41] Gaurav Mittal and Baoyuan Wang. Animating face using disentangled audio representations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3290–3298, 2020. 1 +[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3 +[43] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 3, 7 +[44] Hai X Pham, Samuel Cheung, and Vladimir Pavlovic. Speech-driven 3d facial animation with implicit emotional awareness: a deep learning approach. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 80-88, 2017. 3 +[45] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2 +[46] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 3 +[47] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 1, 2, 7 +[48] Elad Richardson, Matan Sela, and Ron Kimmel. 3d face reconstruction by learning from synthetic data. In 2016 fourth international conference on 3D vision (3DV), pages 460-469. IEEE, 2016. 3 +[49] Elad Richardson, Matan Sela, Roy Or-El, and Ron Kimmel. Learning detailed face reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1259–1268, 2017. 3 +[50] Sami Romdhani and Thomas Vetter. Estimating 3d shape and texture using pixel intensity, edges, specular highlights, texture constraints and a prior. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 2, pages 986-993. IEEE, 2005. 3 +[51] Shunsuke Saito, Tianye Li, and Hao Li. Real-time facial segmentation and performance capture from rgb input. In European conference on computer vision, pages 244–261. Springer, 2016. 2 +[52] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7763-7772, 2019. 3 +[53] Matan Sela, Elad Richardson, and Ron Kimmel. Unrestricted facial geometry reconstruction using image-to-image translation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1576-1585, 2017. 3 +[54] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Mingmin Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In European Conference on Computer Vision, pages 53-70. Springer, 2020. 3 +[55] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Trans. Graph., 36(4):1-13, 2017. 1 +[56] Sarah Taylor, Taehwan Kim, Yisong Yue, Moshe Mahler, James Krahe, Anastasio Garcia Rodriguez, Jessica Hodgins, and Iain Matthews. A deep learning approach for generalized speech animation. ACM Transactions on Graphics (TOG), 36(4):1-11, 2017. 3 +[57] Sarah L Taylor, Moshe Mahler, Barry-John Theobald, and Iain Matthews. Dynamic units of visual speech. In Proceedings of the 11th ACM SIGGRAPH/Eurographics conference on Computer Animation, pages 275-284, 2012. 2 +[58] Ayush Tewari, Florian Bernard, Pablo Garrido, Gaurav Bharaj, Mohamed Elgharib, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Fml: Face model learning from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10812–10822, 2019. 3 +[59] Ayush Tewari, Michael Zollhöfer, Pablo Garrido, Florian Bernard, Hyeongwoo Kim, Patrick Pérez, and Christian Theobalt. Self-supervised multi-level face model learning for monocular reconstruction at over $250\mathrm{~hz}$ . In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2549-2559, 2018. 3 +[60] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1274-1283, 2017. 3 +[61] Balamurugan Thambiraja, Ikhsanul Habibie, Sadegh Aliakbarian, Darren Cosker, Christian Theobalt, and Justus Thies. Imitator: Personalized speech-driven 3d facial animation. arXiv preprint arXiv:2301.00023, 2022. 7 +[62] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3 +[63] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5163-5172, 2017. 3 +[64] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change + +Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700-717. Springer, 2020. 6 +[65] Qianyun Wang, Zhenfeng Fan, and Shihong Xia. 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051, 2021. 3 +[66] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 3 +[67] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljkovic, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 8 +[68] Yuyu Xu, Andrew W Feng, Stacy Marsella, and Ari Shapiro. A practical and configurable lip sync method for games. In Proceedings of Motion on Games, pages 131-140, 2013. 2 +[69] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 3 +[70] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2 +[71] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, pages 9299–9306, 2019. 1 +[72] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2 +[73] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2 +[74] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 3 +[75] Xiangyu Zhu, Zhen Lei, Junjie Yan, Dong Yi, and Stan Z Li. High-fidelity pose and expression normalization for face recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 787-796, 2015. 3 +[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision, pages 250–269. Springer, 2022. 8 \ No newline at end of file diff --git a/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/images.zip b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..05a8f8d9ac17e0a6472de04e714c65c548b5eccf --- /dev/null +++ b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1af34ac0c805a984a596a2d184ac60f7a4257a26f52e5d06827e9f706f8c794d +size 357437 diff --git a/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/layout.json b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6db8df892acc6f8cebd713c9c8a858b2a00c2b5f --- /dev/null +++ b/2023/Speech4Mesh_ Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation/layout.json @@ -0,0 +1,9548 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 520, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 520, + 138 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 520, + 138 + ], + "type": "text", + "content": "Speech4Mesh: Speech-Assisted Monocular 3D Facial Reconstruction for Speech-Driven 3D Facial Animation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "spans": [ + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "text", + "content": "Shan He" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "text", + "content": ", Haonan He" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "text", + "content": ", Shuo Yang" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "text", + "content": ", Xiaoyan Wu" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "text", + "content": ", Pengcheng Xia" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "text", + "content": ", Bing Yin" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "text", + "content": ", Cong Liu" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "text", + "content": ", Lirong Dai" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "text", + "content": ", Chang Xu" + }, + { + "bbox": [ + 126, + 160, + 467, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 189, + 523, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 189, + 523, + 232 + ], + "spans": [ + { + "bbox": [ + 70, + 189, + 523, + 232 + ], + "type": "text", + "content": "1University of Science and Technology of China, 2iFLYTEK Research, 3University of Sydney {shanhe2, hnhe, shuoyang7, xywu10, pcxia, bingyin, congliu2}@iflytek.com lrdai@ustc.edu.cn, c.xu@sydney.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 285, + 290, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 290, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 290, + 621 + ], + "type": "text", + "content": "Recent audio2mesh-based methods have shown promising prospects for speech-driven 3D facial animation tasks. However, some intractable challenges are urgent to be settled. For example, the data-scarcity problem is intrinsically inevitable due to the difficulty of 4D data collection. Besides, current methods generally lack controllability on the animated face. To this end, we propose a novel framework named Speech4Mesh to consecutively generate 4D talking head data and train the audio2mesh network with the reconstructed meshes. In our framework, we first reconstruct the 4D talking head sequence based on the monocular videos. For precise capture of the talking-related variation on the face, we exploit the audio-visual alignment information from the video by employing a contrastive learning scheme. We next can train the audio2mesh network (e.g., FaceFormer) based on the generated 4D data. To get control of the animated talking face, we encode the speaking-unrelated factors (e.g., emotion, etc.) into an emotion embedding for manipulation. Finally, a differentiable renderer guarantees more accurate photometric details of the reconstruction and animation results. Empirical experiments demonstrate that the Speech4Mesh framework can not only outperform state-of-the-art reconstruction methods, especially on the lower-face part but also achieve better animation performance both perceptually and objectively after pre-trained on the synthesized data. Besides, we also verify that the proposed framework is able to explicitly control the emotion of the animated talking face." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 644, + 128, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 128, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 128, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Speech-driven facial animation is aimed at guiding the face model (either 2D image or 3D mesh) to have perceptually rational motion (especially for lip sync) only according to the input audio signal. It is drawing increasingly more" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 258, + 539, + 423 + ], + "blocks": [ + { + "bbox": [ + 313, + 258, + 539, + 423 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 539, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 539, + 423 + ], + "type": "image", + "image_path": "db64bfb2d479e1ede49ea6bdb631142f53d4050c80a55e001b3bd360a71aa2b6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 426, + 547, + 499 + ], + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 499 + ], + "type": "text", + "content": "Figure 1: Overview of the proposed Speech4Mesh framework. In this framework, the speech signal is leveraged for both the reconstruction module and the audio2mesh module. An emotion code is embedded into the audio2mesh module for emotional manipulation. Differentiable rendering is employed for recovering the details of the face." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 504, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 547, + 708 + ], + "type": "text", + "content": "attention with the development of virtual reality, game and film production, etc. There are currently two mainstreams to deal with this task, either from 2D or 3D perspectives. By learning from video data [9, 41, 55, 71], 2D solutions try to generate the talking head images frame by frame. Although sufficient data is available to be used, the synthesized images cannot directly be applied to 3D scenes. Thus, in this paper, we mainly focus on 3D facial animation driven by speech. In recent years, a handful of 3D speech-driven facial animation methods [20, 47, 39, 13, 7, 34] have shown promising performance by directly learning a mapping from audio to 3D avatar mesh deformation with deep neural networks (hereinafter called audio2mesh methods). In this way, some subtle variations on the face mesh can be captured. Generally speaking, for the audio2mesh methods, data quality is one of the most important factors to be considered [20] since they generally demand high-precision" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14192" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "4D talking head data to learn the correspondence from audio embedding to mesh movement in a supervised manner. However, in practice, real 4D data is extremely hard to collect. It is generally reconstructed and registered from synchronized multi-view images or 3D scan frame by frame, which is extraordinarily costly [13, 21]. Therefore, how to deal with the data scarcity problem becomes an urgent agenda to be discussed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 350 + ], + "type": "text", + "content": "In contrast, the accessibility of monocular videos is much higher. The scarcity of 4D data may be mitigated through the use of 3D reconstruction techniques applied to 2D videos. The majority of current monocular 3D face reconstruction methods (e.g., 3DDFA-V2 [27], DECA [22]) are based on the self-supervised training scheme, whose essential guidances are landmark loss and photometric loss. Nonetheless, these losses can only provide 2D information from the image plane, which is relatively weak to capture the complex nonrigid deformation of the lower face region [51, 62]. Therefore, the reconstructed results always demonstrate obvious artifacts and ambiguities between utterance and visual perception, such as mouth funnel, pucker, and rounded vowels, which may considerably limit the precision of the audio2mesh methods [20]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 555 + ], + "type": "text", + "content": "To address the problems mentioned above, we propose a novel framework named Speech4Mesh to consecutively synthesize 4D talking head data and train the audio2mesh network with the reconstructed mesh sequences. An overview of the Speech4Mesh framework is illustrated in Fig. 1. Firstly, to obtain better synthesis results, we propose a novel speech-assisted monocular 3D face reconstruction module to exploit multi-modal correlation from the video data rather than merely from the RGB image, since the mouth geometry and expression are naturally correlated to the speech. Specifically, this module is built based on a backbone reconstruction model DECA [22], one of the state-of-the-art monocular 3D face reconstruction methods. By employing a speech-visual contrastive loss, it can learn the correspondence between audio and visemes to make the reconstructed 3D face to be more reliable and perceptually reasonable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "Secondly, once sufficient 4D data has been acquired, we can train an audio2mesh network to model the deformation of a face mesh based on input speech. For our work, we employed FaceFormer [20] as the backbone for the audio2mesh network. Compared with the vanilla training scheme of FaceFormer, the primary benefit of Speech4Mesh is data sufficiency. In fact, Speech4Mesh provides a low-cost self-supervised approach to generate \"infinite\" 4D training data from infinite RGB videos. Besides, owing to the diversity of the talking head videos, the reconstructed data empowers our model to further disentangle other auxiliary properties of the face (e.g., emotion, etc.) for manipulation. To be specific, we first pre-train Face-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "Former on the reconstructed emotional 4D dataset with the corresponding emotion codes. At the fine-tuning stage, different emotion codes can be embedded into the FaceFormer backbone to realize the emotion control of the animated talking head." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 547, + 156 + ], + "type": "text", + "content": "The main contributions of our work can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 164, + 545, + 396 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 164, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 545, + 236 + ], + "type": "text", + "content": "- We propose Speech4Mesh, a novel framework that tackles the data scarcity issue in speech-driven 3D facial animation by consecutively reconstructing pseudo-4D data and training an audio2mesh network. This framework also provides a referable scheme to alleviate the data scarcity problem of similar 4D tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 243, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 545, + 327 + ], + "type": "text", + "content": "- This framework enables emotional control in the absence of high-quality expressive 4D data by pretraining with the emotional pseudo-4D talking head data reconstructed from 2D videos. And the synthesized emotional 4D talking head dataset can be found in https://github.com/haonanhe/MEAD-3D/tree/main." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 335, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 335, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 335, + 545, + 396 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to utilize speech information as complement for monocular 3D face reconstruction. Empirical experiments show that the exploitation of multi-modal information yields more natural and precise results." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 406, + 392, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 392, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 392, + 418 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 715 + ], + "type": "text", + "content": "Speech-driven 3D Facial Animation: Speech-driven facial animation has been widely explored over the years. Existing methods could be categorized into 2D-based approaches which attempt to operate in 2D pixel space [72, 70, 45, 15, 11, 8], and 3D-based approaches that we focus on. Some works animate 3D models based on visemes through complex rules of coarticulation [19, 73, 68, 57, 12]. Though they could capture accurate lip movements and achieve explicit control over the entire process, such methods usually involve plenty of manual adjustments. Data-driven approaches have been proposed to automatically learn a map from acoustic features to 3D models. Most of them were trained in a supervised manner [20, 13, 47, 6, 34] with multi-view motion capture data [4, 6] or high-resolution 4D scans [34, 13]. [6] searches the matching mouth animation for phonemes based on an Anime Graph structure. [34] proposes an end-to-end convolutional network and uses a latent code to control the emotional state of animated mesh. VOCA [13] achieves speaker-independent facial animation by leveraging a subject-specific latent code. MeshTalk [47] learns a categorical latent space to disentangle audio-correlated and audio-uncorrelated information, which enables it to animate the entire face. FaceFormer [20] first applies Transformer to 3D facial animation to make an autore" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14193" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "pressive prediction of the animated 3D face mesh sequence, which claims to be able to capture accurate lip movements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 99, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 99, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 99, + 288, + 312 + ], + "type": "text", + "content": "However, 4D scans are usually too expensive to access, which leads to the data scarcity issue. There are methods trying to address this problem by using 4D data generated from videos [56, 65, 44]. [44] trains an LSTM-RNN model to regress the shape parameter of the parametric face model tracked from input video frames. [56] uses a sliding window predictor to learn mappings from phoneme sequences to lower facial shape and appearance that an Active Appearance Model (AAM) tracks from video frames. [65] aims to synthesize 3D talking-head with rich emotion. However, their 3D faces tracked from videos are only optimized by 2D landmarks loss, which may lead to inaccurate reconstruction and animation results. In contrast, our method leverages speech features to capture 3D faces with more accurate facial movements. In addition, pre-training on emotion datasets enables our reconstructed 3D face meshes to have emotion states, which leads to 3D facial animation with rich emotion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 288, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 508 + ], + "type": "text", + "content": "Monocular 3D face reconstruction: Monocular 3D face reconstruction aims to track and recover the 3D shape and texture of the human face from monocular 2D data. Early methods often perform an optimization procedure in an analysis-by-synthesis framework, predicting landmarks[75, 62, 30, 5] or local features[29, 50] to regress parameters of 3D Morphable Models (3DMM) [2, 43, 3, 35]. Although optimization-based approaches could recover high-quality 3D faces, they usually involve costly optimization processes. In recent years, learning-based methods were proposed for more efficient reconstruction. Some of them predict 3DMM coefficients [32, 74, 48, 37, 49] directly. To alleviate constraints caused by models' limited geometry, there are also model-free methods directly regressing 3D meshes [16, 18, 23, 33, 53], voxels [31] or Signed Distance Function (SDF) [42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "However, the greatest limitation of learning-based methods is the lack of paired training data. Some methods address this problem by generating synthetic data or using image-model pairs fitted by traditional optimization-based methods [63, 66, 28, 38, 74, 48, 53]. Others have tried to solve it in an unsupervised manner. Most of unsupervised methods try to regress 3DMM parameters by predicting 2D landmarks [17, 22, 40, 52, 54, 58, 59, 60, 69], photometric constraints [17, 22, 25, 54, 59, 60, 69] or multi-image constraints[17, 22, 25, 52, 58]. Perceptual constraints such as face recognition features [17, 22, 25, 54] or emotion features [14] are also utilized to constrain models' learning objects. While unsupervised methods provide a large amount of training data, reconstructing 3D faces from merely 2D information is still underconstrained. To capture more accurate expressions from monocular videos, our framework leverages speech information as another source" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "of constraints, which has a strong relationship with facial movements, especially the mouth area." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 108, + 391, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 391, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 391, + 121 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 129, + 545, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 213 + ], + "type": "text", + "content": "Current monocular 3D face reconstruction methods generally leverage the self-supervised scheme for training. In such a way, a parametric face model and a differentiable renderer are required to recover the 3D face mesh and render the 3D face back to the image for training, respectively. In this section, we review the basic knowledge of the parametric face model and differentiable renderer for reference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 217, + 369, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 369, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 369, + 228 + ], + "type": "text", + "content": "3.1. FLAME" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 236, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 295 + ], + "type": "text", + "content": "In this work, we used FLAME [36] as the prior parametric face model for reconstruction, which is a generic 3D head model using standard vertex-based linear blend skinning (LBS) with corrective blendshapes. Specifically, it consists of " + }, + { + "bbox": [ + 305, + 236, + 545, + 295 + ], + "type": "inline_equation", + "content": "N = 5023" + }, + { + "bbox": [ + 305, + 236, + 545, + 295 + ], + "type": "text", + "content": " vertices and could be defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 383, + 304, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 304, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 383, + 304, + 545, + 319 + ], + "type": "interline_equation", + "content": "M (\\boldsymbol {\\beta}, \\boldsymbol {\\theta}, \\boldsymbol {\\psi}) \\in \\mathbb {R} ^ {3 N}, \\tag {1}", + "image_path": "e16dc43d6661d8afb5d47f7f261c98a0164e86eea4581cd18ab5973bc25916a1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "type": "text", + "content": "which contains shape parameters " + }, + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{|\\beta|}" + }, + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "type": "text", + "content": " accounting for shape variation, expression " + }, + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\psi \\in \\mathbb{R}^{|\\psi|}" + }, + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "type": "text", + "content": " to capture facial expressions, and pose parameters " + }, + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{3K + 3}" + }, + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "type": "text", + "content": " to rotate vertices around " + }, + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 304, + 328, + 545, + 449 + ], + "type": "text", + "content": " joints (neck, jaw, and eyeballs) as well as correcting pose deformations. FLAME provides a geometric constraint and a low-dimensional latent space, making our reconstruction process much easier. Besides, highly disentangled parameters allow us to control the face model's expression or pose independently, which benefits us in both the reconstruction and animation phases." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 449, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 509 + ], + "type": "text", + "content": "We obtain FLAME's texture map through an appearance model " + }, + { + "bbox": [ + 305, + 449, + 545, + 509 + ], + "type": "inline_equation", + "content": "A(\\alpha) \\in \\mathbb{R}^{d \\times d \\times 3}" + }, + { + "bbox": [ + 305, + 449, + 545, + 509 + ], + "type": "text", + "content": " with albedo parameters " + }, + { + "bbox": [ + 305, + 449, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}^{|\\alpha|}" + }, + { + "bbox": [ + 305, + 449, + 545, + 509 + ], + "type": "text", + "content": " as input. This is achieved by converting Basel Face Model's linear albedo space to FLAME's UV layout. More details can be found in [35]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 515, + 446, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 446, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 446, + 528 + ], + "type": "text", + "content": "3.2. Differentiable Rendering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 534, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 534, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 305, + 534, + 545, + 618 + ], + "type": "text", + "content": "3D rendering is a function that maps the 3D object to a 2D image. Differentiable rendering, as the name implies, provides a differentiable rendering function where the gradient of the function with respect to the 3D object parameters can be effectively computed. In this work, we use the differentiable rasterizer from Pytorch3D [46] to render parametric FLAME model " + }, + { + "bbox": [ + 305, + 534, + 545, + 618 + ], + "type": "inline_equation", + "content": "M(\\beta, \\theta, \\psi)" + }, + { + "bbox": [ + 305, + 534, + 545, + 618 + ], + "type": "text", + "content": " to 2D images " + }, + { + "bbox": [ + 305, + 534, + 545, + 618 + ], + "type": "inline_equation", + "content": "I_R" + }, + { + "bbox": [ + 305, + 534, + 545, + 618 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 364, + 628, + 545, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 628, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 364, + 628, + 545, + 641 + ], + "type": "interline_equation", + "content": "R (M (\\boldsymbol {\\beta}, \\boldsymbol {\\theta}, \\psi), \\boldsymbol {\\alpha}, \\mathbf {l}, \\mathbf {c}) \\rightarrow I _ {R}, \\tag {2}", + "image_path": "dfdeca209754f0dde60d43a1fe5fc0eba8d8ed1dc9524b02a9cfc6201c4ccaa5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "text", + "content": "with albedo parameters " + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "text", + "content": ", Spherical Harmonics (SH) lighting parameters " + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{l} = \\left[\\mathbf{l}_1^T, \\dots, \\mathbf{l}_9^T\\right]^T" + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_k \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "text", + "content": " to create shaded textures and camera parameters " + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "text", + "content": " which is the concatenation of isotropic scale " + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}" + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "text", + "content": " and 2D translation " + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 305, + 651, + 545, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14194" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 281, + 236 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 281, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 281, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 281, + 236 + ], + "type": "image", + "image_path": "fa7ddd1d4745c95c0e35e7d58047ae79b6071804af14b24a9967df591f9a8f43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 240, + 288, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 288, + 289 + ], + "type": "text", + "content": "Figure 2: Speech-Assisted Monocular 3D Face Reconstruction. The block in grey color are fixed models or networks, and the Expression Encoder " + }, + { + "bbox": [ + 46, + 240, + 288, + 289 + ], + "type": "inline_equation", + "content": "E_{e}" + }, + { + "bbox": [ + 46, + 240, + 288, + 289 + ], + "type": "text", + "content": " in blue color is the learnable network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 293, + 133, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 133, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 133, + 306 + ], + "type": "text", + "content": "4. Speech4Mesh" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 313, + 287, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 287, + 337 + ], + "type": "text", + "content": "4.1. Speech-Assisted Monocular 3D Face Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 343, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 570 + ], + "type": "text", + "content": "In the proposed Speech4Mesh framework, 3D face sequences are first reconstructed from RGB videos. The detailed structure of our reconstruction module is illustrated in Fig. 2. The overall structure is built on DECA [22], the state-of-the-art method in monocular 3D face reconstruction. DECA learns animatable detailed 3D face models from static RGB images " + }, + { + "bbox": [ + 46, + 343, + 287, + 570 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 343, + 287, + 570 + ], + "type": "text", + "content": " in a two-stage scheme. The coarse stage first predicts parameters to reconstruct a rough face shape based on the FLAME model. The detail stage then generates a UV displacement map conditioned on expression and jaw pose parameters to achieve fine-grained details. Since the detail stage may bring noise to the reconstructed mesh, we only borrow the coarse part as our backbone. In DECA's coarse part, an encoder (ResNet-50) learns to regress 3DMM coefficients merely under visual constraints, including landmark-related losses, photometric loss, perceptual-level loss, and shape consistency loss. We will introduce some of the losses we have used later in the text." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "type": "text", + "content": "Expression encoder: Since DECA already can capture the overall facial features well, we fix the pre-trained coarse encoder " + }, + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "type": "text", + "content": " in DECA for shape, pose (except jaw), camera, albedo, and light parameters. We then train an additional expression encoder " + }, + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "type": "inline_equation", + "content": "E_{e}" + }, + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "type": "text", + "content": " to regress expression parameters " + }, + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "type": "inline_equation", + "content": "\\psi_{e}" + }, + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "type": "text", + "content": " and jaw parameters " + }, + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "type": "inline_equation", + "content": "\\theta_{e}^{jaw}" + }, + { + "bbox": [ + 46, + 571, + 286, + 644 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 647, + 287, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 647, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 125, + 647, + 287, + 662 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {e} ^ {j a w}, \\psi_ {e} = E _ {e} (I). \\tag {3}", + "image_path": "586fcd4b58481ee08fa9f5003b8065e3973fb0e5aba5bd3b82f133760cd7dba3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 666, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 286, + 689 + ], + "type": "text", + "content": "Note that, for computational efficiency, we use MobileNetV2 instead of ResNet-50 as the backbone network." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Only training " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "E_{e}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " simplifies the learning goal and enables us to remove losses that are unrelated to lip motion such as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "identity loss " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "L_{id}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and shape consistency loss " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "L_{sc}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " in DECA, which leads to faster training speed and fewer training resources. Also, speech content would mainly affect mouth-related movements, which are most relevant to expression and jaw pose. The rendering output of the reconstruction part is then denoted as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 152, + 545, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 152, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 339, + 152, + 545, + 167 + ], + "type": "interline_equation", + "content": "R \\left(M \\left(\\boldsymbol {\\beta}, [ \\boldsymbol {\\theta}, \\boldsymbol {\\theta} _ {e} ^ {j a w} ], \\psi_ {e}\\right), \\boldsymbol {\\alpha}, 1, \\mathbf {c}\\right)\\rightarrow I _ {R e}, \\tag {4}", + "image_path": "a54c34f496acd4d1bdf90a49218a0b5289e7cbc1962a56a461bcea739c64919f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 173, + 545, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 198 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 304, + 173, + 545, + 198 + ], + "type": "inline_equation", + "content": "[\\theta, \\theta_{e}^{jaw}]" + }, + { + "bbox": [ + 304, + 173, + 545, + 198 + ], + "type": "text", + "content": " means we replace the original vectors representing jaw pose in " + }, + { + "bbox": [ + 304, + 173, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 173, + 545, + 198 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 173, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\theta_{e}^{jaw}" + }, + { + "bbox": [ + 304, + 173, + 545, + 198 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "text", + "content": "Speech-visual contrastive loss: As discussed in previous sections, the majority of current works only utilize the visual information from utterance videos. While these methods have achieved acceptable reconstruction results, it is still an underconstrained problem to retrieve 3D faces merely from 2D planes. Ignoring information contained in audio can lead to a considerable loss in reconstruction. Thus, we propose to leverage speech constraints here to achieve more precise 3D face reconstruction in a contrastive manner. We calculate the contrastive loss using a minibatch of N speech-visual representation pairs " + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "inline_equation", + "content": "(\\epsilon_s,\\epsilon_v)" + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "inline_equation", + "content": "\\epsilon_v" + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "text", + "content": " is the hidden state before the last 4-layer MLP in expression encoder, which projects " + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "inline_equation", + "content": "\\epsilon_v" + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "text", + "content": " to the target 3DMM space, and " + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "inline_equation", + "content": "\\epsilon_s" + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "text", + "content": " is the audio embedding. In specific, " + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "inline_equation", + "content": "\\epsilon_s" + }, + { + "bbox": [ + 304, + 198, + 546, + 377 + ], + "type": "text", + "content": " is obtained by the pre-trained wav2vec 2.0 [1] model as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 399, + 385, + 545, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 385, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 399, + 385, + 545, + 398 + ], + "type": "interline_equation", + "content": "\\epsilon_ {s} = E _ {s} (S), \\tag {5}", + "image_path": "63e442d085e8a5643be0727feacc6b8a0e0a3f34d93f0e965c754eaef3e44e7d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "type": "text", + "content": "which can extract powerful representations from speech audio. Since we interpolate the frequency of audio features to be two times the video frame rate, we then adopt a linear combination of audio features from time steps " + }, + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "type": "inline_equation", + "content": "(t - \\lfloor T / 2\\rfloor)\\cdot 2" + }, + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "type": "inline_equation", + "content": "(t + \\lfloor T / 2\\rfloor)\\cdot 2" + }, + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "type": "text", + "content": " for frame " + }, + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "type": "text", + "content": " as the speech-visual positive pair, where " + }, + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 407, + 545, + 538 + ], + "type": "text", + "content": " is set to 5 in our experiment. Minimizing the contrastive loss leads to encoders that maximally preserve the mutual information between positive speech-visual pairs. The contrastive loss consists of two terms, in which the first term calculates the speech-to-visual contrastive cost for the i-th pair:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 343, + 545, + 545, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 545, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 343, + 545, + 545, + 575 + ], + "type": "interline_equation", + "content": "\\ell_ {i} ^ {(s \\rightarrow v)} = - \\log \\frac {\\exp \\left(\\left\\langle \\epsilon_ {\\mathbf {i}} ^ {\\mathbf {s}}, \\epsilon_ {\\mathbf {i}} ^ {\\mathbf {v}} \\right\\rangle / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(\\left\\langle \\epsilon_ {\\mathbf {i}} ^ {\\mathbf {s}}, \\epsilon_ {\\mathbf {k}} ^ {\\mathbf {v}} \\right\\rangle / \\tau\\right)} \\tag {6}", + "image_path": "c115d71f0e29369454ce4238f8df454f1baefef502738e5bfa230898e1738f74.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\langle \\epsilon_{\\mathbf{i}}^{\\mathbf{s}},\\epsilon_{\\mathbf{i}}^{\\mathbf{v}}\\rangle" + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "content": " calculates cosine similarity between two vectors, i.e., " + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{v},\\mathbf{u}\\rangle = \\mathbf{v}^\\top \\mathbf{u} / \\| \\mathbf{v}\\| \\| \\mathbf{u}\\|" + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\tau \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "content": " is the temperature parameter, which is fixed to 0.07 in our experiment. While the second term is similar to the first term, it calculates cosine similarity as " + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\langle \\epsilon_{\\mathbf{i}}^{\\mathbf{v}},\\epsilon_{\\mathbf{i}}^{\\mathbf{s}}\\rangle" + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "content": ". The final contrastive loss is the weighted sum of the two terms:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 333, + 661, + 545, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 661, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 333, + 661, + 545, + 693 + ], + "type": "interline_equation", + "content": "L _ {c o n} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\lambda \\ell_ {i} ^ {(s \\rightarrow v)} + (1 - \\lambda) \\ell_ {i} ^ {(v \\rightarrow s)}\\right), \\tag {7}", + "image_path": "62e02030b678ef50062df06f430d90dad678a05c54644a37ddf80447b121d9d5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 701, + 542, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 542, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 542, + 714 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 306, + 701, + 542, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 701, + 542, + 714 + ], + "type": "text", + "content": " to 0.5 in our experiment as a scalar weight." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14195" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "Weighted landmark loss: We use weighted landmark loss to minimize the " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " distance between ground-truth 2D key points and corresponding landmarks projected into the image plane from mesh." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 124, + 287, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 124, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 91, + 124, + 287, + 156 + ], + "type": "interline_equation", + "content": "L _ {l m k} = \\sum_ {i = 1} ^ {6 8} \\lambda_ {i} \\| \\mathbf {k} _ {i} - s \\Pi (M _ {i}) + t \\| _ {1} \\tag {8}", + "image_path": "0cb9fba7cc5ec98efdc05b2d8e06ce3046df964dc8618402bf5922fd047ad209.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "type": "text", + "content": " is the weight of the " + }, + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "type": "text", + "content": "th landmark, " + }, + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "type": "inline_equation", + "content": "\\Pi(M_{i})" + }, + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "type": "text", + "content": " represents projecting 3D meshes " + }, + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 46, + 160, + 287, + 219 + ], + "type": "text", + "content": " to the 2D pixel space. We set weights of landmarks in the mouth, jaw, and nose area higher than others to capture more accurate geometry representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 220, + 266, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 220, + 266, + 232 + ], + "spans": [ + { + "bbox": [ + 59, + 220, + 266, + 232 + ], + "type": "text", + "content": "Eye closure loss: We calculate the eye closure loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 236, + 287, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 236, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 75, + 236, + 287, + 263 + ], + "type": "interline_equation", + "content": "L _ {e y e} = \\sum_ {(i, j) \\in E} \\| \\mathbf {k} _ {i} - \\mathbf {k} _ {j} - s \\Pi (M _ {i} - M _ {j}) \\| _ {1} \\tag {9}", + "image_path": "4fb5e5a22565383469fc07680ace3332b72e3f916982803f4c66fb83c3631a10.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "type": "text", + "content": "to force the relative offset of vertices " + }, + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "type": "inline_equation", + "content": "M_{j}" + }, + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "type": "text", + "content": " on the upper and lower eyelid to be close to the offset of ground-truth eyelid landmarks pairs " + }, + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i" + }, + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_j" + }, + { + "bbox": [ + 46, + 266, + 287, + 315 + ], + "type": "text", + "content": ". In this way, the reconstructed mesh could capture the blink of eye." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 316, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 351 + ], + "type": "text", + "content": "Lip closure loss: The lip closure loss " + }, + { + "bbox": [ + 46, + 316, + 287, + 351 + ], + "type": "inline_equation", + "content": "L_{lp}" + }, + { + "bbox": [ + 46, + 316, + 287, + 351 + ], + "type": "text", + "content": " is similar to the eye closure loss. It measures the relative distance between the upper/lower lip and left/right lip corners keypoint pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 351, + 224, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 351, + 224, + 362 + ], + "spans": [ + { + "bbox": [ + 59, + 351, + 224, + 362 + ], + "type": "text", + "content": "Photometric loss: The photometric loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 367, + 287, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 367, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 107, + 367, + 287, + 381 + ], + "type": "interline_equation", + "content": "L _ {p h o} = \\left\\| V _ {I} \\odot (I - I _ {R e}) \\right\\| _ {1, 1} \\tag {10}", + "image_path": "8fc953300fe3ecc762a99c2b2bae2822b1ff0e6a15bea6489d86e72877ca0395.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 385, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 432 + ], + "type": "text", + "content": "measures the pixel-to-pixel differences between ground-truth training images and rendered images. " + }, + { + "bbox": [ + 46, + 385, + 287, + 432 + ], + "type": "inline_equation", + "content": "V_{I}" + }, + { + "bbox": [ + 46, + 385, + 287, + 432 + ], + "type": "text", + "content": " is the face skin mask with a value 1 in the face skin area and " + }, + { + "bbox": [ + 46, + 385, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 385, + 287, + 432 + ], + "type": "text", + "content": " denotes the Hadamard product." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 433, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 504 + ], + "type": "text", + "content": "Emotion consistency loss: To capture more vivid expressions, we borrow the idea from EMOCA [14] that introduces an emotion consistency loss with a fixed emotion recognition network. We compute the " + }, + { + "bbox": [ + 46, + 433, + 287, + 504 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 433, + 287, + 504 + ], + "type": "text", + "content": " distance between emotion features of input images and output rendered images as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 509, + 287, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 509, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 116, + 509, + 287, + 523 + ], + "type": "interline_equation", + "content": "L _ {e m o} = \\left\\| \\epsilon_ {I} ^ {e m o}, \\epsilon_ {R e} ^ {e m o} \\right\\| _ {2}, \\tag {11}", + "image_path": "a4e5bed55fc34967ac419ba147da9c923085b608a8733481815eb7a5920a5cf6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 526, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 526, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\epsilon_I^{emo}" + }, + { + "bbox": [ + 46, + 526, + 287, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 526, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\epsilon_{Re}^{emo}" + }, + { + "bbox": [ + 46, + 526, + 287, + 550 + ], + "type": "text", + "content": " are produced by a fixed emotion recognition network." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "content": "Regularization term: We constrain the expression parameters " + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\psi_{e}" + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "content": " and jaw parameters " + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\theta_{e}^{jaw}" + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "content": " produced by our expression encoder to be close to the original parameters produced by pre-trained DECA by minimizing their " + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "content": " distance as performed in [24]: " + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\| \\psi_{e} - \\psi \\|_{2}" + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 552, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\left\\| \\theta_e^{jaw} - \\theta^{jaw}\\right\\| _2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 623, + 270, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 623, + 270, + 635 + ], + "spans": [ + { + "bbox": [ + 59, + 623, + 270, + 635 + ], + "type": "text", + "content": "Loss function: To sum up, our optimization goal is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 639, + 287, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 287, + 677 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {r e c o n}} = \\lambda_ {c o n} L _ {c o n} + \\lambda_ {l m k} L _ {l m k} + \\lambda_ {e y e} L _ {e y e} + \\lambda_ {l p} L _ {l p} \\\\ \\lambda_ {p h o} L _ {p h o} + \\lambda_ {e m o} L _ {e m o} + \\lambda_ {r e g} L _ {r e g}, \\tag {12} \\\\ \\end{array}", + "image_path": "3c62102d1c91f653348183995f4e55add213bf5cef4f8463a47cf5d1d63058c0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " with different subscripts are hyper-parameters to balance the corresponding losses. The detailed values of " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "s can be found in the supplemental material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 84 + ], + "type": "text", + "content": "4.2. Speech-Driven Emotional 3D Facial Animation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 90, + 545, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 210 + ], + "type": "text", + "content": "After acquiring sufficient synthesized 4D training data, we can train the audio2mesh module for animation. We use FaceFormer [20], a Transformer-based model to drive 3D face vertices with speech embeddings. FaceFormer [20] is the first work to apply Transformer to 3D facial animation. It has achieved better performances in both realism and lip synchronization than prior methods owing to the Transformer's ability to learn long-term audio dependencies. Specifically, FaceFormer predicts the vertices' movement in an autoregressive manner as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 357, + 220, + 545, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 220, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 357, + 220, + 545, + 233 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} _ {\\mathbf {t}} = \\operatorname {F a c e F o r m e r} \\left(\\hat {\\mathbf {y}} _ {< \\mathbf {t}}, \\mathbf {i} _ {\\mathbf {n}}, \\mathbf {S}\\right), \\tag {13}", + "image_path": "c16d2fc6f1a59ef2e918453fd2bb0dd69c8534f43a4629dd3756b8975b78267c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "text", + "content": " represents the predicted vertices offset at time step " + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_{\\mathbf{n}}" + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "text", + "content": " represents the identity code. Although FaceFormer has achieved remarkable animation results, its performance highly depends on the data quality, since it is trained in a fully supervised scheme where the only supervision is the MSE loss between predicted and ground-truth meshes:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 356, + 324, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 324, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 356, + 324, + 545, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S E}} = \\sum_ {\\mathbf {t} = 1} ^ {\\mathbf {T}} \\sum_ {\\mathbf {v} = 1} ^ {\\mathbf {V}} \\left\\| \\hat {\\mathbf {y}} _ {\\mathbf {t}, \\mathbf {v}} - \\mathbf {y} _ {\\mathbf {t}, \\mathbf {v}} \\right\\| ^ {2}. \\tag {14}", + "image_path": "329458d5cb055a98a096b817a323fdf4fe53e0d19fbd406fe7ef416758ed8968.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": "In such a way, it's hard to capture some facial details due to the registration loss of the scanned 4D training data. To this end, during pre-training with the synthesized data, we further introduce a pixel-level constraint by computing the photometric loss between the rendered images and corresponding real images as follow:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 352, + 447, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 447, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 352, + 447, + 545, + 460 + ], + "type": "interline_equation", + "content": "R \\left(M _ {T} + \\hat {\\mathbf {y}} _ {\\mathbf {t}}, \\boldsymbol {\\theta} _ {\\mathbf {t}}, \\boldsymbol {\\alpha} _ {\\mathbf {t}}, \\mathbf {l} _ {\\mathbf {t}}, \\mathbf {c} _ {\\mathbf {t}}\\right)\\rightarrow I _ {R e} ^ {t}, \\tag {15}", + "image_path": "cd9e0ef45d37aad3f088e78c2d8ba9375edeb50ca8c5f0308db0fc3d29bf4027.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": " is the template face mesh (the mean of all faces in the training set as in FaceFormer). Note that, when performing differentiable rendering, since the output meshes of the FaceFormer network are all aligned, i.e., without any rotation or camera transform, we borrow " + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\theta_t^1" + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "inline_equation", + "content": "l_t" + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 304, + 470, + 545, + 578 + ], + "type": "text", + "content": " estimated by our 3D reconstruction encoders to render the vertices to 2D images. Except for the pixel-to-pixel loss, we also calculate landmark-related losses as in the reconstruction part." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 578, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 661 + ], + "type": "text", + "content": "Furthermore, due to the insufficiency of emotional 4D data, current audio2mesh methods generally lack emotion controllability. To make our meshes more lifelike, we introduce a one-hot emotion code to explicitly control the emotional state. Specifically, during pre-training on the emotional 2D video dataset, we embed the emotion category into the FaceFormer network as follows:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 348, + 672, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 672, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 348, + 672, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} _ {\\mathbf {t}} = \\operatorname {F a c e F o r m e r} \\left(\\hat {\\mathbf {y}} _ {< \\mathbf {t}}, \\mathbf {i} _ {\\mathbf {n}}, \\mathbf {e} _ {\\mathbf {m}}, \\mathbf {S}\\right), \\tag {16}", + "image_path": "e5054f40f33d7600bc9544195b66bb85d9c3948241a5cf5401ba803c02f56367.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": "1With slight abuse of notation, the " + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\theta_t" + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": " here does not include the jaw rotation." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14196" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 532, + 238 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 532, + 238 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 532, + 238 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 532, + 238 + ], + "type": "image", + "image_path": "45bd049caaf13a367ce7d68d980459b57a782213f60b66bcf43945c592b8857a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 297 + ], + "type": "text", + "content": "Figure 3: Comparison of the reconstruction results on MEAD (left) and VoxCeleb2 (right) datasets. The first column on the left is the original RGB images. The reconstructed faces from the second to the sixth column are from 3DDFA-V2, Deep3DFace (Pytorch), DECA (coarse), EMOCA (coarse), and ours, respectively. More qualitative results could be found in the supplementary materials." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 288, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 303, + 288, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{m}}" + }, + { + "bbox": [ + 46, + 303, + 288, + 361 + ], + "type": "text", + "content": " is the one-hot vector projected to a higher dimension space. In this way, during the inference stage, we can achieve speaker-independent emotion-controllable 3D facial animation by embedding different emotion vectors into FaceFormer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 363, + 288, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 288, + 411 + ], + "type": "text", + "content": "To sum up, by employing the self-supervised differentiable rendering scheme and emotion code, our modified FaceFormer model can generate not only more realistic but also emotion-controllable animated face meshes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 423, + 128, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 128, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 128, + 436 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 443, + 105, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 105, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 105, + 454 + ], + "type": "text", + "content": "5.1. Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 509 + ], + "type": "text", + "content": "In this paper, we leveraged RGB video data and 4D mesh data for training the reconstruction module and fine-tuning the audio2mesh module, respectively. A brief introduction of the datasets we used is as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "content": "2D Training Data: MEAD [64], VoxCeleb2 [10]. The MEAD dataset is a talking head corpus with different intensities of emotions. It has a total of 281,400 video clips collected from 60 English-speaking actors and actresses under 8 common emotions with 3 levels of intensity. Note that although the dataset contains videos from different views, we only took the front-face data for training to keep more visible details. VoxCeleb2 is a large-scale talking head video dataset collected from open-source media which covers a huge range of identities, races, languages, and head poses. In practice, we randomly sampled 2,000 identities for training. The images were extracted under 30 and 25 fps for MEAD and VoxCeleb2 videos, respectively. We used wav2vec 2.0 [1] to exact the audio embeddings, in which the output frequency is " + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "inline_equation", + "content": "49\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "content": ". To ensure temporal alignment of the image and audio, we interpolate the audio embeddings to " + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "inline_equation", + "content": "50\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "content": " for the VoxCeleb2 dataset and to " + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "inline_equation", + "content": "60\\mathrm{Hz}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 303, + 400, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 400, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 400, + 313 + ], + "type": "text", + "content": "for the MEAD dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 315, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 545, + 387 + ], + "type": "text", + "content": "4D Fine-Tuning Data: VOCASET [13]. VOCASET is a 4D dataset composed of consecutive scanned 3D heads which are all registered to a 5023 vertices mesh. It contains a total of 480 talking motion sequences from 12 subjects under the frame rate of 60 fps. The training, validation, and testing splits are directly adopted from [20]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 395, + 399, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 399, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 399, + 407 + ], + "type": "text", + "content": "5.2. Training Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 414, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 414, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 414, + 545, + 498 + ], + "type": "text", + "content": "The reconstruction part was trained for a maximum of 20 epochs, using the Adam optimizer and a learning rate of " + }, + { + "bbox": [ + 305, + 414, + 545, + 498 + ], + "type": "inline_equation", + "content": "5e - 5" + }, + { + "bbox": [ + 305, + 414, + 545, + 498 + ], + "type": "text", + "content": ". In one iteration, we used 4 mini-batches of data, each of which contains 16 frames randomly sampled from a single video. We further implied distributed data-parallel technique to collect data from 4 GPUs to enlarge the pool of training pairs for contrastive learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 568 + ], + "type": "text", + "content": "The audio2mesh network was trained using the Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 304, + 498, + 545, + 568 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 304, + 498, + 545, + 568 + ], + "type": "text", + "content": " for a total of 100 epochs. The parameters of the encoder were initialized with the pre-trained wav2vec 2.0 weights and the parameters of TCN were fixed during training. The period " + }, + { + "bbox": [ + 304, + 498, + 545, + 568 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 498, + 545, + 568 + ], + "type": "text", + "content": " for the biased causal multi-head self-attention was set to 30." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": "We pre-trained our audio2mesh network on 4D data produced by our monocular 3D reconstruction network from the MEAD dataset. While it contains 3 levels of emotional intensity, we only use the mildest first level to train. We created 8 one-hot emotion codes (angry, contempt, disgusted, fear, happy, sad, surprised, or neutral) and 42 one-hot speaker identity codes. For each pair of image and its 3D reconstruction in the MEAD dataset, we embedded its corresponding emotion code and speaker ID code to a higher dimension and then concatenated and projected it to the audio2mesh network." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 712 + ], + "type": "text", + "content": "We fine-tuned the audio2mesh network on VOCASET" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14197" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "which contains high-resolution 4D scans. During training, we randomly chose 8 kinds of speaker ID codes defined in the pre-train phase and set the emotion code to neutral." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 118, + 273, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 118, + 273, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 273, + 131 + ], + "type": "text", + "content": "5.3. Analysis of 3D Face Reconstruction Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 137, + 289, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 137, + 289, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 289, + 435 + ], + "type": "text", + "content": "Subjective Evaluation: Visualization of the reconstruction results is demonstrated in Fig. 3. We randomly sampled several frames from MEAD and VoxCeleb2 datasets for comparison. For clear observation of the expression details, we did not apply texture on the face. The sampled frames contain a variety of talking heads with various lip motions, extreme emotions, different genders, head poses, ages, and races. From Fig. 3, we can find that our method apparently outperforms the 3DDFA-V2 and DECA regarding the lip motion as well as the overall appearance. The curly lips and emotional expressions cannot be recovered by these two baselines. EMOCA is an expert in capturing extreme emotions. However, some mild expressions will be significantly exaggerated by EMOCA, which causes unnatural artifacts and inaccurate lip shapes in the reconstruction results. More comparison results with EMOCA could be found in supplementary materials. Note that, we used the Pytorch version of Deep3DFace for comparison, which is claimed to be better than the Tensorflow implementation. As shown in the third column in Fig. 3, the recovered faces by Deep3DFace reveal impressive identity consistency compared with the original RGB image. This may attribute to the identity aggregation mechanism and a different parametric face model (i.e., Basel Face Model [43]). However, the lower face expressiveness is still lower than our proposed method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": "We further conducted a user study to make a perceptual evaluation between our and the competitors' reconstruction results. Specifically, 150 participants were asked to judge a side-by-side image composed of the original image, our reconstruction result, and the competitor's result. The participants can select their favorable one or choose the equal option if they cannot distinguish the difference. 50 samples were randomly selected from the MEAD and Voxceleb2 datasets. The result of the study is summarized in Tab. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": "As we can see from Tab. 1, the result demonstrates an obvious inclination to support our results. It is worth mentioning that about 80 percent of our reconstructed samples are deemed better than or equal with 3DDFA-V2 and DECA methods regarding both full and lower face. EMOCA improves the perception of the results due to its emotion prior. Deep3DFace also achieves better performance as we discussed above. However, owing to the audio-visual alignment, our method achieves better performance at both full-face and the low-face region according to the user study." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Objective Evaluation: Our reconstruction model's ability to capture lip shapes is measured using the lip vertex error as the objective metric, which has been extensively leveraged to evaluate the precision of lip synchronization" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 538, + 157 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 538, + 157 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 538, + 157 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 538, + 157 + ], + "type": "table", + "html": "
Ours vs. CompetitorFace regionFavorability(%) ↑Ours better or equal
EqualComp.Ours
Ours vs. 3DDFA-V2Full5.419.475.280.6
Lower5.421.373.378.7
Ours vs. DECAFull11.418.869.881.2
Lower9.818.571.781.5
Ours vs. EMOCAFull8.438.353.358.9
Lower8.337.254.562.8
Ours vs. Deep3DFaceFull5.941.153.158.9
Lower8.734.257.065.8
", + "image_path": "4cb19f1070bd50aeaffb05dede2642387ca0a7363120434843da6e1bd5ca961c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 350, + 200, + 501, + 227 + ], + "blocks": [ + { + "bbox": [ + 304, + 160, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 304, + 160, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 160, + 545, + 196 + ], + "type": "text", + "content": "Table 1: A user study of the reconstruction results. We compare the participants' favorability of our reconstructions with the competitors' w.r.t. different regions of the face." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 350, + 200, + 501, + 227 + ], + "lines": [ + { + "bbox": [ + 350, + 200, + 501, + 227 + ], + "spans": [ + { + "bbox": [ + 350, + 200, + 501, + 227 + ], + "type": "table", + "html": "
MetricOursEMOCA
Lip Vertex Error ↓0.995e-21.359e-2
", + "image_path": "c381aa38cc2fc73446f2887157cd44d4b041542aaf1ea67aa082d9dfad6b71d1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 334, + 230, + 517, + 241 + ], + "lines": [ + { + "bbox": [ + 334, + 230, + 517, + 241 + ], + "spans": [ + { + "bbox": [ + 334, + 230, + 517, + 241 + ], + "type": "text", + "content": "Table 2: Reconstruction error on VOCASET." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 243, + 545, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 362 + ], + "type": "text", + "content": "[20, 47]. To calculate this error, we determine the maximal " + }, + { + "bbox": [ + 304, + 243, + 545, + 362 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 243, + 545, + 362 + ], + "type": "text", + "content": " distance between the lip vertices of the reconstructed meshes and their corresponding ground truth in the VO-CASET. We compared Speech4Mesh with EMOCA, which performs the best in the subjective evaluation. As presented in Table 2, our method outperforms EMOCA in terms of lower lip vertex error, demonstrating our superior ability to capture accurate lip shapes. On the contrary, EMOCA performs poorly on this metric due to its exaggerated facial expression with contorted mouth movements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 395 + ], + "type": "text", + "content": "5.4. Analysis of Speech-Driven Facial Animation Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 402, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 546, + 676 + ], + "type": "text", + "content": "Subjective Evaluation: We also performed a user study on the speech-driven animation results. Similar to the reconstruction user study, we recruited 150 participants to judge the animation quality of our method and the vanilla FaceFormer from the realism and lip synchronization dimensions. The speeches for evaluation are from FaceFormer's test data (the demo video of FaceFormer) and the VoxCeleb2 dataset. For each dataset, we sampled 56 speeches (7 ids, 8 sentences for each id). Each participant evaluates 8 videos selected at random. The statistic of the user study is shown in Tab. 3. From Tab. 3 we can find that the favorability rating of our Speech4Mesh is consistently higher than that of FaceFormer, especially on the more challenging VoxCeleb2 speeches. We also visualize several representative syllables and corresponding animations in Fig. 4 for demonstration. In Fig. 4, our animated faces are more expressive and perceptually reasonable. We would like to note that MeshTalk cannot be well-trained on VOCASET which in a different topology, as mentioned in [61]. We have refrained from presenting any quantitative or qualitative comparison results here. However, we do incorporate a straightforward comparison in the supplementary videos. It is highly recommended to see those videos for more details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Objective Evaluation: We evaluate the lip vertex error of our proposed method and the vanilla FaceFormer on the test set of VOCASET. For simplicity, the scale we used for eval" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14198" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 287, + 147 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 287, + 147 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 287, + 147 + ], + "type": "table", + "html": "
Test Speech DataCriterionFavorability(%) ↑Ours better or equal
EqualFaceFormerSpeech4Mesh
Speech from FaceFormer DemoRealism24.831.843.468.2
Lip Sync21.936.441.863.6
Speech from VoxCeleb2Realism16.326.157.773.9
Lip Sync16.730.752.569.3
", + "image_path": "e9116db4c11dd58686f94a168a900af0c1ebb2f2585e838bd367959e299b63ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 91, + 192, + 241, + 247 + ], + "blocks": [ + { + "bbox": [ + 46, + 149, + 288, + 187 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 288, + 187 + ], + "type": "text", + "content": "Table 3: A user study of the speech-driven results. We compare the participants' favorability of our results with the FaceFormer with respect to realism and lip synchronization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 192, + 241, + 247 + ], + "lines": [ + { + "bbox": [ + 91, + 192, + 241, + 247 + ], + "spans": [ + { + "bbox": [ + 91, + 192, + 241, + 247 + ], + "type": "table", + "html": "
MethodLip Vertex Error ↓
VOCA6.428e-3
FaceFormer5.355e-3
Speech4Mesh4.863e-3
", + "image_path": "77ec48d62bf1a7affe8a387848a75f25b960203dc8736f042f35166339c17ea3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "type": "text", + "content": "Table 4: Comparison of the lip-sync error between VOCA, the vanilla FaceFormer, and Speech4Mesh." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 277, + 214, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 214, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 214, + 289 + ], + "type": "text", + "content": "uation is the same as the FLAME model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "type": "text", + "content": "Tab. 4 displays the lip vertex errors of the vanilla Face-Former method and our Speech4Mesh method. It can be seen that more than a " + }, + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 46, + 291, + 287, + 351 + ], + "type": "text", + "content": " improvement in lip synchronization is brought by our training framework, which restates the importance of data sufficiency for 4D tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 357, + 149, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 149, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 149, + 369 + ], + "type": "text", + "content": "5.5. Emotion Control" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 376, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 497 + ], + "type": "text", + "content": "As introduced in previous sections, our audio2mesh module can be trained on the reconstructed meshes with different emotions. By employing an emotion code, we can manipulate the originally neural 4D talking heads to have different emotions. As shown in Fig. 5, different emotions (e.g., happy, surprised, sad) are embedded into the talking face driven by the same speech. The result reveals the potentiality of more diverse applications by exploiting the 2D videos under the Speech4Mesh framework. More emotion control results could be found in supplementary videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 502, + 148, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 148, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 148, + 514 + ], + "type": "text", + "content": "5.6. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "content": "To understand the impact of the contrastive loss and the photometric loss on speech-driven animation, we conducted quantitative experiments as shown in Tab. 5. More ablation studies could be found in supplementary materials. For the ablation study of contrastive loss, we pre-trained our audio2mesh module using 4D meshes reconstructed from the MEAD dataset, without incorporating contrastive loss during training. For the ablation study of photometric loss, we simply removed it in the pre-training stage. The quantitative results in Tab. 5. show that pre-training without contrastive loss or photometric loss results in a higher lip vertex error. These results underscore the importance of high-quality reconstructed 4D meshes in our method. They also support the notion that the photometric loss can encourage the audio2mesh network to learn detailed information which can not be captured solely from 4D meshes." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 523, + 178 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 523, + 178 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 523, + 178 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 523, + 178 + ], + "type": "image", + "image_path": "19755bcc633b5ef5f2330fec81ab25631352b285c8cb374bf567f94378903fd7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 179, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 305, + 179, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 305, + 179, + 547, + 215 + ], + "type": "text", + "content": "Figure 4: Representative syllables and corresponding animations. The animation results of our Speech4Mesh and FaceFormer are displayed in the first and second columns." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 354, + 221, + 493, + 350 + ], + "blocks": [ + { + "bbox": [ + 354, + 221, + 493, + 350 + ], + "lines": [ + { + "bbox": [ + 354, + 221, + 493, + 350 + ], + "spans": [ + { + "bbox": [ + 354, + 221, + 493, + 350 + ], + "type": "image", + "image_path": "8753daa8ceb5377c021443de597afbfdc507cd40bf4bda7e7106ba3fcd768100.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 313, + 370, + 537, + 396 + ], + "blocks": [ + { + "bbox": [ + 309, + 350, + 541, + 363 + ], + "lines": [ + { + "bbox": [ + 309, + 350, + 541, + 363 + ], + "spans": [ + { + "bbox": [ + 309, + 350, + 541, + 363 + ], + "type": "text", + "content": "Figure 5: Speech4Mesh samples with different emotions." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 370, + 537, + 396 + ], + "lines": [ + { + "bbox": [ + 313, + 370, + 537, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 537, + 396 + ], + "type": "table", + "html": "
MetricFullwo. Pho.wo. Con.FaceFormer
Lip Vertex Error ↓4.863e-35.233e-35.271e-35.355e-3
", + "image_path": "8f5aa289bb75f7ecfc25db3378732960ed731dbe015ffe65e1f5388db345350f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 399, + 544, + 423 + ], + "lines": [ + { + "bbox": [ + 305, + 399, + 544, + 423 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 544, + 423 + ], + "type": "text", + "content": "Table 5: Ablation study on the photometric loss and contrastive loss w.r.t. the lip-sync error." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 428, + 476, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 476, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 476, + 441 + ], + "type": "text", + "content": "6. Conclusions and Future Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 449, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 629 + ], + "type": "text", + "content": "In this paper, we propose a novel framework named Speech4Mesh to train the audio2mesh network for speech-driven 3D facial animation. In Speech4Mesh, the speech-assisted monocular 3D face reconstruction module can provide significantly more accurate 4D synthesized data for pre-training by leveraging the audio-visual contrastive alignment. By employing the differentiable rendering loss and an emotion code, the modified audio2mesh network can achieve more realistic animation performance and unprecedentedly realize emotion manipulation. Empirical experiments verify the effectiveness and superiority of the Speech4Mesh framework both quantitatively and perceptually. We hope that the proposed Speech4Mesh framework can bring inspiration for similar 4D tasks to address the data scarcity problem." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "In the future, we plan to replace the image-based reconstruction backbone DECA with a video-based reconstruction backbone such as MICA [76], neural head avatar [26] or [67]. We will also leverage more high-quality 2D video data for better animation performance and devise a more effective emotion modeling scheme to improve the expressiveness of expressions." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14199" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems, 33:12449-12460, 2020. 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "text", + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187-194, 1999. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 245 + ], + "type": "text", + "content": "[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5543-5552, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 300 + ], + "type": "text", + "content": "[4] Carlos Busso, Zhigang Deng, Michael Grimm, Ulrich Neumann, and Shrikanth Narayanan. Rigid head motion in expressive speech animation: Analysis and synthesis. IEEE transactions on audio, speech, and language processing, 15(3):1075-1086, 2007. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 301, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 288, + 334 + ], + "type": "text", + "content": "[5] Chen Cao, Qiming Hou, and Kun Zhou. Displaced dynamic expression regression for real-time facial tracking and animation. ACM Trans. Graph, 33(4):1-10, 2014. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "text", + "content": "[6] Yong Cao, Wen C Tien, Petros Faloutsos, and Frédéric Pighin. Expressive speech-driven facial animation. ACM Transactions on Graphics (TOG), 24(4):1283-1302, 2005. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "type": "text", + "content": "[7] Yujin Chai, Yanlin Weng, Lvdi Wang, and Kun Zhou. Speech-driven facial animation with spectral gathering and temporal attention. Frontiers of Computer Science, 16(3):1-10, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 413, + 288, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 457 + ], + "type": "text", + "content": "[8] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520-535, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "type": "text", + "content": "[9] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7832-7841, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "text", + "content": "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. Proc. Interspeech 2018, pages 1086-1090, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 580 + ], + "type": "text", + "content": "[11] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian conference on computer vision, pages 251-263. Springer, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "text", + "content": "[12] Michael M Cohen, Rashid Clark, and Dominic W Massaro. Animated speech: Research progress and applications. In AVSP 2001-International Conference on Auditory-Visual Speech Processing, 2001. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 679 + ], + "type": "text", + "content": "[13] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10101-10111, 2019. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[14] Radek Daneček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In CVPR, pages 20311-20322, 2022. 3, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[15] Dipanjan Das, Sandika Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European conference on computer vision, pages 408-424. Springer, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 185 + ], + "type": "text", + "content": "[16] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multilevel face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 546, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 546, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 546, + 250 + ], + "type": "text", + "content": "[17] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 252, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 546, + 297 + ], + "type": "text", + "content": "[18] Pengfei Dou, Shishir K Shah, and Ioannis A Kakadiaris. End-to-end 3d face reconstruction with deep neural networks. In proceedings of the IEEE conference on computer vision and pattern recognition, pages 5908-5917, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 297, + 546, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 546, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 546, + 341 + ], + "type": "text", + "content": "[19] Pif Edwards, Chris Landreth, Eugene Fume, and Karan Singh. Jali: an animator-centric viseme model for expressive lip synchronization. ACM Trans. Graph, 35(4):1-11, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 342, + 546, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 546, + 398 + ], + "type": "text", + "content": "[20] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 546, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 546, + 443 + ], + "type": "text", + "content": "[21] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE Transactions on Multimedia, 12(6):591-598, 2010. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 444, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 546, + 487 + ], + "type": "text", + "content": "[22] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Trans. Graph., 40(4):1-13, 2021. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 489, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 546, + 544 + ], + "type": "text", + "content": "[23] Yao Feng, Fan Wu, Xiaohu Shao, Yanfeng Wang, and Xi Zhou. Joint 3d face reconstruction and dense alignment with position map regression network. In Proceedings of the European conference on computer vision, pages 534-551, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 545, + 546, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 546, + 601 + ], + "type": "text", + "content": "[24] Panagiotis P Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Visual speech-aware perceptual 3d facial expression reconstruction from videos. arXiv preprint arXiv:2207.11094, 2022.5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 601, + 546, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 546, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 546, + 658 + ], + "type": "text", + "content": "[25] Kyle Genova, Forrester Cole, Aaron Maschinot, Aaron Sarna, Daniel Vlasic, and William T Freeman. Unsupervised training for 3d morphable model regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8377-8386, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 714 + ], + "type": "text", + "content": "[26] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14200" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Jianzhu Guo, Xiangyu Zhu, Yang Yang, Fan Yang, Zhen Lei, and Stan Z Li. Towards fast, accurate and stable 3d dense face alignment. In European Conference on Computer Vision, pages 152-168. Springer, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "text", + "content": "[28] Yudong Guo, Jianfei Cai, Boyi Jiang, Jianmin Zheng, et al. CNN-based real-time dense face reconstruction with inverse-rendered photo-realistic face images. IEEE transactions on pattern analysis and machine intelligence, 41(6):1294-1307, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[29] Patrik Huber, Zhen-Hua Feng, William Christmas, Josef Kittler, and Matthias Ratsch. Fitting 3d morphable face models using local features. In 2015 IEEE international conference on image processing, pages 1195-1199. IEEE, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 282 + ], + "type": "text", + "content": "[30] Patrik Huber, Guosheng Hu, Rafael Tena, Pouria Mortazavian, P Koppen, William J Christmas, Matthias Ratsch, and Josef Kittler. A multiresolution 3d morphable face model and fitting framework. In Proceedings of the 11th international joint conference on computer vision, imaging and computer graphics theory and applications, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 284, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 337 + ], + "type": "text", + "content": "[31] Aaron S Jackson, Adrian Bulat, Vasileios Argyriou, and Georgios Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In Proceedings of the IEEE international conference on computer vision, pages 1031-1039, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "text", + "content": "[32] Amin Jourabloo and Xiaoming Liu. Large-pose face alignment via cnn-based dense 3d model fitting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4188–4196, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 435 + ], + "type": "text", + "content": "[33] Harim Jung, Myeong-Seok Oh, and Seong-Whan Lee. Learning free-form deformation for 3d face reconstruction from in-the-wild images. In 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 2737–2742. IEEE, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[34] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Trans. Graph., 36(4):1-12, 2017. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 514 + ], + "type": "text", + "content": "[35] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 515, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 558 + ], + "type": "text", + "content": "[36] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "type": "text", + "content": "[37] Feng Liu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Joint face alignment and 3d face reconstruction. In European Conference on Computer Vision, pages 545-560. Springer, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "text", + "content": "[38] Feng Liu, Ronghang Zhu, Dan Zeng, Qijun Zhao, and Xiaoming Liu. Disentangling features in 3d face shapes for joint face reconstruction and recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5216-5225, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[39] Jingying Liu, Binyuan Hui, Kun Li, Yunke Liu, Yu-Kun Lai, Yuxiang Zhang, Yebin Liu, and Jingyu Yang. Geometry-guided dense perspective network for speech-driven facial animation. IEEE Transactions on Visualization and Computer Graphics, 2021. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[40] Yaojie Liu, Amin Jourabloo, William Ren, and Xiaoming Liu. Dense face alignment. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1619-1628, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[41] Gaurav Mittal and Baoyuan Wang. Animating face using disentangled audio representations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3290–3298, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[43] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[44] Hai X Pham, Samuel Cheung, and Vladimir Pavlovic. Speech-driven 3d facial animation with implicit emotional awareness: a deep learning approach. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 80-88, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 332, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 384 + ], + "type": "text", + "content": "[45] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[46] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "text", + "content": "[47] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1173-1182, 2021. 1, 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "type": "text", + "content": "[48] Elad Richardson, Matan Sela, and Ron Kimmel. 3d face reconstruction by learning from synthetic data. In 2016 fourth international conference on 3D vision (3DV), pages 460-469. IEEE, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "type": "text", + "content": "[49] Elad Richardson, Matan Sela, Roy Or-El, and Ron Kimmel. Learning detailed face reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1259–1268, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "text", + "content": "[50] Sami Romdhani and Thomas Vetter. Estimating 3d shape and texture using pixel intensity, edges, specular highlights, texture constraints and a prior. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), volume 2, pages 986-993. IEEE, 2005. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[51] Shunsuke Saito, Tianye Li, and Hao Li. Real-time facial segmentation and performance capture from rgb input. In European conference on computer vision, pages 244–261. Springer, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[52] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael J Black. Learning to regress 3d face shape and expression from an image without 3d supervision. In Proceedings of" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14201" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7763-7772, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[53] Matan Sela, Elad Richardson, and Ron Kimmel. Unrestricted facial geometry reconstruction using image-to-image translation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1576-1585, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 194 + ], + "type": "text", + "content": "[54] Jiaxiang Shang, Tianwei Shen, Shiwei Li, Lei Zhou, Mingmin Zhen, Tian Fang, and Long Quan. Self-supervised monocular 3d face reconstruction by occlusion-aware multiview geometry consistency. In European Conference on Computer Vision, pages 53-70. Springer, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "type": "text", + "content": "[55] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learning lip sync from audio. ACM Trans. Graph., 36(4):1-13, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 287, + 293 + ], + "type": "text", + "content": "[56] Sarah Taylor, Taehwan Kim, Yisong Yue, Moshe Mahler, James Krahe, Anastasio Garcia Rodriguez, Jessica Hodgins, and Iain Matthews. A deep learning approach for generalized speech animation. ACM Transactions on Graphics (TOG), 36(4):1-11, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 295, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 295, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 49, + 295, + 287, + 338 + ], + "type": "text", + "content": "[57] Sarah L Taylor, Moshe Mahler, Barry-John Theobald, and Iain Matthews. Dynamic units of visual speech. In Proceedings of the 11th ACM SIGGRAPH/Eurographics conference on Computer Animation, pages 275-284, 2012. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 338, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 287, + 403 + ], + "type": "text", + "content": "[58] Ayush Tewari, Florian Bernard, Pablo Garrido, Gaurav Bharaj, Mohamed Elgharib, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Fml: Face model learning from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10812–10822, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 405, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 405, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 49, + 405, + 287, + 470 + ], + "type": "text", + "content": "[59] Ayush Tewari, Michael Zollhöfer, Pablo Garrido, Florian Bernard, Hyeongwoo Kim, Patrick Pérez, and Christian Theobalt. Self-supervised multi-level face model learning for monocular reconstruction at over " + }, + { + "bbox": [ + 49, + 405, + 287, + 470 + ], + "type": "inline_equation", + "content": "250\\mathrm{~hz}" + }, + { + "bbox": [ + 49, + 405, + 287, + 470 + ], + "type": "text", + "content": ". In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2549-2559, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 471, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 471, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 49, + 471, + 287, + 536 + ], + "type": "text", + "content": "[60] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 1274-1283, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 537, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 287, + 581 + ], + "type": "text", + "content": "[61] Balamurugan Thambiraja, Ikhsanul Habibie, Sadegh Aliakbarian, Darren Cosker, Christian Theobalt, and Justus Thies. Imitator: Personalized speech-driven 3d facial animation. arXiv preprint arXiv:2301.00023, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 582, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 582, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 582, + 287, + 635 + ], + "type": "text", + "content": "[62] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 637, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 637, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 49, + 637, + 287, + 690 + ], + "type": "text", + "content": "[63] Anh Tuan Tran, Tal Hassner, Iacopo Masi, and Gerard Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5163-5172, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[64] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 705 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In European Conference on Computer Vision, pages 700-717. Springer, 2020. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 139 + ], + "type": "text", + "content": "[65] Qianyun Wang, Zhenfeng Fan, and Shihong Xia. 3d-talkemo: Learning to synthesize 3d emotional talking head. arXiv preprint arXiv:2104.12051, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "text", + "content": "[66] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681-3691, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 197, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 260 + ], + "type": "text", + "content": "[67] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Matthew Johnson, Jingjing Shen, Nikola Milosavljevic, Daniel Wilde, Stephan Garbin, Toby Sharp, Ivan Stojiljkovic, et al. 3d face reconstruction with dense landmarks. In European Conference on Computer Vision, pages 160-177. Springer, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 262, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 296 + ], + "type": "text", + "content": "[68] Yuyu Xu, Andrew W Feng, Stacy Marsella, and Ari Shapiro. A practical and configurable lip sync method for games. In Proceedings of Motion on Games, pages 131-140, 2013. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 297, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 360 + ], + "type": "text", + "content": "[69] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 363, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 363, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 308, + 363, + 545, + 406 + ], + "type": "text", + "content": "[70] Ran Yi, Zipeng Ye, Juyong Zhang, Hujun Bao, and Yong-Jin Liu. Audio-driven talking face video generation with learning-based personalized head pose. arXiv preprint arXiv:2002.10137, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 407, + 545, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 407, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 308, + 407, + 545, + 452 + ], + "type": "text", + "content": "[71] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, pages 9299–9306, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 453, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 453, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 453, + 545, + 517 + ], + "type": "text", + "content": "[72] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 519, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 519, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 308, + 519, + 545, + 563 + ], + "type": "text", + "content": "[73] Yang Zhou, Zhan Xu, Chris Landreth, Evangelos Kalogerakis, Subhransu Maji, and Karan Singh. Visemenet: Audio-driven animator-centric speech animation. ACM Transactions on Graphics (TOG), 37(4):1-10, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 563, + 545, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 563, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 308, + 563, + 545, + 607 + ], + "type": "text", + "content": "[74] Xiangyu Zhu, Zhen Lei, Xiaoming Liu, Hailin Shi, and Stan Z Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 146-155, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 609, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 609, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 308, + 609, + 545, + 662 + ], + "type": "text", + "content": "[75] Xiangyu Zhu, Zhen Lei, Junjie Yan, Dong Yi, and Stan Z Li. High-fidelity pose and expression normalization for face recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 787-796, 2015. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 664, + 545, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 664, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 308, + 664, + 545, + 705 + ], + "type": "text", + "content": "[76] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In European Conference on Computer Vision, pages 250–269. Springer, 2022. 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14202" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/55278fbf-9750-4827-94bb-81181843441e_content_list.json b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/55278fbf-9750-4827-94bb-81181843441e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..73b7623fecf78f2903c7764854b0cd411f3e4c74 --- /dev/null +++ b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/55278fbf-9750-4827-94bb-81181843441e_content_list.json @@ -0,0 +1,2032 @@ +[ + { + "type": "text", + "text": "Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution", + "text_level": 1, + "bbox": [ + 76, + 130, + 893, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zixiang Zhao $^{1,2}$ Jiangshe Zhang $^{1*}$ Xiang Gu $^{1}$ Chengli Tan $^{1}$ Shuang Xu $^{3}$ Yulun Zhang $^{2}$ Radu Timofte $^{2,4}$ Luc Van Gool $^{2}$", + "bbox": [ + 227, + 179, + 743, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Xi'an Jiaotong University $^{2}$ Computer Vision Lab, ETH Zürich $^{3}$ Northwestern Polytechnical University $^{4}$ University of Würzburg", + "bbox": [ + 218, + 222, + 750, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zixiangzhao@stu.xjtu.edu.cn, jszhang@mail.xjtu.edu.cn", + "bbox": [ + 250, + 262, + 720, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 310, + 313, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guided depth map super-resolution (GDSR), as a hot topic in multi-modal image processing, aims to upsample low-resolution (LR) depth maps with additional information involved in high-resolution (HR) RGB images from the same scene. The critical step of this task is to effectively extract domain-shared and domain-private RGB/depth features. In addition, three detailed issues, namely blurry edges, noisy surfaces, and over-transferred RGB texture, need to be addressed. In this paper, we propose the Spherical Space feature Decomposition Network (SSDNet) to solve the above issues. To better model cross-modality features, Restormer block-based RGB/depth encoders are employed for extracting local-global features. Then, the extracted features are mapped to the spherical space to complete the separation of private features and the alignment of shared features. Shared features of RGB are fused with the depth features to complete the GDSR task. Subsequently, a spherical contrast refinement (SCR) module is proposed to further address the detail issues. Patches that are classified according to imperfect categories are input into the SCR module, where the patch features are pulled closer to the ground truth and pushed away from the corresponding imperfect samples in the spherical feature space via contrastive learning. Extensive experiments demonstrate that our method can achieve state-of-the-art results on four test datasets, as well as successfully generalize to real-world scenes. The code is available at https://github.com/Zhaozixiang1228/GDSR-SSDNet.", + "bbox": [ + 73, + 342, + 473, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 792, + 209, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth maps, as images that measure the distance of scene points from the sensor, are widely used in autonomous driving [38, 50, 87-89], pose estimation [75, 83], virtual reality [41, 51, 79, 80], and scene understanding [24, 53, 65, 96].", + "bbox": [ + 75, + 818, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/148f4c1fca23c465fc51ee743897246415b87a76156ba44c5d14189b8d474d86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 308, + 696, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdcaae00eaec76888c6faf45e90c5e001f1b0cedefb6a1ce7c34a6911bd7634d.jpg", + "image_caption": [ + "Figure 1: Our SSDNet achieves outstanding performance on the RGBDD dataset for $\\times 4$ , $\\times 8$ , and $\\times 16$ and Middlebury for $\\times 4$ while being computationally efficient." + ], + "image_footnote": [], + "bbox": [ + 500, + 426, + 694, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c97d9af4227f6916876cfdc7aa9d2c94c7e07289c49449ada4345855463cd9f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 308, + 890, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2cef7c8222f66e500997ef27de203d42c9ab3f59dd4bed2bec394d3c72bfc8d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 426, + 888, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the depth maps produced by current consumer-level depth sensors, e.g., Time-of-Flight (ToF) and Kinect cameras, often have the disadvantages of low resolution and noise. These disadvantages are insufficient to meet the requirements of advanced computing vision tasks [52, 76, 93].", + "bbox": [ + 496, + 611, + 893, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To obtain high-resolution depth maps, we naturally hope to accomplish depth map super-resolution (SR) by transferring mature SR models in the RGB image domain [13, 39, 86] to the depth SR task. However, a potential risk is that RGB SR models tend to focus on reconstructing high-frequency image information, such as details and texture. Conversely, for depth images, the objects' depth information is often textureless and piecewise, and more sensitive to unclear edges and noise. Therefore, it is unreasonable to directly apply RGB SR models to the depth SR task. On the other hand, while acquiring the depth map, it is relatively easy to obtain HR and noise-free RGB images in the same scene. Furthermore, there are statistical co-occurrences between the edges in RGB images and the discontinuities in", + "bbox": [ + 496, + 688, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12547", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "depth maps [55]. Therefore, we hope to use the HR RGB images to provide edges and contour information missing in the depth map, and to fuse the multi-modal information to accomplish the LR depth image upsampling.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the era of deep learning (DL), numerous methods have been utilized to learn the mapping between $\\mathrm{LR} \\rightarrow \\mathrm{HR}$ depth maps. These methods succeeded to some extent in modeling the cross-modality features and reconstructing the contour and edge information. However, DL models which rely on natural priors will limit the model's flexibility. Models learning the $\\mathrm{LR} \\rightarrow \\mathrm{HR}$ mapping via data-driven methods are difficult to interpret due to the unclear working mechanism [90]. Thus, effectively extracting and distinguishing domain-shared and domain-private RGB/depth features is still a challenge. While microscopically, the obtained depth images are often plagued by three detail issues: blurry edges, noisy surfaces, and over-transferred RGB texture, all of which affect the display quality of depth maps.", + "bbox": [ + 75, + 154, + 470, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to addressing the above challenges, we expect to limit the solution space by constraining the extracted features and further improve the modeling of the dependencies between different modalities. Based on our observation, RGB images and depth maps contain shared features, such as depth map discontinuities and edge features of RGB objects, which can be aligned in the feature space. In addition, unique private features, such as the distance information of the depth map and the texture of RGB object surface, should be separated conversely. Thus, in the feature space, domain-shared and domain-private features are expected to separate and align respectively, while the distance between HR features and imperfect features with the above-mentioned detail issues should also be pushed away.", + "bbox": [ + 75, + 369, + 470, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above goal can be divided into two steps. First, extracting features through an effective encoder, and then selecting an appropriate distance measure to complete the alignment and separation of features. Considering that CNN-based architectures limit feature extraction capabilities due to content-independent convolution kernels and the lack of global information, we employ the Restormer blocks [84], which has been proved to effectively extract features in the low-level vision domain, as the basic unit of our encoder. For the choice of distance measure, we first think of the Euclidean distance, such as $\\ell_1$ or $\\ell_2$ distance. However, cross-modality features often contain different scales and orders of magnitude, and the Euclidean distance is easily affected by the scale, which makes it difficult to achieve our goal. Recently, with the development of spherical DL models [4, 7, 21, 22], the spherical feature learning is well known by virtue of its advantages over the Euclidean feature learning in many applications, e.g., domain adaptation [21]. Due to the distances between spherical features being regularized, the alignment and separation of features can be done more easily without losing the model's representational", + "bbox": [ + 75, + 583, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "capacity.", + "bbox": [ + 498, + 92, + 562, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "According to the above analysis, we propose our Spherical Space feature Decomposition Network (SSDNet). The specific workflow is displayed in Fig. 2. Our contribution can be summarized in three-fold:", + "bbox": [ + 496, + 108, + 893, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a spherical space feature decomposition framework to model the cross-modality features. The features extracted by the Restormer block are mapped to the spherical space for separation and alignment. This is the first time that the Transformer and spherical-space distance measure are applied to the GDSR task.", + "- The spherical contrast refinement module, cooperating with the imperfect patch classification and the corresponding contrastive learning branch, is proposed to address the possible detail issues in depth maps. This is also the first time that the contrastive learning technique has been used for the GDSR task.", + "- Experiments on four GDSR benchmarks and a real scene dataset demonstrate that our method can generate satisfactory HR depth maps in different scenarios and exhibits good generalization ability." + ], + "bbox": [ + 516, + 169, + 890, + 410 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 428, + 640, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we briefly introduce the GDSR task, and illustrate the Vision Transformer, spherical space DL and contrastive learning techniques utilized in SSDNet.", + "bbox": [ + 498, + 454, + 890, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. GDSR methods", + "text_level": 1, + "bbox": [ + 500, + 513, + 656, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image super-resolution is a popular image processing and computer vision task with too many sub-categories, so we only discuss the GDSR task here. Conventional GDSR methods can be divided into local- [40, 43, 44, 46], global- [12, 17, 36, 48, 77] and learning-based [8, 19, 20, 73, 74] methods, etc. They extract cross-modality information by manually designed filters, optimizing equations, or sparse dictionary learning. With the rapid development of DL, GDSR is further promoted by the CNN-based methods [58, 60–62, 68, 81]. Generally, DL-based methods can be categorized in three groups, i.e., auto-encoder (AE) [23, 30, 35, 78, 90, 92], learnable filter [14, 31, 71], algorithm unfolding [10, 11, 54, 94] and unsupervised [9, 15] methods. AE-based methods learn cross-modality features via shared or private encoders and reconstruct HR depth maps with decoders. In the learnable filter group, the convolution kernels used to extract information are set to be context-dependent and spatially-variant, which improves the flexibility of the model. The algorithmic unfolding group establishes interpretable GDSR models by building a bridge between traditional optimization functions and deep learning. Unsupervised methods change the training paradigm that requires paired LR/HR images, thus solving the difficulty of obtaining paired training data.", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12548", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Spherical deep learning", + "text_level": 1, + "bbox": [ + 76, + 90, + 295, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The works [4, 6, 7] introduce the geometrical neural network or the spherical convolution neural networks for analyzing the spherical signals that are rotation invariant. Gu et al. [22] propose the spherical multi-layer perceptron for aligning feature distributions of different domains in spherical feature space for domain adaptation. Our idea of spherical space feature decomposition is inspired by Gu et al. [22] that shows the advantages of spherical feature learning over the Euclidean feature learning for domain adaptation in the \"high-level\" image classification task. Different from [22], we tackle the \"low-level\" GDSR task by aligning/separating the shared/private features of paired RGB image and depth map, rather than aligning the feature distributions as in [22]. We further propose the spherical contrast refinement for tackling the issues of blurry edges, noisy surfaces, and overtransferred RGB texture for GDSR.", + "bbox": [ + 75, + 113, + 472, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Transformer in vision", + "text_level": 1, + "bbox": [ + 76, + 367, + 281, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer, proposed by Vaswani et al. [64], has now become a popular technology in computer vision. It has achieved dominance in classification [16, 42, 63], object detection [2, 95], segmentation [66, 91], etc. Simultaneously, it has also made progress in low-level vision [1, 18, 33, 67]. Chen et al. [5] proposed IPT based on the standard Transformer and multi-task learning. Liang et al. [37] proposed SwinIR whose core module inherits from Swin Transformer [42]. Recently, Restormer [84] improves the transformer block by the gated-Dconv network and multi-Dconv head transposed attention. By applying self-attention in the feature dimension, the ability of local-global representation learning is maintained while being more friendly to high-resolution input images.", + "bbox": [ + 75, + 391, + 472, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Contrastive Learning", + "text_level": 1, + "bbox": [ + 76, + 614, + 279, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive learning, which aims to address the scarcity of source and target paired data and prioritize self-supervised representation learning [25, 28]. The primary objective is to bring samples close to their positive counterparts and push them away from negative ones within high-dimensional manifolds, and it seeks to increase the distance between samples belonging to different classes. Recently, contrastive learning has gained significant attention in computer vision, particularly in high-level tasks such as object detection [72], video segmentation [3], and object tracking [82]. It has also found applications in low-level tasks such as super-resolution [85], image translation [49], and image dehazing [70].", + "bbox": [ + 75, + 638, + 472, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.5. Comparison with existing approaches", + "text_level": 1, + "bbox": [ + 76, + 830, + 403, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The methods most related to our approach are feature-decomposition DL methods [10, 90], which propose the idea that cross-modality features contain the common and", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "private information. In contrast, our method, for the first time, exploits the Transformer model's powerful ability to model globally dependent features. Meanwhile, the distance measure in spherical space is proved to decompose the cross-modality features more effectively than Euclidean distances. Additionally, instead of only focusing on optimizing the $\\ell_2$ loss between the reconstructed image and ground truth, we give classification-based fine-tuning for image patches under different imperfect conditions, thus further improving the performance of GDSR.", + "bbox": [ + 496, + 90, + 893, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 258, + 591, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first present the specific architecture of our SSDNet. Then, comprehensive descriptions are provided for each module. Finally, the loss function and training details will also be addressed.", + "bbox": [ + 496, + 285, + 893, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 357, + 609, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We can define the input LR depth map, RGB image and the HR depth map as $D_{LR} \\in \\mathbb{R}^{h \\times w}$ , $D_{HR} \\in \\mathbb{R}^{H \\times W}$ and $R \\in \\mathbb{R}^{H \\times W \\times 3}$ , where $\\{H, W\\}$ and $\\{h, w\\}$ are the height and width of input RGB and depth images, respectively. For our SSDNet network, it consists of five modules, namely, the Encoders for RGB and Depth images, the Decoders for RGB and Depth images, and the Spherical Contrast Refinement (SCR) module, which is denoted as $\\mathcal{E}_{\\mathcal{R}}(\\cdot), \\mathcal{E}_{\\mathcal{D}}(\\cdot), \\mathcal{D}_{\\mathcal{R}}(\\cdot), \\mathcal{D}_{\\mathcal{D}}(\\cdot),$ and $S(\\cdot)$ , respectively.", + "bbox": [ + 496, + 381, + 893, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In general, as shown in Fig. 2, SSDNet utilizes $\\mathcal{E}_{\\mathcal{R}}$ and $\\mathcal{E}_{\\mathcal{D}}$ to extract features, which are then projected onto the spherical space for feature separation and alignment. Then, the depth features and shared RGB features are input into $\\mathcal{D}_{\\mathcal{R}}$ to obtain reconstructed depth maps, and RGB features are fed into $\\mathcal{E}_{\\mathcal{R}}$ to get reconstructed RGB images. Finally, the reconstructed depth maps are subsequently provided to the SCR module to refine details and output the final $D_{HR}$ .", + "bbox": [ + 496, + 518, + 893, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The basic unit we employed for feature extraction and image restoration is the Restormer block [84]. The reason we opted for the Restormer block in $\\mathcal{E}_{\\mathcal{R}}$ , $\\mathcal{E}_{\\mathcal{D}}$ , $\\mathcal{D}_{\\mathcal{R}}$ and $\\mathcal{D}_{\\mathcal{D}}$ is because it allows for the extraction of global features from high-resolution input images by utilizing self-attention across feature dimensions [84]. This approach enables the extraction of cross-modality shallow features without significantly increasing computational requirements. For details on the Restormer block architecture, please refer to the supplementary material or the original paper [84].", + "bbox": [ + 496, + 640, + 895, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Spherical space transform", + "text_level": 1, + "bbox": [ + 500, + 815, + 736, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first define the mapping based on Riemannian geometry between Euclidean Space feature maps and Spherical Space feature maps, i.e., $\\mathcal{L}\\mathcal{O}\\mathcal{G}(\\cdot)$ and $\\mathcal{E}\\mathcal{X}\\mathcal{P}(\\cdot)$ , which are employed in the feature decomposition. A schematic diagram for the", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12549", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f5833979bc14144232c832eca99dd6e62464e5feab18daf25a463c5c68bc1565.jpg", + "image_caption": [ + "(a) Workflow of SSDNet", + "(b) Spherical Space Feature Decomposition Module" + ], + "image_footnote": [], + "bbox": [ + 83, + 107, + 612, + 286 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ca3fd5ba5dbe8dcd1c44f364fcd0387216831fd0b40c2fc40d0a4249057d8260.jpg", + "image_caption": [ + "(c) Spherical Contrast Refinement Module", + "Figure 2: (a) The Architecture of our SSDNet method. First, the LR depth map is input into $\\mathcal{E}_{\\mathcal{D}}$ , and the RGB image is input into $\\mathcal{E}_{\\mathcal{R}}$ . The intermediate feature maps are projected onto the spherical space to complete the feature decomposition, i.e., alignment and separation. The shared & private features of the depth map, and the shared features of the RGB image are input into the $\\mathcal{D}_{\\mathcal{D}}$ to reconstruct HR depth map. The RGB features are fed into $\\mathcal{D}_{\\mathcal{R}}$ to reconstruct RGB image. (b) Illustration of spherical space feature decomposition. (c) The module for spherical contrast refinement. (d) The mapping between Euclidean Space and Spherical Space." + ], + "image_footnote": [ + "C: Channel Concatenation", + "$\\mathcal{S}\\{\\cdot ,\\cdot \\} :$ Spherical-space Distance" + ], + "bbox": [ + 83, + 301, + 602, + 513 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e8398cdf5386c8604b0a4cfb2ad9e40e4423b2574deaf82aca19f97d8f1fbd88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 113, + 879, + 351 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b345c9106e9dbe7e98f34cb91e06025a06c6d18592f26fc7168fb3de5976d52b.jpg", + "image_caption": [ + "(d) Illustration of Spherical transform" + ], + "image_footnote": [ + "ATTN: Attention (Feature Extraction)" + ], + "bbox": [ + 620, + 375, + 864, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mappings is shown in Fig. 2d. The distance measure on the spherical space is also defined.", + "bbox": [ + 75, + 619, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1 (Spherical exponential mapping). Given a vector $v$ in $d$ -dimensional Euclidean Space, we define the $(d + 1)$ -dimensional vector $\\bar{v}$ by $\\bar{v} = (v,r)$ where $r$ is a hyperparameter named radius. Then the spherical exponential transform [69] $\\exp_N: T_N\\mathbb{S}_r^{d + 1} \\to \\mathbb{S}_r^{d + 1}$ is defined as", + "bbox": [ + 75, + 657, + 472, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\exp_ {N} (v) = N \\cos \\theta + \\bar {v} \\frac {\\sin \\theta}{\\theta}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 742, + 468, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{S}_r^{d + 1} = \\{x\\in \\mathbb{R}^{d + 1}:||x|| = r\\} \\subset \\mathbb{R}^{d + 1}$ is the $d$ -dimensional spherical space, $N = (0,\\dots ,0,r)\\in \\mathbb{S}_r^{d + 1}$ is the north pole, $\\theta = \\frac{\\|\\vec{v}\\|}{r}$ , and $T_{N}\\mathbb{S}_{r}^{d + 1} = \\{(v,r):v\\in$ $\\mathbb{R}^d\\}$ is the tangent space of $\\mathbb{S}_r^{d + 1}$ . Further, the spherical exponential mapping $\\mathcal{E}\\mathcal{X}\\mathcal{P}:\\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times d}\\to \\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times (d + 1)}$ for an Euclidean feature map $\\Phi$ is defined as", + "bbox": [ + 76, + 780, + 470, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E X P} (\\Phi) [ i, j,: ] = \\exp_ {N} (\\Phi [ i, j,: ]) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Phi [i,j,:]$ is the feature in location $(i,j)$ .", + "bbox": [ + 500, + 618, + 803, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 2 (Spherical logarithmic mapping). Given the spherical feature $x \\in \\mathbb{S}_r^{d+1}$ with $\\|x\\| = r$ , we define the spherical logarithmic transform [69] $\\log_N: \\mathbb{S}_r^{d+1} \\to T_N \\mathbb{S}_r^{d+1}$ by", + "bbox": [ + 498, + 642, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\log_ {N} (x) = \\frac {\\psi}{\\sin \\psi} (x - N \\cos \\psi), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 712, + 890, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N = (0,\\dots ,0,r)\\in \\mathbb{S}_r^{d + 1}$ is the north pole, $\\psi =$ $\\arg \\cos (N^T x / r^2)$ . Further, the spherical logarithmic mapping $\\mathcal{LOG}:\\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times (d + 1)}\\to \\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times d}$ for a spherical feature map $\\Phi$ is defined by", + "bbox": [ + 498, + 750, + 893, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L O G} (\\Phi) [ i, j,: ] = \\mathcal {H} \\left(\\log_ {N} \\left(\\Phi [ i, j,: ]\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 821, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{H}:T_N\\mathbb{S}_r^{d + 1}\\to \\mathbb{R}^d$ is defined by $\\mathcal{H}((v,r)) = v$", + "bbox": [ + 498, + 845, + 862, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 3 (Spherical space distance). Given two spherical feature maps $\\Phi_1, \\Phi_2 \\in \\mathbb{R}^{h' \\times w' \\times (d + 1)}$ with $\\| \\Phi_1[i,j,:] \\| =$", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12550", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\| \\Phi_2[i,j,:]\\| = r$ for any $i,j$ , the spherical space distance of $\\Phi_{1}$ and $\\Phi_{2}$ is defined as", + "bbox": [ + 76, + 90, + 472, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} \\left\\{\\Phi_ {1}, \\Phi_ {2} \\right\\} = \\sum_ {i = 1} ^ {h ^ {\\prime}} \\sum_ {j = 1} ^ {w ^ {\\prime}} 1 - \\frac {1}{r ^ {2}} \\Phi_ {1} [ i, j,: ] ^ {T} \\Phi_ {2} [ i, j,: ]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 127, + 468, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.2 Encoder", + "text_level": 1, + "bbox": [ + 76, + 186, + 191, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the feature extraction of $D_{LR}$ as an example to explain the procedure, and $R$ can be carried out similarly to $D_{LR}$ by changing the subscripts from $D$ to $R$ . First, a $3 \\times 3$ convolution is used to obtain shallow features embedding $\\Phi_D^{(0)}$ . The main body of feature extraction consists of a cascade of $P$ Restormer blocks, and we denote the $p$ -th Restormer block in $\\mathcal{E}_D$ as $\\mathcal{R}_D^{(p)}$ , where $p = 1, \\dots, P$ . The input of each $\\mathcal{R}_D^{(p)}$ is represented by $\\Phi_D^{(p-1)}$ .", + "bbox": [ + 75, + 209, + 468, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the $p$ -th step of feature extraction, $\\Phi_D^{(p-1)}$ passes through $\\mathcal{R}_D^{(p)}$ to obtain the preliminary extracted feature $\\tilde{\\Phi}_D^{(p)}$ . According to the analysis for motivation in Sec. 1, after the multi-head self-attention mechanism, $\\tilde{\\Phi}_D^{(p)}$ should contain shared features in some channels and private features in other channels. Thus, we assume that features in the first $\\frac{dim}{2}$ channels are shared and represent cross-modality information, while features in the latter $\\frac{dim}{2}$ channels are private and represent the characteristics of their own modality. To achieve feature separation and alignment, $\\tilde{\\Phi}_D^{(p)}$ is mapped to the spherical space using the spherical exponential mapping $\\mathcal{E}\\mathcal{X}\\mathcal{P}(\\cdot)$ , and then recovered in the feature domain by the spherical logarithmic mapping $\\mathcal{L}\\mathcal{O}\\mathcal{G}(\\cdot)$ after calculating the feature decomposition loss. Finally, the recovered features are re-concatenated along the channel dimension to obtain $\\Phi_D^{(p)}$ , which will be input to the next $\\mathcal{R}_D^{(p+1)}$ . The feature decomposition loss will be illustrated in subsequent sections. The total feature extraction process of $p$ -th step is:", + "bbox": [ + 75, + 338, + 472, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\Phi} _ {D} ^ {(p)} = \\mathcal {R} _ {D} ^ {(p)} \\left(\\Phi_ {D} ^ {(p - 1)}\\right) \\tag {6a}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 618, + 468, + 641 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\Phi} _ {D, (p)} ^ {a l i g n} = \\tilde {\\Phi} _ {D} ^ {(p)} \\left[ 0: \\frac {d i m}{2} \\right], \\tilde {\\Phi} _ {D, (p)} ^ {s e p n} = \\tilde {\\Phi} _ {D} ^ {(p)} \\left[ \\frac {d i m}{2}: d i m \\right] \\quad (6 b)\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 643, + 468, + 664 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {D, (p)} ^ {\\text {a l i g n}} = \\mathcal {L O G} \\left(\\mathcal {E X P} \\left(\\tilde {\\Phi} _ {D, (p)} ^ {\\text {a l i g n}}\\right)\\right) \\tag {6c}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 665, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {D, (p)} ^ {s e p n} = \\mathcal {L O G} \\left(\\mathcal {E X P} \\left(\\tilde {\\Phi} _ {D, (p)} ^ {s e p n}\\right)\\right) \\tag {6d}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 690, + 468, + 714 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {D} ^ {(p)} = \\operatorname {C a t} \\left(\\Phi_ {D, (p)} ^ {\\text {a l i g n}}, \\Phi_ {D, (p)} ^ {\\text {s e p n}}\\right) \\tag {6e}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 715, + 468, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\{\\tilde{\\Phi}_{D,(p)}^{\\text{align}}, \\tilde{\\Phi}_{D,(p)}^{\\text{sep}}\\}$ and $\\{\\Phi_{D,(p)}^{\\text{align}}, \\Phi_{D,(p)}^{\\text{sep}}\\}$ are the aligned and separated features after/before calculating the feature decomposition loss, and $Cat(\\cdot, \\cdot)$ is the channel concatenation operator.", + "bbox": [ + 75, + 744, + 468, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eventually, the entire encoder feature extraction can be regarded as:", + "bbox": [ + 76, + 806, + 468, + 837 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {D} = \\mathcal {E} _ {D} (D _ {L R}), \\Phi_ {R} = \\mathcal {E} _ {R} (R), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 844, + 468, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_D$ and $\\Phi_R$ are abbreviated forms of $\\Phi_D^{(P)}$ and $\\Phi_D^{(P)}$ . Then $\\Phi_D$ and $\\Phi_R$ will be used as the input to the decoder.", + "bbox": [ + 75, + 866, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.3 Decoder", + "text_level": 1, + "bbox": [ + 500, + 90, + 612, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to Sec. 3.1.2, we obtained features $\\Phi_D$ and $\\Phi_R$ , each containing shared information across $\\frac{dim}{2}$ channels and private information across the other $\\frac{dim}{2}$ channels. The features that we consider helpful for the HR depth reconstruction task are the full depth features $\\Phi_D$ and the shared RGB features $\\Phi_R^{sep_n}$ . Therefore, we concatenate $\\Phi_D$ and $\\Phi_R^{sep_n}$ in channel dimension and input them to $\\mathcal{D}_{\\mathcal{D}}$ to obtain the reconstructed depth image $\\hat{D}_{HR}$ , and input $\\Phi_R$ to $\\mathcal{D}_{\\mathcal{R}}$ to obtain the reconstructed RGB image $\\hat{R}$ , which is:", + "bbox": [ + 496, + 114, + 893, + 253 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} _ {H R} = \\mathcal {D} _ {\\mathcal {D}} \\left(C a t \\left(\\Phi_ {D}, \\Phi_ {R} [ 0: \\frac {d i m}{2} ]\\right)\\right), \\hat {R} = \\mathcal {D} _ {\\mathcal {R}} \\left(\\Phi_ {R}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 258, + 890, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.4 Spherical Contrast Refinement module", + "text_level": 1, + "bbox": [ + 500, + 301, + 825, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After decoding, we have obtained a preliminary estimation of the HR depth map $\\hat{D}_{HR}$ . However, it potentially has some minor issues, such as blurry edges, noisy surfaces and overtransferred RGB texture. Hence, in the Spherical Contrast Refinement (SCR) module, we target the imperfect issues in $\\hat{D}_{HR}$ by contrastive learning, and the diagram is shown in Fig. 2c.", + "bbox": [ + 496, + 325, + 893, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Defect patches classifier. Firstly, we artificially synthesize the \"imperfect image dataset\" using the training set. For example, for an $m \\times m$ image patch from the ground truth in the training set, we can add random noise to make the patch noisy, apply Gaussian blur to it blurry, and add the same-location RGB image to make it texture over-transferred. Patches that are not processed can be regarded as perfect. In this way, we obtain a dataset with labels \"noisy\", \"blurry\", \"texture over-transferred\" and \"perfect\" which can be used to train a defect patches classifier (DPC) based on ResNet34 [26].", + "bbox": [ + 496, + 434, + 893, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Positive and negative samples. After obtaining the decoder output $\\hat{D}_{HR}$ in Eq. (8), we randomly crop it into an $m \\times m$ patch $\\hat{D}_{HR}^{pat}$ , and input it into the well-trained DPC to obtain the imperfect type of $\\hat{D}_{HR}^{pat}$ . After getting the imperfect label, similar to the operation for making the \"imperfect image dataset\", we can transform the corresponding ground truth $D_{HR}$ into an imperfect depth map $\\tilde{D}_{HR}$ . Then, we randomly crop $N$ $m \\times m$ patches from $\\tilde{D}_{HR}$ that are different from the position of $\\hat{D}_{HR}^{pat}$ , and the corresponding imperfect patches $\\tilde{D}_{HR}^{pat}$ , named negative samples, are generated. Meanwhile, we refer to $\\hat{D}_{HR}^{pat}$ as the anchor and the same-location ground truth $D_{HR}^{pat}$ as the positive sample. The $k$ -th positive, anchor and negative samples are represented by $\\mu_k^+$ , $\\mu_k$ , and $\\mu_k^-$ , respectively.", + "bbox": [ + 496, + 604, + 893, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spherical contrast refinement. We input $\\mu_k^+$ , $\\mu_k$ , and $\\mu_k^-$ into $\\mathcal{E}_D$ and get the contrastive refinement loss $\\mathcal{L}_{SCR}$ :", + "bbox": [ + 496, + 825, + 890, + 856 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S C R} = \\sum_ {k = 1} ^ {K} \\frac {\\mathcal {S} \\left\\{\\mathcal {E} _ {D} \\left(\\mu_ {k}\\right) , \\mathcal {E} _ {D} \\left(\\mu_ {k} ^ {+}\\right) \\right\\}}{\\sum_ {n = 1} ^ {N} \\mathcal {S} \\left\\{\\mathcal {E} _ {D} \\left(\\mu_ {k}\\right) , \\mathcal {E} _ {D} \\left(\\mu_ {k} ^ {n -}\\right) \\right\\}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 861, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12551", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3a13cc767acbe83ee5529d7767d0b296829910fd3439594647d2d42af944cee.jpg", + "image_caption": [ + "Figure 3: Exhibition for different settings of SSDNet. Results are evaluated on validation set of NYU V2 for GDSR $\\times 8$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 351, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9662e3b9c28fe80fcb1a999ea16649d5d176e23de01beb25d6bbda70a028650.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 85, + 616, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be66b245e1187c21f7757e9912351079789181d9f17f400cca992173c5e1dbf5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 85, + 880, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76331366bf2358a1cffb1af10dfd18394985fcab9c233436dfc33af1122bd5f4.jpg", + "image_caption": [ + "Figure 4: Visualization of the aligned and separated features." + ], + "image_footnote": [], + "bbox": [ + 158, + 258, + 815, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{S}\\{\\cdot ,\\cdot \\}$ is the spherical space distance in Definition 3, and $\\mu_{k}^{n - }$ is the nth negative sample in $\\mu_k^-$ . Through gradient descent, we minimize the distance between $\\mu_{k}^{+}$ and $\\mu_{k}$ features and maximize that between $\\mu_{k}$ and $\\mu_{k}^{-}$ features. This process fine-tunes $\\mathcal{E}_D$ . However, since the increased training cost for the SCR module, we incorporate it into the regular network training process every few iterations to strike a balance between training efficiency and effectiveness.", + "bbox": [ + 75, + 484, + 470, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Training loss", + "text_level": 1, + "bbox": [ + 76, + 616, + 215, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The training loss in this paper comprises several components: the depth map reconstruction loss $\\mathcal{L}_{pixel}^{D}$ , the RGB reconstruction loss $\\mathcal{L}_{pixel}^{R}$ , the feature decomposition loss $\\mathcal{L}_{dec}$ , and the spherical contrastive refinement loss $\\mathcal{L}_{SCR}$ . We describe each loss separately next.", + "bbox": [ + 75, + 641, + 470, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{pixel}^{D}$ ensures that the estimated depth map $\\hat{D}_{HR}$ output by our SSDNet is close to the ground truth depth map $D_{HR}$ . $\\mathcal{L}_{pixel}^{R}$ ensures that the output $\\hat{R}$ is close to the input $R$ . Although theoretically unrelated to the depth map superresolution task, this loss item is used to guarantee that the semantic information from the RGB image is involved in the shared RGB features, rather than simply generating random noise that is approximate to the depth features to meet the feature decomposition. Specifically,", + "bbox": [ + 75, + 718, + 470, + 854 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p i x e l}} ^ {D} = \\sum_ {k = 1} ^ {K} \\| \\hat {D} _ {H R} ^ {(k)} - D _ {H R} ^ {(k)} \\| _ {2} ^ {2}, \\mathcal {L} _ {\\text {p i x e l}} ^ {R} = \\sum_ {k = 1} ^ {K} \\| \\hat {R} ^ {(k)} - R ^ {(k)} \\| _ {2} ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 866, + 468, + 904 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding the feature decomposition loss $\\mathcal{L}_{dec}$ , we utilize the spherical space distance to enhance the similarity between shared features while reducing the similarity between separated features. We define the specific structure of the feature decomposition loss $\\mathcal{L}_{dec}$ as follows:", + "bbox": [ + 496, + 484, + 893, + 560 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e c} = \\mathcal {L} _ {a l i g n} - \\left(1 - \\mathcal {L} _ {s e p n}\\right) ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 570, + 890, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 500, + 599, + 544, + 611 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e p n}} = \\sum_ {p = 1} ^ {P} \\mathcal {S} \\left\\{\\Phi_ {D, (p)} ^ {\\text {s e p n}}, \\Phi_ {R, (p)} ^ {\\text {s e p n}} \\right\\}, \\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {p = 1} ^ {P} \\mathcal {S} \\left\\{\\Phi_ {D, (p)} ^ {\\text {a l i g n}}, \\Phi_ {R, (p)} ^ {\\text {a l i g n}} \\right\\}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 619, + 898, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike $\\ell_2$ distance, the spherical distance captures relative differences without being influenced by scale. For 1D feature maps with three consecutive pixels: $f_{1} = [0.4,0.5,0.6]$ , $f_{2} = [4,5,6]$ , $f_{3} = [0.6,0.5,0.4]$ , $f_{1}$ and $f_{2}$ exhibit similar pixel-wise increase structures and potentially similar extracted features, while $f_{3}$ differs. Thus, $\\{f_1,f_2\\}$ should be closer in distance and aligned, while $\\{f_1,f_3\\}$ should be separated. However, $\\ell_2\\{f_1,f_2\\} = 7.90$ , $\\ell_2\\{f_1,f_3\\} = 0.28$ while $\\mathcal{S}\\{f_1,f_2\\} = 7.7\\times 10^{-8}$ , $\\mathcal{S}\\{f_1,f_3\\} = 0.02$ . Therefore, the spherical distance is a better choice.", + "bbox": [ + 496, + 672, + 893, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When optimizing $\\mathcal{L}_{dec}$ , since the value range of spherical distance $S\\{\\cdot, \\cdot\\}$ is $[0,2]$ , we have $\\mathcal{L}_{align} \\to 0$ , $\\mathcal{L}_{sepn} \\to 1$ . According to Definition 3, the derivation gives: $\\mathcal{C}\\{\\Phi_{D,(p)}^{align}, \\Phi_{R,(p)}^{align}\\} \\to 1$ and $\\mathcal{C}\\{\\Phi_{D,(p)}^{sepn}, \\Phi_{R,(p)}^{sepn}\\} \\to 0$ , where $\\mathcal{C}$ is the cosine similarity (the last term of Eq. (5)). When", + "bbox": [ + 496, + 823, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12552", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53fdec5ccd7c452001ed4feae975686092e63999dad86a08f65b1fc7286a3f3a.jpg", + "image_caption": [ + "Figure 5: Visual error maps for test image in the Middlebury dataset for $8 \\times$ super-resolution." + ], + "image_footnote": [], + "bbox": [ + 88, + 87, + 883, + 318 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a254881d0939d8a4ba38d74da3fbeeea4a6eca5322d6ff349a5f72df2ecdf33f.jpg", + "image_caption": [ + "Figure 6: Visual error maps for test image in the RGBDD dataset for $16\\times$ super-resolution." + ], + "image_footnote": [], + "bbox": [ + 86, + 349, + 883, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathcal{C} = 1$ , vectors have an angle of 0, indicating higher similarity. If $\\mathcal{C} = 0$ , the vectors are orthogonal, signifying no correlation. By forcing $\\mathcal{C} = 1$ or $\\mathcal{C} = 0$ for aligned/private features mapped to spherical space, we ensure the enhancement/reduction of similarity, which achieves our goal of feature decomposition.", + "bbox": [ + 75, + 613, + 470, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By combining Eqs. (9) to (11), we obtain the total loss used for training, which can be expressed as follows:", + "bbox": [ + 75, + 689, + 468, + 720 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p i x e l}} ^ {D} + \\alpha_ {1} \\mathcal {L} _ {\\text {p i x e l}} ^ {R} + \\alpha_ {2} \\mathcal {L} _ {\\text {d e c}} + \\alpha_ {3} \\mathcal {L} _ {\\text {S C R}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 733, + 468, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that $\\alpha_{3} = 0$ if the current epoch does not require the spherical contrast refinement.", + "bbox": [ + 75, + 765, + 468, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 813, + 200, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section will conduct a comprehensive set of experiments that aim to showcase the effectiveness of our model in addressing the GDSR task. Moreover, we will also provide evidence that substantiates the soundness of our SSDNet.", + "bbox": [ + 75, + 839, + 470, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 612, + 581, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. Our experiments follow the protocol established in [27, 31, 90]. To evaluate our model, we employ four popular benchmarks: Middlebury [29, 56], Lu [45], NYU v2 [57], and RGBDD [27]. Our training and validation sets consist of the first 1000 images of NYU v2 dataset [57], divided into a 9:1 ratio. The last 449 pairs in NYU v2 are utilized as the test dataset. Middlebury [29, 56] (30 pairs), Lu [45] (6 pairs), and RGBDD [27] (405 pairs) are also used as test sets to evaluate the depth map SR ability across different scenes and objects. To furthermore demonstrate the generalization ability of our model to unknown scenarios, we incorporate thereal-world branch of the RGBDD dataset, which comprises 2215/405 pairs of RGB-D images for training/test sets, respectively. Lastly, we use the root-mean-square error (RMSE) metric to evaluate the super-resolution effect of our model.", + "bbox": [ + 496, + 638, + 893, + 880 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. In our experiments, we apply bicu", + "bbox": [ + 500, + 885, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12553", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/583ed6f19d4c5632bc9fbf45adf3a48dbac7acd305b808948ecd71b4daacdfcd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsMiddlebury [29, 56]NYU V2 [57]Lu [45]RGBDD [27]
×4×8×16×4×8×16×4×8×16×4×8×16
PAC [59]1.322.624.581.893.336.781.202.335.191.251.983.49
CUNet [11]1.102.174.331.923.706.780.912.234.991.181.953.45
DKN [31]1.232.124.241.623.266.510.962.165.111.301.963.42
FDKN [31]1.082.174.501.863.586.960.822.105.051.181.913.41
FDSR [27]1.132.084.391.613.185.861.292.195.001.161.823.06
GraphSR [8]1.112.124.431.793.176.020.922.055.151.301.833.12
DCTNet [90]1.102.054.191.593.165.840.881.854.391.081.743.05
Ours1.021.914.021.603.145.860.801.824.771.041.722.92
", + "bbox": [ + 143, + 88, + 831, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4553b3d3a14afc954421868ee5f9a3874071a0185d61caa1a83cd5c55569b839.jpg", + "table_caption": [ + "Table 1: Quantitative comparisons among the SOTA methods and our SSDNet in test datasets. The best and second best values are highlighted by **bold** and **underline**, respectively." + ], + "table_footnote": [], + "table_body": "
Dataset: RGBDD in real-world
MethodsRMSEMethodsRMSEMethodsRMSE
SVLRM [47]8.05DJF [34]7.90DJFR [35]8.01
DKN [31]7.38FDKN [31]7.50FDSR [27]7.50
DCTNet [90]7.37SSDNet7.32
FDSR* [27]5.49DCTNet* [90]5.43SSDNet*5.38
", + "bbox": [ + 81, + 316, + 468, + 402 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/47ce1e7f13528b133ecd957950e3c6a845a532edc1f99adf95e4e93af25eb56a.jpg", + "table_caption": [ + "Table 2: Quantitative comparisons among the SOTA methods and SSDNet on real-world branch of RGBDD. The best and second best RMSEs are highlighted by **bold** and **underline**. Model* means that the model has been fine-tuned on the real-world training data." + ], + "table_footnote": [], + "table_body": "
MethodsPACCUNDKNFDKNFDSRGSRDCTOurs
Time (s)0.200.230.210.040.010.920.080.10
", + "bbox": [ + 78, + 478, + 468, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Results of time-consuming comparison for generating a $640 \\times 480$ HR depth map.", + "bbox": [ + 75, + 518, + 468, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bic down-sampling to the HR depth maps to synthesize the corresponding LR depth maps. During the pre-processing stage, the training samples undergo resizing to $128 \\times 128$ . We train the network with a mini-batch size of 56 for 100 epochs, employing the Adam [32] optimizer with an initial learning rate of $5 \\times 10^{-3}$ . For training and testing, we employ a PC featuring eight NVIDIA GeForce RTX 3090 GPUs. $\\alpha_{1}, \\alpha_{2}$ and $\\alpha_{3}$ in Eq. (13) set to $10^{-2}$ , $10^{-3}$ and $10^{-2}$ in order to balance the magnitudes of each term in the loss. The SCR fine-tuning is performed once every 10 epochs of training. Further details are provided in the supplementary material.", + "bbox": [ + 75, + 574, + 470, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Empirical validation experiment", + "text_level": 1, + "bbox": [ + 76, + 752, + 364, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of network hyper-parameters. In our proposed SSDNet, the number of Restormer Blocks $P$ , and number of dimensions in Restormer $C$ , and the number of negative samples in SCR $N$ are crucial factors in enhancing the super-resolution performance. We conduct experiments on the NYU validation set to study the impact of different combinations of $P, C, N$ . Initially, we fix any two parameters at $C = 64$ , $P = 4$ , and $N = 8$ , respectively, and", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "evaluate the prediction quality of the unfixed parameter at $P = 2,3,4,5,6$ , $C = 32,64,96,128$ , and $N = 2,4,8,16$ . The results are summarized in Fig. 3. We observe that the performance is limited when $P < 4$ , and redundant parameter increases do not yield corresponding improvements when $P > 4$ . Similarly, increasing $C$ to more than 64 does not produce significant benefits but instead increased the training cost and computational expense. For the SCR module, the increase in training burden does not match the improvement in effectiveness when $N$ exceeds 8. Therefore, to ensure a balance between performance and computational cost, we choose $P = 4$ , $C = 64$ , $N = 8$ for subsequent experiments.", + "bbox": [ + 496, + 318, + 893, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of feature decomposition. Aligned and separated features are visualized in Fig. 4. Aligned features extract shared properties (edges/contours) from the depth/RGB image pair. Separated features capture modality-specific details (textures in RGB images, smooth depth distance in depth maps). The visualization aligns with our motivation.", + "bbox": [ + 496, + 505, + 895, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Comparison with SOTA methods", + "text_level": 1, + "bbox": [ + 498, + 609, + 790, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section aims to evaluate the performance of our SS-DNet on several popular benchmarks in Sec. 4.1, and compare our results with the state-of-the-art methods, including PAC [59], CUNet [11], DKN [31], FDKN [31], FDSR [27], GraphSR [8], and DCTNet [90].", + "bbox": [ + 496, + 633, + 893, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Comparison. We present comparisons of the error maps with ground truth depth maps among SOTA methods in Figs. 5 and 6. Intuitively, the depth predictions of SSDNet exhibit lower prediction errors and are closer to the ground truth images, especially in terms of perception and restoration for the edges and contours. Further visual comparisons are shown in the supplementary material.", + "bbox": [ + 496, + 714, + 893, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative results for traditional testsets. The quantitative results with scaling factors of 4, 8, and 16 on four test sets are shown in Tab. 1. Compared with existing methods that achieve good results on certain datasets or superresolution factors, our SSDNet produces satisfactory predic", + "bbox": [ + 496, + 824, + 895, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12554", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/004432a8726c72c87b4fdbae85c3b952b8117fb348684b1f041109eacb91bf69.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Encoder&DecoderLdecForm of LdecLRpixelSCR Module×4×8×16
Exp. IShared✓S✓✓1.232.103.49
Exp. IIPrivate×S✓✓1.171.943.27
Exp. IIIPrivate✓l2✓✓1.151.833.06
Exp. IVPrivate✓S×✓1.161.893.29
Exp. VPrivate✓S✓×1.181.873.36
OursPrivate✓S✓✓1.041.722.92
", + "bbox": [ + 161, + 88, + 812, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4: Results of ablation experiments on the RGBDD test set. Bold indicates the best score in terms of RMSE.", + "bbox": [ + 147, + 215, + 818, + 229 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tions on multiple datasets and various super-resolution scales. This indicates that our model has a better super-resolution performance than previous methods.", + "bbox": [ + 75, + 256, + 470, + 301 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Quantitative results for real-world branch. In addition, following [27, 90], we use the pre-trained $4 \\times$ model to directly test on the real-world branch of the RGBDD dataset to explore the generalization ability in real-world scenarios. We conduct all testing without any additional fine-tuning, and the quantitative results are presented in Tab. 2. Furthermore, we also perform targeted training and testing on this dataset. Regardless of whether fine-tuning or not, SSDNet outperforms previous methods in RMSE, highlighting its powerful generalization ability in handling unknown scenes.", + "bbox": [ + 75, + 305, + 470, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Parameter & running time comparison. We demonstrate the relationship between the number of model learnable parameters vs. RMSE on the RGBDD dataset for $\\times 4$ , $\\times 8$ , and $\\times 16$ and Middlebury for $\\times 4$ in Fig. 1. Our model exhibits a clear advantage over existing methods with relatively fewer parameters. The time-consuming comparison is presented in Tab. 3. They both demonstrate the efficiency of our method for the GDSR task, showing the potential for developing lightweight networks and practical applications in the future.", + "bbox": [ + 75, + 460, + 470, + 597 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 604, + 241, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we validate the design rationality of our SSDNet through ablation experiments based on the RGBDD testset, and present the results in Tab. 4.", + "bbox": [ + 75, + 628, + 468, + 674 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SSDNet architecture. In Exp. I, we share the encoders and decoders for RGB and depth images, i.e., a unified encoder and decoder are used instead of $\\{\\mathcal{E}_D,\\mathcal{E}_R\\}$ and $\\{\\mathcal{D}_D,\\mathcal{D}_R\\}$ . The number of Restormer blocks is increased to ensure that the learnable parameter number is comparable.", + "bbox": [ + 75, + 676, + 470, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Spherical feature decomposition. In Exp. II, we eliminate the entire feature decomposition module, i.e., $\\mathcal{L}_{dec}$ in Eq. (11) will not be employed. In Exp. III, we change the distance measure for $\\mathcal{L}_{sep}$ and $\\mathcal{L}_{align}$ in Eq. (12) from spherical space distance to Euclidean distance, i.e., $\\ell_2$ -loss.", + "bbox": [ + 75, + 757, + 470, + 833 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Reconstruction RGB. In Exp. IV, We eliminate $\\mathcal{L}_{pixel}^{R}$ in Eq. (10) and explore the change for feature extraction capability without the restriction of reconstructing RGB.", + "bbox": [ + 75, + 835, + 470, + 882 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Spherical contrast refinement. In Exp. V, we remove the", + "bbox": [ + 75, + 885, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SCR module and use $\\hat{D}_{HR}$ in Eq. (8) as the final output of our model.", + "bbox": [ + 496, + 255, + 893, + 285 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Analysis. Tab. 4 shows that changing settings leads to performance degradation, validating our model design.", + "bbox": [ + 496, + 292, + 893, + 321 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Exp. I: Unified encoder/decoder ignores modality-specific features, hindering efficient cross-modal feature extraction and causing the largest degradation.", + "bbox": [ + 496, + 321, + 890, + 367 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Exp. II: The absence of $\\mathcal{L}_{dec}$ prevents decomposing the shared/private features.", + "bbox": [ + 496, + 367, + 890, + 396 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Exp. III: $\\ell_2$ distance partially decomposes features but is less effective than $S$ .", + "bbox": [ + 496, + 397, + 890, + 426 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Exp. IV: $\\mathcal{L}_{pixel}^{R}$ preserves semantic information in RGB and avoids excessive adaptation of feature separation/alignment.", + "bbox": [ + 496, + 426, + 890, + 458 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Exp. V: Removal of SCR leads to mentioned detailed issues and performance degradation.", + "bbox": [ + 496, + 458, + 890, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 496, + 508, + 619, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We propose a Spherical Space feature Decomposition network (SSDNet) for guided depth map super-resolution, where a Restormer-based encoder is used to extract the global features of the inputs, and the intermediate features will be mapped to the spherical space to complete the separation/alignment of shared/private features, respectively. Finally, a Restormer-based decoder is to reconstruct the HR depth map. The spherical contrast refinement module is then employed to address the possible detail issues, e.g., blurry edges, noisy surfaces and over-transferred RGB texture. The satisfactory output results and the lightweight size demonstrate the superiority of our approach.", + "bbox": [ + 496, + 536, + 893, + 718 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 737, + 660, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work has been supported by the National Key Research and Development Program of China under grant 2018AAA0102201, the National Natural Science Foundation of China under Grant 61976174 and 12201497, Shaanxi Fundamental Science Research Project for Mathematics and Physics under Grant 22JSQ033, the Fundamental Research Funds for the Central Universities under Grant D5000220060, and partly supported by the Alexander von Humboldt Foundation.", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12555", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. CoRR, abs/2106.06847, 2021. 3", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 3", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021. 3", + "[4] Efrain Castillo-Muniz and Eduardo Bayro-Corrochano. Geometric spherical networks for visual data processing. In IJCNN, 2012. 2, 3", + "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310. CVF / IEEE, 2021. 3", + "[6] Taco Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Convolutional networks for spherical signals. arXiv preprint arXiv:1709.04893, 2017. 3", + "[7] Taco S. Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical CNNs. In ICLR, 2018. 2, 3", + "[8] Riccardo de Lutio, Alexander Becker, Stefano D'Aronco, Stefania Russo, Jan D. Wegner, and Konrad Schindler. Learning graph regularisation for guided super-resolution. In CVPR, pages 1969-1978. IEEE, 2022. 2, 8", + "[9] Riccardo de Lutio, Stefano D'Aronco, Jan Dirk Wegner, and Konrad Schindler. Guided super-resolution as pixel-to-pixel transformation. In ICCV, pages 8828-8836. IEEE, 2019. 2", + "[10] Xin Deng and Pier Luigi Dragotti. Deep coupled ISTA network for multi-modal image super-resolution. IEEE Trans. Image Process., 29:1683-1698, 2020. 2, 3", + "[11] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Trans. Pattern Anal. Mach. Intell., 43(10):3333-3348, 2021. 2, 8", + "[12] James Diebel and Sebastian Thrun. An application of markov random fields to range sensing. In NeurIPS, pages 291-298, 2005. 2", + "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image superresolution. In ECCV, pages 184-199. Springer, 2014. 1", + "[14] Jiangxin Dong, Jinshan Pan, Jimmy S. Ren, Liang Lin, Jinhui Tang, and Ming-Hsuan Yang. Learning spatially variant linear representation models for joint filtering. IEEE Trans. Pattern Anal. Mach. Intell., 44(11):8355-8370, 2022. 2", + "[15] Xiaoyu Dong, Naoto Yokoya, Longguang Wang, and Tatsumi Uezato. Learning mutual modulation for self-supervised cross-modal super-resolution. In ECCV, 2022. 2", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[17] David Ferstl, Christian Reinbacher, René Ranftl, Matthias Rüther, and Horst Bischof. Image guided depth upsampling" + ], + "bbox": [ + 78, + 114, + 470, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "using anisotropic total generalized variation. In ICCV, pages 993-1000, 2013. 2", + "[18] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. RSTT: real-time spatial temporal transformer for space-time video super-resolution. CoRR, abs/2203.14186, 2022. 3", + "[19] Shuhang Gu, Shi Guo, Wangmeng Zuo, Yunjin Chen, Radu Timofte, Luc Van Gool, and Lei Zhang. Learned dynamic guidance for depth image reconstruction. IEEE Trans. Pattern Anal. Mach. Intell., 42(10):2437-2452, 2019. 2", + "[20] Shuhang Gu, Wangmeng Zuo, Shi Guo, Yunjin Chen, Chongyu Chen, and Lei Zhang. Learning dynamic guidance for depth image enhancement. In CVPR, pages 712-721, 2017. 2", + "[21] Xiang Gu, Jian Sun, and Zongben Xu. Spherical space domain adaptation with robust pseudo-label loss. In CVPR, 2020. 2", + "[22] Xiang Gu, Jian Sun, and Zongben Xu. Unsupervised and semi-supervised robust spherical space domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-1, 2022. 2, 3", + "[23] Chunle Guo, Chongyi Li, Jichang Guo, Runmin Cong, Huazhu Fu, and Ping Han. Hierarchical features driven residual learning for depth map super-resolution. IEEE Trans. Image Process., 28(5):2545-2557, 2019. 2", + "[24] Saurabh Gupta, Ross B. Girshick, Pablo Andres Arbeláez, and Jitendra Malik. Learning rich features from RGB-D images for object detection and segmentation. In ECCV, pages 345-360. Springer, 2014. 1", + "[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735. CVF / IEEE, 2020. 3", + "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5", + "[27] Lingzhi He, Hongguang Zhu, Feng Li, Huihui Bai, Runmin Cong, Chunjie Zhang, Chunyu Lin, Meiqin Liu, and Yao Zhao. Towards fast and accurate real-world depth superresolution: Benchmark dataset and baseline. In CVPR, pages 9229-9238, 2021. 7, 8, 9", + "[28] Olivier J. Henaff. Data-efficient image recognition with contrastive predictive coding. In ICML, volume 119, pages 4182-4192. PMLR, 2020. 3", + "[29] Heiko Hirschmüller and Daniel Scharstein. Evaluation of cost functions for stereo matching. In CVPR, 2007. 7, 8", + "[30] Tak-Wai Hui, Chen Change Loy, and Xiaou Tang. Depth map super-resolution by deep multi-scale guidance. In ECCV, pages 353-369. Springer, 2016. 2", + "[31] Beomjun Kim, Jean Ponce, and Bumsub Ham. Deformable kernel networks for joint image filtering. Int. J. Comput. Vis., 129(2):579-600, 2021. 2, 7, 8", + "[32] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 8", + "[33] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. CoRR, abs/2112.10175, 2021. 3", + "[34] Yijun Li, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep joint image filtering. In ECCV, pages 154-169. Springer, 2016. 8", + "[35] Yijun Li, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan" + ], + "bbox": [ + 501, + 93, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12556", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yang. Joint image filtering with deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell., 41(8):1909-1923, 2019. 2, 8", + "[36] Yu Li, Dongbo Min, Minh N. Do, and Jiangbo Lu. Fast guided global interpolation for depth and motion. In ECCV, pages 717-733. Springer, 2016. 2", + "[37] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCVW, pages 1833-1844. IEEE, 2021. 3", + "[38] Miao Liao, Feixiang Lu, Dingfu Zhou, Sibo Zhang, Wei Li, and Ruigang Yang. DVI: depth guided video inpainting for autonomous driving. In ECCV, pages 1-17. Springer, 2020. 1", + "[39] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 1", + "[40] Ming-Yu Liu, Oncel Tuzel, and Yuichi Taguchi. Joint geodesic upsampling of depth images. In CVPR, pages 169-176, 2013. 2", + "[41] Xianming Liu, Deming Zhai, Rong Chen, Xiangyang Ji, Debin Zhao, and Wen Gao. Depth super-resolution via joint color-guided internal and external regularizations. IEEE Trans. Image Process., 28(4):1636-1645, 2019. 1", + "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002. IEEE, 2021. 3", + "[43] Jiajun Lu and David A. Forsyth. Sparse depth super resolution. In CVPR, pages 2245-2253, 2015. 2", + "[44] Jiangbo Lu, Keyang Shi, Dongbo Min, Liang Lin, and Minh N. Do. Cross-based local multipoint filtering. In CVPR, pages 430-437, 2012. 2", + "[45] Si Lu, Xiaofeng Ren, and Feng Liu. Depth enhancement via low-rank matrix completion. In CVPR, pages 3390-3397, 2014. 7, 8", + "[46] Dongbo Min, Jiangbo Lu, and Minh N. Do. Depth video enhancement based on weighted mode filtering. IEEE Trans. Image Process., 21(3):1176-1190, 2012. 2", + "[47] Jinshan Pan, Jiangxin Dong, Jimmy S. J. Ren, Liang Lin, Jinhui Tang, and Ming-Hsuan Yang. Spatially variant linear representation models for joint filtering. In CVPR, pages 1702-1711, 2019. 8", + "[48] Jaesik Park, Hyeongwoo Kim, Yu-Wing Tai, Michael S. Brown, and In-So Kweon. High quality depth map upsampling for 3d-tof cameras. In ICCV, pages 1623-1630, 2011. 2", + "[49] Taesung Park, Alexei A. Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In ECCV (9), volume 12354 of Lecture Notes in Computer Science, pages 319-345. Springer, 2020. 3", + "[50] Wanli Peng, Hao Pan, He Liu, and Yi Sun. IDA-3D: instance-depth-aware 3d object detection from stereo vision for autonomous driving. In CVPR, pages 13012–13021, 2020. 1", + "[51] Haotong Qin, Zhongang Cai, Mingyuan Zhang, Yifu Ding, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Hao Su. Bipoint-net: Binary neural network for point clouds. In ICLR, 2021. 1", + "[52] Haotong Qin, Ruihao Gong, Xianglong Liu, Xiao Bai, Jingkuan Song, and Nicu Sebe. Binary neural networks:" + ], + "bbox": [ + 78, + 90, + 470, + 887 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "A survey. Pattern Recognition, 2020. 1", + "[53] Haotong Qin, Ruihao Gong, Xianglong Liu, Mingzhu Shen, Ziran Wei, Fengwei Yu, and Jingkuan Song. Forward and backward information retention for accurate binary neural networks. In ICCV, 2020. 1", + "[54] Gernot Riegler, David Ferstl, Matthias Rüther, and Horst Bischof. A deep primal-dual network for guided depth superresolution. In BMVC. BMVA Press, 2016. 2", + "[55] Gernot Riegler, Matthias Rüther, and Horst Bischof. Atgv-net: Accurate depth super-resolution. In ECCV, pages 268-284. Springer, 2016. 2", + "[56] Daniel Scharstein and Chris Pal. Learning conditional random fields for stereo. In CVPR, 2007. 7, 8", + "[57] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. In ECCV, pages 746-760. Springer, 2012. 7, 8", + "[58] Xibin Song, Yuchao Dai, Dingfu Zhou, Liu Liu, Wei Li, Hongdong Li, and Ruigang Yang. Channel attention based iterative residual learning for depth map super-resolution. In CVPR, pages 5630-5639, 2020. 2", + "[59] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik G. Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In CVPR, pages 11166-11175, 2019. 8", + "[60] Baoli Sun, Xinchen Ye, Baopu Li, Haojie Li, Zhihui Wang, and Rui Xu. Learning scene structure guidance via cross-task knowledge transfer for single depth super-resolution. In CVPR, pages 7792-7801, 2021. 2", + "[61] Jiaxiang Tang, Xiaokang Chen, and Gang Zeng. Joint implicit image function for guided depth super-resolution. In ACM Multimedia, pages 4390-4399. ACM, 2021.", + "[62] Qi Tang, Runmin Cong, Ronghui Sheng, Lingzhi He, Dan Zhang, Yao Zhao, and Sam Kwong. Bridgenet: A joint learning network of depth map super-resolution and monocular depth estimation. In ACM Multimedia, pages 2148-2157. ACM, 2021. 2", + "[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, volume 139, pages 10347-10357. PMLR, 2021. 3", + "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 3", + "[65] Jiakai Wang, Aishan Liu, Xiao Bai, and Xianglong Liu. Universal adversarial patch attack for automatic checkout using perceptual and attentional bias. IEEE Trans. Image Process., 31:598-611, 2021. 1", + "[66] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 548-558. IEEE, 2021. 3", + "[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. CoRR, abs/2106.03106, 2021. 3", + "[68] Yang Wen, Bin Sheng, Ping Li, Weiyao Lin, and David Dagan Feng. Deep color guided coarse-to-fine convolutional network" + ], + "bbox": [ + 503, + 92, + 892, + 888 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12557", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + " cascade for depth image super-resolution. IEEE Trans. Image Process., 28(2):994-1006, 2019. 2", + "[69] Richard C. Wilson, Edwin R. Hancock, Elzbieta Pekalska, and Robert P. W. Duin. Spherical and hyperbolic embeddings of data. IEEE Trans. Pattern Anal. Mach. Intell., 36(11):2255-2269, 2014. 4", + "[70] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In CVPR, pages 10551-10560. CVF / IEEE, 2021. 3", + "[71] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Fast end-to-end trainable guided filter. In CVPR, pages 1838-1847, 2018. 2", + "[72] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In ICCV, pages 8372-8381. IEEE, 2021. 3", + "[73] Jun Xie, Rogério Schmidt Feris, and Ming-Ting Sun. Edge-guided single depth image super resolution. IEEE Trans. Image Process., 25(1):428-438, 2016. 2", + "[74] Jun Xie, Rogério Schmidt Feris, Shiaw-Shian Yu, and Ming-Ting Sun. Joint super resolution and denoising from a single depth image. IEEE Trans. Multim., 17(9):1525-1537, 2015. 2", + "[75] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2J: anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In ICCV, pages 793-802. IEEE, 2019. 1", + "[76] Ruikang Xu, Mingde Yao, and Zhiwei Xiong. Zero-shot dual-lens super-resolution. In CVPR, pages 9130-9139, 2023. 1", + "[77] Jingyu Yang, Xinchen Ye, Kun Li, Chunping Hou, and Yao Wang. Color-guided depth recovery from RGB-D data using an adaptive autoregressive model. IEEE Trans. Image Process., 23(8):3443-3458, 2014. 2", + "[78] Yuxiang Yang, Qi Cao, Jing Zhang, and Dacheng Tao. CODON: on orchestrating cross-domain attentions for depth super-resolution. Int. J. Comput. Vis., 130(2):267-284, 2022. 2", + "[79] Mingde Yao, Dongliang He, Xin Li, Fu Li, and Zhiwei Xiong. Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1", + "[80] Mingde Yao, Dongliang He, Xin Li, Zhihong Pan, and Zhiwei Xiong. Bidirectional translation between uhd-hdr and hd-sdr videos. IEEE Transactions on Multimedia, 2023. 1", + "[81] Xinchen Ye, Baoli Sun, Zhihui Wang, Jingyu Yang, Rui Xu, Haojie Li, and Baopu Li. Pmbanet: Progressive multi-branch aggregation network for scene depth super-resolution. IEEE Trans. Image Process., 29:7427-7442, 2020. 2", + "[82] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In CVPR, pages 8824-8833. IEEE, 2022. 3", + "[83] Shanxin Yuan, Guillermo Garcia-Hernando, Björn Stenger, Gyeongsik Moon, Ju Yong Chang, Kyoung Mu Lee, Pavlo Molchanov, Jan Kautz, Sina Honari, Liuhao Ge, Junsong Yuan, Xinghao Chen, Guijin Wang, Fan Yang, Kai Akiyama," + ], + "bbox": [ + 78, + 90, + 470, + 890 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yang Wu, Qingfu Wan, Meysam Madadi, Sergio Escalera, Shile Li, Dongheui Lee, Jason Oikonomidis, Antonis A. Argyros, and Tae-Kyun Kim. Depth-based 3d hand pose estimation: From current achievements to future goals. In CVPR, pages 2636-2645, 2018. 1", + "[84] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, pages 5718-5729. IEEE, 2022. 2, 3", + "[85] Jiahui Zhang, Shijian Lu, Fangneng Zhan, and Yingchen Yu. Blind image super-resolution via contrastive representation learning. CoRR, abs/2107.00708, 2021. 3", + "[86] Kai Zhang, Luc Van Gool, and Radu Timofte. Deep unfolding network for image super-resolution. In CVPR, pages 3214-3223, 2020. 1", + "[87] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In CVPR, pages 5906–5916, June 2023. 1", + "[88] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. arXiv preprint arXiv:2305.11443, 2023.", + "[89] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. DDFM: denoising diffusion model for multi-modality image fusion. CoRR, abs/2303.06840, 2023. 1", + "[90] Zixiang Zhao, Jiangshe Zhang, Shuang Xu, Zudi Lin, and Hanspeter Pfister. Discrete cosine transform network for guided depth map super-resolution. In CVPR, pages 5687-5697. IEEE, 2022. 2, 3, 7, 8, 9", + "[91] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H. S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890. CVF / IEEE, 2021. 3", + "[92] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, Zhiwen Chen, and Xiangyang Ji. High-resolution depth maps imaging via attention-based hierarchical multi-modal fusion. IEEE Trans. Image Process., 31:648-663, 2022. 2", + "[93] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, and Xiangyang Ji. Guided depth map super-resolution: A survey. ACM Computing Surveys, 2023. 1", + "[94] Man Zhou, Keyu Yan, Jinshan Pan, Wenqi Ren, Qi Xie, and Xiangyong Cao. Memory-augmented deep unfolding network for guided image super-resolution. Int. J. Comput. Vis., 131(1):215-242, 2023. 2", + "[95] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3", + "[96] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rhemann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew W. Fitzgibbon, Charles T. Loop, Christian Theobalt, and Marc Stamminger. Real-time non-rigid reconstruction using an RGB-D camera. ACM Trans. Graph., 33(4):156:1-156:12, 2014. 1" + ], + "bbox": [ + 503, + 90, + 893, + 892 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12558", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/55278fbf-9750-4827-94bb-81181843441e_model.json b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/55278fbf-9750-4827-94bb-81181843441e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..399fb1baf01eb9834b1a5dc4a7c06faa60ab9d49 --- /dev/null +++ b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/55278fbf-9750-4827-94bb-81181843441e_model.json @@ -0,0 +1,3150 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.131, + 0.895, + 0.156 + ], + "angle": 0, + "content": "Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.18, + 0.744, + 0.218 + ], + "angle": 0, + "content": "Zixiang Zhao\\(^{1,2}\\) Jiangshe Zhang\\(^{1*}\\) Xiang Gu\\(^{1}\\) Chengli Tan\\(^{1}\\) Shuang Xu\\(^{3}\\) Yulun Zhang\\(^{2}\\) Radu Timofte\\(^{2,4}\\) Luc Van Gool\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.223, + 0.751, + 0.261 + ], + "angle": 0, + "content": "\\(^{1}\\)Xi'an Jiaotong University \\(^{2}\\)Computer Vision Lab, ETH Zürich \\(^{3}\\)Northwestern Polytechnical University \\(^{4}\\)University of Würzburg" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.263, + 0.721, + 0.277 + ], + "angle": 0, + "content": "zixiangzhao@stu.xjtu.edu.cn, jszhang@mail.xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.311, + 0.314, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.343, + 0.474, + 0.766 + ], + "angle": 0, + "content": "Guided depth map super-resolution (GDSR), as a hot topic in multi-modal image processing, aims to upsample low-resolution (LR) depth maps with additional information involved in high-resolution (HR) RGB images from the same scene. The critical step of this task is to effectively extract domain-shared and domain-private RGB/depth features. In addition, three detailed issues, namely blurry edges, noisy surfaces, and over-transferred RGB texture, need to be addressed. In this paper, we propose the Spherical Space feature Decomposition Network (SSDNet) to solve the above issues. To better model cross-modality features, Restormer block-based RGB/depth encoders are employed for extracting local-global features. Then, the extracted features are mapped to the spherical space to complete the separation of private features and the alignment of shared features. Shared features of RGB are fused with the depth features to complete the GDSR task. Subsequently, a spherical contrast refinement (SCR) module is proposed to further address the detail issues. Patches that are classified according to imperfect categories are input into the SCR module, where the patch features are pulled closer to the ground truth and pushed away from the corresponding imperfect samples in the spherical feature space via contrastive learning. Extensive experiments demonstrate that our method can achieve state-of-the-art results on four test datasets, as well as successfully generalize to real-world scenes. The code is available at https://github.com/Zhaozixiang1228/GDSR-SSDNet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.793, + 0.21, + 0.809 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.819, + 0.472, + 0.88 + ], + "angle": 0, + "content": "Depth maps, as images that measure the distance of scene points from the sensor, are widely used in autonomous driving [38, 50, 87-89], pose estimation [75, 83], virtual reality [41, 51, 79, 80], and scene understanding [24, 53, 65, 96]." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.309, + 0.697, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.428, + 0.695, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.309, + 0.891, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.428, + 0.89, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.539, + 0.895, + 0.581 + ], + "angle": 0, + "content": "Figure 1: Our SSDNet achieves outstanding performance on the RGBDD dataset for \\(\\times 4\\), \\(\\times 8\\), and \\(\\times 16\\) and Middlebury for \\(\\times 4\\) while being computationally efficient." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.895, + 0.688 + ], + "angle": 0, + "content": "However, the depth maps produced by current consumer-level depth sensors, e.g., Time-of-Flight (ToF) and Kinect cameras, often have the disadvantages of low resolution and noise. These disadvantages are insufficient to meet the requirements of advanced computing vision tasks [52, 76, 93]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.897, + 0.901 + ], + "angle": 0, + "content": "To obtain high-resolution depth maps, we naturally hope to accomplish depth map super-resolution (SR) by transferring mature SR models in the RGB image domain [13, 39, 86] to the depth SR task. However, a potential risk is that RGB SR models tend to focus on reconstructing high-frequency image information, such as details and texture. Conversely, for depth images, the objects' depth information is often textureless and piecewise, and more sensitive to unclear edges and noise. Therefore, it is unreasonable to directly apply RGB SR models to the depth SR task. On the other hand, while acquiring the depth map, it is relatively easy to obtain HR and noise-free RGB images in the same scene. Furthermore, there are statistical co-occurrences between the edges in RGB images and the discontinuities in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12547" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "depth maps [55]. Therefore, we hope to use the HR RGB images to provide edges and contour information missing in the depth map, and to fuse the multi-modal information to accomplish the LR depth image upsampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.471, + 0.367 + ], + "angle": 0, + "content": "In the era of deep learning (DL), numerous methods have been utilized to learn the mapping between \\(\\mathrm{LR} \\rightarrow \\mathrm{HR}\\) depth maps. These methods succeeded to some extent in modeling the cross-modality features and reconstructing the contour and edge information. However, DL models which rely on natural priors will limit the model's flexibility. Models learning the \\(\\mathrm{LR} \\rightarrow \\mathrm{HR}\\) mapping via data-driven methods are difficult to interpret due to the unclear working mechanism [90]. Thus, effectively extracting and distinguishing domain-shared and domain-private RGB/depth features is still a challenge. While microscopically, the obtained depth images are often plagued by three detail issues: blurry edges, noisy surfaces, and over-transferred RGB texture, all of which affect the display quality of depth maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.471, + 0.581 + ], + "angle": 0, + "content": "In response to addressing the above challenges, we expect to limit the solution space by constraining the extracted features and further improve the modeling of the dependencies between different modalities. Based on our observation, RGB images and depth maps contain shared features, such as depth map discontinuities and edge features of RGB objects, which can be aligned in the feature space. In addition, unique private features, such as the distance information of the depth map and the texture of RGB object surface, should be separated conversely. Thus, in the feature space, domain-shared and domain-private features are expected to separate and align respectively, while the distance between HR features and imperfect features with the above-mentioned detail issues should also be pushed away." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The above goal can be divided into two steps. First, extracting features through an effective encoder, and then selecting an appropriate distance measure to complete the alignment and separation of features. Considering that CNN-based architectures limit feature extraction capabilities due to content-independent convolution kernels and the lack of global information, we employ the Restormer blocks [84], which has been proved to effectively extract features in the low-level vision domain, as the basic unit of our encoder. For the choice of distance measure, we first think of the Euclidean distance, such as \\(\\ell_1\\) or \\(\\ell_2\\) distance. However, cross-modality features often contain different scales and orders of magnitude, and the Euclidean distance is easily affected by the scale, which makes it difficult to achieve our goal. Recently, with the development of spherical DL models [4, 7, 21, 22], the spherical feature learning is well known by virtue of its advantages over the Euclidean feature learning in many applications, e.g., domain adaptation [21]. Due to the distances between spherical features being regularized, the alignment and separation of features can be done more easily without losing the model's representational" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.093, + 0.563, + 0.107 + ], + "angle": 0, + "content": "capacity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.109, + 0.894, + 0.167 + ], + "angle": 0, + "content": "According to the above analysis, we propose our Spherical Space feature Decomposition Network (SSDNet). The specific workflow is displayed in Fig. 2. Our contribution can be summarized in three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.17, + 0.892, + 0.26 + ], + "angle": 0, + "content": "- We propose a spherical space feature decomposition framework to model the cross-modality features. The features extracted by the Restormer block are mapped to the spherical space for separation and alignment. This is the first time that the Transformer and spherical-space distance measure are applied to the GDSR task." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.261, + 0.892, + 0.349 + ], + "angle": 0, + "content": "- The spherical contrast refinement module, cooperating with the imperfect patch classification and the corresponding contrastive learning branch, is proposed to address the possible detail issues in depth maps. This is also the first time that the contrastive learning technique has been used for the GDSR task." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.351, + 0.892, + 0.411 + ], + "angle": 0, + "content": "- Experiments on four GDSR benchmarks and a real scene dataset demonstrate that our method can generate satisfactory HR depth maps in different scenarios and exhibits good generalization ability." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.17, + 0.892, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.429, + 0.642, + 0.445 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.455, + 0.892, + 0.502 + ], + "angle": 0, + "content": "In this section, we briefly introduce the GDSR task, and illustrate the Vision Transformer, spherical space DL and contrastive learning techniques utilized in SSDNet." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.514, + 0.657, + 0.53 + ], + "angle": 0, + "content": "2.1. GDSR methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Image super-resolution is a popular image processing and computer vision task with too many sub-categories, so we only discuss the GDSR task here. Conventional GDSR methods can be divided into local- [40, 43, 44, 46], global- [12, 17, 36, 48, 77] and learning-based [8, 19, 20, 73, 74] methods, etc. They extract cross-modality information by manually designed filters, optimizing equations, or sparse dictionary learning. With the rapid development of DL, GDSR is further promoted by the CNN-based methods [58, 60–62, 68, 81]. Generally, DL-based methods can be categorized in three groups, i.e., auto-encoder (AE) [23, 30, 35, 78, 90, 92], learnable filter [14, 31, 71], algorithm unfolding [10, 11, 54, 94] and unsupervised [9, 15] methods. AE-based methods learn cross-modality features via shared or private encoders and reconstruct HR depth maps with decoders. In the learnable filter group, the convolution kernels used to extract information are set to be context-dependent and spatially-variant, which improves the flexibility of the model. The algorithmic unfolding group establishes interpretable GDSR models by building a bridge between traditional optimization functions and deep learning. Unsupervised methods change the training paradigm that requires paired LR/HR images, thus solving the difficulty of obtaining paired training data." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12548" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.296, + 0.108 + ], + "angle": 0, + "content": "2.2. Spherical deep learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.473, + 0.356 + ], + "angle": 0, + "content": "The works [4, 6, 7] introduce the geometrical neural network or the spherical convolution neural networks for analyzing the spherical signals that are rotation invariant. Gu et al. [22] propose the spherical multi-layer perceptron for aligning feature distributions of different domains in spherical feature space for domain adaptation. Our idea of spherical space feature decomposition is inspired by Gu et al. [22] that shows the advantages of spherical feature learning over the Euclidean feature learning for domain adaptation in the \"high-level\" image classification task. Different from [22], we tackle the \"low-level\" GDSR task by aligning/separating the shared/private features of paired RGB image and depth map, rather than aligning the feature distributions as in [22]. We further propose the spherical contrast refinement for tackling the issues of blurry edges, noisy surfaces, and overtransferred RGB texture for GDSR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.368, + 0.282, + 0.383 + ], + "angle": 0, + "content": "2.3. Transformer in vision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.473, + 0.603 + ], + "angle": 0, + "content": "Transformer, proposed by Vaswani et al. [64], has now become a popular technology in computer vision. It has achieved dominance in classification [16, 42, 63], object detection [2, 95], segmentation [66, 91], etc. Simultaneously, it has also made progress in low-level vision [1, 18, 33, 67]. Chen et al. [5] proposed IPT based on the standard Transformer and multi-task learning. Liang et al. [37] proposed SwinIR whose core module inherits from Swin Transformer [42]. Recently, Restormer [84] improves the transformer block by the gated-Dconv network and multi-Dconv head transposed attention. By applying self-attention in the feature dimension, the ability of local-global representation learning is maintained while being more friendly to high-resolution input images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.615, + 0.28, + 0.631 + ], + "angle": 0, + "content": "2.4. Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.473, + 0.821 + ], + "angle": 0, + "content": "Contrastive learning, which aims to address the scarcity of source and target paired data and prioritize self-supervised representation learning [25, 28]. The primary objective is to bring samples close to their positive counterparts and push them away from negative ones within high-dimensional manifolds, and it seeks to increase the distance between samples belonging to different classes. Recently, contrastive learning has gained significant attention in computer vision, particularly in high-level tasks such as object detection [72], video segmentation [3], and object tracking [82]. It has also found applications in low-level tasks such as super-resolution [85], image translation [49], and image dehazing [70]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.405, + 0.849 + ], + "angle": 0, + "content": "2.5. Comparison with existing approaches" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.901 + ], + "angle": 0, + "content": "The methods most related to our approach are feature-decomposition DL methods [10, 90], which propose the idea that cross-modality features contain the common and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.243 + ], + "angle": 0, + "content": "private information. In contrast, our method, for the first time, exploits the Transformer model's powerful ability to model globally dependent features. Meanwhile, the distance measure in spherical space is proved to decompose the cross-modality features more effectively than Euclidean distances. Additionally, instead of only focusing on optimizing the \\(\\ell_2\\) loss between the reconstructed image and ground truth, we give classification-based fine-tuning for image patches under different imperfect conditions, thus further improving the performance of GDSR." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.259, + 0.593, + 0.274 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.894, + 0.345 + ], + "angle": 0, + "content": "In this section, we first present the specific architecture of our SSDNet. Then, comprehensive descriptions are provided for each module. Finally, the loss function and training details will also be addressed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.358, + 0.611, + 0.373 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.895, + 0.519 + ], + "angle": 0, + "content": "We can define the input LR depth map, RGB image and the HR depth map as \\( D_{LR} \\in \\mathbb{R}^{h \\times w} \\), \\( D_{HR} \\in \\mathbb{R}^{H \\times W} \\) and \\( R \\in \\mathbb{R}^{H \\times W \\times 3} \\), where \\( \\{H, W\\} \\) and \\( \\{h, w\\} \\) are the height and width of input RGB and depth images, respectively. For our SSDNet network, it consists of five modules, namely, the Encoders for RGB and Depth images, the Decoders for RGB and Depth images, and the Spherical Contrast Refinement (SCR) module, which is denoted as \\( \\mathcal{E}_{\\mathcal{R}}(\\cdot), \\mathcal{E}_{\\mathcal{D}}(\\cdot), \\mathcal{D}_{\\mathcal{R}}(\\cdot), \\mathcal{D}_{\\mathcal{D}}(\\cdot), \\) and \\( S(\\cdot) \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.895, + 0.64 + ], + "angle": 0, + "content": "In general, as shown in Fig. 2, SSDNet utilizes \\(\\mathcal{E}_{\\mathcal{R}}\\) and \\(\\mathcal{E}_{\\mathcal{D}}\\) to extract features, which are then projected onto the spherical space for feature separation and alignment. Then, the depth features and shared RGB features are input into \\(\\mathcal{D}_{\\mathcal{R}}\\) to obtain reconstructed depth maps, and RGB features are fed into \\(\\mathcal{E}_{\\mathcal{R}}\\) to get reconstructed RGB images. Finally, the reconstructed depth maps are subsequently provided to the SCR module to refine details and output the final \\(D_{HR}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.897, + 0.792 + ], + "angle": 0, + "content": "The basic unit we employed for feature extraction and image restoration is the Restormer block [84]. The reason we opted for the Restormer block in \\(\\mathcal{E}_{\\mathcal{R}}\\), \\(\\mathcal{E}_{\\mathcal{D}}\\), \\(\\mathcal{D}_{\\mathcal{R}}\\) and \\(\\mathcal{D}_{\\mathcal{D}}\\) is because it allows for the extraction of global features from high-resolution input images by utilizing self-attention across feature dimensions [84]. This approach enables the extraction of cross-modality shallow features without significantly increasing computational requirements. For details on the Restormer block architecture, please refer to the supplementary material or the original paper [84]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.738, + 0.831 + ], + "angle": 0, + "content": "3.1.1 Spherical space transform" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We first define the mapping based on Riemannian geometry between Euclidean Space feature maps and Spherical Space feature maps, i.e., \\(\\mathcal{L}\\mathcal{O}\\mathcal{G}(\\cdot)\\) and \\(\\mathcal{E}\\mathcal{X}\\mathcal{P}(\\cdot)\\), which are employed in the feature decomposition. A schematic diagram for the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12549" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.093, + 0.425, + 0.106 + ], + "angle": 0, + "content": "(a) Workflow of SSDNet" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.108, + 0.613, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.29, + 0.48, + 0.302 + ], + "angle": 0, + "content": "(c) Spherical Contrast Refinement Module" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.303, + 0.603, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.094, + 0.885, + 0.106 + ], + "angle": 0, + "content": "(b) Spherical Space Feature Decomposition Module" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.114, + 0.88, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.361, + 0.868, + 0.373 + ], + "angle": 0, + "content": "(d) Illustration of Spherical transform" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.376, + 0.865, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.631, + 0.459, + 0.776, + 0.473 + ], + "angle": 0, + "content": "C: Channel Concatenation" + }, + { + "type": "image_footnote", + "bbox": [ + 0.632, + 0.476, + 0.819, + 0.489 + ], + "angle": 0, + "content": "ATTN: Attention (Feature Extraction)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.632, + 0.491, + 0.8, + 0.504 + ], + "angle": 0, + "content": "\\(\\mathcal{S}\\{\\cdot ,\\cdot \\} :\\) Spherical-space Distance" + }, + { + "type": "list", + "bbox": [ + 0.631, + 0.459, + 0.819, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.523, + 0.894, + 0.594 + ], + "angle": 0, + "content": "Figure 2: (a) The Architecture of our SSDNet method. First, the LR depth map is input into \\(\\mathcal{E}_{\\mathcal{D}}\\), and the RGB image is input into \\(\\mathcal{E}_{\\mathcal{R}}\\). The intermediate feature maps are projected onto the spherical space to complete the feature decomposition, i.e., alignment and separation. The shared & private features of the depth map, and the shared features of the RGB image are input into the \\(\\mathcal{D}_{\\mathcal{D}}\\) to reconstruct HR depth map. The RGB features are fed into \\(\\mathcal{D}_{\\mathcal{R}}\\) to reconstruct RGB image. (b) Illustration of spherical space feature decomposition. (c) The module for spherical contrast refinement. (d) The mapping between Euclidean Space and Spherical Space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.65 + ], + "angle": 0, + "content": "mappings is shown in Fig. 2d. The distance measure on the spherical space is also defined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.473, + 0.735 + ], + "angle": 0, + "content": "Definition 1 (Spherical exponential mapping). Given a vector \\( v \\) in \\( d \\)-dimensional Euclidean Space, we define the \\( (d + 1) \\)-dimensional vector \\( \\bar{v} \\) by \\( \\bar{v} = (v,r) \\) where \\( r \\) is a hyperparameter named radius. Then the spherical exponential transform [69] \\( \\exp_N: T_N\\mathbb{S}_r^{d + 1} \\to \\mathbb{S}_r^{d + 1} \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.743, + 0.47, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\exp_ {N} (v) = N \\cos \\theta + \\bar {v} \\frac {\\sin \\theta}{\\theta}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.781, + 0.471, + 0.875 + ], + "angle": 0, + "content": "where \\(\\mathbb{S}_r^{d + 1} = \\{x\\in \\mathbb{R}^{d + 1}:||x|| = r\\} \\subset \\mathbb{R}^{d + 1}\\) is the \\(d\\) -dimensional spherical space, \\(N = (0,\\dots ,0,r)\\in \\mathbb{S}_r^{d + 1}\\) is the north pole, \\(\\theta = \\frac{\\|\\vec{v}\\|}{r}\\), and \\(T_{N}\\mathbb{S}_{r}^{d + 1} = \\{(v,r):v\\in\\) \\(\\mathbb{R}^d\\}\\) is the tangent space of \\(\\mathbb{S}_r^{d + 1}\\). Further, the spherical exponential mapping \\(\\mathcal{E}\\mathcal{X}\\mathcal{P}:\\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times d}\\to \\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times (d + 1)}\\) for an Euclidean feature map \\(\\Phi\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.885, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E X P} (\\Phi) [ i, j,: ] = \\exp_ {N} (\\Phi [ i, j,: ]) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.619, + 0.804, + 0.636 + ], + "angle": 0, + "content": "where \\(\\Phi [i,j,:]\\) is the feature in location \\((i,j)\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.643, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Definition 2 (Spherical logarithmic mapping). Given the spherical feature \\( x \\in \\mathbb{S}_r^{d+1} \\) with \\( \\|x\\| = r \\), we define the spherical logarithmic transform [69] \\( \\log_N: \\mathbb{S}_r^{d+1} \\to T_N \\mathbb{S}_r^{d+1} \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.713, + 0.892, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\log_ {N} (x) = \\frac {\\psi}{\\sin \\psi} (x - N \\cos \\psi), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.751, + 0.894, + 0.813 + ], + "angle": 0, + "content": "where \\(N = (0,\\dots ,0,r)\\in \\mathbb{S}_r^{d + 1}\\) is the north pole, \\(\\psi =\\) \\(\\arg \\cos (N^T x / r^2)\\). Further, the spherical logarithmic mapping \\(\\mathcal{LOG}:\\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times (d + 1)}\\to \\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times d}\\) for a spherical feature map \\(\\Phi\\) is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.822, + 0.892, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L O G} (\\Phi) [ i, j,: ] = \\mathcal {H} \\left(\\log_ {N} \\left(\\Phi [ i, j,: ]\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.847, + 0.864, + 0.864 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}:T_N\\mathbb{S}_r^{d + 1}\\to \\mathbb{R}^d\\) is defined by \\(\\mathcal{H}((v,r)) = v\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Definition 3 (Spherical space distance). Given two spherical feature maps \\(\\Phi_1, \\Phi_2 \\in \\mathbb{R}^{h' \\times w' \\times (d + 1)}\\) with \\(\\| \\Phi_1[i,j,:] \\| =\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12550" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.123 + ], + "angle": 0, + "content": "\\(\\| \\Phi_2[i,j,:]\\| = r\\) for any \\(i,j\\), the spherical space distance of \\(\\Phi_{1}\\) and \\(\\Phi_{2}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.128, + 0.47, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} \\left\\{\\Phi_ {1}, \\Phi_ {2} \\right\\} = \\sum_ {i = 1} ^ {h ^ {\\prime}} \\sum_ {j = 1} ^ {w ^ {\\prime}} 1 - \\frac {1}{r ^ {2}} \\Phi_ {1} [ i, j,: ] ^ {T} \\Phi_ {2} [ i, j,: ]. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.187, + 0.192, + 0.201 + ], + "angle": 0, + "content": "3.1.2 Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.21, + 0.47, + 0.338 + ], + "angle": 0, + "content": "We use the feature extraction of \\( D_{LR} \\) as an example to explain the procedure, and \\( R \\) can be carried out similarly to \\( D_{LR} \\) by changing the subscripts from \\( D \\) to \\( R \\). First, a \\( 3 \\times 3 \\) convolution is used to obtain shallow features embedding \\( \\Phi_D^{(0)} \\). The main body of feature extraction consists of a cascade of \\( P \\) Restormer blocks, and we denote the \\( p \\)-th Restormer block in \\( \\mathcal{E}_D \\) as \\( \\mathcal{R}_D^{(p)} \\), where \\( p = 1, \\dots, P \\). The input of each \\( \\mathcal{R}_D^{(p)} \\) is represented by \\( \\Phi_D^{(p-1)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.473, + 0.614 + ], + "angle": 0, + "content": "At the \\(p\\)-th step of feature extraction, \\(\\Phi_D^{(p-1)}\\) passes through \\(\\mathcal{R}_D^{(p)}\\) to obtain the preliminary extracted feature \\(\\tilde{\\Phi}_D^{(p)}\\). According to the analysis for motivation in Sec. 1, after the multi-head self-attention mechanism, \\(\\tilde{\\Phi}_D^{(p)}\\) should contain shared features in some channels and private features in other channels. Thus, we assume that features in the first \\(\\frac{dim}{2}\\) channels are shared and represent cross-modality information, while features in the latter \\(\\frac{dim}{2}\\) channels are private and represent the characteristics of their own modality. To achieve feature separation and alignment, \\(\\tilde{\\Phi}_D^{(p)}\\) is mapped to the spherical space using the spherical exponential mapping \\(\\mathcal{E}\\mathcal{X}\\mathcal{P}(\\cdot)\\), and then recovered in the feature domain by the spherical logarithmic mapping \\(\\mathcal{L}\\mathcal{O}\\mathcal{G}(\\cdot)\\) after calculating the feature decomposition loss. Finally, the recovered features are re-concatenated along the channel dimension to obtain \\(\\Phi_D^{(p)}\\), which will be input to the next \\(\\mathcal{R}_D^{(p+1)}\\). The feature decomposition loss will be illustrated in subsequent sections. The total feature extraction process of \\(p\\)-th step is:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.619, + 0.47, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\Phi} _ {D} ^ {(p)} = \\mathcal {R} _ {D} ^ {(p)} \\left(\\Phi_ {D} ^ {(p - 1)}\\right) \\tag {6a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.644, + 0.469, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\Phi} _ {D, (p)} ^ {a l i g n} = \\tilde {\\Phi} _ {D} ^ {(p)} \\left[ 0: \\frac {d i m}{2} \\right], \\tilde {\\Phi} _ {D, (p)} ^ {s e p n} = \\tilde {\\Phi} _ {D} ^ {(p)} \\left[ \\frac {d i m}{2}: d i m \\right] \\quad (6 b)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.666, + 0.469, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {D, (p)} ^ {\\text {a l i g n}} = \\mathcal {L O G} \\left(\\mathcal {E X P} \\left(\\tilde {\\Phi} _ {D, (p)} ^ {\\text {a l i g n}}\\right)\\right) \\tag {6c}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.691, + 0.469, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {D, (p)} ^ {s e p n} = \\mathcal {L O G} \\left(\\mathcal {E X P} \\left(\\tilde {\\Phi} _ {D, (p)} ^ {s e p n}\\right)\\right) \\tag {6d}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.717, + 0.469, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {D} ^ {(p)} = \\operatorname {C a t} \\left(\\Phi_ {D, (p)} ^ {\\text {a l i g n}}, \\Phi_ {D, (p)} ^ {\\text {s e p n}}\\right) \\tag {6e}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.47, + 0.807 + ], + "angle": 0, + "content": "where \\(\\{\\tilde{\\Phi}_{D,(p)}^{\\text{align}}, \\tilde{\\Phi}_{D,(p)}^{\\text{sep}}\\}\\) and \\(\\{\\Phi_{D,(p)}^{\\text{align}}, \\Phi_{D,(p)}^{\\text{sep}}\\}\\) are the aligned and separated features after/before calculating the feature decomposition loss, and \\(Cat(\\cdot, \\cdot)\\) is the channel concatenation operator." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.808, + 0.469, + 0.838 + ], + "angle": 0, + "content": "Eventually, the entire encoder feature extraction can be regarded as:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.845, + 0.469, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {D} = \\mathcal {E} _ {D} (D _ {L R}), \\Phi_ {R} = \\mathcal {E} _ {R} (R), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.867, + 0.472, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Phi_D\\) and \\(\\Phi_R\\) are abbreviated forms of \\(\\Phi_D^{(P)}\\) and \\(\\Phi_D^{(P)}\\). Then \\(\\Phi_D\\) and \\(\\Phi_R\\) will be used as the input to the decoder." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.614, + 0.105 + ], + "angle": 0, + "content": "3.1.3 Decoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.255 + ], + "angle": 0, + "content": "According to Sec. 3.1.2, we obtained features \\(\\Phi_D\\) and \\(\\Phi_R\\), each containing shared information across \\(\\frac{dim}{2}\\) channels and private information across the other \\(\\frac{dim}{2}\\) channels. The features that we consider helpful for the HR depth reconstruction task are the full depth features \\(\\Phi_D\\) and the shared RGB features \\(\\Phi_R^{sep_n}\\). Therefore, we concatenate \\(\\Phi_D\\) and \\(\\Phi_R^{sep_n}\\) in channel dimension and input them to \\(\\mathcal{D}_{\\mathcal{D}}\\) to obtain the reconstructed depth image \\(\\hat{D}_{HR}\\), and input \\(\\Phi_R\\) to \\(\\mathcal{D}_{\\mathcal{R}}\\) to obtain the reconstructed RGB image \\(\\hat{R}\\), which is:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.26, + 0.892, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\hat {D} _ {H R} = \\mathcal {D} _ {\\mathcal {D}} \\left(C a t \\left(\\Phi_ {D}, \\Phi_ {R} [ 0: \\frac {d i m}{2} ]\\right)\\right), \\hat {R} = \\mathcal {D} _ {\\mathcal {R}} \\left(\\Phi_ {R}\\right). \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.303, + 0.826, + 0.317 + ], + "angle": 0, + "content": "3.1.4 Spherical Contrast Refinement module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.894, + 0.431 + ], + "angle": 0, + "content": "After decoding, we have obtained a preliminary estimation of the HR depth map \\(\\hat{D}_{HR}\\). However, it potentially has some minor issues, such as blurry edges, noisy surfaces and overtransferred RGB texture. Hence, in the Spherical Contrast Refinement (SCR) module, we target the imperfect issues in \\(\\hat{D}_{HR}\\) by contrastive learning, and the diagram is shown in Fig. 2c." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.895, + 0.602 + ], + "angle": 0, + "content": "Defect patches classifier. Firstly, we artificially synthesize the \"imperfect image dataset\" using the training set. For example, for an \\( m \\times m \\) image patch from the ground truth in the training set, we can add random noise to make the patch noisy, apply Gaussian blur to it blurry, and add the same-location RGB image to make it texture over-transferred. Patches that are not processed can be regarded as perfect. In this way, we obtain a dataset with labels \"noisy\", \"blurry\", \"texture over-transferred\" and \"perfect\" which can be used to train a defect patches classifier (DPC) based on ResNet34 [26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.894, + 0.822 + ], + "angle": 0, + "content": "Positive and negative samples. After obtaining the decoder output \\(\\hat{D}_{HR}\\) in Eq. (8), we randomly crop it into an \\(m \\times m\\) patch \\(\\hat{D}_{HR}^{pat}\\), and input it into the well-trained DPC to obtain the imperfect type of \\(\\hat{D}_{HR}^{pat}\\). After getting the imperfect label, similar to the operation for making the \"imperfect image dataset\", we can transform the corresponding ground truth \\(D_{HR}\\) into an imperfect depth map \\(\\tilde{D}_{HR}\\). Then, we randomly crop \\(N\\) \\(m \\times m\\) patches from \\(\\tilde{D}_{HR}\\) that are different from the position of \\(\\hat{D}_{HR}^{pat}\\), and the corresponding imperfect patches \\(\\tilde{D}_{HR}^{pat}\\), named negative samples, are generated. Meanwhile, we refer to \\(\\hat{D}_{HR}^{pat}\\) as the anchor and the same-location ground truth \\(D_{HR}^{pat}\\) as the positive sample. The \\(k\\)-th positive, anchor and negative samples are represented by \\(\\mu_k^+\\), \\(\\mu_k\\), and \\(\\mu_k^-\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.857 + ], + "angle": 0, + "content": "Spherical contrast refinement. We input \\(\\mu_k^+\\), \\(\\mu_k\\), and \\(\\mu_k^-\\) into \\(\\mathcal{E}_D\\) and get the contrastive refinement loss \\(\\mathcal{L}_{SCR}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.862, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S C R} = \\sum_ {k = 1} ^ {K} \\frac {\\mathcal {S} \\left\\{\\mathcal {E} _ {D} \\left(\\mu_ {k}\\right) , \\mathcal {E} _ {D} \\left(\\mu_ {k} ^ {+}\\right) \\right\\}}{\\sum_ {n = 1} ^ {N} \\mathcal {S} \\left\\{\\mathcal {E} _ {D} \\left(\\mu_ {k}\\right) , \\mathcal {E} _ {D} \\left(\\mu_ {k} ^ {n -}\\right) \\right\\}}, \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12551" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.086, + 0.352, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.087, + 0.617, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.087, + 0.882, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.232, + 0.842, + 0.247 + ], + "angle": 0, + "content": "Figure 3: Exhibition for different settings of SSDNet. Results are evaluated on validation set of NYU V2 for GDSR \\(\\times 8\\)." + }, + { + "type": "image", + "bbox": [ + 0.159, + 0.259, + 0.816, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.444, + 0.663, + 0.459 + ], + "angle": 0, + "content": "Figure 4: Visualization of the aligned and separated features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.472, + 0.606 + ], + "angle": 0, + "content": "where \\(\\mathcal{S}\\{\\cdot ,\\cdot \\}\\) is the spherical space distance in Definition 3, and \\(\\mu_{k}^{n - }\\) is the nth negative sample in \\(\\mu_k^-\\). Through gradient descent, we minimize the distance between \\(\\mu_{k}^{+}\\) and \\(\\mu_{k}\\) features and maximize that between \\(\\mu_{k}\\) and \\(\\mu_{k}^{-}\\) features. This process fine-tunes \\(\\mathcal{E}_D\\). However, since the increased training cost for the SCR module, we incorporate it into the regular network training process every few iterations to strike a balance between training efficiency and effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.617, + 0.216, + 0.634 + ], + "angle": 0, + "content": "3.2. Training loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.471, + 0.718 + ], + "angle": 0, + "content": "The training loss in this paper comprises several components: the depth map reconstruction loss \\(\\mathcal{L}_{pixel}^{D}\\), the RGB reconstruction loss \\(\\mathcal{L}_{pixel}^{R}\\), the feature decomposition loss \\(\\mathcal{L}_{dec}\\), and the spherical contrastive refinement loss \\(\\mathcal{L}_{SCR}\\). We describe each loss separately next." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.472, + 0.856 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{pixel}^{D}\\) ensures that the estimated depth map \\(\\hat{D}_{HR}\\) output by our SSDNet is close to the ground truth depth map \\(D_{HR}\\). \\(\\mathcal{L}_{pixel}^{R}\\) ensures that the output \\(\\hat{R}\\) is close to the input \\(R\\). Although theoretically unrelated to the depth map superresolution task, this loss item is used to guarantee that the semantic information from the RGB image is involved in the shared RGB features, rather than simply generating random noise that is approximate to the depth features to meet the feature decomposition. Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.867, + 0.47, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p i x e l}} ^ {D} = \\sum_ {k = 1} ^ {K} \\| \\hat {D} _ {H R} ^ {(k)} - D _ {H R} ^ {(k)} \\| _ {2} ^ {2}, \\mathcal {L} _ {\\text {p i x e l}} ^ {R} = \\sum_ {k = 1} ^ {K} \\| \\hat {R} ^ {(k)} - R ^ {(k)} \\| _ {2} ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.895, + 0.561 + ], + "angle": 0, + "content": "Regarding the feature decomposition loss \\(\\mathcal{L}_{dec}\\), we utilize the spherical space distance to enhance the similarity between shared features while reducing the similarity between separated features. We define the specific structure of the feature decomposition loss \\(\\mathcal{L}_{dec}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.571, + 0.892, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e c} = \\mathcal {L} _ {a l i g n} - \\left(1 - \\mathcal {L} _ {s e p n}\\right) ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.6, + 0.545, + 0.612 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.621, + 0.9, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e p n}} = \\sum_ {p = 1} ^ {P} \\mathcal {S} \\left\\{\\Phi_ {D, (p)} ^ {\\text {s e p n}}, \\Phi_ {R, (p)} ^ {\\text {s e p n}} \\right\\}, \\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {p = 1} ^ {P} \\mathcal {S} \\left\\{\\Phi_ {D, (p)} ^ {\\text {a l i g n}}, \\Phi_ {R, (p)} ^ {\\text {a l i g n}} \\right\\}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.895, + 0.822 + ], + "angle": 0, + "content": "Unlike \\(\\ell_2\\) distance, the spherical distance captures relative differences without being influenced by scale. For 1D feature maps with three consecutive pixels: \\(f_{1} = [0.4,0.5,0.6]\\), \\(f_{2} = [4,5,6]\\), \\(f_{3} = [0.6,0.5,0.4]\\), \\(f_{1}\\) and \\(f_{2}\\) exhibit similar pixel-wise increase structures and potentially similar extracted features, while \\(f_{3}\\) differs. Thus, \\(\\{f_1,f_2\\}\\) should be closer in distance and aligned, while \\(\\{f_1,f_3\\}\\) should be separated. However, \\(\\ell_2\\{f_1,f_2\\} = 7.90\\), \\(\\ell_2\\{f_1,f_3\\} = 0.28\\) while \\(\\mathcal{S}\\{f_1,f_2\\} = 7.7\\times 10^{-8}\\), \\(\\mathcal{S}\\{f_1,f_3\\} = 0.02\\). Therefore, the spherical distance is a better choice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.824, + 0.895, + 0.901 + ], + "angle": 0, + "content": "When optimizing \\(\\mathcal{L}_{dec}\\), since the value range of spherical distance \\(S\\{\\cdot, \\cdot\\}\\) is \\([0,2]\\), we have \\(\\mathcal{L}_{align} \\to 0\\), \\(\\mathcal{L}_{sepn} \\to 1\\). According to Definition 3, the derivation gives: \\(\\mathcal{C}\\{\\Phi_{D,(p)}^{align}, \\Phi_{R,(p)}^{align}\\} \\to 1\\) and \\(\\mathcal{C}\\{\\Phi_{D,(p)}^{sepn}, \\Phi_{R,(p)}^{sepn}\\} \\to 0\\), where \\(\\mathcal{C}\\) is the cosine similarity (the last term of Eq. (5)). When" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12552" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.089, + 0.088, + 0.885, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.323, + 0.761, + 0.338 + ], + "angle": 0, + "content": "Figure 5: Visual error maps for test image in the Middlebury dataset for \\(8 \\times\\) super-resolution." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.35, + 0.885, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.213, + 0.573, + 0.754, + 0.587 + ], + "angle": 0, + "content": "Figure 6: Visual error maps for test image in the RGBDD dataset for \\(16\\times\\) super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.688 + ], + "angle": 0, + "content": "\\(\\mathcal{C} = 1\\), vectors have an angle of 0, indicating higher similarity. If \\(\\mathcal{C} = 0\\), the vectors are orthogonal, signifying no correlation. By forcing \\(\\mathcal{C} = 1\\) or \\(\\mathcal{C} = 0\\) for aligned/private features mapped to spherical space, we ensure the enhancement/reduction of similarity, which achieves our goal of feature decomposition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.721 + ], + "angle": 0, + "content": "By combining Eqs. (9) to (11), we obtain the total loss used for training, which can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.734, + 0.47, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p i x e l}} ^ {D} + \\alpha_ {1} \\mathcal {L} _ {\\text {p i x e l}} ^ {R} + \\alpha_ {2} \\mathcal {L} _ {\\text {d e c}} + \\alpha_ {3} \\mathcal {L} _ {\\text {S C R}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.469, + 0.797 + ], + "angle": 0, + "content": "Note that \\(\\alpha_{3} = 0\\) if the current epoch does not require the spherical contrast refinement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.814, + 0.201, + 0.831 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.9 + ], + "angle": 0, + "content": "This section will conduct a comprehensive set of experiments that aim to showcase the effectiveness of our model in addressing the GDSR task. Moreover, we will also provide evidence that substantiates the soundness of our SSDNet." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.582, + 0.63 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.895, + 0.881 + ], + "angle": 0, + "content": "Datasets. Our experiments follow the protocol established in [27, 31, 90]. To evaluate our model, we employ four popular benchmarks: Middlebury [29, 56], Lu [45], NYU v2 [57], and RGBDD [27]. Our training and validation sets consist of the first 1000 images of NYU v2 dataset [57], divided into a 9:1 ratio. The last 449 pairs in NYU v2 are utilized as the test dataset. Middlebury [29, 56] (30 pairs), Lu [45] (6 pairs), and RGBDD [27] (405 pairs) are also used as test sets to evaluate the depth map SR ability across different scenes and objects. To furthermore demonstrate the generalization ability of our model to unknown scenarios, we incorporate thereal-world branch of the RGBDD dataset, which comprises 2215/405 pairs of RGB-D images for training/test sets, respectively. Lastly, we use the root-mean-square error (RMSE) metric to evaluate the super-resolution effect of our model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Implementation details. In our experiments, we apply bicu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12553" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.144, + 0.089, + 0.833, + 0.26 + ], + "angle": 0, + "content": "
MethodsMiddlebury [29, 56]NYU V2 [57]Lu [45]RGBDD [27]
×4×8×16×4×8×16×4×8×16×4×8×16
PAC [59]1.322.624.581.893.336.781.202.335.191.251.983.49
CUNet [11]1.102.174.331.923.706.780.912.234.991.181.953.45
DKN [31]1.232.124.241.623.266.510.962.165.111.301.963.42
FDKN [31]1.082.174.501.863.586.960.822.105.051.181.913.41
FDSR [27]1.132.084.391.613.185.861.292.195.001.161.823.06
GraphSR [8]1.112.124.431.793.176.020.922.055.151.301.833.12
DCTNet [90]1.102.054.191.593.165.840.881.854.391.081.743.05
Ours1.021.914.021.603.145.860.801.824.771.041.722.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.264, + 0.895, + 0.296 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons among the SOTA methods and our SSDNet in test datasets. The best and second best values are highlighted by **bold** and **underline**, respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.317, + 0.47, + 0.404 + ], + "angle": 0, + "content": "
Dataset: RGBDD in real-world
MethodsRMSEMethodsRMSEMethodsRMSE
SVLRM [47]8.05DJF [34]7.90DJFR [35]8.01
DKN [31]7.38FDKN [31]7.50FDSR [27]7.50
DCTNet [90]7.37SSDNet7.32
FDSR* [27]5.49DCTNet* [90]5.43SSDNet*5.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.408, + 0.472, + 0.465 + ], + "angle": 0, + "content": "Table 2: Quantitative comparisons among the SOTA methods and SSDNet on real-world branch of RGBDD. The best and second best RMSEs are highlighted by **bold** and **underline**. Model* means that the model has been fine-tuned on the real-world training data." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.479, + 0.47, + 0.515 + ], + "angle": 0, + "content": "
MethodsPACCUNDKNFDKNFDSRGSRDCTOurs
Time (s)0.200.230.210.040.010.920.080.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.548 + ], + "angle": 0, + "content": "Table 3: Results of time-consuming comparison for generating a \\(640 \\times 480\\) HR depth map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.472, + 0.742 + ], + "angle": 0, + "content": "bic down-sampling to the HR depth maps to synthesize the corresponding LR depth maps. During the pre-processing stage, the training samples undergo resizing to \\(128 \\times 128\\). We train the network with a mini-batch size of 56 for 100 epochs, employing the Adam [32] optimizer with an initial learning rate of \\(5 \\times 10^{-3}\\). For training and testing, we employ a PC featuring eight NVIDIA GeForce RTX 3090 GPUs. \\(\\alpha_{1}, \\alpha_{2}\\) and \\(\\alpha_{3}\\) in Eq. (13) set to \\(10^{-2}\\), \\(10^{-3}\\) and \\(10^{-2}\\) in order to balance the magnitudes of each term in the loss. The SCR fine-tuning is performed once every 10 epochs of training. Further details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.753, + 0.365, + 0.77 + ], + "angle": 0, + "content": "4.2. Empirical validation experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Impact of network hyper-parameters. In our proposed SSDNet, the number of Restormer Blocks \\(P\\), and number of dimensions in Restormer \\(C\\), and the number of negative samples in SCR \\(N\\) are crucial factors in enhancing the super-resolution performance. We conduct experiments on the NYU validation set to study the impact of different combinations of \\(P, C, N\\). Initially, we fix any two parameters at \\(C = 64\\), \\(P = 4\\), and \\(N = 8\\), respectively, and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.502 + ], + "angle": 0, + "content": "evaluate the prediction quality of the unfixed parameter at \\( P = 2,3,4,5,6 \\), \\( C = 32,64,96,128 \\), and \\( N = 2,4,8,16 \\). The results are summarized in Fig. 3. We observe that the performance is limited when \\( P < 4 \\), and redundant parameter increases do not yield corresponding improvements when \\( P > 4 \\). Similarly, increasing \\( C \\) to more than 64 does not produce significant benefits but instead increased the training cost and computational expense. For the SCR module, the increase in training burden does not match the improvement in effectiveness when \\( N \\) exceeds 8. Therefore, to ensure a balance between performance and computational cost, we choose \\( P = 4 \\), \\( C = 64 \\), \\( N = 8 \\) for subsequent experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.897, + 0.598 + ], + "angle": 0, + "content": "Visualization of feature decomposition. Aligned and separated features are visualized in Fig. 4. Aligned features extract shared properties (edges/contours) from the depth/RGB image pair. Separated features capture modality-specific details (textures in RGB images, smooth depth distance in depth maps). The visualization aligns with our motivation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.61, + 0.792, + 0.627 + ], + "angle": 0, + "content": "4.3. Comparison with SOTA methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.895, + 0.711 + ], + "angle": 0, + "content": "This section aims to evaluate the performance of our SS-DNet on several popular benchmarks in Sec. 4.1, and compare our results with the state-of-the-art methods, including PAC [59], CUNet [11], DKN [31], FDKN [31], FDSR [27], GraphSR [8], and DCTNet [90]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.895, + 0.821 + ], + "angle": 0, + "content": "Qualitative Comparison. We present comparisons of the error maps with ground truth depth maps among SOTA methods in Figs. 5 and 6. Intuitively, the depth predictions of SSDNet exhibit lower prediction errors and are closer to the ground truth images, especially in terms of perception and restoration for the edges and contours. Further visual comparisons are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Quantitative results for traditional testsets. The quantitative results with scaling factors of 4, 8, and 16 on four test sets are shown in Tab. 1. Compared with existing methods that achieve good results on certain datasets or superresolution factors, our SSDNet produces satisfactory predic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12554" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.089, + 0.813, + 0.212 + ], + "angle": 0, + "content": "
Encoder&DecoderLdecForm of LdecLRpixelSCR Module×4×8×16
Exp. IShared✓S✓✓1.232.103.49
Exp. IIPrivate×S✓✓1.171.943.27
Exp. IIIPrivate✓l2✓✓1.151.833.06
Exp. IVPrivate✓S×✓1.161.893.29
Exp. VPrivate✓S✓×1.181.873.36
OursPrivate✓S✓✓1.041.722.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.148, + 0.216, + 0.82, + 0.23 + ], + "angle": 0, + "content": "Table 4: Results of ablation experiments on the RGBDD test set. Bold indicates the best score in terms of RMSE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.257, + 0.471, + 0.302 + ], + "angle": 0, + "content": "tions on multiple datasets and various super-resolution scales. This indicates that our model has a better super-resolution performance than previous methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.472, + 0.457 + ], + "angle": 0, + "content": "Quantitative results for real-world branch. In addition, following [27, 90], we use the pre-trained \\(4 \\times\\) model to directly test on the real-world branch of the RGBDD dataset to explore the generalization ability in real-world scenarios. We conduct all testing without any additional fine-tuning, and the quantitative results are presented in Tab. 2. Furthermore, we also perform targeted training and testing on this dataset. Regardless of whether fine-tuning or not, SSDNet outperforms previous methods in RMSE, highlighting its powerful generalization ability in handling unknown scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.472, + 0.598 + ], + "angle": 0, + "content": "Parameter & running time comparison. We demonstrate the relationship between the number of model learnable parameters vs. RMSE on the RGBDD dataset for \\(\\times 4\\), \\(\\times 8\\), and \\(\\times 16\\) and Middlebury for \\(\\times 4\\) in Fig. 1. Our model exhibits a clear advantage over existing methods with relatively fewer parameters. The time-consuming comparison is presented in Tab. 3. They both demonstrate the efficiency of our method for the GDSR task, showing the potential for developing lightweight networks and practical applications in the future." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.242, + 0.621 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.675 + ], + "angle": 0, + "content": "In this section, we validate the design rationality of our SSDNet through ablation experiments based on the RGBDD testset, and present the results in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.471, + 0.755 + ], + "angle": 0, + "content": "SSDNet architecture. In Exp. I, we share the encoders and decoders for RGB and depth images, i.e., a unified encoder and decoder are used instead of \\(\\{\\mathcal{E}_D,\\mathcal{E}_R\\}\\) and \\(\\{\\mathcal{D}_D,\\mathcal{D}_R\\}\\). The number of Restormer blocks is increased to ensure that the learnable parameter number is comparable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.472, + 0.834 + ], + "angle": 0, + "content": "Spherical feature decomposition. In Exp. II, we eliminate the entire feature decomposition module, i.e., \\(\\mathcal{L}_{dec}\\) in Eq. (11) will not be employed. In Exp. III, we change the distance measure for \\(\\mathcal{L}_{sep}\\) and \\(\\mathcal{L}_{align}\\) in Eq. (12) from spherical space distance to Euclidean distance, i.e., \\(\\ell_2\\)-loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.836, + 0.472, + 0.883 + ], + "angle": 0, + "content": "Reconstruction RGB. In Exp. IV, We eliminate \\(\\mathcal{L}_{pixel}^{R}\\) in Eq. (10) and explore the change for feature extraction capability without the restriction of reconstructing RGB." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.886, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Spherical contrast refinement. In Exp. V, we remove the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.894, + 0.286 + ], + "angle": 0, + "content": "SCR module and use \\(\\hat{D}_{HR}\\) in Eq. (8) as the final output of our model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.894, + 0.323 + ], + "angle": 0, + "content": "Analysis. Tab. 4 shows that changing settings leads to performance degradation, validating our model design." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Exp. I: Unified encoder/decoder ignores modality-specific features, hindering efficient cross-modal feature extraction and causing the largest degradation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Exp. II: The absence of \\(\\mathcal{L}_{dec}\\) prevents decomposing the shared/private features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.427 + ], + "angle": 0, + "content": "Exp. III: \\(\\ell_2\\) distance partially decomposes features but is less effective than \\(S\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Exp. IV: \\(\\mathcal{L}_{pixel}^{R}\\) preserves semantic information in RGB and avoids excessive adaptation of feature separation/alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Exp. V: Removal of SCR leads to mentioned detailed issues and performance degradation." + }, + { + "type": "title", + "bbox": [ + 0.498, + 0.509, + 0.62, + 0.525 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.719 + ], + "angle": 0, + "content": "We propose a Spherical Space feature Decomposition network (SSDNet) for guided depth map super-resolution, where a Restormer-based encoder is used to extract the global features of the inputs, and the intermediate features will be mapped to the spherical space to complete the separation/alignment of shared/private features, respectively. Finally, a Restormer-based decoder is to reconstruct the HR depth map. The spherical contrast refinement module is then employed to address the possible detail issues, e.g., blurry edges, noisy surfaces and over-transferred RGB texture. The satisfactory output results and the lightweight size demonstrate the superiority of our approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.738, + 0.661, + 0.755 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "This work has been supported by the National Key Research and Development Program of China under grant 2018AAA0102201, the National Natural Science Foundation of China under Grant 61976174 and 12201497, Shaanxi Fundamental Science Research Project for Mathematics and Physics under Grant 22JSQ033, the Fundamental Research Funds for the Central Universities under Grant D5000220060, and partly supported by the Alexander von Humboldt Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12555" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.154 + ], + "angle": 0, + "content": "[1] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. CoRR, abs/2106.06847, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.21 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.211, + 0.472, + 0.264 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.265, + 0.472, + 0.304 + ], + "angle": 0, + "content": "[4] Efrain Castillo-Muniz and Eduardo Bayro-Corrochano. Geometric spherical networks for visual data processing. In IJCNN, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310. CVF / IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[6] Taco Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Convolutional networks for spherical signals. arXiv preprint arXiv:1709.04893, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.4, + 0.471, + 0.425 + ], + "angle": 0, + "content": "[7] Taco S. Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical CNNs. In ICLR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.471, + 0.48 + ], + "angle": 0, + "content": "[8] Riccardo de Lutio, Alexander Becker, Stefano D'Aronco, Stefania Russo, Jan D. Wegner, and Konrad Schindler. Learning graph regularisation for guided super-resolution. In CVPR, pages 1969-1978. IEEE, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.52 + ], + "angle": 0, + "content": "[9] Riccardo de Lutio, Stefano D'Aronco, Jan Dirk Wegner, and Konrad Schindler. Guided super-resolution as pixel-to-pixel transformation. In ICCV, pages 8828-8836. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.521, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[10] Xin Deng and Pier Luigi Dragotti. Deep coupled ISTA network for multi-modal image super-resolution. IEEE Trans. Image Process., 29:1683-1698, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.56, + 0.471, + 0.613 + ], + "angle": 0, + "content": "[11] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Trans. Pattern Anal. Mach. Intell., 43(10):3333-3348, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.471, + 0.653 + ], + "angle": 0, + "content": "[12] James Diebel and Sebastian Thrun. An application of markov random fields to range sensing. In NeurIPS, pages 291-298, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.654, + 0.471, + 0.694 + ], + "angle": 0, + "content": "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image superresolution. In ECCV, pages 184-199. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.695, + 0.471, + 0.748 + ], + "angle": 0, + "content": "[14] Jiangxin Dong, Jinshan Pan, Jimmy S. Ren, Liang Lin, Jinhui Tang, and Ming-Hsuan Yang. Learning spatially variant linear representation models for joint filtering. IEEE Trans. Pattern Anal. Mach. Intell., 44(11):8355-8370, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[15] Xiaoyu Dong, Naoto Yokoya, Longguang Wang, and Tatsumi Uezato. Learning mutual modulation for self-supervised cross-modal super-resolution. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.87 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.871, + 0.471, + 0.898 + ], + "angle": 0, + "content": "[17] David Ferstl, Christian Reinbacher, René Ranftl, Matthias Rüther, and Horst Bischof. Image guided depth upsampling" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.119 + ], + "angle": 0, + "content": "using anisotropic total generalized variation. In ICCV, pages 993-1000, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.894, + 0.173 + ], + "angle": 0, + "content": "[18] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. RSTT: real-time spatial temporal transformer for space-time video super-resolution. CoRR, abs/2203.14186, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.174, + 0.892, + 0.228 + ], + "angle": 0, + "content": "[19] Shuhang Gu, Shi Guo, Wangmeng Zuo, Yunjin Chen, Radu Timofte, Luc Van Gool, and Lei Zhang. Learned dynamic guidance for depth image reconstruction. IEEE Trans. Pattern Anal. Mach. Intell., 42(10):2437-2452, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.229, + 0.894, + 0.281 + ], + "angle": 0, + "content": "[20] Shuhang Gu, Wangmeng Zuo, Shi Guo, Yunjin Chen, Chongyu Chen, and Lei Zhang. Learning dynamic guidance for depth image enhancement. In CVPR, pages 712-721, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.282, + 0.892, + 0.309 + ], + "angle": 0, + "content": "[21] Xiang Gu, Jian Sun, and Zongben Xu. Spherical space domain adaptation with robust pseudo-label loss. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[22] Xiang Gu, Jian Sun, and Zongben Xu. Unsupervised and semi-supervised robust spherical space domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-1, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[23] Chunle Guo, Chongyi Li, Jichang Guo, Runmin Cong, Huazhu Fu, and Ping Han. Hierarchical features driven residual learning for depth map super-resolution. IEEE Trans. Image Process., 28(5):2545-2557, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.417, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[24] Saurabh Gupta, Ross B. Girshick, Pablo Andres Arbeláez, and Jitendra Malik. Learning rich features from RGB-D images for object detection and segmentation. In ECCV, pages 345-360. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.471, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735. CVF / IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.565, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[27] Lingzhi He, Hongguang Zhu, Feng Li, Huihui Bai, Runmin Cong, Chunjie Zhang, Chunyu Lin, Meiqin Liu, and Yao Zhao. Towards fast and accurate real-world depth superresolution: Benchmark dataset and baseline. In CVPR, pages 9229-9238, 2021. 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[28] Olivier J. Henaff. Data-efficient image recognition with contrastive predictive coding. In ICML, volume 119, pages 4182-4192. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[29] Heiko Hirschmüller and Daniel Scharstein. Evaluation of cost functions for stereo matching. In CVPR, 2007. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[30] Tak-Wai Hui, Chen Change Loy, and Xiaou Tang. Depth map super-resolution by deep multi-scale guidance. In ECCV, pages 353-369. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.74, + 0.892, + 0.78 + ], + "angle": 0, + "content": "[31] Beomjun Kim, Jean Ponce, and Bumsub Ham. Deformable kernel networks for joint image filtering. Int. J. Comput. Vis., 129(2):579-600, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.781, + 0.892, + 0.806 + ], + "angle": 0, + "content": "[32] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.807, + 0.892, + 0.847 + ], + "angle": 0, + "content": "[33] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. CoRR, abs/2112.10175, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.847, + 0.892, + 0.887 + ], + "angle": 0, + "content": "[34] Yijun Li, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep joint image filtering. In ECCV, pages 154-169. Springer, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.888, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[35] Yijun Li, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.094, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12556" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.471, + 0.133 + ], + "angle": 0, + "content": "Yang. Joint image filtering with deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell., 41(8):1909-1923, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.471, + 0.174 + ], + "angle": 0, + "content": "[36] Yu Li, Dongbo Min, Minh N. Do, and Jiangbo Lu. Fast guided global interpolation for depth and motion. In ECCV, pages 717-733. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.174, + 0.471, + 0.227 + ], + "angle": 0, + "content": "[37] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCVW, pages 1833-1844. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.228, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[38] Miao Liao, Feixiang Lu, Dingfu Zhou, Sibo Zhang, Wei Li, and Ruigang Yang. DVI: depth guided video inpainting for autonomous driving. In ECCV, pages 1-17. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.269, + 0.471, + 0.308 + ], + "angle": 0, + "content": "[39] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.471, + 0.349 + ], + "angle": 0, + "content": "[40] Ming-Yu Liu, Oncel Tuzel, and Yuichi Taguchi. Joint geodesic upsampling of depth images. In CVPR, pages 169-176, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[41] Xianming Liu, Deming Zhai, Rong Chen, Xiangyang Ji, Debin Zhao, and Wen Gao. Depth super-resolution via joint color-guided internal and external regularizations. IEEE Trans. Image Process., 28(4):1636-1645, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.471, + 0.483 + ], + "angle": 0, + "content": "[43] Jiajun Lu and David A. Forsyth. Sparse depth super resolution. In CVPR, pages 2245-2253, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[44] Jiangbo Lu, Keyang Shi, Dongbo Min, Liang Lin, and Minh N. Do. Cross-based local multipoint filtering. In CVPR, pages 430-437, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[45] Si Lu, Xiaofeng Ren, and Feng Liu. Depth enhancement via low-rank matrix completion. In CVPR, pages 3390-3397, 2014. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.564, + 0.471, + 0.604 + ], + "angle": 0, + "content": "[46] Dongbo Min, Jiangbo Lu, and Minh N. Do. Depth video enhancement based on weighted mode filtering. IEEE Trans. Image Process., 21(3):1176-1190, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[47] Jinshan Pan, Jiangxin Dong, Jimmy S. J. Ren, Liang Lin, Jinhui Tang, and Ming-Hsuan Yang. Spatially variant linear representation models for joint filtering. In CVPR, pages 1702-1711, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.471, + 0.711 + ], + "angle": 0, + "content": "[48] Jaesik Park, Hyeongwoo Kim, Yu-Wing Tai, Michael S. Brown, and In-So Kweon. High quality depth map upsampling for 3d-tof cameras. In ICCV, pages 1623-1630, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.712, + 0.471, + 0.767 + ], + "angle": 0, + "content": "[49] Taesung Park, Alexei A. Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In ECCV (9), volume 12354 of Lecture Notes in Computer Science, pages 319-345. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.767, + 0.471, + 0.807 + ], + "angle": 0, + "content": "[50] Wanli Peng, Hao Pan, He Liu, and Yi Sun. IDA-3D: instance-depth-aware 3d object detection from stereo vision for autonomous driving. In CVPR, pages 13012–13021, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.808, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[51] Haotong Qin, Zhongang Cai, Mingyuan Zhang, Yifu Ding, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Hao Su. Bipoint-net: Binary neural network for point clouds. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.888 + ], + "angle": 0, + "content": "[52] Haotong Qin, Ruihao Gong, Xianglong Liu, Xiao Bai, Jingkuan Song, and Nicu Sebe. Binary neural networks:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.093, + 0.768, + 0.106 + ], + "angle": 0, + "content": "A survey. Pattern Recognition, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.106, + 0.893, + 0.16 + ], + "angle": 0, + "content": "[53] Haotong Qin, Ruihao Gong, Xianglong Liu, Mingzhu Shen, Ziran Wei, Fengwei Yu, and Jingkuan Song. Forward and backward information retention for accurate binary neural networks. In ICCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.16, + 0.893, + 0.2 + ], + "angle": 0, + "content": "[54] Gernot Riegler, David Ferstl, Matthias Rüther, and Horst Bischof. A deep primal-dual network for guided depth superresolution. In BMVC. BMVA Press, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.201, + 0.893, + 0.24 + ], + "angle": 0, + "content": "[55] Gernot Riegler, Matthias Rüther, and Horst Bischof. Atgv-net: Accurate depth super-resolution. In ECCV, pages 268-284. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.241, + 0.893, + 0.266 + ], + "angle": 0, + "content": "[56] Daniel Scharstein and Chris Pal. Learning conditional random fields for stereo. In CVPR, 2007. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.267, + 0.893, + 0.32 + ], + "angle": 0, + "content": "[57] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. In ECCV, pages 746-760. Springer, 2012. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.893, + 0.375 + ], + "angle": 0, + "content": "[58] Xibin Song, Yuchao Dai, Dingfu Zhou, Liu Liu, Wei Li, Hongdong Li, and Ruigang Yang. Channel attention based iterative residual learning for depth map super-resolution. In CVPR, pages 5630-5639, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.893, + 0.416 + ], + "angle": 0, + "content": "[59] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik G. Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In CVPR, pages 11166-11175, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.416, + 0.893, + 0.469 + ], + "angle": 0, + "content": "[60] Baoli Sun, Xinchen Ye, Baopu Li, Haojie Li, Zhihui Wang, and Rui Xu. Learning scene structure guidance via cross-task knowledge transfer for single depth super-resolution. In CVPR, pages 7792-7801, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.893, + 0.51 + ], + "angle": 0, + "content": "[61] Jiaxiang Tang, Xiaokang Chen, and Gang Zeng. Joint implicit image function for guided depth super-resolution. In ACM Multimedia, pages 4390-4399. ACM, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.51, + 0.893, + 0.577 + ], + "angle": 0, + "content": "[62] Qi Tang, Runmin Cong, Ronghui Sheng, Lingzhi He, Dan Zhang, Yao Zhao, and Sam Kwong. Bridgenet: A joint learning network of depth map super-resolution and monocular depth estimation. In ACM Multimedia, pages 2148-2157. ACM, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.578, + 0.893, + 0.645 + ], + "angle": 0, + "content": "[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, volume 139, pages 10347-10357. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.893, + 0.699 + ], + "angle": 0, + "content": "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.893, + 0.753 + ], + "angle": 0, + "content": "[65] Jiakai Wang, Aishan Liu, Xiao Bai, and Xianglong Liu. Universal adversarial patch attack for automatic checkout using perceptual and attentional bias. IEEE Trans. Image Process., 31:598-611, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.755, + 0.893, + 0.821 + ], + "angle": 0, + "content": "[66] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 548-558. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.822, + 0.893, + 0.862 + ], + "angle": 0, + "content": "[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. CoRR, abs/2106.03106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.863, + 0.893, + 0.889 + ], + "angle": 0, + "content": "[68] Yang Wen, Bin Sheng, Ping Li, Weiyao Lin, and David Dagan Feng. Deep color guided coarse-to-fine convolutional network" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.893, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12557" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": " cascade for depth image super-resolution. IEEE Trans. Image Process., 28(2):994-1006, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.12, + 0.472, + 0.173 + ], + "angle": 0, + "content": "[69] Richard C. Wilson, Edwin R. Hancock, Elzbieta Pekalska, and Robert P. W. Duin. Spherical and hyperbolic embeddings of data. IEEE Trans. Pattern Anal. Mach. Intell., 36(11):2255-2269, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.173, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[70] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In CVPR, pages 10551-10560. CVF / IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.227, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[71] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Fast end-to-end trainable guided filter. In CVPR, pages 1838-1847, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.472, + 0.322 + ], + "angle": 0, + "content": "[72] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In ICCV, pages 8372-8381. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.472, + 0.363 + ], + "angle": 0, + "content": "[73] Jun Xie, Rogério Schmidt Feris, and Ming-Ting Sun. Edge-guided single depth image super resolution. IEEE Trans. Image Process., 25(1):428-438, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.472, + 0.415 + ], + "angle": 0, + "content": "[74] Jun Xie, Rogério Schmidt Feris, Shiaw-Shian Yu, and Ming-Ting Sun. Joint super resolution and denoising from a single depth image. IEEE Trans. Multim., 17(9):1525-1537, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.416, + 0.472, + 0.484 + ], + "angle": 0, + "content": "[75] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2J: anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In ICCV, pages 793-802. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.472, + 0.524 + ], + "angle": 0, + "content": "[76] Ruikang Xu, Mingde Yao, and Zhiwei Xiong. Zero-shot dual-lens super-resolution. In CVPR, pages 9130-9139, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.579 + ], + "angle": 0, + "content": "[77] Jingyu Yang, Xinchen Ye, Kun Li, Chunping Hou, and Yao Wang. Color-guided depth recovery from RGB-D data using an adaptive autoregressive model. IEEE Trans. Image Process., 23(8):3443-3458, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.472, + 0.631 + ], + "angle": 0, + "content": "[78] Yuxiang Yang, Qi Cao, Jing Zhang, and Dacheng Tao. CODON: on orchestrating cross-domain attentions for depth super-resolution. Int. J. Comput. Vis., 130(2):267-284, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.472, + 0.687 + ], + "angle": 0, + "content": "[79] Mingde Yao, Dongliang He, Xin Li, Fu Li, and Zhiwei Xiong. Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.472, + 0.727 + ], + "angle": 0, + "content": "[80] Mingde Yao, Dongliang He, Xin Li, Zhihong Pan, and Zhiwei Xiong. Bidirectional translation between uhd-hdr and hd-sdr videos. IEEE Transactions on Multimedia, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.727, + 0.472, + 0.782 + ], + "angle": 0, + "content": "[81] Xinchen Ye, Baoli Sun, Zhihui Wang, Jingyu Yang, Rui Xu, Haojie Li, and Baopu Li. Pmbanet: Progressive multi-branch aggregation network for scene depth super-resolution. IEEE Trans. Image Process., 29:7427-7442, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.782, + 0.472, + 0.835 + ], + "angle": 0, + "content": "[82] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In CVPR, pages 8824-8833. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.835, + 0.472, + 0.891 + ], + "angle": 0, + "content": "[83] Shanxin Yuan, Guillermo Garcia-Hernando, Björn Stenger, Gyeongsik Moon, Ju Yong Chang, Kyoung Mu Lee, Pavlo Molchanov, Jan Kautz, Sina Honari, Liuhao Ge, Junsong Yuan, Xinghao Chen, Guijin Wang, Fan Yang, Kai Akiyama," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.895, + 0.162 + ], + "angle": 0, + "content": "Yang Wu, Qingfu Wan, Meysam Madadi, Sergio Escalera, Shile Li, Dongheui Lee, Jason Oikonomidis, Antonis A. Argyros, and Tae-Kyun Kim. Depth-based 3d hand pose estimation: From current achievements to future goals. In CVPR, pages 2636-2645, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.162, + 0.895, + 0.215 + ], + "angle": 0, + "content": "[84] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, pages 5718-5729. IEEE, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.215, + 0.895, + 0.256 + ], + "angle": 0, + "content": "[85] Jiahui Zhang, Shijian Lu, Fangneng Zhan, and Yingchen Yu. Blind image super-resolution via contrastive representation learning. CoRR, abs/2107.00708, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.256, + 0.895, + 0.296 + ], + "angle": 0, + "content": "[86] Kai Zhang, Luc Van Gool, and Radu Timofte. Deep unfolding network for image super-resolution. In CVPR, pages 3214-3223, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.296, + 0.895, + 0.363 + ], + "angle": 0, + "content": "[87] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In CVPR, pages 5906–5916, June 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.895, + 0.418 + ], + "angle": 0, + "content": "[88] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. arXiv preprint arXiv:2305.11443, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.895, + 0.485 + ], + "angle": 0, + "content": "[89] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. DDFM: denoising diffusion model for multi-modality image fusion. CoRR, abs/2303.06840, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.895, + 0.54 + ], + "angle": 0, + "content": "[90] Zixiang Zhao, Jiangshe Zhang, Shuang Xu, Zudi Lin, and Hanspeter Pfister. Discrete cosine transform network for guided depth map super-resolution. In CVPR, pages 5687-5697. IEEE, 2022. 2, 3, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.895, + 0.621 + ], + "angle": 0, + "content": "[91] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H. S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890. CVF / IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.621, + 0.895, + 0.676 + ], + "angle": 0, + "content": "[92] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, Zhiwen Chen, and Xiangyang Ji. High-resolution depth maps imaging via attention-based hierarchical multi-modal fusion. IEEE Trans. Image Process., 31:648-663, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.895, + 0.717 + ], + "angle": 0, + "content": "[93] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, and Xiangyang Ji. Guided depth map super-resolution: A survey. ACM Computing Surveys, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.895, + 0.77 + ], + "angle": 0, + "content": "[94] Man Zhou, Keyu Yan, Jinshan Pan, Wenqi Ren, Qi Xie, and Xiangyong Cao. Memory-augmented deep unfolding network for guided image super-resolution. Int. J. Comput. Vis., 131(1):215-242, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.77, + 0.895, + 0.811 + ], + "angle": 0, + "content": "[95] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.811, + 0.895, + 0.893 + ], + "angle": 0, + "content": "[96] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rhemann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew W. Fitzgibbon, Charles T. Loop, Christian Theobalt, and Marc Stamminger. Real-time non-rigid reconstruction using an RGB-D camera. ACM Trans. Graph., 33(4):156:1-156:12, 2014. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.092, + 0.895, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12558" + } + ] +] \ No newline at end of file diff --git a/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/55278fbf-9750-4827-94bb-81181843441e_origin.pdf b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/55278fbf-9750-4827-94bb-81181843441e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3432049d8d95a45a4d6684f02ce56fd1eb11b10f --- /dev/null +++ b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/55278fbf-9750-4827-94bb-81181843441e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c85f03a1c08e6a3d0cdcbca43e82a34ea23d3d402c9de7dd5ab57be3d24d888 +size 5576814 diff --git a/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/full.md b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dad00f6fe9c216830b05094bf9bffa784ec19c04 --- /dev/null +++ b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/full.md @@ -0,0 +1,445 @@ +# Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution + +Zixiang Zhao $^{1,2}$ Jiangshe Zhang $^{1*}$ Xiang Gu $^{1}$ Chengli Tan $^{1}$ Shuang Xu $^{3}$ Yulun Zhang $^{2}$ Radu Timofte $^{2,4}$ Luc Van Gool $^{2}$ + +$^{1}$ Xi'an Jiaotong University $^{2}$ Computer Vision Lab, ETH Zürich $^{3}$ Northwestern Polytechnical University $^{4}$ University of Würzburg + +zixiangzhao@stu.xjtu.edu.cn, jszhang@mail.xjtu.edu.cn + +# Abstract + +Guided depth map super-resolution (GDSR), as a hot topic in multi-modal image processing, aims to upsample low-resolution (LR) depth maps with additional information involved in high-resolution (HR) RGB images from the same scene. The critical step of this task is to effectively extract domain-shared and domain-private RGB/depth features. In addition, three detailed issues, namely blurry edges, noisy surfaces, and over-transferred RGB texture, need to be addressed. In this paper, we propose the Spherical Space feature Decomposition Network (SSDNet) to solve the above issues. To better model cross-modality features, Restormer block-based RGB/depth encoders are employed for extracting local-global features. Then, the extracted features are mapped to the spherical space to complete the separation of private features and the alignment of shared features. Shared features of RGB are fused with the depth features to complete the GDSR task. Subsequently, a spherical contrast refinement (SCR) module is proposed to further address the detail issues. Patches that are classified according to imperfect categories are input into the SCR module, where the patch features are pulled closer to the ground truth and pushed away from the corresponding imperfect samples in the spherical feature space via contrastive learning. Extensive experiments demonstrate that our method can achieve state-of-the-art results on four test datasets, as well as successfully generalize to real-world scenes. The code is available at https://github.com/Zhaozixiang1228/GDSR-SSDNet. + +# 1. Introduction + +Depth maps, as images that measure the distance of scene points from the sensor, are widely used in autonomous driving [38, 50, 87-89], pose estimation [75, 83], virtual reality [41, 51, 79, 80], and scene understanding [24, 53, 65, 96]. + +![](images/148f4c1fca23c465fc51ee743897246415b87a76156ba44c5d14189b8d474d86.jpg) + +![](images/cdcaae00eaec76888c6faf45e90c5e001f1b0cedefb6a1ce7c34a6911bd7634d.jpg) +Figure 1: Our SSDNet achieves outstanding performance on the RGBDD dataset for $\times 4$ , $\times 8$ , and $\times 16$ and Middlebury for $\times 4$ while being computationally efficient. + +![](images/c97d9af4227f6916876cfdc7aa9d2c94c7e07289c49449ada4345855463cd9f6.jpg) + +![](images/2cef7c8222f66e500997ef27de203d42c9ab3f59dd4bed2bec394d3c72bfc8d7.jpg) + +However, the depth maps produced by current consumer-level depth sensors, e.g., Time-of-Flight (ToF) and Kinect cameras, often have the disadvantages of low resolution and noise. These disadvantages are insufficient to meet the requirements of advanced computing vision tasks [52, 76, 93]. + +To obtain high-resolution depth maps, we naturally hope to accomplish depth map super-resolution (SR) by transferring mature SR models in the RGB image domain [13, 39, 86] to the depth SR task. However, a potential risk is that RGB SR models tend to focus on reconstructing high-frequency image information, such as details and texture. Conversely, for depth images, the objects' depth information is often textureless and piecewise, and more sensitive to unclear edges and noise. Therefore, it is unreasonable to directly apply RGB SR models to the depth SR task. On the other hand, while acquiring the depth map, it is relatively easy to obtain HR and noise-free RGB images in the same scene. Furthermore, there are statistical co-occurrences between the edges in RGB images and the discontinuities in + +depth maps [55]. Therefore, we hope to use the HR RGB images to provide edges and contour information missing in the depth map, and to fuse the multi-modal information to accomplish the LR depth image upsampling. + +In the era of deep learning (DL), numerous methods have been utilized to learn the mapping between $\mathrm{LR} \rightarrow \mathrm{HR}$ depth maps. These methods succeeded to some extent in modeling the cross-modality features and reconstructing the contour and edge information. However, DL models which rely on natural priors will limit the model's flexibility. Models learning the $\mathrm{LR} \rightarrow \mathrm{HR}$ mapping via data-driven methods are difficult to interpret due to the unclear working mechanism [90]. Thus, effectively extracting and distinguishing domain-shared and domain-private RGB/depth features is still a challenge. While microscopically, the obtained depth images are often plagued by three detail issues: blurry edges, noisy surfaces, and over-transferred RGB texture, all of which affect the display quality of depth maps. + +In response to addressing the above challenges, we expect to limit the solution space by constraining the extracted features and further improve the modeling of the dependencies between different modalities. Based on our observation, RGB images and depth maps contain shared features, such as depth map discontinuities and edge features of RGB objects, which can be aligned in the feature space. In addition, unique private features, such as the distance information of the depth map and the texture of RGB object surface, should be separated conversely. Thus, in the feature space, domain-shared and domain-private features are expected to separate and align respectively, while the distance between HR features and imperfect features with the above-mentioned detail issues should also be pushed away. + +The above goal can be divided into two steps. First, extracting features through an effective encoder, and then selecting an appropriate distance measure to complete the alignment and separation of features. Considering that CNN-based architectures limit feature extraction capabilities due to content-independent convolution kernels and the lack of global information, we employ the Restormer blocks [84], which has been proved to effectively extract features in the low-level vision domain, as the basic unit of our encoder. For the choice of distance measure, we first think of the Euclidean distance, such as $\ell_1$ or $\ell_2$ distance. However, cross-modality features often contain different scales and orders of magnitude, and the Euclidean distance is easily affected by the scale, which makes it difficult to achieve our goal. Recently, with the development of spherical DL models [4, 7, 21, 22], the spherical feature learning is well known by virtue of its advantages over the Euclidean feature learning in many applications, e.g., domain adaptation [21]. Due to the distances between spherical features being regularized, the alignment and separation of features can be done more easily without losing the model's representational + +capacity. + +According to the above analysis, we propose our Spherical Space feature Decomposition Network (SSDNet). The specific workflow is displayed in Fig. 2. Our contribution can be summarized in three-fold: + +- We propose a spherical space feature decomposition framework to model the cross-modality features. The features extracted by the Restormer block are mapped to the spherical space for separation and alignment. This is the first time that the Transformer and spherical-space distance measure are applied to the GDSR task. +- The spherical contrast refinement module, cooperating with the imperfect patch classification and the corresponding contrastive learning branch, is proposed to address the possible detail issues in depth maps. This is also the first time that the contrastive learning technique has been used for the GDSR task. +- Experiments on four GDSR benchmarks and a real scene dataset demonstrate that our method can generate satisfactory HR depth maps in different scenarios and exhibits good generalization ability. + +# 2. Related Work + +In this section, we briefly introduce the GDSR task, and illustrate the Vision Transformer, spherical space DL and contrastive learning techniques utilized in SSDNet. + +# 2.1. GDSR methods + +Image super-resolution is a popular image processing and computer vision task with too many sub-categories, so we only discuss the GDSR task here. Conventional GDSR methods can be divided into local- [40, 43, 44, 46], global- [12, 17, 36, 48, 77] and learning-based [8, 19, 20, 73, 74] methods, etc. They extract cross-modality information by manually designed filters, optimizing equations, or sparse dictionary learning. With the rapid development of DL, GDSR is further promoted by the CNN-based methods [58, 60–62, 68, 81]. Generally, DL-based methods can be categorized in three groups, i.e., auto-encoder (AE) [23, 30, 35, 78, 90, 92], learnable filter [14, 31, 71], algorithm unfolding [10, 11, 54, 94] and unsupervised [9, 15] methods. AE-based methods learn cross-modality features via shared or private encoders and reconstruct HR depth maps with decoders. In the learnable filter group, the convolution kernels used to extract information are set to be context-dependent and spatially-variant, which improves the flexibility of the model. The algorithmic unfolding group establishes interpretable GDSR models by building a bridge between traditional optimization functions and deep learning. Unsupervised methods change the training paradigm that requires paired LR/HR images, thus solving the difficulty of obtaining paired training data. + +# 2.2. Spherical deep learning + +The works [4, 6, 7] introduce the geometrical neural network or the spherical convolution neural networks for analyzing the spherical signals that are rotation invariant. Gu et al. [22] propose the spherical multi-layer perceptron for aligning feature distributions of different domains in spherical feature space for domain adaptation. Our idea of spherical space feature decomposition is inspired by Gu et al. [22] that shows the advantages of spherical feature learning over the Euclidean feature learning for domain adaptation in the "high-level" image classification task. Different from [22], we tackle the "low-level" GDSR task by aligning/separating the shared/private features of paired RGB image and depth map, rather than aligning the feature distributions as in [22]. We further propose the spherical contrast refinement for tackling the issues of blurry edges, noisy surfaces, and overtransferred RGB texture for GDSR. + +# 2.3. Transformer in vision + +Transformer, proposed by Vaswani et al. [64], has now become a popular technology in computer vision. It has achieved dominance in classification [16, 42, 63], object detection [2, 95], segmentation [66, 91], etc. Simultaneously, it has also made progress in low-level vision [1, 18, 33, 67]. Chen et al. [5] proposed IPT based on the standard Transformer and multi-task learning. Liang et al. [37] proposed SwinIR whose core module inherits from Swin Transformer [42]. Recently, Restormer [84] improves the transformer block by the gated-Dconv network and multi-Dconv head transposed attention. By applying self-attention in the feature dimension, the ability of local-global representation learning is maintained while being more friendly to high-resolution input images. + +# 2.4. Contrastive Learning + +Contrastive learning, which aims to address the scarcity of source and target paired data and prioritize self-supervised representation learning [25, 28]. The primary objective is to bring samples close to their positive counterparts and push them away from negative ones within high-dimensional manifolds, and it seeks to increase the distance between samples belonging to different classes. Recently, contrastive learning has gained significant attention in computer vision, particularly in high-level tasks such as object detection [72], video segmentation [3], and object tracking [82]. It has also found applications in low-level tasks such as super-resolution [85], image translation [49], and image dehazing [70]. + +# 2.5. Comparison with existing approaches + +The methods most related to our approach are feature-decomposition DL methods [10, 90], which propose the idea that cross-modality features contain the common and + +private information. In contrast, our method, for the first time, exploits the Transformer model's powerful ability to model globally dependent features. Meanwhile, the distance measure in spherical space is proved to decompose the cross-modality features more effectively than Euclidean distances. Additionally, instead of only focusing on optimizing the $\ell_2$ loss between the reconstructed image and ground truth, we give classification-based fine-tuning for image patches under different imperfect conditions, thus further improving the performance of GDSR. + +# 3. Method + +In this section, we first present the specific architecture of our SSDNet. Then, comprehensive descriptions are provided for each module. Finally, the loss function and training details will also be addressed. + +# 3.1. Overview + +We can define the input LR depth map, RGB image and the HR depth map as $D_{LR} \in \mathbb{R}^{h \times w}$ , $D_{HR} \in \mathbb{R}^{H \times W}$ and $R \in \mathbb{R}^{H \times W \times 3}$ , where $\{H, W\}$ and $\{h, w\}$ are the height and width of input RGB and depth images, respectively. For our SSDNet network, it consists of five modules, namely, the Encoders for RGB and Depth images, the Decoders for RGB and Depth images, and the Spherical Contrast Refinement (SCR) module, which is denoted as $\mathcal{E}_{\mathcal{R}}(\cdot), \mathcal{E}_{\mathcal{D}}(\cdot), \mathcal{D}_{\mathcal{R}}(\cdot), \mathcal{D}_{\mathcal{D}}(\cdot),$ and $S(\cdot)$ , respectively. + +In general, as shown in Fig. 2, SSDNet utilizes $\mathcal{E}_{\mathcal{R}}$ and $\mathcal{E}_{\mathcal{D}}$ to extract features, which are then projected onto the spherical space for feature separation and alignment. Then, the depth features and shared RGB features are input into $\mathcal{D}_{\mathcal{R}}$ to obtain reconstructed depth maps, and RGB features are fed into $\mathcal{E}_{\mathcal{R}}$ to get reconstructed RGB images. Finally, the reconstructed depth maps are subsequently provided to the SCR module to refine details and output the final $D_{HR}$ . + +The basic unit we employed for feature extraction and image restoration is the Restormer block [84]. The reason we opted for the Restormer block in $\mathcal{E}_{\mathcal{R}}$ , $\mathcal{E}_{\mathcal{D}}$ , $\mathcal{D}_{\mathcal{R}}$ and $\mathcal{D}_{\mathcal{D}}$ is because it allows for the extraction of global features from high-resolution input images by utilizing self-attention across feature dimensions [84]. This approach enables the extraction of cross-modality shallow features without significantly increasing computational requirements. For details on the Restormer block architecture, please refer to the supplementary material or the original paper [84]. + +# 3.1.1 Spherical space transform + +We first define the mapping based on Riemannian geometry between Euclidean Space feature maps and Spherical Space feature maps, i.e., $\mathcal{L}\mathcal{O}\mathcal{G}(\cdot)$ and $\mathcal{E}\mathcal{X}\mathcal{P}(\cdot)$ , which are employed in the feature decomposition. A schematic diagram for the + +![](images/f5833979bc14144232c832eca99dd6e62464e5feab18daf25a463c5c68bc1565.jpg) +(a) Workflow of SSDNet +(b) Spherical Space Feature Decomposition Module + +(c) Spherical Contrast Refinement Module +Figure 2: (a) The Architecture of our SSDNet method. First, the LR depth map is input into $\mathcal{E}_{\mathcal{D}}$ , and the RGB image is input into $\mathcal{E}_{\mathcal{R}}$ . The intermediate feature maps are projected onto the spherical space to complete the feature decomposition, i.e., alignment and separation. The shared & private features of the depth map, and the shared features of the RGB image are input into the $\mathcal{D}_{\mathcal{D}}$ to reconstruct HR depth map. The RGB features are fed into $\mathcal{D}_{\mathcal{R}}$ to reconstruct RGB image. (b) Illustration of spherical space feature decomposition. (c) The module for spherical contrast refinement. (d) The mapping between Euclidean Space and Spherical Space. +![](images/ca3fd5ba5dbe8dcd1c44f364fcd0387216831fd0b40c2fc40d0a4249057d8260.jpg) +C: Channel Concatenation +$\mathcal{S}\{\cdot ,\cdot \} :$ Spherical-space Distance + +![](images/e8398cdf5386c8604b0a4cfb2ad9e40e4423b2574deaf82aca19f97d8f1fbd88.jpg) + +(d) Illustration of Spherical transform +![](images/b345c9106e9dbe7e98f34cb91e06025a06c6d18592f26fc7168fb3de5976d52b.jpg) +ATTN: Attention (Feature Extraction) + +mappings is shown in Fig. 2d. The distance measure on the spherical space is also defined. + +Definition 1 (Spherical exponential mapping). Given a vector $v$ in $d$ -dimensional Euclidean Space, we define the $(d + 1)$ -dimensional vector $\bar{v}$ by $\bar{v} = (v,r)$ where $r$ is a hyperparameter named radius. Then the spherical exponential transform [69] $\exp_N: T_N\mathbb{S}_r^{d + 1} \to \mathbb{S}_r^{d + 1}$ is defined as + +$$ +\exp_ {N} (v) = N \cos \theta + \bar {v} \frac {\sin \theta}{\theta}, \tag {1} +$$ + +where $\mathbb{S}_r^{d + 1} = \{x\in \mathbb{R}^{d + 1}:||x|| = r\} \subset \mathbb{R}^{d + 1}$ is the $d$ -dimensional spherical space, $N = (0,\dots ,0,r)\in \mathbb{S}_r^{d + 1}$ is the north pole, $\theta = \frac{\|\vec{v}\|}{r}$ , and $T_{N}\mathbb{S}_{r}^{d + 1} = \{(v,r):v\in$ $\mathbb{R}^d\}$ is the tangent space of $\mathbb{S}_r^{d + 1}$ . Further, the spherical exponential mapping $\mathcal{E}\mathcal{X}\mathcal{P}:\mathbb{R}^{h^{\prime}\times w^{\prime}\times d}\to \mathbb{R}^{h^{\prime}\times w^{\prime}\times (d + 1)}$ for an Euclidean feature map $\Phi$ is defined as + +$$ +\mathcal {E X P} (\Phi) [ i, j,: ] = \exp_ {N} (\Phi [ i, j,: ]) \tag {2} +$$ + +where $\Phi [i,j,:]$ is the feature in location $(i,j)$ . + +Definition 2 (Spherical logarithmic mapping). Given the spherical feature $x \in \mathbb{S}_r^{d+1}$ with $\|x\| = r$ , we define the spherical logarithmic transform [69] $\log_N: \mathbb{S}_r^{d+1} \to T_N \mathbb{S}_r^{d+1}$ by + +$$ +\log_ {N} (x) = \frac {\psi}{\sin \psi} (x - N \cos \psi), \tag {3} +$$ + +where $N = (0,\dots ,0,r)\in \mathbb{S}_r^{d + 1}$ is the north pole, $\psi =$ $\arg \cos (N^T x / r^2)$ . Further, the spherical logarithmic mapping $\mathcal{LOG}:\mathbb{R}^{h^{\prime}\times w^{\prime}\times (d + 1)}\to \mathbb{R}^{h^{\prime}\times w^{\prime}\times d}$ for a spherical feature map $\Phi$ is defined by + +$$ +\mathcal {L O G} (\Phi) [ i, j,: ] = \mathcal {H} \left(\log_ {N} \left(\Phi [ i, j,: ]\right)\right), \tag {4} +$$ + +where $\mathcal{H}:T_N\mathbb{S}_r^{d + 1}\to \mathbb{R}^d$ is defined by $\mathcal{H}((v,r)) = v$ + +Definition 3 (Spherical space distance). Given two spherical feature maps $\Phi_1, \Phi_2 \in \mathbb{R}^{h' \times w' \times (d + 1)}$ with $\| \Phi_1[i,j,:] \| =$ + +$\| \Phi_2[i,j,:]\| = r$ for any $i,j$ , the spherical space distance of $\Phi_{1}$ and $\Phi_{2}$ is defined as + +$$ +\mathcal {S} \left\{\Phi_ {1}, \Phi_ {2} \right\} = \sum_ {i = 1} ^ {h ^ {\prime}} \sum_ {j = 1} ^ {w ^ {\prime}} 1 - \frac {1}{r ^ {2}} \Phi_ {1} [ i, j,: ] ^ {T} \Phi_ {2} [ i, j,: ]. \tag {5} +$$ + +# 3.1.2 Encoder + +We use the feature extraction of $D_{LR}$ as an example to explain the procedure, and $R$ can be carried out similarly to $D_{LR}$ by changing the subscripts from $D$ to $R$ . First, a $3 \times 3$ convolution is used to obtain shallow features embedding $\Phi_D^{(0)}$ . The main body of feature extraction consists of a cascade of $P$ Restormer blocks, and we denote the $p$ -th Restormer block in $\mathcal{E}_D$ as $\mathcal{R}_D^{(p)}$ , where $p = 1, \dots, P$ . The input of each $\mathcal{R}_D^{(p)}$ is represented by $\Phi_D^{(p-1)}$ . + +At the $p$ -th step of feature extraction, $\Phi_D^{(p-1)}$ passes through $\mathcal{R}_D^{(p)}$ to obtain the preliminary extracted feature $\tilde{\Phi}_D^{(p)}$ . According to the analysis for motivation in Sec. 1, after the multi-head self-attention mechanism, $\tilde{\Phi}_D^{(p)}$ should contain shared features in some channels and private features in other channels. Thus, we assume that features in the first $\frac{dim}{2}$ channels are shared and represent cross-modality information, while features in the latter $\frac{dim}{2}$ channels are private and represent the characteristics of their own modality. To achieve feature separation and alignment, $\tilde{\Phi}_D^{(p)}$ is mapped to the spherical space using the spherical exponential mapping $\mathcal{E}\mathcal{X}\mathcal{P}(\cdot)$ , and then recovered in the feature domain by the spherical logarithmic mapping $\mathcal{L}\mathcal{O}\mathcal{G}(\cdot)$ after calculating the feature decomposition loss. Finally, the recovered features are re-concatenated along the channel dimension to obtain $\Phi_D^{(p)}$ , which will be input to the next $\mathcal{R}_D^{(p+1)}$ . The feature decomposition loss will be illustrated in subsequent sections. The total feature extraction process of $p$ -th step is: + +$$ +\tilde {\Phi} _ {D} ^ {(p)} = \mathcal {R} _ {D} ^ {(p)} \left(\Phi_ {D} ^ {(p - 1)}\right) \tag {6a} +$$ + +$$ +\tilde {\Phi} _ {D, (p)} ^ {a l i g n} = \tilde {\Phi} _ {D} ^ {(p)} \left[ 0: \frac {d i m}{2} \right], \tilde {\Phi} _ {D, (p)} ^ {s e p n} = \tilde {\Phi} _ {D} ^ {(p)} \left[ \frac {d i m}{2}: d i m \right] \quad (6 b) +$$ + +$$ +\Phi_ {D, (p)} ^ {\text {a l i g n}} = \mathcal {L O G} \left(\mathcal {E X P} \left(\tilde {\Phi} _ {D, (p)} ^ {\text {a l i g n}}\right)\right) \tag {6c} +$$ + +$$ +\Phi_ {D, (p)} ^ {s e p n} = \mathcal {L O G} \left(\mathcal {E X P} \left(\tilde {\Phi} _ {D, (p)} ^ {s e p n}\right)\right) \tag {6d} +$$ + +$$ +\Phi_ {D} ^ {(p)} = \operatorname {C a t} \left(\Phi_ {D, (p)} ^ {\text {a l i g n}}, \Phi_ {D, (p)} ^ {\text {s e p n}}\right) \tag {6e} +$$ + +where $\{\tilde{\Phi}_{D,(p)}^{\text{align}}, \tilde{\Phi}_{D,(p)}^{\text{sep}}\}$ and $\{\Phi_{D,(p)}^{\text{align}}, \Phi_{D,(p)}^{\text{sep}}\}$ are the aligned and separated features after/before calculating the feature decomposition loss, and $Cat(\cdot, \cdot)$ is the channel concatenation operator. + +Eventually, the entire encoder feature extraction can be regarded as: + +$$ +\Phi_ {D} = \mathcal {E} _ {D} (D _ {L R}), \Phi_ {R} = \mathcal {E} _ {R} (R), \tag {7} +$$ + +where $\Phi_D$ and $\Phi_R$ are abbreviated forms of $\Phi_D^{(P)}$ and $\Phi_D^{(P)}$ . Then $\Phi_D$ and $\Phi_R$ will be used as the input to the decoder. + +# 3.1.3 Decoder + +According to Sec. 3.1.2, we obtained features $\Phi_D$ and $\Phi_R$ , each containing shared information across $\frac{dim}{2}$ channels and private information across the other $\frac{dim}{2}$ channels. The features that we consider helpful for the HR depth reconstruction task are the full depth features $\Phi_D$ and the shared RGB features $\Phi_R^{sep_n}$ . Therefore, we concatenate $\Phi_D$ and $\Phi_R^{sep_n}$ in channel dimension and input them to $\mathcal{D}_{\mathcal{D}}$ to obtain the reconstructed depth image $\hat{D}_{HR}$ , and input $\Phi_R$ to $\mathcal{D}_{\mathcal{R}}$ to obtain the reconstructed RGB image $\hat{R}$ , which is: + +$$ +\hat {D} _ {H R} = \mathcal {D} _ {\mathcal {D}} \left(C a t \left(\Phi_ {D}, \Phi_ {R} [ 0: \frac {d i m}{2} ]\right)\right), \hat {R} = \mathcal {D} _ {\mathcal {R}} \left(\Phi_ {R}\right). \tag {8} +$$ + +# 3.1.4 Spherical Contrast Refinement module + +After decoding, we have obtained a preliminary estimation of the HR depth map $\hat{D}_{HR}$ . However, it potentially has some minor issues, such as blurry edges, noisy surfaces and overtransferred RGB texture. Hence, in the Spherical Contrast Refinement (SCR) module, we target the imperfect issues in $\hat{D}_{HR}$ by contrastive learning, and the diagram is shown in Fig. 2c. + +Defect patches classifier. Firstly, we artificially synthesize the "imperfect image dataset" using the training set. For example, for an $m \times m$ image patch from the ground truth in the training set, we can add random noise to make the patch noisy, apply Gaussian blur to it blurry, and add the same-location RGB image to make it texture over-transferred. Patches that are not processed can be regarded as perfect. In this way, we obtain a dataset with labels "noisy", "blurry", "texture over-transferred" and "perfect" which can be used to train a defect patches classifier (DPC) based on ResNet34 [26]. + +Positive and negative samples. After obtaining the decoder output $\hat{D}_{HR}$ in Eq. (8), we randomly crop it into an $m \times m$ patch $\hat{D}_{HR}^{pat}$ , and input it into the well-trained DPC to obtain the imperfect type of $\hat{D}_{HR}^{pat}$ . After getting the imperfect label, similar to the operation for making the "imperfect image dataset", we can transform the corresponding ground truth $D_{HR}$ into an imperfect depth map $\tilde{D}_{HR}$ . Then, we randomly crop $N$ $m \times m$ patches from $\tilde{D}_{HR}$ that are different from the position of $\hat{D}_{HR}^{pat}$ , and the corresponding imperfect patches $\tilde{D}_{HR}^{pat}$ , named negative samples, are generated. Meanwhile, we refer to $\hat{D}_{HR}^{pat}$ as the anchor and the same-location ground truth $D_{HR}^{pat}$ as the positive sample. The $k$ -th positive, anchor and negative samples are represented by $\mu_k^+$ , $\mu_k$ , and $\mu_k^-$ , respectively. + +Spherical contrast refinement. We input $\mu_k^+$ , $\mu_k$ , and $\mu_k^-$ into $\mathcal{E}_D$ and get the contrastive refinement loss $\mathcal{L}_{SCR}$ : + +$$ +\mathcal {L} _ {S C R} = \sum_ {k = 1} ^ {K} \frac {\mathcal {S} \left\{\mathcal {E} _ {D} \left(\mu_ {k}\right) , \mathcal {E} _ {D} \left(\mu_ {k} ^ {+}\right) \right\}}{\sum_ {n = 1} ^ {N} \mathcal {S} \left\{\mathcal {E} _ {D} \left(\mu_ {k}\right) , \mathcal {E} _ {D} \left(\mu_ {k} ^ {n -}\right) \right\}}, \tag {9} +$$ + +![](images/c3a13cc767acbe83ee5529d7767d0b296829910fd3439594647d2d42af944cee.jpg) +Figure 3: Exhibition for different settings of SSDNet. Results are evaluated on validation set of NYU V2 for GDSR $\times 8$ . + +![](images/c9662e3b9c28fe80fcb1a999ea16649d5d176e23de01beb25d6bbda70a028650.jpg) + +![](images/be66b245e1187c21f7757e9912351079789181d9f17f400cca992173c5e1dbf5.jpg) + +![](images/76331366bf2358a1cffb1af10dfd18394985fcab9c233436dfc33af1122bd5f4.jpg) +Figure 4: Visualization of the aligned and separated features. + +where $\mathcal{S}\{\cdot ,\cdot \}$ is the spherical space distance in Definition 3, and $\mu_{k}^{n - }$ is the nth negative sample in $\mu_k^-$ . Through gradient descent, we minimize the distance between $\mu_{k}^{+}$ and $\mu_{k}$ features and maximize that between $\mu_{k}$ and $\mu_{k}^{-}$ features. This process fine-tunes $\mathcal{E}_D$ . However, since the increased training cost for the SCR module, we incorporate it into the regular network training process every few iterations to strike a balance between training efficiency and effectiveness. + +# 3.2. Training loss + +The training loss in this paper comprises several components: the depth map reconstruction loss $\mathcal{L}_{pixel}^{D}$ , the RGB reconstruction loss $\mathcal{L}_{pixel}^{R}$ , the feature decomposition loss $\mathcal{L}_{dec}$ , and the spherical contrastive refinement loss $\mathcal{L}_{SCR}$ . We describe each loss separately next. + +$\mathcal{L}_{pixel}^{D}$ ensures that the estimated depth map $\hat{D}_{HR}$ output by our SSDNet is close to the ground truth depth map $D_{HR}$ . $\mathcal{L}_{pixel}^{R}$ ensures that the output $\hat{R}$ is close to the input $R$ . Although theoretically unrelated to the depth map superresolution task, this loss item is used to guarantee that the semantic information from the RGB image is involved in the shared RGB features, rather than simply generating random noise that is approximate to the depth features to meet the feature decomposition. Specifically, + +$$ +\mathcal {L} _ {\text {p i x e l}} ^ {D} = \sum_ {k = 1} ^ {K} \| \hat {D} _ {H R} ^ {(k)} - D _ {H R} ^ {(k)} \| _ {2} ^ {2}, \mathcal {L} _ {\text {p i x e l}} ^ {R} = \sum_ {k = 1} ^ {K} \| \hat {R} ^ {(k)} - R ^ {(k)} \| _ {2} ^ {2}. \tag {10} +$$ + +Regarding the feature decomposition loss $\mathcal{L}_{dec}$ , we utilize the spherical space distance to enhance the similarity between shared features while reducing the similarity between separated features. We define the specific structure of the feature decomposition loss $\mathcal{L}_{dec}$ as follows: + +$$ +\mathcal {L} _ {d e c} = \mathcal {L} _ {a l i g n} - \left(1 - \mathcal {L} _ {s e p n}\right) ^ {2}, \tag {11} +$$ + +where + +$$ +\mathcal {L} _ {\text {s e p n}} = \sum_ {p = 1} ^ {P} \mathcal {S} \left\{\Phi_ {D, (p)} ^ {\text {s e p n}}, \Phi_ {R, (p)} ^ {\text {s e p n}} \right\}, \mathcal {L} _ {\text {a l i g n}} = \sum_ {p = 1} ^ {P} \mathcal {S} \left\{\Phi_ {D, (p)} ^ {\text {a l i g n}}, \Phi_ {R, (p)} ^ {\text {a l i g n}} \right\}. \tag {12} +$$ + +Unlike $\ell_2$ distance, the spherical distance captures relative differences without being influenced by scale. For 1D feature maps with three consecutive pixels: $f_{1} = [0.4,0.5,0.6]$ , $f_{2} = [4,5,6]$ , $f_{3} = [0.6,0.5,0.4]$ , $f_{1}$ and $f_{2}$ exhibit similar pixel-wise increase structures and potentially similar extracted features, while $f_{3}$ differs. Thus, $\{f_1,f_2\}$ should be closer in distance and aligned, while $\{f_1,f_3\}$ should be separated. However, $\ell_2\{f_1,f_2\} = 7.90$ , $\ell_2\{f_1,f_3\} = 0.28$ while $\mathcal{S}\{f_1,f_2\} = 7.7\times 10^{-8}$ , $\mathcal{S}\{f_1,f_3\} = 0.02$ . Therefore, the spherical distance is a better choice. + +When optimizing $\mathcal{L}_{dec}$ , since the value range of spherical distance $S\{\cdot, \cdot\}$ is $[0,2]$ , we have $\mathcal{L}_{align} \to 0$ , $\mathcal{L}_{sepn} \to 1$ . According to Definition 3, the derivation gives: $\mathcal{C}\{\Phi_{D,(p)}^{align}, \Phi_{R,(p)}^{align}\} \to 1$ and $\mathcal{C}\{\Phi_{D,(p)}^{sepn}, \Phi_{R,(p)}^{sepn}\} \to 0$ , where $\mathcal{C}$ is the cosine similarity (the last term of Eq. (5)). When + +![](images/53fdec5ccd7c452001ed4feae975686092e63999dad86a08f65b1fc7286a3f3a.jpg) +Figure 5: Visual error maps for test image in the Middlebury dataset for $8 \times$ super-resolution. + +![](images/a254881d0939d8a4ba38d74da3fbeeea4a6eca5322d6ff349a5f72df2ecdf33f.jpg) +Figure 6: Visual error maps for test image in the RGBDD dataset for $16\times$ super-resolution. + +$\mathcal{C} = 1$ , vectors have an angle of 0, indicating higher similarity. If $\mathcal{C} = 0$ , the vectors are orthogonal, signifying no correlation. By forcing $\mathcal{C} = 1$ or $\mathcal{C} = 0$ for aligned/private features mapped to spherical space, we ensure the enhancement/reduction of similarity, which achieves our goal of feature decomposition. + +By combining Eqs. (9) to (11), we obtain the total loss used for training, which can be expressed as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {p i x e l}} ^ {D} + \alpha_ {1} \mathcal {L} _ {\text {p i x e l}} ^ {R} + \alpha_ {2} \mathcal {L} _ {\text {d e c}} + \alpha_ {3} \mathcal {L} _ {\text {S C R}}. \tag {13} +$$ + +Note that $\alpha_{3} = 0$ if the current epoch does not require the spherical contrast refinement. + +# 4. Experiment + +This section will conduct a comprehensive set of experiments that aim to showcase the effectiveness of our model in addressing the GDSR task. Moreover, we will also provide evidence that substantiates the soundness of our SSDNet. + +# 4.1. Setup + +Datasets. Our experiments follow the protocol established in [27, 31, 90]. To evaluate our model, we employ four popular benchmarks: Middlebury [29, 56], Lu [45], NYU v2 [57], and RGBDD [27]. Our training and validation sets consist of the first 1000 images of NYU v2 dataset [57], divided into a 9:1 ratio. The last 449 pairs in NYU v2 are utilized as the test dataset. Middlebury [29, 56] (30 pairs), Lu [45] (6 pairs), and RGBDD [27] (405 pairs) are also used as test sets to evaluate the depth map SR ability across different scenes and objects. To furthermore demonstrate the generalization ability of our model to unknown scenarios, we incorporate thereal-world branch of the RGBDD dataset, which comprises 2215/405 pairs of RGB-D images for training/test sets, respectively. Lastly, we use the root-mean-square error (RMSE) metric to evaluate the super-resolution effect of our model. + +Implementation details. In our experiments, we apply bicu + +
MethodsMiddlebury [29, 56]NYU V2 [57]Lu [45]RGBDD [27]
×4×8×16×4×8×16×4×8×16×4×8×16
PAC [59]1.322.624.581.893.336.781.202.335.191.251.983.49
CUNet [11]1.102.174.331.923.706.780.912.234.991.181.953.45
DKN [31]1.232.124.241.623.266.510.962.165.111.301.963.42
FDKN [31]1.082.174.501.863.586.960.822.105.051.181.913.41
FDSR [27]1.132.084.391.613.185.861.292.195.001.161.823.06
GraphSR [8]1.112.124.431.793.176.020.922.055.151.301.833.12
DCTNet [90]1.102.054.191.593.165.840.881.854.391.081.743.05
Ours1.021.914.021.603.145.860.801.824.771.041.722.92
+ +Table 1: Quantitative comparisons among the SOTA methods and our SSDNet in test datasets. The best and second best values are highlighted by **bold** and **underline**, respectively. + +
Dataset: RGBDD in real-world
MethodsRMSEMethodsRMSEMethodsRMSE
SVLRM [47]8.05DJF [34]7.90DJFR [35]8.01
DKN [31]7.38FDKN [31]7.50FDSR [27]7.50
DCTNet [90]7.37SSDNet7.32
FDSR* [27]5.49DCTNet* [90]5.43SSDNet*5.38
+ +Table 2: Quantitative comparisons among the SOTA methods and SSDNet on real-world branch of RGBDD. The best and second best RMSEs are highlighted by **bold** and **underline**. Model* means that the model has been fine-tuned on the real-world training data. + +
MethodsPACCUNDKNFDKNFDSRGSRDCTOurs
Time (s)0.200.230.210.040.010.920.080.10
+ +Table 3: Results of time-consuming comparison for generating a $640 \times 480$ HR depth map. + +bic down-sampling to the HR depth maps to synthesize the corresponding LR depth maps. During the pre-processing stage, the training samples undergo resizing to $128 \times 128$ . We train the network with a mini-batch size of 56 for 100 epochs, employing the Adam [32] optimizer with an initial learning rate of $5 \times 10^{-3}$ . For training and testing, we employ a PC featuring eight NVIDIA GeForce RTX 3090 GPUs. $\alpha_{1}, \alpha_{2}$ and $\alpha_{3}$ in Eq. (13) set to $10^{-2}$ , $10^{-3}$ and $10^{-2}$ in order to balance the magnitudes of each term in the loss. The SCR fine-tuning is performed once every 10 epochs of training. Further details are provided in the supplementary material. + +# 4.2. Empirical validation experiment + +Impact of network hyper-parameters. In our proposed SSDNet, the number of Restormer Blocks $P$ , and number of dimensions in Restormer $C$ , and the number of negative samples in SCR $N$ are crucial factors in enhancing the super-resolution performance. We conduct experiments on the NYU validation set to study the impact of different combinations of $P, C, N$ . Initially, we fix any two parameters at $C = 64$ , $P = 4$ , and $N = 8$ , respectively, and + +evaluate the prediction quality of the unfixed parameter at $P = 2,3,4,5,6$ , $C = 32,64,96,128$ , and $N = 2,4,8,16$ . The results are summarized in Fig. 3. We observe that the performance is limited when $P < 4$ , and redundant parameter increases do not yield corresponding improvements when $P > 4$ . Similarly, increasing $C$ to more than 64 does not produce significant benefits but instead increased the training cost and computational expense. For the SCR module, the increase in training burden does not match the improvement in effectiveness when $N$ exceeds 8. Therefore, to ensure a balance between performance and computational cost, we choose $P = 4$ , $C = 64$ , $N = 8$ for subsequent experiments. + +Visualization of feature decomposition. Aligned and separated features are visualized in Fig. 4. Aligned features extract shared properties (edges/contours) from the depth/RGB image pair. Separated features capture modality-specific details (textures in RGB images, smooth depth distance in depth maps). The visualization aligns with our motivation. + +# 4.3. Comparison with SOTA methods + +This section aims to evaluate the performance of our SS-DNet on several popular benchmarks in Sec. 4.1, and compare our results with the state-of-the-art methods, including PAC [59], CUNet [11], DKN [31], FDKN [31], FDSR [27], GraphSR [8], and DCTNet [90]. + +Qualitative Comparison. We present comparisons of the error maps with ground truth depth maps among SOTA methods in Figs. 5 and 6. Intuitively, the depth predictions of SSDNet exhibit lower prediction errors and are closer to the ground truth images, especially in terms of perception and restoration for the edges and contours. Further visual comparisons are shown in the supplementary material. + +Quantitative results for traditional testsets. The quantitative results with scaling factors of 4, 8, and 16 on four test sets are shown in Tab. 1. Compared with existing methods that achieve good results on certain datasets or superresolution factors, our SSDNet produces satisfactory predic + +
Encoder&DecoderLdecForm of LdecLRpixelSCR Module×4×8×16
Exp. IShared✓S✓✓1.232.103.49
Exp. IIPrivate×S✓✓1.171.943.27
Exp. IIIPrivate✓l2✓✓1.151.833.06
Exp. IVPrivate✓S×✓1.161.893.29
Exp. VPrivate✓S✓×1.181.873.36
OursPrivate✓S✓✓1.041.722.92
+ +Table 4: Results of ablation experiments on the RGBDD test set. Bold indicates the best score in terms of RMSE. + +tions on multiple datasets and various super-resolution scales. This indicates that our model has a better super-resolution performance than previous methods. + +Quantitative results for real-world branch. In addition, following [27, 90], we use the pre-trained $4 \times$ model to directly test on the real-world branch of the RGBDD dataset to explore the generalization ability in real-world scenarios. We conduct all testing without any additional fine-tuning, and the quantitative results are presented in Tab. 2. Furthermore, we also perform targeted training and testing on this dataset. Regardless of whether fine-tuning or not, SSDNet outperforms previous methods in RMSE, highlighting its powerful generalization ability in handling unknown scenes. + +Parameter & running time comparison. We demonstrate the relationship between the number of model learnable parameters vs. RMSE on the RGBDD dataset for $\times 4$ , $\times 8$ , and $\times 16$ and Middlebury for $\times 4$ in Fig. 1. Our model exhibits a clear advantage over existing methods with relatively fewer parameters. The time-consuming comparison is presented in Tab. 3. They both demonstrate the efficiency of our method for the GDSR task, showing the potential for developing lightweight networks and practical applications in the future. + +# 4.4. Ablation Studies + +In this section, we validate the design rationality of our SSDNet through ablation experiments based on the RGBDD testset, and present the results in Tab. 4. + +SSDNet architecture. In Exp. I, we share the encoders and decoders for RGB and depth images, i.e., a unified encoder and decoder are used instead of $\{\mathcal{E}_D,\mathcal{E}_R\}$ and $\{\mathcal{D}_D,\mathcal{D}_R\}$ . The number of Restormer blocks is increased to ensure that the learnable parameter number is comparable. + +Spherical feature decomposition. In Exp. II, we eliminate the entire feature decomposition module, i.e., $\mathcal{L}_{dec}$ in Eq. (11) will not be employed. In Exp. III, we change the distance measure for $\mathcal{L}_{sep}$ and $\mathcal{L}_{align}$ in Eq. (12) from spherical space distance to Euclidean distance, i.e., $\ell_2$ -loss. + +Reconstruction RGB. In Exp. IV, We eliminate $\mathcal{L}_{pixel}^{R}$ in Eq. (10) and explore the change for feature extraction capability without the restriction of reconstructing RGB. + +Spherical contrast refinement. In Exp. V, we remove the + +SCR module and use $\hat{D}_{HR}$ in Eq. (8) as the final output of our model. + +Analysis. Tab. 4 shows that changing settings leads to performance degradation, validating our model design. + +Exp. I: Unified encoder/decoder ignores modality-specific features, hindering efficient cross-modal feature extraction and causing the largest degradation. + +Exp. II: The absence of $\mathcal{L}_{dec}$ prevents decomposing the shared/private features. + +Exp. III: $\ell_2$ distance partially decomposes features but is less effective than $S$ . + +Exp. IV: $\mathcal{L}_{pixel}^{R}$ preserves semantic information in RGB and avoids excessive adaptation of feature separation/alignment. + +Exp. V: Removal of SCR leads to mentioned detailed issues and performance degradation. + +# 5. Conclusion + +We propose a Spherical Space feature Decomposition network (SSDNet) for guided depth map super-resolution, where a Restormer-based encoder is used to extract the global features of the inputs, and the intermediate features will be mapped to the spherical space to complete the separation/alignment of shared/private features, respectively. Finally, a Restormer-based decoder is to reconstruct the HR depth map. The spherical contrast refinement module is then employed to address the possible detail issues, e.g., blurry edges, noisy surfaces and over-transferred RGB texture. The satisfactory output results and the lightweight size demonstrate the superiority of our approach. + +# Acknowledgement + +This work has been supported by the National Key Research and Development Program of China under grant 2018AAA0102201, the National Natural Science Foundation of China under Grant 61976174 and 12201497, Shaanxi Fundamental Science Research Project for Mathematics and Physics under Grant 22JSQ033, the Fundamental Research Funds for the Central Universities under Grant D5000220060, and partly supported by the Alexander von Humboldt Foundation. + +# References + +[1] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. CoRR, abs/2106.06847, 2021. 3 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 3 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021. 3 +[4] Efrain Castillo-Muniz and Eduardo Bayro-Corrochano. Geometric spherical networks for visual data processing. In IJCNN, 2012. 2, 3 +[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310. CVF / IEEE, 2021. 3 +[6] Taco Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Convolutional networks for spherical signals. arXiv preprint arXiv:1709.04893, 2017. 3 +[7] Taco S. Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical CNNs. In ICLR, 2018. 2, 3 +[8] Riccardo de Lutio, Alexander Becker, Stefano D'Aronco, Stefania Russo, Jan D. Wegner, and Konrad Schindler. Learning graph regularisation for guided super-resolution. In CVPR, pages 1969-1978. IEEE, 2022. 2, 8 +[9] Riccardo de Lutio, Stefano D'Aronco, Jan Dirk Wegner, and Konrad Schindler. Guided super-resolution as pixel-to-pixel transformation. In ICCV, pages 8828-8836. IEEE, 2019. 2 +[10] Xin Deng and Pier Luigi Dragotti. Deep coupled ISTA network for multi-modal image super-resolution. IEEE Trans. Image Process., 29:1683-1698, 2020. 2, 3 +[11] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Trans. Pattern Anal. Mach. Intell., 43(10):3333-3348, 2021. 2, 8 +[12] James Diebel and Sebastian Thrun. An application of markov random fields to range sensing. In NeurIPS, pages 291-298, 2005. 2 +[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image superresolution. In ECCV, pages 184-199. Springer, 2014. 1 +[14] Jiangxin Dong, Jinshan Pan, Jimmy S. Ren, Liang Lin, Jinhui Tang, and Ming-Hsuan Yang. Learning spatially variant linear representation models for joint filtering. IEEE Trans. Pattern Anal. Mach. Intell., 44(11):8355-8370, 2022. 2 +[15] Xiaoyu Dong, Naoto Yokoya, Longguang Wang, and Tatsumi Uezato. Learning mutual modulation for self-supervised cross-modal super-resolution. In ECCV, 2022. 2 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[17] David Ferstl, Christian Reinbacher, René Ranftl, Matthias Rüther, and Horst Bischof. Image guided depth upsampling + +using anisotropic total generalized variation. In ICCV, pages 993-1000, 2013. 2 +[18] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. RSTT: real-time spatial temporal transformer for space-time video super-resolution. CoRR, abs/2203.14186, 2022. 3 +[19] Shuhang Gu, Shi Guo, Wangmeng Zuo, Yunjin Chen, Radu Timofte, Luc Van Gool, and Lei Zhang. Learned dynamic guidance for depth image reconstruction. IEEE Trans. Pattern Anal. Mach. Intell., 42(10):2437-2452, 2019. 2 +[20] Shuhang Gu, Wangmeng Zuo, Shi Guo, Yunjin Chen, Chongyu Chen, and Lei Zhang. Learning dynamic guidance for depth image enhancement. In CVPR, pages 712-721, 2017. 2 +[21] Xiang Gu, Jian Sun, and Zongben Xu. Spherical space domain adaptation with robust pseudo-label loss. In CVPR, 2020. 2 +[22] Xiang Gu, Jian Sun, and Zongben Xu. Unsupervised and semi-supervised robust spherical space domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-1, 2022. 2, 3 +[23] Chunle Guo, Chongyi Li, Jichang Guo, Runmin Cong, Huazhu Fu, and Ping Han. Hierarchical features driven residual learning for depth map super-resolution. IEEE Trans. Image Process., 28(5):2545-2557, 2019. 2 +[24] Saurabh Gupta, Ross B. Girshick, Pablo Andres Arbeláez, and Jitendra Malik. Learning rich features from RGB-D images for object detection and segmentation. In ECCV, pages 345-360. Springer, 2014. 1 +[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735. CVF / IEEE, 2020. 3 +[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5 +[27] Lingzhi He, Hongguang Zhu, Feng Li, Huihui Bai, Runmin Cong, Chunjie Zhang, Chunyu Lin, Meiqin Liu, and Yao Zhao. Towards fast and accurate real-world depth superresolution: Benchmark dataset and baseline. In CVPR, pages 9229-9238, 2021. 7, 8, 9 +[28] Olivier J. Henaff. Data-efficient image recognition with contrastive predictive coding. In ICML, volume 119, pages 4182-4192. PMLR, 2020. 3 +[29] Heiko Hirschmüller and Daniel Scharstein. Evaluation of cost functions for stereo matching. In CVPR, 2007. 7, 8 +[30] Tak-Wai Hui, Chen Change Loy, and Xiaou Tang. Depth map super-resolution by deep multi-scale guidance. In ECCV, pages 353-369. Springer, 2016. 2 +[31] Beomjun Kim, Jean Ponce, and Bumsub Ham. Deformable kernel networks for joint image filtering. Int. J. Comput. Vis., 129(2):579-600, 2021. 2, 7, 8 +[32] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 8 +[33] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. CoRR, abs/2112.10175, 2021. 3 +[34] Yijun Li, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep joint image filtering. In ECCV, pages 154-169. Springer, 2016. 8 +[35] Yijun Li, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan + +Yang. Joint image filtering with deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell., 41(8):1909-1923, 2019. 2, 8 +[36] Yu Li, Dongbo Min, Minh N. Do, and Jiangbo Lu. Fast guided global interpolation for depth and motion. In ECCV, pages 717-733. Springer, 2016. 2 +[37] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCVW, pages 1833-1844. IEEE, 2021. 3 +[38] Miao Liao, Feixiang Lu, Dingfu Zhou, Sibo Zhang, Wei Li, and Ruigang Yang. DVI: depth guided video inpainting for autonomous driving. In ECCV, pages 1-17. Springer, 2020. 1 +[39] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 1 +[40] Ming-Yu Liu, Oncel Tuzel, and Yuichi Taguchi. Joint geodesic upsampling of depth images. In CVPR, pages 169-176, 2013. 2 +[41] Xianming Liu, Deming Zhai, Rong Chen, Xiangyang Ji, Debin Zhao, and Wen Gao. Depth super-resolution via joint color-guided internal and external regularizations. IEEE Trans. Image Process., 28(4):1636-1645, 2019. 1 +[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002. IEEE, 2021. 3 +[43] Jiajun Lu and David A. Forsyth. Sparse depth super resolution. In CVPR, pages 2245-2253, 2015. 2 +[44] Jiangbo Lu, Keyang Shi, Dongbo Min, Liang Lin, and Minh N. Do. Cross-based local multipoint filtering. In CVPR, pages 430-437, 2012. 2 +[45] Si Lu, Xiaofeng Ren, and Feng Liu. Depth enhancement via low-rank matrix completion. In CVPR, pages 3390-3397, 2014. 7, 8 +[46] Dongbo Min, Jiangbo Lu, and Minh N. Do. Depth video enhancement based on weighted mode filtering. IEEE Trans. Image Process., 21(3):1176-1190, 2012. 2 +[47] Jinshan Pan, Jiangxin Dong, Jimmy S. J. Ren, Liang Lin, Jinhui Tang, and Ming-Hsuan Yang. Spatially variant linear representation models for joint filtering. In CVPR, pages 1702-1711, 2019. 8 +[48] Jaesik Park, Hyeongwoo Kim, Yu-Wing Tai, Michael S. Brown, and In-So Kweon. High quality depth map upsampling for 3d-tof cameras. In ICCV, pages 1623-1630, 2011. 2 +[49] Taesung Park, Alexei A. Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In ECCV (9), volume 12354 of Lecture Notes in Computer Science, pages 319-345. Springer, 2020. 3 +[50] Wanli Peng, Hao Pan, He Liu, and Yi Sun. IDA-3D: instance-depth-aware 3d object detection from stereo vision for autonomous driving. In CVPR, pages 13012–13021, 2020. 1 +[51] Haotong Qin, Zhongang Cai, Mingyuan Zhang, Yifu Ding, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Hao Su. Bipoint-net: Binary neural network for point clouds. In ICLR, 2021. 1 +[52] Haotong Qin, Ruihao Gong, Xianglong Liu, Xiao Bai, Jingkuan Song, and Nicu Sebe. Binary neural networks: + +A survey. Pattern Recognition, 2020. 1 +[53] Haotong Qin, Ruihao Gong, Xianglong Liu, Mingzhu Shen, Ziran Wei, Fengwei Yu, and Jingkuan Song. Forward and backward information retention for accurate binary neural networks. In ICCV, 2020. 1 +[54] Gernot Riegler, David Ferstl, Matthias Rüther, and Horst Bischof. A deep primal-dual network for guided depth superresolution. In BMVC. BMVA Press, 2016. 2 +[55] Gernot Riegler, Matthias Rüther, and Horst Bischof. Atgv-net: Accurate depth super-resolution. In ECCV, pages 268-284. Springer, 2016. 2 +[56] Daniel Scharstein and Chris Pal. Learning conditional random fields for stereo. In CVPR, 2007. 7, 8 +[57] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. In ECCV, pages 746-760. Springer, 2012. 7, 8 +[58] Xibin Song, Yuchao Dai, Dingfu Zhou, Liu Liu, Wei Li, Hongdong Li, and Ruigang Yang. Channel attention based iterative residual learning for depth map super-resolution. In CVPR, pages 5630-5639, 2020. 2 +[59] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik G. Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In CVPR, pages 11166-11175, 2019. 8 +[60] Baoli Sun, Xinchen Ye, Baopu Li, Haojie Li, Zhihui Wang, and Rui Xu. Learning scene structure guidance via cross-task knowledge transfer for single depth super-resolution. In CVPR, pages 7792-7801, 2021. 2 +[61] Jiaxiang Tang, Xiaokang Chen, and Gang Zeng. Joint implicit image function for guided depth super-resolution. In ACM Multimedia, pages 4390-4399. ACM, 2021. +[62] Qi Tang, Runmin Cong, Ronghui Sheng, Lingzhi He, Dan Zhang, Yao Zhao, and Sam Kwong. Bridgenet: A joint learning network of depth map super-resolution and monocular depth estimation. In ACM Multimedia, pages 2148-2157. ACM, 2021. 2 +[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, volume 139, pages 10347-10357. PMLR, 2021. 3 +[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 3 +[65] Jiakai Wang, Aishan Liu, Xiao Bai, and Xianglong Liu. Universal adversarial patch attack for automatic checkout using perceptual and attentional bias. IEEE Trans. Image Process., 31:598-611, 2021. 1 +[66] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 548-558. IEEE, 2021. 3 +[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. CoRR, abs/2106.03106, 2021. 3 +[68] Yang Wen, Bin Sheng, Ping Li, Weiyao Lin, and David Dagan Feng. Deep color guided coarse-to-fine convolutional network + +cascade for depth image super-resolution. IEEE Trans. Image Process., 28(2):994-1006, 2019. 2 +[69] Richard C. Wilson, Edwin R. Hancock, Elzbieta Pekalska, and Robert P. W. Duin. Spherical and hyperbolic embeddings of data. IEEE Trans. Pattern Anal. Mach. Intell., 36(11):2255-2269, 2014. 4 +[70] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In CVPR, pages 10551-10560. CVF / IEEE, 2021. 3 +[71] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Fast end-to-end trainable guided filter. In CVPR, pages 1838-1847, 2018. 2 +[72] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In ICCV, pages 8372-8381. IEEE, 2021. 3 +[73] Jun Xie, Rogério Schmidt Feris, and Ming-Ting Sun. Edge-guided single depth image super resolution. IEEE Trans. Image Process., 25(1):428-438, 2016. 2 +[74] Jun Xie, Rogério Schmidt Feris, Shiaw-Shian Yu, and Ming-Ting Sun. Joint super resolution and denoising from a single depth image. IEEE Trans. Multim., 17(9):1525-1537, 2015. 2 +[75] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2J: anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In ICCV, pages 793-802. IEEE, 2019. 1 +[76] Ruikang Xu, Mingde Yao, and Zhiwei Xiong. Zero-shot dual-lens super-resolution. In CVPR, pages 9130-9139, 2023. 1 +[77] Jingyu Yang, Xinchen Ye, Kun Li, Chunping Hou, and Yao Wang. Color-guided depth recovery from RGB-D data using an adaptive autoregressive model. IEEE Trans. Image Process., 23(8):3443-3458, 2014. 2 +[78] Yuxiang Yang, Qi Cao, Jing Zhang, and Dacheng Tao. CODON: on orchestrating cross-domain attentions for depth super-resolution. Int. J. Comput. Vis., 130(2):267-284, 2022. 2 +[79] Mingde Yao, Dongliang He, Xin Li, Fu Li, and Zhiwei Xiong. Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1 +[80] Mingde Yao, Dongliang He, Xin Li, Zhihong Pan, and Zhiwei Xiong. Bidirectional translation between uhd-hdr and hd-sdr videos. IEEE Transactions on Multimedia, 2023. 1 +[81] Xinchen Ye, Baoli Sun, Zhihui Wang, Jingyu Yang, Rui Xu, Haojie Li, and Baopu Li. Pmbanet: Progressive multi-branch aggregation network for scene depth super-resolution. IEEE Trans. Image Process., 29:7427-7442, 2020. 2 +[82] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In CVPR, pages 8824-8833. IEEE, 2022. 3 +[83] Shanxin Yuan, Guillermo Garcia-Hernando, Björn Stenger, Gyeongsik Moon, Ju Yong Chang, Kyoung Mu Lee, Pavlo Molchanov, Jan Kautz, Sina Honari, Liuhao Ge, Junsong Yuan, Xinghao Chen, Guijin Wang, Fan Yang, Kai Akiyama, + +Yang Wu, Qingfu Wan, Meysam Madadi, Sergio Escalera, Shile Li, Dongheui Lee, Jason Oikonomidis, Antonis A. Argyros, and Tae-Kyun Kim. Depth-based 3d hand pose estimation: From current achievements to future goals. In CVPR, pages 2636-2645, 2018. 1 +[84] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, pages 5718-5729. IEEE, 2022. 2, 3 +[85] Jiahui Zhang, Shijian Lu, Fangneng Zhan, and Yingchen Yu. Blind image super-resolution via contrastive representation learning. CoRR, abs/2107.00708, 2021. 3 +[86] Kai Zhang, Luc Van Gool, and Radu Timofte. Deep unfolding network for image super-resolution. In CVPR, pages 3214-3223, 2020. 1 +[87] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In CVPR, pages 5906–5916, June 2023. 1 +[88] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. arXiv preprint arXiv:2305.11443, 2023. +[89] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. DDFM: denoising diffusion model for multi-modality image fusion. CoRR, abs/2303.06840, 2023. 1 +[90] Zixiang Zhao, Jiangshe Zhang, Shuang Xu, Zudi Lin, and Hanspeter Pfister. Discrete cosine transform network for guided depth map super-resolution. In CVPR, pages 5687-5697. IEEE, 2022. 2, 3, 7, 8, 9 +[91] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H. S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890. CVF / IEEE, 2021. 3 +[92] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, Zhiwen Chen, and Xiangyang Ji. High-resolution depth maps imaging via attention-based hierarchical multi-modal fusion. IEEE Trans. Image Process., 31:648-663, 2022. 2 +[93] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, and Xiangyang Ji. Guided depth map super-resolution: A survey. ACM Computing Surveys, 2023. 1 +[94] Man Zhou, Keyu Yan, Jinshan Pan, Wenqi Ren, Qi Xie, and Xiangyong Cao. Memory-augmented deep unfolding network for guided image super-resolution. Int. J. Comput. Vis., 131(1):215-242, 2023. 2 +[95] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3 +[96] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rhemann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew W. Fitzgibbon, Charles T. Loop, Christian Theobalt, and Marc Stamminger. Real-time non-rigid reconstruction using an RGB-D camera. ACM Trans. Graph., 33(4):156:1-156:12, 2014. 1 \ No newline at end of file diff --git a/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/images.zip b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c54b0809ba47a8f069a9b540c3fdeafca53ace62 --- /dev/null +++ b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9364edb85be34483e84a22b28ca172a967cb4462c8373b6ff0218e5ba218f7a1 +size 846439 diff --git a/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/layout.json b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e8467ddc595c2cee318c6635bd33141fc874b35f --- /dev/null +++ b/2023/Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution/layout.json @@ -0,0 +1,14331 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 47, + 103, + 547, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 547, + 123 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 547, + 123 + ], + "type": "text", + "content": "Spherical Space Feature Decomposition for Guided Depth Map Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "spans": [ + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "text", + "content": "Zixiang Zhao" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "text", + "content": " Jiangshe Zhang" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "text", + "content": " Xiang Gu" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "text", + "content": " Chengli Tan" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "text", + "content": " Shuang Xu" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "text", + "content": " Yulun Zhang" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "text", + "content": " Radu Timofte" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "text", + "content": " Luc Van Gool" + }, + { + "bbox": [ + 139, + 142, + 455, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 176, + 459, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 176, + 459, + 206 + ], + "spans": [ + { + "bbox": [ + 134, + 176, + 459, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 176, + 459, + 206 + ], + "type": "text", + "content": "Xi'an Jiaotong University " + }, + { + "bbox": [ + 134, + 176, + 459, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 176, + 459, + 206 + ], + "type": "text", + "content": "Computer Vision Lab, ETH Zürich " + }, + { + "bbox": [ + 134, + 176, + 459, + 206 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 176, + 459, + 206 + ], + "type": "text", + "content": "Northwestern Polytechnical University " + }, + { + "bbox": [ + 134, + 176, + 459, + 206 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 134, + 176, + 459, + 206 + ], + "type": "text", + "content": "University of Würzburg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 208, + 441, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 208, + 441, + 219 + ], + "spans": [ + { + "bbox": [ + 153, + 208, + 441, + 219 + ], + "type": "text", + "content": "zixiangzhao@stu.xjtu.edu.cn, jszhang@mail.xjtu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 246, + 192, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 246, + 192, + 260 + ], + "spans": [ + { + "bbox": [ + 143, + 246, + 192, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 271, + 290, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 271, + 290, + 606 + ], + "spans": [ + { + "bbox": [ + 45, + 271, + 290, + 606 + ], + "type": "text", + "content": "Guided depth map super-resolution (GDSR), as a hot topic in multi-modal image processing, aims to upsample low-resolution (LR) depth maps with additional information involved in high-resolution (HR) RGB images from the same scene. The critical step of this task is to effectively extract domain-shared and domain-private RGB/depth features. In addition, three detailed issues, namely blurry edges, noisy surfaces, and over-transferred RGB texture, need to be addressed. In this paper, we propose the Spherical Space feature Decomposition Network (SSDNet) to solve the above issues. To better model cross-modality features, Restormer block-based RGB/depth encoders are employed for extracting local-global features. Then, the extracted features are mapped to the spherical space to complete the separation of private features and the alignment of shared features. Shared features of RGB are fused with the depth features to complete the GDSR task. Subsequently, a spherical contrast refinement (SCR) module is proposed to further address the detail issues. Patches that are classified according to imperfect categories are input into the SCR module, where the patch features are pulled closer to the ground truth and pushed away from the corresponding imperfect samples in the spherical feature space via contrastive learning. Extensive experiments demonstrate that our method can achieve state-of-the-art results on four test datasets, as well as successfully generalize to real-world scenes. The code is available at https://github.com/Zhaozixiang1228/GDSR-SSDNet." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 628, + 128, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 128, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 128, + 640 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 648, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 648, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 648, + 288, + 696 + ], + "type": "text", + "content": "Depth maps, as images that measure the distance of scene points from the sensor, are widely used in autonomous driving [38, 50, 87-89], pose estimation [75, 83], virtual reality [41, 51, 79, 80], and scene understanding [24, 53, 65, 96]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 244, + 426, + 338 + ], + "blocks": [ + { + "bbox": [ + 306, + 244, + 426, + 338 + ], + "lines": [ + { + "bbox": [ + 306, + 244, + 426, + 338 + ], + "spans": [ + { + "bbox": [ + 306, + 244, + 426, + 338 + ], + "type": "image", + "image_path": "148f4c1fca23c465fc51ee743897246415b87a76156ba44c5d14189b8d474d86.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 338, + 425, + 423 + ], + "blocks": [ + { + "bbox": [ + 306, + 338, + 425, + 423 + ], + "lines": [ + { + "bbox": [ + 306, + 338, + 425, + 423 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 425, + 423 + ], + "type": "image", + "image_path": "cdcaae00eaec76888c6faf45e90c5e001f1b0cedefb6a1ce7c34a6911bd7634d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "lines": [ + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "type": "text", + "content": "Figure 1: Our SSDNet achieves outstanding performance on the RGBDD dataset for " + }, + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\times 16" + }, + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "type": "text", + "content": " and Middlebury for " + }, + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 305, + 426, + 547, + 460 + ], + "type": "text", + "content": " while being computationally efficient." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 244, + 545, + 338 + ], + "blocks": [ + { + "bbox": [ + 426, + 244, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 426, + 244, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 426, + 244, + 545, + 338 + ], + "type": "image", + "image_path": "c97d9af4227f6916876cfdc7aa9d2c94c7e07289c49449ada4345855463cd9f6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 425, + 338, + 544, + 423 + ], + "blocks": [ + { + "bbox": [ + 425, + 338, + 544, + 423 + ], + "lines": [ + { + "bbox": [ + 425, + 338, + 544, + 423 + ], + "spans": [ + { + "bbox": [ + 425, + 338, + 544, + 423 + ], + "type": "image", + "image_path": "2cef7c8222f66e500997ef27de203d42c9ab3f59dd4bed2bec394d3c72bfc8d7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 484, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 544 + ], + "type": "text", + "content": "However, the depth maps produced by current consumer-level depth sensors, e.g., Time-of-Flight (ToF) and Kinect cameras, often have the disadvantages of low resolution and noise. These disadvantages are insufficient to meet the requirements of advanced computing vision tasks [52, 76, 93]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": "To obtain high-resolution depth maps, we naturally hope to accomplish depth map super-resolution (SR) by transferring mature SR models in the RGB image domain [13, 39, 86] to the depth SR task. However, a potential risk is that RGB SR models tend to focus on reconstructing high-frequency image information, such as details and texture. Conversely, for depth images, the objects' depth information is often textureless and piecewise, and more sensitive to unclear edges and noise. Therefore, it is unreasonable to directly apply RGB SR models to the depth SR task. On the other hand, while acquiring the depth map, it is relatively easy to obtain HR and noise-free RGB images in the same scene. Furthermore, there are statistical co-occurrences between the edges in RGB images and the discontinuities in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12547" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "depth maps [55]. Therefore, we hope to use the HR RGB images to provide edges and contour information missing in the depth map, and to fuse the multi-modal information to accomplish the LR depth image upsampling." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 288, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 288, + 290 + ], + "type": "text", + "content": "In the era of deep learning (DL), numerous methods have been utilized to learn the mapping between " + }, + { + "bbox": [ + 46, + 122, + 288, + 290 + ], + "type": "inline_equation", + "content": "\\mathrm{LR} \\rightarrow \\mathrm{HR}" + }, + { + "bbox": [ + 46, + 122, + 288, + 290 + ], + "type": "text", + "content": " depth maps. These methods succeeded to some extent in modeling the cross-modality features and reconstructing the contour and edge information. However, DL models which rely on natural priors will limit the model's flexibility. Models learning the " + }, + { + "bbox": [ + 46, + 122, + 288, + 290 + ], + "type": "inline_equation", + "content": "\\mathrm{LR} \\rightarrow \\mathrm{HR}" + }, + { + "bbox": [ + 46, + 122, + 288, + 290 + ], + "type": "text", + "content": " mapping via data-driven methods are difficult to interpret due to the unclear working mechanism [90]. Thus, effectively extracting and distinguishing domain-shared and domain-private RGB/depth features is still a challenge. While microscopically, the obtained depth images are often plagued by three detail issues: blurry edges, noisy surfaces, and over-transferred RGB texture, all of which affect the display quality of depth maps." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 288, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 288, + 460 + ], + "type": "text", + "content": "In response to addressing the above challenges, we expect to limit the solution space by constraining the extracted features and further improve the modeling of the dependencies between different modalities. Based on our observation, RGB images and depth maps contain shared features, such as depth map discontinuities and edge features of RGB objects, which can be aligned in the feature space. In addition, unique private features, such as the distance information of the depth map and the texture of RGB object surface, should be separated conversely. Thus, in the feature space, domain-shared and domain-private features are expected to separate and align respectively, while the distance between HR features and imperfect features with the above-mentioned detail issues should also be pushed away." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": "The above goal can be divided into two steps. First, extracting features through an effective encoder, and then selecting an appropriate distance measure to complete the alignment and separation of features. Considering that CNN-based architectures limit feature extraction capabilities due to content-independent convolution kernels and the lack of global information, we employ the Restormer blocks [84], which has been proved to effectively extract features in the low-level vision domain, as the basic unit of our encoder. For the choice of distance measure, we first think of the Euclidean distance, such as " + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": " distance. However, cross-modality features often contain different scales and orders of magnitude, and the Euclidean distance is easily affected by the scale, which makes it difficult to achieve our goal. Recently, with the development of spherical DL models [4, 7, 21, 22], the spherical feature learning is well known by virtue of its advantages over the Euclidean feature learning in many applications, e.g., domain adaptation [21]. Due to the distances between spherical features being regularized, the alignment and separation of features can be done more easily without losing the model's representational" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 73, + 344, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 73, + 344, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 73, + 344, + 84 + ], + "type": "text", + "content": "capacity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 86, + 547, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 86, + 547, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 86, + 547, + 132 + ], + "type": "text", + "content": "According to the above analysis, we propose our Spherical Space feature Decomposition Network (SSDNet). The specific workflow is displayed in Fig. 2. Our contribution can be summarized in three-fold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 134, + 545, + 325 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 316, + 134, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 134, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 134, + 545, + 205 + ], + "type": "text", + "content": "- We propose a spherical space feature decomposition framework to model the cross-modality features. The features extracted by the Restormer block are mapped to the spherical space for separation and alignment. This is the first time that the Transformer and spherical-space distance measure are applied to the GDSR task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 206, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 545, + 276 + ], + "type": "text", + "content": "- The spherical contrast refinement module, cooperating with the imperfect patch classification and the corresponding contrastive learning branch, is proposed to address the possible detail issues in depth maps. This is also the first time that the contrastive learning technique has been used for the GDSR task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 277, + 545, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 545, + 325 + ], + "type": "text", + "content": "- Experiments on four GDSR benchmarks and a real scene dataset demonstrate that our method can generate satisfactory HR depth maps in different scenarios and exhibits good generalization ability." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 339, + 392, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 339, + 392, + 352 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 392, + 352 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 360, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 360, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 360, + 545, + 397 + ], + "type": "text", + "content": "In this section, we briefly introduce the GDSR task, and illustrate the Vision Transformer, spherical space DL and contrastive learning techniques utilized in SSDNet." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 407, + 402, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 402, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 402, + 419 + ], + "type": "text", + "content": "2.1. GDSR methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "Image super-resolution is a popular image processing and computer vision task with too many sub-categories, so we only discuss the GDSR task here. Conventional GDSR methods can be divided into local- [40, 43, 44, 46], global- [12, 17, 36, 48, 77] and learning-based [8, 19, 20, 73, 74] methods, etc. They extract cross-modality information by manually designed filters, optimizing equations, or sparse dictionary learning. With the rapid development of DL, GDSR is further promoted by the CNN-based methods [58, 60–62, 68, 81]. Generally, DL-based methods can be categorized in three groups, i.e., auto-encoder (AE) [23, 30, 35, 78, 90, 92], learnable filter [14, 31, 71], algorithm unfolding [10, 11, 54, 94] and unsupervised [9, 15] methods. AE-based methods learn cross-modality features via shared or private encoders and reconstruct HR depth maps with decoders. In the learnable filter group, the convolution kernels used to extract information are set to be context-dependent and spatially-variant, which improves the flexibility of the model. The algorithmic unfolding group establishes interpretable GDSR models by building a bridge between traditional optimization functions and deep learning. Unsupervised methods change the training paradigm that requires paired LR/HR images, thus solving the difficulty of obtaining paired training data." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12548" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 181, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 181, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 181, + 85 + ], + "type": "text", + "content": "2.2. Spherical deep learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 289, + 281 + ], + "type": "text", + "content": "The works [4, 6, 7] introduce the geometrical neural network or the spherical convolution neural networks for analyzing the spherical signals that are rotation invariant. Gu et al. [22] propose the spherical multi-layer perceptron for aligning feature distributions of different domains in spherical feature space for domain adaptation. Our idea of spherical space feature decomposition is inspired by Gu et al. [22] that shows the advantages of spherical feature learning over the Euclidean feature learning for domain adaptation in the \"high-level\" image classification task. Different from [22], we tackle the \"low-level\" GDSR task by aligning/separating the shared/private features of paired RGB image and depth map, rather than aligning the feature distributions as in [22]. We further propose the spherical contrast refinement for tackling the issues of blurry edges, noisy surfaces, and overtransferred RGB texture for GDSR." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 291, + 172, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 172, + 303 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 172, + 303 + ], + "type": "text", + "content": "2.3. Transformer in vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 310, + 289, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 289, + 477 + ], + "type": "text", + "content": "Transformer, proposed by Vaswani et al. [64], has now become a popular technology in computer vision. It has achieved dominance in classification [16, 42, 63], object detection [2, 95], segmentation [66, 91], etc. Simultaneously, it has also made progress in low-level vision [1, 18, 33, 67]. Chen et al. [5] proposed IPT based on the standard Transformer and multi-task learning. Liang et al. [37] proposed SwinIR whose core module inherits from Swin Transformer [42]. Recently, Restormer [84] improves the transformer block by the gated-Dconv network and multi-Dconv head transposed attention. By applying self-attention in the feature dimension, the ability of local-global representation learning is maintained while being more friendly to high-resolution input images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 487, + 171, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 171, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 171, + 499 + ], + "type": "text", + "content": "2.4. Contrastive Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 289, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 650 + ], + "type": "text", + "content": "Contrastive learning, which aims to address the scarcity of source and target paired data and prioritize self-supervised representation learning [25, 28]. The primary objective is to bring samples close to their positive counterparts and push them away from negative ones within high-dimensional manifolds, and it seeks to increase the distance between samples belonging to different classes. Recently, contrastive learning has gained significant attention in computer vision, particularly in high-level tasks such as object detection [72], video segmentation [3], and object tracking [82]. It has also found applications in low-level tasks such as super-resolution [85], image translation [49], and image dehazing [70]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 247, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 247, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 247, + 672 + ], + "type": "text", + "content": "2.5. Comparison with existing approaches" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "The methods most related to our approach are feature-decomposition DL methods [10, 90], which propose the idea that cross-modality features contain the common and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "type": "text", + "content": "private information. In contrast, our method, for the first time, exploits the Transformer model's powerful ability to model globally dependent features. Meanwhile, the distance measure in spherical space is proved to decompose the cross-modality features more effectively than Euclidean distances. Additionally, instead of only focusing on optimizing the " + }, + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "type": "text", + "content": " loss between the reconstructed image and ground truth, we give classification-based fine-tuning for image patches under different imperfect conditions, thus further improving the performance of GDSR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 205, + 362, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 362, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 362, + 217 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 226, + 547, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 226, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 547, + 273 + ], + "type": "text", + "content": "In this section, we first present the specific architecture of our SSDNet. Then, comprehensive descriptions are provided for each module. Finally, the loss function and training details will also be addressed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 283, + 373, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 373, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 373, + 295 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "text", + "content": "We can define the input LR depth map, RGB image and the HR depth map as " + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "inline_equation", + "content": "D_{LR} \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "inline_equation", + "content": "D_{HR} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\{H, W\\}" + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\{h, w\\}" + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "text", + "content": " are the height and width of input RGB and depth images, respectively. For our SSDNet network, it consists of five modules, namely, the Encoders for RGB and Depth images, the Decoders for RGB and Depth images, and the Spherical Contrast Refinement (SCR) module, which is denoted as " + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{R}}(\\cdot), \\mathcal{E}_{\\mathcal{D}}(\\cdot), \\mathcal{D}_{\\mathcal{R}}(\\cdot), \\mathcal{D}_{\\mathcal{D}}(\\cdot)," + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 304, + 302, + 547, + 411 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "text", + "content": "In general, as shown in Fig. 2, SSDNet utilizes " + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{R}}" + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{D}}" + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "text", + "content": " to extract features, which are then projected onto the spherical space for feature separation and alignment. Then, the depth features and shared RGB features are input into " + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}" + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "text", + "content": " to obtain reconstructed depth maps, and RGB features are fed into " + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{R}}" + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "text", + "content": " to get reconstructed RGB images. Finally, the reconstructed depth maps are subsequently provided to the SCR module to refine details and output the final " + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "inline_equation", + "content": "D_{HR}" + }, + { + "bbox": [ + 304, + 411, + 547, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "type": "text", + "content": "The basic unit we employed for feature extraction and image restoration is the Restormer block [84]. The reason we opted for the Restormer block in " + }, + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{R}}" + }, + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{D}}" + }, + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}" + }, + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{D}}" + }, + { + "bbox": [ + 304, + 507, + 548, + 627 + ], + "type": "text", + "content": " is because it allows for the extraction of global features from high-resolution input images by utilizing self-attention across feature dimensions [84]. This approach enables the extraction of cross-modality shallow features without significantly increasing computational requirements. For details on the Restormer block architecture, please refer to the supplementary material or the original paper [84]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 646, + 451, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 451, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 451, + 658 + ], + "type": "text", + "content": "3.1.1 Spherical space transform" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "We first define the mapping based on Riemannian geometry between Euclidean Space feature maps and Spherical Space feature maps, i.e., " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\mathcal{O}\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\mathcal{X}\\mathcal{P}(\\cdot)" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": ", which are employed in the feature decomposition. A schematic diagram for the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12549" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 85, + 375, + 227 + ], + "blocks": [ + { + "bbox": [ + 168, + 73, + 260, + 83 + ], + "lines": [ + { + "bbox": [ + 168, + 73, + 260, + 83 + ], + "spans": [ + { + "bbox": [ + 168, + 73, + 260, + 83 + ], + "type": "text", + "content": "(a) Workflow of SSDNet" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 85, + 375, + 227 + ], + "lines": [ + { + "bbox": [ + 51, + 85, + 375, + 227 + ], + "spans": [ + { + "bbox": [ + 51, + 85, + 375, + 227 + ], + "type": "image", + "image_path": "f5833979bc14144232c832eca99dd6e62464e5feab18daf25a463c5c68bc1565.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 74, + 541, + 83 + ], + "lines": [ + { + "bbox": [ + 379, + 74, + 541, + 83 + ], + "spans": [ + { + "bbox": [ + 379, + 74, + 541, + 83 + ], + "type": "text", + "content": "(b) Spherical Space Feature Decomposition Module" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 239, + 369, + 407 + ], + "blocks": [ + { + "bbox": [ + 134, + 229, + 293, + 239 + ], + "lines": [ + { + "bbox": [ + 134, + 229, + 293, + 239 + ], + "spans": [ + { + "bbox": [ + 134, + 229, + 293, + 239 + ], + "type": "text", + "content": "(c) Spherical Contrast Refinement Module" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 239, + 369, + 407 + ], + "lines": [ + { + "bbox": [ + 51, + 239, + 369, + 407 + ], + "spans": [ + { + "bbox": [ + 51, + 239, + 369, + 407 + ], + "type": "image", + "image_path": "ca3fd5ba5dbe8dcd1c44f364fcd0387216831fd0b40c2fc40d0a4249057d8260.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 363, + 474, + 374 + ], + "lines": [ + { + "bbox": [ + 386, + 363, + 474, + 374 + ], + "spans": [ + { + "bbox": [ + 386, + 363, + 474, + 374 + ], + "type": "text", + "content": "C: Channel Concatenation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 386, + 388, + 489, + 399 + ], + "lines": [ + { + "bbox": [ + 386, + 388, + 489, + 399 + ], + "spans": [ + { + "bbox": [ + 386, + 388, + 489, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{S}\\{\\cdot ,\\cdot \\} :" + }, + { + "bbox": [ + 386, + 388, + 489, + 399 + ], + "type": "text", + "content": " Spherical-space Distance" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "lines": [ + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "type": "text", + "content": "Figure 2: (a) The Architecture of our SSDNet method. First, the LR depth map is input into " + }, + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "type": "text", + "content": ", and the RGB image is input into " + }, + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "type": "text", + "content": ". The intermediate feature maps are projected onto the spherical space to complete the feature decomposition, i.e., alignment and separation. The shared & private features of the depth map, and the shared features of the RGB image are input into the " + }, + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "type": "text", + "content": " to reconstruct HR depth map. The RGB features are fed into " + }, + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 414, + 547, + 470 + ], + "type": "text", + "content": " to reconstruct RGB image. (b) Illustration of spherical space feature decomposition. (c) The module for spherical contrast refinement. (d) The mapping between Euclidean Space and Spherical Space." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 380, + 90, + 538, + 278 + ], + "blocks": [ + { + "bbox": [ + 380, + 90, + 538, + 278 + ], + "lines": [ + { + "bbox": [ + 380, + 90, + 538, + 278 + ], + "spans": [ + { + "bbox": [ + 380, + 90, + 538, + 278 + ], + "type": "image", + "image_path": "e8398cdf5386c8604b0a4cfb2ad9e40e4423b2574deaf82aca19f97d8f1fbd88.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 380, + 297, + 529, + 355 + ], + "blocks": [ + { + "bbox": [ + 390, + 285, + 531, + 295 + ], + "lines": [ + { + "bbox": [ + 390, + 285, + 531, + 295 + ], + "spans": [ + { + "bbox": [ + 390, + 285, + 531, + 295 + ], + "type": "text", + "content": "(d) Illustration of Spherical transform" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 297, + 529, + 355 + ], + "lines": [ + { + "bbox": [ + 380, + 297, + 529, + 355 + ], + "spans": [ + { + "bbox": [ + 380, + 297, + 529, + 355 + ], + "type": "image", + "image_path": "b345c9106e9dbe7e98f34cb91e06025a06c6d18592f26fc7168fb3de5976d52b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 376, + 501, + 387 + ], + "lines": [ + { + "bbox": [ + 386, + 376, + 501, + 387 + ], + "spans": [ + { + "bbox": [ + 386, + 376, + 501, + 387 + ], + "type": "text", + "content": "ATTN: Attention (Feature Extraction)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 491, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 514 + ], + "type": "text", + "content": "mappings is shown in Fig. 2d. The distance measure on the spherical space is also defined." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "text", + "content": "Definition 1 (Spherical exponential mapping). Given a vector " + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "text", + "content": "-dimensional Euclidean Space, we define the " + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "inline_equation", + "content": "(d + 1)" + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "text", + "content": "-dimensional vector " + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "inline_equation", + "content": "\\bar{v}" + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "inline_equation", + "content": "\\bar{v} = (v,r)" + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "text", + "content": " is a hyperparameter named radius. Then the spherical exponential transform [69] " + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "inline_equation", + "content": "\\exp_N: T_N\\mathbb{S}_r^{d + 1} \\to \\mathbb{S}_r^{d + 1}" + }, + { + "bbox": [ + 46, + 521, + 289, + 582 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 588, + 287, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 588, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 287, + 612 + ], + "type": "interline_equation", + "content": "\\exp_ {N} (v) = N \\cos \\theta + \\bar {v} \\frac {\\sin \\theta}{\\theta}, \\tag {1}", + "image_path": "8f974c760874dbf05ca692f6afc6beee07096fd6af3f8f164dd10be2bbc3fcd1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_r^{d + 1} = \\{x\\in \\mathbb{R}^{d + 1}:||x|| = r\\} \\subset \\mathbb{R}^{d + 1}" + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "text", + "content": " -dimensional spherical space, " + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "inline_equation", + "content": "N = (0,\\dots ,0,r)\\in \\mathbb{S}_r^{d + 1}" + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "text", + "content": " is the north pole, " + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "inline_equation", + "content": "\\theta = \\frac{\\|\\vec{v}\\|}{r}" + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "inline_equation", + "content": "T_{N}\\mathbb{S}_{r}^{d + 1} = \\{(v,r):v\\in" + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d\\}" + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "text", + "content": " is the tangent space of " + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_r^{d + 1}" + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "text", + "content": ". Further, the spherical exponential mapping " + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\mathcal{X}\\mathcal{P}:\\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times d}\\to \\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times (d + 1)}" + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "text", + "content": " for an Euclidean feature map " + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 618, + 288, + 693 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 96, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 96, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {E X P} (\\Phi) [ i, j,: ] = \\exp_ {N} (\\Phi [ i, j,: ]) \\tag {2}", + "image_path": "b8d03a6d416d3c9617a9b526072e90737d56f7cf91d0b74c1c9dba6e5f5710ed.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 490, + 492, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 492, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 492, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 490, + 492, + 503 + ], + "type": "inline_equation", + "content": "\\Phi [i,j,:]" + }, + { + "bbox": [ + 306, + 490, + 492, + 503 + ], + "type": "text", + "content": " is the feature in location " + }, + { + "bbox": [ + 306, + 490, + 492, + 503 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 306, + 490, + 492, + 503 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 509, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 509, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 509, + 545, + 557 + ], + "type": "text", + "content": "Definition 2 (Spherical logarithmic mapping). Given the spherical feature " + }, + { + "bbox": [ + 305, + 509, + 545, + 557 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{S}_r^{d+1}" + }, + { + "bbox": [ + 305, + 509, + 545, + 557 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 509, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\|x\\| = r" + }, + { + "bbox": [ + 305, + 509, + 545, + 557 + ], + "type": "text", + "content": ", we define the spherical logarithmic transform [69] " + }, + { + "bbox": [ + 305, + 509, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\log_N: \\mathbb{S}_r^{d+1} \\to T_N \\mathbb{S}_r^{d+1}" + }, + { + "bbox": [ + 305, + 509, + 545, + 557 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 358, + 564, + 545, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 564, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 358, + 564, + 545, + 587 + ], + "type": "interline_equation", + "content": "\\log_ {N} (x) = \\frac {\\psi}{\\sin \\psi} (x - N \\cos \\psi), \\tag {3}", + "image_path": "0039db19cefa624b1fe7ec041bbb7bd6a4da4c4fef6f6ecc7c578eb78577153e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "inline_equation", + "content": "N = (0,\\dots ,0,r)\\in \\mathbb{S}_r^{d + 1}" + }, + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "text", + "content": " is the north pole, " + }, + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\psi =" + }, + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\arg \\cos (N^T x / r^2)" + }, + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "text", + "content": ". Further, the spherical logarithmic mapping " + }, + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{LOG}:\\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times (d + 1)}\\to \\mathbb{R}^{h^{\\prime}\\times w^{\\prime}\\times d}" + }, + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "text", + "content": " for a spherical feature map " + }, + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 305, + 594, + 547, + 643 + ], + "type": "text", + "content": " is defined by" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 345, + 651, + 545, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 651, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 345, + 651, + 545, + 664 + ], + "type": "interline_equation", + "content": "\\mathcal {L O G} (\\Phi) [ i, j,: ] = \\mathcal {H} \\left(\\log_ {N} \\left(\\Phi [ i, j,: ]\\right)\\right), \\tag {4}", + "image_path": "3d7ac386bee9916bc6193c62dd30768a2cc195415d1cfa699d9196b4d9fb280b.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 670, + 528, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 670, + 528, + 684 + ], + "spans": [ + { + "bbox": [ + 305, + 670, + 528, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 670, + 528, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{H}:T_N\\mathbb{S}_r^{d + 1}\\to \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 670, + 528, + 684 + ], + "type": "text", + "content": " is defined by " + }, + { + "bbox": [ + 305, + 670, + 528, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{H}((v,r)) = v" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Definition 3 (Spherical space distance). Given two spherical feature maps " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_1, \\Phi_2 \\in \\mathbb{R}^{h' \\times w' \\times (d + 1)}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\| \\Phi_1[i,j,:] \\| =" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12550" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "inline_equation", + "content": "\\| \\Phi_2[i,j,:]\\| = r" + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "content": ", the spherical space distance of " + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "inline_equation", + "content": "\\Phi_{1}" + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "inline_equation", + "content": "\\Phi_{2}" + }, + { + "bbox": [ + 47, + 72, + 289, + 97 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 101, + 287, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 101, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 101, + 287, + 135 + ], + "type": "interline_equation", + "content": "\\mathcal {S} \\left\\{\\Phi_ {1}, \\Phi_ {2} \\right\\} = \\sum_ {i = 1} ^ {h ^ {\\prime}} \\sum_ {j = 1} ^ {w ^ {\\prime}} 1 - \\frac {1}{r ^ {2}} \\Phi_ {1} [ i, j,: ] ^ {T} \\Phi_ {2} [ i, j,: ]. \\tag {5}", + "image_path": "2dad15d3faeafea9f1145060e4a857e3794d7b00bc265066866ca0cae5205880.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 148, + 117, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 117, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 117, + 159 + ], + "type": "text", + "content": "3.1.2 Encoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": "We use the feature extraction of " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "D_{LR}" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": " as an example to explain the procedure, and " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": " can be carried out similarly to " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "D_{LR}" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": " by changing the subscripts from " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": ". First, a " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": " convolution is used to obtain shallow features embedding " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "\\Phi_D^{(0)}" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": ". The main body of feature extraction consists of a cascade of " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": " Restormer blocks, and we denote the " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": "-th Restormer block in " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_D" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_D^{(p)}" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "p = 1, \\dots, P" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": ". The input of each " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_D^{(p)}" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": " is represented by " + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "inline_equation", + "content": "\\Phi_D^{(p-1)}" + }, + { + "bbox": [ + 46, + 166, + 287, + 267 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": "At the " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": "-th step of feature extraction, " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\Phi_D^{(p-1)}" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": " passes through " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_D^{(p)}" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": " to obtain the preliminary extracted feature " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\tilde{\\Phi}_D^{(p)}" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": ". According to the analysis for motivation in Sec. 1, after the multi-head self-attention mechanism, " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\tilde{\\Phi}_D^{(p)}" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": " should contain shared features in some channels and private features in other channels. Thus, we assume that features in the first " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\frac{dim}{2}" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": " channels are shared and represent cross-modality information, while features in the latter " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\frac{dim}{2}" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": " channels are private and represent the characteristics of their own modality. To achieve feature separation and alignment, " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\tilde{\\Phi}_D^{(p)}" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": " is mapped to the spherical space using the spherical exponential mapping " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\mathcal{X}\\mathcal{P}(\\cdot)" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": ", and then recovered in the feature domain by the spherical logarithmic mapping " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\mathcal{O}\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": " after calculating the feature decomposition loss. Finally, the recovered features are re-concatenated along the channel dimension to obtain " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\Phi_D^{(p)}" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": ", which will be input to the next " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_D^{(p+1)}" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": ". The feature decomposition loss will be illustrated in subsequent sections. The total feature extraction process of " + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 268, + 289, + 486 + ], + "type": "text", + "content": "-th step is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 490, + 287, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 490, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 66, + 490, + 287, + 508 + ], + "type": "interline_equation", + "content": "\\tilde {\\Phi} _ {D} ^ {(p)} = \\mathcal {R} _ {D} ^ {(p)} \\left(\\Phi_ {D} ^ {(p - 1)}\\right) \\tag {6a}", + "image_path": "5249b7b7e225fbf43fa7bcd6c5d4a074fbef67beb60ae2ca835731d15c829423.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 510, + 287, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 510, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 58, + 510, + 287, + 526 + ], + "type": "interline_equation", + "content": "\\tilde {\\Phi} _ {D, (p)} ^ {a l i g n} = \\tilde {\\Phi} _ {D} ^ {(p)} \\left[ 0: \\frac {d i m}{2} \\right], \\tilde {\\Phi} _ {D, (p)} ^ {s e p n} = \\tilde {\\Phi} _ {D} ^ {(p)} \\left[ \\frac {d i m}{2}: d i m \\right] \\quad (6 b)", + "image_path": "6f144011d16e776b6f7af98b4daf4c55bf5dde8871f3762126530612d81c8e43.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 527, + 287, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 527, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 58, + 527, + 287, + 545 + ], + "type": "interline_equation", + "content": "\\Phi_ {D, (p)} ^ {\\text {a l i g n}} = \\mathcal {L O G} \\left(\\mathcal {E X P} \\left(\\tilde {\\Phi} _ {D, (p)} ^ {\\text {a l i g n}}\\right)\\right) \\tag {6c}", + "image_path": "f831a02be1b87b9f668ca4afee641bb49b13d116b9570b325cd5b27782488652.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 547, + 287, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 547, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 58, + 547, + 287, + 566 + ], + "type": "interline_equation", + "content": "\\Phi_ {D, (p)} ^ {s e p n} = \\mathcal {L O G} \\left(\\mathcal {E X P} \\left(\\tilde {\\Phi} _ {D, (p)} ^ {s e p n}\\right)\\right) \\tag {6d}", + "image_path": "d0ea1ec19ef852e58a1aad728cae8e02310566d8bda18197cb6d7e996b24f022.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 567, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 567, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 66, + 567, + 287, + 586 + ], + "type": "interline_equation", + "content": "\\Phi_ {D} ^ {(p)} = \\operatorname {C a t} \\left(\\Phi_ {D, (p)} ^ {\\text {a l i g n}}, \\Phi_ {D, (p)} ^ {\\text {s e p n}}\\right) \\tag {6e}", + "image_path": "38cb7f836a351c12ed2618ea7ec522ccba293fb912879be01dcf0528c9e47286.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\Phi}_{D,(p)}^{\\text{align}}, \\tilde{\\Phi}_{D,(p)}^{\\text{sep}}\\}" + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\{\\Phi_{D,(p)}^{\\text{align}}, \\Phi_{D,(p)}^{\\text{sep}}\\}" + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": " are the aligned and separated features after/before calculating the feature decomposition loss, and " + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "inline_equation", + "content": "Cat(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": " is the channel concatenation operator." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "content": "Eventually, the entire encoder feature extraction can be regarded as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 95, + 669, + 287, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 669, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 95, + 669, + 287, + 682 + ], + "type": "interline_equation", + "content": "\\Phi_ {D} = \\mathcal {E} _ {D} (D _ {L R}), \\Phi_ {R} = \\mathcal {E} _ {R} (R), \\tag {7}", + "image_path": "f8769962bdd98f727f3bae49e3682b3a3684a9eba854bf23d1c1290e8c117bbf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_D" + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_R" + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "text", + "content": " are abbreviated forms of " + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_D^{(P)}" + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_D^{(P)}" + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_D" + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_R" + }, + { + "bbox": [ + 46, + 686, + 288, + 714 + ], + "type": "text", + "content": " will be used as the input to the decoder." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 375, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 375, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 375, + 83 + ], + "type": "text", + "content": "3.1.3 Decoder" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": "According to Sec. 3.1.2, we obtained features " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\Phi_D" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\Phi_R" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": ", each containing shared information across " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\frac{dim}{2}" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": " channels and private information across the other " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\frac{dim}{2}" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": " channels. The features that we consider helpful for the HR depth reconstruction task are the full depth features " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\Phi_D" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": " and the shared RGB features " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\Phi_R^{sep_n}" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": ". Therefore, we concatenate " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\Phi_D" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\Phi_R^{sep_n}" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": " in channel dimension and input them to " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{D}}" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": " to obtain the reconstructed depth image " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\hat{D}_{HR}" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": ", and input " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\Phi_R" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathcal{R}}" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": " to obtain the reconstructed RGB image " + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\hat{R}" + }, + { + "bbox": [ + 304, + 91, + 547, + 201 + ], + "type": "text", + "content": ", which is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 205, + 545, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 205, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 312, + 205, + 545, + 222 + ], + "type": "interline_equation", + "content": "\\hat {D} _ {H R} = \\mathcal {D} _ {\\mathcal {D}} \\left(C a t \\left(\\Phi_ {D}, \\Phi_ {R} [ 0: \\frac {d i m}{2} ]\\right)\\right), \\hat {R} = \\mathcal {D} _ {\\mathcal {R}} \\left(\\Phi_ {R}\\right). \\tag {8}", + "image_path": "800d9dfad4114829e62c6a4c585eb265d8283641ae3ec9962419e33baf937e27.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 239, + 505, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 505, + 251 + ], + "type": "text", + "content": "3.1.4 Spherical Contrast Refinement module" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 258, + 547, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 547, + 341 + ], + "type": "text", + "content": "After decoding, we have obtained a preliminary estimation of the HR depth map " + }, + { + "bbox": [ + 304, + 258, + 547, + 341 + ], + "type": "inline_equation", + "content": "\\hat{D}_{HR}" + }, + { + "bbox": [ + 304, + 258, + 547, + 341 + ], + "type": "text", + "content": ". However, it potentially has some minor issues, such as blurry edges, noisy surfaces and overtransferred RGB texture. Hence, in the Spherical Contrast Refinement (SCR) module, we target the imperfect issues in " + }, + { + "bbox": [ + 304, + 258, + 547, + 341 + ], + "type": "inline_equation", + "content": "\\hat{D}_{HR}" + }, + { + "bbox": [ + 304, + 258, + 547, + 341 + ], + "type": "text", + "content": " by contrastive learning, and the diagram is shown in Fig. 2c." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 344, + 547, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 547, + 476 + ], + "type": "text", + "content": "Defect patches classifier. Firstly, we artificially synthesize the \"imperfect image dataset\" using the training set. For example, for an " + }, + { + "bbox": [ + 304, + 344, + 547, + 476 + ], + "type": "inline_equation", + "content": "m \\times m" + }, + { + "bbox": [ + 304, + 344, + 547, + 476 + ], + "type": "text", + "content": " image patch from the ground truth in the training set, we can add random noise to make the patch noisy, apply Gaussian blur to it blurry, and add the same-location RGB image to make it texture over-transferred. Patches that are not processed can be regarded as perfect. In this way, we obtain a dataset with labels \"noisy\", \"blurry\", \"texture over-transferred\" and \"perfect\" which can be used to train a defect patches classifier (DPC) based on ResNet34 [26]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": "Positive and negative samples. After obtaining the decoder output " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\hat{D}_{HR}" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": " in Eq. (8), we randomly crop it into an " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "m \\times m" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": " patch " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\hat{D}_{HR}^{pat}" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": ", and input it into the well-trained DPC to obtain the imperfect type of " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\hat{D}_{HR}^{pat}" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": ". After getting the imperfect label, similar to the operation for making the \"imperfect image dataset\", we can transform the corresponding ground truth " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "D_{HR}" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": " into an imperfect depth map " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\tilde{D}_{HR}" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": ". Then, we randomly crop " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "m \\times m" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": " patches from " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\tilde{D}_{HR}" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": " that are different from the position of " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\hat{D}_{HR}^{pat}" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": ", and the corresponding imperfect patches " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\tilde{D}_{HR}^{pat}" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": ", named negative samples, are generated. Meanwhile, we refer to " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\hat{D}_{HR}^{pat}" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": " as the anchor and the same-location ground truth " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "D_{HR}^{pat}" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": " as the positive sample. The " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": "-th positive, anchor and negative samples are represented by " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mu_k^+" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mu_k" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mu_k^-" + }, + { + "bbox": [ + 304, + 479, + 547, + 651 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": "Spherical contrast refinement. We input " + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mu_k^+" + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mu_k" + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mu_k^-" + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_D" + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": " and get the contrastive refinement loss " + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SCR}" + }, + { + "bbox": [ + 304, + 654, + 545, + 678 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 329, + 682, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 682, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 329, + 682, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S C R} = \\sum_ {k = 1} ^ {K} \\frac {\\mathcal {S} \\left\\{\\mathcal {E} _ {D} \\left(\\mu_ {k}\\right) , \\mathcal {E} _ {D} \\left(\\mu_ {k} ^ {+}\\right) \\right\\}}{\\sum_ {n = 1} ^ {N} \\mathcal {S} \\left\\{\\mathcal {E} _ {D} \\left(\\mu_ {k}\\right) , \\mathcal {E} _ {D} \\left(\\mu_ {k} ^ {n -}\\right) \\right\\}}, \\tag {9}", + "image_path": "2ffc0d3593b8a3b40a23713ac3fafaaa39c11345e4e33ec045846cbfaebdf2df.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12551" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 215, + 182 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 215, + 182 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 215, + 182 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 215, + 182 + ], + "type": "image", + "image_path": "c3a13cc767acbe83ee5529d7767d0b296829910fd3439594647d2d42af944cee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 183, + 515, + 195 + ], + "lines": [ + { + "bbox": [ + 77, + 183, + 515, + 195 + ], + "spans": [ + { + "bbox": [ + 77, + 183, + 515, + 195 + ], + "type": "text", + "content": "Figure 3: Exhibition for different settings of SSDNet. Results are evaluated on validation set of NYU V2 for GDSR " + }, + { + "bbox": [ + 77, + 183, + 515, + 195 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 77, + 183, + 515, + 195 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 68, + 377, + 182 + ], + "blocks": [ + { + "bbox": [ + 215, + 68, + 377, + 182 + ], + "lines": [ + { + "bbox": [ + 215, + 68, + 377, + 182 + ], + "spans": [ + { + "bbox": [ + 215, + 68, + 377, + 182 + ], + "type": "image", + "image_path": "c9662e3b9c28fe80fcb1a999ea16649d5d176e23de01beb25d6bbda70a028650.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 378, + 68, + 539, + 182 + ], + "blocks": [ + { + "bbox": [ + 378, + 68, + 539, + 182 + ], + "lines": [ + { + "bbox": [ + 378, + 68, + 539, + 182 + ], + "spans": [ + { + "bbox": [ + 378, + 68, + 539, + 182 + ], + "type": "image", + "image_path": "be66b245e1187c21f7757e9912351079789181d9f17f400cca992173c5e1dbf5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 97, + 205, + 499, + 348 + ], + "blocks": [ + { + "bbox": [ + 97, + 205, + 499, + 348 + ], + "lines": [ + { + "bbox": [ + 97, + 205, + 499, + 348 + ], + "spans": [ + { + "bbox": [ + 97, + 205, + 499, + 348 + ], + "type": "image", + "image_path": "76331366bf2358a1cffb1af10dfd18394985fcab9c233436dfc33af1122bd5f4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 351, + 405, + 363 + ], + "lines": [ + { + "bbox": [ + 187, + 351, + 405, + 363 + ], + "spans": [ + { + "bbox": [ + 187, + 351, + 405, + 363 + ], + "type": "text", + "content": "Figure 4: Visualization of the aligned and separated features." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{S}\\{\\cdot ,\\cdot \\}" + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "text", + "content": " is the spherical space distance in Definition 3, and " + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "inline_equation", + "content": "\\mu_{k}^{n - }" + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "text", + "content": " is the nth negative sample in " + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "inline_equation", + "content": "\\mu_k^-" + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "text", + "content": ". Through gradient descent, we minimize the distance between " + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "inline_equation", + "content": "\\mu_{k}^{+}" + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "text", + "content": " features and maximize that between " + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "inline_equation", + "content": "\\mu_{k}^{-}" + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "text", + "content": " features. This process fine-tunes " + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_D" + }, + { + "bbox": [ + 46, + 384, + 288, + 479 + ], + "type": "text", + "content": ". However, since the increased training cost for the SCR module, we incorporate it into the regular network training process every few iterations to strike a balance between training efficiency and effectiveness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 488, + 132, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 488, + 132, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 488, + 132, + 502 + ], + "type": "text", + "content": "3.2. Training loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "type": "text", + "content": "The training loss in this paper comprises several components: the depth map reconstruction loss " + }, + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pixel}^{D}" + }, + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "type": "text", + "content": ", the RGB reconstruction loss " + }, + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pixel}^{R}" + }, + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "type": "text", + "content": ", the feature decomposition loss " + }, + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}" + }, + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "type": "text", + "content": ", and the spherical contrastive refinement loss " + }, + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SCR}" + }, + { + "bbox": [ + 46, + 508, + 288, + 568 + ], + "type": "text", + "content": ". We describe each loss separately next." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pixel}^{D}" + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": " ensures that the estimated depth map " + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\hat{D}_{HR}" + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": " output by our SSDNet is close to the ground truth depth map " + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "inline_equation", + "content": "D_{HR}" + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pixel}^{R}" + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": " ensures that the output " + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\hat{R}" + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": " is close to the input " + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 569, + 288, + 677 + ], + "type": "text", + "content": ". Although theoretically unrelated to the depth map superresolution task, this loss item is used to guarantee that the semantic information from the RGB image is involved in the shared RGB features, rather than simply generating random noise that is approximate to the depth features to meet the feature decomposition. Specifically," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 51, + 686, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 686, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 51, + 686, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p i x e l}} ^ {D} = \\sum_ {k = 1} ^ {K} \\| \\hat {D} _ {H R} ^ {(k)} - D _ {H R} ^ {(k)} \\| _ {2} ^ {2}, \\mathcal {L} _ {\\text {p i x e l}} ^ {R} = \\sum_ {k = 1} ^ {K} \\| \\hat {R} ^ {(k)} - R ^ {(k)} \\| _ {2} ^ {2}. \\tag {10}", + "image_path": "ff1935ec599ee79e8dfd8bc76df69b1451540f5e8793a24eeaa050a93a70e5f5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 384, + 547, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 444 + ], + "type": "text", + "content": "Regarding the feature decomposition loss " + }, + { + "bbox": [ + 304, + 384, + 547, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}" + }, + { + "bbox": [ + 304, + 384, + 547, + 444 + ], + "type": "text", + "content": ", we utilize the spherical space distance to enhance the similarity between shared features while reducing the similarity between separated features. We define the specific structure of the feature decomposition loss " + }, + { + "bbox": [ + 304, + 384, + 547, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}" + }, + { + "bbox": [ + 304, + 384, + 547, + 444 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 361, + 452, + 545, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 452, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 361, + 452, + 545, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e c} = \\mathcal {L} _ {a l i g n} - \\left(1 - \\mathcal {L} _ {s e p n}\\right) ^ {2}, \\tag {11}", + "image_path": "3b689588ac6c7688d57fac85993c6039e61e2e7cca7a33993ee3d450c8659188.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 475, + 333, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 333, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 333, + 484 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 491, + 550, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 550, + 531 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 550, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e p n}} = \\sum_ {p = 1} ^ {P} \\mathcal {S} \\left\\{\\Phi_ {D, (p)} ^ {\\text {s e p n}}, \\Phi_ {R, (p)} ^ {\\text {s e p n}} \\right\\}, \\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {p = 1} ^ {P} \\mathcal {S} \\left\\{\\Phi_ {D, (p)} ^ {\\text {a l i g n}}, \\Phi_ {R, (p)} ^ {\\text {a l i g n}} \\right\\}. \\tag {12}", + "image_path": "8ad01b8316b90b20e89a352f11962e00336d4438e73cfef8901e8378cde63f9c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": "Unlike " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": " distance, the spherical distance captures relative differences without being influenced by scale. For 1D feature maps with three consecutive pixels: " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "f_{1} = [0.4,0.5,0.6]" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "f_{2} = [4,5,6]" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "f_{3} = [0.6,0.5,0.4]" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": " exhibit similar pixel-wise increase structures and potentially similar extracted features, while " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "f_{3}" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": " differs. Thus, " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\{f_1,f_2\\}" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": " should be closer in distance and aligned, while " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\{f_1,f_3\\}" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": " should be separated. However, " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\ell_2\\{f_1,f_2\\} = 7.90" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\ell_2\\{f_1,f_3\\} = 0.28" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{S}\\{f_1,f_2\\} = 7.7\\times 10^{-8}" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{S}\\{f_1,f_3\\} = 0.02" + }, + { + "bbox": [ + 304, + 533, + 547, + 651 + ], + "type": "text", + "content": ". Therefore, the spherical distance is a better choice." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "text", + "content": "When optimizing " + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}" + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "text", + "content": ", since the value range of spherical distance " + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "inline_equation", + "content": "S\\{\\cdot, \\cdot\\}" + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "inline_equation", + "content": "[0,2]" + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{align} \\to 0" + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sepn} \\to 1" + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "text", + "content": ". According to Definition 3, the derivation gives: " + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}\\{\\Phi_{D,(p)}^{align}, \\Phi_{R,(p)}^{align}\\} \\to 1" + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}\\{\\Phi_{D,(p)}^{sepn}, \\Phi_{R,(p)}^{sepn}\\} \\to 0" + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 652, + 547, + 713 + ], + "type": "text", + "content": " is the cosine similarity (the last term of Eq. (5)). When" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "12552" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 69, + 541, + 252 + ], + "blocks": [ + { + "bbox": [ + 54, + 69, + 541, + 252 + ], + "lines": [ + { + "bbox": [ + 54, + 69, + 541, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 69, + 541, + 252 + ], + "type": "image", + "image_path": "53fdec5ccd7c452001ed4feae975686092e63999dad86a08f65b1fc7286a3f3a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 255, + 465, + 267 + ], + "lines": [ + { + "bbox": [ + 127, + 255, + 465, + 267 + ], + "spans": [ + { + "bbox": [ + 127, + 255, + 465, + 267 + ], + "type": "text", + "content": "Figure 5: Visual error maps for test image in the Middlebury dataset for " + }, + { + "bbox": [ + 127, + 255, + 465, + 267 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 127, + 255, + 465, + 267 + ], + "type": "text", + "content": " super-resolution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 277, + 541, + 449 + ], + "blocks": [ + { + "bbox": [ + 53, + 277, + 541, + 449 + ], + "lines": [ + { + "bbox": [ + 53, + 277, + 541, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 541, + 449 + ], + "type": "image", + "image_path": "a254881d0939d8a4ba38d74da3fbeeea4a6eca5322d6ff349a5f72df2ecdf33f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 453, + 461, + 464 + ], + "lines": [ + { + "bbox": [ + 130, + 453, + 461, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 453, + 461, + 464 + ], + "type": "text", + "content": "Figure 6: Visual error maps for test image in the RGBDD dataset for " + }, + { + "bbox": [ + 130, + 453, + 461, + 464 + ], + "type": "inline_equation", + "content": "16\\times" + }, + { + "bbox": [ + 130, + 453, + 461, + 464 + ], + "type": "text", + "content": " super-resolution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 486, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 1" + }, + { + "bbox": [ + 46, + 486, + 288, + 544 + ], + "type": "text", + "content": ", vectors have an angle of 0, indicating higher similarity. If " + }, + { + "bbox": [ + 46, + 486, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 0" + }, + { + "bbox": [ + 46, + 486, + 288, + 544 + ], + "type": "text", + "content": ", the vectors are orthogonal, signifying no correlation. By forcing " + }, + { + "bbox": [ + 46, + 486, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 1" + }, + { + "bbox": [ + 46, + 486, + 288, + 544 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 486, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = 0" + }, + { + "bbox": [ + 46, + 486, + 288, + 544 + ], + "type": "text", + "content": " for aligned/private features mapped to spherical space, we ensure the enhancement/reduction of similarity, which achieves our goal of feature decomposition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 546, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 571 + ], + "type": "text", + "content": "By combining Eqs. (9) to (11), we obtain the total loss used for training, which can be expressed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 581, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p i x e l}} ^ {D} + \\alpha_ {1} \\mathcal {L} _ {\\text {p i x e l}} ^ {R} + \\alpha_ {2} \\mathcal {L} _ {\\text {d e c}} + \\alpha_ {3} \\mathcal {L} _ {\\text {S C R}}. \\tag {13}", + "image_path": "837c2a8aa72917748be5cd0b6c977982e96203e6239a26eb4fac63b0dd2da79c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 606, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 631 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 46, + 606, + 287, + 631 + ], + "type": "inline_equation", + "content": "\\alpha_{3} = 0" + }, + { + "bbox": [ + 46, + 606, + 287, + 631 + ], + "type": "text", + "content": " if the current epoch does not require the spherical contrast refinement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 644, + 123, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 123, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 123, + 658 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 712 + ], + "type": "text", + "content": "This section will conduct a comprehensive set of experiments that aim to showcase the effectiveness of our model in addressing the GDSR task. Moreover, we will also provide evidence that substantiates the soundness of our SSDNet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 485, + 356, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 356, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 356, + 498 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 506, + 547, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 697 + ], + "type": "text", + "content": "Datasets. Our experiments follow the protocol established in [27, 31, 90]. To evaluate our model, we employ four popular benchmarks: Middlebury [29, 56], Lu [45], NYU v2 [57], and RGBDD [27]. Our training and validation sets consist of the first 1000 images of NYU v2 dataset [57], divided into a 9:1 ratio. The last 449 pairs in NYU v2 are utilized as the test dataset. Middlebury [29, 56] (30 pairs), Lu [45] (6 pairs), and RGBDD [27] (405 pairs) are also used as test sets to evaluate the depth map SR ability across different scenes and objects. To furthermore demonstrate the generalization ability of our model to unknown scenarios, we incorporate thereal-world branch of the RGBDD dataset, which comprises 2215/405 pairs of RGB-D images for training/test sets, respectively. Lastly, we use the root-mean-square error (RMSE) metric to evaluate the super-resolution effect of our model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "type": "text", + "content": "Implementation details. In our experiments, we apply bicu" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12553" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 509, + 205 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 509, + 205 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 509, + 205 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 509, + 205 + ], + "type": "table", + "html": "
MethodsMiddlebury [29, 56]NYU V2 [57]Lu [45]RGBDD [27]
×4×8×16×4×8×16×4×8×16×4×8×16
PAC [59]1.322.624.581.893.336.781.202.335.191.251.983.49
CUNet [11]1.102.174.331.923.706.780.912.234.991.181.953.45
DKN [31]1.232.124.241.623.266.510.962.165.111.301.963.42
FDKN [31]1.082.174.501.863.586.960.822.105.051.181.913.41
FDSR [27]1.132.084.391.613.185.861.292.195.001.161.823.06
GraphSR [8]1.112.124.431.793.176.020.922.055.151.301.833.12
DCTNet [90]1.102.054.191.593.165.840.881.854.391.081.743.05
Ours1.021.914.021.603.145.860.801.824.771.041.722.92
", + "image_path": "583ed6f19d4c5632bc9fbf45adf3a48dbac7acd305b808948ecd71b4daacdfcd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 251, + 287, + 319 + ], + "blocks": [ + { + "bbox": [ + 46, + 209, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 234 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons among the SOTA methods and our SSDNet in test datasets. The best and second best values are highlighted by **bold** and **underline**, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 251, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 50, + 251, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 50, + 251, + 287, + 319 + ], + "type": "table", + "html": "
Dataset: RGBDD in real-world
MethodsRMSEMethodsRMSEMethodsRMSE
SVLRM [47]8.05DJF [34]7.90DJFR [35]8.01
DKN [31]7.38FDKN [31]7.50FDSR [27]7.50
DCTNet [90]7.37SSDNet7.32
FDSR* [27]5.49DCTNet* [90]5.43SSDNet*5.38
", + "image_path": "4553b3d3a14afc954421868ee5f9a3874071a0185d61caa1a83cd5c55569b839.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 379, + 287, + 407 + ], + "blocks": [ + { + "bbox": [ + 46, + 323, + 288, + 368 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 368 + ], + "type": "text", + "content": "Table 2: Quantitative comparisons among the SOTA methods and SSDNet on real-world branch of RGBDD. The best and second best RMSEs are highlighted by **bold** and **underline**. Model* means that the model has been fine-tuned on the real-world training data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 379, + 287, + 407 + ], + "lines": [ + { + "bbox": [ + 48, + 379, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 379, + 287, + 407 + ], + "type": "table", + "html": "
MethodsPACCUNDKNFDKNFDSRGSRDCTOurs
Time (s)0.200.230.210.040.010.920.080.10
", + "image_path": "47ce1e7f13528b133ecd957950e3c6a845a532edc1f99adf95e4e93af25eb56a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "type": "text", + "content": "Table 3: Results of time-consuming comparison for generating a " + }, + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 46, + 411, + 287, + 434 + ], + "type": "text", + "content": " HR depth map." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "text", + "content": "bic down-sampling to the HR depth maps to synthesize the corresponding LR depth maps. During the pre-processing stage, the training samples undergo resizing to " + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "text", + "content": ". We train the network with a mini-batch size of 56 for 100 epochs, employing the Adam [32] optimizer with an initial learning rate of " + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-3}" + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "text", + "content": ". For training and testing, we employ a PC featuring eight NVIDIA GeForce RTX 3090 GPUs. " + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "inline_equation", + "content": "\\alpha_{1}, \\alpha_{2}" + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "inline_equation", + "content": "\\alpha_{3}" + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "text", + "content": " in Eq. (13) set to " + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 46, + 455, + 288, + 587 + ], + "type": "text", + "content": " in order to balance the magnitudes of each term in the loss. The SCR fine-tuning is performed once every 10 epochs of training. Further details are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 596, + 223, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 223, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 223, + 609 + ], + "type": "text", + "content": "4.2. Empirical validation experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Impact of network hyper-parameters. In our proposed SSDNet, the number of Restormer Blocks " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": ", and number of dimensions in Restormer " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": ", and the number of negative samples in SCR " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": " are crucial factors in enhancing the super-resolution performance. We conduct experiments on the NYU validation set to study the impact of different combinations of " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "P, C, N" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": ". Initially, we fix any two parameters at " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "C = 64" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "P = 4" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": ", respectively, and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": "evaluate the prediction quality of the unfixed parameter at " + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "inline_equation", + "content": "P = 2,3,4,5,6" + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "inline_equation", + "content": "C = 32,64,96,128" + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "inline_equation", + "content": "N = 2,4,8,16" + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": ". The results are summarized in Fig. 3. We observe that the performance is limited when " + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "inline_equation", + "content": "P < 4" + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": ", and redundant parameter increases do not yield corresponding improvements when " + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "inline_equation", + "content": "P > 4" + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": ". Similarly, increasing " + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": " to more than 64 does not produce significant benefits but instead increased the training cost and computational expense. For the SCR module, the increase in training burden does not match the improvement in effectiveness when " + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": " exceeds 8. Therefore, to ensure a balance between performance and computational cost, we choose " + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "inline_equation", + "content": "P = 4" + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "inline_equation", + "content": "C = 64" + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 304, + 252, + 547, + 397 + ], + "type": "text", + "content": " for subsequent experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 400, + 548, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 548, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 548, + 473 + ], + "type": "text", + "content": "Visualization of feature decomposition. Aligned and separated features are visualized in Fig. 4. Aligned features extract shared properties (edges/contours) from the depth/RGB image pair. Separated features capture modality-specific details (textures in RGB images, smooth depth distance in depth maps). The visualization aligns with our motivation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 483, + 484, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 484, + 496 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 484, + 496 + ], + "type": "text", + "content": "4.3. Comparison with SOTA methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 502, + 547, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 563 + ], + "type": "text", + "content": "This section aims to evaluate the performance of our SS-DNet on several popular benchmarks in Sec. 4.1, and compare our results with the state-of-the-art methods, including PAC [59], CUNet [11], DKN [31], FDKN [31], FDSR [27], GraphSR [8], and DCTNet [90]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 566, + 547, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 547, + 650 + ], + "type": "text", + "content": "Qualitative Comparison. We present comparisons of the error maps with ground truth depth maps among SOTA methods in Figs. 5 and 6. Intuitively, the depth predictions of SSDNet exhibit lower prediction errors and are closer to the ground truth images, especially in terms of perception and restoration for the edges and contours. Further visual comparisons are shown in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "content": "Quantitative results for traditional testsets. The quantitative results with scaling factors of 4, 8, and 16 on four test sets are shown in Tab. 1. Compared with existing methods that achieve good results on certain datasets or superresolution factors, our SSDNet produces satisfactory predic" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12554" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 497, + 167 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 497, + 167 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 497, + 167 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 497, + 167 + ], + "type": "table", + "html": "
Encoder&DecoderLdecForm of LdecLRpixelSCR Module×4×8×16
Exp. IShared✓S✓✓1.232.103.49
Exp. IIPrivate×S✓✓1.171.943.27
Exp. IIIPrivate✓l2✓✓1.151.833.06
Exp. IVPrivate✓S×✓1.161.893.29
Exp. VPrivate✓S✓×1.181.873.36
OursPrivate✓S✓✓1.041.722.92
", + "image_path": "004432a8726c72c87b4fdbae85c3b952b8117fb348684b1f041109eacb91bf69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 171, + 501, + 182 + ], + "lines": [ + { + "bbox": [ + 90, + 171, + 501, + 182 + ], + "spans": [ + { + "bbox": [ + 90, + 171, + 501, + 182 + ], + "type": "text", + "content": "Table 4: Results of ablation experiments on the RGBDD test set. Bold indicates the best score in terms of RMSE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 203, + 288, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 203, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 288, + 239 + ], + "type": "text", + "content": "tions on multiple datasets and various super-resolution scales. This indicates that our model has a better super-resolution performance than previous methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "content": "Quantitative results for real-world branch. In addition, following [27, 90], we use the pre-trained " + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 242, + 288, + 361 + ], + "type": "text", + "content": " model to directly test on the real-world branch of the RGBDD dataset to explore the generalization ability in real-world scenarios. We conduct all testing without any additional fine-tuning, and the quantitative results are presented in Tab. 2. Furthermore, we also perform targeted training and testing on this dataset. Regardless of whether fine-tuning or not, SSDNet outperforms previous methods in RMSE, highlighting its powerful generalization ability in handling unknown scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "type": "text", + "content": "Parameter & running time comparison. We demonstrate the relationship between the number of model learnable parameters vs. RMSE on the RGBDD dataset for " + }, + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "type": "inline_equation", + "content": "\\times 16" + }, + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "type": "text", + "content": " and Middlebury for " + }, + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 365, + 288, + 473 + ], + "type": "text", + "content": " in Fig. 1. Our model exhibits a clear advantage over existing methods with relatively fewer parameters. The time-consuming comparison is presented in Tab. 3. They both demonstrate the efficiency of our method for the GDSR task, showing the potential for developing lightweight networks and practical applications in the future." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 479, + 148, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 148, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 148, + 491 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 534 + ], + "type": "text", + "content": "In this section, we validate the design rationality of our SSDNet through ablation experiments based on the RGBDD testset, and present the results in Tab. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 536, + 288, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 288, + 597 + ], + "type": "text", + "content": "SSDNet architecture. In Exp. I, we share the encoders and decoders for RGB and depth images, i.e., a unified encoder and decoder are used instead of " + }, + { + "bbox": [ + 46, + 536, + 288, + 597 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}_D,\\mathcal{E}_R\\}" + }, + { + "bbox": [ + 46, + 536, + 288, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 536, + 288, + 597 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_D,\\mathcal{D}_R\\}" + }, + { + "bbox": [ + 46, + 536, + 288, + 597 + ], + "type": "text", + "content": ". The number of Restormer blocks is increased to ensure that the learnable parameter number is comparable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "type": "text", + "content": "Spherical feature decomposition. In Exp. II, we eliminate the entire feature decomposition module, i.e., " + }, + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}" + }, + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "type": "text", + "content": " in Eq. (11) will not be employed. In Exp. III, we change the distance measure for " + }, + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sep}" + }, + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{align}" + }, + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "type": "text", + "content": " in Eq. (12) from spherical space distance to Euclidean distance, i.e., " + }, + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 600, + 288, + 660 + ], + "type": "text", + "content": "-loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 662, + 288, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 662, + 288, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 288, + 699 + ], + "type": "text", + "content": "Reconstruction RGB. In Exp. IV, We eliminate " + }, + { + "bbox": [ + 46, + 662, + 288, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pixel}^{R}" + }, + { + "bbox": [ + 46, + 662, + 288, + 699 + ], + "type": "text", + "content": " in Eq. (10) and explore the change for feature extraction capability without the restriction of reconstructing RGB." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 701, + 288, + 714 + ], + "type": "text", + "content": "Spherical contrast refinement. In Exp. V, we remove the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 202, + 547, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 547, + 226 + ], + "type": "text", + "content": "SCR module and use " + }, + { + "bbox": [ + 304, + 202, + 547, + 226 + ], + "type": "inline_equation", + "content": "\\hat{D}_{HR}" + }, + { + "bbox": [ + 304, + 202, + 547, + 226 + ], + "type": "text", + "content": " in Eq. (8) as the final output of our model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 232, + 547, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 547, + 255 + ], + "type": "text", + "content": "Analysis. Tab. 4 shows that changing settings leads to performance degradation, validating our model design." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 255, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 291 + ], + "type": "text", + "content": "Exp. I: Unified encoder/decoder ignores modality-specific features, hindering efficient cross-modal feature extraction and causing the largest degradation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 291, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 314 + ], + "type": "text", + "content": "Exp. II: The absence of " + }, + { + "bbox": [ + 304, + 291, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}" + }, + { + "bbox": [ + 304, + 291, + 545, + 314 + ], + "type": "text", + "content": " prevents decomposing the shared/private features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 315, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 338 + ], + "type": "text", + "content": "Exp. III: " + }, + { + "bbox": [ + 304, + 315, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 315, + 545, + 338 + ], + "type": "text", + "content": " distance partially decomposes features but is less effective than " + }, + { + "bbox": [ + 304, + 315, + 545, + 338 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 315, + 545, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 338, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 363 + ], + "type": "text", + "content": "Exp. IV: " + }, + { + "bbox": [ + 304, + 338, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pixel}^{R}" + }, + { + "bbox": [ + 304, + 338, + 545, + 363 + ], + "type": "text", + "content": " preserves semantic information in RGB and avoids excessive adaptation of feature separation/alignment." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 363, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 387 + ], + "type": "text", + "content": "Exp. V: Removal of SCR leads to mentioned detailed issues and performance degradation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 403, + 379, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 379, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 379, + 415 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 425, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 569 + ], + "type": "text", + "content": "We propose a Spherical Space feature Decomposition network (SSDNet) for guided depth map super-resolution, where a Restormer-based encoder is used to extract the global features of the inputs, and the intermediate features will be mapped to the spherical space to complete the separation/alignment of shared/private features, respectively. Finally, a Restormer-based decoder is to reconstruct the HR depth map. The spherical contrast refinement module is then employed to address the possible detail issues, e.g., blurry edges, noisy surfaces and over-transferred RGB texture. The satisfactory output results and the lightweight size demonstrate the superiority of our approach." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 584, + 404, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 584, + 404, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 584, + 404, + 597 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "This work has been supported by the National Key Research and Development Program of China under grant 2018AAA0102201, the National Natural Science Foundation of China under Grant 61976174 and 12201497, Shaanxi Fundamental Science Research Project for Mathematics and Physics under Grant 22JSQ033, the Fundamental Research Funds for the Central Universities under Grant D5000220060, and partly supported by the Alexander von Humboldt Foundation." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12555" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 711 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 121 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 121 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 121 + ], + "type": "text", + "content": "[1] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. CoRR, abs/2106.06847, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 166 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 166 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 167, + 288, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 167, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 53, + 167, + 288, + 209 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 209, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 288, + 240 + ], + "type": "text", + "content": "[4] Efrain Castillo-Muniz and Eduardo Bayro-Corrochano. Geometric spherical networks for visual data processing. In IJCNN, 2012. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 241, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 284 + ], + "type": "text", + "content": "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310. CVF / IEEE, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 316 + ], + "type": "text", + "content": "[6] Taco Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Convolutional networks for spherical signals. arXiv preprint arXiv:1709.04893, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 316, + 288, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 316, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 288, + 336 + ], + "type": "text", + "content": "[7] Taco S. Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical CNNs. In ICLR, 2018. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 337, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 288, + 380 + ], + "type": "text", + "content": "[8] Riccardo de Lutio, Alexander Becker, Stefano D'Aronco, Stefania Russo, Jan D. Wegner, and Konrad Schindler. Learning graph regularisation for guided super-resolution. In CVPR, pages 1969-1978. IEEE, 2022. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 288, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 411 + ], + "type": "text", + "content": "[9] Riccardo de Lutio, Stefano D'Aronco, Jan Dirk Wegner, and Konrad Schindler. Guided super-resolution as pixel-to-pixel transformation. In ICCV, pages 8828-8836. IEEE, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 412, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 412, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 288, + 443 + ], + "type": "text", + "content": "[10] Xin Deng and Pier Luigi Dragotti. Deep coupled ISTA network for multi-modal image super-resolution. IEEE Trans. Image Process., 29:1683-1698, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 443, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 288, + 485 + ], + "type": "text", + "content": "[11] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Trans. Pattern Anal. Mach. Intell., 43(10):3333-3348, 2021. 2, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 486, + 288, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 288, + 517 + ], + "type": "text", + "content": "[12] James Diebel and Sebastian Thrun. An application of markov random fields to range sensing. In NeurIPS, pages 291-298, 2005. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 517, + 288, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 517, + 288, + 549 + ], + "spans": [ + { + "bbox": [ + 48, + 517, + 288, + 549 + ], + "type": "text", + "content": "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image superresolution. In ECCV, pages 184-199. Springer, 2014. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 550, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 550, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 550, + 288, + 592 + ], + "type": "text", + "content": "[14] Jiangxin Dong, Jinshan Pan, Jimmy S. Ren, Liang Lin, Jinhui Tang, and Ming-Hsuan Yang. Learning spatially variant linear representation models for joint filtering. IEEE Trans. Pattern Anal. Mach. Intell., 44(11):8355-8370, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 593, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 288, + 624 + ], + "type": "text", + "content": "[15] Xiaoyu Dong, Naoto Yokoya, Longguang Wang, and Tatsumi Uezato. Learning mutual modulation for self-supervised cross-modal super-resolution. In ECCV, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 689, + 288, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 288, + 711 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 288, + 711 + ], + "type": "text", + "content": "[17] David Ferstl, Christian Reinbacher, René Ranftl, Matthias Rüther, and Horst Bischof. Image guided depth upsampling" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 74, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "type": "text", + "content": "using anisotropic total generalized variation. In ICCV, pages 993-1000, 2013. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 95, + 547, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 137 + ], + "type": "text", + "content": "[18] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. RSTT: real-time spatial temporal transformer for space-time video super-resolution. CoRR, abs/2203.14186, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 137, + 545, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 137, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 545, + 180 + ], + "type": "text", + "content": "[19] Shuhang Gu, Shi Guo, Wangmeng Zuo, Yunjin Chen, Radu Timofte, Luc Van Gool, and Lei Zhang. Learned dynamic guidance for depth image reconstruction. IEEE Trans. Pattern Anal. Mach. Intell., 42(10):2437-2452, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 181, + 547, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 181, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 307, + 181, + 547, + 222 + ], + "type": "text", + "content": "[20] Shuhang Gu, Wangmeng Zuo, Shi Guo, Yunjin Chen, Chongyu Chen, and Lei Zhang. Learning dynamic guidance for depth image enhancement. In CVPR, pages 712-721, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 223, + 545, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 223, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 223, + 545, + 244 + ], + "type": "text", + "content": "[21] Xiang Gu, Jian Sun, and Zongben Xu. Spherical space domain adaptation with robust pseudo-label loss. In CVPR, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 244, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 545, + 285 + ], + "type": "text", + "content": "[22] Xiang Gu, Jian Sun, and Zongben Xu. Unsupervised and semi-supervised robust spherical space domain adaptation. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-1, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "text", + "content": "[23] Chunle Guo, Chongyi Li, Jichang Guo, Runmin Cong, Huazhu Fu, and Ping Han. Hierarchical features driven residual learning for depth map super-resolution. IEEE Trans. Image Process., 28(5):2545-2557, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 330, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 330, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 330, + 545, + 373 + ], + "type": "text", + "content": "[24] Saurabh Gupta, Ross B. Girshick, Pablo Andres Arbeláez, and Jitendra Malik. Learning rich features from RGB-D images for object detection and segmentation. In ECCV, pages 345-360. Springer, 2014. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 373, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 415 + ], + "type": "text", + "content": "[25] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735. CVF / IEEE, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 415, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 446 + ], + "type": "text", + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 447, + 545, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 447, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 447, + 545, + 500 + ], + "type": "text", + "content": "[27] Lingzhi He, Hongguang Zhu, Feng Li, Huihui Bai, Runmin Cong, Chunjie Zhang, Chunyu Lin, Meiqin Liu, and Yao Zhao. Towards fast and accurate real-world depth superresolution: Benchmark dataset and baseline. In CVPR, pages 9229-9238, 2021. 7, 8, 9" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "text", + "content": "[28] Olivier J. Henaff. Data-efficient image recognition with contrastive predictive coding. In ICML, volume 119, pages 4182-4192. PMLR, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 533, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 553 + ], + "type": "text", + "content": "[29] Heiko Hirschmüller and Daniel Scharstein. Evaluation of cost functions for stereo matching. In CVPR, 2007. 7, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 586 + ], + "type": "text", + "content": "[30] Tak-Wai Hui, Chen Change Loy, and Xiaou Tang. Depth map super-resolution by deep multi-scale guidance. In ECCV, pages 353-369. Springer, 2016. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 586, + 545, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 586, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 307, + 586, + 545, + 617 + ], + "type": "text", + "content": "[31] Beomjun Kim, Jean Ponce, and Bumsub Ham. Deformable kernel networks for joint image filtering. Int. J. Comput. Vis., 129(2):579-600, 2021. 2, 7, 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 618, + 545, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 618, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 307, + 618, + 545, + 638 + ], + "type": "text", + "content": "[32] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 639, + 545, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 639, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 307, + 639, + 545, + 670 + ], + "type": "text", + "content": "[33] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. CoRR, abs/2112.10175, 2021. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 670, + 545, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 702 + ], + "type": "text", + "content": "[34] Yijun Li, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep joint image filtering. In ECCV, pages 154-169. Springer, 2016. 8" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 703, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 703, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 703, + 545, + 713 + ], + "type": "text", + "content": "[35] Yijun Li, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12556" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 703 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "text", + "content": "Yang. Joint image filtering with deep convolutional networks. IEEE Trans. Pattern Anal. Mach. Intell., 41(8):1909-1923, 2019. 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 137 + ], + "type": "text", + "content": "[36] Yu Li, Dongbo Min, Minh N. Do, and Jiangbo Lu. Fast guided global interpolation for depth and motion. In ECCV, pages 717-733. Springer, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 288, + 179 + ], + "type": "text", + "content": "[37] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCVW, pages 1833-1844. IEEE, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 180, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 180, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 180, + 288, + 213 + ], + "type": "text", + "content": "[38] Miao Liao, Feixiang Lu, Dingfu Zhou, Sibo Zhang, Wei Li, and Ruigang Yang. DVI: depth guided video inpainting for autonomous driving. In ECCV, pages 1-17. Springer, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 213, + 288, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 213, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 288, + 243 + ], + "type": "text", + "content": "[39] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 244, + 288, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 288, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 288, + 276 + ], + "type": "text", + "content": "[40] Ming-Yu Liu, Oncel Tuzel, and Yuichi Taguchi. Joint geodesic upsampling of depth images. In CVPR, pages 169-176, 2013. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 276, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 319 + ], + "type": "text", + "content": "[41] Xianming Liu, Deming Zhai, Rong Chen, Xiangyang Ji, Debin Zhao, and Wen Gao. Depth super-resolution via joint color-guided internal and external regularizations. IEEE Trans. Image Process., 28(4):1636-1645, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 319, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 288, + 361 + ], + "type": "text", + "content": "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002. IEEE, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 361, + 288, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 382 + ], + "type": "text", + "content": "[43] Jiajun Lu and David A. Forsyth. Sparse depth super resolution. In CVPR, pages 2245-2253, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 383, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 414 + ], + "type": "text", + "content": "[44] Jiangbo Lu, Keyang Shi, Dongbo Min, Liang Lin, and Minh N. Do. Cross-based local multipoint filtering. In CVPR, pages 430-437, 2012. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 415, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 445 + ], + "type": "text", + "content": "[45] Si Lu, Xiaofeng Ren, and Feng Liu. Depth enhancement via low-rank matrix completion. In CVPR, pages 3390-3397, 2014. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 446, + 288, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 446, + 288, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 446, + 288, + 478 + ], + "type": "text", + "content": "[46] Dongbo Min, Jiangbo Lu, and Minh N. Do. Depth video enhancement based on weighted mode filtering. IEEE Trans. Image Process., 21(3):1176-1190, 2012. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 479, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 288, + 520 + ], + "type": "text", + "content": "[47] Jinshan Pan, Jiangxin Dong, Jimmy S. J. Ren, Liang Lin, Jinhui Tang, and Ming-Hsuan Yang. Spatially variant linear representation models for joint filtering. In CVPR, pages 1702-1711, 2019. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 521, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 288, + 563 + ], + "type": "text", + "content": "[48] Jaesik Park, Hyeongwoo Kim, Yu-Wing Tai, Michael S. Brown, and In-So Kweon. High quality depth map upsampling for 3d-tof cameras. In ICCV, pages 1623-1630, 2011. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 563, + 288, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 563, + 288, + 607 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 288, + 607 + ], + "type": "text", + "content": "[49] Taesung Park, Alexei A. Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In ECCV (9), volume 12354 of Lecture Notes in Computer Science, pages 319-345. Springer, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 607, + 288, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 607, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 48, + 607, + 288, + 639 + ], + "type": "text", + "content": "[50] Wanli Peng, Hao Pan, He Liu, and Yi Sun. IDA-3D: instance-depth-aware 3d object detection from stereo vision for autonomous driving. In CVPR, pages 13012–13021, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 639, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 639, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 639, + 288, + 680 + ], + "type": "text", + "content": "[51] Haotong Qin, Zhongang Cai, Mingyuan Zhang, Yifu Ding, Haiyu Zhao, Shuai Yi, Xianglong Liu, and Hao Su. Bipoint-net: Binary neural network for point clouds. In ICLR, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 288, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 703 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 703 + ], + "type": "text", + "content": "[52] Haotong Qin, Ruihao Gong, Xianglong Liu, Xiao Bai, Jingkuan Song, and Nicu Sebe. Binary neural networks:" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 546, + 704 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 326, + 73, + 470, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 470, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 470, + 83 + ], + "type": "text", + "content": "A survey. Pattern Recognition, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "type": "text", + "content": "[53] Haotong Qin, Ruihao Gong, Xianglong Liu, Mingzhu Shen, Ziran Wei, Fengwei Yu, and Jingkuan Song. Forward and backward information retention for accurate binary neural networks. In ICCV, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 126, + 546, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 126, + 546, + 158 + ], + "spans": [ + { + "bbox": [ + 308, + 126, + 546, + 158 + ], + "type": "text", + "content": "[54] Gernot Riegler, David Ferstl, Matthias Rüther, and Horst Bischof. A deep primal-dual network for guided depth superresolution. In BMVC. BMVA Press, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 159, + 546, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 159, + 546, + 190 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 546, + 190 + ], + "type": "text", + "content": "[55] Gernot Riegler, Matthias Rüther, and Horst Bischof. Atgv-net: Accurate depth super-resolution. In ECCV, pages 268-284. Springer, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 190, + 546, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 190, + 546, + 210 + ], + "spans": [ + { + "bbox": [ + 308, + 190, + 546, + 210 + ], + "type": "text", + "content": "[56] Daniel Scharstein and Chris Pal. Learning conditional random fields for stereo. In CVPR, 2007. 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 211, + 546, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 211, + 546, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 211, + 546, + 253 + ], + "type": "text", + "content": "[57] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. In ECCV, pages 746-760. Springer, 2012. 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 254, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 546, + 297 + ], + "type": "text", + "content": "[58] Xibin Song, Yuchao Dai, Dingfu Zhou, Liu Liu, Wei Li, Hongdong Li, and Ruigang Yang. Channel attention based iterative residual learning for depth map super-resolution. In CVPR, pages 5630-5639, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 297, + 546, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 546, + 329 + ], + "type": "text", + "content": "[59] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik G. Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In CVPR, pages 11166-11175, 2019. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 329, + 546, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 329, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 329, + 546, + 371 + ], + "type": "text", + "content": "[60] Baoli Sun, Xinchen Ye, Baopu Li, Haojie Li, Zhihui Wang, and Rui Xu. Learning scene structure guidance via cross-task knowledge transfer for single depth super-resolution. In CVPR, pages 7792-7801, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 372, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 546, + 403 + ], + "type": "text", + "content": "[61] Jiaxiang Tang, Xiaokang Chen, and Gang Zeng. Joint implicit image function for guided depth super-resolution. In ACM Multimedia, pages 4390-4399. ACM, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 403, + 546, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 403, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 308, + 403, + 546, + 456 + ], + "type": "text", + "content": "[62] Qi Tang, Runmin Cong, Ronghui Sheng, Lingzhi He, Dan Zhang, Yao Zhao, and Sam Kwong. Bridgenet: A joint learning network of depth map super-resolution and monocular depth estimation. In ACM Multimedia, pages 2148-2157. ACM, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 457, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 457, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 457, + 546, + 510 + ], + "type": "text", + "content": "[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, volume 139, pages 10347-10357. PMLR, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 511, + 546, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 546, + 553 + ], + "type": "text", + "content": "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 554, + 546, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 546, + 596 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 546, + 596 + ], + "type": "text", + "content": "[65] Jiakai Wang, Aishan Liu, Xiao Bai, and Xianglong Liu. Universal adversarial patch attack for automatic checkout using perceptual and attentional bias. IEEE Trans. Image Process., 31:598-611, 2021. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 597, + 546, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 597, + 546, + 650 + ], + "spans": [ + { + "bbox": [ + 308, + 597, + 546, + 650 + ], + "type": "text", + "content": "[66] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 548-558. IEEE, 2021. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 651, + 546, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 651, + 546, + 682 + ], + "spans": [ + { + "bbox": [ + 308, + 651, + 546, + 682 + ], + "type": "text", + "content": "[67] Zhendong Wang, Xiaodong Cun, Jianmin Bao, and Jianzhuang Liu. Uformer: A general u-shaped transformer for image restoration. CoRR, abs/2106.03106, 2021. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 683, + 546, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 683, + 546, + 704 + ], + "spans": [ + { + "bbox": [ + 308, + 683, + 546, + 704 + ], + "type": "text", + "content": "[68] Yang Wen, Bin Sheng, Ping Li, Weiyao Lin, and David Dagan Feng. Deep color guided coarse-to-fine convolutional network" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12557" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 705 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": " cascade for depth image super-resolution. IEEE Trans. Image Process., 28(2):994-1006, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 288, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 288, + 137 + ], + "type": "text", + "content": "[69] Richard C. Wilson, Edwin R. Hancock, Elzbieta Pekalska, and Robert P. W. Duin. Spherical and hyperbolic embeddings of data. IEEE Trans. Pattern Anal. Mach. Intell., 36(11):2255-2269, 2014. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 137, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 288, + 179 + ], + "type": "text", + "content": "[70] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In CVPR, pages 10551-10560. CVF / IEEE, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 179, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 179, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 48, + 179, + 288, + 212 + ], + "type": "text", + "content": "[71] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Fast end-to-end trainable guided filter. In CVPR, pages 1838-1847, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 288, + 255 + ], + "type": "text", + "content": "[72] Enze Xie, Jian Ding, Wenhai Wang, Xiaohang Zhan, Hang Xu, Peize Sun, Zhenguo Li, and Ping Luo. Detco: Unsupervised contrastive learning for object detection. In ICCV, pages 8372-8381. IEEE, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 288, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 288, + 287 + ], + "type": "text", + "content": "[73] Jun Xie, Rogério Schmidt Feris, and Ming-Ting Sun. Edge-guided single depth image super resolution. IEEE Trans. Image Process., 25(1):428-438, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 288, + 328 + ], + "type": "text", + "content": "[74] Jun Xie, Rogério Schmidt Feris, Shiaw-Shian Yu, and Ming-Ting Sun. Joint super resolution and denoising from a single depth image. IEEE Trans. Multim., 17(9):1525-1537, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 329, + 288, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 329, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 48, + 329, + 288, + 383 + ], + "type": "text", + "content": "[75] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2J: anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In ICCV, pages 793-802. IEEE, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 415 + ], + "type": "text", + "content": "[76] Ruikang Xu, Mingde Yao, and Zhiwei Xiong. Zero-shot dual-lens super-resolution. In CVPR, pages 9130-9139, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "text", + "content": "[77] Jingyu Yang, Xinchen Ye, Kun Li, Chunping Hou, and Yao Wang. Color-guided depth recovery from RGB-D data using an adaptive autoregressive model. IEEE Trans. Image Process., 23(8):3443-3458, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 288, + 499 + ], + "type": "text", + "content": "[78] Yuxiang Yang, Qi Cao, Jing Zhang, and Dacheng Tao. CODON: on orchestrating cross-domain attentions for depth super-resolution. Int. J. Comput. Vis., 130(2):267-284, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "type": "text", + "content": "[79] Mingde Yao, Dongliang He, Xin Li, Fu Li, and Zhiwei Xiong. Towards interactive self-supervised denoising. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 288, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 575 + ], + "type": "text", + "content": "[80] Mingde Yao, Dongliang He, Xin Li, Zhihong Pan, and Zhiwei Xiong. Bidirectional translation between uhd-hdr and hd-sdr videos. IEEE Transactions on Multimedia, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 575, + 288, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 575, + 288, + 619 + ], + "spans": [ + { + "bbox": [ + 48, + 575, + 288, + 619 + ], + "type": "text", + "content": "[81] Xinchen Ye, Baoli Sun, Zhihui Wang, Jingyu Yang, Rui Xu, Haojie Li, and Baopu Li. Pmbanet: Progressive multi-branch aggregation network for scene depth super-resolution. IEEE Trans. Image Process., 29:7427-7442, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 619, + 288, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 619, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 48, + 619, + 288, + 661 + ], + "type": "text", + "content": "[82] En Yu, Zhuoling Li, and Shoudong Han. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking. In CVPR, pages 8824-8833. IEEE, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 661, + 288, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 661, + 288, + 705 + ], + "spans": [ + { + "bbox": [ + 48, + 661, + 288, + 705 + ], + "type": "text", + "content": "[83] Shanxin Yuan, Guillermo Garcia-Hernando, Björn Stenger, Gyeongsik Moon, Ju Yong Chang, Kyoung Mu Lee, Pavlo Molchanov, Jan Kautz, Sina Honari, Liuhao Ge, Junsong Yuan, Xinghao Chen, Guijin Wang, Fan Yang, Kai Akiyama," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 72, + 547, + 707 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 72, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 547, + 128 + ], + "type": "text", + "content": "Yang Wu, Qingfu Wan, Meysam Madadi, Sergio Escalera, Shile Li, Dongheui Lee, Jason Oikonomidis, Antonis A. Argyros, and Tae-Kyun Kim. Depth-based 3d hand pose estimation: From current achievements to future goals. In CVPR, pages 2636-2645, 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 128, + 547, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 128, + 547, + 170 + ], + "spans": [ + { + "bbox": [ + 308, + 128, + 547, + 170 + ], + "type": "text", + "content": "[84] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, pages 5718-5729. IEEE, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 170, + 547, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 170, + 547, + 202 + ], + "spans": [ + { + "bbox": [ + 308, + 170, + 547, + 202 + ], + "type": "text", + "content": "[85] Jiahui Zhang, Shijian Lu, Fangneng Zhan, and Yingchen Yu. Blind image super-resolution via contrastive representation learning. CoRR, abs/2107.00708, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 202, + 547, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 202, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 308, + 202, + 547, + 234 + ], + "type": "text", + "content": "[86] Kai Zhang, Luc Van Gool, and Radu Timofte. Deep unfolding network for image super-resolution. In CVPR, pages 3214-3223, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 234, + 547, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 234, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 308, + 234, + 547, + 287 + ], + "type": "text", + "content": "[87] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In CVPR, pages 5906–5916, June 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 287, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 547, + 331 + ], + "type": "text", + "content": "[88] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. arXiv preprint arXiv:2305.11443, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 331, + 547, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 547, + 384 + ], + "type": "text", + "content": "[89] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. DDFM: denoising diffusion model for multi-modality image fusion. CoRR, abs/2303.06840, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 384, + 547, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 547, + 427 + ], + "type": "text", + "content": "[90] Zixiang Zhao, Jiangshe Zhang, Shuang Xu, Zudi Lin, and Hanspeter Pfister. Discrete cosine transform network for guided depth map super-resolution. In CVPR, pages 5687-5697. IEEE, 2022. 2, 3, 7, 8, 9" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 427, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 547, + 491 + ], + "type": "text", + "content": "[91] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H. S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890. CVF / IEEE, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 491, + 547, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 491, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 491, + 547, + 535 + ], + "type": "text", + "content": "[92] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, Zhiwen Chen, and Xiangyang Ji. High-resolution depth maps imaging via attention-based hierarchical multi-modal fusion. IEEE Trans. Image Process., 31:648-663, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 535, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 547, + 567 + ], + "type": "text", + "content": "[93] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, and Xiangyang Ji. Guided depth map super-resolution: A survey. ACM Computing Surveys, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 567, + 547, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 547, + 609 + ], + "type": "text", + "content": "[94] Man Zhou, Keyu Yan, Jinshan Pan, Wenqi Ren, Qi Xie, and Xiangyong Cao. Memory-augmented deep unfolding network for guided image super-resolution. Int. J. Comput. Vis., 131(1):215-242, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 609, + 547, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 609, + 547, + 642 + ], + "spans": [ + { + "bbox": [ + 308, + 609, + 547, + 642 + ], + "type": "text", + "content": "[95] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 642, + 547, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 642, + 547, + 707 + ], + "spans": [ + { + "bbox": [ + 308, + 642, + 547, + 707 + ], + "type": "text", + "content": "[96] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rhemann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew W. Fitzgibbon, Charles T. Loop, Christian Theobalt, and Marc Stamminger. Real-time non-rigid reconstruction using an RGB-D camera. ACM Trans. Graph., 33(4):156:1-156:12, 2014. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12558" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/280659a7-98c7-41f5-b75f-aea246ac862a_content_list.json b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/280659a7-98c7-41f5-b75f-aea246ac862a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3e8332e30b185ba1d14901475202a1d3fcda50fd --- /dev/null +++ b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/280659a7-98c7-41f5-b75f-aea246ac862a_content_list.json @@ -0,0 +1,2227 @@ +[ + { + "type": "text", + "text": "SpinCam: High-Speed Imaging via a Rotating Point-Spread Function", + "text_level": 1, + "bbox": [ + 133, + 130, + 834, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dorian Chan, Mark Sheinin, and Matthew O'Toole", + "bbox": [ + 281, + 181, + 684, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Carnegie Mellon University, Pittsburgh, PA 15213, USA", + "bbox": [ + 261, + 199, + 707, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dychan@andrew.cmu.edu, marksheinin@gmail.com, mpotoole@cmu.edu", + "bbox": [ + 209, + 219, + 759, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-speed cameras are an indispensable tool used for the slow-motion analysis of scenes. However, the fixed bandwidth of any imaging system quickly becomes a bottleneck, resulting in a fundamental trade-off between the camera's spatial and temporal resolutions. In recent years, compressive high-speed imaging systems have been proposed to circumvent these issues by optically encoding the signal and using a reconstruction procedure to recover a video. Our work proposes a novel approach for compressive high-speed imaging based on temporally coding the camera's point-spread function (PSF). By mechanically spinning a diffraction grating in front of a camera, the sensor integrates an image blurred by a PSF that continuously rotates over time. We also propose a deconvolution-based reconstruction algorithm to reconstruct videos from these measurements. Our method achieves superior light efficiency and handles a wider scene class than prior methods. Also, our mechanical design yields flexible temporal resolution that can be easily increased, potentially allowing capture at $192\\mathrm{kHz}$ far higher than prior works. We demonstrate a prototype for various applications, including motion capture and particle image velocimetry (PIV).", + "bbox": [ + 75, + 300, + 473, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 209, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The world is full of high-speed events that are imperceptible to both the human eye and most conventional cameras (e.g., the flicker of electric lights or the motion of fast-moving objects). The traditional approach to capturing such events is using slow-motion photography, i.e., use a camera to shoot images at a high frame rate and display the results at a lower frame rate. Slow-motion photography is often used for artistic effects in film and as a measurement tool in scientific applications. However, despite significant advancements in sensor technology, conventional high-speed cameras tend to be expensive, bulky, and have high throughput requirements for recording visual content at both high frame rates and high spatial resolutions. These hardware limitations set an upper limit to the cameras' capture rates.", + "bbox": [ + 75, + 688, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/637c746f8ed30d0313223c23ab115d9ec8ed05b139f74c8859fef505e38a4bbc.jpg", + "image_caption": [ + "Figure 1. Principle of high-speed imaging with a rotating diffraction grating. A camera captures a long exposure image of a moving scene. During the camera's exposure period, the diffraction grating placed in front of the camera continuously rotates to generate a time-varying point-spread function (PSF) that encodes scene dynamics into a single frame. Solving a deconvolution problem yields a high-speed video from the captured frame." + ], + "image_footnote": [], + "bbox": [ + 519, + 266, + 870, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compressed sensing provides a framework to circumvent such issues. Since videos are highly compressible, it is possible to recover a video from a single snapshot by (i) compressively recording spatiotemporal visual information onto a sensor and (ii) solving a reconstruction problem to recover the video using priors on the visual content. Among the large body of work in compressive high-speed imaging, most techniques temporally modulate the light incident on sensor pixels [48, 12, 20, 22, 9, 34, 32, 35, 14, 16, 19,", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10789", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "13, 49], either through the use of a spatial light modulator (SLM) or novel sensors capable of turning pixels on or off. Alternatively, by using spatial-multiplexing optics (e.g., diffusers, diffraction gratings) to distribute light from any scene point to many camera pixels, it is possible to perform high-speed imaging by using a rolling shutter [2, 50, 26], a reduced region-of-interest (ROI) [39], or a line sensor [40].", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work proposes a fundamentally different approach to compressive high-speed imaging. Similar to past works that make use of spatial-multiplexing optics, we propose to distribute light across a sensor by using a dual-axis diffraction grating, which diffracts light horizontally and vertically (see Fig. 1). Unlike prior work, however, we opt to change the point-spread function (PSF) during the camera's exposure period. Specifically, we use a motor to mechanically rotate the PSF at high speeds, record a single global-shutter frame, and solve a simple deconvolution-based reconstruction procedure to reconstruct scene dynamics. Please refer to Fig. 1 for an overview of the process.", + "bbox": [ + 75, + 198, + 467, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach, like past work using spatial-multiplexing optics [2, 50, 40, 39], is primarily targeted towards spatiotemporally sparse scenes. This imaging regime includes motion capture and particle-image velocimetry (PIV), which see broad application in domains like biomechanics [29], sports science [31], animation and film-making [38], virtual and augmented reality [5], aerodynamics [47], fluid mechanics [36], biomedical engineering [18], oceanography [7], and combustion and engine research [28]. We demonstrate our approach for motion capture and PIV in Sec. 6. Other potential applications for spatiotemporally sparse video include high-speed 3D scanning [24] and capturing the fast flicker of electric bulbs [41].", + "bbox": [ + 75, + 381, + 467, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach combines four key advantages over existing methods targeting spatiotemporally sparse scenes:", + "bbox": [ + 76, + 579, + 467, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(I) Superior light efficiency. In many prior methods, the scene light is dispersed using specialized optics to reach many camera pixels, but only a small subset of camera pixels collect light at each timestamp [2, 40, 39, 50]. Thus, most light is lost by falling outside the 'active' pixels. Conversely, in our work, the signal from different timestamps is encoded by different temporal PSFs, and the entire sensor is left exposed, yielding far superior light efficiency. We show this advantage in simulated experiments in Sec. 3.", + "bbox": [ + 75, + 619, + 467, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(II) Easily increasable temporal resolution. In most prior works, temporal resolution (i.e., FPS) is fundamentally limited by the speed of the respective technology (e.g., the modulation speed of a SLM/sensor [12, 22, 34, 32, 14, 49], or the read-out speed of a camera [2, 40, 39, 50, 9]). In contrast, the temporal resolution of our approach is determined by the mechanical rotation speed and thus can be easily increased by rotating the PSF faster. Thus, we show that our approach can resolve a LED flickering at $96\\mathrm{kHz}$ .", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(III) A wider class of supported scenes. Prior works using spatial-multiplexing optics allocate a fixed amount of pixels for each timestamp (e.g., a single rolling shutter row) [2, 40, 39, 50]. This limits the amount of information recoverable per frame, imposing a spatial sparsity constraint on each high-speed frame individually. Instead, our solution multiplexes the signals from different timestamps together, yielding a sparsity constraint over the entire video. Our method can thus handle temporally sparse but spatially dense scenes, verified in simulated and real experiments.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(IV) Simple hardware. While many prior works require specialized optics (e.g., SLMs [12, 34, 49], piezoelectric stages [20, 16, 19]) or specialized sensors (e.g., coded exposure pixels [22, 14, 32]), our setup can work with any global-shutter camera combined with simple low-cost hardware that is readily available at your local hobby store. We include a DIY guide as part of our supplementary materials.", + "bbox": [ + 496, + 241, + 890, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 498, + 354, + 635, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Spinning Disks & Revolving Mirrors", + "text_level": 1, + "bbox": [ + 500, + 378, + 816, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mechanical spinning disks and revolving mirrors have been used to measure transient phenomena for centuries. In 1834, Wheatstone [51] estimated the velocity of electricity traveling through long copper wires by observing the sparks generated at the ends. Using a revolving mirror, Wheatstone observed a $1\\mu \\mathrm{s}$ time difference, and concluded the velocity of electricity to be 288,000 miles per second. Fizeau and Foucault [25] later measured the speed of light through various apparatuses based on revolving mirrors and spinning cogwheels. For example, in 1848-49, Fizeau measured the transit time of light from a source, to a mirror, and back to the source by using a rotating wheel to modulate the light.", + "bbox": [ + 496, + 402, + 890, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though rare in modern consumer electronics, modulating light mechanically with a spinning element is both faster and lower cost than most modern digital modulation solutions. For example, digital micromirror devices (DMDs) only operate up to $30\\mathrm{kHz}$ , whereas spinning disks can offer rates up to $2.4\\mathrm{MHz}$ [10]. As a result, coded disks have been used for single pixel imaging at rates up to 72 FPS [43, 42, 10], and revolving coded mirrors have been used to reconstruct 1400 frames from a single shot [23]. A spinning camera was used to recover an 11-frame video through scattering media by disambiguating the resulting rotated speckle patterns [44], while a rotating camera coupled with a polarizer generalized the optical-flow equation to include polarization cues [46]. An event camera can be combined with a spinning polarizer for fast, accurate shape-from-polarization [27]. This work explores the use of spinning elements for compressive high-speed video.", + "bbox": [ + 496, + 584, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Pixel-wise Coded Exposure", + "text_level": 1, + "bbox": [ + 500, + 847, + 743, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Coded exposure photography involves temporally modulating the light incident on a sensor, which has long been", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10790", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "used to capture dynamic events [4]. Raskar et al. [33] used a liquid-crystal panel to rapidly open and close the shutter of a camera; the motion blur associated with fast-moving objects could then be reliably removed through deconvolution. This methodology was later extended for recovering full high-speed videos of periodic [48] and generic scenes [13].", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A popular approach to compressive high-speed imaging involves pixel-wise exposure control [12, 9, 34, 32, 35, 22, 14, 20, 16, 19]. In practice, the codes are often binary, and the corresponding reconstruction techniques rely on various priors to reconstruct videos. Example priors include assuming the scene produces periodic motion [48] or satisfies the brightness constancy assumption [34], and might rely on dictionary learning [12, 13] and neural networks [22, 14, 30] to aid the reconstruction procedure.", + "bbox": [ + 75, + 183, + 467, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These techniques lack the (II) temporal resolution and (IV) hardware complexity advantages of our approach since they typically require specialized hardware (e.g., SLMs [12, 34, 49], piezoelectric stages [20, 16, 19]) or specialized sensors [22, 14, 32]) which can be expensive, difficult to acquire, and still be limited in frame rate. Moreover, as shown in the supplementary materials, since these methods do not spread light across multiple pixels, these systems struggle with spatially sparse scenes, which appear in many practical applications of high-speed imaging [40, 39].", + "bbox": [ + 75, + 319, + 467, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Spatial-Multiplexing Optics", + "text_level": 1, + "bbox": [ + 76, + 481, + 326, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An alternative approach to capture space-time volumes (videos) involves spatial multiplexing. By placing an optical element in front of a sensor or camera (e.g., a diffuser or diffraction grating), the light from a single point in the scene spreads to multiple camera pixels according to the optical system's PSF. When combined with a rolling shutter camera that samples individual sensor rows at fast rates [2, 50, 26], it is possible to reconstruct high-speed videos. Alternatively, a limited region-of-interest (ROI) or 1D sensor can also be used to sample measurements at fast rates [40, 39]. By integrating diffraction patterns generated by time-varying coded masks, it is even possible to reconstruct 4D volumes of spacetime from a single snapshot [49].", + "bbox": [ + 75, + 506, + 467, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar ideas have been used in combination with streak cameras, an optoelectronic device used to measure ultra-fast light phenomena at femtosecond time scales. A streak camera spatially multiplexes the scene's temporal response, by deflecting photoelectrons to different points on a sensor as a function of time. While designed to be a line sensor, it is possible to encode 2D spatial information by incorporating a tilted microlens [45] or pinhole array [11]. In Gao et al. [8], the 2D scene at different timestamps is coded with a fixed mask, sheared a temporally varying amount, and then overlayed onto the 2D signal captured by a streak camera.", + "bbox": [ + 75, + 704, + 467, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These methods lack the (I) light efficiency and (III) density advantages of our approach—they typically rely", + "bbox": [ + 76, + 869, + 467, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ff9e9d3158198e325b52bd2d5f64a26e62377ec5dc4c139f65e95491e242c435.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 88, + 834, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/046888cc7066ba023f0954b21e72779acda91a1bc7415bbea8767ec4fb65a379.jpg", + "image_caption": [ + "(a) Captured image", + "(b) Recovered time signal (ms)", + "Figure 2. High-speed reconstruction of a flickering LED. The LED in this experiment flashes at $96\\mathrm{kHz}$ , as driven by an Arduino. (a) The diffraction image encodes the high-frequency patterns of the LED (zoom into inset). (b) This temporally-varying signal can be directly extracted from the image to recover the intensity pattern of the LED. A frequency analysis of the recovered time signal indicates the system's ability to resolve frequencies up to $96\\mathrm{kHz}$ . Given that the Nyquist frequency is half that of the sample rate, the effective framerate for the system is therefore over 192,000 FPS." + ], + "image_footnote": [], + "bbox": [ + 535, + 224, + 846, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on a small subset of camera pixels to reconstruct each frame, \"wasting\" light and limiting recoverable scene density. Furthermore, their capture speed is limited by the maximum speed of the hardware used (i.e., line-scan mode [40], ROIs [39] or rolling shutter [2, 50]). Thus, these methods lack the ability to (II) easily increase temporal resolution.", + "bbox": [ + 496, + 449, + 890, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. High-Speed Imaging using a Time-Varying Point-Spread Function", + "text_level": 1, + "bbox": [ + 500, + 551, + 890, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a scene that consists of a single bright LED in a dark room, where the LED is configured to pulse rapidly over time. Unless it is faster than the blinking LED, a normal camera forms an image of the LED, but provides no information about the LED's temporal dynamics.", + "bbox": [ + 496, + 598, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed approach to imaging high-speed events involves two steps. First, following prior work on diffraction imaging [40, 39], we place a diffraction grating in front of the camera, which blurs the scene with a rainbow PSF. The light from the LED is therefore spread over many additional camera pixels. Second, we propose to rotate the diffraction grating such that the corresponding PSF changes over time. This produces measurements that spatially encode temporally varying signals, which can be used to analyze extremely high-speed events (see Fig. 2 for an example).", + "bbox": [ + 496, + 674, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the image formation model and reconstruction process for recovering a high-speed video from an image captured with a generic time-varying PSF. We then analyze the special case of using a rotating PSF for capturing images.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10791", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/01965abe5cea810eac69f31f79cf91612ca3f8b39ee3975e512799c5cde86b82.jpg", + "image_caption": [ + "Rolling $^+$ diffuser [2] (64 row exposure)" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 217, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dac4d2807bf373bd873263f60d65e44a9e94fbb0c1535f7847efc409ed2656a7.jpg", + "image_caption": [ + "ROI + grating [39] (16 row ROI)" + ], + "image_footnote": [], + "bbox": [ + 218, + 88, + 334, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/227aa7f4891ec1d7d47abb9c09d833cfb578d1c163967a7c7c09aa1f77ba8a1d.jpg", + "image_caption": [ + "Rotating PSF" + ], + "image_footnote": [], + "bbox": [ + 336, + 88, + 454, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d1f555643eb4531ba875b6077d6408e451f399a999d3405af70a087574b15c01.jpg", + "image_caption": [ + "(a) Recovered brightness of time-varying light source" + ], + "image_footnote": [], + "bbox": [ + 84, + 178, + 454, + 255 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f9c5ed59b216b5477c59963999ea4dadeff50591458faab07bede059eb8a3f80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 280, + 215, + 349 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/904dae00d73e8faf386a6baa51ef674475d37a55320431a5a07b391603decf82.jpg", + "image_caption": [ + "Rolling $^+$ diffuser [2] (64 row exposure)", + "(b) Recovery of dense moving scene", + "Figure 3. Light efficiency experiment. In the presence of Poisson and read noise, we simulated the measurements captured by a Canon 5D Mark III sensor [6] with ISO 800. (a) We simulate a dim point light source with a time-varying intensity profile. (b) We simulate the word \"test\" moving across the scene. In both cases, our approach is able to much more accurately resolve the signal, compared to a rolling shutter and diffuser as well as a ROI and grating. We theorize that this is due to the fact that our approach captures more light for every time step, and that the requisite demultiplexing minimizes the effects of noise." + ], + "image_footnote": [], + "bbox": [ + 86, + 351, + 215, + 417 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd01ad50f52be70cca44dd33940637c2862d8338328822c387761c44d661843b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 281, + 334, + 349 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69ea9149a41dea49faf869542d3ee0c232c807ad2a168c1142ced317ca5e6be0.jpg", + "image_caption": [ + "ROI + grating [39] (16 row ROI)" + ], + "image_footnote": [], + "bbox": [ + 218, + 351, + 336, + 417 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b283a5de6050ff094b2edca34f947b1ea19c92d6c60f6ed2a64f33ee79020297.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 281, + 454, + 349 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/136f9d6de205a41cc5d13d5135ed0a542751661873120b57edd42202f53a49ff.jpg", + "image_caption": [ + "Rotating PSF" + ], + "image_footnote": [], + "bbox": [ + 338, + 351, + 455, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Image Formation Model", + "text_level": 1, + "bbox": [ + 76, + 628, + 300, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, for a camera where the PSF changes with time, the forward model can be written as:", + "bbox": [ + 76, + 654, + 468, + 685 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} (x, y) = \\int_ {t = t _ {s}} ^ {t = t _ {s} + E} (\\mathbf {k} (t) * \\mathbf {s} (t)) (x, y) d t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 702, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{I}(x,y)$ is the captured image, $E$ denotes the exposure time of the camera, and $t_s$ denotes the time when the camera started exposing. The symbol $*$ denotes the spatial convolution operator between the scene $\\mathbf{s}(t)$ and the PSF $\\mathbf{k}(t)$ at time $t$ . When discretized to $N_E$ time steps from $t_s$ to $t_s + E$ , this can be rewritten in matrix form as:", + "bbox": [ + 75, + 752, + 468, + 843 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} = \\mathbf {M} \\mathbf {s} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 859, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\left[ \\begin{array}{l l l l} \\mathbf {C} _ {1} & \\mathbf {C} _ {2} & \\dots & \\mathbf {C} _ {N _ {E}} \\end{array} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 878, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{I} \\in \\mathbb{R}_{+}^{N_{x}N_{y}}$ is the vectorized representation of the captured image, $\\mathbf{s} \\in \\mathbb{R}_{+}^{N_{x}N_{y}N_{E}}$ is the vectorized representation of the high-speed scene, and $\\mathbf{C}_k \\in \\mathbb{R}_{+}^{N_xN_y\\times N_xN_y}$ represents the convolution with the kernel at timestep $k$ .", + "bbox": [ + 500, + 88, + 890, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Reconstruction Process", + "text_level": 1, + "bbox": [ + 500, + 167, + 714, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Natural scenes contain sparse gradients, and many target applications for high-speed imaging require the tracking of just sparse points. As a result, we recover a high-speed video with an $\\ell_1$ sparsity prior and an anisotropic total variation prior, solving the following optimization problem:", + "bbox": [ + 498, + 190, + 890, + 266 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} \\quad \\| \\mathbf {I} - \\mathbf {M} \\mathbf {x} \\| ^ {2} + \\mathcal {I} _ {\\mathbb {R} _ {+}} (\\mathbf {x}) + \\lambda_ {\\text {s p a r s i t y}} \\| \\mathbf {x} \\| _ {1} + \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 273, + 890, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {d x} \\| \\mathbf {D} _ {x} \\mathbf {x} \\| _ {1} + \\lambda_ {d y} \\| \\mathbf {D} _ {y} \\mathbf {x} \\| _ {1} + \\lambda_ {d t} \\| \\mathbf {D} _ {t} \\mathbf {x} \\| _ {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 304, + 857, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_{\\mathbb{R}_+}$ denotes a non-negativity prior, $\\mathbf{D}_x$ , $\\mathbf{D}_y$ , and $\\mathbf{D}_t$ compute forward finite differences along the $x$ , $y$ , and $t$ dimensions respectively of the high-speed video $\\mathbf{x}$ , and $\\lambda_{sparsity}$ , $\\lambda_{dx}$ , $\\lambda_{dy}$ , and $\\lambda_{dt}$ weigh the regularization terms.", + "bbox": [ + 500, + 328, + 890, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Analysis", + "text_level": 1, + "bbox": [ + 500, + 398, + 601, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Varying the PSF over time results in fundamentally different properties compared to prior approaches, which we explore in the rest of this section.", + "bbox": [ + 498, + 421, + 890, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Higher light efficiency. Our method, like other prior compressive video works [2, 50, 40, 39], relies on PSFs with broad spatial supports. A broad spatial support spreads the signal from a single point across multiple pixels. Recent approaches record just a subset of these pixels at each timestep [2, 50, 40, 39]. Because fewer pixels encode each frame and light is lost, assuming a standard noise model with Poisson shot noise and additive Gaussian noise, the reconstruction's signal-to-noise ratio is proportionately decreased. In contrast, under a time-varying PSF, the entire frame is always recorded, so a much larger fraction of the PSF (and therefore light) is captured for every scene point during capture. In addition, the signals from different timestamps are mixed together in the final capture, which could improve the quality of the reconstructions under read noise [37]. Thus, this overall increase in light efficiency leads to better performance in resolving noisy scenes. We show this experimentally in Fig. 3 on two simulated cases: a point light source with time-varying intensity, and a denser moving scene. We find that our approach robustly recovers these two scenes, even under low light conditions.", + "bbox": [ + 496, + 470, + 892, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A wider class of scenes. Many past approaches achieve high-speeds by allocating a number of pixels per frames that is much lower than the total number of frame pixels [2, 50, 40, 39]. This causes each frame to undergo lossy compression, yielding a sparsity constraint per frame. For example, for a rolling shutter with exposure set to the row rate, each frame is encoded onto just a single sensor row.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10792", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d432968d16ec4e19a92459dc808b633e0ebdb7bcc216f8a5d04d03ebe5c29026.jpg", + "image_caption": [ + "(a) Simulated scene:", + "(1) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 76, + 101, + 200, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d000aec2211483a15a7d27bbd0fa8d47bb9f062a13a5751925af30ad2830c416.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 101, + 354, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a5e8b014d322cfd0ddde00e5e429c73463001a349f8df40f49b226dc96552969.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 178, + 354, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e901f427caf1a3fd9567a48f45e8ee960a9650bc3d26da70af3e9131e9862719.jpg", + "image_caption": [ + "(2) Rolling + diffuser (1 row exposure)" + ], + "image_footnote": [], + "bbox": [ + 222, + 253, + 354, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e24694fe7297650688b2d64d41043775fcb781b46597a398428cca98e4a58cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 101, + 485, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/188ddd3c8f10b71fadba1178662df2b08f59c5da74d172088e11b5e9c77d61fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 178, + 486, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c5c869ffb06cc00be5774852da6f32a5a0937c4809d2ece88c28517a0edcd158.jpg", + "image_caption": [ + "Rolling $^+$ diffuser [2] (64 row exposure)" + ], + "image_footnote": [], + "bbox": [ + 354, + 253, + 486, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1e3d10399923c49f903784a620cb3889e2b8cbf3b6e9f9363104a08924b03329.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 101, + 616, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/966659a7743685c7cc77b46b2eb15e298a82dfe1ac458d24ed07b7689d092108.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 178, + 617, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/74d0b8820f71c4f8406782d741b6dd9b9a47bfc9ad845e7262c7dea3c6eb58f6.jpg", + "image_caption": [ + "(4) ROI + grating [39] (2 row ROI)" + ], + "image_footnote": [], + "bbox": [ + 488, + 253, + 617, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8f7d9239f844fdbce4e14491ddf9fbf7c38101051b7b1030e26f9fb0ae5f2670.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 101, + 750, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e9979480d83129dcf248ebe41276a2f3f9aca2182a4a3ddbe262f6db4833d8d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 178, + 750, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c852d951ceffc07a5f74246792081b7450de61742ca0d5859c9747dcfdb71d10.jpg", + "image_caption": [ + "(5) ROI + grating [39] (16 row ROI)" + ], + "image_footnote": [], + "bbox": [ + 619, + 253, + 750, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9219cebf810a69b0003e1f37155522c0122e55709aae4bf07eeac749b38dcff9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 101, + 882, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1b2c0d35f239a6724e1fc2ffec104e67f35c9ac041f9c6ea0a519b68148ab01a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 178, + 882, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f1f35ff8cf4a4c469b618a90fb3f5c9d9cc9eb34f4a1c0e25420ad32903fd54c.jpg", + "image_caption": [ + "(6) Rotating PSF" + ], + "image_footnote": [], + "bbox": [ + 751, + 253, + 882, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/485458d0fed3a27c6c7e34ef794be5a2f206169e3fafd6e9b327340b96201c77.jpg", + "image_caption": [ + "(b) Real scene:", + "(1) Reference frame", + "Figure 4. Reconstructing temporally sparse, but spatially dense scenes. (a) A simulated scene of two people dancing. The video is dark for the entire sequence, except for a short sequence of fast-moving frames where the dancers are strobed. (a2)-(a3) We implement a setup based on a dual rolling shutter combined with a diffuser [2]. In (a2), with the exposure time set to match the row rate of the shutter, the reconstruction is unable to well-resolve the dense scene. In (a3), with the exposure time set to 64 times the previous value, the reconstruction is able to recover the dancers, at the cost of temporal resolution [2]. (a4)-(a5) We implement an approach based on fast temporal sampling by using a ROI and a diffraction grating [39]. In (a4), the ROI consists of just two rows, resulting in low quality results. In (a5), by using more bandwidth and exposing 16 rows, such an approach can get better results, but this reduces the recoverable framerate by 8 times in practice. (a6) In contrast, our methodology recovers the dense scene much more accurately compared to (a2) and (a4)—the signal is encoded over a much larger portion of sensor. This does not require increasing the exposure time and therefore decreasing the temporal resolution as in (a3), or increasing the requisite bandwidth (a5). (b) A real world scene, where a spinning fan is strobed twice during the camera exposure at unknown times. Our approach is able to disambiguate the fan rotation at each strobe time." + ], + "image_footnote": [], + "bbox": [ + 81, + 366, + 279, + 478 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/507f6def4fc57260c121ca72a3cdece98ed1569d392c0a7b3e527274bb46bc8d.jpg", + "image_caption": [ + "(2) Rotation-encoded frame" + ], + "image_footnote": [], + "bbox": [ + 282, + 366, + 482, + 478 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/60446f80bb913cb5584e00b4b3a839b7aeb90edf4e3173d8f69666cee91bcd9e.jpg", + "image_caption": [ + "(3) Recovered frames" + ], + "image_footnote": [], + "bbox": [ + 483, + 366, + 684, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, our method allocates the entire frame to encode a plurality of video frames belonging to a fixed time interval. Therefore, our method assumes sparsity in a more general sense, namely sparsity over the entire fixed time interval. The encoded fixed time interval can be either spatially sparse and temporally dense, which fits most prior-work assumptions, or temporally sparse but spatially dense, which prior approaches do not consider due to the fixed bitrate per frame imposed by their low-dimensional encoding.", + "bbox": [ + 75, + 676, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To illustrate how this is an advantage, consider the case of a dense scene that is entirely dark, except for a single time instant with a flash of light. If sampling the scene using a rolling-shutter or line sensor, the flashed frame will be encoded onto just a single row. In contrast, a time-varying", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PSF encodes the flashed frame on the entire frame, wasting no resources encoding the dark frames. Therefore, in this case, a time-varying PSF enables a better reconstruction.", + "bbox": [ + 496, + 676, + 890, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 4, handling temporally sparse, but spatially dense scenes makes our approach very useful under strobed lighting. As shown in real and simulated experiments, a rotating PSF is able to resolve the content within each individual strobe at high spatiotemporal resolution, without any constraints on scene structure between strobes.", + "bbox": [ + 496, + 722, + 890, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Using a Rotating PSF", + "text_level": 1, + "bbox": [ + 500, + 828, + 702, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sec. 3 described the encoding and reconstruction of high-speed video using a time-varying PSF. While a time-varying PSF can be implemented in a number of ways,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10793", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/54be67f8effe5d6f1fd305e1f872fa462dbb6fb9250e5d040959abc2cc5abeae.jpg", + "image_caption": [ + "Figure 5. The PSF and spatial frequency coding of time. A rotating PSF effectively codes the signal from different timesteps into a rotating set of spatial frequencies in the captured image. For instance, a rotating line PSF encodes different timestamps into different lines in Fourier space. The star PSF from a double-axis diffraction grating behaves as a mixture of the point and line cases. Its Fourier transform contains bright line components that provide temporal signal, but remains broadband for spatial resolution." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 467, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the sake of cost and speed, we simply rotate a PSF-generating optical element over the course of the camera exposure. This results in a rotation of the corresponding PSF. In this particular case, $\\mathbf{k}(t)$ can now be rewritten as $\\mathbf{k}(t) = \\mathbf{R}(\\mathbf{k},\\theta_t)$ , where $\\mathbf{k}$ is the stationary PSF of the rotating optical element, $\\theta_{t}$ is the angle of the element at time $t$ , and $\\mathbf{R}(\\mathbf{k},\\theta_t)$ denotes a rotation of $\\mathbf{k}$ by $\\theta_{t}$ . In the rest of this section, we analyze the properties of such a rotating setup.", + "bbox": [ + 75, + 393, + 468, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PSF performance. The selected PSF affects the spatiotemporal reconstruction quality of every video frame. Because each time step is convolved with a rotated PSF, the PSF's rotated Fourier transform determines which parts of the spectrum will encode each time step. To illustrate, an ideal thin lens having a perfect impulse PSF will have perfect spatial resolution because its Fourier transform contributes values in all spectrum frequencies (Fig. 5(a)). It will, however, have poor temporal resolution because every other timestamp will also contribute values across the entire spectrum, making them impossible to distinguish.", + "bbox": [ + 75, + 522, + 468, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast, consider a line PSF (Fig. 5(b)). In the Fourier domain, the PSF remains a line that goes through the origin. When the line PSF rotates, its Fourier spectrum rotates correspondingly, such that each time step is encoded into a different line in Fourier space, without interfering with values from previous timestamps. As a result, the signal at each time step can be easily reconstructed simply by extracting its corresponding line from Fourier space. However, because only a single direction of spatial frequencies is available for reconstruction for every time step, the re", + "bbox": [ + 75, + 689, + 468, + 840 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7efc04c745edaac170df2bcc64f3b366a07f6fd5eace08118a557c7e0ead0e9f.jpg", + "image_caption": [ + "(a) Setup" + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 700, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56dc7d3ed5f2f7dcb5744e972b3c553fbebb00b03fa4ece179945021f1b9090b.jpg", + "image_caption": [ + "(b) Captured PSF", + "Figure 6. Image of our hardware setup, consisting of a global-shutter camera and a circular diffraction grating attached to a 20,000 RPM motor." + ], + "image_footnote": [], + "bbox": [ + 704, + 106, + 898, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sulting recovered frame will have poor spatial resolution.", + "bbox": [ + 500, + 345, + 877, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by these two cases, we use a double-axis diffraction grating for the experiments shown in this work, whose corresponding PSF is shown in Fig. 6(b). This star PSF contains a bright DC spot at the center, intuitively providing spatial resolution. Simultaneously, the PSF also contains line-shaped rainbow lobes that provide temporal information. We visualize this phenomena in Fig. 5(c).", + "bbox": [ + 496, + 361, + 890, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Temporal resolution. The temporal resolution of our system is fundamentally determined by the amount of temporal variation of the PSF between different timestamps, or namely, the minimum angle required to reliably distinguish between two different rotations of the PSF. As a result, the temporal resolution can be increased either by increasing the pixel resolution, the spatial spread of the PSF, or the rotation speed. But since our grating PSF is roughly $90^{\\circ}$ periodic, rotating more than $90^{\\circ}$ during the exposure causes temporal ambiguity and degrades the reconstruction quality.", + "bbox": [ + 496, + 479, + 892, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the supplement, we further analyze PSF performance from the mutual coherence perspective. We also discuss how a time-varying PSF behaves as a form of coded exposure in frequency, as shown by the above Fourier analysis.", + "bbox": [ + 496, + 630, + 890, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Implementation", + "text_level": 1, + "bbox": [ + 500, + 708, + 658, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hardware. An image of our hardware setup is given in Fig. 6. To generate a usable PSF, we cut a 531-lines per millimeter double-axis diffraction grating sheet into a circular disk of radius 6 inches. To ensure that the sheet stayed flat, we taped it onto a laser-cut acrylic disk with a radius of 6 inches. We attached this element to a Puly DC 775 Motor with a shaft coupling to rotate it. We place this entire rotating setup $5\\mathrm{mm}$ in front of the camera's lens. We used a UI-3070CP-C-HQ R2 camera, fitted with a $9\\mathrm{mm}f / 1.4$ lens (Fujinon HF9HA). We also placed a second identical camera beside our system to capture intermittent reference", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1The Fourier transform of the convolution of two functions is the element-wise product of their Fourier transforms. If a function is rotated by $\\theta$ , its Fourier transform is the Fourier transform of the original function, rotated by $\\theta$ .", + "bbox": [ + 75, + 849, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10794", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/562699de3ebf26737e1e23971575f024f3aae293a3bcc6de179811241926ceb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 88, + 285, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/546cc8099894e51416c6cb1002c77f4b397b3be5019dba3d88882d77ba0b9188.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 166, + 285, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f2199849286190feb69e57e6c590b79331dbe38f5ea07da03de5ebd50c4b6a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 244, + 284, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc714076934e57c7a33a400aac193490aa16c1ded52d8970146f8a2c921ac059.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 340, + 284, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f8d12569c3902f258b4c1f56c0356796152cfbe763414fa468f2221eaca6045.jpg", + "image_caption": [ + "(a) Reference start frame" + ], + "image_footnote": [], + "bbox": [ + 88, + 435, + 284, + 530 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec60926cb0bb1ca8ae17254120ec9bebf3eff312c7abebacd20673bf572706e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 88, + 482, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/beda0f73715a8a296e45cbdc2e0623d789a4383bc67f4cf51051a20f8e1ea869.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 166, + 480, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e14fe5640c9ef078c6b81089c755ae7fac784f18c3570cb4bb8164d173a5629b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 243, + 480, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85d9be2171d891c0efa929a139ca1d822d87dcd224504e5dea29b2bd4a57203d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 340, + 480, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/034645de103e544372b6ec03c146fcd76d0254da7b66b058fe607edd29c740c9.jpg", + "image_caption": [ + "(b) Reference end frame" + ], + "image_footnote": [], + "bbox": [ + 287, + 435, + 480, + 530 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84bed79ee189f771c0457c284eaa6dc5bd6847eacd36b35ec7923973f01823a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 88, + 679, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f522b3fce1e253e05adc65308b19972809c54ffc19fd6a3e32abb0dcc1c9b59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 166, + 679, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d1211679be31115af89373bd99fa07b6a9c06a141f34af32cffe651132a6ef1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 243, + 679, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9cf22ae89555790b2d9b9817e36b1ddb49a4203e24aeff90aad7962f2ad65352.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 340, + 679, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e9d4b45b41132e11641964f2f33f04a564b8fac590a340215d708ad31f94413.jpg", + "image_caption": [ + "(c) Rotation-encoded frame", + "Figure 7. High-speed reconstruction of fast moving objects. Row 1: A ball with a black stripe bounces and spins. Row 2: An I rotates $120^{\\circ}$ . Row 3: A toy bullet bouncing off a board. Row 4, 5: Our setup is capable of tracking the movements of both sparse points and dense marker lines for high-speed motion capture." + ], + "image_footnote": [], + "bbox": [ + 483, + 435, + 679, + 530 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e1aee89aa1e75e7b1267cf5d573889caeef574bb1b5299c93baef001b574753.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 89, + 877, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0184d67e45d66823dee45077c4d53c173cfc92fecda4f9f8c66e4ef8ac5bf129.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 166, + 877, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93088bc551ca8fdc2f9a1f8926a303063e02ca7899e266867b187dd11a48d407.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 243, + 877, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ed38485188a7ac5c5fcd92e351904366a0ca25042153962b2300409cb6f1d0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 340, + 877, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d52a0d868b2bfb51bff0591593002071c1a9d87e606d118145b2ce6b4e84bd81.jpg", + "image_caption": [ + "(d) Recovered frames" + ], + "image_footnote": [], + "bbox": [ + 681, + 435, + 877, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames for visualization. Please refer to the supplement for a DIY guide.", + "bbox": [ + 75, + 616, + 468, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Software. To solve Eq. (4), we used a PyTorch implementation of FISTA [3], combined with an approximate proximal operator [15] to handle the anisotropic total variation prior. We ran 10000 iterations of FISTA for each result shown. For most results, we set $\\lambda_{sparsity} = 0$ , $\\lambda_{dx} = \\lambda_{dy} = 1 \\times 10^{-3}$ , and $\\lambda_{dt} = 1 \\times 10^{-3}$ , but we vary the strength of these terms depending on the brightness of the background and the speed of the motion. We list the parameters for each result and provide sample code in the supplemental.", + "bbox": [ + 75, + 652, + 468, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Calibration and capture. To calibrate the time-varying PSF of our imaging system, we process a stationary image of a small retroreflector or LED, and then computationally rotate this image to create the dictionary required for Eq. (4)—please see the supplemental for more details.", + "bbox": [ + 75, + 794, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For every capture, we need to identify the range of angles that are rotated over during the camera exposure in order to", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "solve Eq. (4). This could be done by placing an LED at a known location, and reconstructing which PSFs are present around it. Alternatively, a controllable motor or trigger system could be used to directly provide the requisite angles. In our work, we opt for a unstructured approach, that simply relies on signal being present at the start and end of the camera exposure. Before capture, we first calibrate the motor to rotate roughly $90^{\\circ}$ over the course of the exposure. Then, we simply perform a coarse reconstruction of Eq. (4) using the captured image and a dictionary that spans all $360^{\\circ}$ , and select the best matching $90^{\\circ}$ for a finer reconstruction. Please see the supplemental for an in-depth discussion.", + "bbox": [ + 496, + 616, + 890, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Results", + "text_level": 1, + "bbox": [ + 500, + 814, + 586, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For our real-world reconstructions, we downsampled the input images to $300 \\times 400$ pixels and reconstructed $N_{E} = 146$ frames for every input. For the simulated results (Fig. 3, Fig. 4), we instead used $N_{E} = 150$ to better match", + "bbox": [ + 500, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10795", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28e43121f66334602705b045cf561eb19fd29b878e924710ad249695d3981920.jpg", + "image_caption": [ + "(a) Rotation-encoded frame" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 284, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36b446289bfa65bf028b8c39f266a788242224c793fe5b456be5e7a4610ff48f.jpg", + "image_caption": [ + "(b) Example recovered frame" + ], + "image_footnote": [], + "bbox": [ + 285, + 88, + 482, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6dab6558f77b3a5a4d9aa36bcfc2e4ee898538962ec4e0b48aee4d3ed5a87e67.jpg", + "image_caption": [ + "(c) Flow visualization", + "Figure 8. High-speed particle image velocimetry. Following Sheinin et al. [39], a magnetic stirrer is placed in a tank seeded with tracer particles. (a) Captured diffraction image. (b) Recovered frame using our reconstruction procedure. (c) We visualize a $17.33\\mathrm{ms}$ interval by overlaying the recovered frames. (d) Using OpenPIV [17], we recover the flow between the start and end frame of (c)." + ], + "image_footnote": [], + "bbox": [ + 483, + 88, + 681, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/75d3da9bee7fb70c8070a37ab437c90ea2ce23add0b6717fed5bed470f08e40d.jpg", + "image_caption": [ + "(d) Recovered flow" + ], + "image_footnote": [], + "bbox": [ + 694, + 94, + 872, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a dual rolling shutter setup. In this document, we use color to code a video into a single image; we refer the reader to our supplemental materials for directly viewing the recovered videos. The camera exposure was $22.178\\mathrm{ms}$ , yielding an effective framerate of 6583.10 FPS.", + "bbox": [ + 75, + 282, + 467, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate our methodology on fast scenes on multiple examples (Fig. 7)—a bouncing ball, an \"I\" spinning on a fan, toy bullets bouncing off a board, and letters spinning and falling. In the last row of Fig. 7, we applied our technique to marker-based motion capture. Our technique readily captures a wide variety of marker shapes. In Fig. 4(b), we show a temporally sparse, but spatially dense scene of a fan under strobed illumination.", + "bbox": [ + 75, + 358, + 467, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown by Sheinin et al. [39], low-cost, high-speed imaging could help measure the movement of fluids in particle image velocimetry [1, 21, 52]. In short, small tracer particles are placed into the medium, which can then be tracked to visualize the flow. We apply our setup to this task in Fig. 8, imaging a tank filled with water and a magnetic stirrer. To maximize the visible movement, we increased the exposure time of our camera to $100\\mathrm{ms}$ and proportionately decreased the speed of the motor.", + "bbox": [ + 75, + 479, + 467, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we push the temporal resolution of our approach by imaging an LED pulsed at $96\\mathrm{kHz}$ . For this particular result, we used a $16\\mathrm{mm}$ focal length lens (Edmund Optics #86-571) to increase the scale of the PSF, and rotated the motor at roughly 15,000 RPM. At resolution $2048\\times 1542$ pixels, our methodology is capable of easily resolving this signal as shown in Fig. 2, demonstrating $192\\mathrm{kHz}$ capture.", + "bbox": [ + 75, + 617, + 467, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitations", + "text_level": 1, + "bbox": [ + 76, + 738, + 199, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal unmixing. Rolling shutter or ROI-based approaches [2, 50, 40, 39] code the signal from each timestamp onto a unique set of measurements that only record the signal at that time. In contrast, with a rotating PSF, the signal from different timestamps is mixed together and requires unmixing using Eq. (4). This unmixing can yield degraded results when there are significant nonidealities, e.g., a poorly calibrated PSF, major spatial variation of the PSF, sizable background signal, or when the scene is too dense.", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d9771b55af5f6eab0a9001c5c7ad5ebbb4510ced9c529f3c65a7db406ab85f5.jpg", + "image_caption": [ + "PSF spect.", + "Rec. Mx" + ], + "image_footnote": [], + "bbox": [ + 504, + 279, + 630, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 573, + 280, + 629, + 319 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9de95f1966c497a6c935a12608df5aa7f2dd1074a123afd8b754901f06047179.jpg", + "image_caption": [ + "PSF spect" + ], + "image_footnote": [], + "bbox": [ + 630, + 280, + 697, + 320 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e61f5974bd20e2f1d8c9430ba7b45b0b3586de7cf7e5f208be4b67ab5c6455ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 279, + 756, + 320 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f967a5316a8c5a108e5acd302061bd03aab7b932529b7aaba2d649c2ed26691.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 279, + 883, + 320 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/63cb8deb4620193d4f109942edf6de57c1e902e69b79c7ff9d0d4262f5c47909.jpg", + "image_caption": [ + "Sample recons. frame", + "Original PSF", + "Figure 9. Reconstruction sensitivity to PSF spectrum. We reconstruct Fig. 7, Row 2 with digitally augmented versions of the original real measured PSF, to simulate a PSF calibrated under incorrect lighting or from a different material. The reconstruction quality is fairly robust to these PSF variations." + ], + "image_footnote": [], + "bbox": [ + 504, + 329, + 629, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4e44bc08f088db1da0e63300f474cf41ae9fb1ca599477519e5e49c128e183a.jpg", + "image_caption": [ + "Rec. $\\mathbf{M}\\mathbf{x}$", + "Sample recons. frame", + "Bandpass filtered PSF" + ], + "image_footnote": [], + "bbox": [ + 630, + 329, + 756, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37e30f913d02aa6ada26293d65ec955db406064ef53c5bb841d09b3acb52b099.jpg", + "image_caption": [ + "Rec. Mx", + "Sample recons. frame", + "Notch filtered PSF" + ], + "image_footnote": [], + "bbox": [ + 759, + 329, + 883, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In these cases, the signal from one timestamp may bleed into the reconstructions for other timestamps, causing a loss of temporal resolution.", + "bbox": [ + 498, + 493, + 890, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Saturation. Our method can be especially prone to sensor saturation for several reasons. First, the use of a longer camera exposure may saturate those camera pixels which received scene light over a significant portion of the exposure duration. Secondly, a bright flash of light at a single timestamp could saturate the entire sensor, corrupting the whole video sequence, while only a single row (and therefore timestamp) would be saturated for a rolling shutter or ROI. Thirdly, our grating's PSF contains a 'DC spot' that is much brighter than the rainbow streaks. Therefore, well-exposed rainbow streaks yield a large DC component that might saturate camera pixels. This latter problem could be mitigated by using a higher-quality grating.", + "bbox": [ + 496, + 545, + 890, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spectrum. The PSF of an object, when viewed through a diffraction grating, depends on its reflectance spectrum, which itself depends on the light source spectrum and object material. Our approach assumes that scenes consist of a single reflectance spectrum so that this PSF is known and fixed for every point in the scene. This assumption makes our approach well-suited for vision applications like motion capture or PIV, which use specialized light sources and markers. In these cases, only slight deviations in PSF occur, which our method can handle (see Fig. 9). However, a", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10796", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "more general scene may consist of many diverse materials and light sources. This results in many different reflectance spectrums and corresponding PSFs at the sensor, restricting the applicability of our method.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To address this limitation, the single reflectance spectrum assumption can potentially be relaxed, albeit with tradeoffs. For instance, the reconstruction problem in Eq. (4) can be expanded to explicitly account for these multiple PSFs following Sheinin et al. [39], at the cost of extra computational complexity that scales with the number of possible spectrums. Alternatively, a less wavelength-dependent optic (e.g., a simple diffuser [2]) could be used, such that the PSF does not significantly vary for different wavelengths. However, this loses the additional degrees of information encoded in the different PSF colors.", + "bbox": [ + 75, + 152, + 470, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "8. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 76, + 333, + 346, + 349 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We demonstrated a novel compressive high-speed video methodology that relies on temporally coding the camera's PSF using a spinning diffraction grating. Our approach can turn ordinary cameras, regardless of their designed speed, into high-speed sensors for spatiotemporally sparse scenes. Compared to prior work, our method can tackle different scene types, is more light efficient, and its temporal resolution is easily increased.", + "bbox": [ + 75, + 359, + 468, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our prototype relies on a PSF created by a low-cost diffraction grating. Better PSFs, including ones that are optimized using an end-to-end framework [14, 22, 30], would likely produce superior results. Moreover, there are likely more sophisticated mechanical ways to temporally modulate the PSF, potentially providing better performance. For instance, a fast stepper motor that rotates the PSF at discrete steps, synchronized with the reconstruction resolution, could help minimize the mismatch between the discretized model in Eq. (4) and actual measurements. Additionally, other forms of temporal modulation beyond rotation could also prove useful, such as axially translating the spatial-multiplexing optic to scale the PSF, or using a spatial-light modulator to create arbitrary programmable PSFs.", + "bbox": [ + 75, + 481, + 468, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Additionally, better reconstruction procedures that leverage optical flow [34], dictionary learning [12, 13], or deep learning [14, 22, 30] could significantly improve quality over our naive implementations. Finally, our method is orthogonal to previous compressive video approaches mentioned in Sec. 2; through novel combinations of these techniques, we believe it may be possible to spin out reconstructions at even higher spatial and temporal resolutions.", + "bbox": [ + 75, + 693, + 468, + 813 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Amitabh Vyas and Ioannis Gkioulekas for fruitful discussions, and Justin Macey for lending us the motion capture equipment. Matthew O'Toole was supported by a National Science Foundation CAREER award (IIS 2238485).", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 90, + 596, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ronald J Adrian and Jerry Westerweel. Particle image velocimetry. Number 30. Cambridge university press, 2011. 8", + "[2] Nick Antipa, Patrick Oare, Emrah Bostan, Ren Ng, and Laura Waller. Video from stills: Lensless imaging with rolling shutter. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2019. 2, 3, 4, 5, 8, 9", + "[3] Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences, 2(1):183-202, 2009. 7", + "[4] Joyce Bedi, Douglas Collins, and Harold Eugene Edgerton. Seeing the unseen: Dr. Harold E. Edgerton and the wonders of Strobe Alley. mIt press, 1994. 3", + "[5] Jacky CP Chan, Howard Leung, Jeff KT Tang, and Taku Komura. A virtual reality dance training system using motion capture technology. IEEE transactions on learning technologies, 4(2):187-195, 2010. 2", + "[6] Roger N. Clark. Clarkvision.com. 4", + "[7] P Doron, L Bertuccioli, J Katz, and TR Osborn. Turbulence characteristics and dissipation estimates in the coastal ocean bottom boundary layer from pv data. Journal of Physical Oceanography, 31(8):2108-2134, 2001. 2", + "[8] Liang Gao, Jinyang Liang, Chiye Li, and Lihong V Wang. Single-shot compressed ultrafast photography at one hundred billion frames per second. Nature, 516(7529):74-77, 2014. 3", + "[9] Jinwei Gu, Yasunobu Hitomi, Tomoo Mitsunaga, and Shree Nayar. Coded rolling shutter photography: Flexible spacetime sampling. In 2010 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2010. 2, 3", + "[10] Evgeny Hahamovich, Sagi Monin, Yoav Hazan, and Amir Rosenthal. Single pixel imaging at megahertz switching rates via cyclic hadamard masks. Nature communications, 12(1):1-6, 2021. 2", + "[11] Barmak Heshmat, Guy Satat, Christopher Barsi, and Ramesh Raskar. Single-shot ultrafast imaging using parallax-free alignment with a tilted lenslet array. In CLEO: Science and Innovations, pages STu3E-7. Optica Publishing Group, 2014. 3", + "[12] Yasunobu Hitomi, Jinwei Gu, Mohit Gupta, Tomoo Mitsunaga, and Shree K Nayar. Video from a single coded exposure photograph using a learned over-complete dictionary. In 2011 International Conference on Computer Vision, pages 287-294. IEEE, 2011. 2, 3, 9", + "[13] Jason Holloway, Aswin C Sankaranarayanan, Ashok Veeraraghavan, and Salil Tambe. Flutter shutter video camera for compressive sensing of videos. In 2012 IEEE International Conference on Computational Photography (ICCP), pages 1-9. IEEE, 2012. 2, 3, 9", + "[14] Michael Iliadis, Leonidas Spinoulas, and Aggelos Katsaggelos. Deepbinarymask: Learning a binary mask for video compressive sensing. Digital Signal Processing, 96:102591, 2020. 2, 3, 9" + ], + "bbox": [ + 501, + 114, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10797", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Ulugbek S Kamilov. A parallel proximal algorithm for anisotropic total variation minimization. IEEE Transactions on Image Processing, 26(2):539-548, 2016. 7", + "[16] Roman Koller, Lukas Schmid, Nathan Matsuda, Thomas Niederberger, Leonidas Spinoulas, Oliver Cossairt, Guido Schuster, and Aggelos Katsaggelos. High spatio-temporal resolution video with compressed sensing. Optics express, 23(12):15992-16007, 2015. 2, 3", + "[17] Alex Liberzon, Theo Käufer, Andreas Bauer, Peter Vennemann, and Erich Zimmer. Openpiv/openpivpython: Openpivpython v0.23.4, Jan. 2021. 8", + "[18] Rui Lima, Shigeo Wada, Shuji Tanaka, Motohiro Takeda, Takuji Ishikawa, Ken-ichi Tsubota, Yohsuke Imai, and Takami Yamaguchi. In vitro blood flow in a rectangular pdms microchannel: experimental observations using a confocal micro-piv system. Biomedical microdevices, 10:153-167, 2008. 2", + "[19] Yang Liu, Xin Yuan, Jinli Suo, David J Brady, and Qionghai Dai. Rank minimization for snapshot compressive imaging. IEEE transactions on pattern analysis and machine intelligence, 41(12):2990-3006, 2018. 2, 3", + "[20] Patrick Llull, Xuejun Liao, Xin Yuan, Jianbo Yang, David Kittle, Lawrence Carin, Guillermo Sapiro, and David J Brady. Coded aperture compressive temporal imaging. Optics express, 21(9):10526-10545, 2013. 2, 3", + "[21] LM Lourenco, A Krothapalli, and CA Smith. Particle image velocimetry. In Advances in fluid mechanics measurements, pages 127-199. Springer, 1989. 8", + "[22] Julien NP Martel, Lorenz K Mueller, Stephen J Carey, Piotr Dudek, and Gordon Wetzstein. Neural sensors: Learning pixel exposures for hdr imaging and video compressive sensing with programmable sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1642-1653, 2020. 2, 3, 9", + "[23] Amir Matin and Xu Wang. Compressive coded rotating mirror camera for high-speed imaging. In Photonics, volume 8, page 34. MDPI, 2021. 2", + "[24] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2", + "[25] Albert Abraham Michelson. Experimental determination of the velocity of light: made at the US Naval Academy, Annapolis, volume 1. US Nautical Almanac Office, 1880. 2", + "[26] Kristina Monakhova, Vi Tran, Grace Kuo, and Laura Waller. Untrained networks for compressive lensless photography. Optics Express, 29(13):20913-20929, 2021. 2, 3", + "[27] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 2", + "[28] SHR Müller, B Böhm, M Gleißner, R Grzeszik, S Arndt, and A Dreizler. Flow field measurements in an optically accessible, direct-injection spray-guided internal combustion engine using high-speed piv. *Experiments in fluids*, 48:281-290, 2010. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Gergely Nagymáte and Rita M Kiss. Application of optitrack motion capture systems in human movement analysis: A systematic literature review. *Recent Innovations in Mechatronics*, 5(1):1-9, 2018. 2", + "[30] Cindy M Nguyen, Julien NP Martel, and Gordon Wetzstein. Learning spatially varying pixel exposures for motion deblurring. arXiv preprint arXiv:2204.07267, 2022. 3, 9", + "[31] Basilio Pueo Ortega and José M Jiménez Olmedo. Application of motion capture technology for sport performance analysis. Retos:rientas en Educación física, deporte y recreación, (32):241-247, 2017. 2", + "[32] Travis Portz, Li Zhang, and Hongrui Jiang. Random coded sampling for high-speed hdr video. In IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2013. 2, 3", + "[33] Ramesh Raskar, Amit Agrawal, and Jack Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 3", + "[34] Dikpal Reddy, Ashok Veeraraghavan, and Rama Chellappa. P2c2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336. IEEE, 2011. 2, 3, 9", + "[35] Aswin C, Sankaranarayanan, Pavan K Turaga, Rama Chellappa, and Richard G Baraniuk. Compressive acquisition of linear dynamical systems. SIAM Journal on Imaging Sciences, 6(4):2109-2133, 2013. 2, 3", + "[36] Fulvio Scarano. Tomographic piv: principles and practice. Measurement Science and Technology, 24(1):012001, 2012. 2", + "[37] Yoav Y Schechner, Shree K Nayar, and Peter N Belhumeur. Multiplexing for optimal lighting. IEEE Transactions on pattern analysis and machine intelligence, 29(8):1339-1354, 2007. 4", + "[38] Shubham Sharma, Shubhankar Verma, Mohit Kumar, and Lavanya Sharma. Use of motion capture in 3d animation: motion capture systems, challenges, and recent trends. In 2019 international conference on machine learning, big data, cloud and parallel computing (comitcon), pages 289-294. IEEE, 2019. 2", + "[39] Mark Sheinin, Matthew O'Toole, and Srinivasa G Narasimhan. Deconvolving diffraction for fast imaging of sparse scenes. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2021. 2, 3, 4, 5, 8, 9", + "[40] Mark Sheinin, Dinesh N Reddy, Matthew O'Toole, and Srinivasa G Narasimhan. Diffraction line imaging. In European Conference on Computer Vision, pages 1-16. Springer, 2020. 2, 3, 4, 8", + "[41] Mark Sheinin, Yoav Y Schechner, and Kiriakos N Kutulakos. Computational imaging on the electric grid. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6437-6446, 2017. 2", + "[42] H Shen, L Gan, N Newman, Y Dong, C Li, Y Huang, and YC Shen. Spinning disk for compressive imaging. Optics letters, 37(1):46-48, 2012. 2", + "[43] H Shen, N Newman, L Gan, SC Zhong, Y Huang, and YC Shen. Compressed terahertz imaging system using a spin" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10798", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "disk. In 35th International Conference on infrared, millimeter, and terahertz waves, pages 1-2. IEEE, 2010. 2", + "[44] Yaoyao Shi, Youwen Liu, Wei Sheng, Jiming Wang, and Tong Wu. Speckle rotation decorrelation based single-shot video through scattering media. Optics express, 27(10):14567-14576, 2019. 2", + "[45] Ho Shiraga, M Heya, A Fujishima, O Maegawa, K Shimada, Y Kato, T Yamanaka, and S Nakai. Laser-imploded core structure observed by using two-dimensional x-ray imaging with 10-ps temporal resolution. Review of scientific instruments, 66(1):722-724, 1995. 3", + "[46] Masada Tzabari and Yoav Y Schechner. Polarized opticalflow gyroscope. In Proceedings of the European Conference on Computer Vision, pages 363-381. Springer, 2020. 2", + "[47] BW Van Oudheusden. Piv-based pressure measurement. Measurement Science and Technology, 24(3):032001, 2013. 2", + "[48] Ashok Veeraraghavan, Dikpal Reddy, and Ramesh Raskar. Coded strobing photography: Compressive sensing of high speed periodic videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(4):671-686, 2010. 2, 3", + "[49] Zihao Wang, Leonidas Spinoulas, Kuan He, Lei Tian, Oliver Cossairt, Aggelos K Katsaggelos, and Huajin Chen. Compressive holographic video. Optics express, 25(1):250-262, 2017. 2, 3", + "[50] Gil Weinberg and Ori Katz. 100,000 frames-per-second compressive imaging with a conventional rolling-shutter camera by random point-spread-function engineering. Optics Express, 28(21):30616-30625, 2020. 2, 3, 4, 8", + "[51] Charles Wheatstone. An account of some experiments to measure the velocity of electricity and the duration of electric light. Philosophical Transactions of the Royal Society of London, (124):583-591, 1834. 2", + "[52] Jinhui Xiong, Ramzi Idoughi, Andres A Aguirre-Pablo, Abdulrahman B Aljedaani, Xiong Dun, Qiang Fu, Sigurdur T Thoroddsen, and Wolfgang Heidrich. Rainbow particle imaging velocimetry for dense 3d fluid velocity imaging. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017." + ], + "bbox": [ + 78, + 90, + 468, + 630 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10799", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/280659a7-98c7-41f5-b75f-aea246ac862a_model.json b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/280659a7-98c7-41f5-b75f-aea246ac862a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..64c7ea6e7e4065690f6f0aeb45684557fb8d6cc6 --- /dev/null +++ b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/280659a7-98c7-41f5-b75f-aea246ac862a_model.json @@ -0,0 +1,3093 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.131, + 0.836, + 0.154 + ], + "angle": 0, + "content": "SpinCam: High-Speed Imaging via a Rotating Point-Spread Function" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.182, + 0.686, + 0.198 + ], + "angle": 0, + "content": "Dorian Chan, Mark Sheinin, and Matthew O'Toole" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.2, + 0.709, + 0.217 + ], + "angle": 0, + "content": "Carnegie Mellon University, Pittsburgh, PA 15213, USA" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.22, + 0.76, + 0.234 + ], + "angle": 0, + "content": "dychan@andrew.cmu.edu, marksheinin@gmail.com, mpotoole@cmu.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.635 + ], + "angle": 0, + "content": "High-speed cameras are an indispensable tool used for the slow-motion analysis of scenes. However, the fixed bandwidth of any imaging system quickly becomes a bottleneck, resulting in a fundamental trade-off between the camera's spatial and temporal resolutions. In recent years, compressive high-speed imaging systems have been proposed to circumvent these issues by optically encoding the signal and using a reconstruction procedure to recover a video. Our work proposes a novel approach for compressive high-speed imaging based on temporally coding the camera's point-spread function (PSF). By mechanically spinning a diffraction grating in front of a camera, the sensor integrates an image blurred by a PSF that continuously rotates over time. We also propose a deconvolution-based reconstruction algorithm to reconstruct videos from these measurements. Our method achieves superior light efficiency and handles a wider scene class than prior methods. Also, our mechanical design yields flexible temporal resolution that can be easily increased, potentially allowing capture at \\(192\\mathrm{kHz}\\) far higher than prior works. We demonstrate a prototype for various applications, including motion capture and particle image velocimetry (PIV)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.21, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The world is full of high-speed events that are imperceptible to both the human eye and most conventional cameras (e.g., the flicker of electric lights or the motion of fast-moving objects). The traditional approach to capturing such events is using slow-motion photography, i.e., use a camera to shoot images at a high frame rate and display the results at a lower frame rate. Slow-motion photography is often used for artistic effects in film and as a measurement tool in scientific applications. However, despite significant advancements in sensor technology, conventional high-speed cameras tend to be expensive, bulky, and have high throughput requirements for recording visual content at both high frame rates and high spatial resolutions. These hardware limitations set an upper limit to the cameras' capture rates." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.267, + 0.871, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.641, + 0.895, + 0.74 + ], + "angle": 0, + "content": "Figure 1. Principle of high-speed imaging with a rotating diffraction grating. A camera captures a long exposure image of a moving scene. During the camera's exposure period, the diffraction grating placed in front of the camera continuously rotates to generate a time-varying point-spread function (PSF) that encodes scene dynamics into a single frame. Solving a deconvolution problem yields a high-speed video from the captured frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Compressed sensing provides a framework to circumvent such issues. Since videos are highly compressible, it is possible to recover a video from a single snapshot by (i) compressively recording spatiotemporal visual information onto a sensor and (ii) solving a reconstruction problem to recover the video using priors on the visual content. Among the large body of work in compressive high-speed imaging, most techniques temporally modulate the light incident on sensor pixels [48, 12, 20, 22, 9, 34, 32, 35, 14, 16, 19," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10789" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "13, 49], either through the use of a spatial light modulator (SLM) or novel sensors capable of turning pixels on or off. Alternatively, by using spatial-multiplexing optics (e.g., diffusers, diffraction gratings) to distribute light from any scene point to many camera pixels, it is possible to perform high-speed imaging by using a rolling shutter [2, 50, 26], a reduced region-of-interest (ROI) [39], or a line sensor [40]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.468, + 0.381 + ], + "angle": 0, + "content": "Our work proposes a fundamentally different approach to compressive high-speed imaging. Similar to past works that make use of spatial-multiplexing optics, we propose to distribute light across a sensor by using a dual-axis diffraction grating, which diffracts light horizontally and vertically (see Fig. 1). Unlike prior work, however, we opt to change the point-spread function (PSF) during the camera's exposure period. Specifically, we use a motor to mechanically rotate the PSF at high speeds, record a single global-shutter frame, and solve a simple deconvolution-based reconstruction procedure to reconstruct scene dynamics. Please refer to Fig. 1 for an overview of the process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.468, + 0.579 + ], + "angle": 0, + "content": "Our approach, like past work using spatial-multiplexing optics [2, 50, 40, 39], is primarily targeted towards spatiotemporally sparse scenes. This imaging regime includes motion capture and particle-image velocimetry (PIV), which see broad application in domains like biomechanics [29], sports science [31], animation and film-making [38], virtual and augmented reality [5], aerodynamics [47], fluid mechanics [36], biomedical engineering [18], oceanography [7], and combustion and engine research [28]. We demonstrate our approach for motion capture and PIV in Sec. 6. Other potential applications for spatiotemporally sparse video include high-speed 3D scanning [24] and capturing the fast flicker of electric bulbs [41]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.58, + 0.468, + 0.611 + ], + "angle": 0, + "content": "Our approach combines four key advantages over existing methods targeting spatiotemporally sparse scenes:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.468, + 0.755 + ], + "angle": 0, + "content": "(I) Superior light efficiency. In many prior methods, the scene light is dispersed using specialized optics to reach many camera pixels, but only a small subset of camera pixels collect light at each timestamp [2, 40, 39, 50]. Thus, most light is lost by falling outside the 'active' pixels. Conversely, in our work, the signal from different timestamps is encoded by different temporal PSFs, and the entire sensor is left exposed, yielding far superior light efficiency. We show this advantage in simulated experiments in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "(II) Easily increasable temporal resolution. In most prior works, temporal resolution (i.e., FPS) is fundamentally limited by the speed of the respective technology (e.g., the modulation speed of a SLM/sensor [12, 22, 34, 32, 14, 49], or the read-out speed of a camera [2, 40, 39, 50, 9]). In contrast, the temporal resolution of our approach is determined by the mechanical rotation speed and thus can be easily increased by rotating the PSF faster. Thus, we show that our approach can resolve a LED flickering at \\(96\\mathrm{kHz}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "(III) A wider class of supported scenes. Prior works using spatial-multiplexing optics allocate a fixed amount of pixels for each timestamp (e.g., a single rolling shutter row) [2, 40, 39, 50]. This limits the amount of information recoverable per frame, imposing a spatial sparsity constraint on each high-speed frame individually. Instead, our solution multiplexes the signals from different timestamps together, yielding a sparsity constraint over the entire video. Our method can thus handle temporally sparse but spatially dense scenes, verified in simulated and real experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.242, + 0.892, + 0.348 + ], + "angle": 0, + "content": "(IV) Simple hardware. While many prior works require specialized optics (e.g., SLMs [12, 34, 49], piezoelectric stages [20, 16, 19]) or specialized sensors (e.g., coded exposure pixels [22, 14, 32]), our setup can work with any global-shutter camera combined with simple low-cost hardware that is readily available at your local hobby store. We include a DIY guide as part of our supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.355, + 0.637, + 0.37 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.379, + 0.817, + 0.396 + ], + "angle": 0, + "content": "2.1. Spinning Disks & Revolving Mirrors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Mechanical spinning disks and revolving mirrors have been used to measure transient phenomena for centuries. In 1834, Wheatstone [51] estimated the velocity of electricity traveling through long copper wires by observing the sparks generated at the ends. Using a revolving mirror, Wheatstone observed a \\(1\\mu \\mathrm{s}\\) time difference, and concluded the velocity of electricity to be 288,000 miles per second. Fizeau and Foucault [25] later measured the speed of light through various apparatuses based on revolving mirrors and spinning cogwheels. For example, in 1848-49, Fizeau measured the transit time of light from a source, to a mirror, and back to the source by using a rotating wheel to modulate the light." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Though rare in modern consumer electronics, modulating light mechanically with a spinning element is both faster and lower cost than most modern digital modulation solutions. For example, digital micromirror devices (DMDs) only operate up to \\(30\\mathrm{kHz}\\), whereas spinning disks can offer rates up to \\(2.4\\mathrm{MHz}\\) [10]. As a result, coded disks have been used for single pixel imaging at rates up to 72 FPS [43, 42, 10], and revolving coded mirrors have been used to reconstruct 1400 frames from a single shot [23]. A spinning camera was used to recover an 11-frame video through scattering media by disambiguating the resulting rotated speckle patterns [44], while a rotating camera coupled with a polarizer generalized the optical-flow equation to include polarization cues [46]. An event camera can be combined with a spinning polarizer for fast, accurate shape-from-polarization [27]. This work explores the use of spinning elements for compressive high-speed video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.744, + 0.864 + ], + "angle": 0, + "content": "2.2. Pixel-wise Coded Exposure" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Coded exposure photography involves temporally modulating the light incident on a sensor, which has long been" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10790" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.183 + ], + "angle": 0, + "content": "used to capture dynamic events [4]. Raskar et al. [33] used a liquid-crystal panel to rapidly open and close the shutter of a camera; the motion blur associated with fast-moving objects could then be reliably removed through deconvolution. This methodology was later extended for recovering full high-speed videos of periodic [48] and generic scenes [13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.468, + 0.318 + ], + "angle": 0, + "content": "A popular approach to compressive high-speed imaging involves pixel-wise exposure control [12, 9, 34, 32, 35, 22, 14, 20, 16, 19]. In practice, the codes are often binary, and the corresponding reconstruction techniques rely on various priors to reconstruct videos. Example priors include assuming the scene produces periodic motion [48] or satisfies the brightness constancy assumption [34], and might rely on dictionary learning [12, 13] and neural networks [22, 14, 30] to aid the reconstruction procedure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.468, + 0.471 + ], + "angle": 0, + "content": "These techniques lack the (II) temporal resolution and (IV) hardware complexity advantages of our approach since they typically require specialized hardware (e.g., SLMs [12, 34, 49], piezoelectric stages [20, 16, 19]) or specialized sensors [22, 14, 32]) which can be expensive, difficult to acquire, and still be limited in frame rate. Moreover, as shown in the supplementary materials, since these methods do not spread light across multiple pixels, these systems struggle with spatially sparse scenes, which appear in many practical applications of high-speed imaging [40, 39]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.327, + 0.499 + ], + "angle": 0, + "content": "2.3. Spatial-Multiplexing Optics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.468, + 0.704 + ], + "angle": 0, + "content": "An alternative approach to capture space-time volumes (videos) involves spatial multiplexing. By placing an optical element in front of a sensor or camera (e.g., a diffuser or diffraction grating), the light from a single point in the scene spreads to multiple camera pixels according to the optical system's PSF. When combined with a rolling shutter camera that samples individual sensor rows at fast rates [2, 50, 26], it is possible to reconstruct high-speed videos. Alternatively, a limited region-of-interest (ROI) or 1D sensor can also be used to sample measurements at fast rates [40, 39]. By integrating diffraction patterns generated by time-varying coded masks, it is even possible to reconstruct 4D volumes of spacetime from a single snapshot [49]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.87 + ], + "angle": 0, + "content": "Similar ideas have been used in combination with streak cameras, an optoelectronic device used to measure ultra-fast light phenomena at femtosecond time scales. A streak camera spatially multiplexes the scene's temporal response, by deflecting photoelectrons to different points on a sensor as a function of time. While designed to be a line sensor, it is possible to encode 2D spatial information by incorporating a tilted microlens [45] or pinhole array [11]. In Gao et al. [8], the 2D scene at different timestamps is coded with a fixed mask, sheared a temporally varying amount, and then overlayed onto the 2D signal captured by a streak camera." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "These methods lack the (I) light efficiency and (III) density advantages of our approach—they typically rely" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.089, + 0.835, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.213, + 0.76, + 0.226 + ], + "angle": 0, + "content": "(a) Captured image" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.226, + 0.848, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.289, + 0.798, + 0.303 + ], + "angle": 0, + "content": "(b) Recovered time signal (ms)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Figure 2. High-speed reconstruction of a flickering LED. The LED in this experiment flashes at \\(96\\mathrm{kHz}\\), as driven by an Arduino. (a) The diffraction image encodes the high-frequency patterns of the LED (zoom into inset). (b) This temporally-varying signal can be directly extracted from the image to recover the intensity pattern of the LED. A frequency analysis of the recovered time signal indicates the system's ability to resolve frequencies up to \\(96\\mathrm{kHz}\\). Given that the Nyquist frequency is half that of the sample rate, the effective framerate for the system is therefore over 192,000 FPS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.541 + ], + "angle": 0, + "content": "on a small subset of camera pixels to reconstruct each frame, \"wasting\" light and limiting recoverable scene density. Furthermore, their capture speed is limited by the maximum speed of the hardware used (i.e., line-scan mode [40], ROIs [39] or rolling shutter [2, 50]). Thus, these methods lack the ability to (II) easily increase temporal resolution." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.553, + 0.892, + 0.588 + ], + "angle": 0, + "content": "3. High-Speed Imaging using a Time-Varying Point-Spread Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.891, + 0.674 + ], + "angle": 0, + "content": "Consider a scene that consists of a single bright LED in a dark room, where the LED is configured to pulse rapidly over time. Unless it is faster than the blinking LED, a normal camera forms an image of the LED, but provides no information about the LED's temporal dynamics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Our proposed approach to imaging high-speed events involves two steps. First, following prior work on diffraction imaging [40, 39], we place a diffraction grating in front of the camera, which blurs the scene with a rainbow PSF. The light from the LED is therefore spread over many additional camera pixels. Second, we propose to rotate the diffraction grating such that the corresponding PSF changes over time. This produces measurements that spatially encode temporally varying signals, which can be used to analyze extremely high-speed events (see Fig. 2 for an example)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we describe the image formation model and reconstruction process for recovering a high-speed video from an image captured with a generic time-varying PSF. We then analyze the special case of using a rotating PSF for capturing images." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10791" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.218, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.159, + 0.214, + 0.178 + ], + "angle": 0, + "content": "Rolling \\(^+\\) diffuser [2] (64 row exposure)" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.089, + 0.336, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.159, + 0.328, + 0.178 + ], + "angle": 0, + "content": "ROI + grating [39] (16 row ROI)" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.089, + 0.455, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.159, + 0.432, + 0.17 + ], + "angle": 0, + "content": "Rotating PSF" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.179, + 0.455, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.257, + 0.43, + 0.27 + ], + "angle": 0, + "content": "(a) Recovered brightness of time-varying light source" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.281, + 0.217, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.352, + 0.217, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.42, + 0.215, + 0.441 + ], + "angle": 0, + "content": "Rolling \\(^+\\) diffuser [2] (64 row exposure)" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.282, + 0.336, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.352, + 0.337, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.42, + 0.327, + 0.44 + ], + "angle": 0, + "content": "ROI + grating [39] (16 row ROI)" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.282, + 0.455, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.352, + 0.456, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.42, + 0.432, + 0.431 + ], + "angle": 0, + "content": "Rotating PSF" + }, + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.443, + 0.381, + 0.456 + ], + "angle": 0, + "content": "(b) Recovery of dense moving scene" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.601 + ], + "angle": 0, + "content": "Figure 3. Light efficiency experiment. In the presence of Poisson and read noise, we simulated the measurements captured by a Canon 5D Mark III sensor [6] with ISO 800. (a) We simulate a dim point light source with a time-varying intensity profile. (b) We simulate the word \"test\" moving across the scene. In both cases, our approach is able to much more accurately resolve the signal, compared to a rolling shutter and diffuser as well as a ROI and grating. We theorize that this is due to the fact that our approach captures more light for every time step, and that the requisite demultiplexing minimizes the effects of noise." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.63, + 0.301, + 0.646 + ], + "angle": 0, + "content": "3.1. Image Formation Model" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.655, + 0.47, + 0.686 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, for a camera where the PSF changes with time, the forward model can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.703, + 0.469, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} (x, y) = \\int_ {t = t _ {s}} ^ {t = t _ {s} + E} (\\mathbf {k} (t) * \\mathbf {s} (t)) (x, y) d t, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.469, + 0.844 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}(x,y)\\) is the captured image, \\(E\\) denotes the exposure time of the camera, and \\(t_s\\) denotes the time when the camera started exposing. The symbol \\(*\\) denotes the spatial convolution operator between the scene \\(\\mathbf{s}(t)\\) and the PSF \\(\\mathbf{k}(t)\\) at time \\(t\\). When discretized to \\(N_E\\) time steps from \\(t_s\\) to \\(t_s + E\\), this can be rewritten in matrix form as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.861, + 0.469, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} = \\mathbf {M} \\mathbf {s} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.879, + 0.469, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\left[ \\begin{array}{l l l l} \\mathbf {C} _ {1} & \\mathbf {C} _ {2} & \\dots & \\mathbf {C} _ {N _ {E}} \\end{array} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.089, + 0.892, + 0.16 + ], + "angle": 0, + "content": "where \\(\\mathbf{I} \\in \\mathbb{R}_{+}^{N_{x}N_{y}}\\) is the vectorized representation of the captured image, \\(\\mathbf{s} \\in \\mathbb{R}_{+}^{N_{x}N_{y}N_{E}}\\) is the vectorized representation of the high-speed scene, and \\(\\mathbf{C}_k \\in \\mathbb{R}_{+}^{N_xN_y\\times N_xN_y}\\) represents the convolution with the kernel at timestep \\(k\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.168, + 0.715, + 0.183 + ], + "angle": 0, + "content": "3.2. Reconstruction Process" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.191, + 0.892, + 0.267 + ], + "angle": 0, + "content": "Natural scenes contain sparse gradients, and many target applications for high-speed imaging require the tracking of just sparse points. As a result, we recover a high-speed video with an \\(\\ell_1\\) sparsity prior and an anisotropic total variation prior, solving the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.275, + 0.892, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} \\quad \\| \\mathbf {I} - \\mathbf {M} \\mathbf {x} \\| ^ {2} + \\mathcal {I} _ {\\mathbb {R} _ {+}} (\\mathbf {x}) + \\lambda_ {\\text {s p a r s i t y}} \\| \\mathbf {x} \\| _ {1} + \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.305, + 0.858, + 0.321 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {d x} \\| \\mathbf {D} _ {x} \\mathbf {x} \\| _ {1} + \\lambda_ {d y} \\| \\mathbf {D} _ {y} \\mathbf {x} \\| _ {1} + \\lambda_ {d t} \\| \\mathbf {D} _ {t} \\mathbf {x} \\| _ {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.329, + 0.892, + 0.391 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_{\\mathbb{R}_+}\\) denotes a non-negativity prior, \\(\\mathbf{D}_x\\), \\(\\mathbf{D}_y\\), and \\(\\mathbf{D}_t\\) compute forward finite differences along the \\(x\\), \\(y\\), and \\(t\\) dimensions respectively of the high-speed video \\(\\mathbf{x}\\), and \\(\\lambda_{sparsity}\\), \\(\\lambda_{dx}\\), \\(\\lambda_{dy}\\), and \\(\\lambda_{dt}\\) weigh the regularization terms." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.399, + 0.602, + 0.415 + ], + "angle": 0, + "content": "3.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.422, + 0.892, + 0.467 + ], + "angle": 0, + "content": "Varying the PSF over time results in fundamentally different properties compared to prior approaches, which we explore in the rest of this section." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.893, + 0.79 + ], + "angle": 0, + "content": "Higher light efficiency. Our method, like other prior compressive video works [2, 50, 40, 39], relies on PSFs with broad spatial supports. A broad spatial support spreads the signal from a single point across multiple pixels. Recent approaches record just a subset of these pixels at each timestep [2, 50, 40, 39]. Because fewer pixels encode each frame and light is lost, assuming a standard noise model with Poisson shot noise and additive Gaussian noise, the reconstruction's signal-to-noise ratio is proportionately decreased. In contrast, under a time-varying PSF, the entire frame is always recorded, so a much larger fraction of the PSF (and therefore light) is captured for every scene point during capture. In addition, the signals from different timestamps are mixed together in the final capture, which could improve the quality of the reconstructions under read noise [37]. Thus, this overall increase in light efficiency leads to better performance in resolving noisy scenes. We show this experimentally in Fig. 3 on two simulated cases: a point light source with time-varying intensity, and a denser moving scene. We find that our approach robustly recovers these two scenes, even under low light conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A wider class of scenes. Many past approaches achieve high-speeds by allocating a number of pixels per frames that is much lower than the total number of frame pixels [2, 50, 40, 39]. This causes each frame to undergo lossy compression, yielding a sparsity constraint per frame. For example, for a rolling shutter with exposure set to the row rate, each frame is encoded onto just a single sensor row." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10792" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.088, + 0.223, + 0.102 + ], + "angle": 0, + "content": "(a) Simulated scene:" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.102, + 0.202, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.332, + 0.195, + 0.342 + ], + "angle": 0, + "content": "(1) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.102, + 0.355, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.179, + 0.355, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.255, + 0.355, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.332, + 0.339, + 0.353 + ], + "angle": 0, + "content": "(2) Rolling + diffuser (1 row exposure)" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.102, + 0.486, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.179, + 0.487, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.255, + 0.487, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.332, + 0.487, + 0.353 + ], + "angle": 0, + "content": "Rolling \\(^+\\) diffuser [2] (64 row exposure)" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.102, + 0.617, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.179, + 0.618, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.255, + 0.618, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.332, + 0.612, + 0.353 + ], + "angle": 0, + "content": "(4) ROI + grating [39] (2 row ROI)" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.102, + 0.75, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.179, + 0.75, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.255, + 0.75, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.332, + 0.745, + 0.353 + ], + "angle": 0, + "content": "(5) ROI + grating [39] (16 row ROI)" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.102, + 0.883, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.179, + 0.883, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.255, + 0.883, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.332, + 0.862, + 0.343 + ], + "angle": 0, + "content": "(6) Rotating PSF" + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.354, + 0.184, + 0.367 + ], + "angle": 0, + "content": "(b) Real scene:" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.367, + 0.281, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.48, + 0.248, + 0.492 + ], + "angle": 0, + "content": "(1) Reference frame" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.367, + 0.483, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.48, + 0.475, + 0.492 + ], + "angle": 0, + "content": "(2) Rotation-encoded frame" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.367, + 0.685, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.48, + 0.756, + 0.492 + ], + "angle": 0, + "content": "(3) Recovered frames" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.499, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Figure 4. Reconstructing temporally sparse, but spatially dense scenes. (a) A simulated scene of two people dancing. The video is dark for the entire sequence, except for a short sequence of fast-moving frames where the dancers are strobed. (a2)-(a3) We implement a setup based on a dual rolling shutter combined with a diffuser [2]. In (a2), with the exposure time set to match the row rate of the shutter, the reconstruction is unable to well-resolve the dense scene. In (a3), with the exposure time set to 64 times the previous value, the reconstruction is able to recover the dancers, at the cost of temporal resolution [2]. (a4)-(a5) We implement an approach based on fast temporal sampling by using a ROI and a diffraction grating [39]. In (a4), the ROI consists of just two rows, resulting in low quality results. In (a5), by using more bandwidth and exposing 16 rows, such an approach can get better results, but this reduces the recoverable framerate by 8 times in practice. (a6) In contrast, our methodology recovers the dense scene much more accurately compared to (a2) and (a4)—the signal is encoded over a much larger portion of sensor. This does not require increasing the exposure time and therefore decreasing the temporal resolution as in (a3), or increasing the requisite bandwidth (a5). (b) A real world scene, where a spinning fan is strobed twice during the camera exposure at unknown times. Our approach is able to disambiguate the fan rotation at each strobe time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.469, + 0.812 + ], + "angle": 0, + "content": "In contrast, our method allocates the entire frame to encode a plurality of video frames belonging to a fixed time interval. Therefore, our method assumes sparsity in a more general sense, namely sparsity over the entire fixed time interval. The encoded fixed time interval can be either spatially sparse and temporally dense, which fits most prior-work assumptions, or temporally sparse but spatially dense, which prior approaches do not consider due to the fixed bitrate per frame imposed by their low-dimensional encoding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To illustrate how this is an advantage, consider the case of a dense scene that is entirely dark, except for a single time instant with a flash of light. If sampling the scene using a rolling-shutter or line sensor, the flashed frame will be encoded onto just a single row. In contrast, a time-varying" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.891, + 0.721 + ], + "angle": 0, + "content": "PSF encodes the flashed frame on the entire frame, wasting no resources encoding the dark frames. Therefore, in this case, a time-varying PSF enables a better reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.891, + 0.813 + ], + "angle": 0, + "content": "As shown in Fig. 4, handling temporally sparse, but spatially dense scenes makes our approach very useful under strobed lighting. As shown in real and simulated experiments, a rotating PSF is able to resolve the content within each individual strobe at high spatiotemporal resolution, without any constraints on scene structure between strobes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.704, + 0.846 + ], + "angle": 0, + "content": "4. Using a Rotating PSF" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Sec. 3 described the encoding and reconstruction of high-speed video using a time-varying PSF. While a time-varying PSF can be implemented in a number of ways," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10793" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.468, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.368 + ], + "angle": 0, + "content": "Figure 5. The PSF and spatial frequency coding of time. A rotating PSF effectively codes the signal from different timesteps into a rotating set of spatial frequencies in the captured image. For instance, a rotating line PSF encodes different timestamps into different lines in Fourier space. The star PSF from a double-axis diffraction grating behaves as a mixture of the point and line cases. Its Fourier transform contains bright line components that provide temporal signal, but remains broadband for spatial resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.47, + 0.517 + ], + "angle": 0, + "content": "for the sake of cost and speed, we simply rotate a PSF-generating optical element over the course of the camera exposure. This results in a rotation of the corresponding PSF. In this particular case, \\(\\mathbf{k}(t)\\) can now be rewritten as \\(\\mathbf{k}(t) = \\mathbf{R}(\\mathbf{k},\\theta_t)\\), where \\(\\mathbf{k}\\) is the stationary PSF of the rotating optical element, \\(\\theta_{t}\\) is the angle of the element at time \\(t\\), and \\(\\mathbf{R}(\\mathbf{k},\\theta_t)\\) denotes a rotation of \\(\\mathbf{k}\\) by \\(\\theta_{t}\\). In the rest of this section, we analyze the properties of such a rotating setup." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.689 + ], + "angle": 0, + "content": "PSF performance. The selected PSF affects the spatiotemporal reconstruction quality of every video frame. Because each time step is convolved with a rotated PSF, the PSF's rotated Fourier transform determines which parts of the spectrum will encode each time step. To illustrate, an ideal thin lens having a perfect impulse PSF will have perfect spatial resolution because its Fourier transform contributes values in all spectrum frequencies (Fig. 5(a)). It will, however, have poor temporal resolution because every other timestamp will also contribute values across the entire spectrum, making them impossible to distinguish." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.842 + ], + "angle": 0, + "content": "In contrast, consider a line PSF (Fig. 5(b)). In the Fourier domain, the PSF remains a line that goes through the origin. When the line PSF rotates, its Fourier spectrum rotates correspondingly, such that each time step is encoded into a different line in Fourier space, without interfering with values from previous timestamps. As a result, the signal at each time step can be easily reconstructed simply by extracting its corresponding line from Fourier space. However, because only a single direction of spatial frequencies is available for reconstruction for every time step, the re" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.702, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.258, + 0.635, + 0.272 + ], + "angle": 0, + "content": "(a) Setup" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.107, + 0.9, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.258, + 0.86, + 0.271 + ], + "angle": 0, + "content": "(b) Captured PSF" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.278, + 0.892, + 0.319 + ], + "angle": 0, + "content": "Figure 6. Image of our hardware setup, consisting of a global-shutter camera and a circular diffraction grating attached to a 20,000 RPM motor." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.346, + 0.878, + 0.361 + ], + "angle": 0, + "content": "sulting recovered frame will have poor spatial resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Inspired by these two cases, we use a double-axis diffraction grating for the experiments shown in this work, whose corresponding PSF is shown in Fig. 6(b). This star PSF contains a bright DC spot at the center, intuitively providing spatial resolution. Simultaneously, the PSF also contains line-shaped rainbow lobes that provide temporal information. We visualize this phenomena in Fig. 5(c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.893, + 0.63 + ], + "angle": 0, + "content": "Temporal resolution. The temporal resolution of our system is fundamentally determined by the amount of temporal variation of the PSF between different timestamps, or namely, the minimum angle required to reliably distinguish between two different rotations of the PSF. As a result, the temporal resolution can be increased either by increasing the pixel resolution, the spatial spread of the PSF, or the rotation speed. But since our grating PSF is roughly \\(90^{\\circ}\\) periodic, rotating more than \\(90^{\\circ}\\) during the exposure causes temporal ambiguity and degrades the reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.892, + 0.692 + ], + "angle": 0, + "content": "In the supplement, we further analyze PSF performance from the mutual coherence perspective. We also discuss how a time-varying PSF behaves as a form of coded exposure in frequency, as shown by the above Fourier analysis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.709, + 0.659, + 0.725 + ], + "angle": 0, + "content": "5. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Hardware. An image of our hardware setup is given in Fig. 6. To generate a usable PSF, we cut a 531-lines per millimeter double-axis diffraction grating sheet into a circular disk of radius 6 inches. To ensure that the sheet stayed flat, we taped it onto a laser-cut acrylic disk with a radius of 6 inches. We attached this element to a Puly DC 775 Motor with a shaft coupling to rotate it. We place this entire rotating setup \\(5\\mathrm{mm}\\) in front of the camera's lens. We used a UI-3070CP-C-HQ R2 camera, fitted with a \\(9\\mathrm{mm}f / 1.4\\) lens (Fujinon HF9HA). We also placed a second identical camera beside our system to capture intermittent reference" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.47, + 0.901 + ], + "angle": 0, + "content": "1The Fourier transform of the convolution of two functions is the element-wise product of their Fourier transforms. If a function is rotated by \\(\\theta\\), its Fourier transform is the Fourier transform of the original function, rotated by \\(\\theta\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10794" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.287, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.167, + 0.286, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.245, + 0.285, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.341, + 0.285, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.436, + 0.285, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.532, + 0.27, + 0.545 + ], + "angle": 0, + "content": "(a) Reference start frame" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.089, + 0.483, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.167, + 0.482, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.244, + 0.482, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.341, + 0.482, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.436, + 0.482, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.532, + 0.466, + 0.544 + ], + "angle": 0, + "content": "(b) Reference end frame" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.089, + 0.681, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.167, + 0.681, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.244, + 0.681, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.341, + 0.681, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.436, + 0.681, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.532, + 0.674, + 0.545 + ], + "angle": 0, + "content": "(c) Rotation-encoded frame" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.09, + 0.878, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.167, + 0.878, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.244, + 0.878, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.341, + 0.878, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.436, + 0.878, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.532, + 0.851, + 0.545 + ], + "angle": 0, + "content": "(d) Recovered frames" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.55, + 0.894, + 0.593 + ], + "angle": 0, + "content": "Figure 7. High-speed reconstruction of fast moving objects. Row 1: A ball with a black stripe bounces and spins. Row 2: An I rotates \\(120^{\\circ}\\). Row 3: A toy bullet bouncing off a board. Row 4, 5: Our setup is capable of tracking the movements of both sparse points and dense marker lines for high-speed motion capture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.648 + ], + "angle": 0, + "content": "frames for visualization. Please refer to the supplement for a DIY guide." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Software. To solve Eq. (4), we used a PyTorch implementation of FISTA [3], combined with an approximate proximal operator [15] to handle the anisotropic total variation prior. We ran 10000 iterations of FISTA for each result shown. For most results, we set \\(\\lambda_{sparsity} = 0\\), \\(\\lambda_{dx} = \\lambda_{dy} = 1 \\times 10^{-3}\\), and \\(\\lambda_{dt} = 1 \\times 10^{-3}\\), but we vary the strength of these terms depending on the brightness of the background and the speed of the motion. We list the parameters for each result and provide sample code in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Calibration and capture. To calibrate the time-varying PSF of our imaging system, we process a stationary image of a small retroreflector or LED, and then computationally rotate this image to create the dictionary required for Eq. (4)—please see the supplemental for more details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "For every capture, we need to identify the range of angles that are rotated over during the camera exposure in order to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.892, + 0.799 + ], + "angle": 0, + "content": "solve Eq. (4). This could be done by placing an LED at a known location, and reconstructing which PSFs are present around it. Alternatively, a controllable motor or trigger system could be used to directly provide the requisite angles. In our work, we opt for a unstructured approach, that simply relies on signal being present at the start and end of the camera exposure. Before capture, we first calibrate the motor to rotate roughly \\(90^{\\circ}\\) over the course of the exposure. Then, we simply perform a coarse reconstruction of Eq. (4) using the captured image and a dictionary that spans all \\(360^{\\circ}\\), and select the best matching \\(90^{\\circ}\\) for a finer reconstruction. Please see the supplemental for an in-depth discussion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.587, + 0.83 + ], + "angle": 0, + "content": "6. Results" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "For our real-world reconstructions, we downsampled the input images to \\(300 \\times 400\\) pixels and reconstructed \\(N_{E} = 146\\) frames for every input. For the simulated results (Fig. 3, Fig. 4), we instead used \\(N_{E} = 150\\) to better match" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10795" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.089, + 0.285, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.197, + 0.277, + 0.21 + ], + "angle": 0, + "content": "(a) Rotation-encoded frame" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.089, + 0.483, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.197, + 0.482, + 0.21 + ], + "angle": 0, + "content": "(b) Example recovered frame" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.089, + 0.683, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.198, + 0.656, + 0.21 + ], + "angle": 0, + "content": "(c) Flow visualization" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.095, + 0.874, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.196, + 0.845, + 0.21 + ], + "angle": 0, + "content": "(d) Recovered flow" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.216, + 0.894, + 0.258 + ], + "angle": 0, + "content": "Figure 8. High-speed particle image velocimetry. Following Sheinin et al. [39], a magnetic stirrer is placed in a tank seeded with tracer particles. (a) Captured diffraction image. (b) Recovered frame using our reconstruction procedure. (c) We visualize a \\(17.33\\mathrm{ms}\\) interval by overlaying the recovered frames. (d) Using OpenPIV [17], we recover the flow between the start and end frame of (c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.468, + 0.357 + ], + "angle": 0, + "content": "a dual rolling shutter setup. In this document, we use color to code a video into a single image; we refer the reader to our supplemental materials for directly viewing the recovered videos. The camera exposure was \\(22.178\\mathrm{ms}\\), yielding an effective framerate of 6583.10 FPS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.468, + 0.478 + ], + "angle": 0, + "content": "We demonstrate our methodology on fast scenes on multiple examples (Fig. 7)—a bouncing ball, an \"I\" spinning on a fan, toy bullets bouncing off a board, and letters spinning and falling. In the last row of Fig. 7, we applied our technique to marker-based motion capture. Our technique readily captures a wide variety of marker shapes. In Fig. 4(b), we show a temporally sparse, but spatially dense scene of a fan under strobed illumination." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.468, + 0.616 + ], + "angle": 0, + "content": "As shown by Sheinin et al. [39], low-cost, high-speed imaging could help measure the movement of fluids in particle image velocimetry [1, 21, 52]. In short, small tracer particles are placed into the medium, which can then be tracked to visualize the flow. We apply our setup to this task in Fig. 8, imaging a tank filled with water and a magnetic stirrer. To maximize the visible movement, we increased the exposure time of our camera to \\(100\\mathrm{ms}\\) and proportionately decreased the speed of the motor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.468, + 0.723 + ], + "angle": 0, + "content": "Finally, we push the temporal resolution of our approach by imaging an LED pulsed at \\(96\\mathrm{kHz}\\). For this particular result, we used a \\(16\\mathrm{mm}\\) focal length lens (Edmund Optics #86-571) to increase the scale of the PSF, and rotated the motor at roughly 15,000 RPM. At resolution \\(2048\\times 1542\\) pixels, our methodology is capable of easily resolving this signal as shown in Fig. 2, demonstrating \\(192\\mathrm{kHz}\\) capture." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.739, + 0.2, + 0.755 + ], + "angle": 0, + "content": "7. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Temporal unmixing. Rolling shutter or ROI-based approaches [2, 50, 40, 39] code the signal from each timestamp onto a unique set of measurements that only record the signal at that time. In contrast, with a rotating PSF, the signal from different timestamps is mixed together and requires unmixing using Eq. (4). This unmixing can yield degraded results when there are significant nonidealities, e.g., a poorly calibrated PSF, major spatial variation of the PSF, sizable background signal, or when the scene is too dense." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.28, + 0.631, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.321, + 0.566, + 0.331 + ], + "angle": 0, + "content": "PSF spect." + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.281, + 0.63, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.321, + 0.628, + 0.33 + ], + "angle": 0, + "content": "Rec. Mx" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.281, + 0.698, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.321, + 0.691, + 0.33 + ], + "angle": 0, + "content": "PSF spect" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.28, + 0.758, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.321, + 0.756, + 0.33 + ], + "angle": 0, + "content": "Rec. \\(\\mathbf{M}\\mathbf{x}\\)" + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.28, + 0.884, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.321, + 0.883, + 0.33 + ], + "angle": 0, + "content": "Rec. Mx" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.33, + 0.63, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.373, + 0.628, + 0.384 + ], + "angle": 0, + "content": "Sample recons. frame" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.33, + 0.758, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.373, + 0.754, + 0.384 + ], + "angle": 0, + "content": "Sample recons. frame" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.33, + 0.885, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.373, + 0.884, + 0.384 + ], + "angle": 0, + "content": "Sample recons. frame" + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.389, + 0.604, + 0.401 + ], + "angle": 0, + "content": "Original PSF" + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.388, + 0.754, + 0.4 + ], + "angle": 0, + "content": "Bandpass filtered PSF" + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.388, + 0.872, + 0.4 + ], + "angle": 0, + "content": "Notch filtered PSF" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.406, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Figure 9. Reconstruction sensitivity to PSF spectrum. We reconstruct Fig. 7, Row 2 with digitally augmented versions of the original real measured PSF, to simulate a PSF calibrated under incorrect lighting or from a different material. The reconstruction quality is fairly robust to these PSF variations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.494, + 0.892, + 0.539 + ], + "angle": 0, + "content": "In these cases, the signal from one timestamp may bleed into the reconstructions for other timestamps, causing a loss of temporal resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Saturation. Our method can be especially prone to sensor saturation for several reasons. First, the use of a longer camera exposure may saturate those camera pixels which received scene light over a significant portion of the exposure duration. Secondly, a bright flash of light at a single timestamp could saturate the entire sensor, corrupting the whole video sequence, while only a single row (and therefore timestamp) would be saturated for a rolling shutter or ROI. Thirdly, our grating's PSF contains a 'DC spot' that is much brighter than the rainbow streaks. Therefore, well-exposed rainbow streaks yield a large DC component that might saturate camera pixels. This latter problem could be mitigated by using a higher-quality grating." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Spectrum. The PSF of an object, when viewed through a diffraction grating, depends on its reflectance spectrum, which itself depends on the light source spectrum and object material. Our approach assumes that scenes consist of a single reflectance spectrum so that this PSF is known and fixed for every point in the scene. This assumption makes our approach well-suited for vision applications like motion capture or PIV, which use specialized light sources and markers. In these cases, only slight deviations in PSF occur, which our method can handle (see Fig. 9). However, a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10796" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "more general scene may consist of many diverse materials and light sources. This results in many different reflectance spectrums and corresponding PSFs at the sensor, restricting the applicability of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.318 + ], + "angle": 0, + "content": "To address this limitation, the single reflectance spectrum assumption can potentially be relaxed, albeit with tradeoffs. For instance, the reconstruction problem in Eq. (4) can be expanded to explicitly account for these multiple PSFs following Sheinin et al. [39], at the cost of extra computational complexity that scales with the number of possible spectrums. Alternatively, a less wavelength-dependent optic (e.g., a simple diffuser [2]) could be used, such that the PSF does not significantly vary for different wavelengths. However, this loses the additional degrees of information encoded in the different PSF colors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.334, + 0.347, + 0.35 + ], + "angle": 0, + "content": "8. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.469, + 0.48 + ], + "angle": 0, + "content": "We demonstrated a novel compressive high-speed video methodology that relies on temporally coding the camera's PSF using a spinning diffraction grating. Our approach can turn ordinary cameras, regardless of their designed speed, into high-speed sensors for spatiotemporally sparse scenes. Compared to prior work, our method can tackle different scene types, is more light efficient, and its temporal resolution is easily increased." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.469, + 0.692 + ], + "angle": 0, + "content": "Our prototype relies on a PSF created by a low-cost diffraction grating. Better PSFs, including ones that are optimized using an end-to-end framework [14, 22, 30], would likely produce superior results. Moreover, there are likely more sophisticated mechanical ways to temporally modulate the PSF, potentially providing better performance. For instance, a fast stepper motor that rotates the PSF at discrete steps, synchronized with the reconstruction resolution, could help minimize the mismatch between the discretized model in Eq. (4) and actual measurements. Additionally, other forms of temporal modulation beyond rotation could also prove useful, such as axially translating the spatial-multiplexing optic to scale the PSF, or using a spatial-light modulator to create arbitrary programmable PSFs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.469, + 0.814 + ], + "angle": 0, + "content": "Additionally, better reconstruction procedures that leverage optical flow [34], dictionary learning [12, 13], or deep learning [14, 22, 30] could significantly improve quality over our naive implementations. Finally, our method is orthogonal to previous compressive video approaches mentioned in Sec. 2; through novel combinations of these techniques, we believe it may be possible to spin out reconstructions at even higher spatial and temporal resolutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. We thank Amitabh Vyas and Ioannis Gkioulekas for fruitful discussions, and Justin Macey for lending us the motion capture equipment. Matthew O'Toole was supported by a National Science Foundation CAREER award (IIS 2238485)." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.091, + 0.597, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.116, + 0.892, + 0.155 + ], + "angle": 0, + "content": "[1] Ronald J Adrian and Jerry Westerweel. Particle image velocimetry. Number 30. Cambridge university press, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.159, + 0.894, + 0.228 + ], + "angle": 0, + "content": "[2] Nick Antipa, Patrick Oare, Emrah Bostan, Ren Ng, and Laura Waller. Video from stills: Lensless imaging with rolling shutter. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2019. 2, 3, 4, 5, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.231, + 0.895, + 0.272 + ], + "angle": 0, + "content": "[3] Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences, 2(1):183-202, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.274, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[4] Joyce Bedi, Douglas Collins, and Harold Eugene Edgerton. Seeing the unseen: Dr. Harold E. Edgerton and the wonders of Strobe Alley. mIt press, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.317, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[5] Jacky CP Chan, Howard Leung, Jeff KT Tang, and Taku Komura. A virtual reality dance training system using motion capture technology. IEEE transactions on learning technologies, 4(2):187-195, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.374, + 0.751, + 0.388 + ], + "angle": 0, + "content": "[6] Roger N. Clark. Clarkvision.com. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[7] P Doron, L Bertuccioli, J Katz, and TR Osborn. Turbulence characteristics and dissipation estimates in the coastal ocean bottom boundary layer from pv data. Journal of Physical Oceanography, 31(8):2108-2134, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.447, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[8] Liang Gao, Jinyang Liang, Chiye Li, and Lihong V Wang. Single-shot compressed ultrafast photography at one hundred billion frames per second. Nature, 516(7529):74-77, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.504, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[9] Jinwei Gu, Yasunobu Hitomi, Tomoo Mitsunaga, and Shree Nayar. Coded rolling shutter photography: Flexible spacetime sampling. In 2010 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2010. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.575, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[10] Evgeny Hahamovich, Sagi Monin, Yoav Hazan, and Amir Rosenthal. Single pixel imaging at megahertz switching rates via cyclic hadamard masks. Nature communications, 12(1):1-6, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.633, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[11] Barmak Heshmat, Guy Satat, Christopher Barsi, and Ramesh Raskar. Single-shot ultrafast imaging using parallax-free alignment with a tilted lenslet array. In CLEO: Science and Innovations, pages STu3E-7. Optica Publishing Group, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.704, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[12] Yasunobu Hitomi, Jinwei Gu, Mohit Gupta, Tomoo Mitsunaga, and Shree K Nayar. Video from a single coded exposure photograph using a learned over-complete dictionary. In 2011 International Conference on Computer Vision, pages 287-294. IEEE, 2011. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.774, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[13] Jason Holloway, Aswin C Sankaranarayanan, Ashok Veeraraghavan, and Salil Tambe. Flutter shutter video camera for compressive sensing of videos. In 2012 IEEE International Conference on Computational Photography (ICCP), pages 1-9. IEEE, 2012. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[14] Michael Iliadis, Leonidas Spinoulas, and Aggelos Katsaggelos. Deepbinarymask: Learning a binary mask for video compressive sensing. Digital Signal Processing, 96:102591, 2020. 2, 3, 9" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.116, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10797" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[15] Ulugbek S Kamilov. A parallel proximal algorithm for anisotropic total variation minimization. IEEE Transactions on Image Processing, 26(2):539-548, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[16] Roman Koller, Lukas Schmid, Nathan Matsuda, Thomas Niederberger, Leonidas Spinoulas, Oliver Cossairt, Guido Schuster, and Aggelos Katsaggelos. High spatio-temporal resolution video with compressed sensing. Optics express, 23(12):15992-16007, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[17] Alex Liberzon, Theo Käufer, Andreas Bauer, Peter Vennemann, and Erich Zimmer. Openpiv/openpivpython: Openpivpython v0.23.4, Jan. 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[18] Rui Lima, Shigeo Wada, Shuji Tanaka, Motohiro Takeda, Takuji Ishikawa, Ken-ichi Tsubota, Yohsuke Imai, and Takami Yamaguchi. In vitro blood flow in a rectangular pdms microchannel: experimental observations using a confocal micro-piv system. Biomedical microdevices, 10:153-167, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[19] Yang Liu, Xin Yuan, Jinli Suo, David J Brady, and Qionghai Dai. Rank minimization for snapshot compressive imaging. IEEE transactions on pattern analysis and machine intelligence, 41(12):2990-3006, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[20] Patrick Llull, Xuejun Liao, Xin Yuan, Jianbo Yang, David Kittle, Lawrence Carin, Guillermo Sapiro, and David J Brady. Coded aperture compressive temporal imaging. Optics express, 21(9):10526-10545, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.49 + ], + "angle": 0, + "content": "[21] LM Lourenco, A Krothapalli, and CA Smith. Particle image velocimetry. In Advances in fluid mechanics measurements, pages 127-199. Springer, 1989. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[22] Julien NP Martel, Lorenz K Mueller, Stephen J Carey, Piotr Dudek, and Gordon Wetzstein. Neural sensors: Learning pixel exposures for hdr imaging and video compressive sensing with programmable sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1642-1653, 2020. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[23] Amir Matin and Xu Wang. Compressive coded rotating mirror camera for high-speed imaging. In Photonics, volume 8, page 34. MDPI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[24] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[25] Albert Abraham Michelson. Experimental determination of the velocity of light: made at the US Naval Academy, Annapolis, volume 1. US Nautical Almanac Office, 1880. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[26] Kristina Monakhova, Vi Tran, Grace Kuo, and Laura Waller. Untrained networks for compressive lensless photography. Optics Express, 29(13):20913-20929, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[27] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[28] SHR Müller, B Böhm, M Gleißner, R Grzeszik, S Arndt, and A Dreizler. Flow field measurements in an optically accessible, direct-injection spray-guided internal combustion engine using high-speed piv. *Experiments in fluids*, 48:281-290, 2010. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[29] Gergely Nagymáte and Rita M Kiss. Application of optitrack motion capture systems in human movement analysis: A systematic literature review. *Recent Innovations in Mechatronics*, 5(1):1-9, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[30] Cindy M Nguyen, Julien NP Martel, and Gordon Wetzstein. Learning spatially varying pixel exposures for motion deblurring. arXiv preprint arXiv:2204.07267, 2022. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[31] Basilio Pueo Ortega and José M Jiménez Olmedo. Application of motion capture technology for sport performance analysis. Retos:rientas en Educación física, deporte y recreación, (32):241-247, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[32] Travis Portz, Li Zhang, and Hongrui Jiang. Random coded sampling for high-speed hdr video. In IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[33] Ramesh Raskar, Amit Agrawal, and Jack Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[34] Dikpal Reddy, Ashok Veeraraghavan, and Rama Chellappa. P2c2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336. IEEE, 2011. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[35] Aswin C, Sankaranarayanan, Pavan K Turaga, Rama Chellappa, and Richard G Baraniuk. Compressive acquisition of linear dynamical systems. SIAM Journal on Imaging Sciences, 6(4):2109-2133, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[36] Fulvio Scarano. Tomographic piv: principles and practice. Measurement Science and Technology, 24(1):012001, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[37] Yoav Y Schechner, Shree K Nayar, and Peter N Belhumeur. Multiplexing for optimal lighting. IEEE Transactions on pattern analysis and machine intelligence, 29(8):1339-1354, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[38] Shubham Sharma, Shubhankar Verma, Mohit Kumar, and Lavanya Sharma. Use of motion capture in 3d animation: motion capture systems, challenges, and recent trends. In 2019 international conference on machine learning, big data, cloud and parallel computing (comitcon), pages 289-294. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[39] Mark Sheinin, Matthew O'Toole, and Srinivasa G Narasimhan. Deconvolving diffraction for fast imaging of sparse scenes. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2021. 2, 3, 4, 5, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[40] Mark Sheinin, Dinesh N Reddy, Matthew O'Toole, and Srinivasa G Narasimhan. Diffraction line imaging. In European Conference on Computer Vision, pages 1-16. Springer, 2020. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[41] Mark Sheinin, Yoav Y Schechner, and Kiriakos N Kutulakos. Computational imaging on the electric grid. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6437-6446, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[42] H Shen, L Gan, N Newman, Y Dong, C Li, Y Huang, and YC Shen. Spinning disk for compressive imaging. Optics letters, 37(1):46-48, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[43] H Shen, N Newman, L Gan, SC Zhong, Y Huang, and YC Shen. Compressed terahertz imaging system using a spin" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10798" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "disk. In 35th International Conference on infrared, millimeter, and terahertz waves, pages 1-2. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[44] Yaoyao Shi, Youwen Liu, Wei Sheng, Jiming Wang, and Tong Wu. Speckle rotation decorrelation based single-shot video through scattering media. Optics express, 27(10):14567-14576, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[45] Ho Shiraga, M Heya, A Fujishima, O Maegawa, K Shimada, Y Kato, T Yamanaka, and S Nakai. Laser-imploded core structure observed by using two-dimensional x-ray imaging with 10-ps temporal resolution. Review of scientific instruments, 66(1):722-724, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[46] Masada Tzabari and Yoav Y Schechner. Polarized opticalflow gyroscope. In Proceedings of the European Conference on Computer Vision, pages 363-381. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[47] BW Van Oudheusden. Piv-based pressure measurement. Measurement Science and Technology, 24(3):032001, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[48] Ashok Veeraraghavan, Dikpal Reddy, and Ramesh Raskar. Coded strobing photography: Compressive sensing of high speed periodic videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(4):671-686, 2010. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[49] Zihao Wang, Leonidas Spinoulas, Kuan He, Lei Tian, Oliver Cossairt, Aggelos K Katsaggelos, and Huajin Chen. Compressive holographic video. Optics express, 25(1):250-262, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[50] Gil Weinberg and Ori Katz. 100,000 frames-per-second compressive imaging with a conventional rolling-shutter camera by random point-spread-function engineering. Optics Express, 28(21):30616-30625, 2020. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.47, + 0.559 + ], + "angle": 0, + "content": "[51] Charles Wheatstone. An account of some experiments to measure the velocity of electricity and the duration of electric light. Philosophical Transactions of the Royal Society of London, (124):583-591, 1834. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[52] Jinhui Xiong, Ramzi Idoughi, Andres A Aguirre-Pablo, Abdulrahman B Aljedaani, Xiong Dun, Qiang Fu, Sigurdur T Thoroddsen, and Wolfgang Heidrich. Rainbow particle imaging velocimetry for dense 3d fluid velocity imaging. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10799" + } + ] +] \ No newline at end of file diff --git a/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/280659a7-98c7-41f5-b75f-aea246ac862a_origin.pdf b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/280659a7-98c7-41f5-b75f-aea246ac862a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..274079948609fc1b84ceb098c1f0973ea7e91424 --- /dev/null +++ b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/280659a7-98c7-41f5-b75f-aea246ac862a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:259652e9b9d61988aa8d3bd94a27de7f40d99595d95b66f5c94818da9ad3d7e1 +size 9026774 diff --git a/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/full.md b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/full.md new file mode 100644 index 0000000000000000000000000000000000000000..839a32e98627e6ac7f610a86a635a611cf92a6af --- /dev/null +++ b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/full.md @@ -0,0 +1,429 @@ +# SpinCam: High-Speed Imaging via a Rotating Point-Spread Function + +Dorian Chan, Mark Sheinin, and Matthew O'Toole + +Carnegie Mellon University, Pittsburgh, PA 15213, USA + +dychan@andrew.cmu.edu, marksheinin@gmail.com, mpotoole@cmu.edu + +# Abstract + +High-speed cameras are an indispensable tool used for the slow-motion analysis of scenes. However, the fixed bandwidth of any imaging system quickly becomes a bottleneck, resulting in a fundamental trade-off between the camera's spatial and temporal resolutions. In recent years, compressive high-speed imaging systems have been proposed to circumvent these issues by optically encoding the signal and using a reconstruction procedure to recover a video. Our work proposes a novel approach for compressive high-speed imaging based on temporally coding the camera's point-spread function (PSF). By mechanically spinning a diffraction grating in front of a camera, the sensor integrates an image blurred by a PSF that continuously rotates over time. We also propose a deconvolution-based reconstruction algorithm to reconstruct videos from these measurements. Our method achieves superior light efficiency and handles a wider scene class than prior methods. Also, our mechanical design yields flexible temporal resolution that can be easily increased, potentially allowing capture at $192\mathrm{kHz}$ far higher than prior works. We demonstrate a prototype for various applications, including motion capture and particle image velocimetry (PIV). + +# 1. Introduction + +The world is full of high-speed events that are imperceptible to both the human eye and most conventional cameras (e.g., the flicker of electric lights or the motion of fast-moving objects). The traditional approach to capturing such events is using slow-motion photography, i.e., use a camera to shoot images at a high frame rate and display the results at a lower frame rate. Slow-motion photography is often used for artistic effects in film and as a measurement tool in scientific applications. However, despite significant advancements in sensor technology, conventional high-speed cameras tend to be expensive, bulky, and have high throughput requirements for recording visual content at both high frame rates and high spatial resolutions. These hardware limitations set an upper limit to the cameras' capture rates. + +![](images/637c746f8ed30d0313223c23ab115d9ec8ed05b139f74c8859fef505e38a4bbc.jpg) +Figure 1. Principle of high-speed imaging with a rotating diffraction grating. A camera captures a long exposure image of a moving scene. During the camera's exposure period, the diffraction grating placed in front of the camera continuously rotates to generate a time-varying point-spread function (PSF) that encodes scene dynamics into a single frame. Solving a deconvolution problem yields a high-speed video from the captured frame. + +Compressed sensing provides a framework to circumvent such issues. Since videos are highly compressible, it is possible to recover a video from a single snapshot by (i) compressively recording spatiotemporal visual information onto a sensor and (ii) solving a reconstruction problem to recover the video using priors on the visual content. Among the large body of work in compressive high-speed imaging, most techniques temporally modulate the light incident on sensor pixels [48, 12, 20, 22, 9, 34, 32, 35, 14, 16, 19, + +13, 49], either through the use of a spatial light modulator (SLM) or novel sensors capable of turning pixels on or off. Alternatively, by using spatial-multiplexing optics (e.g., diffusers, diffraction gratings) to distribute light from any scene point to many camera pixels, it is possible to perform high-speed imaging by using a rolling shutter [2, 50, 26], a reduced region-of-interest (ROI) [39], or a line sensor [40]. + +Our work proposes a fundamentally different approach to compressive high-speed imaging. Similar to past works that make use of spatial-multiplexing optics, we propose to distribute light across a sensor by using a dual-axis diffraction grating, which diffracts light horizontally and vertically (see Fig. 1). Unlike prior work, however, we opt to change the point-spread function (PSF) during the camera's exposure period. Specifically, we use a motor to mechanically rotate the PSF at high speeds, record a single global-shutter frame, and solve a simple deconvolution-based reconstruction procedure to reconstruct scene dynamics. Please refer to Fig. 1 for an overview of the process. + +Our approach, like past work using spatial-multiplexing optics [2, 50, 40, 39], is primarily targeted towards spatiotemporally sparse scenes. This imaging regime includes motion capture and particle-image velocimetry (PIV), which see broad application in domains like biomechanics [29], sports science [31], animation and film-making [38], virtual and augmented reality [5], aerodynamics [47], fluid mechanics [36], biomedical engineering [18], oceanography [7], and combustion and engine research [28]. We demonstrate our approach for motion capture and PIV in Sec. 6. Other potential applications for spatiotemporally sparse video include high-speed 3D scanning [24] and capturing the fast flicker of electric bulbs [41]. + +Our approach combines four key advantages over existing methods targeting spatiotemporally sparse scenes: + +(I) Superior light efficiency. In many prior methods, the scene light is dispersed using specialized optics to reach many camera pixels, but only a small subset of camera pixels collect light at each timestamp [2, 40, 39, 50]. Thus, most light is lost by falling outside the 'active' pixels. Conversely, in our work, the signal from different timestamps is encoded by different temporal PSFs, and the entire sensor is left exposed, yielding far superior light efficiency. We show this advantage in simulated experiments in Sec. 3. + +(II) Easily increasable temporal resolution. In most prior works, temporal resolution (i.e., FPS) is fundamentally limited by the speed of the respective technology (e.g., the modulation speed of a SLM/sensor [12, 22, 34, 32, 14, 49], or the read-out speed of a camera [2, 40, 39, 50, 9]). In contrast, the temporal resolution of our approach is determined by the mechanical rotation speed and thus can be easily increased by rotating the PSF faster. Thus, we show that our approach can resolve a LED flickering at $96\mathrm{kHz}$ . + +(III) A wider class of supported scenes. Prior works using spatial-multiplexing optics allocate a fixed amount of pixels for each timestamp (e.g., a single rolling shutter row) [2, 40, 39, 50]. This limits the amount of information recoverable per frame, imposing a spatial sparsity constraint on each high-speed frame individually. Instead, our solution multiplexes the signals from different timestamps together, yielding a sparsity constraint over the entire video. Our method can thus handle temporally sparse but spatially dense scenes, verified in simulated and real experiments. + +(IV) Simple hardware. While many prior works require specialized optics (e.g., SLMs [12, 34, 49], piezoelectric stages [20, 16, 19]) or specialized sensors (e.g., coded exposure pixels [22, 14, 32]), our setup can work with any global-shutter camera combined with simple low-cost hardware that is readily available at your local hobby store. We include a DIY guide as part of our supplementary materials. + +# 2. Related work + +# 2.1. Spinning Disks & Revolving Mirrors + +Mechanical spinning disks and revolving mirrors have been used to measure transient phenomena for centuries. In 1834, Wheatstone [51] estimated the velocity of electricity traveling through long copper wires by observing the sparks generated at the ends. Using a revolving mirror, Wheatstone observed a $1\mu \mathrm{s}$ time difference, and concluded the velocity of electricity to be 288,000 miles per second. Fizeau and Foucault [25] later measured the speed of light through various apparatuses based on revolving mirrors and spinning cogwheels. For example, in 1848-49, Fizeau measured the transit time of light from a source, to a mirror, and back to the source by using a rotating wheel to modulate the light. + +Though rare in modern consumer electronics, modulating light mechanically with a spinning element is both faster and lower cost than most modern digital modulation solutions. For example, digital micromirror devices (DMDs) only operate up to $30\mathrm{kHz}$ , whereas spinning disks can offer rates up to $2.4\mathrm{MHz}$ [10]. As a result, coded disks have been used for single pixel imaging at rates up to 72 FPS [43, 42, 10], and revolving coded mirrors have been used to reconstruct 1400 frames from a single shot [23]. A spinning camera was used to recover an 11-frame video through scattering media by disambiguating the resulting rotated speckle patterns [44], while a rotating camera coupled with a polarizer generalized the optical-flow equation to include polarization cues [46]. An event camera can be combined with a spinning polarizer for fast, accurate shape-from-polarization [27]. This work explores the use of spinning elements for compressive high-speed video. + +# 2.2. Pixel-wise Coded Exposure + +Coded exposure photography involves temporally modulating the light incident on a sensor, which has long been + +used to capture dynamic events [4]. Raskar et al. [33] used a liquid-crystal panel to rapidly open and close the shutter of a camera; the motion blur associated with fast-moving objects could then be reliably removed through deconvolution. This methodology was later extended for recovering full high-speed videos of periodic [48] and generic scenes [13]. + +A popular approach to compressive high-speed imaging involves pixel-wise exposure control [12, 9, 34, 32, 35, 22, 14, 20, 16, 19]. In practice, the codes are often binary, and the corresponding reconstruction techniques rely on various priors to reconstruct videos. Example priors include assuming the scene produces periodic motion [48] or satisfies the brightness constancy assumption [34], and might rely on dictionary learning [12, 13] and neural networks [22, 14, 30] to aid the reconstruction procedure. + +These techniques lack the (II) temporal resolution and (IV) hardware complexity advantages of our approach since they typically require specialized hardware (e.g., SLMs [12, 34, 49], piezoelectric stages [20, 16, 19]) or specialized sensors [22, 14, 32]) which can be expensive, difficult to acquire, and still be limited in frame rate. Moreover, as shown in the supplementary materials, since these methods do not spread light across multiple pixels, these systems struggle with spatially sparse scenes, which appear in many practical applications of high-speed imaging [40, 39]. + +# 2.3. Spatial-Multiplexing Optics + +An alternative approach to capture space-time volumes (videos) involves spatial multiplexing. By placing an optical element in front of a sensor or camera (e.g., a diffuser or diffraction grating), the light from a single point in the scene spreads to multiple camera pixels according to the optical system's PSF. When combined with a rolling shutter camera that samples individual sensor rows at fast rates [2, 50, 26], it is possible to reconstruct high-speed videos. Alternatively, a limited region-of-interest (ROI) or 1D sensor can also be used to sample measurements at fast rates [40, 39]. By integrating diffraction patterns generated by time-varying coded masks, it is even possible to reconstruct 4D volumes of spacetime from a single snapshot [49]. + +Similar ideas have been used in combination with streak cameras, an optoelectronic device used to measure ultra-fast light phenomena at femtosecond time scales. A streak camera spatially multiplexes the scene's temporal response, by deflecting photoelectrons to different points on a sensor as a function of time. While designed to be a line sensor, it is possible to encode 2D spatial information by incorporating a tilted microlens [45] or pinhole array [11]. In Gao et al. [8], the 2D scene at different timestamps is coded with a fixed mask, sheared a temporally varying amount, and then overlayed onto the 2D signal captured by a streak camera. + +These methods lack the (I) light efficiency and (III) density advantages of our approach—they typically rely + +![](images/ff9e9d3158198e325b52bd2d5f64a26e62377ec5dc4c139f65e95491e242c435.jpg) + +![](images/046888cc7066ba023f0954b21e72779acda91a1bc7415bbea8767ec4fb65a379.jpg) +(a) Captured image +(b) Recovered time signal (ms) +Figure 2. High-speed reconstruction of a flickering LED. The LED in this experiment flashes at $96\mathrm{kHz}$ , as driven by an Arduino. (a) The diffraction image encodes the high-frequency patterns of the LED (zoom into inset). (b) This temporally-varying signal can be directly extracted from the image to recover the intensity pattern of the LED. A frequency analysis of the recovered time signal indicates the system's ability to resolve frequencies up to $96\mathrm{kHz}$ . Given that the Nyquist frequency is half that of the sample rate, the effective framerate for the system is therefore over 192,000 FPS. + +on a small subset of camera pixels to reconstruct each frame, "wasting" light and limiting recoverable scene density. Furthermore, their capture speed is limited by the maximum speed of the hardware used (i.e., line-scan mode [40], ROIs [39] or rolling shutter [2, 50]). Thus, these methods lack the ability to (II) easily increase temporal resolution. + +# 3. High-Speed Imaging using a Time-Varying Point-Spread Function + +Consider a scene that consists of a single bright LED in a dark room, where the LED is configured to pulse rapidly over time. Unless it is faster than the blinking LED, a normal camera forms an image of the LED, but provides no information about the LED's temporal dynamics. + +Our proposed approach to imaging high-speed events involves two steps. First, following prior work on diffraction imaging [40, 39], we place a diffraction grating in front of the camera, which blurs the scene with a rainbow PSF. The light from the LED is therefore spread over many additional camera pixels. Second, we propose to rotate the diffraction grating such that the corresponding PSF changes over time. This produces measurements that spatially encode temporally varying signals, which can be used to analyze extremely high-speed events (see Fig. 2 for an example). + +In this section, we describe the image formation model and reconstruction process for recovering a high-speed video from an image captured with a generic time-varying PSF. We then analyze the special case of using a rotating PSF for capturing images. + +![](images/01965abe5cea810eac69f31f79cf91612ca3f8b39ee3975e512799c5cde86b82.jpg) +Rolling $^+$ diffuser [2] (64 row exposure) + +![](images/dac4d2807bf373bd873263f60d65e44a9e94fbb0c1535f7847efc409ed2656a7.jpg) +ROI + grating [39] (16 row ROI) + +![](images/227aa7f4891ec1d7d47abb9c09d833cfb578d1c163967a7c7c09aa1f77ba8a1d.jpg) +Rotating PSF + +![](images/d1f555643eb4531ba875b6077d6408e451f399a999d3405af70a087574b15c01.jpg) +(a) Recovered brightness of time-varying light source + +![](images/f9c5ed59b216b5477c59963999ea4dadeff50591458faab07bede059eb8a3f80.jpg) + +![](images/904dae00d73e8faf386a6baa51ef674475d37a55320431a5a07b391603decf82.jpg) +Rolling $^+$ diffuser [2] (64 row exposure) +(b) Recovery of dense moving scene +Figure 3. Light efficiency experiment. In the presence of Poisson and read noise, we simulated the measurements captured by a Canon 5D Mark III sensor [6] with ISO 800. (a) We simulate a dim point light source with a time-varying intensity profile. (b) We simulate the word "test" moving across the scene. In both cases, our approach is able to much more accurately resolve the signal, compared to a rolling shutter and diffuser as well as a ROI and grating. We theorize that this is due to the fact that our approach captures more light for every time step, and that the requisite demultiplexing minimizes the effects of noise. + +![](images/cd01ad50f52be70cca44dd33940637c2862d8338328822c387761c44d661843b.jpg) + +![](images/69ea9149a41dea49faf869542d3ee0c232c807ad2a168c1142ced317ca5e6be0.jpg) +ROI + grating [39] (16 row ROI) + +![](images/b283a5de6050ff094b2edca34f947b1ea19c92d6c60f6ed2a64f33ee79020297.jpg) + +![](images/136f9d6de205a41cc5d13d5135ed0a542751661873120b57edd42202f53a49ff.jpg) +Rotating PSF + +# 3.1. Image Formation Model + +As illustrated in Fig. 1, for a camera where the PSF changes with time, the forward model can be written as: + +$$ +\mathbf {I} (x, y) = \int_ {t = t _ {s}} ^ {t = t _ {s} + E} (\mathbf {k} (t) * \mathbf {s} (t)) (x, y) d t, \tag {1} +$$ + +where $\mathbf{I}(x,y)$ is the captured image, $E$ denotes the exposure time of the camera, and $t_s$ denotes the time when the camera started exposing. The symbol $*$ denotes the spatial convolution operator between the scene $\mathbf{s}(t)$ and the PSF $\mathbf{k}(t)$ at time $t$ . When discretized to $N_E$ time steps from $t_s$ to $t_s + E$ , this can be rewritten in matrix form as: + +$$ +\mathbf {I} = \mathbf {M} \mathbf {s} \tag {2} +$$ + +$$ +\mathbf {M} = \left[ \begin{array}{l l l l} \mathbf {C} _ {1} & \mathbf {C} _ {2} & \dots & \mathbf {C} _ {N _ {E}} \end{array} \right] \tag {3} +$$ + +where $\mathbf{I} \in \mathbb{R}_{+}^{N_{x}N_{y}}$ is the vectorized representation of the captured image, $\mathbf{s} \in \mathbb{R}_{+}^{N_{x}N_{y}N_{E}}$ is the vectorized representation of the high-speed scene, and $\mathbf{C}_k \in \mathbb{R}_{+}^{N_xN_y\times N_xN_y}$ represents the convolution with the kernel at timestep $k$ . + +# 3.2. Reconstruction Process + +Natural scenes contain sparse gradients, and many target applications for high-speed imaging require the tracking of just sparse points. As a result, we recover a high-speed video with an $\ell_1$ sparsity prior and an anisotropic total variation prior, solving the following optimization problem: + +$$ +\underset {\mathbf {x}} {\operatorname {a r g m i n}} \quad \| \mathbf {I} - \mathbf {M} \mathbf {x} \| ^ {2} + \mathcal {I} _ {\mathbb {R} _ {+}} (\mathbf {x}) + \lambda_ {\text {s p a r s i t y}} \| \mathbf {x} \| _ {1} + \tag {4} +$$ + +$$ +\lambda_ {d x} \| \mathbf {D} _ {x} \mathbf {x} \| _ {1} + \lambda_ {d y} \| \mathbf {D} _ {y} \mathbf {x} \| _ {1} + \lambda_ {d t} \| \mathbf {D} _ {t} \mathbf {x} \| _ {1} +$$ + +where $\mathcal{I}_{\mathbb{R}_+}$ denotes a non-negativity prior, $\mathbf{D}_x$ , $\mathbf{D}_y$ , and $\mathbf{D}_t$ compute forward finite differences along the $x$ , $y$ , and $t$ dimensions respectively of the high-speed video $\mathbf{x}$ , and $\lambda_{sparsity}$ , $\lambda_{dx}$ , $\lambda_{dy}$ , and $\lambda_{dt}$ weigh the regularization terms. + +# 3.3. Analysis + +Varying the PSF over time results in fundamentally different properties compared to prior approaches, which we explore in the rest of this section. + +Higher light efficiency. Our method, like other prior compressive video works [2, 50, 40, 39], relies on PSFs with broad spatial supports. A broad spatial support spreads the signal from a single point across multiple pixels. Recent approaches record just a subset of these pixels at each timestep [2, 50, 40, 39]. Because fewer pixels encode each frame and light is lost, assuming a standard noise model with Poisson shot noise and additive Gaussian noise, the reconstruction's signal-to-noise ratio is proportionately decreased. In contrast, under a time-varying PSF, the entire frame is always recorded, so a much larger fraction of the PSF (and therefore light) is captured for every scene point during capture. In addition, the signals from different timestamps are mixed together in the final capture, which could improve the quality of the reconstructions under read noise [37]. Thus, this overall increase in light efficiency leads to better performance in resolving noisy scenes. We show this experimentally in Fig. 3 on two simulated cases: a point light source with time-varying intensity, and a denser moving scene. We find that our approach robustly recovers these two scenes, even under low light conditions. + +A wider class of scenes. Many past approaches achieve high-speeds by allocating a number of pixels per frames that is much lower than the total number of frame pixels [2, 50, 40, 39]. This causes each frame to undergo lossy compression, yielding a sparsity constraint per frame. For example, for a rolling shutter with exposure set to the row rate, each frame is encoded onto just a single sensor row. + +![](images/d432968d16ec4e19a92459dc808b633e0ebdb7bcc216f8a5d04d03ebe5c29026.jpg) +(a) Simulated scene: +(1) Ground truth + +![](images/d000aec2211483a15a7d27bbd0fa8d47bb9f062a13a5751925af30ad2830c416.jpg) + +![](images/a5e8b014d322cfd0ddde00e5e429c73463001a349f8df40f49b226dc96552969.jpg) + +![](images/e901f427caf1a3fd9567a48f45e8ee960a9650bc3d26da70af3e9131e9862719.jpg) +(2) Rolling + diffuser (1 row exposure) + +![](images/0e24694fe7297650688b2d64d41043775fcb781b46597a398428cca98e4a58cc.jpg) + +![](images/188ddd3c8f10b71fadba1178662df2b08f59c5da74d172088e11b5e9c77d61fa.jpg) + +![](images/c5c869ffb06cc00be5774852da6f32a5a0937c4809d2ece88c28517a0edcd158.jpg) +Rolling $^+$ diffuser [2] (64 row exposure) + +![](images/1e3d10399923c49f903784a620cb3889e2b8cbf3b6e9f9363104a08924b03329.jpg) + +![](images/966659a7743685c7cc77b46b2eb15e298a82dfe1ac458d24ed07b7689d092108.jpg) + +![](images/74d0b8820f71c4f8406782d741b6dd9b9a47bfc9ad845e7262c7dea3c6eb58f6.jpg) +(4) ROI + grating [39] (2 row ROI) + +![](images/8f7d9239f844fdbce4e14491ddf9fbf7c38101051b7b1030e26f9fb0ae5f2670.jpg) + +![](images/e9979480d83129dcf248ebe41276a2f3f9aca2182a4a3ddbe262f6db4833d8d1.jpg) + +![](images/c852d951ceffc07a5f74246792081b7450de61742ca0d5859c9747dcfdb71d10.jpg) +(5) ROI + grating [39] (16 row ROI) + +![](images/9219cebf810a69b0003e1f37155522c0122e55709aae4bf07eeac749b38dcff9.jpg) + +![](images/1b2c0d35f239a6724e1fc2ffec104e67f35c9ac041f9c6ea0a519b68148ab01a.jpg) + +![](images/f1f35ff8cf4a4c469b618a90fb3f5c9d9cc9eb34f4a1c0e25420ad32903fd54c.jpg) +(6) Rotating PSF + +![](images/485458d0fed3a27c6c7e34ef794be5a2f206169e3fafd6e9b327340b96201c77.jpg) +(b) Real scene: +(1) Reference frame +Figure 4. Reconstructing temporally sparse, but spatially dense scenes. (a) A simulated scene of two people dancing. The video is dark for the entire sequence, except for a short sequence of fast-moving frames where the dancers are strobed. (a2)-(a3) We implement a setup based on a dual rolling shutter combined with a diffuser [2]. In (a2), with the exposure time set to match the row rate of the shutter, the reconstruction is unable to well-resolve the dense scene. In (a3), with the exposure time set to 64 times the previous value, the reconstruction is able to recover the dancers, at the cost of temporal resolution [2]. (a4)-(a5) We implement an approach based on fast temporal sampling by using a ROI and a diffraction grating [39]. In (a4), the ROI consists of just two rows, resulting in low quality results. In (a5), by using more bandwidth and exposing 16 rows, such an approach can get better results, but this reduces the recoverable framerate by 8 times in practice. (a6) In contrast, our methodology recovers the dense scene much more accurately compared to (a2) and (a4)—the signal is encoded over a much larger portion of sensor. This does not require increasing the exposure time and therefore decreasing the temporal resolution as in (a3), or increasing the requisite bandwidth (a5). (b) A real world scene, where a spinning fan is strobed twice during the camera exposure at unknown times. Our approach is able to disambiguate the fan rotation at each strobe time. + +![](images/507f6def4fc57260c121ca72a3cdece98ed1569d392c0a7b3e527274bb46bc8d.jpg) +(2) Rotation-encoded frame + +![](images/60446f80bb913cb5584e00b4b3a839b7aeb90edf4e3173d8f69666cee91bcd9e.jpg) +(3) Recovered frames + +In contrast, our method allocates the entire frame to encode a plurality of video frames belonging to a fixed time interval. Therefore, our method assumes sparsity in a more general sense, namely sparsity over the entire fixed time interval. The encoded fixed time interval can be either spatially sparse and temporally dense, which fits most prior-work assumptions, or temporally sparse but spatially dense, which prior approaches do not consider due to the fixed bitrate per frame imposed by their low-dimensional encoding. + +To illustrate how this is an advantage, consider the case of a dense scene that is entirely dark, except for a single time instant with a flash of light. If sampling the scene using a rolling-shutter or line sensor, the flashed frame will be encoded onto just a single row. In contrast, a time-varying + +PSF encodes the flashed frame on the entire frame, wasting no resources encoding the dark frames. Therefore, in this case, a time-varying PSF enables a better reconstruction. + +As shown in Fig. 4, handling temporally sparse, but spatially dense scenes makes our approach very useful under strobed lighting. As shown in real and simulated experiments, a rotating PSF is able to resolve the content within each individual strobe at high spatiotemporal resolution, without any constraints on scene structure between strobes. + +# 4. Using a Rotating PSF + +Sec. 3 described the encoding and reconstruction of high-speed video using a time-varying PSF. While a time-varying PSF can be implemented in a number of ways, + +![](images/54be67f8effe5d6f1fd305e1f872fa462dbb6fb9250e5d040959abc2cc5abeae.jpg) +Figure 5. The PSF and spatial frequency coding of time. A rotating PSF effectively codes the signal from different timesteps into a rotating set of spatial frequencies in the captured image. For instance, a rotating line PSF encodes different timestamps into different lines in Fourier space. The star PSF from a double-axis diffraction grating behaves as a mixture of the point and line cases. Its Fourier transform contains bright line components that provide temporal signal, but remains broadband for spatial resolution. + +for the sake of cost and speed, we simply rotate a PSF-generating optical element over the course of the camera exposure. This results in a rotation of the corresponding PSF. In this particular case, $\mathbf{k}(t)$ can now be rewritten as $\mathbf{k}(t) = \mathbf{R}(\mathbf{k},\theta_t)$ , where $\mathbf{k}$ is the stationary PSF of the rotating optical element, $\theta_{t}$ is the angle of the element at time $t$ , and $\mathbf{R}(\mathbf{k},\theta_t)$ denotes a rotation of $\mathbf{k}$ by $\theta_{t}$ . In the rest of this section, we analyze the properties of such a rotating setup. + +PSF performance. The selected PSF affects the spatiotemporal reconstruction quality of every video frame. Because each time step is convolved with a rotated PSF, the PSF's rotated Fourier transform determines which parts of the spectrum will encode each time step. To illustrate, an ideal thin lens having a perfect impulse PSF will have perfect spatial resolution because its Fourier transform contributes values in all spectrum frequencies (Fig. 5(a)). It will, however, have poor temporal resolution because every other timestamp will also contribute values across the entire spectrum, making them impossible to distinguish. + +In contrast, consider a line PSF (Fig. 5(b)). In the Fourier domain, the PSF remains a line that goes through the origin. When the line PSF rotates, its Fourier spectrum rotates correspondingly, such that each time step is encoded into a different line in Fourier space, without interfering with values from previous timestamps. As a result, the signal at each time step can be easily reconstructed simply by extracting its corresponding line from Fourier space. However, because only a single direction of spatial frequencies is available for reconstruction for every time step, the re + +![](images/7efc04c745edaac170df2bcc64f3b366a07f6fd5eace08118a557c7e0ead0e9f.jpg) +(a) Setup + +![](images/56dc7d3ed5f2f7dcb5744e972b3c553fbebb00b03fa4ece179945021f1b9090b.jpg) +(b) Captured PSF +Figure 6. Image of our hardware setup, consisting of a global-shutter camera and a circular diffraction grating attached to a 20,000 RPM motor. + +sulting recovered frame will have poor spatial resolution. + +Inspired by these two cases, we use a double-axis diffraction grating for the experiments shown in this work, whose corresponding PSF is shown in Fig. 6(b). This star PSF contains a bright DC spot at the center, intuitively providing spatial resolution. Simultaneously, the PSF also contains line-shaped rainbow lobes that provide temporal information. We visualize this phenomena in Fig. 5(c). + +Temporal resolution. The temporal resolution of our system is fundamentally determined by the amount of temporal variation of the PSF between different timestamps, or namely, the minimum angle required to reliably distinguish between two different rotations of the PSF. As a result, the temporal resolution can be increased either by increasing the pixel resolution, the spatial spread of the PSF, or the rotation speed. But since our grating PSF is roughly $90^{\circ}$ periodic, rotating more than $90^{\circ}$ during the exposure causes temporal ambiguity and degrades the reconstruction quality. + +In the supplement, we further analyze PSF performance from the mutual coherence perspective. We also discuss how a time-varying PSF behaves as a form of coded exposure in frequency, as shown by the above Fourier analysis. + +# 5. Implementation + +Hardware. An image of our hardware setup is given in Fig. 6. To generate a usable PSF, we cut a 531-lines per millimeter double-axis diffraction grating sheet into a circular disk of radius 6 inches. To ensure that the sheet stayed flat, we taped it onto a laser-cut acrylic disk with a radius of 6 inches. We attached this element to a Puly DC 775 Motor with a shaft coupling to rotate it. We place this entire rotating setup $5\mathrm{mm}$ in front of the camera's lens. We used a UI-3070CP-C-HQ R2 camera, fitted with a $9\mathrm{mm}f / 1.4$ lens (Fujinon HF9HA). We also placed a second identical camera beside our system to capture intermittent reference + +![](images/562699de3ebf26737e1e23971575f024f3aae293a3bcc6de179811241926ceb0.jpg) + +![](images/546cc8099894e51416c6cb1002c77f4b397b3be5019dba3d88882d77ba0b9188.jpg) + +![](images/2f2199849286190feb69e57e6c590b79331dbe38f5ea07da03de5ebd50c4b6a6.jpg) + +![](images/fc714076934e57c7a33a400aac193490aa16c1ded52d8970146f8a2c921ac059.jpg) + +![](images/0f8d12569c3902f258b4c1f56c0356796152cfbe763414fa468f2221eaca6045.jpg) +(a) Reference start frame + +![](images/ec60926cb0bb1ca8ae17254120ec9bebf3eff312c7abebacd20673bf572706e1.jpg) + +![](images/beda0f73715a8a296e45cbdc2e0623d789a4383bc67f4cf51051a20f8e1ea869.jpg) + +![](images/e14fe5640c9ef078c6b81089c755ae7fac784f18c3570cb4bb8164d173a5629b.jpg) + +![](images/85d9be2171d891c0efa929a139ca1d822d87dcd224504e5dea29b2bd4a57203d.jpg) + +![](images/034645de103e544372b6ec03c146fcd76d0254da7b66b058fe607edd29c740c9.jpg) +(b) Reference end frame + +![](images/84bed79ee189f771c0457c284eaa6dc5bd6847eacd36b35ec7923973f01823a7.jpg) + +![](images/5f522b3fce1e253e05adc65308b19972809c54ffc19fd6a3e32abb0dcc1c9b59.jpg) + +![](images/6d1211679be31115af89373bd99fa07b6a9c06a141f34af32cffe651132a6ef1.jpg) + +![](images/9cf22ae89555790b2d9b9817e36b1ddb49a4203e24aeff90aad7962f2ad65352.jpg) + +![](images/2e9d4b45b41132e11641964f2f33f04a564b8fac590a340215d708ad31f94413.jpg) +(c) Rotation-encoded frame +Figure 7. High-speed reconstruction of fast moving objects. Row 1: A ball with a black stripe bounces and spins. Row 2: An I rotates $120^{\circ}$ . Row 3: A toy bullet bouncing off a board. Row 4, 5: Our setup is capable of tracking the movements of both sparse points and dense marker lines for high-speed motion capture. + +![](images/7e1aee89aa1e75e7b1267cf5d573889caeef574bb1b5299c93baef001b574753.jpg) + +![](images/0184d67e45d66823dee45077c4d53c173cfc92fecda4f9f8c66e4ef8ac5bf129.jpg) + +![](images/93088bc551ca8fdc2f9a1f8926a303063e02ca7899e266867b187dd11a48d407.jpg) + +![](images/6ed38485188a7ac5c5fcd92e351904366a0ca25042153962b2300409cb6f1d0a.jpg) + +![](images/d52a0d868b2bfb51bff0591593002071c1a9d87e606d118145b2ce6b4e84bd81.jpg) +(d) Recovered frames + +frames for visualization. Please refer to the supplement for a DIY guide. + +Software. To solve Eq. (4), we used a PyTorch implementation of FISTA [3], combined with an approximate proximal operator [15] to handle the anisotropic total variation prior. We ran 10000 iterations of FISTA for each result shown. For most results, we set $\lambda_{sparsity} = 0$ , $\lambda_{dx} = \lambda_{dy} = 1 \times 10^{-3}$ , and $\lambda_{dt} = 1 \times 10^{-3}$ , but we vary the strength of these terms depending on the brightness of the background and the speed of the motion. We list the parameters for each result and provide sample code in the supplemental. + +Calibration and capture. To calibrate the time-varying PSF of our imaging system, we process a stationary image of a small retroreflector or LED, and then computationally rotate this image to create the dictionary required for Eq. (4)—please see the supplemental for more details. + +For every capture, we need to identify the range of angles that are rotated over during the camera exposure in order to + +solve Eq. (4). This could be done by placing an LED at a known location, and reconstructing which PSFs are present around it. Alternatively, a controllable motor or trigger system could be used to directly provide the requisite angles. In our work, we opt for a unstructured approach, that simply relies on signal being present at the start and end of the camera exposure. Before capture, we first calibrate the motor to rotate roughly $90^{\circ}$ over the course of the exposure. Then, we simply perform a coarse reconstruction of Eq. (4) using the captured image and a dictionary that spans all $360^{\circ}$ , and select the best matching $90^{\circ}$ for a finer reconstruction. Please see the supplemental for an in-depth discussion. + +# 6. Results + +For our real-world reconstructions, we downsampled the input images to $300 \times 400$ pixels and reconstructed $N_{E} = 146$ frames for every input. For the simulated results (Fig. 3, Fig. 4), we instead used $N_{E} = 150$ to better match + +![](images/28e43121f66334602705b045cf561eb19fd29b878e924710ad249695d3981920.jpg) +(a) Rotation-encoded frame + +![](images/36b446289bfa65bf028b8c39f266a788242224c793fe5b456be5e7a4610ff48f.jpg) +(b) Example recovered frame + +![](images/6dab6558f77b3a5a4d9aa36bcfc2e4ee898538962ec4e0b48aee4d3ed5a87e67.jpg) +(c) Flow visualization +Figure 8. High-speed particle image velocimetry. Following Sheinin et al. [39], a magnetic stirrer is placed in a tank seeded with tracer particles. (a) Captured diffraction image. (b) Recovered frame using our reconstruction procedure. (c) We visualize a $17.33\mathrm{ms}$ interval by overlaying the recovered frames. (d) Using OpenPIV [17], we recover the flow between the start and end frame of (c). + +![](images/75d3da9bee7fb70c8070a37ab437c90ea2ce23add0b6717fed5bed470f08e40d.jpg) +(d) Recovered flow + +a dual rolling shutter setup. In this document, we use color to code a video into a single image; we refer the reader to our supplemental materials for directly viewing the recovered videos. The camera exposure was $22.178\mathrm{ms}$ , yielding an effective framerate of 6583.10 FPS. + +We demonstrate our methodology on fast scenes on multiple examples (Fig. 7)—a bouncing ball, an "I" spinning on a fan, toy bullets bouncing off a board, and letters spinning and falling. In the last row of Fig. 7, we applied our technique to marker-based motion capture. Our technique readily captures a wide variety of marker shapes. In Fig. 4(b), we show a temporally sparse, but spatially dense scene of a fan under strobed illumination. + +As shown by Sheinin et al. [39], low-cost, high-speed imaging could help measure the movement of fluids in particle image velocimetry [1, 21, 52]. In short, small tracer particles are placed into the medium, which can then be tracked to visualize the flow. We apply our setup to this task in Fig. 8, imaging a tank filled with water and a magnetic stirrer. To maximize the visible movement, we increased the exposure time of our camera to $100\mathrm{ms}$ and proportionately decreased the speed of the motor. + +Finally, we push the temporal resolution of our approach by imaging an LED pulsed at $96\mathrm{kHz}$ . For this particular result, we used a $16\mathrm{mm}$ focal length lens (Edmund Optics #86-571) to increase the scale of the PSF, and rotated the motor at roughly 15,000 RPM. At resolution $2048\times 1542$ pixels, our methodology is capable of easily resolving this signal as shown in Fig. 2, demonstrating $192\mathrm{kHz}$ capture. + +# 7. Limitations + +Temporal unmixing. Rolling shutter or ROI-based approaches [2, 50, 40, 39] code the signal from each timestamp onto a unique set of measurements that only record the signal at that time. In contrast, with a rotating PSF, the signal from different timestamps is mixed together and requires unmixing using Eq. (4). This unmixing can yield degraded results when there are significant nonidealities, e.g., a poorly calibrated PSF, major spatial variation of the PSF, sizable background signal, or when the scene is too dense. + +![](images/4d9771b55af5f6eab0a9001c5c7ad5ebbb4510ced9c529f3c65a7db406ab85f5.jpg) +PSF spect. +Rec. Mx + +![](images/9de95f1966c497a6c935a12608df5aa7f2dd1074a123afd8b754901f06047179.jpg) +PSF spect + +![](images/e61f5974bd20e2f1d8c9430ba7b45b0b3586de7cf7e5f208be4b67ab5c6455ed.jpg) + +![](images/0f967a5316a8c5a108e5acd302061bd03aab7b932529b7aaba2d649c2ed26691.jpg) + +![](images/63cb8deb4620193d4f109942edf6de57c1e902e69b79c7ff9d0d4262f5c47909.jpg) +Sample recons. frame +Original PSF +Figure 9. Reconstruction sensitivity to PSF spectrum. We reconstruct Fig. 7, Row 2 with digitally augmented versions of the original real measured PSF, to simulate a PSF calibrated under incorrect lighting or from a different material. The reconstruction quality is fairly robust to these PSF variations. + +![](images/e4e44bc08f088db1da0e63300f474cf41ae9fb1ca599477519e5e49c128e183a.jpg) +Rec. $\mathbf{M}\mathbf{x}$ +Sample recons. frame +Bandpass filtered PSF + +![](images/37e30f913d02aa6ada26293d65ec955db406064ef53c5bb841d09b3acb52b099.jpg) +Rec. Mx +Sample recons. frame +Notch filtered PSF + +In these cases, the signal from one timestamp may bleed into the reconstructions for other timestamps, causing a loss of temporal resolution. + +Saturation. Our method can be especially prone to sensor saturation for several reasons. First, the use of a longer camera exposure may saturate those camera pixels which received scene light over a significant portion of the exposure duration. Secondly, a bright flash of light at a single timestamp could saturate the entire sensor, corrupting the whole video sequence, while only a single row (and therefore timestamp) would be saturated for a rolling shutter or ROI. Thirdly, our grating's PSF contains a 'DC spot' that is much brighter than the rainbow streaks. Therefore, well-exposed rainbow streaks yield a large DC component that might saturate camera pixels. This latter problem could be mitigated by using a higher-quality grating. + +Spectrum. The PSF of an object, when viewed through a diffraction grating, depends on its reflectance spectrum, which itself depends on the light source spectrum and object material. Our approach assumes that scenes consist of a single reflectance spectrum so that this PSF is known and fixed for every point in the scene. This assumption makes our approach well-suited for vision applications like motion capture or PIV, which use specialized light sources and markers. In these cases, only slight deviations in PSF occur, which our method can handle (see Fig. 9). However, a + +more general scene may consist of many diverse materials and light sources. This results in many different reflectance spectrums and corresponding PSFs at the sensor, restricting the applicability of our method. + +To address this limitation, the single reflectance spectrum assumption can potentially be relaxed, albeit with tradeoffs. For instance, the reconstruction problem in Eq. (4) can be expanded to explicitly account for these multiple PSFs following Sheinin et al. [39], at the cost of extra computational complexity that scales with the number of possible spectrums. Alternatively, a less wavelength-dependent optic (e.g., a simple diffuser [2]) could be used, such that the PSF does not significantly vary for different wavelengths. However, this loses the additional degrees of information encoded in the different PSF colors. + +# 8. Conclusion and Future Work + +We demonstrated a novel compressive high-speed video methodology that relies on temporally coding the camera's PSF using a spinning diffraction grating. Our approach can turn ordinary cameras, regardless of their designed speed, into high-speed sensors for spatiotemporally sparse scenes. Compared to prior work, our method can tackle different scene types, is more light efficient, and its temporal resolution is easily increased. + +Our prototype relies on a PSF created by a low-cost diffraction grating. Better PSFs, including ones that are optimized using an end-to-end framework [14, 22, 30], would likely produce superior results. Moreover, there are likely more sophisticated mechanical ways to temporally modulate the PSF, potentially providing better performance. For instance, a fast stepper motor that rotates the PSF at discrete steps, synchronized with the reconstruction resolution, could help minimize the mismatch between the discretized model in Eq. (4) and actual measurements. Additionally, other forms of temporal modulation beyond rotation could also prove useful, such as axially translating the spatial-multiplexing optic to scale the PSF, or using a spatial-light modulator to create arbitrary programmable PSFs. + +Additionally, better reconstruction procedures that leverage optical flow [34], dictionary learning [12, 13], or deep learning [14, 22, 30] could significantly improve quality over our naive implementations. Finally, our method is orthogonal to previous compressive video approaches mentioned in Sec. 2; through novel combinations of these techniques, we believe it may be possible to spin out reconstructions at even higher spatial and temporal resolutions. + +Acknowledgements. We thank Amitabh Vyas and Ioannis Gkioulekas for fruitful discussions, and Justin Macey for lending us the motion capture equipment. Matthew O'Toole was supported by a National Science Foundation CAREER award (IIS 2238485). + +# References + +[1] Ronald J Adrian and Jerry Westerweel. Particle image velocimetry. Number 30. Cambridge university press, 2011. 8 +[2] Nick Antipa, Patrick Oare, Emrah Bostan, Ren Ng, and Laura Waller. Video from stills: Lensless imaging with rolling shutter. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2019. 2, 3, 4, 5, 8, 9 +[3] Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences, 2(1):183-202, 2009. 7 +[4] Joyce Bedi, Douglas Collins, and Harold Eugene Edgerton. Seeing the unseen: Dr. Harold E. Edgerton and the wonders of Strobe Alley. mIt press, 1994. 3 +[5] Jacky CP Chan, Howard Leung, Jeff KT Tang, and Taku Komura. A virtual reality dance training system using motion capture technology. IEEE transactions on learning technologies, 4(2):187-195, 2010. 2 +[6] Roger N. Clark. Clarkvision.com. 4 +[7] P Doron, L Bertuccioli, J Katz, and TR Osborn. Turbulence characteristics and dissipation estimates in the coastal ocean bottom boundary layer from pv data. Journal of Physical Oceanography, 31(8):2108-2134, 2001. 2 +[8] Liang Gao, Jinyang Liang, Chiye Li, and Lihong V Wang. Single-shot compressed ultrafast photography at one hundred billion frames per second. Nature, 516(7529):74-77, 2014. 3 +[9] Jinwei Gu, Yasunobu Hitomi, Tomoo Mitsunaga, and Shree Nayar. Coded rolling shutter photography: Flexible spacetime sampling. In 2010 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2010. 2, 3 +[10] Evgeny Hahamovich, Sagi Monin, Yoav Hazan, and Amir Rosenthal. Single pixel imaging at megahertz switching rates via cyclic hadamard masks. Nature communications, 12(1):1-6, 2021. 2 +[11] Barmak Heshmat, Guy Satat, Christopher Barsi, and Ramesh Raskar. Single-shot ultrafast imaging using parallax-free alignment with a tilted lenslet array. In CLEO: Science and Innovations, pages STu3E-7. Optica Publishing Group, 2014. 3 +[12] Yasunobu Hitomi, Jinwei Gu, Mohit Gupta, Tomoo Mitsunaga, and Shree K Nayar. Video from a single coded exposure photograph using a learned over-complete dictionary. In 2011 International Conference on Computer Vision, pages 287-294. IEEE, 2011. 2, 3, 9 +[13] Jason Holloway, Aswin C Sankaranarayanan, Ashok Veeraraghavan, and Salil Tambe. Flutter shutter video camera for compressive sensing of videos. In 2012 IEEE International Conference on Computational Photography (ICCP), pages 1-9. IEEE, 2012. 2, 3, 9 +[14] Michael Iliadis, Leonidas Spinoulas, and Aggelos Katsaggelos. Deepbinarymask: Learning a binary mask for video compressive sensing. Digital Signal Processing, 96:102591, 2020. 2, 3, 9 + +[15] Ulugbek S Kamilov. A parallel proximal algorithm for anisotropic total variation minimization. IEEE Transactions on Image Processing, 26(2):539-548, 2016. 7 +[16] Roman Koller, Lukas Schmid, Nathan Matsuda, Thomas Niederberger, Leonidas Spinoulas, Oliver Cossairt, Guido Schuster, and Aggelos Katsaggelos. High spatio-temporal resolution video with compressed sensing. Optics express, 23(12):15992-16007, 2015. 2, 3 +[17] Alex Liberzon, Theo Käufer, Andreas Bauer, Peter Vennemann, and Erich Zimmer. Openpiv/openpivpython: Openpivpython v0.23.4, Jan. 2021. 8 +[18] Rui Lima, Shigeo Wada, Shuji Tanaka, Motohiro Takeda, Takuji Ishikawa, Ken-ichi Tsubota, Yohsuke Imai, and Takami Yamaguchi. In vitro blood flow in a rectangular pdms microchannel: experimental observations using a confocal micro-piv system. Biomedical microdevices, 10:153-167, 2008. 2 +[19] Yang Liu, Xin Yuan, Jinli Suo, David J Brady, and Qionghai Dai. Rank minimization for snapshot compressive imaging. IEEE transactions on pattern analysis and machine intelligence, 41(12):2990-3006, 2018. 2, 3 +[20] Patrick Llull, Xuejun Liao, Xin Yuan, Jianbo Yang, David Kittle, Lawrence Carin, Guillermo Sapiro, and David J Brady. Coded aperture compressive temporal imaging. Optics express, 21(9):10526-10545, 2013. 2, 3 +[21] LM Lourenco, A Krothapalli, and CA Smith. Particle image velocimetry. In Advances in fluid mechanics measurements, pages 127-199. Springer, 1989. 8 +[22] Julien NP Martel, Lorenz K Mueller, Stephen J Carey, Piotr Dudek, and Gordon Wetzstein. Neural sensors: Learning pixel exposures for hdr imaging and video compressive sensing with programmable sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1642-1653, 2020. 2, 3, 9 +[23] Amir Matin and Xu Wang. Compressive coded rotating mirror camera for high-speed imaging. In Photonics, volume 8, page 34. MDPI, 2021. 2 +[24] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2 +[25] Albert Abraham Michelson. Experimental determination of the velocity of light: made at the US Naval Academy, Annapolis, volume 1. US Nautical Almanac Office, 1880. 2 +[26] Kristina Monakhova, Vi Tran, Grace Kuo, and Laura Waller. Untrained networks for compressive lensless photography. Optics Express, 29(13):20913-20929, 2021. 2, 3 +[27] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 2 +[28] SHR Müller, B Böhm, M Gleißner, R Grzeszik, S Arndt, and A Dreizler. Flow field measurements in an optically accessible, direct-injection spray-guided internal combustion engine using high-speed piv. *Experiments in fluids*, 48:281-290, 2010. 2 + +[29] Gergely Nagymáte and Rita M Kiss. Application of optitrack motion capture systems in human movement analysis: A systematic literature review. *Recent Innovations in Mechatronics*, 5(1):1-9, 2018. 2 +[30] Cindy M Nguyen, Julien NP Martel, and Gordon Wetzstein. Learning spatially varying pixel exposures for motion deblurring. arXiv preprint arXiv:2204.07267, 2022. 3, 9 +[31] Basilio Pueo Ortega and José M Jiménez Olmedo. Application of motion capture technology for sport performance analysis. Retos:rientas en Educación física, deporte y recreación, (32):241-247, 2017. 2 +[32] Travis Portz, Li Zhang, and Hongrui Jiang. Random coded sampling for high-speed hdr video. In IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2013. 2, 3 +[33] Ramesh Raskar, Amit Agrawal, and Jack Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 3 +[34] Dikpal Reddy, Ashok Veeraraghavan, and Rama Chellappa. P2c2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336. IEEE, 2011. 2, 3, 9 +[35] Aswin C, Sankaranarayanan, Pavan K Turaga, Rama Chellappa, and Richard G Baraniuk. Compressive acquisition of linear dynamical systems. SIAM Journal on Imaging Sciences, 6(4):2109-2133, 2013. 2, 3 +[36] Fulvio Scarano. Tomographic piv: principles and practice. Measurement Science and Technology, 24(1):012001, 2012. 2 +[37] Yoav Y Schechner, Shree K Nayar, and Peter N Belhumeur. Multiplexing for optimal lighting. IEEE Transactions on pattern analysis and machine intelligence, 29(8):1339-1354, 2007. 4 +[38] Shubham Sharma, Shubhankar Verma, Mohit Kumar, and Lavanya Sharma. Use of motion capture in 3d animation: motion capture systems, challenges, and recent trends. In 2019 international conference on machine learning, big data, cloud and parallel computing (comitcon), pages 289-294. IEEE, 2019. 2 +[39] Mark Sheinin, Matthew O'Toole, and Srinivasa G Narasimhan. Deconvolving diffraction for fast imaging of sparse scenes. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2021. 2, 3, 4, 5, 8, 9 +[40] Mark Sheinin, Dinesh N Reddy, Matthew O'Toole, and Srinivasa G Narasimhan. Diffraction line imaging. In European Conference on Computer Vision, pages 1-16. Springer, 2020. 2, 3, 4, 8 +[41] Mark Sheinin, Yoav Y Schechner, and Kiriakos N Kutulakos. Computational imaging on the electric grid. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6437-6446, 2017. 2 +[42] H Shen, L Gan, N Newman, Y Dong, C Li, Y Huang, and YC Shen. Spinning disk for compressive imaging. Optics letters, 37(1):46-48, 2012. 2 +[43] H Shen, N Newman, L Gan, SC Zhong, Y Huang, and YC Shen. Compressed terahertz imaging system using a spin + +disk. In 35th International Conference on infrared, millimeter, and terahertz waves, pages 1-2. IEEE, 2010. 2 +[44] Yaoyao Shi, Youwen Liu, Wei Sheng, Jiming Wang, and Tong Wu. Speckle rotation decorrelation based single-shot video through scattering media. Optics express, 27(10):14567-14576, 2019. 2 +[45] Ho Shiraga, M Heya, A Fujishima, O Maegawa, K Shimada, Y Kato, T Yamanaka, and S Nakai. Laser-imploded core structure observed by using two-dimensional x-ray imaging with 10-ps temporal resolution. Review of scientific instruments, 66(1):722-724, 1995. 3 +[46] Masada Tzabari and Yoav Y Schechner. Polarized opticalflow gyroscope. In Proceedings of the European Conference on Computer Vision, pages 363-381. Springer, 2020. 2 +[47] BW Van Oudheusden. Piv-based pressure measurement. Measurement Science and Technology, 24(3):032001, 2013. 2 +[48] Ashok Veeraraghavan, Dikpal Reddy, and Ramesh Raskar. Coded strobing photography: Compressive sensing of high speed periodic videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(4):671-686, 2010. 2, 3 +[49] Zihao Wang, Leonidas Spinoulas, Kuan He, Lei Tian, Oliver Cossairt, Aggelos K Katsaggelos, and Huajin Chen. Compressive holographic video. Optics express, 25(1):250-262, 2017. 2, 3 +[50] Gil Weinberg and Ori Katz. 100,000 frames-per-second compressive imaging with a conventional rolling-shutter camera by random point-spread-function engineering. Optics Express, 28(21):30616-30625, 2020. 2, 3, 4, 8 +[51] Charles Wheatstone. An account of some experiments to measure the velocity of electricity and the duration of electric light. Philosophical Transactions of the Royal Society of London, (124):583-591, 1834. 2 +[52] Jinhui Xiong, Ramzi Idoughi, Andres A Aguirre-Pablo, Abdulrahman B Aljedaani, Xiong Dun, Qiang Fu, Sigurdur T Thoroddsen, and Wolfgang Heidrich. Rainbow particle imaging velocimetry for dense 3d fluid velocity imaging. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. \ No newline at end of file diff --git a/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/images.zip b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..18d1f2cbf2bc72d1ee7acf5828a031703cfa9e87 --- /dev/null +++ b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:608dbddaca2803993806b93a868910ad8388bf4b4d295f0997ccadaca90bc699 +size 461231 diff --git a/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/layout.json b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d5a2d44f8a0eef137bbad434d403215cc9f0fe --- /dev/null +++ b/2023/SpinCam_ High-Speed Imaging via a Rotating Point-Spread Function/layout.json @@ -0,0 +1,11425 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 511, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 511, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 511, + 121 + ], + "type": "text", + "content": "SpinCam: High-Speed Imaging via a Rotating Point-Spread Function" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 172, + 144, + 419, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 144, + 419, + 156 + ], + "spans": [ + { + "bbox": [ + 172, + 144, + 419, + 156 + ], + "type": "text", + "content": "Dorian Chan, Mark Sheinin, and Matthew O'Toole" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 158, + 433, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 158, + 433, + 171 + ], + "spans": [ + { + "bbox": [ + 160, + 158, + 433, + 171 + ], + "type": "text", + "content": "Carnegie Mellon University, Pittsburgh, PA 15213, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 128, + 174, + 465, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 174, + 465, + 185 + ], + "spans": [ + { + "bbox": [ + 128, + 174, + 465, + 185 + ], + "type": "text", + "content": "dychan@andrew.cmu.edu, marksheinin@gmail.com, mpotoole@cmu.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "text", + "content": "High-speed cameras are an indispensable tool used for the slow-motion analysis of scenes. However, the fixed bandwidth of any imaging system quickly becomes a bottleneck, resulting in a fundamental trade-off between the camera's spatial and temporal resolutions. In recent years, compressive high-speed imaging systems have been proposed to circumvent these issues by optically encoding the signal and using a reconstruction procedure to recover a video. Our work proposes a novel approach for compressive high-speed imaging based on temporally coding the camera's point-spread function (PSF). By mechanically spinning a diffraction grating in front of a camera, the sensor integrates an image blurred by a PSF that continuously rotates over time. We also propose a deconvolution-based reconstruction algorithm to reconstruct videos from these measurements. Our method achieves superior light efficiency and handles a wider scene class than prior methods. Also, our mechanical design yields flexible temporal resolution that can be easily increased, potentially allowing capture at " + }, + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "inline_equation", + "content": "192\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "text", + "content": " far higher than prior works. We demonstrate a prototype for various applications, including motion capture and particle image velocimetry (PIV)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": "The world is full of high-speed events that are imperceptible to both the human eye and most conventional cameras (e.g., the flicker of electric lights or the motion of fast-moving objects). The traditional approach to capturing such events is using slow-motion photography, i.e., use a camera to shoot images at a high frame rate and display the results at a lower frame rate. Slow-motion photography is often used for artistic effects in film and as a measurement tool in scientific applications. However, despite significant advancements in sensor technology, conventional high-speed cameras tend to be expensive, bulky, and have high throughput requirements for recording visual content at both high frame rates and high spatial resolutions. These hardware limitations set an upper limit to the cameras' capture rates." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 211, + 533, + 502 + ], + "blocks": [ + { + "bbox": [ + 318, + 211, + 533, + 502 + ], + "lines": [ + { + "bbox": [ + 318, + 211, + 533, + 502 + ], + "spans": [ + { + "bbox": [ + 318, + 211, + 533, + 502 + ], + "type": "image", + "image_path": "637c746f8ed30d0313223c23ab115d9ec8ed05b139f74c8859fef505e38a4bbc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 507, + 547, + 586 + ], + "lines": [ + { + "bbox": [ + 305, + 507, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 507, + 547, + 586 + ], + "type": "text", + "content": "Figure 1. Principle of high-speed imaging with a rotating diffraction grating. A camera captures a long exposure image of a moving scene. During the camera's exposure period, the diffraction grating placed in front of the camera continuously rotates to generate a time-varying point-spread function (PSF) that encodes scene dynamics into a single frame. Solving a deconvolution problem yields a high-speed video from the captured frame." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Compressed sensing provides a framework to circumvent such issues. Since videos are highly compressible, it is possible to recover a video from a single snapshot by (i) compressively recording spatiotemporal visual information onto a sensor and (ii) solving a reconstruction problem to recover the video using priors on the visual content. Among the large body of work in compressive high-speed imaging, most techniques temporally modulate the light incident on sensor pixels [48, 12, 20, 22, 9, 34, 32, 35, 14, 16, 19," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10789" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "13, 49], either through the use of a spatial light modulator (SLM) or novel sensors capable of turning pixels on or off. Alternatively, by using spatial-multiplexing optics (e.g., diffusers, diffraction gratings) to distribute light from any scene point to many camera pixels, it is possible to perform high-speed imaging by using a rolling shutter [2, 50, 26], a reduced region-of-interest (ROI) [39], or a line sensor [40]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 286, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 286, + 301 + ], + "type": "text", + "content": "Our work proposes a fundamentally different approach to compressive high-speed imaging. Similar to past works that make use of spatial-multiplexing optics, we propose to distribute light across a sensor by using a dual-axis diffraction grating, which diffracts light horizontally and vertically (see Fig. 1). Unlike prior work, however, we opt to change the point-spread function (PSF) during the camera's exposure period. Specifically, we use a motor to mechanically rotate the PSF at high speeds, record a single global-shutter frame, and solve a simple deconvolution-based reconstruction procedure to reconstruct scene dynamics. Please refer to Fig. 1 for an overview of the process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 302, + 286, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 286, + 458 + ], + "type": "text", + "content": "Our approach, like past work using spatial-multiplexing optics [2, 50, 40, 39], is primarily targeted towards spatiotemporally sparse scenes. This imaging regime includes motion capture and particle-image velocimetry (PIV), which see broad application in domains like biomechanics [29], sports science [31], animation and film-making [38], virtual and augmented reality [5], aerodynamics [47], fluid mechanics [36], biomedical engineering [18], oceanography [7], and combustion and engine research [28]. We demonstrate our approach for motion capture and PIV in Sec. 6. Other potential applications for spatiotemporally sparse video include high-speed 3D scanning [24] and capturing the fast flicker of electric bulbs [41]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 459, + 286, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 286, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 286, + 483 + ], + "type": "text", + "content": "Our approach combines four key advantages over existing methods targeting spatiotemporally sparse scenes:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 491, + 286, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 286, + 597 + ], + "type": "text", + "content": "(I) Superior light efficiency. In many prior methods, the scene light is dispersed using specialized optics to reach many camera pixels, but only a small subset of camera pixels collect light at each timestamp [2, 40, 39, 50]. Thus, most light is lost by falling outside the 'active' pixels. Conversely, in our work, the signal from different timestamps is encoded by different temporal PSFs, and the entire sensor is left exposed, yielding far superior light efficiency. We show this advantage in simulated experiments in Sec. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "(II) Easily increasable temporal resolution. In most prior works, temporal resolution (i.e., FPS) is fundamentally limited by the speed of the respective technology (e.g., the modulation speed of a SLM/sensor [12, 22, 34, 32, 14, 49], or the read-out speed of a camera [2, 40, 39, 50, 9]). In contrast, the temporal resolution of our approach is determined by the mechanical rotation speed and thus can be easily increased by rotating the PSF faster. Thus, we show that our approach can resolve a LED flickering at " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "96\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "(III) A wider class of supported scenes. Prior works using spatial-multiplexing optics allocate a fixed amount of pixels for each timestamp (e.g., a single rolling shutter row) [2, 40, 39, 50]. This limits the amount of information recoverable per frame, imposing a spatial sparsity constraint on each high-speed frame individually. Instead, our solution multiplexes the signals from different timestamps together, yielding a sparsity constraint over the entire video. Our method can thus handle temporally sparse but spatially dense scenes, verified in simulated and real experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 545, + 275 + ], + "type": "text", + "content": "(IV) Simple hardware. While many prior works require specialized optics (e.g., SLMs [12, 34, 49], piezoelectric stages [20, 16, 19]) or specialized sensors (e.g., coded exposure pixels [22, 14, 32]), our setup can work with any global-shutter camera combined with simple low-cost hardware that is readily available at your local hobby store. We include a DIY guide as part of our supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 281, + 389, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 389, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 389, + 293 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 300, + 500, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 500, + 313 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 500, + 313 + ], + "type": "text", + "content": "2.1. Spinning Disks & Revolving Mirrors" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 319, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 462 + ], + "type": "text", + "content": "Mechanical spinning disks and revolving mirrors have been used to measure transient phenomena for centuries. In 1834, Wheatstone [51] estimated the velocity of electricity traveling through long copper wires by observing the sparks generated at the ends. Using a revolving mirror, Wheatstone observed a " + }, + { + "bbox": [ + 304, + 319, + 545, + 462 + ], + "type": "inline_equation", + "content": "1\\mu \\mathrm{s}" + }, + { + "bbox": [ + 304, + 319, + 545, + 462 + ], + "type": "text", + "content": " time difference, and concluded the velocity of electricity to be 288,000 miles per second. Fizeau and Foucault [25] later measured the speed of light through various apparatuses based on revolving mirrors and spinning cogwheels. For example, in 1848-49, Fizeau measured the transit time of light from a source, to a mirror, and back to the source by using a rotating wheel to modulate the light." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 463, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 665 + ], + "type": "text", + "content": "Though rare in modern consumer electronics, modulating light mechanically with a spinning element is both faster and lower cost than most modern digital modulation solutions. For example, digital micromirror devices (DMDs) only operate up to " + }, + { + "bbox": [ + 304, + 463, + 545, + 665 + ], + "type": "inline_equation", + "content": "30\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 463, + 545, + 665 + ], + "type": "text", + "content": ", whereas spinning disks can offer rates up to " + }, + { + "bbox": [ + 304, + 463, + 545, + 665 + ], + "type": "inline_equation", + "content": "2.4\\mathrm{MHz}" + }, + { + "bbox": [ + 304, + 463, + 545, + 665 + ], + "type": "text", + "content": " [10]. As a result, coded disks have been used for single pixel imaging at rates up to 72 FPS [43, 42, 10], and revolving coded mirrors have been used to reconstruct 1400 frames from a single shot [23]. A spinning camera was used to recover an 11-frame video through scattering media by disambiguating the resulting rotated speckle patterns [44], while a rotating camera coupled with a polarizer generalized the optical-flow equation to include polarization cues [46]. An event camera can be combined with a spinning polarizer for fast, accurate shape-from-polarization [27]. This work explores the use of spinning elements for compressive high-speed video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 671, + 455, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 455, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 455, + 684 + ], + "type": "text", + "content": "2.2. Pixel-wise Coded Exposure" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Coded exposure photography involves temporally modulating the light incident on a sensor, which has long been" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10790" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "used to capture dynamic events [4]. Raskar et al. [33] used a liquid-crystal panel to rapidly open and close the shutter of a camera; the motion blur associated with fast-moving objects could then be reliably removed through deconvolution. This methodology was later extended for recovering full high-speed videos of periodic [48] and generic scenes [13]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 286, + 251 + ], + "type": "text", + "content": "A popular approach to compressive high-speed imaging involves pixel-wise exposure control [12, 9, 34, 32, 35, 22, 14, 20, 16, 19]. In practice, the codes are often binary, and the corresponding reconstruction techniques rely on various priors to reconstruct videos. Example priors include assuming the scene produces periodic motion [48] or satisfies the brightness constancy assumption [34], and might rely on dictionary learning [12, 13] and neural networks [22, 14, 30] to aid the reconstruction procedure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 253, + 286, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 286, + 373 + ], + "type": "text", + "content": "These techniques lack the (II) temporal resolution and (IV) hardware complexity advantages of our approach since they typically require specialized hardware (e.g., SLMs [12, 34, 49], piezoelectric stages [20, 16, 19]) or specialized sensors [22, 14, 32]) which can be expensive, difficult to acquire, and still be limited in frame rate. Moreover, as shown in the supplementary materials, since these methods do not spread light across multiple pixels, these systems struggle with spatially sparse scenes, which appear in many practical applications of high-speed imaging [40, 39]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 381, + 200, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 200, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 200, + 395 + ], + "type": "text", + "content": "2.3. Spatial-Multiplexing Optics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 286, + 557 + ], + "type": "text", + "content": "An alternative approach to capture space-time volumes (videos) involves spatial multiplexing. By placing an optical element in front of a sensor or camera (e.g., a diffuser or diffraction grating), the light from a single point in the scene spreads to multiple camera pixels according to the optical system's PSF. When combined with a rolling shutter camera that samples individual sensor rows at fast rates [2, 50, 26], it is possible to reconstruct high-speed videos. Alternatively, a limited region-of-interest (ROI) or 1D sensor can also be used to sample measurements at fast rates [40, 39]. By integrating diffraction patterns generated by time-varying coded masks, it is even possible to reconstruct 4D volumes of spacetime from a single snapshot [49]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 689 + ], + "type": "text", + "content": "Similar ideas have been used in combination with streak cameras, an optoelectronic device used to measure ultra-fast light phenomena at femtosecond time scales. A streak camera spatially multiplexes the scene's temporal response, by deflecting photoelectrons to different points on a sensor as a function of time. While designed to be a line sensor, it is possible to encode 2D spatial information by incorporating a tilted microlens [45] or pinhole array [11]. In Gao et al. [8], the 2D scene at different timestamps is coded with a fixed mask, sheared a temporally varying amount, and then overlayed onto the 2D signal captured by a streak camera." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": "These methods lack the (I) light efficiency and (III) density advantages of our approach—they typically rely" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 70, + 511, + 167 + ], + "blocks": [ + { + "bbox": [ + 342, + 70, + 511, + 167 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 511, + 167 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 511, + 167 + ], + "type": "image", + "image_path": "ff9e9d3158198e325b52bd2d5f64a26e62377ec5dc4c139f65e95491e242c435.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 178, + 518, + 228 + ], + "blocks": [ + { + "bbox": [ + 386, + 168, + 465, + 178 + ], + "lines": [ + { + "bbox": [ + 386, + 168, + 465, + 178 + ], + "spans": [ + { + "bbox": [ + 386, + 168, + 465, + 178 + ], + "type": "text", + "content": "(a) Captured image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 178, + 518, + 228 + ], + "lines": [ + { + "bbox": [ + 328, + 178, + 518, + 228 + ], + "spans": [ + { + "bbox": [ + 328, + 178, + 518, + 228 + ], + "type": "image", + "image_path": "046888cc7066ba023f0954b21e72779acda91a1bc7415bbea8767ec4fb65a379.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 228, + 488, + 239 + ], + "lines": [ + { + "bbox": [ + 362, + 228, + 488, + 239 + ], + "spans": [ + { + "bbox": [ + 362, + 228, + 488, + 239 + ], + "type": "text", + "content": "(b) Recovered time signal (ms)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 243, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 342 + ], + "type": "text", + "content": "Figure 2. High-speed reconstruction of a flickering LED. The LED in this experiment flashes at " + }, + { + "bbox": [ + 304, + 243, + 545, + 342 + ], + "type": "inline_equation", + "content": "96\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 243, + 545, + 342 + ], + "type": "text", + "content": ", as driven by an Arduino. (a) The diffraction image encodes the high-frequency patterns of the LED (zoom into inset). (b) This temporally-varying signal can be directly extracted from the image to recover the intensity pattern of the LED. A frequency analysis of the recovered time signal indicates the system's ability to resolve frequencies up to " + }, + { + "bbox": [ + 304, + 243, + 545, + 342 + ], + "type": "inline_equation", + "content": "96\\mathrm{kHz}" + }, + { + "bbox": [ + 304, + 243, + 545, + 342 + ], + "type": "text", + "content": ". Given that the Nyquist frequency is half that of the sample rate, the effective framerate for the system is therefore over 192,000 FPS." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 356, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 428 + ], + "type": "text", + "content": "on a small subset of camera pixels to reconstruct each frame, \"wasting\" light and limiting recoverable scene density. Furthermore, their capture speed is limited by the maximum speed of the hardware used (i.e., line-scan mode [40], ROIs [39] or rolling shutter [2, 50]). Thus, these methods lack the ability to (II) easily increase temporal resolution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 437, + 545, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 545, + 465 + ], + "type": "text", + "content": "3. High-Speed Imaging using a Time-Varying Point-Spread Function" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "content": "Consider a scene that consists of a single bright LED in a dark room, where the LED is configured to pulse rapidly over time. Unless it is faster than the blinking LED, a normal camera forms an image of the LED, but provides no information about the LED's temporal dynamics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 653 + ], + "type": "text", + "content": "Our proposed approach to imaging high-speed events involves two steps. First, following prior work on diffraction imaging [40, 39], we place a diffraction grating in front of the camera, which blurs the scene with a rainbow PSF. The light from the LED is therefore spread over many additional camera pixels. Second, we propose to rotate the diffraction grating such that the corresponding PSF changes over time. This produces measurements that spatially encode temporally varying signals, which can be used to analyze extremely high-speed events (see Fig. 2 for an example)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "In this section, we describe the image formation model and reconstruction process for recovering a high-speed video from an image captured with a generic time-varying PSF. We then analyze the special case of using a rotating PSF for capturing images." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10791" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 133, + 125 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 133, + 125 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 133, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 133, + 125 + ], + "type": "image", + "image_path": "01965abe5cea810eac69f31f79cf91612ca3f8b39ee3975e512799c5cde86b82.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 125, + 130, + 140 + ], + "lines": [ + { + "bbox": [ + 62, + 125, + 130, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 125, + 130, + 140 + ], + "type": "text", + "content": "Rolling " + }, + { + "bbox": [ + 62, + 125, + 130, + 140 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 62, + 125, + 130, + 140 + ], + "type": "text", + "content": " diffuser [2] (64 row exposure)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 134, + 70, + 205, + 125 + ], + "blocks": [ + { + "bbox": [ + 134, + 70, + 205, + 125 + ], + "lines": [ + { + "bbox": [ + 134, + 70, + 205, + 125 + ], + "spans": [ + { + "bbox": [ + 134, + 70, + 205, + 125 + ], + "type": "image", + "image_path": "dac4d2807bf373bd873263f60d65e44a9e94fbb0c1535f7847efc409ed2656a7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 125, + 200, + 140 + ], + "lines": [ + { + "bbox": [ + 139, + 125, + 200, + 140 + ], + "spans": [ + { + "bbox": [ + 139, + 125, + 200, + 140 + ], + "type": "text", + "content": "ROI + grating [39] (16 row ROI)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 206, + 70, + 278, + 125 + ], + "blocks": [ + { + "bbox": [ + 206, + 70, + 278, + 125 + ], + "lines": [ + { + "bbox": [ + 206, + 70, + 278, + 125 + ], + "spans": [ + { + "bbox": [ + 206, + 70, + 278, + 125 + ], + "type": "image", + "image_path": "227aa7f4891ec1d7d47abb9c09d833cfb578d1c163967a7c7c09aa1f77ba8a1d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 125, + 264, + 134 + ], + "lines": [ + { + "bbox": [ + 220, + 125, + 264, + 134 + ], + "spans": [ + { + "bbox": [ + 220, + 125, + 264, + 134 + ], + "type": "text", + "content": "Rotating PSF" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 141, + 278, + 202 + ], + "blocks": [ + { + "bbox": [ + 52, + 141, + 278, + 202 + ], + "lines": [ + { + "bbox": [ + 52, + 141, + 278, + 202 + ], + "spans": [ + { + "bbox": [ + 52, + 141, + 278, + 202 + ], + "type": "image", + "image_path": "d1f555643eb4531ba875b6077d6408e451f399a999d3405af70a087574b15c01.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 203, + 263, + 213 + ], + "lines": [ + { + "bbox": [ + 69, + 203, + 263, + 213 + ], + "spans": [ + { + "bbox": [ + 69, + 203, + 263, + 213 + ], + "type": "text", + "content": "(a) Recovered brightness of time-varying light source" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 53, + 222, + 132, + 277 + ], + "blocks": [ + { + "bbox": [ + 53, + 222, + 132, + 277 + ], + "lines": [ + { + "bbox": [ + 53, + 222, + 132, + 277 + ], + "spans": [ + { + "bbox": [ + 53, + 222, + 132, + 277 + ], + "type": "image", + "image_path": "f9c5ed59b216b5477c59963999ea4dadeff50591458faab07bede059eb8a3f80.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 53, + 278, + 132, + 331 + ], + "blocks": [ + { + "bbox": [ + 53, + 278, + 132, + 331 + ], + "lines": [ + { + "bbox": [ + 53, + 278, + 132, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 278, + 132, + 331 + ], + "type": "image", + "image_path": "904dae00d73e8faf386a6baa51ef674475d37a55320431a5a07b391603decf82.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 332, + 131, + 349 + ], + "lines": [ + { + "bbox": [ + 62, + 332, + 131, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 131, + 349 + ], + "type": "text", + "content": "Rolling " + }, + { + "bbox": [ + 62, + 332, + 131, + 349 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 62, + 332, + 131, + 349 + ], + "type": "text", + "content": " diffuser [2] (64 row exposure)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 350, + 233, + 361 + ], + "lines": [ + { + "bbox": [ + 100, + 350, + 233, + 361 + ], + "spans": [ + { + "bbox": [ + 100, + 350, + 233, + 361 + ], + "type": "text", + "content": "(b) Recovery of dense moving scene" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 365, + 287, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 475 + ], + "type": "text", + "content": "Figure 3. Light efficiency experiment. In the presence of Poisson and read noise, we simulated the measurements captured by a Canon 5D Mark III sensor [6] with ISO 800. (a) We simulate a dim point light source with a time-varying intensity profile. (b) We simulate the word \"test\" moving across the scene. In both cases, our approach is able to much more accurately resolve the signal, compared to a rolling shutter and diffuser as well as a ROI and grating. We theorize that this is due to the fact that our approach captures more light for every time step, and that the requisite demultiplexing minimizes the effects of noise." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 134, + 223, + 205, + 277 + ], + "blocks": [ + { + "bbox": [ + 134, + 223, + 205, + 277 + ], + "lines": [ + { + "bbox": [ + 134, + 223, + 205, + 277 + ], + "spans": [ + { + "bbox": [ + 134, + 223, + 205, + 277 + ], + "type": "image", + "image_path": "cd01ad50f52be70cca44dd33940637c2862d8338328822c387761c44d661843b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 134, + 278, + 206, + 331 + ], + "blocks": [ + { + "bbox": [ + 134, + 278, + 206, + 331 + ], + "lines": [ + { + "bbox": [ + 134, + 278, + 206, + 331 + ], + "spans": [ + { + "bbox": [ + 134, + 278, + 206, + 331 + ], + "type": "image", + "image_path": "69ea9149a41dea49faf869542d3ee0c232c807ad2a168c1142ced317ca5e6be0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 332, + 200, + 348 + ], + "lines": [ + { + "bbox": [ + 138, + 332, + 200, + 348 + ], + "spans": [ + { + "bbox": [ + 138, + 332, + 200, + 348 + ], + "type": "text", + "content": "ROI + grating [39] (16 row ROI)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 206, + 223, + 278, + 277 + ], + "blocks": [ + { + "bbox": [ + 206, + 223, + 278, + 277 + ], + "lines": [ + { + "bbox": [ + 206, + 223, + 278, + 277 + ], + "spans": [ + { + "bbox": [ + 206, + 223, + 278, + 277 + ], + "type": "image", + "image_path": "b283a5de6050ff094b2edca34f947b1ea19c92d6c60f6ed2a64f33ee79020297.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 207, + 278, + 279, + 331 + ], + "blocks": [ + { + "bbox": [ + 207, + 278, + 279, + 331 + ], + "lines": [ + { + "bbox": [ + 207, + 278, + 279, + 331 + ], + "spans": [ + { + "bbox": [ + 207, + 278, + 279, + 331 + ], + "type": "image", + "image_path": "136f9d6de205a41cc5d13d5135ed0a542751661873120b57edd42202f53a49ff.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 332, + 264, + 341 + ], + "lines": [ + { + "bbox": [ + 220, + 332, + 264, + 341 + ], + "spans": [ + { + "bbox": [ + 220, + 332, + 264, + 341 + ], + "type": "text", + "content": "Rotating PSF" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 498, + 184, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 184, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 184, + 511 + ], + "type": "text", + "content": "3.1. Image Formation Model" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 518, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 287, + 543 + ], + "type": "text", + "content": "As illustrated in Fig. 1, for a camera where the PSF changes with time, the forward model can be written as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 79, + 556, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 556, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 79, + 556, + 287, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {I} (x, y) = \\int_ {t = t _ {s}} ^ {t = t _ {s} + E} (\\mathbf {k} (t) * \\mathbf {s} (t)) (x, y) d t, \\tag {1}", + "image_path": "f833aa560d5e76e293f799f299221f25a8f98ba63e6f8371911c1b12134c5f55.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(x,y)" + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": " is the captured image, " + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": " denotes the exposure time of the camera, and " + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": " denotes the time when the camera started exposing. The symbol " + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": " denotes the spatial convolution operator between the scene " + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{s}(t)" + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": " and the PSF " + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{k}(t)" + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": ". When discretized to " + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "inline_equation", + "content": "N_E" + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": " time steps from " + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "inline_equation", + "content": "t_s + E" + }, + { + "bbox": [ + 46, + 596, + 287, + 668 + ], + "type": "text", + "content": ", this can be rewritten in matrix form as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 111, + 681, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 681, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 111, + 681, + 287, + 693 + ], + "type": "interline_equation", + "content": "\\mathbf {I} = \\mathbf {M} \\mathbf {s} \\tag {2}", + "image_path": "501a4094362669882a6c8aa18060d2f194566f16f3506561fb3b59091f4e403c.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 696, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 696, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\left[ \\begin{array}{l l l l} \\mathbf {C} _ {1} & \\mathbf {C} _ {2} & \\dots & \\mathbf {C} _ {N _ {E}} \\end{array} \\right] \\tag {3}", + "image_path": "125a26e4298ea62ad6e681c40d06a1aca1858c3e4939987698e55ad0b38a9bba.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}_{+}^{N_{x}N_{y}}" + }, + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "type": "text", + "content": " is the vectorized representation of the captured image, " + }, + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}_{+}^{N_{x}N_{y}N_{E}}" + }, + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "type": "text", + "content": " is the vectorized representation of the high-speed scene, and " + }, + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_k \\in \\mathbb{R}_{+}^{N_xN_y\\times N_xN_y}" + }, + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "type": "text", + "content": " represents the convolution with the kernel at timestep " + }, + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 70, + 545, + 126 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 133, + 437, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 437, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 437, + 144 + ], + "type": "text", + "content": "3.2. Reconstruction Process" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 151, + 545, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 151, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 305, + 151, + 545, + 211 + ], + "type": "text", + "content": "Natural scenes contain sparse gradients, and many target applications for high-speed imaging require the tracking of just sparse points. As a result, we recover a high-speed video with an " + }, + { + "bbox": [ + 305, + 151, + 545, + 211 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 305, + 151, + 545, + 211 + ], + "type": "text", + "content": " sparsity prior and an anisotropic total variation prior, solving the following optimization problem:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 217, + 545, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 545, + 240 + ], + "type": "interline_equation", + "content": "\\underset {\\mathbf {x}} {\\operatorname {a r g m i n}} \\quad \\| \\mathbf {I} - \\mathbf {M} \\mathbf {x} \\| ^ {2} + \\mathcal {I} _ {\\mathbb {R} _ {+}} (\\mathbf {x}) + \\lambda_ {\\text {s p a r s i t y}} \\| \\mathbf {x} \\| _ {1} + \\tag {4}", + "image_path": "593ed64a676e37e665d10fe38918893058d3d607fd680defef21153876dcaf51.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 354, + 241, + 525, + 254 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 241, + 525, + 254 + ], + "spans": [ + { + "bbox": [ + 354, + 241, + 525, + 254 + ], + "type": "interline_equation", + "content": "\\lambda_ {d x} \\| \\mathbf {D} _ {x} \\mathbf {x} \\| _ {1} + \\lambda_ {d y} \\| \\mathbf {D} _ {y} \\mathbf {x} \\| _ {1} + \\lambda_ {d t} \\| \\mathbf {D} _ {t} \\mathbf {x} \\| _ {1}", + "image_path": "ca04d31290d512739ccaa56edf25b4bd40ac6c66e59f713bfd528e9c4113032f.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathbb{R}_+}" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": " denotes a non-negativity prior, " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_x" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_y" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_t" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": " compute forward finite differences along the " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": " dimensions respectively of the high-speed video " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\lambda_{sparsity}" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\lambda_{dx}" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\lambda_{dy}" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\lambda_{dt}" + }, + { + "bbox": [ + 306, + 260, + 545, + 309 + ], + "type": "text", + "content": " weigh the regularization terms." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 316, + 368, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 368, + 328 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 368, + 328 + ], + "type": "text", + "content": "3.3. Analysis" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 334, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 545, + 369 + ], + "type": "text", + "content": "Varying the PSF over time results in fundamentally different properties compared to prior approaches, which we explore in the rest of this section." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 373, + 546, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 625 + ], + "type": "text", + "content": "Higher light efficiency. Our method, like other prior compressive video works [2, 50, 40, 39], relies on PSFs with broad spatial supports. A broad spatial support spreads the signal from a single point across multiple pixels. Recent approaches record just a subset of these pixels at each timestep [2, 50, 40, 39]. Because fewer pixels encode each frame and light is lost, assuming a standard noise model with Poisson shot noise and additive Gaussian noise, the reconstruction's signal-to-noise ratio is proportionately decreased. In contrast, under a time-varying PSF, the entire frame is always recorded, so a much larger fraction of the PSF (and therefore light) is captured for every scene point during capture. In addition, the signals from different timestamps are mixed together in the final capture, which could improve the quality of the reconstructions under read noise [37]. Thus, this overall increase in light efficiency leads to better performance in resolving noisy scenes. We show this experimentally in Fig. 3 on two simulated cases: a point light source with time-varying intensity, and a denser moving scene. We find that our approach robustly recovers these two scenes, even under low light conditions." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "A wider class of scenes. Many past approaches achieve high-speeds by allocating a number of pixels per frames that is much lower than the total number of frame pixels [2, 50, 40, 39]. This causes each frame to undergo lossy compression, yielding a sparsity constraint per frame. For example, for a rolling shutter with exposure set to the row rate, each frame is encoded onto just a single sensor row." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10792" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 80, + 123, + 262 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 136, + 80 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 136, + 80 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 136, + 80 + ], + "type": "text", + "content": "(a) Simulated scene:" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 80, + 123, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 80, + 123, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 80, + 123, + 262 + ], + "type": "image", + "image_path": "d432968d16ec4e19a92459dc808b633e0ebdb7bcc216f8a5d04d03ebe5c29026.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 262, + 119, + 270 + ], + "lines": [ + { + "bbox": [ + 65, + 262, + 119, + 270 + ], + "spans": [ + { + "bbox": [ + 65, + 262, + 119, + 270 + ], + "type": "text", + "content": "(1) Ground truth" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 137, + 80, + 217, + 140 + ], + "blocks": [ + { + "bbox": [ + 137, + 80, + 217, + 140 + ], + "lines": [ + { + "bbox": [ + 137, + 80, + 217, + 140 + ], + "spans": [ + { + "bbox": [ + 137, + 80, + 217, + 140 + ], + "type": "image", + "image_path": "d000aec2211483a15a7d27bbd0fa8d47bb9f062a13a5751925af30ad2830c416.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 141, + 217, + 201 + ], + "blocks": [ + { + "bbox": [ + 136, + 141, + 217, + 201 + ], + "lines": [ + { + "bbox": [ + 136, + 141, + 217, + 201 + ], + "spans": [ + { + "bbox": [ + 136, + 141, + 217, + 201 + ], + "type": "image", + "image_path": "a5e8b014d322cfd0ddde00e5e429c73463001a349f8df40f49b226dc96552969.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 136, + 201, + 217, + 262 + ], + "blocks": [ + { + "bbox": [ + 136, + 201, + 217, + 262 + ], + "lines": [ + { + "bbox": [ + 136, + 201, + 217, + 262 + ], + "spans": [ + { + "bbox": [ + 136, + 201, + 217, + 262 + ], + "type": "image", + "image_path": "e901f427caf1a3fd9567a48f45e8ee960a9650bc3d26da70af3e9131e9862719.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 262, + 207, + 279 + ], + "lines": [ + { + "bbox": [ + 137, + 262, + 207, + 279 + ], + "spans": [ + { + "bbox": [ + 137, + 262, + 207, + 279 + ], + "type": "text", + "content": "(2) Rolling + diffuser (1 row exposure)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 217, + 80, + 297, + 140 + ], + "blocks": [ + { + "bbox": [ + 217, + 80, + 297, + 140 + ], + "lines": [ + { + "bbox": [ + 217, + 80, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 217, + 80, + 297, + 140 + ], + "type": "image", + "image_path": "0e24694fe7297650688b2d64d41043775fcb781b46597a398428cca98e4a58cc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 217, + 141, + 298, + 201 + ], + "blocks": [ + { + "bbox": [ + 217, + 141, + 298, + 201 + ], + "lines": [ + { + "bbox": [ + 217, + 141, + 298, + 201 + ], + "spans": [ + { + "bbox": [ + 217, + 141, + 298, + 201 + ], + "type": "image", + "image_path": "188ddd3c8f10b71fadba1178662df2b08f59c5da74d172088e11b5e9c77d61fa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 217, + 201, + 298, + 262 + ], + "blocks": [ + { + "bbox": [ + 217, + 201, + 298, + 262 + ], + "lines": [ + { + "bbox": [ + 217, + 201, + 298, + 262 + ], + "spans": [ + { + "bbox": [ + 217, + 201, + 298, + 262 + ], + "type": "image", + "image_path": "c5c869ffb06cc00be5774852da6f32a5a0937c4809d2ece88c28517a0edcd158.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 262, + 298, + 279 + ], + "lines": [ + { + "bbox": [ + 228, + 262, + 298, + 279 + ], + "spans": [ + { + "bbox": [ + 228, + 262, + 298, + 279 + ], + "type": "text", + "content": "Rolling " + }, + { + "bbox": [ + 228, + 262, + 298, + 279 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 228, + 262, + 298, + 279 + ], + "type": "text", + "content": " diffuser [2] (64 row exposure)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 299, + 80, + 377, + 140 + ], + "blocks": [ + { + "bbox": [ + 299, + 80, + 377, + 140 + ], + "lines": [ + { + "bbox": [ + 299, + 80, + 377, + 140 + ], + "spans": [ + { + "bbox": [ + 299, + 80, + 377, + 140 + ], + "type": "image", + "image_path": "1e3d10399923c49f903784a620cb3889e2b8cbf3b6e9f9363104a08924b03329.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 299, + 141, + 378, + 201 + ], + "blocks": [ + { + "bbox": [ + 299, + 141, + 378, + 201 + ], + "lines": [ + { + "bbox": [ + 299, + 141, + 378, + 201 + ], + "spans": [ + { + "bbox": [ + 299, + 141, + 378, + 201 + ], + "type": "image", + "image_path": "966659a7743685c7cc77b46b2eb15e298a82dfe1ac458d24ed07b7689d092108.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 299, + 201, + 378, + 262 + ], + "blocks": [ + { + "bbox": [ + 299, + 201, + 378, + 262 + ], + "lines": [ + { + "bbox": [ + 299, + 201, + 378, + 262 + ], + "spans": [ + { + "bbox": [ + 299, + 201, + 378, + 262 + ], + "type": "image", + "image_path": "74d0b8820f71c4f8406782d741b6dd9b9a47bfc9ad845e7262c7dea3c6eb58f6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 262, + 374, + 279 + ], + "lines": [ + { + "bbox": [ + 302, + 262, + 374, + 279 + ], + "spans": [ + { + "bbox": [ + 302, + 262, + 374, + 279 + ], + "type": "text", + "content": "(4) ROI + grating [39] (2 row ROI)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 379, + 80, + 459, + 140 + ], + "blocks": [ + { + "bbox": [ + 379, + 80, + 459, + 140 + ], + "lines": [ + { + "bbox": [ + 379, + 80, + 459, + 140 + ], + "spans": [ + { + "bbox": [ + 379, + 80, + 459, + 140 + ], + "type": "image", + "image_path": "8f7d9239f844fdbce4e14491ddf9fbf7c38101051b7b1030e26f9fb0ae5f2670.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 379, + 141, + 459, + 201 + ], + "blocks": [ + { + "bbox": [ + 379, + 141, + 459, + 201 + ], + "lines": [ + { + "bbox": [ + 379, + 141, + 459, + 201 + ], + "spans": [ + { + "bbox": [ + 379, + 141, + 459, + 201 + ], + "type": "image", + "image_path": "e9979480d83129dcf248ebe41276a2f3f9aca2182a4a3ddbe262f6db4833d8d1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 379, + 201, + 459, + 262 + ], + "blocks": [ + { + "bbox": [ + 379, + 201, + 459, + 262 + ], + "lines": [ + { + "bbox": [ + 379, + 201, + 459, + 262 + ], + "spans": [ + { + "bbox": [ + 379, + 201, + 459, + 262 + ], + "type": "image", + "image_path": "c852d951ceffc07a5f74246792081b7450de61742ca0d5859c9747dcfdb71d10.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 262, + 455, + 279 + ], + "lines": [ + { + "bbox": [ + 383, + 262, + 455, + 279 + ], + "spans": [ + { + "bbox": [ + 383, + 262, + 455, + 279 + ], + "type": "text", + "content": "(5) ROI + grating [39] (16 row ROI)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 460, + 80, + 540, + 140 + ], + "blocks": [ + { + "bbox": [ + 460, + 80, + 540, + 140 + ], + "lines": [ + { + "bbox": [ + 460, + 80, + 540, + 140 + ], + "spans": [ + { + "bbox": [ + 460, + 80, + 540, + 140 + ], + "type": "image", + "image_path": "9219cebf810a69b0003e1f37155522c0122e55709aae4bf07eeac749b38dcff9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 460, + 141, + 540, + 201 + ], + "blocks": [ + { + "bbox": [ + 460, + 141, + 540, + 201 + ], + "lines": [ + { + "bbox": [ + 460, + 141, + 540, + 201 + ], + "spans": [ + { + "bbox": [ + 460, + 141, + 540, + 201 + ], + "type": "image", + "image_path": "1b2c0d35f239a6724e1fc2ffec104e67f35c9ac041f9c6ea0a519b68148ab01a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 460, + 201, + 540, + 262 + ], + "blocks": [ + { + "bbox": [ + 460, + 201, + 540, + 262 + ], + "lines": [ + { + "bbox": [ + 460, + 201, + 540, + 262 + ], + "spans": [ + { + "bbox": [ + 460, + 201, + 540, + 262 + ], + "type": "image", + "image_path": "f1f35ff8cf4a4c469b618a90fb3f5c9d9cc9eb34f4a1c0e25420ad32903fd54c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 472, + 262, + 527, + 271 + ], + "lines": [ + { + "bbox": [ + 472, + 262, + 527, + 271 + ], + "spans": [ + { + "bbox": [ + 472, + 262, + 527, + 271 + ], + "type": "text", + "content": "(6) Rotating PSF" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 50, + 290, + 171, + 379 + ], + "blocks": [ + { + "bbox": [ + 49, + 280, + 112, + 290 + ], + "lines": [ + { + "bbox": [ + 49, + 280, + 112, + 290 + ], + "spans": [ + { + "bbox": [ + 49, + 280, + 112, + 290 + ], + "type": "text", + "content": "(b) Real scene:" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 290, + 171, + 379 + ], + "lines": [ + { + "bbox": [ + 50, + 290, + 171, + 379 + ], + "spans": [ + { + "bbox": [ + 50, + 290, + 171, + 379 + ], + "type": "image", + "image_path": "485458d0fed3a27c6c7e34ef794be5a2f206169e3fafd6e9b327340b96201c77.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 380, + 151, + 389 + ], + "lines": [ + { + "bbox": [ + 70, + 380, + 151, + 389 + ], + "spans": [ + { + "bbox": [ + 70, + 380, + 151, + 389 + ], + "type": "text", + "content": "(1) Reference frame" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 395, + 545, + 515 + ], + "lines": [ + { + "bbox": [ + 46, + 395, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 545, + 515 + ], + "type": "text", + "content": "Figure 4. Reconstructing temporally sparse, but spatially dense scenes. (a) A simulated scene of two people dancing. The video is dark for the entire sequence, except for a short sequence of fast-moving frames where the dancers are strobed. (a2)-(a3) We implement a setup based on a dual rolling shutter combined with a diffuser [2]. In (a2), with the exposure time set to match the row rate of the shutter, the reconstruction is unable to well-resolve the dense scene. In (a3), with the exposure time set to 64 times the previous value, the reconstruction is able to recover the dancers, at the cost of temporal resolution [2]. (a4)-(a5) We implement an approach based on fast temporal sampling by using a ROI and a diffraction grating [39]. In (a4), the ROI consists of just two rows, resulting in low quality results. In (a5), by using more bandwidth and exposing 16 rows, such an approach can get better results, but this reduces the recoverable framerate by 8 times in practice. (a6) In contrast, our methodology recovers the dense scene much more accurately compared to (a2) and (a4)—the signal is encoded over a much larger portion of sensor. This does not require increasing the exposure time and therefore decreasing the temporal resolution as in (a3), or increasing the requisite bandwidth (a5). (b) A real world scene, where a spinning fan is strobed twice during the camera exposure at unknown times. Our approach is able to disambiguate the fan rotation at each strobe time." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 173, + 290, + 295, + 379 + ], + "blocks": [ + { + "bbox": [ + 173, + 290, + 295, + 379 + ], + "lines": [ + { + "bbox": [ + 173, + 290, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 173, + 290, + 295, + 379 + ], + "type": "image", + "image_path": "507f6def4fc57260c121ca72a3cdece98ed1569d392c0a7b3e527274bb46bc8d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 380, + 290, + 389 + ], + "lines": [ + { + "bbox": [ + 179, + 380, + 290, + 389 + ], + "spans": [ + { + "bbox": [ + 179, + 380, + 290, + 389 + ], + "type": "text", + "content": "(2) Rotation-encoded frame" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 296, + 290, + 419, + 379 + ], + "blocks": [ + { + "bbox": [ + 296, + 290, + 419, + 379 + ], + "lines": [ + { + "bbox": [ + 296, + 290, + 419, + 379 + ], + "spans": [ + { + "bbox": [ + 296, + 290, + 419, + 379 + ], + "type": "image", + "image_path": "60446f80bb913cb5584e00b4b3a839b7aeb90edf4e3173d8f69666cee91bcd9e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 380, + 462, + 389 + ], + "lines": [ + { + "bbox": [ + 375, + 380, + 462, + 389 + ], + "spans": [ + { + "bbox": [ + 375, + 380, + 462, + 389 + ], + "type": "text", + "content": "(3) Recovered frames" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 536, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 643 + ], + "type": "text", + "content": "In contrast, our method allocates the entire frame to encode a plurality of video frames belonging to a fixed time interval. Therefore, our method assumes sparsity in a more general sense, namely sparsity over the entire fixed time interval. The encoded fixed time interval can be either spatially sparse and temporally dense, which fits most prior-work assumptions, or temporally sparse but spatially dense, which prior approaches do not consider due to the fixed bitrate per frame imposed by their low-dimensional encoding." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "To illustrate how this is an advantage, consider the case of a dense scene that is entirely dark, except for a single time instant with a flash of light. If sampling the scene using a rolling-shutter or line sensor, the flashed frame will be encoded onto just a single row. In contrast, a time-varying" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 536, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 571 + ], + "type": "text", + "content": "PSF encodes the flashed frame on the entire frame, wasting no resources encoding the dark frames. Therefore, in this case, a time-varying PSF enables a better reconstruction." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 572, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 545, + 643 + ], + "type": "text", + "content": "As shown in Fig. 4, handling temporally sparse, but spatially dense scenes makes our approach very useful under strobed lighting. As shown in real and simulated experiments, a rotating PSF is able to resolve the content within each individual strobe at high spatiotemporal resolution, without any constraints on scene structure between strobes." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 656, + 430, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 430, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 430, + 670 + ], + "type": "text", + "content": "4. Using a Rotating PSF" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Sec. 3 described the encoding and reconstruction of high-speed video using a time-varying PSF. While a time-varying PSF can be implemented in a number of ways," + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10793" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 286, + 200 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 286, + 200 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 286, + 200 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 286, + 200 + ], + "type": "image", + "image_path": "54be67f8effe5d6f1fd305e1f872fa462dbb6fb9250e5d040959abc2cc5abeae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 289, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 291 + ], + "type": "text", + "content": "Figure 5. The PSF and spatial frequency coding of time. A rotating PSF effectively codes the signal from different timesteps into a rotating set of spatial frequencies in the captured image. For instance, a rotating line PSF encodes different timestamps into different lines in Fourier space. The star PSF from a double-axis diffraction grating behaves as a mixture of the point and line cases. Its Fourier transform contains bright line components that provide temporal signal, but remains broadband for spatial resolution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "text", + "content": "for the sake of cost and speed, we simply rotate a PSF-generating optical element over the course of the camera exposure. This results in a rotation of the corresponding PSF. In this particular case, " + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{k}(t)" + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "text", + "content": " can now be rewritten as " + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{k}(t) = \\mathbf{R}(\\mathbf{k},\\theta_t)" + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "text", + "content": " is the stationary PSF of the rotating optical element, " + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "text", + "content": " is the angle of the element at time " + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\mathbf{k},\\theta_t)" + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "text", + "content": " denotes a rotation of " + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 46, + 312, + 287, + 409 + ], + "type": "text", + "content": ". In the rest of this section, we analyze the properties of such a rotating setup." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "text", + "content": "PSF performance. The selected PSF affects the spatiotemporal reconstruction quality of every video frame. Because each time step is convolved with a rotated PSF, the PSF's rotated Fourier transform determines which parts of the spectrum will encode each time step. To illustrate, an ideal thin lens having a perfect impulse PSF will have perfect spatial resolution because its Fourier transform contributes values in all spectrum frequencies (Fig. 5(a)). It will, however, have poor temporal resolution because every other timestamp will also contribute values across the entire spectrum, making them impossible to distinguish." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 546, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 666 + ], + "type": "text", + "content": "In contrast, consider a line PSF (Fig. 5(b)). In the Fourier domain, the PSF remains a line that goes through the origin. When the line PSF rotates, its Fourier spectrum rotates correspondingly, such that each time step is encoded into a different line in Fourier space, without interfering with values from previous timestamps. As a result, the signal at each time step can be easily reconstructed simply by extracting its corresponding line from Fourier space. However, because only a single direction of spatial frequencies is available for reconstruction for every time step, the re" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 429, + 201 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 429, + 201 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 429, + 201 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 429, + 201 + ], + "type": "image", + "image_path": "7efc04c745edaac170df2bcc64f3b366a07f6fd5eace08118a557c7e0ead0e9f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 204, + 388, + 215 + ], + "lines": [ + { + "bbox": [ + 350, + 204, + 388, + 215 + ], + "spans": [ + { + "bbox": [ + 350, + 204, + 388, + 215 + ], + "type": "text", + "content": "(a) Setup" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 431, + 84, + 550, + 204 + ], + "blocks": [ + { + "bbox": [ + 431, + 84, + 550, + 204 + ], + "lines": [ + { + "bbox": [ + 431, + 84, + 550, + 204 + ], + "spans": [ + { + "bbox": [ + 431, + 84, + 550, + 204 + ], + "type": "image", + "image_path": "56dc7d3ed5f2f7dcb5744e972b3c553fbebb00b03fa4ece179945021f1b9090b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 204, + 526, + 214 + ], + "lines": [ + { + "bbox": [ + 455, + 204, + 526, + 214 + ], + "spans": [ + { + "bbox": [ + 455, + 204, + 526, + 214 + ], + "type": "text", + "content": "(b) Captured PSF" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 220, + 545, + 252 + ], + "lines": [ + { + "bbox": [ + 306, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 545, + 252 + ], + "type": "text", + "content": "Figure 6. Image of our hardware setup, consisting of a global-shutter camera and a circular diffraction grating attached to a 20,000 RPM motor." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 274, + 537, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 537, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 537, + 285 + ], + "type": "text", + "content": "sulting recovered frame will have poor spatial resolution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 370 + ], + "type": "text", + "content": "Inspired by these two cases, we use a double-axis diffraction grating for the experiments shown in this work, whose corresponding PSF is shown in Fig. 6(b). This star PSF contains a bright DC spot at the center, intuitively providing spatial resolution. Simultaneously, the PSF also contains line-shaped rainbow lobes that provide temporal information. We visualize this phenomena in Fig. 5(c)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 380, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 546, + 498 + ], + "type": "text", + "content": "Temporal resolution. The temporal resolution of our system is fundamentally determined by the amount of temporal variation of the PSF between different timestamps, or namely, the minimum angle required to reliably distinguish between two different rotations of the PSF. As a result, the temporal resolution can be increased either by increasing the pixel resolution, the spatial spread of the PSF, or the rotation speed. But since our grating PSF is roughly " + }, + { + "bbox": [ + 304, + 380, + 546, + 498 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 380, + 546, + 498 + ], + "type": "text", + "content": " periodic, rotating more than " + }, + { + "bbox": [ + 304, + 380, + 546, + 498 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 380, + 546, + 498 + ], + "type": "text", + "content": " during the exposure causes temporal ambiguity and degrades the reconstruction quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 499, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 545, + 548 + ], + "type": "text", + "content": "In the supplement, we further analyze PSF performance from the mutual coherence perspective. We also discuss how a time-varying PSF behaves as a form of coded exposure in frequency, as shown by the above Fourier analysis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 561, + 403, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 403, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 403, + 574 + ], + "type": "text", + "content": "5. Implementation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "Hardware. An image of our hardware setup is given in Fig. 6. To generate a usable PSF, we cut a 531-lines per millimeter double-axis diffraction grating sheet into a circular disk of radius 6 inches. To ensure that the sheet stayed flat, we taped it onto a laser-cut acrylic disk with a radius of 6 inches. We attached this element to a Puly DC 775 Motor with a shaft coupling to rotate it. We place this entire rotating setup " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "5\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " in front of the camera's lens. We used a UI-3070CP-C-HQ R2 camera, fitted with a " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "9\\mathrm{mm}f / 1.4" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " lens (Fujinon HF9HA). We also placed a second identical camera beside our system to capture intermittent reference" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "text", + "content": "1The Fourier transform of the convolution of two functions is the element-wise product of their Fourier transforms. If a function is rotated by " + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "text", + "content": ", its Fourier transform is the Fourier transform of the original function, rotated by " + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 673, + 287, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10794" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 175, + 131 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 175, + 131 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 175, + 131 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 175, + 131 + ], + "type": "image", + "image_path": "562699de3ebf26737e1e23971575f024f3aae293a3bcc6de179811241926ceb0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 54, + 132, + 175, + 192 + ], + "blocks": [ + { + "bbox": [ + 54, + 132, + 175, + 192 + ], + "lines": [ + { + "bbox": [ + 54, + 132, + 175, + 192 + ], + "spans": [ + { + "bbox": [ + 54, + 132, + 175, + 192 + ], + "type": "image", + "image_path": "546cc8099894e51416c6cb1002c77f4b397b3be5019dba3d88882d77ba0b9188.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 54, + 194, + 174, + 269 + ], + "blocks": [ + { + "bbox": [ + 54, + 194, + 174, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 194, + 174, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 194, + 174, + 269 + ], + "type": "image", + "image_path": "2f2199849286190feb69e57e6c590b79331dbe38f5ea07da03de5ebd50c4b6a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 54, + 270, + 174, + 344 + ], + "blocks": [ + { + "bbox": [ + 54, + 270, + 174, + 344 + ], + "lines": [ + { + "bbox": [ + 54, + 270, + 174, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 174, + 344 + ], + "type": "image", + "image_path": "fc714076934e57c7a33a400aac193490aa16c1ded52d8970146f8a2c921ac059.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 54, + 345, + 174, + 420 + ], + "blocks": [ + { + "bbox": [ + 54, + 345, + 174, + 420 + ], + "lines": [ + { + "bbox": [ + 54, + 345, + 174, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 345, + 174, + 420 + ], + "type": "image", + "image_path": "0f8d12569c3902f258b4c1f56c0356796152cfbe763414fa468f2221eaca6045.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 421, + 165, + 431 + ], + "lines": [ + { + "bbox": [ + 64, + 421, + 165, + 431 + ], + "spans": [ + { + "bbox": [ + 64, + 421, + 165, + 431 + ], + "type": "text", + "content": "(a) Reference start frame" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 295, + 131 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 295, + 131 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 295, + 131 + ], + "type": "image", + "image_path": "ec60926cb0bb1ca8ae17254120ec9bebf3eff312c7abebacd20673bf572706e1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 176, + 132, + 294, + 192 + ], + "blocks": [ + { + "bbox": [ + 176, + 132, + 294, + 192 + ], + "lines": [ + { + "bbox": [ + 176, + 132, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 176, + 132, + 294, + 192 + ], + "type": "image", + "image_path": "beda0f73715a8a296e45cbdc2e0623d789a4383bc67f4cf51051a20f8e1ea869.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 176, + 193, + 294, + 269 + ], + "blocks": [ + { + "bbox": [ + 176, + 193, + 294, + 269 + ], + "lines": [ + { + "bbox": [ + 176, + 193, + 294, + 269 + ], + "spans": [ + { + "bbox": [ + 176, + 193, + 294, + 269 + ], + "type": "image", + "image_path": "e14fe5640c9ef078c6b81089c755ae7fac784f18c3570cb4bb8164d173a5629b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 176, + 270, + 294, + 344 + ], + "blocks": [ + { + "bbox": [ + 176, + 270, + 294, + 344 + ], + "lines": [ + { + "bbox": [ + 176, + 270, + 294, + 344 + ], + "spans": [ + { + "bbox": [ + 176, + 270, + 294, + 344 + ], + "type": "image", + "image_path": "85d9be2171d891c0efa929a139ca1d822d87dcd224504e5dea29b2bd4a57203d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 176, + 345, + 294, + 420 + ], + "blocks": [ + { + "bbox": [ + 176, + 345, + 294, + 420 + ], + "lines": [ + { + "bbox": [ + 176, + 345, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 176, + 345, + 294, + 420 + ], + "type": "image", + "image_path": "034645de103e544372b6ec03c146fcd76d0254da7b66b058fe607edd29c740c9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 421, + 285, + 430 + ], + "lines": [ + { + "bbox": [ + 186, + 421, + 285, + 430 + ], + "spans": [ + { + "bbox": [ + 186, + 421, + 285, + 430 + ], + "type": "text", + "content": "(b) Reference end frame" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 416, + 131 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 416, + 131 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 416, + 131 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 416, + 131 + ], + "type": "image", + "image_path": "84bed79ee189f771c0457c284eaa6dc5bd6847eacd36b35ec7923973f01823a7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 296, + 132, + 416, + 192 + ], + "blocks": [ + { + "bbox": [ + 296, + 132, + 416, + 192 + ], + "lines": [ + { + "bbox": [ + 296, + 132, + 416, + 192 + ], + "spans": [ + { + "bbox": [ + 296, + 132, + 416, + 192 + ], + "type": "image", + "image_path": "5f522b3fce1e253e05adc65308b19972809c54ffc19fd6a3e32abb0dcc1c9b59.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 296, + 193, + 416, + 269 + ], + "blocks": [ + { + "bbox": [ + 296, + 193, + 416, + 269 + ], + "lines": [ + { + "bbox": [ + 296, + 193, + 416, + 269 + ], + "spans": [ + { + "bbox": [ + 296, + 193, + 416, + 269 + ], + "type": "image", + "image_path": "6d1211679be31115af89373bd99fa07b6a9c06a141f34af32cffe651132a6ef1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 296, + 270, + 416, + 344 + ], + "blocks": [ + { + "bbox": [ + 296, + 270, + 416, + 344 + ], + "lines": [ + { + "bbox": [ + 296, + 270, + 416, + 344 + ], + "spans": [ + { + "bbox": [ + 296, + 270, + 416, + 344 + ], + "type": "image", + "image_path": "9cf22ae89555790b2d9b9817e36b1ddb49a4203e24aeff90aad7962f2ad65352.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 296, + 345, + 416, + 420 + ], + "blocks": [ + { + "bbox": [ + 296, + 345, + 416, + 420 + ], + "lines": [ + { + "bbox": [ + 296, + 345, + 416, + 420 + ], + "spans": [ + { + "bbox": [ + 296, + 345, + 416, + 420 + ], + "type": "image", + "image_path": "2e9d4b45b41132e11641964f2f33f04a564b8fac590a340215d708ad31f94413.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 300, + 421, + 412, + 431 + ], + "lines": [ + { + "bbox": [ + 300, + 421, + 412, + 431 + ], + "spans": [ + { + "bbox": [ + 300, + 421, + 412, + 431 + ], + "type": "text", + "content": "(c) Rotation-encoded frame" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 435, + 547, + 469 + ], + "lines": [ + { + "bbox": [ + 46, + 435, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 547, + 469 + ], + "type": "text", + "content": "Figure 7. High-speed reconstruction of fast moving objects. Row 1: A ball with a black stripe bounces and spins. Row 2: An I rotates " + }, + { + "bbox": [ + 46, + 435, + 547, + 469 + ], + "type": "inline_equation", + "content": "120^{\\circ}" + }, + { + "bbox": [ + 46, + 435, + 547, + 469 + ], + "type": "text", + "content": ". Row 3: A toy bullet bouncing off a board. Row 4, 5: Our setup is capable of tracking the movements of both sparse points and dense marker lines for high-speed motion capture." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 417, + 71, + 537, + 131 + ], + "blocks": [ + { + "bbox": [ + 417, + 71, + 537, + 131 + ], + "lines": [ + { + "bbox": [ + 417, + 71, + 537, + 131 + ], + "spans": [ + { + "bbox": [ + 417, + 71, + 537, + 131 + ], + "type": "image", + "image_path": "7e1aee89aa1e75e7b1267cf5d573889caeef574bb1b5299c93baef001b574753.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 417, + 132, + 537, + 192 + ], + "blocks": [ + { + "bbox": [ + 417, + 132, + 537, + 192 + ], + "lines": [ + { + "bbox": [ + 417, + 132, + 537, + 192 + ], + "spans": [ + { + "bbox": [ + 417, + 132, + 537, + 192 + ], + "type": "image", + "image_path": "0184d67e45d66823dee45077c4d53c173cfc92fecda4f9f8c66e4ef8ac5bf129.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 417, + 193, + 537, + 269 + ], + "blocks": [ + { + "bbox": [ + 417, + 193, + 537, + 269 + ], + "lines": [ + { + "bbox": [ + 417, + 193, + 537, + 269 + ], + "spans": [ + { + "bbox": [ + 417, + 193, + 537, + 269 + ], + "type": "image", + "image_path": "93088bc551ca8fdc2f9a1f8926a303063e02ca7899e266867b187dd11a48d407.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 417, + 270, + 537, + 344 + ], + "blocks": [ + { + "bbox": [ + 417, + 270, + 537, + 344 + ], + "lines": [ + { + "bbox": [ + 417, + 270, + 537, + 344 + ], + "spans": [ + { + "bbox": [ + 417, + 270, + 537, + 344 + ], + "type": "image", + "image_path": "6ed38485188a7ac5c5fcd92e351904366a0ca25042153962b2300409cb6f1d0a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 417, + 345, + 537, + 420 + ], + "blocks": [ + { + "bbox": [ + 417, + 345, + 537, + 420 + ], + "lines": [ + { + "bbox": [ + 417, + 345, + 537, + 420 + ], + "spans": [ + { + "bbox": [ + 417, + 345, + 537, + 420 + ], + "type": "image", + "image_path": "d52a0d868b2bfb51bff0591593002071c1a9d87e606d118145b2ce6b4e84bd81.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 421, + 520, + 431 + ], + "lines": [ + { + "bbox": [ + 433, + 421, + 520, + 431 + ], + "spans": [ + { + "bbox": [ + 433, + 421, + 520, + 431 + ], + "type": "text", + "content": "(d) Recovered frames" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 488, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 513 + ], + "type": "text", + "content": "frames for visualization. Please refer to the supplement for a DIY guide." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "content": "Software. To solve Eq. (4), we used a PyTorch implementation of FISTA [3], combined with an approximate proximal operator [15] to handle the anisotropic total variation prior. We ran 10000 iterations of FISTA for each result shown. For most results, we set " + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\lambda_{sparsity} = 0" + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\lambda_{dx} = \\lambda_{dy} = 1 \\times 10^{-3}" + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\lambda_{dt} = 1 \\times 10^{-3}" + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "content": ", but we vary the strength of these terms depending on the brightness of the background and the speed of the motion. We list the parameters for each result and provide sample code in the supplemental." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 689 + ], + "type": "text", + "content": "Calibration and capture. To calibrate the time-varying PSF of our imaging system, we process a stationary image of a small retroreflector or LED, and then computationally rotate this image to create the dictionary required for Eq. (4)—please see the supplemental for more details." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "For every capture, we need to identify the range of angles that are rotated over during the camera exposure in order to" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "text", + "content": "solve Eq. (4). This could be done by placing an LED at a known location, and reconstructing which PSFs are present around it. Alternatively, a controllable motor or trigger system could be used to directly provide the requisite angles. In our work, we opt for a unstructured approach, that simply relies on signal being present at the start and end of the camera exposure. Before capture, we first calibrate the motor to rotate roughly " + }, + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "text", + "content": " over the course of the exposure. Then, we simply perform a coarse reconstruction of Eq. (4) using the captured image and a dictionary that spans all " + }, + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "text", + "content": ", and select the best matching " + }, + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "text", + "content": " for a finer reconstruction. Please see the supplemental for an in-depth discussion." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 645, + 359, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 359, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 359, + 657 + ], + "type": "text", + "content": "6. Results" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 547, + 715 + ], + "type": "text", + "content": "For our real-world reconstructions, we downsampled the input images to " + }, + { + "bbox": [ + 306, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "300 \\times 400" + }, + { + "bbox": [ + 306, + 665, + 547, + 715 + ], + "type": "text", + "content": " pixels and reconstructed " + }, + { + "bbox": [ + 306, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "N_{E} = 146" + }, + { + "bbox": [ + 306, + 665, + 547, + 715 + ], + "type": "text", + "content": " frames for every input. For the simulated results (Fig. 3, Fig. 4), we instead used " + }, + { + "bbox": [ + 306, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "N_{E} = 150" + }, + { + "bbox": [ + 306, + 665, + 547, + 715 + ], + "type": "text", + "content": " to better match" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10795" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 174, + 155 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 174, + 155 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 174, + 155 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 174, + 155 + ], + "type": "image", + "image_path": "28e43121f66334602705b045cf561eb19fd29b878e924710ad249695d3981920.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 156, + 169, + 166 + ], + "lines": [ + { + "bbox": [ + 58, + 156, + 169, + 166 + ], + "spans": [ + { + "bbox": [ + 58, + 156, + 169, + 166 + ], + "type": "text", + "content": "(a) Rotation-encoded frame" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 70, + 295, + 155 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 295, + 155 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 295, + 155 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 295, + 155 + ], + "type": "image", + "image_path": "36b446289bfa65bf028b8c39f266a788242224c793fe5b456be5e7a4610ff48f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 156, + 294, + 166 + ], + "lines": [ + { + "bbox": [ + 176, + 156, + 294, + 166 + ], + "spans": [ + { + "bbox": [ + 176, + 156, + 294, + 166 + ], + "type": "text", + "content": "(b) Example recovered frame" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 417, + 156 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 417, + 156 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 417, + 156 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 417, + 156 + ], + "type": "image", + "image_path": "6dab6558f77b3a5a4d9aa36bcfc2e4ee898538962ec4e0b48aee4d3ed5a87e67.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 156, + 401, + 166 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 401, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 401, + 166 + ], + "type": "text", + "content": "(c) Flow visualization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 171, + 547, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 547, + 204 + ], + "type": "text", + "content": "Figure 8. High-speed particle image velocimetry. Following Sheinin et al. [39], a magnetic stirrer is placed in a tank seeded with tracer particles. (a) Captured diffraction image. (b) Recovered frame using our reconstruction procedure. (c) We visualize a " + }, + { + "bbox": [ + 46, + 171, + 547, + 204 + ], + "type": "inline_equation", + "content": "17.33\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 171, + 547, + 204 + ], + "type": "text", + "content": " interval by overlaying the recovered frames. (d) Using OpenPIV [17], we recover the flow between the start and end frame of (c)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 425, + 75, + 534, + 150 + ], + "blocks": [ + { + "bbox": [ + 425, + 75, + 534, + 150 + ], + "lines": [ + { + "bbox": [ + 425, + 75, + 534, + 150 + ], + "spans": [ + { + "bbox": [ + 425, + 75, + 534, + 150 + ], + "type": "image", + "image_path": "75d3da9bee7fb70c8070a37ab437c90ea2ce23add0b6717fed5bed470f08e40d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 155, + 517, + 166 + ], + "lines": [ + { + "bbox": [ + 440, + 155, + 517, + 166 + ], + "spans": [ + { + "bbox": [ + 440, + 155, + 517, + 166 + ], + "type": "text", + "content": "(d) Recovered flow" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 224, + 286, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 286, + 282 + ], + "type": "text", + "content": "a dual rolling shutter setup. In this document, we use color to code a video into a single image; we refer the reader to our supplemental materials for directly viewing the recovered videos. The camera exposure was " + }, + { + "bbox": [ + 46, + 224, + 286, + 282 + ], + "type": "inline_equation", + "content": "22.178\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 224, + 286, + 282 + ], + "type": "text", + "content": ", yielding an effective framerate of 6583.10 FPS." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 284, + 286, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 286, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 286, + 378 + ], + "type": "text", + "content": "We demonstrate our methodology on fast scenes on multiple examples (Fig. 7)—a bouncing ball, an \"I\" spinning on a fan, toy bullets bouncing off a board, and letters spinning and falling. In the last row of Fig. 7, we applied our technique to marker-based motion capture. Our technique readily captures a wide variety of marker shapes. In Fig. 4(b), we show a temporally sparse, but spatially dense scene of a fan under strobed illumination." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 380, + 286, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 286, + 487 + ], + "type": "text", + "content": "As shown by Sheinin et al. [39], low-cost, high-speed imaging could help measure the movement of fluids in particle image velocimetry [1, 21, 52]. In short, small tracer particles are placed into the medium, which can then be tracked to visualize the flow. We apply our setup to this task in Fig. 8, imaging a tank filled with water and a magnetic stirrer. To maximize the visible movement, we increased the exposure time of our camera to " + }, + { + "bbox": [ + 46, + 380, + 286, + 487 + ], + "type": "inline_equation", + "content": "100\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 380, + 286, + 487 + ], + "type": "text", + "content": " and proportionately decreased the speed of the motor." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "type": "text", + "content": "Finally, we push the temporal resolution of our approach by imaging an LED pulsed at " + }, + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "type": "inline_equation", + "content": "96\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "type": "text", + "content": ". For this particular result, we used a " + }, + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "type": "inline_equation", + "content": "16\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "type": "text", + "content": " focal length lens (Edmund Optics #86-571) to increase the scale of the PSF, and rotated the motor at roughly 15,000 RPM. At resolution " + }, + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "type": "inline_equation", + "content": "2048\\times 1542" + }, + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "type": "text", + "content": " pixels, our methodology is capable of easily resolving this signal as shown in Fig. 2, demonstrating " + }, + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "type": "inline_equation", + "content": "192\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 489, + 286, + 572 + ], + "type": "text", + "content": " capture." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 585, + 122, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 122, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 122, + 597 + ], + "type": "text", + "content": "7. Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Temporal unmixing. Rolling shutter or ROI-based approaches [2, 50, 40, 39] code the signal from each timestamp onto a unique set of measurements that only record the signal at that time. In contrast, with a rotating PSF, the signal from different timestamps is mixed together and requires unmixing using Eq. (4). This unmixing can yield degraded results when there are significant nonidealities, e.g., a poorly calibrated PSF, major spatial variation of the PSF, sizable background signal, or when the scene is too dense." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 221, + 386, + 254 + ], + "blocks": [ + { + "bbox": [ + 309, + 221, + 386, + 254 + ], + "lines": [ + { + "bbox": [ + 309, + 221, + 386, + 254 + ], + "spans": [ + { + "bbox": [ + 309, + 221, + 386, + 254 + ], + "type": "image", + "image_path": "4d9771b55af5f6eab0a9001c5c7ad5ebbb4510ced9c529f3c65a7db406ab85f5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 254, + 346, + 262 + ], + "lines": [ + { + "bbox": [ + 312, + 254, + 346, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 254, + 346, + 262 + ], + "type": "text", + "content": "PSF spect." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 254, + 384, + 261 + ], + "lines": [ + { + "bbox": [ + 354, + 254, + 384, + 261 + ], + "spans": [ + { + "bbox": [ + 354, + 254, + 384, + 261 + ], + "type": "text", + "content": "Rec. Mx" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 222, + 385, + 253 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 222, + 385, + 253 + ], + "spans": [ + { + "bbox": [ + 351, + 222, + 385, + 253 + ], + "type": "image", + "image_path": "6ea3230f13e119f28e2678d062444013c41bdb2baad3c433a0188f7cb180db04.jpg" + } + ] + } + ], + "index": 17, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 386, + 222, + 427, + 254 + ], + "blocks": [ + { + "bbox": [ + 386, + 222, + 427, + 254 + ], + "lines": [ + { + "bbox": [ + 386, + 222, + 427, + 254 + ], + "spans": [ + { + "bbox": [ + 386, + 222, + 427, + 254 + ], + "type": "image", + "image_path": "9de95f1966c497a6c935a12608df5aa7f2dd1074a123afd8b754901f06047179.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 254, + 422, + 261 + ], + "lines": [ + { + "bbox": [ + 390, + 254, + 422, + 261 + ], + "spans": [ + { + "bbox": [ + 390, + 254, + 422, + 261 + ], + "type": "text", + "content": "PSF spect" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 429, + 221, + 463, + 254 + ], + "blocks": [ + { + "bbox": [ + 429, + 221, + 463, + 254 + ], + "lines": [ + { + "bbox": [ + 429, + 221, + 463, + 254 + ], + "spans": [ + { + "bbox": [ + 429, + 221, + 463, + 254 + ], + "type": "image", + "image_path": "e61f5974bd20e2f1d8c9430ba7b45b0b3586de7cf7e5f208be4b67ab5c6455ed.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 507, + 221, + 541, + 254 + ], + "blocks": [ + { + "bbox": [ + 507, + 221, + 541, + 254 + ], + "lines": [ + { + "bbox": [ + 507, + 221, + 541, + 254 + ], + "spans": [ + { + "bbox": [ + 507, + 221, + 541, + 254 + ], + "type": "image", + "image_path": "0f967a5316a8c5a108e5acd302061bd03aab7b932529b7aaba2d649c2ed26691.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 309, + 261, + 385, + 294 + ], + "blocks": [ + { + "bbox": [ + 309, + 261, + 385, + 294 + ], + "lines": [ + { + "bbox": [ + 309, + 261, + 385, + 294 + ], + "spans": [ + { + "bbox": [ + 309, + 261, + 385, + 294 + ], + "type": "image", + "image_path": "63cb8deb4620193d4f109942edf6de57c1e902e69b79c7ff9d0d4262f5c47909.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 295, + 384, + 304 + ], + "lines": [ + { + "bbox": [ + 312, + 295, + 384, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 295, + 384, + 304 + ], + "type": "text", + "content": "Sample recons. frame" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 308, + 369, + 317 + ], + "lines": [ + { + "bbox": [ + 325, + 308, + 369, + 317 + ], + "spans": [ + { + "bbox": [ + 325, + 308, + 369, + 317 + ], + "type": "text", + "content": "Original PSF" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 321, + 545, + 375 + ], + "lines": [ + { + "bbox": [ + 305, + 321, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 545, + 375 + ], + "type": "text", + "content": "Figure 9. Reconstruction sensitivity to PSF spectrum. We reconstruct Fig. 7, Row 2 with digitally augmented versions of the original real measured PSF, to simulate a PSF calibrated under incorrect lighting or from a different material. The reconstruction quality is fairly robust to these PSF variations." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 386, + 261, + 463, + 294 + ], + "blocks": [ + { + "bbox": [ + 431, + 254, + 462, + 261 + ], + "lines": [ + { + "bbox": [ + 431, + 254, + 462, + 261 + ], + "spans": [ + { + "bbox": [ + 431, + 254, + 462, + 261 + ], + "type": "text", + "content": "Rec. " + }, + { + "bbox": [ + 431, + 254, + 462, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\mathbf{x}" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 261, + 463, + 294 + ], + "lines": [ + { + "bbox": [ + 386, + 261, + 463, + 294 + ], + "spans": [ + { + "bbox": [ + 386, + 261, + 463, + 294 + ], + "type": "image", + "image_path": "e4e44bc08f088db1da0e63300f474cf41ae9fb1ca599477519e5e49c128e183a.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 295, + 461, + 304 + ], + "lines": [ + { + "bbox": [ + 389, + 295, + 461, + 304 + ], + "spans": [ + { + "bbox": [ + 389, + 295, + 461, + 304 + ], + "type": "text", + "content": "Sample recons. frame" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 307, + 461, + 316 + ], + "lines": [ + { + "bbox": [ + 389, + 307, + 461, + 316 + ], + "spans": [ + { + "bbox": [ + 389, + 307, + 461, + 316 + ], + "type": "text", + "content": "Bandpass filtered PSF" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 465, + 261, + 541, + 294 + ], + "blocks": [ + { + "bbox": [ + 507, + 254, + 540, + 261 + ], + "lines": [ + { + "bbox": [ + 507, + 254, + 540, + 261 + ], + "spans": [ + { + "bbox": [ + 507, + 254, + 540, + 261 + ], + "type": "text", + "content": "Rec. Mx" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 465, + 261, + 541, + 294 + ], + "lines": [ + { + "bbox": [ + 465, + 261, + 541, + 294 + ], + "spans": [ + { + "bbox": [ + 465, + 261, + 541, + 294 + ], + "type": "image", + "image_path": "37e30f913d02aa6ada26293d65ec955db406064ef53c5bb841d09b3acb52b099.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 295, + 541, + 304 + ], + "lines": [ + { + "bbox": [ + 470, + 295, + 541, + 304 + ], + "spans": [ + { + "bbox": [ + 470, + 295, + 541, + 304 + ], + "type": "text", + "content": "Sample recons. frame" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 307, + 533, + 316 + ], + "lines": [ + { + "bbox": [ + 471, + 307, + 533, + 316 + ], + "spans": [ + { + "bbox": [ + 471, + 307, + 533, + 316 + ], + "type": "text", + "content": "Notch filtered PSF" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 391, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 391, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 545, + 426 + ], + "type": "text", + "content": "In these cases, the signal from one timestamp may bleed into the reconstructions for other timestamps, causing a loss of temporal resolution." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 432, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 589 + ], + "type": "text", + "content": "Saturation. Our method can be especially prone to sensor saturation for several reasons. First, the use of a longer camera exposure may saturate those camera pixels which received scene light over a significant portion of the exposure duration. Secondly, a bright flash of light at a single timestamp could saturate the entire sensor, corrupting the whole video sequence, while only a single row (and therefore timestamp) would be saturated for a rolling shutter or ROI. Thirdly, our grating's PSF contains a 'DC spot' that is much brighter than the rainbow streaks. Therefore, well-exposed rainbow streaks yield a large DC component that might saturate camera pixels. This latter problem could be mitigated by using a higher-quality grating." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Spectrum. The PSF of an object, when viewed through a diffraction grating, depends on its reflectance spectrum, which itself depends on the light source spectrum and object material. Our approach assumes that scenes consist of a single reflectance spectrum so that this PSF is known and fixed for every point in the scene. This assumption makes our approach well-suited for vision applications like motion capture or PIV, which use specialized light sources and markers. In these cases, only slight deviations in PSF occur, which our method can handle (see Fig. 9). However, a" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10796" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "more general scene may consist of many diverse materials and light sources. This results in many different reflectance spectrums and corresponding PSFs at the sensor, restricting the applicability of our method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 251 + ], + "type": "text", + "content": "To address this limitation, the single reflectance spectrum assumption can potentially be relaxed, albeit with tradeoffs. For instance, the reconstruction problem in Eq. (4) can be expanded to explicitly account for these multiple PSFs following Sheinin et al. [39], at the cost of extra computational complexity that scales with the number of possible spectrums. Alternatively, a less wavelength-dependent optic (e.g., a simple diffuser [2]) could be used, such that the PSF does not significantly vary for different wavelengths. However, this loses the additional degrees of information encoded in the different PSF colors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 264, + 212, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 212, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 212, + 277 + ], + "type": "text", + "content": "8. Conclusion and Future Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 380 + ], + "type": "text", + "content": "We demonstrated a novel compressive high-speed video methodology that relies on temporally coding the camera's PSF using a spinning diffraction grating. Our approach can turn ordinary cameras, regardless of their designed speed, into high-speed sensors for spatiotemporally sparse scenes. Compared to prior work, our method can tackle different scene types, is more light efficient, and its temporal resolution is easily increased." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 548 + ], + "type": "text", + "content": "Our prototype relies on a PSF created by a low-cost diffraction grating. Better PSFs, including ones that are optimized using an end-to-end framework [14, 22, 30], would likely produce superior results. Moreover, there are likely more sophisticated mechanical ways to temporally modulate the PSF, potentially providing better performance. For instance, a fast stepper motor that rotates the PSF at discrete steps, synchronized with the reconstruction resolution, could help minimize the mismatch between the discretized model in Eq. (4) and actual measurements. Additionally, other forms of temporal modulation beyond rotation could also prove useful, such as axially translating the spatial-multiplexing optic to scale the PSF, or using a spatial-light modulator to create arbitrary programmable PSFs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 549, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 287, + 644 + ], + "type": "text", + "content": "Additionally, better reconstruction procedures that leverage optical flow [34], dictionary learning [12, 13], or deep learning [14, 22, 30] could significantly improve quality over our naive implementations. Finally, our method is orthogonal to previous compressive video approaches mentioned in Sec. 2; through novel combinations of these techniques, we believe it may be possible to spin out reconstructions at even higher spatial and temporal resolutions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "Acknowledgements. We thank Amitabh Vyas and Ioannis Gkioulekas for fruitful discussions, and Justin Macey for lending us the motion capture equipment. Matthew O'Toole was supported by a National Science Foundation CAREER award (IIS 2238485)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 91, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 311, + 91, + 545, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 545, + 122 + ], + "type": "text", + "content": "[1] Ronald J Adrian and Jerry Westerweel. Particle image velocimetry. Number 30. Cambridge university press, 2011. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 125, + 547, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 125, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 311, + 125, + 547, + 180 + ], + "type": "text", + "content": "[2] Nick Antipa, Patrick Oare, Emrah Bostan, Ren Ng, and Laura Waller. Video from stills: Lensless imaging with rolling shutter. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2019. 2, 3, 4, 5, 8, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 182, + 547, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 182, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 311, + 182, + 547, + 215 + ], + "type": "text", + "content": "[3] Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences, 2(1):183-202, 2009. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 217, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 217, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 311, + 217, + 545, + 249 + ], + "type": "text", + "content": "[4] Joyce Bedi, Douglas Collins, and Harold Eugene Edgerton. Seeing the unseen: Dr. Harold E. Edgerton and the wonders of Strobe Alley. mIt press, 1994. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 251, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 251, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 311, + 251, + 545, + 294 + ], + "type": "text", + "content": "[5] Jacky CP Chan, Howard Leung, Jeff KT Tang, and Taku Komura. A virtual reality dance training system using motion capture technology. IEEE transactions on learning technologies, 4(2):187-195, 2010. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 296, + 459, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 296, + 459, + 307 + ], + "spans": [ + { + "bbox": [ + 311, + 296, + 459, + 307 + ], + "type": "text", + "content": "[6] Roger N. Clark. Clarkvision.com. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 311, + 308, + 545, + 352 + ], + "type": "text", + "content": "[7] P Doron, L Bertuccioli, J Katz, and TR Osborn. Turbulence characteristics and dissipation estimates in the coastal ocean bottom boundary layer from pv data. Journal of Physical Oceanography, 31(8):2108-2134, 2001. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 354, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 354, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 311, + 354, + 545, + 396 + ], + "type": "text", + "content": "[8] Liang Gao, Jinyang Liang, Chiye Li, and Lihong V Wang. Single-shot compressed ultrafast photography at one hundred billion frames per second. Nature, 516(7529):74-77, 2014. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 311, + 399, + 545, + 453 + ], + "type": "text", + "content": "[9] Jinwei Gu, Yasunobu Hitomi, Tomoo Mitsunaga, and Shree Nayar. Coded rolling shutter photography: Flexible spacetime sampling. In 2010 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2010. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "text", + "content": "[10] Evgeny Hahamovich, Sagi Monin, Yoav Hazan, and Amir Rosenthal. Single pixel imaging at megahertz switching rates via cyclic hadamard masks. Nature communications, 12(1):1-6, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 501, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 554 + ], + "type": "text", + "content": "[11] Barmak Heshmat, Guy Satat, Christopher Barsi, and Ramesh Raskar. Single-shot ultrafast imaging using parallax-free alignment with a tilted lenslet array. In CLEO: Science and Innovations, pages STu3E-7. Optica Publishing Group, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "type": "text", + "content": "[12] Yasunobu Hitomi, Jinwei Gu, Mohit Gupta, Tomoo Mitsunaga, and Shree K Nayar. Video from a single coded exposure photograph using a learned over-complete dictionary. In 2011 International Conference on Computer Vision, pages 287-294. IEEE, 2011. 2, 3, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[13] Jason Holloway, Aswin C Sankaranarayanan, Ashok Veeraraghavan, and Salil Tambe. Flutter shutter video camera for compressive sensing of videos. In 2012 IEEE International Conference on Computational Photography (ICCP), pages 1-9. IEEE, 2012. 2, 3, 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[14] Michael Iliadis, Leonidas Spinoulas, and Aggelos Katsaggelos. Deepbinarymask: Learning a binary mask for video compressive sensing. Digital Signal Processing, 96:102591, 2020. 2, 3, 9" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10797" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[15] Ulugbek S Kamilov. A parallel proximal algorithm for anisotropic total variation minimization. IEEE Transactions on Image Processing, 26(2):539-548, 2016. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[16] Roman Koller, Lukas Schmid, Nathan Matsuda, Thomas Niederberger, Leonidas Spinoulas, Oliver Cossairt, Guido Schuster, and Aggelos Katsaggelos. High spatio-temporal resolution video with compressed sensing. Optics express, 23(12):15992-16007, 2015. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "type": "text", + "content": "[17] Alex Liberzon, Theo Käufer, Andreas Bauer, Peter Vennemann, and Erich Zimmer. Openpiv/openpivpython: Openpivpython v0.23.4, Jan. 2021. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 262 + ], + "type": "text", + "content": "[18] Rui Lima, Shigeo Wada, Shuji Tanaka, Motohiro Takeda, Takuji Ishikawa, Ken-ichi Tsubota, Yohsuke Imai, and Takami Yamaguchi. In vitro blood flow in a rectangular pdms microchannel: experimental observations using a confocal micro-piv system. Biomedical microdevices, 10:153-167, 2008. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[19] Yang Liu, Xin Yuan, Jinli Suo, David J Brady, and Qionghai Dai. Rank minimization for snapshot compressive imaging. IEEE transactions on pattern analysis and machine intelligence, 41(12):2990-3006, 2018. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "text", + "content": "[20] Patrick Llull, Xuejun Liao, Xin Yuan, Jianbo Yang, David Kittle, Lawrence Carin, Guillermo Sapiro, and David J Brady. Coded aperture compressive temporal imaging. Optics express, 21(9):10526-10545, 2013. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 388 + ], + "type": "text", + "content": "[21] LM Lourenco, A Krothapalli, and CA Smith. Particle image velocimetry. In Advances in fluid mechanics measurements, pages 127-199. Springer, 1989. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 453 + ], + "type": "text", + "content": "[22] Julien NP Martel, Lorenz K Mueller, Stephen J Carey, Piotr Dudek, and Gordon Wetzstein. Neural sensors: Learning pixel exposures for hdr imaging and video compressive sensing with programmable sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1642-1653, 2020. 2, 3, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "text", + "content": "[23] Amir Matin and Xu Wang. Compressive coded rotating mirror camera for high-speed imaging. In Photonics, volume 8, page 34. MDPI, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[24] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. Mc3d: Motion contrast 3d scanning. In 2015 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 567 + ], + "type": "text", + "content": "[25] Albert Abraham Michelson. Experimental determination of the velocity of light: made at the US Naval Academy, Annapolis, volume 1. US Nautical Almanac Office, 1880. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "type": "text", + "content": "[26] Kristina Monakhova, Vi Tran, Grace Kuo, and Laura Waller. Untrained networks for compressive lensless photography. Optics Express, 29(13):20913-20929, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 656 + ], + "type": "text", + "content": "[27] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1547-1556, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[28] SHR Müller, B Böhm, M Gleißner, R Grzeszik, S Arndt, and A Dreizler. Flow field measurements in an optically accessible, direct-injection spray-guided internal combustion engine using high-speed piv. *Experiments in fluids*, 48:281-290, 2010. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[29] Gergely Nagymáte and Rita M Kiss. Application of optitrack motion capture systems in human movement analysis: A systematic literature review. *Recent Innovations in Mechatronics*, 5(1):1-9, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[30] Cindy M Nguyen, Julien NP Martel, and Gordon Wetzstein. Learning spatially varying pixel exposures for motion deblurring. arXiv preprint arXiv:2204.07267, 2022. 3, 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "text", + "content": "[31] Basilio Pueo Ortega and José M Jiménez Olmedo. Application of motion capture technology for sport performance analysis. Retos:rientas en Educación física, deporte y recreación, (32):241-247, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "text", + "content": "[32] Travis Portz, Li Zhang, and Hongrui Jiang. Random coded sampling for high-speed hdr video. In IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2013. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 239, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 281 + ], + "type": "text", + "content": "[33] Ramesh Raskar, Amit Agrawal, and Jack Tumblin. Coded exposure photography: motion deblurring using fluttered shutter. In Acm Siggraph 2006 Papers, pages 795-804. 2006. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "text", + "content": "[34] Dikpal Reddy, Ashok Veeraraghavan, and Rama Chellappa. P2c2: Programmable pixel compressive camera for high speed imaging. In CVPR 2011, pages 329-336. IEEE, 2011. 2, 3, 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "text", + "content": "[35] Aswin C, Sankaranarayanan, Pavan K Turaga, Rama Chellappa, and Richard G Baraniuk. Compressive acquisition of linear dynamical systems. SIAM Journal on Imaging Sciences, 6(4):2109-2133, 2013. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "type": "text", + "content": "[36] Fulvio Scarano. Tomographic piv: principles and practice. Measurement Science and Technology, 24(1):012001, 2012. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 447 + ], + "type": "text", + "content": "[37] Yoav Y Schechner, Shree K Nayar, and Peter N Belhumeur. Multiplexing for optimal lighting. IEEE Transactions on pattern analysis and machine intelligence, 29(8):1339-1354, 2007. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 449, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 514 + ], + "type": "text", + "content": "[38] Shubham Sharma, Shubhankar Verma, Mohit Kumar, and Lavanya Sharma. Use of motion capture in 3d animation: motion capture systems, challenges, and recent trends. In 2019 international conference on machine learning, big data, cloud and parallel computing (comitcon), pages 289-294. IEEE, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "type": "text", + "content": "[39] Mark Sheinin, Matthew O'Toole, and Srinivasa G Narasimhan. Deconvolving diffraction for fast imaging of sparse scenes. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-10. IEEE, 2021. 2, 3, 4, 5, 8, 9" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 613 + ], + "type": "text", + "content": "[40] Mark Sheinin, Dinesh N Reddy, Matthew O'Toole, and Srinivasa G Narasimhan. Diffraction line imaging. In European Conference on Computer Vision, pages 1-16. Springer, 2020. 2, 3, 4, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "type": "text", + "content": "[41] Mark Sheinin, Yoav Y Schechner, and Kiriakos N Kutulakos. Computational imaging on the electric grid. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6437-6446, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "text", + "content": "[42] H Shen, L Gan, N Newman, Y Dong, C Li, Y Huang, and YC Shen. Spinning disk for compressive imaging. Optics letters, 37(1):46-48, 2012. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 692, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 714 + ], + "type": "text", + "content": "[43] H Shen, N Newman, L Gan, SC Zhong, Y Huang, and YC Shen. Compressed terahertz imaging system using a spin" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10798" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 499 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "disk. In 35th International Conference on infrared, millimeter, and terahertz waves, pages 1-2. IEEE, 2010. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[44] Yaoyao Shi, Youwen Liu, Wei Sheng, Jiming Wang, and Tong Wu. Speckle rotation decorrelation based single-shot video through scattering media. Optics express, 27(10):14567-14576, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "text", + "content": "[45] Ho Shiraga, M Heya, A Fujishima, O Maegawa, K Shimada, Y Kato, T Yamanaka, and S Nakai. Laser-imploded core structure observed by using two-dimensional x-ray imaging with 10-ps temporal resolution. Review of scientific instruments, 66(1):722-724, 1995. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 230 + ], + "type": "text", + "content": "[46] Masada Tzabari and Yoav Y Schechner. Polarized opticalflow gyroscope. In Proceedings of the European Conference on Computer Vision, pages 363-381. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 261 + ], + "type": "text", + "content": "[47] BW Van Oudheusden. Piv-based pressure measurement. Measurement Science and Technology, 24(3):032001, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[48] Ashok Veeraraghavan, Dikpal Reddy, and Ramesh Raskar. Coded strobing photography: Compressive sensing of high speed periodic videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(4):671-686, 2010. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "text", + "content": "[49] Zihao Wang, Leonidas Spinoulas, Kuan He, Lei Tian, Oliver Cossairt, Aggelos K Katsaggelos, and Huajin Chen. Compressive holographic video. Optics express, 25(1):250-262, 2017. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "text", + "content": "[50] Gil Weinberg and Ori Katz. 100,000 frames-per-second compressive imaging with a conventional rolling-shutter camera by random point-spread-function engineering. Optics Express, 28(21):30616-30625, 2020. 2, 3, 4, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "text", + "content": "[51] Charles Wheatstone. An account of some experiments to measure the velocity of electricity and the duration of electric light. Philosophical Transactions of the Royal Society of London, (124):583-591, 1834. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "text", + "content": "[52] Jinhui Xiong, Ramzi Idoughi, Andres A Aguirre-Pablo, Abdulrahman B Aljedaani, Xiong Dun, Qiang Fu, Sigurdur T Thoroddsen, and Wolfgang Heidrich. Rainbow particle imaging velocimetry for dense 3d fluid velocity imaging. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10799" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/4b8d0490-50f4-45bc-9fad-a0ac515c1ca0_content_list.json b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/4b8d0490-50f4-45bc-9fad-a0ac515c1ca0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..12779448c9ebd2009f39b59cfa952a7bf455215b --- /dev/null +++ b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/4b8d0490-50f4-45bc-9fad-a0ac515c1ca0_content_list.json @@ -0,0 +1,1976 @@ +[ + { + "type": "text", + "text": "SportsMOT: A Large Multi-Object Tracking Dataset in Multiple Sports Scenes", + "text_level": 1, + "bbox": [ + 86, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yutao Cui* Chenkai Zeng* Xiaoyu Zhao* Yichun Yang* Gangshan Wu Limin Wang $^{\\text{☑}}$ State Key Laboratory for Novel Software Technology, Nanjing University, China", + "bbox": [ + 96, + 180, + 872, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/MCG-NJU/SportsMOT", + "bbox": [ + 266, + 227, + 697, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 294, + 313, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-object tracking (MOT) in sports scenes plays a critical role in gathering players statistics, supporting further applications, such as automatic tactical analysis. Yet existing MOT benchmarks cast little attention on this domain. In this work, we present a new large-scale multi-object tracking dataset in multiple sports scenes, coined as SportsMOT, where all players on the court are supposed to be tracked. It consists of 240 video sequences, over 150K frames (almost $15 \\times$ MOT17) and over 1.6M bounding boxes ( $3 \\times$ MOT17) collected from 3 sports categories, including basketball, volleyball and football. Our dataset is characterized with two key properties: 1) fast and variable-speed motion and 2) similar yet distinguishable appearance. We expect SportsMOT to encourage the MOT trackers to promote in both motion-based association and appearance-based association. We benchmark several state-of-the-art trackers and reveal the key challenge of SportsMOT lies in object association. To alleviate the issue, we further propose a new multi-object tracking framework, termed as MixSort, introducing a MixFormer-like structure as an auxiliary association model to prevailing tracking-by-detection trackers. By integrating the customized appearance-based association with the original motion-based association, MixSort achieves state-of-the-art performance on SportsMOT and MOT17. Based on MixSort, we give an in-depth analysis and provide some profound insights into SportsMOT.", + "bbox": [ + 75, + 325, + 472, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 762, + 209, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-object tracking (MOT) has been a fundamental computer vision task for recent decades, aiming to locate the objects and associate them in video sequences. Researchers have cast much focus on various practical use cases like crowded street scenes [9, 27], static dancing scenes [33] and driving scenarios [14], achieving consid-", + "bbox": [ + 75, + 787, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "erable progress [2, 4, 30, 37, 38, 40, 41] in MOT. MOT for sports scenes however is overlooked, where typically only the players on the court should be tracked for further analysis, such as counting the players' running distance or average speed and automatic tactical analysis.", + "bbox": [ + 500, + 280, + 890, + 356 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generally, prevailing state-of-the-art trackers [1,6,11,39, 44] consist of several components to accomplish the tracking task: objects localization module, motion based objects association module and appearance based association module. Biased to the data distribution of specific human tracking benchmarks, e.g. MOT17 [27], MOT20 [9] and DanceTrack [33], the components of these trackers have difficulty adapting to sports scenes. Firstly, motivated by surveillance or self-driving applications, current human tracking benchmarks provide tracks for almost all persons in the scenes. While for sports scenes like basketball or football games, generally only the players on the court are what we focus on, hence a specialized training platform is required to make the detectors suitable for sports scenes. More importantly, in MOT17 and MOT20, these trackers highlight Kalman Filter [16] based IoU matching for object association, due to the slow and regular motion of pedestrians. DanceTrack highlights diverse motion rather than fast movement [33], that is, dancers frequently switch the motion direction and relative position. However in sports scenes, we observe fast and variable-speed movement of objects on adjacent frames, i.e. players usually possess high speed and frequently change their running speed in professional sports events, thus constituting barriers in existing motion based association. For instance, as visualized in Fig. 1, the adjacent IoU and Kalman-Filter-based IoU in sports scenes remain lower than that on MOT17 and DanceTrack (More detailed comparison can be found in Fig. 2 and Fig. 3). As a consequence, more suitable motion based association for sports scenes is required. Additionally, compared to the MOT17 and MOT20 datasets in street scenes, the objects appearances in sports scenes are less distinguishable, since not only the players inherently are in similar clothes but also the players are frequently blurred caused by fast camera motion or targets motion. Different from DanceTrack, where generally the dancers are in almost the same clothes and", + "bbox": [ + 496, + 357, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. ☑ : Corresponding author.", + "bbox": [ + 99, + 887, + 357, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9921", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6093b651a013e26892c96857ae8238d2a7d15e423f7e960ccd6e9ceb0ec2943.jpg", + "image_caption": [ + "(a) Basketball in SportsMOT" + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 470, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae829d78234f11771cd13c87b41f86c1f0f414483014f7199098186c96303f81.jpg", + "image_caption": [ + "(c) MOT17" + ], + "image_footnote": [], + "bbox": [ + 496, + 89, + 864, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5834c172bde20e9fdfbbd3ba7f2b91e339f0295db95cbf0f7e7331f7e8a7b9a5.jpg", + "image_caption": [ + "(b) Football in SportsMOT" + ], + "image_footnote": [], + "bbox": [ + 102, + 167, + 470, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd5d3028c2d7dea6394a2e18565223a3633a4cb0662063701a63849d7b625af6.jpg", + "image_caption": [ + "(d) DanceTrack" + ], + "image_footnote": [], + "bbox": [ + 496, + 167, + 864, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/29d4c55a59ed743e82bd19e5eb1c3fa30f0b6ddec4e7163085ed4f341f60d639.jpg", + "image_caption": [ + "Figure 1. Sampled sequences from the categories of basketball and football of SportsMOT, MOT17 and DanceTrack. There exist two key properties of SportsMOT: 1) fast and variable-speed motion, i.e. players usually possess high speed and frequently change their running speed (the visualized adjacent IoU and Kalman Filter based adjacent IoU can indicate the property); 2) similar yet distinguishable appearance, that is, players in sports scenes inherently wear jackets with different numbers and usually display distinct postures. We expect SportsMOT to encourage the MOT trackers to promote in both motion-based association and appearance based association." + ], + "image_footnote": [], + "bbox": [ + 176, + 250, + 470, + 323 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2a40cd93e332510d9c72f1b185d943206cfc13081a1a8b1d7b8de3ff234478c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 323, + 349, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4297c956d949b511450566b2ed26f61d00c336736fcd7e35d5666fb2e9d13c83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 323, + 362, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2abb806f47620e8beef110e381485a4b327f968b047b721745244f9f1db4ceae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 323, + 379, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1833f6c580ee1ddde54002f1d5e9cf3816232ab2ec2d9d3434f49d957e16c880.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 323, + 392, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9430a6b413733f12ef19ee98ccd5eee007ca0b6a32986190248f6fb3a54718a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 323, + 415, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f30e7e86769324880e288f2e2f7512c63a20be63907eac872f6bfbcea724bdad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 323, + 429, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b3f4e10e4f9d68178ce708f523e32b8b74be86608036591837eaeba152acd74d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 323, + 442, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/91ca6723bb13a3da58896b395bf520d2f3146e9f1bbc60862c4fa1af6ca0b498.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 323, + 459, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3c1ca75ed8bd64dc7c3c3c5c812dcd764f71e7cb6c32e57ba401aeabd5d19b3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 323, + 473, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cd290579af55a967b1cbd553c1a82900df4c34a784077437d13905e5f0fa6298.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 250, + 789, + 323 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e9822f6aa6b70b5bea6cb69c7b512fc035dbe5c6abe4e828d88bac9696416f1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 323, + 539, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a8764d7458b8c915de171a02a608a23acf8617f09137e2eef3ca6db89356f5db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 323, + 553, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae65fde53686d5266e33ab420be215e9ffab2ad998f7f98d62a2bd8ee0dc242c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 323, + 566, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4448b7920f3b15788934e4960792dda92224160680a089b7d1f1d553f7046448.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 323, + 581, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9df2ffcd97e0e8978fc913c7f77daacc9ef40d0e022dd90f86e665b8e4ecad9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 323, + 596, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1da734336ee7669b09d167ef069f1bfcd244473a9d4870a2b5d8fdce1b47e14e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 323, + 612, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d15aa12fa1e6d47a082a35f3f8f1456fa49b175a258003d7af7d9b4f481e1b42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 323, + 630, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "thus having indistinguishable appearance, players in sports scenes inherently wear jackets with different numbers and usually display distinct postures. Thereby, we argue that objects in sports scenes are with similar yet distinguishable appearance, which necessitates the appearance model developing more discriminative and extensive representations.", + "bbox": [ + 75, + 417, + 468, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the analysis above, to advance the development of tracking and sports analysis, we propose a multi-object tracking dataset in sports scenes, termed as SportsMOT. The dataset is large-scale, high-quality and contains dense annotations for every player on the court in various sports scenes. It consists of 240 videos, over 150K frames (almost $15 \\times$ MOT17 [27]) and over 1.6M bounding boxes ( $3 \\times$ MOT17 [27]) collected from 3 categories of sports, including basketball, volleyball and football. To provide a platform for making the trackers suitable for sports scenes, we split the dataset into training, validation and test subsets, consisting of 45, 45 and 150 video sequences respectively. There exist two core properties of SportsMOT: (1) fast and variable-speed motion, requiring more suitable motion modeling association; (2) similar yet distinguishable appearance, which necessitates the appearance model developing more discriminative and extensive representations. Altogether, we expect SportsMOT to encourage the trackers to promote in both the certain aspects, i.e. motion based association and appearance based association.", + "bbox": [ + 75, + 515, + 470, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the large-scale multi-object tracking dataset SportsMOT, we benchmark some recent tracking approaches and retrain all of them on the training split. We observe IDF1 and AssA metrics are lower than that on MOT17 while the DetA is quite high, indicating that the main chal", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lenge of SportsMOT lies in objects association rather than objects localization. To alleviate the issue, we propose a new multi-object tracking framework, dubbed as MixSort, with introducing a MixFormer-like [8] structure as appearance based association to prevailing tracking-by-detection trackers (e.g. ByteTrack [44], OC-SORT [6]). By integrating the original motion based objects association and the designed appearance based association, the performance gets boosted on both SportsMOT and MOT17 benchmarks. Based on MixSort, we perform extensive exploration studies and provide some profound insights into SportsMOT.", + "bbox": [ + 496, + 417, + 893, + 584 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows:", + "bbox": [ + 519, + 588, + 859, + 602 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We build a new large-scale multi-object tracking dataset in diverse sports scenes, SportsMOT, equipped with two key properties of 1) fast and variable-speed motion and 2) similar yet distinguishable appearance, aiming to advance the development of both tracking and sports analysis.", + "- We benchmark some prevailing trackers on SportsMOT, which reveals that the key challenges lie in objects association and hopefully can facilitate further research.", + "- We propose a new multi-object tracking framework MixSort, with introducing a MixFormer-like structure as appearance based association model to prevailing tracking-by-detection trackers, so as to boost the objects association. Based on MixSort, we perform extensive studies and provide some profound insights into SportsMOT." + ], + "bbox": [ + 517, + 614, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9922", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/163adefa5a4800ef308cfd0f727b55beb06107bb98d33a73f1a93e5dfae100e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetVideosFramesLength (s)BboxTracks
MOT171411,235463292,7331,342
MOT20813,4105351,652,0403,456
DanceTrack100105,8555,292-990
SportsMOT240150,3796,0151,629,4903,401
", + "bbox": [ + 81, + 89, + 467, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of statistics between existing human MOT datasets and our SportsMOT.", + "bbox": [ + 76, + 167, + 467, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 200, + 217, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-object tracking datasets. Existing Multi-object tracking datasets usually focus on different scenes, such as autonomous driving, pedestrians on roads, and dancing. For autonomous driving, there are KITTI [14], KITTI360 [22] and BDD100K [42], which focus on pedestrians and vehicles. Besides, other datasets, focusing on only the pedestrians, collect the videos from static and moving cameras. One of the earliest is PETS [12], but it's too simple in some scenes. MOT15 [20] proposes the first large-scale benchmark for Multi-object tracking, followed by MOT17 [27] and MOT20 [9]. It's worth noting that MOT20 focuses on extremely crowded scenes where many pedestrians are occluded, increasing the tracking difficulty greatly in both detection and association. Recently, DanceTrack [33], focusing on dancing scenes, is proposed to encourage trackers to rely less on visual discrimination and depend more on motion analysis. The emphasized properties are uniform appearance and diverse motion. While in SportsMOT, the appearance is similar yet distinguishable, and the players' motion is fast and with variable speed. We expect SportsMOT to encourage algorithms to promote in both appearance and motion association. Besides, SoccerNet-Tracking [7] is presented to track elements in football scenarios. The main difference lies in that, it only contains soccer scenes and tracks almost all elements (players, goalkeepers, referees, balls) on the court without distinction. While SportsMOT contains three types of sports where the objects in basketball scenes are more crowded and thus more challenging (refer to Section 5.3), and only focuses on the players to support further statistics and tactical analysis.", + "bbox": [ + 76, + 227, + 468, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object association in tracking. Association is a very important task in tracking, where trackers need to associate detections in new frames with existing tracks. For most of the trackers, a similarity matrix (or cost matrix) between new detections and tracks is computed based commonly on motion and appearance cues, which is later fed into Hungarian algorithm [18] to perform association. For example, SORT [4] uses Kalman Filter to predict the location of objects and computes the IoU of detected and predicted bounding boxes as similarity matrix. IOU-Tracker [5] directly computes the IoU without prediction. ByteTrack [44] adds an association phase for detection with low confidence score, which can boost the performance. OC-SORT [6] tries to address the limitation of Kalman Filter.", + "bbox": [ + 76, + 688, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Appearance cues also play an important role in association. DeepSORT [39] crops the detection from frame images, which are then used by networks to generate re-ID features. Then the motion cues and distance of re-ID features are fused to perform association. FairMOT [45] uses a re-ID branch on a backbone shared by the detection branch to generate re-ID features. In CenterTrack [46], the previous frame is used to help the prediction of tracks. Recently, Transformer [36] is used by some work such as TrackFormer [26] and MOTR [43] to boost the association quality. Our proposed MixSort integrates the motion and appearance cues with a motion modeling component and the designed MixFormer-like structure respectively.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SportsMOT Dataset", + "text_level": 1, + "bbox": [ + 498, + 301, + 692, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset Construction", + "text_level": 1, + "bbox": [ + 500, + 325, + 699, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Collection. We select three worldwide famous sports, football, basketball, and volleyball, and collect videos of high-quality professional games including NCAA, Premier League, and Olympics from Multi-Sports [21], which is a large dataset in sports area focusing on spatio-temporal action localization. Each category has typical players' formations and motion patterns, and they can effectively represent the diversity of sports scenarios. Only the overhead shots of sports game scenes are used, guaranteeing certain extreme situations do not occur. The proposed dataset consists of 240 video sequences in total, each of which is 720P and 25 FPS. Following the principles of multi-object tracking, each video clip is manually checked to ensure that there are no abrupt viewpoint switches within the video.", + "bbox": [ + 496, + 348, + 890, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Annotation Pipeline. We annotate the collected videos according to the following guidelines.", + "bbox": [ + 500, + 579, + 888, + 609 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The entire athlete's limbs and torso, excluding any other objects like balls touching the athlete's body, are required to be annotated.", + "- The annotators are asked to predict the bounding box of the athlete in the case of occlusion, as long as the athletes have a visible part of body. However, if half of the athletes' torso is outside the view, annotators should just skip them.", + "- We ask the annotators to confirm that each player has a unique ID throughout the whole clip." + ], + "bbox": [ + 517, + 612, + 888, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide a customized labeling tool for SportsMOT and a corresponding manual book to annotators. Once they start annotating a new object, the labeling tool automatically assigns a new ID to the object and propagates the bounding box of previous state to the current state, with the help of the single object tracker KCF [15]. Then the generated bounding boxes should be refined by the annotators, so as to improve annotation quality. After carefully reviewing each", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9923", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fda639178cdbf7cb39b2d0186c2e29e82e4fdb3256ac906a590a448d557cbf14.jpg", + "image_caption": [ + "Figure 2. IoU on adjacent frames. (a) Compared to MOT17 and DanceTrack, SportsMOT has a lower score, indicating that objects have faster motion. (b) In SportsMOT, the category of football has the lowest IoU score, which means that football players often have fast motion." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 274, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19c98d15d90c27ace1fbc4181a2f4a722e50374c38bf6c46523b085980628a0f.jpg", + "image_caption": [ + "Figure 3. Kalman-Filter-based IoU on adjacent frames. (a) Compared to MOT17 and DanceTrack, SportsMOT has a lower score, indicating that objects have more variable-speed motion. (b) In SportsMOT, the category of football has the lowest Kalman-Filter-based IoU score." + ], + "image_footnote": [], + "bbox": [ + 282, + 90, + 464, + 196 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/6603b96078f85aad876bab17afa6e8a138015f6c7effa0ce4c2321c408da02c5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryFramesTracksTrack gap len.Track len.Bboxes per frame
Basketball845.41068.7767.99.1
Volleyball360.41238.2335.911.2
Football673.920.5116.1422.112.8
Total626.614.296.6479.110.8
", + "bbox": [ + 81, + 275, + 467, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Detailed statistics of the three categories in SportsMOT.", + "bbox": [ + 78, + 375, + 465, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "annotation result, we refine the bounding boxes and IDs that do not satisfy the standards, hence building a high-quality dataset. Finally, the bounding boxes with too small size, i.e. $w < 5$ or $h < 5$ , are deleted.", + "bbox": [ + 75, + 395, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dataset Statistic", + "text_level": 1, + "bbox": [ + 76, + 465, + 238, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. SportsMOT is a large-scale and high-quality MOT dataset, aiming to advance the development of both sports analysis and multi-object tracking. Table 1 compares the statistics of SportsMOT with the prevailing human tracking datasets, including MOT17, MOT20 and DanceTrack. According to the statistics, SportsMOT has a large number of bounding boxes of over 1.6M, which is comparable to MOT20 and significantly larger than MOT17. Besides, SportsMOT has a large number of video clips, tracks $(2.5 \\times \\text{MOT17}, 3.4 \\times \\text{DanceTrack})$ , frames $(13.4 \\times \\text{MOT17})$ . As shown in Table 2, we also compare the basic statistics of each category of SportsMOT. SportsMOT solely provides fine annotations of the players on the court, which are supposed to be tracked for further analysis. To provide a platform for making the trackers suitable for sports scenes, we split the dataset into training, validation and test subsets, consisting of 45, 45 and 150 video sequences respectively.", + "bbox": [ + 75, + 488, + 468, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fast and Variable-Speed Motion. Motion cues play an important role in object association for multi-object tracking. The existing human-tracking datasets (e.g. MOT17, MOT20 and DanceTrack) generally have certain motion patterns that are distinct with sports scenes, constituting barriers in players tracking. For instance, in MOT17 and MOT20, pedestrians are featured by linear motion with constant speed, which is easily hacked by association strategies with constant velocity assumption. Besides, Dance", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3e823abf075bdab709ff5becb3f1653f6da32e08434215420dc3a059251b821a.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 700, + 183 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/51856aaa134953264ab6c1f433fa41092026c734f9df2e8964a94753142cd4a2.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 702, + 90, + 885, + 183 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7a4a0510c929571eca65970e4f2a745773160dad264ace4225c0b90b778f46c0.jpg", + "image_caption": [ + "(a) MOT17" + ], + "image_footnote": [], + "bbox": [ + 501, + 273, + 629, + 345 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/798f7bbe3ddb04b21dab554643dad22f5e368bec787b7d09db73d8b8989211e8.jpg", + "image_caption": [ + "(b) SportsMOT", + "Figure 4. Visualization of re-ID features from sampled videos in MOT17, SportsMOT and DanceTrack dataset using t-SNE [35]. The same object is coded by the same color. It indicates that object appearance of SportsMOT is less distinguishable than that of MOT17, while more distinguishable than that of DanceTrack. We expect the appearance model to capture more discriminative and extensive representation for object association." + ], + "image_footnote": [], + "bbox": [ + 632, + 273, + 756, + 345 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/adbef16cd430c0f29d8def26cc5bc827c6c379c79588c8efdb54a39eae1d233d.jpg", + "image_caption": [ + "(c) DanceTrack" + ], + "image_footnote": [], + "bbox": [ + 761, + 273, + 885, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Track highlights diverse motion rather than fast motion, that is, dancers usually move in more diverse directions with relatively low speed. In contrast, SportsMOT has distinct motion patterns, i.e. fast and variable-speed motion, where the players typically move fast with their running speed or camera speed frequently changing. As illustrated in Fig. 2, among the three datasets, SportsMOT has the lowest IoU score of the objects bounding boxes on adjacent frames, indicating the fast movement. We use the ground truth of previous frames for Kalman Filter prediction. The result and current ground truth are used to calculate Kalman-Filter-based IoU. Seen from Fig. 3, SportsMOT also has the lowest Kalman-Filter-based IoU score on adjacent frames, which suggests that the motion can not be easily modeled by prevailing methods due to the variable-speed movement. Specifically, football has the smallest adjacent IoU and Kalman-Filter-based IoU, which is closely related to the fast running speed, abrupt acceleration or stops. It poses a major challenge for trackers based on simple motion assumptions and also encourages them to model object motion in more dynamic and adaptive ways.", + "bbox": [ + 496, + 472, + 890, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar yet Distinguishable Appearance. Object appearance is another kind of cue on which MOT trackers often rely to distinguish different objects. In MOT17 and MOT20, pedestrians are usually distinct in body size and wear different clothes, yielding discriminative visual features. In contrast, the objects in DanceTrack typically wear nearly identical outfits, leading to indistinguishable", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9924", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c7f690504ba803de7ddee35c2c84b4ec3054b9ccbd9dcfda39837a90587c3b76.jpg", + "image_caption": [ + "Figure 5. Paradigm for computing visual similarity matrix of tracks and detections in MixSort. The center of ground-truth target detection is marked with red dots, the others are green. The blue dashed box indicates the cropped search region. The blank part of similarity vector means that for detections not in the search region, the corresponding value is set to 0." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 465, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "appearances. Thereby, DanceTrack highlights solely relying on motion-based association rather than appearance-based association. In SportsMOT, the players also have very similar appearances. However, the players wear jackets with different numbers and usually display distinct postures, thus resulting in similar yet distinguishable appearance. In Fig. 4, we provide visualization of re-ID feature from sampled videos in MOT17, SportsMOT and DanceTrack dataset using t-SNE [35]. It implies that the re-ID features in SportsMOT are similar yet distinguishable compared to MOT17 and DanceTrack. We aim to encourage the trackers to learn more discriminative visual representations for more robust object association.", + "bbox": [ + 75, + 292, + 472, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. MOTA [3] is the main metric for existing MOT evaluations. However, MOTA focuses more on measuring the accuracy of detection. To highlight the performance of object association, we recommend HOTA [25], AssA and IDF1 [31] as the major evaluation metrics in SportsMOT dataset. HOTA aims to measure the accuracy of detection and association equally.", + "bbox": [ + 75, + 494, + 468, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Multi-Object Tracking on SportsMOT", + "text_level": 1, + "bbox": [ + 76, + 614, + 423, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present our proposed multi-object tracking framework, called MixSort. This framework is designed to enhance the appearance-based association performance and can be applied to any trackers that follow the tracking-by-detection paradigm, such as ByteTrack [44] and OC-SORT [6].", + "bbox": [ + 75, + 640, + 468, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. MixFormer for Appearance-based Association", + "text_level": 1, + "bbox": [ + 76, + 739, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MixFormer. In this paragraph, we discuss the use of Mix-Former [8] in our proposed framework MixSort. Mix-Former is designed to extract target-specific discriminative features and perform extensive communication between the target and search area, therefore, it is the key component that enables MixSort to compute visual similarities between the templates of tracked objects and detected objects in the search region of the current frames. The original Mix-Former uses a corner-based localization head to predict the", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1a4bb390bb15982083ff7954ca0b4d2d03aeb55eb0c63a0e7fb161bfd97d936.jpg", + "image_caption": [ + "Figure 6. The pipeline of MixSort. We use motion model and MixFormer to generate fused similarity matrix for association." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "top-left and bottom-right corners of the input template in the search region. However, we modify the corner head by using a heatmap prediction head that predicts the center of the template and generates a confidence heatmap. This allows us to compute the similarity between the detection and the template. To make MixSort suitable for multi-object tracking and accelerate inference speed, we reduce the number of mixed attention modules in MixFormer from 12 to 4. The steps involved in computing the visual similarity matrix are illustrated in Figure 5.", + "bbox": [ + 496, + 247, + 893, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Association Strategy. In order to perform association between detections and existing tracks, we use a mixed similarity matrix generated by computing the visual similarity between the target template and the detected objects in the search region of the current frame. Specifically, we obtain the heatmap response at the center of each detection as its visual similarity to the template. The resulting similarity matrix is then combined with the IoU matrix using the Hungarian Algorithm.", + "bbox": [ + 496, + 402, + 890, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To start, for each existing track $t$ , we use the Kalman Filter to predict its new location. Then, we crop the current frame centered at the predicted location with a certain scale to obtain the search region $s$ . By feeding $s$ and the template $t$ into MixFormer, we generate a heatmap $\\mathcal{H}$ that represents the similarity between the template and search region.", + "bbox": [ + 496, + 539, + 890, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, for each detection $d$ whose center is in the search region $s$ , we set its similarity to track $t$ as the response in the heatmap $\\mathcal{H}$ ; the similarity values of other detections are set to 0. Finally, we fuse the visual similarity and the IoU score to obtain the mixed similarity matrix", + "bbox": [ + 496, + 630, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM = \\alpha \\cdot \\operatorname {I o U} + (1 - \\alpha) \\cdot V \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 718, + 890, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is the weight coefficient and $V$ represents the visual similarity matrix calculated using MixFormer.", + "bbox": [ + 498, + 746, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. MixSort Tracking", + "text_level": 1, + "bbox": [ + 500, + 786, + 674, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the tracking-by-detection paradigm, the pipeline of MixSort can be generalized as follows. As shown in Figure 6, we first obtain detections using an object detector. Then, we employ a motion model (e.g., Kalman Filter) to predict new locations of existing tracks. Based on the new locations and templates of tracks, we compute a", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9925", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/40449ebccba8dd7450f2dc44743207f66e3eb6fddb03905e176d3e8ff826a857.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training SetupHOTA↑IDF1↑AssA↑MOTA↑DetA↑LocA↑IDs↓Frag↓
CenterTrack [46]Train62.760.048.090.882.190.8104815750
FairMOT [45]Train49.353.534.786.470.283.9992821673
QDTrack [29]Train60.462.347.290.177.588.0637711850
TransTrack [34]Train68.971.557.592.682.791.049929994
GTR [47]Train54.555.845.967.964.889.0956714525
ByteTrack [44]Train62.869.851.294.177.185.632674499
OC-SORT [6]Train71.972.259.894.586.492.430933474
ByteTrackTrain+Val64.171.452.395.978.585.730894216
OC-SORTTrain+Val73.774.061.596.588.592.727283144
MixSort-ByteTrain+Val65.7 (+1.6)74.1 (+2.7)54.8 (+2.5)96.278.885.724724009
MixSort-OCTrain+Val74.1 (+0.4)74.4 (+0.4)62.0 (+0.5)96.588.592.727813199
", + "bbox": [ + 81, + 89, + 893, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Tracking performance of investigated algorithms on our proposed SportsMOT. The best results are shown in bold.", + "bbox": [ + 119, + 285, + 846, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fused similarity matrix as described above and use it to associate tracks and detections by the Hungarian Algorithm. Finally, for matched tracks and detections, we update the online templates. For unmatched tracks, we keep them until the threshold is reached. For unmatched detections with confidence scores higher than the threshold, new tracks are initialized.", + "bbox": [ + 75, + 318, + 470, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 433, + 290, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. We only consider the training of MixFormer here since the detector remains the same as initial method (e.g. ByteTrack). The original MixFormer is trained on SOT datasets, so we first modify the format of ground truth of MOT datasets, that is, converting the ground truth trajectory of every single player into TrackingNet format [28].", + "bbox": [ + 75, + 455, + 468, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each ground-truth bounding box, we compute its corresponding center location $(c_x, c_y)$ in the low-resolution heatmap. Following CornerNet [19], the ground-truth heatmap response is generated using 2D Gaussian kernel:", + "bbox": [ + 75, + 547, + 468, + 608 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {x y} = \\exp \\left(- \\frac {\\left(x - c _ {x}\\right) ^ {2} + \\left(y - c _ {y}\\right) ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 617, + 468, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma$ is adaptive to the size of the bounding box. The training loss is a pixel-wise logistic regression with focal loss [23]:", + "bbox": [ + 75, + 652, + 468, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = - \\sum_ {x y} \\left\\{ \\begin{array}{l l} (1 - \\hat {h} _ {x y}) ^ {\\gamma} \\log (\\hat {h} _ {x y}) & , h _ {x y} = 1; \\\\ (1 - h _ {x y}) ^ {\\beta} (\\hat {h} _ {x y}) ^ {\\gamma} \\log (1 - \\hat {h} _ {x y}) & , \\text {o t h e r w i s e}. \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 705, + 468, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma$ and $\\beta$ are hyper-parameters in focal loss, and we set $\\gamma = 2$ , $\\beta = 4$ following CornerNet.", + "bbox": [ + 75, + 761, + 468, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. For each track, we maintain only one template for keeping a balance between accuracy and speed. When a detection is matched to an existing track, we directly replace the original template to the new detection, if and only if the ratio of its uncovered (i.e. overlapping with any detected objects) area is larger than a certain threshold, so as to reduce the impact of misleading representations.", + "bbox": [ + 75, + 794, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments and Analysis", + "text_level": 1, + "bbox": [ + 498, + 316, + 743, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 342, + 676, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Split. In benchmark experiments, we follow the default split described in Sec. 3.2. In exploration study, we split the original MOT17 training set into two sets, used for training and validation respectively following CenterTrack.", + "bbox": [ + 496, + 364, + 890, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Following ByteTrack and OC-SORT, we use YOLOX [13] as our detector. Using COCO-pretrained model as the initialized weights, we first train the model on CrowdHuman [32] for 80 epochs and then train on SportsMOT for another 80 epochs. The remaining settings are the same as that in ByteTrack.", + "bbox": [ + 496, + 429, + 890, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For MixFormer, we initialize the backbone with the model trained on VOT datasets and then fine-tune it on SportsMOT for 300 epochs with learning rate initialized as $1e - 4$ and decreased to $1e - 5$ at epoch 200. The optimizer is ADAM [17] with weight decay $10^{-4}$ . The sizes of search images and templates are $224 \\times 224$ and $96 \\times 96$ respectively. The max sample interval is set to 10. For every tracking result, we apply linear interpolation as post-processing, with maximum gap set to 20.", + "bbox": [ + 496, + 520, + 890, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Benchmark Results", + "text_level": 1, + "bbox": [ + 500, + 665, + 687, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate several representative methods of three kinds on our dataset. ByteTrack [44], OC-SORT [6] and QDTrack [29] are trackers in tracking-by-detection paradigm. CenterTrack [46] and FairMOT [45] perform joint detection and tracking in one stage. TransTrack [34] and GTR [47] are trackers based on Transformer. Most of the current best multi-object tracking algorithms belong to tracking-by-detection paradigm, however, due to the separation of detection and tracking, the information cannot be shared completely. Joint-detection-and-tracking paradigm couples the two modules, with the goal of boosting the performance of each. Transformer-based-tracking methods are relatively new but have achieved great performance. Despite its huge potential, the model complexity and calcula", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9926", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9b51716ecd2c75621a6bff1f443e753cf70e24f0b1d6c3d8cf4b2306c3cdfb0b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
α10.90.80.70.60.50.40.30.20.10
basketball65.966.166.265.965.364.863.660.756.747.126.9
volleyball76.076.876.576.476.576.976.475.573.769.140.9
football71.972.472.372.472.572.873.272.972.671.465.7
", + "bbox": [ + 78, + 89, + 470, + 138 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion cost are much higher, resulting in large memory and long training time.", + "bbox": [ + 75, + 196, + 468, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All training settings including the number of epochs and change of learning rate are consistent with original papers. According to different default settings, we follow the commonly used pretraining datasets, such as CrowdHuman [32], COCO [24] and ImageNet [10], and apply SportsMOT-train with or without other datasets for finetuning for different methods. We compare the results in Tab. 3.", + "bbox": [ + 75, + 227, + 468, + 332 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In sports scenes, the clear appearance and sparse density of objects allow current mature detection frameworks to generate bounding boxes with high accuracy. However, specialized detectors need to be trained for not detecting audience and referees. The key challenges are fast speed and motion blur, which forces us to pay more attention to improve the association performance. Besides, The wide range of HOTA and MOTA denotes SportsMOT is more distinguishable among different kinds of algorithms.", + "bbox": [ + 75, + 333, + 468, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Exploration Study", + "text_level": 1, + "bbox": [ + 76, + 481, + 254, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform extensive studies on the proposed MixSort and SportsMOT.", + "bbox": [ + 76, + 505, + 468, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the proposed association module. We evaluate the effectiveness of MixSort by applying it to two state-of-the-art trackers, OC-SORT [6] and ByteTrack [44], which follow the tracking-by-detection paradigm and useYOLOX as their detector. The evaluation is conducted on the SportsMOT test set, and the results are presented in Tab. 3. Our experiments show that MixSort significantly improves the performance of both trackers, with OC-SORT achieving a 0.4 HOTA increase and ByteTrack achieving a 1.6 HOTA increase on SportsMOT. This demonstrates the effectiveness of MixSort in enhancing the association.", + "bbox": [ + 75, + 544, + 468, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Appearance-based vs. Motion-based association. We have demonstrated that MixSort can improve association performance. In this paragraph, we investigate the impact of the fusion weight $\\alpha$ on the ability of appearance cues to aid in conventional motion-based association. We evaluate OC-SORT with MixSort on the three categories in the SportsMOT test set using $\\alpha$ values ranging from 1 to 0 in Eq. (1). The results, presented in Tab. 4, reveal that pure motion-based association $(\\alpha = 1)$ outperforms pure appearance-based association $(\\alpha = 0)$ in all categories, underscoring the significance of motion cues in sports scenes. Moreover, fused association surpasses both", + "bbox": [ + 75, + 719, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3d6c6405994a2d1737bf45dd184849f8efd2e09005eae4f5fea6cb0ba4b66834.jpg", + "table_caption": [ + "Table 4. Comparison of the HOTA metric of basketball, volley and football under different fusion parameters $\\alpha$ on SportsMOT test set. The models are trained on SportsMOT training set." + ], + "table_footnote": [], + "table_body": "
IoUMotionMix.HOTA↑IDF1↑AssA↑MOTA↑IDs↓
✓71.571.258.195.94329
✓64.263.948.791.125947
✓✓64.171.452.395.93089
✓✓73.874.461.696.63203
✓✓✓65.774.154.896.12469
", + "bbox": [ + 504, + 89, + 890, + 174 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results of the ablation experiment on SportsMOT test set. IoU means computing IoU between detections and the last location of existing tracks for association, while Motion means using Kalman filter to predict the location of tracks. The models are trained on SportsMOT training and validation set.", + "bbox": [ + 496, + 180, + 893, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pure motion-based and appearance-based association, suggesting that both motion and appearance cues should be considered jointly for optimal results.", + "bbox": [ + 496, + 260, + 890, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our analysis of the three categories reveals that appearance cues provide the most significant improvement for football videos (+1.3), followed by volleyball (+0.9) and basketball (+0.3). Combining this with Fig. 2, which indicates that the adjacent IoU of football games is the smallest (i.e., motion is fastest) among the three categories, we can conclude that scenes with faster motion are more dependent on appearance cues.", + "bbox": [ + 496, + 306, + 890, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study on MixSort. We ablate important components of our tracker (MixSort based on ByteTrack) including IoU, Motion (Kalman Filter) and MixSort on SportsMOT test set. The results are presented in Tab. 5. Surprisingly, we found that simple IoU without using motion prediction outperformed IoU with motion prediction by a large margin (from 64.1 HOTA to 71.5 HOTA), indicating that the Kalman filter, which assumes linear motion models, performed poorly on SportsMOT, where the motion patterns are far more complex than in previous datasets.", + "bbox": [ + 496, + 436, + 890, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we observed that MixSort played a crucial role in boosting the performance of the tracker significantly. By fusing the IoU and MixSort cues, our method achieved the best performance of 73.8 HOTA compared to 71.5 HOTA of simple IoU in our experiments.", + "bbox": [ + 496, + 588, + 890, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on different categories of SportsMOT. While the three SportsMOT categories share some common characteristics such as fast motion and similar appearance, they also have distinct features due to the different types of games. In this section, we analyze the results of our experiments on the three categories.", + "bbox": [ + 496, + 672, + 890, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first use the best HOTA metric from Tab. 4 to represent the overall difficulty of a category. Based on this metric, we find that basketball videos (66.17) are the most difficult, followed by football (73.19), and finally volleyball (76.91).", + "bbox": [ + 496, + 763, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we consider the appearance-based association (i.e. $\\alpha = 0$ in Tab. 4). We observe a notable gap between the HOTA metrics of the three categories, where basketball (26.94) has a much lower HOTA than volleyball (40.94) and football (65.65). Similarly, basketball remains the most dif", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9927", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/661f297b75b98168e34e939782c8ba6d1382fb3eccf786427fe1532acc4a4e1b.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Comparison of the state-of-the-art methods under the \"private detector\" protocol on MOT17-test set and MOT17-val set." + ], + "table_body": "
MOT17-testMOT17-val
HOTA↑IDF1↑AssA↑MOTA↑DetA↑IDs↓HOTA↑IDF1↑AssA↑MOTA↑DetA↑IDs↓
QDTrack [29]53.966.352.768.755.63378------
MOTR [43]57.268.455.871.958.92115------
GTR [47]59.171.557.075.361.6285963.075.966.271.360.4-
ByteTrack [44]63.177.362.080.364.52196-79.7-76.7-159
OC-SORT [6]63.277.563.478.063.2195068.079.369.977.9--
MixSort-Byte64.078.764.279.364.1223569.481.171.379.968.2155
MixSort-OC63.477.863.278.963.8150969.280.671.578.967.4135
", + "bbox": [ + 81, + 88, + 893, + 232 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/61b5d2727d778ab042e9df19bdc3b39655adcf02fc0c19da4ee435276cd9e999.jpg", + "table_caption": [], + "table_footnote": [ + "Table 7. Comparison of ByteTrack and MixSort-Byte on DanceTrack validation set. For MixSort-Byte, the fuse parameter $\\alpha$ is 0.9, which results in the highest HOTA among $\\{0.6, 0.7, 0.8, 0.9, 0.95\\}$ . The models are trained on DanceTrack training set." + ], + "table_body": "
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
ByteTrack47.151.931.588.270.5
MixSort-Byte46.753.031.985.868.6
", + "bbox": [ + 76, + 263, + 470, + 321 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eb961e073f1e181486560719f633b935881de90317bedecf40a9c0bff59af2f7.jpg", + "table_caption": [], + "table_footnote": [ + "Table 8. Results of MixSort-Byte on SportsMOT and SoccerNet test set. The models are trained on SportsMOT training and validation set and SoccerNet training set respectively and the hyperparameters are the same." + ], + "table_body": "
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
SportsMOToverall65.774.154.896.178.8
basketball60.867.846.897.379.1
volleyball72.587.066.896.578.7
football66.473.656.394.978.5
SoccerNet62.973.955.587.871.5
", + "bbox": [ + 81, + 388, + 467, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ficult in the motion-based association (i.e. $\\alpha = 1$ in Tab. 4), while the volleyball becomes the easiest.", + "bbox": [ + 76, + 544, + 468, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We believe that the differences in difficulty arise from several factors, including the size of the game court and the degree of physical confrontation among players. For instance, basketball scenes are played on smaller courts and involve more physical contact between players than football scenes. This can lead to more occlusion and blur in basketball videos, making the association task more challenging than in football scenes.", + "bbox": [ + 75, + 574, + 468, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with DanceTrack. We evaluate MixSort-Byte on the DanceTrack validation set and compare the results with that of ByteTrack as shown in Tab. 7. Unlike the results on SportsMOT where MixSort brings significant improvement, on DanceTrack the original ByteTrack performs better instead, with HOTA, MOTA and DetA metrics all higher than MixSort-Byte. This indicates that appearances in our proposed dataset SportsMOT are similar yet distinguishable, while those in DanceTrack are much harder to distinguish. Therefore SportsMOT highlights both the motion-based and appearance-based associations.", + "bbox": [ + 75, + 699, + 468, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with SoccerNet. We conducted experiments to compare SoccerNet that focuses only on soccer", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aac8399450c94c8ea389639d9bd2f86908d6e164afab2479dd419a26d2f799f3.jpg", + "table_caption": [], + "table_footnote": [ + "Table 9. Comparison of ByteTrack, ByteTrack with ReID model and MixSort-Byte on SportsMOT test set. The ReID model is the same as in DeepSORT [4] and all the models are trained on SportsMOT training set and the best results are shown in **bold**." + ], + "table_body": "
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
ByteTrack64.171.452.395.978.5
ByteTrack+ReID64.872.253.496.178.8
MixSort-Byte65.774.154.896.178.8
", + "bbox": [ + 500, + 263, + 893, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scenes with SportsMOT. Results shown in Tab. 8 suggest that SportsMOT is a challenging dataset with varying levels of difficulty across different sports. Specifically, basketball is proved to be the most difficult with the lowest HOTA of 60.8, while volleyball and football are relatively easier. Notably, MixSort obtains higher HOTA on overall SportsMOT than on SoccerNet, which is mainly because all the elements on the court are to be tracked in SoccerNet, leading to more false detections and much lower DetA (71.5 vs. 78.8). Yet, with a higher AssA on SoccerNet despite these false detections, it underscores SportsMOT's challenging association and significance for sports tracking.", + "bbox": [ + 496, + 401, + 893, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MixSort vs. ReID models. To validate the effectiveness of the MixFormer-like model in MixSort, we present a comparison with ReID models in Tab. 9. The results show that ByteTrack with the ReID model outperforms the original ByteTrack, highlighting the value of appearance-based association on SportsMOT. Notably, MixSort-Byte further improves upon ByteTrack with ReID by 0.9, 1.9, and 1.4 in HOTA, IDF1, and AssA. This underscores MixSort's superior appearance modeling, capable of extracting broader and more discriminative representations and facilitating efficient offline learning.", + "bbox": [ + 496, + 587, + 893, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More results on MOT17 The propose MixSort can be applied to any trackers following this paradigm and achieve state-of-the-art performance on SportsMOT. To further validate the effectiveness of MixSort, we compare MixSort with the prevailing trackers on MOT17 validation set and test set under the private detection protocol in Tab. 6. MixSort-Byte and MixSort-OC outperform these trackers in HOTA, IDF1.", + "bbox": [ + 496, + 758, + 893, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detailed Challenging attributes of SportsMOT. For a comprehensive statistical analysis and performance evalua", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9928", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e157b0e3a1e26d2b3563c1e1448f68ae26a0e19a9576c60c21e55fae15e6e7ce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Attr.Definition of sequenceAttr.Definition of target box
CMAbrupt motion of the cameraLRThe target box is smaller than a threshold
OVThe target completely leaves the video frameFOCThe area of target box intersecting with others is larger than 0.9
MBThe target region is blurred due to target or camera motionPOCThe area of target box intersecting with others is in range [0.4, 0.9]
FMThe total average IoU of target boxes in adjacent frames is smaller than a thresholdDEFThe sum of the relative changes in the length and width of the target boxes in adjacent frames is greater than a threshold
", + "bbox": [ + 76, + 89, + 890, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 10. Definition of attributes. For the attributes in the right half, we first calculate the proportion of boxes in each sequence that meet the criteria defined in the table. If this proportion exceeds a certain threshold, we consider the sequence to exhibit that attribute.", + "bbox": [ + 76, + 194, + 892, + 223 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5783a89997d202b754f3a9522ff5a26882a5a3a8b3faed16a52d02743e75962e.jpg", + "image_caption": [ + "(a) Total" + ], + "image_footnote": [], + "bbox": [ + 80, + 232, + 256, + 339 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/27f7dfbc3f9b4352efb65f2c4c065077efbc316870797f9e725588a64e0cbfc9.jpg", + "image_caption": [ + "(b) Category-wise" + ], + "image_footnote": [], + "bbox": [ + 259, + 232, + 892, + 339 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0e422629dbcde8ff0be1e8f3c96621cf300de85f01125fe2607b6724e5026aea.jpg", + "image_caption": [ + "(a) Total", + "Figure 8. Average HOTA of sequences in each attribute on SportsMOT dataset. MixSort-OC is employed for analysis." + ], + "image_footnote": [], + "bbox": [ + 80, + 377, + 256, + 484 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b075346d36606c1fafc8fb953395c42679dcb504ebc501fa02522599071d1088.jpg", + "image_caption": [ + "Figure 7. Distribution of sequences in each attribute on SportsMOT dataset.", + "(b) Category-wise" + ], + "image_footnote": [], + "bbox": [ + 261, + 377, + 892, + 484 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tion of the SportsMOT dataset, we categorize and present eight distinct challenging attributes: camera motion (CM), out of view (OV), motion blur (MB), fast motion (FM), low resolution (LR), full occlusion (FOC), partial occlusion (POC), and large deformation (DEF). The definitions for each attribute are specified in Tab. 10, with their corresponding distribution illustrated in Fig. 7. Human selection is used to identify the CM, OV, and MB attributes, while metrics calculation helps determine the others. The diversity and complexity of the scenarios featured in Fig. 7a suggest that our three chosen sports games provide a representative snapshot of broader sports analysis based on tracking states. Furthermore, in comparison to SoccerNetTracking [7], our SportsMOT offers a more diverse set of challenges observed in sports. These challenges range from the frequent occlusion seen in basketball and volleyball, to the low resolution typical of football, as depicted in Fig. 7b. We also assess the difficulty level posed by different attributes in the following.", + "bbox": [ + 75, + 526, + 472, + 815 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Difficulty analysis based on HOTA scores. We employ the average HOTA scores of sequences corresponding to each attribute as an indicator of that attribute's difficulty. The results are depicted in Fig. 8. The low HOTA score for full occlusion, evident in Fig. 8a, highlights its challeng-", + "bbox": [ + 73, + 824, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ing nature in MOT within the sports field. Furthermore, an analysis of average HOTA scores shows that basketball consistently has lower scores than volleyball and football for all attributes (Fig. 8b), indicating its increased complexity.", + "bbox": [ + 496, + 526, + 890, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 604, + 617, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we have introduced SportsMOT, a large-scale multi-object tracking dataset in sports scenes. SportsMOT is characterized with two key properties: 1) fast and variable-speed motion and 2) similar yet distinguishable appearance. We have empirically investigated several prevailing MOT trackers on the SportsMOT dataset. We have also proposed a new MOT framework MixSort, introducing a MixFormer-like association module. Hopefully, SportsMOT can provide a platform for facilitating both sports analysis and multi-object tracking.", + "bbox": [ + 496, + 631, + 890, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization.", + "bbox": [ + 496, + 794, + 892, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9929", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 1", + "[2] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 941-951, 2019. 1", + "[3] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008. 5", + "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 1, 3, 8", + "[5] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In 2017 14th IEEE international conference on advanced video and signal based surveillance (AVSS), pages 1-6. IEEE, 2017. 3", + "[6] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9686-9696, 2023. 1, 2, 3, 5, 6, 7, 8", + "[7] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3491-3502, 2022. 3, 9", + "[8] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13608-13618, 2022. 2, 5", + "[9] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 3", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7", + "[11] Yunhao Du, Yang Song, Bo Yang, and Yanyun Zhao. Strongsort: Make deepsort great again. arXiv preprint arXiv:2202.13514, 2022. 1", + "[12] James Ferryman and Ali Shahrokni. Pets2009: Dataset and challenge. In 2009 Twelfth IEEE international workshop on performance evaluation of tracking and surveillance, pages 1-6. IEEE, 2009. 3", + "[13] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 3", + "[15] Sion Hannuna, Massimo Camplani, Jake Hall, Majid Mirmehdi, Dima Damen, Tilo Burghardt, Adeline Paiement, and Lili Tao. Ds-kcf: a real-time tracker for rgb-d data. Journal of Real-Time Image Processing, 16(5):1439-1458, 2019. 3", + "[16] Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82(1):35-45, 1960. 1", + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3", + "[19] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 6", + "[20] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942, 2015. 3", + "[21] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13536-13545, October 2021. 3", + "[22] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3", + "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7", + "[25] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129(2):548-578, 2021. 5", + "[26] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8844-8854, 2022. 3", + "[27] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016. 1, 2, 3", + "[28] Matthias Müller, Adel Bibi, Silvio Giancola, Salman Al-Subaihi, and Bernard Ghanem. Trackingnet: A large-scale" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9930", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dataset and benchmark for object tracking in the wild. In Proceedings of the European Conference on Computer Vision, ECCV, 2018. 6", + "[29] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 164-173, 2021. 6, 8", + "[30] Jinlong Peng, Changan Wang, Fangbin Wan, Yang Wu, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Chained-tracker: Chaining paired attentive regression results for end-to-end joint multiple-object detection and tracking. In European conference on computer vision, pages 145-161. Springer, 2020. 1", + "[31] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 5", + "[32] Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xiangyu Zhang, and Jian Sun. Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123, 2018. 6, 7", + "[33] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1, 3", + "[34] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 6", + "[35] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 4, 5", + "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[37] Qiang Wang, Yun Zheng, Pan Pan, and Yinghui Xu. Multiple object tracking with correlation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3876-3886, 2021. 1", + "[38] Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In European Conference on Computer Vision, pages 107-122. Springer, 2020. 1", + "[39] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017. 1, 3", + "[40] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12352-12361, 2021. 1", + "[41] Yihong Xu, Yutong Ban, Guillaume Delorme, Chuang Gan, Daniela Rus, and Xavier Alameda-Pineda. Transcenter:" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Transformers with dense queries for multiple-object tracking. arXiv preprint arXiv:2103.15145, 2021. 1", + "[42] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 3", + "[43] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision, pages 659–675. Springer, 2022. 3, 8", + "[44] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In European Conference on Computer Vision, pages 1-21. Springer, 2022. 1, 2, 3, 5, 6, 7, 8", + "[45] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129(11):3069-3087, 2021. 3, 6", + "[46] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In European Conference on Computer Vision, pages 474-490. Springer, 2020. 3, 6", + "[47] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8771-8780, 2022. 6, 8" + ], + "bbox": [ + 501, + 92, + 893, + 486 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9931", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/4b8d0490-50f4-45bc-9fad-a0ac515c1ca0_model.json b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/4b8d0490-50f4-45bc-9fad-a0ac515c1ca0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..652e3ce9bd1b8b95ea72405ae060c566f4feffa5 --- /dev/null +++ b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/4b8d0490-50f4-45bc-9fad-a0ac515c1ca0_model.json @@ -0,0 +1,2653 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.087, + 0.131, + 0.885, + 0.154 + ], + "angle": 0, + "content": "SportsMOT: A Large Multi-Object Tracking Dataset in Multiple Sports Scenes" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.181, + 0.873, + 0.222 + ], + "angle": 0, + "content": "Yutao Cui* Chenkai Zeng* Xiaoyu Zhao* Yichun Yang* Gangshan Wu Limin Wang \\(^{\\text{☑}}\\) State Key Laboratory for Novel Software Technology, Nanjing University, China" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.228, + 0.698, + 0.246 + ], + "angle": 0, + "content": "https://github.com/MCG-NJU/SportsMOT" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.295, + 0.314, + 0.312 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.473, + 0.735 + ], + "angle": 0, + "content": "Multi-object tracking (MOT) in sports scenes plays a critical role in gathering players statistics, supporting further applications, such as automatic tactical analysis. Yet existing MOT benchmarks cast little attention on this domain. In this work, we present a new large-scale multi-object tracking dataset in multiple sports scenes, coined as SportsMOT, where all players on the court are supposed to be tracked. It consists of 240 video sequences, over 150K frames (almost \\(15 \\times\\) MOT17) and over 1.6M bounding boxes (\\(3 \\times\\) MOT17) collected from 3 sports categories, including basketball, volleyball and football. Our dataset is characterized with two key properties: 1) fast and variable-speed motion and 2) similar yet distinguishable appearance. We expect SportsMOT to encourage the MOT trackers to promote in both motion-based association and appearance-based association. We benchmark several state-of-the-art trackers and reveal the key challenge of SportsMOT lies in object association. To alleviate the issue, we further propose a new multi-object tracking framework, termed as MixSort, introducing a MixFormer-like structure as an auxiliary association model to prevailing tracking-by-detection trackers. By integrating the customized appearance-based association with the original motion-based association, MixSort achieves state-of-the-art performance on SportsMOT and MOT17. Based on MixSort, we give an in-depth analysis and provide some profound insights into SportsMOT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.21, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Multi-object tracking (MOT) has been a fundamental computer vision task for recent decades, aiming to locate the objects and associate them in video sequences. Researchers have cast much focus on various practical use cases like crowded street scenes [9, 27], static dancing scenes [33] and driving scenarios [14], achieving consid-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.281, + 0.892, + 0.357 + ], + "angle": 0, + "content": "erable progress [2, 4, 30, 37, 38, 40, 41] in MOT. MOT for sports scenes however is overlooked, where typically only the players on the court should be tracked for further analysis, such as counting the players' running distance or average speed and automatic tactical analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Generally, prevailing state-of-the-art trackers [1,6,11,39, 44] consist of several components to accomplish the tracking task: objects localization module, motion based objects association module and appearance based association module. Biased to the data distribution of specific human tracking benchmarks, e.g. MOT17 [27], MOT20 [9] and DanceTrack [33], the components of these trackers have difficulty adapting to sports scenes. Firstly, motivated by surveillance or self-driving applications, current human tracking benchmarks provide tracks for almost all persons in the scenes. While for sports scenes like basketball or football games, generally only the players on the court are what we focus on, hence a specialized training platform is required to make the detectors suitable for sports scenes. More importantly, in MOT17 and MOT20, these trackers highlight Kalman Filter [16] based IoU matching for object association, due to the slow and regular motion of pedestrians. DanceTrack highlights diverse motion rather than fast movement [33], that is, dancers frequently switch the motion direction and relative position. However in sports scenes, we observe fast and variable-speed movement of objects on adjacent frames, i.e. players usually possess high speed and frequently change their running speed in professional sports events, thus constituting barriers in existing motion based association. For instance, as visualized in Fig. 1, the adjacent IoU and Kalman-Filter-based IoU in sports scenes remain lower than that on MOT17 and DanceTrack (More detailed comparison can be found in Fig. 2 and Fig. 3). As a consequence, more suitable motion based association for sports scenes is required. Additionally, compared to the MOT17 and MOT20 datasets in street scenes, the objects appearances in sports scenes are less distinguishable, since not only the players inherently are in similar clothes but also the players are frequently blurred caused by fast camera motion or targets motion. Different from DanceTrack, where generally the dancers are in almost the same clothes and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.888, + 0.359, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. ☑ : Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "9921" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.472, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.155, + 0.343, + 0.166 + ], + "angle": 0, + "content": "(a) Basketball in SportsMOT" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.09, + 0.865, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.156, + 0.706, + 0.166 + ], + "angle": 0, + "content": "(c) MOT17" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.168, + 0.472, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.235, + 0.339, + 0.246 + ], + "angle": 0, + "content": "(b) Football in SportsMOT" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.168, + 0.865, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.235, + 0.715, + 0.246 + ], + "angle": 0, + "content": "(d) DanceTrack" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.25, + 0.472, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.324, + 0.351, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.324, + 0.364, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.324, + 0.38, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.324, + 0.393, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.324, + 0.416, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.324, + 0.43, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.324, + 0.444, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.324, + 0.46, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.324, + 0.474, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.25, + 0.79, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.324, + 0.54, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.324, + 0.554, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.324, + 0.568, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.324, + 0.583, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.324, + 0.597, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.324, + 0.613, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.324, + 0.631, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.895, + 0.412 + ], + "angle": 0, + "content": "Figure 1. Sampled sequences from the categories of basketball and football of SportsMOT, MOT17 and DanceTrack. There exist two key properties of SportsMOT: 1) fast and variable-speed motion, i.e. players usually possess high speed and frequently change their running speed (the visualized adjacent IoU and Kalman Filter based adjacent IoU can indicate the property); 2) similar yet distinguishable appearance, that is, players in sports scenes inherently wear jackets with different numbers and usually display distinct postures. We expect SportsMOT to encourage the MOT trackers to promote in both motion-based association and appearance based association." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.51 + ], + "angle": 0, + "content": "thus having indistinguishable appearance, players in sports scenes inherently wear jackets with different numbers and usually display distinct postures. Thereby, we argue that objects in sports scenes are with similar yet distinguishable appearance, which necessitates the appearance model developing more discriminative and extensive representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.471, + 0.82 + ], + "angle": 0, + "content": "Considering the analysis above, to advance the development of tracking and sports analysis, we propose a multi-object tracking dataset in sports scenes, termed as SportsMOT. The dataset is large-scale, high-quality and contains dense annotations for every player on the court in various sports scenes. It consists of 240 videos, over 150K frames (almost \\(15 \\times\\) MOT17 [27]) and over 1.6M bounding boxes (\\(3 \\times\\) MOT17 [27]) collected from 3 categories of sports, including basketball, volleyball and football. To provide a platform for making the trackers suitable for sports scenes, we split the dataset into training, validation and test subsets, consisting of 45, 45 and 150 video sequences respectively. There exist two core properties of SportsMOT: (1) fast and variable-speed motion, requiring more suitable motion modeling association; (2) similar yet distinguishable appearance, which necessitates the appearance model developing more discriminative and extensive representations. Altogether, we expect SportsMOT to encourage the trackers to promote in both the certain aspects, i.e. motion based association and appearance based association." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Given the large-scale multi-object tracking dataset SportsMOT, we benchmark some recent tracking approaches and retrain all of them on the training split. We observe IDF1 and AssA metrics are lower than that on MOT17 while the DetA is quite high, indicating that the main chal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.895, + 0.585 + ], + "angle": 0, + "content": "lenge of SportsMOT lies in objects association rather than objects localization. To alleviate the issue, we propose a new multi-object tracking framework, dubbed as MixSort, with introducing a MixFormer-like [8] structure as appearance based association to prevailing tracking-by-detection trackers (e.g. ByteTrack [44], OC-SORT [6]). By integrating the original motion based objects association and the designed appearance based association, the performance gets boosted on both SportsMOT and MOT17 benchmarks. Based on MixSort, we perform extensive exploration studies and provide some profound insights into SportsMOT." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.589, + 0.86, + 0.603 + ], + "angle": 0, + "content": "The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.615, + 0.892, + 0.707 + ], + "angle": 0, + "content": "- We build a new large-scale multi-object tracking dataset in diverse sports scenes, SportsMOT, equipped with two key properties of 1) fast and variable-speed motion and 2) similar yet distinguishable appearance, aiming to advance the development of both tracking and sports analysis." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.72, + 0.892, + 0.78 + ], + "angle": 0, + "content": "- We benchmark some prevailing trackers on SportsMOT, which reveals that the key challenges lie in objects association and hopefully can facilitate further research." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- We propose a new multi-object tracking framework MixSort, with introducing a MixFormer-like structure as appearance based association model to prevailing tracking-by-detection trackers, so as to boost the objects association. Based on MixSort, we perform extensive studies and provide some profound insights into SportsMOT." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.615, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9922" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.164 + ], + "angle": 0, + "content": "
DatasetVideosFramesLength (s)BboxTracks
MOT171411,235463292,7331,342
MOT20813,4105351,652,0403,456
DanceTrack100105,8555,292-990
SportsMOT240150,3796,0151,629,4903,401
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.169, + 0.468, + 0.197 + ], + "angle": 0, + "content": "Table 1. Comparison of statistics between existing human MOT datasets and our SportsMOT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.202, + 0.218, + 0.218 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.228, + 0.47, + 0.681 + ], + "angle": 0, + "content": "Multi-object tracking datasets. Existing Multi-object tracking datasets usually focus on different scenes, such as autonomous driving, pedestrians on roads, and dancing. For autonomous driving, there are KITTI [14], KITTI360 [22] and BDD100K [42], which focus on pedestrians and vehicles. Besides, other datasets, focusing on only the pedestrians, collect the videos from static and moving cameras. One of the earliest is PETS [12], but it's too simple in some scenes. MOT15 [20] proposes the first large-scale benchmark for Multi-object tracking, followed by MOT17 [27] and MOT20 [9]. It's worth noting that MOT20 focuses on extremely crowded scenes where many pedestrians are occluded, increasing the tracking difficulty greatly in both detection and association. Recently, DanceTrack [33], focusing on dancing scenes, is proposed to encourage trackers to rely less on visual discrimination and depend more on motion analysis. The emphasized properties are uniform appearance and diverse motion. While in SportsMOT, the appearance is similar yet distinguishable, and the players' motion is fast and with variable speed. We expect SportsMOT to encourage algorithms to promote in both appearance and motion association. Besides, SoccerNet-Tracking [7] is presented to track elements in football scenarios. The main difference lies in that, it only contains soccer scenes and tracks almost all elements (players, goalkeepers, referees, balls) on the court without distinction. While SportsMOT contains three types of sports where the objects in basketball scenes are more crowded and thus more challenging (refer to Section 5.3), and only focuses on the players to support further statistics and tactical analysis." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Object association in tracking. Association is a very important task in tracking, where trackers need to associate detections in new frames with existing tracks. For most of the trackers, a similarity matrix (or cost matrix) between new detections and tracks is computed based commonly on motion and appearance cues, which is later fed into Hungarian algorithm [18] to perform association. For example, SORT [4] uses Kalman Filter to predict the location of objects and computes the IoU of detected and predicted bounding boxes as similarity matrix. IOU-Tracker [5] directly computes the IoU without prediction. ByteTrack [44] adds an association phase for detection with low confidence score, which can boost the performance. OC-SORT [6] tries to address the limitation of Kalman Filter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Appearance cues also play an important role in association. DeepSORT [39] crops the detection from frame images, which are then used by networks to generate re-ID features. Then the motion cues and distance of re-ID features are fused to perform association. FairMOT [45] uses a re-ID branch on a backbone shared by the detection branch to generate re-ID features. In CenterTrack [46], the previous frame is used to help the prediction of tracks. Recently, Transformer [36] is used by some work such as TrackFormer [26] and MOTR [43] to boost the association quality. Our proposed MixSort integrates the motion and appearance cues with a motion modeling component and the designed MixFormer-like structure respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.302, + 0.694, + 0.318 + ], + "angle": 0, + "content": "3. SportsMOT Dataset" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.326, + 0.7, + 0.341 + ], + "angle": 0, + "content": "3.1. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.892, + 0.575 + ], + "angle": 0, + "content": "Video Collection. We select three worldwide famous sports, football, basketball, and volleyball, and collect videos of high-quality professional games including NCAA, Premier League, and Olympics from Multi-Sports [21], which is a large dataset in sports area focusing on spatio-temporal action localization. Each category has typical players' formations and motion patterns, and they can effectively represent the diversity of sports scenarios. Only the overhead shots of sports game scenes are used, guaranteeing certain extreme situations do not occur. The proposed dataset consists of 240 video sequences in total, each of which is 720P and 25 FPS. Following the principles of multi-object tracking, each video clip is manually checked to ensure that there are no abrupt viewpoint switches within the video." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.58, + 0.89, + 0.611 + ], + "angle": 0, + "content": "Annotation Pipeline. We annotate the collected videos according to the following guidelines." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.613, + 0.89, + 0.657 + ], + "angle": 0, + "content": "- The entire athlete's limbs and torso, excluding any other objects like balls touching the athlete's body, are required to be annotated." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.661, + 0.89, + 0.736 + ], + "angle": 0, + "content": "- The annotators are asked to predict the bounding box of the athlete in the case of occlusion, as long as the athletes have a visible part of body. However, if half of the athletes' torso is outside the view, annotators should just skip them." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.74, + 0.89, + 0.769 + ], + "angle": 0, + "content": "- We ask the annotators to confirm that each player has a unique ID throughout the whole clip." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.613, + 0.89, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We provide a customized labeling tool for SportsMOT and a corresponding manual book to annotators. Once they start annotating a new object, the labeling tool automatically assigns a new ID to the object and propagates the bounding box of previous state to the current state, with the help of the single object tracker KCF [15]. Then the generated bounding boxes should be refined by the annotators, so as to improve annotation quality. After carefully reviewing each" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9923" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.275, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.092, + 0.465, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.199, + 0.47, + 0.268 + ], + "angle": 0, + "content": "Figure 2. IoU on adjacent frames. (a) Compared to MOT17 and DanceTrack, SportsMOT has a lower score, indicating that objects have faster motion. (b) In SportsMOT, the category of football has the lowest IoU score, which means that football players often have fast motion." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.276, + 0.468, + 0.37 + ], + "angle": 0, + "content": "
CategoryFramesTracksTrack gap len.Track len.Bboxes per frame
Basketball845.41068.7767.99.1
Volleyball360.41238.2335.911.2
Football673.920.5116.1422.112.8
Total626.614.296.6479.110.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.375, + 0.466, + 0.391 + ], + "angle": 0, + "content": "Table 2. Detailed statistics of the three categories in SportsMOT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.469, + 0.456 + ], + "angle": 0, + "content": "annotation result, we refine the bounding boxes and IDs that do not satisfy the standards, hence building a high-quality dataset. Finally, the bounding boxes with too small size, i.e. \\( w < 5 \\) or \\( h < 5 \\), are deleted." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.239, + 0.48 + ], + "angle": 0, + "content": "3.2. Dataset Statistic" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.761 + ], + "angle": 0, + "content": "Overview. SportsMOT is a large-scale and high-quality MOT dataset, aiming to advance the development of both sports analysis and multi-object tracking. Table 1 compares the statistics of SportsMOT with the prevailing human tracking datasets, including MOT17, MOT20 and DanceTrack. According to the statistics, SportsMOT has a large number of bounding boxes of over 1.6M, which is comparable to MOT20 and significantly larger than MOT17. Besides, SportsMOT has a large number of video clips, tracks \\((2.5 \\times \\text{MOT17}, 3.4 \\times \\text{DanceTrack})\\), frames \\((13.4 \\times \\text{MOT17})\\). As shown in Table 2, we also compare the basic statistics of each category of SportsMOT. SportsMOT solely provides fine annotations of the players on the court, which are supposed to be tracked for further analysis. To provide a platform for making the trackers suitable for sports scenes, we split the dataset into training, validation and test subsets, consisting of 45, 45 and 150 video sequences respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Fast and Variable-Speed Motion. Motion cues play an important role in object association for multi-object tracking. The existing human-tracking datasets (e.g. MOT17, MOT20 and DanceTrack) generally have certain motion patterns that are distinct with sports scenes, constituting barriers in players tracking. For instance, in MOT17 and MOT20, pedestrians are featured by linear motion with constant speed, which is easily hacked by association strategies with constant velocity assumption. Besides, Dance" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.701, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.186, + 0.629, + 0.196 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.092, + 0.887, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.186, + 0.816, + 0.196 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.266 + ], + "angle": 0, + "content": "Figure 3. Kalman-Filter-based IoU on adjacent frames. (a) Compared to MOT17 and DanceTrack, SportsMOT has a lower score, indicating that objects have more variable-speed motion. (b) In SportsMOT, the category of football has the lowest Kalman-Filter-based IoU score." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.274, + 0.63, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.349, + 0.598, + 0.362 + ], + "angle": 0, + "content": "(a) MOT17" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.275, + 0.758, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.349, + 0.735, + 0.362 + ], + "angle": 0, + "content": "(b) SportsMOT" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.275, + 0.886, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.349, + 0.864, + 0.361 + ], + "angle": 0, + "content": "(c) DanceTrack" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Figure 4. Visualization of re-ID features from sampled videos in MOT17, SportsMOT and DanceTrack dataset using t-SNE [35]. The same object is coded by the same color. It indicates that object appearance of SportsMOT is less distinguishable than that of MOT17, while more distinguishable than that of DanceTrack. We expect the appearance model to capture more discriminative and extensive representation for object association." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Track highlights diverse motion rather than fast motion, that is, dancers usually move in more diverse directions with relatively low speed. In contrast, SportsMOT has distinct motion patterns, i.e. fast and variable-speed motion, where the players typically move fast with their running speed or camera speed frequently changing. As illustrated in Fig. 2, among the three datasets, SportsMOT has the lowest IoU score of the objects bounding boxes on adjacent frames, indicating the fast movement. We use the ground truth of previous frames for Kalman Filter prediction. The result and current ground truth are used to calculate Kalman-Filter-based IoU. Seen from Fig. 3, SportsMOT also has the lowest Kalman-Filter-based IoU score on adjacent frames, which suggests that the motion can not be easily modeled by prevailing methods due to the variable-speed movement. Specifically, football has the smallest adjacent IoU and Kalman-Filter-based IoU, which is closely related to the fast running speed, abrupt acceleration or stops. It poses a major challenge for trackers based on simple motion assumptions and also encourages them to model object motion in more dynamic and adaptive ways." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Similar yet Distinguishable Appearance. Object appearance is another kind of cue on which MOT trackers often rely to distinguish different objects. In MOT17 and MOT20, pedestrians are usually distinct in body size and wear different clothes, yielding discriminative visual features. In contrast, the objects in DanceTrack typically wear nearly identical outfits, leading to indistinguishable" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9924" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.466, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.201, + 0.473, + 0.287 + ], + "angle": 0, + "content": "Figure 5. Paradigm for computing visual similarity matrix of tracks and detections in MixSort. The center of ground-truth target detection is marked with red dots, the others are green. The blue dashed box indicates the cropped search region. The blank part of similarity vector means that for detections not in the search region, the corresponding value is set to 0." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.473, + 0.492 + ], + "angle": 0, + "content": "appearances. Thereby, DanceTrack highlights solely relying on motion-based association rather than appearance-based association. In SportsMOT, the players also have very similar appearances. However, the players wear jackets with different numbers and usually display distinct postures, thus resulting in similar yet distinguishable appearance. In Fig. 4, we provide visualization of re-ID feature from sampled videos in MOT17, SportsMOT and DanceTrack dataset using t-SNE [35]. It implies that the re-ID features in SportsMOT are similar yet distinguishable compared to MOT17 and DanceTrack. We aim to encourage the trackers to learn more discriminative visual representations for more robust object association." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Evaluation Metrics. MOTA [3] is the main metric for existing MOT evaluations. However, MOTA focuses more on measuring the accuracy of detection. To highlight the performance of object association, we recommend HOTA [25], AssA and IDF1 [31] as the major evaluation metrics in SportsMOT dataset. HOTA aims to measure the accuracy of detection and association equally." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.615, + 0.424, + 0.634 + ], + "angle": 0, + "content": "4. Multi-Object Tracking on SportsMOT" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.732 + ], + "angle": 0, + "content": "In this section, we present our proposed multi-object tracking framework, called MixSort. This framework is designed to enhance the appearance-based association performance and can be applied to any trackers that follow the tracking-by-detection paradigm, such as ByteTrack [44] and OC-SORT [6]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.47, + 0.759 + ], + "angle": 0, + "content": "4.1. MixFormer for Appearance-based Association" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "MixFormer. In this paragraph, we discuss the use of Mix-Former [8] in our proposed framework MixSort. Mix-Former is designed to extract target-specific discriminative features and perform extensive communication between the target and search area, therefore, it is the key component that enables MixSort to compute visual similarities between the templates of tracked objects and detected objects in the search region of the current frames. The original Mix-Former uses a corner-based localization head to predict the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.211, + 0.894, + 0.241 + ], + "angle": 0, + "content": "Figure 6. The pipeline of MixSort. We use motion model and MixFormer to generate fused similarity matrix for association." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.895, + 0.401 + ], + "angle": 0, + "content": "top-left and bottom-right corners of the input template in the search region. However, we modify the corner head by using a heatmap prediction head that predicts the center of the template and generates a confidence heatmap. This allows us to compute the similarity between the detection and the template. To make MixSort suitable for multi-object tracking and accelerate inference speed, we reduce the number of mixed attention modules in MixFormer from 12 to 4. The steps involved in computing the visual similarity matrix are illustrated in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.892, + 0.54 + ], + "angle": 0, + "content": "Association Strategy. In order to perform association between detections and existing tracks, we use a mixed similarity matrix generated by computing the visual similarity between the target template and the detected objects in the search region of the current frame. Specifically, we obtain the heatmap response at the center of each detection as its visual similarity to the template. The resulting similarity matrix is then combined with the IoU matrix using the Hungarian Algorithm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.631 + ], + "angle": 0, + "content": "To start, for each existing track \\( t \\), we use the Kalman Filter to predict its new location. Then, we crop the current frame centered at the predicted location with a certain scale to obtain the search region \\( s \\). By feeding \\( s \\) and the template \\( t \\) into MixFormer, we generate a heatmap \\( \\mathcal{H} \\) that represents the similarity between the template and search region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.892, + 0.707 + ], + "angle": 0, + "content": "Next, for each detection \\(d\\) whose center is in the search region \\(s\\), we set its similarity to track \\(t\\) as the response in the heatmap \\(\\mathcal{H}\\); the similarity values of other detections are set to 0. Finally, we fuse the visual similarity and the IoU score to obtain the mixed similarity matrix" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.719, + 0.892, + 0.736 + ], + "angle": 0, + "content": "\\[\nM = \\alpha \\cdot \\operatorname {I o U} + (1 - \\alpha) \\cdot V \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.747, + 0.892, + 0.779 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the weight coefficient and \\(V\\) represents the visual similarity matrix calculated using MixFormer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.676, + 0.804 + ], + "angle": 0, + "content": "4.2. MixSort Tracking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Based on the tracking-by-detection paradigm, the pipeline of MixSort can be generalized as follows. As shown in Figure 6, we first obtain detections using an object detector. Then, we employ a motion model (e.g., Kalman Filter) to predict new locations of existing tracks. Based on the new locations and templates of tracks, we compute a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9925" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.894, + 0.284 + ], + "angle": 0, + "content": "
Training SetupHOTA↑IDF1↑AssA↑MOTA↑DetA↑LocA↑IDs↓Frag↓
CenterTrack [46]Train62.760.048.090.882.190.8104815750
FairMOT [45]Train49.353.534.786.470.283.9992821673
QDTrack [29]Train60.462.347.290.177.588.0637711850
TransTrack [34]Train68.971.557.592.682.791.049929994
GTR [47]Train54.555.845.967.964.889.0956714525
ByteTrack [44]Train62.869.851.294.177.185.632674499
OC-SORT [6]Train71.972.259.894.586.492.430933474
ByteTrackTrain+Val64.171.452.395.978.585.730894216
OC-SORTTrain+Val73.774.061.596.588.592.727283144
MixSort-ByteTrain+Val65.7 (+1.6)74.1 (+2.7)54.8 (+2.5)96.278.885.724724009
MixSort-OCTrain+Val74.1 (+0.4)74.4 (+0.4)62.0 (+0.5)96.588.592.727813199
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.286, + 0.848, + 0.301 + ], + "angle": 0, + "content": "Table 3. Tracking performance of investigated algorithms on our proposed SportsMOT. The best results are shown in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.471, + 0.425 + ], + "angle": 0, + "content": "fused similarity matrix as described above and use it to associate tracks and detections by the Hungarian Algorithm. Finally, for matched tracks and detections, we update the online templates. For unmatched tracks, we keep them until the threshold is reached. For unmatched detections with confidence scores higher than the threshold, new tracks are initialized." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.434, + 0.292, + 0.45 + ], + "angle": 0, + "content": "4.3. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.547 + ], + "angle": 0, + "content": "Training. We only consider the training of MixFormer here since the detector remains the same as initial method (e.g. ByteTrack). The original MixFormer is trained on SOT datasets, so we first modify the format of ground truth of MOT datasets, that is, converting the ground truth trajectory of every single player into TrackingNet format [28]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.609 + ], + "angle": 0, + "content": "For each ground-truth bounding box, we compute its corresponding center location \\((c_x, c_y)\\) in the low-resolution heatmap. Following CornerNet [19], the ground-truth heatmap response is generated using 2D Gaussian kernel:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.618, + 0.47, + 0.65 + ], + "angle": 0, + "content": "\\[\nh _ {x y} = \\exp \\left(- \\frac {\\left(x - c _ {x}\\right) ^ {2} + \\left(y - c _ {y}\\right) ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.698 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is adaptive to the size of the bounding box. The training loss is a pixel-wise logistic regression with focal loss [23]:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.707, + 0.469, + 0.761 + ], + "angle": 0, + "content": "\\[\nL = - \\sum_ {x y} \\left\\{ \\begin{array}{l l} (1 - \\hat {h} _ {x y}) ^ {\\gamma} \\log (\\hat {h} _ {x y}) & , h _ {x y} = 1; \\\\ (1 - h _ {x y}) ^ {\\beta} (\\hat {h} _ {x y}) ^ {\\gamma} \\log (1 - \\hat {h} _ {x y}) & , \\text {o t h e r w i s e}. \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.792 + ], + "angle": 0, + "content": "where \\(\\gamma\\) and \\(\\beta\\) are hyper-parameters in focal loss, and we set \\(\\gamma = 2\\), \\(\\beta = 4\\) following CornerNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Inference. For each track, we maintain only one template for keeping a balance between accuracy and speed. When a detection is matched to an existing track, we directly replace the original template to the new detection, if and only if the ratio of its uncovered (i.e. overlapping with any detected objects) area is larger than a certain threshold, so as to reduce the impact of misleading representations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.318, + 0.745, + 0.335 + ], + "angle": 0, + "content": "5. Experiments and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.343, + 0.677, + 0.359 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.427 + ], + "angle": 0, + "content": "Dataset Split. In benchmark experiments, we follow the default split described in Sec. 3.2. In exploration study, we split the original MOT17 training set into two sets, used for training and validation respectively following CenterTrack." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Implementation Details. Following ByteTrack and OC-SORT, we use YOLOX [13] as our detector. Using COCO-pretrained model as the initialized weights, we first train the model on CrowdHuman [32] for 80 epochs and then train on SportsMOT for another 80 epochs. The remaining settings are the same as that in ByteTrack." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.657 + ], + "angle": 0, + "content": "For MixFormer, we initialize the backbone with the model trained on VOT datasets and then fine-tune it on SportsMOT for 300 epochs with learning rate initialized as \\(1e - 4\\) and decreased to \\(1e - 5\\) at epoch 200. The optimizer is ADAM [17] with weight decay \\(10^{-4}\\). The sizes of search images and templates are \\(224 \\times 224\\) and \\(96 \\times 96\\) respectively. The max sample interval is set to 10. For every tracking result, we apply linear interpolation as post-processing, with maximum gap set to 20." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.688, + 0.681 + ], + "angle": 0, + "content": "5.2. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluate several representative methods of three kinds on our dataset. ByteTrack [44], OC-SORT [6] and QDTrack [29] are trackers in tracking-by-detection paradigm. CenterTrack [46] and FairMOT [45] perform joint detection and tracking in one stage. TransTrack [34] and GTR [47] are trackers based on Transformer. Most of the current best multi-object tracking algorithms belong to tracking-by-detection paradigm, however, due to the separation of detection and tracking, the information cannot be shared completely. Joint-detection-and-tracking paradigm couples the two modules, with the goal of boosting the performance of each. Transformer-based-tracking methods are relatively new but have achieved great performance. Despite its huge potential, the model complexity and calcula" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9926" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.472, + 0.139 + ], + "angle": 0, + "content": "
α10.90.80.70.60.50.40.30.20.10
basketball65.966.166.265.965.364.863.660.756.747.126.9
volleyball76.076.876.576.476.576.976.475.573.769.140.9
football71.972.472.372.472.572.873.272.972.671.465.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.145, + 0.47, + 0.187 + ], + "angle": 0, + "content": "Table 4. Comparison of the HOTA metric of basketball, volley and football under different fusion parameters \\(\\alpha\\) on SportsMOT test set. The models are trained on SportsMOT training set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.197, + 0.47, + 0.226 + ], + "angle": 0, + "content": "tion cost are much higher, resulting in large memory and long training time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.469, + 0.333 + ], + "angle": 0, + "content": "All training settings including the number of epochs and change of learning rate are consistent with original papers. According to different default settings, we follow the commonly used pretraining datasets, such as CrowdHuman [32], COCO [24] and ImageNet [10], and apply SportsMOT-train with or without other datasets for finetuning for different methods. We compare the results in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.47, + 0.471 + ], + "angle": 0, + "content": "In sports scenes, the clear appearance and sparse density of objects allow current mature detection frameworks to generate bounding boxes with high accuracy. However, specialized detectors need to be trained for not detecting audience and referees. The key challenges are fast speed and motion blur, which forces us to pay more attention to improve the association performance. Besides, The wide range of HOTA and MOTA denotes SportsMOT is more distinguishable among different kinds of algorithms." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.255, + 0.499 + ], + "angle": 0, + "content": "5.3. Exploration Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.506, + 0.47, + 0.537 + ], + "angle": 0, + "content": "In this section, we perform extensive studies on the proposed MixSort and SportsMOT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.47, + 0.712 + ], + "angle": 0, + "content": "Effectiveness of the proposed association module. We evaluate the effectiveness of MixSort by applying it to two state-of-the-art trackers, OC-SORT [6] and ByteTrack [44], which follow the tracking-by-detection paradigm and useYOLOX as their detector. The evaluation is conducted on the SportsMOT test set, and the results are presented in Tab. 3. Our experiments show that MixSort significantly improves the performance of both trackers, with OC-SORT achieving a 0.4 HOTA increase and ByteTrack achieving a 1.6 HOTA increase on SportsMOT. This demonstrates the effectiveness of MixSort in enhancing the association." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Appearance-based vs. Motion-based association. We have demonstrated that MixSort can improve association performance. In this paragraph, we investigate the impact of the fusion weight \\(\\alpha\\) on the ability of appearance cues to aid in conventional motion-based association. We evaluate OC-SORT with MixSort on the three categories in the SportsMOT test set using \\(\\alpha\\) values ranging from 1 to 0 in Eq. (1). The results, presented in Tab. 4, reveal that pure motion-based association \\((\\alpha = 1)\\) outperforms pure appearance-based association \\((\\alpha = 0)\\) in all categories, underscoring the significance of motion cues in sports scenes. Moreover, fused association surpasses both" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.892, + 0.175 + ], + "angle": 0, + "content": "
IoUMotionMix.HOTA↑IDF1↑AssA↑MOTA↑IDs↓
✓71.571.258.195.94329
✓64.263.948.791.125947
✓✓64.171.452.395.93089
✓✓73.874.461.696.63203
✓✓✓65.774.154.896.12469
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.181, + 0.894, + 0.251 + ], + "angle": 0, + "content": "Table 5. Results of the ablation experiment on SportsMOT test set. IoU means computing IoU between detections and the last location of existing tracks for association, while Motion means using Kalman filter to predict the location of tracks. The models are trained on SportsMOT training and validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.306 + ], + "angle": 0, + "content": "pure motion-based and appearance-based association, suggesting that both motion and appearance cues should be considered jointly for optimal results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.429 + ], + "angle": 0, + "content": "Our analysis of the three categories reveals that appearance cues provide the most significant improvement for football videos (+1.3), followed by volleyball (+0.9) and basketball (+0.3). Combining this with Fig. 2, which indicates that the adjacent IoU of football games is the smallest (i.e., motion is fastest) among the three categories, we can conclude that scenes with faster motion are more dependent on appearance cues." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.588 + ], + "angle": 0, + "content": "Ablation study on MixSort. We ablate important components of our tracker (MixSort based on ByteTrack) including IoU, Motion (Kalman Filter) and MixSort on SportsMOT test set. The results are presented in Tab. 5. Surprisingly, we found that simple IoU without using motion prediction outperformed IoU with motion prediction by a large margin (from 64.1 HOTA to 71.5 HOTA), indicating that the Kalman filter, which assumes linear motion models, performed poorly on SportsMOT, where the motion patterns are far more complex than in previous datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.665 + ], + "angle": 0, + "content": "Furthermore, we observed that MixSort played a crucial role in boosting the performance of the tracker significantly. By fusing the IoU and MixSort cues, our method achieved the best performance of 73.8 HOTA compared to 71.5 HOTA of simple IoU in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Analysis on different categories of SportsMOT. While the three SportsMOT categories share some common characteristics such as fast motion and similar appearance, they also have distinct features due to the different types of games. In this section, we analyze the results of our experiments on the three categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.825 + ], + "angle": 0, + "content": "We first use the best HOTA metric from Tab. 4 to represent the overall difficulty of a category. Based on this metric, we find that basketball videos (66.17) are the most difficult, followed by football (73.19), and finally volleyball (76.91)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Next, we consider the appearance-based association (i.e. \\(\\alpha = 0\\) in Tab. 4). We observe a notable gap between the HOTA metrics of the three categories, where basketball (26.94) has a much lower HOTA than volleyball (40.94) and football (65.65). Similarly, basketball remains the most dif" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9927" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.233 + ], + "angle": 0, + "content": "
MOT17-testMOT17-val
HOTA↑IDF1↑AssA↑MOTA↑DetA↑IDs↓HOTA↑IDF1↑AssA↑MOTA↑DetA↑IDs↓
QDTrack [29]53.966.352.768.755.63378------
MOTR [43]57.268.455.871.958.92115------
GTR [47]59.171.557.075.361.6285963.075.966.271.360.4-
ByteTrack [44]63.177.362.080.364.52196-79.7-76.7-159
OC-SORT [6]63.277.563.478.063.2195068.079.369.977.9--
MixSort-Byte64.078.764.279.364.1223569.481.171.379.968.2155
MixSort-OC63.477.863.278.963.8150969.280.671.578.967.4135
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.103, + 0.234, + 0.867, + 0.248 + ], + "angle": 0, + "content": "Table 6. Comparison of the state-of-the-art methods under the \"private detector\" protocol on MOT17-test set and MOT17-val set." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.265, + 0.472, + 0.322 + ], + "angle": 0, + "content": "
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
ByteTrack47.151.931.588.270.5
MixSort-Byte46.753.031.985.868.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.326, + 0.47, + 0.383 + ], + "angle": 0, + "content": "Table 7. Comparison of ByteTrack and MixSort-Byte on DanceTrack validation set. For MixSort-Byte, the fuse parameter \\(\\alpha\\) is 0.9, which results in the highest HOTA among \\(\\{0.6, 0.7, 0.8, 0.9, 0.95\\}\\). The models are trained on DanceTrack training set." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.389, + 0.468, + 0.476 + ], + "angle": 0, + "content": "
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
SportsMOToverall65.774.154.896.178.8
basketball60.867.846.897.379.1
volleyball72.587.066.896.578.7
football66.473.656.394.978.5
SoccerNet62.973.955.587.871.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.48, + 0.47, + 0.536 + ], + "angle": 0, + "content": "Table 8. Results of MixSort-Byte on SportsMOT and SoccerNet test set. The models are trained on SportsMOT training and validation set and SoccerNet training set respectively and the hyperparameters are the same." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.545, + 0.469, + 0.574 + ], + "angle": 0, + "content": "ficult in the motion-based association (i.e. \\(\\alpha = 1\\) in Tab. 4), while the volleyball becomes the easiest." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.695 + ], + "angle": 0, + "content": "We believe that the differences in difficulty arise from several factors, including the size of the game court and the degree of physical confrontation among players. For instance, basketball scenes are played on smaller courts and involve more physical contact between players than football scenes. This can lead to more occlusion and blur in basketball videos, making the association task more challenging than in football scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Comparison with DanceTrack. We evaluate MixSort-Byte on the DanceTrack validation set and compare the results with that of ByteTrack as shown in Tab. 7. Unlike the results on SportsMOT where MixSort brings significant improvement, on DanceTrack the original ByteTrack performs better instead, with HOTA, MOTA and DetA metrics all higher than MixSort-Byte. This indicates that appearances in our proposed dataset SportsMOT are similar yet distinguishable, while those in DanceTrack are much harder to distinguish. Therefore SportsMOT highlights both the motion-based and appearance-based associations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Comparison with SoccerNet. We conducted experiments to compare SoccerNet that focuses only on soccer" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.265, + 0.894, + 0.332 + ], + "angle": 0, + "content": "
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
ByteTrack64.171.452.395.978.5
ByteTrack+ReID64.872.253.496.178.8
MixSort-Byte65.774.154.896.178.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.337, + 0.894, + 0.393 + ], + "angle": 0, + "content": "Table 9. Comparison of ByteTrack, ByteTrack with ReID model and MixSort-Byte on SportsMOT test set. The ReID model is the same as in DeepSORT [4] and all the models are trained on SportsMOT training set and the best results are shown in **bold**." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.895, + 0.584 + ], + "angle": 0, + "content": "scenes with SportsMOT. Results shown in Tab. 8 suggest that SportsMOT is a challenging dataset with varying levels of difficulty across different sports. Specifically, basketball is proved to be the most difficult with the lowest HOTA of 60.8, while volleyball and football are relatively easier. Notably, MixSort obtains higher HOTA on overall SportsMOT than on SoccerNet, which is mainly because all the elements on the court are to be tracked in SoccerNet, leading to more false detections and much lower DetA (71.5 vs. 78.8). Yet, with a higher AssA on SoccerNet despite these false detections, it underscores SportsMOT's challenging association and significance for sports tracking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.895, + 0.756 + ], + "angle": 0, + "content": "MixSort vs. ReID models. To validate the effectiveness of the MixFormer-like model in MixSort, we present a comparison with ReID models in Tab. 9. The results show that ByteTrack with the ReID model outperforms the original ByteTrack, highlighting the value of appearance-based association on SportsMOT. Notably, MixSort-Byte further improves upon ByteTrack with ReID by 0.9, 1.9, and 1.4 in HOTA, IDF1, and AssA. This underscores MixSort's superior appearance modeling, capable of extracting broader and more discriminative representations and facilitating efficient offline learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.759, + 0.894, + 0.867 + ], + "angle": 0, + "content": "More results on MOT17 The propose MixSort can be applied to any trackers following this paradigm and achieve state-of-the-art performance on SportsMOT. To further validate the effectiveness of MixSort, we compare MixSort with the prevailing trackers on MOT17 validation set and test set under the private detection protocol in Tab. 6. MixSort-Byte and MixSort-OC outperform these trackers in HOTA, IDF1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Detailed Challenging attributes of SportsMOT. For a comprehensive statistical analysis and performance evalua" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9928" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.09, + 0.892, + 0.193 + ], + "angle": 0, + "content": "
Attr.Definition of sequenceAttr.Definition of target box
CMAbrupt motion of the cameraLRThe target box is smaller than a threshold
OVThe target completely leaves the video frameFOCThe area of target box intersecting with others is larger than 0.9
MBThe target region is blurred due to target or camera motionPOCThe area of target box intersecting with others is in range [0.4, 0.9]
FMThe total average IoU of target boxes in adjacent frames is smaller than a thresholdDEFThe sum of the relative changes in the length and width of the target boxes in adjacent frames is greater than a threshold
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.195, + 0.893, + 0.224 + ], + "angle": 0, + "content": "Table 10. Definition of attributes. For the attributes in the right half, we first calculate the proportion of boxes in each sequence that meet the criteria defined in the table. If this proportion exceeds a certain threshold, we consider the sequence to exhibit that attribute." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.233, + 0.257, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.344, + 0.193, + 0.356 + ], + "angle": 0, + "content": "(a) Total" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.233, + 0.893, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.344, + 0.626, + 0.356 + ], + "angle": 0, + "content": "(b) Category-wise" + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.359, + 0.708, + 0.373 + ], + "angle": 0, + "content": "Figure 7. Distribution of sequences in each attribute on SportsMOT dataset." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.378, + 0.258, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.49, + 0.193, + 0.501 + ], + "angle": 0, + "content": "(a) Total" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.378, + 0.893, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.489, + 0.626, + 0.502 + ], + "angle": 0, + "content": "(b) Category-wise" + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.504, + 0.833, + 0.519 + ], + "angle": 0, + "content": "Figure 8. Average HOTA of sequences in each attribute on SportsMOT dataset. MixSort-OC is employed for analysis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.527, + 0.473, + 0.816 + ], + "angle": 0, + "content": "tion of the SportsMOT dataset, we categorize and present eight distinct challenging attributes: camera motion (CM), out of view (OV), motion blur (MB), fast motion (FM), low resolution (LR), full occlusion (FOC), partial occlusion (POC), and large deformation (DEF). The definitions for each attribute are specified in Tab. 10, with their corresponding distribution illustrated in Fig. 7. Human selection is used to identify the CM, OV, and MB attributes, while metrics calculation helps determine the others. The diversity and complexity of the scenarios featured in Fig. 7a suggest that our three chosen sports games provide a representative snapshot of broader sports analysis based on tracking states. Furthermore, in comparison to SoccerNetTracking [7], our SportsMOT offers a more diverse set of challenges observed in sports. These challenges range from the frequent occlusion seen in basketball and volleyball, to the low resolution typical of football, as depicted in Fig. 7b. We also assess the difficulty level posed by different attributes in the following." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Difficulty analysis based on HOTA scores. We employ the average HOTA scores of sequences corresponding to each attribute as an indicator of that attribute's difficulty. The results are depicted in Fig. 8. The low HOTA score for full occlusion, evident in Fig. 8a, highlights its challeng-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.589 + ], + "angle": 0, + "content": "ing nature in MOT within the sports field. Furthermore, an analysis of average HOTA scores shows that basketball consistently has lower scores than volleyball and football for all attributes (Fig. 8b), indicating its increased complexity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.619, + 0.621 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.785 + ], + "angle": 0, + "content": "In this paper, we have introduced SportsMOT, a large-scale multi-object tracking dataset in sports scenes. SportsMOT is characterized with two key properties: 1) fast and variable-speed motion and 2) similar yet distinguishable appearance. We have empirically investigated several prevailing MOT trackers on the SportsMOT dataset. We have also proposed a new MOT framework MixSort, introducing a MixFormer-like association module. Hopefully, SportsMOT can provide a platform for facilitating both sports analysis and multi-object tracking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9929" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.471, + 0.215 + ], + "angle": 0, + "content": "[2] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 941-951, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.471, + 0.271 + ], + "angle": 0, + "content": "[3] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.33 + ], + "angle": 0, + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[5] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In 2017 14th IEEE international conference on advanced video and signal based surveillance (AVSS), pages 1-6. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.472, + 0.471 + ], + "angle": 0, + "content": "[6] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9686-9696, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.471, + 0.557 + ], + "angle": 0, + "content": "[7] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3491-3502, 2022. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.471, + 0.627 + ], + "angle": 0, + "content": "[8] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13608-13618, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.471, + 0.699 + ], + "angle": 0, + "content": "[9] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.701, + 0.471, + 0.757 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.758, + 0.471, + 0.799 + ], + "angle": 0, + "content": "[11] Yunhao Du, Yang Song, Bo Yang, and Yanyun Zhao. Strongsort: Make deepsort great again. arXiv preprint arXiv:2202.13514, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.856 + ], + "angle": 0, + "content": "[12] James Ferryman and Ali Shahrokni. Pets2009: Dataset and challenge. In 2009 Twelfth IEEE international workshop on performance evaluation of tracking and surveillance, pages 1-6. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[13] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Sion Hannuna, Massimo Camplani, Jake Hall, Majid Mirmehdi, Dima Damen, Tilo Burghardt, Adeline Paiement, and Lili Tao. Ds-kcf: a real-time tracker for rgb-d data. Journal of Real-Time Image Processing, 16(5):1439-1458, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[16] Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82(1):35-45, 1960. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[19] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[20] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[21] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13536-13545, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[22] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[25] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129(2):548-578, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[26] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8844-8854, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[27] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Matthias Müller, Adel Bibi, Silvio Giancola, Salman Al-Subaihi, and Bernard Ghanem. Trackingnet: A large-scale" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9930" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "dataset and benchmark for object tracking in the wild. In Proceedings of the European Conference on Computer Vision, ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[29] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 164-173, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[30] Jinlong Peng, Changan Wang, Fangbin Wan, Yang Wu, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Chained-tracker: Chaining paired attentive regression results for end-to-end joint multiple-object detection and tracking. In European conference on computer vision, pages 145-161. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.347 + ], + "angle": 0, + "content": "[31] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[32] Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xiangyu Zhang, and Jian Sun. Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.471, + 0.473 + ], + "angle": 0, + "content": "[33] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.471, + 0.53 + ], + "angle": 0, + "content": "[34] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.471, + 0.573 + ], + "angle": 0, + "content": "[35] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.471, + 0.63 + ], + "angle": 0, + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[37] Qiang Wang, Yun Zheng, Pan Pan, and Yinghui Xu. Multiple object tracking with correlation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3876-3886, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.471, + 0.744 + ], + "angle": 0, + "content": "[38] Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In European Conference on Computer Vision, pages 107-122. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[39] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[40] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12352-12361, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[41] Yihong Xu, Yutong Ban, Guillaume Delorme, Chuang Gan, Daniela Rus, and Xavier Alameda-Pineda. Transcenter:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Transformers with dense queries for multiple-object tracking. arXiv preprint arXiv:2103.15145, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[42] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[43] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision, pages 659–675. Springer, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[44] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In European Conference on Computer Vision, pages 1-21. Springer, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[45] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129(11):3069-3087, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[46] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In European Conference on Computer Vision, pages 474-490. Springer, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.894, + 0.487 + ], + "angle": 0, + "content": "[47] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8771-8780, 2022. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9931" + } + ] +] \ No newline at end of file diff --git a/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/4b8d0490-50f4-45bc-9fad-a0ac515c1ca0_origin.pdf b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/4b8d0490-50f4-45bc-9fad-a0ac515c1ca0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a1d1e4101d83e85480184e90d26b34442b1482f --- /dev/null +++ b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/4b8d0490-50f4-45bc-9fad-a0ac515c1ca0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5531df8227e463aac6b1348ae192c7f819a3a3873668d6cc8ead8b87e43d68e6 +size 6341407 diff --git a/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/full.md b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8e7b3f9feaeefe615bdc152dac76bc7f49e5eb51 --- /dev/null +++ b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/full.md @@ -0,0 +1,373 @@ +# SportsMOT: A Large Multi-Object Tracking Dataset in Multiple Sports Scenes + +Yutao Cui* Chenkai Zeng* Xiaoyu Zhao* Yichun Yang* Gangshan Wu Limin Wang $^{\text{☑}}$ State Key Laboratory for Novel Software Technology, Nanjing University, China + +https://github.com/MCG-NJU/SportsMOT + +# Abstract + +Multi-object tracking (MOT) in sports scenes plays a critical role in gathering players statistics, supporting further applications, such as automatic tactical analysis. Yet existing MOT benchmarks cast little attention on this domain. In this work, we present a new large-scale multi-object tracking dataset in multiple sports scenes, coined as SportsMOT, where all players on the court are supposed to be tracked. It consists of 240 video sequences, over 150K frames (almost $15 \times$ MOT17) and over 1.6M bounding boxes ( $3 \times$ MOT17) collected from 3 sports categories, including basketball, volleyball and football. Our dataset is characterized with two key properties: 1) fast and variable-speed motion and 2) similar yet distinguishable appearance. We expect SportsMOT to encourage the MOT trackers to promote in both motion-based association and appearance-based association. We benchmark several state-of-the-art trackers and reveal the key challenge of SportsMOT lies in object association. To alleviate the issue, we further propose a new multi-object tracking framework, termed as MixSort, introducing a MixFormer-like structure as an auxiliary association model to prevailing tracking-by-detection trackers. By integrating the customized appearance-based association with the original motion-based association, MixSort achieves state-of-the-art performance on SportsMOT and MOT17. Based on MixSort, we give an in-depth analysis and provide some profound insights into SportsMOT. + +# 1. Introduction + +Multi-object tracking (MOT) has been a fundamental computer vision task for recent decades, aiming to locate the objects and associate them in video sequences. Researchers have cast much focus on various practical use cases like crowded street scenes [9, 27], static dancing scenes [33] and driving scenarios [14], achieving consid- + +erable progress [2, 4, 30, 37, 38, 40, 41] in MOT. MOT for sports scenes however is overlooked, where typically only the players on the court should be tracked for further analysis, such as counting the players' running distance or average speed and automatic tactical analysis. + +Generally, prevailing state-of-the-art trackers [1,6,11,39, 44] consist of several components to accomplish the tracking task: objects localization module, motion based objects association module and appearance based association module. Biased to the data distribution of specific human tracking benchmarks, e.g. MOT17 [27], MOT20 [9] and DanceTrack [33], the components of these trackers have difficulty adapting to sports scenes. Firstly, motivated by surveillance or self-driving applications, current human tracking benchmarks provide tracks for almost all persons in the scenes. While for sports scenes like basketball or football games, generally only the players on the court are what we focus on, hence a specialized training platform is required to make the detectors suitable for sports scenes. More importantly, in MOT17 and MOT20, these trackers highlight Kalman Filter [16] based IoU matching for object association, due to the slow and regular motion of pedestrians. DanceTrack highlights diverse motion rather than fast movement [33], that is, dancers frequently switch the motion direction and relative position. However in sports scenes, we observe fast and variable-speed movement of objects on adjacent frames, i.e. players usually possess high speed and frequently change their running speed in professional sports events, thus constituting barriers in existing motion based association. For instance, as visualized in Fig. 1, the adjacent IoU and Kalman-Filter-based IoU in sports scenes remain lower than that on MOT17 and DanceTrack (More detailed comparison can be found in Fig. 2 and Fig. 3). As a consequence, more suitable motion based association for sports scenes is required. Additionally, compared to the MOT17 and MOT20 datasets in street scenes, the objects appearances in sports scenes are less distinguishable, since not only the players inherently are in similar clothes but also the players are frequently blurred caused by fast camera motion or targets motion. Different from DanceTrack, where generally the dancers are in almost the same clothes and + +![](images/c6093b651a013e26892c96857ae8238d2a7d15e423f7e960ccd6e9ceb0ec2943.jpg) +(a) Basketball in SportsMOT + +![](images/ae829d78234f11771cd13c87b41f86c1f0f414483014f7199098186c96303f81.jpg) +(c) MOT17 + +![](images/5834c172bde20e9fdfbbd3ba7f2b91e339f0295db95cbf0f7e7331f7e8a7b9a5.jpg) +(b) Football in SportsMOT + +![](images/dd5d3028c2d7dea6394a2e18565223a3633a4cb0662063701a63849d7b625af6.jpg) +(d) DanceTrack + +![](images/29d4c55a59ed743e82bd19e5eb1c3fa30f0b6ddec4e7163085ed4f341f60d639.jpg) +Figure 1. Sampled sequences from the categories of basketball and football of SportsMOT, MOT17 and DanceTrack. There exist two key properties of SportsMOT: 1) fast and variable-speed motion, i.e. players usually possess high speed and frequently change their running speed (the visualized adjacent IoU and Kalman Filter based adjacent IoU can indicate the property); 2) similar yet distinguishable appearance, that is, players in sports scenes inherently wear jackets with different numbers and usually display distinct postures. We expect SportsMOT to encourage the MOT trackers to promote in both motion-based association and appearance based association. + +![](images/2a40cd93e332510d9c72f1b185d943206cfc13081a1a8b1d7b8de3ff234478c4.jpg) + +![](images/4297c956d949b511450566b2ed26f61d00c336736fcd7e35d5666fb2e9d13c83.jpg) + +![](images/2abb806f47620e8beef110e381485a4b327f968b047b721745244f9f1db4ceae.jpg) + +![](images/1833f6c580ee1ddde54002f1d5e9cf3816232ab2ec2d9d3434f49d957e16c880.jpg) + +![](images/9430a6b413733f12ef19ee98ccd5eee007ca0b6a32986190248f6fb3a54718a7.jpg) + +![](images/f30e7e86769324880e288f2e2f7512c63a20be63907eac872f6bfbcea724bdad.jpg) + +![](images/b3f4e10e4f9d68178ce708f523e32b8b74be86608036591837eaeba152acd74d.jpg) + +![](images/91ca6723bb13a3da58896b395bf520d2f3146e9f1bbc60862c4fa1af6ca0b498.jpg) + +![](images/3c1ca75ed8bd64dc7c3c3c5c812dcd764f71e7cb6c32e57ba401aeabd5d19b3f.jpg) + +![](images/cd290579af55a967b1cbd553c1a82900df4c34a784077437d13905e5f0fa6298.jpg) + +![](images/e9822f6aa6b70b5bea6cb69c7b512fc035dbe5c6abe4e828d88bac9696416f1c.jpg) + +![](images/a8764d7458b8c915de171a02a608a23acf8617f09137e2eef3ca6db89356f5db.jpg) + +![](images/ae65fde53686d5266e33ab420be215e9ffab2ad998f7f98d62a2bd8ee0dc242c.jpg) + +![](images/4448b7920f3b15788934e4960792dda92224160680a089b7d1f1d553f7046448.jpg) + +![](images/c9df2ffcd97e0e8978fc913c7f77daacc9ef40d0e022dd90f86e665b8e4ecad9.jpg) + +![](images/1da734336ee7669b09d167ef069f1bfcd244473a9d4870a2b5d8fdce1b47e14e.jpg) + +![](images/d15aa12fa1e6d47a082a35f3f8f1456fa49b175a258003d7af7d9b4f481e1b42.jpg) + +thus having indistinguishable appearance, players in sports scenes inherently wear jackets with different numbers and usually display distinct postures. Thereby, we argue that objects in sports scenes are with similar yet distinguishable appearance, which necessitates the appearance model developing more discriminative and extensive representations. + +Considering the analysis above, to advance the development of tracking and sports analysis, we propose a multi-object tracking dataset in sports scenes, termed as SportsMOT. The dataset is large-scale, high-quality and contains dense annotations for every player on the court in various sports scenes. It consists of 240 videos, over 150K frames (almost $15 \times$ MOT17 [27]) and over 1.6M bounding boxes ( $3 \times$ MOT17 [27]) collected from 3 categories of sports, including basketball, volleyball and football. To provide a platform for making the trackers suitable for sports scenes, we split the dataset into training, validation and test subsets, consisting of 45, 45 and 150 video sequences respectively. There exist two core properties of SportsMOT: (1) fast and variable-speed motion, requiring more suitable motion modeling association; (2) similar yet distinguishable appearance, which necessitates the appearance model developing more discriminative and extensive representations. Altogether, we expect SportsMOT to encourage the trackers to promote in both the certain aspects, i.e. motion based association and appearance based association. + +Given the large-scale multi-object tracking dataset SportsMOT, we benchmark some recent tracking approaches and retrain all of them on the training split. We observe IDF1 and AssA metrics are lower than that on MOT17 while the DetA is quite high, indicating that the main chal + +lenge of SportsMOT lies in objects association rather than objects localization. To alleviate the issue, we propose a new multi-object tracking framework, dubbed as MixSort, with introducing a MixFormer-like [8] structure as appearance based association to prevailing tracking-by-detection trackers (e.g. ByteTrack [44], OC-SORT [6]). By integrating the original motion based objects association and the designed appearance based association, the performance gets boosted on both SportsMOT and MOT17 benchmarks. Based on MixSort, we perform extensive exploration studies and provide some profound insights into SportsMOT. + +The main contributions are summarized as follows: + +- We build a new large-scale multi-object tracking dataset in diverse sports scenes, SportsMOT, equipped with two key properties of 1) fast and variable-speed motion and 2) similar yet distinguishable appearance, aiming to advance the development of both tracking and sports analysis. +- We benchmark some prevailing trackers on SportsMOT, which reveals that the key challenges lie in objects association and hopefully can facilitate further research. +- We propose a new multi-object tracking framework MixSort, with introducing a MixFormer-like structure as appearance based association model to prevailing tracking-by-detection trackers, so as to boost the objects association. Based on MixSort, we perform extensive studies and provide some profound insights into SportsMOT. + +
DatasetVideosFramesLength (s)BboxTracks
MOT171411,235463292,7331,342
MOT20813,4105351,652,0403,456
DanceTrack100105,8555,292-990
SportsMOT240150,3796,0151,629,4903,401
+ +Table 1. Comparison of statistics between existing human MOT datasets and our SportsMOT. + +# 2. Related Work + +Multi-object tracking datasets. Existing Multi-object tracking datasets usually focus on different scenes, such as autonomous driving, pedestrians on roads, and dancing. For autonomous driving, there are KITTI [14], KITTI360 [22] and BDD100K [42], which focus on pedestrians and vehicles. Besides, other datasets, focusing on only the pedestrians, collect the videos from static and moving cameras. One of the earliest is PETS [12], but it's too simple in some scenes. MOT15 [20] proposes the first large-scale benchmark for Multi-object tracking, followed by MOT17 [27] and MOT20 [9]. It's worth noting that MOT20 focuses on extremely crowded scenes where many pedestrians are occluded, increasing the tracking difficulty greatly in both detection and association. Recently, DanceTrack [33], focusing on dancing scenes, is proposed to encourage trackers to rely less on visual discrimination and depend more on motion analysis. The emphasized properties are uniform appearance and diverse motion. While in SportsMOT, the appearance is similar yet distinguishable, and the players' motion is fast and with variable speed. We expect SportsMOT to encourage algorithms to promote in both appearance and motion association. Besides, SoccerNet-Tracking [7] is presented to track elements in football scenarios. The main difference lies in that, it only contains soccer scenes and tracks almost all elements (players, goalkeepers, referees, balls) on the court without distinction. While SportsMOT contains three types of sports where the objects in basketball scenes are more crowded and thus more challenging (refer to Section 5.3), and only focuses on the players to support further statistics and tactical analysis. + +Object association in tracking. Association is a very important task in tracking, where trackers need to associate detections in new frames with existing tracks. For most of the trackers, a similarity matrix (or cost matrix) between new detections and tracks is computed based commonly on motion and appearance cues, which is later fed into Hungarian algorithm [18] to perform association. For example, SORT [4] uses Kalman Filter to predict the location of objects and computes the IoU of detected and predicted bounding boxes as similarity matrix. IOU-Tracker [5] directly computes the IoU without prediction. ByteTrack [44] adds an association phase for detection with low confidence score, which can boost the performance. OC-SORT [6] tries to address the limitation of Kalman Filter. + +Appearance cues also play an important role in association. DeepSORT [39] crops the detection from frame images, which are then used by networks to generate re-ID features. Then the motion cues and distance of re-ID features are fused to perform association. FairMOT [45] uses a re-ID branch on a backbone shared by the detection branch to generate re-ID features. In CenterTrack [46], the previous frame is used to help the prediction of tracks. Recently, Transformer [36] is used by some work such as TrackFormer [26] and MOTR [43] to boost the association quality. Our proposed MixSort integrates the motion and appearance cues with a motion modeling component and the designed MixFormer-like structure respectively. + +# 3. SportsMOT Dataset + +# 3.1. Dataset Construction + +Video Collection. We select three worldwide famous sports, football, basketball, and volleyball, and collect videos of high-quality professional games including NCAA, Premier League, and Olympics from Multi-Sports [21], which is a large dataset in sports area focusing on spatio-temporal action localization. Each category has typical players' formations and motion patterns, and they can effectively represent the diversity of sports scenarios. Only the overhead shots of sports game scenes are used, guaranteeing certain extreme situations do not occur. The proposed dataset consists of 240 video sequences in total, each of which is 720P and 25 FPS. Following the principles of multi-object tracking, each video clip is manually checked to ensure that there are no abrupt viewpoint switches within the video. + +Annotation Pipeline. We annotate the collected videos according to the following guidelines. + +- The entire athlete's limbs and torso, excluding any other objects like balls touching the athlete's body, are required to be annotated. +- The annotators are asked to predict the bounding box of the athlete in the case of occlusion, as long as the athletes have a visible part of body. However, if half of the athletes' torso is outside the view, annotators should just skip them. +- We ask the annotators to confirm that each player has a unique ID throughout the whole clip. + +We provide a customized labeling tool for SportsMOT and a corresponding manual book to annotators. Once they start annotating a new object, the labeling tool automatically assigns a new ID to the object and propagates the bounding box of previous state to the current state, with the help of the single object tracker KCF [15]. Then the generated bounding boxes should be refined by the annotators, so as to improve annotation quality. After carefully reviewing each + +![](images/fda639178cdbf7cb39b2d0186c2e29e82e4fdb3256ac906a590a448d557cbf14.jpg) +Figure 2. IoU on adjacent frames. (a) Compared to MOT17 and DanceTrack, SportsMOT has a lower score, indicating that objects have faster motion. (b) In SportsMOT, the category of football has the lowest IoU score, which means that football players often have fast motion. + +![](images/19c98d15d90c27ace1fbc4181a2f4a722e50374c38bf6c46523b085980628a0f.jpg) +Figure 3. Kalman-Filter-based IoU on adjacent frames. (a) Compared to MOT17 and DanceTrack, SportsMOT has a lower score, indicating that objects have more variable-speed motion. (b) In SportsMOT, the category of football has the lowest Kalman-Filter-based IoU score. + +
CategoryFramesTracksTrack gap len.Track len.Bboxes per frame
Basketball845.41068.7767.99.1
Volleyball360.41238.2335.911.2
Football673.920.5116.1422.112.8
Total626.614.296.6479.110.8
+ +Table 2. Detailed statistics of the three categories in SportsMOT. + +annotation result, we refine the bounding boxes and IDs that do not satisfy the standards, hence building a high-quality dataset. Finally, the bounding boxes with too small size, i.e. $w < 5$ or $h < 5$ , are deleted. + +# 3.2. Dataset Statistic + +Overview. SportsMOT is a large-scale and high-quality MOT dataset, aiming to advance the development of both sports analysis and multi-object tracking. Table 1 compares the statistics of SportsMOT with the prevailing human tracking datasets, including MOT17, MOT20 and DanceTrack. According to the statistics, SportsMOT has a large number of bounding boxes of over 1.6M, which is comparable to MOT20 and significantly larger than MOT17. Besides, SportsMOT has a large number of video clips, tracks $(2.5 \times \text{MOT17}, 3.4 \times \text{DanceTrack})$ , frames $(13.4 \times \text{MOT17})$ . As shown in Table 2, we also compare the basic statistics of each category of SportsMOT. SportsMOT solely provides fine annotations of the players on the court, which are supposed to be tracked for further analysis. To provide a platform for making the trackers suitable for sports scenes, we split the dataset into training, validation and test subsets, consisting of 45, 45 and 150 video sequences respectively. + +Fast and Variable-Speed Motion. Motion cues play an important role in object association for multi-object tracking. The existing human-tracking datasets (e.g. MOT17, MOT20 and DanceTrack) generally have certain motion patterns that are distinct with sports scenes, constituting barriers in players tracking. For instance, in MOT17 and MOT20, pedestrians are featured by linear motion with constant speed, which is easily hacked by association strategies with constant velocity assumption. Besides, Dance + +![](images/3e823abf075bdab709ff5becb3f1653f6da32e08434215420dc3a059251b821a.jpg) +(a) + +![](images/51856aaa134953264ab6c1f433fa41092026c734f9df2e8964a94753142cd4a2.jpg) +(b) + +![](images/7a4a0510c929571eca65970e4f2a745773160dad264ace4225c0b90b778f46c0.jpg) +(a) MOT17 + +![](images/798f7bbe3ddb04b21dab554643dad22f5e368bec787b7d09db73d8b8989211e8.jpg) +(b) SportsMOT +Figure 4. Visualization of re-ID features from sampled videos in MOT17, SportsMOT and DanceTrack dataset using t-SNE [35]. The same object is coded by the same color. It indicates that object appearance of SportsMOT is less distinguishable than that of MOT17, while more distinguishable than that of DanceTrack. We expect the appearance model to capture more discriminative and extensive representation for object association. + +![](images/adbef16cd430c0f29d8def26cc5bc827c6c379c79588c8efdb54a39eae1d233d.jpg) +(c) DanceTrack + +Track highlights diverse motion rather than fast motion, that is, dancers usually move in more diverse directions with relatively low speed. In contrast, SportsMOT has distinct motion patterns, i.e. fast and variable-speed motion, where the players typically move fast with their running speed or camera speed frequently changing. As illustrated in Fig. 2, among the three datasets, SportsMOT has the lowest IoU score of the objects bounding boxes on adjacent frames, indicating the fast movement. We use the ground truth of previous frames for Kalman Filter prediction. The result and current ground truth are used to calculate Kalman-Filter-based IoU. Seen from Fig. 3, SportsMOT also has the lowest Kalman-Filter-based IoU score on adjacent frames, which suggests that the motion can not be easily modeled by prevailing methods due to the variable-speed movement. Specifically, football has the smallest adjacent IoU and Kalman-Filter-based IoU, which is closely related to the fast running speed, abrupt acceleration or stops. It poses a major challenge for trackers based on simple motion assumptions and also encourages them to model object motion in more dynamic and adaptive ways. + +Similar yet Distinguishable Appearance. Object appearance is another kind of cue on which MOT trackers often rely to distinguish different objects. In MOT17 and MOT20, pedestrians are usually distinct in body size and wear different clothes, yielding discriminative visual features. In contrast, the objects in DanceTrack typically wear nearly identical outfits, leading to indistinguishable + +![](images/c7f690504ba803de7ddee35c2c84b4ec3054b9ccbd9dcfda39837a90587c3b76.jpg) +Figure 5. Paradigm for computing visual similarity matrix of tracks and detections in MixSort. The center of ground-truth target detection is marked with red dots, the others are green. The blue dashed box indicates the cropped search region. The blank part of similarity vector means that for detections not in the search region, the corresponding value is set to 0. + +appearances. Thereby, DanceTrack highlights solely relying on motion-based association rather than appearance-based association. In SportsMOT, the players also have very similar appearances. However, the players wear jackets with different numbers and usually display distinct postures, thus resulting in similar yet distinguishable appearance. In Fig. 4, we provide visualization of re-ID feature from sampled videos in MOT17, SportsMOT and DanceTrack dataset using t-SNE [35]. It implies that the re-ID features in SportsMOT are similar yet distinguishable compared to MOT17 and DanceTrack. We aim to encourage the trackers to learn more discriminative visual representations for more robust object association. + +Evaluation Metrics. MOTA [3] is the main metric for existing MOT evaluations. However, MOTA focuses more on measuring the accuracy of detection. To highlight the performance of object association, we recommend HOTA [25], AssA and IDF1 [31] as the major evaluation metrics in SportsMOT dataset. HOTA aims to measure the accuracy of detection and association equally. + +# 4. Multi-Object Tracking on SportsMOT + +In this section, we present our proposed multi-object tracking framework, called MixSort. This framework is designed to enhance the appearance-based association performance and can be applied to any trackers that follow the tracking-by-detection paradigm, such as ByteTrack [44] and OC-SORT [6]. + +# 4.1. MixFormer for Appearance-based Association + +MixFormer. In this paragraph, we discuss the use of Mix-Former [8] in our proposed framework MixSort. Mix-Former is designed to extract target-specific discriminative features and perform extensive communication between the target and search area, therefore, it is the key component that enables MixSort to compute visual similarities between the templates of tracked objects and detected objects in the search region of the current frames. The original Mix-Former uses a corner-based localization head to predict the + +![](images/d1a4bb390bb15982083ff7954ca0b4d2d03aeb55eb0c63a0e7fb161bfd97d936.jpg) +Figure 6. The pipeline of MixSort. We use motion model and MixFormer to generate fused similarity matrix for association. + +top-left and bottom-right corners of the input template in the search region. However, we modify the corner head by using a heatmap prediction head that predicts the center of the template and generates a confidence heatmap. This allows us to compute the similarity between the detection and the template. To make MixSort suitable for multi-object tracking and accelerate inference speed, we reduce the number of mixed attention modules in MixFormer from 12 to 4. The steps involved in computing the visual similarity matrix are illustrated in Figure 5. + +Association Strategy. In order to perform association between detections and existing tracks, we use a mixed similarity matrix generated by computing the visual similarity between the target template and the detected objects in the search region of the current frame. Specifically, we obtain the heatmap response at the center of each detection as its visual similarity to the template. The resulting similarity matrix is then combined with the IoU matrix using the Hungarian Algorithm. + +To start, for each existing track $t$ , we use the Kalman Filter to predict its new location. Then, we crop the current frame centered at the predicted location with a certain scale to obtain the search region $s$ . By feeding $s$ and the template $t$ into MixFormer, we generate a heatmap $\mathcal{H}$ that represents the similarity between the template and search region. + +Next, for each detection $d$ whose center is in the search region $s$ , we set its similarity to track $t$ as the response in the heatmap $\mathcal{H}$ ; the similarity values of other detections are set to 0. Finally, we fuse the visual similarity and the IoU score to obtain the mixed similarity matrix + +$$ +M = \alpha \cdot \operatorname {I o U} + (1 - \alpha) \cdot V \tag {1} +$$ + +where $\alpha$ is the weight coefficient and $V$ represents the visual similarity matrix calculated using MixFormer. + +# 4.2. MixSort Tracking + +Based on the tracking-by-detection paradigm, the pipeline of MixSort can be generalized as follows. As shown in Figure 6, we first obtain detections using an object detector. Then, we employ a motion model (e.g., Kalman Filter) to predict new locations of existing tracks. Based on the new locations and templates of tracks, we compute a + +
Training SetupHOTA↑IDF1↑AssA↑MOTA↑DetA↑LocA↑IDs↓Frag↓
CenterTrack [46]Train62.760.048.090.882.190.8104815750
FairMOT [45]Train49.353.534.786.470.283.9992821673
QDTrack [29]Train60.462.347.290.177.588.0637711850
TransTrack [34]Train68.971.557.592.682.791.049929994
GTR [47]Train54.555.845.967.964.889.0956714525
ByteTrack [44]Train62.869.851.294.177.185.632674499
OC-SORT [6]Train71.972.259.894.586.492.430933474
ByteTrackTrain+Val64.171.452.395.978.585.730894216
OC-SORTTrain+Val73.774.061.596.588.592.727283144
MixSort-ByteTrain+Val65.7 (+1.6)74.1 (+2.7)54.8 (+2.5)96.278.885.724724009
MixSort-OCTrain+Val74.1 (+0.4)74.4 (+0.4)62.0 (+0.5)96.588.592.727813199
+ +Table 3. Tracking performance of investigated algorithms on our proposed SportsMOT. The best results are shown in bold. + +fused similarity matrix as described above and use it to associate tracks and detections by the Hungarian Algorithm. Finally, for matched tracks and detections, we update the online templates. For unmatched tracks, we keep them until the threshold is reached. For unmatched detections with confidence scores higher than the threshold, new tracks are initialized. + +# 4.3. Training and Inference + +Training. We only consider the training of MixFormer here since the detector remains the same as initial method (e.g. ByteTrack). The original MixFormer is trained on SOT datasets, so we first modify the format of ground truth of MOT datasets, that is, converting the ground truth trajectory of every single player into TrackingNet format [28]. + +For each ground-truth bounding box, we compute its corresponding center location $(c_x, c_y)$ in the low-resolution heatmap. Following CornerNet [19], the ground-truth heatmap response is generated using 2D Gaussian kernel: + +$$ +h _ {x y} = \exp \left(- \frac {\left(x - c _ {x}\right) ^ {2} + \left(y - c _ {y}\right) ^ {2}}{2 \sigma^ {2}}\right) \tag {2} +$$ + +where $\sigma$ is adaptive to the size of the bounding box. The training loss is a pixel-wise logistic regression with focal loss [23]: + +$$ +L = - \sum_ {x y} \left\{ \begin{array}{l l} (1 - \hat {h} _ {x y}) ^ {\gamma} \log (\hat {h} _ {x y}) & , h _ {x y} = 1; \\ (1 - h _ {x y}) ^ {\beta} (\hat {h} _ {x y}) ^ {\gamma} \log (1 - \hat {h} _ {x y}) & , \text {o t h e r w i s e}. \end{array} \right. \tag {3} +$$ + +where $\gamma$ and $\beta$ are hyper-parameters in focal loss, and we set $\gamma = 2$ , $\beta = 4$ following CornerNet. + +Inference. For each track, we maintain only one template for keeping a balance between accuracy and speed. When a detection is matched to an existing track, we directly replace the original template to the new detection, if and only if the ratio of its uncovered (i.e. overlapping with any detected objects) area is larger than a certain threshold, so as to reduce the impact of misleading representations. + +# 5. Experiments and Analysis + +# 5.1. Experiment Setup + +Dataset Split. In benchmark experiments, we follow the default split described in Sec. 3.2. In exploration study, we split the original MOT17 training set into two sets, used for training and validation respectively following CenterTrack. + +Implementation Details. Following ByteTrack and OC-SORT, we use YOLOX [13] as our detector. Using COCO-pretrained model as the initialized weights, we first train the model on CrowdHuman [32] for 80 epochs and then train on SportsMOT for another 80 epochs. The remaining settings are the same as that in ByteTrack. + +For MixFormer, we initialize the backbone with the model trained on VOT datasets and then fine-tune it on SportsMOT for 300 epochs with learning rate initialized as $1e - 4$ and decreased to $1e - 5$ at epoch 200. The optimizer is ADAM [17] with weight decay $10^{-4}$ . The sizes of search images and templates are $224 \times 224$ and $96 \times 96$ respectively. The max sample interval is set to 10. For every tracking result, we apply linear interpolation as post-processing, with maximum gap set to 20. + +# 5.2. Benchmark Results + +We evaluate several representative methods of three kinds on our dataset. ByteTrack [44], OC-SORT [6] and QDTrack [29] are trackers in tracking-by-detection paradigm. CenterTrack [46] and FairMOT [45] perform joint detection and tracking in one stage. TransTrack [34] and GTR [47] are trackers based on Transformer. Most of the current best multi-object tracking algorithms belong to tracking-by-detection paradigm, however, due to the separation of detection and tracking, the information cannot be shared completely. Joint-detection-and-tracking paradigm couples the two modules, with the goal of boosting the performance of each. Transformer-based-tracking methods are relatively new but have achieved great performance. Despite its huge potential, the model complexity and calcula + +
α10.90.80.70.60.50.40.30.20.10
basketball65.966.166.265.965.364.863.660.756.747.126.9
volleyball76.076.876.576.476.576.976.475.573.769.140.9
football71.972.472.372.472.572.873.272.972.671.465.7
+ +tion cost are much higher, resulting in large memory and long training time. + +All training settings including the number of epochs and change of learning rate are consistent with original papers. According to different default settings, we follow the commonly used pretraining datasets, such as CrowdHuman [32], COCO [24] and ImageNet [10], and apply SportsMOT-train with or without other datasets for finetuning for different methods. We compare the results in Tab. 3. + +In sports scenes, the clear appearance and sparse density of objects allow current mature detection frameworks to generate bounding boxes with high accuracy. However, specialized detectors need to be trained for not detecting audience and referees. The key challenges are fast speed and motion blur, which forces us to pay more attention to improve the association performance. Besides, The wide range of HOTA and MOTA denotes SportsMOT is more distinguishable among different kinds of algorithms. + +# 5.3. Exploration Study + +In this section, we perform extensive studies on the proposed MixSort and SportsMOT. + +Effectiveness of the proposed association module. We evaluate the effectiveness of MixSort by applying it to two state-of-the-art trackers, OC-SORT [6] and ByteTrack [44], which follow the tracking-by-detection paradigm and useYOLOX as their detector. The evaluation is conducted on the SportsMOT test set, and the results are presented in Tab. 3. Our experiments show that MixSort significantly improves the performance of both trackers, with OC-SORT achieving a 0.4 HOTA increase and ByteTrack achieving a 1.6 HOTA increase on SportsMOT. This demonstrates the effectiveness of MixSort in enhancing the association. + +Appearance-based vs. Motion-based association. We have demonstrated that MixSort can improve association performance. In this paragraph, we investigate the impact of the fusion weight $\alpha$ on the ability of appearance cues to aid in conventional motion-based association. We evaluate OC-SORT with MixSort on the three categories in the SportsMOT test set using $\alpha$ values ranging from 1 to 0 in Eq. (1). The results, presented in Tab. 4, reveal that pure motion-based association $(\alpha = 1)$ outperforms pure appearance-based association $(\alpha = 0)$ in all categories, underscoring the significance of motion cues in sports scenes. Moreover, fused association surpasses both + +Table 4. Comparison of the HOTA metric of basketball, volley and football under different fusion parameters $\alpha$ on SportsMOT test set. The models are trained on SportsMOT training set. + +
IoUMotionMix.HOTA↑IDF1↑AssA↑MOTA↑IDs↓
✓71.571.258.195.94329
✓64.263.948.791.125947
✓✓64.171.452.395.93089
✓✓73.874.461.696.63203
✓✓✓65.774.154.896.12469
+ +Table 5. Results of the ablation experiment on SportsMOT test set. IoU means computing IoU between detections and the last location of existing tracks for association, while Motion means using Kalman filter to predict the location of tracks. The models are trained on SportsMOT training and validation set. + +pure motion-based and appearance-based association, suggesting that both motion and appearance cues should be considered jointly for optimal results. + +Our analysis of the three categories reveals that appearance cues provide the most significant improvement for football videos (+1.3), followed by volleyball (+0.9) and basketball (+0.3). Combining this with Fig. 2, which indicates that the adjacent IoU of football games is the smallest (i.e., motion is fastest) among the three categories, we can conclude that scenes with faster motion are more dependent on appearance cues. + +Ablation study on MixSort. We ablate important components of our tracker (MixSort based on ByteTrack) including IoU, Motion (Kalman Filter) and MixSort on SportsMOT test set. The results are presented in Tab. 5. Surprisingly, we found that simple IoU without using motion prediction outperformed IoU with motion prediction by a large margin (from 64.1 HOTA to 71.5 HOTA), indicating that the Kalman filter, which assumes linear motion models, performed poorly on SportsMOT, where the motion patterns are far more complex than in previous datasets. + +Furthermore, we observed that MixSort played a crucial role in boosting the performance of the tracker significantly. By fusing the IoU and MixSort cues, our method achieved the best performance of 73.8 HOTA compared to 71.5 HOTA of simple IoU in our experiments. + +Analysis on different categories of SportsMOT. While the three SportsMOT categories share some common characteristics such as fast motion and similar appearance, they also have distinct features due to the different types of games. In this section, we analyze the results of our experiments on the three categories. + +We first use the best HOTA metric from Tab. 4 to represent the overall difficulty of a category. Based on this metric, we find that basketball videos (66.17) are the most difficult, followed by football (73.19), and finally volleyball (76.91). + +Next, we consider the appearance-based association (i.e. $\alpha = 0$ in Tab. 4). We observe a notable gap between the HOTA metrics of the three categories, where basketball (26.94) has a much lower HOTA than volleyball (40.94) and football (65.65). Similarly, basketball remains the most dif + +
MOT17-testMOT17-val
HOTA↑IDF1↑AssA↑MOTA↑DetA↑IDs↓HOTA↑IDF1↑AssA↑MOTA↑DetA↑IDs↓
QDTrack [29]53.966.352.768.755.63378------
MOTR [43]57.268.455.871.958.92115------
GTR [47]59.171.557.075.361.6285963.075.966.271.360.4-
ByteTrack [44]63.177.362.080.364.52196-79.7-76.7-159
OC-SORT [6]63.277.563.478.063.2195068.079.369.977.9--
MixSort-Byte64.078.764.279.364.1223569.481.171.379.968.2155
MixSort-OC63.477.863.278.963.8150969.280.671.578.967.4135
+ +Table 6. Comparison of the state-of-the-art methods under the "private detector" protocol on MOT17-test set and MOT17-val set. + +
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
ByteTrack47.151.931.588.270.5
MixSort-Byte46.753.031.985.868.6
+ +Table 7. Comparison of ByteTrack and MixSort-Byte on DanceTrack validation set. For MixSort-Byte, the fuse parameter $\alpha$ is 0.9, which results in the highest HOTA among $\{0.6, 0.7, 0.8, 0.9, 0.95\}$ . The models are trained on DanceTrack training set. + +
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
SportsMOToverall65.774.154.896.178.8
basketball60.867.846.897.379.1
volleyball72.587.066.896.578.7
football66.473.656.394.978.5
SoccerNet62.973.955.587.871.5
+ +Table 8. Results of MixSort-Byte on SportsMOT and SoccerNet test set. The models are trained on SportsMOT training and validation set and SoccerNet training set respectively and the hyperparameters are the same. + +ficult in the motion-based association (i.e. $\alpha = 1$ in Tab. 4), while the volleyball becomes the easiest. + +We believe that the differences in difficulty arise from several factors, including the size of the game court and the degree of physical confrontation among players. For instance, basketball scenes are played on smaller courts and involve more physical contact between players than football scenes. This can lead to more occlusion and blur in basketball videos, making the association task more challenging than in football scenes. + +Comparison with DanceTrack. We evaluate MixSort-Byte on the DanceTrack validation set and compare the results with that of ByteTrack as shown in Tab. 7. Unlike the results on SportsMOT where MixSort brings significant improvement, on DanceTrack the original ByteTrack performs better instead, with HOTA, MOTA and DetA metrics all higher than MixSort-Byte. This indicates that appearances in our proposed dataset SportsMOT are similar yet distinguishable, while those in DanceTrack are much harder to distinguish. Therefore SportsMOT highlights both the motion-based and appearance-based associations. + +Comparison with SoccerNet. We conducted experiments to compare SoccerNet that focuses only on soccer + +
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
ByteTrack64.171.452.395.978.5
ByteTrack+ReID64.872.253.496.178.8
MixSort-Byte65.774.154.896.178.8
+ +Table 9. Comparison of ByteTrack, ByteTrack with ReID model and MixSort-Byte on SportsMOT test set. The ReID model is the same as in DeepSORT [4] and all the models are trained on SportsMOT training set and the best results are shown in **bold**. + +scenes with SportsMOT. Results shown in Tab. 8 suggest that SportsMOT is a challenging dataset with varying levels of difficulty across different sports. Specifically, basketball is proved to be the most difficult with the lowest HOTA of 60.8, while volleyball and football are relatively easier. Notably, MixSort obtains higher HOTA on overall SportsMOT than on SoccerNet, which is mainly because all the elements on the court are to be tracked in SoccerNet, leading to more false detections and much lower DetA (71.5 vs. 78.8). Yet, with a higher AssA on SoccerNet despite these false detections, it underscores SportsMOT's challenging association and significance for sports tracking. + +MixSort vs. ReID models. To validate the effectiveness of the MixFormer-like model in MixSort, we present a comparison with ReID models in Tab. 9. The results show that ByteTrack with the ReID model outperforms the original ByteTrack, highlighting the value of appearance-based association on SportsMOT. Notably, MixSort-Byte further improves upon ByteTrack with ReID by 0.9, 1.9, and 1.4 in HOTA, IDF1, and AssA. This underscores MixSort's superior appearance modeling, capable of extracting broader and more discriminative representations and facilitating efficient offline learning. + +More results on MOT17 The propose MixSort can be applied to any trackers following this paradigm and achieve state-of-the-art performance on SportsMOT. To further validate the effectiveness of MixSort, we compare MixSort with the prevailing trackers on MOT17 validation set and test set under the private detection protocol in Tab. 6. MixSort-Byte and MixSort-OC outperform these trackers in HOTA, IDF1. + +Detailed Challenging attributes of SportsMOT. For a comprehensive statistical analysis and performance evalua + +
Attr.Definition of sequenceAttr.Definition of target box
CMAbrupt motion of the cameraLRThe target box is smaller than a threshold
OVThe target completely leaves the video frameFOCThe area of target box intersecting with others is larger than 0.9
MBThe target region is blurred due to target or camera motionPOCThe area of target box intersecting with others is in range [0.4, 0.9]
FMThe total average IoU of target boxes in adjacent frames is smaller than a thresholdDEFThe sum of the relative changes in the length and width of the target boxes in adjacent frames is greater than a threshold
+ +Table 10. Definition of attributes. For the attributes in the right half, we first calculate the proportion of boxes in each sequence that meet the criteria defined in the table. If this proportion exceeds a certain threshold, we consider the sequence to exhibit that attribute. + +![](images/5783a89997d202b754f3a9522ff5a26882a5a3a8b3faed16a52d02743e75962e.jpg) +(a) Total + +![](images/27f7dfbc3f9b4352efb65f2c4c065077efbc316870797f9e725588a64e0cbfc9.jpg) +(b) Category-wise + +![](images/0e422629dbcde8ff0be1e8f3c96621cf300de85f01125fe2607b6724e5026aea.jpg) +(a) Total +Figure 8. Average HOTA of sequences in each attribute on SportsMOT dataset. MixSort-OC is employed for analysis. + +![](images/b075346d36606c1fafc8fb953395c42679dcb504ebc501fa02522599071d1088.jpg) +Figure 7. Distribution of sequences in each attribute on SportsMOT dataset. +(b) Category-wise + +tion of the SportsMOT dataset, we categorize and present eight distinct challenging attributes: camera motion (CM), out of view (OV), motion blur (MB), fast motion (FM), low resolution (LR), full occlusion (FOC), partial occlusion (POC), and large deformation (DEF). The definitions for each attribute are specified in Tab. 10, with their corresponding distribution illustrated in Fig. 7. Human selection is used to identify the CM, OV, and MB attributes, while metrics calculation helps determine the others. The diversity and complexity of the scenarios featured in Fig. 7a suggest that our three chosen sports games provide a representative snapshot of broader sports analysis based on tracking states. Furthermore, in comparison to SoccerNetTracking [7], our SportsMOT offers a more diverse set of challenges observed in sports. These challenges range from the frequent occlusion seen in basketball and volleyball, to the low resolution typical of football, as depicted in Fig. 7b. We also assess the difficulty level posed by different attributes in the following. + +Difficulty analysis based on HOTA scores. We employ the average HOTA scores of sequences corresponding to each attribute as an indicator of that attribute's difficulty. The results are depicted in Fig. 8. The low HOTA score for full occlusion, evident in Fig. 8a, highlights its challeng- + +ing nature in MOT within the sports field. Furthermore, an analysis of average HOTA scores shows that basketball consistently has lower scores than volleyball and football for all attributes (Fig. 8b), indicating its increased complexity. + +# 6. Conclusion + +In this paper, we have introduced SportsMOT, a large-scale multi-object tracking dataset in sports scenes. SportsMOT is characterized with two key properties: 1) fast and variable-speed motion and 2) similar yet distinguishable appearance. We have empirically investigated several prevailing MOT trackers on the SportsMOT dataset. We have also proposed a new MOT framework MixSort, introducing a MixFormer-like association module. Hopefully, SportsMOT can provide a platform for facilitating both sports analysis and multi-object tracking. + +Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization. + +# References + +[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 1 +[2] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 941-951, 2019. 1 +[3] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008. 5 +[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 1, 3, 8 +[5] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In 2017 14th IEEE international conference on advanced video and signal based surveillance (AVSS), pages 1-6. IEEE, 2017. 3 +[6] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9686-9696, 2023. 1, 2, 3, 5, 6, 7, 8 +[7] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3491-3502, 2022. 3, 9 +[8] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13608-13618, 2022. 2, 5 +[9] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 3 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7 +[11] Yunhao Du, Yang Song, Bo Yang, and Yanyun Zhao. Strongsort: Make deepsort great again. arXiv preprint arXiv:2202.13514, 2022. 1 +[12] James Ferryman and Ali Shahrokni. Pets2009: Dataset and challenge. In 2009 Twelfth IEEE international workshop on performance evaluation of tracking and surveillance, pages 1-6. IEEE, 2009. 3 +[13] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6 + +[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 3 +[15] Sion Hannuna, Massimo Camplani, Jake Hall, Majid Mirmehdi, Dima Damen, Tilo Burghardt, Adeline Paiement, and Lili Tao. Ds-kcf: a real-time tracker for rgb-d data. Journal of Real-Time Image Processing, 16(5):1439-1458, 2019. 3 +[16] Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82(1):35-45, 1960. 1 +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3 +[19] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 6 +[20] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942, 2015. 3 +[21] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13536-13545, October 2021. 3 +[22] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3 +[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6 +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7 +[25] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129(2):548-578, 2021. 5 +[26] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8844-8854, 2022. 3 +[27] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016. 1, 2, 3 +[28] Matthias Müller, Adel Bibi, Silvio Giancola, Salman Al-Subaihi, and Bernard Ghanem. Trackingnet: A large-scale + +dataset and benchmark for object tracking in the wild. In Proceedings of the European Conference on Computer Vision, ECCV, 2018. 6 +[29] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 164-173, 2021. 6, 8 +[30] Jinlong Peng, Changan Wang, Fangbin Wan, Yang Wu, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Chained-tracker: Chaining paired attentive regression results for end-to-end joint multiple-object detection and tracking. In European conference on computer vision, pages 145-161. Springer, 2020. 1 +[31] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 5 +[32] Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xiangyu Zhang, and Jian Sun. Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123, 2018. 6, 7 +[33] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1, 3 +[34] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 6 +[35] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 4, 5 +[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[37] Qiang Wang, Yun Zheng, Pan Pan, and Yinghui Xu. Multiple object tracking with correlation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3876-3886, 2021. 1 +[38] Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In European Conference on Computer Vision, pages 107-122. Springer, 2020. 1 +[39] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017. 1, 3 +[40] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12352-12361, 2021. 1 +[41] Yihong Xu, Yutong Ban, Guillaume Delorme, Chuang Gan, Daniela Rus, and Xavier Alameda-Pineda. Transcenter: + +Transformers with dense queries for multiple-object tracking. arXiv preprint arXiv:2103.15145, 2021. 1 +[42] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 3 +[43] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision, pages 659–675. Springer, 2022. 3, 8 +[44] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In European Conference on Computer Vision, pages 1-21. Springer, 2022. 1, 2, 3, 5, 6, 7, 8 +[45] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129(11):3069-3087, 2021. 3, 6 +[46] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In European Conference on Computer Vision, pages 474-490. Springer, 2020. 3, 6 +[47] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8771-8780, 2022. 6, 8 \ No newline at end of file diff --git a/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/images.zip b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6a33a2cfb4c6caf97b978a5e284380458c1a88e8 --- /dev/null +++ b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e2473718223f4794523a4857934c6af630f25df37c840912c4c6f198a11668a +size 654292 diff --git a/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/layout.json b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae6aa144eb1d7daabd98201707de956462a3d608 --- /dev/null +++ b/2023/SportsMOT_ A Large Multi-Object Tracking Dataset in Multiple Sports Scenes/layout.json @@ -0,0 +1,9253 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 541, + 121 + ], + "type": "text", + "content": "SportsMOT: A Large Multi-Object Tracking Dataset in Multiple Sports Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 143, + 534, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 143, + 534, + 175 + ], + "spans": [ + { + "bbox": [ + 59, + 143, + 534, + 175 + ], + "type": "text", + "content": "Yutao Cui* Chenkai Zeng* Xiaoyu Zhao* Yichun Yang* Gangshan Wu Limin Wang " + }, + { + "bbox": [ + 59, + 143, + 534, + 175 + ], + "type": "inline_equation", + "content": "^{\\text{☑}}" + }, + { + "bbox": [ + 59, + 143, + 534, + 175 + ], + "type": "text", + "content": " State Key Laboratory for Novel Software Technology, Nanjing University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 163, + 180, + 427, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 180, + 427, + 194 + ], + "spans": [ + { + "bbox": [ + 163, + 180, + 427, + 194 + ], + "type": "text", + "content": "https://github.com/MCG-NJU/SportsMOT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 233, + 192, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 233, + 192, + 247 + ], + "spans": [ + { + "bbox": [ + 143, + 233, + 192, + 247 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 258, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 289, + 582 + ], + "type": "text", + "content": "Multi-object tracking (MOT) in sports scenes plays a critical role in gathering players statistics, supporting further applications, such as automatic tactical analysis. Yet existing MOT benchmarks cast little attention on this domain. In this work, we present a new large-scale multi-object tracking dataset in multiple sports scenes, coined as SportsMOT, where all players on the court are supposed to be tracked. It consists of 240 video sequences, over 150K frames (almost " + }, + { + "bbox": [ + 46, + 258, + 289, + 582 + ], + "type": "inline_equation", + "content": "15 \\times" + }, + { + "bbox": [ + 46, + 258, + 289, + 582 + ], + "type": "text", + "content": " MOT17) and over 1.6M bounding boxes (" + }, + { + "bbox": [ + 46, + 258, + 289, + 582 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 258, + 289, + 582 + ], + "type": "text", + "content": " MOT17) collected from 3 sports categories, including basketball, volleyball and football. Our dataset is characterized with two key properties: 1) fast and variable-speed motion and 2) similar yet distinguishable appearance. We expect SportsMOT to encourage the MOT trackers to promote in both motion-based association and appearance-based association. We benchmark several state-of-the-art trackers and reveal the key challenge of SportsMOT lies in object association. To alleviate the issue, we further propose a new multi-object tracking framework, termed as MixSort, introducing a MixFormer-like structure as an auxiliary association model to prevailing tracking-by-detection trackers. By integrating the customized appearance-based association with the original motion-based association, MixSort achieves state-of-the-art performance on SportsMOT and MOT17. Based on MixSort, we give an in-depth analysis and provide some profound insights into SportsMOT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 624, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 696 + ], + "type": "text", + "content": "Multi-object tracking (MOT) has been a fundamental computer vision task for recent decades, aiming to locate the objects and associate them in video sequences. Researchers have cast much focus on various practical use cases like crowded street scenes [9, 27], static dancing scenes [33] and driving scenarios [14], achieving consid-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 222, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 222, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 306, + 222, + 545, + 282 + ], + "type": "text", + "content": "erable progress [2, 4, 30, 37, 38, 40, 41] in MOT. MOT for sports scenes however is overlooked, where typically only the players on the court should be tracked for further analysis, such as counting the players' running distance or average speed and automatic tactical analysis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 283, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 715 + ], + "type": "text", + "content": "Generally, prevailing state-of-the-art trackers [1,6,11,39, 44] consist of several components to accomplish the tracking task: objects localization module, motion based objects association module and appearance based association module. Biased to the data distribution of specific human tracking benchmarks, e.g. MOT17 [27], MOT20 [9] and DanceTrack [33], the components of these trackers have difficulty adapting to sports scenes. Firstly, motivated by surveillance or self-driving applications, current human tracking benchmarks provide tracks for almost all persons in the scenes. While for sports scenes like basketball or football games, generally only the players on the court are what we focus on, hence a specialized training platform is required to make the detectors suitable for sports scenes. More importantly, in MOT17 and MOT20, these trackers highlight Kalman Filter [16] based IoU matching for object association, due to the slow and regular motion of pedestrians. DanceTrack highlights diverse motion rather than fast movement [33], that is, dancers frequently switch the motion direction and relative position. However in sports scenes, we observe fast and variable-speed movement of objects on adjacent frames, i.e. players usually possess high speed and frequently change their running speed in professional sports events, thus constituting barriers in existing motion based association. For instance, as visualized in Fig. 1, the adjacent IoU and Kalman-Filter-based IoU in sports scenes remain lower than that on MOT17 and DanceTrack (More detailed comparison can be found in Fig. 2 and Fig. 3). As a consequence, more suitable motion based association for sports scenes is required. Additionally, compared to the MOT17 and MOT20 datasets in street scenes, the objects appearances in sports scenes are less distinguishable, since not only the players inherently are in similar clothes but also the players are frequently blurred caused by fast camera motion or targets motion. Different from DanceTrack, where generally the dancers are in almost the same clothes and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 703, + 219, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 219, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 219, + 713 + ], + "type": "text", + "content": "* Equal contribution. ☑ : Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "9921" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 288, + 121 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 288, + 121 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 288, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 288, + 121 + ], + "type": "image", + "image_path": "c6093b651a013e26892c96857ae8238d2a7d15e423f7e960ccd6e9ceb0ec2943.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 122, + 209, + 131 + ], + "lines": [ + { + "bbox": [ + 138, + 122, + 209, + 131 + ], + "spans": [ + { + "bbox": [ + 138, + 122, + 209, + 131 + ], + "type": "text", + "content": "(a) Basketball in SportsMOT" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 71, + 529, + 121 + ], + "blocks": [ + { + "bbox": [ + 304, + 71, + 529, + 121 + ], + "lines": [ + { + "bbox": [ + 304, + 71, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 529, + 121 + ], + "type": "image", + "image_path": "ae829d78234f11771cd13c87b41f86c1f0f414483014f7199098186c96303f81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 123, + 432, + 131 + ], + "lines": [ + { + "bbox": [ + 403, + 123, + 432, + 131 + ], + "spans": [ + { + "bbox": [ + 403, + 123, + 432, + 131 + ], + "type": "text", + "content": "(c) MOT17" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 133, + 288, + 183 + ], + "blocks": [ + { + "bbox": [ + 63, + 133, + 288, + 183 + ], + "lines": [ + { + "bbox": [ + 63, + 133, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 63, + 133, + 288, + 183 + ], + "type": "image", + "image_path": "5834c172bde20e9fdfbbd3ba7f2b91e339f0295db95cbf0f7e7331f7e8a7b9a5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 186, + 207, + 194 + ], + "lines": [ + { + "bbox": [ + 140, + 186, + 207, + 194 + ], + "spans": [ + { + "bbox": [ + 140, + 186, + 207, + 194 + ], + "type": "text", + "content": "(b) Football in SportsMOT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 133, + 529, + 183 + ], + "blocks": [ + { + "bbox": [ + 304, + 133, + 529, + 183 + ], + "lines": [ + { + "bbox": [ + 304, + 133, + 529, + 183 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 529, + 183 + ], + "type": "image", + "image_path": "dd5d3028c2d7dea6394a2e18565223a3633a4cb0662063701a63849d7b625af6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 186, + 437, + 194 + ], + "lines": [ + { + "bbox": [ + 399, + 186, + 437, + 194 + ], + "spans": [ + { + "bbox": [ + 399, + 186, + 437, + 194 + ], + "type": "text", + "content": "(d) DanceTrack" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 108, + 198, + 288, + 256 + ], + "blocks": [ + { + "bbox": [ + 108, + 198, + 288, + 256 + ], + "lines": [ + { + "bbox": [ + 108, + 198, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 108, + 198, + 288, + 256 + ], + "type": "image", + "image_path": "29d4c55a59ed743e82bd19e5eb1c3fa30f0b6ddec4e7163085ed4f341f60d639.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 547, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 547, + 326 + ], + "type": "text", + "content": "Figure 1. Sampled sequences from the categories of basketball and football of SportsMOT, MOT17 and DanceTrack. There exist two key properties of SportsMOT: 1) fast and variable-speed motion, i.e. players usually possess high speed and frequently change their running speed (the visualized adjacent IoU and Kalman Filter based adjacent IoU can indicate the property); 2) similar yet distinguishable appearance, that is, players in sports scenes inherently wear jackets with different numbers and usually display distinct postures. We expect SportsMOT to encourage the MOT trackers to promote in both motion-based association and appearance based association." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 206, + 256, + 214, + 262 + ], + "blocks": [ + { + "bbox": [ + 206, + 256, + 214, + 262 + ], + "lines": [ + { + "bbox": [ + 206, + 256, + 214, + 262 + ], + "spans": [ + { + "bbox": [ + 206, + 256, + 214, + 262 + ], + "type": "image", + "image_path": "2a40cd93e332510d9c72f1b185d943206cfc13081a1a8b1d7b8de3ff234478c4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 215, + 256, + 222, + 262 + ], + "blocks": [ + { + "bbox": [ + 215, + 256, + 222, + 262 + ], + "lines": [ + { + "bbox": [ + 215, + 256, + 222, + 262 + ], + "spans": [ + { + "bbox": [ + 215, + 256, + 222, + 262 + ], + "type": "image", + "image_path": "4297c956d949b511450566b2ed26f61d00c336736fcd7e35d5666fb2e9d13c83.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 223, + 256, + 232, + 262 + ], + "blocks": [ + { + "bbox": [ + 223, + 256, + 232, + 262 + ], + "lines": [ + { + "bbox": [ + 223, + 256, + 232, + 262 + ], + "spans": [ + { + "bbox": [ + 223, + 256, + 232, + 262 + ], + "type": "image", + "image_path": "2abb806f47620e8beef110e381485a4b327f968b047b721745244f9f1db4ceae.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 233, + 256, + 240, + 262 + ], + "blocks": [ + { + "bbox": [ + 233, + 256, + 240, + 262 + ], + "lines": [ + { + "bbox": [ + 233, + 256, + 240, + 262 + ], + "spans": [ + { + "bbox": [ + 233, + 256, + 240, + 262 + ], + "type": "image", + "image_path": "1833f6c580ee1ddde54002f1d5e9cf3816232ab2ec2d9d3434f49d957e16c880.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 246, + 256, + 254, + 262 + ], + "blocks": [ + { + "bbox": [ + 246, + 256, + 254, + 262 + ], + "lines": [ + { + "bbox": [ + 246, + 256, + 254, + 262 + ], + "spans": [ + { + "bbox": [ + 246, + 256, + 254, + 262 + ], + "type": "image", + "image_path": "9430a6b413733f12ef19ee98ccd5eee007ca0b6a32986190248f6fb3a54718a7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 255, + 256, + 263, + 262 + ], + "blocks": [ + { + "bbox": [ + 255, + 256, + 263, + 262 + ], + "lines": [ + { + "bbox": [ + 255, + 256, + 263, + 262 + ], + "spans": [ + { + "bbox": [ + 255, + 256, + 263, + 262 + ], + "type": "image", + "image_path": "f30e7e86769324880e288f2e2f7512c63a20be63907eac872f6bfbcea724bdad.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 263, + 256, + 271, + 262 + ], + "blocks": [ + { + "bbox": [ + 263, + 256, + 271, + 262 + ], + "lines": [ + { + "bbox": [ + 263, + 256, + 271, + 262 + ], + "spans": [ + { + "bbox": [ + 263, + 256, + 271, + 262 + ], + "type": "image", + "image_path": "b3f4e10e4f9d68178ce708f523e32b8b74be86608036591837eaeba152acd74d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 272, + 256, + 281, + 262 + ], + "blocks": [ + { + "bbox": [ + 272, + 256, + 281, + 262 + ], + "lines": [ + { + "bbox": [ + 272, + 256, + 281, + 262 + ], + "spans": [ + { + "bbox": [ + 272, + 256, + 281, + 262 + ], + "type": "image", + "image_path": "91ca6723bb13a3da58896b395bf520d2f3146e9f1bbc60862c4fa1af6ca0b498.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 282, + 256, + 290, + 262 + ], + "blocks": [ + { + "bbox": [ + 282, + 256, + 290, + 262 + ], + "lines": [ + { + "bbox": [ + 282, + 256, + 290, + 262 + ], + "spans": [ + { + "bbox": [ + 282, + 256, + 290, + 262 + ], + "type": "image", + "image_path": "3c1ca75ed8bd64dc7c3c3c5c812dcd764f71e7cb6c32e57ba401aeabd5d19b3f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 290, + 198, + 483, + 256 + ], + "blocks": [ + { + "bbox": [ + 290, + 198, + 483, + 256 + ], + "lines": [ + { + "bbox": [ + 290, + 198, + 483, + 256 + ], + "spans": [ + { + "bbox": [ + 290, + 198, + 483, + 256 + ], + "type": "image", + "image_path": "cd290579af55a967b1cbd553c1a82900df4c34a784077437d13905e5f0fa6298.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 321, + 256, + 330, + 262 + ], + "blocks": [ + { + "bbox": [ + 321, + 256, + 330, + 262 + ], + "lines": [ + { + "bbox": [ + 321, + 256, + 330, + 262 + ], + "spans": [ + { + "bbox": [ + 321, + 256, + 330, + 262 + ], + "type": "image", + "image_path": "e9822f6aa6b70b5bea6cb69c7b512fc035dbe5c6abe4e828d88bac9696416f1c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 331, + 256, + 339, + 262 + ], + "blocks": [ + { + "bbox": [ + 331, + 256, + 339, + 262 + ], + "lines": [ + { + "bbox": [ + 331, + 256, + 339, + 262 + ], + "spans": [ + { + "bbox": [ + 331, + 256, + 339, + 262 + ], + "type": "image", + "image_path": "a8764d7458b8c915de171a02a608a23acf8617f09137e2eef3ca6db89356f5db.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 339, + 256, + 347, + 262 + ], + "blocks": [ + { + "bbox": [ + 339, + 256, + 347, + 262 + ], + "lines": [ + { + "bbox": [ + 339, + 256, + 347, + 262 + ], + "spans": [ + { + "bbox": [ + 339, + 256, + 347, + 262 + ], + "type": "image", + "image_path": "ae65fde53686d5266e33ab420be215e9ffab2ad998f7f98d62a2bd8ee0dc242c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 348, + 256, + 356, + 262 + ], + "blocks": [ + { + "bbox": [ + 348, + 256, + 356, + 262 + ], + "lines": [ + { + "bbox": [ + 348, + 256, + 356, + 262 + ], + "spans": [ + { + "bbox": [ + 348, + 256, + 356, + 262 + ], + "type": "image", + "image_path": "4448b7920f3b15788934e4960792dda92224160680a089b7d1f1d553f7046448.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 357, + 256, + 365, + 262 + ], + "blocks": [ + { + "bbox": [ + 357, + 256, + 365, + 262 + ], + "lines": [ + { + "bbox": [ + 357, + 256, + 365, + 262 + ], + "spans": [ + { + "bbox": [ + 357, + 256, + 365, + 262 + ], + "type": "image", + "image_path": "c9df2ffcd97e0e8978fc913c7f77daacc9ef40d0e022dd90f86e665b8e4ecad9.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 367, + 256, + 375, + 262 + ], + "blocks": [ + { + "bbox": [ + 367, + 256, + 375, + 262 + ], + "lines": [ + { + "bbox": [ + 367, + 256, + 375, + 262 + ], + "spans": [ + { + "bbox": [ + 367, + 256, + 375, + 262 + ], + "type": "image", + "image_path": "1da734336ee7669b09d167ef069f1bfcd244473a9d4870a2b5d8fdce1b47e14e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 375, + 256, + 386, + 262 + ], + "blocks": [ + { + "bbox": [ + 375, + 256, + 386, + 262 + ], + "lines": [ + { + "bbox": [ + 375, + 256, + 386, + 262 + ], + "spans": [ + { + "bbox": [ + 375, + 256, + 386, + 262 + ], + "type": "image", + "image_path": "d15aa12fa1e6d47a082a35f3f8f1456fa49b175a258003d7af7d9b4f481e1b42.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "content": "thus having indistinguishable appearance, players in sports scenes inherently wear jackets with different numbers and usually display distinct postures. Thereby, we argue that objects in sports scenes are with similar yet distinguishable appearance, which necessitates the appearance model developing more discriminative and extensive representations." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 408, + 288, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 288, + 649 + ], + "type": "text", + "content": "Considering the analysis above, to advance the development of tracking and sports analysis, we propose a multi-object tracking dataset in sports scenes, termed as SportsMOT. The dataset is large-scale, high-quality and contains dense annotations for every player on the court in various sports scenes. It consists of 240 videos, over 150K frames (almost " + }, + { + "bbox": [ + 46, + 408, + 288, + 649 + ], + "type": "inline_equation", + "content": "15 \\times" + }, + { + "bbox": [ + 46, + 408, + 288, + 649 + ], + "type": "text", + "content": " MOT17 [27]) and over 1.6M bounding boxes (" + }, + { + "bbox": [ + 46, + 408, + 288, + 649 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 408, + 288, + 649 + ], + "type": "text", + "content": " MOT17 [27]) collected from 3 categories of sports, including basketball, volleyball and football. To provide a platform for making the trackers suitable for sports scenes, we split the dataset into training, validation and test subsets, consisting of 45, 45 and 150 video sequences respectively. There exist two core properties of SportsMOT: (1) fast and variable-speed motion, requiring more suitable motion modeling association; (2) similar yet distinguishable appearance, which necessitates the appearance model developing more discriminative and extensive representations. Altogether, we expect SportsMOT to encourage the trackers to promote in both the certain aspects, i.e. motion based association and appearance based association." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "Given the large-scale multi-object tracking dataset SportsMOT, we benchmark some recent tracking approaches and retrain all of them on the training split. We observe IDF1 and AssA metrics are lower than that on MOT17 while the DetA is quite high, indicating that the main chal" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 331, + 547, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 463 + ], + "type": "text", + "content": "lenge of SportsMOT lies in objects association rather than objects localization. To alleviate the issue, we propose a new multi-object tracking framework, dubbed as MixSort, with introducing a MixFormer-like [8] structure as appearance based association to prevailing tracking-by-detection trackers (e.g. ByteTrack [44], OC-SORT [6]). By integrating the original motion based objects association and the designed appearance based association, the performance gets boosted on both SportsMOT and MOT17 benchmarks. Based on MixSort, we perform extensive exploration studies and provide some profound insights into SportsMOT." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 318, + 466, + 526, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 466, + 526, + 477 + ], + "spans": [ + { + "bbox": [ + 318, + 466, + 526, + 477 + ], + "type": "text", + "content": "The main contributions are summarized as follows:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 487, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 317, + 487, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 487, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 317, + 487, + 545, + 559 + ], + "type": "text", + "content": "- We build a new large-scale multi-object tracking dataset in diverse sports scenes, SportsMOT, equipped with two key properties of 1) fast and variable-speed motion and 2) similar yet distinguishable appearance, aiming to advance the development of both tracking and sports analysis." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 570, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 545, + 617 + ], + "type": "text", + "content": "- We benchmark some prevailing trackers on SportsMOT, which reveals that the key challenges lie in objects association and hopefully can facilitate further research." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 629, + 545, + 713 + ], + "type": "text", + "content": "- We propose a new multi-object tracking framework MixSort, with introducing a MixFormer-like structure as appearance based association model to prevailing tracking-by-detection trackers, so as to boost the objects association. Based on MixSort, we perform extensive studies and provide some profound insights into SportsMOT." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9922" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 286, + 129 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 129 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 129 + ], + "type": "table", + "html": "
DatasetVideosFramesLength (s)BboxTracks
MOT171411,235463292,7331,342
MOT20813,4105351,652,0403,456
DanceTrack100105,8555,292-990
SportsMOT240150,3796,0151,629,4903,401
", + "image_path": "163adefa5a4800ef308cfd0f727b55beb06107bb98d33a73f1a93e5dfae100e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 133, + 286, + 156 + ], + "lines": [ + { + "bbox": [ + 47, + 133, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 286, + 156 + ], + "type": "text", + "content": "Table 1. Comparison of statistics between existing human MOT datasets and our SportsMOT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 159, + 133, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 159, + 133, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 159, + 133, + 172 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 180, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 287, + 539 + ], + "type": "text", + "content": "Multi-object tracking datasets. Existing Multi-object tracking datasets usually focus on different scenes, such as autonomous driving, pedestrians on roads, and dancing. For autonomous driving, there are KITTI [14], KITTI360 [22] and BDD100K [42], which focus on pedestrians and vehicles. Besides, other datasets, focusing on only the pedestrians, collect the videos from static and moving cameras. One of the earliest is PETS [12], but it's too simple in some scenes. MOT15 [20] proposes the first large-scale benchmark for Multi-object tracking, followed by MOT17 [27] and MOT20 [9]. It's worth noting that MOT20 focuses on extremely crowded scenes where many pedestrians are occluded, increasing the tracking difficulty greatly in both detection and association. Recently, DanceTrack [33], focusing on dancing scenes, is proposed to encourage trackers to rely less on visual discrimination and depend more on motion analysis. The emphasized properties are uniform appearance and diverse motion. While in SportsMOT, the appearance is similar yet distinguishable, and the players' motion is fast and with variable speed. We expect SportsMOT to encourage algorithms to promote in both appearance and motion association. Besides, SoccerNet-Tracking [7] is presented to track elements in football scenarios. The main difference lies in that, it only contains soccer scenes and tracks almost all elements (players, goalkeepers, referees, balls) on the court without distinction. While SportsMOT contains three types of sports where the objects in basketball scenes are more crowded and thus more challenging (refer to Section 5.3), and only focuses on the players to support further statistics and tactical analysis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 287, + 713 + ], + "type": "text", + "content": "Object association in tracking. Association is a very important task in tracking, where trackers need to associate detections in new frames with existing tracks. For most of the trackers, a similarity matrix (or cost matrix) between new detections and tracks is computed based commonly on motion and appearance cues, which is later fed into Hungarian algorithm [18] to perform association. For example, SORT [4] uses Kalman Filter to predict the location of objects and computes the IoU of detected and predicted bounding boxes as similarity matrix. IOU-Tracker [5] directly computes the IoU without prediction. ByteTrack [44] adds an association phase for detection with low confidence score, which can boost the performance. OC-SORT [6] tries to address the limitation of Kalman Filter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "Appearance cues also play an important role in association. DeepSORT [39] crops the detection from frame images, which are then used by networks to generate re-ID features. Then the motion cues and distance of re-ID features are fused to perform association. FairMOT [45] uses a re-ID branch on a backbone shared by the detection branch to generate re-ID features. In CenterTrack [46], the previous frame is used to help the prediction of tracks. Recently, Transformer [36] is used by some work such as TrackFormer [26] and MOTR [43] to boost the association quality. Our proposed MixSort integrates the motion and appearance cues with a motion modeling component and the designed MixFormer-like structure respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 239, + 424, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 424, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 424, + 251 + ], + "type": "text", + "content": "3. SportsMOT Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 258, + 428, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 428, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 428, + 270 + ], + "type": "text", + "content": "3.1. Dataset Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 276, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 545, + 455 + ], + "type": "text", + "content": "Video Collection. We select three worldwide famous sports, football, basketball, and volleyball, and collect videos of high-quality professional games including NCAA, Premier League, and Olympics from Multi-Sports [21], which is a large dataset in sports area focusing on spatio-temporal action localization. Each category has typical players' formations and motion patterns, and they can effectively represent the diversity of sports scenarios. Only the overhead shots of sports game scenes are used, guaranteeing certain extreme situations do not occur. The proposed dataset consists of 240 video sequences in total, each of which is 720P and 25 FPS. Following the principles of multi-object tracking, each video clip is manually checked to ensure that there are no abrupt viewpoint switches within the video." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 459, + 544, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 544, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 544, + 483 + ], + "type": "text", + "content": "Annotation Pipeline. We annotate the collected videos according to the following guidelines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 485, + 544, + 609 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 317, + 485, + 544, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 485, + 544, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 485, + 544, + 520 + ], + "type": "text", + "content": "- The entire athlete's limbs and torso, excluding any other objects like balls touching the athlete's body, are required to be annotated." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 523, + 544, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 544, + 582 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 544, + 582 + ], + "type": "text", + "content": "- The annotators are asked to predict the bounding box of the athlete in the case of occlusion, as long as the athletes have a visible part of body. However, if half of the athletes' torso is outside the view, annotators should just skip them." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 586, + 544, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 586, + 544, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 586, + 544, + 609 + ], + "type": "text", + "content": "- We ask the annotators to confirm that each player has a unique ID throughout the whole clip." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We provide a customized labeling tool for SportsMOT and a corresponding manual book to annotators. Once they start annotating a new object, the labeling tool automatically assigns a new ID to the object and propagates the bounding box of previous state to the current state, with the help of the single object tracker KCF [15]. Then the generated bounding boxes should be refined by the annotators, so as to improve annotation quality. After carefully reviewing each" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9923" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 168, + 156 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 168, + 156 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 168, + 156 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 168, + 156 + ], + "type": "image", + "image_path": "fda639178cdbf7cb39b2d0186c2e29e82e4fdb3256ac906a590a448d557cbf14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 157, + 287, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 157, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 287, + 212 + ], + "type": "text", + "content": "Figure 2. IoU on adjacent frames. (a) Compared to MOT17 and DanceTrack, SportsMOT has a lower score, indicating that objects have faster motion. (b) In SportsMOT, the category of football has the lowest IoU score, which means that football players often have fast motion." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 284, + 156 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 284, + 156 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 284, + 156 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 284, + 156 + ], + "type": "image", + "image_path": "19c98d15d90c27ace1fbc4181a2f4a722e50374c38bf6c46523b085980628a0f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 156, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 210 + ], + "type": "text", + "content": "Figure 3. Kalman-Filter-based IoU on adjacent frames. (a) Compared to MOT17 and DanceTrack, SportsMOT has a lower score, indicating that objects have more variable-speed motion. (b) In SportsMOT, the category of football has the lowest Kalman-Filter-based IoU score." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 218, + 286, + 293 + ], + "blocks": [ + { + "bbox": [ + 50, + 218, + 286, + 293 + ], + "lines": [ + { + "bbox": [ + 50, + 218, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 50, + 218, + 286, + 293 + ], + "type": "table", + "html": "
CategoryFramesTracksTrack gap len.Track len.Bboxes per frame
Basketball845.41068.7767.99.1
Volleyball360.41238.2335.911.2
Football673.920.5116.1422.112.8
Total626.614.296.6479.110.8
", + "image_path": "6603b96078f85aad876bab17afa6e8a138015f6c7effa0ce4c2321c408da02c5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 285, + 309 + ], + "lines": [ + { + "bbox": [ + 48, + 297, + 285, + 309 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 285, + 309 + ], + "type": "text", + "content": "Table 2. Detailed statistics of the three categories in SportsMOT." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 313, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 361 + ], + "type": "text", + "content": "annotation result, we refine the bounding boxes and IDs that do not satisfy the standards, hence building a high-quality dataset. Finally, the bounding boxes with too small size, i.e. " + }, + { + "bbox": [ + 46, + 313, + 287, + 361 + ], + "type": "inline_equation", + "content": "w < 5" + }, + { + "bbox": [ + 46, + 313, + 287, + 361 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 313, + 287, + 361 + ], + "type": "inline_equation", + "content": "h < 5" + }, + { + "bbox": [ + 46, + 313, + 287, + 361 + ], + "type": "text", + "content": ", are deleted." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 369, + 146, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 146, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 146, + 380 + ], + "type": "text", + "content": "3.2. Dataset Statistic" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 387, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 602 + ], + "type": "text", + "content": "Overview. SportsMOT is a large-scale and high-quality MOT dataset, aiming to advance the development of both sports analysis and multi-object tracking. Table 1 compares the statistics of SportsMOT with the prevailing human tracking datasets, including MOT17, MOT20 and DanceTrack. According to the statistics, SportsMOT has a large number of bounding boxes of over 1.6M, which is comparable to MOT20 and significantly larger than MOT17. Besides, SportsMOT has a large number of video clips, tracks " + }, + { + "bbox": [ + 46, + 387, + 287, + 602 + ], + "type": "inline_equation", + "content": "(2.5 \\times \\text{MOT17}, 3.4 \\times \\text{DanceTrack})" + }, + { + "bbox": [ + 46, + 387, + 287, + 602 + ], + "type": "text", + "content": ", frames " + }, + { + "bbox": [ + 46, + 387, + 287, + 602 + ], + "type": "inline_equation", + "content": "(13.4 \\times \\text{MOT17})" + }, + { + "bbox": [ + 46, + 387, + 287, + 602 + ], + "type": "text", + "content": ". As shown in Table 2, we also compare the basic statistics of each category of SportsMOT. SportsMOT solely provides fine annotations of the players on the court, which are supposed to be tracked for further analysis. To provide a platform for making the trackers suitable for sports scenes, we split the dataset into training, validation and test subsets, consisting of 45, 45 and 150 video sequences respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Fast and Variable-Speed Motion. Motion cues play an important role in object association for multi-object tracking. The existing human-tracking datasets (e.g. MOT17, MOT20 and DanceTrack) generally have certain motion patterns that are distinct with sports scenes, constituting barriers in players tracking. For instance, in MOT17 and MOT20, pedestrians are featured by linear motion with constant speed, which is easily hacked by association strategies with constant velocity assumption. Besides, Dance" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 429, + 145 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 429, + 145 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 429, + 145 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 429, + 145 + ], + "type": "image", + "image_path": "3e823abf075bdab709ff5becb3f1653f6da32e08434215420dc3a059251b821a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 147, + 384, + 155 + ], + "lines": [ + { + "bbox": [ + 374, + 147, + 384, + 155 + ], + "spans": [ + { + "bbox": [ + 374, + 147, + 384, + 155 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 430, + 72, + 542, + 145 + ], + "blocks": [ + { + "bbox": [ + 430, + 72, + 542, + 145 + ], + "lines": [ + { + "bbox": [ + 430, + 72, + 542, + 145 + ], + "spans": [ + { + "bbox": [ + 430, + 72, + 542, + 145 + ], + "type": "image", + "image_path": "51856aaa134953264ab6c1f433fa41092026c734f9df2e8964a94753142cd4a2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 147, + 499, + 155 + ], + "lines": [ + { + "bbox": [ + 488, + 147, + 499, + 155 + ], + "spans": [ + { + "bbox": [ + 488, + 147, + 499, + 155 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 217, + 385, + 274 + ], + "blocks": [ + { + "bbox": [ + 307, + 217, + 385, + 274 + ], + "lines": [ + { + "bbox": [ + 307, + 217, + 385, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 385, + 274 + ], + "type": "image", + "image_path": "7a4a0510c929571eca65970e4f2a745773160dad264ace4225c0b90b778f46c0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 276, + 365, + 286 + ], + "lines": [ + { + "bbox": [ + 328, + 276, + 365, + 286 + ], + "spans": [ + { + "bbox": [ + 328, + 276, + 365, + 286 + ], + "type": "text", + "content": "(a) MOT17" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 387, + 217, + 463, + 274 + ], + "blocks": [ + { + "bbox": [ + 387, + 217, + 463, + 274 + ], + "lines": [ + { + "bbox": [ + 387, + 217, + 463, + 274 + ], + "spans": [ + { + "bbox": [ + 387, + 217, + 463, + 274 + ], + "type": "image", + "image_path": "798f7bbe3ddb04b21dab554643dad22f5e368bec787b7d09db73d8b8989211e8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 276, + 449, + 286 + ], + "lines": [ + { + "bbox": [ + 399, + 276, + 449, + 286 + ], + "spans": [ + { + "bbox": [ + 399, + 276, + 449, + 286 + ], + "type": "text", + "content": "(b) SportsMOT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 290, + 545, + 367 + ], + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 367 + ], + "type": "text", + "content": "Figure 4. Visualization of re-ID features from sampled videos in MOT17, SportsMOT and DanceTrack dataset using t-SNE [35]. The same object is coded by the same color. It indicates that object appearance of SportsMOT is less distinguishable than that of MOT17, while more distinguishable than that of DanceTrack. We expect the appearance model to capture more discriminative and extensive representation for object association." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 466, + 217, + 542, + 274 + ], + "blocks": [ + { + "bbox": [ + 466, + 217, + 542, + 274 + ], + "lines": [ + { + "bbox": [ + 466, + 217, + 542, + 274 + ], + "spans": [ + { + "bbox": [ + 466, + 217, + 542, + 274 + ], + "type": "image", + "image_path": "adbef16cd430c0f29d8def26cc5bc827c6c379c79588c8efdb54a39eae1d233d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 276, + 528, + 285 + ], + "lines": [ + { + "bbox": [ + 477, + 276, + 528, + 285 + ], + "spans": [ + { + "bbox": [ + 477, + 276, + 528, + 285 + ], + "type": "text", + "content": "(c) DanceTrack" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 374, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 625 + ], + "type": "text", + "content": "Track highlights diverse motion rather than fast motion, that is, dancers usually move in more diverse directions with relatively low speed. In contrast, SportsMOT has distinct motion patterns, i.e. fast and variable-speed motion, where the players typically move fast with their running speed or camera speed frequently changing. As illustrated in Fig. 2, among the three datasets, SportsMOT has the lowest IoU score of the objects bounding boxes on adjacent frames, indicating the fast movement. We use the ground truth of previous frames for Kalman Filter prediction. The result and current ground truth are used to calculate Kalman-Filter-based IoU. Seen from Fig. 3, SportsMOT also has the lowest Kalman-Filter-based IoU score on adjacent frames, which suggests that the motion can not be easily modeled by prevailing methods due to the variable-speed movement. Specifically, football has the smallest adjacent IoU and Kalman-Filter-based IoU, which is closely related to the fast running speed, abrupt acceleration or stops. It poses a major challenge for trackers based on simple motion assumptions and also encourages them to model object motion in more dynamic and adaptive ways." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Similar yet Distinguishable Appearance. Object appearance is another kind of cue on which MOT trackers often rely to distinguish different objects. In MOT17 and MOT20, pedestrians are usually distinct in body size and wear different clothes, yielding discriminative visual features. In contrast, the objects in DanceTrack typically wear nearly identical outfits, leading to indistinguishable" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9924" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 285, + 158 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 285, + 158 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 285, + 158 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 285, + 158 + ], + "type": "image", + "image_path": "c7f690504ba803de7ddee35c2c84b4ec3054b9ccbd9dcfda39837a90587c3b76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 159, + 289, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 289, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 289, + 227 + ], + "type": "text", + "content": "Figure 5. Paradigm for computing visual similarity matrix of tracks and detections in MixSort. The center of ground-truth target detection is marked with red dots, the others are green. The blue dashed box indicates the cropped search region. The blank part of similarity vector means that for detections not in the search region, the corresponding value is set to 0." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 232, + 289, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 289, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 289, + 389 + ], + "type": "text", + "content": "appearances. Thereby, DanceTrack highlights solely relying on motion-based association rather than appearance-based association. In SportsMOT, the players also have very similar appearances. However, the players wear jackets with different numbers and usually display distinct postures, thus resulting in similar yet distinguishable appearance. In Fig. 4, we provide visualization of re-ID feature from sampled videos in MOT17, SportsMOT and DanceTrack dataset using t-SNE [35]. It implies that the re-ID features in SportsMOT are similar yet distinguishable compared to MOT17 and DanceTrack. We aim to encourage the trackers to learn more discriminative visual representations for more robust object association." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 392, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 477 + ], + "type": "text", + "content": "Evaluation Metrics. MOTA [3] is the main metric for existing MOT evaluations. However, MOTA focuses more on measuring the accuracy of detection. To highlight the performance of object association, we recommend HOTA [25], AssA and IDF1 [31] as the major evaluation metrics in SportsMOT dataset. HOTA aims to measure the accuracy of detection and association equally." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 487, + 259, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 259, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 259, + 502 + ], + "type": "text", + "content": "4. Multi-Object Tracking on SportsMOT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 507, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 579 + ], + "type": "text", + "content": "In this section, we present our proposed multi-object tracking framework, called MixSort. This framework is designed to enhance the appearance-based association performance and can be applied to any trackers that follow the tracking-by-detection paradigm, such as ByteTrack [44] and OC-SORT [6]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 287, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 601 + ], + "type": "text", + "content": "4.1. MixFormer for Appearance-based Association" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "MixFormer. In this paragraph, we discuss the use of Mix-Former [8] in our proposed framework MixSort. Mix-Former is designed to extract target-specific discriminative features and perform extensive communication between the target and search area, therefore, it is the key component that enables MixSort to compute visual similarities between the templates of tracked objects and detected objects in the search region of the current frames. The original Mix-Former uses a corner-based localization head to predict the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 163 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 163 + ], + "type": "image", + "image_path": "d1a4bb390bb15982083ff7954ca0b4d2d03aeb55eb0c63a0e7fb161bfd97d936.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 167, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 305, + 167, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 547, + 190 + ], + "type": "text", + "content": "Figure 6. The pipeline of MixSort. We use motion model and MixFormer to generate fused similarity matrix for association." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 196, + 547, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 547, + 317 + ], + "type": "text", + "content": "top-left and bottom-right corners of the input template in the search region. However, we modify the corner head by using a heatmap prediction head that predicts the center of the template and generates a confidence heatmap. This allows us to compute the similarity between the detection and the template. To make MixSort suitable for multi-object tracking and accelerate inference speed, we reduce the number of mixed attention modules in MixFormer from 12 to 4. The steps involved in computing the visual similarity matrix are illustrated in Figure 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 319, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 427 + ], + "type": "text", + "content": "Association Strategy. In order to perform association between detections and existing tracks, we use a mixed similarity matrix generated by computing the visual similarity between the target template and the detected objects in the search region of the current frame. Specifically, we obtain the heatmap response at the center of each detection as its visual similarity to the template. The resulting similarity matrix is then combined with the IoU matrix using the Hungarian Algorithm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "text", + "content": "To start, for each existing track " + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "text", + "content": ", we use the Kalman Filter to predict its new location. Then, we crop the current frame centered at the predicted location with a certain scale to obtain the search region " + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "text", + "content": ". By feeding " + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "text", + "content": " and the template " + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "text", + "content": " into MixFormer, we generate a heatmap " + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 427, + 545, + 499 + ], + "type": "text", + "content": " that represents the similarity between the template and search region." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "text", + "content": "Next, for each detection " + }, + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "text", + "content": " whose center is in the search region " + }, + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "text", + "content": ", we set its similarity to track " + }, + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "text", + "content": " as the response in the heatmap " + }, + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 499, + 545, + 559 + ], + "type": "text", + "content": "; the similarity values of other detections are set to 0. Finally, we fuse the visual similarity and the IoU score to obtain the mixed similarity matrix" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 367, + 569, + 545, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 569, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 367, + 569, + 545, + 582 + ], + "type": "interline_equation", + "content": "M = \\alpha \\cdot \\operatorname {I o U} + (1 - \\alpha) \\cdot V \\tag {1}", + "image_path": "5a966584cc9448bc2439ec228ef81784c10f1f2d1c204228bc101b9f046e78ad.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 591, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 591, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 305, + 591, + 545, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 591, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 591, + 545, + 616 + ], + "type": "text", + "content": " is the weight coefficient and " + }, + { + "bbox": [ + 305, + 591, + 545, + 616 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 591, + 545, + 616 + ], + "type": "text", + "content": " represents the visual similarity matrix calculated using MixFormer." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 623, + 413, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 413, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 413, + 636 + ], + "type": "text", + "content": "4.2. MixSort Tracking" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Based on the tracking-by-detection paradigm, the pipeline of MixSort can be generalized as follows. As shown in Figure 6, we first obtain detections using an object detector. Then, we employ a motion model (e.g., Kalman Filter) to predict new locations of existing tracks. Based on the new locations and templates of tracks, we compute a" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9925" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 547, + 224 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 547, + 224 + ], + "type": "table", + "html": "
Training SetupHOTA↑IDF1↑AssA↑MOTA↑DetA↑LocA↑IDs↓Frag↓
CenterTrack [46]Train62.760.048.090.882.190.8104815750
FairMOT [45]Train49.353.534.786.470.283.9992821673
QDTrack [29]Train60.462.347.290.177.588.0637711850
TransTrack [34]Train68.971.557.592.682.791.049929994
GTR [47]Train54.555.845.967.964.889.0956714525
ByteTrack [44]Train62.869.851.294.177.185.632674499
OC-SORT [6]Train71.972.259.894.586.492.430933474
ByteTrackTrain+Val64.171.452.395.978.585.730894216
OC-SORTTrain+Val73.774.061.596.588.592.727283144
MixSort-ByteTrain+Val65.7 (+1.6)74.1 (+2.7)54.8 (+2.5)96.278.885.724724009
MixSort-OCTrain+Val74.1 (+0.4)74.4 (+0.4)62.0 (+0.5)96.588.592.727813199
", + "image_path": "40449ebccba8dd7450f2dc44743207f66e3eb6fddb03905e176d3e8ff826a857.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 73, + 226, + 518, + 238 + ], + "lines": [ + { + "bbox": [ + 73, + 226, + 518, + 238 + ], + "spans": [ + { + "bbox": [ + 73, + 226, + 518, + 238 + ], + "type": "text", + "content": "Table 3. Tracking performance of investigated algorithms on our proposed SportsMOT. The best results are shown in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 252, + 288, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 288, + 336 + ], + "type": "text", + "content": "fused similarity matrix as described above and use it to associate tracks and detections by the Hungarian Algorithm. Finally, for matched tracks and detections, we update the online templates. For unmatched tracks, we keep them until the threshold is reached. For unmatched detections with confidence scores higher than the threshold, new tracks are initialized." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 343, + 178, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 178, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 178, + 356 + ], + "type": "text", + "content": "4.3. Training and Inference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 433 + ], + "type": "text", + "content": "Training. We only consider the training of MixFormer here since the detector remains the same as initial method (e.g. ByteTrack). The original MixFormer is trained on SOT datasets, so we first modify the format of ground truth of MOT datasets, that is, converting the ground truth trajectory of every single player into TrackingNet format [28]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 434, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 482 + ], + "type": "text", + "content": "For each ground-truth bounding box, we compute its corresponding center location " + }, + { + "bbox": [ + 46, + 434, + 287, + 482 + ], + "type": "inline_equation", + "content": "(c_x, c_y)" + }, + { + "bbox": [ + 46, + 434, + 287, + 482 + ], + "type": "text", + "content": " in the low-resolution heatmap. Following CornerNet [19], the ground-truth heatmap response is generated using 2D Gaussian kernel:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 489, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 489, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 88, + 489, + 287, + 514 + ], + "type": "interline_equation", + "content": "h _ {x y} = \\exp \\left(- \\frac {\\left(x - c _ {x}\\right) ^ {2} + \\left(y - c _ {y}\\right) ^ {2}}{2 \\sigma^ {2}}\\right) \\tag {2}", + "image_path": "b016bc679bef8e560df13993f228dea1ff6d127aec4b858871fe2626d1172c05.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 517, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 517, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 517, + 287, + 552 + ], + "type": "text", + "content": " is adaptive to the size of the bounding box. The training loss is a pixel-wise logistic regression with focal loss [23]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 559, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 287, + 602 + ], + "type": "interline_equation", + "content": "L = - \\sum_ {x y} \\left\\{ \\begin{array}{l l} (1 - \\hat {h} _ {x y}) ^ {\\gamma} \\log (\\hat {h} _ {x y}) & , h _ {x y} = 1; \\\\ (1 - h _ {x y}) ^ {\\beta} (\\hat {h} _ {x y}) ^ {\\gamma} \\log (1 - \\hat {h} _ {x y}) & , \\text {o t h e r w i s e}. \\end{array} \\right. \\tag {3}", + "image_path": "c97163b3e51b0962782a8acdbe9ce999cf23efe1245cbf1000987c352e5c0125.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "text", + "content": " are hyper-parameters in focal loss, and we set " + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\beta = 4" + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "text", + "content": " following CornerNet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "Inference. For each track, we maintain only one template for keeping a balance between accuracy and speed. When a detection is matched to an existing track, we directly replace the original template to the new detection, if and only if the ratio of its uncovered (i.e. overlapping with any detected objects) area is larger than a certain threshold, so as to reduce the impact of misleading representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 251, + 455, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 251, + 455, + 265 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 455, + 265 + ], + "type": "text", + "content": "5. Experiments and Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 271, + 414, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 271, + 414, + 284 + ], + "spans": [ + { + "bbox": [ + 306, + 271, + 414, + 284 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 289, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 338 + ], + "type": "text", + "content": "Dataset Split. In benchmark experiments, we follow the default split described in Sec. 3.2. In exploration study, we split the original MOT17 training set into two sets, used for training and validation respectively following CenterTrack." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 340, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 412 + ], + "type": "text", + "content": "Implementation Details. Following ByteTrack and OC-SORT, we use YOLOX [13] as our detector. Using COCO-pretrained model as the initialized weights, we first train the model on CrowdHuman [32] for 80 epochs and then train on SportsMOT for another 80 epochs. The remaining settings are the same as that in ByteTrack." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "text", + "content": "For MixFormer, we initialize the backbone with the model trained on VOT datasets and then fine-tune it on SportsMOT for 300 epochs with learning rate initialized as " + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "text", + "content": " and decreased to " + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "inline_equation", + "content": "1e - 5" + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "text", + "content": " at epoch 200. The optimizer is ADAM [17] with weight decay " + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "text", + "content": ". The sizes of search images and templates are " + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 304, + 412, + 545, + 520 + ], + "type": "text", + "content": " respectively. The max sample interval is set to 10. For every tracking result, we apply linear interpolation as post-processing, with maximum gap set to 20." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 527, + 421, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 421, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 421, + 539 + ], + "type": "text", + "content": "5.2. Benchmark Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "We evaluate several representative methods of three kinds on our dataset. ByteTrack [44], OC-SORT [6] and QDTrack [29] are trackers in tracking-by-detection paradigm. CenterTrack [46] and FairMOT [45] perform joint detection and tracking in one stage. TransTrack [34] and GTR [47] are trackers based on Transformer. Most of the current best multi-object tracking algorithms belong to tracking-by-detection paradigm, however, due to the separation of detection and tracking, the information cannot be shared completely. Joint-detection-and-tracking paradigm couples the two modules, with the goal of boosting the performance of each. Transformer-based-tracking methods are relatively new but have achieved great performance. Despite its huge potential, the model complexity and calcula" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9926" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 288, + 110 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 288, + 110 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 288, + 110 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 288, + 110 + ], + "type": "table", + "html": "
α10.90.80.70.60.50.40.30.20.10
basketball65.966.166.265.965.364.863.660.756.747.126.9
volleyball76.076.876.576.476.576.976.475.573.769.140.9
football71.972.472.372.472.572.873.272.972.671.465.7
", + "image_path": "9b51716ecd2c75621a6bff1f443e753cf70e24f0b1d6c3d8cf4b2306c3cdfb0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 287, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 178 + ], + "type": "text", + "content": "tion cost are much higher, resulting in large memory and long training time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 180, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 263 + ], + "type": "text", + "content": "All training settings including the number of epochs and change of learning rate are consistent with original papers. According to different default settings, we follow the commonly used pretraining datasets, such as CrowdHuman [32], COCO [24] and ImageNet [10], and apply SportsMOT-train with or without other datasets for finetuning for different methods. We compare the results in Tab. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 264, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 287, + 373 + ], + "type": "text", + "content": "In sports scenes, the clear appearance and sparse density of objects allow current mature detection frameworks to generate bounding boxes with high accuracy. However, specialized detectors need to be trained for not detecting audience and referees. The key challenges are fast speed and motion blur, which forces us to pay more attention to improve the association performance. Besides, The wide range of HOTA and MOTA denotes SportsMOT is more distinguishable among different kinds of algorithms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 381, + 156, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 156, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 156, + 395 + ], + "type": "text", + "content": "5.3. Exploration Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 400, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 287, + 425 + ], + "type": "text", + "content": "In this section, we perform extensive studies on the proposed MixSort and SportsMOT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 431, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 563 + ], + "type": "text", + "content": "Effectiveness of the proposed association module. We evaluate the effectiveness of MixSort by applying it to two state-of-the-art trackers, OC-SORT [6] and ByteTrack [44], which follow the tracking-by-detection paradigm and useYOLOX as their detector. The evaluation is conducted on the SportsMOT test set, and the results are presented in Tab. 3. Our experiments show that MixSort significantly improves the performance of both trackers, with OC-SORT achieving a 0.4 HOTA increase and ByteTrack achieving a 1.6 HOTA increase on SportsMOT. This demonstrates the effectiveness of MixSort in enhancing the association." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "text", + "content": "Appearance-based vs. Motion-based association. We have demonstrated that MixSort can improve association performance. In this paragraph, we investigate the impact of the fusion weight " + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "text", + "content": " on the ability of appearance cues to aid in conventional motion-based association. We evaluate OC-SORT with MixSort on the three categories in the SportsMOT test set using " + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "text", + "content": " values ranging from 1 to 0 in Eq. (1). The results, presented in Tab. 4, reveal that pure motion-based association " + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "inline_equation", + "content": "(\\alpha = 1)" + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "text", + "content": " outperforms pure appearance-based association " + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "inline_equation", + "content": "(\\alpha = 0)" + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "text", + "content": " in all categories, underscoring the significance of motion cues in sports scenes. Moreover, fused association surpasses both" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 138 + ], + "blocks": [ + { + "bbox": [ + 47, + 114, + 287, + 148 + ], + "lines": [ + { + "bbox": [ + 47, + 114, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 287, + 148 + ], + "type": "text", + "content": "Table 4. Comparison of the HOTA metric of basketball, volley and football under different fusion parameters " + }, + { + "bbox": [ + 47, + 114, + 287, + 148 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 114, + 287, + 148 + ], + "type": "text", + "content": " on SportsMOT test set. The models are trained on SportsMOT training set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 71, + 545, + 138 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 138 + ], + "type": "table", + "html": "
IoUMotionMix.HOTA↑IDF1↑AssA↑MOTA↑IDs↓
✓71.571.258.195.94329
✓64.263.948.791.125947
✓✓64.171.452.395.93089
✓✓73.874.461.696.63203
✓✓✓65.774.154.896.12469
", + "image_path": "3d6c6405994a2d1737bf45dd184849f8efd2e09005eae4f5fea6cb0ba4b66834.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 143, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 304, + 143, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 143, + 547, + 198 + ], + "type": "text", + "content": "Table 5. Results of the ablation experiment on SportsMOT test set. IoU means computing IoU between detections and the last location of existing tracks for association, while Motion means using Kalman filter to predict the location of tracks. The models are trained on SportsMOT training and validation set." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 206, + 545, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 242 + ], + "type": "text", + "content": "pure motion-based and appearance-based association, suggesting that both motion and appearance cues should be considered jointly for optimal results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 243, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 339 + ], + "type": "text", + "content": "Our analysis of the three categories reveals that appearance cues provide the most significant improvement for football videos (+1.3), followed by volleyball (+0.9) and basketball (+0.3). Combining this with Fig. 2, which indicates that the adjacent IoU of football games is the smallest (i.e., motion is fastest) among the three categories, we can conclude that scenes with faster motion are more dependent on appearance cues." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 346, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 465 + ], + "type": "text", + "content": "Ablation study on MixSort. We ablate important components of our tracker (MixSort based on ByteTrack) including IoU, Motion (Kalman Filter) and MixSort on SportsMOT test set. The results are presented in Tab. 5. Surprisingly, we found that simple IoU without using motion prediction outperformed IoU with motion prediction by a large margin (from 64.1 HOTA to 71.5 HOTA), indicating that the Kalman filter, which assumes linear motion models, performed poorly on SportsMOT, where the motion patterns are far more complex than in previous datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 466, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 526 + ], + "type": "text", + "content": "Furthermore, we observed that MixSort played a crucial role in boosting the performance of the tracker significantly. By fusing the IoU and MixSort cues, our method achieved the best performance of 73.8 HOTA compared to 71.5 HOTA of simple IoU in our experiments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 604 + ], + "type": "text", + "content": "Analysis on different categories of SportsMOT. While the three SportsMOT categories share some common characteristics such as fast motion and similar appearance, they also have distinct features due to the different types of games. In this section, we analyze the results of our experiments on the three categories." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "content": "We first use the best HOTA metric from Tab. 4 to represent the overall difficulty of a category. Based on this metric, we find that basketball videos (66.17) are the most difficult, followed by football (73.19), and finally volleyball (76.91)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Next, we consider the appearance-based association (i.e. " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " in Tab. 4). We observe a notable gap between the HOTA metrics of the three categories, where basketball (26.94) has a much lower HOTA than volleyball (40.94) and football (65.65). Similarly, basketball remains the most dif" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9927" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 184 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 184 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 184 + ], + "type": "table", + "html": "
MOT17-testMOT17-val
HOTA↑IDF1↑AssA↑MOTA↑DetA↑IDs↓HOTA↑IDF1↑AssA↑MOTA↑DetA↑IDs↓
QDTrack [29]53.966.352.768.755.63378------
MOTR [43]57.268.455.871.958.92115------
GTR [47]59.171.557.075.361.6285963.075.966.271.360.4-
ByteTrack [44]63.177.362.080.364.52196-79.7-76.7-159
OC-SORT [6]63.277.563.478.063.2195068.079.369.977.9--
MixSort-Byte64.078.764.279.364.1223569.481.171.379.968.2155
MixSort-OC63.477.863.278.963.8150969.280.671.578.967.4135
", + "image_path": "661f297b75b98168e34e939782c8ba6d1382fb3eccf786427fe1532acc4a4e1b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 63, + 185, + 530, + 196 + ], + "lines": [ + { + "bbox": [ + 63, + 185, + 530, + 196 + ], + "spans": [ + { + "bbox": [ + 63, + 185, + 530, + 196 + ], + "type": "text", + "content": "Table 6. Comparison of the state-of-the-art methods under the \"private detector\" protocol on MOT17-test set and MOT17-val set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 209, + 288, + 255 + ], + "blocks": [ + { + "bbox": [ + 47, + 209, + 288, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 209, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 288, + 255 + ], + "type": "table", + "html": "
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
ByteTrack47.151.931.588.270.5
MixSort-Byte46.753.031.985.868.6
", + "image_path": "61b5d2727d778ab042e9df19bdc3b39655adcf02fc0c19da4ee435276cd9e999.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 258, + 287, + 303 + ], + "lines": [ + { + "bbox": [ + 47, + 258, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 287, + 303 + ], + "type": "text", + "content": "Table 7. Comparison of ByteTrack and MixSort-Byte on DanceTrack validation set. For MixSort-Byte, the fuse parameter " + }, + { + "bbox": [ + 47, + 258, + 287, + 303 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 258, + 287, + 303 + ], + "type": "text", + "content": " is 0.9, which results in the highest HOTA among " + }, + { + "bbox": [ + 47, + 258, + 287, + 303 + ], + "type": "inline_equation", + "content": "\\{0.6, 0.7, 0.8, 0.9, 0.95\\}" + }, + { + "bbox": [ + 47, + 258, + 287, + 303 + ], + "type": "text", + "content": ". The models are trained on DanceTrack training set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 308, + 286, + 376 + ], + "blocks": [ + { + "bbox": [ + 50, + 308, + 286, + 376 + ], + "lines": [ + { + "bbox": [ + 50, + 308, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 50, + 308, + 286, + 376 + ], + "type": "table", + "html": "
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
SportsMOToverall65.774.154.896.178.8
basketball60.867.846.897.379.1
volleyball72.587.066.896.578.7
football66.473.656.394.978.5
SoccerNet62.973.955.587.871.5
", + "image_path": "eb961e073f1e181486560719f633b935881de90317bedecf40a9c0bff59af2f7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 380, + 287, + 424 + ], + "lines": [ + { + "bbox": [ + 47, + 380, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 287, + 424 + ], + "type": "text", + "content": "Table 8. Results of MixSort-Byte on SportsMOT and SoccerNet test set. The models are trained on SportsMOT training and validation set and SoccerNet training set respectively and the hyperparameters are the same." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 431, + 287, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 287, + 454 + ], + "type": "text", + "content": "ficult in the motion-based association (i.e. " + }, + { + "bbox": [ + 47, + 431, + 287, + 454 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 47, + 431, + 287, + 454 + ], + "type": "text", + "content": " in Tab. 4), while the volleyball becomes the easiest." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 455, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 550 + ], + "type": "text", + "content": "We believe that the differences in difficulty arise from several factors, including the size of the game court and the degree of physical confrontation among players. For instance, basketball scenes are played on smaller courts and involve more physical contact between players than football scenes. This can lead to more occlusion and blur in basketball videos, making the association task more challenging than in football scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 554, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 685 + ], + "type": "text", + "content": "Comparison with DanceTrack. We evaluate MixSort-Byte on the DanceTrack validation set and compare the results with that of ByteTrack as shown in Tab. 7. Unlike the results on SportsMOT where MixSort brings significant improvement, on DanceTrack the original ByteTrack performs better instead, with HOTA, MOTA and DetA metrics all higher than MixSort-Byte. This indicates that appearances in our proposed dataset SportsMOT are similar yet distinguishable, while those in DanceTrack are much harder to distinguish. Therefore SportsMOT highlights both the motion-based and appearance-based associations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Comparison with SoccerNet. We conducted experiments to compare SoccerNet that focuses only on soccer" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 209, + 547, + 262 + ], + "blocks": [ + { + "bbox": [ + 306, + 209, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 306, + 209, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 547, + 262 + ], + "type": "table", + "html": "
HOTA↑IDF1↑AssA↑MOTA↑DetA↑
ByteTrack64.171.452.395.978.5
ByteTrack+ReID64.872.253.496.178.8
MixSort-Byte65.774.154.896.178.8
", + "image_path": "aac8399450c94c8ea389639d9bd2f86908d6e164afab2479dd419a26d2f799f3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 266, + 547, + 311 + ], + "lines": [ + { + "bbox": [ + 305, + 266, + 547, + 311 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 547, + 311 + ], + "type": "text", + "content": "Table 9. Comparison of ByteTrack, ByteTrack with ReID model and MixSort-Byte on SportsMOT test set. The ReID model is the same as in DeepSORT [4] and all the models are trained on SportsMOT training set and the best results are shown in **bold**." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 318, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 547, + 462 + ], + "type": "text", + "content": "scenes with SportsMOT. Results shown in Tab. 8 suggest that SportsMOT is a challenging dataset with varying levels of difficulty across different sports. Specifically, basketball is proved to be the most difficult with the lowest HOTA of 60.8, while volleyball and football are relatively easier. Notably, MixSort obtains higher HOTA on overall SportsMOT than on SoccerNet, which is mainly because all the elements on the court are to be tracked in SoccerNet, leading to more false detections and much lower DetA (71.5 vs. 78.8). Yet, with a higher AssA on SoccerNet despite these false detections, it underscores SportsMOT's challenging association and significance for sports tracking." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 465, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 547, + 598 + ], + "type": "text", + "content": "MixSort vs. ReID models. To validate the effectiveness of the MixFormer-like model in MixSort, we present a comparison with ReID models in Tab. 9. The results show that ByteTrack with the ReID model outperforms the original ByteTrack, highlighting the value of appearance-based association on SportsMOT. Notably, MixSort-Byte further improves upon ByteTrack with ReID by 0.9, 1.9, and 1.4 in HOTA, IDF1, and AssA. This underscores MixSort's superior appearance modeling, capable of extracting broader and more discriminative representations and facilitating efficient offline learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 601, + 547, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 547, + 686 + ], + "type": "text", + "content": "More results on MOT17 The propose MixSort can be applied to any trackers following this paradigm and achieve state-of-the-art performance on SportsMOT. To further validate the effectiveness of MixSort, we compare MixSort with the prevailing trackers on MOT17 validation set and test set under the private detection protocol in Tab. 6. MixSort-Byte and MixSort-OC outperform these trackers in HOTA, IDF1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Detailed Challenging attributes of SportsMOT. For a comprehensive statistical analysis and performance evalua" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9928" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 71, + 545, + 152 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 152 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 152 + ], + "type": "table", + "html": "
Attr.Definition of sequenceAttr.Definition of target box
CMAbrupt motion of the cameraLRThe target box is smaller than a threshold
OVThe target completely leaves the video frameFOCThe area of target box intersecting with others is larger than 0.9
MBThe target region is blurred due to target or camera motionPOCThe area of target box intersecting with others is in range [0.4, 0.9]
FMThe total average IoU of target boxes in adjacent frames is smaller than a thresholdDEFThe sum of the relative changes in the length and width of the target boxes in adjacent frames is greater than a threshold
", + "image_path": "e157b0e3a1e26d2b3563c1e1448f68ae26a0e19a9576c60c21e55fae15e6e7ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 154, + 546, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 154, + 546, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 546, + 177 + ], + "type": "text", + "content": "Table 10. Definition of attributes. For the attributes in the right half, we first calculate the proportion of boxes in each sequence that meet the criteria defined in the table. If this proportion exceeds a certain threshold, we consider the sequence to exhibit that attribute." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 184, + 157, + 269 + ], + "blocks": [ + { + "bbox": [ + 49, + 184, + 157, + 269 + ], + "lines": [ + { + "bbox": [ + 49, + 184, + 157, + 269 + ], + "spans": [ + { + "bbox": [ + 49, + 184, + 157, + 269 + ], + "type": "image", + "image_path": "5783a89997d202b754f3a9522ff5a26882a5a3a8b3faed16a52d02743e75962e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 272, + 118, + 281 + ], + "lines": [ + { + "bbox": [ + 89, + 272, + 118, + 281 + ], + "spans": [ + { + "bbox": [ + 89, + 272, + 118, + 281 + ], + "type": "text", + "content": "(a) Total" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 184, + 546, + 269 + ], + "blocks": [ + { + "bbox": [ + 159, + 184, + 546, + 269 + ], + "lines": [ + { + "bbox": [ + 159, + 184, + 546, + 269 + ], + "spans": [ + { + "bbox": [ + 159, + 184, + 546, + 269 + ], + "type": "image", + "image_path": "27f7dfbc3f9b4352efb65f2c4c065077efbc316870797f9e725588a64e0cbfc9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 272, + 383, + 281 + ], + "lines": [ + { + "bbox": [ + 323, + 272, + 383, + 281 + ], + "spans": [ + { + "bbox": [ + 323, + 272, + 383, + 281 + ], + "type": "text", + "content": "(b) Category-wise" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 299, + 157, + 384 + ], + "blocks": [ + { + "bbox": [ + 49, + 299, + 157, + 384 + ], + "lines": [ + { + "bbox": [ + 49, + 299, + 157, + 384 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 157, + 384 + ], + "type": "image", + "image_path": "0e422629dbcde8ff0be1e8f3c96621cf300de85f01125fe2607b6724e5026aea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 388, + 118, + 396 + ], + "lines": [ + { + "bbox": [ + 89, + 388, + 118, + 396 + ], + "spans": [ + { + "bbox": [ + 89, + 388, + 118, + 396 + ], + "type": "text", + "content": "(a) Total" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 399, + 509, + 411 + ], + "lines": [ + { + "bbox": [ + 82, + 399, + 509, + 411 + ], + "spans": [ + { + "bbox": [ + 82, + 399, + 509, + 411 + ], + "type": "text", + "content": "Figure 8. Average HOTA of sequences in each attribute on SportsMOT dataset. MixSort-OC is employed for analysis." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 160, + 299, + 546, + 384 + ], + "blocks": [ + { + "bbox": [ + 159, + 284, + 433, + 295 + ], + "lines": [ + { + "bbox": [ + 159, + 284, + 433, + 295 + ], + "spans": [ + { + "bbox": [ + 159, + 284, + 433, + 295 + ], + "type": "text", + "content": "Figure 7. Distribution of sequences in each attribute on SportsMOT dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 299, + 546, + 384 + ], + "lines": [ + { + "bbox": [ + 160, + 299, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 160, + 299, + 546, + 384 + ], + "type": "image", + "image_path": "b075346d36606c1fafc8fb953395c42679dcb504ebc501fa02522599071d1088.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 387, + 383, + 397 + ], + "lines": [ + { + "bbox": [ + 323, + 387, + 383, + 397 + ], + "spans": [ + { + "bbox": [ + 323, + 387, + 383, + 397 + ], + "type": "text", + "content": "(b) Category-wise" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 417, + 289, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 417, + 289, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 289, + 646 + ], + "type": "text", + "content": "tion of the SportsMOT dataset, we categorize and present eight distinct challenging attributes: camera motion (CM), out of view (OV), motion blur (MB), fast motion (FM), low resolution (LR), full occlusion (FOC), partial occlusion (POC), and large deformation (DEF). The definitions for each attribute are specified in Tab. 10, with their corresponding distribution illustrated in Fig. 7. Human selection is used to identify the CM, OV, and MB attributes, while metrics calculation helps determine the others. The diversity and complexity of the scenarios featured in Fig. 7a suggest that our three chosen sports games provide a representative snapshot of broader sports analysis based on tracking states. Furthermore, in comparison to SoccerNetTracking [7], our SportsMOT offers a more diverse set of challenges observed in sports. These challenges range from the frequent occlusion seen in basketball and volleyball, to the low resolution typical of football, as depicted in Fig. 7b. We also assess the difficulty level posed by different attributes in the following." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 653, + 287, + 715 + ], + "type": "text", + "content": "Difficulty analysis based on HOTA scores. We employ the average HOTA scores of sequences corresponding to each attribute as an indicator of that attribute's difficulty. The results are depicted in Fig. 8. The low HOTA score for full occlusion, evident in Fig. 8a, highlights its challeng-" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 417, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 466 + ], + "type": "text", + "content": "ing nature in MOT within the sports field. Furthermore, an analysis of average HOTA scores shows that basketball consistently has lower scores than volleyball and football for all attributes (Fig. 8b), indicating its increased complexity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 479, + 378, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 378, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 378, + 491 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 500, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 621 + ], + "type": "text", + "content": "In this paper, we have introduced SportsMOT, a large-scale multi-object tracking dataset in sports scenes. SportsMOT is characterized with two key properties: 1) fast and variable-speed motion and 2) similar yet distinguishable appearance. We have empirically investigated several prevailing MOT trackers on the SportsMOT dataset. We have also proposed a new MOT framework MixSort, introducing a MixFormer-like association module. Hopefully, SportsMOT can provide a platform for facilitating both sports analysis and multi-object tracking." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9929" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "text", + "content": "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 170 + ], + "type": "text", + "content": "[2] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 941-951, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 214 + ], + "type": "text", + "content": "[3] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 261 + ], + "type": "text", + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 1, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 316 + ], + "type": "text", + "content": "[5] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In 2017 14th IEEE international conference on advanced video and signal based surveillance (AVSS), pages 1-6. IEEE, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "text", + "content": "[6] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris Kitani. Observation-centric sort: Rethinking sort for robust multi-object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9686-9696, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 441 + ], + "type": "text", + "content": "[7] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3491-3502, 2022. 3, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 288, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 496 + ], + "type": "text", + "content": "[8] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13608-13618, 2022. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 288, + 553 + ], + "type": "text", + "content": "[9] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 632 + ], + "type": "text", + "content": "[11] Yunhao Du, Yang Song, Bo Yang, and Yanyun Zhao. Strongsort: Make deepsort great again. arXiv preprint arXiv:2202.13514, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 677 + ], + "type": "text", + "content": "[12] James Ferryman and Ali Shahrokni. Pets2009: Dataset and challenge. In 2009 Twelfth IEEE international workshop on performance evaluation of tracking and surveillance, pages 1-6. IEEE, 2009. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[13] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[15] Sion Hannuna, Massimo Camplani, Jake Hall, Majid Mirmehdi, Dima Damen, Tilo Burghardt, Adeline Paiement, and Lili Tao. Ds-kcf: a real-time tracker for rgb-d data. Journal of Real-Time Image Processing, 16(5):1439-1458, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "text", + "content": "[16] Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 82(1):35-45, 1960. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "text", + "content": "[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 294 + ], + "type": "text", + "content": "[19] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 337 + ], + "type": "text", + "content": "[20] Laura Leal-Taixe, Anton Milan, Ian Reid, Stefan Roth, and Konrad Schindler. Motchallenge 2015: Towards a benchmark for multi-target tracking. arXiv preprint arXiv:1504.01942, 2015. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "type": "text", + "content": "[21] Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang. Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13536-13545, October 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "text", + "content": "[22] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "type": "text", + "content": "[25] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129(2):548-578, 2021. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 603, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 656 + ], + "type": "text", + "content": "[26] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8844-8854, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[27] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016. 1, 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "type": "text", + "content": "[28] Matthias Müller, Adel Bibi, Silvio Giancola, Salman Al-Subaihi, and Bernard Ghanem. Trackingnet: A large-scale" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9930" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "dataset and benchmark for object tracking in the wild. In Proceedings of the European Conference on Computer Vision, ECCV, 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "text", + "content": "[29] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 164-173, 2021. 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 229 + ], + "type": "text", + "content": "[30] Jinlong Peng, Changan Wang, Fangbin Wan, Yang Wu, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Chained-tracker: Chaining paired attentive regression results for end-to-end joint multiple-object detection and tracking. In European conference on computer vision, pages 145-161. Springer, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "type": "text", + "content": "[31] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 319 + ], + "type": "text", + "content": "[32] Shuai Shao, Zijian Zhao, Boxun Li, Tete Xiao, Gang Yu, Xiangyu Zhang, and Jian Sun. Crowdhuman: A benchmark for detecting human in a crowd. arXiv preprint arXiv:1805.00123, 2018. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 374 + ], + "type": "text", + "content": "[33] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 419 + ], + "type": "text", + "content": "[34] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 421, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 288, + 453 + ], + "type": "text", + "content": "[35] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 288, + 498 + ], + "type": "text", + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "type": "text", + "content": "[37] Qiang Wang, Yun Zheng, Pan Pan, and Yinghui Xu. Multiple object tracking with correlation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3876-3886, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 589 + ], + "type": "text", + "content": "[38] Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, and Shengjin Wang. Towards real-time multi-object tracking. In European Conference on Computer Vision, pages 107-122. Springer, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 634 + ], + "type": "text", + "content": "[39] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[40] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12352-12361, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[41] Yihong Xu, Yutong Ban, Guillaume Delorme, Chuang Gan, Daniela Rus, and Xavier Alameda-Pineda. Transcenter:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 385 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Transformers with dense queries for multiple-object tracking. arXiv preprint arXiv:2103.15145, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 161 + ], + "type": "text", + "content": "[42] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[43] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision, pages 659–675. Springer, 2022. 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[44] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In European Conference on Computer Vision, pages 1-21. Springer, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 307 + ], + "type": "text", + "content": "[45] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129(11):3069-3087, 2021. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "text", + "content": "[46] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In European Conference on Computer Vision, pages 474-490. Springer, 2020. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 342, + 547, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 547, + 385 + ], + "type": "text", + "content": "[47] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8771-8780, 2022. 6, 8" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9931" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/8d4ee68a-1062-4c0b-81d9-4f01bf36823f_content_list.json b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/8d4ee68a-1062-4c0b-81d9-4f01bf36823f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bab0647d01815f82d304b2c97a5d2afa1b28df9e --- /dev/null +++ b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/8d4ee68a-1062-4c0b-81d9-4f01bf36823f_content_list.json @@ -0,0 +1,3203 @@ +[ + { + "type": "text", + "text": "Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet", + "text_level": 1, + "bbox": [ + 99, + 128, + 870, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yannic Neuhaus", + "bbox": [ + 138, + 203, + 272, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maximilian Augustin", + "bbox": [ + 313, + 203, + 486, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Valentyn Boreiko", + "bbox": [ + 529, + 203, + 671, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthias Hein", + "bbox": [ + 715, + 203, + 830, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tübingen AI Center - University of Tübingen", + "bbox": [ + 302, + 220, + 666, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benchmark performance of deep learning classifiers alone is not a reliable predictor for the performance of a deployed model. In particular, if the image classifier has picked up spurious features in the training data, its predictions can fail in unexpected ways. In this paper, we develop a framework that allows us to systematically identify spurious features in large datasets like ImageNet. It is based on our neural PCA components and their visualization. Previous work on spurious features often operates in toy settings or requires costly pixel-wise annotations. In contrast, we work with ImageNet and validate our results by showing that presence of the harmful spurious feature of a class alone is sufficient to trigger the prediction of that class. We introduce the novel dataset \"Spurious ImageNet\" which allows to measure the reliance of any ImageNet classifier on harmful spurious features. Moreover, we introduce SpuFix as a simple mitigation method to reduce the dependence of any ImageNet classifier on previously identified harmful spurious features without requiring additional labels or retraining of the model. We provide code and data at https://github.com/YanNeu/spuriousImagenet.", + "bbox": [ + 76, + 305, + 473, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 648, + 209, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has led to tremendous progress in image classification [37, 50] and natural language processing [14]. Over the years, however, it has become apparent, that evaluating predictive performance on a fixed test set is not necessarily indicative of the performance when image classifiers are deployed in the wild. Several potential failure cases have been discovered. This starts with a lack of robustness due to image corruptions [31], adversarial perturbations [68], and arbitrary predictions on out-of-distribution inputs [45, 32, 30]. In this paper, we consider the problem of identifying and debugging image classifiers from spurious features [2]. Spurious features in image classification are features that co-occur with the actual class object and are picked up by the classifier. In the worst case, they lead to shortcut learning [25], where only the spurious but not the", + "bbox": [ + 75, + 674, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spurious Features in Training Data", + "bbox": [ + 578, + 273, + 813, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bird feeder", + "bbox": [ + 526, + 289, + 601, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "graffiti", + "bbox": [ + 632, + 289, + 678, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eucalyptus", + "bbox": [ + 707, + 289, + 779, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "label", + "bbox": [ + 815, + 289, + 849, + 301 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e837ae86e2d771d5286c9eb2a22c0f9fbc3d3a30bd4fa5dc8d07eff2f74a74d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 303, + 609, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7ab984c8c6ee8f1e9e913a5c4b7eb2342d45b5dba6fd82527e0b400521da1f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 303, + 699, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d06c41754196ab95ce2b2602d999b7e23aaaf3eec95d35183a2d02b233c11b52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 303, + 789, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6bb75441b7caa1ae1fdb35a463ae277a98cbca646e9806df92900c645f19b6d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 303, + 875, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hummingbird", + "bbox": [ + 517, + 372, + 611, + 387 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Freight Car", + "bbox": [ + 617, + 372, + 694, + 387 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Koala", + "bbox": [ + 723, + 372, + 764, + 386 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hard Disc", + "bbox": [ + 797, + 372, + 867, + 386 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Images from the web with spurious feature but no class features classified as class below", + "bbox": [ + 547, + 393, + 846, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3fe0b619236e6498566704a9b635cb5dc304dc36671ec78be38991bb951c99e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 422, + 607, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9555789428e3d27b20db3b5883a60a98a52c9698b51351a2a852c4d1bf7ceec9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 422, + 699, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a15bba2a618e26b41249dcfa87a907ebacbb4b5627cd067b1fde31a871b1ce8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 422, + 789, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/963df71285c03eaa75050a0317ab26178288277db62bcdca8efecbf19517f2a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 422, + 875, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d050b9811842346325d13ed718729889cd50fa44f41bf4d6611ca5caad206e79.jpg", + "image_caption": [ + "Figure 1: Top: Examples of spurious features found via our neural PCA components but not in previous study [61]. Bottom: We validate our spurious features by mining images from the web showing only the spurious feature but not the class. They are classified by four ImageNet models as the corresponding class. Some of them even contain ImageNet classes (bees on feeder, grasshopper in leaves)." + ], + "image_footnote": [], + "bbox": [ + 519, + 491, + 607, + 560 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10e87f53c103305227963b7100d97c882846e1a51d6f6f2f4a213bd99712bb54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 491, + 699, + 559 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86c91affd6b9f248ab60af7e0f3d2620e962cb443f18e5c5718e3a37d97c62d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 491, + 789, + 559 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00e18b10c73c49299fd1881ca4e8a1bb66343739cd0c083d2a8cb4690e4a7429.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 491, + 875, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hummingbird", + "bbox": [ + 516, + 561, + 611, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Freight Car", + "bbox": [ + 617, + 561, + 694, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Koala", + "bbox": [ + 723, + 561, + 764, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hard Disc", + "bbox": [ + 797, + 561, + 867, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "correct feature is associated with the class, e.g., [86] found that a pneumonia detector's bad generalization across hospitals was caused by the neural network learning to identify the hospital where the training data originated from. A weaker form of spurious feature (at least from a learning perspective) is the case when the classifier picks up the correct class features, e.g., of a hummingbird, but additionally associates a spurious feature, e.g., a bird feeder, with the class as they appear together on a subset of the training set. This becomes a harmful spurious feature if only the spurious feature without the class feature is sufficient to trigger the classification of that class, see Fig. 1 for an illustration of", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20235", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such spurious features found via our method. Harmful spurious features are difficult to detect and thus can easily go unnoticed, leading to unexpected behaviour of a deployed image classifier.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper we make the following key contributions:", + "bbox": [ + 76, + 152, + 437, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- we develop a pipeline for the detection of harmful spurious features with little human supervision based on our class-wise neural PCA (NPCA) components of an adversarially robust classifier together with their Neural PCA Feature Visualization (NPFV).", + "- unlike prior work, which used masking images or pixelwise annotations, we validate our found spurious features by using our NPCA components to find real images containing only the spurious feature but not the class object.", + "- using these images we create the dataset \"Spurious ImageNet\" and propose a measure for dependence on spurious features. We do a large-scale evaluation of state-of-the-art (SOTA) ImageNet models. We show that the spurious features found for the robust model generalize to non-robust classifiers. Moreover, we analyze the influence of different training setups, e.g. pre-training on ImageNet21k or larger datasets like LAION.", + "- we develop SpuFix, a technique to mitigate the dependence on identified harmful spurious features without requiring new labels or retraining, and show how to transfer it to any ImageNet classifier. SpuFix consistently improves the dependence on harmful spurious features even for SOTA models with negligible impact on test accuracy." + ], + "bbox": [ + 76, + 167, + 468, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 527, + 212, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When classifiers in safety-critical systems such as healthcare or autonomous driving are deployed in the wild [3], it is important to discover potential failure cases before release. Prior work has focused on corruption [31], adversarial robustness [11, 68, 43], and out-of-distribution detection [45, 32, 30]. There is less work on spurious features, although their potential harm might be higher.", + "bbox": [ + 75, + 551, + 468, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spurious features: It has been noted early on that classifiers show reliance on spurious features [15] e.g., a cow on the beach is not recognized [9] due to the missing spurious feature of grass. Other forms of spurious features have been reported in the classification of skin lesions [12], pneumonia [86], traffic signs [67], and object recognition [88]. Moreover, it has been shown that deep neural networks are biased towards texture [26] and background context [80], see [25] for an overview. [58] argues theoretically that spurious features are picked up due to a simplicity bias.", + "bbox": [ + 75, + 659, + 468, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detection of spurious features has been achieved using human label-intense pixel-wise annotations [48, 59, 60]. In [78], they use sparsity regularization to enforce a more interpretable model and use it for finding spurious features. [4] propose a complex pipeline to detect spurious features. While they scale to ImageNet, their analysis is lim", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ited to a few spurious features for a subset of 100 classes. [61, 44, 62] do a search on full ImageNet based on class-weighted \"neural maps\". The neural maps are used to add noise to \"spurious\" resp. \"core\" features but no significant difference in classification performance is observed. It remains unclear if their found spurious features are harmful, that is the feature alone triggers the decision for that class.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interpretability methods: In recent years several interpretability methods have been proposed e.g., attribution methods such as GradCAM [57], Shapley values [42], Relevance Propagation [7], and LIME [51]. The use of these methods for the detection of spurious features has been analyzed in [2, 1] with mixed success and it has been argued that interpretability methods are not robust [19, 64, 27]. However, attribution methods work better for robust classifiers due to more interpretable gradients [22]. Another technique is counterfactual explanations [76, 75] which are difficult to generate for images due to the similarity to adversarial examples [68]. Thus visual counterfactual explanations are realized via manipulation of a latent space [56] or in image space [53, 6, 13] for an adversarily robust classifier. Visual counterfactuals for non-robust classifiers using diffusion models [35, 65, 33, 18] have been proposed in [5]. ImageNet: ImageNet [52] suffers from several shortcomings: apart from an inherent dataset bias [71], semantically overlapping or even identical class pairs were reported [34, 73, 10], e.g., two classes \"maillot\", \"sunglass\" vs \"sun-glasses\", \"notebook\" vs \"laptop\" etc. We disregard such trivial cases of dataset contamination and focus on classes with harmful spurious features, in particular ones where only a small portion of the training set is contaminated.", + "bbox": [ + 496, + 198, + 890, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Spurious features", + "text_level": 1, + "bbox": [ + 500, + 573, + 669, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A proper definition of spurious features is difficult. We describe two settings of harmful spurious features which appear in this paper. We denote by $C_k$ the set of all images containing objects belonging to class $k$ (assuming for simplicity that we have a deterministic problem and ignoring multi-labels). Let $S$ be the set of all images containing a feature $s$ (e.g., a bird feeder). It is a correlated feature for class $k$ when $C_k \\cap S$ and $S \\setminus C_k$ are non-empty, i.e., the feature occurs frequently with the class object but there is no causal implication that appearance of $s$ implies the appearance of the class object (a bird feeder in the image does not imply presence of a hummingbird). A correlated feature becomes a spurious feature when the classifier picks it up as feature of this class. Not every spurious feature is immediately harmful, even humans use context information [25] to get more confident in a decision. However, a spurious feature is harmful if the spurious feature alone is enough to trigger the decision for the corresponding class without the class object being present in the image. We consider two scenarios for a harmful spurious feature shown in Fig. 2.", + "bbox": [ + 496, + 598, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20236", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c3e3e4aa0171ff564873305b533c9a4932d01f6cbdcd10514a58794c49e17e3.jpg", + "image_caption": [ + "Figure 2: Type of harmful Spurious Feature: Left: The spurious feature $s$ is taken up by the classifier which predicts class $k$ on $C_k \\cup S$ instead only on $C_k$ . Right: The spurious feature $s$ is shared between classes but appears more often in class $k$ . The classifier associates $S$ with class $k$ and thus predicts class $k$ also on $C_l \\cap S$ instead of class $l$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 472, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spurious Class Extension: For this type of spurious feature (left in Fig. 2) the classifier picks up the spurious feature $s$ for class $k$ and predicts the class $k$ even on $S \\setminus C_k$ with high confidence (prediction of \"hummingbird\" for images showing a bird feeder but no hummingbird). The classifier predicts class $k$ beyond its actual domain $C_k$ and thus we call this a spurious class extension. While this spurious feature does not necessarily hurt in terms of test performance it can easily lead to completely unexpected behavior when the classifier is deployed in the wild.", + "bbox": [ + 75, + 371, + 468, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spurious Shared Feature: Here, two classes $C_k$ and $C_l$ share a spurious feature $s$ (e.g., \"water jet\" for the classes \"fireboat\" and \"fountain\"). As there are more training images with feature $s$ in $C_k$ than in $C_l$ the classifier associates $S$ with class $k$ and predicts class $k$ for $S \\cap C_l$ .", + "bbox": [ + 76, + 522, + 468, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The two types of harmful spurious features are not exclusive. A shared spurious feature $s$ can at the same time lead to a spurious class extension, e.g., the object bird feeder leads to a spurious class extension of the class \"hummingbird\" (see Fig. 4) to images of bird feeders without hummingbirds. In the training set the hummingbird feeder appears only in images of class \"hummingbird\" but the hummingbird feeder has parts which mimic flowers and flowers are a shared spurious feature with bees. In Fig. 4 right top row, images of bees on a bird feeder are classified as \"hummingbird\" instead of \"bee\", so the spurious feature is strong enough to override the decision for the true class \"bee\" (spurious shared feature).", + "bbox": [ + 75, + 598, + 470, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Finding spurious features via neural PCA and associated feature visualizations", + "text_level": 1, + "bbox": [ + 76, + 825, + 468, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we define our class-wise neural PCA (NPCA) which allows us to find diverse subpopulations in the train-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing data, e.g., we checked that the bird feeder for \"hummingbird\" is visible in $15\\%$ of the training images (component 2), while another $15\\%$ contain a part of it (component 3), see Fig. 3 or, for more examples, App. F. Then we introduce our neural PCA feature visualization (which requires an adversarially robust model) and how we select NPCA components for human inspection. The identification of spurious features requires minimal human supervision, and our effective setup allows us to screen all ImageNet classes.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarily robust model: Similar to [61], we use an adversarily robust model to find spurious features in ImageNet. The reason for this is that robust models have generative properties [74, 53, 6, 61, 13] in the sense that maximizing the predicted probability of a class in a neighborhood of an image leads to semantically meaningful changes. They also have more informative gradients [22] and thus attribution maps such as GradCAM [57] work better. We use the multiple-norm robust model of [13] as they claim it has the best generative properties. The generative properties of robust models are mainly used for the neural PCA feature visualization where we maximize the NPCA component of a class starting from a gray image. If a spurious feature appears without the class object, this is a strong indicator of a harmful spurious feature. A non-robust model would only produce semantically meaningless adversarial noise, hence we need a robust model for this part of our detection pipeline. Our detected spurious features are not specific to the robust model. We show that SOTA ImageNet models share the same spurious features (Fig. 4 and Sec. 7.2).", + "bbox": [ + 496, + 227, + 892, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Class-wise neural PCA: Let $(x_{i},y_{i})_{i = 1}^{N}$ be the training set, where $y_{i}\\in \\{1,\\dots ,K\\}$ and $K$ is the number of classes. We consider features of the penultimate layer $\\phi (x)\\in \\mathbb{R}^D$ of a neural network for an input $x$ . For a given class $k$ and its associated weights $w_{k}\\in \\mathbb{R}^{D}$ in the final layer, we define", + "bbox": [ + 496, + 529, + 890, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {k} (x) = w _ {k} \\odot \\phi (x). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 612, + 890, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ is the componentwise product. Let $b\\in \\mathbb{R}^K$ be the bias vector of the final layer then the logit $f_{k}$ of class $k$ is:", + "bbox": [ + 496, + 633, + 890, + 665 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {k} (x) = \\sum_ {j = 1} ^ {D} w _ {k j} \\phi (x) _ {j} + b _ {k} = \\langle \\mathbf {1}, \\psi_ {k} (x) \\rangle + b _ {k}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 670, + 890, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $I_{k}$ be the index set of the training set of class $k$ and $\\bar{\\psi}_k$ the class-wise mean, $\\bar{\\psi}_k = \\frac{1}{|I_k|}\\sum_{s\\in I_k}\\psi_k(x_s)$ . The class-wise neural PCA allows us to identify variations in the set $\\{\\psi (x_r)\\}_{r\\in I_k}$ arising due to small subpopulations in the training set. In the class-wise neural PCA, we compute eigenvectors of the class-wise covariance matrix,", + "bbox": [ + 496, + 704, + 890, + 795 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {s \\in I _ {k}} \\left(\\psi_ {k} \\left(x _ {s}\\right) - \\bar {\\psi} _ {k}\\right) \\left(\\psi_ {k} \\left(x _ {s}\\right) - \\bar {\\psi} _ {k}\\right) ^ {T}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 801, + 890, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The eigenvectors, $v_{1},\\ldots ,v_{D}$ form an orthonormal basis of $\\mathbb{R}^D$ and we write $\\psi_k(x) - \\bar{\\psi}_k$ in this basis,", + "bbox": [ + 496, + 840, + 890, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {k} (x) - \\bar {\\psi} _ {k} = \\sum_ {l = 1} ^ {D} v _ {l} \\left\\langle \\psi_ {k} (x) - \\bar {\\psi} _ {k}, v _ {l} \\right\\rangle , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 877, + 890, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20237", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/28e7dffbbe321d2281ad33a08966585590676aa800f49524184302807e2e3714.jpg", + "image_caption": [ + "Top 5 neural PCA components (ours)", + "NNE" + ], + "image_footnote": [], + "bbox": [ + 98, + 104, + 467, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e7a194e1a7bc8b42a74bddc476398b68002e531e40dc608a4776ecce579427c.jpg", + "image_caption": [ + "Max. act." + ], + "image_footnote": [], + "bbox": [ + 98, + 164, + 467, + 223 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d7988fe9f7416a4b9e2570316c78aa88f91ffed005e878461d29d50ad061318.jpg", + "image_caption": [ + "Heeannnnnne", + "Figure 3: Top 5 Neural PCA components for class hummingbird: first row shows our neural PCA feature visualization (NPFV), second row shows four most activating training images of each NPCA component and the last row GradCAM for the NPCA component. Our NPCA components capture different subpopulations in the training data. Comp. 2 is identified as spurious feature \"bird feeder\", see NPFV and most activating training images (see also Fig. 4)." + ], + "image_footnote": [], + "bbox": [ + 98, + 224, + 467, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and define", + "bbox": [ + 76, + 436, + 150, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {l} ^ {(k)} (x) = \\left\\langle \\mathbf {1}, v _ {l} \\right\\rangle \\left\\langle \\psi_ {k} (x) - \\bar {\\psi} _ {k}, v _ {l} \\right\\rangle , \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 460, + 468, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The logit $f_{k}(x)$ of the $k$ -th class can then be written as", + "bbox": [ + 76, + 491, + 437, + 507 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {k} (x) = \\sum_ {l = 1} ^ {D} \\alpha_ {l} ^ {(k)} (x) + \\left\\langle \\mathbf {1}, \\bar {\\psi} _ {k} \\right\\rangle + b _ {k}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 518, + 468, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus for a given $x$ , we can interpret $\\alpha_l^{(k)}(x)$ as the contribution of the neural PCA component $l$ of class $k$ to the logit $f_k(x)$ of class $k$ since the term $\\langle \\mathbf{1}, \\bar{\\psi}_k \\rangle + b_k$ is constant for all inputs. Based on this we introduce a mitigation technique for spurious features without retraining in Sec. 5.", + "bbox": [ + 75, + 558, + 468, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural PCA Feature Visualization: To identify semantic features corresponding to our neural PCA component $l$ , we show the training images which attain the maximal values of $\\alpha_{l}^{(k)}(x)$ . Additionally, we generate an image $z_{l}^{(k)}$ , which we call the Neural PCA Feature Visualization (NPFV) of feature $l$ of class $k$ , by maximizing $\\alpha_{l}^{(k)}(x)$ :", + "bbox": [ + 76, + 650, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {l} ^ {(k)} = g + \\operatorname * {a r g m a x} _ {\\| \\delta \\| _ {2} \\leq \\epsilon} \\alpha_ {l} ^ {(k)} (g + \\delta),\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 761, + 406, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g$ is a gray image (all channels equal to 0.5). Thus we maximize the feature $\\alpha_{l}^{(k)}$ outgoing from a non-informative and unbiased initialization $g$ . The optimization problem is solved using adaptive projected gradient descent (APGD) [16] with 200 steps. The budget for changes, $\\epsilon = 30$ , is small to avoid the overactivation of feature attacks maximizing the output of individual neurons [21, 63, 61], see", + "bbox": [ + 75, + 792, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 8. In Fig. 4 we show for each identified spurious feature, the corresponding NPFV, e.g., for \"hummingbird\" one can see the bird feeder but no hummingbird. The NPCA components together with the maximally activating training images are in principle sufficient to identify spurious features, but the NPFV is very useful to judge how harmful a spurious feature is. Therefore, we use an adversarily robust model, since a non-robust model would yield semantically meaningless adversarial noise as NPFV.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Selection of neural PCA components for human inspection: The penultimate layer of the robust ResNet50 we are using has 2048 neurons. Thus it is infeasible (and unnecessary) to investigate all neural PCA components. A strong criterion that one has found a harmful spurious feature is i) the NPFV shows mainly the spurious feature and not the class, and ii) the NPFV has high confidence. If ii) is not satisfied, then the NPFV is a spurious feature the classifier may have picked up, but it is not harmful in the sense that this feature alone causes the classifier to choose that class. Moreover, we noticed that the eigenvalues of the neural PCA, and the corresponding $\\alpha$ values, decay quickly. Thus we compute the NPFV for the top 128 neural PCA components (having maximal variance) and then select the ten components which realize the highest confidence for their NPFV in the corresponding class. Note that we do not optimize the confidence when generating the NPFV but only $\\alpha_{l}^{(k)}(x)$ which is part of the logit of the $k$ -th class.", + "bbox": [ + 496, + 227, + 892, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Identification of spurious neural PCA components via human supervision: For each ImageNet class $k$ we show the human labeler the top 10 components. For each component $l$ we show the NPFV $z_{l}^{(k)}$ and the 5 training images $x_{r}$ of class $k$ with the largest values of $\\alpha_{l}^{(k)}(x_{r})$ . Moreover, we compute GradCAM [57] images for the NPFV and the five training images using the NPCA component $\\alpha_{l}^{(k)}$ as score. The human marks a component as spurious if i) the NPFV shows dominantly an object not belonging to the class ii) the five training images show consistently this object, iii) the GradCAM activations are primarily not on the class object. The setup shown to the human labeler can be seen in App. B. The labeling of one class takes on average about 45 seconds, so the full labeling of all ImageNet classes took about 13 hours. The human labeler (researcher in machine learning) found in total 337 spurious components. Another human labeler checked all of them and removed spurious features in case of disagreement, resulting in 322 spurious features from 230 ImageNet classes.", + "bbox": [ + 496, + 503, + 892, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. SpuFix - Mitigation of spurious features", + "text_level": 1, + "bbox": [ + 498, + 813, + 857, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once the spurious features are identified, the question is how one can mitigate that the classifier relies on them. One way is to identify the training images containing the spurious feature and then discard or downweight them during", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20238", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b14bf69b497362b8af6a3e938c5160b68291741a643fcbb5b3878d8411f4b9c6.jpg", + "image_caption": [ + "Hummingbird - Random train. images (confidence / $\\alpha_{k}$ )", + "0.93/1.7" + ], + "image_footnote": [], + "bbox": [ + 80, + 104, + 158, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bc0b28bfc916a5957df536af5be3e4e59f8da09d64090eb1ce52be150f1eefc1.jpg", + "image_caption": [ + "1.00/-0.9 0.96/-1.0" + ], + "image_footnote": [], + "bbox": [ + 160, + 104, + 238, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/42061f85986488c370fe1403148bc244f77aae006c86035a5278b75e93aec6f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 104, + 320, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/564e85099f5fcb75517ed03ec10f9e2ac1281b1b7e7912168419afa7c82aa0cc.jpg", + "image_caption": [ + "Max. activating train. images - NPCA Comp. 2", + "1.00/5.9" + ], + "image_footnote": [], + "bbox": [ + 240, + 198, + 320, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0bf1c17e1415b8b5a086a0934782e84692922b518147e332cea369bafebf3ab8.jpg", + "image_caption": [ + "0.99/2.2 1.00/1.54" + ], + "image_footnote": [], + "bbox": [ + 320, + 104, + 400, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 320, + 199, + 400, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/73649e84b888d8a170558282f72b9119f8421478e23746fb0ce9ef2c06dbe182.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 104, + 480, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1.00/1.54", + "bbox": [ + 406, + 167, + 477, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A Comp. 2", + "bbox": [ + 403, + 183, + 477, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee9abd0b3cf89ae4ab7540cd4d1cd641be18143c2c94b6ad93311dc845855134.jpg", + "image_caption": [ + "NPFV-2", + "1.00/9.7" + ], + "image_footnote": [], + "bbox": [ + 80, + 199, + 158, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/55e73764a49276f04235ea41a34d73c5ff3ea4a2234cfa7f10a29b5f3038a146.jpg", + "image_caption": [ + "1.00/7.5" + ], + "image_footnote": [], + "bbox": [ + 160, + 199, + 238, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 240, + 199, + 320, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1.00/5.9", + "bbox": [ + 248, + 262, + 310, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 320, + 199, + 400, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1.00/5.6", + "bbox": [ + 328, + 263, + 393, + 275 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3d2ff902933bd3e2b9f8becf427f2361c4641ab084f16473623a2518546840ba.jpg", + "image_caption": [ + "1.00/5.6", + "1.00/5.6" + ], + "image_footnote": [], + "bbox": [ + 400, + 199, + 480, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4e979465e870cbdf4b8c413f94e50d01f2cc1144bf0a83b1fa561a53ba386bd3.jpg", + "image_caption": [ + "Freight car - Random train. images (confidence / $\\alpha_{k}$ )", + "1.00/7.7" + ], + "image_footnote": [], + "bbox": [ + 80, + 292, + 158, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9abcadc71d9090855e26081cec5d5e52e7d20d3a6f1164ae18f32d15e63b0936.jpg", + "image_caption": [ + "1.00/3.4" + ], + "image_footnote": [], + "bbox": [ + 160, + 292, + 238, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bfab20a774e62a2065a524bccb6bbb1005463907e4162fdbf08d4e3b7c979fdf.jpg", + "image_caption": [ + "1.00/6.3" + ], + "image_footnote": [], + "bbox": [ + 240, + 292, + 320, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c9d9314d3941c262721f2249d8ba298e2cd439a066dab189d9e00c9f2f64d789.jpg", + "image_caption": [ + "0.98/-0.8" + ], + "image_footnote": [], + "bbox": [ + 320, + 292, + 400, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d39e205b71a0709e2c941d21fbb09aaaea7fe6bfacd5ee5685853c2f541637cf.jpg", + "image_caption": [ + "1.00/4.2" + ], + "image_footnote": [], + "bbox": [ + 400, + 292, + 480, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b37ebabe2a3afb713344d4567307ea6abea8d8e96a9f3b6e78002694ff20e621.jpg", + "image_caption": [ + "NPFV-1", + "1.00/12.1" + ], + "image_footnote": [], + "bbox": [ + 80, + 387, + 158, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fcf21de7695b8da5ddc4d84e03a64bb687d7a503f9ce1ca22201048369cd0429.jpg", + "image_caption": [ + "Max. activating train. images - NPCA Comp. 1", + "1.00/10.9" + ], + "image_footnote": [], + "bbox": [ + 160, + 387, + 240, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/23f8319e0d23f30392514d7c469a3d05e68f2dec9dadd3f632a11759cf32069e.jpg", + "image_caption": [ + "1.00/10.4" + ], + "image_footnote": [], + "bbox": [ + 240, + 387, + 320, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/106f86f1b02a225d4b39bb2158a317c5f3d170608b0545500c50bae7bbb27d5a.jpg", + "image_caption": [ + "1.00/10.2" + ], + "image_footnote": [], + "bbox": [ + 320, + 387, + 400, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b9b526175825f8302e0707717326217793a880498bb8ae8553e22713ebedd99d.jpg", + "image_caption": [ + "1.00/10.2" + ], + "image_footnote": [], + "bbox": [ + 400, + 387, + 480, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2542e369ce891f1da3a82d5484e8df2745e02c6b49db9f61e0f2089dd515db28.jpg", + "image_caption": [ + "Koala - Random train. images (confidence / $\\alpha_{k}$ )", + "1.00/0.77" + ], + "image_footnote": [], + "bbox": [ + 80, + 481, + 158, + 542 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/099c91b963703a915f76eda313baf1f8e8a46cdf32e90a5e52ef889e16630a38.jpg", + "image_caption": [ + "0.87/2.4" + ], + "image_footnote": [], + "bbox": [ + 160, + 481, + 240, + 542 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/43de58373554e8da48d3e2bc2058799b89fb568a5016021084561b503b60e9e7.jpg", + "image_caption": [ + "1.00/0.4" + ], + "image_footnote": [], + "bbox": [ + 240, + 481, + 320, + 542 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7b2a0498eabd530d3c6f7b82940e61a1a2509f790afa1a19cb7a5a0996bcd07b.jpg", + "image_caption": [ + "1.00/0.0" + ], + "image_footnote": [], + "bbox": [ + 320, + 481, + 400, + 542 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/38c8d0820678ff378e6b0cd711b93ea5ac9053e887f130c0f3bbd8bcadfeeb3e.jpg", + "image_caption": [ + "0.95/0.5" + ], + "image_footnote": [], + "bbox": [ + 400, + 481, + 480, + 542 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7e0f88858e23beed70e24a8579ae8f549b6b8e5498604d4e5c3464bc1a3ccccd.jpg", + "image_caption": [ + "NPFV-3", + "1.00/5.5" + ], + "image_footnote": [], + "bbox": [ + 80, + 575, + 158, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/85a9cf365a5d70f37b0937c99fcfd657361c01785806145e0dd65ecb21021fc8.jpg", + "image_caption": [ + "Max. activating train. images - NPCA Comp. 3", + "1.00/4.6" + ], + "image_footnote": [], + "bbox": [ + 160, + 575, + 240, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aabeef1685499126ba04e6f2a069bc3f04d0fd96b00998392151e6971559cc56.jpg", + "image_caption": [ + "1.00/4.5" + ], + "image_footnote": [], + "bbox": [ + 240, + 575, + 320, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f37200399a4dd0a1a1bee69188f721d122733fa978bd68c0f3a9f389b0ee62f0.jpg", + "image_caption": [ + "1.00/4.4" + ], + "image_footnote": [], + "bbox": [ + 320, + 575, + 400, + 637 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f587b437982882523d294ee93ba63294ec6cc9a6a21d34c28513b19fc46115a1.jpg", + "image_caption": [ + "0.86/4.3" + ], + "image_footnote": [], + "bbox": [ + 400, + 575, + 480, + 637 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/28d10134535d18af7dcf2d59ba7e5d842b09057d7df5bfec65eb4113412512d1.jpg", + "image_caption": [ + "Fireboat - Random train. images (confidence / $\\alpha_{k}$ )", + "0.88/-1.1" + ], + "image_footnote": [], + "bbox": [ + 80, + 670, + 158, + 731 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f2ca3287a934247402b3047e7345d2bccbe2defbbc945ea230deb08ee3c3b73c.jpg", + "image_caption": [ + "0.95/0.5" + ], + "image_footnote": [], + "bbox": [ + 160, + 670, + 238, + 731 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/10f679df621f8554213f997271b63ad8b6fcf08b6b7632967059e16b4f1d6cbd.jpg", + "image_caption": [ + "0.12/-0.6" + ], + "image_footnote": [], + "bbox": [ + 240, + 670, + 320, + 731 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8973b07c4a97153d5e9146d29f562ef8abf618d7845386b5b77d31e6f66de1b7.jpg", + "image_caption": [ + "0.84/-0.2" + ], + "image_footnote": [], + "bbox": [ + 320, + 670, + 400, + 731 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/36f8cc08837596bca683088b426cd5f0b7b2e11f5e95ee1a3b301586c2d03acf.jpg", + "image_caption": [ + "0.02/-1.1" + ], + "image_footnote": [], + "bbox": [ + 400, + 670, + 480, + 731 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fa054c68d416f25b599c682b3410ff02d5cf5b94ff1a664d6c40d47704dad8e1.jpg", + "image_caption": [ + "NPFV-2", + "1.00/5.5", + "Figure 4: Spurious features (ImageNet): found by human labeling of our neural PCA components. For each class we show 5 random train. images (top left), the neural PCA Feature Visual. (NPFV) and 4 most activating train. images for the spurious feature component (bottom left). Right: four ImageNet models classify images showing only the spurious feature but no class object as this class." + ], + "image_footnote": [], + "bbox": [ + 80, + 763, + 158, + 824 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91890ab85f24c0fe4ff3e4d87cf6586c12f238c95bf345d5210ec281cde8c3e4.jpg", + "image_caption": [ + "Max. activating train. images - NPCA Comp. 2", + "0.42/4.1" + ], + "image_footnote": [], + "bbox": [ + 160, + 763, + 240, + 824 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2d85f28e5420f90157fa94845c66fd9fb4d4e510ff66fc405576d654d0a3909a.jpg", + "image_caption": [ + "1.00/4.0" + ], + "image_footnote": [], + "bbox": [ + 240, + 763, + 320, + 824 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/93e8be9bbfbdfcbf20a9e73ee64c590f1c0d0f77b4df71860dac4da49d02ac85.jpg", + "image_caption": [ + "1.00/3.9" + ], + "image_footnote": [], + "bbox": [ + 320, + 763, + 400, + 824 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a33b3698a9b9bf9ec8d6be3f4ba84af1597e4b1320e4ade818bb8ea9016f2e2a.jpg", + "image_caption": [ + "1.00/3.9" + ], + "image_footnote": [], + "bbox": [ + 400, + 763, + 480, + 824 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a45101c80cd815b58da987d5c24eb8cc7467832f0b79b8ff6280ecc1cc92d5bf.jpg", + "image_caption": [ + "Images with spurious bird feeder but no hummingbird", + "0.94/5.7" + ], + "image_footnote": [], + "bbox": [ + 501, + 104, + 580, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a80f1d75e79d106d0c8c264cc6927f495f0e31e4d252bedaf73e10468a47772.jpg", + "image_caption": [ + "0.94/3.4" + ], + "image_footnote": [], + "bbox": [ + 581, + 104, + 660, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2125c00b429315f1ae61534ff1e8810e53af39f0e1e0403aa0d738e5382e8de1.jpg", + "image_caption": [ + "0.82/2.9" + ], + "image_footnote": [], + "bbox": [ + 661, + 104, + 738, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3e511056306d77c041e2d2bedd8de69bccd3cc66a080d4053953d0aa0f6033a3.jpg", + "image_caption": [ + "0.91/5.6" + ], + "image_footnote": [], + "bbox": [ + 740, + 104, + 820, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/06196a4a233b5176ab789b24e32008598fac17ae3656c05b544b2b2a8773a617.jpg", + "image_caption": [ + "0.91/4.7" + ], + "image_footnote": [], + "bbox": [ + 821, + 104, + 900, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/edc3193bf8177af534bf2d4a751fcb1c5fad50d517725c2ffcac95807adc870a.jpg", + "image_caption": [ + "all classified as humming bird by four ImageNet models", + "0.86/4.3" + ], + "image_footnote": [], + "bbox": [ + 501, + 198, + 580, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee7284a760a207d058f8128dc592dbb8be7117ac1c127953c23efa56ba4f39e1.jpg", + "image_caption": [ + "0.81/3.5" + ], + "image_footnote": [], + "bbox": [ + 581, + 198, + 660, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/acf64746e60be599c2c47372211bc4e0dd74fa501a30b96ec72507e8aee97826.jpg", + "image_caption": [ + "0.89/3.3" + ], + "image_footnote": [], + "bbox": [ + 661, + 198, + 738, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2abd1f654cc64097e2ebf3229076a5525b3f04014995084534c3cc5adbddc412.jpg", + "image_caption": [ + "0.89/3.8" + ], + "image_footnote": [], + "bbox": [ + 740, + 198, + 820, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3607ca6620d48cf89461c71bdc2b4d885649ec1a1ad2ae54d78c1dc9ad27e744.jpg", + "image_caption": [ + "0.78/5.91" + ], + "image_footnote": [], + "bbox": [ + 821, + 198, + 900, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/262cfde3f1b1b79b83e2e26f16d2506c66fc6a90aeb75608aa7d14554bda8741.jpg", + "image_caption": [ + "Images with spurious grafitti but no freight car", + "0.97/3.5" + ], + "image_footnote": [], + "bbox": [ + 501, + 292, + 580, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/60a9a85f1c8ca0e02b8d4edf4ead405387013f9f69b5020ccbfee9a1b1557f52.jpg", + "image_caption": [ + "0.79/4.0" + ], + "image_footnote": [], + "bbox": [ + 581, + 292, + 660, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/76e56ea2c33cc00427a7e546d8363a91a75072aa98798556b58c29822ae72383.jpg", + "image_caption": [ + "0.81/2.4" + ], + "image_footnote": [], + "bbox": [ + 661, + 292, + 738, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/57dee91e1490077c8b1dd710f3f532ea5811d60104ceef2245d5236238366c74.jpg", + "image_caption": [ + "0.85/3.2" + ], + "image_footnote": [], + "bbox": [ + 740, + 292, + 820, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b43ee7265dc02cca8f65dd9204d7e5b6e07e295b14d14fedd59c59b1ad683b99.jpg", + "image_caption": [ + "0.88/2.6" + ], + "image_footnote": [], + "bbox": [ + 821, + 292, + 900, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4e7140d05f47c85caef4c21a641ee3cdd70ed87a0f5ee698c50314669f9a5ed5.jpg", + "image_caption": [ + "all classified as freight car by four ImageNet models", + "0.85/2.5" + ], + "image_footnote": [], + "bbox": [ + 501, + 387, + 580, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c765abbfa0845be76cfba9bc4314b4a671a86dd65c0eee290423e949c101756.jpg", + "image_caption": [ + "0.86/2.5" + ], + "image_footnote": [], + "bbox": [ + 581, + 387, + 660, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/271004bd3fd2d0688020e1fae53de0d933064f13017c960bb7ad9840f3b8a3f2.jpg", + "image_caption": [ + "0.90/2.3" + ], + "image_footnote": [], + "bbox": [ + 661, + 387, + 738, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/63240b2ea26dd0e19553f2428c0b456d15ab83512661cc740427511cc9713ebb.jpg", + "image_caption": [ + "0.82/2.2" + ], + "image_footnote": [], + "bbox": [ + 740, + 387, + 820, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/682e474921ec27b310ef72061a462208f2a81668ae224208a102796055f4bccd.jpg", + "image_caption": [ + "0.87/2.2" + ], + "image_footnote": [], + "bbox": [ + 821, + 387, + 900, + 446 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7cc871558c2f74dee577490a040eb115083fdc3a1eb2923c0b95f3275e771343.jpg", + "image_caption": [ + "Images with spurious eucalyptus/plants but no koala", + "0.49/3.5" + ], + "image_footnote": [], + "bbox": [ + 501, + 481, + 580, + 542 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d9ed6648e6e7ec103f36ec6ef269baa9d8fb476d2c799d461b1622c9f6d0389a.jpg", + "image_caption": [ + "0.61/3.1" + ], + "image_footnote": [], + "bbox": [ + 581, + 481, + 660, + 542 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13fc59d130ba54cbeb33bb7879ca0346e9c93fd71b30d839afdd5f713dfd35d6.jpg", + "image_caption": [ + "0.36/3.1" + ], + "image_footnote": [], + "bbox": [ + 661, + 481, + 738, + 542 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2a962ccc0109ae8bdf64c42eeb97112c411021f75caf04cf924183cc02ca1e9f.jpg", + "image_caption": [ + "0.36/3.1" + ], + "image_footnote": [], + "bbox": [ + 740, + 481, + 820, + 542 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fb2d9fedf1c00962f3a8ec7fab10f710347d785b7826f1e9f1dab34aa3102473.jpg", + "image_caption": [ + "0.69/2.7" + ], + "image_footnote": [], + "bbox": [ + 821, + 481, + 900, + 542 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ec0ce2d1a664953e5608d6b6070ad07de40d8ae64d8af6c90f0e37df146a5a15.jpg", + "image_caption": [ + "all classified as koala by four ImageNet models", + "0.56/2.6" + ], + "image_footnote": [], + "bbox": [ + 501, + 575, + 580, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fef7e6a9fe2d10335af4e7728a604c8d30f401e872a0b1d6129d41fe1088b2a7.jpg", + "image_caption": [ + "0.62/2.5" + ], + "image_footnote": [], + "bbox": [ + 581, + 575, + 660, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c6a74f224f78c5b62042650c42fe903570e5a1b792771e345f869fa79c355cf9.jpg", + "image_caption": [ + "0.74/2.5" + ], + "image_footnote": [], + "bbox": [ + 661, + 575, + 738, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e1ac2a48b5b267a5f9f1093a7a3007ddfba246d63efbbb94e4697652f0f54be7.jpg", + "image_caption": [ + "0.72/2.4" + ], + "image_footnote": [], + "bbox": [ + 740, + 575, + 820, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a0daf6d3b2d23f26bb5ad7a69324b618fb81cc9c899f2cf52ae2139f715669f8.jpg", + "image_caption": [ + "0.66/2.2" + ], + "image_footnote": [], + "bbox": [ + 821, + 575, + 900, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/539f197c9df300a1aa9e765daa8d1af929468c7716c3c3b51dea548ae04dddc1.jpg", + "image_caption": [ + "Images with spurious water jet but no fireboat", + "0.63/2.3" + ], + "image_footnote": [], + "bbox": [ + 501, + 670, + 580, + 731 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/28b216cd46e16f027887cf4dffcf9a136cb940f482ebefb3db4519920803a214.jpg", + "image_caption": [ + "0.71/2.2" + ], + "image_footnote": [], + "bbox": [ + 581, + 670, + 660, + 731 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/66c383237bb2b0f4faa016135376e8c11ad7ff76c853f68d1c53d07adf2b222e.jpg", + "image_caption": [ + "0.84/2.2" + ], + "image_footnote": [], + "bbox": [ + 661, + 670, + 738, + 731 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2e74305ae404761c3c77dfd49a05a9731b27bfd41aeac5fd6fd6a2983548f4b6.jpg", + "image_caption": [ + "0.65/2.1" + ], + "image_footnote": [], + "bbox": [ + 740, + 670, + 820, + 731 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1d50568e613b91952708f5aba8784e3606eeb8792a7aa51f94dd57b17ffaf90a.jpg", + "image_caption": [ + "0.63/1.9" + ], + "image_footnote": [], + "bbox": [ + 821, + 670, + 900, + 731 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b0e378ac19e854bde16c8c16b81016e8bd6d2062686b3156b52ea4c629b31439.jpg", + "image_caption": [ + "all classified as fireboat by four ImageNet models", + "0.53/1.9" + ], + "image_footnote": [], + "bbox": [ + 501, + 763, + 580, + 824 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/494d868e345a740ca98e1534f8c2d1a4f19efe7c7b1ec1748cfda15a9b3ad39e.jpg", + "image_caption": [ + "0.79/1.9" + ], + "image_footnote": [], + "bbox": [ + 581, + 763, + 660, + 824 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/71073f348182fd929fb4a5ab6345d59d1d824aef912ad5ff5b6628923f879d12.jpg", + "image_caption": [ + "0.78/1.8" + ], + "image_footnote": [], + "bbox": [ + 661, + 763, + 738, + 824 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3fa638d12c61fb5772f77e9553ee27edc4abbc4a025a4dc50b658973ecc02481.jpg", + "image_caption": [ + "0.79/1.8" + ], + "image_footnote": [], + "bbox": [ + 740, + 763, + 820, + 824 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b9ca131bbf96e9cf4b8c4e9a39bcbd814f4f36181e146dbf343b815b8d441331.jpg", + "image_caption": [ + "0.76/1.8" + ], + "image_footnote": [], + "bbox": [ + 821, + 763, + 900, + 824 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20239", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "training. However, this would require relabeling all spurious ImageNet classes which is not feasible. We could order the training set according to the value $\\alpha_{l}^{(k)}$ of the corresponding neural PCA component which indicates how much of the spurious feature an image contains. While this would speed up the process significantly, it would still require a significant amount of manual relabeling. Can one do it also without any additional labeling? Yes, as described in Sec. 4 we can rewrite the logit of the $k$ -th class as", + "bbox": [ + 75, + 90, + 472, + 229 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {k} (x) = \\sum_ {l = 1} ^ {D} \\alpha_ {l} ^ {(k)} (x) + \\left\\langle \\mathbf {1}, \\bar {\\psi} _ {k} \\right\\rangle + b _ {k}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 237, + 468, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a spurious component $l$ of class $k$ , we use $\\min \\{\\alpha_l^{(k)}, 0\\}$ instead of $\\alpha_l^{(k)}$ to remove its positive contribution from the logit (negative contributions are semantically different). After removal of the spurious features, the new logit becomes", + "bbox": [ + 76, + 273, + 468, + 339 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {k} ^ {S p u F i x} (x) = f _ {k} (x) - \\sum_ {l \\in \\mathcal {S} _ {k}} \\max \\left\\{\\alpha_ {l} ^ {(k)} (x), 0 \\right\\} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 347, + 468, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S_{k}$ is the set of spurious NPCA components of class $k$ . We denote this method as SpuFix. It significantly reduces dependence on spurious features, see Sec. 7.3.", + "bbox": [ + 76, + 387, + 468, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transfer of SpuFix to any ImageNet classifier: As described in Sec. 3, harmful spurious features are a result of subpopulations in the training data. While not every spurious correlation will be picked up by every model, most of our detected spurious features generalize to a wide range of classifiers (see Sec. 7). In the following, we show how SpuFix can be transferred to any given ImageNet classifier $\\tilde{f}$ for which $\\tilde{f}_k$ denotes the logit and $\\tilde{\\psi}_k$ the weighted penultimate layer of class $k$ . The goal is to find a direction $b$ in the weighted feature space $\\tilde{\\psi}_k$ of $\\tilde{f}$ for every spurious NPCA component $l$ corresponding to the eigenvector $v_l$ of the original model, resp. $\\alpha_l^{(k)}(x)$ , and then truncate its positive component. To find this direction we maximize the covariance of the projection onto $b$ and $\\alpha_l^{(k)}$ over the training images of class $k$ :", + "bbox": [ + 75, + 450, + 472, + 681 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nb _ {l} ^ {(k)} = \\underset {\\| b \\| _ {2} = 1} {\\operatorname {a r g m a x}} \\sum_ {s \\in I _ {k}} \\left\\langle b, \\tilde {\\psi} _ {k} \\left(x _ {s}\\right) - \\bar {\\tilde {\\psi}} _ {k} \\right\\rangle \\alpha_ {l} ^ {(k)} \\left(x _ {s}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 689, + 468, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which has a closed form solution", + "bbox": [ + 76, + 732, + 297, + 746 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nb _ {l} ^ {(k)} = \\frac {\\sum_ {s \\in I _ {k}} \\left(\\tilde {\\psi} _ {k} \\left(x _ {s}\\right) - \\bar {\\tilde {\\psi}} _ {k}\\right) \\alpha_ {l} ^ {(k)} \\left(x _ {s}\\right)}{\\left\\| \\sum_ {s \\in I _ {k}} \\left(\\tilde {\\psi} _ {k} \\left(x _ {s}\\right) - \\bar {\\tilde {\\psi}} _ {k}\\right) \\alpha_ {l} ^ {(k)} \\left(x _ {s}\\right) \\right\\| _ {2}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 752, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast to the eigenvectors $v_{l}$ , the matched vectors $b_{l}^{(k)}$ are not necessarily orthogonal. Thus before truncation the centered features $\\tilde{\\psi}_k(x) - \\bar{\\tilde{\\psi}}_k$ need to be projected onto the subspace spanned by the $b_{l}^{(k)}$ and represented in the non-orthogonal basis $\\{b_{l}^{(k)}\\}_{l\\in S_{k}}$ . We denote this representation", + "bbox": [ + 75, + 813, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by $P^{(k)}(x)$ (details in C.1). The logit of class $k$ of the SpuFix version of $\\tilde{f}$ is then:", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} _ {k} ^ {S p u F i x} (x) = \\tilde {f} _ {k} (x) - \\sum_ {l \\in \\mathcal {S} _ {k}} \\max \\left\\{\\langle \\mathbf {1}, b _ {l} ^ {(k)} \\rangle P _ {l} ^ {(k)} (x), 0 \\right\\}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 131, + 890, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the case where $\\tilde{f} = f$ (the robust model), we recover the original SpuFix truncation in (8) (see C.2). It turns out that SpuFix is even effective when the architecture is quite different from the ResNet50 we used for detection, e.g. ViT or VOLO, see Sec. 7.3 and Table 1 and 2.", + "bbox": [ + 496, + 178, + 893, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Comparison to neural features of [61]", + "text_level": 1, + "bbox": [ + 498, + 268, + 839, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our NPCA framework to the method of [61] to detect spurious features for ImageNet. As model, they use a $\\ell_2$ -robust ResNet50. Let $J_{k}$ be the set of training images classified as class $k$ . [61] define the $j$ -th component $m_j^{(k)}$ of the class-wise mean over predictions,", + "bbox": [ + 496, + 292, + 890, + 375 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nm ^ {(k)} = \\frac {1}{\\left| J _ {k} \\right|} \\sum_ {x _ {s} \\in J _ {k}} \\psi_ {k} \\left(x _ {s}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 385, + 890, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as the importance of the $j$ -th neuron for class $k$ .1 Then, they order the neurons of the penultimate layer according to the score $m_j^{(k)}$ and consider the top-5 neurons of each class. The main difference to our approach is that they assume single neurons with maximal influence on the mean are capturing spurious features whereas our NPCA components are linear combinations of neurons that capture the variance around the mean. Given that the ResNet50 has only 2048 neurons for 1000 classes, some neurons are labeled as core and spurious feature for multiple classes simultaneously, even though the images are quite different. A major advantage of NPCA is that due to the orthogonality of the PCA components, we identify diverse subpopulations in the training data. As [61] use no constraints for the neurons, they often find very similar subpopulations. Hence, one may miss spurious subpopulations when only checking the top-5 components, see Fig. 5. Another difference is that they maximize the score $m_j^{(k)}$ for the training images $x_r$", + "bbox": [ + 496, + 429, + 892, + 708 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {j} ^ {(k)} = x _ {r} + \\mathrm {a r g m a x} _ {\\| \\delta \\| _ {2} \\leq \\epsilon} m _ {j} ^ {(k)} (x _ {r} + \\delta),\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 718, + 839, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "whereas we maximize our NPCA component $\\alpha_l^{(k)}(x)$ starting from a gray image to introduce no bias. Thus, we check if the classifier produces the spurious feature and not only enhances it on an image showing it already.", + "bbox": [ + 496, + 753, + 890, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The third difference is that [61] want to identify any spurious feature while our goal is to find harmful ones. Thus, they use weaker criteria for deciding if a neuron shows a spurious feature: main criterion is if the neural activation", + "bbox": [ + 496, + 816, + 890, + 877 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1The top-5 neurons do not change when using $I_{k}$ instead of $J_{k}$", + "bbox": [ + 517, + 886, + 854, + 900 + ], + "page_idx": 5 + }, + { + "type": "footer", + "text": "20240", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/07923a3e5e2377b02bbc92739db36494da0d9c91fd2f8284e531fe75ea1bd88d.jpg", + "image_caption": [ + "Top 5 neural PCA components (ours)", + "Figure 5: Comparison of top-5 NPCA and top-5 neurons for class hummingbird for the robust model of [61]: Our NPCA components identify diverse subpopulations in the training set whereas the top neurons show similar ones and the spurious bird feeder is not detected, see also App. A." + ], + "image_footnote": [], + "bbox": [ + 76, + 104, + 478, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "map based on $m_{j}^{(k)}$ is off the class object according to a majority vote of 5 human labelers. The visualizations $w_{j}^{(k)}$ are only used if the activation maps are inconclusive. In contrast, we require i) our NFPV to show the spurious feature, ii) the GradCAM based on $\\alpha_{l}^{(k)}$ highlights mainly the spurious feature, and iii) our human labelers have to agree that the NPCA component is a harmful spurious feature.", + "bbox": [ + 75, + 460, + 468, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As our criteria are more strict (in particular, that the NFPV shows the spurious feature is a strong criterion), it is not surprising that [61] find more spurious features (630 in 357 classes) than we do (322 in 230 classes). Moreover, the employed models are different and we examine top-10 NPCA components whereas they check top-5 neurons. Thus, the comparison is difficult and our novel dataset \"SpuriousImageNet\" could be biased towards spurious features which only we found. Hence, we compute our top-5 NPCA components for their robust ResNet50 for a direct comparison to their top-5 neurons and their found spurious features. In Fig. 5 we compare them for the class hummingbird where they do not find the spurious feature \"bird feeder\". In general, we observe that our found subpopulations are more diverse and thus we find more spurious features than they do when using their weaker criteria. In App. E we do an extensive comparison for all classes.", + "bbox": [ + 75, + 574, + 470, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Experiments", + "text_level": 1, + "bbox": [ + 76, + 844, + 209, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we provide a qualitative and quantitative evaluation of our 322 detected spurious features in Ima", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "geNet, see Sec. 4. For the quantitative evaluation, we create the dataset \"Spurious ImageNet\", which allows checking the reliance of a given ImageNet classifier on spurious features. We also evaluate our mitigation strategy \"SpuFix\" which does not require additional labels or retraining of the classifier and can be transferred to other image classifiers.", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1. Qualitative evaluation", + "text_level": 1, + "bbox": [ + 500, + 194, + 705, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the qualitative evaluation, we visualize some of our found 322 spurious features, see Fig. 4. For each class, we show five random training images, the NPFV, and the four most activating training images of the neural PCA component labeled to be spurious. Additionally, we always show ten images which only show the spurious feature but not the actual class e.g., only the bird feeder (spurious) but no hummingbird (class). All ten images are classified as the corresponding class for the robust classifier we have used to compute the NPCA components and three non-robust ImageNet classifiers (ResNext101, Eff.Net B5, ConvNext-B, see also Tab. 2). This shows that our spurious features generalize from the robust classifier to SOTA ImageNet classifiers, indicating that the found spurious features are mainly due to the design of the training set, rather than failures in model training. Our novel validation by collecting real images with the spurious feature but without the class object which are consistently classified as this class directly shows the impact of harmful spurious features and has the advantage that it does not introduce artifacts via masking nor requires expensive pixel-wise segmentations.", + "bbox": [ + 496, + 218, + 892, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image Collection: The images showing only the spurious feature were obtained by sorting the 9 million images of OpenImages [38] by the value $\\alpha_l^{(k)}(x)$ of the neural PCA component. We check the top 625 retrieved images classified by the robust classifier as the corresponding class if they are all classified as the same class by the additional non-robust classifiers and do not show the corresponding class. This is a quite strict criterion as spurious features can be shared across classes, e.g., twigs for birds, and thus agreement of classifiers is not granted and the images can show the spurious feature and the true class. Nevertheless, this procedure yields between 77 (\"hummingbird\") and 179 (\"freight car\") images of which we show a selection. For \"hummingbird\", a lot of these images show red flowers (without a hummingbird) which makes sense as the NPFV displays features of red flowers and due to the bias that OpenImages does not contain many images of hummingbird feeders. In these cases, we additionally retrieve Flickr images with appropriate text queries e.g., \"hummingbird feeder\" and filter them.", + "bbox": [ + 496, + 536, + 892, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spurious Class Extension: For \"hummingbird\", \"freight car\", and \"koala\" the spurious features significantly extend the predictions beyond the actual class (see Fig. 2). Bird feeders are classified as hummingbirds, graf", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20241", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4563f6e3923208960d0af87f591719c55f3f480f7ecd87599ed63d9f3efe52e4.jpg", + "image_caption": [ + "Figure 6: Spurious Score: we plot the AUC of different models for 7 out of 100 classes in \"Spurious ImageNet\" (see Fig 13-15): hummingbird (bird feeder/red flowers), freight car (graffiti), koala (plants/trees), fireboat (water jet), flagpole (US flag), gondola (house/river), and bookshop (storefront). The spurious features for hummingbird, freight car and koala which were not detected in [61, 62] are also spurious for their robust ResNet50 [61]. Training on ImageNet 21k or Fine-tuning from 21k (1kFT21k) decreases dependence on harmful spurious features but classes like flagpole and bookshop remain strongly affected. Our cheap mitigation strategy SpuFix improves the AUCs significantly for both robust ResNet50 but also improves the ViT-B variants trained/fine-tuned on ImageNet1k, especially for the difficult classes flagpole and bookshop." + ], + "image_footnote": [], + "bbox": [ + 86, + 93, + 883, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b51854110c5f9577a2e18122ed2a08190de46fc889f1d6c0d18446e8e6427c13.jpg", + "image_caption": [ + "hummingbird (red flowers)", + "fireboat" + ], + "image_footnote": [], + "bbox": [ + 106, + 367, + 205, + 443 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f6a05639cea9b15a673e3da497bd21175863f24e67b1836acae247343f22319.jpg", + "image_caption": [ + "(water jet)", + "Figure 7: Spurious ImageNet: sample images from the dataset for 6 out of 100 classes showing the spurious feature but not the class object, see also App. G." + ], + "image_footnote": [], + "bbox": [ + 106, + 476, + 205, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eed06027d1d78c3d80f163024a2ef6b152593b0359f6fda13b18a2ff780eb396.jpg", + "image_caption": [ + "freight car (graffiti)" + ], + "image_footnote": [], + "bbox": [ + 220, + 367, + 321, + 443 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c6536a6220f471cfa4105ae0c1c1345cf028c60727e243c8859ae4894691d0e.jpg", + "image_caption": [ + "flagpole (US flag)" + ], + "image_footnote": [], + "bbox": [ + 220, + 476, + 321, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7860febbf72c9ffcd5b66b6bba9296c151a3373c7dbbe608d3df22208bf00745.jpg", + "image_caption": [ + "koala (plants/trees)" + ], + "image_footnote": [], + "bbox": [ + 338, + 367, + 436, + 441 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21bf0e1b5da35e544868e770485cc819e022f017481c6cb0c5bb08570e1e56b5.jpg", + "image_caption": [ + "gondola (house/river)" + ], + "image_footnote": [], + "bbox": [ + 338, + 476, + 437, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fiti as freight cars, and (eucalyptus) plants as koalas. This class extension cannot be detected by monitoring test performance and thus is likely to be noticed only after deployment. For \"hummingbird\", we see in Fig. 4 two images with bees on the bird feeder where \"bee\" is an ImageNet class (also a grasshopper for \"koala\"). Nevertheless, the spurious \"bird feeder\" feature of \"hummingbird\" overrules \"bee\" even though no hummingbird is present.", + "bbox": [ + 75, + 628, + 467, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spurious Shared Feature: The spurious feature \"water jet\" is shared among the classes \"fireboat\" and \"fountain\". It appears more frequently for \"fireboat\" (see Fig. 2) which leads to an in-distribution shift where now a large number of images of the \"fountain\"-class with a water jet are wrongly classified as \"fireboat\". The spurious feature \"water jet\" for \"fireboat\" has been found also in the Salient ImageNet dataset [61, 62]. However, they did not find spurious features for freight car and koala (in App. A we do a comparison). More examples are in App. F.", + "bbox": [ + 75, + 750, + 467, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.2. The Spurious ImageNet dataset", + "text_level": 1, + "bbox": [ + 500, + 338, + 777, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A key contribution of this paper is our novel evaluation of spurious features for image classifiers without requiring pixel-wise annotations [48, 59] or having to rely on the validity of neural heatmaps [61]. Instead, we use images from OpenImages to show that images only containing the spurious feature but not the class object are classified as this class. This has the advantage that we consider real images and thus provide a realistic impression of the performance of ImageNet classifiers in the wild. Adding noise [61] or masking [48, 44] image regions requires pixel-wise accurate annotations which are labor-expensive, masking only the object still contains shape information, and using masks avoiding this, e.g., a bounding box around the object, can hide a significant portion of the image which is unrealistic.", + "bbox": [ + 496, + 366, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To allow for a quantitative analysis of the influence of spurious features on ImageNet classifiers, we collected images similar to the ones shown to illustrate the spurious features in Fig. 4. The images are chosen such that they show the spurious feature but not the class object. The only difference is that we relax the classification condition and only require two of the four classifiers (robust ResNet50, ResNext101, EfficientNet-B5, ConvNext-B) to predict the corresponding class. We select 100 of our spurious features and for each collect 75 images from the top-ranked images in OpenImages according to the value of $\\alpha_{l}^{(k)}$ for which two human labelers agree that they contain the spurious feature but not class $k$ and two out of four classifiers predict class $k$ . We call the dataset Spurious ImageNet as it allows to check the dependence on spurious features with real images for ImageNet classifiers, see Fig. 7 and App. G for samples.", + "bbox": [ + 496, + 577, + 890, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spurious Score: A classifier $f$ not relying on the spurious feature should predict a low probability for class $k$ for the Spurious ImageNet samples, especially compared to ImageNet test set images of class $k$ . Thus, for each class, we measure the AUC (area under the curve) for the separation", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20242", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of images with the spurious features but not showing class $k$ versus test set images of class $k$ according to the predicted probability for class $k$ . A classifier not depending on the spurious feature should attain a perfect AUC of 1, whereas a value significantly below 1 shows strong reliance. We report the mean AUC (mAUC) over all 100 classes in Tab. 1. All ImageNet models trained only on ImageNet1k are heavily influenced by spurious features. Thus, spurious features are mainly a problem of the training set rather than the classifier, and spurious features found with an adversarially robust model transfer to other ImageNet classifiers.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pre-training on larger datasets: Some spurious features such as flag (flag pole), bird feeder (hummingbird), and eucalyptus (koala) are classes in ImageNet21k. Therefore, they should no longer be spurious for the other classes. Thus, we test if ImageNet1k-classifiers fine-tuned from an ImageNet21k model are less reliant on spurious features. The results in Tab. 1 and Fig. 6 suggest that the influence of spurious features is damped but they are far from being free of them. To check how much is lost due to fine-tuning we evaluate a ViT-B trained on ImageNet21k which has a mean AUC of 0.931 whereas the fine-tuned model has 0.917. This shows that fine-tuning does not hurt much. While finetuning from ImageNet21k improves the mean AUC, for several classes the dependence on spurious features is still significant, see also Fig. 16 how one has to be careful in the interpretation of higher AUC values. In addition to ImageNet21k, we also evaluate models trained on other large image datasets (JFT-300M[28], YFFC-100M, 1B Instagram[84], MIM[24], LAION-2B and LAION-400M[54]) using self-supervised learning or which are based on CLIP [49]. However, these models also do not achieve better spurious scores (Tab. 1 and Tab. 2). We evaluate a large number of SOTA models in App. D.", + "bbox": [ + 75, + 257, + 472, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7.3. Evaluation of mitigation technique SpuFix", + "text_level": 1, + "bbox": [ + 76, + 618, + 439, + 636 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fixing spurious features is a non-trivial task and can require a substantial labeling effort. We evaluate our SpuFix from Sec. 5 that does not require retraining or additional labels. The positive effect of this fix of spurious features (SpuFix) can be seen in Tab. 1 and Fig. 6. Compared to the original robust ResNet50 with a spurious mAUC of 0.630, the SpuFix version has a significantly better spurious mAUC of 0.763. Test set accuracy reduces by $0.6\\%$ but this is a rather positive effect, as several of the additional errors arise since the robust ResNet50 uses spurious features for its decision, e.g., for classes like \"balance beam\" or \"puck\" the class object is often not visible in the cropped test set images. In Table 2 we provide a large scale evaluation of the transfer of SpuFix to SOTA ImageNet models. We observe a consistently better mAUC on Spurious ImageNet, even for very large models fine-tuned from 21k or trained on other large datasets, e.g. SpuFix improves the mAUC of VOLO-", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/48ef1e4ec4310dcf32c609bedc1f34f6e9cb769f7bf8483cfb2aa4efab7adc2c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelOriginalSpuFix
INet Acc. ↑SpurIN mAUC ↑INet Acc. ↑SpurIN mAUC ↑
ImageNet1k
Rob. ResNet5057.4%0.63056.8%0.763
Rob. ResNet50[61]57.9%0.65157.2%0.764
ConvNeXt-L[41]84.8%0.80384.8%0.819
ViT-B AugReg[66]81.1%0.85081.1%0.859
VOLO-D5 512[85]87.1%0.88287.1%0.907
ImageNet21kFT1k
EfficientNetv2-L[70]86.8%0.89386.8%0.898
ConvNeXt-L[41]87.0%0.91087.0%0.913
ViT-B AugReg[66]86.0%0.91785.9%0.925
BEiT-L\\16[8]88.6%0.92188.6%0.927
LAION-2B
CNeXt-L CLIP 384[49]87.8%0.87987.9%0.884
ViT-L\\14 CLIP[49]88.2%0.91288.2%0.914
MIM
EVA-G\\14 CLIP 560[24]89.8%0.91989.8%0.925
ImageNet21k
ConvNeXt-L[41]-0.943-0.943
ViT-B AugReg[66]-0.931-0.931
", + "bbox": [ + 500, + 87, + 898, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Quantitative Evaluation on Spurious ImageNet: ImageNet classifiers of different training modalities depend on spurious features in varying strength. The mAUC is the mean of AUCs for the separation of images containing the spurious feature but not class $k$ versus test images of class $k$ with the predicted probability of class $k$ as score.", + "bbox": [ + 496, + 439, + 893, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "D5 (87.1% acc.) trained only on 1k by 2.5%, or by 0.6% for EVA-G $\\backslash$ 14 CLIP 560 trained on MIM (91.9% acc.), as well as BeiT-L $\\backslash$ 16 fine-tuned from 21k (88.6% acc.) by 0.6%. Thus even SOTA models profit from our SpuFix with negligible difference in accuracy ( $\\leq$ 0.1%) and thus the use of SpuFix is recommended for any ImageNet model.", + "bbox": [ + 496, + 547, + 893, + 640 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 655, + 619, + 671 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have shown that large-scale identification of spurious features is feasible with our neural PCA components and neural PCA feature visualizations. With \"Spurious ImageNet\" we introduced a novel dataset to evaluate the dependence of ImageNet classifiers on spurious features based on real images. We demonstrated that our SpuFix method mitigates the dependence on harmful spurious features for any ImageNet classifier without costly labeling or re-training.", + "bbox": [ + 496, + 681, + 893, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements We acknowledge support by the DFG, Project number 390727645, the Carl Zeiss Foundation, project \"Certification and Foundations of Safe Machine Learning Systems in Healthcare\". The authors thank the IMPRS-IS for supporting YN.", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20243", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Julius Adebayo, Michael Muelly, Hal Abelson, and Been Kim. Post hoc explanations may be ineffective for detecting unknown spurious correlation. In ICLR, 2022.", + "[2] Julius Adebayo, Michael Muelly, Ilaria Liccardi, and Been Kim. Debugging tests for model explanations. In NeurIPS, 2020.", + "[3] D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Mane. Concrete problems in AI safety. arXiv:1606.06565v2, 2016.", + "[4] Christopher J. Anders, Leander Weber, David Neumann, Wojciech Samek, Klaus-Robert Müller, and Sebastian Lapuschkin. Finding and removing clever hans: Using explanation methods to debug and improve deep models. Information Fusion, 77:261-295, 2022.", + "[5] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. In NeurIPS, 2022.", + "[6] Maximilian Augustin, Alexander Meinke, and Matthias Hein. Adversarial robustness on in- and out-distribution improves explainability. In ECCV, 2020.", + "[7] David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja Hansen, and Klaus-Robert Müller. How to explain individual classification decisions. JMLR, 2010.", + "[8] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022.", + "[9] Sara Beery, Grant van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, 2018.", + "[10] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with ImageNet? arXiv:2006.07159, 2020.", + "[11] B. Biggio, I. Corona, D. Maiorca, B. Nelson, N. Srndic, P. Laskov, G. Giacinto, and F. Roli. Evasion attacks against machine learning at test time. In ECML/PPKD, 2013.", + "[12] Alceu Bissoto, Eduardo Valle, and Sandra Avila. Debiasing skin lesion datasets and models? not so fast. arXiv:2004.11457, 2020.", + "[13] Valentyn Boreiko, Maximilian Augustin, Francesco Croce, Philipp Berens, and Matthias Hein. Sparse visual counterfactual explanations in image space. In $GCPR$ , 2022.", + "[14] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020.", + "[15] Myung Jin Choi, Antonio Torralba, and Alan S. Willsky. Context models and out-of-context objects. Pattern Recognition Letters, 33(7):853-862, 2012.", + "[16] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020." + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR workshop, pages 702-703, 2020.", + "[18] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021.", + "[19] Ann-Kathrin Dombrowski, Maximillian Alber, Christopher Anders, Marcel Ackermann, Klaus-Robert Müller, and Pan Kessel. Explanations can be manipulated and geometry is to blame. In NeurIPS, 2019.", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021.", + "[21] Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. Adversarial robustness as a prior for learned representations, 2019.", + "[22] Christian Etmann, Sebastian Lunz, Peter Maass, and Carola-Bibiane Schonlieb. On the connection between adversarial robustness and saliency map interpretability. In ICML, 2019.", + "[23] Sabri Eyuboglu, Maya Varma, Khaled Saab, Jean-Benoit Delbrouck, Christopher Lee-Messer, Jared Dunnmon, James Zou, and Christopher Ré. Domino: Discovering systematic errors with cross-modal embeddings. In ICLR, 2022.", + "[24] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636, 2022.", + "[25] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020.", + "[26] R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019.", + "[27] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In AAAI, 2019.", + "[28] A Gupta, C Sun, A Shrivastava, and S Singh. Revisiting the unreasonable effectiveness of data. arXiv preprint arXiv:1707.02968, 2017.", + "[29] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016.", + "[30] M. Hein, M. Andriushchenko, and J. Bitterwolf. Why ReLU networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In CVPR, 2019.", + "[31] D. Hendrycks and T. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019.", + "[32] D. Hendrycks and K. Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017.", + "[33] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20244", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Sara Hooker, Yann Dauphin, Aaron Courville, and Andrea Frome. Selective brain damage: Measuring the disparate impact of model pruning. In ICLR, 2020.", + "[35] Pieter Abbeel Jonathan Ho, Ajay Jain. Denoising diffusion probabilistic models. In NeurIPS, 2020.", + "[36] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In ECCV, 2020.", + "[37] A. Kurakin, I. J. Goodfellow, and S. Bengio. Adversarial examples in the physical world. In *ICLR Workshop*, 2017.", + "[38] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020.", + "[39] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022.", + "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021.", + "[41] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022.", + "[42] Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In NeurIPS, 2017.", + "[43] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018.", + "[44] Mazda Moayeri, Sahil Singla, and Soheil Feizi. Hard ImageNet: Segmentations for objects with strong spurious cues. In NeurIPS Datasets and Benchmarks Track, 2022.", + "[45] A. Nguyen, J. Yosinski, and J. Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In CVPR, 2015.", + "[46] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022.", + "[47] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. Proc. CVPR, 2022.", + "[48] Gregory Plumb, Marco Tulio Ribeiro, and Ameet Talwalkar. Finding and fixing spurious patterns with explanations. Transactions on Machine Learning Research (TMLR), 2022.", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021.", + "[50] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In ICML, 2019." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\": Explaining the predictions of any classifier. In KDD, 2016.", + "[52] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. License: No license specified.", + "[53] Shibani Santurkar, Dimitris Tsipras, Brandon Tran, Andrew Ilyas, Logan Engstrom, and Aleksander Madry. Image synthesis with a single (robust) classifier. In NeurIPS, 2019.", + "[54] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022.", + "[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.", + "[56] Kathryn Schutte, Olivier Moindrot, Paul Hérent, Jean-Baptiste Schiratti, and Simon Jégou. Using stylegan for visual interpretability of deep learning models on medical images. In NeurIPS Workshop, 2020.", + "[57] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017.", + "[58] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. In NeurIPS, 2020.", + "[59] Rakshith Shetty, Bernt Schiele, and Mario Fritz. Not using the car to see the sidewalk-quantifying and controlling the effects of context in classification and segmentation. In CVPR, 2019.", + "[60] Krishna Kumar Singh, Dhruv Mahajan, Kristen Grauman, Yong Jae Lee, Matt Feiszli, and Deepti Ghadiyaram. Don't judge an object by its context: learning to overcome contextual bias. In CVPR, 2020.", + "[61] Sahil Singla and Soheil Feizi. Salient imagenet: How to discover spurious features in deep learning? In ICLR, 2022.", + "[62] Sahil Singla, Mazda Moayeri, and Soheil Feizi. Core risk minimization using salient imagenet. arXiv:2203.15566, 2022.", + "[63] Sahil Singla, Besmira Nushi, Shital Shah, Ece Kamar, and Eric Horvitz. Understanding failures of deep networks via robust feature extraction. In CVPR, 2021.", + "[64] Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju. Fooling lime and shap: Adversarial attacks on post hoc explanation methods. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 180-186, 2020.", + "[65] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20245", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. In TMLR, 2022.", + "[67] Pierre Stock and Moustapha Cisse. Convnets and ImageNet beyond accuracy: Understanding mistakes and uncovering biases. In ECCV, 2018.", + "[68] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014.", + "[69] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019.", + "[70] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In ICML, 2021.", + "[71] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR, pages 1521-1528, 2011.", + "[72] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022.", + "[73] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. From ImageNet to image classification: Contextualizing progress on benchmarks. In ICML, 2020.", + "[74] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019.", + "[75] Sahil Verma, John P. Dickerson, and Keegan Hines. Counterfactual explanations for machine learning: A review. arXiv:2010.10596, 2020.", + "[76] Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the GDPR. Harvard Journal of Law & Technology, 2018.", + "[77] Ross Wightman. Pytorch image models. https://github.com/ rwightman/pytorch-image-models, 2019.", + "[78] Eric Wong, Shibani Santurkar, and Aleksander Madry. Leveraging sparse linear layers for debuggable deep networks. In ICML, 2021.", + "[79] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023.", + "[80] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. In ICLR, 2021.", + "[81] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020.", + "[82] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017.", + "[83] Cong Xu and Min Yang. Adversarial momentum-contrastive pre-training. arXiv preprint, arXiv:2012.13154v2, 2020.", + "[84] I Zeki Yalniz, Herve Jégou, Kan Chen, Manohar Paluri, and Dhruv Mahajan. Billion-scale semi-supervised learning for image classification. arXiv preprint arXiv:1905.00546, 2019." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision observer for visual recognition. PAMI, 2022.", + "[86] John R. Zech, Marcus A. Badgeley, Manway Liu, Anthony B. Costa, Joseph J. Titano, and Eric Karl Oermann. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLOS Medicine, 15(11):1-17, 2018.", + "[87] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018.", + "[88] Zhuotun Zhu, Lingxi Xie, and Alan L. Yuille. Object recognition with and without objects. In *IJCAI*, 2017." + ], + "bbox": [ + 501, + 90, + 890, + 276 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20246", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/8d4ee68a-1062-4c0b-81d9-4f01bf36823f_model.json b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/8d4ee68a-1062-4c0b-81d9-4f01bf36823f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e27efb2b4bc81e75627df78e8ece5f07a41d230 --- /dev/null +++ b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/8d4ee68a-1062-4c0b-81d9-4f01bf36823f_model.json @@ -0,0 +1,4976 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.13, + 0.872, + 0.177 + ], + "angle": 0, + "content": "Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.204, + 0.273, + 0.221 + ], + "angle": 0, + "content": "Yannic Neuhaus" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.204, + 0.487, + 0.222 + ], + "angle": 0, + "content": "Maximilian Augustin" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.204, + 0.672, + 0.221 + ], + "angle": 0, + "content": "Valentyn Boreiko" + }, + { + "type": "text", + "bbox": [ + 0.716, + 0.204, + 0.831, + 0.221 + ], + "angle": 0, + "content": "Matthias Hein" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.222, + 0.667, + 0.24 + ], + "angle": 0, + "content": "Tübingen AI Center - University of Tübingen" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.306, + 0.474, + 0.624 + ], + "angle": 0, + "content": "Benchmark performance of deep learning classifiers alone is not a reliable predictor for the performance of a deployed model. In particular, if the image classifier has picked up spurious features in the training data, its predictions can fail in unexpected ways. In this paper, we develop a framework that allows us to systematically identify spurious features in large datasets like ImageNet. It is based on our neural PCA components and their visualization. Previous work on spurious features often operates in toy settings or requires costly pixel-wise annotations. In contrast, we work with ImageNet and validate our results by showing that presence of the harmful spurious feature of a class alone is sufficient to trigger the prediction of that class. We introduce the novel dataset \"Spurious ImageNet\" which allows to measure the reliance of any ImageNet classifier on harmful spurious features. Moreover, we introduce SpuFix as a simple mitigation method to reduce the dependence of any ImageNet classifier on previously identified harmful spurious features without requiring additional labels or retraining of the model. We provide code and data at https://github.com/YanNeu/spuriousImagenet." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.649, + 0.21, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Deep learning has led to tremendous progress in image classification [37, 50] and natural language processing [14]. Over the years, however, it has become apparent, that evaluating predictive performance on a fixed test set is not necessarily indicative of the performance when image classifiers are deployed in the wild. Several potential failure cases have been discovered. This starts with a lack of robustness due to image corruptions [31], adversarial perturbations [68], and arbitrary predictions on out-of-distribution inputs [45, 32, 30]. In this paper, we consider the problem of identifying and debugging image classifiers from spurious features [2]. Spurious features in image classification are features that co-occur with the actual class object and are picked up by the classifier. In the worst case, they lead to shortcut learning [25], where only the spurious but not the" + }, + { + "type": "text", + "bbox": [ + 0.58, + 0.274, + 0.815, + 0.289 + ], + "angle": 0, + "content": "Spurious Features in Training Data" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.29, + 0.602, + 0.303 + ], + "angle": 0, + "content": "bird feeder" + }, + { + "type": "text", + "bbox": [ + 0.633, + 0.29, + 0.679, + 0.303 + ], + "angle": 0, + "content": "graffiti" + }, + { + "type": "text", + "bbox": [ + 0.709, + 0.29, + 0.781, + 0.303 + ], + "angle": 0, + "content": "eucalyptus" + }, + { + "type": "text", + "bbox": [ + 0.816, + 0.29, + 0.85, + 0.303 + ], + "angle": 0, + "content": "label" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.304, + 0.61, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.304, + 0.7, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.304, + 0.79, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.304, + 0.877, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.373, + 0.612, + 0.388 + ], + "angle": 0, + "content": "Hummingbird" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.373, + 0.696, + 0.388 + ], + "angle": 0, + "content": "Freight Car" + }, + { + "type": "text", + "bbox": [ + 0.725, + 0.373, + 0.766, + 0.387 + ], + "angle": 0, + "content": "Koala" + }, + { + "type": "text", + "bbox": [ + 0.799, + 0.373, + 0.868, + 0.387 + ], + "angle": 0, + "content": "Hard Disc" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.394, + 0.847, + 0.422 + ], + "angle": 0, + "content": "Images from the web with spurious feature but no class features classified as class below" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.423, + 0.609, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.423, + 0.7, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.423, + 0.79, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.423, + 0.877, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.492, + 0.609, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.492, + 0.7, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.492, + 0.79, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.492, + 0.877, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.563, + 0.612, + 0.577 + ], + "angle": 0, + "content": "Hummingbird" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.563, + 0.696, + 0.577 + ], + "angle": 0, + "content": "Freight Car" + }, + { + "type": "text", + "bbox": [ + 0.725, + 0.563, + 0.766, + 0.576 + ], + "angle": 0, + "content": "Koala" + }, + { + "type": "text", + "bbox": [ + 0.799, + 0.563, + 0.868, + 0.576 + ], + "angle": 0, + "content": "Hard Disc" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.59, + 0.895, + 0.697 + ], + "angle": 0, + "content": "Figure 1: Top: Examples of spurious features found via our neural PCA components but not in previous study [61]. Bottom: We validate our spurious features by mining images from the web showing only the spurious feature but not the class. They are classified by four ImageNet models as the corresponding class. Some of them even contain ImageNet classes (bees on feeder, grasshopper in leaves)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "correct feature is associated with the class, e.g., [86] found that a pneumonia detector's bad generalization across hospitals was caused by the neural network learning to identify the hospital where the training data originated from. A weaker form of spurious feature (at least from a learning perspective) is the case when the classifier picks up the correct class features, e.g., of a hummingbird, but additionally associates a spurious feature, e.g., a bird feeder, with the class as they appear together on a subset of the training set. This becomes a harmful spurious feature if only the spurious feature without the class feature is sufficient to trigger the classification of that class, see Fig. 1 for an illustration of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20235" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "such spurious features found via our method. Harmful spurious features are difficult to detect and thus can easily go unnoticed, leading to unexpected behaviour of a deployed image classifier." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.153, + 0.439, + 0.167 + ], + "angle": 0, + "content": "In this paper we make the following key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.168, + 0.469, + 0.243 + ], + "angle": 0, + "content": "- we develop a pipeline for the detection of harmful spurious features with little human supervision based on our class-wise neural PCA (NPCA) components of an adversarially robust classifier together with their Neural PCA Feature Visualization (NPFV)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.243, + 0.469, + 0.303 + ], + "angle": 0, + "content": "- unlike prior work, which used masking images or pixelwise annotations, we validate our found spurious features by using our NPCA components to find real images containing only the spurious feature but not the class object." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.304, + 0.469, + 0.424 + ], + "angle": 0, + "content": "- using these images we create the dataset \"Spurious ImageNet\" and propose a measure for dependence on spurious features. We do a large-scale evaluation of state-of-the-art (SOTA) ImageNet models. We show that the spurious features found for the robust model generalize to non-robust classifiers. Moreover, we analyze the influence of different training setups, e.g. pre-training on ImageNet21k or larger datasets like LAION." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.424, + 0.469, + 0.514 + ], + "angle": 0, + "content": "- we develop SpuFix, a technique to mitigate the dependence on identified harmful spurious features without requiring new labels or retraining, and show how to transfer it to any ImageNet classifier. SpuFix consistently improves the dependence on harmful spurious features even for SOTA models with negligible impact on test accuracy." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.168, + 0.469, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.528, + 0.214, + 0.543 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.659 + ], + "angle": 0, + "content": "When classifiers in safety-critical systems such as healthcare or autonomous driving are deployed in the wild [3], it is important to discover potential failure cases before release. Prior work has focused on corruption [31], adversarial robustness [11, 68, 43], and out-of-distribution detection [45, 32, 30]. There is less work on spurious features, although their potential harm might be higher." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.811 + ], + "angle": 0, + "content": "Spurious features: It has been noted early on that classifiers show reliance on spurious features [15] e.g., a cow on the beach is not recognized [9] due to the missing spurious feature of grass. Other forms of spurious features have been reported in the classification of skin lesions [12], pneumonia [86], traffic signs [67], and object recognition [88]. Moreover, it has been shown that deep neural networks are biased towards texture [26] and background context [80], see [25] for an overview. [58] argues theoretically that spurious features are picked up due to a simplicity bias." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Detection of spurious features has been achieved using human label-intense pixel-wise annotations [48, 59, 60]. In [78], they use sparsity regularization to enforce a more interpretable model and use it for finding spurious features. [4] propose a complex pipeline to detect spurious features. While they scale to ImageNet, their analysis is lim" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "ited to a few spurious features for a subset of 100 classes. [61, 44, 62] do a search on full ImageNet based on class-weighted \"neural maps\". The neural maps are used to add noise to \"spurious\" resp. \"core\" features but no significant difference in classification performance is observed. It remains unclear if their found spurious features are harmful, that is the feature alone triggers the decision for that class." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Interpretability methods: In recent years several interpretability methods have been proposed e.g., attribution methods such as GradCAM [57], Shapley values [42], Relevance Propagation [7], and LIME [51]. The use of these methods for the detection of spurious features has been analyzed in [2, 1] with mixed success and it has been argued that interpretability methods are not robust [19, 64, 27]. However, attribution methods work better for robust classifiers due to more interpretable gradients [22]. Another technique is counterfactual explanations [76, 75] which are difficult to generate for images due to the similarity to adversarial examples [68]. Thus visual counterfactual explanations are realized via manipulation of a latent space [56] or in image space [53, 6, 13] for an adversarily robust classifier. Visual counterfactuals for non-robust classifiers using diffusion models [35, 65, 33, 18] have been proposed in [5]. ImageNet: ImageNet [52] suffers from several shortcomings: apart from an inherent dataset bias [71], semantically overlapping or even identical class pairs were reported [34, 73, 10], e.g., two classes \"maillot\", \"sunglass\" vs \"sun-glasses\", \"notebook\" vs \"laptop\" etc. We disregard such trivial cases of dataset contamination and focus on classes with harmful spurious features, in particular ones where only a small portion of the training set is contaminated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.671, + 0.59 + ], + "angle": 0, + "content": "3. Spurious features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A proper definition of spurious features is difficult. We describe two settings of harmful spurious features which appear in this paper. We denote by \\( C_k \\) the set of all images containing objects belonging to class \\( k \\) (assuming for simplicity that we have a deterministic problem and ignoring multi-labels). Let \\( S \\) be the set of all images containing a feature \\( s \\) (e.g., a bird feeder). It is a correlated feature for class \\( k \\) when \\( C_k \\cap S \\) and \\( S \\setminus C_k \\) are non-empty, i.e., the feature occurs frequently with the class object but there is no causal implication that appearance of \\( s \\) implies the appearance of the class object (a bird feeder in the image does not imply presence of a hummingbird). A correlated feature becomes a spurious feature when the classifier picks it up as feature of this class. Not every spurious feature is immediately harmful, even humans use context information [25] to get more confident in a decision. However, a spurious feature is harmful if the spurious feature alone is enough to trigger the decision for the corresponding class without the class object being present in the image. We consider two scenarios for a harmful spurious feature shown in Fig. 2." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20236" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.473, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.354 + ], + "angle": 0, + "content": "Figure 2: Type of harmful Spurious Feature: Left: The spurious feature \\( s \\) is taken up by the classifier which predicts class \\( k \\) on \\( C_k \\cup S \\) instead only on \\( C_k \\). Right: The spurious feature \\( s \\) is shared between classes but appears more often in class \\( k \\). The classifier associates \\( S \\) with class \\( k \\) and thus predicts class \\( k \\) also on \\( C_l \\cap S \\) instead of class \\( l \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.522 + ], + "angle": 0, + "content": "Spurious Class Extension: For this type of spurious feature (left in Fig. 2) the classifier picks up the spurious feature \\( s \\) for class \\( k \\) and predicts the class \\( k \\) even on \\( S \\setminus C_k \\) with high confidence (prediction of \"hummingbird\" for images showing a bird feeder but no hummingbird). The classifier predicts class \\( k \\) beyond its actual domain \\( C_k \\) and thus we call this a spurious class extension. While this spurious feature does not necessarily hurt in terms of test performance it can easily lead to completely unexpected behavior when the classifier is deployed in the wild." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.523, + 0.47, + 0.599 + ], + "angle": 0, + "content": "Spurious Shared Feature: Here, two classes \\( C_k \\) and \\( C_l \\) share a spurious feature \\( s \\) (e.g., \"water jet\" for the classes \"fireboat\" and \"fountain\"). As there are more training images with feature \\( s \\) in \\( C_k \\) than in \\( C_l \\) the classifier associates \\( S \\) with class \\( k \\) and predicts class \\( k \\) for \\( S \\cap C_l \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.794 + ], + "angle": 0, + "content": "The two types of harmful spurious features are not exclusive. A shared spurious feature \\( s \\) can at the same time lead to a spurious class extension, e.g., the object bird feeder leads to a spurious class extension of the class \"hummingbird\" (see Fig. 4) to images of bird feeders without hummingbirds. In the training set the hummingbird feeder appears only in images of class \"hummingbird\" but the hummingbird feeder has parts which mimic flowers and flowers are a shared spurious feature with bees. In Fig. 4 right top row, images of bees on a bird feeder are classified as \"hummingbird\" instead of \"bee\", so the spurious feature is strong enough to override the decision for the true class \"bee\" (spurious shared feature)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.827, + 0.469, + 0.86 + ], + "angle": 0, + "content": "4. Finding spurious features via neural PCA and associated feature visualizations" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "First, we define our class-wise neural PCA (NPCA) which allows us to find diverse subpopulations in the train-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "ing data, e.g., we checked that the bird feeder for \"hummingbird\" is visible in \\(15\\%\\) of the training images (component 2), while another \\(15\\%\\) contain a part of it (component 3), see Fig. 3 or, for more examples, App. F. Then we introduce our neural PCA feature visualization (which requires an adversarially robust model) and how we select NPCA components for human inspection. The identification of spurious features requires minimal human supervision, and our effective setup allows us to screen all ImageNet classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.893, + 0.529 + ], + "angle": 0, + "content": "Adversarily robust model: Similar to [61], we use an adversarily robust model to find spurious features in ImageNet. The reason for this is that robust models have generative properties [74, 53, 6, 61, 13] in the sense that maximizing the predicted probability of a class in a neighborhood of an image leads to semantically meaningful changes. They also have more informative gradients [22] and thus attribution maps such as GradCAM [57] work better. We use the multiple-norm robust model of [13] as they claim it has the best generative properties. The generative properties of robust models are mainly used for the neural PCA feature visualization where we maximize the NPCA component of a class starting from a gray image. If a spurious feature appears without the class object, this is a strong indicator of a harmful spurious feature. A non-robust model would only produce semantically meaningless adversarial noise, hence we need a robust model for this part of our detection pipeline. Our detected spurious features are not specific to the robust model. We show that SOTA ImageNet models share the same spurious features (Fig. 4 and Sec. 7.2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.605 + ], + "angle": 0, + "content": "Class-wise neural PCA: Let \\((x_{i},y_{i})_{i = 1}^{N}\\) be the training set, where \\(y_{i}\\in \\{1,\\dots ,K\\}\\) and \\(K\\) is the number of classes. We consider features of the penultimate layer \\(\\phi (x)\\in \\mathbb{R}^D\\) of a neural network for an input \\(x\\). For a given class \\(k\\) and its associated weights \\(w_{k}\\in \\mathbb{R}^{D}\\) in the final layer, we define" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.613, + 0.891, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\psi_ {k} (x) = w _ {k} \\odot \\phi (x). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.891, + 0.666 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the componentwise product. Let \\(b\\in \\mathbb{R}^K\\) be the bias vector of the final layer then the logit \\(f_{k}\\) of class \\(k\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.671, + 0.891, + 0.7 + ], + "angle": 0, + "content": "\\[\nf _ {k} (x) = \\sum_ {j = 1} ^ {D} w _ {k j} \\phi (x) _ {j} + b _ {k} = \\langle \\mathbf {1}, \\psi_ {k} (x) \\rangle + b _ {k}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.796 + ], + "angle": 0, + "content": "Let \\(I_{k}\\) be the index set of the training set of class \\(k\\) and \\(\\bar{\\psi}_k\\) the class-wise mean, \\(\\bar{\\psi}_k = \\frac{1}{|I_k|}\\sum_{s\\in I_k}\\psi_k(x_s)\\). The class-wise neural PCA allows us to identify variations in the set \\(\\{\\psi (x_r)\\}_{r\\in I_k}\\) arising due to small subpopulations in the training set. In the class-wise neural PCA, we compute eigenvectors of the class-wise covariance matrix," + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.802, + 0.891, + 0.835 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {s \\in I _ {k}} \\left(\\psi_ {k} \\left(x _ {s}\\right) - \\bar {\\psi} _ {k}\\right) \\left(\\psi_ {k} \\left(x _ {s}\\right) - \\bar {\\psi} _ {k}\\right) ^ {T}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.842, + 0.892, + 0.872 + ], + "angle": 0, + "content": "The eigenvectors, \\(v_{1},\\ldots ,v_{D}\\) form an orthonormal basis of \\(\\mathbb{R}^D\\) and we write \\(\\psi_k(x) - \\bar{\\psi}_k\\) in this basis," + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.878, + 0.891, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\psi_ {k} (x) - \\bar {\\psi} _ {k} = \\sum_ {l = 1} ^ {D} v _ {l} \\left\\langle \\psi_ {k} (x) - \\bar {\\psi} _ {k}, v _ {l} \\right\\rangle , \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20237" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.09, + 0.406, + 0.104 + ], + "angle": 0, + "content": "Top 5 neural PCA components (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.119, + 0.097, + 0.152 + ], + "angle": 0, + "content": "NNE" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.105, + 0.468, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.17, + 0.097, + 0.222 + ], + "angle": 0, + "content": "Max. act." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.165, + 0.468, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.228, + 0.097, + 0.278 + ], + "angle": 0, + "content": "Heeannnnnne" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.226, + 0.468, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.471, + 0.421 + ], + "angle": 0, + "content": "Figure 3: Top 5 Neural PCA components for class hummingbird: first row shows our neural PCA feature visualization (NPFV), second row shows four most activating training images of each NPCA component and the last row GradCAM for the NPCA component. Our NPCA components capture different subpopulations in the training data. Comp. 2 is identified as spurious feature \"bird feeder\", see NPFV and most activating training images (see also Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.437, + 0.151, + 0.451 + ], + "angle": 0, + "content": "and define" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.461, + 0.47, + 0.483 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {l} ^ {(k)} (x) = \\left\\langle \\mathbf {1}, v _ {l} \\right\\rangle \\left\\langle \\psi_ {k} (x) - \\bar {\\psi} _ {k}, v _ {l} \\right\\rangle , \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.492, + 0.439, + 0.508 + ], + "angle": 0, + "content": "The logit \\( f_{k}(x) \\) of the \\( k \\)-th class can then be written as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.519, + 0.469, + 0.547 + ], + "angle": 0, + "content": "\\[\nf _ {k} (x) = \\sum_ {l = 1} ^ {D} \\alpha_ {l} ^ {(k)} (x) + \\left\\langle \\mathbf {1}, \\bar {\\psi} _ {k} \\right\\rangle + b _ {k}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Thus for a given \\(x\\), we can interpret \\(\\alpha_l^{(k)}(x)\\) as the contribution of the neural PCA component \\(l\\) of class \\(k\\) to the logit \\(f_k(x)\\) of class \\(k\\) since the term \\(\\langle \\mathbf{1}, \\bar{\\psi}_k \\rangle + b_k\\) is constant for all inputs. Based on this we introduce a mitigation technique for spurious features without retraining in Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.651, + 0.47, + 0.751 + ], + "angle": 0, + "content": "Neural PCA Feature Visualization: To identify semantic features corresponding to our neural PCA component \\(l\\), we show the training images which attain the maximal values of \\(\\alpha_{l}^{(k)}(x)\\). Additionally, we generate an image \\(z_{l}^{(k)}\\), which we call the Neural PCA Feature Visualization (NPFV) of feature \\(l\\) of class \\(k\\), by maximizing \\(\\alpha_{l}^{(k)}(x)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.762, + 0.407, + 0.785 + ], + "angle": 0, + "content": "\\[\nz _ {l} ^ {(k)} = g + \\operatorname * {a r g m a x} _ {\\| \\delta \\| _ {2} \\leq \\epsilon} \\alpha_ {l} ^ {(k)} (g + \\delta),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\( g \\) is a gray image (all channels equal to 0.5). Thus we maximize the feature \\( \\alpha_{l}^{(k)} \\) outgoing from a non-informative and unbiased initialization \\( g \\). The optimization problem is solved using adaptive projected gradient descent (APGD) [16] with 200 steps. The budget for changes, \\( \\epsilon = 30 \\), is small to avoid the overactivation of feature attacks maximizing the output of individual neurons [21, 63, 61], see" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Fig. 8. In Fig. 4 we show for each identified spurious feature, the corresponding NPFV, e.g., for \"hummingbird\" one can see the bird feeder but no hummingbird. The NPCA components together with the maximally activating training images are in principle sufficient to identify spurious features, but the NPFV is very useful to judge how harmful a spurious feature is. Therefore, we use an adversarily robust model, since a non-robust model would yield semantically meaningless adversarial noise as NPFV." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.893, + 0.502 + ], + "angle": 0, + "content": "Selection of neural PCA components for human inspection: The penultimate layer of the robust ResNet50 we are using has 2048 neurons. Thus it is infeasible (and unnecessary) to investigate all neural PCA components. A strong criterion that one has found a harmful spurious feature is i) the NPFV shows mainly the spurious feature and not the class, and ii) the NPFV has high confidence. If ii) is not satisfied, then the NPFV is a spurious feature the classifier may have picked up, but it is not harmful in the sense that this feature alone causes the classifier to choose that class. Moreover, we noticed that the eigenvalues of the neural PCA, and the corresponding \\(\\alpha\\) values, decay quickly. Thus we compute the NPFV for the top 128 neural PCA components (having maximal variance) and then select the ten components which realize the highest confidence for their NPFV in the corresponding class. Note that we do not optimize the confidence when generating the NPFV but only \\(\\alpha_{l}^{(k)}(x)\\) which is part of the logit of the \\(k\\)-th class." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.893, + 0.798 + ], + "angle": 0, + "content": "Identification of spurious neural PCA components via human supervision: For each ImageNet class \\( k \\) we show the human labeler the top 10 components. For each component \\( l \\) we show the NPFV \\( z_{l}^{(k)} \\) and the 5 training images \\( x_{r} \\) of class \\( k \\) with the largest values of \\( \\alpha_{l}^{(k)}(x_{r}) \\). Moreover, we compute GradCAM [57] images for the NPFV and the five training images using the NPCA component \\( \\alpha_{l}^{(k)} \\) as score. The human marks a component as spurious if i) the NPFV shows dominantly an object not belonging to the class ii) the five training images show consistently this object, iii) the GradCAM activations are primarily not on the class object. The setup shown to the human labeler can be seen in App. B. The labeling of one class takes on average about 45 seconds, so the full labeling of all ImageNet classes took about 13 hours. The human labeler (researcher in machine learning) found in total 337 spurious components. Another human labeler checked all of them and removed spurious features in case of disagreement, resulting in 322 spurious features from 230 ImageNet classes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.814, + 0.859, + 0.832 + ], + "angle": 0, + "content": "5. SpuFix - Mitigation of spurious features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Once the spurious features are identified, the question is how one can mitigate that the classifier relies on them. One way is to identify the training images containing the spurious feature and then discard or downweight them during" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20238" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.09, + 0.461, + 0.104 + ], + "angle": 0, + "content": "Hummingbird - Random train. images (confidence / \\(\\alpha_{k}\\))" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.105, + 0.16, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.169, + 0.155, + 0.181 + ], + "angle": 0, + "content": "0.93/1.7" + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.184, + 0.151, + 0.196 + ], + "angle": 0, + "content": "NPFV-2" + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.105, + 0.24, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.169, + 0.321, + 0.182 + ], + "angle": 0, + "content": "1.00/-0.9 0.96/-1.0" + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.184, + 0.478, + 0.198 + ], + "angle": 0, + "content": "Max. activating train. images - NPCA Comp. 2" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.105, + 0.321, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.199, + 0.321, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.263, + 0.312, + 0.276 + ], + "angle": 0, + "content": "1.00/5.9" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.105, + 0.401, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.169, + 0.478, + 0.182 + ], + "angle": 0, + "content": "0.99/2.2 1.00/1.54" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.2, + 0.401, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.263, + 0.394, + 0.276 + ], + "angle": 0, + "content": "1.00/5.6" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.105, + 0.481, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.169, + 0.478, + 0.182 + ], + "angle": 0, + "content": "1.00/1.54" + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.184, + 0.478, + 0.198 + ], + "angle": 0, + "content": "A Comp. 2" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.2, + 0.16, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.263, + 0.155, + 0.276 + ], + "angle": 0, + "content": "1.00/9.7" + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.2, + 0.24, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.263, + 0.234, + 0.276 + ], + "angle": 0, + "content": "1.00/7.5" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.2, + 0.321, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.263, + 0.312, + 0.276 + ], + "angle": 0, + "content": "1.00/5.9" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.2, + 0.401, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.264, + 0.394, + 0.276 + ], + "angle": 0, + "content": "1.00/5.6" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.2, + 0.481, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.264, + 0.475, + 0.276 + ], + "angle": 0, + "content": "1.00/5.6" + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.279, + 0.44, + 0.293 + ], + "angle": 0, + "content": "Freight car - Random train. images (confidence / \\(\\alpha_{k}\\))" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.294, + 0.16, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.357, + 0.155, + 0.37 + ], + "angle": 0, + "content": "1.00/7.7" + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.372, + 0.15, + 0.385 + ], + "angle": 0, + "content": "NPFV-1" + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.294, + 0.24, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.357, + 0.234, + 0.37 + ], + "angle": 0, + "content": "1.00/3.4" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.294, + 0.321, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.357, + 0.312, + 0.37 + ], + "angle": 0, + "content": "1.00/6.3" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.294, + 0.401, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.357, + 0.401, + 0.37 + ], + "angle": 0, + "content": "0.98/-0.8" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.294, + 0.481, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.357, + 0.475, + 0.37 + ], + "angle": 0, + "content": "1.00/4.2" + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.372, + 0.475, + 0.387 + ], + "angle": 0, + "content": "Max. activating train. images - NPCA Comp. 1" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.388, + 0.16, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.452, + 0.156, + 0.464 + ], + "angle": 0, + "content": "1.00/12.1" + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.388, + 0.241, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.452, + 0.237, + 0.464 + ], + "angle": 0, + "content": "1.00/10.9" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.388, + 0.321, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.452, + 0.317, + 0.464 + ], + "angle": 0, + "content": "1.00/10.4" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.388, + 0.401, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.452, + 0.4, + 0.464 + ], + "angle": 0, + "content": "1.00/10.2" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.388, + 0.481, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.452, + 0.477, + 0.464 + ], + "angle": 0, + "content": "1.00/10.2" + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.468, + 0.401, + 0.481 + ], + "angle": 0, + "content": "Koala - Random train. images (confidence / \\(\\alpha_{k}\\))" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.482, + 0.16, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.547, + 0.157, + 0.559 + ], + "angle": 0, + "content": "1.00/0.77" + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.482, + 0.241, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.547, + 0.233, + 0.559 + ], + "angle": 0, + "content": "0.87/2.4" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.482, + 0.321, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.547, + 0.312, + 0.559 + ], + "angle": 0, + "content": "1.00/0.4" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.482, + 0.401, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.547, + 0.394, + 0.559 + ], + "angle": 0, + "content": "1.00/0.0" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.482, + 0.481, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.547, + 0.475, + 0.559 + ], + "angle": 0, + "content": "0.95/0.5" + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.562, + 0.15, + 0.574 + ], + "angle": 0, + "content": "NPFV-3" + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.562, + 0.477, + 0.576 + ], + "angle": 0, + "content": "Max. activating train. images - NPCA Comp. 3" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.577, + 0.16, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.64, + 0.155, + 0.653 + ], + "angle": 0, + "content": "1.00/5.5" + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.577, + 0.241, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.641, + 0.234, + 0.653 + ], + "angle": 0, + "content": "1.00/4.6" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.577, + 0.321, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.641, + 0.313, + 0.653 + ], + "angle": 0, + "content": "1.00/4.5" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.577, + 0.401, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.641, + 0.394, + 0.653 + ], + "angle": 0, + "content": "1.00/4.4" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.577, + 0.481, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.641, + 0.475, + 0.653 + ], + "angle": 0, + "content": "0.86/4.3" + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.657, + 0.42, + 0.67 + ], + "angle": 0, + "content": "Fireboat - Random train. images (confidence / \\(\\alpha_{k}\\))" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.671, + 0.16, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.735, + 0.158, + 0.747 + ], + "angle": 0, + "content": "0.88/-1.1" + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.671, + 0.24, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.735, + 0.233, + 0.747 + ], + "angle": 0, + "content": "0.95/0.5" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.671, + 0.321, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.735, + 0.319, + 0.747 + ], + "angle": 0, + "content": "0.12/-0.6" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.671, + 0.401, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.735, + 0.4, + 0.747 + ], + "angle": 0, + "content": "0.84/-0.2" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.671, + 0.481, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.735, + 0.476, + 0.747 + ], + "angle": 0, + "content": "0.02/-1.1" + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.751, + 0.15, + 0.763 + ], + "angle": 0, + "content": "NPFV-2" + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.751, + 0.478, + 0.765 + ], + "angle": 0, + "content": "Max. activating train. images - NPCA Comp. 2" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.765, + 0.16, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.829, + 0.155, + 0.842 + ], + "angle": 0, + "content": "1.00/5.5" + }, + { + "type": "image", + "bbox": [ + 0.161, + 0.765, + 0.241, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.829, + 0.233, + 0.842 + ], + "angle": 0, + "content": "0.42/4.1" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.765, + 0.321, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.829, + 0.314, + 0.842 + ], + "angle": 0, + "content": "1.00/4.0" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.765, + 0.401, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.829, + 0.394, + 0.842 + ], + "angle": 0, + "content": "1.00/3.9" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.765, + 0.481, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.829, + 0.475, + 0.842 + ], + "angle": 0, + "content": "1.00/3.9" + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.09, + 0.886, + 0.104 + ], + "angle": 0, + "content": "Images with spurious bird feeder but no hummingbird" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.105, + 0.581, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.169, + 0.575, + 0.181 + ], + "angle": 0, + "content": "0.94/5.7" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.105, + 0.661, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.169, + 0.654, + 0.181 + ], + "angle": 0, + "content": "0.94/3.4" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.105, + 0.74, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.169, + 0.734, + 0.181 + ], + "angle": 0, + "content": "0.82/2.9" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.105, + 0.821, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.169, + 0.814, + 0.181 + ], + "angle": 0, + "content": "0.91/5.6" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.105, + 0.901, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.169, + 0.894, + 0.181 + ], + "angle": 0, + "content": "0.91/4.7" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.184, + 0.889, + 0.198 + ], + "angle": 0, + "content": "all classified as humming bird by four ImageNet models" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.199, + 0.581, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.263, + 0.574, + 0.276 + ], + "angle": 0, + "content": "0.86/4.3" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.199, + 0.661, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.263, + 0.654, + 0.276 + ], + "angle": 0, + "content": "0.81/3.5" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.199, + 0.74, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.263, + 0.734, + 0.276 + ], + "angle": 0, + "content": "0.89/3.3" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.199, + 0.821, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.263, + 0.815, + 0.276 + ], + "angle": 0, + "content": "0.89/3.8" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.199, + 0.901, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.264, + 0.896, + 0.276 + ], + "angle": 0, + "content": "0.78/5.91" + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.28, + 0.86, + 0.293 + ], + "angle": 0, + "content": "Images with spurious grafitti but no freight car" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.294, + 0.581, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.357, + 0.574, + 0.37 + ], + "angle": 0, + "content": "0.97/3.5" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.294, + 0.661, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.357, + 0.654, + 0.37 + ], + "angle": 0, + "content": "0.79/4.0" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.294, + 0.74, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.357, + 0.734, + 0.37 + ], + "angle": 0, + "content": "0.81/2.4" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.294, + 0.821, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.357, + 0.814, + 0.37 + ], + "angle": 0, + "content": "0.85/3.2" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.294, + 0.901, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.357, + 0.895, + 0.37 + ], + "angle": 0, + "content": "0.88/2.6" + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.373, + 0.876, + 0.387 + ], + "angle": 0, + "content": "all classified as freight car by four ImageNet models" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.388, + 0.581, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.452, + 0.574, + 0.464 + ], + "angle": 0, + "content": "0.85/2.5" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.388, + 0.661, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.452, + 0.654, + 0.464 + ], + "angle": 0, + "content": "0.86/2.5" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.388, + 0.74, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.452, + 0.734, + 0.464 + ], + "angle": 0, + "content": "0.90/2.3" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.388, + 0.821, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.452, + 0.814, + 0.464 + ], + "angle": 0, + "content": "0.82/2.2" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.388, + 0.901, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.452, + 0.895, + 0.464 + ], + "angle": 0, + "content": "0.87/2.2" + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.468, + 0.879, + 0.482 + ], + "angle": 0, + "content": "Images with spurious eucalyptus/plants but no koala" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.482, + 0.581, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.547, + 0.574, + 0.559 + ], + "angle": 0, + "content": "0.49/3.5" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.482, + 0.661, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.547, + 0.654, + 0.559 + ], + "angle": 0, + "content": "0.61/3.1" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.482, + 0.74, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.547, + 0.734, + 0.559 + ], + "angle": 0, + "content": "0.36/3.1" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.482, + 0.821, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.547, + 0.813, + 0.559 + ], + "angle": 0, + "content": "0.36/3.1" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.482, + 0.901, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.547, + 0.895, + 0.559 + ], + "angle": 0, + "content": "0.69/2.7" + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.563, + 0.858, + 0.575 + ], + "angle": 0, + "content": "all classified as koala by four ImageNet models" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.577, + 0.581, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.64, + 0.574, + 0.653 + ], + "angle": 0, + "content": "0.56/2.6" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.577, + 0.661, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.64, + 0.654, + 0.653 + ], + "angle": 0, + "content": "0.62/2.5" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.577, + 0.74, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.64, + 0.734, + 0.653 + ], + "angle": 0, + "content": "0.74/2.5" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.577, + 0.821, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.64, + 0.814, + 0.653 + ], + "angle": 0, + "content": "0.72/2.4" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.577, + 0.901, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.64, + 0.895, + 0.653 + ], + "angle": 0, + "content": "0.66/2.2" + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.657, + 0.857, + 0.671 + ], + "angle": 0, + "content": "Images with spurious water jet but no fireboat" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.671, + 0.581, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.735, + 0.574, + 0.747 + ], + "angle": 0, + "content": "0.63/2.3" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.671, + 0.661, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.735, + 0.654, + 0.747 + ], + "angle": 0, + "content": "0.71/2.2" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.671, + 0.74, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.735, + 0.734, + 0.747 + ], + "angle": 0, + "content": "0.84/2.2" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.671, + 0.821, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.735, + 0.813, + 0.747 + ], + "angle": 0, + "content": "0.65/2.1" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.671, + 0.901, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.735, + 0.895, + 0.747 + ], + "angle": 0, + "content": "0.63/1.9" + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.751, + 0.867, + 0.765 + ], + "angle": 0, + "content": "all classified as fireboat by four ImageNet models" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.765, + 0.581, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.829, + 0.574, + 0.842 + ], + "angle": 0, + "content": "0.53/1.9" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.765, + 0.661, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.829, + 0.654, + 0.842 + ], + "angle": 0, + "content": "0.79/1.9" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.765, + 0.74, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.829, + 0.734, + 0.842 + ], + "angle": 0, + "content": "0.78/1.8" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.765, + 0.821, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.829, + 0.814, + 0.842 + ], + "angle": 0, + "content": "0.79/1.8" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.765, + 0.901, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.829, + 0.895, + 0.842 + ], + "angle": 0, + "content": "0.76/1.8" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.856, + 0.895, + 0.917 + ], + "angle": 0, + "content": "Figure 4: Spurious features (ImageNet): found by human labeling of our neural PCA components. For each class we show 5 random train. images (top left), the neural PCA Feature Visual. (NPFV) and 4 most activating train. images for the spurious feature component (bottom left). Right: four ImageNet models classify images showing only the spurious feature but no class object as this class." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20239" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.231 + ], + "angle": 0, + "content": "training. However, this would require relabeling all spurious ImageNet classes which is not feasible. We could order the training set according to the value \\(\\alpha_{l}^{(k)}\\) of the corresponding neural PCA component which indicates how much of the spurious feature an image contains. While this would speed up the process significantly, it would still require a significant amount of manual relabeling. Can one do it also without any additional labeling? Yes, as described in Sec. 4 we can rewrite the logit of the \\(k\\)-th class as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.238, + 0.47, + 0.265 + ], + "angle": 0, + "content": "\\[\nf _ {k} (x) = \\sum_ {l = 1} ^ {D} \\alpha_ {l} ^ {(k)} (x) + \\left\\langle \\mathbf {1}, \\bar {\\psi} _ {k} \\right\\rangle + b _ {k}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.274, + 0.469, + 0.34 + ], + "angle": 0, + "content": "For a spurious component \\(l\\) of class \\(k\\), we use \\(\\min \\{\\alpha_l^{(k)}, 0\\}\\) instead of \\(\\alpha_l^{(k)}\\) to remove its positive contribution from the logit (negative contributions are semantically different). After removal of the spurious features, the new logit becomes" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.348, + 0.469, + 0.381 + ], + "angle": 0, + "content": "\\[\nf _ {k} ^ {S p u F i x} (x) = f _ {k} (x) - \\sum_ {l \\in \\mathcal {S} _ {k}} \\max \\left\\{\\alpha_ {l} ^ {(k)} (x), 0 \\right\\} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.388, + 0.47, + 0.435 + ], + "angle": 0, + "content": "where \\( S_{k} \\) is the set of spurious NPCA components of class \\( k \\). We denote this method as SpuFix. It significantly reduces dependence on spurious features, see Sec. 7.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.473, + 0.683 + ], + "angle": 0, + "content": "Transfer of SpuFix to any ImageNet classifier: As described in Sec. 3, harmful spurious features are a result of subpopulations in the training data. While not every spurious correlation will be picked up by every model, most of our detected spurious features generalize to a wide range of classifiers (see Sec. 7). In the following, we show how SpuFix can be transferred to any given ImageNet classifier \\(\\tilde{f}\\) for which \\(\\tilde{f}_k\\) denotes the logit and \\(\\tilde{\\psi}_k\\) the weighted penultimate layer of class \\(k\\). The goal is to find a direction \\(b\\) in the weighted feature space \\(\\tilde{\\psi}_k\\) of \\(\\tilde{f}\\) for every spurious NPCA component \\(l\\) corresponding to the eigenvector \\(v_l\\) of the original model, resp. \\(\\alpha_l^{(k)}(x)\\), and then truncate its positive component. To find this direction we maximize the covariance of the projection onto \\(b\\) and \\(\\alpha_l^{(k)}\\) over the training images of class \\(k\\):" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.69, + 0.47, + 0.725 + ], + "angle": 0, + "content": "\\[\nb _ {l} ^ {(k)} = \\underset {\\| b \\| _ {2} = 1} {\\operatorname {a r g m a x}} \\sum_ {s \\in I _ {k}} \\left\\langle b, \\tilde {\\psi} _ {k} \\left(x _ {s}\\right) - \\bar {\\tilde {\\psi}} _ {k} \\right\\rangle \\alpha_ {l} ^ {(k)} \\left(x _ {s}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.733, + 0.299, + 0.747 + ], + "angle": 0, + "content": "which has a closed form solution" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.753, + 0.469, + 0.807 + ], + "angle": 0, + "content": "\\[\nb _ {l} ^ {(k)} = \\frac {\\sum_ {s \\in I _ {k}} \\left(\\tilde {\\psi} _ {k} \\left(x _ {s}\\right) - \\bar {\\tilde {\\psi}} _ {k}\\right) \\alpha_ {l} ^ {(k)} \\left(x _ {s}\\right)}{\\left\\| \\sum_ {s \\in I _ {k}} \\left(\\tilde {\\psi} _ {k} \\left(x _ {s}\\right) - \\bar {\\tilde {\\psi}} _ {k}\\right) \\alpha_ {l} ^ {(k)} \\left(x _ {s}\\right) \\right\\| _ {2}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.814, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In contrast to the eigenvectors \\( v_{l} \\), the matched vectors \\( b_{l}^{(k)} \\) are not necessarily orthogonal. Thus before truncation the centered features \\( \\tilde{\\psi}_k(x) - \\bar{\\tilde{\\psi}}_k \\) need to be projected onto the subspace spanned by the \\( b_{l}^{(k)} \\) and represented in the non-orthogonal basis \\( \\{b_{l}^{(k)}\\}_{l\\in S_{k}} \\). We denote this representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.122 + ], + "angle": 0, + "content": "by \\( P^{(k)}(x) \\) (details in C.1). The logit of class \\( k \\) of the SpuFix version of \\( \\tilde{f} \\) is then:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.132, + 0.892, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} _ {k} ^ {S p u F i x} (x) = \\tilde {f} _ {k} (x) - \\sum_ {l \\in \\mathcal {S} _ {k}} \\max \\left\\{\\langle \\mathbf {1}, b _ {l} ^ {(k)} \\rangle P _ {l} ^ {(k)} (x), 0 \\right\\}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.179, + 0.894, + 0.254 + ], + "angle": 0, + "content": "In the case where \\(\\tilde{f} = f\\) (the robust model), we recover the original SpuFix truncation in (8) (see C.2). It turns out that SpuFix is even effective when the architecture is quite different from the ResNet50 we used for detection, e.g. ViT or VOLO, see Sec. 7.3 and Table 1 and 2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.269, + 0.84, + 0.287 + ], + "angle": 0, + "content": "6. Comparison to neural features of [61]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.892, + 0.375 + ], + "angle": 0, + "content": "We compare our NPCA framework to the method of [61] to detect spurious features for ImageNet. As model, they use a \\(\\ell_2\\)-robust ResNet50. Let \\(J_{k}\\) be the set of training images classified as class \\(k\\). [61] define the \\(j\\)-th component \\(m_j^{(k)}\\) of the class-wise mean over predictions," + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.386, + 0.892, + 0.419 + ], + "angle": 0, + "content": "\\[\nm ^ {(k)} = \\frac {1}{\\left| J _ {k} \\right|} \\sum_ {x _ {s} \\in J _ {k}} \\psi_ {k} \\left(x _ {s}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.893, + 0.709 + ], + "angle": 0, + "content": "as the importance of the \\(j\\)-th neuron for class \\(k\\).1 Then, they order the neurons of the penultimate layer according to the score \\(m_j^{(k)}\\) and consider the top-5 neurons of each class. The main difference to our approach is that they assume single neurons with maximal influence on the mean are capturing spurious features whereas our NPCA components are linear combinations of neurons that capture the variance around the mean. Given that the ResNet50 has only 2048 neurons for 1000 classes, some neurons are labeled as core and spurious feature for multiple classes simultaneously, even though the images are quite different. A major advantage of NPCA is that due to the orthogonality of the PCA components, we identify diverse subpopulations in the training data. As [61] use no constraints for the neurons, they often find very similar subpopulations. Hence, one may miss spurious subpopulations when only checking the top-5 components, see Fig. 5. Another difference is that they maximize the score \\(m_j^{(k)}\\) for the training images \\(x_r\\)" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.719, + 0.84, + 0.743 + ], + "angle": 0, + "content": "\\[\nw _ {j} ^ {(k)} = x _ {r} + \\mathrm {a r g m a x} _ {\\| \\delta \\| _ {2} \\leq \\epsilon} m _ {j} ^ {(k)} (x _ {r} + \\delta),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.754, + 0.892, + 0.817 + ], + "angle": 0, + "content": "whereas we maximize our NPCA component \\(\\alpha_l^{(k)}(x)\\) starting from a gray image to introduce no bias. Thus, we check if the classifier produces the spurious feature and not only enhances it on an image showing it already." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.892, + 0.878 + ], + "angle": 0, + "content": "The third difference is that [61] want to identify any spurious feature while our goal is to find harmful ones. Thus, they use weaker criteria for deciding if a neuron shows a spurious feature: main criterion is if the neural activation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.855, + 0.901 + ], + "angle": 0, + "content": "1The top-5 neurons do not change when using \\(I_{k}\\) instead of \\(J_{k}\\)" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20240" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.09, + 0.411, + 0.104 + ], + "angle": 0, + "content": "Top 5 neural PCA components (ours)" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.106, + 0.48, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.369, + 0.471, + 0.446 + ], + "angle": 0, + "content": "Figure 5: Comparison of top-5 NPCA and top-5 neurons for class hummingbird for the robust model of [61]: Our NPCA components identify diverse subpopulations in the training set whereas the top neurons show similar ones and the spurious bird feeder is not detected, see also App. A." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.575 + ], + "angle": 0, + "content": "map based on \\( m_{j}^{(k)} \\) is off the class object according to a majority vote of 5 human labelers. The visualizations \\( w_{j}^{(k)} \\) are only used if the activation maps are inconclusive. In contrast, we require i) our NFPV to show the spurious feature, ii) the GradCAM based on \\( \\alpha_{l}^{(k)} \\) highlights mainly the spurious feature, and iii) our human labelers have to agree that the NPCA component is a harmful spurious feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.471, + 0.833 + ], + "angle": 0, + "content": "As our criteria are more strict (in particular, that the NFPV shows the spurious feature is a strong criterion), it is not surprising that [61] find more spurious features (630 in 357 classes) than we do (322 in 230 classes). Moreover, the employed models are different and we examine top-10 NPCA components whereas they check top-5 neurons. Thus, the comparison is difficult and our novel dataset \"SpuriousImageNet\" could be biased towards spurious features which only we found. Hence, we compute our top-5 NPCA components for their robust ResNet50 for a direct comparison to their top-5 neurons and their found spurious features. In Fig. 5 we compare them for the class hummingbird where they do not find the spurious feature \"bird feeder\". In general, we observe that our found subpopulations are more diverse and thus we find more spurious features than they do when using their weaker criteria. In App. E we do an extensive comparison for all classes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.21, + 0.862 + ], + "angle": 0, + "content": "7. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we provide a qualitative and quantitative evaluation of our 322 detected spurious features in Ima" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "geNet, see Sec. 4. For the quantitative evaluation, we create the dataset \"Spurious ImageNet\", which allows checking the reliance of a given ImageNet classifier on spurious features. We also evaluate our mitigation strategy \"SpuFix\" which does not require additional labels or retraining of the classifier and can be transferred to other image classifiers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.195, + 0.707, + 0.211 + ], + "angle": 0, + "content": "7.1. Qualitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.893, + 0.537 + ], + "angle": 0, + "content": "For the qualitative evaluation, we visualize some of our found 322 spurious features, see Fig. 4. For each class, we show five random training images, the NPFV, and the four most activating training images of the neural PCA component labeled to be spurious. Additionally, we always show ten images which only show the spurious feature but not the actual class e.g., only the bird feeder (spurious) but no hummingbird (class). All ten images are classified as the corresponding class for the robust classifier we have used to compute the NPCA components and three non-robust ImageNet classifiers (ResNext101, Eff.Net B5, ConvNext-B, see also Tab. 2). This shows that our spurious features generalize from the robust classifier to SOTA ImageNet classifiers, indicating that the found spurious features are mainly due to the design of the training set, rather than failures in model training. Our novel validation by collecting real images with the spurious feature but without the class object which are consistently classified as this class directly shows the impact of harmful spurious features and has the advantage that it does not introduce artifacts via masking nor requires expensive pixel-wise segmentations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.893, + 0.84 + ], + "angle": 0, + "content": "Image Collection: The images showing only the spurious feature were obtained by sorting the 9 million images of OpenImages [38] by the value \\(\\alpha_l^{(k)}(x)\\) of the neural PCA component. We check the top 625 retrieved images classified by the robust classifier as the corresponding class if they are all classified as the same class by the additional non-robust classifiers and do not show the corresponding class. This is a quite strict criterion as spurious features can be shared across classes, e.g., twigs for birds, and thus agreement of classifiers is not granted and the images can show the spurious feature and the true class. Nevertheless, this procedure yields between 77 (\"hummingbird\") and 179 (\"freight car\") images of which we show a selection. For \"hummingbird\", a lot of these images show red flowers (without a hummingbird) which makes sense as the NPFV displays features of red flowers and due to the bias that OpenImages does not contain many images of hummingbird feeders. In these cases, we additionally retrieve Flickr images with appropriate text queries e.g., \"hummingbird feeder\" and filter them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Spurious Class Extension: For \"hummingbird\", \"freight car\", and \"koala\" the spurious features significantly extend the predictions beyond the actual class (see Fig. 2). Bird feeders are classified as hummingbirds, graf" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20241" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.094, + 0.885, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.206, + 0.895, + 0.314 + ], + "angle": 0, + "content": "Figure 6: Spurious Score: we plot the AUC of different models for 7 out of 100 classes in \"Spurious ImageNet\" (see Fig 13-15): hummingbird (bird feeder/red flowers), freight car (graffiti), koala (plants/trees), fireboat (water jet), flagpole (US flag), gondola (house/river), and bookshop (storefront). The spurious features for hummingbird, freight car and koala which were not detected in [61, 62] are also spurious for their robust ResNet50 [61]. Training on ImageNet 21k or Fine-tuning from 21k (1kFT21k) decreases dependence on harmful spurious features but classes like flagpole and bookshop remain strongly affected. Our cheap mitigation strategy SpuFix improves the AUCs significantly for both robust ResNet50 but also improves the ViT-B variants trained/fine-tuned on ImageNet1k, especially for the difficult classes flagpole and bookshop." + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.337, + 0.206, + 0.367 + ], + "angle": 0, + "content": "hummingbird (red flowers)" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.368, + 0.206, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.448, + 0.184, + 0.459 + ], + "angle": 0, + "content": "fireboat" + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.464, + 0.191, + 0.476 + ], + "angle": 0, + "content": "(water jet)" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.477, + 0.206, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.338, + 0.308, + 0.367 + ], + "angle": 0, + "content": "freight car (graffiti)" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.368, + 0.322, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.448, + 0.301, + 0.476 + ], + "angle": 0, + "content": "flagpole (US flag)" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.477, + 0.322, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.339, + 0.432, + 0.367 + ], + "angle": 0, + "content": "koala (plants/trees)" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.368, + 0.437, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.448, + 0.431, + 0.476 + ], + "angle": 0, + "content": "gondola (house/river)" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.477, + 0.438, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.607 + ], + "angle": 0, + "content": "Figure 7: Spurious ImageNet: sample images from the dataset for 6 out of 100 classes showing the spurious feature but not the class object, see also App. G." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.75 + ], + "angle": 0, + "content": "fiti as freight cars, and (eucalyptus) plants as koalas. This class extension cannot be detected by monitoring test performance and thus is likely to be noticed only after deployment. For \"hummingbird\", we see in Fig. 4 two images with bees on the bird feeder where \"bee\" is an ImageNet class (also a grasshopper for \"koala\"). Nevertheless, the spurious \"bird feeder\" feature of \"hummingbird\" overrules \"bee\" even though no hummingbird is present." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Spurious Shared Feature: The spurious feature \"water jet\" is shared among the classes \"fireboat\" and \"fountain\". It appears more frequently for \"fireboat\" (see Fig. 2) which leads to an in-distribution shift where now a large number of images of the \"fountain\"-class with a water jet are wrongly classified as \"fireboat\". The spurious feature \"water jet\" for \"fireboat\" has been found also in the Salient ImageNet dataset [61, 62]. However, they did not find spurious features for freight car and koala (in App. A we do a comparison). More examples are in App. F." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.339, + 0.779, + 0.355 + ], + "angle": 0, + "content": "7.2. The Spurious ImageNet dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.577 + ], + "angle": 0, + "content": "A key contribution of this paper is our novel evaluation of spurious features for image classifiers without requiring pixel-wise annotations [48, 59] or having to rely on the validity of neural heatmaps [61]. Instead, we use images from OpenImages to show that images only containing the spurious feature but not the class object are classified as this class. This has the advantage that we consider real images and thus provide a realistic impression of the performance of ImageNet classifiers in the wild. Adding noise [61] or masking [48, 44] image regions requires pixel-wise accurate annotations which are labor-expensive, masking only the object still contains shape information, and using masks avoiding this, e.g., a bounding box around the object, can hide a significant portion of the image which is unrealistic." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.821 + ], + "angle": 0, + "content": "To allow for a quantitative analysis of the influence of spurious features on ImageNet classifiers, we collected images similar to the ones shown to illustrate the spurious features in Fig. 4. The images are chosen such that they show the spurious feature but not the class object. The only difference is that we relax the classification condition and only require two of the four classifiers (robust ResNet50, ResNext101, EfficientNet-B5, ConvNext-B) to predict the corresponding class. We select 100 of our spurious features and for each collect 75 images from the top-ranked images in OpenImages according to the value of \\(\\alpha_{l}^{(k)}\\) for which two human labelers agree that they contain the spurious feature but not class \\(k\\) and two out of four classifiers predict class \\(k\\). We call the dataset Spurious ImageNet as it allows to check the dependence on spurious features with real images for ImageNet classifiers, see Fig. 7 and App. G for samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Spurious Score: A classifier \\( f \\) not relying on the spurious feature should predict a low probability for class \\( k \\) for the Spurious ImageNet samples, especially compared to ImageNet test set images of class \\( k \\). Thus, for each class, we measure the AUC (area under the curve) for the separation" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20242" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "of images with the spurious features but not showing class \\( k \\) versus test set images of class \\( k \\) according to the predicted probability for class \\( k \\). A classifier not depending on the spurious feature should attain a perfect AUC of 1, whereas a value significantly below 1 shows strong reliance. We report the mean AUC (mAUC) over all 100 classes in Tab. 1. All ImageNet models trained only on ImageNet1k are heavily influenced by spurious features. Thus, spurious features are mainly a problem of the training set rather than the classifier, and spurious features found with an adversarially robust model transfer to other ImageNet classifiers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.606 + ], + "angle": 0, + "content": "Pre-training on larger datasets: Some spurious features such as flag (flag pole), bird feeder (hummingbird), and eucalyptus (koala) are classes in ImageNet21k. Therefore, they should no longer be spurious for the other classes. Thus, we test if ImageNet1k-classifiers fine-tuned from an ImageNet21k model are less reliant on spurious features. The results in Tab. 1 and Fig. 6 suggest that the influence of spurious features is damped but they are far from being free of them. To check how much is lost due to fine-tuning we evaluate a ViT-B trained on ImageNet21k which has a mean AUC of 0.931 whereas the fine-tuned model has 0.917. This shows that fine-tuning does not hurt much. While finetuning from ImageNet21k improves the mean AUC, for several classes the dependence on spurious features is still significant, see also Fig. 16 how one has to be careful in the interpretation of higher AUC values. In addition to ImageNet21k, we also evaluate models trained on other large image datasets (JFT-300M[28], YFFC-100M, 1B Instagram[84], MIM[24], LAION-2B and LAION-400M[54]) using self-supervised learning or which are based on CLIP [49]. However, these models also do not achieve better spurious scores (Tab. 1 and Tab. 2). We evaluate a large number of SOTA models in App. D." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.441, + 0.637 + ], + "angle": 0, + "content": "7.3. Evaluation of mitigation technique SpuFix" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Fixing spurious features is a non-trivial task and can require a substantial labeling effort. We evaluate our SpuFix from Sec. 5 that does not require retraining or additional labels. The positive effect of this fix of spurious features (SpuFix) can be seen in Tab. 1 and Fig. 6. Compared to the original robust ResNet50 with a spurious mAUC of 0.630, the SpuFix version has a significantly better spurious mAUC of 0.763. Test set accuracy reduces by \\(0.6\\%\\) but this is a rather positive effect, as several of the additional errors arise since the robust ResNet50 uses spurious features for its decision, e.g., for classes like \"balance beam\" or \"puck\" the class object is often not visible in the cropped test set images. In Table 2 we provide a large scale evaluation of the transfer of SpuFix to SOTA ImageNet models. We observe a consistently better mAUC on Spurious ImageNet, even for very large models fine-tuned from 21k or trained on other large datasets, e.g. SpuFix improves the mAUC of VOLO-" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.088, + 0.9, + 0.43 + ], + "angle": 0, + "content": "
ModelOriginalSpuFix
INet Acc. ↑SpurIN mAUC ↑INet Acc. ↑SpurIN mAUC ↑
ImageNet1k
Rob. ResNet5057.4%0.63056.8%0.763
Rob. ResNet50[61]57.9%0.65157.2%0.764
ConvNeXt-L[41]84.8%0.80384.8%0.819
ViT-B AugReg[66]81.1%0.85081.1%0.859
VOLO-D5 512[85]87.1%0.88287.1%0.907
ImageNet21kFT1k
EfficientNetv2-L[70]86.8%0.89386.8%0.898
ConvNeXt-L[41]87.0%0.91087.0%0.913
ViT-B AugReg[66]86.0%0.91785.9%0.925
BEiT-L\\16[8]88.6%0.92188.6%0.927
LAION-2B
CNeXt-L CLIP 384[49]87.8%0.87987.9%0.884
ViT-L\\14 CLIP[49]88.2%0.91288.2%0.914
MIM
EVA-G\\14 CLIP 560[24]89.8%0.91989.8%0.925
ImageNet21k
ConvNeXt-L[41]-0.943-0.943
ViT-B AugReg[66]-0.931-0.931
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.44, + 0.895, + 0.532 + ], + "angle": 0, + "content": "Table 1: Quantitative Evaluation on Spurious ImageNet: ImageNet classifiers of different training modalities depend on spurious features in varying strength. The mAUC is the mean of AUCs for the separation of images containing the spurious feature but not class \\( k \\) versus test images of class \\( k \\) with the predicted probability of class \\( k \\) as score." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.895, + 0.641 + ], + "angle": 0, + "content": "D5 (87.1% acc.) trained only on 1k by 2.5%, or by 0.6% for EVA-G\\(\\backslash\\)14 CLIP 560 trained on MIM (91.9% acc.), as well as BeiT-L\\(\\backslash\\)16 fine-tuned from 21k (88.6% acc.) by 0.6%. Thus even SOTA models profit from our SpuFix with negligible difference in accuracy (\\(\\leq\\) 0.1%) and thus the use of SpuFix is recommended for any ImageNet model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.656, + 0.62, + 0.672 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.895, + 0.803 + ], + "angle": 0, + "content": "We have shown that large-scale identification of spurious features is feasible with our neural PCA components and neural PCA feature visualizations. With \"Spurious ImageNet\" we introduced a novel dataset to evaluate the dependence of ImageNet classifiers on spurious features based on real images. We demonstrated that our SpuFix method mitigates the dependence on harmful spurious features for any ImageNet classifier without costly labeling or re-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements We acknowledge support by the DFG, Project number 390727645, the Carl Zeiss Foundation, project \"Certification and Foundations of Safe Machine Learning Systems in Healthcare\". The authors thank the IMPRS-IS for supporting YN." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20243" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Julius Adebayo, Michael Muelly, Hal Abelson, and Been Kim. Post hoc explanations may be ineffective for detecting unknown spurious correlation. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.468, + 0.197 + ], + "angle": 0, + "content": "[2] Julius Adebayo, Michael Muelly, Ilaria Liccardi, and Been Kim. Debugging tests for model explanations. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.2, + 0.468, + 0.24 + ], + "angle": 0, + "content": "[3] D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Mane. Concrete problems in AI safety. arXiv:1606.06565v2, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.242, + 0.468, + 0.31 + ], + "angle": 0, + "content": "[4] Christopher J. Anders, Leander Weber, David Neumann, Wojciech Samek, Klaus-Robert Müller, and Sebastian Lapuschkin. Finding and removing clever hans: Using explanation methods to debug and improve deep models. Information Fusion, 77:261-295, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.312, + 0.468, + 0.352 + ], + "angle": 0, + "content": "[5] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.354, + 0.468, + 0.395 + ], + "angle": 0, + "content": "[6] Maximilian Augustin, Alexander Meinke, and Matthias Hein. Adversarial robustness on in- and out-distribution improves explainability. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.397, + 0.468, + 0.45 + ], + "angle": 0, + "content": "[7] David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja Hansen, and Klaus-Robert Müller. How to explain individual classification decisions. JMLR, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.453, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[8] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.481, + 0.468, + 0.508 + ], + "angle": 0, + "content": "[9] Sara Beery, Grant van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.51, + 0.468, + 0.55 + ], + "angle": 0, + "content": "[10] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with ImageNet? arXiv:2006.07159, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.468, + 0.592 + ], + "angle": 0, + "content": "[11] B. Biggio, I. Corona, D. Maiorca, B. Nelson, N. Srndic, P. Laskov, G. Giacinto, and F. Roli. Evasion attacks against machine learning at test time. In ECML/PPKD, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.468, + 0.634 + ], + "angle": 0, + "content": "[12] Alceu Bissoto, Eduardo Valle, and Sandra Avila. Debiasing skin lesion datasets and models? not so fast. arXiv:2004.11457, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.468, + 0.677 + ], + "angle": 0, + "content": "[13] Valentyn Boreiko, Maximilian Augustin, Francesco Croce, Philipp Berens, and Matthias Hein. Sparse visual counterfactual explanations in image space. In \\(GCPR\\), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[14] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[15] Myung Jin Choi, Antonio Torralba, and Alan S. Willsky. Context models and out-of-context objects. Pattern Recognition Letters, 33(7):853-862, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[16] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR workshop, pages 702-703, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[18] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[19] Ann-Kathrin Dombrowski, Maximillian Alber, Christopher Anders, Marcel Ackermann, Klaus-Robert Müller, and Pan Kessel. Explanations can be manipulated and geometry is to blame. In NeurIPS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[21] Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. Adversarial robustness as a prior for learned representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[22] Christian Etmann, Sebastian Lunz, Peter Maass, and Carola-Bibiane Schonlieb. On the connection between adversarial robustness and saliency map interpretability. In ICML, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[23] Sabri Eyuboglu, Maya Varma, Khaled Saab, Jean-Benoit Delbrouck, Christopher Lee-Messer, Jared Dunnmon, James Zou, and Christopher Ré. Domino: Discovering systematic errors with cross-modal embeddings. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[24] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.517, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[25] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[26] R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.631, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[27] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In AAAI, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[28] A Gupta, C Sun, A Shrivastava, and S Singh. Revisiting the unreasonable effectiveness of data. arXiv preprint arXiv:1707.02968, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[29] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[30] M. Hein, M. Andriushchenko, and J. Bitterwolf. Why ReLU networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[31] D. Hendrycks and T. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[32] D. Hendrycks and K. Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20244" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[34] Sara Hooker, Yann Dauphin, Aaron Courville, and Andrea Frome. Selective brain damage: Measuring the disparate impact of model pruning. In ICLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.163 + ], + "angle": 0, + "content": "[35] Pieter Abbeel Jonathan Ho, Ajay Jain. Denoising diffusion probabilistic models. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.165, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[36] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[37] A. Kurakin, I. J. Goodfellow, and S. Bengio. Adversarial examples in the physical world. In *ICLR Workshop*, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[38] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[39] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.447, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[41] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.49, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[42] Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[43] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[44] Mazda Moayeri, Sahil Singla, and Soheil Feizi. Hard ImageNet: Segmentations for objects with strong spurious cues. In NeurIPS Datasets and Benchmarks Track, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.604, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[45] A. Nguyen, J. Yosinski, and J. Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In CVPR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.648, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[46] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.69, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[47] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. Proc. CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.733, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[48] Gregory Plumb, Marco Tulio Ribeiro, and Ameet Talwalkar. Finding and fixing spurious patterns with explanations. Transactions on Machine Learning Research (TMLR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[50] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In ICML, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[51] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\": Explaining the predictions of any classifier. In KDD, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[52] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. License: No license specified." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[53] Shibani Santurkar, Dimitris Tsipras, Brandon Tran, Andrew Ilyas, Logan Engstrom, and Aleksander Madry. Image synthesis with a single (robust) classifier. In NeurIPS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[54] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[56] Kathryn Schutte, Olivier Moindrot, Paul Hérent, Jean-Baptiste Schiratti, and Simon Jégou. Using stylegan for visual interpretability of deep learning models on medical images. In NeurIPS Workshop, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[57] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[58] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[59] Rakshith Shetty, Bernt Schiele, and Mario Fritz. Not using the car to see the sidewalk-quantifying and controlling the effects of context in classification and segmentation. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[60] Krishna Kumar Singh, Dhruv Mahajan, Kristen Grauman, Yong Jae Lee, Matt Feiszli, and Deepti Ghadiyaram. Don't judge an object by its context: learning to overcome contextual bias. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[61] Sahil Singla and Soheil Feizi. Salient imagenet: How to discover spurious features in deep learning? In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[62] Sahil Singla, Mazda Moayeri, and Soheil Feizi. Core risk minimization using salient imagenet. arXiv:2203.15566, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[63] Sahil Singla, Besmira Nushi, Shital Shah, Ece Kamar, and Eric Horvitz. Understanding failures of deep networks via robust feature extraction. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[64] Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju. Fooling lime and shap: Adversarial attacks on post hoc explanation methods. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 180-186, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[65] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20245" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[66] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. In TMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[67] Pierre Stock and Moustapha Cisse. Convnets and ImageNet beyond accuracy: Understanding mistakes and uncovering biases. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.47, + 0.235 + ], + "angle": 0, + "content": "[68] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.264 + ], + "angle": 0, + "content": "[69] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.47, + 0.294 + ], + "angle": 0, + "content": "[70] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.47, + 0.323 + ], + "angle": 0, + "content": "[71] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR, pages 1521-1528, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.47, + 0.353 + ], + "angle": 0, + "content": "[72] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[73] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. From ImageNet to image classification: Contextualizing progress on benchmarks. In ICML, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[74] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[75] Sahil Verma, John P. Dickerson, and Keegan Hines. Counterfactual explanations for machine learning: A review. arXiv:2010.10596, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.499, + 0.47, + 0.553 + ], + "angle": 0, + "content": "[76] Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the GDPR. Harvard Journal of Law & Technology, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.47, + 0.584 + ], + "angle": 0, + "content": "[77] Ross Wightman. Pytorch image models. https://github.com/ rwightman/pytorch-image-models, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.47, + 0.626 + ], + "angle": 0, + "content": "[78] Eric Wong, Shibani Santurkar, and Aleksander Madry. Leveraging sparse linear layers for debuggable deep networks. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.47, + 0.684 + ], + "angle": 0, + "content": "[79] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.47, + 0.727 + ], + "angle": 0, + "content": "[80] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[81] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[82] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[83] Cong Xu and Min Yang. Adversarial momentum-contrastive pre-training. arXiv preprint, arXiv:2012.13154v2, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[84] I Zeki Yalniz, Herve Jégou, Kan Chen, Manohar Paluri, and Dhruv Mahajan. Billion-scale semi-supervised learning for image classification. arXiv preprint arXiv:1905.00546, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[85] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision observer for visual recognition. PAMI, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[86] John R. Zech, Marcus A. Badgeley, Manway Liu, Anthony B. Costa, Joseph J. Titano, and Eric Karl Oermann. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLOS Medicine, 15(11):1-17, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[87] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[88] Zhuotun Zhu, Lingxi Xie, and Alan L. Yuille. Object recognition with and without objects. In *IJCAI*, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20246" + } + ] +] \ No newline at end of file diff --git a/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/8d4ee68a-1062-4c0b-81d9-4f01bf36823f_origin.pdf b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/8d4ee68a-1062-4c0b-81d9-4f01bf36823f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c41746e0f0fe13525359e18cc7cf447ba07d5369 --- /dev/null +++ b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/8d4ee68a-1062-4c0b-81d9-4f01bf36823f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30684cc89a6e7ed7f4a8379bb572e69d1a56c936e433d62c699402e263a7d9b7 +size 1781668 diff --git a/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/full.md b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/full.md new file mode 100644 index 0000000000000000000000000000000000000000..af978ba225f4edca16b284854fa6955b850e6cbc --- /dev/null +++ b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/full.md @@ -0,0 +1,675 @@ +# Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet + +Yannic Neuhaus + +Maximilian Augustin + +Valentyn Boreiko + +Matthias Hein + +Tübingen AI Center - University of Tübingen + +# Abstract + +Benchmark performance of deep learning classifiers alone is not a reliable predictor for the performance of a deployed model. In particular, if the image classifier has picked up spurious features in the training data, its predictions can fail in unexpected ways. In this paper, we develop a framework that allows us to systematically identify spurious features in large datasets like ImageNet. It is based on our neural PCA components and their visualization. Previous work on spurious features often operates in toy settings or requires costly pixel-wise annotations. In contrast, we work with ImageNet and validate our results by showing that presence of the harmful spurious feature of a class alone is sufficient to trigger the prediction of that class. We introduce the novel dataset "Spurious ImageNet" which allows to measure the reliance of any ImageNet classifier on harmful spurious features. Moreover, we introduce SpuFix as a simple mitigation method to reduce the dependence of any ImageNet classifier on previously identified harmful spurious features without requiring additional labels or retraining of the model. We provide code and data at https://github.com/YanNeu/spuriousImagenet. + +# 1. Introduction + +Deep learning has led to tremendous progress in image classification [37, 50] and natural language processing [14]. Over the years, however, it has become apparent, that evaluating predictive performance on a fixed test set is not necessarily indicative of the performance when image classifiers are deployed in the wild. Several potential failure cases have been discovered. This starts with a lack of robustness due to image corruptions [31], adversarial perturbations [68], and arbitrary predictions on out-of-distribution inputs [45, 32, 30]. In this paper, we consider the problem of identifying and debugging image classifiers from spurious features [2]. Spurious features in image classification are features that co-occur with the actual class object and are picked up by the classifier. In the worst case, they lead to shortcut learning [25], where only the spurious but not the + +Spurious Features in Training Data + +bird feeder + +graffiti + +eucalyptus + +label + +![](images/7e837ae86e2d771d5286c9eb2a22c0f9fbc3d3a30bd4fa5dc8d07eff2f74a74d.jpg) + +![](images/d7ab984c8c6ee8f1e9e913a5c4b7eb2342d45b5dba6fd82527e0b400521da1f7.jpg) + +![](images/d06c41754196ab95ce2b2602d999b7e23aaaf3eec95d35183a2d02b233c11b52.jpg) + +![](images/6bb75441b7caa1ae1fdb35a463ae277a98cbca646e9806df92900c645f19b6d8.jpg) + +Hummingbird + +Freight Car + +Koala + +Hard Disc + +Images from the web with spurious feature but no class features classified as class below + +![](images/f3fe0b619236e6498566704a9b635cb5dc304dc36671ec78be38991bb951c99e.jpg) + +![](images/9555789428e3d27b20db3b5883a60a98a52c9698b51351a2a852c4d1bf7ceec9.jpg) + +![](images/a15bba2a618e26b41249dcfa87a907ebacbb4b5627cd067b1fde31a871b1ce8d.jpg) + +![](images/963df71285c03eaa75050a0317ab26178288277db62bcdca8efecbf19517f2a9.jpg) + +![](images/d050b9811842346325d13ed718729889cd50fa44f41bf4d6611ca5caad206e79.jpg) +Figure 1: Top: Examples of spurious features found via our neural PCA components but not in previous study [61]. Bottom: We validate our spurious features by mining images from the web showing only the spurious feature but not the class. They are classified by four ImageNet models as the corresponding class. Some of them even contain ImageNet classes (bees on feeder, grasshopper in leaves). + +![](images/10e87f53c103305227963b7100d97c882846e1a51d6f6f2f4a213bd99712bb54.jpg) + +![](images/86c91affd6b9f248ab60af7e0f3d2620e962cb443f18e5c5718e3a37d97c62d5.jpg) + +![](images/00e18b10c73c49299fd1881ca4e8a1bb66343739cd0c083d2a8cb4690e4a7429.jpg) + +Hummingbird + +Freight Car + +Koala + +Hard Disc + +correct feature is associated with the class, e.g., [86] found that a pneumonia detector's bad generalization across hospitals was caused by the neural network learning to identify the hospital where the training data originated from. A weaker form of spurious feature (at least from a learning perspective) is the case when the classifier picks up the correct class features, e.g., of a hummingbird, but additionally associates a spurious feature, e.g., a bird feeder, with the class as they appear together on a subset of the training set. This becomes a harmful spurious feature if only the spurious feature without the class feature is sufficient to trigger the classification of that class, see Fig. 1 for an illustration of + +such spurious features found via our method. Harmful spurious features are difficult to detect and thus can easily go unnoticed, leading to unexpected behaviour of a deployed image classifier. + +In this paper we make the following key contributions: + +- we develop a pipeline for the detection of harmful spurious features with little human supervision based on our class-wise neural PCA (NPCA) components of an adversarially robust classifier together with their Neural PCA Feature Visualization (NPFV). +- unlike prior work, which used masking images or pixelwise annotations, we validate our found spurious features by using our NPCA components to find real images containing only the spurious feature but not the class object. +- using these images we create the dataset "Spurious ImageNet" and propose a measure for dependence on spurious features. We do a large-scale evaluation of state-of-the-art (SOTA) ImageNet models. We show that the spurious features found for the robust model generalize to non-robust classifiers. Moreover, we analyze the influence of different training setups, e.g. pre-training on ImageNet21k or larger datasets like LAION. +- we develop SpuFix, a technique to mitigate the dependence on identified harmful spurious features without requiring new labels or retraining, and show how to transfer it to any ImageNet classifier. SpuFix consistently improves the dependence on harmful spurious features even for SOTA models with negligible impact on test accuracy. + +# 2. Related work + +When classifiers in safety-critical systems such as healthcare or autonomous driving are deployed in the wild [3], it is important to discover potential failure cases before release. Prior work has focused on corruption [31], adversarial robustness [11, 68, 43], and out-of-distribution detection [45, 32, 30]. There is less work on spurious features, although their potential harm might be higher. + +Spurious features: It has been noted early on that classifiers show reliance on spurious features [15] e.g., a cow on the beach is not recognized [9] due to the missing spurious feature of grass. Other forms of spurious features have been reported in the classification of skin lesions [12], pneumonia [86], traffic signs [67], and object recognition [88]. Moreover, it has been shown that deep neural networks are biased towards texture [26] and background context [80], see [25] for an overview. [58] argues theoretically that spurious features are picked up due to a simplicity bias. + +Detection of spurious features has been achieved using human label-intense pixel-wise annotations [48, 59, 60]. In [78], they use sparsity regularization to enforce a more interpretable model and use it for finding spurious features. [4] propose a complex pipeline to detect spurious features. While they scale to ImageNet, their analysis is lim + +ited to a few spurious features for a subset of 100 classes. [61, 44, 62] do a search on full ImageNet based on class-weighted "neural maps". The neural maps are used to add noise to "spurious" resp. "core" features but no significant difference in classification performance is observed. It remains unclear if their found spurious features are harmful, that is the feature alone triggers the decision for that class. + +Interpretability methods: In recent years several interpretability methods have been proposed e.g., attribution methods such as GradCAM [57], Shapley values [42], Relevance Propagation [7], and LIME [51]. The use of these methods for the detection of spurious features has been analyzed in [2, 1] with mixed success and it has been argued that interpretability methods are not robust [19, 64, 27]. However, attribution methods work better for robust classifiers due to more interpretable gradients [22]. Another technique is counterfactual explanations [76, 75] which are difficult to generate for images due to the similarity to adversarial examples [68]. Thus visual counterfactual explanations are realized via manipulation of a latent space [56] or in image space [53, 6, 13] for an adversarily robust classifier. Visual counterfactuals for non-robust classifiers using diffusion models [35, 65, 33, 18] have been proposed in [5]. ImageNet: ImageNet [52] suffers from several shortcomings: apart from an inherent dataset bias [71], semantically overlapping or even identical class pairs were reported [34, 73, 10], e.g., two classes "maillot", "sunglass" vs "sun-glasses", "notebook" vs "laptop" etc. We disregard such trivial cases of dataset contamination and focus on classes with harmful spurious features, in particular ones where only a small portion of the training set is contaminated. + +# 3. Spurious features + +A proper definition of spurious features is difficult. We describe two settings of harmful spurious features which appear in this paper. We denote by $C_k$ the set of all images containing objects belonging to class $k$ (assuming for simplicity that we have a deterministic problem and ignoring multi-labels). Let $S$ be the set of all images containing a feature $s$ (e.g., a bird feeder). It is a correlated feature for class $k$ when $C_k \cap S$ and $S \setminus C_k$ are non-empty, i.e., the feature occurs frequently with the class object but there is no causal implication that appearance of $s$ implies the appearance of the class object (a bird feeder in the image does not imply presence of a hummingbird). A correlated feature becomes a spurious feature when the classifier picks it up as feature of this class. Not every spurious feature is immediately harmful, even humans use context information [25] to get more confident in a decision. However, a spurious feature is harmful if the spurious feature alone is enough to trigger the decision for the corresponding class without the class object being present in the image. We consider two scenarios for a harmful spurious feature shown in Fig. 2. + +![](images/1c3e3e4aa0171ff564873305b533c9a4932d01f6cbdcd10514a58794c49e17e3.jpg) +Figure 2: Type of harmful Spurious Feature: Left: The spurious feature $s$ is taken up by the classifier which predicts class $k$ on $C_k \cup S$ instead only on $C_k$ . Right: The spurious feature $s$ is shared between classes but appears more often in class $k$ . The classifier associates $S$ with class $k$ and thus predicts class $k$ also on $C_l \cap S$ instead of class $l$ . + +Spurious Class Extension: For this type of spurious feature (left in Fig. 2) the classifier picks up the spurious feature $s$ for class $k$ and predicts the class $k$ even on $S \setminus C_k$ with high confidence (prediction of "hummingbird" for images showing a bird feeder but no hummingbird). The classifier predicts class $k$ beyond its actual domain $C_k$ and thus we call this a spurious class extension. While this spurious feature does not necessarily hurt in terms of test performance it can easily lead to completely unexpected behavior when the classifier is deployed in the wild. + +Spurious Shared Feature: Here, two classes $C_k$ and $C_l$ share a spurious feature $s$ (e.g., "water jet" for the classes "fireboat" and "fountain"). As there are more training images with feature $s$ in $C_k$ than in $C_l$ the classifier associates $S$ with class $k$ and predicts class $k$ for $S \cap C_l$ . + +The two types of harmful spurious features are not exclusive. A shared spurious feature $s$ can at the same time lead to a spurious class extension, e.g., the object bird feeder leads to a spurious class extension of the class "hummingbird" (see Fig. 4) to images of bird feeders without hummingbirds. In the training set the hummingbird feeder appears only in images of class "hummingbird" but the hummingbird feeder has parts which mimic flowers and flowers are a shared spurious feature with bees. In Fig. 4 right top row, images of bees on a bird feeder are classified as "hummingbird" instead of "bee", so the spurious feature is strong enough to override the decision for the true class "bee" (spurious shared feature). + +# 4. Finding spurious features via neural PCA and associated feature visualizations + +First, we define our class-wise neural PCA (NPCA) which allows us to find diverse subpopulations in the train- + +ing data, e.g., we checked that the bird feeder for "hummingbird" is visible in $15\%$ of the training images (component 2), while another $15\%$ contain a part of it (component 3), see Fig. 3 or, for more examples, App. F. Then we introduce our neural PCA feature visualization (which requires an adversarially robust model) and how we select NPCA components for human inspection. The identification of spurious features requires minimal human supervision, and our effective setup allows us to screen all ImageNet classes. + +Adversarily robust model: Similar to [61], we use an adversarily robust model to find spurious features in ImageNet. The reason for this is that robust models have generative properties [74, 53, 6, 61, 13] in the sense that maximizing the predicted probability of a class in a neighborhood of an image leads to semantically meaningful changes. They also have more informative gradients [22] and thus attribution maps such as GradCAM [57] work better. We use the multiple-norm robust model of [13] as they claim it has the best generative properties. The generative properties of robust models are mainly used for the neural PCA feature visualization where we maximize the NPCA component of a class starting from a gray image. If a spurious feature appears without the class object, this is a strong indicator of a harmful spurious feature. A non-robust model would only produce semantically meaningless adversarial noise, hence we need a robust model for this part of our detection pipeline. Our detected spurious features are not specific to the robust model. We show that SOTA ImageNet models share the same spurious features (Fig. 4 and Sec. 7.2). + +Class-wise neural PCA: Let $(x_{i},y_{i})_{i = 1}^{N}$ be the training set, where $y_{i}\in \{1,\dots ,K\}$ and $K$ is the number of classes. We consider features of the penultimate layer $\phi (x)\in \mathbb{R}^D$ of a neural network for an input $x$ . For a given class $k$ and its associated weights $w_{k}\in \mathbb{R}^{D}$ in the final layer, we define + +$$ +\psi_ {k} (x) = w _ {k} \odot \phi (x). \tag {1} +$$ + +where $\odot$ is the componentwise product. Let $b\in \mathbb{R}^K$ be the bias vector of the final layer then the logit $f_{k}$ of class $k$ is: + +$$ +f _ {k} (x) = \sum_ {j = 1} ^ {D} w _ {k j} \phi (x) _ {j} + b _ {k} = \langle \mathbf {1}, \psi_ {k} (x) \rangle + b _ {k}. \tag {2} +$$ + +Let $I_{k}$ be the index set of the training set of class $k$ and $\bar{\psi}_k$ the class-wise mean, $\bar{\psi}_k = \frac{1}{|I_k|}\sum_{s\in I_k}\psi_k(x_s)$ . The class-wise neural PCA allows us to identify variations in the set $\{\psi (x_r)\}_{r\in I_k}$ arising due to small subpopulations in the training set. In the class-wise neural PCA, we compute eigenvectors of the class-wise covariance matrix, + +$$ +C = \sum_ {s \in I _ {k}} \left(\psi_ {k} \left(x _ {s}\right) - \bar {\psi} _ {k}\right) \left(\psi_ {k} \left(x _ {s}\right) - \bar {\psi} _ {k}\right) ^ {T}. \tag {3} +$$ + +The eigenvectors, $v_{1},\ldots ,v_{D}$ form an orthonormal basis of $\mathbb{R}^D$ and we write $\psi_k(x) - \bar{\psi}_k$ in this basis, + +$$ +\psi_ {k} (x) - \bar {\psi} _ {k} = \sum_ {l = 1} ^ {D} v _ {l} \left\langle \psi_ {k} (x) - \bar {\psi} _ {k}, v _ {l} \right\rangle , \tag {4} +$$ + +![](images/28e7dffbbe321d2281ad33a08966585590676aa800f49524184302807e2e3714.jpg) +Top 5 neural PCA components (ours) +NNE + +![](images/5e7a194e1a7bc8b42a74bddc476398b68002e531e40dc608a4776ecce579427c.jpg) +Max. act. + +![](images/4d7988fe9f7416a4b9e2570316c78aa88f91ffed005e878461d29d50ad061318.jpg) +Heeannnnnne +Figure 3: Top 5 Neural PCA components for class hummingbird: first row shows our neural PCA feature visualization (NPFV), second row shows four most activating training images of each NPCA component and the last row GradCAM for the NPCA component. Our NPCA components capture different subpopulations in the training data. Comp. 2 is identified as spurious feature "bird feeder", see NPFV and most activating training images (see also Fig. 4). + +and define + +$$ +\alpha_ {l} ^ {(k)} (x) = \left\langle \mathbf {1}, v _ {l} \right\rangle \left\langle \psi_ {k} (x) - \bar {\psi} _ {k}, v _ {l} \right\rangle , \tag {5} +$$ + +The logit $f_{k}(x)$ of the $k$ -th class can then be written as + +$$ +f _ {k} (x) = \sum_ {l = 1} ^ {D} \alpha_ {l} ^ {(k)} (x) + \left\langle \mathbf {1}, \bar {\psi} _ {k} \right\rangle + b _ {k}. \tag {6} +$$ + +Thus for a given $x$ , we can interpret $\alpha_l^{(k)}(x)$ as the contribution of the neural PCA component $l$ of class $k$ to the logit $f_k(x)$ of class $k$ since the term $\langle \mathbf{1}, \bar{\psi}_k \rangle + b_k$ is constant for all inputs. Based on this we introduce a mitigation technique for spurious features without retraining in Sec. 5. + +Neural PCA Feature Visualization: To identify semantic features corresponding to our neural PCA component $l$ , we show the training images which attain the maximal values of $\alpha_{l}^{(k)}(x)$ . Additionally, we generate an image $z_{l}^{(k)}$ , which we call the Neural PCA Feature Visualization (NPFV) of feature $l$ of class $k$ , by maximizing $\alpha_{l}^{(k)}(x)$ : + +$$ +z _ {l} ^ {(k)} = g + \operatorname * {a r g m a x} _ {\| \delta \| _ {2} \leq \epsilon} \alpha_ {l} ^ {(k)} (g + \delta), +$$ + +where $g$ is a gray image (all channels equal to 0.5). Thus we maximize the feature $\alpha_{l}^{(k)}$ outgoing from a non-informative and unbiased initialization $g$ . The optimization problem is solved using adaptive projected gradient descent (APGD) [16] with 200 steps. The budget for changes, $\epsilon = 30$ , is small to avoid the overactivation of feature attacks maximizing the output of individual neurons [21, 63, 61], see + +Fig. 8. In Fig. 4 we show for each identified spurious feature, the corresponding NPFV, e.g., for "hummingbird" one can see the bird feeder but no hummingbird. The NPCA components together with the maximally activating training images are in principle sufficient to identify spurious features, but the NPFV is very useful to judge how harmful a spurious feature is. Therefore, we use an adversarily robust model, since a non-robust model would yield semantically meaningless adversarial noise as NPFV. + +Selection of neural PCA components for human inspection: The penultimate layer of the robust ResNet50 we are using has 2048 neurons. Thus it is infeasible (and unnecessary) to investigate all neural PCA components. A strong criterion that one has found a harmful spurious feature is i) the NPFV shows mainly the spurious feature and not the class, and ii) the NPFV has high confidence. If ii) is not satisfied, then the NPFV is a spurious feature the classifier may have picked up, but it is not harmful in the sense that this feature alone causes the classifier to choose that class. Moreover, we noticed that the eigenvalues of the neural PCA, and the corresponding $\alpha$ values, decay quickly. Thus we compute the NPFV for the top 128 neural PCA components (having maximal variance) and then select the ten components which realize the highest confidence for their NPFV in the corresponding class. Note that we do not optimize the confidence when generating the NPFV but only $\alpha_{l}^{(k)}(x)$ which is part of the logit of the $k$ -th class. + +Identification of spurious neural PCA components via human supervision: For each ImageNet class $k$ we show the human labeler the top 10 components. For each component $l$ we show the NPFV $z_{l}^{(k)}$ and the 5 training images $x_{r}$ of class $k$ with the largest values of $\alpha_{l}^{(k)}(x_{r})$ . Moreover, we compute GradCAM [57] images for the NPFV and the five training images using the NPCA component $\alpha_{l}^{(k)}$ as score. The human marks a component as spurious if i) the NPFV shows dominantly an object not belonging to the class ii) the five training images show consistently this object, iii) the GradCAM activations are primarily not on the class object. The setup shown to the human labeler can be seen in App. B. The labeling of one class takes on average about 45 seconds, so the full labeling of all ImageNet classes took about 13 hours. The human labeler (researcher in machine learning) found in total 337 spurious components. Another human labeler checked all of them and removed spurious features in case of disagreement, resulting in 322 spurious features from 230 ImageNet classes. + +# 5. SpuFix - Mitigation of spurious features + +Once the spurious features are identified, the question is how one can mitigate that the classifier relies on them. One way is to identify the training images containing the spurious feature and then discard or downweight them during + +![](images/b14bf69b497362b8af6a3e938c5160b68291741a643fcbb5b3878d8411f4b9c6.jpg) +Hummingbird - Random train. images (confidence / $\alpha_{k}$ ) +0.93/1.7 + +![](images/bc0b28bfc916a5957df536af5be3e4e59f8da09d64090eb1ce52be150f1eefc1.jpg) +1.00/-0.9 0.96/-1.0 + +![](images/42061f85986488c370fe1403148bc244f77aae006c86035a5278b75e93aec6f1.jpg) + +![](images/564e85099f5fcb75517ed03ec10f9e2ac1281b1b7e7912168419afa7c82aa0cc.jpg) +Max. activating train. images - NPCA Comp. 2 +1.00/5.9 + +![](images/0bf1c17e1415b8b5a086a0934782e84692922b518147e332cea369bafebf3ab8.jpg) +0.99/2.2 1.00/1.54 + +![](images/73649e84b888d8a170558282f72b9119f8421478e23746fb0ce9ef2c06dbe182.jpg) + +1.00/1.54 + +A Comp. 2 + +![](images/ee9abd0b3cf89ae4ab7540cd4d1cd641be18143c2c94b6ad93311dc845855134.jpg) +NPFV-2 +1.00/9.7 + +![](images/55e73764a49276f04235ea41a34d73c5ff3ea4a2234cfa7f10a29b5f3038a146.jpg) +1.00/7.5 + +1.00/5.9 + +1.00/5.6 + +![](images/3d2ff902933bd3e2b9f8becf427f2361c4641ab084f16473623a2518546840ba.jpg) +1.00/5.6 +1.00/5.6 + +![](images/4e979465e870cbdf4b8c413f94e50d01f2cc1144bf0a83b1fa561a53ba386bd3.jpg) +Freight car - Random train. images (confidence / $\alpha_{k}$ ) +1.00/7.7 + +![](images/9abcadc71d9090855e26081cec5d5e52e7d20d3a6f1164ae18f32d15e63b0936.jpg) +1.00/3.4 + +![](images/bfab20a774e62a2065a524bccb6bbb1005463907e4162fdbf08d4e3b7c979fdf.jpg) +1.00/6.3 + +![](images/c9d9314d3941c262721f2249d8ba298e2cd439a066dab189d9e00c9f2f64d789.jpg) +0.98/-0.8 + +![](images/d39e205b71a0709e2c941d21fbb09aaaea7fe6bfacd5ee5685853c2f541637cf.jpg) +1.00/4.2 + +![](images/b37ebabe2a3afb713344d4567307ea6abea8d8e96a9f3b6e78002694ff20e621.jpg) +NPFV-1 +1.00/12.1 + +![](images/fcf21de7695b8da5ddc4d84e03a64bb687d7a503f9ce1ca22201048369cd0429.jpg) +Max. activating train. images - NPCA Comp. 1 +1.00/10.9 + +![](images/23f8319e0d23f30392514d7c469a3d05e68f2dec9dadd3f632a11759cf32069e.jpg) +1.00/10.4 + +![](images/106f86f1b02a225d4b39bb2158a317c5f3d170608b0545500c50bae7bbb27d5a.jpg) +1.00/10.2 + +![](images/b9b526175825f8302e0707717326217793a880498bb8ae8553e22713ebedd99d.jpg) +1.00/10.2 + +![](images/2542e369ce891f1da3a82d5484e8df2745e02c6b49db9f61e0f2089dd515db28.jpg) +Koala - Random train. images (confidence / $\alpha_{k}$ ) +1.00/0.77 + +![](images/099c91b963703a915f76eda313baf1f8e8a46cdf32e90a5e52ef889e16630a38.jpg) +0.87/2.4 + +![](images/43de58373554e8da48d3e2bc2058799b89fb568a5016021084561b503b60e9e7.jpg) +1.00/0.4 + +![](images/7b2a0498eabd530d3c6f7b82940e61a1a2509f790afa1a19cb7a5a0996bcd07b.jpg) +1.00/0.0 + +![](images/38c8d0820678ff378e6b0cd711b93ea5ac9053e887f130c0f3bbd8bcadfeeb3e.jpg) +0.95/0.5 + +![](images/7e0f88858e23beed70e24a8579ae8f549b6b8e5498604d4e5c3464bc1a3ccccd.jpg) +NPFV-3 +1.00/5.5 + +![](images/85a9cf365a5d70f37b0937c99fcfd657361c01785806145e0dd65ecb21021fc8.jpg) +Max. activating train. images - NPCA Comp. 3 +1.00/4.6 + +![](images/aabeef1685499126ba04e6f2a069bc3f04d0fd96b00998392151e6971559cc56.jpg) +1.00/4.5 + +![](images/f37200399a4dd0a1a1bee69188f721d122733fa978bd68c0f3a9f389b0ee62f0.jpg) +1.00/4.4 + +![](images/f587b437982882523d294ee93ba63294ec6cc9a6a21d34c28513b19fc46115a1.jpg) +0.86/4.3 + +![](images/28d10134535d18af7dcf2d59ba7e5d842b09057d7df5bfec65eb4113412512d1.jpg) +Fireboat - Random train. images (confidence / $\alpha_{k}$ ) +0.88/-1.1 + +![](images/f2ca3287a934247402b3047e7345d2bccbe2defbbc945ea230deb08ee3c3b73c.jpg) +0.95/0.5 + +![](images/10f679df621f8554213f997271b63ad8b6fcf08b6b7632967059e16b4f1d6cbd.jpg) +0.12/-0.6 + +![](images/8973b07c4a97153d5e9146d29f562ef8abf618d7845386b5b77d31e6f66de1b7.jpg) +0.84/-0.2 + +![](images/36f8cc08837596bca683088b426cd5f0b7b2e11f5e95ee1a3b301586c2d03acf.jpg) +0.02/-1.1 + +![](images/fa054c68d416f25b599c682b3410ff02d5cf5b94ff1a664d6c40d47704dad8e1.jpg) +NPFV-2 +1.00/5.5 +Figure 4: Spurious features (ImageNet): found by human labeling of our neural PCA components. For each class we show 5 random train. images (top left), the neural PCA Feature Visual. (NPFV) and 4 most activating train. images for the spurious feature component (bottom left). Right: four ImageNet models classify images showing only the spurious feature but no class object as this class. + +![](images/91890ab85f24c0fe4ff3e4d87cf6586c12f238c95bf345d5210ec281cde8c3e4.jpg) +Max. activating train. images - NPCA Comp. 2 +0.42/4.1 + +![](images/2d85f28e5420f90157fa94845c66fd9fb4d4e510ff66fc405576d654d0a3909a.jpg) +1.00/4.0 + +![](images/93e8be9bbfbdfcbf20a9e73ee64c590f1c0d0f77b4df71860dac4da49d02ac85.jpg) +1.00/3.9 + +![](images/a33b3698a9b9bf9ec8d6be3f4ba84af1597e4b1320e4ade818bb8ea9016f2e2a.jpg) +1.00/3.9 + +![](images/a45101c80cd815b58da987d5c24eb8cc7467832f0b79b8ff6280ecc1cc92d5bf.jpg) +Images with spurious bird feeder but no hummingbird +0.94/5.7 + +![](images/4a80f1d75e79d106d0c8c264cc6927f495f0e31e4d252bedaf73e10468a47772.jpg) +0.94/3.4 + +![](images/2125c00b429315f1ae61534ff1e8810e53af39f0e1e0403aa0d738e5382e8de1.jpg) +0.82/2.9 + +![](images/3e511056306d77c041e2d2bedd8de69bccd3cc66a080d4053953d0aa0f6033a3.jpg) +0.91/5.6 + +![](images/06196a4a233b5176ab789b24e32008598fac17ae3656c05b544b2b2a8773a617.jpg) +0.91/4.7 + +![](images/edc3193bf8177af534bf2d4a751fcb1c5fad50d517725c2ffcac95807adc870a.jpg) +all classified as humming bird by four ImageNet models +0.86/4.3 + +![](images/ee7284a760a207d058f8128dc592dbb8be7117ac1c127953c23efa56ba4f39e1.jpg) +0.81/3.5 + +![](images/acf64746e60be599c2c47372211bc4e0dd74fa501a30b96ec72507e8aee97826.jpg) +0.89/3.3 + +![](images/2abd1f654cc64097e2ebf3229076a5525b3f04014995084534c3cc5adbddc412.jpg) +0.89/3.8 + +![](images/3607ca6620d48cf89461c71bdc2b4d885649ec1a1ad2ae54d78c1dc9ad27e744.jpg) +0.78/5.91 + +![](images/262cfde3f1b1b79b83e2e26f16d2506c66fc6a90aeb75608aa7d14554bda8741.jpg) +Images with spurious grafitti but no freight car +0.97/3.5 + +![](images/60a9a85f1c8ca0e02b8d4edf4ead405387013f9f69b5020ccbfee9a1b1557f52.jpg) +0.79/4.0 + +![](images/76e56ea2c33cc00427a7e546d8363a91a75072aa98798556b58c29822ae72383.jpg) +0.81/2.4 + +![](images/57dee91e1490077c8b1dd710f3f532ea5811d60104ceef2245d5236238366c74.jpg) +0.85/3.2 + +![](images/b43ee7265dc02cca8f65dd9204d7e5b6e07e295b14d14fedd59c59b1ad683b99.jpg) +0.88/2.6 + +![](images/4e7140d05f47c85caef4c21a641ee3cdd70ed87a0f5ee698c50314669f9a5ed5.jpg) +all classified as freight car by four ImageNet models +0.85/2.5 + +![](images/5c765abbfa0845be76cfba9bc4314b4a671a86dd65c0eee290423e949c101756.jpg) +0.86/2.5 + +![](images/271004bd3fd2d0688020e1fae53de0d933064f13017c960bb7ad9840f3b8a3f2.jpg) +0.90/2.3 + +![](images/63240b2ea26dd0e19553f2428c0b456d15ab83512661cc740427511cc9713ebb.jpg) +0.82/2.2 + +![](images/682e474921ec27b310ef72061a462208f2a81668ae224208a102796055f4bccd.jpg) +0.87/2.2 + +![](images/7cc871558c2f74dee577490a040eb115083fdc3a1eb2923c0b95f3275e771343.jpg) +Images with spurious eucalyptus/plants but no koala +0.49/3.5 + +![](images/d9ed6648e6e7ec103f36ec6ef269baa9d8fb476d2c799d461b1622c9f6d0389a.jpg) +0.61/3.1 + +![](images/13fc59d130ba54cbeb33bb7879ca0346e9c93fd71b30d839afdd5f713dfd35d6.jpg) +0.36/3.1 + +![](images/2a962ccc0109ae8bdf64c42eeb97112c411021f75caf04cf924183cc02ca1e9f.jpg) +0.36/3.1 + +![](images/fb2d9fedf1c00962f3a8ec7fab10f710347d785b7826f1e9f1dab34aa3102473.jpg) +0.69/2.7 + +![](images/ec0ce2d1a664953e5608d6b6070ad07de40d8ae64d8af6c90f0e37df146a5a15.jpg) +all classified as koala by four ImageNet models +0.56/2.6 + +![](images/fef7e6a9fe2d10335af4e7728a604c8d30f401e872a0b1d6129d41fe1088b2a7.jpg) +0.62/2.5 + +![](images/c6a74f224f78c5b62042650c42fe903570e5a1b792771e345f869fa79c355cf9.jpg) +0.74/2.5 + +![](images/e1ac2a48b5b267a5f9f1093a7a3007ddfba246d63efbbb94e4697652f0f54be7.jpg) +0.72/2.4 + +![](images/a0daf6d3b2d23f26bb5ad7a69324b618fb81cc9c899f2cf52ae2139f715669f8.jpg) +0.66/2.2 + +![](images/539f197c9df300a1aa9e765daa8d1af929468c7716c3c3b51dea548ae04dddc1.jpg) +Images with spurious water jet but no fireboat +0.63/2.3 + +![](images/28b216cd46e16f027887cf4dffcf9a136cb940f482ebefb3db4519920803a214.jpg) +0.71/2.2 + +![](images/66c383237bb2b0f4faa016135376e8c11ad7ff76c853f68d1c53d07adf2b222e.jpg) +0.84/2.2 + +![](images/2e74305ae404761c3c77dfd49a05a9731b27bfd41aeac5fd6fd6a2983548f4b6.jpg) +0.65/2.1 + +![](images/1d50568e613b91952708f5aba8784e3606eeb8792a7aa51f94dd57b17ffaf90a.jpg) +0.63/1.9 + +![](images/b0e378ac19e854bde16c8c16b81016e8bd6d2062686b3156b52ea4c629b31439.jpg) +all classified as fireboat by four ImageNet models +0.53/1.9 + +![](images/494d868e345a740ca98e1534f8c2d1a4f19efe7c7b1ec1748cfda15a9b3ad39e.jpg) +0.79/1.9 + +![](images/71073f348182fd929fb4a5ab6345d59d1d824aef912ad5ff5b6628923f879d12.jpg) +0.78/1.8 + +![](images/3fa638d12c61fb5772f77e9553ee27edc4abbc4a025a4dc50b658973ecc02481.jpg) +0.79/1.8 + +![](images/b9ca131bbf96e9cf4b8c4e9a39bcbd814f4f36181e146dbf343b815b8d441331.jpg) +0.76/1.8 + +training. However, this would require relabeling all spurious ImageNet classes which is not feasible. We could order the training set according to the value $\alpha_{l}^{(k)}$ of the corresponding neural PCA component which indicates how much of the spurious feature an image contains. While this would speed up the process significantly, it would still require a significant amount of manual relabeling. Can one do it also without any additional labeling? Yes, as described in Sec. 4 we can rewrite the logit of the $k$ -th class as + +$$ +f _ {k} (x) = \sum_ {l = 1} ^ {D} \alpha_ {l} ^ {(k)} (x) + \left\langle \mathbf {1}, \bar {\psi} _ {k} \right\rangle + b _ {k}. \tag {7} +$$ + +For a spurious component $l$ of class $k$ , we use $\min \{\alpha_l^{(k)}, 0\}$ instead of $\alpha_l^{(k)}$ to remove its positive contribution from the logit (negative contributions are semantically different). After removal of the spurious features, the new logit becomes + +$$ +f _ {k} ^ {S p u F i x} (x) = f _ {k} (x) - \sum_ {l \in \mathcal {S} _ {k}} \max \left\{\alpha_ {l} ^ {(k)} (x), 0 \right\} \tag {8} +$$ + +where $S_{k}$ is the set of spurious NPCA components of class $k$ . We denote this method as SpuFix. It significantly reduces dependence on spurious features, see Sec. 7.3. + +Transfer of SpuFix to any ImageNet classifier: As described in Sec. 3, harmful spurious features are a result of subpopulations in the training data. While not every spurious correlation will be picked up by every model, most of our detected spurious features generalize to a wide range of classifiers (see Sec. 7). In the following, we show how SpuFix can be transferred to any given ImageNet classifier $\tilde{f}$ for which $\tilde{f}_k$ denotes the logit and $\tilde{\psi}_k$ the weighted penultimate layer of class $k$ . The goal is to find a direction $b$ in the weighted feature space $\tilde{\psi}_k$ of $\tilde{f}$ for every spurious NPCA component $l$ corresponding to the eigenvector $v_l$ of the original model, resp. $\alpha_l^{(k)}(x)$ , and then truncate its positive component. To find this direction we maximize the covariance of the projection onto $b$ and $\alpha_l^{(k)}$ over the training images of class $k$ : + +$$ +b _ {l} ^ {(k)} = \underset {\| b \| _ {2} = 1} {\operatorname {a r g m a x}} \sum_ {s \in I _ {k}} \left\langle b, \tilde {\psi} _ {k} \left(x _ {s}\right) - \bar {\tilde {\psi}} _ {k} \right\rangle \alpha_ {l} ^ {(k)} \left(x _ {s}\right). \tag {9} +$$ + +which has a closed form solution + +$$ +b _ {l} ^ {(k)} = \frac {\sum_ {s \in I _ {k}} \left(\tilde {\psi} _ {k} \left(x _ {s}\right) - \bar {\tilde {\psi}} _ {k}\right) \alpha_ {l} ^ {(k)} \left(x _ {s}\right)}{\left\| \sum_ {s \in I _ {k}} \left(\tilde {\psi} _ {k} \left(x _ {s}\right) - \bar {\tilde {\psi}} _ {k}\right) \alpha_ {l} ^ {(k)} \left(x _ {s}\right) \right\| _ {2}}. \tag {10} +$$ + +In contrast to the eigenvectors $v_{l}$ , the matched vectors $b_{l}^{(k)}$ are not necessarily orthogonal. Thus before truncation the centered features $\tilde{\psi}_k(x) - \bar{\tilde{\psi}}_k$ need to be projected onto the subspace spanned by the $b_{l}^{(k)}$ and represented in the non-orthogonal basis $\{b_{l}^{(k)}\}_{l\in S_{k}}$ . We denote this representation + +by $P^{(k)}(x)$ (details in C.1). The logit of class $k$ of the SpuFix version of $\tilde{f}$ is then: + +$$ +\tilde {f} _ {k} ^ {S p u F i x} (x) = \tilde {f} _ {k} (x) - \sum_ {l \in \mathcal {S} _ {k}} \max \left\{\langle \mathbf {1}, b _ {l} ^ {(k)} \rangle P _ {l} ^ {(k)} (x), 0 \right\}. \tag {11} +$$ + +In the case where $\tilde{f} = f$ (the robust model), we recover the original SpuFix truncation in (8) (see C.2). It turns out that SpuFix is even effective when the architecture is quite different from the ResNet50 we used for detection, e.g. ViT or VOLO, see Sec. 7.3 and Table 1 and 2. + +# 6. Comparison to neural features of [61] + +We compare our NPCA framework to the method of [61] to detect spurious features for ImageNet. As model, they use a $\ell_2$ -robust ResNet50. Let $J_{k}$ be the set of training images classified as class $k$ . [61] define the $j$ -th component $m_j^{(k)}$ of the class-wise mean over predictions, + +$$ +m ^ {(k)} = \frac {1}{\left| J _ {k} \right|} \sum_ {x _ {s} \in J _ {k}} \psi_ {k} \left(x _ {s}\right), \tag {12} +$$ + +as the importance of the $j$ -th neuron for class $k$ .1 Then, they order the neurons of the penultimate layer according to the score $m_j^{(k)}$ and consider the top-5 neurons of each class. The main difference to our approach is that they assume single neurons with maximal influence on the mean are capturing spurious features whereas our NPCA components are linear combinations of neurons that capture the variance around the mean. Given that the ResNet50 has only 2048 neurons for 1000 classes, some neurons are labeled as core and spurious feature for multiple classes simultaneously, even though the images are quite different. A major advantage of NPCA is that due to the orthogonality of the PCA components, we identify diverse subpopulations in the training data. As [61] use no constraints for the neurons, they often find very similar subpopulations. Hence, one may miss spurious subpopulations when only checking the top-5 components, see Fig. 5. Another difference is that they maximize the score $m_j^{(k)}$ for the training images $x_r$ + +$$ +w _ {j} ^ {(k)} = x _ {r} + \mathrm {a r g m a x} _ {\| \delta \| _ {2} \leq \epsilon} m _ {j} ^ {(k)} (x _ {r} + \delta), +$$ + +whereas we maximize our NPCA component $\alpha_l^{(k)}(x)$ starting from a gray image to introduce no bias. Thus, we check if the classifier produces the spurious feature and not only enhances it on an image showing it already. + +The third difference is that [61] want to identify any spurious feature while our goal is to find harmful ones. Thus, they use weaker criteria for deciding if a neuron shows a spurious feature: main criterion is if the neural activation + +![](images/07923a3e5e2377b02bbc92739db36494da0d9c91fd2f8284e531fe75ea1bd88d.jpg) +Top 5 neural PCA components (ours) +Figure 5: Comparison of top-5 NPCA and top-5 neurons for class hummingbird for the robust model of [61]: Our NPCA components identify diverse subpopulations in the training set whereas the top neurons show similar ones and the spurious bird feeder is not detected, see also App. A. + +map based on $m_{j}^{(k)}$ is off the class object according to a majority vote of 5 human labelers. The visualizations $w_{j}^{(k)}$ are only used if the activation maps are inconclusive. In contrast, we require i) our NFPV to show the spurious feature, ii) the GradCAM based on $\alpha_{l}^{(k)}$ highlights mainly the spurious feature, and iii) our human labelers have to agree that the NPCA component is a harmful spurious feature. + +As our criteria are more strict (in particular, that the NFPV shows the spurious feature is a strong criterion), it is not surprising that [61] find more spurious features (630 in 357 classes) than we do (322 in 230 classes). Moreover, the employed models are different and we examine top-10 NPCA components whereas they check top-5 neurons. Thus, the comparison is difficult and our novel dataset "SpuriousImageNet" could be biased towards spurious features which only we found. Hence, we compute our top-5 NPCA components for their robust ResNet50 for a direct comparison to their top-5 neurons and their found spurious features. In Fig. 5 we compare them for the class hummingbird where they do not find the spurious feature "bird feeder". In general, we observe that our found subpopulations are more diverse and thus we find more spurious features than they do when using their weaker criteria. In App. E we do an extensive comparison for all classes. + +# 7. Experiments + +In this section, we provide a qualitative and quantitative evaluation of our 322 detected spurious features in Ima + +geNet, see Sec. 4. For the quantitative evaluation, we create the dataset "Spurious ImageNet", which allows checking the reliance of a given ImageNet classifier on spurious features. We also evaluate our mitigation strategy "SpuFix" which does not require additional labels or retraining of the classifier and can be transferred to other image classifiers. + +# 7.1. Qualitative evaluation + +For the qualitative evaluation, we visualize some of our found 322 spurious features, see Fig. 4. For each class, we show five random training images, the NPFV, and the four most activating training images of the neural PCA component labeled to be spurious. Additionally, we always show ten images which only show the spurious feature but not the actual class e.g., only the bird feeder (spurious) but no hummingbird (class). All ten images are classified as the corresponding class for the robust classifier we have used to compute the NPCA components and three non-robust ImageNet classifiers (ResNext101, Eff.Net B5, ConvNext-B, see also Tab. 2). This shows that our spurious features generalize from the robust classifier to SOTA ImageNet classifiers, indicating that the found spurious features are mainly due to the design of the training set, rather than failures in model training. Our novel validation by collecting real images with the spurious feature but without the class object which are consistently classified as this class directly shows the impact of harmful spurious features and has the advantage that it does not introduce artifacts via masking nor requires expensive pixel-wise segmentations. + +Image Collection: The images showing only the spurious feature were obtained by sorting the 9 million images of OpenImages [38] by the value $\alpha_l^{(k)}(x)$ of the neural PCA component. We check the top 625 retrieved images classified by the robust classifier as the corresponding class if they are all classified as the same class by the additional non-robust classifiers and do not show the corresponding class. This is a quite strict criterion as spurious features can be shared across classes, e.g., twigs for birds, and thus agreement of classifiers is not granted and the images can show the spurious feature and the true class. Nevertheless, this procedure yields between 77 ("hummingbird") and 179 ("freight car") images of which we show a selection. For "hummingbird", a lot of these images show red flowers (without a hummingbird) which makes sense as the NPFV displays features of red flowers and due to the bias that OpenImages does not contain many images of hummingbird feeders. In these cases, we additionally retrieve Flickr images with appropriate text queries e.g., "hummingbird feeder" and filter them. + +Spurious Class Extension: For "hummingbird", "freight car", and "koala" the spurious features significantly extend the predictions beyond the actual class (see Fig. 2). Bird feeders are classified as hummingbirds, graf + +![](images/4563f6e3923208960d0af87f591719c55f3f480f7ecd87599ed63d9f3efe52e4.jpg) +Figure 6: Spurious Score: we plot the AUC of different models for 7 out of 100 classes in "Spurious ImageNet" (see Fig 13-15): hummingbird (bird feeder/red flowers), freight car (graffiti), koala (plants/trees), fireboat (water jet), flagpole (US flag), gondola (house/river), and bookshop (storefront). The spurious features for hummingbird, freight car and koala which were not detected in [61, 62] are also spurious for their robust ResNet50 [61]. Training on ImageNet 21k or Fine-tuning from 21k (1kFT21k) decreases dependence on harmful spurious features but classes like flagpole and bookshop remain strongly affected. Our cheap mitigation strategy SpuFix improves the AUCs significantly for both robust ResNet50 but also improves the ViT-B variants trained/fine-tuned on ImageNet1k, especially for the difficult classes flagpole and bookshop. + +![](images/b51854110c5f9577a2e18122ed2a08190de46fc889f1d6c0d18446e8e6427c13.jpg) +hummingbird (red flowers) +fireboat + +![](images/4f6a05639cea9b15a673e3da497bd21175863f24e67b1836acae247343f22319.jpg) +(water jet) +Figure 7: Spurious ImageNet: sample images from the dataset for 6 out of 100 classes showing the spurious feature but not the class object, see also App. G. + +![](images/eed06027d1d78c3d80f163024a2ef6b152593b0359f6fda13b18a2ff780eb396.jpg) +freight car (graffiti) + +![](images/6c6536a6220f471cfa4105ae0c1c1345cf028c60727e243c8859ae4894691d0e.jpg) +flagpole (US flag) + +![](images/7860febbf72c9ffcd5b66b6bba9296c151a3373c7dbbe608d3df22208bf00745.jpg) +koala (plants/trees) + +![](images/21bf0e1b5da35e544868e770485cc819e022f017481c6cb0c5bb08570e1e56b5.jpg) +gondola (house/river) + +fiti as freight cars, and (eucalyptus) plants as koalas. This class extension cannot be detected by monitoring test performance and thus is likely to be noticed only after deployment. For "hummingbird", we see in Fig. 4 two images with bees on the bird feeder where "bee" is an ImageNet class (also a grasshopper for "koala"). Nevertheless, the spurious "bird feeder" feature of "hummingbird" overrules "bee" even though no hummingbird is present. + +Spurious Shared Feature: The spurious feature "water jet" is shared among the classes "fireboat" and "fountain". It appears more frequently for "fireboat" (see Fig. 2) which leads to an in-distribution shift where now a large number of images of the "fountain"-class with a water jet are wrongly classified as "fireboat". The spurious feature "water jet" for "fireboat" has been found also in the Salient ImageNet dataset [61, 62]. However, they did not find spurious features for freight car and koala (in App. A we do a comparison). More examples are in App. F. + +# 7.2. The Spurious ImageNet dataset + +A key contribution of this paper is our novel evaluation of spurious features for image classifiers without requiring pixel-wise annotations [48, 59] or having to rely on the validity of neural heatmaps [61]. Instead, we use images from OpenImages to show that images only containing the spurious feature but not the class object are classified as this class. This has the advantage that we consider real images and thus provide a realistic impression of the performance of ImageNet classifiers in the wild. Adding noise [61] or masking [48, 44] image regions requires pixel-wise accurate annotations which are labor-expensive, masking only the object still contains shape information, and using masks avoiding this, e.g., a bounding box around the object, can hide a significant portion of the image which is unrealistic. + +To allow for a quantitative analysis of the influence of spurious features on ImageNet classifiers, we collected images similar to the ones shown to illustrate the spurious features in Fig. 4. The images are chosen such that they show the spurious feature but not the class object. The only difference is that we relax the classification condition and only require two of the four classifiers (robust ResNet50, ResNext101, EfficientNet-B5, ConvNext-B) to predict the corresponding class. We select 100 of our spurious features and for each collect 75 images from the top-ranked images in OpenImages according to the value of $\alpha_{l}^{(k)}$ for which two human labelers agree that they contain the spurious feature but not class $k$ and two out of four classifiers predict class $k$ . We call the dataset Spurious ImageNet as it allows to check the dependence on spurious features with real images for ImageNet classifiers, see Fig. 7 and App. G for samples. + +Spurious Score: A classifier $f$ not relying on the spurious feature should predict a low probability for class $k$ for the Spurious ImageNet samples, especially compared to ImageNet test set images of class $k$ . Thus, for each class, we measure the AUC (area under the curve) for the separation + +of images with the spurious features but not showing class $k$ versus test set images of class $k$ according to the predicted probability for class $k$ . A classifier not depending on the spurious feature should attain a perfect AUC of 1, whereas a value significantly below 1 shows strong reliance. We report the mean AUC (mAUC) over all 100 classes in Tab. 1. All ImageNet models trained only on ImageNet1k are heavily influenced by spurious features. Thus, spurious features are mainly a problem of the training set rather than the classifier, and spurious features found with an adversarially robust model transfer to other ImageNet classifiers. + +Pre-training on larger datasets: Some spurious features such as flag (flag pole), bird feeder (hummingbird), and eucalyptus (koala) are classes in ImageNet21k. Therefore, they should no longer be spurious for the other classes. Thus, we test if ImageNet1k-classifiers fine-tuned from an ImageNet21k model are less reliant on spurious features. The results in Tab. 1 and Fig. 6 suggest that the influence of spurious features is damped but they are far from being free of them. To check how much is lost due to fine-tuning we evaluate a ViT-B trained on ImageNet21k which has a mean AUC of 0.931 whereas the fine-tuned model has 0.917. This shows that fine-tuning does not hurt much. While finetuning from ImageNet21k improves the mean AUC, for several classes the dependence on spurious features is still significant, see also Fig. 16 how one has to be careful in the interpretation of higher AUC values. In addition to ImageNet21k, we also evaluate models trained on other large image datasets (JFT-300M[28], YFFC-100M, 1B Instagram[84], MIM[24], LAION-2B and LAION-400M[54]) using self-supervised learning or which are based on CLIP [49]. However, these models also do not achieve better spurious scores (Tab. 1 and Tab. 2). We evaluate a large number of SOTA models in App. D. + +# 7.3. Evaluation of mitigation technique SpuFix + +Fixing spurious features is a non-trivial task and can require a substantial labeling effort. We evaluate our SpuFix from Sec. 5 that does not require retraining or additional labels. The positive effect of this fix of spurious features (SpuFix) can be seen in Tab. 1 and Fig. 6. Compared to the original robust ResNet50 with a spurious mAUC of 0.630, the SpuFix version has a significantly better spurious mAUC of 0.763. Test set accuracy reduces by $0.6\%$ but this is a rather positive effect, as several of the additional errors arise since the robust ResNet50 uses spurious features for its decision, e.g., for classes like "balance beam" or "puck" the class object is often not visible in the cropped test set images. In Table 2 we provide a large scale evaluation of the transfer of SpuFix to SOTA ImageNet models. We observe a consistently better mAUC on Spurious ImageNet, even for very large models fine-tuned from 21k or trained on other large datasets, e.g. SpuFix improves the mAUC of VOLO- + +
ModelOriginalSpuFix
INet Acc. ↑SpurIN mAUC ↑INet Acc. ↑SpurIN mAUC ↑
ImageNet1k
Rob. ResNet5057.4%0.63056.8%0.763
Rob. ResNet50[61]57.9%0.65157.2%0.764
ConvNeXt-L[41]84.8%0.80384.8%0.819
ViT-B AugReg[66]81.1%0.85081.1%0.859
VOLO-D5 512[85]87.1%0.88287.1%0.907
ImageNet21kFT1k
EfficientNetv2-L[70]86.8%0.89386.8%0.898
ConvNeXt-L[41]87.0%0.91087.0%0.913
ViT-B AugReg[66]86.0%0.91785.9%0.925
BEiT-L\16[8]88.6%0.92188.6%0.927
LAION-2B
CNeXt-L CLIP 384[49]87.8%0.87987.9%0.884
ViT-L\14 CLIP[49]88.2%0.91288.2%0.914
MIM
EVA-G\14 CLIP 560[24]89.8%0.91989.8%0.925
ImageNet21k
ConvNeXt-L[41]-0.943-0.943
ViT-B AugReg[66]-0.931-0.931
+ +Table 1: Quantitative Evaluation on Spurious ImageNet: ImageNet classifiers of different training modalities depend on spurious features in varying strength. The mAUC is the mean of AUCs for the separation of images containing the spurious feature but not class $k$ versus test images of class $k$ with the predicted probability of class $k$ as score. + +D5 (87.1% acc.) trained only on 1k by 2.5%, or by 0.6% for EVA-G $\backslash$ 14 CLIP 560 trained on MIM (91.9% acc.), as well as BeiT-L $\backslash$ 16 fine-tuned from 21k (88.6% acc.) by 0.6%. Thus even SOTA models profit from our SpuFix with negligible difference in accuracy ( $\leq$ 0.1%) and thus the use of SpuFix is recommended for any ImageNet model. + +# 8. Conclusion + +We have shown that large-scale identification of spurious features is feasible with our neural PCA components and neural PCA feature visualizations. With "Spurious ImageNet" we introduced a novel dataset to evaluate the dependence of ImageNet classifiers on spurious features based on real images. We demonstrated that our SpuFix method mitigates the dependence on harmful spurious features for any ImageNet classifier without costly labeling or re-training. + +Acknowledgements We acknowledge support by the DFG, Project number 390727645, the Carl Zeiss Foundation, project "Certification and Foundations of Safe Machine Learning Systems in Healthcare". The authors thank the IMPRS-IS for supporting YN. + +# References + +[1] Julius Adebayo, Michael Muelly, Hal Abelson, and Been Kim. Post hoc explanations may be ineffective for detecting unknown spurious correlation. In ICLR, 2022. +[2] Julius Adebayo, Michael Muelly, Ilaria Liccardi, and Been Kim. Debugging tests for model explanations. In NeurIPS, 2020. +[3] D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Mane. Concrete problems in AI safety. arXiv:1606.06565v2, 2016. +[4] Christopher J. Anders, Leander Weber, David Neumann, Wojciech Samek, Klaus-Robert Müller, and Sebastian Lapuschkin. Finding and removing clever hans: Using explanation methods to debug and improve deep models. Information Fusion, 77:261-295, 2022. +[5] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. In NeurIPS, 2022. +[6] Maximilian Augustin, Alexander Meinke, and Matthias Hein. Adversarial robustness on in- and out-distribution improves explainability. In ECCV, 2020. +[7] David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja Hansen, and Klaus-Robert Müller. How to explain individual classification decisions. JMLR, 2010. +[8] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. +[9] Sara Beery, Grant van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, 2018. +[10] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with ImageNet? arXiv:2006.07159, 2020. +[11] B. Biggio, I. Corona, D. Maiorca, B. Nelson, N. Srndic, P. Laskov, G. Giacinto, and F. Roli. Evasion attacks against machine learning at test time. In ECML/PPKD, 2013. +[12] Alceu Bissoto, Eduardo Valle, and Sandra Avila. Debiasing skin lesion datasets and models? not so fast. arXiv:2004.11457, 2020. +[13] Valentyn Boreiko, Maximilian Augustin, Francesco Croce, Philipp Berens, and Matthias Hein. Sparse visual counterfactual explanations in image space. In $GCPR$ , 2022. +[14] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020. +[15] Myung Jin Choi, Antonio Torralba, and Alan S. Willsky. Context models and out-of-context objects. Pattern Recognition Letters, 33(7):853-862, 2012. +[16] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. + +[17] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR workshop, pages 702-703, 2020. +[18] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. +[19] Ann-Kathrin Dombrowski, Maximillian Alber, Christopher Anders, Marcel Ackermann, Klaus-Robert Müller, and Pan Kessel. Explanations can be manipulated and geometry is to blame. In NeurIPS, 2019. +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. +[21] Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. Adversarial robustness as a prior for learned representations, 2019. +[22] Christian Etmann, Sebastian Lunz, Peter Maass, and Carola-Bibiane Schonlieb. On the connection between adversarial robustness and saliency map interpretability. In ICML, 2019. +[23] Sabri Eyuboglu, Maya Varma, Khaled Saab, Jean-Benoit Delbrouck, Christopher Lee-Messer, Jared Dunnmon, James Zou, and Christopher Ré. Domino: Discovering systematic errors with cross-modal embeddings. In ICLR, 2022. +[24] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636, 2022. +[25] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. +[26] R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. +[27] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In AAAI, 2019. +[28] A Gupta, C Sun, A Shrivastava, and S Singh. Revisiting the unreasonable effectiveness of data. arXiv preprint arXiv:1707.02968, 2017. +[29] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. +[30] M. Hein, M. Andriushchenko, and J. Bitterwolf. Why ReLU networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In CVPR, 2019. +[31] D. Hendrycks and T. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019. +[32] D. Hendrycks and K. Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. +[33] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. + +[34] Sara Hooker, Yann Dauphin, Aaron Courville, and Andrea Frome. Selective brain damage: Measuring the disparate impact of model pruning. In ICLR, 2020. +[35] Pieter Abbeel Jonathan Ho, Ajay Jain. Denoising diffusion probabilistic models. In NeurIPS, 2020. +[36] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In ECCV, 2020. +[37] A. Kurakin, I. J. Goodfellow, and S. Bengio. Adversarial examples in the physical world. In *ICLR Workshop*, 2017. +[38] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. +[39] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. +[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. +[41] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. +[42] Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In NeurIPS, 2017. +[43] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. +[44] Mazda Moayeri, Sahil Singla, and Soheil Feizi. Hard ImageNet: Segmentations for objects with strong spurious cues. In NeurIPS Datasets and Benchmarks Track, 2022. +[45] A. Nguyen, J. Yosinski, and J. Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In CVPR, 2015. +[46] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. +[47] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. Proc. CVPR, 2022. +[48] Gregory Plumb, Marco Tulio Ribeiro, and Ameet Talwalkar. Finding and fixing spurious patterns with explanations. Transactions on Machine Learning Research (TMLR), 2022. +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. +[50] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In ICML, 2019. + +[51] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. "why should i trust you?": Explaining the predictions of any classifier. In KDD, 2016. +[52] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. License: No license specified. +[53] Shibani Santurkar, Dimitris Tsipras, Brandon Tran, Andrew Ilyas, Logan Engstrom, and Aleksander Madry. Image synthesis with a single (robust) classifier. In NeurIPS, 2019. +[54] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. +[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. +[56] Kathryn Schutte, Olivier Moindrot, Paul Hérent, Jean-Baptiste Schiratti, and Simon Jégou. Using stylegan for visual interpretability of deep learning models on medical images. In NeurIPS Workshop, 2020. +[57] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. +[58] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. In NeurIPS, 2020. +[59] Rakshith Shetty, Bernt Schiele, and Mario Fritz. Not using the car to see the sidewalk-quantifying and controlling the effects of context in classification and segmentation. In CVPR, 2019. +[60] Krishna Kumar Singh, Dhruv Mahajan, Kristen Grauman, Yong Jae Lee, Matt Feiszli, and Deepti Ghadiyaram. Don't judge an object by its context: learning to overcome contextual bias. In CVPR, 2020. +[61] Sahil Singla and Soheil Feizi. Salient imagenet: How to discover spurious features in deep learning? In ICLR, 2022. +[62] Sahil Singla, Mazda Moayeri, and Soheil Feizi. Core risk minimization using salient imagenet. arXiv:2203.15566, 2022. +[63] Sahil Singla, Besmira Nushi, Shital Shah, Ece Kamar, and Eric Horvitz. Understanding failures of deep networks via robust feature extraction. In CVPR, 2021. +[64] Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju. Fooling lime and shap: Adversarial attacks on post hoc explanation methods. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 180-186, 2020. +[65] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. + +[66] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. In TMLR, 2022. +[67] Pierre Stock and Moustapha Cisse. Convnets and ImageNet beyond accuracy: Understanding mistakes and uncovering biases. In ECCV, 2018. +[68] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014. +[69] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. +[70] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In ICML, 2021. +[71] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR, pages 1521-1528, 2011. +[72] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. +[73] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. From ImageNet to image classification: Contextualizing progress on benchmarks. In ICML, 2020. +[74] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019. +[75] Sahil Verma, John P. Dickerson, and Keegan Hines. Counterfactual explanations for machine learning: A review. arXiv:2010.10596, 2020. +[76] Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the GDPR. Harvard Journal of Law & Technology, 2018. +[77] Ross Wightman. Pytorch image models. https://github.com/ rwightman/pytorch-image-models, 2019. +[78] Eric Wong, Shibani Santurkar, and Aleksander Madry. Leveraging sparse linear layers for debuggable deep networks. In ICML, 2021. +[79] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023. +[80] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. In ICLR, 2021. +[81] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. +[82] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. +[83] Cong Xu and Min Yang. Adversarial momentum-contrastive pre-training. arXiv preprint, arXiv:2012.13154v2, 2020. +[84] I Zeki Yalniz, Herve Jégou, Kan Chen, Manohar Paluri, and Dhruv Mahajan. Billion-scale semi-supervised learning for image classification. arXiv preprint arXiv:1905.00546, 2019. + +[85] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision observer for visual recognition. PAMI, 2022. +[86] John R. Zech, Marcus A. Badgeley, Manway Liu, Anthony B. Costa, Joseph J. Titano, and Eric Karl Oermann. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLOS Medicine, 15(11):1-17, 2018. +[87] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. +[88] Zhuotun Zhu, Lingxi Xie, and Alan L. Yuille. Object recognition with and without objects. In *IJCAI*, 2017. \ No newline at end of file diff --git a/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/images.zip b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fc4fcb48babc8866b47de4fdf3c8d3eda784b561 --- /dev/null +++ b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57eaf71dac6587a0031e1f5abd288c81c1b0ce598fb32146a359cab847a5f003 +size 939368 diff --git a/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/layout.json b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8874164d8ba12a3cf8f0495b6a0adad098aab24b --- /dev/null +++ b/2023/Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet/layout.json @@ -0,0 +1,19178 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 533, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 533, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 533, + 140 + ], + "type": "text", + "content": "Spurious Features Everywhere - Large-Scale Detection of Harmful Spurious Features in ImageNet" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 161, + 167, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 161, + 167, + 175 + ], + "spans": [ + { + "bbox": [ + 85, + 161, + 167, + 175 + ], + "type": "text", + "content": "Yannic Neuhaus" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 161, + 298, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 161, + 298, + 175 + ], + "spans": [ + { + "bbox": [ + 192, + 161, + 298, + 175 + ], + "type": "text", + "content": "Maximilian Augustin" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 324, + 161, + 411, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 161, + 411, + 175 + ], + "spans": [ + { + "bbox": [ + 324, + 161, + 411, + 175 + ], + "type": "text", + "content": "Valentyn Boreiko" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 438, + 161, + 508, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 161, + 508, + 175 + ], + "spans": [ + { + "bbox": [ + 438, + 161, + 508, + 175 + ], + "type": "text", + "content": "Matthias Hein" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 185, + 175, + 408, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 175, + 408, + 190 + ], + "spans": [ + { + "bbox": [ + 185, + 175, + 408, + 190 + ], + "type": "text", + "content": "Tübingen AI Center - University of Tübingen" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 242, + 290, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 242, + 290, + 494 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 290, + 494 + ], + "type": "text", + "content": "Benchmark performance of deep learning classifiers alone is not a reliable predictor for the performance of a deployed model. In particular, if the image classifier has picked up spurious features in the training data, its predictions can fail in unexpected ways. In this paper, we develop a framework that allows us to systematically identify spurious features in large datasets like ImageNet. It is based on our neural PCA components and their visualization. Previous work on spurious features often operates in toy settings or requires costly pixel-wise annotations. In contrast, we work with ImageNet and validate our results by showing that presence of the harmful spurious feature of a class alone is sufficient to trigger the prediction of that class. We introduce the novel dataset \"Spurious ImageNet\" which allows to measure the reliance of any ImageNet classifier on harmful spurious features. Moreover, we introduce SpuFix as a simple mitigation method to reduce the dependence of any ImageNet classifier on previously identified harmful spurious features without requiring additional labels or retraining of the model. We provide code and data at https://github.com/YanNeu/spuriousImagenet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": "Deep learning has led to tremendous progress in image classification [37, 50] and natural language processing [14]. Over the years, however, it has become apparent, that evaluating predictive performance on a fixed test set is not necessarily indicative of the performance when image classifiers are deployed in the wild. Several potential failure cases have been discovered. This starts with a lack of robustness due to image corruptions [31], adversarial perturbations [68], and arbitrary predictions on out-of-distribution inputs [45, 32, 30]. In this paper, we consider the problem of identifying and debugging image classifiers from spurious features [2]. Spurious features in image classification are features that co-occur with the actual class object and are picked up by the classifier. In the worst case, they lead to shortcut learning [25], where only the spurious but not the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 217, + 498, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 217, + 498, + 228 + ], + "spans": [ + { + "bbox": [ + 354, + 217, + 498, + 228 + ], + "type": "text", + "content": "Spurious Features in Training Data" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 322, + 229, + 368, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 229, + 368, + 239 + ], + "spans": [ + { + "bbox": [ + 322, + 229, + 368, + 239 + ], + "type": "text", + "content": "bird feeder" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 387, + 229, + 415, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 229, + 415, + 239 + ], + "spans": [ + { + "bbox": [ + 387, + 229, + 415, + 239 + ], + "type": "text", + "content": "graffiti" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 433, + 229, + 477, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 433, + 229, + 477, + 239 + ], + "spans": [ + { + "bbox": [ + 433, + 229, + 477, + 239 + ], + "type": "text", + "content": "eucalyptus" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 499, + 229, + 520, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 499, + 229, + 520, + 239 + ], + "spans": [ + { + "bbox": [ + 499, + 229, + 520, + 239 + ], + "type": "text", + "content": "label" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 318, + 240, + 373, + 293 + ], + "blocks": [ + { + "bbox": [ + 318, + 240, + 373, + 293 + ], + "lines": [ + { + "bbox": [ + 318, + 240, + 373, + 293 + ], + "spans": [ + { + "bbox": [ + 318, + 240, + 373, + 293 + ], + "type": "image", + "image_path": "7e837ae86e2d771d5286c9eb2a22c0f9fbc3d3a30bd4fa5dc8d07eff2f74a74d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 375, + 240, + 428, + 293 + ], + "blocks": [ + { + "bbox": [ + 375, + 240, + 428, + 293 + ], + "lines": [ + { + "bbox": [ + 375, + 240, + 428, + 293 + ], + "spans": [ + { + "bbox": [ + 375, + 240, + 428, + 293 + ], + "type": "image", + "image_path": "d7ab984c8c6ee8f1e9e913a5c4b7eb2342d45b5dba6fd82527e0b400521da1f7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 429, + 240, + 483, + 293 + ], + "blocks": [ + { + "bbox": [ + 429, + 240, + 483, + 293 + ], + "lines": [ + { + "bbox": [ + 429, + 240, + 483, + 293 + ], + "spans": [ + { + "bbox": [ + 429, + 240, + 483, + 293 + ], + "type": "image", + "image_path": "d06c41754196ab95ce2b2602d999b7e23aaaf3eec95d35183a2d02b233c11b52.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 484, + 240, + 536, + 293 + ], + "blocks": [ + { + "bbox": [ + 484, + 240, + 536, + 293 + ], + "lines": [ + { + "bbox": [ + 484, + 240, + 536, + 293 + ], + "spans": [ + { + "bbox": [ + 484, + 240, + 536, + 293 + ], + "type": "image", + "image_path": "6bb75441b7caa1ae1fdb35a463ae277a98cbca646e9806df92900c645f19b6d8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 295, + 374, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 374, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 374, + 307 + ], + "type": "text", + "content": "Hummingbird" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 378, + 295, + 425, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 295, + 425, + 307 + ], + "spans": [ + { + "bbox": [ + 378, + 295, + 425, + 307 + ], + "type": "text", + "content": "Freight Car" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 443, + 295, + 468, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 295, + 468, + 306 + ], + "spans": [ + { + "bbox": [ + 443, + 295, + 468, + 306 + ], + "type": "text", + "content": "Koala" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 488, + 295, + 531, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 488, + 295, + 531, + 306 + ], + "spans": [ + { + "bbox": [ + 488, + 295, + 531, + 306 + ], + "type": "text", + "content": "Hard Disc" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 335, + 312, + 518, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 312, + 518, + 334 + ], + "spans": [ + { + "bbox": [ + 335, + 312, + 518, + 334 + ], + "type": "text", + "content": "Images from the web with spurious feature but no class features classified as class below" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 318, + 335, + 372, + 388 + ], + "blocks": [ + { + "bbox": [ + 318, + 335, + 372, + 388 + ], + "lines": [ + { + "bbox": [ + 318, + 335, + 372, + 388 + ], + "spans": [ + { + "bbox": [ + 318, + 335, + 372, + 388 + ], + "type": "image", + "image_path": "f3fe0b619236e6498566704a9b635cb5dc304dc36671ec78be38991bb951c99e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 375, + 335, + 428, + 388 + ], + "blocks": [ + { + "bbox": [ + 375, + 335, + 428, + 388 + ], + "lines": [ + { + "bbox": [ + 375, + 335, + 428, + 388 + ], + "spans": [ + { + "bbox": [ + 375, + 335, + 428, + 388 + ], + "type": "image", + "image_path": "9555789428e3d27b20db3b5883a60a98a52c9698b51351a2a852c4d1bf7ceec9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 429, + 335, + 483, + 388 + ], + "blocks": [ + { + "bbox": [ + 429, + 335, + 483, + 388 + ], + "lines": [ + { + "bbox": [ + 429, + 335, + 483, + 388 + ], + "spans": [ + { + "bbox": [ + 429, + 335, + 483, + 388 + ], + "type": "image", + "image_path": "a15bba2a618e26b41249dcfa87a907ebacbb4b5627cd067b1fde31a871b1ce8d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 484, + 335, + 536, + 388 + ], + "blocks": [ + { + "bbox": [ + 484, + 335, + 536, + 388 + ], + "lines": [ + { + "bbox": [ + 484, + 335, + 536, + 388 + ], + "spans": [ + { + "bbox": [ + 484, + 335, + 536, + 388 + ], + "type": "image", + "image_path": "963df71285c03eaa75050a0317ab26178288277db62bcdca8efecbf19517f2a9.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 318, + 389, + 372, + 444 + ], + "blocks": [ + { + "bbox": [ + 318, + 389, + 372, + 444 + ], + "lines": [ + { + "bbox": [ + 318, + 389, + 372, + 444 + ], + "spans": [ + { + "bbox": [ + 318, + 389, + 372, + 444 + ], + "type": "image", + "image_path": "d050b9811842346325d13ed718729889cd50fa44f41bf4d6611ca5caad206e79.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 467, + 547, + 552 + ], + "lines": [ + { + "bbox": [ + 305, + 467, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 547, + 552 + ], + "type": "text", + "content": "Figure 1: Top: Examples of spurious features found via our neural PCA components but not in previous study [61]. Bottom: We validate our spurious features by mining images from the web showing only the spurious feature but not the class. They are classified by four ImageNet models as the corresponding class. Some of them even contain ImageNet classes (bees on feeder, grasshopper in leaves)." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 375, + 389, + 428, + 443 + ], + "blocks": [ + { + "bbox": [ + 375, + 389, + 428, + 443 + ], + "lines": [ + { + "bbox": [ + 375, + 389, + 428, + 443 + ], + "spans": [ + { + "bbox": [ + 375, + 389, + 428, + 443 + ], + "type": "image", + "image_path": "10e87f53c103305227963b7100d97c882846e1a51d6f6f2f4a213bd99712bb54.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 429, + 389, + 483, + 443 + ], + "blocks": [ + { + "bbox": [ + 429, + 389, + 483, + 443 + ], + "lines": [ + { + "bbox": [ + 429, + 389, + 483, + 443 + ], + "spans": [ + { + "bbox": [ + 429, + 389, + 483, + 443 + ], + "type": "image", + "image_path": "86c91affd6b9f248ab60af7e0f3d2620e962cb443f18e5c5718e3a37d97c62d5.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 484, + 389, + 536, + 443 + ], + "blocks": [ + { + "bbox": [ + 484, + 389, + 536, + 443 + ], + "lines": [ + { + "bbox": [ + 484, + 389, + 536, + 443 + ], + "spans": [ + { + "bbox": [ + 484, + 389, + 536, + 443 + ], + "type": "image", + "image_path": "00e18b10c73c49299fd1881ca4e8a1bb66343739cd0c083d2a8cb4690e4a7429.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 445, + 374, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 374, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 374, + 456 + ], + "type": "text", + "content": "Hummingbird" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 378, + 445, + 425, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 445, + 425, + 456 + ], + "spans": [ + { + "bbox": [ + 378, + 445, + 425, + 456 + ], + "type": "text", + "content": "Freight Car" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 443, + 445, + 468, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 445, + 468, + 456 + ], + "spans": [ + { + "bbox": [ + 443, + 445, + 468, + 456 + ], + "type": "text", + "content": "Koala" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 488, + 445, + 531, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 488, + 445, + 531, + 456 + ], + "spans": [ + { + "bbox": [ + 488, + 445, + 531, + 456 + ], + "type": "text", + "content": "Hard Disc" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "correct feature is associated with the class, e.g., [86] found that a pneumonia detector's bad generalization across hospitals was caused by the neural network learning to identify the hospital where the training data originated from. A weaker form of spurious feature (at least from a learning perspective) is the case when the classifier picks up the correct class features, e.g., of a hummingbird, but additionally associates a spurious feature, e.g., a bird feeder, with the class as they appear together on a subset of the training set. This becomes a harmful spurious feature if only the spurious feature without the class feature is sufficient to trigger the classification of that class, see Fig. 1 for an illustration of" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20235" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "such spurious features found via our method. Harmful spurious features are difficult to detect and thus can easily go unnoticed, leading to unexpected behaviour of a deployed image classifier." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 268, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 268, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 268, + 132 + ], + "type": "text", + "content": "In this paper we make the following key contributions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 133, + 287, + 407 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 133, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 287, + 192 + ], + "type": "text", + "content": "- we develop a pipeline for the detection of harmful spurious features with little human supervision based on our class-wise neural PCA (NPCA) components of an adversarially robust classifier together with their Neural PCA Feature Visualization (NPFV)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 192, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 287, + 239 + ], + "type": "text", + "content": "- unlike prior work, which used masking images or pixelwise annotations, we validate our found spurious features by using our NPCA components to find real images containing only the spurious feature but not the class object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 240, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 287, + 335 + ], + "type": "text", + "content": "- using these images we create the dataset \"Spurious ImageNet\" and propose a measure for dependence on spurious features. We do a large-scale evaluation of state-of-the-art (SOTA) ImageNet models. We show that the spurious features found for the robust model generalize to non-robust classifiers. Moreover, we analyze the influence of different training setups, e.g. pre-training on ImageNet21k or larger datasets like LAION." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 335, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 407 + ], + "type": "text", + "content": "- we develop SpuFix, a technique to mitigate the dependence on identified harmful spurious features without requiring new labels or retraining, and show how to transfer it to any ImageNet classifier. SpuFix consistently improves the dependence on harmful spurious features even for SOTA models with negligible impact on test accuracy." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 418, + 130, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 130, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 130, + 430 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": "When classifiers in safety-critical systems such as healthcare or autonomous driving are deployed in the wild [3], it is important to discover potential failure cases before release. Prior work has focused on corruption [31], adversarial robustness [11, 68, 43], and out-of-distribution detection [45, 32, 30]. There is less work on spurious features, although their potential harm might be higher." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 522, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 642 + ], + "type": "text", + "content": "Spurious features: It has been noted early on that classifiers show reliance on spurious features [15] e.g., a cow on the beach is not recognized [9] due to the missing spurious feature of grass. Other forms of spurious features have been reported in the classification of skin lesions [12], pneumonia [86], traffic signs [67], and object recognition [88]. Moreover, it has been shown that deep neural networks are biased towards texture [26] and background context [80], see [25] for an overview. [58] argues theoretically that spurious features are picked up due to a simplicity bias." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Detection of spurious features has been achieved using human label-intense pixel-wise annotations [48, 59, 60]. In [78], they use sparsity regularization to enforce a more interpretable model and use it for finding spurious features. [4] propose a complex pipeline to detect spurious features. While they scale to ImageNet, their analysis is lim" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "ited to a few spurious features for a subset of 100 classes. [61, 44, 62] do a search on full ImageNet based on class-weighted \"neural maps\". The neural maps are used to add noise to \"spurious\" resp. \"core\" features but no significant difference in classification performance is observed. It remains unclear if their found spurious features are harmful, that is the feature alone triggers the decision for that class." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 157, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 443 + ], + "type": "text", + "content": "Interpretability methods: In recent years several interpretability methods have been proposed e.g., attribution methods such as GradCAM [57], Shapley values [42], Relevance Propagation [7], and LIME [51]. The use of these methods for the detection of spurious features has been analyzed in [2, 1] with mixed success and it has been argued that interpretability methods are not robust [19, 64, 27]. However, attribution methods work better for robust classifiers due to more interpretable gradients [22]. Another technique is counterfactual explanations [76, 75] which are difficult to generate for images due to the similarity to adversarial examples [68]. Thus visual counterfactual explanations are realized via manipulation of a latent space [56] or in image space [53, 6, 13] for an adversarily robust classifier. Visual counterfactuals for non-robust classifiers using diffusion models [35, 65, 33, 18] have been proposed in [5]. ImageNet: ImageNet [52] suffers from several shortcomings: apart from an inherent dataset bias [71], semantically overlapping or even identical class pairs were reported [34, 73, 10], e.g., two classes \"maillot\", \"sunglass\" vs \"sun-glasses\", \"notebook\" vs \"laptop\" etc. We disregard such trivial cases of dataset contamination and focus on classes with harmful spurious features, in particular ones where only a small portion of the training set is contaminated." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 454, + 410, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 410, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 410, + 467 + ], + "type": "text", + "content": "3. Spurious features" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": "A proper definition of spurious features is difficult. We describe two settings of harmful spurious features which appear in this paper. We denote by " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " the set of all images containing objects belonging to class " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " (assuming for simplicity that we have a deterministic problem and ignoring multi-labels). Let " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " be the set of all images containing a feature " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " (e.g., a bird feeder). It is a correlated feature for class " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "C_k \\cap S" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "S \\setminus C_k" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " are non-empty, i.e., the feature occurs frequently with the class object but there is no causal implication that appearance of " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " implies the appearance of the class object (a bird feeder in the image does not imply presence of a hummingbird). A correlated feature becomes a spurious feature when the classifier picks it up as feature of this class. Not every spurious feature is immediately harmful, even humans use context information [25] to get more confident in a decision. However, a spurious feature is harmful if the spurious feature alone is enough to trigger the decision for the corresponding class without the class object being present in the image. We consider two scenarios for a harmful spurious feature shown in Fig. 2." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20236" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 289, + 198 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 289, + 198 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 289, + 198 + ], + "type": "image", + "image_path": "1c3e3e4aa0171ff564873305b533c9a4932d01f6cbdcd10514a58794c49e17e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": "Figure 2: Type of harmful Spurious Feature: Left: The spurious feature " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": " is taken up by the classifier which predicts class " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "C_k \\cup S" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": " instead only on " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": ". Right: The spurious feature " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": " is shared between classes but appears more often in class " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": ". The classifier associates " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": " with class " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": " and thus predicts class " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": " also on " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "C_l \\cap S" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": " instead of class " + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "text", + "content": "Spurious Class Extension: For this type of spurious feature (left in Fig. 2) the classifier picks up the spurious feature " + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "text", + "content": " for class " + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "text", + "content": " and predicts the class " + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "text", + "content": " even on " + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "inline_equation", + "content": "S \\setminus C_k" + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "text", + "content": " with high confidence (prediction of \"hummingbird\" for images showing a bird feeder but no hummingbird). The classifier predicts class " + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "text", + "content": " beyond its actual domain " + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 46, + 294, + 287, + 413 + ], + "type": "text", + "content": " and thus we call this a spurious class extension. While this spurious feature does not necessarily hurt in terms of test performance it can easily lead to completely unexpected behavior when the classifier is deployed in the wild." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": "Spurious Shared Feature: Here, two classes " + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": " share a spurious feature " + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": " (e.g., \"water jet\" for the classes \"fireboat\" and \"fountain\"). As there are more training images with feature " + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": " than in " + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": " the classifier associates " + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": " with class " + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": " and predicts class " + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "inline_equation", + "content": "S \\cap C_l" + }, + { + "bbox": [ + 47, + 414, + 287, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 628 + ], + "type": "text", + "content": "The two types of harmful spurious features are not exclusive. A shared spurious feature " + }, + { + "bbox": [ + 46, + 474, + 288, + 628 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 474, + 288, + 628 + ], + "type": "text", + "content": " can at the same time lead to a spurious class extension, e.g., the object bird feeder leads to a spurious class extension of the class \"hummingbird\" (see Fig. 4) to images of bird feeders without hummingbirds. In the training set the hummingbird feeder appears only in images of class \"hummingbird\" but the hummingbird feeder has parts which mimic flowers and flowers are a shared spurious feature with bees. In Fig. 4 right top row, images of bees on a bird feeder are classified as \"hummingbird\" instead of \"bee\", so the spurious feature is strong enough to override the decision for the true class \"bee\" (spurious shared feature)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 654, + 287, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 681 + ], + "type": "text", + "content": "4. Finding spurious features via neural PCA and associated feature visualizations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "First, we define our class-wise neural PCA (NPCA) which allows us to find diverse subpopulations in the train-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "ing data, e.g., we checked that the bird feeder for \"hummingbird\" is visible in " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " of the training images (component 2), while another " + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": " contain a part of it (component 3), see Fig. 3 or, for more examples, App. F. Then we introduce our neural PCA feature visualization (which requires an adversarially robust model) and how we select NPCA components for human inspection. The identification of spurious features requires minimal human supervision, and our effective setup allows us to screen all ImageNet classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 180, + 546, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 546, + 418 + ], + "type": "text", + "content": "Adversarily robust model: Similar to [61], we use an adversarily robust model to find spurious features in ImageNet. The reason for this is that robust models have generative properties [74, 53, 6, 61, 13] in the sense that maximizing the predicted probability of a class in a neighborhood of an image leads to semantically meaningful changes. They also have more informative gradients [22] and thus attribution maps such as GradCAM [57] work better. We use the multiple-norm robust model of [13] as they claim it has the best generative properties. The generative properties of robust models are mainly used for the neural PCA feature visualization where we maximize the NPCA component of a class starting from a gray image. If a spurious feature appears without the class object, this is a strong indicator of a harmful spurious feature. A non-robust model would only produce semantically meaningless adversarial noise, hence we need a robust model for this part of our detection pipeline. Our detected spurious features are not specific to the robust model. We show that SOTA ImageNet models share the same spurious features (Fig. 4 and Sec. 7.2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "text", + "content": "Class-wise neural PCA: Let " + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})_{i = 1}^{N}" + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "text", + "content": " be the training set, where " + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "inline_equation", + "content": "y_{i}\\in \\{1,\\dots ,K\\}" + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "text", + "content": " is the number of classes. We consider features of the penultimate layer " + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\phi (x)\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "text", + "content": " of a neural network for an input " + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "text", + "content": ". For a given class " + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "text", + "content": " and its associated weights " + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "inline_equation", + "content": "w_{k}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 304, + 419, + 545, + 479 + ], + "type": "text", + "content": " in the final layer, we define" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 381, + 485, + 545, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 485, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 381, + 485, + 545, + 497 + ], + "type": "interline_equation", + "content": "\\psi_ {k} (x) = w _ {k} \\odot \\phi (x). \\tag {1}", + "image_path": "7f3424870c732d901ae2cc13ae5df4050a0c0c9fa849946caf9abb9b4c0d7db1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "text", + "content": " is the componentwise product. Let " + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "inline_equation", + "content": "b\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "text", + "content": " be the bias vector of the final layer then the logit " + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "inline_equation", + "content": "f_{k}" + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 502, + 545, + 527 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 531, + 545, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 545, + 554 + ], + "type": "interline_equation", + "content": "f _ {k} (x) = \\sum_ {j = 1} ^ {D} w _ {k j} \\phi (x) _ {j} + b _ {k} = \\langle \\mathbf {1}, \\psi_ {k} (x) \\rangle + b _ {k}. \\tag {2}", + "image_path": "f486ff289218e4c5d1f894f60f36ac3af28d1615daa924e9b54332f71cf0011a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "inline_equation", + "content": "I_{k}" + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "content": " be the index set of the training set of class " + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "inline_equation", + "content": "\\bar{\\psi}_k" + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "content": " the class-wise mean, " + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "inline_equation", + "content": "\\bar{\\psi}_k = \\frac{1}{|I_k|}\\sum_{s\\in I_k}\\psi_k(x_s)" + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "content": ". The class-wise neural PCA allows us to identify variations in the set " + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "inline_equation", + "content": "\\{\\psi (x_r)\\}_{r\\in I_k}" + }, + { + "bbox": [ + 304, + 558, + 545, + 630 + ], + "type": "text", + "content": " arising due to small subpopulations in the training set. In the class-wise neural PCA, we compute eigenvectors of the class-wise covariance matrix," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 635, + 545, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 635, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 337, + 635, + 545, + 661 + ], + "type": "interline_equation", + "content": "C = \\sum_ {s \\in I _ {k}} \\left(\\psi_ {k} \\left(x _ {s}\\right) - \\bar {\\psi} _ {k}\\right) \\left(\\psi_ {k} \\left(x _ {s}\\right) - \\bar {\\psi} _ {k}\\right) ^ {T}. \\tag {3}", + "image_path": "358c7d470207250b23b37a7e2d38977fb3ec469a07b6ba2a3579406c5d32e77b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 666, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 690 + ], + "type": "text", + "content": "The eigenvectors, " + }, + { + "bbox": [ + 304, + 666, + 545, + 690 + ], + "type": "inline_equation", + "content": "v_{1},\\ldots ,v_{D}" + }, + { + "bbox": [ + 304, + 666, + 545, + 690 + ], + "type": "text", + "content": " form an orthonormal basis of " + }, + { + "bbox": [ + 304, + 666, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 666, + 545, + 690 + ], + "type": "text", + "content": " and we write " + }, + { + "bbox": [ + 304, + 666, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\psi_k(x) - \\bar{\\psi}_k" + }, + { + "bbox": [ + 304, + 666, + 545, + 690 + ], + "type": "text", + "content": " in this basis," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 695, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 695, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 336, + 695, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\psi_ {k} (x) - \\bar {\\psi} _ {k} = \\sum_ {l = 1} ^ {D} v _ {l} \\left\\langle \\psi_ {k} (x) - \\bar {\\psi} _ {k}, v _ {l} \\right\\rangle , \\tag {4}", + "image_path": "55532fd446c7bd113c9da98a76fc77d7f0eaee94966b7eca37358a0af797546a.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20237" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 83, + 286, + 129 + ], + "blocks": [ + { + "bbox": [ + 96, + 71, + 248, + 82 + ], + "lines": [ + { + "bbox": [ + 96, + 71, + 248, + 82 + ], + "spans": [ + { + "bbox": [ + 96, + 71, + 248, + 82 + ], + "type": "text", + "content": "Top 5 neural PCA components (ours)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 94, + 59, + 120 + ], + "lines": [ + { + "bbox": [ + 49, + 94, + 59, + 120 + ], + "spans": [ + { + "bbox": [ + 49, + 94, + 59, + 120 + ], + "type": "text", + "content": "NNE" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 83, + 286, + 129 + ], + "lines": [ + { + "bbox": [ + 60, + 83, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 60, + 83, + 286, + 129 + ], + "type": "image", + "image_path": "28e7dffbbe321d2281ad33a08966585590676aa800f49524184302807e2e3714.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 130, + 286, + 177 + ], + "blocks": [ + { + "bbox": [ + 50, + 134, + 59, + 175 + ], + "lines": [ + { + "bbox": [ + 50, + 134, + 59, + 175 + ], + "spans": [ + { + "bbox": [ + 50, + 134, + 59, + 175 + ], + "type": "text", + "content": "Max. act." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 130, + 286, + 177 + ], + "lines": [ + { + "bbox": [ + 60, + 130, + 286, + 177 + ], + "spans": [ + { + "bbox": [ + 60, + 130, + 286, + 177 + ], + "type": "image", + "image_path": "5e7a194e1a7bc8b42a74bddc476398b68002e531e40dc608a4776ecce579427c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 178, + 286, + 225 + ], + "blocks": [ + { + "bbox": [ + 50, + 180, + 59, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 180, + 59, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 180, + 59, + 220 + ], + "type": "text", + "content": "Heeannnnnne" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 178, + 286, + 225 + ], + "lines": [ + { + "bbox": [ + 60, + 178, + 286, + 225 + ], + "spans": [ + { + "bbox": [ + 60, + 178, + 286, + 225 + ], + "type": "image", + "image_path": "4d7988fe9f7416a4b9e2570316c78aa88f91ffed005e878461d29d50ad061318.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 288, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 333 + ], + "type": "text", + "content": "Figure 3: Top 5 Neural PCA components for class hummingbird: first row shows our neural PCA feature visualization (NPFV), second row shows four most activating training images of each NPCA component and the last row GradCAM for the NPCA component. Our NPCA components capture different subpopulations in the training data. Comp. 2 is identified as spurious feature \"bird feeder\", see NPFV and most activating training images (see also Fig. 4)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 346, + 92, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 92, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 92, + 357 + ], + "type": "text", + "content": "and define" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 92, + 365, + 287, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 365, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 92, + 365, + 287, + 382 + ], + "type": "interline_equation", + "content": "\\alpha_ {l} ^ {(k)} (x) = \\left\\langle \\mathbf {1}, v _ {l} \\right\\rangle \\left\\langle \\psi_ {k} (x) - \\bar {\\psi} _ {k}, v _ {l} \\right\\rangle , \\tag {5}", + "image_path": "45a0d6b77162e3be74e117409563f7582f11c32663755fc73b667b9cc3d0c770.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 389, + 268, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 268, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 268, + 402 + ], + "type": "text", + "content": "The logit " + }, + { + "bbox": [ + 47, + 389, + 268, + 402 + ], + "type": "inline_equation", + "content": "f_{k}(x)" + }, + { + "bbox": [ + 47, + 389, + 268, + 402 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 47, + 389, + 268, + 402 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 389, + 268, + 402 + ], + "type": "text", + "content": "-th class can then be written as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 83, + 411, + 287, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 411, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 83, + 411, + 287, + 433 + ], + "type": "interline_equation", + "content": "f _ {k} (x) = \\sum_ {l = 1} ^ {D} \\alpha_ {l} ^ {(k)} (x) + \\left\\langle \\mathbf {1}, \\bar {\\psi} _ {k} \\right\\rangle + b _ {k}. \\tag {6}", + "image_path": "05a678f3681ee0a2d9c3e5fc6f1230bd48e38e2489ec67190c003dd657fe4231.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "text", + "content": "Thus for a given " + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "text", + "content": ", we can interpret " + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\alpha_l^{(k)}(x)" + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "text", + "content": " as the contribution of the neural PCA component " + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "text", + "content": " to the logit " + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "inline_equation", + "content": "f_k(x)" + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "text", + "content": " since the term " + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{1}, \\bar{\\psi}_k \\rangle + b_k" + }, + { + "bbox": [ + 46, + 442, + 287, + 504 + ], + "type": "text", + "content": " is constant for all inputs. Based on this we introduce a mitigation technique for spurious features without retraining in Sec. 5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "text", + "content": "Neural PCA Feature Visualization: To identify semantic features corresponding to our neural PCA component " + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "text", + "content": ", we show the training images which attain the maximal values of " + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{(k)}(x)" + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "text", + "content": ". Additionally, we generate an image " + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "inline_equation", + "content": "z_{l}^{(k)}" + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "text", + "content": ", which we call the Neural PCA Feature Visualization (NPFV) of feature " + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "text", + "content": ", by maximizing " + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{(k)}(x)" + }, + { + "bbox": [ + 47, + 515, + 287, + 594 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 85, + 603, + 249, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 603, + 249, + 621 + ], + "spans": [ + { + "bbox": [ + 85, + 603, + 249, + 621 + ], + "type": "interline_equation", + "content": "z _ {l} ^ {(k)} = g + \\operatorname * {a r g m a x} _ {\\| \\delta \\| _ {2} \\leq \\epsilon} \\alpha_ {l} ^ {(k)} (g + \\delta),", + "image_path": "dae3f23fe5c822498e72e29e88fed225bd803e08a1f54957978cd8f467ddcbaa.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": " is a gray image (all channels equal to 0.5). Thus we maximize the feature " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{(k)}" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": " outgoing from a non-informative and unbiased initialization " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": ". The optimization problem is solved using adaptive projected gradient descent (APGD) [16] with 200 steps. The budget for changes, " + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon = 30" + }, + { + "bbox": [ + 46, + 628, + 288, + 714 + ], + "type": "text", + "content": ", is small to avoid the overactivation of feature attacks maximizing the output of individual neurons [21, 63, 61], see" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "Fig. 8. In Fig. 4 we show for each identified spurious feature, the corresponding NPFV, e.g., for \"hummingbird\" one can see the bird feeder but no hummingbird. The NPCA components together with the maximally activating training images are in principle sufficient to identify spurious features, but the NPFV is very useful to judge how harmful a spurious feature is. Therefore, we use an adversarily robust model, since a non-robust model would yield semantically meaningless adversarial noise as NPFV." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 180, + 546, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 546, + 397 + ], + "type": "text", + "content": "Selection of neural PCA components for human inspection: The penultimate layer of the robust ResNet50 we are using has 2048 neurons. Thus it is infeasible (and unnecessary) to investigate all neural PCA components. A strong criterion that one has found a harmful spurious feature is i) the NPFV shows mainly the spurious feature and not the class, and ii) the NPFV has high confidence. If ii) is not satisfied, then the NPFV is a spurious feature the classifier may have picked up, but it is not harmful in the sense that this feature alone causes the classifier to choose that class. Moreover, we noticed that the eigenvalues of the neural PCA, and the corresponding " + }, + { + "bbox": [ + 304, + 180, + 546, + 397 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 180, + 546, + 397 + ], + "type": "text", + "content": " values, decay quickly. Thus we compute the NPFV for the top 128 neural PCA components (having maximal variance) and then select the ten components which realize the highest confidence for their NPFV in the corresponding class. Note that we do not optimize the confidence when generating the NPFV but only " + }, + { + "bbox": [ + 304, + 180, + 546, + 397 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{(k)}(x)" + }, + { + "bbox": [ + 304, + 180, + 546, + 397 + ], + "type": "text", + "content": " which is part of the logit of the " + }, + { + "bbox": [ + 304, + 180, + 546, + 397 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 180, + 546, + 397 + ], + "type": "text", + "content": "-th class." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "text", + "content": "Identification of spurious neural PCA components via human supervision: For each ImageNet class " + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "text", + "content": " we show the human labeler the top 10 components. For each component " + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "text", + "content": " we show the NPFV " + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "inline_equation", + "content": "z_{l}^{(k)}" + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "text", + "content": " and the 5 training images " + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "text", + "content": " with the largest values of " + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{(k)}(x_{r})" + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "text", + "content": ". Moreover, we compute GradCAM [57] images for the NPFV and the five training images using the NPCA component " + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{(k)}" + }, + { + "bbox": [ + 304, + 399, + 546, + 632 + ], + "type": "text", + "content": " as score. The human marks a component as spurious if i) the NPFV shows dominantly an object not belonging to the class ii) the five training images show consistently this object, iii) the GradCAM activations are primarily not on the class object. The setup shown to the human labeler can be seen in App. B. The labeling of one class takes on average about 45 seconds, so the full labeling of all ImageNet classes took about 13 hours. The human labeler (researcher in machine learning) found in total 337 spurious components. Another human labeler checked all of them and removed spurious features in case of disagreement, resulting in 322 spurious features from 230 ImageNet classes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 644, + 525, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 644, + 525, + 658 + ], + "spans": [ + { + "bbox": [ + 305, + 644, + 525, + 658 + ], + "type": "text", + "content": "5. SpuFix - Mitigation of spurious features" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Once the spurious features are identified, the question is how one can mitigate that the classifier relies on them. One way is to identify the training images containing the spurious feature and then discard or downweight them during" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20238" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 83, + 97, + 131 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 282, + 82 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 282, + 82 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 282, + 82 + ], + "type": "text", + "content": "Hummingbird - Random train. images (confidence / " + }, + { + "bbox": [ + 49, + 71, + 282, + 82 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 49, + 71, + 282, + 82 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 83, + 97, + 131 + ], + "lines": [ + { + "bbox": [ + 49, + 83, + 97, + 131 + ], + "spans": [ + { + "bbox": [ + 49, + 83, + 97, + 131 + ], + "type": "image", + "image_path": "b14bf69b497362b8af6a3e938c5160b68291741a643fcbb5b3878d8411f4b9c6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 133, + 94, + 143 + ], + "lines": [ + { + "bbox": [ + 53, + 133, + 94, + 143 + ], + "spans": [ + { + "bbox": [ + 53, + 133, + 94, + 143 + ], + "type": "text", + "content": "0.93/1.7" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 98, + 83, + 146, + 131 + ], + "blocks": [ + { + "bbox": [ + 98, + 83, + 146, + 131 + ], + "lines": [ + { + "bbox": [ + 98, + 83, + 146, + 131 + ], + "spans": [ + { + "bbox": [ + 98, + 83, + 146, + 131 + ], + "type": "image", + "image_path": "bc0b28bfc916a5957df536af5be3e4e59f8da09d64090eb1ce52be150f1eefc1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 133, + 196, + 144 + ], + "lines": [ + { + "bbox": [ + 99, + 133, + 196, + 144 + ], + "spans": [ + { + "bbox": [ + 99, + 133, + 196, + 144 + ], + "type": "text", + "content": "1.00/-0.9 0.96/-1.0" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 147, + 83, + 196, + 131 + ], + "blocks": [ + { + "bbox": [ + 147, + 83, + 196, + 131 + ], + "lines": [ + { + "bbox": [ + 147, + 83, + 196, + 131 + ], + "spans": [ + { + "bbox": [ + 147, + 83, + 196, + 131 + ], + "type": "image", + "image_path": "42061f85986488c370fe1403148bc244f77aae006c86035a5278b75e93aec6f1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 147, + 157, + 196, + 205 + ], + "blocks": [ + { + "bbox": [ + 100, + 145, + 292, + 156 + ], + "lines": [ + { + "bbox": [ + 100, + 145, + 292, + 156 + ], + "spans": [ + { + "bbox": [ + 100, + 145, + 292, + 156 + ], + "type": "text", + "content": "Max. activating train. images - NPCA Comp. 2" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 147, + 157, + 196, + 205 + ], + "lines": [ + { + "bbox": [ + 147, + 157, + 196, + 205 + ], + "spans": [ + { + "bbox": [ + 147, + 157, + 196, + 205 + ], + "type": "image", + "image_path": "564e85099f5fcb75517ed03ec10f9e2ac1281b1b7e7912168419afa7c82aa0cc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 208, + 190, + 218 + ], + "lines": [ + { + "bbox": [ + 151, + 208, + 190, + 218 + ], + "spans": [ + { + "bbox": [ + 151, + 208, + 190, + 218 + ], + "type": "text", + "content": "1.00/5.9" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 196, + 83, + 245, + 131 + ], + "blocks": [ + { + "bbox": [ + 196, + 83, + 245, + 131 + ], + "lines": [ + { + "bbox": [ + 196, + 83, + 245, + 131 + ], + "spans": [ + { + "bbox": [ + 196, + 83, + 245, + 131 + ], + "type": "image", + "image_path": "0bf1c17e1415b8b5a086a0934782e84692922b518147e332cea369bafebf3ab8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 133, + 292, + 144 + ], + "lines": [ + { + "bbox": [ + 200, + 133, + 292, + 144 + ], + "spans": [ + { + "bbox": [ + 200, + 133, + 292, + 144 + ], + "type": "text", + "content": "0.99/2.2 1.00/1.54" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 196, + 158, + 245, + 205 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 158, + 245, + 205 + ], + "spans": [ + { + "bbox": [ + 196, + 158, + 245, + 205 + ], + "type": "image", + "image_path": "33acc809fb7ba8b787e503c8d6330ef958485f44472545609b4764b6ce73a32e.jpg" + } + ] + } + ], + "index": 12, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 245, + 83, + 294, + 131 + ], + "blocks": [ + { + "bbox": [ + 245, + 83, + 294, + 131 + ], + "lines": [ + { + "bbox": [ + 245, + 83, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 245, + 83, + 294, + 131 + ], + "type": "image", + "image_path": "73649e84b888d8a170558282f72b9119f8421478e23746fb0ce9ef2c06dbe182.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 249, + 133, + 292, + 144 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 133, + 292, + 144 + ], + "spans": [ + { + "bbox": [ + 249, + 133, + 292, + 144 + ], + "type": "text", + "content": "1.00/1.54" + } + ] + } + ], + "index": 15, + "type": "text" + }, + { + "bbox": [ + 247, + 145, + 292, + 156 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 145, + 292, + 156 + ], + "spans": [ + { + "bbox": [ + 247, + 145, + 292, + 156 + ], + "type": "text", + "content": "A Comp. 2" + } + ] + } + ], + "index": 16, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 158, + 97, + 205 + ], + "blocks": [ + { + "bbox": [ + 56, + 145, + 92, + 155 + ], + "lines": [ + { + "bbox": [ + 56, + 145, + 92, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 92, + 155 + ], + "type": "text", + "content": "NPFV-2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 158, + 97, + 205 + ], + "lines": [ + { + "bbox": [ + 49, + 158, + 97, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 158, + 97, + 205 + ], + "type": "image", + "image_path": "ee9abd0b3cf89ae4ab7540cd4d1cd641be18143c2c94b6ad93311dc845855134.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 208, + 94, + 218 + ], + "lines": [ + { + "bbox": [ + 54, + 208, + 94, + 218 + ], + "spans": [ + { + "bbox": [ + 54, + 208, + 94, + 218 + ], + "type": "text", + "content": "1.00/9.7" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 98, + 158, + 146, + 205 + ], + "blocks": [ + { + "bbox": [ + 98, + 158, + 146, + 205 + ], + "lines": [ + { + "bbox": [ + 98, + 158, + 146, + 205 + ], + "spans": [ + { + "bbox": [ + 98, + 158, + 146, + 205 + ], + "type": "image", + "image_path": "55e73764a49276f04235ea41a34d73c5ff3ea4a2234cfa7f10a29b5f3038a146.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 208, + 143, + 218 + ], + "lines": [ + { + "bbox": [ + 103, + 208, + 143, + 218 + ], + "spans": [ + { + "bbox": [ + 103, + 208, + 143, + 218 + ], + "type": "text", + "content": "1.00/7.5" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 147, + 158, + 196, + 205 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 158, + 196, + 205 + ], + "spans": [ + { + "bbox": [ + 147, + 158, + 196, + 205 + ], + "type": "image", + "image_path": "27b340f79544bc8e5c9c57b3f860b4ca9452ea6d6903278b25920058113d76b6.jpg" + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "bbox": [ + 152, + 208, + 190, + 218 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 208, + 190, + 218 + ], + "spans": [ + { + "bbox": [ + 152, + 208, + 190, + 218 + ], + "type": "text", + "content": "1.00/5.9" + } + ] + } + ], + "index": 22, + "type": "text" + }, + { + "bbox": [ + 196, + 158, + 245, + 205 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 158, + 245, + 205 + ], + "spans": [ + { + "bbox": [ + 196, + 158, + 245, + 205 + ], + "type": "image", + "image_path": "33acc809fb7ba8b787e503c8d6330ef958485f44472545609b4764b6ce73a32e.jpg" + } + ] + } + ], + "index": 23, + "type": "text" + }, + { + "bbox": [ + 201, + 209, + 241, + 218 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 209, + 241, + 218 + ], + "spans": [ + { + "bbox": [ + 201, + 209, + 241, + 218 + ], + "type": "text", + "content": "1.00/5.6" + } + ] + } + ], + "index": 24, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 245, + 158, + 294, + 205 + ], + "blocks": [ + { + "bbox": [ + 201, + 208, + 241, + 218 + ], + "lines": [ + { + "bbox": [ + 201, + 208, + 241, + 218 + ], + "spans": [ + { + "bbox": [ + 201, + 208, + 241, + 218 + ], + "type": "text", + "content": "1.00/5.6" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 245, + 158, + 294, + 205 + ], + "lines": [ + { + "bbox": [ + 245, + 158, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 245, + 158, + 294, + 205 + ], + "type": "image", + "image_path": "3d2ff902933bd3e2b9f8becf427f2361c4641ab084f16473623a2518546840ba.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 209, + 290, + 218 + ], + "lines": [ + { + "bbox": [ + 250, + 209, + 290, + 218 + ], + "spans": [ + { + "bbox": [ + 250, + 209, + 290, + 218 + ], + "type": "text", + "content": "1.00/5.6" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 49, + 232, + 97, + 280 + ], + "blocks": [ + { + "bbox": [ + 49, + 220, + 269, + 232 + ], + "lines": [ + { + "bbox": [ + 49, + 220, + 269, + 232 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 269, + 232 + ], + "type": "text", + "content": "Freight car - Random train. images (confidence / " + }, + { + "bbox": [ + 49, + 220, + 269, + 232 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 49, + 220, + 269, + 232 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 232, + 97, + 280 + ], + "lines": [ + { + "bbox": [ + 49, + 232, + 97, + 280 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 97, + 280 + ], + "type": "image", + "image_path": "4e979465e870cbdf4b8c413f94e50d01f2cc1144bf0a83b1fa561a53ba386bd3.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 282, + 94, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 282, + 94, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 282, + 94, + 293 + ], + "type": "text", + "content": "1.00/7.7" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 98, + 232, + 146, + 280 + ], + "blocks": [ + { + "bbox": [ + 98, + 232, + 146, + 280 + ], + "lines": [ + { + "bbox": [ + 98, + 232, + 146, + 280 + ], + "spans": [ + { + "bbox": [ + 98, + 232, + 146, + 280 + ], + "type": "image", + "image_path": "9abcadc71d9090855e26081cec5d5e52e7d20d3a6f1164ae18f32d15e63b0936.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 282, + 143, + 293 + ], + "lines": [ + { + "bbox": [ + 103, + 282, + 143, + 293 + ], + "spans": [ + { + "bbox": [ + 103, + 282, + 143, + 293 + ], + "type": "text", + "content": "1.00/3.4" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 147, + 232, + 196, + 280 + ], + "blocks": [ + { + "bbox": [ + 147, + 232, + 196, + 280 + ], + "lines": [ + { + "bbox": [ + 147, + 232, + 196, + 280 + ], + "spans": [ + { + "bbox": [ + 147, + 232, + 196, + 280 + ], + "type": "image", + "image_path": "bfab20a774e62a2065a524bccb6bbb1005463907e4162fdbf08d4e3b7c979fdf.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 282, + 190, + 293 + ], + "lines": [ + { + "bbox": [ + 152, + 282, + 190, + 293 + ], + "spans": [ + { + "bbox": [ + 152, + 282, + 190, + 293 + ], + "type": "text", + "content": "1.00/6.3" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 196, + 232, + 245, + 280 + ], + "blocks": [ + { + "bbox": [ + 196, + 232, + 245, + 280 + ], + "lines": [ + { + "bbox": [ + 196, + 232, + 245, + 280 + ], + "spans": [ + { + "bbox": [ + 196, + 232, + 245, + 280 + ], + "type": "image", + "image_path": "c9d9314d3941c262721f2249d8ba298e2cd439a066dab189d9e00c9f2f64d789.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 282, + 245, + 293 + ], + "lines": [ + { + "bbox": [ + 200, + 282, + 245, + 293 + ], + "spans": [ + { + "bbox": [ + 200, + 282, + 245, + 293 + ], + "type": "text", + "content": "0.98/-0.8" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 245, + 232, + 294, + 280 + ], + "blocks": [ + { + "bbox": [ + 245, + 232, + 294, + 280 + ], + "lines": [ + { + "bbox": [ + 245, + 232, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 245, + 232, + 294, + 280 + ], + "type": "image", + "image_path": "d39e205b71a0709e2c941d21fbb09aaaea7fe6bfacd5ee5685853c2f541637cf.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 282, + 290, + 293 + ], + "lines": [ + { + "bbox": [ + 249, + 282, + 290, + 293 + ], + "spans": [ + { + "bbox": [ + 249, + 282, + 290, + 293 + ], + "type": "text", + "content": "1.00/4.2" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 49, + 307, + 97, + 354 + ], + "blocks": [ + { + "bbox": [ + 56, + 294, + 91, + 304 + ], + "lines": [ + { + "bbox": [ + 56, + 294, + 91, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 91, + 304 + ], + "type": "text", + "content": "NPFV-1" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 307, + 97, + 354 + ], + "lines": [ + { + "bbox": [ + 49, + 307, + 97, + 354 + ], + "spans": [ + { + "bbox": [ + 49, + 307, + 97, + 354 + ], + "type": "image", + "image_path": "b37ebabe2a3afb713344d4567307ea6abea8d8e96a9f3b6e78002694ff20e621.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 357, + 95, + 367 + ], + "lines": [ + { + "bbox": [ + 53, + 357, + 95, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 95, + 367 + ], + "type": "text", + "content": "1.00/12.1" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 98, + 307, + 147, + 354 + ], + "blocks": [ + { + "bbox": [ + 100, + 294, + 290, + 306 + ], + "lines": [ + { + "bbox": [ + 100, + 294, + 290, + 306 + ], + "spans": [ + { + "bbox": [ + 100, + 294, + 290, + 306 + ], + "type": "text", + "content": "Max. activating train. images - NPCA Comp. 1" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 98, + 307, + 147, + 354 + ], + "lines": [ + { + "bbox": [ + 98, + 307, + 147, + 354 + ], + "spans": [ + { + "bbox": [ + 98, + 307, + 147, + 354 + ], + "type": "image", + "image_path": "fcf21de7695b8da5ddc4d84e03a64bb687d7a503f9ce1ca22201048369cd0429.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 357, + 145, + 367 + ], + "lines": [ + { + "bbox": [ + 102, + 357, + 145, + 367 + ], + "spans": [ + { + "bbox": [ + 102, + 357, + 145, + 367 + ], + "type": "text", + "content": "1.00/10.9" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 147, + 307, + 196, + 354 + ], + "blocks": [ + { + "bbox": [ + 147, + 307, + 196, + 354 + ], + "lines": [ + { + "bbox": [ + 147, + 307, + 196, + 354 + ], + "spans": [ + { + "bbox": [ + 147, + 307, + 196, + 354 + ], + "type": "image", + "image_path": "23f8319e0d23f30392514d7c469a3d05e68f2dec9dadd3f632a11759cf32069e.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 357, + 194, + 367 + ], + "lines": [ + { + "bbox": [ + 151, + 357, + 194, + 367 + ], + "spans": [ + { + "bbox": [ + 151, + 357, + 194, + 367 + ], + "type": "text", + "content": "1.00/10.4" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 196, + 307, + 245, + 354 + ], + "blocks": [ + { + "bbox": [ + 196, + 307, + 245, + 354 + ], + "lines": [ + { + "bbox": [ + 196, + 307, + 245, + 354 + ], + "spans": [ + { + "bbox": [ + 196, + 307, + 245, + 354 + ], + "type": "image", + "image_path": "106f86f1b02a225d4b39bb2158a317c5f3d170608b0545500c50bae7bbb27d5a.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 357, + 244, + 367 + ], + "lines": [ + { + "bbox": [ + 200, + 357, + 244, + 367 + ], + "spans": [ + { + "bbox": [ + 200, + 357, + 244, + 367 + ], + "type": "text", + "content": "1.00/10.2" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 245, + 307, + 294, + 354 + ], + "blocks": [ + { + "bbox": [ + 245, + 307, + 294, + 354 + ], + "lines": [ + { + "bbox": [ + 245, + 307, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 245, + 307, + 294, + 354 + ], + "type": "image", + "image_path": "b9b526175825f8302e0707717326217793a880498bb8ae8553e22713ebedd99d.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 357, + 291, + 367 + ], + "lines": [ + { + "bbox": [ + 249, + 357, + 291, + 367 + ], + "spans": [ + { + "bbox": [ + 249, + 357, + 291, + 367 + ], + "type": "text", + "content": "1.00/10.2" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 49, + 381, + 97, + 430 + ], + "blocks": [ + { + "bbox": [ + 49, + 370, + 245, + 380 + ], + "lines": [ + { + "bbox": [ + 49, + 370, + 245, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 370, + 245, + 380 + ], + "type": "text", + "content": "Koala - Random train. images (confidence / " + }, + { + "bbox": [ + 49, + 370, + 245, + 380 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 49, + 370, + 245, + 380 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 381, + 97, + 430 + ], + "lines": [ + { + "bbox": [ + 49, + 381, + 97, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 381, + 97, + 430 + ], + "type": "image", + "image_path": "2542e369ce891f1da3a82d5484e8df2745e02c6b49db9f61e0f2089dd515db28.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 433, + 96, + 442 + ], + "lines": [ + { + "bbox": [ + 53, + 433, + 96, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 96, + 442 + ], + "type": "text", + "content": "1.00/0.77" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 98, + 381, + 147, + 430 + ], + "blocks": [ + { + "bbox": [ + 98, + 381, + 147, + 430 + ], + "lines": [ + { + "bbox": [ + 98, + 381, + 147, + 430 + ], + "spans": [ + { + "bbox": [ + 98, + 381, + 147, + 430 + ], + "type": "image", + "image_path": "099c91b963703a915f76eda313baf1f8e8a46cdf32e90a5e52ef889e16630a38.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 433, + 142, + 442 + ], + "lines": [ + { + "bbox": [ + 102, + 433, + 142, + 442 + ], + "spans": [ + { + "bbox": [ + 102, + 433, + 142, + 442 + ], + "type": "text", + "content": "0.87/2.4" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 147, + 381, + 196, + 430 + ], + "blocks": [ + { + "bbox": [ + 147, + 381, + 196, + 430 + ], + "lines": [ + { + "bbox": [ + 147, + 381, + 196, + 430 + ], + "spans": [ + { + "bbox": [ + 147, + 381, + 196, + 430 + ], + "type": "image", + "image_path": "43de58373554e8da48d3e2bc2058799b89fb568a5016021084561b503b60e9e7.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 433, + 190, + 442 + ], + "lines": [ + { + "bbox": [ + 152, + 433, + 190, + 442 + ], + "spans": [ + { + "bbox": [ + 152, + 433, + 190, + 442 + ], + "type": "text", + "content": "1.00/0.4" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 196, + 381, + 245, + 430 + ], + "blocks": [ + { + "bbox": [ + 196, + 381, + 245, + 430 + ], + "lines": [ + { + "bbox": [ + 196, + 381, + 245, + 430 + ], + "spans": [ + { + "bbox": [ + 196, + 381, + 245, + 430 + ], + "type": "image", + "image_path": "7b2a0498eabd530d3c6f7b82940e61a1a2509f790afa1a19cb7a5a0996bcd07b.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 433, + 241, + 442 + ], + "lines": [ + { + "bbox": [ + 201, + 433, + 241, + 442 + ], + "spans": [ + { + "bbox": [ + 201, + 433, + 241, + 442 + ], + "type": "text", + "content": "1.00/0.0" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 245, + 381, + 294, + 430 + ], + "blocks": [ + { + "bbox": [ + 245, + 381, + 294, + 430 + ], + "lines": [ + { + "bbox": [ + 245, + 381, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 245, + 381, + 294, + 430 + ], + "type": "image", + "image_path": "38c8d0820678ff378e6b0cd711b93ea5ac9053e887f130c0f3bbd8bcadfeeb3e.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 433, + 290, + 442 + ], + "lines": [ + { + "bbox": [ + 249, + 433, + 290, + 442 + ], + "spans": [ + { + "bbox": [ + 249, + 433, + 290, + 442 + ], + "type": "text", + "content": "0.95/0.5" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 49, + 456, + 97, + 504 + ], + "blocks": [ + { + "bbox": [ + 56, + 445, + 91, + 454 + ], + "lines": [ + { + "bbox": [ + 56, + 445, + 91, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 91, + 454 + ], + "type": "text", + "content": "NPFV-3" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 456, + 97, + 504 + ], + "lines": [ + { + "bbox": [ + 49, + 456, + 97, + 504 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 97, + 504 + ], + "type": "image", + "image_path": "7e0f88858e23beed70e24a8579ae8f549b6b8e5498604d4e5c3464bc1a3ccccd.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 506, + 94, + 517 + ], + "lines": [ + { + "bbox": [ + 54, + 506, + 94, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 506, + 94, + 517 + ], + "type": "text", + "content": "1.00/5.5" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 98, + 456, + 147, + 504 + ], + "blocks": [ + { + "bbox": [ + 101, + 445, + 291, + 456 + ], + "lines": [ + { + "bbox": [ + 101, + 445, + 291, + 456 + ], + "spans": [ + { + "bbox": [ + 101, + 445, + 291, + 456 + ], + "type": "text", + "content": "Max. activating train. images - NPCA Comp. 3" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 98, + 456, + 147, + 504 + ], + "lines": [ + { + "bbox": [ + 98, + 456, + 147, + 504 + ], + "spans": [ + { + "bbox": [ + 98, + 456, + 147, + 504 + ], + "type": "image", + "image_path": "85a9cf365a5d70f37b0937c99fcfd657361c01785806145e0dd65ecb21021fc8.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 507, + 143, + 517 + ], + "lines": [ + { + "bbox": [ + 104, + 507, + 143, + 517 + ], + "spans": [ + { + "bbox": [ + 104, + 507, + 143, + 517 + ], + "type": "text", + "content": "1.00/4.6" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 147, + 456, + 196, + 504 + ], + "blocks": [ + { + "bbox": [ + 147, + 456, + 196, + 504 + ], + "lines": [ + { + "bbox": [ + 147, + 456, + 196, + 504 + ], + "spans": [ + { + "bbox": [ + 147, + 456, + 196, + 504 + ], + "type": "image", + "image_path": "aabeef1685499126ba04e6f2a069bc3f04d0fd96b00998392151e6971559cc56.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 507, + 191, + 517 + ], + "lines": [ + { + "bbox": [ + 153, + 507, + 191, + 517 + ], + "spans": [ + { + "bbox": [ + 153, + 507, + 191, + 517 + ], + "type": "text", + "content": "1.00/4.5" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 196, + 456, + 245, + 505 + ], + "blocks": [ + { + "bbox": [ + 196, + 456, + 245, + 505 + ], + "lines": [ + { + "bbox": [ + 196, + 456, + 245, + 505 + ], + "spans": [ + { + "bbox": [ + 196, + 456, + 245, + 505 + ], + "type": "image", + "image_path": "f37200399a4dd0a1a1bee69188f721d122733fa978bd68c0f3a9f389b0ee62f0.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 507, + 241, + 517 + ], + "lines": [ + { + "bbox": [ + 201, + 507, + 241, + 517 + ], + "spans": [ + { + "bbox": [ + 201, + 507, + 241, + 517 + ], + "type": "text", + "content": "1.00/4.4" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 245, + 456, + 294, + 505 + ], + "blocks": [ + { + "bbox": [ + 245, + 456, + 294, + 505 + ], + "lines": [ + { + "bbox": [ + 245, + 456, + 294, + 505 + ], + "spans": [ + { + "bbox": [ + 245, + 456, + 294, + 505 + ], + "type": "image", + "image_path": "f587b437982882523d294ee93ba63294ec6cc9a6a21d34c28513b19fc46115a1.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 507, + 290, + 517 + ], + "lines": [ + { + "bbox": [ + 249, + 507, + 290, + 517 + ], + "spans": [ + { + "bbox": [ + 249, + 507, + 290, + 517 + ], + "type": "text", + "content": "0.86/4.3" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 49, + 531, + 97, + 579 + ], + "blocks": [ + { + "bbox": [ + 49, + 520, + 257, + 530 + ], + "lines": [ + { + "bbox": [ + 49, + 520, + 257, + 530 + ], + "spans": [ + { + "bbox": [ + 49, + 520, + 257, + 530 + ], + "type": "text", + "content": "Fireboat - Random train. images (confidence / " + }, + { + "bbox": [ + 49, + 520, + 257, + 530 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 49, + 520, + 257, + 530 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 531, + 97, + 579 + ], + "lines": [ + { + "bbox": [ + 49, + 531, + 97, + 579 + ], + "spans": [ + { + "bbox": [ + 49, + 531, + 97, + 579 + ], + "type": "image", + "image_path": "28d10134535d18af7dcf2d59ba7e5d842b09057d7df5bfec65eb4113412512d1.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 582, + 96, + 591 + ], + "lines": [ + { + "bbox": [ + 50, + 582, + 96, + 591 + ], + "spans": [ + { + "bbox": [ + 50, + 582, + 96, + 591 + ], + "type": "text", + "content": "0.88/-1.1" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_caption" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 98, + 531, + 146, + 579 + ], + "blocks": [ + { + "bbox": [ + 98, + 531, + 146, + 579 + ], + "lines": [ + { + "bbox": [ + 98, + 531, + 146, + 579 + ], + "spans": [ + { + "bbox": [ + 98, + 531, + 146, + 579 + ], + "type": "image", + "image_path": "f2ca3287a934247402b3047e7345d2bccbe2defbbc945ea230deb08ee3c3b73c.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 582, + 142, + 591 + ], + "lines": [ + { + "bbox": [ + 102, + 582, + 142, + 591 + ], + "spans": [ + { + "bbox": [ + 102, + 582, + 142, + 591 + ], + "type": "text", + "content": "0.95/0.5" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_caption" + } + ], + "index": 76 + }, + { + "type": "image", + "bbox": [ + 147, + 531, + 196, + 579 + ], + "blocks": [ + { + "bbox": [ + 147, + 531, + 196, + 579 + ], + "lines": [ + { + "bbox": [ + 147, + 531, + 196, + 579 + ], + "spans": [ + { + "bbox": [ + 147, + 531, + 196, + 579 + ], + "type": "image", + "image_path": "10f679df621f8554213f997271b63ad8b6fcf08b6b7632967059e16b4f1d6cbd.jpg" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 582, + 195, + 591 + ], + "lines": [ + { + "bbox": [ + 151, + 582, + 195, + 591 + ], + "spans": [ + { + "bbox": [ + 151, + 582, + 195, + 591 + ], + "type": "text", + "content": "0.12/-0.6" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_caption" + } + ], + "index": 78 + }, + { + "type": "image", + "bbox": [ + 196, + 531, + 245, + 579 + ], + "blocks": [ + { + "bbox": [ + 196, + 531, + 245, + 579 + ], + "lines": [ + { + "bbox": [ + 196, + 531, + 245, + 579 + ], + "spans": [ + { + "bbox": [ + 196, + 531, + 245, + 579 + ], + "type": "image", + "image_path": "8973b07c4a97153d5e9146d29f562ef8abf618d7845386b5b77d31e6f66de1b7.jpg" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 582, + 244, + 591 + ], + "lines": [ + { + "bbox": [ + 200, + 582, + 244, + 591 + ], + "spans": [ + { + "bbox": [ + 200, + 582, + 244, + 591 + ], + "type": "text", + "content": "0.84/-0.2" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_caption" + } + ], + "index": 80 + }, + { + "type": "image", + "bbox": [ + 245, + 531, + 294, + 579 + ], + "blocks": [ + { + "bbox": [ + 245, + 531, + 294, + 579 + ], + "lines": [ + { + "bbox": [ + 245, + 531, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 245, + 531, + 294, + 579 + ], + "type": "image", + "image_path": "36f8cc08837596bca683088b426cd5f0b7b2e11f5e95ee1a3b301586c2d03acf.jpg" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 582, + 291, + 591 + ], + "lines": [ + { + "bbox": [ + 249, + 582, + 291, + 591 + ], + "spans": [ + { + "bbox": [ + 249, + 582, + 291, + 591 + ], + "type": "text", + "content": "0.02/-1.1" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_caption" + } + ], + "index": 82 + }, + { + "type": "image", + "bbox": [ + 49, + 605, + 97, + 653 + ], + "blocks": [ + { + "bbox": [ + 56, + 594, + 91, + 604 + ], + "lines": [ + { + "bbox": [ + 56, + 594, + 91, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 594, + 91, + 604 + ], + "type": "text", + "content": "NPFV-2" + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 605, + 97, + 653 + ], + "lines": [ + { + "bbox": [ + 49, + 605, + 97, + 653 + ], + "spans": [ + { + "bbox": [ + 49, + 605, + 97, + 653 + ], + "type": "image", + "image_path": "fa054c68d416f25b599c682b3410ff02d5cf5b94ff1a664d6c40d47704dad8e1.jpg" + } + ] + } + ], + "index": 86, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 656, + 94, + 666 + ], + "lines": [ + { + "bbox": [ + 54, + 656, + 94, + 666 + ], + "spans": [ + { + "bbox": [ + 54, + 656, + 94, + 666 + ], + "type": "text", + "content": "1.00/5.5" + } + ] + } + ], + "index": 87, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 677, + 547, + 726 + ], + "lines": [ + { + "bbox": [ + 46, + 677, + 547, + 726 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 547, + 726 + ], + "type": "text", + "content": "Figure 4: Spurious features (ImageNet): found by human labeling of our neural PCA components. For each class we show 5 random train. images (top left), the neural PCA Feature Visual. (NPFV) and 4 most activating train. images for the spurious feature component (bottom left). Right: four ImageNet models classify images showing only the spurious feature but no class object as this class." + } + ] + } + ], + "index": 184, + "angle": 0, + "type": "image_caption" + } + ], + "index": 86 + }, + { + "type": "image", + "bbox": [ + 98, + 605, + 147, + 653 + ], + "blocks": [ + { + "bbox": [ + 101, + 594, + 292, + 605 + ], + "lines": [ + { + "bbox": [ + 101, + 594, + 292, + 605 + ], + "spans": [ + { + "bbox": [ + 101, + 594, + 292, + 605 + ], + "type": "text", + "content": "Max. activating train. images - NPCA Comp. 2" + } + ] + } + ], + "index": 85, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 98, + 605, + 147, + 653 + ], + "lines": [ + { + "bbox": [ + 98, + 605, + 147, + 653 + ], + "spans": [ + { + "bbox": [ + 98, + 605, + 147, + 653 + ], + "type": "image", + "image_path": "91890ab85f24c0fe4ff3e4d87cf6586c12f238c95bf345d5210ec281cde8c3e4.jpg" + } + ] + } + ], + "index": 88, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 656, + 142, + 666 + ], + "lines": [ + { + "bbox": [ + 104, + 656, + 142, + 666 + ], + "spans": [ + { + "bbox": [ + 104, + 656, + 142, + 666 + ], + "type": "text", + "content": "0.42/4.1" + } + ] + } + ], + "index": 89, + "angle": 0, + "type": "image_caption" + } + ], + "index": 88 + }, + { + "type": "image", + "bbox": [ + 147, + 605, + 196, + 653 + ], + "blocks": [ + { + "bbox": [ + 147, + 605, + 196, + 653 + ], + "lines": [ + { + "bbox": [ + 147, + 605, + 196, + 653 + ], + "spans": [ + { + "bbox": [ + 147, + 605, + 196, + 653 + ], + "type": "image", + "image_path": "2d85f28e5420f90157fa94845c66fd9fb4d4e510ff66fc405576d654d0a3909a.jpg" + } + ] + } + ], + "index": 90, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 656, + 192, + 666 + ], + "lines": [ + { + "bbox": [ + 153, + 656, + 192, + 666 + ], + "spans": [ + { + "bbox": [ + 153, + 656, + 192, + 666 + ], + "type": "text", + "content": "1.00/4.0" + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "image_caption" + } + ], + "index": 90 + }, + { + "type": "image", + "bbox": [ + 196, + 605, + 245, + 653 + ], + "blocks": [ + { + "bbox": [ + 196, + 605, + 245, + 653 + ], + "lines": [ + { + "bbox": [ + 196, + 605, + 245, + 653 + ], + "spans": [ + { + "bbox": [ + 196, + 605, + 245, + 653 + ], + "type": "image", + "image_path": "93e8be9bbfbdfcbf20a9e73ee64c590f1c0d0f77b4df71860dac4da49d02ac85.jpg" + } + ] + } + ], + "index": 92, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 656, + 241, + 666 + ], + "lines": [ + { + "bbox": [ + 201, + 656, + 241, + 666 + ], + "spans": [ + { + "bbox": [ + 201, + 656, + 241, + 666 + ], + "type": "text", + "content": "1.00/3.9" + } + ] + } + ], + "index": 93, + "angle": 0, + "type": "image_caption" + } + ], + "index": 92 + }, + { + "type": "image", + "bbox": [ + 245, + 605, + 294, + 653 + ], + "blocks": [ + { + "bbox": [ + 245, + 605, + 294, + 653 + ], + "lines": [ + { + "bbox": [ + 245, + 605, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 245, + 605, + 294, + 653 + ], + "type": "image", + "image_path": "a33b3698a9b9bf9ec8d6be3f4ba84af1597e4b1320e4ade818bb8ea9016f2e2a.jpg" + } + ] + } + ], + "index": 94, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 656, + 290, + 666 + ], + "lines": [ + { + "bbox": [ + 249, + 656, + 290, + 666 + ], + "spans": [ + { + "bbox": [ + 249, + 656, + 290, + 666 + ], + "type": "text", + "content": "1.00/3.9" + } + ] + } + ], + "index": 95, + "angle": 0, + "type": "image_caption" + } + ], + "index": 94 + }, + { + "type": "image", + "bbox": [ + 307, + 83, + 355, + 131 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 542, + 82 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 542, + 82 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 542, + 82 + ], + "type": "text", + "content": "Images with spurious bird feeder but no hummingbird" + } + ] + } + ], + "index": 96, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 83, + 355, + 131 + ], + "lines": [ + { + "bbox": [ + 307, + 83, + 355, + 131 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 355, + 131 + ], + "type": "image", + "image_path": "a45101c80cd815b58da987d5c24eb8cc7467832f0b79b8ff6280ecc1cc92d5bf.jpg" + } + ] + } + ], + "index": 97, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 133, + 351, + 143 + ], + "lines": [ + { + "bbox": [ + 310, + 133, + 351, + 143 + ], + "spans": [ + { + "bbox": [ + 310, + 133, + 351, + 143 + ], + "type": "text", + "content": "0.94/5.7" + } + ] + } + ], + "index": 98, + "angle": 0, + "type": "image_caption" + } + ], + "index": 97 + }, + { + "type": "image", + "bbox": [ + 356, + 83, + 404, + 131 + ], + "blocks": [ + { + "bbox": [ + 356, + 83, + 404, + 131 + ], + "lines": [ + { + "bbox": [ + 356, + 83, + 404, + 131 + ], + "spans": [ + { + "bbox": [ + 356, + 83, + 404, + 131 + ], + "type": "image", + "image_path": "4a80f1d75e79d106d0c8c264cc6927f495f0e31e4d252bedaf73e10468a47772.jpg" + } + ] + } + ], + "index": 99, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 133, + 400, + 143 + ], + "lines": [ + { + "bbox": [ + 359, + 133, + 400, + 143 + ], + "spans": [ + { + "bbox": [ + 359, + 133, + 400, + 143 + ], + "type": "text", + "content": "0.94/3.4" + } + ] + } + ], + "index": 100, + "angle": 0, + "type": "image_caption" + } + ], + "index": 99 + }, + { + "type": "image", + "bbox": [ + 405, + 83, + 452, + 131 + ], + "blocks": [ + { + "bbox": [ + 405, + 83, + 452, + 131 + ], + "lines": [ + { + "bbox": [ + 405, + 83, + 452, + 131 + ], + "spans": [ + { + "bbox": [ + 405, + 83, + 452, + 131 + ], + "type": "image", + "image_path": "2125c00b429315f1ae61534ff1e8810e53af39f0e1e0403aa0d738e5382e8de1.jpg" + } + ] + } + ], + "index": 101, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 133, + 449, + 143 + ], + "lines": [ + { + "bbox": [ + 409, + 133, + 449, + 143 + ], + "spans": [ + { + "bbox": [ + 409, + 133, + 449, + 143 + ], + "type": "text", + "content": "0.82/2.9" + } + ] + } + ], + "index": 102, + "angle": 0, + "type": "image_caption" + } + ], + "index": 101 + }, + { + "type": "image", + "bbox": [ + 453, + 83, + 502, + 131 + ], + "blocks": [ + { + "bbox": [ + 453, + 83, + 502, + 131 + ], + "lines": [ + { + "bbox": [ + 453, + 83, + 502, + 131 + ], + "spans": [ + { + "bbox": [ + 453, + 83, + 502, + 131 + ], + "type": "image", + "image_path": "3e511056306d77c041e2d2bedd8de69bccd3cc66a080d4053953d0aa0f6033a3.jpg" + } + ] + } + ], + "index": 103, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 133, + 498, + 143 + ], + "lines": [ + { + "bbox": [ + 458, + 133, + 498, + 143 + ], + "spans": [ + { + "bbox": [ + 458, + 133, + 498, + 143 + ], + "type": "text", + "content": "0.91/5.6" + } + ] + } + ], + "index": 104, + "angle": 0, + "type": "image_caption" + } + ], + "index": 103 + }, + { + "type": "image", + "bbox": [ + 503, + 83, + 551, + 131 + ], + "blocks": [ + { + "bbox": [ + 503, + 83, + 551, + 131 + ], + "lines": [ + { + "bbox": [ + 503, + 83, + 551, + 131 + ], + "spans": [ + { + "bbox": [ + 503, + 83, + 551, + 131 + ], + "type": "image", + "image_path": "06196a4a233b5176ab789b24e32008598fac17ae3656c05b544b2b2a8773a617.jpg" + } + ] + } + ], + "index": 105, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 133, + 547, + 143 + ], + "lines": [ + { + "bbox": [ + 507, + 133, + 547, + 143 + ], + "spans": [ + { + "bbox": [ + 507, + 133, + 547, + 143 + ], + "type": "text", + "content": "0.91/4.7" + } + ] + } + ], + "index": 106, + "angle": 0, + "type": "image_caption" + } + ], + "index": 105 + }, + { + "type": "image", + "bbox": [ + 307, + 157, + 355, + 205 + ], + "blocks": [ + { + "bbox": [ + 313, + 145, + 544, + 156 + ], + "lines": [ + { + "bbox": [ + 313, + 145, + 544, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 544, + 156 + ], + "type": "text", + "content": "all classified as humming bird by four ImageNet models" + } + ] + } + ], + "index": 107, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 157, + 355, + 205 + ], + "lines": [ + { + "bbox": [ + 307, + 157, + 355, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 157, + 355, + 205 + ], + "type": "image", + "image_path": "edc3193bf8177af534bf2d4a751fcb1c5fad50d517725c2ffcac95807adc870a.jpg" + } + ] + } + ], + "index": 108, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 208, + 351, + 218 + ], + "lines": [ + { + "bbox": [ + 310, + 208, + 351, + 218 + ], + "spans": [ + { + "bbox": [ + 310, + 208, + 351, + 218 + ], + "type": "text", + "content": "0.86/4.3" + } + ] + } + ], + "index": 109, + "angle": 0, + "type": "image_caption" + } + ], + "index": 108 + }, + { + "type": "image", + "bbox": [ + 356, + 157, + 404, + 205 + ], + "blocks": [ + { + "bbox": [ + 356, + 157, + 404, + 205 + ], + "lines": [ + { + "bbox": [ + 356, + 157, + 404, + 205 + ], + "spans": [ + { + "bbox": [ + 356, + 157, + 404, + 205 + ], + "type": "image", + "image_path": "ee7284a760a207d058f8128dc592dbb8be7117ac1c127953c23efa56ba4f39e1.jpg" + } + ] + } + ], + "index": 110, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 208, + 400, + 218 + ], + "lines": [ + { + "bbox": [ + 359, + 208, + 400, + 218 + ], + "spans": [ + { + "bbox": [ + 359, + 208, + 400, + 218 + ], + "type": "text", + "content": "0.81/3.5" + } + ] + } + ], + "index": 111, + "angle": 0, + "type": "image_caption" + } + ], + "index": 110 + }, + { + "type": "image", + "bbox": [ + 405, + 157, + 452, + 205 + ], + "blocks": [ + { + "bbox": [ + 405, + 157, + 452, + 205 + ], + "lines": [ + { + "bbox": [ + 405, + 157, + 452, + 205 + ], + "spans": [ + { + "bbox": [ + 405, + 157, + 452, + 205 + ], + "type": "image", + "image_path": "acf64746e60be599c2c47372211bc4e0dd74fa501a30b96ec72507e8aee97826.jpg" + } + ] + } + ], + "index": 112, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 208, + 449, + 218 + ], + "lines": [ + { + "bbox": [ + 409, + 208, + 449, + 218 + ], + "spans": [ + { + "bbox": [ + 409, + 208, + 449, + 218 + ], + "type": "text", + "content": "0.89/3.3" + } + ] + } + ], + "index": 113, + "angle": 0, + "type": "image_caption" + } + ], + "index": 112 + }, + { + "type": "image", + "bbox": [ + 453, + 157, + 502, + 205 + ], + "blocks": [ + { + "bbox": [ + 453, + 157, + 502, + 205 + ], + "lines": [ + { + "bbox": [ + 453, + 157, + 502, + 205 + ], + "spans": [ + { + "bbox": [ + 453, + 157, + 502, + 205 + ], + "type": "image", + "image_path": "2abd1f654cc64097e2ebf3229076a5525b3f04014995084534c3cc5adbddc412.jpg" + } + ] + } + ], + "index": 114, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 208, + 498, + 218 + ], + "lines": [ + { + "bbox": [ + 458, + 208, + 498, + 218 + ], + "spans": [ + { + "bbox": [ + 458, + 208, + 498, + 218 + ], + "type": "text", + "content": "0.89/3.8" + } + ] + } + ], + "index": 115, + "angle": 0, + "type": "image_caption" + } + ], + "index": 114 + }, + { + "type": "image", + "bbox": [ + 503, + 157, + 551, + 205 + ], + "blocks": [ + { + "bbox": [ + 503, + 157, + 551, + 205 + ], + "lines": [ + { + "bbox": [ + 503, + 157, + 551, + 205 + ], + "spans": [ + { + "bbox": [ + 503, + 157, + 551, + 205 + ], + "type": "image", + "image_path": "3607ca6620d48cf89461c71bdc2b4d885649ec1a1ad2ae54d78c1dc9ad27e744.jpg" + } + ] + } + ], + "index": 116, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 209, + 548, + 218 + ], + "lines": [ + { + "bbox": [ + 505, + 209, + 548, + 218 + ], + "spans": [ + { + "bbox": [ + 505, + 209, + 548, + 218 + ], + "type": "text", + "content": "0.78/5.91" + } + ] + } + ], + "index": 117, + "angle": 0, + "type": "image_caption" + } + ], + "index": 116 + }, + { + "type": "image", + "bbox": [ + 307, + 232, + 355, + 280 + ], + "blocks": [ + { + "bbox": [ + 331, + 221, + 526, + 232 + ], + "lines": [ + { + "bbox": [ + 331, + 221, + 526, + 232 + ], + "spans": [ + { + "bbox": [ + 331, + 221, + 526, + 232 + ], + "type": "text", + "content": "Images with spurious grafitti but no freight car" + } + ] + } + ], + "index": 118, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 232, + 355, + 280 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 355, + 280 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 355, + 280 + ], + "type": "image", + "image_path": "262cfde3f1b1b79b83e2e26f16d2506c66fc6a90aeb75608aa7d14554bda8741.jpg" + } + ] + } + ], + "index": 119, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 282, + 351, + 293 + ], + "lines": [ + { + "bbox": [ + 310, + 282, + 351, + 293 + ], + "spans": [ + { + "bbox": [ + 310, + 282, + 351, + 293 + ], + "type": "text", + "content": "0.97/3.5" + } + ] + } + ], + "index": 120, + "angle": 0, + "type": "image_caption" + } + ], + "index": 119 + }, + { + "type": "image", + "bbox": [ + 356, + 232, + 404, + 280 + ], + "blocks": [ + { + "bbox": [ + 356, + 232, + 404, + 280 + ], + "lines": [ + { + "bbox": [ + 356, + 232, + 404, + 280 + ], + "spans": [ + { + "bbox": [ + 356, + 232, + 404, + 280 + ], + "type": "image", + "image_path": "60a9a85f1c8ca0e02b8d4edf4ead405387013f9f69b5020ccbfee9a1b1557f52.jpg" + } + ] + } + ], + "index": 121, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 282, + 400, + 293 + ], + "lines": [ + { + "bbox": [ + 359, + 282, + 400, + 293 + ], + "spans": [ + { + "bbox": [ + 359, + 282, + 400, + 293 + ], + "type": "text", + "content": "0.79/4.0" + } + ] + } + ], + "index": 122, + "angle": 0, + "type": "image_caption" + } + ], + "index": 121 + }, + { + "type": "image", + "bbox": [ + 405, + 232, + 452, + 280 + ], + "blocks": [ + { + "bbox": [ + 405, + 232, + 452, + 280 + ], + "lines": [ + { + "bbox": [ + 405, + 232, + 452, + 280 + ], + "spans": [ + { + "bbox": [ + 405, + 232, + 452, + 280 + ], + "type": "image", + "image_path": "76e56ea2c33cc00427a7e546d8363a91a75072aa98798556b58c29822ae72383.jpg" + } + ] + } + ], + "index": 123, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 282, + 449, + 293 + ], + "lines": [ + { + "bbox": [ + 409, + 282, + 449, + 293 + ], + "spans": [ + { + "bbox": [ + 409, + 282, + 449, + 293 + ], + "type": "text", + "content": "0.81/2.4" + } + ] + } + ], + "index": 124, + "angle": 0, + "type": "image_caption" + } + ], + "index": 123 + }, + { + "type": "image", + "bbox": [ + 453, + 232, + 502, + 280 + ], + "blocks": [ + { + "bbox": [ + 453, + 232, + 502, + 280 + ], + "lines": [ + { + "bbox": [ + 453, + 232, + 502, + 280 + ], + "spans": [ + { + "bbox": [ + 453, + 232, + 502, + 280 + ], + "type": "image", + "image_path": "57dee91e1490077c8b1dd710f3f532ea5811d60104ceef2245d5236238366c74.jpg" + } + ] + } + ], + "index": 125, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 282, + 498, + 293 + ], + "lines": [ + { + "bbox": [ + 458, + 282, + 498, + 293 + ], + "spans": [ + { + "bbox": [ + 458, + 282, + 498, + 293 + ], + "type": "text", + "content": "0.85/3.2" + } + ] + } + ], + "index": 126, + "angle": 0, + "type": "image_caption" + } + ], + "index": 125 + }, + { + "type": "image", + "bbox": [ + 503, + 232, + 551, + 280 + ], + "blocks": [ + { + "bbox": [ + 503, + 232, + 551, + 280 + ], + "lines": [ + { + "bbox": [ + 503, + 232, + 551, + 280 + ], + "spans": [ + { + "bbox": [ + 503, + 232, + 551, + 280 + ], + "type": "image", + "image_path": "b43ee7265dc02cca8f65dd9204d7e5b6e07e295b14d14fedd59c59b1ad683b99.jpg" + } + ] + } + ], + "index": 127, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 282, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 507, + 282, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 507, + 282, + 547, + 293 + ], + "type": "text", + "content": "0.88/2.6" + } + ] + } + ], + "index": 128, + "angle": 0, + "type": "image_caption" + } + ], + "index": 127 + }, + { + "type": "image", + "bbox": [ + 307, + 307, + 355, + 354 + ], + "blocks": [ + { + "bbox": [ + 321, + 295, + 536, + 306 + ], + "lines": [ + { + "bbox": [ + 321, + 295, + 536, + 306 + ], + "spans": [ + { + "bbox": [ + 321, + 295, + 536, + 306 + ], + "type": "text", + "content": "all classified as freight car by four ImageNet models" + } + ] + } + ], + "index": 129, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 307, + 355, + 354 + ], + "lines": [ + { + "bbox": [ + 307, + 307, + 355, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 307, + 355, + 354 + ], + "type": "image", + "image_path": "4e7140d05f47c85caef4c21a641ee3cdd70ed87a0f5ee698c50314669f9a5ed5.jpg" + } + ] + } + ], + "index": 130, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 357, + 351, + 367 + ], + "lines": [ + { + "bbox": [ + 310, + 357, + 351, + 367 + ], + "spans": [ + { + "bbox": [ + 310, + 357, + 351, + 367 + ], + "type": "text", + "content": "0.85/2.5" + } + ] + } + ], + "index": 131, + "angle": 0, + "type": "image_caption" + } + ], + "index": 130 + }, + { + "type": "image", + "bbox": [ + 356, + 307, + 404, + 354 + ], + "blocks": [ + { + "bbox": [ + 356, + 307, + 404, + 354 + ], + "lines": [ + { + "bbox": [ + 356, + 307, + 404, + 354 + ], + "spans": [ + { + "bbox": [ + 356, + 307, + 404, + 354 + ], + "type": "image", + "image_path": "5c765abbfa0845be76cfba9bc4314b4a671a86dd65c0eee290423e949c101756.jpg" + } + ] + } + ], + "index": 132, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 357, + 400, + 367 + ], + "lines": [ + { + "bbox": [ + 359, + 357, + 400, + 367 + ], + "spans": [ + { + "bbox": [ + 359, + 357, + 400, + 367 + ], + "type": "text", + "content": "0.86/2.5" + } + ] + } + ], + "index": 133, + "angle": 0, + "type": "image_caption" + } + ], + "index": 132 + }, + { + "type": "image", + "bbox": [ + 405, + 307, + 452, + 354 + ], + "blocks": [ + { + "bbox": [ + 405, + 307, + 452, + 354 + ], + "lines": [ + { + "bbox": [ + 405, + 307, + 452, + 354 + ], + "spans": [ + { + "bbox": [ + 405, + 307, + 452, + 354 + ], + "type": "image", + "image_path": "271004bd3fd2d0688020e1fae53de0d933064f13017c960bb7ad9840f3b8a3f2.jpg" + } + ] + } + ], + "index": 134, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 357, + 449, + 367 + ], + "lines": [ + { + "bbox": [ + 409, + 357, + 449, + 367 + ], + "spans": [ + { + "bbox": [ + 409, + 357, + 449, + 367 + ], + "type": "text", + "content": "0.90/2.3" + } + ] + } + ], + "index": 135, + "angle": 0, + "type": "image_caption" + } + ], + "index": 134 + }, + { + "type": "image", + "bbox": [ + 453, + 307, + 502, + 354 + ], + "blocks": [ + { + "bbox": [ + 453, + 307, + 502, + 354 + ], + "lines": [ + { + "bbox": [ + 453, + 307, + 502, + 354 + ], + "spans": [ + { + "bbox": [ + 453, + 307, + 502, + 354 + ], + "type": "image", + "image_path": "63240b2ea26dd0e19553f2428c0b456d15ab83512661cc740427511cc9713ebb.jpg" + } + ] + } + ], + "index": 136, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 357, + 498, + 367 + ], + "lines": [ + { + "bbox": [ + 458, + 357, + 498, + 367 + ], + "spans": [ + { + "bbox": [ + 458, + 357, + 498, + 367 + ], + "type": "text", + "content": "0.82/2.2" + } + ] + } + ], + "index": 137, + "angle": 0, + "type": "image_caption" + } + ], + "index": 136 + }, + { + "type": "image", + "bbox": [ + 503, + 307, + 551, + 354 + ], + "blocks": [ + { + "bbox": [ + 503, + 307, + 551, + 354 + ], + "lines": [ + { + "bbox": [ + 503, + 307, + 551, + 354 + ], + "spans": [ + { + "bbox": [ + 503, + 307, + 551, + 354 + ], + "type": "image", + "image_path": "682e474921ec27b310ef72061a462208f2a81668ae224208a102796055f4bccd.jpg" + } + ] + } + ], + "index": 138, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 357, + 547, + 367 + ], + "lines": [ + { + "bbox": [ + 507, + 357, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 507, + 357, + 547, + 367 + ], + "type": "text", + "content": "0.87/2.2" + } + ] + } + ], + "index": 139, + "angle": 0, + "type": "image_caption" + } + ], + "index": 138 + }, + { + "type": "image", + "bbox": [ + 307, + 381, + 355, + 430 + ], + "blocks": [ + { + "bbox": [ + 320, + 370, + 537, + 381 + ], + "lines": [ + { + "bbox": [ + 320, + 370, + 537, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 370, + 537, + 381 + ], + "type": "text", + "content": "Images with spurious eucalyptus/plants but no koala" + } + ] + } + ], + "index": 140, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 381, + 355, + 430 + ], + "lines": [ + { + "bbox": [ + 307, + 381, + 355, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 355, + 430 + ], + "type": "image", + "image_path": "7cc871558c2f74dee577490a040eb115083fdc3a1eb2923c0b95f3275e771343.jpg" + } + ] + } + ], + "index": 141, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 433, + 351, + 442 + ], + "lines": [ + { + "bbox": [ + 310, + 433, + 351, + 442 + ], + "spans": [ + { + "bbox": [ + 310, + 433, + 351, + 442 + ], + "type": "text", + "content": "0.49/3.5" + } + ] + } + ], + "index": 142, + "angle": 0, + "type": "image_caption" + } + ], + "index": 141 + }, + { + "type": "image", + "bbox": [ + 356, + 381, + 404, + 430 + ], + "blocks": [ + { + "bbox": [ + 356, + 381, + 404, + 430 + ], + "lines": [ + { + "bbox": [ + 356, + 381, + 404, + 430 + ], + "spans": [ + { + "bbox": [ + 356, + 381, + 404, + 430 + ], + "type": "image", + "image_path": "d9ed6648e6e7ec103f36ec6ef269baa9d8fb476d2c799d461b1622c9f6d0389a.jpg" + } + ] + } + ], + "index": 143, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 433, + 400, + 442 + ], + "lines": [ + { + "bbox": [ + 359, + 433, + 400, + 442 + ], + "spans": [ + { + "bbox": [ + 359, + 433, + 400, + 442 + ], + "type": "text", + "content": "0.61/3.1" + } + ] + } + ], + "index": 144, + "angle": 0, + "type": "image_caption" + } + ], + "index": 143 + }, + { + "type": "image", + "bbox": [ + 405, + 381, + 452, + 430 + ], + "blocks": [ + { + "bbox": [ + 405, + 381, + 452, + 430 + ], + "lines": [ + { + "bbox": [ + 405, + 381, + 452, + 430 + ], + "spans": [ + { + "bbox": [ + 405, + 381, + 452, + 430 + ], + "type": "image", + "image_path": "13fc59d130ba54cbeb33bb7879ca0346e9c93fd71b30d839afdd5f713dfd35d6.jpg" + } + ] + } + ], + "index": 145, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 433, + 449, + 442 + ], + "lines": [ + { + "bbox": [ + 409, + 433, + 449, + 442 + ], + "spans": [ + { + "bbox": [ + 409, + 433, + 449, + 442 + ], + "type": "text", + "content": "0.36/3.1" + } + ] + } + ], + "index": 146, + "angle": 0, + "type": "image_caption" + } + ], + "index": 145 + }, + { + "type": "image", + "bbox": [ + 453, + 381, + 502, + 430 + ], + "blocks": [ + { + "bbox": [ + 453, + 381, + 502, + 430 + ], + "lines": [ + { + "bbox": [ + 453, + 381, + 502, + 430 + ], + "spans": [ + { + "bbox": [ + 453, + 381, + 502, + 430 + ], + "type": "image", + "image_path": "2a962ccc0109ae8bdf64c42eeb97112c411021f75caf04cf924183cc02ca1e9f.jpg" + } + ] + } + ], + "index": 147, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 433, + 497, + 442 + ], + "lines": [ + { + "bbox": [ + 459, + 433, + 497, + 442 + ], + "spans": [ + { + "bbox": [ + 459, + 433, + 497, + 442 + ], + "type": "text", + "content": "0.36/3.1" + } + ] + } + ], + "index": 148, + "angle": 0, + "type": "image_caption" + } + ], + "index": 147 + }, + { + "type": "image", + "bbox": [ + 503, + 381, + 551, + 430 + ], + "blocks": [ + { + "bbox": [ + 503, + 381, + 551, + 430 + ], + "lines": [ + { + "bbox": [ + 503, + 381, + 551, + 430 + ], + "spans": [ + { + "bbox": [ + 503, + 381, + 551, + 430 + ], + "type": "image", + "image_path": "fb2d9fedf1c00962f3a8ec7fab10f710347d785b7826f1e9f1dab34aa3102473.jpg" + } + ] + } + ], + "index": 149, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 433, + 547, + 442 + ], + "lines": [ + { + "bbox": [ + 507, + 433, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 507, + 433, + 547, + 442 + ], + "type": "text", + "content": "0.69/2.7" + } + ] + } + ], + "index": 150, + "angle": 0, + "type": "image_caption" + } + ], + "index": 149 + }, + { + "type": "image", + "bbox": [ + 307, + 456, + 355, + 504 + ], + "blocks": [ + { + "bbox": [ + 332, + 445, + 525, + 455 + ], + "lines": [ + { + "bbox": [ + 332, + 445, + 525, + 455 + ], + "spans": [ + { + "bbox": [ + 332, + 445, + 525, + 455 + ], + "type": "text", + "content": "all classified as koala by four ImageNet models" + } + ] + } + ], + "index": 151, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 456, + 355, + 504 + ], + "lines": [ + { + "bbox": [ + 307, + 456, + 355, + 504 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 355, + 504 + ], + "type": "image", + "image_path": "ec0ce2d1a664953e5608d6b6070ad07de40d8ae64d8af6c90f0e37df146a5a15.jpg" + } + ] + } + ], + "index": 152, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 506, + 351, + 517 + ], + "lines": [ + { + "bbox": [ + 310, + 506, + 351, + 517 + ], + "spans": [ + { + "bbox": [ + 310, + 506, + 351, + 517 + ], + "type": "text", + "content": "0.56/2.6" + } + ] + } + ], + "index": 153, + "angle": 0, + "type": "image_caption" + } + ], + "index": 152 + }, + { + "type": "image", + "bbox": [ + 356, + 456, + 404, + 504 + ], + "blocks": [ + { + "bbox": [ + 356, + 456, + 404, + 504 + ], + "lines": [ + { + "bbox": [ + 356, + 456, + 404, + 504 + ], + "spans": [ + { + "bbox": [ + 356, + 456, + 404, + 504 + ], + "type": "image", + "image_path": "fef7e6a9fe2d10335af4e7728a604c8d30f401e872a0b1d6129d41fe1088b2a7.jpg" + } + ] + } + ], + "index": 154, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 506, + 400, + 517 + ], + "lines": [ + { + "bbox": [ + 359, + 506, + 400, + 517 + ], + "spans": [ + { + "bbox": [ + 359, + 506, + 400, + 517 + ], + "type": "text", + "content": "0.62/2.5" + } + ] + } + ], + "index": 155, + "angle": 0, + "type": "image_caption" + } + ], + "index": 154 + }, + { + "type": "image", + "bbox": [ + 405, + 456, + 452, + 504 + ], + "blocks": [ + { + "bbox": [ + 405, + 456, + 452, + 504 + ], + "lines": [ + { + "bbox": [ + 405, + 456, + 452, + 504 + ], + "spans": [ + { + "bbox": [ + 405, + 456, + 452, + 504 + ], + "type": "image", + "image_path": "c6a74f224f78c5b62042650c42fe903570e5a1b792771e345f869fa79c355cf9.jpg" + } + ] + } + ], + "index": 156, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 506, + 449, + 517 + ], + "lines": [ + { + "bbox": [ + 409, + 506, + 449, + 517 + ], + "spans": [ + { + "bbox": [ + 409, + 506, + 449, + 517 + ], + "type": "text", + "content": "0.74/2.5" + } + ] + } + ], + "index": 157, + "angle": 0, + "type": "image_caption" + } + ], + "index": 156 + }, + { + "type": "image", + "bbox": [ + 453, + 456, + 502, + 504 + ], + "blocks": [ + { + "bbox": [ + 453, + 456, + 502, + 504 + ], + "lines": [ + { + "bbox": [ + 453, + 456, + 502, + 504 + ], + "spans": [ + { + "bbox": [ + 453, + 456, + 502, + 504 + ], + "type": "image", + "image_path": "e1ac2a48b5b267a5f9f1093a7a3007ddfba246d63efbbb94e4697652f0f54be7.jpg" + } + ] + } + ], + "index": 158, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 506, + 498, + 517 + ], + "lines": [ + { + "bbox": [ + 459, + 506, + 498, + 517 + ], + "spans": [ + { + "bbox": [ + 459, + 506, + 498, + 517 + ], + "type": "text", + "content": "0.72/2.4" + } + ] + } + ], + "index": 159, + "angle": 0, + "type": "image_caption" + } + ], + "index": 158 + }, + { + "type": "image", + "bbox": [ + 503, + 456, + 551, + 504 + ], + "blocks": [ + { + "bbox": [ + 503, + 456, + 551, + 504 + ], + "lines": [ + { + "bbox": [ + 503, + 456, + 551, + 504 + ], + "spans": [ + { + "bbox": [ + 503, + 456, + 551, + 504 + ], + "type": "image", + "image_path": "a0daf6d3b2d23f26bb5ad7a69324b618fb81cc9c899f2cf52ae2139f715669f8.jpg" + } + ] + } + ], + "index": 160, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 506, + 547, + 517 + ], + "lines": [ + { + "bbox": [ + 507, + 506, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 507, + 506, + 547, + 517 + ], + "type": "text", + "content": "0.66/2.2" + } + ] + } + ], + "index": 161, + "angle": 0, + "type": "image_caption" + } + ], + "index": 160 + }, + { + "type": "image", + "bbox": [ + 307, + 531, + 355, + 579 + ], + "blocks": [ + { + "bbox": [ + 334, + 520, + 524, + 531 + ], + "lines": [ + { + "bbox": [ + 334, + 520, + 524, + 531 + ], + "spans": [ + { + "bbox": [ + 334, + 520, + 524, + 531 + ], + "type": "text", + "content": "Images with spurious water jet but no fireboat" + } + ] + } + ], + "index": 162, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 531, + 355, + 579 + ], + "lines": [ + { + "bbox": [ + 307, + 531, + 355, + 579 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 355, + 579 + ], + "type": "image", + "image_path": "539f197c9df300a1aa9e765daa8d1af929468c7716c3c3b51dea548ae04dddc1.jpg" + } + ] + } + ], + "index": 163, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 582, + 351, + 591 + ], + "lines": [ + { + "bbox": [ + 310, + 582, + 351, + 591 + ], + "spans": [ + { + "bbox": [ + 310, + 582, + 351, + 591 + ], + "type": "text", + "content": "0.63/2.3" + } + ] + } + ], + "index": 164, + "angle": 0, + "type": "image_caption" + } + ], + "index": 163 + }, + { + "type": "image", + "bbox": [ + 356, + 531, + 404, + 579 + ], + "blocks": [ + { + "bbox": [ + 356, + 531, + 404, + 579 + ], + "lines": [ + { + "bbox": [ + 356, + 531, + 404, + 579 + ], + "spans": [ + { + "bbox": [ + 356, + 531, + 404, + 579 + ], + "type": "image", + "image_path": "28b216cd46e16f027887cf4dffcf9a136cb940f482ebefb3db4519920803a214.jpg" + } + ] + } + ], + "index": 165, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 582, + 400, + 591 + ], + "lines": [ + { + "bbox": [ + 359, + 582, + 400, + 591 + ], + "spans": [ + { + "bbox": [ + 359, + 582, + 400, + 591 + ], + "type": "text", + "content": "0.71/2.2" + } + ] + } + ], + "index": 166, + "angle": 0, + "type": "image_caption" + } + ], + "index": 165 + }, + { + "type": "image", + "bbox": [ + 405, + 531, + 452, + 579 + ], + "blocks": [ + { + "bbox": [ + 405, + 531, + 452, + 579 + ], + "lines": [ + { + "bbox": [ + 405, + 531, + 452, + 579 + ], + "spans": [ + { + "bbox": [ + 405, + 531, + 452, + 579 + ], + "type": "image", + "image_path": "66c383237bb2b0f4faa016135376e8c11ad7ff76c853f68d1c53d07adf2b222e.jpg" + } + ] + } + ], + "index": 167, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 582, + 449, + 591 + ], + "lines": [ + { + "bbox": [ + 409, + 582, + 449, + 591 + ], + "spans": [ + { + "bbox": [ + 409, + 582, + 449, + 591 + ], + "type": "text", + "content": "0.84/2.2" + } + ] + } + ], + "index": 168, + "angle": 0, + "type": "image_caption" + } + ], + "index": 167 + }, + { + "type": "image", + "bbox": [ + 453, + 531, + 502, + 579 + ], + "blocks": [ + { + "bbox": [ + 453, + 531, + 502, + 579 + ], + "lines": [ + { + "bbox": [ + 453, + 531, + 502, + 579 + ], + "spans": [ + { + "bbox": [ + 453, + 531, + 502, + 579 + ], + "type": "image", + "image_path": "2e74305ae404761c3c77dfd49a05a9731b27bfd41aeac5fd6fd6a2983548f4b6.jpg" + } + ] + } + ], + "index": 169, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 582, + 497, + 591 + ], + "lines": [ + { + "bbox": [ + 459, + 582, + 497, + 591 + ], + "spans": [ + { + "bbox": [ + 459, + 582, + 497, + 591 + ], + "type": "text", + "content": "0.65/2.1" + } + ] + } + ], + "index": 170, + "angle": 0, + "type": "image_caption" + } + ], + "index": 169 + }, + { + "type": "image", + "bbox": [ + 503, + 531, + 551, + 579 + ], + "blocks": [ + { + "bbox": [ + 503, + 531, + 551, + 579 + ], + "lines": [ + { + "bbox": [ + 503, + 531, + 551, + 579 + ], + "spans": [ + { + "bbox": [ + 503, + 531, + 551, + 579 + ], + "type": "image", + "image_path": "1d50568e613b91952708f5aba8784e3606eeb8792a7aa51f94dd57b17ffaf90a.jpg" + } + ] + } + ], + "index": 171, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 582, + 547, + 591 + ], + "lines": [ + { + "bbox": [ + 507, + 582, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 507, + 582, + 547, + 591 + ], + "type": "text", + "content": "0.63/1.9" + } + ] + } + ], + "index": 172, + "angle": 0, + "type": "image_caption" + } + ], + "index": 171 + }, + { + "type": "image", + "bbox": [ + 307, + 605, + 355, + 653 + ], + "blocks": [ + { + "bbox": [ + 328, + 594, + 530, + 605 + ], + "lines": [ + { + "bbox": [ + 328, + 594, + 530, + 605 + ], + "spans": [ + { + "bbox": [ + 328, + 594, + 530, + 605 + ], + "type": "text", + "content": "all classified as fireboat by four ImageNet models" + } + ] + } + ], + "index": 173, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 605, + 355, + 653 + ], + "lines": [ + { + "bbox": [ + 307, + 605, + 355, + 653 + ], + "spans": [ + { + "bbox": [ + 307, + 605, + 355, + 653 + ], + "type": "image", + "image_path": "b0e378ac19e854bde16c8c16b81016e8bd6d2062686b3156b52ea4c629b31439.jpg" + } + ] + } + ], + "index": 174, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 656, + 351, + 666 + ], + "lines": [ + { + "bbox": [ + 310, + 656, + 351, + 666 + ], + "spans": [ + { + "bbox": [ + 310, + 656, + 351, + 666 + ], + "type": "text", + "content": "0.53/1.9" + } + ] + } + ], + "index": 175, + "angle": 0, + "type": "image_caption" + } + ], + "index": 174 + }, + { + "type": "image", + "bbox": [ + 356, + 605, + 404, + 653 + ], + "blocks": [ + { + "bbox": [ + 356, + 605, + 404, + 653 + ], + "lines": [ + { + "bbox": [ + 356, + 605, + 404, + 653 + ], + "spans": [ + { + "bbox": [ + 356, + 605, + 404, + 653 + ], + "type": "image", + "image_path": "494d868e345a740ca98e1534f8c2d1a4f19efe7c7b1ec1748cfda15a9b3ad39e.jpg" + } + ] + } + ], + "index": 176, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 656, + 400, + 666 + ], + "lines": [ + { + "bbox": [ + 359, + 656, + 400, + 666 + ], + "spans": [ + { + "bbox": [ + 359, + 656, + 400, + 666 + ], + "type": "text", + "content": "0.79/1.9" + } + ] + } + ], + "index": 177, + "angle": 0, + "type": "image_caption" + } + ], + "index": 176 + }, + { + "type": "image", + "bbox": [ + 405, + 605, + 452, + 653 + ], + "blocks": [ + { + "bbox": [ + 405, + 605, + 452, + 653 + ], + "lines": [ + { + "bbox": [ + 405, + 605, + 452, + 653 + ], + "spans": [ + { + "bbox": [ + 405, + 605, + 452, + 653 + ], + "type": "image", + "image_path": "71073f348182fd929fb4a5ab6345d59d1d824aef912ad5ff5b6628923f879d12.jpg" + } + ] + } + ], + "index": 178, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 656, + 449, + 666 + ], + "lines": [ + { + "bbox": [ + 409, + 656, + 449, + 666 + ], + "spans": [ + { + "bbox": [ + 409, + 656, + 449, + 666 + ], + "type": "text", + "content": "0.78/1.8" + } + ] + } + ], + "index": 179, + "angle": 0, + "type": "image_caption" + } + ], + "index": 178 + }, + { + "type": "image", + "bbox": [ + 453, + 605, + 502, + 653 + ], + "blocks": [ + { + "bbox": [ + 453, + 605, + 502, + 653 + ], + "lines": [ + { + "bbox": [ + 453, + 605, + 502, + 653 + ], + "spans": [ + { + "bbox": [ + 453, + 605, + 502, + 653 + ], + "type": "image", + "image_path": "3fa638d12c61fb5772f77e9553ee27edc4abbc4a025a4dc50b658973ecc02481.jpg" + } + ] + } + ], + "index": 180, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 656, + 498, + 666 + ], + "lines": [ + { + "bbox": [ + 459, + 656, + 498, + 666 + ], + "spans": [ + { + "bbox": [ + 459, + 656, + 498, + 666 + ], + "type": "text", + "content": "0.79/1.8" + } + ] + } + ], + "index": 181, + "angle": 0, + "type": "image_caption" + } + ], + "index": 180 + }, + { + "type": "image", + "bbox": [ + 503, + 605, + 551, + 653 + ], + "blocks": [ + { + "bbox": [ + 503, + 605, + 551, + 653 + ], + "lines": [ + { + "bbox": [ + 503, + 605, + 551, + 653 + ], + "spans": [ + { + "bbox": [ + 503, + 605, + 551, + 653 + ], + "type": "image", + "image_path": "b9ca131bbf96e9cf4b8c4e9a39bcbd814f4f36181e146dbf343b815b8d441331.jpg" + } + ] + } + ], + "index": 182, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 656, + 547, + 666 + ], + "lines": [ + { + "bbox": [ + 507, + 656, + 547, + 666 + ], + "spans": [ + { + "bbox": [ + 507, + 656, + 547, + 666 + ], + "type": "text", + "content": "0.76/1.8" + } + ] + } + ], + "index": 183, + "angle": 0, + "type": "image_caption" + } + ], + "index": 182 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20239" + } + ] + } + ], + "index": 185 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 182 + ], + "type": "text", + "content": "training. However, this would require relabeling all spurious ImageNet classes which is not feasible. We could order the training set according to the value " + }, + { + "bbox": [ + 46, + 72, + 289, + 182 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{(k)}" + }, + { + "bbox": [ + 46, + 72, + 289, + 182 + ], + "type": "text", + "content": " of the corresponding neural PCA component which indicates how much of the spurious feature an image contains. While this would speed up the process significantly, it would still require a significant amount of manual relabeling. Can one do it also without any additional labeling? Yes, as described in Sec. 4 we can rewrite the logit of the " + }, + { + "bbox": [ + 46, + 72, + 289, + 182 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 182 + ], + "type": "text", + "content": "-th class as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 188, + 287, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 188, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 83, + 188, + 287, + 209 + ], + "type": "interline_equation", + "content": "f _ {k} (x) = \\sum_ {l = 1} ^ {D} \\alpha_ {l} ^ {(k)} (x) + \\left\\langle \\mathbf {1}, \\bar {\\psi} _ {k} \\right\\rangle + b _ {k}. \\tag {7}", + "image_path": "45f23e75fb7f5611d5b4cc322d18125e659f280304f6362f4f49d8db7743a565.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "type": "text", + "content": "For a spurious component " + }, + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "type": "inline_equation", + "content": "\\min \\{\\alpha_l^{(k)}, 0\\}" + }, + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "type": "inline_equation", + "content": "\\alpha_l^{(k)}" + }, + { + "bbox": [ + 47, + 217, + 287, + 269 + ], + "type": "text", + "content": " to remove its positive contribution from the logit (negative contributions are semantically different). After removal of the spurious features, the new logit becomes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 275, + 287, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 275, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 73, + 275, + 287, + 301 + ], + "type": "interline_equation", + "content": "f _ {k} ^ {S p u F i x} (x) = f _ {k} (x) - \\sum_ {l \\in \\mathcal {S} _ {k}} \\max \\left\\{\\alpha_ {l} ^ {(k)} (x), 0 \\right\\} \\tag {8}", + "image_path": "397b926bc1d1c9d105df03de3e273ba83ede8fc56411e6e17637453706179979.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "text", + "content": " is the set of spurious NPCA components of class " + }, + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 307, + 287, + 344 + ], + "type": "text", + "content": ". We denote this method as SpuFix. It significantly reduces dependence on spurious features, see Sec. 7.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": "Transfer of SpuFix to any ImageNet classifier: As described in Sec. 3, harmful spurious features are a result of subpopulations in the training data. While not every spurious correlation will be picked up by every model, most of our detected spurious features generalize to a wide range of classifiers (see Sec. 7). In the following, we show how SpuFix can be transferred to any given ImageNet classifier " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": " for which " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "\\tilde{f}_k" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": " denotes the logit and " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "\\tilde{\\psi}_k" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": " the weighted penultimate layer of class " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": ". The goal is to find a direction " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": " in the weighted feature space " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "\\tilde{\\psi}_k" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": " for every spurious NPCA component " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": " corresponding to the eigenvector " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "v_l" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": " of the original model, resp. " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "\\alpha_l^{(k)}(x)" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": ", and then truncate its positive component. To find this direction we maximize the covariance of the projection onto " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "\\alpha_l^{(k)}" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": " over the training images of class " + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 357, + 289, + 540 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 546, + 287, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 546, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 61, + 546, + 287, + 574 + ], + "type": "interline_equation", + "content": "b _ {l} ^ {(k)} = \\underset {\\| b \\| _ {2} = 1} {\\operatorname {a r g m a x}} \\sum_ {s \\in I _ {k}} \\left\\langle b, \\tilde {\\psi} _ {k} \\left(x _ {s}\\right) - \\bar {\\tilde {\\psi}} _ {k} \\right\\rangle \\alpha_ {l} ^ {(k)} \\left(x _ {s}\\right). \\tag {9}", + "image_path": "4b3de780570b54f973f637f44b932bf9e4056ae1d1f1947cec13305f01f5e87c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 580, + 182, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 182, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 182, + 591 + ], + "type": "text", + "content": "which has a closed form solution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 596, + 287, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 596, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 66, + 596, + 287, + 639 + ], + "type": "interline_equation", + "content": "b _ {l} ^ {(k)} = \\frac {\\sum_ {s \\in I _ {k}} \\left(\\tilde {\\psi} _ {k} \\left(x _ {s}\\right) - \\bar {\\tilde {\\psi}} _ {k}\\right) \\alpha_ {l} ^ {(k)} \\left(x _ {s}\\right)}{\\left\\| \\sum_ {s \\in I _ {k}} \\left(\\tilde {\\psi} _ {k} \\left(x _ {s}\\right) - \\bar {\\tilde {\\psi}} _ {k}\\right) \\alpha_ {l} ^ {(k)} \\left(x _ {s}\\right) \\right\\| _ {2}}. \\tag {10}", + "image_path": "4a11817fd24d590dd3519354abf604f6dcb74f51d2c11b86f9dc654a5cdef67a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "text", + "content": "In contrast to the eigenvectors " + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "text", + "content": ", the matched vectors " + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "inline_equation", + "content": "b_{l}^{(k)}" + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "text", + "content": " are not necessarily orthogonal. Thus before truncation the centered features " + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\psi}_k(x) - \\bar{\\tilde{\\psi}}_k" + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "text", + "content": " need to be projected onto the subspace spanned by the " + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "inline_equation", + "content": "b_{l}^{(k)}" + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "text", + "content": " and represented in the non-orthogonal basis " + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{b_{l}^{(k)}\\}_{l\\in S_{k}}" + }, + { + "bbox": [ + 46, + 644, + 287, + 715 + ], + "type": "text", + "content": ". We denote this representation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "by " + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "P^{(k)}(x)" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": " (details in C.1). The logit of class " + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": " of the SpuFix version of " + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": " is then:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 104, + 545, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 104, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 310, + 104, + 545, + 140 + ], + "type": "interline_equation", + "content": "\\tilde {f} _ {k} ^ {S p u F i x} (x) = \\tilde {f} _ {k} (x) - \\sum_ {l \\in \\mathcal {S} _ {k}} \\max \\left\\{\\langle \\mathbf {1}, b _ {l} ^ {(k)} \\rangle P _ {l} ^ {(k)} (x), 0 \\right\\}. \\tag {11}", + "image_path": "196c48d62d396ac9c0c246df343040e49dcc5df51aa310e584eafbb012c71f11.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 141, + 547, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 141, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 304, + 141, + 547, + 201 + ], + "type": "text", + "content": "In the case where " + }, + { + "bbox": [ + 304, + 141, + 547, + 201 + ], + "type": "inline_equation", + "content": "\\tilde{f} = f" + }, + { + "bbox": [ + 304, + 141, + 547, + 201 + ], + "type": "text", + "content": " (the robust model), we recover the original SpuFix truncation in (8) (see C.2). It turns out that SpuFix is even effective when the architecture is quite different from the ResNet50 we used for detection, e.g. ViT or VOLO, see Sec. 7.3 and Table 1 and 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 213, + 514, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 514, + 227 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 514, + 227 + ], + "type": "text", + "content": "6. Comparison to neural features of [61]" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "text", + "content": "We compare our NPCA framework to the method of [61] to detect spurious features for ImageNet. As model, they use a " + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "text", + "content": "-robust ResNet50. Let " + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "inline_equation", + "content": "J_{k}" + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "text", + "content": " be the set of training images classified as class " + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "text", + "content": ". [61] define the " + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "text", + "content": "-th component " + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "inline_equation", + "content": "m_j^{(k)}" + }, + { + "bbox": [ + 304, + 232, + 545, + 297 + ], + "type": "text", + "content": " of the class-wise mean over predictions," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 305, + 545, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 305, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 361, + 305, + 545, + 331 + ], + "type": "interline_equation", + "content": "m ^ {(k)} = \\frac {1}{\\left| J _ {k} \\right|} \\sum_ {x _ {s} \\in J _ {k}} \\psi_ {k} \\left(x _ {s}\\right), \\tag {12}", + "image_path": "f89dd068635c6d8f9140ed80af123cf6c8ba632d602e618c71fd60e817cf9f61.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "text", + "content": "as the importance of the " + }, + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "text", + "content": "-th neuron for class " + }, + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "text", + "content": ".1 Then, they order the neurons of the penultimate layer according to the score " + }, + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "inline_equation", + "content": "m_j^{(k)}" + }, + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "text", + "content": " and consider the top-5 neurons of each class. The main difference to our approach is that they assume single neurons with maximal influence on the mean are capturing spurious features whereas our NPCA components are linear combinations of neurons that capture the variance around the mean. Given that the ResNet50 has only 2048 neurons for 1000 classes, some neurons are labeled as core and spurious feature for multiple classes simultaneously, even though the images are quite different. A major advantage of NPCA is that due to the orthogonality of the PCA components, we identify diverse subpopulations in the training data. As [61] use no constraints for the neurons, they often find very similar subpopulations. Hence, one may miss spurious subpopulations when only checking the top-5 components, see Fig. 5. Another difference is that they maximize the score " + }, + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "inline_equation", + "content": "m_j^{(k)}" + }, + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "text", + "content": " for the training images " + }, + { + "bbox": [ + 304, + 340, + 546, + 561 + ], + "type": "inline_equation", + "content": "x_r" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 338, + 569, + 514, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 569, + 514, + 588 + ], + "spans": [ + { + "bbox": [ + 338, + 569, + 514, + 588 + ], + "type": "interline_equation", + "content": "w _ {j} ^ {(k)} = x _ {r} + \\mathrm {a r g m a x} _ {\\| \\delta \\| _ {2} \\leq \\epsilon} m _ {j} ^ {(k)} (x _ {r} + \\delta),", + "image_path": "c0cf1cbfa46ae31ead0eb1034ae65340fab2818b667449cfc13ab40c4df6224a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 597, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 647 + ], + "type": "text", + "content": "whereas we maximize our NPCA component " + }, + { + "bbox": [ + 304, + 597, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\alpha_l^{(k)}(x)" + }, + { + "bbox": [ + 304, + 597, + 545, + 647 + ], + "type": "text", + "content": " starting from a gray image to introduce no bias. Thus, we check if the classifier produces the spurious feature and not only enhances it on an image showing it already." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 647, + 545, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 695 + ], + "type": "text", + "content": "The third difference is that [61] want to identify any spurious feature while our goal is to find harmful ones. Thus, they use weaker criteria for deciding if a neuron shows a spurious feature: main criterion is if the neural activation" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 523, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 523, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 523, + 713 + ], + "type": "text", + "content": "1The top-5 neurons do not change when using " + }, + { + "bbox": [ + 317, + 702, + 523, + 713 + ], + "type": "inline_equation", + "content": "I_{k}" + }, + { + "bbox": [ + 317, + 702, + 523, + 713 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 317, + 702, + 523, + 713 + ], + "type": "inline_equation", + "content": "J_{k}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20240" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 83, + 293, + 286 + ], + "blocks": [ + { + "bbox": [ + 99, + 71, + 251, + 82 + ], + "lines": [ + { + "bbox": [ + 99, + 71, + 251, + 82 + ], + "spans": [ + { + "bbox": [ + 99, + 71, + 251, + 82 + ], + "type": "text", + "content": "Top 5 neural PCA components (ours)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 83, + 293, + 286 + ], + "lines": [ + { + "bbox": [ + 47, + 83, + 293, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 83, + 293, + 286 + ], + "type": "image", + "image_path": "07923a3e5e2377b02bbc92739db36494da0d9c91fd2f8284e531fe75ea1bd88d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 292, + 288, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 288, + 353 + ], + "type": "text", + "content": "Figure 5: Comparison of top-5 NPCA and top-5 neurons for class hummingbird for the robust model of [61]: Our NPCA components identify diverse subpopulations in the training set whereas the top neurons show similar ones and the spurious bird feeder is not detected, see also App. A." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 365, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 455 + ], + "type": "text", + "content": "map based on " + }, + { + "bbox": [ + 46, + 365, + 287, + 455 + ], + "type": "inline_equation", + "content": "m_{j}^{(k)}" + }, + { + "bbox": [ + 46, + 365, + 287, + 455 + ], + "type": "text", + "content": " is off the class object according to a majority vote of 5 human labelers. The visualizations " + }, + { + "bbox": [ + 46, + 365, + 287, + 455 + ], + "type": "inline_equation", + "content": "w_{j}^{(k)}" + }, + { + "bbox": [ + 46, + 365, + 287, + 455 + ], + "type": "text", + "content": " are only used if the activation maps are inconclusive. In contrast, we require i) our NFPV to show the spurious feature, ii) the GradCAM based on " + }, + { + "bbox": [ + 46, + 365, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{(k)}" + }, + { + "bbox": [ + 46, + 365, + 287, + 455 + ], + "type": "text", + "content": " highlights mainly the spurious feature, and iii) our human labelers have to agree that the NPCA component is a harmful spurious feature." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 455, + 288, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 288, + 659 + ], + "type": "text", + "content": "As our criteria are more strict (in particular, that the NFPV shows the spurious feature is a strong criterion), it is not surprising that [61] find more spurious features (630 in 357 classes) than we do (322 in 230 classes). Moreover, the employed models are different and we examine top-10 NPCA components whereas they check top-5 neurons. Thus, the comparison is difficult and our novel dataset \"SpuriousImageNet\" could be biased towards spurious features which only we found. Hence, we compute our top-5 NPCA components for their robust ResNet50 for a direct comparison to their top-5 neurons and their found spurious features. In Fig. 5 we compare them for the class hummingbird where they do not find the spurious feature \"bird feeder\". In general, we observe that our found subpopulations are more diverse and thus we find more spurious features than they do when using their weaker criteria. In App. E we do an extensive comparison for all classes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "type": "text", + "content": "7. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "In this section, we provide a qualitative and quantitative evaluation of our 322 detected spurious features in Ima" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "geNet, see Sec. 4. For the quantitative evaluation, we create the dataset \"Spurious ImageNet\", which allows checking the reliance of a given ImageNet classifier on spurious features. We also evaluate our mitigation strategy \"SpuFix\" which does not require additional labels or retraining of the classifier and can be transferred to other image classifiers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 154, + 432, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 154, + 432, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 154, + 432, + 167 + ], + "type": "text", + "content": "7.1. Qualitative evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 173, + 546, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 546, + 425 + ], + "type": "text", + "content": "For the qualitative evaluation, we visualize some of our found 322 spurious features, see Fig. 4. For each class, we show five random training images, the NPFV, and the four most activating training images of the neural PCA component labeled to be spurious. Additionally, we always show ten images which only show the spurious feature but not the actual class e.g., only the bird feeder (spurious) but no hummingbird (class). All ten images are classified as the corresponding class for the robust classifier we have used to compute the NPCA components and three non-robust ImageNet classifiers (ResNext101, Eff.Net B5, ConvNext-B, see also Tab. 2). This shows that our spurious features generalize from the robust classifier to SOTA ImageNet classifiers, indicating that the found spurious features are mainly due to the design of the training set, rather than failures in model training. Our novel validation by collecting real images with the spurious feature but without the class object which are consistently classified as this class directly shows the impact of harmful spurious features and has the advantage that it does not introduce artifacts via masking nor requires expensive pixel-wise segmentations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 425, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 546, + 665 + ], + "type": "text", + "content": "Image Collection: The images showing only the spurious feature were obtained by sorting the 9 million images of OpenImages [38] by the value " + }, + { + "bbox": [ + 304, + 425, + 546, + 665 + ], + "type": "inline_equation", + "content": "\\alpha_l^{(k)}(x)" + }, + { + "bbox": [ + 304, + 425, + 546, + 665 + ], + "type": "text", + "content": " of the neural PCA component. We check the top 625 retrieved images classified by the robust classifier as the corresponding class if they are all classified as the same class by the additional non-robust classifiers and do not show the corresponding class. This is a quite strict criterion as spurious features can be shared across classes, e.g., twigs for birds, and thus agreement of classifiers is not granted and the images can show the spurious feature and the true class. Nevertheless, this procedure yields between 77 (\"hummingbird\") and 179 (\"freight car\") images of which we show a selection. For \"hummingbird\", a lot of these images show red flowers (without a hummingbird) which makes sense as the NPFV displays features of red flowers and due to the bias that OpenImages does not contain many images of hummingbird feeders. In these cases, we additionally retrieve Flickr images with appropriate text queries e.g., \"hummingbird feeder\" and filter them." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Spurious Class Extension: For \"hummingbird\", \"freight car\", and \"koala\" the spurious features significantly extend the predictions beyond the actual class (see Fig. 2). Bird feeders are classified as hummingbirds, graf" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20241" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 74, + 541, + 150 + ], + "blocks": [ + { + "bbox": [ + 53, + 74, + 541, + 150 + ], + "lines": [ + { + "bbox": [ + 53, + 74, + 541, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 74, + 541, + 150 + ], + "type": "image", + "image_path": "4563f6e3923208960d0af87f591719c55f3f480f7ecd87599ed63d9f3efe52e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 163, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 163, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 547, + 248 + ], + "type": "text", + "content": "Figure 6: Spurious Score: we plot the AUC of different models for 7 out of 100 classes in \"Spurious ImageNet\" (see Fig 13-15): hummingbird (bird feeder/red flowers), freight car (graffiti), koala (plants/trees), fireboat (water jet), flagpole (US flag), gondola (house/river), and bookshop (storefront). The spurious features for hummingbird, freight car and koala which were not detected in [61, 62] are also spurious for their robust ResNet50 [61]. Training on ImageNet 21k or Fine-tuning from 21k (1kFT21k) decreases dependence on harmful spurious features but classes like flagpole and bookshop remain strongly affected. Our cheap mitigation strategy SpuFix improves the AUCs significantly for both robust ResNet50 but also improves the ViT-B variants trained/fine-tuned on ImageNet1k, especially for the difficult classes flagpole and bookshop." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 291, + 126, + 351 + ], + "blocks": [ + { + "bbox": [ + 65, + 266, + 126, + 290 + ], + "lines": [ + { + "bbox": [ + 65, + 266, + 126, + 290 + ], + "spans": [ + { + "bbox": [ + 65, + 266, + 126, + 290 + ], + "type": "text", + "content": "hummingbird (red flowers)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 291, + 126, + 351 + ], + "lines": [ + { + "bbox": [ + 65, + 291, + 126, + 351 + ], + "spans": [ + { + "bbox": [ + 65, + 291, + 126, + 351 + ], + "type": "image", + "image_path": "b51854110c5f9577a2e18122ed2a08190de46fc889f1d6c0d18446e8e6427c13.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 354, + 112, + 363 + ], + "lines": [ + { + "bbox": [ + 78, + 354, + 112, + 363 + ], + "spans": [ + { + "bbox": [ + 78, + 354, + 112, + 363 + ], + "type": "text", + "content": "fireboat" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 65, + 377, + 126, + 437 + ], + "blocks": [ + { + "bbox": [ + 74, + 367, + 116, + 376 + ], + "lines": [ + { + "bbox": [ + 74, + 367, + 116, + 376 + ], + "spans": [ + { + "bbox": [ + 74, + 367, + 116, + 376 + ], + "type": "text", + "content": "(water jet)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 377, + 126, + 437 + ], + "lines": [ + { + "bbox": [ + 65, + 377, + 126, + 437 + ], + "spans": [ + { + "bbox": [ + 65, + 377, + 126, + 437 + ], + "type": "image", + "image_path": "4f6a05639cea9b15a673e3da497bd21175863f24e67b1836acae247343f22319.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 444, + 287, + 480 + ], + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 480 + ], + "type": "text", + "content": "Figure 7: Spurious ImageNet: sample images from the dataset for 6 out of 100 classes showing the spurious feature but not the class object, see also App. G." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 135, + 291, + 197, + 351 + ], + "blocks": [ + { + "bbox": [ + 145, + 267, + 188, + 290 + ], + "lines": [ + { + "bbox": [ + 145, + 267, + 188, + 290 + ], + "spans": [ + { + "bbox": [ + 145, + 267, + 188, + 290 + ], + "type": "text", + "content": "freight car (graffiti)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 291, + 197, + 351 + ], + "lines": [ + { + "bbox": [ + 135, + 291, + 197, + 351 + ], + "spans": [ + { + "bbox": [ + 135, + 291, + 197, + 351 + ], + "type": "image", + "image_path": "eed06027d1d78c3d80f163024a2ef6b152593b0359f6fda13b18a2ff780eb396.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 135, + 377, + 197, + 437 + ], + "blocks": [ + { + "bbox": [ + 149, + 354, + 184, + 376 + ], + "lines": [ + { + "bbox": [ + 149, + 354, + 184, + 376 + ], + "spans": [ + { + "bbox": [ + 149, + 354, + 184, + 376 + ], + "type": "text", + "content": "flagpole (US flag)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 377, + 197, + 437 + ], + "lines": [ + { + "bbox": [ + 135, + 377, + 197, + 437 + ], + "spans": [ + { + "bbox": [ + 135, + 377, + 197, + 437 + ], + "type": "image", + "image_path": "6c6536a6220f471cfa4105ae0c1c1345cf028c60727e243c8859ae4894691d0e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 207, + 291, + 267, + 350 + ], + "blocks": [ + { + "bbox": [ + 210, + 268, + 264, + 290 + ], + "lines": [ + { + "bbox": [ + 210, + 268, + 264, + 290 + ], + "spans": [ + { + "bbox": [ + 210, + 268, + 264, + 290 + ], + "type": "text", + "content": "koala (plants/trees)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 207, + 291, + 267, + 350 + ], + "lines": [ + { + "bbox": [ + 207, + 291, + 267, + 350 + ], + "spans": [ + { + "bbox": [ + 207, + 291, + 267, + 350 + ], + "type": "image", + "image_path": "7860febbf72c9ffcd5b66b6bba9296c151a3373c7dbbe608d3df22208bf00745.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 207, + 377, + 268, + 437 + ], + "blocks": [ + { + "bbox": [ + 211, + 354, + 263, + 376 + ], + "lines": [ + { + "bbox": [ + 211, + 354, + 263, + 376 + ], + "spans": [ + { + "bbox": [ + 211, + 354, + 263, + 376 + ], + "type": "text", + "content": "gondola (house/river)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 207, + 377, + 268, + 437 + ], + "lines": [ + { + "bbox": [ + 207, + 377, + 268, + 437 + ], + "spans": [ + { + "bbox": [ + 207, + 377, + 268, + 437 + ], + "type": "image", + "image_path": "21bf0e1b5da35e544868e770485cc819e022f017481c6cb0c5bb08570e1e56b5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 498, + 286, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 594 + ], + "type": "text", + "content": "fiti as freight cars, and (eucalyptus) plants as koalas. This class extension cannot be detected by monitoring test performance and thus is likely to be noticed only after deployment. For \"hummingbird\", we see in Fig. 4 two images with bees on the bird feeder where \"bee\" is an ImageNet class (also a grasshopper for \"koala\"). Nevertheless, the spurious \"bird feeder\" feature of \"hummingbird\" overrules \"bee\" even though no hummingbird is present." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 714 + ], + "type": "text", + "content": "Spurious Shared Feature: The spurious feature \"water jet\" is shared among the classes \"fireboat\" and \"fountain\". It appears more frequently for \"fireboat\" (see Fig. 2) which leads to an in-distribution shift where now a large number of images of the \"fountain\"-class with a water jet are wrongly classified as \"fireboat\". The spurious feature \"water jet\" for \"fireboat\" has been found also in the Salient ImageNet dataset [61, 62]. However, they did not find spurious features for freight car and koala (in App. A we do a comparison). More examples are in App. F." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 268, + 476, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 476, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 476, + 281 + ], + "type": "text", + "content": "7.2. The Spurious ImageNet dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 290, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 456 + ], + "type": "text", + "content": "A key contribution of this paper is our novel evaluation of spurious features for image classifiers without requiring pixel-wise annotations [48, 59] or having to rely on the validity of neural heatmaps [61]. Instead, we use images from OpenImages to show that images only containing the spurious feature but not the class object are classified as this class. This has the advantage that we consider real images and thus provide a realistic impression of the performance of ImageNet classifiers in the wild. Adding noise [61] or masking [48, 44] image regions requires pixel-wise accurate annotations which are labor-expensive, masking only the object still contains shape information, and using masks avoiding this, e.g., a bounding box around the object, can hide a significant portion of the image which is unrealistic." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 457, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 650 + ], + "type": "text", + "content": "To allow for a quantitative analysis of the influence of spurious features on ImageNet classifiers, we collected images similar to the ones shown to illustrate the spurious features in Fig. 4. The images are chosen such that they show the spurious feature but not the class object. The only difference is that we relax the classification condition and only require two of the four classifiers (robust ResNet50, ResNext101, EfficientNet-B5, ConvNext-B) to predict the corresponding class. We select 100 of our spurious features and for each collect 75 images from the top-ranked images in OpenImages according to the value of " + }, + { + "bbox": [ + 304, + 457, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{(k)}" + }, + { + "bbox": [ + 304, + 457, + 545, + 650 + ], + "type": "text", + "content": " for which two human labelers agree that they contain the spurious feature but not class " + }, + { + "bbox": [ + 304, + 457, + 545, + 650 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 457, + 545, + 650 + ], + "type": "text", + "content": " and two out of four classifiers predict class " + }, + { + "bbox": [ + 304, + 457, + 545, + 650 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 457, + 545, + 650 + ], + "type": "text", + "content": ". We call the dataset Spurious ImageNet as it allows to check the dependence on spurious features with real images for ImageNet classifiers, see Fig. 7 and App. G for samples." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Spurious Score: A classifier " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " not relying on the spurious feature should predict a low probability for class " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " for the Spurious ImageNet samples, especially compared to ImageNet test set images of class " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Thus, for each class, we measure the AUC (area under the curve) for the separation" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20242" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "of images with the spurious features but not showing class " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " versus test set images of class " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " according to the predicted probability for class " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ". A classifier not depending on the spurious feature should attain a perfect AUC of 1, whereas a value significantly below 1 shows strong reliance. We report the mean AUC (mAUC) over all 100 classes in Tab. 1. All ImageNet models trained only on ImageNet1k are heavily influenced by spurious features. Thus, spurious features are mainly a problem of the training set rather than the classifier, and spurious features found with an adversarially robust model transfer to other ImageNet classifiers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 479 + ], + "type": "text", + "content": "Pre-training on larger datasets: Some spurious features such as flag (flag pole), bird feeder (hummingbird), and eucalyptus (koala) are classes in ImageNet21k. Therefore, they should no longer be spurious for the other classes. Thus, we test if ImageNet1k-classifiers fine-tuned from an ImageNet21k model are less reliant on spurious features. The results in Tab. 1 and Fig. 6 suggest that the influence of spurious features is damped but they are far from being free of them. To check how much is lost due to fine-tuning we evaluate a ViT-B trained on ImageNet21k which has a mean AUC of 0.931 whereas the fine-tuned model has 0.917. This shows that fine-tuning does not hurt much. While finetuning from ImageNet21k improves the mean AUC, for several classes the dependence on spurious features is still significant, see also Fig. 16 how one has to be careful in the interpretation of higher AUC values. In addition to ImageNet21k, we also evaluate models trained on other large image datasets (JFT-300M[28], YFFC-100M, 1B Instagram[84], MIM[24], LAION-2B and LAION-400M[54]) using self-supervised learning or which are based on CLIP [49]. However, these models also do not achieve better spurious scores (Tab. 1 and Tab. 2). We evaluate a large number of SOTA models in App. D." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 490, + 269, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 269, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 269, + 504 + ], + "type": "text", + "content": "7.3. Evaluation of mitigation technique SpuFix" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Fixing spurious features is a non-trivial task and can require a substantial labeling effort. We evaluate our SpuFix from Sec. 5 that does not require retraining or additional labels. The positive effect of this fix of spurious features (SpuFix) can be seen in Tab. 1 and Fig. 6. Compared to the original robust ResNet50 with a spurious mAUC of 0.630, the SpuFix version has a significantly better spurious mAUC of 0.763. Test set accuracy reduces by " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " but this is a rather positive effect, as several of the additional errors arise since the robust ResNet50 uses spurious features for its decision, e.g., for classes like \"balance beam\" or \"puck\" the class object is often not visible in the cropped test set images. In Table 2 we provide a large scale evaluation of the transfer of SpuFix to SOTA ImageNet models. We observe a consistently better mAUC on Spurious ImageNet, even for very large models fine-tuned from 21k or trained on other large datasets, e.g. SpuFix improves the mAUC of VOLO-" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 306, + 69, + 550, + 340 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 550, + 340 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 550, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 550, + 340 + ], + "type": "table", + "html": "
ModelOriginalSpuFix
INet Acc. ↑SpurIN mAUC ↑INet Acc. ↑SpurIN mAUC ↑
ImageNet1k
Rob. ResNet5057.4%0.63056.8%0.763
Rob. ResNet50[61]57.9%0.65157.2%0.764
ConvNeXt-L[41]84.8%0.80384.8%0.819
ViT-B AugReg[66]81.1%0.85081.1%0.859
VOLO-D5 512[85]87.1%0.88287.1%0.907
ImageNet21kFT1k
EfficientNetv2-L[70]86.8%0.89386.8%0.898
ConvNeXt-L[41]87.0%0.91087.0%0.913
ViT-B AugReg[66]86.0%0.91785.9%0.925
BEiT-L\\16[8]88.6%0.92188.6%0.927
LAION-2B
CNeXt-L CLIP 384[49]87.8%0.87987.9%0.884
ViT-L\\14 CLIP[49]88.2%0.91288.2%0.914
MIM
EVA-G\\14 CLIP 560[24]89.8%0.91989.8%0.925
ImageNet21k
ConvNeXt-L[41]-0.943-0.943
ViT-B AugReg[66]-0.931-0.931
", + "image_path": "48ef1e4ec4310dcf32c609bedc1f34f6e9cb769f7bf8483cfb2aa4efab7adc2c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 348, + 547, + 421 + ], + "lines": [ + { + "bbox": [ + 304, + 348, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 547, + 421 + ], + "type": "text", + "content": "Table 1: Quantitative Evaluation on Spurious ImageNet: ImageNet classifiers of different training modalities depend on spurious features in varying strength. The mAUC is the mean of AUCs for the separation of images containing the spurious feature but not class " + }, + { + "bbox": [ + 304, + 348, + 547, + 421 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 348, + 547, + 421 + ], + "type": "text", + "content": " versus test images of class " + }, + { + "bbox": [ + 304, + 348, + 547, + 421 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 348, + 547, + 421 + ], + "type": "text", + "content": " with the predicted probability of class " + }, + { + "bbox": [ + 304, + 348, + 547, + 421 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 348, + 547, + 421 + ], + "type": "text", + "content": " as score." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "text", + "content": "D5 (87.1% acc.) trained only on 1k by 2.5%, or by 0.6% for EVA-G" + }, + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "inline_equation", + "content": "\\backslash" + }, + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "text", + "content": "14 CLIP 560 trained on MIM (91.9% acc.), as well as BeiT-L" + }, + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "inline_equation", + "content": "\\backslash" + }, + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "text", + "content": "16 fine-tuned from 21k (88.6% acc.) by 0.6%. Thus even SOTA models profit from our SpuFix with negligible difference in accuracy (" + }, + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "inline_equation", + "content": "\\leq" + }, + { + "bbox": [ + 304, + 434, + 547, + 507 + ], + "type": "text", + "content": " 0.1%) and thus the use of SpuFix is recommended for any ImageNet model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 519, + 379, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 519, + 379, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 519, + 379, + 532 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 540, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 547, + 635 + ], + "type": "text", + "content": "We have shown that large-scale identification of spurious features is feasible with our neural PCA components and neural PCA feature visualizations. With \"Spurious ImageNet\" we introduced a novel dataset to evaluate the dependence of ImageNet classifiers on spurious features based on real images. We demonstrated that our SpuFix method mitigates the dependence on harmful spurious features for any ImageNet classifier without costly labeling or re-training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgements We acknowledge support by the DFG, Project number 390727645, the Carl Zeiss Foundation, project \"Certification and Foundations of Safe Machine Learning Systems in Healthcare\". The authors thank the IMPRS-IS for supporting YN." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20243" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Julius Adebayo, Michael Muelly, Hal Abelson, and Been Kim. Post hoc explanations may be ineffective for detecting unknown spurious correlation. In ICLR, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 286, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 286, + 156 + ], + "type": "text", + "content": "[2] Julius Adebayo, Michael Muelly, Ilaria Liccardi, and Been Kim. Debugging tests for model explanations. In NeurIPS, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 286, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 286, + 190 + ], + "type": "text", + "content": "[3] D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Mane. Concrete problems in AI safety. arXiv:1606.06565v2, 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 286, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 286, + 245 + ], + "type": "text", + "content": "[4] Christopher J. Anders, Leander Weber, David Neumann, Wojciech Samek, Klaus-Robert Müller, and Sebastian Lapuschkin. Finding and removing clever hans: Using explanation methods to debug and improve deep models. Information Fusion, 77:261-295, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 286, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 278 + ], + "type": "text", + "content": "[5] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. In NeurIPS, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 286, + 312 + ], + "type": "text", + "content": "[6] Maximilian Augustin, Alexander Meinke, and Matthias Hein. Adversarial robustness on in- and out-distribution improves explainability. In ECCV, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 314, + 286, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 286, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 286, + 356 + ], + "type": "text", + "content": "[7] David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja Hansen, and Klaus-Robert Müller. How to explain individual classification decisions. JMLR, 2010." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 358, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 286, + 379 + ], + "type": "text", + "content": "[8] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 286, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 402 + ], + "type": "text", + "content": "[9] Sara Beery, Grant van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 403, + 286, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 286, + 435 + ], + "type": "text", + "content": "[10] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with ImageNet? arXiv:2006.07159, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 437, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 286, + 468 + ], + "type": "text", + "content": "[11] B. Biggio, I. Corona, D. Maiorca, B. Nelson, N. Srndic, P. Laskov, G. Giacinto, and F. Roli. Evasion attacks against machine learning at test time. In ECML/PPKD, 2013." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 470, + 286, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 286, + 502 + ], + "type": "text", + "content": "[12] Alceu Bissoto, Eduardo Valle, and Sandra Avila. Debiasing skin lesion datasets and models? not so fast. arXiv:2004.11457, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 503, + 286, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 286, + 536 + ], + "type": "text", + "content": "[13] Valentyn Boreiko, Maximilian Augustin, Francesco Croce, Philipp Berens, and Matthias Hein. Sparse visual counterfactual explanations in image space. In " + }, + { + "bbox": [ + 48, + 503, + 286, + 536 + ], + "type": "inline_equation", + "content": "GCPR" + }, + { + "bbox": [ + 48, + 503, + 286, + 536 + ], + "type": "text", + "content": ", 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 537, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 286, + 645 + ], + "type": "text", + "content": "[14] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 679 + ], + "type": "text", + "content": "[15] Myung Jin Choi, Antonio Torralba, and Alan S. Willsky. Context models and out-of-context objects. Pattern Recognition Letters, 33(7):853-862, 2012." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "text", + "content": "[16] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[17] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR workshop, pages 702-703, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 118, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 140 + ], + "type": "text", + "content": "[18] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[19] Ann-Kathrin Dombrowski, Maximillian Alber, Christopher Anders, Marcel Ackermann, Klaus-Robert Müller, and Pan Kessel. Explanations can be manipulated and geometry is to blame. In NeurIPS, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 274 + ], + "type": "text", + "content": "[21] Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. Adversarial robustness as a prior for learned representations, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "text", + "content": "[22] Christian Etmann, Sebastian Lunz, Peter Maass, and Carola-Bibiane Schonlieb. On the connection between adversarial robustness and saliency map interpretability. In ICML, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "text", + "content": "[23] Sabri Eyuboglu, Maya Varma, Khaled Saab, Jean-Benoit Delbrouck, Christopher Lee-Messer, Jared Dunnmon, James Zou, and Christopher Ré. Domino: Discovering systematic errors with cross-modal embeddings. In ICLR, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 354, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 407 + ], + "type": "text", + "content": "[24] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 409, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 409, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 409, + 545, + 453 + ], + "type": "text", + "content": "[25] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 454, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 545, + 497 + ], + "type": "text", + "content": "[26] R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 499, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 499, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 499, + 545, + 520 + ], + "type": "text", + "content": "[27] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In AAAI, 2019." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 522, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 553 + ], + "type": "text", + "content": "[28] A Gupta, C Sun, A Shrivastava, and S Singh. Revisiting the unreasonable effectiveness of data. arXiv preprint arXiv:1707.02968, 2017." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 556, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 578 + ], + "type": "text", + "content": "[29] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 579, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 621 + ], + "type": "text", + "content": "[30] M. Hein, M. Andriushchenko, and J. Bitterwolf. Why ReLU networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In CVPR, 2019." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 655 + ], + "type": "text", + "content": "[31] D. Hendrycks and T. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[32] D. Hendrycks and K. Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[33] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20244" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[34] Sara Hooker, Yann Dauphin, Aaron Courville, and Andrea Frome. Selective brain damage: Measuring the disparate impact of model pruning. In ICLR, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 129 + ], + "type": "text", + "content": "[35] Pieter Abbeel Jonathan Ho, Ajay Jain. Denoising diffusion probabilistic models. In NeurIPS, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 130, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 130, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 130, + 287, + 173 + ], + "type": "text", + "content": "[36] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In ECCV, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 196 + ], + "type": "text", + "content": "[37] A. Kurakin, I. J. Goodfellow, and S. Bengio. Adversarial examples in the physical world. In *ICLR Workshop*, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 198, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 262 + ], + "type": "text", + "content": "[38] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 264, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 307 + ], + "type": "text", + "content": "[39] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 287, + 352 + ], + "type": "text", + "content": "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 354, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 287, + 386 + ], + "type": "text", + "content": "[41] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 388, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 409 + ], + "type": "text", + "content": "[42] Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In NeurIPS, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 443 + ], + "type": "text", + "content": "[43] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 445, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 476 + ], + "type": "text", + "content": "[44] Mazda Moayeri, Sahil Singla, and Soheil Feizi. Hard ImageNet: Segmentations for objects with strong spurious cues. In NeurIPS Datasets and Benchmarks Track, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 478, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 478, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 49, + 478, + 287, + 511 + ], + "type": "text", + "content": "[45] A. Nguyen, J. Yosinski, and J. Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In CVPR, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 513, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 513, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 513, + 287, + 544 + ], + "type": "text", + "content": "[46] Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 546, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 546, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 546, + 287, + 578 + ], + "type": "text", + "content": "[47] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. Proc. CVPR, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 580, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 580, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 580, + 287, + 622 + ], + "type": "text", + "content": "[48] Gregory Plumb, Marco Tulio Ribeiro, and Ameet Talwalkar. Finding and fixing spurious patterns with explanations. Transactions on Machine Learning Research (TMLR), 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 678 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 287, + 712 + ], + "type": "text", + "content": "[50] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In ICML, 2019." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[51] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\": Explaining the predictions of any classifier. In KDD, 2016." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[52] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 2015. License: No license specified." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "text", + "content": "[53] Shibani Santurkar, Dimitris Tsipras, Brandon Tran, Andrew Ilyas, Logan Engstrom, and Aleksander Madry. Image synthesis with a single (robust) classifier. In NeurIPS, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "text", + "content": "[54] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "type": "text", + "content": "[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "text", + "content": "[56] Kathryn Schutte, Olivier Moindrot, Paul Hérent, Jean-Baptiste Schiratti, and Simon Jégou. Using stylegan for visual interpretability of deep learning models on medical images. In NeurIPS Workshop, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "type": "text", + "content": "[57] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 430 + ], + "type": "text", + "content": "[58] Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks. In NeurIPS, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 433, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 475 + ], + "type": "text", + "content": "[59] Rakshith Shetty, Bernt Schiele, and Mario Fritz. Not using the car to see the sidewalk-quantifying and controlling the effects of context in classification and segmentation. In CVPR, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 478, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 520 + ], + "type": "text", + "content": "[60] Krishna Kumar Singh, Dhruv Mahajan, Kristen Grauman, Yong Jae Lee, Matt Feiszli, and Deepti Ghadiyaram. Don't judge an object by its context: learning to overcome contextual bias. In CVPR, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 522, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 544 + ], + "type": "text", + "content": "[61] Sahil Singla and Soheil Feizi. Salient imagenet: How to discover spurious features in deep learning? In ICLR, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "text", + "content": "[62] Sahil Singla, Mazda Moayeri, and Soheil Feizi. Core risk minimization using salient imagenet. arXiv:2203.15566, 2022." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 580, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 612 + ], + "type": "text", + "content": "[63] Sahil Singla, Besmira Nushi, Shital Shah, Ece Kamar, and Eric Horvitz. Understanding failures of deep networks via robust feature extraction. In CVPR, 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "text", + "content": "[64] Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju. Fooling lime and shap: Adversarial attacks on post hoc explanation methods. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 180-186, 2020." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[65] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20245" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[66] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. In TMLR, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[67] Pierre Stock and Moustapha Cisse. Convnets and ImageNet beyond accuracy: Understanding mistakes and uncovering biases. In ECCV, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 186 + ], + "type": "text", + "content": "[68] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 209 + ], + "type": "text", + "content": "[69] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 210, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 232 + ], + "type": "text", + "content": "[70] Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In ICML, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 233, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 255 + ], + "type": "text", + "content": "[71] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR, pages 1521-1528, 2011." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 257, + 287, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 257, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 287, + 279 + ], + "type": "text", + "content": "[72] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "type": "text", + "content": "[73] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. From ImageNet to image classification: Contextualizing progress on benchmarks. In ICML, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "type": "text", + "content": "[74] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "text", + "content": "[75] Sahil Verma, John P. Dickerson, and Keegan Hines. Counterfactual explanations for machine learning: A review. arXiv:2010.10596, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 395, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 395, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 395, + 287, + 437 + ], + "type": "text", + "content": "[76] Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the GDPR. Harvard Journal of Law & Technology, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 440, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 462 + ], + "type": "text", + "content": "[77] Ross Wightman. Pytorch image models. https://github.com/ rwightman/pytorch-image-models, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 463, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 495 + ], + "type": "text", + "content": "[78] Eric Wong, Shibani Santurkar, and Aleksander Madry. Leveraging sparse linear layers for debuggable deep networks. In ICML, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 498, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 541 + ], + "type": "text", + "content": "[79] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 575 + ], + "type": "text", + "content": "[80] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or signal: The role of image backgrounds in object recognition. In ICLR, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 609 + ], + "type": "text", + "content": "[81] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "text", + "content": "[82] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 646, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 668 + ], + "type": "text", + "content": "[83] Cong Xu and Min Yang. Adversarial momentum-contrastive pre-training. arXiv preprint, arXiv:2012.13154v2, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[84] I Zeki Yalniz, Herve Jégou, Kan Chen, Manohar Paluri, and Dhruv Mahajan. Billion-scale semi-supervised learning for image classification. arXiv preprint arXiv:1905.00546, 2019." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 219 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[85] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision observer for visual recognition. PAMI, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[86] John R. Zech, Marcus A. Badgeley, Manway Liu, Anthony B. Costa, Joseph J. Titano, and Eric Karl Oermann. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLOS Medicine, 15(11):1-17, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[87] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 197, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 219 + ], + "type": "text", + "content": "[88] Zhuotun Zhu, Lingxi Xie, and Alan L. Yuille. Object recognition with and without objects. In *IJCAI*, 2017." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20246" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/3ce1c749-00ff-4ed7-ac11-5982848d2220_content_list.json b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/3ce1c749-00ff-4ed7-ac11-5982848d2220_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c441822ac323b928a1075b33befa5c9b495ee32 --- /dev/null +++ b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/3ce1c749-00ff-4ed7-ac11-5982848d2220_content_list.json @@ -0,0 +1,1527 @@ +[ + { + "type": "text", + "text": "Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation", + "text_level": 1, + "bbox": [ + 186, + 130, + 782, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunpeng Zhai $^{1}$ , Peixi Peng $^{2,3,*}$ , Yifan Zhao $^{1}$ , Yangru Huang $^{1}$ , Yonghong Tian $^{1,2,3,*}$ $^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University, Beijing, China \n $^{2}$ School of Electronic and Computer Engineering, Peking University Shenzhen Graduate School, Shenzhen, China \n $^{3}$ Peng Cheng Laboratory, Shenzhen, China", + "bbox": [ + 161, + 202, + 815, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ypzhai, pxpeng, zhaoyf, yhtian}@pku.edu.cn, yrhuang@stu.pku.edu.cn", + "bbox": [ + 187, + 311, + 779, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 361, + 313, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-based reinforcement learning (RL) depends on discriminative representation encoders to abstract the observation states. Despite the great success of increasing CNN parameters for many supervised computer vision tasks, reinforcement learning with temporal-difference (TD) losses cannot benefit from it in most complex environments. In this paper, we analyze that the training instability arises from the oscillating self-overfitting of the heavy-optimizable encoder. We argue that serious oscillation will occur to the parameters when enforced to fit the sensitive TD targets, causing uncertain drifting of the latent state space and thus transmitting these perturbations to the policy learning. To alleviate this phenomenon, we propose a novel asymmetric interactive cooperation approach with the interaction between a heavy-optimizable encoder and a supportive light-optimizable encoder, in which both their advantages are integrated including the highly discriminative capability as well as the training stability. We also present a greedy bootstrapping optimization to isolate the visual perturbations from policy learning, where representation and policy are trained sufficiently by turns. Finally, we demonstrate the effectiveness of our method in utilizing larger visual models by first-person highway driving task CARLA and Vizdoom environments.", + "bbox": [ + 75, + 393, + 473, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 787, + 209, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning complex control from high-dimensional observations such as images is significant for many real-world applications [37, 28]. It puts forward higher requirements for the representation capability of visual encoder models,", + "bbox": [ + 75, + 813, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3416461f6b614d61f15d3450008213d0cc1ecd711d5c084989a3cbdb9549c191.jpg", + "image_caption": [ + "a) Returns during training" + ], + "image_footnote": [], + "bbox": [ + 511, + 369, + 692, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4415384ff751e746add5f4526c56413e1ac2d8534a4703b9498b9597d93b8a1a.jpg", + "image_caption": [ + "b) Parameter oscillation", + "Figure 1. Comparison of DeepMDP agents on CARLA with light- and heavy-weight visual encoders in episode returns (a) and parameter oscillation $\\left(\\sum_{t}^{T}\\|\\nabla\\theta_{t}\\|\\right)/\\|\\sum_{t}^{T}\\nabla\\theta_{t}\\|$ (b). Compared with 4-layer CNN, larger models like ResNet-18 do not improve the RL performance as in supervised learning. The deterioration results from the oscillating self-overfitting of the larger model, which occurs in neither supervised learning nor RL with a lightweight encoder." + ], + "image_footnote": [], + "bbox": [ + 697, + 369, + 872, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "especially in some complex scenes such as self-driving [5] and robot controlling [29]. The last decade has witnessed impressive progress in computer vision by training large-scale networks [18] as they increase the search space of possible solutions. However, such parameter increment of visual models cannot directly benefit the reinforcement learning, and even leads to deterioration of training. For example, as shown in Fig. 1(a), we illustrated the training curves of DeepMDP [7] agents that use different CNN networks as visual encoders on the self-driving environment CARLA [5]. The result shows that using a larger model, e.g., ResNet-18 [18], leads to unstable training and achieves distinctly lower returns than the lighter model with only four convolutional layers.", + "bbox": [ + 496, + 657, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To investigate this phenomenon, we quantified the oscillation of the network parameters by introducing an in-", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "207", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "indicator calculated by the ratio of accumulation of modulus length of the gradient to the module lengths of cumulative gradient within $T$ training steps, $(\\sum_{t}^{T}\\|\\nabla\\theta_{t}\\|)/\\|\\sum_{t}^{T}\\nabla\\theta_{t}\\|$ , where $\\nabla$ is gradient operator and $\\theta_{t}$ is parameters of the last convolution layer in the $t$ -th step. A higher indicator means worse oscillation and instability of parameters during training. As shown in Fig. 1(b), we compared the oscillation in three experiments including i) training ResNet-18 on RL, ii) training 4-layer CNN on RL, and iii) training ResNet-18 on supervised learning (SL) with CIFAR-100 [24]. ResNet-18 suffers much more serious oscillation when training on RL, resulting in deteriorated performance. However, such oscillation occurs neither on SL with ResNet-18 nor on RL with the lighter 4-layer CNN. We name this phenomenon as oscillating self-overfitting, which particularly results from a pathological concurrence of the overfitting capability of large models and the sensitive learning targets of the temporal-difference (TD) loss [25]. Specifically, when the heavyweight parameters are enforced to fit the sensitive TD targets which are partially generated by themselves with a bootstrapping formulation, contradictory gradients will propagate back to oscillate the parameters. This phenomenon results in uncertain drifting of the state space and transmits these perturbations to policy learning. Therefore, stabilizing visual encoders with amounts of parameters from TD losses for performance gain in RL is still an open challenge.", + "bbox": [ + 76, + 90, + 472, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From the perspective of the relation between the encoder and the TD objective, existing representation learning for RL can be categorized into two groups, as illustrated in Fig. 2. One approach jointly learns representation with policy by the TD loss [16, 7, 23]. It efficiently learns the long-term expected returns with light encoders while is incompatible with heavy-optimizable encoders due to the oscillating self-overfitting. Another group of approaches decouples representation learning from RL to avoid instability, where the encoder is learned only by auxiliary dynamic predictive losses [11, 32]. However, it is inaccessible to the expected returns from the bootstrapping objective in RL. In this paper, we propose a novel asymmetric interactive cooperation for representation learning in RL. To take both the advantages of representation capability and the stability for TD targets, it separately trains a main heavy-optimizable encoder and a supportive light-optimizable encoder by auxiliary tasks and TD losses, respectively. And the asymmetric interaction is simultaneously conducted between them to effectively exchange their knowledge from each other, where the heavy one transfer the representation capability by parameter momentum and the light one transfer the long-term expected returns by topological distillation. Hence, the heavy encoder is equipped with the capability of latent state abstraction without oscillating by the TD objective. Moreover, we present a greedy bootstrapping optimization for", + "bbox": [ + 76, + 508, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/64212d11eddc8d9940c85d488b7263a97b323a3bd425821c5f64fa713b2956a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 90, + 694, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5afeef562035408a31e9ad91cf5676a319b1071ef47342e6208ceef07017fcd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 890, + 209 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3cf6e94cf456eed5a8b19103076a63cf391a455c2e6f6b5673719ddfd659339e.jpg", + "image_caption": [ + "Figure 2. Illustrations of different learning paradigms of visual encoders. a) Joint Learning trains the visual encoder end-to-end with reinforcement loss. b) Decoupled learning trains the encoder only by auxiliary tasks. c) The proposed AIC trains a supportive light-optimizable encoder with the expected returns information from TD targets and transfers it to the main visual encoder." + ], + "image_footnote": [], + "bbox": [ + 503, + 212, + 890, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "further stability of training, where representation and policy are trained sufficiently by turns.", + "bbox": [ + 496, + 460, + 890, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper can be summarized in three aspects. First, it investigates the phenomenon of oscillating self-overfitting that leads to deterioration in RL with heavy-optimizable encoders, and proposes a novel asymmetric interactive cooperation to alleviate it. Second, it presents a topological distillation between latent state spaces to learn state abstraction by interactive cooperation without any explicit labels. Third, it presents a greedy bootstrapping optimization for further stability of training. Experiments demonstrate a significant performance gain over the complex and realistic environments of CARLA and Vizdoom.", + "bbox": [ + 496, + 493, + 892, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 691, + 640, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Jointly Representation Learning with RL. Since learning control directly from high-dimension visual observations is hard to converge, many approaches learn the encoder jointly with RL and elaborate auxiliary tasks [39, 38, 26, 6, 21, 34], as shown in Fig. 2(a). The first line of methods utilizes self-supervision to learn the visual invariances [17, 16, 15, 35]. CURL [27] introduced a contrastive loss to learn discriminative features from raw pixels. The second line learns to abstract the state by preserving the property of Markov decision processes (MDPs) [12, 19, 2]. DeepMDP [7] learns the latent space with the same dynamic transition as the environment. DBC [40] learns invariant representations with", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "208", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3c9056d7dbf55b31eb1475a3f0a4f1e33b0fa55b0cbcbd08cc4a1f9992bd3112.jpg", + "image_caption": [ + "Figure 3. Flowchart of the proposed asymmetric interactive cooperation. The representation learning consists of two components including momentum adaptor tuning and topological distilled transferring. In the momentum adaptor tuning, the adaptor is optimized by RL to learn the information of expected returns while the momentum encoder is updated by moving average from the online encoder. In topology distilled transferring, the online encoder is trained by three auxiliary tasks while transferring the information of expected returns by topology distillation from the momentum adaptor encoder. The policy is learned by freezing the online encoder." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 885, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the bisimulation distance. The third line of methods designs regularization to stabilize the training process [23]. DrQ [23] and SVEA [16] propose data regularization using image augmentation of random shift or random convolution. A-LIX [3] adaptively regularizes the convolutional features to prevent overfitting while it doesn't study the heavy visual models. The common drawback of these methods is the difficulty for the heavy-optimizable encoders due to oscillating self-overfitting caused by TD losses, resulting in more serious instability. Other recent works have concerned the stability of large models in RL [25, 31], but they focus on a different component of policy learning instead of representation. We also note recent work RL with Transformer while they only contain slight parameters [4, 20] or rely on supervised imitation learning [36].", + "bbox": [ + 75, + 525, + 472, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decoupled Representation Learning from RL. Recent methods proposed to decouple representation learning from TD losses for efficient training [30, 31], in Fig. 2(b). A typical approach is the world model [9, 8]. It first pretrains an encoder using Variational Autoencoder [33] by the collected rollouts from a random policy and then trains policy with the encoder frozen. However, the encoder does not perform well since the pretraining data is of different distribution from those in policy learning. Recent world model", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "class methods [14, 30, 32, 13, 1], such as Dreamer [11], turn to learn the encoder and the policy alternatively step-by-step, in which the encoder is updated by the newly collected data. Theoretically, these approaches can train any large model stably. But there still exists a bottleneck, that is, the encoder is isolated from the long-term expected returns hidden in the TD losses, which is significant for the state abstraction with respect to the downstream controls. In our work, we tackle this problem by interacting with a supportive light-optimizable encoder, which learns from TD losses stably and transfers its knowledge to the main encoder, as shown in Fig. 2(c).", + "bbox": [ + 496, + 525, + 893, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Approach", + "text_level": 1, + "bbox": [ + 498, + 724, + 728, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 752, + 609, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Asymmetric interactive cooperation alternatively trains a light-optimizable supportive encoder with TD losses to learn the long-term expected returns without oscillating self-overfitting, and a heavy-optimizable main encoder with auxiliary tasks to capture stronger representation capability. Simultaneously, two kinds of interaction are conducted between them to exchange their knowledge including:", + "bbox": [ + 496, + 777, + 890, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Light $\\leftarrow$ Heavy: parameter momentum to absorb the", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "209", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "representation capability, as in Sec. 3.2.", + "bbox": [ + 76, + 90, + 341, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Light $\\rightarrow$ Heavy: topological distillation to learn the long-term expected returns, as in Sec. 3.3.", + "bbox": [ + 75, + 106, + 467, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Beyond them, it presents a greed bootstrapping framework for AIC to alleviate the drifting of the state space for further stability in Sec. 3.4. The flowchart of AIC is illustrated in Fig. 3.", + "bbox": [ + 75, + 136, + 467, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Momentum Adaptor Tuning", + "text_level": 1, + "bbox": [ + 76, + 205, + 331, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attaching adaptor to momentum. To efficiently learn the long-term expected returns, the light-optimizable encoder should satisfy the following two conditions: 1) It only contains a few learnable parameters so that it converges without oscillating self-overfitting. 2) It can utilize the middle-level representation of the heavy-optimizable encoder rather than relying only on the raw observation input. To fulfill the above conditions, we implement the light-optimizable encoder by applying learnable CNN adaptors $\\beta$ to the momentum network $\\theta^m$ of the online heavy encoder $\\theta^o$ . As illustrated in Fig. 3, each $3\\times 3$ convolutional layer in the momentum encoder is connected with a $1\\times 1$ convolutional adaptor module in a parallel manner. Thus the result of each layer is formulated by,", + "bbox": [ + 75, + 229, + 468, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {l + 1} = \\rho \\left(x _ {l}; \\theta_ {l} ^ {m}, \\beta_ {l}\\right) = \\theta_ {l} ^ {m} * x _ {l} + \\beta_ {l} * x _ {l}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 453, + 468, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{l}$ is the input tensor of the $l$ -th layer and $*$ denotes the convolution operator. $\\theta_{l}^{m}$ and $\\beta_{l}$ are the weights of the momentum model and the adaptor, respectively. We name this light-optimizable encoder as momentum adaptor encoder, referring to $f^{m}$ . It interacts with the online heavy encoder $f^{o}$ by parameter momentum to utilize its representation. In each iteration step, the momentum model is updated by moving average of the online encoder $f^{o}$ ,", + "bbox": [ + 75, + 481, + 468, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {m} \\leftarrow \\tau \\theta^ {m} + (1 - \\tau) \\theta^ {o}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 614, + 468, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta^o$ is the parameters of online encoder and $\\tau \\in [0,1]$ is the momentum factor.", + "bbox": [ + 75, + 642, + 467, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptor Tuning. The adaptor is updated with RL by maximizing the expected return,", + "bbox": [ + 75, + 672, + 467, + 702 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\beta^ {*} = \\arg \\max _ {\\beta} \\mathbb {E} \\left[ \\sum_ {t} ^ {\\infty} \\gamma^ {t} R \\left(s _ {t}, a _ {t}, s _ {t + 1}\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 713, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is the discount factor and $R$ is the reward. $s_t$ and $a_{t}$ are the state and action in time step $t$ . Given a batch of transitions $\\{(\\mathbf{o}_t,a_t,r_t,\\mathbf{o}_{t + 1})\\}$ from the replay buffer $\\mathcal{B}$ , $\\mathbf{o}_t\\in \\mathbb{R}^{C\\times H\\times W}$ is the observation stacking multiple frames. $H$ , $W$ are the height and width of the image and $C$ is the channel number. We first projected the observation from images to a latent state $\\mathbf{s}^m = f^m (\\mathbf{o}_t;\\theta^m,\\beta)$ with the momentum adaptor encoder $f^{m}$ . To tune the adaptor with the reinforcement learning, we optimize the predicted state by", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4754f2ecbb5ab2b0ca074510761630ccf7b5dc160c5e7ec2524856184dcaa515.jpg", + "image_caption": [ + "Figure 4. Illustration of the distance between state topology in different latent spaces. $\\psi^m (\\cdot)$ and $\\psi^o (\\cdot)$ denote the similarity function of states." + ], + "image_footnote": [], + "bbox": [ + 529, + 92, + 890, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "conditioning it on a learnable critic head $h_c$ and an actor head $h_a$ for SAC [10] algorithms. Both two heads are implemented by 3-layer MLPs. Then the critic takes state $\\mathbf{s}_t^m$ and action $a_t$ as inputs and learns to predict $Q(\\mathbf{s}_t^m, a_t)$ with the critic loss,", + "bbox": [ + 496, + 296, + 890, + 369 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c r i t i c}} = \\left(Q \\left(\\mathbf {s} _ {t}, a _ {t}\\right) - y \\left(r _ {t}, \\mathbf {s} _ {t + 1}\\right)\\right) ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 381, + 890, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the TD target is given by", + "bbox": [ + 500, + 410, + 710, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny \\left(r _ {t}, \\mathbf {s} _ {t + 1}\\right) = r _ {t} + \\gamma \\left(\\bar {Q} \\left(\\mathbf {s} _ {t + 1}, \\hat {a} _ {t + 1}\\right) - \\alpha_ {r} \\log \\pi \\left(\\hat {a} _ {t + 1} \\mid \\mathbf {s} _ {t + 1}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 436, + 898, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\bar{Q}$ is the target critic by moving average, and $\\pi (\\hat{a}_{t + 1}|\\mathbf{s}_{t + 1})$ is the policy distribution predicted by the actor head. $\\alpha_{r}$ is the temperature parameter in SAC. $\\hat{a}_{t + 1}\\sim \\pi (\\cdot |\\mathbf{s}_{t + 1})$ . The actor learns to generate the action that maximizes the Qvalue predicted by the critic,", + "bbox": [ + 498, + 467, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a c t o r}} = \\alpha_ {r} \\log \\pi (\\hat {a} _ {t} | \\mathbf {s} _ {t}) - Q (\\mathbf {s} _ {t}, \\hat {a} _ {t}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 554, + 890, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By jointly training the adaptor $\\beta$ with the critic $h_c$ and actor $h_a$ , the information of long-term expected returns will be efficiently learned to the adaptor, resulting in a better state abstraction for the control task.", + "bbox": [ + 498, + 580, + 890, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Topology Distilled Transferring", + "text_level": 1, + "bbox": [ + 500, + 650, + 777, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Topology of abstracted states. Since the information of expected returns is stably captured by the momentum adaptor $f^m$ , the core issue is how to transfer that knowledge into the online heavy-optimizable main encoder $f^o$ , making its representation more adaptive to policy learning. To this end, we introduce a teacher-student distillation manner where the momentum adaptor encoder $f^m(\\cdot; \\theta^m, \\beta)$ and the online main encoder $f^o(\\cdot; \\theta^o)$ are considered as the teacher and student, respectively. Learning state abstraction for control is to project the observed states into a latent space, where the adjacency among states is consistent with the transitions in the real environment. For instance, the states near the same time-step in the real environment should also be adjacent in the feature space. Hence, we transfer the capability of state abstraction by distillation of the adjacency, i.e.,", + "bbox": [ + 496, + 672, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "210", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "topological relation in the latent state space. To this end, we first define topological relation $\\mathcal{T}(S)$ in the state space $S$ as the set of similarity between every state pairs $(\\mathbf{s}_i,\\mathbf{s}_j)$ in the data,", + "bbox": [ + 75, + 90, + 468, + 148 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} (S) = \\left\\{\\psi \\left(\\mathbf {s} _ {i}, \\mathbf {s} _ {j}\\right) \\mid \\left(\\mathbf {s} _ {i}, \\mathbf {s} _ {j}\\right) \\in S \\right\\} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 152, + 468, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi (\\mathbf{s}_i,\\mathbf{s}_j)$ denotes the similarity function of states. This set of similarities reveals the structural information of the state space, thus preserving the representation capability of the states. Our objective is to minimize the distance between the teacher's topological relation $\\mathcal{T}(S^m)$ and the student's $\\mathcal{T}(S^o)$ , as illustrated in Fig. 4,", + "bbox": [ + 75, + 176, + 472, + 268 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta^ {m *} = \\arg \\min _ {\\theta^ {m}} D _ {\\mathcal {T}} (\\mathcal {T} (S ^ {m}), \\mathcal {T} (S ^ {o})) \\\\ = \\arg \\min _ {\\theta^ {m}} \\mathbb {E} _ {\\mathbf {o} _ {i}, \\mathbf {o} _ {j} \\sim \\mathcal {B}} \\left[ \\operatorname {D i s t} \\left(\\psi^ {m} \\left(\\mathbf {s} _ {i} ^ {m}, \\mathbf {s} _ {j} ^ {m}\\right), \\psi^ {o} \\left(\\mathbf {s} _ {i} ^ {o}, \\mathbf {s} _ {j} ^ {o}\\right)\\right) \\right]. \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 75, + 279, + 477, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$s_i^m = f^m (o_i|\\theta^m,\\beta), s_i^o = f^o (o_i|\\theta^o), \\psi^m (\\cdot), \\psi^o (\\cdot)$ are the similarity functions among states of $S^{m}, S^{o}$ , respectively.", + "bbox": [ + 75, + 339, + 468, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Topological distillation. Here we introduce the process of topological distillation in detail. Given a batch of observations $\\{\\mathbf{o}_t\\}$ , we encode them to latent states $\\{\\mathbf{s}_t^m\\}$ and $\\{\\mathbf{s}_t^o\\}$ with the momentum adaptor encoder $f^{m}(\\cdot ;\\theta^{m},\\beta)$ and the online main encoder $f^{o}(\\cdot ;\\theta^{o})$ , respectively. In this paper, we suppose the states are abstracted in a high-dimensional Euclidean space, and use $L2$ distance as the similarity function $\\psi^m (\\cdot) = \\| \\cdot \\| ^2$ for the teacher. Then the topological relation within a batch is formulated as an adjacent matrix of the state features from the teacher, referring to $\\mathbf{A}^m\\in \\mathbb{R}^{N\\times N}$ . $N$ is the batch size. $\\mathbf{A}_{[i,j]}^m = \\psi^m (\\mathbf{s}_i^m,\\mathbf{s}_j^m) = \\| \\mathbf{s}_i^m -\\mathbf{s}_j^m\\| ^2$ , where $[i,j]$ denotes the i-th row and the j-th column in the matrix $\\mathbf{A}^m$ . Moreover, due to the unknown scale of the latent state, the matrix is normalized by its average value as $\\| \\mathbf{A}^m\\|$ ,", + "bbox": [ + 75, + 371, + 472, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\mathbf {A} ^ {m} \\right\\| _ {[ i, j ]} = \\mathbf {A} _ {[ i, j ]} ^ {m} / \\left(\\frac {1}{N ^ {2}} \\sum_ {m, n} \\mathbf {A} _ {[ m, n ]} ^ {m}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 612, + 468, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, directly minimizing the distance between topological relations with the same similarity function $\\psi^o$ as $\\psi^m$ will prevent the student online encoder from other basic discrimination. To flexibly distill knowledge and avoid disturbing the basic representation, the similarity function of the student's state space is defined by the $L2$ distance after a linear transformation $\\phi$ with weight $\\mathbf{W}$ ,", + "bbox": [ + 75, + 660, + 468, + 767 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi^ {o} \\left(\\mathbf {s} _ {i} ^ {o}, \\mathbf {s} _ {j} ^ {o}\\right) = \\| \\phi \\left(\\mathbf {s} _ {i} ^ {o}\\right) - \\phi \\left(\\mathbf {s} _ {j} ^ {o}\\right) \\| ^ {2} = \\| \\mathbf {W} ^ {T} \\mathbf {s} _ {i} ^ {o} - \\mathbf {W} ^ {T} \\mathbf {s} _ {j} ^ {o} \\| ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 777, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "And the intra-batch adjacent matrix of the online encoder's state space is computed as $\\mathbf{A}^o$ , where $\\mathbf{A}_{[i,j]}^o = \\psi^o (\\mathbf{s}_i^o,\\mathbf{s}_j^o)$ . Similar normalization as in Eq. 9 is also performed to $\\mathbf{A}^o$ , referring to $\\| \\mathbf{A}^o\\|$ . We distill the information of long-term expected returns from $\\mathbf{s}^m$ to $\\mathbf{s}^o$ by minimizing the L2 distance between their intra-batch similarity matrices $\\mathbf{A}^m$ and", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{A}^o$ . The transferring loss is defined by", + "bbox": [ + 500, + 90, + 758, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a n s}} = \\frac {1}{N ^ {2}} \\sum_ {(i, j) \\in [ 1, N ]} \\operatorname {D i s t} \\left(\\mathbf {A} _ {[ i, j ]} ^ {m}, \\mathbf {A} _ {[ i, j ]} ^ {o}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 114, + 890, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Dist}(u,v) = \\max \\left((u - v)^2 -\\epsilon ,0\\right)$ . Note that $\\epsilon$ is a small loose factor, making the distillation learning more stable to converge.", + "bbox": [ + 496, + 164, + 890, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic prediction. Although the online encoder indirectly learns the long-term expected returns by the above distillation, its potential in representation capability is far from being developed. To this end, the online encoder is learned to capture other basic characters including environment dynamics as well as visual perception by using additional auxiliary tasks. As illustrated in Fig. 3, the predicted states $\\mathbf{s}_t^o$ are conditioned on three auxiliary heads: a reward predictive head $\\mathcal{R}$ , a state predictive head $\\mathcal{P}$ and an observation decoder head $\\mathcal{G}$ . Taking the current states $\\mathbf{s}_t^o$ and actions $a_{t}$ as inputs, the $\\mathcal{R}$ and $\\mathcal{P}$ aim to predict the reward $\\mathcal{R}(\\mathbf{s}_t^o,a_t)$ and next state $\\mathcal{P}(\\mathbf{s}_t^o,a_t)$ , and thus learn a latent state space of which dynamic is consistent with the environment. Dynamic predictive loss is defined as", + "bbox": [ + 496, + 210, + 892, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {d y n a m i c}} = \\mathcal {L} _ {\\text {r e w a r d}} + \\mathcal {L} _ {\\text {s t a t e}} \\\\ = D _ {r} \\left(\\mathcal {R} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right), r _ {t}\\right) + D _ {s} \\left(\\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right), \\mathbf {s} _ {t + 1} ^ {o}\\right)) \\\\ = \\left(\\mathcal {R} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right) - r _ {t}\\right) ^ {2} + \\left\\| \\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right) - \\mathbf {s} _ {t + 1} ^ {o}\\right) \\| ^ {2}. \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 431, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To guarantee temporal capacity, the decoder $\\mathcal{G}$ is trained to predict the next observation with the predicted next state $\\hat{s}_{t + 1}^o = \\mathcal{P}(\\mathbf{s}_t^o,a_t)$ . The decoding loss is defined as", + "bbox": [ + 496, + 501, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {d e c} = D _ {o} \\left(\\hat {\\mathbf {o}} _ {t + 1}, \\mathbf {o} _ {t + 1}\\right) \\\\ = D _ {o} \\left(\\mathcal {G} \\left(\\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, \\mathbf {a} _ {t}\\right)\\right), \\mathbf {o} _ {t + 1}\\right) \\quad H W \\tag {13} \\\\ = \\frac {1}{H W} \\sum_ {h, w = 1} ^ {H, W} \\left(\\mathcal {G} \\left(\\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right)\\right) ^ {[ h, w ]} - \\mathbf {o} _ {t + 1} ^ {[ h, w ]}\\right) ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 558, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the overall loss of the topology distilled transferring is formulated by", + "bbox": [ + 496, + 650, + 890, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e n c o d e r}} = \\mathcal {L} _ {\\text {t r a n s}} + \\mathcal {L} _ {\\text {d y n a m i c}} + \\mathcal {L} _ {\\text {o b s}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 691, + 890, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the online main encoder and auxiliary heads are jointly learned with the adaptor frozen.", + "bbox": [ + 496, + 717, + 890, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Greedy Bootstrapping for AIC", + "text_level": 1, + "bbox": [ + 500, + 756, + 769, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In most deep reinforcement learning approaches, the representation module and policy module are trained in a simultaneous manner. That is, while the policy model is trained from the reinforcement signals the encoder is also simultaneously changing. This phenomenon will disturb policy improvement due to the drifting of state representation. To tackle this problem, this section presents a greedy bootstrapping framework for AIC, which consists of two", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "211", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Greedy Bootstrapping for AIC" + ], + "code_body": "1 Initialize the replay buffer $\\mathcal{B}$ with random episodes. \n2 while not converged do \n3 // Representation learning \n4 Build data-loader with the replay buffer $\\mathcal{B}$ \n5 for update step $c = 1\\dots C$ do \n6 Draw next batch of transitions $\\{\\left(\\mathbf{o}_t,a_t,r_t,\\mathbf{o}_{t + 1}\\right)\\}$ \n7 //Momentum Adaptor Tuning \n8 Update momentum encoder by moving average. \n9 Update adaptor by optimizing Eq. 4. \n10 //Topology Distilled Transferring \n11 Update online encoder by optimizing Eq. 14. \n12 end \n13 // Policy learning (encoder frozen) \n14 for update step $c_{p} = 1\\dots S$ do \n15 Draw transition sequences $\\{\\left(\\mathbf{o}_t,a_t,r_t,\\mathbf{o}_{t + 1}\\right)\\} \\sim \\mathcal{B}$ \n16 Update critic and actor parameters. \n17 end \n18 Add new experiences to the replay buffer $\\mathcal{B}$ \n19 end", + "bbox": [ + 81, + 112, + 436, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "alternative training stages for the representation and policy, respectively. In each representation stage, the encoder keeps training until convergence, and so does the policy in the other stage. The framework is described in detail in Algorithm 1.", + "bbox": [ + 75, + 380, + 468, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Greedy representation learning. Over the latest collected replay buffer, the encoder is trained by iterating all the images in each epoch. Different from sampling a random batch in each iteration like most other methods, our ergodic sampling strategy brings stable training without ignorance of any crucial samples. In each stage of representation learning, the encoder is trained for several epochs until it is converged.", + "bbox": [ + 75, + 455, + 468, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Greedy policy learning. With the main encoder trained sufficiently, the policy is learned continuously by freezing the encoder and conditioned on its output state. Beneficial from the stable and discriminative latent state, the critic and actor are trained efficiently from the reinforcement signal by SAC algorithms and tend to produce better behaviors. The policy is learned continuously for enough steps and then used to interact with the environment for new trajectories.", + "bbox": [ + 75, + 578, + 468, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 714, + 200, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Environments Setup", + "text_level": 1, + "bbox": [ + 76, + 739, + 269, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Benchmarks.** We evaluate our method on two reinforcement learning environments with complex and realistic observations, i.e., CARLA [5] and Vizdoom [22], with continuous and discrete action space, respectively. Both tasks are developed in 3D physical scenes with a complex state space, which calls for deeper visual encoders with stronger representation capability.", + "bbox": [ + 75, + 763, + 468, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We adopt ResNet-18 as the backbone of the visual encoder. Before the last pooling layer, the", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc868855c08398cd6ccc0507c6ed4b1f9735667953c63895640f3d37f92142de.jpg", + "image_caption": [ + "Figure 5. Comparison with other baseline approaches on the CARLA environment." + ], + "image_footnote": [], + "bbox": [ + 519, + 95, + 874, + 271 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6d661de121118efee8834a7077d2d2a7b7374bfcf17d77497586e3fa5dfdc8f1.jpg", + "image_caption": [ + "Figure 6. Ablation Study of our method using ResNet-18 as the visual encoder on the CARLA environment." + ], + "image_footnote": [], + "bbox": [ + 521, + 327, + 870, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "feature map is flattened and then projected into 1024-dim state features with a linear layer and layer norm operation. The momentum factor $\\tau$ is set to 0.05. We use Adam to optimize the parameters of all modules in our approach.", + "bbox": [ + 496, + 565, + 890, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. CARLA Autonomous Driving Environment", + "text_level": 1, + "bbox": [ + 498, + 635, + 869, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental settings. In the autonomous driving task, we construct a highway driving scenario with 20 other vehicles of different models using the CARLA simulator. We use five cameras on the roof of the ego-vehicle, each with a 60-degree view and obtaining images of $84 \\times 84$ pixels. We concatenate them together for observation of $84 \\times 420$ pixels. We stack the last three frames as the inputs. Following DBC [40], the reward is formulated as $r_t = v_{ego}^T \\hat{u}_h \\cdot \\Delta t - \\xi_1 \\cdot \\mathbb{I} - \\xi_2 \\cdot |steer|$ , where $v_{ego}$ is the velocity vector of the ego-vehicle, projected onto the highway's unit vector $\\hat{u}_h$ , and multiplied by time discretization $\\Delta t = 0.05$ to measure highway progression in meters. Impulse $\\mathbb{I} \\in \\mathbb{R}^+$ is caused by collisions, and a steering penalty steer $\\in [-1,1]$ facilitates lane-keeping. The hyperparameters $\\xi_1$ and $\\xi_2$ are set to $10^{-4}$ and 1, respectively.", + "bbox": [ + 496, + 657, + 890, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with other methods. We compare our ap", + "bbox": [ + 500, + 886, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "212", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6977a11160d6966730f280a0113e56266738595547c05a1eacd92b104bfe8d68.jpg", + "image_caption": [ + "Figure 7. Comparison of other representation-decoupled methods on the CARLA environment." + ], + "image_footnote": [], + "bbox": [ + 102, + 99, + 444, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "proach with seven baselines, including joint representation learning methods, i.e., CURL [27], SVEA [16], Deep-MDP [7], DBC [40]; representation decoupled methods, i.e., DreamerV2 [13], Iso-Dream [32], and the baseline SAC [10]. As shown in Fig. 5, our AIC outperforms all baselines by large margins. Compared with joint representation learning, the superior performance can be attributed to that our method stably learns a more discriminative state abstraction with deeper architectures, while avoiding the oscillating self-overfitting. Representation decoupled methods train encoders only on auxiliary tasks without TD losses and thus usually require more time to converge. However, AIC effectively learns the long-term expected returns by interaction with the light-optimizable momentum adaptor encoder, resulting in faster convergence.", + "bbox": [ + 75, + 337, + 468, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. We conduct ablation studies to validate the effectiveness of the proposed topology distilled transferring and the greedy bootstrapping strategy. All the experiments train the ResNet-18 as the encoder. Fig. 6 shows that SAC fails to learn an effective policy with the heavy-optimizable encoder since the state features are deteriorated by the serious oscillating self-overfitting. By applying the dynamic loss and decoder loss, the deterioration is slightly alleviated, since the encoder is encouraged to learn the inherent information of environment dynamics by the targets which are more stable than TD. However, disturbed by the oscillating gradients of the TD loss, the encoder is still hard to converge. Compared to it, the greedy bootstrapping strategy achieves significant improvement, denoted as AIC $w \\backslash o L_{trans}$ , because it interrupts the backward propagation of the unstable gradient and simultaneously it trains the encoder and policy stably and sufficiently. AIC distinctly outperforms the AIC $w \\backslash o L_{trans}$ due to the topology distilled transferring, which additionally incorporates the long-term expected return information into the encoder, making it more adaptive for the downstream controlling.", + "bbox": [ + 75, + 566, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with other decoupling methods. To further", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/643cda3bc755d17d8704662373cdb1563fbc35fa1b1e0134cba03c931b239ce3.jpg", + "image_caption": [ + "Figure 8. Comparison with different architectures of the encoder on the CARLA environment." + ], + "image_footnote": [], + "bbox": [ + 526, + 99, + 867, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f09ebae0686f1312673628b57fdb906cb41a840eb90460e55cb1d330c3477370.jpg", + "image_caption": [ + "Figure 9. Comparison of other stabilizing methods using ResNet-18 on the CARLA environment." + ], + "image_footnote": [], + "bbox": [ + 524, + 332, + 867, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "evaluate the greedy bootstrapping framework, we compare it with other two typical methods that decouple representation learning from RL. Initial decoupling first collects a set of transitions using a random policy and then pretrains the encoder with auxiliary tasks as in Eq. 12 and 13. Afterward, the policy is trained by SAC with the encoder frozen. As shown in Fig. 7, the initial decoupling method performs high-variance results because the performance significantly depends on whether the distribution of the collected pretraining data is similar to those from a learned policy. Stepping decoupling learns the encoder and the policy step by step, respectively. The encoder is trained by transitions from the same data distribution as the policy. However, since the encoder keeps updating during policy learning, the latent state space suffers from unstable drifting, making it difficult to learn policy from the volatile state abstraction. The greedy bootstrapping strategy surpasses both two approaches because it guarantees a determined latent state space that is conducive to the inductive learning of policy. Moreover, it learns the encoder by traversing all the transitions in the replay buffer, prevented from ignorance of some critical images or overfitting to parts of data.", + "bbox": [ + 500, + 568, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "213", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/80a806827e13993b0ce0ce2473d85965cb39a46bbed939721f59dc1ab49c045d.jpg", + "table_caption": [ + "Table 1. Comparison with other methods using ResNet-18 on Medikit of Vizdoom environment." + ], + "table_footnote": [], + "table_body": "
-DQN (Baseline)DrQSVEACURLDeepMDPAIC (Ours)
Episode Returns (mean/std)888±5841361 ± 1004435±185408±2031026±8071686±1283
Min returns252252792428456
Max returns244442001008113637104800
Survival Time (mean/std)64 ±2685±4345±844±972±3796±53
Min Time262627233926
Max Time1352106874191210
", + "bbox": [ + 124, + 101, + 846, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Improvements on different encoders. We evaluate our approach with different architectures as the visual encoder, including a light CNN of 4-layer convolution (Conv-4), and three heavyweight models, i.e., VGG-11, ResNet-18, and ResNet34. For each architecture, we conduct three experiments using SAC, SAC with auxiliary tasks (dynamic loss and decoder loss as in Eq. 12 and 13), and our AIC approach. The comparison is illustrated in Fig. 8. Firstly, all heavyweight models fail to learn effective policies using SAC, where only TD loss is used. The results indicate the oscillating self-overfitting is consistent for heavy models. Secondly, auxiliary tasks improve the performance of SAC with heavyweight encoders, but they are still inferior to the light encoder like Conv-4. The oscillating self-overfitting seriously prevents the encoder from sufficient learning. Finally, AIC consistently improves the performance with all heavy encoders by large margins, as well as the light encoder of Conv-4. It shows a consistent rank of performance that these architectures have achieved in supervised learning tasks. It demonstrates that our method makes full use of the representation capability of heavyweight visual encoders through interactive cooperation.", + "bbox": [ + 75, + 263, + 472, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with stabilizing methods. To evaluate AIC for stabilizing the training of heavy-optimizable encoders, we compare it with SVEA and DrQ using ResNet-18 as the backbone. As shown in Fig. 9, both two methods perform not well, since the increase of parameters significantly exacerbates the challenge of stabilizing. However, AIC exceeds them by large margins, indicating its capability to stably learn heavy models.", + "bbox": [ + 75, + 597, + 468, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative analysis. In order to evaluate the effectiveness of our approach in alleviating the oscillating self-overfitting, we visualize the updating trajectories of the encoder parameters over a set of training steps by TSNE. We use ResNet-18 as the encoder backbone and choose the weights of the last convolution layer for visualization. The comparison between DeepMDP with the TD loss and AIC is shown in Fig. 10, where each point represents the parameters in a step and the color changes from deep to light over the training steps. As we can see, the parameters with the TD loss of the DeepMDP agent suffer from significant oscillation during training. Different from that, the parameters of AIC show a dis", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92a2dc787f62f3400c0431aa8810243c3af0590866542fd6caf95b92f9762793.jpg", + "image_caption": [ + "Figure 10. Parameter updating trajectories of the ResNet-18 encoder over training steps." + ], + "image_footnote": [], + "bbox": [ + 506, + 263, + 691, + 425 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e5972fe737d309c73cc5428bc929eda398f47e27d6794ca1ee8ff3f0d72b6e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 263, + 885, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tinct optimizing direction and only have slight oscillation in a short time. The qualitative results demonstrate AIC effectively alleviates the oscillating self-overfitting for the heavy-optimizable encoder in visual reinforcement learning.", + "bbox": [ + 498, + 484, + 890, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Vizdoom Medikit Environment", + "text_level": 1, + "bbox": [ + 500, + 558, + 774, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental settings. Vizdoom is based on the first-person perspective video game in a semi-realistic 3D world. We use the difficult medikit collecting scenario, where the agent is spawned in a random spot of a maze with an acid surface, which slowly, but constantly, takes away the agent's life. The agent needs to collect medikits and avoid blue vials with poison, both of which appear in random places during the episode. In each step, the agent is allowed to move (forward/backward), or turn (left/right). The reward is formulated as $r_t = 1 - \\xi_3 \\cdot E_m - \\xi_4 \\cdot E_v - \\xi_5 \\cdot E_d$ , where $E_m$ , $E_v$ , $E_d$ denote the events of getting a medikit, a vial and to death, respectively. $\\xi_3 = \\xi_4 = \\xi_5 = 100$ .", + "bbox": [ + 496, + 582, + 890, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Superior performance with heavy encoders. To evaluate our approach, especially with deep deep encoders, we compare it with five methods. All of them use ResNet-18 as backbones. Since the Vizdoom is of discrete action space, we adopt DQN as the baseline. As shown in Table 1, AIC performs superiorly to the compared methods. CURL and SVEA perform poorly, since the observations of such a fine-grained collecting task of first-person perspective are sensitive to random cropping of CURL. And the random conv of", + "bbox": [ + 496, + 765, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "214", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SVEA is not applicable for the low-contrast images in this scenario. However, AIC is able to effectively train heavy models under different conditions.", + "bbox": [ + 76, + 90, + 468, + 135 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 148, + 194, + 165 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we investigate the phenomenon of oscillating self-overfitting which results in the deterioration of RL when training heavy-optimizable encoders. To alleviate this problem, we propose a novel asymmetric interactive cooperation approach. Through interaction between a main heavy-optimizable encoder and a supportive light-optimizable encoder, both advantages of them are integrated including the highly discriminative capability and the training stability. We further present a greedy bootstrapping framework for further stability. Experimental results of CARLA and Vizdoom show that AIC achieves competitive performance over complex and realistic environments.", + "bbox": [ + 75, + 174, + 472, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 368, + 243, + 386 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is partially supported by grants from the Key-Area Research and Development Program of Guangdong Province under contact No. 2020B0101380001 and grants from the National Natural Science Foundation of China under contract No. 62027804, No. 61825101, No. 62088102 and No. 62202010. The computing resources of Pengcheng Cloudbrain are used in this research.", + "bbox": [ + 75, + 393, + 468, + 498 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 511, + 173, + 526 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Homanga Bharadhwaj, Mohammad Babaeizadeh, Dumitru Erhan, and Sergey Levine. Information prioritization through empowerment in visual model-based rl. arXiv preprint arXiv:2204.08585, 2022.", + "[2] Pablo Samuel Castro. Scalable methods for computing state similarity in deterministic markov decision processes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10069-10076, 2020.", + "[3] Edoardo Cetin, Philip J Ball, Steve Roberts, and Oya Celiktutan. Stabilizing off-policy deep reinforcement learning from pixels. arXiv preprint arXiv:2207.00986, 2022.", + "[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021.", + "[5] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017.", + "[6] Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 2786-2793. IEEE, 2017.", + "[7] Carles Gelada, Saurabh Kumar, Jacob Buckman, Ofir Nachum, and Marc G Bellemare. Deepmdp: Learning" + ], + "bbox": [ + 86, + 536, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "continuous latent space models for representation learning. In International Conference on Machine Learning, pages 2170-2179. PMLR, 2019.", + "[8] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018.", + "[9] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018.", + "[10] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018.", + "[11] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019.", + "[12] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International conference on machine learning, pages 2555-2565. PMLR, 2019.", + "[13] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020.", + "[14] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023.", + "[15] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. arXiv preprint arXiv:2007.04309, 2020.", + "[16] Nicklas Hansen, Hao Su, and Xiaolong Wang. Stabilizing deep q-learning with convnets and vision transformers under data augmentation. 2021.", + "[17] Nicklas Hansen and Xiaolong Wang. Generalization in reinforcement learning by soft data augmentation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13611-13617. IEEE, 2021.", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[19] Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016.", + "[20] Michael Janner, Qiyang Li, and Sergey Levine. Offline reinforcement learning as one big sequence modeling problem. Advances in neural information processing systems, 34:1273-1286, 2021.", + "[21] Lukasz Kaiser, Mohammad Babaeizadeh, Piotr Milos, Blazej Osinski, Roy H Campbell, Konrad Czechowski, Dumitru Erhan, Chelsea Finn, Piotr Kozakowski, Sergey Levine, et al. Model-based reinforcement learning for atari. arXiv preprint arXiv:1903.00374, 2019." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "215", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Michal Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Vizdoom: A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE conference on computational intelligence and games (CIG), pages 1-8. IEEE, 2016.", + "[23] Ilya Kostrikov, Denis Yarats, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. arXiv preprint arXiv:2004.13649, 2020.", + "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.", + "[25] Aviral Kumar, Rishabh Agarwal, Dibya Ghosh, and Sergey Levine. Implicit under-parameterization inhibits data-efficient deep reinforcement learning. arXiv preprint arXiv:2010.14498, 2020.", + "[26] Misha Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. Advances in neural information processing systems, 33:19884-19895, 2020.", + "[27] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In International Conference on Machine Learning, pages 5639-5650. PMLR, 2020.", + "[28] Alex X. Lee, Anusha Nagabandi, Pieter Abbeel, and Sergey Levine. Stochastic latent actor-critic: Deep reinforcement learning with a latent variable model. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 741-752. Curran Associates, Inc., 2020.", + "[29] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. The Journal of Machine Learning Research, 17(1):1334-1373, 2016.", + "[30] Masashi Okada and Tadahiro Taniguchi. Dreaming: Model-based reinforcement learning by latent imagination without reconstruction. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4209-4215. IEEE, 2021.", + "[31] Kei Ota, Devesh K Jha, and Asako Kanezaki. Training larger networks for deep reinforcement learning. arXiv preprint arXiv:2102.07920, 2021.", + "[32] Minting Pan, Xiangming Zhu, Yunbo Wang, and Xiaokang Yang. Iso-dream: Isolating and leveraging noncontrollable visual dynamics in world models. In Advances in Neural Information Processing Systems.", + "[33] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014.", + "[34] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International Conference on Machine Learning, pages 8583-8592. PMLR, 2020.", + "[35] Evan Shelhamer, Parsa Mahmoudieh, Max Argus, and Trevor Darrell. Loss is its own reward: Self-supervision for reinforcement learning. arXiv preprint arXiv:1612.07307, 2016." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Oriol Vinyals, Igor Babuschkin, Wojciech M Czarnecki, Michael Mathieu, Andrew Dudzik, Junyoung Chung, David H Choi, Richard Powell, Timo Ewalds, Petko Georgiev, et al. Grandmaster level in starcraft ii using multiagent reinforcement learning. Nature, 575(7782):350-354, 2019.", + "[37] Manuel Watter, Jost Springenberg, Joschka Boedecker, and Martin Riedmiller. Embed to control: A locally linear latent dynamics model for control from raw images. Advances in neural information processing systems, 28, 2015.", + "[38] Denis Yarats, Amy Zhang, Ilya Kostrikov, Brandon Amos, Joelle Pineau, and Rob Fergus. Improving sample efficiency in model-free reinforcement learning from images. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10674-10681, 2021.", + "[39] Tao Yu, Zhizheng Zhang, Cuiling Lan, Yan Lu, and Zhibo Chen. Mask-based latent reconstruction for reinforcement learning. In Advances in Neural Information Processing Systems.", + "[40] Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant representations for reinforcement learning without reconstruction. arXiv preprint arXiv:2006.10742, 2020." + ], + "bbox": [ + 501, + 90, + 892, + 415 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "216", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/3ce1c749-00ff-4ed7-ac11-5982848d2220_model.json b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/3ce1c749-00ff-4ed7-ac11-5982848d2220_model.json new file mode 100644 index 0000000000000000000000000000000000000000..de8e73401a18e204d9d951979db46a46407642ce --- /dev/null +++ b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/3ce1c749-00ff-4ed7-ac11-5982848d2220_model.json @@ -0,0 +1,2013 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.784, + 0.176 + ], + "angle": 0, + "content": "Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.203, + 0.816, + 0.31 + ], + "angle": 0, + "content": "Yunpeng Zhai\\(^{1}\\), Peixi Peng\\(^{2,3,*}\\), Yifan Zhao\\(^{1}\\), Yangru Huang\\(^{1}\\), Yonghong Tian\\(^{1,2,3,*}\\) \n\\(^{1}\\)National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University, Beijing, China \n\\(^{2}\\)School of Electronic and Computer Engineering, Peking University Shenzhen Graduate School, Shenzhen, China \n\\(^{3}\\)Peng Cheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.313, + 0.78, + 0.327 + ], + "angle": 0, + "content": "{ypzhai, pxpeng, zhaoyf, yhtian}@pku.edu.cn, yrhuang@stu.pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.362, + 0.314, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.474, + 0.758 + ], + "angle": 0, + "content": "Vision-based reinforcement learning (RL) depends on discriminative representation encoders to abstract the observation states. Despite the great success of increasing CNN parameters for many supervised computer vision tasks, reinforcement learning with temporal-difference (TD) losses cannot benefit from it in most complex environments. In this paper, we analyze that the training instability arises from the oscillating self-overfitting of the heavy-optimizable encoder. We argue that serious oscillation will occur to the parameters when enforced to fit the sensitive TD targets, causing uncertain drifting of the latent state space and thus transmitting these perturbations to the policy learning. To alleviate this phenomenon, we propose a novel asymmetric interactive cooperation approach with the interaction between a heavy-optimizable encoder and a supportive light-optimizable encoder, in which both their advantages are integrated including the highly discriminative capability as well as the training stability. We also present a greedy bootstrapping optimization to isolate the visual perturbations from policy learning, where representation and policy are trained sufficiently by turns. Finally, we demonstrate the effectiveness of our method in utilizing larger visual models by first-person highway driving task CARLA and Vizdoom environments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.788, + 0.21, + 0.804 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.814, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Learning complex control from high-dimensional observations such as images is significant for many real-world applications [37, 28]. It puts forward higher requirements for the representation capability of visual encoder models," + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.37, + 0.693, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.505, + 0.671, + 0.517 + ], + "angle": 0, + "content": "a) Returns during training" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.37, + 0.874, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.505, + 0.847, + 0.516 + ], + "angle": 0, + "content": "b) Parameter oscillation" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.633 + ], + "angle": 0, + "content": "Figure 1. Comparison of DeepMDP agents on CARLA with light- and heavy-weight visual encoders in episode returns (a) and parameter oscillation \\(\\left(\\sum_{t}^{T}\\|\\nabla\\theta_{t}\\|\\right)/\\|\\sum_{t}^{T}\\nabla\\theta_{t}\\|\\) (b). Compared with 4-layer CNN, larger models like ResNet-18 do not improve the RL performance as in supervised learning. The deterioration results from the oscillating self-overfitting of the larger model, which occurs in neither supervised learning nor RL with a lightweight encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.87 + ], + "angle": 0, + "content": "especially in some complex scenes such as self-driving [5] and robot controlling [29]. The last decade has witnessed impressive progress in computer vision by training large-scale networks [18] as they increase the search space of possible solutions. However, such parameter increment of visual models cannot directly benefit the reinforcement learning, and even leads to deterioration of training. For example, as shown in Fig. 1(a), we illustrated the training curves of DeepMDP [7] agents that use different CNN networks as visual encoders on the self-driving environment CARLA [5]. The result shows that using a larger model, e.g., ResNet-18 [18], leads to unstable training and achieves distinctly lower returns than the lighter model with only four convolutional layers." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To investigate this phenomenon, we quantified the oscillation of the network parameters by introducing an in-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "207" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.501 + ], + "angle": 0, + "content": "indicator calculated by the ratio of accumulation of modulus length of the gradient to the module lengths of cumulative gradient within \\(T\\) training steps, \\((\\sum_{t}^{T}\\|\\nabla\\theta_{t}\\|)/\\|\\sum_{t}^{T}\\nabla\\theta_{t}\\|\\), where \\(\\nabla\\) is gradient operator and \\(\\theta_{t}\\) is parameters of the last convolution layer in the \\(t\\)-th step. A higher indicator means worse oscillation and instability of parameters during training. As shown in Fig. 1(b), we compared the oscillation in three experiments including i) training ResNet-18 on RL, ii) training 4-layer CNN on RL, and iii) training ResNet-18 on supervised learning (SL) with CIFAR-100 [24]. ResNet-18 suffers much more serious oscillation when training on RL, resulting in deteriorated performance. However, such oscillation occurs neither on SL with ResNet-18 nor on RL with the lighter 4-layer CNN. We name this phenomenon as oscillating self-overfitting, which particularly results from a pathological concurrence of the overfitting capability of large models and the sensitive learning targets of the temporal-difference (TD) loss [25]. Specifically, when the heavyweight parameters are enforced to fit the sensitive TD targets which are partially generated by themselves with a bootstrapping formulation, contradictory gradients will propagate back to oscillate the parameters. This phenomenon results in uncertain drifting of the state space and transmits these perturbations to policy learning. Therefore, stabilizing visual encoders with amounts of parameters from TD losses for performance gain in RL is still an open challenge." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.509, + 0.473, + 0.904 + ], + "angle": 0, + "content": "From the perspective of the relation between the encoder and the TD objective, existing representation learning for RL can be categorized into two groups, as illustrated in Fig. 2. One approach jointly learns representation with policy by the TD loss [16, 7, 23]. It efficiently learns the long-term expected returns with light encoders while is incompatible with heavy-optimizable encoders due to the oscillating self-overfitting. Another group of approaches decouples representation learning from RL to avoid instability, where the encoder is learned only by auxiliary dynamic predictive losses [11, 32]. However, it is inaccessible to the expected returns from the bootstrapping objective in RL. In this paper, we propose a novel asymmetric interactive cooperation for representation learning in RL. To take both the advantages of representation capability and the stability for TD targets, it separately trains a main heavy-optimizable encoder and a supportive light-optimizable encoder by auxiliary tasks and TD losses, respectively. And the asymmetric interaction is simultaneously conducted between them to effectively exchange their knowledge from each other, where the heavy one transfer the representation capability by parameter momentum and the light one transfer the long-term expected returns by topological distillation. Hence, the heavy encoder is equipped with the capability of latent state abstraction without oscillating by the TD objective. Moreover, we present a greedy bootstrapping optimization for" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.695, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.091, + 0.891, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.213, + 0.891, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.351, + 0.895, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Illustrations of different learning paradigms of visual encoders. a) Joint Learning trains the visual encoder end-to-end with reinforcement loss. b) Decoupled learning trains the encoder only by auxiliary tasks. c) The proposed AIC trains a supportive light-optimizable encoder with the expected returns information from TD targets and transfers it to the main visual encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.492 + ], + "angle": 0, + "content": "further stability of training, where representation and policy are trained sufficiently by turns." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.893, + 0.676 + ], + "angle": 0, + "content": "The main contributions of this paper can be summarized in three aspects. First, it investigates the phenomenon of oscillating self-overfitting that leads to deterioration in RL with heavy-optimizable encoders, and proposes a novel asymmetric interactive cooperation to alleviate it. Second, it presents a topological distillation between latent state spaces to learn state abstraction by interactive cooperation without any explicit labels. Third, it presents a greedy bootstrapping optimization for further stability of training. Experiments demonstrate a significant performance gain over the complex and realistic environments of CARLA and Vizdoom." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.693, + 0.642, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Jointly Representation Learning with RL. Since learning control directly from high-dimension visual observations is hard to converge, many approaches learn the encoder jointly with RL and elaborate auxiliary tasks [39, 38, 26, 6, 21, 34], as shown in Fig. 2(a). The first line of methods utilizes self-supervision to learn the visual invariances [17, 16, 15, 35]. CURL [27] introduced a contrastive loss to learn discriminative features from raw pixels. The second line learns to abstract the state by preserving the property of Markov decision processes (MDPs) [12, 19, 2]. DeepMDP [7] learns the latent space with the same dynamic transition as the environment. DBC [40] learns invariant representations with" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "208" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.887, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.431, + 0.896, + 0.504 + ], + "angle": 0, + "content": "Figure 3. Flowchart of the proposed asymmetric interactive cooperation. The representation learning consists of two components including momentum adaptor tuning and topological distilled transferring. In the momentum adaptor tuning, the adaptor is optimized by RL to learn the information of expected returns while the momentum encoder is updated by moving average from the online encoder. In topology distilled transferring, the online encoder is trained by three auxiliary tasks while transferring the information of expected returns by topology distillation from the momentum adaptor encoder. The policy is learned by freezing the online encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.473, + 0.753 + ], + "angle": 0, + "content": "the bisimulation distance. The third line of methods designs regularization to stabilize the training process [23]. DrQ [23] and SVEA [16] propose data regularization using image augmentation of random shift or random convolution. A-LIX [3] adaptively regularizes the convolutional features to prevent overfitting while it doesn't study the heavy visual models. The common drawback of these methods is the difficulty for the heavy-optimizable encoders due to oscillating self-overfitting caused by TD losses, resulting in more serious instability. Other recent works have concerned the stability of large models in RL [25, 31], but they focus on a different component of policy learning instead of representation. We also note recent work RL with Transformer while they only contain slight parameters [4, 20] or rely on supervised imitation learning [36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Decoupled Representation Learning from RL. Recent methods proposed to decouple representation learning from TD losses for efficient training [30, 31], in Fig. 2(b). A typical approach is the world model [9, 8]. It first pretrains an encoder using Variational Autoencoder [33] by the collected rollouts from a random policy and then trains policy with the encoder frozen. However, the encoder does not perform well since the pretraining data is of different distribution from those in policy learning. Recent world model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.895, + 0.709 + ], + "angle": 0, + "content": "class methods [14, 30, 32, 13, 1], such as Dreamer [11], turn to learn the encoder and the policy alternatively step-by-step, in which the encoder is updated by the newly collected data. Theoretically, these approaches can train any large model stably. But there still exists a bottleneck, that is, the encoder is isolated from the long-term expected returns hidden in the TD losses, which is significant for the state abstraction with respect to the downstream controls. In our work, we tackle this problem by interacting with a supportive light-optimizable encoder, which learns from TD losses stably and transfers its knowledge to the main encoder, as shown in Fig. 2(c)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.726, + 0.73, + 0.745 + ], + "angle": 0, + "content": "3. The Proposed Approach" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.753, + 0.611, + 0.768 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.778, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Asymmetric interactive cooperation alternatively trains a light-optimizable supportive encoder with TD losses to learn the long-term expected returns without oscillating self-overfitting, and a heavy-optimizable main encoder with auxiliary tasks to capture stronger representation capability. Simultaneously, two kinds of interaction are conducted between them to exchange their knowledge including:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "- Light \\(\\leftarrow\\) Heavy: parameter momentum to absorb the" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "209" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.342, + 0.106 + ], + "angle": 0, + "content": "representation capability, as in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.137 + ], + "angle": 0, + "content": "- Light \\(\\rightarrow\\) Heavy: topological distillation to learn the long-term expected returns, as in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.197 + ], + "angle": 0, + "content": "Beyond them, it presents a greed bootstrapping framework for AIC to alleviate the drifting of the state space for further stability in Sec. 3.4. The flowchart of AIC is illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.207, + 0.332, + 0.224 + ], + "angle": 0, + "content": "3.2. Momentum Adaptor Tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Attaching adaptor to momentum. To efficiently learn the long-term expected returns, the light-optimizable encoder should satisfy the following two conditions: 1) It only contains a few learnable parameters so that it converges without oscillating self-overfitting. 2) It can utilize the middle-level representation of the heavy-optimizable encoder rather than relying only on the raw observation input. To fulfill the above conditions, we implement the light-optimizable encoder by applying learnable CNN adaptors \\(\\beta\\) to the momentum network \\(\\theta^m\\) of the online heavy encoder \\(\\theta^o\\). As illustrated in Fig. 3, each \\(3\\times 3\\) convolutional layer in the momentum encoder is connected with a \\(1\\times 1\\) convolutional adaptor module in a parallel manner. Thus the result of each layer is formulated by," + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.454, + 0.469, + 0.47 + ], + "angle": 0, + "content": "\\[\nx _ {l + 1} = \\rho \\left(x _ {l}; \\theta_ {l} ^ {m}, \\beta_ {l}\\right) = \\theta_ {l} ^ {m} * x _ {l} + \\beta_ {l} * x _ {l}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.469, + 0.603 + ], + "angle": 0, + "content": "where \\( x_{l} \\) is the input tensor of the \\( l \\)-th layer and \\( * \\) denotes the convolution operator. \\( \\theta_{l}^{m} \\) and \\( \\beta_{l} \\) are the weights of the momentum model and the adaptor, respectively. We name this light-optimizable encoder as momentum adaptor encoder, referring to \\( f^{m} \\). It interacts with the online heavy encoder \\( f^{o} \\) by parameter momentum to utilize its representation. In each iteration step, the momentum model is updated by moving average of the online encoder \\( f^{o} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.615, + 0.469, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\theta^ {m} \\leftarrow \\tau \\theta^ {m} + (1 - \\tau) \\theta^ {o}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.468, + 0.672 + ], + "angle": 0, + "content": "where \\(\\theta^o\\) is the parameters of online encoder and \\(\\tau \\in [0,1]\\) is the momentum factor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.468, + 0.703 + ], + "angle": 0, + "content": "Adaptor Tuning. The adaptor is updated with RL by maximizing the expected return," + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.714, + 0.469, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\beta^ {*} = \\arg \\max _ {\\beta} \\mathbb {E} \\left[ \\sum_ {t} ^ {\\infty} \\gamma^ {t} R \\left(s _ {t}, a _ {t}, s _ {t + 1}\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the discount factor and \\(R\\) is the reward. \\(s_t\\) and \\(a_{t}\\) are the state and action in time step \\(t\\). Given a batch of transitions \\(\\{(\\mathbf{o}_t,a_t,r_t,\\mathbf{o}_{t + 1})\\}\\) from the replay buffer \\(\\mathcal{B}\\), \\(\\mathbf{o}_t\\in \\mathbb{R}^{C\\times H\\times W}\\) is the observation stacking multiple frames. \\(H\\), \\(W\\) are the height and width of the image and \\(C\\) is the channel number. We first projected the observation from images to a latent state \\(\\mathbf{s}^m = f^m (\\mathbf{o}_t;\\theta^m,\\beta)\\) with the momentum adaptor encoder \\(f^{m}\\). To tune the adaptor with the reinforcement learning, we optimize the predicted state by" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.093, + 0.891, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.233, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Illustration of the distance between state topology in different latent spaces. \\(\\psi^m (\\cdot)\\) and \\(\\psi^o (\\cdot)\\) denote the similarity function of states." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.371 + ], + "angle": 0, + "content": "conditioning it on a learnable critic head \\( h_c \\) and an actor head \\( h_a \\) for SAC [10] algorithms. Both two heads are implemented by 3-layer MLPs. Then the critic takes state \\( \\mathbf{s}_t^m \\) and action \\( a_t \\) as inputs and learns to predict \\( Q(\\mathbf{s}_t^m, a_t) \\) with the critic loss," + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.382, + 0.891, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c r i t i c}} = \\left(Q \\left(\\mathbf {s} _ {t}, a _ {t}\\right) - y \\left(r _ {t}, \\mathbf {s} _ {t + 1}\\right)\\right) ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.411, + 0.712, + 0.426 + ], + "angle": 0, + "content": "where the TD target is given by" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.437, + 0.9, + 0.468 + ], + "angle": 0, + "content": "\\[\ny \\left(r _ {t}, \\mathbf {s} _ {t + 1}\\right) = r _ {t} + \\gamma \\left(\\bar {Q} \\left(\\mathbf {s} _ {t + 1}, \\hat {a} _ {t + 1}\\right) - \\alpha_ {r} \\log \\pi \\left(\\hat {a} _ {t + 1} \\mid \\mathbf {s} _ {t + 1}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.468, + 0.891, + 0.543 + ], + "angle": 0, + "content": "\\(\\bar{Q}\\) is the target critic by moving average, and \\(\\pi (\\hat{a}_{t + 1}|\\mathbf{s}_{t + 1})\\) is the policy distribution predicted by the actor head. \\(\\alpha_{r}\\) is the temperature parameter in SAC. \\(\\hat{a}_{t + 1}\\sim \\pi (\\cdot |\\mathbf{s}_{t + 1})\\) . The actor learns to generate the action that maximizes the Qvalue predicted by the critic," + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.555, + 0.891, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a c t o r}} = \\alpha_ {r} \\log \\pi (\\hat {a} _ {t} | \\mathbf {s} _ {t}) - Q (\\mathbf {s} _ {t}, \\hat {a} _ {t}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.582, + 0.891, + 0.641 + ], + "angle": 0, + "content": "By jointly training the adaptor \\(\\beta\\) with the critic \\(h_c\\) and actor \\(h_a\\), the information of long-term expected returns will be efficiently learned to the adaptor, resulting in a better state abstraction for the control task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.779, + 0.668 + ], + "angle": 0, + "content": "3.3. Topology Distilled Transferring" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Topology of abstracted states. Since the information of expected returns is stably captured by the momentum adaptor \\( f^m \\), the core issue is how to transfer that knowledge into the online heavy-optimizable main encoder \\( f^o \\), making its representation more adaptive to policy learning. To this end, we introduce a teacher-student distillation manner where the momentum adaptor encoder \\( f^m(\\cdot; \\theta^m, \\beta) \\) and the online main encoder \\( f^o(\\cdot; \\theta^o) \\) are considered as the teacher and student, respectively. Learning state abstraction for control is to project the observed states into a latent space, where the adjacency among states is consistent with the transitions in the real environment. For instance, the states near the same time-step in the real environment should also be adjacent in the feature space. Hence, we transfer the capability of state abstraction by distillation of the adjacency, i.e.," + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "210" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.15 + ], + "angle": 0, + "content": "topological relation in the latent state space. To this end, we first define topological relation \\(\\mathcal{T}(S)\\) in the state space \\(S\\) as the set of similarity between every state pairs \\((\\mathbf{s}_i,\\mathbf{s}_j)\\) in the data," + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.153, + 0.469, + 0.17 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} (S) = \\left\\{\\psi \\left(\\mathbf {s} _ {i}, \\mathbf {s} _ {j}\\right) \\mid \\left(\\mathbf {s} _ {i}, \\mathbf {s} _ {j}\\right) \\in S \\right\\} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.178, + 0.473, + 0.27 + ], + "angle": 0, + "content": "where \\(\\psi (\\mathbf{s}_i,\\mathbf{s}_j)\\) denotes the similarity function of states. This set of similarities reveals the structural information of the state space, thus preserving the representation capability of the states. Our objective is to minimize the distance between the teacher's topological relation \\(\\mathcal{T}(S^m)\\) and the student's \\(\\mathcal{T}(S^o)\\), as illustrated in Fig. 4," + }, + { + "type": "equation", + "bbox": [ + 0.076, + 0.28, + 0.478, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta^ {m *} = \\arg \\min _ {\\theta^ {m}} D _ {\\mathcal {T}} (\\mathcal {T} (S ^ {m}), \\mathcal {T} (S ^ {o})) \\\\ = \\arg \\min _ {\\theta^ {m}} \\mathbb {E} _ {\\mathbf {o} _ {i}, \\mathbf {o} _ {j} \\sim \\mathcal {B}} \\left[ \\operatorname {D i s t} \\left(\\psi^ {m} \\left(\\mathbf {s} _ {i} ^ {m}, \\mathbf {s} _ {j} ^ {m}\\right), \\psi^ {o} \\left(\\mathbf {s} _ {i} ^ {o}, \\mathbf {s} _ {j} ^ {o}\\right)\\right) \\right]. \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.371 + ], + "angle": 0, + "content": "\\(s_i^m = f^m (o_i|\\theta^m,\\beta), s_i^o = f^o (o_i|\\theta^o), \\psi^m (\\cdot), \\psi^o (\\cdot)\\) are the similarity functions among states of \\(S^{m}, S^{o}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.473, + 0.603 + ], + "angle": 0, + "content": "Topological distillation. Here we introduce the process of topological distillation in detail. Given a batch of observations \\(\\{\\mathbf{o}_t\\}\\), we encode them to latent states \\(\\{\\mathbf{s}_t^m\\}\\) and \\(\\{\\mathbf{s}_t^o\\}\\) with the momentum adaptor encoder \\(f^{m}(\\cdot ;\\theta^{m},\\beta)\\) and the online main encoder \\(f^{o}(\\cdot ;\\theta^{o})\\), respectively. In this paper, we suppose the states are abstracted in a high-dimensional Euclidean space, and use \\(L2\\) distance as the similarity function \\(\\psi^m (\\cdot) = \\| \\cdot \\| ^2\\) for the teacher. Then the topological relation within a batch is formulated as an adjacent matrix of the state features from the teacher, referring to \\(\\mathbf{A}^m\\in \\mathbb{R}^{N\\times N}\\). \\(N\\) is the batch size. \\(\\mathbf{A}_{[i,j]}^m = \\psi^m (\\mathbf{s}_i^m,\\mathbf{s}_j^m) = \\| \\mathbf{s}_i^m -\\mathbf{s}_j^m\\| ^2\\), where \\([i,j]\\) denotes the i-th row and the j-th column in the matrix \\(\\mathbf{A}^m\\). Moreover, due to the unknown scale of the latent state, the matrix is normalized by its average value as \\(\\| \\mathbf{A}^m\\|\\)," + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.613, + 0.469, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\mathbf {A} ^ {m} \\right\\| _ {[ i, j ]} = \\mathbf {A} _ {[ i, j ]} ^ {m} / \\left(\\frac {1}{N ^ {2}} \\sum_ {m, n} \\mathbf {A} _ {[ m, n ]} ^ {m}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.47, + 0.768 + ], + "angle": 0, + "content": "However, directly minimizing the distance between topological relations with the same similarity function \\(\\psi^o\\) as \\(\\psi^m\\) will prevent the student online encoder from other basic discrimination. To flexibly distill knowledge and avoid disturbing the basic representation, the similarity function of the student's state space is defined by the \\(L2\\) distance after a linear transformation \\(\\phi\\) with weight \\(\\mathbf{W}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.778, + 0.469, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\psi^ {o} \\left(\\mathbf {s} _ {i} ^ {o}, \\mathbf {s} _ {j} ^ {o}\\right) = \\| \\phi \\left(\\mathbf {s} _ {i} ^ {o}\\right) - \\phi \\left(\\mathbf {s} _ {j} ^ {o}\\right) \\| ^ {2} = \\| \\mathbf {W} ^ {T} \\mathbf {s} _ {i} ^ {o} - \\mathbf {W} ^ {T} \\mathbf {s} _ {j} ^ {o} \\| ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.901 + ], + "angle": 0, + "content": "And the intra-batch adjacent matrix of the online encoder's state space is computed as \\(\\mathbf{A}^o\\), where \\(\\mathbf{A}_{[i,j]}^o = \\psi^o (\\mathbf{s}_i^o,\\mathbf{s}_j^o)\\). Similar normalization as in Eq. 9 is also performed to \\(\\mathbf{A}^o\\), referring to \\(\\| \\mathbf{A}^o\\|\\). We distill the information of long-term expected returns from \\(\\mathbf{s}^m\\) to \\(\\mathbf{s}^o\\) by minimizing the L2 distance between their intra-batch similarity matrices \\(\\mathbf{A}^m\\) and" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.092, + 0.759, + 0.107 + ], + "angle": 0, + "content": "\\(\\mathbf{A}^o\\). The transferring loss is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.116, + 0.892, + 0.155 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a n s}} = \\frac {1}{N ^ {2}} \\sum_ {(i, j) \\in [ 1, N ]} \\operatorname {D i s t} \\left(\\mathbf {A} _ {[ i, j ]} ^ {m}, \\mathbf {A} _ {[ i, j ]} ^ {o}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.165, + 0.892, + 0.21 + ], + "angle": 0, + "content": "where \\(\\mathrm{Dist}(u,v) = \\max \\left((u - v)^2 -\\epsilon ,0\\right)\\). Note that \\(\\epsilon\\) is a small loose factor, making the distillation learning more stable to converge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.211, + 0.893, + 0.422 + ], + "angle": 0, + "content": "Dynamic prediction. Although the online encoder indirectly learns the long-term expected returns by the above distillation, its potential in representation capability is far from being developed. To this end, the online encoder is learned to capture other basic characters including environment dynamics as well as visual perception by using additional auxiliary tasks. As illustrated in Fig. 3, the predicted states \\(\\mathbf{s}_t^o\\) are conditioned on three auxiliary heads: a reward predictive head \\(\\mathcal{R}\\), a state predictive head \\(\\mathcal{P}\\) and an observation decoder head \\(\\mathcal{G}\\). Taking the current states \\(\\mathbf{s}_t^o\\) and actions \\(a_{t}\\) as inputs, the \\(\\mathcal{R}\\) and \\(\\mathcal{P}\\) aim to predict the reward \\(\\mathcal{R}(\\mathbf{s}_t^o,a_t)\\) and next state \\(\\mathcal{P}(\\mathbf{s}_t^o,a_t)\\), and thus learn a latent state space of which dynamic is consistent with the environment. Dynamic predictive loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.433, + 0.891, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {d y n a m i c}} = \\mathcal {L} _ {\\text {r e w a r d}} + \\mathcal {L} _ {\\text {s t a t e}} \\\\ = D _ {r} \\left(\\mathcal {R} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right), r _ {t}\\right) + D _ {s} \\left(\\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right), \\mathbf {s} _ {t + 1} ^ {o}\\right)) \\\\ = \\left(\\mathcal {R} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right) - r _ {t}\\right) ^ {2} + \\left\\| \\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right) - \\mathbf {s} _ {t + 1} ^ {o}\\right) \\| ^ {2}. \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.548 + ], + "angle": 0, + "content": "To guarantee temporal capacity, the decoder \\(\\mathcal{G}\\) is trained to predict the next observation with the predicted next state \\(\\hat{s}_{t + 1}^o = \\mathcal{P}(\\mathbf{s}_t^o,a_t)\\). The decoding loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.559, + 0.891, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {d e c} = D _ {o} \\left(\\hat {\\mathbf {o}} _ {t + 1}, \\mathbf {o} _ {t + 1}\\right) \\\\ = D _ {o} \\left(\\mathcal {G} \\left(\\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, \\mathbf {a} _ {t}\\right)\\right), \\mathbf {o} _ {t + 1}\\right) \\quad H W \\tag {13} \\\\ = \\frac {1}{H W} \\sum_ {h, w = 1} ^ {H, W} \\left(\\mathcal {G} \\left(\\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right)\\right) ^ {[ h, w ]} - \\mathbf {o} _ {t + 1} ^ {[ h, w ]}\\right) ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.681 + ], + "angle": 0, + "content": "Then the overall loss of the topology distilled transferring is formulated by" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.692, + 0.891, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e n c o d e r}} = \\mathcal {L} _ {\\text {t r a n s}} + \\mathcal {L} _ {\\text {d y n a m i c}} + \\mathcal {L} _ {\\text {o b s}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.749 + ], + "angle": 0, + "content": "where the online main encoder and auxiliary heads are jointly learned with the adaptor frozen." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.771, + 0.773 + ], + "angle": 0, + "content": "3.4. Greedy Bootstrapping for AIC" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In most deep reinforcement learning approaches, the representation module and policy module are trained in a simultaneous manner. That is, while the policy model is trained from the reinforcement signals the encoder is also simultaneously changing. This phenomenon will disturb policy improvement due to the drifting of state representation. To tackle this problem, this section presents a greedy bootstrapping framework for AIC, which consists of two" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "211" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.089, + 0.096, + 0.386, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Greedy Bootstrapping for AIC" + }, + { + "type": "algorithm", + "bbox": [ + 0.082, + 0.113, + 0.437, + 0.345 + ], + "angle": 0, + "content": "1 Initialize the replay buffer \\(\\mathcal{B}\\) with random episodes. \n2 while not converged do \n3 // Representation learning \n4 Build data-loader with the replay buffer \\(\\mathcal{B}\\) \n5 for update step \\(c = 1\\dots C\\) do \n6 Draw next batch of transitions \\(\\{\\left(\\mathbf{o}_t,a_t,r_t,\\mathbf{o}_{t + 1}\\right)\\}\\) \n7 //Momentum Adaptor Tuning \n8 Update momentum encoder by moving average. \n9 Update adaptor by optimizing Eq. 4. \n10 //Topology Distilled Transferring \n11 Update online encoder by optimizing Eq. 14. \n12 end \n13 // Policy learning (encoder frozen) \n14 for update step \\(c_{p} = 1\\dots S\\) do \n15 Draw transition sequences \\(\\{\\left(\\mathbf{o}_t,a_t,r_t,\\mathbf{o}_{t + 1}\\right)\\} \\sim \\mathcal{B}\\) \n16 Update critic and actor parameters. \n17 end \n18 Add new experiences to the replay buffer \\(\\mathcal{B}\\) \n19 end" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.456 + ], + "angle": 0, + "content": "alternative training stages for the representation and policy, respectively. In each representation stage, the encoder keeps training until convergence, and so does the policy in the other stage. The framework is described in detail in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.577 + ], + "angle": 0, + "content": "Greedy representation learning. Over the latest collected replay buffer, the encoder is trained by iterating all the images in each epoch. Different from sampling a random batch in each iteration like most other methods, our ergodic sampling strategy brings stable training without ignorance of any crucial samples. In each stage of representation learning, the encoder is trained for several epochs until it is converged." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.469, + 0.7 + ], + "angle": 0, + "content": "Greedy policy learning. With the main encoder trained sufficiently, the policy is learned continuously by freezing the encoder and conditioned on its output state. Beneficial from the stable and discriminative latent state, the critic and actor are trained efficiently from the reinforcement signal by SAC algorithms and tend to produce better behaviors. The policy is learned continuously for enough steps and then used to interact with the environment for new trajectories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.715, + 0.202, + 0.732 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.27, + 0.758 + ], + "angle": 0, + "content": "4.1. Environments Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.869 + ], + "angle": 0, + "content": "**Benchmarks.** We evaluate our method on two reinforcement learning environments with complex and realistic observations, i.e., CARLA [5] and Vizdoom [22], with continuous and discrete action space, respectively. Both tasks are developed in 3D physical scenes with a complex state space, which calls for deeper visual encoders with stronger representation capability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Implementation details. We adopt ResNet-18 as the backbone of the visual encoder. Before the last pooling layer, the" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.097, + 0.875, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.282, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Figure 5. Comparison with other baseline approaches on the CARLA environment." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.328, + 0.872, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.515, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Figure 6. Ablation Study of our method using ResNet-18 as the visual encoder on the CARLA environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.627 + ], + "angle": 0, + "content": "feature map is flattened and then projected into 1024-dim state features with a linear layer and layer norm operation. The momentum factor \\(\\tau\\) is set to 0.05. We use Adam to optimize the parameters of all modules in our approach." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.636, + 0.87, + 0.652 + ], + "angle": 0, + "content": "4.2. CARLA Autonomous Driving Environment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Experimental settings. In the autonomous driving task, we construct a highway driving scenario with 20 other vehicles of different models using the CARLA simulator. We use five cameras on the roof of the ego-vehicle, each with a 60-degree view and obtaining images of \\(84 \\times 84\\) pixels. We concatenate them together for observation of \\(84 \\times 420\\) pixels. We stack the last three frames as the inputs. Following DBC [40], the reward is formulated as \\(r_t = v_{ego}^T \\hat{u}_h \\cdot \\Delta t - \\xi_1 \\cdot \\mathbb{I} - \\xi_2 \\cdot |steer|\\), where \\(v_{ego}\\) is the velocity vector of the ego-vehicle, projected onto the highway's unit vector \\(\\hat{u}_h\\), and multiplied by time discretization \\(\\Delta t = 0.05\\) to measure highway progression in meters. Impulse \\(\\mathbb{I} \\in \\mathbb{R}^+\\) is caused by collisions, and a steering penalty steer \\(\\in [-1,1]\\) facilitates lane-keeping. The hyperparameters \\(\\xi_1\\) and \\(\\xi_2\\) are set to \\(10^{-4}\\) and 1, respectively." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.887, + 0.891, + 0.902 + ], + "angle": 0, + "content": "Comparison with other methods. We compare our ap" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "212" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.1, + 0.445, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.282, + 0.47, + 0.31 + ], + "angle": 0, + "content": "Figure 7. Comparison of other representation-decoupled methods on the CARLA environment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.47, + 0.565 + ], + "angle": 0, + "content": "proach with seven baselines, including joint representation learning methods, i.e., CURL [27], SVEA [16], Deep-MDP [7], DBC [40]; representation decoupled methods, i.e., DreamerV2 [13], Iso-Dream [32], and the baseline SAC [10]. As shown in Fig. 5, our AIC outperforms all baselines by large margins. Compared with joint representation learning, the superior performance can be attributed to that our method stably learns a more discriminative state abstraction with deeper architectures, while avoiding the oscillating self-overfitting. Representation decoupled methods train encoders only on auxiliary tasks without TD losses and thus usually require more time to converge. However, AIC effectively learns the long-term expected returns by interaction with the light-optimizable momentum adaptor encoder, resulting in faster convergence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Ablation Study. We conduct ablation studies to validate the effectiveness of the proposed topology distilled transferring and the greedy bootstrapping strategy. All the experiments train the ResNet-18 as the encoder. Fig. 6 shows that SAC fails to learn an effective policy with the heavy-optimizable encoder since the state features are deteriorated by the serious oscillating self-overfitting. By applying the dynamic loss and decoder loss, the deterioration is slightly alleviated, since the encoder is encouraged to learn the inherent information of environment dynamics by the targets which are more stable than TD. However, disturbed by the oscillating gradients of the TD loss, the encoder is still hard to converge. Compared to it, the greedy bootstrapping strategy achieves significant improvement, denoted as AIC \\( w \\backslash o L_{trans} \\), because it interrupts the backward propagation of the unstable gradient and simultaneously it trains the encoder and policy stably and sufficiently. AIC distinctly outperforms the AIC \\( w \\backslash o L_{trans} \\) due to the topology distilled transferring, which additionally incorporates the long-term expected return information into the encoder, making it more adaptive for the downstream controlling." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Comparison with other decoupling methods. To further" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.1, + 0.868, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.282, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Figure 8. Comparison with different architectures of the encoder on the CARLA environment." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.333, + 0.868, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.516, + 0.89, + 0.543 + ], + "angle": 0, + "content": "Figure 9. Comparison of other stabilizing methods using ResNet-18 on the CARLA environment." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.569, + 0.892, + 0.901 + ], + "angle": 0, + "content": "evaluate the greedy bootstrapping framework, we compare it with other two typical methods that decouple representation learning from RL. Initial decoupling first collects a set of transitions using a random policy and then pretrains the encoder with auxiliary tasks as in Eq. 12 and 13. Afterward, the policy is trained by SAC with the encoder frozen. As shown in Fig. 7, the initial decoupling method performs high-variance results because the performance significantly depends on whether the distribution of the collected pretraining data is similar to those from a learned policy. Stepping decoupling learns the encoder and the policy step by step, respectively. The encoder is trained by transitions from the same data distribution as the policy. However, since the encoder keeps updating during policy learning, the latent state space suffers from unstable drifting, making it difficult to learn policy from the volatile state abstraction. The greedy bootstrapping strategy surpasses both two approaches because it guarantees a determined latent state space that is conducive to the inductive learning of policy. Moreover, it learns the encoder by traversing all the transitions in the replay buffer, prevented from ignorance of some critical images or overfitting to parts of data." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "213" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.2, + 0.088, + 0.768, + 0.102 + ], + "angle": 0, + "content": "Table 1. Comparison with other methods using ResNet-18 on Medikit of Vizdoom environment." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.102, + 0.847, + 0.227 + ], + "angle": 0, + "content": "
-DQN (Baseline)DrQSVEACURLDeepMDPAIC (Ours)
Episode Returns (mean/std)888±5841361 ± 1004435±185408±2031026±8071686±1283
Min returns252252792428456
Max returns244442001008113637104800
Survival Time (mean/std)64 ±2685±4345±844±972±3796±53
Min Time262627233926
Max Time1352106874191210
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.473, + 0.597 + ], + "angle": 0, + "content": "Improvements on different encoders. We evaluate our approach with different architectures as the visual encoder, including a light CNN of 4-layer convolution (Conv-4), and three heavyweight models, i.e., VGG-11, ResNet-18, and ResNet34. For each architecture, we conduct three experiments using SAC, SAC with auxiliary tasks (dynamic loss and decoder loss as in Eq. 12 and 13), and our AIC approach. The comparison is illustrated in Fig. 8. Firstly, all heavyweight models fail to learn effective policies using SAC, where only TD loss is used. The results indicate the oscillating self-overfitting is consistent for heavy models. Secondly, auxiliary tasks improve the performance of SAC with heavyweight encoders, but they are still inferior to the light encoder like Conv-4. The oscillating self-overfitting seriously prevents the encoder from sufficient learning. Finally, AIC consistently improves the performance with all heavy encoders by large margins, as well as the light encoder of Conv-4. It shows a consistent rank of performance that these architectures have achieved in supervised learning tasks. It demonstrates that our method makes full use of the representation capability of heavyweight visual encoders through interactive cooperation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Comparison with stabilizing methods. To evaluate AIC for stabilizing the training of heavy-optimizable encoders, we compare it with SVEA and DrQ using ResNet-18 as the backbone. As shown in Fig. 9, both two methods perform not well, since the increase of parameters significantly exacerbates the challenge of stabilizing. However, AIC exceeds them by large margins, indicating its capability to stably learn heavy models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Qualitative analysis. In order to evaluate the effectiveness of our approach in alleviating the oscillating self-overfitting, we visualize the updating trajectories of the encoder parameters over a set of training steps by TSNE. We use ResNet-18 as the encoder backbone and choose the weights of the last convolution layer for visualization. The comparison between DeepMDP with the TD loss and AIC is shown in Fig. 10, where each point represents the parameters in a step and the color changes from deep to light over the training steps. As we can see, the parameters with the TD loss of the DeepMDP agent suffer from significant oscillation during training. Different from that, the parameters of AIC show a dis" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.265, + 0.692, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.265, + 0.886, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.431, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Figure 10. Parameter updating trajectories of the ResNet-18 encoder over training steps." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.486, + 0.892, + 0.547 + ], + "angle": 0, + "content": "tinct optimizing direction and only have slight oscillation in a short time. The qualitative results demonstrate AIC effectively alleviates the oscillating self-overfitting for the heavy-optimizable encoder in visual reinforcement learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.776, + 0.573 + ], + "angle": 0, + "content": "4.3. Vizdoom Medikit Environment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Experimental settings. Vizdoom is based on the first-person perspective video game in a semi-realistic 3D world. We use the difficult medikit collecting scenario, where the agent is spawned in a random spot of a maze with an acid surface, which slowly, but constantly, takes away the agent's life. The agent needs to collect medikits and avoid blue vials with poison, both of which appear in random places during the episode. In each step, the agent is allowed to move (forward/backward), or turn (left/right). The reward is formulated as \\( r_t = 1 - \\xi_3 \\cdot E_m - \\xi_4 \\cdot E_v - \\xi_5 \\cdot E_d \\), where \\( E_m \\), \\( E_v \\), \\( E_d \\) denote the events of getting a medikit, a vial and to death, respectively. \\( \\xi_3 = \\xi_4 = \\xi_5 = 100 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Superior performance with heavy encoders. To evaluate our approach, especially with deep deep encoders, we compare it with five methods. All of them use ResNet-18 as backbones. Since the Vizdoom is of discrete action space, we adopt DQN as the baseline. As shown in Table 1, AIC performs superiorly to the compared methods. CURL and SVEA perform poorly, since the observations of such a fine-grained collecting task of first-person perspective are sensitive to random cropping of CURL. And the random conv of" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "214" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "SVEA is not applicable for the low-contrast images in this scenario. However, AIC is able to effectively train heavy models under different conditions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.149, + 0.196, + 0.166 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.175, + 0.473, + 0.356 + ], + "angle": 0, + "content": "In this paper, we investigate the phenomenon of oscillating self-overfitting which results in the deterioration of RL when training heavy-optimizable encoders. To alleviate this problem, we propose a novel asymmetric interactive cooperation approach. Through interaction between a main heavy-optimizable encoder and a supportive light-optimizable encoder, both advantages of them are integrated including the highly discriminative capability and the training stability. We further present a greedy bootstrapping framework for further stability. Experimental results of CARLA and Vizdoom show that AIC achieves competitive performance over complex and realistic environments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.369, + 0.245, + 0.387 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.47, + 0.499 + ], + "angle": 0, + "content": "This work is partially supported by grants from the Key-Area Research and Development Program of Guangdong Province under contact No. 2020B0101380001 and grants from the National Natural Science Foundation of China under contract No. 62027804, No. 61825101, No. 62088102 and No. 62202010. The computing resources of Pengcheng Cloudbrain are used in this research." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.512, + 0.174, + 0.527 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[1] Homanga Bharadhwaj, Mohammad Babaeizadeh, Dumitru Erhan, and Sergey Levine. Information prioritization through empowerment in visual model-based rl. arXiv preprint arXiv:2204.08585, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[2] Pablo Samuel Castro. Scalable methods for computing state similarity in deterministic markov decision processes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10069-10076, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.65, + 0.468, + 0.69 + ], + "angle": 0, + "content": "[3] Edoardo Cetin, Philip J Ball, Steve Roberts, and Oya Celiktutan. Stabilizing off-policy deep reinforcement learning from pixels. arXiv preprint arXiv:2207.00986, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.691, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.761, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[6] Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 2786-2793. IEEE, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[7] Carles Gelada, Saurabh Kumar, Jacob Buckman, Ofir Nachum, and Marc G Bellemare. Deepmdp: Learning" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.537, + 0.469, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "continuous latent space models for representation learning. In International Conference on Machine Learning, pages 2170-2179. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[8] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.206, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[9] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[10] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[11] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[12] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International conference on machine learning, pages 2555-2565. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[13] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[14] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[15] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. arXiv preprint arXiv:2007.04309, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[16] Nicklas Hansen, Hao Su, and Xiaolong Wang. Stabilizing deep q-learning with convnets and vision transformers under data augmentation. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[17] Nicklas Hansen and Xiaolong Wang. Generalization in reinforcement learning by soft data augmentation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13611-13617. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[19] Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[20] Michael Janner, Qiyang Li, and Sergey Levine. Offline reinforcement learning as one big sequence modeling problem. Advances in neural information processing systems, 34:1273-1286, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[21] Lukasz Kaiser, Mohammad Babaeizadeh, Piotr Milos, Blazej Osinski, Roy H Campbell, Konrad Czechowski, Dumitru Erhan, Chelsea Finn, Piotr Kozakowski, Sergey Levine, et al. Model-based reinforcement learning for atari. arXiv preprint arXiv:1903.00374, 2019." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "215" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[22] Michal Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Vizdoom: A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE conference on computational intelligence and games (CIG), pages 1-8. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[23] Ilya Kostrikov, Denis Yarats, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. arXiv preprint arXiv:2004.13649, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.247 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[25] Aviral Kumar, Rishabh Agarwal, Dibya Ghosh, and Sergey Levine. Implicit under-parameterization inhibits data-efficient deep reinforcement learning. arXiv preprint arXiv:2010.14498, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[26] Misha Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. Advances in neural information processing systems, 33:19884-19895, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[27] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In International Conference on Machine Learning, pages 5639-5650. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.47, + 0.497 + ], + "angle": 0, + "content": "[28] Alex X. Lee, Anusha Nagabandi, Pieter Abbeel, and Sergey Levine. Stochastic latent actor-critic: Deep reinforcement learning with a latent variable model. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 741-752. Curran Associates, Inc., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[29] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. The Journal of Machine Learning Research, 17(1):1334-1373, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.47, + 0.621 + ], + "angle": 0, + "content": "[30] Masashi Okada and Tadahiro Taniguchi. Dreaming: Model-based reinforcement learning by latent imagination without reconstruction. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4209-4215. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.663 + ], + "angle": 0, + "content": "[31] Kei Ota, Devesh K Jha, and Asako Kanezaki. Training larger networks for deep reinforcement learning. arXiv preprint arXiv:2102.07920, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.72 + ], + "angle": 0, + "content": "[32] Minting Pan, Xiangming Zhu, Yunbo Wang, and Xiaokang Yang. Iso-dream: Isolating and leveraging noncontrollable visual dynamics in world models. In Advances in Neural Information Processing Systems." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.47, + 0.776 + ], + "angle": 0, + "content": "[33] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[34] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International Conference on Machine Learning, pages 8583-8592. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[35] Evan Shelhamer, Parsa Mahmoudieh, Max Argus, and Trevor Darrell. Loss is its own reward: Self-supervision for reinforcement learning. arXiv preprint arXiv:1612.07307, 2016." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.174 + ], + "angle": 0, + "content": "[36] Oriol Vinyals, Igor Babuschkin, Wojciech M Czarnecki, Michael Mathieu, Andrew Dudzik, Junyoung Chung, David H Choi, Richard Powell, Timo Ewalds, Petko Georgiev, et al. Grandmaster level in starcraft ii using multiagent reinforcement learning. Nature, 575(7782):350-354, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.176, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[37] Manuel Watter, Jost Springenberg, Joschka Boedecker, and Martin Riedmiller. Embed to control: A locally linear latent dynamics model for control from raw images. Advances in neural information processing systems, 28, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[38] Denis Yarats, Amy Zhang, Ilya Kostrikov, Brandon Amos, Joelle Pineau, and Rob Fergus. Improving sample efficiency in model-free reinforcement learning from images. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10674-10681, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.893, + 0.358 + ], + "angle": 0, + "content": "[39] Tao Yu, Zhizheng Zhang, Cuiling Lan, Yan Lu, and Zhibo Chen. Mask-based latent reconstruction for reinforcement learning. In Advances in Neural Information Processing Systems." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.36, + 0.893, + 0.416 + ], + "angle": 0, + "content": "[40] Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant representations for reinforcement learning without reconstruction. arXiv preprint arXiv:2006.10742, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "216" + } + ] +] \ No newline at end of file diff --git a/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/3ce1c749-00ff-4ed7-ac11-5982848d2220_origin.pdf b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/3ce1c749-00ff-4ed7-ac11-5982848d2220_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd1d1d9210d84eb0c816a38679277e44c8ce923c --- /dev/null +++ b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/3ce1c749-00ff-4ed7-ac11-5982848d2220_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0e85a7b6c769fde2830513d4bcbc3f9274b925378878e771b641fce992da085 +size 4227108 diff --git a/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/full.md b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75c271ef9e15b79bc6c4f477fa990b92a0dd82eb --- /dev/null +++ b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/full.md @@ -0,0 +1,325 @@ +# Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation + +Yunpeng Zhai $^{1}$ , Peixi Peng $^{2,3,*}$ , Yifan Zhao $^{1}$ , Yangru Huang $^{1}$ , Yonghong Tian $^{1,2,3,*}$ $^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University, Beijing, China + $^{2}$ School of Electronic and Computer Engineering, Peking University Shenzhen Graduate School, Shenzhen, China + $^{3}$ Peng Cheng Laboratory, Shenzhen, China + +{ypzhai, pxpeng, zhaoyf, yhtian}@pku.edu.cn, yrhuang@stu.pku.edu.cn + +# Abstract + +Vision-based reinforcement learning (RL) depends on discriminative representation encoders to abstract the observation states. Despite the great success of increasing CNN parameters for many supervised computer vision tasks, reinforcement learning with temporal-difference (TD) losses cannot benefit from it in most complex environments. In this paper, we analyze that the training instability arises from the oscillating self-overfitting of the heavy-optimizable encoder. We argue that serious oscillation will occur to the parameters when enforced to fit the sensitive TD targets, causing uncertain drifting of the latent state space and thus transmitting these perturbations to the policy learning. To alleviate this phenomenon, we propose a novel asymmetric interactive cooperation approach with the interaction between a heavy-optimizable encoder and a supportive light-optimizable encoder, in which both their advantages are integrated including the highly discriminative capability as well as the training stability. We also present a greedy bootstrapping optimization to isolate the visual perturbations from policy learning, where representation and policy are trained sufficiently by turns. Finally, we demonstrate the effectiveness of our method in utilizing larger visual models by first-person highway driving task CARLA and Vizdoom environments. + +# 1. Introduction + +Learning complex control from high-dimensional observations such as images is significant for many real-world applications [37, 28]. It puts forward higher requirements for the representation capability of visual encoder models, + +![](images/3416461f6b614d61f15d3450008213d0cc1ecd711d5c084989a3cbdb9549c191.jpg) +a) Returns during training + +![](images/4415384ff751e746add5f4526c56413e1ac2d8534a4703b9498b9597d93b8a1a.jpg) +b) Parameter oscillation +Figure 1. Comparison of DeepMDP agents on CARLA with light- and heavy-weight visual encoders in episode returns (a) and parameter oscillation $\left(\sum_{t}^{T}\|\nabla\theta_{t}\|\right)/\|\sum_{t}^{T}\nabla\theta_{t}\|$ (b). Compared with 4-layer CNN, larger models like ResNet-18 do not improve the RL performance as in supervised learning. The deterioration results from the oscillating self-overfitting of the larger model, which occurs in neither supervised learning nor RL with a lightweight encoder. + +especially in some complex scenes such as self-driving [5] and robot controlling [29]. The last decade has witnessed impressive progress in computer vision by training large-scale networks [18] as they increase the search space of possible solutions. However, such parameter increment of visual models cannot directly benefit the reinforcement learning, and even leads to deterioration of training. For example, as shown in Fig. 1(a), we illustrated the training curves of DeepMDP [7] agents that use different CNN networks as visual encoders on the self-driving environment CARLA [5]. The result shows that using a larger model, e.g., ResNet-18 [18], leads to unstable training and achieves distinctly lower returns than the lighter model with only four convolutional layers. + +To investigate this phenomenon, we quantified the oscillation of the network parameters by introducing an in- + +indicator calculated by the ratio of accumulation of modulus length of the gradient to the module lengths of cumulative gradient within $T$ training steps, $(\sum_{t}^{T}\|\nabla\theta_{t}\|)/\|\sum_{t}^{T}\nabla\theta_{t}\|$ , where $\nabla$ is gradient operator and $\theta_{t}$ is parameters of the last convolution layer in the $t$ -th step. A higher indicator means worse oscillation and instability of parameters during training. As shown in Fig. 1(b), we compared the oscillation in three experiments including i) training ResNet-18 on RL, ii) training 4-layer CNN on RL, and iii) training ResNet-18 on supervised learning (SL) with CIFAR-100 [24]. ResNet-18 suffers much more serious oscillation when training on RL, resulting in deteriorated performance. However, such oscillation occurs neither on SL with ResNet-18 nor on RL with the lighter 4-layer CNN. We name this phenomenon as oscillating self-overfitting, which particularly results from a pathological concurrence of the overfitting capability of large models and the sensitive learning targets of the temporal-difference (TD) loss [25]. Specifically, when the heavyweight parameters are enforced to fit the sensitive TD targets which are partially generated by themselves with a bootstrapping formulation, contradictory gradients will propagate back to oscillate the parameters. This phenomenon results in uncertain drifting of the state space and transmits these perturbations to policy learning. Therefore, stabilizing visual encoders with amounts of parameters from TD losses for performance gain in RL is still an open challenge. + +From the perspective of the relation between the encoder and the TD objective, existing representation learning for RL can be categorized into two groups, as illustrated in Fig. 2. One approach jointly learns representation with policy by the TD loss [16, 7, 23]. It efficiently learns the long-term expected returns with light encoders while is incompatible with heavy-optimizable encoders due to the oscillating self-overfitting. Another group of approaches decouples representation learning from RL to avoid instability, where the encoder is learned only by auxiliary dynamic predictive losses [11, 32]. However, it is inaccessible to the expected returns from the bootstrapping objective in RL. In this paper, we propose a novel asymmetric interactive cooperation for representation learning in RL. To take both the advantages of representation capability and the stability for TD targets, it separately trains a main heavy-optimizable encoder and a supportive light-optimizable encoder by auxiliary tasks and TD losses, respectively. And the asymmetric interaction is simultaneously conducted between them to effectively exchange their knowledge from each other, where the heavy one transfer the representation capability by parameter momentum and the light one transfer the long-term expected returns by topological distillation. Hence, the heavy encoder is equipped with the capability of latent state abstraction without oscillating by the TD objective. Moreover, we present a greedy bootstrapping optimization for + +![](images/64212d11eddc8d9940c85d488b7263a97b323a3bd425821c5f64fa713b2956a8.jpg) + +![](images/5afeef562035408a31e9ad91cf5676a319b1071ef47342e6208ceef07017fcd7.jpg) + +![](images/3cf6e94cf456eed5a8b19103076a63cf391a455c2e6f6b5673719ddfd659339e.jpg) +Figure 2. Illustrations of different learning paradigms of visual encoders. a) Joint Learning trains the visual encoder end-to-end with reinforcement loss. b) Decoupled learning trains the encoder only by auxiliary tasks. c) The proposed AIC trains a supportive light-optimizable encoder with the expected returns information from TD targets and transfers it to the main visual encoder. + +further stability of training, where representation and policy are trained sufficiently by turns. + +The main contributions of this paper can be summarized in three aspects. First, it investigates the phenomenon of oscillating self-overfitting that leads to deterioration in RL with heavy-optimizable encoders, and proposes a novel asymmetric interactive cooperation to alleviate it. Second, it presents a topological distillation between latent state spaces to learn state abstraction by interactive cooperation without any explicit labels. Third, it presents a greedy bootstrapping optimization for further stability of training. Experiments demonstrate a significant performance gain over the complex and realistic environments of CARLA and Vizdoom. + +# 2. Related Work + +Jointly Representation Learning with RL. Since learning control directly from high-dimension visual observations is hard to converge, many approaches learn the encoder jointly with RL and elaborate auxiliary tasks [39, 38, 26, 6, 21, 34], as shown in Fig. 2(a). The first line of methods utilizes self-supervision to learn the visual invariances [17, 16, 15, 35]. CURL [27] introduced a contrastive loss to learn discriminative features from raw pixels. The second line learns to abstract the state by preserving the property of Markov decision processes (MDPs) [12, 19, 2]. DeepMDP [7] learns the latent space with the same dynamic transition as the environment. DBC [40] learns invariant representations with + +![](images/3c9056d7dbf55b31eb1475a3f0a4f1e33b0fa55b0cbcbd08cc4a1f9992bd3112.jpg) +Figure 3. Flowchart of the proposed asymmetric interactive cooperation. The representation learning consists of two components including momentum adaptor tuning and topological distilled transferring. In the momentum adaptor tuning, the adaptor is optimized by RL to learn the information of expected returns while the momentum encoder is updated by moving average from the online encoder. In topology distilled transferring, the online encoder is trained by three auxiliary tasks while transferring the information of expected returns by topology distillation from the momentum adaptor encoder. The policy is learned by freezing the online encoder. + +the bisimulation distance. The third line of methods designs regularization to stabilize the training process [23]. DrQ [23] and SVEA [16] propose data regularization using image augmentation of random shift or random convolution. A-LIX [3] adaptively regularizes the convolutional features to prevent overfitting while it doesn't study the heavy visual models. The common drawback of these methods is the difficulty for the heavy-optimizable encoders due to oscillating self-overfitting caused by TD losses, resulting in more serious instability. Other recent works have concerned the stability of large models in RL [25, 31], but they focus on a different component of policy learning instead of representation. We also note recent work RL with Transformer while they only contain slight parameters [4, 20] or rely on supervised imitation learning [36]. + +Decoupled Representation Learning from RL. Recent methods proposed to decouple representation learning from TD losses for efficient training [30, 31], in Fig. 2(b). A typical approach is the world model [9, 8]. It first pretrains an encoder using Variational Autoencoder [33] by the collected rollouts from a random policy and then trains policy with the encoder frozen. However, the encoder does not perform well since the pretraining data is of different distribution from those in policy learning. Recent world model + +class methods [14, 30, 32, 13, 1], such as Dreamer [11], turn to learn the encoder and the policy alternatively step-by-step, in which the encoder is updated by the newly collected data. Theoretically, these approaches can train any large model stably. But there still exists a bottleneck, that is, the encoder is isolated from the long-term expected returns hidden in the TD losses, which is significant for the state abstraction with respect to the downstream controls. In our work, we tackle this problem by interacting with a supportive light-optimizable encoder, which learns from TD losses stably and transfers its knowledge to the main encoder, as shown in Fig. 2(c). + +# 3. The Proposed Approach + +# 3.1. Overview + +Asymmetric interactive cooperation alternatively trains a light-optimizable supportive encoder with TD losses to learn the long-term expected returns without oscillating self-overfitting, and a heavy-optimizable main encoder with auxiliary tasks to capture stronger representation capability. Simultaneously, two kinds of interaction are conducted between them to exchange their knowledge including: + +- Light $\leftarrow$ Heavy: parameter momentum to absorb the + +representation capability, as in Sec. 3.2. + +- Light $\rightarrow$ Heavy: topological distillation to learn the long-term expected returns, as in Sec. 3.3. + +Beyond them, it presents a greed bootstrapping framework for AIC to alleviate the drifting of the state space for further stability in Sec. 3.4. The flowchart of AIC is illustrated in Fig. 3. + +# 3.2. Momentum Adaptor Tuning + +Attaching adaptor to momentum. To efficiently learn the long-term expected returns, the light-optimizable encoder should satisfy the following two conditions: 1) It only contains a few learnable parameters so that it converges without oscillating self-overfitting. 2) It can utilize the middle-level representation of the heavy-optimizable encoder rather than relying only on the raw observation input. To fulfill the above conditions, we implement the light-optimizable encoder by applying learnable CNN adaptors $\beta$ to the momentum network $\theta^m$ of the online heavy encoder $\theta^o$ . As illustrated in Fig. 3, each $3\times 3$ convolutional layer in the momentum encoder is connected with a $1\times 1$ convolutional adaptor module in a parallel manner. Thus the result of each layer is formulated by, + +$$ +x _ {l + 1} = \rho \left(x _ {l}; \theta_ {l} ^ {m}, \beta_ {l}\right) = \theta_ {l} ^ {m} * x _ {l} + \beta_ {l} * x _ {l}, \tag {1} +$$ + +where $x_{l}$ is the input tensor of the $l$ -th layer and $*$ denotes the convolution operator. $\theta_{l}^{m}$ and $\beta_{l}$ are the weights of the momentum model and the adaptor, respectively. We name this light-optimizable encoder as momentum adaptor encoder, referring to $f^{m}$ . It interacts with the online heavy encoder $f^{o}$ by parameter momentum to utilize its representation. In each iteration step, the momentum model is updated by moving average of the online encoder $f^{o}$ , + +$$ +\theta^ {m} \leftarrow \tau \theta^ {m} + (1 - \tau) \theta^ {o}, \tag {2} +$$ + +where $\theta^o$ is the parameters of online encoder and $\tau \in [0,1]$ is the momentum factor. + +Adaptor Tuning. The adaptor is updated with RL by maximizing the expected return, + +$$ +\beta^ {*} = \arg \max _ {\beta} \mathbb {E} \left[ \sum_ {t} ^ {\infty} \gamma^ {t} R \left(s _ {t}, a _ {t}, s _ {t + 1}\right) \right], \tag {3} +$$ + +where $\gamma$ is the discount factor and $R$ is the reward. $s_t$ and $a_{t}$ are the state and action in time step $t$ . Given a batch of transitions $\{(\mathbf{o}_t,a_t,r_t,\mathbf{o}_{t + 1})\}$ from the replay buffer $\mathcal{B}$ , $\mathbf{o}_t\in \mathbb{R}^{C\times H\times W}$ is the observation stacking multiple frames. $H$ , $W$ are the height and width of the image and $C$ is the channel number. We first projected the observation from images to a latent state $\mathbf{s}^m = f^m (\mathbf{o}_t;\theta^m,\beta)$ with the momentum adaptor encoder $f^{m}$ . To tune the adaptor with the reinforcement learning, we optimize the predicted state by + +![](images/4754f2ecbb5ab2b0ca074510761630ccf7b5dc160c5e7ec2524856184dcaa515.jpg) +Figure 4. Illustration of the distance between state topology in different latent spaces. $\psi^m (\cdot)$ and $\psi^o (\cdot)$ denote the similarity function of states. + +conditioning it on a learnable critic head $h_c$ and an actor head $h_a$ for SAC [10] algorithms. Both two heads are implemented by 3-layer MLPs. Then the critic takes state $\mathbf{s}_t^m$ and action $a_t$ as inputs and learns to predict $Q(\mathbf{s}_t^m, a_t)$ with the critic loss, + +$$ +\mathcal {L} _ {\text {c r i t i c}} = \left(Q \left(\mathbf {s} _ {t}, a _ {t}\right) - y \left(r _ {t}, \mathbf {s} _ {t + 1}\right)\right) ^ {2}, \tag {4} +$$ + +where the TD target is given by + +$$ +y \left(r _ {t}, \mathbf {s} _ {t + 1}\right) = r _ {t} + \gamma \left(\bar {Q} \left(\mathbf {s} _ {t + 1}, \hat {a} _ {t + 1}\right) - \alpha_ {r} \log \pi \left(\hat {a} _ {t + 1} \mid \mathbf {s} _ {t + 1}\right)\right). \tag {5} +$$ + +$\bar{Q}$ is the target critic by moving average, and $\pi (\hat{a}_{t + 1}|\mathbf{s}_{t + 1})$ is the policy distribution predicted by the actor head. $\alpha_{r}$ is the temperature parameter in SAC. $\hat{a}_{t + 1}\sim \pi (\cdot |\mathbf{s}_{t + 1})$ . The actor learns to generate the action that maximizes the Qvalue predicted by the critic, + +$$ +\mathcal {L} _ {\text {a c t o r}} = \alpha_ {r} \log \pi (\hat {a} _ {t} | \mathbf {s} _ {t}) - Q (\mathbf {s} _ {t}, \hat {a} _ {t}). \tag {6} +$$ + +By jointly training the adaptor $\beta$ with the critic $h_c$ and actor $h_a$ , the information of long-term expected returns will be efficiently learned to the adaptor, resulting in a better state abstraction for the control task. + +# 3.3. Topology Distilled Transferring + +Topology of abstracted states. Since the information of expected returns is stably captured by the momentum adaptor $f^m$ , the core issue is how to transfer that knowledge into the online heavy-optimizable main encoder $f^o$ , making its representation more adaptive to policy learning. To this end, we introduce a teacher-student distillation manner where the momentum adaptor encoder $f^m(\cdot; \theta^m, \beta)$ and the online main encoder $f^o(\cdot; \theta^o)$ are considered as the teacher and student, respectively. Learning state abstraction for control is to project the observed states into a latent space, where the adjacency among states is consistent with the transitions in the real environment. For instance, the states near the same time-step in the real environment should also be adjacent in the feature space. Hence, we transfer the capability of state abstraction by distillation of the adjacency, i.e., + +topological relation in the latent state space. To this end, we first define topological relation $\mathcal{T}(S)$ in the state space $S$ as the set of similarity between every state pairs $(\mathbf{s}_i,\mathbf{s}_j)$ in the data, + +$$ +\mathcal {T} (S) = \left\{\psi \left(\mathbf {s} _ {i}, \mathbf {s} _ {j}\right) \mid \left(\mathbf {s} _ {i}, \mathbf {s} _ {j}\right) \in S \right\} \tag {7} +$$ + +where $\psi (\mathbf{s}_i,\mathbf{s}_j)$ denotes the similarity function of states. This set of similarities reveals the structural information of the state space, thus preserving the representation capability of the states. Our objective is to minimize the distance between the teacher's topological relation $\mathcal{T}(S^m)$ and the student's $\mathcal{T}(S^o)$ , as illustrated in Fig. 4, + +$$ +\begin{array}{l} \theta^ {m *} = \arg \min _ {\theta^ {m}} D _ {\mathcal {T}} (\mathcal {T} (S ^ {m}), \mathcal {T} (S ^ {o})) \\ = \arg \min _ {\theta^ {m}} \mathbb {E} _ {\mathbf {o} _ {i}, \mathbf {o} _ {j} \sim \mathcal {B}} \left[ \operatorname {D i s t} \left(\psi^ {m} \left(\mathbf {s} _ {i} ^ {m}, \mathbf {s} _ {j} ^ {m}\right), \psi^ {o} \left(\mathbf {s} _ {i} ^ {o}, \mathbf {s} _ {j} ^ {o}\right)\right) \right]. \tag {8} \\ \end{array} +$$ + +$s_i^m = f^m (o_i|\theta^m,\beta), s_i^o = f^o (o_i|\theta^o), \psi^m (\cdot), \psi^o (\cdot)$ are the similarity functions among states of $S^{m}, S^{o}$ , respectively. + +Topological distillation. Here we introduce the process of topological distillation in detail. Given a batch of observations $\{\mathbf{o}_t\}$ , we encode them to latent states $\{\mathbf{s}_t^m\}$ and $\{\mathbf{s}_t^o\}$ with the momentum adaptor encoder $f^{m}(\cdot ;\theta^{m},\beta)$ and the online main encoder $f^{o}(\cdot ;\theta^{o})$ , respectively. In this paper, we suppose the states are abstracted in a high-dimensional Euclidean space, and use $L2$ distance as the similarity function $\psi^m (\cdot) = \| \cdot \| ^2$ for the teacher. Then the topological relation within a batch is formulated as an adjacent matrix of the state features from the teacher, referring to $\mathbf{A}^m\in \mathbb{R}^{N\times N}$ . $N$ is the batch size. $\mathbf{A}_{[i,j]}^m = \psi^m (\mathbf{s}_i^m,\mathbf{s}_j^m) = \| \mathbf{s}_i^m -\mathbf{s}_j^m\| ^2$ , where $[i,j]$ denotes the i-th row and the j-th column in the matrix $\mathbf{A}^m$ . Moreover, due to the unknown scale of the latent state, the matrix is normalized by its average value as $\| \mathbf{A}^m\|$ , + +$$ +\left\| \mathbf {A} ^ {m} \right\| _ {[ i, j ]} = \mathbf {A} _ {[ i, j ]} ^ {m} / \left(\frac {1}{N ^ {2}} \sum_ {m, n} \mathbf {A} _ {[ m, n ]} ^ {m}\right). \tag {9} +$$ + +However, directly minimizing the distance between topological relations with the same similarity function $\psi^o$ as $\psi^m$ will prevent the student online encoder from other basic discrimination. To flexibly distill knowledge and avoid disturbing the basic representation, the similarity function of the student's state space is defined by the $L2$ distance after a linear transformation $\phi$ with weight $\mathbf{W}$ , + +$$ +\psi^ {o} \left(\mathbf {s} _ {i} ^ {o}, \mathbf {s} _ {j} ^ {o}\right) = \| \phi \left(\mathbf {s} _ {i} ^ {o}\right) - \phi \left(\mathbf {s} _ {j} ^ {o}\right) \| ^ {2} = \| \mathbf {W} ^ {T} \mathbf {s} _ {i} ^ {o} - \mathbf {W} ^ {T} \mathbf {s} _ {j} ^ {o} \| ^ {2}. \tag {10} +$$ + +And the intra-batch adjacent matrix of the online encoder's state space is computed as $\mathbf{A}^o$ , where $\mathbf{A}_{[i,j]}^o = \psi^o (\mathbf{s}_i^o,\mathbf{s}_j^o)$ . Similar normalization as in Eq. 9 is also performed to $\mathbf{A}^o$ , referring to $\| \mathbf{A}^o\|$ . We distill the information of long-term expected returns from $\mathbf{s}^m$ to $\mathbf{s}^o$ by minimizing the L2 distance between their intra-batch similarity matrices $\mathbf{A}^m$ and + +$\mathbf{A}^o$ . The transferring loss is defined by + +$$ +\mathcal {L} _ {\text {t r a n s}} = \frac {1}{N ^ {2}} \sum_ {(i, j) \in [ 1, N ]} \operatorname {D i s t} \left(\mathbf {A} _ {[ i, j ]} ^ {m}, \mathbf {A} _ {[ i, j ]} ^ {o}\right), \tag {11} +$$ + +where $\mathrm{Dist}(u,v) = \max \left((u - v)^2 -\epsilon ,0\right)$ . Note that $\epsilon$ is a small loose factor, making the distillation learning more stable to converge. + +Dynamic prediction. Although the online encoder indirectly learns the long-term expected returns by the above distillation, its potential in representation capability is far from being developed. To this end, the online encoder is learned to capture other basic characters including environment dynamics as well as visual perception by using additional auxiliary tasks. As illustrated in Fig. 3, the predicted states $\mathbf{s}_t^o$ are conditioned on three auxiliary heads: a reward predictive head $\mathcal{R}$ , a state predictive head $\mathcal{P}$ and an observation decoder head $\mathcal{G}$ . Taking the current states $\mathbf{s}_t^o$ and actions $a_{t}$ as inputs, the $\mathcal{R}$ and $\mathcal{P}$ aim to predict the reward $\mathcal{R}(\mathbf{s}_t^o,a_t)$ and next state $\mathcal{P}(\mathbf{s}_t^o,a_t)$ , and thus learn a latent state space of which dynamic is consistent with the environment. Dynamic predictive loss is defined as + +$$ +\begin{array}{l} \mathcal {L} _ {\text {d y n a m i c}} = \mathcal {L} _ {\text {r e w a r d}} + \mathcal {L} _ {\text {s t a t e}} \\ = D _ {r} \left(\mathcal {R} \left(\mathbf {s} _ {t} ^ {o}, a _ {t}\right), r _ {t}\right) + D _ {s} \left(\mathcal {P} \left(\mathbf {s} _ {t} ^ {o}, a _ {t}\right), \mathbf {s} _ {t + 1} ^ {o}\right)) \\ = \left(\mathcal {R} \left(\mathbf {s} _ {t} ^ {o}, a _ {t}\right) - r _ {t}\right) ^ {2} + \left\| \mathcal {P} \left(\mathbf {s} _ {t} ^ {o}, a _ {t}\right) - \mathbf {s} _ {t + 1} ^ {o}\right) \| ^ {2}. \tag {12} \\ \end{array} +$$ + +To guarantee temporal capacity, the decoder $\mathcal{G}$ is trained to predict the next observation with the predicted next state $\hat{s}_{t + 1}^o = \mathcal{P}(\mathbf{s}_t^o,a_t)$ . The decoding loss is defined as + +$$ +\begin{array}{l} \mathcal {L} _ {d e c} = D _ {o} \left(\hat {\mathbf {o}} _ {t + 1}, \mathbf {o} _ {t + 1}\right) \\ = D _ {o} \left(\mathcal {G} \left(\mathcal {P} \left(\mathbf {s} _ {t} ^ {o}, \mathbf {a} _ {t}\right)\right), \mathbf {o} _ {t + 1}\right) \quad H W \tag {13} \\ = \frac {1}{H W} \sum_ {h, w = 1} ^ {H, W} \left(\mathcal {G} \left(\mathcal {P} \left(\mathbf {s} _ {t} ^ {o}, a _ {t}\right)\right) ^ {[ h, w ]} - \mathbf {o} _ {t + 1} ^ {[ h, w ]}\right) ^ {2}. \\ \end{array} +$$ + +Then the overall loss of the topology distilled transferring is formulated by + +$$ +\mathcal {L} _ {\text {e n c o d e r}} = \mathcal {L} _ {\text {t r a n s}} + \mathcal {L} _ {\text {d y n a m i c}} + \mathcal {L} _ {\text {o b s}}, \tag {14} +$$ + +where the online main encoder and auxiliary heads are jointly learned with the adaptor frozen. + +# 3.4. Greedy Bootstrapping for AIC + +In most deep reinforcement learning approaches, the representation module and policy module are trained in a simultaneous manner. That is, while the policy model is trained from the reinforcement signals the encoder is also simultaneously changing. This phenomenon will disturb policy improvement due to the drifting of state representation. To tackle this problem, this section presents a greedy bootstrapping framework for AIC, which consists of two + +Algorithm 1: Greedy Bootstrapping for AIC +1 Initialize the replay buffer $\mathcal{B}$ with random episodes. +2 while not converged do +3 // Representation learning +4 Build data-loader with the replay buffer $\mathcal{B}$ +5 for update step $c = 1\dots C$ do +6 Draw next batch of transitions $\{\left(\mathbf{o}_t,a_t,r_t,\mathbf{o}_{t + 1}\right)\}$ +7 //Momentum Adaptor Tuning +8 Update momentum encoder by moving average. +9 Update adaptor by optimizing Eq. 4. +10 //Topology Distilled Transferring +11 Update online encoder by optimizing Eq. 14. +12 end +13 // Policy learning (encoder frozen) +14 for update step $c_{p} = 1\dots S$ do +15 Draw transition sequences $\{\left(\mathbf{o}_t,a_t,r_t,\mathbf{o}_{t + 1}\right)\} \sim \mathcal{B}$ +16 Update critic and actor parameters. +17 end +18 Add new experiences to the replay buffer $\mathcal{B}$ +19 end + +alternative training stages for the representation and policy, respectively. In each representation stage, the encoder keeps training until convergence, and so does the policy in the other stage. The framework is described in detail in Algorithm 1. + +Greedy representation learning. Over the latest collected replay buffer, the encoder is trained by iterating all the images in each epoch. Different from sampling a random batch in each iteration like most other methods, our ergodic sampling strategy brings stable training without ignorance of any crucial samples. In each stage of representation learning, the encoder is trained for several epochs until it is converged. + +Greedy policy learning. With the main encoder trained sufficiently, the policy is learned continuously by freezing the encoder and conditioned on its output state. Beneficial from the stable and discriminative latent state, the critic and actor are trained efficiently from the reinforcement signal by SAC algorithms and tend to produce better behaviors. The policy is learned continuously for enough steps and then used to interact with the environment for new trajectories. + +# 4. Experiment + +# 4.1. Environments Setup + +**Benchmarks.** We evaluate our method on two reinforcement learning environments with complex and realistic observations, i.e., CARLA [5] and Vizdoom [22], with continuous and discrete action space, respectively. Both tasks are developed in 3D physical scenes with a complex state space, which calls for deeper visual encoders with stronger representation capability. + +Implementation details. We adopt ResNet-18 as the backbone of the visual encoder. Before the last pooling layer, the + +![](images/fc868855c08398cd6ccc0507c6ed4b1f9735667953c63895640f3d37f92142de.jpg) +Figure 5. Comparison with other baseline approaches on the CARLA environment. + +![](images/6d661de121118efee8834a7077d2d2a7b7374bfcf17d77497586e3fa5dfdc8f1.jpg) +Figure 6. Ablation Study of our method using ResNet-18 as the visual encoder on the CARLA environment. + +feature map is flattened and then projected into 1024-dim state features with a linear layer and layer norm operation. The momentum factor $\tau$ is set to 0.05. We use Adam to optimize the parameters of all modules in our approach. + +# 4.2. CARLA Autonomous Driving Environment + +Experimental settings. In the autonomous driving task, we construct a highway driving scenario with 20 other vehicles of different models using the CARLA simulator. We use five cameras on the roof of the ego-vehicle, each with a 60-degree view and obtaining images of $84 \times 84$ pixels. We concatenate them together for observation of $84 \times 420$ pixels. We stack the last three frames as the inputs. Following DBC [40], the reward is formulated as $r_t = v_{ego}^T \hat{u}_h \cdot \Delta t - \xi_1 \cdot \mathbb{I} - \xi_2 \cdot |steer|$ , where $v_{ego}$ is the velocity vector of the ego-vehicle, projected onto the highway's unit vector $\hat{u}_h$ , and multiplied by time discretization $\Delta t = 0.05$ to measure highway progression in meters. Impulse $\mathbb{I} \in \mathbb{R}^+$ is caused by collisions, and a steering penalty steer $\in [-1,1]$ facilitates lane-keeping. The hyperparameters $\xi_1$ and $\xi_2$ are set to $10^{-4}$ and 1, respectively. + +Comparison with other methods. We compare our ap + +![](images/6977a11160d6966730f280a0113e56266738595547c05a1eacd92b104bfe8d68.jpg) +Figure 7. Comparison of other representation-decoupled methods on the CARLA environment. + +proach with seven baselines, including joint representation learning methods, i.e., CURL [27], SVEA [16], Deep-MDP [7], DBC [40]; representation decoupled methods, i.e., DreamerV2 [13], Iso-Dream [32], and the baseline SAC [10]. As shown in Fig. 5, our AIC outperforms all baselines by large margins. Compared with joint representation learning, the superior performance can be attributed to that our method stably learns a more discriminative state abstraction with deeper architectures, while avoiding the oscillating self-overfitting. Representation decoupled methods train encoders only on auxiliary tasks without TD losses and thus usually require more time to converge. However, AIC effectively learns the long-term expected returns by interaction with the light-optimizable momentum adaptor encoder, resulting in faster convergence. + +Ablation Study. We conduct ablation studies to validate the effectiveness of the proposed topology distilled transferring and the greedy bootstrapping strategy. All the experiments train the ResNet-18 as the encoder. Fig. 6 shows that SAC fails to learn an effective policy with the heavy-optimizable encoder since the state features are deteriorated by the serious oscillating self-overfitting. By applying the dynamic loss and decoder loss, the deterioration is slightly alleviated, since the encoder is encouraged to learn the inherent information of environment dynamics by the targets which are more stable than TD. However, disturbed by the oscillating gradients of the TD loss, the encoder is still hard to converge. Compared to it, the greedy bootstrapping strategy achieves significant improvement, denoted as AIC $w \backslash o L_{trans}$ , because it interrupts the backward propagation of the unstable gradient and simultaneously it trains the encoder and policy stably and sufficiently. AIC distinctly outperforms the AIC $w \backslash o L_{trans}$ due to the topology distilled transferring, which additionally incorporates the long-term expected return information into the encoder, making it more adaptive for the downstream controlling. + +Comparison with other decoupling methods. To further + +![](images/643cda3bc755d17d8704662373cdb1563fbc35fa1b1e0134cba03c931b239ce3.jpg) +Figure 8. Comparison with different architectures of the encoder on the CARLA environment. + +![](images/f09ebae0686f1312673628b57fdb906cb41a840eb90460e55cb1d330c3477370.jpg) +Figure 9. Comparison of other stabilizing methods using ResNet-18 on the CARLA environment. + +evaluate the greedy bootstrapping framework, we compare it with other two typical methods that decouple representation learning from RL. Initial decoupling first collects a set of transitions using a random policy and then pretrains the encoder with auxiliary tasks as in Eq. 12 and 13. Afterward, the policy is trained by SAC with the encoder frozen. As shown in Fig. 7, the initial decoupling method performs high-variance results because the performance significantly depends on whether the distribution of the collected pretraining data is similar to those from a learned policy. Stepping decoupling learns the encoder and the policy step by step, respectively. The encoder is trained by transitions from the same data distribution as the policy. However, since the encoder keeps updating during policy learning, the latent state space suffers from unstable drifting, making it difficult to learn policy from the volatile state abstraction. The greedy bootstrapping strategy surpasses both two approaches because it guarantees a determined latent state space that is conducive to the inductive learning of policy. Moreover, it learns the encoder by traversing all the transitions in the replay buffer, prevented from ignorance of some critical images or overfitting to parts of data. + +Table 1. Comparison with other methods using ResNet-18 on Medikit of Vizdoom environment. + +
-DQN (Baseline)DrQSVEACURLDeepMDPAIC (Ours)
Episode Returns (mean/std)888±5841361 ± 1004435±185408±2031026±8071686±1283
Min returns252252792428456
Max returns244442001008113637104800
Survival Time (mean/std)64 ±2685±4345±844±972±3796±53
Min Time262627233926
Max Time1352106874191210
+ +Improvements on different encoders. We evaluate our approach with different architectures as the visual encoder, including a light CNN of 4-layer convolution (Conv-4), and three heavyweight models, i.e., VGG-11, ResNet-18, and ResNet34. For each architecture, we conduct three experiments using SAC, SAC with auxiliary tasks (dynamic loss and decoder loss as in Eq. 12 and 13), and our AIC approach. The comparison is illustrated in Fig. 8. Firstly, all heavyweight models fail to learn effective policies using SAC, where only TD loss is used. The results indicate the oscillating self-overfitting is consistent for heavy models. Secondly, auxiliary tasks improve the performance of SAC with heavyweight encoders, but they are still inferior to the light encoder like Conv-4. The oscillating self-overfitting seriously prevents the encoder from sufficient learning. Finally, AIC consistently improves the performance with all heavy encoders by large margins, as well as the light encoder of Conv-4. It shows a consistent rank of performance that these architectures have achieved in supervised learning tasks. It demonstrates that our method makes full use of the representation capability of heavyweight visual encoders through interactive cooperation. + +Comparison with stabilizing methods. To evaluate AIC for stabilizing the training of heavy-optimizable encoders, we compare it with SVEA and DrQ using ResNet-18 as the backbone. As shown in Fig. 9, both two methods perform not well, since the increase of parameters significantly exacerbates the challenge of stabilizing. However, AIC exceeds them by large margins, indicating its capability to stably learn heavy models. + +Qualitative analysis. In order to evaluate the effectiveness of our approach in alleviating the oscillating self-overfitting, we visualize the updating trajectories of the encoder parameters over a set of training steps by TSNE. We use ResNet-18 as the encoder backbone and choose the weights of the last convolution layer for visualization. The comparison between DeepMDP with the TD loss and AIC is shown in Fig. 10, where each point represents the parameters in a step and the color changes from deep to light over the training steps. As we can see, the parameters with the TD loss of the DeepMDP agent suffer from significant oscillation during training. Different from that, the parameters of AIC show a dis + +![](images/92a2dc787f62f3400c0431aa8810243c3af0590866542fd6caf95b92f9762793.jpg) +Figure 10. Parameter updating trajectories of the ResNet-18 encoder over training steps. + +![](images/4e5972fe737d309c73cc5428bc929eda398f47e27d6794ca1ee8ff3f0d72b6e6.jpg) + +tinct optimizing direction and only have slight oscillation in a short time. The qualitative results demonstrate AIC effectively alleviates the oscillating self-overfitting for the heavy-optimizable encoder in visual reinforcement learning. + +# 4.3. Vizdoom Medikit Environment + +Experimental settings. Vizdoom is based on the first-person perspective video game in a semi-realistic 3D world. We use the difficult medikit collecting scenario, where the agent is spawned in a random spot of a maze with an acid surface, which slowly, but constantly, takes away the agent's life. The agent needs to collect medikits and avoid blue vials with poison, both of which appear in random places during the episode. In each step, the agent is allowed to move (forward/backward), or turn (left/right). The reward is formulated as $r_t = 1 - \xi_3 \cdot E_m - \xi_4 \cdot E_v - \xi_5 \cdot E_d$ , where $E_m$ , $E_v$ , $E_d$ denote the events of getting a medikit, a vial and to death, respectively. $\xi_3 = \xi_4 = \xi_5 = 100$ . + +Superior performance with heavy encoders. To evaluate our approach, especially with deep deep encoders, we compare it with five methods. All of them use ResNet-18 as backbones. Since the Vizdoom is of discrete action space, we adopt DQN as the baseline. As shown in Table 1, AIC performs superiorly to the compared methods. CURL and SVEA perform poorly, since the observations of such a fine-grained collecting task of first-person perspective are sensitive to random cropping of CURL. And the random conv of + +SVEA is not applicable for the low-contrast images in this scenario. However, AIC is able to effectively train heavy models under different conditions. + +# 5. Conclusion + +In this paper, we investigate the phenomenon of oscillating self-overfitting which results in the deterioration of RL when training heavy-optimizable encoders. To alleviate this problem, we propose a novel asymmetric interactive cooperation approach. Through interaction between a main heavy-optimizable encoder and a supportive light-optimizable encoder, both advantages of them are integrated including the highly discriminative capability and the training stability. We further present a greedy bootstrapping framework for further stability. Experimental results of CARLA and Vizdoom show that AIC achieves competitive performance over complex and realistic environments. + +# Acknowledgements + +This work is partially supported by grants from the Key-Area Research and Development Program of Guangdong Province under contact No. 2020B0101380001 and grants from the National Natural Science Foundation of China under contract No. 62027804, No. 61825101, No. 62088102 and No. 62202010. The computing resources of Pengcheng Cloudbrain are used in this research. + +# References + +[1] Homanga Bharadhwaj, Mohammad Babaeizadeh, Dumitru Erhan, and Sergey Levine. Information prioritization through empowerment in visual model-based rl. arXiv preprint arXiv:2204.08585, 2022. +[2] Pablo Samuel Castro. Scalable methods for computing state similarity in deterministic markov decision processes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10069-10076, 2020. +[3] Edoardo Cetin, Philip J Ball, Steve Roberts, and Oya Celiktutan. Stabilizing off-policy deep reinforcement learning from pixels. arXiv preprint arXiv:2207.00986, 2022. +[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. +[5] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. +[6] Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 2786-2793. IEEE, 2017. +[7] Carles Gelada, Saurabh Kumar, Jacob Buckman, Ofir Nachum, and Marc G Bellemare. Deepmdp: Learning + +continuous latent space models for representation learning. In International Conference on Machine Learning, pages 2170-2179. PMLR, 2019. +[8] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018. +[9] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. +[10] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018. +[11] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019. +[12] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International conference on machine learning, pages 2555-2565. PMLR, 2019. +[13] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. +[14] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. +[15] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. arXiv preprint arXiv:2007.04309, 2020. +[16] Nicklas Hansen, Hao Su, and Xiaolong Wang. Stabilizing deep q-learning with convnets and vision transformers under data augmentation. 2021. +[17] Nicklas Hansen and Xiaolong Wang. Generalization in reinforcement learning by soft data augmentation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13611-13617. IEEE, 2021. +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[19] Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016. +[20] Michael Janner, Qiyang Li, and Sergey Levine. Offline reinforcement learning as one big sequence modeling problem. Advances in neural information processing systems, 34:1273-1286, 2021. +[21] Lukasz Kaiser, Mohammad Babaeizadeh, Piotr Milos, Blazej Osinski, Roy H Campbell, Konrad Czechowski, Dumitru Erhan, Chelsea Finn, Piotr Kozakowski, Sergey Levine, et al. Model-based reinforcement learning for atari. arXiv preprint arXiv:1903.00374, 2019. + +[22] Michal Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Vizdoom: A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE conference on computational intelligence and games (CIG), pages 1-8. IEEE, 2016. +[23] Ilya Kostrikov, Denis Yarats, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. arXiv preprint arXiv:2004.13649, 2020. +[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. +[25] Aviral Kumar, Rishabh Agarwal, Dibya Ghosh, and Sergey Levine. Implicit under-parameterization inhibits data-efficient deep reinforcement learning. arXiv preprint arXiv:2010.14498, 2020. +[26] Misha Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. Advances in neural information processing systems, 33:19884-19895, 2020. +[27] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In International Conference on Machine Learning, pages 5639-5650. PMLR, 2020. +[28] Alex X. Lee, Anusha Nagabandi, Pieter Abbeel, and Sergey Levine. Stochastic latent actor-critic: Deep reinforcement learning with a latent variable model. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 741-752. Curran Associates, Inc., 2020. +[29] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. The Journal of Machine Learning Research, 17(1):1334-1373, 2016. +[30] Masashi Okada and Tadahiro Taniguchi. Dreaming: Model-based reinforcement learning by latent imagination without reconstruction. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4209-4215. IEEE, 2021. +[31] Kei Ota, Devesh K Jha, and Asako Kanezaki. Training larger networks for deep reinforcement learning. arXiv preprint arXiv:2102.07920, 2021. +[32] Minting Pan, Xiangming Zhu, Yunbo Wang, and Xiaokang Yang. Iso-dream: Isolating and leveraging noncontrollable visual dynamics in world models. In Advances in Neural Information Processing Systems. +[33] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014. +[34] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International Conference on Machine Learning, pages 8583-8592. PMLR, 2020. +[35] Evan Shelhamer, Parsa Mahmoudieh, Max Argus, and Trevor Darrell. Loss is its own reward: Self-supervision for reinforcement learning. arXiv preprint arXiv:1612.07307, 2016. + +[36] Oriol Vinyals, Igor Babuschkin, Wojciech M Czarnecki, Michael Mathieu, Andrew Dudzik, Junyoung Chung, David H Choi, Richard Powell, Timo Ewalds, Petko Georgiev, et al. Grandmaster level in starcraft ii using multiagent reinforcement learning. Nature, 575(7782):350-354, 2019. +[37] Manuel Watter, Jost Springenberg, Joschka Boedecker, and Martin Riedmiller. Embed to control: A locally linear latent dynamics model for control from raw images. Advances in neural information processing systems, 28, 2015. +[38] Denis Yarats, Amy Zhang, Ilya Kostrikov, Brandon Amos, Joelle Pineau, and Rob Fergus. Improving sample efficiency in model-free reinforcement learning from images. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10674-10681, 2021. +[39] Tao Yu, Zhizheng Zhang, Cuiling Lan, Yan Lu, and Zhibo Chen. Mask-based latent reconstruction for reinforcement learning. In Advances in Neural Information Processing Systems. +[40] Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant representations for reinforcement learning without reconstruction. arXiv preprint arXiv:2006.10742, 2020. \ No newline at end of file diff --git a/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/images.zip b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..79d40446881898311bc54e697fefec96fbc137b5 --- /dev/null +++ b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6217ea448e20a21e3c819e23ee5abd86a9828ff0b38f0d4d18f8fd968cfccac +size 529410 diff --git a/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/layout.json b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7dfda72eb2c0b08edbcb09f5afb238d154aa9b15 --- /dev/null +++ b/2023/Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation/layout.json @@ -0,0 +1,8901 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 479, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 479, + 139 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 479, + 139 + ], + "type": "text", + "content": "Stabilizing Visual Reinforcement Learning via Asymmetric Interactive Cooperation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "spans": [ + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "text", + "content": "Yunpeng Zhai" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "text", + "content": ", Peixi Peng" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "inline_equation", + "content": "^{2,3,*}" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "text", + "content": ", Yifan Zhao" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "text", + "content": ", Yangru Huang" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "text", + "content": ", Yonghong Tian" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "inline_equation", + "content": "^{1,2,3,*}" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "text", + "content": "National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University, Beijing, China \n" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "text", + "content": "School of Electronic and Computer Engineering, Peking University Shenzhen Graduate School, Shenzhen, China \n" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 160, + 499, + 245 + ], + "type": "text", + "content": "Peng Cheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 247, + 477, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 247, + 477, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 247, + 477, + 258 + ], + "type": "text", + "content": "{ypzhai, pxpeng, zhaoyf, yhtian}@pku.edu.cn, yrhuang@stu.pku.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 312, + 290, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 290, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 290, + 600 + ], + "type": "text", + "content": "Vision-based reinforcement learning (RL) depends on discriminative representation encoders to abstract the observation states. Despite the great success of increasing CNN parameters for many supervised computer vision tasks, reinforcement learning with temporal-difference (TD) losses cannot benefit from it in most complex environments. In this paper, we analyze that the training instability arises from the oscillating self-overfitting of the heavy-optimizable encoder. We argue that serious oscillation will occur to the parameters when enforced to fit the sensitive TD targets, causing uncertain drifting of the latent state space and thus transmitting these perturbations to the policy learning. To alleviate this phenomenon, we propose a novel asymmetric interactive cooperation approach with the interaction between a heavy-optimizable encoder and a supportive light-optimizable encoder, in which both their advantages are integrated including the highly discriminative capability as well as the training stability. We also present a greedy bootstrapping optimization to isolate the visual perturbations from policy learning, where representation and policy are trained sufficiently by turns. Finally, we demonstrate the effectiveness of our method in utilizing larger visual models by first-person highway driving task CARLA and Vizdoom environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 624, + 128, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 128, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 128, + 636 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 644, + 287, + 693 + ], + "type": "text", + "content": "Learning complex control from high-dimensional observations such as images is significant for many real-world applications [37, 28]. It puts forward higher requirements for the representation capability of visual encoder models," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 293, + 424, + 398 + ], + "blocks": [ + { + "bbox": [ + 313, + 293, + 424, + 398 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 424, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 424, + 398 + ], + "type": "image", + "image_path": "3416461f6b614d61f15d3450008213d0cc1ecd711d5c084989a3cbdb9549c191.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 399, + 410, + 409 + ], + "lines": [ + { + "bbox": [ + 331, + 399, + 410, + 409 + ], + "spans": [ + { + "bbox": [ + 331, + 399, + 410, + 409 + ], + "type": "text", + "content": "a) Returns during training" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 427, + 293, + 534, + 398 + ], + "blocks": [ + { + "bbox": [ + 427, + 293, + 534, + 398 + ], + "lines": [ + { + "bbox": [ + 427, + 293, + 534, + 398 + ], + "spans": [ + { + "bbox": [ + 427, + 293, + 534, + 398 + ], + "type": "image", + "image_path": "4415384ff751e746add5f4526c56413e1ac2d8534a4703b9498b9597d93b8a1a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 399, + 518, + 408 + ], + "lines": [ + { + "bbox": [ + 445, + 399, + 518, + 408 + ], + "spans": [ + { + "bbox": [ + 445, + 399, + 518, + 408 + ], + "type": "text", + "content": "b) Parameter oscillation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 413, + 547, + 501 + ], + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 501 + ], + "type": "text", + "content": "Figure 1. Comparison of DeepMDP agents on CARLA with light- and heavy-weight visual encoders in episode returns (a) and parameter oscillation " + }, + { + "bbox": [ + 304, + 413, + 547, + 501 + ], + "type": "inline_equation", + "content": "\\left(\\sum_{t}^{T}\\|\\nabla\\theta_{t}\\|\\right)/\\|\\sum_{t}^{T}\\nabla\\theta_{t}\\|" + }, + { + "bbox": [ + 304, + 413, + 547, + 501 + ], + "type": "text", + "content": " (b). Compared with 4-layer CNN, larger models like ResNet-18 do not improve the RL performance as in supervised learning. The deterioration results from the oscillating self-overfitting of the larger model, which occurs in neither supervised learning nor RL with a lightweight encoder." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "content": "especially in some complex scenes such as self-driving [5] and robot controlling [29]. The last decade has witnessed impressive progress in computer vision by training large-scale networks [18] as they increase the search space of possible solutions. However, such parameter increment of visual models cannot directly benefit the reinforcement learning, and even leads to deterioration of training. For example, as shown in Fig. 1(a), we illustrated the training curves of DeepMDP [7] agents that use different CNN networks as visual encoders on the self-driving environment CARLA [5]. The result shows that using a larger model, e.g., ResNet-18 [18], leads to unstable training and achieves distinctly lower returns than the lighter model with only four convolutional layers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "To investigate this phenomenon, we quantified the oscillation of the network parameters by introducing an in-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "207" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": "indicator calculated by the ratio of accumulation of modulus length of the gradient to the module lengths of cumulative gradient within " + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": " training steps, " + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "inline_equation", + "content": "(\\sum_{t}^{T}\\|\\nabla\\theta_{t}\\|)/\\|\\sum_{t}^{T}\\nabla\\theta_{t}\\|" + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": " is gradient operator and " + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": " is parameters of the last convolution layer in the " + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": "-th step. A higher indicator means worse oscillation and instability of parameters during training. As shown in Fig. 1(b), we compared the oscillation in three experiments including i) training ResNet-18 on RL, ii) training 4-layer CNN on RL, and iii) training ResNet-18 on supervised learning (SL) with CIFAR-100 [24]. ResNet-18 suffers much more serious oscillation when training on RL, resulting in deteriorated performance. However, such oscillation occurs neither on SL with ResNet-18 nor on RL with the lighter 4-layer CNN. We name this phenomenon as oscillating self-overfitting, which particularly results from a pathological concurrence of the overfitting capability of large models and the sensitive learning targets of the temporal-difference (TD) loss [25]. Specifically, when the heavyweight parameters are enforced to fit the sensitive TD targets which are partially generated by themselves with a bootstrapping formulation, contradictory gradients will propagate back to oscillate the parameters. This phenomenon results in uncertain drifting of the state space and transmits these perturbations to policy learning. Therefore, stabilizing visual encoders with amounts of parameters from TD losses for performance gain in RL is still an open challenge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 403, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 289, + 715 + ], + "type": "text", + "content": "From the perspective of the relation between the encoder and the TD objective, existing representation learning for RL can be categorized into two groups, as illustrated in Fig. 2. One approach jointly learns representation with policy by the TD loss [16, 7, 23]. It efficiently learns the long-term expected returns with light encoders while is incompatible with heavy-optimizable encoders due to the oscillating self-overfitting. Another group of approaches decouples representation learning from RL to avoid instability, where the encoder is learned only by auxiliary dynamic predictive losses [11, 32]. However, it is inaccessible to the expected returns from the bootstrapping objective in RL. In this paper, we propose a novel asymmetric interactive cooperation for representation learning in RL. To take both the advantages of representation capability and the stability for TD targets, it separately trains a main heavy-optimizable encoder and a supportive light-optimizable encoder by auxiliary tasks and TD losses, respectively. And the asymmetric interaction is simultaneously conducted between them to effectively exchange their knowledge from each other, where the heavy one transfer the representation capability by parameter momentum and the light one transfer the long-term expected returns by topological distillation. Hence, the heavy encoder is equipped with the capability of latent state abstraction without oscillating by the TD objective. Moreover, we present a greedy bootstrapping optimization for" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 425, + 166 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 425, + 166 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 425, + 166 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 425, + 166 + ], + "type": "image", + "image_path": "64212d11eddc8d9940c85d488b7263a97b323a3bd425821c5f64fa713b2956a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 545, + 166 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 545, + 166 + ], + "type": "image", + "image_path": "5afeef562035408a31e9ad91cf5676a319b1071ef47342e6208ceef07017fcd7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 168, + 545, + 274 + ], + "blocks": [ + { + "bbox": [ + 308, + 168, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 308, + 168, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 168, + 545, + 274 + ], + "type": "image", + "image_path": "3cf6e94cf456eed5a8b19103076a63cf391a455c2e6f6b5673719ddfd659339e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 277, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 344 + ], + "type": "text", + "content": "Figure 2. Illustrations of different learning paradigms of visual encoders. a) Joint Learning trains the visual encoder end-to-end with reinforcement loss. b) Decoupled learning trains the encoder only by auxiliary tasks. c) The proposed AIC trains a supportive light-optimizable encoder with the expected returns information from TD targets and transfers it to the main visual encoder." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 365, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 389 + ], + "type": "text", + "content": "further stability of training, where representation and policy are trained sufficiently by turns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 391, + 546, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 546, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 546, + 535 + ], + "type": "text", + "content": "The main contributions of this paper can be summarized in three aspects. First, it investigates the phenomenon of oscillating self-overfitting that leads to deterioration in RL with heavy-optimizable encoders, and proposes a novel asymmetric interactive cooperation to alleviate it. Second, it presents a topological distillation between latent state spaces to learn state abstraction by interactive cooperation without any explicit labels. Third, it presents a greedy bootstrapping optimization for further stability of training. Experiments demonstrate a significant performance gain over the complex and realistic environments of CARLA and Vizdoom." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 548, + 392, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 392, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 392, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Jointly Representation Learning with RL. Since learning control directly from high-dimension visual observations is hard to converge, many approaches learn the encoder jointly with RL and elaborate auxiliary tasks [39, 38, 26, 6, 21, 34], as shown in Fig. 2(a). The first line of methods utilizes self-supervision to learn the visual invariances [17, 16, 15, 35]. CURL [27] introduced a contrastive loss to learn discriminative features from raw pixels. The second line learns to abstract the state by preserving the property of Markov decision processes (MDPs) [12, 19, 2]. DeepMDP [7] learns the latent space with the same dynamic transition as the environment. DBC [40] learns invariant representations with" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "208" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 542, + 339 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 542, + 339 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 542, + 339 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 542, + 339 + ], + "type": "image", + "image_path": "3c9056d7dbf55b31eb1475a3f0a4f1e33b0fa55b0cbcbd08cc4a1f9992bd3112.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 341, + 548, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 341, + 548, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 548, + 399 + ], + "type": "text", + "content": "Figure 3. Flowchart of the proposed asymmetric interactive cooperation. The representation learning consists of two components including momentum adaptor tuning and topological distilled transferring. In the momentum adaptor tuning, the adaptor is optimized by RL to learn the information of expected returns while the momentum encoder is updated by moving average from the online encoder. In topology distilled transferring, the online encoder is trained by three auxiliary tasks while transferring the information of expected returns by topology distillation from the momentum adaptor encoder. The policy is learned by freezing the online encoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 416, + 289, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 289, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 289, + 596 + ], + "type": "text", + "content": "the bisimulation distance. The third line of methods designs regularization to stabilize the training process [23]. DrQ [23] and SVEA [16] propose data regularization using image augmentation of random shift or random convolution. A-LIX [3] adaptively regularizes the convolutional features to prevent overfitting while it doesn't study the heavy visual models. The common drawback of these methods is the difficulty for the heavy-optimizable encoders due to oscillating self-overfitting caused by TD losses, resulting in more serious instability. Other recent works have concerned the stability of large models in RL [25, 31], but they focus on a different component of policy learning instead of representation. We also note recent work RL with Transformer while they only contain slight parameters [4, 20] or rely on supervised imitation learning [36]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Decoupled Representation Learning from RL. Recent methods proposed to decouple representation learning from TD losses for efficient training [30, 31], in Fig. 2(b). A typical approach is the world model [9, 8]. It first pretrains an encoder using Variational Autoencoder [33] by the collected rollouts from a random policy and then trains policy with the encoder frozen. However, the encoder does not perform well since the pretraining data is of different distribution from those in policy learning. Recent world model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 416, + 547, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 547, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 547, + 561 + ], + "type": "text", + "content": "class methods [14, 30, 32, 13, 1], such as Dreamer [11], turn to learn the encoder and the policy alternatively step-by-step, in which the encoder is updated by the newly collected data. Theoretically, these approaches can train any large model stably. But there still exists a bottleneck, that is, the encoder is isolated from the long-term expected returns hidden in the TD losses, which is significant for the state abstraction with respect to the downstream controls. In our work, we tackle this problem by interacting with a supportive light-optimizable encoder, which learns from TD losses stably and transfers its knowledge to the main encoder, as shown in Fig. 2(c)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 574, + 446, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 574, + 446, + 590 + ], + "spans": [ + { + "bbox": [ + 305, + 574, + 446, + 590 + ], + "type": "text", + "content": "3. The Proposed Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 596, + 373, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 596, + 373, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 596, + 373, + 608 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "content": "Asymmetric interactive cooperation alternatively trains a light-optimizable supportive encoder with TD losses to learn the long-term expected returns without oscillating self-overfitting, and a heavy-optimizable main encoder with auxiliary tasks to capture stronger representation capability. Simultaneously, two kinds of interaction are conducted between them to exchange their knowledge including:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "- Light " + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": " Heavy: parameter momentum to absorb the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "209" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 209, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 209, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 209, + 83 + ], + "type": "text", + "content": "representation capability, as in Sec. 3.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 108 + ], + "type": "text", + "content": "- Light " + }, + { + "bbox": [ + 46, + 84, + 286, + 108 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 84, + 286, + 108 + ], + "type": "text", + "content": " Heavy: topological distillation to learn the long-term expected returns, as in Sec. 3.3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 108, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 156 + ], + "type": "text", + "content": "Beyond them, it presents a greed bootstrapping framework for AIC to alleviate the drifting of the state space for further stability in Sec. 3.4. The flowchart of AIC is illustrated in Fig. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 163, + 203, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 203, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 203, + 177 + ], + "type": "text", + "content": "3.2. Momentum Adaptor Tuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "text", + "content": "Attaching adaptor to momentum. To efficiently learn the long-term expected returns, the light-optimizable encoder should satisfy the following two conditions: 1) It only contains a few learnable parameters so that it converges without oscillating self-overfitting. 2) It can utilize the middle-level representation of the heavy-optimizable encoder rather than relying only on the raw observation input. To fulfill the above conditions, we implement the light-optimizable encoder by applying learnable CNN adaptors " + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "text", + "content": " to the momentum network " + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\theta^m" + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "text", + "content": " of the online heavy encoder " + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\theta^o" + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "text", + "content": ". As illustrated in Fig. 3, each " + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "text", + "content": " convolutional layer in the momentum encoder is connected with a " + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 46, + 182, + 287, + 350 + ], + "type": "text", + "content": " convolutional adaptor module in a parallel manner. Thus the result of each layer is formulated by," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 359, + 287, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 359, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 80, + 359, + 287, + 372 + ], + "type": "interline_equation", + "content": "x _ {l + 1} = \\rho \\left(x _ {l}; \\theta_ {l} ^ {m}, \\beta_ {l}\\right) = \\theta_ {l} ^ {m} * x _ {l} + \\beta_ {l} * x _ {l}, \\tag {1}", + "image_path": "df689edbc99441a7e0a56856fb41c3d044987b1b4f8a8233916c00183d07659e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": " is the input tensor of the " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": "-th layer and " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": " denotes the convolution operator. " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\theta_{l}^{m}" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\beta_{l}" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": " are the weights of the momentum model and the adaptor, respectively. We name this light-optimizable encoder as momentum adaptor encoder, referring to " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "f^{m}" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": ". It interacts with the online heavy encoder " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "f^{o}" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": " by parameter momentum to utilize its representation. In each iteration step, the momentum model is updated by moving average of the online encoder " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "f^{o}" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 487, + 287, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 487, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 287, + 499 + ], + "type": "interline_equation", + "content": "\\theta^ {m} \\leftarrow \\tau \\theta^ {m} + (1 - \\tau) \\theta^ {o}, \\tag {2}", + "image_path": "39f2791065e77256f840c6eb532c29732775c7e51b4ea97cf33221acc66597d5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 509, + 286, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 286, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 286, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 509, + 286, + 532 + ], + "type": "inline_equation", + "content": "\\theta^o" + }, + { + "bbox": [ + 46, + 509, + 286, + 532 + ], + "type": "text", + "content": " is the parameters of online encoder and " + }, + { + "bbox": [ + 46, + 509, + 286, + 532 + ], + "type": "inline_equation", + "content": "\\tau \\in [0,1]" + }, + { + "bbox": [ + 46, + 509, + 286, + 532 + ], + "type": "text", + "content": " is the momentum factor." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 533, + 286, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 556 + ], + "type": "text", + "content": "Adaptor Tuning. The adaptor is updated with RL by maximizing the expected return," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 79, + 565, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 565, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 79, + 565, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\beta^ {*} = \\arg \\max _ {\\beta} \\mathbb {E} \\left[ \\sum_ {t} ^ {\\infty} \\gamma^ {t} R \\left(s _ {t}, a _ {t}, s _ {t + 1}\\right) \\right], \\tag {3}", + "image_path": "66c5613ed26b2a0575e8c034c808c3969f2f42523457091c8dbb9b7954b68fe8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " is the discount factor and " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " is the reward. " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " are the state and action in time step " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". Given a batch of transitions " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{o}_t,a_t,r_t,\\mathbf{o}_{t + 1})\\}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " from the replay buffer " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_t\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " is the observation stacking multiple frames. " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " are the height and width of the image and " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " is the channel number. We first projected the observation from images to a latent state " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^m = f^m (\\mathbf{o}_t;\\theta^m,\\beta)" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " with the momentum adaptor encoder " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "f^{m}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". To tune the adaptor with the reinforcement learning, we optimize the predicted state by" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 324, + 73, + 545, + 175 + ], + "blocks": [ + { + "bbox": [ + 324, + 73, + 545, + 175 + ], + "lines": [ + { + "bbox": [ + 324, + 73, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 324, + 73, + 545, + 175 + ], + "type": "image", + "image_path": "4754f2ecbb5ab2b0ca074510761630ccf7b5dc160c5e7ec2524856184dcaa515.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 184, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 305, + 184, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 184, + 545, + 216 + ], + "type": "text", + "content": "Figure 4. Illustration of the distance between state topology in different latent spaces. " + }, + { + "bbox": [ + 305, + 184, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\psi^m (\\cdot)" + }, + { + "bbox": [ + 305, + 184, + 545, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 184, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\psi^o (\\cdot)" + }, + { + "bbox": [ + 305, + 184, + 545, + 216 + ], + "type": "text", + "content": " denote the similarity function of states." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "text", + "content": "conditioning it on a learnable critic head " + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "inline_equation", + "content": "h_c" + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "text", + "content": " and an actor head " + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "text", + "content": " for SAC [10] algorithms. Both two heads are implemented by 3-layer MLPs. Then the critic takes state " + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t^m" + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "text", + "content": " and action " + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "text", + "content": " as inputs and learns to predict " + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "inline_equation", + "content": "Q(\\mathbf{s}_t^m, a_t)" + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "text", + "content": " with the critic loss," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 349, + 302, + 545, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 302, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 349, + 302, + 545, + 317 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c r i t i c}} = \\left(Q \\left(\\mathbf {s} _ {t}, a _ {t}\\right) - y \\left(r _ {t}, \\mathbf {s} _ {t + 1}\\right)\\right) ^ {2}, \\tag {4}", + "image_path": "5523381ddcf3ad0b0c50419dcf552beac1a17a9df5c3821840906b839dafd9a1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 325, + 435, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 325, + 435, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 325, + 435, + 337 + ], + "type": "text", + "content": "where the TD target is given by" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 346, + 550, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 550, + 370 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 550, + 370 + ], + "type": "interline_equation", + "content": "y \\left(r _ {t}, \\mathbf {s} _ {t + 1}\\right) = r _ {t} + \\gamma \\left(\\bar {Q} \\left(\\mathbf {s} _ {t + 1}, \\hat {a} _ {t + 1}\\right) - \\alpha_ {r} \\log \\pi \\left(\\hat {a} _ {t + 1} \\mid \\mathbf {s} _ {t + 1}\\right)\\right). \\tag {5}", + "image_path": "6501fbaa39dbf2dc117a90a4508996630ae4b5e66ec7d2567652d87745b4336f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 370, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\bar{Q}" + }, + { + "bbox": [ + 305, + 370, + 545, + 430 + ], + "type": "text", + "content": " is the target critic by moving average, and " + }, + { + "bbox": [ + 305, + 370, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\pi (\\hat{a}_{t + 1}|\\mathbf{s}_{t + 1})" + }, + { + "bbox": [ + 305, + 370, + 545, + 430 + ], + "type": "text", + "content": " is the policy distribution predicted by the actor head. " + }, + { + "bbox": [ + 305, + 370, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\alpha_{r}" + }, + { + "bbox": [ + 305, + 370, + 545, + 430 + ], + "type": "text", + "content": " is the temperature parameter in SAC. " + }, + { + "bbox": [ + 305, + 370, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\hat{a}_{t + 1}\\sim \\pi (\\cdot |\\mathbf{s}_{t + 1})" + }, + { + "bbox": [ + 305, + 370, + 545, + 430 + ], + "type": "text", + "content": " . The actor learns to generate the action that maximizes the Qvalue predicted by the critic," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 348, + 439, + 545, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 439, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 348, + 439, + 545, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a c t o r}} = \\alpha_ {r} \\log \\pi (\\hat {a} _ {t} | \\mathbf {s} _ {t}) - Q (\\mathbf {s} _ {t}, \\hat {a} _ {t}). \\tag {6}", + "image_path": "bf7b4f500138c39b6cb3cdaf7a82a17ebccadb23c3400df4c7e75c4976ba716b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 460, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 545, + 507 + ], + "type": "text", + "content": "By jointly training the adaptor " + }, + { + "bbox": [ + 305, + 460, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 460, + 545, + 507 + ], + "type": "text", + "content": " with the critic " + }, + { + "bbox": [ + 305, + 460, + 545, + 507 + ], + "type": "inline_equation", + "content": "h_c" + }, + { + "bbox": [ + 305, + 460, + 545, + 507 + ], + "type": "text", + "content": " and actor " + }, + { + "bbox": [ + 305, + 460, + 545, + 507 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 305, + 460, + 545, + 507 + ], + "type": "text", + "content": ", the information of long-term expected returns will be efficiently learned to the adaptor, resulting in a better state abstraction for the control task." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 515, + 476, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 476, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 476, + 529 + ], + "type": "text", + "content": "3.3. Topology Distilled Transferring" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": "Topology of abstracted states. Since the information of expected returns is stably captured by the momentum adaptor " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "f^m" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": ", the core issue is how to transfer that knowledge into the online heavy-optimizable main encoder " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "f^o" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": ", making its representation more adaptive to policy learning. To this end, we introduce a teacher-student distillation manner where the momentum adaptor encoder " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "f^m(\\cdot; \\theta^m, \\beta)" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": " and the online main encoder " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "f^o(\\cdot; \\theta^o)" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": " are considered as the teacher and student, respectively. Learning state abstraction for control is to project the observed states into a latent space, where the adjacency among states is consistent with the transitions in the real environment. For instance, the states near the same time-step in the real environment should also be adjacent in the feature space. Hence, we transfer the capability of state abstraction by distillation of the adjacency, i.e.," + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "210" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": "topological relation in the latent state space. To this end, we first define topological relation " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(S)" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": " in the state space " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": " as the set of similarity between every state pairs " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "(\\mathbf{s}_i,\\mathbf{s}_j)" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": " in the data," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 100, + 121, + 287, + 134 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 121, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 100, + 121, + 287, + 134 + ], + "type": "interline_equation", + "content": "\\mathcal {T} (S) = \\left\\{\\psi \\left(\\mathbf {s} _ {i}, \\mathbf {s} _ {j}\\right) \\mid \\left(\\mathbf {s} _ {i}, \\mathbf {s} _ {j}\\right) \\in S \\right\\} \\tag {7}", + "image_path": "b7d65e6ea9850efe007a76c4c7a3f42c6b443968deb6297fa019775cdb762998.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 140, + 289, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 140, + 289, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 289, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 140, + 289, + 213 + ], + "type": "inline_equation", + "content": "\\psi (\\mathbf{s}_i,\\mathbf{s}_j)" + }, + { + "bbox": [ + 46, + 140, + 289, + 213 + ], + "type": "text", + "content": " denotes the similarity function of states. This set of similarities reveals the structural information of the state space, thus preserving the representation capability of the states. Our objective is to minimize the distance between the teacher's topological relation " + }, + { + "bbox": [ + 46, + 140, + 289, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(S^m)" + }, + { + "bbox": [ + 46, + 140, + 289, + 213 + ], + "type": "text", + "content": " and the student's " + }, + { + "bbox": [ + 46, + 140, + 289, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(S^o)" + }, + { + "bbox": [ + 46, + 140, + 289, + 213 + ], + "type": "text", + "content": ", as illustrated in Fig. 4," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 221, + 292, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 292, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 292, + 269 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta^ {m *} = \\arg \\min _ {\\theta^ {m}} D _ {\\mathcal {T}} (\\mathcal {T} (S ^ {m}), \\mathcal {T} (S ^ {o})) \\\\ = \\arg \\min _ {\\theta^ {m}} \\mathbb {E} _ {\\mathbf {o} _ {i}, \\mathbf {o} _ {j} \\sim \\mathcal {B}} \\left[ \\operatorname {D i s t} \\left(\\psi^ {m} \\left(\\mathbf {s} _ {i} ^ {m}, \\mathbf {s} _ {j} ^ {m}\\right), \\psi^ {o} \\left(\\mathbf {s} _ {i} ^ {o}, \\mathbf {s} _ {j} ^ {o}\\right)\\right) \\right]. \\tag {8} \\\\ \\end{array}", + "image_path": "173a90fbd9cb2e311faf35ec5e3f7118700b182aadbe59773b7b8c7b28f8be74.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 269, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 293 + ], + "type": "inline_equation", + "content": "s_i^m = f^m (o_i|\\theta^m,\\beta), s_i^o = f^o (o_i|\\theta^o), \\psi^m (\\cdot), \\psi^o (\\cdot)" + }, + { + "bbox": [ + 46, + 269, + 287, + 293 + ], + "type": "text", + "content": " are the similarity functions among states of " + }, + { + "bbox": [ + 46, + 269, + 287, + 293 + ], + "type": "inline_equation", + "content": "S^{m}, S^{o}" + }, + { + "bbox": [ + 46, + 269, + 287, + 293 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": "Topological distillation. Here we introduce the process of topological distillation in detail. Given a batch of observations " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{o}_t\\}" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": ", we encode them to latent states " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_t^m\\}" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_t^o\\}" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": " with the momentum adaptor encoder " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "f^{m}(\\cdot ;\\theta^{m},\\beta)" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": " and the online main encoder " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "f^{o}(\\cdot ;\\theta^{o})" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": ", respectively. In this paper, we suppose the states are abstracted in a high-dimensional Euclidean space, and use " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": " distance as the similarity function " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "\\psi^m (\\cdot) = \\| \\cdot \\| ^2" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": " for the teacher. Then the topological relation within a batch is formulated as an adjacent matrix of the state features from the teacher, referring to " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^m\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": " is the batch size. " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{[i,j]}^m = \\psi^m (\\mathbf{s}_i^m,\\mathbf{s}_j^m) = \\| \\mathbf{s}_i^m -\\mathbf{s}_j^m\\| ^2" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "[i,j]" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": " denotes the i-th row and the j-th column in the matrix " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^m" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": ". Moreover, due to the unknown scale of the latent state, the matrix is normalized by its average value as " + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{A}^m\\|" + }, + { + "bbox": [ + 46, + 294, + 289, + 477 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 485, + 287, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 485, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 88, + 485, + 287, + 513 + ], + "type": "interline_equation", + "content": "\\left\\| \\mathbf {A} ^ {m} \\right\\| _ {[ i, j ]} = \\mathbf {A} _ {[ i, j ]} ^ {m} / \\left(\\frac {1}{N ^ {2}} \\sum_ {m, n} \\mathbf {A} _ {[ m, n ]} ^ {m}\\right). \\tag {9}", + "image_path": "1be6f2db7f6d11ce3aa1d642ecb72b6467f7ab03c1eab3d34317da4cb68042bd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "text", + "content": "However, directly minimizing the distance between topological relations with the same similarity function " + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\psi^o" + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\psi^m" + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "text", + "content": " will prevent the student online encoder from other basic discrimination. To flexibly distill knowledge and avoid disturbing the basic representation, the similarity function of the student's state space is defined by the " + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "text", + "content": " distance after a linear transformation " + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "text", + "content": " with weight " + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 523, + 287, + 608 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 616, + 287, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 616, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 53, + 616, + 287, + 641 + ], + "type": "interline_equation", + "content": "\\psi^ {o} \\left(\\mathbf {s} _ {i} ^ {o}, \\mathbf {s} _ {j} ^ {o}\\right) = \\| \\phi \\left(\\mathbf {s} _ {i} ^ {o}\\right) - \\phi \\left(\\mathbf {s} _ {j} ^ {o}\\right) \\| ^ {2} = \\| \\mathbf {W} ^ {T} \\mathbf {s} _ {i} ^ {o} - \\mathbf {W} ^ {T} \\mathbf {s} _ {j} ^ {o} \\| ^ {2}. \\tag {10}", + "image_path": "73d0b46d4eea827d033cf5bffcb54aadcab25fbf1b729e0c9f15b8a75fee3da6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "And the intra-batch adjacent matrix of the online encoder's state space is computed as " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^o" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{[i,j]}^o = \\psi^o (\\mathbf{s}_i^o,\\mathbf{s}_j^o)" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": ". Similar normalization as in Eq. 9 is also performed to " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^o" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": ", referring to " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{A}^o\\|" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": ". We distill the information of long-term expected returns from " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^m" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^o" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " by minimizing the L2 distance between their intra-batch similarity matrices " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^m" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 464, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 464, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 464, + 84 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^o" + }, + { + "bbox": [ + 306, + 72, + 464, + 84 + ], + "type": "text", + "content": ". The transferring loss is defined by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 91, + 545, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 91, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 328, + 91, + 545, + 122 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a n s}} = \\frac {1}{N ^ {2}} \\sum_ {(i, j) \\in [ 1, N ]} \\operatorname {D i s t} \\left(\\mathbf {A} _ {[ i, j ]} ^ {m}, \\mathbf {A} _ {[ i, j ]} ^ {o}\\right), \\tag {11}", + "image_path": "78ef1b1599833f3771c8a88945d7a79158df42cdca15544d5f6d05019fe93e4b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 130, + 545, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 130, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 304, + 130, + 545, + 166 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 130, + 545, + 166 + ], + "type": "inline_equation", + "content": "\\mathrm{Dist}(u,v) = \\max \\left((u - v)^2 -\\epsilon ,0\\right)" + }, + { + "bbox": [ + 304, + 130, + 545, + 166 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 304, + 130, + 545, + 166 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 130, + 545, + 166 + ], + "type": "text", + "content": " is a small loose factor, making the distillation learning more stable to converge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": "Dynamic prediction. Although the online encoder indirectly learns the long-term expected returns by the above distillation, its potential in representation capability is far from being developed. To this end, the online encoder is learned to capture other basic characters including environment dynamics as well as visual perception by using additional auxiliary tasks. As illustrated in Fig. 3, the predicted states " + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t^o" + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": " are conditioned on three auxiliary heads: a reward predictive head " + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": ", a state predictive head " + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": " and an observation decoder head " + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": ". Taking the current states " + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t^o" + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": " and actions " + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": " as inputs, the " + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": " aim to predict the reward " + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\mathbf{s}_t^o,a_t)" + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": " and next state " + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathbf{s}_t^o,a_t)" + }, + { + "bbox": [ + 304, + 167, + 546, + 334 + ], + "type": "text", + "content": ", and thus learn a latent state space of which dynamic is consistent with the environment. Dynamic predictive loss is defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 342, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 312, + 342, + 545, + 396 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {d y n a m i c}} = \\mathcal {L} _ {\\text {r e w a r d}} + \\mathcal {L} _ {\\text {s t a t e}} \\\\ = D _ {r} \\left(\\mathcal {R} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right), r _ {t}\\right) + D _ {s} \\left(\\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right), \\mathbf {s} _ {t + 1} ^ {o}\\right)) \\\\ = \\left(\\mathcal {R} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right) - r _ {t}\\right) ^ {2} + \\left\\| \\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right) - \\mathbf {s} _ {t + 1} ^ {o}\\right) \\| ^ {2}. \\tag {12} \\\\ \\end{array}", + "image_path": "f037ac249c12fdee561277b14a6418d60c902c6a9f6e18c25752f4aa9553d27d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "content": "To guarantee temporal capacity, the decoder " + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "content": " is trained to predict the next observation with the predicted next state " + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\hat{s}_{t + 1}^o = \\mathcal{P}(\\mathbf{s}_t^o,a_t)" + }, + { + "bbox": [ + 304, + 397, + 545, + 434 + ], + "type": "text", + "content": ". The decoding loss is defined as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 442, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {d e c} = D _ {o} \\left(\\hat {\\mathbf {o}} _ {t + 1}, \\mathbf {o} _ {t + 1}\\right) \\\\ = D _ {o} \\left(\\mathcal {G} \\left(\\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, \\mathbf {a} _ {t}\\right)\\right), \\mathbf {o} _ {t + 1}\\right) \\quad H W \\tag {13} \\\\ = \\frac {1}{H W} \\sum_ {h, w = 1} ^ {H, W} \\left(\\mathcal {G} \\left(\\mathcal {P} \\left(\\mathbf {s} _ {t} ^ {o}, a _ {t}\\right)\\right) ^ {[ h, w ]} - \\mathbf {o} _ {t + 1} ^ {[ h, w ]}\\right) ^ {2}. \\\\ \\end{array}", + "image_path": "f7dc81c14c2809fb7238d2306d06b1fdebd4c0b4cacc1c658c8e5f229ff84e57.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 515, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 539 + ], + "type": "text", + "content": "Then the overall loss of the topology distilled transferring is formulated by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 343, + 548, + 545, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 548, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 343, + 548, + 545, + 560 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e n c o d e r}} = \\mathcal {L} _ {\\text {t r a n s}} + \\mathcal {L} _ {\\text {d y n a m i c}} + \\mathcal {L} _ {\\text {o b s}}, \\tag {14}", + "image_path": "1209eae1990ac5eca63933faef353647f5eb3f73a71ce62c3dcadd1932767ca9.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 568, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 593 + ], + "type": "text", + "content": "where the online main encoder and auxiliary heads are jointly learned with the adaptor frozen." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 599, + 471, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 471, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 471, + 612 + ], + "type": "text", + "content": "3.4. Greedy Bootstrapping for AIC" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "In most deep reinforcement learning approaches, the representation module and policy module are trained in a simultaneous manner. That is, while the policy model is trained from the reinforcement signals the encoder is also simultaneously changing. This phenomenon will disturb policy improvement due to the drifting of state representation. To tackle this problem, this section presents a greedy bootstrapping framework for AIC, which consists of two" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "211" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 50, + 89, + 267, + 273 + ], + "blocks": [ + { + "bbox": [ + 54, + 76, + 236, + 87 + ], + "lines": [ + { + "bbox": [ + 54, + 76, + 236, + 87 + ], + "spans": [ + { + "bbox": [ + 54, + 76, + 236, + 87 + ], + "type": "text", + "content": "Algorithm 1: Greedy Bootstrapping for AIC" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "lines": [ + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "spans": [ + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "text", + "content": "1 Initialize the replay buffer " + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "text", + "content": " with random episodes. \n2 while not converged do \n3 // Representation learning \n4 Build data-loader with the replay buffer " + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "text", + "content": " \n5 for update step " + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "inline_equation", + "content": "c = 1\\dots C" + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "text", + "content": " do \n6 Draw next batch of transitions " + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "inline_equation", + "content": "\\{\\left(\\mathbf{o}_t,a_t,r_t,\\mathbf{o}_{t + 1}\\right)\\}" + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "text", + "content": " \n7 //Momentum Adaptor Tuning \n8 Update momentum encoder by moving average. \n9 Update adaptor by optimizing Eq. 4. \n10 //Topology Distilled Transferring \n11 Update online encoder by optimizing Eq. 14. \n12 end \n13 // Policy learning (encoder frozen) \n14 for update step " + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "inline_equation", + "content": "c_{p} = 1\\dots S" + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "text", + "content": " do \n15 Draw transition sequences " + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "inline_equation", + "content": "\\{\\left(\\mathbf{o}_t,a_t,r_t,\\mathbf{o}_{t + 1}\\right)\\} \\sim \\mathcal{B}" + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "text", + "content": " \n16 Update critic and actor parameters. \n17 end \n18 Add new experiences to the replay buffer " + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 50, + 89, + 267, + 273 + ], + "type": "text", + "content": " \n19 end" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 301, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 361 + ], + "type": "text", + "content": "alternative training stages for the representation and policy, respectively. In each representation stage, the encoder keeps training until convergence, and so does the policy in the other stage. The framework is described in detail in Algorithm 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 456 + ], + "type": "text", + "content": "Greedy representation learning. Over the latest collected replay buffer, the encoder is trained by iterating all the images in each epoch. Different from sampling a random batch in each iteration like most other methods, our ergodic sampling strategy brings stable training without ignorance of any crucial samples. In each stage of representation learning, the encoder is trained for several epochs until it is converged." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 458, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 554 + ], + "type": "text", + "content": "Greedy policy learning. With the main encoder trained sufficiently, the policy is learned continuously by freezing the encoder and conditioned on its output state. Beneficial from the stable and discriminative latent state, the critic and actor are trained efficiently from the reinforcement signal by SAC algorithms and tend to produce better behaviors. The policy is learned continuously for enough steps and then used to interact with the environment for new trajectories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 566, + 123, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 123, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 123, + 579 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 165, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 165, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 165, + 600 + ], + "type": "text", + "content": "4.1. Environments Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": "**Benchmarks.** We evaluate our method on two reinforcement learning environments with complex and realistic observations, i.e., CARLA [5] and Vizdoom [22], with continuous and discrete action space, respectively. Both tasks are developed in 3D physical scenes with a complex state space, which calls for deeper visual encoders with stronger representation capability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Implementation details. We adopt ResNet-18 as the backbone of the visual encoder. Before the last pooling layer, the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 76, + 535, + 215 + ], + "blocks": [ + { + "bbox": [ + 318, + 76, + 535, + 215 + ], + "lines": [ + { + "bbox": [ + 318, + 76, + 535, + 215 + ], + "spans": [ + { + "bbox": [ + 318, + 76, + 535, + 215 + ], + "type": "image", + "image_path": "fc868855c08398cd6ccc0507c6ed4b1f9735667953c63895640f3d37f92142de.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 223, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 305, + 223, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 545, + 245 + ], + "type": "text", + "content": "Figure 5. Comparison with other baseline approaches on the CARLA environment." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 259, + 533, + 399 + ], + "blocks": [ + { + "bbox": [ + 319, + 259, + 533, + 399 + ], + "lines": [ + { + "bbox": [ + 319, + 259, + 533, + 399 + ], + "spans": [ + { + "bbox": [ + 319, + 259, + 533, + 399 + ], + "type": "image", + "image_path": "6d661de121118efee8834a7077d2d2a7b7374bfcf17d77497586e3fa5dfdc8f1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 407, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 430 + ], + "type": "text", + "content": "Figure 6. Ablation Study of our method using ResNet-18 as the visual encoder on the CARLA environment." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 448, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 496 + ], + "type": "text", + "content": "feature map is flattened and then projected into 1024-dim state features with a linear layer and layer norm operation. The momentum factor " + }, + { + "bbox": [ + 304, + 448, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 448, + 545, + 496 + ], + "type": "text", + "content": " is set to 0.05. We use Adam to optimize the parameters of all modules in our approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 503, + 532, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 532, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 532, + 516 + ], + "type": "text", + "content": "4.2. CARLA Autonomous Driving Environment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": "Experimental settings. In the autonomous driving task, we construct a highway driving scenario with 20 other vehicles of different models using the CARLA simulator. We use five cameras on the roof of the ego-vehicle, each with a 60-degree view and obtaining images of " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "84 \\times 84" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " pixels. We concatenate them together for observation of " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "84 \\times 420" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " pixels. We stack the last three frames as the inputs. Following DBC [40], the reward is formulated as " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "r_t = v_{ego}^T \\hat{u}_h \\cdot \\Delta t - \\xi_1 \\cdot \\mathbb{I} - \\xi_2 \\cdot |steer|" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "v_{ego}" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " is the velocity vector of the ego-vehicle, projected onto the highway's unit vector " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\hat{u}_h" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": ", and multiplied by time discretization " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\Delta t = 0.05" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " to measure highway progression in meters. Impulse " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbb{I} \\in \\mathbb{R}^+" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " is caused by collisions, and a steering penalty steer " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\in [-1,1]" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " facilitates lane-keeping. The hyperparameters " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\xi_1" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\xi_2" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " and 1, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 702, + 545, + 714 + ], + "type": "text", + "content": "Comparison with other methods. We compare our ap" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "212" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 79, + 272, + 210 + ], + "blocks": [ + { + "bbox": [ + 63, + 79, + 272, + 210 + ], + "lines": [ + { + "bbox": [ + 63, + 79, + 272, + 210 + ], + "spans": [ + { + "bbox": [ + 63, + 79, + 272, + 210 + ], + "type": "image", + "image_path": "6977a11160d6966730f280a0113e56266738595547c05a1eacd92b104bfe8d68.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 223, + 287, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 245 + ], + "type": "text", + "content": "Figure 7. Comparison of other representation-decoupled methods on the CARLA environment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 287, + 447 + ], + "type": "text", + "content": "proach with seven baselines, including joint representation learning methods, i.e., CURL [27], SVEA [16], Deep-MDP [7], DBC [40]; representation decoupled methods, i.e., DreamerV2 [13], Iso-Dream [32], and the baseline SAC [10]. As shown in Fig. 5, our AIC outperforms all baselines by large margins. Compared with joint representation learning, the superior performance can be attributed to that our method stably learns a more discriminative state abstraction with deeper architectures, while avoiding the oscillating self-overfitting. Representation decoupled methods train encoders only on auxiliary tasks without TD losses and thus usually require more time to converge. However, AIC effectively learns the long-term expected returns by interaction with the light-optimizable momentum adaptor encoder, resulting in faster convergence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 700 + ], + "type": "text", + "content": "Ablation Study. We conduct ablation studies to validate the effectiveness of the proposed topology distilled transferring and the greedy bootstrapping strategy. All the experiments train the ResNet-18 as the encoder. Fig. 6 shows that SAC fails to learn an effective policy with the heavy-optimizable encoder since the state features are deteriorated by the serious oscillating self-overfitting. By applying the dynamic loss and decoder loss, the deterioration is slightly alleviated, since the encoder is encouraged to learn the inherent information of environment dynamics by the targets which are more stable than TD. However, disturbed by the oscillating gradients of the TD loss, the encoder is still hard to converge. Compared to it, the greedy bootstrapping strategy achieves significant improvement, denoted as AIC " + }, + { + "bbox": [ + 46, + 449, + 287, + 700 + ], + "type": "inline_equation", + "content": "w \\backslash o L_{trans}" + }, + { + "bbox": [ + 46, + 449, + 287, + 700 + ], + "type": "text", + "content": ", because it interrupts the backward propagation of the unstable gradient and simultaneously it trains the encoder and policy stably and sufficiently. AIC distinctly outperforms the AIC " + }, + { + "bbox": [ + 46, + 449, + 287, + 700 + ], + "type": "inline_equation", + "content": "w \\backslash o L_{trans}" + }, + { + "bbox": [ + 46, + 449, + 287, + 700 + ], + "type": "text", + "content": " due to the topology distilled transferring, which additionally incorporates the long-term expected return information into the encoder, making it more adaptive for the downstream controlling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Comparison with other decoupling methods. To further" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 79, + 531, + 212 + ], + "blocks": [ + { + "bbox": [ + 322, + 79, + 531, + 212 + ], + "lines": [ + { + "bbox": [ + 322, + 79, + 531, + 212 + ], + "spans": [ + { + "bbox": [ + 322, + 79, + 531, + 212 + ], + "type": "image", + "image_path": "643cda3bc755d17d8704662373cdb1563fbc35fa1b1e0134cba03c931b239ce3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 223, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 306, + 223, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 306, + 223, + 545, + 245 + ], + "type": "text", + "content": "Figure 8. Comparison with different architectures of the encoder on the CARLA environment." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 321, + 263, + 531, + 396 + ], + "blocks": [ + { + "bbox": [ + 321, + 263, + 531, + 396 + ], + "lines": [ + { + "bbox": [ + 321, + 263, + 531, + 396 + ], + "spans": [ + { + "bbox": [ + 321, + 263, + 531, + 396 + ], + "type": "image", + "image_path": "f09ebae0686f1312673628b57fdb906cb41a840eb90460e55cb1d330c3477370.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 408, + 544, + 430 + ], + "lines": [ + { + "bbox": [ + 306, + 408, + 544, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 544, + 430 + ], + "type": "text", + "content": "Figure 9. Comparison of other stabilizing methods using ResNet-18 on the CARLA environment." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 450, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 545, + 713 + ], + "type": "text", + "content": "evaluate the greedy bootstrapping framework, we compare it with other two typical methods that decouple representation learning from RL. Initial decoupling first collects a set of transitions using a random policy and then pretrains the encoder with auxiliary tasks as in Eq. 12 and 13. Afterward, the policy is trained by SAC with the encoder frozen. As shown in Fig. 7, the initial decoupling method performs high-variance results because the performance significantly depends on whether the distribution of the collected pretraining data is similar to those from a learned policy. Stepping decoupling learns the encoder and the policy step by step, respectively. The encoder is trained by transitions from the same data distribution as the policy. However, since the encoder keeps updating during policy learning, the latent state space suffers from unstable drifting, making it difficult to learn policy from the volatile state abstraction. The greedy bootstrapping strategy surpasses both two approaches because it guarantees a determined latent state space that is conducive to the inductive learning of policy. Moreover, it learns the encoder by traversing all the transitions in the replay buffer, prevented from ignorance of some critical images or overfitting to parts of data." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "213" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 80, + 518, + 179 + ], + "blocks": [ + { + "bbox": [ + 122, + 69, + 470, + 80 + ], + "lines": [ + { + "bbox": [ + 122, + 69, + 470, + 80 + ], + "spans": [ + { + "bbox": [ + 122, + 69, + 470, + 80 + ], + "type": "text", + "content": "Table 1. Comparison with other methods using ResNet-18 on Medikit of Vizdoom environment." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 80, + 518, + 179 + ], + "lines": [ + { + "bbox": [ + 76, + 80, + 518, + 179 + ], + "spans": [ + { + "bbox": [ + 76, + 80, + 518, + 179 + ], + "type": "table", + "html": "
-DQN (Baseline)DrQSVEACURLDeepMDPAIC (Ours)
Episode Returns (mean/std)888±5841361 ± 1004435±185408±2031026±8071686±1283
Min returns252252792428456
Max returns244442001008113637104800
Survival Time (mean/std)64 ±2685±4345±844±972±3796±53
Min Time262627233926
Max Time1352106874191210
", + "image_path": "80a806827e13993b0ce0ce2473d85965cb39a46bbed939721f59dc1ab49c045d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 289, + 472 + ], + "type": "text", + "content": "Improvements on different encoders. We evaluate our approach with different architectures as the visual encoder, including a light CNN of 4-layer convolution (Conv-4), and three heavyweight models, i.e., VGG-11, ResNet-18, and ResNet34. For each architecture, we conduct three experiments using SAC, SAC with auxiliary tasks (dynamic loss and decoder loss as in Eq. 12 and 13), and our AIC approach. The comparison is illustrated in Fig. 8. Firstly, all heavyweight models fail to learn effective policies using SAC, where only TD loss is used. The results indicate the oscillating self-overfitting is consistent for heavy models. Secondly, auxiliary tasks improve the performance of SAC with heavyweight encoders, but they are still inferior to the light encoder like Conv-4. The oscillating self-overfitting seriously prevents the encoder from sufficient learning. Finally, AIC consistently improves the performance with all heavy encoders by large margins, as well as the light encoder of Conv-4. It shows a consistent rank of performance that these architectures have achieved in supervised learning tasks. It demonstrates that our method makes full use of the representation capability of heavyweight visual encoders through interactive cooperation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 473, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 568 + ], + "type": "text", + "content": "Comparison with stabilizing methods. To evaluate AIC for stabilizing the training of heavy-optimizable encoders, we compare it with SVEA and DrQ using ResNet-18 as the backbone. As shown in Fig. 9, both two methods perform not well, since the increase of parameters significantly exacerbates the challenge of stabilizing. However, AIC exceeds them by large margins, indicating its capability to stably learn heavy models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Qualitative analysis. In order to evaluate the effectiveness of our approach in alleviating the oscillating self-overfitting, we visualize the updating trajectories of the encoder parameters over a set of training steps by TSNE. We use ResNet-18 as the encoder backbone and choose the weights of the last convolution layer for visualization. The comparison between DeepMDP with the TD loss and AIC is shown in Fig. 10, where each point represents the parameters in a step and the color changes from deep to light over the training steps. As we can see, the parameters with the TD loss of the DeepMDP agent suffer from significant oscillation during training. Different from that, the parameters of AIC show a dis" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 209, + 423, + 337 + ], + "blocks": [ + { + "bbox": [ + 310, + 209, + 423, + 337 + ], + "lines": [ + { + "bbox": [ + 310, + 209, + 423, + 337 + ], + "spans": [ + { + "bbox": [ + 310, + 209, + 423, + 337 + ], + "type": "image", + "image_path": "92a2dc787f62f3400c0431aa8810243c3af0590866542fd6caf95b92f9762793.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 341, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 305, + 341, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 341, + 545, + 364 + ], + "type": "text", + "content": "Figure 10. Parameter updating trajectories of the ResNet-18 encoder over training steps." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 425, + 209, + 542, + 337 + ], + "blocks": [ + { + "bbox": [ + 425, + 209, + 542, + 337 + ], + "lines": [ + { + "bbox": [ + 425, + 209, + 542, + 337 + ], + "spans": [ + { + "bbox": [ + 425, + 209, + 542, + 337 + ], + "type": "image", + "image_path": "4e5972fe737d309c73cc5428bc929eda398f47e27d6794ca1ee8ff3f0d72b6e6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 384, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 433 + ], + "type": "text", + "content": "tinct optimizing direction and only have slight oscillation in a short time. The qualitative results demonstrate AIC effectively alleviates the oscillating self-overfitting for the heavy-optimizable encoder in visual reinforcement learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 442, + 474, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 474, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 474, + 453 + ], + "type": "text", + "content": "4.3. Vizdoom Medikit Environment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "text", + "content": "Experimental settings. Vizdoom is based on the first-person perspective video game in a semi-realistic 3D world. We use the difficult medikit collecting scenario, where the agent is spawned in a random spot of a maze with an acid surface, which slowly, but constantly, takes away the agent's life. The agent needs to collect medikits and avoid blue vials with poison, both of which appear in random places during the episode. In each step, the agent is allowed to move (forward/backward), or turn (left/right). The reward is formulated as " + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "inline_equation", + "content": "r_t = 1 - \\xi_3 \\cdot E_m - \\xi_4 \\cdot E_v - \\xi_5 \\cdot E_d" + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "inline_equation", + "content": "E_m" + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "inline_equation", + "content": "E_v" + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "inline_equation", + "content": "E_d" + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "text", + "content": " denote the events of getting a medikit, a vial and to death, respectively. " + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\xi_3 = \\xi_4 = \\xi_5 = 100" + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": "Superior performance with heavy encoders. To evaluate our approach, especially with deep deep encoders, we compare it with five methods. All of them use ResNet-18 as backbones. Since the Vizdoom is of discrete action space, we adopt DQN as the baseline. As shown in Table 1, AIC performs superiorly to the compared methods. CURL and SVEA perform poorly, since the observations of such a fine-grained collecting task of first-person perspective are sensitive to random cropping of CURL. And the random conv of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "214" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": "SVEA is not applicable for the low-contrast images in this scenario. However, AIC is able to effectively train heavy models under different conditions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 118, + 119, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 118, + 119, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 119, + 131 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 138, + 289, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 138, + 289, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 289, + 281 + ], + "type": "text", + "content": "In this paper, we investigate the phenomenon of oscillating self-overfitting which results in the deterioration of RL when training heavy-optimizable encoders. To alleviate this problem, we propose a novel asymmetric interactive cooperation approach. Through interaction between a main heavy-optimizable encoder and a supportive light-optimizable encoder, both advantages of them are integrated including the highly discriminative capability and the training stability. We further present a greedy bootstrapping framework for further stability. Experimental results of CARLA and Vizdoom show that AIC achieves competitive performance over complex and realistic environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 292, + 149, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 149, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 149, + 306 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 312, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 395 + ], + "type": "text", + "content": "This work is partially supported by grants from the Key-Area Research and Development Program of Guangdong Province under contact No. 2020B0101380001 and grants from the National Natural Science Foundation of China under contract No. 62027804, No. 61825101, No. 62088102 and No. 62202010. The computing resources of Pengcheng Cloudbrain are used in this research." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 405, + 106, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 106, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 106, + 417 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 425, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "text", + "content": "[1] Homanga Bharadhwaj, Mohammad Babaeizadeh, Dumitru Erhan, and Sergey Levine. Information prioritization through empowerment in visual model-based rl. arXiv preprint arXiv:2204.08585, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 469, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 514 + ], + "type": "text", + "content": "[2] Pablo Samuel Castro. Scalable methods for computing state similarity in deterministic markov decision processes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10069-10076, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 514, + 286, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 286, + 546 + ], + "type": "text", + "content": "[3] Edoardo Cetin, Philip J Ball, Steve Roberts, and Oya Celiktutan. Stabilizing off-policy deep reinforcement learning from pixels. arXiv preprint arXiv:2207.00986, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 547, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 547, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 547, + 287, + 602 + ], + "type": "text", + "content": "[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 602, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 287, + 645 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "text", + "content": "[6] Chelsea Finn and Sergey Levine. Deep visual foresight for planning robot motion. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 2786-2793. IEEE, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "text", + "content": "[7] Carles Gelada, Saurabh Kumar, Jacob Buckman, Ofir Nachum, and Marc G Bellemare. Deepmdp: Learning" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "continuous latent space models for representation learning. In International Conference on Machine Learning, pages 2170-2179. PMLR, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 311, + 107, + 545, + 161 + ], + "type": "text", + "content": "[8] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 163, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 163, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 311, + 163, + 545, + 185 + ], + "type": "text", + "content": "[9] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 239 + ], + "type": "text", + "content": "[10] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[11] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[12] Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International conference on machine learning, pages 2555-2565. PMLR, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "text", + "content": "[13] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[14] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[15] Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. arXiv preprint arXiv:2007.04309, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 477 + ], + "type": "text", + "content": "[16] Nicklas Hansen, Hao Su, and Xiaolong Wang. Stabilizing deep q-learning with convnets and vision transformers under data augmentation. 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[17] Nicklas Hansen and Xiaolong Wang. Generalization in reinforcement learning by soft data augmentation. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13611-13617. IEEE, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "text", + "content": "[19] Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. arXiv preprint arXiv:1611.05397, 2016." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "text", + "content": "[20] Michael Janner, Qiyang Li, and Sergey Levine. Offline reinforcement learning as one big sequence modeling problem. Advances in neural information processing systems, 34:1273-1286, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[21] Lukasz Kaiser, Mohammad Babaeizadeh, Piotr Milos, Blazej Osinski, Roy H Campbell, Konrad Czechowski, Dumitru Erhan, Chelsea Finn, Piotr Kozakowski, Sergey Levine, et al. Model-based reinforcement learning for atari. arXiv preprint arXiv:1903.00374, 2019." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "215" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[22] Michal Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Vizdoom: A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE conference on computational intelligence and games (CIG), pages 1-8. IEEE, 2016." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "type": "text", + "content": "[23] Ilya Kostrikov, Denis Yarats, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. arXiv preprint arXiv:2004.13649, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 195 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "text", + "content": "[25] Aviral Kumar, Rishabh Agarwal, Dibya Ghosh, and Sergey Levine. Implicit under-parameterization inhibits data-efficient deep reinforcement learning. arXiv preprint arXiv:2010.14498, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "text", + "content": "[26] Misha Laskin, Kimin Lee, Adam Stooke, Lerrel Pinto, Pieter Abbeel, and Aravind Srinivas. Reinforcement learning with augmented data. Advances in neural information processing systems, 33:19884-19895, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 327 + ], + "type": "text", + "content": "[27] Michael Laskin, Aravind Srinivas, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In International Conference on Machine Learning, pages 5639-5650. PMLR, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 393 + ], + "type": "text", + "content": "[28] Alex X. Lee, Anusha Nagabandi, Pieter Abbeel, and Sergey Levine. Stochastic latent actor-critic: Deep reinforcement learning with a latent variable model. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 741-752. Curran Associates, Inc., 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[29] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. The Journal of Machine Learning Research, 17(1):1334-1373, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[30] Masashi Okada and Tadahiro Taniguchi. Dreaming: Model-based reinforcement learning by latent imagination without reconstruction. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4209-4215. IEEE, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "type": "text", + "content": "[31] Kei Ota, Devesh K Jha, and Asako Kanezaki. Training larger networks for deep reinforcement learning. arXiv preprint arXiv:2102.07920, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 570 + ], + "type": "text", + "content": "[32] Minting Pan, Xiangming Zhu, Yunbo Wang, and Xiaokang Yang. Iso-dream: Isolating and leveraging noncontrollable visual dynamics in world models. In Advances in Neural Information Processing Systems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 571, + 287, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 614 + ], + "type": "text", + "content": "[33] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning, pages 1278-1286. PMLR, 2014." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[34] Ramanan Sekar, Oleh Rybkin, Kostas Daniilidis, Pieter Abbeel, Danijar Hafner, and Deepak Pathak. Planning to explore via self-supervised world models. In International Conference on Machine Learning, pages 8583-8592. PMLR, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[35] Evan Shelhamer, Parsa Mahmoudieh, Max Argus, and Trevor Darrell. Loss is its own reward: Self-supervision for reinforcement learning. arXiv preprint arXiv:1612.07307, 2016." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 329 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 137 + ], + "type": "text", + "content": "[36] Oriol Vinyals, Igor Babuschkin, Wojciech M Czarnecki, Michael Mathieu, Andrew Dudzik, Junyoung Chung, David H Choi, Richard Powell, Timo Ewalds, Petko Georgiev, et al. Grandmaster level in starcraft ii using multiagent reinforcement learning. Nature, 575(7782):350-354, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 139, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 139, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 546, + 184 + ], + "type": "text", + "content": "[37] Manuel Watter, Jost Springenberg, Joschka Boedecker, and Martin Riedmiller. Embed to control: A locally linear latent dynamics model for control from raw images. Advances in neural information processing systems, 28, 2015." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 239 + ], + "type": "text", + "content": "[38] Denis Yarats, Amy Zhang, Ilya Kostrikov, Brandon Amos, Joelle Pineau, and Rob Fergus. Improving sample efficiency in model-free reinforcement learning from images. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10674-10681, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 240, + 546, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 546, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 546, + 283 + ], + "type": "text", + "content": "[39] Tao Yu, Zhizheng Zhang, Cuiling Lan, Yan Lu, and Zhibo Chen. Mask-based latent reconstruction for reinforcement learning. In Advances in Neural Information Processing Systems." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 285, + 546, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 546, + 329 + ], + "type": "text", + "content": "[40] Amy Zhang, Rowan McAllister, Roberto Calandra, Yarin Gal, and Sergey Levine. Learning invariant representations for reinforcement learning without reconstruction. arXiv preprint arXiv:2006.10742, 2020." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "216" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Stable Cluster Discrimination for Deep Clustering/9f881270-2ff5-4f87-a2db-8feef1f47071_content_list.json b/2023/Stable Cluster Discrimination for Deep Clustering/9f881270-2ff5-4f87-a2db-8feef1f47071_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..32a0ddedc6ebe0e2f43253104645e0ec5decbfea --- /dev/null +++ b/2023/Stable Cluster Discrimination for Deep Clustering/9f881270-2ff5-4f87-a2db-8feef1f47071_content_list.json @@ -0,0 +1,2035 @@ +[ + { + "type": "text", + "text": "Stable Cluster Discrimination for Deep Clustering", + "text_level": 1, + "bbox": [ + 230, + 130, + 740, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Qian \nAlibaba Group, Bellevue, WA 98004, USA \nqi.qian@alibaba-inc.com", + "bbox": [ + 313, + 181, + 656, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep clustering can optimize representations of instances (i.e., representation learning) and explore the inherent data distribution (i.e., clustering) simultaneously, which demonstrates a superior performance over conventional clustering methods with given features. However, the coupled objective implies a trivial solution that all instances collapse to the uniform features. To tackle the challenge, a two-stage training strategy is developed for decoupling, where it introduces an additional pre-training stage for representation learning and then fine-tunes the obtained model for clustering. Meanwhile, one-stage methods are developed mainly for representation learning rather than clustering, where various constraints for cluster assignments are designed to avoid collapsing explicitly. Despite the success of these methods, an appropriate learning objective tailored for deep clustering has not been investigated sufficiently. In this work, we first show that the prevalent discrimination task in supervised learning is unstable for one-stage clustering due to the lack of ground-truth labels and positive instances for certain clusters in each mini-batch. To mitigate the issue, a novel stable cluster discrimination (SeCu) task is proposed and a new hardness-aware clustering criterion can be obtained accordingly. Moreover, a global entropy constraint for cluster assignments is studied with efficient optimization. Extensive experiments are conducted on benchmark data sets and ImageNet. SeCu achieves state-of-the-art performance on all of them, which demonstrates the effectiveness of one-stage deep clustering.", + "bbox": [ + 75, + 301, + 472, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clustering is a fundamental task in unsupervised learning. Given features of instances, the unlabeled data set will be partitioned into multiple clusters, where instances from the same cluster are similar according to the measurement defined by a distance function. With fixed features, most of research efforts focus on studying appropriate distance functions and ingenious algorithms have been proposed by different measurements, e.g., k-means clustering [24], spec", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d4651c447ce90e196cf23908d71313dffa6dc0b8fd10b8d4047a7253127216a.jpg", + "image_caption": [ + "Figure 1: An illustration of the proposed method. 10 data points are randomly sampled from two different Gaussian distributions, respectively. Points with the same color are from the same distribution, while squares denote the corresponding cluster centers obtained by different methods. Star indicates the mis-classified data. Unlike k-means that assigns the uniform weight for different instances, our method considers the hardness of instances and is better for discrimination based clustering." + ], + "image_footnote": [], + "bbox": [ + 517, + 270, + 686, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c1fa25d5bce1822d8cfe9a6b49da87f78229c6e3b50478b87e1a82cd3551977.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 268, + 870, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tral clustering [28], subspace clustering [13], etc.", + "bbox": [ + 498, + 564, + 823, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the development of deep neural networks, deep learning is capable of learning representations from raw materials and demonstrates the dominating performance on various supervised tasks [22]. Thereafter, it is also introduced to clustering recently [4, 11, 14, 19, 20, 23, 29, 30, 32]. Unlike the conventional clustering, representations of instances are learned with cluster assignments and centers simultaneously in deep clustering, which is more flexible to capture the data distribution. However, the coupled objective can result in a trivial solution that all instances collapse to the uniform features [4] and designing appropriate clustering criterion in the new scenario becomes challenging.", + "bbox": [ + 496, + 580, + 890, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many deep clustering algorithms [11, 14, 32] exploit a two-stage training strategy to decouple representation learning and clustering to avoid collapsing. The recent progress in unsupervised representation learning [2, 4, 7, 16, 17, 26] shows that informative features capturing the semantic similarity between instances without collapsing can be obtained by pre-training a large set of unlabeled data. Inspired by the observation, those methods leverage a pre-training stage and then focus on developing algorithms to optimize clus", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16645", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tering by fine-tuning a pre-trained model in the second stage. By refining the relations between nearest neighbors obtained from the pre-training stage, two-stage methods achieve a significantly better performance than one-stage ones without sufficient representation learning [19, 20].", + "bbox": [ + 78, + 90, + 467, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the objective of pre-training can be inconsistent with that of clustering, which results in a sub-optimal performance for deep clustering. Note that different from supervised learning where the objective is explicitly defined by labels, that for unsupervised representation learning is arbitrary and various pretext tasks have been proposed, e.g., instance discrimination [7, 17], cluster discrimination [4, 26], masked modeling [2, 16], etc. Most of existing two-stage clustering methods adopt instance discrimination, i.e., SimCLR [7], for pre-training, whereas it aims to identify each instance as an individual class and the objective is different from clustering that aims to group instances from the same cluster. Moreover, SwAV [5] demonstrates that clustering itself is also an effective pretext task for representation learning. Therefore, we focus on facilitating one-stage deep clustering that optimizes representations and clustering simultaneously in this work.", + "bbox": [ + 78, + 169, + 467, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most of existing one-stage methods are proposed solely for representation learning. To tackle the collapsing issue, research efforts are mainly devoted to developing appropriate constraints for cluster assignments, especially for online deep clustering. For example, [1, 5] apply the balanced constraint that each cluster has the same number of instances for clustering. [26] further relaxes the balanced constraint to a lower-bound size constraint that limits the minimal size of clusters and demonstrates a more flexible assignment.", + "bbox": [ + 78, + 428, + 467, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After obtaining cluster assignments as pseudo labels, representations and cluster centers can be optimized as in the supervised scenario. For example, [5, 6] learn the encoder network and cluster centers by solving a classification problem with the standard cross entropy loss. [26] has the same classification task for representation learning while adopting k-means for updating cluster centers. Although these methods achieve a satisfied performance on representation learning, a learning objective tailored for deep clustering has not attracted sufficient attentions.", + "bbox": [ + 78, + 566, + 467, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we investigate the effective learning task for one-stage deep clustering. By analyzing the standard cross entropy loss for supervised learning, we find that it can be unstable for unsupervised learning. Concretely, the gradient for updating cluster centers consists of two ingredients: gradient from positive instances of the target cluster and that from irrelevant negative ones. However, with a limited mini-batch size in stochastic gradient descent (SGD), there can be no positive ones for a large proportion of clusters (e.g., $90\\%$ on ImageNet) at each iteration. Due to the lack of positive instances, the influence from negative ones is dominating. Unlike supervised learning where labels are fixed,", + "bbox": [ + 78, + 719, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the cluster assignments can change during training in deep clustering. Therefore, the noise from the large variance of negative instances will be accumulated, which makes the optimization unstable.", + "bbox": [ + 503, + 90, + 890, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the problem, we propose a stable cluster discrimination (SeCu) task for deep clustering, which stops the gradient from negative instances for updating cluster centers in the cross entropy loss. Compared with k-means in [26], where positive instances have the uniform weight for updating centers, SeCu as a discrimination task considers the hardness of instances and a large weight will be assigned to hard instances for updating. Fig. 1 illustrates the hardness-aware clustering criterion implied by SeCu. Besides, we improve the cluster assignment by developing an entropy constraint that regularizes the entropy of assignments over the entire data set. Compared to the optimization with the size constraint [26], our method can reduce the number of variables and hyper-parameters, and thus make the learning more convenient. The main contributions of this work can be summarized as follows.", + "bbox": [ + 503, + 152, + 890, + 392 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel task is proposed for one-stage deep clustering. SeCu tailors the supervised cross entropy loss by eliminating the influence from the negative instances in learning cluster centers, which makes the training stable when ground-truth labels are unavailable.", + "- A global entropy constraint is exploited to balance the size of clusters and an efficient closed-form solution is developed for online assignment with the constraint.", + "- A simple framework with a single loss function and encoder is introduced for deep clustering in Eqn. 6. The proposed method is evaluated with the standard protocol on benchmark data sets and ImageNet [27]. The superior performance of SeCu confirms its effectiveness for deep clustering." + ], + "bbox": [ + 521, + 406, + 890, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 657, + 638, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we briefly review unsupervised representation learning and then focus on clustering.", + "bbox": [ + 503, + 684, + 888, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Unsupervised Representation Learning", + "text_level": 1, + "bbox": [ + 503, + 726, + 834, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By leveraging massive unlabeled data, unsupervised representation learning can learn a pre-trained model that encoders the semantic information from data and helps downstream tasks with fine-tuning. Compared with supervised learning containing labels, one major challenge in unsupervised learning is defining appropriate positive pairs for representation learning. SimCLR [7] and MoCo [17] apply instance discrimination as the objective that generates positive pairs from diverse views of the same instance and considers other instances as the negative ones. Then, BYOL [15]", + "bbox": [ + 503, + 750, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16646", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and SimSiam [8] demonstrate that with appropriate neural network architectures, the applicable models can be obtained with only positive pairs. Besides the work optimizing instance space, Barlow Twins [31] defines positive pairs in feature space and also learns effective pre-trained models. Recently, learning with masked modeling [2, 16] achieves success on the specific backbone of vision transformer [12] and shows the better fine-tuning performance than the generic objective on the ResNet [18].", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Deep Clustering", + "text_level": 1, + "bbox": [ + 76, + 237, + 238, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared with instance discrimination, cluster discrimination aims to obtain positive pairs consisting of different instances for representation learning. Therefore, it has to learn representations and relations between instances simultaneously. DeepCluster [4] obtains membership of instances with an offline k-means and then optimizes representations alternately. Some work focuses on representation learning and introduces the balanced constraint that each cluster has the same number of instances for cluster assignments [1, 5] to mitigate the collapsing problem. Besides, DINO [6] applies an additional momentum encoder to further improve the representation. However, the constraint assumes a well balanced distribution, which is hard to capture the ground-truth data distribution. Thereafter, CoKe [26] relaxes the constraint to only lower-bound the minimal size of clusters to model the distribution more flexibly.", + "bbox": [ + 75, + 260, + 472, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides representation learning, there are many works developed solely for better clustering [11, 14, 20, 23, 29, 30, 32]. To handle the collapsing issue, two-stage methods [11, 14, 32] leverage the representations from a pretrained model to obtain nearest neighbors and fine-tune the model for clustering accordingly. In addition, the objective of pre-training can be included for clustering as multitask learning [23]. The work closest to ours is CoKe [26] that optimizes representations and clustering simultaneously. Compared with the conventional k-means in CoKe, we propose a novel stable cluster discrimination objective tailored for one-stage deep clustering. Moreover, an entropy constraint is introduced to reduce the number of parameters while demonstrating the competitive performance.", + "bbox": [ + 75, + 503, + 472, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Stable Cluster Discrimination", + "text_level": 1, + "bbox": [ + 76, + 729, + 352, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Cluster Discrimination", + "text_level": 1, + "bbox": [ + 76, + 753, + 290, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a data set with $N$ images $\\{x_{i}\\}_{i = 1}^{N}$ , if the corresponding labels are accessible as $\\{y_{i}\\}_{i = 1}^{N}$ , the literature from distance metric learning [25] shows that appropriate representations of examples can be learned by optimizing a classification task as", + "bbox": [ + 75, + 776, + 470, + 852 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\sum_ {i = 1} ^ {N} \\ell \\left(x _ {i}, y _ {i}; \\theta\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 862, + 339, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\ell (\\cdot)$ is the cross entropy loss with the normalized Softmax operator as suggested in [25]. $\\theta$ denotes the parameters of the deep neural network.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For unsupervised learning that the label information is unavailable, the objective of cluster discrimination with $K$ clusters can be written as", + "bbox": [ + 496, + 137, + 893, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {f}, \\left\\{\\mathbf {w} _ {j} \\right\\}, y _ {i} \\in \\Delta} \\mathcal {L} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} - y _ {i, j} \\log \\left(p _ {i, j}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 191, + 890, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y_{i}$ denotes the learnable label for $x_{i}$ and $\\Delta = \\{y_i|\\sum_{j = 1}^{K}y_{i,j} = 1,\\forall j,y_{i,j}\\in \\{0,1\\} \\}$ . With the normalized Softmax operator, the prediction $p_{i,j}$ can be computed", + "bbox": [ + 500, + 247, + 893, + 292 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, j} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {j} / \\lambda\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {k} / \\lambda\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 305, + 890, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_i = f(x_i)$ and $f(\\cdot)$ denotes the encoder network. $\\theta_{f}$ contains parameters of $f$ . $\\{\\mathbf{w}_j\\}_{j=1}^K$ consists of cluster centers. $\\lambda$ is the parameter of temperature and $\\forall i, j, \\|\\mathbf{x}_i\\|_2 = \\|\\mathbf{w}_j\\|_2 = 1$ by normalization.", + "bbox": [ + 496, + 354, + 890, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared with the supervised counterpart, the problem in Eqn. 1 has to optimize cluster assignments $\\{y\\}$ , cluster centers $\\{\\mathbf{w}\\}$ , and representation encoder network $f$ simultaneously. While most of existing work focuses on optimizing $\\{y\\}$ with different constraints to avoid collapsing, this work investigates the learning objective for $\\{\\mathbf{w}\\}$ and $f$ , and a new clustering criterion is introduced accordingly.", + "bbox": [ + 496, + 416, + 890, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Stable Loss for Optimization with Mini-batch", + "text_level": 1, + "bbox": [ + 498, + 532, + 883, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike supervised learning, where labels are fixed for examples, cluster assignments $\\{y\\}$ in unsupervised learning are dynamic with the training of instance representations and cluster centers. Therefore, the original cross entropy loss becomes unstable for unsupervised discrimination. The problem can be elaborated by analyzing the updating criterion for cluster centers.", + "bbox": [ + 496, + 556, + 890, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Letting $y_{i}$ denote the label of $\\mathbf{x}_i$ , the gradient of $\\mathbf{w}_j$ from the standard cross entropy loss in Eqn. 1 can be computed as", + "bbox": [ + 496, + 662, + 890, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {w} _ {j}} \\mathcal {L} = \\frac {1}{\\lambda} \\left(\\sum_ {i: y _ {i} = j} \\left(p _ {i, j} - 1\\right) \\mathbf {x} _ {i} + \\sum_ {k: y _ {k} \\neq j} p _ {k, j} \\mathbf {x} _ {k}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 718, + 854, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The former term is to pull the center $\\mathbf{w}_j$ to the assigned instances and the latter one is to push it away from instances of other clusters. However, the following Propositions show that this updating is unstable for deep clustering. The detailed proof can be found in the supplementary.", + "bbox": [ + 496, + 767, + 890, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. When sampling a mini-batch of $b$ instances for $K$ clusters, there are no positive instances for at least $K - b$ clusters.", + "bbox": [ + 496, + 854, + 893, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16647", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 2. Let $\\mathrm{Var}_{pos}$ and $\\mathrm{Var}_{neg}$ be the variance of sampled positive and negative instances, respectively. Assuming that each instance has unit norm and the norm of the cluster mean is $a$ , we have $\\mathrm{Var}_{neg} = \\mathcal{O}\\left(\\frac{1}{1 - a^2}\\right)\\mathrm{Var}_{pos}$ .", + "bbox": [ + 76, + 90, + 472, + 154 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark Proposition 1 implies that with cross entropy loss, a large proportion of cluster centers will be updated only by negative instances. Proposition 2 indicates that the variance of sampled negative instances is much larger than that of positive ones when each cluster is compact as $a$ approaching 1. Due to the small size of a mini-batch in training deep neural networks, the variance cannot be reduced sufficiently.", + "bbox": [ + 75, + 169, + 468, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When $K = 10,000$ and $b = 1,024$ as in our multiclustering settings for ImageNet, at least about $90\\%$ centers will only access a mini-batch of negative instances at each iteration. Without ground-truth labels, the bias in updating will be accumulated and mislead the learning process.", + "bbox": [ + 76, + 289, + 468, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate the problem, we propose to eliminate the direction of negative instances from gradient for stable training as", + "bbox": [ + 76, + 364, + 468, + 410 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {w} _ {j}} \\mathcal {L} = \\frac {1}{\\lambda} \\sum_ {i: y _ {i} = j} (p _ {i, j} - 1) \\mathbf {x} _ {i}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 409, + 372, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The corresponding stable cluster discrimination loss becomes", + "bbox": [ + 76, + 449, + 468, + 478 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\ell_ {\\mathrm {S e C u}} \\left(x _ {i}, y _ {i}\\right) = \\\\ - \\log \\left(\\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right)}{\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right) + \\sum_ {k : k \\neq y _ {i}} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\tilde {\\mathbf {w}} _ {k} / \\lambda\\right)}\\right) \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 488, + 468, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{w}}_k$ denotes $\\mathbf{w}_k$ with the stop-gradient operator. Compared with the standard cross entropy loss, cluster centers in SeCu loss will be updated only by positive instances, which is more stable for deep clustering when optimizing with mini-batches.", + "bbox": [ + 76, + 559, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, the analysis in the following theorem demonstrates the novel hardness-aware clustering criterion for deep clustering implied by the proposed objective.", + "bbox": [ + 76, + 633, + 468, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. When fixing $\\{y_i\\}$ and $\\{\\mathbf{x}_i\\}$ , let $\\{\\mathbf{w}^*\\}$ be the optimal solution of the problem with loss function in Eqn. 3 and assume $\\forall i, \\| \\mathbf{x}_i\\|_2 = 1; \\forall j, \\| \\mathbf{w}_j\\|_2 = 1$ , then we have", + "bbox": [ + 76, + 686, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {j} ^ {*} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} \\left(\\frac {\\sum_ {i : y _ {i} = j} (1 - p _ {i , j}) \\mathbf {x} _ {i}}{\\sum_ {i : y _ {i} = j} 1 - p _ {i , j}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 739, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Pi_{\\| \\mathbf{w}\\| _2 = 1}$ projects the vector to the unit norm.", + "bbox": [ + 76, + 784, + 421, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark For k-means in CoKe [26], cluster centers will be averaged over all assigned instances with the uniform weight", + "bbox": [ + 76, + 816, + 468, + 861 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {j} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} \\left(\\frac {\\sum_ {i : y _ {i} = j} \\mathbf {x} _ {i}}{\\sum_ {i} 1 (y _ {i} = j)}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 868, + 468, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $1(\\cdot)$ is the indicator function. On the contrary, our proposal considers the weight of each instance by its hardness, i.e., $p_{i,y_i}$ . By assigning a large weight to hard instances, the corresponding center can capture the distribution of hard instances better, which is illustrated in Fig. 1.", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our formulation also implies CoKe as a special case. Concretely, by increasing the temperature $\\lambda$ for updating $\\mathbf{w}$ to be infinite, we have $p_{i,j} = 1 / K$ and Eqn. 4 will degenerate to Eqn. 5.", + "bbox": [ + 496, + 167, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, besides SGD, Eqn. 4 also suggests an alternative updating strategy for $\\mathbf{w}$ . Since $p_{i,j}$ is computed by $\\mathbf{w}$ , it may still require multiple iterations with Eqn. 4 for convergence. However, the representations of instances are also updated by training and we can keep a single update for centers at each iteration and improve centers along with the learning of representations. Compared with SGD for learning cluster centers, the strategy eliminates the learning rate for centers with a comparable performance as shown in the ablation study.", + "bbox": [ + 496, + 228, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the learning objective SeCu, we will elaborate the proposed deep clustering method in next subsection.", + "bbox": [ + 496, + 378, + 890, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. SeCu for Deep Clustering", + "text_level": 1, + "bbox": [ + 500, + 420, + 733, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the proposed loss function, the objective of stable cluster discrimination for deep clustering can be written as", + "bbox": [ + 496, + 444, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {f}, \\left\\{\\mathbf {w} _ {j} \\right\\}, y _ {i} \\in \\Delta} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} \\left(x _ {i}, y _ {i}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 500, + 772, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad h _ {m} (Y) \\geq b _ {m}, \\quad m = 1, \\dots , M \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 542, + 890, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Y = [y_{1},\\ldots ,y_{N}]$ , $M$ denotes the number of constraints, and $h_m(\\cdot)$ is the $m$ -th constraint for cluster assignment. Considering that variables are entangled, the problem is solved alternately.", + "bbox": [ + 500, + 571, + 890, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Update of $\\theta_{f}$ First, when fixing $\\{y^{t - 1}\\}$ and $\\{\\mathbf{w}^{t - 1}\\}$ from the last epoch, the sub-problem for representation learning in the $t$ -th epoch becomes", + "bbox": [ + 496, + 652, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {f}} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} (x _ {i}, y _ {i} ^ {t - 1})\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 712, + 777, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the SeCu loss degenerates to the standard cross entropy loss with fixed $\\{\\mathbf{w}\\}$ and can be optimized with SGD. The one-hot label $y^{t - 1}$ is kept from the $(t - 1)$ -th epoch, which makes the optimization consistent between adjacent epochs but the updating for representations may be delayed. To incorporate the information from the current epoch, we include two views of augmentations for the individual instance at each iteration, which is prevalent for representation learning [7, 17, 26].", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16648", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $x_{i}^{1}$ and $x_{i}^{2}$ be two perturbed views of the original image with random augmentations, and the prediction is", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, j} ^ {\\mathrm {s}} = \\frac {\\exp (\\mathbf {x} _ {i} ^ {\\mathrm {s} \\top} \\mathbf {w} _ {j} ^ {t - 1} / \\lambda)}{\\sum_ {j} ^ {K} \\exp (\\mathbf {x} _ {i} ^ {\\mathrm {s} \\top} \\mathbf {w} _ {j} ^ {t - 1} / \\lambda)}; s = \\{1, 2 \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 127, + 421, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_i^s = f_t(x_i^s)$ is extracted by the current encoder. The soft label for each view can be obtained as [26]", + "bbox": [ + 76, + 172, + 468, + 203 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i} ^ {1} = \\tau y _ {i} ^ {t - 1} + (1 - \\tau) p _ {i} ^ {2}; y _ {i} ^ {2} = \\tau y _ {i} ^ {t - 1} + (1 - \\tau) p _ {i} ^ {1} (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 209, + 468, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The soft label contains the prediction from the other view of the same instance, which optimizes the consistency between different views at the same iteration. The problem with two-view optimization can be written as", + "bbox": [ + 76, + 233, + 468, + 294 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {f}} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} (x _ {i} ^ {1}, y _ {i} ^ {1}) + \\ell_ {\\mathrm {S e C u}} (x _ {i} ^ {2}, y _ {i} ^ {2})\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 300, + 403, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Update of $y$ When fixing $\\mathbf{x}_i^t$ and $\\{\\mathbf{w}^t\\}$ , the cluster assignment can be updated by solving the problem", + "bbox": [ + 76, + 358, + 468, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log (p _ {i, j})\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 395, + 338, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad h _ {m} (Y) \\geq b _ {m}, \\quad m = 1, \\dots , M\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 441, + 398, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_{i,j}$ is defined by $\\mathbf{x}_i^t$ and $\\mathbf{w}^t$ as in Eqn. 2. Without the constraints $\\{h_m\\}$ , the objective implies a greedy solution that assigns each instance to the most related cluster. It can incur the trivial solution of collapsing, especially for online deep clustering, where each instance can only be accessed once in each epoch and cluster assignments cannot be refined with multiple iterations over the entire data. Therefore, many effective constraints are developed to mitigate the challenge.", + "bbox": [ + 76, + 464, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we investigate the size constraint that is prevalent in existing methods [5, 26]. Following [26], the cluster size can be lower-bounded as", + "bbox": [ + 76, + 599, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log (p _ {i, j})\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 648, + 328, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad \\sum_ {i} y _ {i, j} \\geq \\gamma N / K, \\quad j = 1, \\dots , K \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 695, + 468, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is the proportion to the average size and $\\gamma = 1$ implies the balanced constraint. Let $\\rho_{j}$ denote dual variables for the $j$ -th lower-bound constraint. The problem becomes", + "bbox": [ + 76, + 732, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\rho : \\rho \\geq 0} \\min _ {y _ {i} \\in \\Delta} - \\sum_ {i} \\sum_ {j} y _ {i, j} \\log (p _ {i, j}) - \\sum_ {j} \\rho_ {j} (\\sum_ {i} y _ {i, j} - \\gamma N / K)\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 782, + 478, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When a mini-batch of instances arrive, the cluster assignment for each instance can be obtained via a closed-form solution", + "bbox": [ + 76, + 821, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i, j} ^ {t} = \\left\\{ \\begin{array}{l l} 1 & j = \\arg \\min _ {j} - \\log \\left(p _ {i, j}\\right) - \\rho_ {j} \\\\ 0 & o. w. \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 869, + 468, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the dual variables will be updated by stochastic gradient ascent. More details can be found in [26].", + "bbox": [ + 498, + 90, + 888, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The size constraint is capable of avoiding collapsing problem explicitly, but it introduces additional dual variables to help assignment and has to optimize them in training. To simplify the optimization, a global entropy constraint is investigated to balance the distribution over all clusters. Compared with the size constraint, the bound for the cluster size is implicit with the entropy constraint. Nevertheless, as we will demonstrate, it can help reduce the number of variables and hyper-parameters, which is more friendly for users.", + "bbox": [ + 496, + 122, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the set of cluster assignments $\\{y\\}$ , the entropy of the whole data set can be defined as", + "bbox": [ + 498, + 273, + 890, + 304 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH (y) = - \\sum_ {j = 1} ^ {K} \\frac {\\sum_ {i} ^ {N} y _ {i , j}}{N} \\log (\\frac {\\sum_ {i} ^ {N} y _ {i , j}}{N})\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 315, + 825, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the entropy as the regularization, the objective can be written as", + "bbox": [ + 498, + 371, + 888, + 400 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log (p _ {i, j}) \\quad s. t. \\quad H (y) \\geq \\gamma\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 412, + 852, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared with the problem consisting of $K$ constraints in Eqn. 8, there is only a single constraint that controls the size of all clusters simultaneously. According to the dual theory [3], the problem is equivalent to", + "bbox": [ + 498, + 469, + 890, + 530 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log \\left(p _ {i, j}\\right) - \\alpha H (y) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 542, + 890, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since one-hot labels of all instances are kept in memory, the optimal solution can be obtained by enumerating. When optimizing the label for $x_{i}$ and fixing other labels, the closed-form solution for $y_{i}$ is", + "bbox": [ + 498, + 599, + 888, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i, j} ^ {t} = \\left\\{ \\begin{array}{l l} 1 & j = \\arg \\min _ {j} - \\log \\left(p _ {i, j}\\right) - \\alpha H \\left(y ^ {t - 1}, y _ {i: j}\\right) \\\\ 0 & o. w. \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 672, + 893, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H(y^{t - 1},y_{i:j})$ replaces the previous label $y_{i}$ of $\\mathbf{x}_i$ by $j$ as $H(y^{t - 1},y_{i:j}) = H([y_1^{t - 1},\\ldots ,y_{i - 1}^{t - 1},y_{i + 1}^{t - 1},\\ldots ,y_n^{t - 1};y_{i,j} = 1])$ . The convergence is guaranteed as in the following theorem.", + "bbox": [ + 498, + 736, + 890, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2. Let $\\mathcal{L}(Y)$ denote the objective in Eqn. 10. If updating cluster assignments sequentially according to Eqn. 11, we have $\\mathcal{L}(Y^t) \\leq \\mathcal{L}(Y^{t-1})$ .", + "bbox": [ + 498, + 810, + 890, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. It is directly from the optimality of the closed-form solution in Eqn. 11. $\\square$", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16649", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark While some deep clustering methods apply the entropy regularization for learning, it is defined over a minibatch of instances and optimized by SGD [14]. On the contrary, our method leverages the entropy of the entire data set that can capture the global distribution better. Moreover, the entropy in our method can be maximized with a closed-form solution, which is more efficient for learning.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the two-view optimization, the updating becomes", + "bbox": [ + 96, + 198, + 468, + 213 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nj = \\arg \\min _ {j} - (\\log (p _ {i, j} ^ {1}) + \\log (p _ {i, j} ^ {2})) / 2 - \\alpha H (y ^ {t - 1}, y _ {i: j})\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 239, + 465, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Update of w With fixed $\\mathbf{x}_i^t$ and the updated pseudo label $\\{y^{t}\\}$ , cluster centers can be optimized by minimizing SeCu loss over two views of augmentations by SGD", + "bbox": [ + 76, + 286, + 468, + 333 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\{\\mathbf {w} _ {j} \\}} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} (\\mathbf {x} _ {i} ^ {1}, y _ {i} ^ {t}) + \\ell_ {\\mathrm {S e C u}} (\\mathbf {x} _ {i} ^ {2}, y _ {i} ^ {t})\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 344, + 403, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alternatively, $\\mathbf{w}$ can be updated by a closed-form solution", + "bbox": [ + 76, + 398, + 460, + 414 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {j} ^ {t: B} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} (\\frac {\\sum_ {i : y _ {i} = j} ^ {B} (1 - p _ {i , j}) \\mathbf {x} _ {i} ^ {t}}{\\sum_ {i : y _ {i} = j} ^ {B} 1 - p _ {i , j}})\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 425, + 408, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $B$ denotes the total number of instances received in the current epoch.", + "bbox": [ + 76, + 479, + 468, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed method is easy to implement and the whole framework is illustrated in Alg. 1. Compared with the supervised discrimination, our main computational overhead is from cluster assignment, which is negligible due to the online strategy.", + "bbox": [ + 75, + 511, + 468, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 602, + 207, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance of the proposed method is evaluated on CIFAR-10 [21], CIFAR-100 [21], STL-10[10] and ImageNet [27]. To demonstrate the generalization performance of SeCu, we follow the setting in SCAN [14] that trains and evaluates models on the standard train/test split provided by the original data set. For STL-10 that contains an additional noisy unlabeled data set, we first learn the model on the training and noisy set, and then on the target training data as suggested by other methods [14, 32]. The standard metrics for clustering, including clustering accuracy (ACC), normalized mutual information (NMI) and adjusted rand index (ARI) are reported for evaluation. For a fair comparison, we follow the common practice to configure the architecture and training protocol for our method. Concretely, ResNet-50 [18] is applied on ImageNet, while ResNet-18 is adopted for other data sets. The key settings for ResNet-18 are elaborated as follows, while those of ResNet-50 follow [26]. More details can be found in the supplementary.", + "bbox": [ + 75, + 628, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Pseudo-code of Stable Cluster Discrimination (SeCu) for One-stage Deep Clustering." + ], + "code_body": "f: encoder network\n# w: cluster centers\n# w_p: centers from the last epoch\n# y: list of pseudo one-hot labels (Nx1)\n# tau: weight for labels from last epoch\n# lambda: temperature\n# keep last cluster centers before each epoch\nw_p = w_detach()\n# train one epoch\nfor x in loader: # load a minibatch with b samples\nx_1, x_2 = f(aug(x)), f(aug(x)) # two random views\ny_x = y(x_id) # retrieve label from last epoch\n# compute prediction for each view\np_1 = softmax(x_1 @ w_p / lambda)\np_2 = softmax(x_2 @ w_p / lambda)\n# obtain soft label for discrimination\ny_1 = tau * y_x + (1-tau) * p_2\ny_2 = tau * y_x + (1-tau) * p_1\n# loss for representation learning\nloss_x = (SeCu(p_1, y_1) + SeCu(p_2, y_2)) / 2\n# update prediction for clustering\np_1 = softmax(x_1detach() @ w / lambda)\np_2 = softmax(x_2detach() @ w / lambda)\n# update cluster assignments with constraints\ny(x_id) = y_x = cluster_assign(p_1, p_2)\n# loss for cluster centers\nloss_w = (SeCu(p_1, y_x) + SeCu(p_2, y_x)) / 2\n# update encoder and cluster centers\nloss = loss_x + loss_w\nloss_backward()", + "guess_lang": "txt", + "bbox": [ + 501, + 130, + 879, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architecture Besides backbone network, a 2-layer MLP head is attached and the output dimension is 128. The similar architecture is adopted in [26, 32]. With the MLP head, the total number of parameters is almost the same as the original ResNet-18. The discussion about the MLP head can be found in Sec. 4.1.1. After the MLP head, the learned representation will be classified by a fully-connected (FC) layer that encodes cluster centers. The size of the FC layer is $128 \\times K$ . As a clustering method, we let $K$ be the number of ground-truth classes as in [14] and have the direct prediction from the model as cluster assignments for evaluation. Following [1, 14], 10 different classification heads are applied as multi-clustering, which benefits the target clustering. To avoid the problem of selecting an appropriate head for evaluation when 10 heads have the same $K$ , we have a varying $cK$ as the number of clusters for each head, where $c \\in \\{1, \\dots, 10\\}$ . The prediction from the head with $c = 1$ is reported for comparison.", + "bbox": [ + 496, + 452, + 890, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimization To reduce efforts of parameter tuning and reuse the parameter of [26], we have $\\mathbf{x}_i^\\top \\mathbf{w}_j$ in lie of $\\log(p_{i,j})$ in Eqn. 9 and 11 for assignment. Searching the parameter for the latter one can achieve the similar performance. Before training the encoder network, one epoch is adopted to initialize cluster assignments and centers. The encoder network is optimized by SGD with a batch size of 128. The momentum and learning rate are set to be 0.9 and 0.2, respectively. Moreover, the first 10 epochs are applied for warm-up and the cosine decay for learning rate", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16650", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is adopted subsequently. For CIFAR-10, the model is optimized by 400 epochs and cluster centers are learned by SGD with a constant learning rate of 1.2. For other challenging data sets, epochs and learning rate for centers are 800 and 0.8, respectively. $\\tau$ for the soft label as in Eqn. 7 and the temperature $\\lambda$ are fixed as 0.2 and 0.05. For the size constraint, the lower-bound parameter $\\gamma$ and the learning rate of dual variables $\\eta_{\\rho}$ are set to be 0.9 and 0.1, respectively. The only parameter $\\alpha$ for the proposed global entropy constraint is set to be $6N / 50$ , where $N$ is the total number of instances. The weight is scaled according to the ablation study on CIFAR-10 as illustrated in Sec. 4.1.4.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Augmentation Augmentation is essential for the success of unsupervised representation learning [7]. For a fair comparison, we apply the prevalent settings in existing work [7, 15]. Concretely, random crop, color jitter, random gray scale, Gaussian blur, solarize and random horizontal flip are included as augmentations. Considering that the resolution of images in CIFAR and STL-10 is much smaller than those in ImageNet, we let the minimal scale of random crop be 0.3, 0.2, and 0.2 for CIFAR-10, CIFAR-100, and STL-10, respectively. That for ImageNet is kept as 0.08. Other parameters are the same for all data sets and many recent work share the similar augmentation [9, 15, 26, 32].", + "bbox": [ + 75, + 292, + 468, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All experiments on small data sets are implemented on a single V100, while 8 GPUs are sufficient for ImageNet.", + "bbox": [ + 76, + 474, + 468, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 513, + 230, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To illustrate the behavior of the proposed method, we study the effect of parameters in SeCu on CIFAR-10 with the size constraint.", + "bbox": [ + 76, + 537, + 468, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.1 Effect of MLP Head", + "text_level": 1, + "bbox": [ + 76, + 603, + 272, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The MLP head shows better generalization than the FC layer in representation learning [7], but its application in deep clustering has not been evaluated systematically. We compare different architectures of MLP head and summarize results in Table 1. Besides the standard MLP layer for projection, there can be another MLP layer attached to the projection MLP head, which is referred as the prediction head [15].", + "bbox": [ + 75, + 628, + 468, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, with the standard architecture denoted by 0 for projection and prediction, SeCu is already $3.7\\%$ better than SCAN on ACC and it confirms that the one-stage deep clustering strategy can learn more consistent features than two-stage methods. If including a single layer for SeCu, the performance can be improved by $2.1\\%$ , which shows that adding a layer with the reduced dimension in head is helpful for deep clustering. Moreover, due to the dimension reduction from the projection layer, the size of cluster centers decreases and the total number of parameters in the model", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2a02e61286c88b32768304a15ed718c7b44bae0cbf29397cddb360cad381db36.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Proj#Pred#ParaACCNMIARI
0011.45M85.576.273.1
1011.30M87.678.776.7
2011.57M88.179.477.6
3011.83M88.179.477.6
3211.97M88.279.677.8
", + "bbox": [ + 526, + 88, + 867, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is even smaller than that of the original network. When increasing the number of layers and having a 2-layer MLP head, ACC can be further improved by $0.5\\%$ and the model size is only slightly increased. After that, there is no significant gain by applying more complicated MLP head. It is because that the 2-layer MLP head is sufficient for small data sets and we will fix 2-layer MLP as the head for the rest experiments except ImageNet, where 3-layer projection and 2-layer prediction are applied as in [26].", + "bbox": [ + 496, + 299, + 893, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.2 Effect of Negative Instances", + "text_level": 1, + "bbox": [ + 500, + 460, + 748, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After the ablation experiments for the architecture, we study the effect of the proposed stable cluster discrimination loss in Eqn. 3. Table 2 compares it to the standard cross entropy loss (CE) that keeps the gradient from negative instances.", + "bbox": [ + 496, + 488, + 890, + 550 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/42097bc2101780547c2338135441593327958447e73e84bfdb62cf182289d71e.jpg", + "table_caption": [ + "Table 1: Comparison of MLP head on CIFAR-10. \"Proj\" and \"#Pred\" indicate the number of layers in MLP for projection head and prediction head, respectively. ACC (\\%), NMI (\\%) and ARI (\\%) are reported. Baseline ACC of SCAN [14] is $81.8\\%$ ." + ], + "table_footnote": [], + "table_body": "
Loss#Max#MinACCNMIARI
CE6,4864,50151.650.235.0
SeCu5,1904,55688.179.477.6
", + "bbox": [ + 532, + 563, + 859, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Comparison of loss functions on CIFAR-10. \"Max\" and \"Min\" indicate the maximal and minimal size of obtained clusters.", + "bbox": [ + 496, + 623, + 890, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we can observe that the proposed SeCu loss outperforms the standard cross entropy loss by a large margin of $36.5\\%$ on ACC, which confirms the benefit of stable training. Then, by investigating the distribution of learned clusters, we find that the minimal size of clusters obtained from different loss functions is almost the same due to the lower-bound constraint, while the maximal size varies. The largest cluster obtained by SeCu contains 5,190 instances, which is close to the ground-truth distribution of CIFAR-10. On the contrary, the dominating cluster found by CE loss has 6,486 instances that is significantly different from the uniform distribution in CIFAR-10. It is because that the negative instances in CE loss can perturb the learning of centers and result in sub-optimal cluster assignments.", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16651", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.3 Effect of Updating Criterion for Cluster Centers", + "text_level": 1, + "bbox": [ + 76, + 90, + 467, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The cluster centers in SeCu can be optimized by SGD or updated by a closed-form solution as in Eqn. 4. We compare these two strategies and summarize the performance in Table 3.", + "bbox": [ + 76, + 114, + 467, + 174 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/482a7b4599f579f1f67e9001b5f3075ef079b7dc9d7ac5d35f587d859c0fbb3b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsACCNMIARI
SeCu-CF87.4±0.578.5±0.676.4±0.9
SeCu-SGD88.1±0.379.3±0.477.5±0.4
", + "bbox": [ + 102, + 183, + 444, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let \"SeCu-CF\" and \"SeCu-SGD\" be the closed-form update and the SGD optimization, respectively. We can observe that SeCu-CF has the similar performance to SeCu-SGD, which shows the effectiveness of the closed-form updating. Moreover, SeCu-SGD has a smaller standard deviation than SeCu-CF due to the smoothness from the momentum in SGD. Compared with SeCu-CF, SeCu-SGD has an additional parameter of the learning rate for centers. Hence, SeCu-CF can be applied to evaluate the preliminary performance of SeCu with less tuning efforts.", + "bbox": [ + 75, + 290, + 467, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1.4 Effect of $\\alpha$ in Entropy Constraint", + "text_level": 1, + "bbox": [ + 76, + 459, + 362, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides the size constraint, the proposed entropy constraint is also effective for balancing cluster assignments. Table 4 demonstrates the results with different $\\alpha$ .", + "bbox": [ + 76, + 482, + 467, + 527 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9e8fb517f4e49ee3d58bfa2fa6982a42982ca87a3b2db7618b111a38e9ee7652.jpg", + "table_caption": [ + "Table 3: Comparison of optimization for cluster centers on CIFAR-10. Mean±std over 8 trails are reported." + ], + "table_footnote": [], + "table_body": "
α#Max#MinACCNMIARI
20,0005,0564,91385.376.373.1
8,0005,1154,65887.979.077.1
6,0005,1204,57988.079.477.4
4,0005,3764,32484.975.472.2
10028,575019.938.717.0
050,000010.00.00.0
", + "bbox": [ + 102, + 536, + 444, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Comparison of $\\alpha$ in Eqn. 10 on CIFAR-10.", + "bbox": [ + 101, + 656, + 442, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, a large $\\alpha$ will regularize the distribution to be uniform and result in a sub-optimal performance. By decreasing the weight, the assignment becomes more flexible and a desired performance can be observed when $\\alpha = 6,000$ . However, a small $\\alpha$ will lead to an imbalanced distribution, which is inappropriate for the balanced data sets such as CIFAR and STL. $\\alpha = 0$ discards the constraint and incurs collapsing. Evidently, the entropy constraint can balance the size of clusters effectively", + "bbox": [ + 75, + 689, + 467, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the entropy is defined on the whole data set and CIFAR-10 contains 50,000 instances for training, $\\alpha$ will be scaled as $6N / 50$ for different data sets, where $N$ denotes the total number of instances in the training set. More ablation experiments can be found in the supplementary.", + "bbox": [ + 75, + 825, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art", + "text_level": 1, + "bbox": [ + 500, + 90, + 795, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After the ablation study, experiments are conducted on benchmark data sets to compare with state-of-the-art methods. The results averaged over 8 trials and the best performance among them are reported for our method. SeCu with size constraint and entropy constraint are denoted as \"SeCu-Size\" and \"SeCu-Entropy\", respectively. Table 5 shows the performance of different methods, where two-stage methods with a pre-training phase are marked for \"Two-stage\".", + "bbox": [ + 496, + 114, + 890, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, as a one-stage method, SeCu outperforms two-stage methods by a margin of about $3\\%$ over all metrics on CIFAR-10 and CIFAR-100-20. It demonstrates that learning representations and clustering simultaneously is essential for deep clustering. The two-stage method NNM[11] shows the comparable performance to SeCu on STL-10. It is because that STL-10 only constrains 5,000 target instances for training and both of NNM and SeCu already achieve the accuracy of supervised learning. Second, compared with the one-stage method without sufficient representation learning, i.e., PICA, SeCu demonstrates a better performance by about $20\\%$ on CIFAR-10 and CIFAR-100-20, and $10\\%$ on STL-10, which shows the importance of optimizing representations for instances. Third, with the same size constraint, SeCu surpasses CoKe by $4.3\\%$ for ARI on CIFAR-10, which confirms the effectiveness of the proposed learning objective. Finally, compared with size constraint in SeCu-Size, SeCu-Entropy demonstrates a competitive performance over all data sets. Since the entropy constraint is optimized by enumeration without introducing any auxiliary variable, it can be an substitute to the size constraint for cluster assignments.", + "bbox": [ + 496, + 234, + 890, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After the clustering phase, self-labeling [14] is an effective strategy to further improve the performance. We try it for SeCu-Size with the setting in [14] and report the result in Table 5. Interestingly, self-labeling is also effective for SeCu and helps approach the supervised accuracy on CIFAR-10. As a one-stage method, SeCu has the strong augmentation for representation learning, which may introduce additional noise for clustering. Self-labeling has the weak augmentation for fine-tuning and refines clustering.", + "bbox": [ + 496, + 566, + 890, + 703 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/491d7827b83f03d6386245ce0ed1eeb26bde28b70ca843b4fdb4f4e6619b3a49.jpg", + "image_caption": [ + "CIFAR-10" + ], + "image_footnote": [], + "bbox": [ + 506, + 718, + 687, + 767 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99998451b7ab208cdebf4ae06a508fc39cea43172bf9a543156afa1cf4c377dd.jpg", + "image_caption": [ + "STL-10", + "Figure 2: Illustration of exemplar images." + ], + "image_footnote": [], + "bbox": [ + 506, + 777, + 687, + 838 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76e829472ebecd858455b17dde3d09d35a61b7996e19c5ca6b98d529626617e7.jpg", + "image_caption": [ + "CIFAR-100-20" + ], + "image_footnote": [], + "bbox": [ + 689, + 718, + 885, + 838 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16652", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5fe4766ffc752a311ea0d90b0d2c3f0cd9d46847c7d3336f0c28e7e57aec29a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTwo-stageCIFAR10CIFAR100-20STL10
ACCNMIARIACCNMIARIACCNMIARI
Supervised93.886.287.080.068.063.280.665.963.1
DeepCluster [4]37.4--18.9--33.4--
IIC [20]61.751.141.125.722.511.759.649.639.7
PICA [19] (best)69.659.151.233.731.017.171.361.153.1
Pretext [7]+k-means✓65.959.850.939.540.223.965.860.450.6
SCAN [14] (mean)✓81.871.266.542.244.126.775.565.459.0
NNM [11]✓84.374.870.947.748.431.680.869.465.0
GCC [32]✓85.676.472.847.247.230.578.868.463.1
CoKe [26] (mean)85.776.673.249.749.133.5---
SeCu-Size (mean)88.179.377.550.050.735.080.269.463.9
SeCu-Size (best)88.579.978.251.651.636.081.470.765.7
SeCu-Entropy (mean)88.079.277.249.950.634.179.568.763.0
SeCu-Entropy (best)88.279.777.751.251.434.980.569.964.4
with self-labeling:
SCAN [14]✓88.379.777.250.748.633.380.969.864.6
GCC [32]✓90.1--52.3--83.3--
SeCu93.086.185.755.255.139.783.673.369.3
", + "bbox": [ + 130, + 88, + 839, + 385 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Exemplar Images Finally, we show exemplar images from clusters obtained by SeCu for different data sets. Fig. 2 illustrates the images that are close to cluster centers in CIFAR-10, CIFAR-100-20, and STL-10, respectively. Evidently, the proposed method can obtain the ground-truth class structure for CIFAR-10 and STL-10. Moreover, with a large intra-class variance, SeCu can still observe appropriate clusters for CIFAR-100-20.", + "bbox": [ + 75, + 438, + 468, + 556 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Considering that labels of super-classes in CIFAR-100-20 contain vague concepts such as \"vehicles 1\" and \"vehicles 2\", we include results on the target 100 classes in Table 6, which may benefit future research. The experiment settings for CIFAR-100 are identical to that for CIFAR-100-20, except the learning rate of centers, which is reduced to 0.01.", + "bbox": [ + 75, + 559, + 468, + 662 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/028a8dfb6023c0a1a3921f65091fc82bc87f41952f27adef5412bd6cc3b073fb.jpg", + "table_caption": [ + "Table 5: Comparison of clustering methods on benchmark data sets. \"Two-stage\" denotes an additional pre-training stage." + ], + "table_footnote": [], + "table_body": "
MethodsACCNMIARI
SeCu-Size (mean)46.461.732.3
SeCu-Size (best)47.161.932.6
SeCu-Entropy (mean)45.360.731.5
SeCu-Entropy (best)46.761.032.2
SeCu†51.365.237.1
", + "bbox": [ + 114, + 672, + 433, + 767 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3. Comparison on ImageNet", + "text_level": 1, + "bbox": [ + 76, + 832, + 310, + 848 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Now we evaluate SeCu on the challenging ImageNet data set in Table 7, where the cost of many two-stage methods becomes intractable and only available baselines are", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "included. First, SeCu achieves $51.4\\%$ ACC on ImageNet, which is better than SCAN with self-labeling by more than $11\\%$ . With the same 10 clustering heads and size constraint, SeCu is still $3.8\\%$ better than CoKe on ACC. Compared with the objective in CoKe, SeCu is tailored for deep clustering and the comparison confirms the efficacy of our method. With self-labeling, SeCu shows state-of-the-art performance of $53.5\\%$ ACC on ImageNet. It demonstrates that our method is applicable for large-scale data sets.", + "bbox": [ + 496, + 436, + 890, + 575 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/4271b570cdbf2fd5ebfcb6acb5b202013abd2a478e506400fd1b67dc6b0fad43.jpg", + "table_caption": [ + "Table 6: Results of SeCu on CIFAR-100. † denotes a method with self-labeling." + ], + "table_footnote": [], + "table_body": "
MethodsACCNMIARI
SCAN†39.972.027.5
CoKe‡47.676.235.6
SeCu51.478.039.7
SeCu†53.579.441.9
", + "bbox": [ + 580, + 587, + 812, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 7: Comparison on ImageNet. ${}^{ \\dagger }$ denotes a method with self-labeling. ${}^{ \\ddagger }$ is our reproduction with 10 clustering heads.", + "bbox": [ + 496, + 676, + 890, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 739, + 617, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, a novel stable cluster discrimination task that considers the hardness of instances is proposed for one-stage deep clustering. To avoid collapsing in representations, a global entropy constraint with theoretical guarantee is investigated. Finally, the comparison with state-of-the-art methods confirms the effectiveness of our proposal. While recent work shows that leveraging nearest neighbors improves deep clustering [11, 14, 32], incorporating such information in SeCu can be our future work.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16653", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020.", + "[2] Hangbo Bao, Li Dong, and Furu Wei. Beit: BERT pretraining of image transformers. CoRR, abs/2106.08254, 2021.", + "[3] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004.", + "[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018.", + "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020.", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021.", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, volume 119, pages 1597-1607, 2020.", + "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758. Computer Vision Foundation / IEEE, 2021.", + "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. CoRR, abs/2104.02057, 2021.", + "[10] Adam Coates, Andrew Y. Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In AISTATS, volume 15, pages 215-223. JMLR.org, 2011.", + "[11] Zhiyuan Dang, Cheng Deng, Xu Yang, Kun Wei, and Heng Huang. Nearest neighbor matching for deep clustering. In CVPR, pages 13693-13702. Computer Vision Foundation / IEEE, 2021.", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021.", + "[13] Ehsan Elhamifar and René Vidal. Sparse subspace clustering: Algorithm, theory, and applications. IEEE Trans. Pattern Anal. Mach. Intell., 35(11):2765-2781, 2013.", + "[14] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. SCAN: learning to classify images without labels. In ECCV, volume 12355, pages 268-285. Springer, 2020.", + "[15] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent - A new approach to self-supervised learning. In NeurIPS, 2020." + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021.", + "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020.", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016.", + "[19] Jiabo Huang, Shaogang Gong, and Xiatian Zhu. Deep semantic clustering by partition confidence maximisation. In CVPR, pages 8846-8855. Computer Vision Foundation / IEEE, 2020.", + "[20] Xu Ji, Andrea Vedaldi, and João F. Henriques. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9864-9873. IEEE, 2019.", + "[21] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009.", + "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, pages 1106-1114, 2012.", + "[23] Yunfan Li, Peng Hu, Jerry Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In AAAI, pages 8547-8555. AAAI Press, 2021.", + "[24] Stuart P. Lloyd. Least squares quantization in PCM. IEEE Trans. Inf. Theory, 28(2):129-136, 1982.", + "[25] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In ICCV, pages 6449-6457, 2019.", + "[26] Qi Qian, Yuanhong Xu, Juhua Hu, Hao Li, and Rong Jin. Unsupervised visual representation learning by online constrained k-means. In CVPR, pages 16619-16628. IEEE, 2022.", + "[27] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Fei-Fei Li. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis., 115(3):211-252, 2015.", + "[28] Ulrike von Luxburg. A tutorial on spectral clustering. Stat. Comput., 17(4):395-416, 2007.", + "[29] Junyuan Xie, Ross B. Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In ICML, volume 48, pages 478-487. JMLR.org, 2016.", + "[30] Bo Yang, Xiao Fu, Nicholas D. Sidiropoulos, and Mingyi Hong. Towards k-means-friendly spaces: Simultaneous deep learning and clustering. In ICML, 2017.", + "[31] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, volume 139, pages 12310-12320. PMLR, 2021.", + "[32] Huasong Zhong, Jianlong Wu, Chong Chen, Jianqiang Huang, Minghua Deng, Liqiang Nie, Zhouchen Lin, and Xian-Sheng Hua. Graph contrastive clustering. In ICCV, pages 9204-9213. IEEE, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 873 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16654", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Stable Cluster Discrimination for Deep Clustering/9f881270-2ff5-4f87-a2db-8feef1f47071_model.json b/2023/Stable Cluster Discrimination for Deep Clustering/9f881270-2ff5-4f87-a2db-8feef1f47071_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f182013d129ea44550bf0af0f6418771f81f38b1 --- /dev/null +++ b/2023/Stable Cluster Discrimination for Deep Clustering/9f881270-2ff5-4f87-a2db-8feef1f47071_model.json @@ -0,0 +1,2409 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.232, + 0.131, + 0.741, + 0.154 + ], + "angle": 0, + "content": "Stable Cluster Discrimination for Deep Clustering" + }, + { + "type": "text", + "bbox": [ + 0.314, + 0.182, + 0.658, + 0.234 + ], + "angle": 0, + "content": "Qi Qian \nAlibaba Group, Bellevue, WA 98004, USA \nqi.qian@alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.473, + 0.725 + ], + "angle": 0, + "content": "Deep clustering can optimize representations of instances (i.e., representation learning) and explore the inherent data distribution (i.e., clustering) simultaneously, which demonstrates a superior performance over conventional clustering methods with given features. However, the coupled objective implies a trivial solution that all instances collapse to the uniform features. To tackle the challenge, a two-stage training strategy is developed for decoupling, where it introduces an additional pre-training stage for representation learning and then fine-tunes the obtained model for clustering. Meanwhile, one-stage methods are developed mainly for representation learning rather than clustering, where various constraints for cluster assignments are designed to avoid collapsing explicitly. Despite the success of these methods, an appropriate learning objective tailored for deep clustering has not been investigated sufficiently. In this work, we first show that the prevalent discrimination task in supervised learning is unstable for one-stage clustering due to the lack of ground-truth labels and positive instances for certain clusters in each mini-batch. To mitigate the issue, a novel stable cluster discrimination (SeCu) task is proposed and a new hardness-aware clustering criterion can be obtained accordingly. Moreover, a global entropy constraint for cluster assignments is studied with efficient optimization. Extensive experiments are conducted on benchmark data sets and ImageNet. SeCu achieves state-of-the-art performance on all of them, which demonstrates the effectiveness of one-stage deep clustering." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.21, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Clustering is a fundamental task in unsupervised learning. Given features of instances, the unlabeled data set will be partitioned into multiple clusters, where instances from the same cluster are similar according to the measurement defined by a distance function. With fixed features, most of research efforts focus on studying appropriate distance functions and ingenious algorithms have been proposed by different measurements, e.g., k-means clustering [24], spec" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.271, + 0.687, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.27, + 0.871, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.398, + 0.895, + 0.536 + ], + "angle": 0, + "content": "Figure 1: An illustration of the proposed method. 10 data points are randomly sampled from two different Gaussian distributions, respectively. Points with the same color are from the same distribution, while squares denote the corresponding cluster centers obtained by different methods. Star indicates the mis-classified data. Unlike k-means that assigns the uniform weight for different instances, our method considers the hardness of instances and is better for discrimination based clustering." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.565, + 0.825, + 0.581 + ], + "angle": 0, + "content": "tral clustering [28], subspace clustering [13], etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.763 + ], + "angle": 0, + "content": "With the development of deep neural networks, deep learning is capable of learning representations from raw materials and demonstrates the dominating performance on various supervised tasks [22]. Thereafter, it is also introduced to clustering recently [4, 11, 14, 19, 20, 23, 29, 30, 32]. Unlike the conventional clustering, representations of instances are learned with cluster assignments and centers simultaneously in deep clustering, which is more flexible to capture the data distribution. However, the coupled objective can result in a trivial solution that all instances collapse to the uniform features [4] and designing appropriate clustering criterion in the new scenario becomes challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Many deep clustering algorithms [11, 14, 32] exploit a two-stage training strategy to decouple representation learning and clustering to avoid collapsing. The recent progress in unsupervised representation learning [2, 4, 7, 16, 17, 26] shows that informative features capturing the semantic similarity between instances without collapsing can be obtained by pre-training a large set of unlabeled data. Inspired by the observation, those methods leverage a pre-training stage and then focus on developing algorithms to optimize clus" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16645" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "tering by fine-tuning a pre-trained model in the second stage. By refining the relations between nearest neighbors obtained from the pre-training stage, two-stage methods achieve a significantly better performance than one-stage ones without sufficient representation learning [19, 20]." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.17, + 0.468, + 0.425 + ], + "angle": 0, + "content": "However, the objective of pre-training can be inconsistent with that of clustering, which results in a sub-optimal performance for deep clustering. Note that different from supervised learning where the objective is explicitly defined by labels, that for unsupervised representation learning is arbitrary and various pretext tasks have been proposed, e.g., instance discrimination [7, 17], cluster discrimination [4, 26], masked modeling [2, 16], etc. Most of existing two-stage clustering methods adopt instance discrimination, i.e., SimCLR [7], for pre-training, whereas it aims to identify each instance as an individual class and the objective is different from clustering that aims to group instances from the same cluster. Moreover, SwAV [5] demonstrates that clustering itself is also an effective pretext task for representation learning. Therefore, we focus on facilitating one-stage deep clustering that optimizes representations and clustering simultaneously in this work." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.429, + 0.468, + 0.564 + ], + "angle": 0, + "content": "Most of existing one-stage methods are proposed solely for representation learning. To tackle the collapsing issue, research efforts are mainly devoted to developing appropriate constraints for cluster assignments, especially for online deep clustering. For example, [1, 5] apply the balanced constraint that each cluster has the same number of instances for clustering. [26] further relaxes the balanced constraint to a lower-bound size constraint that limits the minimal size of clusters and demonstrates a more flexible assignment." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.567, + 0.468, + 0.716 + ], + "angle": 0, + "content": "After obtaining cluster assignments as pseudo labels, representations and cluster centers can be optimized as in the supervised scenario. For example, [5, 6] learn the encoder network and cluster centers by solving a classification problem with the standard cross entropy loss. [26] has the same classification task for representation learning while adopting k-means for updating cluster centers. Although these methods achieve a satisfied performance on representation learning, a learning objective tailored for deep clustering has not attracted sufficient attentions." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.9 + ], + "angle": 0, + "content": "In this work, we investigate the effective learning task for one-stage deep clustering. By analyzing the standard cross entropy loss for supervised learning, we find that it can be unstable for unsupervised learning. Concretely, the gradient for updating cluster centers consists of two ingredients: gradient from positive instances of the target cluster and that from irrelevant negative ones. However, with a limited mini-batch size in stochastic gradient descent (SGD), there can be no positive ones for a large proportion of clusters (e.g., \\(90\\%\\) on ImageNet) at each iteration. Due to the lack of positive instances, the influence from negative ones is dominating. Unlike supervised learning where labels are fixed," + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.891, + 0.151 + ], + "angle": 0, + "content": "the cluster assignments can change during training in deep clustering. Therefore, the noise from the large variance of negative instances will be accumulated, which makes the optimization unstable." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.153, + 0.891, + 0.393 + ], + "angle": 0, + "content": "To mitigate the problem, we propose a stable cluster discrimination (SeCu) task for deep clustering, which stops the gradient from negative instances for updating cluster centers in the cross entropy loss. Compared with k-means in [26], where positive instances have the uniform weight for updating centers, SeCu as a discrimination task considers the hardness of instances and a large weight will be assigned to hard instances for updating. Fig. 1 illustrates the hardness-aware clustering criterion implied by SeCu. Besides, we improve the cluster assignment by developing an entropy constraint that regularizes the entropy of assignments over the entire data set. Compared to the optimization with the size constraint [26], our method can reduce the number of variables and hyper-parameters, and thus make the learning more convenient. The main contributions of this work can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.407, + 0.891, + 0.481 + ], + "angle": 0, + "content": "- A novel task is proposed for one-stage deep clustering. SeCu tailors the supervised cross entropy loss by eliminating the influence from the negative instances in learning cluster centers, which makes the training stable when ground-truth labels are unavailable." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.495, + 0.891, + 0.539 + ], + "angle": 0, + "content": "- A global entropy constraint is exploited to balance the size of clusters and an efficient closed-form solution is developed for online assignment with the constraint." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.552, + 0.891, + 0.642 + ], + "angle": 0, + "content": "- A simple framework with a single loss function and encoder is introduced for deep clustering in Eqn. 6. The proposed method is evaluated with the standard protocol on benchmark data sets and ImageNet [27]. The superior performance of SeCu confirms its effectiveness for deep clustering." + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.407, + 0.891, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.659, + 0.64, + 0.674 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.685, + 0.89, + 0.714 + ], + "angle": 0, + "content": "In this section, we briefly review unsupervised representation learning and then focus on clustering." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.727, + 0.836, + 0.742 + ], + "angle": 0, + "content": "2.1. Unsupervised Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.75, + 0.891, + 0.9 + ], + "angle": 0, + "content": "By leveraging massive unlabeled data, unsupervised representation learning can learn a pre-trained model that encoders the semantic information from data and helps downstream tasks with fine-tuning. Compared with supervised learning containing labels, one major challenge in unsupervised learning is defining appropriate positive pairs for representation learning. SimCLR [7] and MoCo [17] apply instance discrimination as the objective that generates positive pairs from diverse views of the same instance and considers other instances as the negative ones. Then, BYOL [15]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16646" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "and SimSiam [8] demonstrate that with appropriate neural network architectures, the applicable models can be obtained with only positive pairs. Besides the work optimizing instance space, Barlow Twins [31] defines positive pairs in feature space and also learns effective pre-trained models. Recently, learning with masked modeling [2, 16] achieves success on the specific backbone of vision transformer [12] and shows the better fine-tuning performance than the generic objective on the ResNet [18]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.238, + 0.24, + 0.254 + ], + "angle": 0, + "content": "2.2. Deep Clustering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.502 + ], + "angle": 0, + "content": "Compared with instance discrimination, cluster discrimination aims to obtain positive pairs consisting of different instances for representation learning. Therefore, it has to learn representations and relations between instances simultaneously. DeepCluster [4] obtains membership of instances with an offline k-means and then optimizes representations alternately. Some work focuses on representation learning and introduces the balanced constraint that each cluster has the same number of instances for cluster assignments [1, 5] to mitigate the collapsing problem. Besides, DINO [6] applies an additional momentum encoder to further improve the representation. However, the constraint assumes a well balanced distribution, which is hard to capture the ground-truth data distribution. Thereafter, CoKe [26] relaxes the constraint to only lower-bound the minimal size of clusters to model the distribution more flexibly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.473, + 0.716 + ], + "angle": 0, + "content": "Besides representation learning, there are many works developed solely for better clustering [11, 14, 20, 23, 29, 30, 32]. To handle the collapsing issue, two-stage methods [11, 14, 32] leverage the representations from a pretrained model to obtain nearest neighbors and fine-tune the model for clustering accordingly. In addition, the objective of pre-training can be included for clustering as multitask learning [23]. The work closest to ours is CoKe [26] that optimizes representations and clustering simultaneously. Compared with the conventional k-means in CoKe, we propose a novel stable cluster discrimination objective tailored for one-stage deep clustering. Moreover, an entropy constraint is introduced to reduce the number of parameters while demonstrating the competitive performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.73, + 0.353, + 0.745 + ], + "angle": 0, + "content": "3. Stable Cluster Discrimination" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.754, + 0.291, + 0.769 + ], + "angle": 0, + "content": "3.1. Cluster Discrimination" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.472, + 0.853 + ], + "angle": 0, + "content": "Given a data set with \\(N\\) images \\(\\{x_{i}\\}_{i = 1}^{N}\\), if the corresponding labels are accessible as \\(\\{y_{i}\\}_{i = 1}^{N}\\), the literature from distance metric learning [25] shows that appropriate representations of examples can be learned by optimizing a classification task as" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.863, + 0.341, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\sum_ {i = 1} ^ {N} \\ell \\left(x _ {i}, y _ {i}; \\theta\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "where \\(\\ell (\\cdot)\\) is the cross entropy loss with the normalized Softmax operator as suggested in [25]. \\(\\theta\\) denotes the parameters of the deep neural network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.894, + 0.182 + ], + "angle": 0, + "content": "For unsupervised learning that the label information is unavailable, the objective of cluster discrimination with \\( K \\) clusters can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.192, + 0.892, + 0.236 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {f}, \\left\\{\\mathbf {w} _ {j} \\right\\}, y _ {i} \\in \\Delta} \\mathcal {L} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} - y _ {i, j} \\log \\left(p _ {i, j}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.248, + 0.894, + 0.294 + ], + "angle": 0, + "content": "where \\(y_{i}\\) denotes the learnable label for \\(x_{i}\\) and \\(\\Delta = \\{y_i|\\sum_{j = 1}^{K}y_{i,j} = 1,\\forall j,y_{i,j}\\in \\{0,1\\} \\}\\). With the normalized Softmax operator, the prediction \\(p_{i,j}\\) can be computed" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.306, + 0.892, + 0.345 + ], + "angle": 0, + "content": "\\[\np _ {i, j} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {j} / \\lambda\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {k} / \\lambda\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.416 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_i = f(x_i)\\) and \\(f(\\cdot)\\) denotes the encoder network. \\(\\theta_{f}\\) contains parameters of \\(f\\). \\(\\{\\mathbf{w}_j\\}_{j=1}^K\\) consists of cluster centers. \\(\\lambda\\) is the parameter of temperature and \\(\\forall i, j, \\|\\mathbf{x}_i\\|_2 = \\|\\mathbf{w}_j\\|_2 = 1\\) by normalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Compared with the supervised counterpart, the problem in Eqn. 1 has to optimize cluster assignments \\(\\{y\\}\\), cluster centers \\(\\{\\mathbf{w}\\}\\), and representation encoder network \\(f\\) simultaneously. While most of existing work focuses on optimizing \\(\\{y\\}\\) with different constraints to avoid collapsing, this work investigates the learning objective for \\(\\{\\mathbf{w}\\}\\) and \\(f\\), and a new clustering criterion is introduced accordingly." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.534, + 0.885, + 0.55 + ], + "angle": 0, + "content": "3.2. Stable Loss for Optimization with Mini-batch" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.662 + ], + "angle": 0, + "content": "Unlike supervised learning, where labels are fixed for examples, cluster assignments \\(\\{y\\}\\) in unsupervised learning are dynamic with the training of instance representations and cluster centers. Therefore, the original cross entropy loss becomes unstable for unsupervised discrimination. The problem can be elaborated by analyzing the updating criterion for cluster centers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.707 + ], + "angle": 0, + "content": "Letting \\( y_{i} \\) denote the label of \\( \\mathbf{x}_i \\), the gradient of \\( \\mathbf{w}_j \\) from the standard cross entropy loss in Eqn. 1 can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.719, + 0.856, + 0.757 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {w} _ {j}} \\mathcal {L} = \\frac {1}{\\lambda} \\left(\\sum_ {i: y _ {i} = j} \\left(p _ {i, j} - 1\\right) \\mathbf {x} _ {i} + \\sum_ {k: y _ {k} \\neq j} p _ {k, j} \\mathbf {x} _ {k}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.768, + 0.892, + 0.845 + ], + "angle": 0, + "content": "The former term is to pull the center \\(\\mathbf{w}_j\\) to the assigned instances and the latter one is to push it away from instances of other clusters. However, the following Propositions show that this updating is unstable for deep clustering. The detailed proof can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Proposition 1. When sampling a mini-batch of \\( b \\) instances for \\( K \\) clusters, there are no positive instances for at least \\( K - b \\) clusters." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16647" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.155 + ], + "angle": 0, + "content": "Proposition 2. Let \\(\\mathrm{Var}_{pos}\\) and \\(\\mathrm{Var}_{neg}\\) be the variance of sampled positive and negative instances, respectively. Assuming that each instance has unit norm and the norm of the cluster mean is \\(a\\), we have \\(\\mathrm{Var}_{neg} = \\mathcal{O}\\left(\\frac{1}{1 - a^2}\\right)\\mathrm{Var}_{pos}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.47, + 0.29 + ], + "angle": 0, + "content": "Remark Proposition 1 implies that with cross entropy loss, a large proportion of cluster centers will be updated only by negative instances. Proposition 2 indicates that the variance of sampled negative instances is much larger than that of positive ones when each cluster is compact as \\(a\\) approaching 1. Due to the small size of a mini-batch in training deep neural networks, the variance cannot be reduced sufficiently." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.29, + 0.469, + 0.366 + ], + "angle": 0, + "content": "When \\( K = 10,000 \\) and \\( b = 1,024 \\) as in our multiclustering settings for ImageNet, at least about \\( 90\\% \\) centers will only access a mini-batch of negative instances at each iteration. Without ground-truth labels, the bias in updating will be accumulated and mislead the learning process." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.469, + 0.411 + ], + "angle": 0, + "content": "To mitigate the problem, we propose to eliminate the direction of negative instances from gradient for stable training as" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.41, + 0.374, + 0.447 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {w} _ {j}} \\mathcal {L} = \\frac {1}{\\lambda} \\sum_ {i: y _ {i} = j} (p _ {i, j} - 1) \\mathbf {x} _ {i}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.45, + 0.469, + 0.479 + ], + "angle": 0, + "content": "The corresponding stable cluster discrimination loss becomes" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.489, + 0.469, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\ell_ {\\mathrm {S e C u}} \\left(x _ {i}, y _ {i}\\right) = \\\\ - \\log \\left(\\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right)}{\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right) + \\sum_ {k : k \\neq y _ {i}} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\tilde {\\mathbf {w}} _ {k} / \\lambda\\right)}\\right) \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.56, + 0.469, + 0.634 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{w}}_k\\) denotes \\(\\mathbf{w}_k\\) with the stop-gradient operator. Compared with the standard cross entropy loss, cluster centers in SeCu loss will be updated only by positive instances, which is more stable for deep clustering when optimizing with mini-batches." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.635, + 0.469, + 0.681 + ], + "angle": 0, + "content": "In addition, the analysis in the following theorem demonstrates the novel hardness-aware clustering criterion for deep clustering implied by the proposed objective." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.687, + 0.469, + 0.734 + ], + "angle": 0, + "content": "Theorem 1. When fixing \\(\\{y_i\\}\\) and \\(\\{\\mathbf{x}_i\\}\\), let \\(\\{\\mathbf{w}^*\\}\\) be the optimal solution of the problem with loss function in Eqn. 3 and assume \\(\\forall i, \\| \\mathbf{x}_i\\|_2 = 1; \\forall j, \\| \\mathbf{w}_j\\|_2 = 1\\), then we have" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.74, + 0.469, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {j} ^ {*} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} \\left(\\frac {\\sum_ {i : y _ {i} = j} (1 - p _ {i , j}) \\mathbf {x} _ {i}}{\\sum_ {i : y _ {i} = j} 1 - p _ {i , j}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.785, + 0.422, + 0.801 + ], + "angle": 0, + "content": "where \\(\\Pi_{\\| \\mathbf{w}\\| _2 = 1}\\) projects the vector to the unit norm." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.469, + 0.862 + ], + "angle": 0, + "content": "Remark For k-means in CoKe [26], cluster centers will be averaged over all assigned instances with the uniform weight" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.869, + 0.469, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {j} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} \\left(\\frac {\\sum_ {i : y _ {i} = j} \\mathbf {x} _ {i}}{\\sum_ {i} 1 (y _ {i} = j)}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "where \\(1(\\cdot)\\) is the indicator function. On the contrary, our proposal considers the weight of each instance by its hardness, i.e., \\(p_{i,y_i}\\). By assigning a large weight to hard instances, the corresponding center can capture the distribution of hard instances better, which is illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.228 + ], + "angle": 0, + "content": "Our formulation also implies CoKe as a special case. Concretely, by increasing the temperature \\(\\lambda\\) for updating \\(\\mathbf{w}\\) to be infinite, we have \\(p_{i,j} = 1 / K\\) and Eqn. 4 will degenerate to Eqn. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Finally, besides SGD, Eqn. 4 also suggests an alternative updating strategy for \\(\\mathbf{w}\\). Since \\(p_{i,j}\\) is computed by \\(\\mathbf{w}\\), it may still require multiple iterations with Eqn. 4 for convergence. However, the representations of instances are also updated by training and we can keep a single update for centers at each iteration and improve centers along with the learning of representations. Compared with SGD for learning cluster centers, the strategy eliminates the learning rate for centers with a comparable performance as shown in the ablation study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Given the learning objective SeCu, we will elaborate the proposed deep clustering method in next subsection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.734, + 0.437 + ], + "angle": 0, + "content": "3.3. SeCu for Deep Clustering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.476 + ], + "angle": 0, + "content": "With the proposed loss function, the objective of stable cluster discrimination for deep clustering can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.5, + 0.773, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {f}, \\left\\{\\mathbf {w} _ {j} \\right\\}, y _ {i} \\in \\Delta} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} \\left(x _ {i}, y _ {i}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.544, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad h _ {m} (Y) \\geq b _ {m}, \\quad m = 1, \\dots , M \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.573, + 0.892, + 0.634 + ], + "angle": 0, + "content": "where \\( Y = [y_{1},\\ldots ,y_{N}] \\), \\( M \\) denotes the number of constraints, and \\( h_m(\\cdot) \\) is the \\( m \\)-th constraint for cluster assignment. Considering that variables are entangled, the problem is solved alternately." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Update of \\(\\theta_{f}\\) First, when fixing \\(\\{y^{t - 1}\\}\\) and \\(\\{\\mathbf{w}^{t - 1}\\}\\) from the last epoch, the sub-problem for representation learning in the \\(t\\)-th epoch becomes" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.713, + 0.778, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {f}} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} (x _ {i}, y _ {i} ^ {t - 1})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where the SeCu loss degenerates to the standard cross entropy loss with fixed \\(\\{\\mathbf{w}\\}\\) and can be optimized with SGD. The one-hot label \\(y^{t - 1}\\) is kept from the \\((t - 1)\\)-th epoch, which makes the optimization consistent between adjacent epochs but the updating for representations may be delayed. To incorporate the information from the current epoch, we include two views of augmentations for the individual instance at each iteration, which is prevalent for representation learning [7, 17, 26]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16648" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.122 + ], + "angle": 0, + "content": "Let \\( x_{i}^{1} \\) and \\( x_{i}^{2} \\) be two perturbed views of the original image with random augmentations, and the prediction is" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.128, + 0.423, + 0.169 + ], + "angle": 0, + "content": "\\[\np _ {i, j} ^ {\\mathrm {s}} = \\frac {\\exp (\\mathbf {x} _ {i} ^ {\\mathrm {s} \\top} \\mathbf {w} _ {j} ^ {t - 1} / \\lambda)}{\\sum_ {j} ^ {K} \\exp (\\mathbf {x} _ {i} ^ {\\mathrm {s} \\top} \\mathbf {w} _ {j} ^ {t - 1} / \\lambda)}; s = \\{1, 2 \\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.174, + 0.469, + 0.204 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_i^s = f_t(x_i^s)\\) is extracted by the current encoder. The soft label for each view can be obtained as [26]" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.21, + 0.469, + 0.23 + ], + "angle": 0, + "content": "\\[\ny _ {i} ^ {1} = \\tau y _ {i} ^ {t - 1} + (1 - \\tau) p _ {i} ^ {2}; y _ {i} ^ {2} = \\tau y _ {i} ^ {t - 1} + (1 - \\tau) p _ {i} ^ {1} (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.234, + 0.469, + 0.295 + ], + "angle": 0, + "content": "The soft label contains the prediction from the other view of the same instance, which optimizes the consistency between different views at the same iteration. The problem with two-view optimization can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.301, + 0.405, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {f}} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} (x _ {i} ^ {1}, y _ {i} ^ {1}) + \\ell_ {\\mathrm {S e C u}} (x _ {i} ^ {2}, y _ {i} ^ {2})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.469, + 0.39 + ], + "angle": 0, + "content": "Update of \\(y\\) When fixing \\(\\mathbf{x}_i^t\\) and \\(\\{\\mathbf{w}^t\\}\\), the cluster assignment can be updated by solving the problem" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.396, + 0.339, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log (p _ {i, j})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.442, + 0.4, + 0.457 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad h _ {m} (Y) \\geq b _ {m}, \\quad m = 1, \\dots , M\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.465, + 0.469, + 0.601 + ], + "angle": 0, + "content": "where \\( p_{i,j} \\) is defined by \\( \\mathbf{x}_i^t \\) and \\( \\mathbf{w}^t \\) as in Eqn. 2. Without the constraints \\( \\{h_m\\} \\), the objective implies a greedy solution that assigns each instance to the most related cluster. It can incur the trivial solution of collapsing, especially for online deep clustering, where each instance can only be accessed once in each epoch and cluster assignments cannot be refined with multiple iterations over the entire data. Therefore, many effective constraints are developed to mitigate the challenge." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.601, + 0.469, + 0.645 + ], + "angle": 0, + "content": "First, we investigate the size constraint that is prevalent in existing methods [5, 26]. Following [26], the cluster size can be lower-bounded as" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.65, + 0.329, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log (p _ {i, j})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.696, + 0.469, + 0.727 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad \\sum_ {i} y _ {i, j} \\geq \\gamma N / K, \\quad j = 1, \\dots , K \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.733, + 0.469, + 0.779 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the proportion to the average size and \\(\\gamma = 1\\) implies the balanced constraint. Let \\(\\rho_{j}\\) denote dual variables for the \\(j\\)-th lower-bound constraint. The problem becomes" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.784, + 0.479, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\rho : \\rho \\geq 0} \\min _ {y _ {i} \\in \\Delta} - \\sum_ {i} \\sum_ {j} y _ {i, j} \\log (p _ {i, j}) - \\sum_ {j} \\rho_ {j} (\\sum_ {i} y _ {i, j} - \\gamma N / K)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.823, + 0.469, + 0.867 + ], + "angle": 0, + "content": "When a mini-batch of instances arrive, the cluster assignment for each instance can be obtained via a closed-form solution" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.871, + 0.469, + 0.905 + ], + "angle": 0, + "content": "\\[\ny _ {i, j} ^ {t} = \\left\\{ \\begin{array}{l l} 1 & j = \\arg \\min _ {j} - \\log \\left(p _ {i, j}\\right) - \\rho_ {j} \\\\ 0 & o. w. \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "Then, the dual variables will be updated by stochastic gradient ascent. More details can be found in [26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.892, + 0.273 + ], + "angle": 0, + "content": "The size constraint is capable of avoiding collapsing problem explicitly, but it introduces additional dual variables to help assignment and has to optimize them in training. To simplify the optimization, a global entropy constraint is investigated to balance the distribution over all clusters. Compared with the size constraint, the bound for the cluster size is implicit with the entropy constraint. Nevertheless, as we will demonstrate, it can help reduce the number of variables and hyper-parameters, which is more friendly for users." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.275, + 0.892, + 0.305 + ], + "angle": 0, + "content": "Given the set of cluster assignments \\(\\{y\\}\\), the entropy of the whole data set can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.316, + 0.826, + 0.359 + ], + "angle": 0, + "content": "\\[\nH (y) = - \\sum_ {j = 1} ^ {K} \\frac {\\sum_ {i} ^ {N} y _ {i , j}}{N} \\log (\\frac {\\sum_ {i} ^ {N} y _ {i , j}}{N})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.372, + 0.89, + 0.401 + ], + "angle": 0, + "content": "With the entropy as the regularization, the objective can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.413, + 0.854, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log (p _ {i, j}) \\quad s. t. \\quad H (y) \\geq \\gamma\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Compared with the problem consisting of \\( K \\) constraints in Eqn. 8, there is only a single constraint that controls the size of all clusters simultaneously. According to the dual theory [3], the problem is equivalent to" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.544, + 0.891, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log \\left(p _ {i, j}\\right) - \\alpha H (y) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.601, + 0.89, + 0.661 + ], + "angle": 0, + "content": "Since one-hot labels of all instances are kept in memory, the optimal solution can be obtained by enumerating. When optimizing the label for \\( x_{i} \\) and fixing other labels, the closed-form solution for \\( y_{i} \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.674, + 0.894, + 0.723 + ], + "angle": 0, + "content": "\\[\ny _ {i, j} ^ {t} = \\left\\{ \\begin{array}{l l} 1 & j = \\arg \\min _ {j} - \\log \\left(p _ {i, j}\\right) - \\alpha H \\left(y ^ {t - 1}, y _ {i: j}\\right) \\\\ 0 & o. w. \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.737, + 0.892, + 0.799 + ], + "angle": 0, + "content": "where \\( H(y^{t - 1},y_{i:j}) \\) replaces the previous label \\( y_{i} \\) of \\( \\mathbf{x}_i \\) by \\( j \\) as \\( H(y^{t - 1},y_{i:j}) = H([y_1^{t - 1},\\ldots ,y_{i - 1}^{t - 1},y_{i + 1}^{t - 1},\\ldots ,y_n^{t - 1};y_{i,j} = 1]) \\). The convergence is guaranteed as in the following theorem." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.858 + ], + "angle": 0, + "content": "Theorem 2. Let \\(\\mathcal{L}(Y)\\) denote the objective in Eqn. 10. If updating cluster assignments sequentially according to Eqn. 11, we have \\(\\mathcal{L}(Y^t) \\leq \\mathcal{L}(Y^{t-1})\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Proof. It is directly from the optimality of the closed-form solution in Eqn. 11. \\(\\square\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16649" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "Remark While some deep clustering methods apply the entropy regularization for learning, it is defined over a minibatch of instances and optimized by SGD [14]. On the contrary, our method leverages the entropy of the entire data set that can capture the global distribution better. Moreover, the entropy in our method can be maximized with a closed-form solution, which is more efficient for learning." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.199, + 0.469, + 0.214 + ], + "angle": 0, + "content": "With the two-view optimization, the updating becomes" + }, + { + "type": "equation", + "bbox": [ + 0.079, + 0.24, + 0.466, + 0.265 + ], + "angle": 0, + "content": "\\[\nj = \\arg \\min _ {j} - (\\log (p _ {i, j} ^ {1}) + \\log (p _ {i, j} ^ {2})) / 2 - \\alpha H (y ^ {t - 1}, y _ {i: j})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.287, + 0.469, + 0.334 + ], + "angle": 0, + "content": "Update of w With fixed \\(\\mathbf{x}_i^t\\) and the updated pseudo label \\(\\{y^{t}\\}\\), cluster centers can be optimized by minimizing SeCu loss over two views of augmentations by SGD" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.345, + 0.404, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\{\\mathbf {w} _ {j} \\}} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} (\\mathbf {x} _ {i} ^ {1}, y _ {i} ^ {t}) + \\ell_ {\\mathrm {S e C u}} (\\mathbf {x} _ {i} ^ {2}, y _ {i} ^ {t})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.399, + 0.462, + 0.415 + ], + "angle": 0, + "content": "Alternatively, \\(\\mathbf{w}\\) can be updated by a closed-form solution" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.426, + 0.409, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {j} ^ {t: B} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} (\\frac {\\sum_ {i : y _ {i} = j} ^ {B} (1 - p _ {i , j}) \\mathbf {x} _ {i} ^ {t}}{\\sum_ {i : y _ {i} = j} ^ {B} 1 - p _ {i , j}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.48, + 0.469, + 0.51 + ], + "angle": 0, + "content": "where \\(B\\) denotes the total number of instances received in the current epoch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.469, + 0.587 + ], + "angle": 0, + "content": "The proposed method is easy to implement and the whole framework is illustrated in Alg. 1. Compared with the supervised discrimination, our main computational overhead is from cluster assignment, which is negligible due to the online strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.603, + 0.209, + 0.62 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The performance of the proposed method is evaluated on CIFAR-10 [21], CIFAR-100 [21], STL-10[10] and ImageNet [27]. To demonstrate the generalization performance of SeCu, we follow the setting in SCAN [14] that trains and evaluates models on the standard train/test split provided by the original data set. For STL-10 that contains an additional noisy unlabeled data set, we first learn the model on the training and noisy set, and then on the target training data as suggested by other methods [14, 32]. The standard metrics for clustering, including clustering accuracy (ACC), normalized mutual information (NMI) and adjusted rand index (ARI) are reported for evaluation. For a fair comparison, we follow the common practice to configure the architecture and training protocol for our method. Concretely, ResNet-50 [18] is applied on ImageNet, while ResNet-18 is adopted for other data sets. The key settings for ResNet-18 are elaborated as follows, while those of ResNet-50 follow [26]. More details can be found in the supplementary." + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.892, + 0.122 + ], + "angle": 0, + "content": "Algorithm 1 Pseudo-code of Stable Cluster Discrimination (SeCu) for One-stage Deep Clustering." + }, + { + "type": "code", + "bbox": [ + 0.502, + 0.131, + 0.88, + 0.416 + ], + "angle": 0, + "content": "f: encoder network\n# w: cluster centers\n# w_p: centers from the last epoch\n# y: list of pseudo one-hot labels (Nx1)\n# tau: weight for labels from last epoch\n# lambda: temperature\n# keep last cluster centers before each epoch\nw_p = w_detach()\n# train one epoch\nfor x in loader: # load a minibatch with b samples\nx_1, x_2 = f(aug(x)), f(aug(x)) # two random views\ny_x = y(x_id) # retrieve label from last epoch\n# compute prediction for each view\np_1 = softmax(x_1 @ w_p / lambda)\np_2 = softmax(x_2 @ w_p / lambda)\n# obtain soft label for discrimination\ny_1 = tau * y_x + (1-tau) * p_2\ny_2 = tau * y_x + (1-tau) * p_1\n# loss for representation learning\nloss_x = (SeCu(p_1, y_1) + SeCu(p_2, y_2)) / 2\n# update prediction for clustering\np_1 = softmax(x_1detach() @ w / lambda)\np_2 = softmax(x_2detach() @ w / lambda)\n# update cluster assignments with constraints\ny(x_id) = y_x = cluster_assign(p_1, p_2)\n# loss for cluster centers\nloss_w = (SeCu(p_1, y_x) + SeCu(p_2, y_x)) / 2\n# update encoder and cluster centers\nloss = loss_x + loss_w\nloss_backward()" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.726 + ], + "angle": 0, + "content": "Architecture Besides backbone network, a 2-layer MLP head is attached and the output dimension is 128. The similar architecture is adopted in [26, 32]. With the MLP head, the total number of parameters is almost the same as the original ResNet-18. The discussion about the MLP head can be found in Sec. 4.1.1. After the MLP head, the learned representation will be classified by a fully-connected (FC) layer that encodes cluster centers. The size of the FC layer is \\(128 \\times K\\). As a clustering method, we let \\(K\\) be the number of ground-truth classes as in [14] and have the direct prediction from the model as cluster assignments for evaluation. Following [1, 14], 10 different classification heads are applied as multi-clustering, which benefits the target clustering. To avoid the problem of selecting an appropriate head for evaluation when 10 heads have the same \\(K\\), we have a varying \\(cK\\) as the number of clusters for each head, where \\(c \\in \\{1, \\dots, 10\\}\\). The prediction from the head with \\(c = 1\\) is reported for comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Optimization To reduce efforts of parameter tuning and reuse the parameter of [26], we have \\(\\mathbf{x}_i^\\top \\mathbf{w}_j\\) in lie of \\(\\log(p_{i,j})\\) in Eqn. 9 and 11 for assignment. Searching the parameter for the latter one can achieve the similar performance. Before training the encoder network, one epoch is adopted to initialize cluster assignments and centers. The encoder network is optimized by SGD with a batch size of 128. The momentum and learning rate are set to be 0.9 and 0.2, respectively. Moreover, the first 10 epochs are applied for warm-up and the cosine decay for learning rate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16650" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "is adopted subsequently. For CIFAR-10, the model is optimized by 400 epochs and cluster centers are learned by SGD with a constant learning rate of 1.2. For other challenging data sets, epochs and learning rate for centers are 800 and 0.8, respectively. \\(\\tau\\) for the soft label as in Eqn. 7 and the temperature \\(\\lambda\\) are fixed as 0.2 and 0.05. For the size constraint, the lower-bound parameter \\(\\gamma\\) and the learning rate of dual variables \\(\\eta_{\\rho}\\) are set to be 0.9 and 0.1, respectively. The only parameter \\(\\alpha\\) for the proposed global entropy constraint is set to be \\(6N / 50\\), where \\(N\\) is the total number of instances. The weight is scaled according to the ablation study on CIFAR-10 as illustrated in Sec. 4.1.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.47, + 0.475 + ], + "angle": 0, + "content": "Augmentation Augmentation is essential for the success of unsupervised representation learning [7]. For a fair comparison, we apply the prevalent settings in existing work [7, 15]. Concretely, random crop, color jitter, random gray scale, Gaussian blur, solarize and random horizontal flip are included as augmentations. Considering that the resolution of images in CIFAR and STL-10 is much smaller than those in ImageNet, we let the minimal scale of random crop be 0.3, 0.2, and 0.2 for CIFAR-10, CIFAR-100, and STL-10, respectively. That for ImageNet is kept as 0.08. Other parameters are the same for all data sets and many recent work share the similar augmentation [9, 15, 26, 32]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.475, + 0.47, + 0.506 + ], + "angle": 0, + "content": "All experiments on small data sets are implemented on a single V100, while 8 GPUs are sufficient for ImageNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.231, + 0.532 + ], + "angle": 0, + "content": "4.1. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.539, + 0.47, + 0.584 + ], + "angle": 0, + "content": "To illustrate the behavior of the proposed method, we study the effect of parameters in SeCu on CIFAR-10 with the size constraint." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.273, + 0.619 + ], + "angle": 0, + "content": "4.1.1 Effect of MLP Head" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.749 + ], + "angle": 0, + "content": "The MLP head shows better generalization than the FC layer in representation learning [7], but its application in deep clustering has not been evaluated systematically. We compare different architectures of MLP head and summarize results in Table 1. Besides the standard MLP layer for projection, there can be another MLP layer attached to the projection MLP head, which is referred as the prediction head [15]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "First, with the standard architecture denoted by 0 for projection and prediction, SeCu is already \\(3.7\\%\\) better than SCAN on ACC and it confirms that the one-stage deep clustering strategy can learn more consistent features than two-stage methods. If including a single layer for SeCu, the performance can be improved by \\(2.1\\%\\), which shows that adding a layer with the reduced dimension in head is helpful for deep clustering. Moreover, due to the dimension reduction from the projection layer, the size of cluster centers decreases and the total number of parameters in the model" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.868, + 0.184 + ], + "angle": 0, + "content": "
#Proj#Pred#ParaACCNMIARI
0011.45M85.576.273.1
1011.30M87.678.776.7
2011.57M88.179.477.6
3011.83M88.179.477.6
3211.97M88.279.677.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.194, + 0.895, + 0.27 + ], + "angle": 0, + "content": "Table 1: Comparison of MLP head on CIFAR-10. \"Proj\" and \"#Pred\" indicate the number of layers in MLP for projection head and prediction head, respectively. ACC (\\%), NMI (\\%) and ARI (\\%) are reported. Baseline ACC of SCAN [14] is \\(81.8\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.437 + ], + "angle": 0, + "content": "is even smaller than that of the original network. When increasing the number of layers and having a 2-layer MLP head, ACC can be further improved by \\(0.5\\%\\) and the model size is only slightly increased. After that, there is no significant gain by applying more complicated MLP head. It is because that the 2-layer MLP head is sufficient for small data sets and we will fix 2-layer MLP as the head for the rest experiments except ImageNet, where 3-layer projection and 2-layer prediction are applied as in [26]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.462, + 0.749, + 0.478 + ], + "angle": 0, + "content": "4.1.2 Effect of Negative Instances" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.551 + ], + "angle": 0, + "content": "After the ablation experiments for the architecture, we study the effect of the proposed stable cluster discrimination loss in Eqn. 3. Table 2 compares it to the standard cross entropy loss (CE) that keeps the gradient from negative instances." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.564, + 0.86, + 0.615 + ], + "angle": 0, + "content": "
Loss#Max#MinACCNMIARI
CE6,4864,50151.650.235.0
SeCu5,1904,55688.179.477.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.67 + ], + "angle": 0, + "content": "Table 2: Comparison of loss functions on CIFAR-10. \"Max\" and \"Min\" indicate the maximal and minimal size of obtained clusters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "First, we can observe that the proposed SeCu loss outperforms the standard cross entropy loss by a large margin of \\(36.5\\%\\) on ACC, which confirms the benefit of stable training. Then, by investigating the distribution of learned clusters, we find that the minimal size of clusters obtained from different loss functions is almost the same due to the lower-bound constraint, while the maximal size varies. The largest cluster obtained by SeCu contains 5,190 instances, which is close to the ground-truth distribution of CIFAR-10. On the contrary, the dominating cluster found by CE loss has 6,486 instances that is significantly different from the uniform distribution in CIFAR-10. It is because that the negative instances in CE loss can perturb the learning of centers and result in sub-optimal cluster assignments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16651" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.108 + ], + "angle": 0, + "content": "4.1.3 Effect of Updating Criterion for Cluster Centers" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.468, + 0.175 + ], + "angle": 0, + "content": "The cluster centers in SeCu can be optimized by SGD or updated by a closed-form solution as in Eqn. 4. We compare these two strategies and summarize the performance in Table 3." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.184, + 0.445, + 0.233 + ], + "angle": 0, + "content": "
MethodsACCNMIARI
SeCu-CF87.4±0.578.5±0.676.4±0.9
SeCu-SGD88.1±0.379.3±0.477.5±0.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.244, + 0.468, + 0.274 + ], + "angle": 0, + "content": "Table 3: Comparison of optimization for cluster centers on CIFAR-10. Mean±std over 8 trails are reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.468, + 0.442 + ], + "angle": 0, + "content": "Let \"SeCu-CF\" and \"SeCu-SGD\" be the closed-form update and the SGD optimization, respectively. We can observe that SeCu-CF has the similar performance to SeCu-SGD, which shows the effectiveness of the closed-form updating. Moreover, SeCu-SGD has a smaller standard deviation than SeCu-CF due to the smoothness from the momentum in SGD. Compared with SeCu-CF, SeCu-SGD has an additional parameter of the learning rate for centers. Hence, SeCu-CF can be applied to evaluate the preliminary performance of SeCu with less tuning efforts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.46, + 0.364, + 0.475 + ], + "angle": 0, + "content": "4.1.4 Effect of \\(\\alpha\\) in Entropy Constraint" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.483, + 0.468, + 0.528 + ], + "angle": 0, + "content": "Besides the size constraint, the proposed entropy constraint is also effective for balancing cluster assignments. Table 4 demonstrates the results with different \\(\\alpha\\)." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.537, + 0.445, + 0.646 + ], + "angle": 0, + "content": "
α#Max#MinACCNMIARI
20,0005,0564,91385.376.373.1
8,0005,1154,65887.979.077.1
6,0005,1204,57988.079.477.4
4,0005,3764,32484.975.472.2
10028,575019.938.717.0
050,000010.00.00.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.657, + 0.443, + 0.672 + ], + "angle": 0, + "content": "Table 4: Comparison of \\( \\alpha \\) in Eqn. 10 on CIFAR-10." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.468, + 0.825 + ], + "angle": 0, + "content": "First, a large \\(\\alpha\\) will regularize the distribution to be uniform and result in a sub-optimal performance. By decreasing the weight, the assignment becomes more flexible and a desired performance can be observed when \\(\\alpha = 6,000\\). However, a small \\(\\alpha\\) will lead to an imbalanced distribution, which is inappropriate for the balanced data sets such as CIFAR and STL. \\(\\alpha = 0\\) discards the constraint and incurs collapsing. Evidently, the entropy constraint can balance the size of clusters effectively" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Since the entropy is defined on the whole data set and CIFAR-10 contains 50,000 instances for training, \\(\\alpha\\) will be scaled as \\(6N / 50\\) for different data sets, where \\(N\\) denotes the total number of instances in the training set. More ablation experiments can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.797, + 0.107 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.235 + ], + "angle": 0, + "content": "After the ablation study, experiments are conducted on benchmark data sets to compare with state-of-the-art methods. The results averaged over 8 trials and the best performance among them are reported for our method. SeCu with size constraint and entropy constraint are denoted as \"SeCu-Size\" and \"SeCu-Entropy\", respectively. Table 5 shows the performance of different methods, where two-stage methods with a pre-training phase are marked for \"Two-stage\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.567 + ], + "angle": 0, + "content": "First, as a one-stage method, SeCu outperforms two-stage methods by a margin of about \\(3\\%\\) over all metrics on CIFAR-10 and CIFAR-100-20. It demonstrates that learning representations and clustering simultaneously is essential for deep clustering. The two-stage method NNM[11] shows the comparable performance to SeCu on STL-10. It is because that STL-10 only constrains 5,000 target instances for training and both of NNM and SeCu already achieve the accuracy of supervised learning. Second, compared with the one-stage method without sufficient representation learning, i.e., PICA, SeCu demonstrates a better performance by about \\(20\\%\\) on CIFAR-10 and CIFAR-100-20, and \\(10\\%\\) on STL-10, which shows the importance of optimizing representations for instances. Third, with the same size constraint, SeCu surpasses CoKe by \\(4.3\\%\\) for ARI on CIFAR-10, which confirms the effectiveness of the proposed learning objective. Finally, compared with size constraint in SeCu-Size, SeCu-Entropy demonstrates a competitive performance over all data sets. Since the entropy constraint is optimized by enumeration without introducing any auxiliary variable, it can be an substitute to the size constraint for cluster assignments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.704 + ], + "angle": 0, + "content": "After the clustering phase, self-labeling [14] is an effective strategy to further improve the performance. We try it for SeCu-Size with the setting in [14] and report the result in Table 5. Interestingly, self-labeling is also effective for SeCu and helps approach the supervised accuracy on CIFAR-10. As a one-stage method, SeCu has the strong augmentation for representation learning, which may introduce additional noise for clustering. Self-labeling has the weak augmentation for fine-tuning and refines clustering." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.719, + 0.689, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.768, + 0.631, + 0.778 + ], + "angle": 0, + "content": "CIFAR-10" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.778, + 0.688, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.842, + 0.623, + 0.853 + ], + "angle": 0, + "content": "STL-10" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.719, + 0.887, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.842, + 0.834, + 0.853 + ], + "angle": 0, + "content": "CIFAR-100-20" + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.866, + 0.833, + 0.881 + ], + "angle": 0, + "content": "Figure 2: Illustration of exemplar images." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16652" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.841, + 0.386 + ], + "angle": 0, + "content": "
MethodsTwo-stageCIFAR10CIFAR100-20STL10
ACCNMIARIACCNMIARIACCNMIARI
Supervised93.886.287.080.068.063.280.665.963.1
DeepCluster [4]37.4--18.9--33.4--
IIC [20]61.751.141.125.722.511.759.649.639.7
PICA [19] (best)69.659.151.233.731.017.171.361.153.1
Pretext [7]+k-means✓65.959.850.939.540.223.965.860.450.6
SCAN [14] (mean)✓81.871.266.542.244.126.775.565.459.0
NNM [11]✓84.374.870.947.748.431.680.869.465.0
GCC [32]✓85.676.472.847.247.230.578.868.463.1
CoKe [26] (mean)85.776.673.249.749.133.5---
SeCu-Size (mean)88.179.377.550.050.735.080.269.463.9
SeCu-Size (best)88.579.978.251.651.636.081.470.765.7
SeCu-Entropy (mean)88.079.277.249.950.634.179.568.763.0
SeCu-Entropy (best)88.279.777.751.251.434.980.569.964.4
with self-labeling:
SCAN [14]✓88.379.777.250.748.633.380.969.864.6
GCC [32]✓90.1--52.3--83.3--
SeCu93.086.185.755.255.139.783.673.369.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.396, + 0.885, + 0.412 + ], + "angle": 0, + "content": "Table 5: Comparison of clustering methods on benchmark data sets. \"Two-stage\" denotes an additional pre-training stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.558 + ], + "angle": 0, + "content": "Exemplar Images Finally, we show exemplar images from clusters obtained by SeCu for different data sets. Fig. 2 illustrates the images that are close to cluster centers in CIFAR-10, CIFAR-100-20, and STL-10, respectively. Evidently, the proposed method can obtain the ground-truth class structure for CIFAR-10 and STL-10. Moreover, with a large intra-class variance, SeCu can still observe appropriate clusters for CIFAR-100-20." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.664 + ], + "angle": 0, + "content": "Considering that labels of super-classes in CIFAR-100-20 contain vague concepts such as \"vehicles 1\" and \"vehicles 2\", we include results on the target 100 classes in Table 6, which may benefit future research. The experiment settings for CIFAR-100 are identical to that for CIFAR-100-20, except the learning rate of centers, which is reduced to 0.01." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.673, + 0.434, + 0.768 + ], + "angle": 0, + "content": "
MethodsACCNMIARI
SeCu-Size (mean)46.461.732.3
SeCu-Size (best)47.161.932.6
SeCu-Entropy (mean)45.360.731.5
SeCu-Entropy (best)46.761.032.2
SeCu†51.365.237.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.778, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Table 6: Results of SeCu on CIFAR-100. † denotes a method with self-labeling." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.312, + 0.849 + ], + "angle": 0, + "content": "4.3. Comparison on ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Now we evaluate SeCu on the challenging ImageNet data set in Table 7, where the cost of many two-stage methods becomes intractable and only available baselines are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.576 + ], + "angle": 0, + "content": "included. First, SeCu achieves \\(51.4\\%\\) ACC on ImageNet, which is better than SCAN with self-labeling by more than \\(11\\%\\). With the same 10 clustering heads and size constraint, SeCu is still \\(3.8\\%\\) better than CoKe on ACC. Compared with the objective in CoKe, SeCu is tailored for deep clustering and the comparison confirms the efficacy of our method. With self-labeling, SeCu shows state-of-the-art performance of \\(53.5\\%\\) ACC on ImageNet. It demonstrates that our method is applicable for large-scale data sets." + }, + { + "type": "table", + "bbox": [ + 0.581, + 0.588, + 0.813, + 0.667 + ], + "angle": 0, + "content": "
MethodsACCNMIARI
SCAN†39.972.027.5
CoKe‡47.676.235.6
SeCu51.478.039.7
SeCu†53.579.441.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.678, + 0.892, + 0.71 + ], + "angle": 0, + "content": "Table 7: Comparison on ImageNet. \\( {}^{ \\dagger } \\) denotes a method with self-labeling. \\( {}^{ \\ddagger } \\) is our reproduction with 10 clustering heads." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.619, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, a novel stable cluster discrimination task that considers the hardness of instances is proposed for one-stage deep clustering. To avoid collapsing in representations, a global entropy constraint with theoretical guarantee is investigated. Finally, the comparison with state-of-the-art methods confirms the effectiveness of our proposal. While recent work shows that leveraging nearest neighbors improves deep clustering [11, 14, 32], incorporating such information in SeCu can be our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16653" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.157 + ], + "angle": 0, + "content": "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.468, + 0.2 + ], + "angle": 0, + "content": "[2] Hangbo Bao, Li Dong, and Furu Wei. Beit: BERT pretraining of image transformers. CoRR, abs/2106.08254, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.468, + 0.231 + ], + "angle": 0, + "content": "[3] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.468, + 0.272 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.275, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.333, + 0.468, + 0.387 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.39, + 0.468, + 0.444 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, volume 119, pages 1597-1607, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.447, + 0.468, + 0.488 + ], + "angle": 0, + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758. Computer Vision Foundation / IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.49, + 0.468, + 0.53 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. CoRR, abs/2104.02057, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.468, + 0.574 + ], + "angle": 0, + "content": "[10] Adam Coates, Andrew Y. Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In AISTATS, volume 15, pages 215-223. JMLR.org, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.468, + 0.63 + ], + "angle": 0, + "content": "[11] Zhiyuan Dang, Cheng Deng, Xu Yang, Kun Wei, and Heng Huang. Nearest neighbor matching for deep clustering. In CVPR, pages 13693-13702. Computer Vision Foundation / IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[13] Ehsan Elhamifar and René Vidal. Sparse subspace clustering: Algorithm, theory, and applications. IEEE Trans. Pattern Anal. Mach. Intell., 35(11):2765-2781, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[14] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. SCAN: learning to classify images without labels. In ECCV, volume 12355, pages 268-285. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[15] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent - A new approach to self-supervised learning. In NeurIPS, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[19] Jiabo Huang, Shaogang Gong, and Xiatian Zhu. Deep semantic clustering by partition confidence maximisation. In CVPR, pages 8846-8855. Computer Vision Foundation / IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[20] Xu Ji, Andrea Vedaldi, and João F. Henriques. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9864-9873. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[21] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, pages 1106-1114, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[23] Yunfan Li, Peng Hu, Jerry Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In AAAI, pages 8547-8555. AAAI Press, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[24] Stuart P. Lloyd. Least squares quantization in PCM. IEEE Trans. Inf. Theory, 28(2):129-136, 1982." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.478, + 0.892, + 0.519 + ], + "angle": 0, + "content": "[25] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In ICCV, pages 6449-6457, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.522, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[26] Qi Qian, Yuanhong Xu, Juhua Hu, Hao Li, and Rong Jin. Unsupervised visual representation learning by online constrained k-means. In CVPR, pages 16619-16628. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.578, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[27] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Fei-Fei Li. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis., 115(3):211-252, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.649, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[28] Ulrike von Luxburg. A tutorial on spectral clustering. Stat. Comput., 17(4):395-416, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.677, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[29] Junyuan Xie, Ross B. Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In ICML, volume 48, pages 478-487. JMLR.org, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[30] Bo Yang, Xiao Fu, Nicholas D. Sidiropoulos, and Mingyi Hong. Towards k-means-friendly spaces: Simultaneous deep learning and clustering. In ICML, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[31] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, volume 139, pages 12310-12320. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.874 + ], + "angle": 0, + "content": "[32] Huasong Zhong, Jianlong Wu, Chong Chen, Jianqiang Huang, Minghua Deng, Liqiang Nie, Zhouchen Lin, and Xian-Sheng Hua. Graph contrastive clustering. In ICCV, pages 9204-9213. IEEE, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.874 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16654" + } + ] +] \ No newline at end of file diff --git a/2023/Stable Cluster Discrimination for Deep Clustering/9f881270-2ff5-4f87-a2db-8feef1f47071_origin.pdf b/2023/Stable Cluster Discrimination for Deep Clustering/9f881270-2ff5-4f87-a2db-8feef1f47071_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9cd695cc8b3edd64a1aa5c5af3280df79c4ff11b --- /dev/null +++ b/2023/Stable Cluster Discrimination for Deep Clustering/9f881270-2ff5-4f87-a2db-8feef1f47071_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd974cfaf320aca275082f3cb6c8447feea7aef860fbfc671dc0c4f1ef700563 +size 521526 diff --git a/2023/Stable Cluster Discrimination for Deep Clustering/full.md b/2023/Stable Cluster Discrimination for Deep Clustering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7098f35ed1200b78e1ad278bfcd5db9345ffdc47 --- /dev/null +++ b/2023/Stable Cluster Discrimination for Deep Clustering/full.md @@ -0,0 +1,452 @@ +# Stable Cluster Discrimination for Deep Clustering + +Qi Qian +Alibaba Group, Bellevue, WA 98004, USA +qi.qian@alibaba-inc.com + +# Abstract + +Deep clustering can optimize representations of instances (i.e., representation learning) and explore the inherent data distribution (i.e., clustering) simultaneously, which demonstrates a superior performance over conventional clustering methods with given features. However, the coupled objective implies a trivial solution that all instances collapse to the uniform features. To tackle the challenge, a two-stage training strategy is developed for decoupling, where it introduces an additional pre-training stage for representation learning and then fine-tunes the obtained model for clustering. Meanwhile, one-stage methods are developed mainly for representation learning rather than clustering, where various constraints for cluster assignments are designed to avoid collapsing explicitly. Despite the success of these methods, an appropriate learning objective tailored for deep clustering has not been investigated sufficiently. In this work, we first show that the prevalent discrimination task in supervised learning is unstable for one-stage clustering due to the lack of ground-truth labels and positive instances for certain clusters in each mini-batch. To mitigate the issue, a novel stable cluster discrimination (SeCu) task is proposed and a new hardness-aware clustering criterion can be obtained accordingly. Moreover, a global entropy constraint for cluster assignments is studied with efficient optimization. Extensive experiments are conducted on benchmark data sets and ImageNet. SeCu achieves state-of-the-art performance on all of them, which demonstrates the effectiveness of one-stage deep clustering. + +# 1. Introduction + +Clustering is a fundamental task in unsupervised learning. Given features of instances, the unlabeled data set will be partitioned into multiple clusters, where instances from the same cluster are similar according to the measurement defined by a distance function. With fixed features, most of research efforts focus on studying appropriate distance functions and ingenious algorithms have been proposed by different measurements, e.g., k-means clustering [24], spec + +![](images/7d4651c447ce90e196cf23908d71313dffa6dc0b8fd10b8d4047a7253127216a.jpg) +Figure 1: An illustration of the proposed method. 10 data points are randomly sampled from two different Gaussian distributions, respectively. Points with the same color are from the same distribution, while squares denote the corresponding cluster centers obtained by different methods. Star indicates the mis-classified data. Unlike k-means that assigns the uniform weight for different instances, our method considers the hardness of instances and is better for discrimination based clustering. + +![](images/9c1fa25d5bce1822d8cfe9a6b49da87f78229c6e3b50478b87e1a82cd3551977.jpg) + +tral clustering [28], subspace clustering [13], etc. + +With the development of deep neural networks, deep learning is capable of learning representations from raw materials and demonstrates the dominating performance on various supervised tasks [22]. Thereafter, it is also introduced to clustering recently [4, 11, 14, 19, 20, 23, 29, 30, 32]. Unlike the conventional clustering, representations of instances are learned with cluster assignments and centers simultaneously in deep clustering, which is more flexible to capture the data distribution. However, the coupled objective can result in a trivial solution that all instances collapse to the uniform features [4] and designing appropriate clustering criterion in the new scenario becomes challenging. + +Many deep clustering algorithms [11, 14, 32] exploit a two-stage training strategy to decouple representation learning and clustering to avoid collapsing. The recent progress in unsupervised representation learning [2, 4, 7, 16, 17, 26] shows that informative features capturing the semantic similarity between instances without collapsing can be obtained by pre-training a large set of unlabeled data. Inspired by the observation, those methods leverage a pre-training stage and then focus on developing algorithms to optimize clus + +tering by fine-tuning a pre-trained model in the second stage. By refining the relations between nearest neighbors obtained from the pre-training stage, two-stage methods achieve a significantly better performance than one-stage ones without sufficient representation learning [19, 20]. + +However, the objective of pre-training can be inconsistent with that of clustering, which results in a sub-optimal performance for deep clustering. Note that different from supervised learning where the objective is explicitly defined by labels, that for unsupervised representation learning is arbitrary and various pretext tasks have been proposed, e.g., instance discrimination [7, 17], cluster discrimination [4, 26], masked modeling [2, 16], etc. Most of existing two-stage clustering methods adopt instance discrimination, i.e., SimCLR [7], for pre-training, whereas it aims to identify each instance as an individual class and the objective is different from clustering that aims to group instances from the same cluster. Moreover, SwAV [5] demonstrates that clustering itself is also an effective pretext task for representation learning. Therefore, we focus on facilitating one-stage deep clustering that optimizes representations and clustering simultaneously in this work. + +Most of existing one-stage methods are proposed solely for representation learning. To tackle the collapsing issue, research efforts are mainly devoted to developing appropriate constraints for cluster assignments, especially for online deep clustering. For example, [1, 5] apply the balanced constraint that each cluster has the same number of instances for clustering. [26] further relaxes the balanced constraint to a lower-bound size constraint that limits the minimal size of clusters and demonstrates a more flexible assignment. + +After obtaining cluster assignments as pseudo labels, representations and cluster centers can be optimized as in the supervised scenario. For example, [5, 6] learn the encoder network and cluster centers by solving a classification problem with the standard cross entropy loss. [26] has the same classification task for representation learning while adopting k-means for updating cluster centers. Although these methods achieve a satisfied performance on representation learning, a learning objective tailored for deep clustering has not attracted sufficient attentions. + +In this work, we investigate the effective learning task for one-stage deep clustering. By analyzing the standard cross entropy loss for supervised learning, we find that it can be unstable for unsupervised learning. Concretely, the gradient for updating cluster centers consists of two ingredients: gradient from positive instances of the target cluster and that from irrelevant negative ones. However, with a limited mini-batch size in stochastic gradient descent (SGD), there can be no positive ones for a large proportion of clusters (e.g., $90\%$ on ImageNet) at each iteration. Due to the lack of positive instances, the influence from negative ones is dominating. Unlike supervised learning where labels are fixed, + +the cluster assignments can change during training in deep clustering. Therefore, the noise from the large variance of negative instances will be accumulated, which makes the optimization unstable. + +To mitigate the problem, we propose a stable cluster discrimination (SeCu) task for deep clustering, which stops the gradient from negative instances for updating cluster centers in the cross entropy loss. Compared with k-means in [26], where positive instances have the uniform weight for updating centers, SeCu as a discrimination task considers the hardness of instances and a large weight will be assigned to hard instances for updating. Fig. 1 illustrates the hardness-aware clustering criterion implied by SeCu. Besides, we improve the cluster assignment by developing an entropy constraint that regularizes the entropy of assignments over the entire data set. Compared to the optimization with the size constraint [26], our method can reduce the number of variables and hyper-parameters, and thus make the learning more convenient. The main contributions of this work can be summarized as follows. + +- A novel task is proposed for one-stage deep clustering. SeCu tailors the supervised cross entropy loss by eliminating the influence from the negative instances in learning cluster centers, which makes the training stable when ground-truth labels are unavailable. +- A global entropy constraint is exploited to balance the size of clusters and an efficient closed-form solution is developed for online assignment with the constraint. +- A simple framework with a single loss function and encoder is introduced for deep clustering in Eqn. 6. The proposed method is evaluated with the standard protocol on benchmark data sets and ImageNet [27]. The superior performance of SeCu confirms its effectiveness for deep clustering. + +# 2. Related Work + +In this section, we briefly review unsupervised representation learning and then focus on clustering. + +# 2.1. Unsupervised Representation Learning + +By leveraging massive unlabeled data, unsupervised representation learning can learn a pre-trained model that encoders the semantic information from data and helps downstream tasks with fine-tuning. Compared with supervised learning containing labels, one major challenge in unsupervised learning is defining appropriate positive pairs for representation learning. SimCLR [7] and MoCo [17] apply instance discrimination as the objective that generates positive pairs from diverse views of the same instance and considers other instances as the negative ones. Then, BYOL [15] + +and SimSiam [8] demonstrate that with appropriate neural network architectures, the applicable models can be obtained with only positive pairs. Besides the work optimizing instance space, Barlow Twins [31] defines positive pairs in feature space and also learns effective pre-trained models. Recently, learning with masked modeling [2, 16] achieves success on the specific backbone of vision transformer [12] and shows the better fine-tuning performance than the generic objective on the ResNet [18]. + +# 2.2. Deep Clustering + +Compared with instance discrimination, cluster discrimination aims to obtain positive pairs consisting of different instances for representation learning. Therefore, it has to learn representations and relations between instances simultaneously. DeepCluster [4] obtains membership of instances with an offline k-means and then optimizes representations alternately. Some work focuses on representation learning and introduces the balanced constraint that each cluster has the same number of instances for cluster assignments [1, 5] to mitigate the collapsing problem. Besides, DINO [6] applies an additional momentum encoder to further improve the representation. However, the constraint assumes a well balanced distribution, which is hard to capture the ground-truth data distribution. Thereafter, CoKe [26] relaxes the constraint to only lower-bound the minimal size of clusters to model the distribution more flexibly. + +Besides representation learning, there are many works developed solely for better clustering [11, 14, 20, 23, 29, 30, 32]. To handle the collapsing issue, two-stage methods [11, 14, 32] leverage the representations from a pretrained model to obtain nearest neighbors and fine-tune the model for clustering accordingly. In addition, the objective of pre-training can be included for clustering as multitask learning [23]. The work closest to ours is CoKe [26] that optimizes representations and clustering simultaneously. Compared with the conventional k-means in CoKe, we propose a novel stable cluster discrimination objective tailored for one-stage deep clustering. Moreover, an entropy constraint is introduced to reduce the number of parameters while demonstrating the competitive performance. + +# 3. Stable Cluster Discrimination + +# 3.1. Cluster Discrimination + +Given a data set with $N$ images $\{x_{i}\}_{i = 1}^{N}$ , if the corresponding labels are accessible as $\{y_{i}\}_{i = 1}^{N}$ , the literature from distance metric learning [25] shows that appropriate representations of examples can be learned by optimizing a classification task as + +$$ +\min _ {\theta} \sum_ {i = 1} ^ {N} \ell \left(x _ {i}, y _ {i}; \theta\right) +$$ + +where $\ell (\cdot)$ is the cross entropy loss with the normalized Softmax operator as suggested in [25]. $\theta$ denotes the parameters of the deep neural network. + +For unsupervised learning that the label information is unavailable, the objective of cluster discrimination with $K$ clusters can be written as + +$$ +\min _ {\theta_ {f}, \left\{\mathbf {w} _ {j} \right\}, y _ {i} \in \Delta} \mathcal {L} = \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {K} - y _ {i, j} \log \left(p _ {i, j}\right) \tag {1} +$$ + +where $y_{i}$ denotes the learnable label for $x_{i}$ and $\Delta = \{y_i|\sum_{j = 1}^{K}y_{i,j} = 1,\forall j,y_{i,j}\in \{0,1\} \}$ . With the normalized Softmax operator, the prediction $p_{i,j}$ can be computed + +$$ +p _ {i, j} = \frac {\exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {j} / \lambda\right)}{\sum_ {k = 1} ^ {K} \exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {k} / \lambda\right)} \tag {2} +$$ + +where $\mathbf{x}_i = f(x_i)$ and $f(\cdot)$ denotes the encoder network. $\theta_{f}$ contains parameters of $f$ . $\{\mathbf{w}_j\}_{j=1}^K$ consists of cluster centers. $\lambda$ is the parameter of temperature and $\forall i, j, \|\mathbf{x}_i\|_2 = \|\mathbf{w}_j\|_2 = 1$ by normalization. + +Compared with the supervised counterpart, the problem in Eqn. 1 has to optimize cluster assignments $\{y\}$ , cluster centers $\{\mathbf{w}\}$ , and representation encoder network $f$ simultaneously. While most of existing work focuses on optimizing $\{y\}$ with different constraints to avoid collapsing, this work investigates the learning objective for $\{\mathbf{w}\}$ and $f$ , and a new clustering criterion is introduced accordingly. + +# 3.2. Stable Loss for Optimization with Mini-batch + +Unlike supervised learning, where labels are fixed for examples, cluster assignments $\{y\}$ in unsupervised learning are dynamic with the training of instance representations and cluster centers. Therefore, the original cross entropy loss becomes unstable for unsupervised discrimination. The problem can be elaborated by analyzing the updating criterion for cluster centers. + +Letting $y_{i}$ denote the label of $\mathbf{x}_i$ , the gradient of $\mathbf{w}_j$ from the standard cross entropy loss in Eqn. 1 can be computed as + +$$ +\nabla_ {\mathbf {w} _ {j}} \mathcal {L} = \frac {1}{\lambda} \left(\sum_ {i: y _ {i} = j} \left(p _ {i, j} - 1\right) \mathbf {x} _ {i} + \sum_ {k: y _ {k} \neq j} p _ {k, j} \mathbf {x} _ {k}\right) +$$ + +The former term is to pull the center $\mathbf{w}_j$ to the assigned instances and the latter one is to push it away from instances of other clusters. However, the following Propositions show that this updating is unstable for deep clustering. The detailed proof can be found in the supplementary. + +Proposition 1. When sampling a mini-batch of $b$ instances for $K$ clusters, there are no positive instances for at least $K - b$ clusters. + +Proposition 2. Let $\mathrm{Var}_{pos}$ and $\mathrm{Var}_{neg}$ be the variance of sampled positive and negative instances, respectively. Assuming that each instance has unit norm and the norm of the cluster mean is $a$ , we have $\mathrm{Var}_{neg} = \mathcal{O}\left(\frac{1}{1 - a^2}\right)\mathrm{Var}_{pos}$ . + +Remark Proposition 1 implies that with cross entropy loss, a large proportion of cluster centers will be updated only by negative instances. Proposition 2 indicates that the variance of sampled negative instances is much larger than that of positive ones when each cluster is compact as $a$ approaching 1. Due to the small size of a mini-batch in training deep neural networks, the variance cannot be reduced sufficiently. + +When $K = 10,000$ and $b = 1,024$ as in our multiclustering settings for ImageNet, at least about $90\%$ centers will only access a mini-batch of negative instances at each iteration. Without ground-truth labels, the bias in updating will be accumulated and mislead the learning process. + +To mitigate the problem, we propose to eliminate the direction of negative instances from gradient for stable training as + +$$ +\nabla_ {\mathbf {w} _ {j}} \mathcal {L} = \frac {1}{\lambda} \sum_ {i: y _ {i} = j} (p _ {i, j} - 1) \mathbf {x} _ {i} +$$ + +The corresponding stable cluster discrimination loss becomes + +$$ +\begin{array}{l} \ell_ {\mathrm {S e C u}} \left(x _ {i}, y _ {i}\right) = \\ - \log \left(\frac {\exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {y _ {i}} / \lambda\right)}{\exp \left(\mathbf {x} _ {i} ^ {\top} \mathbf {w} _ {y _ {i}} / \lambda\right) + \sum_ {k : k \neq y _ {i}} \exp \left(\mathbf {x} _ {i} ^ {\top} \tilde {\mathbf {w}} _ {k} / \lambda\right)}\right) \tag {3} \\ \end{array} +$$ + +where $\tilde{\mathbf{w}}_k$ denotes $\mathbf{w}_k$ with the stop-gradient operator. Compared with the standard cross entropy loss, cluster centers in SeCu loss will be updated only by positive instances, which is more stable for deep clustering when optimizing with mini-batches. + +In addition, the analysis in the following theorem demonstrates the novel hardness-aware clustering criterion for deep clustering implied by the proposed objective. + +Theorem 1. When fixing $\{y_i\}$ and $\{\mathbf{x}_i\}$ , let $\{\mathbf{w}^*\}$ be the optimal solution of the problem with loss function in Eqn. 3 and assume $\forall i, \| \mathbf{x}_i\|_2 = 1; \forall j, \| \mathbf{w}_j\|_2 = 1$ , then we have + +$$ +\mathbf {w} _ {j} ^ {*} = \Pi_ {\| \mathbf {w} \| _ {2} = 1} \left(\frac {\sum_ {i : y _ {i} = j} (1 - p _ {i , j}) \mathbf {x} _ {i}}{\sum_ {i : y _ {i} = j} 1 - p _ {i , j}}\right) \tag {4} +$$ + +where $\Pi_{\| \mathbf{w}\| _2 = 1}$ projects the vector to the unit norm. + +Remark For k-means in CoKe [26], cluster centers will be averaged over all assigned instances with the uniform weight + +$$ +\mathbf {w} _ {j} = \Pi_ {\| \mathbf {w} \| _ {2} = 1} \left(\frac {\sum_ {i : y _ {i} = j} \mathbf {x} _ {i}}{\sum_ {i} 1 (y _ {i} = j)}\right) \tag {5} +$$ + +where $1(\cdot)$ is the indicator function. On the contrary, our proposal considers the weight of each instance by its hardness, i.e., $p_{i,y_i}$ . By assigning a large weight to hard instances, the corresponding center can capture the distribution of hard instances better, which is illustrated in Fig. 1. + +Our formulation also implies CoKe as a special case. Concretely, by increasing the temperature $\lambda$ for updating $\mathbf{w}$ to be infinite, we have $p_{i,j} = 1 / K$ and Eqn. 4 will degenerate to Eqn. 5. + +Finally, besides SGD, Eqn. 4 also suggests an alternative updating strategy for $\mathbf{w}$ . Since $p_{i,j}$ is computed by $\mathbf{w}$ , it may still require multiple iterations with Eqn. 4 for convergence. However, the representations of instances are also updated by training and we can keep a single update for centers at each iteration and improve centers along with the learning of representations. Compared with SGD for learning cluster centers, the strategy eliminates the learning rate for centers with a comparable performance as shown in the ablation study. + +Given the learning objective SeCu, we will elaborate the proposed deep clustering method in next subsection. + +# 3.3. SeCu for Deep Clustering + +With the proposed loss function, the objective of stable cluster discrimination for deep clustering can be written as + +$$ +\min _ {\theta_ {f}, \left\{\mathbf {w} _ {j} \right\}, y _ {i} \in \Delta} \sum_ {i = 1} ^ {N} \ell_ {\mathrm {S e C u}} \left(x _ {i}, y _ {i}\right) +$$ + +$$ +s. t. \quad h _ {m} (Y) \geq b _ {m}, \quad m = 1, \dots , M \tag {6} +$$ + +where $Y = [y_{1},\ldots ,y_{N}]$ , $M$ denotes the number of constraints, and $h_m(\cdot)$ is the $m$ -th constraint for cluster assignment. Considering that variables are entangled, the problem is solved alternately. + +Update of $\theta_{f}$ First, when fixing $\{y^{t - 1}\}$ and $\{\mathbf{w}^{t - 1}\}$ from the last epoch, the sub-problem for representation learning in the $t$ -th epoch becomes + +$$ +\min _ {\theta_ {f}} \sum_ {i = 1} ^ {N} \ell_ {\mathrm {S e C u}} (x _ {i}, y _ {i} ^ {t - 1}) +$$ + +where the SeCu loss degenerates to the standard cross entropy loss with fixed $\{\mathbf{w}\}$ and can be optimized with SGD. The one-hot label $y^{t - 1}$ is kept from the $(t - 1)$ -th epoch, which makes the optimization consistent between adjacent epochs but the updating for representations may be delayed. To incorporate the information from the current epoch, we include two views of augmentations for the individual instance at each iteration, which is prevalent for representation learning [7, 17, 26]. + +Let $x_{i}^{1}$ and $x_{i}^{2}$ be two perturbed views of the original image with random augmentations, and the prediction is + +$$ +p _ {i, j} ^ {\mathrm {s}} = \frac {\exp (\mathbf {x} _ {i} ^ {\mathrm {s} \top} \mathbf {w} _ {j} ^ {t - 1} / \lambda)}{\sum_ {j} ^ {K} \exp (\mathbf {x} _ {i} ^ {\mathrm {s} \top} \mathbf {w} _ {j} ^ {t - 1} / \lambda)}; s = \{1, 2 \} +$$ + +where $\mathbf{x}_i^s = f_t(x_i^s)$ is extracted by the current encoder. The soft label for each view can be obtained as [26] + +$$ +y _ {i} ^ {1} = \tau y _ {i} ^ {t - 1} + (1 - \tau) p _ {i} ^ {2}; y _ {i} ^ {2} = \tau y _ {i} ^ {t - 1} + (1 - \tau) p _ {i} ^ {1} (7) +$$ + +The soft label contains the prediction from the other view of the same instance, which optimizes the consistency between different views at the same iteration. The problem with two-view optimization can be written as + +$$ +\min _ {\theta_ {f}} \sum_ {i = 1} ^ {N} \ell_ {\mathrm {S e C u}} (x _ {i} ^ {1}, y _ {i} ^ {1}) + \ell_ {\mathrm {S e C u}} (x _ {i} ^ {2}, y _ {i} ^ {2}) +$$ + +Update of $y$ When fixing $\mathbf{x}_i^t$ and $\{\mathbf{w}^t\}$ , the cluster assignment can be updated by solving the problem + +$$ +\min _ {y _ {i} \in \Delta} - \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {K} y _ {i, j} \log (p _ {i, j}) +$$ + +$$ +s. t. \quad h _ {m} (Y) \geq b _ {m}, \quad m = 1, \dots , M +$$ + +where $p_{i,j}$ is defined by $\mathbf{x}_i^t$ and $\mathbf{w}^t$ as in Eqn. 2. Without the constraints $\{h_m\}$ , the objective implies a greedy solution that assigns each instance to the most related cluster. It can incur the trivial solution of collapsing, especially for online deep clustering, where each instance can only be accessed once in each epoch and cluster assignments cannot be refined with multiple iterations over the entire data. Therefore, many effective constraints are developed to mitigate the challenge. + +First, we investigate the size constraint that is prevalent in existing methods [5, 26]. Following [26], the cluster size can be lower-bounded as + +$$ +\min _ {y _ {i} \in \Delta} - \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {K} y _ {i, j} \log (p _ {i, j}) +$$ + +$$ +s. t. \quad \sum_ {i} y _ {i, j} \geq \gamma N / K, \quad j = 1, \dots , K \tag {8} +$$ + +where $\gamma$ is the proportion to the average size and $\gamma = 1$ implies the balanced constraint. Let $\rho_{j}$ denote dual variables for the $j$ -th lower-bound constraint. The problem becomes + +$$ +\max _ {\rho : \rho \geq 0} \min _ {y _ {i} \in \Delta} - \sum_ {i} \sum_ {j} y _ {i, j} \log (p _ {i, j}) - \sum_ {j} \rho_ {j} (\sum_ {i} y _ {i, j} - \gamma N / K) +$$ + +When a mini-batch of instances arrive, the cluster assignment for each instance can be obtained via a closed-form solution + +$$ +y _ {i, j} ^ {t} = \left\{ \begin{array}{l l} 1 & j = \arg \min _ {j} - \log \left(p _ {i, j}\right) - \rho_ {j} \\ 0 & o. w. \end{array} \right. \tag {9} +$$ + +Then, the dual variables will be updated by stochastic gradient ascent. More details can be found in [26]. + +The size constraint is capable of avoiding collapsing problem explicitly, but it introduces additional dual variables to help assignment and has to optimize them in training. To simplify the optimization, a global entropy constraint is investigated to balance the distribution over all clusters. Compared with the size constraint, the bound for the cluster size is implicit with the entropy constraint. Nevertheless, as we will demonstrate, it can help reduce the number of variables and hyper-parameters, which is more friendly for users. + +Given the set of cluster assignments $\{y\}$ , the entropy of the whole data set can be defined as + +$$ +H (y) = - \sum_ {j = 1} ^ {K} \frac {\sum_ {i} ^ {N} y _ {i , j}}{N} \log (\frac {\sum_ {i} ^ {N} y _ {i , j}}{N}) +$$ + +With the entropy as the regularization, the objective can be written as + +$$ +\min _ {y _ {i} \in \Delta} - \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {K} y _ {i, j} \log (p _ {i, j}) \quad s. t. \quad H (y) \geq \gamma +$$ + +Compared with the problem consisting of $K$ constraints in Eqn. 8, there is only a single constraint that controls the size of all clusters simultaneously. According to the dual theory [3], the problem is equivalent to + +$$ +\min _ {y _ {i} \in \Delta} - \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {K} y _ {i, j} \log \left(p _ {i, j}\right) - \alpha H (y) \tag {10} +$$ + +Since one-hot labels of all instances are kept in memory, the optimal solution can be obtained by enumerating. When optimizing the label for $x_{i}$ and fixing other labels, the closed-form solution for $y_{i}$ is + +$$ +y _ {i, j} ^ {t} = \left\{ \begin{array}{l l} 1 & j = \arg \min _ {j} - \log \left(p _ {i, j}\right) - \alpha H \left(y ^ {t - 1}, y _ {i: j}\right) \\ 0 & o. w. \end{array} \right. \tag {11} +$$ + +where $H(y^{t - 1},y_{i:j})$ replaces the previous label $y_{i}$ of $\mathbf{x}_i$ by $j$ as $H(y^{t - 1},y_{i:j}) = H([y_1^{t - 1},\ldots ,y_{i - 1}^{t - 1},y_{i + 1}^{t - 1},\ldots ,y_n^{t - 1};y_{i,j} = 1])$ . The convergence is guaranteed as in the following theorem. + +Theorem 2. Let $\mathcal{L}(Y)$ denote the objective in Eqn. 10. If updating cluster assignments sequentially according to Eqn. 11, we have $\mathcal{L}(Y^t) \leq \mathcal{L}(Y^{t-1})$ . + +Proof. It is directly from the optimality of the closed-form solution in Eqn. 11. $\square$ + +Remark While some deep clustering methods apply the entropy regularization for learning, it is defined over a minibatch of instances and optimized by SGD [14]. On the contrary, our method leverages the entropy of the entire data set that can capture the global distribution better. Moreover, the entropy in our method can be maximized with a closed-form solution, which is more efficient for learning. + +With the two-view optimization, the updating becomes + +$$ +j = \arg \min _ {j} - (\log (p _ {i, j} ^ {1}) + \log (p _ {i, j} ^ {2})) / 2 - \alpha H (y ^ {t - 1}, y _ {i: j}) +$$ + +Update of w With fixed $\mathbf{x}_i^t$ and the updated pseudo label $\{y^{t}\}$ , cluster centers can be optimized by minimizing SeCu loss over two views of augmentations by SGD + +$$ +\min _ {\{\mathbf {w} _ {j} \}} \sum_ {i = 1} ^ {N} \ell_ {\mathrm {S e C u}} (\mathbf {x} _ {i} ^ {1}, y _ {i} ^ {t}) + \ell_ {\mathrm {S e C u}} (\mathbf {x} _ {i} ^ {2}, y _ {i} ^ {t}) +$$ + +Alternatively, $\mathbf{w}$ can be updated by a closed-form solution + +$$ +\mathbf {w} _ {j} ^ {t: B} = \Pi_ {\| \mathbf {w} \| _ {2} = 1} (\frac {\sum_ {i : y _ {i} = j} ^ {B} (1 - p _ {i , j}) \mathbf {x} _ {i} ^ {t}}{\sum_ {i : y _ {i} = j} ^ {B} 1 - p _ {i , j}}) +$$ + +where $B$ denotes the total number of instances received in the current epoch. + +The proposed method is easy to implement and the whole framework is illustrated in Alg. 1. Compared with the supervised discrimination, our main computational overhead is from cluster assignment, which is negligible due to the online strategy. + +# 4. Experiments + +The performance of the proposed method is evaluated on CIFAR-10 [21], CIFAR-100 [21], STL-10[10] and ImageNet [27]. To demonstrate the generalization performance of SeCu, we follow the setting in SCAN [14] that trains and evaluates models on the standard train/test split provided by the original data set. For STL-10 that contains an additional noisy unlabeled data set, we first learn the model on the training and noisy set, and then on the target training data as suggested by other methods [14, 32]. The standard metrics for clustering, including clustering accuracy (ACC), normalized mutual information (NMI) and adjusted rand index (ARI) are reported for evaluation. For a fair comparison, we follow the common practice to configure the architecture and training protocol for our method. Concretely, ResNet-50 [18] is applied on ImageNet, while ResNet-18 is adopted for other data sets. The key settings for ResNet-18 are elaborated as follows, while those of ResNet-50 follow [26]. More details can be found in the supplementary. + +Algorithm 1 Pseudo-code of Stable Cluster Discrimination (SeCu) for One-stage Deep Clustering. +```txt +f: encoder network +# w: cluster centers +# w_p: centers from the last epoch +# y: list of pseudo one-hot labels (Nx1) +# tau: weight for labels from last epoch +# lambda: temperature +# keep last cluster centers before each epoch +w_p = w_detach() +# train one epoch +for x in loader: # load a minibatch with b samples +x_1, x_2 = f(aug(x)), f(aug(x)) # two random views +y_x = y(x_id) # retrieve label from last epoch +# compute prediction for each view +p_1 = softmax(x_1 @ w_p / lambda) +p_2 = softmax(x_2 @ w_p / lambda) +# obtain soft label for discrimination +y_1 = tau * y_x + (1-tau) * p_2 +y_2 = tau * y_x + (1-tau) * p_1 +# loss for representation learning +loss_x = (SeCu(p_1, y_1) + SeCu(p_2, y_2)) / 2 +# update prediction for clustering +p_1 = softmax(x_1detach() @ w / lambda) +p_2 = softmax(x_2detach() @ w / lambda) +# update cluster assignments with constraints +y(x_id) = y_x = cluster_assign(p_1, p_2) +# loss for cluster centers +loss_w = (SeCu(p_1, y_x) + SeCu(p_2, y_x)) / 2 +# update encoder and cluster centers +loss = loss_x + loss_w +loss_backward() +``` + +Architecture Besides backbone network, a 2-layer MLP head is attached and the output dimension is 128. The similar architecture is adopted in [26, 32]. With the MLP head, the total number of parameters is almost the same as the original ResNet-18. The discussion about the MLP head can be found in Sec. 4.1.1. After the MLP head, the learned representation will be classified by a fully-connected (FC) layer that encodes cluster centers. The size of the FC layer is $128 \times K$ . As a clustering method, we let $K$ be the number of ground-truth classes as in [14] and have the direct prediction from the model as cluster assignments for evaluation. Following [1, 14], 10 different classification heads are applied as multi-clustering, which benefits the target clustering. To avoid the problem of selecting an appropriate head for evaluation when 10 heads have the same $K$ , we have a varying $cK$ as the number of clusters for each head, where $c \in \{1, \dots, 10\}$ . The prediction from the head with $c = 1$ is reported for comparison. + +Optimization To reduce efforts of parameter tuning and reuse the parameter of [26], we have $\mathbf{x}_i^\top \mathbf{w}_j$ in lie of $\log(p_{i,j})$ in Eqn. 9 and 11 for assignment. Searching the parameter for the latter one can achieve the similar performance. Before training the encoder network, one epoch is adopted to initialize cluster assignments and centers. The encoder network is optimized by SGD with a batch size of 128. The momentum and learning rate are set to be 0.9 and 0.2, respectively. Moreover, the first 10 epochs are applied for warm-up and the cosine decay for learning rate + +is adopted subsequently. For CIFAR-10, the model is optimized by 400 epochs and cluster centers are learned by SGD with a constant learning rate of 1.2. For other challenging data sets, epochs and learning rate for centers are 800 and 0.8, respectively. $\tau$ for the soft label as in Eqn. 7 and the temperature $\lambda$ are fixed as 0.2 and 0.05. For the size constraint, the lower-bound parameter $\gamma$ and the learning rate of dual variables $\eta_{\rho}$ are set to be 0.9 and 0.1, respectively. The only parameter $\alpha$ for the proposed global entropy constraint is set to be $6N / 50$ , where $N$ is the total number of instances. The weight is scaled according to the ablation study on CIFAR-10 as illustrated in Sec. 4.1.4. + +Augmentation Augmentation is essential for the success of unsupervised representation learning [7]. For a fair comparison, we apply the prevalent settings in existing work [7, 15]. Concretely, random crop, color jitter, random gray scale, Gaussian blur, solarize and random horizontal flip are included as augmentations. Considering that the resolution of images in CIFAR and STL-10 is much smaller than those in ImageNet, we let the minimal scale of random crop be 0.3, 0.2, and 0.2 for CIFAR-10, CIFAR-100, and STL-10, respectively. That for ImageNet is kept as 0.08. Other parameters are the same for all data sets and many recent work share the similar augmentation [9, 15, 26, 32]. + +All experiments on small data sets are implemented on a single V100, while 8 GPUs are sufficient for ImageNet. + +# 4.1. Ablation Study + +To illustrate the behavior of the proposed method, we study the effect of parameters in SeCu on CIFAR-10 with the size constraint. + +# 4.1.1 Effect of MLP Head + +The MLP head shows better generalization than the FC layer in representation learning [7], but its application in deep clustering has not been evaluated systematically. We compare different architectures of MLP head and summarize results in Table 1. Besides the standard MLP layer for projection, there can be another MLP layer attached to the projection MLP head, which is referred as the prediction head [15]. + +First, with the standard architecture denoted by 0 for projection and prediction, SeCu is already $3.7\%$ better than SCAN on ACC and it confirms that the one-stage deep clustering strategy can learn more consistent features than two-stage methods. If including a single layer for SeCu, the performance can be improved by $2.1\%$ , which shows that adding a layer with the reduced dimension in head is helpful for deep clustering. Moreover, due to the dimension reduction from the projection layer, the size of cluster centers decreases and the total number of parameters in the model + +
#Proj#Pred#ParaACCNMIARI
0011.45M85.576.273.1
1011.30M87.678.776.7
2011.57M88.179.477.6
3011.83M88.179.477.6
3211.97M88.279.677.8
+ +is even smaller than that of the original network. When increasing the number of layers and having a 2-layer MLP head, ACC can be further improved by $0.5\%$ and the model size is only slightly increased. After that, there is no significant gain by applying more complicated MLP head. It is because that the 2-layer MLP head is sufficient for small data sets and we will fix 2-layer MLP as the head for the rest experiments except ImageNet, where 3-layer projection and 2-layer prediction are applied as in [26]. + +# 4.1.2 Effect of Negative Instances + +After the ablation experiments for the architecture, we study the effect of the proposed stable cluster discrimination loss in Eqn. 3. Table 2 compares it to the standard cross entropy loss (CE) that keeps the gradient from negative instances. + +Table 1: Comparison of MLP head on CIFAR-10. "Proj" and "#Pred" indicate the number of layers in MLP for projection head and prediction head, respectively. ACC (\%), NMI (\%) and ARI (\%) are reported. Baseline ACC of SCAN [14] is $81.8\%$ . + +
Loss#Max#MinACCNMIARI
CE6,4864,50151.650.235.0
SeCu5,1904,55688.179.477.6
+ +Table 2: Comparison of loss functions on CIFAR-10. "Max" and "Min" indicate the maximal and minimal size of obtained clusters. + +First, we can observe that the proposed SeCu loss outperforms the standard cross entropy loss by a large margin of $36.5\%$ on ACC, which confirms the benefit of stable training. Then, by investigating the distribution of learned clusters, we find that the minimal size of clusters obtained from different loss functions is almost the same due to the lower-bound constraint, while the maximal size varies. The largest cluster obtained by SeCu contains 5,190 instances, which is close to the ground-truth distribution of CIFAR-10. On the contrary, the dominating cluster found by CE loss has 6,486 instances that is significantly different from the uniform distribution in CIFAR-10. It is because that the negative instances in CE loss can perturb the learning of centers and result in sub-optimal cluster assignments. + +# 4.1.3 Effect of Updating Criterion for Cluster Centers + +The cluster centers in SeCu can be optimized by SGD or updated by a closed-form solution as in Eqn. 4. We compare these two strategies and summarize the performance in Table 3. + +
MethodsACCNMIARI
SeCu-CF87.4±0.578.5±0.676.4±0.9
SeCu-SGD88.1±0.379.3±0.477.5±0.4
+ +Let "SeCu-CF" and "SeCu-SGD" be the closed-form update and the SGD optimization, respectively. We can observe that SeCu-CF has the similar performance to SeCu-SGD, which shows the effectiveness of the closed-form updating. Moreover, SeCu-SGD has a smaller standard deviation than SeCu-CF due to the smoothness from the momentum in SGD. Compared with SeCu-CF, SeCu-SGD has an additional parameter of the learning rate for centers. Hence, SeCu-CF can be applied to evaluate the preliminary performance of SeCu with less tuning efforts. + +# 4.1.4 Effect of $\alpha$ in Entropy Constraint + +Besides the size constraint, the proposed entropy constraint is also effective for balancing cluster assignments. Table 4 demonstrates the results with different $\alpha$ . + +Table 3: Comparison of optimization for cluster centers on CIFAR-10. Mean±std over 8 trails are reported. + +
α#Max#MinACCNMIARI
20,0005,0564,91385.376.373.1
8,0005,1154,65887.979.077.1
6,0005,1204,57988.079.477.4
4,0005,3764,32484.975.472.2
10028,575019.938.717.0
050,000010.00.00.0
+ +Table 4: Comparison of $\alpha$ in Eqn. 10 on CIFAR-10. + +First, a large $\alpha$ will regularize the distribution to be uniform and result in a sub-optimal performance. By decreasing the weight, the assignment becomes more flexible and a desired performance can be observed when $\alpha = 6,000$ . However, a small $\alpha$ will lead to an imbalanced distribution, which is inappropriate for the balanced data sets such as CIFAR and STL. $\alpha = 0$ discards the constraint and incurs collapsing. Evidently, the entropy constraint can balance the size of clusters effectively + +Since the entropy is defined on the whole data set and CIFAR-10 contains 50,000 instances for training, $\alpha$ will be scaled as $6N / 50$ for different data sets, where $N$ denotes the total number of instances in the training set. More ablation experiments can be found in the supplementary. + +# 4.2. Comparison with State-of-the-Art + +After the ablation study, experiments are conducted on benchmark data sets to compare with state-of-the-art methods. The results averaged over 8 trials and the best performance among them are reported for our method. SeCu with size constraint and entropy constraint are denoted as "SeCu-Size" and "SeCu-Entropy", respectively. Table 5 shows the performance of different methods, where two-stage methods with a pre-training phase are marked for "Two-stage". + +First, as a one-stage method, SeCu outperforms two-stage methods by a margin of about $3\%$ over all metrics on CIFAR-10 and CIFAR-100-20. It demonstrates that learning representations and clustering simultaneously is essential for deep clustering. The two-stage method NNM[11] shows the comparable performance to SeCu on STL-10. It is because that STL-10 only constrains 5,000 target instances for training and both of NNM and SeCu already achieve the accuracy of supervised learning. Second, compared with the one-stage method without sufficient representation learning, i.e., PICA, SeCu demonstrates a better performance by about $20\%$ on CIFAR-10 and CIFAR-100-20, and $10\%$ on STL-10, which shows the importance of optimizing representations for instances. Third, with the same size constraint, SeCu surpasses CoKe by $4.3\%$ for ARI on CIFAR-10, which confirms the effectiveness of the proposed learning objective. Finally, compared with size constraint in SeCu-Size, SeCu-Entropy demonstrates a competitive performance over all data sets. Since the entropy constraint is optimized by enumeration without introducing any auxiliary variable, it can be an substitute to the size constraint for cluster assignments. + +After the clustering phase, self-labeling [14] is an effective strategy to further improve the performance. We try it for SeCu-Size with the setting in [14] and report the result in Table 5. Interestingly, self-labeling is also effective for SeCu and helps approach the supervised accuracy on CIFAR-10. As a one-stage method, SeCu has the strong augmentation for representation learning, which may introduce additional noise for clustering. Self-labeling has the weak augmentation for fine-tuning and refines clustering. + +![](images/491d7827b83f03d6386245ce0ed1eeb26bde28b70ca843b4fdb4f4e6619b3a49.jpg) +CIFAR-10 + +![](images/99998451b7ab208cdebf4ae06a508fc39cea43172bf9a543156afa1cf4c377dd.jpg) +STL-10 +Figure 2: Illustration of exemplar images. + +![](images/76e829472ebecd858455b17dde3d09d35a61b7996e19c5ca6b98d529626617e7.jpg) +CIFAR-100-20 + +
MethodsTwo-stageCIFAR10CIFAR100-20STL10
ACCNMIARIACCNMIARIACCNMIARI
Supervised93.886.287.080.068.063.280.665.963.1
DeepCluster [4]37.4--18.9--33.4--
IIC [20]61.751.141.125.722.511.759.649.639.7
PICA [19] (best)69.659.151.233.731.017.171.361.153.1
Pretext [7]+k-means✓65.959.850.939.540.223.965.860.450.6
SCAN [14] (mean)✓81.871.266.542.244.126.775.565.459.0
NNM [11]✓84.374.870.947.748.431.680.869.465.0
GCC [32]✓85.676.472.847.247.230.578.868.463.1
CoKe [26] (mean)85.776.673.249.749.133.5---
SeCu-Size (mean)88.179.377.550.050.735.080.269.463.9
SeCu-Size (best)88.579.978.251.651.636.081.470.765.7
SeCu-Entropy (mean)88.079.277.249.950.634.179.568.763.0
SeCu-Entropy (best)88.279.777.751.251.434.980.569.964.4
with self-labeling:
SCAN [14]✓88.379.777.250.748.633.380.969.864.6
GCC [32]✓90.1--52.3--83.3--
SeCu93.086.185.755.255.139.783.673.369.3
+ +Exemplar Images Finally, we show exemplar images from clusters obtained by SeCu for different data sets. Fig. 2 illustrates the images that are close to cluster centers in CIFAR-10, CIFAR-100-20, and STL-10, respectively. Evidently, the proposed method can obtain the ground-truth class structure for CIFAR-10 and STL-10. Moreover, with a large intra-class variance, SeCu can still observe appropriate clusters for CIFAR-100-20. + +Considering that labels of super-classes in CIFAR-100-20 contain vague concepts such as "vehicles 1" and "vehicles 2", we include results on the target 100 classes in Table 6, which may benefit future research. The experiment settings for CIFAR-100 are identical to that for CIFAR-100-20, except the learning rate of centers, which is reduced to 0.01. + +Table 5: Comparison of clustering methods on benchmark data sets. "Two-stage" denotes an additional pre-training stage. + +
MethodsACCNMIARI
SeCu-Size (mean)46.461.732.3
SeCu-Size (best)47.161.932.6
SeCu-Entropy (mean)45.360.731.5
SeCu-Entropy (best)46.761.032.2
SeCu†51.365.237.1
+ +# 4.3. Comparison on ImageNet + +Now we evaluate SeCu on the challenging ImageNet data set in Table 7, where the cost of many two-stage methods becomes intractable and only available baselines are + +included. First, SeCu achieves $51.4\%$ ACC on ImageNet, which is better than SCAN with self-labeling by more than $11\%$ . With the same 10 clustering heads and size constraint, SeCu is still $3.8\%$ better than CoKe on ACC. Compared with the objective in CoKe, SeCu is tailored for deep clustering and the comparison confirms the efficacy of our method. With self-labeling, SeCu shows state-of-the-art performance of $53.5\%$ ACC on ImageNet. It demonstrates that our method is applicable for large-scale data sets. + +Table 6: Results of SeCu on CIFAR-100. † denotes a method with self-labeling. + +
MethodsACCNMIARI
SCAN†39.972.027.5
CoKe‡47.676.235.6
SeCu51.478.039.7
SeCu†53.579.441.9
+ +Table 7: Comparison on ImageNet. ${}^{ \dagger }$ denotes a method with self-labeling. ${}^{ \ddagger }$ is our reproduction with 10 clustering heads. + +# 5. Conclusion + +In this work, a novel stable cluster discrimination task that considers the hardness of instances is proposed for one-stage deep clustering. To avoid collapsing in representations, a global entropy constraint with theoretical guarantee is investigated. Finally, the comparison with state-of-the-art methods confirms the effectiveness of our proposal. While recent work shows that leveraging nearest neighbors improves deep clustering [11, 14, 32], incorporating such information in SeCu can be our future work. + +# References + +[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. +[2] Hangbo Bao, Li Dong, and Furu Wei. Beit: BERT pretraining of image transformers. CoRR, abs/2106.08254, 2021. +[3] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. +[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. +[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021. +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, volume 119, pages 1597-1607, 2020. +[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758. Computer Vision Foundation / IEEE, 2021. +[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. CoRR, abs/2104.02057, 2021. +[10] Adam Coates, Andrew Y. Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In AISTATS, volume 15, pages 215-223. JMLR.org, 2011. +[11] Zhiyuan Dang, Cheng Deng, Xu Yang, Kun Wei, and Heng Huang. Nearest neighbor matching for deep clustering. In CVPR, pages 13693-13702. Computer Vision Foundation / IEEE, 2021. +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021. +[13] Ehsan Elhamifar and René Vidal. Sparse subspace clustering: Algorithm, theory, and applications. IEEE Trans. Pattern Anal. Mach. Intell., 35(11):2765-2781, 2013. +[14] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. SCAN: learning to classify images without labels. In ECCV, volume 12355, pages 268-285. Springer, 2020. +[15] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent - A new approach to self-supervised learning. In NeurIPS, 2020. + +[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021. +[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020. +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. +[19] Jiabo Huang, Shaogang Gong, and Xiatian Zhu. Deep semantic clustering by partition confidence maximisation. In CVPR, pages 8846-8855. Computer Vision Foundation / IEEE, 2020. +[20] Xu Ji, Andrea Vedaldi, and João F. Henriques. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9864-9873. IEEE, 2019. +[21] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009. +[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, pages 1106-1114, 2012. +[23] Yunfan Li, Peng Hu, Jerry Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In AAAI, pages 8547-8555. AAAI Press, 2021. +[24] Stuart P. Lloyd. Least squares quantization in PCM. IEEE Trans. Inf. Theory, 28(2):129-136, 1982. +[25] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In ICCV, pages 6449-6457, 2019. +[26] Qi Qian, Yuanhong Xu, Juhua Hu, Hao Li, and Rong Jin. Unsupervised visual representation learning by online constrained k-means. In CVPR, pages 16619-16628. IEEE, 2022. +[27] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Fei-Fei Li. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis., 115(3):211-252, 2015. +[28] Ulrike von Luxburg. A tutorial on spectral clustering. Stat. Comput., 17(4):395-416, 2007. +[29] Junyuan Xie, Ross B. Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In ICML, volume 48, pages 478-487. JMLR.org, 2016. +[30] Bo Yang, Xiao Fu, Nicholas D. Sidiropoulos, and Mingyi Hong. Towards k-means-friendly spaces: Simultaneous deep learning and clustering. In ICML, 2017. +[31] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, volume 139, pages 12310-12320. PMLR, 2021. +[32] Huasong Zhong, Jianlong Wu, Chong Chen, Jianqiang Huang, Minghua Deng, Liqiang Nie, Zhouchen Lin, and Xian-Sheng Hua. Graph contrastive clustering. In ICCV, pages 9204-9213. IEEE, 2021. \ No newline at end of file diff --git a/2023/Stable Cluster Discrimination for Deep Clustering/images.zip b/2023/Stable Cluster Discrimination for Deep Clustering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4451fd64c4c8cdf81e952df7099531f806e0e061 --- /dev/null +++ b/2023/Stable Cluster Discrimination for Deep Clustering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b899b0835aa8582d669a5d5bc1e0418f3a669b43f9104975d5d58acfa141b38f +size 477109 diff --git a/2023/Stable Cluster Discrimination for Deep Clustering/layout.json b/2023/Stable Cluster Discrimination for Deep Clustering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a2ada4be0ce9916256935c8da9ad5667150504b6 --- /dev/null +++ b/2023/Stable Cluster Discrimination for Deep Clustering/layout.json @@ -0,0 +1,10122 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 103, + 453, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 103, + 453, + 121 + ], + "spans": [ + { + "bbox": [ + 141, + 103, + 453, + 121 + ], + "type": "text", + "content": "Stable Cluster Discrimination for Deep Clustering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 144, + 402, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 144, + 402, + 185 + ], + "spans": [ + { + "bbox": [ + 192, + 144, + 402, + 185 + ], + "type": "text", + "content": "Qi Qian \nAlibaba Group, Bellevue, WA 98004, USA \nqi.qian@alibaba-inc.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 289, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 574 + ], + "type": "text", + "content": "Deep clustering can optimize representations of instances (i.e., representation learning) and explore the inherent data distribution (i.e., clustering) simultaneously, which demonstrates a superior performance over conventional clustering methods with given features. However, the coupled objective implies a trivial solution that all instances collapse to the uniform features. To tackle the challenge, a two-stage training strategy is developed for decoupling, where it introduces an additional pre-training stage for representation learning and then fine-tunes the obtained model for clustering. Meanwhile, one-stage methods are developed mainly for representation learning rather than clustering, where various constraints for cluster assignments are designed to avoid collapsing explicitly. Despite the success of these methods, an appropriate learning objective tailored for deep clustering has not been investigated sufficiently. In this work, we first show that the prevalent discrimination task in supervised learning is unstable for one-stage clustering due to the lack of ground-truth labels and positive instances for certain clusters in each mini-batch. To mitigate the issue, a novel stable cluster discrimination (SeCu) task is proposed and a new hardness-aware clustering criterion can be obtained accordingly. Moreover, a global entropy constraint for cluster assignments is studied with efficient optimization. Extensive experiments are conducted on benchmark data sets and ImageNet. SeCu achieves state-of-the-art performance on all of them, which demonstrates the effectiveness of one-stage deep clustering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "Clustering is a fundamental task in unsupervised learning. Given features of instances, the unlabeled data set will be partitioned into multiple clusters, where instances from the same cluster are similar according to the measurement defined by a distance function. With fixed features, most of research efforts focus on studying appropriate distance functions and ingenious algorithms have been proposed by different measurements, e.g., k-means clustering [24], spec" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 214, + 420, + 308 + ], + "blocks": [ + { + "bbox": [ + 317, + 214, + 420, + 308 + ], + "lines": [ + { + "bbox": [ + 317, + 214, + 420, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 214, + 420, + 308 + ], + "type": "image", + "image_path": "7d4651c447ce90e196cf23908d71313dffa6dc0b8fd10b8d4047a7253127216a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 315, + 547, + 424 + ], + "lines": [ + { + "bbox": [ + 304, + 315, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 547, + 424 + ], + "type": "text", + "content": "Figure 1: An illustration of the proposed method. 10 data points are randomly sampled from two different Gaussian distributions, respectively. Points with the same color are from the same distribution, while squares denote the corresponding cluster centers obtained by different methods. Star indicates the mis-classified data. Unlike k-means that assigns the uniform weight for different instances, our method considers the hardness of instances and is better for discrimination based clustering." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 429, + 213, + 533, + 308 + ], + "blocks": [ + { + "bbox": [ + 429, + 213, + 533, + 308 + ], + "lines": [ + { + "bbox": [ + 429, + 213, + 533, + 308 + ], + "spans": [ + { + "bbox": [ + 429, + 213, + 533, + 308 + ], + "type": "image", + "image_path": "9c1fa25d5bce1822d8cfe9a6b49da87f78229c6e3b50478b87e1a82cd3551977.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 447, + 504, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 447, + 504, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 447, + 504, + 460 + ], + "type": "text", + "content": "tral clustering [28], subspace clustering [13], etc." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 460, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 604 + ], + "type": "text", + "content": "With the development of deep neural networks, deep learning is capable of learning representations from raw materials and demonstrates the dominating performance on various supervised tasks [22]. Thereafter, it is also introduced to clustering recently [4, 11, 14, 19, 20, 23, 29, 30, 32]. Unlike the conventional clustering, representations of instances are learned with cluster assignments and centers simultaneously in deep clustering, which is more flexible to capture the data distribution. However, the coupled objective can result in a trivial solution that all instances collapse to the uniform features [4] and designing appropriate clustering criterion in the new scenario becomes challenging." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Many deep clustering algorithms [11, 14, 32] exploit a two-stage training strategy to decouple representation learning and clustering to avoid collapsing. The recent progress in unsupervised representation learning [2, 4, 7, 16, 17, 26] shows that informative features capturing the semantic similarity between instances without collapsing can be obtained by pre-training a large set of unlabeled data. Inspired by the observation, those methods leverage a pre-training stage and then focus on developing algorithms to optimize clus" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16645" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 132 + ], + "type": "text", + "content": "tering by fine-tuning a pre-trained model in the second stage. By refining the relations between nearest neighbors obtained from the pre-training stage, two-stage methods achieve a significantly better performance than one-stage ones without sufficient representation learning [19, 20]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 134, + 286, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 134, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 134, + 286, + 336 + ], + "type": "text", + "content": "However, the objective of pre-training can be inconsistent with that of clustering, which results in a sub-optimal performance for deep clustering. Note that different from supervised learning where the objective is explicitly defined by labels, that for unsupervised representation learning is arbitrary and various pretext tasks have been proposed, e.g., instance discrimination [7, 17], cluster discrimination [4, 26], masked modeling [2, 16], etc. Most of existing two-stage clustering methods adopt instance discrimination, i.e., SimCLR [7], for pre-training, whereas it aims to identify each instance as an individual class and the objective is different from clustering that aims to group instances from the same cluster. Moreover, SwAV [5] demonstrates that clustering itself is also an effective pretext task for representation learning. Therefore, we focus on facilitating one-stage deep clustering that optimizes representations and clustering simultaneously in this work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 339, + 286, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 339, + 286, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 339, + 286, + 446 + ], + "type": "text", + "content": "Most of existing one-stage methods are proposed solely for representation learning. To tackle the collapsing issue, research efforts are mainly devoted to developing appropriate constraints for cluster assignments, especially for online deep clustering. For example, [1, 5] apply the balanced constraint that each cluster has the same number of instances for clustering. [26] further relaxes the balanced constraint to a lower-bound size constraint that limits the minimal size of clusters and demonstrates a more flexible assignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 449, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 286, + 567 + ], + "type": "text", + "content": "After obtaining cluster assignments as pseudo labels, representations and cluster centers can be optimized as in the supervised scenario. For example, [5, 6] learn the encoder network and cluster centers by solving a classification problem with the standard cross entropy loss. [26] has the same classification task for representation learning while adopting k-means for updating cluster centers. Although these methods achieve a satisfied performance on representation learning, a learning objective tailored for deep clustering has not attracted sufficient attentions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "type": "text", + "content": "In this work, we investigate the effective learning task for one-stage deep clustering. By analyzing the standard cross entropy loss for supervised learning, we find that it can be unstable for unsupervised learning. Concretely, the gradient for updating cluster centers consists of two ingredients: gradient from positive instances of the target cluster and that from irrelevant negative ones. However, with a limited mini-batch size in stochastic gradient descent (SGD), there can be no positive ones for a large proportion of clusters (e.g., " + }, + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 48, + 570, + 286, + 712 + ], + "type": "text", + "content": " on ImageNet) at each iteration. Due to the lack of positive instances, the influence from negative ones is dominating. Unlike supervised learning where labels are fixed," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 119 + ], + "type": "text", + "content": "the cluster assignments can change during training in deep clustering. Therefore, the noise from the large variance of negative instances will be accumulated, which makes the optimization unstable." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 121, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 121, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 308, + 121, + 545, + 311 + ], + "type": "text", + "content": "To mitigate the problem, we propose a stable cluster discrimination (SeCu) task for deep clustering, which stops the gradient from negative instances for updating cluster centers in the cross entropy loss. Compared with k-means in [26], where positive instances have the uniform weight for updating centers, SeCu as a discrimination task considers the hardness of instances and a large weight will be assigned to hard instances for updating. Fig. 1 illustrates the hardness-aware clustering criterion implied by SeCu. Besides, we improve the cluster assignment by developing an entropy constraint that regularizes the entropy of assignments over the entire data set. Compared to the optimization with the size constraint [26], our method can reduce the number of variables and hyper-parameters, and thus make the learning more convenient. The main contributions of this work can be summarized as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 319, + 322, + 545, + 508 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 319, + 322, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 322, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 319, + 322, + 545, + 380 + ], + "type": "text", + "content": "- A novel task is proposed for one-stage deep clustering. SeCu tailors the supervised cross entropy loss by eliminating the influence from the negative instances in learning cluster centers, which makes the training stable when ground-truth labels are unavailable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 392, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 392, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 319, + 392, + 545, + 426 + ], + "type": "text", + "content": "- A global entropy constraint is exploited to balance the size of clusters and an efficient closed-form solution is developed for online assignment with the constraint." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 437, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 437, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 319, + 437, + 545, + 508 + ], + "type": "text", + "content": "- A simple framework with a single loss function and encoder is introduced for deep clustering in Eqn. 6. The proposed method is evaluated with the standard protocol on benchmark data sets and ImageNet [27]. The superior performance of SeCu confirms its effectiveness for deep clustering." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 308, + 521, + 391, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 391, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 391, + 533 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 542, + 544, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 542, + 544, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 542, + 544, + 565 + ], + "type": "text", + "content": "In this section, we briefly review unsupervised representation learning and then focus on clustering." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 575, + 511, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 575, + 511, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 575, + 511, + 587 + ], + "type": "text", + "content": "2.1. Unsupervised Representation Learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 594, + 545, + 712 + ], + "type": "text", + "content": "By leveraging massive unlabeled data, unsupervised representation learning can learn a pre-trained model that encoders the semantic information from data and helps downstream tasks with fine-tuning. Compared with supervised learning containing labels, one major challenge in unsupervised learning is defining appropriate positive pairs for representation learning. SimCLR [7] and MoCo [17] apply instance discrimination as the objective that generates positive pairs from diverse views of the same instance and considers other instances as the negative ones. Then, BYOL [15]" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "16646" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "and SimSiam [8] demonstrate that with appropriate neural network architectures, the applicable models can be obtained with only positive pairs. Besides the work optimizing instance space, Barlow Twins [31] defines positive pairs in feature space and also learns effective pre-trained models. Recently, learning with masked modeling [2, 16] achieves success on the specific backbone of vision transformer [12] and shows the better fine-tuning performance than the generic objective on the ResNet [18]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 188, + 146, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 188, + 146, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 188, + 146, + 201 + ], + "type": "text", + "content": "2.2. Deep Clustering" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 206, + 289, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 397 + ], + "type": "text", + "content": "Compared with instance discrimination, cluster discrimination aims to obtain positive pairs consisting of different instances for representation learning. Therefore, it has to learn representations and relations between instances simultaneously. DeepCluster [4] obtains membership of instances with an offline k-means and then optimizes representations alternately. Some work focuses on representation learning and introduces the balanced constraint that each cluster has the same number of instances for cluster assignments [1, 5] to mitigate the collapsing problem. Besides, DINO [6] applies an additional momentum encoder to further improve the representation. However, the constraint assumes a well balanced distribution, which is hard to capture the ground-truth data distribution. Thereafter, CoKe [26] relaxes the constraint to only lower-bound the minimal size of clusters to model the distribution more flexibly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 399, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 289, + 567 + ], + "type": "text", + "content": "Besides representation learning, there are many works developed solely for better clustering [11, 14, 20, 23, 29, 30, 32]. To handle the collapsing issue, two-stage methods [11, 14, 32] leverage the representations from a pretrained model to obtain nearest neighbors and fine-tune the model for clustering accordingly. In addition, the objective of pre-training can be included for clustering as multitask learning [23]. The work closest to ours is CoKe [26] that optimizes representations and clustering simultaneously. Compared with the conventional k-means in CoKe, we propose a novel stable cluster discrimination objective tailored for one-stage deep clustering. Moreover, an entropy constraint is introduced to reduce the number of parameters while demonstrating the competitive performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 578, + 216, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 216, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 216, + 590 + ], + "type": "text", + "content": "3. Stable Cluster Discrimination" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 597, + 178, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 178, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 178, + 609 + ], + "type": "text", + "content": "3.1. Cluster Discrimination" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "text", + "content": "Given a data set with " + }, + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "inline_equation", + "content": "\\{x_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "text", + "content": ", if the corresponding labels are accessible as " + }, + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "inline_equation", + "content": "\\{y_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 46, + 615, + 288, + 675 + ], + "type": "text", + "content": ", the literature from distance metric learning [25] shows that appropriate representations of examples can be learned by optimizing a classification task as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 683, + 208, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 683, + 208, + 715 + ], + "spans": [ + { + "bbox": [ + 126, + 683, + 208, + 715 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\sum_ {i = 1} ^ {N} \\ell \\left(x _ {i}, y _ {i}; \\theta\\right)", + "image_path": "45b986514e6ba9471168a07c4f1ec9d2af0ff044dd7fc7a63acdb31524208e50.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\ell (\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": " is the cross entropy loss with the normalized Softmax operator as suggested in [25]. " + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": " denotes the parameters of the deep neural network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 109, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 547, + 144 + ], + "type": "text", + "content": "For unsupervised learning that the label information is unavailable, the objective of cluster discrimination with " + }, + { + "bbox": [ + 304, + 109, + 547, + 144 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 109, + 547, + 144 + ], + "type": "text", + "content": " clusters can be written as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 332, + 152, + 545, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 152, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 332, + 152, + 545, + 186 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {f}, \\left\\{\\mathbf {w} _ {j} \\right\\}, y _ {i} \\in \\Delta} \\mathcal {L} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} - y _ {i, j} \\log \\left(p _ {i, j}\\right) \\tag {1}", + "image_path": "ba7e4652abed087ff7d9792717372c242bce005925e79a82391b8e0a5e0137cf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "type": "text", + "content": " denotes the learnable label for " + }, + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "type": "inline_equation", + "content": "\\Delta = \\{y_i|\\sum_{j = 1}^{K}y_{i,j} = 1,\\forall j,y_{i,j}\\in \\{0,1\\} \\}" + }, + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "type": "text", + "content": ". With the normalized Softmax operator, the prediction " + }, + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 306, + 196, + 547, + 232 + ], + "type": "text", + "content": " can be computed" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 242, + 545, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 242, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 354, + 242, + 545, + 273 + ], + "type": "interline_equation", + "content": "p _ {i, j} = \\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {j} / \\lambda\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {k} / \\lambda\\right)} \\tag {2}", + "image_path": "083069355adeeed3f06802cc42ca1c1042101eb137cc6cdd0f77a9b07f9b63dc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = f(x_i)" + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "content": " denotes the encoder network. " + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\theta_{f}" + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "content": " contains parameters of " + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_j\\}_{j=1}^K" + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "content": " consists of cluster centers. " + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "content": " is the parameter of temperature and " + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\forall i, j, \\|\\mathbf{x}_i\\|_2 = \\|\\mathbf{w}_j\\|_2 = 1" + }, + { + "bbox": [ + 304, + 281, + 545, + 329 + ], + "type": "text", + "content": " by normalization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "text", + "content": "Compared with the supervised counterpart, the problem in Eqn. 1 has to optimize cluster assignments " + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "inline_equation", + "content": "\\{y\\}" + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "text", + "content": ", cluster centers " + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}\\}" + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "text", + "content": ", and representation encoder network " + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "text", + "content": " simultaneously. While most of existing work focuses on optimizing " + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "inline_equation", + "content": "\\{y\\}" + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "text", + "content": " with different constraints to avoid collapsing, this work investigates the learning objective for " + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}\\}" + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 330, + 545, + 414 + ], + "type": "text", + "content": ", and a new clustering criterion is introduced accordingly." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 422, + 541, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 541, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 541, + 435 + ], + "type": "text", + "content": "3.2. Stable Loss for Optimization with Mini-batch" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 441, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 524 + ], + "type": "text", + "content": "Unlike supervised learning, where labels are fixed for examples, cluster assignments " + }, + { + "bbox": [ + 304, + 441, + 545, + 524 + ], + "type": "inline_equation", + "content": "\\{y\\}" + }, + { + "bbox": [ + 304, + 441, + 545, + 524 + ], + "type": "text", + "content": " in unsupervised learning are dynamic with the training of instance representations and cluster centers. Therefore, the original cross entropy loss becomes unstable for unsupervised discrimination. The problem can be elaborated by analyzing the updating criterion for cluster centers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 525, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 559 + ], + "type": "text", + "content": "Letting " + }, + { + "bbox": [ + 304, + 525, + 545, + 559 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 304, + 525, + 545, + 559 + ], + "type": "text", + "content": " denote the label of " + }, + { + "bbox": [ + 304, + 525, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 525, + 545, + 559 + ], + "type": "text", + "content": ", the gradient of " + }, + { + "bbox": [ + 304, + 525, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j" + }, + { + "bbox": [ + 304, + 525, + 545, + 559 + ], + "type": "text", + "content": " from the standard cross entropy loss in Eqn. 1 can be computed as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 328, + 569, + 523, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 569, + 523, + 599 + ], + "spans": [ + { + "bbox": [ + 328, + 569, + 523, + 599 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {w} _ {j}} \\mathcal {L} = \\frac {1}{\\lambda} \\left(\\sum_ {i: y _ {i} = j} \\left(p _ {i, j} - 1\\right) \\mathbf {x} _ {i} + \\sum_ {k: y _ {k} \\neq j} p _ {k, j} \\mathbf {x} _ {k}\\right)", + "image_path": "63e26c96657140a637acf8cf0dcb67668bf1d47aa28e13ac895f5d76e5eb87c1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 608, + 545, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 608, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 608, + 545, + 669 + ], + "type": "text", + "content": "The former term is to pull the center " + }, + { + "bbox": [ + 304, + 608, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_j" + }, + { + "bbox": [ + 304, + 608, + 545, + 669 + ], + "type": "text", + "content": " to the assigned instances and the latter one is to push it away from instances of other clusters. However, the following Propositions show that this updating is unstable for deep clustering. The detailed proof can be found in the supplementary." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "text", + "content": "Proposition 1. When sampling a mini-batch of " + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "text", + "content": " instances for " + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "text", + "content": " clusters, there are no positive instances for at least " + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "inline_equation", + "content": "K - b" + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "text", + "content": " clusters." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16647" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "type": "text", + "content": "Proposition 2. Let " + }, + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "type": "inline_equation", + "content": "\\mathrm{Var}_{pos}" + }, + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "type": "inline_equation", + "content": "\\mathrm{Var}_{neg}" + }, + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "type": "text", + "content": " be the variance of sampled positive and negative instances, respectively. Assuming that each instance has unit norm and the norm of the cluster mean is " + }, + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "type": "inline_equation", + "content": "\\mathrm{Var}_{neg} = \\mathcal{O}\\left(\\frac{1}{1 - a^2}\\right)\\mathrm{Var}_{pos}" + }, + { + "bbox": [ + 47, + 72, + 289, + 122 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 287, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 229 + ], + "type": "text", + "content": "Remark Proposition 1 implies that with cross entropy loss, a large proportion of cluster centers will be updated only by negative instances. Proposition 2 indicates that the variance of sampled negative instances is much larger than that of positive ones when each cluster is compact as " + }, + { + "bbox": [ + 46, + 134, + 287, + 229 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 46, + 134, + 287, + 229 + ], + "type": "text", + "content": " approaching 1. Due to the small size of a mini-batch in training deep neural networks, the variance cannot be reduced sufficiently." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "type": "inline_equation", + "content": "K = 10,000" + }, + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "type": "inline_equation", + "content": "b = 1,024" + }, + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "type": "text", + "content": " as in our multiclustering settings for ImageNet, at least about " + }, + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "type": "text", + "content": " centers will only access a mini-batch of negative instances at each iteration. Without ground-truth labels, the bias in updating will be accumulated and mislead the learning process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 289, + 287, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 287, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 287, + 325 + ], + "type": "text", + "content": "To mitigate the problem, we propose to eliminate the direction of negative instances from gradient for stable training as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 324, + 228, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 324, + 228, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 228, + 354 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {w} _ {j}} \\mathcal {L} = \\frac {1}{\\lambda} \\sum_ {i: y _ {i} = j} (p _ {i, j} - 1) \\mathbf {x} _ {i}", + "image_path": "164c93d37b1628423d1dd8bef8122637faedcb0b11708eecaacd0721e720fd56.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 356, + 287, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 287, + 379 + ], + "type": "text", + "content": "The corresponding stable cluster discrimination loss becomes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 387, + 287, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 387, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 61, + 387, + 287, + 438 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\ell_ {\\mathrm {S e C u}} \\left(x _ {i}, y _ {i}\\right) = \\\\ - \\log \\left(\\frac {\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right)}{\\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\mathbf {w} _ {y _ {i}} / \\lambda\\right) + \\sum_ {k : k \\neq y _ {i}} \\exp \\left(\\mathbf {x} _ {i} ^ {\\top} \\tilde {\\mathbf {w}} _ {k} / \\lambda\\right)}\\right) \\tag {3} \\\\ \\end{array}", + "image_path": "cf073ef7b4056a1d52cafecb990513629530ee1001acea79c0c98b641c0e1efd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 443, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 287, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 443, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{w}}_k" + }, + { + "bbox": [ + 47, + 443, + 287, + 502 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 47, + 443, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_k" + }, + { + "bbox": [ + 47, + 443, + 287, + 502 + ], + "type": "text", + "content": " with the stop-gradient operator. Compared with the standard cross entropy loss, cluster centers in SeCu loss will be updated only by positive instances, which is more stable for deep clustering when optimizing with mini-batches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 502, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 287, + 539 + ], + "type": "text", + "content": "In addition, the analysis in the following theorem demonstrates the novel hardness-aware clustering criterion for deep clustering implied by the proposed objective." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "type": "text", + "content": "Theorem 1. When fixing " + }, + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\{y_i\\}" + }, + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}" + }, + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}^*\\}" + }, + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "type": "text", + "content": " be the optimal solution of the problem with loss function in Eqn. 3 and assume " + }, + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\forall i, \\| \\mathbf{x}_i\\|_2 = 1; \\forall j, \\| \\mathbf{w}_j\\|_2 = 1" + }, + { + "bbox": [ + 47, + 544, + 287, + 581 + ], + "type": "text", + "content": ", then we have" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 586, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 586, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 77, + 586, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {j} ^ {*} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} \\left(\\frac {\\sum_ {i : y _ {i} = j} (1 - p _ {i , j}) \\mathbf {x} _ {i}}{\\sum_ {i : y _ {i} = j} 1 - p _ {i , j}}\\right) \\tag {4}", + "image_path": "09979db8b263d5a186e4ee288155e3709e96a5af7e0b5f774f18b2028e3a0959.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 621, + 258, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 258, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 258, + 634 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 621, + 258, + 634 + ], + "type": "inline_equation", + "content": "\\Pi_{\\| \\mathbf{w}\\| _2 = 1}" + }, + { + "bbox": [ + 47, + 621, + 258, + 634 + ], + "type": "text", + "content": " projects the vector to the unit norm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 647, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 682 + ], + "type": "text", + "content": "Remark For k-means in CoKe [26], cluster centers will be averaged over all assigned instances with the uniform weight" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 91, + 688, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 688, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 91, + 688, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {j} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} \\left(\\frac {\\sum_ {i : y _ {i} = j} \\mathbf {x} _ {i}}{\\sum_ {i} 1 (y _ {i} = j)}\\right) \\tag {5}", + "image_path": "6cc284eb11b6c07bd69abe3af8dcf04e08591adbe3c63b216c64a5751d656983.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "1(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": " is the indicator function. On the contrary, our proposal considers the weight of each instance by its hardness, i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "p_{i,y_i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": ". By assigning a large weight to hard instances, the corresponding center can capture the distribution of hard instances better, which is illustrated in Fig. 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "type": "text", + "content": "Our formulation also implies CoKe as a special case. Concretely, by increasing the temperature " + }, + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "type": "text", + "content": " for updating " + }, + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "type": "text", + "content": " to be infinite, we have " + }, + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "type": "inline_equation", + "content": "p_{i,j} = 1 / K" + }, + { + "bbox": [ + 304, + 133, + 545, + 180 + ], + "type": "text", + "content": " and Eqn. 4 will degenerate to Eqn. 5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "text", + "content": "Finally, besides SGD, Eqn. 4 also suggests an alternative updating strategy for " + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "text", + "content": " is computed by " + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 181, + 545, + 300 + ], + "type": "text", + "content": ", it may still require multiple iterations with Eqn. 4 for convergence. However, the representations of instances are also updated by training and we can keep a single update for centers at each iteration and improve centers along with the learning of representations. Compared with SGD for learning cluster centers, the strategy eliminates the learning rate for centers with a comparable performance as shown in the ablation study." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 300, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 324 + ], + "type": "text", + "content": "Given the learning objective SeCu, we will elaborate the proposed deep clustering method in next subsection." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 333, + 449, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 449, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 449, + 346 + ], + "type": "text", + "content": "3.3. SeCu for Deep Clustering" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 352, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 376 + ], + "type": "text", + "content": "With the proposed loss function, the objective of stable cluster discrimination for deep clustering can be written as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 349, + 396, + 473, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 396, + 473, + 428 + ], + "spans": [ + { + "bbox": [ + 349, + 396, + 473, + 428 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {f}, \\left\\{\\mathbf {w} _ {j} \\right\\}, y _ {i} \\in \\Delta} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} \\left(x _ {i}, y _ {i}\\right)", + "image_path": "2e9226f71a7e1a4123a2de5adf9efdd46d3e404b2ba47b6e268835ea2d9eec29.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 349, + 430, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 430, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 349, + 430, + 545, + 444 + ], + "type": "interline_equation", + "content": "s. t. \\quad h _ {m} (Y) \\geq b _ {m}, \\quad m = 1, \\dots , M \\tag {6}", + "image_path": "52f5026f22a71119afa23e4c8a8a09424ee282f81d4842f3e605a0b8ad7b1dc5.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "type": "inline_equation", + "content": "Y = [y_{1},\\ldots ,y_{N}]" + }, + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "type": "text", + "content": " denotes the number of constraints, and " + }, + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "type": "inline_equation", + "content": "h_m(\\cdot)" + }, + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 306, + 453, + 545, + 502 + ], + "type": "text", + "content": "-th constraint for cluster assignment. Considering that variables are entangled, the problem is solved alternately." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "type": "text", + "content": "Update of " + }, + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\theta_{f}" + }, + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "type": "text", + "content": " First, when fixing " + }, + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\{y^{t - 1}\\}" + }, + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}^{t - 1}\\}" + }, + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "type": "text", + "content": " from the last epoch, the sub-problem for representation learning in the " + }, + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 517, + 545, + 554 + ], + "type": "text", + "content": "-th epoch becomes" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 375, + 564, + 476, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 564, + 476, + 597 + ], + "spans": [ + { + "bbox": [ + 375, + 564, + 476, + 597 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {f}} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} (x _ {i}, y _ {i} ^ {t - 1})", + "image_path": "7aab3511d0f2e3bd9d9738f26a90097b74c147a2c3a162251b9bfea78fde85fc.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "where the SeCu loss degenerates to the standard cross entropy loss with fixed " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and can be optimized with SGD. The one-hot label " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "y^{t - 1}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " is kept from the " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "-th epoch, which makes the optimization consistent between adjacent epochs but the updating for representations may be delayed. To incorporate the information from the current epoch, we include two views of augmentations for the individual instance at each iteration, which is prevalent for representation learning [7, 17, 26]." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16648" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "x_{i}^{1}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "x_{i}^{2}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " be two perturbed views of the original image with random augmentations, and the prediction is" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 101, + 258, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 101, + 258, + 133 + ], + "spans": [ + { + "bbox": [ + 75, + 101, + 258, + 133 + ], + "type": "interline_equation", + "content": "p _ {i, j} ^ {\\mathrm {s}} = \\frac {\\exp (\\mathbf {x} _ {i} ^ {\\mathrm {s} \\top} \\mathbf {w} _ {j} ^ {t - 1} / \\lambda)}{\\sum_ {j} ^ {K} \\exp (\\mathbf {x} _ {i} ^ {\\mathrm {s} \\top} \\mathbf {w} _ {j} ^ {t - 1} / \\lambda)}; s = \\{1, 2 \\}", + "image_path": "59058cb884b3397f88ce3db0839e7bf788408bad3147a4eb487675d83785be79.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^s = f_t(x_i^s)" + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "text", + "content": " is extracted by the current encoder. The soft label for each view can be obtained as [26]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 166, + 287, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 166, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 52, + 166, + 287, + 182 + ], + "type": "interline_equation", + "content": "y _ {i} ^ {1} = \\tau y _ {i} ^ {t - 1} + (1 - \\tau) p _ {i} ^ {2}; y _ {i} ^ {2} = \\tau y _ {i} ^ {t - 1} + (1 - \\tau) p _ {i} ^ {1} (7)", + "image_path": "c969667c14d9437639050cb5ee480c51225bf3520c961e72e37dbe8404261a62.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 185, + 287, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 185, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 185, + 287, + 233 + ], + "type": "text", + "content": "The soft label contains the prediction from the other view of the same instance, which optimizes the consistency between different views at the same iteration. The problem with two-view optimization can be written as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 238, + 247, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 238, + 247, + 271 + ], + "spans": [ + { + "bbox": [ + 87, + 238, + 247, + 271 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {f}} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} (x _ {i} ^ {1}, y _ {i} ^ {1}) + \\ell_ {\\mathrm {S e C u}} (x _ {i} ^ {2}, y _ {i} ^ {2})", + "image_path": "fe6ae987ef8e0f4666d962e8514ee618b12cf447245170608fb29506819dbef6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "content": "Update of " + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "content": " When fixing " + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t" + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}^t\\}" + }, + { + "bbox": [ + 47, + 284, + 287, + 308 + ], + "type": "text", + "content": ", the cluster assignment can be updated by solving the problem" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 313, + 207, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 313, + 207, + 346 + ], + "spans": [ + { + "bbox": [ + 91, + 313, + 207, + 346 + ], + "type": "interline_equation", + "content": "\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log (p _ {i, j})", + "image_path": "5d0627f092ed4f9c7cbd381b895e9bed8717b40cd1e0c294ad1f4697de0f17ca.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 90, + 350, + 244, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 350, + 244, + 361 + ], + "spans": [ + { + "bbox": [ + 90, + 350, + 244, + 361 + ], + "type": "interline_equation", + "content": "s. t. \\quad h _ {m} (Y) \\geq b _ {m}, \\quad m = 1, \\dots , M", + "image_path": "90afb6d4fe475b1add7e4d6ded200c855d8ebcbd21fb9974f119653df910c411.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "type": "text", + "content": " is defined by " + }, + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t" + }, + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^t" + }, + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "type": "text", + "content": " as in Eqn. 2. Without the constraints " + }, + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\{h_m\\}" + }, + { + "bbox": [ + 47, + 368, + 287, + 475 + ], + "type": "text", + "content": ", the objective implies a greedy solution that assigns each instance to the most related cluster. It can incur the trivial solution of collapsing, especially for online deep clustering, where each instance can only be accessed once in each epoch and cluster assignments cannot be refined with multiple iterations over the entire data. Therefore, many effective constraints are developed to mitigate the challenge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "content": "First, we investigate the size constraint that is prevalent in existing methods [5, 26]. Following [26], the cluster size can be lower-bounded as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 84, + 514, + 201, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 514, + 201, + 548 + ], + "spans": [ + { + "bbox": [ + 84, + 514, + 201, + 548 + ], + "type": "interline_equation", + "content": "\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log (p _ {i, j})", + "image_path": "12945029d107a4f32fd9296b3ac94ca4d7871366403ff5aee83ae710116bf803.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 83, + 551, + 287, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 551, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 83, + 551, + 287, + 575 + ], + "type": "interline_equation", + "content": "s. t. \\quad \\sum_ {i} y _ {i, j} \\geq \\gamma N / K, \\quad j = 1, \\dots , K \\tag {8}", + "image_path": "3d67c8838355756e773e089efbe8a6dab944537b5d6f16d16ef91ea47dbac386.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "type": "text", + "content": " is the proportion to the average size and " + }, + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\gamma = 1" + }, + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "type": "text", + "content": " implies the balanced constraint. Let " + }, + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\rho_{j}" + }, + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "type": "text", + "content": " denote dual variables for the " + }, + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 580, + 287, + 616 + ], + "type": "text", + "content": "-th lower-bound constraint. The problem becomes" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 620, + 293, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 293, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 293, + 647 + ], + "type": "interline_equation", + "content": "\\max _ {\\rho : \\rho \\geq 0} \\min _ {y _ {i} \\in \\Delta} - \\sum_ {i} \\sum_ {j} y _ {i, j} \\log (p _ {i, j}) - \\sum_ {j} \\rho_ {j} (\\sum_ {i} y _ {i, j} - \\gamma N / K)", + "image_path": "a26e4f6c2967430e41e37556e7b6adb7b0938c1c5827417212e3efa5fced0715.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 651, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 686 + ], + "type": "text", + "content": "When a mini-batch of instances arrive, the cluster assignment for each instance can be obtained via a closed-form solution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 58, + 689, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 287, + 716 + ], + "type": "interline_equation", + "content": "y _ {i, j} ^ {t} = \\left\\{ \\begin{array}{l l} 1 & j = \\arg \\min _ {j} - \\log \\left(p _ {i, j}\\right) - \\rho_ {j} \\\\ 0 & o. w. \\end{array} \\right. \\tag {9}", + "image_path": "517d9b4901fd6deb525763f2cd92df8c2d51a8ad981826eebddcb3bff1367fc1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": "Then, the dual variables will be updated by stochastic gradient ascent. More details can be found in [26]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 97, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 545, + 216 + ], + "type": "text", + "content": "The size constraint is capable of avoiding collapsing problem explicitly, but it introduces additional dual variables to help assignment and has to optimize them in training. To simplify the optimization, a global entropy constraint is investigated to balance the distribution over all clusters. Compared with the size constraint, the bound for the cluster size is implicit with the entropy constraint. Nevertheless, as we will demonstrate, it can help reduce the number of variables and hyper-parameters, which is more friendly for users." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 217, + 545, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 241 + ], + "type": "text", + "content": "Given the set of cluster assignments " + }, + { + "bbox": [ + 305, + 217, + 545, + 241 + ], + "type": "inline_equation", + "content": "\\{y\\}" + }, + { + "bbox": [ + 305, + 217, + 545, + 241 + ], + "type": "text", + "content": ", the entropy of the whole data set can be defined as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 346, + 250, + 505, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 250, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 346, + 250, + 505, + 284 + ], + "type": "interline_equation", + "content": "H (y) = - \\sum_ {j = 1} ^ {K} \\frac {\\sum_ {i} ^ {N} y _ {i , j}}{N} \\log (\\frac {\\sum_ {i} ^ {N} y _ {i , j}}{N})", + "image_path": "ce5fea4eb43d46227f6768b3c2c3b6741a091d92e94d6281be9453129447e8a9.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 294, + 544, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 294, + 544, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 544, + 317 + ], + "type": "text", + "content": "With the entropy as the regularization, the objective can be written as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 330, + 327, + 522, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 327, + 522, + 361 + ], + "spans": [ + { + "bbox": [ + 330, + 327, + 522, + 361 + ], + "type": "interline_equation", + "content": "\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log (p _ {i, j}) \\quad s. t. \\quad H (y) \\geq \\gamma", + "image_path": "bc41aa7a8a6b6be4cc2cd0a422935c8c5df1af8f18c70e798709789a46d16869.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 372, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 420 + ], + "type": "text", + "content": "Compared with the problem consisting of " + }, + { + "bbox": [ + 305, + 372, + 545, + 420 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 372, + 545, + 420 + ], + "type": "text", + "content": " constraints in Eqn. 8, there is only a single constraint that controls the size of all clusters simultaneously. According to the dual theory [3], the problem is equivalent to" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 337, + 430, + 545, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 430, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 337, + 430, + 545, + 464 + ], + "type": "interline_equation", + "content": "\\min _ {y _ {i} \\in \\Delta} - \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {K} y _ {i, j} \\log \\left(p _ {i, j}\\right) - \\alpha H (y) \\tag {10}", + "image_path": "9efb46a6f1788e6e2b70f95aefc13e8eb1c667a02a9ec0dab8090a671624675c.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 475, + 544, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 475, + 544, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 544, + 523 + ], + "type": "text", + "content": "Since one-hot labels of all instances are kept in memory, the optimal solution can be obtained by enumerating. When optimizing the label for " + }, + { + "bbox": [ + 305, + 475, + 544, + 523 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 475, + 544, + 523 + ], + "type": "text", + "content": " and fixing other labels, the closed-form solution for " + }, + { + "bbox": [ + 305, + 475, + 544, + 523 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 305, + 475, + 544, + 523 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 533, + 547, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 547, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 547, + 572 + ], + "type": "interline_equation", + "content": "y _ {i, j} ^ {t} = \\left\\{ \\begin{array}{l l} 1 & j = \\arg \\min _ {j} - \\log \\left(p _ {i, j}\\right) - \\alpha H \\left(y ^ {t - 1}, y _ {i: j}\\right) \\\\ 0 & o. w. \\end{array} \\right. \\tag {11}", + "image_path": "cb5aafeba946b652761bc5fb85c40596a700e52d6149d00a72e40ac8cc834c08.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "inline_equation", + "content": "H(y^{t - 1},y_{i:j})" + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "text", + "content": " replaces the previous label " + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "inline_equation", + "content": "H(y^{t - 1},y_{i:j}) = H([y_1^{t - 1},\\ldots ,y_{i - 1}^{t - 1},y_{i + 1}^{t - 1},\\ldots ,y_n^{t - 1};y_{i,j} = 1])" + }, + { + "bbox": [ + 305, + 583, + 545, + 632 + ], + "type": "text", + "content": ". The convergence is guaranteed as in the following theorem." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 642, + 545, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 679 + ], + "type": "text", + "content": "Theorem 2. Let " + }, + { + "bbox": [ + 305, + 642, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(Y)" + }, + { + "bbox": [ + 305, + 642, + 545, + 679 + ], + "type": "text", + "content": " denote the objective in Eqn. 10. If updating cluster assignments sequentially according to Eqn. 11, we have " + }, + { + "bbox": [ + 305, + 642, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(Y^t) \\leq \\mathcal{L}(Y^{t-1})" + }, + { + "bbox": [ + 305, + 642, + 545, + 679 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Proof. It is directly from the optimality of the closed-form solution in Eqn. 11. " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\square" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16649" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "Remark While some deep clustering methods apply the entropy regularization for learning, it is defined over a minibatch of instances and optimized by SGD [14]. On the contrary, our method leverages the entropy of the entire data set that can capture the global distribution better. Moreover, the entropy in our method can be maximized with a closed-form solution, which is more efficient for learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 157, + 287, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 157, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 59, + 157, + 287, + 169 + ], + "type": "text", + "content": "With the two-view optimization, the updating becomes" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 190, + 285, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 190, + 285, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 190, + 285, + 209 + ], + "type": "interline_equation", + "content": "j = \\arg \\min _ {j} - (\\log (p _ {i, j} ^ {1}) + \\log (p _ {i, j} ^ {2})) / 2 - \\alpha H (y ^ {t - 1}, y _ {i: j})", + "image_path": "241235d95972c136f3d022f2c65e1e54c63b8123ef19140b46203e3025f9f7e4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 227, + 287, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 287, + 264 + ], + "type": "text", + "content": "Update of w With fixed " + }, + { + "bbox": [ + 47, + 227, + 287, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t" + }, + { + "bbox": [ + 47, + 227, + 287, + 264 + ], + "type": "text", + "content": " and the updated pseudo label " + }, + { + "bbox": [ + 47, + 227, + 287, + 264 + ], + "type": "inline_equation", + "content": "\\{y^{t}\\}" + }, + { + "bbox": [ + 47, + 227, + 287, + 264 + ], + "type": "text", + "content": ", cluster centers can be optimized by minimizing SeCu loss over two views of augmentations by SGD" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 273, + 247, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 273, + 247, + 306 + ], + "spans": [ + { + "bbox": [ + 87, + 273, + 247, + 306 + ], + "type": "interline_equation", + "content": "\\min _ {\\{\\mathbf {w} _ {j} \\}} \\sum_ {i = 1} ^ {N} \\ell_ {\\mathrm {S e C u}} (\\mathbf {x} _ {i} ^ {1}, y _ {i} ^ {t}) + \\ell_ {\\mathrm {S e C u}} (\\mathbf {x} _ {i} ^ {2}, y _ {i} ^ {t})", + "image_path": "035d3fc2c25ea4dff7b89fae11bfab2ef3f022b90eaf0aa71ecb3ea515df9389.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 316, + 282, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 282, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 282, + 328 + ], + "type": "text", + "content": "Alternatively, " + }, + { + "bbox": [ + 47, + 316, + 282, + 328 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 47, + 316, + 282, + 328 + ], + "type": "text", + "content": " can be updated by a closed-form solution" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 337, + 250, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 337, + 250, + 372 + ], + "spans": [ + { + "bbox": [ + 83, + 337, + 250, + 372 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {j} ^ {t: B} = \\Pi_ {\\| \\mathbf {w} \\| _ {2} = 1} (\\frac {\\sum_ {i : y _ {i} = j} ^ {B} (1 - p _ {i , j}) \\mathbf {x} _ {i} ^ {t}}{\\sum_ {i : y _ {i} = j} ^ {B} 1 - p _ {i , j}})", + "image_path": "df8be421bccc09f884bd28a0ba8124b0f3eab7e80b1c2a8360797f207cf9d303.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 380, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 287, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 380, + 287, + 403 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 380, + 287, + 403 + ], + "type": "text", + "content": " denotes the total number of instances received in the current epoch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 405, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 464 + ], + "type": "text", + "content": "The proposed method is easy to implement and the whole framework is illustrated in Alg. 1. Compared with the supervised discrimination, our main computational overhead is from cluster assignment, which is negligible due to the online strategy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 477, + 127, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 127, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 127, + 491 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": "The performance of the proposed method is evaluated on CIFAR-10 [21], CIFAR-100 [21], STL-10[10] and ImageNet [27]. To demonstrate the generalization performance of SeCu, we follow the setting in SCAN [14] that trains and evaluates models on the standard train/test split provided by the original data set. For STL-10 that contains an additional noisy unlabeled data set, we first learn the model on the training and noisy set, and then on the target training data as suggested by other methods [14, 32]. The standard metrics for clustering, including clustering accuracy (ACC), normalized mutual information (NMI) and adjusted rand index (ARI) are reported for evaluation. For a fair comparison, we follow the common practice to configure the architecture and training protocol for our method. Concretely, ResNet-50 [18] is applied on ImageNet, while ResNet-18 is adopted for other data sets. The key settings for ResNet-18 are elaborated as follows, while those of ResNet-50 follow [26]. More details can be found in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "type": "code", + "bbox": [ + 307, + 103, + 538, + 329 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "Algorithm 1 Pseudo-code of Stable Cluster Discrimination (SeCu) for One-stage Deep Clustering." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 307, + 103, + 538, + 329 + ], + "lines": [ + { + "bbox": [ + 307, + 103, + 538, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 103, + 538, + 329 + ], + "type": "text", + "content": "f: encoder network\n# w: cluster centers\n# w_p: centers from the last epoch\n# y: list of pseudo one-hot labels (Nx1)\n# tau: weight for labels from last epoch\n# lambda: temperature\n# keep last cluster centers before each epoch\nw_p = w_detach()\n# train one epoch\nfor x in loader: # load a minibatch with b samples\nx_1, x_2 = f(aug(x)), f(aug(x)) # two random views\ny_x = y(x_id) # retrieve label from last epoch\n# compute prediction for each view\np_1 = softmax(x_1 @ w_p / lambda)\np_2 = softmax(x_2 @ w_p / lambda)\n# obtain soft label for discrimination\ny_1 = tau * y_x + (1-tau) * p_2\ny_2 = tau * y_x + (1-tau) * p_1\n# loss for representation learning\nloss_x = (SeCu(p_1, y_1) + SeCu(p_2, y_2)) / 2\n# update prediction for clustering\np_1 = softmax(x_1detach() @ w / lambda)\np_2 = softmax(x_2detach() @ w / lambda)\n# update cluster assignments with constraints\ny(x_id) = y_x = cluster_assign(p_1, p_2)\n# loss for cluster centers\nloss_w = (SeCu(p_1, y_x) + SeCu(p_2, y_x)) / 2\n# update encoder and cluster centers\nloss = loss_x + loss_w\nloss_backward()" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "text", + "content": "Architecture Besides backbone network, a 2-layer MLP head is attached and the output dimension is 128. The similar architecture is adopted in [26, 32]. With the MLP head, the total number of parameters is almost the same as the original ResNet-18. The discussion about the MLP head can be found in Sec. 4.1.1. After the MLP head, the learned representation will be classified by a fully-connected (FC) layer that encodes cluster centers. The size of the FC layer is " + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "inline_equation", + "content": "128 \\times K" + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "text", + "content": ". As a clustering method, we let " + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "text", + "content": " be the number of ground-truth classes as in [14] and have the direct prediction from the model as cluster assignments for evaluation. Following [1, 14], 10 different classification heads are applied as multi-clustering, which benefits the target clustering. To avoid the problem of selecting an appropriate head for evaluation when 10 heads have the same " + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "text", + "content": ", we have a varying " + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "inline_equation", + "content": "cK" + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "text", + "content": " as the number of clusters for each head, where " + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "inline_equation", + "content": "c \\in \\{1, \\dots, 10\\}" + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "text", + "content": ". The prediction from the head with " + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "inline_equation", + "content": "c = 1" + }, + { + "bbox": [ + 304, + 358, + 545, + 574 + ], + "type": "text", + "content": " is reported for comparison." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Optimization To reduce efforts of parameter tuning and reuse the parameter of [26], we have " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^\\top \\mathbf{w}_j" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " in lie of " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\log(p_{i,j})" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " in Eqn. 9 and 11 for assignment. Searching the parameter for the latter one can achieve the similar performance. Before training the encoder network, one epoch is adopted to initialize cluster assignments and centers. The encoder network is optimized by SGD with a batch size of 128. The momentum and learning rate are set to be 0.9 and 0.2, respectively. Moreover, the first 10 epochs are applied for warm-up and the cosine decay for learning rate" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16650" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "is adopted subsequently. For CIFAR-10, the model is optimized by 400 epochs and cluster centers are learned by SGD with a constant learning rate of 1.2. For other challenging data sets, epochs and learning rate for centers are 800 and 0.8, respectively. " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " for the soft label as in Eqn. 7 and the temperature " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " are fixed as 0.2 and 0.05. For the size constraint, the lower-bound parameter " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " and the learning rate of dual variables " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\eta_{\\rho}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " are set to be 0.9 and 0.1, respectively. The only parameter " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " for the proposed global entropy constraint is set to be " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "6N / 50" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " is the total number of instances. The weight is scaled according to the ablation study on CIFAR-10 as illustrated in Sec. 4.1.4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 232, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 376 + ], + "type": "text", + "content": "Augmentation Augmentation is essential for the success of unsupervised representation learning [7]. For a fair comparison, we apply the prevalent settings in existing work [7, 15]. Concretely, random crop, color jitter, random gray scale, Gaussian blur, solarize and random horizontal flip are included as augmentations. Considering that the resolution of images in CIFAR and STL-10 is much smaller than those in ImageNet, we let the minimal scale of random crop be 0.3, 0.2, and 0.2 for CIFAR-10, CIFAR-100, and STL-10, respectively. That for ImageNet is kept as 0.08. Other parameters are the same for all data sets and many recent work share the similar augmentation [9, 15, 26, 32]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 400 + ], + "type": "text", + "content": "All experiments on small data sets are implemented on a single V100, while 8 GPUs are sufficient for ImageNet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 141, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 141, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 141, + 421 + ], + "type": "text", + "content": "4.1. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 426, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 287, + 462 + ], + "type": "text", + "content": "To illustrate the behavior of the proposed method, we study the effect of parameters in SeCu on CIFAR-10 with the size constraint." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 478, + 167, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 167, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 167, + 490 + ], + "type": "text", + "content": "4.1.1 Effect of MLP Head" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 593 + ], + "type": "text", + "content": "The MLP head shows better generalization than the FC layer in representation learning [7], but its application in deep clustering has not been evaluated systematically. We compare different architectures of MLP head and summarize results in Table 1. Besides the standard MLP layer for projection, there can be another MLP layer attached to the projection MLP head, which is referred as the prediction head [15]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "First, with the standard architecture denoted by 0 for projection and prediction, SeCu is already " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " better than SCAN on ACC and it confirms that the one-stage deep clustering strategy can learn more consistent features than two-stage methods. If including a single layer for SeCu, the performance can be improved by " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ", which shows that adding a layer with the reduced dimension in head is helpful for deep clustering. Moreover, due to the dimension reduction from the projection layer, the size of cluster centers decreases and the total number of parameters in the model" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 531, + 145 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 531, + 145 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 531, + 145 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 531, + 145 + ], + "type": "table", + "html": "
#Proj#Pred#ParaACCNMIARI
0011.45M85.576.273.1
1011.30M87.678.776.7
2011.57M88.179.477.6
3011.83M88.179.477.6
3211.97M88.279.677.8
", + "image_path": "2a02e61286c88b32768304a15ed718c7b44bae0cbf29397cddb360cad381db36.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 237, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 346 + ], + "type": "text", + "content": "is even smaller than that of the original network. When increasing the number of layers and having a 2-layer MLP head, ACC can be further improved by " + }, + { + "bbox": [ + 304, + 237, + 547, + 346 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 304, + 237, + 547, + 346 + ], + "type": "text", + "content": " and the model size is only slightly increased. After that, there is no significant gain by applying more complicated MLP head. It is because that the 2-layer MLP head is sufficient for small data sets and we will fix 2-layer MLP as the head for the rest experiments except ImageNet, where 3-layer projection and 2-layer prediction are applied as in [26]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 365, + 458, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 458, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 458, + 378 + ], + "type": "text", + "content": "4.1.2 Effect of Negative Instances" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 387, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 436 + ], + "type": "text", + "content": "After the ablation experiments for the architecture, we study the effect of the proposed stable cluster discrimination loss in Eqn. 3. Table 2 compares it to the standard cross entropy loss (CE) that keeps the gradient from negative instances." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 326, + 446, + 526, + 487 + ], + "blocks": [ + { + "bbox": [ + 304, + 153, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 304, + 153, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 547, + 213 + ], + "type": "text", + "content": "Table 1: Comparison of MLP head on CIFAR-10. \"Proj\" and \"#Pred\" indicate the number of layers in MLP for projection head and prediction head, respectively. ACC (\\%), NMI (\\%) and ARI (\\%) are reported. Baseline ACC of SCAN [14] is " + }, + { + "bbox": [ + 304, + 153, + 547, + 213 + ], + "type": "inline_equation", + "content": "81.8\\%" + }, + { + "bbox": [ + 304, + 153, + 547, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 446, + 526, + 487 + ], + "lines": [ + { + "bbox": [ + 326, + 446, + 526, + 487 + ], + "spans": [ + { + "bbox": [ + 326, + 446, + 526, + 487 + ], + "type": "table", + "html": "
Loss#Max#MinACCNMIARI
CE6,4864,50151.650.235.0
SeCu5,1904,55688.179.477.6
", + "image_path": "42097bc2101780547c2338135441593327958447e73e84bfdb62cf182289d71e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 494, + 545, + 530 + ], + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 530 + ], + "type": "text", + "content": "Table 2: Comparison of loss functions on CIFAR-10. \"Max\" and \"Min\" indicate the maximal and minimal size of obtained clusters." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "First, we can observe that the proposed SeCu loss outperforms the standard cross entropy loss by a large margin of " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "36.5\\%" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " on ACC, which confirms the benefit of stable training. Then, by investigating the distribution of learned clusters, we find that the minimal size of clusters obtained from different loss functions is almost the same due to the lower-bound constraint, while the maximal size varies. The largest cluster obtained by SeCu contains 5,190 instances, which is close to the ground-truth distribution of CIFAR-10. On the contrary, the dominating cluster found by CE loss has 6,486 instances that is significantly different from the uniform distribution in CIFAR-10. It is because that the negative instances in CE loss can perturb the learning of centers and result in sub-optimal cluster assignments." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16651" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 85 + ], + "type": "text", + "content": "4.1.3 Effect of Updating Criterion for Cluster Centers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 286, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 286, + 138 + ], + "type": "text", + "content": "The cluster centers in SeCu can be optimized by SGD or updated by a closed-form solution as in Eqn. 4. We compare these two strategies and summarize the performance in Table 3." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 63, + 145, + 272, + 184 + ], + "blocks": [ + { + "bbox": [ + 63, + 145, + 272, + 184 + ], + "lines": [ + { + "bbox": [ + 63, + 145, + 272, + 184 + ], + "spans": [ + { + "bbox": [ + 63, + 145, + 272, + 184 + ], + "type": "table", + "html": "
MethodsACCNMIARI
SeCu-CF87.4±0.578.5±0.676.4±0.9
SeCu-SGD88.1±0.379.3±0.477.5±0.4
", + "image_path": "482a7b4599f579f1f67e9001b5f3075ef079b7dc9d7ac5d35f587d859c0fbb3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 230, + 286, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 286, + 350 + ], + "type": "text", + "content": "Let \"SeCu-CF\" and \"SeCu-SGD\" be the closed-form update and the SGD optimization, respectively. We can observe that SeCu-CF has the similar performance to SeCu-SGD, which shows the effectiveness of the closed-form updating. Moreover, SeCu-SGD has a smaller standard deviation than SeCu-CF due to the smoothness from the momentum in SGD. Compared with SeCu-CF, SeCu-SGD has an additional parameter of the learning rate for centers. Hence, SeCu-CF can be applied to evaluate the preliminary performance of SeCu with less tuning efforts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 364, + 222, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 222, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 222, + 376 + ], + "type": "text", + "content": "4.1.4 Effect of " + }, + { + "bbox": [ + 47, + 364, + 222, + 376 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 364, + 222, + 376 + ], + "type": "text", + "content": " in Entropy Constraint" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 382, + 286, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 286, + 418 + ], + "type": "text", + "content": "Besides the size constraint, the proposed entropy constraint is also effective for balancing cluster assignments. Table 4 demonstrates the results with different " + }, + { + "bbox": [ + 47, + 382, + 286, + 418 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 382, + 286, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 63, + 425, + 272, + 511 + ], + "blocks": [ + { + "bbox": [ + 47, + 193, + 286, + 217 + ], + "lines": [ + { + "bbox": [ + 47, + 193, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 286, + 217 + ], + "type": "text", + "content": "Table 3: Comparison of optimization for cluster centers on CIFAR-10. Mean±std over 8 trails are reported." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 425, + 272, + 511 + ], + "lines": [ + { + "bbox": [ + 63, + 425, + 272, + 511 + ], + "spans": [ + { + "bbox": [ + 63, + 425, + 272, + 511 + ], + "type": "table", + "html": "
α#Max#MinACCNMIARI
20,0005,0564,91385.376.373.1
8,0005,1154,65887.979.077.1
6,0005,1204,57988.079.477.4
4,0005,3764,32484.975.472.2
10028,575019.938.717.0
050,000010.00.00.0
", + "image_path": "9e8fb517f4e49ee3d58bfa2fa6982a42982ca87a3b2db7618b111a38e9ee7652.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 520, + 271, + 532 + ], + "lines": [ + { + "bbox": [ + 62, + 520, + 271, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 271, + 532 + ], + "type": "text", + "content": "Table 4: Comparison of " + }, + { + "bbox": [ + 62, + 520, + 271, + 532 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 62, + 520, + 271, + 532 + ], + "type": "text", + "content": " in Eqn. 10 on CIFAR-10." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "type": "text", + "content": "First, a large " + }, + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "type": "text", + "content": " will regularize the distribution to be uniform and result in a sub-optimal performance. By decreasing the weight, the assignment becomes more flexible and a desired performance can be observed when " + }, + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "type": "inline_equation", + "content": "\\alpha = 6,000" + }, + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "type": "text", + "content": ". However, a small " + }, + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "type": "text", + "content": " will lead to an imbalanced distribution, which is inappropriate for the balanced data sets such as CIFAR and STL. " + }, + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 46, + 546, + 286, + 653 + ], + "type": "text", + "content": " discards the constraint and incurs collapsing. Evidently, the entropy constraint can balance the size of clusters effectively" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": "Since the entropy is defined on the whole data set and CIFAR-10 contains 50,000 instances for training, " + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": " will be scaled as " + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "inline_equation", + "content": "6N / 50" + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": " for different data sets, where " + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": " denotes the total number of instances in the training set. More ablation experiments can be found in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 487, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 487, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 487, + 84 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 186 + ], + "type": "text", + "content": "After the ablation study, experiments are conducted on benchmark data sets to compare with state-of-the-art methods. The results averaged over 8 trials and the best performance among them are reported for our method. SeCu with size constraint and entropy constraint are denoted as \"SeCu-Size\" and \"SeCu-Entropy\", respectively. Table 5 shows the performance of different methods, where two-stage methods with a pre-training phase are marked for \"Two-stage\"." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "text", + "content": "First, as a one-stage method, SeCu outperforms two-stage methods by a margin of about " + }, + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "text", + "content": " over all metrics on CIFAR-10 and CIFAR-100-20. It demonstrates that learning representations and clustering simultaneously is essential for deep clustering. The two-stage method NNM[11] shows the comparable performance to SeCu on STL-10. It is because that STL-10 only constrains 5,000 target instances for training and both of NNM and SeCu already achieve the accuracy of supervised learning. Second, compared with the one-stage method without sufficient representation learning, i.e., PICA, SeCu demonstrates a better performance by about " + }, + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "text", + "content": " on CIFAR-10 and CIFAR-100-20, and " + }, + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "text", + "content": " on STL-10, which shows the importance of optimizing representations for instances. Third, with the same size constraint, SeCu surpasses CoKe by " + }, + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 304, + 186, + 545, + 449 + ], + "type": "text", + "content": " for ARI on CIFAR-10, which confirms the effectiveness of the proposed learning objective. Finally, compared with size constraint in SeCu-Size, SeCu-Entropy demonstrates a competitive performance over all data sets. Since the entropy constraint is optimized by enumeration without introducing any auxiliary variable, it can be an substitute to the size constraint for cluster assignments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 557 + ], + "type": "text", + "content": "After the clustering phase, self-labeling [14] is an effective strategy to further improve the performance. We try it for SeCu-Size with the setting in [14] and report the result in Table 5. Interestingly, self-labeling is also effective for SeCu and helps approach the supervised accuracy on CIFAR-10. As a one-stage method, SeCu has the strong augmentation for representation learning, which may introduce additional noise for clustering. Self-labeling has the weak augmentation for fine-tuning and refines clustering." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 310, + 569, + 421, + 608 + ], + "blocks": [ + { + "bbox": [ + 310, + 569, + 421, + 608 + ], + "lines": [ + { + "bbox": [ + 310, + 569, + 421, + 608 + ], + "spans": [ + { + "bbox": [ + 310, + 569, + 421, + 608 + ], + "type": "image", + "image_path": "491d7827b83f03d6386245ce0ed1eeb26bde28b70ca843b4fdb4f4e6619b3a49.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 608, + 386, + 616 + ], + "lines": [ + { + "bbox": [ + 349, + 608, + 386, + 616 + ], + "spans": [ + { + "bbox": [ + 349, + 608, + 386, + 616 + ], + "type": "text", + "content": "CIFAR-10" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 310, + 616, + 421, + 664 + ], + "blocks": [ + { + "bbox": [ + 310, + 616, + 421, + 664 + ], + "lines": [ + { + "bbox": [ + 310, + 616, + 421, + 664 + ], + "spans": [ + { + "bbox": [ + 310, + 616, + 421, + 664 + ], + "type": "image", + "image_path": "99998451b7ab208cdebf4ae06a508fc39cea43172bf9a543156afa1cf4c377dd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 666, + 381, + 675 + ], + "lines": [ + { + "bbox": [ + 353, + 666, + 381, + 675 + ], + "spans": [ + { + "bbox": [ + 353, + 666, + 381, + 675 + ], + "type": "text", + "content": "STL-10" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 685, + 509, + 697 + ], + "lines": [ + { + "bbox": [ + 340, + 685, + 509, + 697 + ], + "spans": [ + { + "bbox": [ + 340, + 685, + 509, + 697 + ], + "type": "text", + "content": "Figure 2: Illustration of exemplar images." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 422, + 569, + 542, + 664 + ], + "blocks": [ + { + "bbox": [ + 422, + 569, + 542, + 664 + ], + "lines": [ + { + "bbox": [ + 422, + 569, + 542, + 664 + ], + "spans": [ + { + "bbox": [ + 422, + 569, + 542, + 664 + ], + "type": "image", + "image_path": "76e829472ebecd858455b17dde3d09d35a61b7996e19c5ca6b98d529626617e7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 666, + 510, + 675 + ], + "lines": [ + { + "bbox": [ + 457, + 666, + 510, + 675 + ], + "spans": [ + { + "bbox": [ + 457, + 666, + 510, + 675 + ], + "type": "text", + "content": "CIFAR-100-20" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16652" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 514, + 305 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 514, + 305 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 514, + 305 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 514, + 305 + ], + "type": "table", + "html": "
MethodsTwo-stageCIFAR10CIFAR100-20STL10
ACCNMIARIACCNMIARIACCNMIARI
Supervised93.886.287.080.068.063.280.665.963.1
DeepCluster [4]37.4--18.9--33.4--
IIC [20]61.751.141.125.722.511.759.649.639.7
PICA [19] (best)69.659.151.233.731.017.171.361.153.1
Pretext [7]+k-means✓65.959.850.939.540.223.965.860.450.6
SCAN [14] (mean)✓81.871.266.542.244.126.775.565.459.0
NNM [11]✓84.374.870.947.748.431.680.869.465.0
GCC [32]✓85.676.472.847.247.230.578.868.463.1
CoKe [26] (mean)85.776.673.249.749.133.5---
SeCu-Size (mean)88.179.377.550.050.735.080.269.463.9
SeCu-Size (best)88.579.978.251.651.636.081.470.765.7
SeCu-Entropy (mean)88.079.277.249.950.634.179.568.763.0
SeCu-Entropy (best)88.279.777.751.251.434.980.569.964.4
with self-labeling:
SCAN [14]✓88.379.777.250.748.633.380.969.864.6
GCC [32]✓90.1--52.3--83.3--
SeCu93.086.185.755.255.139.783.673.369.3
", + "image_path": "5fe4766ffc752a311ea0d90b0d2c3f0cd9d46847c7d3336f0c28e7e57aec29a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 347, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 441 + ], + "type": "text", + "content": "Exemplar Images Finally, we show exemplar images from clusters obtained by SeCu for different data sets. Fig. 2 illustrates the images that are close to cluster centers in CIFAR-10, CIFAR-100-20, and STL-10, respectively. Evidently, the proposed method can obtain the ground-truth class structure for CIFAR-10 and STL-10. Moreover, with a large intra-class variance, SeCu can still observe appropriate clusters for CIFAR-100-20." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 443, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 525 + ], + "type": "text", + "content": "Considering that labels of super-classes in CIFAR-100-20 contain vague concepts such as \"vehicles 1\" and \"vehicles 2\", we include results on the target 100 classes in Table 6, which may benefit future research. The experiment settings for CIFAR-100 are identical to that for CIFAR-100-20, except the learning rate of centers, which is reduced to 0.01." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 70, + 533, + 265, + 608 + ], + "blocks": [ + { + "bbox": [ + 51, + 313, + 541, + 326 + ], + "lines": [ + { + "bbox": [ + 51, + 313, + 541, + 326 + ], + "spans": [ + { + "bbox": [ + 51, + 313, + 541, + 326 + ], + "type": "text", + "content": "Table 5: Comparison of clustering methods on benchmark data sets. \"Two-stage\" denotes an additional pre-training stage." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 533, + 265, + 608 + ], + "lines": [ + { + "bbox": [ + 70, + 533, + 265, + 608 + ], + "spans": [ + { + "bbox": [ + 70, + 533, + 265, + 608 + ], + "type": "table", + "html": "
MethodsACCNMIARI
SeCu-Size (mean)46.461.732.3
SeCu-Size (best)47.161.932.6
SeCu-Entropy (mean)45.360.731.5
SeCu-Entropy (best)46.761.032.2
SeCu†51.365.237.1
", + "image_path": "028a8dfb6023c0a1a3921f65091fc82bc87f41952f27adef5412bd6cc3b073fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 659, + 190, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 190, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 190, + 672 + ], + "type": "text", + "content": "4.3. Comparison on ImageNet" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Now we evaluate SeCu on the challenging ImageNet data set in Table 7, where the cost of many two-stage methods becomes intractable and only available baselines are" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "type": "text", + "content": "included. First, SeCu achieves " + }, + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "type": "inline_equation", + "content": "51.4\\%" + }, + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "type": "text", + "content": " ACC on ImageNet, which is better than SCAN with self-labeling by more than " + }, + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "type": "text", + "content": ". With the same 10 clustering heads and size constraint, SeCu is still " + }, + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "type": "text", + "content": " better than CoKe on ACC. Compared with the objective in CoKe, SeCu is tailored for deep clustering and the comparison confirms the efficacy of our method. With self-labeling, SeCu shows state-of-the-art performance of " + }, + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "type": "inline_equation", + "content": "53.5\\%" + }, + { + "bbox": [ + 304, + 346, + 545, + 456 + ], + "type": "text", + "content": " ACC on ImageNet. It demonstrates that our method is applicable for large-scale data sets." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 355, + 465, + 497, + 528 + ], + "blocks": [ + { + "bbox": [ + 46, + 616, + 287, + 641 + ], + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 641 + ], + "type": "text", + "content": "Table 6: Results of SeCu on CIFAR-100. † denotes a method with self-labeling." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 355, + 465, + 497, + 528 + ], + "lines": [ + { + "bbox": [ + 355, + 465, + 497, + 528 + ], + "spans": [ + { + "bbox": [ + 355, + 465, + 497, + 528 + ], + "type": "table", + "html": "
MethodsACCNMIARI
SCAN†39.972.027.5
CoKe‡47.676.235.6
SeCu51.478.039.7
SeCu†53.579.441.9
", + "image_path": "4271b570cdbf2fd5ebfcb6acb5b202013abd2a478e506400fd1b67dc6b0fad43.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 536, + 545, + 562 + ], + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 562 + ], + "type": "text", + "content": "Table 7: Comparison on ImageNet. " + }, + { + "bbox": [ + 304, + 536, + 545, + 562 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 304, + 536, + 545, + 562 + ], + "type": "text", + "content": " denotes a method with self-labeling. " + }, + { + "bbox": [ + 304, + 536, + 545, + 562 + ], + "type": "inline_equation", + "content": "{}^{ \\ddagger }" + }, + { + "bbox": [ + 304, + 536, + 545, + 562 + ], + "type": "text", + "content": " is our reproduction with 10 clustering heads." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "In this work, a novel stable cluster discrimination task that considers the hardness of instances is proposed for one-stage deep clustering. To avoid collapsing in representations, a global entropy constraint with theoretical guarantee is investigated. Finally, the comparison with state-of-the-art methods confirms the effectiveness of our proposal. While recent work shows that leveraging nearest neighbors improves deep clustering [11, 14, 32], incorporating such information in SeCu can be our future work." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16653" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "type": "text", + "content": "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 286, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 286, + 158 + ], + "type": "text", + "content": "[2] Hangbo Bao, Li Dong, and Furu Wei. Beit: BERT pretraining of image transformers. CoRR, abs/2106.08254, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 286, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 286, + 182 + ], + "type": "text", + "content": "[3] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 184, + 286, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 286, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 286, + 215 + ], + "type": "text", + "content": "[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 217, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 286, + 261 + ], + "type": "text", + "content": "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 263, + 286, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 263, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 263, + 286, + 306 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640. IEEE, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 308, + 286, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 286, + 351 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, volume 119, pages 1597-1607, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 354, + 286, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 286, + 386 + ], + "type": "text", + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758. Computer Vision Foundation / IEEE, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 388, + 286, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 286, + 419 + ], + "type": "text", + "content": "[9] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. CoRR, abs/2104.02057, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 422, + 286, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 286, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 286, + 454 + ], + "type": "text", + "content": "[10] Adam Coates, Andrew Y. Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In AISTATS, volume 15, pages 215-223. JMLR.org, 2011." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 456, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 286, + 498 + ], + "type": "text", + "content": "[11] Zhiyuan Dang, Cheng Deng, Xu Yang, Kun Wei, and Heng Huang. Nearest neighbor matching for deep clustering. In CVPR, pages 13693-13702. Computer Vision Foundation / IEEE, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 502, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 286, + 567 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 569, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 286, + 601 + ], + "type": "text", + "content": "[13] Ehsan Elhamifar and René Vidal. Sparse subspace clustering: Algorithm, theory, and applications. IEEE Trans. Pattern Anal. Mach. Intell., 35(11):2765-2781, 2013." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "type": "text", + "content": "[14] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. SCAN: learning to classify images without labels. In ECCV, volume 12355, pages 268-285. Springer, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 713 + ], + "type": "text", + "content": "[15] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent - A new approach to self-supervised learning. In NeurIPS, 2020." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 692 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. CoRR, abs/2111.06377, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9726-9735, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[19] Jiabo Huang, Shaogang Gong, and Xiatian Zhu. Deep semantic clustering by partition confidence maximisation. In CVPR, pages 8846-8855. Computer Vision Foundation / IEEE, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "text", + "content": "[20] Xu Ji, Andrea Vedaldi, and João F. Henriques. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, pages 9864-9873. IEEE, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 265, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 286 + ], + "type": "text", + "content": "[21] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 288, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 320 + ], + "type": "text", + "content": "[22] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In NeurIPS, pages 1106-1114, 2012." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 322, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 354 + ], + "type": "text", + "content": "[23] Yunfan Li, Peng Hu, Jerry Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. In AAAI, pages 8547-8555. AAAI Press, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 376 + ], + "type": "text", + "content": "[24] Stuart P. Lloyd. Least squares quantization in PCM. IEEE Trans. Inf. Theory, 28(2):129-136, 1982." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 378, + 545, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 378, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 307, + 378, + 545, + 411 + ], + "type": "text", + "content": "[25] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In ICCV, pages 6449-6457, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 413, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 413, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 413, + 545, + 454 + ], + "type": "text", + "content": "[26] Qi Qian, Yuanhong Xu, Juhua Hu, Hao Li, and Rong Jin. Unsupervised visual representation learning by online constrained k-means. In CVPR, pages 16619-16628. IEEE, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 457, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 457, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 457, + 545, + 511 + ], + "type": "text", + "content": "[27] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Fei-Fei Li. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis., 115(3):211-252, 2015." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 514, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 534 + ], + "type": "text", + "content": "[28] Ulrike von Luxburg. A tutorial on spectral clustering. Stat. Comput., 17(4):395-416, 2007." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 536, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 545, + 568 + ], + "type": "text", + "content": "[29] Junyuan Xie, Ross B. Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In ICML, volume 48, pages 478-487. JMLR.org, 2016." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "text", + "content": "[30] Bo Yang, Xiao Fu, Nicholas D. Sidiropoulos, and Mingyi Hong. Towards k-means-friendly spaces: Simultaneous deep learning and clustering. In ICML, 2017." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 604, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 604, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 545, + 646 + ], + "type": "text", + "content": "[31] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, volume 139, pages 12310-12320. PMLR, 2021." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 647, + 545, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 692 + ], + "type": "text", + "content": "[32] Huasong Zhong, Jianlong Wu, Chong Chen, Jianqiang Huang, Minghua Deng, Liqiang Nie, Zhouchen Lin, and Xian-Sheng Hua. Graph contrastive clustering. In ICCV, pages 9204-9213. IEEE, 2021." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16654" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/fdb72a4a-94ed-4829-ab6f-8cf80d64f7f5_content_list.json b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/fdb72a4a-94ed-4829-ab6f-8cf80d64f7f5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5e92e35b5e9059e3f7cb7762fad67ed520901d49 --- /dev/null +++ b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/fdb72a4a-94ed-4829-ab6f-8cf80d64f7f5_content_list.json @@ -0,0 +1,1707 @@ +[ + { + "type": "text", + "text": "Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations", + "text_level": 1, + "bbox": [ + 106, + 128, + 864, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuewei Yang, Hai Li, Yiran Chen", + "bbox": [ + 351, + 203, + 617, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Duke University", + "bbox": [ + 418, + 222, + 550, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Durham, USA", + "bbox": [ + 426, + 239, + 542, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yuwei.yang@duke.edu, hai.li@duke.edu, yiran.chen@duke.edu", + "bbox": [ + 225, + 258, + 743, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, discriminative self-supervised methods have made significant strides in advancing various visual tasks. The central idea of learning a data encoder that is robust to data distortions/augmentations is straightforward yet highly effective. Although many studies have demonstrated the empirical success of various learning methods, the resulting learned representations can exhibit instability and hinder downstream performance. In this study, we analyze discriminative self-supervised methods from a causal perspective to explain these unstable behaviors and propose solutions to overcome them. Our approach draws inspiration from prior works that empirically demonstrate the ability of discriminative self-supervised methods to demix ground truth causal sources to some extent. Unlike previous work on causality-empowered representation learning, we do not apply our solutions during the training process but rather during the inference process to improve time efficiency. Through experiments on both controlled image datasets and realistic image datasets, we show that our proposed solutions, which involve tempering a linear transformation with controlled synthetic data, are effective in addressing these issues.", + "bbox": [ + 75, + 342, + 473, + 675 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 707, + 209, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning generalized representation with unlabeled data is a challenging task in various fields, but Self-Supervised Learning (SSL) has recently demonstrated remarkable success in learning semantic invariant representations without labels [40, 41, 53]. There are two main types of self-supervised learning (SSL) based on the pretext task used: generative and discriminative SSL, with generative SSL reconstructing altered or distorted data to its original input [9, 28, 31, 59, 65, 71] and early discriminative SSL predicting easily designed labels and task-specific representations that are not very generalizable [25, 57, 75]. More recent", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/210d2f49e331f76ac6e8ec4cb2e6ed001d0b54603621d409dbce501dd562d9ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 306, + 839, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77582930b2b339d2c9e59af355f71daa3dbb052ba51491edb704d4a8f5985767.jpg", + "image_caption": [ + "Figure 1: During the training of discriminative SSL, aligning positive representations will be robust to the changes applied as augmentations (red arrows). However, during the inference stage, one small change in the data variable (such as view angle) will result in an unexpected degradation on downstream performance." + ], + "image_footnote": [], + "bbox": [ + 557, + 419, + 838, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "discriminative SSL trains the model to identify similarities and differences between pairs of augmented examples [7, 10, 11, 26, 29, 74]. The success of SSL in deep image models has resulted in progress in other data modalities [53, 52, 54, 61, 62] and attention-based models like transformers [12, 8, 49, 72]. Recent discriminative SSL aims to learn content and semantic invariant representations that are robust to data augmentations, but the learned representations can be unstable when one subtle factor of the data is changed to a value that is not accessible through all augmentations. To avoid the high cost of incorporating all possible subtle changes during training, insights are needed to uncover the root cause of instability and find a solution to prevent performance deterioration during inference. Figure 1 summarizes this deterioration effect.", + "bbox": [ + 496, + 657, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Causality [60] is a vital tool to investigate the causal", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16109", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "relationships between variables in observational data, and can uncover the underlying causal factors that explain unexpected model behavior due to changes in the environment. Independent Component Analysis (ICA) is often used to disentangle sources in unsupervised training [36, 37, 39, 43, 45], and causal analysis has been applied in follow-up works [76, 67] to examine the empirical success of SSL under an ICA framework. However, while these works identify factors that contribute to SSL's success, they do not address the problem's unstable mode, which can cause a severe performance drop when underlying factors shift slightly to an unseen environment. Some works [56, 27, 48] attempt to incorporate causality during the training process to identify and alleviate the impact of such shifts, but this approach is time-costly and only marginally improves performance compared to non-causal SSL methods. A more time-efficient and accessible solution would be to simply reverse the unstable shift during inference.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We aim to address the issue of unstable behavior during the inference stage by building upon previous theories of successful InfoNCE-facilitated contrastive SSL and extending it to all recent SSL methods with additional assumptions and constraints. Drawing inspiration from the relationship between the ground truth positive pairs distribution and learned positive pairs distribution, we demonstrate that the approximated transformation between the ground truth representations and learned representations is orthogonal to the augmentations applied during training. However, a change in the data factor/variable that violates the conditions for successful SSL can cause a corresponding shift in the inferred representation, resulting in a decline in downstream performance. This change of data factor/variable can be a change in the background, texture, or view angles etc. To overcome this issue, we propose learning targeted transformations that regularize the violating shift and restore performance on the unseen data shift. This approach effectively avoids the undesirable behavior and improves performance on previously unseen data shifts.", + "bbox": [ + 75, + 363, + 472, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are following:", + "bbox": [ + 96, + 666, + 410, + 681 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Through the use of a comparable data generation process in prior research, we show that All current SSL techniques benefit from the alignment of positive pairs.", + "- Through our alternative derivation of the transformation matrix between the ground truth representation and the learned representation, we have shown that the augmentations applied during training are orthogonal to the resulting matrix.", + "- By interpreting a change in the data variable causally, we propose two solutions focusing on inference to modify the negative shift in representation space caused by such a change." + ], + "bbox": [ + 94, + 694, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We validate the proposed solutions by conducting experiments on both controlled and realistic datasets, providing evidence for their efficacy during the inference stage without retraining the pretrained models.", + "bbox": [ + 517, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 165, + 640, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Discriminative SSL learns invariant representations from positive pairs of unlabeled data samples, but previous attempts to use trivial labels like colors[75], rotations[25], and patch positions[57] offer minimal benefits for complex downstream tasks due to their easy augmentations. Recent discriminative SSL apply random augmentations to generate two views of an image sample and train an encoder to extract representations for maximizing similarity between the paired augmentations. SimCLR and MoCo[10, 29] are pioneer SSL works that maximize the cosine similarity between positive pairs and minimize the cosine similarity between negative pairs via optimizing the InfoNCE loss[58]. Immense resources are used to enforce a large number of negative samples since a larger number can tighten the upper bound of the mutual information between positive pairs[58]. Later advancement of discriminative SSL excludes the notion of negative pairs by only aligning positive pairs and preventing representation collapse through various regularizations. BYOL[26] predicts an Exponential-Moving-Averaged (EMA) representation of one view with a projected representation of another view. SimSiam[11] maximizes similarity between a projected representation and a detached representation of two positive samples. Unlike previous work focusing optimization on an instance level, Barlow Twins[74] encourages high similarity in corresponding feature dimensions and discourages redundancy across different feature dimensions between two views of a data sample. Detailed formulations are exhibited in A.", + "bbox": [ + 496, + 190, + 893, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mutual Information is a different perspective on the behaviour of discriminative SSL. Referred to InfoMax principle[51], the MI between different transformations of a data sample is maximized via optimizing the InfoNCE loss[2, 32, 44]. Though showing theoretical relation between optimizing InfoNCE and maximizing MI between positive pairs, the underlying factors instructing the behaviour of different SSL methods are not explored. Nonlinear ICA[38], on the other hand, captures complex data structures of SSL methods by disentangling underlying factors via minimizing the mutual information between learned representations and the original data input[17, 73, 20]. Other works associate the nonlinear ICA objectives with the contrastive SSL so the MI between positive pairs are maximized and negative pairs are minimized.", + "bbox": [ + 496, + 613, + 893, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other researchers have explored causality and causal inference as a means of understanding the success of discriminative SSL. Prior work has focused on partitioning the InfoNCE loss to alignment between the positive pairs and", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16110", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "uniformity between aggregations of all positive clusters[70]. By formulating a data generation process, [76] empirically explains that networks optimized via InfoNCE infer an orthogonal transformation of the ground truth latent representations. Furthermore [67] validates that that augmentations used in both generative and discriminative SSL isolate the content factor from the style factor. Our theory and work draw great inspirations from these two work. However, instead of solely focusing on InfoNCE-driven SSL and two factors, this work extends the framework to all recent discriminative SSL methods and identifies reasons for unstable circumstances analytically. We also propose methods to nullify the negative effects of unstable representations during inference.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain Adaptation is a strategy to bridge the gap between the model performance on a source domain and that on a target domain [4, 68]. Feature adaptation methods try to learn a new feature representation that is more invariant to the domain shift[66, 19, 15, 50], while instance adaptation methods try to reweight the importance of the labeled source examples to better align with the target domain[55, 6, 34]. Recent studies also implement a contrastive framework to learn a shared latent space between the source and target domains by maximizing the agreement between representations of corresponding samples while minimizing the agreement between representations of noncorresponding samples [69, 63, 42]. Unlike these works focusing on adapting to a target domain for better performance, we investigate the underlying causal factors for the performance gap and based on the findings we propose easy solutions to connect the gap.", + "bbox": [ + 75, + 301, + 473, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Theory", + "text_level": 1, + "bbox": [ + 76, + 571, + 163, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we build on previous data generation process [76, 67] (3.1) and show that current SSL techniques benefit from the alignment of positive pairs (3.2). During a deeper dive into the generation process, we show the relation between the ground truth representations and the inferred representations, and this relation is an linear transformation matrix that is orthogonal to the augmentations applied during training (3.3). Finally, we disclose the causal reason for the unstable behaviour caused by a change in the data variable during the inference stage and propose analytical solutions to address this unstable issue (3.4).", + "bbox": [ + 75, + 597, + 470, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem formulation", + "text_level": 1, + "bbox": [ + 76, + 771, + 277, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data generation is assumed to be a generation function that takes ground truth latent representation as the input to generate an observation data/image.", + "bbox": [ + 75, + 794, + 468, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, we assume that the marginal distribution of sampling ground truth representations $\\mathbf{z} \\subseteq \\mathcal{Z} \\in \\mathbb{R}^{d_1}$ w.r.t. a class is uniform on a unit sphere $\\mathbb{S}^{d_1 - 1}$ . An injective generation function $g(.) : \\mathbb{R}^{d_1} \\to \\mathbb{R}^d$ takes a ground truth rep-", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/99cf6be555efe2964c0444eec817c0c6c366220f40744e28d9f136218478a52f.jpg", + "image_caption": [ + "Figure 2: A causal graph for data generation process." + ], + "image_footnote": [], + "bbox": [ + 642, + 87, + 766, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resentation and generate an observation sample $\\mathbf{x} \\in \\mathbb{R}^d$ , $\\mathbf{x} = g(\\mathbf{z})$ . Different variables, denoted as a set $V = \\{V_i\\}$ and class/content variable $c$ , constitute the values of each dimension in the ground truth representation. These variables can be direct causes or confounding factors, $\\mathbf{z} = (c, \\{V_i\\})$ . Specific examples include view angles, object size, background colors, etc. A simple causal graph depicts this relationship is shown in Figure 2. So the general generation process is as following:", + "bbox": [ + 496, + 268, + 893, + 407 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {z} = (c, V _ {*})) \\sim \\frac {1}{| \\mathcal {Z} |} \\quad \\mathbf {x} = g (\\mathbf {z}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 416, + 890, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To sample a positive example w.r.t the same class, we assume the conditional distribution is a von Mises-Fisher (vMF) distribution [21]:", + "bbox": [ + 496, + 458, + 893, + 503 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (\\tilde {\\mathbf {z}} = (c, \\tilde {V} _ {*}) | \\mathbf {z} = (c, V _ {*})) = C _ {p} ^ {- 1} e ^ {\\kappa_ {1} \\mathbf {z} ^ {\\top} \\tilde {\\mathbf {z}}} \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 513, + 890, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C_p = \\int e^{\\kappa_1\\mathbf{z}^\\top \\tilde{\\mathbf{z}}}d\\tilde{\\mathbf{z}}$ is a normalization constant and $\\kappa$ is a concentration parameter.", + "bbox": [ + 496, + 547, + 890, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Representation learning is a process that a feature encoder, $f(\\cdot):\\mathbb{R}^d\\to \\mathbb{R}^{d_2}$ , extracts representations from two positive observations $f(\\tilde{\\mathbf{x}}) = f\\circ g(\\tilde{\\mathbf{z}})$ , $f(\\mathbf{x}) = f\\circ g(\\mathbf{z})$ and a distribution associated with the encoder $f$ through $h = f\\circ g$ is:", + "bbox": [ + 496, + 580, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq _ {h} (\\tilde {\\mathbf {z}} | \\mathbf {z}) = C _ {q} ^ {- 1} (\\mathbf {z}) e ^ {\\kappa_ {2} h (\\tilde {\\mathbf {z}}) ^ {\\top} h (\\mathbf {z})} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 654, + 890, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $C_q(\\mathbf{z}) = \\int e^{\\kappa_2 h(\\tilde{\\mathbf{z}})^\\top h(\\mathbf{z})} d\\tilde{\\mathbf{z}}$ be the normalization term. Optimizing any discriminative SSL objective will maximize the similarity between these positive pairs. An example of a well constructed objective is the InfoNCE loss:", + "bbox": [ + 496, + 681, + 890, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n f o N C E} = \\mathbb {E} \\left[ - l o g \\frac {e ^ {f (\\tilde {\\mathbf {x}}) ^ {\\top} f (\\mathbf {x}) / \\tau}}{e ^ {f (\\tilde {\\mathbf {x}}) ^ {\\top} f (\\mathbf {x}) / \\tau} + \\sum_ {i = 1} ^ {K} e ^ {f \\left(\\mathbf {x} _ {i} ^ {-}\\right) ^ {\\top} f (\\mathbf {x}) / \\tau}} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 755, + 890, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{x}}$ is a positive example w.r.t. $\\mathbf{x}$ in the observation space and $\\{\\mathbf{x}_i^-\\}_{1}^{K}$ are $K$ samples from distributions of all observations. The global minimum of (4) is reached when", + "bbox": [ + 496, + 806, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1The mapping of representations on a hypersphere may be different to Barlow Twins methodology, but as shown in [74], normalize representations on a unit sphere also works under Barlow Twins.", + "bbox": [ + 496, + 862, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16111", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the cosine similarity between positive pairs is maximized and the cosine similarity between all negative pairs is minimized. In the following section, we will show that discriminative SSL including InfoNCE-driven and non InfoNCE-driven (EMA and Siamese with a predictor) follows strict rules of alignment to maximize the similarity.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Alignment in discriminative SSL", + "text_level": 1, + "bbox": [ + 76, + 191, + 366, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section we will combine theories stated in [76, 67, 70] so that the general factors of successful discriminative SSL can be summarized. Additionally, instead of focusing on just InfoNCE SSL variation and content block-identifiability [35, 47], we extend the combined theory to demonstrate that all discriminative SSL benefit from alignment between positive representations.", + "bbox": [ + 75, + 215, + 470, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3.1 With a data generation process described in 3.1, all discriminative SSL objectives have an alignment loss function between positive pairs from the network:", + "bbox": [ + 76, + 335, + 472, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\left\\| f (\\mathbf {x}) - f (\\tilde {\\mathbf {x}}) \\right\\| _ {2} ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 393, + 468, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This is a weaker statement than Theorem 4.4 in [67] since we only focus on the alignment term. For analysis on the regularization term on the network entropy, see B.", + "bbox": [ + 75, + 424, + 468, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. For InfoNCE-driven SSL (SimCLR and MoCo), as derived in [68, 76], the InfoNCE loss converges to an alignment term and a uniformity term as the number of negative samples approaches infinity. (See B for full details).", + "bbox": [ + 75, + 469, + 468, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For EMA-based SSL methods (BYOL), a predictor $p$ is associated with the online network so the SSL objective becomes $\\mathcal{L} = \\mathbb{E}_{(\\mathbf{x},\\tilde{\\mathbf{x}})}\\| p\\circ f(\\mathbf{x},\\theta) - f(\\tilde{\\mathbf{x}},\\xi)\\| _2^2$ , where $\\xi^t = \\alpha \\xi^{t - 1} + (1 - \\alpha)\\theta^t$ is target network parameter and $\\theta$ is the online network parameter. Denote $p^\\prime = p\\circ f$ . By adding and subtracting $p^\\prime (\\tilde{\\mathbf{x}},\\theta)$ we derive:", + "bbox": [ + 76, + 531, + 472, + 623 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\| p ^ {\\prime} (\\mathbf {x}, \\theta) - p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) + p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - f (\\tilde {\\mathbf {x}}, \\xi) \\| _ {2} ^ {2} \\\\ = \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\| p ^ {\\prime} (\\mathbf {x}, \\theta) - p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) \\| _ {2} ^ {2} + \\underset {\\tilde {\\mathbf {x}}} {\\mathbb {E}} \\| p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - f (\\tilde {\\mathbf {x}}, \\xi) \\| _ {2} ^ {2} \\\\ - 2 \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\left[ \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - p ^ {\\prime} (\\mathbf {x}, \\theta)\\right) ^ {\\top} \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - f (\\tilde {\\mathbf {x}}, \\xi)\\right) \\right] \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 632, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $f(\\mathbf{x},\\theta)$ and $f(\\tilde{\\mathbf{x}},\\xi)$ maps in the same space $\\mathbb{R}^{d_2}$ , $p$ can be considered as a bijective linear transformation within $\\mathbb{R}^{b_2}$ . In fact, the performance difference between a linear predictor and non-linear predictor is subtle. The global minimizer of (6) must align network output w.r.t to $(\\mathbf{x},\\tilde{\\mathbf{x}})$ with the first term in (6) and align outputs from different networks with second term in (6). Hence completes the proof.", + "bbox": [ + 75, + 748, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For Siamese network with a predictor (SimSiam), a similar approach can reformulate the objective as it is a special case when $f(\\tilde{\\mathbf{x}},\\xi) = f(\\tilde{\\mathbf{x}},\\theta)$ and a stop-gradient is", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "applied on the $f(\\tilde{\\mathbf{x}},\\theta)$ . Hence by substituting $s_{\\mathcal{G}}(f(\\tilde{\\mathbf{x}},\\theta))$ with $f(\\tilde{\\mathbf{x}},\\xi)$ in (6), we complete the proof by:", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\| p ^ {\\prime} (\\mathbf {x}, \\theta) - p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) \\| _ {2} ^ {2} \\\\ + \\underset {\\tilde {\\mathbf {x}}} {\\mathbb {E}} \\| p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - \\operatorname {s g} (f (\\tilde {\\mathbf {x}}, \\theta)) \\| _ {2} ^ {2} \\\\ - 2 \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\left[ \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - p ^ {\\prime} (\\mathbf {x}, \\theta)\\right) ^ {\\top} \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - \\operatorname {s g} (f (\\tilde {\\mathbf {x}}, \\theta))\\right) \\right] \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 132, + 890, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the third term in (6) and (7) can be formulated by the differential entropy of $H(p^{\\prime}(.))$ hence prevent representation collapse.", + "bbox": [ + 498, + 244, + 890, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For Barlow Twins, the diagonal of the cross-correlation matrix $C_{ii}$ is the cosine similarity between positive pairs. Hence completes the proof by:", + "bbox": [ + 498, + 290, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {i} ^ {d _ {2}} \\left(1 - C _ {i i}\\right) ^ {2} = \\sum_ {i} \\left(1 - \\left(f (\\mathbf {x}) ^ {\\top} f (\\tilde {\\mathbf {x}})\\right) _ {i} / \\left(d _ {2} - 1\\right)\\right) ^ {2} (8) \\\\ = \\sum_ {i} \\left(\\| f (\\mathbf {x}) _ {i} - f (\\tilde {\\mathbf {x}}) _ {i} \\| _ {2} ^ {2} / \\left(2 * \\left(d _ {2} - 1\\right)\\right) ^ {2} \\right. (9) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 349, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Transformation of the ground-truth factors is orthogonal to the applied augmentations", + "text_level": 1, + "bbox": [ + 500, + 436, + 890, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By demonstrating that all discriminative SSL have a alignment loss term, the transformation between the ground truth representations and the inferred representations can be derived as in [76]. But different to [76], our derivation of minimization of cross entropy is assumed to be a lower bound since we only include the alignment term and the uniformity term is always positive [14, 5]. However, with all SSL objectives there are additional terms to maximize the output entropy of the model (some described in 3.2). So optimizing SSL objectives as a complete loss function will minimize the cross entropy $\\mathbb{E}[H(p(.|\\mathbf{z}),q_h(.|\\mathbf{z}))]$", + "bbox": [ + 496, + 474, + 892, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3.2 By considering the generation conditional distribution as $p(\\tilde{\\mathbf{z}}|\\mathbf{z}) = C_p^{-1}e^{\\kappa_1\\mathbf{z}^\\top \\tilde{\\mathbf{z}}}$ , the inferred conditional distribution $q_h(\\tilde{\\mathbf{z}}|\\mathbf{z})$ can match $p(\\tilde{\\mathbf{z}}|\\mathbf{z})$ by minimizing $\\| h(\\mathbf{z}) - h(\\tilde{\\mathbf{z}})\\|_2^2$ and $\\forall \\mathbf{z}, \\tilde{\\mathbf{z}} : \\kappa \\mathbf{z}^\\top \\tilde{\\mathbf{z}} = h(\\mathbf{z})^\\top \\tilde{\\mathbf{z}}$ with $h = f \\circ g$ or $h = p \\circ f \\circ g$ maps onto a hypersphere with radius $\\sqrt{\\kappa_1 / \\kappa_2}$ .", + "bbox": [ + 498, + 657, + 893, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof exactly follows Proposition 1 in [76] just with minor modification on the concentration term $\\kappa$ and $h$ so that Theorem 3.2 can apply to non-InfoNCE SSL. A global minimizer of the alignment term $\\| h(\\mathbf{z}) - h(\\tilde{\\mathbf{z}})\\| _2^2$ will also minimize the cross entropy of between $p(\\tilde{\\mathbf{z}} |\\mathbf{z})$ in (2) and $q_{h}(\\tilde{\\mathbf{z}} |\\mathbf{z})$ in (3). This indicates that minimizers of SSL objective alignment maintain the dot product. Then we can use Proposition 2 in [76] directly to show that $h$ is an orthogonal linear transformation.", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16112", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3.3 Assume the data generation process (cf. 3.1), a model parameterized by $h = f \\circ g$ or $h = p \\circ f \\circ g$ ( $h: \\mathbb{R}^{d_1} \\to \\mathbb{R}^{d_2}$ ) that minimizes the alignment term in all discriminative SSL objectives: $\\| h(\\mathbf{z}) - h(\\tilde{\\mathbf{z}}) \\|_2^2$ as (15), $h$ is an orthogonal transformation: $h(\\tilde{\\mathbf{z}}) = \\mathbf{A}\\tilde{\\mathbf{z}}$ where $\\mathbf{A}$ is an orthogonal matrix.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proof follows that a function $h$ minimizes the alignment will also minimize the cross entropy between the ground truth conditional distribution $p(\\tilde{\\mathbf{z}}|\\mathbf{z})$ and the inferred conditional distribution $q_{h}(\\tilde{\\mathbf{z}}|\\mathbf{z})$ . Therefore if $h$ is isometric w.r.t the dot product as indicated in Theorem 3.2 then $h(\\tilde{\\mathbf{z}}) = \\mathbf{A}\\tilde{\\mathbf{z}}$ according to Proposition 2 in [76].", + "bbox": [ + 76, + 191, + 472, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 3.4 Assume augmentations applied during the training can be represented as a change in the ground truth representations, i.e. $\\tilde{\\mathbf{x}} = g(\\tilde{\\mathbf{z}}(c,\\tilde{V}*))$ and change in data variables induces a shift in the ground truth representations $\\delta \\mathbf{z} = (c,\\tilde{V}_*) - (c,V_*)$ , then $\\delta \\mathbf{z}$ is in orthogonal to $\\mathbf{A}$ i.e. $\\mathbf{A}\\delta \\mathbf{z} = 0$ if a discriminative SSL objective is optimized.", + "bbox": [ + 76, + 292, + 472, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. With augmentations normally applied during SSL such as color distortions, rotations, random cropping, and etc., one can view the alteration as a change of a variable in the data variable (color, view angles, sizes). This change of variable under the generation framework described in 3.1 will result in a change in $V = \\{V_{i}\\}$ since $c$ is not changed. Regardless of structure of the causal graph shown in Figure 2, the change in $V$ can be reflected in the ground truth representation space as $\\tilde{\\mathbf{z}} = \\mathbf{z} + \\delta \\mathbf{z}$ . A global minimizer of any discriminative SSL objective will minimize the alignment term $h(\\mathbf{z}) = h(\\tilde{\\mathbf{z}}) = h(\\mathbf{z} + \\delta \\mathbf{z})$ . According to Theorem 3.3 we derive:", + "bbox": [ + 75, + 392, + 472, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh (\\mathbf {z}) = \\mathbf {A} \\mathbf {z} = \\mathbf {A} (\\mathbf {z} + \\delta \\mathbf {z}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 583, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} \\delta \\mathbf {z} = 0 \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 602, + 468, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\mathbf{A}$ being a linear orthogonal matrix. This indicates that the transformation $A$ is learned to annul the change in data variables or the effect of augmentations applied.", + "bbox": [ + 76, + 627, + 468, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Reason and solutions for unstable change in data variables", + "text_level": 1, + "bbox": [ + 76, + 681, + 468, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In section 3.3 we show that generalized representation is robust to augmentations since $\\mathbf{A}$ , the linear transformation between the ground truth representation and inferred representations, is orthogonal to augmentation applied during the training. And if a change in the data variable reflects a shift, $\\delta \\mathbf{z}$ , in the ground truth representations and the newly inferred representation is stable, meaning $\\delta \\mathbf{z}$ will be absorbed by the transformation matrix $\\mathbf{A}$ , then the shift in the ground truth representation corresponds to an augmentation that is applied during the training. However, when $\\delta \\mathbf{z}$ appears in the range of $\\mathbf{A}$ and $\\mathbf{A} \\delta \\mathbf{z} \\neq 0$ , the resultant inferred representation can be unstable and lead to performance drop", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on downstream models (denote as $D(\\mathbf{z})$ ). We quantify this deterioration on $D$ by:", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm \\left(D (h (\\mathbf {z})) _ {\\text {s t a b l e}}\\right) - m \\left(D \\left(h (\\tilde {\\mathbf {z}}) _ {\\text {u n s t a b l e}}\\right) \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 132, + 890, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m(.)$ is a metric on an outcome of the downstream task. In example of $D$ being a linear classifier, $m(.)$ can be the probability of predicting the target class (prediction score) or the proportion of correct predictions (accuracy). In order to overcome this deterioration, we propose two methods, namely Robust Dimensions and Stable Inference Mapping:", + "bbox": [ + 496, + 160, + 893, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Robust Dimensions: Under a stable condition, $D(f(\\mathbf{x}))^2 = D(\\mathbf{A}\\mathbf{z})$ , each dimension in the inferred representation $f(\\mathbf{x})$ is a linear combination of dimensions of the ground truth representation. The dimensions contributing most to $D()$ should be robust to unstable shift $\\delta \\mathbf{z}_{\\text{unstable}}$ . In other words, most robust dimensions of $f(\\mathbf{x})$ should be also robust in $f(\\tilde{\\mathbf{x}})$ where $\\tilde{\\mathbf{x}} = g(\\mathbf{z} + \\delta \\mathbf{z}_{\\text{unstable}})$ as some dimensions of $\\mathbf{A} \\delta \\mathbf{z}_{\\text{unstable}}$ will be zero. Hence identifying most important dimensions in $f(\\mathbf{x})_{\\text{stable}}$ and pass through the same dimensions of $f(\\tilde{\\mathbf{x}})_{\\text{unstable}}$ should alleviate the deterioration by making $m(D(h(\\mathbf{z}))_{\\text{stable}})_{\\{\\dim\\}} = m(D(h(\\tilde{\\mathbf{z}})_{\\text{unstable}}))_{\\{\\dim\\}}$ where $\\{\\dim\\}$ is a set of most robust dimensions. In example of $D$ being a linear classifier, the contribution of each dimension can be calculated by $W_c^\\top f(\\mathbf{x})$ where $W_c^\\top$ is the Jacobian of the linear classifier w.r.t target class $c$ .", + "bbox": [ + 511, + 275, + 893, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2. Stable Inference Mapping: Since $f(\\mathbf{x})_{\\text{stable}} = -f(\\tilde{\\mathbf{x}})_{\\text{unstable}} = -\\mathbf{A} \\delta \\mathbf{z}$ , we can learn another linear transformation $\\mathbf{F}$ to absorb $\\delta \\mathbf{z}$ . Especially, we want to learn $\\mathbf{F} f(\\tilde{\\mathbf{z}})_{\\text{unstable}} = \\mathbf{F} \\mathbf{A} \\mathbf{z} + \\mathbf{F} \\mathbf{A} \\delta \\mathbf{z}$ such that the additional $\\mathbf{F}$ will not only set $\\mathbf{A} \\delta z$ to 0, but also make stable representation more robust to the unstable shift by assuring $\\mathbf{F} \\mathbf{A}$ orthogonal to $\\delta \\mathbf{z}_{\\text{unstable}}$ in addition to the augmentations applied during training. Formally, we model a linear layer $l(f(\\tilde{\\mathbf{x}})_{\\text{unstable}}) = f(\\mathbf{x})_{\\text{stable}} - f(\\tilde{\\mathbf{x}})_{\\text{unstable}}$ , hence during the inference $f(\\mathbf{x}) + l(f(\\mathbf{x}))$ is used for downstream task.", + "bbox": [ + 511, + 541, + 893, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Relation to Causal Inference Since we only have observations in the image space, we consider the augmentations or changes of a data variable as interventions on the ground truth representations: $\\tilde{\\mathbf{x}} = g(\\tilde{\\mathbf{z}} |(c,do(V_i = v_i))$ . With an access to ground truth representation, we can evaluate treatment-control effect i.e. $Pr(f(g(\\tilde{\\mathbf{z}} |(c,do(V_i = v_{stable}))) - Pr(f(g(\\tilde{\\mathbf{z}} |(c,do(V_i = v_{unstable})))$ . Without any access to the ground truth representations, we can evaluate the average treatment effect [33] by $\\mathbb{E}[D(f(\\mathbf{x})_{stable})-$", + "bbox": [ + 496, + 717, + 893, + 854 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2Note that in general $f$ contains a projector. However we exclude the notion of the projector to simplify the problem and [10, 29, 26] show that a projector is not necessary in SSL.", + "bbox": [ + 496, + 862, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16113", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$D(f(\\tilde{\\mathbf{x}})_{\\text{unstable}})]$ via synthesizing manual data samples $\\tilde{\\mathbf{x}}_i = g(\\tilde{\\mathbf{z}} | (c, \\text{do} (V_i = v_i))$ .", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments and Discussions", + "text_level": 1, + "bbox": [ + 76, + 137, + 346, + 155 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section we evaluate our solutions to unstable shift in data variable on two datasets: Causal3DIdent and ImageNet pretrained SSL and corresponding linear classifiers as the downstream task.", + "bbox": [ + 76, + 164, + 468, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Causal3DIdent", + "text_level": 1, + "bbox": [ + 76, + 236, + 230, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eec2275428e8faf79fdde2e91d218a79d519c3d158d53b9f7ff570a4c454a55e.jpg", + "image_caption": [ + "Figure 3: Causal graph for $V$ in Causal3DIdent." + ], + "image_footnote": [], + "bbox": [ + 153, + 268, + 393, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[76] develops the dataset that 7 objects in the dataset that each $224 \\times 224$ image is associated with an object class and a 10-dimensional latent representation. These 10 dimensions correspond to 10 data variables. A causal graph imposed on these variables is shown in Figure 3. As the dependency shown in Figure 3, an object is placed at $pos_{obj} = (x,y,z)$ with $rot_{obj} = (\\phi, \\theta, \\psi)$ and $hue_{o}bj$ under a spot light at an angle $pos_{spl}$ with color $hue_{spl}$ on a background with color $hue_{bg}$ . Detailed information in C.", + "bbox": [ + 75, + 407, + 468, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To simulate unstable change of variable that is not accessible during the training, a range of some data variables is hold out and hence this portion of data is treated as potential unstable samples that causes the performance drop on the linear classifier. Since all data variables are truncated in range [-1,1], we hold out the edge value(s) to further portray 'unexpected' values during the inference time. We select 8 dimensions to intervene, namely: $z$ in object position3, all 3 object rotation angles, spot light position, and all 3 hue variables. Both training and testing data are sampled and for detailed sampling procedures refer to C.", + "bbox": [ + 75, + 545, + 468, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SSL Experiment Setup ResNet18[30] is the backbone of the feature encoder $f$ . Same augmentations in [10] are applied during training. An Adam optimizer with a learning rate at 0.0001 and a weight decay at 0.00001 is optimized for all discriminative SSL. Hyperparameters for each SSL is presented in C. The dimension of the inferred representation space is set to 128. The network is trained for 20 epochs on intervened data. Then a linear classifier is trained on the frozen representations of the network with a same optimizer for 10 epochs.", + "bbox": [ + 75, + 710, + 468, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To illustrate the simulation results in the deterioration of the downstream performance, for each testing data that has the same data variable distribution as the training data, we search in the ground truth representation space for 5 nearest representations in the hold-out test distribution when $n$ dimension of the testing data is shifted to a random value in the hold-out range of corresponding data variable. Within the corresponding 5 images representing a shift in $n$ dimensions that is not seen during the training, we record the lowest performance to fulfill the deterioration scenario. In Figure 4 the score and accuracy are averaged over all combinations of $\\binom{8}{n}$ . When $n = 0$ , this indicates the performance of SSL on the testing data with the same data variable distribution as the training data. The performance is exceptional since hold-out values of some data variables may be covered by the augmentations applied ( $z$ covered by random cop, and hue colors covered by color distortions). A will be orthogonal to changes in these data variables. In fact in C we show that the performance difference between seen distribution and unseen distribution in testing data is comparatively small. As illustrated in Figure 4, even when only one variable is changed to a unseen value, there is a large drop in both accuracy and prediction score (20% in accuracy and 30% in prediction score for SimCLR). With more tangled changes in data variables, the unstable representations result in poorer downstream performance. Since the selection of the data variables may be too complex due to the dependency, we also validate the same issue on selecting only children nodes in Figure 3. And also we visualize the latent shift between stable and unstable examples. See C for more results.", + "bbox": [ + 500, + 90, + 890, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robust Dimensions For each pair of testing data $\\mathbf{x}_{\\text{stable}}$ with seen data variable distribution and the selected data $\\mathbf{x}_{\\text{unstable}}$ among the 5 nearest neighbours when changing $n$ dimensions to unseen distribution values, we apply the Jacobian of the linear classifier w.r.t the target class $W_{c}^{\\top}$ on the stable representation $f(\\mathbf{x})_{\\text{stable}}$ to identify the top $k$ most important dimensions and pass the same dimensions of $f(\\tilde{\\mathbf{x}})_{\\text{unstable}}$ to $D$ to evaluate the performance. As shown in Figure 5, the accuracy only deteriorate slightly when top $90\\%$ most important features of $f(\\tilde{\\mathbf{x}})_{\\text{unstable}}$ are selected for the downstream task. This is true even when all 8 variables are shifted. This suggests that $\\mathbf{A}$ is orthogonal to changes in ground truth representation $\\mathbf{z}$ in most dimensions. This high percentage of dimensions may be due to $f$ optimizing the SSL objectives to a high level and the augmentations applied covers some of the hold-out variable values. Interestingly, there are cases where passing the top $k\\%$ (around $40\\%$ ) important dimensions results in higher performance than stable representations (uninterrupted in Figure 5). However, as expected, including more less important dimensions where the unseen shift in the data variable results in non-zero adjustment ( $\\mathbf{A} \\delta \\mathbf{z} \\neq 0$ ) initiates the", + "bbox": [ + 498, + 568, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "${}^{3}$ Only $z$ is altered because most of deep vision models are translation invariant.", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16114", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/21cc225abf728803edbfd65cec6a68b81ee71a8a6786f030d15c49190fc88e6a.jpg", + "image_caption": [ + "(a) Deterioration in Accuracy" + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 483, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2fb9f3265c94903183cf9b34812e200e1b9ef60bae532cf6d45e99763d04480.jpg", + "image_caption": [ + "(b) Deterioration in Score", + "Figure 4: Deterioration of unstable changing in data variables on all SSL. The more unexpected changes occur in the data variables, the more severe the deterioration." + ], + "image_footnote": [], + "bbox": [ + 486, + 89, + 875, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "deterioration phenomenon we observe.", + "bbox": [ + 76, + 275, + 334, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c1ed25508b7d95acd0189f359673b4a542b7f92f937c9ba9d4059d09b8521cb.jpg", + "image_caption": [ + "(a) Improvement on Accuracy" + ], + "image_footnote": [], + "bbox": [ + 78, + 316, + 486, + 443 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6199777575f6d3f01cb0c27e3d47513d3562264b32832d64ca7ccd7d152bd88.jpg", + "image_caption": [ + "(b) Improvement on Score", + "Figure 5: By identifying and passing top $k\\%$ important dimensions, the deterioration in both accuracy and prediction score is significantly alleviated." + ], + "image_footnote": [], + "bbox": [ + 81, + 465, + 483, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stable Inference Mapping As shown in 3.4, a linear transformation $\\mathbf{F}:\\mathbb{R}^{d_2}\\to \\mathbb{R}^{d_2}$ is trained to cancel the effect $\\mathbf{A}\\delta \\mathbf{z}$ and also further improve the robustness of transformation on the ground truth representation $\\mathbf{Az}$ . In this experiment, we match each training data with a random member in the 5 nearest representations when only one dimension is changed to a unstable value. A linear layer is trained with the same optimizer for 10 epochs on the training pairs. In Table 1, all accuracy for unstable examples increases significantly (except SimSiam) after learning $\\mathbf{F}$ . However, since the $f$ is very close to the global minimizer of the alignment term, the improvement on stable examples cannot be observed.", + "bbox": [ + 75, + 703, + 472, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3dc36aef3ab7feb50c4307ca308b89f1cfd534cbe4f182666557cb2f111af614.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
xstablex unstable
w/o Fw/ Fw/o Fw/ F
SimCLR0.9960.9980.8330.889
MoCo0.9920.9920.8000.849
BYOL0.9960.9980.8860.920
SimSiam0.9980.9990.9190.928
Barlow Twins0.9910.9950.8180.862
", + "bbox": [ + 522, + 273, + 870, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: The effect of $\\mathbf{F}$ on both stable and unstable samples. The accuracy is average over 3 random seeds.", + "bbox": [ + 498, + 395, + 890, + 426 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a06d872cf06ffeae1ff707465a95603fe0cb65a1b43151084ed33d857fed1a38.jpg", + "image_caption": [ + "Figure 6: Examples of ObjectNet dataset $(1^{st}$ row), Stylized ImageNet dataset $(2^{nd}$ row), Synthetic dataset $(3^{rd}$ row)." + ], + "image_footnote": [], + "bbox": [ + 581, + 443, + 812, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. ImageNet", + "text_level": 1, + "bbox": [ + 500, + 739, + 611, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate our findings on a larger scale and more realistic settings, we apply proposed solutions on ImageNet[16] with various altered and synthetic datasets as unstable shift in the data variable. See Figure 6 and D for more information on ObjectNet, Stylized-ImageNet, Synthetic dataset.", + "bbox": [ + 496, + 763, + 890, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ObjectNet[3] is a collection of objects that are intentionally placed at an unusual view angle and backgrounds, so that the bias learned by the model with usual data variables is more prominent when testing on ObjectNet. With focus-", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16115", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing on 113 classes overlapping with ImageNet classes, we evaluate our second proposed solution explicitly on ObjectNet since the shift presented in ObjectNet is very unstable and it is very challenging to overcome the negative effect of the shift.", + "bbox": [ + 75, + 90, + 470, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stylized ImageNet[24] change the style of original ImageNet image to a random artistic style. With this drastic shift in the data variable, the analysis of the first solution is more insightful on the robustness of dimensions of $\\mathbf{A}$ on a very different shift in data variable.", + "bbox": [ + 75, + 167, + 468, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Synthetic Data follows synthetic procedure in [18] where object is masked on a background at a location with a rotation angle. We explore the benefit of this synthesized dataset at three modes: background, location, rotation where the target variable is randomly sampled with the other two variables fixed. Additionally, we explore texture as an independent variable by masking 'texturized' objects in [23]. We set total number of updating steps per epoch as 100 with batch size 256. This means we select total number of 512000 synthesized images every epoch.", + "bbox": [ + 75, + 244, + 468, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiment Setup ResNet 50 pretrained on ImageNet and a linear classifier fintuned via SimCLR, BYOL, and SimSiam are tested with both proposed solutions. For Stable Inference Mapping, a linear layer is optimized with an Adam optimizer for 10 epochs on the synthetic dataset.", + "bbox": [ + 75, + 398, + 468, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robust Dimensions For each of ImageNet validation data sample, we stylized the image to a random artistic fashion. We observe the dramatic performance difference between the ImageNet stable images and Stylized unstable ImageNet images $(\\mathbf{x}_{stable}, \\mathbf{x}_{unstable})$ . The result is shown in Figure 7. For SimCLR, passing the top $10\\%$ important dimensions can close a small performance gap between stable representations and unstable representations. Nonetheless, all SSL seem to be sensible to the strong style change as they shorten the difference between ImageNet to an modest extent. This is as expected since the Stylized ImageNet changes multiple variables to an extreme value.", + "bbox": [ + 75, + 476, + 468, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Stable Inference Mapping As described in Synthetic Data, we explore the benefit of our second proposed method to evaluate on a dataset that most of samples are unstable according to learned $\\mathbf{A}$ . At each training step, with other variables randomly fixed, 10 images with random target variable values are generated. The pair with maximum $m(D(\\mathbf{x})) - m(D(\\tilde{\\mathbf{x}}))$ is selected to train the linear transformation. In Table 2, inferring $\\mathbf{F}$ via controlling location produces least improvement. This is expected since the network is robust to translation by design. While background, rotation, and texture improves the performance considerably with the consideration on the training time. However, in D we show that training the model longer using Stable Inference Mapping yield less favorable results since the improvement is less significant and starts to saturate at around 30 epochs.", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ce5deaa8597f4d839b7edf29964ce209f5b1c963cba81d281dbb1832ab57ae0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 890, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef24a24ca9ebf8c9b8309833600ec769da956d94d114a591c3b9ed67c0842140.jpg", + "image_caption": [ + "(a) Improvement on Accuracy", + "(b) Improvement on Score", + "Figure 7: By identifying and passing top $k\\%$ important dimensions, the deterioration in both accuracy and prediction score is alleviated." + ], + "image_footnote": [], + "bbox": [ + 501, + 224, + 890, + 335 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8e40c42063bd03e289ada23d88398d34c0518222c5b8f86437faee2ea22db0de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SimCLRBYOLSimSiam
w/o F10.1214.0411.15
background12.3415.4312.45
location10.3514.1311.17
rotation12.0415.8812.38
texture12.7915.0613.11
", + "bbox": [ + 535, + 433, + 857, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: Comparison of inferring different $\\mathbf{F}$ on ObjectNet with a target data variable random trials.", + "bbox": [ + 498, + 541, + 890, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 500, + 601, + 622, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Though we identify the root cause of unstable inference for discriminative SSL by constructing a causal framework inspired by the prior work, the proposed solutions are constrained and limited to be applied on realistic applications. Robust Dimensions involves establishing a correspondence between stable and unstable instances on a one-to-one basis, enabling the identification of dimensions contributing to stability. On the other hand, Stable Inference Mapping necessitates a collection of unstable instances with a specific alteration in a particular group of data variables. Within the Causal3dIdent dataset, both solutions can be assessed using the same unstable instances. In more realistic datasets, achieving a one-to-one correspondence is feasible, and manipulation of one group of data variables can be accomplished using synthetic data. However, any assessments with involving artificially generated images might introduce some level of uncertainty. In a realistic setup, since training samples are not directly observ-", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16116", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "able during the inference stage, simple interventions on inference samples may not effectively separate the unstable variables from the stable ones. Consequently, the potential benefits of the proposed solutions in realistic datasets are undermined.", + "bbox": [ + 76, + 90, + 472, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 181, + 204, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In conclusion, this paper has proposed a novel approach to address the issue of unstable behavior during the inference stage in SSL methods. By building on the previous theories of successful InfoNCE-facilitated contrastive SSL and extending it to recent SSL methods, we have demonstrated that a change in the data factor can result in a shift in the inferred representation, leading to a decline in downstream performance. We have proposed learning targeted transformations that regularize the violating shift and restore performance on the unseen data shift. Our experiments on both controlled and realistic datasets have shown the efficacy of our proposed solutions. These contributions provide a better understanding of SSL methods and offer a promising solution to the problem of unstable behavior during the inference stage. We hope that our work will inspire further research in this area and lead to improved SSL methods that are more robust to changes in data factors.", + "bbox": [ + 76, + 205, + 472, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 478, + 264, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work was funded in part by NSF CNS-2112562 and IIS-2140247.", + "bbox": [ + 76, + 503, + 468, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 547, + 174, + 564 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ibrahim Ahmad and Pi-Erh Lin. A nonparametric estimation of the entropy for absolutely continuous distributions (corresp.). IEEE Transactions on Information Theory, 22(3):372-375, 1976. 14", + "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2", + "[3] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 7", + "[4] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006. 3", + "[5] Sergiy V Borodachov, Douglas P Hardin, and Edward B Saff. Discrete energy on rectifiable sets, volume 3. Springer, 2019. 4", + "[6] Yue Cao, Mingsheng Long, and Jianmin Wang. Unsupervised domain adaptation with distribution matching ma" + ], + "bbox": [ + 86, + 573, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "chines. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 3", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 1", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1", + "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 5, 6, 13", + "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1, 2", + "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 1", + "[13] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 17", + "[14] Henry Cohn and Abhinav Kumar. Universally optimal distribution of points on spheres. Journal of the American Mathematical Society, 20(1):99-148, 2007. 4", + "[15] Hal Daumé III. Frustratingly easy domain adaptation. arXiv preprint arXiv:0907.1815, 2009. 3", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7", + "[17] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516, 2014. 2", + "[18] Josip Djolonga, Jessica Yung, Michael Tschannen, Rob Romijnders, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Matthias Minderer, Alexander D'Amour, Dan Moldovan, Sylvain Gelly, Neil Houlsby, Xiaohua Zhai, and Mario Lucic. On robustness and transferability of convolutional neural networks, 2020. 8, 16", + "[19] Lixin Duan, Dong Xu, and Ivor Tsang. Learning with augmented features for heterogeneous domain adaptation. arXiv preprint arXiv:1206.4660, 2012. 3", + "[20] Deniz Erdogmus, Kenneth E Hild, Yadunandana N Rao, and Jose C Principe. Minimax mutual information approach for independent component analysis. Neural Computation, 16(6):1235-1252, 2004. 2" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16117", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Ronald Aylmer Fisher. Dispersion on a sphere. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 217(1130):295-305, 1953. 3", + "[22] Bernhard C. Geiger and Gernot Kubin. On the information loss in memoryless systems: The multivariate case. CoRR, abs/1109.4856, 2011. 14", + "[23] Robert Geirhos, Claudio Michaelis, Felix A. Wichmann, Patricia Rubisch, Matthias Bethge, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. 7th International Conference on Learning Representations, ICLR 2019, (c):1-22, 2019. 8", + "[24] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 8", + "[25] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 1, 2", + "[26] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 5", + "[27] Jeff Z. HaoChen, Colin Wei, Adrien Gaidon, and Tengyu Ma. Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Loss. Advances in Neural Information Processing Systems, 7(NeurIPS):5000-5011, 2021. 2", + "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1", + "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 5", + "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[31] Geoffrey E Hinton and Richard Zemel. Autoencoders, minimum description length and helmholtz free energy. Advances in neural information processing systems, 6, 1993. 1", + "[32] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670, 2018. 2", + "[33] Paul W Holland. Statistics and causal inference. Journal of the American statistical Association, 81(396):945-960, 1986. 5", + "[34] Tzu Ming Harry Hsu, Wei Yu Chen, Cheng-An Hou, Yao-Hung Hubert Tsai, Yi-Ren Yeh, and Yu-Chiang Frank Wang." + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Unsupervised domain adaptation with imbalanced cross-domain data. In Proceedings of the IEEE International Conference on Computer Vision, pages 4121-4129, 2015. 3", + "[35] Aapo Hyvarinen and Patrik Hoyer. Emergence of phase-and shift-invariant features by decomposition of natural images into independent feature subspaces. Neural computation, 12(7):1705-1720, 2000. 4", + "[36] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and Nonlinear ICA. Advances in Neural Information Processing Systems, (Nips):3772-3780, 2016. 2", + "[37] Aapo Hyvarinen and Hiroshi Morioka. Nonlinear ICA of temporally dependent stationary sources. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 2017. 2", + "[38] Aapo Hyvarinen and Petteri Pajunen. Nonlinear independent component analysis: Existence and uniqueness results. Neural networks, 12(3):429-439, 1999. 2", + "[39] Aapo Hyvarinen, Hiroaki Sasaki, and Richard E. Turner. Nonlinear ICA using auxiliary variables and generalized contrastive learning. AISTATS 2019 - 22nd International Conference on Artificial Intelligence and Statistics, 89, 2020. 2", + "[40] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1):2, 2020. 1", + "[41] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 1", + "[42] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4893-4902, 2019. 3", + "[43] Ilyes Khemakhem, Diederik P. Kingma, Ricardo Pio Monti, and Aapo Hyvarinen. Variational Autoencoders and Nonlinear ICA: A Unifying Framework. 108(2015), 2019. 2", + "[44] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 2", + "[45] David Klindt, Lukas Schott, Yash Sharma, Ivan Ustyuzhaninov, Wieland Brendel, Matthias Bethge, and Dylan Paiton. Towards Nonlinear Disentanglement in Natural Data with Temporal Sparse Coding. pages 1-51, 2020. 2", + "[46] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 16", + "[47] Quoc V Le, Will Y Zou, Serena Y Yeung, and Andrew Y Ng. Learning hierarchical invariant spatio-temporal features" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16118", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for action recognition with independent subspace analysis. In CVPR 2011, pages 3361-3368. IEEE, 2011. 4", + "[48] Jason D. Lee, Qi Lei, Nikunj Saunshi, and Jiacheng Zhuo. Predicting What You Already Know Helps: Provable Self-Supervised Learning. Advances in Neural Information Processing Systems, 1(NeurIPS):309-323, 2021. 2", + "[49] Chunyuan Li, Jianwei Yang, Pengchuan Zhang, Mei Gao, Bin Xiao, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Efficient Self-supervised Vision Transformers for Representation Learning. 3:1-27, 2021. 1", + "[50] Wen Li, Lixin Duan, Dong Xu, and Ivor W Tsang. Learning with augmented features for supervised and semi-supervised heterogeneous domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 36(6):1134-1148, 2013. 3", + "[51] Ralph Linsker. Self-organization in a perceptual network. Computer, 21(3):105–117, 1988. 2", + "[52] Shuo Liu, Adria Mallol-Ragolta, Emilia Parada-Cabaleiro, Kun Qian, Xin Jing, Alexander Kathan, Bin Hu, and Bjöern W Schuller. Audio self-supervised learning: A survey. Patterns, 3(12):100616, 2022. 1", + "[53] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering, 35(1):857-876, 2021. 1", + "[54] Yixin Liu, Ming Jin, Shirui Pan, Chuan Zhou, Yu Zheng, Feng Xia, and Philip Yu. Graph self-supervised learning: A survey. IEEE Transactions on Knowledge and Data Engineering, 2022. 1", + "[55] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer joint matching for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1410-1417, 2014. 3", + "[56] Jovana Mitrovic, Brian McWilliams, Jacob Walker, Lars Buesing, and Charles Blundell. Representation Learning via Invariant Causal Mechanisms. pages 1-21, 2020. 2", + "[57] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI, pages 69-84. Springer, 2016. 1, 2", + "[58] Aaron Van Den Oord. Representation Learning with Contrastive Predictive Coding. 2", + "[59] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 1", + "[60] Judea Pearl. Causality: Models, reasoning, and inference, second edition. 2011. 1", + "[61] Madeline C Schiappa, Yogesh S Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ACM Computing Surveys, 2022. 1", + "[62] Saeed Shurrab and Rehab Duwairi. Self-supervised learning methods and applications in medical imaging analysis: A survey. PeerJ Computer Science, 8:e1045, 2022. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Mamatha Thota and Georgios Leontidis. Contrastive domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2209-2218, 2021. 3", + "[64] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 15", + "[65] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 1", + "[66] Riccardo Volpi, Pietro Morerio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5495-5504, 2018. 3", + "[67] Kügelgen Von, Julius and, Wieland Brendel, Bernhard Schölkopf, Michel Besserve, and Francesco Locatello. Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style Luigi Gresele. Yash Sharma, 3(NeurIPS):4, 2021. 2, 3, 4, 15", + "[68] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3, 4", + "[69] Rui Wang, Zuxuan Wu, Zejia Weng, Jingjing Chen, Guo-Jun Qi, and Yu-Gang Jiang. Cross-domain contrastive learning for unsupervised domain adaptation. IEEE Transactions on Multimedia, 2022. 3", + "[70] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. 37th International Conference on Machine Learning, ICML 2020, PartF168147-13:9871-9881, 2020. 3, 4", + "[71] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 1", + "[72] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-Supervised Learning with Swin Transformers. pages 1-8, 2021. 1", + "[73] Detai Xin, Tatsuya Komatsu, Shinnosuke Takamichi, and Hiroshi Saruwatari. Disentangled speaker and language representations using mutual information minimization and domain adaptation for cross-lingual tts. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6608-6612. IEEE, 2021. 2", + "[74] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 1, 2, 3, 14, 15", + "[75] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, Octo" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16119", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ber 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 1, 2", + "[76] Roland S. Zimmermann, Yash Sharma, Steffen Schneider, Matthias Bethge, and Wieland Brendel. Contrastive Learning Inverts the Data Generating Process. 2021. 2, 3, 4, 5, 6, 15" + ], + "bbox": [ + 78, + 90, + 467, + 174 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16120", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/fdb72a4a-94ed-4829-ab6f-8cf80d64f7f5_model.json b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/fdb72a4a-94ed-4829-ab6f-8cf80d64f7f5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..046e30d7785311d9f572908a9e438c45df6c09e7 --- /dev/null +++ b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/fdb72a4a-94ed-4829-ab6f-8cf80d64f7f5_model.json @@ -0,0 +1,2611 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.865, + 0.176 + ], + "angle": 0, + "content": "Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.204, + 0.619, + 0.222 + ], + "angle": 0, + "content": "Yuewei Yang, Hai Li, Yiran Chen" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.223, + 0.551, + 0.24 + ], + "angle": 0, + "content": "Duke University" + }, + { + "type": "text", + "bbox": [ + 0.427, + 0.24, + 0.544, + 0.256 + ], + "angle": 0, + "content": "Durham, USA" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.26, + 0.744, + 0.275 + ], + "angle": 0, + "content": "yuwei.yang@duke.edu, hai.li@duke.edu, yiran.chen@duke.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.474, + 0.676 + ], + "angle": 0, + "content": "In recent years, discriminative self-supervised methods have made significant strides in advancing various visual tasks. The central idea of learning a data encoder that is robust to data distortions/augmentations is straightforward yet highly effective. Although many studies have demonstrated the empirical success of various learning methods, the resulting learned representations can exhibit instability and hinder downstream performance. In this study, we analyze discriminative self-supervised methods from a causal perspective to explain these unstable behaviors and propose solutions to overcome them. Our approach draws inspiration from prior works that empirically demonstrate the ability of discriminative self-supervised methods to demix ground truth causal sources to some extent. Unlike previous work on causality-empowered representation learning, we do not apply our solutions during the training process but rather during the inference process to improve time efficiency. Through experiments on both controlled image datasets and realistic image datasets, we show that our proposed solutions, which involve tempering a linear transformation with controlled synthetic data, are effective in addressing these issues." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.708, + 0.21, + 0.724 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Learning generalized representation with unlabeled data is a challenging task in various fields, but Self-Supervised Learning (SSL) has recently demonstrated remarkable success in learning semantic invariant representations without labels [40, 41, 53]. There are two main types of self-supervised learning (SSL) based on the pretext task used: generative and discriminative SSL, with generative SSL reconstructing altered or distorted data to its original input [9, 28, 31, 59, 65, 71] and early discriminative SSL predicting easily designed labels and task-specific representations that are not very generalizable [25, 57, 75]. More recent" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.308, + 0.841, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.42, + 0.839, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.54, + 0.895, + 0.632 + ], + "angle": 0, + "content": "Figure 1: During the training of discriminative SSL, aligning positive representations will be robust to the changes applied as augmentations (red arrows). However, during the inference stage, one small change in the data variable (such as view angle) will result in an unexpected degradation on downstream performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.884 + ], + "angle": 0, + "content": "discriminative SSL trains the model to identify similarities and differences between pairs of augmented examples [7, 10, 11, 26, 29, 74]. The success of SSL in deep image models has resulted in progress in other data modalities [53, 52, 54, 61, 62] and attention-based models like transformers [12, 8, 49, 72]. Recent discriminative SSL aims to learn content and semantic invariant representations that are robust to data augmentations, but the learned representations can be unstable when one subtle factor of the data is changed to a value that is not accessible through all augmentations. To avoid the high cost of incorporating all possible subtle changes during training, insights are needed to uncover the root cause of instability and find a solution to prevent performance deterioration during inference. Figure 1 summarizes this deterioration effect." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Causality [60] is a vital tool to investigate the causal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16109" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "relationships between variables in observational data, and can uncover the underlying causal factors that explain unexpected model behavior due to changes in the environment. Independent Component Analysis (ICA) is often used to disentangle sources in unsupervised training [36, 37, 39, 43, 45], and causal analysis has been applied in follow-up works [76, 67] to examine the empirical success of SSL under an ICA framework. However, while these works identify factors that contribute to SSL's success, they do not address the problem's unstable mode, which can cause a severe performance drop when underlying factors shift slightly to an unseen environment. Some works [56, 27, 48] attempt to incorporate causality during the training process to identify and alleviate the impact of such shifts, but this approach is time-costly and only marginally improves performance compared to non-causal SSL methods. A more time-efficient and accessible solution would be to simply reverse the unstable shift during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.666 + ], + "angle": 0, + "content": "We aim to address the issue of unstable behavior during the inference stage by building upon previous theories of successful InfoNCE-facilitated contrastive SSL and extending it to all recent SSL methods with additional assumptions and constraints. Drawing inspiration from the relationship between the ground truth positive pairs distribution and learned positive pairs distribution, we demonstrate that the approximated transformation between the ground truth representations and learned representations is orthogonal to the augmentations applied during training. However, a change in the data factor/variable that violates the conditions for successful SSL can cause a corresponding shift in the inferred representation, resulting in a decline in downstream performance. This change of data factor/variable can be a change in the background, texture, or view angles etc. To overcome this issue, we propose learning targeted transformations that regularize the violating shift and restore performance on the unseen data shift. This approach effectively avoids the undesirable behavior and improves performance on previously unseen data shifts." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.667, + 0.411, + 0.682 + ], + "angle": 0, + "content": "To summarize, our contributions are following:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.695, + 0.469, + 0.74 + ], + "angle": 0, + "content": "- Through the use of a comparable data generation process in prior research, we show that All current SSL techniques benefit from the alignment of positive pairs." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.752, + 0.469, + 0.828 + ], + "angle": 0, + "content": "- Through our alternative derivation of the transformation matrix between the ground truth representation and the learned representation, we have shown that the augmentations applied during training are orthogonal to the resulting matrix." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- By interpreting a change in the data variable causally, we propose two solutions focusing on inference to modify the negative shift in representation space caused by such a change." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.695, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "- We validate the proposed solutions by conducting experiments on both controlled and realistic datasets, providing evidence for their efficacy during the inference stage without retraining the pretrained models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.166, + 0.642, + 0.182 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.191, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Discriminative SSL learns invariant representations from positive pairs of unlabeled data samples, but previous attempts to use trivial labels like colors[75], rotations[25], and patch positions[57] offer minimal benefits for complex downstream tasks due to their easy augmentations. Recent discriminative SSL apply random augmentations to generate two views of an image sample and train an encoder to extract representations for maximizing similarity between the paired augmentations. SimCLR and MoCo[10, 29] are pioneer SSL works that maximize the cosine similarity between positive pairs and minimize the cosine similarity between negative pairs via optimizing the InfoNCE loss[58]. Immense resources are used to enforce a large number of negative samples since a larger number can tighten the upper bound of the mutual information between positive pairs[58]. Later advancement of discriminative SSL excludes the notion of negative pairs by only aligning positive pairs and preventing representation collapse through various regularizations. BYOL[26] predicts an Exponential-Moving-Averaged (EMA) representation of one view with a projected representation of another view. SimSiam[11] maximizes similarity between a projected representation and a detached representation of two positive samples. Unlike previous work focusing optimization on an instance level, Barlow Twins[74] encourages high similarity in corresponding feature dimensions and discourages redundancy across different feature dimensions between two views of a data sample. Detailed formulations are exhibited in A." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Mutual Information is a different perspective on the behaviour of discriminative SSL. Referred to InfoMax principle[51], the MI between different transformations of a data sample is maximized via optimizing the InfoNCE loss[2, 32, 44]. Though showing theoretical relation between optimizing InfoNCE and maximizing MI between positive pairs, the underlying factors instructing the behaviour of different SSL methods are not explored. Nonlinear ICA[38], on the other hand, captures complex data structures of SSL methods by disentangling underlying factors via minimizing the mutual information between learned representations and the original data input[17, 73, 20]. Other works associate the nonlinear ICA objectives with the contrastive SSL so the MI between positive pairs are maximized and negative pairs are minimized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Other researchers have explored causality and causal inference as a means of understanding the success of discriminative SSL. Prior work has focused on partitioning the InfoNCE loss to alignment between the positive pairs and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "16110" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "uniformity between aggregations of all positive clusters[70]. By formulating a data generation process, [76] empirically explains that networks optimized via InfoNCE infer an orthogonal transformation of the ground truth latent representations. Furthermore [67] validates that that augmentations used in both generative and discriminative SSL isolate the content factor from the style factor. Our theory and work draw great inspirations from these two work. However, instead of solely focusing on InfoNCE-driven SSL and two factors, this work extends the framework to all recent discriminative SSL methods and identifies reasons for unstable circumstances analytically. We also propose methods to nullify the negative effects of unstable representations during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.474, + 0.562 + ], + "angle": 0, + "content": "Domain Adaptation is a strategy to bridge the gap between the model performance on a source domain and that on a target domain [4, 68]. Feature adaptation methods try to learn a new feature representation that is more invariant to the domain shift[66, 19, 15, 50], while instance adaptation methods try to reweight the importance of the labeled source examples to better align with the target domain[55, 6, 34]. Recent studies also implement a contrastive framework to learn a shared latent space between the source and target domains by maximizing the agreement between representations of corresponding samples while minimizing the agreement between representations of noncorresponding samples [69, 63, 42]. Unlike these works focusing on adapting to a target domain for better performance, we investigate the underlying causal factors for the performance gap and based on the findings we propose easy solutions to connect the gap." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.165, + 0.589 + ], + "angle": 0, + "content": "3. Theory" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.471, + 0.765 + ], + "angle": 0, + "content": "In this section, we build on previous data generation process [76, 67] (3.1) and show that current SSL techniques benefit from the alignment of positive pairs (3.2). During a deeper dive into the generation process, we show the relation between the ground truth representations and the inferred representations, and this relation is an linear transformation matrix that is orthogonal to the augmentations applied during training (3.3). Finally, we disclose the causal reason for the unstable behaviour caused by a change in the data variable during the inference stage and propose analytical solutions to address this unstable issue (3.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.279, + 0.788 + ], + "angle": 0, + "content": "3.1. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.841 + ], + "angle": 0, + "content": "Data generation is assumed to be a generation function that takes ground truth latent representation as the input to generate an observation data/image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Formally, we assume that the marginal distribution of sampling ground truth representations \\(\\mathbf{z} \\subseteq \\mathcal{Z} \\in \\mathbb{R}^{d_1}\\) w.r.t. a class is uniform on a unit sphere \\(\\mathbb{S}^{d_1 - 1}\\). An injective generation function \\(g(.) : \\mathbb{R}^{d_1} \\to \\mathbb{R}^d\\) takes a ground truth rep-" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.088, + 0.767, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.227, + 0.871, + 0.243 + ], + "angle": 0, + "content": "Figure 2: A causal graph for data generation process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.895, + 0.408 + ], + "angle": 0, + "content": "resentation and generate an observation sample \\(\\mathbf{x} \\in \\mathbb{R}^d\\), \\(\\mathbf{x} = g(\\mathbf{z})\\). Different variables, denoted as a set \\(V = \\{V_i\\}\\) and class/content variable \\(c\\), constitute the values of each dimension in the ground truth representation. These variables can be direct causes or confounding factors, \\(\\mathbf{z} = (c, \\{V_i\\})\\). Specific examples include view angles, object size, background colors, etc. A simple causal graph depicts this relationship is shown in Figure 2. So the general generation process is as following:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.417, + 0.892, + 0.451 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {z} = (c, V _ {*})) \\sim \\frac {1}{| \\mathcal {Z} |} \\quad \\mathbf {x} = g (\\mathbf {z}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.895, + 0.505 + ], + "angle": 0, + "content": "To sample a positive example w.r.t the same class, we assume the conditional distribution is a von Mises-Fisher (vMF) distribution [21]:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.515, + 0.892, + 0.538 + ], + "angle": 0, + "content": "\\[\np (\\tilde {\\mathbf {z}} = (c, \\tilde {V} _ {*}) | \\mathbf {z} = (c, V _ {*})) = C _ {p} ^ {- 1} e ^ {\\kappa_ {1} \\mathbf {z} ^ {\\top} \\tilde {\\mathbf {z}}} \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.582 + ], + "angle": 0, + "content": "where \\(C_p = \\int e^{\\kappa_1\\mathbf{z}^\\top \\tilde{\\mathbf{z}}}d\\tilde{\\mathbf{z}}\\) is a normalization constant and \\(\\kappa\\) is a concentration parameter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Representation learning is a process that a feature encoder, \\( f(\\cdot):\\mathbb{R}^d\\to \\mathbb{R}^{d_2} \\), extracts representations from two positive observations \\( f(\\tilde{\\mathbf{x}}) = f\\circ g(\\tilde{\\mathbf{z}}) \\), \\( f(\\mathbf{x}) = f\\circ g(\\mathbf{z}) \\) and a distribution associated with the encoder \\( f \\) through \\( h = f\\circ g \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.655, + 0.892, + 0.677 + ], + "angle": 0, + "content": "\\[\nq _ {h} (\\tilde {\\mathbf {z}} | \\mathbf {z}) = C _ {q} ^ {- 1} (\\mathbf {z}) e ^ {\\kappa_ {2} h (\\tilde {\\mathbf {z}}) ^ {\\top} h (\\mathbf {z})} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.746 + ], + "angle": 0, + "content": "with \\( C_q(\\mathbf{z}) = \\int e^{\\kappa_2 h(\\tilde{\\mathbf{z}})^\\top h(\\mathbf{z})} d\\tilde{\\mathbf{z}} \\) be the normalization term. Optimizing any discriminative SSL objective will maximize the similarity between these positive pairs. An example of a well constructed objective is the InfoNCE loss:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.756, + 0.892, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n f o N C E} = \\mathbb {E} \\left[ - l o g \\frac {e ^ {f (\\tilde {\\mathbf {x}}) ^ {\\top} f (\\mathbf {x}) / \\tau}}{e ^ {f (\\tilde {\\mathbf {x}}) ^ {\\top} f (\\mathbf {x}) / \\tau} + \\sum_ {i = 1} ^ {K} e ^ {f \\left(\\mathbf {x} _ {i} ^ {-}\\right) ^ {\\top} f (\\mathbf {x}) / \\tau}} \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.892, + 0.854 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{x}}\\) is a positive example w.r.t. \\(\\mathbf{x}\\) in the observation space and \\(\\{\\mathbf{x}_i^-\\}_{1}^{K}\\) are \\(K\\) samples from distributions of all observations. The global minimum of (4) is reached when" + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1The mapping of representations on a hypersphere may be different to Barlow Twins methodology, but as shown in [74], normalize representations on a unit sphere also works under Barlow Twins." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16111" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "the cosine similarity between positive pairs is maximized and the cosine similarity between all negative pairs is minimized. In the following section, we will show that discriminative SSL including InfoNCE-driven and non InfoNCE-driven (EMA and Siamese with a predictor) follows strict rules of alignment to maximize the similarity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.193, + 0.367, + 0.21 + ], + "angle": 0, + "content": "3.2. Alignment in discriminative SSL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.471, + 0.324 + ], + "angle": 0, + "content": "In this section we will combine theories stated in [76, 67, 70] so that the general factors of successful discriminative SSL can be summarized. Additionally, instead of focusing on just InfoNCE SSL variation and content block-identifiability [35, 47], we extend the combined theory to demonstrate that all discriminative SSL benefit from alignment between positive representations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.473, + 0.385 + ], + "angle": 0, + "content": "Theorem 3.1 With a data generation process described in 3.1, all discriminative SSL objectives have an alignment loss function between positive pairs from the network:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.394, + 0.47, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\left\\| f (\\mathbf {x}) - f (\\tilde {\\mathbf {x}}) \\right\\| _ {2} ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.47 + ], + "angle": 0, + "content": "This is a weaker statement than Theorem 4.4 in [67] since we only focus on the alignment term. For analysis on the regularization term on the network entropy, see B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.531 + ], + "angle": 0, + "content": "Proof. For InfoNCE-driven SSL (SimCLR and MoCo), as derived in [68, 76], the InfoNCE loss converges to an alignment term and a uniformity term as the number of negative samples approaches infinity. (See B for full details)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.532, + 0.473, + 0.624 + ], + "angle": 0, + "content": "For EMA-based SSL methods (BYOL), a predictor \\( p \\) is associated with the online network so the SSL objective becomes \\( \\mathcal{L} = \\mathbb{E}_{(\\mathbf{x},\\tilde{\\mathbf{x}})}\\| p\\circ f(\\mathbf{x},\\theta) - f(\\tilde{\\mathbf{x}},\\xi)\\| _2^2 \\), where \\( \\xi^t = \\alpha \\xi^{t - 1} + (1 - \\alpha)\\theta^t \\) is target network parameter and \\( \\theta \\) is the online network parameter. Denote \\( p^\\prime = p\\circ f \\). By adding and subtracting \\( p^\\prime (\\tilde{\\mathbf{x}},\\theta) \\) we derive:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.633, + 0.47, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\| p ^ {\\prime} (\\mathbf {x}, \\theta) - p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) + p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - f (\\tilde {\\mathbf {x}}, \\xi) \\| _ {2} ^ {2} \\\\ = \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\| p ^ {\\prime} (\\mathbf {x}, \\theta) - p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) \\| _ {2} ^ {2} + \\underset {\\tilde {\\mathbf {x}}} {\\mathbb {E}} \\| p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - f (\\tilde {\\mathbf {x}}, \\xi) \\| _ {2} ^ {2} \\\\ - 2 \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\left[ \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - p ^ {\\prime} (\\mathbf {x}, \\theta)\\right) ^ {\\top} \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - f (\\tilde {\\mathbf {x}}, \\xi)\\right) \\right] \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Since \\( f(\\mathbf{x},\\theta) \\) and \\( f(\\tilde{\\mathbf{x}},\\xi) \\) maps in the same space \\( \\mathbb{R}^{d_2} \\), \\( p \\) can be considered as a bijective linear transformation within \\( \\mathbb{R}^{b_2} \\). In fact, the performance difference between a linear predictor and non-linear predictor is subtle. The global minimizer of (6) must align network output w.r.t to \\( (\\mathbf{x},\\tilde{\\mathbf{x}}) \\) with the first term in (6) and align outputs from different networks with second term in (6). Hence completes the proof." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "For Siamese network with a predictor (SimSiam), a similar approach can reformulate the objective as it is a special case when \\( f(\\tilde{\\mathbf{x}},\\xi) = f(\\tilde{\\mathbf{x}},\\theta) \\) and a stop-gradient is" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "applied on the \\(f(\\tilde{\\mathbf{x}},\\theta)\\). Hence by substituting \\(s_{\\mathcal{G}}(f(\\tilde{\\mathbf{x}},\\theta))\\) with \\(f(\\tilde{\\mathbf{x}},\\xi)\\) in (6), we complete the proof by:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.133, + 0.892, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\| p ^ {\\prime} (\\mathbf {x}, \\theta) - p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) \\| _ {2} ^ {2} \\\\ + \\underset {\\tilde {\\mathbf {x}}} {\\mathbb {E}} \\| p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - \\operatorname {s g} (f (\\tilde {\\mathbf {x}}, \\theta)) \\| _ {2} ^ {2} \\\\ - 2 \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\left[ \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - p ^ {\\prime} (\\mathbf {x}, \\theta)\\right) ^ {\\top} \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - \\operatorname {s g} (f (\\tilde {\\mathbf {x}}, \\theta))\\right) \\right] \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.245, + 0.892, + 0.291 + ], + "angle": 0, + "content": "Note that the third term in (6) and (7) can be formulated by the differential entropy of \\( H(p^{\\prime}(.)) \\) hence prevent representation collapse." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.291, + 0.892, + 0.337 + ], + "angle": 0, + "content": "For Barlow Twins, the diagonal of the cross-correlation matrix \\( C_{ii} \\) is the cosine similarity between positive pairs. Hence completes the proof by:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.35, + 0.892, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {i} ^ {d _ {2}} \\left(1 - C _ {i i}\\right) ^ {2} = \\sum_ {i} \\left(1 - \\left(f (\\mathbf {x}) ^ {\\top} f (\\tilde {\\mathbf {x}})\\right) _ {i} / \\left(d _ {2} - 1\\right)\\right) ^ {2} (8) \\\\ = \\sum_ {i} \\left(\\| f (\\mathbf {x}) _ {i} - f (\\tilde {\\mathbf {x}}) _ {i} \\| _ {2} ^ {2} / \\left(2 * \\left(d _ {2} - 1\\right)\\right) ^ {2} \\right. (9) \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.437, + 0.892, + 0.469 + ], + "angle": 0, + "content": "3.3. Transformation of the ground-truth factors is orthogonal to the applied augmentations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.893, + 0.644 + ], + "angle": 0, + "content": "By demonstrating that all discriminative SSL have a alignment loss term, the transformation between the ground truth representations and the inferred representations can be derived as in [76]. But different to [76], our derivation of minimization of cross entropy is assumed to be a lower bound since we only include the alignment term and the uniformity term is always positive [14, 5]. However, with all SSL objectives there are additional terms to maximize the output entropy of the model (some described in 3.2). So optimizing SSL objectives as a complete loss function will minimize the cross entropy \\(\\mathbb{E}[H(p(.|\\mathbf{z}),q_h(.|\\mathbf{z}))]\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.658, + 0.894, + 0.752 + ], + "angle": 0, + "content": "Theorem 3.2 By considering the generation conditional distribution as \\( p(\\tilde{\\mathbf{z}}|\\mathbf{z}) = C_p^{-1}e^{\\kappa_1\\mathbf{z}^\\top \\tilde{\\mathbf{z}}} \\), the inferred conditional distribution \\( q_h(\\tilde{\\mathbf{z}}|\\mathbf{z}) \\) can match \\( p(\\tilde{\\mathbf{z}}|\\mathbf{z}) \\) by minimizing \\( \\| h(\\mathbf{z}) - h(\\tilde{\\mathbf{z}})\\|_2^2 \\) and \\( \\forall \\mathbf{z}, \\tilde{\\mathbf{z}} : \\kappa \\mathbf{z}^\\top \\tilde{\\mathbf{z}} = h(\\mathbf{z})^\\top \\tilde{\\mathbf{z}} \\) with \\( h = f \\circ g \\) or \\( h = p \\circ f \\circ g \\) maps onto a hypersphere with radius \\( \\sqrt{\\kappa_1 / \\kappa_2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The proof exactly follows Proposition 1 in [76] just with minor modification on the concentration term \\(\\kappa\\) and \\(h\\) so that Theorem 3.2 can apply to non-InfoNCE SSL. A global minimizer of the alignment term \\(\\| h(\\mathbf{z}) - h(\\tilde{\\mathbf{z}})\\| _2^2\\) will also minimize the cross entropy of between \\(p(\\tilde{\\mathbf{z}} |\\mathbf{z})\\) in (2) and \\(q_{h}(\\tilde{\\mathbf{z}} |\\mathbf{z})\\) in (3). This indicates that minimizers of SSL objective alignment maintain the dot product. Then we can use Proposition 2 in [76] directly to show that \\(h\\) is an orthogonal linear transformation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16112" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "Theorem 3.3 Assume the data generation process (cf. 3.1), a model parameterized by \\( h = f \\circ g \\) or \\( h = p \\circ f \\circ g \\) (\\( h: \\mathbb{R}^{d_1} \\to \\mathbb{R}^{d_2} \\)) that minimizes the alignment term in all discriminative SSL objectives: \\( \\| h(\\mathbf{z}) - h(\\tilde{\\mathbf{z}}) \\|_2^2 \\) as (15), \\( h \\) is an orthogonal transformation: \\( h(\\tilde{\\mathbf{z}}) = \\mathbf{A}\\tilde{\\mathbf{z}} \\) where \\( \\mathbf{A} \\) is an orthogonal matrix." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.192, + 0.473, + 0.284 + ], + "angle": 0, + "content": "The proof follows that a function \\( h \\) minimizes the alignment will also minimize the cross entropy between the ground truth conditional distribution \\( p(\\tilde{\\mathbf{z}}|\\mathbf{z}) \\) and the inferred conditional distribution \\( q_{h}(\\tilde{\\mathbf{z}}|\\mathbf{z}) \\). Therefore if \\( h \\) is isometric w.r.t the dot product as indicated in Theorem 3.2 then \\( h(\\tilde{\\mathbf{z}}) = \\mathbf{A}\\tilde{\\mathbf{z}} \\) according to Proposition 2 in [76]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.293, + 0.473, + 0.385 + ], + "angle": 0, + "content": "Theorem 3.4 Assume augmentations applied during the training can be represented as a change in the ground truth representations, i.e. \\(\\tilde{\\mathbf{x}} = g(\\tilde{\\mathbf{z}}(c,\\tilde{V}*))\\) and change in data variables induces a shift in the ground truth representations \\(\\delta \\mathbf{z} = (c,\\tilde{V}_*) - (c,V_*)\\), then \\(\\delta \\mathbf{z}\\) is in orthogonal to \\(\\mathbf{A}\\) i.e. \\(\\mathbf{A}\\delta \\mathbf{z} = 0\\) if a discriminative SSL objective is optimized." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.473, + 0.573 + ], + "angle": 0, + "content": "Proof. With augmentations normally applied during SSL such as color distortions, rotations, random cropping, and etc., one can view the alteration as a change of a variable in the data variable (color, view angles, sizes). This change of variable under the generation framework described in 3.1 will result in a change in \\( V = \\{V_{i}\\} \\) since \\( c \\) is not changed. Regardless of structure of the causal graph shown in Figure 2, the change in \\( V \\) can be reflected in the ground truth representation space as \\( \\tilde{\\mathbf{z}} = \\mathbf{z} + \\delta \\mathbf{z} \\). A global minimizer of any discriminative SSL objective will minimize the alignment term \\( h(\\mathbf{z}) = h(\\tilde{\\mathbf{z}}) = h(\\mathbf{z} + \\delta \\mathbf{z}) \\). According to Theorem 3.3 we derive:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.584, + 0.469, + 0.6 + ], + "angle": 0, + "content": "\\[\nh (\\mathbf {z}) = \\mathbf {A} \\mathbf {z} = \\mathbf {A} (\\mathbf {z} + \\delta \\mathbf {z}) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.234, + 0.603, + 0.469, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} \\delta \\mathbf {z} = 0 \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.47, + 0.675 + ], + "angle": 0, + "content": "with \\(\\mathbf{A}\\) being a linear orthogonal matrix. This indicates that the transformation \\(A\\) is learned to annul the change in data variables or the effect of augmentations applied." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.47, + 0.713 + ], + "angle": 0, + "content": "3.4. Reason and solutions for unstable change in data variables" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In section 3.3 we show that generalized representation is robust to augmentations since \\(\\mathbf{A}\\), the linear transformation between the ground truth representation and inferred representations, is orthogonal to augmentation applied during the training. And if a change in the data variable reflects a shift, \\(\\delta \\mathbf{z}\\), in the ground truth representations and the newly inferred representation is stable, meaning \\(\\delta \\mathbf{z}\\) will be absorbed by the transformation matrix \\(\\mathbf{A}\\), then the shift in the ground truth representation corresponds to an augmentation that is applied during the training. However, when \\(\\delta \\mathbf{z}\\) appears in the range of \\(\\mathbf{A}\\) and \\(\\mathbf{A} \\delta \\mathbf{z} \\neq 0\\), the resultant inferred representation can be unstable and lead to performance drop" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "on downstream models (denote as \\(D(\\mathbf{z})\\)). We quantify this deterioration on \\(D\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.133, + 0.892, + 0.15 + ], + "angle": 0, + "content": "\\[\nm \\left(D (h (\\mathbf {z})) _ {\\text {s t a b l e}}\\right) - m \\left(D \\left(h (\\tilde {\\mathbf {z}}) _ {\\text {u n s t a b l e}}\\right) \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.161, + 0.894, + 0.268 + ], + "angle": 0, + "content": "where \\( m(.) \\) is a metric on an outcome of the downstream task. In example of \\( D \\) being a linear classifier, \\( m(.) \\) can be the probability of predicting the target class (prediction score) or the proportion of correct predictions (accuracy). In order to overcome this deterioration, we propose two methods, namely Robust Dimensions and Stable Inference Mapping:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.276, + 0.895, + 0.534 + ], + "angle": 0, + "content": "1. Robust Dimensions: Under a stable condition, \\( D(f(\\mathbf{x}))^2 = D(\\mathbf{A}\\mathbf{z}) \\), each dimension in the inferred representation \\( f(\\mathbf{x}) \\) is a linear combination of dimensions of the ground truth representation. The dimensions contributing most to \\( D() \\) should be robust to unstable shift \\( \\delta \\mathbf{z}_{\\text{unstable}} \\). In other words, most robust dimensions of \\( f(\\mathbf{x}) \\) should be also robust in \\( f(\\tilde{\\mathbf{x}}) \\) where \\( \\tilde{\\mathbf{x}} = g(\\mathbf{z} + \\delta \\mathbf{z}_{\\text{unstable}}) \\) as some dimensions of \\( \\mathbf{A} \\delta \\mathbf{z}_{\\text{unstable}} \\) will be zero. Hence identifying most important dimensions in \\( f(\\mathbf{x})_{\\text{stable}} \\) and pass through the same dimensions of \\( f(\\tilde{\\mathbf{x}})_{\\text{unstable}} \\) should alleviate the deterioration by making \\( m(D(h(\\mathbf{z}))_{\\text{stable}})_{\\{\\dim\\}} = m(D(h(\\tilde{\\mathbf{z}})_{\\text{unstable}}))_{\\{\\dim\\}} \\) where \\( \\{\\dim\\} \\) is a set of most robust dimensions. In example of \\( D \\) being a linear classifier, the contribution of each dimension can be calculated by \\( W_c^\\top f(\\mathbf{x}) \\) where \\( W_c^\\top \\) is the Jacobian of the linear classifier w.r.t target class \\( c \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.542, + 0.895, + 0.71 + ], + "angle": 0, + "content": "2. Stable Inference Mapping: Since \\( f(\\mathbf{x})_{\\text{stable}} = -f(\\tilde{\\mathbf{x}})_{\\text{unstable}} = -\\mathbf{A} \\delta \\mathbf{z} \\), we can learn another linear transformation \\( \\mathbf{F} \\) to absorb \\( \\delta \\mathbf{z} \\). Especially, we want to learn \\( \\mathbf{F} f(\\tilde{\\mathbf{z}})_{\\text{unstable}} = \\mathbf{F} \\mathbf{A} \\mathbf{z} + \\mathbf{F} \\mathbf{A} \\delta \\mathbf{z} \\) such that the additional \\( \\mathbf{F} \\) will not only set \\( \\mathbf{A} \\delta z \\) to 0, but also make stable representation more robust to the unstable shift by assuring \\( \\mathbf{F} \\mathbf{A} \\) orthogonal to \\( \\delta \\mathbf{z}_{\\text{unstable}} \\) in addition to the augmentations applied during training. Formally, we model a linear layer \\( l(f(\\tilde{\\mathbf{x}})_{\\text{unstable}}) = f(\\mathbf{x})_{\\text{stable}} - f(\\tilde{\\mathbf{x}})_{\\text{unstable}} \\), hence during the inference \\( f(\\mathbf{x}) + l(f(\\mathbf{x})) \\) is used for downstream task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.856 + ], + "angle": 0, + "content": "Relation to Causal Inference Since we only have observations in the image space, we consider the augmentations or changes of a data variable as interventions on the ground truth representations: \\(\\tilde{\\mathbf{x}} = g(\\tilde{\\mathbf{z}} |(c,do(V_i = v_i))\\). With an access to ground truth representation, we can evaluate treatment-control effect i.e. \\(Pr(f(g(\\tilde{\\mathbf{z}} |(c,do(V_i = v_{stable}))) - Pr(f(g(\\tilde{\\mathbf{z}} |(c,do(V_i = v_{unstable})))\\). Without any access to the ground truth representations, we can evaluate the average treatment effect [33] by \\(\\mathbb{E}[D(f(\\mathbf{x})_{stable})-\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.863, + 0.894, + 0.901 + ], + "angle": 0, + "content": "2Note that in general \\( f \\) contains a projector. However we exclude the notion of the projector to simplify the problem and [10, 29, 26] show that a projector is not necessary in SSL." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16113" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.122 + ], + "angle": 0, + "content": "\\(D(f(\\tilde{\\mathbf{x}})_{\\text{unstable}})]\\) via synthesizing manual data samples \\(\\tilde{\\mathbf{x}}_i = g(\\tilde{\\mathbf{z}} | (c, \\text{do} (V_i = v_i))\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.138, + 0.347, + 0.156 + ], + "angle": 0, + "content": "4. Experiments and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.165, + 0.47, + 0.224 + ], + "angle": 0, + "content": "In this section we evaluate our solutions to unstable shift in data variable on two datasets: Causal3DIdent and ImageNet pretrained SSL and corresponding linear classifiers as the downstream task." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.237, + 0.232, + 0.252 + ], + "angle": 0, + "content": "4.1. Causal3DIdent" + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.27, + 0.395, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.372, + 0.431, + 0.387 + ], + "angle": 0, + "content": "Figure 3: Causal graph for \\(V\\) in Causal3DIdent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.47, + 0.544 + ], + "angle": 0, + "content": "[76] develops the dataset that 7 objects in the dataset that each \\(224 \\times 224\\) image is associated with an object class and a 10-dimensional latent representation. These 10 dimensions correspond to 10 data variables. A causal graph imposed on these variables is shown in Figure 3. As the dependency shown in Figure 3, an object is placed at \\(pos_{obj} = (x,y,z)\\) with \\(rot_{obj} = (\\phi, \\theta, \\psi)\\) and \\(hue_{o}bj\\) under a spot light at an angle \\(pos_{spl}\\) with color \\(hue_{spl}\\) on a background with color \\(hue_{bg}\\). Detailed information in C." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.711 + ], + "angle": 0, + "content": "To simulate unstable change of variable that is not accessible during the training, a range of some data variables is hold out and hence this portion of data is treated as potential unstable samples that causes the performance drop on the linear classifier. Since all data variables are truncated in range [-1,1], we hold out the edge value(s) to further portray 'unexpected' values during the inference time. We select 8 dimensions to intervene, namely: \\( z \\) in object position3, all 3 object rotation angles, spot light position, and all 3 hue variables. Both training and testing data are sampled and for detailed sampling procedures refer to C." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.47, + 0.863 + ], + "angle": 0, + "content": "SSL Experiment Setup ResNet18[30] is the backbone of the feature encoder \\( f \\). Same augmentations in [10] are applied during training. An Adam optimizer with a learning rate at 0.0001 and a weight decay at 0.00001 is optimized for all discriminative SSL. Hyperparameters for each SSL is presented in C. The dimension of the inferred representation space is set to 128. The network is trained for 20 epochs on intervened data. Then a linear classifier is trained on the frozen representations of the network with a same optimizer for 10 epochs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.559 + ], + "angle": 0, + "content": "To illustrate the simulation results in the deterioration of the downstream performance, for each testing data that has the same data variable distribution as the training data, we search in the ground truth representation space for 5 nearest representations in the hold-out test distribution when \\( n \\) dimension of the testing data is shifted to a random value in the hold-out range of corresponding data variable. Within the corresponding 5 images representing a shift in \\( n \\) dimensions that is not seen during the training, we record the lowest performance to fulfill the deterioration scenario. In Figure 4 the score and accuracy are averaged over all combinations of \\( \\binom{8}{n} \\). When \\( n = 0 \\), this indicates the performance of SSL on the testing data with the same data variable distribution as the training data. The performance is exceptional since hold-out values of some data variables may be covered by the augmentations applied (\\( z \\) covered by random cop, and hue colors covered by color distortions). A will be orthogonal to changes in these data variables. In fact in C we show that the performance difference between seen distribution and unseen distribution in testing data is comparatively small. As illustrated in Figure 4, even when only one variable is changed to a unseen value, there is a large drop in both accuracy and prediction score (20% in accuracy and 30% in prediction score for SimCLR). With more tangled changes in data variables, the unstable representations result in poorer downstream performance. Since the selection of the data variables may be too complex due to the dependency, we also validate the same issue on selecting only children nodes in Figure 3. And also we visualize the latent shift between stable and unstable examples. See C for more results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.569, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Robust Dimensions For each pair of testing data \\(\\mathbf{x}_{\\text{stable}}\\) with seen data variable distribution and the selected data \\(\\mathbf{x}_{\\text{unstable}}\\) among the 5 nearest neighbours when changing \\(n\\) dimensions to unseen distribution values, we apply the Jacobian of the linear classifier w.r.t the target class \\(W_{c}^{\\top}\\) on the stable representation \\(f(\\mathbf{x})_{\\text{stable}}\\) to identify the top \\(k\\) most important dimensions and pass the same dimensions of \\(f(\\tilde{\\mathbf{x}})_{\\text{unstable}}\\) to \\(D\\) to evaluate the performance. As shown in Figure 5, the accuracy only deteriorate slightly when top \\(90\\%\\) most important features of \\(f(\\tilde{\\mathbf{x}})_{\\text{unstable}}\\) are selected for the downstream task. This is true even when all 8 variables are shifted. This suggests that \\(\\mathbf{A}\\) is orthogonal to changes in ground truth representation \\(\\mathbf{z}\\) in most dimensions. This high percentage of dimensions may be due to \\(f\\) optimizing the SSL objectives to a high level and the augmentations applied covers some of the hold-out variable values. Interestingly, there are cases where passing the top \\(k\\%\\) (around \\(40\\%\\)) important dimensions results in higher performance than stable representations (uninterrupted in Figure 5). However, as expected, including more less important dimensions where the unseen shift in the data variable results in non-zero adjustment (\\(\\mathbf{A} \\delta \\mathbf{z} \\neq 0\\)) initiates the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.468, + 0.899 + ], + "angle": 0, + "content": "\\( {}^{3} \\) Only \\( z \\) is altered because most of deep vision models are translation invariant." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16114" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.485, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.193, + 0.378, + 0.207 + ], + "angle": 0, + "content": "(a) Deterioration in Accuracy" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.09, + 0.877, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.193, + 0.759, + 0.206 + ], + "angle": 0, + "content": "(b) Deterioration in Score" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.219, + 0.895, + 0.249 + ], + "angle": 0, + "content": "Figure 4: Deterioration of unstable changing in data variables on all SSL. The more unexpected changes occur in the data variables, the more severe the deterioration." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.276, + 0.336, + 0.292 + ], + "angle": 0, + "content": "deterioration phenomenon we observe." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.318, + 0.487, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.451, + 0.374, + 0.463 + ], + "angle": 0, + "content": "(a) Improvement on Accuracy" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.466, + 0.485, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.596, + 0.362, + 0.609 + ], + "angle": 0, + "content": "(b) Improvement on Score" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.667 + ], + "angle": 0, + "content": "Figure 5: By identifying and passing top \\( k\\% \\) important dimensions, the deterioration in both accuracy and prediction score is significantly alleviated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.9 + ], + "angle": 0, + "content": "Stable Inference Mapping As shown in 3.4, a linear transformation \\(\\mathbf{F}:\\mathbb{R}^{d_2}\\to \\mathbb{R}^{d_2}\\) is trained to cancel the effect \\(\\mathbf{A}\\delta \\mathbf{z}\\) and also further improve the robustness of transformation on the ground truth representation \\(\\mathbf{Az}\\). In this experiment, we match each training data with a random member in the 5 nearest representations when only one dimension is changed to a unstable value. A linear layer is trained with the same optimizer for 10 epochs on the training pairs. In Table 1, all accuracy for unstable examples increases significantly (except SimSiam) after learning \\(\\mathbf{F}\\). However, since the \\(f\\) is very close to the global minimizer of the alignment term, the improvement on stable examples cannot be observed." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.274, + 0.871, + 0.386 + ], + "angle": 0, + "content": "
xstablex unstable
w/o Fw/ Fw/o Fw/ F
SimCLR0.9960.9980.8330.889
MoCo0.9920.9920.8000.849
BYOL0.9960.9980.8860.920
SimSiam0.9980.9990.9190.928
Barlow Twins0.9910.9950.8180.862
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.396, + 0.892, + 0.427 + ], + "angle": 0, + "content": "Table 1: The effect of \\(\\mathbf{F}\\) on both stable and unstable samples. The accuracy is average over 3 random seeds." + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.444, + 0.813, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.684, + 0.892, + 0.714 + ], + "angle": 0, + "content": "Figure 6: Examples of ObjectNet dataset \\((1^{st}\\) row), Stylized ImageNet dataset \\((2^{nd}\\) row), Synthetic dataset \\((3^{rd}\\) row)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.612, + 0.758 + ], + "angle": 0, + "content": "4.2. ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.841 + ], + "angle": 0, + "content": "To validate our findings on a larger scale and more realistic settings, we apply proposed solutions on ImageNet[16] with various altered and synthetic datasets as unstable shift in the data variable. See Figure 6 and D for more information on ObjectNet, Stylized-ImageNet, Synthetic dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ObjectNet[3] is a collection of objects that are intentionally placed at an unusual view angle and backgrounds, so that the bias learned by the model with usual data variables is more prominent when testing on ObjectNet. With focus-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16115" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.166 + ], + "angle": 0, + "content": "ing on 113 classes overlapping with ImageNet classes, we evaluate our second proposed solution explicitly on ObjectNet since the shift presented in ObjectNet is very unstable and it is very challenging to overcome the negative effect of the shift." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.47, + 0.244 + ], + "angle": 0, + "content": "Stylized ImageNet[24] change the style of original ImageNet image to a random artistic style. With this drastic shift in the data variable, the analysis of the first solution is more insightful on the robustness of dimensions of \\(\\mathbf{A}\\) on a very different shift in data variable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.469, + 0.398 + ], + "angle": 0, + "content": "Synthetic Data follows synthetic procedure in [18] where object is masked on a background at a location with a rotation angle. We explore the benefit of this synthesized dataset at three modes: background, location, rotation where the target variable is randomly sampled with the other two variables fixed. Additionally, we explore texture as an independent variable by masking 'texturized' objects in [23]. We set total number of updating steps per epoch as 100 with batch size 256. This means we select total number of 512000 synthesized images every epoch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.469, + 0.476 + ], + "angle": 0, + "content": "Experiment Setup ResNet 50 pretrained on ImageNet and a linear classifier fintuned via SimCLR, BYOL, and SimSiam are tested with both proposed solutions. For Stable Inference Mapping, a linear layer is optimized with an Adam optimizer for 10 epochs on the synthetic dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.469, + 0.657 + ], + "angle": 0, + "content": "Robust Dimensions For each of ImageNet validation data sample, we stylized the image to a random artistic fashion. We observe the dramatic performance difference between the ImageNet stable images and Stylized unstable ImageNet images \\((\\mathbf{x}_{stable}, \\mathbf{x}_{unstable})\\). The result is shown in Figure 7. For SimCLR, passing the top \\(10\\%\\) important dimensions can close a small performance gap between stable representations and unstable representations. Nonetheless, all SSL seem to be sensible to the strong style change as they shorten the difference between ImageNet to an modest extent. This is as expected since the Stylized ImageNet changes multiple variables to an extreme value." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Stable Inference Mapping As described in Synthetic Data, we explore the benefit of our second proposed method to evaluate on a dataset that most of samples are unstable according to learned \\(\\mathbf{A}\\). At each training step, with other variables randomly fixed, 10 images with random target variable values are generated. The pair with maximum \\(m(D(\\mathbf{x})) - m(D(\\tilde{\\mathbf{x}}))\\) is selected to train the linear transformation. In Table 2, inferring \\(\\mathbf{F}\\) via controlling location produces least improvement. This is expected since the network is robust to translation by design. While background, rotation, and texture improves the performance considerably with the consideration on the training time. However, in D we show that training the model longer using Stable Inference Mapping yield less favorable results since the improvement is less significant and starts to saturate at around 30 epochs." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.892, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.21, + 0.788, + 0.224 + ], + "angle": 0, + "content": "(a) Improvement on Accuracy" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.226, + 0.892, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.345, + 0.776, + 0.358 + ], + "angle": 0, + "content": "(b) Improvement on Score" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.371, + 0.892, + 0.415 + ], + "angle": 0, + "content": "Figure 7: By identifying and passing top \\( k\\% \\) important dimensions, the deterioration in both accuracy and prediction score is alleviated." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.434, + 0.858, + 0.531 + ], + "angle": 0, + "content": "
SimCLRBYOLSimSiam
w/o F10.1214.0411.15
background12.3415.4312.45
location10.3514.1311.17
rotation12.0415.8812.38
texture12.7915.0613.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.542, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Table 2: Comparison of inferring different \\( \\mathbf{F} \\) on ObjectNet with a target data variable random trials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.623, + 0.618 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Though we identify the root cause of unstable inference for discriminative SSL by constructing a causal framework inspired by the prior work, the proposed solutions are constrained and limited to be applied on realistic applications. Robust Dimensions involves establishing a correspondence between stable and unstable instances on a one-to-one basis, enabling the identification of dimensions contributing to stability. On the other hand, Stable Inference Mapping necessitates a collection of unstable instances with a specific alteration in a particular group of data variables. Within the Causal3dIdent dataset, both solutions can be assessed using the same unstable instances. In more realistic datasets, achieving a one-to-one correspondence is feasible, and manipulation of one group of data variables can be accomplished using synthetic data. However, any assessments with involving artificially generated images might introduce some level of uncertainty. In a realistic setup, since training samples are not directly observ-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16116" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "able during the inference stage, simple interventions on inference samples may not effectively separate the unstable variables from the stable ones. Consequently, the potential benefits of the proposed solutions in realistic datasets are undermined." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.182, + 0.205, + 0.198 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.207, + 0.473, + 0.466 + ], + "angle": 0, + "content": "In conclusion, this paper has proposed a novel approach to address the issue of unstable behavior during the inference stage in SSL methods. By building on the previous theories of successful InfoNCE-facilitated contrastive SSL and extending it to recent SSL methods, we have demonstrated that a change in the data factor can result in a shift in the inferred representation, leading to a decline in downstream performance. We have proposed learning targeted transformations that regularize the violating shift and restore performance on the unseen data shift. Our experiments on both controlled and realistic datasets have shown the efficacy of our proposed solutions. These contributions provide a better understanding of SSL methods and offer a promising solution to the problem of unstable behavior during the inference stage. We hope that our work will inspire further research in this area and lead to improved SSL methods that are more robust to changes in data factors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.479, + 0.266, + 0.496 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.504, + 0.47, + 0.534 + ], + "angle": 0, + "content": "The work was funded in part by NSF CNS-2112562 and IIS-2140247." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.549, + 0.175, + 0.565 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.629 + ], + "angle": 0, + "content": "[1] Ibrahim Ahmad and Pi-Erh Lin. A nonparametric estimation of the entropy for absolutely continuous distributions (corresp.). IEEE Transactions on Information Theory, 22(3):372-375, 1976. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.688, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[3] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.773, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[4] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.83, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[5] Sergiy V Borodachov, Douglas P Hardin, and Edward B Saff. Discrete energy on rectifiable sets, volume 3. Springer, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[6] Yue Cao, Mingsheng Long, and Jianmin Wang. Unsupervised domain adaptation with distribution matching ma" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "chines. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.121, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.19, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 5, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[13] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[14] Henry Cohn and Abhinav Kumar. Universally optimal distribution of points on spheres. Journal of the American Mathematical Society, 20(1):99-148, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[15] Hal Daumé III. Frustratingly easy domain adaptation. arXiv preprint arXiv:0907.1815, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[17] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[18] Josip Djolonga, Jessica Yung, Michael Tschannen, Rob Romijnders, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Matthias Minderer, Alexander D'Amour, Dan Moldovan, Sylvain Gelly, Neil Houlsby, Xiaohua Zhai, and Mario Lucic. On robustness and transferability of convolutional neural networks, 2020. 8, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[19] Lixin Duan, Dong Xu, and Ivor Tsang. Learning with augmented features for heterogeneous domain adaptation. arXiv preprint arXiv:1206.4660, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[20] Deniz Erdogmus, Kenneth E Hild, Yadunandana N Rao, and Jose C Principe. Minimax mutual information approach for independent component analysis. Neural Computation, 16(6):1235-1252, 2004. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16117" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[21] Ronald Aylmer Fisher. Dispersion on a sphere. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 217(1130):295-305, 1953. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[22] Bernhard C. Geiger and Gernot Kubin. On the information loss in memoryless systems: The multivariate case. CoRR, abs/1109.4856, 2011. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[23] Robert Geirhos, Claudio Michaelis, Felix A. Wichmann, Patricia Rubisch, Matthias Bethge, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. 7th International Conference on Learning Representations, ICLR 2019, (c):1-22, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[24] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[25] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[26] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[27] Jeff Z. HaoChen, Colin Wei, Adrien Gaidon, and Tengyu Ma. Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Loss. Advances in Neural Information Processing Systems, 7(NeurIPS):5000-5011, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[31] Geoffrey E Hinton and Richard Zemel. Autoencoders, minimum description length and helmholtz free energy. Advances in neural information processing systems, 6, 1993. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[32] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[33] Paul W Holland. Statistics and causal inference. Journal of the American statistical Association, 81(396):945-960, 1986. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[34] Tzu Ming Harry Hsu, Wei Yu Chen, Cheng-An Hou, Yao-Hung Hubert Tsai, Yi-Ren Yeh, and Yu-Chiang Frank Wang." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Unsupervised domain adaptation with imbalanced cross-domain data. In Proceedings of the IEEE International Conference on Computer Vision, pages 4121-4129, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[35] Aapo Hyvarinen and Patrik Hoyer. Emergence of phase-and shift-invariant features by decomposition of natural images into independent feature subspaces. Neural computation, 12(7):1705-1720, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[36] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and Nonlinear ICA. Advances in Neural Information Processing Systems, (Nips):3772-3780, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[37] Aapo Hyvarinen and Hiroshi Morioka. Nonlinear ICA of temporally dependent stationary sources. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[38] Aapo Hyvarinen and Petteri Pajunen. Nonlinear independent component analysis: Existence and uniqueness results. Neural networks, 12(3):429-439, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[39] Aapo Hyvarinen, Hiroaki Sasaki, and Richard E. Turner. Nonlinear ICA using auxiliary variables and generalized contrastive learning. AISTATS 2019 - 22nd International Conference on Artificial Intelligence and Statistics, 89, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[40] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1):2, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[41] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[42] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4893-4902, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[43] Ilyes Khemakhem, Diederik P. Kingma, Ricardo Pio Monti, and Aapo Hyvarinen. Variational Autoencoders and Nonlinear ICA: A Unifying Framework. 108(2015), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.649, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[44] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[45] David Klindt, Lukas Schott, Yash Sharma, Ivan Ustyuzhaninov, Wieland Brendel, Matthias Bethge, and Dylan Paiton. Towards Nonlinear Disentanglement in Natural Data with Temporal Sparse Coding. pages 1-51, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[46] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[47] Quoc V Le, Will Y Zou, Serena Y Yeung, and Andrew Y Ng. Learning hierarchical invariant spatio-temporal features" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16118" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "for action recognition with independent subspace analysis. In CVPR 2011, pages 3361-3368. IEEE, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[48] Jason D. Lee, Qi Lei, Nikunj Saunshi, and Jiacheng Zhuo. Predicting What You Already Know Helps: Provable Self-Supervised Learning. Advances in Neural Information Processing Systems, 1(NeurIPS):309-323, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[49] Chunyuan Li, Jianwei Yang, Pengchuan Zhang, Mei Gao, Bin Xiao, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Efficient Self-supervised Vision Transformers for Representation Learning. 3:1-27, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[50] Wen Li, Lixin Duan, Dong Xu, and Ivor W Tsang. Learning with augmented features for supervised and semi-supervised heterogeneous domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 36(6):1134-1148, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[51] Ralph Linsker. Self-organization in a perceptual network. Computer, 21(3):105–117, 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[52] Shuo Liu, Adria Mallol-Ragolta, Emilia Parada-Cabaleiro, Kun Qian, Xin Jing, Alexander Kathan, Bin Hu, and Bjöern W Schuller. Audio self-supervised learning: A survey. Patterns, 3(12):100616, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[53] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering, 35(1):857-876, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[54] Yixin Liu, Ming Jin, Shirui Pan, Chuan Zhou, Yu Zheng, Feng Xia, and Philip Yu. Graph self-supervised learning: A survey. IEEE Transactions on Knowledge and Data Engineering, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[55] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer joint matching for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1410-1417, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[56] Jovana Mitrovic, Brian McWilliams, Jacob Walker, Lars Buesing, and Charles Blundell. Representation Learning via Invariant Causal Mechanisms. pages 1-21, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[57] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI, pages 69-84. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[58] Aaron Van Den Oord. Representation Learning with Contrastive Predictive Coding. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[59] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[60] Judea Pearl. Causality: Models, reasoning, and inference, second edition. 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[61] Madeline C Schiappa, Yogesh S Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ACM Computing Surveys, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[62] Saeed Shurrab and Rehab Duwairi. Self-supervised learning methods and applications in medical imaging analysis: A survey. PeerJ Computer Science, 8:e1045, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[63] Mamatha Thota and Georgios Leontidis. Contrastive domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2209-2218, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[64] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[65] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[66] Riccardo Volpi, Pietro Morerio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5495-5504, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[67] Kügelgen Von, Julius and, Wieland Brendel, Bernhard Schölkopf, Michel Besserve, and Francesco Locatello. Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style Luigi Gresele. Yash Sharma, 3(NeurIPS):4, 2021. 2, 3, 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[68] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[69] Rui Wang, Zuxuan Wu, Zejia Weng, Jingjing Chen, Guo-Jun Qi, and Yu-Gang Jiang. Cross-domain contrastive learning for unsupervised domain adaptation. IEEE Transactions on Multimedia, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[70] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. 37th International Conference on Machine Learning, ICML 2020, PartF168147-13:9871-9881, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[71] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[72] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-Supervised Learning with Swin Transformers. pages 1-8, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.691, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[73] Detai Xin, Tatsuya Komatsu, Shinnosuke Takamichi, and Hiroshi Saruwatari. Disentangled speaker and language representations using mutual information minimization and domain adaptation for cross-lingual tts. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6608-6612. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[74] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 1, 2, 3, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[75] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, Octo" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16119" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "ber 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[76] Roland S. Zimmermann, Yash Sharma, Steffen Schneider, Matthias Bethge, and Wieland Brendel. Contrastive Learning Inverts the Data Generating Process. 2021. 2, 3, 4, 5, 6, 15" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16120" + } + ] +] \ No newline at end of file diff --git a/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/fdb72a4a-94ed-4829-ab6f-8cf80d64f7f5_origin.pdf b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/fdb72a4a-94ed-4829-ab6f-8cf80d64f7f5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..08fb75e247e015c85bf9d9fb5549249731ac203c --- /dev/null +++ b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/fdb72a4a-94ed-4829-ab6f-8cf80d64f7f5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0660734f0202b91fe3bd389a96c8ba4bf3ab4588bec1a01c7e4c8c9aba6e1d1 +size 2214428 diff --git a/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/full.md b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bd10563150d6d537f1c66a326378268cc807bad3 --- /dev/null +++ b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/full.md @@ -0,0 +1,353 @@ +# Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations + +Yuewei Yang, Hai Li, Yiran Chen + +Duke University + +Durham, USA + +yuwei.yang@duke.edu, hai.li@duke.edu, yiran.chen@duke.edu + +# Abstract + +In recent years, discriminative self-supervised methods have made significant strides in advancing various visual tasks. The central idea of learning a data encoder that is robust to data distortions/augmentations is straightforward yet highly effective. Although many studies have demonstrated the empirical success of various learning methods, the resulting learned representations can exhibit instability and hinder downstream performance. In this study, we analyze discriminative self-supervised methods from a causal perspective to explain these unstable behaviors and propose solutions to overcome them. Our approach draws inspiration from prior works that empirically demonstrate the ability of discriminative self-supervised methods to demix ground truth causal sources to some extent. Unlike previous work on causality-empowered representation learning, we do not apply our solutions during the training process but rather during the inference process to improve time efficiency. Through experiments on both controlled image datasets and realistic image datasets, we show that our proposed solutions, which involve tempering a linear transformation with controlled synthetic data, are effective in addressing these issues. + +# 1. Introduction + +Learning generalized representation with unlabeled data is a challenging task in various fields, but Self-Supervised Learning (SSL) has recently demonstrated remarkable success in learning semantic invariant representations without labels [40, 41, 53]. There are two main types of self-supervised learning (SSL) based on the pretext task used: generative and discriminative SSL, with generative SSL reconstructing altered or distorted data to its original input [9, 28, 31, 59, 65, 71] and early discriminative SSL predicting easily designed labels and task-specific representations that are not very generalizable [25, 57, 75]. More recent + +![](images/210d2f49e331f76ac6e8ec4cb2e6ed001d0b54603621d409dbce501dd562d9ee.jpg) + +![](images/77582930b2b339d2c9e59af355f71daa3dbb052ba51491edb704d4a8f5985767.jpg) +Figure 1: During the training of discriminative SSL, aligning positive representations will be robust to the changes applied as augmentations (red arrows). However, during the inference stage, one small change in the data variable (such as view angle) will result in an unexpected degradation on downstream performance. + +discriminative SSL trains the model to identify similarities and differences between pairs of augmented examples [7, 10, 11, 26, 29, 74]. The success of SSL in deep image models has resulted in progress in other data modalities [53, 52, 54, 61, 62] and attention-based models like transformers [12, 8, 49, 72]. Recent discriminative SSL aims to learn content and semantic invariant representations that are robust to data augmentations, but the learned representations can be unstable when one subtle factor of the data is changed to a value that is not accessible through all augmentations. To avoid the high cost of incorporating all possible subtle changes during training, insights are needed to uncover the root cause of instability and find a solution to prevent performance deterioration during inference. Figure 1 summarizes this deterioration effect. + +Causality [60] is a vital tool to investigate the causal + +relationships between variables in observational data, and can uncover the underlying causal factors that explain unexpected model behavior due to changes in the environment. Independent Component Analysis (ICA) is often used to disentangle sources in unsupervised training [36, 37, 39, 43, 45], and causal analysis has been applied in follow-up works [76, 67] to examine the empirical success of SSL under an ICA framework. However, while these works identify factors that contribute to SSL's success, they do not address the problem's unstable mode, which can cause a severe performance drop when underlying factors shift slightly to an unseen environment. Some works [56, 27, 48] attempt to incorporate causality during the training process to identify and alleviate the impact of such shifts, but this approach is time-costly and only marginally improves performance compared to non-causal SSL methods. A more time-efficient and accessible solution would be to simply reverse the unstable shift during inference. + +We aim to address the issue of unstable behavior during the inference stage by building upon previous theories of successful InfoNCE-facilitated contrastive SSL and extending it to all recent SSL methods with additional assumptions and constraints. Drawing inspiration from the relationship between the ground truth positive pairs distribution and learned positive pairs distribution, we demonstrate that the approximated transformation between the ground truth representations and learned representations is orthogonal to the augmentations applied during training. However, a change in the data factor/variable that violates the conditions for successful SSL can cause a corresponding shift in the inferred representation, resulting in a decline in downstream performance. This change of data factor/variable can be a change in the background, texture, or view angles etc. To overcome this issue, we propose learning targeted transformations that regularize the violating shift and restore performance on the unseen data shift. This approach effectively avoids the undesirable behavior and improves performance on previously unseen data shifts. + +To summarize, our contributions are following: + +- Through the use of a comparable data generation process in prior research, we show that All current SSL techniques benefit from the alignment of positive pairs. +- Through our alternative derivation of the transformation matrix between the ground truth representation and the learned representation, we have shown that the augmentations applied during training are orthogonal to the resulting matrix. +- By interpreting a change in the data variable causally, we propose two solutions focusing on inference to modify the negative shift in representation space caused by such a change. + +- We validate the proposed solutions by conducting experiments on both controlled and realistic datasets, providing evidence for their efficacy during the inference stage without retraining the pretrained models. + +# 2. Related Work + +Discriminative SSL learns invariant representations from positive pairs of unlabeled data samples, but previous attempts to use trivial labels like colors[75], rotations[25], and patch positions[57] offer minimal benefits for complex downstream tasks due to their easy augmentations. Recent discriminative SSL apply random augmentations to generate two views of an image sample and train an encoder to extract representations for maximizing similarity between the paired augmentations. SimCLR and MoCo[10, 29] are pioneer SSL works that maximize the cosine similarity between positive pairs and minimize the cosine similarity between negative pairs via optimizing the InfoNCE loss[58]. Immense resources are used to enforce a large number of negative samples since a larger number can tighten the upper bound of the mutual information between positive pairs[58]. Later advancement of discriminative SSL excludes the notion of negative pairs by only aligning positive pairs and preventing representation collapse through various regularizations. BYOL[26] predicts an Exponential-Moving-Averaged (EMA) representation of one view with a projected representation of another view. SimSiam[11] maximizes similarity between a projected representation and a detached representation of two positive samples. Unlike previous work focusing optimization on an instance level, Barlow Twins[74] encourages high similarity in corresponding feature dimensions and discourages redundancy across different feature dimensions between two views of a data sample. Detailed formulations are exhibited in A. + +Mutual Information is a different perspective on the behaviour of discriminative SSL. Referred to InfoMax principle[51], the MI between different transformations of a data sample is maximized via optimizing the InfoNCE loss[2, 32, 44]. Though showing theoretical relation between optimizing InfoNCE and maximizing MI between positive pairs, the underlying factors instructing the behaviour of different SSL methods are not explored. Nonlinear ICA[38], on the other hand, captures complex data structures of SSL methods by disentangling underlying factors via minimizing the mutual information between learned representations and the original data input[17, 73, 20]. Other works associate the nonlinear ICA objectives with the contrastive SSL so the MI between positive pairs are maximized and negative pairs are minimized. + +Other researchers have explored causality and causal inference as a means of understanding the success of discriminative SSL. Prior work has focused on partitioning the InfoNCE loss to alignment between the positive pairs and + +uniformity between aggregations of all positive clusters[70]. By formulating a data generation process, [76] empirically explains that networks optimized via InfoNCE infer an orthogonal transformation of the ground truth latent representations. Furthermore [67] validates that that augmentations used in both generative and discriminative SSL isolate the content factor from the style factor. Our theory and work draw great inspirations from these two work. However, instead of solely focusing on InfoNCE-driven SSL and two factors, this work extends the framework to all recent discriminative SSL methods and identifies reasons for unstable circumstances analytically. We also propose methods to nullify the negative effects of unstable representations during inference. + +Domain Adaptation is a strategy to bridge the gap between the model performance on a source domain and that on a target domain [4, 68]. Feature adaptation methods try to learn a new feature representation that is more invariant to the domain shift[66, 19, 15, 50], while instance adaptation methods try to reweight the importance of the labeled source examples to better align with the target domain[55, 6, 34]. Recent studies also implement a contrastive framework to learn a shared latent space between the source and target domains by maximizing the agreement between representations of corresponding samples while minimizing the agreement between representations of noncorresponding samples [69, 63, 42]. Unlike these works focusing on adapting to a target domain for better performance, we investigate the underlying causal factors for the performance gap and based on the findings we propose easy solutions to connect the gap. + +# 3. Theory + +In this section, we build on previous data generation process [76, 67] (3.1) and show that current SSL techniques benefit from the alignment of positive pairs (3.2). During a deeper dive into the generation process, we show the relation between the ground truth representations and the inferred representations, and this relation is an linear transformation matrix that is orthogonal to the augmentations applied during training (3.3). Finally, we disclose the causal reason for the unstable behaviour caused by a change in the data variable during the inference stage and propose analytical solutions to address this unstable issue (3.4). + +# 3.1. Problem formulation + +Data generation is assumed to be a generation function that takes ground truth latent representation as the input to generate an observation data/image. + +Formally, we assume that the marginal distribution of sampling ground truth representations $\mathbf{z} \subseteq \mathcal{Z} \in \mathbb{R}^{d_1}$ w.r.t. a class is uniform on a unit sphere $\mathbb{S}^{d_1 - 1}$ . An injective generation function $g(.) : \mathbb{R}^{d_1} \to \mathbb{R}^d$ takes a ground truth rep- + +![](images/99cf6be555efe2964c0444eec817c0c6c366220f40744e28d9f136218478a52f.jpg) +Figure 2: A causal graph for data generation process. + +resentation and generate an observation sample $\mathbf{x} \in \mathbb{R}^d$ , $\mathbf{x} = g(\mathbf{z})$ . Different variables, denoted as a set $V = \{V_i\}$ and class/content variable $c$ , constitute the values of each dimension in the ground truth representation. These variables can be direct causes or confounding factors, $\mathbf{z} = (c, \{V_i\})$ . Specific examples include view angles, object size, background colors, etc. A simple causal graph depicts this relationship is shown in Figure 2. So the general generation process is as following: + +$$ +p (\mathbf {z} = (c, V _ {*})) \sim \frac {1}{| \mathcal {Z} |} \quad \mathbf {x} = g (\mathbf {z}) \tag {1} +$$ + +To sample a positive example w.r.t the same class, we assume the conditional distribution is a von Mises-Fisher (vMF) distribution [21]: + +$$ +p (\tilde {\mathbf {z}} = (c, \tilde {V} _ {*}) | \mathbf {z} = (c, V _ {*})) = C _ {p} ^ {- 1} e ^ {\kappa_ {1} \mathbf {z} ^ {\top} \tilde {\mathbf {z}}} \qquad (2) +$$ + +where $C_p = \int e^{\kappa_1\mathbf{z}^\top \tilde{\mathbf{z}}}d\tilde{\mathbf{z}}$ is a normalization constant and $\kappa$ is a concentration parameter. + +Representation learning is a process that a feature encoder, $f(\cdot):\mathbb{R}^d\to \mathbb{R}^{d_2}$ , extracts representations from two positive observations $f(\tilde{\mathbf{x}}) = f\circ g(\tilde{\mathbf{z}})$ , $f(\mathbf{x}) = f\circ g(\mathbf{z})$ and a distribution associated with the encoder $f$ through $h = f\circ g$ is: + +$$ +q _ {h} (\tilde {\mathbf {z}} | \mathbf {z}) = C _ {q} ^ {- 1} (\mathbf {z}) e ^ {\kappa_ {2} h (\tilde {\mathbf {z}}) ^ {\top} h (\mathbf {z})} \tag {3} +$$ + +with $C_q(\mathbf{z}) = \int e^{\kappa_2 h(\tilde{\mathbf{z}})^\top h(\mathbf{z})} d\tilde{\mathbf{z}}$ be the normalization term. Optimizing any discriminative SSL objective will maximize the similarity between these positive pairs. An example of a well constructed objective is the InfoNCE loss: + +$$ +\mathcal {L} _ {i n f o N C E} = \mathbb {E} \left[ - l o g \frac {e ^ {f (\tilde {\mathbf {x}}) ^ {\top} f (\mathbf {x}) / \tau}}{e ^ {f (\tilde {\mathbf {x}}) ^ {\top} f (\mathbf {x}) / \tau} + \sum_ {i = 1} ^ {K} e ^ {f \left(\mathbf {x} _ {i} ^ {-}\right) ^ {\top} f (\mathbf {x}) / \tau}} \right] \tag {4} +$$ + +where $\tilde{\mathbf{x}}$ is a positive example w.r.t. $\mathbf{x}$ in the observation space and $\{\mathbf{x}_i^-\}_{1}^{K}$ are $K$ samples from distributions of all observations. The global minimum of (4) is reached when + +the cosine similarity between positive pairs is maximized and the cosine similarity between all negative pairs is minimized. In the following section, we will show that discriminative SSL including InfoNCE-driven and non InfoNCE-driven (EMA and Siamese with a predictor) follows strict rules of alignment to maximize the similarity. + +# 3.2. Alignment in discriminative SSL + +In this section we will combine theories stated in [76, 67, 70] so that the general factors of successful discriminative SSL can be summarized. Additionally, instead of focusing on just InfoNCE SSL variation and content block-identifiability [35, 47], we extend the combined theory to demonstrate that all discriminative SSL benefit from alignment between positive representations. + +Theorem 3.1 With a data generation process described in 3.1, all discriminative SSL objectives have an alignment loss function between positive pairs from the network: + +$$ +\mathcal {L} _ {\text {a l i g n}} = \left\| f (\mathbf {x}) - f (\tilde {\mathbf {x}}) \right\| _ {2} ^ {2} \tag {5} +$$ + +This is a weaker statement than Theorem 4.4 in [67] since we only focus on the alignment term. For analysis on the regularization term on the network entropy, see B. + +Proof. For InfoNCE-driven SSL (SimCLR and MoCo), as derived in [68, 76], the InfoNCE loss converges to an alignment term and a uniformity term as the number of negative samples approaches infinity. (See B for full details). + +For EMA-based SSL methods (BYOL), a predictor $p$ is associated with the online network so the SSL objective becomes $\mathcal{L} = \mathbb{E}_{(\mathbf{x},\tilde{\mathbf{x}})}\| p\circ f(\mathbf{x},\theta) - f(\tilde{\mathbf{x}},\xi)\| _2^2$ , where $\xi^t = \alpha \xi^{t - 1} + (1 - \alpha)\theta^t$ is target network parameter and $\theta$ is the online network parameter. Denote $p^\prime = p\circ f$ . By adding and subtracting $p^\prime (\tilde{\mathbf{x}},\theta)$ we derive: + +$$ +\begin{array}{l} \mathcal {L} = \underset {(\mathbf {x}, \tilde {\mathbf {x}})} {\mathbb {E}} \| p ^ {\prime} (\mathbf {x}, \theta) - p ^ {\prime} (\tilde {\mathbf {x}}, \theta) + p ^ {\prime} (\tilde {\mathbf {x}}, \theta) - f (\tilde {\mathbf {x}}, \xi) \| _ {2} ^ {2} \\ = \underset {(\mathbf {x}, \tilde {\mathbf {x}})} {\mathbb {E}} \| p ^ {\prime} (\mathbf {x}, \theta) - p ^ {\prime} (\tilde {\mathbf {x}}, \theta) \| _ {2} ^ {2} + \underset {\tilde {\mathbf {x}}} {\mathbb {E}} \| p ^ {\prime} (\tilde {\mathbf {x}}, \theta) - f (\tilde {\mathbf {x}}, \xi) \| _ {2} ^ {2} \\ - 2 \underset {(\mathbf {x}, \tilde {\mathbf {x}})} {\mathbb {E}} \left[ \left(p ^ {\prime} (\tilde {\mathbf {x}}, \theta) - p ^ {\prime} (\mathbf {x}, \theta)\right) ^ {\top} \left(p ^ {\prime} (\tilde {\mathbf {x}}, \theta) - f (\tilde {\mathbf {x}}, \xi)\right) \right] \tag {6} \\ \end{array} +$$ + +Since $f(\mathbf{x},\theta)$ and $f(\tilde{\mathbf{x}},\xi)$ maps in the same space $\mathbb{R}^{d_2}$ , $p$ can be considered as a bijective linear transformation within $\mathbb{R}^{b_2}$ . In fact, the performance difference between a linear predictor and non-linear predictor is subtle. The global minimizer of (6) must align network output w.r.t to $(\mathbf{x},\tilde{\mathbf{x}})$ with the first term in (6) and align outputs from different networks with second term in (6). Hence completes the proof. + +For Siamese network with a predictor (SimSiam), a similar approach can reformulate the objective as it is a special case when $f(\tilde{\mathbf{x}},\xi) = f(\tilde{\mathbf{x}},\theta)$ and a stop-gradient is + +applied on the $f(\tilde{\mathbf{x}},\theta)$ . Hence by substituting $s_{\mathcal{G}}(f(\tilde{\mathbf{x}},\theta))$ with $f(\tilde{\mathbf{x}},\xi)$ in (6), we complete the proof by: + +$$ +\begin{array}{l} \mathcal {L} = \underset {(\mathbf {x}, \tilde {\mathbf {x}})} {\mathbb {E}} \| p ^ {\prime} (\mathbf {x}, \theta) - p ^ {\prime} (\tilde {\mathbf {x}}, \theta) \| _ {2} ^ {2} \\ + \underset {\tilde {\mathbf {x}}} {\mathbb {E}} \| p ^ {\prime} (\tilde {\mathbf {x}}, \theta) - \operatorname {s g} (f (\tilde {\mathbf {x}}, \theta)) \| _ {2} ^ {2} \\ - 2 \underset {(\mathbf {x}, \tilde {\mathbf {x}})} {\mathbb {E}} \left[ \left(p ^ {\prime} (\tilde {\mathbf {x}}, \theta) - p ^ {\prime} (\mathbf {x}, \theta)\right) ^ {\top} \left(p ^ {\prime} (\tilde {\mathbf {x}}, \theta) - \operatorname {s g} (f (\tilde {\mathbf {x}}, \theta))\right) \right] \tag {7} \\ \end{array} +$$ + +Note that the third term in (6) and (7) can be formulated by the differential entropy of $H(p^{\prime}(.))$ hence prevent representation collapse. + +For Barlow Twins, the diagonal of the cross-correlation matrix $C_{ii}$ is the cosine similarity between positive pairs. Hence completes the proof by: + +$$ +\begin{array}{l} \sum_ {i} ^ {d _ {2}} \left(1 - C _ {i i}\right) ^ {2} = \sum_ {i} \left(1 - \left(f (\mathbf {x}) ^ {\top} f (\tilde {\mathbf {x}})\right) _ {i} / \left(d _ {2} - 1\right)\right) ^ {2} (8) \\ = \sum_ {i} \left(\| f (\mathbf {x}) _ {i} - f (\tilde {\mathbf {x}}) _ {i} \| _ {2} ^ {2} / \left(2 * \left(d _ {2} - 1\right)\right) ^ {2} \right. (9) \\ \end{array} +$$ + +# 3.3. Transformation of the ground-truth factors is orthogonal to the applied augmentations + +By demonstrating that all discriminative SSL have a alignment loss term, the transformation between the ground truth representations and the inferred representations can be derived as in [76]. But different to [76], our derivation of minimization of cross entropy is assumed to be a lower bound since we only include the alignment term and the uniformity term is always positive [14, 5]. However, with all SSL objectives there are additional terms to maximize the output entropy of the model (some described in 3.2). So optimizing SSL objectives as a complete loss function will minimize the cross entropy $\mathbb{E}[H(p(.|\mathbf{z}),q_h(.|\mathbf{z}))]$ + +Theorem 3.2 By considering the generation conditional distribution as $p(\tilde{\mathbf{z}}|\mathbf{z}) = C_p^{-1}e^{\kappa_1\mathbf{z}^\top \tilde{\mathbf{z}}}$ , the inferred conditional distribution $q_h(\tilde{\mathbf{z}}|\mathbf{z})$ can match $p(\tilde{\mathbf{z}}|\mathbf{z})$ by minimizing $\| h(\mathbf{z}) - h(\tilde{\mathbf{z}})\|_2^2$ and $\forall \mathbf{z}, \tilde{\mathbf{z}} : \kappa \mathbf{z}^\top \tilde{\mathbf{z}} = h(\mathbf{z})^\top \tilde{\mathbf{z}}$ with $h = f \circ g$ or $h = p \circ f \circ g$ maps onto a hypersphere with radius $\sqrt{\kappa_1 / \kappa_2}$ . + +The proof exactly follows Proposition 1 in [76] just with minor modification on the concentration term $\kappa$ and $h$ so that Theorem 3.2 can apply to non-InfoNCE SSL. A global minimizer of the alignment term $\| h(\mathbf{z}) - h(\tilde{\mathbf{z}})\| _2^2$ will also minimize the cross entropy of between $p(\tilde{\mathbf{z}} |\mathbf{z})$ in (2) and $q_{h}(\tilde{\mathbf{z}} |\mathbf{z})$ in (3). This indicates that minimizers of SSL objective alignment maintain the dot product. Then we can use Proposition 2 in [76] directly to show that $h$ is an orthogonal linear transformation. + +Theorem 3.3 Assume the data generation process (cf. 3.1), a model parameterized by $h = f \circ g$ or $h = p \circ f \circ g$ ( $h: \mathbb{R}^{d_1} \to \mathbb{R}^{d_2}$ ) that minimizes the alignment term in all discriminative SSL objectives: $\| h(\mathbf{z}) - h(\tilde{\mathbf{z}}) \|_2^2$ as (15), $h$ is an orthogonal transformation: $h(\tilde{\mathbf{z}}) = \mathbf{A}\tilde{\mathbf{z}}$ where $\mathbf{A}$ is an orthogonal matrix. + +The proof follows that a function $h$ minimizes the alignment will also minimize the cross entropy between the ground truth conditional distribution $p(\tilde{\mathbf{z}}|\mathbf{z})$ and the inferred conditional distribution $q_{h}(\tilde{\mathbf{z}}|\mathbf{z})$ . Therefore if $h$ is isometric w.r.t the dot product as indicated in Theorem 3.2 then $h(\tilde{\mathbf{z}}) = \mathbf{A}\tilde{\mathbf{z}}$ according to Proposition 2 in [76]. + +Theorem 3.4 Assume augmentations applied during the training can be represented as a change in the ground truth representations, i.e. $\tilde{\mathbf{x}} = g(\tilde{\mathbf{z}}(c,\tilde{V}*))$ and change in data variables induces a shift in the ground truth representations $\delta \mathbf{z} = (c,\tilde{V}_*) - (c,V_*)$ , then $\delta \mathbf{z}$ is in orthogonal to $\mathbf{A}$ i.e. $\mathbf{A}\delta \mathbf{z} = 0$ if a discriminative SSL objective is optimized. + +Proof. With augmentations normally applied during SSL such as color distortions, rotations, random cropping, and etc., one can view the alteration as a change of a variable in the data variable (color, view angles, sizes). This change of variable under the generation framework described in 3.1 will result in a change in $V = \{V_{i}\}$ since $c$ is not changed. Regardless of structure of the causal graph shown in Figure 2, the change in $V$ can be reflected in the ground truth representation space as $\tilde{\mathbf{z}} = \mathbf{z} + \delta \mathbf{z}$ . A global minimizer of any discriminative SSL objective will minimize the alignment term $h(\mathbf{z}) = h(\tilde{\mathbf{z}}) = h(\mathbf{z} + \delta \mathbf{z})$ . According to Theorem 3.3 we derive: + +$$ +h (\mathbf {z}) = \mathbf {A} \mathbf {z} = \mathbf {A} (\mathbf {z} + \delta \mathbf {z}) \tag {10} +$$ + +$$ +\mathbf {A} \delta \mathbf {z} = 0 \tag {11} +$$ + +with $\mathbf{A}$ being a linear orthogonal matrix. This indicates that the transformation $A$ is learned to annul the change in data variables or the effect of augmentations applied. + +# 3.4. Reason and solutions for unstable change in data variables + +In section 3.3 we show that generalized representation is robust to augmentations since $\mathbf{A}$ , the linear transformation between the ground truth representation and inferred representations, is orthogonal to augmentation applied during the training. And if a change in the data variable reflects a shift, $\delta \mathbf{z}$ , in the ground truth representations and the newly inferred representation is stable, meaning $\delta \mathbf{z}$ will be absorbed by the transformation matrix $\mathbf{A}$ , then the shift in the ground truth representation corresponds to an augmentation that is applied during the training. However, when $\delta \mathbf{z}$ appears in the range of $\mathbf{A}$ and $\mathbf{A} \delta \mathbf{z} \neq 0$ , the resultant inferred representation can be unstable and lead to performance drop + +on downstream models (denote as $D(\mathbf{z})$ ). We quantify this deterioration on $D$ by: + +$$ +m \left(D (h (\mathbf {z})) _ {\text {s t a b l e}}\right) - m \left(D \left(h (\tilde {\mathbf {z}}) _ {\text {u n s t a b l e}}\right) \right. \tag {12} +$$ + +where $m(.)$ is a metric on an outcome of the downstream task. In example of $D$ being a linear classifier, $m(.)$ can be the probability of predicting the target class (prediction score) or the proportion of correct predictions (accuracy). In order to overcome this deterioration, we propose two methods, namely Robust Dimensions and Stable Inference Mapping: + +1. Robust Dimensions: Under a stable condition, $D(f(\mathbf{x}))^2 = D(\mathbf{A}\mathbf{z})$ , each dimension in the inferred representation $f(\mathbf{x})$ is a linear combination of dimensions of the ground truth representation. The dimensions contributing most to $D()$ should be robust to unstable shift $\delta \mathbf{z}_{\text{unstable}}$ . In other words, most robust dimensions of $f(\mathbf{x})$ should be also robust in $f(\tilde{\mathbf{x}})$ where $\tilde{\mathbf{x}} = g(\mathbf{z} + \delta \mathbf{z}_{\text{unstable}})$ as some dimensions of $\mathbf{A} \delta \mathbf{z}_{\text{unstable}}$ will be zero. Hence identifying most important dimensions in $f(\mathbf{x})_{\text{stable}}$ and pass through the same dimensions of $f(\tilde{\mathbf{x}})_{\text{unstable}}$ should alleviate the deterioration by making $m(D(h(\mathbf{z}))_{\text{stable}})_{\{\dim\}} = m(D(h(\tilde{\mathbf{z}})_{\text{unstable}}))_{\{\dim\}}$ where $\{\dim\}$ is a set of most robust dimensions. In example of $D$ being a linear classifier, the contribution of each dimension can be calculated by $W_c^\top f(\mathbf{x})$ where $W_c^\top$ is the Jacobian of the linear classifier w.r.t target class $c$ . + +2. Stable Inference Mapping: Since $f(\mathbf{x})_{\text{stable}} = -f(\tilde{\mathbf{x}})_{\text{unstable}} = -\mathbf{A} \delta \mathbf{z}$ , we can learn another linear transformation $\mathbf{F}$ to absorb $\delta \mathbf{z}$ . Especially, we want to learn $\mathbf{F} f(\tilde{\mathbf{z}})_{\text{unstable}} = \mathbf{F} \mathbf{A} \mathbf{z} + \mathbf{F} \mathbf{A} \delta \mathbf{z}$ such that the additional $\mathbf{F}$ will not only set $\mathbf{A} \delta z$ to 0, but also make stable representation more robust to the unstable shift by assuring $\mathbf{F} \mathbf{A}$ orthogonal to $\delta \mathbf{z}_{\text{unstable}}$ in addition to the augmentations applied during training. Formally, we model a linear layer $l(f(\tilde{\mathbf{x}})_{\text{unstable}}) = f(\mathbf{x})_{\text{stable}} - f(\tilde{\mathbf{x}})_{\text{unstable}}$ , hence during the inference $f(\mathbf{x}) + l(f(\mathbf{x}))$ is used for downstream task. + +Relation to Causal Inference Since we only have observations in the image space, we consider the augmentations or changes of a data variable as interventions on the ground truth representations: $\tilde{\mathbf{x}} = g(\tilde{\mathbf{z}} |(c,do(V_i = v_i))$ . With an access to ground truth representation, we can evaluate treatment-control effect i.e. $Pr(f(g(\tilde{\mathbf{z}} |(c,do(V_i = v_{stable}))) - Pr(f(g(\tilde{\mathbf{z}} |(c,do(V_i = v_{unstable})))$ . Without any access to the ground truth representations, we can evaluate the average treatment effect [33] by $\mathbb{E}[D(f(\mathbf{x})_{stable})-$ + +$D(f(\tilde{\mathbf{x}})_{\text{unstable}})]$ via synthesizing manual data samples $\tilde{\mathbf{x}}_i = g(\tilde{\mathbf{z}} | (c, \text{do} (V_i = v_i))$ . + +# 4. Experiments and Discussions + +In this section we evaluate our solutions to unstable shift in data variable on two datasets: Causal3DIdent and ImageNet pretrained SSL and corresponding linear classifiers as the downstream task. + +# 4.1. Causal3DIdent + +![](images/eec2275428e8faf79fdde2e91d218a79d519c3d158d53b9f7ff570a4c454a55e.jpg) +Figure 3: Causal graph for $V$ in Causal3DIdent. + +[76] develops the dataset that 7 objects in the dataset that each $224 \times 224$ image is associated with an object class and a 10-dimensional latent representation. These 10 dimensions correspond to 10 data variables. A causal graph imposed on these variables is shown in Figure 3. As the dependency shown in Figure 3, an object is placed at $pos_{obj} = (x,y,z)$ with $rot_{obj} = (\phi, \theta, \psi)$ and $hue_{o}bj$ under a spot light at an angle $pos_{spl}$ with color $hue_{spl}$ on a background with color $hue_{bg}$ . Detailed information in C. + +To simulate unstable change of variable that is not accessible during the training, a range of some data variables is hold out and hence this portion of data is treated as potential unstable samples that causes the performance drop on the linear classifier. Since all data variables are truncated in range [-1,1], we hold out the edge value(s) to further portray 'unexpected' values during the inference time. We select 8 dimensions to intervene, namely: $z$ in object position3, all 3 object rotation angles, spot light position, and all 3 hue variables. Both training and testing data are sampled and for detailed sampling procedures refer to C. + +SSL Experiment Setup ResNet18[30] is the backbone of the feature encoder $f$ . Same augmentations in [10] are applied during training. An Adam optimizer with a learning rate at 0.0001 and a weight decay at 0.00001 is optimized for all discriminative SSL. Hyperparameters for each SSL is presented in C. The dimension of the inferred representation space is set to 128. The network is trained for 20 epochs on intervened data. Then a linear classifier is trained on the frozen representations of the network with a same optimizer for 10 epochs. + +To illustrate the simulation results in the deterioration of the downstream performance, for each testing data that has the same data variable distribution as the training data, we search in the ground truth representation space for 5 nearest representations in the hold-out test distribution when $n$ dimension of the testing data is shifted to a random value in the hold-out range of corresponding data variable. Within the corresponding 5 images representing a shift in $n$ dimensions that is not seen during the training, we record the lowest performance to fulfill the deterioration scenario. In Figure 4 the score and accuracy are averaged over all combinations of $\binom{8}{n}$ . When $n = 0$ , this indicates the performance of SSL on the testing data with the same data variable distribution as the training data. The performance is exceptional since hold-out values of some data variables may be covered by the augmentations applied ( $z$ covered by random cop, and hue colors covered by color distortions). A will be orthogonal to changes in these data variables. In fact in C we show that the performance difference between seen distribution and unseen distribution in testing data is comparatively small. As illustrated in Figure 4, even when only one variable is changed to a unseen value, there is a large drop in both accuracy and prediction score (20% in accuracy and 30% in prediction score for SimCLR). With more tangled changes in data variables, the unstable representations result in poorer downstream performance. Since the selection of the data variables may be too complex due to the dependency, we also validate the same issue on selecting only children nodes in Figure 3. And also we visualize the latent shift between stable and unstable examples. See C for more results. + +Robust Dimensions For each pair of testing data $\mathbf{x}_{\text{stable}}$ with seen data variable distribution and the selected data $\mathbf{x}_{\text{unstable}}$ among the 5 nearest neighbours when changing $n$ dimensions to unseen distribution values, we apply the Jacobian of the linear classifier w.r.t the target class $W_{c}^{\top}$ on the stable representation $f(\mathbf{x})_{\text{stable}}$ to identify the top $k$ most important dimensions and pass the same dimensions of $f(\tilde{\mathbf{x}})_{\text{unstable}}$ to $D$ to evaluate the performance. As shown in Figure 5, the accuracy only deteriorate slightly when top $90\%$ most important features of $f(\tilde{\mathbf{x}})_{\text{unstable}}$ are selected for the downstream task. This is true even when all 8 variables are shifted. This suggests that $\mathbf{A}$ is orthogonal to changes in ground truth representation $\mathbf{z}$ in most dimensions. This high percentage of dimensions may be due to $f$ optimizing the SSL objectives to a high level and the augmentations applied covers some of the hold-out variable values. Interestingly, there are cases where passing the top $k\%$ (around $40\%$ ) important dimensions results in higher performance than stable representations (uninterrupted in Figure 5). However, as expected, including more less important dimensions where the unseen shift in the data variable results in non-zero adjustment ( $\mathbf{A} \delta \mathbf{z} \neq 0$ ) initiates the + +![](images/21cc225abf728803edbfd65cec6a68b81ee71a8a6786f030d15c49190fc88e6a.jpg) +(a) Deterioration in Accuracy + +![](images/e2fb9f3265c94903183cf9b34812e200e1b9ef60bae532cf6d45e99763d04480.jpg) +(b) Deterioration in Score +Figure 4: Deterioration of unstable changing in data variables on all SSL. The more unexpected changes occur in the data variables, the more severe the deterioration. + +deterioration phenomenon we observe. + +![](images/9c1ed25508b7d95acd0189f359673b4a542b7f92f937c9ba9d4059d09b8521cb.jpg) +(a) Improvement on Accuracy + +![](images/d6199777575f6d3f01cb0c27e3d47513d3562264b32832d64ca7ccd7d152bd88.jpg) +(b) Improvement on Score +Figure 5: By identifying and passing top $k\%$ important dimensions, the deterioration in both accuracy and prediction score is significantly alleviated. + +Stable Inference Mapping As shown in 3.4, a linear transformation $\mathbf{F}:\mathbb{R}^{d_2}\to \mathbb{R}^{d_2}$ is trained to cancel the effect $\mathbf{A}\delta \mathbf{z}$ and also further improve the robustness of transformation on the ground truth representation $\mathbf{Az}$ . In this experiment, we match each training data with a random member in the 5 nearest representations when only one dimension is changed to a unstable value. A linear layer is trained with the same optimizer for 10 epochs on the training pairs. In Table 1, all accuracy for unstable examples increases significantly (except SimSiam) after learning $\mathbf{F}$ . However, since the $f$ is very close to the global minimizer of the alignment term, the improvement on stable examples cannot be observed. + +
xstablex unstable
w/o Fw/ Fw/o Fw/ F
SimCLR0.9960.9980.8330.889
MoCo0.9920.9920.8000.849
BYOL0.9960.9980.8860.920
SimSiam0.9980.9990.9190.928
Barlow Twins0.9910.9950.8180.862
+ +Table 1: The effect of $\mathbf{F}$ on both stable and unstable samples. The accuracy is average over 3 random seeds. + +![](images/a06d872cf06ffeae1ff707465a95603fe0cb65a1b43151084ed33d857fed1a38.jpg) +Figure 6: Examples of ObjectNet dataset $(1^{st}$ row), Stylized ImageNet dataset $(2^{nd}$ row), Synthetic dataset $(3^{rd}$ row). + +# 4.2. ImageNet + +To validate our findings on a larger scale and more realistic settings, we apply proposed solutions on ImageNet[16] with various altered and synthetic datasets as unstable shift in the data variable. See Figure 6 and D for more information on ObjectNet, Stylized-ImageNet, Synthetic dataset. + +ObjectNet[3] is a collection of objects that are intentionally placed at an unusual view angle and backgrounds, so that the bias learned by the model with usual data variables is more prominent when testing on ObjectNet. With focus- + +ing on 113 classes overlapping with ImageNet classes, we evaluate our second proposed solution explicitly on ObjectNet since the shift presented in ObjectNet is very unstable and it is very challenging to overcome the negative effect of the shift. + +Stylized ImageNet[24] change the style of original ImageNet image to a random artistic style. With this drastic shift in the data variable, the analysis of the first solution is more insightful on the robustness of dimensions of $\mathbf{A}$ on a very different shift in data variable. + +Synthetic Data follows synthetic procedure in [18] where object is masked on a background at a location with a rotation angle. We explore the benefit of this synthesized dataset at three modes: background, location, rotation where the target variable is randomly sampled with the other two variables fixed. Additionally, we explore texture as an independent variable by masking 'texturized' objects in [23]. We set total number of updating steps per epoch as 100 with batch size 256. This means we select total number of 512000 synthesized images every epoch. + +Experiment Setup ResNet 50 pretrained on ImageNet and a linear classifier fintuned via SimCLR, BYOL, and SimSiam are tested with both proposed solutions. For Stable Inference Mapping, a linear layer is optimized with an Adam optimizer for 10 epochs on the synthetic dataset. + +Robust Dimensions For each of ImageNet validation data sample, we stylized the image to a random artistic fashion. We observe the dramatic performance difference between the ImageNet stable images and Stylized unstable ImageNet images $(\mathbf{x}_{stable}, \mathbf{x}_{unstable})$ . The result is shown in Figure 7. For SimCLR, passing the top $10\%$ important dimensions can close a small performance gap between stable representations and unstable representations. Nonetheless, all SSL seem to be sensible to the strong style change as they shorten the difference between ImageNet to an modest extent. This is as expected since the Stylized ImageNet changes multiple variables to an extreme value. + +Stable Inference Mapping As described in Synthetic Data, we explore the benefit of our second proposed method to evaluate on a dataset that most of samples are unstable according to learned $\mathbf{A}$ . At each training step, with other variables randomly fixed, 10 images with random target variable values are generated. The pair with maximum $m(D(\mathbf{x})) - m(D(\tilde{\mathbf{x}}))$ is selected to train the linear transformation. In Table 2, inferring $\mathbf{F}$ via controlling location produces least improvement. This is expected since the network is robust to translation by design. While background, rotation, and texture improves the performance considerably with the consideration on the training time. However, in D we show that training the model longer using Stable Inference Mapping yield less favorable results since the improvement is less significant and starts to saturate at around 30 epochs. + +![](images/0ce5deaa8597f4d839b7edf29964ce209f5b1c963cba81d281dbb1832ab57ae0.jpg) + +![](images/ef24a24ca9ebf8c9b8309833600ec769da956d94d114a591c3b9ed67c0842140.jpg) +(a) Improvement on Accuracy +(b) Improvement on Score +Figure 7: By identifying and passing top $k\%$ important dimensions, the deterioration in both accuracy and prediction score is alleviated. + +
SimCLRBYOLSimSiam
w/o F10.1214.0411.15
background12.3415.4312.45
location10.3514.1311.17
rotation12.0415.8812.38
texture12.7915.0613.11
+ +Table 2: Comparison of inferring different $\mathbf{F}$ on ObjectNet with a target data variable random trials. + +# 5. Limitations + +Though we identify the root cause of unstable inference for discriminative SSL by constructing a causal framework inspired by the prior work, the proposed solutions are constrained and limited to be applied on realistic applications. Robust Dimensions involves establishing a correspondence between stable and unstable instances on a one-to-one basis, enabling the identification of dimensions contributing to stability. On the other hand, Stable Inference Mapping necessitates a collection of unstable instances with a specific alteration in a particular group of data variables. Within the Causal3dIdent dataset, both solutions can be assessed using the same unstable instances. In more realistic datasets, achieving a one-to-one correspondence is feasible, and manipulation of one group of data variables can be accomplished using synthetic data. However, any assessments with involving artificially generated images might introduce some level of uncertainty. In a realistic setup, since training samples are not directly observ- + +able during the inference stage, simple interventions on inference samples may not effectively separate the unstable variables from the stable ones. Consequently, the potential benefits of the proposed solutions in realistic datasets are undermined. + +# 6. Conclusions + +In conclusion, this paper has proposed a novel approach to address the issue of unstable behavior during the inference stage in SSL methods. By building on the previous theories of successful InfoNCE-facilitated contrastive SSL and extending it to recent SSL methods, we have demonstrated that a change in the data factor can result in a shift in the inferred representation, leading to a decline in downstream performance. We have proposed learning targeted transformations that regularize the violating shift and restore performance on the unseen data shift. Our experiments on both controlled and realistic datasets have shown the efficacy of our proposed solutions. These contributions provide a better understanding of SSL methods and offer a promising solution to the problem of unstable behavior during the inference stage. We hope that our work will inspire further research in this area and lead to improved SSL methods that are more robust to changes in data factors. + +# 7. Acknowledgements + +The work was funded in part by NSF CNS-2112562 and IIS-2140247. + +# References + +[1] Ibrahim Ahmad and Pi-Erh Lin. A nonparametric estimation of the entropy for absolutely continuous distributions (corresp.). IEEE Transactions on Information Theory, 22(3):372-375, 1976. 14 +[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2 +[3] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 7 +[4] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006. 3 +[5] Sergiy V Borodachov, Douglas P Hardin, and Edward B Saff. Discrete energy on rectifiable sets, volume 3. Springer, 2019. 4 +[6] Yue Cao, Mingsheng Long, and Jianmin Wang. Unsupervised domain adaptation with distribution matching ma + +chines. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 3 +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 1 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1 +[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 5, 6, 13 +[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1, 2 +[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 1 +[13] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 17 +[14] Henry Cohn and Abhinav Kumar. Universally optimal distribution of points on spheres. Journal of the American Mathematical Society, 20(1):99-148, 2007. 4 +[15] Hal Daumé III. Frustratingly easy domain adaptation. arXiv preprint arXiv:0907.1815, 2009. 3 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7 +[17] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516, 2014. 2 +[18] Josip Djolonga, Jessica Yung, Michael Tschannen, Rob Romijnders, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Matthias Minderer, Alexander D'Amour, Dan Moldovan, Sylvain Gelly, Neil Houlsby, Xiaohua Zhai, and Mario Lucic. On robustness and transferability of convolutional neural networks, 2020. 8, 16 +[19] Lixin Duan, Dong Xu, and Ivor Tsang. Learning with augmented features for heterogeneous domain adaptation. arXiv preprint arXiv:1206.4660, 2012. 3 +[20] Deniz Erdogmus, Kenneth E Hild, Yadunandana N Rao, and Jose C Principe. Minimax mutual information approach for independent component analysis. Neural Computation, 16(6):1235-1252, 2004. 2 + +[21] Ronald Aylmer Fisher. Dispersion on a sphere. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 217(1130):295-305, 1953. 3 +[22] Bernhard C. Geiger and Gernot Kubin. On the information loss in memoryless systems: The multivariate case. CoRR, abs/1109.4856, 2011. 14 +[23] Robert Geirhos, Claudio Michaelis, Felix A. Wichmann, Patricia Rubisch, Matthias Bethge, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. 7th International Conference on Learning Representations, ICLR 2019, (c):1-22, 2019. 8 +[24] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 8 +[25] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 1, 2 +[26] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 5 +[27] Jeff Z. HaoChen, Colin Wei, Adrien Gaidon, and Tengyu Ma. Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Loss. Advances in Neural Information Processing Systems, 7(NeurIPS):5000-5011, 2021. 2 +[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1 +[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 5 +[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[31] Geoffrey E Hinton and Richard Zemel. Autoencoders, minimum description length and helmholtz free energy. Advances in neural information processing systems, 6, 1993. 1 +[32] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670, 2018. 2 +[33] Paul W Holland. Statistics and causal inference. Journal of the American statistical Association, 81(396):945-960, 1986. 5 +[34] Tzu Ming Harry Hsu, Wei Yu Chen, Cheng-An Hou, Yao-Hung Hubert Tsai, Yi-Ren Yeh, and Yu-Chiang Frank Wang. + +Unsupervised domain adaptation with imbalanced cross-domain data. In Proceedings of the IEEE International Conference on Computer Vision, pages 4121-4129, 2015. 3 +[35] Aapo Hyvarinen and Patrik Hoyer. Emergence of phase-and shift-invariant features by decomposition of natural images into independent feature subspaces. Neural computation, 12(7):1705-1720, 2000. 4 +[36] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and Nonlinear ICA. Advances in Neural Information Processing Systems, (Nips):3772-3780, 2016. 2 +[37] Aapo Hyvarinen and Hiroshi Morioka. Nonlinear ICA of temporally dependent stationary sources. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 2017. 2 +[38] Aapo Hyvarinen and Petteri Pajunen. Nonlinear independent component analysis: Existence and uniqueness results. Neural networks, 12(3):429-439, 1999. 2 +[39] Aapo Hyvarinen, Hiroaki Sasaki, and Richard E. Turner. Nonlinear ICA using auxiliary variables and generalized contrastive learning. AISTATS 2019 - 22nd International Conference on Artificial Intelligence and Statistics, 89, 2020. 2 +[40] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1):2, 2020. 1 +[41] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 1 +[42] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4893-4902, 2019. 3 +[43] Ilyes Khemakhem, Diederik P. Kingma, Ricardo Pio Monti, and Aapo Hyvarinen. Variational Autoencoders and Nonlinear ICA: A Unifying Framework. 108(2015), 2019. 2 +[44] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 2 +[45] David Klindt, Lukas Schott, Yash Sharma, Ivan Ustyuzhaninov, Wieland Brendel, Matthias Bethge, and Dylan Paiton. Towards Nonlinear Disentanglement in Natural Data with Temporal Sparse Coding. pages 1-51, 2020. 2 +[46] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 16 +[47] Quoc V Le, Will Y Zou, Serena Y Yeung, and Andrew Y Ng. Learning hierarchical invariant spatio-temporal features + +for action recognition with independent subspace analysis. In CVPR 2011, pages 3361-3368. IEEE, 2011. 4 +[48] Jason D. Lee, Qi Lei, Nikunj Saunshi, and Jiacheng Zhuo. Predicting What You Already Know Helps: Provable Self-Supervised Learning. Advances in Neural Information Processing Systems, 1(NeurIPS):309-323, 2021. 2 +[49] Chunyuan Li, Jianwei Yang, Pengchuan Zhang, Mei Gao, Bin Xiao, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Efficient Self-supervised Vision Transformers for Representation Learning. 3:1-27, 2021. 1 +[50] Wen Li, Lixin Duan, Dong Xu, and Ivor W Tsang. Learning with augmented features for supervised and semi-supervised heterogeneous domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 36(6):1134-1148, 2013. 3 +[51] Ralph Linsker. Self-organization in a perceptual network. Computer, 21(3):105–117, 1988. 2 +[52] Shuo Liu, Adria Mallol-Ragolta, Emilia Parada-Cabaleiro, Kun Qian, Xin Jing, Alexander Kathan, Bin Hu, and Bjöern W Schuller. Audio self-supervised learning: A survey. Patterns, 3(12):100616, 2022. 1 +[53] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering, 35(1):857-876, 2021. 1 +[54] Yixin Liu, Ming Jin, Shirui Pan, Chuan Zhou, Yu Zheng, Feng Xia, and Philip Yu. Graph self-supervised learning: A survey. IEEE Transactions on Knowledge and Data Engineering, 2022. 1 +[55] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer joint matching for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1410-1417, 2014. 3 +[56] Jovana Mitrovic, Brian McWilliams, Jacob Walker, Lars Buesing, and Charles Blundell. Representation Learning via Invariant Causal Mechanisms. pages 1-21, 2020. 2 +[57] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI, pages 69-84. Springer, 2016. 1, 2 +[58] Aaron Van Den Oord. Representation Learning with Contrastive Predictive Coding. 2 +[59] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 1 +[60] Judea Pearl. Causality: Models, reasoning, and inference, second edition. 2011. 1 +[61] Madeline C Schiappa, Yogesh S Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ACM Computing Surveys, 2022. 1 +[62] Saeed Shurrab and Rehab Duwairi. Self-supervised learning methods and applications in medical imaging analysis: A survey. PeerJ Computer Science, 8:e1045, 2022. 1 + +[63] Mamatha Thota and Georgios Leontidis. Contrastive domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2209-2218, 2021. 3 +[64] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 15 +[65] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 1 +[66] Riccardo Volpi, Pietro Morerio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5495-5504, 2018. 3 +[67] Kügelgen Von, Julius and, Wieland Brendel, Bernhard Schölkopf, Michel Besserve, and Francesco Locatello. Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style Luigi Gresele. Yash Sharma, 3(NeurIPS):4, 2021. 2, 3, 4, 15 +[68] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3, 4 +[69] Rui Wang, Zuxuan Wu, Zejia Weng, Jingjing Chen, Guo-Jun Qi, and Yu-Gang Jiang. Cross-domain contrastive learning for unsupervised domain adaptation. IEEE Transactions on Multimedia, 2022. 3 +[70] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. 37th International Conference on Machine Learning, ICML 2020, PartF168147-13:9871-9881, 2020. 3, 4 +[71] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 1 +[72] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-Supervised Learning with Swin Transformers. pages 1-8, 2021. 1 +[73] Detai Xin, Tatsuya Komatsu, Shinnosuke Takamichi, and Hiroshi Saruwatari. Disentangled speaker and language representations using mutual information minimization and domain adaptation for cross-lingual tts. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6608-6612. IEEE, 2021. 2 +[74] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 1, 2, 3, 14, 15 +[75] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, Octo + +ber 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 1, 2 +[76] Roland S. Zimmermann, Yash Sharma, Steffen Schneider, Matthias Bethge, and Wieland Brendel. Contrastive Learning Inverts the Data Generating Process. 2021. 2, 3, 4, 5, 6, 15 \ No newline at end of file diff --git a/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/images.zip b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f31eccb5bd1569c9015bbafbe03e0b7117a9e4ff --- /dev/null +++ b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1ac93525525bfcc9c1f0722efb6d2238c0f269e6b7c77a939ea98ca45e3c69a +size 374235 diff --git a/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/layout.json b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5260669394b51c12d40fb69a61a29008b551d37f --- /dev/null +++ b/2023/Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations/layout.json @@ -0,0 +1,11310 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 529, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 529, + 139 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 529, + 139 + ], + "type": "text", + "content": "Stable and Causal Inference for Discriminative Self-supervised Deep Visual Representations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 215, + 161, + 378, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 161, + 378, + 175 + ], + "spans": [ + { + "bbox": [ + 215, + 161, + 378, + 175 + ], + "type": "text", + "content": "Yuewei Yang, Hai Li, Yiran Chen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 256, + 176, + 337, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 176, + 337, + 190 + ], + "spans": [ + { + "bbox": [ + 256, + 176, + 337, + 190 + ], + "type": "text", + "content": "Duke University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 190, + 332, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 190, + 332, + 202 + ], + "spans": [ + { + "bbox": [ + 261, + 190, + 332, + 202 + ], + "type": "text", + "content": "Durham, USA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 205, + 455, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 205, + 455, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 205, + 455, + 217 + ], + "type": "text", + "content": "yuwei.yang@duke.edu, hai.li@duke.edu, yiran.chen@duke.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 271, + 290, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 290, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 290, + 535 + ], + "type": "text", + "content": "In recent years, discriminative self-supervised methods have made significant strides in advancing various visual tasks. The central idea of learning a data encoder that is robust to data distortions/augmentations is straightforward yet highly effective. Although many studies have demonstrated the empirical success of various learning methods, the resulting learned representations can exhibit instability and hinder downstream performance. In this study, we analyze discriminative self-supervised methods from a causal perspective to explain these unstable behaviors and propose solutions to overcome them. Our approach draws inspiration from prior works that empirically demonstrate the ability of discriminative self-supervised methods to demix ground truth causal sources to some extent. Unlike previous work on causality-empowered representation learning, we do not apply our solutions during the training process but rather during the inference process to improve time efficiency. Through experiments on both controlled image datasets and realistic image datasets, we show that our proposed solutions, which involve tempering a linear transformation with controlled synthetic data, are effective in addressing these issues." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 560, + 128, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 128, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 128, + 573 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Learning generalized representation with unlabeled data is a challenging task in various fields, but Self-Supervised Learning (SSL) has recently demonstrated remarkable success in learning semantic invariant representations without labels [40, 41, 53]. There are two main types of self-supervised learning (SSL) based on the pretext task used: generative and discriminative SSL, with generative SSL reconstructing altered or distorted data to its original input [9, 28, 31, 59, 65, 71] and early discriminative SSL predicting easily designed labels and task-specific representations that are not very generalizable [25, 57, 75]. More recent" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 340, + 243, + 514, + 331 + ], + "blocks": [ + { + "bbox": [ + 340, + 243, + 514, + 331 + ], + "lines": [ + { + "bbox": [ + 340, + 243, + 514, + 331 + ], + "spans": [ + { + "bbox": [ + 340, + 243, + 514, + 331 + ], + "type": "image", + "image_path": "210d2f49e331f76ac6e8ec4cb2e6ed001d0b54603621d409dbce501dd562d9ee.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 341, + 332, + 513, + 420 + ], + "blocks": [ + { + "bbox": [ + 341, + 332, + 513, + 420 + ], + "lines": [ + { + "bbox": [ + 341, + 332, + 513, + 420 + ], + "spans": [ + { + "bbox": [ + 341, + 332, + 513, + 420 + ], + "type": "image", + "image_path": "77582930b2b339d2c9e59af355f71daa3dbb052ba51491edb704d4a8f5985767.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "lines": [ + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 547, + 500 + ], + "type": "text", + "content": "Figure 1: During the training of discriminative SSL, aligning positive representations will be robust to the changes applied as augmentations (red arrows). However, during the inference stage, one small change in the data variable (such as view angle) will result in an unexpected degradation on downstream performance." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 700 + ], + "type": "text", + "content": "discriminative SSL trains the model to identify similarities and differences between pairs of augmented examples [7, 10, 11, 26, 29, 74]. The success of SSL in deep image models has resulted in progress in other data modalities [53, 52, 54, 61, 62] and attention-based models like transformers [12, 8, 49, 72]. Recent discriminative SSL aims to learn content and semantic invariant representations that are robust to data augmentations, but the learned representations can be unstable when one subtle factor of the data is changed to a value that is not accessible through all augmentations. To avoid the high cost of incorporating all possible subtle changes during training, insights are needed to uncover the root cause of instability and find a solution to prevent performance deterioration during inference. Figure 1 summarizes this deterioration effect." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Causality [60] is a vital tool to investigate the causal" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16109" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "relationships between variables in observational data, and can uncover the underlying causal factors that explain unexpected model behavior due to changes in the environment. Independent Component Analysis (ICA) is often used to disentangle sources in unsupervised training [36, 37, 39, 43, 45], and causal analysis has been applied in follow-up works [76, 67] to examine the empirical success of SSL under an ICA framework. However, while these works identify factors that contribute to SSL's success, they do not address the problem's unstable mode, which can cause a severe performance drop when underlying factors shift slightly to an unseen environment. Some works [56, 27, 48] attempt to incorporate causality during the training process to identify and alleviate the impact of such shifts, but this approach is time-costly and only marginally improves performance compared to non-causal SSL methods. A more time-efficient and accessible solution would be to simply reverse the unstable shift during inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 527 + ], + "type": "text", + "content": "We aim to address the issue of unstable behavior during the inference stage by building upon previous theories of successful InfoNCE-facilitated contrastive SSL and extending it to all recent SSL methods with additional assumptions and constraints. Drawing inspiration from the relationship between the ground truth positive pairs distribution and learned positive pairs distribution, we demonstrate that the approximated transformation between the ground truth representations and learned representations is orthogonal to the augmentations applied during training. However, a change in the data factor/variable that violates the conditions for successful SSL can cause a corresponding shift in the inferred representation, resulting in a decline in downstream performance. This change of data factor/variable can be a change in the background, texture, or view angles etc. To overcome this issue, we propose learning targeted transformations that regularize the violating shift and restore performance on the unseen data shift. This approach effectively avoids the undesirable behavior and improves performance on previously unseen data shifts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 528, + 251, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 528, + 251, + 540 + ], + "spans": [ + { + "bbox": [ + 59, + 528, + 251, + 540 + ], + "type": "text", + "content": "To summarize, our contributions are following:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 550, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 550, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 550, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 59, + 550, + 287, + 586 + ], + "type": "text", + "content": "- Through the use of a comparable data generation process in prior research, we show that All current SSL techniques benefit from the alignment of positive pairs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 595, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 595, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 59, + 595, + 287, + 655 + ], + "type": "text", + "content": "- Through our alternative derivation of the transformation matrix between the ground truth representation and the learned representation, we have shown that the augmentations applied during training are orthogonal to the resulting matrix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 287, + 713 + ], + "type": "text", + "content": "- By interpreting a change in the data variable causally, we propose two solutions focusing on inference to modify the negative shift in representation space caused by such a change." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 120 + ], + "type": "text", + "content": "- We validate the proposed solutions by conducting experiments on both controlled and realistic datasets, providing evidence for their efficacy during the inference stage without retraining the pretrained models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 131, + 392, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 131, + 392, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 131, + 392, + 144 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 151, + 547, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 151, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 547, + 486 + ], + "type": "text", + "content": "Discriminative SSL learns invariant representations from positive pairs of unlabeled data samples, but previous attempts to use trivial labels like colors[75], rotations[25], and patch positions[57] offer minimal benefits for complex downstream tasks due to their easy augmentations. Recent discriminative SSL apply random augmentations to generate two views of an image sample and train an encoder to extract representations for maximizing similarity between the paired augmentations. SimCLR and MoCo[10, 29] are pioneer SSL works that maximize the cosine similarity between positive pairs and minimize the cosine similarity between negative pairs via optimizing the InfoNCE loss[58]. Immense resources are used to enforce a large number of negative samples since a larger number can tighten the upper bound of the mutual information between positive pairs[58]. Later advancement of discriminative SSL excludes the notion of negative pairs by only aligning positive pairs and preventing representation collapse through various regularizations. BYOL[26] predicts an Exponential-Moving-Averaged (EMA) representation of one view with a projected representation of another view. SimSiam[11] maximizes similarity between a projected representation and a detached representation of two positive samples. Unlike previous work focusing optimization on an instance level, Barlow Twins[74] encourages high similarity in corresponding feature dimensions and discourages redundancy across different feature dimensions between two views of a data sample. Detailed formulations are exhibited in A." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 665 + ], + "type": "text", + "content": "Mutual Information is a different perspective on the behaviour of discriminative SSL. Referred to InfoMax principle[51], the MI between different transformations of a data sample is maximized via optimizing the InfoNCE loss[2, 32, 44]. Though showing theoretical relation between optimizing InfoNCE and maximizing MI between positive pairs, the underlying factors instructing the behaviour of different SSL methods are not explored. Nonlinear ICA[38], on the other hand, captures complex data structures of SSL methods by disentangling underlying factors via minimizing the mutual information between learned representations and the original data input[17, 73, 20]. Other works associate the nonlinear ICA objectives with the contrastive SSL so the MI between positive pairs are maximized and negative pairs are minimized." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Other researchers have explored causality and causal inference as a means of understanding the success of discriminative SSL. Prior work has focused on partitioning the InfoNCE loss to alignment between the positive pairs and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16110" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "uniformity between aggregations of all positive clusters[70]. By formulating a data generation process, [76] empirically explains that networks optimized via InfoNCE infer an orthogonal transformation of the ground truth latent representations. Furthermore [67] validates that that augmentations used in both generative and discriminative SSL isolate the content factor from the style factor. Our theory and work draw great inspirations from these two work. However, instead of solely focusing on InfoNCE-driven SSL and two factors, this work extends the framework to all recent discriminative SSL methods and identifies reasons for unstable circumstances analytically. We also propose methods to nullify the negative effects of unstable representations during inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 290, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 445 + ], + "type": "text", + "content": "Domain Adaptation is a strategy to bridge the gap between the model performance on a source domain and that on a target domain [4, 68]. Feature adaptation methods try to learn a new feature representation that is more invariant to the domain shift[66, 19, 15, 50], while instance adaptation methods try to reweight the importance of the labeled source examples to better align with the target domain[55, 6, 34]. Recent studies also implement a contrastive framework to learn a shared latent space between the source and target domains by maximizing the agreement between representations of corresponding samples while minimizing the agreement between representations of noncorresponding samples [69, 63, 42]. Unlike these works focusing on adapting to a target domain for better performance, we investigate the underlying causal factors for the performance gap and based on the findings we propose easy solutions to connect the gap." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 453, + 100, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 100, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 100, + 466 + ], + "type": "text", + "content": "3. Theory" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 473, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 605 + ], + "type": "text", + "content": "In this section, we build on previous data generation process [76, 67] (3.1) and show that current SSL techniques benefit from the alignment of positive pairs (3.2). During a deeper dive into the generation process, we show the relation between the ground truth representations and the inferred representations, and this relation is an linear transformation matrix that is orthogonal to the augmentations applied during training (3.3). Finally, we disclose the causal reason for the unstable behaviour caused by a change in the data variable during the inference stage and propose analytical solutions to address this unstable issue (3.4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 611, + 170, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 170, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 170, + 624 + ], + "type": "text", + "content": "3.1. Problem formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 666 + ], + "type": "text", + "content": "Data generation is assumed to be a generation function that takes ground truth latent representation as the input to generate an observation data/image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Formally, we assume that the marginal distribution of sampling ground truth representations " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\subseteq \\mathcal{Z} \\in \\mathbb{R}^{d_1}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " w.r.t. a class is uniform on a unit sphere " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^{d_1 - 1}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": ". An injective generation function " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "g(.) : \\mathbb{R}^{d_1} \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " takes a ground truth rep-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 393, + 69, + 469, + 173 + ], + "blocks": [ + { + "bbox": [ + 393, + 69, + 469, + 173 + ], + "lines": [ + { + "bbox": [ + 393, + 69, + 469, + 173 + ], + "spans": [ + { + "bbox": [ + 393, + 69, + 469, + 173 + ], + "type": "image", + "image_path": "99cf6be555efe2964c0444eec817c0c6c366220f40744e28d9f136218478a52f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 179, + 533, + 192 + ], + "lines": [ + { + "bbox": [ + 318, + 179, + 533, + 192 + ], + "spans": [ + { + "bbox": [ + 318, + 179, + 533, + 192 + ], + "type": "text", + "content": "Figure 2: A causal graph for data generation process." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "text", + "content": "resentation and generate an observation sample " + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = g(\\mathbf{z})" + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "text", + "content": ". Different variables, denoted as a set " + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "inline_equation", + "content": "V = \\{V_i\\}" + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "text", + "content": " and class/content variable " + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "text", + "content": ", constitute the values of each dimension in the ground truth representation. These variables can be direct causes or confounding factors, " + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = (c, \\{V_i\\})" + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "text", + "content": ". Specific examples include view angles, object size, background colors, etc. A simple causal graph depicts this relationship is shown in Figure 2. So the general generation process is as following:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 342, + 330, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 330, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 342, + 330, + 545, + 357 + ], + "type": "interline_equation", + "content": "p (\\mathbf {z} = (c, V _ {*})) \\sim \\frac {1}{| \\mathcal {Z} |} \\quad \\mathbf {x} = g (\\mathbf {z}) \\tag {1}", + "image_path": "940b5bdd146d646ab4ab30358e57e73c5c5d27c02e4669ad2224d00a56af8447.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 363, + 547, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 399 + ], + "type": "text", + "content": "To sample a positive example w.r.t the same class, we assume the conditional distribution is a von Mises-Fisher (vMF) distribution [21]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 340, + 407, + 545, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 407, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 340, + 407, + 545, + 426 + ], + "type": "interline_equation", + "content": "p (\\tilde {\\mathbf {z}} = (c, \\tilde {V} _ {*}) | \\mathbf {z} = (c, V _ {*})) = C _ {p} ^ {- 1} e ^ {\\kappa_ {1} \\mathbf {z} ^ {\\top} \\tilde {\\mathbf {z}}} \\qquad (2)", + "image_path": "f0dcc5fc81a049bde0038e5e0d70df79e852920eaadc649fc8029138860a6439.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 434, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 434, + 545, + 460 + ], + "type": "inline_equation", + "content": "C_p = \\int e^{\\kappa_1\\mathbf{z}^\\top \\tilde{\\mathbf{z}}}d\\tilde{\\mathbf{z}}" + }, + { + "bbox": [ + 304, + 434, + 545, + 460 + ], + "type": "text", + "content": " is a normalization constant and " + }, + { + "bbox": [ + 304, + 434, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 434, + 545, + 460 + ], + "type": "text", + "content": " is a concentration parameter." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "text", + "content": "Representation learning is a process that a feature encoder, " + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "inline_equation", + "content": "f(\\cdot):\\mathbb{R}^d\\to \\mathbb{R}^{d_2}" + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "text", + "content": ", extracts representations from two positive observations " + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "inline_equation", + "content": "f(\\tilde{\\mathbf{x}}) = f\\circ g(\\tilde{\\mathbf{z}})" + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}) = f\\circ g(\\mathbf{z})" + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "text", + "content": " and a distribution associated with the encoder " + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "inline_equation", + "content": "h = f\\circ g" + }, + { + "bbox": [ + 304, + 460, + 545, + 518 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 518, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 518, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 361, + 518, + 545, + 536 + ], + "type": "interline_equation", + "content": "q _ {h} (\\tilde {\\mathbf {z}} | \\mathbf {z}) = C _ {q} ^ {- 1} (\\mathbf {z}) e ^ {\\kappa_ {2} h (\\tilde {\\mathbf {z}}) ^ {\\top} h (\\mathbf {z})} \\tag {3}", + "image_path": "3feb5be250a39e7c482ef08445ffd8bb5d84e1c97e26b7387247c12447e13b95.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 540, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 590 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 540, + 545, + 590 + ], + "type": "inline_equation", + "content": "C_q(\\mathbf{z}) = \\int e^{\\kappa_2 h(\\tilde{\\mathbf{z}})^\\top h(\\mathbf{z})} d\\tilde{\\mathbf{z}}" + }, + { + "bbox": [ + 304, + 540, + 545, + 590 + ], + "type": "text", + "content": " be the normalization term. Optimizing any discriminative SSL objective will maximize the similarity between these positive pairs. An example of a well constructed objective is the InfoNCE loss:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 598, + 545, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 545, + 639 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n f o N C E} = \\mathbb {E} \\left[ - l o g \\frac {e ^ {f (\\tilde {\\mathbf {x}}) ^ {\\top} f (\\mathbf {x}) / \\tau}}{e ^ {f (\\tilde {\\mathbf {x}}) ^ {\\top} f (\\mathbf {x}) / \\tau} + \\sum_ {i = 1} ^ {K} e ^ {f \\left(\\mathbf {x} _ {i} ^ {-}\\right) ^ {\\top} f (\\mathbf {x}) / \\tau}} \\right] \\tag {4}", + "image_path": "a08ca7f19ed3cc107787a0b6b433a1c1a1e4cf59e332ac86459bb52277f651be.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": " is a positive example w.r.t. " + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": " in the observation space and " + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^-\\}_{1}^{K}" + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": " samples from distributions of all observations. The global minimum of (4) is reached when" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 683, + 545, + 713 + ], + "type": "text", + "content": "1The mapping of representations on a hypersphere may be different to Barlow Twins methodology, but as shown in [74], normalize representations on a unit sphere also works under Barlow Twins." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16111" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "the cosine similarity between positive pairs is maximized and the cosine similarity between all negative pairs is minimized. In the following section, we will show that discriminative SSL including InfoNCE-driven and non InfoNCE-driven (EMA and Siamese with a predictor) follows strict rules of alignment to maximize the similarity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 224, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 224, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 224, + 166 + ], + "type": "text", + "content": "3.2. Alignment in discriminative SSL" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 288, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 288, + 256 + ], + "type": "text", + "content": "In this section we will combine theories stated in [76, 67, 70] so that the general factors of successful discriminative SSL can be summarized. Additionally, instead of focusing on just InfoNCE SSL variation and content block-identifiability [35, 47], we extend the combined theory to demonstrate that all discriminative SSL benefit from alignment between positive representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 266, + 289, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 289, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 289, + 304 + ], + "type": "text", + "content": "Theorem 3.1 With a data generation process described in 3.1, all discriminative SSL objectives have an alignment loss function between positive pairs from the network:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 312, + 287, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 312, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 112, + 312, + 287, + 327 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\left\\| f (\\mathbf {x}) - f (\\tilde {\\mathbf {x}}) \\right\\| _ {2} ^ {2} \\tag {5}", + "image_path": "8b67b3a7cc8effe3d4294592412f6de73d9d4de4a0691ab330dad60e38050779.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 336, + 287, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 372 + ], + "type": "text", + "content": "This is a weaker statement than Theorem 4.4 in [67] since we only focus on the alignment term. For analysis on the regularization term on the network entropy, see B." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 372, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 420 + ], + "type": "text", + "content": "Proof. For InfoNCE-driven SSL (SimCLR and MoCo), as derived in [68, 76], the InfoNCE loss converges to an alignment term and a uniformity term as the number of negative samples approaches infinity. (See B for full details)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "text", + "content": "For EMA-based SSL methods (BYOL), a predictor " + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "text", + "content": " is associated with the online network so the SSL objective becomes " + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathbb{E}_{(\\mathbf{x},\\tilde{\\mathbf{x}})}\\| p\\circ f(\\mathbf{x},\\theta) - f(\\tilde{\\mathbf{x}},\\xi)\\| _2^2" + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "inline_equation", + "content": "\\xi^t = \\alpha \\xi^{t - 1} + (1 - \\alpha)\\theta^t" + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "text", + "content": " is target network parameter and " + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "text", + "content": " is the online network parameter. Denote " + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "inline_equation", + "content": "p^\\prime = p\\circ f" + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "text", + "content": ". By adding and subtracting " + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "inline_equation", + "content": "p^\\prime (\\tilde{\\mathbf{x}},\\theta)" + }, + { + "bbox": [ + 47, + 421, + 289, + 494 + ], + "type": "text", + "content": " we derive:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 501, + 287, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 501, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 52, + 501, + 287, + 582 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\| p ^ {\\prime} (\\mathbf {x}, \\theta) - p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) + p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - f (\\tilde {\\mathbf {x}}, \\xi) \\| _ {2} ^ {2} \\\\ = \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\| p ^ {\\prime} (\\mathbf {x}, \\theta) - p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) \\| _ {2} ^ {2} + \\underset {\\tilde {\\mathbf {x}}} {\\mathbb {E}} \\| p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - f (\\tilde {\\mathbf {x}}, \\xi) \\| _ {2} ^ {2} \\\\ - 2 \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\left[ \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - p ^ {\\prime} (\\mathbf {x}, \\theta)\\right) ^ {\\top} \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - f (\\tilde {\\mathbf {x}}, \\xi)\\right) \\right] \\tag {6} \\\\ \\end{array}", + "image_path": "84f06ffca76499f5b2b50e996dacecde15454f7a185b8eea0da75d46c5f847bf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x},\\theta)" + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "inline_equation", + "content": "f(\\tilde{\\mathbf{x}},\\xi)" + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "text", + "content": " maps in the same space " + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_2}" + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "text", + "content": " can be considered as a bijective linear transformation within " + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{b_2}" + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "text", + "content": ". In fact, the performance difference between a linear predictor and non-linear predictor is subtle. The global minimizer of (6) must align network output w.r.t to " + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},\\tilde{\\mathbf{x}})" + }, + { + "bbox": [ + 46, + 593, + 287, + 677 + ], + "type": "text", + "content": " with the first term in (6) and align outputs from different networks with second term in (6). Hence completes the proof." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "For Siamese network with a predictor (SimSiam), a similar approach can reformulate the objective as it is a special case when " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "f(\\tilde{\\mathbf{x}},\\xi) = f(\\tilde{\\mathbf{x}},\\theta)" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " and a stop-gradient is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "applied on the " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "f(\\tilde{\\mathbf{x}},\\theta)" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": ". Hence by substituting " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "s_{\\mathcal{G}}(f(\\tilde{\\mathbf{x}},\\theta))" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "f(\\tilde{\\mathbf{x}},\\xi)" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " in (6), we complete the proof by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 105, + 545, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 105, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 105, + 545, + 184 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\| p ^ {\\prime} (\\mathbf {x}, \\theta) - p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) \\| _ {2} ^ {2} \\\\ + \\underset {\\tilde {\\mathbf {x}}} {\\mathbb {E}} \\| p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - \\operatorname {s g} (f (\\tilde {\\mathbf {x}}, \\theta)) \\| _ {2} ^ {2} \\\\ - 2 \\underset {(\\mathbf {x}, \\tilde {\\mathbf {x}})} {\\mathbb {E}} \\left[ \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - p ^ {\\prime} (\\mathbf {x}, \\theta)\\right) ^ {\\top} \\left(p ^ {\\prime} (\\tilde {\\mathbf {x}}, \\theta) - \\operatorname {s g} (f (\\tilde {\\mathbf {x}}, \\theta))\\right) \\right] \\tag {7} \\\\ \\end{array}", + "image_path": "0534870d9ba542191afd146db6769d9ae9fb7d3e1392ad6130092d2227bdfd4f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 194, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 545, + 230 + ], + "type": "text", + "content": "Note that the third term in (6) and (7) can be formulated by the differential entropy of " + }, + { + "bbox": [ + 305, + 194, + 545, + 230 + ], + "type": "inline_equation", + "content": "H(p^{\\prime}(.))" + }, + { + "bbox": [ + 305, + 194, + 545, + 230 + ], + "type": "text", + "content": " hence prevent representation collapse." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 230, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 230, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 230, + 545, + 266 + ], + "type": "text", + "content": "For Barlow Twins, the diagonal of the cross-correlation matrix " + }, + { + "bbox": [ + 305, + 230, + 545, + 266 + ], + "type": "inline_equation", + "content": "C_{ii}" + }, + { + "bbox": [ + 305, + 230, + 545, + 266 + ], + "type": "text", + "content": " is the cosine similarity between positive pairs. Hence completes the proof by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 277, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 277, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 312, + 277, + 545, + 335 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {i} ^ {d _ {2}} \\left(1 - C _ {i i}\\right) ^ {2} = \\sum_ {i} \\left(1 - \\left(f (\\mathbf {x}) ^ {\\top} f (\\tilde {\\mathbf {x}})\\right) _ {i} / \\left(d _ {2} - 1\\right)\\right) ^ {2} (8) \\\\ = \\sum_ {i} \\left(\\| f (\\mathbf {x}) _ {i} - f (\\tilde {\\mathbf {x}}) _ {i} \\| _ {2} ^ {2} / \\left(2 * \\left(d _ {2} - 1\\right)\\right) ^ {2} \\right. (9) \\\\ \\end{array}", + "image_path": "e35c1848b14483fad9ea69f020818bd5e55d94022366d214fa68afcf975b32bb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 346, + 545, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 545, + 371 + ], + "type": "text", + "content": "3.3. Transformation of the ground-truth factors is orthogonal to the applied augmentations" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 376, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 510 + ], + "type": "text", + "content": "By demonstrating that all discriminative SSL have a alignment loss term, the transformation between the ground truth representations and the inferred representations can be derived as in [76]. But different to [76], our derivation of minimization of cross entropy is assumed to be a lower bound since we only include the alignment term and the uniformity term is always positive [14, 5]. However, with all SSL objectives there are additional terms to maximize the output entropy of the model (some described in 3.2). So optimizing SSL objectives as a complete loss function will minimize the cross entropy " + }, + { + "bbox": [ + 304, + 376, + 546, + 510 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[H(p(.|\\mathbf{z}),q_h(.|\\mathbf{z}))]" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "text", + "content": "Theorem 3.2 By considering the generation conditional distribution as " + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\mathbf{z}}|\\mathbf{z}) = C_p^{-1}e^{\\kappa_1\\mathbf{z}^\\top \\tilde{\\mathbf{z}}}" + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "text", + "content": ", the inferred conditional distribution " + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "inline_equation", + "content": "q_h(\\tilde{\\mathbf{z}}|\\mathbf{z})" + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "text", + "content": " can match " + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\mathbf{z}}|\\mathbf{z})" + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\| h(\\mathbf{z}) - h(\\tilde{\\mathbf{z}})\\|_2^2" + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\forall \\mathbf{z}, \\tilde{\\mathbf{z}} : \\kappa \\mathbf{z}^\\top \\tilde{\\mathbf{z}} = h(\\mathbf{z})^\\top \\tilde{\\mathbf{z}}" + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "inline_equation", + "content": "h = f \\circ g" + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "inline_equation", + "content": "h = p \\circ f \\circ g" + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "text", + "content": " maps onto a hypersphere with radius " + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\sqrt{\\kappa_1 / \\kappa_2}" + }, + { + "bbox": [ + 305, + 521, + 547, + 595 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "The proof exactly follows Proposition 1 in [76] just with minor modification on the concentration term " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " so that Theorem 3.2 can apply to non-InfoNCE SSL. A global minimizer of the alignment term " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\| h(\\mathbf{z}) - h(\\tilde{\\mathbf{z}})\\| _2^2" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " will also minimize the cross entropy of between " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\mathbf{z}} |\\mathbf{z})" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " in (2) and " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "q_{h}(\\tilde{\\mathbf{z}} |\\mathbf{z})" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " in (3). This indicates that minimizers of SSL objective alignment maintain the dot product. Then we can use Proposition 2 in [76] directly to show that " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " is an orthogonal linear transformation." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16112" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "Theorem 3.3 Assume the data generation process (cf. 3.1), a model parameterized by " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "h = f \\circ g" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "h = p \\circ f \\circ g" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "h: \\mathbb{R}^{d_1} \\to \\mathbb{R}^{d_2}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": ") that minimizes the alignment term in all discriminative SSL objectives: " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\| h(\\mathbf{z}) - h(\\tilde{\\mathbf{z}}) \\|_2^2" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " as (15), " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " is an orthogonal transformation: " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "h(\\tilde{\\mathbf{z}}) = \\mathbf{A}\\tilde{\\mathbf{z}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " is an orthogonal matrix." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "text", + "content": "The proof follows that a function " + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "text", + "content": " minimizes the alignment will also minimize the cross entropy between the ground truth conditional distribution " + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "inline_equation", + "content": "p(\\tilde{\\mathbf{z}}|\\mathbf{z})" + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "text", + "content": " and the inferred conditional distribution " + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "inline_equation", + "content": "q_{h}(\\tilde{\\mathbf{z}}|\\mathbf{z})" + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "text", + "content": ". Therefore if " + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "text", + "content": " is isometric w.r.t the dot product as indicated in Theorem 3.2 then " + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "inline_equation", + "content": "h(\\tilde{\\mathbf{z}}) = \\mathbf{A}\\tilde{\\mathbf{z}}" + }, + { + "bbox": [ + 47, + 152, + 289, + 224 + ], + "type": "text", + "content": " according to Proposition 2 in [76]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "text", + "content": "Theorem 3.4 Assume augmentations applied during the training can be represented as a change in the ground truth representations, i.e. " + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} = g(\\tilde{\\mathbf{z}}(c,\\tilde{V}*))" + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "text", + "content": " and change in data variables induces a shift in the ground truth representations " + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{z} = (c,\\tilde{V}_*) - (c,V_*)" + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{z}" + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "text", + "content": " is in orthogonal to " + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "text", + "content": " i.e. " + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\delta \\mathbf{z} = 0" + }, + { + "bbox": [ + 47, + 232, + 289, + 304 + ], + "type": "text", + "content": " if a discriminative SSL objective is optimized." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "text", + "content": "Proof. With augmentations normally applied during SSL such as color distortions, rotations, random cropping, and etc., one can view the alteration as a change of a variable in the data variable (color, view angles, sizes). This change of variable under the generation framework described in 3.1 will result in a change in " + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "inline_equation", + "content": "V = \\{V_{i}\\}" + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "text", + "content": " is not changed. Regardless of structure of the causal graph shown in Figure 2, the change in " + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "text", + "content": " can be reflected in the ground truth representation space as " + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}} = \\mathbf{z} + \\delta \\mathbf{z}" + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "text", + "content": ". A global minimizer of any discriminative SSL objective will minimize the alignment term " + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "inline_equation", + "content": "h(\\mathbf{z}) = h(\\tilde{\\mathbf{z}}) = h(\\mathbf{z} + \\delta \\mathbf{z})" + }, + { + "bbox": [ + 46, + 311, + 289, + 453 + ], + "type": "text", + "content": ". According to Theorem 3.3 we derive:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 462, + 287, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 462, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 113, + 462, + 287, + 475 + ], + "type": "interline_equation", + "content": "h (\\mathbf {z}) = \\mathbf {A} \\mathbf {z} = \\mathbf {A} (\\mathbf {z} + \\delta \\mathbf {z}) \\tag {10}", + "image_path": "c6c1a43206a469dd4bd9474e45475496f30bd45800acc7f1f8e45b6a2eaea5d1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 477, + 287, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 477, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 143, + 477, + 287, + 489 + ], + "type": "interline_equation", + "content": "\\mathbf {A} \\delta \\mathbf {z} = 0 \\tag {11}", + "image_path": "09ac0d2114b1ed959cd001c1603720fd06924297f9137de82c8b744d31d8f30f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 497, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 287, + 534 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 497, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 47, + 497, + 287, + 534 + ], + "type": "text", + "content": " being a linear orthogonal matrix. This indicates that the transformation " + }, + { + "bbox": [ + 47, + 497, + 287, + 534 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 497, + 287, + 534 + ], + "type": "text", + "content": " is learned to annul the change in data variables or the effect of augmentations applied." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 540, + 287, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 287, + 564 + ], + "type": "text", + "content": "3.4. Reason and solutions for unstable change in data variables" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "In section 3.3 we show that generalized representation is robust to augmentations since " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", the linear transformation between the ground truth representation and inferred representations, is orthogonal to augmentation applied during the training. And if a change in the data variable reflects a shift, " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{z}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", in the ground truth representations and the newly inferred representation is stable, meaning " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{z}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " will be absorbed by the transformation matrix " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", then the shift in the ground truth representation corresponds to an augmentation that is applied during the training. However, when " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{z}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " appears in the range of " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\delta \\mathbf{z} \\neq 0" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": ", the resultant inferred representation can be unstable and lead to performance drop" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "on downstream models (denote as " + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "D(\\mathbf{z})" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "). We quantify this deterioration on " + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 342, + 105, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 105, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 342, + 105, + 545, + 118 + ], + "type": "interline_equation", + "content": "m \\left(D (h (\\mathbf {z})) _ {\\text {s t a b l e}}\\right) - m \\left(D \\left(h (\\tilde {\\mathbf {z}}) _ {\\text {u n s t a b l e}}\\right) \\right. \\tag {12}", + "image_path": "715938490bc52b53b3c1c11377cb21f9819c31d8b792fd845793acb466354ace.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "inline_equation", + "content": "m(.)" + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": " is a metric on an outcome of the downstream task. In example of " + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": " being a linear classifier, " + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "inline_equation", + "content": "m(.)" + }, + { + "bbox": [ + 304, + 127, + 547, + 212 + ], + "type": "text", + "content": " can be the probability of predicting the target class (prediction score) or the proportion of correct predictions (accuracy). In order to overcome this deterioration, we propose two methods, namely Robust Dimensions and Stable Inference Mapping:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": "1. Robust Dimensions: Under a stable condition, " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "D(f(\\mathbf{x}))^2 = D(\\mathbf{A}\\mathbf{z})" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": ", each dimension in the inferred representation " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " is a linear combination of dimensions of the ground truth representation. The dimensions contributing most to " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "D()" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " should be robust to unstable shift " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{z}_{\\text{unstable}}" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": ". In other words, most robust dimensions of " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " should be also robust in " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "f(\\tilde{\\mathbf{x}})" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} = g(\\mathbf{z} + \\delta \\mathbf{z}_{\\text{unstable}})" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " as some dimensions of " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\delta \\mathbf{z}_{\\text{unstable}}" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " will be zero. Hence identifying most important dimensions in " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})_{\\text{stable}}" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " and pass through the same dimensions of " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "f(\\tilde{\\mathbf{x}})_{\\text{unstable}}" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " should alleviate the deterioration by making " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "m(D(h(\\mathbf{z}))_{\\text{stable}})_{\\{\\dim\\}} = m(D(h(\\tilde{\\mathbf{z}})_{\\text{unstable}}))_{\\{\\dim\\}}" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\{\\dim\\}" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " is a set of most robust dimensions. In example of " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " being a linear classifier, the contribution of each dimension can be calculated by " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "W_c^\\top f(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "W_c^\\top" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": " is the Jacobian of the linear classifier w.r.t target class " + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 218, + 547, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": "2. Stable Inference Mapping: Since " + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})_{\\text{stable}} = -f(\\tilde{\\mathbf{x}})_{\\text{unstable}} = -\\mathbf{A} \\delta \\mathbf{z}" + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": ", we can learn another linear transformation " + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": " to absorb " + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{z}" + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": ". Especially, we want to learn " + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{F} f(\\tilde{\\mathbf{z}})_{\\text{unstable}} = \\mathbf{F} \\mathbf{A} \\mathbf{z} + \\mathbf{F} \\mathbf{A} \\delta \\mathbf{z}" + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": " such that the additional " + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": " will not only set " + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\delta z" + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": " to 0, but also make stable representation more robust to the unstable shift by assuring " + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\mathbf{A}" + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": " orthogonal to " + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{z}_{\\text{unstable}}" + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": " in addition to the augmentations applied during training. Formally, we model a linear layer " + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "inline_equation", + "content": "l(f(\\tilde{\\mathbf{x}})_{\\text{unstable}}) = f(\\mathbf{x})_{\\text{stable}} - f(\\tilde{\\mathbf{x}})_{\\text{unstable}}" + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": ", hence during the inference " + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}) + l(f(\\mathbf{x}))" + }, + { + "bbox": [ + 313, + 429, + 547, + 562 + ], + "type": "text", + "content": " is used for downstream task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": "Relation to Causal Inference Since we only have observations in the image space, we consider the augmentations or changes of a data variable as interventions on the ground truth representations: " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} = g(\\tilde{\\mathbf{z}} |(c,do(V_i = v_i))" + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": ". With an access to ground truth representation, we can evaluate treatment-control effect i.e. " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "Pr(f(g(\\tilde{\\mathbf{z}} |(c,do(V_i = v_{stable}))) - Pr(f(g(\\tilde{\\mathbf{z}} |(c,do(V_i = v_{unstable})))" + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "text", + "content": ". Without any access to the ground truth representations, we can evaluate the average treatment effect [33] by " + }, + { + "bbox": [ + 304, + 568, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[D(f(\\mathbf{x})_{stable})-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 683, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 683, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 683, + 547, + 713 + ], + "type": "text", + "content": "2Note that in general " + }, + { + "bbox": [ + 304, + 683, + 547, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 683, + 547, + 713 + ], + "type": "text", + "content": " contains a projector. However we exclude the notion of the projector to simplify the problem and [10, 29, 26] show that a projector is not necessary in SSL." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16113" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "D(f(\\tilde{\\mathbf{x}})_{\\text{unstable}})]" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " via synthesizing manual data samples " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i = g(\\tilde{\\mathbf{z}} | (c, \\text{do} (V_i = v_i))" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 109, + 212, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 109, + 212, + 123 + ], + "spans": [ + { + "bbox": [ + 47, + 109, + 212, + 123 + ], + "type": "text", + "content": "4. Experiments and Discussions" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 130, + 287, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 130, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 130, + 287, + 177 + ], + "type": "text", + "content": "In this section we evaluate our solutions to unstable shift in data variable on two datasets: Causal3DIdent and ImageNet pretrained SSL and corresponding linear classifiers as the downstream task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 187, + 141, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 187, + 141, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 187, + 141, + 199 + ], + "type": "text", + "content": "4.1. Causal3DIdent" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 94, + 213, + 241, + 286 + ], + "blocks": [ + { + "bbox": [ + 94, + 213, + 241, + 286 + ], + "lines": [ + { + "bbox": [ + 94, + 213, + 241, + 286 + ], + "spans": [ + { + "bbox": [ + 94, + 213, + 241, + 286 + ], + "type": "image", + "image_path": "eec2275428e8faf79fdde2e91d218a79d519c3d158d53b9f7ff570a4c454a55e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 294, + 263, + 306 + ], + "lines": [ + { + "bbox": [ + 69, + 294, + 263, + 306 + ], + "spans": [ + { + "bbox": [ + 69, + 294, + 263, + 306 + ], + "type": "text", + "content": "Figure 3: Causal graph for " + }, + { + "bbox": [ + 69, + 294, + 263, + 306 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 69, + 294, + 263, + 306 + ], + "type": "text", + "content": " in Causal3DIdent." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "text", + "content": "[76] develops the dataset that 7 objects in the dataset that each " + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "text", + "content": " image is associated with an object class and a 10-dimensional latent representation. These 10 dimensions correspond to 10 data variables. A causal graph imposed on these variables is shown in Figure 3. As the dependency shown in Figure 3, an object is placed at " + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "inline_equation", + "content": "pos_{obj} = (x,y,z)" + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "inline_equation", + "content": "rot_{obj} = (\\phi, \\theta, \\psi)" + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "inline_equation", + "content": "hue_{o}bj" + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "text", + "content": " under a spot light at an angle " + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "inline_equation", + "content": "pos_{spl}" + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "text", + "content": " with color " + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "inline_equation", + "content": "hue_{spl}" + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "text", + "content": " on a background with color " + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "inline_equation", + "content": "hue_{bg}" + }, + { + "bbox": [ + 46, + 323, + 287, + 430 + ], + "type": "text", + "content": ". Detailed information in C." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 432, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 563 + ], + "type": "text", + "content": "To simulate unstable change of variable that is not accessible during the training, a range of some data variables is hold out and hence this portion of data is treated as potential unstable samples that causes the performance drop on the linear classifier. Since all data variables are truncated in range [-1,1], we hold out the edge value(s) to further portray 'unexpected' values during the inference time. We select 8 dimensions to intervene, namely: " + }, + { + "bbox": [ + 46, + 432, + 287, + 563 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 432, + 287, + 563 + ], + "type": "text", + "content": " in object position3, all 3 object rotation angles, spot light position, and all 3 hue variables. Both training and testing data are sampled and for detailed sampling procedures refer to C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 563, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 683 + ], + "type": "text", + "content": "SSL Experiment Setup ResNet18[30] is the backbone of the feature encoder " + }, + { + "bbox": [ + 46, + 563, + 287, + 683 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 563, + 287, + 683 + ], + "type": "text", + "content": ". Same augmentations in [10] are applied during training. An Adam optimizer with a learning rate at 0.0001 and a weight decay at 0.00001 is optimized for all discriminative SSL. Hyperparameters for each SSL is presented in C. The dimension of the inferred representation space is set to 128. The network is trained for 20 epochs on intervened data. Then a linear classifier is trained on the frozen representations of the network with a same optimizer for 10 epochs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "text", + "content": "To illustrate the simulation results in the deterioration of the downstream performance, for each testing data that has the same data variable distribution as the training data, we search in the ground truth representation space for 5 nearest representations in the hold-out test distribution when " + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "text", + "content": " dimension of the testing data is shifted to a random value in the hold-out range of corresponding data variable. Within the corresponding 5 images representing a shift in " + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "text", + "content": " dimensions that is not seen during the training, we record the lowest performance to fulfill the deterioration scenario. In Figure 4 the score and accuracy are averaged over all combinations of " + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "inline_equation", + "content": "\\binom{8}{n}" + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "inline_equation", + "content": "n = 0" + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "text", + "content": ", this indicates the performance of SSL on the testing data with the same data variable distribution as the training data. The performance is exceptional since hold-out values of some data variables may be covered by the augmentations applied (" + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 306, + 72, + 545, + 442 + ], + "type": "text", + "content": " covered by random cop, and hue colors covered by color distortions). A will be orthogonal to changes in these data variables. In fact in C we show that the performance difference between seen distribution and unseen distribution in testing data is comparatively small. As illustrated in Figure 4, even when only one variable is changed to a unseen value, there is a large drop in both accuracy and prediction score (20% in accuracy and 30% in prediction score for SimCLR). With more tangled changes in data variables, the unstable representations result in poorer downstream performance. Since the selection of the data variables may be too complex due to the dependency, we also validate the same issue on selecting only children nodes in Figure 3. And also we visualize the latent shift between stable and unstable examples. See C for more results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": "Robust Dimensions For each pair of testing data " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text{stable}}" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " with seen data variable distribution and the selected data " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text{unstable}}" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " among the 5 nearest neighbours when changing " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " dimensions to unseen distribution values, we apply the Jacobian of the linear classifier w.r.t the target class " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "W_{c}^{\\top}" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " on the stable representation " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})_{\\text{stable}}" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " to identify the top " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " most important dimensions and pass the same dimensions of " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "f(\\tilde{\\mathbf{x}})_{\\text{unstable}}" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " to evaluate the performance. As shown in Figure 5, the accuracy only deteriorate slightly when top " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " most important features of " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "f(\\tilde{\\mathbf{x}})_{\\text{unstable}}" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " are selected for the downstream task. This is true even when all 8 variables are shifted. This suggests that " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " is orthogonal to changes in ground truth representation " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " in most dimensions. This high percentage of dimensions may be due to " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " optimizing the SSL objectives to a high level and the augmentations applied covers some of the hold-out variable values. Interestingly, there are cases where passing the top " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": " (around " + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": ") important dimensions results in higher performance than stable representations (uninterrupted in Figure 5). However, as expected, including more less important dimensions where the unseen shift in the data variable results in non-zero adjustment (" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\delta \\mathbf{z} \\neq 0" + }, + { + "bbox": [ + 305, + 450, + 545, + 713 + ], + "type": "text", + "content": ") initiates the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": " Only " + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": " is altered because most of deep vision models are translation invariant." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16114" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 296, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 296, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 296, + 148 + ], + "type": "image", + "image_path": "21cc225abf728803edbfd65cec6a68b81ee71a8a6786f030d15c49190fc88e6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 152, + 231, + 163 + ], + "lines": [ + { + "bbox": [ + 123, + 152, + 231, + 163 + ], + "spans": [ + { + "bbox": [ + 123, + 152, + 231, + 163 + ], + "type": "text", + "content": "(a) Deterioration in Accuracy" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 298, + 71, + 536, + 148 + ], + "blocks": [ + { + "bbox": [ + 298, + 71, + 536, + 148 + ], + "lines": [ + { + "bbox": [ + 298, + 71, + 536, + 148 + ], + "spans": [ + { + "bbox": [ + 298, + 71, + 536, + 148 + ], + "type": "image", + "image_path": "e2fb9f3265c94903183cf9b34812e200e1b9ef60bae532cf6d45e99763d04480.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 152, + 464, + 163 + ], + "lines": [ + { + "bbox": [ + 369, + 152, + 464, + 163 + ], + "spans": [ + { + "bbox": [ + 369, + 152, + 464, + 163 + ], + "type": "text", + "content": "(b) Deterioration in Score" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 173, + 547, + 197 + ], + "lines": [ + { + "bbox": [ + 46, + 173, + 547, + 197 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 547, + 197 + ], + "type": "text", + "content": "Figure 4: Deterioration of unstable changing in data variables on all SSL. The more unexpected changes occur in the data variables, the more severe the deterioration." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 218, + 205, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 205, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 205, + 231 + ], + "type": "text", + "content": "deterioration phenomenon we observe." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 48, + 251, + 298, + 351 + ], + "blocks": [ + { + "bbox": [ + 48, + 251, + 298, + 351 + ], + "lines": [ + { + "bbox": [ + 48, + 251, + 298, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 298, + 351 + ], + "type": "image", + "image_path": "9c1ed25508b7d95acd0189f359673b4a542b7f92f937c9ba9d4059d09b8521cb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 357, + 228, + 366 + ], + "lines": [ + { + "bbox": [ + 118, + 357, + 228, + 366 + ], + "spans": [ + { + "bbox": [ + 118, + 357, + 228, + 366 + ], + "type": "text", + "content": "(a) Improvement on Accuracy" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 369, + 296, + 466 + ], + "blocks": [ + { + "bbox": [ + 50, + 369, + 296, + 466 + ], + "lines": [ + { + "bbox": [ + 50, + 369, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 50, + 369, + 296, + 466 + ], + "type": "image", + "image_path": "d6199777575f6d3f01cb0c27e3d47513d3562264b32832d64ca7ccd7d152bd88.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 472, + 221, + 482 + ], + "lines": [ + { + "bbox": [ + 124, + 472, + 221, + 482 + ], + "spans": [ + { + "bbox": [ + 124, + 472, + 221, + 482 + ], + "type": "text", + "content": "(b) Improvement on Score" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "text", + "content": "Figure 5: By identifying and passing top " + }, + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 46, + 491, + 287, + 528 + ], + "type": "text", + "content": " important dimensions, the deterioration in both accuracy and prediction score is significantly alleviated." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "text", + "content": "Stable Inference Mapping As shown in 3.4, a linear transformation " + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{F}:\\mathbb{R}^{d_2}\\to \\mathbb{R}^{d_2}" + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "text", + "content": " is trained to cancel the effect " + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\delta \\mathbf{z}" + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "text", + "content": " and also further improve the robustness of transformation on the ground truth representation " + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{Az}" + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "text", + "content": ". In this experiment, we match each training data with a random member in the 5 nearest representations when only one dimension is changed to a unstable value. A linear layer is trained with the same optimizer for 10 epochs on the training pairs. In Table 1, all accuracy for unstable examples increases significantly (except SimSiam) after learning " + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "text", + "content": ". However, since the " + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 557, + 289, + 712 + ], + "type": "text", + "content": " is very close to the global minimizer of the alignment term, the improvement on stable examples cannot be observed." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 320, + 217, + 533, + 305 + ], + "blocks": [ + { + "bbox": [ + 320, + 217, + 533, + 305 + ], + "lines": [ + { + "bbox": [ + 320, + 217, + 533, + 305 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 533, + 305 + ], + "type": "table", + "html": "
xstablex unstable
w/o Fw/ Fw/o Fw/ F
SimCLR0.9960.9980.8330.889
MoCo0.9920.9920.8000.849
BYOL0.9960.9980.8860.920
SimSiam0.9980.9990.9190.928
Barlow Twins0.9910.9950.8180.862
", + "image_path": "3dc36aef3ab7feb50c4307ca308b89f1cfd534cbe4f182666557cb2f111af614.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "content": "Table 1: The effect of " + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "content": " on both stable and unstable samples. The accuracy is average over 3 random seeds." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 356, + 351, + 497, + 531 + ], + "blocks": [ + { + "bbox": [ + 356, + 351, + 497, + 531 + ], + "lines": [ + { + "bbox": [ + 356, + 351, + 497, + 531 + ], + "spans": [ + { + "bbox": [ + 356, + 351, + 497, + 531 + ], + "type": "image", + "image_path": "a06d872cf06ffeae1ff707465a95603fe0cb65a1b43151084ed33d857fed1a38.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 541, + 545, + 565 + ], + "lines": [ + { + "bbox": [ + 306, + 541, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 541, + 545, + 565 + ], + "type": "text", + "content": "Figure 6: Examples of ObjectNet dataset " + }, + { + "bbox": [ + 306, + 541, + 545, + 565 + ], + "type": "inline_equation", + "content": "(1^{st}" + }, + { + "bbox": [ + 306, + 541, + 545, + 565 + ], + "type": "text", + "content": " row), Stylized ImageNet dataset " + }, + { + "bbox": [ + 306, + 541, + 545, + 565 + ], + "type": "inline_equation", + "content": "(2^{nd}" + }, + { + "bbox": [ + 306, + 541, + 545, + 565 + ], + "type": "text", + "content": " row), Synthetic dataset " + }, + { + "bbox": [ + 306, + 541, + 545, + 565 + ], + "type": "inline_equation", + "content": "(3^{rd}" + }, + { + "bbox": [ + 306, + 541, + 545, + 565 + ], + "type": "text", + "content": " row)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 586, + 374, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 374, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 374, + 600 + ], + "type": "text", + "content": "4.2. ImageNet" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 666 + ], + "type": "text", + "content": "To validate our findings on a larger scale and more realistic settings, we apply proposed solutions on ImageNet[16] with various altered and synthetic datasets as unstable shift in the data variable. See Figure 6 and D for more information on ObjectNet, Stylized-ImageNet, Synthetic dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "ObjectNet[3] is a collection of objects that are intentionally placed at an unusual view angle and backgrounds, so that the bias learned by the model with usual data variables is more prominent when testing on ObjectNet. With focus-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16115" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "content": "ing on 113 classes overlapping with ImageNet classes, we evaluate our second proposed solution explicitly on ObjectNet since the shift presented in ObjectNet is very unstable and it is very challenging to overcome the negative effect of the shift." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 193 + ], + "type": "text", + "content": "Stylized ImageNet[24] change the style of original ImageNet image to a random artistic style. With this drastic shift in the data variable, the analysis of the first solution is more insightful on the robustness of dimensions of " + }, + { + "bbox": [ + 46, + 133, + 287, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 133, + 287, + 193 + ], + "type": "text", + "content": " on a very different shift in data variable." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 315 + ], + "type": "text", + "content": "Synthetic Data follows synthetic procedure in [18] where object is masked on a background at a location with a rotation angle. We explore the benefit of this synthesized dataset at three modes: background, location, rotation where the target variable is randomly sampled with the other two variables fixed. Additionally, we explore texture as an independent variable by masking 'texturized' objects in [23]. We set total number of updating steps per epoch as 100 with batch size 256. This means we select total number of 512000 synthesized images every epoch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 376 + ], + "type": "text", + "content": "Experiment Setup ResNet 50 pretrained on ImageNet and a linear classifier fintuned via SimCLR, BYOL, and SimSiam are tested with both proposed solutions. For Stable Inference Mapping, a linear layer is optimized with an Adam optimizer for 10 epochs on the synthetic dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": "Robust Dimensions For each of ImageNet validation data sample, we stylized the image to a random artistic fashion. We observe the dramatic performance difference between the ImageNet stable images and Stylized unstable ImageNet images " + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_{stable}, \\mathbf{x}_{unstable})" + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": ". The result is shown in Figure 7. For SimCLR, passing the top " + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 377, + 287, + 520 + ], + "type": "text", + "content": " important dimensions can close a small performance gap between stable representations and unstable representations. Nonetheless, all SSL seem to be sensible to the strong style change as they shorten the difference between ImageNet to an modest extent. This is as expected since the Stylized ImageNet changes multiple variables to an extreme value." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "Stable Inference Mapping As described in Synthetic Data, we explore the benefit of our second proposed method to evaluate on a dataset that most of samples are unstable according to learned " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": ". At each training step, with other variables randomly fixed, 10 images with random target variable values are generated. The pair with maximum " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "m(D(\\mathbf{x})) - m(D(\\tilde{\\mathbf{x}}))" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " is selected to train the linear transformation. In Table 2, inferring " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " via controlling location produces least improvement. This is expected since the network is robust to translation by design. While background, rotation, and texture improves the performance considerably with the consideration on the training time. However, in D we show that training the model longer using Stable Inference Mapping yield less favorable results since the improvement is less significant and starts to saturate at around 30 epochs." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 545, + 160 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 160 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 160 + ], + "type": "image", + "image_path": "0ce5deaa8597f4d839b7edf29964ce209f5b1c963cba81d281dbb1832ab57ae0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 178, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 372, + 166, + 482, + 177 + ], + "lines": [ + { + "bbox": [ + 372, + 166, + 482, + 177 + ], + "spans": [ + { + "bbox": [ + 372, + 166, + 482, + 177 + ], + "type": "text", + "content": "(a) Improvement on Accuracy" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 178, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 307, + 178, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 178, + 545, + 266 + ], + "type": "image", + "image_path": "ef24a24ca9ebf8c9b8309833600ec769da956d94d114a591c3b9ed67c0842140.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 273, + 474, + 283 + ], + "lines": [ + { + "bbox": [ + 378, + 273, + 474, + 283 + ], + "spans": [ + { + "bbox": [ + 378, + 273, + 474, + 283 + ], + "type": "text", + "content": "(b) Improvement on Score" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 293, + 545, + 328 + ], + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 328 + ], + "type": "text", + "content": "Figure 7: By identifying and passing top " + }, + { + "bbox": [ + 305, + 293, + 545, + 328 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 305, + 293, + 545, + 328 + ], + "type": "text", + "content": " important dimensions, the deterioration in both accuracy and prediction score is alleviated." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 328, + 343, + 525, + 420 + ], + "blocks": [ + { + "bbox": [ + 328, + 343, + 525, + 420 + ], + "lines": [ + { + "bbox": [ + 328, + 343, + 525, + 420 + ], + "spans": [ + { + "bbox": [ + 328, + 343, + 525, + 420 + ], + "type": "table", + "html": "
SimCLRBYOLSimSiam
w/o F10.1214.0411.15
background12.3415.4312.45
location10.3514.1311.17
rotation12.0415.8812.38
texture12.7915.0613.11
", + "image_path": "8e40c42063bd03e289ada23d88398d34c0518222c5b8f86437faee2ea22db0de.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 429, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 305, + 429, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 545, + 453 + ], + "type": "text", + "content": "Table 2: Comparison of inferring different " + }, + { + "bbox": [ + 305, + 429, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 305, + 429, + 545, + 453 + ], + "type": "text", + "content": " on ObjectNet with a target data variable random trials." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 476, + 381, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 381, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 381, + 489 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Though we identify the root cause of unstable inference for discriminative SSL by constructing a causal framework inspired by the prior work, the proposed solutions are constrained and limited to be applied on realistic applications. Robust Dimensions involves establishing a correspondence between stable and unstable instances on a one-to-one basis, enabling the identification of dimensions contributing to stability. On the other hand, Stable Inference Mapping necessitates a collection of unstable instances with a specific alteration in a particular group of data variables. Within the Causal3dIdent dataset, both solutions can be assessed using the same unstable instances. In more realistic datasets, achieving a one-to-one correspondence is feasible, and manipulation of one group of data variables can be accomplished using synthetic data. However, any assessments with involving artificially generated images might introduce some level of uncertainty. In a realistic setup, since training samples are not directly observ-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16116" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 133 + ], + "type": "text", + "content": "able during the inference stage, simple interventions on inference samples may not effectively separate the unstable variables from the stable ones. Consequently, the potential benefits of the proposed solutions in realistic datasets are undermined." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 144, + 125, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 144, + 125, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 125, + 156 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 289, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 289, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 289, + 369 + ], + "type": "text", + "content": "In conclusion, this paper has proposed a novel approach to address the issue of unstable behavior during the inference stage in SSL methods. By building on the previous theories of successful InfoNCE-facilitated contrastive SSL and extending it to recent SSL methods, we have demonstrated that a change in the data factor can result in a shift in the inferred representation, leading to a decline in downstream performance. We have proposed learning targeted transformations that regularize the violating shift and restore performance on the unseen data shift. Our experiments on both controlled and realistic datasets have shown the efficacy of our proposed solutions. These contributions provide a better understanding of SSL methods and offer a promising solution to the problem of unstable behavior during the inference stage. We hope that our work will inspire further research in this area and lead to improved SSL methods that are more robust to changes in data factors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 379, + 162, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 379, + 162, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 379, + 162, + 392 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 399, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 287, + 422 + ], + "type": "text", + "content": "The work was funded in part by NSF CNS-2112562 and IIS-2140247." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 434, + 107, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 107, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 107, + 447 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 454, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 454, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 498 + ], + "type": "text", + "content": "[1] Ibrahim Ahmad and Pi-Erh Lin. A nonparametric estimation of the entropy for absolutely continuous distributions (corresp.). IEEE Transactions on Information Theory, 22(3):372-375, 1976. 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 499, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 287, + 544 + ], + "type": "text", + "content": "[2] Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 544, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 610 + ], + "type": "text", + "content": "[3] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 612, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 612, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 53, + 612, + 287, + 655 + ], + "type": "text", + "content": "[4] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 657, + 287, + 689 + ], + "type": "text", + "content": "[5] Sergiy V Borodachov, Douglas P Hardin, and Edward B Saff. Discrete energy on rectifiable sets, volume 3. Springer, 2019. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 714 + ], + "type": "text", + "content": "[6] Yue Cao, Mingsheng Long, and Jianmin Wang. Unsupervised domain adaptation with distribution matching ma" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "chines. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 95, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 95, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 95, + 547, + 148 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 150, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 150, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 150, + 545, + 205 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "type": "text", + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 250, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 303 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 5, 6, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "type": "text", + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[12] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640-9649, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 394, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 436 + ], + "type": "text", + "content": "[13] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 17" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 437, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 470 + ], + "type": "text", + "content": "[14] Henry Cohn and Abhinav Kumar. Universally optimal distribution of points on spheres. Journal of the American Mathematical Society, 20(1):99-148, 2007. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 491 + ], + "type": "text", + "content": "[15] Hal Daumé III. Frustratingly easy domain adaptation. arXiv preprint arXiv:0907.1815, 2009. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 569 + ], + "type": "text", + "content": "[17] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516, 2014. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 571, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 635 + ], + "type": "text", + "content": "[18] Josip Djolonga, Jessica Yung, Michael Tschannen, Rob Romijnders, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Matthias Minderer, Alexander D'Amour, Dan Moldovan, Sylvain Gelly, Neil Houlsby, Xiaohua Zhai, and Mario Lucic. On robustness and transferability of convolutional neural networks, 2020. 8, 16" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "text", + "content": "[19] Lixin Duan, Dong Xu, and Ivor Tsang. Learning with augmented features for heterogeneous domain adaptation. arXiv preprint arXiv:1206.4660, 2012. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[20] Deniz Erdogmus, Kenneth E Hild, Yadunandana N Rao, and Jose C Principe. Minimax mutual information approach for independent component analysis. Neural Computation, 16(6):1235-1252, 2004. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16117" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 106 + ], + "type": "text", + "content": "[21] Ronald Aylmer Fisher. Dispersion on a sphere. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 217(1130):295-305, 1953. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "text", + "content": "[22] Bernhard C. Geiger and Gernot Kubin. On the information loss in memoryless systems: The multivariate case. CoRR, abs/1109.4856, 2011. 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 206 + ], + "type": "text", + "content": "[23] Robert Geirhos, Claudio Michaelis, Felix A. Wichmann, Patricia Rubisch, Matthias Bethge, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. 7th International Conference on Learning Representations, ICLR 2019, (c):1-22, 2019. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[24] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[25] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 364 + ], + "type": "text", + "content": "[26] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "text", + "content": "[27] Jeff Z. HaoChen, Colin Wei, Adrien Gaidon, and Tengyu Ma. Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Loss. Advances in Neural Information Processing Systems, 7(NeurIPS):5000-5011, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "type": "text", + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 466, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 520 + ], + "type": "text", + "content": "[29] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "text", + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "text", + "content": "[31] Geoffrey E Hinton and Richard Zemel. Autoencoders, minimum description length and helmholtz free energy. Advances in neural information processing systems, 6, 1993. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "text", + "content": "[32] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[33] Paul W Holland. Statistics and causal inference. Journal of the American statistical Association, 81(396):945-960, 1986. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[34] Tzu Ming Harry Hsu, Wei Yu Chen, Cheng-An Hou, Yao-Hung Hubert Tsai, Yi-Ren Yeh, and Yu-Chiang Frank Wang." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Unsupervised domain adaptation with imbalanced cross-domain data. In Proceedings of the IEEE International Conference on Computer Vision, pages 4121-4129, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[35] Aapo Hyvarinen and Patrik Hoyer. Emergence of phase-and shift-invariant features by decomposition of natural images into independent feature subspaces. Neural computation, 12(7):1705-1720, 2000. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[36] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and Nonlinear ICA. Advances in Neural Information Processing Systems, (Nips):3772-3780, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "type": "text", + "content": "[37] Aapo Hyvarinen and Hiroshi Morioka. Nonlinear ICA of temporally dependent stationary sources. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 274 + ], + "type": "text", + "content": "[38] Aapo Hyvarinen and Petteri Pajunen. Nonlinear independent component analysis: Existence and uniqueness results. Neural networks, 12(3):429-439, 1999. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "type": "text", + "content": "[39] Aapo Hyvarinen, Hiroaki Sasaki, and Richard E. Turner. Nonlinear ICA using auxiliary variables and generalized contrastive learning. AISTATS 2019 - 22nd International Conference on Artificial Intelligence and Statistics, 89, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 333, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 375 + ], + "type": "text", + "content": "[40] Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Banerjee, and Fillia Makedon. A survey on contrastive self-supervised learning. Technologies, 9(1):2, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "type": "text", + "content": "[41] Longlong Jing and Yingli Tian. Self-supervised visual feature learning with deep neural networks: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(11):4037-4058, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "text", + "content": "[42] Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann. Contrastive adaptation network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4893-4902, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 479, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 511 + ], + "type": "text", + "content": "[43] Ilyes Khemakhem, Diederik P. Kingma, Ricardo Pio Monti, and Aapo Hyvarinen. Variational Autoencoders and Nonlinear ICA: A Unifying Framework. 108(2015), 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 514, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 566 + ], + "type": "text", + "content": "[44] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[45] David Klindt, Lukas Schott, Yash Sharma, Ivan Ustyuzhaninov, Wieland Brendel, Matthias Bethge, and Dylan Paiton. Towards Nonlinear Disentanglement in Natural Data with Temporal Sparse Coding. pages 1-51, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 689 + ], + "type": "text", + "content": "[46] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 16" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[47] Quoc V Le, Will Y Zou, Serena Y Yeung, and Andrew Y Ng. Learning hierarchical invariant spatio-temporal features" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16118" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "for action recognition with independent subspace analysis. In CVPR 2011, pages 3361-3368. IEEE, 2011. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[48] Jason D. Lee, Qi Lei, Nikunj Saunshi, and Jiacheng Zhuo. Predicting What You Already Know Helps: Provable Self-Supervised Learning. Advances in Neural Information Processing Systems, 1(NeurIPS):309-323, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[49] Chunyuan Li, Jianwei Yang, Pengchuan Zhang, Mei Gao, Bin Xiao, Xiyang Dai, Lu Yuan, and Jianfeng Gao. Efficient Self-supervised Vision Transformers for Representation Learning. 3:1-27, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "text", + "content": "[50] Wen Li, Lixin Duan, Dong Xu, and Ivor W Tsang. Learning with augmented features for supervised and semi-supervised heterogeneous domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 36(6):1134-1148, 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 286, + 262 + ], + "type": "text", + "content": "[51] Ralph Linsker. Self-organization in a perceptual network. Computer, 21(3):105–117, 1988. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[52] Shuo Liu, Adria Mallol-Ragolta, Emilia Parada-Cabaleiro, Kun Qian, Xin Jing, Alexander Kathan, Bin Hu, and Bjöern W Schuller. Audio self-supervised learning: A survey. Patterns, 3(12):100616, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 353 + ], + "type": "text", + "content": "[53] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering, 35(1):857-876, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[54] Yixin Liu, Ming Jin, Shirui Pan, Chuan Zhou, Yu Zheng, Feng Xia, and Philip Yu. Graph self-supervised learning: A survey. IEEE Transactions on Knowledge and Data Engineering, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[55] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer joint matching for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1410-1417, 2014. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "text", + "content": "[56] Jovana Mitrovic, Brian McWilliams, Jacob Walker, Lars Buesing, and Charles Blundell. Representation Learning via Invariant Causal Mechanisms. pages 1-21, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "text", + "content": "[57] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI, pages 69-84. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 567 + ], + "type": "text", + "content": "[58] Aaron Van Den Oord. Representation Learning with Contrastive Predictive Coding. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "text", + "content": "[59] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 644 + ], + "type": "text", + "content": "[60] Judea Pearl. Causality: Models, reasoning, and inference, second edition. 2011. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[61] Madeline C Schiappa, Yogesh S Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ACM Computing Surveys, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[62] Saeed Shurrab and Rehab Duwairi. Self-supervised learning methods and applications in medical imaging analysis: A survey. PeerJ Computer Science, 8:e1045, 2022. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[63] Mamatha Thota and Georgios Leontidis. Contrastive domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2209-2218, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[64] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 15" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "text", + "content": "[65] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 262 + ], + "type": "text", + "content": "[66] Riccardo Volpi, Pietro Morerio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5495-5504, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[67] Kügelgen Von, Julius and, Wieland Brendel, Bernhard Schölkopf, Michel Besserve, and Francesco Locatello. Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style Luigi Gresele. Yash Sharma, 3(NeurIPS):4, 2021. 2, 3, 4, 15" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 321, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 352 + ], + "type": "text", + "content": "[68] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 398 + ], + "type": "text", + "content": "[69] Rui Wang, Zuxuan Wu, Zejia Weng, Jingjing Chen, Guo-Jun Qi, and Yu-Gang Jiang. Cross-domain contrastive learning for unsupervised domain adaptation. IEEE Transactions on Multimedia, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "text", + "content": "[70] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. 37th International Conference on Machine Learning, ICML 2020, PartF168147-13:9871-9881, 2020. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[71] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668-14678, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 513, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 545 + ], + "type": "text", + "content": "[72] Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-Supervised Learning with Swin Transformers. pages 1-8, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 547, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 547, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 547, + 545, + 621 + ], + "type": "text", + "content": "[73] Detai Xin, Tatsuya Komatsu, Shinnosuke Takamichi, and Hiroshi Saruwatari. Disentangled speaker and language representations using mutual information minimization and domain adaptation for cross-lingual tts. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6608-6612. IEEE, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 677 + ], + "type": "text", + "content": "[74] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 1, 2, 3, 14, 15" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[75] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, Octo" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16119" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 138 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "ber 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 138 + ], + "type": "text", + "content": "[76] Roland S. Zimmermann, Yash Sharma, Steffen Schneider, Matthias Bethge, and Wieland Brendel. Contrastive Learning Inverts the Data Generating Process. 2021. 2, 3, 4, 5, 6, 15" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16120" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/bcc42e2b-6089-4c27-a5c2-cdec53620ca0_content_list.json b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/bcc42e2b-6089-4c27-a5c2-cdec53620ca0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c580e7384aa45b5cb2807d899fc3e68d9652e962 --- /dev/null +++ b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/bcc42e2b-6089-4c27-a5c2-cdec53620ca0_content_list.json @@ -0,0 +1,1770 @@ +[ + { + "type": "text", + "text": "StableVideo: Text-driven Consistency-aware Diffusion Video Editing", + "text_level": 1, + "bbox": [ + 137, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenhao Chai* Xun Guo² Gaoang Wang¹ Yan Lu²", + "bbox": [ + 246, + 179, + 720, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Zhejiang University 2 Microsoft Research Asia", + "bbox": [ + 282, + 204, + 687, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wenhaochai.19, gaoangwang}@intl.zju.edu.cn, {xunguo, yanlu}@microsoft.com", + "bbox": [ + 158, + 233, + 810, + 248 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69efd36d46271bdad74017f4791141c4003d8b37d502048e67705fe660cd0d42.jpg", + "image_caption": [ + "Figure 1: Editing results of StableVideo. The input video (top row) contains long range motion and viewpoint changing. Our approach performs stable editing according to the prompt of \"Orange SUV in sunny snow winter\" on foreground and background. In the edited video (bottom row), the \"orange SUV\" maintains high geometric and temporal consistency, although the viewpoints keep changing." + ], + "image_footnote": [], + "bbox": [ + 84, + 258, + 885, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 467, + 313, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based methods can generate realistic images and videos, but they struggle to edit existing objects in a video while preserving their appearance over time. This prevents diffusion models from being applied to natural video editing in practical scenarios. In this paper, we tackle this problem by introducing temporal dependency to existing text-driven diffusion models, which allows them to generate consistent appearance for the edited objects. Specifically, we develop a novel inter-frame propagation mechanism for diffusion video editing, which leverages the concept of layered representations to propagate the appearance information from one frame to the next. We then build up a text-driven video editing framework based on this mechanism, namely StableVideo, which can achieve consistency-aware video editing. Extensive experiments demonstrate the strong editing capability of our approach. Compared with state-of-the-art video editing methods, our approach shows superior qualitative and quantitative results. Our code is available at this https URL.", + "bbox": [ + 75, + 500, + 470, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 816, + 207, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed significant progress in extensive computer vision tasks taken by deep learning. Nev-", + "bbox": [ + 76, + 844, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ertheless, natural video editing, which aims at manipulating the appearance of target objects and scenes, still faces two essential challenges that are deterministic to the editing quality: the generator equipped with rich prior knowledge that consistently produces high-fidelity edited contents adhering faithfully to the original geometry of the target objects, and the propagator that disseminates the edited contents throughout the entire video while keeping highly temporal consistency.", + "bbox": [ + 500, + 468, + 890, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The flourish of text-driven generative diffusion models pre-trained on large-scale image and language data [34, 14, 53, 16, 41, 5] provides impressive generation quality. Several diffusion-based methods achieve good performance in image editing [2, 31], but few methods have tried to apply diffusion models in video editing, since it is challenging to modify existing objects while preserving their appearance over the entire video [49, 12, 26]. Dreamix [27] proposes a solution to generate consistent video according to input image/video and prompts. However, it focuses more on generating smooth motions, e.g., pose and camera movements, rather than maintaining geometric consistency of the objects across time. Moreover, such video diffusion models often suffer from huge computing complexity which is not friendly for practical applications.", + "bbox": [ + 496, + 608, + 892, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural layered atlas (NLA) [24, 23] tries to tackle the temporal continuity problem by decomposing the video into a set of atlas layers, each of which describes one target object to be edited. For each atlas layer, the positions of the", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The work was done when the author was with MSRA as an intern.", + "bbox": [ + 94, + 886, + 450, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23040", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "video are mapped into the corresponding 2D positions in it, so that semantically correspondent pixels over the whole video can be represented by the same atlas position. Instead of frame-by-frame editing, NLA edits atlas layers to ensure that the modifications can be precisely mapped back to video frames for temporal smoothness. Text2LIVE [1] provides a text-driven appearance manipulation solution of adding additional edit layers on atlases, in which a specific generator for the edit layers is trained. Although it achieves good results with strict structure preserved, it is not able to apply thorough editing. Moreover, the specifically trained generator also limits the richness of the generated contents.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This brings up the question: Could text-driven diffusion video editing achieve high temporal consistency? Intuitively, employing text-driven diffusion models to edit the atlases corresponding to the target objects could reach such goal. However, this gives rise to drawbacks rather than benefits. Being the summary of the whole video, atlases always have distorted appearance due to the viewpoint and camera movement, which are required to be specifically pretrained and generated as in [1]. Diffusion models may fail in generating satisfied atlas pixels in many cases, so that the corresponding edited frames will also be contaminated. To answer the question, we present two concepts for utilizing diffusion models in video editing. Firstly, instead of editing the atlases directly, we propose to update the atlases via editing key video frames. Secondly, we introduce temporal dependency constraints for diffusion models to generate objects with consistent appearance across time.", + "bbox": [ + 75, + 273, + 472, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on analysis above, we present a novel diffusion video editing approach, StableVideo, to perform consistency-aware video editing. In specific, we propose two effective technologies for this purpose. Firstly, to edit the objects with consistent appearance, we design an interframe propagation mechanism on top of the existing diffusion model [55], which can generate new objects with coherent geometry across time. Secondly, to achieve temporal consistency by leveraging NLA, we design an aggregation network to generate the edited atlases from the key frames. We then build up a text-driven diffusion-based framework, which provides high-quality natural video editing. We conduct extensive qualitative and quantitative experiments to demonstrate the capability of our approach. Compared with state-of-the-art methods, our approach achieves superior results with much lower complexity.", + "bbox": [ + 75, + 530, + 472, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we present the following contributions:", + "bbox": [ + 96, + 772, + 441, + 787 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To our best knowledge, we are the first to solve the consistency problem of diffusion video editing by considering the concept of layered atlas approaches, which provides an efficient and effective way for this topic.", + "- We present a new video editing framework which can manipulate the appearance of the objects with high ge" + ], + "bbox": [ + 94, + 797, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ometry and appearance consistency across time. Our method can be easily applied to other text-driven diffusion models.", + "bbox": [ + 529, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We conduct extensive experiments on a variety of natural videos, which shows superior editing performance compared with state-of-the-art methods.", + "bbox": [ + 517, + 146, + 890, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 213, + 640, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion for Image Editing", + "text_level": 1, + "bbox": [ + 500, + 238, + 748, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The editing of natural images is an important task in the field of computer vision that has been widely studied. Prior to the emergence of diffusion models [42, 15], many GAN-based approaches [11, 10, 30, 32, 48] have achieved good results. There are also some works focus on low-level editing [52, 6, 51]. The advent of diffusion models has made it possible to achieve even higher quality and more diverse edited contents. SDEdit [26] adds noise and corruptions to an input image and uses diffusion models to reverse the process for image editing, while suffering from the loss of fidelity. Prompt-to-Prompt [12] and Plug-and-Play [45] perform semantic editing by mixing activations from original and target text prompts. InstructPix2Pix [2] applies semantic editing at test time and personalizes the model through finetuning and optimization to learn a special token describing the content. UniTune [46] and Imagic [21] finetune on a single image for better editability while maintaining good fidelity. There are also some works exploring the controllability [55, 18, 39, 22, 4, 3] and personalization [9, 38, 29] of diffusion-based generation. Our proposed video editing method leverages existing image editing methods that can preserve the structures, such as [55, 31, 45] and [28].", + "bbox": [ + 496, + 262, + 890, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion for Video Editing", + "text_level": 1, + "bbox": [ + 500, + 603, + 745, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to image editing, video editing is more challenging for diffusion-based methods for geometric and temporal consistency. Tune-a-Video [49] inflates a text-to-image model for video editing. However, since temporal correlations are not fully considered, the editing results suffer from inconsistency of geometry and motion. Dreamix [27] develops a text-to-video backbone for motion editing while maintaining temporal consistency. There are also some works based on video generation like [8, 40, 13, 54, 56] and [17]. Unlike these approaches, our purpose is to enable diffusion models to perform appearance editing with both geometric and temporal consistency.", + "bbox": [ + 496, + 627, + 890, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Temporal Propagation in Video Editing", + "text_level": 1, + "bbox": [ + 500, + 816, + 839, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal propagation plays an important role in natural video process, since it is the essential factor for temporal consistency. Some methods rely on key frames [19, 44, 50] or optical flow [37] to propagate contents between frames.", + "bbox": [ + 496, + 840, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23041", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/233f6ea506e4a63c7c632db96f4dde7989e030a53d016bc9e053ab0f0be9244f.jpg", + "image_caption": [ + "Figure 2: Framework of the proposed StableVideo. The input video is first fed into NLA [24] to generate foreground and background atlases using the pre-trained model. $\\mathcal{G}_b$ is the diffusion model used to edit background atlas, and $\\mathcal{G}_f$ is used to edit the foreground key frames. Note that $\\mathcal{G}_b$ and $\\mathcal{G}_f$ share the same weights, but accept different conditions. We employ depth information, extracted by MiDaS [35], for $\\mathcal{G}_b$ to maintain the consistency between the foreground motion and the environment, while structure guidance is used for $\\mathcal{G}_f$ to keep geometric consistency between the new generated foreground and the old one. After being edited, the foreground and background are blended together to reconstruct the edited frames." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 888, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another bunch of methods are to achieve consistent interframe editing by forming a compressed representation of a video. Omnimatte [24, 25] estimate RGBA layers for target subject and scene effects for each frame independently, but cannot achieve consistent propagation of contents along temporal direction. Atlas [1, 20] tackles this problem by decomposing the video into unified 2D atlas layers for each target. This approach allows contents to be applied to the global summarized 2D atlases and mapped back to the video, achieving temporal consistency with minimal effort. Inspired by the concept of atlas approach, we employ the pre-trained neural layered atlas model to solve the inconsistency problem in diffusion video editing, thereby achieving high-quality editing results with temporal coherence.", + "bbox": [ + 75, + 542, + 472, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 768, + 169, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the pipeline of our proposed StableVideo. We utilize NLA [20] as the propagator for consistent video editing. Specifically, we conduct foreground and background editing separately. For foreground editing, we adopt key frame editing to generate atlas layers with high quality and the inter-frame propagation module to ensure better geometric and temporal consistency. The edited key frames", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are then mapped to partial atlases and aggregated by the aggregation network to produce the edited foreground atlas. It is noteworthy that our approach can also handle more than one foreground layers.", + "bbox": [ + 496, + 542, + 893, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 498, + 617, + 702, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ the pre-trained NLA model [24] to propagate the edited contents to ensure that the target objects and scenes can maintain homogeneous appearances and motions across the entire video. The concept of NLA is to decompose the input video into layered representations, namely foreground atlas and background atlas, which globally summarize the correlated pixels for the foreground and the background, respectively. Three mapping networks, i.e., $\\mathcal{M}^b (\\cdot),\\mathcal{M}^f (\\cdot)$ and $\\mathcal{M}^{\\alpha}(\\cdot)$ , are provided for this purpose. Given an input video $I$ , for each frame $I_{i}$ , we obtain the mapping relationships of the atlas in the background and the foreground with respect to the pixel coordinate system, named as $\\mathrm{UV}^b (\\cdot)$ and $\\mathrm{UV}^f (\\cdot)$ , as well as the foreground opacity $\\alpha_{i}$ on the pixel coordinate system, formulated as:", + "bbox": [ + 496, + 642, + 893, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {U V} _ {i} ^ {b} (\\cdot) = \\mathcal {M} ^ {b} (I _ {i}), \\operatorname {U V} _ {i} ^ {f} (\\cdot) = \\mathcal {M} ^ {f} (I _ {i}), \\alpha_ {i} = \\mathcal {M} ^ {\\alpha} (I _ {i}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 867, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23042", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d1d5f71c9c1edeafaef3c740071afd189040955fe537477a7189873addd98839.jpg", + "image_caption": [ + "Figure 3: Inter-frame propagation for foreground editing. We use two edited key frames, $E_{i-1}$ and $E_i$ , to illustrate the process more clearly. The structure guidance and the text prompt is added into the denoising UNet via the concatenation and cross-attention mechanism respectively." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 472, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, we formulate the mappings from the atlas representation of the background $A^b$ and the foreground $A^f$ , to the pixel coordinate systems of $B_i$ and $F_i$ :", + "bbox": [ + 76, + 460, + 468, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nB _ {i} = \\mathrm {U V} _ {i} ^ {b} (A ^ {b}), F _ {i} = \\mathrm {U V} _ {i} ^ {f} (A ^ {f}). \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 515, + 468, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method achieves geometrical consistent editing by fixing the mappings of $\\mathrm{UV}^b$ and $\\mathrm{UV}^f$ , and generating the edited atlases of $A^b$ and $A^f$ . We adopt a pre-trained latent diffusion model [36] with guided conditions as our generator, namely $\\mathcal{G}^b (\\cdot)$ and $\\mathcal{G}^f (\\cdot)$ . Note that we are not simply editing the foreground and the background atlases directly. We apply inter-frame propagation mechanism on the editing process of foreground atlas. More details are explained in Sec. 3.2. After that, the entire video $I$ can be reconstructed frame by frame as the following equation:", + "bbox": [ + 75, + 542, + 468, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} = \\alpha_ {i} \\circ \\operatorname {U V} _ {i} ^ {f} \\left(\\mathcal {G} ^ {f} \\left(A ^ {f}\\right)\\right) + \\left(1 - \\alpha_ {i}\\right) \\circ \\operatorname {U V} _ {i} ^ {b} \\left(\\mathcal {G} ^ {b} \\left(A ^ {b}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 703, + 468, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\circ$ denotes pixel-wise product.", + "bbox": [ + 76, + 732, + 316, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Inter-frame Propagation", + "text_level": 1, + "bbox": [ + 76, + 756, + 303, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we further elaborate on how inter-frame propagation mechanism helps consistent foreground editing. One of the major challenges for diffusion models is to generate video contents with temporal consistency. Existing state-of-the-art text-driven diffusion methods [55, 28] can maintain the similar geometry between the target objects and the generated ones for image editing by adding structure conditions. However, the situation is different for", + "bbox": [ + 75, + 780, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/73b0403a69007c24d9ccc44e07281e8acd878e2db2a0e475e65f552827e011c7.jpg", + "image_caption": [ + "Figure 4: Training process of aggregation network. We employ a simple 3D network to aggregate the partial atlases generated from the edited key frames. For each training iteration, the sampled key frames are edited and mapped to partial atlases by $UV_{i}^{f}$ . The partial atlases are then fed into the Aggregation Network to generate the edit atlas, which is mapped back by $(UV_{i}^{f})^{-1}$ to generate the reconstructed frames. A $L_{1}$ loss is used to guarantee the aggregation consistency between the edited key frames and the reconstructed frames." + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "videos. Generating temporally consistent geometry needs to handle some uncertain changes across time, e.g., motion and deformation, which can not be supported by them. We tackle this problem by introducing a conditional denoising process to enable the diffusion models to consider both the structure of the current frame and the appearance information from previous frame, thereby sequentially generating new objects with geometric consistency across time. In specific, we employ canny edge as structure guidance, which is also adopted by existing diffusion methods [55]. Another important question is how to propagate the information of one object across frames to achieve consistent appearance. With the help of NLA, we can transfer the appearance features of the overlapping parts of previous frame to the next frame. Inspired by SDEdit [26] and ILVR [7], we further use a process of adding noise and denoising to obtain a more complete output. We illustrate this generation process in Fig. 3. The inter-frame propagation method is only applied on the foreground objects.", + "bbox": [ + 496, + 412, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first select $N$ foreground key frames from the original video $I$ , ensuring that 1) there is significant overlap between the adjacent frames, and 2) these key frames capture the appearance of all faces of the object. Given a generator $\\mathcal{G}^f(\\cdot)$ and a text prompt $T$ , we edit the first frame $F_0$ in pixel coordinate system with its structure condition $C_0$ as the extra guidance:", + "bbox": [ + 496, + 700, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {0} = \\mathcal {G} ^ {f} (T, C _ {0}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 814, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_0$ represents the editing result. Then for the remaining key frames, we propagate the editing result from the previous key frame $E_{i-1}$ to obtain the one of the current key frame $E_i$ . To be specific, we map $E_{i-1}$ from the pixel", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23043", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Foreground: a polar bear; Background: north pole.", + "bbox": [ + 91, + 89, + 434, + 104 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5374b288a1d996eb4934bc8d6613e530dd9d45f6e25d58094167a26c16ef2695.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 106, + 885, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Foreground: a car with graffiti; Background: Miami city.", + "bbox": [ + 91, + 244, + 473, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8d280b1ac9351d46e152ca0e06d145bf8790b7d18180b2e6ccb7377982a96b51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 262, + 885, + 323 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18c530d0c83dfaf44954e1f81390142a04906fe3aae616f2d7930fa9c8931c2f.jpg", + "image_caption": [ + "Figure 5: Compositing editing. We demonstrate two editing examples with non-rigid and rigid foreground objects. Our approach can well preserve the geometry of \"bear\" and \"car\" across frames." + ], + "image_footnote": [], + "bbox": [ + 86, + 327, + 885, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "coordinate to atlas $A_{i - 1}^{E}$ and then map it back to the pixel coordinate in the current frame $i$ as $\\hat{E}_i$ with multiplying the opacity $\\alpha_{i}$ , formulated by:", + "bbox": [ + 75, + 457, + 470, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i - 1} ^ {f} = \\left(\\mathrm {U V} _ {i - 1} ^ {f}\\right) ^ {- 1} \\left(E _ {i - 1}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 517, + 468, + 545 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} _ {i} = \\alpha_ {i} \\circ \\mathrm {U V} _ {i} ^ {f} (A _ {i - 1} ^ {f}).\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 539, + 354, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is noteworthy that the entire video shares the same foreground atlas $A^f$ for each target object, where the subscript $i$ represents an incomplete partial atlas in frame $i$ . We then use the atlas to propagate the pixel values from previous frame to their corresponding positions in the current frame to obtain an incomplete partial appearance $\\hat{E}_i$ .", + "bbox": [ + 75, + 570, + 468, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the partial appearance $\\hat{E}_i$ , we first encode it with VQ-VAE [47] to get the latent representation $\\hat{Z}_i$ , and then add noise to it with Variance Preserving Stochastic Differential Equation (VP-SDE), formulated as:", + "bbox": [ + 75, + 661, + 468, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Z} _ {i} (t _ {0}) = \\alpha (t _ {0}) \\hat {Z} _ {t} (0) + \\sigma (t _ {0}) \\mathbf {z}, \\quad \\mathbf {z} \\sim \\mathcal {N} (0, \\mathbf {I}) \\quad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 733, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma(t_0)$ and $\\alpha(t_0)$ are two scalar functions that satisfy $\\alpha^2(t) + \\sigma^2(t) = 1, \\forall t \\in (0,1]$ , and $t_0 \\in [0,1]$ is a hyper parameter of the noise strength. Then we apply denoising process $\\hat{Z}_i(t_0)$ under the condition guidance from both text prompt $T$ and structure guidance $C_i$ to get the latent representation $Z_i$ . Finally, we decode the latent representation to $Z_i$ propagate editing result $E_i$ . Our experiments further demonstrate that this mechanism can achieve good propagation results without training or fine-tuning the model.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Aggregation Network", + "text_level": 1, + "bbox": [ + 498, + 455, + 702, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from [1] and [24], our approach edits video frames rather than atlases, which have the chance to achieve more information of different viewpoints. This brings two advantages. Firstly, the geometries and pixels from different viewpoints provide more details of the target objects, allowing the diffusion model to generate the edited content with higher fidelity. Secondly, this alleviates the risk of failure editing due to the potential wrong mapping from the atlas to the video frames. We then aggregate the edited key frames by using a simple yet effective two-layer 2D convolution network with skip connection as shown in Fig. 4. Our goal is to guarantee that the aggregated atlas is highly aligned with the original one, in terms of locations, so that appearance edit will not affect the geometric consistency and the temporal continuity. Reconstruction loss, $\\mathcal{L}_{rec}$ , between the edited and reconstructed key frames is employed in the training process as:", + "bbox": [ + 496, + 479, + 890, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\sum_ {i = 1} ^ {N} | | E _ {i} - \\mathrm {U V} _ {i} ^ {f} \\left(A ^ {f}\\right) | | _ {1}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 744, + 890, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the number of key frames.", + "bbox": [ + 500, + 794, + 753, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 821, + 633, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 847, + 707, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, we implement our approach over Stable Diffusion [36]. Despite there are several image-based meth", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23044", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Background: desert scene, pyramid.", + "bbox": [ + 89, + 90, + 334, + 106 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/663f260a0da3dadbde6d448edb13f668868800720d77f84c661d8fa4a70eb0cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 108, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/07465c1bd9aab0c9f7176266cb42fbb26f2c4ab76d3855cebb98797db63e9e85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 172, + 887, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Background: magma scene, crack road.", + "bbox": [ + 88, + 244, + 357, + 260 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af16a53aa50d2e156307ee56123b0389f296ec98af50aa54aec663fbb21d2c7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 263, + 887, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fff5268a65ad885b9b793e22cb71867ab1f553add3d2c532e89176d50e4e329c.jpg", + "image_caption": [ + "Figure 6: Background replacement. Since our approach can effectively maintain the geometry of the foreground, it can perform background replacement while maintaining geometric consistency of depth and temporal continuity of perspective." + ], + "image_footnote": [], + "bbox": [ + 81, + 330, + 887, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ods [31, 45, 28] can perform structure-preserving editing, we choose the canny condition branch from [55] as the structure guidance for the proposed inter-frame propagation in our method. We apply our method on several videos from DAVIS [33], with each video containing a moving object in $50 \\sim 70$ frames. The image resolution is set to $768 \\times 432$ , and the resolution of foreground atlas is set to $2000 \\times 2000$ . We employ DDIM [43] sampler with 20 steps. In this case, our method requires only $\\sim 10$ GB GPU memory and takes $\\sim 30$ seconds for each video in a single NVIDIA A40 GPU.", + "bbox": [ + 75, + 445, + 472, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Editing Results", + "text_level": 1, + "bbox": [ + 76, + 604, + 232, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We tested our method for various editing types among several videos. Here we demonstrate several scenarios:", + "bbox": [ + 75, + 628, + 468, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compositing editing. Our method can edit foreground and background separately to achieve high-quality and semantically matched editing results as shown in Fig. 5.", + "bbox": [ + 75, + 657, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Background replacement. Due to the separate editing capabilities enabled by NLA, as well as the diverse and high-quality editing brought by the diffusion model, our method achieves high-quality video background replacement while maintaining geometric consistency of depth and temporal continuity of perspective, as shown in Fig. 6.", + "bbox": [ + 75, + 703, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Style transfer. Our method accomplishes a wide range of style transfers, while simultaneously ensuring temporal consistency, as shown in Fig. 7.", + "bbox": [ + 75, + 794, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison to Prior Arts", + "text_level": 1, + "bbox": [ + 76, + 847, + 312, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our editing results with state-of-the-art atlas-based method Text2LIVE [1] and diffusion-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "based method Tune-A-Video [49].", + "bbox": [ + 498, + 445, + 728, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to Text2LIVE [1]. As shown in Fig. 8, given the text prompt \"an orange SUV\", Text2LIVE shows incomplete editing, while our method demonstrates a much more holistic result. It is because our method employs key frame editing, while Text2LIVE creates a new layer to edit the atlas somehow directly. The potential failure editing in atlas will lead to the error propagation to the entire video. In addition, compared to Text2LIVE, we can achieve higher quality and richer content with faster inference speed.", + "bbox": [ + 496, + 460, + 892, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to Tune-A-Video [49]. As shown in Fig. 9, vanilla video diffusion models like Tune-A-Video often fail in video editing. While it adeptly captures the semantic information of text prompts, it struggles to preserve consistency in video layout and object geometry.", + "bbox": [ + 496, + 598, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistency analysis. To the best of our knowledge, there is currently no widely accepted metric for evaluating the geometric and temporal consistency of videos. In this paper, we employ motion consistency of dense optical flow and deviation consistency of the edited video frames for this metric. For motion consistency, we employ the Farneback algorithm in OpenCV to calculate the average L2 distance of dense optical flow between the edited and original videos. We select the \"car-turn\" video from DAVIS and apply the same prompt to all methods. The experiment is repeated several times to obtain the average number. Our method shows better stability than Tune-A-Video as shown in Tab. 1. For deviation consistency, we conduct experiments on 1) CLIP score: target text faithfulness 2) LPIPS-P: deviation from the original video frames and 3) LPIPS-T:", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23045", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Foreground: a panda; Both: in the style of Chinese painting.", + "bbox": [ + 91, + 90, + 493, + 106 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2483d949fe37b74d3efe3b5e38a5fb432eb0613d37e570b2f406012a13d75489.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 108, + 885, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b27afc08e297416387f9b62914f53cfb2b581d2ad4f314cd2c4d873bbe01c49c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 172, + 885, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Both: in the style of Vincent Willem van Gogh's Starry Night.", + "bbox": [ + 91, + 247, + 506, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1484075a1af04971cc6cb3b66f09a8529a63a4b8a693e7fcbe6b82825a3976a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 263, + 885, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e3f99ea0d6a2d87f317804c6c8fc9219ed830982162baa413adf955312ef81e.jpg", + "image_caption": [ + "Figure 7: Style transfer. Our approach achieves diverse video style transfer while ensuring high temporal consistency." + ], + "image_footnote": [], + "bbox": [ + 84, + 328, + 885, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc3358983057c125467417a34372a52f5eeb97de8c0dc1adfda43f9afc96c097.jpg", + "image_caption": [ + "Text2LIVE" + ], + "image_footnote": [], + "bbox": [ + 83, + 445, + 467, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20d3965b105977ffad705ff40b4d8152352a8dd9e425e9ec43b1b59018c0ae39.jpg", + "image_caption": [ + "StableVideo (ours)", + "Figure 8: Comparison to Text2LIVE. Foreground prompt: an orange SUV. Our method achieves more holistic editing results on the foreground.", + "Figure 9: Comparison to Tune-A-Video. Prompt: an orange SUV in sunny snowy winter, cabins. Our method achieves much more consistent editing results." + ], + "image_footnote": [], + "bbox": [ + 83, + 563, + 467, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "deviation between adjacent frames, as shown in Tab. 2. Our method achieves comparable CLIP score and much lower deviations, which shows the effectiveness and stability.", + "bbox": [ + 75, + 746, + 470, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 801, + 230, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the necessity of the key frame editing, we apply editing in atlas layer directly for the foreground as a simple baseline. The atlas might not be so deformed for human perception, but it significantly affects the diffusion models. Fig. 10 shows an example, where obvious deformation and", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c27349148b31ba5c0a543bf3cf3170d60fc119b40ceedd0b8c2bd7a819cd3d29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 443, + 888, + 487 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd446f702b699b438815e52e29f595deb71ad72f2d4b8fc6c33dcce9da2c93e5.jpg", + "image_caption": [ + "original video", + "Tune-A-Video" + ], + "image_footnote": [], + "bbox": [ + 501, + 503, + 888, + 547 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f34f2c5c00fd2dcc8563c1047e9810c3d1d0e3f61c15dde62227e918e188d704.jpg", + "image_caption": [ + "StableVideo (ours)" + ], + "image_footnote": [], + "bbox": [ + 501, + 566, + 888, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Editing atlas", + "text_level": 1, + "bbox": [ + 501, + 696, + 571, + 708 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/521d3b6d18cd5b0421b906fd8e238abf0b9610e5a3dcd36d40667d82b5143cca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 709, + 890, + 739 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9c7691af5ef240ba069b8bca46d2fdefdeda2a138b7bc586d68963728772eb4.jpg", + "image_caption": [ + "Editing key frames", + "Figure 10: Ablation study on directly editing the atlas. The deformation in atlas affects the diffusion models." + ], + "image_footnote": [], + "bbox": [ + 501, + 753, + 890, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "inconsistency exist.", + "bbox": [ + 500, + 854, + 630, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct extensive ablation study on inter-frame propagation module. The objective of this module is to", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23046", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c0a28fcb55e54d53a4d4f1119b1c641b2279c0a787ce20d26d64c03ca8a7d01.jpg", + "image_caption": [ + "random generation" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 465, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e3febb87f7f45873f359d1a48971cde4ea62b3c14fa5a9321d0830fdf31f60c.jpg", + "image_caption": [ + "fix initial latent" + ], + "image_footnote": [], + "bbox": [ + 81, + 154, + 465, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e14a33a891338e9e0fdd62d378fe1ec2aaa6fc7e9070efb55adf7f80f924251.jpg", + "image_caption": [ + "sequential generation" + ], + "image_footnote": [], + "bbox": [ + 81, + 218, + 465, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/beab8b47099ef54fbe07f10d3baf812d34f57500f54739c585205db3e0d82d73.jpg", + "image_caption": [ + "Inter-frame Propagation (ours)" + ], + "image_footnote": [], + "bbox": [ + 81, + 282, + 465, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8453c216cf43082f1bc4176039f6fe32e174706e58aa900f729298f460d61381.jpg", + "image_caption": [ + "Figure 11: Ablation study on inter-frame propagation module. Foreground prompt: an orange SUV. Our method achieves excellent consistency in key frame appearance.", + "Figure 12: Binary search on hyper parameter $t_0$ in interframe propagation module. The appearance of previous frame is shown on the right side. The sequence on the left shows the generated results of the current frame as a function of $t_0$ . The value of $t_0$ around 0.8 is a good sweet spot.", + "Fig. 12. We observe that as $t_0$ gradually increases, the generated results become more realistic but gradually lose their match with the appearance of the previous frame. Using binary search, we find out that a reasonable trade-off between fidelity and realism for $t_0$ lies around 0.8." + ], + "image_footnote": [], + "bbox": [ + 81, + 415, + 467, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "maintain the geometry of the foreground when editing key frames. Firstly, we consider four different settings for editing key frames as shown in Fig. 11.", + "bbox": [ + 75, + 561, + 468, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Random generation. Each key frame only shares the same text prompt with the others. In this case, there are significant differences among the generated key frames.", + "bbox": [ + 75, + 609, + 468, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fix initial latent. Unlike starting from random noise every time we edit, we start generating each key frame from the same latent noise and share the text prompt. In this case, there is higher similarity in the content generated for each frame, but the consistency is still not satisfactory.", + "bbox": [ + 75, + 655, + 468, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sequential generation. Furthermore, we concatenate the latent noise between frames. Specifically, we apply imaged-to-image translation between frames. This method still cannot guarantee consistency since the appearances of the objects between the two frames do not match.", + "bbox": [ + 75, + 732, + 468, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inter-frame propagation (ours). Our final approach is to employ partial atlas to geometrically align the appearances between two frames, followed by a process of adding noise and then apply denoising process.", + "bbox": [ + 75, + 808, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, We also conducted experiments on the selection of the hyper parameter $t_0$ in this module as shown in", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7c4693f0e970415c09253db6c9bfec5576a7b6646f7292e94de91d9a6d527be4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodForeground EditingComposite Editing
Text2LIVE [1]1.997.39
Tune-A-Video [49]4.6312.74
StableVideo (ours)3.3411.48
", + "bbox": [ + 524, + 89, + 867, + 154 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3fabaf00bd616a351c6e297b1414dcd7b3e27f68cee4de4b0ed206fa2f81550d.jpg", + "table_caption": [ + "Table 1: Consistency of dense optical flow. Original video: car-turn. Foreground prompt (column 1): an orange suv. Composite prompt (column 2): a car driving in winter." + ], + "table_footnote": [], + "table_body": "
MethodCLIP (↑)LPIPS-P (↓)LPIPS-T (↓)
Tune-A-Video [49]0.27870.63460.1851
StableVideo (ours)0.27130.16130.0386
", + "bbox": [ + 524, + 224, + 867, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: Quantitative results. Original video: car-turn. Prompt: an orange SUV in the winter.", + "bbox": [ + 498, + 291, + 888, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95117529b88d127ce7f948386fcec03fe288ebe71d53380335f9a7a79db417dc.jpg", + "image_caption": [ + "Figure 13: Failure case. Videos with non-rigid deformation may lead to failure editing, since the movement of the object is more difficult to be well captured in this case." + ], + "image_footnote": [], + "bbox": [ + 501, + 335, + 890, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Future Works", + "text_level": 1, + "bbox": [ + 500, + 556, + 779, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Firstly, our method is constrained by NLA. Learning atlas layers may fail for non-rigid objects with significant structural deformation as shown in Fig. 13. While we can mitigate this by dividing long videos into short clips where the objects can be considered to be rigid, it is still not feasible to address every single case. Secondly, our method is constrained by the capabilities of the diffusion models, which may struggle with specific scenarios such as human or animals. Besides, it may be better to optimize the diffusion model with the objective of aligning the generated contents to the reconstructed ones.", + "bbox": [ + 496, + 582, + 890, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 762, + 617, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a text-driven diffusion video editing approach. To solve the consistency problem for diffusion models in foreground object editing, we propose a interframe propagation mechanism and an atlas aggregation network. We conducted extensive experiments and demonstrated the superior qualitative and quantitative results of our method compared to state-of-the-art approaches.", + "bbox": [ + 496, + 787, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23047", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 2, 3, 5, 6, 8", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 1, 2", + "[3] Shidong Cao, Wenhao Chai, Shengyu Hao, and Gaoang Wang. Image reference-guided fashion design with structure-aware transfer by diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3524-3528, 2023. 2", + "[4] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. arXiv preprint arXiv:2302.06826, 2023. 2", + "[5] Wenhao Chai and Gaoang Wang. Deep vision multimodal learning: Methodology, benchmark, and trend. Applied Sciences, 12(13):6588, 2022. 1", + "[6] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2", + "[7] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 4", + "[8] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 2", + "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[10] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2", + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2", + "[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:210.02303, 2022. 1", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2", + "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1", + "[17] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. 2", + "[18] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 2", + "[19] Ondřej Jamriška, Šárka Sochorová, Ondřej Texler, Michal Lukáč, Jakub Fiser, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Stylizing video by example. ACM Transactions on Graphics (TOG), 38(4):1-11, 2019. 2", + "[20] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 3", + "[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2", + "[22] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. arXiv preprint arXiv:2301.07093, 2023. 2", + "[23] Weng Fei Low and Gim Hee Lee. Minimal neural atlas: Parameterizing complex surfaces with minimal charts and distortion. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 465-481. Springer, 2022. 1", + "[24] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. arXiv preprint arXiv:2009.07833, 2020. 1, 3, 5", + "[25] Erika Lu, Forrester Cole, Weidi Xie, Tali Dekel, William T Freeman, Andrew Zisserman, and Michael Rubinstein. Associating objects and their effects in video through coordination games. In Advances in Neural Information Processing Systems. 3", + "[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2, 4", + "[27] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23048", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2, 4, 6", + "[29] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2", + "[30] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 2", + "[31] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. arXiv preprint arXiv:2302.03027, 2023. 1, 2, 6", + "[32] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2", + "[33] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 6", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1", + "[35] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 3", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 4, 5", + "[37] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In *Pattern Recognition: 38th German Conference*, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings 38, pages 26-36. Springer, 2016. 2", + "[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2", + "[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2", + "[40] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual," + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[41] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1", + "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations. 2", + "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6", + "[44] Ondrej Texler, David Futschik, Michal Kucera, Ondrej Jamriška, Šárka Sochorová, Menclei Chai, Sergey Tulyakov, and Daniel Sykora. Interactive video stylization using few-shot patch-based training. ACM Transactions on Graphics (TOG), 39(4):73-1, 2020. 2", + "[45] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 6", + "[46] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 2", + "[47] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5", + "[48] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2", + "[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 1, 2, 6, 8", + "[50] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 357-374. Springer, 2022. 2", + "[51] Ren Yang, Radu Timofte, Xin Li, Qi Zhang, Lin Zhang, Fanglong Liu, Dongliang He, Fu Li, He Zheng, Weihang Yuan, et al. Aim 2022 challenge on super-resolution of compressed image and video: Dataset, methods and results. In European Conference on Computer Vision, pages 174-202. Springer, 2022. 2", + "[52] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2", + "[53] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. Magvit:" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23049", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Masked generative video transformer. arXiv preprint arXiv:2212.05199, 2022. 1", + "[54] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. arXiv preprint arXiv:2302.07685, 2023. 2", + "[55] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2, 4, 6", + "[56] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 261 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23050", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/bcc42e2b-6089-4c27-a5c2-cdec53620ca0_model.json b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/bcc42e2b-6089-4c27-a5c2-cdec53620ca0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..935f70143221b77dd210c2135386108168ad716b --- /dev/null +++ b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/bcc42e2b-6089-4c27-a5c2-cdec53620ca0_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "StableVideo: Text-driven Consistency-aware Diffusion Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.18, + 0.721, + 0.201 + ], + "angle": 0, + "content": "Wenhao Chai* Xun Guo² Gaoang Wang¹ Yan Lu²" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.205, + 0.688, + 0.224 + ], + "angle": 0, + "content": "1 Zhejiang University 2 Microsoft Research Asia" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.234, + 0.812, + 0.249 + ], + "angle": 0, + "content": "{wenhaochai.19, gaoangwang}@intl.zju.edu.cn, {xunguo, yanlu}@microsoft.com" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.259, + 0.887, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.393, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Figure 1: Editing results of StableVideo. The input video (top row) contains long range motion and viewpoint changing. Our approach performs stable editing according to the prompt of \"Orange SUV in sunny snow winter\" on foreground and background. In the edited video (bottom row), the \"orange SUV\" maintains high geometric and temporal consistency, although the viewpoints keep changing." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.468, + 0.314, + 0.483 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.472, + 0.788 + ], + "angle": 0, + "content": "Diffusion-based methods can generate realistic images and videos, but they struggle to edit existing objects in a video while preserving their appearance over time. This prevents diffusion models from being applied to natural video editing in practical scenarios. In this paper, we tackle this problem by introducing temporal dependency to existing text-driven diffusion models, which allows them to generate consistent appearance for the edited objects. Specifically, we develop a novel inter-frame propagation mechanism for diffusion video editing, which leverages the concept of layered representations to propagate the appearance information from one frame to the next. We then build up a text-driven video editing framework based on this mechanism, namely StableVideo, which can achieve consistency-aware video editing. Extensive experiments demonstrate the strong editing capability of our approach. Compared with state-of-the-art video editing methods, our approach shows superior qualitative and quantitative results. Our code is available at this https URL." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.818, + 0.208, + 0.833 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.845, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Recent years have witnessed significant progress in extensive computer vision tasks taken by deep learning. Nev-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.469, + 0.892, + 0.606 + ], + "angle": 0, + "content": "ertheless, natural video editing, which aims at manipulating the appearance of target objects and scenes, still faces two essential challenges that are deterministic to the editing quality: the generator equipped with rich prior knowledge that consistently produces high-fidelity edited contents adhering faithfully to the original geometry of the target objects, and the propagator that disseminates the edited contents throughout the entire video while keeping highly temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.893, + 0.837 + ], + "angle": 0, + "content": "The flourish of text-driven generative diffusion models pre-trained on large-scale image and language data [34, 14, 53, 16, 41, 5] provides impressive generation quality. Several diffusion-based methods achieve good performance in image editing [2, 31], but few methods have tried to apply diffusion models in video editing, since it is challenging to modify existing objects while preserving their appearance over the entire video [49, 12, 26]. Dreamix [27] proposes a solution to generate consistent video according to input image/video and prompts. However, it focuses more on generating smooth motions, e.g., pose and camera movements, rather than maintaining geometric consistency of the objects across time. Moreover, such video diffusion models often suffer from huge computing complexity which is not friendly for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Neural layered atlas (NLA) [24, 23] tries to tackle the temporal continuity problem by decomposing the video into a set of atlas layers, each of which describes one target object to be edited. For each atlas layer, the positions of the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.452, + 0.9 + ], + "angle": 0, + "content": "*The work was done when the author was with MSRA as an intern." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "23040" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "video are mapped into the corresponding 2D positions in it, so that semantically correspondent pixels over the whole video can be represented by the same atlas position. Instead of frame-by-frame editing, NLA edits atlas layers to ensure that the modifications can be precisely mapped back to video frames for temporal smoothness. Text2LIVE [1] provides a text-driven appearance manipulation solution of adding additional edit layers on atlases, in which a specific generator for the edit layers is trained. Although it achieves good results with strict structure preserved, it is not able to apply thorough editing. Moreover, the specifically trained generator also limits the richness of the generated contents." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.529 + ], + "angle": 0, + "content": "This brings up the question: Could text-driven diffusion video editing achieve high temporal consistency? Intuitively, employing text-driven diffusion models to edit the atlases corresponding to the target objects could reach such goal. However, this gives rise to drawbacks rather than benefits. Being the summary of the whole video, atlases always have distorted appearance due to the viewpoint and camera movement, which are required to be specifically pretrained and generated as in [1]. Diffusion models may fail in generating satisfied atlas pixels in many cases, so that the corresponding edited frames will also be contaminated. To answer the question, we present two concepts for utilizing diffusion models in video editing. Firstly, instead of editing the atlases directly, we propose to update the atlases via editing key video frames. Secondly, we introduce temporal dependency constraints for diffusion models to generate objects with consistent appearance across time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.473, + 0.773 + ], + "angle": 0, + "content": "Based on analysis above, we present a novel diffusion video editing approach, StableVideo, to perform consistency-aware video editing. In specific, we propose two effective technologies for this purpose. Firstly, to edit the objects with consistent appearance, we design an interframe propagation mechanism on top of the existing diffusion model [55], which can generate new objects with coherent geometry across time. Secondly, to achieve temporal consistency by leveraging NLA, we design an aggregation network to generate the edited atlases from the key frames. We then build up a text-driven diffusion-based framework, which provides high-quality natural video editing. We conduct extensive qualitative and quantitative experiments to demonstrate the capability of our approach. Compared with state-of-the-art methods, our approach achieves superior results with much lower complexity." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.773, + 0.442, + 0.788 + ], + "angle": 0, + "content": "In summary, we present the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.798, + 0.47, + 0.86 + ], + "angle": 0, + "content": "- To our best knowledge, we are the first to solve the consistency problem of diffusion video editing by considering the concept of layered atlas approaches, which provides an efficient and effective way for this topic." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We present a new video editing framework which can manipulate the appearance of the objects with high ge" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.798, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "ometry and appearance consistency across time. Our method can be easily applied to other text-driven diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.147, + 0.892, + 0.192 + ], + "angle": 0, + "content": "- We conduct extensive experiments on a variety of natural videos, which shows superior editing performance compared with state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.214, + 0.642, + 0.23 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.239, + 0.749, + 0.256 + ], + "angle": 0, + "content": "2.1. Diffusion for Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.892, + 0.596 + ], + "angle": 0, + "content": "The editing of natural images is an important task in the field of computer vision that has been widely studied. Prior to the emergence of diffusion models [42, 15], many GAN-based approaches [11, 10, 30, 32, 48] have achieved good results. There are also some works focus on low-level editing [52, 6, 51]. The advent of diffusion models has made it possible to achieve even higher quality and more diverse edited contents. SDEdit [26] adds noise and corruptions to an input image and uses diffusion models to reverse the process for image editing, while suffering from the loss of fidelity. Prompt-to-Prompt [12] and Plug-and-Play [45] perform semantic editing by mixing activations from original and target text prompts. InstructPix2Pix [2] applies semantic editing at test time and personalizes the model through finetuning and optimization to learn a special token describing the content. UniTune [46] and Imagic [21] finetune on a single image for better editability while maintaining good fidelity. There are also some works exploring the controllability [55, 18, 39, 22, 4, 3] and personalization [9, 38, 29] of diffusion-based generation. Our proposed video editing method leverages existing image editing methods that can preserve the structures, such as [55, 31, 45] and [28]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.746, + 0.621 + ], + "angle": 0, + "content": "2.2. Diffusion for Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Compared to image editing, video editing is more challenging for diffusion-based methods for geometric and temporal consistency. Tune-a-Video [49] inflates a text-to-image model for video editing. However, since temporal correlations are not fully considered, the editing results suffer from inconsistency of geometry and motion. Dreamix [27] develops a text-to-video backbone for motion editing while maintaining temporal consistency. There are also some works based on video generation like [8, 40, 13, 54, 56] and [17]. Unlike these approaches, our purpose is to enable diffusion models to perform appearance editing with both geometric and temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.84, + 0.835 + ], + "angle": 0, + "content": "2.3. Temporal Propagation in Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Temporal propagation plays an important role in natural video process, since it is the essential factor for temporal consistency. Some methods rely on key frames [19, 44, 50] or optical flow [37] to propagate contents between frames." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "23041" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.889, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.425, + 0.896, + 0.519 + ], + "angle": 0, + "content": "Figure 2: Framework of the proposed StableVideo. The input video is first fed into NLA [24] to generate foreground and background atlases using the pre-trained model. \\(\\mathcal{G}_b\\) is the diffusion model used to edit background atlas, and \\(\\mathcal{G}_f\\) is used to edit the foreground key frames. Note that \\(\\mathcal{G}_b\\) and \\(\\mathcal{G}_f\\) share the same weights, but accept different conditions. We employ depth information, extracted by MiDaS [35], for \\(\\mathcal{G}_b\\) to maintain the consistency between the foreground motion and the environment, while structure guidance is used for \\(\\mathcal{G}_f\\) to keep geometric consistency between the new generated foreground and the old one. After being edited, the foreground and background are blended together to reconstruct the edited frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.473, + 0.756 + ], + "angle": 0, + "content": "Another bunch of methods are to achieve consistent interframe editing by forming a compressed representation of a video. Omnimatte [24, 25] estimate RGBA layers for target subject and scene effects for each frame independently, but cannot achieve consistent propagation of contents along temporal direction. Atlas [1, 20] tackles this problem by decomposing the video into unified 2D atlas layers for each target. This approach allows contents to be applied to the global summarized 2D atlases and mapped back to the video, achieving temporal consistency with minimal effort. Inspired by the concept of atlas approach, we employ the pre-trained neural layered atlas model to solve the inconsistency problem in diffusion video editing, thereby achieving high-quality editing results with temporal coherence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.17, + 0.785 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Fig. 2 shows the pipeline of our proposed StableVideo. We utilize NLA [20] as the propagator for consistent video editing. Specifically, we conduct foreground and background editing separately. For foreground editing, we adopt key frame editing to generate atlas layers with high quality and the inter-frame propagation module to ensure better geometric and temporal consistency. The edited key frames" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.895, + 0.605 + ], + "angle": 0, + "content": "are then mapped to partial atlases and aggregated by the aggregation network to produce the edited foreground atlas. It is noteworthy that our approach can also handle more than one foreground layers." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.618, + 0.704, + 0.633 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.895, + 0.856 + ], + "angle": 0, + "content": "We employ the pre-trained NLA model [24] to propagate the edited contents to ensure that the target objects and scenes can maintain homogeneous appearances and motions across the entire video. The concept of NLA is to decompose the input video into layered representations, namely foreground atlas and background atlas, which globally summarize the correlated pixels for the foreground and the background, respectively. Three mapping networks, i.e., \\(\\mathcal{M}^b (\\cdot),\\mathcal{M}^f (\\cdot)\\) and \\(\\mathcal{M}^{\\alpha}(\\cdot)\\), are provided for this purpose. Given an input video \\(I\\), for each frame \\(I_{i}\\), we obtain the mapping relationships of the atlas in the background and the foreground with respect to the pixel coordinate system, named as \\(\\mathrm{UV}^b (\\cdot)\\) and \\(\\mathrm{UV}^f (\\cdot)\\), as well as the foreground opacity \\(\\alpha_{i}\\) on the pixel coordinate system, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.868, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\operatorname {U V} _ {i} ^ {b} (\\cdot) = \\mathcal {M} ^ {b} (I _ {i}), \\operatorname {U V} _ {i} ^ {f} (\\cdot) = \\mathcal {M} ^ {f} (I _ {i}), \\alpha_ {i} = \\mathcal {M} ^ {\\alpha} (I _ {i}). \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "23042" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.087, + 0.473, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.437 + ], + "angle": 0, + "content": "Figure 3: Inter-frame propagation for foreground editing. We use two edited key frames, \\( E_{i-1} \\) and \\( E_i \\), to illustrate the process more clearly. The structure guidance and the text prompt is added into the denoising UNet via the concatenation and cross-attention mechanism respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.462, + 0.47, + 0.507 + ], + "angle": 0, + "content": "After that, we formulate the mappings from the atlas representation of the background \\( A^b \\) and the foreground \\( A^f \\), to the pixel coordinate systems of \\( B_i \\) and \\( F_i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.516, + 0.47, + 0.535 + ], + "angle": 0, + "content": "\\[\nB _ {i} = \\mathrm {U V} _ {i} ^ {b} (A ^ {b}), F _ {i} = \\mathrm {U V} _ {i} ^ {f} (A ^ {f}). \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.696 + ], + "angle": 0, + "content": "Our method achieves geometrical consistent editing by fixing the mappings of \\(\\mathrm{UV}^b\\) and \\(\\mathrm{UV}^f\\), and generating the edited atlases of \\(A^b\\) and \\(A^f\\). We adopt a pre-trained latent diffusion model [36] with guided conditions as our generator, namely \\(\\mathcal{G}^b (\\cdot)\\) and \\(\\mathcal{G}^f (\\cdot)\\). Note that we are not simply editing the foreground and the background atlases directly. We apply inter-frame propagation mechanism on the editing process of foreground atlas. More details are explained in Sec. 3.2. After that, the entire video \\(I\\) can be reconstructed frame by frame as the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.704, + 0.469, + 0.723 + ], + "angle": 0, + "content": "\\[\nI _ {i} = \\alpha_ {i} \\circ \\operatorname {U V} _ {i} ^ {f} \\left(\\mathcal {G} ^ {f} \\left(A ^ {f}\\right)\\right) + \\left(1 - \\alpha_ {i}\\right) \\circ \\operatorname {U V} _ {i} ^ {b} \\left(\\mathcal {G} ^ {b} \\left(A ^ {b}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.733, + 0.318, + 0.749 + ], + "angle": 0, + "content": "where \\(\\circ\\) denotes pixel-wise product." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.304, + 0.773 + ], + "angle": 0, + "content": "3.2. Inter-frame Propagation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In this section, we further elaborate on how inter-frame propagation mechanism helps consistent foreground editing. One of the major challenges for diffusion models is to generate video contents with temporal consistency. Existing state-of-the-art text-driven diffusion methods [55, 28] can maintain the similar geometry between the target objects and the generated ones for image editing by adding structure conditions. However, the situation is different for" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.891, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.238, + 0.895, + 0.389 + ], + "angle": 0, + "content": "Figure 4: Training process of aggregation network. We employ a simple 3D network to aggregate the partial atlases generated from the edited key frames. For each training iteration, the sampled key frames are edited and mapped to partial atlases by \\( UV_{i}^{f} \\). The partial atlases are then fed into the Aggregation Network to generate the edit atlas, which is mapped back by \\( (UV_{i}^{f})^{-1} \\) to generate the reconstructed frames. A \\( L_{1} \\) loss is used to guarantee the aggregation consistency between the edited key frames and the reconstructed frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.7 + ], + "angle": 0, + "content": "videos. Generating temporally consistent geometry needs to handle some uncertain changes across time, e.g., motion and deformation, which can not be supported by them. We tackle this problem by introducing a conditional denoising process to enable the diffusion models to consider both the structure of the current frame and the appearance information from previous frame, thereby sequentially generating new objects with geometric consistency across time. In specific, we employ canny edge as structure guidance, which is also adopted by existing diffusion methods [55]. Another important question is how to propagate the information of one object across frames to achieve consistent appearance. With the help of NLA, we can transfer the appearance features of the overlapping parts of previous frame to the next frame. Inspired by SDEdit [26] and ILVR [7], we further use a process of adding noise and denoising to obtain a more complete output. We illustrate this generation process in Fig. 3. The inter-frame propagation method is only applied on the foreground objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Specifically, we first select \\(N\\) foreground key frames from the original video \\(I\\), ensuring that 1) there is significant overlap between the adjacent frames, and 2) these key frames capture the appearance of all faces of the object. Given a generator \\(\\mathcal{G}^f(\\cdot)\\) and a text prompt \\(T\\), we edit the first frame \\(F_0\\) in pixel coordinate system with its structure condition \\(C_0\\) as the extra guidance:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.815, + 0.891, + 0.832 + ], + "angle": 0, + "content": "\\[\nE _ {0} = \\mathcal {G} ^ {f} (T, C _ {0}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( E_0 \\) represents the editing result. Then for the remaining key frames, we propagate the editing result from the previous key frame \\( E_{i-1} \\) to obtain the one of the current key frame \\( E_i \\). To be specific, we map \\( E_{i-1} \\) from the pixel" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23043" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.093, + 0.09, + 0.435, + 0.105 + ], + "angle": 0, + "content": "Foreground: a polar bear; Background: north pole." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.107, + 0.887, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.246, + 0.475, + 0.262 + ], + "angle": 0, + "content": "Foreground: a car with graffiti; Background: Miami city." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.263, + 0.887, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.328, + 0.887, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.401, + 0.895, + 0.434 + ], + "angle": 0, + "content": "Figure 5: Compositing editing. We demonstrate two editing examples with non-rigid and rigid foreground objects. Our approach can well preserve the geometry of \"bear\" and \"car\" across frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.471, + 0.507 + ], + "angle": 0, + "content": "coordinate to atlas \\(A_{i - 1}^{E}\\) and then map it back to the pixel coordinate in the current frame \\(i\\) as \\(\\hat{E}_i\\) with multiplying the opacity \\(\\alpha_{i}\\), formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.518, + 0.469, + 0.546 + ], + "angle": 0, + "content": "\\[\nA _ {i - 1} ^ {f} = \\left(\\mathrm {U V} _ {i - 1} ^ {f}\\right) ^ {- 1} \\left(E _ {i - 1}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.54, + 0.355, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\hat {E} _ {i} = \\alpha_ {i} \\circ \\mathrm {U V} _ {i} ^ {f} (A _ {i - 1} ^ {f}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.47, + 0.661 + ], + "angle": 0, + "content": "It is noteworthy that the entire video shares the same foreground atlas \\( A^f \\) for each target object, where the subscript \\( i \\) represents an incomplete partial atlas in frame \\( i \\). We then use the atlas to propagate the pixel values from previous frame to their corresponding positions in the current frame to obtain an incomplete partial appearance \\( \\hat{E}_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.662, + 0.47, + 0.723 + ], + "angle": 0, + "content": "Given the partial appearance \\(\\hat{E}_i\\), we first encode it with VQ-VAE [47] to get the latent representation \\(\\hat{Z}_i\\), and then add noise to it with Variance Preserving Stochastic Differential Equation (VP-SDE), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.734, + 0.47, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\hat {Z} _ {i} (t _ {0}) = \\alpha (t _ {0}) \\hat {Z} _ {t} (0) + \\sigma (t _ {0}) \\mathbf {z}, \\quad \\mathbf {z} \\sim \\mathcal {N} (0, \\mathbf {I}) \\quad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\(\\sigma(t_0)\\) and \\(\\alpha(t_0)\\) are two scalar functions that satisfy \\(\\alpha^2(t) + \\sigma^2(t) = 1, \\forall t \\in (0,1]\\), and \\(t_0 \\in [0,1]\\) is a hyper parameter of the noise strength. Then we apply denoising process \\(\\hat{Z}_i(t_0)\\) under the condition guidance from both text prompt \\(T\\) and structure guidance \\(C_i\\) to get the latent representation \\(Z_i\\). Finally, we decode the latent representation to \\(Z_i\\) propagate editing result \\(E_i\\). Our experiments further demonstrate that this mechanism can achieve good propagation results without training or fine-tuning the model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.457, + 0.704, + 0.474 + ], + "angle": 0, + "content": "3.3. Aggregation Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Different from [1] and [24], our approach edits video frames rather than atlases, which have the chance to achieve more information of different viewpoints. This brings two advantages. Firstly, the geometries and pixels from different viewpoints provide more details of the target objects, allowing the diffusion model to generate the edited content with higher fidelity. Secondly, this alleviates the risk of failure editing due to the potential wrong mapping from the atlas to the video frames. We then aggregate the edited key frames by using a simple yet effective two-layer 2D convolution network with skip connection as shown in Fig. 4. Our goal is to guarantee that the aggregated atlas is highly aligned with the original one, in terms of locations, so that appearance edit will not affect the geometric consistency and the temporal continuity. Reconstruction loss, \\(\\mathcal{L}_{rec}\\), between the edited and reconstructed key frames is employed in the training process as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.746, + 0.892, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\sum_ {i = 1} ^ {N} | | E _ {i} - \\mathrm {U V} _ {i} ^ {f} \\left(A ^ {f}\\right) | | _ {1}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.795, + 0.754, + 0.81 + ], + "angle": 0, + "content": "where \\(N\\) is the number of key frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.823, + 0.634, + 0.84 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.709, + 0.865 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In practice, we implement our approach over Stable Diffusion [36]. Despite there are several image-based meth" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "23044" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.335, + 0.107 + ], + "angle": 0, + "content": "Background: desert scene, pyramid." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.109, + 0.891, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.174, + 0.888, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.246, + 0.359, + 0.261 + ], + "angle": 0, + "content": "Background: magma scene, crack road." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.265, + 0.888, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.332, + 0.888, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.406, + 0.893, + 0.438 + ], + "angle": 0, + "content": "Figure 6: Background replacement. Since our approach can effectively maintain the geometry of the foreground, it can perform background replacement while maintaining geometric consistency of depth and temporal continuity of perspective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.473, + 0.598 + ], + "angle": 0, + "content": "ods [31, 45, 28] can perform structure-preserving editing, we choose the canny condition branch from [55] as the structure guidance for the proposed inter-frame propagation in our method. We apply our method on several videos from DAVIS [33], with each video containing a moving object in \\(50 \\sim 70\\) frames. The image resolution is set to \\(768 \\times 432\\), and the resolution of foreground atlas is set to \\(2000 \\times 2000\\). We employ DDIM [43] sampler with 20 steps. In this case, our method requires only \\(\\sim 10\\) GB GPU memory and takes \\(\\sim 30\\) seconds for each video in a single NVIDIA A40 GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.233, + 0.621 + ], + "angle": 0, + "content": "4.2. Editing Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.657 + ], + "angle": 0, + "content": "We tested our method for various editing types among several videos. Here we demonstrate several scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.703 + ], + "angle": 0, + "content": "Compositing editing. Our method can edit foreground and background separately to achieve high-quality and semantically matched editing results as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Background replacement. Due to the separate editing capabilities enabled by NLA, as well as the diverse and high-quality editing brought by the diffusion model, our method achieves high-quality video background replacement while maintaining geometric consistency of depth and temporal continuity of perspective, as shown in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Style transfer. Our method accomplishes a wide range of style transfers, while simultaneously ensuring temporal consistency, as shown in Fig. 7." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.313, + 0.864 + ], + "angle": 0, + "content": "4.3. Comparison to Prior Arts" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this section, we compare our editing results with state-of-the-art atlas-based method Text2LIVE [1] and diffusion-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.446, + 0.729, + 0.461 + ], + "angle": 0, + "content": "based method Tune-A-Video [49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.893, + 0.598 + ], + "angle": 0, + "content": "Comparison to Text2LIVE [1]. As shown in Fig. 8, given the text prompt \"an orange SUV\", Text2LIVE shows incomplete editing, while our method demonstrates a much more holistic result. It is because our method employs key frame editing, while Text2LIVE creates a new layer to edit the atlas somehow directly. The potential failure editing in atlas will lead to the error propagation to the entire video. In addition, compared to Text2LIVE, we can achieve higher quality and richer content with faster inference speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Comparison to Tune-A-Video [49]. As shown in Fig. 9, vanilla video diffusion models like Tune-A-Video often fail in video editing. While it adeptly captures the semantic information of text prompts, it struggles to preserve consistency in video layout and object geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Consistency analysis. To the best of our knowledge, there is currently no widely accepted metric for evaluating the geometric and temporal consistency of videos. In this paper, we employ motion consistency of dense optical flow and deviation consistency of the edited video frames for this metric. For motion consistency, we employ the Farneback algorithm in OpenCV to calculate the average L2 distance of dense optical flow between the edited and original videos. We select the \"car-turn\" video from DAVIS and apply the same prompt to all methods. The experiment is repeated several times to obtain the average number. Our method shows better stability than Tune-A-Video as shown in Tab. 1. For deviation consistency, we conduct experiments on 1) CLIP score: target text faithfulness 2) LPIPS-P: deviation from the original video frames and 3) LPIPS-T:" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23045" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.093, + 0.091, + 0.495, + 0.107 + ], + "angle": 0, + "content": "Foreground: a panda; Both: in the style of Chinese painting." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.109, + 0.887, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.174, + 0.887, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.508, + 0.264 + ], + "angle": 0, + "content": "Both: in the style of Vincent Willem van Gogh's Starry Night." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.265, + 0.887, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.329, + 0.887, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.403, + 0.873, + 0.42 + ], + "angle": 0, + "content": "Figure 7: Style transfer. Our approach achieves diverse video style transfer while ensuring high temporal consistency." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.446, + 0.468, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.544, + 0.312, + 0.556 + ], + "angle": 0, + "content": "Text2LIVE" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.564, + 0.468, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.662, + 0.334, + 0.675 + ], + "angle": 0, + "content": "StableVideo (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.688, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Figure 8: Comparison to Text2LIVE. Foreground prompt: an orange SUV. Our method achieves more holistic editing results on the foreground." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.471, + 0.793 + ], + "angle": 0, + "content": "deviation between adjacent frames, as shown in Tab. 2. Our method achieves comparable CLIP score and much lower deviations, which shows the effectiveness and stability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.231, + 0.818 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To verify the necessity of the key frame editing, we apply editing in atlas layer directly for the foreground as a simple baseline. The atlas might not be so deformed for human perception, but it significantly affects the diffusion models. Fig. 10 shows an example, where obvious deformation and" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.444, + 0.89, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.49, + 0.743, + 0.503 + ], + "angle": 0, + "content": "original video" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.505, + 0.89, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.551, + 0.745, + 0.563 + ], + "angle": 0, + "content": "Tune-A-Video" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.567, + 0.89, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.612, + 0.76, + 0.626 + ], + "angle": 0, + "content": "StableVideo (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.636, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Figure 9: Comparison to Tune-A-Video. Prompt: an orange SUV in sunny snowy winter, cabins. Our method achieves much more consistent editing results." + }, + { + "type": "title", + "bbox": [ + 0.503, + 0.697, + 0.573, + 0.709 + ], + "angle": 0, + "content": "Editing atlas" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.71, + 0.892, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.742, + 0.606, + 0.754 + ], + "angle": 0, + "content": "Editing key frames" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.754, + 0.892, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.797, + 0.891, + 0.826 + ], + "angle": 0, + "content": "Figure 10: Ablation study on directly editing the atlas. The deformation in atlas affects the diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.631, + 0.87 + ], + "angle": 0, + "content": "inconsistency exist." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We also conduct extensive ablation study on inter-frame propagation module. The objective of this module is to" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23046" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.466, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.135, + 0.333, + 0.147 + ], + "angle": 0, + "content": "random generation" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.155, + 0.466, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.199, + 0.323, + 0.21 + ], + "angle": 0, + "content": "fix initial latent" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.219, + 0.466, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.264, + 0.341, + 0.276 + ], + "angle": 0, + "content": "sequential generation" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.284, + 0.466, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.181, + 0.327, + 0.366, + 0.34 + ], + "angle": 0, + "content": "Inter-frame Propagation (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.47, + 0.398 + ], + "angle": 0, + "content": "Figure 11: Ablation study on inter-frame propagation module. Foreground prompt: an orange SUV. Our method achieves excellent consistency in key frame appearance." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.416, + 0.468, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.47, + 0.469, + 0.547 + ], + "angle": 0, + "content": "Figure 12: Binary search on hyper parameter \\( t_0 \\) in interframe propagation module. The appearance of previous frame is shown on the right side. The sequence on the left shows the generated results of the current frame as a function of \\( t_0 \\). The value of \\( t_0 \\) around 0.8 is a good sweet spot." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.469, + 0.608 + ], + "angle": 0, + "content": "maintain the geometry of the foreground when editing key frames. Firstly, we consider four different settings for editing key frames as shown in Fig. 11." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.469, + 0.655 + ], + "angle": 0, + "content": "Random generation. Each key frame only shares the same text prompt with the others. In this case, there are significant differences among the generated key frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.469, + 0.732 + ], + "angle": 0, + "content": "Fix initial latent. Unlike starting from random noise every time we edit, we start generating each key frame from the same latent noise and share the text prompt. In this case, there is higher similarity in the content generated for each frame, but the consistency is still not satisfactory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.469, + 0.808 + ], + "angle": 0, + "content": "Sequential generation. Furthermore, we concatenate the latent noise between frames. Specifically, we apply imaged-to-image translation between frames. This method still cannot guarantee consistency since the appearances of the objects between the two frames do not match." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.809, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Inter-frame propagation (ours). Our final approach is to employ partial atlas to geometrically align the appearances between two frames, followed by a process of adding noise and then apply denoising process." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In addition, We also conducted experiments on the selection of the hyper parameter \\( t_0 \\) in this module as shown in" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.09, + 0.869, + 0.155 + ], + "angle": 0, + "content": "
MethodForeground EditingComposite Editing
Text2LIVE [1]1.997.39
Tune-A-Video [49]4.6312.74
StableVideo (ours)3.3411.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.165, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Table 1: Consistency of dense optical flow. Original video: car-turn. Foreground prompt (column 1): an orange suv. Composite prompt (column 2): a car driving in winter." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.225, + 0.869, + 0.281 + ], + "angle": 0, + "content": "
MethodCLIP (↑)LPIPS-P (↓)LPIPS-T (↓)
Tune-A-Video [49]0.27870.63460.1851
StableVideo (ours)0.27130.16130.0386
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.292, + 0.89, + 0.322 + ], + "angle": 0, + "content": "Table 2: Quantitative results. Original video: car-turn. Prompt: an orange SUV in the winter." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.337, + 0.891, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.408, + 0.892, + 0.455 + ], + "angle": 0, + "content": "Figure 13: Failure case. Videos with non-rigid deformation may lead to failure editing, since the movement of the object is more difficult to be well captured in this case." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.469, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Fig. 12. We observe that as \\( t_0 \\) gradually increases, the generated results become more realistic but gradually lose their match with the appearance of the previous frame. Using binary search, we find out that a reasonable trade-off between fidelity and realism for \\( t_0 \\) lies around 0.8." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.78, + 0.573 + ], + "angle": 0, + "content": "5. Limitations and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Firstly, our method is constrained by NLA. Learning atlas layers may fail for non-rigid objects with significant structural deformation as shown in Fig. 13. While we can mitigate this by dividing long videos into short clips where the objects can be considered to be rigid, it is still not feasible to address every single case. Secondly, our method is constrained by the capabilities of the diffusion models, which may struggle with specific scenarios such as human or animals. Besides, it may be better to optimize the diffusion model with the objective of aligning the generated contents to the reconstructed ones." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.619, + 0.779 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.788, + 0.892, + 0.895 + ], + "angle": 0, + "content": "We have proposed a text-driven diffusion video editing approach. To solve the consistency problem for diffusion models in foreground object editing, we propose a interframe propagation mechanism and an atlas aggregation network. We conducted extensive experiments and demonstrated the superior qualitative and quantitative results of our method compared to state-of-the-art approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "23047" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.314 + ], + "angle": 0, + "content": "[3] Shidong Cao, Wenhao Chai, Shengyu Hao, and Gaoang Wang. Image reference-guided fashion design with structure-aware transfer by diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3524-3528, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[4] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. arXiv preprint arXiv:2302.06826, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.415 + ], + "angle": 0, + "content": "[5] Wenhao Chai and Gaoang Wang. Deep vision multimodal learning: Methodology, benchmark, and trend. Applied Sciences, 12(13):6588, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.472, + 0.472 + ], + "angle": 0, + "content": "[6] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.472, + 0.529 + ], + "angle": 0, + "content": "[7] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[8] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[10] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[14] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:210.02303, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[18] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[19] Ondřej Jamriška, Šárka Sochorová, Ondřej Texler, Michal Lukáč, Jakub Fiser, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Stylizing video by example. ACM Transactions on Graphics (TOG), 38(4):1-11, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[20] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[22] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. arXiv preprint arXiv:2301.07093, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.647 + ], + "angle": 0, + "content": "[23] Weng Fei Low and Gim Hee Lee. Minimal neural atlas: Parameterizing complex surfaces with minimal charts and distortion. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 465-481. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[24] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. arXiv preprint arXiv:2009.07833, 2020. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[25] Erika Lu, Forrester Cole, Weidi Xie, Tali Dekel, William T Freeman, Andrew Zisserman, and Michael Rubinstein. Associating objects and their effects in video through coordination games. In Advances in Neural Information Processing Systems. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "23048" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[29] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[30] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[31] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. arXiv preprint arXiv:2302.03027, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[32] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.447 + ], + "angle": 0, + "content": "[33] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.47, + 0.531 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[35] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[37] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In *Pattern Recognition: 38th German Conference*, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings 38, pages 26-36. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.903 + ], + "angle": 0, + "content": "[40] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[41] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[44] Ondrej Texler, David Futschik, Michal Kucera, Ondrej Jamriška, Šárka Sochorová, Menclei Chai, Sergey Tulyakov, and Daniel Sykora. Interactive video stylization using few-shot patch-based training. ACM Transactions on Graphics (TOG), 39(4):73-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[45] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[46] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[47] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[48] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[50] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 357-374. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[51] Ren Yang, Radu Timofte, Xin Li, Qi Zhang, Lin Zhang, Fanglong Liu, Dongliang He, Fu Li, He Zheng, Weihang Yuan, et al. Aim 2022 challenge on super-resolution of compressed image and video: Dataset, methods and results. In European Conference on Computer Vision, pages 174-202. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[52] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[53] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. Magvit:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "23049" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Masked generative video transformer. arXiv preprint arXiv:2212.05199, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.163 + ], + "angle": 0, + "content": "[54] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. arXiv preprint arXiv:2302.07685, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[55] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[56] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23050" + } + ] +] \ No newline at end of file diff --git a/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/bcc42e2b-6089-4c27-a5c2-cdec53620ca0_origin.pdf b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/bcc42e2b-6089-4c27-a5c2-cdec53620ca0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..235022c90dcd88ff07986f90ae6f5356273696b6 --- /dev/null +++ b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/bcc42e2b-6089-4c27-a5c2-cdec53620ca0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a902d0c5e982ad6c6e7b2564f6d6b9037192499e9ae8144e831fb0864432c73 +size 8497102 diff --git a/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/full.md b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..db2842291d46c49be80374f69c946b6e7397925d --- /dev/null +++ b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/full.md @@ -0,0 +1,351 @@ +# StableVideo: Text-driven Consistency-aware Diffusion Video Editing + +Wenhao Chai* Xun Guo² Gaoang Wang¹ Yan Lu² + +1 Zhejiang University 2 Microsoft Research Asia + +{wenhaochai.19, gaoangwang}@intl.zju.edu.cn, {xunguo, yanlu}@microsoft.com + +![](images/69efd36d46271bdad74017f4791141c4003d8b37d502048e67705fe660cd0d42.jpg) +Figure 1: Editing results of StableVideo. The input video (top row) contains long range motion and viewpoint changing. Our approach performs stable editing according to the prompt of "Orange SUV in sunny snow winter" on foreground and background. In the edited video (bottom row), the "orange SUV" maintains high geometric and temporal consistency, although the viewpoints keep changing. + +# Abstract + +Diffusion-based methods can generate realistic images and videos, but they struggle to edit existing objects in a video while preserving their appearance over time. This prevents diffusion models from being applied to natural video editing in practical scenarios. In this paper, we tackle this problem by introducing temporal dependency to existing text-driven diffusion models, which allows them to generate consistent appearance for the edited objects. Specifically, we develop a novel inter-frame propagation mechanism for diffusion video editing, which leverages the concept of layered representations to propagate the appearance information from one frame to the next. We then build up a text-driven video editing framework based on this mechanism, namely StableVideo, which can achieve consistency-aware video editing. Extensive experiments demonstrate the strong editing capability of our approach. Compared with state-of-the-art video editing methods, our approach shows superior qualitative and quantitative results. Our code is available at this https URL. + +# 1. Introduction + +Recent years have witnessed significant progress in extensive computer vision tasks taken by deep learning. Nev- + +ertheless, natural video editing, which aims at manipulating the appearance of target objects and scenes, still faces two essential challenges that are deterministic to the editing quality: the generator equipped with rich prior knowledge that consistently produces high-fidelity edited contents adhering faithfully to the original geometry of the target objects, and the propagator that disseminates the edited contents throughout the entire video while keeping highly temporal consistency. + +The flourish of text-driven generative diffusion models pre-trained on large-scale image and language data [34, 14, 53, 16, 41, 5] provides impressive generation quality. Several diffusion-based methods achieve good performance in image editing [2, 31], but few methods have tried to apply diffusion models in video editing, since it is challenging to modify existing objects while preserving their appearance over the entire video [49, 12, 26]. Dreamix [27] proposes a solution to generate consistent video according to input image/video and prompts. However, it focuses more on generating smooth motions, e.g., pose and camera movements, rather than maintaining geometric consistency of the objects across time. Moreover, such video diffusion models often suffer from huge computing complexity which is not friendly for practical applications. + +Neural layered atlas (NLA) [24, 23] tries to tackle the temporal continuity problem by decomposing the video into a set of atlas layers, each of which describes one target object to be edited. For each atlas layer, the positions of the + +video are mapped into the corresponding 2D positions in it, so that semantically correspondent pixels over the whole video can be represented by the same atlas position. Instead of frame-by-frame editing, NLA edits atlas layers to ensure that the modifications can be precisely mapped back to video frames for temporal smoothness. Text2LIVE [1] provides a text-driven appearance manipulation solution of adding additional edit layers on atlases, in which a specific generator for the edit layers is trained. Although it achieves good results with strict structure preserved, it is not able to apply thorough editing. Moreover, the specifically trained generator also limits the richness of the generated contents. + +This brings up the question: Could text-driven diffusion video editing achieve high temporal consistency? Intuitively, employing text-driven diffusion models to edit the atlases corresponding to the target objects could reach such goal. However, this gives rise to drawbacks rather than benefits. Being the summary of the whole video, atlases always have distorted appearance due to the viewpoint and camera movement, which are required to be specifically pretrained and generated as in [1]. Diffusion models may fail in generating satisfied atlas pixels in many cases, so that the corresponding edited frames will also be contaminated. To answer the question, we present two concepts for utilizing diffusion models in video editing. Firstly, instead of editing the atlases directly, we propose to update the atlases via editing key video frames. Secondly, we introduce temporal dependency constraints for diffusion models to generate objects with consistent appearance across time. + +Based on analysis above, we present a novel diffusion video editing approach, StableVideo, to perform consistency-aware video editing. In specific, we propose two effective technologies for this purpose. Firstly, to edit the objects with consistent appearance, we design an interframe propagation mechanism on top of the existing diffusion model [55], which can generate new objects with coherent geometry across time. Secondly, to achieve temporal consistency by leveraging NLA, we design an aggregation network to generate the edited atlases from the key frames. We then build up a text-driven diffusion-based framework, which provides high-quality natural video editing. We conduct extensive qualitative and quantitative experiments to demonstrate the capability of our approach. Compared with state-of-the-art methods, our approach achieves superior results with much lower complexity. + +In summary, we present the following contributions: + +- To our best knowledge, we are the first to solve the consistency problem of diffusion video editing by considering the concept of layered atlas approaches, which provides an efficient and effective way for this topic. +- We present a new video editing framework which can manipulate the appearance of the objects with high ge + +ometry and appearance consistency across time. Our method can be easily applied to other text-driven diffusion models. + +- We conduct extensive experiments on a variety of natural videos, which shows superior editing performance compared with state-of-the-art methods. + +# 2. Related Work + +# 2.1. Diffusion for Image Editing + +The editing of natural images is an important task in the field of computer vision that has been widely studied. Prior to the emergence of diffusion models [42, 15], many GAN-based approaches [11, 10, 30, 32, 48] have achieved good results. There are also some works focus on low-level editing [52, 6, 51]. The advent of diffusion models has made it possible to achieve even higher quality and more diverse edited contents. SDEdit [26] adds noise and corruptions to an input image and uses diffusion models to reverse the process for image editing, while suffering from the loss of fidelity. Prompt-to-Prompt [12] and Plug-and-Play [45] perform semantic editing by mixing activations from original and target text prompts. InstructPix2Pix [2] applies semantic editing at test time and personalizes the model through finetuning and optimization to learn a special token describing the content. UniTune [46] and Imagic [21] finetune on a single image for better editability while maintaining good fidelity. There are also some works exploring the controllability [55, 18, 39, 22, 4, 3] and personalization [9, 38, 29] of diffusion-based generation. Our proposed video editing method leverages existing image editing methods that can preserve the structures, such as [55, 31, 45] and [28]. + +# 2.2. Diffusion for Video Editing + +Compared to image editing, video editing is more challenging for diffusion-based methods for geometric and temporal consistency. Tune-a-Video [49] inflates a text-to-image model for video editing. However, since temporal correlations are not fully considered, the editing results suffer from inconsistency of geometry and motion. Dreamix [27] develops a text-to-video backbone for motion editing while maintaining temporal consistency. There are also some works based on video generation like [8, 40, 13, 54, 56] and [17]. Unlike these approaches, our purpose is to enable diffusion models to perform appearance editing with both geometric and temporal consistency. + +# 2.3. Temporal Propagation in Video Editing + +Temporal propagation plays an important role in natural video process, since it is the essential factor for temporal consistency. Some methods rely on key frames [19, 44, 50] or optical flow [37] to propagate contents between frames. + +![](images/233f6ea506e4a63c7c632db96f4dde7989e030a53d016bc9e053ab0f0be9244f.jpg) +Figure 2: Framework of the proposed StableVideo. The input video is first fed into NLA [24] to generate foreground and background atlases using the pre-trained model. $\mathcal{G}_b$ is the diffusion model used to edit background atlas, and $\mathcal{G}_f$ is used to edit the foreground key frames. Note that $\mathcal{G}_b$ and $\mathcal{G}_f$ share the same weights, but accept different conditions. We employ depth information, extracted by MiDaS [35], for $\mathcal{G}_b$ to maintain the consistency between the foreground motion and the environment, while structure guidance is used for $\mathcal{G}_f$ to keep geometric consistency between the new generated foreground and the old one. After being edited, the foreground and background are blended together to reconstruct the edited frames. + +Another bunch of methods are to achieve consistent interframe editing by forming a compressed representation of a video. Omnimatte [24, 25] estimate RGBA layers for target subject and scene effects for each frame independently, but cannot achieve consistent propagation of contents along temporal direction. Atlas [1, 20] tackles this problem by decomposing the video into unified 2D atlas layers for each target. This approach allows contents to be applied to the global summarized 2D atlases and mapped back to the video, achieving temporal consistency with minimal effort. Inspired by the concept of atlas approach, we employ the pre-trained neural layered atlas model to solve the inconsistency problem in diffusion video editing, thereby achieving high-quality editing results with temporal coherence. + +# 3. Method + +Fig. 2 shows the pipeline of our proposed StableVideo. We utilize NLA [20] as the propagator for consistent video editing. Specifically, we conduct foreground and background editing separately. For foreground editing, we adopt key frame editing to generate atlas layers with high quality and the inter-frame propagation module to ensure better geometric and temporal consistency. The edited key frames + +are then mapped to partial atlases and aggregated by the aggregation network to produce the edited foreground atlas. It is noteworthy that our approach can also handle more than one foreground layers. + +# 3.1. Problem Formulation + +We employ the pre-trained NLA model [24] to propagate the edited contents to ensure that the target objects and scenes can maintain homogeneous appearances and motions across the entire video. The concept of NLA is to decompose the input video into layered representations, namely foreground atlas and background atlas, which globally summarize the correlated pixels for the foreground and the background, respectively. Three mapping networks, i.e., $\mathcal{M}^b (\cdot),\mathcal{M}^f (\cdot)$ and $\mathcal{M}^{\alpha}(\cdot)$ , are provided for this purpose. Given an input video $I$ , for each frame $I_{i}$ , we obtain the mapping relationships of the atlas in the background and the foreground with respect to the pixel coordinate system, named as $\mathrm{UV}^b (\cdot)$ and $\mathrm{UV}^f (\cdot)$ , as well as the foreground opacity $\alpha_{i}$ on the pixel coordinate system, formulated as: + +$$ +\operatorname {U V} _ {i} ^ {b} (\cdot) = \mathcal {M} ^ {b} (I _ {i}), \operatorname {U V} _ {i} ^ {f} (\cdot) = \mathcal {M} ^ {f} (I _ {i}), \alpha_ {i} = \mathcal {M} ^ {\alpha} (I _ {i}). \tag {1} +$$ + +![](images/d1d5f71c9c1edeafaef3c740071afd189040955fe537477a7189873addd98839.jpg) +Figure 3: Inter-frame propagation for foreground editing. We use two edited key frames, $E_{i-1}$ and $E_i$ , to illustrate the process more clearly. The structure guidance and the text prompt is added into the denoising UNet via the concatenation and cross-attention mechanism respectively. + +After that, we formulate the mappings from the atlas representation of the background $A^b$ and the foreground $A^f$ , to the pixel coordinate systems of $B_i$ and $F_i$ : + +$$ +B _ {i} = \mathrm {U V} _ {i} ^ {b} (A ^ {b}), F _ {i} = \mathrm {U V} _ {i} ^ {f} (A ^ {f}). \qquad (2) +$$ + +Our method achieves geometrical consistent editing by fixing the mappings of $\mathrm{UV}^b$ and $\mathrm{UV}^f$ , and generating the edited atlases of $A^b$ and $A^f$ . We adopt a pre-trained latent diffusion model [36] with guided conditions as our generator, namely $\mathcal{G}^b (\cdot)$ and $\mathcal{G}^f (\cdot)$ . Note that we are not simply editing the foreground and the background atlases directly. We apply inter-frame propagation mechanism on the editing process of foreground atlas. More details are explained in Sec. 3.2. After that, the entire video $I$ can be reconstructed frame by frame as the following equation: + +$$ +I _ {i} = \alpha_ {i} \circ \operatorname {U V} _ {i} ^ {f} \left(\mathcal {G} ^ {f} \left(A ^ {f}\right)\right) + \left(1 - \alpha_ {i}\right) \circ \operatorname {U V} _ {i} ^ {b} \left(\mathcal {G} ^ {b} \left(A ^ {b}\right)\right), \tag {3} +$$ + +where $\circ$ denotes pixel-wise product. + +# 3.2. Inter-frame Propagation + +In this section, we further elaborate on how inter-frame propagation mechanism helps consistent foreground editing. One of the major challenges for diffusion models is to generate video contents with temporal consistency. Existing state-of-the-art text-driven diffusion methods [55, 28] can maintain the similar geometry between the target objects and the generated ones for image editing by adding structure conditions. However, the situation is different for + +![](images/73b0403a69007c24d9ccc44e07281e8acd878e2db2a0e475e65f552827e011c7.jpg) +Figure 4: Training process of aggregation network. We employ a simple 3D network to aggregate the partial atlases generated from the edited key frames. For each training iteration, the sampled key frames are edited and mapped to partial atlases by $UV_{i}^{f}$ . The partial atlases are then fed into the Aggregation Network to generate the edit atlas, which is mapped back by $(UV_{i}^{f})^{-1}$ to generate the reconstructed frames. A $L_{1}$ loss is used to guarantee the aggregation consistency between the edited key frames and the reconstructed frames. + +videos. Generating temporally consistent geometry needs to handle some uncertain changes across time, e.g., motion and deformation, which can not be supported by them. We tackle this problem by introducing a conditional denoising process to enable the diffusion models to consider both the structure of the current frame and the appearance information from previous frame, thereby sequentially generating new objects with geometric consistency across time. In specific, we employ canny edge as structure guidance, which is also adopted by existing diffusion methods [55]. Another important question is how to propagate the information of one object across frames to achieve consistent appearance. With the help of NLA, we can transfer the appearance features of the overlapping parts of previous frame to the next frame. Inspired by SDEdit [26] and ILVR [7], we further use a process of adding noise and denoising to obtain a more complete output. We illustrate this generation process in Fig. 3. The inter-frame propagation method is only applied on the foreground objects. + +Specifically, we first select $N$ foreground key frames from the original video $I$ , ensuring that 1) there is significant overlap between the adjacent frames, and 2) these key frames capture the appearance of all faces of the object. Given a generator $\mathcal{G}^f(\cdot)$ and a text prompt $T$ , we edit the first frame $F_0$ in pixel coordinate system with its structure condition $C_0$ as the extra guidance: + +$$ +E _ {0} = \mathcal {G} ^ {f} (T, C _ {0}), \tag {4} +$$ + +where $E_0$ represents the editing result. Then for the remaining key frames, we propagate the editing result from the previous key frame $E_{i-1}$ to obtain the one of the current key frame $E_i$ . To be specific, we map $E_{i-1}$ from the pixel + +Foreground: a polar bear; Background: north pole. + +![](images/5374b288a1d996eb4934bc8d6613e530dd9d45f6e25d58094167a26c16ef2695.jpg) + +Foreground: a car with graffiti; Background: Miami city. + +![](images/8d280b1ac9351d46e152ca0e06d145bf8790b7d18180b2e6ccb7377982a96b51.jpg) + +![](images/18c530d0c83dfaf44954e1f81390142a04906fe3aae616f2d7930fa9c8931c2f.jpg) +Figure 5: Compositing editing. We demonstrate two editing examples with non-rigid and rigid foreground objects. Our approach can well preserve the geometry of "bear" and "car" across frames. + +coordinate to atlas $A_{i - 1}^{E}$ and then map it back to the pixel coordinate in the current frame $i$ as $\hat{E}_i$ with multiplying the opacity $\alpha_{i}$ , formulated by: + +$$ +A _ {i - 1} ^ {f} = \left(\mathrm {U V} _ {i - 1} ^ {f}\right) ^ {- 1} \left(E _ {i - 1}\right), \tag {5} +$$ + +$$ +\hat {E} _ {i} = \alpha_ {i} \circ \mathrm {U V} _ {i} ^ {f} (A _ {i - 1} ^ {f}). +$$ + +It is noteworthy that the entire video shares the same foreground atlas $A^f$ for each target object, where the subscript $i$ represents an incomplete partial atlas in frame $i$ . We then use the atlas to propagate the pixel values from previous frame to their corresponding positions in the current frame to obtain an incomplete partial appearance $\hat{E}_i$ . + +Given the partial appearance $\hat{E}_i$ , we first encode it with VQ-VAE [47] to get the latent representation $\hat{Z}_i$ , and then add noise to it with Variance Preserving Stochastic Differential Equation (VP-SDE), formulated as: + +$$ +\hat {Z} _ {i} (t _ {0}) = \alpha (t _ {0}) \hat {Z} _ {t} (0) + \sigma (t _ {0}) \mathbf {z}, \quad \mathbf {z} \sim \mathcal {N} (0, \mathbf {I}) \quad (6) +$$ + +where $\sigma(t_0)$ and $\alpha(t_0)$ are two scalar functions that satisfy $\alpha^2(t) + \sigma^2(t) = 1, \forall t \in (0,1]$ , and $t_0 \in [0,1]$ is a hyper parameter of the noise strength. Then we apply denoising process $\hat{Z}_i(t_0)$ under the condition guidance from both text prompt $T$ and structure guidance $C_i$ to get the latent representation $Z_i$ . Finally, we decode the latent representation to $Z_i$ propagate editing result $E_i$ . Our experiments further demonstrate that this mechanism can achieve good propagation results without training or fine-tuning the model. + +# 3.3. Aggregation Network + +Different from [1] and [24], our approach edits video frames rather than atlases, which have the chance to achieve more information of different viewpoints. This brings two advantages. Firstly, the geometries and pixels from different viewpoints provide more details of the target objects, allowing the diffusion model to generate the edited content with higher fidelity. Secondly, this alleviates the risk of failure editing due to the potential wrong mapping from the atlas to the video frames. We then aggregate the edited key frames by using a simple yet effective two-layer 2D convolution network with skip connection as shown in Fig. 4. Our goal is to guarantee that the aggregated atlas is highly aligned with the original one, in terms of locations, so that appearance edit will not affect the geometric consistency and the temporal continuity. Reconstruction loss, $\mathcal{L}_{rec}$ , between the edited and reconstructed key frames is employed in the training process as: + +$$ +\mathcal {L} _ {r e c} = \sum_ {i = 1} ^ {N} | | E _ {i} - \mathrm {U V} _ {i} ^ {f} \left(A ^ {f}\right) | | _ {1}. \tag {7} +$$ + +where $N$ is the number of key frames. + +# 4. Experiments + +# 4.1. Experimental Settings + +In practice, we implement our approach over Stable Diffusion [36]. Despite there are several image-based meth + +Background: desert scene, pyramid. + +![](images/663f260a0da3dadbde6d448edb13f668868800720d77f84c661d8fa4a70eb0cb.jpg) + +![](images/07465c1bd9aab0c9f7176266cb42fbb26f2c4ab76d3855cebb98797db63e9e85.jpg) + +Background: magma scene, crack road. + +![](images/af16a53aa50d2e156307ee56123b0389f296ec98af50aa54aec663fbb21d2c7b.jpg) + +![](images/fff5268a65ad885b9b793e22cb71867ab1f553add3d2c532e89176d50e4e329c.jpg) +Figure 6: Background replacement. Since our approach can effectively maintain the geometry of the foreground, it can perform background replacement while maintaining geometric consistency of depth and temporal continuity of perspective. + +ods [31, 45, 28] can perform structure-preserving editing, we choose the canny condition branch from [55] as the structure guidance for the proposed inter-frame propagation in our method. We apply our method on several videos from DAVIS [33], with each video containing a moving object in $50 \sim 70$ frames. The image resolution is set to $768 \times 432$ , and the resolution of foreground atlas is set to $2000 \times 2000$ . We employ DDIM [43] sampler with 20 steps. In this case, our method requires only $\sim 10$ GB GPU memory and takes $\sim 30$ seconds for each video in a single NVIDIA A40 GPU. + +# 4.2. Editing Results + +We tested our method for various editing types among several videos. Here we demonstrate several scenarios: + +Compositing editing. Our method can edit foreground and background separately to achieve high-quality and semantically matched editing results as shown in Fig. 5. + +Background replacement. Due to the separate editing capabilities enabled by NLA, as well as the diverse and high-quality editing brought by the diffusion model, our method achieves high-quality video background replacement while maintaining geometric consistency of depth and temporal continuity of perspective, as shown in Fig. 6. + +Style transfer. Our method accomplishes a wide range of style transfers, while simultaneously ensuring temporal consistency, as shown in Fig. 7. + +# 4.3. Comparison to Prior Arts + +In this section, we compare our editing results with state-of-the-art atlas-based method Text2LIVE [1] and diffusion- + +based method Tune-A-Video [49]. + +Comparison to Text2LIVE [1]. As shown in Fig. 8, given the text prompt "an orange SUV", Text2LIVE shows incomplete editing, while our method demonstrates a much more holistic result. It is because our method employs key frame editing, while Text2LIVE creates a new layer to edit the atlas somehow directly. The potential failure editing in atlas will lead to the error propagation to the entire video. In addition, compared to Text2LIVE, we can achieve higher quality and richer content with faster inference speed. + +Comparison to Tune-A-Video [49]. As shown in Fig. 9, vanilla video diffusion models like Tune-A-Video often fail in video editing. While it adeptly captures the semantic information of text prompts, it struggles to preserve consistency in video layout and object geometry. + +Consistency analysis. To the best of our knowledge, there is currently no widely accepted metric for evaluating the geometric and temporal consistency of videos. In this paper, we employ motion consistency of dense optical flow and deviation consistency of the edited video frames for this metric. For motion consistency, we employ the Farneback algorithm in OpenCV to calculate the average L2 distance of dense optical flow between the edited and original videos. We select the "car-turn" video from DAVIS and apply the same prompt to all methods. The experiment is repeated several times to obtain the average number. Our method shows better stability than Tune-A-Video as shown in Tab. 1. For deviation consistency, we conduct experiments on 1) CLIP score: target text faithfulness 2) LPIPS-P: deviation from the original video frames and 3) LPIPS-T: + +Foreground: a panda; Both: in the style of Chinese painting. + +![](images/2483d949fe37b74d3efe3b5e38a5fb432eb0613d37e570b2f406012a13d75489.jpg) + +![](images/b27afc08e297416387f9b62914f53cfb2b581d2ad4f314cd2c4d873bbe01c49c.jpg) + +Both: in the style of Vincent Willem van Gogh's Starry Night. + +![](images/1484075a1af04971cc6cb3b66f09a8529a63a4b8a693e7fcbe6b82825a3976a6.jpg) + +![](images/9e3f99ea0d6a2d87f317804c6c8fc9219ed830982162baa413adf955312ef81e.jpg) +Figure 7: Style transfer. Our approach achieves diverse video style transfer while ensuring high temporal consistency. + +![](images/fc3358983057c125467417a34372a52f5eeb97de8c0dc1adfda43f9afc96c097.jpg) +Text2LIVE + +![](images/20d3965b105977ffad705ff40b4d8152352a8dd9e425e9ec43b1b59018c0ae39.jpg) +StableVideo (ours) +Figure 8: Comparison to Text2LIVE. Foreground prompt: an orange SUV. Our method achieves more holistic editing results on the foreground. +Figure 9: Comparison to Tune-A-Video. Prompt: an orange SUV in sunny snowy winter, cabins. Our method achieves much more consistent editing results. + +deviation between adjacent frames, as shown in Tab. 2. Our method achieves comparable CLIP score and much lower deviations, which shows the effectiveness and stability. + +# 4.4. Ablation Study + +To verify the necessity of the key frame editing, we apply editing in atlas layer directly for the foreground as a simple baseline. The atlas might not be so deformed for human perception, but it significantly affects the diffusion models. Fig. 10 shows an example, where obvious deformation and + +![](images/c27349148b31ba5c0a543bf3cf3170d60fc119b40ceedd0b8c2bd7a819cd3d29.jpg) + +![](images/fd446f702b699b438815e52e29f595deb71ad72f2d4b8fc6c33dcce9da2c93e5.jpg) +original video +Tune-A-Video + +![](images/f34f2c5c00fd2dcc8563c1047e9810c3d1d0e3f61c15dde62227e918e188d704.jpg) +StableVideo (ours) + +# Editing atlas + +![](images/521d3b6d18cd5b0421b906fd8e238abf0b9610e5a3dcd36d40667d82b5143cca.jpg) + +![](images/b9c7691af5ef240ba069b8bca46d2fdefdeda2a138b7bc586d68963728772eb4.jpg) +Editing key frames +Figure 10: Ablation study on directly editing the atlas. The deformation in atlas affects the diffusion models. + +inconsistency exist. + +We also conduct extensive ablation study on inter-frame propagation module. The objective of this module is to + +![](images/4c0a28fcb55e54d53a4d4f1119b1c641b2279c0a787ce20d26d64c03ca8a7d01.jpg) +random generation + +![](images/1e3febb87f7f45873f359d1a48971cde4ea62b3c14fa5a9321d0830fdf31f60c.jpg) +fix initial latent + +![](images/0e14a33a891338e9e0fdd62d378fe1ec2aaa6fc7e9070efb55adf7f80f924251.jpg) +sequential generation + +![](images/beab8b47099ef54fbe07f10d3baf812d34f57500f54739c585205db3e0d82d73.jpg) +Inter-frame Propagation (ours) + +![](images/8453c216cf43082f1bc4176039f6fe32e174706e58aa900f729298f460d61381.jpg) +Figure 11: Ablation study on inter-frame propagation module. Foreground prompt: an orange SUV. Our method achieves excellent consistency in key frame appearance. +Figure 12: Binary search on hyper parameter $t_0$ in interframe propagation module. The appearance of previous frame is shown on the right side. The sequence on the left shows the generated results of the current frame as a function of $t_0$ . The value of $t_0$ around 0.8 is a good sweet spot. +Fig. 12. We observe that as $t_0$ gradually increases, the generated results become more realistic but gradually lose their match with the appearance of the previous frame. Using binary search, we find out that a reasonable trade-off between fidelity and realism for $t_0$ lies around 0.8. + +maintain the geometry of the foreground when editing key frames. Firstly, we consider four different settings for editing key frames as shown in Fig. 11. + +Random generation. Each key frame only shares the same text prompt with the others. In this case, there are significant differences among the generated key frames. + +Fix initial latent. Unlike starting from random noise every time we edit, we start generating each key frame from the same latent noise and share the text prompt. In this case, there is higher similarity in the content generated for each frame, but the consistency is still not satisfactory. + +Sequential generation. Furthermore, we concatenate the latent noise between frames. Specifically, we apply imaged-to-image translation between frames. This method still cannot guarantee consistency since the appearances of the objects between the two frames do not match. + +Inter-frame propagation (ours). Our final approach is to employ partial atlas to geometrically align the appearances between two frames, followed by a process of adding noise and then apply denoising process. + +In addition, We also conducted experiments on the selection of the hyper parameter $t_0$ in this module as shown in + +
MethodForeground EditingComposite Editing
Text2LIVE [1]1.997.39
Tune-A-Video [49]4.6312.74
StableVideo (ours)3.3411.48
+ +Table 1: Consistency of dense optical flow. Original video: car-turn. Foreground prompt (column 1): an orange suv. Composite prompt (column 2): a car driving in winter. + +
MethodCLIP (↑)LPIPS-P (↓)LPIPS-T (↓)
Tune-A-Video [49]0.27870.63460.1851
StableVideo (ours)0.27130.16130.0386
+ +Table 2: Quantitative results. Original video: car-turn. Prompt: an orange SUV in the winter. + +![](images/95117529b88d127ce7f948386fcec03fe288ebe71d53380335f9a7a79db417dc.jpg) +Figure 13: Failure case. Videos with non-rigid deformation may lead to failure editing, since the movement of the object is more difficult to be well captured in this case. + +# 5. Limitations and Future Works + +Firstly, our method is constrained by NLA. Learning atlas layers may fail for non-rigid objects with significant structural deformation as shown in Fig. 13. While we can mitigate this by dividing long videos into short clips where the objects can be considered to be rigid, it is still not feasible to address every single case. Secondly, our method is constrained by the capabilities of the diffusion models, which may struggle with specific scenarios such as human or animals. Besides, it may be better to optimize the diffusion model with the objective of aligning the generated contents to the reconstructed ones. + +# 6. Conclusion + +We have proposed a text-driven diffusion video editing approach. To solve the consistency problem for diffusion models in foreground object editing, we propose a interframe propagation mechanism and an atlas aggregation network. We conducted extensive experiments and demonstrated the superior qualitative and quantitative results of our method compared to state-of-the-art approaches. + +# References + +[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 2, 3, 5, 6, 8 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 1, 2 +[3] Shidong Cao, Wenhao Chai, Shengyu Hao, and Gaoang Wang. Image reference-guided fashion design with structure-aware transfer by diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3524-3528, 2023. 2 +[4] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. arXiv preprint arXiv:2302.06826, 2023. 2 +[5] Wenhao Chai and Gaoang Wang. Deep vision multimodal learning: Methodology, benchmark, and trend. Applied Sciences, 12(13):6588, 2022. 1 +[6] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2 +[7] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 4 +[8] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 2 +[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[10] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2 +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2 +[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 + +[14] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:210.02303, 2022. 1 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2 +[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1 +[17] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. 2 +[18] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 2 +[19] Ondřej Jamriška, Šárka Sochorová, Ondřej Texler, Michal Lukáč, Jakub Fiser, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Stylizing video by example. ACM Transactions on Graphics (TOG), 38(4):1-11, 2019. 2 +[20] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 3 +[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2 +[22] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. arXiv preprint arXiv:2301.07093, 2023. 2 +[23] Weng Fei Low and Gim Hee Lee. Minimal neural atlas: Parameterizing complex surfaces with minimal charts and distortion. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 465-481. Springer, 2022. 1 +[24] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. arXiv preprint arXiv:2009.07833, 2020. 1, 3, 5 +[25] Erika Lu, Forrester Cole, Weidi Xie, Tali Dekel, William T Freeman, Andrew Zisserman, and Michael Rubinstein. Associating objects and their effects in video through coordination games. In Advances in Neural Information Processing Systems. 3 +[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2, 4 +[27] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1, 2 + +[28] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2, 4, 6 +[29] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2 +[30] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 2 +[31] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. arXiv preprint arXiv:2302.03027, 2023. 1, 2, 6 +[32] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2 +[33] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 6 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1 +[35] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 3 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 4, 5 +[37] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In *Pattern Recognition: 38th German Conference*, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings 38, pages 26-36. Springer, 2016. 2 +[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2 +[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2 +[40] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, + +Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[41] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1 +[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations. 2 +[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6 +[44] Ondrej Texler, David Futschik, Michal Kucera, Ondrej Jamriška, Šárka Sochorová, Menclei Chai, Sergey Tulyakov, and Daniel Sykora. Interactive video stylization using few-shot patch-based training. ACM Transactions on Graphics (TOG), 39(4):73-1, 2020. 2 +[45] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 6 +[46] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 2 +[47] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5 +[48] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2 +[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 1, 2, 6, 8 +[50] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 357-374. Springer, 2022. 2 +[51] Ren Yang, Radu Timofte, Xin Li, Qi Zhang, Lin Zhang, Fanglong Liu, Dongliang He, Fu Li, He Zheng, Weihang Yuan, et al. Aim 2022 challenge on super-resolution of compressed image and video: Dataset, methods and results. In European Conference on Computer Vision, pages 174-202. Springer, 2022. 2 +[52] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2 +[53] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. Magvit: + +Masked generative video transformer. arXiv preprint arXiv:2212.05199, 2022. 1 +[54] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. arXiv preprint arXiv:2302.07685, 2023. 2 +[55] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2, 4, 6 +[56] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2 \ No newline at end of file diff --git a/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/images.zip b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..548f49ac3e633a3053af46a42eed59edfd98ed76 --- /dev/null +++ b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea74fc381604b1647967bbb8067a70fc5adb80281af1a487debdd280bdba6734 +size 975059 diff --git a/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/layout.json b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..78e29392b1b3e155e1a60419d756ee66d3199432 --- /dev/null +++ b/2023/StableVideo_ Text-driven Consistency-aware Diffusion Video Editing/layout.json @@ -0,0 +1,9334 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 84, + 103, + 509, + 121 + ], + "type": "text", + "content": "StableVideo: Text-driven Consistency-aware Diffusion Video Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 142, + 441, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 142, + 441, + 159 + ], + "spans": [ + { + "bbox": [ + 151, + 142, + 441, + 159 + ], + "type": "text", + "content": "Wenhao Chai* Xun Guo² Gaoang Wang¹ Yan Lu²" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 162, + 421, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 162, + 421, + 177 + ], + "spans": [ + { + "bbox": [ + 173, + 162, + 421, + 177 + ], + "type": "text", + "content": "1 Zhejiang University 2 Microsoft Research Asia" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 185, + 496, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 185, + 496, + 197 + ], + "spans": [ + { + "bbox": [ + 97, + 185, + 496, + 197 + ], + "type": "text", + "content": "{wenhaochai.19, gaoangwang}@intl.zju.edu.cn, {xunguo, yanlu}@microsoft.com" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 205, + 542, + 301 + ], + "blocks": [ + { + "bbox": [ + 52, + 205, + 542, + 301 + ], + "lines": [ + { + "bbox": [ + 52, + 205, + 542, + 301 + ], + "spans": [ + { + "bbox": [ + 52, + 205, + 542, + 301 + ], + "type": "image", + "image_path": "69efd36d46271bdad74017f4791141c4003d8b37d502048e67705fe660cd0d42.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 311, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 311, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 546, + 361 + ], + "type": "text", + "content": "Figure 1: Editing results of StableVideo. The input video (top row) contains long range motion and viewpoint changing. Our approach performs stable editing according to the prompt of \"Orange SUV in sunny snow winter\" on foreground and background. In the edited video (bottom row), the \"orange SUV\" maintains high geometric and temporal consistency, although the viewpoints keep changing." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 370, + 192, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 370, + 192, + 382 + ], + "spans": [ + { + "bbox": [ + 143, + 370, + 192, + 382 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 396, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 288, + 624 + ], + "type": "text", + "content": "Diffusion-based methods can generate realistic images and videos, but they struggle to edit existing objects in a video while preserving their appearance over time. This prevents diffusion models from being applied to natural video editing in practical scenarios. In this paper, we tackle this problem by introducing temporal dependency to existing text-driven diffusion models, which allows them to generate consistent appearance for the edited objects. Specifically, we develop a novel inter-frame propagation mechanism for diffusion video editing, which leverages the concept of layered representations to propagate the appearance information from one frame to the next. We then build up a text-driven video editing framework based on this mechanism, namely StableVideo, which can achieve consistency-aware video editing. Extensive experiments demonstrate the strong editing capability of our approach. Compared with state-of-the-art video editing methods, our approach shows superior qualitative and quantitative results. Our code is available at this https URL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "type": "text", + "content": "Recent years have witnessed significant progress in extensive computer vision tasks taken by deep learning. Nev-" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 371, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 545, + 479 + ], + "type": "text", + "content": "ertheless, natural video editing, which aims at manipulating the appearance of target objects and scenes, still faces two essential challenges that are deterministic to the editing quality: the generator equipped with rich prior knowledge that consistently produces high-fidelity edited contents adhering faithfully to the original geometry of the target objects, and the propagator that disseminates the edited contents throughout the entire video while keeping highly temporal consistency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 482, + 546, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 546, + 662 + ], + "type": "text", + "content": "The flourish of text-driven generative diffusion models pre-trained on large-scale image and language data [34, 14, 53, 16, 41, 5] provides impressive generation quality. Several diffusion-based methods achieve good performance in image editing [2, 31], but few methods have tried to apply diffusion models in video editing, since it is challenging to modify existing objects while preserving their appearance over the entire video [49, 12, 26]. Dreamix [27] proposes a solution to generate consistent video according to input image/video and prompts. However, it focuses more on generating smooth motions, e.g., pose and camera movements, rather than maintaining geometric consistency of the objects across time. Moreover, such video diffusion models often suffer from huge computing complexity which is not friendly for practical applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Neural layered atlas (NLA) [24, 23] tries to tackle the temporal continuity problem by decomposing the video into a set of atlas layers, each of which describes one target object to be edited. For each atlas layer, the positions of the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 276, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 276, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 276, + 712 + ], + "type": "text", + "content": "*The work was done when the author was with MSRA as an intern." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23040" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "video are mapped into the corresponding 2D positions in it, so that semantically correspondent pixels over the whole video can be represented by the same atlas position. Instead of frame-by-frame editing, NLA edits atlas layers to ensure that the modifications can be precisely mapped back to video frames for temporal smoothness. Text2LIVE [1] provides a text-driven appearance manipulation solution of adding additional edit layers on atlases, in which a specific generator for the edit layers is trained. Although it achieves good results with strict structure preserved, it is not able to apply thorough editing. Moreover, the specifically trained generator also limits the richness of the generated contents." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 418 + ], + "type": "text", + "content": "This brings up the question: Could text-driven diffusion video editing achieve high temporal consistency? Intuitively, employing text-driven diffusion models to edit the atlases corresponding to the target objects could reach such goal. However, this gives rise to drawbacks rather than benefits. Being the summary of the whole video, atlases always have distorted appearance due to the viewpoint and camera movement, which are required to be specifically pretrained and generated as in [1]. Diffusion models may fail in generating satisfied atlas pixels in many cases, so that the corresponding edited frames will also be contaminated. To answer the question, we present two concepts for utilizing diffusion models in video editing. Firstly, instead of editing the atlases directly, we propose to update the atlases via editing key video frames. Secondly, we introduce temporal dependency constraints for diffusion models to generate objects with consistent appearance across time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 420, + 289, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 289, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 289, + 612 + ], + "type": "text", + "content": "Based on analysis above, we present a novel diffusion video editing approach, StableVideo, to perform consistency-aware video editing. In specific, we propose two effective technologies for this purpose. Firstly, to edit the objects with consistent appearance, we design an interframe propagation mechanism on top of the existing diffusion model [55], which can generate new objects with coherent geometry across time. Secondly, to achieve temporal consistency by leveraging NLA, we design an aggregation network to generate the edited atlases from the key frames. We then build up a text-driven diffusion-based framework, which provides high-quality natural video editing. We conduct extensive qualitative and quantitative experiments to demonstrate the capability of our approach. Compared with state-of-the-art methods, our approach achieves superior results with much lower complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 612, + 270, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 612, + 270, + 624 + ], + "spans": [ + { + "bbox": [ + 59, + 612, + 270, + 624 + ], + "type": "text", + "content": "In summary, we present the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 632, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 632, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 632, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 58, + 632, + 287, + 681 + ], + "type": "text", + "content": "- To our best knowledge, we are the first to solve the consistency problem of diffusion video editing by considering the concept of layered atlas approaches, which provides an efficient and effective way for this topic." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "content": "- We present a new video editing framework which can manipulate the appearance of the objects with high ge" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 108 + ], + "type": "text", + "content": "ometry and appearance consistency across time. Our method can be easily applied to other text-driven diffusion models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 116, + 545, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 116, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 116, + 545, + 152 + ], + "type": "text", + "content": "- We conduct extensive experiments on a variety of natural videos, which shows superior editing performance compared with state-of-the-art methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 169, + 392, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 169, + 392, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 169, + 392, + 182 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 189, + 458, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 189, + 458, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 189, + 458, + 202 + ], + "type": "text", + "content": "2.1. Diffusion for Image Editing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 208, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 545, + 472 + ], + "type": "text", + "content": "The editing of natural images is an important task in the field of computer vision that has been widely studied. Prior to the emergence of diffusion models [42, 15], many GAN-based approaches [11, 10, 30, 32, 48] have achieved good results. There are also some works focus on low-level editing [52, 6, 51]. The advent of diffusion models has made it possible to achieve even higher quality and more diverse edited contents. SDEdit [26] adds noise and corruptions to an input image and uses diffusion models to reverse the process for image editing, while suffering from the loss of fidelity. Prompt-to-Prompt [12] and Plug-and-Play [45] perform semantic editing by mixing activations from original and target text prompts. InstructPix2Pix [2] applies semantic editing at test time and personalizes the model through finetuning and optimization to learn a special token describing the content. UniTune [46] and Imagic [21] finetune on a single image for better editability while maintaining good fidelity. There are also some works exploring the controllability [55, 18, 39, 22, 4, 3] and personalization [9, 38, 29] of diffusion-based generation. Our proposed video editing method leverages existing image editing methods that can preserve the structures, such as [55, 31, 45] and [28]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 478, + 456, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 456, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 456, + 491 + ], + "type": "text", + "content": "2.2. Diffusion for Video Editing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 497, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 640 + ], + "type": "text", + "content": "Compared to image editing, video editing is more challenging for diffusion-based methods for geometric and temporal consistency. Tune-a-Video [49] inflates a text-to-image model for video editing. However, since temporal correlations are not fully considered, the editing results suffer from inconsistency of geometry and motion. Dreamix [27] develops a text-to-video backbone for motion editing while maintaining temporal consistency. There are also some works based on video generation like [8, 40, 13, 54, 56] and [17]. Unlike these approaches, our purpose is to enable diffusion models to perform appearance editing with both geometric and temporal consistency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 647, + 514, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 514, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 514, + 661 + ], + "type": "text", + "content": "2.3. Temporal Propagation in Video Editing" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 714 + ], + "type": "text", + "content": "Temporal propagation plays an important role in natural video process, since it is the essential factor for temporal consistency. Some methods rely on key frames [19, 44, 50] or optical flow [37] to propagate contents between frames." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23041" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 544, + 326 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 544, + 326 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 544, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 544, + 326 + ], + "type": "image", + "image_path": "233f6ea506e4a63c7c632db96f4dde7989e030a53d016bc9e053ab0f0be9244f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "text", + "content": "Figure 2: Framework of the proposed StableVideo. The input video is first fed into NLA [24] to generate foreground and background atlases using the pre-trained model. " + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_b" + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "text", + "content": " is the diffusion model used to edit background atlas, and " + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_f" + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "text", + "content": " is used to edit the foreground key frames. Note that " + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_b" + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_f" + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "text", + "content": " share the same weights, but accept different conditions. We employ depth information, extracted by MiDaS [35], for " + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_b" + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "text", + "content": " to maintain the consistency between the foreground motion and the environment, while structure guidance is used for " + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_f" + }, + { + "bbox": [ + 46, + 336, + 548, + 411 + ], + "type": "text", + "content": " to keep geometric consistency between the new generated foreground and the old one. After being edited, the foreground and background are blended together to reconstruct the edited frames." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 430, + 289, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 289, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 289, + 598 + ], + "type": "text", + "content": "Another bunch of methods are to achieve consistent interframe editing by forming a compressed representation of a video. Omnimatte [24, 25] estimate RGBA layers for target subject and scene effects for each frame independently, but cannot achieve consistent propagation of contents along temporal direction. Atlas [1, 20] tackles this problem by decomposing the video into unified 2D atlas layers for each target. This approach allows contents to be applied to the global summarized 2D atlases and mapped back to the video, achieving temporal consistency with minimal effort. Inspired by the concept of atlas approach, we employ the pre-trained neural layered atlas model to solve the inconsistency problem in diffusion video editing, thereby achieving high-quality editing results with temporal coherence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 609, + 104, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 104, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 104, + 621 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Fig. 2 shows the pipeline of our proposed StableVideo. We utilize NLA [20] as the propagator for consistent video editing. Specifically, we conduct foreground and background editing separately. For foreground editing, we adopt key frame editing to generate atlas layers with high quality and the inter-frame propagation module to ensure better geometric and temporal consistency. The edited key frames" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 430, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 479 + ], + "type": "text", + "content": "are then mapped to partial atlases and aggregated by the aggregation network to produce the edited foreground atlas. It is noteworthy that our approach can also handle more than one foreground layers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 489, + 430, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 489, + 430, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 489, + 430, + 501 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "text", + "content": "We employ the pre-trained NLA model [24] to propagate the edited contents to ensure that the target objects and scenes can maintain homogeneous appearances and motions across the entire video. The concept of NLA is to decompose the input video into layered representations, namely foreground atlas and background atlas, which globally summarize the correlated pixels for the foreground and the background, respectively. Three mapping networks, i.e., " + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^b (\\cdot),\\mathcal{M}^f (\\cdot)" + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\alpha}(\\cdot)" + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "text", + "content": ", are provided for this purpose. Given an input video " + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "text", + "content": ", for each frame " + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "text", + "content": ", we obtain the mapping relationships of the atlas in the background and the foreground with respect to the pixel coordinate system, named as " + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{UV}^b (\\cdot)" + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{UV}^f (\\cdot)" + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "text", + "content": ", as well as the foreground opacity " + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 304, + 509, + 547, + 677 + ], + "type": "text", + "content": " on the pixel coordinate system, formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 687, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 687, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 687, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\operatorname {U V} _ {i} ^ {b} (\\cdot) = \\mathcal {M} ^ {b} (I _ {i}), \\operatorname {U V} _ {i} ^ {f} (\\cdot) = \\mathcal {M} ^ {f} (I _ {i}), \\alpha_ {i} = \\mathcal {M} ^ {\\alpha} (I _ {i}). \\tag {1}", + "image_path": "666378780d39917af33a0895d96862170f81cb2fac8f0c037673c95c34b24915.jpg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23042" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 289, + 276 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 289, + 276 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 289, + 276 + ], + "type": "image", + "image_path": "d1d5f71c9c1edeafaef3c740071afd189040955fe537477a7189873addd98839.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 284, + 287, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 346 + ], + "type": "text", + "content": "Figure 3: Inter-frame propagation for foreground editing. We use two edited key frames, " + }, + { + "bbox": [ + 46, + 284, + 287, + 346 + ], + "type": "inline_equation", + "content": "E_{i-1}" + }, + { + "bbox": [ + 46, + 284, + 287, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 284, + 287, + 346 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 46, + 284, + 287, + 346 + ], + "type": "text", + "content": ", to illustrate the process more clearly. The structure guidance and the text prompt is added into the denoising UNet via the concatenation and cross-attention mechanism respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "type": "text", + "content": "After that, we formulate the mappings from the atlas representation of the background " + }, + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "type": "inline_equation", + "content": "A^b" + }, + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "type": "text", + "content": " and the foreground " + }, + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "type": "inline_equation", + "content": "A^f" + }, + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "type": "text", + "content": ", to the pixel coordinate systems of " + }, + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "type": "inline_equation", + "content": "B_i" + }, + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 47, + 365, + 287, + 401 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 408, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 408, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 99, + 408, + 287, + 423 + ], + "type": "interline_equation", + "content": "B _ {i} = \\mathrm {U V} _ {i} ^ {b} (A ^ {b}), F _ {i} = \\mathrm {U V} _ {i} ^ {f} (A ^ {f}). \\qquad (2)", + "image_path": "94ca79f5ed87e4987c3b8ddd86cf95faa3f402d95c9efeb319e32f16da0b5c38.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "text", + "content": "Our method achieves geometrical consistent editing by fixing the mappings of " + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\mathrm{UV}^b" + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\mathrm{UV}^f" + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "text", + "content": ", and generating the edited atlases of " + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "inline_equation", + "content": "A^b" + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "inline_equation", + "content": "A^f" + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "text", + "content": ". We adopt a pre-trained latent diffusion model [36] with guided conditions as our generator, namely " + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^b (\\cdot)" + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^f (\\cdot)" + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "text", + "content": ". Note that we are not simply editing the foreground and the background atlases directly. We apply inter-frame propagation mechanism on the editing process of foreground atlas. More details are explained in Sec. 3.2. After that, the entire video " + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 430, + 287, + 551 + ], + "type": "text", + "content": " can be reconstructed frame by frame as the following equation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 557, + 287, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 557, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 52, + 557, + 287, + 572 + ], + "type": "interline_equation", + "content": "I _ {i} = \\alpha_ {i} \\circ \\operatorname {U V} _ {i} ^ {f} \\left(\\mathcal {G} ^ {f} \\left(A ^ {f}\\right)\\right) + \\left(1 - \\alpha_ {i}\\right) \\circ \\operatorname {U V} _ {i} ^ {b} \\left(\\mathcal {G} ^ {b} \\left(A ^ {b}\\right)\\right), \\tag {3}", + "image_path": "9cceb23487618e11b700eaf3c78113d93389bf1a3964c6a6c801666d61af103a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 580, + 194, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 194, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 194, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 580, + 194, + 593 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 47, + 580, + 194, + 593 + ], + "type": "text", + "content": " denotes pixel-wise product." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 186, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 186, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 186, + 612 + ], + "type": "text", + "content": "3.2. Inter-frame Propagation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "content": "In this section, we further elaborate on how inter-frame propagation mechanism helps consistent foreground editing. One of the major challenges for diffusion models is to generate video contents with temporal consistency. Existing state-of-the-art text-driven diffusion methods [55, 28] can maintain the similar geometry between the target objects and the generated ones for image editing by adding structure conditions. However, the situation is different for" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 545, + 180 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 545, + 180 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 545, + 180 + ], + "type": "image", + "image_path": "73b0403a69007c24d9ccc44e07281e8acd878e2db2a0e475e65f552827e011c7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 188, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 304, + 188, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 547, + 308 + ], + "type": "text", + "content": "Figure 4: Training process of aggregation network. We employ a simple 3D network to aggregate the partial atlases generated from the edited key frames. For each training iteration, the sampled key frames are edited and mapped to partial atlases by " + }, + { + "bbox": [ + 304, + 188, + 547, + 308 + ], + "type": "inline_equation", + "content": "UV_{i}^{f}" + }, + { + "bbox": [ + 304, + 188, + 547, + 308 + ], + "type": "text", + "content": ". The partial atlases are then fed into the Aggregation Network to generate the edit atlas, which is mapped back by " + }, + { + "bbox": [ + 304, + 188, + 547, + 308 + ], + "type": "inline_equation", + "content": "(UV_{i}^{f})^{-1}" + }, + { + "bbox": [ + 304, + 188, + 547, + 308 + ], + "type": "text", + "content": " to generate the reconstructed frames. A " + }, + { + "bbox": [ + 304, + 188, + 547, + 308 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 188, + 547, + 308 + ], + "type": "text", + "content": " loss is used to guarantee the aggregation consistency between the edited key frames and the reconstructed frames." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 554 + ], + "type": "text", + "content": "videos. Generating temporally consistent geometry needs to handle some uncertain changes across time, e.g., motion and deformation, which can not be supported by them. We tackle this problem by introducing a conditional denoising process to enable the diffusion models to consider both the structure of the current frame and the appearance information from previous frame, thereby sequentially generating new objects with geometric consistency across time. In specific, we employ canny edge as structure guidance, which is also adopted by existing diffusion methods [55]. Another important question is how to propagate the information of one object across frames to achieve consistent appearance. With the help of NLA, we can transfer the appearance features of the overlapping parts of previous frame to the next frame. Inspired by SDEdit [26] and ILVR [7], we further use a process of adding noise and denoising to obtain a more complete output. We illustrate this generation process in Fig. 3. The inter-frame propagation method is only applied on the foreground objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": "Specifically, we first select " + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": " foreground key frames from the original video " + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": ", ensuring that 1) there is significant overlap between the adjacent frames, and 2) these key frames capture the appearance of all faces of the object. Given a generator " + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^f(\\cdot)" + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": " and a text prompt " + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": ", we edit the first frame " + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": " in pixel coordinate system with its structure condition " + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "inline_equation", + "content": "C_0" + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": " as the extra guidance:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 389, + 645, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 645, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 389, + 645, + 545, + 658 + ], + "type": "interline_equation", + "content": "E _ {0} = \\mathcal {G} ^ {f} (T, C _ {0}), \\tag {4}", + "image_path": "462c2f9a40b4a3f79ad1e62141d4ae53c3ea29b9606b7e36bb56d1a0cedf4f16.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_0" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " represents the editing result. Then for the remaining key frames, we propagate the editing result from the previous key frame " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{i-1}" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " to obtain the one of the current key frame " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": ". To be specific, we map " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "E_{i-1}" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " from the pixel" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23043" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 266, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 266, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 266, + 83 + ], + "type": "text", + "content": "Foreground: a polar bear; Background: north pole." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 84, + 542, + 185 + ], + "blocks": [ + { + "bbox": [ + 53, + 84, + 542, + 185 + ], + "lines": [ + { + "bbox": [ + 53, + 84, + 542, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 84, + 542, + 185 + ], + "type": "image", + "image_path": "5374b288a1d996eb4934bc8d6613e530dd9d45f6e25d58094167a26c16ef2695.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 290, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 290, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 290, + 207 + ], + "type": "text", + "content": "Foreground: a car with graffiti; Background: Miami city." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 208, + 542, + 256 + ], + "blocks": [ + { + "bbox": [ + 53, + 208, + 542, + 256 + ], + "lines": [ + { + "bbox": [ + 53, + 208, + 542, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 542, + 256 + ], + "type": "image", + "image_path": "8d280b1ac9351d46e152ca0e06d145bf8790b7d18180b2e6ccb7377982a96b51.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 53, + 259, + 542, + 308 + ], + "blocks": [ + { + "bbox": [ + 53, + 259, + 542, + 308 + ], + "lines": [ + { + "bbox": [ + 53, + 259, + 542, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 259, + 542, + 308 + ], + "type": "image", + "image_path": "18c530d0c83dfaf44954e1f81390142a04906fe3aae616f2d7930fa9c8931c2f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 317, + 547, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 317, + 547, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 547, + 343 + ], + "type": "text", + "content": "Figure 5: Compositing editing. We demonstrate two editing examples with non-rigid and rigid foreground objects. Our approach can well preserve the geometry of \"bear\" and \"car\" across frames." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "type": "text", + "content": "coordinate to atlas " + }, + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "type": "inline_equation", + "content": "A_{i - 1}^{E}" + }, + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "type": "text", + "content": " and then map it back to the pixel coordinate in the current frame " + }, + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "type": "inline_equation", + "content": "\\hat{E}_i" + }, + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "type": "text", + "content": " with multiplying the opacity " + }, + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 46, + 362, + 288, + 401 + ], + "type": "text", + "content": ", formulated by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 410, + 287, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 410, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 108, + 410, + 287, + 432 + ], + "type": "interline_equation", + "content": "A _ {i - 1} ^ {f} = \\left(\\mathrm {U V} _ {i - 1} ^ {f}\\right) ^ {- 1} \\left(E _ {i - 1}\\right), \\tag {5}", + "image_path": "49dd50a7acafe7da3eb02e1e6f67910ddba749cc729164f564456dfdd1e0e145.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 427, + 217, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 427, + 217, + 443 + ], + "spans": [ + { + "bbox": [ + 121, + 427, + 217, + 443 + ], + "type": "interline_equation", + "content": "\\hat {E} _ {i} = \\alpha_ {i} \\circ \\mathrm {U V} _ {i} ^ {f} (A _ {i - 1} ^ {f}).", + "image_path": "e1a0e46604c7203359a2a22e7009e433b9f75fa2dccdad8eb8da1e039d753ee5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "type": "text", + "content": "It is noteworthy that the entire video shares the same foreground atlas " + }, + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "type": "inline_equation", + "content": "A^f" + }, + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "type": "text", + "content": " for each target object, where the subscript " + }, + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "type": "text", + "content": " represents an incomplete partial atlas in frame " + }, + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "type": "text", + "content": ". We then use the atlas to propagate the pixel values from previous frame to their corresponding positions in the current frame to obtain an incomplete partial appearance " + }, + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\hat{E}_i" + }, + { + "bbox": [ + 46, + 452, + 287, + 523 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 524, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 524, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 524, + 287, + 572 + ], + "type": "text", + "content": "Given the partial appearance " + }, + { + "bbox": [ + 46, + 524, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\hat{E}_i" + }, + { + "bbox": [ + 46, + 524, + 287, + 572 + ], + "type": "text", + "content": ", we first encode it with VQ-VAE [47] to get the latent representation " + }, + { + "bbox": [ + 46, + 524, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\hat{Z}_i" + }, + { + "bbox": [ + 46, + 524, + 287, + 572 + ], + "type": "text", + "content": ", and then add noise to it with Variance Preserving Stochastic Differential Equation (VP-SDE), formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 71, + 581, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 581, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 71, + 581, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\hat {Z} _ {i} (t _ {0}) = \\alpha (t _ {0}) \\hat {Z} _ {t} (0) + \\sigma (t _ {0}) \\mathbf {z}, \\quad \\mathbf {z} \\sim \\mathcal {N} (0, \\mathbf {I}) \\quad (6)", + "image_path": "61cb9dbed839ded789c853c5ada4442614cdfca846d3288ae8e725fce946f29e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\sigma(t_0)" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\alpha(t_0)" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " are two scalar functions that satisfy " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\alpha^2(t) + \\sigma^2(t) = 1, \\forall t \\in (0,1]" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "t_0 \\in [0,1]" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " is a hyper parameter of the noise strength. Then we apply denoising process " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{Z}_i(t_0)" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " under the condition guidance from both text prompt " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " and structure guidance " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " to get the latent representation " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "Z_i" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ". Finally, we decode the latent representation to " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "Z_i" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " propagate editing result " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ". Our experiments further demonstrate that this mechanism can achieve good propagation results without training or fine-tuning the model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 361, + 430, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 430, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 430, + 375 + ], + "type": "text", + "content": "3.3. Aggregation Network" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 380, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 583 + ], + "type": "text", + "content": "Different from [1] and [24], our approach edits video frames rather than atlases, which have the chance to achieve more information of different viewpoints. This brings two advantages. Firstly, the geometries and pixels from different viewpoints provide more details of the target objects, allowing the diffusion model to generate the edited content with higher fidelity. Secondly, this alleviates the risk of failure editing due to the potential wrong mapping from the atlas to the video frames. We then aggregate the edited key frames by using a simple yet effective two-layer 2D convolution network with skip connection as shown in Fig. 4. Our goal is to guarantee that the aggregated atlas is highly aligned with the original one, in terms of locations, so that appearance edit will not affect the geometric consistency and the temporal continuity. Reconstruction loss, " + }, + { + "bbox": [ + 304, + 380, + 545, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 304, + 380, + 545, + 583 + ], + "type": "text", + "content": ", between the edited and reconstructed key frames is employed in the training process as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 590, + 545, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 590, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 359, + 590, + 545, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\sum_ {i = 1} ^ {N} | | E _ {i} - \\mathrm {U V} _ {i} ^ {f} \\left(A ^ {f}\\right) | | _ {1}. \\tag {7}", + "image_path": "af4442a28b91f7f165fd55877291cbfd356967c37dd99ea17a180e2a24c1d8da.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 629, + 461, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 629, + 461, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 629, + 461, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 629, + 461, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 629, + 461, + 641 + ], + "type": "text", + "content": " is the number of key frames." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 651, + 388, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 388, + 665 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 388, + 665 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 671, + 433, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 433, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 433, + 685 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "In practice, we implement our approach over Stable Diffusion [36]. Despite there are several image-based meth" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23044" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 205, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 205, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 205, + 84 + ], + "type": "text", + "content": "Background: desert scene, pyramid." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 86, + 545, + 133 + ], + "blocks": [ + { + "bbox": [ + 50, + 86, + 545, + 133 + ], + "lines": [ + { + "bbox": [ + 50, + 86, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 50, + 86, + 545, + 133 + ], + "type": "image", + "image_path": "663f260a0da3dadbde6d448edb13f668868800720d77f84c661d8fa4a70eb0cb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 137, + 543, + 186 + ], + "blocks": [ + { + "bbox": [ + 50, + 137, + 543, + 186 + ], + "lines": [ + { + "bbox": [ + 50, + 137, + 543, + 186 + ], + "spans": [ + { + "bbox": [ + 50, + 137, + 543, + 186 + ], + "type": "image", + "image_path": "07465c1bd9aab0c9f7176266cb42fbb26f2c4ab76d3855cebb98797db63e9e85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 194, + 219, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 194, + 219, + 206 + ], + "spans": [ + { + "bbox": [ + 54, + 194, + 219, + 206 + ], + "type": "text", + "content": "Background: magma scene, crack road." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 209, + 543, + 258 + ], + "blocks": [ + { + "bbox": [ + 50, + 209, + 543, + 258 + ], + "lines": [ + { + "bbox": [ + 50, + 209, + 543, + 258 + ], + "spans": [ + { + "bbox": [ + 50, + 209, + 543, + 258 + ], + "type": "image", + "image_path": "af16a53aa50d2e156307ee56123b0389f296ec98af50aa54aec663fbb21d2c7b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 262, + 543, + 311 + ], + "blocks": [ + { + "bbox": [ + 50, + 262, + 543, + 311 + ], + "lines": [ + { + "bbox": [ + 50, + 262, + 543, + 311 + ], + "spans": [ + { + "bbox": [ + 50, + 262, + 543, + 311 + ], + "type": "image", + "image_path": "fff5268a65ad885b9b793e22cb71867ab1f553add3d2c532e89176d50e4e329c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 321, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 546, + 346 + ], + "type": "text", + "content": "Figure 6: Background replacement. Since our approach can effectively maintain the geometry of the foreground, it can perform background replacement while maintaining geometric consistency of depth and temporal continuity of perspective." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "text", + "content": "ods [31, 45, 28] can perform structure-preserving editing, we choose the canny condition branch from [55] as the structure guidance for the proposed inter-frame propagation in our method. We apply our method on several videos from DAVIS [33], with each video containing a moving object in " + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "inline_equation", + "content": "50 \\sim 70" + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "text", + "content": " frames. The image resolution is set to " + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "inline_equation", + "content": "768 \\times 432" + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "text", + "content": ", and the resolution of foreground atlas is set to " + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "inline_equation", + "content": "2000 \\times 2000" + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "text", + "content": ". We employ DDIM [43] sampler with 20 steps. In this case, our method requires only " + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\sim 10" + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "text", + "content": " GB GPU memory and takes " + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\sim 30" + }, + { + "bbox": [ + 46, + 353, + 289, + 473 + ], + "type": "text", + "content": " seconds for each video in a single NVIDIA A40 GPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 479, + 142, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 142, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 142, + 491 + ], + "type": "text", + "content": "4.2. Editing Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 498, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 520 + ], + "type": "text", + "content": "We tested our method for various editing types among several videos. Here we demonstrate several scenarios:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "text", + "content": "Compositing editing. Our method can edit foreground and background separately to achieve high-quality and semantically matched editing results as shown in Fig. 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 557, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 628 + ], + "type": "text", + "content": "Background replacement. Due to the separate editing capabilities enabled by NLA, as well as the diverse and high-quality editing brought by the diffusion model, our method achieves high-quality video background replacement while maintaining geometric consistency of depth and temporal continuity of perspective, as shown in Fig. 6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 629, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 665 + ], + "type": "text", + "content": "Style transfer. Our method accomplishes a wide range of style transfers, while simultaneously ensuring temporal consistency, as shown in Fig. 7." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 671, + 191, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 191, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 191, + 684 + ], + "type": "text", + "content": "4.3. Comparison to Prior Arts" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In this section, we compare our editing results with state-of-the-art atlas-based method Text2LIVE [1] and diffusion-" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 353, + 446, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 353, + 446, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 446, + 365 + ], + "type": "text", + "content": "based method Tune-A-Video [49]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 365, + 546, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 473 + ], + "type": "text", + "content": "Comparison to Text2LIVE [1]. As shown in Fig. 8, given the text prompt \"an orange SUV\", Text2LIVE shows incomplete editing, while our method demonstrates a much more holistic result. It is because our method employs key frame editing, while Text2LIVE creates a new layer to edit the atlas somehow directly. The potential failure editing in atlas will lead to the error propagation to the entire video. In addition, compared to Text2LIVE, we can achieve higher quality and richer content with faster inference speed." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "content": "Comparison to Tune-A-Video [49]. As shown in Fig. 9, vanilla video diffusion models like Tune-A-Video often fail in video editing. While it adeptly captures the semantic information of text prompts, it struggles to preserve consistency in video layout and object geometry." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Consistency analysis. To the best of our knowledge, there is currently no widely accepted metric for evaluating the geometric and temporal consistency of videos. In this paper, we employ motion consistency of dense optical flow and deviation consistency of the edited video frames for this metric. For motion consistency, we employ the Farneback algorithm in OpenCV to calculate the average L2 distance of dense optical flow between the edited and original videos. We select the \"car-turn\" video from DAVIS and apply the same prompt to all methods. The experiment is repeated several times to obtain the average number. Our method shows better stability than Tune-A-Video as shown in Tab. 1. For deviation consistency, we conduct experiments on 1) CLIP score: target text faithfulness 2) LPIPS-P: deviation from the original video frames and 3) LPIPS-T:" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23045" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 302, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 302, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 302, + 84 + ], + "type": "text", + "content": "Foreground: a panda; Both: in the style of Chinese painting." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 86, + 542, + 133 + ], + "blocks": [ + { + "bbox": [ + 52, + 86, + 542, + 133 + ], + "lines": [ + { + "bbox": [ + 52, + 86, + 542, + 133 + ], + "spans": [ + { + "bbox": [ + 52, + 86, + 542, + 133 + ], + "type": "image", + "image_path": "2483d949fe37b74d3efe3b5e38a5fb432eb0613d37e570b2f406012a13d75489.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 137, + 542, + 186 + ], + "blocks": [ + { + "bbox": [ + 52, + 137, + 542, + 186 + ], + "lines": [ + { + "bbox": [ + 52, + 137, + 542, + 186 + ], + "spans": [ + { + "bbox": [ + 52, + 137, + 542, + 186 + ], + "type": "image", + "image_path": "b27afc08e297416387f9b62914f53cfb2b581d2ad4f314cd2c4d873bbe01c49c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 196, + 310, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 310, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 310, + 209 + ], + "type": "text", + "content": "Both: in the style of Vincent Willem van Gogh's Starry Night." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 209, + 542, + 258 + ], + "blocks": [ + { + "bbox": [ + 52, + 209, + 542, + 258 + ], + "lines": [ + { + "bbox": [ + 52, + 209, + 542, + 258 + ], + "spans": [ + { + "bbox": [ + 52, + 209, + 542, + 258 + ], + "type": "image", + "image_path": "1484075a1af04971cc6cb3b66f09a8529a63a4b8a693e7fcbe6b82825a3976a6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 260, + 542, + 309 + ], + "blocks": [ + { + "bbox": [ + 52, + 260, + 542, + 309 + ], + "lines": [ + { + "bbox": [ + 52, + 260, + 542, + 309 + ], + "spans": [ + { + "bbox": [ + 52, + 260, + 542, + 309 + ], + "type": "image", + "image_path": "9e3f99ea0d6a2d87f317804c6c8fc9219ed830982162baa413adf955312ef81e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 319, + 534, + 332 + ], + "lines": [ + { + "bbox": [ + 57, + 319, + 534, + 332 + ], + "spans": [ + { + "bbox": [ + 57, + 319, + 534, + 332 + ], + "type": "text", + "content": "Figure 7: Style transfer. Our approach achieves diverse video style transfer while ensuring high temporal consistency." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 51, + 353, + 286, + 430 + ], + "blocks": [ + { + "bbox": [ + 51, + 353, + 286, + 430 + ], + "lines": [ + { + "bbox": [ + 51, + 353, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 51, + 353, + 286, + 430 + ], + "type": "image", + "image_path": "fc3358983057c125467417a34372a52f5eeb97de8c0dc1adfda43f9afc96c097.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 430, + 190, + 440 + ], + "lines": [ + { + "bbox": [ + 145, + 430, + 190, + 440 + ], + "spans": [ + { + "bbox": [ + 145, + 430, + 190, + 440 + ], + "type": "text", + "content": "Text2LIVE" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 51, + 446, + 286, + 524 + ], + "blocks": [ + { + "bbox": [ + 51, + 446, + 286, + 524 + ], + "lines": [ + { + "bbox": [ + 51, + 446, + 286, + 524 + ], + "spans": [ + { + "bbox": [ + 51, + 446, + 286, + 524 + ], + "type": "image", + "image_path": "20d3965b105977ffad705ff40b4d8152352a8dd9e425e9ec43b1b59018c0ae39.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 524, + 204, + 534 + ], + "lines": [ + { + "bbox": [ + 130, + 524, + 204, + 534 + ], + "spans": [ + { + "bbox": [ + 130, + 524, + 204, + 534 + ], + "type": "text", + "content": "StableVideo (ours)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 544, + 287, + 581 + ], + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 581 + ], + "type": "text", + "content": "Figure 8: Comparison to Text2LIVE. Foreground prompt: an orange SUV. Our method achieves more holistic editing results on the foreground." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 503, + 545, + 540 + ], + "lines": [ + { + "bbox": [ + 305, + 503, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 545, + 540 + ], + "type": "text", + "content": "Figure 9: Comparison to Tune-A-Video. Prompt: an orange SUV in sunny snowy winter, cabins. Our method achieves much more consistent editing results." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 591, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 288, + 628 + ], + "type": "text", + "content": "deviation between adjacent frames, as shown in Tab. 2. Our method achieves comparable CLIP score and much lower deviations, which shows the effectiveness and stability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 635, + 141, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 141, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 141, + 647 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "To verify the necessity of the key frame editing, we apply editing in atlas layer directly for the foreground as a simple baseline. The atlas might not be so deformed for human perception, but it significantly affects the diffusion models. Fig. 10 shows an example, where obvious deformation and" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 351, + 544, + 386 + ], + "blocks": [ + { + "bbox": [ + 307, + 351, + 544, + 386 + ], + "lines": [ + { + "bbox": [ + 307, + 351, + 544, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 351, + 544, + 386 + ], + "type": "image", + "image_path": "c27349148b31ba5c0a543bf3cf3170d60fc119b40ceedd0b8c2bd7a819cd3d29.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 399, + 544, + 434 + ], + "blocks": [ + { + "bbox": [ + 399, + 388, + 454, + 398 + ], + "lines": [ + { + "bbox": [ + 399, + 388, + 454, + 398 + ], + "spans": [ + { + "bbox": [ + 399, + 388, + 454, + 398 + ], + "type": "text", + "content": "original video" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 399, + 544, + 434 + ], + "lines": [ + { + "bbox": [ + 307, + 399, + 544, + 434 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 544, + 434 + ], + "type": "image", + "image_path": "fd446f702b699b438815e52e29f595deb71ad72f2d4b8fc6c33dcce9da2c93e5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 436, + 455, + 445 + ], + "lines": [ + { + "bbox": [ + 399, + 436, + 455, + 445 + ], + "spans": [ + { + "bbox": [ + 399, + 436, + 455, + 445 + ], + "type": "text", + "content": "Tune-A-Video" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 449, + 544, + 483 + ], + "blocks": [ + { + "bbox": [ + 307, + 449, + 544, + 483 + ], + "lines": [ + { + "bbox": [ + 307, + 449, + 544, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 544, + 483 + ], + "type": "image", + "image_path": "f34f2c5c00fd2dcc8563c1047e9810c3d1d0e3f61c15dde62227e918e188d704.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 484, + 465, + 495 + ], + "lines": [ + { + "bbox": [ + 389, + 484, + 465, + 495 + ], + "spans": [ + { + "bbox": [ + 389, + 484, + 465, + 495 + ], + "type": "text", + "content": "StableVideo (ours)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 552, + 350, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 552, + 350, + 561 + ], + "spans": [ + { + "bbox": [ + 307, + 552, + 350, + 561 + ], + "type": "text", + "content": "Editing atlas" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 307, + 562, + 545, + 586 + ], + "blocks": [ + { + "bbox": [ + 307, + 562, + 545, + 586 + ], + "lines": [ + { + "bbox": [ + 307, + 562, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 562, + 545, + 586 + ], + "type": "image", + "image_path": "521d3b6d18cd5b0421b906fd8e238abf0b9610e5a3dcd36d40667d82b5143cca.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 307, + 597, + 545, + 621 + ], + "blocks": [ + { + "bbox": [ + 308, + 587, + 370, + 597 + ], + "lines": [ + { + "bbox": [ + 308, + 587, + 370, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 370, + 597 + ], + "type": "text", + "content": "Editing key frames" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 597, + 545, + 621 + ], + "lines": [ + { + "bbox": [ + 307, + 597, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 597, + 545, + 621 + ], + "type": "image", + "image_path": "b9c7691af5ef240ba069b8bca46d2fdefdeda2a138b7bc586d68963728772eb4.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 631, + 545, + 654 + ], + "lines": [ + { + "bbox": [ + 305, + 631, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 305, + 631, + 545, + 654 + ], + "type": "text", + "content": "Figure 10: Ablation study on directly editing the atlas. The deformation in atlas affects the diffusion models." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 677, + 386, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 386, + 689 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 386, + 689 + ], + "type": "text", + "content": "inconsistency exist." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "We also conduct extensive ablation study on inter-frame propagation module. The objective of this module is to" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23046" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 285, + 105 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 285, + 105 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 285, + 105 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 285, + 105 + ], + "type": "image", + "image_path": "4c0a28fcb55e54d53a4d4f1119b1c641b2279c0a787ce20d26d64c03ca8a7d01.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 106, + 203, + 116 + ], + "lines": [ + { + "bbox": [ + 133, + 106, + 203, + 116 + ], + "spans": [ + { + "bbox": [ + 133, + 106, + 203, + 116 + ], + "type": "text", + "content": "random generation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 122, + 285, + 156 + ], + "blocks": [ + { + "bbox": [ + 50, + 122, + 285, + 156 + ], + "lines": [ + { + "bbox": [ + 50, + 122, + 285, + 156 + ], + "spans": [ + { + "bbox": [ + 50, + 122, + 285, + 156 + ], + "type": "image", + "image_path": "1e3febb87f7f45873f359d1a48971cde4ea62b3c14fa5a9321d0830fdf31f60c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 157, + 197, + 166 + ], + "lines": [ + { + "bbox": [ + 140, + 157, + 197, + 166 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 197, + 166 + ], + "type": "text", + "content": "fix initial latent" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 173, + 285, + 206 + ], + "blocks": [ + { + "bbox": [ + 50, + 173, + 285, + 206 + ], + "lines": [ + { + "bbox": [ + 50, + 173, + 285, + 206 + ], + "spans": [ + { + "bbox": [ + 50, + 173, + 285, + 206 + ], + "type": "image", + "image_path": "0e14a33a891338e9e0fdd62d378fe1ec2aaa6fc7e9070efb55adf7f80f924251.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 209, + 208, + 218 + ], + "lines": [ + { + "bbox": [ + 129, + 209, + 208, + 218 + ], + "spans": [ + { + "bbox": [ + 129, + 209, + 208, + 218 + ], + "type": "text", + "content": "sequential generation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 224, + 285, + 258 + ], + "blocks": [ + { + "bbox": [ + 50, + 224, + 285, + 258 + ], + "lines": [ + { + "bbox": [ + 50, + 224, + 285, + 258 + ], + "spans": [ + { + "bbox": [ + 50, + 224, + 285, + 258 + ], + "type": "image", + "image_path": "beab8b47099ef54fbe07f10d3baf812d34f57500f54739c585205db3e0d82d73.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 258, + 223, + 269 + ], + "lines": [ + { + "bbox": [ + 110, + 258, + 223, + 269 + ], + "spans": [ + { + "bbox": [ + 110, + 258, + 223, + 269 + ], + "type": "text", + "content": "Inter-frame Propagation (ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 329, + 286, + 363 + ], + "blocks": [ + { + "bbox": [ + 46, + 278, + 287, + 315 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 315 + ], + "type": "text", + "content": "Figure 11: Ablation study on inter-frame propagation module. Foreground prompt: an orange SUV. Our method achieves excellent consistency in key frame appearance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 329, + 286, + 363 + ], + "lines": [ + { + "bbox": [ + 50, + 329, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 50, + 329, + 286, + 363 + ], + "type": "image", + "image_path": "8453c216cf43082f1bc4176039f6fe32e174706e58aa900f729298f460d61381.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "type": "text", + "content": "Figure 12: Binary search on hyper parameter " + }, + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "type": "text", + "content": " in interframe propagation module. The appearance of previous frame is shown on the right side. The sequence on the left shows the generated results of the current frame as a function of " + }, + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "type": "text", + "content": ". The value of " + }, + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 372, + 287, + 433 + ], + "type": "text", + "content": " around 0.8 is a good sweet spot." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 371, + 545, + 431 + ], + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 431 + ], + "type": "text", + "content": "Fig. 12. We observe that as " + }, + { + "bbox": [ + 305, + 371, + 545, + 431 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 305, + 371, + 545, + 431 + ], + "type": "text", + "content": " gradually increases, the generated results become more realistic but gradually lose their match with the appearance of the previous frame. Using binary search, we find out that a reasonable trade-off between fidelity and realism for " + }, + { + "bbox": [ + 305, + 371, + 545, + 431 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 305, + 371, + 545, + 431 + ], + "type": "text", + "content": " lies around 0.8." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 481 + ], + "type": "text", + "content": "maintain the geometry of the foreground when editing key frames. Firstly, we consider four different settings for editing key frames as shown in Fig. 11." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 483, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 518 + ], + "type": "text", + "content": "Random generation. Each key frame only shares the same text prompt with the others. In this case, there are significant differences among the generated key frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 579 + ], + "type": "text", + "content": "Fix initial latent. Unlike starting from random noise every time we edit, we start generating each key frame from the same latent noise and share the text prompt. In this case, there is higher similarity in the content generated for each frame, but the consistency is still not satisfactory." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 580, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 639 + ], + "type": "text", + "content": "Sequential generation. Furthermore, we concatenate the latent noise between frames. Specifically, we apply imaged-to-image translation between frames. This method still cannot guarantee consistency since the appearances of the objects between the two frames do not match." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 640, + 287, + 689 + ], + "type": "text", + "content": "Inter-frame propagation (ours). Our final approach is to employ partial atlas to geometrically align the appearances between two frames, followed by a process of adding noise and then apply denoising process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In addition, We also conducted experiments on the selection of the hyper parameter " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " in this module as shown in" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 321, + 71, + 531, + 122 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 531, + 122 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 531, + 122 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 531, + 122 + ], + "type": "table", + "html": "
MethodForeground EditingComposite Editing
Text2LIVE [1]1.997.39
Tune-A-Video [49]4.6312.74
StableVideo (ours)3.3411.48
", + "image_path": "7c4693f0e970415c09253db6c9bfec5576a7b6646f7292e94de91d9a6d527be4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 321, + 178, + 531, + 222 + ], + "blocks": [ + { + "bbox": [ + 305, + 130, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 305, + 130, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 130, + 545, + 167 + ], + "type": "text", + "content": "Table 1: Consistency of dense optical flow. Original video: car-turn. Foreground prompt (column 1): an orange suv. Composite prompt (column 2): a car driving in winter." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 178, + 531, + 222 + ], + "lines": [ + { + "bbox": [ + 321, + 178, + 531, + 222 + ], + "spans": [ + { + "bbox": [ + 321, + 178, + 531, + 222 + ], + "type": "table", + "html": "
MethodCLIP (↑)LPIPS-P (↓)LPIPS-T (↓)
Tune-A-Video [49]0.27870.63460.1851
StableVideo (ours)0.27130.16130.0386
", + "image_path": "3fabaf00bd616a351c6e297b1414dcd7b3e27f68cee4de4b0ed206fa2f81550d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 231, + 544, + 255 + ], + "lines": [ + { + "bbox": [ + 305, + 231, + 544, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 544, + 255 + ], + "type": "text", + "content": "Table 2: Quantitative results. Original video: car-turn. Prompt: an orange SUV in the winter." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 266, + 545, + 314 + ], + "blocks": [ + { + "bbox": [ + 307, + 266, + 545, + 314 + ], + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 314 + ], + "type": "image", + "image_path": "95117529b88d127ce7f948386fcec03fe288ebe71d53380335f9a7a79db417dc.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 323, + 545, + 360 + ], + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 360 + ], + "type": "text", + "content": "Figure 13: Failure case. Videos with non-rigid deformation may lead to failure editing, since the movement of the object is more difficult to be well captured in this case." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 441, + 477, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 477, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 477, + 453 + ], + "type": "text", + "content": "5. Limitations and Future Works" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 461, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 593 + ], + "type": "text", + "content": "Firstly, our method is constrained by NLA. Learning atlas layers may fail for non-rigid objects with significant structural deformation as shown in Fig. 13. While we can mitigate this by dividing long videos into short clips where the objects can be considered to be rigid, it is still not feasible to address every single case. Secondly, our method is constrained by the capabilities of the diffusion models, which may struggle with specific scenarios such as human or animals. Besides, it may be better to optimize the diffusion model with the objective of aligning the generated contents to the reconstructed ones." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 604, + 378, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 378, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 378, + 616 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "type": "text", + "content": "We have proposed a text-driven diffusion video editing approach. To solve the consistency problem for diffusion models in foreground object editing, we propose a interframe propagation mechanism and an atlas aggregation network. We conducted extensive experiments and demonstrated the superior qualitative and quantitative results of our method compared to state-of-the-art approaches." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23047" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 248 + ], + "type": "text", + "content": "[3] Shidong Cao, Wenhao Chai, Shengyu Hao, and Gaoang Wang. Image reference-guided fashion design with structure-aware transfer by diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3524-3528, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 294 + ], + "type": "text", + "content": "[4] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. arXiv preprint arXiv:2302.06826, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 328 + ], + "type": "text", + "content": "[5] Wenhao Chai and Gaoang Wang. Deep vision multimodal learning: Methodology, benchmark, and trend. Applied Sciences, 12(13):6588, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 373 + ], + "type": "text", + "content": "[6] Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen, Jun Shi, and Jingchun Zhou. Snowformer: Scale-aware transformer via context interaction for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "text", + "content": "[7] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 288, + 464 + ], + "type": "text", + "content": "[8] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 288, + 520 + ], + "type": "text", + "content": "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "text", + "content": "[10] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 612 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[14] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:210.02303, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 196 + ], + "type": "text", + "content": "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "text", + "content": "[17] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "text", + "content": "[18] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "text", + "content": "[19] Ondřej Jamriška, Šárka Sochorová, Ondřej Texler, Michal Lukáč, Jakub Fiser, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Stylizing video by example. ACM Transactions on Graphics (TOG), 38(4):1-11, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "text", + "content": "[20] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 366, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 410 + ], + "type": "text", + "content": "[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "type": "text", + "content": "[22] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. arXiv preprint arXiv:2301.07093, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 456, + 545, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 512 + ], + "type": "text", + "content": "[23] Weng Fei Low and Gim Hee Lee. Minimal neural atlas: Parameterizing complex surfaces with minimal charts and distortion. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 465-481. Springer, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 513, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 556 + ], + "type": "text", + "content": "[24] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. arXiv preprint arXiv:2009.07833, 2020. 1, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[25] Erika Lu, Forrester Cole, Weidi Xie, Tali Dekel, William T Freeman, Andrew Zisserman, and Michael Rubinstein. Associating objects and their effects in video through coordination games. In Advances in Neural Information Processing Systems. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "text", + "content": "[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23048" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 163 + ], + "type": "text", + "content": "[29] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "text", + "content": "[30] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 253 + ], + "type": "text", + "content": "[31] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. arXiv preprint arXiv:2302.03027, 2023. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "text", + "content": "[32] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 354 + ], + "type": "text", + "content": "[33] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 420 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 477 + ], + "type": "text", + "content": "[35] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 533 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 589 + ], + "type": "text", + "content": "[37] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In *Pattern Recognition: 38th German Conference*, GCPR 2016, Hannover, Germany, September 12-15, 2016, Proceedings 38, pages 26-36. Springer, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "text", + "content": "[38] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[39] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[40] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "type": "text", + "content": "[41] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 162, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 195 + ], + "type": "text", + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "text", + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "type": "text", + "content": "[44] Ondrej Texler, David Futschik, Michal Kucera, Ondrej Jamriška, Šárka Sochorová, Menclei Chai, Sergey Tulyakov, and Daniel Sykora. Interactive video stylization using few-shot patch-based training. ACM Transactions on Graphics (TOG), 39(4):73-1, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 545, + 327 + ], + "type": "text", + "content": "[45] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "text", + "content": "[46] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 371, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 404 + ], + "type": "text", + "content": "[47] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "text", + "content": "[48] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "type": "text", + "content": "[49] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 514, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 570 + ], + "type": "text", + "content": "[50] Yiran Xu, Badour AlBahar, and Jia-Bin Huang. Temporally consistent semantic video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 357-374. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 571, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 635 + ], + "type": "text", + "content": "[51] Ren Yang, Radu Timofte, Xin Li, Qi Zhang, Lin Zhang, Fanglong Liu, Dongliang He, Fu Li, He Zheng, Weihang Yuan, et al. Aim 2022 challenge on super-resolution of compressed image and video: Dataset, methods and results. In European Conference on Computer Vision, pages 174-202. Springer, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 680 + ], + "type": "text", + "content": "[52] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 681, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 714 + ], + "type": "text", + "content": "[53] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. Magvit:" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23049" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 207 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Masked generative video transformer. arXiv preprint arXiv:2212.05199, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[54] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. arXiv preprint arXiv:2302.07685, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 162 + ], + "type": "text", + "content": "[55] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "type": "text", + "content": "[56] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23050" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/0e05e637-ac63-484f-9954-22b778c7c120_content_list.json b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/0e05e637-ac63-484f-9954-22b778c7c120_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6ea5bd4458be23d110e2c55523716bb4b69fe7dc --- /dev/null +++ b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/0e05e637-ac63-484f-9954-22b778c7c120_content_list.json @@ -0,0 +1,1570 @@ +[ + { + "type": "text", + "text": "StageInteractor: Query-based Object Detector with Cross-stage Interaction", + "text_level": 1, + "bbox": [ + 102, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yao Teng $^{1}$ Haisong Liu $^{1}$ Sheng Guo $^{3}$ Limin Wang $^{1,2,\\boxtimes}$", + "bbox": [ + 236, + 178, + 732, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1State Key Laboratory for Novel Software Technology, Nanjing University, China", + "bbox": [ + 163, + 198, + 805, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Shanghai AI Lab, China $^{3}$ MYbank, Ant Group, China", + "bbox": [ + 256, + 215, + 712, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous object detectors make predictions based on dense grid points or numerous preset anchors. Most of these detectors are trained with one-to-many label assignment strategies. On the contrary, recent query-based object detectors are based a sparse set of learnable queries refined by a series of decoder layers. The one-to-one label assignment is independently applied on each layer for deep supervision during training. Despite the great success of query-based object detection, however, this vanilla one-to-one label assignment strategy requires the detectors to have strong fine-grained discrimination and modeling capacity. In this paper, we propose a new query-based object detector with cross-stage interaction, coined as StageInteractor. During the forward pass, we come up with an efficient way to improve this modeling ability by reusing dynamic operators with lightweight adapters. As for the label assignment, a cross-stage label assigner is designed to improve the one-to-one label assignment. With this assigner, the training target class labels are gathered across stages and then reallocated to proper predictions at each decoder layer. On MS COCO benchmark, our model improves the baseline counterpart by 2.2 AP, and achieves a 44.8 AP with ResNet-50 as backbone, 100 queries and 12 training epochs. With longer training time and 300 queries, StageInteractor achieves 51.3 AP and 52.7 AP with ResNeXt-101-DCN and Swin-S, respectively. The code and models are made available at https://github.com/MCG-NJU/StageInteractor.", + "bbox": [ + 75, + 301, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 756, + 209, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection is a fundamental task in computer vision and acts as a cornerstone for many downstream tasks [48, 9]. It aims to localize and categorize all object instances in an image. Over the past few decades, dense spatial prior has been widely applied in various detectors. These detectors make predictions based on either", + "bbox": [ + 75, + 782, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3cff15939ecfe44854a861c2fc35019913918325c89f59b3b28ea3a9d657c2e.jpg", + "image_caption": [ + "Figure 1: Convergence curves of our model and other query-based object detectors [4, 68, 46, 15] with ResNet-50 [21] on MS COCO [31] minival set." + ], + "image_footnote": [], + "bbox": [ + 522, + 284, + 859, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a large quantity of pre-defined anchors covering the whole image [18, 41, 3, 33, 30, 40] or dense grid points in the feature map of this image [26, 13, 65, 49, 39, 57]. To deliver supervision signals to these detectors, most works employ the one-to-many label assignment strategy [62, 66, 25, 17, 16, 59] (i.e., the classification label and localization target of one ground-truth object could be assigned to multiple object predictions). Although this paradigm is widely used in object detection, it suffers from redundant and near-duplicate predictions due to such label assignment [45], and thus relies heavily on specific post-processing algorithms for duplicate removal [2, 23, 22].", + "bbox": [ + 496, + 537, + 893, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, DETR [4] and its variants [68, 46, 15, 7, 32, 14, 37, 8] open a new area of object detection. These query-based object detectors get rid of the dense prior and view object detection as a set prediction problem. Specifically, they use a sparse set of queries (i.e., learnable embeddings) to progressively capture the characteristics and location of objects with the help of a series of decoder layers. In each layer, image features are sampled and fused into the input queries via attention-like operation [51, 68] or dynamic mixing [46, 15]. Then, the transformed queries are decoded into object predictions and also serve as the inputs of next layer. As for the training of this paradigm, a kind of one-to", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\boxtimes$ Corresponding author (lmwang@nju.edu.cn).", + "bbox": [ + 99, + 886, + 364, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6577", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d676776bfd94b7ade44a472c0e0d0623b09f118a160fcdaaf50dd962f25e16fa.jpg", + "image_caption": [ + "Figure 2: The results of label assignment at various stages. The green box denotes the ground-truth object Person. The red and white boxes denote object prediction derived from two different queries. Pos and Neg denote the positive sample and the negative sample, respectively. (a) The white box is assigned with the ground-truth object Person by bipartite matching at the first stage, while the red box is not. But the opposite is true for the sixth stage. (b) With our cross-stage label assigner, the red box in the first stage can be assigned with the ground-truth Person." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "one label assignment (i.e., each ground-truth object has to be assigned with only one prediction), termed as bipartite matching, is independently adopted on each decoder layer for deep supervision [61, 1, 27]. For inference, only the high-confidence outputs from the last layer are taken for evaluation.", + "bbox": [ + 75, + 431, + 467, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, this one-to-one label assignment requires the detectors to have strong fine-grained discrimination and modeling capacity. On one hand, this strict bipartite matching would ask the detector to capture details to distinguish the predictions. For example, as shown in Fig. 2a, although the predicted boxes of a query (colored in red) cover most of the ground-truth object (Person) at each decoder layer, this object is assigned to the boxes of different queries at different stages (Stage1, white box; Stage6, red box). In other words, only one of the predicted boxes (red or white) can become the positive sample at each stage. To distinguish these boxes, the detector needs strong fine-grained discriminability to extract high-quality features from the image. Unfortunately, this goal is hard to be fully accomplished. As a result, we are motivated to directly modify the supervision of each decoder layer, i.e., introducing additional supervision from other stages to assist the training of this layer, shown in Fig. 2b. In addition, the large modeling capacity is vital to the fine-grained discrimination. To efficiently improve this modeling ability, we resort to adding lightweight modules and reusing heavy dynamic operators in decoder layers.", + "bbox": [ + 75, + 523, + 467, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above analysis, in this paper, we present a new paradigm, i.e., query-based object detector with cross-stage interaction, coined as StageInteractor. This interac", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion of our method lies in two aspects: cross-stage label assignment and cross-stage dynamic filter reuse. Specifically, during the label assignment, a cross-stage label assigner is applied subsequent to the vanilla bipartite matching in each decoder layer. This assigner collects the results of bipartite matching across stages, and then reallocate proper training target labels for each object prediction according to the query index and a score. As for the forward pass, in each decoder layer, we reuse the heavy dynamic operators in preceding stages and add lightweight modules to increase the modeling capacity at a relatively low cost.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments show that our model with ResNet-50 [21] as backbone can achieve 44.8 AP on MS COCO validation set under the simple setting of 100 queries, with $27.5\\mathrm{AP}_s$ , $48.0\\mathrm{AP}_m$ and $61.3\\mathrm{AP}_l$ on small, medium and large object detection, respectively. Equipped with $3\\times$ training time, 300 queries and more data augmentation in line with other query-based detectors, our model can achieve 49.9 AP, 51.3 AP, and 52.7 AP with ResNet-101, ResNeXt-101-DCN [55, 67], and Swin-S [34] as backbones, under the setting of single scale and single model testing. Our model significantly outperforms the previous methods, as shown in Fig. 1, and it has become a new state-of-the-art query-based object detector.", + "bbox": [ + 496, + 257, + 890, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 467, + 640, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DETR [4] is an end-to-end query-based object detector without hand-crafted designs such as preset anchors and non-maximum suppression (NMS), but it suffers from many problems, such as slow training convergence and unstable label assignment. To handle these challenges, a large quantity of works have been proposed. In this part, we divide these methods into two categories: modification of architectures and improvement of training procedures.", + "bbox": [ + 496, + 492, + 890, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Architecture. The vanilla DETR took a transformer encoder-decoder architecture [51] as the detection head, and used a set of object queries, content vectors, to encode the priors of object detection datasets. In Deformable DETR [68], the attention operator [51] was replaced with multi-scale deformable attention module, and iterative bounding box refinement and a two-stage design which enables the detector to adaptively generate queries with a dense prior were also introduced. In Sparse R-CNN [46], object query is decoupled into a content vector and an explicit bounding box, and image features are progressively aggregated into content vectors by ROAlign [20] with these boxes and dynamic convolution. Moreover, Sparse R-CNN is only a decoder architecture without any transformer encoder. There were also many works like Conditional DETR [36], DAB-DETR [32], SMCA [14] and REGO [8] studying how to introduce spatial priors for proper features to accelerate convergence. Adamixer [15] improved Sparse R-CNN via deformable operators and spa", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6578", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9acd87135f908b3db965cc710c03ac6cd37e3535da3b6f0b99bf6bc693e5991.jpg", + "image_caption": [ + "Figure 3: Overview. The cross-stage interaction incorporates two parts: cross-stage label assignment and cross-stage dynamic filter reuse. During the forward propagation, dynamic filters in each stage of decode layer are reused in the subsequent stages, i.e., we stack them with specific lightweight adapters to increase the depth of each decoder layer. As for the label assignment, our cross-stage label assigner gathers the results of bipartite matching across stages, and then selects proper target labels as supervision." + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 539, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0f85d807661aabe556f3cd3a8bf3c58c8d398bbd1662ad61984d5b87835e55c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 88, + 888, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tial dynamic convolution to increase the adaptability. There were also works like DDQ [63] which combines the dense and sparse queries to guarantee high recall. In this paper, we improve query-based object detector from a new perspective of the scalability [64, 44, 52] of the decoder, i.e., we reuse dynamic operators among decoder layers to capture more diverse and complex representations.", + "bbox": [ + 75, + 303, + 470, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training Procedure. In vanilla DETR, a set prediction loss [4] is adopted for training. Recently, many papers have analyzed how to accelerate the convergence of DETR via improved training procedures. To verify whether the instability of Hungarian loss slow down the convergence, Sun et al. [47] utilized a matching distillation with a pre-trained DETR providing label assignment to train another model, and they found this instability only influenced the convergence in the early few epochs. DN-DETR [28] presented a denoising training strategy where a set of additional noised ground-truth objects were passed through the decoder to reconstruct the corresponding raw objects. DINO-DETR [60] improved DN-DETR via contrastive denoising training for hard negative samples. Group DETR [6] introduced multiple groups of object queries for the global one-to-many label assignment during training, but maintained one-to-one label assignment in each group, and thus Group DETR could achieve the ability of duplicate removal with one group of queries. Hybrid Matching [24] was also proposed to combine one-to-one and one-to-many label assignment into one query-based object detector with a large number of queries, and it had three types of implementations: hybrid branch scheme (one-to-many matching for one group of queries, one-to-one matching for another), hybrid epoch scheme (one-to-many matching in early epochs, one-to-one matching in late epochs) and hybrid layer scheme (one-to-many matching in early layers, one-to-one in late layers). Different from the previous methods, in this paper, we focus on the calibration of label assignment without adding additional queries. We collect training target labels across stages by a cross-stage label assigner, and then select proper targets to act as the supervision of each object prediction.", + "bbox": [ + 75, + 417, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 500, + 301, + 689, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we focus on the cross-stage interaction in query-based object detectors because it can well mitigate the misalignment between the decoders and supervisions in an object detector. We first revisit the state-of-the-art query-based object detectors, especially AdaMixer [15], and then elaborate on our proposed cross-stage interaction.", + "bbox": [ + 496, + 329, + 890, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary on query-based object detectors", + "text_level": 1, + "bbox": [ + 498, + 434, + 877, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generally, the architecture of query-based object detectors is composed by four parts: object queries, a backbone, a series of encoders and decoders. Distinctively, Adamixer removes the encoders and still maintains the desired performance. As shown in Fig. 4, it consists of object queries, a backbone and a series of decoders.", + "bbox": [ + 496, + 459, + 890, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Query. The initial object queries is just a set of learnable embeddings. Recent object detectors [32, 46, 15] decompose them into content vectors and positional vectors. The content vector is a vector $\\mathbf{v} \\in \\mathbb{R}^D$ . The positional vector is presented in the format of box coordinates. For example, in AdaMixer, it contains the information about the center point, scale and aspect ratio of an individual box.", + "bbox": [ + 496, + 551, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decoder Layer. In a query-based detector, the decoder layers are stacked multiple times to form a cascade structure. Each decoder layer is generally composed of three components: a multi-head self-attention (MHSA), a dynamic interaction module, and feed-forward networks (FFNs). DETR-like models use the multi-head cross-attention for the dynamic interaction, while AdaMixer adopts a feature sampler and a dynamic mixing module, as shown in Fig. 4. The object queries are sequentially passed through these modules. Specifically, the queries are first processed by the multi-head self-attention module. Then, its outputs, the updated content vectors, together with the positional vectors are fed into the feature sampler. In this sampler, each query is allocated with a unique group of regional multi-scale image features, i.e., sampled features, by using the queries and bilinear interpolation. Subsequently,", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6579", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d80e662046a197fe209c3993ebdd5e61a89ffc487ec784cc4050b1ea9a882bc5.jpg", + "image_caption": [ + "Figure 4: Overview of AdaMixer." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 467, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the adaptive mixing is performed on the sampled features with dynamic filters generated from the content vectors. Its outputs are aggregated into the vectors. Last, by means of FFNs, the queries updated by these modules can be decoded into object predictions, i.e., the relative scaling and offsets to the positional vectors (bounding boxes), and the classification score vectors. These updated queries also serve as inputs of the next stage. Note that any query has and only has one corresponding prediction in every decoder layer. Thus, we simply represent the predictions derived from the same initial query with one index, i.e., the query index.", + "bbox": [ + 75, + 279, + 468, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training and testing. In each decoder layer of a query-based detector, a bipartite matching is directly adopted on the ground-truth objects and the predictions. After some predictions match the ground-truth, these predictions are deemed as positive samples (assigned with foreground classification labels and supervision for localization) while others are the negative (assigned with the background label). Focal loss [30] serves as the classification loss and GIoU loss [42] with $\\ell_1$ loss acts as the localization loss. During inference, only the outputs with high confidence from the last layer are used for evaluation.", + "bbox": [ + 75, + 446, + 468, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Cross-stage Label Assignment", + "text_level": 1, + "bbox": [ + 76, + 619, + 344, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate the training difficulty caused by bipartite matching in query-based object detectors, we propose a new cross-stage label assigner to modify the results of the preceding bipartite matching. As depicted in Fig. 3, our assigner first gathers these assignments across the stages (i.e., decoder layers), and then selects appropriate training target class labels for each prediction.", + "bbox": [ + 75, + 643, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When gathering training targets, the cross-stage label assigner forces each prediction to only access the targets of predictions sharing the same query index (defined in Sec. 3.1) with it. The motivation behind this constraint is that the supervision of a single query may vary across stages, even though its predicted boxes across stages continuously cover the same ground-truth object, shown in Fig. 2a. To alleviate this inconsistency for each query, we leverage its assigned targets provided by bipartite matching from multiple stages. When introducing targets, we use a", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7a8d37b6df55ca5408e976e4f83102b0599ad18c3e360c7b81f66f6c12f14198.jpg", + "image_caption": [], + "image_footnote": [ + "(a) Given a stage $i$ , we enumerate other decoder layers whose indexes are denoted as $j$ , and select their assigned targets as the supervision for the stage $i$ according to Eq. (1)." + ], + "bbox": [ + 506, + 89, + 880, + 334 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d523c3f0e66355bc89928b2248a72ce66850c60fa638d93147a77ca870ca0e60.jpg", + "image_caption": [ + "Figure 5: The process of our cross-stage label assignment." + ], + "image_footnote": [ + "(b) The elements in each candidate set are formed into the final supervision." + ], + "bbox": [ + 596, + 396, + 787, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "score to determine whether a match between a target and a prediction is suitable to be shared across stages.", + "bbox": [ + 498, + 612, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm. Thanks to Focal loss [30] with binary cross entropy loss in prevailing query-based object detectors [15, 46, 28, 60], the multi-class classification task can be viewed as the binary classification on each category, and there is no need to consider a separate background category when training. More importantly, our cross-stage label assignment can be conducted on each category independently. The specific algorithm is as follows:", + "bbox": [ + 496, + 643, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our cross-stage label assignment is performed between two stages with one type of condition. First, as shown in Fig. 5a, given a stage $i$ and its predictions $\\{p_1^i,\\dots,p_N^i\\}$ , we traverse other stages with indexes in the range $[\\alpha_i,\\beta_i]$ . Taking the stage $j$ as an example, we obtain its predictions $\\{p_1^j,\\dots,p_N^j\\}$ . Second, we also create an empty candidate bag for each prediction of the stage $i$ . We perform bipartite matching between the predictions of the stage $j$ and the ground-truth objects. Since this matching is an one-to-one", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6580", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ab7ce1873e091d395012a653a5fc2a0423eddca5dee83f1a890bf1327c304811.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAPAP50AP75
Baseline43.061.546.1
Hybrid Layer†41.8 (-1.2)60.5 (-1.0)44.7 (-1.4)
Ours44.8 (+1.8)63.0 (+1.5)48.4 (+2.3)
", + "bbox": [ + 78, + 88, + 464, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: We reproduce $(\\dagger)$ the hybrid matching scheme [24], and compare it with our cross-stage label assigner with 100 queries.", + "bbox": [ + 76, + 169, + 468, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "label assignment, each ground-truth object corresponds to a prediction with a specific query index at the stage $j$ . Third, we align the predictions from these two stages according to the query indexes, and thus each ground-truth object also corresponds to a candidate bag. We then select the training targets from the stage $j$ as the candidate supervisions according to the condition:", + "bbox": [ + 75, + 241, + 468, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\vartheta_ {i} (q, t) \\geq \\eta_ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 358, + 468, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\vartheta_{i}(q,t)$ denotes a score between the query $q$ and the ground-truth object $t$ at the stage $i$ . Here, the object $t$ needs to be assigned to query $q$ at stage $\\Phi_j$ by the bipartite matching. $\\eta_{i}$ denotes a threshold. In practice, we follow the classical setting [41], where the IoU as the score and the threshold is set as 0.5. Thus, if the IoU is satisfied, the ground-truth objects are added to the corresponding candidate bags. Finally, after traversing all the stages in the range $[\\alpha_i,\\beta_i]$ , we use the updated candidate bags to provide supervision for the stage $i$ . As shown in Figure 5b, we convert the labels in each candidate bag into one-hot vectors, and then merge them into one vector. The resulting vector serves as classification supervision through the binary cross entropy loss (Focal loss [30] in practice).", + "bbox": [ + 75, + 386, + 468, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. Although our label assignment seems to have something in common with existing works like TSP-RCNN [47] and Hybrid Layer Matching [24], the discrepancy between their design and ours cannot be ignored: (1) in TSP-RCNN, first, the idea from Faster-R-CNN [41] is directly borrowed into the set prediction problem (i.e., a ground-truth object is only assigned to the proposal that shares an IoU score greater than 0.5 with it). TSP-RCNN adopts such strategy for both the classification and localization. Differently, we can apply it only for the classification task, shown as Tab. 12. Second, TSP-RCNN adopts such strategy with dense proposals for both the classification and localization. Differently, we apply it with sparse queries. (2) for Hybrid Layer Matching, ground-truth objects are simply replicated in the first four stages, and then a bipartite matching is conducted. Differently, we do not modify the set of ground-truth objects. We gather results of the bipartite matching across stages, and then only select some targets as supervision. Also, we observe that Hybrid Layer Matching is incompatible with the models given few", + "bbox": [ + 75, + 598, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9ceda2cabc2cba71defc0e8528e54bdb9e9d2cf5b760567e72d3ee5db7a40327.jpg", + "image_caption": [ + "Figure 6: The dynamic mixing with filter reusing on sampled features. The content vectors dynamically generate filters through linear layers. These filters are used for mixing on sampled features, and then are stored into a dynamic filter bank for subsequent stages. Previous filters stored in the bank are also reused for mixing." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "queries. In Tab. 1, we implement it with the basic setting of our detector with 100 queries. The results show that it is totally infeasible under such a circumstance. In a nutshell, our approach is greatly different from these methods.", + "bbox": [ + 496, + 373, + 890, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Cross-stage Dynamic Filter Reuse", + "text_level": 1, + "bbox": [ + 500, + 441, + 795, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The model capacity is essential for neural networks to capture complex representations [44, 64]. Since one of the major characteristics shared by query-based detectors is a series of decoder layers, we resort to modifying the decoder to improve this capacity.", + "bbox": [ + 496, + 464, + 890, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A straightforward way of this modification is adding the attention-like operators. The prevailing DETR-like detectors [32, 28, 60] perform deformable cross-attention [68] once along the spatial dimension in each stage, so directly adding this attention is feasible. By contrast, other detectors [46, 15] like AdaMixer perform more than one dynamic mixing at each stage, and they require a huge number of parameters [53] to generate the corresponding dynamic filters. Taking the channel mixing as an example, in each decoder layer, the specific process of generating a filter is as follows:", + "bbox": [ + 496, + 541, + 890, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i} = \\mathbf {W} _ {0} ^ {(i)} + \\sum_ {d = 1} ^ {D} \\mathbf {W} _ {d} ^ {(i)} \\mathbf {v} _ {i, q, d} \\in \\mathbb {R} ^ {D _ {C} \\times D _ {C}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 699, + 890, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_{i,q,d}$ denotes the $d$ -th element of the content vector of query $q$ at the stage $i$ , $\\mathbf{W}_d^{(i)} \\in \\mathbb{R}^{D_C \\times D_C}$ denotes a learnable weight matrix, and the number of such matrices in a stage is $D$ , $D_C$ denotes the channel dimension of the input features, and $\\mathbf{M}_i$ serves as the kernel for channel mixing ( $1 \\times 1$ convolution) [50]. Only the parameters used to generate a single filter have already been more than $D \\times D_C^2$ . Thereby, it is impractical to directly stack more decoders given limited resources. Fortunately, there are a large quantity of disposable dynamic filters in these cascade decoder", + "bbox": [ + 496, + 746, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6581", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "layers, so these operators have the potential to be reused with lightweight modules across stages.", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As depicted in Fig. 6, we propose a cascade dynamic mixing module with a filter bank as an extension of the original channel mixing. The filters generated in each stage are stored in the filter bank for future use. Given a stage, the filter adapters update these stored filters with the ones from the current stage. In each adapter, the process of updating filters is as follows:", + "bbox": [ + 73, + 122, + 467, + 227 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {w} _ {j} ^ {(1)} = \\sigma \\left(\\mathbf {W} _ {j} ^ {(1)} \\mathbf {v} _ {i, q}\\right) \\in \\mathbb {R} ^ {D _ {C}}, \\mathbf {w} _ {j} ^ {(2)} = \\sigma \\left(\\mathbf {W} _ {j} ^ {(2)} \\mathbf {v} _ {i, q}\\right) \\in \\mathbb {R} ^ {D _ {C}}, \\\\ \\mathbf {M} _ {j, i} ^ {\\prime} = \\left(\\mathbf {w} _ {j} ^ {(1)} \\cdot \\mathbf {w} _ {j} ^ {(2) ^ {\\top}}\\right) \\odot \\mathbf {M} _ {j} + \\left(\\mathbf {1} - \\mathbf {w} _ {j} ^ {(1)} \\cdot \\mathbf {w} _ {j} ^ {(2) ^ {\\top}}\\right) \\odot \\mathbf {M} _ {i}, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 234, + 467, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where 1 denotes an all-ones matrix, $\\mathbf{W}_j^{(1)},\\mathbf{W}_j^{(2)}$ denote learnable weights, $\\sigma (\\cdot)$ denotes the sigmoid function, $\\odot$ denotes the element-wise product, $\\mathbf{M}_j$ denotes a previous filter, $j\\in [\\gamma_i,i)$ $\\gamma_{i}$ is a lower bound, $\\mathbf{M}_{j,i}^{\\prime}$ is used for subsequent modules, and the number of these updated filters is $\\Delta \\gamma_{i} = i - \\gamma_{i}$ .We empirically find more filters can lead to more performance gain, so $\\Delta \\gamma_{i} = i - 1$ is the best choice. Note that this adapter only costs $2D_C^2$ parameters, which is more lightweight than the process of generating a filter from scratch as Eq. (2). Then, the updated filters are used in the cascade dynamic mixing. In this structure, since each dynamic mixing is followed by a layer normalization with an activation function, we insert a lightweight linear layer between every two dynamic mixing, consistent with [21].", + "bbox": [ + 75, + 287, + 467, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, we can modify the original spatial mixing with the dynamic filter reusing. Unlike the channel mixing, the original dynamic spatial mixing costs a lot of computational resources on its output features due to its expansion [43] on spatial dimension. To tackle this problem, we also adopt the bank to store the filters generated for spatial mixing. Then, in each decoder layer, we update these filters with adapters and concatenate them with the current filter along the output dimension. This new filter is used to perform spatial mixing only once, rather than iteratively. This approach allows us to reduce the parameters of the filter generator of the spatial mixing, i.e., we employ the filters from preceding stages to replace a proportion of the filter from the current layer.", + "bbox": [ + 75, + 503, + 467, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 755, + 294, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experiments are performed on the MS COCO [31] object detection dataset, where the train2017 split and the val2017 split are for training and testing. All of our experiments on AdaMixer are based on mmdetection codebase [5]. The experiments on DN-DETR and DINO are based on DETREX codebase [11]. The convolutional neural networks [21, 56] or visiontrasformers [19, 34, 12] can be taken as the backbone network. 8 RTX 2080ti with 11G", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d81cfd0d17cbf01c930f3823276e9c5c86db84b7ee94258e4ce0f60f1c1747a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
detectorAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
FCOS [49]38.757.441.822.942.550.1
Cascade R-CNN [3]40.458.944.122.843.754.0
GFocalV2 [29]41.158.844.923.544.953.3
BorderDet [38]41.459.444.523.645.154.6
Dynamic Head [10]42.660.146.426.146.856.0
DETR [4]20.036.219.36.020.532.2
Deform-DETR [68]35.153.637.718.238.548.7
Sparse R-CNN [46]37.956.040.520.740.053.5
AdaMixer [15]42.761.545.924.745.459.2
\\( AdaMixer^† \\)[15]45.064.248.627.947.861.1
StageInteractor44.863.048.427.548.061.3
StageInteractor\\( ^† \\)46.965.251.130.049.762.3
", + "bbox": [ + 501, + 89, + 890, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: $1 \\times$ training scheme (12 epochs) performance of various detectors on COCO minival set with ResNet-50. 100 object queries is the default setting in our method. † denotes 500 queries.", + "bbox": [ + 498, + 313, + 890, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/35a19d461e427d7c9ff6b05192a23a45d0956bc4fe4eb9b21b8ffd82d13d236e.jpg", + "image_caption": [ + "Figure 7: The instability of the assigned labels." + ], + "image_footnote": [], + "bbox": [ + 503, + 382, + 885, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GPU memory are enough to train our model with ResNet-50 and 100 queries. For larger backbones or more queries, we resort to 8 V100 with 32G. The cross-stage label assignment is applied on each decoder layer. The reuse of spatial dynamic filters do not perform on the first two decoder layers. The weighted sum of these losses are for training, and the loss weights are in line with those of AdaMixer in mmdetection codebase [5] and those of DETRs in DETREX codebase [11]. AdamW [35] is taken as the optimizer.", + "bbox": [ + 496, + 577, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to State-of-the-Art Detectors", + "text_level": 1, + "bbox": [ + 500, + 724, + 854, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model significantly outperforms the previous methods, shown in Fig. 1, and it has become a new state-of-the-art query-based object detector. As shown in Tab. 2, with $1 \\times$ training scheme and ResNet-50 [21] as backbone, our detector outperforms various methods on COCO minival set [31] with even with 100 queries. Our model with ResNet-50 as backbone can achieve 44.8 AP on MS COCO validation set under the basic setting of 100 object queries, with $27.5\\mathrm{AP}_s$ , $48.0\\mathrm{AP}_m$ and $61.3\\mathrm{AP}_l$ on small, medium and large object detection, respectively. When", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6582", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/33fd8ea9a64bc69a8cff878af494b1892d5846efa667f506061563e636c6a080.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DetectorBackboneEncoder/FPNEpochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
DETR [4]ResNet-50-DC5TransformerEnc50043.363.145.922.547.361.1
SMCA [14]ResNet-50TransformerEnc5043.763.647.224.247.060.4
Deformable DETR [68]ResNet-50DeformTransEnc5043.862.647.726.447.158.0
Anchor DETR [54]ResNet-50-DC5DecoupTransEnc5044.264.747.524.748.260.6
Efficient DETR [58]ResNet-50DeformTransEnc3645.163.149.128.348.459.0
Conditional DETR [36]ResNet-50-DC5TransformerEnc10845.165.448.525.349.062.2
Sparse R-CNN [46]ResNet-50FPN3645.063.448.226.947.259.5
REGO [8]ResNet-50DeformTransEnc5047.666.851.629.650.662.3
DAB-D-DETR [32]ResNet-50DeformTransEnc5046.866.050.429.149.862.3
DN-DAB-D-DETR [28]ResNet-50DeformTransEnc1243.461.947.224.846.859.4
DN-DAB-D-DETR [28]ResNet-50DeformTransEnc5048.667.452.731.052.063.7
AdaMixer [15]ResNet-50-1244.163.147.829.547.058.8
AdaMixer [15]ResNet-50-2446.765.950.529.749.761.5
AdaMixer [15]ResNet-50-3647.066.051.130.150.261.8
StageInteractorResNet-50-1246.364.350.629.849.660.8
StageInteractorResNet-50-2448.366.652.931.751.463.3
StageInteractorResNet-50-3648.967.453.431.751.864.3
StageInteractor*ResNet-50-3650.866.855.934.054.666.2
DETR [4]ResNet-101-DC5TransformerEnc50044.964.747.723.749.562.3
SMCA [14]ResNet-101TransformerEnc5044.465.248.024.348.561.0
Efficient DETR [58]ResNet-101DeformTransEnc3645.764.149.528.249.160.2
Conditional DETR [36]ResNet-101-DC5TransformerEnc10845.966.849.527.250.363.3
Sparse R-CNN [46]ResNet-101FPN3646.464.649.528.348.361.6
REGO [8]ResNet-101DeformTransEnc5048.567.052.429.552.064.4
AdaMixer [15]ResNet-101-3648.067.052.430.051.263.7
StageInteractorResNet-101-3649.968.654.633.053.665.4
REGO [8]ResNeXt-101DeformTransEnc5049.167.553.130.052.665.0
AdaMixer [15]ResNeXt-101-DCN-3649.568.953.931.352.366.3
StageInteractorResNeXt-101-DCN-3651.370.256.033.254.666.9
AdaMixer [15]Swin-S-3651.371.255.734.254.667.3
StageInteractorSwin-S-3652.771.757.736.156.267.7
", + "bbox": [ + 78, + 87, + 890, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: The performance of various query-based detectors on MS COCO minival set with longer training scheme and single scale testing. The number of queries defaults to 300 in our method. * denotes 900 queries with more sampling points.", + "bbox": [ + 75, + 609, + 892, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "equipped with 500 queries, our detector performs better, and it can achieve 46.9 AP. As depicted in Tab. 3, equipped with $3 \\times$ training time, 300 queries and more data augmentation in line with other query-based object detectors, our model can achieve 48.9 AP, 49.9 AP, 51.3 AP, and 52.7 AP with ResNet-50, ResNet-101, ResNeXt-101-DCN [55, 67], and Swin-S [34] as backbones, under the setting of single scale and single model testing. Moreover, if we extend the quantity of queries into 900 and use tricks (i.e., adding more sampling points at each stage), our model can achieve 50.8 AP with ResNet-50.", + "bbox": [ + 75, + 667, + 470, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More importantly, our designs can be applied on DETR-like detectors such as DN-DETR [28] and DINO [60]. Unlike the cross-stage label assignment, our cross-stage filter reuse is tailored for AdaMixer [15]. To achieve larger model", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "capacity, we opt to simply double the cross-attention because this attention-like operation is more lightweight than the dynamic mixing. In Tab. 4, our designs yield more than $+0.5\\mathrm{AP}$ for all detectors. To verify the effectiveness of our cross-stage label assignment on the larger backbones, we use DINO with Swin-B [34] for experiments. As shown in Tab. 5, our cross-stage label assignment can still yield $+0.4\\mathrm{AP}$ on DINO with Swin-B as backbone with 12 epochs.", + "bbox": [ + 496, + 667, + 890, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To find the reason for the lower performance gain on DE-TRs than that on AdaMixer, we present the instability of assigned labels (i.e., the probability of a ground-truth object transferring from one query to another across stages) as curves in Fig. 7. This figure shows that Deformable-DETR is more stable, and we attribute this to its powerful transformer encoder which provides high-quality features for the", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6583", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2c35aa06ef98b2b64e2e8447c1c67665c7b5ebbcf19274e6edb0f67dabfda48b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DetectorEpochAP\\( AP_{50} \\)\\( AP_{75} \\)
DINO [60]1249.266.953.7
DINO + CSLA1249.767.054.1
DINO + CSLA + DCA1250.067.454.9
DINO [60]2450.668.555.5
DINO + CSLA2451.068.755.6
DINO + CSLA + DCA2451.369.256.1
Deform-DETR [68]5046.164.949.9
Deform-DETR + CSLA5046.865.050.9
DN-DETR [28]5044.765.347.6
DN-DETR + CSLA5045.565.349.0
H-DETR [24]1248.666.353.2
H-DETR + CSLA1249.266.353.7
", + "bbox": [ + 80, + 88, + 467, + 327 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f931d3a492883f8da238b7283ade05037ebe312e01fcc7469ce359d7acd2e4e0.jpg", + "table_caption": [ + "Table 4: The performance of DINO [60] on MS COCOMinival set with ResNet-50. CSLA: cross-stage label assignment. DCA: dual cross-attention." + ], + "table_footnote": [], + "table_body": "
DetectorBackboneAPAP50AP75
DINO [60]Swin-B55.874.460.7
DINO [60] + CSLASwin-B56.274.761.3
", + "bbox": [ + 80, + 391, + 465, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "decoder.", + "bbox": [ + 76, + 517, + 137, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 542, + 240, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Because the computational resource is limited, ResNet-50 is employed as our backbone network, the number of queries is set as 100, and $1 \\times$ training scheme is used for our ablation studies. Due to the simplicity, we adopt AdaMixer as the baseline for ablation studies.", + "bbox": [ + 76, + 566, + 468, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effectiveness of our proposed modules. In Tab. 6, we conduct experiments to verify the effectiveness of our proposed modules. The results show that our modules lead to $+2.2$ AP gain, and each of them boosts the performance.", + "bbox": [ + 76, + 642, + 468, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The components of our label assigner. We verify how the components of our cross-stage label assigner influences the performance in Tab. 7. In the first two lines, the results show that directly adding more supervisions based on the query index only brings marginal gains. According to the last two lines, we find that while using scores like IoU can boost the performance, using both IoU and the cross-stage labels leads to better results.", + "bbox": [ + 75, + 703, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The number of reused spatial filters. We explore the effectiveness of reusing spatial filters and the results are in Tab. 8. We find a certain number of reused filters can bring a slight performance gain. This may result from the easier optimization brought by the fewer model parameters.", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9c80a7f6324223a9d1778b858173e17c23b1dde8d715c914ac392b8beb54fdad.jpg", + "table_caption": [ + "Table 5: Cross-stage label assignment on DINO [60] with Swin-B [34] as backbone." + ], + "table_footnote": [], + "table_body": "
CSLAReuseAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
42.661.445.724.445.758.2
✓43.061.546.125.045.759.1
✓44.162.347.625.547.560.3
✓✓44.863.048.427.548.061.3
", + "bbox": [ + 504, + 88, + 885, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: The effectiveness of our proposed modules. CSLA: cross-stage label assignment.", + "bbox": [ + 498, + 184, + 890, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference speed and training memory use. As depicted in Tab. 9, with one TITAN XP and one batch size, the speed of our model is $11.6\\mathrm{img / s}$ while that of the baseline is $13.5\\mathrm{img / s}$ , i.e., only $1.16\\times$ slower. When we set the batch size as 2 for training, the additional operation only costs about 0.3G GPU memory (about $5\\%$ ).", + "bbox": [ + 496, + 244, + 890, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The number of additional channel mixing. We conduct ablation studies on how many previous filters are required for the channel mixing, and report performance in Tab. 10. We do experiments on the $\\max\\{\\Delta \\gamma_i\\}$ . This means there are at most $\\max\\{\\Delta \\gamma_i\\}$ filters reused for the $i$ -th stage. We find more filters can bring more performance gain. The results also show the scalability of the decoder layers in this framework.", + "bbox": [ + 496, + 337, + 892, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Selection of filters on mixing. We explore whether the previous filters with adapters are suitable for our new channel mixing. As shown in Tab. 11, we find that using the adapters to fuse previous filters with the current ones is most suitable, which ensures both the adaptability of each stage and the diversity [64] of filters. More importantly, the existence of the additional dynamic mixing is necessary.", + "bbox": [ + 496, + 458, + 890, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Modifying the localization supervision. In our cross-stage label assigner, only classification labels are used for gathering, while the supervision for localization is unchanged. Thus, we explore its influence by consistently updating the supervision of classification and localization, i.e. selecting the ground-truth boxes that have the greatest IoU with predicted boxes across stages. The results are shown in Tab. 12, and we find that the performance under the two settings is very close, so we do not modify the localization part for simplicity.", + "bbox": [ + 496, + 566, + 890, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Application scope of the cross-stage label assigner. For the $i$ -th stage, the application scope of cross-stage label assigner is $[\\alpha_i,\\beta_i]$ . In this part, we explore the appropriate values of $\\alpha_{i}$ and $\\beta_{i}$ . As shown in Tab. 13, we find that the best setting is $[i - 1,L]$ . Yet if too many previous ground-truth labels are included, like $\\alpha_{i} = 1$ , this hinders the last decoder layer to learn how to remove duplicates. Moreover, we conduct experiments to verify whether the cross-stage label assigner needs to be applied on all stages, because existing works [4, 24] demonstrate that only early stages can be applied with one-to-many label assignment. On the contrary, as shown in the last line of Tab. 13, we find that our", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6584", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1793b56f59a7e5a5d8bceaef8119dbe571e3f4e969196a069c53e86a18c84345.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CrossScoreAPAP50AP75APsAPmAPl
✓43.061.546.125.045.759.1
43.161.946.525.345.859.7
✓44.062.147.626.047.659.3
✓✓44.863.048.427.548.061.3
", + "bbox": [ + 84, + 88, + 455, + 179 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ea255782737af3fc1029c4788d3d636d5fc36164baf928d2a6fd4912e0eb986f.jpg", + "table_caption": [ + "Table 7: The effectiveness of the components of the cross-stage label assigner." + ], + "table_footnote": [], + "table_body": "
NsAPAP50AP75APsAPmAPl
044.562.548.127.448.060.6
144.863.048.427.548.061.3
344.462.448.026.247.460.6
", + "bbox": [ + 94, + 232, + 447, + 309 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/be1f99253be9cf41d45711e4692aed6dfa7311e5a38d78211aa7e44c614fa16f.jpg", + "table_caption": [ + "Table 8: The number of reused spatial filters." + ], + "table_footnote": [], + "table_body": "
MethodAPSpeed (img/s)Memory (GB)
Baseline42.613.56.0
Ours44.811.66.3
", + "bbox": [ + 94, + 345, + 447, + 407 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/9f77c26e4fa3e6639a1214f076b14d84512a1f1caf990824e0b46bb72fd0920f.jpg", + "table_caption": [ + "Table 9: The inference speed and training GPU memory use." + ], + "table_footnote": [], + "table_body": "
\\( \\max\\{\\Delta\\gamma_i\\} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
044.162.347.625.547.560.3
144.262.347.625.847.560.6
244.262.347.926.148.060.4
344.662.848.326.747.560.7
444.863.048.427.548.061.3
", + "bbox": [ + 93, + 460, + 450, + 566 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2a26540812679dccedbedb93e93f3c70b985c4603dc325014c2851ced24acb5e.jpg", + "table_caption": [ + "Table 10: The number of additional channel mixing." + ], + "table_footnote": [], + "table_body": "
FilterAPAP50AP75APsAPmAPl
Pre43.561.746.825.347.059.0
44.261.947.626.147.660.0
Cur44.562.747.926.547.760.5
Adp44.863.048.427.548.061.3
", + "bbox": [ + 89, + 604, + 454, + 696 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "cross-stage label assigner can also be applied on the last two stages and this brings performance gain. We speculate that this module helps remove some inappropriate supervision caused by vanilla bipartite matching which has a shortage of constraining IoU, and this module provides some proper supervision from the previous stage.", + "bbox": [ + 75, + 762, + 467, + 853 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The threshold of IoU. As shown in Tab. 14, we find that using a threshold of 0.5 is the best choice to select appropriate labels.", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6f48b89f3881d9329d590a371c6daa2339d3d1bf56480f4df8dd622285dc040f.jpg", + "table_caption": [ + "Table 11: Usage of filters. Pre, Cur: directly using previous and current filters. Adp: using adapters." + ], + "table_footnote": [], + "table_body": "
Loc.APAP50AP75APsAPmAPl
44.862.848.426.748.260.7
✓44.863.048.427.548.061.3
", + "bbox": [ + 514, + 88, + 874, + 148 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b0b7fe0d86bb85d67f49ce6a4190f48c5ba2662ee3f16449afaf23cdc5f3192c.jpg", + "table_caption": [ + "Table 12: Adding localization supervision in cross-stage label assigner." + ], + "table_footnote": [], + "table_body": "
Range of iαiβiAPAP50AP75APsAPmAPl
[1,L]ii44.062.147.626.047.659.3
[1,L]i-1i44.162.247.725.947.559.8
[1,L]i-1i+144.162.448.026.547.360.1
[1,L]i-1L44.863.048.427.548.061.3
[1,L]1L43.360.746.825.747.258.6
[1,L-2]i-1L44.362.447.726.147.660.4
", + "bbox": [ + 501, + 202, + 890, + 325 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e5dd0923f6a6ce4aac1ff87fc5c67be42d85c5d4d444aa1556c9602d1f2d3a41.jpg", + "table_caption": [ + "Table 13: The application scope of the cross-stage label assigner for each stage." + ], + "table_footnote": [], + "table_body": "
IoU thresAPAP50AP75APsAPmAPl
0.343.862.447.226.346.860.3
0.744.461.148.426.047.561.5
0.544.863.048.427.548.061.3
", + "bbox": [ + 521, + 377, + 867, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 14: The threshold of IoU in cross-stage label assignment.", + "bbox": [ + 500, + 455, + 890, + 483 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 513, + 617, + 529 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we have presented a new fast-converging query-based object detector with cross-stage interaction, termed as StageInteractor, to alleviate inconsistency between the one-to-one label assignment and improve the modeling capacity. Our proposed cross-stage label assigner gathers assigned labels across stages, and then reassigns proper labels to predictions in each stage. We also accumulate dynamic filter across stages and reuse them with lightweight modules to increase the modeling capacity without introducing too many parameters. With these two unique designs, StageInteractor significantly outperforms the previous methods, and our proposed label assignment can boost the performance of various query-based object detectors.", + "bbox": [ + 496, + 541, + 890, + 751 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization. The authors would like to thank Ziteng Gao for helpful discussions.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6585", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. Character-level language modeling with deeper self-attention. In Proceedings of the AAAI conference on artificial intelligence, pages 3159–3166, 2019. 2", + "[2] Navaneeth Bodla, Bharat Singh, Rama Chellappa, and Larry S Davis. Soft-nms-improving object detection with one line of code. In Proceedings of the IEEE international conference on computer vision, pages 5561-5569, 2017. 1", + "[3] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: high quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 1, 6", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European conference on computer vision, pages 213-229, 2020. 1, 2, 3, 6, 7, 8", + "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6", + "[6] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 3", + "[7] Xiaokang Chen, Fangyun Wei, Gang Zeng, and Jingdong Wang. Conditional detr v2: Efficient detection transformer with box queries. arXiv preprint arXiv:2207.08914, 2022. 1", + "[8] Zhe Chen, Jing Zhang, and Dacheng Tao. Recurrent glimpse-based decoder for detection with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5260-5269, 2022. 1, 2, 7", + "[9] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16372-16382, 2021. 1", + "[10] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 6", + "[11] detrex contributors. detrex: An research platform for transformer-based object detection algorithms, 2022. 6", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019. 1", + "[14] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3621–3630, 2021. 1, 2, 7", + "[15] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 1, 2, 3, 4, 5, 6, 7", + "[16] Ziteng Gao, Limin Wang, and Gangshan Wu. Mutual supervision for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3641-3650, 2021. 1", + "[17] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 303-312, 2021. 1", + "[18] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 1", + "[19] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 6", + "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 6", + "[22] Yihui He, Chenchen Zhu, Jianren Wang, Marios Savvides, and Xiangyu Zhang. Bounding box regression with uncertainty for accurate object detection. In Proceedings of the iee/ cvf conference on computer vision and pattern recognition, pages 2888-2897, 2019. 1", + "[23] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3588-3597, 2018. 1", + "[24] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 3, 5, 8", + "[25] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In European Conference on Computer Vision, pages 355-371. Springer, 2020. 1", + "[26] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision, pages 734-750, 2018. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6586", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou Zhang, and Zhuowen Tu. Deeply-supervised nets. In Artificial intelligence and statistics, pages 562-570. PMLR, 2015. 2", + "[28] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 3, 4, 5, 7, 8", + "[29] Xiang Li, Wenhai Wang, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss v2: Learning reliable localization quality estimation for dense object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11632-11641, 2021. 6", + "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 1, 4, 5", + "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 6", + "[32] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 3, 5, 7", + "[33] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pages 21-37. Springer, 2016. 1", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 6, 7, 8", + "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[36] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2, 7", + "[37] Duy-Kien Nguyen, Jihong Ju, Olaf Booij, Martin R Oswald, and Cees GM Snoek. Boxer: Box-attention for 2d and 3d transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4773–4782, 2022. 1", + "[38] Han Qiu, Yuchen Ma, Zeming Li, Songtao Liu, and Jian Sun. Borderdet: Border feature for dense object detection. In European Conference on Computer Vision, pages 549-564. Springer, 2020. 6", + "[39] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 1", + "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 5", + "[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 4", + "[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6", + "[44] Zhiqiang Shen, Zechun Liu, and Eric Xing. Sliced recursive transformer. arXiv preprint arXiv:2111.05297, 2021. 3, 5", + "[45] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In International Conference on Machine Learning, pages 9934-9944. PMLR, 2021. 1", + "[46] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 1, 2, 3, 4, 5, 6, 7", + "[47] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3611-3620, 2021. 3, 5", + "[48] Yao Teng and Limin Wang. Structured sparse r-cnn for direct scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19437-19446, 2022. 1", + "[49] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 1, 6", + "[50] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261–24272, 2021. 5", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2", + "[52] Hongyu Wang, Shuming Ma, Li Dong, Shaohan Huang, Dongdong Zhang, and Furu Wei. Deepnet: Scaling transformers to 1,000 layers. arXiv preprint arXiv:2203.00555, 2022.3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6587", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Wen Wang, Jing Zhang, Yang Cao, Yongliang Shen, and Dacheng Tao. Towards data-efficient detection transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 88-105. Springer, 2022. 5", + "[54] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 7", + "[55] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500, 2017. 2, 7", + "[56] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 6", + "[57] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9657-9666, 2019. 1", + "[58] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 7", + "[59] Mohsen Zand, Ali Etemad, and Michael Greenspan. Object-box: From centers to boxes for anchor-free object detection. arXiv preprint arXiv:2207.06985, 2022. 1", + "[60] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5, 7, 8", + "[61] Linfeng Zhang, Xin Chen, Junbo Zhang, Runpei Dong, and Kaisheng Ma. Contrastive deep supervision. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 2", + "[62] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020. 1", + "[63] Shilong Zhang, Xinjiang Wang, Jiaqi Wang, Jiangmiao Pang, Chengqi Lyu, Wenwei Zhang, Ping Luo, and Kai Chen. Dense distinct query for end-to-end object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7329-7338, 2023. 3", + "[64] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 3, 5, 8", + "[65] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 1", + "[66] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020.1" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 2, 7", + "[68] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1, 2, 5, 6, 7, 8" + ], + "bbox": [ + 501, + 90, + 892, + 204 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6588", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/0e05e637-ac63-484f-9954-22b778c7c120_model.json b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/0e05e637-ac63-484f-9954-22b778c7c120_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c0ee6ea4458a7b233b35fccb2a11aa057e73ae20 --- /dev/null +++ b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/0e05e637-ac63-484f-9954-22b778c7c120_model.json @@ -0,0 +1,2347 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "StageInteractor: Query-based Object Detector with Cross-stage Interaction" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.179, + 0.733, + 0.199 + ], + "angle": 0, + "content": "Yao Teng\\(^{1}\\) Haisong Liu\\(^{1}\\) Sheng Guo\\(^{3}\\) Limin Wang\\(^{1,2,\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.199, + 0.807, + 0.217 + ], + "angle": 0, + "content": "1State Key Laboratory for Novel Software Technology, Nanjing University, China" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.217, + 0.713, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)Shanghai AI Lab, China \\(^{3}\\)MYbank, Ant Group, China" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.286 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.725 + ], + "angle": 0, + "content": "Previous object detectors make predictions based on dense grid points or numerous preset anchors. Most of these detectors are trained with one-to-many label assignment strategies. On the contrary, recent query-based object detectors are based a sparse set of learnable queries refined by a series of decoder layers. The one-to-one label assignment is independently applied on each layer for deep supervision during training. Despite the great success of query-based object detection, however, this vanilla one-to-one label assignment strategy requires the detectors to have strong fine-grained discrimination and modeling capacity. In this paper, we propose a new query-based object detector with cross-stage interaction, coined as StageInteractor. During the forward pass, we come up with an efficient way to improve this modeling ability by reusing dynamic operators with lightweight adapters. As for the label assignment, a cross-stage label assigner is designed to improve the one-to-one label assignment. With this assigner, the training target class labels are gathered across stages and then reallocated to proper predictions at each decoder layer. On MS COCO benchmark, our model improves the baseline counterpart by 2.2 AP, and achieves a 44.8 AP with ResNet-50 as backbone, 100 queries and 12 training epochs. With longer training time and 300 queries, StageInteractor achieves 51.3 AP and 52.7 AP with ResNeXt-101-DCN and Swin-S, respectively. The code and models are made available at https://github.com/MCG-NJU/StageInteractor." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.757, + 0.21, + 0.772 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.783, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Object detection is a fundamental task in computer vision and acts as a cornerstone for many downstream tasks [48, 9]. It aims to localize and categorize all object instances in an image. Over the past few decades, dense spatial prior has been widely applied in various detectors. These detectors make predictions based on either" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.285, + 0.861, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.466, + 0.895, + 0.512 + ], + "angle": 0, + "content": "Figure 1: Convergence curves of our model and other query-based object detectors [4, 68, 46, 15] with ResNet-50 [21] on MS COCO [31] minival set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.719 + ], + "angle": 0, + "content": "a large quantity of pre-defined anchors covering the whole image [18, 41, 3, 33, 30, 40] or dense grid points in the feature map of this image [26, 13, 65, 49, 39, 57]. To deliver supervision signals to these detectors, most works employ the one-to-many label assignment strategy [62, 66, 25, 17, 16, 59] (i.e., the classification label and localization target of one ground-truth object could be assigned to multiple object predictions). Although this paradigm is widely used in object detection, it suffers from redundant and near-duplicate predictions due to such label assignment [45], and thus relies heavily on specific post-processing algorithms for duplicate removal [2, 23, 22]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Recently, DETR [4] and its variants [68, 46, 15, 7, 32, 14, 37, 8] open a new area of object detection. These query-based object detectors get rid of the dense prior and view object detection as a set prediction problem. Specifically, they use a sparse set of queries (i.e., learnable embeddings) to progressively capture the characteristics and location of objects with the help of a series of decoder layers. In each layer, image features are sampled and fused into the input queries via attention-like operation [51, 68] or dynamic mixing [46, 15]. Then, the transformed queries are decoded into object predictions and also serve as the inputs of next layer. As for the training of this paradigm, a kind of one-to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.365, + 0.902 + ], + "angle": 0, + "content": "\\(\\boxtimes\\) Corresponding author (lmwang@nju.edu.cn)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "6577" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.47, + 0.418 + ], + "angle": 0, + "content": "Figure 2: The results of label assignment at various stages. The green box denotes the ground-truth object Person. The red and white boxes denote object prediction derived from two different queries. Pos and Neg denote the positive sample and the negative sample, respectively. (a) The white box is assigned with the ground-truth object Person by bipartite matching at the first stage, while the red box is not. But the opposite is true for the sixth stage. (b) With our cross-stage label assigner, the red box in the first stage can be assigned with the ground-truth Person." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.468, + 0.522 + ], + "angle": 0, + "content": "one label assignment (i.e., each ground-truth object has to be assigned with only one prediction), termed as bipartite matching, is independently adopted on each decoder layer for deep supervision [61, 1, 27]. For inference, only the high-confidence outputs from the last layer are taken for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.468, + 0.853 + ], + "angle": 0, + "content": "However, this one-to-one label assignment requires the detectors to have strong fine-grained discrimination and modeling capacity. On one hand, this strict bipartite matching would ask the detector to capture details to distinguish the predictions. For example, as shown in Fig. 2a, although the predicted boxes of a query (colored in red) cover most of the ground-truth object (Person) at each decoder layer, this object is assigned to the boxes of different queries at different stages (Stage1, white box; Stage6, red box). In other words, only one of the predicted boxes (red or white) can become the positive sample at each stage. To distinguish these boxes, the detector needs strong fine-grained discriminability to extract high-quality features from the image. Unfortunately, this goal is hard to be fully accomplished. As a result, we are motivated to directly modify the supervision of each decoder layer, i.e., introducing additional supervision from other stages to assist the training of this layer, shown in Fig. 2b. In addition, the large modeling capacity is vital to the fine-grained discrimination. To efficiently improve this modeling ability, we resort to adding lightweight modules and reusing heavy dynamic operators in decoder layers." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Based on the above analysis, in this paper, we present a new paradigm, i.e., query-based object detector with cross-stage interaction, coined as StageInteractor. This interac" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "tion of our method lies in two aspects: cross-stage label assignment and cross-stage dynamic filter reuse. Specifically, during the label assignment, a cross-stage label assigner is applied subsequent to the vanilla bipartite matching in each decoder layer. This assigner collects the results of bipartite matching across stages, and then reallocate proper training target labels for each object prediction according to the query index and a score. As for the forward pass, in each decoder layer, we reuse the heavy dynamic operators in preceding stages and add lightweight modules to increase the modeling capacity at a relatively low cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.455 + ], + "angle": 0, + "content": "Experiments show that our model with ResNet-50 [21] as backbone can achieve 44.8 AP on MS COCO validation set under the simple setting of 100 queries, with \\(27.5\\mathrm{AP}_s\\), \\(48.0\\mathrm{AP}_m\\) and \\(61.3\\mathrm{AP}_l\\) on small, medium and large object detection, respectively. Equipped with \\(3\\times\\) training time, 300 queries and more data augmentation in line with other query-based detectors, our model can achieve 49.9 AP, 51.3 AP, and 52.7 AP with ResNet-101, ResNeXt-101-DCN [55, 67], and Swin-S [34] as backbones, under the setting of single scale and single model testing. Our model significantly outperforms the previous methods, as shown in Fig. 1, and it has become a new state-of-the-art query-based object detector." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.642, + 0.484 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.614 + ], + "angle": 0, + "content": "DETR [4] is an end-to-end query-based object detector without hand-crafted designs such as preset anchors and non-maximum suppression (NMS), but it suffers from many problems, such as slow training convergence and unstable label assignment. To handle these challenges, a large quantity of works have been proposed. In this part, we divide these methods into two categories: modification of architectures and improvement of training procedures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Architecture. The vanilla DETR took a transformer encoder-decoder architecture [51] as the detection head, and used a set of object queries, content vectors, to encode the priors of object detection datasets. In Deformable DETR [68], the attention operator [51] was replaced with multi-scale deformable attention module, and iterative bounding box refinement and a two-stage design which enables the detector to adaptively generate queries with a dense prior were also introduced. In Sparse R-CNN [46], object query is decoupled into a content vector and an explicit bounding box, and image features are progressively aggregated into content vectors by ROAlign [20] with these boxes and dynamic convolution. Moreover, Sparse R-CNN is only a decoder architecture without any transformer encoder. There were also many works like Conditional DETR [36], DAB-DETR [32], SMCA [14] and REGO [8] studying how to introduce spatial priors for proper features to accelerate convergence. Adamixer [15] improved Sparse R-CNN via deformable operators and spa" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6578" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.54, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.089, + 0.889, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.202, + 0.894, + 0.279 + ], + "angle": 0, + "content": "Figure 3: Overview. The cross-stage interaction incorporates two parts: cross-stage label assignment and cross-stage dynamic filter reuse. During the forward propagation, dynamic filters in each stage of decode layer are reused in the subsequent stages, i.e., we stack them with specific lightweight adapters to increase the depth of each decoder layer. As for the label assignment, our cross-stage label assigner gathers the results of bipartite matching across stages, and then selects proper target labels as supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.471, + 0.412 + ], + "angle": 0, + "content": "tial dynamic convolution to increase the adaptability. There were also works like DDQ [63] which combines the dense and sparse queries to guarantee high recall. In this paper, we improve query-based object detector from a new perspective of the scalability [64, 44, 52] of the decoder, i.e., we reuse dynamic operators among decoder layers to capture more diverse and complex representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Training Procedure. In vanilla DETR, a set prediction loss [4] is adopted for training. Recently, many papers have analyzed how to accelerate the convergence of DETR via improved training procedures. To verify whether the instability of Hungarian loss slow down the convergence, Sun et al. [47] utilized a matching distillation with a pre-trained DETR providing label assignment to train another model, and they found this instability only influenced the convergence in the early few epochs. DN-DETR [28] presented a denoising training strategy where a set of additional noised ground-truth objects were passed through the decoder to reconstruct the corresponding raw objects. DINO-DETR [60] improved DN-DETR via contrastive denoising training for hard negative samples. Group DETR [6] introduced multiple groups of object queries for the global one-to-many label assignment during training, but maintained one-to-one label assignment in each group, and thus Group DETR could achieve the ability of duplicate removal with one group of queries. Hybrid Matching [24] was also proposed to combine one-to-one and one-to-many label assignment into one query-based object detector with a large number of queries, and it had three types of implementations: hybrid branch scheme (one-to-many matching for one group of queries, one-to-one matching for another), hybrid epoch scheme (one-to-many matching in early epochs, one-to-one matching in late epochs) and hybrid layer scheme (one-to-many matching in early layers, one-to-one in late layers). Different from the previous methods, in this paper, we focus on the calibration of label assignment without adding additional queries. We collect training target labels across stages by a cross-stage label assigner, and then select proper targets to act as the supervision of each object prediction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.303, + 0.691, + 0.321 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.422 + ], + "angle": 0, + "content": "In this paper, we focus on the cross-stage interaction in query-based object detectors because it can well mitigate the misalignment between the decoders and supervisions in an object detector. We first revisit the state-of-the-art query-based object detectors, especially AdaMixer [15], and then elaborate on our proposed cross-stage interaction." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.435, + 0.878, + 0.452 + ], + "angle": 0, + "content": "3.1. Preliminary on query-based object detectors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.55 + ], + "angle": 0, + "content": "Generally, the architecture of query-based object detectors is composed by four parts: object queries, a backbone, a series of encoders and decoders. Distinctively, Adamixer removes the encoders and still maintains the desired performance. As shown in Fig. 4, it consists of object queries, a backbone and a series of decoders." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Object Query. The initial object queries is just a set of learnable embeddings. Recent object detectors [32, 46, 15] decompose them into content vectors and positional vectors. The content vector is a vector \\(\\mathbf{v} \\in \\mathbb{R}^D\\). The positional vector is presented in the format of box coordinates. For example, in AdaMixer, it contains the information about the center point, scale and aspect ratio of an individual box." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Decoder Layer. In a query-based detector, the decoder layers are stacked multiple times to form a cascade structure. Each decoder layer is generally composed of three components: a multi-head self-attention (MHSA), a dynamic interaction module, and feed-forward networks (FFNs). DETR-like models use the multi-head cross-attention for the dynamic interaction, while AdaMixer adopts a feature sampler and a dynamic mixing module, as shown in Fig. 4. The object queries are sequentially passed through these modules. Specifically, the queries are first processed by the multi-head self-attention module. Then, its outputs, the updated content vectors, together with the positional vectors are fed into the feature sampler. In this sampler, each query is allocated with a unique group of regional multi-scale image features, i.e., sampled features, by using the queries and bilinear interpolation. Subsequently," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6579" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.468, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.24, + 0.384, + 0.254 + ], + "angle": 0, + "content": "Figure 4: Overview of AdaMixer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.47, + 0.446 + ], + "angle": 0, + "content": "the adaptive mixing is performed on the sampled features with dynamic filters generated from the content vectors. Its outputs are aggregated into the vectors. Last, by means of FFNs, the queries updated by these modules can be decoded into object predictions, i.e., the relative scaling and offsets to the positional vectors (bounding boxes), and the classification score vectors. These updated queries also serve as inputs of the next stage. Note that any query has and only has one corresponding prediction in every decoder layer. Thus, we simply represent the predictions derived from the same initial query with one index, i.e., the query index." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.469, + 0.611 + ], + "angle": 0, + "content": "Training and testing. In each decoder layer of a query-based detector, a bipartite matching is directly adopted on the ground-truth objects and the predictions. After some predictions match the ground-truth, these predictions are deemed as positive samples (assigned with foreground classification labels and supervision for localization) while others are the negative (assigned with the background label). Focal loss [30] serves as the classification loss and GIoU loss [42] with \\(\\ell_1\\) loss acts as the localization loss. During inference, only the outputs with high confidence from the last layer are used for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.345, + 0.637 + ], + "angle": 0, + "content": "3.2. Cross-stage Label Assignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.749 + ], + "angle": 0, + "content": "To mitigate the training difficulty caused by bipartite matching in query-based object detectors, we propose a new cross-stage label assigner to modify the results of the preceding bipartite matching. As depicted in Fig. 3, our assigner first gathers these assignments across the stages (i.e., decoder layers), and then selects appropriate training target class labels for each prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "When gathering training targets, the cross-stage label assigner forces each prediction to only access the targets of predictions sharing the same query index (defined in Sec. 3.1) with it. The motivation behind this constraint is that the supervision of a single query may vary across stages, even though its predicted boxes across stages continuously cover the same ground-truth object, shown in Fig. 2a. To alleviate this inconsistency for each query, we leverage its assigned targets provided by bipartite matching from multiple stages. When introducing targets, we use a" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.882, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.501, + 0.341, + 0.89, + 0.384 + ], + "angle": 0, + "content": "(a) Given a stage \\(i\\), we enumerate other decoder layers whose indexes are denoted as \\(j\\), and select their assigned targets as the supervision for the stage \\(i\\) according to Eq. (1)." + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.397, + 0.789, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.5, + 0.535, + 0.888, + 0.563 + ], + "angle": 0, + "content": "(b) The elements in each candidate set are formed into the final supervision." + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.569, + 0.888, + 0.585 + ], + "angle": 0, + "content": "Figure 5: The process of our cross-stage label assignment." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.613, + 0.891, + 0.643 + ], + "angle": 0, + "content": "score to determine whether a match between a target and a prediction is suitable to be shared across stages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Algorithm. Thanks to Focal loss [30] with binary cross entropy loss in prevailing query-based object detectors [15, 46, 28, 60], the multi-class classification task can be viewed as the binary classification on each category, and there is no need to consider a separate background category when training. More importantly, our cross-stage label assignment can be conducted on each category independently. The specific algorithm is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our cross-stage label assignment is performed between two stages with one type of condition. First, as shown in Fig. 5a, given a stage \\(i\\) and its predictions \\(\\{p_1^i,\\dots,p_N^i\\}\\), we traverse other stages with indexes in the range \\([\\alpha_i,\\beta_i]\\). Taking the stage \\(j\\) as an example, we obtain its predictions \\(\\{p_1^j,\\dots,p_N^j\\}\\). Second, we also create an empty candidate bag for each prediction of the stage \\(i\\). We perform bipartite matching between the predictions of the stage \\(j\\) and the ground-truth objects. Since this matching is an one-to-one" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6580" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.465, + 0.166 + ], + "angle": 0, + "content": "
MethodAPAP50AP75
Baseline43.061.546.1
Hybrid Layer†41.8 (-1.2)60.5 (-1.0)44.7 (-1.4)
Ours44.8 (+1.8)63.0 (+1.5)48.4 (+2.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.17, + 0.469, + 0.216 + ], + "angle": 0, + "content": "Table 1: We reproduce \\((\\dagger)\\) the hybrid matching scheme [24], and compare it with our cross-stage label assigner with 100 queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.242, + 0.469, + 0.348 + ], + "angle": 0, + "content": "label assignment, each ground-truth object corresponds to a prediction with a specific query index at the stage \\( j \\). Third, we align the predictions from these two stages according to the query indexes, and thus each ground-truth object also corresponds to a candidate bag. We then select the training targets from the stage \\( j \\) as the candidate supervisions according to the condition:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.359, + 0.469, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\vartheta_ {i} (q, t) \\geq \\eta_ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.469, + 0.599 + ], + "angle": 0, + "content": "where \\(\\vartheta_{i}(q,t)\\) denotes a score between the query \\(q\\) and the ground-truth object \\(t\\) at the stage \\(i\\). Here, the object \\(t\\) needs to be assigned to query \\(q\\) at stage \\(\\Phi_j\\) by the bipartite matching. \\(\\eta_{i}\\) denotes a threshold. In practice, we follow the classical setting [41], where the IoU as the score and the threshold is set as 0.5. Thus, if the IoU is satisfied, the ground-truth objects are added to the corresponding candidate bags. Finally, after traversing all the stages in the range \\([\\alpha_i,\\beta_i]\\), we use the updated candidate bags to provide supervision for the stage \\(i\\). As shown in Figure 5b, we convert the labels in each candidate bag into one-hot vectors, and then merge them into one vector. The resulting vector serves as classification supervision through the binary cross entropy loss (Focal loss [30] in practice)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Discussion. Although our label assignment seems to have something in common with existing works like TSP-RCNN [47] and Hybrid Layer Matching [24], the discrepancy between their design and ours cannot be ignored: (1) in TSP-RCNN, first, the idea from Faster-R-CNN [41] is directly borrowed into the set prediction problem (i.e., a ground-truth object is only assigned to the proposal that shares an IoU score greater than 0.5 with it). TSP-RCNN adopts such strategy for both the classification and localization. Differently, we can apply it only for the classification task, shown as Tab. 12. Second, TSP-RCNN adopts such strategy with dense proposals for both the classification and localization. Differently, we apply it with sparse queries. (2) for Hybrid Layer Matching, ground-truth objects are simply replicated in the first four stages, and then a bipartite matching is conducted. Differently, we do not modify the set of ground-truth objects. We gather results of the bipartite matching across stages, and then only select some targets as supervision. Also, we observe that Hybrid Layer Matching is incompatible with the models given few" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.09, + 0.891, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Figure 6: The dynamic mixing with filter reusing on sampled features. The content vectors dynamically generate filters through linear layers. These filters are used for mixing on sampled features, and then are stored into a dynamic filter bank for subsequent stages. Previous filters stored in the bank are also reused for mixing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.435 + ], + "angle": 0, + "content": "queries. In Tab. 1, we implement it with the basic setting of our detector with 100 queries. The results show that it is totally infeasible under such a circumstance. In a nutshell, our approach is greatly different from these methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.442, + 0.796, + 0.459 + ], + "angle": 0, + "content": "3.3. Cross-stage Dynamic Filter Reuse" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.892, + 0.542 + ], + "angle": 0, + "content": "The model capacity is essential for neural networks to capture complex representations [44, 64]. Since one of the major characteristics shared by query-based detectors is a series of decoder layers, we resort to modifying the decoder to improve this capacity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.693 + ], + "angle": 0, + "content": "A straightforward way of this modification is adding the attention-like operators. The prevailing DETR-like detectors [32, 28, 60] perform deformable cross-attention [68] once along the spatial dimension in each stage, so directly adding this attention is feasible. By contrast, other detectors [46, 15] like AdaMixer perform more than one dynamic mixing at each stage, and they require a huge number of parameters [53] to generate the corresponding dynamic filters. Taking the channel mixing as an example, in each decoder layer, the specific process of generating a filter is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.7, + 0.892, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i} = \\mathbf {W} _ {0} ^ {(i)} + \\sum_ {d = 1} ^ {D} \\mathbf {W} _ {d} ^ {(i)} \\mathbf {v} _ {i, q, d} \\in \\mathbb {R} ^ {D _ {C} \\times D _ {C}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_{i,q,d}\\) denotes the \\(d\\)-th element of the content vector of query \\(q\\) at the stage \\(i\\), \\(\\mathbf{W}_d^{(i)} \\in \\mathbb{R}^{D_C \\times D_C}\\) denotes a learnable weight matrix, and the number of such matrices in a stage is \\(D\\), \\(D_C\\) denotes the channel dimension of the input features, and \\(\\mathbf{M}_i\\) serves as the kernel for channel mixing (\\(1 \\times 1\\) convolution) [50]. Only the parameters used to generate a single filter have already been more than \\(D \\times D_C^2\\). Thereby, it is impractical to directly stack more decoders given limited resources. Fortunately, there are a large quantity of disposable dynamic filters in these cascade decoder" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6581" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "layers, so these operators have the potential to be reused with lightweight modules across stages." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.123, + 0.468, + 0.228 + ], + "angle": 0, + "content": "As depicted in Fig. 6, we propose a cascade dynamic mixing module with a filter bank as an extension of the original channel mixing. The filters generated in each stage are stored in the filter bank for future use. Given a stage, the filter adapters update these stored filters with the ones from the current stage. In each adapter, the process of updating filters is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.236, + 0.468, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {w} _ {j} ^ {(1)} = \\sigma \\left(\\mathbf {W} _ {j} ^ {(1)} \\mathbf {v} _ {i, q}\\right) \\in \\mathbb {R} ^ {D _ {C}}, \\mathbf {w} _ {j} ^ {(2)} = \\sigma \\left(\\mathbf {W} _ {j} ^ {(2)} \\mathbf {v} _ {i, q}\\right) \\in \\mathbb {R} ^ {D _ {C}}, \\\\ \\mathbf {M} _ {j, i} ^ {\\prime} = \\left(\\mathbf {w} _ {j} ^ {(1)} \\cdot \\mathbf {w} _ {j} ^ {(2) ^ {\\top}}\\right) \\odot \\mathbf {M} _ {j} + \\left(\\mathbf {1} - \\mathbf {w} _ {j} ^ {(1)} \\cdot \\mathbf {w} _ {j} ^ {(2) ^ {\\top}}\\right) \\odot \\mathbf {M} _ {i}, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.468, + 0.502 + ], + "angle": 0, + "content": "where 1 denotes an all-ones matrix, \\(\\mathbf{W}_j^{(1)},\\mathbf{W}_j^{(2)}\\) denote learnable weights, \\(\\sigma (\\cdot)\\) denotes the sigmoid function, \\(\\odot\\) denotes the element-wise product, \\(\\mathbf{M}_j\\) denotes a previous filter, \\(j\\in [\\gamma_i,i)\\) \\(\\gamma_{i}\\) is a lower bound, \\(\\mathbf{M}_{j,i}^{\\prime}\\) is used for subsequent modules, and the number of these updated filters is \\(\\Delta \\gamma_{i} = i - \\gamma_{i}\\) .We empirically find more filters can lead to more performance gain, so \\(\\Delta \\gamma_{i} = i - 1\\) is the best choice. Note that this adapter only costs \\(2D_C^2\\) parameters, which is more lightweight than the process of generating a filter from scratch as Eq. (2). Then, the updated filters are used in the cascade dynamic mixing. In this structure, since each dynamic mixing is followed by a layer normalization with an activation function, we insert a lightweight linear layer between every two dynamic mixing, consistent with [21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.468, + 0.716 + ], + "angle": 0, + "content": "Moreover, we can modify the original spatial mixing with the dynamic filter reusing. Unlike the channel mixing, the original dynamic spatial mixing costs a lot of computational resources on its output features due to its expansion [43] on spatial dimension. To tackle this problem, we also adopt the bank to store the filters generated for spatial mixing. Then, in each decoder layer, we update these filters with adapters and concatenate them with the current filter along the output dimension. This new filter is used to perform spatial mixing only once, rather than iteratively. This approach allows us to reduce the parameters of the filter generator of the spatial mixing, i.e., we employ the filters from preceding stages to replace a proportion of the filter from the current layer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.731, + 0.21, + 0.748 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.295, + 0.772 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The experiments are performed on the MS COCO [31] object detection dataset, where the train2017 split and the val2017 split are for training and testing. All of our experiments on AdaMixer are based on mmdetection codebase [5]. The experiments on DN-DETR and DINO are based on DETREX codebase [11]. The convolutional neural networks [21, 56] or visiontrasformers [19, 34, 12] can be taken as the backbone network. 8 RTX 2080ti with 11G" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.309 + ], + "angle": 0, + "content": "
detectorAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
FCOS [49]38.757.441.822.942.550.1
Cascade R-CNN [3]40.458.944.122.843.754.0
GFocalV2 [29]41.158.844.923.544.953.3
BorderDet [38]41.459.444.523.645.154.6
Dynamic Head [10]42.660.146.426.146.856.0
DETR [4]20.036.219.36.020.532.2
Deform-DETR [68]35.153.637.718.238.548.7
Sparse R-CNN [46]37.956.040.520.740.053.5
AdaMixer [15]42.761.545.924.745.459.2
\\( AdaMixer^† \\)[15]45.064.248.627.947.861.1
StageInteractor44.863.048.427.548.061.3
StageInteractor\\( ^† \\)46.965.251.130.049.762.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.314, + 0.892, + 0.375 + ], + "angle": 0, + "content": "Table 2: \\(1 \\times\\) training scheme (12 epochs) performance of various detectors on COCO minival set with ResNet-50. 100 object queries is the default setting in our method. † denotes 500 queries." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.383, + 0.886, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.542, + 0.851, + 0.557 + ], + "angle": 0, + "content": "Figure 7: The instability of the assigned labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.716 + ], + "angle": 0, + "content": "GPU memory are enough to train our model with ResNet-50 and 100 queries. For larger backbones or more queries, we resort to 8 V100 with 32G. The cross-stage label assignment is applied on each decoder layer. The reuse of spatial dynamic filters do not perform on the first two decoder layers. The weighted sum of these losses are for training, and the loss weights are in line with those of AdaMixer in mmdetection codebase [5] and those of DETRs in DETREX codebase [11]. AdamW [35] is taken as the optimizer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.855, + 0.742 + ], + "angle": 0, + "content": "4.2. Comparison to State-of-the-Art Detectors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our model significantly outperforms the previous methods, shown in Fig. 1, and it has become a new state-of-the-art query-based object detector. As shown in Tab. 2, with \\(1 \\times\\) training scheme and ResNet-50 [21] as backbone, our detector outperforms various methods on COCO minival set [31] with even with 100 queries. Our model with ResNet-50 as backbone can achieve 44.8 AP on MS COCO validation set under the basic setting of 100 object queries, with \\(27.5\\mathrm{AP}_s\\), \\(48.0\\mathrm{AP}_m\\) and \\(61.3\\mathrm{AP}_l\\) on small, medium and large object detection, respectively. When" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6582" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.088, + 0.891, + 0.608 + ], + "angle": 0, + "content": "
DetectorBackboneEncoder/FPNEpochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
DETR [4]ResNet-50-DC5TransformerEnc50043.363.145.922.547.361.1
SMCA [14]ResNet-50TransformerEnc5043.763.647.224.247.060.4
Deformable DETR [68]ResNet-50DeformTransEnc5043.862.647.726.447.158.0
Anchor DETR [54]ResNet-50-DC5DecoupTransEnc5044.264.747.524.748.260.6
Efficient DETR [58]ResNet-50DeformTransEnc3645.163.149.128.348.459.0
Conditional DETR [36]ResNet-50-DC5TransformerEnc10845.165.448.525.349.062.2
Sparse R-CNN [46]ResNet-50FPN3645.063.448.226.947.259.5
REGO [8]ResNet-50DeformTransEnc5047.666.851.629.650.662.3
DAB-D-DETR [32]ResNet-50DeformTransEnc5046.866.050.429.149.862.3
DN-DAB-D-DETR [28]ResNet-50DeformTransEnc1243.461.947.224.846.859.4
DN-DAB-D-DETR [28]ResNet-50DeformTransEnc5048.667.452.731.052.063.7
AdaMixer [15]ResNet-50-1244.163.147.829.547.058.8
AdaMixer [15]ResNet-50-2446.765.950.529.749.761.5
AdaMixer [15]ResNet-50-3647.066.051.130.150.261.8
StageInteractorResNet-50-1246.364.350.629.849.660.8
StageInteractorResNet-50-2448.366.652.931.751.463.3
StageInteractorResNet-50-3648.967.453.431.751.864.3
StageInteractor*ResNet-50-3650.866.855.934.054.666.2
DETR [4]ResNet-101-DC5TransformerEnc50044.964.747.723.749.562.3
SMCA [14]ResNet-101TransformerEnc5044.465.248.024.348.561.0
Efficient DETR [58]ResNet-101DeformTransEnc3645.764.149.528.249.160.2
Conditional DETR [36]ResNet-101-DC5TransformerEnc10845.966.849.527.250.363.3
Sparse R-CNN [46]ResNet-101FPN3646.464.649.528.348.361.6
REGO [8]ResNet-101DeformTransEnc5048.567.052.429.552.064.4
AdaMixer [15]ResNet-101-3648.067.052.430.051.263.7
StageInteractorResNet-101-3649.968.654.633.053.665.4
REGO [8]ResNeXt-101DeformTransEnc5049.167.553.130.052.665.0
AdaMixer [15]ResNeXt-101-DCN-3649.568.953.931.352.366.3
StageInteractorResNeXt-101-DCN-3651.370.256.033.254.666.9
AdaMixer [15]Swin-S-3651.371.255.734.254.667.3
StageInteractorSwin-S-3652.771.757.736.156.267.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.611, + 0.893, + 0.645 + ], + "angle": 0, + "content": "Table 3: The performance of various query-based detectors on MS COCO minival set with longer training scheme and single scale testing. The number of queries defaults to 300 in our method. * denotes 900 queries with more sampling points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.471, + 0.835 + ], + "angle": 0, + "content": "equipped with 500 queries, our detector performs better, and it can achieve 46.9 AP. As depicted in Tab. 3, equipped with \\(3 \\times\\) training time, 300 queries and more data augmentation in line with other query-based object detectors, our model can achieve 48.9 AP, 49.9 AP, 51.3 AP, and 52.7 AP with ResNet-50, ResNet-101, ResNeXt-101-DCN [55, 67], and Swin-S [34] as backbones, under the setting of single scale and single model testing. Moreover, if we extend the quantity of queries into 900 and use tricks (i.e., adding more sampling points at each stage), our model can achieve 50.8 AP with ResNet-50." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "More importantly, our designs can be applied on DETR-like detectors such as DN-DETR [28] and DINO [60]. Unlike the cross-stage label assignment, our cross-stage filter reuse is tailored for AdaMixer [15]. To achieve larger model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.79 + ], + "angle": 0, + "content": "capacity, we opt to simply double the cross-attention because this attention-like operation is more lightweight than the dynamic mixing. In Tab. 4, our designs yield more than \\(+0.5\\mathrm{AP}\\) for all detectors. To verify the effectiveness of our cross-stage label assignment on the larger backbones, we use DINO with Swin-B [34] for experiments. As shown in Tab. 5, our cross-stage label assignment can still yield \\(+0.4\\mathrm{AP}\\) on DINO with Swin-B as backbone with 12 epochs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To find the reason for the lower performance gain on DE-TRs than that on AdaMixer, we present the instability of assigned labels (i.e., the probability of a ground-truth object transferring from one query to another across stages) as curves in Fig. 7. This figure shows that Deformable-DETR is more stable, and we attribute this to its powerful transformer encoder which provides high-quality features for the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6583" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.328 + ], + "angle": 0, + "content": "
DetectorEpochAP\\( AP_{50} \\)\\( AP_{75} \\)
DINO [60]1249.266.953.7
DINO + CSLA1249.767.054.1
DINO + CSLA + DCA1250.067.454.9
DINO [60]2450.668.555.5
DINO + CSLA2451.068.755.6
DINO + CSLA + DCA2451.369.256.1
Deform-DETR [68]5046.164.949.9
Deform-DETR + CSLA5046.865.050.9
DN-DETR [28]5044.765.347.6
DN-DETR + CSLA5045.565.349.0
H-DETR [24]1248.666.353.2
H-DETR + CSLA1249.266.353.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.331, + 0.47, + 0.378 + ], + "angle": 0, + "content": "Table 4: The performance of DINO [60] on MS COCOMinival set with ResNet-50. CSLA: cross-stage label assignment. DCA: dual cross-attention." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.392, + 0.466, + 0.455 + ], + "angle": 0, + "content": "
DetectorBackboneAPAP50AP75
DINO [60]Swin-B55.874.460.7
DINO [60] + CSLASwin-B56.274.761.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.458, + 0.469, + 0.489 + ], + "angle": 0, + "content": "Table 5: Cross-stage label assignment on DINO [60] with Swin-B [34] as backbone." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.518, + 0.138, + 0.531 + ], + "angle": 0, + "content": "decoder." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.543, + 0.241, + 0.558 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.567, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Because the computational resource is limited, ResNet-50 is employed as our backbone network, the number of queries is set as 100, and \\(1 \\times\\) training scheme is used for our ablation studies. Due to the simplicity, we adopt AdaMixer as the baseline for ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.643, + 0.47, + 0.704 + ], + "angle": 0, + "content": "The effectiveness of our proposed modules. In Tab. 6, we conduct experiments to verify the effectiveness of our proposed modules. The results show that our modules lead to \\(+2.2\\) AP gain, and each of them boosts the performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.824 + ], + "angle": 0, + "content": "The components of our label assigner. We verify how the components of our cross-stage label assigner influences the performance in Tab. 7. In the first two lines, the results show that directly adding more supervisions based on the query index only brings marginal gains. According to the last two lines, we find that while using scores like IoU can boost the performance, using both IoU and the cross-stage labels leads to better results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.903 + ], + "angle": 0, + "content": "The number of reused spatial filters. We explore the effectiveness of reusing spatial filters and the results are in Tab. 8. We find a certain number of reused filters can bring a slight performance gain. This may result from the easier optimization brought by the fewer model parameters." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.886, + 0.181 + ], + "angle": 0, + "content": "
CSLAReuseAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
42.661.445.724.445.758.2
✓43.061.546.125.045.759.1
✓44.162.347.625.547.560.3
✓✓44.863.048.427.548.061.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.185, + 0.892, + 0.216 + ], + "angle": 0, + "content": "Table 6: The effectiveness of our proposed modules. CSLA: cross-stage label assignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.336 + ], + "angle": 0, + "content": "Inference speed and training memory use. As depicted in Tab. 9, with one TITAN XP and one batch size, the speed of our model is \\(11.6\\mathrm{img / s}\\) while that of the baseline is \\(13.5\\mathrm{img / s}\\), i.e., only \\(1.16\\times\\) slower. When we set the batch size as 2 for training, the additional operation only costs about 0.3G GPU memory (about \\(5\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.893, + 0.459 + ], + "angle": 0, + "content": "The number of additional channel mixing. We conduct ablation studies on how many previous filters are required for the channel mixing, and report performance in Tab. 10. We do experiments on the \\(\\max\\{\\Delta \\gamma_i\\}\\). This means there are at most \\(\\max\\{\\Delta \\gamma_i\\}\\) filters reused for the \\(i\\)-th stage. We find more filters can bring more performance gain. The results also show the scalability of the decoder layers in this framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Selection of filters on mixing. We explore whether the previous filters with adapters are suitable for our new channel mixing. As shown in Tab. 11, we find that using the adapters to fuse previous filters with the current ones is most suitable, which ensures both the adaptability of each stage and the diversity [64] of filters. More importantly, the existence of the additional dynamic mixing is necessary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Modifying the localization supervision. In our cross-stage label assigner, only classification labels are used for gathering, while the supervision for localization is unchanged. Thus, we explore its influence by consistently updating the supervision of classification and localization, i.e. selecting the ground-truth boxes that have the greatest IoU with predicted boxes across stages. The results are shown in Tab. 12, and we find that the performance under the two settings is very close, so we do not modify the localization part for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Application scope of the cross-stage label assigner. For the \\(i\\)-th stage, the application scope of cross-stage label assigner is \\([\\alpha_i,\\beta_i]\\). In this part, we explore the appropriate values of \\(\\alpha_{i}\\) and \\(\\beta_{i}\\). As shown in Tab. 13, we find that the best setting is \\([i - 1,L]\\). Yet if too many previous ground-truth labels are included, like \\(\\alpha_{i} = 1\\), this hinders the last decoder layer to learn how to remove duplicates. Moreover, we conduct experiments to verify whether the cross-stage label assigner needs to be applied on all stages, because existing works [4, 24] demonstrate that only early stages can be applied with one-to-many label assignment. On the contrary, as shown in the last line of Tab. 13, we find that our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6584" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.457, + 0.18 + ], + "angle": 0, + "content": "
CrossScoreAPAP50AP75APsAPmAPl
✓43.061.546.125.045.759.1
43.161.946.525.345.859.7
✓44.062.147.626.047.659.3
✓✓44.863.048.427.548.061.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.185, + 0.468, + 0.217 + ], + "angle": 0, + "content": "Table 7: The effectiveness of the components of the cross-stage label assigner." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.233, + 0.449, + 0.31 + ], + "angle": 0, + "content": "
NsAPAP50AP75APsAPmAPl
044.562.548.127.448.060.6
144.863.048.427.548.061.3
344.462.448.026.247.460.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.314, + 0.421, + 0.33 + ], + "angle": 0, + "content": "Table 8: The number of reused spatial filters." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.347, + 0.449, + 0.409 + ], + "angle": 0, + "content": "
MethodAPSpeed (img/s)Memory (GB)
Baseline42.613.56.0
Ours44.811.66.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.414, + 0.468, + 0.443 + ], + "angle": 0, + "content": "Table 9: The inference speed and training GPU memory use." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.462, + 0.451, + 0.568 + ], + "angle": 0, + "content": "
\\( \\max\\{\\Delta\\gamma_i\\} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
044.162.347.625.547.560.3
144.262.347.625.847.560.6
244.262.347.926.148.060.4
344.662.848.326.747.560.7
444.863.048.427.548.061.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.574, + 0.446, + 0.589 + ], + "angle": 0, + "content": "Table 10: The number of additional channel mixing." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.606, + 0.455, + 0.698 + ], + "angle": 0, + "content": "
FilterAPAP50AP75APsAPmAPl
Pre43.561.746.825.347.059.0
44.261.947.626.147.660.0
Cur44.562.747.926.547.760.5
Adp44.863.048.427.548.061.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.703, + 0.468, + 0.733 + ], + "angle": 0, + "content": "Table 11: Usage of filters. Pre, Cur: directly using previous and current filters. Adp: using adapters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.468, + 0.854 + ], + "angle": 0, + "content": "cross-stage label assigner can also be applied on the last two stages and this brings performance gain. We speculate that this module helps remove some inappropriate supervision caused by vanilla bipartite matching which has a shortage of constraining IoU, and this module provides some proper supervision from the previous stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The threshold of IoU. As shown in Tab. 14, we find that using a threshold of 0.5 is the best choice to select appropriate labels." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.875, + 0.15 + ], + "angle": 0, + "content": "
Loc.APAP50AP75APsAPmAPl
44.862.848.426.748.260.7
✓44.863.048.427.548.061.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.155, + 0.892, + 0.186 + ], + "angle": 0, + "content": "Table 12: Adding localization supervision in cross-stage label assigner." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.203, + 0.891, + 0.326 + ], + "angle": 0, + "content": "
Range of iαiβiAPAP50AP75APsAPmAPl
[1,L]ii44.062.147.626.047.659.3
[1,L]i-1i44.162.247.725.947.559.8
[1,L]i-1i+144.162.448.026.547.360.1
[1,L]i-1L44.863.048.427.548.061.3
[1,L]1L43.360.746.825.747.258.6
[1,L-2]i-1L44.362.447.726.147.660.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.331, + 0.891, + 0.361 + ], + "angle": 0, + "content": "Table 13: The application scope of the cross-stage label assigner for each stage." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.378, + 0.868, + 0.451 + ], + "angle": 0, + "content": "
IoU thresAPAP50AP75APsAPmAPl
0.343.862.447.226.346.860.3
0.744.461.148.426.047.561.5
0.544.863.048.427.548.061.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.456, + 0.891, + 0.484 + ], + "angle": 0, + "content": "Table 14: The threshold of IoU in cross-stage label assignment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.618, + 0.53 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.752 + ], + "angle": 0, + "content": "In this paper, we have presented a new fast-converging query-based object detector with cross-stage interaction, termed as StageInteractor, to alleviate inconsistency between the one-to-one label assignment and improve the modeling capacity. Our proposed cross-stage label assigner gathers assigned labels across stages, and then reassigns proper labels to predictions in each stage. We also accumulate dynamic filter across stages and reuse them with lightweight modules to increase the modeling capacity without introducing too many parameters. With these two unique designs, StageInteractor significantly outperforms the previous methods, and our proposed label assignment can boost the performance of various query-based object detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization. The authors would like to thank Ziteng Gao for helpful discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6585" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. Character-level language modeling with deeper self-attention. In Proceedings of the AAAI conference on artificial intelligence, pages 3159–3166, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.229 + ], + "angle": 0, + "content": "[2] Navaneeth Bodla, Bharat Singh, Rama Chellappa, and Larry S Davis. Soft-nms-improving object detection with one line of code. In Proceedings of the IEEE international conference on computer vision, pages 5561-5569, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[3] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: high quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.289, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European conference on computer vision, pages 213-229, 2020. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.47, + 0.47 + ], + "angle": 0, + "content": "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.469, + 0.527 + ], + "angle": 0, + "content": "[6] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[7] Xiaokang Chen, Fangyun Wei, Gang Zeng, and Jingdong Wang. Conditional detr v2: Efficient detection transformer with box queries. arXiv preprint arXiv:2207.08914, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[8] Zhe Chen, Jing Zhang, and Dacheng Tao. Recurrent glimpse-based decoder for detection with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5260-5269, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[9] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16372-16382, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[10] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[11] detrex contributors. detrex: An research platform for transformer-based object detection algorithms, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[14] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3621–3630, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[15] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[16] Ziteng Gao, Limin Wang, and Gangshan Wu. Mutual supervision for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3641-3650, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[17] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 303-312, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.424, + 0.892, + 0.464 + ], + "angle": 0, + "content": "[18] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.468, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[19] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[22] Yihui He, Chenchen Zhu, Jianren Wang, Marios Savvides, and Xiangyu Zhang. Bounding box regression with uncertainty for accurate object detection. In Proceedings of the iee/ cvf conference on computer vision and pattern recognition, pages 2888-2897, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[23] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3588-3597, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[24] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[25] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In European Conference on Computer Vision, pages 355-371. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision, pages 734-750, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6586" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[27] Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou Zhang, and Zhuowen Tu. Deeply-supervised nets. In Artificial intelligence and statistics, pages 562-570. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[28] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.469, + 0.292 + ], + "angle": 0, + "content": "[29] Xiang Li, Wenhai Wang, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss v2: Learning reliable localization quality estimation for dense object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11632-11641, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.469, + 0.349 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.424, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[32] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[33] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pages 21-37. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.469, + 0.61 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.469, + 0.713 + ], + "angle": 0, + "content": "[36] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.783 + ], + "angle": 0, + "content": "[37] Duy-Kien Nguyen, Jihong Ju, Olaf Booij, Martin R Oswald, and Cees GM Snoek. Boxer: Box-attention for 2d and 3d transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4773–4782, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[38] Han Qiu, Yuchen Ma, Zeming Li, Songtao Liu, and Jian Sun. Borderdet: Border feature for dense object detection. In European Conference on Computer Vision, pages 549-564. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[40] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[44] Zhiqiang Shen, Zechun Liu, and Eric Xing. Sliced recursive transformer. arXiv preprint arXiv:2111.05297, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[45] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In International Conference on Machine Learning, pages 9934-9944. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[46] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[47] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3611-3620, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[48] Yao Teng and Limin Wang. Structured sparse r-cnn for direct scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19437-19446, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[49] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[50] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261–24272, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[52] Hongyu Wang, Shuming Ma, Li Dong, Shaohan Huang, Dongdong Zhang, and Furu Wei. Deepnet: Scaling transformers to 1,000 layers. arXiv preprint arXiv:2203.00555, 2022.3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "6587" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[53] Wen Wang, Jing Zhang, Yang Cao, Yongliang Shen, and Dacheng Tao. Towards data-efficient detection transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 88-105. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[54] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[55] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.355 + ], + "angle": 0, + "content": "[56] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.357, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[57] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9657-9666, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[58] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.496 + ], + "angle": 0, + "content": "[59] Mohsen Zand, Ali Etemad, and Michael Greenspan. Object-box: From centers to boxes for anchor-free object detection. arXiv preprint arXiv:2207.06985, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[60] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[61] Linfeng Zhang, Xin Chen, Junbo Zhang, Runpei Dong, and Kaisheng Ma. Contrastive deep supervision. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[62] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[63] Shilong Zhang, Xinjiang Wang, Jiaqi Wang, Jiangmiao Pang, Chengqi Lyu, Wenwei Zhang, Ping Luo, and Kai Chen. Dense distinct query for end-to-end object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7329-7338, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[64] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[65] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[66] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020.1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.148 + ], + "angle": 0, + "content": "[67] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.205 + ], + "angle": 0, + "content": "[68] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1, 2, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6588" + } + ] +] \ No newline at end of file diff --git a/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/0e05e637-ac63-484f-9954-22b778c7c120_origin.pdf b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/0e05e637-ac63-484f-9954-22b778c7c120_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c4ca8e0186c8c0c799627d912d0989b4030f86fb --- /dev/null +++ b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/0e05e637-ac63-484f-9954-22b778c7c120_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:078d4fe446ea001b99afc86ae0092fd64253e2c1ed9b11d6f57116b6cd930932 +size 894701 diff --git a/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/full.md b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d3357e6886728e45fb603f8dc3b76eca149128fc --- /dev/null +++ b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/full.md @@ -0,0 +1,314 @@ +# StageInteractor: Query-based Object Detector with Cross-stage Interaction + +Yao Teng $^{1}$ Haisong Liu $^{1}$ Sheng Guo $^{3}$ Limin Wang $^{1,2,\boxtimes}$ + +1State Key Laboratory for Novel Software Technology, Nanjing University, China + +$^{2}$ Shanghai AI Lab, China $^{3}$ MYbank, Ant Group, China + +# Abstract + +Previous object detectors make predictions based on dense grid points or numerous preset anchors. Most of these detectors are trained with one-to-many label assignment strategies. On the contrary, recent query-based object detectors are based a sparse set of learnable queries refined by a series of decoder layers. The one-to-one label assignment is independently applied on each layer for deep supervision during training. Despite the great success of query-based object detection, however, this vanilla one-to-one label assignment strategy requires the detectors to have strong fine-grained discrimination and modeling capacity. In this paper, we propose a new query-based object detector with cross-stage interaction, coined as StageInteractor. During the forward pass, we come up with an efficient way to improve this modeling ability by reusing dynamic operators with lightweight adapters. As for the label assignment, a cross-stage label assigner is designed to improve the one-to-one label assignment. With this assigner, the training target class labels are gathered across stages and then reallocated to proper predictions at each decoder layer. On MS COCO benchmark, our model improves the baseline counterpart by 2.2 AP, and achieves a 44.8 AP with ResNet-50 as backbone, 100 queries and 12 training epochs. With longer training time and 300 queries, StageInteractor achieves 51.3 AP and 52.7 AP with ResNeXt-101-DCN and Swin-S, respectively. The code and models are made available at https://github.com/MCG-NJU/StageInteractor. + +# 1. Introduction + +Object detection is a fundamental task in computer vision and acts as a cornerstone for many downstream tasks [48, 9]. It aims to localize and categorize all object instances in an image. Over the past few decades, dense spatial prior has been widely applied in various detectors. These detectors make predictions based on either + +![](images/f3cff15939ecfe44854a861c2fc35019913918325c89f59b3b28ea3a9d657c2e.jpg) +Figure 1: Convergence curves of our model and other query-based object detectors [4, 68, 46, 15] with ResNet-50 [21] on MS COCO [31] minival set. + +a large quantity of pre-defined anchors covering the whole image [18, 41, 3, 33, 30, 40] or dense grid points in the feature map of this image [26, 13, 65, 49, 39, 57]. To deliver supervision signals to these detectors, most works employ the one-to-many label assignment strategy [62, 66, 25, 17, 16, 59] (i.e., the classification label and localization target of one ground-truth object could be assigned to multiple object predictions). Although this paradigm is widely used in object detection, it suffers from redundant and near-duplicate predictions due to such label assignment [45], and thus relies heavily on specific post-processing algorithms for duplicate removal [2, 23, 22]. + +Recently, DETR [4] and its variants [68, 46, 15, 7, 32, 14, 37, 8] open a new area of object detection. These query-based object detectors get rid of the dense prior and view object detection as a set prediction problem. Specifically, they use a sparse set of queries (i.e., learnable embeddings) to progressively capture the characteristics and location of objects with the help of a series of decoder layers. In each layer, image features are sampled and fused into the input queries via attention-like operation [51, 68] or dynamic mixing [46, 15]. Then, the transformed queries are decoded into object predictions and also serve as the inputs of next layer. As for the training of this paradigm, a kind of one-to + +![](images/d676776bfd94b7ade44a472c0e0d0623b09f118a160fcdaaf50dd962f25e16fa.jpg) +Figure 2: The results of label assignment at various stages. The green box denotes the ground-truth object Person. The red and white boxes denote object prediction derived from two different queries. Pos and Neg denote the positive sample and the negative sample, respectively. (a) The white box is assigned with the ground-truth object Person by bipartite matching at the first stage, while the red box is not. But the opposite is true for the sixth stage. (b) With our cross-stage label assigner, the red box in the first stage can be assigned with the ground-truth Person. + +one label assignment (i.e., each ground-truth object has to be assigned with only one prediction), termed as bipartite matching, is independently adopted on each decoder layer for deep supervision [61, 1, 27]. For inference, only the high-confidence outputs from the last layer are taken for evaluation. + +However, this one-to-one label assignment requires the detectors to have strong fine-grained discrimination and modeling capacity. On one hand, this strict bipartite matching would ask the detector to capture details to distinguish the predictions. For example, as shown in Fig. 2a, although the predicted boxes of a query (colored in red) cover most of the ground-truth object (Person) at each decoder layer, this object is assigned to the boxes of different queries at different stages (Stage1, white box; Stage6, red box). In other words, only one of the predicted boxes (red or white) can become the positive sample at each stage. To distinguish these boxes, the detector needs strong fine-grained discriminability to extract high-quality features from the image. Unfortunately, this goal is hard to be fully accomplished. As a result, we are motivated to directly modify the supervision of each decoder layer, i.e., introducing additional supervision from other stages to assist the training of this layer, shown in Fig. 2b. In addition, the large modeling capacity is vital to the fine-grained discrimination. To efficiently improve this modeling ability, we resort to adding lightweight modules and reusing heavy dynamic operators in decoder layers. + +Based on the above analysis, in this paper, we present a new paradigm, i.e., query-based object detector with cross-stage interaction, coined as StageInteractor. This interac + +tion of our method lies in two aspects: cross-stage label assignment and cross-stage dynamic filter reuse. Specifically, during the label assignment, a cross-stage label assigner is applied subsequent to the vanilla bipartite matching in each decoder layer. This assigner collects the results of bipartite matching across stages, and then reallocate proper training target labels for each object prediction according to the query index and a score. As for the forward pass, in each decoder layer, we reuse the heavy dynamic operators in preceding stages and add lightweight modules to increase the modeling capacity at a relatively low cost. + +Experiments show that our model with ResNet-50 [21] as backbone can achieve 44.8 AP on MS COCO validation set under the simple setting of 100 queries, with $27.5\mathrm{AP}_s$ , $48.0\mathrm{AP}_m$ and $61.3\mathrm{AP}_l$ on small, medium and large object detection, respectively. Equipped with $3\times$ training time, 300 queries and more data augmentation in line with other query-based detectors, our model can achieve 49.9 AP, 51.3 AP, and 52.7 AP with ResNet-101, ResNeXt-101-DCN [55, 67], and Swin-S [34] as backbones, under the setting of single scale and single model testing. Our model significantly outperforms the previous methods, as shown in Fig. 1, and it has become a new state-of-the-art query-based object detector. + +# 2. Related Work + +DETR [4] is an end-to-end query-based object detector without hand-crafted designs such as preset anchors and non-maximum suppression (NMS), but it suffers from many problems, such as slow training convergence and unstable label assignment. To handle these challenges, a large quantity of works have been proposed. In this part, we divide these methods into two categories: modification of architectures and improvement of training procedures. + +Architecture. The vanilla DETR took a transformer encoder-decoder architecture [51] as the detection head, and used a set of object queries, content vectors, to encode the priors of object detection datasets. In Deformable DETR [68], the attention operator [51] was replaced with multi-scale deformable attention module, and iterative bounding box refinement and a two-stage design which enables the detector to adaptively generate queries with a dense prior were also introduced. In Sparse R-CNN [46], object query is decoupled into a content vector and an explicit bounding box, and image features are progressively aggregated into content vectors by ROAlign [20] with these boxes and dynamic convolution. Moreover, Sparse R-CNN is only a decoder architecture without any transformer encoder. There were also many works like Conditional DETR [36], DAB-DETR [32], SMCA [14] and REGO [8] studying how to introduce spatial priors for proper features to accelerate convergence. Adamixer [15] improved Sparse R-CNN via deformable operators and spa + +![](images/a9acd87135f908b3db965cc710c03ac6cd37e3535da3b6f0b99bf6bc693e5991.jpg) +Figure 3: Overview. The cross-stage interaction incorporates two parts: cross-stage label assignment and cross-stage dynamic filter reuse. During the forward propagation, dynamic filters in each stage of decode layer are reused in the subsequent stages, i.e., we stack them with specific lightweight adapters to increase the depth of each decoder layer. As for the label assignment, our cross-stage label assigner gathers the results of bipartite matching across stages, and then selects proper target labels as supervision. + +![](images/0f85d807661aabe556f3cd3a8bf3c58c8d398bbd1662ad61984d5b87835e55c5.jpg) + +tial dynamic convolution to increase the adaptability. There were also works like DDQ [63] which combines the dense and sparse queries to guarantee high recall. In this paper, we improve query-based object detector from a new perspective of the scalability [64, 44, 52] of the decoder, i.e., we reuse dynamic operators among decoder layers to capture more diverse and complex representations. + +Training Procedure. In vanilla DETR, a set prediction loss [4] is adopted for training. Recently, many papers have analyzed how to accelerate the convergence of DETR via improved training procedures. To verify whether the instability of Hungarian loss slow down the convergence, Sun et al. [47] utilized a matching distillation with a pre-trained DETR providing label assignment to train another model, and they found this instability only influenced the convergence in the early few epochs. DN-DETR [28] presented a denoising training strategy where a set of additional noised ground-truth objects were passed through the decoder to reconstruct the corresponding raw objects. DINO-DETR [60] improved DN-DETR via contrastive denoising training for hard negative samples. Group DETR [6] introduced multiple groups of object queries for the global one-to-many label assignment during training, but maintained one-to-one label assignment in each group, and thus Group DETR could achieve the ability of duplicate removal with one group of queries. Hybrid Matching [24] was also proposed to combine one-to-one and one-to-many label assignment into one query-based object detector with a large number of queries, and it had three types of implementations: hybrid branch scheme (one-to-many matching for one group of queries, one-to-one matching for another), hybrid epoch scheme (one-to-many matching in early epochs, one-to-one matching in late epochs) and hybrid layer scheme (one-to-many matching in early layers, one-to-one in late layers). Different from the previous methods, in this paper, we focus on the calibration of label assignment without adding additional queries. We collect training target labels across stages by a cross-stage label assigner, and then select proper targets to act as the supervision of each object prediction. + +# 3. Proposed Approach + +In this paper, we focus on the cross-stage interaction in query-based object detectors because it can well mitigate the misalignment between the decoders and supervisions in an object detector. We first revisit the state-of-the-art query-based object detectors, especially AdaMixer [15], and then elaborate on our proposed cross-stage interaction. + +# 3.1. Preliminary on query-based object detectors + +Generally, the architecture of query-based object detectors is composed by four parts: object queries, a backbone, a series of encoders and decoders. Distinctively, Adamixer removes the encoders and still maintains the desired performance. As shown in Fig. 4, it consists of object queries, a backbone and a series of decoders. + +Object Query. The initial object queries is just a set of learnable embeddings. Recent object detectors [32, 46, 15] decompose them into content vectors and positional vectors. The content vector is a vector $\mathbf{v} \in \mathbb{R}^D$ . The positional vector is presented in the format of box coordinates. For example, in AdaMixer, it contains the information about the center point, scale and aspect ratio of an individual box. + +Decoder Layer. In a query-based detector, the decoder layers are stacked multiple times to form a cascade structure. Each decoder layer is generally composed of three components: a multi-head self-attention (MHSA), a dynamic interaction module, and feed-forward networks (FFNs). DETR-like models use the multi-head cross-attention for the dynamic interaction, while AdaMixer adopts a feature sampler and a dynamic mixing module, as shown in Fig. 4. The object queries are sequentially passed through these modules. Specifically, the queries are first processed by the multi-head self-attention module. Then, its outputs, the updated content vectors, together with the positional vectors are fed into the feature sampler. In this sampler, each query is allocated with a unique group of regional multi-scale image features, i.e., sampled features, by using the queries and bilinear interpolation. Subsequently, + +![](images/d80e662046a197fe209c3993ebdd5e61a89ffc487ec784cc4050b1ea9a882bc5.jpg) +Figure 4: Overview of AdaMixer. + +the adaptive mixing is performed on the sampled features with dynamic filters generated from the content vectors. Its outputs are aggregated into the vectors. Last, by means of FFNs, the queries updated by these modules can be decoded into object predictions, i.e., the relative scaling and offsets to the positional vectors (bounding boxes), and the classification score vectors. These updated queries also serve as inputs of the next stage. Note that any query has and only has one corresponding prediction in every decoder layer. Thus, we simply represent the predictions derived from the same initial query with one index, i.e., the query index. + +Training and testing. In each decoder layer of a query-based detector, a bipartite matching is directly adopted on the ground-truth objects and the predictions. After some predictions match the ground-truth, these predictions are deemed as positive samples (assigned with foreground classification labels and supervision for localization) while others are the negative (assigned with the background label). Focal loss [30] serves as the classification loss and GIoU loss [42] with $\ell_1$ loss acts as the localization loss. During inference, only the outputs with high confidence from the last layer are used for evaluation. + +# 3.2. Cross-stage Label Assignment + +To mitigate the training difficulty caused by bipartite matching in query-based object detectors, we propose a new cross-stage label assigner to modify the results of the preceding bipartite matching. As depicted in Fig. 3, our assigner first gathers these assignments across the stages (i.e., decoder layers), and then selects appropriate training target class labels for each prediction. + +When gathering training targets, the cross-stage label assigner forces each prediction to only access the targets of predictions sharing the same query index (defined in Sec. 3.1) with it. The motivation behind this constraint is that the supervision of a single query may vary across stages, even though its predicted boxes across stages continuously cover the same ground-truth object, shown in Fig. 2a. To alleviate this inconsistency for each query, we leverage its assigned targets provided by bipartite matching from multiple stages. When introducing targets, we use a + +![](images/7a8d37b6df55ca5408e976e4f83102b0599ad18c3e360c7b81f66f6c12f14198.jpg) +(a) Given a stage $i$ , we enumerate other decoder layers whose indexes are denoted as $j$ , and select their assigned targets as the supervision for the stage $i$ according to Eq. (1). + +Figure 5: The process of our cross-stage label assignment. +![](images/d523c3f0e66355bc89928b2248a72ce66850c60fa638d93147a77ca870ca0e60.jpg) +(b) The elements in each candidate set are formed into the final supervision. + +score to determine whether a match between a target and a prediction is suitable to be shared across stages. + +Algorithm. Thanks to Focal loss [30] with binary cross entropy loss in prevailing query-based object detectors [15, 46, 28, 60], the multi-class classification task can be viewed as the binary classification on each category, and there is no need to consider a separate background category when training. More importantly, our cross-stage label assignment can be conducted on each category independently. The specific algorithm is as follows: + +Our cross-stage label assignment is performed between two stages with one type of condition. First, as shown in Fig. 5a, given a stage $i$ and its predictions $\{p_1^i,\dots,p_N^i\}$ , we traverse other stages with indexes in the range $[\alpha_i,\beta_i]$ . Taking the stage $j$ as an example, we obtain its predictions $\{p_1^j,\dots,p_N^j\}$ . Second, we also create an empty candidate bag for each prediction of the stage $i$ . We perform bipartite matching between the predictions of the stage $j$ and the ground-truth objects. Since this matching is an one-to-one + +
MethodAPAP50AP75
Baseline43.061.546.1
Hybrid Layer†41.8 (-1.2)60.5 (-1.0)44.7 (-1.4)
Ours44.8 (+1.8)63.0 (+1.5)48.4 (+2.3)
+ +Table 1: We reproduce $(\dagger)$ the hybrid matching scheme [24], and compare it with our cross-stage label assigner with 100 queries. + +label assignment, each ground-truth object corresponds to a prediction with a specific query index at the stage $j$ . Third, we align the predictions from these two stages according to the query indexes, and thus each ground-truth object also corresponds to a candidate bag. We then select the training targets from the stage $j$ as the candidate supervisions according to the condition: + +$$ +\vartheta_ {i} (q, t) \geq \eta_ {i}, \tag {1} +$$ + +where $\vartheta_{i}(q,t)$ denotes a score between the query $q$ and the ground-truth object $t$ at the stage $i$ . Here, the object $t$ needs to be assigned to query $q$ at stage $\Phi_j$ by the bipartite matching. $\eta_{i}$ denotes a threshold. In practice, we follow the classical setting [41], where the IoU as the score and the threshold is set as 0.5. Thus, if the IoU is satisfied, the ground-truth objects are added to the corresponding candidate bags. Finally, after traversing all the stages in the range $[\alpha_i,\beta_i]$ , we use the updated candidate bags to provide supervision for the stage $i$ . As shown in Figure 5b, we convert the labels in each candidate bag into one-hot vectors, and then merge them into one vector. The resulting vector serves as classification supervision through the binary cross entropy loss (Focal loss [30] in practice). + +Discussion. Although our label assignment seems to have something in common with existing works like TSP-RCNN [47] and Hybrid Layer Matching [24], the discrepancy between their design and ours cannot be ignored: (1) in TSP-RCNN, first, the idea from Faster-R-CNN [41] is directly borrowed into the set prediction problem (i.e., a ground-truth object is only assigned to the proposal that shares an IoU score greater than 0.5 with it). TSP-RCNN adopts such strategy for both the classification and localization. Differently, we can apply it only for the classification task, shown as Tab. 12. Second, TSP-RCNN adopts such strategy with dense proposals for both the classification and localization. Differently, we apply it with sparse queries. (2) for Hybrid Layer Matching, ground-truth objects are simply replicated in the first four stages, and then a bipartite matching is conducted. Differently, we do not modify the set of ground-truth objects. We gather results of the bipartite matching across stages, and then only select some targets as supervision. Also, we observe that Hybrid Layer Matching is incompatible with the models given few + +![](images/9ceda2cabc2cba71defc0e8528e54bdb9e9d2cf5b760567e72d3ee5db7a40327.jpg) +Figure 6: The dynamic mixing with filter reusing on sampled features. The content vectors dynamically generate filters through linear layers. These filters are used for mixing on sampled features, and then are stored into a dynamic filter bank for subsequent stages. Previous filters stored in the bank are also reused for mixing. + +queries. In Tab. 1, we implement it with the basic setting of our detector with 100 queries. The results show that it is totally infeasible under such a circumstance. In a nutshell, our approach is greatly different from these methods. + +# 3.3. Cross-stage Dynamic Filter Reuse + +The model capacity is essential for neural networks to capture complex representations [44, 64]. Since one of the major characteristics shared by query-based detectors is a series of decoder layers, we resort to modifying the decoder to improve this capacity. + +A straightforward way of this modification is adding the attention-like operators. The prevailing DETR-like detectors [32, 28, 60] perform deformable cross-attention [68] once along the spatial dimension in each stage, so directly adding this attention is feasible. By contrast, other detectors [46, 15] like AdaMixer perform more than one dynamic mixing at each stage, and they require a huge number of parameters [53] to generate the corresponding dynamic filters. Taking the channel mixing as an example, in each decoder layer, the specific process of generating a filter is as follows: + +$$ +\mathbf {M} _ {i} = \mathbf {W} _ {0} ^ {(i)} + \sum_ {d = 1} ^ {D} \mathbf {W} _ {d} ^ {(i)} \mathbf {v} _ {i, q, d} \in \mathbb {R} ^ {D _ {C} \times D _ {C}}, \tag {2} +$$ + +where $\mathbf{v}_{i,q,d}$ denotes the $d$ -th element of the content vector of query $q$ at the stage $i$ , $\mathbf{W}_d^{(i)} \in \mathbb{R}^{D_C \times D_C}$ denotes a learnable weight matrix, and the number of such matrices in a stage is $D$ , $D_C$ denotes the channel dimension of the input features, and $\mathbf{M}_i$ serves as the kernel for channel mixing ( $1 \times 1$ convolution) [50]. Only the parameters used to generate a single filter have already been more than $D \times D_C^2$ . Thereby, it is impractical to directly stack more decoders given limited resources. Fortunately, there are a large quantity of disposable dynamic filters in these cascade decoder + +layers, so these operators have the potential to be reused with lightweight modules across stages. + +As depicted in Fig. 6, we propose a cascade dynamic mixing module with a filter bank as an extension of the original channel mixing. The filters generated in each stage are stored in the filter bank for future use. Given a stage, the filter adapters update these stored filters with the ones from the current stage. In each adapter, the process of updating filters is as follows: + +$$ +\begin{array}{l} \mathbf {w} _ {j} ^ {(1)} = \sigma \left(\mathbf {W} _ {j} ^ {(1)} \mathbf {v} _ {i, q}\right) \in \mathbb {R} ^ {D _ {C}}, \mathbf {w} _ {j} ^ {(2)} = \sigma \left(\mathbf {W} _ {j} ^ {(2)} \mathbf {v} _ {i, q}\right) \in \mathbb {R} ^ {D _ {C}}, \\ \mathbf {M} _ {j, i} ^ {\prime} = \left(\mathbf {w} _ {j} ^ {(1)} \cdot \mathbf {w} _ {j} ^ {(2) ^ {\top}}\right) \odot \mathbf {M} _ {j} + \left(\mathbf {1} - \mathbf {w} _ {j} ^ {(1)} \cdot \mathbf {w} _ {j} ^ {(2) ^ {\top}}\right) \odot \mathbf {M} _ {i}, \tag {3} \\ \end{array} +$$ + +where 1 denotes an all-ones matrix, $\mathbf{W}_j^{(1)},\mathbf{W}_j^{(2)}$ denote learnable weights, $\sigma (\cdot)$ denotes the sigmoid function, $\odot$ denotes the element-wise product, $\mathbf{M}_j$ denotes a previous filter, $j\in [\gamma_i,i)$ $\gamma_{i}$ is a lower bound, $\mathbf{M}_{j,i}^{\prime}$ is used for subsequent modules, and the number of these updated filters is $\Delta \gamma_{i} = i - \gamma_{i}$ .We empirically find more filters can lead to more performance gain, so $\Delta \gamma_{i} = i - 1$ is the best choice. Note that this adapter only costs $2D_C^2$ parameters, which is more lightweight than the process of generating a filter from scratch as Eq. (2). Then, the updated filters are used in the cascade dynamic mixing. In this structure, since each dynamic mixing is followed by a layer normalization with an activation function, we insert a lightweight linear layer between every two dynamic mixing, consistent with [21]. + +Moreover, we can modify the original spatial mixing with the dynamic filter reusing. Unlike the channel mixing, the original dynamic spatial mixing costs a lot of computational resources on its output features due to its expansion [43] on spatial dimension. To tackle this problem, we also adopt the bank to store the filters generated for spatial mixing. Then, in each decoder layer, we update these filters with adapters and concatenate them with the current filter along the output dimension. This new filter is used to perform spatial mixing only once, rather than iteratively. This approach allows us to reduce the parameters of the filter generator of the spatial mixing, i.e., we employ the filters from preceding stages to replace a proportion of the filter from the current layer. + +# 4. Experiments + +# 4.1. Implementation Details + +The experiments are performed on the MS COCO [31] object detection dataset, where the train2017 split and the val2017 split are for training and testing. All of our experiments on AdaMixer are based on mmdetection codebase [5]. The experiments on DN-DETR and DINO are based on DETREX codebase [11]. The convolutional neural networks [21, 56] or visiontrasformers [19, 34, 12] can be taken as the backbone network. 8 RTX 2080ti with 11G + +
detectorAP\( AP_{50} \)\( AP_{75} \)\( AP_s \)\( AP_m \)\( AP_l \)
FCOS [49]38.757.441.822.942.550.1
Cascade R-CNN [3]40.458.944.122.843.754.0
GFocalV2 [29]41.158.844.923.544.953.3
BorderDet [38]41.459.444.523.645.154.6
Dynamic Head [10]42.660.146.426.146.856.0
DETR [4]20.036.219.36.020.532.2
Deform-DETR [68]35.153.637.718.238.548.7
Sparse R-CNN [46]37.956.040.520.740.053.5
AdaMixer [15]42.761.545.924.745.459.2
\( AdaMixer^† \)[15]45.064.248.627.947.861.1
StageInteractor44.863.048.427.548.061.3
StageInteractor\( ^† \)46.965.251.130.049.762.3
+ +Table 2: $1 \times$ training scheme (12 epochs) performance of various detectors on COCO minival set with ResNet-50. 100 object queries is the default setting in our method. † denotes 500 queries. + +![](images/35a19d461e427d7c9ff6b05192a23a45d0956bc4fe4eb9b21b8ffd82d13d236e.jpg) +Figure 7: The instability of the assigned labels. + +GPU memory are enough to train our model with ResNet-50 and 100 queries. For larger backbones or more queries, we resort to 8 V100 with 32G. The cross-stage label assignment is applied on each decoder layer. The reuse of spatial dynamic filters do not perform on the first two decoder layers. The weighted sum of these losses are for training, and the loss weights are in line with those of AdaMixer in mmdetection codebase [5] and those of DETRs in DETREX codebase [11]. AdamW [35] is taken as the optimizer. + +# 4.2. Comparison to State-of-the-Art Detectors + +Our model significantly outperforms the previous methods, shown in Fig. 1, and it has become a new state-of-the-art query-based object detector. As shown in Tab. 2, with $1 \times$ training scheme and ResNet-50 [21] as backbone, our detector outperforms various methods on COCO minival set [31] with even with 100 queries. Our model with ResNet-50 as backbone can achieve 44.8 AP on MS COCO validation set under the basic setting of 100 object queries, with $27.5\mathrm{AP}_s$ , $48.0\mathrm{AP}_m$ and $61.3\mathrm{AP}_l$ on small, medium and large object detection, respectively. When + +
DetectorBackboneEncoder/FPNEpochsAP\( AP_{50} \)\( AP_{75} \)\( AP_s \)\( AP_m \)\( AP_l \)
DETR [4]ResNet-50-DC5TransformerEnc50043.363.145.922.547.361.1
SMCA [14]ResNet-50TransformerEnc5043.763.647.224.247.060.4
Deformable DETR [68]ResNet-50DeformTransEnc5043.862.647.726.447.158.0
Anchor DETR [54]ResNet-50-DC5DecoupTransEnc5044.264.747.524.748.260.6
Efficient DETR [58]ResNet-50DeformTransEnc3645.163.149.128.348.459.0
Conditional DETR [36]ResNet-50-DC5TransformerEnc10845.165.448.525.349.062.2
Sparse R-CNN [46]ResNet-50FPN3645.063.448.226.947.259.5
REGO [8]ResNet-50DeformTransEnc5047.666.851.629.650.662.3
DAB-D-DETR [32]ResNet-50DeformTransEnc5046.866.050.429.149.862.3
DN-DAB-D-DETR [28]ResNet-50DeformTransEnc1243.461.947.224.846.859.4
DN-DAB-D-DETR [28]ResNet-50DeformTransEnc5048.667.452.731.052.063.7
AdaMixer [15]ResNet-50-1244.163.147.829.547.058.8
AdaMixer [15]ResNet-50-2446.765.950.529.749.761.5
AdaMixer [15]ResNet-50-3647.066.051.130.150.261.8
StageInteractorResNet-50-1246.364.350.629.849.660.8
StageInteractorResNet-50-2448.366.652.931.751.463.3
StageInteractorResNet-50-3648.967.453.431.751.864.3
StageInteractor*ResNet-50-3650.866.855.934.054.666.2
DETR [4]ResNet-101-DC5TransformerEnc50044.964.747.723.749.562.3
SMCA [14]ResNet-101TransformerEnc5044.465.248.024.348.561.0
Efficient DETR [58]ResNet-101DeformTransEnc3645.764.149.528.249.160.2
Conditional DETR [36]ResNet-101-DC5TransformerEnc10845.966.849.527.250.363.3
Sparse R-CNN [46]ResNet-101FPN3646.464.649.528.348.361.6
REGO [8]ResNet-101DeformTransEnc5048.567.052.429.552.064.4
AdaMixer [15]ResNet-101-3648.067.052.430.051.263.7
StageInteractorResNet-101-3649.968.654.633.053.665.4
REGO [8]ResNeXt-101DeformTransEnc5049.167.553.130.052.665.0
AdaMixer [15]ResNeXt-101-DCN-3649.568.953.931.352.366.3
StageInteractorResNeXt-101-DCN-3651.370.256.033.254.666.9
AdaMixer [15]Swin-S-3651.371.255.734.254.667.3
StageInteractorSwin-S-3652.771.757.736.156.267.7
+ +Table 3: The performance of various query-based detectors on MS COCO minival set with longer training scheme and single scale testing. The number of queries defaults to 300 in our method. * denotes 900 queries with more sampling points. + +equipped with 500 queries, our detector performs better, and it can achieve 46.9 AP. As depicted in Tab. 3, equipped with $3 \times$ training time, 300 queries and more data augmentation in line with other query-based object detectors, our model can achieve 48.9 AP, 49.9 AP, 51.3 AP, and 52.7 AP with ResNet-50, ResNet-101, ResNeXt-101-DCN [55, 67], and Swin-S [34] as backbones, under the setting of single scale and single model testing. Moreover, if we extend the quantity of queries into 900 and use tricks (i.e., adding more sampling points at each stage), our model can achieve 50.8 AP with ResNet-50. + +More importantly, our designs can be applied on DETR-like detectors such as DN-DETR [28] and DINO [60]. Unlike the cross-stage label assignment, our cross-stage filter reuse is tailored for AdaMixer [15]. To achieve larger model + +capacity, we opt to simply double the cross-attention because this attention-like operation is more lightweight than the dynamic mixing. In Tab. 4, our designs yield more than $+0.5\mathrm{AP}$ for all detectors. To verify the effectiveness of our cross-stage label assignment on the larger backbones, we use DINO with Swin-B [34] for experiments. As shown in Tab. 5, our cross-stage label assignment can still yield $+0.4\mathrm{AP}$ on DINO with Swin-B as backbone with 12 epochs. + +To find the reason for the lower performance gain on DE-TRs than that on AdaMixer, we present the instability of assigned labels (i.e., the probability of a ground-truth object transferring from one query to another across stages) as curves in Fig. 7. This figure shows that Deformable-DETR is more stable, and we attribute this to its powerful transformer encoder which provides high-quality features for the + +
DetectorEpochAP\( AP_{50} \)\( AP_{75} \)
DINO [60]1249.266.953.7
DINO + CSLA1249.767.054.1
DINO + CSLA + DCA1250.067.454.9
DINO [60]2450.668.555.5
DINO + CSLA2451.068.755.6
DINO + CSLA + DCA2451.369.256.1
Deform-DETR [68]5046.164.949.9
Deform-DETR + CSLA5046.865.050.9
DN-DETR [28]5044.765.347.6
DN-DETR + CSLA5045.565.349.0
H-DETR [24]1248.666.353.2
H-DETR + CSLA1249.266.353.7
+ +Table 4: The performance of DINO [60] on MS COCOMinival set with ResNet-50. CSLA: cross-stage label assignment. DCA: dual cross-attention. + +
DetectorBackboneAPAP50AP75
DINO [60]Swin-B55.874.460.7
DINO [60] + CSLASwin-B56.274.761.3
+ +decoder. + +# 4.3. Ablation Studies + +Because the computational resource is limited, ResNet-50 is employed as our backbone network, the number of queries is set as 100, and $1 \times$ training scheme is used for our ablation studies. Due to the simplicity, we adopt AdaMixer as the baseline for ablation studies. + +The effectiveness of our proposed modules. In Tab. 6, we conduct experiments to verify the effectiveness of our proposed modules. The results show that our modules lead to $+2.2$ AP gain, and each of them boosts the performance. + +The components of our label assigner. We verify how the components of our cross-stage label assigner influences the performance in Tab. 7. In the first two lines, the results show that directly adding more supervisions based on the query index only brings marginal gains. According to the last two lines, we find that while using scores like IoU can boost the performance, using both IoU and the cross-stage labels leads to better results. + +The number of reused spatial filters. We explore the effectiveness of reusing spatial filters and the results are in Tab. 8. We find a certain number of reused filters can bring a slight performance gain. This may result from the easier optimization brought by the fewer model parameters. + +Table 5: Cross-stage label assignment on DINO [60] with Swin-B [34] as backbone. + +
CSLAReuseAP\( AP_{50} \)\( AP_{75} \)\( AP_s \)\( AP_m \)\( AP_l \)
42.661.445.724.445.758.2
✓43.061.546.125.045.759.1
✓44.162.347.625.547.560.3
✓✓44.863.048.427.548.061.3
+ +Table 6: The effectiveness of our proposed modules. CSLA: cross-stage label assignment. + +Inference speed and training memory use. As depicted in Tab. 9, with one TITAN XP and one batch size, the speed of our model is $11.6\mathrm{img / s}$ while that of the baseline is $13.5\mathrm{img / s}$ , i.e., only $1.16\times$ slower. When we set the batch size as 2 for training, the additional operation only costs about 0.3G GPU memory (about $5\%$ ). + +The number of additional channel mixing. We conduct ablation studies on how many previous filters are required for the channel mixing, and report performance in Tab. 10. We do experiments on the $\max\{\Delta \gamma_i\}$ . This means there are at most $\max\{\Delta \gamma_i\}$ filters reused for the $i$ -th stage. We find more filters can bring more performance gain. The results also show the scalability of the decoder layers in this framework. + +Selection of filters on mixing. We explore whether the previous filters with adapters are suitable for our new channel mixing. As shown in Tab. 11, we find that using the adapters to fuse previous filters with the current ones is most suitable, which ensures both the adaptability of each stage and the diversity [64] of filters. More importantly, the existence of the additional dynamic mixing is necessary. + +Modifying the localization supervision. In our cross-stage label assigner, only classification labels are used for gathering, while the supervision for localization is unchanged. Thus, we explore its influence by consistently updating the supervision of classification and localization, i.e. selecting the ground-truth boxes that have the greatest IoU with predicted boxes across stages. The results are shown in Tab. 12, and we find that the performance under the two settings is very close, so we do not modify the localization part for simplicity. + +Application scope of the cross-stage label assigner. For the $i$ -th stage, the application scope of cross-stage label assigner is $[\alpha_i,\beta_i]$ . In this part, we explore the appropriate values of $\alpha_{i}$ and $\beta_{i}$ . As shown in Tab. 13, we find that the best setting is $[i - 1,L]$ . Yet if too many previous ground-truth labels are included, like $\alpha_{i} = 1$ , this hinders the last decoder layer to learn how to remove duplicates. Moreover, we conduct experiments to verify whether the cross-stage label assigner needs to be applied on all stages, because existing works [4, 24] demonstrate that only early stages can be applied with one-to-many label assignment. On the contrary, as shown in the last line of Tab. 13, we find that our + +
CrossScoreAPAP50AP75APsAPmAPl
✓43.061.546.125.045.759.1
43.161.946.525.345.859.7
✓44.062.147.626.047.659.3
✓✓44.863.048.427.548.061.3
+ +Table 7: The effectiveness of the components of the cross-stage label assigner. + +
NsAPAP50AP75APsAPmAPl
044.562.548.127.448.060.6
144.863.048.427.548.061.3
344.462.448.026.247.460.6
+ +Table 8: The number of reused spatial filters. + +
MethodAPSpeed (img/s)Memory (GB)
Baseline42.613.56.0
Ours44.811.66.3
+ +Table 9: The inference speed and training GPU memory use. + +
\( \max\{\Delta\gamma_i\} \)AP\( AP_{50} \)\( AP_{75} \)\( AP_s \)\( AP_m \)\( AP_l \)
044.162.347.625.547.560.3
144.262.347.625.847.560.6
244.262.347.926.148.060.4
344.662.848.326.747.560.7
444.863.048.427.548.061.3
+ +Table 10: The number of additional channel mixing. + +
FilterAPAP50AP75APsAPmAPl
Pre43.561.746.825.347.059.0
44.261.947.626.147.660.0
Cur44.562.747.926.547.760.5
Adp44.863.048.427.548.061.3
+ +cross-stage label assigner can also be applied on the last two stages and this brings performance gain. We speculate that this module helps remove some inappropriate supervision caused by vanilla bipartite matching which has a shortage of constraining IoU, and this module provides some proper supervision from the previous stage. + +The threshold of IoU. As shown in Tab. 14, we find that using a threshold of 0.5 is the best choice to select appropriate labels. + +Table 11: Usage of filters. Pre, Cur: directly using previous and current filters. Adp: using adapters. + +
Loc.APAP50AP75APsAPmAPl
44.862.848.426.748.260.7
✓44.863.048.427.548.061.3
+ +Table 12: Adding localization supervision in cross-stage label assigner. + +
Range of iαiβiAPAP50AP75APsAPmAPl
[1,L]ii44.062.147.626.047.659.3
[1,L]i-1i44.162.247.725.947.559.8
[1,L]i-1i+144.162.448.026.547.360.1
[1,L]i-1L44.863.048.427.548.061.3
[1,L]1L43.360.746.825.747.258.6
[1,L-2]i-1L44.362.447.726.147.660.4
+ +Table 13: The application scope of the cross-stage label assigner for each stage. + +
IoU thresAPAP50AP75APsAPmAPl
0.343.862.447.226.346.860.3
0.744.461.148.426.047.561.5
0.544.863.048.427.548.061.3
+ +Table 14: The threshold of IoU in cross-stage label assignment. + +# 5. Conclusion + +In this paper, we have presented a new fast-converging query-based object detector with cross-stage interaction, termed as StageInteractor, to alleviate inconsistency between the one-to-one label assignment and improve the modeling capacity. Our proposed cross-stage label assigner gathers assigned labels across stages, and then reassigns proper labels to predictions in each stage. We also accumulate dynamic filter across stages and reuse them with lightweight modules to increase the modeling capacity without introducing too many parameters. With these two unique designs, StageInteractor significantly outperforms the previous methods, and our proposed label assignment can boost the performance of various query-based object detectors. + +Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization. The authors would like to thank Ziteng Gao for helpful discussions. + +# References + +[1] Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. Character-level language modeling with deeper self-attention. In Proceedings of the AAAI conference on artificial intelligence, pages 3159–3166, 2019. 2 +[2] Navaneeth Bodla, Bharat Singh, Rama Chellappa, and Larry S Davis. Soft-nms-improving object detection with one line of code. In Proceedings of the IEEE international conference on computer vision, pages 5561-5569, 2017. 1 +[3] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: high quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 1, 6 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European conference on computer vision, pages 213-229, 2020. 1, 2, 3, 6, 7, 8 +[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6 +[6] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 3 +[7] Xiaokang Chen, Fangyun Wei, Gang Zeng, and Jingdong Wang. Conditional detr v2: Efficient detection transformer with box queries. arXiv preprint arXiv:2207.08914, 2022. 1 +[8] Zhe Chen, Jing Zhang, and Dacheng Tao. Recurrent glimpse-based decoder for detection with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5260-5269, 2022. 1, 2, 7 +[9] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16372-16382, 2021. 1 +[10] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 6 +[11] detrex contributors. detrex: An research platform for transformer-based object detection algorithms, 2022. 6 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 + +[13] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019. 1 +[14] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3621–3630, 2021. 1, 2, 7 +[15] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 1, 2, 3, 4, 5, 6, 7 +[16] Ziteng Gao, Limin Wang, and Gangshan Wu. Mutual supervision for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3641-3650, 2021. 1 +[17] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 303-312, 2021. 1 +[18] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 1 +[19] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 6 +[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 6 +[22] Yihui He, Chenchen Zhu, Jianren Wang, Marios Savvides, and Xiangyu Zhang. Bounding box regression with uncertainty for accurate object detection. In Proceedings of the iee/ cvf conference on computer vision and pattern recognition, pages 2888-2897, 2019. 1 +[23] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3588-3597, 2018. 1 +[24] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 3, 5, 8 +[25] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In European Conference on Computer Vision, pages 355-371. Springer, 2020. 1 +[26] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision, pages 734-750, 2018. 1 + +[27] Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou Zhang, and Zhuowen Tu. Deeply-supervised nets. In Artificial intelligence and statistics, pages 562-570. PMLR, 2015. 2 +[28] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 3, 4, 5, 7, 8 +[29] Xiang Li, Wenhai Wang, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss v2: Learning reliable localization quality estimation for dense object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11632-11641, 2021. 6 +[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 1, 4, 5 +[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 6 +[32] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 3, 5, 7 +[33] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pages 21-37. Springer, 2016. 1 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 6, 7, 8 +[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[36] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2, 7 +[37] Duy-Kien Nguyen, Jihong Ju, Olaf Booij, Martin R Oswald, and Cees GM Snoek. Boxer: Box-attention for 2d and 3d transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4773–4782, 2022. 1 +[38] Han Qiu, Yuchen Ma, Zeming Li, Songtao Liu, and Jian Sun. Borderdet: Border feature for dense object detection. In European Conference on Computer Vision, pages 549-564. Springer, 2020. 6 +[39] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1 + +[40] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 1 +[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 5 +[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 4 +[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6 +[44] Zhiqiang Shen, Zechun Liu, and Eric Xing. Sliced recursive transformer. arXiv preprint arXiv:2111.05297, 2021. 3, 5 +[45] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In International Conference on Machine Learning, pages 9934-9944. PMLR, 2021. 1 +[46] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 1, 2, 3, 4, 5, 6, 7 +[47] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3611-3620, 2021. 3, 5 +[48] Yao Teng and Limin Wang. Structured sparse r-cnn for direct scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19437-19446, 2022. 1 +[49] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 1, 6 +[50] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261–24272, 2021. 5 +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2 +[52] Hongyu Wang, Shuming Ma, Li Dong, Shaohan Huang, Dongdong Zhang, and Furu Wei. Deepnet: Scaling transformers to 1,000 layers. arXiv preprint arXiv:2203.00555, 2022.3 + +[53] Wen Wang, Jing Zhang, Yang Cao, Yongliang Shen, and Dacheng Tao. Towards data-efficient detection transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 88-105. Springer, 2022. 5 +[54] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 7 +[55] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500, 2017. 2, 7 +[56] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 6 +[57] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9657-9666, 2019. 1 +[58] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 7 +[59] Mohsen Zand, Ali Etemad, and Michael Greenspan. Object-box: From centers to boxes for anchor-free object detection. arXiv preprint arXiv:2207.06985, 2022. 1 +[60] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5, 7, 8 +[61] Linfeng Zhang, Xin Chen, Junbo Zhang, Runpei Dong, and Kaisheng Ma. Contrastive deep supervision. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 2 +[62] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020. 1 +[63] Shilong Zhang, Xinjiang Wang, Jiaqi Wang, Jiangmiao Pang, Chengqi Lyu, Wenwei Zhang, Ping Luo, and Kai Chen. Dense distinct query for end-to-end object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7329-7338, 2023. 3 +[64] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 3, 5, 8 +[65] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 1 +[66] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020.1 + +[67] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 2, 7 +[68] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1, 2, 5, 6, 7, 8 \ No newline at end of file diff --git a/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/images.zip b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4ccb17dcbbdeea1cf1c20e8a91d7794afc427fdd --- /dev/null +++ b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f9cda9b25b2f333bde434941d463c129760626b130dcc62e01911c10ccd3194 +size 905940 diff --git a/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/layout.json b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4bc370cb92c4fd008cb506915ff0795f662d4627 --- /dev/null +++ b/2023/StageInteractor_ Query-based Object Detector with Cross-stage Interaction/layout.json @@ -0,0 +1,8820 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 531, + 121 + ], + "type": "text", + "content": "StageInteractor: Query-based Object Detector with Cross-stage Interaction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 141, + 448, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 141, + 448, + 157 + ], + "spans": [ + { + "bbox": [ + 145, + 141, + 448, + 157 + ], + "type": "text", + "content": "Yao Teng" + }, + { + "bbox": [ + 145, + 141, + 448, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 141, + 448, + 157 + ], + "type": "text", + "content": " Haisong Liu" + }, + { + "bbox": [ + 145, + 141, + 448, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 141, + 448, + 157 + ], + "type": "text", + "content": " Sheng Guo" + }, + { + "bbox": [ + 145, + 141, + 448, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 141, + 448, + 157 + ], + "type": "text", + "content": " Limin Wang" + }, + { + "bbox": [ + 145, + 141, + 448, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,\\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 157, + 493, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 157, + 493, + 171 + ], + "spans": [ + { + "bbox": [ + 100, + 157, + 493, + 171 + ], + "type": "text", + "content": "1State Key Laboratory for Novel Software Technology, Nanjing University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 157, + 171, + 436, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 171, + 436, + 186 + ], + "spans": [ + { + "bbox": [ + 157, + 171, + 436, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 171, + 436, + 186 + ], + "type": "text", + "content": "Shanghai AI Lab, China " + }, + { + "bbox": [ + 157, + 171, + 436, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 171, + 436, + 186 + ], + "type": "text", + "content": "MYbank, Ant Group, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 239, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 574 + ], + "type": "text", + "content": "Previous object detectors make predictions based on dense grid points or numerous preset anchors. Most of these detectors are trained with one-to-many label assignment strategies. On the contrary, recent query-based object detectors are based a sparse set of learnable queries refined by a series of decoder layers. The one-to-one label assignment is independently applied on each layer for deep supervision during training. Despite the great success of query-based object detection, however, this vanilla one-to-one label assignment strategy requires the detectors to have strong fine-grained discrimination and modeling capacity. In this paper, we propose a new query-based object detector with cross-stage interaction, coined as StageInteractor. During the forward pass, we come up with an efficient way to improve this modeling ability by reusing dynamic operators with lightweight adapters. As for the label assignment, a cross-stage label assigner is designed to improve the one-to-one label assignment. With this assigner, the training target class labels are gathered across stages and then reallocated to proper predictions at each decoder layer. On MS COCO benchmark, our model improves the baseline counterpart by 2.2 AP, and achieves a 44.8 AP with ResNet-50 as backbone, 100 queries and 12 training epochs. With longer training time and 300 queries, StageInteractor achieves 51.3 AP and 52.7 AP with ResNeXt-101-DCN and Swin-S, respectively. The code and models are made available at https://github.com/MCG-NJU/StageInteractor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 128, + 611 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 287, + 693 + ], + "type": "text", + "content": "Object detection is a fundamental task in computer vision and acts as a cornerstone for many downstream tasks [48, 9]. It aims to localize and categorize all object instances in an image. Over the past few decades, dense spatial prior has been widely applied in various detectors. These detectors make predictions based on either" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 225, + 526, + 360 + ], + "blocks": [ + { + "bbox": [ + 320, + 225, + 526, + 360 + ], + "lines": [ + { + "bbox": [ + 320, + 225, + 526, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 225, + 526, + 360 + ], + "type": "image", + "image_path": "f3cff15939ecfe44854a861c2fc35019913918325c89f59b3b28ea3a9d657c2e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 369, + 547, + 405 + ], + "lines": [ + { + "bbox": [ + 305, + 369, + 547, + 405 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 547, + 405 + ], + "type": "text", + "content": "Figure 1: Convergence curves of our model and other query-based object detectors [4, 68, 46, 15] with ResNet-50 [21] on MS COCO [31] minival set." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "content": "a large quantity of pre-defined anchors covering the whole image [18, 41, 3, 33, 30, 40] or dense grid points in the feature map of this image [26, 13, 65, 49, 39, 57]. To deliver supervision signals to these detectors, most works employ the one-to-many label assignment strategy [62, 66, 25, 17, 16, 59] (i.e., the classification label and localization target of one ground-truth object could be assigned to multiple object predictions). Although this paradigm is widely used in object detection, it suffers from redundant and near-duplicate predictions due to such label assignment [45], and thus relies heavily on specific post-processing algorithms for duplicate removal [2, 23, 22]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Recently, DETR [4] and its variants [68, 46, 15, 7, 32, 14, 37, 8] open a new area of object detection. These query-based object detectors get rid of the dense prior and view object detection as a set prediction problem. Specifically, they use a sparse set of queries (i.e., learnable embeddings) to progressively capture the characteristics and location of objects with the help of a series of decoder layers. In each layer, image features are sampled and fused into the input queries via attention-like operation [51, 68] or dynamic mixing [46, 15]. Then, the transformed queries are decoded into object predictions and also serve as the inputs of next layer. As for the training of this paradigm, a kind of one-to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 223, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 223, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 223, + 714 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 61, + 702, + 223, + 714 + ], + "type": "text", + "content": " Corresponding author (lmwang@nju.edu.cn)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "6577" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 202 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 202 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 202 + ], + "type": "image", + "image_path": "d676776bfd94b7ade44a472c0e0d0623b09f118a160fcdaaf50dd962f25e16fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 211, + 287, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 287, + 331 + ], + "type": "text", + "content": "Figure 2: The results of label assignment at various stages. The green box denotes the ground-truth object Person. The red and white boxes denote object prediction derived from two different queries. Pos and Neg denote the positive sample and the negative sample, respectively. (a) The white box is assigned with the ground-truth object Person by bipartite matching at the first stage, while the red box is not. But the opposite is true for the sixth stage. (b) With our cross-stage label assigner, the red box in the first stage can be assigned with the ground-truth Person." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 342, + 286, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 286, + 413 + ], + "type": "text", + "content": "one label assignment (i.e., each ground-truth object has to be assigned with only one prediction), termed as bipartite matching, is independently adopted on each decoder layer for deep supervision [61, 1, 27]. For inference, only the high-confidence outputs from the last layer are taken for evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 286, + 675 + ], + "type": "text", + "content": "However, this one-to-one label assignment requires the detectors to have strong fine-grained discrimination and modeling capacity. On one hand, this strict bipartite matching would ask the detector to capture details to distinguish the predictions. For example, as shown in Fig. 2a, although the predicted boxes of a query (colored in red) cover most of the ground-truth object (Person) at each decoder layer, this object is assigned to the boxes of different queries at different stages (Stage1, white box; Stage6, red box). In other words, only one of the predicted boxes (red or white) can become the positive sample at each stage. To distinguish these boxes, the detector needs strong fine-grained discriminability to extract high-quality features from the image. Unfortunately, this goal is hard to be fully accomplished. As a result, we are motivated to directly modify the supervision of each decoder layer, i.e., introducing additional supervision from other stages to assist the training of this layer, shown in Fig. 2b. In addition, the large modeling capacity is vital to the fine-grained discrimination. To efficiently improve this modeling ability, we resort to adding lightweight modules and reusing heavy dynamic operators in decoder layers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Based on the above analysis, in this paper, we present a new paradigm, i.e., query-based object detector with cross-stage interaction, coined as StageInteractor. This interac" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "tion of our method lies in two aspects: cross-stage label assignment and cross-stage dynamic filter reuse. Specifically, during the label assignment, a cross-stage label assigner is applied subsequent to the vanilla bipartite matching in each decoder layer. This assigner collects the results of bipartite matching across stages, and then reallocate proper training target labels for each object prediction according to the query index and a score. As for the forward pass, in each decoder layer, we reuse the heavy dynamic operators in preceding stages and add lightweight modules to increase the modeling capacity at a relatively low cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "type": "text", + "content": "Experiments show that our model with ResNet-50 [21] as backbone can achieve 44.8 AP on MS COCO validation set under the simple setting of 100 queries, with " + }, + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "type": "inline_equation", + "content": "27.5\\mathrm{AP}_s" + }, + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "type": "inline_equation", + "content": "48.0\\mathrm{AP}_m" + }, + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "type": "inline_equation", + "content": "61.3\\mathrm{AP}_l" + }, + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "type": "text", + "content": " on small, medium and large object detection, respectively. Equipped with " + }, + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 304, + 204, + 545, + 360 + ], + "type": "text", + "content": " training time, 300 queries and more data augmentation in line with other query-based detectors, our model can achieve 49.9 AP, 51.3 AP, and 52.7 AP with ResNet-101, ResNeXt-101-DCN [55, 67], and Swin-S [34] as backbones, under the setting of single scale and single model testing. Our model significantly outperforms the previous methods, as shown in Fig. 1, and it has become a new state-of-the-art query-based object detector." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "content": "DETR [4] is an end-to-end query-based object detector without hand-crafted designs such as preset anchors and non-maximum suppression (NMS), but it suffers from many problems, such as slow training convergence and unstable label assignment. To handle these challenges, a large quantity of works have been proposed. In this part, we divide these methods into two categories: modification of architectures and improvement of training procedures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Architecture. The vanilla DETR took a transformer encoder-decoder architecture [51] as the detection head, and used a set of object queries, content vectors, to encode the priors of object detection datasets. In Deformable DETR [68], the attention operator [51] was replaced with multi-scale deformable attention module, and iterative bounding box refinement and a two-stage design which enables the detector to adaptively generate queries with a dense prior were also introduced. In Sparse R-CNN [46], object query is decoupled into a content vector and an explicit bounding box, and image features are progressively aggregated into content vectors by ROAlign [20] with these boxes and dynamic convolution. Moreover, Sparse R-CNN is only a decoder architecture without any transformer encoder. There were also many works like Conditional DETR [36], DAB-DETR [32], SMCA [14] and REGO [8] studying how to introduce spatial priors for proper features to accelerate convergence. Adamixer [15] improved Sparse R-CNN via deformable operators and spa" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6578" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 330, + 152 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 330, + 152 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 330, + 152 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 330, + 152 + ], + "type": "image", + "image_path": "a9acd87135f908b3db965cc710c03ac6cd37e3535da3b6f0b99bf6bc693e5991.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 159, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 547, + 220 + ], + "type": "text", + "content": "Figure 3: Overview. The cross-stage interaction incorporates two parts: cross-stage label assignment and cross-stage dynamic filter reuse. During the forward propagation, dynamic filters in each stage of decode layer are reused in the subsequent stages, i.e., we stack them with specific lightweight adapters to increase the depth of each decoder layer. As for the label assignment, our cross-stage label assigner gathers the results of bipartite matching across stages, and then selects proper target labels as supervision." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 330, + 70, + 544, + 152 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 544, + 152 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 544, + 152 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 544, + 152 + ], + "type": "image", + "image_path": "0f85d807661aabe556f3cd3a8bf3c58c8d398bbd1662ad61984d5b87835e55c5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 288, + 326 + ], + "type": "text", + "content": "tial dynamic convolution to increase the adaptability. There were also works like DDQ [63] which combines the dense and sparse queries to guarantee high recall. In this paper, we improve query-based object detector from a new perspective of the scalability [64, 44, 52] of the decoder, i.e., we reuse dynamic operators among decoder layers to capture more diverse and complex representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 288, + 713 + ], + "type": "text", + "content": "Training Procedure. In vanilla DETR, a set prediction loss [4] is adopted for training. Recently, many papers have analyzed how to accelerate the convergence of DETR via improved training procedures. To verify whether the instability of Hungarian loss slow down the convergence, Sun et al. [47] utilized a matching distillation with a pre-trained DETR providing label assignment to train another model, and they found this instability only influenced the convergence in the early few epochs. DN-DETR [28] presented a denoising training strategy where a set of additional noised ground-truth objects were passed through the decoder to reconstruct the corresponding raw objects. DINO-DETR [60] improved DN-DETR via contrastive denoising training for hard negative samples. Group DETR [6] introduced multiple groups of object queries for the global one-to-many label assignment during training, but maintained one-to-one label assignment in each group, and thus Group DETR could achieve the ability of duplicate removal with one group of queries. Hybrid Matching [24] was also proposed to combine one-to-one and one-to-many label assignment into one query-based object detector with a large number of queries, and it had three types of implementations: hybrid branch scheme (one-to-many matching for one group of queries, one-to-one matching for another), hybrid epoch scheme (one-to-many matching in early epochs, one-to-one matching in late epochs) and hybrid layer scheme (one-to-many matching in early layers, one-to-one in late layers). Different from the previous methods, in this paper, we focus on the calibration of label assignment without adding additional queries. We collect training target labels across stages by a cross-stage label assigner, and then select proper targets to act as the supervision of each object prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 239, + 422, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 422, + 254 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 422, + 254 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 261, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 545, + 334 + ], + "type": "text", + "content": "In this paper, we focus on the cross-stage interaction in query-based object detectors because it can well mitigate the misalignment between the decoders and supervisions in an object detector. We first revisit the state-of-the-art query-based object detectors, especially AdaMixer [15], and then elaborate on our proposed cross-stage interaction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 344, + 537, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 344, + 537, + 357 + ], + "spans": [ + { + "bbox": [ + 305, + 344, + 537, + 357 + ], + "type": "text", + "content": "3.1. Preliminary on query-based object detectors" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 364, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 435 + ], + "type": "text", + "content": "Generally, the architecture of query-based object detectors is composed by four parts: object queries, a backbone, a series of encoders and decoders. Distinctively, Adamixer removes the encoders and still maintains the desired performance. As shown in Fig. 4, it consists of object queries, a backbone and a series of decoders." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "content": "Object Query. The initial object queries is just a set of learnable embeddings. Recent object detectors [32, 46, 15] decompose them into content vectors and positional vectors. The content vector is a vector " + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "content": ". The positional vector is presented in the format of box coordinates. For example, in AdaMixer, it contains the information about the center point, scale and aspect ratio of an individual box." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "Decoder Layer. In a query-based detector, the decoder layers are stacked multiple times to form a cascade structure. Each decoder layer is generally composed of three components: a multi-head self-attention (MHSA), a dynamic interaction module, and feed-forward networks (FFNs). DETR-like models use the multi-head cross-attention for the dynamic interaction, while AdaMixer adopts a feature sampler and a dynamic mixing module, as shown in Fig. 4. The object queries are sequentially passed through these modules. Specifically, the queries are first processed by the multi-head self-attention module. Then, its outputs, the updated content vectors, together with the positional vectors are fed into the feature sampler. In this sampler, each query is allocated with a unique group of regional multi-scale image features, i.e., sampled features, by using the queries and bilinear interpolation. Subsequently," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6579" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 286, + 180 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 286, + 180 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 286, + 180 + ], + "type": "image", + "image_path": "d80e662046a197fe209c3993ebdd5e61a89ffc487ec784cc4050b1ea9a882bc5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 190, + 235, + 201 + ], + "lines": [ + { + "bbox": [ + 97, + 190, + 235, + 201 + ], + "spans": [ + { + "bbox": [ + 97, + 190, + 235, + 201 + ], + "type": "text", + "content": "Figure 4: Overview of AdaMixer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 221, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 287, + 353 + ], + "type": "text", + "content": "the adaptive mixing is performed on the sampled features with dynamic filters generated from the content vectors. Its outputs are aggregated into the vectors. Last, by means of FFNs, the queries updated by these modules can be decoded into object predictions, i.e., the relative scaling and offsets to the positional vectors (bounding boxes), and the classification score vectors. These updated queries also serve as inputs of the next stage. Note that any query has and only has one corresponding prediction in every decoder layer. Thus, we simply represent the predictions derived from the same initial query with one index, i.e., the query index." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 483 + ], + "type": "text", + "content": "Training and testing. In each decoder layer of a query-based detector, a bipartite matching is directly adopted on the ground-truth objects and the predictions. After some predictions match the ground-truth, these predictions are deemed as positive samples (assigned with foreground classification labels and supervision for localization) while others are the negative (assigned with the background label). Focal loss [30] serves as the classification loss and GIoU loss [42] with " + }, + { + "bbox": [ + 46, + 354, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 354, + 287, + 483 + ], + "type": "text", + "content": " loss acts as the localization loss. During inference, only the outputs with high confidence from the last layer are used for evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 491, + 211, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 211, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 211, + 504 + ], + "type": "text", + "content": "3.2. Cross-stage Label Assignment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 593 + ], + "type": "text", + "content": "To mitigate the training difficulty caused by bipartite matching in query-based object detectors, we propose a new cross-stage label assigner to modify the results of the preceding bipartite matching. As depicted in Fig. 3, our assigner first gathers these assignments across the stages (i.e., decoder layers), and then selects appropriate training target class labels for each prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "When gathering training targets, the cross-stage label assigner forces each prediction to only access the targets of predictions sharing the same query index (defined in Sec. 3.1) with it. The motivation behind this constraint is that the supervision of a single query may vary across stages, even though its predicted boxes across stages continuously cover the same ground-truth object, shown in Fig. 2a. To alleviate this inconsistency for each query, we leverage its assigned targets provided by bipartite matching from multiple stages. When introducing targets, we use a" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 539, + 265 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 539, + 265 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 539, + 265 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 539, + 265 + ], + "type": "image", + "image_path": "7a8d37b6df55ca5408e976e4f83102b0599ad18c3e360c7b81f66f6c12f14198.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 270, + 544, + 304 + ], + "lines": [ + { + "bbox": [ + 306, + 270, + 544, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 544, + 304 + ], + "type": "text", + "content": "(a) Given a stage " + }, + { + "bbox": [ + 306, + 270, + 544, + 304 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 270, + 544, + 304 + ], + "type": "text", + "content": ", we enumerate other decoder layers whose indexes are denoted as " + }, + { + "bbox": [ + 306, + 270, + 544, + 304 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 306, + 270, + 544, + 304 + ], + "type": "text", + "content": ", and select their assigned targets as the supervision for the stage " + }, + { + "bbox": [ + 306, + 270, + 544, + 304 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 270, + 544, + 304 + ], + "type": "text", + "content": " according to Eq. (1)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 365, + 314, + 482, + 419 + ], + "blocks": [ + { + "bbox": [ + 365, + 314, + 482, + 419 + ], + "lines": [ + { + "bbox": [ + 365, + 314, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 365, + 314, + 482, + 419 + ], + "type": "image", + "image_path": "d523c3f0e66355bc89928b2248a72ce66850c60fa638d93147a77ca870ca0e60.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 423, + 543, + 445 + ], + "lines": [ + { + "bbox": [ + 306, + 423, + 543, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 543, + 445 + ], + "type": "text", + "content": "(b) The elements in each candidate set are formed into the final supervision." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 307, + 450, + 543, + 463 + ], + "lines": [ + { + "bbox": [ + 307, + 450, + 543, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 450, + 543, + 463 + ], + "type": "text", + "content": "Figure 5: The process of our cross-stage label assignment." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 485, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 485, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 545, + 509 + ], + "type": "text", + "content": "score to determine whether a match between a target and a prediction is suitable to be shared across stages." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": "Algorithm. Thanks to Focal loss [30] with binary cross entropy loss in prevailing query-based object detectors [15, 46, 28, 60], the multi-class classification task can be viewed as the binary classification on each category, and there is no need to consider a separate background category when training. More importantly, our cross-stage label assignment can be conducted on each category independently. The specific algorithm is as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Our cross-stage label assignment is performed between two stages with one type of condition. First, as shown in Fig. 5a, given a stage " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and its predictions " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{p_1^i,\\dots,p_N^i\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", we traverse other stages with indexes in the range " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "[\\alpha_i,\\beta_i]" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". Taking the stage " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " as an example, we obtain its predictions " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{p_1^j,\\dots,p_N^j\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". Second, we also create an empty candidate bag for each prediction of the stage " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". We perform bipartite matching between the predictions of the stage " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and the ground-truth objects. Since this matching is an one-to-one" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6580" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 284, + 131 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 284, + 131 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 284, + 131 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 284, + 131 + ], + "type": "table", + "html": "
MethodAPAP50AP75
Baseline43.061.546.1
Hybrid Layer†41.8 (-1.2)60.5 (-1.0)44.7 (-1.4)
Ours44.8 (+1.8)63.0 (+1.5)48.4 (+2.3)
", + "image_path": "ab7ce1873e091d395012a653a5fc2a0423eddca5dee83f1a890bf1327c304811.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 134, + 287, + 171 + ], + "lines": [ + { + "bbox": [ + 47, + 134, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 287, + 171 + ], + "type": "text", + "content": "Table 1: We reproduce " + }, + { + "bbox": [ + 47, + 134, + 287, + 171 + ], + "type": "inline_equation", + "content": "(\\dagger)" + }, + { + "bbox": [ + 47, + 134, + 287, + 171 + ], + "type": "text", + "content": " the hybrid matching scheme [24], and compare it with our cross-stage label assigner with 100 queries." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "content": "label assignment, each ground-truth object corresponds to a prediction with a specific query index at the stage " + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "content": ". Third, we align the predictions from these two stages according to the query indexes, and thus each ground-truth object also corresponds to a candidate bag. We then select the training targets from the stage " + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 191, + 287, + 275 + ], + "type": "text", + "content": " as the candidate supervisions according to the condition:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 284, + 287, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 284, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 137, + 284, + 287, + 298 + ], + "type": "interline_equation", + "content": "\\vartheta_ {i} (q, t) \\geq \\eta_ {i}, \\tag {1}", + "image_path": "5ceab5dc529735fe2dee890cc5118fc60cd01ed4a014ffe2bb72ba92a760b75e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\vartheta_{i}(q,t)" + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": " denotes a score between the query " + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": " and the ground-truth object " + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": " at the stage " + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": ". Here, the object " + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": " needs to be assigned to query " + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": " at stage " + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\Phi_j" + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": " by the bipartite matching. " + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\eta_{i}" + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": " denotes a threshold. In practice, we follow the classical setting [41], where the IoU as the score and the threshold is set as 0.5. Thus, if the IoU is satisfied, the ground-truth objects are added to the corresponding candidate bags. Finally, after traversing all the stages in the range " + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "inline_equation", + "content": "[\\alpha_i,\\beta_i]" + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": ", we use the updated candidate bags to provide supervision for the stage " + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 306, + 287, + 474 + ], + "type": "text", + "content": ". As shown in Figure 5b, we convert the labels in each candidate bag into one-hot vectors, and then merge them into one vector. The resulting vector serves as classification supervision through the binary cross entropy loss (Focal loss [30] in practice)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": "Discussion. Although our label assignment seems to have something in common with existing works like TSP-RCNN [47] and Hybrid Layer Matching [24], the discrepancy between their design and ours cannot be ignored: (1) in TSP-RCNN, first, the idea from Faster-R-CNN [41] is directly borrowed into the set prediction problem (i.e., a ground-truth object is only assigned to the proposal that shares an IoU score greater than 0.5 with it). TSP-RCNN adopts such strategy for both the classification and localization. Differently, we can apply it only for the classification task, shown as Tab. 12. Second, TSP-RCNN adopts such strategy with dense proposals for both the classification and localization. Differently, we apply it with sparse queries. (2) for Hybrid Layer Matching, ground-truth objects are simply replicated in the first four stages, and then a bipartite matching is conducted. Differently, we do not modify the set of ground-truth objects. We gather results of the bipartite matching across stages, and then only select some targets as supervision. Also, we observe that Hybrid Layer Matching is incompatible with the models given few" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 545, + 198 + ], + "type": "image", + "image_path": "9ceda2cabc2cba71defc0e8528e54bdb9e9d2cf5b760567e72d3ee5db7a40327.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 205, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 277 + ], + "type": "text", + "content": "Figure 6: The dynamic mixing with filter reusing on sampled features. The content vectors dynamically generate filters through linear layers. These filters are used for mixing on sampled features, and then are stored into a dynamic filter bank for subsequent stages. Previous filters stored in the bank are also reused for mixing." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "type": "text", + "content": "queries. In Tab. 1, we implement it with the basic setting of our detector with 100 queries. The results show that it is totally infeasible under such a circumstance. In a nutshell, our approach is greatly different from these methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 350, + 487, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 487, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 487, + 363 + ], + "type": "text", + "content": "3.3. Cross-stage Dynamic Filter Reuse" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 368, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 545, + 429 + ], + "type": "text", + "content": "The model capacity is essential for neural networks to capture complex representations [44, 64]. Since one of the major characteristics shared by query-based detectors is a series of decoder layers, we resort to modifying the decoder to improve this capacity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 548 + ], + "type": "text", + "content": "A straightforward way of this modification is adding the attention-like operators. The prevailing DETR-like detectors [32, 28, 60] perform deformable cross-attention [68] once along the spatial dimension in each stage, so directly adding this attention is feasible. By contrast, other detectors [46, 15] like AdaMixer perform more than one dynamic mixing at each stage, and they require a huge number of parameters [53] to generate the corresponding dynamic filters. Taking the channel mixing as an example, in each decoder layer, the specific process of generating a filter is as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 336, + 554, + 545, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 554, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 336, + 554, + 545, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i} = \\mathbf {W} _ {0} ^ {(i)} + \\sum_ {d = 1} ^ {D} \\mathbf {W} _ {d} ^ {(i)} \\mathbf {v} _ {i, q, d} \\in \\mathbb {R} ^ {D _ {C} \\times D _ {C}}, \\tag {2}", + "image_path": "24e4ec6f8300be3bb96cc6f1751b6f035802482e009df4ba7b9b734b441edb11.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{i,q,d}" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": "-th element of the content vector of query " + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": " at the stage " + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_d^{(i)} \\in \\mathbb{R}^{D_C \\times D_C}" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": " denotes a learnable weight matrix, and the number of such matrices in a stage is " + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "inline_equation", + "content": "D_C" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": " denotes the channel dimension of the input features, and " + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": " serves as the kernel for channel mixing (" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": " convolution) [50]. Only the parameters used to generate a single filter have already been more than " + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "inline_equation", + "content": "D \\times D_C^2" + }, + { + "bbox": [ + 304, + 591, + 546, + 713 + ], + "type": "text", + "content": ". Thereby, it is impractical to directly stack more decoders given limited resources. Fortunately, there are a large quantity of disposable dynamic filters in these cascade decoder" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6581" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "layers, so these operators have the potential to be reused with lightweight modules across stages." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 97, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 97, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 45, + 97, + 286, + 180 + ], + "type": "text", + "content": "As depicted in Fig. 6, we propose a cascade dynamic mixing module with a filter bank as an extension of the original channel mixing. The filters generated in each stage are stored in the filter bank for future use. Given a stage, the filter adapters update these stored filters with the ones from the current stage. In each adapter, the process of updating filters is as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 186, + 286, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 186, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 51, + 186, + 286, + 228 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {w} _ {j} ^ {(1)} = \\sigma \\left(\\mathbf {W} _ {j} ^ {(1)} \\mathbf {v} _ {i, q}\\right) \\in \\mathbb {R} ^ {D _ {C}}, \\mathbf {w} _ {j} ^ {(2)} = \\sigma \\left(\\mathbf {W} _ {j} ^ {(2)} \\mathbf {v} _ {i, q}\\right) \\in \\mathbb {R} ^ {D _ {C}}, \\\\ \\mathbf {M} _ {j, i} ^ {\\prime} = \\left(\\mathbf {w} _ {j} ^ {(1)} \\cdot \\mathbf {w} _ {j} ^ {(2) ^ {\\top}}\\right) \\odot \\mathbf {M} _ {j} + \\left(\\mathbf {1} - \\mathbf {w} _ {j} ^ {(1)} \\cdot \\mathbf {w} _ {j} ^ {(2) ^ {\\top}}\\right) \\odot \\mathbf {M} _ {i}, \\tag {3} \\\\ \\end{array}", + "image_path": "4f402f930f9fb998f088cf6487dffef07658ad4a47e677c520db35fa0d2d2617.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "content": "where 1 denotes an all-ones matrix, " + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_j^{(1)},\\mathbf{W}_j^{(2)}" + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "content": " denote learnable weights, " + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "content": " denotes the sigmoid function, " + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "content": " denotes the element-wise product, " + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_j" + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "content": " denotes a previous filter, " + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "inline_equation", + "content": "j\\in [\\gamma_i,i)" + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "inline_equation", + "content": "\\gamma_{i}" + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "content": " is a lower bound, " + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{j,i}^{\\prime}" + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "content": " is used for subsequent modules, and the number of these updated filters is " + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "inline_equation", + "content": "\\Delta \\gamma_{i} = i - \\gamma_{i}" + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "content": " .We empirically find more filters can lead to more performance gain, so " + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "inline_equation", + "content": "\\Delta \\gamma_{i} = i - 1" + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "content": " is the best choice. Note that this adapter only costs " + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "inline_equation", + "content": "2D_C^2" + }, + { + "bbox": [ + 46, + 228, + 286, + 397 + ], + "type": "text", + "content": " parameters, which is more lightweight than the process of generating a filter from scratch as Eq. (2). Then, the updated filters are used in the cascade dynamic mixing. In this structure, since each dynamic mixing is followed by a layer normalization with an activation function, we insert a lightweight linear layer between every two dynamic mixing, consistent with [21]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 399, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 286, + 567 + ], + "type": "text", + "content": "Moreover, we can modify the original spatial mixing with the dynamic filter reusing. Unlike the channel mixing, the original dynamic spatial mixing costs a lot of computational resources on its output features due to its expansion [43] on spatial dimension. To tackle this problem, we also adopt the bank to store the filters generated for spatial mixing. Then, in each decoder layer, we update these filters with adapters and concatenate them with the current filter along the output dimension. This new filter is used to perform spatial mixing only once, rather than iteratively. This approach allows us to reduce the parameters of the filter generator of the spatial mixing, i.e., we employ the filters from preceding stages to replace a proportion of the filter from the current layer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 578, + 128, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 592 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 592 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 598, + 180, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 180, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 180, + 611 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "The experiments are performed on the MS COCO [31] object detection dataset, where the train2017 split and the val2017 split are for training and testing. All of our experiments on AdaMixer are based on mmdetection codebase [5]. The experiments on DN-DETR and DINO are based on DETREX codebase [11]. The convolutional neural networks [21, 56] or visiontrasformers [19, 34, 12] can be taken as the backbone network. 8 RTX 2080ti with 11G" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 545, + 244 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 244 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 244 + ], + "type": "table", + "html": "
detectorAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
FCOS [49]38.757.441.822.942.550.1
Cascade R-CNN [3]40.458.944.122.843.754.0
GFocalV2 [29]41.158.844.923.544.953.3
BorderDet [38]41.459.444.523.645.154.6
Dynamic Head [10]42.660.146.426.146.856.0
DETR [4]20.036.219.36.020.532.2
Deform-DETR [68]35.153.637.718.238.548.7
Sparse R-CNN [46]37.956.040.520.740.053.5
AdaMixer [15]42.761.545.924.745.459.2
\\( AdaMixer^† \\)[15]45.064.248.627.947.861.1
StageInteractor44.863.048.427.548.061.3
StageInteractor\\( ^† \\)46.965.251.130.049.762.3
", + "image_path": "d81cfd0d17cbf01c930f3823276e9c5c86db84b7ee94258e4ce0f60f1c1747a5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 248, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 305, + 248, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 248, + 545, + 297 + ], + "type": "text", + "content": "Table 2: " + }, + { + "bbox": [ + 305, + 248, + 545, + 297 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 305, + 248, + 545, + 297 + ], + "type": "text", + "content": " training scheme (12 epochs) performance of various detectors on COCO minival set with ResNet-50. 100 object queries is the default setting in our method. † denotes 500 queries." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 308, + 303, + 542, + 425 + ], + "blocks": [ + { + "bbox": [ + 308, + 303, + 542, + 425 + ], + "lines": [ + { + "bbox": [ + 308, + 303, + 542, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 303, + 542, + 425 + ], + "type": "image", + "image_path": "35a19d461e427d7c9ff6b05192a23a45d0956bc4fe4eb9b21b8ffd82d13d236e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 429, + 520, + 441 + ], + "lines": [ + { + "bbox": [ + 329, + 429, + 520, + 441 + ], + "spans": [ + { + "bbox": [ + 329, + 429, + 520, + 441 + ], + "type": "text", + "content": "Figure 7: The instability of the assigned labels." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 457, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 567 + ], + "type": "text", + "content": "GPU memory are enough to train our model with ResNet-50 and 100 queries. For larger backbones or more queries, we resort to 8 V100 with 32G. The cross-stage label assignment is applied on each decoder layer. The reuse of spatial dynamic filters do not perform on the first two decoder layers. The weighted sum of these losses are for training, and the loss weights are in line with those of AdaMixer in mmdetection codebase [5] and those of DETRs in DETREX codebase [11]. AdamW [35] is taken as the optimizer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 574, + 523, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 523, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 523, + 587 + ], + "type": "text", + "content": "4.2. Comparison to State-of-the-Art Detectors" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Our model significantly outperforms the previous methods, shown in Fig. 1, and it has become a new state-of-the-art query-based object detector. As shown in Tab. 2, with " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " training scheme and ResNet-50 [21] as backbone, our detector outperforms various methods on COCO minival set [31] with even with 100 queries. Our model with ResNet-50 as backbone can achieve 44.8 AP on MS COCO validation set under the basic setting of 100 object queries, with " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "27.5\\mathrm{AP}_s" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "48.0\\mathrm{AP}_m" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "61.3\\mathrm{AP}_l" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " on small, medium and large object detection, respectively. When" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6582" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 69, + 545, + 481 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 545, + 481 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 545, + 481 + ], + "type": "table", + "html": "
DetectorBackboneEncoder/FPNEpochsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
DETR [4]ResNet-50-DC5TransformerEnc50043.363.145.922.547.361.1
SMCA [14]ResNet-50TransformerEnc5043.763.647.224.247.060.4
Deformable DETR [68]ResNet-50DeformTransEnc5043.862.647.726.447.158.0
Anchor DETR [54]ResNet-50-DC5DecoupTransEnc5044.264.747.524.748.260.6
Efficient DETR [58]ResNet-50DeformTransEnc3645.163.149.128.348.459.0
Conditional DETR [36]ResNet-50-DC5TransformerEnc10845.165.448.525.349.062.2
Sparse R-CNN [46]ResNet-50FPN3645.063.448.226.947.259.5
REGO [8]ResNet-50DeformTransEnc5047.666.851.629.650.662.3
DAB-D-DETR [32]ResNet-50DeformTransEnc5046.866.050.429.149.862.3
DN-DAB-D-DETR [28]ResNet-50DeformTransEnc1243.461.947.224.846.859.4
DN-DAB-D-DETR [28]ResNet-50DeformTransEnc5048.667.452.731.052.063.7
AdaMixer [15]ResNet-50-1244.163.147.829.547.058.8
AdaMixer [15]ResNet-50-2446.765.950.529.749.761.5
AdaMixer [15]ResNet-50-3647.066.051.130.150.261.8
StageInteractorResNet-50-1246.364.350.629.849.660.8
StageInteractorResNet-50-2448.366.652.931.751.463.3
StageInteractorResNet-50-3648.967.453.431.751.864.3
StageInteractor*ResNet-50-3650.866.855.934.054.666.2
DETR [4]ResNet-101-DC5TransformerEnc50044.964.747.723.749.562.3
SMCA [14]ResNet-101TransformerEnc5044.465.248.024.348.561.0
Efficient DETR [58]ResNet-101DeformTransEnc3645.764.149.528.249.160.2
Conditional DETR [36]ResNet-101-DC5TransformerEnc10845.966.849.527.250.363.3
Sparse R-CNN [46]ResNet-101FPN3646.464.649.528.348.361.6
REGO [8]ResNet-101DeformTransEnc5048.567.052.429.552.064.4
AdaMixer [15]ResNet-101-3648.067.052.430.051.263.7
StageInteractorResNet-101-3649.968.654.633.053.665.4
REGO [8]ResNeXt-101DeformTransEnc5049.167.553.130.052.665.0
AdaMixer [15]ResNeXt-101-DCN-3649.568.953.931.352.366.3
StageInteractorResNeXt-101-DCN-3651.370.256.033.254.666.9
AdaMixer [15]Swin-S-3651.371.255.734.254.667.3
StageInteractorSwin-S-3652.771.757.736.156.267.7
", + "image_path": "33fd8ea9a64bc69a8cff878af494b1892d5846efa667f506061563e636c6a080.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 483, + 546, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 483, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 546, + 510 + ], + "type": "text", + "content": "Table 3: The performance of various query-based detectors on MS COCO minival set with longer training scheme and single scale testing. The number of queries defaults to 300 in our method. * denotes 900 queries with more sampling points." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 529, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 288, + 661 + ], + "type": "text", + "content": "equipped with 500 queries, our detector performs better, and it can achieve 46.9 AP. As depicted in Tab. 3, equipped with " + }, + { + "bbox": [ + 46, + 529, + 288, + 661 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 529, + 288, + 661 + ], + "type": "text", + "content": " training time, 300 queries and more data augmentation in line with other query-based object detectors, our model can achieve 48.9 AP, 49.9 AP, 51.3 AP, and 52.7 AP with ResNet-50, ResNet-101, ResNeXt-101-DCN [55, 67], and Swin-S [34] as backbones, under the setting of single scale and single model testing. Moreover, if we extend the quantity of queries into 900 and use tricks (i.e., adding more sampling points at each stage), our model can achieve 50.8 AP with ResNet-50." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "More importantly, our designs can be applied on DETR-like detectors such as DN-DETR [28] and DINO [60]. Unlike the cross-stage label assignment, our cross-stage filter reuse is tailored for AdaMixer [15]. To achieve larger model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "text", + "content": "capacity, we opt to simply double the cross-attention because this attention-like operation is more lightweight than the dynamic mixing. In Tab. 4, our designs yield more than " + }, + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "inline_equation", + "content": "+0.5\\mathrm{AP}" + }, + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "text", + "content": " for all detectors. To verify the effectiveness of our cross-stage label assignment on the larger backbones, we use DINO with Swin-B [34] for experiments. As shown in Tab. 5, our cross-stage label assignment can still yield " + }, + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "inline_equation", + "content": "+0.4\\mathrm{AP}" + }, + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "text", + "content": " on DINO with Swin-B as backbone with 12 epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "To find the reason for the lower performance gain on DE-TRs than that on AdaMixer, we present the instability of assigned labels (i.e., the probability of a ground-truth object transferring from one query to another across stages) as curves in Fig. 7. This figure shows that Deformable-DETR is more stable, and we attribute this to its powerful transformer encoder which provides high-quality features for the" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6583" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 286, + 259 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 259 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 259 + ], + "type": "table", + "html": "
DetectorEpochAP\\( AP_{50} \\)\\( AP_{75} \\)
DINO [60]1249.266.953.7
DINO + CSLA1249.767.054.1
DINO + CSLA + DCA1250.067.454.9
DINO [60]2450.668.555.5
DINO + CSLA2451.068.755.6
DINO + CSLA + DCA2451.369.256.1
Deform-DETR [68]5046.164.949.9
Deform-DETR + CSLA5046.865.050.9
DN-DETR [28]5044.765.347.6
DN-DETR + CSLA5045.565.349.0
H-DETR [24]1248.666.353.2
H-DETR + CSLA1249.266.353.7
", + "image_path": "2c35aa06ef98b2b64e2e8447c1c67665c7b5ebbcf19274e6edb0f67dabfda48b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 310, + 285, + 360 + ], + "blocks": [ + { + "bbox": [ + 47, + 262, + 287, + 299 + ], + "lines": [ + { + "bbox": [ + 47, + 262, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 287, + 299 + ], + "type": "text", + "content": "Table 4: The performance of DINO [60] on MS COCOMinival set with ResNet-50. CSLA: cross-stage label assignment. DCA: dual cross-attention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 310, + 285, + 360 + ], + "lines": [ + { + "bbox": [ + 49, + 310, + 285, + 360 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 285, + 360 + ], + "type": "table", + "html": "
DetectorBackboneAPAP50AP75
DINO [60]Swin-B55.874.460.7
DINO [60] + CSLASwin-B56.274.761.3
", + "image_path": "f931d3a492883f8da238b7283ade05037ebe312e01fcc7469ce359d7acd2e4e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 410, + 84, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 84, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 84, + 420 + ], + "type": "text", + "content": "decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 430, + 147, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 147, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 147, + 441 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 449, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 287, + 508 + ], + "type": "text", + "content": "Because the computational resource is limited, ResNet-50 is employed as our backbone network, the number of queries is set as 100, and " + }, + { + "bbox": [ + 47, + 449, + 287, + 508 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 47, + 449, + 287, + 508 + ], + "type": "text", + "content": " training scheme is used for our ablation studies. Due to the simplicity, we adopt AdaMixer as the baseline for ablation studies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 509, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 287, + 557 + ], + "type": "text", + "content": "The effectiveness of our proposed modules. In Tab. 6, we conduct experiments to verify the effectiveness of our proposed modules. The results show that our modules lead to " + }, + { + "bbox": [ + 47, + 509, + 287, + 557 + ], + "type": "inline_equation", + "content": "+2.2" + }, + { + "bbox": [ + 47, + 509, + 287, + 557 + ], + "type": "text", + "content": " AP gain, and each of them boosts the performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 652 + ], + "type": "text", + "content": "The components of our label assigner. We verify how the components of our cross-stage label assigner influences the performance in Tab. 7. In the first two lines, the results show that directly adding more supervisions based on the query index only brings marginal gains. According to the last two lines, we find that while using scores like IoU can boost the performance, using both IoU and the cross-stage labels leads to better results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "The number of reused spatial filters. We explore the effectiveness of reusing spatial filters and the results are in Tab. 8. We find a certain number of reused filters can bring a slight performance gain. This may result from the easier optimization brought by the fewer model parameters." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 542, + 143 + ], + "blocks": [ + { + "bbox": [ + 47, + 362, + 287, + 387 + ], + "lines": [ + { + "bbox": [ + 47, + 362, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 287, + 387 + ], + "type": "text", + "content": "Table 5: Cross-stage label assignment on DINO [60] with Swin-B [34] as backbone." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 542, + 143 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 542, + 143 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 542, + 143 + ], + "type": "table", + "html": "
CSLAReuseAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
42.661.445.724.445.758.2
✓43.061.546.125.045.759.1
✓44.162.347.625.547.560.3
✓✓44.863.048.427.548.061.3
", + "image_path": "9c80a7f6324223a9d1778b858173e17c23b1dde8d715c914ac392b8beb54fdad.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 146, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 305, + 146, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 305, + 146, + 545, + 171 + ], + "type": "text", + "content": "Table 6: The effectiveness of our proposed modules. CSLA: cross-stage label assignment." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "type": "text", + "content": "Inference speed and training memory use. As depicted in Tab. 9, with one TITAN XP and one batch size, the speed of our model is " + }, + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "type": "inline_equation", + "content": "11.6\\mathrm{img / s}" + }, + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "type": "text", + "content": " while that of the baseline is " + }, + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "type": "inline_equation", + "content": "13.5\\mathrm{img / s}" + }, + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "type": "text", + "content": ", i.e., only " + }, + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "type": "inline_equation", + "content": "1.16\\times" + }, + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "type": "text", + "content": " slower. When we set the batch size as 2 for training, the additional operation only costs about 0.3G GPU memory (about " + }, + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 194, + 545, + 266 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 267, + 546, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 546, + 363 + ], + "type": "text", + "content": "The number of additional channel mixing. We conduct ablation studies on how many previous filters are required for the channel mixing, and report performance in Tab. 10. We do experiments on the " + }, + { + "bbox": [ + 304, + 267, + 546, + 363 + ], + "type": "inline_equation", + "content": "\\max\\{\\Delta \\gamma_i\\}" + }, + { + "bbox": [ + 304, + 267, + 546, + 363 + ], + "type": "text", + "content": ". This means there are at most " + }, + { + "bbox": [ + 304, + 267, + 546, + 363 + ], + "type": "inline_equation", + "content": "\\max\\{\\Delta \\gamma_i\\}" + }, + { + "bbox": [ + 304, + 267, + 546, + 363 + ], + "type": "text", + "content": " filters reused for the " + }, + { + "bbox": [ + 304, + 267, + 546, + 363 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 267, + 546, + 363 + ], + "type": "text", + "content": "-th stage. We find more filters can bring more performance gain. The results also show the scalability of the decoder layers in this framework." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 449 + ], + "type": "text", + "content": "Selection of filters on mixing. We explore whether the previous filters with adapters are suitable for our new channel mixing. As shown in Tab. 11, we find that using the adapters to fuse previous filters with the current ones is most suitable, which ensures both the adaptability of each stage and the diversity [64] of filters. More importantly, the existence of the additional dynamic mixing is necessary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 568 + ], + "type": "text", + "content": "Modifying the localization supervision. In our cross-stage label assigner, only classification labels are used for gathering, while the supervision for localization is unchanged. Thus, we explore its influence by consistently updating the supervision of classification and localization, i.e. selecting the ground-truth boxes that have the greatest IoU with predicted boxes across stages. The results are shown in Tab. 12, and we find that the performance under the two settings is very close, so we do not modify the localization part for simplicity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Application scope of the cross-stage label assigner. For the " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "-th stage, the application scope of cross-stage label assigner is " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "[\\alpha_i,\\beta_i]" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". In this part, we explore the appropriate values of " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". As shown in Tab. 13, we find that the best setting is " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "[i - 1,L]" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". Yet if too many previous ground-truth labels are included, like " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = 1" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ", this hinders the last decoder layer to learn how to remove duplicates. Moreover, we conduct experiments to verify whether the cross-stage label assigner needs to be applied on all stages, because existing works [4, 24] demonstrate that only early stages can be applied with one-to-many label assignment. On the contrary, as shown in the last line of Tab. 13, we find that our" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6584" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 279, + 142 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 279, + 142 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 279, + 142 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 279, + 142 + ], + "type": "table", + "html": "
CrossScoreAPAP50AP75APsAPmAPl
✓43.061.546.125.045.759.1
43.161.946.525.345.859.7
✓44.062.147.626.047.659.3
✓✓44.863.048.427.548.061.3
", + "image_path": "1793b56f59a7e5a5d8bceaef8119dbe571e3f4e969196a069c53e86a18c84345.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 184, + 274, + 245 + ], + "blocks": [ + { + "bbox": [ + 47, + 146, + 286, + 171 + ], + "lines": [ + { + "bbox": [ + 47, + 146, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 286, + 171 + ], + "type": "text", + "content": "Table 7: The effectiveness of the components of the cross-stage label assigner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 184, + 274, + 245 + ], + "lines": [ + { + "bbox": [ + 58, + 184, + 274, + 245 + ], + "spans": [ + { + "bbox": [ + 58, + 184, + 274, + 245 + ], + "type": "table", + "html": "
NsAPAP50AP75APsAPmAPl
044.562.548.127.448.060.6
144.863.048.427.548.061.3
344.462.448.026.247.460.6
", + "image_path": "ea255782737af3fc1029c4788d3d636d5fc36164baf928d2a6fd4912e0eb986f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 274, + 274, + 323 + ], + "blocks": [ + { + "bbox": [ + 75, + 248, + 257, + 261 + ], + "lines": [ + { + "bbox": [ + 75, + 248, + 257, + 261 + ], + "spans": [ + { + "bbox": [ + 75, + 248, + 257, + 261 + ], + "type": "text", + "content": "Table 8: The number of reused spatial filters." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 274, + 274, + 323 + ], + "lines": [ + { + "bbox": [ + 58, + 274, + 274, + 323 + ], + "spans": [ + { + "bbox": [ + 58, + 274, + 274, + 323 + ], + "type": "table", + "html": "
MethodAPSpeed (img/s)Memory (GB)
Baseline42.613.56.0
Ours44.811.66.3
", + "image_path": "be1f99253be9cf41d45711e4692aed6dfa7311e5a38d78211aa7e44c614fa16f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 57, + 365, + 276, + 449 + ], + "blocks": [ + { + "bbox": [ + 47, + 327, + 286, + 350 + ], + "lines": [ + { + "bbox": [ + 47, + 327, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 286, + 350 + ], + "type": "text", + "content": "Table 9: The inference speed and training GPU memory use." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 365, + 276, + 449 + ], + "lines": [ + { + "bbox": [ + 57, + 365, + 276, + 449 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 276, + 449 + ], + "type": "table", + "html": "
\\( \\max\\{\\Delta\\gamma_i\\} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
044.162.347.625.547.560.3
144.262.347.625.847.560.6
244.262.347.926.148.060.4
344.662.848.326.747.560.7
444.863.048.427.548.061.3
", + "image_path": "9f77c26e4fa3e6639a1214f076b14d84512a1f1caf990824e0b46bb72fd0920f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 55, + 479, + 278, + 552 + ], + "blocks": [ + { + "bbox": [ + 61, + 454, + 272, + 466 + ], + "lines": [ + { + "bbox": [ + 61, + 454, + 272, + 466 + ], + "spans": [ + { + "bbox": [ + 61, + 454, + 272, + 466 + ], + "type": "text", + "content": "Table 10: The number of additional channel mixing." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 479, + 278, + 552 + ], + "lines": [ + { + "bbox": [ + 55, + 479, + 278, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 278, + 552 + ], + "type": "table", + "html": "
FilterAPAP50AP75APsAPmAPl
Pre43.561.746.825.347.059.0
44.261.947.626.147.660.0
Cur44.562.747.926.547.760.5
Adp44.863.048.427.548.061.3
", + "image_path": "2a26540812679dccedbedb93e93f3c70b985c4603dc325014c2851ced24acb5e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 604, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 286, + 676 + ], + "type": "text", + "content": "cross-stage label assigner can also be applied on the last two stages and this brings performance gain. We speculate that this module helps remove some inappropriate supervision caused by vanilla bipartite matching which has a shortage of constraining IoU, and this module provides some proper supervision from the previous stage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "The threshold of IoU. As shown in Tab. 14, we find that using a threshold of 0.5 is the best choice to select appropriate labels." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 535, + 118 + ], + "blocks": [ + { + "bbox": [ + 47, + 556, + 286, + 580 + ], + "lines": [ + { + "bbox": [ + 47, + 556, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 286, + 580 + ], + "type": "text", + "content": "Table 11: Usage of filters. Pre, Cur: directly using previous and current filters. Adp: using adapters." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 535, + 118 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 535, + 118 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 535, + 118 + ], + "type": "table", + "html": "
Loc.APAP50AP75APsAPmAPl
44.862.848.426.748.260.7
✓44.863.048.427.548.061.3
", + "image_path": "6f48b89f3881d9329d590a371c6daa2339d3d1bf56480f4df8dd622285dc040f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 307, + 160, + 545, + 258 + ], + "blocks": [ + { + "bbox": [ + 306, + 122, + 545, + 147 + ], + "lines": [ + { + "bbox": [ + 306, + 122, + 545, + 147 + ], + "spans": [ + { + "bbox": [ + 306, + 122, + 545, + 147 + ], + "type": "text", + "content": "Table 12: Adding localization supervision in cross-stage label assigner." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 160, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 307, + 160, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 307, + 160, + 545, + 258 + ], + "type": "table", + "html": "
Range of iαiβiAPAP50AP75APsAPmAPl
[1,L]ii44.062.147.626.047.659.3
[1,L]i-1i44.162.247.725.947.559.8
[1,L]i-1i+144.162.448.026.547.360.1
[1,L]i-1L44.863.048.427.548.061.3
[1,L]1L43.360.746.825.747.258.6
[1,L-2]i-1L44.362.447.726.147.660.4
", + "image_path": "b0b7fe0d86bb85d67f49ce6a4190f48c5ba2662ee3f16449afaf23cdc5f3192c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 319, + 299, + 531, + 357 + ], + "blocks": [ + { + "bbox": [ + 306, + 262, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 306, + 262, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 545, + 285 + ], + "type": "text", + "content": "Table 13: The application scope of the cross-stage label assigner for each stage." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 299, + 531, + 357 + ], + "lines": [ + { + "bbox": [ + 319, + 299, + 531, + 357 + ], + "spans": [ + { + "bbox": [ + 319, + 299, + 531, + 357 + ], + "type": "table", + "html": "
IoU thresAPAP50AP75APsAPmAPl
0.343.862.447.226.346.860.3
0.744.461.148.426.047.561.5
0.544.863.048.427.548.061.3
", + "image_path": "e5dd0923f6a6ce4aac1ff87fc5c67be42d85c5d4d444aa1556c9602d1f2d3a41.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 361, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 306, + 361, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 545, + 383 + ], + "type": "text", + "content": "Table 14: The threshold of IoU in cross-stage label assignment." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 407, + 378, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 378, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 378, + 419 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 429, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 595 + ], + "type": "text", + "content": "In this paper, we have presented a new fast-converging query-based object detector with cross-stage interaction, termed as StageInteractor, to alleviate inconsistency between the one-to-one label assignment and improve the modeling capacity. Our proposed cross-stage label assigner gathers assigned labels across stages, and then reassigns proper labels to predictions in each stage. We also accumulate dynamic filter across stages and reuse them with lightweight modules to increase the modeling capacity without introducing too many parameters. With these two unique designs, StageInteractor significantly outperforms the previous methods, and our proposed label assignment can boost the performance of various query-based object detectors." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by National Key R&D Program of China (No. 2022ZD0160900), National Natural Science Foundation of China (No. 62076119, No. 61921006), Fundamental Research Funds for the Central Universities (No. 020214380091, No. 020214380099), and Collaborative Innovation Center of Novel Software Technology and Industrialization. The authors would like to thank Ziteng Gao for helpful discussions." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6585" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. Character-level language modeling with deeper self-attention. In Proceedings of the AAAI conference on artificial intelligence, pages 3159–3166, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "type": "text", + "content": "[2] Navaneeth Bodla, Bharat Singh, Rama Chellappa, and Larry S Davis. Soft-nms-improving object detection with one line of code. In Proceedings of the IEEE international conference on computer vision, pages 5561-5569, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 226 + ], + "type": "text", + "content": "[3] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: high quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 282 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proceedings of the European conference on computer vision, pages 213-229, 2020. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 372 + ], + "type": "text", + "content": "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 374, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 287, + 417 + ], + "type": "text", + "content": "[6] Qiang Chen, Xiaokang Chen, Gang Zeng, and Jingdong Wang. Group detr: Fast training convergence with decoupled one-to-many label assignment. arXiv preprint arXiv:2207.13085, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 419, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 287, + 453 + ], + "type": "text", + "content": "[7] Xiaokang Chen, Fangyun Wei, Gang Zeng, and Jingdong Wang. Conditional detr v2: Efficient detection transformer with box queries. arXiv preprint arXiv:2207.08914, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 454, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 507 + ], + "type": "text", + "content": "[8] Zhe Chen, Jing Zhang, and Dacheng Tao. Recurrent glimpse-based decoder for detection with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5260-5269, 2022. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 565 + ], + "type": "text", + "content": "[9] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16372-16382, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "type": "text", + "content": "[10] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 646 + ], + "type": "text", + "content": "[11] detrex contributors. detrex: An research platform for transformer-based object detection algorithms, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6569-6578, 2019. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[14] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3621–3630, 2021. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "type": "text", + "content": "[15] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 243, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 287 + ], + "type": "text", + "content": "[16] Ziteng Gao, Limin Wang, and Gangshan Wu. Mutual supervision for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3641-3650, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 289, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 334 + ], + "type": "text", + "content": "[17] Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 303-312, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 335, + 545, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 335, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 307, + 335, + 545, + 367 + ], + "type": "text", + "content": "[18] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 370, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 370, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 545, + 403 + ], + "type": "text", + "content": "[19] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 405, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 438 + ], + "type": "text", + "content": "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 486, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 540 + ], + "type": "text", + "content": "[22] Yihui He, Chenchen Zhu, Jianren Wang, Marios Savvides, and Xiangyu Zhang. Bounding box regression with uncertainty for accurate object detection. In Proceedings of the iee/ cvf conference on computer vision and pattern recognition, pages 2888-2897, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 543, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 586 + ], + "type": "text", + "content": "[23] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3588-3597, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "text", + "content": "[24] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. arXiv preprint arXiv:2207.13080, 2022. 3, 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "text", + "content": "[25] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In European Conference on Computer Vision, pages 355-371. Springer, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision, pages 734-750, 2018. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6586" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[27] Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou Zhang, and Zhuowen Tu. Deeply-supervised nets. In Artificial intelligence and statistics, pages 562-570. PMLR, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "type": "text", + "content": "[28] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13619-13627, 2022. 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 231 + ], + "type": "text", + "content": "[29] Xiang Li, Wenhai Wang, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss v2: Learning reliable localization quality estimation for dense object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11632-11641, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 233, + 287, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 276 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 1, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 279, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 334 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 335, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 335, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 48, + 335, + 287, + 379 + ], + "type": "text", + "content": "[32] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 426 + ], + "type": "text", + "content": "[33] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pages 21-37. Springer, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 428, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 287, + 483 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 485, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 507 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "text", + "content": "[36] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 620 + ], + "type": "text", + "content": "[37] Duy-Kien Nguyen, Jihong Ju, Olaf Booij, Martin R Oswald, and Cees GM Snoek. Boxer: Box-attention for 2d and 3d transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4773–4782, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "type": "text", + "content": "[38] Han Qiu, Yuchen Ma, Zeming Li, Songtao Liu, and Jian Sun. Borderdet: Border feature for dense object detection. In European Conference on Computer Vision, pages 549-564. Springer, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[39] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[40] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 150 + ], + "type": "text", + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "type": "text", + "content": "[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 276, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 297 + ], + "type": "text", + "content": "[44] Zhiqiang Shen, Zechun Liu, and Eric Xing. Sliced recursive transformer. arXiv preprint arXiv:2111.05297, 2021. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "text", + "content": "[45] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In International Conference on Machine Learning, pages 9934-9944. PMLR, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 344, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 409 + ], + "type": "text", + "content": "[46] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "text", + "content": "[47] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3611-3620, 2021. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "text", + "content": "[48] Yao Teng and Limin Wang. Structured sparse r-cnn for direct scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19437-19446, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 513, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 555 + ], + "type": "text", + "content": "[49] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 621 + ], + "type": "text", + "content": "[50] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261–24272, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 668 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[52] Hongyu Wang, Shuming Ma, Li Dong, Shaohan Huang, Dongdong Zhang, and Furu Wei. Deepnet: Scaling transformers to 1,000 layers. arXiv preprint arXiv:2203.00555, 2022.3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6587" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[53] Wen Wang, Jing Zhang, Yang Cao, Yongliang Shen, and Dacheng Tao. Towards data-efficient detection transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 88-105. Springer, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[54] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 226 + ], + "type": "text", + "content": "[55] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500, 2017. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "type": "text", + "content": "[56] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 282, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 282, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 287, + 327 + ], + "type": "text", + "content": "[57] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9657-9666, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 358 + ], + "type": "text", + "content": "[58] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "text", + "content": "[59] Mohsen Zand, Ali Etemad, and Michael Greenspan. Object-box: From centers to boxes for anchor-free object detection. arXiv preprint arXiv:2207.06985, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 446 + ], + "type": "text", + "content": "[60] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 448, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 287, + 491 + ], + "type": "text", + "content": "[61] Linfeng Zhang, Xin Chen, Junbo Zhang, Runpei Dong, and Kaisheng Ma. Contrastive deep supervision. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 547 + ], + "type": "text", + "content": "[62] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "text", + "content": "[63] Shilong Zhang, Xinjiang Wang, Jiaqi Wang, Jiangmiao Pang, Chengqi Lyu, Wenwei Zhang, Ping Luo, and Kai Chen. Dense distinct query for end-to-end object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7329-7338, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[64] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 3, 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 669 + ], + "type": "text", + "content": "[65] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[66] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020.1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 162 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "type": "text", + "content": "[67] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "type": "text", + "content": "[68] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6588" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/396c3711-d204-4df2-8e69-9bdf93d4fbff_content_list.json b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/396c3711-d204-4df2-8e69-9bdf93d4fbff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..34743b92c2eddea520e0ef1097875d2b3c6bdcdf --- /dev/null +++ b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/396c3711-d204-4df2-8e69-9bdf93d4fbff_content_list.json @@ -0,0 +1,2541 @@ +[ + { + "type": "text", + "text": "Steered Diffusion: A Generalized Framework for Plug-and-Play Conditional Image Synthesis", + "text_level": 1, + "bbox": [ + 161, + 114, + 808, + 159 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nithin Gopalakrishnan Nair $^{1*}$ Anoop Cherian $^{2}$ Suhas Lohit $^{2}$ Ye Wang $^{2}$ Toshiaki Koike-Akino $^{2}$ Vishal M. Patel $^{1}$ Tim K. Marks $^{2}$", + "bbox": [ + 181, + 169, + 785, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Johns Hopkins University $^{2}$ Mitsubishi Electric Research Laboratories (MERL)", + "bbox": [ + 236, + 207, + 730, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ngopala2,vpatel36}@jhu.edu {acherian,slohit,yewang,koike,tmarks}@merl.com", + "bbox": [ + 225, + 227, + 743, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://merl.com/demos/steered-diffusion", + "bbox": [ + 303, + 244, + 661, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6d43ea9497aeeff66941815a9e222105e0545c582c8aecb02294b764894edb9.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 284, + 166, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20d52534047cca05edb7e9035eb3076a6d555bfd8b0c15b5ee41f1108a4dce05.jpg", + "image_caption": [ + "(a) Inpainting" + ], + "image_footnote": [], + "bbox": [ + 171, + 284, + 256, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5db5887effc534890999561fdd8db83c9371a6401f12f3cfde3ab2741ab5b153.jpg", + "image_caption": [ + "Generated Samples" + ], + "image_footnote": [], + "bbox": [ + 258, + 284, + 343, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01a8e0bf55bb1684b1cb8a53e68a63e6e47b64622decfafc6a291b2f55ed9a83.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 354, + 282, + 439, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34589c9422d014e2a197dab34129bfca1142c3286f1b6424a7ea0504dde6a37b.jpg", + "image_caption": [ + "(b) Colorization", + "Generated Samples" + ], + "image_footnote": [], + "bbox": [ + 447, + 282, + 531, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1394637a8ae4cb0854b93fe08a225eb627d625fe1296f43a859d5a2ae318c74f.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 604, + 268, + 717, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af87089ed381715a43e164e486bc3c1664fbd1f1c0517bf17685f1ac7dd3d634.jpg", + "image_caption": [ + "Generated Samples" + ], + "image_footnote": [], + "bbox": [ + 710, + 268, + 815, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7abccc03400d49abfc04759453e279268a5a911abcfe0f787977545a64c18f90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 268, + 893, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/099e2a4b78121cfcb5bd3b78bf502bad6315dd5208be214dc824afac59dc613f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 431, + 166, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ffe046b2525aa3da7e78401245d277b7ecc7e10ba4c515b507d0ec1616c858f4.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 78, + 515, + 165, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/020da89ebd90ba9e9f33ebfd93abe775b7a9eabb1c5a1de0b49f55a87411b312.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 448, + 256, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac63b95cefee6a5dd25e6eaa8d264a4d37f40d44d0d725aac4248d4c71ee302e.jpg", + "image_caption": [ + "Generated Samples" + ], + "image_footnote": [], + "bbox": [ + 169, + 515, + 256, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a376821d67a3b1580a4bf17597e6915715ce5ea4b6bcad893daf31bf43137631.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 448, + 343, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae4da69c0abf6bc633990e8f92d5852b7eb128e2656adb312c6791fcb766ad34.jpg", + "image_caption": [ + "Figure 1. An illustration of various applications of our method. We use a diffusion model trained unconditionally and condition using our proposed algorithm only during the test time. We present the results on six tasks: (a) image inpainting, (b) colorization, (c) image super-resolution, (d) semantic generation, (e) identity replication, and (f) text-based image editing. In part (f), the text prompts for the first and second columns, respectively, are \"This person has blonde hair\" and \"This person has wavy hair.\"" + ], + "image_footnote": [], + "bbox": [ + 258, + 515, + 341, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e14d42300e62a0cfa860f8f36a4b23a96a8fe235f42c6f437a187eb663448197.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 448, + 439, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a887ce85cac79f97bade9936416150ad4a6d1c7e151c6da753fdb2f538da4c5.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 354, + 515, + 439, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59a2addfe500d8a7bd65a28becf9b2722a52d97de454491352ab9856ef91b866.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 448, + 531, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/28d0ba19b215f092bf1b19ca7fbe0a3c85bd2701eb72eeb788a21bae5bee6635.jpg", + "image_caption": [ + "Generated Samples" + ], + "image_footnote": [], + "bbox": [ + 447, + 515, + 532, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/091addb3904f01a135e8e72bb484a3c8dfe956d27319d450c5e97abc3cac0fc2.jpg", + "image_caption": [ + "(f) Text-based editing" + ], + "image_footnote": [], + "bbox": [ + 630, + 449, + 715, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40d434215f88d8ed34d74528dd777296a85dc411a100324caae98ff515e311cb.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 630, + 515, + 715, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce17f26b769a984c8ecebf62ed7714e9357e9d413c6169a1dc83115b81719973.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 449, + 807, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f4a9b88b7ec46662e3d0290fa7951984a2b503a4d582bb6a21e30be36fc9292.jpg", + "image_caption": [ + "Generated Samples" + ], + "image_footnote": [], + "bbox": [ + 722, + 515, + 807, + 579 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a419a4b54bc0cacd25c0d6a1c3b30ba6f63672ac702bf9840d8f4bf5782fb93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 449, + 893, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f15468105f903a50aa454a42249cb876ab534bbd78a325f79e41f027d80f2935.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 515, + 893, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 259, + 666, + 339, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conditional generative models typically demand large annotated training sets to achieve high-quality synthesis. As a result, there has been significant interest in designing models that perform plug-and-play generation, i.e., to use a predefined or pretrained model, which is not explicitly trained on the generative task, to guide the generative process (e.g., using language). However, such guidance is typically useful only towards synthesizing high-level semantics rather than editing fine-grained details as in image-to-image translation tasks. To this end, and capitalizing on the powerful fine-grained generative control offered by the recent diffusion-based generative", + "bbox": [ + 75, + 696, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models, we introduce Steered Diffusion, a generalized framework for photorealistic zero-shot conditional image generation using a diffusion model trained for unconditional generation. The key idea is to steer the image generation of the diffusion model at inference time via designing a loss using a pre-trained inverse model that characterizes the conditional task. This loss modulates the sampling trajectory of the diffusion process. Our framework allows for easy incorporation of multiple conditions during inference. We present experiments using steered diffusion on several tasks including inpainting, colorization, text-guided semantic editing, and image super-resolution. Our results demonstrate clear qualitative and quantitative improvements over state-of-the-art diffusion-based plug-and-play models while adding negligible additional computational cost.", + "bbox": [ + 496, + 667, + 893, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during internship at MERL.", + "bbox": [ + 99, + 887, + 318, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20850", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep diffusion-based probabilistic generative models [8, 14, 40] are quickly emerging as one of the most powerful methods to synthesize high-quality content and have shown the potential to revolutionize content creation not only in computer vision, but also in many other areas including speech, audio, and language. Such models (e.g., ImagGen [36], Stable Diffusion [34]) have demonstrated outstanding synthesis results in conditional generation tasks, such as text-conditioned image synthesis [2, 33] and image reconstruction [31, 35, 38]. However, these models do not typically possess zero-shot conditional generative abilities when used directly (zero-shot capabilities as are commonly seen in language foundation models such as GPT-3 [4]), and often demand large amounts of annotated and paired (multimodal) data for conditional generation, which may be challenging to obtain [15].", + "bbox": [ + 76, + 114, + 468, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One way to circumvent this need for large annotated training sets is to leverage predefined models as plug-and-play modules [12, 24, 25] in an otherwise unconditionally trained diffusion model. Specifically, in such plug-and-play models, a model is first trained in an unconditional setting (without labels). During inference, the plug-and-play modules (networks separately trained for a particular conditional task, e.g., image captioning) are incorporated in the reverse diffusion process to produce intermediate samples guided in the Markov chain in specific directions to satisfy the desired condition. Prior works, such as [25, 26], have proposed similar methods in which the authors derive text- or class-conditioned samples from Generative Adversarial Networks (GANs) [11] that were trained without labels. To achieve this, they iteratively refine the noise input of the GAN until the desired sample satisfies the condition. Very recently, Grakios et al. [12] proposed a diffusion-based plug-and-play method that enables using unconditional diffusion models for conditional generation utilizing class labels. Both these methods are specifically designed for tasks involving label-level semantics. However, these methods do not address the usage of unconditional models for general image-to-image translation tasks, which require synthesizing visual content conditioned on fine-grained details in the source image. There are also works that propose diffusion models for image-to-image translation, such as for image super-resolution and inpainting [5, 20]; however, these methods are task-specific and do not generalize well to new tasks or new types of inverse problems (as demonstrated in Section 5). In this work, we present a generic framework that can generalize to any image-to-image translation task.", + "bbox": [ + 76, + 356, + 470, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we derive the necessary theory and formulate an algorithm, which we call Steered Diffusion, for diffusion-based image editing and image-to-image translation; our model is subsequently validated on a wide range of tasks. Steered diffusion is motivated by the energy-based", + "bbox": [ + 76, + 825, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c1801592f0ca9a188469e33ae593f7b8df6decc559534a3f1f23fad90c633641.jpg", + "image_caption": [ + "Figure 2. An illustration of the difference between existing plug-and-play generation approaches (e.g., [12]) and the proposed approach. Existing plug-and-play works operate with an energy function $V$ of the noisy latent $x_{t}$ . In contrast, our model uses the implicit prediction of the diffusion model (i.e., a coarse estimate of the clean image $x_{0}$ ) in its energy function $V_{1}$ , which allows the use of any pre-trained network for steering. In addition, our model provides a looping mechanism $V_{2}$ , which iterates $N$ times at each timestep $t$ to enhance generation quality." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "formulation of diffusion probabilistic models [10]. In general, inference in a generative model can be thought of as deriving samples from a learned distribution. Recall that every probability density function can be formulated as an energy field that describes an unnormalized estimate of how the distribution density varies in space [13,25]. If one needs to find points in space that are the closest match to a given condition, one can utilize gradient-based optimization algorithms to find points in the field that have the highest density value for the condition. The gradient-based optimization scheme can be viewed as a modulation of the energy toward the desired direction. Previous work has utilized this idea on GANs [25, 26] and obtained reasonable results for label-based generation tasks. Due to their model structure, diffusion models are ideal candidates for such an energy modulation. One key challenge remains to design a good energy estimator that is robust to all noise levels. Previously, classifier-based guidance [8, 27] has been proposed and thought of as an energy modulation utilizing a pretrained classifier trained on noisy images. This poses a limitation that the guiding function should be noise-robust. In this work, we propose an alternative solution that does not need noise-robust networks but could use any network by utilizing the diffusion model as an implicit denoiser. Figure 2 gives a brief overview of how our approach is different from existing methods.", + "bbox": [ + 496, + 426, + 893, + 818 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present experiments using steered diffusion on multiple conditional generative tasks on faces as well as generic images as portrayed in Figure 1. We present results on (i) identity replication [7], (ii) semantic image generation [29], (iii) linear inverse problems [21], and (iv) text-conditioned", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20851", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image editing. Although our method is generic, for evaluations we perform experiments on faces. Before presenting our framework in detail, we now summarize the key contributions of our work:", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose steered diffusion, a general plug-and-play framework that can utilize various pre-existing models to steer an unconditional diffusion model.", + "- We present the first work applicable to both label-level synthesis and image-to-image translation tasks and demonstrate its effectiveness for various applications.", + "- We propose an implicit conditioning-based sampling strategy that significantly boosts the performance of conditional sampling from unconditional diffusion models compared with previous methods.", + "- We introduce a new strategy that uses multiple steps of projected gradient descent to improve sample quality." + ], + "bbox": [ + 94, + 160, + 468, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 76, + 354, + 205, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Related Work", + "text_level": 1, + "bbox": [ + 76, + 378, + 220, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early works on unpaired image-to-image translation utilize a cycle consistency loss between the input and the target domains [9, 46]. Newer works, such as [16], have introduced a contrastive learning-based approach where a contrastive loss between corresponding patches of the input and target domains is minimized. The consistency-based method often fails to generate photorealistic images; hence, conditional generative models are preferred when labeled data are available. A few works [37, 38] utilize diffusion models for conditional image-to-image translation because of their photorealistic generation quality.", + "bbox": [ + 75, + 402, + 468, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Guiding diffusion models during inference time has been explored by several works, such as [24]. The first method that proposed inference-time conditioning [8] uses a pretrained noise-robust classifier to guide the inference of an unconditional model. GLIDE [27] proposed a method for conditioning using text. Earlier work in plug-and-play modelling for generative models utilized GANs and performed iterative refinement on the latent space of GANs [25]. This method uses a predefined classifier or text captioning network to estimate a loss between the desired label output or text caption and the one generated from the GAN generator. This loss is backpropagated to refine the noise input of the GAN iteratively until the generator predicts the desired output. Recently, [12] proposed a method that uses diffusion models as a plug-and-play prior for class-conditioned generation. Several works have addressed the task of image-to-image translation using unconditional diffusion models [1,5,18,20], but each of these proposes a task-specific inference scheme. For example, ILVR [5] performs image super-resolution, and Reinpaint [20] performs image inpainting. Blended diffusion [1] proposes a method for text-conditioned image editing. DDRM [18] proposes an inference-time scheme", + "bbox": [ + 75, + 569, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "offering a general solution for linear inverse problems such as colorization and super-resolution.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Concurrent Work", + "text_level": 1, + "bbox": [ + 500, + 130, + 671, + 143 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In concurrent work to ours that also explores zero-shot conditional generation using diffusion models, [3] used a text to image model [34] and a two-step forward and backward universal guidance process, but it works well only after heavy optimization on network-based inverse problems such as semantic generation and identity generation. Another concurrent work [42] explored the usage of unconditional diffusion models for linear inverse problems using a pseudo-inverse model. In contrast to all these prior works, our steered diffusion algorithm generalizes well to both image-to-image translation tasks and high-level label-based generation tasks.", + "bbox": [ + 496, + 152, + 893, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Denoising Diffusion Probabilistic Models", + "text_level": 1, + "bbox": [ + 500, + 329, + 849, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising diffusion probabilistic models (DDPMs) [14, 40] belong to a class of generative models in which the model learns the distribution of data through a Markovian sampling process. DDPMs consist of a forward process and a reverse process. Let $x_{t}$ denote the latent state of an input image at timestep $t$ in a diffusion process. The sampling operation $q(\\cdot)$ for the forward process in DDPM is defined as:", + "bbox": [ + 496, + 352, + 890, + 458 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) := \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 468, + 890, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\{\\beta_t\\}$ is a predefined variance schedule and $I$ is the identity matrix.", + "bbox": [ + 496, + 497, + 890, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The forward process can be considered as a noising operation, where the next state $x_{t}$ is obtained from the current state $x_{t-1}$ by adding a small amount of Gaussian noise according to the sampled timestep. The state $x_{t}$ at timestep $t$ can also be sampled directly from the initial state $x_{0}$ , using:", + "bbox": [ + 496, + 527, + 893, + 603 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {0}\\right) := \\mathcal {N} \\left(x _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) I\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 614, + 890, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "or equivalently,", + "bbox": [ + 500, + 642, + 604, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = x _ {0} \\sqrt {\\bar {\\alpha} _ {t}} + \\epsilon \\sqrt {1 - \\bar {\\alpha} _ {t}}, \\quad \\epsilon \\sim \\mathcal {N} (0, I), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 667, + 890, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s$ and $\\alpha_t = 1 - \\beta_t$ .", + "bbox": [ + 496, + 694, + 759, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [40], it is shown that if the number of time steps is large and the increment in $\\{\\beta_i\\}$ is small, then each step in the reverse sampling process can also be approximated by a Gaussian. If $\\mu_{\\theta}$ and $\\Sigma_{\\theta}$ respectively denote the mean and the covariance of this Gaussian, modelled via neural networks with parameters $\\theta$ , then each reverse step samples the state $x_{t-1}$ according to:", + "bbox": [ + 496, + 710, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) := \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 827, + 890, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The parameters $\\theta$ are obtained by minimizing the variational lower bound of the negative log-likelihood of the data distribution.", + "bbox": [ + 496, + 854, + 893, + 898 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "20852", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f0e06ea56d133f147010aea7036c8bdbb7ac94684f5fc7400b7ec82ea2f0386e.jpg", + "image_caption": [ + "Figure 3. An illustration of Steered Diffusion. During each step of the sampling process the implicit prediction is steered to the direction of the condition using a steering network or predefined function. Note that this figure is only for illustrating the idea and does not show the actual sampled images; in the actual sampling, the steering process is much more gradual, not sudden as portrayed in this image." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 588, + 263 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2e5c8ae020418efe7901e29213e5df5d82d4260eb0794667150e284e33207bd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 88, + 874, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 314, + 250, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Steered Diffusion at Inference Time", + "text_level": 1, + "bbox": [ + 76, + 339, + 387, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our work is motivated by the energy-based formulation of diffusion models. For any probability density function, the corresponding energy-based model (EBM) is defined by:", + "bbox": [ + 75, + 362, + 472, + 409 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (x) = \\frac {\\exp (- V (x))}{Z}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 429, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V(x)$ denotes the corresponding energy function across states $x$ , and $Z$ denotes a normalization constant. To derive samples from this distribution, one can utilize the Langevin equation [39] describing the state transition of a particle in the presence of an energy field. For diffusion models the sampling step is", + "bbox": [ + 75, + 469, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = x _ {t} - \\nabla_ {x _ {t}} \\log p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) + \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 585, + 468, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The term $\\nabla_{x_t}\\log p_\\theta (x_{t - 1}|x_t)$ is called the score function of the density $p_{\\theta}(x_{t - 1}|x_t)$ . One key advantage of the energy-based formulation is that it allows modulation of the energy function to satisfy given criteria. This was initially introduced as classifier guidance [8], which allows label conditional sampling from an unconditionally trained diffusion model utilizing a noise-robust classifier. In the remaining part of this section, we motivate how we can extend the functionality of unconditional diffusion models to conditional tasks. Consider a conditional sampling scenario based on a condition $c$ , for sampling from a state $x_{t}$ to state $x_{t - 1}$ . The conditional transition probability $p_{\\theta}(x_{t - 1}|x_t,c)$ can be decomposed as", + "bbox": [ + 75, + 612, + 472, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, c\\right) \\propto \\frac {p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) p \\left(c \\mid x _ {t - 1}\\right)}{p \\left(c \\mid x _ {t}\\right)}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 829, + 468, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, for any timestep $t$ , the effective score for conditional transition can be found by utilizing using the log", + "bbox": [ + 76, + 869, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of probability density in the EBM formulations (5) of the individual densities and can be represented as", + "bbox": [ + 498, + 316, + 890, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t - 1} | x _ {t}, c) =\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 356, + 679, + 373 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {x _ {t}} \\log p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) - \\nabla_ {x _ {t}} V _ {1} \\left(x _ {t}, c\\right) + \\nabla_ {x _ {t}} V _ {2} \\left(x _ {t - 1}, c\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 375, + 890, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V_{1}$ and $V_{2}$ are the corresponding energy functions that model the conditional distributions of $x_{t}$ and $x_{t-1}$ given a condition $c$ . Specifically, they project the higher dimensional $x_{t}$ to the lower dimensional space of $c$ and measure the distance between the mapped value and $c$ . The better this particular measure, the more effectively it can be used to generate conditional samples from an unconditional model. Using Eq. (8), the conditional sampling equation for the reverse process is", + "bbox": [ + 496, + 416, + 893, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} ^ {(t)} (x _ {t}, t)\\right) \\\\ - \\nabla_ {x _ {t}} V _ {1} (x _ {t}, c) + \\nabla_ {x _ {t}} V _ {2} (x _ {t - 1}, c) + \\sigma_ {t} \\epsilon . \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 561, + 890, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\epsilon_{\\theta}^{(t)}(x_t,t)$ is the network prediction at a timestep $t$ and $\\sigma_{t}$ is the corresponding variance of the reverse step. The formulation in Eq. (9) shows that the energy function requires a functional mapping from a noisy $x_{t}$ to $c$ .", + "bbox": [ + 496, + 626, + 893, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In many applications of interest, the mapping function from $x_{t}$ to $c$ is complex and can be modeled effectively using deep networks. For example, in image-to-image translation tasks such as image generation from semantic maps [29], the image is $x_{t}$ and the semantic map is the condition $c$ . Similarly, for text to image generation, the image is $x_{t}$ and $c$ corresponds to the text. Ideally, we would like to employ an existing (pre-trained) deep neural network for the mapping from $x_{t}$ to $c$ . However, deep networks are usually trained on clean images, which limits the usability of existing pretrained deep networks for mapping directly from the noisy image $x_{t}$ to $c$ . One workaround would be to use noise-robust networks to map from $x_{t}$ to $c$ , but training noise-robust networks for conditional mapping can be computationally", + "bbox": [ + 496, + 689, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20853", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "expensive. Moreover, a network that is trained with multiple different noise levels often results in lower mapping performance, as it cannot denoise all noise levels accurately; we validate this claim experimentally in Section 5.6. Alternatively, one could include two mapping functions: a first that denoises $x_{t}$ , and a second that maps from the denoised image to $c$ . Rather than training a seperate denoising network, however, we realized that diffusion models are inherently trained as denoisers, and reconstruction quality improves as time proceeds in the reverse sampling of the diffusion process. Because of this capacity, we can use a reverse sampling step to make coarse predictions of the denoised image from any time step $t$ .", + "bbox": [ + 75, + 90, + 472, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hence, we modify our original energy expression (8) to:", + "bbox": [ + 96, + 287, + 470, + 304 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t - 1} | x _ {t}, c) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t - 1} | x _ {t}) \\\\ - \\nabla_ {x _ {t}} V _ {1} \\left(x _ {0 \\mid t}, c\\right) - \\delta_ {1} + \\nabla_ {x _ {t}} V _ {2} \\left(x _ {0 \\mid t - 1}, c\\right) + \\delta_ {2}, \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 330, + 470, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, we define the implicit step prediction $x_{0|t}$ as:", + "bbox": [ + 76, + 393, + 423, + 410 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {0 \\mid t} = \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (x _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 422, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we assume $x_{t}$ and $x_{t-1}$ are first denoised to $x_{0|t}$ and $x_{0|t-1}$ , respectively. The terms $\\delta_{1}$ and $\\delta_{2}$ capture the errors arising from the shift in the domain from $x_{t}$ to $x_{0}$ and from $x_{t-1}$ to $x_{0}$ ; for large $t$ , $\\delta_{1} \\approx \\delta_{2}$ as the implicit predictions at nearby steps tend to be similar.", + "bbox": [ + 76, + 470, + 468, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in our experiments, the energy function should be selected according to the task. An easy way to choose the energy function is by looking at the training loss of the mapping network. For example, in the case of semantic generation, a good energy function is the cross-entropy loss between the predicted semantic map at any timestep and the input semantic map. In the case of identity replication, a good choice of regularization would be the negative cosine similarity score between the embeddings from a recognition network for the input and target image. In the case of text-to-image generation, it would be CLIP loss [32]. As a rule of thumb, an energy function could be chosen easily by looking at the loss function used to train the pre-trained network (or an inverse function) that maps from the image $x$ to the condition $c$ .", + "bbox": [ + 75, + 547, + 472, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Revisiting Sampling in Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 785, + 423, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain a closed-form expression for plugging the energy-based formulation into the reverse sampling process efficiently, we take inspiration from DDIM [41] and revisit the reverse sampling operation of diffusion models. From $p_{\\theta}(x_{1:T})$ , one can generate a sample $x_{t-1}$ from a sample $x_{t}$ by:", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Steered Diffusion" + ], + "code_body": "Input: Energy function $V$ , condition $c$ \n1: $x_{T} \\sim \\mathcal{N}(x_{T}; 0, I)$ \n2: for $t = T - 1, \\ldots, 1$ do \n3: for $n = N, \\ldots, 1$ do \n4: $\\epsilon \\sim \\mathcal{N}(\\epsilon; 0, I)$ \n5: $x_{0|t} = \\frac{x_{t} - \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon_{\\theta}^{(t)}(x_{t})}{\\sqrt{\\bar{\\alpha}_{t}}}$ \n6: Compute $x_{0|t}^{\\text{feas}}$ with $V, c$ using Eq. (15) \n7: if $(n > 1)$ then \n8: Compute $x_{t}^{uc}$ using Eq. (13) \n9: else \n10: Compute $x_{t-1}^{uc}$ using Eq. (13) \n11: end if \n12: end for \n13: end for \n14: return $x_{0}$", + "bbox": [ + 500, + 107, + 862, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t - 1} ^ {\\mathrm {u c}} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\cdot \\underbrace {\\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (x _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}}} _ {\\text {‘ p r e d i c t e d x _ {0} ”}} \\\\ + \\underbrace {\\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\cdot \\epsilon_ {\\theta} ^ {(t)} \\left(x _ {t}\\right)} _ {\\text {d i r e c t i o n p o i n t i n g t o} x _ {t} ”} + \\underbrace {\\sigma_ {t} \\epsilon} _ {\\text {r a n d o m n o i s e}}, \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 396, + 892, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as in Song et al. [41]. Using Eq. (11), we can rewrite the unconditional sampling step Eq. (12) in terms of $x_{0|t}$ :", + "bbox": [ + 498, + 508, + 890, + 540 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t - 1} ^ {\\mathrm {u c}} = \\sqrt {\\bar {\\alpha} _ {t - 1}} x _ {0 | t} + \\\\ \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\cdot \\frac {x _ {t} - \\sqrt {\\bar {\\alpha} _ {t}} x _ {0 | t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} + \\sigma_ {t} \\epsilon , \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 547, + 892, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here the superscript uc denotes the unconditional sample, which is obtained without any steering while transitioning from $x_{t}$ to $x_{t-1}$ . The conditional sampling step (8) can then be rewritten as", + "bbox": [ + 498, + 609, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = x _ {t - 1} ^ {\\mathrm {u c}} - \\nabla_ {x _ {t}} V _ {1} \\left(x _ {0 \\mid t}, c\\right) + \\nabla_ {x _ {t}} V _ {2} \\left(x _ {0 \\mid t - 1}, c\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 678, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through this, we can modulate $x_0$ directly, as $x_{0|t}$ is also a function of $\\epsilon_{\\theta}^{(t)}(x_t)$ . Following Eq. (14), a rough estimate of the desired $x_0$ for conditional sampling, represented by $x_{0|t}^{\\mathrm{feas}}$ , can be obtained using", + "bbox": [ + 498, + 704, + 890, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {0 | t} ^ {\\text {f e a s}} = x _ {0 | t} - k (t) \\nabla_ {x _ {t}} \\left(V _ {1} \\left(x _ {0 | t}, c\\right) - V _ {2} \\left(x _ {0 | t - 1}, c\\right)\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 780, + 890, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k(t)$ is a scaling factor defining the strength of regularization. We call the process of finding $x_{0|t}^{\\mathrm{feas}}$ from $x_{0|t}$ Implicit Steering Control (ISC), and we call the new sampling process steered diffusion. The exact algorithm is illustrated in Fig. 3 and explained in Algorithm 1.", + "bbox": [ + 496, + 824, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20854", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4e5898a5209a4adea25b9568a2695fdb4231e7516881f8935b027324939e0843.jpg", + "image_caption": [ + "Figure 4. An illustration of the steering function for linear inverse problems. For linear inverse problems, the component of the implicit prediction along the degradation direction can be replaced by the ground truth condition." + ], + "image_footnote": [], + "bbox": [ + 127, + 88, + 395, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Tips for Improved Performance", + "text_level": 1, + "bbox": [ + 76, + 319, + 366, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Linear Inverse Problems", + "text_level": 1, + "bbox": [ + 76, + 344, + 302, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For optimization-based inverse problems such as text-to-image generation and semantic map to natural image generation, the exact mapping function is not always available. On the other hand, for linear inverse problems such as colorization, super-resolution, and image inpainting, the mapping is simply a linear function, and we can write Eq. (15) more simply. In these cases, the exact mapping function to the latent space of the condition is known. Hence, one can decompose the implicit prediction at each timestep along the direction of the condition and simply replace this component by its desired ideal condition, i.e., if our predicted sample needs to map to a condition $c$ , then the modified implicit prediction step becomes", + "bbox": [ + 75, + 368, + 470, + 565 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx _ {0 | t} ^ {\\text {f e a s}} = x _ {0 | t} + k (t) \\left(D (y) - D \\left(x _ {0 | t}\\right)\\right), \\quad \\text {w h e r e} c = D (y). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 575, + 468, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $y$ is the clean image and $D$ is the known degradation model. Our sampling procedure ensures that the series of operations preserve the consistency of domains of $x_{t-1}$ and $x_{0|t}^{\\mathrm{feas}}$ to the original data distribution at the corresponding timesteps. An illustration is shown in Figure 4.", + "bbox": [ + 75, + 623, + 468, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Multi-Step Implicit Modulation", + "text_level": 1, + "bbox": [ + 76, + 710, + 354, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiments show (see Fig. 7) that performing refinement using Eq. (15) on the implicit step prediction multiple times for each timestep significantly boosts the conditioning quality for more ill-posed conditions such as image inpainting and colorization. A similar observation was also found by [20]. Specifically, at a particular timestep $t$ , we iterate the procedure of steering towards the next sampling step $x_{t - 1}$ and then adding noise to return to $x_{t}$ . We present the corresponding algorithm in Algorithm 1. An example is shown in Fig. 7, where more realistic images are generated using the multiple-step sampling scheme row labeled \"OURS multi.\"", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectively, the $V_{2}$ term in Eq. (12) can be thought of as enabling a multistep sampling in which we modulate the current step by looking ahead to the next sampling step. On a careful analysis of Eq. (7), i.e. from the score contribution due to the different regularization functions, we can see that the term for $\\nabla_{x_t}V_1(x_t,c)$ modulates $x_{t}$ based on its current state, and the term from $\\nabla_{x_t}V_2(x_{t - 1},c)$ is a look-ahead correction where the derivative with respect to the future prediction is found. This is exactly what happens in the case of looping back from $x_{t - 1}$ to $x_{t}$ , where $x_{t}$ is modulated iteratively by looking forward to what the future prediction would be.", + "bbox": [ + 496, + 90, + 890, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Choosing the Scaling Factor $k(t)$", + "text_level": 1, + "bbox": [ + 500, + 280, + 782, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Eq. (15), $k(t)$ denotes the strength of the regularization constraint. A very small $k(t)$ would denote no effective regularization, and a large $k$ would lead to the diffusion process going out of the latent space manifold. Since the derivative of the regularization function by itself is a score value, similar to the normal scaling value of the score function, the appropriate time-varying normalization factor is $\\sqrt{1 - \\bar{\\alpha}_t}$ . The exact value for $k(t)$ for each task is defined in Table 1. For linear inverse problems we use a constant $k(t) = 1$ , which provided the best results.", + "bbox": [ + 496, + 303, + 893, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 465, + 632, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of our network qualitatively and quantitatively using four image-to-image translation tasks—semantic layout to face image translation, face inpainting, face colorization, and face super-resolution—as well as two high-level vision tasks: identity-based image generation and text-guided image editing. Unlike existing approaches, our method is completely zero-shot and applies to a wide variety of tasks. We compare with other diffusion-based approaches best applicable for each task for a fair evaluation. We also compare the semantic layout to image translation performance with that of task-specific unsupervised methods. We choose the unconditional model released by [5] as the unconditional pretrained diffusion model in all of our experiments with faces. Note that the sampling scheme in ILVR [5] and Repaint [20] can be thought of as happening at time $t$ rather than at the implicit step as in our method. Hence, comparing these methods for super-resolution and inpainting in our experiments below can be considered an additional ablation study highlighting the improvement from our implicit sampling.", + "bbox": [ + 496, + 491, + 893, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 801, + 717, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For our experiments, we utilize pixel-level unconditional diffusion models. For faces, we utilize the model trained on the FFHQ dataset [17] that was released by [5]. For generic images, we use the model trained on ImageNet [6] released in ADM [8]. All our experiments use 100 steps of sampling.", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20855", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9108f10b96e8b300083c4dd80ebd8ddbd145f0df0f753b2d635d6751fb597ce2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLinear inverse problemsComplex inverse problems
ColorizationInpaintingSuper-resolutionSemantic GenerationIdentity ReplicationImage Editing
V1(D(x0|t) - c)2(D(x0|t) - c)2(D(x0|t) - c)2BCE(D(x0),c)1 - D(x0|t).D(c)3(D1(x0|t),D1(c1))2 + 2000CS(D2(x0|t),c2)
V2(D(x0|t-1) - c)2(D(x0|t-1) - c)20000
DGrayscaleMaskingDownscalingFARL [45]FARL recognizer [45]VGG Face [30] & FARL CLIP [32]
N331111
k(t)1112000√1-αt3000√1-αt10√1-αt & 1500√1-αt
", + "bbox": [ + 93, + 80, + 879, + 175 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Parameter set for each application.", + "bbox": [ + 354, + 180, + 612, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/70bca2bf639c04e835e046524b143f7e51da70c6cd9ca0cc437d0ff13fd9d523.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons for semantic generation." + ], + "image_footnote": [], + "bbox": [ + 80, + 217, + 465, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5860c116ea3730c61b86594e20ad8df02348eb991d591d82264a81b00f4ff28.jpg", + "image_caption": [ + "Figure 6. Results on $8 \\times$ super resolution." + ], + "image_footnote": [], + "bbox": [ + 78, + 431, + 473, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Semantic Face Generation", + "text_level": 1, + "bbox": [ + 76, + 693, + 313, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate how our method performs in generic image-to-image translation tasks, we evaluate our method's performance for the task of semantic layout to face generation. We utilize the CelebA dataset for this. To generate the semantic labels, we use facer [45] and create 11 label classes for each face. Since there is no other unconditional model that can perform fully test-time semantic generation, to evaluate the performance of our method, we compare with fully unsupervised image translation methods: CycleGAN [46], CUT [28], and ILVR [5].", + "bbox": [ + 75, + 717, + 470, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The corresponding qualitative results are shown in Fig. 5. It is clear that CUT, CycleGAN, and ILVR produce unrealis-", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/654e90b6fef1a0f14d94c6c9d443918f73b16fc15334d90fbab349901832da28.jpg", + "image_caption": [ + "Figure 7. Qualitative comparisons for colorization. The row labeled \"OURS multi\" refers to the use of multi-step sampling, as described in Sec. 4.2." + ], + "image_footnote": [], + "bbox": [ + 501, + 217, + 900, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tic facial images with huge artifacts or create low-resolution faces. In contrast, our method always creates good-quality realistic faces. We present the quantitative results in Table 2. The table shows that our method obtains the best FID scores of all methods and obtains the best mIOU score among the inference-time techniques.", + "bbox": [ + 498, + 529, + 893, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Face Super-Resolution", + "text_level": 1, + "bbox": [ + 500, + 633, + 709, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the performance for the face super-resolution task using the CelebA dataset [19]. As baselines, we utilize fully inference-time methods in which no task-specific training is used. As the first baseline, we choose PULSE [22], a self-supervised upsampling technique utilizing GANs. As the next comparison method, we choose ILVR [5] which, like our method, performs super-resolution utilizing an unconditional pre-trained diffusion model. However, in ILVR, sampling happens at timestep $t$ rather than at the implicit step in our algorithm. In total, we utilize 300 images for evaluations. We present some qualitative results in Fig. 6. For ILVR [5], we utilize 100 timesteps of sampling, the same as in our case. As we can see, PULSE [22] and ILVR [5] are unable to restore the correct identity and also contain blur artifacts after restoration. On the other hand, steered diffusion (our method) is able to restore photorealistic facial", + "bbox": [ + 496, + 659, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20856", + "bbox": [ + 478, + 926, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c32960c85d337eea0ed8f47ad4c429c5a8fd92905cea08a253cc3f26dd7f8727.jpg", + "image_caption": [ + "Degraded" + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 169, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/602d5d8b76d6172338380de68d8a08d359761e2b0fe212f717fddf7b169ce31f.jpg", + "image_caption": [ + "Reinpaint" + ], + "image_footnote": [], + "bbox": [ + 171, + 88, + 256, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/802d6320a6cddebb523b7a0d87582acc75e8c6d72f5adbe7863ad3bfbfbede4e.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 258, + 88, + 341, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dfd65f5801bb9fc4af5f6616bee6ed63015fd6946eb1ec7bb29f3c0685e1daa1.jpg", + "image_caption": [ + "Degraded" + ], + "image_footnote": [], + "bbox": [ + 354, + 88, + 439, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/030f5d602bb1d81bf21a81ea1db0b0d79cf2c0b272484358566032d6c0096008.jpg", + "image_caption": [ + "Reinpaint" + ], + "image_footnote": [], + "bbox": [ + 441, + 88, + 524, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57b700bc89d80244a40413e5ff38da989002a58b0bb533458d34953ba1cb55a1.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 526, + 88, + 609, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/725dca099ff18d343af72eae1e33dab367deccaf1cf95a4de63e2c8192894d94.jpg", + "image_caption": [ + "Degraded" + ], + "image_footnote": [], + "bbox": [ + 624, + 88, + 707, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/813654a93234b9249024a775f851c410b488edf9211d53f6c9b140703d38e5b1.jpg", + "image_caption": [ + "Reinpaint" + ], + "image_footnote": [], + "bbox": [ + 709, + 88, + 792, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae4bd3a4ec785891155b8b122e4de4938853850135c8bb07a64f69e26ad3409d.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 794, + 88, + 877, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51aeb79b549932b0b45d635ce0b8c6c65877f3309c074c7aebda79c8910270bc.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 88, + 316, + 200, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7421b32858f0af6c759bce17f41674721722c5940f5f96c29daf50485b6caae1.jpg", + "image_caption": [ + "Figure 8. Qualitative comparisons for inpainting for thin, medium and thick masks.", + "Photo of a young man", + "Figure 9. Qualitative comparisons for text-based image editing." + ], + "image_footnote": [], + "bbox": [ + 202, + 316, + 312, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee5c2da4602c1a169b122944814d8dfb3ebe342cbf2a4421c958e0fab9b9907c.jpg", + "image_caption": [ + "She has wavy hair" + ], + "image_footnote": [], + "bbox": [ + 316, + 316, + 426, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3872586b42c28021953312560a0dfdce680576be43ea63ccacdd72d20c6d15b.jpg", + "image_caption": [ + "She has blonde hair" + ], + "image_footnote": [], + "bbox": [ + 429, + 316, + 539, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92f90481741da69482c90776ffe37526a062ab53b03b9b81e0c7e52012818997.jpg", + "image_caption": [ + "Photo of an old woman" + ], + "image_footnote": [], + "bbox": [ + 542, + 316, + 653, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/100df6cfac2b29c71d71547141a2ac171c4b8b3564748f8bced0ab66d51925da.jpg", + "image_caption": [ + "She is angry." + ], + "image_footnote": [], + "bbox": [ + 656, + 316, + 766, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de93655a1478376b937650bcf3dfb0024126ada5cde7324d976d4060020daa2a.jpg", + "image_caption": [ + "She is sad." + ], + "image_footnote": [], + "bbox": [ + 769, + 316, + 880, + 402 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/100e40b3af285c85e296194374b6bcabd540bf43c0c62b93f4f6e6e9fae2d751.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Trained MethodsInference-Time Methods
MethodCUT [16]GCGAN [9]CycleGAN [46]ILVR [5]Ours
FID (↓)49.3458.8052.70107.4641.38
mIoU(↑)0.6470.6990.7230.3770.532
", + "bbox": [ + 81, + 453, + 482, + 505 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/db49e1db65362909bebd5256b8646c98aec70afb8e515c75a94de035853b29b9.jpg", + "table_caption": [ + "Table 2. Quantitative results for semantic generation" + ], + "table_footnote": [], + "table_body": "
MethodFID ↓LPIPS↓NIQE↓PSNR↑SSIM↑
Bicubic130.30.441912.0323.850.642
ILVR [5]62.240.41647.3820.540.5527
PULSE [22]84.670.43655.0421.080.5285
Ours51.190.25939.2426.020.711
", + "bbox": [ + 101, + 539, + 447, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "images. The qualitative evaluations are presented in Table 3; our method yields a 0.18 improvement in perceptual similarity, 6.95 dB improvement in PSNR, and 0.24 improvement in SSIM versus all of the other comparison methods.", + "bbox": [ + 75, + 651, + 470, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Face Colorization", + "text_level": 1, + "bbox": [ + 76, + 724, + 248, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a baseline method, we modify ILVR [5] to suit the task of colorization. For this, rather than performing the constraint at every step, we start the sampling process from a noised grayscale image and enforce consistency between the generated and original grayscale images. In total, we utilize 300 images for evaluation. The corresponding results can be seen in Fig. 7. As we can see, our method is able to reconstruct photorealistic faces with naturalistic colours compared to ILVR [5]. The corresponding quantitative metrics are presented in Table. 4. We get a significant boost in", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a77a2fc208928cff06a693efbac9af639fd5c943d8bf7d2df89e6c871f50402c.jpg", + "table_caption": [ + "Table 3. Quantitative results for super-resolution" + ], + "table_footnote": [], + "table_body": "
MethodFID ↓LPIPS↓NIQE↓
Grayscale69.270.27815.07
ILVR [5]67.660.52707.54
Ours49.720.33115.91
", + "bbox": [ + 555, + 459, + 841, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative results for colorization", + "bbox": [ + 563, + 547, + 826, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performance, with an FID score of 19, LPIPS [44] score of 0.19, and NIQE [23] score of 1.5 for our method.", + "bbox": [ + 500, + 590, + 890, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Inpainting, Image Editing, Identity replication", + "text_level": 1, + "bbox": [ + 500, + 642, + 890, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method can also utilize multiple conditions simultaneously; we provide an illustration in Figure 9 where we condition with an identity-preserving network and a text caption simultaneously. From the figure, one can see that our method can generalize well to a diverse range of captions. For preserving identity, we used the VGGFace network [30]. We utilize FARL [45], which is pre-trained with face and corresponding text pairs to enforce the captions. For generic identity replication as in Figure 1 we use FARL face embedder.", + "bbox": [ + 496, + 669, + 893, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For our image inpainting experiments, we use the subset released by [43] and evaluate three different kinds of masks. Our method obtains better results than existing baselines across all mask variations. Qualitative and quantitative results are shown in Fig. 8 and Table 5, respectively.", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20857", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/52a4b99ca9b91a9892f50ff9737cc4928aa75d2a42abb215ad19dd8141fcd039.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodThinMediumThick
FID ↓LPIPS↓PSNR↑SSIM↑FID ↓LPIPS↓PSNR↑SSIM↑FID ↓LPIPS↓PSNR↑SSIM↑
Degraded371.030.6767.8470.201258.860.6247.5270.228231.930.5857.6060.249
Reinpaint [20]43.350.30417.990.67153.710.39913.710.55852.400.40712.780.530
Ours30.850.18324.920.83335.390.22021.550.78640.120.24218.870.705
", + "bbox": [ + 81, + 82, + 915, + 171 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5. Quantitative results for inpainting.", + "bbox": [ + 354, + 175, + 611, + 189 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/95dfe810af322fca7b4b1bac59f8a960198de272be940e59b7fcd3e232d7ba1b.jpg", + "image_caption": [ + "Semantic Map" + ], + "image_footnote": [], + "bbox": [ + 99, + 195, + 207, + 279 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6ef6bed403bad1cc2de2f0a937e2e1cc49054aedbbd015f769b5a7223479b43a.jpg", + "image_caption": [ + "$K = 0$" + ], + "image_footnote": [], + "bbox": [ + 210, + 196, + 316, + 279 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5a8da2c8b25ec356836003d0422438696a32eaf74994819f738bd615c3f14965.jpg", + "image_caption": [ + "$K = 200$" + ], + "image_footnote": [], + "bbox": [ + 320, + 196, + 426, + 279 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/39776518c35ced6dadd9567bbc7d63f30cd034ab6eb129a6bef2efd734b7e7c9.jpg", + "image_caption": [ + "$K = 2000$" + ], + "image_footnote": [], + "bbox": [ + 429, + 196, + 535, + 279 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9fe8229619d4dabac5ceb27c0dfe7c01e720f985c4e2d9bf8bcad14d34c4e437.jpg", + "image_caption": [ + "$K = 20000$" + ], + "image_footnote": [], + "bbox": [ + 537, + 196, + 645, + 279 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/45b50e4c730c9d0c9cc337074bdbdb5a545e74a5c8477d06b51b595a63c26a8d.jpg", + "image_caption": [ + "$K = 2\\times 10^{5}$" + ], + "image_footnote": [], + "bbox": [ + 645, + 196, + 754, + 279 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f3a11d08557b20d4c6b65b0bf11c871cf948fc5c452a16692fd262a6f45f9402.jpg", + "image_caption": [ + "$K = 2\\times 10^{6}$" + ], + "image_footnote": [], + "bbox": [ + 756, + 196, + 864, + 279 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/460407a823d71c159d1e69645223c962e949711ee7975ac5c707b604921e3286.jpg", + "image_caption": [ + "Figure 10. Figure showing sample variation with scaling factor $k(t) = K\\sqrt{1 - \\bar{\\alpha}_t}$" + ], + "image_footnote": [], + "bbox": [ + 133, + 315, + 223, + 386 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c28bc45cccfa5c4f281514b4728aa05ac4da72c8a66c2d0b7645905b390a7a20.jpg", + "image_caption": [ + "Degraded", + "Figure 11. Qualitative comparisons for Colorization." + ], + "image_footnote": [], + "bbox": [ + 133, + 386, + 223, + 455 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1958e5e661498a87cd4213ae5721fe08dde99c3bac20488d6167fbfc1961ccaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 315, + 316, + 386 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/21c2dd39d7514dc2a2819893e1eec540787b7b3010f1f6d08ef02c82810c38d8.jpg", + "image_caption": [ + "Noise robust" + ], + "image_footnote": [], + "bbox": [ + 225, + 386, + 316, + 455 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/aa819edcc972238ffeafe4650c8411ca86b50fabf1b570e7454ccceeebea724e.jpg", + "image_caption": [ + "OURS" + ], + "image_footnote": [], + "bbox": [ + 318, + 315, + 410, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.6. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 494, + 228, + 510 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effect of scaling factor $k(t)$ for semantic generation: In this section, we analyze how the scaling factor affects the quality of the sample in the case of complex conditioning of semantic generation. Fig. 10 shows the variation of sample quality starting from the same initial noise with different scaling factors $k(t) = K\\sqrt{1 - \\bar{\\alpha}_t}$ . The sample quality is bad for very low scaling factors, and for very high scaling factors, the diffusion process escapes the manifold of natural face images. We show the variation in sample quality for a fixed scaling factor versus a time-varying scaling factor in Fig. 12, which demonstrates that a time-varying scale factor produces more realistic samples. This is because the effective variance of the noise scheduling, which effectively controls the amount of regularization possible at a particular timestep, reduces as the generation process proceeds. Hence a larger tweak is permissible at the early steps of diffusion, and only very small tweaks are permitted in the later steps.", + "bbox": [ + 75, + 520, + 468, + 776 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Noise-Robust classifier: To validate the claims in section 3.1, we train a noise-robust inverse mapper for the task of colorization and show the output of the noise-robust classifier and the diffusion outputs for different noise levels in Fig. 11. The noise-robust classifier fails to preserve key details that our approach preserves.", + "bbox": [ + 75, + 777, + 470, + 868 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations Although our method can generalize to a wide series of tasks, one limitation that persists is the value of the", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/11599771e86de2750d03f850a0761e63cc549093484283f4e74ec94f72d0b1d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 315, + 650, + 392 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/581a0b3ad128f0c504ad2c9a07a89366a257c779b6f900e908f349b097659604.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 392, + 650, + 468 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/155d76b370e7d19f4906d3b1b93b541d94e9dda62caad6b49872d0a3e1630c38.jpg", + "image_caption": [ + "Figure 12. A comparison of a time-varying scaling factor with non-time-varying." + ], + "image_footnote": [], + "bbox": [ + 537, + 469, + 650, + 545 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/83c47f3b33a6f470e1141945df08cffba35fac5f75eee27e4eeea8a40ebafaa3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 315, + 750, + 392 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/39195c3e30d02512f8e9b3642dbdeeb5612ba3b384c6daac900f68cf3ed42dcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 392, + 750, + 468 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fdecf438744e32334487383ecf40e743a6844701d723eb75b80a6fba975bd5de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 469, + 751, + 545 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c4b6624c3e8a2a14fdddc4d53d8340e72512d53bb9d7171c984ec7731b3c085a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 315, + 849, + 392 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2e3e83c1181c3cf3c8db0f94bd524bd67bf99d6a09a57d8d06b6f457afbb496c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 392, + 849, + 468 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1b98a9b384664e4bfecbf32be1ceb4043f0a0758040b2a278c39a8c8280dca23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 469, + 849, + 545 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "scaling factor $k(t)$ . The value of $k(t)$ has to be empirically found based on the task, but once a few images are used to tune the value of $k(t)$ , the model generalizes well to other conditioning images of the same task. Like any other conditional generation models that can perform image editing, our method also has societal impacts, and care must be taken in applying these methods.", + "bbox": [ + 496, + 582, + 893, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 695, + 617, + 709 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose the first framework for plug- and-play conditional generation that can generalize well to both image-to-image translation tasks and label-based generation tasks. For this, we use the energy-based formulation of diffusion models and modulate the inference process using a task-specific predefined network or other preexisting function. Furthermore, we introduce a novel implicit sampling-based technique that improves the sampling quality across multiple tasks. We performed experiments on various tasks to show that our method can generalize across multiple tasks and outperforms existing methods that do not require additional training.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20858", + "bbox": [ + 478, + 926, + 517, + 936 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2", + "[3] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. arXiv preprint arXiv:2302.07121, 2023. 3", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2", + "[5] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 2, 3, 6, 7, 8", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6", + "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 4, 6", + "[9] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, Kun Zhang, and Dacheng Tao. Geometry-Consistent Generative Adversarial Networks for One-Sided Unsupervised Domain Mapping. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3, 8", + "[10] Ruiqi Gao, Yang Song, Ben Poole, Ying Nian Wu, and Diederik P Kingma. Learning energy-based models by diffusion recovery likelihood. arXiv preprint arXiv:2012.08125, 2020. 2", + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[12] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. arXiv preprint arXiv:2206.09012, 2022. 2, 3", + "[13] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. arXiv preprint arXiv:1912.03263, 2019. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3", + "[15] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts gans. arXiv preprint arXiv:2112.05130, 2021. 2", + "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3, 8", + "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6", + "[18] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793, 2022. 3", + "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), December 2015. 7", + "[20] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 3, 6, 9", + "[21] Kangfu Mei, Nithin Gopalakrishnan Nair, and Vishal M Patel. Bi-noising diffusion: Towards conditional diffusion models with generative restoration priors. arXiv preprint arXiv:2212.07352, 2022. 2", + "[22] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2437-2445, 2020. 7, 8", + "[23] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 8", + "[24] Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, and Vishal M Patel. Unite and conquer: Plug & play multimodal synthesis using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6070-6079, 2023. 2, 3", + "[25] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4467-4477, 2017. 2, 3", + "[26] Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks. Advances in neural information processing systems, 29, 2016. 2", + "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20859", + "bbox": [ + 478, + 926, + 517, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 7", + "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 2, 4", + "[30] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In *British Machine Vision Conference*, 2015. 7, 8", + "[31] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 7", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3", + "[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2", + "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2", + "[37] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021. 3", + "[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3", + "[39] Ken Sekimoto. Langevin equation and thermodynamics. Progress of Theoretical Physics Supplement, 130:17-27, 1998. 4", + "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5", + "[42] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2022. 3", + "[43] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 8", + "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8", + "[45] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. arXiv preprint arXiv:2112.03109, 2021. 7, 8", + "[46] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In IEEE International Conference on Computer Vision (ICCV), 2017. 3, 7, 8" + ], + "bbox": [ + 501, + 92, + 893, + 458 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20860", + "bbox": [ + 478, + 926, + 519, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/396c3711-d204-4df2-8e69-9bdf93d4fbff_model.json b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/396c3711-d204-4df2-8e69-9bdf93d4fbff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..db7e7a9531f862d16b67613034b1f8b8079381b2 --- /dev/null +++ b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/396c3711-d204-4df2-8e69-9bdf93d4fbff_model.json @@ -0,0 +1,3368 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.116, + 0.809, + 0.16 + ], + "angle": 0, + "content": "Steered Diffusion: A Generalized Framework for Plug-and-Play Conditional Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.17, + 0.787, + 0.205 + ], + "angle": 0, + "content": "Nithin Gopalakrishnan Nair\\(^{1*}\\) Anoop Cherian\\(^{2}\\) Suhas Lohit\\(^{2}\\) Ye Wang\\(^{2}\\) Toshiaki Koike-Akino\\(^{2}\\) Vishal M. Patel\\(^{1}\\) Tim K. Marks\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.208, + 0.732, + 0.224 + ], + "angle": 0, + "content": "\\(^{1}\\) Johns Hopkins University \\(^{2}\\) Mitsubishi Electric Research Laboratories (MERL)" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.228, + 0.745, + 0.241 + ], + "angle": 0, + "content": "{ngopala2,vpatel36}@jhu.edu {acherian,slohit,yewang,koike,tmarks}@merl.com" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.245, + 0.662, + 0.258 + ], + "angle": 0, + "content": "https://merl.com/demos/steered-diffusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.269, + 0.232, + 0.281 + ], + "angle": 0, + "content": "(a) Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.285, + 0.167, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.416, + 0.149, + 0.427 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.285, + 0.257, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.416, + 0.313, + 0.427 + ], + "angle": 0, + "content": "Generated Samples" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.285, + 0.344, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.269, + 0.514, + 0.28 + ], + "angle": 0, + "content": "(b) Colorization" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.284, + 0.441, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.416, + 0.422, + 0.427 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.284, + 0.532, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.416, + 0.586, + 0.427 + ], + "angle": 0, + "content": "Generated Samples" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.269, + 0.718, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.416, + 0.696, + 0.427 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.269, + 0.816, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.416, + 0.859, + 0.427 + ], + "angle": 0, + "content": "Generated Samples" + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.269, + 0.895, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.434, + 0.82, + 0.446 + ], + "angle": 0, + "content": "(f) Text-based editing" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.433, + 0.167, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.516, + 0.166, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.581, + 0.149, + 0.592 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.449, + 0.257, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.516, + 0.257, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.581, + 0.313, + 0.592 + ], + "angle": 0, + "content": "Generated Samples" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.449, + 0.344, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.516, + 0.343, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.449, + 0.441, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.516, + 0.441, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.581, + 0.422, + 0.592 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.449, + 0.532, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.516, + 0.533, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.581, + 0.586, + 0.592 + ], + "angle": 0, + "content": "Generated Samples" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.45, + 0.716, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.516, + 0.716, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.581, + 0.696, + 0.592 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.45, + 0.808, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.516, + 0.808, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.581, + 0.859, + 0.592 + ], + "angle": 0, + "content": "Generated Samples" + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.45, + 0.895, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.516, + 0.895, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.598, + 0.893, + 0.655 + ], + "angle": 0, + "content": "Figure 1. An illustration of various applications of our method. We use a diffusion model trained unconditionally and condition using our proposed algorithm only during the test time. We present the results on six tasks: (a) image inpainting, (b) colorization, (c) image super-resolution, (d) semantic generation, (e) identity replication, and (f) text-based image editing. In part (f), the text prompts for the first and second columns, respectively, are \"This person has blonde hair\" and \"This person has wavy hair.\"" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.667, + 0.34, + 0.682 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Conditional generative models typically demand large annotated training sets to achieve high-quality synthesis. As a result, there has been significant interest in designing models that perform plug-and-play generation, i.e., to use a predefined or pretrained model, which is not explicitly trained on the generative task, to guide the generative process (e.g., using language). However, such guidance is typically useful only towards synthesizing high-level semantics rather than editing fine-grained details as in image-to-image translation tasks. To this end, and capitalizing on the powerful fine-grained generative control offered by the recent diffusion-based generative" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.668, + 0.895, + 0.895 + ], + "angle": 0, + "content": "models, we introduce Steered Diffusion, a generalized framework for photorealistic zero-shot conditional image generation using a diffusion model trained for unconditional generation. The key idea is to steer the image generation of the diffusion model at inference time via designing a loss using a pre-trained inverse model that characterizes the conditional task. This loss modulates the sampling trajectory of the diffusion process. Our framework allows for easy incorporation of multiple conditions during inference. We present experiments using steered diffusion on several tasks including inpainting, colorization, text-guided semantic editing, and image super-resolution. Our results demonstrate clear qualitative and quantitative improvements over state-of-the-art diffusion-based plug-and-play models while adding negligible additional computational cost." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.32, + 0.901 + ], + "angle": 0, + "content": "* Work done during internship at MERL." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20850" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.208, + 0.105 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.356 + ], + "angle": 0, + "content": "Deep diffusion-based probabilistic generative models [8, 14, 40] are quickly emerging as one of the most powerful methods to synthesize high-quality content and have shown the potential to revolutionize content creation not only in computer vision, but also in many other areas including speech, audio, and language. Such models (e.g., ImagGen [36], Stable Diffusion [34]) have demonstrated outstanding synthesis results in conditional generation tasks, such as text-conditioned image synthesis [2, 33] and image reconstruction [31, 35, 38]. However, these models do not typically possess zero-shot conditional generative abilities when used directly (zero-shot capabilities as are commonly seen in language foundation models such as GPT-3 [4]), and often demand large amounts of annotated and paired (multimodal) data for conditional generation, which may be challenging to obtain [15]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.357, + 0.471, + 0.825 + ], + "angle": 0, + "content": "One way to circumvent this need for large annotated training sets is to leverage predefined models as plug-and-play modules [12, 24, 25] in an otherwise unconditionally trained diffusion model. Specifically, in such plug-and-play models, a model is first trained in an unconditional setting (without labels). During inference, the plug-and-play modules (networks separately trained for a particular conditional task, e.g., image captioning) are incorporated in the reverse diffusion process to produce intermediate samples guided in the Markov chain in specific directions to satisfy the desired condition. Prior works, such as [25, 26], have proposed similar methods in which the authors derive text- or class-conditioned samples from Generative Adversarial Networks (GANs) [11] that were trained without labels. To achieve this, they iteratively refine the noise input of the GAN until the desired sample satisfies the condition. Very recently, Grakios et al. [12] proposed a diffusion-based plug-and-play method that enables using unconditional diffusion models for conditional generation utilizing class labels. Both these methods are specifically designed for tasks involving label-level semantics. However, these methods do not address the usage of unconditional models for general image-to-image translation tasks, which require synthesizing visual content conditioned on fine-grained details in the source image. There are also works that propose diffusion models for image-to-image translation, such as for image super-resolution and inpainting [5, 20]; however, these methods are task-specific and do not generalize well to new tasks or new types of inverse problems (as demonstrated in Section 5). In this work, we present a generic framework that can generalize to any image-to-image translation task." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this paper, we derive the necessary theory and formulate an algorithm, which we call Steered Diffusion, for diffusion-based image editing and image-to-image translation; our model is subsequently validated on a wide range of tasks. Steered diffusion is motivated by the energy-based" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.891, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.284, + 0.893, + 0.41 + ], + "angle": 0, + "content": "Figure 2. An illustration of the difference between existing plug-and-play generation approaches (e.g., [12]) and the proposed approach. Existing plug-and-play works operate with an energy function \\( V \\) of the noisy latent \\( x_{t} \\). In contrast, our model uses the implicit prediction of the diffusion model (i.e., a coarse estimate of the clean image \\( x_{0} \\)) in its energy function \\( V_{1} \\), which allows the use of any pre-trained network for steering. In addition, our model provides a looping mechanism \\( V_{2} \\), which iterates \\( N \\) times at each timestep \\( t \\) to enhance generation quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.895, + 0.819 + ], + "angle": 0, + "content": "formulation of diffusion probabilistic models [10]. In general, inference in a generative model can be thought of as deriving samples from a learned distribution. Recall that every probability density function can be formulated as an energy field that describes an unnormalized estimate of how the distribution density varies in space [13,25]. If one needs to find points in space that are the closest match to a given condition, one can utilize gradient-based optimization algorithms to find points in the field that have the highest density value for the condition. The gradient-based optimization scheme can be viewed as a modulation of the energy toward the desired direction. Previous work has utilized this idea on GANs [25, 26] and obtained reasonable results for label-based generation tasks. Due to their model structure, diffusion models are ideal candidates for such an energy modulation. One key challenge remains to design a good energy estimator that is robust to all noise levels. Previously, classifier-based guidance [8, 27] has been proposed and thought of as an energy modulation utilizing a pretrained classifier trained on noisy images. This poses a limitation that the guiding function should be noise-robust. In this work, we propose an alternative solution that does not need noise-robust networks but could use any network by utilizing the diffusion model as an implicit denoiser. Figure 2 gives a brief overview of how our approach is different from existing methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We present experiments using steered diffusion on multiple conditional generative tasks on faces as well as generic images as portrayed in Figure 1. We present results on (i) identity replication [7], (ii) semantic image generation [29], (iii) linear inverse problems [21], and (iv) text-conditioned" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "20851" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "image editing. Although our method is generic, for evaluations we perform experiments on faces. Before presenting our framework in detail, we now summarize the key contributions of our work:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.161, + 0.469, + 0.204 + ], + "angle": 0, + "content": "- We propose steered diffusion, a general plug-and-play framework that can utilize various pre-existing models to steer an unconditional diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.206, + 0.469, + 0.251 + ], + "angle": 0, + "content": "- We present the first work applicable to both label-level synthesis and image-to-image translation tasks and demonstrate its effectiveness for various applications." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.252, + 0.469, + 0.311 + ], + "angle": 0, + "content": "- We propose an implicit conditioning-based sampling strategy that significantly boosts the performance of conditional sampling from unconditional diffusion models compared with previous methods." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.312, + 0.469, + 0.342 + ], + "angle": 0, + "content": "- We introduce a new strategy that uses multiple steps of projected gradient descent to improve sample quality." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.161, + 0.469, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.355, + 0.206, + 0.372 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.379, + 0.221, + 0.394 + ], + "angle": 0, + "content": "2.1. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.569 + ], + "angle": 0, + "content": "Early works on unpaired image-to-image translation utilize a cycle consistency loss between the input and the target domains [9, 46]. Newer works, such as [16], have introduced a contrastive learning-based approach where a contrastive loss between corresponding patches of the input and target domains is minimized. The consistency-based method often fails to generate photorealistic images; hence, conditional generative models are preferred when labeled data are available. A few works [37, 38] utilize diffusion models for conditional image-to-image translation because of their photorealistic generation quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Guiding diffusion models during inference time has been explored by several works, such as [24]. The first method that proposed inference-time conditioning [8] uses a pretrained noise-robust classifier to guide the inference of an unconditional model. GLIDE [27] proposed a method for conditioning using text. Earlier work in plug-and-play modelling for generative models utilized GANs and performed iterative refinement on the latent space of GANs [25]. This method uses a predefined classifier or text captioning network to estimate a loss between the desired label output or text caption and the one generated from the GAN generator. This loss is backpropagated to refine the noise input of the GAN iteratively until the generator predicts the desired output. Recently, [12] proposed a method that uses diffusion models as a plug-and-play prior for class-conditioned generation. Several works have addressed the task of image-to-image translation using unconditional diffusion models [1,5,18,20], but each of these proposes a task-specific inference scheme. For example, ILVR [5] performs image super-resolution, and Reinpaint [20] performs image inpainting. Blended diffusion [1] proposes a method for text-conditioned image editing. DDRM [18] proposes an inference-time scheme" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "offering a general solution for linear inverse problems such as colorization and super-resolution." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.131, + 0.673, + 0.145 + ], + "angle": 0, + "content": "2.2. Concurrent Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.894, + 0.32 + ], + "angle": 0, + "content": "In concurrent work to ours that also explores zero-shot conditional generation using diffusion models, [3] used a text to image model [34] and a two-step forward and backward universal guidance process, but it works well only after heavy optimization on network-based inverse problems such as semantic generation and identity generation. Another concurrent work [42] explored the usage of unconditional diffusion models for linear inverse problems using a pseudo-inverse model. In contrast to all these prior works, our steered diffusion algorithm generalizes well to both image-to-image translation tasks and high-level label-based generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.33, + 0.85, + 0.345 + ], + "angle": 0, + "content": "2.3. Denoising Diffusion Probabilistic Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Denoising diffusion probabilistic models (DDPMs) [14, 40] belong to a class of generative models in which the model learns the distribution of data through a Markovian sampling process. DDPMs consist of a forward process and a reverse process. Let \\( x_{t} \\) denote the latent state of an input image at timestep \\( t \\) in a diffusion process. The sampling operation \\( q(\\cdot) \\) for the forward process in DDPM is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.469, + 0.892, + 0.488 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) := \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.891, + 0.528 + ], + "angle": 0, + "content": "where \\(\\{\\beta_t\\}\\) is a predefined variance schedule and \\(I\\) is the identity matrix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.894, + 0.604 + ], + "angle": 0, + "content": "The forward process can be considered as a noising operation, where the next state \\( x_{t} \\) is obtained from the current state \\( x_{t-1} \\) by adding a small amount of Gaussian noise according to the sampled timestep. The state \\( x_{t} \\) at timestep \\( t \\) can also be sampled directly from the initial state \\( x_{0} \\), using:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.615, + 0.892, + 0.633 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {0}\\right) := \\mathcal {N} \\left(x _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) I\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.643, + 0.606, + 0.658 + ], + "angle": 0, + "content": "or equivalently," + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.668, + 0.892, + 0.686 + ], + "angle": 0, + "content": "\\[\nx _ {t} = x _ {0} \\sqrt {\\bar {\\alpha} _ {t}} + \\epsilon \\sqrt {1 - \\bar {\\alpha} _ {t}}, \\quad \\epsilon \\sim \\mathcal {N} (0, I), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.761, + 0.712 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s\\) and \\(\\alpha_t = 1 - \\beta_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.817 + ], + "angle": 0, + "content": "In [40], it is shown that if the number of time steps is large and the increment in \\(\\{\\beta_i\\}\\) is small, then each step in the reverse sampling process can also be approximated by a Gaussian. If \\(\\mu_{\\theta}\\) and \\(\\Sigma_{\\theta}\\) respectively denote the mean and the covariance of this Gaussian, modelled via neural networks with parameters \\(\\theta\\), then each reverse step samples the state \\(x_{t-1}\\) according to:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.828, + 0.892, + 0.846 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) := \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.899 + ], + "angle": 0, + "content": "The parameters \\(\\theta\\) are obtained by minimizing the variational lower bound of the negative log-likelihood of the data distribution." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "20852" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.589, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.089, + 0.875, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.895, + 0.31 + ], + "angle": 0, + "content": "Figure 3. An illustration of Steered Diffusion. During each step of the sampling process the implicit prediction is steered to the direction of the condition using a steering network or predefined function. Note that this figure is only for illustrating the idea and does not show the actual sampled images; in the actual sampling, the steering process is much more gradual, not sudden as portrayed in this image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.315, + 0.251, + 0.332 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.34, + 0.388, + 0.357 + ], + "angle": 0, + "content": "3.1. Steered Diffusion at Inference Time" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.473, + 0.41 + ], + "angle": 0, + "content": "Our work is motivated by the energy-based formulation of diffusion models. For any probability density function, the corresponding energy-based model (EBM) is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.43, + 0.47, + 0.462 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (x) = \\frac {\\exp (- V (x))}{Z}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.562 + ], + "angle": 0, + "content": "where \\( V(x) \\) denotes the corresponding energy function across states \\( x \\), and \\( Z \\) denotes a normalization constant. To derive samples from this distribution, one can utilize the Langevin equation [39] describing the state transition of a particle in the presence of an energy field. For diffusion models the sampling step is" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.586, + 0.47, + 0.603 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = x _ {t} - \\nabla_ {x _ {t}} \\log p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) + \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.473, + 0.81 + ], + "angle": 0, + "content": "The term \\(\\nabla_{x_t}\\log p_\\theta (x_{t - 1}|x_t)\\) is called the score function of the density \\(p_{\\theta}(x_{t - 1}|x_t)\\). One key advantage of the energy-based formulation is that it allows modulation of the energy function to satisfy given criteria. This was initially introduced as classifier guidance [8], which allows label conditional sampling from an unconditionally trained diffusion model utilizing a noise-robust classifier. In the remaining part of this section, we motivate how we can extend the functionality of unconditional diffusion models to conditional tasks. Consider a conditional sampling scenario based on a condition \\(c\\), for sampling from a state \\(x_{t}\\) to state \\(x_{t - 1}\\). The conditional transition probability \\(p_{\\theta}(x_{t - 1}|x_t,c)\\) can be decomposed as" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.83, + 0.47, + 0.864 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, c\\right) \\propto \\frac {p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) p \\left(c \\mid x _ {t - 1}\\right)}{p \\left(c \\mid x _ {t}\\right)}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Hence, for any timestep \\( t \\), the effective score for conditional transition can be found by utilizing using the log" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.317, + 0.892, + 0.348 + ], + "angle": 0, + "content": "of probability density in the EBM formulations (5) of the individual densities and can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.357, + 0.68, + 0.374 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t - 1} | x _ {t}, c) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.376, + 0.892, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {x _ {t}} \\log p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) - \\nabla_ {x _ {t}} V _ {1} \\left(x _ {t}, c\\right) + \\nabla_ {x _ {t}} V _ {2} \\left(x _ {t - 1}, c\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.895, + 0.553 + ], + "angle": 0, + "content": "where \\( V_{1} \\) and \\( V_{2} \\) are the corresponding energy functions that model the conditional distributions of \\( x_{t} \\) and \\( x_{t-1} \\) given a condition \\( c \\). Specifically, they project the higher dimensional \\( x_{t} \\) to the lower dimensional space of \\( c \\) and measure the distance between the mapped value and \\( c \\). The better this particular measure, the more effectively it can be used to generate conditional samples from an unconditional model. Using Eq. (8), the conditional sampling equation for the reverse process is" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.562, + 0.892, + 0.616 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} ^ {(t)} (x _ {t}, t)\\right) \\\\ - \\nabla_ {x _ {t}} V _ {1} (x _ {t}, c) + \\nabla_ {x _ {t}} V _ {2} (x _ {t - 1}, c) + \\sigma_ {t} \\epsilon . \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.894, + 0.69 + ], + "angle": 0, + "content": "Here \\(\\epsilon_{\\theta}^{(t)}(x_t,t)\\) is the network prediction at a timestep \\(t\\) and \\(\\sigma_{t}\\) is the corresponding variance of the reverse step. The formulation in Eq. (9) shows that the energy function requires a functional mapping from a noisy \\(x_{t}\\) to \\(c\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In many applications of interest, the mapping function from \\( x_{t} \\) to \\( c \\) is complex and can be modeled effectively using deep networks. For example, in image-to-image translation tasks such as image generation from semantic maps [29], the image is \\( x_{t} \\) and the semantic map is the condition \\( c \\). Similarly, for text to image generation, the image is \\( x_{t} \\) and \\( c \\) corresponds to the text. Ideally, we would like to employ an existing (pre-trained) deep neural network for the mapping from \\( x_{t} \\) to \\( c \\). However, deep networks are usually trained on clean images, which limits the usability of existing pretrained deep networks for mapping directly from the noisy image \\( x_{t} \\) to \\( c \\). One workaround would be to use noise-robust networks to map from \\( x_{t} \\) to \\( c \\), but training noise-robust networks for conditional mapping can be computationally" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "20853" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.287 + ], + "angle": 0, + "content": "expensive. Moreover, a network that is trained with multiple different noise levels often results in lower mapping performance, as it cannot denoise all noise levels accurately; we validate this claim experimentally in Section 5.6. Alternatively, one could include two mapping functions: a first that denoises \\( x_{t} \\), and a second that maps from the denoised image to \\( c \\). Rather than training a seperate denoising network, however, we realized that diffusion models are inherently trained as denoisers, and reconstruction quality improves as time proceeds in the reverse sampling of the diffusion process. Because of this capacity, we can use a reverse sampling step to make coarse predictions of the denoised image from any time step \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.289, + 0.472, + 0.305 + ], + "angle": 0, + "content": "Hence, we modify our original energy expression (8) to:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.332, + 0.471, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t - 1} | x _ {t}, c) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t - 1} | x _ {t}) \\\\ - \\nabla_ {x _ {t}} V _ {1} \\left(x _ {0 \\mid t}, c\\right) - \\delta_ {1} + \\nabla_ {x _ {t}} V _ {2} \\left(x _ {0 \\mid t - 1}, c\\right) + \\delta_ {2}, \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.424, + 0.411 + ], + "angle": 0, + "content": "where, we define the implicit step prediction \\(x_{0|t}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.424, + 0.47, + 0.462 + ], + "angle": 0, + "content": "\\[\nx _ {0 \\mid t} = \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (x _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.472, + 0.47, + 0.547 + ], + "angle": 0, + "content": "Here, we assume \\( x_{t} \\) and \\( x_{t-1} \\) are first denoised to \\( x_{0|t} \\) and \\( x_{0|t-1} \\), respectively. The terms \\( \\delta_{1} \\) and \\( \\delta_{2} \\) capture the errors arising from the shift in the domain from \\( x_{t} \\) to \\( x_{0} \\) and from \\( x_{t-1} \\) to \\( x_{0} \\); for large \\( t \\), \\( \\delta_{1} \\approx \\delta_{2} \\) as the implicit predictions at nearby steps tend to be similar." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.473, + 0.775 + ], + "angle": 0, + "content": "As shown in our experiments, the energy function should be selected according to the task. An easy way to choose the energy function is by looking at the training loss of the mapping network. For example, in the case of semantic generation, a good energy function is the cross-entropy loss between the predicted semantic map at any timestep and the input semantic map. In the case of identity replication, a good choice of regularization would be the negative cosine similarity score between the embeddings from a recognition network for the input and target image. In the case of text-to-image generation, it would be CLIP loss [32]. As a rule of thumb, an energy function could be chosen easily by looking at the loss function used to train the pre-trained network (or an inverse function) that maps from the image \\( x \\) to the condition \\( c \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.424, + 0.803 + ], + "angle": 0, + "content": "3.2. Revisiting Sampling in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To obtain a closed-form expression for plugging the energy-based formulation into the reverse sampling process efficiently, we take inspiration from DDIM [41] and revisit the reverse sampling operation of diffusion models. From \\( p_{\\theta}(x_{1:T}) \\), one can generate a sample \\( x_{t-1} \\) from a sample \\( x_{t} \\) by:" + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.71, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Steered Diffusion" + }, + { + "type": "algorithm", + "bbox": [ + 0.5, + 0.108, + 0.864, + 0.35 + ], + "angle": 0, + "content": "Input: Energy function \\(V\\), condition \\(c\\) \n1: \\(x_{T} \\sim \\mathcal{N}(x_{T}; 0, I)\\) \n2: for \\(t = T - 1, \\ldots, 1\\) do \n3: for \\(n = N, \\ldots, 1\\) do \n4: \\(\\epsilon \\sim \\mathcal{N}(\\epsilon; 0, I)\\) \n5: \\(x_{0|t} = \\frac{x_{t} - \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon_{\\theta}^{(t)}(x_{t})}{\\sqrt{\\bar{\\alpha}_{t}}}\\) \n6: Compute \\(x_{0|t}^{\\text{feas}}\\) with \\(V, c\\) using Eq. (15) \n7: if \\((n > 1)\\) then \n8: Compute \\(x_{t}^{uc}\\) using Eq. (13) \n9: else \n10: Compute \\(x_{t-1}^{uc}\\) using Eq. (13) \n11: end if \n12: end for \n13: end for \n14: return \\(x_{0}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.397, + 0.893, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t - 1} ^ {\\mathrm {u c}} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\cdot \\underbrace {\\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (x _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}}} _ {\\text {‘ p r e d i c t e d x _ {0} ”}} \\\\ + \\underbrace {\\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\cdot \\epsilon_ {\\theta} ^ {(t)} \\left(x _ {t}\\right)} _ {\\text {d i r e c t i o n p o i n t i n g t o} x _ {t} ”} + \\underbrace {\\sigma_ {t} \\epsilon} _ {\\text {r a n d o m n o i s e}}, \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.509, + 0.892, + 0.541 + ], + "angle": 0, + "content": "as in Song et al. [41]. Using Eq. (11), we can rewrite the unconditional sampling step Eq. (12) in terms of \\( x_{0|t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.548, + 0.893, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t - 1} ^ {\\mathrm {u c}} = \\sqrt {\\bar {\\alpha} _ {t - 1}} x _ {0 | t} + \\\\ \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\cdot \\frac {x _ {t} - \\sqrt {\\bar {\\alpha} _ {t}} x _ {0 | t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} + \\sigma_ {t} \\epsilon , \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.61, + 0.892, + 0.669 + ], + "angle": 0, + "content": "Here the superscript uc denotes the unconditional sample, which is obtained without any steering while transitioning from \\( x_{t} \\) to \\( x_{t-1} \\). The conditional sampling step (8) can then be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.679, + 0.892, + 0.697 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = x _ {t - 1} ^ {\\mathrm {u c}} - \\nabla_ {x _ {t}} V _ {1} \\left(x _ {0 \\mid t}, c\\right) + \\nabla_ {x _ {t}} V _ {2} \\left(x _ {0 \\mid t - 1}, c\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Through this, we can modulate \\( x_0 \\) directly, as \\( x_{0|t} \\) is also a function of \\( \\epsilon_{\\theta}^{(t)}(x_t) \\). Following Eq. (14), a rough estimate of the desired \\( x_0 \\) for conditional sampling, represented by \\( x_{0|t}^{\\mathrm{feas}} \\), can be obtained using" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.781, + 0.892, + 0.815 + ], + "angle": 0, + "content": "\\[\nx _ {0 | t} ^ {\\text {f e a s}} = x _ {0 | t} - k (t) \\nabla_ {x _ {t}} \\left(V _ {1} \\left(x _ {0 | t}, c\\right) - V _ {2} \\left(x _ {0 | t - 1}, c\\right)\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.902 + ], + "angle": 0, + "content": "where \\( k(t) \\) is a scaling factor defining the strength of regularization. We call the process of finding \\( x_{0|t}^{\\mathrm{feas}} \\) from \\( x_{0|t} \\) Implicit Steering Control (ISC), and we call the new sampling process steered diffusion. The exact algorithm is illustrated in Fig. 3 and explained in Algorithm 1." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.52, + 0.938 + ], + "angle": 0, + "content": "20854" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.089, + 0.396, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.471, + 0.311 + ], + "angle": 0, + "content": "Figure 4. An illustration of the steering function for linear inverse problems. For linear inverse problems, the component of the implicit prediction along the degradation direction can be replaced by the ground truth condition." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.32, + 0.367, + 0.337 + ], + "angle": 0, + "content": "4. Tips for Improved Performance" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.345, + 0.303, + 0.36 + ], + "angle": 0, + "content": "4.1. Linear Inverse Problems" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.471, + 0.566 + ], + "angle": 0, + "content": "For optimization-based inverse problems such as text-to-image generation and semantic map to natural image generation, the exact mapping function is not always available. On the other hand, for linear inverse problems such as colorization, super-resolution, and image inpainting, the mapping is simply a linear function, and we can write Eq. (15) more simply. In these cases, the exact mapping function to the latent space of the condition is known. Hence, one can decompose the implicit prediction at each timestep along the direction of the condition and simply replace this component by its desired ideal condition, i.e., if our predicted sample needs to map to a condition \\( c \\), then the modified implicit prediction step becomes" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.577, + 0.47, + 0.612 + ], + "angle": 0, + "content": "\\[\nx _ {0 | t} ^ {\\text {f e a s}} = x _ {0 | t} + k (t) \\left(D (y) - D \\left(x _ {0 | t}\\right)\\right), \\quad \\text {w h e r e} c = D (y). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.701 + ], + "angle": 0, + "content": "Here, \\( y \\) is the clean image and \\( D \\) is the known degradation model. Our sampling procedure ensures that the series of operations preserve the consistency of domains of \\( x_{t-1} \\) and \\( x_{0|t}^{\\mathrm{feas}} \\) to the original data distribution at the corresponding timesteps. An illustration is shown in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.356, + 0.727 + ], + "angle": 0, + "content": "4.2. Multi-Step Implicit Modulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Our experiments show (see Fig. 7) that performing refinement using Eq. (15) on the implicit step prediction multiple times for each timestep significantly boosts the conditioning quality for more ill-posed conditions such as image inpainting and colorization. A similar observation was also found by [20]. Specifically, at a particular timestep \\( t \\), we iterate the procedure of steering towards the next sampling step \\( x_{t - 1} \\) and then adding noise to return to \\( x_{t} \\). We present the corresponding algorithm in Algorithm 1. An example is shown in Fig. 7, where more realistic images are generated using the multiple-step sampling scheme row labeled \"OURS multi.\"" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Effectively, the \\( V_{2} \\) term in Eq. (12) can be thought of as enabling a multistep sampling in which we modulate the current step by looking ahead to the next sampling step. On a careful analysis of Eq. (7), i.e. from the score contribution due to the different regularization functions, we can see that the term for \\( \\nabla_{x_t}V_1(x_t,c) \\) modulates \\( x_{t} \\) based on its current state, and the term from \\( \\nabla_{x_t}V_2(x_{t - 1},c) \\) is a look-ahead correction where the derivative with respect to the future prediction is found. This is exactly what happens in the case of looping back from \\( x_{t - 1} \\) to \\( x_{t} \\), where \\( x_{t} \\) is modulated iteratively by looking forward to what the future prediction would be." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.281, + 0.783, + 0.297 + ], + "angle": 0, + "content": "4.3. Choosing the Scaling Factor \\( k(t) \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.895, + 0.456 + ], + "angle": 0, + "content": "In Eq. (15), \\( k(t) \\) denotes the strength of the regularization constraint. A very small \\( k(t) \\) would denote no effective regularization, and a large \\( k \\) would lead to the diffusion process going out of the latent space manifold. Since the derivative of the regularization function by itself is a score value, similar to the normal scaling value of the score function, the appropriate time-varying normalization factor is \\( \\sqrt{1 - \\bar{\\alpha}_t} \\). The exact value for \\( k(t) \\) for each task is defined in Table 1. For linear inverse problems we use a constant \\( k(t) = 1 \\), which provided the best results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.467, + 0.633, + 0.484 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.895, + 0.795 + ], + "angle": 0, + "content": "We evaluate the performance of our network qualitatively and quantitatively using four image-to-image translation tasks—semantic layout to face image translation, face inpainting, face colorization, and face super-resolution—as well as two high-level vision tasks: identity-based image generation and text-guided image editing. Unlike existing approaches, our method is completely zero-shot and applies to a wide variety of tasks. We compare with other diffusion-based approaches best applicable for each task for a fair evaluation. We also compare the semantic layout to image translation performance with that of task-specific unsupervised methods. We choose the unconditional model released by [5] as the unconditional pretrained diffusion model in all of our experiments with faces. Note that the sampling scheme in ILVR [5] and Repaint [20] can be thought of as happening at time \\( t \\) rather than at the implicit step as in our method. Hence, comparing these methods for super-resolution and inpainting in our experiments below can be considered an additional ablation study highlighting the improvement from our implicit sampling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.718, + 0.818 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "For our experiments, we utilize pixel-level unconditional diffusion models. For faces, we utilize the model trained on the FFHQ dataset [17] that was released by [5]. For generic images, we use the model trained on ImageNet [6] released in ADM [8]. All our experiments use 100 steps of sampling." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "20855" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.082, + 0.88, + 0.176 + ], + "angle": 0, + "content": "
MethodLinear inverse problemsComplex inverse problems
ColorizationInpaintingSuper-resolutionSemantic GenerationIdentity ReplicationImage Editing
V1(D(x0|t) - c)2(D(x0|t) - c)2(D(x0|t) - c)2BCE(D(x0),c)1 - D(x0|t).D(c)3(D1(x0|t),D1(c1))2 + 2000CS(D2(x0|t),c2)
V2(D(x0|t-1) - c)2(D(x0|t-1) - c)20000
DGrayscaleMaskingDownscalingFARL [45]FARL recognizer [45]VGG Face [30] & FARL CLIP [32]
N331111
k(t)1112000√1-αt3000√1-αt10√1-αt & 1500√1-αt
" + }, + { + "type": "table_caption", + "bbox": [ + 0.355, + 0.181, + 0.614, + 0.195 + ], + "angle": 0, + "content": "Table 1. Parameter set for each application." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.218, + 0.466, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.408, + 0.447, + 0.422 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons for semantic generation." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.432, + 0.475, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.673, + 0.396, + 0.687 + ], + "angle": 0, + "content": "Figure 6. Results on \\(8 \\times\\) super resolution." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.315, + 0.709 + ], + "angle": 0, + "content": "5.2. Semantic Face Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.471, + 0.868 + ], + "angle": 0, + "content": "To evaluate how our method performs in generic image-to-image translation tasks, we evaluate our method's performance for the task of semantic layout to face generation. We utilize the CelebA dataset for this. To generate the semantic labels, we use facer [45] and create 11 label classes for each face. Since there is no other unconditional model that can perform fully test-time semantic generation, to evaluate the performance of our method, we compare with fully unsupervised image translation methods: CycleGAN [46], CUT [28], and ILVR [5]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The corresponding qualitative results are shown in Fig. 5. It is clear that CUT, CycleGAN, and ILVR produce unrealis-" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.218, + 0.901, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.467, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons for colorization. The row labeled \"OURS multi\" refers to the use of multi-step sampling, as described in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.53, + 0.894, + 0.621 + ], + "angle": 0, + "content": "tic facial images with huge artifacts or create low-resolution faces. In contrast, our method always creates good-quality realistic faces. We present the quantitative results in Table 2. The table shows that our method obtains the best FID scores of all methods and obtains the best mIOU score among the inference-time techniques." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.71, + 0.65 + ], + "angle": 0, + "content": "5.3. Face Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We evaluate the performance for the face super-resolution task using the CelebA dataset [19]. As baselines, we utilize fully inference-time methods in which no task-specific training is used. As the first baseline, we choose PULSE [22], a self-supervised upsampling technique utilizing GANs. As the next comparison method, we choose ILVR [5] which, like our method, performs super-resolution utilizing an unconditional pre-trained diffusion model. However, in ILVR, sampling happens at timestep \\( t \\) rather than at the implicit step in our algorithm. In total, we utilize 300 images for evaluations. We present some qualitative results in Fig. 6. For ILVR [5], we utilize 100 timesteps of sampling, the same as in our case. As we can see, PULSE [22] and ILVR [5] are unable to restore the correct identity and also contain blur artifacts after restoration. On the other hand, steered diffusion (our method) is able to restore photorealistic facial" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "20856" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.171, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.284, + 0.154, + 0.293 + ], + "angle": 0, + "content": "Degraded" + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.089, + 0.257, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.284, + 0.238, + 0.293 + ], + "angle": 0, + "content": "Reinpaint" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.089, + 0.343, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.284, + 0.312, + 0.293 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.089, + 0.44, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.284, + 0.422, + 0.294 + ], + "angle": 0, + "content": "Degraded" + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.089, + 0.525, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.284, + 0.506, + 0.294 + ], + "angle": 0, + "content": "Reinpaint" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.611, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.284, + 0.58, + 0.293 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.089, + 0.709, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.284, + 0.69, + 0.294 + ], + "angle": 0, + "content": "Degraded" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.793, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.284, + 0.775, + 0.294 + ], + "angle": 0, + "content": "Reinpaint" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.089, + 0.879, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.284, + 0.848, + 0.293 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.296, + 0.731, + 0.309 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons for inpainting for thin, medium and thick masks." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.317, + 0.202, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.406, + 0.182, + 0.417 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.318, + 0.313, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.406, + 0.311, + 0.417 + ], + "angle": 0, + "content": "Photo of a young man" + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.318, + 0.428, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.406, + 0.415, + 0.417 + ], + "angle": 0, + "content": "She has wavy hair" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.318, + 0.54, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.406, + 0.532, + 0.416 + ], + "angle": 0, + "content": "She has blonde hair" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.318, + 0.655, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.406, + 0.653, + 0.416 + ], + "angle": 0, + "content": "Photo of an old woman" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.318, + 0.767, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.406, + 0.741, + 0.417 + ], + "angle": 0, + "content": "She is angry." + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.318, + 0.882, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.406, + 0.849, + 0.416 + ], + "angle": 0, + "content": "She is sad." + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.43, + 0.672, + 0.444 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparisons for text-based image editing." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.454, + 0.483, + 0.506 + ], + "angle": 0, + "content": "
Trained MethodsInference-Time Methods
MethodCUT [16]GCGAN [9]CycleGAN [46]ILVR [5]Ours
FID (↓)49.3458.8052.70107.4641.38
mIoU(↑)0.6470.6990.7230.3770.532
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.51, + 0.427, + 0.523 + ], + "angle": 0, + "content": "Table 2. Quantitative results for semantic generation" + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.54, + 0.449, + 0.615 + ], + "angle": 0, + "content": "
MethodFID ↓LPIPS↓NIQE↓PSNR↑SSIM↑
Bicubic130.30.441912.0323.850.642
ILVR [5]62.240.41647.3820.540.5527
PULSE [22]84.670.43655.0421.080.5285
Ours51.190.25939.2426.020.711
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.62, + 0.418, + 0.633 + ], + "angle": 0, + "content": "Table 3. Quantitative results for super-resolution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.471, + 0.713 + ], + "angle": 0, + "content": "images. The qualitative evaluations are presented in Table 3; our method yields a 0.18 improvement in perceptual similarity, 6.95 dB improvement in PSNR, and 0.24 improvement in SSIM versus all of the other comparison methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.249, + 0.74 + ], + "angle": 0, + "content": "5.4. Face Colorization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "As a baseline method, we modify ILVR [5] to suit the task of colorization. For this, rather than performing the constraint at every step, we start the sampling process from a noised grayscale image and enforce consistency between the generated and original grayscale images. In total, we utilize 300 images for evaluation. The corresponding results can be seen in Fig. 7. As we can see, our method is able to reconstruct photorealistic faces with naturalistic colours compared to ILVR [5]. The corresponding quantitative metrics are presented in Table. 4. We get a significant boost in" + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.46, + 0.842, + 0.538 + ], + "angle": 0, + "content": "
MethodFID ↓LPIPS↓NIQE↓
Grayscale69.270.27815.07
ILVR [5]67.660.52707.54
Ours49.720.33115.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.564, + 0.548, + 0.828, + 0.562 + ], + "angle": 0, + "content": "Table 4. Quantitative results for colorization" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.592, + 0.892, + 0.621 + ], + "angle": 0, + "content": "performance, with an FID score of 19, LPIPS [44] score of 0.19, and NIQE [23] score of 1.5 for our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.643, + 0.891, + 0.66 + ], + "angle": 0, + "content": "5.5. Inpainting, Image Editing, Identity replication" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.895, + 0.82 + ], + "angle": 0, + "content": "Our method can also utilize multiple conditions simultaneously; we provide an illustration in Figure 9 where we condition with an identity-preserving network and a text caption simultaneously. From the figure, one can see that our method can generalize well to a diverse range of captions. For preserving identity, we used the VGGFace network [30]. We utilize FARL [45], which is pre-trained with face and corresponding text pairs to enforce the captions. For generic identity replication as in Figure 1 we use FARL face embedder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "For our image inpainting experiments, we use the subset released by [43] and evaluate three different kinds of masks. Our method obtains better results than existing baselines across all mask variations. Qualitative and quantitative results are shown in Fig. 8 and Table 5, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "20857" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.083, + 0.916, + 0.172 + ], + "angle": 0, + "content": "
MethodThinMediumThick
FID ↓LPIPS↓PSNR↑SSIM↑FID ↓LPIPS↓PSNR↑SSIM↑FID ↓LPIPS↓PSNR↑SSIM↑
Degraded371.030.6767.8470.201258.860.6247.5270.228231.930.5857.6060.249
Reinpaint [20]43.350.30417.990.67153.710.39913.710.55852.400.40712.780.530
Ours30.850.18324.920.83335.390.22021.550.78640.120.24218.870.705
" + }, + { + "type": "table_caption", + "bbox": [ + 0.356, + 0.176, + 0.612, + 0.19 + ], + "angle": 0, + "content": "Table 5. Quantitative results for inpainting." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.196, + 0.209, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.283, + 0.19, + 0.294 + ], + "angle": 0, + "content": "Semantic Map" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.197, + 0.318, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.283, + 0.284, + 0.293 + ], + "angle": 0, + "content": "\\(K = 0\\)" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.197, + 0.428, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.283, + 0.399, + 0.293 + ], + "angle": 0, + "content": "\\(K = 200\\)" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.197, + 0.536, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.283, + 0.513, + 0.293 + ], + "angle": 0, + "content": "\\(K = 2000\\)" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.197, + 0.646, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.283, + 0.624, + 0.293 + ], + "angle": 0, + "content": "\\(K = 20000\\)" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.197, + 0.755, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.283, + 0.738, + 0.293 + ], + "angle": 0, + "content": "\\(K = 2\\times 10^{5}\\)" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.197, + 0.865, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.283, + 0.847, + 0.293 + ], + "angle": 0, + "content": "\\(K = 2\\times 10^{6}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.296, + 0.73, + 0.31 + ], + "angle": 0, + "content": "Figure 10. Figure showing sample variation with scaling factor \\( k(t) = K\\sqrt{1 - \\bar{\\alpha}_t} \\)" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.316, + 0.225, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.387, + 0.225, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.459, + 0.203, + 0.469 + ], + "angle": 0, + "content": "Degraded" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.316, + 0.318, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.387, + 0.318, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.459, + 0.302, + 0.469 + ], + "angle": 0, + "content": "Noise robust" + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.316, + 0.411, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.459, + 0.382, + 0.469 + ], + "angle": 0, + "content": "OURS" + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.471, + 0.429, + 0.485 + ], + "angle": 0, + "content": "Figure 11. Qualitative comparisons for Colorization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.23, + 0.511 + ], + "angle": 0, + "content": "5.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Effect of scaling factor \\( k(t) \\) for semantic generation: In this section, we analyze how the scaling factor affects the quality of the sample in the case of complex conditioning of semantic generation. Fig. 10 shows the variation of sample quality starting from the same initial noise with different scaling factors \\( k(t) = K\\sqrt{1 - \\bar{\\alpha}_t} \\). The sample quality is bad for very low scaling factors, and for very high scaling factors, the diffusion process escapes the manifold of natural face images. We show the variation in sample quality for a fixed scaling factor versus a time-varying scaling factor in Fig. 12, which demonstrates that a time-varying scale factor produces more realistic samples. This is because the effective variance of the noise scheduling, which effectively controls the amount of regularization possible at a particular timestep, reduces as the generation process proceeds. Hence a larger tweak is permissible at the early steps of diffusion, and only very small tweaks are permitted in the later steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.471, + 0.869 + ], + "angle": 0, + "content": "Noise-Robust classifier: To validate the claims in section 3.1, we train a noise-robust inverse mapper for the task of colorization and show the output of the noise-robust classifier and the diffusion outputs for different noise levels in Fig. 11. The noise-robust classifier fails to preserve key details that our approach preserves." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Limitations Although our method can generalize to a wide series of tasks, one limitation that persists is the value of the" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.316, + 0.651, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.393, + 0.651, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.47, + 0.651, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.316, + 0.751, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.393, + 0.751, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.47, + 0.752, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.316, + 0.85, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.393, + 0.851, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.47, + 0.851, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.55, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Figure 12. A comparison of a time-varying scaling factor with non-time-varying." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.894, + 0.689 + ], + "angle": 0, + "content": "scaling factor \\( k(t) \\). The value of \\( k(t) \\) has to be empirically found based on the task, but once a few images are used to tune the value of \\( k(t) \\), the model generalizes well to other conditioning images of the same task. Like any other conditional generation models that can perform image editing, our method also has societal impacts, and care must be taken in applying these methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.619, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose the first framework for plug- and-play conditional generation that can generalize well to both image-to-image translation tasks and label-based generation tasks. For this, we use the energy-based formulation of diffusion models and modulate the inference process using a task-specific predefined network or other preexisting function. Furthermore, we introduce a novel implicit sampling-based technique that improves the sampling quality across multiple tasks. We performed experiments on various tasks to show that our method can generalize across multiple tasks and outperforms existing methods that do not require additional training." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "20858" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[3] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. arXiv preprint arXiv:2302.07121, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.472, + 0.428 + ], + "angle": 0, + "content": "[5] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.472, + 0.556 + ], + "angle": 0, + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.472, + 0.601 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.603, + 0.472, + 0.672 + ], + "angle": 0, + "content": "[9] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, Kun Zhang, and Dacheng Tao. Geometry-Consistent Generative Adversarial Networks for One-Sided Unsupervised Domain Mapping. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.673, + 0.472, + 0.728 + ], + "angle": 0, + "content": "[10] Ruiqi Gao, Yang Song, Ben Poole, Ying Nian Wu, and Diederik P Kingma. Learning energy-based models by diffusion recovery likelihood. arXiv preprint arXiv:2012.08125, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.731, + 0.472, + 0.786 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[12] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. arXiv preprint arXiv:2206.09012, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[13] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. arXiv preprint arXiv:1912.03263, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[15] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts gans. arXiv preprint arXiv:2112.05130, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.893, + 0.288 + ], + "angle": 0, + "content": "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.893, + 0.328 + ], + "angle": 0, + "content": "[18] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.894, + 0.384 + ], + "angle": 0, + "content": "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), December 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[20] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 3, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.894, + 0.51 + ], + "angle": 0, + "content": "[21] Kangfu Mei, Nithin Gopalakrishnan Nair, and Vishal M Patel. Bi-noising diffusion: Towards conditional diffusion models with generative restoration priors. arXiv preprint arXiv:2212.07352, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[22] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2437-2445, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.894, + 0.623 + ], + "angle": 0, + "content": "[23] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.894, + 0.692 + ], + "angle": 0, + "content": "[24] Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, and Vishal M Patel. Unite and conquer: Plug & play multimodal synthesis using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6070-6079, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.894, + 0.762 + ], + "angle": 0, + "content": "[25] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4467-4477, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.894, + 0.83 + ], + "angle": 0, + "content": "[26] Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "20859" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[28] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.472, + 0.272 + ], + "angle": 0, + "content": "[30] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In *British Machine Vision Conference*, 2015. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[31] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.472, + 0.48 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.472, + 0.524 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.593 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.472, + 0.676 + ], + "angle": 0, + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.472, + 0.733 + ], + "angle": 0, + "content": "[37] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.472, + 0.789 + ], + "angle": 0, + "content": "[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[39] Ken Sekimoto. Langevin equation and thermodynamics. Progress of Theoretical Physics Supplement, 130:17-27, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.133 + ], + "angle": 0, + "content": "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[42] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[43] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.893, + 0.33 + ], + "angle": 0, + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.894, + 0.401 + ], + "angle": 0, + "content": "[45] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. arXiv preprint arXiv:2112.03109, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.893, + 0.459 + ], + "angle": 0, + "content": "[46] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In IEEE International Conference on Computer Vision (ICCV), 2017. 3, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "20860" + } + ] +] \ No newline at end of file diff --git a/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/396c3711-d204-4df2-8e69-9bdf93d4fbff_origin.pdf b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/396c3711-d204-4df2-8e69-9bdf93d4fbff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d9e30829fcb2779c754f984f49ea75e3a106c881 --- /dev/null +++ b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/396c3711-d204-4df2-8e69-9bdf93d4fbff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:924e265d7460b3a5822dfc03af393c195a9864c199881ec6e84158f239610727 +size 10275112 diff --git a/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/full.md b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6a3ca1a86e6da769cd92698c7eb32a0fecd1d2b3 --- /dev/null +++ b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/full.md @@ -0,0 +1,518 @@ +# Steered Diffusion: A Generalized Framework for Plug-and-Play Conditional Image Synthesis + +Nithin Gopalakrishnan Nair $^{1*}$ Anoop Cherian $^{2}$ Suhas Lohit $^{2}$ Ye Wang $^{2}$ Toshiaki Koike-Akino $^{2}$ Vishal M. Patel $^{1}$ Tim K. Marks $^{2}$ + +$^{1}$ Johns Hopkins University $^{2}$ Mitsubishi Electric Research Laboratories (MERL) + +{ngopala2,vpatel36}@jhu.edu {acherian,slohit,yewang,koike,tmarks}@merl.com + +https://merl.com/demos/steered-diffusion + +![](images/d6d43ea9497aeeff66941815a9e222105e0545c582c8aecb02294b764894edb9.jpg) +Input + +![](images/20d52534047cca05edb7e9035eb3076a6d555bfd8b0c15b5ee41f1108a4dce05.jpg) +(a) Inpainting + +![](images/5db5887effc534890999561fdd8db83c9371a6401f12f3cfde3ab2741ab5b153.jpg) +Generated Samples + +![](images/01a8e0bf55bb1684b1cb8a53e68a63e6e47b64622decfafc6a291b2f55ed9a83.jpg) +Input + +![](images/34589c9422d014e2a197dab34129bfca1142c3286f1b6424a7ea0504dde6a37b.jpg) +(b) Colorization +Generated Samples + +![](images/1394637a8ae4cb0854b93fe08a225eb627d625fe1296f43a859d5a2ae318c74f.jpg) +Input + +![](images/af87089ed381715a43e164e486bc3c1664fbd1f1c0517bf17685f1ac7dd3d634.jpg) +Generated Samples + +![](images/7abccc03400d49abfc04759453e279268a5a911abcfe0f787977545a64c18f90.jpg) + +![](images/099e2a4b78121cfcb5bd3b78bf502bad6315dd5208be214dc824afac59dc613f.jpg) + +![](images/ffe046b2525aa3da7e78401245d277b7ecc7e10ba4c515b507d0ec1616c858f4.jpg) +Input + +![](images/020da89ebd90ba9e9f33ebfd93abe775b7a9eabb1c5a1de0b49f55a87411b312.jpg) + +![](images/ac63b95cefee6a5dd25e6eaa8d264a4d37f40d44d0d725aac4248d4c71ee302e.jpg) +Generated Samples + +![](images/a376821d67a3b1580a4bf17597e6915715ce5ea4b6bcad893daf31bf43137631.jpg) + +![](images/ae4da69c0abf6bc633990e8f92d5852b7eb128e2656adb312c6791fcb766ad34.jpg) +Figure 1. An illustration of various applications of our method. We use a diffusion model trained unconditionally and condition using our proposed algorithm only during the test time. We present the results on six tasks: (a) image inpainting, (b) colorization, (c) image super-resolution, (d) semantic generation, (e) identity replication, and (f) text-based image editing. In part (f), the text prompts for the first and second columns, respectively, are "This person has blonde hair" and "This person has wavy hair." + +![](images/e14d42300e62a0cfa860f8f36a4b23a96a8fe235f42c6f437a187eb663448197.jpg) + +![](images/3a887ce85cac79f97bade9936416150ad4a6d1c7e151c6da753fdb2f538da4c5.jpg) +Input + +![](images/59a2addfe500d8a7bd65a28becf9b2722a52d97de454491352ab9856ef91b866.jpg) + +![](images/28d0ba19b215f092bf1b19ca7fbe0a3c85bd2701eb72eeb788a21bae5bee6635.jpg) +Generated Samples + +![](images/091addb3904f01a135e8e72bb484a3c8dfe956d27319d450c5e97abc3cac0fc2.jpg) +(f) Text-based editing + +![](images/40d434215f88d8ed34d74528dd777296a85dc411a100324caae98ff515e311cb.jpg) +Input + +![](images/ce17f26b769a984c8ecebf62ed7714e9357e9d413c6169a1dc83115b81719973.jpg) + +![](images/5f4a9b88b7ec46662e3d0290fa7951984a2b503a4d582bb6a21e30be36fc9292.jpg) +Generated Samples + +![](images/2a419a4b54bc0cacd25c0d6a1c3b30ba6f63672ac702bf9840d8f4bf5782fb93.jpg) + +![](images/f15468105f903a50aa454a42249cb876ab534bbd78a325f79e41f027d80f2935.jpg) + +# Abstract + +Conditional generative models typically demand large annotated training sets to achieve high-quality synthesis. As a result, there has been significant interest in designing models that perform plug-and-play generation, i.e., to use a predefined or pretrained model, which is not explicitly trained on the generative task, to guide the generative process (e.g., using language). However, such guidance is typically useful only towards synthesizing high-level semantics rather than editing fine-grained details as in image-to-image translation tasks. To this end, and capitalizing on the powerful fine-grained generative control offered by the recent diffusion-based generative + +models, we introduce Steered Diffusion, a generalized framework for photorealistic zero-shot conditional image generation using a diffusion model trained for unconditional generation. The key idea is to steer the image generation of the diffusion model at inference time via designing a loss using a pre-trained inverse model that characterizes the conditional task. This loss modulates the sampling trajectory of the diffusion process. Our framework allows for easy incorporation of multiple conditions during inference. We present experiments using steered diffusion on several tasks including inpainting, colorization, text-guided semantic editing, and image super-resolution. Our results demonstrate clear qualitative and quantitative improvements over state-of-the-art diffusion-based plug-and-play models while adding negligible additional computational cost. + +# 1. Introduction + +Deep diffusion-based probabilistic generative models [8, 14, 40] are quickly emerging as one of the most powerful methods to synthesize high-quality content and have shown the potential to revolutionize content creation not only in computer vision, but also in many other areas including speech, audio, and language. Such models (e.g., ImagGen [36], Stable Diffusion [34]) have demonstrated outstanding synthesis results in conditional generation tasks, such as text-conditioned image synthesis [2, 33] and image reconstruction [31, 35, 38]. However, these models do not typically possess zero-shot conditional generative abilities when used directly (zero-shot capabilities as are commonly seen in language foundation models such as GPT-3 [4]), and often demand large amounts of annotated and paired (multimodal) data for conditional generation, which may be challenging to obtain [15]. + +One way to circumvent this need for large annotated training sets is to leverage predefined models as plug-and-play modules [12, 24, 25] in an otherwise unconditionally trained diffusion model. Specifically, in such plug-and-play models, a model is first trained in an unconditional setting (without labels). During inference, the plug-and-play modules (networks separately trained for a particular conditional task, e.g., image captioning) are incorporated in the reverse diffusion process to produce intermediate samples guided in the Markov chain in specific directions to satisfy the desired condition. Prior works, such as [25, 26], have proposed similar methods in which the authors derive text- or class-conditioned samples from Generative Adversarial Networks (GANs) [11] that were trained without labels. To achieve this, they iteratively refine the noise input of the GAN until the desired sample satisfies the condition. Very recently, Grakios et al. [12] proposed a diffusion-based plug-and-play method that enables using unconditional diffusion models for conditional generation utilizing class labels. Both these methods are specifically designed for tasks involving label-level semantics. However, these methods do not address the usage of unconditional models for general image-to-image translation tasks, which require synthesizing visual content conditioned on fine-grained details in the source image. There are also works that propose diffusion models for image-to-image translation, such as for image super-resolution and inpainting [5, 20]; however, these methods are task-specific and do not generalize well to new tasks or new types of inverse problems (as demonstrated in Section 5). In this work, we present a generic framework that can generalize to any image-to-image translation task. + +In this paper, we derive the necessary theory and formulate an algorithm, which we call Steered Diffusion, for diffusion-based image editing and image-to-image translation; our model is subsequently validated on a wide range of tasks. Steered diffusion is motivated by the energy-based + +![](images/c1801592f0ca9a188469e33ae593f7b8df6decc559534a3f1f23fad90c633641.jpg) +Figure 2. An illustration of the difference between existing plug-and-play generation approaches (e.g., [12]) and the proposed approach. Existing plug-and-play works operate with an energy function $V$ of the noisy latent $x_{t}$ . In contrast, our model uses the implicit prediction of the diffusion model (i.e., a coarse estimate of the clean image $x_{0}$ ) in its energy function $V_{1}$ , which allows the use of any pre-trained network for steering. In addition, our model provides a looping mechanism $V_{2}$ , which iterates $N$ times at each timestep $t$ to enhance generation quality. + +formulation of diffusion probabilistic models [10]. In general, inference in a generative model can be thought of as deriving samples from a learned distribution. Recall that every probability density function can be formulated as an energy field that describes an unnormalized estimate of how the distribution density varies in space [13,25]. If one needs to find points in space that are the closest match to a given condition, one can utilize gradient-based optimization algorithms to find points in the field that have the highest density value for the condition. The gradient-based optimization scheme can be viewed as a modulation of the energy toward the desired direction. Previous work has utilized this idea on GANs [25, 26] and obtained reasonable results for label-based generation tasks. Due to their model structure, diffusion models are ideal candidates for such an energy modulation. One key challenge remains to design a good energy estimator that is robust to all noise levels. Previously, classifier-based guidance [8, 27] has been proposed and thought of as an energy modulation utilizing a pretrained classifier trained on noisy images. This poses a limitation that the guiding function should be noise-robust. In this work, we propose an alternative solution that does not need noise-robust networks but could use any network by utilizing the diffusion model as an implicit denoiser. Figure 2 gives a brief overview of how our approach is different from existing methods. + +We present experiments using steered diffusion on multiple conditional generative tasks on faces as well as generic images as portrayed in Figure 1. We present results on (i) identity replication [7], (ii) semantic image generation [29], (iii) linear inverse problems [21], and (iv) text-conditioned + +image editing. Although our method is generic, for evaluations we perform experiments on faces. Before presenting our framework in detail, we now summarize the key contributions of our work: + +- We propose steered diffusion, a general plug-and-play framework that can utilize various pre-existing models to steer an unconditional diffusion model. +- We present the first work applicable to both label-level synthesis and image-to-image translation tasks and demonstrate its effectiveness for various applications. +- We propose an implicit conditioning-based sampling strategy that significantly boosts the performance of conditional sampling from unconditional diffusion models compared with previous methods. +- We introduce a new strategy that uses multiple steps of projected gradient descent to improve sample quality. + +# 2. Background + +# 2.1. Related Work + +Early works on unpaired image-to-image translation utilize a cycle consistency loss between the input and the target domains [9, 46]. Newer works, such as [16], have introduced a contrastive learning-based approach where a contrastive loss between corresponding patches of the input and target domains is minimized. The consistency-based method often fails to generate photorealistic images; hence, conditional generative models are preferred when labeled data are available. A few works [37, 38] utilize diffusion models for conditional image-to-image translation because of their photorealistic generation quality. + +Guiding diffusion models during inference time has been explored by several works, such as [24]. The first method that proposed inference-time conditioning [8] uses a pretrained noise-robust classifier to guide the inference of an unconditional model. GLIDE [27] proposed a method for conditioning using text. Earlier work in plug-and-play modelling for generative models utilized GANs and performed iterative refinement on the latent space of GANs [25]. This method uses a predefined classifier or text captioning network to estimate a loss between the desired label output or text caption and the one generated from the GAN generator. This loss is backpropagated to refine the noise input of the GAN iteratively until the generator predicts the desired output. Recently, [12] proposed a method that uses diffusion models as a plug-and-play prior for class-conditioned generation. Several works have addressed the task of image-to-image translation using unconditional diffusion models [1,5,18,20], but each of these proposes a task-specific inference scheme. For example, ILVR [5] performs image super-resolution, and Reinpaint [20] performs image inpainting. Blended diffusion [1] proposes a method for text-conditioned image editing. DDRM [18] proposes an inference-time scheme + +offering a general solution for linear inverse problems such as colorization and super-resolution. + +# 2.2. Concurrent Work + +In concurrent work to ours that also explores zero-shot conditional generation using diffusion models, [3] used a text to image model [34] and a two-step forward and backward universal guidance process, but it works well only after heavy optimization on network-based inverse problems such as semantic generation and identity generation. Another concurrent work [42] explored the usage of unconditional diffusion models for linear inverse problems using a pseudo-inverse model. In contrast to all these prior works, our steered diffusion algorithm generalizes well to both image-to-image translation tasks and high-level label-based generation tasks. + +# 2.3. Denoising Diffusion Probabilistic Models + +Denoising diffusion probabilistic models (DDPMs) [14, 40] belong to a class of generative models in which the model learns the distribution of data through a Markovian sampling process. DDPMs consist of a forward process and a reverse process. Let $x_{t}$ denote the latent state of an input image at timestep $t$ in a diffusion process. The sampling operation $q(\cdot)$ for the forward process in DDPM is defined as: + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) := \mathcal {N} \left(x _ {t}; \sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} I\right), \tag {1} +$$ + +where $\{\beta_t\}$ is a predefined variance schedule and $I$ is the identity matrix. + +The forward process can be considered as a noising operation, where the next state $x_{t}$ is obtained from the current state $x_{t-1}$ by adding a small amount of Gaussian noise according to the sampled timestep. The state $x_{t}$ at timestep $t$ can also be sampled directly from the initial state $x_{0}$ , using: + +$$ +q \left(x _ {t} \mid x _ {0}\right) := \mathcal {N} \left(x _ {t}; \sqrt {\bar {\alpha} _ {t}} x _ {0}, (1 - \bar {\alpha} _ {t}) I\right), \tag {2} +$$ + +or equivalently, + +$$ +x _ {t} = x _ {0} \sqrt {\bar {\alpha} _ {t}} + \epsilon \sqrt {1 - \bar {\alpha} _ {t}}, \quad \epsilon \sim \mathcal {N} (0, I), \tag {3} +$$ + +where $\bar{\alpha}_t = \prod_{s=1}^t \alpha_s$ and $\alpha_t = 1 - \beta_t$ . + +In [40], it is shown that if the number of time steps is large and the increment in $\{\beta_i\}$ is small, then each step in the reverse sampling process can also be approximated by a Gaussian. If $\mu_{\theta}$ and $\Sigma_{\theta}$ respectively denote the mean and the covariance of this Gaussian, modelled via neural networks with parameters $\theta$ , then each reverse step samples the state $x_{t-1}$ according to: + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) := \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}, t\right), \Sigma_ {\theta} \left(x _ {t}, t\right)\right). \tag {4} +$$ + +The parameters $\theta$ are obtained by minimizing the variational lower bound of the negative log-likelihood of the data distribution. + +![](images/f0e06ea56d133f147010aea7036c8bdbb7ac94684f5fc7400b7ec82ea2f0386e.jpg) +Figure 3. An illustration of Steered Diffusion. During each step of the sampling process the implicit prediction is steered to the direction of the condition using a steering network or predefined function. Note that this figure is only for illustrating the idea and does not show the actual sampled images; in the actual sampling, the steering process is much more gradual, not sudden as portrayed in this image. + +![](images/2e5c8ae020418efe7901e29213e5df5d82d4260eb0794667150e284e33207bd2.jpg) + +# 3. Proposed Method + +# 3.1. Steered Diffusion at Inference Time + +Our work is motivated by the energy-based formulation of diffusion models. For any probability density function, the corresponding energy-based model (EBM) is defined by: + +$$ +p _ {\theta} (x) = \frac {\exp (- V (x))}{Z}, \tag {5} +$$ + +where $V(x)$ denotes the corresponding energy function across states $x$ , and $Z$ denotes a normalization constant. To derive samples from this distribution, one can utilize the Langevin equation [39] describing the state transition of a particle in the presence of an energy field. For diffusion models the sampling step is + +$$ +x _ {t - 1} = x _ {t} - \nabla_ {x _ {t}} \log p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) + \epsilon , \epsilon \sim \mathcal {N} (0, I). \tag {6} +$$ + +The term $\nabla_{x_t}\log p_\theta (x_{t - 1}|x_t)$ is called the score function of the density $p_{\theta}(x_{t - 1}|x_t)$ . One key advantage of the energy-based formulation is that it allows modulation of the energy function to satisfy given criteria. This was initially introduced as classifier guidance [8], which allows label conditional sampling from an unconditionally trained diffusion model utilizing a noise-robust classifier. In the remaining part of this section, we motivate how we can extend the functionality of unconditional diffusion models to conditional tasks. Consider a conditional sampling scenario based on a condition $c$ , for sampling from a state $x_{t}$ to state $x_{t - 1}$ . The conditional transition probability $p_{\theta}(x_{t - 1}|x_t,c)$ can be decomposed as + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}, c\right) \propto \frac {p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) p \left(c \mid x _ {t - 1}\right)}{p \left(c \mid x _ {t}\right)}. \tag {7} +$$ + +Hence, for any timestep $t$ , the effective score for conditional transition can be found by utilizing using the log + +of probability density in the EBM formulations (5) of the individual densities and can be represented as + +$$ +\nabla_ {x _ {t}} \log p _ {\theta} (x _ {t - 1} | x _ {t}, c) = +$$ + +$$ +\nabla_ {x _ {t}} \log p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) - \nabla_ {x _ {t}} V _ {1} \left(x _ {t}, c\right) + \nabla_ {x _ {t}} V _ {2} \left(x _ {t - 1}, c\right), \tag {8} +$$ + +where $V_{1}$ and $V_{2}$ are the corresponding energy functions that model the conditional distributions of $x_{t}$ and $x_{t-1}$ given a condition $c$ . Specifically, they project the higher dimensional $x_{t}$ to the lower dimensional space of $c$ and measure the distance between the mapped value and $c$ . The better this particular measure, the more effectively it can be used to generate conditional samples from an unconditional model. Using Eq. (8), the conditional sampling equation for the reverse process is + +$$ +\begin{array}{l} x _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(x _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} ^ {(t)} (x _ {t}, t)\right) \\ - \nabla_ {x _ {t}} V _ {1} (x _ {t}, c) + \nabla_ {x _ {t}} V _ {2} (x _ {t - 1}, c) + \sigma_ {t} \epsilon . \tag {9} \\ \end{array} +$$ + +Here $\epsilon_{\theta}^{(t)}(x_t,t)$ is the network prediction at a timestep $t$ and $\sigma_{t}$ is the corresponding variance of the reverse step. The formulation in Eq. (9) shows that the energy function requires a functional mapping from a noisy $x_{t}$ to $c$ . + +In many applications of interest, the mapping function from $x_{t}$ to $c$ is complex and can be modeled effectively using deep networks. For example, in image-to-image translation tasks such as image generation from semantic maps [29], the image is $x_{t}$ and the semantic map is the condition $c$ . Similarly, for text to image generation, the image is $x_{t}$ and $c$ corresponds to the text. Ideally, we would like to employ an existing (pre-trained) deep neural network for the mapping from $x_{t}$ to $c$ . However, deep networks are usually trained on clean images, which limits the usability of existing pretrained deep networks for mapping directly from the noisy image $x_{t}$ to $c$ . One workaround would be to use noise-robust networks to map from $x_{t}$ to $c$ , but training noise-robust networks for conditional mapping can be computationally + +expensive. Moreover, a network that is trained with multiple different noise levels often results in lower mapping performance, as it cannot denoise all noise levels accurately; we validate this claim experimentally in Section 5.6. Alternatively, one could include two mapping functions: a first that denoises $x_{t}$ , and a second that maps from the denoised image to $c$ . Rather than training a seperate denoising network, however, we realized that diffusion models are inherently trained as denoisers, and reconstruction quality improves as time proceeds in the reverse sampling of the diffusion process. Because of this capacity, we can use a reverse sampling step to make coarse predictions of the denoised image from any time step $t$ . + +Hence, we modify our original energy expression (8) to: + +$$ +\begin{array}{l} \nabla_ {x _ {t}} \log p _ {\theta} (x _ {t - 1} | x _ {t}, c) = \nabla_ {x _ {t}} \log p _ {\theta} (x _ {t - 1} | x _ {t}) \\ - \nabla_ {x _ {t}} V _ {1} \left(x _ {0 \mid t}, c\right) - \delta_ {1} + \nabla_ {x _ {t}} V _ {2} \left(x _ {0 \mid t - 1}, c\right) + \delta_ {2}, \tag {10} \\ \end{array} +$$ + +where, we define the implicit step prediction $x_{0|t}$ as: + +$$ +x _ {0 \mid t} = \frac {x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} ^ {(t)} (x _ {t})}{\sqrt {\bar {\alpha} _ {t}}}. \tag {11} +$$ + +Here, we assume $x_{t}$ and $x_{t-1}$ are first denoised to $x_{0|t}$ and $x_{0|t-1}$ , respectively. The terms $\delta_{1}$ and $\delta_{2}$ capture the errors arising from the shift in the domain from $x_{t}$ to $x_{0}$ and from $x_{t-1}$ to $x_{0}$ ; for large $t$ , $\delta_{1} \approx \delta_{2}$ as the implicit predictions at nearby steps tend to be similar. + +As shown in our experiments, the energy function should be selected according to the task. An easy way to choose the energy function is by looking at the training loss of the mapping network. For example, in the case of semantic generation, a good energy function is the cross-entropy loss between the predicted semantic map at any timestep and the input semantic map. In the case of identity replication, a good choice of regularization would be the negative cosine similarity score between the embeddings from a recognition network for the input and target image. In the case of text-to-image generation, it would be CLIP loss [32]. As a rule of thumb, an energy function could be chosen easily by looking at the loss function used to train the pre-trained network (or an inverse function) that maps from the image $x$ to the condition $c$ . + +# 3.2. Revisiting Sampling in Diffusion Models + +To obtain a closed-form expression for plugging the energy-based formulation into the reverse sampling process efficiently, we take inspiration from DDIM [41] and revisit the reverse sampling operation of diffusion models. From $p_{\theta}(x_{1:T})$ , one can generate a sample $x_{t-1}$ from a sample $x_{t}$ by: + +Algorithm 1 Steered Diffusion +Input: Energy function $V$ , condition $c$ +1: $x_{T} \sim \mathcal{N}(x_{T}; 0, I)$ +2: for $t = T - 1, \ldots, 1$ do +3: for $n = N, \ldots, 1$ do +4: $\epsilon \sim \mathcal{N}(\epsilon; 0, I)$ +5: $x_{0|t} = \frac{x_{t} - \sqrt{1 - \bar{\alpha}_{t}} \epsilon_{\theta}^{(t)}(x_{t})}{\sqrt{\bar{\alpha}_{t}}}$ +6: Compute $x_{0|t}^{\text{feas}}$ with $V, c$ using Eq. (15) +7: if $(n > 1)$ then +8: Compute $x_{t}^{uc}$ using Eq. (13) +9: else +10: Compute $x_{t-1}^{uc}$ using Eq. (13) +11: end if +12: end for +13: end for +14: return $x_{0}$ + +$$ +\begin{array}{l} x _ {t - 1} ^ {\mathrm {u c}} = \sqrt {\bar {\alpha} _ {t - 1}} \cdot \underbrace {\frac {x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} ^ {(t)} (x _ {t})}{\sqrt {\bar {\alpha} _ {t}}}} _ {\text {‘ p r e d i c t e d x _ {0} ”}} \\ + \underbrace {\sqrt {1 - \bar {\alpha} _ {t - 1} - \sigma_ {t} ^ {2}} \cdot \epsilon_ {\theta} ^ {(t)} \left(x _ {t}\right)} _ {\text {d i r e c t i o n p o i n t i n g t o} x _ {t} ”} + \underbrace {\sigma_ {t} \epsilon} _ {\text {r a n d o m n o i s e}}, \tag {12} \\ \end{array} +$$ + +as in Song et al. [41]. Using Eq. (11), we can rewrite the unconditional sampling step Eq. (12) in terms of $x_{0|t}$ : + +$$ +\begin{array}{l} x _ {t - 1} ^ {\mathrm {u c}} = \sqrt {\bar {\alpha} _ {t - 1}} x _ {0 | t} + \\ \sqrt {1 - \bar {\alpha} _ {t - 1} - \sigma_ {t} ^ {2}} \cdot \frac {x _ {t} - \sqrt {\bar {\alpha} _ {t}} x _ {0 | t}}{\sqrt {1 - \bar {\alpha} _ {t}}} + \sigma_ {t} \epsilon , \tag {13} \\ \end{array} +$$ + +Here the superscript uc denotes the unconditional sample, which is obtained without any steering while transitioning from $x_{t}$ to $x_{t-1}$ . The conditional sampling step (8) can then be rewritten as + +$$ +x _ {t - 1} = x _ {t - 1} ^ {\mathrm {u c}} - \nabla_ {x _ {t}} V _ {1} \left(x _ {0 \mid t}, c\right) + \nabla_ {x _ {t}} V _ {2} \left(x _ {0 \mid t - 1}, c\right). \tag {14} +$$ + +Through this, we can modulate $x_0$ directly, as $x_{0|t}$ is also a function of $\epsilon_{\theta}^{(t)}(x_t)$ . Following Eq. (14), a rough estimate of the desired $x_0$ for conditional sampling, represented by $x_{0|t}^{\mathrm{feas}}$ , can be obtained using + +$$ +x _ {0 | t} ^ {\text {f e a s}} = x _ {0 | t} - k (t) \nabla_ {x _ {t}} \left(V _ {1} \left(x _ {0 | t}, c\right) - V _ {2} \left(x _ {0 | t - 1}, c\right)\right), \tag {15} +$$ + +where $k(t)$ is a scaling factor defining the strength of regularization. We call the process of finding $x_{0|t}^{\mathrm{feas}}$ from $x_{0|t}$ Implicit Steering Control (ISC), and we call the new sampling process steered diffusion. The exact algorithm is illustrated in Fig. 3 and explained in Algorithm 1. + +![](images/4e5898a5209a4adea25b9568a2695fdb4231e7516881f8935b027324939e0843.jpg) +Figure 4. An illustration of the steering function for linear inverse problems. For linear inverse problems, the component of the implicit prediction along the degradation direction can be replaced by the ground truth condition. + +# 4. Tips for Improved Performance + +# 4.1. Linear Inverse Problems + +For optimization-based inverse problems such as text-to-image generation and semantic map to natural image generation, the exact mapping function is not always available. On the other hand, for linear inverse problems such as colorization, super-resolution, and image inpainting, the mapping is simply a linear function, and we can write Eq. (15) more simply. In these cases, the exact mapping function to the latent space of the condition is known. Hence, one can decompose the implicit prediction at each timestep along the direction of the condition and simply replace this component by its desired ideal condition, i.e., if our predicted sample needs to map to a condition $c$ , then the modified implicit prediction step becomes + +$$ +x _ {0 | t} ^ {\text {f e a s}} = x _ {0 | t} + k (t) \left(D (y) - D \left(x _ {0 | t}\right)\right), \quad \text {w h e r e} c = D (y). \tag {16} +$$ + +Here, $y$ is the clean image and $D$ is the known degradation model. Our sampling procedure ensures that the series of operations preserve the consistency of domains of $x_{t-1}$ and $x_{0|t}^{\mathrm{feas}}$ to the original data distribution at the corresponding timesteps. An illustration is shown in Figure 4. + +# 4.2. Multi-Step Implicit Modulation + +Our experiments show (see Fig. 7) that performing refinement using Eq. (15) on the implicit step prediction multiple times for each timestep significantly boosts the conditioning quality for more ill-posed conditions such as image inpainting and colorization. A similar observation was also found by [20]. Specifically, at a particular timestep $t$ , we iterate the procedure of steering towards the next sampling step $x_{t - 1}$ and then adding noise to return to $x_{t}$ . We present the corresponding algorithm in Algorithm 1. An example is shown in Fig. 7, where more realistic images are generated using the multiple-step sampling scheme row labeled "OURS multi." + +Effectively, the $V_{2}$ term in Eq. (12) can be thought of as enabling a multistep sampling in which we modulate the current step by looking ahead to the next sampling step. On a careful analysis of Eq. (7), i.e. from the score contribution due to the different regularization functions, we can see that the term for $\nabla_{x_t}V_1(x_t,c)$ modulates $x_{t}$ based on its current state, and the term from $\nabla_{x_t}V_2(x_{t - 1},c)$ is a look-ahead correction where the derivative with respect to the future prediction is found. This is exactly what happens in the case of looping back from $x_{t - 1}$ to $x_{t}$ , where $x_{t}$ is modulated iteratively by looking forward to what the future prediction would be. + +# 4.3. Choosing the Scaling Factor $k(t)$ + +In Eq. (15), $k(t)$ denotes the strength of the regularization constraint. A very small $k(t)$ would denote no effective regularization, and a large $k$ would lead to the diffusion process going out of the latent space manifold. Since the derivative of the regularization function by itself is a score value, similar to the normal scaling value of the score function, the appropriate time-varying normalization factor is $\sqrt{1 - \bar{\alpha}_t}$ . The exact value for $k(t)$ for each task is defined in Table 1. For linear inverse problems we use a constant $k(t) = 1$ , which provided the best results. + +# 5. Experiments + +We evaluate the performance of our network qualitatively and quantitatively using four image-to-image translation tasks—semantic layout to face image translation, face inpainting, face colorization, and face super-resolution—as well as two high-level vision tasks: identity-based image generation and text-guided image editing. Unlike existing approaches, our method is completely zero-shot and applies to a wide variety of tasks. We compare with other diffusion-based approaches best applicable for each task for a fair evaluation. We also compare the semantic layout to image translation performance with that of task-specific unsupervised methods. We choose the unconditional model released by [5] as the unconditional pretrained diffusion model in all of our experiments with faces. Note that the sampling scheme in ILVR [5] and Repaint [20] can be thought of as happening at time $t$ rather than at the implicit step as in our method. Hence, comparing these methods for super-resolution and inpainting in our experiments below can be considered an additional ablation study highlighting the improvement from our implicit sampling. + +# 5.1. Implementation Details + +For our experiments, we utilize pixel-level unconditional diffusion models. For faces, we utilize the model trained on the FFHQ dataset [17] that was released by [5]. For generic images, we use the model trained on ImageNet [6] released in ADM [8]. All our experiments use 100 steps of sampling. + +
MethodLinear inverse problemsComplex inverse problems
ColorizationInpaintingSuper-resolutionSemantic GenerationIdentity ReplicationImage Editing
V1(D(x0|t) - c)2(D(x0|t) - c)2(D(x0|t) - c)2BCE(D(x0),c)1 - D(x0|t).D(c)3(D1(x0|t),D1(c1))2 + 2000CS(D2(x0|t),c2)
V2(D(x0|t-1) - c)2(D(x0|t-1) - c)20000
DGrayscaleMaskingDownscalingFARL [45]FARL recognizer [45]VGG Face [30] & FARL CLIP [32]
N331111
k(t)1112000√1-αt3000√1-αt10√1-αt & 1500√1-αt
+ +Table 1. Parameter set for each application. + +![](images/70bca2bf639c04e835e046524b143f7e51da70c6cd9ca0cc437d0ff13fd9d523.jpg) +Figure 5. Qualitative comparisons for semantic generation. + +![](images/d5860c116ea3730c61b86594e20ad8df02348eb991d591d82264a81b00f4ff28.jpg) +Figure 6. Results on $8 \times$ super resolution. + +# 5.2. Semantic Face Generation + +To evaluate how our method performs in generic image-to-image translation tasks, we evaluate our method's performance for the task of semantic layout to face generation. We utilize the CelebA dataset for this. To generate the semantic labels, we use facer [45] and create 11 label classes for each face. Since there is no other unconditional model that can perform fully test-time semantic generation, to evaluate the performance of our method, we compare with fully unsupervised image translation methods: CycleGAN [46], CUT [28], and ILVR [5]. + +The corresponding qualitative results are shown in Fig. 5. It is clear that CUT, CycleGAN, and ILVR produce unrealis- + +![](images/654e90b6fef1a0f14d94c6c9d443918f73b16fc15334d90fbab349901832da28.jpg) +Figure 7. Qualitative comparisons for colorization. The row labeled "OURS multi" refers to the use of multi-step sampling, as described in Sec. 4.2. + +tic facial images with huge artifacts or create low-resolution faces. In contrast, our method always creates good-quality realistic faces. We present the quantitative results in Table 2. The table shows that our method obtains the best FID scores of all methods and obtains the best mIOU score among the inference-time techniques. + +# 5.3. Face Super-Resolution + +We evaluate the performance for the face super-resolution task using the CelebA dataset [19]. As baselines, we utilize fully inference-time methods in which no task-specific training is used. As the first baseline, we choose PULSE [22], a self-supervised upsampling technique utilizing GANs. As the next comparison method, we choose ILVR [5] which, like our method, performs super-resolution utilizing an unconditional pre-trained diffusion model. However, in ILVR, sampling happens at timestep $t$ rather than at the implicit step in our algorithm. In total, we utilize 300 images for evaluations. We present some qualitative results in Fig. 6. For ILVR [5], we utilize 100 timesteps of sampling, the same as in our case. As we can see, PULSE [22] and ILVR [5] are unable to restore the correct identity and also contain blur artifacts after restoration. On the other hand, steered diffusion (our method) is able to restore photorealistic facial + +![](images/c32960c85d337eea0ed8f47ad4c429c5a8fd92905cea08a253cc3f26dd7f8727.jpg) +Degraded + +![](images/602d5d8b76d6172338380de68d8a08d359761e2b0fe212f717fddf7b169ce31f.jpg) +Reinpaint + +![](images/802d6320a6cddebb523b7a0d87582acc75e8c6d72f5adbe7863ad3bfbfbede4e.jpg) +Ours + +![](images/dfd65f5801bb9fc4af5f6616bee6ed63015fd6946eb1ec7bb29f3c0685e1daa1.jpg) +Degraded + +![](images/030f5d602bb1d81bf21a81ea1db0b0d79cf2c0b272484358566032d6c0096008.jpg) +Reinpaint + +![](images/57b700bc89d80244a40413e5ff38da989002a58b0bb533458d34953ba1cb55a1.jpg) +Ours + +![](images/725dca099ff18d343af72eae1e33dab367deccaf1cf95a4de63e2c8192894d94.jpg) +Degraded + +![](images/813654a93234b9249024a775f851c410b488edf9211d53f6c9b140703d38e5b1.jpg) +Reinpaint + +![](images/ae4bd3a4ec785891155b8b122e4de4938853850135c8bb07a64f69e26ad3409d.jpg) +Ours + +![](images/51aeb79b549932b0b45d635ce0b8c6c65877f3309c074c7aebda79c8910270bc.jpg) +Original Image + +![](images/7421b32858f0af6c759bce17f41674721722c5940f5f96c29daf50485b6caae1.jpg) +Figure 8. Qualitative comparisons for inpainting for thin, medium and thick masks. +Photo of a young man +Figure 9. Qualitative comparisons for text-based image editing. + +![](images/ee5c2da4602c1a169b122944814d8dfb3ebe342cbf2a4421c958e0fab9b9907c.jpg) +She has wavy hair + +![](images/a3872586b42c28021953312560a0dfdce680576be43ea63ccacdd72d20c6d15b.jpg) +She has blonde hair + +![](images/92f90481741da69482c90776ffe37526a062ab53b03b9b81e0c7e52012818997.jpg) +Photo of an old woman + +![](images/100df6cfac2b29c71d71547141a2ac171c4b8b3564748f8bced0ab66d51925da.jpg) +She is angry. + +![](images/de93655a1478376b937650bcf3dfb0024126ada5cde7324d976d4060020daa2a.jpg) +She is sad. + +
Trained MethodsInference-Time Methods
MethodCUT [16]GCGAN [9]CycleGAN [46]ILVR [5]Ours
FID (↓)49.3458.8052.70107.4641.38
mIoU(↑)0.6470.6990.7230.3770.532
+ +Table 2. Quantitative results for semantic generation + +
MethodFID ↓LPIPS↓NIQE↓PSNR↑SSIM↑
Bicubic130.30.441912.0323.850.642
ILVR [5]62.240.41647.3820.540.5527
PULSE [22]84.670.43655.0421.080.5285
Ours51.190.25939.2426.020.711
+ +images. The qualitative evaluations are presented in Table 3; our method yields a 0.18 improvement in perceptual similarity, 6.95 dB improvement in PSNR, and 0.24 improvement in SSIM versus all of the other comparison methods. + +# 5.4. Face Colorization + +As a baseline method, we modify ILVR [5] to suit the task of colorization. For this, rather than performing the constraint at every step, we start the sampling process from a noised grayscale image and enforce consistency between the generated and original grayscale images. In total, we utilize 300 images for evaluation. The corresponding results can be seen in Fig. 7. As we can see, our method is able to reconstruct photorealistic faces with naturalistic colours compared to ILVR [5]. The corresponding quantitative metrics are presented in Table. 4. We get a significant boost in + +Table 3. Quantitative results for super-resolution + +
MethodFID ↓LPIPS↓NIQE↓
Grayscale69.270.27815.07
ILVR [5]67.660.52707.54
Ours49.720.33115.91
+ +Table 4. Quantitative results for colorization + +performance, with an FID score of 19, LPIPS [44] score of 0.19, and NIQE [23] score of 1.5 for our method. + +# 5.5. Inpainting, Image Editing, Identity replication + +Our method can also utilize multiple conditions simultaneously; we provide an illustration in Figure 9 where we condition with an identity-preserving network and a text caption simultaneously. From the figure, one can see that our method can generalize well to a diverse range of captions. For preserving identity, we used the VGGFace network [30]. We utilize FARL [45], which is pre-trained with face and corresponding text pairs to enforce the captions. For generic identity replication as in Figure 1 we use FARL face embedder. + +For our image inpainting experiments, we use the subset released by [43] and evaluate three different kinds of masks. Our method obtains better results than existing baselines across all mask variations. Qualitative and quantitative results are shown in Fig. 8 and Table 5, respectively. + +
MethodThinMediumThick
FID ↓LPIPS↓PSNR↑SSIM↑FID ↓LPIPS↓PSNR↑SSIM↑FID ↓LPIPS↓PSNR↑SSIM↑
Degraded371.030.6767.8470.201258.860.6247.5270.228231.930.5857.6060.249
Reinpaint [20]43.350.30417.990.67153.710.39913.710.55852.400.40712.780.530
Ours30.850.18324.920.83335.390.22021.550.78640.120.24218.870.705
+ +Table 5. Quantitative results for inpainting. + +![](images/95dfe810af322fca7b4b1bac59f8a960198de272be940e59b7fcd3e232d7ba1b.jpg) +Semantic Map + +![](images/6ef6bed403bad1cc2de2f0a937e2e1cc49054aedbbd015f769b5a7223479b43a.jpg) +$K = 0$ + +![](images/5a8da2c8b25ec356836003d0422438696a32eaf74994819f738bd615c3f14965.jpg) +$K = 200$ + +![](images/39776518c35ced6dadd9567bbc7d63f30cd034ab6eb129a6bef2efd734b7e7c9.jpg) +$K = 2000$ + +![](images/9fe8229619d4dabac5ceb27c0dfe7c01e720f985c4e2d9bf8bcad14d34c4e437.jpg) +$K = 20000$ + +![](images/45b50e4c730c9d0c9cc337074bdbdb5a545e74a5c8477d06b51b595a63c26a8d.jpg) +$K = 2\times 10^{5}$ + +![](images/f3a11d08557b20d4c6b65b0bf11c871cf948fc5c452a16692fd262a6f45f9402.jpg) +$K = 2\times 10^{6}$ + +![](images/460407a823d71c159d1e69645223c962e949711ee7975ac5c707b604921e3286.jpg) +Figure 10. Figure showing sample variation with scaling factor $k(t) = K\sqrt{1 - \bar{\alpha}_t}$ + +![](images/c28bc45cccfa5c4f281514b4728aa05ac4da72c8a66c2d0b7645905b390a7a20.jpg) +Degraded +Figure 11. Qualitative comparisons for Colorization. + +![](images/1958e5e661498a87cd4213ae5721fe08dde99c3bac20488d6167fbfc1961ccaf.jpg) + +![](images/21c2dd39d7514dc2a2819893e1eec540787b7b3010f1f6d08ef02c82810c38d8.jpg) +Noise robust + +![](images/aa819edcc972238ffeafe4650c8411ca86b50fabf1b570e7454ccceeebea724e.jpg) +OURS + +# 5.6. Ablation Study + +Effect of scaling factor $k(t)$ for semantic generation: In this section, we analyze how the scaling factor affects the quality of the sample in the case of complex conditioning of semantic generation. Fig. 10 shows the variation of sample quality starting from the same initial noise with different scaling factors $k(t) = K\sqrt{1 - \bar{\alpha}_t}$ . The sample quality is bad for very low scaling factors, and for very high scaling factors, the diffusion process escapes the manifold of natural face images. We show the variation in sample quality for a fixed scaling factor versus a time-varying scaling factor in Fig. 12, which demonstrates that a time-varying scale factor produces more realistic samples. This is because the effective variance of the noise scheduling, which effectively controls the amount of regularization possible at a particular timestep, reduces as the generation process proceeds. Hence a larger tweak is permissible at the early steps of diffusion, and only very small tweaks are permitted in the later steps. + +Noise-Robust classifier: To validate the claims in section 3.1, we train a noise-robust inverse mapper for the task of colorization and show the output of the noise-robust classifier and the diffusion outputs for different noise levels in Fig. 11. The noise-robust classifier fails to preserve key details that our approach preserves. + +Limitations Although our method can generalize to a wide series of tasks, one limitation that persists is the value of the + +![](images/11599771e86de2750d03f850a0761e63cc549093484283f4e74ec94f72d0b1d5.jpg) + +![](images/581a0b3ad128f0c504ad2c9a07a89366a257c779b6f900e908f349b097659604.jpg) + +![](images/155d76b370e7d19f4906d3b1b93b541d94e9dda62caad6b49872d0a3e1630c38.jpg) +Figure 12. A comparison of a time-varying scaling factor with non-time-varying. + +![](images/83c47f3b33a6f470e1141945df08cffba35fac5f75eee27e4eeea8a40ebafaa3.jpg) + +![](images/39195c3e30d02512f8e9b3642dbdeeb5612ba3b384c6daac900f68cf3ed42dcc.jpg) + +![](images/fdecf438744e32334487383ecf40e743a6844701d723eb75b80a6fba975bd5de.jpg) + +![](images/c4b6624c3e8a2a14fdddc4d53d8340e72512d53bb9d7171c984ec7731b3c085a.jpg) + +![](images/2e3e83c1181c3cf3c8db0f94bd524bd67bf99d6a09a57d8d06b6f457afbb496c.jpg) + +![](images/1b98a9b384664e4bfecbf32be1ceb4043f0a0758040b2a278c39a8c8280dca23.jpg) + +scaling factor $k(t)$ . The value of $k(t)$ has to be empirically found based on the task, but once a few images are used to tune the value of $k(t)$ , the model generalizes well to other conditioning images of the same task. Like any other conditional generation models that can perform image editing, our method also has societal impacts, and care must be taken in applying these methods. + +# 6. Conclusion + +In this paper, we propose the first framework for plug- and-play conditional generation that can generalize well to both image-to-image translation tasks and label-based generation tasks. For this, we use the energy-based formulation of diffusion models and modulate the inference process using a task-specific predefined network or other preexisting function. Furthermore, we introduce a novel implicit sampling-based technique that improves the sampling quality across multiple tasks. We performed experiments on various tasks to show that our method can generalize across multiple tasks and outperforms existing methods that do not require additional training. + +# References + +[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2 +[3] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. arXiv preprint arXiv:2302.07121, 2023. 3 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2 +[5] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 2, 3, 6, 7, 8 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6 +[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 4, 6 +[9] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, Kun Zhang, and Dacheng Tao. Geometry-Consistent Generative Adversarial Networks for One-Sided Unsupervised Domain Mapping. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3, 8 +[10] Ruiqi Gao, Yang Song, Ben Poole, Ying Nian Wu, and Diederik P Kingma. Learning energy-based models by diffusion recovery likelihood. arXiv preprint arXiv:2012.08125, 2020. 2 +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[12] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. arXiv preprint arXiv:2206.09012, 2022. 2, 3 +[13] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. arXiv preprint arXiv:1912.03263, 2019. 2 + +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3 +[15] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts gans. arXiv preprint arXiv:2112.05130, 2021. 2 +[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3, 8 +[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6 +[18] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793, 2022. 3 +[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), December 2015. 7 +[20] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 3, 6, 9 +[21] Kangfu Mei, Nithin Gopalakrishnan Nair, and Vishal M Patel. Bi-noising diffusion: Towards conditional diffusion models with generative restoration priors. arXiv preprint arXiv:2212.07352, 2022. 2 +[22] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2437-2445, 2020. 7, 8 +[23] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 8 +[24] Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, and Vishal M Patel. Unite and conquer: Plug & play multimodal synthesis using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6070-6079, 2023. 2, 3 +[25] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4467-4477, 2017. 2, 3 +[26] Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks. Advances in neural information processing systems, 29, 2016. 2 +[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3 + +[28] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 7 +[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 2, 4 +[30] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In *British Machine Vision Conference*, 2015. 7, 8 +[31] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 7 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3 +[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2 +[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2 +[37] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021. 3 +[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3 +[39] Ken Sekimoto. Langevin equation and thermodynamics. Progress of Theoretical Physics Supplement, 130:17-27, 1998. 4 +[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2, 3 + +[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 +[42] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2022. 3 +[43] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 8 +[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8 +[45] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. arXiv preprint arXiv:2112.03109, 2021. 7, 8 +[46] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In IEEE International Conference on Computer Vision (ICCV), 2017. 3, 7, 8 \ No newline at end of file diff --git a/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/images.zip b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9c3b1a9e932b3b9aed6bab330fabf84c11aea653 --- /dev/null +++ b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6681f98fd0875a6eb307bc90279758797adb43f7a727173f68eaad91e1064ed7 +size 887003 diff --git a/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/layout.json b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ca163e23fe77bda70bdab4349f3eb4a550e1f68b --- /dev/null +++ b/2023/Steered Diffusion_ A Generalized Framework for Plug-and-Play Conditional Image Synthesis/layout.json @@ -0,0 +1,13825 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 91, + 495, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 91, + 495, + 126 + ], + "spans": [ + { + "bbox": [ + 99, + 91, + 495, + 126 + ], + "type": "text", + "content": "Steered Diffusion: A Generalized Framework for Plug-and-Play Conditional Image Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "text", + "content": "Nithin Gopalakrishnan Nair" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "text", + "content": " Anoop Cherian" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "text", + "content": " Suhas Lohit" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "text", + "content": " Ye Wang" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "text", + "content": " Toshiaki Koike-Akino" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "text", + "content": " Vishal M. Patel" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "text", + "content": " Tim K. Marks" + }, + { + "bbox": [ + 111, + 134, + 481, + 162 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 164, + 447, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 164, + 447, + 177 + ], + "spans": [ + { + "bbox": [ + 145, + 164, + 447, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 164, + 447, + 177 + ], + "type": "text", + "content": " Johns Hopkins University " + }, + { + "bbox": [ + 145, + 164, + 447, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 164, + 447, + 177 + ], + "type": "text", + "content": " Mitsubishi Electric Research Laboratories (MERL)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 180, + 455, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 180, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 180, + 455, + 190 + ], + "type": "text", + "content": "{ngopala2,vpatel36}@jhu.edu {acherian,slohit,yewang,koike,tmarks}@merl.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 186, + 194, + 405, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 194, + 405, + 204 + ], + "spans": [ + { + "bbox": [ + 186, + 194, + 405, + 204 + ], + "type": "text", + "content": "https://merl.com/demos/steered-diffusion" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 225, + 102, + 328 + ], + "blocks": [ + { + "bbox": [ + 49, + 225, + 102, + 328 + ], + "lines": [ + { + "bbox": [ + 49, + 225, + 102, + 328 + ], + "spans": [ + { + "bbox": [ + 49, + 225, + 102, + 328 + ], + "type": "image", + "image_path": "d6d43ea9497aeeff66941815a9e222105e0545c582c8aecb02294b764894edb9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 329, + 91, + 338 + ], + "lines": [ + { + "bbox": [ + 72, + 329, + 91, + 338 + ], + "spans": [ + { + "bbox": [ + 72, + 329, + 91, + 338 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 105, + 225, + 157, + 328 + ], + "blocks": [ + { + "bbox": [ + 96, + 213, + 141, + 222 + ], + "lines": [ + { + "bbox": [ + 96, + 213, + 141, + 222 + ], + "spans": [ + { + "bbox": [ + 96, + 213, + 141, + 222 + ], + "type": "text", + "content": "(a) Inpainting" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 105, + 225, + 157, + 328 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 157, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 157, + 328 + ], + "type": "image", + "image_path": "20d52534047cca05edb7e9035eb3076a6d555bfd8b0c15b5ee41f1108a4dce05.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 158, + 225, + 210, + 328 + ], + "blocks": [ + { + "bbox": [ + 127, + 329, + 191, + 338 + ], + "lines": [ + { + "bbox": [ + 127, + 329, + 191, + 338 + ], + "spans": [ + { + "bbox": [ + 127, + 329, + 191, + 338 + ], + "type": "text", + "content": "Generated Samples" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 158, + 225, + 210, + 328 + ], + "lines": [ + { + "bbox": [ + 158, + 225, + 210, + 328 + ], + "spans": [ + { + "bbox": [ + 158, + 225, + 210, + 328 + ], + "type": "image", + "image_path": "5db5887effc534890999561fdd8db83c9371a6401f12f3cfde3ab2741ab5b153.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 217, + 224, + 269, + 328 + ], + "blocks": [ + { + "bbox": [ + 217, + 224, + 269, + 328 + ], + "lines": [ + { + "bbox": [ + 217, + 224, + 269, + 328 + ], + "spans": [ + { + "bbox": [ + 217, + 224, + 269, + 328 + ], + "type": "image", + "image_path": "01a8e0bf55bb1684b1cb8a53e68a63e6e47b64622decfafc6a291b2f55ed9a83.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 329, + 258, + 338 + ], + "lines": [ + { + "bbox": [ + 239, + 329, + 258, + 338 + ], + "spans": [ + { + "bbox": [ + 239, + 329, + 258, + 338 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 274, + 224, + 325, + 328 + ], + "blocks": [ + { + "bbox": [ + 261, + 213, + 314, + 221 + ], + "lines": [ + { + "bbox": [ + 261, + 213, + 314, + 221 + ], + "spans": [ + { + "bbox": [ + 261, + 213, + 314, + 221 + ], + "type": "text", + "content": "(b) Colorization" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 274, + 224, + 325, + 328 + ], + "lines": [ + { + "bbox": [ + 274, + 224, + 325, + 328 + ], + "spans": [ + { + "bbox": [ + 274, + 224, + 325, + 328 + ], + "type": "image", + "image_path": "34589c9422d014e2a197dab34129bfca1142c3286f1b6424a7ea0504dde6a37b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 329, + 358, + 338 + ], + "lines": [ + { + "bbox": [ + 294, + 329, + 358, + 338 + ], + "spans": [ + { + "bbox": [ + 294, + 329, + 358, + 338 + ], + "type": "text", + "content": "Generated Samples" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 370, + 213, + 439, + 328 + ], + "blocks": [ + { + "bbox": [ + 370, + 213, + 439, + 328 + ], + "lines": [ + { + "bbox": [ + 370, + 213, + 439, + 328 + ], + "spans": [ + { + "bbox": [ + 370, + 213, + 439, + 328 + ], + "type": "image", + "image_path": "1394637a8ae4cb0854b93fe08a225eb627d625fe1296f43a859d5a2ae318c74f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 329, + 425, + 338 + ], + "lines": [ + { + "bbox": [ + 406, + 329, + 425, + 338 + ], + "spans": [ + { + "bbox": [ + 406, + 329, + 425, + 338 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 435, + 213, + 499, + 328 + ], + "blocks": [ + { + "bbox": [ + 435, + 213, + 499, + 328 + ], + "lines": [ + { + "bbox": [ + 435, + 213, + 499, + 328 + ], + "spans": [ + { + "bbox": [ + 435, + 213, + 499, + 328 + ], + "type": "image", + "image_path": "af87089ed381715a43e164e486bc3c1664fbd1f1c0517bf17685f1ac7dd3d634.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 329, + 525, + 338 + ], + "lines": [ + { + "bbox": [ + 462, + 329, + 525, + 338 + ], + "spans": [ + { + "bbox": [ + 462, + 329, + 525, + 338 + ], + "type": "text", + "content": "Generated Samples" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 493, + 213, + 547, + 328 + ], + "blocks": [ + { + "bbox": [ + 493, + 213, + 547, + 328 + ], + "lines": [ + { + "bbox": [ + 493, + 213, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 493, + 213, + 547, + 328 + ], + "type": "image", + "image_path": "7abccc03400d49abfc04759453e279268a5a911abcfe0f787977545a64c18f90.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 48, + 342, + 102, + 407 + ], + "blocks": [ + { + "bbox": [ + 48, + 342, + 102, + 407 + ], + "lines": [ + { + "bbox": [ + 48, + 342, + 102, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 102, + 407 + ], + "type": "image", + "image_path": "099e2a4b78121cfcb5bd3b78bf502bad6315dd5208be214dc824afac59dc613f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 48, + 408, + 101, + 459 + ], + "blocks": [ + { + "bbox": [ + 48, + 408, + 101, + 459 + ], + "lines": [ + { + "bbox": [ + 48, + 408, + 101, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 101, + 459 + ], + "type": "image", + "image_path": "ffe046b2525aa3da7e78401245d277b7ecc7e10ba4c515b507d0ec1616c858f4.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 460, + 91, + 468 + ], + "lines": [ + { + "bbox": [ + 72, + 460, + 91, + 468 + ], + "spans": [ + { + "bbox": [ + 72, + 460, + 91, + 468 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 105, + 355, + 157, + 407 + ], + "blocks": [ + { + "bbox": [ + 105, + 355, + 157, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 355, + 157, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 157, + 407 + ], + "type": "image", + "image_path": "020da89ebd90ba9e9f33ebfd93abe775b7a9eabb1c5a1de0b49f55a87411b312.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 104, + 408, + 157, + 459 + ], + "blocks": [ + { + "bbox": [ + 104, + 408, + 157, + 459 + ], + "lines": [ + { + "bbox": [ + 104, + 408, + 157, + 459 + ], + "spans": [ + { + "bbox": [ + 104, + 408, + 157, + 459 + ], + "type": "image", + "image_path": "ac63b95cefee6a5dd25e6eaa8d264a4d37f40d44d0d725aac4248d4c71ee302e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 460, + 191, + 468 + ], + "lines": [ + { + "bbox": [ + 127, + 460, + 191, + 468 + ], + "spans": [ + { + "bbox": [ + 127, + 460, + 191, + 468 + ], + "type": "text", + "content": "Generated Samples" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 158, + 355, + 210, + 407 + ], + "blocks": [ + { + "bbox": [ + 158, + 355, + 210, + 407 + ], + "lines": [ + { + "bbox": [ + 158, + 355, + 210, + 407 + ], + "spans": [ + { + "bbox": [ + 158, + 355, + 210, + 407 + ], + "type": "image", + "image_path": "a376821d67a3b1580a4bf17597e6915715ce5ea4b6bcad893daf31bf43137631.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 158, + 408, + 209, + 459 + ], + "blocks": [ + { + "bbox": [ + 158, + 408, + 209, + 459 + ], + "lines": [ + { + "bbox": [ + 158, + 408, + 209, + 459 + ], + "spans": [ + { + "bbox": [ + 158, + 408, + 209, + 459 + ], + "type": "image", + "image_path": "ae4da69c0abf6bc633990e8f92d5852b7eb128e2656adb312c6791fcb766ad34.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 473, + 546, + 518 + ], + "lines": [ + { + "bbox": [ + 46, + 473, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 546, + 518 + ], + "type": "text", + "content": "Figure 1. An illustration of various applications of our method. We use a diffusion model trained unconditionally and condition using our proposed algorithm only during the test time. We present the results on six tasks: (a) image inpainting, (b) colorization, (c) image super-resolution, (d) semantic generation, (e) identity replication, and (f) text-based image editing. In part (f), the text prompts for the first and second columns, respectively, are \"This person has blonde hair\" and \"This person has wavy hair.\"" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 217, + 355, + 269, + 407 + ], + "blocks": [ + { + "bbox": [ + 217, + 355, + 269, + 407 + ], + "lines": [ + { + "bbox": [ + 217, + 355, + 269, + 407 + ], + "spans": [ + { + "bbox": [ + 217, + 355, + 269, + 407 + ], + "type": "image", + "image_path": "e14d42300e62a0cfa860f8f36a4b23a96a8fe235f42c6f437a187eb663448197.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 217, + 408, + 269, + 459 + ], + "blocks": [ + { + "bbox": [ + 217, + 408, + 269, + 459 + ], + "lines": [ + { + "bbox": [ + 217, + 408, + 269, + 459 + ], + "spans": [ + { + "bbox": [ + 217, + 408, + 269, + 459 + ], + "type": "image", + "image_path": "3a887ce85cac79f97bade9936416150ad4a6d1c7e151c6da753fdb2f538da4c5.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 460, + 258, + 468 + ], + "lines": [ + { + "bbox": [ + 239, + 460, + 258, + 468 + ], + "spans": [ + { + "bbox": [ + 239, + 460, + 258, + 468 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 274, + 355, + 325, + 407 + ], + "blocks": [ + { + "bbox": [ + 274, + 355, + 325, + 407 + ], + "lines": [ + { + "bbox": [ + 274, + 355, + 325, + 407 + ], + "spans": [ + { + "bbox": [ + 274, + 355, + 325, + 407 + ], + "type": "image", + "image_path": "59a2addfe500d8a7bd65a28becf9b2722a52d97de454491352ab9856ef91b866.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 274, + 408, + 326, + 459 + ], + "blocks": [ + { + "bbox": [ + 274, + 408, + 326, + 459 + ], + "lines": [ + { + "bbox": [ + 274, + 408, + 326, + 459 + ], + "spans": [ + { + "bbox": [ + 274, + 408, + 326, + 459 + ], + "type": "image", + "image_path": "28d0ba19b215f092bf1b19ca7fbe0a3c85bd2701eb72eeb788a21bae5bee6635.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 460, + 358, + 468 + ], + "lines": [ + { + "bbox": [ + 294, + 460, + 358, + 468 + ], + "spans": [ + { + "bbox": [ + 294, + 460, + 358, + 468 + ], + "type": "text", + "content": "Generated Samples" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 386, + 356, + 438, + 407 + ], + "blocks": [ + { + "bbox": [ + 430, + 343, + 501, + 353 + ], + "lines": [ + { + "bbox": [ + 430, + 343, + 501, + 353 + ], + "spans": [ + { + "bbox": [ + 430, + 343, + 501, + 353 + ], + "type": "text", + "content": "(f) Text-based editing" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 356, + 438, + 407 + ], + "lines": [ + { + "bbox": [ + 386, + 356, + 438, + 407 + ], + "spans": [ + { + "bbox": [ + 386, + 356, + 438, + 407 + ], + "type": "image", + "image_path": "091addb3904f01a135e8e72bb484a3c8dfe956d27319d450c5e97abc3cac0fc2.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 386, + 408, + 438, + 459 + ], + "blocks": [ + { + "bbox": [ + 386, + 408, + 438, + 459 + ], + "lines": [ + { + "bbox": [ + 386, + 408, + 438, + 459 + ], + "spans": [ + { + "bbox": [ + 386, + 408, + 438, + 459 + ], + "type": "image", + "image_path": "40d434215f88d8ed34d74528dd777296a85dc411a100324caae98ff515e311cb.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 460, + 425, + 468 + ], + "lines": [ + { + "bbox": [ + 406, + 460, + 425, + 468 + ], + "spans": [ + { + "bbox": [ + 406, + 460, + 425, + 468 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 442, + 356, + 494, + 407 + ], + "blocks": [ + { + "bbox": [ + 442, + 356, + 494, + 407 + ], + "lines": [ + { + "bbox": [ + 442, + 356, + 494, + 407 + ], + "spans": [ + { + "bbox": [ + 442, + 356, + 494, + 407 + ], + "type": "image", + "image_path": "ce17f26b769a984c8ecebf62ed7714e9357e9d413c6169a1dc83115b81719973.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 442, + 408, + 494, + 459 + ], + "blocks": [ + { + "bbox": [ + 442, + 408, + 494, + 459 + ], + "lines": [ + { + "bbox": [ + 442, + 408, + 494, + 459 + ], + "spans": [ + { + "bbox": [ + 442, + 408, + 494, + 459 + ], + "type": "image", + "image_path": "5f4a9b88b7ec46662e3d0290fa7951984a2b503a4d582bb6a21e30be36fc9292.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 460, + 525, + 468 + ], + "lines": [ + { + "bbox": [ + 462, + 460, + 525, + 468 + ], + "spans": [ + { + "bbox": [ + 462, + 460, + 525, + 468 + ], + "type": "text", + "content": "Generated Samples" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 495, + 356, + 547, + 407 + ], + "blocks": [ + { + "bbox": [ + 495, + 356, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 495, + 356, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 495, + 356, + 547, + 407 + ], + "type": "image", + "image_path": "2a419a4b54bc0cacd25c0d6a1c3b30ba6f63672ac702bf9840d8f4bf5782fb93.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 495, + 408, + 547, + 459 + ], + "blocks": [ + { + "bbox": [ + 495, + 408, + 547, + 459 + ], + "lines": [ + { + "bbox": [ + 495, + 408, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 495, + 408, + 547, + 459 + ], + "type": "image", + "image_path": "f15468105f903a50aa454a42249cb876ab534bbd78a325f79e41f027d80f2935.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "bbox": [ + 159, + 528, + 208, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 528, + 208, + 540 + ], + "spans": [ + { + "bbox": [ + 159, + 528, + 208, + 540 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 46, + 552, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 288, + 696 + ], + "type": "text", + "content": "Conditional generative models typically demand large annotated training sets to achieve high-quality synthesis. As a result, there has been significant interest in designing models that perform plug-and-play generation, i.e., to use a predefined or pretrained model, which is not explicitly trained on the generative task, to guide the generative process (e.g., using language). However, such guidance is typically useful only towards synthesizing high-level semantics rather than editing fine-grained details as in image-to-image translation tasks. To this end, and capitalizing on the powerful fine-grained generative control offered by the recent diffusion-based generative" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 529, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 547, + 708 + ], + "type": "text", + "content": "models, we introduce Steered Diffusion, a generalized framework for photorealistic zero-shot conditional image generation using a diffusion model trained for unconditional generation. The key idea is to steer the image generation of the diffusion model at inference time via designing a loss using a pre-trained inverse model that characterizes the conditional task. This loss modulates the sampling trajectory of the diffusion process. Our framework allows for easy incorporation of multiple conditions during inference. We present experiments using steered diffusion on several tasks including inpainting, colorization, text-guided semantic editing, and image super-resolution. Our results demonstrate clear qualitative and quantitative improvements over state-of-the-art diffusion-based plug-and-play models while adding negligible additional computational cost." + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 703, + 195, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 195, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 195, + 713 + ], + "type": "text", + "content": "* Work done during internship at MERL." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20850" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 281 + ], + "type": "text", + "content": "Deep diffusion-based probabilistic generative models [8, 14, 40] are quickly emerging as one of the most powerful methods to synthesize high-quality content and have shown the potential to revolutionize content creation not only in computer vision, but also in many other areas including speech, audio, and language. Such models (e.g., ImagGen [36], Stable Diffusion [34]) have demonstrated outstanding synthesis results in conditional generation tasks, such as text-conditioned image synthesis [2, 33] and image reconstruction [31, 35, 38]. However, these models do not typically possess zero-shot conditional generative abilities when used directly (zero-shot capabilities as are commonly seen in language foundation models such as GPT-3 [4]), and often demand large amounts of annotated and paired (multimodal) data for conditional generation, which may be challenging to obtain [15]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 282, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 288, + 653 + ], + "type": "text", + "content": "One way to circumvent this need for large annotated training sets is to leverage predefined models as plug-and-play modules [12, 24, 25] in an otherwise unconditionally trained diffusion model. Specifically, in such plug-and-play models, a model is first trained in an unconditional setting (without labels). During inference, the plug-and-play modules (networks separately trained for a particular conditional task, e.g., image captioning) are incorporated in the reverse diffusion process to produce intermediate samples guided in the Markov chain in specific directions to satisfy the desired condition. Prior works, such as [25, 26], have proposed similar methods in which the authors derive text- or class-conditioned samples from Generative Adversarial Networks (GANs) [11] that were trained without labels. To achieve this, they iteratively refine the noise input of the GAN until the desired sample satisfies the condition. Very recently, Grakios et al. [12] proposed a diffusion-based plug-and-play method that enables using unconditional diffusion models for conditional generation utilizing class labels. Both these methods are specifically designed for tasks involving label-level semantics. However, these methods do not address the usage of unconditional models for general image-to-image translation tasks, which require synthesizing visual content conditioned on fine-grained details in the source image. There are also works that propose diffusion models for image-to-image translation, such as for image super-resolution and inpainting [5, 20]; however, these methods are task-specific and do not generalize well to new tasks or new types of inverse problems (as demonstrated in Section 5). In this work, we present a generic framework that can generalize to any image-to-image translation task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we derive the necessary theory and formulate an algorithm, which we call Steered Diffusion, for diffusion-based image editing and image-to-image translation; our model is subsequently validated on a wide range of tasks. Steered diffusion is motivated by the energy-based" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 216 + ], + "type": "image", + "image_path": "c1801592f0ca9a188469e33ae593f7b8df6decc559534a3f1f23fad90c633641.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "lines": [ + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "text", + "content": "Figure 2. An illustration of the difference between existing plug-and-play generation approaches (e.g., [12]) and the proposed approach. Existing plug-and-play works operate with an energy function " + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "text", + "content": " of the noisy latent " + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "text", + "content": ". In contrast, our model uses the implicit prediction of the diffusion model (i.e., a coarse estimate of the clean image " + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "text", + "content": ") in its energy function " + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "text", + "content": ", which allows the use of any pre-trained network for steering. In addition, our model provides a looping mechanism " + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "inline_equation", + "content": "V_{2}" + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "text", + "content": ", which iterates " + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "text", + "content": " times at each timestep " + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 224, + 546, + 324 + ], + "type": "text", + "content": " to enhance generation quality." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 338, + 547, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 648 + ], + "type": "text", + "content": "formulation of diffusion probabilistic models [10]. In general, inference in a generative model can be thought of as deriving samples from a learned distribution. Recall that every probability density function can be formulated as an energy field that describes an unnormalized estimate of how the distribution density varies in space [13,25]. If one needs to find points in space that are the closest match to a given condition, one can utilize gradient-based optimization algorithms to find points in the field that have the highest density value for the condition. The gradient-based optimization scheme can be viewed as a modulation of the energy toward the desired direction. Previous work has utilized this idea on GANs [25, 26] and obtained reasonable results for label-based generation tasks. Due to their model structure, diffusion models are ideal candidates for such an energy modulation. One key challenge remains to design a good energy estimator that is robust to all noise levels. Previously, classifier-based guidance [8, 27] has been proposed and thought of as an energy modulation utilizing a pretrained classifier trained on noisy images. This poses a limitation that the guiding function should be noise-robust. In this work, we propose an alternative solution that does not need noise-robust networks but could use any network by utilizing the diffusion model as an implicit denoiser. Figure 2 gives a brief overview of how our approach is different from existing methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "We present experiments using steered diffusion on multiple conditional generative tasks on faces as well as generic images as portrayed in Figure 1. We present results on (i) identity replication [7], (ii) semantic image generation [29], (iii) linear inverse problems [21], and (iv) text-conditioned" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20851" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "image editing. Although our method is generic, for evaluations we perform experiments on faces. Before presenting our framework in detail, we now summarize the key contributions of our work:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 127, + 287, + 270 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 127, + 287, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 127, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 58, + 127, + 287, + 161 + ], + "type": "text", + "content": "- We propose steered diffusion, a general plug-and-play framework that can utilize various pre-existing models to steer an unconditional diffusion model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 163, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 163, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 163, + 287, + 198 + ], + "type": "text", + "content": "- We present the first work applicable to both label-level synthesis and image-to-image translation tasks and demonstrate its effectiveness for various applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 199, + 287, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 199, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 58, + 199, + 287, + 246 + ], + "type": "text", + "content": "- We propose an implicit conditioning-based sampling strategy that significantly boosts the performance of conditional sampling from unconditional diffusion models compared with previous methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 247, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 247, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 287, + 270 + ], + "type": "text", + "content": "- We introduce a new strategy that uses multiple steps of projected gradient descent to improve sample quality." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 281, + 126, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 126, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 126, + 294 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 300, + 135, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 135, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 135, + 312 + ], + "type": "text", + "content": "2.1. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 450 + ], + "type": "text", + "content": "Early works on unpaired image-to-image translation utilize a cycle consistency loss between the input and the target domains [9, 46]. Newer works, such as [16], have introduced a contrastive learning-based approach where a contrastive loss between corresponding patches of the input and target domains is minimized. The consistency-based method often fails to generate photorealistic images; hence, conditional generative models are preferred when labeled data are available. A few works [37, 38] utilize diffusion models for conditional image-to-image translation because of their photorealistic generation quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 451, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 288, + 714 + ], + "type": "text", + "content": "Guiding diffusion models during inference time has been explored by several works, such as [24]. The first method that proposed inference-time conditioning [8] uses a pretrained noise-robust classifier to guide the inference of an unconditional model. GLIDE [27] proposed a method for conditioning using text. Earlier work in plug-and-play modelling for generative models utilized GANs and performed iterative refinement on the latent space of GANs [25]. This method uses a predefined classifier or text captioning network to estimate a loss between the desired label output or text caption and the one generated from the GAN generator. This loss is backpropagated to refine the noise input of the GAN iteratively until the generator predicts the desired output. Recently, [12] proposed a method that uses diffusion models as a plug-and-play prior for class-conditioned generation. Several works have addressed the task of image-to-image translation using unconditional diffusion models [1,5,18,20], but each of these proposes a task-specific inference scheme. For example, ILVR [5] performs image super-resolution, and Reinpaint [20] performs image inpainting. Blended diffusion [1] proposes a method for text-conditioned image editing. DDRM [18] proposes an inference-time scheme" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "offering a general solution for linear inverse problems such as colorization and super-resolution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 103, + 411, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 103, + 411, + 114 + ], + "spans": [ + { + "bbox": [ + 306, + 103, + 411, + 114 + ], + "type": "text", + "content": "2.2. Concurrent Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 121, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 547, + 253 + ], + "type": "text", + "content": "In concurrent work to ours that also explores zero-shot conditional generation using diffusion models, [3] used a text to image model [34] and a two-step forward and backward universal guidance process, but it works well only after heavy optimization on network-based inverse problems such as semantic generation and identity generation. Another concurrent work [42] explored the usage of unconditional diffusion models for linear inverse problems using a pseudo-inverse model. In contrast to all these prior works, our steered diffusion algorithm generalizes well to both image-to-image translation tasks and high-level label-based generation tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 261, + 520, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 261, + 520, + 273 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 520, + 273 + ], + "type": "text", + "content": "2.3. Denoising Diffusion Probabilistic Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "text", + "content": "Denoising diffusion probabilistic models (DDPMs) [14, 40] belong to a class of generative models in which the model learns the distribution of data through a Markovian sampling process. DDPMs consist of a forward process and a reverse process. Let " + }, + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "text", + "content": " denote the latent state of an input image at timestep " + }, + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "text", + "content": " in a diffusion process. The sampling operation " + }, + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "text", + "content": " for the forward process in DDPM is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 371, + 545, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 371, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 339, + 371, + 545, + 386 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) := \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), \\tag {1}", + "image_path": "15fdb527099de67fdfbe4ffc3adb8e3064aa7a04806462c131d6532390b291c1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 394, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 394, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}" + }, + { + "bbox": [ + 304, + 394, + 545, + 418 + ], + "type": "text", + "content": " is a predefined variance schedule and " + }, + { + "bbox": [ + 304, + 394, + 545, + 418 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 394, + 545, + 418 + ], + "type": "text", + "content": " is the identity matrix." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "text", + "content": "The forward process can be considered as a noising operation, where the next state " + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "text", + "content": " is obtained from the current state " + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "text", + "content": " by adding a small amount of Gaussian noise according to the sampled timestep. The state " + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "text", + "content": " can also be sampled directly from the initial state " + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 304, + 418, + 547, + 478 + ], + "type": "text", + "content": ", using:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 344, + 487, + 545, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 487, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 344, + 487, + 545, + 501 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {0}\\right) := \\mathcal {N} \\left(x _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) I\\right), \\tag {2}", + "image_path": "412b30958975f1583eae6fae1b5da663d53f0899411efed6c7465d788237f0d5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 509, + 370, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 370, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 370, + 521 + ], + "type": "text", + "content": "or equivalently," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 339, + 529, + 545, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 529, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 339, + 529, + 545, + 543 + ], + "type": "interline_equation", + "content": "x _ {t} = x _ {0} \\sqrt {\\bar {\\alpha} _ {t}} + \\epsilon \\sqrt {1 - \\bar {\\alpha} _ {t}}, \\quad \\epsilon \\sim \\mathcal {N} (0, I), \\tag {3}", + "image_path": "ec8d5d01f082de9e4083efd9f0f0dfc4568fc205d5646e3ebb913f2470bac324.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 550, + 465, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 465, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 465, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 550, + 465, + 563 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s" + }, + { + "bbox": [ + 304, + 550, + 465, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 550, + 465, + 563 + ], + "type": "inline_equation", + "content": "\\alpha_t = 1 - \\beta_t" + }, + { + "bbox": [ + 304, + 550, + 465, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": "In [40], it is shown that if the number of time steps is large and the increment in " + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\{\\beta_i\\}" + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": " is small, then each step in the reverse sampling process can also be approximated by a Gaussian. If " + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}" + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": " respectively denote the mean and the covariance of this Gaussian, modelled via neural networks with parameters " + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": ", then each reverse step samples the state " + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": " according to:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 330, + 655, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 655, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 330, + 655, + 545, + 670 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) := \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right). \\tag {4}", + "image_path": "d82e50d9c1af38f305dec899351eeee5e2890c0e25fbabddebc5fad668304a80.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "text", + "content": "The parameters " + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 677, + 547, + 712 + ], + "type": "text", + "content": " are obtained by minimizing the variational lower bound of the negative log-likelihood of the data distribution." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "20852" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 360, + 209 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 360, + 209 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 360, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 360, + 209 + ], + "type": "image", + "image_path": "f0e06ea56d133f147010aea7036c8bdbb7ac94684f5fc7400b7ec82ea2f0386e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 210, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 547, + 245 + ], + "type": "text", + "content": "Figure 3. An illustration of Steered Diffusion. During each step of the sampling process the implicit prediction is steered to the direction of the condition using a steering network or predefined function. Note that this figure is only for illustrating the idea and does not show the actual sampled images; in the actual sampling, the steering process is much more gradual, not sudden as portrayed in this image." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 368, + 70, + 535, + 209 + ], + "blocks": [ + { + "bbox": [ + 368, + 70, + 535, + 209 + ], + "lines": [ + { + "bbox": [ + 368, + 70, + 535, + 209 + ], + "spans": [ + { + "bbox": [ + 368, + 70, + 535, + 209 + ], + "type": "image", + "image_path": "2e5c8ae020418efe7901e29213e5df5d82d4260eb0794667150e284e33207bd2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 249, + 153, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 249, + 153, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 153, + 262 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 269, + 237, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 237, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 237, + 282 + ], + "type": "text", + "content": "3.1. Steered Diffusion at Inference Time" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 287, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 289, + 324 + ], + "type": "text", + "content": "Our work is motivated by the energy-based formulation of diffusion models. For any probability density function, the corresponding energy-based model (EBM) is defined by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 340, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 340, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 113, + 340, + 287, + 365 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (x) = \\frac {\\exp (- V (x))}{Z}, \\tag {5}", + "image_path": "12a59ee767f43d01df99fc9e480c51d2f13cc3a053946cf496cf09e4b2b012d5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 372, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 372, + 287, + 445 + ], + "type": "inline_equation", + "content": "V(x)" + }, + { + "bbox": [ + 46, + 372, + 287, + 445 + ], + "type": "text", + "content": " denotes the corresponding energy function across states " + }, + { + "bbox": [ + 46, + 372, + 287, + 445 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 372, + 287, + 445 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 372, + 287, + 445 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 372, + 287, + 445 + ], + "type": "text", + "content": " denotes a normalization constant. To derive samples from this distribution, one can utilize the Langevin equation [39] describing the state transition of a particle in the presence of an energy field. For diffusion models the sampling step is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 464, + 287, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 287, + 477 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = x _ {t} - \\nabla_ {x _ {t}} \\log p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) + \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I). \\tag {6}", + "image_path": "aad60d4cb968df9bc58bdc2a6fd99918de4e70239dbb6a6f0d46b3a89583bdb9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "text", + "content": "The term " + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}\\log p_\\theta (x_{t - 1}|x_t)" + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "text", + "content": " is called the score function of the density " + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t)" + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "text", + "content": ". One key advantage of the energy-based formulation is that it allows modulation of the energy function to satisfy given criteria. This was initially introduced as classifier guidance [8], which allows label conditional sampling from an unconditionally trained diffusion model utilizing a noise-robust classifier. In the remaining part of this section, we motivate how we can extend the functionality of unconditional diffusion models to conditional tasks. Consider a conditional sampling scenario based on a condition " + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "text", + "content": ", for sampling from a state " + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "text", + "content": " to state " + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "text", + "content": ". The conditional transition probability " + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t,c)" + }, + { + "bbox": [ + 46, + 485, + 289, + 641 + ], + "type": "text", + "content": " can be decomposed as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 82, + 657, + 287, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 657, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 82, + 657, + 287, + 684 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, c\\right) \\propto \\frac {p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) p \\left(c \\mid x _ {t - 1}\\right)}{p \\left(c \\mid x _ {t}\\right)}. \\tag {7}", + "image_path": "9195bc20a2690b06f69cf33ccd363d4a4f2ee2ece4502267d297c601ea916031.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "content": "Hence, for any timestep " + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "content": ", the effective score for conditional transition can be found by utilizing using the log" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 251, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 275 + ], + "type": "text", + "content": "of probability density in the EBM formulations (5) of the individual densities and can be represented as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 282, + 416, + 296 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 282, + 416, + 296 + ], + "spans": [ + { + "bbox": [ + 312, + 282, + 416, + 296 + ], + "type": "interline_equation", + "content": "\\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t - 1} | x _ {t}, c) =", + "image_path": "9a0c70a9b23259414b7630ab4a76baa6a6ca5a8000cdc57ea7930e5bb832c752.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 297, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 545, + 322 + ], + "type": "interline_equation", + "content": "\\nabla_ {x _ {t}} \\log p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) - \\nabla_ {x _ {t}} V _ {1} \\left(x _ {t}, c\\right) + \\nabla_ {x _ {t}} V _ {2} \\left(x _ {t - 1}, c\\right), \\tag {8}", + "image_path": "b5c76a23855da688313a940c9b6c329ab1f51022a6e1eeed40abdf239cf53b3d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "inline_equation", + "content": "V_{2}" + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "content": " are the corresponding energy functions that model the conditional distributions of " + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "content": " given a condition " + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "content": ". Specifically, they project the higher dimensional " + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "content": " to the lower dimensional space of " + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "content": " and measure the distance between the mapped value and " + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 330, + 547, + 437 + ], + "type": "text", + "content": ". The better this particular measure, the more effectively it can be used to generate conditional samples from an unconditional model. Using Eq. (8), the conditional sampling equation for the reverse process is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 445, + 545, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 318, + 445, + 545, + 487 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} ^ {(t)} (x _ {t}, t)\\right) \\\\ - \\nabla_ {x _ {t}} V _ {1} (x _ {t}, c) + \\nabla_ {x _ {t}} V _ {2} (x _ {t - 1}, c) + \\sigma_ {t} \\epsilon . \\tag {9} \\\\ \\end{array}", + "image_path": "b4d45e8396084d2a868847e3d592cff2c46f4ba33b7f7feb6fe8c7ab8d171110.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{(t)}(x_t,t)" + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "text", + "content": " is the network prediction at a timestep " + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "text", + "content": " is the corresponding variance of the reverse step. The formulation in Eq. (9) shows that the energy function requires a functional mapping from a noisy " + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 496, + 547, + 546 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": "In many applications of interest, the mapping function from " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": " is complex and can be modeled effectively using deep networks. For example, in image-to-image translation tasks such as image generation from semantic maps [29], the image is " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": " and the semantic map is the condition " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": ". Similarly, for text to image generation, the image is " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": " corresponds to the text. Ideally, we would like to employ an existing (pre-trained) deep neural network for the mapping from " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": ". However, deep networks are usually trained on clean images, which limits the usability of existing pretrained deep networks for mapping directly from the noisy image " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": ". One workaround would be to use noise-robust networks to map from " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": ", but training noise-robust networks for conditional mapping can be computationally" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "20853" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": "expensive. Moreover, a network that is trained with multiple different noise levels often results in lower mapping performance, as it cannot denoise all noise levels accurately; we validate this claim experimentally in Section 5.6. Alternatively, one could include two mapping functions: a first that denoises " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": ", and a second that maps from the denoised image to " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": ". Rather than training a seperate denoising network, however, we realized that diffusion models are inherently trained as denoisers, and reconstruction quality improves as time proceeds in the reverse sampling of the diffusion process. Because of this capacity, we can use a reverse sampling step to make coarse predictions of the denoised image from any time step " + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 228, + 288, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 228, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 59, + 228, + 288, + 241 + ], + "type": "text", + "content": "Hence, we modify our original energy expression (8) to:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 262, + 288, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 262, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 58, + 262, + 288, + 302 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t - 1} | x _ {t}, c) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t - 1} | x _ {t}) \\\\ - \\nabla_ {x _ {t}} V _ {1} \\left(x _ {0 \\mid t}, c\\right) - \\delta_ {1} + \\nabla_ {x _ {t}} V _ {2} \\left(x _ {0 \\mid t - 1}, c\\right) + \\delta_ {2}, \\tag {10} \\\\ \\end{array}", + "image_path": "f137542b64418113fd5bac352daa84e9b0cddf25bb21aab15b43b9f02d49239d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 259, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 259, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 259, + 325 + ], + "type": "text", + "content": "where, we define the implicit step prediction " + }, + { + "bbox": [ + 47, + 312, + 259, + 325 + ], + "type": "inline_equation", + "content": "x_{0|t}" + }, + { + "bbox": [ + 47, + 312, + 259, + 325 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 103, + 335, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 335, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 103, + 335, + 287, + 365 + ], + "type": "interline_equation", + "content": "x _ {0 \\mid t} = \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (x _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {11}", + "image_path": "b8467f844cb86a76454ce538e4a109f8207fd106323d69c7f1898a5e30fbb744.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": "Here, we assume " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": " are first denoised to " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "x_{0|t}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "x_{0|t-1}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": ", respectively. The terms " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\delta_{2}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": " capture the errors arising from the shift in the domain from " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": " and from " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": "; for large " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "inline_equation", + "content": "\\delta_{1} \\approx \\delta_{2}" + }, + { + "bbox": [ + 47, + 373, + 287, + 433 + ], + "type": "text", + "content": " as the implicit predictions at nearby steps tend to be similar." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 434, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 289, + 613 + ], + "type": "text", + "content": "As shown in our experiments, the energy function should be selected according to the task. An easy way to choose the energy function is by looking at the training loss of the mapping network. For example, in the case of semantic generation, a good energy function is the cross-entropy loss between the predicted semantic map at any timestep and the input semantic map. In the case of identity replication, a good choice of regularization would be the negative cosine similarity score between the embeddings from a recognition network for the input and target image. In the case of text-to-image generation, it would be CLIP loss [32]. As a rule of thumb, an energy function could be chosen easily by looking at the loss function used to train the pre-trained network (or an inverse function) that maps from the image " + }, + { + "bbox": [ + 46, + 434, + 289, + 613 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 434, + 289, + 613 + ], + "type": "text", + "content": " to the condition " + }, + { + "bbox": [ + 46, + 434, + 289, + 613 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 434, + 289, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 622, + 259, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 259, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 259, + 635 + ], + "type": "text", + "content": "3.2. Revisiting Sampling in Diffusion Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "To obtain a closed-form expression for plugging the energy-based formulation into the reverse sampling process efficiently, we take inspiration from DDIM [41] and revisit the reverse sampling operation of diffusion models. From " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{1:T})" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": ", one can generate a sample " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " from a sample " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 306, + 85, + 528, + 277 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "type": "text", + "content": "Algorithm 1 Steered Diffusion" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "lines": [ + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": "Input: Energy function " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": ", condition " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(x_{T}; 0, I)" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "t = T - 1, \\ldots, 1" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " do \n3: for " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "n = N, \\ldots, 1" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " do \n4: " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\epsilon; 0, I)" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "x_{0|t} = \\frac{x_{t} - \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon_{\\theta}^{(t)}(x_{t})}{\\sqrt{\\bar{\\alpha}_{t}}}" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " \n6: Compute " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "x_{0|t}^{\\text{feas}}" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "V, c" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " using Eq. (15) \n7: if " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "(n > 1)" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " then \n8: Compute " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "x_{t}^{uc}" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " using Eq. (13) \n9: else \n10: Compute " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "x_{t-1}^{uc}" + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "text", + "content": " using Eq. (13) \n11: end if \n12: end for \n13: end for \n14: return " + }, + { + "bbox": [ + 306, + 85, + 528, + 277 + ], + "type": "inline_equation", + "content": "x_{0}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 325, + 314, + 546, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 314, + 546, + 397 + ], + "spans": [ + { + "bbox": [ + 325, + 314, + 546, + 397 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t - 1} ^ {\\mathrm {u c}} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\cdot \\underbrace {\\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} ^ {(t)} (x _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}}} _ {\\text {‘ p r e d i c t e d x _ {0} ”}} \\\\ + \\underbrace {\\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\cdot \\epsilon_ {\\theta} ^ {(t)} \\left(x _ {t}\\right)} _ {\\text {d i r e c t i o n p o i n t i n g t o} x _ {t} ”} + \\underbrace {\\sigma_ {t} \\epsilon} _ {\\text {r a n d o m n o i s e}}, \\tag {12} \\\\ \\end{array}", + "image_path": "8b8ec47b6c1fc3a24dac21d310401738bb6849c44abd4e6e64dd004b9be1707c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 403, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 545, + 428 + ], + "type": "text", + "content": "as in Song et al. [41]. Using Eq. (11), we can rewrite the unconditional sampling step Eq. (12) in terms of " + }, + { + "bbox": [ + 305, + 403, + 545, + 428 + ], + "type": "inline_equation", + "content": "x_{0|t}" + }, + { + "bbox": [ + 305, + 403, + 545, + 428 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 434, + 546, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 434, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 318, + 434, + 546, + 477 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t - 1} ^ {\\mathrm {u c}} = \\sqrt {\\bar {\\alpha} _ {t - 1}} x _ {0 | t} + \\\\ \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\cdot \\frac {x _ {t} - \\sqrt {\\bar {\\alpha} _ {t}} x _ {0 | t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} + \\sigma_ {t} \\epsilon , \\tag {13} \\\\ \\end{array}", + "image_path": "10a6cc7905bb632415419667c8a22d11d8a75a128a77e1d5e4faa5bfc43894eb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 483, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 545, + 529 + ], + "type": "text", + "content": "Here the superscript uc denotes the unconditional sample, which is obtained without any steering while transitioning from " + }, + { + "bbox": [ + 305, + 483, + 545, + 529 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 305, + 483, + 545, + 529 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 483, + 545, + 529 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 305, + 483, + 545, + 529 + ], + "type": "text", + "content": ". The conditional sampling step (8) can then be rewritten as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 537, + 545, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 545, + 552 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = x _ {t - 1} ^ {\\mathrm {u c}} - \\nabla_ {x _ {t}} V _ {1} \\left(x _ {0 \\mid t}, c\\right) + \\nabla_ {x _ {t}} V _ {2} \\left(x _ {0 \\mid t - 1}, c\\right). \\tag {14}", + "image_path": "ef12ad274a9c4e948500214cf9fd1014d6548459b7a6c0deaac62ad1a49cad73.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "text", + "content": "Through this, we can modulate " + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "text", + "content": " directly, as " + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "inline_equation", + "content": "x_{0|t}" + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "text", + "content": " is also a function of " + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{(t)}(x_t)" + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "text", + "content": ". Following Eq. (14), a rough estimate of the desired " + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "text", + "content": " for conditional sampling, represented by " + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "inline_equation", + "content": "x_{0|t}^{\\mathrm{feas}}" + }, + { + "bbox": [ + 305, + 558, + 545, + 611 + ], + "type": "text", + "content": ", can be obtained using" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 618, + 545, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 618, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 315, + 618, + 545, + 645 + ], + "type": "interline_equation", + "content": "x _ {0 | t} ^ {\\text {f e a s}} = x _ {0 | t} - k (t) \\nabla_ {x _ {t}} \\left(V _ {1} \\left(x _ {0 | t}, c\\right) - V _ {2} \\left(x _ {0 | t - 1}, c\\right)\\right), \\tag {15}", + "image_path": "31f89e650dc4244fb0e215730c097df943750c02554ced9de984187735240505.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "k(t)" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " is a scaling factor defining the strength of regularization. We call the process of finding " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{0|t}^{\\mathrm{feas}}" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "x_{0|t}" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " Implicit Steering Control (ISC), and we call the new sampling process steered diffusion. The exact algorithm is illustrated in Fig. 3 and explained in Algorithm 1." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "20854" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 70, + 242, + 199 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 242, + 199 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 242, + 199 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 242, + 199 + ], + "type": "image", + "image_path": "4e5898a5209a4adea25b9568a2695fdb4231e7516881f8935b027324939e0843.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 202, + 288, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 288, + 246 + ], + "type": "text", + "content": "Figure 4. An illustration of the steering function for linear inverse problems. For linear inverse problems, the component of the implicit prediction along the degradation direction can be replaced by the ground truth condition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 253, + 224, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 224, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 224, + 266 + ], + "type": "text", + "content": "4. Tips for Improved Performance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 273, + 185, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 273, + 185, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 273, + 185, + 285 + ], + "type": "text", + "content": "4.1. Linear Inverse Problems" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 292, + 288, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 288, + 448 + ], + "type": "text", + "content": "For optimization-based inverse problems such as text-to-image generation and semantic map to natural image generation, the exact mapping function is not always available. On the other hand, for linear inverse problems such as colorization, super-resolution, and image inpainting, the mapping is simply a linear function, and we can write Eq. (15) more simply. In these cases, the exact mapping function to the latent space of the condition is known. Hence, one can decompose the implicit prediction at each timestep along the direction of the condition and simply replace this component by its desired ideal condition, i.e., if our predicted sample needs to map to a condition " + }, + { + "bbox": [ + 46, + 292, + 288, + 448 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 292, + 288, + 448 + ], + "type": "text", + "content": ", then the modified implicit prediction step becomes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 456, + 287, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 484 + ], + "type": "interline_equation", + "content": "x _ {0 | t} ^ {\\text {f e a s}} = x _ {0 | t} + k (t) \\left(D (y) - D \\left(x _ {0 | t}\\right)\\right), \\quad \\text {w h e r e} c = D (y). \\tag {16}", + "image_path": "e2053e0eb4992a51289c41eee11d8211777372a0327e255649c7184234a78063.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "type": "text", + "content": " is the clean image and " + }, + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "type": "text", + "content": " is the known degradation model. Our sampling procedure ensures that the series of operations preserve the consistency of domains of " + }, + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "type": "inline_equation", + "content": "x_{0|t}^{\\mathrm{feas}}" + }, + { + "bbox": [ + 46, + 494, + 287, + 555 + ], + "type": "text", + "content": " to the original data distribution at the corresponding timesteps. An illustration is shown in Figure 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 217, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 217, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 217, + 575 + ], + "type": "text", + "content": "4.2. Multi-Step Implicit Modulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Our experiments show (see Fig. 7) that performing refinement using Eq. (15) on the implicit step prediction multiple times for each timestep significantly boosts the conditioning quality for more ill-posed conditions such as image inpainting and colorization. A similar observation was also found by [20]. Specifically, at a particular timestep " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ", we iterate the procedure of steering towards the next sampling step " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " and then adding noise to return to " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ". We present the corresponding algorithm in Algorithm 1. An example is shown in Fig. 7, where more realistic images are generated using the multiple-step sampling scheme row labeled \"OURS multi.\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": "Effectively, the " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "V_{2}" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " term in Eq. (12) can be thought of as enabling a multistep sampling in which we modulate the current step by looking ahead to the next sampling step. On a careful analysis of Eq. (7), i.e. from the score contribution due to the different regularization functions, we can see that the term for " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}V_1(x_t,c)" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " modulates " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " based on its current state, and the term from " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}V_2(x_{t - 1},c)" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " is a look-ahead correction where the derivative with respect to the future prediction is found. This is exactly what happens in the case of looping back from " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " is modulated iteratively by looking forward to what the future prediction would be." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 222, + 479, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 222, + 479, + 235 + ], + "spans": [ + { + "bbox": [ + 306, + 222, + 479, + 235 + ], + "type": "text", + "content": "4.3. Choosing the Scaling Factor " + }, + { + "bbox": [ + 306, + 222, + 479, + 235 + ], + "type": "inline_equation", + "content": "k(t)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": "In Eq. (15), " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "k(t)" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": " denotes the strength of the regularization constraint. A very small " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "k(t)" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": " would denote no effective regularization, and a large " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": " would lead to the diffusion process going out of the latent space manifold. Since the derivative of the regularization function by itself is a score value, similar to the normal scaling value of the score function, the appropriate time-varying normalization factor is " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\sqrt{1 - \\bar{\\alpha}_t}" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": ". The exact value for " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "k(t)" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": " for each task is defined in Table 1. For linear inverse problems we use a constant " + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "inline_equation", + "content": "k(t) = 1" + }, + { + "bbox": [ + 304, + 240, + 547, + 361 + ], + "type": "text", + "content": ", which provided the best results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 369, + 387, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 387, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 387, + 383 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 389, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 547, + 629 + ], + "type": "text", + "content": "We evaluate the performance of our network qualitatively and quantitatively using four image-to-image translation tasks—semantic layout to face image translation, face inpainting, face colorization, and face super-resolution—as well as two high-level vision tasks: identity-based image generation and text-guided image editing. Unlike existing approaches, our method is completely zero-shot and applies to a wide variety of tasks. We compare with other diffusion-based approaches best applicable for each task for a fair evaluation. We also compare the semantic layout to image translation performance with that of task-specific unsupervised methods. We choose the unconditional model released by [5] as the unconditional pretrained diffusion model in all of our experiments with faces. Note that the sampling scheme in ILVR [5] and Repaint [20] can be thought of as happening at time " + }, + { + "bbox": [ + 304, + 389, + 547, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 389, + 547, + 629 + ], + "type": "text", + "content": " rather than at the implicit step as in our method. Hence, comparing these methods for super-resolution and inpainting in our experiments below can be considered an additional ablation study highlighting the improvement from our implicit sampling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 635, + 439, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 439, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 439, + 647 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "For our experiments, we utilize pixel-level unconditional diffusion models. For faces, we utilize the model trained on the FFHQ dataset [17] that was released by [5]. For generic images, we use the model trained on ImageNet [6] released in ADM [8]. All our experiments use 100 steps of sampling." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "20855" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 64, + 538, + 139 + ], + "blocks": [ + { + "bbox": [ + 57, + 64, + 538, + 139 + ], + "lines": [ + { + "bbox": [ + 57, + 64, + 538, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 64, + 538, + 139 + ], + "type": "table", + "html": "
MethodLinear inverse problemsComplex inverse problems
ColorizationInpaintingSuper-resolutionSemantic GenerationIdentity ReplicationImage Editing
V1(D(x0|t) - c)2(D(x0|t) - c)2(D(x0|t) - c)2BCE(D(x0),c)1 - D(x0|t).D(c)3(D1(x0|t),D1(c1))2 + 2000CS(D2(x0|t),c2)
V2(D(x0|t-1) - c)2(D(x0|t-1) - c)20000
DGrayscaleMaskingDownscalingFARL [45]FARL recognizer [45]VGG Face [30] & FARL CLIP [32]
N331111
k(t)1112000√1-αt3000√1-αt10√1-αt & 1500√1-αt
", + "image_path": "9108f10b96e8b300083c4dd80ebd8ddbd145f0df0f753b2d635d6751fb597ce2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 217, + 143, + 375, + 154 + ], + "lines": [ + { + "bbox": [ + 217, + 143, + 375, + 154 + ], + "spans": [ + { + "bbox": [ + 217, + 143, + 375, + 154 + ], + "type": "text", + "content": "Table 1. Parameter set for each application." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 172, + 285, + 322 + ], + "blocks": [ + { + "bbox": [ + 49, + 172, + 285, + 322 + ], + "lines": [ + { + "bbox": [ + 49, + 172, + 285, + 322 + ], + "spans": [ + { + "bbox": [ + 49, + 172, + 285, + 322 + ], + "type": "image", + "image_path": "70bca2bf639c04e835e046524b143f7e51da70c6cd9ca0cc437d0ff13fd9d523.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 323, + 273, + 334 + ], + "lines": [ + { + "bbox": [ + 59, + 323, + 273, + 334 + ], + "spans": [ + { + "bbox": [ + 59, + 323, + 273, + 334 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons for semantic generation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 342, + 290, + 529 + ], + "blocks": [ + { + "bbox": [ + 48, + 342, + 290, + 529 + ], + "lines": [ + { + "bbox": [ + 48, + 342, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 290, + 529 + ], + "type": "image", + "image_path": "d5860c116ea3730c61b86594e20ad8df02348eb991d591d82264a81b00f4ff28.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 533, + 242, + 544 + ], + "lines": [ + { + "bbox": [ + 91, + 533, + 242, + 544 + ], + "spans": [ + { + "bbox": [ + 91, + 533, + 242, + 544 + ], + "type": "text", + "content": "Figure 6. Results on " + }, + { + "bbox": [ + 91, + 533, + 242, + 544 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 91, + 533, + 242, + 544 + ], + "type": "text", + "content": " super resolution." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 549, + 192, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 192, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 192, + 561 + ], + "type": "text", + "content": "5.2. Semantic Face Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 568, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 288, + 687 + ], + "type": "text", + "content": "To evaluate how our method performs in generic image-to-image translation tasks, we evaluate our method's performance for the task of semantic layout to face generation. We utilize the CelebA dataset for this. To generate the semantic labels, we use facer [45] and create 11 label classes for each face. Since there is no other unconditional model that can perform fully test-time semantic generation, to evaluate the performance of our method, we compare with fully unsupervised image translation methods: CycleGAN [46], CUT [28], and ILVR [5]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "The corresponding qualitative results are shown in Fig. 5. It is clear that CUT, CycleGAN, and ILVR produce unrealis-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 172, + 551, + 360 + ], + "blocks": [ + { + "bbox": [ + 307, + 172, + 551, + 360 + ], + "lines": [ + { + "bbox": [ + 307, + 172, + 551, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 551, + 360 + ], + "type": "image", + "image_path": "654e90b6fef1a0f14d94c6c9d443918f73b16fc15334d90fbab349901832da28.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 369, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 401 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons for colorization. The row labeled \"OURS multi\" refers to the use of multi-step sampling, as described in Sec. 4.2." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 419, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 419, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 547, + 491 + ], + "type": "text", + "content": "tic facial images with huge artifacts or create low-resolution faces. In contrast, our method always creates good-quality realistic faces. We present the quantitative results in Table 2. The table shows that our method obtains the best FID scores of all methods and obtains the best mIOU score among the inference-time techniques." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 502, + 434, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 434, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 434, + 514 + ], + "type": "text", + "content": "5.3. Face Super-Resolution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": "We evaluate the performance for the face super-resolution task using the CelebA dataset [19]. As baselines, we utilize fully inference-time methods in which no task-specific training is used. As the first baseline, we choose PULSE [22], a self-supervised upsampling technique utilizing GANs. As the next comparison method, we choose ILVR [5] which, like our method, performs super-resolution utilizing an unconditional pre-trained diffusion model. However, in ILVR, sampling happens at timestep " + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": " rather than at the implicit step in our algorithm. In total, we utilize 300 images for evaluations. We present some qualitative results in Fig. 6. For ILVR [5], we utilize 100 timesteps of sampling, the same as in our case. As we can see, PULSE [22] and ILVR [5] are unable to restore the correct identity and also contain blur artifacts after restoration. On the other hand, steered diffusion (our method) is able to restore photorealistic facial" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "text", + "content": "20856" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 104, + 224 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 104, + 224 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 104, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 104, + 224 + ], + "type": "image", + "image_path": "c32960c85d337eea0ed8f47ad4c429c5a8fd92905cea08a253cc3f26dd7f8727.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 224, + 94, + 232 + ], + "lines": [ + { + "bbox": [ + 65, + 224, + 94, + 232 + ], + "spans": [ + { + "bbox": [ + 65, + 224, + 94, + 232 + ], + "type": "text", + "content": "Degraded" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 105, + 70, + 157, + 224 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 157, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 157, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 157, + 224 + ], + "type": "image", + "image_path": "602d5d8b76d6172338380de68d8a08d359761e2b0fe212f717fddf7b169ce31f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 224, + 145, + 232 + ], + "lines": [ + { + "bbox": [ + 116, + 224, + 145, + 232 + ], + "spans": [ + { + "bbox": [ + 116, + 224, + 145, + 232 + ], + "type": "text", + "content": "Reinpaint" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 158, + 70, + 209, + 224 + ], + "blocks": [ + { + "bbox": [ + 158, + 70, + 209, + 224 + ], + "lines": [ + { + "bbox": [ + 158, + 70, + 209, + 224 + ], + "spans": [ + { + "bbox": [ + 158, + 70, + 209, + 224 + ], + "type": "image", + "image_path": "802d6320a6cddebb523b7a0d87582acc75e8c6d72f5adbe7863ad3bfbfbede4e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 224, + 190, + 232 + ], + "lines": [ + { + "bbox": [ + 176, + 224, + 190, + 232 + ], + "spans": [ + { + "bbox": [ + 176, + 224, + 190, + 232 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 217, + 70, + 269, + 224 + ], + "blocks": [ + { + "bbox": [ + 217, + 70, + 269, + 224 + ], + "lines": [ + { + "bbox": [ + 217, + 70, + 269, + 224 + ], + "spans": [ + { + "bbox": [ + 217, + 70, + 269, + 224 + ], + "type": "image", + "image_path": "dfd65f5801bb9fc4af5f6616bee6ed63015fd6946eb1ec7bb29f3c0685e1daa1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 224, + 258, + 232 + ], + "lines": [ + { + "bbox": [ + 229, + 224, + 258, + 232 + ], + "spans": [ + { + "bbox": [ + 229, + 224, + 258, + 232 + ], + "type": "text", + "content": "Degraded" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 270, + 70, + 321, + 224 + ], + "blocks": [ + { + "bbox": [ + 270, + 70, + 321, + 224 + ], + "lines": [ + { + "bbox": [ + 270, + 70, + 321, + 224 + ], + "spans": [ + { + "bbox": [ + 270, + 70, + 321, + 224 + ], + "type": "image", + "image_path": "030f5d602bb1d81bf21a81ea1db0b0d79cf2c0b272484358566032d6c0096008.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 224, + 309, + 232 + ], + "lines": [ + { + "bbox": [ + 282, + 224, + 309, + 232 + ], + "spans": [ + { + "bbox": [ + 282, + 224, + 309, + 232 + ], + "type": "text", + "content": "Reinpaint" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 373, + 224 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 373, + 224 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 373, + 224 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 373, + 224 + ], + "type": "image", + "image_path": "57b700bc89d80244a40413e5ff38da989002a58b0bb533458d34953ba1cb55a1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 224, + 354, + 232 + ], + "lines": [ + { + "bbox": [ + 340, + 224, + 354, + 232 + ], + "spans": [ + { + "bbox": [ + 340, + 224, + 354, + 232 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 382, + 70, + 433, + 224 + ], + "blocks": [ + { + "bbox": [ + 382, + 70, + 433, + 224 + ], + "lines": [ + { + "bbox": [ + 382, + 70, + 433, + 224 + ], + "spans": [ + { + "bbox": [ + 382, + 70, + 433, + 224 + ], + "type": "image", + "image_path": "725dca099ff18d343af72eae1e33dab367deccaf1cf95a4de63e2c8192894d94.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 224, + 422, + 232 + ], + "lines": [ + { + "bbox": [ + 393, + 224, + 422, + 232 + ], + "spans": [ + { + "bbox": [ + 393, + 224, + 422, + 232 + ], + "type": "text", + "content": "Degraded" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 485, + 224 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 485, + 224 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 485, + 224 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 485, + 224 + ], + "type": "image", + "image_path": "813654a93234b9249024a775f851c410b488edf9211d53f6c9b140703d38e5b1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 224, + 474, + 232 + ], + "lines": [ + { + "bbox": [ + 446, + 224, + 474, + 232 + ], + "spans": [ + { + "bbox": [ + 446, + 224, + 474, + 232 + ], + "type": "text", + "content": "Reinpaint" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 486, + 70, + 537, + 224 + ], + "blocks": [ + { + "bbox": [ + 486, + 70, + 537, + 224 + ], + "lines": [ + { + "bbox": [ + 486, + 70, + 537, + 224 + ], + "spans": [ + { + "bbox": [ + 486, + 70, + 537, + 224 + ], + "type": "image", + "image_path": "ae4bd3a4ec785891155b8b122e4de4938853850135c8bb07a64f69e26ad3409d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 224, + 518, + 232 + ], + "lines": [ + { + "bbox": [ + 504, + 224, + 518, + 232 + ], + "spans": [ + { + "bbox": [ + 504, + 224, + 518, + 232 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 54, + 251, + 123, + 319 + ], + "blocks": [ + { + "bbox": [ + 54, + 251, + 123, + 319 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 123, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 123, + 319 + ], + "type": "image", + "image_path": "51aeb79b549932b0b45d635ce0b8c6c65877f3309c074c7aebda79c8910270bc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 321, + 111, + 330 + ], + "lines": [ + { + "bbox": [ + 66, + 321, + 111, + 330 + ], + "spans": [ + { + "bbox": [ + 66, + 321, + 111, + 330 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 124, + 251, + 191, + 319 + ], + "blocks": [ + { + "bbox": [ + 145, + 234, + 447, + 244 + ], + "lines": [ + { + "bbox": [ + 145, + 234, + 447, + 244 + ], + "spans": [ + { + "bbox": [ + 145, + 234, + 447, + 244 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons for inpainting for thin, medium and thick masks." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 124, + 251, + 191, + 319 + ], + "lines": [ + { + "bbox": [ + 124, + 251, + 191, + 319 + ], + "spans": [ + { + "bbox": [ + 124, + 251, + 191, + 319 + ], + "type": "image", + "image_path": "7421b32858f0af6c759bce17f41674721722c5940f5f96c29daf50485b6caae1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 321, + 190, + 330 + ], + "lines": [ + { + "bbox": [ + 126, + 321, + 190, + 330 + ], + "spans": [ + { + "bbox": [ + 126, + 321, + 190, + 330 + ], + "type": "text", + "content": "Photo of a young man" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 340, + 411, + 351 + ], + "lines": [ + { + "bbox": [ + 181, + 340, + 411, + 351 + ], + "spans": [ + { + "bbox": [ + 181, + 340, + 411, + 351 + ], + "type": "text", + "content": "Figure 9. Qualitative comparisons for text-based image editing." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 194, + 251, + 261, + 319 + ], + "blocks": [ + { + "bbox": [ + 194, + 251, + 261, + 319 + ], + "lines": [ + { + "bbox": [ + 194, + 251, + 261, + 319 + ], + "spans": [ + { + "bbox": [ + 194, + 251, + 261, + 319 + ], + "type": "image", + "image_path": "ee5c2da4602c1a169b122944814d8dfb3ebe342cbf2a4421c958e0fab9b9907c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 321, + 253, + 330 + ], + "lines": [ + { + "bbox": [ + 201, + 321, + 253, + 330 + ], + "spans": [ + { + "bbox": [ + 201, + 321, + 253, + 330 + ], + "type": "text", + "content": "She has wavy hair" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 263, + 251, + 330, + 319 + ], + "blocks": [ + { + "bbox": [ + 263, + 251, + 330, + 319 + ], + "lines": [ + { + "bbox": [ + 263, + 251, + 330, + 319 + ], + "spans": [ + { + "bbox": [ + 263, + 251, + 330, + 319 + ], + "type": "image", + "image_path": "a3872586b42c28021953312560a0dfdce680576be43ea63ccacdd72d20c6d15b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 321, + 325, + 329 + ], + "lines": [ + { + "bbox": [ + 269, + 321, + 325, + 329 + ], + "spans": [ + { + "bbox": [ + 269, + 321, + 325, + 329 + ], + "type": "text", + "content": "She has blonde hair" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 332, + 251, + 400, + 319 + ], + "blocks": [ + { + "bbox": [ + 332, + 251, + 400, + 319 + ], + "lines": [ + { + "bbox": [ + 332, + 251, + 400, + 319 + ], + "spans": [ + { + "bbox": [ + 332, + 251, + 400, + 319 + ], + "type": "image", + "image_path": "92f90481741da69482c90776ffe37526a062ab53b03b9b81e0c7e52012818997.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 321, + 399, + 329 + ], + "lines": [ + { + "bbox": [ + 333, + 321, + 399, + 329 + ], + "spans": [ + { + "bbox": [ + 333, + 321, + 399, + 329 + ], + "type": "text", + "content": "Photo of an old woman" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 402, + 251, + 469, + 319 + ], + "blocks": [ + { + "bbox": [ + 402, + 251, + 469, + 319 + ], + "lines": [ + { + "bbox": [ + 402, + 251, + 469, + 319 + ], + "spans": [ + { + "bbox": [ + 402, + 251, + 469, + 319 + ], + "type": "image", + "image_path": "100df6cfac2b29c71d71547141a2ac171c4b8b3564748f8bced0ab66d51925da.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 321, + 453, + 330 + ], + "lines": [ + { + "bbox": [ + 417, + 321, + 453, + 330 + ], + "spans": [ + { + "bbox": [ + 417, + 321, + 453, + 330 + ], + "type": "text", + "content": "She is angry." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 471, + 251, + 539, + 319 + ], + "blocks": [ + { + "bbox": [ + 471, + 251, + 539, + 319 + ], + "lines": [ + { + "bbox": [ + 471, + 251, + 539, + 319 + ], + "spans": [ + { + "bbox": [ + 471, + 251, + 539, + 319 + ], + "type": "image", + "image_path": "de93655a1478376b937650bcf3dfb0024126ada5cde7324d976d4060020daa2a.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 321, + 519, + 329 + ], + "lines": [ + { + "bbox": [ + 490, + 321, + 519, + 329 + ], + "spans": [ + { + "bbox": [ + 490, + 321, + 519, + 329 + ], + "type": "text", + "content": "She is sad." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 50, + 359, + 295, + 400 + ], + "blocks": [ + { + "bbox": [ + 50, + 359, + 295, + 400 + ], + "lines": [ + { + "bbox": [ + 50, + 359, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 50, + 359, + 295, + 400 + ], + "type": "table", + "html": "
Trained MethodsInference-Time Methods
MethodCUT [16]GCGAN [9]CycleGAN [46]ILVR [5]Ours
FID (↓)49.3458.8052.70107.4641.38
mIoU(↑)0.6470.6990.7230.3770.532
", + "image_path": "100e40b3af285c85e296194374b6bcabd540bf43c0c62b93f4f6e6e9fae2d751.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_body" + } + ], + "index": 34 + }, + { + "type": "table", + "bbox": [ + 62, + 427, + 274, + 487 + ], + "blocks": [ + { + "bbox": [ + 70, + 403, + 261, + 414 + ], + "lines": [ + { + "bbox": [ + 70, + 403, + 261, + 414 + ], + "spans": [ + { + "bbox": [ + 70, + 403, + 261, + 414 + ], + "type": "text", + "content": "Table 2. Quantitative results for semantic generation" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 427, + 274, + 487 + ], + "lines": [ + { + "bbox": [ + 62, + 427, + 274, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 274, + 487 + ], + "type": "table", + "html": "
MethodFID ↓LPIPS↓NIQE↓PSNR↑SSIM↑
Bicubic130.30.441912.0323.850.642
ILVR [5]62.240.41647.3820.540.5527
PULSE [22]84.670.43655.0421.080.5285
Ours51.190.25939.2426.020.711
", + "image_path": "db49e1db65362909bebd5256b8646c98aec70afb8e515c75a94de035853b29b9.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "table_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 516, + 288, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 288, + 564 + ], + "type": "text", + "content": "images. The qualitative evaluations are presented in Table 3; our method yields a 0.18 improvement in perceptual similarity, 6.95 dB improvement in PSNR, and 0.24 improvement in SSIM versus all of the other comparison methods." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 47, + 574, + 152, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 152, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 152, + 586 + ], + "type": "text", + "content": "5.4. Face Colorization" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "As a baseline method, we modify ILVR [5] to suit the task of colorization. For this, rather than performing the constraint at every step, we start the sampling process from a noised grayscale image and enforce consistency between the generated and original grayscale images. In total, we utilize 300 images for evaluation. The corresponding results can be seen in Fig. 7. As we can see, our method is able to reconstruct photorealistic faces with naturalistic colours compared to ILVR [5]. The corresponding quantitative metrics are presented in Table. 4. We get a significant boost in" + } + ] + } + ], + "index": 40 + }, + { + "type": "table", + "bbox": [ + 340, + 364, + 515, + 426 + ], + "blocks": [ + { + "bbox": [ + 78, + 491, + 255, + 501 + ], + "lines": [ + { + "bbox": [ + 78, + 491, + 255, + 501 + ], + "spans": [ + { + "bbox": [ + 78, + 491, + 255, + 501 + ], + "type": "text", + "content": "Table 3. Quantitative results for super-resolution" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 364, + 515, + 426 + ], + "lines": [ + { + "bbox": [ + 340, + 364, + 515, + 426 + ], + "spans": [ + { + "bbox": [ + 340, + 364, + 515, + 426 + ], + "type": "table", + "html": "
MethodFID ↓LPIPS↓NIQE↓
Grayscale69.270.27815.07
ILVR [5]67.660.52707.54
Ours49.720.33115.91
", + "image_path": "a77a2fc208928cff06a693efbac9af639fd5c943d8bf7d2df89e6c871f50402c.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 345, + 434, + 506, + 445 + ], + "lines": [ + { + "bbox": [ + 345, + 434, + 506, + 445 + ], + "spans": [ + { + "bbox": [ + 345, + 434, + 506, + 445 + ], + "type": "text", + "content": "Table 4. Quantitative results for colorization" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 468, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 545, + 491 + ], + "type": "text", + "content": "performance, with an FID score of 19, LPIPS [44] score of 0.19, and NIQE [23] score of 1.5 for our method." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 306, + 509, + 545, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 545, + 522 + ], + "type": "text", + "content": "5.5. Inpainting, Image Editing, Identity replication" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 530, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 547, + 649 + ], + "type": "text", + "content": "Our method can also utilize multiple conditions simultaneously; we provide an illustration in Figure 9 where we condition with an identity-preserving network and a text caption simultaneously. From the figure, one can see that our method can generalize well to a diverse range of captions. For preserving identity, we used the VGGFace network [30]. We utilize FARL [45], which is pre-trained with face and corresponding text pairs to enforce the captions. For generic identity replication as in Figure 1 we use FARL face embedder." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "For our image inpainting experiments, we use the subset released by [43] and evaluate three different kinds of masks. Our method obtains better results than existing baselines across all mask variations. Qualitative and quantitative results are shown in Fig. 8 and Table 5, respectively." + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "20857" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 65, + 560, + 136 + ], + "blocks": [ + { + "bbox": [ + 50, + 65, + 560, + 136 + ], + "lines": [ + { + "bbox": [ + 50, + 65, + 560, + 136 + ], + "spans": [ + { + "bbox": [ + 50, + 65, + 560, + 136 + ], + "type": "table", + "html": "
MethodThinMediumThick
FID ↓LPIPS↓PSNR↑SSIM↑FID ↓LPIPS↓PSNR↑SSIM↑FID ↓LPIPS↓PSNR↑SSIM↑
Degraded371.030.6767.8470.201258.860.6247.5270.228231.930.5857.6060.249
Reinpaint [20]43.350.30417.990.67153.710.39913.710.55852.400.40712.780.530
Ours30.850.18324.920.83335.390.22021.550.78640.120.24218.870.705
", + "image_path": "52a4b99ca9b91a9892f50ff9737cc4928aa75d2a42abb215ad19dd8141fcd039.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 217, + 139, + 374, + 150 + ], + "lines": [ + { + "bbox": [ + 217, + 139, + 374, + 150 + ], + "spans": [ + { + "bbox": [ + 217, + 139, + 374, + 150 + ], + "type": "text", + "content": "Table 5. Quantitative results for inpainting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 61, + 155, + 127, + 221 + ], + "blocks": [ + { + "bbox": [ + 61, + 155, + 127, + 221 + ], + "lines": [ + { + "bbox": [ + 61, + 155, + 127, + 221 + ], + "spans": [ + { + "bbox": [ + 61, + 155, + 127, + 221 + ], + "type": "image", + "image_path": "95dfe810af322fca7b4b1bac59f8a960198de272be940e59b7fcd3e232d7ba1b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 224, + 116, + 232 + ], + "lines": [ + { + "bbox": [ + 74, + 224, + 116, + 232 + ], + "spans": [ + { + "bbox": [ + 74, + 224, + 116, + 232 + ], + "type": "text", + "content": "Semantic Map" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 129, + 156, + 194, + 221 + ], + "blocks": [ + { + "bbox": [ + 129, + 156, + 194, + 221 + ], + "lines": [ + { + "bbox": [ + 129, + 156, + 194, + 221 + ], + "spans": [ + { + "bbox": [ + 129, + 156, + 194, + 221 + ], + "type": "image", + "image_path": "6ef6bed403bad1cc2de2f0a937e2e1cc49054aedbbd015f769b5a7223479b43a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 224, + 173, + 232 + ], + "lines": [ + { + "bbox": [ + 151, + 224, + 173, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 224, + 173, + 232 + ], + "type": "inline_equation", + "content": "K = 0" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 196, + 156, + 261, + 221 + ], + "blocks": [ + { + "bbox": [ + 196, + 156, + 261, + 221 + ], + "lines": [ + { + "bbox": [ + 196, + 156, + 261, + 221 + ], + "spans": [ + { + "bbox": [ + 196, + 156, + 261, + 221 + ], + "type": "image", + "image_path": "5a8da2c8b25ec356836003d0422438696a32eaf74994819f738bd615c3f14965.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 224, + 244, + 232 + ], + "lines": [ + { + "bbox": [ + 214, + 224, + 244, + 232 + ], + "spans": [ + { + "bbox": [ + 214, + 224, + 244, + 232 + ], + "type": "inline_equation", + "content": "K = 200" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 263, + 156, + 328, + 221 + ], + "blocks": [ + { + "bbox": [ + 263, + 156, + 328, + 221 + ], + "lines": [ + { + "bbox": [ + 263, + 156, + 328, + 221 + ], + "spans": [ + { + "bbox": [ + 263, + 156, + 328, + 221 + ], + "type": "image", + "image_path": "39776518c35ced6dadd9567bbc7d63f30cd034ab6eb129a6bef2efd734b7e7c9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 224, + 313, + 232 + ], + "lines": [ + { + "bbox": [ + 279, + 224, + 313, + 232 + ], + "spans": [ + { + "bbox": [ + 279, + 224, + 313, + 232 + ], + "type": "inline_equation", + "content": "K = 2000" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 329, + 156, + 395, + 221 + ], + "blocks": [ + { + "bbox": [ + 329, + 156, + 395, + 221 + ], + "lines": [ + { + "bbox": [ + 329, + 156, + 395, + 221 + ], + "spans": [ + { + "bbox": [ + 329, + 156, + 395, + 221 + ], + "type": "image", + "image_path": "9fe8229619d4dabac5ceb27c0dfe7c01e720f985c4e2d9bf8bcad14d34c4e437.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 224, + 381, + 232 + ], + "lines": [ + { + "bbox": [ + 343, + 224, + 381, + 232 + ], + "spans": [ + { + "bbox": [ + 343, + 224, + 381, + 232 + ], + "type": "inline_equation", + "content": "K = 20000" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 395, + 156, + 462, + 221 + ], + "blocks": [ + { + "bbox": [ + 395, + 156, + 462, + 221 + ], + "lines": [ + { + "bbox": [ + 395, + 156, + 462, + 221 + ], + "spans": [ + { + "bbox": [ + 395, + 156, + 462, + 221 + ], + "type": "image", + "image_path": "45b50e4c730c9d0c9cc337074bdbdb5a545e74a5c8477d06b51b595a63c26a8d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 224, + 451, + 232 + ], + "lines": [ + { + "bbox": [ + 407, + 224, + 451, + 232 + ], + "spans": [ + { + "bbox": [ + 407, + 224, + 451, + 232 + ], + "type": "inline_equation", + "content": "K = 2\\times 10^{5}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 463, + 156, + 529, + 221 + ], + "blocks": [ + { + "bbox": [ + 463, + 156, + 529, + 221 + ], + "lines": [ + { + "bbox": [ + 463, + 156, + 529, + 221 + ], + "spans": [ + { + "bbox": [ + 463, + 156, + 529, + 221 + ], + "type": "image", + "image_path": "f3a11d08557b20d4c6b65b0bf11c871cf948fc5c452a16692fd262a6f45f9402.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 224, + 518, + 232 + ], + "lines": [ + { + "bbox": [ + 474, + 224, + 518, + 232 + ], + "spans": [ + { + "bbox": [ + 474, + 224, + 518, + 232 + ], + "type": "inline_equation", + "content": "K = 2\\times 10^{6}" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 82, + 250, + 137, + 306 + ], + "blocks": [ + { + "bbox": [ + 146, + 234, + 446, + 245 + ], + "lines": [ + { + "bbox": [ + 146, + 234, + 446, + 245 + ], + "spans": [ + { + "bbox": [ + 146, + 234, + 446, + 245 + ], + "type": "text", + "content": "Figure 10. Figure showing sample variation with scaling factor " + }, + { + "bbox": [ + 146, + 234, + 446, + 245 + ], + "type": "inline_equation", + "content": "k(t) = K\\sqrt{1 - \\bar{\\alpha}_t}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 250, + 137, + 306 + ], + "lines": [ + { + "bbox": [ + 82, + 250, + 137, + 306 + ], + "spans": [ + { + "bbox": [ + 82, + 250, + 137, + 306 + ], + "type": "image", + "image_path": "460407a823d71c159d1e69645223c962e949711ee7975ac5c707b604921e3286.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 82, + 306, + 137, + 361 + ], + "blocks": [ + { + "bbox": [ + 82, + 306, + 137, + 361 + ], + "lines": [ + { + "bbox": [ + 82, + 306, + 137, + 361 + ], + "spans": [ + { + "bbox": [ + 82, + 306, + 137, + 361 + ], + "type": "image", + "image_path": "c28bc45cccfa5c4f281514b4728aa05ac4da72c8a66c2d0b7645905b390a7a20.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 363, + 124, + 371 + ], + "lines": [ + { + "bbox": [ + 95, + 363, + 124, + 371 + ], + "spans": [ + { + "bbox": [ + 95, + 363, + 124, + 371 + ], + "type": "text", + "content": "Degraded" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 373, + 262, + 384 + ], + "lines": [ + { + "bbox": [ + 70, + 373, + 262, + 384 + ], + "spans": [ + { + "bbox": [ + 70, + 373, + 262, + 384 + ], + "type": "text", + "content": "Figure 11. Qualitative comparisons for Colorization." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 138, + 250, + 194, + 306 + ], + "blocks": [ + { + "bbox": [ + 138, + 250, + 194, + 306 + ], + "lines": [ + { + "bbox": [ + 138, + 250, + 194, + 306 + ], + "spans": [ + { + "bbox": [ + 138, + 250, + 194, + 306 + ], + "type": "image", + "image_path": "1958e5e661498a87cd4213ae5721fe08dde99c3bac20488d6167fbfc1961ccaf.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 138, + 306, + 194, + 361 + ], + "blocks": [ + { + "bbox": [ + 138, + 306, + 194, + 361 + ], + "lines": [ + { + "bbox": [ + 138, + 306, + 194, + 361 + ], + "spans": [ + { + "bbox": [ + 138, + 306, + 194, + 361 + ], + "type": "image", + "image_path": "21c2dd39d7514dc2a2819893e1eec540787b7b3010f1f6d08ef02c82810c38d8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 363, + 184, + 371 + ], + "lines": [ + { + "bbox": [ + 148, + 363, + 184, + 371 + ], + "spans": [ + { + "bbox": [ + 148, + 363, + 184, + 371 + ], + "type": "text", + "content": "Noise robust" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 195, + 250, + 251, + 361 + ], + "blocks": [ + { + "bbox": [ + 195, + 250, + 251, + 361 + ], + "lines": [ + { + "bbox": [ + 195, + 250, + 251, + 361 + ], + "spans": [ + { + "bbox": [ + 195, + 250, + 251, + 361 + ], + "type": "image", + "image_path": "aa819edcc972238ffeafe4650c8411ca86b50fabf1b570e7454ccceeebea724e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 363, + 233, + 371 + ], + "lines": [ + { + "bbox": [ + 213, + 363, + 233, + 371 + ], + "spans": [ + { + "bbox": [ + 213, + 363, + 233, + 371 + ], + "type": "text", + "content": "OURS" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 392, + 140, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 140, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 140, + 404 + ], + "type": "text", + "content": "5.6. Ablation Study" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 412, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 615 + ], + "type": "text", + "content": "Effect of scaling factor " + }, + { + "bbox": [ + 46, + 412, + 287, + 615 + ], + "type": "inline_equation", + "content": "k(t)" + }, + { + "bbox": [ + 46, + 412, + 287, + 615 + ], + "type": "text", + "content": " for semantic generation: In this section, we analyze how the scaling factor affects the quality of the sample in the case of complex conditioning of semantic generation. Fig. 10 shows the variation of sample quality starting from the same initial noise with different scaling factors " + }, + { + "bbox": [ + 46, + 412, + 287, + 615 + ], + "type": "inline_equation", + "content": "k(t) = K\\sqrt{1 - \\bar{\\alpha}_t}" + }, + { + "bbox": [ + 46, + 412, + 287, + 615 + ], + "type": "text", + "content": ". The sample quality is bad for very low scaling factors, and for very high scaling factors, the diffusion process escapes the manifold of natural face images. We show the variation in sample quality for a fixed scaling factor versus a time-varying scaling factor in Fig. 12, which demonstrates that a time-varying scale factor produces more realistic samples. This is because the effective variance of the noise scheduling, which effectively controls the amount of regularization possible at a particular timestep, reduces as the generation process proceeds. Hence a larger tweak is permissible at the early steps of diffusion, and only very small tweaks are permitted in the later steps." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 616, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 288, + 688 + ], + "type": "text", + "content": "Noise-Robust classifier: To validate the claims in section 3.1, we train a noise-robust inverse mapper for the task of colorization and show the output of the noise-robust classifier and the diffusion outputs for different noise levels in Fig. 11. The noise-robust classifier fails to preserve key details that our approach preserves." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Limitations Although our method can generalize to a wide series of tasks, one limitation that persists is the value of the" + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 328, + 250, + 398, + 311 + ], + "blocks": [ + { + "bbox": [ + 328, + 250, + 398, + 311 + ], + "lines": [ + { + "bbox": [ + 328, + 250, + 398, + 311 + ], + "spans": [ + { + "bbox": [ + 328, + 250, + 398, + 311 + ], + "type": "image", + "image_path": "11599771e86de2750d03f850a0761e63cc549093484283f4e74ec94f72d0b1d5.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 329, + 311, + 398, + 371 + ], + "blocks": [ + { + "bbox": [ + 329, + 311, + 398, + 371 + ], + "lines": [ + { + "bbox": [ + 329, + 311, + 398, + 371 + ], + "spans": [ + { + "bbox": [ + 329, + 311, + 398, + 371 + ], + "type": "image", + "image_path": "581a0b3ad128f0c504ad2c9a07a89366a257c779b6f900e908f349b097659604.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 329, + 372, + 398, + 432 + ], + "blocks": [ + { + "bbox": [ + 329, + 372, + 398, + 432 + ], + "lines": [ + { + "bbox": [ + 329, + 372, + 398, + 432 + ], + "spans": [ + { + "bbox": [ + 329, + 372, + 398, + 432 + ], + "type": "image", + "image_path": "155d76b370e7d19f4906d3b1b93b541d94e9dda62caad6b49872d0a3e1630c38.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 435, + 545, + 458 + ], + "lines": [ + { + "bbox": [ + 306, + 435, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 545, + 458 + ], + "type": "text", + "content": "Figure 12. A comparison of a time-varying scaling factor with non-time-varying." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 399, + 250, + 459, + 311 + ], + "blocks": [ + { + "bbox": [ + 399, + 250, + 459, + 311 + ], + "lines": [ + { + "bbox": [ + 399, + 250, + 459, + 311 + ], + "spans": [ + { + "bbox": [ + 399, + 250, + 459, + 311 + ], + "type": "image", + "image_path": "83c47f3b33a6f470e1141945df08cffba35fac5f75eee27e4eeea8a40ebafaa3.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 399, + 311, + 459, + 371 + ], + "blocks": [ + { + "bbox": [ + 399, + 311, + 459, + 371 + ], + "lines": [ + { + "bbox": [ + 399, + 311, + 459, + 371 + ], + "spans": [ + { + "bbox": [ + 399, + 311, + 459, + 371 + ], + "type": "image", + "image_path": "39195c3e30d02512f8e9b3642dbdeeb5612ba3b384c6daac900f68cf3ed42dcc.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 399, + 372, + 460, + 432 + ], + "blocks": [ + { + "bbox": [ + 399, + 372, + 460, + 432 + ], + "lines": [ + { + "bbox": [ + 399, + 372, + 460, + 432 + ], + "spans": [ + { + "bbox": [ + 399, + 372, + 460, + 432 + ], + "type": "image", + "image_path": "fdecf438744e32334487383ecf40e743a6844701d723eb75b80a6fba975bd5de.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 460, + 250, + 520, + 311 + ], + "blocks": [ + { + "bbox": [ + 460, + 250, + 520, + 311 + ], + "lines": [ + { + "bbox": [ + 460, + 250, + 520, + 311 + ], + "spans": [ + { + "bbox": [ + 460, + 250, + 520, + 311 + ], + "type": "image", + "image_path": "c4b6624c3e8a2a14fdddc4d53d8340e72512d53bb9d7171c984ec7731b3c085a.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 460, + 311, + 520, + 371 + ], + "blocks": [ + { + "bbox": [ + 460, + 311, + 520, + 371 + ], + "lines": [ + { + "bbox": [ + 460, + 311, + 520, + 371 + ], + "spans": [ + { + "bbox": [ + 460, + 311, + 520, + 371 + ], + "type": "image", + "image_path": "2e3e83c1181c3cf3c8db0f94bd524bd67bf99d6a09a57d8d06b6f457afbb496c.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 460, + 372, + 520, + 432 + ], + "blocks": [ + { + "bbox": [ + 460, + 372, + 520, + 432 + ], + "lines": [ + { + "bbox": [ + 460, + 372, + 520, + 432 + ], + "spans": [ + { + "bbox": [ + 460, + 372, + 520, + 432 + ], + "type": "image", + "image_path": "1b98a9b384664e4bfecbf32be1ceb4043f0a0758040b2a278c39a8c8280dca23.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 461, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 545 + ], + "type": "text", + "content": "scaling factor " + }, + { + "bbox": [ + 304, + 461, + 547, + 545 + ], + "type": "inline_equation", + "content": "k(t)" + }, + { + "bbox": [ + 304, + 461, + 547, + 545 + ], + "type": "text", + "content": ". The value of " + }, + { + "bbox": [ + 304, + 461, + 547, + 545 + ], + "type": "inline_equation", + "content": "k(t)" + }, + { + "bbox": [ + 304, + 461, + 547, + 545 + ], + "type": "text", + "content": " has to be empirically found based on the task, but once a few images are used to tune the value of " + }, + { + "bbox": [ + 304, + 461, + 547, + 545 + ], + "type": "inline_equation", + "content": "k(t)" + }, + { + "bbox": [ + 304, + 461, + 547, + 545 + ], + "type": "text", + "content": ", the model generalizes well to other conditioning images of the same task. Like any other conditional generation models that can perform image editing, our method also has societal impacts, and care must be taken in applying these methods." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 551, + 378, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 378, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 378, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we propose the first framework for plug- and-play conditional generation that can generalize well to both image-to-image translation tasks and label-based generation tasks. For this, we use the energy-based formulation of diffusion models and modulate the inference process using a task-specific predefined network or other preexisting function. Furthermore, we introduce a novel implicit sampling-based technique that improves the sampling quality across multiple tasks. We performed experiments on various tasks to show that our method can generalize across multiple tasks and outperforms existing methods that do not require additional training." + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "text", + "content": "20858" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. arXiv preprint arXiv:2302.07121, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 338 + ], + "type": "text", + "content": "[5] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 288, + 385 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 288, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 440 + ], + "type": "text", + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 288, + 475 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 477, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 288, + 532 + ], + "type": "text", + "content": "[9] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, Kun Zhang, and Dacheng Tao. Geometry-Consistent Generative Adversarial Networks for One-Sided Unsupervised Domain Mapping. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 576 + ], + "type": "text", + "content": "[10] Ruiqi Gao, Yang Song, Ben Poole, Ying Nian Wu, and Diederik P Kingma. Learning energy-based models by diffusion recovery likelihood. arXiv preprint arXiv:2012.08125, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 622 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 657 + ], + "type": "text", + "content": "[12] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. arXiv preprint arXiv:2206.09012, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[13] Will Grathwohl, Kuan-Chieh Wang, Jorn-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. arXiv preprint arXiv:1912.03263, 2019. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "type": "text", + "content": "[15] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts gans. arXiv preprint arXiv:2112.05130, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 547, + 184 + ], + "type": "text", + "content": "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 185, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 546, + 228 + ], + "type": "text", + "content": "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 546, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 546, + 259 + ], + "type": "text", + "content": "[18] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 547, + 304 + ], + "type": "text", + "content": "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), December 2015. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 305, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 546, + 360 + ], + "type": "text", + "content": "[20] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 3, 6, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 361, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 547, + 403 + ], + "type": "text", + "content": "[21] Kangfu Mei, Nithin Gopalakrishnan Nair, and Vishal M Patel. Bi-noising diffusion: Towards conditional diffusion models with generative restoration priors. arXiv preprint arXiv:2212.07352, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 547, + 459 + ], + "type": "text", + "content": "[22] Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. Pulse: Self-supervised photo upsampling via latent space exploration of generative models. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 2437-2445, 2020. 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 460, + 547, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 547, + 493 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 547, + 493 + ], + "type": "text", + "content": "[23] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 494, + 547, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 547, + 548 + ], + "type": "text", + "content": "[24] Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, and Vishal M Patel. Unite and conquer: Plug & play multimodal synthesis using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6070-6079, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 547, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 547, + 603 + ], + "type": "text", + "content": "[25] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4467-4477, 2017. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 604, + 547, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 547, + 657 + ], + "type": "text", + "content": "[26] Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "type": "text", + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "text", + "content": "20859" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[28] Taesung Park, Alexei A Efros, Richard Zhang, and Jun-Yan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 215 + ], + "type": "text", + "content": "[30] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In *British Machine Vision Conference*, 2015. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "type": "text", + "content": "[31] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 337 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 380 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 288, + 415 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 535 + ], + "type": "text", + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 536, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 580 + ], + "type": "text", + "content": "[37] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "text", + "content": "[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 656 + ], + "type": "text", + "content": "[39] Ken Sekimoto. Langevin equation and thermodynamics. Progress of Theoretical Physics Supplement, 130:17-27, 1998. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 363 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 105 + ], + "type": "text", + "content": "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "text", + "content": "[42] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 206 + ], + "type": "text", + "content": "[43] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 207, + 546, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 546, + 261 + ], + "type": "text", + "content": "[44] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 263, + 547, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 547, + 317 + ], + "type": "text", + "content": "[45] Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual-linguistic manner. arXiv preprint arXiv:2112.03109, 2021. 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 319, + 546, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 546, + 363 + ], + "type": "text", + "content": "[46] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In IEEE International Conference on Computer Vision (ICCV), 2017. 3, 7, 8" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "text", + "content": "20860" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/3c8d0cf6-dc7d-44ba-a804-37f576c9cfcc_content_list.json b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/3c8d0cf6-dc7d-44ba-a804-37f576c9cfcc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..568a87b55134ba8650641d431697160cbc0aed89 --- /dev/null +++ b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/3c8d0cf6-dc7d-44ba-a804-37f576c9cfcc_content_list.json @@ -0,0 +1,2078 @@ +[ + { + "type": "text", + "text": "StegaNeRF: Embedding Invisible Information within Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenxin Li $^{1*}$ , Brandon Y. Feng $^{2*}$ , Zhiwen Fan $^{3*}$ , Panwang Pan $^{4}$ , Zhangyang Wang $^{3}$ $^{1}$ Chinese University of Hong Kong, $^{2}$ University of Maryland, $^{3}$ University of Texas at Austin, $^{4}$ ByteDance", + "bbox": [ + 156, + 179, + 810, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in neural rendering have paved the way for a future marked by the widespread distribution of visual data through the sharing of Neural Radiance Field (NeRF) model weights. However, while established techniques exist for embedding ownership or copyright information within conventional visual data such as images and videos, the challenges posed by the emerging NeRF format have remained unaddressed. In this paper, we introduce StegaNeRF, an innovative approach for steganographic information embedding within NeRF renderings. We have meticulously developed an optimization framework that enables precise retrieval of hidden information from images generated by NeRF, while ensuring the original visual quality of the rendered images to remain intact. Through rigorous experimentation, we assess the efficacy of our methodology across various potential deployment scenarios. Furthermore, we delve into the insights gleaned from our analysis. StegaNeRF represents an initial foray into the intriguing realm of infusing NeRF renderings with customizable, imperceptible, and recoverable information, all while minimizing any discernible impact on the rendered images. For more details, please visit our project page: https://xggnet.github.io/StegaNeRF/", + "bbox": [ + 73, + 301, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 680, + 209, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implicit neural representation (INR) is an emerging concept where the network describes the data through its weights [14, 24, 40, 42, 45, 58, 65]. After training, the INR weights can then be used for content distribution, streaming, and even downstream inference tasks, all without sending or storing the original data. Arguably the most prominent INR is Neural Radiance Fields (NeRF) [42], where a network learns a continuous function mapping spatial coordinates to density and color. Due to its lightweight size and superb quality, NeRF has immense potential for 3D content representation in future vision and graphics applications.", + "bbox": [ + 75, + 707, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc98755da4cfee183d3e717dde8e24380c2629c36ff498bf79fe2a72540a46a9.jpg", + "image_caption": [ + "Figure 1. We introduce the new problem of NeRF steganography: hiding information in NeRF renderings. Our proposed framework, StegaNeRF, can embed and recover customized hidden information while preserving the original NeRF rendering quality." + ], + "image_footnote": [], + "bbox": [ + 501, + 270, + 890, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While there is a plethora of work dedicated towards better quality [4,63,72,91], faster rendering [43,56,61,71,83], and sparse view reconstruction [10, 11, 44, 57, 80, 81, 85, 88], in this paper, we look beyond the horizon and explore a new question: Can we achieve steganography with NeRF?", + "bbox": [ + 496, + 455, + 892, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Established digital steganography method [12] focus on embedding hidden messages in 2D images. The recent growth of deep learning and social media platforms further gives rise to many practical use cases of image steganography. As countless images and videos are shared online and even used to train deep learning models, 2D steganography methods [2,3] allow users and data providers to protect copyright, embed ownership, and prevent content abuse.", + "bbox": [ + 496, + 534, + 893, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Now, with the ongoing advances in 3D representations powered by NeRF, we envision a future where people frequently share their captured 3D content online just as they are currently sharing 2D images and videos online. Moreover, we are curious to explore the following research questions: 1) Injecting information into 2D images for copyright or ownership identification is common, but can we preserve such information when people share and render 3D scenes through NeRF? 2) NeRF can represent large-scale real-world scenes with training images taken by different people, but can we preserve these multiple source identities in the NeRF renderings to reflect the collaborative efforts required to reconstruct these 3D scenes? 3) Common image steganography methods embed either a hidden image or a message string into a given image, but can we allow different modalities of the hidden signal in NeRF steganography?", + "bbox": [ + 496, + 659, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 886, + 204, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "441", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Driven by these questions, we formulate a framework to embed customizable, imperceptible, and recoverable information in NeRF renderings without sacrificing the visual quality. Fig. 1 presents an overview of our proposed framework, dubbed StegaNeRF. Unlike the traditional image steganography that embeds hidden signals only into a specific source image, we wish to recover the same intended hidden signal from NeRF rendered at arbitrary viewpoints.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite many established works on 2D steganography and hidden watermarks for image and video [2, 3, 9, 25, 36, 64], naively applying 2D steganography on the NeRF training images is not practical, since the embedded information easily gets lost in the actual NeRF renderings. In contrast, our framework enables reliable extraction of the hidden signal from NeRF renderings. During NeRF training, we jointly optimize a decoder network to extract hidden information from the NeRF renderings. To minimize the negative impact on the visual quality of NeRF, we identify weights with low impact on rendering and introduce a gradient masking strategy to steer the hidden steganographic information towards those low-importance weights. Extensive experimental results validate StegaNeRF balances between the rendering quality of novel views and the high-fidelity transmission of the concealed information.", + "bbox": [ + 75, + 212, + 470, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "StegaNeRF presents the first exploration of hiding information in NeRF models for ownership identification, and our contributions can be summarized as follows:", + "bbox": [ + 76, + 454, + 468, + 498 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the new problem of NeRF steganography and present the first effort to embed customizable, imperceptible, and recoverable information in NeRF.", + "- We propose an adaptive gradient masking strategy to steer the injected hidden information towards the less important NeRF weights, balancing the objectives of steganography and rendering quality.", + "- We empirically validate the proposed framework on a diverse set of 3D scenes with different camera layouts and scene characteristics, obtaining high recovery accuracy without sacrificing rendering quality.", + "- We explore various scenarios applying StegaNeRF for ownership identification, with the additional support of multiple identities and multi-modal signals." + ], + "bbox": [ + 96, + 510, + 468, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 768, + 217, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields Following the success of NeRF [42], numerous techniques have been incorporated including ray re-parameterizations [4, 5], explicit spatial data structures [16, 23, 34, 43, 61, 84], caching and distillation [19, 22, 54, 69], ray-based representations [1, 15, 59], geometric primitives [33, 35], large-scale scenes [39, 63, 79, 89], generalizable architectures [60, 68, 73], and dynamic", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "settings [18, 32, 46, 51, 78]. Unlike prior works, this paper explores the uncharted problem of embedding information in NeRF renderings, with critical implications for copyright protection and ownership preservation. As early-stage NeRF-based products already become available [37,49] and more are expected, we believe now is the right time to open up the exploration of NeRF steganography.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Steganography Steganography hides intended signals as invisible watermarks (e.g., hyperlinks, images) within the cover media called carriers (e.g., images, video and audio) [12, 26]. Classical methods focus on seeking and altering the least significant bits (LSBs) [17, 47, 62, 76] and transforming domain techniques [7, 8, 52, 70]. Prior research also uses deep neural networks for steganography [2, 3, 21, 66]. Among them, DeepStega [2] conceals a hidden image within an equal-sized carrier image. Subsequent works [25, 36] use invertible networks to improve the performance of deep image steganography. Another line of work conceals information in other carrier media like audio [13, 20, 82] and video [38, 55, 75]. The above advances all play a critical part in the era when traditional media formats like images and videos are dominant. However, as MLP-based neural representations of 3D scenes are gaining momentum to become a major format of visual data, extending steganography to NeRF is bound to become an important problem in the upcoming future.", + "bbox": [ + 496, + 204, + 892, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lifting Steganography to 3D Prior to NeRF, meshes are commonly used to represent 3D shapes. Early pioneers apply steganography 3D meshes [6, 50] for copyright protection when meshes are exchanged and edited. More recent work [77] has also explored embedding multi-plane images within a JPEG-compressed image. Their problem can be regarded as a special case of 2D steganography, hiding multiple 2D images inside a single 2D image. In contrast, we try to hide a natural image into a 3D scene representation (NeRF), fundamentally differing from these prior methods where 2D images act as the carrier of hidden information.", + "bbox": [ + 496, + 500, + 890, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 679, + 589, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of StegaNeRF is to inject customized (steganographic) information into the NeRF weights with imperceptible visual changes when rendering. Given a NeRF with weights $\\theta_0$ and the information $I$ to be injected, when we render with $\\theta$ on any viewpoint, we hope the injected $I$ can be recovered by a decoder $F_{\\psi}$ with learnable weights $\\psi$ .", + "bbox": [ + 496, + 703, + 890, + 795 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A seemingly obvious solution is to use prior image steganography methods to 1) inject $I$ on training images, 2) train NeRF on those images with embedded information, 3) apply their provided $F_{\\psi}$ to extract $I$ from NeRF-rendered images. This approach has been successfully applied for GAN fingerprinting [86]. However, it fails in the NeRF context (see Fig. 3), where the subtle changes induced by prior", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "442", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b29770537380d4b638d8474caadb3fcc7f31d0cbff08beb0c0e4ccfa34f9979e.jpg", + "image_caption": [ + "Figure 2. StegaNeRF training overview. At the first stage, we optimize $\\theta_0$ with standard NeRF training. At the second stage, we initialize $\\theta$ with $\\theta_0$ and optimize for the steganography objectives. We train the decoder $F_{\\psi}$ to recover hidden information from StegaNeRF renderings and no hidden information from original NeRF renderings. We introduce Classifier Guided Recovery to improve the accuracy of recovered information, and Adaptive Gradient Masking to balance between steganography ability and rendering visual quality." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 617, + 272 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ad4fe3df72f82e271759a77cc7e3347be07d4bfc2df6cafe49d17756ee69e18c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 89, + 890, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "steganography methods easily get smoothed out, inhibiting the decoder from identifying the subtle patterns necessary for information recovery.", + "bbox": [ + 75, + 367, + 468, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Such shortcomings of off-the-shelf steganographic methods are not surprising since they are developed for the traditional setting where 2D images are the ultimate form of visual information. In contrast, our problem setting involves the new concept of INR as the underlying representation, and 2D images are just the final output rendered by NeRF.", + "bbox": [ + 75, + 412, + 468, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Two-Stage Optimization", + "text_level": 1, + "bbox": [ + 76, + 512, + 302, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recognizing the formulation difference due to the emergence of implicit representation with NeRF, we move away from the traditional 2D image-based pipeline that trains an encoder network to inject subtle changes to the given 2D images. Instead, we incorporate the steganography objective into the gradient-based learning process of NeRF.", + "bbox": [ + 75, + 536, + 468, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We re-design the training as a two-stage optimization. The first stage is the original NeRF training procedure, involving the standard photometric loss between the rendered and ground truth pixels to guarantee the visual quality. After finishing the weights update $\\theta_0$ at the first stage, we dedicate the second stage to obtain the final NeRF weights $\\theta$ containing steganographic information. We introduce several techniques to achieve robust information recovery with imperceptible impact on the rendered images. The training workflow is depicted in Fig. 2 and Alg. 1.", + "bbox": [ + 75, + 627, + 468, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Information Recovery", + "text_level": 1, + "bbox": [ + 76, + 786, + 282, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $P$ denote the camera pose at which we render an image from a NeRF network. We want to recover $I$ from the rendered image $\\theta(P)$ . Importantly, we also want to avoid false positives on images $\\theta_0(P)$ rendered by the original network without steganography ability, even if the images rendered by $\\theta_0$ and $\\theta$ look visually identical. Therefore, we", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f3e1e4ae25f1e21df1c01ce78e682cc666d0c0f30d5607ea1ba813a143312571.jpg", + "image_caption": [ + "Figure 3. Results of applying prior 2D steganography methods, LSB [9] (top) and DeepStega [2] (bottom). From left to right, we show (a) training image after applying 2D steganography, (b) residual error of (a) over ground truth, (c) hidden image recovered from (a), (d) residual error of NeRF rendering at the pose of (a), and (e) hidden image recovered from (d). Prior 2D steganography methods fail in the NeRF context since the hidden information injected in training images mostly disappear in NeRF renderings." + ], + "image_footnote": [], + "bbox": [ + 501, + 367, + 594, + 487 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cbe4fda3609a084f42a8a14db3790f72757d45490303723720d58d477593ac29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 368, + 687, + 491 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9f560c7b3ab4013bb5972d53e73e75ad83c244fd7deacb4a8179578a4eee9aa3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 368, + 733, + 491 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/714106e3959ffa6ca22a7f1d49393f175a822b87c47e862453cedeb389222dbe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 368, + 836, + 491 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7788ee5077cd1a502af07201b4d246dc962101c0ce7ce1cb52d330e9b22434c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 844, + 368, + 883, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "optimize $\\theta$ to minimize the following contrastive loss terms:", + "bbox": [ + 500, + 642, + 888, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e c} ^ {+} = \\left| F _ {\\psi} (\\boldsymbol {\\theta} (P)) - I \\right|, \\quad \\mathcal {L} _ {d e c} ^ {-} = \\left| F _ {\\psi} \\left(\\boldsymbol {\\theta} _ {0} (P)\\right) - \\varnothing \\right|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 685, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\varnothing$ is an empty image with all zeros. Effectively, $L_{dec}^{+}$ serves as a regularization term, facilitating the decoder to recover the embedded image pattern based on renderings obtained from a StegaNeRF. In contrast, $L_{dec}^{-}$ prevents the decoder from falsely producing any plausible image pattern when given renderings from a Standard NeRF that is without hidden signals.", + "bbox": [ + 496, + 713, + 890, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-Guided Recovery The decoder $F_{\\psi}$ is easily implemented as a U-Net to decode $I$ as the form of 2D images, but accurately recovering all the details in $I$ might be challenging. Therefore, we additionally train a classifier", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "443", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Train StegaNeRF on a single scene", + "text_level": 1, + "bbox": [ + 76, + 89, + 398, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data: Training images $\\{Y_i\\}$ with poses $\\{P_i\\}$ , hidden information $I$ , learning rate $\\eta = [\\eta_0, \\eta_1]$", + "bbox": [ + 76, + 108, + 467, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Steganographic NeRF $\\theta$ and decoder $F_{\\psi}$", + "bbox": [ + 76, + 138, + 406, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initialize NeRF $\\theta_0$ and decoder network $F_{\\psi}$", + "bbox": [ + 76, + 154, + 367, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimize $\\theta_0$ on $\\{Y_i, P_{ij}\\}$ with standard NeRF training", + "bbox": [ + 76, + 169, + 431, + 184 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute mask $m$ for $\\theta_0$ as in Eq. (3)", + "bbox": [ + 76, + 185, + 328, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for each training iteration $t$ do", + "bbox": [ + 76, + 200, + 281, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Randomly sample a training pose $P_{i}$ and render $\\theta (P_i)$", + "bbox": [ + 91, + 215, + 450, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute stega. losses $\\mathcal{L}_{dec}^{c}$ , $\\mathcal{L}_{dec}^{+}$ , $\\mathcal{L}_{dec}^{-}$ as Eq. (1), (2)", + "bbox": [ + 91, + 232, + 450, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute $\\mathcal{L}_{rgb}$ and combine total loss $\\mathcal{L}$ as in Eq. (4)", + "bbox": [ + 91, + 247, + 441, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Update $\\pmb{\\theta}$ with $\\eta_0\\cdot (\\frac{\\partial\\mathcal{L}}{\\partial\\pmb{\\theta}}\\odot \\pmb {m})$ and $F_{\\psi}$ with $\\eta_{1}\\cdot \\frac{\\partial\\mathcal{L}}{\\partial\\psi}$", + "bbox": [ + 91, + 261, + 429, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "text_level": 1, + "bbox": [ + 76, + 280, + 132, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "network $F_{\\psi}^{c}$ to solve the easier task of classifying whether the given NeRF rendering contains hidden information. $F_{\\psi}^{c}$ is optimized by the following cross-entropy loss [30, 67]:", + "bbox": [ + 75, + 319, + 468, + 366 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e c} ^ {c} = - \\log \\left(F _ {\\psi} ^ {c} (\\boldsymbol {\\theta} (P))\\right) - \\log \\left(1 - F _ {\\psi} ^ {c} \\left(\\boldsymbol {\\theta} _ {0} (P)\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 375, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then use its prediction to guide the process of decoding pixel-wise information by adding the classification output as input to $F_{\\psi}$ , such that $F_{\\psi}(x) = F_{\\psi}\\big(x,F_{\\psi}^{c}(x)\\big)$ .", + "bbox": [ + 75, + 402, + 468, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the above discussion focuses on hiding images, our framework can be easily extended to embed other modalities like strings, text, or even audio, all of which may be represented as 1D vectors. We can simply modify the architecture of $F_{\\psi}$ to have a 1D prediction branch.", + "bbox": [ + 75, + 449, + 468, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Preserving Perceptual Identity", + "text_level": 1, + "bbox": [ + 76, + 534, + 348, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we want to hide information without affecting the visual perception of the rendered output, an intuitive regularization is to penalize how much $\\theta$ deviates from $\\theta_0$ . However, we find that naively summing the deviation penalty across all weights makes it difficult for the NeRF network to adjust its weights for the steganographic objective. Instead, motivated by the fact that INR weights are not equally important and exhibit strong sparsity [28, 87], we propose an adaptive gradient masking strategy to encode the hidden information on specific weights' groups.", + "bbox": [ + 75, + 556, + 468, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given the initial set of weights $\\theta_0\\in \\mathbb{R}^N$ , we compute a soft importance mask $m\\in \\mathbb{R}^{N}$ . Specifically, for each weight $\\pmb {w}_i\\in \\pmb {\\theta}_0$ , we compute:", + "bbox": [ + 75, + 707, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {m} _ {i} = \\frac {\\left| \\boldsymbol {w} _ {i} \\right| ^ {- \\alpha}}{\\sum_ {i} ^ {N} \\left| \\boldsymbol {w} _ {i} \\right| ^ {- \\alpha}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 763, + 468, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the magnitude $|\\cdot|$ of each model weight $w_{i}$ serves as its individual importance indicator, and $\\alpha > 0$ is a scaling factor of power controlling the relative distribution of importance across the weights. We mask the gradient as $\\frac{\\partial\\mathcal{L}}{\\partial\\theta} \\odot m$ when optimizing $\\theta$ based on the total loss $\\mathcal{L}$ in the second stage, where $\\odot$ is a Hadamard product. Effectively,", + "bbox": [ + 76, + 809, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 2 Typical usage scenario of StegaNeRF", + "text_level": 1, + "bbox": [ + 500, + 90, + 838, + 106 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Alice captures some images of a 3D scene", + "2: Alice trains a StegaNeRF to hide a personalized image", + "3: Alice shares the model $\\theta$ online for other people to enjoy and explore the 3D scene themselves", + "4: Bob grabs the model $\\theta$ and reposts it with his own account without crediting Alice or asking for permission", + "5: Alice sees Bob's post, deploys the decoder $F_{\\psi}$ , and verifies the owner of $\\theta$ is Alice, not Bob", + "6: Bob takes down the post or gets banned for copyright infringement" + ], + "bbox": [ + 509, + 109, + 888, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "more significant weights are \"masked out\" to minimize the impact of steganographic learning on the rendered visual quality.", + "bbox": [ + 496, + 290, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We retain the standard photometric error in steganography learning to maintain NeRF rendering fidelity: $\\mathcal{L}_{rgb} = |\\pmb {\\theta}(P) - \\pmb {\\theta}_0(P)|$ . The overall training loss at the second stage can be formulated as follows:", + "bbox": [ + 496, + 335, + 890, + 395 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {0} \\mathcal {L} _ {d e c} ^ {c} + \\lambda_ {1} \\mathcal {L} _ {d e c} ^ {+} + \\lambda_ {2} \\mathcal {L} _ {d e c} ^ {-} + \\lambda_ {3} \\mathcal {L} _ {r g b}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 405, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 444, + 632, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Implementation Details.", + "text_level": 1, + "bbox": [ + 500, + 469, + 720, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset. We use common datasets LLFF [41] and NeRF-Synthetic [42], with forward scenes {flower, fern, fortress, room} from LLFF and $360^{\\circ}$ scenes {lego, drums, chair} from NeRF-Synthetic. We further experiment on the Brandenburg Gate scene from NeRF-W dataset [39], with over 800 views of in-the-wild collected online.", + "bbox": [ + 496, + 492, + 890, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. On LLFF and NeRF-Synthetic, we adopt Plenoxels [83] as the NeRF backbone architecture for efficiency. On NeRF-W, we use the available PyTorch implementation [53]. The first stage of training is performed according to the standard recipes of those implementations. We then perform the second stage of steganography training for 55 epochs unless otherwise noted. On LLFF and NeRF-W, we downsize the training images by 4 times following common practice, and we use the original size on NeRF-Synthetic.", + "bbox": [ + 496, + 583, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For hyper-parameters in Eq. (4), we set the weight of NeRF reconstruction error $\\lambda_3 = 1$ for all experiments. We set $\\lambda_0 = 0.01, \\lambda_1 = 0.5, \\lambda_2 = 0.5$ for all scenes on LLFF dataset, and $\\lambda_0 = 0.1, \\lambda_1 = 1, \\lambda_2 = 1$ for the scenes in NeRF-Synthetic, and $\\lambda_0 = 0.05, \\lambda_1 = 1, \\lambda_2 = 1$ for NeRF-W. For Eq. (3), we set $\\alpha = 3$ for all scenes. We run experiments on one NVIDIA A100 GPU.", + "bbox": [ + 496, + 734, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation. We assume a typical authentication scenario shown in Alg.2. We consider the recovery quality of the embedded information, including the metrics of classification accuracy (Acc.) of the classifier, and the structural sim", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "444", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39a3a01589af864221c2ab7e726b6772ec2530d5c9bd57a39b03ddab82e30774.jpg", + "image_caption": [ + "Figure 4. Results on the NeRF-Synthetic dataset. Within each column, we show the StegaNeRF rendering, residual error from the initial NeRF rendering, and the recovered hidden image. We show the SSIM for the StegaNeRF renderings and the recovered hidden images." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 329 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3159b1a50b982899fb8f35af09c185828f3d695bf7258d91be553d0322758564.jpg", + "table_caption": [ + "Table 1. Quantitative results of StegaNeRF rendering and hidden information recovery. Standard NeRF is the initial NeRF $\\theta_0$ with standard training (upper-bound rendering performance). Prior 2D steganography fails after NeRF training while StegaNeRF successfully embeds and recovers hidden information with minimal impact on the rendering quality. Results are averaged over the selected LLFF and NeRF-Synthetic scenes." + ], + "table_footnote": [], + "table_body": "
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
Standard NeRF28.150.85750.137750.00N/A
LSB [9]28.130.85690.1452N/A0.0132
DeepStega [2]26.940.84320.1641N/A0.2098
StegaNeRF (Ours)28.130.85630.1460100.00.9730
Standard NeRF31.130.96060.031050.00N/A
LSB [9]31.120.96040.0310N/A0.0830
DeepStega [2]31.130.96060.0313N/A0.2440
StegaNeRF (Ours)30.960.95830.029099.720.9677
", + "bbox": [ + 80, + 502, + 468, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ality (SSIM) [74] of the hidden image recovered by the decoder. We evaluate the final NeRF renderings with PSNR, SSIM and LPIPS [90]. All the metrics are computed on the test set and averaged over all the scenes and embedded images.", + "bbox": [ + 75, + 676, + 470, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Case I: Embedding in a Single Scene.", + "text_level": 1, + "bbox": [ + 76, + 762, + 398, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first explore the preliminary case of ownership identification on a specific NeRF scene. We select random images from ImageNet [27] as the hidden information to be injected to NeRF renderings.", + "bbox": [ + 75, + 786, + 470, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Failure of 2D Baseline Due to the lack of prior study on NeRF steganography, we consider a baseline from 2D image steganography by training NeRF from scratch with", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8b68c8e6da47b18c73cc15677a49f22252674ba923b88fbb7bd44da5922218b5.jpg", + "table_caption": [ + "Table 2. Quantitative results of NeRF rendering and hidden information recovery. We consider two conditions, embedding each scene with a common hidden image (One-for-All) or a scene-specific hidden image (One-for-All). We report the quality difference compared to the single-scene settings as $\\Delta_{SSIM}^{(10^{-2})}$ . Results are averaged over the selected LLFF scenes." + ], + "table_footnote": [], + "table_body": "
SettingNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS ↓Δ(10-2)SSIM↑Acc. (%)↑SSIM↑Δ(10-2)SSIM↑
One-for-All24.990.80130.1786-0.10100.000.9860+0.19
One-for-One24.990.80160.1779-0.07100.000.9122-7.19
One-for-All27.900.85130.1236+0.01100.000.9844-0.76
One-for-One27.900.85150.1195+0.03100.000.9448-4.45
One-for-All30.270.84980.1289+0.02100.000.9430-5.42
One-for-One30.120.84800.1302-0.16100.000.9102-8.67
", + "bbox": [ + 503, + 491, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the watermarked images. We implement two off-the-shelf steganography methods including a traditional machine learning approach called Least Significant Bit (LSB [9]), and a deep learning pipeline as DeepStega [2]. An ideal case is that the embedded information can be recovered from the synthesized novel views, indicating the successful NeRF steganography. However, as can be seen in Fig. 3, the embedded information containing the hidden images cannot be recovered from the NeRF renderings. By analyzing the residual maps of training views (between GT training views and the watermarked) and novel views (between GT novel views and the actual rendering), we observe the subtle residuals to recover the hidden information are smoothed out in NeRF renderings, and similar failures occur on other datasets as well. Therefore, for the rest of our study, we mainly focus on analyzing our new framework that performs direct steganography on 3D NeRF.", + "bbox": [ + 496, + 657, + 893, + 915 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "445", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d6c84e49f3bc15605767706ce87f049e7e32deb570f4eed16671ed82011e52b3.jpg", + "image_caption": [ + "(a) Per-Scene" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 176, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/befd6ae597219dc81b1e27ac75f189b933195b1162d5bf30874010a2636db60c.jpg", + "image_caption": [ + "Figure 5. Results on three multi-scene settings. Within each column, we show the StegaNeRF rendering, residual error from the initial NeRF rendering, and recovered hidden image. We show the SSIM for the StegaNeRF renderings and the recovered hidden images." + ], + "image_footnote": [], + "bbox": [ + 176, + 90, + 269, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/724703a416a23b70eef375d130184438f3bd0ba5aa8f779ac91b975f1e840cf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 90, + 330, + 309 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/69a753bd6b77b0add7d695a2f3857e3ff481eb79efde28c14fad54228efcafd4.jpg", + "image_caption": [ + "(b) Multi-Scene: One-for-All" + ], + "image_footnote": [], + "bbox": [ + 351, + 90, + 446, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08c10b53ec3c44fc6e5f96b9244d77e6432ce2929d6faa68b72d4b46ba75ba34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 90, + 542, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/175cfe58e9400866ce5654d7f117139b85151c87270308225607d98fb9a223d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 90, + 607, + 309 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f331cae1e498cc8136eaf4476c507bae58d36a3c26bac8c217349b7d91e064ce.jpg", + "image_caption": [ + "(c) Multi-Scene: One-for-One" + ], + "image_footnote": [], + "bbox": [ + 620, + 90, + 717, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3f6877407ef268b8eca057516f723311e7953bb0ff9658bae81d22c0d4dc5ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 90, + 813, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2aa717d3e47396a8532de09b61ec7888b48cbd0022effaa387079d033e90d281.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 90, + 875, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results Tab. 1 contains quantitative results on LLFF and NeRF-Synthetic scenes. While NeRF trained by 2D steganography methods hardly recovers the embedded information, StegaNeRF accurately recovers the hidden image with minimal impact on the rendering quality measured by PSNR. Fig. 4 provides qualitative results of StegaNeRF on three embedded images on the NeRF-Synthetic scenes. An interesting observation is the common regions where hidden information (high residual error) emerges in the renderings, e.g., the right rear side of lego and left handrail of chair. We also notice that, within each scene, these regions are persistent across multiple viewpoints.", + "bbox": [ + 75, + 390, + 472, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Case II: Embedding in Multiple Scenes.", + "text_level": 1, + "bbox": [ + 76, + 590, + 416, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings We extend our steganography scheme to embed information within multiple scenes at once. Specially, we use three LLFF scenes $\\{flower, fern, fortress\\}$ to test the two sub-settings of multi-scene embedding with: (1) One-for-All, a common hidden image and (2) One-for-One scene-specific hidden images. All scenes share the same decoder $F_{\\psi}$ and classifier $F_{\\psi}^{c}$ . The difference between the two sub-settings is the number of hidden images that $F_{\\psi}$ and $F_{\\psi}^{c}$ need to identify and recover. We sample one scene for every training epoch, and due to the increased data amount, we increase training epochs until convergence.", + "bbox": [ + 75, + 617, + 468, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results Tab. 2 provides quantitative results on the multi-scene setting. The performance drop compared to single-scene training is sometimes noticeable, but it is not surprising due to the inherent requirement of per-scene fitting for our NeRF framework. Fig. 5 presents visual comparisons of multi-scene steganography against single-scene scheme.", + "bbox": [ + 75, + 810, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Case III: Embedding Multiple Identities.", + "text_level": 1, + "bbox": [ + 500, + 388, + 846, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings Constructing NeRF of large-scale cultural landmarks is a promising application, and community contributions are crucial to form the training images. Since every NeRF prediction is indebted to some particular training images, it would be meaningful to somehow identify the contributing users in the rendered output. Specifically, we present a proof-of-concept based on the following scenario: Given a collection of user-uploaded training images, our task is to apply StegaNeRF to ensure the final NeRF renderings hide subtle information about the relevant identities whose uploaded images help generate the current rendering.", + "bbox": [ + 496, + 417, + 890, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To simulate this multi-user scenario in the public NeRF-W dataset, we randomly select $M$ anchor views with different viewing positions, and then find the $K$ nearest neighbour views to each anchor to form their respective clusters. We set $M = 3$ , $K = 20$ in experiments. We assume a common contributor identity for each cluster, and we want the NeRF rendering to contain customized hidden information about those $M$ contributors when we render within the spatial regions spanned by their own cluster. Thus, our classifier network $F_{\\phi}^{c}$ is modified to output $M$ -class cluster predictions and another class for identity outside of the $M$ clusters. Since the decoder should extract no information for views outside of those $M$ clusters to prevent false positives, we also compute $\\mathcal{L}_{dec}^{-}$ (1) and $\\mathcal{L}_{dec}^{c}$ (2) for those poses.", + "bbox": [ + 496, + 589, + 892, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results We employ the same network backbone as NeRF-W [66] to handle in-the-wild images with different time and lighting effects. Fig. 6 presents qualitative results of embedding multiple identities in a collaborative large-scale NeRF.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "446", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4483d0a9b5b2ae4bbaa21689583e9c4057846bf953d6cb41c605d43855bb3761.jpg", + "image_caption": [ + "Figure 6. Qualitative results of NeRF steganography on NeRF-W dataset. Within each block we show the StegaNeRF rendering, the initial NeRF rendering, the residual error between StegaNeRF and initial, and the recovered hidden image. At the bottom right, we provide the SSIM computed with their respective ground truth. \"Unmarked Identity\" denotes the remaining training views not related to any of the considered users, and the decoder does not recover meaningful information from renderings at those poses." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 885, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6115d281af9e8a99be15960170776e7ca09f2a999dcbcd896e05df93fa141114.jpg", + "image_caption": [ + "SSIM=0.9709" + ], + "image_footnote": [], + "bbox": [ + 81, + 380, + 230, + 467 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d5f6e9d56b3afd50151fcc36ad4dcec880d1a82313045d06d4edd2f8e2f9035.jpg", + "image_caption": [ + "$\\mathrm{SSIM} = 0.9688$", + "`The Owner Of These Horns'", + "Figure 7. Multi-modal hidden information. We show StegaNeRF rendering, recovered hidden images, audio (waveform and spectrum), and text. Metrics for hidden information in each modality are labeled respectively. For the StegaNeRF rendering, we report both SSIM and the relative change from initial NeRF renderings." + ], + "image_footnote": [], + "bbox": [ + 81, + 472, + 230, + 559 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eedc427949d4424f0b28fc773a892c6d11e3e4f4b53c383f4d44c06db78ba2ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 238, + 380, + 305, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2541dbb6af9e408ec055d3b7aa36a413e49dd16066289336cdf21de842b81e68.jpg", + "image_caption": [ + "PSNR=76.22" + ], + "image_footnote": [], + "bbox": [ + 308, + 380, + 387, + 426 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02d720779d3fa8852ba90d4ba95f2f643b5a09b5111a41a74c02efec62665ad8.jpg", + "image_caption": [ + "$= 76.22$" + ], + "image_footnote": [], + "bbox": [ + 392, + 382, + 460, + 426 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a241df4d59a9e60a03f5aee43ecadc2716e334fb69ad0b10dfb1489752cbc058.jpg", + "image_caption": [ + "Acc. $= 100.0\\%$" + ], + "image_footnote": [], + "bbox": [ + 251, + 449, + 377, + 459 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2293670317ccd800ee78f7c0eac6594955cd76bd38ba75102b9ae3ced8a275c.jpg", + "image_caption": [ + "Hidden Image" + ], + "image_footnote": [], + "bbox": [ + 245, + 479, + 303, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08db40a38819a1606a7343bc42a8e6965c6f8254d53a8e17f139afba53b17090.jpg", + "image_caption": [ + "Hidden Audio", + "PSNR=62.84", + "Hidden Text" + ], + "image_footnote": [], + "bbox": [ + 313, + 479, + 387, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/448a38526e8ec86cf3093219f717c7da6ff576251a316449e63b928245c351b3.jpg", + "image_caption": [ + "Acc. $= 100.0\\%$" + ], + "image_footnote": [], + "bbox": [ + 393, + 479, + 465, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "62.84", + "bbox": [ + 395, + 523, + 419, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Case IV: Embedding Multi-Modal Information.", + "text_level": 1, + "bbox": [ + 76, + 667, + 468, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings We further show the potential of StegaNeRF in embedding multi-modal hidden information, such as images, audio, and text. We modify the decoder network to build a modal-specific decoder for each modality.", + "bbox": [ + 75, + 690, + 468, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results Fig. 7 shows recovered multi-modal embedded signals in the trex and horns scenes from LLFF. Evidently, the StegaNeRF framework can easily extend to embed multi-modal information with high recovery performance without scarifying the rendering quality.", + "bbox": [ + 75, + 760, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 847, + 241, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effect of removing each component of StegaNeRF is presented in Tab. 3. Standard NeRF uses the initial NeRF", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e30ed4eedca9a7442494c7b930a04e9fe048cbd3b6d9ac5f6a756206fc45b446.jpg", + "table_caption": [ + "Table 3. Ablation study of different components of StegaNeRF. Results are averaged on the selected LLFF scenes." + ], + "table_footnote": [], + "table_body": "
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
StegaNeRF28.130.85680.1430100.00.9486
No Classifier26.850.80770.2417N/A0.5629
No Classifier Guided27.120.82390.2073100.00.6716
No Gradient Masking27.860.83750.1710100.00.8822
No Soft Masking28.050.85580.152694.440.8751
Standard NeRF28.150.85750.137750.00N/A
", + "bbox": [ + 503, + 417, + 890, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rendering. No Classifier removes the classifier $F_{\\psi}^{c}$ , while No Classifier Guided retains the classification task (hence impact on NeRF rendering) but does not condition the decoder on classifier output. No Gradient Masking removes the proposed masking strategy (Sec. 3.3), and No Soft Masking uses the binary mask [29, 31] with a threshold of 0.5. Our StegaNeRF makes a good balance between the rendering quality and the decoding accuracy. Fig. 8 shows the visual impact of removing each component. It appears that when any component is removed, the performance drops accordingly, revealing the effectiveness of our design.", + "bbox": [ + 496, + 573, + 890, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 9 shows the impact of varying gradient masking strategies. We can see a different degree of affected by ratio variation on hidden recovery and rendering quality. Fig. 10 reports the performance of StegaNeRF against the common perturbations including JPEG compression and Gaussian noise. The lines show the average accuracies across selected scenes and the shaded regions indicate the range of 0.5 standard deviation. It appears that the hidden recovery of StegaNeRF is robust to various JPEG compression ratios and Gaussian blur degradation.", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "447", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77d0c1748fe17cc0f6e4a4972819f344c08065aed7646d5e7cb7b23d4f1b7448.jpg", + "image_caption": [ + "(a) Ground Truth", + "(b) No Classifier", + "(c) No Classifier Guided", + "(d) No Gradient Masking", + "(e) No Soft Masking", + "(f) StegaNeRF" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 885, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/583281c660dc64ea12b2dbccb5f16738971b1b9143cdac21cac82a8d98033ca4.jpg", + "image_caption": [ + "(a) Varying $\\alpha$", + "Figure 9. Analysis of gradient masking. (a) Varying $\\alpha$ from Eq. (3). (b) Varying portions of masked out weights from the gradient updates at steganography learning. We provide the SSIM of rendered views (blue) and recovered hidden images (green)." + ], + "image_footnote": [], + "bbox": [ + 84, + 306, + 261, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d787ff1f65773bcd571152f75a43b24c34cd1bec45c942feda514970b7ea221.jpg", + "image_caption": [ + "(b) Varying masking ratio" + ], + "image_footnote": [], + "bbox": [ + 285, + 306, + 455, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/11847d5d8073c01050251fecacb27c3f6503f1f979df108f36ad6e8dcee6d404.jpg", + "image_caption": [ + "(a) Varying JPEG compression rate", + "Figure 10. Analysis of robustness over (a) JPEG compression and (b) Gaussian blur. We provide the SSIM of rendered views (blue) and recovered hidden images (green)." + ], + "image_footnote": [], + "bbox": [ + 91, + 522, + 261, + 623 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c17513acd083b98e6cc55af05ab707a660cc625073f8f3e1f94cffd0c3a1bfd4.jpg", + "image_caption": [ + "(b) Varying Gaussian blur std.", + "Figure 11. Results using a universal decoder to embed (a) multiple images for the flower scene and (b) a common image into different scenes from the LLFF dataset. The successful application under these two setups suggests that StegaNeRF can generalize across different scenes, user identities, and hidden images." + ], + "image_footnote": [], + "bbox": [ + 285, + 522, + 459, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. More Discussions", + "text_level": 1, + "bbox": [ + 76, + 722, + 246, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How useful is steganography for NeRF? Although NeRF-based 3D content has yet to become mainstream, we believe it will play a major future role not only for social platforms, but also for 3D vision research and applications. As people upload their personal NeRF models online for viewing purposes, NeRF steganography for ownership identification is apparently an important feature. Moreover, future 3D vision research will likely demand large-scale NeRF datasets of real-world images, and in this context, NeRF steganography can be a crucial tool for responsible", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/496dc80a0391c8141c4ec35a8aefd20db91e807e929c522b0a9e2bc3868f8838.jpg", + "image_caption": [ + "Figure 8. Impact on visual quality when changing different components of the proposed StegaNeRF framework as in Tab. 3. We show the SSIM for the renderings and the recovered hidden images.", + "(a) Varying #watermark per scene" + ], + "image_footnote": [], + "bbox": [ + 504, + 292, + 700, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a453434f03bfc17b213a281f2a129ffa0e0c7fdc33b91a7ad23e2eb97f953732.jpg", + "image_caption": [ + "(b) Varying #scene per watermark" + ], + "image_footnote": [], + "bbox": [ + 704, + 292, + 901, + 414 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/25681a2bf618e4efda29e9dcad8c81ee20f70b9dc5bddd3864e537a23790e152.jpg", + "table_caption": [ + "Table 4. The further empirical study on StegaNeRF. Results are averaged on the selected LLFF scenes." + ], + "table_footnote": [], + "table_body": "
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
StegaNeRF28.130.85680.1430100.00.9486
No Empty Sampling28.120.85570.1466100.00.9364
NeRF Frozen28.150.85750.137750.000.3975
Standard NeRF28.150.85750.137750.00N/A
", + "bbox": [ + 503, + 579, + 890, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and ethical uses of training data and deployed models.", + "bbox": [ + 500, + 695, + 856, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hidden or Memorized? StegaNeRF follows the standard setup of NeRF which prescribes per-scene training. Therefore, it may be reasonable to wonder whether the decoder is trained to \"memorize\" the hidden image without it really being injected into NeRF. To address this concern, we perform an experiment where the NeRF weights are frozen after the first stage and only update the decoder of StegaNeRF. If the decoder actually memorizes the image by itself, then we should expect good recovery performance under this setup. However, as shown in Tab. 4, with NeRF Frozen, we observe a significant performance degradation. Therefore, we can verify that the decoder alone cannot trivially memorize", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "448", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the hidden image, and the information is actually hidden in the NeRF renderings. Furthermore, we extend StegaNeRF to generalize across different scenes and hidden images while sharing the same universal decoder. Fig. 11a shows the rendering and decoding quality of embedding the different numbers of hidden images to a scene (flower). Fig. 11b shows the performance of embedding a common image into a varying number of scenes from the LLFF dataset. While the performance slightly decreases in the generalized setting, it still remains at a comparable level.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Why not just share the rendered images instead of NeRF weights? Directly sharing the NeRF model gives end users the freedom to render at continuous viewpoints, which fundamentally differs from the traditional setting centered around discrete 2D image frames. It would not be realistic to assume people will stick with sharing 2D images instead of 3D weights, and therefore, it is necessary for the research community to see beyond 2D steganography and explore 3D NeRF steganography.", + "bbox": [ + 75, + 250, + 472, + 387 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Robustness: As shown in Fig. 12, in addition to using a resolution of $128 \\times 128$ for hidden images in most of our experiments, we also tested resolutions ranging from $96 \\times 96$ to $224 \\times 224$ . We found that increasing the size of embedded images had no noticeable impact on the performance of StegaNeRF, with only a $0.6\\%$ drop in rendering PSNR and a $1.2\\%$ reduction in hidden SSIM. We also demonstrated the robustness of StegaNeRF to different sampling approaches during rendering, such as avoiding sampling in empty areas [43, 48]. Instead of considering all the points on a ray during rendering, we disregard the contribution of points to the integration along the entire ray if their density falls below a threshold, like 0.001. As shown in Tab. 4, replacing the conventional rendering strategy with No Empty Sampling has minimal effect on the performance of StegaNeRF, revealing our robustness to different rendering fashions.", + "bbox": [ + 75, + 393, + 472, + 636 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/bd9ca95fb354943a142ca30d30f3b387482ba8c3fc3057cc4465b710876f9dc1.jpg", + "image_caption": [ + "Figure 12. Ablation on the resolutions of embedded images by the average results over the selected scenes from the LLFF dataset." + ], + "image_footnote": [], + "bbox": [ + 76, + 645, + 269, + 736 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6c2503ce199e0e20cbfe939a3d93590a725f4789d6d3c5f848b9bce2a363ad3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 645, + 467, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations: The current implementation is limited by the computational efficiency of NeRF, since the steganographic training stage requires performing gradient descent updates through the NeRF model. Another limitation is that, the recovery accuracy of the injected hidden information varies across different 3D scenes. More systematic analysis is required to understand the underlying factors that make some scenes easier to hide information than others.", + "bbox": [ + 75, + 780, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 90, + 619, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "With the progression of NeRF-based content, now serving as a pivotal element in diverse 3D applications, it is incumbent upon the research community to proactively tackle the imminent requirement to embed information within NeRF models as they are exchanged and disseminated across online platforms. This study introduces StegaNeRF, a pioneering framework for the automated embedding of steganographic information within NeRF renderings. We outline the essential components essential for achieving this functionality, unveiling potential challenges and revealing insights that can be instrumental for subsequent developers. By shedding light on the prospect of ownership identification within the realm of NeRF, this paper underscores the need for increased attention and concerted efforts toward addressing related challenges and avenues.", + "bbox": [ + 496, + 114, + 893, + 342 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 354, + 596, + 369 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Jia-Bin Huang, Michael Zollhoefer, Johannes Kopf, and Changil Kim. Learning Neural Light Fields With Ray-Space Embedding Networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19787-19797, 2022. 2", + "[2] Shumeet Baluja. Hiding Images in Plain Sight: Deep Steganography. Advances in Neural Information Processing Systems, 30, 2017. 1, 2, 3, 5", + "[3] Shumeet Baluja. Hiding Images Within Images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1685-1697, 2019. 1, 2", + "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-Nerf: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2", + "[5] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5460–5469, 2022. 2", + "[6] Oliver Benedens. Geometry-Based Watermarking of 3D Models. IEEE Computer Graphics and Applications, 19:46-55, 1999. 2", + "[7] Ning Bi, Qiyu Sun, Daren Huang, Zhihua Yang, and Jiwu Huang. Robust Image Watermarking Based on Multiband Wavelets and Empirical Mode Decomposition. IEEE Transactions on Image Processing, 16(8):1956-1966, 2007. 2", + "[8] Rajarathnam Chandramouli, Mehdi Kharrazi, and Nasir Memon. Image Steganography and Steganalysis: Concepts and Practice. In International Workshop on Digital Watermarking, pages 35-49. Springer, Springer, 2003. 2", + "[9] Chin-Chen Chang, Ju-Yuan Hsiao, and Chi-Shiang Chan. Finding Optimal Least-Significant-Bit Substitution in Image Hiding by Dynamic Programming Strategy. Pattern Recognition, 36(7):1583-1595, 2003. 2, 3, 5" + ], + "bbox": [ + 506, + 378, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "449", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast Generalizable Radiance Field Reconstruction From Multi-View Stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124–14133, 2021. 1", + "[11] Di Chen, Yu Liu, Lianghua Huang, Bin Wang, and Pan Pan. GeoAug: Data Augmentation for Few-Shot NeRF With Geometry Constraints. In European Conference on Computer Vision, pages 322-337. Springer, Springer, 2022. 1", + "[12] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital Watermarking and Steganography. Morgan Kaufmann, 2007. 1, 2", + "[13] Wenxue Cui, Shaohui Liu, Feng Jiang, Yongliang Liu, and Debin Zhao. Multi-Stage Residual Hiding for Image-Into-Audio Steganography. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2832-2836, 2020. 2", + "[14] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. In European Conference on Computer Vision, pages 636-654. Springer, 2022. 1", + "[15] Brandon Yushan Feng and Amitabh Varshney. SIGNET: Efficient Neural Representation for Light Fields. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14204-14213, 2021. 2", + "[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance Fields Without Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2", + "[17] Jessica Fridrich, Miroslav Goljan, and Rui Du. Detecting LSB Steganography in Color, and Gray-Scale Images. IEEE Multimedia, 8(4):22-28, 2001. 2", + "[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic Neural Radiance Fields for Monocular 4D Facial Avatar Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2", + "[19] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2", + "[20] Margarita Geleta, Cristina Punti, Kevin McGuinness, Jordi Pons, Cristian Canton, and Xavier Giro-i Nieto. PixInWav: Residual Steganography for Hiding Pixels in Audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2485-2489, 2022. 2", + "[21] Jamie Hayes and George Danezis. Generating Steganographic Images Via Adversarial Training. Advances in Neural Information Processing Systems, 30, 2017. 2", + "[22] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. EfficientNeRF Efficient Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12902-12911, 2022. 2", + "[24] Xudong Huang, Wei Li, Jie Hu, Hanting Chen, and Yunhe Wang. Refsr-nerf: Towards high fidelity and super resolution view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8244–8253, 2023. 1", + "[25] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. HiNet: Deep Image Hiding by Invertible Network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4733-4742, 2021. 2", + "[26] Gary C Kessler and Chet Hosmer. An Overview of Steganography. Advances in Computers, 83:51-107, 2011. 2", + "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification With Deep Convolutional Neural Networks. Communications of the ACM, 60(6):84-90, 2017. 5", + "[28] Jaeho Lee, Jihoon Tack, Namhoon Lee, and Jinwoo Shin. Meta-Learning Sparse Implicit Neural Representations. Advances in Neural Information Processing Systems, 34:11769-11780, 2021. 4", + "[29] Chenxin Li, Mingbao Lin, Zhiyuan Ding, Nie Lin, Yihong Zhuang, Yue Huang, Xinghao Ding, and Liujuan Cao. Knowledge condensation distillation. In European Conference on Computer Vision, pages 19-35. Springer, 2022. 7", + "[30] Chenxin Li, Wenao Ma, Liyan Sun, Xinghao Ding, Yue Huang, Guisheng Wang, and Yizhou Yu. Hierarchical deep network with uncertainty-aware semi-supervised learning for vessel segmentation. Neural Computing and Applications, pages 1-14, 2022. 4", + "[31] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 61-65. IEEE, 2021. 7", + "[32] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2", + "[33] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Y. Wang, and Shenlong Wang. NeurMiPs: Neural Mixture of Planar Experts for View Synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15681-15691, 2022. 2", + "[34] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural Sparse Voxel Fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2", + "[35] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason M. Saragih. Mixture of Volumetric Primitives for Efficient Neural Rendering. ACM Transactions on Graphics (TOG), 40:1 - 13, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "450", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Shao-Ping Lu, Rong Wang, Tao Zhong, and Paul L Rosin. Large-Capacity Image Steganography Based on Invertible Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10816-10825, 2021. 2", + "[37] Luma AI. https://lumalabs.ai/.2", + "[38] Xiyang Luo, Yinxiao Li, Huiwen Chang, Ce Liu, Peyman Milanfar, and Feng Yang. Dvmark: A Deep Multiscale Framework for Video Watermarking. ArXiv Preprint ArXiv:2104.12734, 2021. 2", + "[39] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210–7219, 2021. 2, 4", + "[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy Networks: Learning 3d Reconstruction in Function Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 1", + "[41] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local Light Field Fusion: Practical View Synthesis With Prescriptive Sampling Guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 4", + "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing Scenes As Neural Radiance Fields for View Synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 4", + "[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives With a Multiresolution Hash Encoding. ACM Transactions on Graphics (TOG), 41:1-15, 2022. 1, 2, 9", + "[44] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing Neural Radiance Fields for View Synthesis From Sparse Inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 1", + "[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning Continuous Signed Distance Functions for Shape Representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 1", + "[46] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2", + "[47] Tomás Pevny, Tomás Filler, and Patrick Bas. Using High-Dimensional Image Models to Perform Highly Undetectable Steganography. In International Workshop on Information Hiding, pages 161-177. Springer, Springer, 2010. 2", + "[48] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. In 2021 Interna" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on 3D Vision (3DV), pages 1106-1114. IEEE, 2021.9", + "[49] Polycam. https://poly.cam/.2", + "[50] Emil Praun, Hugues Hoppe, and Adam Finkelstein. Robust Mesh Watermarking. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, pages 49-56, 1999. 2", + "[51] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2", + "[52] Yinlong Qian, Jing Dong, Wei Wang, and Tieniu Tan. Deep Learning for Steganalysis Via Convolutional Neural Networks. In Media Watermarking, Security, and Forensics 2015, volume 9409, pages 171-180. SPIE, SPIE, 2015. 2", + "[53] Chen Quei-An. Nerf-pl: A Pytorch-Lighting Implementation of NeRF. https://github.com/kwea123/nerf_pl/.4", + "[54] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding Up Neural Radiance Fields With Thousands of Tiny Mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2", + "[55] Mennatallah M Sadek, Amal S Khalifa, and Mostafa GM Mostafa. Video Steganography: A Comprehensive Review. Multimedia Tools and Applications, 74(17):7063-7094, 2015. 2", + "[56] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-Aware Image Synthesis With Sparse Voxel Grids. ArXiv Preprint ArXiv:2206.07695, 2022. 1", + "[57] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 1", + "[58] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit Neural Representations With Periodic Activation Functions. Advances in Neural Information Processing Systems, 33, 2020. 1", + "[59] Vincent Sitzmann, Semon Rezchikov, William T. Freeman, Joshua B. Tenenbaum, and Frédo Durand. Light Field Networks: Neural Scene Representations With Single-Evaluation Rendering. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[60] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 2", + "[61] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct Voxel Grid Optimization: Super-Fast Convergence for Radiance Fields Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1, 2", + "[62] Abdelfatah A Tamimi, Ayman M Abdalla, and Omaima Al-Allaf. Hiding an Image Inside Another Image Using Variable-Rate Steganography. International Journal of Advanced Computer Science and Applications (IJACSA), 4(10), 2013. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "451", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-Nerf: Scalable Large Scene Neural View Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248-8258, 2022. 1, 2", + "[64] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible Hyperlinks in Physical Photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2126, 2020. 2", + "[65] Matthew Tancik, Pratul P Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T Barron, and Ren Ng. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. Advances in Neural Information Processing Systems, 33, 2020. 1", + "[66] Weixuan Tang, Shunquan Tan, Bin Li, and Jiwu Huang. Automatic Steganographic Distortion Learning Using a Generative Adversarial Network. IEEE Signal Processing Letters, 24(10):1547-1551, 2017. 2, 6", + "[67] Jeroen JG Van Merrienboer, Liesbeth Kester, and Fred Paas. Teaching complex rather than simple tasks: Balancing intrinsic and germane load to enhance transfer of learning. Applied Cognitive Psychology: The Official Journal of the Society for Applied Research in Memory and Cognition, 20(3):343-352, 2006. 4", + "[68] Mukund Varma, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that nef needs? In The Eleventh International Conference on Learning Representations, 2022. 2", + "[69] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and S. Tulyakov. R2L: Distilling Neural Radiance Field to Neural Light Field for Efficient Novel View Synthesis. In European Conference on Computer Vision. Springer, Springer, 2022. 2", + "[70] Huaiqing Wang and Shuozhong Wang. Cyber Warfare: Steganography vs. Steganalysis. Communications of the ACM, 47(10):76-82, 2004. 2", + "[71] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier PlenOctrees for Dynamic Radiance Field Rendering in Real-Time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 1", + "[72] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning Neural Implicit Surfaces by Volume Rendering for Multi-View Reconstruction. Advances in Neural Information Processing Systems, 34, 2021. 1", + "[73] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning Multi-View Image-Based Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690–4699, 2021. 2", + "[74] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image Quality Assessment: From Error Visibility" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "to Structural Similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5", + "[75] Xinyu Weng, Yongzhi Li, Lu Chi, and Yadong Mu. High-Capacity Convolutional Video Steganography With Temporal Residual Modeling. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, pages 87-95, 2019. 2", + "[76] Raymond B Wolfgang and Edward J Delp. A Watermark for Digital Images. In Proceedings of 3rd IEEE International Conference on Image Processing, volume 3, pages 219-222. IEEE, IEEE, 1996. 2", + "[77] Yue Wu, Guotao Meng, and Qifeng Chen. Embedding Novel Views in a Single JPEG Image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14519-14527, 2021. 2", + "[78] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2", + "[79] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. BungeeNeRF: Progressive Neural Radiance Field for Extreme Multi-Scale Scene Rendering. In The European Conference on Computer Vision (ECCV), volume 2, 2022. 2", + "[80] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1", + "[81] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4479–4489, 2023. 1", + "[82] Hyukryul Yang, Hao Ouyang, Vladlen Koltun, and Qifeng Chen. Hiding Video in Audio Via Reversible Generative Models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1100-1109, 2019. 2", + "[83] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5491–5500, 2022. 1, 4", + "[84] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for Real-Time Rendering of Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2", + "[85] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. Pixelnerf: Neural Radiance Fields From One or Few Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1", + "[86] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14448-14457, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "452", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A Structured Dictionary Perspective on Implicit Neural Representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 4", + "[88] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray Priors Through Reprojection: Improving Neural Radiance Fields for Novel View Extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18376-18386, 2022. 1", + "[89] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and Improving Neural Radiance Fields. ArXiv Preprint ArXiv:2010.07492, 2020. 2", + "[90] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features As a Perceptual Metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5", + "[91] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul E. Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural Factorization of Shape and Reflectance Under an Unknown Illumination. ACM Trans. Graph, 40:237:1-237:18, 2021. 1" + ], + "bbox": [ + 78, + 90, + 468, + 428 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "453", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/3c8d0cf6-dc7d-44ba-a804-37f576c9cfcc_model.json b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/3c8d0cf6-dc7d-44ba-a804-37f576c9cfcc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e0575c3e41cdd877bf8317ac35c5f899f651c01 --- /dev/null +++ b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/3c8d0cf6-dc7d-44ba-a804-37f576c9cfcc_model.json @@ -0,0 +1,3394 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "StegaNeRF: Embedding Invisible Information within Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.18, + 0.812, + 0.236 + ], + "angle": 0, + "content": "Chenxin Li\\(^{1*}\\), Brandon Y. Feng\\(^{2*}\\), Zhiwen Fan\\(^{3*}\\), Panwang Pan\\(^{4}\\), Zhangyang Wang\\(^{3}\\) \n\\(^{1}\\)Chinese University of Hong Kong, \\(^{2}\\)University of Maryland, \\(^{3}\\)University of Texas at Austin, \\(^{4}\\)ByteDance" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.286 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.302, + 0.474, + 0.651 + ], + "angle": 0, + "content": "Recent advancements in neural rendering have paved the way for a future marked by the widespread distribution of visual data through the sharing of Neural Radiance Field (NeRF) model weights. However, while established techniques exist for embedding ownership or copyright information within conventional visual data such as images and videos, the challenges posed by the emerging NeRF format have remained unaddressed. In this paper, we introduce StegaNeRF, an innovative approach for steganographic information embedding within NeRF renderings. We have meticulously developed an optimization framework that enables precise retrieval of hidden information from images generated by NeRF, while ensuring the original visual quality of the rendered images to remain intact. Through rigorous experimentation, we assess the efficacy of our methodology across various potential deployment scenarios. Furthermore, we delve into the insights gleaned from our analysis. StegaNeRF represents an initial foray into the intriguing realm of infusing NeRF renderings with customizable, imperceptible, and recoverable information, all while minimizing any discernible impact on the rendered images. For more details, please visit our project page: https://xggnet.github.io/StegaNeRF/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.681, + 0.21, + 0.698 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Implicit neural representation (INR) is an emerging concept where the network describes the data through its weights [14, 24, 40, 42, 45, 58, 65]. After training, the INR weights can then be used for content distribution, streaming, and even downstream inference tasks, all without sending or storing the original data. Arguably the most prominent INR is Neural Radiance Fields (NeRF) [42], where a network learns a continuous function mapping spatial coordinates to density and color. Due to its lightweight size and superb quality, NeRF has immense potential for 3D content representation in future vision and graphics applications." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.271, + 0.891, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.373, + 0.894, + 0.43 + ], + "angle": 0, + "content": "Figure 1. We introduce the new problem of NeRF steganography: hiding information in NeRF renderings. Our proposed framework, StegaNeRF, can embed and recover customized hidden information while preserving the original NeRF rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.893, + 0.533 + ], + "angle": 0, + "content": "While there is a plethora of work dedicated towards better quality [4,63,72,91], faster rendering [43,56,61,71,83], and sparse view reconstruction [10, 11, 44, 57, 80, 81, 85, 88], in this paper, we look beyond the horizon and explore a new question: Can we achieve steganography with NeRF?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Established digital steganography method [12] focus on embedding hidden messages in 2D images. The recent growth of deep learning and social media platforms further gives rise to many practical use cases of image steganography. As countless images and videos are shared online and even used to train deep learning models, 2D steganography methods [2,3] allow users and data providers to protect copyright, embed ownership, and prevent content abuse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Now, with the ongoing advances in 3D representations powered by NeRF, we envision a future where people frequently share their captured 3D content online just as they are currently sharing 2D images and videos online. Moreover, we are curious to explore the following research questions: 1) Injecting information into 2D images for copyright or ownership identification is common, but can we preserve such information when people share and render 3D scenes through NeRF? 2) NeRF can represent large-scale real-world scenes with training images taken by different people, but can we preserve these multiple source identities in the NeRF renderings to reflect the collaborative efforts required to reconstruct these 3D scenes? 3) Common image steganography methods embed either a hidden image or a message string into a given image, but can we allow different modalities of the hidden signal in NeRF steganography?" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.205, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "441" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "Driven by these questions, we formulate a framework to embed customizable, imperceptible, and recoverable information in NeRF renderings without sacrificing the visual quality. Fig. 1 presents an overview of our proposed framework, dubbed StegaNeRF. Unlike the traditional image steganography that embeds hidden signals only into a specific source image, we wish to recover the same intended hidden signal from NeRF rendered at arbitrary viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.452 + ], + "angle": 0, + "content": "Despite many established works on 2D steganography and hidden watermarks for image and video [2, 3, 9, 25, 36, 64], naively applying 2D steganography on the NeRF training images is not practical, since the embedded information easily gets lost in the actual NeRF renderings. In contrast, our framework enables reliable extraction of the hidden signal from NeRF renderings. During NeRF training, we jointly optimize a decoder network to extract hidden information from the NeRF renderings. To minimize the negative impact on the visual quality of NeRF, we identify weights with low impact on rendering and introduce a gradient masking strategy to steer the hidden steganographic information towards those low-importance weights. Extensive experimental results validate StegaNeRF balances between the rendering quality of novel views and the high-fidelity transmission of the concealed information." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.455, + 0.47, + 0.499 + ], + "angle": 0, + "content": "StegaNeRF presents the first exploration of hiding information in NeRF models for ownership identification, and our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.511, + 0.468, + 0.556 + ], + "angle": 0, + "content": "- We introduce the new problem of NeRF steganography and present the first effort to embed customizable, imperceptible, and recoverable information in NeRF." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.567, + 0.469, + 0.628 + ], + "angle": 0, + "content": "- We propose an adaptive gradient masking strategy to steer the injected hidden information towards the less important NeRF weights, balancing the objectives of steganography and rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.639, + 0.469, + 0.699 + ], + "angle": 0, + "content": "- We empirically validate the proposed framework on a diverse set of 3D scenes with different camera layouts and scene characteristics, obtaining high recovery accuracy without sacrificing rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.71, + 0.469, + 0.755 + ], + "angle": 0, + "content": "- We explore various scenarios applying StegaNeRF for ownership identification, with the additional support of multiple identities and multi-modal signals." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.511, + 0.469, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.218, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Neural Radiance Fields Following the success of NeRF [42], numerous techniques have been incorporated including ray re-parameterizations [4, 5], explicit spatial data structures [16, 23, 34, 43, 61, 84], caching and distillation [19, 22, 54, 69], ray-based representations [1, 15, 59], geometric primitives [33, 35], large-scale scenes [39, 63, 79, 89], generalizable architectures [60, 68, 73], and dynamic" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "settings [18, 32, 46, 51, 78]. Unlike prior works, this paper explores the uncharted problem of embedding information in NeRF renderings, with critical implications for copyright protection and ownership preservation. As early-stage NeRF-based products already become available [37,49] and more are expected, we believe now is the right time to open up the exploration of NeRF steganography." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.205, + 0.893, + 0.493 + ], + "angle": 0, + "content": "Image Steganography Steganography hides intended signals as invisible watermarks (e.g., hyperlinks, images) within the cover media called carriers (e.g., images, video and audio) [12, 26]. Classical methods focus on seeking and altering the least significant bits (LSBs) [17, 47, 62, 76] and transforming domain techniques [7, 8, 52, 70]. Prior research also uses deep neural networks for steganography [2, 3, 21, 66]. Among them, DeepStega [2] conceals a hidden image within an equal-sized carrier image. Subsequent works [25, 36] use invertible networks to improve the performance of deep image steganography. Another line of work conceals information in other carrier media like audio [13, 20, 82] and video [38, 55, 75]. The above advances all play a critical part in the era when traditional media formats like images and videos are dominant. However, as MLP-based neural representations of 3D scenes are gaining momentum to become a major format of visual data, extending steganography to NeRF is bound to become an important problem in the upcoming future." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.667 + ], + "angle": 0, + "content": "Lifting Steganography to 3D Prior to NeRF, meshes are commonly used to represent 3D shapes. Early pioneers apply steganography 3D meshes [6, 50] for copyright protection when meshes are exchanged and edited. More recent work [77] has also explored embedding multi-plane images within a JPEG-compressed image. Their problem can be regarded as a special case of 2D steganography, hiding multiple 2D images inside a single 2D image. In contrast, we try to hide a natural image into a 3D scene representation (NeRF), fundamentally differing from these prior methods where 2D images act as the carrier of hidden information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.59, + 0.695 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.796 + ], + "angle": 0, + "content": "The goal of StegaNeRF is to inject customized (steganographic) information into the NeRF weights with imperceptible visual changes when rendering. Given a NeRF with weights \\(\\theta_0\\) and the information \\(I\\) to be injected, when we render with \\(\\theta\\) on any viewpoint, we hope the injected \\(I\\) can be recovered by a decoder \\(F_{\\psi}\\) with learnable weights \\(\\psi\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A seemingly obvious solution is to use prior image steganography methods to 1) inject \\(I\\) on training images, 2) train NeRF on those images with embedded information, 3) apply their provided \\(F_{\\psi}\\) to extract \\(I\\) from NeRF-rendered images. This approach has been successfully applied for GAN fingerprinting [86]. However, it fails in the NeRF context (see Fig. 3), where the subtle changes induced by prior" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "442" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.619, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.09, + 0.891, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.893, + 0.342 + ], + "angle": 0, + "content": "Figure 2. StegaNeRF training overview. At the first stage, we optimize \\(\\theta_0\\) with standard NeRF training. At the second stage, we initialize \\(\\theta\\) with \\(\\theta_0\\) and optimize for the steganography objectives. We train the decoder \\(F_{\\psi}\\) to recover hidden information from StegaNeRF renderings and no hidden information from original NeRF renderings. We introduce Classifier Guided Recovery to improve the accuracy of recovered information, and Adaptive Gradient Masking to balance between steganography ability and rendering visual quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.469, + 0.413 + ], + "angle": 0, + "content": "steganography methods easily get smoothed out, inhibiting the decoder from identifying the subtle patterns necessary for information recovery." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.469, + 0.504 + ], + "angle": 0, + "content": "Such shortcomings of off-the-shelf steganographic methods are not surprising since they are developed for the traditional setting where 2D images are the ultimate form of visual information. In contrast, our problem setting involves the new concept of INR as the underlying representation, and 2D images are just the final output rendered by NeRF." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.303, + 0.53 + ], + "angle": 0, + "content": "3.1. Two-Stage Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.469, + 0.627 + ], + "angle": 0, + "content": "Recognizing the formulation difference due to the emergence of implicit representation with NeRF, we move away from the traditional 2D image-based pipeline that trains an encoder network to inject subtle changes to the given 2D images. Instead, we incorporate the steganography objective into the gradient-based learning process of NeRF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.469, + 0.779 + ], + "angle": 0, + "content": "We re-design the training as a two-stage optimization. The first stage is the original NeRF training procedure, involving the standard photometric loss between the rendered and ground truth pixels to guarantee the visual quality. After finishing the weights update \\(\\theta_0\\) at the first stage, we dedicate the second stage to obtain the final NeRF weights \\(\\theta\\) containing steganographic information. We introduce several techniques to achieve robust information recovery with imperceptible impact on the rendered images. The training workflow is depicted in Fig. 2 and Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.283, + 0.803 + ], + "angle": 0, + "content": "3.2. Information Recovery" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Let \\(P\\) denote the camera pose at which we render an image from a NeRF network. We want to recover \\(I\\) from the rendered image \\(\\theta(P)\\). Importantly, we also want to avoid false positives on images \\(\\theta_0(P)\\) rendered by the original network without steganography ability, even if the images rendered by \\(\\theta_0\\) and \\(\\theta\\) look visually identical. Therefore, we" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.368, + 0.596, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.369, + 0.689, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.369, + 0.735, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.369, + 0.838, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.845, + 0.369, + 0.884, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.505, + 0.892, + 0.617 + ], + "angle": 0, + "content": "Figure 3. Results of applying prior 2D steganography methods, LSB [9] (top) and DeepStega [2] (bottom). From left to right, we show (a) training image after applying 2D steganography, (b) residual error of (a) over ground truth, (c) hidden image recovered from (a), (d) residual error of NeRF rendering at the pose of (a), and (e) hidden image recovered from (d). Prior 2D steganography methods fail in the NeRF context since the hidden information injected in training images mostly disappear in NeRF renderings." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.643, + 0.89, + 0.658 + ], + "angle": 0, + "content": "optimize \\(\\theta\\) to minimize the following contrastive loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.686, + 0.891, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e c} ^ {+} = \\left| F _ {\\psi} (\\boldsymbol {\\theta} (P)) - I \\right|, \\quad \\mathcal {L} _ {d e c} ^ {-} = \\left| F _ {\\psi} \\left(\\boldsymbol {\\theta} _ {0} (P)\\right) - \\varnothing \\right|, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.821 + ], + "angle": 0, + "content": "where \\(\\varnothing\\) is an empty image with all zeros. Effectively, \\(L_{dec}^{+}\\) serves as a regularization term, facilitating the decoder to recover the embedded image pattern based on renderings obtained from a StegaNeRF. In contrast, \\(L_{dec}^{-}\\) prevents the decoder from falsely producing any plausible image pattern when given renderings from a Standard NeRF that is without hidden signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Classifier-Guided Recovery The decoder \\(F_{\\psi}\\) is easily implemented as a U-Net to decode \\(I\\) as the form of 2D images, but accurately recovering all the details in \\(I\\) might be challenging. Therefore, we additionally train a classifier" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "443" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.4, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Train StegaNeRF on a single scene" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.109, + 0.468, + 0.139 + ], + "angle": 0, + "content": "Data: Training images \\(\\{Y_i\\}\\) with poses \\(\\{P_i\\}\\), hidden information \\(I\\), learning rate \\(\\eta = [\\eta_0, \\eta_1]\\)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.14, + 0.408, + 0.154 + ], + "angle": 0, + "content": "Output: Steganographic NeRF \\(\\theta\\) and decoder \\(F_{\\psi}\\)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.155, + 0.369, + 0.169 + ], + "angle": 0, + "content": "Initialize NeRF \\(\\theta_0\\) and decoder network \\(F_{\\psi}\\)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.17, + 0.432, + 0.185 + ], + "angle": 0, + "content": "Optimize \\(\\theta_0\\) on \\(\\{Y_i, P_{ij}\\}\\) with standard NeRF training" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.186, + 0.329, + 0.2 + ], + "angle": 0, + "content": "Compute mask \\(m\\) for \\(\\theta_0\\) as in Eq. (3)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.201, + 0.282, + 0.214 + ], + "angle": 0, + "content": "for each training iteration \\(t\\) do" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.217, + 0.451, + 0.232 + ], + "angle": 0, + "content": "Randomly sample a training pose \\(P_{i}\\) and render \\(\\theta (P_i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.233, + 0.451, + 0.248 + ], + "angle": 0, + "content": "Compute stega. losses \\(\\mathcal{L}_{dec}^{c}\\), \\(\\mathcal{L}_{dec}^{+}\\), \\(\\mathcal{L}_{dec}^{-}\\) as Eq. (1), (2)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.442, + 0.262 + ], + "angle": 0, + "content": "Compute \\(\\mathcal{L}_{rgb}\\) and combine total loss \\(\\mathcal{L}\\) as in Eq. (4)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.262, + 0.43, + 0.28 + ], + "angle": 0, + "content": "Update \\(\\pmb{\\theta}\\) with \\(\\eta_0\\cdot (\\frac{\\partial\\mathcal{L}}{\\partial\\pmb{\\theta}}\\odot \\pmb {m})\\) and \\(F_{\\psi}\\) with \\(\\eta_{1}\\cdot \\frac{\\partial\\mathcal{L}}{\\partial\\psi}\\)" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.281, + 0.133, + 0.294 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.47, + 0.367 + ], + "angle": 0, + "content": "network \\( F_{\\psi}^{c} \\) to solve the easier task of classifying whether the given NeRF rendering contains hidden information. \\( F_{\\psi}^{c} \\) is optimized by the following cross-entropy loss [30, 67]:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.376, + 0.469, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e c} ^ {c} = - \\log \\left(F _ {\\psi} ^ {c} (\\boldsymbol {\\theta} (P))\\right) - \\log \\left(1 - F _ {\\psi} ^ {c} \\left(\\boldsymbol {\\theta} _ {0} (P)\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.469, + 0.45 + ], + "angle": 0, + "content": "We then use its prediction to guide the process of decoding pixel-wise information by adding the classification output as input to \\( F_{\\psi} \\), such that \\( F_{\\psi}(x) = F_{\\psi}\\big(x,F_{\\psi}^{c}(x)\\big) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.469, + 0.526 + ], + "angle": 0, + "content": "Although the above discussion focuses on hiding images, our framework can be easily extended to embed other modalities like strings, text, or even audio, all of which may be represented as 1D vectors. We can simply modify the architecture of \\( F_{\\psi} \\) to have a 1D prediction branch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.349, + 0.551 + ], + "angle": 0, + "content": "3.3. Preserving Perceptual Identity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.469, + 0.708 + ], + "angle": 0, + "content": "Since we want to hide information without affecting the visual perception of the rendered output, an intuitive regularization is to penalize how much \\(\\theta\\) deviates from \\(\\theta_0\\). However, we find that naively summing the deviation penalty across all weights makes it difficult for the NeRF network to adjust its weights for the steganographic objective. Instead, motivated by the fact that INR weights are not equally important and exhibit strong sparsity [28, 87], we propose an adaptive gradient masking strategy to encode the hidden information on specific weights' groups." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.469, + 0.755 + ], + "angle": 0, + "content": "Formally, given the initial set of weights \\(\\theta_0\\in \\mathbb{R}^N\\) , we compute a soft importance mask \\(m\\in \\mathbb{R}^{N}\\) . Specifically, for each weight \\(\\pmb {w}_i\\in \\pmb {\\theta}_0\\) , we compute:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.764, + 0.469, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {m} _ {i} = \\frac {\\left| \\boldsymbol {w} _ {i} \\right| ^ {- \\alpha}}{\\sum_ {i} ^ {N} \\left| \\boldsymbol {w} _ {i} \\right| ^ {- \\alpha}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where the magnitude \\(|\\cdot|\\) of each model weight \\(w_{i}\\) serves as its individual importance indicator, and \\(\\alpha > 0\\) is a scaling factor of power controlling the relative distribution of importance across the weights. We mask the gradient as \\(\\frac{\\partial\\mathcal{L}}{\\partial\\theta} \\odot m\\) when optimizing \\(\\theta\\) based on the total loss \\(\\mathcal{L}\\) in the second stage, where \\(\\odot\\) is a Hadamard product. Effectively," + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.839, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 Typical usage scenario of StegaNeRF" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.111, + 0.809, + 0.125 + ], + "angle": 0, + "content": "1: Alice captures some images of a 3D scene" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.127, + 0.89, + 0.141 + ], + "angle": 0, + "content": "2: Alice trains a StegaNeRF to hide a personalized image" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.142, + 0.89, + 0.17 + ], + "angle": 0, + "content": "3: Alice shares the model \\(\\theta\\) online for other people to enjoy and explore the 3D scene themselves" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.171, + 0.89, + 0.201 + ], + "angle": 0, + "content": "4: Bob grabs the model \\(\\theta\\) and reposts it with his own account without crediting Alice or asking for permission" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.202, + 0.89, + 0.23 + ], + "angle": 0, + "content": "5: Alice sees Bob's post, deploys the decoder \\(F_{\\psi}\\), and verifies the owner of \\(\\theta\\) is Alice, not Bob" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.232, + 0.89, + 0.262 + ], + "angle": 0, + "content": "6: Bob takes down the post or gets banned for copyright infringement" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.111, + 0.89, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.336 + ], + "angle": 0, + "content": "more significant weights are \"masked out\" to minimize the impact of steganographic learning on the rendered visual quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.396 + ], + "angle": 0, + "content": "We retain the standard photometric error in steganography learning to maintain NeRF rendering fidelity: \\(\\mathcal{L}_{rgb} = |\\pmb {\\theta}(P) - \\pmb {\\theta}_0(P)|\\). The overall training loss at the second stage can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.406, + 0.892, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {0} \\mathcal {L} _ {d e c} ^ {c} + \\lambda_ {1} \\mathcal {L} _ {d e c} ^ {+} + \\lambda_ {2} \\mathcal {L} _ {d e c} ^ {-} + \\lambda_ {3} \\mathcal {L} _ {r g b}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.445, + 0.633, + 0.463 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.722, + 0.487 + ], + "angle": 0, + "content": "4.1. Implementation Details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Dataset. We use common datasets LLFF [41] and NeRF-Synthetic [42], with forward scenes {flower, fern, fortress, room} from LLFF and \\(360^{\\circ}\\) scenes {lego, drums, chair} from NeRF-Synthetic. We further experiment on the Brandenburg Gate scene from NeRF-W dataset [39], with over 800 views of in-the-wild collected online." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Training. On LLFF and NeRF-Synthetic, we adopt Plenoxels [83] as the NeRF backbone architecture for efficiency. On NeRF-W, we use the available PyTorch implementation [53]. The first stage of training is performed according to the standard recipes of those implementations. We then perform the second stage of steganography training for 55 epochs unless otherwise noted. On LLFF and NeRF-W, we downsize the training images by 4 times following common practice, and we use the original size on NeRF-Synthetic." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.84 + ], + "angle": 0, + "content": "For hyper-parameters in Eq. (4), we set the weight of NeRF reconstruction error \\(\\lambda_3 = 1\\) for all experiments. We set \\(\\lambda_0 = 0.01, \\lambda_1 = 0.5, \\lambda_2 = 0.5\\) for all scenes on LLFF dataset, and \\(\\lambda_0 = 0.1, \\lambda_1 = 1, \\lambda_2 = 1\\) for the scenes in NeRF-Synthetic, and \\(\\lambda_0 = 0.05, \\lambda_1 = 1, \\lambda_2 = 1\\) for NeRF-W. For Eq. (3), we set \\(\\alpha = 3\\) for all scenes. We run experiments on one NVIDIA A100 GPU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation. We assume a typical authentication scenario shown in Alg.2. We consider the recovery quality of the embedded information, including the metrics of classification accuracy (Acc.) of the classifier, and the structural sim" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "444" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.891, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.895, + 0.371 + ], + "angle": 0, + "content": "Figure 4. Results on the NeRF-Synthetic dataset. Within each column, we show the StegaNeRF rendering, residual error from the initial NeRF rendering, and the recovered hidden image. We show the SSIM for the StegaNeRF renderings and the recovered hidden images." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.394, + 0.475, + 0.493 + ], + "angle": 0, + "content": "Table 1. Quantitative results of StegaNeRF rendering and hidden information recovery. Standard NeRF is the initial NeRF \\(\\theta_0\\) with standard training (upper-bound rendering performance). Prior 2D steganography fails after NeRF training while StegaNeRF successfully embeds and recovers hidden information with minimal impact on the rendering quality. Results are averaged over the selected LLFF and NeRF-Synthetic scenes." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.503, + 0.47, + 0.651 + ], + "angle": 0, + "content": "
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
Standard NeRF28.150.85750.137750.00N/A
LSB [9]28.130.85690.1452N/A0.0132
DeepStega [2]26.940.84320.1641N/A0.2098
StegaNeRF (Ours)28.130.85630.1460100.00.9730
Standard NeRF31.130.96060.031050.00N/A
LSB [9]31.120.96040.0310N/A0.0830
DeepStega [2]31.130.96060.0313N/A0.2440
StegaNeRF (Ours)30.960.95830.029099.720.9677
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.471, + 0.754 + ], + "angle": 0, + "content": "ality (SSIM) [74] of the hidden image recovered by the decoder. We evaluate the final NeRF renderings with PSNR, SSIM and LPIPS [90]. All the metrics are computed on the test set and averaged over all the scenes and embedded images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.4, + 0.781 + ], + "angle": 0, + "content": "4.2. Case I: Embedding in a Single Scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.471, + 0.849 + ], + "angle": 0, + "content": "We first explore the preliminary case of ownership identification on a specific NeRF scene. We select random images from ImageNet [27] as the hidden information to be injected to NeRF renderings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Failure of 2D Baseline Due to the lack of prior study on NeRF steganography, we consider a baseline from 2D image steganography by training NeRF from scratch with" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.394, + 0.895, + 0.482 + ], + "angle": 0, + "content": "Table 2. Quantitative results of NeRF rendering and hidden information recovery. We consider two conditions, embedding each scene with a common hidden image (One-for-All) or a scene-specific hidden image (One-for-All). We report the quality difference compared to the single-scene settings as \\(\\Delta_{SSIM}^{(10^{-2})}\\). Results are averaged over the selected LLFF scenes." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.492, + 0.891, + 0.593 + ], + "angle": 0, + "content": "
SettingNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS ↓Δ(10-2)SSIM↑Acc. (%)↑SSIM↑Δ(10-2)SSIM↑
One-for-All24.990.80130.1786-0.10100.000.9860+0.19
One-for-One24.990.80160.1779-0.07100.000.9122-7.19
One-for-All27.900.85130.1236+0.01100.000.9844-0.76
One-for-One27.900.85150.1195+0.03100.000.9448-4.45
One-for-All30.270.84980.1289+0.02100.000.9430-5.42
One-for-One30.120.84800.1302-0.16100.000.9102-8.67
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.895, + 0.916 + ], + "angle": 0, + "content": "the watermarked images. We implement two off-the-shelf steganography methods including a traditional machine learning approach called Least Significant Bit (LSB [9]), and a deep learning pipeline as DeepStega [2]. An ideal case is that the embedded information can be recovered from the synthesized novel views, indicating the successful NeRF steganography. However, as can be seen in Fig. 3, the embedded information containing the hidden images cannot be recovered from the NeRF renderings. By analyzing the residual maps of training views (between GT training views and the watermarked) and novel views (between GT novel views and the actual rendering), we observe the subtle residuals to recover the hidden information are smoothed out in NeRF renderings, and similar failures occur on other datasets as well. Therefore, for the rest of our study, we mainly focus on analyzing our new framework that performs direct steganography on 3D NeRF." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "445" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.177, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.092, + 0.271, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.092, + 0.331, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.316, + 0.245, + 0.328 + ], + "angle": 0, + "content": "(a) Per-Scene" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.092, + 0.447, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.092, + 0.543, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.092, + 0.609, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.315, + 0.565, + 0.328 + ], + "angle": 0, + "content": "(b) Multi-Scene: One-for-All" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.092, + 0.718, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.092, + 0.814, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.092, + 0.877, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.315, + 0.835, + 0.328 + ], + "angle": 0, + "content": "(c) Multi-Scene: One-for-One" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.342, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Figure 5. Results on three multi-scene settings. Within each column, we show the StegaNeRF rendering, residual error from the initial NeRF rendering, and recovered hidden image. We show the SSIM for the StegaNeRF renderings and the recovered hidden images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.473, + 0.573 + ], + "angle": 0, + "content": "Results Tab. 1 contains quantitative results on LLFF and NeRF-Synthetic scenes. While NeRF trained by 2D steganography methods hardly recovers the embedded information, StegaNeRF accurately recovers the hidden image with minimal impact on the rendering quality measured by PSNR. Fig. 4 provides qualitative results of StegaNeRF on three embedded images on the NeRF-Synthetic scenes. An interesting observation is the common regions where hidden information (high residual error) emerges in the renderings, e.g., the right rear side of lego and left handrail of chair. We also notice that, within each scene, these regions are persistent across multiple viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.418, + 0.608 + ], + "angle": 0, + "content": "4.3. Case II: Embedding in Multiple Scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.47, + 0.785 + ], + "angle": 0, + "content": "Settings We extend our steganography scheme to embed information within multiple scenes at once. Specially, we use three LLFF scenes \\(\\{flower, fern, fortress\\}\\) to test the two sub-settings of multi-scene embedding with: (1) One-for-All, a common hidden image and (2) One-for-One scene-specific hidden images. All scenes share the same decoder \\(F_{\\psi}\\) and classifier \\(F_{\\psi}^{c}\\). The difference between the two sub-settings is the number of hidden images that \\(F_{\\psi}\\) and \\(F_{\\psi}^{c}\\) need to identify and recover. We sample one scene for every training epoch, and due to the increased data amount, we increase training epochs until convergence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Results Tab. 2 provides quantitative results on the multi-scene setting. The performance drop compared to single-scene training is sometimes noticeable, but it is not surprising due to the inherent requirement of per-scene fitting for our NeRF framework. Fig. 5 presents visual comparisons of multi-scene steganography against single-scene scheme." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.39, + 0.848, + 0.407 + ], + "angle": 0, + "content": "4.4. Case III: Embedding Multiple Identities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.892, + 0.585 + ], + "angle": 0, + "content": "Settings Constructing NeRF of large-scale cultural landmarks is a promising application, and community contributions are crucial to form the training images. Since every NeRF prediction is indebted to some particular training images, it would be meaningful to somehow identify the contributing users in the rendered output. Specifically, we present a proof-of-concept based on the following scenario: Given a collection of user-uploaded training images, our task is to apply StegaNeRF to ensure the final NeRF renderings hide subtle information about the relevant identities whose uploaded images help generate the current rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.893, + 0.804 + ], + "angle": 0, + "content": "To simulate this multi-user scenario in the public NeRF-W dataset, we randomly select \\(M\\) anchor views with different viewing positions, and then find the \\(K\\) nearest neighbour views to each anchor to form their respective clusters. We set \\(M = 3\\), \\(K = 20\\) in experiments. We assume a common contributor identity for each cluster, and we want the NeRF rendering to contain customized hidden information about those \\(M\\) contributors when we render within the spatial regions spanned by their own cluster. Thus, our classifier network \\(F_{\\phi}^{c}\\) is modified to output \\(M\\)-class cluster predictions and another class for identity outside of the \\(M\\) clusters. Since the decoder should extract no information for views outside of those \\(M\\) clusters to prevent false positives, we also compute \\(\\mathcal{L}_{dec}^{-}\\) (1) and \\(\\mathcal{L}_{dec}^{c}\\) (2) for those poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results We employ the same network backbone as NeRF-W [66] to handle in-the-wild images with different time and lighting effects. Fig. 6 presents qualitative results of embedding multiple identities in a collaborative large-scale NeRF." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "446" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.094, + 0.887, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.893, + 0.355 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of NeRF steganography on NeRF-W dataset. Within each block we show the StegaNeRF rendering, the initial NeRF rendering, the residual error between StegaNeRF and initial, and the recovered hidden image. At the bottom right, we provide the SSIM computed with their respective ground truth. \"Unmarked Identity\" denotes the remaining training views not related to any of the considered users, and the decoder does not recover meaningful information from renderings at those poses." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.381, + 0.232, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.473, + 0.232, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.381, + 0.306, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.432, + 0.302, + 0.441 + ], + "angle": 0, + "content": "SSIM=0.9709" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.381, + 0.388, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.432, + 0.419, + 0.44 + ], + "angle": 0, + "content": "PSNR=76.22" + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.383, + 0.461, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.432, + 0.445, + 0.44 + ], + "angle": 0, + "content": "\\(= 76.22\\)" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.45, + 0.379, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.45, + 0.445, + 0.459 + ], + "angle": 0, + "content": "Acc. \\(= 100.0\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.469, + 0.307, + 0.478 + ], + "angle": 0, + "content": "Hidden Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.469, + 0.422, + 0.477 + ], + "angle": 0, + "content": "Hidden Audio" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.481, + 0.304, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.524, + 0.3, + 0.532 + ], + "angle": 0, + "content": "\\(\\mathrm{SSIM} = 0.9688\\)" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.48, + 0.388, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.524, + 0.42, + 0.533 + ], + "angle": 0, + "content": "PSNR=62.84" + }, + { + "type": "image", + "bbox": [ + 0.394, + 0.48, + 0.467, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.524, + 0.42, + 0.533 + ], + "angle": 0, + "content": "62.84" + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.536, + 0.345, + 0.544 + ], + "angle": 0, + "content": "Hidden Text" + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.547, + 0.379, + 0.556 + ], + "angle": 0, + "content": "`The Owner Of These Horns'" + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.547, + 0.452, + 0.556 + ], + "angle": 0, + "content": "Acc. \\(= 100.0\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.648 + ], + "angle": 0, + "content": "Figure 7. Multi-modal hidden information. We show StegaNeRF rendering, recovered hidden images, audio (waveform and spectrum), and text. Metrics for hidden information in each modality are labeled respectively. For the StegaNeRF rendering, we report both SSIM and the relative change from initial NeRF renderings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.668, + 0.469, + 0.684 + ], + "angle": 0, + "content": "4.5. Case IV: Embedding Multi-Modal Information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.47, + 0.752 + ], + "angle": 0, + "content": "Settings We further show the potential of StegaNeRF in embedding multi-modal hidden information, such as images, audio, and text. We modify the decoder network to build a modal-specific decoder for each modality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Results Fig. 7 shows recovered multi-modal embedded signals in the trex and horns scenes from LLFF. Evidently, the StegaNeRF framework can easily extend to embed multi-modal information with high recovery performance without scarifying the rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.242, + 0.862 + ], + "angle": 0, + "content": "4.6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The effect of removing each component of StegaNeRF is presented in Tab. 3. Standard NeRF uses the initial NeRF" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.379, + 0.891, + 0.407 + ], + "angle": 0, + "content": "Table 3. Ablation study of different components of StegaNeRF. Results are averaged on the selected LLFF scenes." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.535 + ], + "angle": 0, + "content": "
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
StegaNeRF28.130.85680.1430100.00.9486
No Classifier26.850.80770.2417N/A0.5629
No Classifier Guided27.120.82390.2073100.00.6716
No Gradient Masking27.860.83750.1710100.00.8822
No Soft Masking28.050.85580.152694.440.8751
Standard NeRF28.150.85750.137750.00N/A
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.892, + 0.741 + ], + "angle": 0, + "content": "rendering. No Classifier removes the classifier \\( F_{\\psi}^{c} \\), while No Classifier Guided retains the classification task (hence impact on NeRF rendering) but does not condition the decoder on classifier output. No Gradient Masking removes the proposed masking strategy (Sec. 3.3), and No Soft Masking uses the binary mask [29, 31] with a threshold of 0.5. Our StegaNeRF makes a good balance between the rendering quality and the decoding accuracy. Fig. 8 shows the visual impact of removing each component. It appears that when any component is removed, the performance drops accordingly, revealing the effectiveness of our design." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Fig. 9 shows the impact of varying gradient masking strategies. We can see a different degree of affected by ratio variation on hidden recovery and rendering quality. Fig. 10 reports the performance of StegaNeRF against the common perturbations including JPEG compression and Gaussian noise. The lines show the average accuracies across selected scenes and the shaded regions indicate the range of 0.5 standard deviation. It appears that the hidden recovery of StegaNeRF is robust to various JPEG compression ratios and Gaussian blur degradation." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "447" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.887, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.215, + 0.193, + 0.227 + ], + "angle": 0, + "content": "(a) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.215, + 0.323, + 0.227 + ], + "angle": 0, + "content": "(b) No Classifier" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.215, + 0.475, + 0.227 + ], + "angle": 0, + "content": "(c) No Classifier Guided" + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.215, + 0.618, + 0.227 + ], + "angle": 0, + "content": "(d) No Gradient Masking" + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.215, + 0.743, + 0.227 + ], + "angle": 0, + "content": "(e) No Soft Masking" + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.215, + 0.858, + 0.227 + ], + "angle": 0, + "content": "(f) StegaNeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.892, + 0.271 + ], + "angle": 0, + "content": "Figure 8. Impact on visual quality when changing different components of the proposed StegaNeRF framework as in Tab. 3. We show the SSIM for the renderings and the recovered hidden images." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.307, + 0.262, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.412, + 0.216, + 0.424 + ], + "angle": 0, + "content": "(a) Varying \\(\\alpha\\)" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.307, + 0.457, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.412, + 0.447, + 0.424 + ], + "angle": 0, + "content": "(b) Varying masking ratio" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.436, + 0.47, + 0.492 + ], + "angle": 0, + "content": "Figure 9. Analysis of gradient masking. (a) Varying \\(\\alpha\\) from Eq. (3). (b) Varying portions of masked out weights from the gradient updates at steganography learning. We provide the SSIM of rendered views (blue) and recovered hidden images (green)." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.523, + 0.263, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.627, + 0.274, + 0.638 + ], + "angle": 0, + "content": "(a) Varying JPEG compression rate" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.523, + 0.46, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.627, + 0.46, + 0.638 + ], + "angle": 0, + "content": "(b) Varying Gaussian blur std." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.693 + ], + "angle": 0, + "content": "Figure 10. Analysis of robustness over (a) JPEG compression and (b) Gaussian blur. We provide the SSIM of rendered views (blue) and recovered hidden images (green)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.723, + 0.248, + 0.737 + ], + "angle": 0, + "content": "5. More Discussions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "How useful is steganography for NeRF? Although NeRF-based 3D content has yet to become mainstream, we believe it will play a major future role not only for social platforms, but also for 3D vision research and applications. As people upload their personal NeRF models online for viewing purposes, NeRF steganography for ownership identification is apparently an important feature. Moreover, future 3D vision research will likely demand large-scale NeRF datasets of real-world images, and in this context, NeRF steganography can be a crucial tool for responsible" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.294, + 0.702, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.432, + 0.697, + 0.445 + ], + "angle": 0, + "content": "(a) Varying #watermark per scene" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.294, + 0.903, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.432, + 0.891, + 0.445 + ], + "angle": 0, + "content": "(b) Varying #scene per watermark" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.457, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Figure 11. Results using a universal decoder to embed (a) multiple images for the flower scene and (b) a common image into different scenes from the LLFF dataset. The successful application under these two setups suggests that StegaNeRF can generalize across different scenes, user identities, and hidden images." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.541, + 0.891, + 0.569 + ], + "angle": 0, + "content": "Table 4. The further empirical study on StegaNeRF. Results are averaged on the selected LLFF scenes." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.58, + 0.892, + 0.676 + ], + "angle": 0, + "content": "
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
StegaNeRF28.130.85680.1430100.00.9486
No Empty Sampling28.120.85570.1466100.00.9364
NeRF Frozen28.150.85750.137750.000.3975
Standard NeRF28.150.85750.137750.00N/A
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.696, + 0.857, + 0.711 + ], + "angle": 0, + "content": "and ethical uses of training data and deployed models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Hidden or Memorized? StegaNeRF follows the standard setup of NeRF which prescribes per-scene training. Therefore, it may be reasonable to wonder whether the decoder is trained to \"memorize\" the hidden image without it really being injected into NeRF. To address this concern, we perform an experiment where the NeRF weights are frozen after the first stage and only update the decoder of StegaNeRF. If the decoder actually memorizes the image by itself, then we should expect good recovery performance under this setup. However, as shown in Tab. 4, with NeRF Frozen, we observe a significant performance degradation. Therefore, we can verify that the decoder alone cannot trivially memorize" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "448" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "the hidden image, and the information is actually hidden in the NeRF renderings. Furthermore, we extend StegaNeRF to generalize across different scenes and hidden images while sharing the same universal decoder. Fig. 11a shows the rendering and decoding quality of embedding the different numbers of hidden images to a scene (flower). Fig. 11b shows the performance of embedding a common image into a varying number of scenes from the LLFF dataset. While the performance slightly decreases in the generalized setting, it still remains at a comparable level." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.473, + 0.388 + ], + "angle": 0, + "content": "Why not just share the rendered images instead of NeRF weights? Directly sharing the NeRF model gives end users the freedom to render at continuous viewpoints, which fundamentally differs from the traditional setting centered around discrete 2D image frames. It would not be realistic to assume people will stick with sharing 2D images instead of 3D weights, and therefore, it is necessary for the research community to see beyond 2D steganography and explore 3D NeRF steganography." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.473, + 0.637 + ], + "angle": 0, + "content": "Robustness: As shown in Fig. 12, in addition to using a resolution of \\(128 \\times 128\\) for hidden images in most of our experiments, we also tested resolutions ranging from \\(96 \\times 96\\) to \\(224 \\times 224\\). We found that increasing the size of embedded images had no noticeable impact on the performance of StegaNeRF, with only a \\(0.6\\%\\) drop in rendering PSNR and a \\(1.2\\%\\) reduction in hidden SSIM. We also demonstrated the robustness of StegaNeRF to different sampling approaches during rendering, such as avoiding sampling in empty areas [43, 48]. Instead of considering all the points on a ray during rendering, we disregard the contribution of points to the integration along the entire ray if their density falls below a threshold, like 0.001. As shown in Tab. 4, replacing the conventional rendering strategy with No Empty Sampling has minimal effect on the performance of StegaNeRF, revealing our robustness to different rendering fashions." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.646, + 0.27, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.646, + 0.468, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.771 + ], + "angle": 0, + "content": "Figure 12. Ablation on the resolutions of embedded images by the average results over the selected scenes from the LLFF dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Limitations: The current implementation is limited by the computational efficiency of NeRF, since the steganographic training stage requires performing gradient descent updates through the NeRF model. Another limitation is that, the recovery accuracy of the injected hidden information varies across different 3D scenes. More systematic analysis is required to understand the underlying factors that make some scenes easier to hide information than others." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.62, + 0.106 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.343 + ], + "angle": 0, + "content": "With the progression of NeRF-based content, now serving as a pivotal element in diverse 3D applications, it is incumbent upon the research community to proactively tackle the imminent requirement to embed information within NeRF models as they are exchanged and disseminated across online platforms. This study introduces StegaNeRF, a pioneering framework for the automated embedding of steganographic information within NeRF renderings. We outline the essential components essential for achieving this functionality, unveiling potential challenges and revealing insights that can be instrumental for subsequent developers. By shedding light on the prospect of ownership identification within the realm of NeRF, this paper underscores the need for increased attention and concerted efforts toward addressing related challenges and avenues." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.355, + 0.597, + 0.371 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.38, + 0.895, + 0.45 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Jia-Bin Huang, Michael Zollhoefer, Johannes Kopf, and Changil Kim. Learning Neural Light Fields With Ray-Space Embedding Networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19787-19797, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.451, + 0.895, + 0.492 + ], + "angle": 0, + "content": "[2] Shumeet Baluja. Hiding Images in Plain Sight: Deep Steganography. Advances in Neural Information Processing Systems, 30, 2017. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.493, + 0.895, + 0.535 + ], + "angle": 0, + "content": "[3] Shumeet Baluja. Hiding Images Within Images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1685-1697, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.536, + 0.895, + 0.619 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-Nerf: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.62, + 0.895, + 0.689 + ], + "angle": 0, + "content": "[5] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5460–5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.69, + 0.895, + 0.731 + ], + "angle": 0, + "content": "[6] Oliver Benedens. Geometry-Based Watermarking of 3D Models. IEEE Computer Graphics and Applications, 19:46-55, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.733, + 0.895, + 0.788 + ], + "angle": 0, + "content": "[7] Ning Bi, Qiyu Sun, Daren Huang, Zhihua Yang, and Jiwu Huang. Robust Image Watermarking Based on Multiband Wavelets and Empirical Mode Decomposition. IEEE Transactions on Image Processing, 16(8):1956-1966, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.789, + 0.895, + 0.845 + ], + "angle": 0, + "content": "[8] Rajarathnam Chandramouli, Mehdi Kharrazi, and Nasir Memon. Image Steganography and Steganalysis: Concepts and Practice. In International Workshop on Digital Watermarking, pages 35-49. Springer, Springer, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.846, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[9] Chin-Chen Chang, Ju-Yuan Hsiao, and Chi-Shiang Chan. Finding Optimal Least-Significant-Bit Substitution in Image Hiding by Dynamic Programming Strategy. Pattern Recognition, 36(7):1583-1595, 2003. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.38, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "449" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast Generalizable Radiance Field Reconstruction From Multi-View Stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124–14133, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[11] Di Chen, Yu Liu, Lianghua Huang, Bin Wang, and Pan Pan. GeoAug: Data Augmentation for Few-Shot NeRF With Geometry Constraints. In European Conference on Computer Vision, pages 322-337. Springer, Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[12] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital Watermarking and Steganography. Morgan Kaufmann, 2007. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[13] Wenxue Cui, Shaohui Liu, Feng Jiang, Yongliang Liu, and Debin Zhao. Multi-Stage Residual Hiding for Image-Into-Audio Steganography. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2832-2836, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[14] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. In European Conference on Computer Vision, pages 636-654. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[15] Brandon Yushan Feng and Amitabh Varshney. SIGNET: Efficient Neural Representation for Light Fields. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14204-14213, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance Fields Without Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[17] Jessica Fridrich, Miroslav Goljan, and Rui Du. Detecting LSB Steganography in Color, and Gray-Scale Images. IEEE Multimedia, 8(4):22-28, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic Neural Radiance Fields for Monocular 4D Facial Avatar Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[19] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[20] Margarita Geleta, Cristina Punti, Kevin McGuinness, Jordi Pons, Cristian Canton, and Xavier Giro-i Nieto. PixInWav: Residual Steganography for Hiding Pixels in Audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2485-2489, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[21] Jamie Hayes and George Danezis. Generating Steganographic Images Via Adversarial Training. Advances in Neural Information Processing Systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[22] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[23] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. EfficientNeRF Efficient Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12902-12911, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[24] Xudong Huang, Wei Li, Jie Hu, Hanting Chen, and Yunhe Wang. Refsr-nerf: Towards high fidelity and super resolution view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8244–8253, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[25] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. HiNet: Deep Image Hiding by Invertible Network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4733-4742, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[26] Gary C Kessler and Chet Hosmer. An Overview of Steganography. Advances in Computers, 83:51-107, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification With Deep Convolutional Neural Networks. Communications of the ACM, 60(6):84-90, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[28] Jaeho Lee, Jihoon Tack, Namhoon Lee, and Jinwoo Shin. Meta-Learning Sparse Implicit Neural Representations. Advances in Neural Information Processing Systems, 34:11769-11780, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.463 + ], + "angle": 0, + "content": "[29] Chenxin Li, Mingbao Lin, Zhiyuan Ding, Nie Lin, Yihong Zhuang, Yue Huang, Xinghao Ding, and Liujuan Cao. Knowledge condensation distillation. In European Conference on Computer Vision, pages 19-35. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.465, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[30] Chenxin Li, Wenao Ma, Liyan Sun, Xinghao Ding, Yue Huang, Guisheng Wang, and Yizhou Yu. Hierarchical deep network with uncertainty-aware semi-supervised learning for vessel segmentation. Neural Computing and Applications, pages 1-14, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[31] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 61-65. IEEE, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[32] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[33] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Y. Wang, and Shenlong Wang. NeurMiPs: Neural Mixture of Planar Experts for View Synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15681-15691, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[34] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural Sparse Voxel Fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[35] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason M. Saragih. Mixture of Volumetric Primitives for Efficient Neural Rendering. ACM Transactions on Graphics (TOG), 40:1 - 13, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "450" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[36] Shao-Ping Lu, Rong Wang, Tao Zhong, and Paul L Rosin. Large-Capacity Image Steganography Based on Invertible Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10816-10825, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.163, + 0.37, + 0.176 + ], + "angle": 0, + "content": "[37] Luma AI. https://lumalabs.ai/.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.178, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[38] Xiyang Luo, Yinxiao Li, Huiwen Chang, Ce Liu, Peyman Milanfar, and Feng Yang. Dvmark: A Deep Multiscale Framework for Video Watermarking. ArXiv Preprint ArXiv:2104.12734, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.233, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[39] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210–7219, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.317, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy Networks: Learning 3d Reconstruction in Function Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.386, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[41] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local Light Field Fusion: Practical View Synthesis With Prescriptive Sampling Guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.457, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing Scenes As Neural Radiance Fields for View Synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.525, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives With a Multiresolution Hash Encoding. ACM Transactions on Graphics (TOG), 41:1-15, 2022. 1, 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.581, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[44] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing Neural Radiance Fields for View Synthesis From Sparse Inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.665, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning Continuous Signed Distance Functions for Shape Representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.748, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[46] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[47] Tomás Pevny, Tomás Filler, and Patrick Bas. Using High-Dimensional Image Models to Perform Highly Undetectable Steganography. In International Workshop on Information Hiding, pages 161-177. Springer, Springer, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[48] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. In 2021 Interna" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.891, + 0.119 + ], + "angle": 0, + "content": "tional Conference on 3D Vision (3DV), pages 1106-1114. IEEE, 2021.9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.764, + 0.135 + ], + "angle": 0, + "content": "[49] Polycam. https://poly.cam/.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[50] Emil Praun, Hugues Hoppe, and Adam Finkelstein. Robust Mesh Watermarking. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, pages 49-56, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[51] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[52] Yinlong Qian, Jing Dong, Wei Wang, and Tieniu Tan. Deep Learning for Steganalysis Via Convolutional Neural Networks. In Media Watermarking, Security, and Forensics 2015, volume 9409, pages 171-180. SPIE, SPIE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[53] Chen Quei-An. Nerf-pl: A Pytorch-Lighting Implementation of NeRF. https://github.com/kwea123/nerf_pl/.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[54] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding Up Neural Radiance Fields With Thousands of Tiny Mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[55] Mennatallah M Sadek, Amal S Khalifa, and Mostafa GM Mostafa. Video Steganography: A Comprehensive Review. Multimedia Tools and Applications, 74(17):7063-7094, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[56] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-Aware Image Synthesis With Sparse Voxel Grids. ArXiv Preprint ArXiv:2206.07695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[57] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[58] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit Neural Representations With Periodic Activation Functions. Advances in Neural Information Processing Systems, 33, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[59] Vincent Sitzmann, Semon Rezchikov, William T. Freeman, Joshua B. Tenenbaum, and Frédo Durand. Light Field Networks: Neural Scene Representations With Single-Evaluation Rendering. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[60] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[61] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct Voxel Grid Optimization: Super-Fast Convergence for Radiance Fields Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[62] Abdelfatah A Tamimi, Ayman M Abdalla, and Omaima Al-Allaf. Hiding an Image Inside Another Image Using Variable-Rate Steganography. International Journal of Advanced Computer Science and Applications (IJACSA), 4(10), 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "451" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-Nerf: Scalable Large Scene Neural View Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248-8258, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[64] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible Hyperlinks in Physical Photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2126, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[65] Matthew Tancik, Pratul P Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T Barron, and Ren Ng. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. Advances in Neural Information Processing Systems, 33, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[66] Weixuan Tang, Shunquan Tan, Bin Li, and Jiwu Huang. Automatic Steganographic Distortion Learning Using a Generative Adversarial Network. IEEE Signal Processing Letters, 24(10):1547-1551, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[67] Jeroen JG Van Merrienboer, Liesbeth Kester, and Fred Paas. Teaching complex rather than simple tasks: Balancing intrinsic and germane load to enhance transfer of learning. Applied Cognitive Psychology: The Official Journal of the Society for Applied Research in Memory and Cognition, 20(3):343-352, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[68] Mukund Varma, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that nef needs? In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[69] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and S. Tulyakov. R2L: Distilling Neural Radiance Field to Neural Light Field for Efficient Novel View Synthesis. In European Conference on Computer Vision. Springer, Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.631 + ], + "angle": 0, + "content": "[70] Huaiqing Wang and Shuozhong Wang. Cyber Warfare: Steganography vs. Steganalysis. Communications of the ACM, 47(10):76-82, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[71] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier PlenOctrees for Dynamic Radiance Field Rendering in Real-Time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[72] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning Neural Implicit Surfaces by Volume Rendering for Multi-View Reconstruction. Advances in Neural Information Processing Systems, 34, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[73] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning Multi-View Image-Based Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690–4699, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[74] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image Quality Assessment: From Error Visibility" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "to Structural Similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[75] Xinyu Weng, Yongzhi Li, Lu Chi, and Yadong Mu. High-Capacity Convolutional Video Steganography With Temporal Residual Modeling. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, pages 87-95, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[76] Raymond B Wolfgang and Edward J Delp. A Watermark for Digital Images. In Proceedings of 3rd IEEE International Conference on Image Processing, volume 3, pages 219-222. IEEE, IEEE, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[77] Yue Wu, Guotao Meng, and Qifeng Chen. Embedding Novel Views in a Single JPEG Image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14519-14527, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.303, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[78] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[79] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. BungeeNeRF: Progressive Neural Radiance Field for Extreme Multi-Scale Scene Rendering. In The European Conference on Computer Vision (ECCV), volume 2, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[80] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[81] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4479–4489, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.637 + ], + "angle": 0, + "content": "[82] Hyukryul Yang, Hao Ouyang, Vladlen Koltun, and Qifeng Chen. Hiding Video in Audio Via Reversible Generative Models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1100-1109, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[83] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5491–5500, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[84] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for Real-Time Rendering of Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.832 + ], + "angle": 0, + "content": "[85] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. Pixelnerf: Neural Radiance Fields From One or Few Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[86] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14448-14457, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "452" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[87] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A Structured Dictionary Perspective on Implicit Neural Representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.247 + ], + "angle": 0, + "content": "[88] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray Priors Through Reprojection: Improving Neural Radiance Fields for Novel View Extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18376-18386, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[89] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and Improving Neural Radiance Fields. ArXiv Preprint ArXiv:2010.07492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[90] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features As a Perceptual Metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[91] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul E. Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural Factorization of Shape and Reflectance Under an Unknown Illumination. ACM Trans. Graph, 40:237:1-237:18, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "453" + } + ] +] \ No newline at end of file diff --git a/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/3c8d0cf6-dc7d-44ba-a804-37f576c9cfcc_origin.pdf b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/3c8d0cf6-dc7d-44ba-a804-37f576c9cfcc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bfede0acf87d50bbd86feadf925c849d33804c97 --- /dev/null +++ b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/3c8d0cf6-dc7d-44ba-a804-37f576c9cfcc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dac49880bbf4547ad6cf594740f173a8c57010c1cdee884921a5e11884a9a24b +size 741203 diff --git a/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/full.md b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e8259e3b2e394f43ae65d89488457fc7e8503f5d --- /dev/null +++ b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/full.md @@ -0,0 +1,436 @@ +# StegaNeRF: Embedding Invisible Information within Neural Radiance Fields + +Chenxin Li $^{1*}$ , Brandon Y. Feng $^{2*}$ , Zhiwen Fan $^{3*}$ , Panwang Pan $^{4}$ , Zhangyang Wang $^{3}$ $^{1}$ Chinese University of Hong Kong, $^{2}$ University of Maryland, $^{3}$ University of Texas at Austin, $^{4}$ ByteDance + +# Abstract + +Recent advancements in neural rendering have paved the way for a future marked by the widespread distribution of visual data through the sharing of Neural Radiance Field (NeRF) model weights. However, while established techniques exist for embedding ownership or copyright information within conventional visual data such as images and videos, the challenges posed by the emerging NeRF format have remained unaddressed. In this paper, we introduce StegaNeRF, an innovative approach for steganographic information embedding within NeRF renderings. We have meticulously developed an optimization framework that enables precise retrieval of hidden information from images generated by NeRF, while ensuring the original visual quality of the rendered images to remain intact. Through rigorous experimentation, we assess the efficacy of our methodology across various potential deployment scenarios. Furthermore, we delve into the insights gleaned from our analysis. StegaNeRF represents an initial foray into the intriguing realm of infusing NeRF renderings with customizable, imperceptible, and recoverable information, all while minimizing any discernible impact on the rendered images. For more details, please visit our project page: https://xggnet.github.io/StegaNeRF/ + +# 1. Introduction + +Implicit neural representation (INR) is an emerging concept where the network describes the data through its weights [14, 24, 40, 42, 45, 58, 65]. After training, the INR weights can then be used for content distribution, streaming, and even downstream inference tasks, all without sending or storing the original data. Arguably the most prominent INR is Neural Radiance Fields (NeRF) [42], where a network learns a continuous function mapping spatial coordinates to density and color. Due to its lightweight size and superb quality, NeRF has immense potential for 3D content representation in future vision and graphics applications. + +![](images/fc98755da4cfee183d3e717dde8e24380c2629c36ff498bf79fe2a72540a46a9.jpg) +Figure 1. We introduce the new problem of NeRF steganography: hiding information in NeRF renderings. Our proposed framework, StegaNeRF, can embed and recover customized hidden information while preserving the original NeRF rendering quality. + +While there is a plethora of work dedicated towards better quality [4,63,72,91], faster rendering [43,56,61,71,83], and sparse view reconstruction [10, 11, 44, 57, 80, 81, 85, 88], in this paper, we look beyond the horizon and explore a new question: Can we achieve steganography with NeRF? + +Established digital steganography method [12] focus on embedding hidden messages in 2D images. The recent growth of deep learning and social media platforms further gives rise to many practical use cases of image steganography. As countless images and videos are shared online and even used to train deep learning models, 2D steganography methods [2,3] allow users and data providers to protect copyright, embed ownership, and prevent content abuse. + +Now, with the ongoing advances in 3D representations powered by NeRF, we envision a future where people frequently share their captured 3D content online just as they are currently sharing 2D images and videos online. Moreover, we are curious to explore the following research questions: 1) Injecting information into 2D images for copyright or ownership identification is common, but can we preserve such information when people share and render 3D scenes through NeRF? 2) NeRF can represent large-scale real-world scenes with training images taken by different people, but can we preserve these multiple source identities in the NeRF renderings to reflect the collaborative efforts required to reconstruct these 3D scenes? 3) Common image steganography methods embed either a hidden image or a message string into a given image, but can we allow different modalities of the hidden signal in NeRF steganography? + +Driven by these questions, we formulate a framework to embed customizable, imperceptible, and recoverable information in NeRF renderings without sacrificing the visual quality. Fig. 1 presents an overview of our proposed framework, dubbed StegaNeRF. Unlike the traditional image steganography that embeds hidden signals only into a specific source image, we wish to recover the same intended hidden signal from NeRF rendered at arbitrary viewpoints. + +Despite many established works on 2D steganography and hidden watermarks for image and video [2, 3, 9, 25, 36, 64], naively applying 2D steganography on the NeRF training images is not practical, since the embedded information easily gets lost in the actual NeRF renderings. In contrast, our framework enables reliable extraction of the hidden signal from NeRF renderings. During NeRF training, we jointly optimize a decoder network to extract hidden information from the NeRF renderings. To minimize the negative impact on the visual quality of NeRF, we identify weights with low impact on rendering and introduce a gradient masking strategy to steer the hidden steganographic information towards those low-importance weights. Extensive experimental results validate StegaNeRF balances between the rendering quality of novel views and the high-fidelity transmission of the concealed information. + +StegaNeRF presents the first exploration of hiding information in NeRF models for ownership identification, and our contributions can be summarized as follows: + +- We introduce the new problem of NeRF steganography and present the first effort to embed customizable, imperceptible, and recoverable information in NeRF. +- We propose an adaptive gradient masking strategy to steer the injected hidden information towards the less important NeRF weights, balancing the objectives of steganography and rendering quality. +- We empirically validate the proposed framework on a diverse set of 3D scenes with different camera layouts and scene characteristics, obtaining high recovery accuracy without sacrificing rendering quality. +- We explore various scenarios applying StegaNeRF for ownership identification, with the additional support of multiple identities and multi-modal signals. + +# 2. Related Work + +Neural Radiance Fields Following the success of NeRF [42], numerous techniques have been incorporated including ray re-parameterizations [4, 5], explicit spatial data structures [16, 23, 34, 43, 61, 84], caching and distillation [19, 22, 54, 69], ray-based representations [1, 15, 59], geometric primitives [33, 35], large-scale scenes [39, 63, 79, 89], generalizable architectures [60, 68, 73], and dynamic + +settings [18, 32, 46, 51, 78]. Unlike prior works, this paper explores the uncharted problem of embedding information in NeRF renderings, with critical implications for copyright protection and ownership preservation. As early-stage NeRF-based products already become available [37,49] and more are expected, we believe now is the right time to open up the exploration of NeRF steganography. + +Image Steganography Steganography hides intended signals as invisible watermarks (e.g., hyperlinks, images) within the cover media called carriers (e.g., images, video and audio) [12, 26]. Classical methods focus on seeking and altering the least significant bits (LSBs) [17, 47, 62, 76] and transforming domain techniques [7, 8, 52, 70]. Prior research also uses deep neural networks for steganography [2, 3, 21, 66]. Among them, DeepStega [2] conceals a hidden image within an equal-sized carrier image. Subsequent works [25, 36] use invertible networks to improve the performance of deep image steganography. Another line of work conceals information in other carrier media like audio [13, 20, 82] and video [38, 55, 75]. The above advances all play a critical part in the era when traditional media formats like images and videos are dominant. However, as MLP-based neural representations of 3D scenes are gaining momentum to become a major format of visual data, extending steganography to NeRF is bound to become an important problem in the upcoming future. + +Lifting Steganography to 3D Prior to NeRF, meshes are commonly used to represent 3D shapes. Early pioneers apply steganography 3D meshes [6, 50] for copyright protection when meshes are exchanged and edited. More recent work [77] has also explored embedding multi-plane images within a JPEG-compressed image. Their problem can be regarded as a special case of 2D steganography, hiding multiple 2D images inside a single 2D image. In contrast, we try to hide a natural image into a 3D scene representation (NeRF), fundamentally differing from these prior methods where 2D images act as the carrier of hidden information. + +# 3. Method + +The goal of StegaNeRF is to inject customized (steganographic) information into the NeRF weights with imperceptible visual changes when rendering. Given a NeRF with weights $\theta_0$ and the information $I$ to be injected, when we render with $\theta$ on any viewpoint, we hope the injected $I$ can be recovered by a decoder $F_{\psi}$ with learnable weights $\psi$ . + +A seemingly obvious solution is to use prior image steganography methods to 1) inject $I$ on training images, 2) train NeRF on those images with embedded information, 3) apply their provided $F_{\psi}$ to extract $I$ from NeRF-rendered images. This approach has been successfully applied for GAN fingerprinting [86]. However, it fails in the NeRF context (see Fig. 3), where the subtle changes induced by prior + +![](images/b29770537380d4b638d8474caadb3fcc7f31d0cbff08beb0c0e4ccfa34f9979e.jpg) +Figure 2. StegaNeRF training overview. At the first stage, we optimize $\theta_0$ with standard NeRF training. At the second stage, we initialize $\theta$ with $\theta_0$ and optimize for the steganography objectives. We train the decoder $F_{\psi}$ to recover hidden information from StegaNeRF renderings and no hidden information from original NeRF renderings. We introduce Classifier Guided Recovery to improve the accuracy of recovered information, and Adaptive Gradient Masking to balance between steganography ability and rendering visual quality. + +![](images/ad4fe3df72f82e271759a77cc7e3347be07d4bfc2df6cafe49d17756ee69e18c.jpg) + +steganography methods easily get smoothed out, inhibiting the decoder from identifying the subtle patterns necessary for information recovery. + +Such shortcomings of off-the-shelf steganographic methods are not surprising since they are developed for the traditional setting where 2D images are the ultimate form of visual information. In contrast, our problem setting involves the new concept of INR as the underlying representation, and 2D images are just the final output rendered by NeRF. + +# 3.1. Two-Stage Optimization + +Recognizing the formulation difference due to the emergence of implicit representation with NeRF, we move away from the traditional 2D image-based pipeline that trains an encoder network to inject subtle changes to the given 2D images. Instead, we incorporate the steganography objective into the gradient-based learning process of NeRF. + +We re-design the training as a two-stage optimization. The first stage is the original NeRF training procedure, involving the standard photometric loss between the rendered and ground truth pixels to guarantee the visual quality. After finishing the weights update $\theta_0$ at the first stage, we dedicate the second stage to obtain the final NeRF weights $\theta$ containing steganographic information. We introduce several techniques to achieve robust information recovery with imperceptible impact on the rendered images. The training workflow is depicted in Fig. 2 and Alg. 1. + +# 3.2. Information Recovery + +Let $P$ denote the camera pose at which we render an image from a NeRF network. We want to recover $I$ from the rendered image $\theta(P)$ . Importantly, we also want to avoid false positives on images $\theta_0(P)$ rendered by the original network without steganography ability, even if the images rendered by $\theta_0$ and $\theta$ look visually identical. Therefore, we + +![](images/f3e1e4ae25f1e21df1c01ce78e682cc666d0c0f30d5607ea1ba813a143312571.jpg) +Figure 3. Results of applying prior 2D steganography methods, LSB [9] (top) and DeepStega [2] (bottom). From left to right, we show (a) training image after applying 2D steganography, (b) residual error of (a) over ground truth, (c) hidden image recovered from (a), (d) residual error of NeRF rendering at the pose of (a), and (e) hidden image recovered from (d). Prior 2D steganography methods fail in the NeRF context since the hidden information injected in training images mostly disappear in NeRF renderings. + +![](images/cbe4fda3609a084f42a8a14db3790f72757d45490303723720d58d477593ac29.jpg) + +![](images/9f560c7b3ab4013bb5972d53e73e75ad83c244fd7deacb4a8179578a4eee9aa3.jpg) + +![](images/714106e3959ffa6ca22a7f1d49393f175a822b87c47e862453cedeb389222dbe.jpg) + +![](images/7788ee5077cd1a502af07201b4d246dc962101c0ce7ce1cb52d330e9b22434c1.jpg) + +optimize $\theta$ to minimize the following contrastive loss terms: + +$$ +\mathcal {L} _ {d e c} ^ {+} = \left| F _ {\psi} (\boldsymbol {\theta} (P)) - I \right|, \quad \mathcal {L} _ {d e c} ^ {-} = \left| F _ {\psi} \left(\boldsymbol {\theta} _ {0} (P)\right) - \varnothing \right|, \tag {1} +$$ + +where $\varnothing$ is an empty image with all zeros. Effectively, $L_{dec}^{+}$ serves as a regularization term, facilitating the decoder to recover the embedded image pattern based on renderings obtained from a StegaNeRF. In contrast, $L_{dec}^{-}$ prevents the decoder from falsely producing any plausible image pattern when given renderings from a Standard NeRF that is without hidden signals. + +Classifier-Guided Recovery The decoder $F_{\psi}$ is easily implemented as a U-Net to decode $I$ as the form of 2D images, but accurately recovering all the details in $I$ might be challenging. Therefore, we additionally train a classifier + +# Algorithm 1 Train StegaNeRF on a single scene + +Data: Training images $\{Y_i\}$ with poses $\{P_i\}$ , hidden information $I$ , learning rate $\eta = [\eta_0, \eta_1]$ + +Output: Steganographic NeRF $\theta$ and decoder $F_{\psi}$ + +Initialize NeRF $\theta_0$ and decoder network $F_{\psi}$ + +Optimize $\theta_0$ on $\{Y_i, P_{ij}\}$ with standard NeRF training + +Compute mask $m$ for $\theta_0$ as in Eq. (3) + +for each training iteration $t$ do + +Randomly sample a training pose $P_{i}$ and render $\theta (P_i)$ + +Compute stega. losses $\mathcal{L}_{dec}^{c}$ , $\mathcal{L}_{dec}^{+}$ , $\mathcal{L}_{dec}^{-}$ as Eq. (1), (2) + +Compute $\mathcal{L}_{rgb}$ and combine total loss $\mathcal{L}$ as in Eq. (4) + +Update $\pmb{\theta}$ with $\eta_0\cdot (\frac{\partial\mathcal{L}}{\partial\pmb{\theta}}\odot \pmb {m})$ and $F_{\psi}$ with $\eta_{1}\cdot \frac{\partial\mathcal{L}}{\partial\psi}$ + +# end for + +network $F_{\psi}^{c}$ to solve the easier task of classifying whether the given NeRF rendering contains hidden information. $F_{\psi}^{c}$ is optimized by the following cross-entropy loss [30, 67]: + +$$ +\mathcal {L} _ {d e c} ^ {c} = - \log \left(F _ {\psi} ^ {c} (\boldsymbol {\theta} (P))\right) - \log \left(1 - F _ {\psi} ^ {c} \left(\boldsymbol {\theta} _ {0} (P)\right)\right). \tag {2} +$$ + +We then use its prediction to guide the process of decoding pixel-wise information by adding the classification output as input to $F_{\psi}$ , such that $F_{\psi}(x) = F_{\psi}\big(x,F_{\psi}^{c}(x)\big)$ . + +Although the above discussion focuses on hiding images, our framework can be easily extended to embed other modalities like strings, text, or even audio, all of which may be represented as 1D vectors. We can simply modify the architecture of $F_{\psi}$ to have a 1D prediction branch. + +# 3.3. Preserving Perceptual Identity + +Since we want to hide information without affecting the visual perception of the rendered output, an intuitive regularization is to penalize how much $\theta$ deviates from $\theta_0$ . However, we find that naively summing the deviation penalty across all weights makes it difficult for the NeRF network to adjust its weights for the steganographic objective. Instead, motivated by the fact that INR weights are not equally important and exhibit strong sparsity [28, 87], we propose an adaptive gradient masking strategy to encode the hidden information on specific weights' groups. + +Formally, given the initial set of weights $\theta_0\in \mathbb{R}^N$ , we compute a soft importance mask $m\in \mathbb{R}^{N}$ . Specifically, for each weight $\pmb {w}_i\in \pmb {\theta}_0$ , we compute: + +$$ +\boldsymbol {m} _ {i} = \frac {\left| \boldsymbol {w} _ {i} \right| ^ {- \alpha}}{\sum_ {i} ^ {N} \left| \boldsymbol {w} _ {i} \right| ^ {- \alpha}}, \tag {3} +$$ + +where the magnitude $|\cdot|$ of each model weight $w_{i}$ serves as its individual importance indicator, and $\alpha > 0$ is a scaling factor of power controlling the relative distribution of importance across the weights. We mask the gradient as $\frac{\partial\mathcal{L}}{\partial\theta} \odot m$ when optimizing $\theta$ based on the total loss $\mathcal{L}$ in the second stage, where $\odot$ is a Hadamard product. Effectively, + +# Algorithm 2 Typical usage scenario of StegaNeRF + +1: Alice captures some images of a 3D scene +2: Alice trains a StegaNeRF to hide a personalized image +3: Alice shares the model $\theta$ online for other people to enjoy and explore the 3D scene themselves +4: Bob grabs the model $\theta$ and reposts it with his own account without crediting Alice or asking for permission +5: Alice sees Bob's post, deploys the decoder $F_{\psi}$ , and verifies the owner of $\theta$ is Alice, not Bob +6: Bob takes down the post or gets banned for copyright infringement + +more significant weights are "masked out" to minimize the impact of steganographic learning on the rendered visual quality. + +We retain the standard photometric error in steganography learning to maintain NeRF rendering fidelity: $\mathcal{L}_{rgb} = |\pmb {\theta}(P) - \pmb {\theta}_0(P)|$ . The overall training loss at the second stage can be formulated as follows: + +$$ +\mathcal {L} = \lambda_ {0} \mathcal {L} _ {d e c} ^ {c} + \lambda_ {1} \mathcal {L} _ {d e c} ^ {+} + \lambda_ {2} \mathcal {L} _ {d e c} ^ {-} + \lambda_ {3} \mathcal {L} _ {r g b}. \tag {4} +$$ + +# 4. Experiments + +# 4.1. Implementation Details. + +Dataset. We use common datasets LLFF [41] and NeRF-Synthetic [42], with forward scenes {flower, fern, fortress, room} from LLFF and $360^{\circ}$ scenes {lego, drums, chair} from NeRF-Synthetic. We further experiment on the Brandenburg Gate scene from NeRF-W dataset [39], with over 800 views of in-the-wild collected online. + +Training. On LLFF and NeRF-Synthetic, we adopt Plenoxels [83] as the NeRF backbone architecture for efficiency. On NeRF-W, we use the available PyTorch implementation [53]. The first stage of training is performed according to the standard recipes of those implementations. We then perform the second stage of steganography training for 55 epochs unless otherwise noted. On LLFF and NeRF-W, we downsize the training images by 4 times following common practice, and we use the original size on NeRF-Synthetic. + +For hyper-parameters in Eq. (4), we set the weight of NeRF reconstruction error $\lambda_3 = 1$ for all experiments. We set $\lambda_0 = 0.01, \lambda_1 = 0.5, \lambda_2 = 0.5$ for all scenes on LLFF dataset, and $\lambda_0 = 0.1, \lambda_1 = 1, \lambda_2 = 1$ for the scenes in NeRF-Synthetic, and $\lambda_0 = 0.05, \lambda_1 = 1, \lambda_2 = 1$ for NeRF-W. For Eq. (3), we set $\alpha = 3$ for all scenes. We run experiments on one NVIDIA A100 GPU. + +Evaluation. We assume a typical authentication scenario shown in Alg.2. We consider the recovery quality of the embedded information, including the metrics of classification accuracy (Acc.) of the classifier, and the structural sim + +![](images/39a3a01589af864221c2ab7e726b6772ec2530d5c9bd57a39b03ddab82e30774.jpg) +Figure 4. Results on the NeRF-Synthetic dataset. Within each column, we show the StegaNeRF rendering, residual error from the initial NeRF rendering, and the recovered hidden image. We show the SSIM for the StegaNeRF renderings and the recovered hidden images. + +Table 1. Quantitative results of StegaNeRF rendering and hidden information recovery. Standard NeRF is the initial NeRF $\theta_0$ with standard training (upper-bound rendering performance). Prior 2D steganography fails after NeRF training while StegaNeRF successfully embeds and recovers hidden information with minimal impact on the rendering quality. Results are averaged over the selected LLFF and NeRF-Synthetic scenes. + +
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
Standard NeRF28.150.85750.137750.00N/A
LSB [9]28.130.85690.1452N/A0.0132
DeepStega [2]26.940.84320.1641N/A0.2098
StegaNeRF (Ours)28.130.85630.1460100.00.9730
Standard NeRF31.130.96060.031050.00N/A
LSB [9]31.120.96040.0310N/A0.0830
DeepStega [2]31.130.96060.0313N/A0.2440
StegaNeRF (Ours)30.960.95830.029099.720.9677
+ +ality (SSIM) [74] of the hidden image recovered by the decoder. We evaluate the final NeRF renderings with PSNR, SSIM and LPIPS [90]. All the metrics are computed on the test set and averaged over all the scenes and embedded images. + +# 4.2. Case I: Embedding in a Single Scene. + +We first explore the preliminary case of ownership identification on a specific NeRF scene. We select random images from ImageNet [27] as the hidden information to be injected to NeRF renderings. + +Failure of 2D Baseline Due to the lack of prior study on NeRF steganography, we consider a baseline from 2D image steganography by training NeRF from scratch with + +Table 2. Quantitative results of NeRF rendering and hidden information recovery. We consider two conditions, embedding each scene with a common hidden image (One-for-All) or a scene-specific hidden image (One-for-All). We report the quality difference compared to the single-scene settings as $\Delta_{SSIM}^{(10^{-2})}$ . Results are averaged over the selected LLFF scenes. + +
SettingNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS ↓Δ(10-2)SSIM↑Acc. (%)↑SSIM↑Δ(10-2)SSIM↑
One-for-All24.990.80130.1786-0.10100.000.9860+0.19
One-for-One24.990.80160.1779-0.07100.000.9122-7.19
One-for-All27.900.85130.1236+0.01100.000.9844-0.76
One-for-One27.900.85150.1195+0.03100.000.9448-4.45
One-for-All30.270.84980.1289+0.02100.000.9430-5.42
One-for-One30.120.84800.1302-0.16100.000.9102-8.67
+ +the watermarked images. We implement two off-the-shelf steganography methods including a traditional machine learning approach called Least Significant Bit (LSB [9]), and a deep learning pipeline as DeepStega [2]. An ideal case is that the embedded information can be recovered from the synthesized novel views, indicating the successful NeRF steganography. However, as can be seen in Fig. 3, the embedded information containing the hidden images cannot be recovered from the NeRF renderings. By analyzing the residual maps of training views (between GT training views and the watermarked) and novel views (between GT novel views and the actual rendering), we observe the subtle residuals to recover the hidden information are smoothed out in NeRF renderings, and similar failures occur on other datasets as well. Therefore, for the rest of our study, we mainly focus on analyzing our new framework that performs direct steganography on 3D NeRF. + +![](images/d6c84e49f3bc15605767706ce87f049e7e32deb570f4eed16671ed82011e52b3.jpg) +(a) Per-Scene + +![](images/befd6ae597219dc81b1e27ac75f189b933195b1162d5bf30874010a2636db60c.jpg) +Figure 5. Results on three multi-scene settings. Within each column, we show the StegaNeRF rendering, residual error from the initial NeRF rendering, and recovered hidden image. We show the SSIM for the StegaNeRF renderings and the recovered hidden images. + +![](images/724703a416a23b70eef375d130184438f3bd0ba5aa8f779ac91b975f1e840cf9.jpg) + +![](images/69a753bd6b77b0add7d695a2f3857e3ff481eb79efde28c14fad54228efcafd4.jpg) +(b) Multi-Scene: One-for-All + +![](images/08c10b53ec3c44fc6e5f96b9244d77e6432ce2929d6faa68b72d4b46ba75ba34.jpg) + +![](images/175cfe58e9400866ce5654d7f117139b85151c87270308225607d98fb9a223d8.jpg) + +![](images/f331cae1e498cc8136eaf4476c507bae58d36a3c26bac8c217349b7d91e064ce.jpg) +(c) Multi-Scene: One-for-One + +![](images/c3f6877407ef268b8eca057516f723311e7953bb0ff9658bae81d22c0d4dc5ef.jpg) + +![](images/2aa717d3e47396a8532de09b61ec7888b48cbd0022effaa387079d033e90d281.jpg) + +Results Tab. 1 contains quantitative results on LLFF and NeRF-Synthetic scenes. While NeRF trained by 2D steganography methods hardly recovers the embedded information, StegaNeRF accurately recovers the hidden image with minimal impact on the rendering quality measured by PSNR. Fig. 4 provides qualitative results of StegaNeRF on three embedded images on the NeRF-Synthetic scenes. An interesting observation is the common regions where hidden information (high residual error) emerges in the renderings, e.g., the right rear side of lego and left handrail of chair. We also notice that, within each scene, these regions are persistent across multiple viewpoints. + +# 4.3. Case II: Embedding in Multiple Scenes. + +Settings We extend our steganography scheme to embed information within multiple scenes at once. Specially, we use three LLFF scenes $\{flower, fern, fortress\}$ to test the two sub-settings of multi-scene embedding with: (1) One-for-All, a common hidden image and (2) One-for-One scene-specific hidden images. All scenes share the same decoder $F_{\psi}$ and classifier $F_{\psi}^{c}$ . The difference between the two sub-settings is the number of hidden images that $F_{\psi}$ and $F_{\psi}^{c}$ need to identify and recover. We sample one scene for every training epoch, and due to the increased data amount, we increase training epochs until convergence. + +Results Tab. 2 provides quantitative results on the multi-scene setting. The performance drop compared to single-scene training is sometimes noticeable, but it is not surprising due to the inherent requirement of per-scene fitting for our NeRF framework. Fig. 5 presents visual comparisons of multi-scene steganography against single-scene scheme. + +# 4.4. Case III: Embedding Multiple Identities. + +Settings Constructing NeRF of large-scale cultural landmarks is a promising application, and community contributions are crucial to form the training images. Since every NeRF prediction is indebted to some particular training images, it would be meaningful to somehow identify the contributing users in the rendered output. Specifically, we present a proof-of-concept based on the following scenario: Given a collection of user-uploaded training images, our task is to apply StegaNeRF to ensure the final NeRF renderings hide subtle information about the relevant identities whose uploaded images help generate the current rendering. + +To simulate this multi-user scenario in the public NeRF-W dataset, we randomly select $M$ anchor views with different viewing positions, and then find the $K$ nearest neighbour views to each anchor to form their respective clusters. We set $M = 3$ , $K = 20$ in experiments. We assume a common contributor identity for each cluster, and we want the NeRF rendering to contain customized hidden information about those $M$ contributors when we render within the spatial regions spanned by their own cluster. Thus, our classifier network $F_{\phi}^{c}$ is modified to output $M$ -class cluster predictions and another class for identity outside of the $M$ clusters. Since the decoder should extract no information for views outside of those $M$ clusters to prevent false positives, we also compute $\mathcal{L}_{dec}^{-}$ (1) and $\mathcal{L}_{dec}^{c}$ (2) for those poses. + +Results We employ the same network backbone as NeRF-W [66] to handle in-the-wild images with different time and lighting effects. Fig. 6 presents qualitative results of embedding multiple identities in a collaborative large-scale NeRF. + +![](images/4483d0a9b5b2ae4bbaa21689583e9c4057846bf953d6cb41c605d43855bb3761.jpg) +Figure 6. Qualitative results of NeRF steganography on NeRF-W dataset. Within each block we show the StegaNeRF rendering, the initial NeRF rendering, the residual error between StegaNeRF and initial, and the recovered hidden image. At the bottom right, we provide the SSIM computed with their respective ground truth. "Unmarked Identity" denotes the remaining training views not related to any of the considered users, and the decoder does not recover meaningful information from renderings at those poses. + +![](images/6115d281af9e8a99be15960170776e7ca09f2a999dcbcd896e05df93fa141114.jpg) +SSIM=0.9709 + +![](images/7d5f6e9d56b3afd50151fcc36ad4dcec880d1a82313045d06d4edd2f8e2f9035.jpg) +$\mathrm{SSIM} = 0.9688$ +`The Owner Of These Horns' +Figure 7. Multi-modal hidden information. We show StegaNeRF rendering, recovered hidden images, audio (waveform and spectrum), and text. Metrics for hidden information in each modality are labeled respectively. For the StegaNeRF rendering, we report both SSIM and the relative change from initial NeRF renderings. + +![](images/eedc427949d4424f0b28fc773a892c6d11e3e4f4b53c383f4d44c06db78ba2ba.jpg) + +![](images/2541dbb6af9e408ec055d3b7aa36a413e49dd16066289336cdf21de842b81e68.jpg) +PSNR=76.22 + +![](images/02d720779d3fa8852ba90d4ba95f2f643b5a09b5111a41a74c02efec62665ad8.jpg) +$= 76.22$ + +![](images/a241df4d59a9e60a03f5aee43ecadc2716e334fb69ad0b10dfb1489752cbc058.jpg) +Acc. $= 100.0\%$ + +![](images/c2293670317ccd800ee78f7c0eac6594955cd76bd38ba75102b9ae3ced8a275c.jpg) +Hidden Image + +![](images/08db40a38819a1606a7343bc42a8e6965c6f8254d53a8e17f139afba53b17090.jpg) +Hidden Audio +PSNR=62.84 +Hidden Text + +![](images/448a38526e8ec86cf3093219f717c7da6ff576251a316449e63b928245c351b3.jpg) +Acc. $= 100.0\%$ + +62.84 + +# 4.5. Case IV: Embedding Multi-Modal Information. + +Settings We further show the potential of StegaNeRF in embedding multi-modal hidden information, such as images, audio, and text. We modify the decoder network to build a modal-specific decoder for each modality. + +Results Fig. 7 shows recovered multi-modal embedded signals in the trex and horns scenes from LLFF. Evidently, the StegaNeRF framework can easily extend to embed multi-modal information with high recovery performance without scarifying the rendering quality. + +# 4.6. Ablation Studies + +The effect of removing each component of StegaNeRF is presented in Tab. 3. Standard NeRF uses the initial NeRF + +Table 3. Ablation study of different components of StegaNeRF. Results are averaged on the selected LLFF scenes. + +
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
StegaNeRF28.130.85680.1430100.00.9486
No Classifier26.850.80770.2417N/A0.5629
No Classifier Guided27.120.82390.2073100.00.6716
No Gradient Masking27.860.83750.1710100.00.8822
No Soft Masking28.050.85580.152694.440.8751
Standard NeRF28.150.85750.137750.00N/A
+ +rendering. No Classifier removes the classifier $F_{\psi}^{c}$ , while No Classifier Guided retains the classification task (hence impact on NeRF rendering) but does not condition the decoder on classifier output. No Gradient Masking removes the proposed masking strategy (Sec. 3.3), and No Soft Masking uses the binary mask [29, 31] with a threshold of 0.5. Our StegaNeRF makes a good balance between the rendering quality and the decoding accuracy. Fig. 8 shows the visual impact of removing each component. It appears that when any component is removed, the performance drops accordingly, revealing the effectiveness of our design. + +Fig. 9 shows the impact of varying gradient masking strategies. We can see a different degree of affected by ratio variation on hidden recovery and rendering quality. Fig. 10 reports the performance of StegaNeRF against the common perturbations including JPEG compression and Gaussian noise. The lines show the average accuracies across selected scenes and the shaded regions indicate the range of 0.5 standard deviation. It appears that the hidden recovery of StegaNeRF is robust to various JPEG compression ratios and Gaussian blur degradation. + +![](images/77d0c1748fe17cc0f6e4a4972819f344c08065aed7646d5e7cb7b23d4f1b7448.jpg) +(a) Ground Truth +(b) No Classifier +(c) No Classifier Guided +(d) No Gradient Masking +(e) No Soft Masking +(f) StegaNeRF + +![](images/583281c660dc64ea12b2dbccb5f16738971b1b9143cdac21cac82a8d98033ca4.jpg) +(a) Varying $\alpha$ +Figure 9. Analysis of gradient masking. (a) Varying $\alpha$ from Eq. (3). (b) Varying portions of masked out weights from the gradient updates at steganography learning. We provide the SSIM of rendered views (blue) and recovered hidden images (green). + +![](images/0d787ff1f65773bcd571152f75a43b24c34cd1bec45c942feda514970b7ea221.jpg) +(b) Varying masking ratio + +![](images/11847d5d8073c01050251fecacb27c3f6503f1f979df108f36ad6e8dcee6d404.jpg) +(a) Varying JPEG compression rate +Figure 10. Analysis of robustness over (a) JPEG compression and (b) Gaussian blur. We provide the SSIM of rendered views (blue) and recovered hidden images (green). + +![](images/c17513acd083b98e6cc55af05ab707a660cc625073f8f3e1f94cffd0c3a1bfd4.jpg) +(b) Varying Gaussian blur std. +Figure 11. Results using a universal decoder to embed (a) multiple images for the flower scene and (b) a common image into different scenes from the LLFF dataset. The successful application under these two setups suggests that StegaNeRF can generalize across different scenes, user identities, and hidden images. + +# 5. More Discussions + +How useful is steganography for NeRF? Although NeRF-based 3D content has yet to become mainstream, we believe it will play a major future role not only for social platforms, but also for 3D vision research and applications. As people upload their personal NeRF models online for viewing purposes, NeRF steganography for ownership identification is apparently an important feature. Moreover, future 3D vision research will likely demand large-scale NeRF datasets of real-world images, and in this context, NeRF steganography can be a crucial tool for responsible + +![](images/496dc80a0391c8141c4ec35a8aefd20db91e807e929c522b0a9e2bc3868f8838.jpg) +Figure 8. Impact on visual quality when changing different components of the proposed StegaNeRF framework as in Tab. 3. We show the SSIM for the renderings and the recovered hidden images. +(a) Varying #watermark per scene + +![](images/a453434f03bfc17b213a281f2a129ffa0e0c7fdc33b91a7ad23e2eb97f953732.jpg) +(b) Varying #scene per watermark + +Table 4. The further empirical study on StegaNeRF. Results are averaged on the selected LLFF scenes. + +
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
StegaNeRF28.130.85680.1430100.00.9486
No Empty Sampling28.120.85570.1466100.00.9364
NeRF Frozen28.150.85750.137750.000.3975
Standard NeRF28.150.85750.137750.00N/A
+ +and ethical uses of training data and deployed models. + +Hidden or Memorized? StegaNeRF follows the standard setup of NeRF which prescribes per-scene training. Therefore, it may be reasonable to wonder whether the decoder is trained to "memorize" the hidden image without it really being injected into NeRF. To address this concern, we perform an experiment where the NeRF weights are frozen after the first stage and only update the decoder of StegaNeRF. If the decoder actually memorizes the image by itself, then we should expect good recovery performance under this setup. However, as shown in Tab. 4, with NeRF Frozen, we observe a significant performance degradation. Therefore, we can verify that the decoder alone cannot trivially memorize + +the hidden image, and the information is actually hidden in the NeRF renderings. Furthermore, we extend StegaNeRF to generalize across different scenes and hidden images while sharing the same universal decoder. Fig. 11a shows the rendering and decoding quality of embedding the different numbers of hidden images to a scene (flower). Fig. 11b shows the performance of embedding a common image into a varying number of scenes from the LLFF dataset. While the performance slightly decreases in the generalized setting, it still remains at a comparable level. + +Why not just share the rendered images instead of NeRF weights? Directly sharing the NeRF model gives end users the freedom to render at continuous viewpoints, which fundamentally differs from the traditional setting centered around discrete 2D image frames. It would not be realistic to assume people will stick with sharing 2D images instead of 3D weights, and therefore, it is necessary for the research community to see beyond 2D steganography and explore 3D NeRF steganography. + +Robustness: As shown in Fig. 12, in addition to using a resolution of $128 \times 128$ for hidden images in most of our experiments, we also tested resolutions ranging from $96 \times 96$ to $224 \times 224$ . We found that increasing the size of embedded images had no noticeable impact on the performance of StegaNeRF, with only a $0.6\%$ drop in rendering PSNR and a $1.2\%$ reduction in hidden SSIM. We also demonstrated the robustness of StegaNeRF to different sampling approaches during rendering, such as avoiding sampling in empty areas [43, 48]. Instead of considering all the points on a ray during rendering, we disregard the contribution of points to the integration along the entire ray if their density falls below a threshold, like 0.001. As shown in Tab. 4, replacing the conventional rendering strategy with No Empty Sampling has minimal effect on the performance of StegaNeRF, revealing our robustness to different rendering fashions. + +![](images/bd9ca95fb354943a142ca30d30f3b387482ba8c3fc3057cc4465b710876f9dc1.jpg) +Figure 12. Ablation on the resolutions of embedded images by the average results over the selected scenes from the LLFF dataset. + +![](images/6c2503ce199e0e20cbfe939a3d93590a725f4789d6d3c5f848b9bce2a363ad3f.jpg) + +Limitations: The current implementation is limited by the computational efficiency of NeRF, since the steganographic training stage requires performing gradient descent updates through the NeRF model. Another limitation is that, the recovery accuracy of the injected hidden information varies across different 3D scenes. More systematic analysis is required to understand the underlying factors that make some scenes easier to hide information than others. + +# 6. Conclusion + +With the progression of NeRF-based content, now serving as a pivotal element in diverse 3D applications, it is incumbent upon the research community to proactively tackle the imminent requirement to embed information within NeRF models as they are exchanged and disseminated across online platforms. This study introduces StegaNeRF, a pioneering framework for the automated embedding of steganographic information within NeRF renderings. We outline the essential components essential for achieving this functionality, unveiling potential challenges and revealing insights that can be instrumental for subsequent developers. By shedding light on the prospect of ownership identification within the realm of NeRF, this paper underscores the need for increased attention and concerted efforts toward addressing related challenges and avenues. + +# References + +[1] Benjamin Attal, Jia-Bin Huang, Michael Zollhoefer, Johannes Kopf, and Changil Kim. Learning Neural Light Fields With Ray-Space Embedding Networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19787-19797, 2022. 2 +[2] Shumeet Baluja. Hiding Images in Plain Sight: Deep Steganography. Advances in Neural Information Processing Systems, 30, 2017. 1, 2, 3, 5 +[3] Shumeet Baluja. Hiding Images Within Images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1685-1697, 2019. 1, 2 +[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-Nerf: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2 +[5] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5460–5469, 2022. 2 +[6] Oliver Benedens. Geometry-Based Watermarking of 3D Models. IEEE Computer Graphics and Applications, 19:46-55, 1999. 2 +[7] Ning Bi, Qiyu Sun, Daren Huang, Zhihua Yang, and Jiwu Huang. Robust Image Watermarking Based on Multiband Wavelets and Empirical Mode Decomposition. IEEE Transactions on Image Processing, 16(8):1956-1966, 2007. 2 +[8] Rajarathnam Chandramouli, Mehdi Kharrazi, and Nasir Memon. Image Steganography and Steganalysis: Concepts and Practice. In International Workshop on Digital Watermarking, pages 35-49. Springer, Springer, 2003. 2 +[9] Chin-Chen Chang, Ju-Yuan Hsiao, and Chi-Shiang Chan. Finding Optimal Least-Significant-Bit Substitution in Image Hiding by Dynamic Programming Strategy. Pattern Recognition, 36(7):1583-1595, 2003. 2, 3, 5 + +[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast Generalizable Radiance Field Reconstruction From Multi-View Stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124–14133, 2021. 1 +[11] Di Chen, Yu Liu, Lianghua Huang, Bin Wang, and Pan Pan. GeoAug: Data Augmentation for Few-Shot NeRF With Geometry Constraints. In European Conference on Computer Vision, pages 322-337. Springer, Springer, 2022. 1 +[12] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital Watermarking and Steganography. Morgan Kaufmann, 2007. 1, 2 +[13] Wenxue Cui, Shaohui Liu, Feng Jiang, Yongliang Liu, and Debin Zhao. Multi-Stage Residual Hiding for Image-Into-Audio Steganography. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2832-2836, 2020. 2 +[14] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. In European Conference on Computer Vision, pages 636-654. Springer, 2022. 1 +[15] Brandon Yushan Feng and Amitabh Varshney. SIGNET: Efficient Neural Representation for Light Fields. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14204-14213, 2021. 2 +[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance Fields Without Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2 +[17] Jessica Fridrich, Miroslav Goljan, and Rui Du. Detecting LSB Steganography in Color, and Gray-Scale Images. IEEE Multimedia, 8(4):22-28, 2001. 2 +[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic Neural Radiance Fields for Monocular 4D Facial Avatar Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2 +[19] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2 +[20] Margarita Geleta, Cristina Punti, Kevin McGuinness, Jordi Pons, Cristian Canton, and Xavier Giro-i Nieto. PixInWav: Residual Steganography for Hiding Pixels in Audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2485-2489, 2022. 2 +[21] Jamie Hayes and George Danezis. Generating Steganographic Images Via Adversarial Training. Advances in Neural Information Processing Systems, 30, 2017. 2 +[22] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2 + +[23] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. EfficientNeRF Efficient Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12902-12911, 2022. 2 +[24] Xudong Huang, Wei Li, Jie Hu, Hanting Chen, and Yunhe Wang. Refsr-nerf: Towards high fidelity and super resolution view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8244–8253, 2023. 1 +[25] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. HiNet: Deep Image Hiding by Invertible Network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4733-4742, 2021. 2 +[26] Gary C Kessler and Chet Hosmer. An Overview of Steganography. Advances in Computers, 83:51-107, 2011. 2 +[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification With Deep Convolutional Neural Networks. Communications of the ACM, 60(6):84-90, 2017. 5 +[28] Jaeho Lee, Jihoon Tack, Namhoon Lee, and Jinwoo Shin. Meta-Learning Sparse Implicit Neural Representations. Advances in Neural Information Processing Systems, 34:11769-11780, 2021. 4 +[29] Chenxin Li, Mingbao Lin, Zhiyuan Ding, Nie Lin, Yihong Zhuang, Yue Huang, Xinghao Ding, and Liujuan Cao. Knowledge condensation distillation. In European Conference on Computer Vision, pages 19-35. Springer, 2022. 7 +[30] Chenxin Li, Wenao Ma, Liyan Sun, Xinghao Ding, Yue Huang, Guisheng Wang, and Yizhou Yu. Hierarchical deep network with uncertainty-aware semi-supervised learning for vessel segmentation. Neural Computing and Applications, pages 1-14, 2022. 4 +[31] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 61-65. IEEE, 2021. 7 +[32] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2 +[33] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Y. Wang, and Shenlong Wang. NeurMiPs: Neural Mixture of Planar Experts for View Synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15681-15691, 2022. 2 +[34] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural Sparse Voxel Fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2 +[35] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason M. Saragih. Mixture of Volumetric Primitives for Efficient Neural Rendering. ACM Transactions on Graphics (TOG), 40:1 - 13, 2021. 2 + +[36] Shao-Ping Lu, Rong Wang, Tao Zhong, and Paul L Rosin. Large-Capacity Image Steganography Based on Invertible Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10816-10825, 2021. 2 +[37] Luma AI. https://lumalabs.ai/.2 +[38] Xiyang Luo, Yinxiao Li, Huiwen Chang, Ce Liu, Peyman Milanfar, and Feng Yang. Dvmark: A Deep Multiscale Framework for Video Watermarking. ArXiv Preprint ArXiv:2104.12734, 2021. 2 +[39] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210–7219, 2021. 2, 4 +[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy Networks: Learning 3d Reconstruction in Function Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 1 +[41] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local Light Field Fusion: Practical View Synthesis With Prescriptive Sampling Guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 4 +[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing Scenes As Neural Radiance Fields for View Synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 4 +[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives With a Multiresolution Hash Encoding. ACM Transactions on Graphics (TOG), 41:1-15, 2022. 1, 2, 9 +[44] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing Neural Radiance Fields for View Synthesis From Sparse Inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 1 +[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning Continuous Signed Distance Functions for Shape Representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 1 +[46] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2 +[47] Tomás Pevny, Tomás Filler, and Patrick Bas. Using High-Dimensional Image Models to Perform Highly Undetectable Steganography. In International Workshop on Information Hiding, pages 161-177. Springer, Springer, 2010. 2 +[48] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. In 2021 Interna + +tional Conference on 3D Vision (3DV), pages 1106-1114. IEEE, 2021.9 +[49] Polycam. https://poly.cam/.2 +[50] Emil Praun, Hugues Hoppe, and Adam Finkelstein. Robust Mesh Watermarking. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, pages 49-56, 1999. 2 +[51] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2 +[52] Yinlong Qian, Jing Dong, Wei Wang, and Tieniu Tan. Deep Learning for Steganalysis Via Convolutional Neural Networks. In Media Watermarking, Security, and Forensics 2015, volume 9409, pages 171-180. SPIE, SPIE, 2015. 2 +[53] Chen Quei-An. Nerf-pl: A Pytorch-Lighting Implementation of NeRF. https://github.com/kwea123/nerf_pl/.4 +[54] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding Up Neural Radiance Fields With Thousands of Tiny Mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2 +[55] Mennatallah M Sadek, Amal S Khalifa, and Mostafa GM Mostafa. Video Steganography: A Comprehensive Review. Multimedia Tools and Applications, 74(17):7063-7094, 2015. 2 +[56] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-Aware Image Synthesis With Sparse Voxel Grids. ArXiv Preprint ArXiv:2206.07695, 2022. 1 +[57] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 1 +[58] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit Neural Representations With Periodic Activation Functions. Advances in Neural Information Processing Systems, 33, 2020. 1 +[59] Vincent Sitzmann, Semon Rezchikov, William T. Freeman, Joshua B. Tenenbaum, and Frédo Durand. Light Field Networks: Neural Scene Representations With Single-Evaluation Rendering. Advances in Neural Information Processing Systems, 34, 2021. 2 +[60] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 2 +[61] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct Voxel Grid Optimization: Super-Fast Convergence for Radiance Fields Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1, 2 +[62] Abdelfatah A Tamimi, Ayman M Abdalla, and Omaima Al-Allaf. Hiding an Image Inside Another Image Using Variable-Rate Steganography. International Journal of Advanced Computer Science and Applications (IJACSA), 4(10), 2013. 2 + +[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-Nerf: Scalable Large Scene Neural View Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248-8258, 2022. 1, 2 +[64] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible Hyperlinks in Physical Photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2126, 2020. 2 +[65] Matthew Tancik, Pratul P Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T Barron, and Ren Ng. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. Advances in Neural Information Processing Systems, 33, 2020. 1 +[66] Weixuan Tang, Shunquan Tan, Bin Li, and Jiwu Huang. Automatic Steganographic Distortion Learning Using a Generative Adversarial Network. IEEE Signal Processing Letters, 24(10):1547-1551, 2017. 2, 6 +[67] Jeroen JG Van Merrienboer, Liesbeth Kester, and Fred Paas. Teaching complex rather than simple tasks: Balancing intrinsic and germane load to enhance transfer of learning. Applied Cognitive Psychology: The Official Journal of the Society for Applied Research in Memory and Cognition, 20(3):343-352, 2006. 4 +[68] Mukund Varma, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that nef needs? In The Eleventh International Conference on Learning Representations, 2022. 2 +[69] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and S. Tulyakov. R2L: Distilling Neural Radiance Field to Neural Light Field for Efficient Novel View Synthesis. In European Conference on Computer Vision. Springer, Springer, 2022. 2 +[70] Huaiqing Wang and Shuozhong Wang. Cyber Warfare: Steganography vs. Steganalysis. Communications of the ACM, 47(10):76-82, 2004. 2 +[71] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier PlenOctrees for Dynamic Radiance Field Rendering in Real-Time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 1 +[72] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning Neural Implicit Surfaces by Volume Rendering for Multi-View Reconstruction. Advances in Neural Information Processing Systems, 34, 2021. 1 +[73] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning Multi-View Image-Based Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690–4699, 2021. 2 +[74] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image Quality Assessment: From Error Visibility + +to Structural Similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5 +[75] Xinyu Weng, Yongzhi Li, Lu Chi, and Yadong Mu. High-Capacity Convolutional Video Steganography With Temporal Residual Modeling. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, pages 87-95, 2019. 2 +[76] Raymond B Wolfgang and Edward J Delp. A Watermark for Digital Images. In Proceedings of 3rd IEEE International Conference on Image Processing, volume 3, pages 219-222. IEEE, IEEE, 1996. 2 +[77] Yue Wu, Guotao Meng, and Qifeng Chen. Embedding Novel Views in a Single JPEG Image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14519-14527, 2021. 2 +[78] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2 +[79] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. BungeeNeRF: Progressive Neural Radiance Field for Extreme Multi-Scale Scene Rendering. In The European Conference on Computer Vision (ECCV), volume 2, 2022. 2 +[80] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1 +[81] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4479–4489, 2023. 1 +[82] Hyukryul Yang, Hao Ouyang, Vladlen Koltun, and Qifeng Chen. Hiding Video in Audio Via Reversible Generative Models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1100-1109, 2019. 2 +[83] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5491–5500, 2022. 1, 4 +[84] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for Real-Time Rendering of Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2 +[85] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. Pixelnerf: Neural Radiance Fields From One or Few Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1 +[86] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14448-14457, 2021. 2 + +[87] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A Structured Dictionary Perspective on Implicit Neural Representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 4 +[88] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray Priors Through Reprojection: Improving Neural Radiance Fields for Novel View Extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18376-18386, 2022. 1 +[89] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and Improving Neural Radiance Fields. ArXiv Preprint ArXiv:2010.07492, 2020. 2 +[90] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features As a Perceptual Metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5 +[91] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul E. Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural Factorization of Shape and Reflectance Under an Unknown Illumination. ACM Trans. Graph, 40:237:1-237:18, 2021. 1 \ No newline at end of file diff --git a/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/images.zip b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d4d3efe875ca0ae8e86e69e88419f7a53ad4db71 --- /dev/null +++ b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02203886783fde23cd9a7e14f6c2eff2731c54d21a64daafca73d1924eb351dc +size 715088 diff --git a/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/layout.json b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e743222e76be3bb9a40f29bbcf902b30f3c3594a --- /dev/null +++ b/2023/StegaNeRF_ Embedding Invisible Information within Neural Radiance Fields/layout.json @@ -0,0 +1,12807 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "text", + "content": "StegaNeRF: Embedding Invisible Information within Neural Radiance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "spans": [ + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "text", + "content": "Chenxin Li" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "text", + "content": ", Brandon Y. Feng" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "text", + "content": ", Zhiwen Fan" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "text", + "content": ", Panwang Pan" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "text", + "content": ", Zhangyang Wang" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "text", + "content": "Chinese University of Hong Kong, " + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "text", + "content": "University of Maryland, " + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "text", + "content": "University of Texas at Austin, " + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 142, + 496, + 186 + ], + "type": "text", + "content": "ByteDance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 239, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 239, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 45, + 239, + 290, + 515 + ], + "type": "text", + "content": "Recent advancements in neural rendering have paved the way for a future marked by the widespread distribution of visual data through the sharing of Neural Radiance Field (NeRF) model weights. However, while established techniques exist for embedding ownership or copyright information within conventional visual data such as images and videos, the challenges posed by the emerging NeRF format have remained unaddressed. In this paper, we introduce StegaNeRF, an innovative approach for steganographic information embedding within NeRF renderings. We have meticulously developed an optimization framework that enables precise retrieval of hidden information from images generated by NeRF, while ensuring the original visual quality of the rendered images to remain intact. Through rigorous experimentation, we assess the efficacy of our methodology across various potential deployment scenarios. Furthermore, we delve into the insights gleaned from our analysis. StegaNeRF represents an initial foray into the intriguing realm of infusing NeRF renderings with customizable, imperceptible, and recoverable information, all while minimizing any discernible impact on the rendered images. For more details, please visit our project page: https://xggnet.github.io/StegaNeRF/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 128, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 560, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 288, + 693 + ], + "type": "text", + "content": "Implicit neural representation (INR) is an emerging concept where the network describes the data through its weights [14, 24, 40, 42, 45, 58, 65]. After training, the INR weights can then be used for content distribution, streaming, and even downstream inference tasks, all without sending or storing the original data. Arguably the most prominent INR is Neural Radiance Fields (NeRF) [42], where a network learns a continuous function mapping spatial coordinates to density and color. Due to its lightweight size and superb quality, NeRF has immense potential for 3D content representation in future vision and graphics applications." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 214, + 545, + 285 + ], + "blocks": [ + { + "bbox": [ + 307, + 214, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 307, + 214, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 214, + 545, + 285 + ], + "type": "image", + "image_path": "fc98755da4cfee183d3e717dde8e24380c2629c36ff498bf79fe2a72540a46a9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 295, + 547, + 340 + ], + "lines": [ + { + "bbox": [ + 305, + 295, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 547, + 340 + ], + "type": "text", + "content": "Figure 1. We introduce the new problem of NeRF steganography: hiding information in NeRF renderings. Our proposed framework, StegaNeRF, can embed and recover customized hidden information while preserving the original NeRF rendering quality." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 361, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 422 + ], + "type": "text", + "content": "While there is a plethora of work dedicated towards better quality [4,63,72,91], faster rendering [43,56,61,71,83], and sparse view reconstruction [10, 11, 44, 57, 80, 81, 85, 88], in this paper, we look beyond the horizon and explore a new question: Can we achieve steganography with NeRF?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 423, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 547, + 519 + ], + "type": "text", + "content": "Established digital steganography method [12] focus on embedding hidden messages in 2D images. The recent growth of deep learning and social media platforms further gives rise to many practical use cases of image steganography. As countless images and videos are shared online and even used to train deep learning models, 2D steganography methods [2,3] allow users and data providers to protect copyright, embed ownership, and prevent content abuse." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": "Now, with the ongoing advances in 3D representations powered by NeRF, we envision a future where people frequently share their captured 3D content online just as they are currently sharing 2D images and videos online. Moreover, we are curious to explore the following research questions: 1) Injecting information into 2D images for copyright or ownership identification is common, but can we preserve such information when people share and render 3D scenes through NeRF? 2) NeRF can represent large-scale real-world scenes with training images taken by different people, but can we preserve these multiple source identities in the NeRF renderings to reflect the collaborative efforts required to reconstruct these 3D scenes? 3) Common image steganography methods embed either a hidden image or a message string into a given image, but can we allow different modalities of the hidden signal in NeRF steganography?" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 125, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 125, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 125, + 713 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "441" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "Driven by these questions, we formulate a framework to embed customizable, imperceptible, and recoverable information in NeRF renderings without sacrificing the visual quality. Fig. 1 presents an overview of our proposed framework, dubbed StegaNeRF. Unlike the traditional image steganography that embeds hidden signals only into a specific source image, we wish to recover the same intended hidden signal from NeRF rendered at arbitrary viewpoints." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 357 + ], + "type": "text", + "content": "Despite many established works on 2D steganography and hidden watermarks for image and video [2, 3, 9, 25, 36, 64], naively applying 2D steganography on the NeRF training images is not practical, since the embedded information easily gets lost in the actual NeRF renderings. In contrast, our framework enables reliable extraction of the hidden signal from NeRF renderings. During NeRF training, we jointly optimize a decoder network to extract hidden information from the NeRF renderings. To minimize the negative impact on the visual quality of NeRF, we identify weights with low impact on rendering and introduce a gradient masking strategy to steer the hidden steganographic information towards those low-importance weights. Extensive experimental results validate StegaNeRF balances between the rendering quality of novel views and the high-fidelity transmission of the concealed information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 360, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 287, + 395 + ], + "type": "text", + "content": "StegaNeRF presents the first exploration of hiding information in NeRF models for ownership identification, and our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 404, + 287, + 597 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 404, + 286, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 404, + 286, + 440 + ], + "spans": [ + { + "bbox": [ + 59, + 404, + 286, + 440 + ], + "type": "text", + "content": "- We introduce the new problem of NeRF steganography and present the first effort to embed customizable, imperceptible, and recoverable information in NeRF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 449, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 449, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 59, + 449, + 287, + 497 + ], + "type": "text", + "content": "- We propose an adaptive gradient masking strategy to steer the injected hidden information towards the less important NeRF weights, balancing the objectives of steganography and rendering quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 506, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 506, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 59, + 506, + 287, + 553 + ], + "type": "text", + "content": "- We empirically validate the proposed framework on a diverse set of 3D scenes with different camera layouts and scene characteristics, obtaining high recovery accuracy without sacrificing rendering quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 562, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 562, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 59, + 562, + 287, + 597 + ], + "type": "text", + "content": "- We explore various scenarios applying StegaNeRF for ownership identification, with the additional support of multiple identities and multi-modal signals." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 609, + 133, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 133, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 133, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Neural Radiance Fields Following the success of NeRF [42], numerous techniques have been incorporated including ray re-parameterizations [4, 5], explicit spatial data structures [16, 23, 34, 43, 61, 84], caching and distillation [19, 22, 54, 69], ray-based representations [1, 15, 59], geometric primitives [33, 35], large-scale scenes [39, 63, 79, 89], generalizable architectures [60, 68, 73], and dynamic" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "settings [18, 32, 46, 51, 78]. Unlike prior works, this paper explores the uncharted problem of embedding information in NeRF renderings, with critical implications for copyright protection and ownership preservation. As early-stage NeRF-based products already become available [37,49] and more are expected, we believe now is the right time to open up the exploration of NeRF steganography." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 162, + 546, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 162, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 162, + 546, + 390 + ], + "type": "text", + "content": "Image Steganography Steganography hides intended signals as invisible watermarks (e.g., hyperlinks, images) within the cover media called carriers (e.g., images, video and audio) [12, 26]. Classical methods focus on seeking and altering the least significant bits (LSBs) [17, 47, 62, 76] and transforming domain techniques [7, 8, 52, 70]. Prior research also uses deep neural networks for steganography [2, 3, 21, 66]. Among them, DeepStega [2] conceals a hidden image within an equal-sized carrier image. Subsequent works [25, 36] use invertible networks to improve the performance of deep image steganography. Another line of work conceals information in other carrier media like audio [13, 20, 82] and video [38, 55, 75]. The above advances all play a critical part in the era when traditional media formats like images and videos are dominant. However, as MLP-based neural representations of 3D scenes are gaining momentum to become a major format of visual data, extending steganography to NeRF is bound to become an important problem in the upcoming future." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 396, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 528 + ], + "type": "text", + "content": "Lifting Steganography to 3D Prior to NeRF, meshes are commonly used to represent 3D shapes. Early pioneers apply steganography 3D meshes [6, 50] for copyright protection when meshes are exchanged and edited. More recent work [77] has also explored embedding multi-plane images within a JPEG-compressed image. Their problem can be regarded as a special case of 2D steganography, hiding multiple 2D images inside a single 2D image. In contrast, we try to hide a natural image into a 3D scene representation (NeRF), fundamentally differing from these prior methods where 2D images act as the carrier of hidden information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 538, + 361, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 361, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 361, + 550 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "text", + "content": "The goal of StegaNeRF is to inject customized (steganographic) information into the NeRF weights with imperceptible visual changes when rendering. Given a NeRF with weights " + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "text", + "content": " and the information " + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "text", + "content": " to be injected, when we render with " + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "text", + "content": " on any viewpoint, we hope the injected " + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "text", + "content": " can be recovered by a decoder " + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "text", + "content": " with learnable weights " + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "A seemingly obvious solution is to use prior image steganography methods to 1) inject " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " on training images, 2) train NeRF on those images with embedded information, 3) apply their provided " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " to extract " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " from NeRF-rendered images. This approach has been successfully applied for GAN fingerprinting [86]. However, it fails in the NeRF context (see Fig. 3), where the subtle changes induced by prior" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "442" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 378, + 216 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 378, + 216 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 378, + 216 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 378, + 216 + ], + "type": "image", + "image_path": "b29770537380d4b638d8474caadb3fcc7f31d0cbff08beb0c0e4ccfa34f9979e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "type": "text", + "content": "Figure 2. StegaNeRF training overview. At the first stage, we optimize " + }, + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "type": "text", + "content": " with standard NeRF training. At the second stage, we initialize " + }, + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "type": "text", + "content": " and optimize for the steganography objectives. We train the decoder " + }, + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 46, + 225, + 546, + 270 + ], + "type": "text", + "content": " to recover hidden information from StegaNeRF renderings and no hidden information from original NeRF renderings. We introduce Classifier Guided Recovery to improve the accuracy of recovered information, and Adaptive Gradient Masking to balance between steganography ability and rendering visual quality." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 381, + 71, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 381, + 71, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 381, + 71, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 381, + 71, + 545, + 216 + ], + "type": "image", + "image_path": "ad4fe3df72f82e271759a77cc7e3347be07d4bfc2df6cafe49d17756ee69e18c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 327 + ], + "type": "text", + "content": "steganography methods easily get smoothed out, inhibiting the decoder from identifying the subtle patterns necessary for information recovery." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 399 + ], + "type": "text", + "content": "Such shortcomings of off-the-shelf steganographic methods are not surprising since they are developed for the traditional setting where 2D images are the ultimate form of visual information. In contrast, our problem setting involves the new concept of INR as the underlying representation, and 2D images are just the final output rendered by NeRF." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 406, + 185, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 185, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 185, + 419 + ], + "type": "text", + "content": "3.1. Two-Stage Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 425, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 496 + ], + "type": "text", + "content": "Recognizing the formulation difference due to the emergence of implicit representation with NeRF, we move away from the traditional 2D image-based pipeline that trains an encoder network to inject subtle changes to the given 2D images. Instead, we incorporate the steganography objective into the gradient-based learning process of NeRF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 497, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 616 + ], + "type": "text", + "content": "We re-design the training as a two-stage optimization. The first stage is the original NeRF training procedure, involving the standard photometric loss between the rendered and ground truth pixels to guarantee the visual quality. After finishing the weights update " + }, + { + "bbox": [ + 46, + 497, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 497, + 287, + 616 + ], + "type": "text", + "content": " at the first stage, we dedicate the second stage to obtain the final NeRF weights " + }, + { + "bbox": [ + 46, + 497, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 497, + 287, + 616 + ], + "type": "text", + "content": " containing steganographic information. We introduce several techniques to achieve robust information recovery with imperceptible impact on the rendered images. The training workflow is depicted in Fig. 2 and Alg. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 623, + 173, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 173, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 173, + 635 + ], + "type": "text", + "content": "3.2. Information Recovery" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " denote the camera pose at which we render an image from a NeRF network. We want to recover " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " from the rendered image " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\theta(P)" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": ". Importantly, we also want to avoid false positives on images " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\theta_0(P)" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " rendered by the original network without steganography ability, even if the images rendered by " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " look visually identical. Therefore, we" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 291, + 364, + 386 + ], + "blocks": [ + { + "bbox": [ + 307, + 291, + 364, + 386 + ], + "lines": [ + { + "bbox": [ + 307, + 291, + 364, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 364, + 386 + ], + "type": "image", + "image_path": "f3e1e4ae25f1e21df1c01ce78e682cc666d0c0f30d5607ea1ba813a143312571.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 399, + 545, + 488 + ], + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 488 + ], + "type": "text", + "content": "Figure 3. Results of applying prior 2D steganography methods, LSB [9] (top) and DeepStega [2] (bottom). From left to right, we show (a) training image after applying 2D steganography, (b) residual error of (a) over ground truth, (c) hidden image recovered from (a), (d) residual error of NeRF rendering at the pose of (a), and (e) hidden image recovered from (d). Prior 2D steganography methods fail in the NeRF context since the hidden information injected in training images mostly disappear in NeRF renderings." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 365, + 292, + 421, + 389 + ], + "blocks": [ + { + "bbox": [ + 365, + 292, + 421, + 389 + ], + "lines": [ + { + "bbox": [ + 365, + 292, + 421, + 389 + ], + "spans": [ + { + "bbox": [ + 365, + 292, + 421, + 389 + ], + "type": "image", + "image_path": "cbe4fda3609a084f42a8a14db3790f72757d45490303723720d58d477593ac29.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 425, + 292, + 449, + 389 + ], + "blocks": [ + { + "bbox": [ + 425, + 292, + 449, + 389 + ], + "lines": [ + { + "bbox": [ + 425, + 292, + 449, + 389 + ], + "spans": [ + { + "bbox": [ + 425, + 292, + 449, + 389 + ], + "type": "image", + "image_path": "9f560c7b3ab4013bb5972d53e73e75ad83c244fd7deacb4a8179578a4eee9aa3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 457, + 292, + 512, + 389 + ], + "blocks": [ + { + "bbox": [ + 457, + 292, + 512, + 389 + ], + "lines": [ + { + "bbox": [ + 457, + 292, + 512, + 389 + ], + "spans": [ + { + "bbox": [ + 457, + 292, + 512, + 389 + ], + "type": "image", + "image_path": "714106e3959ffa6ca22a7f1d49393f175a822b87c47e862453cedeb389222dbe.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 517, + 292, + 541, + 389 + ], + "blocks": [ + { + "bbox": [ + 517, + 292, + 541, + 389 + ], + "lines": [ + { + "bbox": [ + 517, + 292, + 541, + 389 + ], + "spans": [ + { + "bbox": [ + 517, + 292, + 541, + 389 + ], + "type": "image", + "image_path": "7788ee5077cd1a502af07201b4d246dc962101c0ce7ce1cb52d330e9b22434c1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 509, + 544, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 544, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 544, + 521 + ], + "type": "text", + "content": "optimize " + }, + { + "bbox": [ + 306, + 509, + 544, + 521 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 306, + 509, + 544, + 521 + ], + "type": "text", + "content": " to minimize the following contrastive loss terms:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 543, + 545, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 543, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 310, + 543, + 545, + 557 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e c} ^ {+} = \\left| F _ {\\psi} (\\boldsymbol {\\theta} (P)) - I \\right|, \\quad \\mathcal {L} _ {d e c} ^ {-} = \\left| F _ {\\psi} \\left(\\boldsymbol {\\theta} _ {0} (P)\\right) - \\varnothing \\right|, \\tag {1}", + "image_path": "8bdfacf377c677ac7cdf9b5cb16dbdd373c367728d91e50488c3946c0a628b84.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": " is an empty image with all zeros. Effectively, " + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "inline_equation", + "content": "L_{dec}^{+}" + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": " serves as a regularization term, facilitating the decoder to recover the embedded image pattern based on renderings obtained from a StegaNeRF. In contrast, " + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "inline_equation", + "content": "L_{dec}^{-}" + }, + { + "bbox": [ + 304, + 565, + 545, + 650 + ], + "type": "text", + "content": " prevents the decoder from falsely producing any plausible image pattern when given renderings from a Standard NeRF that is without hidden signals." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "Classifier-Guided Recovery The decoder " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " is easily implemented as a U-Net to decode " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " as the form of 2D images, but accurately recovering all the details in " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " might be challenging. Therefore, we additionally train a classifier" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "443" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 244, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 244, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 244, + 84 + ], + "type": "text", + "content": "Algorithm 1 Train StegaNeRF on a single scene" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 86, + 286, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 86, + 286, + 110 + ], + "spans": [ + { + "bbox": [ + 47, + 86, + 286, + 110 + ], + "type": "text", + "content": "Data: Training images " + }, + { + "bbox": [ + 47, + 86, + 286, + 110 + ], + "type": "inline_equation", + "content": "\\{Y_i\\}" + }, + { + "bbox": [ + 47, + 86, + 286, + 110 + ], + "type": "text", + "content": " with poses " + }, + { + "bbox": [ + 47, + 86, + 286, + 110 + ], + "type": "inline_equation", + "content": "\\{P_i\\}" + }, + { + "bbox": [ + 47, + 86, + 286, + 110 + ], + "type": "text", + "content": ", hidden information " + }, + { + "bbox": [ + 47, + 86, + 286, + 110 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 86, + 286, + 110 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 47, + 86, + 286, + 110 + ], + "type": "inline_equation", + "content": "\\eta = [\\eta_0, \\eta_1]" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 110, + 249, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 249, + 121 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 249, + 121 + ], + "type": "text", + "content": "Output: Steganographic NeRF " + }, + { + "bbox": [ + 47, + 110, + 249, + 121 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 110, + 249, + 121 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 47, + 110, + 249, + 121 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 122, + 225, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 225, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 225, + 133 + ], + "type": "text", + "content": "Initialize NeRF " + }, + { + "bbox": [ + 47, + 122, + 225, + 133 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 47, + 122, + 225, + 133 + ], + "type": "text", + "content": " and decoder network " + }, + { + "bbox": [ + 47, + 122, + 225, + 133 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 134, + 264, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 134, + 264, + 146 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 264, + 146 + ], + "type": "text", + "content": "Optimize " + }, + { + "bbox": [ + 47, + 134, + 264, + 146 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 47, + 134, + 264, + 146 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 47, + 134, + 264, + 146 + ], + "type": "inline_equation", + "content": "\\{Y_i, P_{ij}\\}" + }, + { + "bbox": [ + 47, + 134, + 264, + 146 + ], + "type": "text", + "content": " with standard NeRF training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 147, + 201, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 147, + 201, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 147, + 201, + 158 + ], + "type": "text", + "content": "Compute mask " + }, + { + "bbox": [ + 47, + 147, + 201, + 158 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 147, + 201, + 158 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 147, + 201, + 158 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 47, + 147, + 201, + 158 + ], + "type": "text", + "content": " as in Eq. (3)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 159, + 172, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 159, + 172, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 159, + 172, + 169 + ], + "type": "text", + "content": "for each training iteration " + }, + { + "bbox": [ + 47, + 159, + 172, + 169 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 159, + 172, + 169 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 171, + 276, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 276, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 276, + 183 + ], + "type": "text", + "content": "Randomly sample a training pose " + }, + { + "bbox": [ + 56, + 171, + 276, + 183 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 56, + 171, + 276, + 183 + ], + "type": "text", + "content": " and render " + }, + { + "bbox": [ + 56, + 171, + 276, + 183 + ], + "type": "inline_equation", + "content": "\\theta (P_i)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 184, + 276, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 276, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 276, + 196 + ], + "type": "text", + "content": "Compute stega. losses " + }, + { + "bbox": [ + 56, + 184, + 276, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}^{c}" + }, + { + "bbox": [ + 56, + 184, + 276, + 196 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 184, + 276, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}^{+}" + }, + { + "bbox": [ + 56, + 184, + 276, + 196 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 184, + 276, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}^{-}" + }, + { + "bbox": [ + 56, + 184, + 276, + 196 + ], + "type": "text", + "content": " as Eq. (1), (2)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 196, + 270, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 270, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 270, + 207 + ], + "type": "text", + "content": "Compute " + }, + { + "bbox": [ + 56, + 196, + 270, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 56, + 196, + 270, + 207 + ], + "type": "text", + "content": " and combine total loss " + }, + { + "bbox": [ + 56, + 196, + 270, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 56, + 196, + 270, + 207 + ], + "type": "text", + "content": " as in Eq. (4)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 207, + 263, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 263, + 221 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 263, + 221 + ], + "type": "text", + "content": "Update " + }, + { + "bbox": [ + 56, + 207, + 263, + 221 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 56, + 207, + 263, + 221 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 56, + 207, + 263, + 221 + ], + "type": "inline_equation", + "content": "\\eta_0\\cdot (\\frac{\\partial\\mathcal{L}}{\\partial\\pmb{\\theta}}\\odot \\pmb {m})" + }, + { + "bbox": [ + 56, + 207, + 263, + 221 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 207, + 263, + 221 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 56, + 207, + 263, + 221 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 56, + 207, + 263, + 221 + ], + "type": "inline_equation", + "content": "\\eta_{1}\\cdot \\frac{\\partial\\mathcal{L}}{\\partial\\psi}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 222, + 81, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 222, + 81, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 222, + 81, + 232 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "type": "text", + "content": "network " + }, + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "type": "inline_equation", + "content": "F_{\\psi}^{c}" + }, + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "type": "text", + "content": " to solve the easier task of classifying whether the given NeRF rendering contains hidden information. " + }, + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "type": "inline_equation", + "content": "F_{\\psi}^{c}" + }, + { + "bbox": [ + 46, + 253, + 287, + 290 + ], + "type": "text", + "content": " is optimized by the following cross-entropy loss [30, 67]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 297, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 287, + 312 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e c} ^ {c} = - \\log \\left(F _ {\\psi} ^ {c} (\\boldsymbol {\\theta} (P))\\right) - \\log \\left(1 - F _ {\\psi} ^ {c} \\left(\\boldsymbol {\\theta} _ {0} (P)\\right)\\right). \\tag {2}", + "image_path": "d8129c9ed6ba3478dfde69f6f25d752a94a96d4613322f589b87f6ed5125f450.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "text", + "content": "We then use its prediction to guide the process of decoding pixel-wise information by adding the classification output as input to " + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "inline_equation", + "content": "F_{\\psi}(x) = F_{\\psi}\\big(x,F_{\\psi}^{c}(x)\\big)" + }, + { + "bbox": [ + 46, + 319, + 287, + 356 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 356, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 287, + 416 + ], + "type": "text", + "content": "Although the above discussion focuses on hiding images, our framework can be easily extended to embed other modalities like strings, text, or even audio, all of which may be represented as 1D vectors. We can simply modify the architecture of " + }, + { + "bbox": [ + 46, + 356, + 287, + 416 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 46, + 356, + 287, + 416 + ], + "type": "text", + "content": " to have a 1D prediction branch." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 423, + 213, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 213, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 213, + 436 + ], + "type": "text", + "content": "3.3. Preserving Perceptual Identity" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 441, + 287, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 560 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 560 + ], + "type": "text", + "content": "Since we want to hide information without affecting the visual perception of the rendered output, an intuitive regularization is to penalize how much " + }, + { + "bbox": [ + 46, + 441, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 441, + 287, + 560 + ], + "type": "text", + "content": " deviates from " + }, + { + "bbox": [ + 46, + 441, + 287, + 560 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 441, + 287, + 560 + ], + "type": "text", + "content": ". However, we find that naively summing the deviation penalty across all weights makes it difficult for the NeRF network to adjust its weights for the steganographic objective. Instead, motivated by the fact that INR weights are not equally important and exhibit strong sparsity [28, 87], we propose an adaptive gradient masking strategy to encode the hidden information on specific weights' groups." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "type": "text", + "content": "Formally, given the initial set of weights " + }, + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\theta_0\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "type": "text", + "content": " , we compute a soft importance mask " + }, + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "type": "inline_equation", + "content": "m\\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "type": "text", + "content": " . Specifically, for each weight " + }, + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\pmb {w}_i\\in \\pmb {\\theta}_0" + }, + { + "bbox": [ + 46, + 560, + 287, + 597 + ], + "type": "text", + "content": " , we compute:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 124, + 605, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 605, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 124, + 605, + 287, + 634 + ], + "type": "interline_equation", + "content": "\\boldsymbol {m} _ {i} = \\frac {\\left| \\boldsymbol {w} _ {i} \\right| ^ {- \\alpha}}{\\sum_ {i} ^ {N} \\left| \\boldsymbol {w} _ {i} \\right| ^ {- \\alpha}}, \\tag {3}", + "image_path": "895e5cc502bf784068f3adc8a467b395e37ff1199e3e55c306ba754f20ff5e85.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": "where the magnitude " + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": " of each model weight " + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": " serves as its individual importance indicator, and " + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha > 0" + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": " is a scaling factor of power controlling the relative distribution of importance across the weights. We mask the gradient as " + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial\\theta} \\odot m" + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": " when optimizing " + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": " based on the total loss " + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": " in the second stage, where " + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 641, + 287, + 714 + ], + "type": "text", + "content": " is a Hadamard product. Effectively," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 72, + 513, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 513, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 513, + 84 + ], + "type": "text", + "content": "Algorithm 2 Typical usage scenario of StegaNeRF" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 87, + 544, + 207 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 87, + 495, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 87, + 495, + 99 + ], + "spans": [ + { + "bbox": [ + 312, + 87, + 495, + 99 + ], + "type": "text", + "content": "1: Alice captures some images of a 3D scene" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 100, + 544, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 100, + 544, + 111 + ], + "spans": [ + { + "bbox": [ + 312, + 100, + 544, + 111 + ], + "type": "text", + "content": "2: Alice trains a StegaNeRF to hide a personalized image" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 112, + 544, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 112, + 544, + 134 + ], + "spans": [ + { + "bbox": [ + 312, + 112, + 544, + 134 + ], + "type": "text", + "content": "3: Alice shares the model " + }, + { + "bbox": [ + 312, + 112, + 544, + 134 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 312, + 112, + 544, + 134 + ], + "type": "text", + "content": " online for other people to enjoy and explore the 3D scene themselves" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 135, + 544, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 135, + 544, + 159 + ], + "spans": [ + { + "bbox": [ + 312, + 135, + 544, + 159 + ], + "type": "text", + "content": "4: Bob grabs the model " + }, + { + "bbox": [ + 312, + 135, + 544, + 159 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 312, + 135, + 544, + 159 + ], + "type": "text", + "content": " and reposts it with his own account without crediting Alice or asking for permission" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 159, + 544, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 159, + 544, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 159, + 544, + 182 + ], + "type": "text", + "content": "5: Alice sees Bob's post, deploys the decoder " + }, + { + "bbox": [ + 312, + 159, + 544, + 182 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 312, + 159, + 544, + 182 + ], + "type": "text", + "content": ", and verifies the owner of " + }, + { + "bbox": [ + 312, + 159, + 544, + 182 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 312, + 159, + 544, + 182 + ], + "type": "text", + "content": " is Alice, not Bob" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 183, + 544, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 183, + 544, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 183, + 544, + 207 + ], + "type": "text", + "content": "6: Bob takes down the post or gets banned for copyright infringement" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 230, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 266 + ], + "type": "text", + "content": "more significant weights are \"masked out\" to minimize the impact of steganographic learning on the rendered visual quality." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 266, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 313 + ], + "type": "text", + "content": "We retain the standard photometric error in steganography learning to maintain NeRF rendering fidelity: " + }, + { + "bbox": [ + 304, + 266, + 545, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb} = |\\pmb {\\theta}(P) - \\pmb {\\theta}_0(P)|" + }, + { + "bbox": [ + 304, + 266, + 545, + 313 + ], + "type": "text", + "content": ". The overall training loss at the second stage can be formulated as follows:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 335, + 321, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 321, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 335, + 321, + 545, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {0} \\mathcal {L} _ {d e c} ^ {c} + \\lambda_ {1} \\mathcal {L} _ {d e c} ^ {+} + \\lambda_ {2} \\mathcal {L} _ {d e c} ^ {-} + \\lambda_ {3} \\mathcal {L} _ {r g b}. \\tag {4}", + "image_path": "96a60134b6b4f7789f68ede5838c6b0b779824d5114d2071fd6f6d9290cd4c0f.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 352, + 387, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 352, + 387, + 366 + ], + "spans": [ + { + "bbox": [ + 306, + 352, + 387, + 366 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 372, + 441, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 441, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 441, + 385 + ], + "type": "text", + "content": "4.1. Implementation Details." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 390, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 461 + ], + "type": "text", + "content": "Dataset. We use common datasets LLFF [41] and NeRF-Synthetic [42], with forward scenes {flower, fern, fortress, room} from LLFF and " + }, + { + "bbox": [ + 304, + 390, + 545, + 461 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 390, + 545, + 461 + ], + "type": "text", + "content": " scenes {lego, drums, chair} from NeRF-Synthetic. We further experiment on the Brandenburg Gate scene from NeRF-W dataset [39], with over 800 views of in-the-wild collected online." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "type": "text", + "content": "Training. On LLFF and NeRF-Synthetic, we adopt Plenoxels [83] as the NeRF backbone architecture for efficiency. On NeRF-W, we use the available PyTorch implementation [53]. The first stage of training is performed according to the standard recipes of those implementations. We then perform the second stage of steganography training for 55 epochs unless otherwise noted. On LLFF and NeRF-W, we downsize the training images by 4 times following common practice, and we use the original size on NeRF-Synthetic." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": "For hyper-parameters in Eq. (4), we set the weight of NeRF reconstruction error " + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_3 = 1" + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": " for all experiments. We set " + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_0 = 0.01, \\lambda_1 = 0.5, \\lambda_2 = 0.5" + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": " for all scenes on LLFF dataset, and " + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_0 = 0.1, \\lambda_1 = 1, \\lambda_2 = 1" + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": " for the scenes in NeRF-Synthetic, and " + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_0 = 0.05, \\lambda_1 = 1, \\lambda_2 = 1" + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": " for NeRF-W. For Eq. (3), we set " + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\alpha = 3" + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": " for all scenes. We run experiments on one NVIDIA A100 GPU." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Evaluation. We assume a typical authentication scenario shown in Alg.2. We consider the recovery quality of the embedded information, including the metrics of classification accuracy (Acc.) of the classifier, and the structural sim" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "444" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 261 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 261 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 261 + ], + "type": "image", + "image_path": "39a3a01589af864221c2ab7e726b6772ec2530d5c9bd57a39b03ddab82e30774.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 547, + 293 + ], + "type": "text", + "content": "Figure 4. Results on the NeRF-Synthetic dataset. Within each column, we show the StegaNeRF rendering, residual error from the initial NeRF rendering, and the recovered hidden image. We show the SSIM for the StegaNeRF renderings and the recovered hidden images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 398, + 287, + 515 + ], + "blocks": [ + { + "bbox": [ + 46, + 312, + 290, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 290, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 290, + 390 + ], + "type": "text", + "content": "Table 1. Quantitative results of StegaNeRF rendering and hidden information recovery. Standard NeRF is the initial NeRF " + }, + { + "bbox": [ + 46, + 312, + 290, + 390 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 312, + 290, + 390 + ], + "type": "text", + "content": " with standard training (upper-bound rendering performance). Prior 2D steganography fails after NeRF training while StegaNeRF successfully embeds and recovers hidden information with minimal impact on the rendering quality. Results are averaged over the selected LLFF and NeRF-Synthetic scenes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 398, + 287, + 515 + ], + "lines": [ + { + "bbox": [ + 49, + 398, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 49, + 398, + 287, + 515 + ], + "type": "table", + "html": "
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
Standard NeRF28.150.85750.137750.00N/A
LSB [9]28.130.85690.1452N/A0.0132
DeepStega [2]26.940.84320.1641N/A0.2098
StegaNeRF (Ours)28.130.85630.1460100.00.9730
Standard NeRF31.130.96060.031050.00N/A
LSB [9]31.120.96040.0310N/A0.0830
DeepStega [2]31.130.96060.0313N/A0.2440
StegaNeRF (Ours)30.960.95830.029099.720.9677
", + "image_path": "3159b1a50b982899fb8f35af09c185828f3d695bf7258d91be553d0322758564.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 536, + 288, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 288, + 597 + ], + "type": "text", + "content": "ality (SSIM) [74] of the hidden image recovered by the decoder. We evaluate the final NeRF renderings with PSNR, SSIM and LPIPS [90]. All the metrics are computed on the test set and averaged over all the scenes and embedded images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 604, + 244, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 244, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 244, + 618 + ], + "type": "text", + "content": "4.2. Case I: Embedding in a Single Scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 623, + 288, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 288, + 672 + ], + "type": "text", + "content": "We first explore the preliminary case of ownership identification on a specific NeRF scene. We select random images from ImageNet [27] as the hidden information to be injected to NeRF renderings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Failure of 2D Baseline Due to the lack of prior study on NeRF steganography, we consider a baseline from 2D image steganography by training NeRF from scratch with" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 389, + 545, + 469 + ], + "blocks": [ + { + "bbox": [ + 304, + 312, + 547, + 381 + ], + "lines": [ + { + "bbox": [ + 304, + 312, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 547, + 381 + ], + "type": "text", + "content": "Table 2. Quantitative results of NeRF rendering and hidden information recovery. We consider two conditions, embedding each scene with a common hidden image (One-for-All) or a scene-specific hidden image (One-for-All). We report the quality difference compared to the single-scene settings as " + }, + { + "bbox": [ + 304, + 312, + 547, + 381 + ], + "type": "inline_equation", + "content": "\\Delta_{SSIM}^{(10^{-2})}" + }, + { + "bbox": [ + 304, + 312, + 547, + 381 + ], + "type": "text", + "content": ". Results are averaged over the selected LLFF scenes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 389, + 545, + 469 + ], + "lines": [ + { + "bbox": [ + 308, + 389, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 389, + 545, + 469 + ], + "type": "table", + "html": "
SettingNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS ↓Δ(10-2)SSIM↑Acc. (%)↑SSIM↑Δ(10-2)SSIM↑
One-for-All24.990.80130.1786-0.10100.000.9860+0.19
One-for-One24.990.80160.1779-0.07100.000.9122-7.19
One-for-All27.900.85130.1236+0.01100.000.9844-0.76
One-for-One27.900.85150.1195+0.03100.000.9448-4.45
One-for-All30.270.84980.1289+0.02100.000.9430-5.42
One-for-One30.120.84800.1302-0.16100.000.9102-8.67
", + "image_path": "8b68c8e6da47b18c73cc15677a49f22252674ba923b88fbb7bd44da5922218b5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 547, + 725 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 725 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 725 + ], + "type": "text", + "content": "the watermarked images. We implement two off-the-shelf steganography methods including a traditional machine learning approach called Least Significant Bit (LSB [9]), and a deep learning pipeline as DeepStega [2]. An ideal case is that the embedded information can be recovered from the synthesized novel views, indicating the successful NeRF steganography. However, as can be seen in Fig. 3, the embedded information containing the hidden images cannot be recovered from the NeRF renderings. By analyzing the residual maps of training views (between GT training views and the watermarked) and novel views (between GT novel views and the actual rendering), we observe the subtle residuals to recover the hidden information are smoothed out in NeRF renderings, and similar failures occur on other datasets as well. Therefore, for the rest of our study, we mainly focus on analyzing our new framework that performs direct steganography on 3D NeRF." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "445" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 108, + 246 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 108, + 246 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 108, + 246 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 108, + 246 + ], + "type": "image", + "image_path": "d6c84e49f3bc15605767706ce87f049e7e32deb570f4eed16671ed82011e52b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 250, + 149, + 259 + ], + "lines": [ + { + "bbox": [ + 100, + 250, + 149, + 259 + ], + "spans": [ + { + "bbox": [ + 100, + 250, + 149, + 259 + ], + "type": "text", + "content": "(a) Per-Scene" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 72, + 165, + 246 + ], + "blocks": [ + { + "bbox": [ + 108, + 72, + 165, + 246 + ], + "lines": [ + { + "bbox": [ + 108, + 72, + 165, + 246 + ], + "spans": [ + { + "bbox": [ + 108, + 72, + 165, + 246 + ], + "type": "image", + "image_path": "befd6ae597219dc81b1e27ac75f189b933195b1162d5bf30874010a2636db60c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 545, + 293 + ], + "type": "text", + "content": "Figure 5. Results on three multi-scene settings. Within each column, we show the StegaNeRF rendering, residual error from the initial NeRF rendering, and recovered hidden image. We show the SSIM for the StegaNeRF renderings and the recovered hidden images." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 167, + 72, + 202, + 245 + ], + "blocks": [ + { + "bbox": [ + 167, + 72, + 202, + 245 + ], + "lines": [ + { + "bbox": [ + 167, + 72, + 202, + 245 + ], + "spans": [ + { + "bbox": [ + 167, + 72, + 202, + 245 + ], + "type": "image", + "image_path": "724703a416a23b70eef375d130184438f3bd0ba5aa8f779ac91b975f1e840cf9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 72, + 273, + 246 + ], + "blocks": [ + { + "bbox": [ + 215, + 72, + 273, + 246 + ], + "lines": [ + { + "bbox": [ + 215, + 72, + 273, + 246 + ], + "spans": [ + { + "bbox": [ + 215, + 72, + 273, + 246 + ], + "type": "image", + "image_path": "69a753bd6b77b0add7d695a2f3857e3ff481eb79efde28c14fad54228efcafd4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 249, + 345, + 259 + ], + "lines": [ + { + "bbox": [ + 242, + 249, + 345, + 259 + ], + "spans": [ + { + "bbox": [ + 242, + 249, + 345, + 259 + ], + "type": "text", + "content": "(b) Multi-Scene: One-for-All" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 274, + 72, + 332, + 246 + ], + "blocks": [ + { + "bbox": [ + 274, + 72, + 332, + 246 + ], + "lines": [ + { + "bbox": [ + 274, + 72, + 332, + 246 + ], + "spans": [ + { + "bbox": [ + 274, + 72, + 332, + 246 + ], + "type": "image", + "image_path": "08c10b53ec3c44fc6e5f96b9244d77e6432ce2929d6faa68b72d4b46ba75ba34.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 334, + 72, + 372, + 245 + ], + "blocks": [ + { + "bbox": [ + 334, + 72, + 372, + 245 + ], + "lines": [ + { + "bbox": [ + 334, + 72, + 372, + 245 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 372, + 245 + ], + "type": "image", + "image_path": "175cfe58e9400866ce5654d7f117139b85151c87270308225607d98fb9a223d8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 380, + 72, + 439, + 246 + ], + "blocks": [ + { + "bbox": [ + 380, + 72, + 439, + 246 + ], + "lines": [ + { + "bbox": [ + 380, + 72, + 439, + 246 + ], + "spans": [ + { + "bbox": [ + 380, + 72, + 439, + 246 + ], + "type": "image", + "image_path": "f331cae1e498cc8136eaf4476c507bae58d36a3c26bac8c217349b7d91e064ce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 249, + 511, + 259 + ], + "lines": [ + { + "bbox": [ + 404, + 249, + 511, + 259 + ], + "spans": [ + { + "bbox": [ + 404, + 249, + 511, + 259 + ], + "type": "text", + "content": "(c) Multi-Scene: One-for-One" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 498, + 246 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 498, + 246 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 498, + 246 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 498, + 246 + ], + "type": "image", + "image_path": "c3f6877407ef268b8eca057516f723311e7953bb0ff9658bae81d22c0d4dc5ef.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 500, + 72, + 536, + 245 + ], + "blocks": [ + { + "bbox": [ + 500, + 72, + 536, + 245 + ], + "lines": [ + { + "bbox": [ + 500, + 72, + 536, + 245 + ], + "spans": [ + { + "bbox": [ + 500, + 72, + 536, + 245 + ], + "type": "image", + "image_path": "2aa717d3e47396a8532de09b61ec7888b48cbd0022effaa387079d033e90d281.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 309, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 289, + 453 + ], + "type": "text", + "content": "Results Tab. 1 contains quantitative results on LLFF and NeRF-Synthetic scenes. While NeRF trained by 2D steganography methods hardly recovers the embedded information, StegaNeRF accurately recovers the hidden image with minimal impact on the rendering quality measured by PSNR. Fig. 4 provides qualitative results of StegaNeRF on three embedded images on the NeRF-Synthetic scenes. An interesting observation is the common regions where hidden information (high residual error) emerges in the renderings, e.g., the right rear side of lego and left handrail of chair. We also notice that, within each scene, these regions are persistent across multiple viewpoints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 468, + 255, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 255, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 255, + 481 + ], + "type": "text", + "content": "4.3. Case II: Embedding in Multiple Scenes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "text", + "content": "Settings We extend our steganography scheme to embed information within multiple scenes at once. Specially, we use three LLFF scenes " + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "inline_equation", + "content": "\\{flower, fern, fortress\\}" + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "text", + "content": " to test the two sub-settings of multi-scene embedding with: (1) One-for-All, a common hidden image and (2) One-for-One scene-specific hidden images. All scenes share the same decoder " + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "text", + "content": " and classifier " + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "inline_equation", + "content": "F_{\\psi}^{c}" + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "text", + "content": ". The difference between the two sub-settings is the number of hidden images that " + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "inline_equation", + "content": "F_{\\psi}" + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "inline_equation", + "content": "F_{\\psi}^{c}" + }, + { + "bbox": [ + 46, + 489, + 287, + 621 + ], + "type": "text", + "content": " need to identify and recover. We sample one scene for every training epoch, and due to the increased data amount, we increase training epochs until convergence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": "Results Tab. 2 provides quantitative results on the multi-scene setting. The performance drop compared to single-scene training is sometimes noticeable, but it is not surprising due to the inherent requirement of per-scene fitting for our NeRF framework. Fig. 5 presents visual comparisons of multi-scene steganography against single-scene scheme." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 308, + 518, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 518, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 518, + 322 + ], + "type": "text", + "content": "4.4. Case III: Embedding Multiple Identities." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 331, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 463 + ], + "type": "text", + "content": "Settings Constructing NeRF of large-scale cultural landmarks is a promising application, and community contributions are crucial to form the training images. Since every NeRF prediction is indebted to some particular training images, it would be meaningful to somehow identify the contributing users in the rendered output. Specifically, we present a proof-of-concept based on the following scenario: Given a collection of user-uploaded training images, our task is to apply StegaNeRF to ensure the final NeRF renderings hide subtle information about the relevant identities whose uploaded images help generate the current rendering." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": "To simulate this multi-user scenario in the public NeRF-W dataset, we randomly select " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": " anchor views with different viewing positions, and then find the " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": " nearest neighbour views to each anchor to form their respective clusters. We set " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "K = 20" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": " in experiments. We assume a common contributor identity for each cluster, and we want the NeRF rendering to contain customized hidden information about those " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": " contributors when we render within the spatial regions spanned by their own cluster. Thus, our classifier network " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "F_{\\phi}^{c}" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": " is modified to output " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": "-class cluster predictions and another class for identity outside of the " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": " clusters. Since the decoder should extract no information for views outside of those " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": " clusters to prevent false positives, we also compute " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}^{-}" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": " (1) and " + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dec}^{c}" + }, + { + "bbox": [ + 304, + 467, + 546, + 636 + ], + "type": "text", + "content": " (2) for those poses." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Results We employ the same network backbone as NeRF-W [66] to handle in-the-wild images with different time and lighting effects. Fig. 6 presents qualitative results of embedding multiple identities in a collaborative large-scale NeRF." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "446" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 542, + 223 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 542, + 223 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 542, + 223 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 542, + 223 + ], + "type": "image", + "image_path": "4483d0a9b5b2ae4bbaa21689583e9c4057846bf953d6cb41c605d43855bb3761.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 546, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 546, + 281 + ], + "type": "text", + "content": "Figure 6. Qualitative results of NeRF steganography on NeRF-W dataset. Within each block we show the StegaNeRF rendering, the initial NeRF rendering, the residual error between StegaNeRF and initial, and the recovered hidden image. At the bottom right, we provide the SSIM computed with their respective ground truth. \"Unmarked Identity\" denotes the remaining training views not related to any of the considered users, and the decoder does not recover meaningful information from renderings at those poses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 301, + 141, + 370 + ], + "blocks": [ + { + "bbox": [ + 50, + 301, + 141, + 370 + ], + "lines": [ + { + "bbox": [ + 50, + 301, + 141, + 370 + ], + "spans": [ + { + "bbox": [ + 50, + 301, + 141, + 370 + ], + "type": "image", + "image_path": "6115d281af9e8a99be15960170776e7ca09f2a999dcbcd896e05df93fa141114.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 342, + 184, + 349 + ], + "lines": [ + { + "bbox": [ + 149, + 342, + 184, + 349 + ], + "spans": [ + { + "bbox": [ + 149, + 342, + 184, + 349 + ], + "type": "text", + "content": "SSIM=0.9709" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 374, + 141, + 443 + ], + "blocks": [ + { + "bbox": [ + 50, + 374, + 141, + 443 + ], + "lines": [ + { + "bbox": [ + 50, + 374, + 141, + 443 + ], + "spans": [ + { + "bbox": [ + 50, + 374, + 141, + 443 + ], + "type": "image", + "image_path": "7d5f6e9d56b3afd50151fcc36ad4dcec880d1a82313045d06d4edd2f8e2f9035.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 415, + 183, + 421 + ], + "lines": [ + { + "bbox": [ + 149, + 415, + 183, + 421 + ], + "spans": [ + { + "bbox": [ + 149, + 415, + 183, + 421 + ], + "type": "inline_equation", + "content": "\\mathrm{SSIM} = 0.9688" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 433, + 231, + 440 + ], + "lines": [ + { + "bbox": [ + 153, + 433, + 231, + 440 + ], + "spans": [ + { + "bbox": [ + 153, + 433, + 231, + 440 + ], + "type": "text", + "content": "`The Owner Of These Horns'" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 457, + 287, + 513 + ], + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 513 + ], + "type": "text", + "content": "Figure 7. Multi-modal hidden information. We show StegaNeRF rendering, recovered hidden images, audio (waveform and spectrum), and text. Metrics for hidden information in each modality are labeled respectively. For the StegaNeRF rendering, we report both SSIM and the relative change from initial NeRF renderings." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 146, + 301, + 187, + 339 + ], + "blocks": [ + { + "bbox": [ + 146, + 301, + 187, + 339 + ], + "lines": [ + { + "bbox": [ + 146, + 301, + 187, + 339 + ], + "spans": [ + { + "bbox": [ + 146, + 301, + 187, + 339 + ], + "type": "image", + "image_path": "eedc427949d4424f0b28fc773a892c6d11e3e4f4b53c383f4d44c06db78ba2ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 189, + 301, + 237, + 338 + ], + "blocks": [ + { + "bbox": [ + 189, + 301, + 237, + 338 + ], + "lines": [ + { + "bbox": [ + 189, + 301, + 237, + 338 + ], + "spans": [ + { + "bbox": [ + 189, + 301, + 237, + 338 + ], + "type": "image", + "image_path": "2541dbb6af9e408ec055d3b7aa36a413e49dd16066289336cdf21de842b81e68.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 342, + 256, + 348 + ], + "lines": [ + { + "bbox": [ + 223, + 342, + 256, + 348 + ], + "spans": [ + { + "bbox": [ + 223, + 342, + 256, + 348 + ], + "type": "text", + "content": "PSNR=76.22" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 240, + 303, + 282, + 338 + ], + "blocks": [ + { + "bbox": [ + 240, + 303, + 282, + 338 + ], + "lines": [ + { + "bbox": [ + 240, + 303, + 282, + 338 + ], + "spans": [ + { + "bbox": [ + 240, + 303, + 282, + 338 + ], + "type": "image", + "image_path": "02d720779d3fa8852ba90d4ba95f2f643b5a09b5111a41a74c02efec62665ad8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 342, + 272, + 348 + ], + "lines": [ + { + "bbox": [ + 239, + 342, + 272, + 348 + ], + "spans": [ + { + "bbox": [ + 239, + 342, + 272, + 348 + ], + "type": "inline_equation", + "content": "= 76.22" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 154, + 356, + 231, + 364 + ], + "blocks": [ + { + "bbox": [ + 154, + 356, + 231, + 364 + ], + "lines": [ + { + "bbox": [ + 154, + 356, + 231, + 364 + ], + "spans": [ + { + "bbox": [ + 154, + 356, + 231, + 364 + ], + "type": "image", + "image_path": "a241df4d59a9e60a03f5aee43ecadc2716e334fb69ad0b10dfb1489752cbc058.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 356, + 272, + 363 + ], + "lines": [ + { + "bbox": [ + 239, + 356, + 272, + 363 + ], + "spans": [ + { + "bbox": [ + 239, + 356, + 272, + 363 + ], + "type": "text", + "content": "Acc. " + }, + { + "bbox": [ + 239, + 356, + 272, + 363 + ], + "type": "inline_equation", + "content": "= 100.0\\%" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 150, + 380, + 186, + 411 + ], + "blocks": [ + { + "bbox": [ + 151, + 371, + 187, + 378 + ], + "lines": [ + { + "bbox": [ + 151, + 371, + 187, + 378 + ], + "spans": [ + { + "bbox": [ + 151, + 371, + 187, + 378 + ], + "type": "text", + "content": "Hidden Image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 150, + 380, + 186, + 411 + ], + "lines": [ + { + "bbox": [ + 150, + 380, + 186, + 411 + ], + "spans": [ + { + "bbox": [ + 150, + 380, + 186, + 411 + ], + "type": "image", + "image_path": "c2293670317ccd800ee78f7c0eac6594955cd76bd38ba75102b9ae3ced8a275c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 192, + 380, + 237, + 412 + ], + "blocks": [ + { + "bbox": [ + 222, + 371, + 258, + 377 + ], + "lines": [ + { + "bbox": [ + 222, + 371, + 258, + 377 + ], + "spans": [ + { + "bbox": [ + 222, + 371, + 258, + 377 + ], + "type": "text", + "content": "Hidden Audio" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 380, + 237, + 412 + ], + "lines": [ + { + "bbox": [ + 192, + 380, + 237, + 412 + ], + "spans": [ + { + "bbox": [ + 192, + 380, + 237, + 412 + ], + "type": "image", + "image_path": "08db40a38819a1606a7343bc42a8e6965c6f8254d53a8e17f139afba53b17090.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 415, + 257, + 422 + ], + "lines": [ + { + "bbox": [ + 225, + 415, + 257, + 422 + ], + "spans": [ + { + "bbox": [ + 225, + 415, + 257, + 422 + ], + "type": "text", + "content": "PSNR=62.84" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 180, + 424, + 211, + 430 + ], + "lines": [ + { + "bbox": [ + 180, + 424, + 211, + 430 + ], + "spans": [ + { + "bbox": [ + 180, + 424, + 211, + 430 + ], + "type": "text", + "content": "Hidden Text" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 241, + 380, + 285, + 412 + ], + "blocks": [ + { + "bbox": [ + 241, + 380, + 285, + 412 + ], + "lines": [ + { + "bbox": [ + 241, + 380, + 285, + 412 + ], + "spans": [ + { + "bbox": [ + 241, + 380, + 285, + 412 + ], + "type": "image", + "image_path": "448a38526e8ec86cf3093219f717c7da6ff576251a316449e63b928245c351b3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 433, + 276, + 440 + ], + "lines": [ + { + "bbox": [ + 244, + 433, + 276, + 440 + ], + "spans": [ + { + "bbox": [ + 244, + 433, + 276, + 440 + ], + "type": "text", + "content": "Acc. " + }, + { + "bbox": [ + 244, + 433, + 276, + 440 + ], + "type": "inline_equation", + "content": "= 100.0\\%" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 242, + 415, + 257, + 422 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 415, + 257, + 422 + ], + "spans": [ + { + "bbox": [ + 242, + 415, + 257, + 422 + ], + "type": "text", + "content": "62.84" + } + ] + } + ], + "index": 19, + "type": "text" + }, + { + "bbox": [ + 47, + 529, + 287, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 287, + 541 + ], + "type": "text", + "content": "4.5. Case IV: Embedding Multi-Modal Information." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 547, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 287, + 595 + ], + "type": "text", + "content": "Settings We further show the potential of StegaNeRF in embedding multi-modal hidden information, such as images, audio, and text. We modify the decoder network to build a modal-specific decoder for each modality." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 602, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 664 + ], + "type": "text", + "content": "Results Fig. 7 shows recovered multi-modal embedded signals in the trex and horns scenes from LLFF. Evidently, the StegaNeRF framework can easily extend to embed multi-modal information with high recovery performance without scarifying the rendering quality." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 47, + 671, + 148, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 148, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 148, + 682 + ], + "type": "text", + "content": "4.6. Ablation Studies" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "The effect of removing each component of StegaNeRF is presented in Tab. 3. Standard NeRF uses the initial NeRF" + } + ] + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 308, + 331, + 545, + 423 + ], + "blocks": [ + { + "bbox": [ + 306, + 300, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 322 + ], + "type": "text", + "content": "Table 3. Ablation study of different components of StegaNeRF. Results are averaged on the selected LLFF scenes." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 331, + 545, + 423 + ], + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 423 + ], + "type": "table", + "html": "
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
StegaNeRF28.130.85680.1430100.00.9486
No Classifier26.850.80770.2417N/A0.5629
No Classifier Guided27.120.82390.2073100.00.6716
No Gradient Masking27.860.83750.1710100.00.8822
No Soft Masking28.050.85580.152694.440.8751
Standard NeRF28.150.85750.137750.00N/A
", + "image_path": "e30ed4eedca9a7442494c7b930a04e9fe048cbd3b6d9ac5f6a756206fc45b446.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "content": "rendering. No Classifier removes the classifier " + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "inline_equation", + "content": "F_{\\psi}^{c}" + }, + { + "bbox": [ + 304, + 454, + 545, + 586 + ], + "type": "text", + "content": ", while No Classifier Guided retains the classification task (hence impact on NeRF rendering) but does not condition the decoder on classifier output. No Gradient Masking removes the proposed masking strategy (Sec. 3.3), and No Soft Masking uses the binary mask [29, 31] with a threshold of 0.5. Our StegaNeRF makes a good balance between the rendering quality and the decoding accuracy. Fig. 8 shows the visual impact of removing each component. It appears that when any component is removed, the performance drops accordingly, revealing the effectiveness of our design." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Fig. 9 shows the impact of varying gradient masking strategies. We can see a different degree of affected by ratio variation on hidden recovery and rendering quality. Fig. 10 reports the performance of StegaNeRF against the common perturbations including JPEG compression and Gaussian noise. The lines show the average accuracies across selected scenes and the shaded regions indicate the range of 0.5 standard deviation. It appears that the hidden recovery of StegaNeRF is robust to various JPEG compression ratios and Gaussian blur degradation." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "447" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 542, + 167 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 542, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 542, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 542, + 167 + ], + "type": "image", + "image_path": "77d0c1748fe17cc0f6e4a4972819f344c08065aed7646d5e7cb7b23d4f1b7448.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 170, + 118, + 179 + ], + "lines": [ + { + "bbox": [ + 60, + 170, + 118, + 179 + ], + "spans": [ + { + "bbox": [ + 60, + 170, + 118, + 179 + ], + "type": "text", + "content": "(a) Ground Truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 170, + 197, + 179 + ], + "lines": [ + { + "bbox": [ + 140, + 170, + 197, + 179 + ], + "spans": [ + { + "bbox": [ + 140, + 170, + 197, + 179 + ], + "type": "text", + "content": "(b) No Classifier" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 170, + 290, + 179 + ], + "lines": [ + { + "bbox": [ + 209, + 170, + 290, + 179 + ], + "spans": [ + { + "bbox": [ + 209, + 170, + 290, + 179 + ], + "type": "text", + "content": "(c) No Classifier Guided" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 297, + 170, + 378, + 179 + ], + "lines": [ + { + "bbox": [ + 297, + 170, + 378, + 179 + ], + "spans": [ + { + "bbox": [ + 297, + 170, + 378, + 179 + ], + "type": "text", + "content": "(d) No Gradient Masking" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 385, + 170, + 454, + 179 + ], + "lines": [ + { + "bbox": [ + 385, + 170, + 454, + 179 + ], + "spans": [ + { + "bbox": [ + 385, + 170, + 454, + 179 + ], + "type": "text", + "content": "(e) No Soft Masking" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 476, + 170, + 525, + 179 + ], + "lines": [ + { + "bbox": [ + 476, + 170, + 525, + 179 + ], + "spans": [ + { + "bbox": [ + 476, + 170, + 525, + 179 + ], + "type": "text", + "content": "(f) StegaNeRF" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 243, + 160, + 323 + ], + "blocks": [ + { + "bbox": [ + 52, + 243, + 160, + 323 + ], + "lines": [ + { + "bbox": [ + 52, + 243, + 160, + 323 + ], + "spans": [ + { + "bbox": [ + 52, + 243, + 160, + 323 + ], + "type": "image", + "image_path": "583281c660dc64ea12b2dbccb5f16738971b1b9143cdac21cac82a8d98033ca4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 326, + 132, + 335 + ], + "lines": [ + { + "bbox": [ + 87, + 326, + 132, + 335 + ], + "spans": [ + { + "bbox": [ + 87, + 326, + 132, + 335 + ], + "type": "text", + "content": "(a) Varying " + }, + { + "bbox": [ + 87, + 326, + 132, + 335 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 345, + 287, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 345, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 287, + 389 + ], + "type": "text", + "content": "Figure 9. Analysis of gradient masking. (a) Varying " + }, + { + "bbox": [ + 46, + 345, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 345, + 287, + 389 + ], + "type": "text", + "content": " from Eq. (3). (b) Varying portions of masked out weights from the gradient updates at steganography learning. We provide the SSIM of rendered views (blue) and recovered hidden images (green)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 175, + 243, + 279, + 323 + ], + "blocks": [ + { + "bbox": [ + 175, + 243, + 279, + 323 + ], + "lines": [ + { + "bbox": [ + 175, + 243, + 279, + 323 + ], + "spans": [ + { + "bbox": [ + 175, + 243, + 279, + 323 + ], + "type": "image", + "image_path": "0d787ff1f65773bcd571152f75a43b24c34cd1bec45c942feda514970b7ea221.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 326, + 273, + 335 + ], + "lines": [ + { + "bbox": [ + 189, + 326, + 273, + 335 + ], + "spans": [ + { + "bbox": [ + 189, + 326, + 273, + 335 + ], + "type": "text", + "content": "(b) Varying masking ratio" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 56, + 414, + 160, + 494 + ], + "blocks": [ + { + "bbox": [ + 56, + 414, + 160, + 494 + ], + "lines": [ + { + "bbox": [ + 56, + 414, + 160, + 494 + ], + "spans": [ + { + "bbox": [ + 56, + 414, + 160, + 494 + ], + "type": "image", + "image_path": "11847d5d8073c01050251fecacb27c3f6503f1f979df108f36ad6e8dcee6d404.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 496, + 167, + 505 + ], + "lines": [ + { + "bbox": [ + 54, + 496, + 167, + 505 + ], + "spans": [ + { + "bbox": [ + 54, + 496, + 167, + 505 + ], + "type": "text", + "content": "(a) Varying JPEG compression rate" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 515, + 287, + 548 + ], + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 548 + ], + "type": "text", + "content": "Figure 10. Analysis of robustness over (a) JPEG compression and (b) Gaussian blur. We provide the SSIM of rendered views (blue) and recovered hidden images (green)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 175, + 414, + 281, + 494 + ], + "blocks": [ + { + "bbox": [ + 175, + 414, + 281, + 494 + ], + "lines": [ + { + "bbox": [ + 175, + 414, + 281, + 494 + ], + "spans": [ + { + "bbox": [ + 175, + 414, + 281, + 494 + ], + "type": "image", + "image_path": "c17513acd083b98e6cc55af05ab707a660cc625073f8f3e1f94cffd0c3a1bfd4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 496, + 281, + 505 + ], + "lines": [ + { + "bbox": [ + 184, + 496, + 281, + 505 + ], + "spans": [ + { + "bbox": [ + 184, + 496, + 281, + 505 + ], + "type": "text", + "content": "(b) Varying Gaussian blur std." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 361, + 545, + 416 + ], + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 416 + ], + "type": "text", + "content": "Figure 11. Results using a universal decoder to embed (a) multiple images for the flower scene and (b) a common image into different scenes from the LLFF dataset. The successful application under these two setups suggests that StegaNeRF can generalize across different scenes, user identities, and hidden images." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 572, + 151, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 151, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 151, + 583 + ], + "type": "text", + "content": "5. More Discussions" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "How useful is steganography for NeRF? Although NeRF-based 3D content has yet to become mainstream, we believe it will play a major future role not only for social platforms, but also for 3D vision research and applications. As people upload their personal NeRF models online for viewing purposes, NeRF steganography for ownership identification is apparently an important feature. Moreover, future 3D vision research will likely demand large-scale NeRF datasets of real-world images, and in this context, NeRF steganography can be a crucial tool for responsible" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 309, + 232, + 429, + 328 + ], + "blocks": [ + { + "bbox": [ + 46, + 192, + 545, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 545, + 214 + ], + "type": "text", + "content": "Figure 8. Impact on visual quality when changing different components of the proposed StegaNeRF framework as in Tab. 3. We show the SSIM for the renderings and the recovered hidden images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 232, + 429, + 328 + ], + "lines": [ + { + "bbox": [ + 309, + 232, + 429, + 328 + ], + "spans": [ + { + "bbox": [ + 309, + 232, + 429, + 328 + ], + "type": "image", + "image_path": "496dc80a0391c8141c4ec35a8aefd20db91e807e929c522b0a9e2bc3868f8838.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 342, + 426, + 352 + ], + "lines": [ + { + "bbox": [ + 309, + 342, + 426, + 352 + ], + "spans": [ + { + "bbox": [ + 309, + 342, + 426, + 352 + ], + "type": "text", + "content": "(a) Varying #watermark per scene" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 431, + 232, + 552, + 328 + ], + "blocks": [ + { + "bbox": [ + 431, + 232, + 552, + 328 + ], + "lines": [ + { + "bbox": [ + 431, + 232, + 552, + 328 + ], + "spans": [ + { + "bbox": [ + 431, + 232, + 552, + 328 + ], + "type": "image", + "image_path": "a453434f03bfc17b213a281f2a129ffa0e0c7fdc33b91a7ad23e2eb97f953732.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 342, + 545, + 352 + ], + "lines": [ + { + "bbox": [ + 435, + 342, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 435, + 342, + 545, + 352 + ], + "type": "text", + "content": "(b) Varying #scene per watermark" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 308, + 459, + 545, + 535 + ], + "blocks": [ + { + "bbox": [ + 306, + 428, + 545, + 450 + ], + "lines": [ + { + "bbox": [ + 306, + 428, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 545, + 450 + ], + "type": "text", + "content": "Table 4. The further empirical study on StegaNeRF. Results are averaged on the selected LLFF scenes." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 459, + 545, + 535 + ], + "lines": [ + { + "bbox": [ + 308, + 459, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 545, + 535 + ], + "type": "table", + "html": "
MethodNeRF RenderingHidden Recovery
PSNR↑SSIM↑LPIPS↓Acc. (%)↑SSIM↑
StegaNeRF28.130.85680.1430100.00.9486
No Empty Sampling28.120.85570.1466100.00.9364
NeRF Frozen28.150.85750.137750.000.3975
Standard NeRF28.150.85750.137750.00N/A
", + "image_path": "25681a2bf618e4efda29e9dcad8c81ee20f70b9dc5bddd3864e537a23790e152.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 551, + 524, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 524, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 524, + 563 + ], + "type": "text", + "content": "and ethical uses of training data and deployed models." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Hidden or Memorized? StegaNeRF follows the standard setup of NeRF which prescribes per-scene training. Therefore, it may be reasonable to wonder whether the decoder is trained to \"memorize\" the hidden image without it really being injected into NeRF. To address this concern, we perform an experiment where the NeRF weights are frozen after the first stage and only update the decoder of StegaNeRF. If the decoder actually memorizes the image by itself, then we should expect good recovery performance under this setup. However, as shown in Tab. 4, with NeRF Frozen, we observe a significant performance degradation. Therefore, we can verify that the decoder alone cannot trivially memorize" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "448" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "the hidden image, and the information is actually hidden in the NeRF renderings. Furthermore, we extend StegaNeRF to generalize across different scenes and hidden images while sharing the same universal decoder. Fig. 11a shows the rendering and decoding quality of embedding the different numbers of hidden images to a scene (flower). Fig. 11b shows the performance of embedding a common image into a varying number of scenes from the LLFF dataset. While the performance slightly decreases in the generalized setting, it still remains at a comparable level." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 198, + 289, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 289, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 289, + 307 + ], + "type": "text", + "content": "Why not just share the rendered images instead of NeRF weights? Directly sharing the NeRF model gives end users the freedom to render at continuous viewpoints, which fundamentally differs from the traditional setting centered around discrete 2D image frames. It would not be realistic to assume people will stick with sharing 2D images instead of 3D weights, and therefore, it is necessary for the research community to see beyond 2D steganography and explore 3D NeRF steganography." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "text", + "content": "Robustness: As shown in Fig. 12, in addition to using a resolution of " + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "text", + "content": " for hidden images in most of our experiments, we also tested resolutions ranging from " + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "text", + "content": ". We found that increasing the size of embedded images had no noticeable impact on the performance of StegaNeRF, with only a " + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "text", + "content": " drop in rendering PSNR and a " + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 312, + 289, + 504 + ], + "type": "text", + "content": " reduction in hidden SSIM. We also demonstrated the robustness of StegaNeRF to different sampling approaches during rendering, such as avoiding sampling in empty areas [43, 48]. Instead of considering all the points on a ray during rendering, we disregard the contribution of points to the integration along the entire ray if their density falls below a threshold, like 0.001. As shown in Tab. 4, replacing the conventional rendering strategy with No Empty Sampling has minimal effect on the performance of StegaNeRF, revealing our robustness to different rendering fashions." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 47, + 511, + 165, + 583 + ], + "blocks": [ + { + "bbox": [ + 47, + 511, + 165, + 583 + ], + "lines": [ + { + "bbox": [ + 47, + 511, + 165, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 165, + 583 + ], + "type": "image", + "image_path": "bd9ca95fb354943a142ca30d30f3b387482ba8c3fc3057cc4465b710876f9dc1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 587, + 287, + 610 + ], + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 610 + ], + "type": "text", + "content": "Figure 12. Ablation on the resolutions of embedded images by the average results over the selected scenes from the LLFF dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 167, + 511, + 286, + 583 + ], + "blocks": [ + { + "bbox": [ + 167, + 511, + 286, + 583 + ], + "lines": [ + { + "bbox": [ + 167, + 511, + 286, + 583 + ], + "spans": [ + { + "bbox": [ + 167, + 511, + 286, + 583 + ], + "type": "image", + "image_path": "6c2503ce199e0e20cbfe939a3d93590a725f4789d6d3c5f848b9bce2a363ad3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 713 + ], + "type": "text", + "content": "Limitations: The current implementation is limited by the computational efficiency of NeRF, since the steganographic training stage requires performing gradient descent updates through the NeRF model. Another limitation is that, the recovery accuracy of the injected hidden information varies across different 3D scenes. More systematic analysis is required to understand the underlying factors that make some scenes easier to hide information than others." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 379, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 379, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 379, + 83 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 547, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 271 + ], + "type": "text", + "content": "With the progression of NeRF-based content, now serving as a pivotal element in diverse 3D applications, it is incumbent upon the research community to proactively tackle the imminent requirement to embed information within NeRF models as they are exchanged and disseminated across online platforms. This study introduces StegaNeRF, a pioneering framework for the automated embedding of steganographic information within NeRF renderings. We outline the essential components essential for achieving this functionality, unveiling potential challenges and revealing insights that can be instrumental for subsequent developers. By shedding light on the prospect of ownership identification within the realm of NeRF, this paper underscores the need for increased attention and concerted efforts toward addressing related challenges and avenues." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 281, + 365, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 365, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 365, + 293 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 300, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 310, + 300, + 547, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 300, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 310, + 300, + 547, + 356 + ], + "type": "text", + "content": "[1] Benjamin Attal, Jia-Bin Huang, Michael Zollhoefer, Johannes Kopf, and Changil Kim. Learning Neural Light Fields With Ray-Space Embedding Networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19787-19797, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 357, + 547, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 357, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 311, + 357, + 547, + 389 + ], + "type": "text", + "content": "[2] Shumeet Baluja. Hiding Images in Plain Sight: Deep Steganography. Advances in Neural Information Processing Systems, 30, 2017. 1, 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 390, + 547, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 390, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 311, + 390, + 547, + 423 + ], + "type": "text", + "content": "[3] Shumeet Baluja. Hiding Images Within Images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1685-1697, 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 424, + 547, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 424, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 311, + 424, + 547, + 490 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-Nerf: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 491, + 547, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 491, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 311, + 491, + 547, + 545 + ], + "type": "text", + "content": "[5] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5460–5469, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 546, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 546, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 311, + 546, + 547, + 578 + ], + "type": "text", + "content": "[6] Oliver Benedens. Geometry-Based Watermarking of 3D Models. IEEE Computer Graphics and Applications, 19:46-55, 1999. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 580, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 580, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 311, + 580, + 547, + 624 + ], + "type": "text", + "content": "[7] Ning Bi, Qiyu Sun, Daren Huang, Zhihua Yang, and Jiwu Huang. Robust Image Watermarking Based on Multiband Wavelets and Empirical Mode Decomposition. IEEE Transactions on Image Processing, 16(8):1956-1966, 2007. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 624, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 624, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 311, + 624, + 547, + 669 + ], + "type": "text", + "content": "[8] Rajarathnam Chandramouli, Mehdi Kharrazi, and Nasir Memon. Image Steganography and Steganalysis: Concepts and Practice. In International Workshop on Digital Watermarking, pages 35-49. Springer, Springer, 2003. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 670, + 547, + 713 + ], + "type": "text", + "content": "[9] Chin-Chen Chang, Ju-Yuan Hsiao, and Chi-Shiang Chan. Finding Optimal Least-Significant-Bit Substitution in Image Hiding by Dynamic Programming Strategy. Pattern Recognition, 36(7):1583-1595, 2003. 2, 3, 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "449" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "text", + "content": "[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast Generalizable Radiance Field Reconstruction From Multi-View Stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124–14133, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 174 + ], + "type": "text", + "content": "[11] Di Chen, Yu Liu, Lianghua Huang, Bin Wang, and Pan Pan. GeoAug: Data Augmentation for Few-Shot NeRF With Geometry Constraints. In European Conference on Computer Vision, pages 322-337. Springer, Springer, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 207 + ], + "type": "text", + "content": "[12] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital Watermarking and Steganography. Morgan Kaufmann, 2007. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "text", + "content": "[13] Wenxue Cui, Shaohui Liu, Feng Jiang, Yongliang Liu, and Debin Zhao. Multi-Stage Residual Hiding for Image-Into-Audio Steganography. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2832-2836, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[14] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. In European Conference on Computer Vision, pages 636-654. Springer, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "text", + "content": "[15] Brandon Yushan Feng and Amitabh Varshney. SIGNET: Efficient Neural Representation for Light Fields. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14204-14213, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 409 + ], + "type": "text", + "content": "[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenumxels: Radiance Fields Without Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501-5510, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "text", + "content": "[17] Jessica Fridrich, Miroslav Goljan, and Rui Du. Detecting LSB Steganography in Color, and Gray-Scale Images. IEEE Multimedia, 8(4):22-28, 2001. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 499 + ], + "type": "text", + "content": "[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic Neural Radiance Fields for Monocular 4D Facial Avatar Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "text", + "content": "[19] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 621 + ], + "type": "text", + "content": "[20] Margarita Geleta, Cristina Punti, Kevin McGuinness, Jordi Pons, Cristian Canton, and Xavier Giro-i Nieto. PixInWav: Residual Steganography for Hiding Pixels in Audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2485-2489, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 657 + ], + "type": "text", + "content": "[21] Jamie Hayes and George Danezis. Generating Steganographic Images Via Adversarial Training. Advances in Neural Information Processing Systems, 30, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[22] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[23] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. EfficientNeRF Efficient Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12902-12911, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[24] Xudong Huang, Wei Li, Jie Hu, Hanting Chen, and Yunhe Wang. Refsr-nerf: Towards high fidelity and super resolution view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8244–8253, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[25] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. HiNet: Deep Image Hiding by Invertible Network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4733-4742, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "type": "text", + "content": "[26] Gary C Kessler and Chet Hosmer. An Overview of Steganography. Advances in Computers, 83:51-107, 2011. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "type": "text", + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet Classification With Deep Convolutional Neural Networks. Communications of the ACM, 60(6):84-90, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "text", + "content": "[28] Jaeho Lee, Jihoon Tack, Namhoon Lee, and Jinwoo Shin. Meta-Learning Sparse Implicit Neural Representations. Advances in Neural Information Processing Systems, 34:11769-11780, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 322, + 545, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 366 + ], + "type": "text", + "content": "[29] Chenxin Li, Mingbao Lin, Zhiyuan Ding, Nie Lin, Yihong Zhuang, Yue Huang, Xinghao Ding, and Liujuan Cao. Knowledge condensation distillation. In European Conference on Computer Vision, pages 19-35. Springer, 2022. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 368, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 368, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 368, + 545, + 422 + ], + "type": "text", + "content": "[30] Chenxin Li, Wenao Ma, Liyan Sun, Xinghao Ding, Yue Huang, Guisheng Wang, and Yizhou Yu. Hierarchical deep network with uncertainty-aware semi-supervised learning for vessel segmentation. Neural Computing and Applications, pages 1-14, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 423, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 545, + 487 + ], + "type": "text", + "content": "[31] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 61-65. IEEE, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 490, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 555 + ], + "type": "text", + "content": "[32] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "type": "text", + "content": "[33] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Y. Wang, and Shenlong Wang. NeurMiPs: Neural Mixture of Planar Experts for View Synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15681-15691, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "text", + "content": "[34] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural Sparse Voxel Fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[35] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason M. Saragih. Mixture of Volumetric Primitives for Efficient Neural Rendering. ACM Transactions on Graphics (TOG), 40:1 - 13, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "450" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[36] Shao-Ping Lu, Rong Wang, Tao Zhong, and Paul L Rosin. Large-Capacity Image Steganography Based on Invertible Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10816-10825, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 226, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 226, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 226, + 139 + ], + "type": "text", + "content": "[37] Luma AI. https://lumalabs.ai/.2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 183 + ], + "type": "text", + "content": "[38] Xiyang Luo, Yinxiao Li, Huiwen Chang, Ce Liu, Peyman Milanfar, and Feng Yang. Dvmark: A Deep Multiscale Framework for Video Watermarking. ArXiv Preprint ArXiv:2104.12734, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 184, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 184, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 49, + 184, + 287, + 249 + ], + "type": "text", + "content": "[39] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210–7219, 2021. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 251, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 251, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 251, + 287, + 304 + ], + "type": "text", + "content": "[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy Networks: Learning 3d Reconstruction in Function Space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 305, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 305, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 49, + 305, + 287, + 360 + ], + "type": "text", + "content": "[41] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local Light Field Fusion: Practical View Synthesis With Prescriptive Sampling Guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 361, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 287, + 414 + ], + "type": "text", + "content": "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing Scenes As Neural Radiance Fields for View Synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 49, + 415, + 287, + 459 + ], + "type": "text", + "content": "[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant Neural Graphics Primitives With a Multiresolution Hash Encoding. ACM Transactions on Graphics (TOG), 41:1-15, 2022. 1, 2, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 460, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 287, + 525 + ], + "type": "text", + "content": "[44] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing Neural Radiance Fields for View Synthesis From Sparse Inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 526, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 526, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 49, + 526, + 287, + 590 + ], + "type": "text", + "content": "[45] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning Continuous Signed Distance Functions for Shape Representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 592, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 592, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 49, + 592, + 287, + 646 + ], + "type": "text", + "content": "[46] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 647, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 691 + ], + "type": "text", + "content": "[47] Tomás Pevny, Tomás Filler, and Patrick Bas. Using High-Dimensional Image Models to Perform Highly Undetectable Steganography. In International Workshop on Information Hiding, pages 161-177. Springer, Springer, 2010. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[48] Martin Piala and Ronald Clark. Terminerf: Ray termination prediction for efficient neural rendering. In 2021 Interna" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "tional Conference on 3D Vision (3DV), pages 1106-1114. IEEE, 2021.9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 467, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 467, + 106 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 467, + 106 + ], + "type": "text", + "content": "[49] Polycam. https://poly.cam/.2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[50] Emil Praun, Hugues Hoppe, and Adam Finkelstein. Robust Mesh Watermarking. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, pages 49-56, 1999. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "type": "text", + "content": "[51] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 250 + ], + "type": "text", + "content": "[52] Yinlong Qian, Jing Dong, Wei Wang, and Tieniu Tan. Deep Learning for Steganalysis Via Convolutional Neural Networks. In Media Watermarking, Security, and Forensics 2015, volume 9409, pages 171-180. SPIE, SPIE, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 251, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 282 + ], + "type": "text", + "content": "[53] Chen Quei-An. Nerf-pl: A Pytorch-Lighting Implementation of NeRF. https://github.com/kwea123/nerf_pl/.4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "type": "text", + "content": "[54] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding Up Neural Radiance Fields With Thousands of Tiny Mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[55] Mennatallah M Sadek, Amal S Khalifa, and Mostafa GM Mostafa. Video Steganography: A Comprehensive Review. Multimedia Tools and Applications, 74(17):7063-7094, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "text", + "content": "[56] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-Aware Image Synthesis With Sparse Voxel Grids. ArXiv Preprint ArXiv:2206.07695, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "text", + "content": "[57] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[58] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit Neural Representations With Periodic Activation Functions. Advances in Neural Information Processing Systems, 33, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "text", + "content": "[59] Vincent Sitzmann, Semon Rezchikov, William T. Freeman, Joshua B. Tenenbaum, and Frédo Durand. Light Field Networks: Neural Scene Representations With Single-Evaluation Rendering. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "text", + "content": "[60] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "text", + "content": "[61] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct Voxel Grid Optimization: Super-Fast Convergence for Radiance Fields Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[62] Abdelfatah A Tamimi, Ayman M Abdalla, and Omaima Al-Allaf. Hiding an Image Inside Another Image Using Variable-Rate Steganography. International Journal of Advanced Computer Science and Applications (IJACSA), 4(10), 2013. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "451" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[63] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-Nerf: Scalable Large Scene Neural View Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248-8258, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 185 + ], + "type": "text", + "content": "[64] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible Hyperlinks in Physical Photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2126, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 252 + ], + "type": "text", + "content": "[65] Matthew Tancik, Pratul P Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T Barron, and Ren Ng. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. Advances in Neural Information Processing Systems, 33, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 253, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 297 + ], + "type": "text", + "content": "[66] Weixuan Tang, Shunquan Tan, Bin Li, and Jiwu Huang. Automatic Steganographic Distortion Learning Using a Generative Adversarial Network. IEEE Signal Processing Letters, 24(10):1547-1551, 2017. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 363 + ], + "type": "text", + "content": "[67] Jeroen JG Van Merrienboer, Liesbeth Kester, and Fred Paas. Teaching complex rather than simple tasks: Balancing intrinsic and germane load to enhance transfer of learning. Applied Cognitive Psychology: The Official Journal of the Society for Applied Research in Memory and Cognition, 20(3):343-352, 2006. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "text", + "content": "[68] Mukund Varma, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that nef needs? In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 465 + ], + "type": "text", + "content": "[69] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and S. Tulyakov. R2L: Distilling Neural Radiance Field to Neural Light Field for Efficient Novel View Synthesis. In European Conference on Computer Vision. Springer, Springer, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "text", + "content": "[70] Huaiqing Wang and Shuozhong Wang. Cyber Warfare: Steganography vs. Steganalysis. Communications of the ACM, 47(10):76-82, 2004. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 501, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 567 + ], + "type": "text", + "content": "[71] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier PlenOctrees for Dynamic Radiance Field Rendering in Real-Time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "text", + "content": "[72] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning Neural Implicit Surfaces by Volume Rendering for Multi-View Reconstruction. Advances in Neural Information Processing Systems, 34, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 690 + ], + "type": "text", + "content": "[73] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning Multi-View Image-Based Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690–4699, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[74] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image Quality Assessment: From Error Visibility" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "to Structural Similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 148 + ], + "type": "text", + "content": "[75] Xinyu Weng, Yongzhi Li, Lu Chi, and Yadong Mu. High-Capacity Convolutional Video Steganography With Temporal Residual Modeling. In Proceedings of the 2019 on International Conference on Multimedia Retrieval, pages 87-95, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 150, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 547, + 194 + ], + "type": "text", + "content": "[76] Raymond B Wolfgang and Edward J Delp. A Watermark for Digital Images. In Proceedings of 3rd IEEE International Conference on Image Processing, volume 3, pages 219-222. IEEE, IEEE, 1996. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 195, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 547, + 239 + ], + "type": "text", + "content": "[77] Yue Wu, Guotao Meng, and Qifeng Chen. Embedding Novel Views in a Single JPEG Image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14519-14527, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "text", + "content": "[78] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 294, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 349 + ], + "type": "text", + "content": "[79] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. BungeeNeRF: Progressive Neural Radiance Field for Extreme Multi-Scale Scene Rendering. In The European Conference on Computer Vision (ECCV), volume 2, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[80] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 404, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 547, + 459 + ], + "type": "text", + "content": "[81] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4479–4489, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 460, + 545, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 504 + ], + "type": "text", + "content": "[82] Hyukryul Yang, Hao Ouyang, Vladlen Koltun, and Qifeng Chen. Hiding Video in Audio Via Reversible Generative Models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1100-1109, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 505, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 558 + ], + "type": "text", + "content": "[83] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance Fields Without Neural Networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5491–5500, 2022. 1, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[84] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for Real-Time Rendering of Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[85] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. Pixelnerf: Neural Radiance Fields From One or Few Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 659, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 547, + 713 + ], + "type": "text", + "content": "[86] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14448-14457, 2021. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "452" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 339 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[87] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A Structured Dictionary Perspective on Implicit Neural Representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 195 + ], + "type": "text", + "content": "[88] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray Priors Through Reprojection: Improving Neural Radiance Fields for Novel View Extrapolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18376-18386, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 228 + ], + "type": "text", + "content": "[89] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and Improving Neural Radiance Fields. ArXiv Preprint ArXiv:2010.07492, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "text", + "content": "[90] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features As a Perceptual Metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 339 + ], + "type": "text", + "content": "[91] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul E. Debevec, William T. Freeman, and Jonathan T. Barron. NeR-Factor: Neural Factorization of Shape and Reflectance Under an Unknown Illumination. ACM Trans. Graph, 40:237:1-237:18, 2021. 1" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "453" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/36fc9ee6-a13a-40ed-bd76-54644b9d43ea_content_list.json b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/36fc9ee6-a13a-40ed-bd76-54644b9d43ea_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..faa928682adcbcad4c8aba1b9aba1e7d810f761a --- /dev/null +++ b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/36fc9ee6-a13a-40ed-bd76-54644b9d43ea_content_list.json @@ -0,0 +1,1751 @@ +[ + { + "type": "text", + "text": "Stochastic Segmentation with Conditional Categorical Diffusion Models", + "text_level": 1, + "bbox": [ + 122, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lukas Zbinden*", + "bbox": [ + 240, + 181, + 370, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adrian Thomas Huber", + "bbox": [ + 192, + 199, + 372, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lars Doorenbos*", + "bbox": [ + 410, + 181, + 547, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Raphael Sznitman", + "bbox": [ + 411, + 199, + 558, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Theodoros Pissas", + "bbox": [ + 588, + 181, + 728, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pablo Márquez-Neila", + "bbox": [ + 599, + 199, + 771, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Bern, Bern, Switzerland", + "bbox": [ + 333, + 223, + 635, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lukas.zbinden, lars.doorenbos, theodoros.pissas, raphael.sznitman, pablo.marquez}@unibe.ch", + "bbox": [ + 99, + 244, + 867, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 294, + 313, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation has made significant progress in recent years thanks to deep neural networks, but the common objective of generating a single segmentation output that accurately matches the image's content may not be suitable for safety-critical domains such as medical diagnostics and autonomous driving. Instead, multiple possible correct segmentation maps may be required to reflect the true distribution of annotation maps. In this context, stochastic semantic segmentation methods must learn to predict conditional distributions of labels given the image, but this is challenging due to the typically multimodal distributions, high-dimensional output spaces, and limited annotation data. To address these challenges, we propose a conditional categorical diffusion model (CCDM) for semantic segmentation based on Denoising Diffusion Probabilistic Models. Our model is conditioned to the input image, enabling it to generate multiple segmentation label maps that account for the aleatoric uncertainty arising from divergent ground truth annotations. Our experimental results show that CCDM achieves state-of-the-art performance on LIDC, a stochastic semantic segmentation dataset, and outperforms established baselines on the classical segmentation dataset Cityscapes.", + "bbox": [ + 76, + 325, + 473, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 700, + 209, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation has significantly progressed in recent years due to powerful deep neural networks. For most methods, the key objective is to generate a single segmentation output that accurately matches the image's content. However, this may not be suitable for safety-critical domains such as medical diagnostics and autonomous driving, as images in these applications often suffer from inherent ambiguity or annotations that have differences in opinion. In these cases, generating a single coherent segmentation may be hopeless to fully describe the set of correct", + "bbox": [ + 75, + 726, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2860b575fdb0e7c4132b9ff61a9052200b804cf1558791aad29fd5ee10b52049.jpg", + "image_caption": [ + "Figure 1: Examples from the LIDC dataset, where expert radiologists were asked to annotate lung nodules. Despite their expertise, they disagree significantly on many cases. Standard segmentation networks fail to capture these variations, thereby giving a false sense of confidence in model predictions. Our approach learns the distribution of possible labels, allowing us to generate realistic and diverse segmentations." + ], + "image_footnote": [], + "bbox": [ + 511, + 297, + 883, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "labeling.", + "bbox": [ + 500, + 595, + 560, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instead, multiple possible correct segmentation maps may be required to reflect the true distribution of annotations. For instance, Fig. 1 illustrates the task of lung nodule segmentation from CT scans where expert annotators provide multiple valid segmentation maps. In this context, stochastic semantic segmentation methods must learn to predict conditional distributions of labels given the image. Doing so is challenging, however, as the distribution is typically multimodal, the output space is high-dimensional, and annotation data is limited.", + "bbox": [ + 496, + 611, + 890, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Denoising Diffusion Probabilistic Models (DDPMs) appear well-suited to overcome these challenges. DDPMs have recently drawn strong interest in computer vision as a framework for learning complex distributions in high-dimensional spaces. After achieving state-of-the-art performance on image synthesis [13], they have been successfully extended to solve tasks such as text-to-image generation [41], counterfactual explanation generation [24], inpainting [34], but also image classification [56] and seman", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 96, + 886, + 202, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1119", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tic segmentation [1, 3, 48] amongst others.", + "bbox": [ + 76, + 90, + 359, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While DDPMs were originally formulated as probabilistic models able to learn high-dimensional data distributions of discrete and ordered variables (e.g., RGB pixel values), re-formulations and modifications that allow for categorical variables (e.g., labels) [21] are one of the key reasons why DDPMs are being explored in a broad range of computer vision tasks [12]. Specifically, the ability to model the spatial distribution of categorical variables is well suited for numerous computer vision tasks, including semantic segmentation [6, 8, 10, 14, 16, 17, 27, 31, 33, 54, 55]. Yet until now, segmentation methods using DDPMs have relied on the original discrete and ordered formulation and different heuristics to yield categorical outputs [1, 3, 48]. Consequently, the potential advantages of adopting diffusion models of categorical variables for stochastic image segmentation are still unknown.", + "bbox": [ + 75, + 106, + 467, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of the above, we propose a conditional categorical diffusion model (CCDM) for semantic segmentation based on DDPMs, which models both the observed and the latent variables as categorical distributions. This enables the model to explicitly generate labels maps of discrete, unordered variables, thereby circumventing the need for switching between continuous and discrete domains, as in previous methods. The model is conditioned to the input image, making it possible to generate multiple segmentation label maps that account for the aleatoric uncertainty arising from image ambiguity. We show experimentally that our approach achieves state-of-the-art performance on LIDC, a stochastic semantic segmentation dataset, according to several performance measures. Moreover, when applied to the classical segmentation dataset Cityscapes, our method provides competitive results, outperforming established baselines.", + "bbox": [ + 75, + 348, + 467, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are the following:", + "bbox": [ + 96, + 604, + 454, + 619 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a conditional categorical diffusion model capable of learning the label distribution given an input image that can be used to produce diverse segmentation samples that capture aleatoric uncertainty.", + "- For the task of learning a multi-rater semantic segmentation label distribution, our method achieves state-of-the-art performance on LIDC, being the first diffusion-based approach proposed for this task.", + "- We report competitive performance on a challenging semantic segmentation task, Cityscapes, outperforming several established baselines using a lightweight model that also leverages an off-the-shelf pre-trained feature extractor." + ], + "bbox": [ + 96, + 628, + 467, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 859, + 212, + 875 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stochastic segmentation: Methods for stochastic se", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "matic segmentation aim at capturing the aleatoric uncertainty and inherent unpredictability of the labels used for segmentation. Different frameworks have been proposed to yield segmentations according to the underlying label distribution.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Initial works aimed at equipping a standard U-Net [40] with a probabilistic element to generate multiple predictions for the same image, typically accomplished by adding a conditional variational autoencoder (cVAE) [45], where the low-dimensional latent space of the cVAE encodes the possible segmentation variants. In [28], samples from this latent space are upscaled and concatenated at the last layer of the U-Net. Multiple methods extend this set-up to a hierarchical version [4, 29, 53]. Other works use normalizing flows to allow for a more expressive distribution than the Gaussian distribution in the cVAE [43, 46], switch to a discrete latent space [37], or add variational dropout and use the inter-grader variability directly as a training target [23].", + "bbox": [ + 496, + 167, + 890, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several other methods do not rely on the probabilistic U-Net. Monteiro et al. [35] propose a network that uses a low-rank multivariate normal distribution to model the logit distribution. Kassapis et al. [25] leverage adversarial training to learn possible label maps based on the logits of a trained segmentation network. Zhang et al. [52] employ an autoregressive PixelCNN to model the conditional distribution between pixels. Finally, Gao et al. [15] use a mixture of stochastic experts, where each expert network estimates a mode of the uncertainty, and a gating network predicts the probabilities that an input image is segmented by one of the experts. Our method is the first to explore the use of categorical diffusion models for stochastic segmentation.", + "bbox": [ + 496, + 367, + 890, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models: Generative diffusion models [44] have drawn much attention following their popularization by [19]. Since then, diffusion models have been successfully applied to various domains, such as image generation, restoration, and super-resolution [12].", + "bbox": [ + 496, + 564, + 890, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More central to the work presented here, a few methods have attempted to apply diffusion models to semantic segmentation. Baranchuk et al. [3] first train diffusion models to generate images, then use multilayer perceptrons (MLP) on its features to predict the class label. Other works focus on binary segmentation with conditional diffusion models [1, 48]. These methods generate single-channel continuous samples conditioned on the input image and obtain binary segmentation masks by thresholding the result. Directly applying continuous diffusion is also done in [49, 50]. Chen et al. [9] generate discrete data with continuous diffusion models by encoding categorical data into bits and modeling these bits as real numbers.", + "bbox": [ + 496, + 642, + 890, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hoogeboom et al. [21] propose multinomial diffusion, a variation of diffusion models designed for categorical data. Subsequently, multinomial diffusion has been applied to discrete use cases, such as for tabular data [30], the latent", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1120", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "space of vector-quantized variational auto-encoders [11, 22] or text [21]. They can also generate segmentation maps in the unconditional setting at a very small resolution $(32 \\times 64)$ [21]. Instead, we focus on the unexplored conditional case and demonstrate results at significantly higher resolutions (up to $256 \\times 512$ ).", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 196, + 168, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now introduce our approach by first framing the problem setting and defining the necessary notation. We then describe categorical diffusion models and the conditioning procedure to produce stochastic semantic segmentation via diffusion.", + "bbox": [ + 75, + 220, + 468, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background and notation", + "text_level": 1, + "bbox": [ + 76, + 306, + 308, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A denoising diffusion probabilistic model (DDPM) is a latent variable model $p_{\\theta}(\\mathbf{x}_0) = \\int p_{\\theta}(\\mathbf{x}_{0:T})d\\mathbf{x}_{1:T}$ describing the distribution of an observable variable $\\mathbf{x}_0 \\in \\mathbb{R}^D$ using a collection of $T$ latent variables $\\{\\mathbf{x}_t\\}_{t=1}^T$ with the same dimensionality as $\\mathbf{x}_0$ . The joint distribution is modeled as a Markov chain $p_{\\theta}(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T)\\prod_{t=1}^T p_{\\theta}(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t)$ , which is commonly known as the reverse process. The initial $p(\\mathbf{x}_T)$ is set to a known, tractable distribution such as the Gaussian distribution, while the transition distribution $p_{\\theta}$ , parameterized by $\\theta$ , is the trainable component of the model. Training a DDPM aims to approximate $p_{\\theta}(\\mathbf{x}_0)$ to an empirical distribution $q(\\mathbf{x}_0)$ defined by a collection of samples (e.g., images from the real world). To that end, training minimizes the cross-entropy between both distributions,", + "bbox": [ + 75, + 330, + 470, + 542 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0})} \\left[ - \\log p _ {\\theta} (\\mathbf {x} _ {0}) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 554, + 468, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is intractable as it requires marginalizing over the latent variables. Instead, a tractable distribution $q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0)$ is introduced and used as an approximation to the intractable true posterior $p(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0)$ to define the evidence lower bound (ELBO),", + "bbox": [ + 75, + 587, + 468, + 662 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\log p _ {\\theta} (\\mathbf {x} _ {0}) \\geq \\mathbb {E} _ {\\mathbf {x} _ {1: T} \\sim q \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right)} \\left[ \\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0 : T}\\right)}{q \\left(\\mathbf {x} _ {1 : T} \\mid \\mathbf {x} _ {0}\\right)} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 672, + 468, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the expectation is approximated by Monte Carlo sampling. The lower bound is tight when the approximate posterior $q$ equals the real posterior. Maximizing the ELBO over samples from $q(\\mathbf{x}_0)$ minimizes the cross-entropy loss of Eq. (1).", + "bbox": [ + 75, + 717, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key difference between DDPMs and other latent variable models is that the approximate posterior $q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0)$ is fixed and not learnable. DDPMs model this distribution as a Markov chain $q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) = \\prod_{t=1}^{T} q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1})$ , known as the forward process. The transition distribution $q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1})$ is chosen to be a tractable distribution that allows efficient sampling from $q(\\mathbf{x}_t \\mid \\mathbf{x}_0)$ for", + "bbox": [ + 75, + 792, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "any $t$ . The only constraint in the design of a DDPM is that $q(\\mathbf{x}_T \\mid \\mathbf{x}_0) \\approx p(\\mathbf{x}_T)$ .", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The original DDPM [19] modeled the transition distributions of the forward and the reverse processes as Gaussian with diagonal covariance matrices, and $p(\\mathbf{x}_T)$ as a standard multivariate normal. However, these assumptions are inadequate when the elements of $\\mathbf{x}_0$ belong to discrete, unordered sets, as in the task of image segmentation.", + "bbox": [ + 496, + 121, + 893, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Categorical diffusion model", + "text_level": 1, + "bbox": [ + 498, + 220, + 750, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now consider the denoising diffusion formulation to learn complex distributions of discrete image labelings. The observable variable $\\mathbf{x}_0\\in \\mathcal{L}^D$ is categorical, where $D$ is the number of pixels of the image and $\\mathcal{L} = \\{1,\\dots ,L\\}$ is the set of discrete labels that can be assigned to each pixel. Following [21], we consider that all latent variables in $\\mathbf{x}_{1:T}$ are also categorical and that the transition distributions for the forward and reverse processes are modeled as categorical distributions. For the forward process, the transition distribution acts element-wise over the previous state $\\mathbf{x}_{t - 1}$ to produce the parameters of the distribution for $\\mathbf{x}_t$ as,", + "bbox": [ + 496, + 244, + 890, + 411 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\prod_ {d = 1} ^ {D} q \\left(\\mathbf {x} _ {t} [ d ] \\mid \\mathbf {x} _ {t - 1} [ d ]\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 421, + 890, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t[d]$ indicates the label at time $t$ and pixel $d$ . In the following discussion, we will use $x_{t}\\in \\mathcal{L}$ to refer to the label of a single pixel $d$ , and we will drop the index $d$ for clarity. The pixel-wise transition distribution $q(x_{t}\\mid x_{t - 1})$ gives the element-wise probability of the next label given the previous label as,", + "bbox": [ + 496, + 473, + 890, + 565 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {C} \\left(x _ {t}; \\frac {\\beta_ {t}}{L} \\mathbf {1} + \\left(1 - \\beta_ {t}\\right) \\mathbf {e} _ {x _ {t - 1}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 575, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{1} = (1,\\dots ,1)^T$ $\\mathbf{e}_{\\ell}$ is the one-hot encoding vector with 1 in position $\\ell$ and 0 elsewhere, and the hyperparameter $\\alpha_{t} = 1 - \\beta_{t}\\in (0,1)$ indicates the probability of keeping the label unchanged. $\\mathcal{C}(x;\\mathbf{p})$ denotes the categorical distribution with parameter vector $\\mathbf{p}\\in [0,1]^L$ . From the properties of categorical distributions, $\\mathcal{C}(x\\mid \\mathbf{p}) = \\mathbf{p}[x]$ and $\\sum_{x}\\mathbf{p}[x] = 1$", + "bbox": [ + 496, + 619, + 890, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The transition distribution of the forward process can be composed as,", + "bbox": [ + 498, + 726, + 890, + 756 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {C} \\left(x _ {t}; \\frac {1 - \\bar {\\alpha} _ {t}}{L} \\mathbf {1} + \\bar {\\alpha} _ {t} \\mathbf {e} _ {x _ {0}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 766, + 890, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $\\bar{\\alpha}_t = \\prod_{\\tau=1}^t \\alpha_\\tau$ , which enables efficient sampling of elements from the Markov chain at any location $t$ . Finally, the posterior of the transition distribution can be computed with the previous formulas by applying Bayes rule,", + "bbox": [ + 496, + 811, + 890, + 873 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) = \\mathcal {C} \\left(x _ {t - 1}; \\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1121", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3bb8a669fdf7b65a888888f22a0a0e90483cab35204622ead8bd7bb6b9c87d01.jpg", + "image_caption": [ + "Figure 2: Illustration of the reverse process of our method. The conditional categorical diffusion model (CCDM) receives as input an image $I$ and a categorical label map $\\mathbf{x}_T^{(i)}$ sampled from the categorical uniform noise. The reverse process of the CCDM generates a label map $\\mathbf{x}_0^{(i)}$ , which is a sample from the learned distribution $p(\\mathbf{x}_0 \\mid I)$ . When repeated for $N$ samples, we obtain an empirical approximation to the multimodal label distribution for the image $I$ , learned from the annotations of multiple expert raters." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 467, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with,", + "bbox": [ + 76, + 364, + 114, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right) = \\frac {1}{\\tilde {\\pi}} \\left(\\frac {\\beta_ {t}}{L} \\mathbf {1} + \\alpha_ {t} \\mathbf {e} _ {x _ {t}}\\right) \\odot \\left(\\frac {1 - \\bar {\\alpha} _ {t - 1}}{L} \\mathbf {1} + \\bar {\\alpha} _ {t - 1} \\mathbf {e} _ {x _ {0}}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 388, + 478, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\tilde{\\pi} = \\frac{1 - \\bar{\\alpha}_t}{L} +\\bar{\\alpha}_t\\cdot \\delta_{xt}^{x_0}$ , where $\\delta$ is the Kronecker delta.", + "bbox": [ + 76, + 434, + 455, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The transition distribution of the reverse process is also an element-wise categorical distribution,", + "bbox": [ + 76, + 450, + 468, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\prod_ {d = 1} ^ {D} \\mathcal {C} \\left(x _ {t - 1}; \\hat {\\mathbf {p}} _ {t - 1}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 489, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{t - 1} = \\mathbf{x}_{t - 1}[d]$ and $\\hat{\\mathbf{p}}_{t - 1}$ are the label and the estimated parameter vector, respectively, at pixel $d$ . Unlike the forward process, the parameter vector for the pixel $d$ is not computed considering only the element $d$ of $\\mathbf{x}_t$ . Instead, it is modeled as a function $f: \\mathcal{L}^D \\to [0,1]^{D\\times L}$ that incorporates context by considering the entire label map $\\mathbf{x}_t$ to produce a collection of $D$ probability distributions for $\\mathbf{x}_{t - 1}$ , which we refer to as $\\hat{\\mathbf{P}}_{t - 1} \\in [0,1]^{D\\times L}$ with $\\hat{\\mathbf{p}}_{t - 1} = \\hat{\\mathbf{P}}_{t - 1}[d]$ .", + "bbox": [ + 75, + 541, + 468, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While it is possible to use a neural network to estimate $\\hat{\\mathbf{P}}_{t - 1}$ , Ho et al. [19] suggested that a consistent output space for the network led to enhanced performance. Following this idea, we train a network $f_{\\theta}$ , parameterized by $\\theta$ , to compute $\\hat{\\mathbf{P}}_0 = f_\\theta (\\mathbf{x}_t,t)\\in [0,1]^{D\\times L}$ by receiving a label map $\\mathbf{x}_t$ and the step $t$ . We then transform the parameter vector for each pixel, $\\hat{\\mathbf{p}}_0 = \\hat{\\mathbf{P}}_0[d]$ to the parameter vector $\\hat{\\mathbf{p}}_{t - 1}$ for the same pixel of $\\mathbf{x}_{t - 1}$ as,", + "bbox": [ + 75, + 678, + 468, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {C} \\left(x _ {t - 1}; \\hat {\\mathrm {p}} _ {t - 1}\\right) = (9) \\\\ = \\sum_ {x _ {0}} q \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) \\cdot \\mathcal {C} \\left(x _ {0}; \\hat {\\mathbf {p}} _ {0}\\right) (10) \\\\ = \\sum_ {x _ {0}} \\mathcal {C} \\left(x _ {t - 1}; \\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right)\\right) \\cdot \\mathcal {C} \\left(x _ {0}; \\hat {\\mathbf {p}} _ {0}\\right), (11) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 809, + 468, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from which,", + "bbox": [ + 500, + 90, + 584, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {p}} _ {t - 1} = \\sum_ {x _ {0} \\in \\mathcal {L}} \\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right) \\cdot \\hat {\\mathbf {p}} _ {0} \\left[ x _ {0} \\right], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 109, + 890, + 142 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we have omitted the pixel indices $d$ for clarity. This transformation is not necessary when $t = 1$ , as then $\\hat{\\mathbf{p}}_{t - 1} = \\hat{\\mathbf{p}}_0$ computed by $f_{\\theta}$ . It is also possible to perform this computation in parallel for every pixel to efficiently obtain $\\hat{\\mathbf{P}}_{t - 1}$ . Note that the result of Eq. (12) differs from the parameter vector computed in [21], where the ill-defined expression $\\hat{\\mathbf{p}}_{t - 1} = \\pi (x_t,\\hat{\\mathbf{x}}_0)$ is employed.", + "bbox": [ + 496, + 148, + 893, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Conditional categorical diffusion", + "text_level": 1, + "bbox": [ + 500, + 262, + 787, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In stochastic segmentation, the label map $\\mathbf{x}_0$ for an image $I$ is modeled by a distribution $q(\\mathbf{x}_0\\mid I)$ . This distribution is often too complex to be properly approximated as a product of pixel-wise categorical distributions. We use a conditional categorical diffusion model $p(\\mathbf{x}_0\\mid I)$ (CCDM) to model the potentially complex interactions between labels and pixels.", + "bbox": [ + 496, + 285, + 890, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When conditioning the categorical diffusion model on an image, the forward process remains unchanged, $q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0, I) = q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0)$ , as any latent variable is conditionally independent of the image given any previous variable. On the other hand, the reverse process needs to incorporate the dependency on the image in its transition distribution, $p_{\\theta}(\\mathbf{x}_{0:T} \\mid I) = p(\\mathbf{x}_T \\mid I) \\prod_{t=1}^{T} p_{\\theta}(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t, I)$ . In practice, this dependency is enforced by an additional input to the neural network $f_{\\theta}(\\mathbf{x}_t, t, I)$ .", + "bbox": [ + 496, + 391, + 892, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 500, + 535, + 604, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training is performed by maximizing the ELBO of Eq. (2). Reorganizing terms and distributing expectations for variance reduction, we express the ELBO as a sum of three terms:", + "bbox": [ + 496, + 558, + 893, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log p _ {\\theta} (\\mathbf {x} _ {0} \\mid I) \\geq \\\\ \\mathbb {E} _ {\\mathbf {x} _ {1} \\sim q \\left(\\mathbf {x} _ {1} \\mid \\mathbf {x} _ {0}\\right)} [ \\log p _ {\\theta} (\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {1}, I) ] (13) \\\\ - \\sum_ {t = 2} ^ {T} \\mathbb {E} _ {\\mathbf {x} _ {t} \\sim q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right)} [ K L (q \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {x} _ {0}\\right) \\| p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, I\\right)) ] (14) \\\\ - K L \\left(q \\left(\\mathbf {x} _ {T} \\mid \\mathbf {x} _ {0}\\right) \\| p \\left(\\mathbf {x} _ {T} \\mid I\\right)\\right). (15) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 626, + 890, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first two terms can be optimized by standard gradient ascent. We approximate the expectations with Monte Carlo sampling with a single sample. The sum over the time variable $t$ is also approximated by a single uniform sample over $\\{1, \\dots, T\\}$ . The KL divergence of the second term is the sum of pixel-wise KL divergences,", + "bbox": [ + 496, + 750, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK L (q \\| p) = \\sum_ {d = 1} ^ {D} K L \\left(q \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) \\| p _ {\\theta} \\left(x _ {t - 1} \\mid \\mathbf {x} _ {t}, I\\right)\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 845, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1122", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1 Training a CCDM with $T$ steps \nRequire: Training data expressed as the empirical distribu-. tion $q(\\mathbf{x}_0,I) = q(\\mathbf{x}_0\\mid I)q(I)$ repeat t \\~ Uniform( $\\{1,\\dots,T\\}$ 1 $\\begin{array}{rl} & {\\boldsymbol {I}\\sim q(\\boldsymbol {I})}\\\\ & {\\mathbf{x}_0\\sim q(\\mathbf{x}_0\\mid \\boldsymbol {I})}\\\\ & {\\mathbf{x}_t\\sim q(\\mathbf{x}_t|\\mathbf{x}_0)}\\\\ & {\\hat{\\mathbf{P}}_0\\leftarrow f_\\theta (\\mathbf{x}_t,I,t)} \\end{array}$ ▷ shape $D\\times L$ if $t > 1$ then $\\triangleright$ Pixel-wise application of Eq. (12) $\\hat{\\mathbf{p}}_{t - 1}\\gets \\sum_{x_0\\in \\mathcal{L}}\\pi (x_t,x_0)\\cdot \\hat{\\mathbf{p}}_0[x_0]\\quad \\triangleright$ shape L $\\triangleright$ Compute KL with Eq. (8) and (16) $\\ell \\leftarrow KL(q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0)\\| p_\\theta (\\mathbf{x}_{t - 1}|\\mathbf{x}_t,I))$ else $\\ell \\leftarrow -\\sum_d\\log \\mathcal{C}(x_0\\mid \\hat{\\mathbf{P}}_0[d])$ end if $\\theta \\leftarrow \\theta -\\nabla_{\\theta}\\ell$ ▷ Gradient descent until converged", + "bbox": [ + 76, + 88, + 472, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the parameter vectors of distributions $q$ and $p$ are computed with Eqs. (7) and (12), respectively. Alg. 1 shows the complete training procedure.", + "bbox": [ + 75, + 412, + 468, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The third term of Eq. (15) does not depend on the learnable parameters $\\theta$ and is ignored during training. It is optimized by the design of the categorical diffusion model. Since the forward process converges as", + "bbox": [ + 75, + 457, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {t \\rightarrow \\infty} q \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {C} \\left(x; \\frac {1}{L}\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 527, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we fix $p(\\mathbf{x}_T\\mid I)$ to the element-wise uniform distribution,", + "bbox": [ + 76, + 571, + 460, + 588 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(x _ {T} \\mid I\\right) = p \\left(x _ {T}\\right) = \\mathcal {C} \\left(x _ {T}; \\frac {1}{L}\\right). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 598, + 468, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This ensures that $p(\\mathbf{x}_T \\mid I) \\approx q(\\mathbf{x}_T \\mid \\mathbf{x}_0)$ , making the third term of the ELBO close to zero.", + "bbox": [ + 75, + 641, + 468, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At inference, the CCDM samples from $p(\\mathbf{x}_0 \\mid I)$ to generate label maps for a given image $I$ , which is achieved by traversing the Markov chain of the reverse process as outlined in Alg. 2 and illustrated in Fig. 2. To minimize the noise of the generated label maps, the CCDM selects the label with maximum probability instead of sampling from $\\mathcal{C}(x_0 \\mid \\hat{\\mathbf{p}}_0)$ in the final step.", + "bbox": [ + 75, + 671, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Architecture of $f_{\\theta}$", + "text_level": 1, + "bbox": [ + 76, + 786, + 250, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As described above, the neural network $f_{\\theta}$ receives a label map $\\mathbf{x}_t$ , a time step $t$ , and an image $I$ to estimate the probability parameters for $\\mathbf{x}_0$ . Its base design is a U-Net-like architecture [13] with self-attention modules at the three innermost layers of the encoder and the decoder [13]. The network processes the input label map represented as", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Inference from a CCDM with $T$ steps" + ], + "code_body": "Require: Input image $I$ , $f_{\\theta}$ a network trained with Alg. 1 \n $\\mathbf{x}_T \\sim \\mathcal{C}^D(x_T; \\frac{1}{L})$ $\\mathbf{x}_{\\text{prev}} \\gets \\mathbf{x}_T \\quad \\triangleright$ Stores interm. and final prediction \nfor $t = T, \\dots, 1$ do \n $\\hat{\\mathbf{P}}_0 \\gets f_{\\theta}(\\mathbf{x}_{\\text{prev}}, I, t)$ \nif $t > 1$ then \n $\\triangleright$ Pixel-wise application of Eq. (12) \n $\\hat{\\mathbf{p}}_{t-1} \\gets \\sum_{x_0 \\in \\mathcal{L}} \\pi(x_t, x_0) \\cdot \\hat{\\mathbf{p}}_0[x_0]$ $\\mathbf{x}_{\\text{prev}} \\sim \\prod_d \\mathcal{C}(x_{t-1} | \\hat{\\mathbf{p}}_{t-1})$ \nelse \n $\\triangleright$ Final prediction \n $\\mathbf{x}_{\\text{prev}} \\gets \\arg \\max_{x_0 \\in \\mathcal{L}} \\hat{\\mathbf{P}}_0[:, x_0]$ \nend if \nend for", + "bbox": [ + 500, + 109, + 890, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a binary tensor with $L$ channels encoding the label of each pixel as a one-hot vector. Parameters of the network are shared for all values of $t$ . The step variable $t$ is encoded with the standard transformer sinusoidal position embedding [19] and concatenated as additional channels to the input tensor and to the feature maps of intermediate layers. Similarly, information from the input image $I$ is presented to the network as raw pixel values concatenated to the input tensor as additional channels. In some experiments we used a pre-trained transformer architecture Dino-ViT [5] to extract informative visual features from the image $I$ . In those cases, the extracted features were concatenated to the feature map of the third level of the U-Net encoder, which corresponds to a spatial shape equal to $\\frac{1}{8}$ the shape of the input image.", + "bbox": [ + 496, + 347, + 890, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 585, + 633, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In all our experiments, we set $T = 250$ and the collection of $\\beta_{t}$ are set following the cosine schedule proposed in [36]. We evaluate our method on two tasks described below.", + "bbox": [ + 498, + 612, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Segmentation with multiple annotations", + "text_level": 1, + "bbox": [ + 498, + 666, + 841, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset The Lung Image Database Consortium (LIDC) [2] binary segmentation dataset consists of 1'018 three dimensional chest CT scans of patients with lung cancer. Lung nodules of each volume are annotated by four expert raters from a pool of 12, yielding large differences in annotations in some cases. We extract nodule-centered slices from the CT volumes and treat each slice as an independent image.", + "bbox": [ + 496, + 688, + 890, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While LIDC is the standard benchmark of stochastic segmentation methods to date (e.g. [4, 15, 23, 25, 28, 29, 35, 43, 53, 54]), experimental configurations (pre-processing, training/validation/test splits, metrics) vastly differ across the literature. We conduct our experiments on the two most prominent LIDC splits and report results on both separately.", + "bbox": [ + 496, + 810, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1123", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5a040d9ebf88f60512da1f6aeaf4beab6df9cdcd8486300417bab3d5e24200b6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLIDCv1LIDCv2
GED16GED32GED50GED100HM-IoU16HM-IoU32GED16GED50GED100HM-IoU16
Prob. Unet [28]0.310±0.01-0.303±0.01+-0.252±0.004†0.552±0.00-0.548±0.00+0.320±0.030‡-0.252±‡0.500±0.030‡
HProb. Unet [29]0.270±0.01----0.530±0.01--0.270±0.010‡--0.530±0.01
PhiSeg [4]0.262±0.00-0.247±0.00+-0.224±0.004†0.586±0.00-0.595±0.00+----
SSN [35]0.259±0.00-0.243±0.01+-0.225±0.0020.558±0.00-0.555±0.01+----
cFlow [43]-0.225±0.01+---0.584±0.00+----
CAR [25]---0.228±0.009--0.264±0.0020.248±0.0040.243±0.0040.592±0.005
JProb. Unet [53]-0.206±0.00---0.647±0.010.262±0.00--0.585±0.00
PixelSeg [52]0.243±0.01---0.614±0.00-0.260±0.00--0.587±0.01
MoSE [15]0.218±0.003-0.195±0.0020.189±0.0020.624±0.004-----
AB [9]0.213±0.0010.196±0.0020.193±0.002-0.614±0.0010.619±0.001----
CIMD [38]0.234±0.0050.218±0.0050.210±0.005-0.587±0.0010.592±0.002----
CCDM (ours)0.212±0.0020.194±0.0010.187±0.0020.183±0.0020.623±0.0020.631±0.0020.239±0.0030.216±0.0030.210±0.0030.598±0.001
", + "bbox": [ + 81, + 89, + 890, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Quantitative results on LIDCv1 and LIDCv2, with the methods ordered by year. Bold and underlined indicate best and second best per column, respectively. Our results are over 3 seeds. For GED, lower is better; for HM-IoU, higher is better. No method, including ours, uses pre-trained weights. Results for CIMD [38] and AB [9] are ours. All other scores are taken from their original papers, except $(^{+})$ from [53], $(-)$ from [53], $(\\dagger)$ from [35], $\\ddagger$ from [25].", + "bbox": [ + 75, + 290, + 892, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The first, referred to as LIDCv1, is used in [4, 15, 35, 53]. LIDCv1 comprises 15'096 slices, divided into training, validation, and testing sets with the ratio $60:20:20$ . The second, LIDCv2, is used in [25, 28] and contains 12'816 images with the ratio $70:15:15$ .", + "bbox": [ + 75, + 377, + 468, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics We measure the performances with the Generalised Energy Distance (GED) and the Hungarian-Matched Intersection over Union (HM-IoU) [15, 25, 29]. Both metrics measure the difference between the distributions of generated and ground-truth label maps. We denote the metrics computed with $n$ samples using a subscript, i.e., $\\mathrm{GED}_n$ and $\\mathrm{HM - IoU}_n$ , and we set $n$ to common values found in the literature. Note that higher number of samples yield more precise estimates.", + "bbox": [ + 75, + 472, + 468, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We compare our approach to eleven recent stochastic segmentation methods: probabilistic U-Net (Prob. Unet) [28], hierarchical probabilistic U-Net (HProb. Unet) [29], PhiSeg [4], stochastic segmentation network (SSN) [35], conditional normalizing flow (cFlow) [43], calibrated adversarial refinement (CAR) [25], joint probabilistic U-Net (JProb. Unet) [53], PixelSeg [52], mixture of stochastic experts (MoSE) [15], analog bits (AB) [9], and collectively intelligent medical diffusion (CIMD) [38].", + "bbox": [ + 75, + 628, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following standard practice, we use random horizontal and vertical flipping and random rotations of $0^{\\circ}$ , $90^{\\circ}$ , $180^{\\circ}$ and $270^{\\circ}$ for data augmentation. The resolution of the input images is $128 \\times 128$ . We trained our method with the Adam optimizer [26] until convergence of the GED metric on the validation set, a polynomial learning rate scheduling starting from $1e^{-4}$ and ending with $1e^{-6}$ , and batch size of 64. We applied Polyak averaging with $\\alpha = 0.99995$ .", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Segmentation with a single annotation", + "text_level": 1, + "bbox": [ + 498, + 377, + 828, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate our method with Cityscapes, a classical multi-class segmentation dataset where each image is annotated with a single label map. It comprises 2'975 RGB images of urban scenes for training and 500 images for validation, with each image labeled using 19 possible classes.", + "bbox": [ + 496, + 400, + 890, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our approach to several established baselines using the validation set: DeepLabv3 [7], HRNet [47], and UPerNet [51], with both ResNet [18] and Swin [32] backbones.", + "bbox": [ + 496, + 477, + 890, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides our standard method, which performs image conditioning by concatenating the raw pixel values as channels of the input tensor, we also included in our comparison a variant CCDM-Dino which leverages pre-trained DinoViT features [5] as additional conditioning concatenated to intermediate feature maps of our model's encoder.", + "bbox": [ + 496, + 537, + 890, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments are conducted separately for two different image resolutions: $128 \\times 256$ and $256 \\times 512$ . For all reported methods, we first resize the images to a fixed resolution and then apply color jittering, random flipping, and standard ImageNet intensity normalization as data augmentation. All baselines are trained for 500 epochs with a batch size of 32, with optimizers, learning rate schedules, and weight decay settings as reported in their respective publications (reported in detail in the supplementary material).", + "bbox": [ + 496, + 628, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method was trained for 800 epochs with a batch size of 32 at $128 \\times 256$ and of 16 at $256 \\times 512$ , using the Adam optimizer [26] with a learning rate of $1e^{-4}$ linearly decayed to $1e^{-6}$ . We applied Polyak averaging with $\\alpha = 0.999$ .", + "bbox": [ + 496, + 765, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance is measured with the mean intersection-over-union (mIoU). Unlike GED and HM-IoU, the metric mIoU is incompatible with multiple label maps per image. During inference, CCDM generates multiple label maps per image that are subsequently fused into a single label map for", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1124", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performance assessment. We found that fusing by averaging the predicted probabilities resulted in superior performances compared to fusing by majority vote.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 76, + 151, + 163, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. LIDC", + "text_level": 1, + "bbox": [ + 76, + 175, + 160, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report performances on LIDCv1 and LIDCv2 in Tab. 1 and qualitative results in Fig. 3. Due to the lack of consistent evaluation protocols, we use a total of 10 metrics, thereby covering all the baselines and allowing for direct comparisons.", + "bbox": [ + 75, + 199, + 470, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our CCDM reaches the best performance for eight out of the ten metrics, despite its relatively small size, with 9M parameters compared to, e.g., the 42M parameters of MoSE. CCDM also outperforms recent continuous diffusion models for segmentation, including AB [9] (9M parameters) and CIMD [38] (24M parameters). On HM-IoU16, the CCDM has a lower mean performance than MoSE by 0.001, but with only half the standard deviation. The JProb. Unet reaches a higher HM-IoU32 than all other methods, despite being considerably worse for GED32 than our CCDM. Furthermore, on LIDCv2, the JProb. Unet achieves only the third-best score on GED16, and fourth-best on HM-IoU16. This result indicates how comparing results obtained on different LIDC versions with each other can be misleading.", + "bbox": [ + 75, + 276, + 470, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 3 presents qualitative results from our method. In columns (g)-(l), we see that our CCDM generates a distribution of samples that captures the annotation variability created by the four expert raters. Further, as seen in the bottom example, the CCDM also generates empty samples according to the annotations (b)-(e).", + "bbox": [ + 75, + 488, + 468, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reduced number of time steps for sampling: During inference, traversing the $T$ steps of the reverse process makes sampling from DDPMs slow. A straightforward solution [36] involves traversing only a subset of nodes of the reverse process, $\\{\\mathbf{x}_{k\\tau} : \\tau \\in \\{0, \\dots, T / k\\}\\}$ , reducing the number of steps by a factor $k$ . This technique accelerates inference at the expense of reduced performance. To illustrate the trade-offs between performance and speed, Fig. 5 presents the evolution of $\\mathrm{GED}_{16}$ and $\\mathrm{HM-IoU}_{16}$ as the number of inference steps is reduced. As expected, CCDMs perform best when the number of training and inference steps are equal, but a reasonable increase in speed without a large sacrifice in performance is possible.", + "bbox": [ + 75, + 579, + 470, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Cityscapes", + "text_level": 1, + "bbox": [ + 76, + 785, + 196, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental comparisons on Cityscapes are presented in Tab. 2, and qualitative examples are provided in Fig. 4. Experiments at $128 \\times 256$ demonstrate that CCDM-Dino outperforms all other methods, even when only a single sample is used. CCDM-raw also remains competitive, being outperformed only by one baseline (UPerNet+Swin", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/91c93a2351a1e273850fa90ca7c121c16028415037f759da82f29ae4317d4d43.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmIoU final (best)
ArchitectureBackbone#params128 × 256256 × 512
DeepLabv3 [7]ResNet50 (✓)39m43.4 (44.1)58.6 (59.2)
DeepLabv3 [7]ResNet101 (✓)58m43.8 (45.5)59.2 (59.8)
UPerNet [51]ResNet101 (✓)83m45.5 (47.1)60.7 (61.2)
HRNet [47]w48v2 (✓)70m48.2 (49.5)63.3 (64.2)
UPerNet [32]Swin-Tiny (✓)58m54.2 (55.9)65.5 (66.0)
CCDM (ours)-
samples=130m53.260.3
samples=530m55.462.0
samples=1030m56.262.4
CCDM (ours)Dino ViT-S (†)
samples=130m + 20M55.564.0
samples=530m + 20M56.965.4
samples=1030m + 20M57.365.8
", + "bbox": [ + 504, + 88, + 877, + 275 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0f674ce2554c6f0fc566e989dbc0ef7cf94233fc1d12fe5793e116a65829a945.jpg", + "table_caption": [ + "Table 2: Results on Cityscapes-val for resolutions $128 \\times 256$ and $256 \\times 512$ . **Bold** and **underlined** indicate best and second best per column, respectively. ( $\\checkmark$ ) and ( $\\dagger$ ) indicate supervised and self-supervised pretraining of the backbone, respectively. Gray indicates pretrained, non-finetuned parameters. We report final performance for our method and baselines. For the latter we also provide best achieved performance during training (in parenthesis). For CCDM methods, the field samples indicate the number of generated samples for label map fusion, as explained in Sect 4.2." + ], + "table_footnote": [], + "table_body": "
CCDM CapacitymIoU(128 × 256)
#paramsUNet Levelssamples=1samples=5samples=10
5.4M437.839.740.6
7.5M544.748.348.5
22M451.654.053.6
30M553.255.456.2
", + "bbox": [ + 557, + 472, + 836, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Effect of increasing CCDM capacity (without feature conditioning).", + "bbox": [ + 498, + 553, + 890, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tiny), despite using only between $36\\%$ and $51\\%$ of the parameters of other models. Similarly, at $256 \\times 512$ , CCDM-Dino outperforms four of the baselines with a single sample, lags behind UPerNet+Swin-Tiny only by 0.1 percent points with 5 samples, and outperforms all baselines with 10 samples. As expected, averaging across more samples improves performance for both CCDM-raw and CCDM-Dino, albeit with diminishing gains. Furthermore, the addition of Dino features boosts single-sample performance by 2.3 percent points at $128 \\times 256$ , and 3.7 percent points at $256 \\times 512$ , hinting the greater value of adding feature conditioning for generating segmentation at a higher resolution.", + "bbox": [ + 496, + 632, + 893, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CCDM Capacity: Tab. 3(b) demonstrates the effect of increasing the capacity of CCDM. Using more U-Net feature levels, and increasing the number of parameters by doubling the number of channels per level, increases the performance regardless of the number of samples used for inference.", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1125", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/124261e915105a28395a1b92f13a05f954f52fb1b05d1e01069d9d8447e93011.jpg", + "image_caption": [ + "Figure 3: Qualitative results on four LIDC images with our method. (a) shows the image, (b)-(e) its four labels, (f) the mean prediction of our CCDM over six predictions, and (g)-(l) six individual predictions." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 870, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ddc294190fcc727fc452af9ee35de2171ac5486e599827f1e79a0f9c7552f6e8.jpg", + "image_caption": [ + "Figure 4: Qualitative comparisons on Cityscapes. All methods are trained and tested at a resolution of $256 \\times 512$ . Our method produces structures with greater visual realism than other baselines. This is especially noticeable inside the marked regions." + ], + "image_footnote": [], + "bbox": [ + 84, + 378, + 883, + 599 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/460bab6719033622606de27821b83948ac968035bd01ee6e9cb79eb974dce2f3.jpg", + "image_caption": [ + "Figure 5: LIDC GED and HM-IoU versus the number of sampling steps on LIDC. Evaluated on 500 random test images using 16 samples each, over 3 seeds." + ], + "image_footnote": [], + "bbox": [ + 148, + 666, + 395, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 667, + 617, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced conditional categorical diffusion models (CCDMs) that are capable of effectively modeling pixel-level semantic distributions. Notably, and contrary to standard deterministic segmentation approaches, our model can produce diverse samples given an input image, thereby capturing the aleatoric uncertainty. Our method learns a multimodal label distribution of segmentations, induced by annotations from multiple expert raters, for which it achieves state-of-the-art results on a challenging medical imaging dataset, LIDC. Additionally, we demonstrate that it can achieve competitive performance on a standard multi-class semantic segmentation benchmark, Cityscapes, by outperforming several established, heavily engineered baselines", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1126", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "despite using significantly fewer parameters.", + "bbox": [ + 76, + 90, + 374, + 106 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "One limitation of our method is the requirement of several iterations for producing a sample, which is a common shortcoming of diffusion models. Accelerating sampling constitutes a crucial research direction, orthogonal to the present work. Finally, resolution scaling remains notoriously difficult for diffusion models, with successful examples relying on massive computational resources to train cascades of models that gradually increase resolution [20, 42] or operate on the latent space of existing embedding methods for continuous data (e.g. images) [39] that are not available for categorical data.", + "bbox": [ + 75, + 107, + 468, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 289, + 243, + 306 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially funded by the University of Bern, Swiss National Science Foundation Grants #320030-188591, #200021-192285 and #200021-191983.", + "bbox": [ + 76, + 315, + 470, + 361 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 377, + 174, + 393 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tomer Amit, Eliya Nachmani, Tal Shaharbany, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021. 2", + "[2] Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al. The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Medical physics, 38(2):915-931, 2011. 5", + "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 2", + "[4] Christian F Baumgartner, Kerem C Tezcan, Krishna Chaitanya, Andreas M Hotker, Urs J Muehlematter, Khoschy Schawkat, Anton S Becker, Olivio Donati, and Ender Konukoglu. Phiseg: Capturing uncertainty in medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part II 22, pages 119-127. Springer, 2019. 2, 5, 6", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5, 6", + "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2", + "[7] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for seman-" + ], + "bbox": [ + 86, + 402, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7", + "[8] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2", + "[9] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022. 2, 6, 7", + "[10] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems, 34:9355-9366, 2021. 2", + "[11] Max Cohen, Guillaume Quispe, Sylvain Le Corff, Charles Ollion, and Eric Moulines. Diffusion bridges vector quantized variational autoencoders. arXiv preprint arXiv:2202.04895, 2022. 3", + "[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. arXiv preprint arXiv:2209.04747, 2022. 2", + "[13] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. CoRR, abs/2105.05233, 2021. 1, 5", + "[14] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019. 2", + "[15] Zhitong Gao, Yuong Chen, Chuyu Zhang, and Xuming He. Modeling multimodal aleatoric uncertainty in segmentation with mixture of stochastic expert. arXiv preprint arXiv:2212.07328, 2022. 2, 5, 6", + "[16] Jiaqi Gu, Hyoukjun Kwon, Dilin Wang, Wei Ye, Meng Li, Yu-Hsin Chen, Liangzhen Lai, Vikas Chandra, and David Z Pan. Multi-scale high-resolution vision transformer for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12094-12103, 2022. 2", + "[17] Adam W Harley, Konstantinos G Derpanis, and Iasonas Kokkinos. Segmentation-aware convolutional networks using local attention masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5038-5047, 2017. 2", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 6", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 4, 5", + "[20] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 9" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1127", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34:12454-12465, 2021. 2, 3, 4", + "[22] Minghui Hu, Yujie Wang, Tat-Jen Cham, Jianfei Yang, and Ponnuthurai N Suganthan. Global context with discrete diffusion in vector quantised modelling for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11502-11511, 2022. 3", + "[23] Shi Hu, Daniel Worrall, Stefan Knegt, Bas Veeling, Henkjan Huisman, and Max Welling. Supervised uncertainty quantification for segmentation with multiple annotations. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22, pages 137–145. Springer, 2019. 2, 5", + "[24] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. arXiv preprint arXiv:2203.15636, 2022. 1", + "[25] Elias Kassapis, Georgi Dikov, Deepak K Gupta, and Cedric Nugteren. Calibrated adversarial refinement for stochastic semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7057-7067, 2021. 2, 5, 6", + "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference for Learning Representations, 2015. 6", + "[27] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019. 2", + "[28] Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic u-net for segmentation of ambiguous images. Advances in neural information processing systems, 31, 2018. 2, 5, 6", + "[29] Simon AA Kohl, Bernardino Romera-Paredes, Klaus H Maier-Hein, Danilo Jimenez Rezende, SM Eslami, Pushmeet Kohli, Andrew Zisserman, and Olaf Ronneberger. A hierarchical probabilistic u-net for modeling multi-scale ambiguities. arXiv preprint arXiv:1905.13077, 2019. 2, 5, 6", + "[30] Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. Tabddpm: Modelling tabular data with diffusion models. arXiv preprint arXiv:2209.15421, 2022. 2", + "[31] Lielei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep hierarchical semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1246-1257, 2022. 2", + "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7", + "[33] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Pro-" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 2", + "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1", + "[35] Miguel Monteiro, Loic Le Folgoc, Daniel Coelho de Castro, Nick Pawlowski, Bernardo Marques, Konstantinos Kamnitsas, Mark van der Wilk, and Ben Glocker. Stochastic segmentation networks: Modelling spatially correlated aleatoric uncertainty. Advances in Neural Information Processing Systems, 33:12756-12767, 2020. 2, 5, 6", + "[36] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 5.7", + "[37] Di Qiu and Lok Ming Lui. Modal uncertainty estimation via discrete latent representation. arXiv preprint arXiv:2007.12858, 2020. 2", + "[38] Aimon Rahman, Jeya Maria Jose Valanarasu, Ilker Hacihaliloglu, and Vishal M Patel. Ambiguous medical image segmentation using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11536-11546, 2023. 6, 7", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 9", + "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2", + "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1", + "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 9", + "[43] Raghavendra Selvan, Frederik Faye, Jon Middleton, and Akshay Pai. Uncertainty quantification in medical image segmentation with normalizing flows. In Machine Learning in Medical Imaging: 11th International Workshop, MLMI 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4, 2020, Proceedings 11, pages 80-90. Springer, 2020. 2, 5, 6", + "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1128", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2", + "[45] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 2", + "[46] MM Amaan Valiuddin, Christiaan GA Viviers, Ruud JG van Sloun, Peter HN de With, and Fons van der Sommen. Improving aleatoric uncertainty quantification in multi-annotated medical image segmentation with normalizing flows. In Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, and Perinatal Imaging, Placental and Preterm Image Analysis: 3rd International Workshop, UNSURE 2021, and 6th International Workshop, PIPPI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 3, pages 75–88. Springer, 2021. 2", + "[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. TPAMI, 2019. 6, 7", + "[48] Julia Wolleb, Robin Sandkühler, Florentin Bieder, Philippe Valmaggia, and Philippe C Cattin. Diffusion models for implicit image segmentation ensembles. arXiv preprint arXiv:2112.03145, 2021. 2", + "[49] Junde Wu, Huihui Fang, Yu Zhang, Yehui Yang, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. arXiv preprint arXiv:2211.00611, 2022. 2", + "[50] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, and Yanwu Xu. Medsegdiff-v2: Diffusion based medical image segmentation with transformer. arXiv preprint arXiv:2301.11798, 2023. 2", + "[51] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018. 6, 7", + "[52] Wei Zhang, Xiaohong Zhang, Sheng Huang, Yuting Lu, and Kun Wang. Pixelseg: Pixel-by-pixel stochastic semantic segmentation for ambiguous medical images. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4742-4750, 2022. 2, 6", + "[53] Wei Zhang, Xiaohong Zhang, Sheng Huang, Yuting Lu, and Kun Wang. A probabilistic model for controlling diversity and accuracy of ambiguous medical image segmentation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4751-4759, 2022. 2, 5, 6", + "[54] Yifan Zhang, Bo Pang, and Cewu Lu. Semantic segmentation by early region proxy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1258-1268, 2022. 2, 5", + "[55] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[56] Roland S Zimmermann, Lukas Schott, Yang Song, Benjamin A Dunn, and David A Klindt. Score-based generative classifiers. arXiv preprint arXiv:2110.00473, 2021. 1", + "bbox": [ + 501, + 90, + 890, + 133 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1129", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/36fc9ee6-a13a-40ed-bd76-54644b9d43ea_model.json b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/36fc9ee6-a13a-40ed-bd76-54644b9d43ea_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f9ad0e829dd0693911d8eb81980f70f3eca4cd9e --- /dev/null +++ b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/36fc9ee6-a13a-40ed-bd76-54644b9d43ea_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.131, + 0.848, + 0.153 + ], + "angle": 0, + "content": "Stochastic Segmentation with Conditional Categorical Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.182, + 0.371, + 0.198 + ], + "angle": 0, + "content": "Lukas Zbinden*" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.2, + 0.374, + 0.216 + ], + "angle": 0, + "content": "Adrian Thomas Huber" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.182, + 0.548, + 0.198 + ], + "angle": 0, + "content": "Lars Doorenbos*" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.2, + 0.56, + 0.217 + ], + "angle": 0, + "content": "Raphael Sznitman" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.182, + 0.73, + 0.198 + ], + "angle": 0, + "content": "Theodoros Pissas" + }, + { + "type": "text", + "bbox": [ + 0.6, + 0.2, + 0.772, + 0.217 + ], + "angle": 0, + "content": "Pablo Márquez-Neila" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.224, + 0.637, + 0.242 + ], + "angle": 0, + "content": "University of Bern, Bern, Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.245, + 0.868, + 0.26 + ], + "angle": 0, + "content": "{lukas.zbinden, lars.doorenbos, theodoros.pissas, raphael.sznitman, pablo.marquez}@unibe.ch" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.295, + 0.314, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.326, + 0.474, + 0.675 + ], + "angle": 0, + "content": "Semantic segmentation has made significant progress in recent years thanks to deep neural networks, but the common objective of generating a single segmentation output that accurately matches the image's content may not be suitable for safety-critical domains such as medical diagnostics and autonomous driving. Instead, multiple possible correct segmentation maps may be required to reflect the true distribution of annotation maps. In this context, stochastic semantic segmentation methods must learn to predict conditional distributions of labels given the image, but this is challenging due to the typically multimodal distributions, high-dimensional output spaces, and limited annotation data. To address these challenges, we propose a conditional categorical diffusion model (CCDM) for semantic segmentation based on Denoising Diffusion Probabilistic Models. Our model is conditioned to the input image, enabling it to generate multiple segmentation label maps that account for the aleatoric uncertainty arising from divergent ground truth annotations. Our experimental results show that CCDM achieves state-of-the-art performance on LIDC, a stochastic semantic segmentation dataset, and outperforms established baselines on the classical segmentation dataset Cityscapes." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.702, + 0.21, + 0.717 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Semantic segmentation has significantly progressed in recent years due to powerful deep neural networks. For most methods, the key objective is to generate a single segmentation output that accurately matches the image's content. However, this may not be suitable for safety-critical domains such as medical diagnostics and autonomous driving, as images in these applications often suffer from inherent ambiguity or annotations that have differences in opinion. In these cases, generating a single coherent segmentation may be hopeless to fully describe the set of correct" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.298, + 0.885, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.445, + 0.895, + 0.566 + ], + "angle": 0, + "content": "Figure 1: Examples from the LIDC dataset, where expert radiologists were asked to annotate lung nodules. Despite their expertise, they disagree significantly on many cases. Standard segmentation networks fail to capture these variations, thereby giving a false sense of confidence in model predictions. Our approach learns the distribution of possible labels, allowing us to generate realistic and diverse segmentations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.596, + 0.562, + 0.611 + ], + "angle": 0, + "content": "labeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Instead, multiple possible correct segmentation maps may be required to reflect the true distribution of annotations. For instance, Fig. 1 illustrates the task of lung nodule segmentation from CT scans where expert annotators provide multiple valid segmentation maps. In this context, stochastic semantic segmentation methods must learn to predict conditional distributions of labels given the image. Doing so is challenging, however, as the distribution is typically multimodal, the output space is high-dimensional, and annotation data is limited." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Denoising Diffusion Probabilistic Models (DDPMs) appear well-suited to overcome these challenges. DDPMs have recently drawn strong interest in computer vision as a framework for learning complex distributions in high-dimensional spaces. After achieving state-of-the-art performance on image synthesis [13], they have been successfully extended to solve tasks such as text-to-image generation [41], counterfactual explanation generation [24], inpainting [34], but also image classification [56] and seman" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.887, + 0.204, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1119" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.36, + 0.106 + ], + "angle": 0, + "content": "tic segmentation [1, 3, 48] amongst others." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.346 + ], + "angle": 0, + "content": "While DDPMs were originally formulated as probabilistic models able to learn high-dimensional data distributions of discrete and ordered variables (e.g., RGB pixel values), re-formulations and modifications that allow for categorical variables (e.g., labels) [21] are one of the key reasons why DDPMs are being explored in a broad range of computer vision tasks [12]. Specifically, the ability to model the spatial distribution of categorical variables is well suited for numerous computer vision tasks, including semantic segmentation [6, 8, 10, 14, 16, 17, 27, 31, 33, 54, 55]. Yet until now, segmentation methods using DDPMs have relied on the original discrete and ordered formulation and different heuristics to yield categorical outputs [1, 3, 48]. Consequently, the potential advantages of adopting diffusion models of categorical variables for stochastic image segmentation are still unknown." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.468, + 0.603 + ], + "angle": 0, + "content": "In light of the above, we propose a conditional categorical diffusion model (CCDM) for semantic segmentation based on DDPMs, which models both the observed and the latent variables as categorical distributions. This enables the model to explicitly generate labels maps of discrete, unordered variables, thereby circumventing the need for switching between continuous and discrete domains, as in previous methods. The model is conditioned to the input image, making it possible to generate multiple segmentation label maps that account for the aleatoric uncertainty arising from image ambiguity. We show experimentally that our approach achieves state-of-the-art performance on LIDC, a stochastic semantic segmentation dataset, according to several performance measures. Moreover, when applied to the classical segmentation dataset Cityscapes, our method provides competitive results, outperforming established baselines." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.605, + 0.455, + 0.621 + ], + "angle": 0, + "content": "In summary, our main contributions are the following:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.63, + 0.468, + 0.69 + ], + "angle": 0, + "content": "- We propose a conditional categorical diffusion model capable of learning the label distribution given an input image that can be used to produce diverse segmentation samples that capture aleatoric uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.701, + 0.468, + 0.761 + ], + "angle": 0, + "content": "- For the task of learning a multi-rater semantic segmentation label distribution, our method achieves state-of-the-art performance on LIDC, being the first diffusion-based approach proposed for this task." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.772, + 0.468, + 0.845 + ], + "angle": 0, + "content": "- We report competitive performance on a challenging semantic segmentation task, Cityscapes, outperforming several established baselines using a lightweight model that also leverages an off-the-shelf pre-trained feature extractor." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.63, + 0.468, + 0.845 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.861, + 0.214, + 0.875 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Stochastic segmentation: Methods for stochastic se" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "matic segmentation aim at capturing the aleatoric uncertainty and inherent unpredictability of the labels used for segmentation. Different frameworks have been proposed to yield segmentations according to the underlying label distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Initial works aimed at equipping a standard U-Net [40] with a probabilistic element to generate multiple predictions for the same image, typically accomplished by adding a conditional variational autoencoder (cVAE) [45], where the low-dimensional latent space of the cVAE encodes the possible segmentation variants. In [28], samples from this latent space are upscaled and concatenated at the last layer of the U-Net. Multiple methods extend this set-up to a hierarchical version [4, 29, 53]. Other works use normalizing flows to allow for a more expressive distribution than the Gaussian distribution in the cVAE [43, 46], switch to a discrete latent space [37], or add variational dropout and use the inter-grader variability directly as a training target [23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Several other methods do not rely on the probabilistic U-Net. Monteiro et al. [35] propose a network that uses a low-rank multivariate normal distribution to model the logit distribution. Kassapis et al. [25] leverage adversarial training to learn possible label maps based on the logits of a trained segmentation network. Zhang et al. [52] employ an autoregressive PixelCNN to model the conditional distribution between pixels. Finally, Gao et al. [15] use a mixture of stochastic experts, where each expert network estimates a mode of the uncertainty, and a gating network predicts the probabilities that an input image is segmented by one of the experts. Our method is the first to explore the use of categorical diffusion models for stochastic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Diffusion models: Generative diffusion models [44] have drawn much attention following their popularization by [19]. Since then, diffusion models have been successfully applied to various domains, such as image generation, restoration, and super-resolution [12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.838 + ], + "angle": 0, + "content": "More central to the work presented here, a few methods have attempted to apply diffusion models to semantic segmentation. Baranchuk et al. [3] first train diffusion models to generate images, then use multilayer perceptrons (MLP) on its features to predict the class label. Other works focus on binary segmentation with conditional diffusion models [1, 48]. These methods generate single-channel continuous samples conditioned on the input image and obtain binary segmentation masks by thresholding the result. Directly applying continuous diffusion is also done in [49, 50]. Chen et al. [9] generate discrete data with continuous diffusion models by encoding categorical data into bits and modeling these bits as real numbers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Hoogeboom et al. [21] propose multinomial diffusion, a variation of diffusion models designed for categorical data. Subsequently, multinomial diffusion has been applied to discrete use cases, such as for tabular data [30], the latent" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1120" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "space of vector-quantized variational auto-encoders [11, 22] or text [21]. They can also generate segmentation maps in the unconditional setting at a very small resolution \\((32 \\times 64)\\) [21]. Instead, we focus on the unexplored conditional case and demonstrate results at significantly higher resolutions (up to \\(256 \\times 512\\))." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.197, + 0.169, + 0.212 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.47, + 0.298 + ], + "angle": 0, + "content": "We now introduce our approach by first framing the problem setting and defining the necessary notation. We then describe categorical diffusion models and the conditioning procedure to produce stochastic semantic segmentation via diffusion." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.308, + 0.31, + 0.323 + ], + "angle": 0, + "content": "3.1. Background and notation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.331, + 0.471, + 0.543 + ], + "angle": 0, + "content": "A denoising diffusion probabilistic model (DDPM) is a latent variable model \\( p_{\\theta}(\\mathbf{x}_0) = \\int p_{\\theta}(\\mathbf{x}_{0:T})d\\mathbf{x}_{1:T} \\) describing the distribution of an observable variable \\( \\mathbf{x}_0 \\in \\mathbb{R}^D \\) using a collection of \\( T \\) latent variables \\( \\{\\mathbf{x}_t\\}_{t=1}^T \\) with the same dimensionality as \\( \\mathbf{x}_0 \\). The joint distribution is modeled as a Markov chain \\( p_{\\theta}(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T)\\prod_{t=1}^T p_{\\theta}(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t) \\), which is commonly known as the reverse process. The initial \\( p(\\mathbf{x}_T) \\) is set to a known, tractable distribution such as the Gaussian distribution, while the transition distribution \\( p_{\\theta} \\), parameterized by \\( \\theta \\), is the trainable component of the model. Training a DDPM aims to approximate \\( p_{\\theta}(\\mathbf{x}_0) \\) to an empirical distribution \\( q(\\mathbf{x}_0) \\) defined by a collection of samples (e.g., images from the real world). To that end, training minimizes the cross-entropy between both distributions," + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.555, + 0.469, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0})} \\left[ - \\log p _ {\\theta} (\\mathbf {x} _ {0}) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.588, + 0.47, + 0.664 + ], + "angle": 0, + "content": "which is intractable as it requires marginalizing over the latent variables. Instead, a tractable distribution \\( q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) \\) is introduced and used as an approximation to the intractable true posterior \\( p(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) \\) to define the evidence lower bound (ELBO)," + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.674, + 0.47, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\log p _ {\\theta} (\\mathbf {x} _ {0}) \\geq \\mathbb {E} _ {\\mathbf {x} _ {1: T} \\sim q \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right)} \\left[ \\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0 : T}\\right)}{q \\left(\\mathbf {x} _ {1 : T} \\mid \\mathbf {x} _ {0}\\right)} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.469, + 0.793 + ], + "angle": 0, + "content": "where the expectation is approximated by Monte Carlo sampling. The lower bound is tight when the approximate posterior \\( q \\) equals the real posterior. Maximizing the ELBO over samples from \\( q(\\mathbf{x}_0) \\) minimizes the cross-entropy loss of Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The key difference between DDPMs and other latent variable models is that the approximate posterior \\( q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) \\) is fixed and not learnable. DDPMs model this distribution as a Markov chain \\( q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) = \\prod_{t=1}^{T} q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1}) \\), known as the forward process. The transition distribution \\( q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1}) \\) is chosen to be a tractable distribution that allows efficient sampling from \\( q(\\mathbf{x}_t \\mid \\mathbf{x}_0) \\) for" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "any \\(t\\). The only constraint in the design of a DDPM is that \\(q(\\mathbf{x}_T \\mid \\mathbf{x}_0) \\approx p(\\mathbf{x}_T)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.213 + ], + "angle": 0, + "content": "The original DDPM [19] modeled the transition distributions of the forward and the reverse processes as Gaussian with diagonal covariance matrices, and \\( p(\\mathbf{x}_T) \\) as a standard multivariate normal. However, these assumptions are inadequate when the elements of \\( \\mathbf{x}_0 \\) belong to discrete, unordered sets, as in the task of image segmentation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.222, + 0.75, + 0.238 + ], + "angle": 0, + "content": "3.2. Categorical diffusion model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.412 + ], + "angle": 0, + "content": "We now consider the denoising diffusion formulation to learn complex distributions of discrete image labelings. The observable variable \\(\\mathbf{x}_0\\in \\mathcal{L}^D\\) is categorical, where \\(D\\) is the number of pixels of the image and \\(\\mathcal{L} = \\{1,\\dots ,L\\}\\) is the set of discrete labels that can be assigned to each pixel. Following [21], we consider that all latent variables in \\(\\mathbf{x}_{1:T}\\) are also categorical and that the transition distributions for the forward and reverse processes are modeled as categorical distributions. For the forward process, the transition distribution acts element-wise over the previous state \\(\\mathbf{x}_{t - 1}\\) to produce the parameters of the distribution for \\(\\mathbf{x}_t\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.422, + 0.892, + 0.464 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\prod_ {d = 1} ^ {D} q \\left(\\mathbf {x} _ {t} [ d ] \\mid \\mathbf {x} _ {t - 1} [ d ]\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.566 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t[d]\\) indicates the label at time \\(t\\) and pixel \\(d\\). In the following discussion, we will use \\(x_{t}\\in \\mathcal{L}\\) to refer to the label of a single pixel \\(d\\), and we will drop the index \\(d\\) for clarity. The pixel-wise transition distribution \\(q(x_{t}\\mid x_{t - 1})\\) gives the element-wise probability of the next label given the previous label as," + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.576, + 0.892, + 0.61 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {C} \\left(x _ {t}; \\frac {\\beta_ {t}}{L} \\mathbf {1} + \\left(1 - \\beta_ {t}\\right) \\mathbf {e} _ {x _ {t - 1}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.727 + ], + "angle": 0, + "content": "where \\(\\mathbf{1} = (1,\\dots ,1)^T\\) \\(\\mathbf{e}_{\\ell}\\) is the one-hot encoding vector with 1 in position \\(\\ell\\) and 0 elsewhere, and the hyperparameter \\(\\alpha_{t} = 1 - \\beta_{t}\\in (0,1)\\) indicates the probability of keeping the label unchanged. \\(\\mathcal{C}(x;\\mathbf{p})\\) denotes the categorical distribution with parameter vector \\(\\mathbf{p}\\in [0,1]^L\\) . From the properties of categorical distributions, \\(\\mathcal{C}(x\\mid \\mathbf{p}) = \\mathbf{p}[x]\\) and \\(\\sum_{x}\\mathbf{p}[x] = 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.727, + 0.892, + 0.757 + ], + "angle": 0, + "content": "The transition distribution of the forward process can be composed as," + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.767, + 0.892, + 0.801 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {C} \\left(x _ {t}; \\frac {1 - \\bar {\\alpha} _ {t}}{L} \\mathbf {1} + \\bar {\\alpha} _ {t} \\mathbf {e} _ {x _ {0}}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.812, + 0.892, + 0.874 + ], + "angle": 0, + "content": "with \\(\\bar{\\alpha}_t = \\prod_{\\tau=1}^t \\alpha_\\tau\\), which enables efficient sampling of elements from the Markov chain at any location \\(t\\). Finally, the posterior of the transition distribution can be computed with the previous formulas by applying Bayes rule," + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) = \\mathcal {C} \\left(x _ {t - 1}; \\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1121" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.468, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.199, + 0.473, + 0.34 + ], + "angle": 0, + "content": "Figure 2: Illustration of the reverse process of our method. The conditional categorical diffusion model (CCDM) receives as input an image \\(I\\) and a categorical label map \\(\\mathbf{x}_T^{(i)}\\) sampled from the categorical uniform noise. The reverse process of the CCDM generates a label map \\(\\mathbf{x}_0^{(i)}\\), which is a sample from the learned distribution \\(p(\\mathbf{x}_0 \\mid I)\\). When repeated for \\(N\\) samples, we obtain an empirical approximation to the multimodal label distribution for the image \\(I\\), learned from the annotations of multiple expert raters." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.116, + 0.38 + ], + "angle": 0, + "content": "with," + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.389, + 0.479, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right) = \\frac {1}{\\tilde {\\pi}} \\left(\\frac {\\beta_ {t}}{L} \\mathbf {1} + \\alpha_ {t} \\mathbf {e} _ {x _ {t}}\\right) \\odot \\left(\\frac {1 - \\bar {\\alpha} _ {t - 1}}{L} \\mathbf {1} + \\bar {\\alpha} _ {t - 1} \\mathbf {e} _ {x _ {0}}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.435, + 0.456, + 0.452 + ], + "angle": 0, + "content": "and \\(\\tilde{\\pi} = \\frac{1 - \\bar{\\alpha}_t}{L} +\\bar{\\alpha}_t\\cdot \\delta_{xt}^{x_0}\\), where \\(\\delta\\) is the Kronecker delta." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.452, + 0.469, + 0.48 + ], + "angle": 0, + "content": "The transition distribution of the reverse process is also an element-wise categorical distribution," + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.491, + 0.47, + 0.533 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\prod_ {d = 1} ^ {D} \\mathcal {C} \\left(x _ {t - 1}; \\hat {\\mathbf {p}} _ {t - 1}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.679 + ], + "angle": 0, + "content": "where \\( x_{t - 1} = \\mathbf{x}_{t - 1}[d] \\) and \\( \\hat{\\mathbf{p}}_{t - 1} \\) are the label and the estimated parameter vector, respectively, at pixel \\( d \\). Unlike the forward process, the parameter vector for the pixel \\( d \\) is not computed considering only the element \\( d \\) of \\( \\mathbf{x}_t \\). Instead, it is modeled as a function \\( f: \\mathcal{L}^D \\to [0,1]^{D\\times L} \\) that incorporates context by considering the entire label map \\( \\mathbf{x}_t \\) to produce a collection of \\( D \\) probability distributions for \\( \\mathbf{x}_{t - 1} \\), which we refer to as \\( \\hat{\\mathbf{P}}_{t - 1} \\in [0,1]^{D\\times L} \\) with \\( \\hat{\\mathbf{p}}_{t - 1} = \\hat{\\mathbf{P}}_{t - 1}[d] \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.47, + 0.8 + ], + "angle": 0, + "content": "While it is possible to use a neural network to estimate \\(\\hat{\\mathbf{P}}_{t - 1}\\), Ho et al. [19] suggested that a consistent output space for the network led to enhanced performance. Following this idea, we train a network \\(f_{\\theta}\\), parameterized by \\(\\theta\\), to compute \\(\\hat{\\mathbf{P}}_0 = f_\\theta (\\mathbf{x}_t,t)\\in [0,1]^{D\\times L}\\) by receiving a label map \\(\\mathbf{x}_t\\) and the step \\(t\\). We then transform the parameter vector for each pixel, \\(\\hat{\\mathbf{p}}_0 = \\hat{\\mathbf{P}}_0[d]\\) to the parameter vector \\(\\hat{\\mathbf{p}}_{t - 1}\\) for the same pixel of \\(\\mathbf{x}_{t - 1}\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.81, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {C} \\left(x _ {t - 1}; \\hat {\\mathrm {p}} _ {t - 1}\\right) = (9) \\\\ = \\sum_ {x _ {0}} q \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) \\cdot \\mathcal {C} \\left(x _ {0}; \\hat {\\mathbf {p}} _ {0}\\right) (10) \\\\ = \\sum_ {x _ {0}} \\mathcal {C} \\left(x _ {t - 1}; \\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right)\\right) \\cdot \\mathcal {C} \\left(x _ {0}; \\hat {\\mathbf {p}} _ {0}\\right), (11) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.586, + 0.105 + ], + "angle": 0, + "content": "from which," + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.111, + 0.892, + 0.143 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {p}} _ {t - 1} = \\sum_ {x _ {0} \\in \\mathcal {L}} \\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right) \\cdot \\hat {\\mathbf {p}} _ {0} \\left[ x _ {0} \\right], \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.15, + 0.894, + 0.257 + ], + "angle": 0, + "content": "where we have omitted the pixel indices \\(d\\) for clarity. This transformation is not necessary when \\(t = 1\\), as then \\(\\hat{\\mathbf{p}}_{t - 1} = \\hat{\\mathbf{p}}_0\\) computed by \\(f_{\\theta}\\). It is also possible to perform this computation in parallel for every pixel to efficiently obtain \\(\\hat{\\mathbf{P}}_{t - 1}\\). Note that the result of Eq. (12) differs from the parameter vector computed in [21], where the ill-defined expression \\(\\hat{\\mathbf{p}}_{t - 1} = \\pi (x_t,\\hat{\\mathbf{x}}_0)\\) is employed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.263, + 0.788, + 0.28 + ], + "angle": 0, + "content": "3.3. Conditional categorical diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.892, + 0.391 + ], + "angle": 0, + "content": "In stochastic segmentation, the label map \\(\\mathbf{x}_0\\) for an image \\(I\\) is modeled by a distribution \\(q(\\mathbf{x}_0\\mid I)\\). This distribution is often too complex to be properly approximated as a product of pixel-wise categorical distributions. We use a conditional categorical diffusion model \\(p(\\mathbf{x}_0\\mid I)\\) (CCDM) to model the potentially complex interactions between labels and pixels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.893, + 0.528 + ], + "angle": 0, + "content": "When conditioning the categorical diffusion model on an image, the forward process remains unchanged, \\( q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0, I) = q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) \\), as any latent variable is conditionally independent of the image given any previous variable. On the other hand, the reverse process needs to incorporate the dependency on the image in its transition distribution, \\( p_{\\theta}(\\mathbf{x}_{0:T} \\mid I) = p(\\mathbf{x}_T \\mid I) \\prod_{t=1}^{T} p_{\\theta}(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t, I) \\). In practice, this dependency is enforced by an additional input to the neural network \\( f_{\\theta}(\\mathbf{x}_t, t, I) \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.536, + 0.605, + 0.552 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.894, + 0.618 + ], + "angle": 0, + "content": "Training is performed by maximizing the ELBO of Eq. (2). Reorganizing terms and distributing expectations for variance reduction, we express the ELBO as a sum of three terms:" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.627, + 0.892, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log p _ {\\theta} (\\mathbf {x} _ {0} \\mid I) \\geq \\\\ \\mathbb {E} _ {\\mathbf {x} _ {1} \\sim q \\left(\\mathbf {x} _ {1} \\mid \\mathbf {x} _ {0}\\right)} [ \\log p _ {\\theta} (\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {1}, I) ] (13) \\\\ - \\sum_ {t = 2} ^ {T} \\mathbb {E} _ {\\mathbf {x} _ {t} \\sim q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right)} [ K L (q \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {x} _ {0}\\right) \\| p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, I\\right)) ] (14) \\\\ - K L \\left(q \\left(\\mathbf {x} _ {T} \\mid \\mathbf {x} _ {0}\\right) \\| p \\left(\\mathbf {x} _ {T} \\mid I\\right)\\right). (15) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.84 + ], + "angle": 0, + "content": "The first two terms can be optimized by standard gradient ascent. We approximate the expectations with Monte Carlo sampling with a single sample. The sum over the time variable \\( t \\) is also approximated by a single uniform sample over \\( \\{1, \\dots, T\\} \\). The KL divergence of the second term is the sum of pixel-wise KL divergences," + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.847, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\nK L (q \\| p) = \\sum_ {d = 1} ^ {D} K L \\left(q \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) \\| p _ {\\theta} \\left(x _ {t - 1} \\mid \\mathbf {x} _ {t}, I\\right)\\right), \\tag {16}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1122" + } + ], + [ + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.089, + 0.473, + 0.386 + ], + "angle": 0, + "content": "Algorithm 1 Training a CCDM with \\(T\\) steps \nRequire: Training data expressed as the empirical distribu-. tion \\(q(\\mathbf{x}_0,I) = q(\\mathbf{x}_0\\mid I)q(I)\\) repeat t \\~ Uniform( \\(\\{1,\\dots,T\\}\\) 1 \\(\\begin{array}{rl} & {\\boldsymbol {I}\\sim q(\\boldsymbol {I})}\\\\ & {\\mathbf{x}_0\\sim q(\\mathbf{x}_0\\mid \\boldsymbol {I})}\\\\ & {\\mathbf{x}_t\\sim q(\\mathbf{x}_t|\\mathbf{x}_0)}\\\\ & {\\hat{\\mathbf{P}}_0\\leftarrow f_\\theta (\\mathbf{x}_t,I,t)} \\end{array}\\) ▷ shape \\(D\\times L\\) if \\(t > 1\\) then \\(\\triangleright\\) Pixel-wise application of Eq. (12) \\(\\hat{\\mathbf{p}}_{t - 1}\\gets \\sum_{x_0\\in \\mathcal{L}}\\pi (x_t,x_0)\\cdot \\hat{\\mathbf{p}}_0[x_0]\\quad \\triangleright\\) shape L \\(\\triangleright\\) Compute KL with Eq. (8) and (16) \\(\\ell \\leftarrow KL(q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0)\\| p_\\theta (\\mathbf{x}_{t - 1}|\\mathbf{x}_t,I))\\) else \\(\\ell \\leftarrow -\\sum_d\\log \\mathcal{C}(x_0\\mid \\hat{\\mathbf{P}}_0[d])\\) end if \\(\\theta \\leftarrow \\theta -\\nabla_{\\theta}\\ell\\) ▷ Gradient descent until converged" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.458 + ], + "angle": 0, + "content": "where the parameter vectors of distributions \\(q\\) and \\(p\\) are computed with Eqs. (7) and (12), respectively. Alg. 1 shows the complete training procedure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.52 + ], + "angle": 0, + "content": "The third term of Eq. (15) does not depend on the learnable parameters \\(\\theta\\) and is ignored during training. It is optimized by the design of the categorical diffusion model. Since the forward process converges as" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.529, + 0.47, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\lim _ {t \\rightarrow \\infty} q \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {C} \\left(x; \\frac {1}{L}\\right), \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.573, + 0.462, + 0.589 + ], + "angle": 0, + "content": "we fix \\(p(\\mathbf{x}_T\\mid I)\\) to the element-wise uniform distribution," + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.599, + 0.47, + 0.632 + ], + "angle": 0, + "content": "\\[\np \\left(x _ {T} \\mid I\\right) = p \\left(x _ {T}\\right) = \\mathcal {C} \\left(x _ {T}; \\frac {1}{L}\\right). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.671 + ], + "angle": 0, + "content": "This ensures that \\( p(\\mathbf{x}_T \\mid I) \\approx q(\\mathbf{x}_T \\mid \\mathbf{x}_0) \\), making the third term of the ELBO close to zero." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.779 + ], + "angle": 0, + "content": "At inference, the CCDM samples from \\( p(\\mathbf{x}_0 \\mid I) \\) to generate label maps for a given image \\( I \\), which is achieved by traversing the Markov chain of the reverse process as outlined in Alg. 2 and illustrated in Fig. 2. To minimize the noise of the generated label maps, the CCDM selects the label with maximum probability instead of sampling from \\( \\mathcal{C}(x_0 \\mid \\hat{\\mathbf{p}}_0) \\) in the final step." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.25, + 0.803 + ], + "angle": 0, + "content": "3.5. Architecture of \\( f_{\\theta} \\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As described above, the neural network \\( f_{\\theta} \\) receives a label map \\( \\mathbf{x}_t \\), a time step \\( t \\), and an image \\( I \\) to estimate the probability parameters for \\( \\mathbf{x}_0 \\). Its base design is a U-Net-like architecture [13] with self-attention modules at the three innermost layers of the encoder and the decoder [13]. The network processes the input label map represented as" + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.842, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 Inference from a CCDM with \\(T\\) steps" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.11, + 0.892, + 0.324 + ], + "angle": 0, + "content": "Require: Input image \\(I\\), \\(f_{\\theta}\\) a network trained with Alg. 1 \n\\(\\mathbf{x}_T \\sim \\mathcal{C}^D(x_T; \\frac{1}{L})\\) \n\\(\\mathbf{x}_{\\text{prev}} \\gets \\mathbf{x}_T \\quad \\triangleright\\) Stores interm. and final prediction \nfor \\(t = T, \\dots, 1\\) do \n\\(\\hat{\\mathbf{P}}_0 \\gets f_{\\theta}(\\mathbf{x}_{\\text{prev}}, I, t)\\) \nif \\(t > 1\\) then \n\\(\\triangleright\\) Pixel-wise application of Eq. (12) \n\\(\\hat{\\mathbf{p}}_{t-1} \\gets \\sum_{x_0 \\in \\mathcal{L}} \\pi(x_t, x_0) \\cdot \\hat{\\mathbf{p}}_0[x_0]\\) \n\\(\\mathbf{x}_{\\text{prev}} \\sim \\prod_d \\mathcal{C}(x_{t-1} | \\hat{\\mathbf{p}}_{t-1})\\) \nelse \n\\(\\triangleright\\) Final prediction \n\\(\\mathbf{x}_{\\text{prev}} \\gets \\arg \\max_{x_0 \\in \\mathcal{L}} \\hat{\\mathbf{P}}_0[:, x_0]\\) \nend if \nend for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.575 + ], + "angle": 0, + "content": "a binary tensor with \\(L\\) channels encoding the label of each pixel as a one-hot vector. Parameters of the network are shared for all values of \\(t\\). The step variable \\(t\\) is encoded with the standard transformer sinusoidal position embedding [19] and concatenated as additional channels to the input tensor and to the feature maps of intermediate layers. Similarly, information from the input image \\(I\\) is presented to the network as raw pixel values concatenated to the input tensor as additional channels. In some experiments we used a pre-trained transformer architecture Dino-ViT [5] to extract informative visual features from the image \\(I\\). In those cases, the extracted features were concatenated to the feature map of the third level of the U-Net encoder, which corresponds to a spatial shape equal to \\(\\frac{1}{8}\\) the shape of the input image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.634, + 0.604 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.613, + 0.892, + 0.658 + ], + "angle": 0, + "content": "In all our experiments, we set \\( T = 250 \\) and the collection of \\( \\beta_{t} \\) are set following the cosine schedule proposed in [36]. We evaluate our method on two tasks described below." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.667, + 0.843, + 0.683 + ], + "angle": 0, + "content": "4.1. Segmentation with multiple annotations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Dataset The Lung Image Database Consortium (LIDC) [2] binary segmentation dataset consists of 1'018 three dimensional chest CT scans of patients with lung cancer. Lung nodules of each volume are annotated by four expert raters from a pool of 12, yielding large differences in annotations in some cases. We extract nodule-centered slices from the CT volumes and treat each slice as an independent image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.902 + ], + "angle": 0, + "content": "While LIDC is the standard benchmark of stochastic segmentation methods to date (e.g. [4, 15, 23, 25, 28, 29, 35, 43, 53, 54]), experimental configurations (pre-processing, training/validation/test splits, metrics) vastly differ across the literature. We conduct our experiments on the two most prominent LIDC splits and report results on both separately." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1123" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.269 + ], + "angle": 0, + "content": "
MethodLIDCv1LIDCv2
GED16GED32GED50GED100HM-IoU16HM-IoU32GED16GED50GED100HM-IoU16
Prob. Unet [28]0.310±0.01-0.303±0.01+-0.252±0.004†0.552±0.00-0.548±0.00+0.320±0.030‡-0.252±‡0.500±0.030‡
HProb. Unet [29]0.270±0.01----0.530±0.01--0.270±0.010‡--0.530±0.01
PhiSeg [4]0.262±0.00-0.247±0.00+-0.224±0.004†0.586±0.00-0.595±0.00+----
SSN [35]0.259±0.00-0.243±0.01+-0.225±0.0020.558±0.00-0.555±0.01+----
cFlow [43]-0.225±0.01+---0.584±0.00+----
CAR [25]---0.228±0.009--0.264±0.0020.248±0.0040.243±0.0040.592±0.005
JProb. Unet [53]-0.206±0.00---0.647±0.010.262±0.00--0.585±0.00
PixelSeg [52]0.243±0.01---0.614±0.00-0.260±0.00--0.587±0.01
MoSE [15]0.218±0.003-0.195±0.0020.189±0.0020.624±0.004-----
AB [9]0.213±0.0010.196±0.0020.193±0.002-0.614±0.0010.619±0.001----
CIMD [38]0.234±0.0050.218±0.0050.210±0.005-0.587±0.0010.592±0.002----
CCDM (ours)0.212±0.0020.194±0.0010.187±0.0020.183±0.0020.623±0.0020.631±0.0020.239±0.0030.216±0.0030.210±0.0030.598±0.001
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.291, + 0.893, + 0.352 + ], + "angle": 0, + "content": "Table 1: Quantitative results on LIDCv1 and LIDCv2, with the methods ordered by year. Bold and underlined indicate best and second best per column, respectively. Our results are over 3 seeds. For GED, lower is better; for HM-IoU, higher is better. No method, including ours, uses pre-trained weights. Results for CIMD [38] and AB [9] are ours. All other scores are taken from their original papers, except \\((^{+})\\) from [53], \\((-)\\) from [53], \\((\\dagger)\\) from [35], \\(\\ddagger\\) from [25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.453 + ], + "angle": 0, + "content": "The first, referred to as LIDCv1, is used in [4, 15, 35, 53]. LIDCv1 comprises 15'096 slices, divided into training, validation, and testing sets with the ratio \\(60:20:20\\). The second, LIDCv2, is used in [25, 28] and contains 12'816 images with the ratio \\(70:15:15\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.609 + ], + "angle": 0, + "content": "Metrics We measure the performances with the Generalised Energy Distance (GED) and the Hungarian-Matched Intersection over Union (HM-IoU) [15, 25, 29]. Both metrics measure the difference between the distributions of generated and ground-truth label maps. We denote the metrics computed with \\( n \\) samples using a subscript, i.e., \\( \\mathrm{GED}_n \\) and \\( \\mathrm{HM - IoU}_n \\), and we set \\( n \\) to common values found in the literature. Note that higher number of samples yield more precise estimates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Baselines We compare our approach to eleven recent stochastic segmentation methods: probabilistic U-Net (Prob. Unet) [28], hierarchical probabilistic U-Net (HProb. Unet) [29], PhiSeg [4], stochastic segmentation network (SSN) [35], conditional normalizing flow (cFlow) [43], calibrated adversarial refinement (CAR) [25], joint probabilistic U-Net (JProb. Unet) [53], PixelSeg [52], mixture of stochastic experts (MoSE) [15], analog bits (AB) [9], and collectively intelligent medical diffusion (CIMD) [38]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Following standard practice, we use random horizontal and vertical flipping and random rotations of \\(0^{\\circ}\\), \\(90^{\\circ}\\), \\(180^{\\circ}\\) and \\(270^{\\circ}\\) for data augmentation. The resolution of the input images is \\(128 \\times 128\\). We trained our method with the Adam optimizer [26] until convergence of the GED metric on the validation set, a polynomial learning rate scheduling starting from \\(1e^{-4}\\) and ending with \\(1e^{-6}\\), and batch size of 64. We applied Polyak averaging with \\(\\alpha = 0.99995\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.378, + 0.829, + 0.394 + ], + "angle": 0, + "content": "4.2. Segmentation with a single annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.477 + ], + "angle": 0, + "content": "We also evaluate our method with Cityscapes, a classical multi-class segmentation dataset where each image is annotated with a single label map. It comprises 2'975 RGB images of urban scenes for training and 500 images for validation, with each image labeled using 19 possible classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.536 + ], + "angle": 0, + "content": "We compare our approach to several established baselines using the validation set: DeepLabv3 [7], HRNet [47], and UPerNet [51], with both ResNet [18] and Swin [32] backbones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Besides our standard method, which performs image conditioning by concatenating the raw pixel values as channels of the input tensor, we also included in our comparison a variant CCDM-Dino which leverages pre-trained DinoViT features [5] as additional conditioning concatenated to intermediate feature maps of our model's encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Experiments are conducted separately for two different image resolutions: \\(128 \\times 256\\) and \\(256 \\times 512\\). For all reported methods, we first resize the images to a fixed resolution and then apply color jittering, random flipping, and standard ImageNet intensity normalization as data augmentation. All baselines are trained for 500 epochs with a batch size of 32, with optimizers, learning rate schedules, and weight decay settings as reported in their respective publications (reported in detail in the supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Our method was trained for 800 epochs with a batch size of 32 at \\(128 \\times 256\\) and of 16 at \\(256 \\times 512\\), using the Adam optimizer [26] with a learning rate of \\(1e^{-4}\\) linearly decayed to \\(1e^{-6}\\). We applied Polyak averaging with \\(\\alpha = 0.999\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Performance is measured with the mean intersection-over-union (mIoU). Unlike GED and HM-IoU, the metric mIoU is incompatible with multiple label maps per image. During inference, CCDM generates multiple label maps per image that are subsequently fused into a single label map for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1124" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "performance assessment. We found that fusing by averaging the predicted probabilities resulted in superior performances compared to fusing by majority vote." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.152, + 0.165, + 0.168 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.176, + 0.161, + 0.192 + ], + "angle": 0, + "content": "5.1. LIDC" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.471, + 0.276 + ], + "angle": 0, + "content": "We report performances on LIDCv1 and LIDCv2 in Tab. 1 and qualitative results in Fig. 3. Due to the lack of consistent evaluation protocols, we use a total of 10 metrics, thereby covering all the baselines and allowing for direct comparisons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.471, + 0.488 + ], + "angle": 0, + "content": "Our CCDM reaches the best performance for eight out of the ten metrics, despite its relatively small size, with 9M parameters compared to, e.g., the 42M parameters of MoSE. CCDM also outperforms recent continuous diffusion models for segmentation, including AB [9] (9M parameters) and CIMD [38] (24M parameters). On HM-IoU16, the CCDM has a lower mean performance than MoSE by 0.001, but with only half the standard deviation. The JProb. Unet reaches a higher HM-IoU32 than all other methods, despite being considerably worse for GED32 than our CCDM. Furthermore, on LIDCv2, the JProb. Unet achieves only the third-best score on GED16, and fourth-best on HM-IoU16. This result indicates how comparing results obtained on different LIDC versions with each other can be misleading." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.579 + ], + "angle": 0, + "content": "Fig. 3 presents qualitative results from our method. In columns (g)-(l), we see that our CCDM generates a distribution of samples that captures the annotation variability created by the four expert raters. Further, as seen in the bottom example, the CCDM also generates empty samples according to the annotations (b)-(e)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.471, + 0.777 + ], + "angle": 0, + "content": "Reduced number of time steps for sampling: During inference, traversing the \\(T\\) steps of the reverse process makes sampling from DDPMs slow. A straightforward solution [36] involves traversing only a subset of nodes of the reverse process, \\(\\{\\mathbf{x}_{k\\tau} : \\tau \\in \\{0, \\dots, T / k\\}\\}\\), reducing the number of steps by a factor \\(k\\). This technique accelerates inference at the expense of reduced performance. To illustrate the trade-offs between performance and speed, Fig. 5 presents the evolution of \\(\\mathrm{GED}_{16}\\) and \\(\\mathrm{HM-IoU}_{16}\\) as the number of inference steps is reduced. As expected, CCDMs perform best when the number of training and inference steps are equal, but a reasonable increase in speed without a large sacrifice in performance is possible." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.197, + 0.803 + ], + "angle": 0, + "content": "5.2. Cityscapes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Experimental comparisons on Cityscapes are presented in Tab. 2, and qualitative examples are provided in Fig. 4. Experiments at \\(128 \\times 256\\) demonstrate that CCDM-Dino outperforms all other methods, even when only a single sample is used. CCDM-raw also remains competitive, being outperformed only by one baseline (UPerNet+Swin" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.878, + 0.276 + ], + "angle": 0, + "content": "
MethodmIoU final (best)
ArchitectureBackbone#params128 × 256256 × 512
DeepLabv3 [7]ResNet50 (✓)39m43.4 (44.1)58.6 (59.2)
DeepLabv3 [7]ResNet101 (✓)58m43.8 (45.5)59.2 (59.8)
UPerNet [51]ResNet101 (✓)83m45.5 (47.1)60.7 (61.2)
HRNet [47]w48v2 (✓)70m48.2 (49.5)63.3 (64.2)
UPerNet [32]Swin-Tiny (✓)58m54.2 (55.9)65.5 (66.0)
CCDM (ours)-
samples=130m53.260.3
samples=530m55.462.0
samples=1030m56.262.4
CCDM (ours)Dino ViT-S (†)
samples=130m + 20M55.564.0
samples=530m + 20M56.965.4
samples=1030m + 20M57.365.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.286, + 0.894, + 0.438 + ], + "angle": 0, + "content": "Table 2: Results on Cityscapes-val for resolutions \\(128 \\times 256\\) and \\(256 \\times 512\\). **Bold** and **underlined** indicate best and second best per column, respectively. (\\(\\checkmark\\)) and (\\(\\dagger\\)) indicate supervised and self-supervised pretraining of the backbone, respectively. Gray indicates pretrained, non-finetuned parameters. We report final performance for our method and baselines. For the latter we also provide best achieved performance during training (in parenthesis). For CCDM methods, the field samples indicate the number of generated samples for label map fusion, as explained in Sect 4.2." + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.473, + 0.838, + 0.544 + ], + "angle": 0, + "content": "
CCDM CapacitymIoU(128 × 256)
#paramsUNet Levelssamples=1samples=5samples=10
5.4M437.839.740.6
7.5M544.748.348.5
22M451.654.053.6
30M553.255.456.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.554, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Table 3: Effect of increasing CCDM capacity (without feature conditioning)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.894, + 0.816 + ], + "angle": 0, + "content": "Tiny), despite using only between \\(36\\%\\) and \\(51\\%\\) of the parameters of other models. Similarly, at \\(256 \\times 512\\), CCDM-Dino outperforms four of the baselines with a single sample, lags behind UPerNet+Swin-Tiny only by 0.1 percent points with 5 samples, and outperforms all baselines with 10 samples. As expected, averaging across more samples improves performance for both CCDM-raw and CCDM-Dino, albeit with diminishing gains. Furthermore, the addition of Dino features boosts single-sample performance by 2.3 percent points at \\(128 \\times 256\\), and 3.7 percent points at \\(256 \\times 512\\), hinting the greater value of adding feature conditioning for generating segmentation at a higher resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "CCDM Capacity: Tab. 3(b) demonstrates the effect of increasing the capacity of CCDM. Using more U-Net feature levels, and increasing the number of parameters by doubling the number of channels per level, increases the performance regardless of the number of samples used for inference." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1125" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.872, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Figure 3: Qualitative results on four LIDC images with our method. (a) shows the image, (b)-(e) its four labels, (f) the mean prediction of our CCDM over six predictions, and (g)-(l) six individual predictions." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.379, + 0.885, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.613, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Figure 4: Qualitative comparisons on Cityscapes. All methods are trained and tested at a resolution of \\(256 \\times 512\\). Our method produces structures with greater visual realism than other baselines. This is especially noticeable inside the marked regions." + }, + { + "type": "image", + "bbox": [ + 0.15, + 0.667, + 0.396, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.807, + 0.471, + 0.853 + ], + "angle": 0, + "content": "Figure 5: LIDC GED and HM-IoU versus the number of sampling steps on LIDC. Evaluated on 500 random test images using 16 samples each, over 3 seeds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.669, + 0.619, + 0.685 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduced conditional categorical diffusion models (CCDMs) that are capable of effectively modeling pixel-level semantic distributions. Notably, and contrary to standard deterministic segmentation approaches, our model can produce diverse samples given an input image, thereby capturing the aleatoric uncertainty. Our method learns a multimodal label distribution of segmentations, induced by annotations from multiple expert raters, for which it achieves state-of-the-art results on a challenging medical imaging dataset, LIDC. Additionally, we demonstrate that it can achieve competitive performance on a standard multi-class semantic segmentation benchmark, Cityscapes, by outperforming several established, heavily engineered baselines" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1126" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.375, + 0.107 + ], + "angle": 0, + "content": "despite using significantly fewer parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.47, + 0.274 + ], + "angle": 0, + "content": "One limitation of our method is the requirement of several iterations for producing a sample, which is a common shortcoming of diffusion models. Accelerating sampling constitutes a crucial research direction, orthogonal to the present work. Finally, resolution scaling remains notoriously difficult for diffusion models, with successful examples relying on massive computational resources to train cascades of models that gradually increase resolution [20, 42] or operate on the latent space of existing embedding methods for continuous data (e.g. images) [39] that are not available for categorical data." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.29, + 0.245, + 0.308 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.316, + 0.471, + 0.362 + ], + "angle": 0, + "content": "This work was partially funded by the University of Bern, Swiss National Science Foundation Grants #320030-188591, #200021-192285 and #200021-191983." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.378, + 0.175, + 0.394 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.404, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[1] Tomer Amit, Eliya Nachmani, Tal Shaharbany, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.447, + 0.472, + 0.544 + ], + "angle": 0, + "content": "[2] Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al. The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Medical physics, 38(2):915-931, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.47, + 0.601 + ], + "angle": 0, + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.604, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[4] Christian F Baumgartner, Kerem C Tezcan, Krishna Chaitanya, Andreas M Hotker, Urs J Muehlematter, Khoschy Schawkat, Anton S Becker, Olivio Donati, and Ender Konukoglu. Phiseg: Capturing uncertainty in medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part II 22, pages 119-127. Springer, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.731, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.802, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for seman-" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.404, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.123, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[8] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.195, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[9] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.252, + 0.892, + 0.321 + ], + "angle": 0, + "content": "[10] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems, 34:9355-9366, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[11] Max Cohen, Guillaume Quispe, Sylvain Le Corff, Charles Ollion, and Eric Moulines. Diffusion bridges vector quantized variational autoencoders. arXiv preprint arXiv:2202.04895, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. arXiv preprint arXiv:2209.04747, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.454 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. CoRR, abs/2105.05233, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[14] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[15] Zhitong Gao, Yuong Chen, Chuyu Zhang, and Xuming He. Modeling multimodal aleatoric uncertainty in segmentation with mixture of stochastic expert. arXiv preprint arXiv:2212.07328, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[16] Jiaqi Gu, Hyoukjun Kwon, Dilin Wang, Wei Ye, Meng Li, Yu-Hsin Chen, Liangzhen Lai, Vikas Chandra, and David Z Pan. Multi-scale high-resolution vision transformer for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12094-12103, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[17] Adam W Harley, Konstantinos G Derpanis, and Iasonas Kokkinos. Segmentation-aware convolutional networks using local attention masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5038-5047, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 9" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1127" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[21] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34:12454-12465, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.244 + ], + "angle": 0, + "content": "[22] Minghui Hu, Yujie Wang, Tat-Jen Cham, Jianfei Yang, and Ponnuthurai N Suganthan. Global context with discrete diffusion in vector quantised modelling for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11502-11511, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[23] Shi Hu, Daniel Worrall, Stefan Knegt, Bas Veeling, Henkjan Huisman, and Max Welling. Supervised uncertainty quantification for segmentation with multiple annotations. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22, pages 137–145. Springer, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[24] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. arXiv preprint arXiv:2203.15636, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[25] Elias Kassapis, Georgi Dikov, Deepak K Gupta, and Cedric Nugteren. Calibrated adversarial refinement for stochastic semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7057-7067, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference for Learning Representations, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[27] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[28] Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic u-net for segmentation of ambiguous images. Advances in neural information processing systems, 31, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.705 + ], + "angle": 0, + "content": "[29] Simon AA Kohl, Bernardino Romera-Paredes, Klaus H Maier-Hein, Danilo Jimenez Rezende, SM Eslami, Pushmeet Kohli, Andrew Zisserman, and Olaf Ronneberger. A hierarchical probabilistic u-net for modeling multi-scale ambiguities. arXiv preprint arXiv:1905.13077, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[30] Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. Tabddpm: Modelling tabular data with diffusion models. arXiv preprint arXiv:2209.15421, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[31] Lielei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep hierarchical semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1246-1257, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[33] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Pro-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[35] Miguel Monteiro, Loic Le Folgoc, Daniel Coelho de Castro, Nick Pawlowski, Bernardo Marques, Konstantinos Kamnitsas, Mark van der Wilk, and Ben Glocker. Stochastic segmentation networks: Modelling spatially correlated aleatoric uncertainty. Advances in Neural Information Processing Systems, 33:12756-12767, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[36] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 5.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[37] Di Qiu and Lok Ming Lui. Modal uncertainty estimation via discrete latent representation. arXiv preprint arXiv:2007.12858, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[38] Aimon Rahman, Jeya Maria Jose Valanarasu, Ilker Hacihaliloglu, and Vishal M Patel. Ambiguous medical image segmentation using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11536-11546, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[43] Raghavendra Selvan, Frederik Faye, Jon Middleton, and Akshay Pai. Uncertainty quantification in medical image segmentation with normalizing flows. In Machine Learning in Medical Imaging: 11th International Workshop, MLMI 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4, 2020, Proceedings 11, pages 80-90. Springer, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1128" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[45] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[46] MM Amaan Valiuddin, Christiaan GA Viviers, Ruud JG van Sloun, Peter HN de With, and Fons van der Sommen. Improving aleatoric uncertainty quantification in multi-annotated medical image segmentation with normalizing flows. In Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, and Perinatal Imaging, Placental and Preterm Image Analysis: 3rd International Workshop, UNSURE 2021, and 6th International Workshop, PIPPI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 3, pages 75–88. Springer, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. TPAMI, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[48] Julia Wolleb, Robin Sandkühler, Florentin Bieder, Philippe Valmaggia, and Philippe C Cattin. Diffusion models for implicit image segmentation ensembles. arXiv preprint arXiv:2112.03145, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[49] Junde Wu, Huihui Fang, Yu Zhang, Yehui Yang, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. arXiv preprint arXiv:2211.00611, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[50] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, and Yanwu Xu. Medsegdiff-v2: Diffusion based medical image segmentation with transformer. arXiv preprint arXiv:2301.11798, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[51] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[52] Wei Zhang, Xiaohong Zhang, Sheng Huang, Yuting Lu, and Kun Wang. Pixelseg: Pixel-by-pixel stochastic semantic segmentation for ambiguous medical images. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4742-4750, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[53] Wei Zhang, Xiaohong Zhang, Sheng Huang, Yuting Lu, and Kun Wang. A probabilistic model for controlling diversity and accuracy of ambiguous medical image segmentation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4751-4759, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[54] Yifan Zhang, Bo Pang, and Cewu Lu. Semantic segmentation by early region proxy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1258-1268, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[55] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[56] Roland S Zimmermann, Lukas Schott, Yang Song, Benjamin A Dunn, and David A Klindt. Score-based generative classifiers. arXiv preprint arXiv:2110.00473, 2021. 1" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1129" + } + ] +] \ No newline at end of file diff --git a/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/36fc9ee6-a13a-40ed-bd76-54644b9d43ea_origin.pdf b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/36fc9ee6-a13a-40ed-bd76-54644b9d43ea_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3cabc2942e36022f627dd91e2f2a7f5228424f21 --- /dev/null +++ b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/36fc9ee6-a13a-40ed-bd76-54644b9d43ea_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:622dfe9bc29a2a147457b81e6771b1efeda73ec7f8662c471e697d390c102c81 +size 2344130 diff --git a/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/full.md b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..44cc6e1c2602e37bcd1dba119272543969e4837e --- /dev/null +++ b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/full.md @@ -0,0 +1,366 @@ +# Stochastic Segmentation with Conditional Categorical Diffusion Models + +Lukas Zbinden* + +Adrian Thomas Huber + +Lars Doorenbos* + +Raphael Sznitman + +Theodoros Pissas + +Pablo Márquez-Neila + +University of Bern, Bern, Switzerland + +{lukas.zbinden, lars.doorenbos, theodoros.pissas, raphael.sznitman, pablo.marquez}@unibe.ch + +# Abstract + +Semantic segmentation has made significant progress in recent years thanks to deep neural networks, but the common objective of generating a single segmentation output that accurately matches the image's content may not be suitable for safety-critical domains such as medical diagnostics and autonomous driving. Instead, multiple possible correct segmentation maps may be required to reflect the true distribution of annotation maps. In this context, stochastic semantic segmentation methods must learn to predict conditional distributions of labels given the image, but this is challenging due to the typically multimodal distributions, high-dimensional output spaces, and limited annotation data. To address these challenges, we propose a conditional categorical diffusion model (CCDM) for semantic segmentation based on Denoising Diffusion Probabilistic Models. Our model is conditioned to the input image, enabling it to generate multiple segmentation label maps that account for the aleatoric uncertainty arising from divergent ground truth annotations. Our experimental results show that CCDM achieves state-of-the-art performance on LIDC, a stochastic semantic segmentation dataset, and outperforms established baselines on the classical segmentation dataset Cityscapes. + +# 1. Introduction + +Semantic segmentation has significantly progressed in recent years due to powerful deep neural networks. For most methods, the key objective is to generate a single segmentation output that accurately matches the image's content. However, this may not be suitable for safety-critical domains such as medical diagnostics and autonomous driving, as images in these applications often suffer from inherent ambiguity or annotations that have differences in opinion. In these cases, generating a single coherent segmentation may be hopeless to fully describe the set of correct + +![](images/2860b575fdb0e7c4132b9ff61a9052200b804cf1558791aad29fd5ee10b52049.jpg) +Figure 1: Examples from the LIDC dataset, where expert radiologists were asked to annotate lung nodules. Despite their expertise, they disagree significantly on many cases. Standard segmentation networks fail to capture these variations, thereby giving a false sense of confidence in model predictions. Our approach learns the distribution of possible labels, allowing us to generate realistic and diverse segmentations. + +labeling. + +Instead, multiple possible correct segmentation maps may be required to reflect the true distribution of annotations. For instance, Fig. 1 illustrates the task of lung nodule segmentation from CT scans where expert annotators provide multiple valid segmentation maps. In this context, stochastic semantic segmentation methods must learn to predict conditional distributions of labels given the image. Doing so is challenging, however, as the distribution is typically multimodal, the output space is high-dimensional, and annotation data is limited. + +Denoising Diffusion Probabilistic Models (DDPMs) appear well-suited to overcome these challenges. DDPMs have recently drawn strong interest in computer vision as a framework for learning complex distributions in high-dimensional spaces. After achieving state-of-the-art performance on image synthesis [13], they have been successfully extended to solve tasks such as text-to-image generation [41], counterfactual explanation generation [24], inpainting [34], but also image classification [56] and seman + +tic segmentation [1, 3, 48] amongst others. + +While DDPMs were originally formulated as probabilistic models able to learn high-dimensional data distributions of discrete and ordered variables (e.g., RGB pixel values), re-formulations and modifications that allow for categorical variables (e.g., labels) [21] are one of the key reasons why DDPMs are being explored in a broad range of computer vision tasks [12]. Specifically, the ability to model the spatial distribution of categorical variables is well suited for numerous computer vision tasks, including semantic segmentation [6, 8, 10, 14, 16, 17, 27, 31, 33, 54, 55]. Yet until now, segmentation methods using DDPMs have relied on the original discrete and ordered formulation and different heuristics to yield categorical outputs [1, 3, 48]. Consequently, the potential advantages of adopting diffusion models of categorical variables for stochastic image segmentation are still unknown. + +In light of the above, we propose a conditional categorical diffusion model (CCDM) for semantic segmentation based on DDPMs, which models both the observed and the latent variables as categorical distributions. This enables the model to explicitly generate labels maps of discrete, unordered variables, thereby circumventing the need for switching between continuous and discrete domains, as in previous methods. The model is conditioned to the input image, making it possible to generate multiple segmentation label maps that account for the aleatoric uncertainty arising from image ambiguity. We show experimentally that our approach achieves state-of-the-art performance on LIDC, a stochastic semantic segmentation dataset, according to several performance measures. Moreover, when applied to the classical segmentation dataset Cityscapes, our method provides competitive results, outperforming established baselines. + +In summary, our main contributions are the following: + +- We propose a conditional categorical diffusion model capable of learning the label distribution given an input image that can be used to produce diverse segmentation samples that capture aleatoric uncertainty. +- For the task of learning a multi-rater semantic segmentation label distribution, our method achieves state-of-the-art performance on LIDC, being the first diffusion-based approach proposed for this task. +- We report competitive performance on a challenging semantic segmentation task, Cityscapes, outperforming several established baselines using a lightweight model that also leverages an off-the-shelf pre-trained feature extractor. + +# 2. Related work + +Stochastic segmentation: Methods for stochastic se + +matic segmentation aim at capturing the aleatoric uncertainty and inherent unpredictability of the labels used for segmentation. Different frameworks have been proposed to yield segmentations according to the underlying label distribution. + +Initial works aimed at equipping a standard U-Net [40] with a probabilistic element to generate multiple predictions for the same image, typically accomplished by adding a conditional variational autoencoder (cVAE) [45], where the low-dimensional latent space of the cVAE encodes the possible segmentation variants. In [28], samples from this latent space are upscaled and concatenated at the last layer of the U-Net. Multiple methods extend this set-up to a hierarchical version [4, 29, 53]. Other works use normalizing flows to allow for a more expressive distribution than the Gaussian distribution in the cVAE [43, 46], switch to a discrete latent space [37], or add variational dropout and use the inter-grader variability directly as a training target [23]. + +Several other methods do not rely on the probabilistic U-Net. Monteiro et al. [35] propose a network that uses a low-rank multivariate normal distribution to model the logit distribution. Kassapis et al. [25] leverage adversarial training to learn possible label maps based on the logits of a trained segmentation network. Zhang et al. [52] employ an autoregressive PixelCNN to model the conditional distribution between pixels. Finally, Gao et al. [15] use a mixture of stochastic experts, where each expert network estimates a mode of the uncertainty, and a gating network predicts the probabilities that an input image is segmented by one of the experts. Our method is the first to explore the use of categorical diffusion models for stochastic segmentation. + +Diffusion models: Generative diffusion models [44] have drawn much attention following their popularization by [19]. Since then, diffusion models have been successfully applied to various domains, such as image generation, restoration, and super-resolution [12]. + +More central to the work presented here, a few methods have attempted to apply diffusion models to semantic segmentation. Baranchuk et al. [3] first train diffusion models to generate images, then use multilayer perceptrons (MLP) on its features to predict the class label. Other works focus on binary segmentation with conditional diffusion models [1, 48]. These methods generate single-channel continuous samples conditioned on the input image and obtain binary segmentation masks by thresholding the result. Directly applying continuous diffusion is also done in [49, 50]. Chen et al. [9] generate discrete data with continuous diffusion models by encoding categorical data into bits and modeling these bits as real numbers. + +Hoogeboom et al. [21] propose multinomial diffusion, a variation of diffusion models designed for categorical data. Subsequently, multinomial diffusion has been applied to discrete use cases, such as for tabular data [30], the latent + +space of vector-quantized variational auto-encoders [11, 22] or text [21]. They can also generate segmentation maps in the unconditional setting at a very small resolution $(32 \times 64)$ [21]. Instead, we focus on the unexplored conditional case and demonstrate results at significantly higher resolutions (up to $256 \times 512$ ). + +# 3. Method + +We now introduce our approach by first framing the problem setting and defining the necessary notation. We then describe categorical diffusion models and the conditioning procedure to produce stochastic semantic segmentation via diffusion. + +# 3.1. Background and notation + +A denoising diffusion probabilistic model (DDPM) is a latent variable model $p_{\theta}(\mathbf{x}_0) = \int p_{\theta}(\mathbf{x}_{0:T})d\mathbf{x}_{1:T}$ describing the distribution of an observable variable $\mathbf{x}_0 \in \mathbb{R}^D$ using a collection of $T$ latent variables $\{\mathbf{x}_t\}_{t=1}^T$ with the same dimensionality as $\mathbf{x}_0$ . The joint distribution is modeled as a Markov chain $p_{\theta}(\mathbf{x}_{0:T}) = p(\mathbf{x}_T)\prod_{t=1}^T p_{\theta}(\mathbf{x}_{t-1} \mid \mathbf{x}_t)$ , which is commonly known as the reverse process. The initial $p(\mathbf{x}_T)$ is set to a known, tractable distribution such as the Gaussian distribution, while the transition distribution $p_{\theta}$ , parameterized by $\theta$ , is the trainable component of the model. Training a DDPM aims to approximate $p_{\theta}(\mathbf{x}_0)$ to an empirical distribution $q(\mathbf{x}_0)$ defined by a collection of samples (e.g., images from the real world). To that end, training minimizes the cross-entropy between both distributions, + +$$ +\min _ {\theta} \mathbb {E} _ {\mathbf {x} _ {0} \sim q (\mathbf {x} _ {0})} \left[ - \log p _ {\theta} (\mathbf {x} _ {0}) \right], \tag {1} +$$ + +which is intractable as it requires marginalizing over the latent variables. Instead, a tractable distribution $q(\mathbf{x}_{1:T} \mid \mathbf{x}_0)$ is introduced and used as an approximation to the intractable true posterior $p(\mathbf{x}_{1:T} \mid \mathbf{x}_0)$ to define the evidence lower bound (ELBO), + +$$ +\log p _ {\theta} (\mathbf {x} _ {0}) \geq \mathbb {E} _ {\mathbf {x} _ {1: T} \sim q \left(\mathbf {x} _ {1: T} \mid \mathbf {x} _ {0}\right)} \left[ \log \frac {p _ {\theta} \left(\mathbf {x} _ {0 : T}\right)}{q \left(\mathbf {x} _ {1 : T} \mid \mathbf {x} _ {0}\right)} \right], \tag {2} +$$ + +where the expectation is approximated by Monte Carlo sampling. The lower bound is tight when the approximate posterior $q$ equals the real posterior. Maximizing the ELBO over samples from $q(\mathbf{x}_0)$ minimizes the cross-entropy loss of Eq. (1). + +The key difference between DDPMs and other latent variable models is that the approximate posterior $q(\mathbf{x}_{1:T} \mid \mathbf{x}_0)$ is fixed and not learnable. DDPMs model this distribution as a Markov chain $q(\mathbf{x}_{1:T} \mid \mathbf{x}_0) = \prod_{t=1}^{T} q(\mathbf{x}_t \mid \mathbf{x}_{t-1})$ , known as the forward process. The transition distribution $q(\mathbf{x}_t \mid \mathbf{x}_{t-1})$ is chosen to be a tractable distribution that allows efficient sampling from $q(\mathbf{x}_t \mid \mathbf{x}_0)$ for + +any $t$ . The only constraint in the design of a DDPM is that $q(\mathbf{x}_T \mid \mathbf{x}_0) \approx p(\mathbf{x}_T)$ . + +The original DDPM [19] modeled the transition distributions of the forward and the reverse processes as Gaussian with diagonal covariance matrices, and $p(\mathbf{x}_T)$ as a standard multivariate normal. However, these assumptions are inadequate when the elements of $\mathbf{x}_0$ belong to discrete, unordered sets, as in the task of image segmentation. + +# 3.2. Categorical diffusion model + +We now consider the denoising diffusion formulation to learn complex distributions of discrete image labelings. The observable variable $\mathbf{x}_0\in \mathcal{L}^D$ is categorical, where $D$ is the number of pixels of the image and $\mathcal{L} = \{1,\dots ,L\}$ is the set of discrete labels that can be assigned to each pixel. Following [21], we consider that all latent variables in $\mathbf{x}_{1:T}$ are also categorical and that the transition distributions for the forward and reverse processes are modeled as categorical distributions. For the forward process, the transition distribution acts element-wise over the previous state $\mathbf{x}_{t - 1}$ to produce the parameters of the distribution for $\mathbf{x}_t$ as, + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {t - 1}\right) = \prod_ {d = 1} ^ {D} q \left(\mathbf {x} _ {t} [ d ] \mid \mathbf {x} _ {t - 1} [ d ]\right), \tag {3} +$$ + +where $\mathbf{x}_t[d]$ indicates the label at time $t$ and pixel $d$ . In the following discussion, we will use $x_{t}\in \mathcal{L}$ to refer to the label of a single pixel $d$ , and we will drop the index $d$ for clarity. The pixel-wise transition distribution $q(x_{t}\mid x_{t - 1})$ gives the element-wise probability of the next label given the previous label as, + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) = \mathcal {C} \left(x _ {t}; \frac {\beta_ {t}}{L} \mathbf {1} + \left(1 - \beta_ {t}\right) \mathbf {e} _ {x _ {t - 1}}\right), \tag {4} +$$ + +where $\mathbf{1} = (1,\dots ,1)^T$ $\mathbf{e}_{\ell}$ is the one-hot encoding vector with 1 in position $\ell$ and 0 elsewhere, and the hyperparameter $\alpha_{t} = 1 - \beta_{t}\in (0,1)$ indicates the probability of keeping the label unchanged. $\mathcal{C}(x;\mathbf{p})$ denotes the categorical distribution with parameter vector $\mathbf{p}\in [0,1]^L$ . From the properties of categorical distributions, $\mathcal{C}(x\mid \mathbf{p}) = \mathbf{p}[x]$ and $\sum_{x}\mathbf{p}[x] = 1$ + +The transition distribution of the forward process can be composed as, + +$$ +q \left(x _ {t} \mid x _ {0}\right) = \mathcal {C} \left(x _ {t}; \frac {1 - \bar {\alpha} _ {t}}{L} \mathbf {1} + \bar {\alpha} _ {t} \mathbf {e} _ {x _ {0}}\right) \tag {5} +$$ + +with $\bar{\alpha}_t = \prod_{\tau=1}^t \alpha_\tau$ , which enables efficient sampling of elements from the Markov chain at any location $t$ . Finally, the posterior of the transition distribution can be computed with the previous formulas by applying Bayes rule, + +$$ +q \left(x _ {t - 1} \mid x _ {t}, x _ {0}\right) = \mathcal {C} \left(x _ {t - 1}; \boldsymbol {\pi} \left(x _ {t}, x _ {0}\right)\right), \tag {6} +$$ + +![](images/3bb8a669fdf7b65a888888f22a0a0e90483cab35204622ead8bd7bb6b9c87d01.jpg) +Figure 2: Illustration of the reverse process of our method. The conditional categorical diffusion model (CCDM) receives as input an image $I$ and a categorical label map $\mathbf{x}_T^{(i)}$ sampled from the categorical uniform noise. The reverse process of the CCDM generates a label map $\mathbf{x}_0^{(i)}$ , which is a sample from the learned distribution $p(\mathbf{x}_0 \mid I)$ . When repeated for $N$ samples, we obtain an empirical approximation to the multimodal label distribution for the image $I$ , learned from the annotations of multiple expert raters. + +with, + +$$ +\boldsymbol {\pi} \left(x _ {t}, x _ {0}\right) = \frac {1}{\tilde {\pi}} \left(\frac {\beta_ {t}}{L} \mathbf {1} + \alpha_ {t} \mathbf {e} _ {x _ {t}}\right) \odot \left(\frac {1 - \bar {\alpha} _ {t - 1}}{L} \mathbf {1} + \bar {\alpha} _ {t - 1} \mathbf {e} _ {x _ {0}}\right) \tag {7} +$$ + +and $\tilde{\pi} = \frac{1 - \bar{\alpha}_t}{L} +\bar{\alpha}_t\cdot \delta_{xt}^{x_0}$ , where $\delta$ is the Kronecker delta. + +The transition distribution of the reverse process is also an element-wise categorical distribution, + +$$ +p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right) = \prod_ {d = 1} ^ {D} \mathcal {C} \left(x _ {t - 1}; \hat {\mathbf {p}} _ {t - 1}\right), \tag {8} +$$ + +where $x_{t - 1} = \mathbf{x}_{t - 1}[d]$ and $\hat{\mathbf{p}}_{t - 1}$ are the label and the estimated parameter vector, respectively, at pixel $d$ . Unlike the forward process, the parameter vector for the pixel $d$ is not computed considering only the element $d$ of $\mathbf{x}_t$ . Instead, it is modeled as a function $f: \mathcal{L}^D \to [0,1]^{D\times L}$ that incorporates context by considering the entire label map $\mathbf{x}_t$ to produce a collection of $D$ probability distributions for $\mathbf{x}_{t - 1}$ , which we refer to as $\hat{\mathbf{P}}_{t - 1} \in [0,1]^{D\times L}$ with $\hat{\mathbf{p}}_{t - 1} = \hat{\mathbf{P}}_{t - 1}[d]$ . + +While it is possible to use a neural network to estimate $\hat{\mathbf{P}}_{t - 1}$ , Ho et al. [19] suggested that a consistent output space for the network led to enhanced performance. Following this idea, we train a network $f_{\theta}$ , parameterized by $\theta$ , to compute $\hat{\mathbf{P}}_0 = f_\theta (\mathbf{x}_t,t)\in [0,1]^{D\times L}$ by receiving a label map $\mathbf{x}_t$ and the step $t$ . We then transform the parameter vector for each pixel, $\hat{\mathbf{p}}_0 = \hat{\mathbf{P}}_0[d]$ to the parameter vector $\hat{\mathbf{p}}_{t - 1}$ for the same pixel of $\mathbf{x}_{t - 1}$ as, + +$$ +\begin{array}{l} \mathcal {C} \left(x _ {t - 1}; \hat {\mathrm {p}} _ {t - 1}\right) = (9) \\ = \sum_ {x _ {0}} q \left(x _ {t - 1} \mid x _ {t}, x _ {0}\right) \cdot \mathcal {C} \left(x _ {0}; \hat {\mathbf {p}} _ {0}\right) (10) \\ = \sum_ {x _ {0}} \mathcal {C} \left(x _ {t - 1}; \boldsymbol {\pi} \left(x _ {t}, x _ {0}\right)\right) \cdot \mathcal {C} \left(x _ {0}; \hat {\mathbf {p}} _ {0}\right), (11) \\ \end{array} +$$ + +from which, + +$$ +\hat {\mathbf {p}} _ {t - 1} = \sum_ {x _ {0} \in \mathcal {L}} \boldsymbol {\pi} \left(x _ {t}, x _ {0}\right) \cdot \hat {\mathbf {p}} _ {0} \left[ x _ {0} \right], \tag {12} +$$ + +where we have omitted the pixel indices $d$ for clarity. This transformation is not necessary when $t = 1$ , as then $\hat{\mathbf{p}}_{t - 1} = \hat{\mathbf{p}}_0$ computed by $f_{\theta}$ . It is also possible to perform this computation in parallel for every pixel to efficiently obtain $\hat{\mathbf{P}}_{t - 1}$ . Note that the result of Eq. (12) differs from the parameter vector computed in [21], where the ill-defined expression $\hat{\mathbf{p}}_{t - 1} = \pi (x_t,\hat{\mathbf{x}}_0)$ is employed. + +# 3.3. Conditional categorical diffusion + +In stochastic segmentation, the label map $\mathbf{x}_0$ for an image $I$ is modeled by a distribution $q(\mathbf{x}_0\mid I)$ . This distribution is often too complex to be properly approximated as a product of pixel-wise categorical distributions. We use a conditional categorical diffusion model $p(\mathbf{x}_0\mid I)$ (CCDM) to model the potentially complex interactions between labels and pixels. + +When conditioning the categorical diffusion model on an image, the forward process remains unchanged, $q(\mathbf{x}_{1:T} \mid \mathbf{x}_0, I) = q(\mathbf{x}_{1:T} \mid \mathbf{x}_0)$ , as any latent variable is conditionally independent of the image given any previous variable. On the other hand, the reverse process needs to incorporate the dependency on the image in its transition distribution, $p_{\theta}(\mathbf{x}_{0:T} \mid I) = p(\mathbf{x}_T \mid I) \prod_{t=1}^{T} p_{\theta}(\mathbf{x}_{t-1} \mid \mathbf{x}_t, I)$ . In practice, this dependency is enforced by an additional input to the neural network $f_{\theta}(\mathbf{x}_t, t, I)$ . + +# 3.4. Training + +Training is performed by maximizing the ELBO of Eq. (2). Reorganizing terms and distributing expectations for variance reduction, we express the ELBO as a sum of three terms: + +$$ +\begin{array}{l} \log p _ {\theta} (\mathbf {x} _ {0} \mid I) \geq \\ \mathbb {E} _ {\mathbf {x} _ {1} \sim q \left(\mathbf {x} _ {1} \mid \mathbf {x} _ {0}\right)} [ \log p _ {\theta} (\mathbf {x} _ {0} \mid \mathbf {x} _ {1}, I) ] (13) \\ - \sum_ {t = 2} ^ {T} \mathbb {E} _ {\mathbf {x} _ {t} \sim q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {0}\right)} [ K L (q \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}, \mathbf {x} _ {0}\right) \| p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}, I\right)) ] (14) \\ - K L \left(q \left(\mathbf {x} _ {T} \mid \mathbf {x} _ {0}\right) \| p \left(\mathbf {x} _ {T} \mid I\right)\right). (15) \\ \end{array} +$$ + +The first two terms can be optimized by standard gradient ascent. We approximate the expectations with Monte Carlo sampling with a single sample. The sum over the time variable $t$ is also approximated by a single uniform sample over $\{1, \dots, T\}$ . The KL divergence of the second term is the sum of pixel-wise KL divergences, + +$$ +K L (q \| p) = \sum_ {d = 1} ^ {D} K L \left(q \left(x _ {t - 1} \mid x _ {t}, x _ {0}\right) \| p _ {\theta} \left(x _ {t - 1} \mid \mathbf {x} _ {t}, I\right)\right), \tag {16} +$$ + +Algorithm 1 Training a CCDM with $T$ steps +Require: Training data expressed as the empirical distribu-. tion $q(\mathbf{x}_0,I) = q(\mathbf{x}_0\mid I)q(I)$ repeat t \~ Uniform( $\{1,\dots,T\}$ 1 $\begin{array}{rl} & {\boldsymbol {I}\sim q(\boldsymbol {I})}\\ & {\mathbf{x}_0\sim q(\mathbf{x}_0\mid \boldsymbol {I})}\\ & {\mathbf{x}_t\sim q(\mathbf{x}_t|\mathbf{x}_0)}\\ & {\hat{\mathbf{P}}_0\leftarrow f_\theta (\mathbf{x}_t,I,t)} \end{array}$ ▷ shape $D\times L$ if $t > 1$ then $\triangleright$ Pixel-wise application of Eq. (12) $\hat{\mathbf{p}}_{t - 1}\gets \sum_{x_0\in \mathcal{L}}\pi (x_t,x_0)\cdot \hat{\mathbf{p}}_0[x_0]\quad \triangleright$ shape L $\triangleright$ Compute KL with Eq. (8) and (16) $\ell \leftarrow KL(q(\mathbf{x}_{t - 1}|\mathbf{x}_t,\mathbf{x}_0)\| p_\theta (\mathbf{x}_{t - 1}|\mathbf{x}_t,I))$ else $\ell \leftarrow -\sum_d\log \mathcal{C}(x_0\mid \hat{\mathbf{P}}_0[d])$ end if $\theta \leftarrow \theta -\nabla_{\theta}\ell$ ▷ Gradient descent until converged + +where the parameter vectors of distributions $q$ and $p$ are computed with Eqs. (7) and (12), respectively. Alg. 1 shows the complete training procedure. + +The third term of Eq. (15) does not depend on the learnable parameters $\theta$ and is ignored during training. It is optimized by the design of the categorical diffusion model. Since the forward process converges as + +$$ +\lim _ {t \rightarrow \infty} q \left(x _ {t} \mid x _ {0}\right) = \mathcal {C} \left(x; \frac {1}{L}\right), \tag {17} +$$ + +we fix $p(\mathbf{x}_T\mid I)$ to the element-wise uniform distribution, + +$$ +p \left(x _ {T} \mid I\right) = p \left(x _ {T}\right) = \mathcal {C} \left(x _ {T}; \frac {1}{L}\right). \tag {18} +$$ + +This ensures that $p(\mathbf{x}_T \mid I) \approx q(\mathbf{x}_T \mid \mathbf{x}_0)$ , making the third term of the ELBO close to zero. + +At inference, the CCDM samples from $p(\mathbf{x}_0 \mid I)$ to generate label maps for a given image $I$ , which is achieved by traversing the Markov chain of the reverse process as outlined in Alg. 2 and illustrated in Fig. 2. To minimize the noise of the generated label maps, the CCDM selects the label with maximum probability instead of sampling from $\mathcal{C}(x_0 \mid \hat{\mathbf{p}}_0)$ in the final step. + +# 3.5. Architecture of $f_{\theta}$ + +As described above, the neural network $f_{\theta}$ receives a label map $\mathbf{x}_t$ , a time step $t$ , and an image $I$ to estimate the probability parameters for $\mathbf{x}_0$ . Its base design is a U-Net-like architecture [13] with self-attention modules at the three innermost layers of the encoder and the decoder [13]. The network processes the input label map represented as + +Algorithm 2 Inference from a CCDM with $T$ steps +Require: Input image $I$ , $f_{\theta}$ a network trained with Alg. 1 + $\mathbf{x}_T \sim \mathcal{C}^D(x_T; \frac{1}{L})$ $\mathbf{x}_{\text{prev}} \gets \mathbf{x}_T \quad \triangleright$ Stores interm. and final prediction +for $t = T, \dots, 1$ do + $\hat{\mathbf{P}}_0 \gets f_{\theta}(\mathbf{x}_{\text{prev}}, I, t)$ +if $t > 1$ then + $\triangleright$ Pixel-wise application of Eq. (12) + $\hat{\mathbf{p}}_{t-1} \gets \sum_{x_0 \in \mathcal{L}} \pi(x_t, x_0) \cdot \hat{\mathbf{p}}_0[x_0]$ $\mathbf{x}_{\text{prev}} \sim \prod_d \mathcal{C}(x_{t-1} | \hat{\mathbf{p}}_{t-1})$ +else + $\triangleright$ Final prediction + $\mathbf{x}_{\text{prev}} \gets \arg \max_{x_0 \in \mathcal{L}} \hat{\mathbf{P}}_0[:, x_0]$ +end if +end for + +a binary tensor with $L$ channels encoding the label of each pixel as a one-hot vector. Parameters of the network are shared for all values of $t$ . The step variable $t$ is encoded with the standard transformer sinusoidal position embedding [19] and concatenated as additional channels to the input tensor and to the feature maps of intermediate layers. Similarly, information from the input image $I$ is presented to the network as raw pixel values concatenated to the input tensor as additional channels. In some experiments we used a pre-trained transformer architecture Dino-ViT [5] to extract informative visual features from the image $I$ . In those cases, the extracted features were concatenated to the feature map of the third level of the U-Net encoder, which corresponds to a spatial shape equal to $\frac{1}{8}$ the shape of the input image. + +# 4. Experiments + +In all our experiments, we set $T = 250$ and the collection of $\beta_{t}$ are set following the cosine schedule proposed in [36]. We evaluate our method on two tasks described below. + +# 4.1. Segmentation with multiple annotations + +Dataset The Lung Image Database Consortium (LIDC) [2] binary segmentation dataset consists of 1'018 three dimensional chest CT scans of patients with lung cancer. Lung nodules of each volume are annotated by four expert raters from a pool of 12, yielding large differences in annotations in some cases. We extract nodule-centered slices from the CT volumes and treat each slice as an independent image. + +While LIDC is the standard benchmark of stochastic segmentation methods to date (e.g. [4, 15, 23, 25, 28, 29, 35, 43, 53, 54]), experimental configurations (pre-processing, training/validation/test splits, metrics) vastly differ across the literature. We conduct our experiments on the two most prominent LIDC splits and report results on both separately. + +
MethodLIDCv1LIDCv2
GED16GED32GED50GED100HM-IoU16HM-IoU32GED16GED50GED100HM-IoU16
Prob. Unet [28]0.310±0.01-0.303±0.01+-0.252±0.004†0.552±0.00-0.548±0.00+0.320±0.030‡-0.252±‡0.500±0.030‡
HProb. Unet [29]0.270±0.01----0.530±0.01--0.270±0.010‡--0.530±0.01
PhiSeg [4]0.262±0.00-0.247±0.00+-0.224±0.004†0.586±0.00-0.595±0.00+----
SSN [35]0.259±0.00-0.243±0.01+-0.225±0.0020.558±0.00-0.555±0.01+----
cFlow [43]-0.225±0.01+---0.584±0.00+----
CAR [25]---0.228±0.009--0.264±0.0020.248±0.0040.243±0.0040.592±0.005
JProb. Unet [53]-0.206±0.00---0.647±0.010.262±0.00--0.585±0.00
PixelSeg [52]0.243±0.01---0.614±0.00-0.260±0.00--0.587±0.01
MoSE [15]0.218±0.003-0.195±0.0020.189±0.0020.624±0.004-----
AB [9]0.213±0.0010.196±0.0020.193±0.002-0.614±0.0010.619±0.001----
CIMD [38]0.234±0.0050.218±0.0050.210±0.005-0.587±0.0010.592±0.002----
CCDM (ours)0.212±0.0020.194±0.0010.187±0.0020.183±0.0020.623±0.0020.631±0.0020.239±0.0030.216±0.0030.210±0.0030.598±0.001
+ +Table 1: Quantitative results on LIDCv1 and LIDCv2, with the methods ordered by year. Bold and underlined indicate best and second best per column, respectively. Our results are over 3 seeds. For GED, lower is better; for HM-IoU, higher is better. No method, including ours, uses pre-trained weights. Results for CIMD [38] and AB [9] are ours. All other scores are taken from their original papers, except $(^{+})$ from [53], $(-)$ from [53], $(\dagger)$ from [35], $\ddagger$ from [25]. + +The first, referred to as LIDCv1, is used in [4, 15, 35, 53]. LIDCv1 comprises 15'096 slices, divided into training, validation, and testing sets with the ratio $60:20:20$ . The second, LIDCv2, is used in [25, 28] and contains 12'816 images with the ratio $70:15:15$ . + +Metrics We measure the performances with the Generalised Energy Distance (GED) and the Hungarian-Matched Intersection over Union (HM-IoU) [15, 25, 29]. Both metrics measure the difference between the distributions of generated and ground-truth label maps. We denote the metrics computed with $n$ samples using a subscript, i.e., $\mathrm{GED}_n$ and $\mathrm{HM - IoU}_n$ , and we set $n$ to common values found in the literature. Note that higher number of samples yield more precise estimates. + +Baselines We compare our approach to eleven recent stochastic segmentation methods: probabilistic U-Net (Prob. Unet) [28], hierarchical probabilistic U-Net (HProb. Unet) [29], PhiSeg [4], stochastic segmentation network (SSN) [35], conditional normalizing flow (cFlow) [43], calibrated adversarial refinement (CAR) [25], joint probabilistic U-Net (JProb. Unet) [53], PixelSeg [52], mixture of stochastic experts (MoSE) [15], analog bits (AB) [9], and collectively intelligent medical diffusion (CIMD) [38]. + +Following standard practice, we use random horizontal and vertical flipping and random rotations of $0^{\circ}$ , $90^{\circ}$ , $180^{\circ}$ and $270^{\circ}$ for data augmentation. The resolution of the input images is $128 \times 128$ . We trained our method with the Adam optimizer [26] until convergence of the GED metric on the validation set, a polynomial learning rate scheduling starting from $1e^{-4}$ and ending with $1e^{-6}$ , and batch size of 64. We applied Polyak averaging with $\alpha = 0.99995$ . + +# 4.2. Segmentation with a single annotation + +We also evaluate our method with Cityscapes, a classical multi-class segmentation dataset where each image is annotated with a single label map. It comprises 2'975 RGB images of urban scenes for training and 500 images for validation, with each image labeled using 19 possible classes. + +We compare our approach to several established baselines using the validation set: DeepLabv3 [7], HRNet [47], and UPerNet [51], with both ResNet [18] and Swin [32] backbones. + +Besides our standard method, which performs image conditioning by concatenating the raw pixel values as channels of the input tensor, we also included in our comparison a variant CCDM-Dino which leverages pre-trained DinoViT features [5] as additional conditioning concatenated to intermediate feature maps of our model's encoder. + +Experiments are conducted separately for two different image resolutions: $128 \times 256$ and $256 \times 512$ . For all reported methods, we first resize the images to a fixed resolution and then apply color jittering, random flipping, and standard ImageNet intensity normalization as data augmentation. All baselines are trained for 500 epochs with a batch size of 32, with optimizers, learning rate schedules, and weight decay settings as reported in their respective publications (reported in detail in the supplementary material). + +Our method was trained for 800 epochs with a batch size of 32 at $128 \times 256$ and of 16 at $256 \times 512$ , using the Adam optimizer [26] with a learning rate of $1e^{-4}$ linearly decayed to $1e^{-6}$ . We applied Polyak averaging with $\alpha = 0.999$ . + +Performance is measured with the mean intersection-over-union (mIoU). Unlike GED and HM-IoU, the metric mIoU is incompatible with multiple label maps per image. During inference, CCDM generates multiple label maps per image that are subsequently fused into a single label map for + +performance assessment. We found that fusing by averaging the predicted probabilities resulted in superior performances compared to fusing by majority vote. + +# 5. Results + +# 5.1. LIDC + +We report performances on LIDCv1 and LIDCv2 in Tab. 1 and qualitative results in Fig. 3. Due to the lack of consistent evaluation protocols, we use a total of 10 metrics, thereby covering all the baselines and allowing for direct comparisons. + +Our CCDM reaches the best performance for eight out of the ten metrics, despite its relatively small size, with 9M parameters compared to, e.g., the 42M parameters of MoSE. CCDM also outperforms recent continuous diffusion models for segmentation, including AB [9] (9M parameters) and CIMD [38] (24M parameters). On HM-IoU16, the CCDM has a lower mean performance than MoSE by 0.001, but with only half the standard deviation. The JProb. Unet reaches a higher HM-IoU32 than all other methods, despite being considerably worse for GED32 than our CCDM. Furthermore, on LIDCv2, the JProb. Unet achieves only the third-best score on GED16, and fourth-best on HM-IoU16. This result indicates how comparing results obtained on different LIDC versions with each other can be misleading. + +Fig. 3 presents qualitative results from our method. In columns (g)-(l), we see that our CCDM generates a distribution of samples that captures the annotation variability created by the four expert raters. Further, as seen in the bottom example, the CCDM also generates empty samples according to the annotations (b)-(e). + +Reduced number of time steps for sampling: During inference, traversing the $T$ steps of the reverse process makes sampling from DDPMs slow. A straightforward solution [36] involves traversing only a subset of nodes of the reverse process, $\{\mathbf{x}_{k\tau} : \tau \in \{0, \dots, T / k\}\}$ , reducing the number of steps by a factor $k$ . This technique accelerates inference at the expense of reduced performance. To illustrate the trade-offs between performance and speed, Fig. 5 presents the evolution of $\mathrm{GED}_{16}$ and $\mathrm{HM-IoU}_{16}$ as the number of inference steps is reduced. As expected, CCDMs perform best when the number of training and inference steps are equal, but a reasonable increase in speed without a large sacrifice in performance is possible. + +# 5.2. Cityscapes + +Experimental comparisons on Cityscapes are presented in Tab. 2, and qualitative examples are provided in Fig. 4. Experiments at $128 \times 256$ demonstrate that CCDM-Dino outperforms all other methods, even when only a single sample is used. CCDM-raw also remains competitive, being outperformed only by one baseline (UPerNet+Swin + +
MethodmIoU final (best)
ArchitectureBackbone#params128 × 256256 × 512
DeepLabv3 [7]ResNet50 (✓)39m43.4 (44.1)58.6 (59.2)
DeepLabv3 [7]ResNet101 (✓)58m43.8 (45.5)59.2 (59.8)
UPerNet [51]ResNet101 (✓)83m45.5 (47.1)60.7 (61.2)
HRNet [47]w48v2 (✓)70m48.2 (49.5)63.3 (64.2)
UPerNet [32]Swin-Tiny (✓)58m54.2 (55.9)65.5 (66.0)
CCDM (ours)-
samples=130m53.260.3
samples=530m55.462.0
samples=1030m56.262.4
CCDM (ours)Dino ViT-S (†)
samples=130m + 20M55.564.0
samples=530m + 20M56.965.4
samples=1030m + 20M57.365.8
+ +Table 2: Results on Cityscapes-val for resolutions $128 \times 256$ and $256 \times 512$ . **Bold** and **underlined** indicate best and second best per column, respectively. ( $\checkmark$ ) and ( $\dagger$ ) indicate supervised and self-supervised pretraining of the backbone, respectively. Gray indicates pretrained, non-finetuned parameters. We report final performance for our method and baselines. For the latter we also provide best achieved performance during training (in parenthesis). For CCDM methods, the field samples indicate the number of generated samples for label map fusion, as explained in Sect 4.2. + +
CCDM CapacitymIoU(128 × 256)
#paramsUNet Levelssamples=1samples=5samples=10
5.4M437.839.740.6
7.5M544.748.348.5
22M451.654.053.6
30M553.255.456.2
+ +Table 3: Effect of increasing CCDM capacity (without feature conditioning). + +Tiny), despite using only between $36\%$ and $51\%$ of the parameters of other models. Similarly, at $256 \times 512$ , CCDM-Dino outperforms four of the baselines with a single sample, lags behind UPerNet+Swin-Tiny only by 0.1 percent points with 5 samples, and outperforms all baselines with 10 samples. As expected, averaging across more samples improves performance for both CCDM-raw and CCDM-Dino, albeit with diminishing gains. Furthermore, the addition of Dino features boosts single-sample performance by 2.3 percent points at $128 \times 256$ , and 3.7 percent points at $256 \times 512$ , hinting the greater value of adding feature conditioning for generating segmentation at a higher resolution. + +CCDM Capacity: Tab. 3(b) demonstrates the effect of increasing the capacity of CCDM. Using more U-Net feature levels, and increasing the number of parameters by doubling the number of channels per level, increases the performance regardless of the number of samples used for inference. + +![](images/124261e915105a28395a1b92f13a05f954f52fb1b05d1e01069d9d8447e93011.jpg) +Figure 3: Qualitative results on four LIDC images with our method. (a) shows the image, (b)-(e) its four labels, (f) the mean prediction of our CCDM over six predictions, and (g)-(l) six individual predictions. + +![](images/ddc294190fcc727fc452af9ee35de2171ac5486e599827f1e79a0f9c7552f6e8.jpg) +Figure 4: Qualitative comparisons on Cityscapes. All methods are trained and tested at a resolution of $256 \times 512$ . Our method produces structures with greater visual realism than other baselines. This is especially noticeable inside the marked regions. + +![](images/460bab6719033622606de27821b83948ac968035bd01ee6e9cb79eb974dce2f3.jpg) +Figure 5: LIDC GED and HM-IoU versus the number of sampling steps on LIDC. Evaluated on 500 random test images using 16 samples each, over 3 seeds. + +# 6. Conclusion + +We introduced conditional categorical diffusion models (CCDMs) that are capable of effectively modeling pixel-level semantic distributions. Notably, and contrary to standard deterministic segmentation approaches, our model can produce diverse samples given an input image, thereby capturing the aleatoric uncertainty. Our method learns a multimodal label distribution of segmentations, induced by annotations from multiple expert raters, for which it achieves state-of-the-art results on a challenging medical imaging dataset, LIDC. Additionally, we demonstrate that it can achieve competitive performance on a standard multi-class semantic segmentation benchmark, Cityscapes, by outperforming several established, heavily engineered baselines + +despite using significantly fewer parameters. + +One limitation of our method is the requirement of several iterations for producing a sample, which is a common shortcoming of diffusion models. Accelerating sampling constitutes a crucial research direction, orthogonal to the present work. Finally, resolution scaling remains notoriously difficult for diffusion models, with successful examples relying on massive computational resources to train cascades of models that gradually increase resolution [20, 42] or operate on the latent space of existing embedding methods for continuous data (e.g. images) [39] that are not available for categorical data. + +# Acknowledgements + +This work was partially funded by the University of Bern, Swiss National Science Foundation Grants #320030-188591, #200021-192285 and #200021-191983. + +# References + +[1] Tomer Amit, Eliya Nachmani, Tal Shaharbany, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021. 2 +[2] Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al. The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Medical physics, 38(2):915-931, 2011. 5 +[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 2 +[4] Christian F Baumgartner, Kerem C Tezcan, Krishna Chaitanya, Andreas M Hotker, Urs J Muehlematter, Khoschy Schawkat, Anton S Becker, Olivio Donati, and Ender Konukoglu. Phiseg: Capturing uncertainty in medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part II 22, pages 119-127. Springer, 2019. 2, 5, 6 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5, 6 +[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2 +[7] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for seman- + +tic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7 +[8] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2 +[9] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022. 2, 6, 7 +[10] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems, 34:9355-9366, 2021. 2 +[11] Max Cohen, Guillaume Quispe, Sylvain Le Corff, Charles Ollion, and Eric Moulines. Diffusion bridges vector quantized variational autoencoders. arXiv preprint arXiv:2202.04895, 2022. 3 +[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. arXiv preprint arXiv:2209.04747, 2022. 2 +[13] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. CoRR, abs/2105.05233, 2021. 1, 5 +[14] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019. 2 +[15] Zhitong Gao, Yuong Chen, Chuyu Zhang, and Xuming He. Modeling multimodal aleatoric uncertainty in segmentation with mixture of stochastic expert. arXiv preprint arXiv:2212.07328, 2022. 2, 5, 6 +[16] Jiaqi Gu, Hyoukjun Kwon, Dilin Wang, Wei Ye, Meng Li, Yu-Hsin Chen, Liangzhen Lai, Vikas Chandra, and David Z Pan. Multi-scale high-resolution vision transformer for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12094-12103, 2022. 2 +[17] Adam W Harley, Konstantinos G Derpanis, and Iasonas Kokkinos. Segmentation-aware convolutional networks using local attention masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5038-5047, 2017. 2 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 6 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 4, 5 +[20] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 9 + +[21] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34:12454-12465, 2021. 2, 3, 4 +[22] Minghui Hu, Yujie Wang, Tat-Jen Cham, Jianfei Yang, and Ponnuthurai N Suganthan. Global context with discrete diffusion in vector quantised modelling for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11502-11511, 2022. 3 +[23] Shi Hu, Daniel Worrall, Stefan Knegt, Bas Veeling, Henkjan Huisman, and Max Welling. Supervised uncertainty quantification for segmentation with multiple annotations. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22, pages 137–145. Springer, 2019. 2, 5 +[24] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. arXiv preprint arXiv:2203.15636, 2022. 1 +[25] Elias Kassapis, Georgi Dikov, Deepak K Gupta, and Cedric Nugteren. Calibrated adversarial refinement for stochastic semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7057-7067, 2021. 2, 5, 6 +[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference for Learning Representations, 2015. 6 +[27] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019. 2 +[28] Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic u-net for segmentation of ambiguous images. Advances in neural information processing systems, 31, 2018. 2, 5, 6 +[29] Simon AA Kohl, Bernardino Romera-Paredes, Klaus H Maier-Hein, Danilo Jimenez Rezende, SM Eslami, Pushmeet Kohli, Andrew Zisserman, and Olaf Ronneberger. A hierarchical probabilistic u-net for modeling multi-scale ambiguities. arXiv preprint arXiv:1905.13077, 2019. 2, 5, 6 +[30] Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. Tabddpm: Modelling tabular data with diffusion models. arXiv preprint arXiv:2209.15421, 2022. 2 +[31] Lielei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep hierarchical semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1246-1257, 2022. 2 +[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7 +[33] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Pro- + +ceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 2 +[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1 +[35] Miguel Monteiro, Loic Le Folgoc, Daniel Coelho de Castro, Nick Pawlowski, Bernardo Marques, Konstantinos Kamnitsas, Mark van der Wilk, and Ben Glocker. Stochastic segmentation networks: Modelling spatially correlated aleatoric uncertainty. Advances in Neural Information Processing Systems, 33:12756-12767, 2020. 2, 5, 6 +[36] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 5.7 +[37] Di Qiu and Lok Ming Lui. Modal uncertainty estimation via discrete latent representation. arXiv preprint arXiv:2007.12858, 2020. 2 +[38] Aimon Rahman, Jeya Maria Jose Valanarasu, Ilker Hacihaliloglu, and Vishal M Patel. Ambiguous medical image segmentation using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11536-11546, 2023. 6, 7 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 9 +[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2 +[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1 +[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 9 +[43] Raghavendra Selvan, Frederik Faye, Jon Middleton, and Akshay Pai. Uncertainty quantification in medical image segmentation with normalizing flows. In Machine Learning in Medical Imaging: 11th International Workshop, MLMI 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4, 2020, Proceedings 11, pages 80-90. Springer, 2020. 2, 5, 6 +[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using + +nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2 +[45] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 2 +[46] MM Amaan Valiuddin, Christiaan GA Viviers, Ruud JG van Sloun, Peter HN de With, and Fons van der Sommen. Improving aleatoric uncertainty quantification in multi-annotated medical image segmentation with normalizing flows. In Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, and Perinatal Imaging, Placental and Preterm Image Analysis: 3rd International Workshop, UNSURE 2021, and 6th International Workshop, PIPPI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 3, pages 75–88. Springer, 2021. 2 +[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. TPAMI, 2019. 6, 7 +[48] Julia Wolleb, Robin Sandkühler, Florentin Bieder, Philippe Valmaggia, and Philippe C Cattin. Diffusion models for implicit image segmentation ensembles. arXiv preprint arXiv:2112.03145, 2021. 2 +[49] Junde Wu, Huihui Fang, Yu Zhang, Yehui Yang, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. arXiv preprint arXiv:2211.00611, 2022. 2 +[50] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, and Yanwu Xu. Medsegdiff-v2: Diffusion based medical image segmentation with transformer. arXiv preprint arXiv:2301.11798, 2023. 2 +[51] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018. 6, 7 +[52] Wei Zhang, Xiaohong Zhang, Sheng Huang, Yuting Lu, and Kun Wang. Pixelseg: Pixel-by-pixel stochastic semantic segmentation for ambiguous medical images. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4742-4750, 2022. 2, 6 +[53] Wei Zhang, Xiaohong Zhang, Sheng Huang, Yuting Lu, and Kun Wang. A probabilistic model for controlling diversity and accuracy of ambiguous medical image segmentation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4751-4759, 2022. 2, 5, 6 +[54] Yifan Zhang, Bo Pang, and Cewu Lu. Semantic segmentation by early region proxy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1258-1268, 2022. 2, 5 +[55] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 2 + +[56] Roland S Zimmermann, Lukas Schott, Yang Song, Benjamin A Dunn, and David A Klindt. Score-based generative classifiers. arXiv preprint arXiv:2110.00473, 2021. 1 \ No newline at end of file diff --git a/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/images.zip b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5c1e079f3a7fe93887dd4c1732dd024584d2f56b --- /dev/null +++ b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c30ff6603643600708769a8feb0c70eb24e01970f0953b0b5529febb15b9f65c +size 412728 diff --git a/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/layout.json b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0285bc00f0e13f649394702a7ea96801e6d44917 --- /dev/null +++ b/2023/Stochastic Segmentation with Conditional Categorical Diffusion Models/layout.json @@ -0,0 +1,10764 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "text", + "content": "Stochastic Segmentation with Conditional Categorical Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 144, + 227, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 144, + 227, + 156 + ], + "spans": [ + { + "bbox": [ + 147, + 144, + 227, + 156 + ], + "type": "text", + "content": "Lukas Zbinden*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 158, + 228, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 158, + 228, + 171 + ], + "spans": [ + { + "bbox": [ + 118, + 158, + 228, + 171 + ], + "type": "text", + "content": "Adrian Thomas Huber" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 144, + 335, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 144, + 335, + 156 + ], + "spans": [ + { + "bbox": [ + 251, + 144, + 335, + 156 + ], + "type": "text", + "content": "Lars Doorenbos*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 252, + 158, + 342, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 158, + 342, + 171 + ], + "spans": [ + { + "bbox": [ + 252, + 158, + 342, + 171 + ], + "type": "text", + "content": "Raphael Sznitman" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 360, + 144, + 446, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 144, + 446, + 156 + ], + "spans": [ + { + "bbox": [ + 360, + 144, + 446, + 156 + ], + "type": "text", + "content": "Theodoros Pissas" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 367, + 158, + 472, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 158, + 472, + 171 + ], + "spans": [ + { + "bbox": [ + 367, + 158, + 472, + 171 + ], + "type": "text", + "content": "Pablo Márquez-Neila" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 204, + 177, + 389, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 177, + 389, + 191 + ], + "spans": [ + { + "bbox": [ + 204, + 177, + 389, + 191 + ], + "type": "text", + "content": "University of Bern, Bern, Switzerland" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 194, + 531, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 531, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 531, + 205 + ], + "type": "text", + "content": "{lukas.zbinden, lars.doorenbos, theodoros.pissas, raphael.sznitman, pablo.marquez}@unibe.ch" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 233, + 192, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 233, + 192, + 245 + ], + "spans": [ + { + "bbox": [ + 143, + 233, + 192, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 258, + 290, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 290, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 290, + 534 + ], + "type": "text", + "content": "Semantic segmentation has made significant progress in recent years thanks to deep neural networks, but the common objective of generating a single segmentation output that accurately matches the image's content may not be suitable for safety-critical domains such as medical diagnostics and autonomous driving. Instead, multiple possible correct segmentation maps may be required to reflect the true distribution of annotation maps. In this context, stochastic semantic segmentation methods must learn to predict conditional distributions of labels given the image, but this is challenging due to the typically multimodal distributions, high-dimensional output spaces, and limited annotation data. To address these challenges, we propose a conditional categorical diffusion model (CCDM) for semantic segmentation based on Denoising Diffusion Probabilistic Models. Our model is conditioned to the input image, enabling it to generate multiple segmentation label maps that account for the aleatoric uncertainty arising from divergent ground truth annotations. Our experimental results show that CCDM achieves state-of-the-art performance on LIDC, a stochastic semantic segmentation dataset, and outperforms established baselines on the classical segmentation dataset Cityscapes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 575, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 288, + 696 + ], + "type": "text", + "content": "Semantic segmentation has significantly progressed in recent years due to powerful deep neural networks. For most methods, the key objective is to generate a single segmentation output that accurately matches the image's content. However, this may not be suitable for safety-critical domains such as medical diagnostics and autonomous driving, as images in these applications often suffer from inherent ambiguity or annotations that have differences in opinion. In these cases, generating a single coherent segmentation may be hopeless to fully describe the set of correct" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 313, + 236, + 541, + 344 + ], + "blocks": [ + { + "bbox": [ + 313, + 236, + 541, + 344 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 541, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 541, + 344 + ], + "type": "image", + "image_path": "2860b575fdb0e7c4132b9ff61a9052200b804cf1558791aad29fd5ee10b52049.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 352, + 547, + 448 + ], + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 448 + ], + "type": "text", + "content": "Figure 1: Examples from the LIDC dataset, where expert radiologists were asked to annotate lung nodules. Despite their expertise, they disagree significantly on many cases. Standard segmentation networks fail to capture these variations, thereby giving a false sense of confidence in model predictions. Our approach learns the distribution of possible labels, allowing us to generate realistic and diverse segmentations." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 472, + 343, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 343, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 343, + 483 + ], + "type": "text", + "content": "labeling." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 484, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 604 + ], + "type": "text", + "content": "Instead, multiple possible correct segmentation maps may be required to reflect the true distribution of annotations. For instance, Fig. 1 illustrates the task of lung nodule segmentation from CT scans where expert annotators provide multiple valid segmentation maps. In this context, stochastic semantic segmentation methods must learn to predict conditional distributions of labels given the image. Doing so is challenging, however, as the distribution is typically multimodal, the output space is high-dimensional, and annotation data is limited." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Denoising Diffusion Probabilistic Models (DDPMs) appear well-suited to overcome these challenges. DDPMs have recently drawn strong interest in computer vision as a framework for learning complex distributions in high-dimensional spaces. After achieving state-of-the-art performance on image synthesis [13], they have been successfully extended to solve tasks such as text-to-image generation [41], counterfactual explanation generation [24], inpainting [34], but also image classification [56] and seman" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 702, + 124, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 124, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 124, + 712 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1119" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 220, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 220, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 220, + 83 + ], + "type": "text", + "content": "tic segmentation [1, 3, 48] amongst others." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 274 + ], + "type": "text", + "content": "While DDPMs were originally formulated as probabilistic models able to learn high-dimensional data distributions of discrete and ordered variables (e.g., RGB pixel values), re-formulations and modifications that allow for categorical variables (e.g., labels) [21] are one of the key reasons why DDPMs are being explored in a broad range of computer vision tasks [12]. Specifically, the ability to model the spatial distribution of categorical variables is well suited for numerous computer vision tasks, including semantic segmentation [6, 8, 10, 14, 16, 17, 27, 31, 33, 54, 55]. Yet until now, segmentation methods using DDPMs have relied on the original discrete and ordered formulation and different heuristics to yield categorical outputs [1, 3, 48]. Consequently, the potential advantages of adopting diffusion models of categorical variables for stochastic image segmentation are still unknown." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 276, + 286, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 286, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 286, + 477 + ], + "type": "text", + "content": "In light of the above, we propose a conditional categorical diffusion model (CCDM) for semantic segmentation based on DDPMs, which models both the observed and the latent variables as categorical distributions. This enables the model to explicitly generate labels maps of discrete, unordered variables, thereby circumventing the need for switching between continuous and discrete domains, as in previous methods. The model is conditioned to the input image, making it possible to generate multiple segmentation label maps that account for the aleatoric uncertainty arising from image ambiguity. We show experimentally that our approach achieves state-of-the-art performance on LIDC, a stochastic semantic segmentation dataset, according to several performance measures. Moreover, when applied to the classical segmentation dataset Cityscapes, our method provides competitive results, outperforming established baselines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 479, + 278, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 479, + 278, + 491 + ], + "spans": [ + { + "bbox": [ + 59, + 479, + 278, + 491 + ], + "type": "text", + "content": "In summary, our main contributions are the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 498, + 286, + 669 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 498, + 286, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 498, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 59, + 498, + 286, + 546 + ], + "type": "text", + "content": "- We propose a conditional categorical diffusion model capable of learning the label distribution given an input image that can be used to produce diverse segmentation samples that capture aleatoric uncertainty." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 555, + 286, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 555, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 59, + 555, + 286, + 602 + ], + "type": "text", + "content": "- For the task of learning a multi-rater semantic segmentation label distribution, our method achieves state-of-the-art performance on LIDC, being the first diffusion-based approach proposed for this task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 611, + 286, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 611, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 59, + 611, + 286, + 669 + ], + "type": "text", + "content": "- We report competitive performance on a challenging semantic segmentation task, Cityscapes, outperforming several established baselines using a lightweight model that also leverages an off-the-shelf pre-trained feature extractor." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 681, + 130, + 693 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 681, + 130, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 681, + 130, + 693 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "Stochastic segmentation: Methods for stochastic se" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "matic segmentation aim at capturing the aleatoric uncertainty and inherent unpredictability of the labels used for segmentation. Different frameworks have been proposed to yield segmentations according to the underlying label distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 133, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 289 + ], + "type": "text", + "content": "Initial works aimed at equipping a standard U-Net [40] with a probabilistic element to generate multiple predictions for the same image, typically accomplished by adding a conditional variational autoencoder (cVAE) [45], where the low-dimensional latent space of the cVAE encodes the possible segmentation variants. In [28], samples from this latent space are upscaled and concatenated at the last layer of the U-Net. Multiple methods extend this set-up to a hierarchical version [4, 29, 53]. Other works use normalizing flows to allow for a more expressive distribution than the Gaussian distribution in the cVAE [43, 46], switch to a discrete latent space [37], or add variational dropout and use the inter-grader variability directly as a training target [23]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 291, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 445 + ], + "type": "text", + "content": "Several other methods do not rely on the probabilistic U-Net. Monteiro et al. [35] propose a network that uses a low-rank multivariate normal distribution to model the logit distribution. Kassapis et al. [25] leverage adversarial training to learn possible label maps based on the logits of a trained segmentation network. Zhang et al. [52] employ an autoregressive PixelCNN to model the conditional distribution between pixels. Finally, Gao et al. [15] use a mixture of stochastic experts, where each expert network estimates a mode of the uncertainty, and a gating network predicts the probabilities that an input image is segmented by one of the experts. Our method is the first to explore the use of categorical diffusion models for stochastic segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 447, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 506 + ], + "type": "text", + "content": "Diffusion models: Generative diffusion models [44] have drawn much attention following their popularization by [19]. Since then, diffusion models have been successfully applied to various domains, such as image generation, restoration, and super-resolution [12]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 509, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 663 + ], + "type": "text", + "content": "More central to the work presented here, a few methods have attempted to apply diffusion models to semantic segmentation. Baranchuk et al. [3] first train diffusion models to generate images, then use multilayer perceptrons (MLP) on its features to predict the class label. Other works focus on binary segmentation with conditional diffusion models [1, 48]. These methods generate single-channel continuous samples conditioned on the input image and obtain binary segmentation masks by thresholding the result. Directly applying continuous diffusion is also done in [49, 50]. Chen et al. [9] generate discrete data with continuous diffusion models by encoding categorical data into bits and modeling these bits as real numbers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Hoogeboom et al. [21] propose multinomial diffusion, a variation of diffusion models designed for categorical data. Subsequently, multinomial diffusion has been applied to discrete use cases, such as for tabular data [30], the latent" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1120" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "space of vector-quantized variational auto-encoders [11, 22] or text [21]. They can also generate segmentation maps in the unconditional setting at a very small resolution " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "(32 \\times 64)" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " [21]. Instead, we focus on the unexplored conditional case and demonstrate results at significantly higher resolutions (up to " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 103, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 103, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 103, + 167 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 236 + ], + "type": "text", + "content": "We now introduce our approach by first framing the problem setting and defining the necessary notation. We then describe categorical diffusion models and the conditioning procedure to produce stochastic semantic segmentation via diffusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 243, + 189, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 189, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 189, + 255 + ], + "type": "text", + "content": "3.1. Background and notation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": "A denoising diffusion probabilistic model (DDPM) is a latent variable model " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0) = \\int p_{\\theta}(\\mathbf{x}_{0:T})d\\mathbf{x}_{1:T}" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": " describing the distribution of an observable variable " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": " using a collection of " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": " latent variables " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": " with the same dimensionality as " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": ". The joint distribution is modeled as a Markov chain " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T)\\prod_{t=1}^T p_{\\theta}(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": ", which is commonly known as the reverse process. The initial " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T)" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": " is set to a known, tractable distribution such as the Gaussian distribution, while the transition distribution " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": ", is the trainable component of the model. Training a DDPM aims to approximate " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": " to an empirical distribution " + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 262, + 288, + 430 + ], + "type": "text", + "content": " defined by a collection of samples (e.g., images from the real world). To that end, training minimizes the cross-entropy between both distributions," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 439, + 287, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 439, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 104, + 439, + 287, + 456 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0})} \\left[ - \\log p _ {\\theta} (\\mathbf {x} _ {0}) \\right], \\tag {1}", + "image_path": "ca34ef3b445ea8b14839a047a4b012ababb1441cf78bdbb46d010f4e0b272f55.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 465, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 465, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 287, + 525 + ], + "type": "text", + "content": "which is intractable as it requires marginalizing over the latent variables. Instead, a tractable distribution " + }, + { + "bbox": [ + 46, + 465, + 287, + 525 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 465, + 287, + 525 + ], + "type": "text", + "content": " is introduced and used as an approximation to the intractable true posterior " + }, + { + "bbox": [ + 46, + 465, + 287, + 525 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 465, + 287, + 525 + ], + "type": "text", + "content": " to define the evidence lower bound (ELBO)," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 533, + 287, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 287, + 561 + ], + "type": "interline_equation", + "content": "\\log p _ {\\theta} (\\mathbf {x} _ {0}) \\geq \\mathbb {E} _ {\\mathbf {x} _ {1: T} \\sim q \\left(\\mathbf {x} _ {1: T} \\mid \\mathbf {x} _ {0}\\right)} \\left[ \\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0 : T}\\right)}{q \\left(\\mathbf {x} _ {1 : T} \\mid \\mathbf {x} _ {0}\\right)} \\right], \\tag {2}", + "image_path": "a100443548f3c7fe60448ddd57d708f7147f672d40b9c5b54896a51b63ac60f4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "type": "text", + "content": "where the expectation is approximated by Monte Carlo sampling. The lower bound is tight when the approximate posterior " + }, + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "type": "text", + "content": " equals the real posterior. Maximizing the ELBO over samples from " + }, + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "type": "text", + "content": " minimizes the cross-entropy loss of Eq. (1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "content": "The key difference between DDPMs and other latent variable models is that the approximate posterior " + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "content": " is fixed and not learnable. DDPMs model this distribution as a Markov chain " + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) = \\prod_{t=1}^{T} q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1})" + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "content": ", known as the forward process. The transition distribution " + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1})" + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "content": " is chosen to be a tractable distribution that allows efficient sampling from " + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 628, + 287, + 714 + ], + "type": "text", + "content": " for" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "any " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ". The only constraint in the design of a DDPM is that " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_T \\mid \\mathbf{x}_0) \\approx p(\\mathbf{x}_T)" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 96, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 168 + ], + "type": "text", + "content": "The original DDPM [19] modeled the transition distributions of the forward and the reverse processes as Gaussian with diagonal covariance matrices, and " + }, + { + "bbox": [ + 304, + 96, + 547, + 168 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T)" + }, + { + "bbox": [ + 304, + 96, + 547, + 168 + ], + "type": "text", + "content": " as a standard multivariate normal. However, these assumptions are inadequate when the elements of " + }, + { + "bbox": [ + 304, + 96, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 96, + 547, + 168 + ], + "type": "text", + "content": " belong to discrete, unordered sets, as in the task of image segmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 175, + 459, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 175, + 459, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 459, + 188 + ], + "type": "text", + "content": "3.2. Categorical diffusion model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": "We now consider the denoising diffusion formulation to learn complex distributions of discrete image labelings. The observable variable " + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0\\in \\mathcal{L}^D" + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": " is categorical, where " + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": " is the number of pixels of the image and " + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\{1,\\dots ,L\\}" + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": " is the set of discrete labels that can be assigned to each pixel. Following [21], we consider that all latent variables in " + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{1:T}" + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": " are also categorical and that the transition distributions for the forward and reverse processes are modeled as categorical distributions. For the forward process, the transition distribution acts element-wise over the previous state " + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}" + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": " to produce the parameters of the distribution for " + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 334, + 545, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 334, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 347, + 334, + 545, + 367 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {t - 1}\\right) = \\prod_ {d = 1} ^ {D} q \\left(\\mathbf {x} _ {t} [ d ] \\mid \\mathbf {x} _ {t - 1} [ d ]\\right), \\tag {3}", + "image_path": "927e73ec184ca80ff259fa8bd2f4aea17cc212787698d15fe5195719687eae3c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t[d]" + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": " indicates the label at time " + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": " and pixel " + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": ". In the following discussion, we will use " + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "inline_equation", + "content": "x_{t}\\in \\mathcal{L}" + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": " to refer to the label of a single pixel " + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": ", and we will drop the index " + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": " for clarity. The pixel-wise transition distribution " + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "inline_equation", + "content": "q(x_{t}\\mid x_{t - 1})" + }, + { + "bbox": [ + 304, + 375, + 545, + 448 + ], + "type": "text", + "content": " gives the element-wise probability of the next label given the previous label as," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 456, + 545, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 456, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 328, + 456, + 545, + 483 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {C} \\left(x _ {t}; \\frac {\\beta_ {t}}{L} \\mathbf {1} + \\left(1 - \\beta_ {t}\\right) \\mathbf {e} _ {x _ {t - 1}}\\right), \\tag {4}", + "image_path": "fa664082ebf6ae22fc41b398f84e135305eb012dae49e9389a0f5be60fafa8aa.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{1} = (1,\\dots ,1)^T" + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\ell}" + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "text", + "content": " is the one-hot encoding vector with 1 in position " + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "text", + "content": " and 0 elsewhere, and the hyperparameter " + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}\\in (0,1)" + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "text", + "content": " indicates the probability of keeping the label unchanged. " + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(x;\\mathbf{p})" + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "text", + "content": " denotes the categorical distribution with parameter vector " + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{p}\\in [0,1]^L" + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "text", + "content": " . From the properties of categorical distributions, " + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(x\\mid \\mathbf{p}) = \\mathbf{p}[x]" + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 491, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\sum_{x}\\mathbf{p}[x] = 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 575, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 575, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 575, + 545, + 599 + ], + "type": "text", + "content": "The transition distribution of the forward process can be composed as," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 343, + 607, + 545, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 607, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 343, + 607, + 545, + 634 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {C} \\left(x _ {t}; \\frac {1 - \\bar {\\alpha} _ {t}}{L} \\mathbf {1} + \\bar {\\alpha} _ {t} \\mathbf {e} _ {x _ {0}}\\right) \\tag {5}", + "image_path": "2f2daf6c703f5da70b046d6bcdadb46f9a665d755d77f14b6530b00c6bae8f99.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 643, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 643, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 304, + 643, + 545, + 692 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 643, + 545, + 692 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{\\tau=1}^t \\alpha_\\tau" + }, + { + "bbox": [ + 304, + 643, + 545, + 692 + ], + "type": "text", + "content": ", which enables efficient sampling of elements from the Markov chain at any location " + }, + { + "bbox": [ + 304, + 643, + 545, + 692 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 643, + 545, + 692 + ], + "type": "text", + "content": ". Finally, the posterior of the transition distribution can be computed with the previous formulas by applying Bayes rule," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 343, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 343, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) = \\mathcal {C} \\left(x _ {t - 1}; \\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right)\\right), \\tag {6}", + "image_path": "b3d0f0ad77652407c5f274be8e77eaff8678c8ac81cd6f35ef0758d809339f8f.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "1121" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 286, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 148 + ], + "type": "image", + "image_path": "3bb8a669fdf7b65a888888f22a0a0e90483cab35204622ead8bd7bb6b9c87d01.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "text", + "content": "Figure 2: Illustration of the reverse process of our method. The conditional categorical diffusion model (CCDM) receives as input an image " + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "text", + "content": " and a categorical label map " + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{(i)}" + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "text", + "content": " sampled from the categorical uniform noise. The reverse process of the CCDM generates a label map " + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^{(i)}" + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "text", + "content": ", which is a sample from the learned distribution " + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid I)" + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "text", + "content": ". When repeated for " + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "text", + "content": " samples, we obtain an empirical approximation to the multimodal label distribution for the image " + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 157, + 289, + 269 + ], + "type": "text", + "content": ", learned from the annotations of multiple expert raters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 289, + 70, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 70, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 70, + 300 + ], + "type": "text", + "content": "with," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 308, + 293, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 293, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 293, + 344 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right) = \\frac {1}{\\tilde {\\pi}} \\left(\\frac {\\beta_ {t}}{L} \\mathbf {1} + \\alpha_ {t} \\mathbf {e} _ {x _ {t}}\\right) \\odot \\left(\\frac {1 - \\bar {\\alpha} _ {t - 1}}{L} \\mathbf {1} + \\bar {\\alpha} _ {t - 1} \\mathbf {e} _ {x _ {0}}\\right) \\tag {7}", + "image_path": "ff9ece4fc05f7fc33ad8dc1af303e25963845d063c34e99b85be7ad055bcf8d2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 344, + 279, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 344, + 279, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 344, + 279, + 357 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 47, + 344, + 279, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{\\pi} = \\frac{1 - \\bar{\\alpha}_t}{L} +\\bar{\\alpha}_t\\cdot \\delta_{xt}^{x_0}" + }, + { + "bbox": [ + 47, + 344, + 279, + 357 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 344, + 279, + 357 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 47, + 344, + 279, + 357 + ], + "type": "text", + "content": " is the Kronecker delta." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 357, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 380 + ], + "type": "text", + "content": "The transition distribution of the reverse process is also an element-wise categorical distribution," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 92, + 388, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 388, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 92, + 388, + 287, + 422 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\prod_ {d = 1} ^ {D} \\mathcal {C} \\left(x _ {t - 1}; \\hat {\\mathbf {p}} _ {t - 1}\\right), \\tag {8}", + "image_path": "af98b8309c444acd384b60264c67c5ef6ebfe8b2533f5f4f1b7f728693217a0e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "x_{t - 1} = \\mathbf{x}_{t - 1}[d]" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_{t - 1}" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": " are the label and the estimated parameter vector, respectively, at pixel " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": ". Unlike the forward process, the parameter vector for the pixel " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": " is not computed considering only the element " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": ". Instead, it is modeled as a function " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "f: \\mathcal{L}^D \\to [0,1]^{D\\times L}" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": " that incorporates context by considering the entire label map " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": " to produce a collection of " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": " probability distributions for " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": ", which we refer to as " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_{t - 1} \\in [0,1]^{D\\times L}" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_{t - 1} = \\hat{\\mathbf{P}}_{t - 1}[d]" + }, + { + "bbox": [ + 46, + 429, + 287, + 537 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "content": "While it is possible to use a neural network to estimate " + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_{t - 1}" + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "content": ", Ho et al. [19] suggested that a consistent output space for the network led to enhanced performance. Following this idea, we train a network " + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "content": ", to compute " + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_0 = f_\\theta (\\mathbf{x}_t,t)\\in [0,1]^{D\\times L}" + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "content": " by receiving a label map " + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "content": " and the step " + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "content": ". We then transform the parameter vector for each pixel, " + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_0 = \\hat{\\mathbf{P}}_0[d]" + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "content": " to the parameter vector " + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_{t - 1}" + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "content": " for the same pixel of " + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}" + }, + { + "bbox": [ + 46, + 537, + 287, + 633 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 87, + 641, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 641, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 87, + 641, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {C} \\left(x _ {t - 1}; \\hat {\\mathrm {p}} _ {t - 1}\\right) = (9) \\\\ = \\sum_ {x _ {0}} q \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) \\cdot \\mathcal {C} \\left(x _ {0}; \\hat {\\mathbf {p}} _ {0}\\right) (10) \\\\ = \\sum_ {x _ {0}} \\mathcal {C} \\left(x _ {t - 1}; \\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right)\\right) \\cdot \\mathcal {C} \\left(x _ {0}; \\hat {\\mathbf {p}} _ {0}\\right), (11) \\\\ \\end{array}", + "image_path": "7d96b0a1f94bc8177f892f12a24ef23242531df0157dac8518ad3ae1903ead2e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 358, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 358, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 358, + 83 + ], + "type": "text", + "content": "from which," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 87, + 545, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 87, + 545, + 113 + ], + "spans": [ + { + "bbox": [ + 359, + 87, + 545, + 113 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {p}} _ {t - 1} = \\sum_ {x _ {0} \\in \\mathcal {L}} \\boldsymbol {\\pi} \\left(x _ {t}, x _ {0}\\right) \\cdot \\hat {\\mathbf {p}} _ {0} \\left[ x _ {0} \\right], \\tag {12}", + "image_path": "449ba583d9907b2835693a360b94fb2c1c7b16f7387091e976e0cda50e68e9d6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "text", + "content": "where we have omitted the pixel indices " + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "text", + "content": " for clarity. This transformation is not necessary when " + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "text", + "content": ", as then " + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_{t - 1} = \\hat{\\mathbf{p}}_0" + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "text", + "content": " computed by " + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "text", + "content": ". It is also possible to perform this computation in parallel for every pixel to efficiently obtain " + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_{t - 1}" + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "text", + "content": ". Note that the result of Eq. (12) differs from the parameter vector computed in [21], where the ill-defined expression " + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_{t - 1} = \\pi (x_t,\\hat{\\mathbf{x}}_0)" + }, + { + "bbox": [ + 304, + 118, + 547, + 203 + ], + "type": "text", + "content": " is employed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 208, + 482, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 208, + 482, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 208, + 482, + 221 + ], + "type": "text", + "content": "3.3. Conditional categorical diffusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "type": "text", + "content": "In stochastic segmentation, the label map " + }, + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "type": "text", + "content": " for an image " + }, + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "type": "text", + "content": " is modeled by a distribution " + }, + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_0\\mid I)" + }, + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "type": "text", + "content": ". This distribution is often too complex to be properly approximated as a product of pixel-wise categorical distributions. We use a conditional categorical diffusion model " + }, + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0\\mid I)" + }, + { + "bbox": [ + 304, + 226, + 545, + 309 + ], + "type": "text", + "content": " (CCDM) to model the potentially complex interactions between labels and pixels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 310, + 546, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 546, + 418 + ], + "type": "text", + "content": "When conditioning the categorical diffusion model on an image, the forward process remains unchanged, " + }, + { + "bbox": [ + 304, + 310, + 546, + 418 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0, I) = q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 304, + 310, + 546, + 418 + ], + "type": "text", + "content": ", as any latent variable is conditionally independent of the image given any previous variable. On the other hand, the reverse process needs to incorporate the dependency on the image in its transition distribution, " + }, + { + "bbox": [ + 304, + 310, + 546, + 418 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{0:T} \\mid I) = p(\\mathbf{x}_T \\mid I) \\prod_{t=1}^{T} p_{\\theta}(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t, I)" + }, + { + "bbox": [ + 304, + 310, + 546, + 418 + ], + "type": "text", + "content": ". In practice, this dependency is enforced by an additional input to the neural network " + }, + { + "bbox": [ + 304, + 310, + 546, + 418 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x}_t, t, I)" + }, + { + "bbox": [ + 304, + 310, + 546, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 424, + 370, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 370, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 370, + 437 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 442, + 547, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 547, + 489 + ], + "type": "text", + "content": "Training is performed by maximizing the ELBO of Eq. (2). Reorganizing terms and distributing expectations for variance reduction, we express the ELBO as a sum of three terms:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 309, + 496, + 545, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 496, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 309, + 496, + 545, + 588 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log p _ {\\theta} (\\mathbf {x} _ {0} \\mid I) \\geq \\\\ \\mathbb {E} _ {\\mathbf {x} _ {1} \\sim q \\left(\\mathbf {x} _ {1} \\mid \\mathbf {x} _ {0}\\right)} [ \\log p _ {\\theta} (\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {1}, I) ] (13) \\\\ - \\sum_ {t = 2} ^ {T} \\mathbb {E} _ {\\mathbf {x} _ {t} \\sim q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right)} [ K L (q \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, \\mathbf {x} _ {0}\\right) \\| p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}, I\\right)) ] (14) \\\\ - K L \\left(q \\left(\\mathbf {x} _ {T} \\mid \\mathbf {x} _ {0}\\right) \\| p \\left(\\mathbf {x} _ {T} \\mid I\\right)\\right). (15) \\\\ \\end{array}", + "image_path": "cc57ea0d722c19303853d61ec3bffd1725269953b43b72709bd2027babe19475.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": "The first two terms can be optimized by standard gradient ascent. We approximate the expectations with Monte Carlo sampling with a single sample. The sum over the time variable " + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": " is also approximated by a single uniform sample over " + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\{1, \\dots, T\\}" + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": ". The KL divergence of the second term is the sum of pixel-wise KL divergences," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 670, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 670, + 545, + 713 + ], + "type": "interline_equation", + "content": "K L (q \\| p) = \\sum_ {d = 1} ^ {D} K L \\left(q \\left(x _ {t - 1} \\mid x _ {t}, x _ {0}\\right) \\| p _ {\\theta} \\left(x _ {t - 1} \\mid \\mathbf {x} _ {t}, I\\right)\\right), \\tag {16}", + "image_path": "650a7bdbbbb89807f5cff2e4c41adc623cc42ec5089b80f26b30589bcf6724e6.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1122" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 70, + 289, + 305 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": "Algorithm 1 Training a CCDM with " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " steps \nRequire: Training data expressed as the empirical distribu-. tion " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_0,I) = q(\\mathbf{x}_0\\mid I)q(I)" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " repeat t \\~ Uniform( " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\{1,\\dots,T\\}" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " 1 " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\begin{array}{rl} & {\\boldsymbol {I}\\sim q(\\boldsymbol {I})}\\\\ & {\\mathbf{x}_0\\sim q(\\mathbf{x}_0\\mid \\boldsymbol {I})}\\\\ & {\\mathbf{x}_t\\sim q(\\mathbf{x}_t|\\mathbf{x}_0)}\\\\ & {\\hat{\\mathbf{P}}_0\\leftarrow f_\\theta (\\mathbf{x}_t,I,t)} \\end{array}" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " ▷ shape " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "D\\times L" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " Pixel-wise application of Eq. (12) " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_{t - 1}\\gets \\sum_{x_0\\in \\mathcal{L}}\\pi (x_t,x_0)\\cdot \\hat{\\mathbf{p}}_0[x_0]\\quad \\triangleright" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " shape L " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " Compute KL with Eq. (8) and (16) " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\ell \\leftarrow KL(q(\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0)\\| p_\\theta (\\mathbf{x}_{t - 1}|\\mathbf{x}_t,I))" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " else " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\ell \\leftarrow -\\sum_d\\log \\mathcal{C}(x_0\\mid \\hat{\\mathbf{P}}_0[d])" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " end if " + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "inline_equation", + "content": "\\theta \\leftarrow \\theta -\\nabla_{\\theta}\\ell" + }, + { + "bbox": [ + 47, + 70, + 289, + 305 + ], + "type": "text", + "content": " ▷ Gradient descent until converged" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 327, + 287, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 362 + ], + "type": "text", + "content": "where the parameter vectors of distributions " + }, + { + "bbox": [ + 46, + 327, + 287, + 362 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 327, + 287, + 362 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 327, + 287, + 362 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 327, + 287, + 362 + ], + "type": "text", + "content": " are computed with Eqs. (7) and (12), respectively. Alg. 1 shows the complete training procedure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 362, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 411 + ], + "type": "text", + "content": "The third term of Eq. (15) does not depend on the learnable parameters " + }, + { + "bbox": [ + 46, + 362, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 362, + 287, + 411 + ], + "type": "text", + "content": " and is ignored during training. It is optimized by the design of the categorical diffusion model. Since the forward process converges as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 418, + 287, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 418, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 287, + 445 + ], + "type": "interline_equation", + "content": "\\lim _ {t \\rightarrow \\infty} q \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {C} \\left(x; \\frac {1}{L}\\right), \\tag {17}", + "image_path": "db35e99fa892217e10b8c9569e78e437bea9b2b50eb4cba7c257bb03074badf3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 453, + 282, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 282, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 282, + 466 + ], + "type": "text", + "content": "we fix " + }, + { + "bbox": [ + 47, + 453, + 282, + 466 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T\\mid I)" + }, + { + "bbox": [ + 47, + 453, + 282, + 466 + ], + "type": "text", + "content": " to the element-wise uniform distribution," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 474, + 287, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 474, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 93, + 474, + 287, + 500 + ], + "type": "interline_equation", + "content": "p \\left(x _ {T} \\mid I\\right) = p \\left(x _ {T}\\right) = \\mathcal {C} \\left(x _ {T}; \\frac {1}{L}\\right). \\tag {18}", + "image_path": "1de247ed9976408791864e0a80b20ce39d9829be05302c4222e83c1324f01ace.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 508, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 531 + ], + "type": "text", + "content": "This ensures that " + }, + { + "bbox": [ + 46, + 508, + 287, + 531 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T \\mid I) \\approx q(\\mathbf{x}_T \\mid \\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 508, + 287, + 531 + ], + "type": "text", + "content": ", making the third term of the ELBO close to zero." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "text", + "content": "At inference, the CCDM samples from " + }, + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0 \\mid I)" + }, + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "text", + "content": " to generate label maps for a given image " + }, + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "text", + "content": ", which is achieved by traversing the Markov chain of the reverse process as outlined in Alg. 2 and illustrated in Fig. 2. To minimize the noise of the generated label maps, the CCDM selects the label with maximum probability instead of sampling from " + }, + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(x_0 \\mid \\hat{\\mathbf{p}}_0)" + }, + { + "bbox": [ + 46, + 532, + 287, + 616 + ], + "type": "text", + "content": " in the final step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 623, + 153, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 153, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 153, + 635 + ], + "type": "text", + "content": "3.5. Architecture of " + }, + { + "bbox": [ + 47, + 623, + 153, + 635 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "As described above, the neural network " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " receives a label map " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": ", a time step " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": ", and an image " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " to estimate the probability parameters for " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": ". Its base design is a U-Net-like architecture [13] with self-attention modules at the three innermost layers of the encoder and the decoder [13]. The network processes the input label map represented as" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 306, + 87, + 545, + 256 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 515, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 515, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 515, + 84 + ], + "type": "text", + "content": "Algorithm 2 Inference from a CCDM with " + }, + { + "bbox": [ + 306, + 72, + 515, + 84 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 72, + 515, + 84 + ], + "type": "text", + "content": " steps" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": "Require: Input image " + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": " a network trained with Alg. 1 \n" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{C}^D(x_T; \\frac{1}{L})" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text{prev}} \\gets \\mathbf{x}_T \\quad \\triangleright" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": " Stores interm. and final prediction \nfor " + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "t = T, \\dots, 1" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_0 \\gets f_{\\theta}(\\mathbf{x}_{\\text{prev}}, I, t)" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": " \nif " + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": " then \n" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": " Pixel-wise application of Eq. (12) \n" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_{t-1} \\gets \\sum_{x_0 \\in \\mathcal{L}} \\pi(x_t, x_0) \\cdot \\hat{\\mathbf{p}}_0[x_0]" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text{prev}} \\sim \\prod_d \\mathcal{C}(x_{t-1} | \\hat{\\mathbf{p}}_{t-1})" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": " \nelse \n" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": " Final prediction \n" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text{prev}} \\gets \\arg \\max_{x_0 \\in \\mathcal{L}} \\hat{\\mathbf{P}}_0[:, x_0]" + }, + { + "bbox": [ + 306, + 87, + 545, + 256 + ], + "type": "text", + "content": " \nend if \nend for" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "text", + "content": "a binary tensor with " + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "text", + "content": " channels encoding the label of each pixel as a one-hot vector. Parameters of the network are shared for all values of " + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "text", + "content": ". The step variable " + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "text", + "content": " is encoded with the standard transformer sinusoidal position embedding [19] and concatenated as additional channels to the input tensor and to the feature maps of intermediate layers. Similarly, information from the input image " + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "text", + "content": " is presented to the network as raw pixel values concatenated to the input tensor as additional channels. In some experiments we used a pre-trained transformer architecture Dino-ViT [5] to extract informative visual features from the image " + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "text", + "content": ". In those cases, the extracted features were concatenated to the feature map of the third level of the U-Net encoder, which corresponds to a spatial shape equal to " + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 304, + 275, + 545, + 455 + ], + "type": "text", + "content": " the shape of the input image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 464, + 388, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 388, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 388, + 478 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 485, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 485, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 545, + 521 + ], + "type": "text", + "content": "In all our experiments, we set " + }, + { + "bbox": [ + 305, + 485, + 545, + 521 + ], + "type": "inline_equation", + "content": "T = 250" + }, + { + "bbox": [ + 305, + 485, + 545, + 521 + ], + "type": "text", + "content": " and the collection of " + }, + { + "bbox": [ + 305, + 485, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 305, + 485, + 545, + 521 + ], + "type": "text", + "content": " are set following the cosine schedule proposed in [36]. We evaluate our method on two tasks described below." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 528, + 515, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 528, + 515, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 528, + 515, + 540 + ], + "type": "text", + "content": "4.1. Segmentation with multiple annotations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 641 + ], + "type": "text", + "content": "Dataset The Lung Image Database Consortium (LIDC) [2] binary segmentation dataset consists of 1'018 three dimensional chest CT scans of patients with lung cancer. Lung nodules of each volume are annotated by four expert raters from a pool of 12, yielding large differences in annotations in some cases. We extract nodule-centered slices from the CT volumes and treat each slice as an independent image." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "content": "While LIDC is the standard benchmark of stochastic segmentation methods to date (e.g. [4, 15, 23, 25, 28, 29, 35, 43, 53, 54]), experimental configurations (pre-processing, training/validation/test splits, metrics) vastly differ across the literature. We conduct our experiments on the two most prominent LIDC splits and report results on both separately." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1123" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 213 + ], + "type": "table", + "html": "
MethodLIDCv1LIDCv2
GED16GED32GED50GED100HM-IoU16HM-IoU32GED16GED50GED100HM-IoU16
Prob. Unet [28]0.310±0.01-0.303±0.01+-0.252±0.004†0.552±0.00-0.548±0.00+0.320±0.030‡-0.252±‡0.500±0.030‡
HProb. Unet [29]0.270±0.01----0.530±0.01--0.270±0.010‡--0.530±0.01
PhiSeg [4]0.262±0.00-0.247±0.00+-0.224±0.004†0.586±0.00-0.595±0.00+----
SSN [35]0.259±0.00-0.243±0.01+-0.225±0.0020.558±0.00-0.555±0.01+----
cFlow [43]-0.225±0.01+---0.584±0.00+----
CAR [25]---0.228±0.009--0.264±0.0020.248±0.0040.243±0.0040.592±0.005
JProb. Unet [53]-0.206±0.00---0.647±0.010.262±0.00--0.585±0.00
PixelSeg [52]0.243±0.01---0.614±0.00-0.260±0.00--0.587±0.01
MoSE [15]0.218±0.003-0.195±0.0020.189±0.0020.624±0.004-----
AB [9]0.213±0.0010.196±0.0020.193±0.002-0.614±0.0010.619±0.001----
CIMD [38]0.234±0.0050.218±0.0050.210±0.005-0.587±0.0010.592±0.002----
CCDM (ours)0.212±0.0020.194±0.0010.187±0.0020.183±0.0020.623±0.0020.631±0.0020.239±0.0030.216±0.0030.210±0.0030.598±0.001
", + "image_path": "5a040d9ebf88f60512da1f6aeaf4beab6df9cdcd8486300417bab3d5e24200b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "type": "text", + "content": "Table 1: Quantitative results on LIDCv1 and LIDCv2, with the methods ordered by year. Bold and underlined indicate best and second best per column, respectively. Our results are over 3 seeds. For GED, lower is better; for HM-IoU, higher is better. No method, including ours, uses pre-trained weights. Results for CIMD [38] and AB [9] are ours. All other scores are taken from their original papers, except " + }, + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "type": "inline_equation", + "content": "(^{+})" + }, + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "type": "text", + "content": " from [53], " + }, + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "type": "inline_equation", + "content": "(-)" + }, + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "type": "text", + "content": " from [53], " + }, + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "type": "inline_equation", + "content": "(\\dagger)" + }, + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "type": "text", + "content": " from [35], " + }, + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 230, + 546, + 278 + ], + "type": "text", + "content": " from [25]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 299, + 287, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 358 + ], + "type": "text", + "content": "The first, referred to as LIDCv1, is used in [4, 15, 35, 53]. LIDCv1 comprises 15'096 slices, divided into training, validation, and testing sets with the ratio " + }, + { + "bbox": [ + 46, + 299, + 287, + 358 + ], + "type": "inline_equation", + "content": "60:20:20" + }, + { + "bbox": [ + 46, + 299, + 287, + 358 + ], + "type": "text", + "content": ". The second, LIDCv2, is used in [25, 28] and contains 12'816 images with the ratio " + }, + { + "bbox": [ + 46, + 299, + 287, + 358 + ], + "type": "inline_equation", + "content": "70:15:15" + }, + { + "bbox": [ + 46, + 299, + 287, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "text", + "content": "Metrics We measure the performances with the Generalised Energy Distance (GED) and the Hungarian-Matched Intersection over Union (HM-IoU) [15, 25, 29]. Both metrics measure the difference between the distributions of generated and ground-truth label maps. We denote the metrics computed with " + }, + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "text", + "content": " samples using a subscript, i.e., " + }, + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\mathrm{GED}_n" + }, + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "inline_equation", + "content": "\\mathrm{HM - IoU}_n" + }, + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "text", + "content": ", and we set " + }, + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 374, + 287, + 482 + ], + "type": "text", + "content": " to common values found in the literature. Note that higher number of samples yield more precise estimates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": "Baselines We compare our approach to eleven recent stochastic segmentation methods: probabilistic U-Net (Prob. Unet) [28], hierarchical probabilistic U-Net (HProb. Unet) [29], PhiSeg [4], stochastic segmentation network (SSN) [35], conditional normalizing flow (cFlow) [43], calibrated adversarial refinement (CAR) [25], joint probabilistic U-Net (JProb. Unet) [53], PixelSeg [52], mixture of stochastic experts (MoSE) [15], analog bits (AB) [9], and collectively intelligent medical diffusion (CIMD) [38]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "Following standard practice, we use random horizontal and vertical flipping and random rotations of " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "270^{\\circ}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " for data augmentation. The resolution of the input images is " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ". We trained our method with the Adam optimizer [26] until convergence of the GED metric on the validation set, a polynomial learning rate scheduling starting from " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " and ending with " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "1e^{-6}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ", and batch size of 64. We applied Polyak averaging with " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 0.99995" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 299, + 507, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 507, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 507, + 312 + ], + "type": "text", + "content": "4.2. Segmentation with a single annotation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 317, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 377 + ], + "type": "text", + "content": "We also evaluate our method with Cityscapes, a classical multi-class segmentation dataset where each image is annotated with a single label map. It comprises 2'975 RGB images of urban scenes for training and 500 images for validation, with each image labeled using 19 possible classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 378, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 424 + ], + "type": "text", + "content": "We compare our approach to several established baselines using the validation set: DeepLabv3 [7], HRNet [47], and UPerNet [51], with both ResNet [18] and Swin [32] backbones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 497 + ], + "type": "text", + "content": "Besides our standard method, which performs image conditioning by concatenating the raw pixel values as channels of the input tensor, we also included in our comparison a variant CCDM-Dino which leverages pre-trained DinoViT features [5] as additional conditioning concatenated to intermediate feature maps of our model's encoder." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "content": "Experiments are conducted separately for two different image resolutions: " + }, + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "inline_equation", + "content": "128 \\times 256" + }, + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "content": ". For all reported methods, we first resize the images to a fixed resolution and then apply color jittering, random flipping, and standard ImageNet intensity normalization as data augmentation. All baselines are trained for 500 epochs with a batch size of 32, with optimizers, learning rate schedules, and weight decay settings as reported in their respective publications (reported in detail in the supplementary material)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "text", + "content": "Our method was trained for 800 epochs with a batch size of 32 at " + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "inline_equation", + "content": "128 \\times 256" + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "text", + "content": " and of 16 at " + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "text", + "content": ", using the Adam optimizer [26] with a learning rate of " + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "text", + "content": " linearly decayed to " + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "inline_equation", + "content": "1e^{-6}" + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "text", + "content": ". We applied Polyak averaging with " + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\alpha = 0.999" + }, + { + "bbox": [ + 304, + 606, + 545, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Performance is measured with the mean intersection-over-union (mIoU). Unlike GED and HM-IoU, the metric mIoU is incompatible with multiple label maps per image. During inference, CCDM generates multiple label maps per image that are subsequently fused into a single label map for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1124" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "performance assessment. We found that fusing by averaging the predicted probabilities resulted in superior performances compared to fusing by majority vote." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 120, + 100, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 120, + 100, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 120, + 100, + 133 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 139, + 98, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 139, + 98, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 98, + 152 + ], + "type": "text", + "content": "5.1. LIDC" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 158, + 288, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 288, + 218 + ], + "type": "text", + "content": "We report performances on LIDCv1 and LIDCv2 in Tab. 1 and qualitative results in Fig. 3. Due to the lack of consistent evaluation protocols, we use a total of 10 metrics, thereby covering all the baselines and allowing for direct comparisons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 219, + 288, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 288, + 386 + ], + "type": "text", + "content": "Our CCDM reaches the best performance for eight out of the ten metrics, despite its relatively small size, with 9M parameters compared to, e.g., the 42M parameters of MoSE. CCDM also outperforms recent continuous diffusion models for segmentation, including AB [9] (9M parameters) and CIMD [38] (24M parameters). On HM-IoU16, the CCDM has a lower mean performance than MoSE by 0.001, but with only half the standard deviation. The JProb. Unet reaches a higher HM-IoU32 than all other methods, despite being considerably worse for GED32 than our CCDM. Furthermore, on LIDCv2, the JProb. Unet achieves only the third-best score on GED16, and fourth-best on HM-IoU16. This result indicates how comparing results obtained on different LIDC versions with each other can be misleading." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 387, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 458 + ], + "type": "text", + "content": "Fig. 3 presents qualitative results from our method. In columns (g)-(l), we see that our CCDM generates a distribution of samples that captures the annotation variability created by the four expert raters. Further, as seen in the bottom example, the CCDM also generates empty samples according to the annotations (b)-(e)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "text", + "content": "Reduced number of time steps for sampling: During inference, traversing the " + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "text", + "content": " steps of the reverse process makes sampling from DDPMs slow. A straightforward solution [36] involves traversing only a subset of nodes of the reverse process, " + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{k\\tau} : \\tau \\in \\{0, \\dots, T / k\\}\\}" + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "text", + "content": ", reducing the number of steps by a factor " + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "text", + "content": ". This technique accelerates inference at the expense of reduced performance. To illustrate the trade-offs between performance and speed, Fig. 5 presents the evolution of " + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{GED}_{16}" + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{HM-IoU}_{16}" + }, + { + "bbox": [ + 46, + 459, + 288, + 615 + ], + "type": "text", + "content": " as the number of inference steps is reduced. As expected, CCDMs perform best when the number of training and inference steps are equal, but a reasonable increase in speed without a large sacrifice in performance is possible." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 622, + 120, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 120, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 120, + 635 + ], + "type": "text", + "content": "5.2. Cityscapes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Experimental comparisons on Cityscapes are presented in Tab. 2, and qualitative examples are provided in Fig. 4. Experiments at " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "128 \\times 256" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " demonstrate that CCDM-Dino outperforms all other methods, even when only a single sample is used. CCDM-raw also remains competitive, being outperformed only by one baseline (UPerNet+Swin" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 537, + 218 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 537, + 218 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 537, + 218 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 537, + 218 + ], + "type": "table", + "html": "
MethodmIoU final (best)
ArchitectureBackbone#params128 × 256256 × 512
DeepLabv3 [7]ResNet50 (✓)39m43.4 (44.1)58.6 (59.2)
DeepLabv3 [7]ResNet101 (✓)58m43.8 (45.5)59.2 (59.8)
UPerNet [51]ResNet101 (✓)83m45.5 (47.1)60.7 (61.2)
HRNet [47]w48v2 (✓)70m48.2 (49.5)63.3 (64.2)
UPerNet [32]Swin-Tiny (✓)58m54.2 (55.9)65.5 (66.0)
CCDM (ours)-
samples=130m53.260.3
samples=530m55.462.0
samples=1030m56.262.4
CCDM (ours)Dino ViT-S (†)
samples=130m + 20M55.564.0
samples=530m + 20M56.965.4
samples=1030m + 20M57.365.8
", + "image_path": "91c93a2351a1e273850fa90ca7c121c16028415037f759da82f29ae4317d4d43.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 341, + 374, + 512, + 430 + ], + "blocks": [ + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "type": "text", + "content": "Table 2: Results on Cityscapes-val for resolutions " + }, + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "type": "inline_equation", + "content": "128 \\times 256" + }, + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "type": "text", + "content": ". **Bold** and **underlined** indicate best and second best per column, respectively. (" + }, + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "type": "text", + "content": ") and (" + }, + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 226, + 547, + 346 + ], + "type": "text", + "content": ") indicate supervised and self-supervised pretraining of the backbone, respectively. Gray indicates pretrained, non-finetuned parameters. We report final performance for our method and baselines. For the latter we also provide best achieved performance during training (in parenthesis). For CCDM methods, the field samples indicate the number of generated samples for label map fusion, as explained in Sect 4.2." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 341, + 374, + 512, + 430 + ], + "lines": [ + { + "bbox": [ + 341, + 374, + 512, + 430 + ], + "spans": [ + { + "bbox": [ + 341, + 374, + 512, + 430 + ], + "type": "table", + "html": "
CCDM CapacitymIoU(128 × 256)
#paramsUNet Levelssamples=1samples=5samples=10
5.4M437.839.740.6
7.5M544.748.348.5
22M451.654.053.6
30M553.255.456.2
", + "image_path": "0f674ce2554c6f0fc566e989dbc0ef7cf94233fc1d12fe5793e116a65829a945.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 438, + 545, + 462 + ], + "lines": [ + { + "bbox": [ + 305, + 438, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 545, + 462 + ], + "type": "text", + "content": "Table 3: Effect of increasing CCDM capacity (without feature conditioning)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "text", + "content": "Tiny), despite using only between " + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "text", + "content": " of the parameters of other models. Similarly, at " + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "text", + "content": ", CCDM-Dino outperforms four of the baselines with a single sample, lags behind UPerNet+Swin-Tiny only by 0.1 percent points with 5 samples, and outperforms all baselines with 10 samples. As expected, averaging across more samples improves performance for both CCDM-raw and CCDM-Dino, albeit with diminishing gains. Furthermore, the addition of Dino features boosts single-sample performance by 2.3 percent points at " + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "inline_equation", + "content": "128 \\times 256" + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "text", + "content": ", and 3.7 percent points at " + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 304, + 501, + 547, + 646 + ], + "type": "text", + "content": ", hinting the greater value of adding feature conditioning for generating segmentation at a higher resolution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "CCDM Capacity: Tab. 3(b) demonstrates the effect of increasing the capacity of CCDM. Using more U-Net feature levels, and increasing the number of parameters by doubling the number of channels per level, increases the performance regardless of the number of samples used for inference." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1125" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 533, + 252 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 533, + 252 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 533, + 252 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 533, + 252 + ], + "type": "image", + "image_path": "124261e915105a28395a1b92f13a05f954f52fb1b05d1e01069d9d8447e93011.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 288 + ], + "type": "text", + "content": "Figure 3: Qualitative results on four LIDC images with our method. (a) shows the image, (b)-(e) its four labels, (f) the mean prediction of our CCDM over six predictions, and (g)-(l) six individual predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 300, + 541, + 475 + ], + "blocks": [ + { + "bbox": [ + 52, + 300, + 541, + 475 + ], + "lines": [ + { + "bbox": [ + 52, + 300, + 541, + 475 + ], + "spans": [ + { + "bbox": [ + 52, + 300, + 541, + 475 + ], + "type": "image", + "image_path": "ddc294190fcc727fc452af9ee35de2171ac5486e599827f1e79a0f9c7552f6e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 485, + 546, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 485, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 546, + 510 + ], + "type": "text", + "content": "Figure 4: Qualitative comparisons on Cityscapes. All methods are trained and tested at a resolution of " + }, + { + "bbox": [ + 46, + 485, + 546, + 510 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 46, + 485, + 546, + 510 + ], + "type": "text", + "content": ". Our method produces structures with greater visual realism than other baselines. This is especially noticeable inside the marked regions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 91, + 528, + 242, + 630 + ], + "blocks": [ + { + "bbox": [ + 91, + 528, + 242, + 630 + ], + "lines": [ + { + "bbox": [ + 91, + 528, + 242, + 630 + ], + "spans": [ + { + "bbox": [ + 91, + 528, + 242, + 630 + ], + "type": "image", + "image_path": "460bab6719033622606de27821b83948ac968035bd01ee6e9cb79eb974dce2f3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 639, + 288, + 675 + ], + "lines": [ + { + "bbox": [ + 46, + 639, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 288, + 675 + ], + "type": "text", + "content": "Figure 5: LIDC GED and HM-IoU versus the number of sampling steps on LIDC. Evaluated on 500 random test images using 16 samples each, over 3 seeds." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 529, + 378, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 378, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 378, + 542 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "We introduced conditional categorical diffusion models (CCDMs) that are capable of effectively modeling pixel-level semantic distributions. Notably, and contrary to standard deterministic segmentation approaches, our model can produce diverse samples given an input image, thereby capturing the aleatoric uncertainty. Our method learns a multimodal label distribution of segmentations, induced by annotations from multiple expert raters, for which it achieves state-of-the-art results on a challenging medical imaging dataset, LIDC. Additionally, we demonstrate that it can achieve competitive performance on a standard multi-class semantic segmentation benchmark, Cityscapes, by outperforming several established, heavily engineered baselines" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1126" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 229, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 229, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 229, + 84 + ], + "type": "text", + "content": "despite using significantly fewer parameters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 287, + 217 + ], + "type": "text", + "content": "One limitation of our method is the requirement of several iterations for producing a sample, which is a common shortcoming of diffusion models. Accelerating sampling constitutes a crucial research direction, orthogonal to the present work. Finally, resolution scaling remains notoriously difficult for diffusion models, with successful examples relying on massive computational resources to train cascades of models that gradually increase resolution [20, 42] or operate on the latent space of existing embedding methods for continuous data (e.g. images) [39] that are not available for categorical data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 229, + 149, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 149, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 149, + 243 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 250, + 288, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 250, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 288, + 286 + ], + "type": "text", + "content": "This work was partially funded by the University of Bern, Swiss National Science Foundation Grants #320030-188591, #200021-192285 and #200021-191983." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 299, + 107, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 107, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 107, + 312 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 319, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 319, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 287, + 352 + ], + "type": "text", + "content": "[1] Tomer Amit, Eliya Nachmani, Tal Shaharbany, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 354, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 288, + 430 + ], + "type": "text", + "content": "[2] Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, et al. The lung image database consortium (lidc) and image database resource initiative (idri): a completed reference database of lung nodules on ct scans. Medical physics, 38(2):915-931, 2011. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 433, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 287, + 475 + ], + "type": "text", + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 478, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 287, + 576 + ], + "type": "text", + "content": "[4] Christian F Baumgartner, Kerem C Tezcan, Krishna Chaitanya, Andreas M Hotker, Urs J Muehlematter, Khoschy Schawkat, Anton S Becker, Olivio Donati, and Ender Konukoglu. Phiseg: Capturing uncertainty in medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part II 22, pages 119-127. Springer, 2019. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 578, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 287, + 633 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 287, + 689 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for seman-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 97, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 97, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 312, + 97, + 545, + 152 + ], + "type": "text", + "content": "[8] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 154, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 154, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 154, + 545, + 196 + ], + "type": "text", + "content": "[9] Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 199, + 545, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 199, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 199, + 545, + 254 + ], + "type": "text", + "content": "[10] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. Advances in Neural Information Processing Systems, 34:9355-9366, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 300 + ], + "type": "text", + "content": "[11] Max Cohen, Guillaume Quispe, Sylvain Le Corff, Charles Ollion, and Eric Moulines. Diffusion bridges vector quantized variational autoencoders. arXiv preprint arXiv:2202.04895, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 302, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 335 + ], + "type": "text", + "content": "[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. arXiv preprint arXiv:2209.04747, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 337, + 545, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 359 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. CoRR, abs/2105.05233, 2021. 1, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 415 + ], + "type": "text", + "content": "[14] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 418, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 460 + ], + "type": "text", + "content": "[15] Zhitong Gao, Yuong Chen, Chuyu Zhang, and Xuming He. Modeling multimodal aleatoric uncertainty in segmentation with mixture of stochastic expert. arXiv preprint arXiv:2212.07328, 2022. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 529 + ], + "type": "text", + "content": "[16] Jiaqi Gu, Hyoukjun Kwon, Dilin Wang, Wei Ye, Meng Li, Yu-Hsin Chen, Liangzhen Lai, Vikas Chandra, and David Z Pan. Multi-scale high-resolution vision transformer for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12094-12103, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 532, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 585 + ], + "type": "text", + "content": "[17] Adam W Harley, Konstantinos G Derpanis, and Iasonas Kokkinos. Segmentation-aware convolutional networks using local attention masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5038-5047, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "text", + "content": "[20] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 9" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1127" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[21] Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forre, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in Neural Information Processing Systems, 34:12454-12465, 2021. 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 193 + ], + "type": "text", + "content": "[22] Minghui Hu, Yujie Wang, Tat-Jen Cham, Jianfei Yang, and Ponnuthurai N Suganthan. Global context with discrete diffusion in vector quantised modelling for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11502-11511, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 194, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 270 + ], + "type": "text", + "content": "[23] Shi Hu, Daniel Worrall, Stefan Knegt, Bas Veeling, Henkjan Huisman, and Max Welling. Supervised uncertainty quantification for segmentation with multiple annotations. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22, pages 137–145. Springer, 2019. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 271, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 288, + 304 + ], + "type": "text", + "content": "[24] Guillaume Jeanneret, Loïc Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. arXiv preprint arXiv:2203.15636, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "type": "text", + "content": "[25] Elias Kassapis, Georgi Dikov, Deepak K Gupta, and Cedric Nugteren. Calibrated adversarial refinement for stochastic semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7057-7067, 2021. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 360, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 392 + ], + "type": "text", + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference for Learning Representations, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 437 + ], + "type": "text", + "content": "[27] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 502 + ], + "type": "text", + "content": "[28] Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic u-net for segmentation of ambiguous images. Advances in neural information processing systems, 31, 2018. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 503, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 558 + ], + "type": "text", + "content": "[29] Simon AA Kohl, Bernardino Romera-Paredes, Klaus H Maier-Hein, Danilo Jimenez Rezende, SM Eslami, Pushmeet Kohli, Andrew Zisserman, and Olaf Ronneberger. A hierarchical probabilistic u-net for modeling multi-scale ambiguities. arXiv preprint arXiv:1905.13077, 2019. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 559, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 591 + ], + "type": "text", + "content": "[30] Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. Tabddpm: Modelling tabular data with diffusion models. arXiv preprint arXiv:2209.15421, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "text", + "content": "[31] Lielei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep hierarchical semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1246-1257, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "text", + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[33] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Pro-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "text", + "content": "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 150, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 215 + ], + "type": "text", + "content": "[35] Miguel Monteiro, Loic Le Folgoc, Daniel Coelho de Castro, Nick Pawlowski, Bernardo Marques, Konstantinos Kamnitsas, Mark van der Wilk, and Ben Glocker. Stochastic segmentation networks: Modelling spatially correlated aleatoric uncertainty. Advances in Neural Information Processing Systems, 33:12756-12767, 2020. 2, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "type": "text", + "content": "[36] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 5.7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "text", + "content": "[37] Di Qiu and Lok Ming Lui. Modal uncertainty estimation via discrete latent representation. arXiv preprint arXiv:2007.12858, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 294, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 349 + ], + "type": "text", + "content": "[38] Aimon Rahman, Jeya Maria Jose Valanarasu, Ilker Hacihaliloglu, and Vishal M Patel. Ambiguous medical image segmentation using diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11536-11546, 2023. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 404 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "text", + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 525 + ], + "type": "text", + "content": "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 526, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 613 + ], + "type": "text", + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 614, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 689 + ], + "type": "text", + "content": "[43] Raghavendra Selvan, Frederik Faye, Jon Middleton, and Akshay Pai. Uncertainty quantification in medical image segmentation with normalizing flows. In Machine Learning in Medical Imaging: 11th International Workshop, MLMI 2020, Held in Conjunction with MICCAI 2020, Lima, Peru, October 4, 2020, Proceedings 11, pages 80-90. Springer, 2020. 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "type": "text", + "content": "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1128" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[45] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 274 + ], + "type": "text", + "content": "[46] MM Amaan Valiuddin, Christiaan GA Viviers, Ruud JG van Sloun, Peter HN de With, and Fons van der Sommen. Improving aleatoric uncertainty quantification in multi-annotated medical image segmentation with normalizing flows. In Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, and Perinatal Imaging, Placental and Preterm Image Analysis: 3rd International Workshop, UNSURE 2021, and 6th International Workshop, PIPPI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 3, pages 75–88. Springer, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 275, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 329 + ], + "type": "text", + "content": "[47] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, Wenyu Liu, and Bin Xiao. Deep high-resolution representation learning for visual recognition. TPAMI, 2019. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[48] Julia Wolleb, Robin Sandkühler, Florentin Bieder, Philippe Valmaggia, and Philippe C Cattin. Diffusion models for implicit image segmentation ensembles. arXiv preprint arXiv:2112.03145, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 376, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 418 + ], + "type": "text", + "content": "[49] Junde Wu, Huihui Fang, Yu Zhang, Yehui Yang, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. arXiv preprint arXiv:2211.00611, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[50] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, and Yanwu Xu. Medsegdiff-v2: Diffusion based medical image segmentation with transformer. arXiv preprint arXiv:2301.11798, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[51] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[52] Wei Zhang, Xiaohong Zhang, Sheng Huang, Yuting Lu, and Kun Wang. Pixelseg: Pixel-by-pixel stochastic semantic segmentation for ambiguous medical images. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4742-4750, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 623 + ], + "type": "text", + "content": "[53] Wei Zhang, Xiaohong Zhang, Sheng Huang, Yuting Lu, and Kun Wang. A probabilistic model for controlling diversity and accuracy of ambiguous medical image segmentation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4751-4759, 2022. 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[54] Yifan Zhang, Bo Pang, and Cewu Lu. Semantic segmentation by early region proxy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1258-1268, 2022. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[55] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[56] Roland S Zimmermann, Lukas Schott, Yang Song, Benjamin A Dunn, and David A Klindt. Score-based generative classifiers. arXiv preprint arXiv:2110.00473, 2021. 1" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1129" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Story Visualization by Online Text Augmentation with Context Memory/b1180507-958c-45ce-9ffe-139cc0187a6e_content_list.json b/2023/Story Visualization by Online Text Augmentation with Context Memory/b1180507-958c-45ce-9ffe-139cc0187a6e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..36b0948ec01f1be33610549c3c39e2071816e3f2 --- /dev/null +++ b/2023/Story Visualization by Online Text Augmentation with Context Memory/b1180507-958c-45ce-9ffe-139cc0187a6e_content_list.json @@ -0,0 +1,1658 @@ +[ + { + "type": "text", + "text": "Story Visualization by Online Text Augmentation with Context Memory", + "text_level": 1, + "bbox": [ + 120, + 130, + 848, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daechul Ahn $^{1,\\S}$ Daneul Kim $^{2}$ Gwangmo Song $^{3}$ Seung Hwan Kim $^{3}$ Honglak Lee $^{3,4}$ Dongyeop Kang $^{5}$ Jonghyun Choi $^{1,\\dagger}$", + "bbox": [ + 176, + 179, + 790, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Yonsei University $^{2}$ GIST $^{3}$ LG AI Research $^{4}$ University of Michigan $^{5}$ University of Minnesota $\\{\\mathrm{dcahn, jc}\\} @$ yonsei.ac.kr flytodk98@gm.gist.ac.kr gwangmo.song@lgresearch.ai skcruise@gmail.com honglak@eecs.umich.edu dongyeop@umn.edu", + "bbox": [ + 102, + 227, + 870, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 323, + 313, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Story visualization (SV) is a challenging text-to-image generation task for the difficulty of not only rendering visual details from the text descriptions but also encoding a long-term context across multiple sentences. While prior efforts mostly focus on generating a semantically relevant image for each sentence, encoding a context spread across the given paragraph to generate contextually convincing images (e.g., with a correct character or with a proper background of the scene) remains a challenge. To this end, we propose a novel memory architecture for the Bi-directional Transformer framework with an online text augmentation that generates multiple pseudo-descriptions as supplementary supervision during training for better generalization to the language variation at inference. In extensive experiments on the two popular SV benchmarks, i.e., the Pororo-SV and Flintstones-SV, the proposed method significantly outperforms the state of the arts in various metrics including FID, character F1, frame accuracy, BLEU-2/3, and R-precision with similar or less computational complexity.", + "bbox": [ + 75, + 356, + 473, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 672, + 209, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Story visualization (SV) [17] is a task of generating a sequence of images from a paragraph, i.e., a sequence of natural language sentences. It is challenging for the requirement of rendering the visual details in images with convincing background of a scene – seasonal elements, environmental objects such as table, location, and the proper character appearing, which here we refer to as context, spread across the given text sentences. Specifically, it needs to encode implicit context presented in the given sentences since each one often omit visual details (i.e., they may be spread over the sentences) necessary to generate a semantically correct im", + "bbox": [ + 75, + 698, + 470, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fd9a836d27890f7568aea7b4d5486157e940b4cdceec2ecfcc90611105eeaf7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 323, + 888, + 453 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1df1c4aca7b4b3bf57e58ddf623eb289f08b0c7e3d4bc98353703be65c9d7dfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 460, + 887, + 592 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ae9701ae8dcc4ac7bd78f0737dcd051f2db78d57e94f325ff2dbd2a0f922b43.jpg", + "image_caption": [ + "Figure 1. Linguistic variations in story visualization and the overview of the proposed method. (a) An example of a story data with various linguistic variations (Var.#) for each image. (b) Modeling temporal context spread across sentences by our context memory (Sec. 3.1). (c) Addressing linguistic variations by online text augmentation from each image for every epoch (Sec. 3.2). (d) Benefit of the proposed context memory (left) and online text augmentation (right)." + ], + "image_footnote": [], + "bbox": [ + 503, + 595, + 890, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age. For example, we can think of Fig. 1-(d), where Sent.#1 (i.e., sentence 1) and Sent.#2 (i.e., sentence 2) are given as sequences. After Sent.#1 is given, generated image by Sent.#2 often exhibits a background that is not semantically", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "§: work done while interning at LG AI Research. †: corresponding author.", + "bbox": [ + 73, + 875, + 470, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Code: https://github.com/yonseivnl/cmota", + "bbox": [ + 73, + 887, + 383, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3125", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "correct [17,22,23]. However, if we use contextual information given in Sent.#1, \"The woods are covered with snow\", it leads to correct image with matching background.", + "bbox": [ + 75, + 90, + 470, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, widely used benchmark datasets for story visualization task [9, 17] provide a single text-image pair for training and inference, mostly due to the annotation budget constraint. This prevents the model from learning language variations, and thus harms the linguistic generalization performance of the model.", + "bbox": [ + 75, + 137, + 468, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges without requiring large scale data and models, we propose a new memory scheme in bi-directional Transformer for encoding the context which generates pseudo-texts in an online fashion to address linguistic variations at inference. We call our model as Context Memory and Online Text Augmentation or CMOTA for short. We empirically validate that our model outperforms the state-of-the-art SV methods by large margins on various metrics evaluated with widely used benchmarks in the literature, i.e., Pororo-SV and Flintstones-SV.", + "bbox": [ + 75, + 228, + 470, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that while large pre-trained models [3,5-7, 31, 43] have shown great success in synthesizing an image or a video from a language description [13, 32, 38], huge computational complexity and large training data makes the models prohibited. Moreover, although we propose and evaluate the model for the standard benchmark datasets without large pretraining data trained with a large model, it would be interesting to apply and evaluate the proposal in large models for further improvement.", + "bbox": [ + 75, + 380, + 470, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 96, + 516, + 387, + 530 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new memory architecture for Transformer to selectively make use of contexts in a story paragraph.", + "- For better generalization of linguistic variations in the given paragraph at inference, we generate pseudo-texts and augment them in an online fashion for richer linguistic supervision.", + "- Our model significantly outperforms prior arts (even some hyper-scale models) by large margins in five evaluation metrics with similar or less computational complexity." + ], + "bbox": [ + 96, + 537, + 468, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 708, + 218, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Story visualization. StoryGAN [17] is one of the recent methods that utilized a story-level discriminator to improve global consistency in generated images. CP-CSV [39] disentangles figure and background information to enhance character consistency. In order to improve global semantic matching between paragraph and generated image sequence, DuCo-StoryGAN [23] presents a pre-trained video captioner as an auxiliary loss along with other design improvements on top of StoryGAN. More recently, VLC-StoryGAN [22] utilizes constituency parse-trees and common sense knowledge to improve consistency and an object-level feedback loop to", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "improve image quality. Another recent work VP-CSV [2] is a two-stage approach, i.e., 1) character generation and 2) background completion, using Transformer model to address this task. We discuss these in more detail in the supplementary material for the space sake. But the prior arts largely neglect to encode the story narrative, which is our primary contribution here.", + "bbox": [ + 496, + 90, + 890, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Very recently, Maharana et al. propose a new task setup of story continuation; using first image as a condition. They fine-tune the large model DALL-E [32] for the SV, which they call StoryDALL-E [24]. Ours differs from it in several aspects as follows. We have explicit memory connections between adjacent image generators using the context memory to globally encode the sentences. In contrast, [24] utilizes global story embedding as additional input for understanding context. Further, while it is huge in size (1.3B parameters), trained with 14 million text-image pair, ours are much larger (97M parameters) and outperforms it in multiple metrics by large margins; even in the image quality metric, FID. Please refer to Sec. 4.2 for empirical comparisons.", + "bbox": [ + 496, + 196, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-video generation. Similar to story visualization, text-to-video generation also generates multiple frames from a given text. Since the pioneering work of Sync-DRAW [25] in text-to-video generation, [9, 18, 27] utilize generative adversarial network for high-quality image generation.", + "bbox": [ + 496, + 405, + 890, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, high-quality video generation models are proposed, including GODIVA [46] and NUWA [47]. Furthermore, recent studies generate high-resolution videos, making sequential frames in high-quality [11, 38]. However, most of state-of-the-art text-to-video model [11, 38] generates a video from a single sentence, mostly having consistent backgrounds. Very recently, there is a method proposed to generate a video from a long paragraph [45]. Although they generate the video in a long time horizon, they require both a mega-scale model trained with huge data and a detailed paragraph where each sentence is describing the scene that are close in time. In contrast, story visualization requires generating frames arbitrarily distant in time (i.e., so-called 'key-frames') corresponding to different sentences, requiring to generate an image sequence that have contextually convincing background.", + "bbox": [ + 496, + 481, + 892, + 722 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image generation. Text-to-image generation is a sub-problem of story visualization, with literature focusing on semantic relevance and resolution improvements. Recently, text-based image synthesis has been greatly improved with the help of a vast amount of training data with a hyper-scale model including DALL-E [32] and its successor DALL-E2, CogView [5] and Make-A-Scene [8] using a sketch input.", + "bbox": [ + 496, + 734, + 890, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although text-to-image generation models generates very high-quality images, it may lack encoding the context, metaphoric sentences spread acorss multiple sentences. In", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3126", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "addition, naively using state-of-the-art text-to-image generation models is computationally prohibited. For example, diffusion-based models [31, 34] have hyper-scale model size, (e.g., Imagen [34] parameter count of 2-B, DALL-E2 [31] parameter count of 3.5-B) making it non-trivial for applying it in a wide range of inference scenarios that may not have the sufficient computing resource. Here, we consider relatively light architectures as our base model for computational efficiency. More discussion are in the supplement.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer using memory. To mitigate context fragmentation issue [4], i.e., losing long-term dependency over a data stream in context, there are efforts to encode long-term contexts in generating an image sequence. [4, 15] adopt a recurrent path into transformer architecture. Specifically, the modeling of new data segments is conditioned on historical hidden states produced in the previous time step and uses highly summarized memory states. Unlike the conventional memory architectures, we propose a novel memory that has a dense connection from the past with attentive weighting schemes for their better usage (Sec. 3.1).", + "bbox": [ + 75, + 241, + 472, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Online augmentation. While offline data augmentation that prepares data outside of learning process is prevalent in many literature [1, 12, 19], online-augmentation that depends on training is seldom explored especially in story visualization task [40]. [35, 40] propose an online augmentation for image classification and VQA task. [26] uses bi-level optimization in image classification. In medical domain, [48] investigate the online augmentation for personalized image based Histopathology diagnosis [48]. Note that we show the benefit of online augmentation over the offline augmentation in the SV for the first time in this literature.", + "bbox": [ + 75, + 421, + 472, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 602, + 187, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are multiple challenges in story visualization, including (1) generating semantically natural images without artifacts from a description, (2) encoding the context (e.g., consistent background of a scene - seasonal elements, environmental objects such as table and location, and consistent characters) spread across the sentences in a given paragraph [17] and (3) addressing the linguistic variation at the inference time (i.e., a given text description may not be in different writing style).", + "bbox": [ + 75, + 628, + 470, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For photo-realistic image generation from a language description, we recently witness unprecedented improvements in the quality of generated images by the help of large scale model [3,5-7, 31, 43]. To leverage the large model's benefits for the story visualization task, we may use it to generate an image sequence by gradually concatenating sentences to visualize the story (i.e., use the first sentence to generate the first image, use a concatenated sentence of the first and the second to generate the second, and so on). Although this", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3955a06e84f176a10a7ea75926fdecc9626cee52112529d11c1c239d5cc11879.jpg", + "image_caption": [ + "Figure 2. Bi-directional (between text and image modality) Transformer for text or image 'sequence' generation with the proposed memory. Our bi-directional 'base' model can simultaneously generate a sequence of both text and image from $t = 1 \\ldots 5$ with the proposed memory (Sec. 3.1). We use a segment embedding to indicate input modality and enable output of any modality." + ], + "image_footnote": [], + "bbox": [ + 504, + 85, + 694, + 262 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3b70e431c038215fce69a5c897e9a5815e323660565cd9c0737be7af8c8f5b4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 85, + 895, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "may generate a single photo-realistic image, it is not able to capture the context sparsely spread across the sentences as story progresses (see supplement for more discussion).", + "bbox": [ + 496, + 375, + 890, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A more involved way of using the large model for SV is to fine-tune it for the downstream SV or using a video generation models. However, as the most high-performing models are huge in size and the codes and the pre-trained models are not publicly available, the computational cost and reproducibility reduces practicality.", + "bbox": [ + 496, + 419, + 890, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a given dataset without using large extra training data, we propose a new memory module to better encode past information with a Transformer by an attentively weighted densely connected architecture (Sec. 3.1). We further propose to generate pseudo-texts during the learning process to augment (online-augmentation) for better linguistic generalization without requiring large external data by learning the bi-directional Transformer [14, 30, 33] in both directions of generating images from texts and vice versa as illustrated in Fig. 2 (Sec. 3.2).", + "bbox": [ + 496, + 510, + 893, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Base model. As shown in Fig. 2, we use bi-directional (i.e., multi-modal) transformer that iteratively generates images and texts in both ways. Similar to [32], the image tokens are sequentially predicted from the input text sequences by the Transformer. Then, the decoder of the VQ-VAE [42] translates the predicted image tokens into image sequence. The text tokens are also sequentially predicted from the input image token sequence by the same Transformer.", + "bbox": [ + 496, + 674, + 893, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Particularly, for the bi-directional multi-modal generation, i.e., generating simultaneously text and image from the unified architecture, we add two embeddings; a positional embedding for absolute position between tokens and a segment embedding for distinguishing source and target. Tokens of a text $\\left(\\{t_{1},\\ldots ,t_{m}\\}\\right)$ and an image $\\left(\\{z_{1},\\ldots ,z_{n}\\}\\right)$ $(m,n:\\#$ of tokens for text and image) are fed into the Trans", + "bbox": [ + 496, + 795, + 895, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3127", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ad8c26c496bb10c7838063ee34de59e890db959c6be088b6191e309a03c38613.jpg", + "image_caption": [ + "Figure 3. Proposed memory module for the bi-directional Transformer. Memory module updates current memory $M_{t}$ by utilizing memory attention and GRU operation with propagated memory, $M_{t-1}$ , and previously calculated hidden state, $H_{t}^{l}$ . To produce image, we propose a novel memory attention (Attn.) mask denoted by blue squares to select only text tokens for less bias to previously generated image tokens. Within the memory attention mask, white-colored squares filter out the content, i.e., image (img) token, whereas the blue squares allow the text (txt) token to propagate through for memory interaction." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 472, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "former to predict tokens in the other modality in multiple epochs with the following objective function written as:", + "bbox": [ + 75, + 383, + 468, + 414 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {j, t 2 i} = \\sum_ {k = 1} ^ {n} - \\ln p _ {j} (z _ {k} | t _ {1}, \\dots , t _ {m}, z _ {1}, \\dots , z _ {k - 1}),\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 425, + 421, + 464 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {j, i 2 t} = \\sum_ {k = 1} ^ {m} - \\ln p _ {j} \\left(t _ {k} \\mid z _ {1}, \\dots , z _ {n}, t _ {1}, \\dots , t _ {k - 1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 467, + 468, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {j} = \\mathcal {L} _ {j, t 2 i} + \\lambda_ {1} \\mathcal {L} _ {j, i 2 t},\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 508, + 284, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_j(\\cdot)$ is a likelihood of $j$ -th generated tokens in one modality given the other modality, $L_{j,t2i}$ is a loss corresponding to $j$ -th text-to-image generation (i.e., negative log likelihood), $L_{j,i2t}$ is vice versa. $\\lambda_1$ is a balancing hyperparameter.", + "bbox": [ + 75, + 536, + 468, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train the model, we iteratively train the generation model in each modality multiple times per each epoch.", + "bbox": [ + 75, + 612, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Context Memory", + "text_level": 1, + "bbox": [ + 76, + 651, + 243, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To encode the context and calculate the propagated memory, as depicted in Fig. 3, we first apply cross attention between the current hidden state, $H_{t}^{l} \\in \\mathcal{R}^{T_{c} \\times d}$ calculated from $(l - 1)$ -th transformer layer and memory state at time $(t - 1)$ , $M_{t - 1} \\in \\mathcal{R}^{T_{M} \\times d}$ ( $T_{c}$ : # tokens, $T_{M}$ : # memory states, $d$ : hidden state dimension). We then obtain $S_{t} = \\text{Attn}(M_{t - 1}, H_{t}, H_{t})$ , where $\\text{Attn}(Q, K, V) \\coloneqq \\text{Softmax}\\left(\\frac{QK^{T}}{\\sqrt{d_{q}}}\\right)V$ and $d_{q}$ is a query dimension ( $Q$ ), with the memory attention mask depicted in the blue box in Fig. 3.", + "bbox": [ + 75, + 674, + 472, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, we apply a memory mask in the attention operation to select text tokens as memory content (depicted as blue-shaded grid cells in the right of Fig. 3) because including image content as memory could be strong constraint for text-to-image generation (see empirical studies of the", + "bbox": [ + 75, + 825, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/896d69dc3fe965aaaea659183cc0f3e253e467e30dd31e1970137e125d9668c9.jpg", + "image_caption": [ + "(a) Conventional memory connection (Dai et al., ACL'19/Lei et al., ACL'20) Layer #" + ], + "image_footnote": [], + "bbox": [ + 500, + 103, + 885, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bed5f34fa3f0b433bf9403ca37139612e6068ddffadc1f83c8bd2baec37415b7.jpg", + "image_caption": [ + "(b) Proposed memory connection", + "Figure 4. Comparison of our memory connection scheme to the conventional one [4,15]. We depict the memory connections in the multi-layer Transformer architecture per each sentence input (unfolded in a row). Unlike (a) the previous memory architecture having a single memory path from the immediate past in all layers, (b) ours has a dense connection from all the pasts to be attentively weighted (denoted here by $\\bar{M}$ 's) in only the last layer. We discuss its empirical benefit in Sec. 4.2.2." + ], + "image_footnote": [], + "bbox": [ + 500, + 251, + 885, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "proposed mask in the supplementary material). Then we feed the $S_{t}$ and $M_{t-1}$ to the GRU in order to compute the information to be stored in the memory, $M_{t}$ , propagated to $(t+1)$ .", + "bbox": [ + 496, + 525, + 892, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory connection. Following [4, 15], we stack multiple Transformer's layers, depicted in Fig. 4-(b). Compared to conventional memory updating architectures with serial hidden state connections in all layers (Fig. 4-(a)), our approach only has memory connections in the last layer, which we call it as partial-level memory augmented (PMA). This is because the later layers (closer to the last layer) achieve better representations with higher-level features, abstract and structured representations [28, 29, 41] with improved computational efficiency, which is unlike prior works (see Sec. 4.2.2 and supp for further discussion).", + "bbox": [ + 496, + 603, + 893, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, not all historic information is equally important for generating an image at a time step. Similar to the masked self-attention [43], we attentively weight the past information for better modeling of sparse context as:", + "bbox": [ + 496, + 770, + 893, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {M} _ {1: (t - 1)} = A t t n (M _ {(t - 1)}, M _ {[ 1: (t - 2) ]}, M _ {[ 1: (t - 2) ]}),\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 843, + 846, + 862 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {M} _ {(t - 1)} = \\left[ M _ {t - 1}; \\bar {M} _ {1: (t - 1)} \\right], \\quad (3 \\leq t \\leq 5), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 864, + 890, + 882 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {t} ^ {l} = \\operatorname {A t t n} \\left(H _ {t} ^ {l}, \\left[ H _ {t} ^ {l}; \\tilde {M} _ {(t - 1)} \\right], \\left[ H _ {t} ^ {l}; \\tilde {M} _ {(t - 1)} \\right]\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 885, + 854, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3128", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{(t - 1)}$ is a memory at time $(t - 1)$ and $[1:(t - 2)]$ refers to the concatenation from time 1 to $(t - 2)$ as depicted in Fig. 4-(b). By doing so, we fuse the contextual information into current hidden state $H_{t}$ . At $t = 2$ , we use $M_{1}$ , instead of $\\tilde{M}_{1}$ as the $M_{1}$ is only available. We call it as attentively weighted memory (AWM).", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Online Text-Augmentation", + "text_level": 1, + "bbox": [ + 76, + 191, + 318, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vedamtam et al. argue that multiple descriptions for an image help generalization as they address language variations presented in descriptions [44]. Hence, a number of image captioning datasets [10, 20, 37] provide multiple natural language directives, obtained by multiple human annotators, in both training and evaluation splits. But the SV benchmark datasets [17] provide only a single sentence per an image.", + "bbox": [ + 75, + 214, + 470, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the linguistic variations of a text input at inference process, we first consider to generate a pseudo text by a well-trained image-to-text generation model, which we refer it as offline-augmentation [16]. But the offline augmentation generates only a single sentence, which may not provide sufficient diversity. Instead, we propose to generate multiple pseudo-texts and augment them in an online fashion when training our model to increase the diversity. We call it online text augmentation, depicted in Fig. 5.", + "bbox": [ + 75, + 320, + 470, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thanks to our bi-directional multi-modal architecture, we can naturally integrate the process of generating pseudotexts to the process of learning image-to-text model and the text-to-image generation model as depicted in Fig. 5. In the early epochs, less meaningful sentences are generated, but as training progresses, more meaningful sentences are generated (see orange box in Fig. 5). As a side-product, by supervising the model learning with intermediate goals at each time step, we expect to expedite the convergence of learning. When we use the online text augmentation, we can rewrite the objective as:", + "bbox": [ + 75, + 455, + 472, + 623 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {j, p t 2 i} = \\sum_ {k = 1} ^ {n} - \\ln p _ {j} \\left(z _ {k} \\mid \\hat {t} _ {1}, \\dots , \\hat {t} _ {m}, z _ {1}, \\dots , z _ {k - 1}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 633, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {j} = \\mathcal {L} _ {j, t 2 i} + \\lambda_ {1} \\mathcal {L} _ {j, i 2 t} + \\lambda_ {2} \\mathcal {L} _ {j, p t 2 i},\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 676, + 369, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{j,pt2i}$ is the additional loss with the augmented pseudo-texts and $\\mathcal{L}_{j,t2i},\\mathcal{L}_{j,i2t}$ are defined in Eq. 1 and $\\lambda_{1}$ and $\\lambda_{2}$ are balancing hyper-parameters. $\\hat{t}$ means the pseudotext token, predicted during online augmentation without gradient flow. Fig. 6 shows pseudo-texts generated by our method. Detailed training procedure and more examples are in the supplementary material.", + "bbox": [ + 75, + 703, + 470, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 821, + 209, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 847, + 267, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use two popular benchmark datasets for evaluating the task of story visualization. Following [23], we use", + "bbox": [ + 75, + 869, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/04d01c9788f45c18214ccf229c06ae564bd6c3ce9a6ce0020c44d05e099441e7.jpg", + "image_caption": [ + "Figure 5. Online text augmentation. We iteratively augment the pseudo-text for better linguistic generalization. (1) generating a pseudo-text at each epoch for each ground-truth image, without gradient flow. (2) generating an image for every epoch. (3,4) learn the text-to-image model with $L_{CE}$ . In early epochs, it generates less meaningful pseudo-text (thus discard them), but produces pseudotexts matching with the input image as the training progresses." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pororo-SV dataset for story visualization task without data overlap. Following prior work [22], we use Flintstones-SV which was originally exploited in the text-to-video synthesis task [9].", + "bbox": [ + 496, + 329, + 893, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. For the evaluation metrics, we use Fréchet Inception Distance (FID) (used in [22, 39]) for visual quality of the generated images, character classification score (Char. F1, Frm. Acc.) [17] for character consistency, and global semantic matching (B-2/3, R-Prec.) [22, 23] between generated image sequence and story paragraph.. We elaborate on the details of the datasets and implementation in the supplementary material for the sake of space.", + "bbox": [ + 496, + 404, + 893, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our method to a number of prior arts including state-of-the-art story visualization methods such as StoryGAN [17], CPCSV [39], DuCo-StoryGAN [23], VLC-StoryGAN [22] and VP-CSV [2] for both quantitative and qualitative analysis in the story visualization setting. Moreover, we compare [24] in the story continuation setting.", + "bbox": [ + 496, + 539, + 893, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. To train our model, we first tokenize the text and image inputs. Particularly, the image inputs are encoded by the encoder of the VQ-VAE [42] into image tokens and decoded by the decoder of the VQ-VAE for generating image sequence. Once the inputs are tokenized, we append two special embedding tokens to the text and image tokens. These tokens signify the beginning of each modality, denoted as 'SOS' (start of sentence) and 'SOI' (start of image). Subsequently, we add two embeddings: (1) a positional embedding to indicate absolute position of each token and (2) a segment embedding to distinguish source and target modality, e.g., for text-to-image generation task, the source corresponds to the text while the target corresponds to the image. Through the utilization of segment embeddings to distinguish between source and target, we enable the model to generate target data from source data", + "bbox": [ + 496, + 657, + 895, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3129", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/50f14c46cbb54df636f9d29a4eaf74fcb6fb20a0475a951b141c0b65cbd9d59b.jpg", + "image_caption": [ + "Figure 6. Generated pseudo-texts by the proposed method during training. We depict the generated pseudo-texts in training procedure. We utilize them as supplementary supervisions to address linguistic diversity of a sentence at inference." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 851, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "within a unified architecture, as illustrated in Fig. 2. Here, we set the token length for each modality as $T_{text} = 80$ and $T_{image} = 256$ (i.e., $16 \\times 16$ ), the hidden dimension size to 512, the number of transformer layer to 6, and the number of attention heads to 16, thereafter the number of trainable parameters approximately 93.7-M as shown in Tab. 1. Particularly, we generate an image per each sentence without any sampling process. We train the CMOTA using AdamW [21] optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.95$ , $\\epsilon = 1e - 8$ , weight decay multiplier with $1e - 2$ and learning rate with $4.5e - 6$ multiplied by batch size. Moreover, we set the hyper-parameters for loss balancing in Eq. 3 as $\\lambda_{1} = 1.0$ and $\\lambda_{2} = 0.5$ and the number of memory state, $T_{M}$ as 1. Finally, to generate a high-resolution image, we modified the VQ-VAE with the additional trainable parameters (i.e., 7.1-M shown in Tab. 1). This modification allows for the generation of high-resolution images, increasing the image resolution from $64 \\times 64$ to $128 \\times 128$ .", + "bbox": [ + 75, + 335, + 472, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 622, + 279, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Comparison with the State of the Arts", + "text_level": 1, + "bbox": [ + 76, + 647, + 395, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 1, we summarize the performance of the prior arts and the proposed CMOTA in Pororo-SV [17] and Flintstones-SV [9] dataset in the various metrics. By default, all methods generate $64 \\times 64$ images along with the higher resolution $128 \\times 128$ images. CMOTA outperforms existing methods by a large margin in both benchmarks in most of metrics.", + "bbox": [ + 75, + 672, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The very recent method VP-CSV [2] performs on par to our CMOTA while ours still outperforms in FID, BLEU and R-precision, implying that CMOTA generates high quality images maintaining global semantic matching between story paragraph and images better than the VP-CSV. The lower performance than the VP-CSV in Char. F1 score and Frm. Acc. is attributed to the specialized character-centric module that only focuses the model to generate accurate characters", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for each image at the expense of the other performance metrics. In addition, our CMOTA-HR model even outperforms the VP-CSV [2] by large margin without the specialized character-centric module.", + "bbox": [ + 496, + 335, + 893, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Low resolution to high resolution. Not surprisingly, CMOTA-HR outperforms CMOTA with default (low-) resolution (i.e., $64 \\times 64$ ) in almost all metrics. Interestingly, however, VLC-StoryGAN-HR [22] shows similar performance to $64 \\times 64$ model in Frame accuracy, BLEU, R-precision, and drops performance in all other metrics. For the increased FID scores by the VLC-StoryGAN when goes to high resolution, we hypothesize that the constituency parse trees of input sentences in the VLC-StoryGAN increases the difficulty in generating visual details as the images contain more visual details in high resolution.", + "bbox": [ + 496, + 414, + 893, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we observe that the performance drop of FID in both our CMOTA and VLC, when goes to high resolution. It is because as high-resolution image requires the model to depict fine-grained details, generation task would be more difficult compared to the low-resolution image generation. In particular, the drop of FID in Flintstones-SV is larger than that in Pororo-SV. It is because there are more complicated visual details in the Flintstones-SV than the Pororo-SV dataset. Note that the FID drops by our method from low-resolution to high resolution is less than the VLC in Pororo-SV. Furthermore, in terms of computational cost, our CMOTA gains performance boost by using high-resolution effectively with just adding relatively small computational cost on top of original $64 \\times 64$ model.", + "bbox": [ + 496, + 580, + 893, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Character consistency. Although our method does not explicitly enforce the character semantic preservation, the proposed memory delivers the semantic across the sentence implicitly. We here investigate the character context preservation performance by comparing our method to [22, 23] on per-character classification F1-score on test split of Pororo", + "bbox": [ + 496, + 809, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3130", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d70b33be9327ff9e0c7bbf0048cec972c3d50cf94273b94683246068411cd1e6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetResolutionMethods# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
Pororo-SV64 × 64StoryGAN [17]-158.0618.599.343.24 / 1.221.51 ± 0.15
CP-CSV [39]-140.2421.7810.033.25 / 1.221.76 ± 0.04
DuCo-StoryGAN [23]101M96.5138.0113.973.68 / 1.343.56 ± 0.04
VLC-StoryGAN [22]100M84.9643.0217.363.80 / 1.443.28 ± 0.00
VP-CSV [2]-65.5156.8425.874.45 / 1.806.95 ± 0.00
CMOTA (Ours)96.6M52.1353.2524.724.58 / 1.907.34 ± 0.03
128 × 128VLC-StoryGAN-HR†[22]102.6M97.0840.3617.173.89 / 1.583.47 ± 0.03
CMOTA-HR (Ours)103.7M52.7758.8628.895.45 / 2.3416.36 ± 0.05
Flinstones-SV64 × 64StoryGAN [17]-127.1946.2032.9613.87 / 7.831.72 ± 0.18
DuCo-StoryGAN [23]101M78.0254.9236.3415.48 / 9.172.64 ± 0.17
VLC-StoryGAN‡[22]100M72.8758.8139.18--
CMOTA (Ours)96.6M36.7179.7466.0119.85 / 12.9810.50 ± 0.35
128 × 128CMOTA-HR (Ours)103.7M54.8186.4474.0622.18 / 15.1725.71 ± 0.70
", + "bbox": [ + 99, + 88, + 870, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison with the state of the arts. On the test split of Pororo-SV and Flintstones-SV. # Params. refers to the number of trainable parameters. Char. F1 refers to character F1 score. Frm. Acc. refers to frame accuracy. R-Prec. refers to R-Precision. ↓ indicates 'lower the better' and ↑ indicates 'higher the better'. Experiments are done in both $64 \\times 64$ and $128 \\times 128$ resolutions. 'HR' refers to its high resolution $(128 \\times 128)$ version. † indicates our results with author's implementation (https://github.com/adymaharana/VLCStoryGan).‡ indicates the absolute values that we computed by the given relative values in [22].", + "bbox": [ + 75, + 303, + 893, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e8a7fb8bd7ea9252edc4ce8f6894437c847fffec2e2857fa70e8a47d1aa3b63.jpg", + "image_caption": [ + "Figure 7. Per-character Classification F1-score on the Test Split of Pororo-SV Dataset. Normalized character (Char.) frequency is shown in red line. It represents the relative number of characters compared to the count of Pororo in training dataset. While DuCo/VLC-StoryGAN shows degraded performance as the character frequency decreases, CMOTA outperforms them across all the characters." + ], + "image_footnote": [], + "bbox": [ + 78, + 390, + 468, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SV dataset and normalized character frequency in train split dataset as illustrated in Fig. 7. Especially, it is interesting to compare with [2] as it is explicitly proposed for that purpose.", + "bbox": [ + 75, + 655, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in the figure, our method outperforms prior arts [22, 23] by large margins. Although we did not have any special module for character information preserving, unlike [2], our model maintains such information over multiple sentences, even outperforming the image quality denoted as FID by $-13.38$ .", + "bbox": [ + 75, + 700, + 470, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 815, + 246, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benefit of the proposed architectural components. To investigate the benefit of each proposed architectural component, we build up the full model with the components from the vanilla, single-directional transformer without memory", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as summarized in Tab. 2.", + "bbox": [ + 498, + 392, + 666, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using the proposed partial-level connected memory architecture only with the vanilla transformer (Tr.), the generation performance improves in all metrics compared to the Tr. Also, if we compare it with all-level connected memory architecture [15], PMA shows better performance in all metrics as in Tab. 3 (see supplementary material for more experimental analysis for various memory connection schemes). Employing attentional weighted memory (AWM) improves performance overall with slight degradation in FID (+0.12). The bi-directional learning scheme (Bi.) improves performance in all metrics except the character classification related metrics, i.e., Char.F1 and Frm. Acc. It is because there are captions that does not contain characters' name, just mentioning characters as 'friends', making the model difficult to generate specific characters.", + "bbox": [ + 496, + 407, + 895, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benefit of the text augmentation. By augmenting the generated text in offline manner (i.e., using an image captioning model pretrained on the dataset of interest), we observe overall performance increase in all metrics. We hypothesize its reason as that the pseudo-text may convey contextual information such as characters' existence or description of background, thereby making the generation model to be robust to language variation as shown in a first row of Tab. 4. By the online augmentation of pseudo texts (i.e., gradually updating text generator, thus generating a diverse and gradually varying set of pseudo-texts), the performance further improves, as shown in Tab. 2 and Tab. 4.", + "bbox": [ + 496, + 646, + 895, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3 Comparison with Large-Scale Models", + "text_level": 1, + "bbox": [ + 498, + 845, + 818, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Despite the unfairness of comparison to the large-scale models due to the model size, we compare CMOTA to larger", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3131", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/db2291dadd07fd827f99f5dd46b03f11167a0f062e87c60e0a0ba3141f9384dc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
VLC-StoryGAN [22]100M84.9643.0217.363.28 / 1.443.28
Single Directional Transformer (Tr.)93.7M63.8845.4818.444.18 / 1.695.67
+ Partial-level Memory Connection (PMA)95.8M59.0549.7221.794.41 / 1.776.28
+ Attentional Weighted Memory (AWM)96.6M56.9850.1522.014.50 / 1.916.79
+ Bi-directional Training (Bi.)96.6M54.6951.2722.154.52 / 1.847.12
+ Online Augmentation (Full model)96.6M52.1353.2524.724.58 / 1.907.34
", + "bbox": [ + 119, + 88, + 846, + 204 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7ccea10ab79183a8f57bc2047a8707d02a0b417812a509d14e887a2e04e12a88.jpg", + "table_caption": [ + "Table 2. Benefit of the proposed components. On the test split of Pororo-SV Dataset. As a baseline, we use a single directional transformer-based text-to-image generation model. We gradually add each of proposed components to investigate its effect for improving performance, i.e., partial-level memory connection (Sec. 3.1), attentively weighted memory (Sec. 3.1), bi-directional training (Sec 3.2), and online-augmentation (Sec. 3.2)." + ], + "table_footnote": [], + "table_body": "
Memory Architecture# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
All-Level Connection [15]118M61.2347.2119.214.21 / 1.726.08
Partial-level Connection (Ours)95.8M59.0549.7221.794.41 / 1.776.28
", + "bbox": [ + 78, + 280, + 467, + 316 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/707442da1c3ea159cf1074d80e9bcef2d2e7f93220a5f70c3ad02f227fca328a.jpg", + "table_caption": [ + "Table 3. Benefit of the proposed memory connection scheme. Proposed scheme outperforms conventional memory connection that uses all-level connections [15] (Fig. 4-(a)) with less number of parameters (test split of Pororo-SV)." + ], + "table_footnote": [], + "table_body": "
Augmentation# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
Offline96.6M54.5151.3222.314.50 / 1.907.09
Online (Ours)96.6M52.1353.2524.724.58 / 1.907.34
", + "bbox": [ + 78, + 392, + 467, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models [24] in Tab. 5. it is surprising to see that CMOTA outperforms on an evaluation metric of character classification (i.e., Char.F1 and Frm. Acc.), regardless of using prompt-tuning or fine-tuning of StoryDALL-E.", + "bbox": [ + 75, + 506, + 468, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "But as expected, the image quality denoted as FID is worse than the StoryDALL-E thanks to the huge size of the model and training dataset. Nevertheless, FID score is slightly better (Pororo-SV) and on-par (Flintstones-SV) compared to 'prompt-tune' that updates $30\\%$ of parameters [24]. We discuss more about this in the supplementary material.", + "bbox": [ + 75, + 569, + 470, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Human Preference Study", + "text_level": 1, + "bbox": [ + 76, + 676, + 307, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct a larger scale (than the prior arts) human survey using the Amazon Mechanical Turk platform to qualitatively compare the generated image sequence by our method over VLC-StoryGAN [22] on three criteria, i.e., visual quality, temporal consistency, and semantic relevancy between generated images and descriptions, following [17, 22, 23, 39]. Unlike the previous arts that employ 2 to 9 human judges [2, 17, 22, 23, 39], we employ 100 judges for statistically more reliable results.", + "bbox": [ + 75, + 702, + 468, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in the top table in Tab. 6, it is clear that human subjects prefer the generated images by our model over those from VLC-StoryGAN [22] on winning ratio(\\%). We discuss more details about this study in the supplementary material.", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a5dc334913ec539608d0fb890250f80b34043a7cb60e8a282df43eddd0307997.jpg", + "table_caption": [ + "Table 4. Benefit of the online augmentation. On the Pororo-SV test split. The proposed online augmentation outperforms the offline augmentation using a pretrained captioner [15]." + ], + "table_footnote": [], + "table_body": "
Methods# Params.FID↓Char. F1↑Frm. Acc↑
Porre-SVStoryDALL-E (prompt-tune)1.3B61.2329.6811.65
StoryDALL-E (fine-tune)1.3B25.9036.9717.26
MEGA-StoryDALL-E (fine-tune)2.8B23.4839.9118.01
CMOTA (Ours)96.6M55.2651.4822.73
Flinstones-SVStoryDALL-E (prompt-tune)1.3B53.7142.3832.54
StoryDALL-E (fine-tune)1.3B26.4973.4355.19
MEGA-StoryDALL-E (fine-tune)2.8B23.5874.2654.68
CMOTA (Ours)96.6M58.5979.7562.98
", + "bbox": [ + 504, + 280, + 888, + 393 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c54a5ca024a0a5c644ae22248274be7ec0410ee936b004ed3a833a402ae2ab9f.jpg", + "table_caption": [ + "Table 5. Quantitative comparisons to large scale models. On the test split of Pororo-SV and Flintstones-SV dataset. All models are implemented on top of the large pretrained transformer (i.e., StoryDALL-E [24] pretrained on 14 million and MEGAStoryDALL-E [24] pretrained on 15 million from Conceptual Caption dataset [36]). In [24], the 'prompt-tune' update $30\\%$ of model parameters compared to full fine-tune'. Without pretraining, we use a much small model compared to the proir arts, i.e., 96.6M (CMOTA) vs. 1.3B (StoryDALL-E)." + ], + "table_footnote": [], + "table_body": "
ResolutionAttributeVLC-SG [22]TieOurs
64 × 64 (Low-Res.)Visual Quality27.8%8.6%63.6%
Temproal Consistency24.3%8.3%59.0%
Semantic Relevance31.2%10.8%57.9%
128 × 128 (High-Res.)Visual Quality21.9%1.5%76.6%
Temporal Consistency21.8%2.5%75.7%
Semantic Relevance23.6%1.7%74.6%
", + "bbox": [ + 503, + 540, + 888, + 642 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0e463c1abe7075372df2f4de2590212a883f82bfa7ab0570fa2a968d4c1856bd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CMOTA w/o Mem.TieCMOTA
Temporal Consistency34.5%4.3%61.2%
", + "bbox": [ + 503, + 651, + 888, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Human preference studies. (Top) With 100 judges in the Amazon Mechanical Turk, on Pororo-SV test split dataset. Win $(\\%)$ refers to the $\\%$ times one model is preferred over the others. \"Tie\" refers to the same. Ours are clearly preferred over the VLC-SG (refers to VLC-StoryGAN) [22]. (Bottom) For the model using our memory (Mem.) module, contrasting results with and without the memory module that considers the contexts spread across the sentences, our method significantly improves human preference in temporal consistency.", + "bbox": [ + 496, + 700, + 890, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human preference for memory usage. We further investigate the benefit of using the proposed memory module (Sec. 3.1) by the human study with 100 annotators in Amazon Mechanical Turk. For the comparative analysis between", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3132", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CMOTA and CMOTA-w/o-Memory, we use 30 randomly sampled image sequences on Pororo-SV test set to conduct A/B test following [39]. Particularly, the annotators need to consider the temporal consistency with semantic relevance between generated image sequence and captions. As depicted in the bottom table in Tab. 6, we observe that $61.2\\%$ of annotators prefer our CMOTA with memory module over the one without it. We can conclude that the memory module has a great deal in creating temporally consistent image sequence with respect to semantic relevancy.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 253, + 269, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We now qualitatively analyze the quality of generated image sequences in Fig. 8 on the Pororo-SV dataset's test split. The top row shows the image sequence of ground truth, the two rows (2-3) contain prior works [17,23] and the final row is the image sequence generated by CMOTA and its high resolution version (CMOTA-HR). CMOTA generates a semantically more plausible image sequence with better visual quality, compared to prior works [22, 23]. Particularly, the CMOTA-HR demonstrates the ability to generate visual details more effectively even as the size of the image increases, while maintaining similar semantics.", + "bbox": [ + 75, + 276, + 470, + 443 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Furthermore, we qualitatively investigate the advantage of using memory architecture and summarize the results in Fig. 9. As shown in the figure, the proposed memory architecture generates a semantically more plausible image sequence with proper context, e.g., background; the CMOTA without memory fails to capture proper background context (shown in dotted red box) since a single sentence could be interpreted in many ways. In contrast, CMOTA (with the context memory) generates an image sequence with plausible background with characters preserved without an explicit character-centric model.", + "bbox": [ + 75, + 443, + 473, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 625, + 194, + 640 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We propose to better encode semantic context, e.g., plausible background and characters, for story visualization using a new memory architecture in a multi-modal bi-directional Transformer. We further propose an online text augmentation training scheme to generate pseudo-text descriptions as an intermediate supervision for addressing linguistic diversity in the texts at inference. The proposed method generates a temporally coherent and semantically relevant image sequence for each sentence in the given text paragraph.", + "bbox": [ + 75, + 650, + 470, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The proposed method outperforms prior works by a large margin on various metrics on the two popular SV benchmark datasets, and also outperforms some of hyper-scale models in multiple semantic understanding metrics. Although computationally prohibited and the pre-trained models are not publicly available, it would be intriguing to apply the proposed memory module and the online augmentation scheme to a large model.", + "bbox": [ + 75, + 787, + 470, + 907 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/707bec45ef572e62b2f3e99daa00148fa57ea05e04ad390e3f96aebc6e8b8ff2.jpg", + "image_caption": [ + "Figure 8. Qualitative results compared to prior arts. On the Pororo-SV's test split. We compare our CMOTA to the prior arts including DuCo-StoryGAN and VLC-StoryGAN. Ours generates a semantically more plausible and temporally coherent image sequence compared to the prior arts. All images except CMOTA-HR $(128\\times 128)$ are generated with resolution of $64\\times 64$ for comparison." + ], + "image_footnote": [], + "bbox": [ + 511, + 90, + 880, + 414 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/37b9e587e9dd43b4f188deca6e4b7bfb988c0d7b4421954911bb2ffc4745a3c3.jpg", + "image_caption": [ + "Figure 9. Benefit of memory module in encoding background context. Our CMOTA generates both semantically and visually plausible image sequences. Without memory module, inconsistent images (i.e., abrupt background change) are generated as indicated by the red dotted box." + ], + "image_footnote": [], + "bbox": [ + 511, + 521, + 883, + 743 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This work is partly supported by the NRF grant (No.2022R1A2C4002300) $25\\%$ , IITP grants (No.2020-0-01361, AI GS Program (Yonsei University) $5 \\%$ , No.2021-0-02068, AI Innovation Hub $5 \\%$ 2022-0-00077 $15\\%$ , 2022-0-00113 $15\\%$ , 2022-0-00959 $15\\%$ , 2022-0-00871 $10\\%$ , 2022-0-00951 $10\\%$ ) funded by the Korea government (MSIT).", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3133", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexander Buslaev, Vladimir I. Iglovikov, Eugene Khvedchenya, Alex Parinov, Mikhail Druzhinin, and Alexandr A. Kalinin. Albumimations: Fast and flexible image augmentations. Information, 11(2), 2020. 3", + "[2] Hong Chen, Rujun Han, Te-Lin Wu, Hideki Nakayama, and Nanyun Peng. Character-centric story visualization via visual planning and token alignment. ArXiv, 2022. 2, 5, 6, 7, 8", + "[3] Mark Chen, Alec Radford, Jeff Wu, Heewoo Jun, Prafulla Dhariwal, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 2, 3", + "[4] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov. Transformer-XL: Attentive language models beyond a fixed-length context. In ACL, 2019. 3, 4", + "[5] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. In Advances in Neural Information Processing Systems, 2021. 2, 3", + "[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. ArXiv, 2022. 2, 3", + "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv, 2022. 2, 3", + "[8] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, 2022. 2", + "[9] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 2, 5, 6", + "[10] Danna Gurari, Qing Li, Chi Lin, Yinan Zhao, Anhong Guo, Abigale Stangl, and Jeffrey P. Bigham. Vizwiz-priv: A dataset for recognizing the presence and purpose of private visual information in images taken by blind people. In CVPR, June 2019. 5", + "[11] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. ArXiv, 2022. 2", + "[12] Philip T. G. Jackson, Amir Atapour-Abarghouei, Stephen Bonner, T. Breckon, and Boguslaw Obara. Style augmentation: Data augmentation via style randomization. In CVPR Workshops, 2019. 3", + "[13] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. IEEE Access, 2020. 2", + "[14] Taehoon Kim, Gwangmo Song, Sihaeng Lee, Sangyun Kim, Yewon Seo, Soonyoung Lee, Seung Hwan Kim, Honglak Lee, and Kyunghoon Bae. L-verse: Bidirectional generation between image and text. In CVPR, 2022. 3", + "[15] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L Berg, and Mohit Bansal. Mart: Memory-augmented recurrent transformer for coherent video paragraph captioning. In ACL, 2020. 3, 4, 7, 8" + ], + "bbox": [ + 78, + 114, + 472, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5", + "[17] Yitong Li, Zhe Gan, Yelong Shen, Jingjing Liu, Yu Cheng, Yuexin Wu, Lawrence Carin, David Carlson, and Jianfeng Gao. Storygan: A sequential conditional gan for story visualization. In CVPR, 2019. 1, 2, 3, 5, 6, 7, 8, 9", + "[18] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 2", + "[19] Swee Kiat Lim, Yi Loo, Ngoc-Trung Tran, Ngai-Man Cheung, Gemma Roig, and Yuval Elovici. Doping: Generative data augmentation for unsupervised anomaly detection with gan. 2018 IEEE International Conference on Data Mining (ICDM), pages 1122–1127, 2018. 3", + "[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5", + "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6", + "[22] Adyasha Maharana and Mohit Bansal. Integrating visuospatial, linguistic and commonsense structure into story visualization. In EMNLP, 2021. 2, 5, 6, 7, 8, 9", + "[23] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Improving generation and evaluation of visual stories via semantic consistency. In NAACL-HLT, 2021. 2, 5, 6, 7, 8, 9", + "[24] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Storydall-e: Adapting pretrained text-to-image transformers for story continuation. ArXiv, 2022. 2, 5, 8", + "[25] Gaurav Mittal, Tanya Marwah, and Vineeth N. Balasubramanian. Sync-draw: Automatic video generation using deep recurrent attentive architectures. In ACM Multimedia, 2017. 2", + "[26] Saypraseuth Mounsaveng, Issam Laradji, Ismail Ben Ayed, David Vazquez, and Marco Pedersoli. Learning data augmentation with online bilevel optimization for image classification. arXiv, 2020. 3", + "[27] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 2", + "[28] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, 2019. 4", + "[29] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 4", + "[30] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Learn, imagine and create: Text-to-image generation from prior knowledge. In NeurIPS, 2019. 3", + "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv, 2022. 2, 3", + "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2, 3" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3134", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 3", + "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. 2022. 3", + "[35] Meet Shah, Xinlei Chen, Marcus Rohrbach, and Devi Parikh. Cycle-consistency for robust visual question answering. CVPR, 2019. 3", + "[36] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 8", + "[37] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, 2020. 5", + "[38] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, 2022. 2", + "[39] Yun-Zhu Song, Zhi Rui Tam, Hung-Jen Chen, Huiao-Han Lu, and Hong-Han Shuai. Character-preserving coherent story visualization. In ECCV, 2020. 2, 5, 7, 8, 9", + "[40] Zhiqiang Tang, Yunhe Gao, Leonid Karlinsky, Prasanna Sattigeri, Rogerio Feris, and Dimitris Metaxas. Onlineaugment: Online data augmentation with less domain knowledge. arXiv, 2020. 3", + "[41] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 4", + "[42] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NIPS, 2017. 3, 5", + "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2, 3, 4", + "[44] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, June 2015. 5", + "[45] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv, 2022. 2", + "[46] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions, 2021. 2", + "[47] Chenfei Wu, Jian Liang, Lei Ji, F. Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pre-training for neural visual world creation. In ECCV, 2022. 2", + "[48] Zhiyue Wu, Yijie Wang, Haibo Mi, Hongzuo Xu, Wei Zhang, and Lanlan Feng. Oada: An online data augmentation method for raw histopathology images. ICONIP, 2021. 3" + ], + "bbox": [ + 78, + 90, + 470, + 893 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3135", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Story Visualization by Online Text Augmentation with Context Memory/b1180507-958c-45ce-9ffe-139cc0187a6e_model.json b/2023/Story Visualization by Online Text Augmentation with Context Memory/b1180507-958c-45ce-9ffe-139cc0187a6e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0fa14190ffba143ef366f4bd8fb0d3a8b59d61b5 --- /dev/null +++ b/2023/Story Visualization by Online Text Augmentation with Context Memory/b1180507-958c-45ce-9ffe-139cc0187a6e_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.131, + 0.849, + 0.154 + ], + "angle": 0, + "content": "Story Visualization by Online Text Augmentation with Context Memory" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.18, + 0.792, + 0.22 + ], + "angle": 0, + "content": "Daechul Ahn\\(^{1,\\S}\\) Daneul Kim\\(^{2}\\) Gwangmo Song\\(^{3}\\) Seung Hwan Kim\\(^{3}\\) Honglak Lee\\(^{3,4}\\) Dongyeop Kang\\(^{5}\\) Jonghyun Choi\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.228, + 0.871, + 0.29 + ], + "angle": 0, + "content": "\\(^{1}\\)Yonsei University \\(^{2}\\)GIST \\(^{3}\\)LG AI Research \\(^{4}\\)University of Michigan \\(^{5}\\)University of Minnesota \\(\\{\\mathrm{dcahn, jc}\\} @\\)yonsei.ac.kr flytodk98@gm.gist.ac.kr gwangmo.song@lgresearch.ai skcruise@gmail.com honglak@eecs.umich.edu dongyeop@umn.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.324, + 0.314, + 0.34 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.474, + 0.645 + ], + "angle": 0, + "content": "Story visualization (SV) is a challenging text-to-image generation task for the difficulty of not only rendering visual details from the text descriptions but also encoding a long-term context across multiple sentences. While prior efforts mostly focus on generating a semantically relevant image for each sentence, encoding a context spread across the given paragraph to generate contextually convincing images (e.g., with a correct character or with a proper background of the scene) remains a challenge. To this end, we propose a novel memory architecture for the Bi-directional Transformer framework with an online text augmentation that generates multiple pseudo-descriptions as supplementary supervision during training for better generalization to the language variation at inference. In extensive experiments on the two popular SV benchmarks, i.e., the Pororo-SV and Flintstones-SV, the proposed method significantly outperforms the state of the arts in various metrics including FID, character F1, frame accuracy, BLEU-2/3, and R-precision with similar or less computational complexity." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.673, + 0.21, + 0.688 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.472, + 0.866 + ], + "angle": 0, + "content": "Story visualization (SV) [17] is a task of generating a sequence of images from a paragraph, i.e., a sequence of natural language sentences. It is challenging for the requirement of rendering the visual details in images with convincing background of a scene – seasonal elements, environmental objects such as table, location, and the proper character appearing, which here we refer to as context, spread across the given text sentences. Specifically, it needs to encode implicit context presented in the given sentences since each one often omit visual details (i.e., they may be spread over the sentences) necessary to generate a semantically correct im" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.324, + 0.889, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.461, + 0.888, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.596, + 0.891, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.717, + 0.895, + 0.828 + ], + "angle": 0, + "content": "Figure 1. Linguistic variations in story visualization and the overview of the proposed method. (a) An example of a story data with various linguistic variations (Var.#) for each image. (b) Modeling temporal context spread across sentences by our context memory (Sec. 3.1). (c) Addressing linguistic variations by online text augmentation from each image for every epoch (Sec. 3.2). (d) Benefit of the proposed context memory (left) and online text augmentation (right)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "age. For example, we can think of Fig. 1-(d), where Sent.#1 (i.e., sentence 1) and Sent.#2 (i.e., sentence 2) are given as sequences. After Sent.#1 is given, generated image by Sent.#2 often exhibits a background that is not semantically" + }, + { + "type": "page_footnote", + "bbox": [ + 0.075, + 0.875, + 0.472, + 0.888 + ], + "angle": 0, + "content": "§: work done while interning at LG AI Research. †: corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.075, + 0.888, + 0.385, + 0.9 + ], + "angle": 0, + "content": "Code: https://github.com/yonseivnl/cmota" + }, + { + "type": "list", + "bbox": [ + 0.075, + 0.875, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3125" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.137 + ], + "angle": 0, + "content": "correct [17,22,23]. However, if we use contextual information given in Sent.#1, \"The woods are covered with snow\", it leads to correct image with matching background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.47, + 0.228 + ], + "angle": 0, + "content": "In addition, widely used benchmark datasets for story visualization task [9, 17] provide a single text-image pair for training and inference, mostly due to the annotation budget constraint. This prevents the model from learning language variations, and thus harms the linguistic generalization performance of the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.471, + 0.379 + ], + "angle": 0, + "content": "To address the aforementioned challenges without requiring large scale data and models, we propose a new memory scheme in bi-directional Transformer for encoding the context which generates pseudo-texts in an online fashion to address linguistic variations at inference. We call our model as Context Memory and Online Text Augmentation or CMOTA for short. We empirically validate that our model outperforms the state-of-the-art SV methods by large margins on various metrics evaluated with widely used benchmarks in the literature, i.e., Pororo-SV and Flintstones-SV." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.471, + 0.516 + ], + "angle": 0, + "content": "Note that while large pre-trained models [3,5-7, 31, 43] have shown great success in synthesizing an image or a video from a language description [13, 32, 38], huge computational complexity and large training data makes the models prohibited. Moreover, although we propose and evaluate the model for the standard benchmark datasets without large pretraining data trained with a large model, it would be interesting to apply and evaluate the proposal in large models for further improvement." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.517, + 0.388, + 0.531 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.538, + 0.47, + 0.567 + ], + "angle": 0, + "content": "- We propose a new memory architecture for Transformer to selectively make use of contexts in a story paragraph." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.57, + 0.47, + 0.63 + ], + "angle": 0, + "content": "- For better generalization of linguistic variations in the given paragraph at inference, we generate pseudo-texts and augment them in an online fashion for richer linguistic supervision." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.633, + 0.47, + 0.693 + ], + "angle": 0, + "content": "- Our model significantly outperforms prior arts (even some hyper-scale models) by large margins in five evaluation metrics with similar or less computational complexity." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.538, + 0.47, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.709, + 0.219, + 0.724 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Story visualization. StoryGAN [17] is one of the recent methods that utilized a story-level discriminator to improve global consistency in generated images. CP-CSV [39] disentangles figure and background information to enhance character consistency. In order to improve global semantic matching between paragraph and generated image sequence, DuCo-StoryGAN [23] presents a pre-trained video captioner as an auxiliary loss along with other design improvements on top of StoryGAN. More recently, VLC-StoryGAN [22] utilizes constituency parse-trees and common sense knowledge to improve consistency and an object-level feedback loop to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.196 + ], + "angle": 0, + "content": "improve image quality. Another recent work VP-CSV [2] is a two-stage approach, i.e., 1) character generation and 2) background completion, using Transformer model to address this task. We discuss these in more detail in the supplementary material for the space sake. But the prior arts largely neglect to encode the story narrative, which is our primary contribution here." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.394 + ], + "angle": 0, + "content": "Very recently, Maharana et al. propose a new task setup of story continuation; using first image as a condition. They fine-tune the large model DALL-E [32] for the SV, which they call StoryDALL-E [24]. Ours differs from it in several aspects as follows. We have explicit memory connections between adjacent image generators using the context memory to globally encode the sentences. In contrast, [24] utilizes global story embedding as additional input for understanding context. Further, while it is huge in size (1.3B parameters), trained with 14 million text-image pair, ours are much larger (97M parameters) and outperforms it in multiple metrics by large margins; even in the image quality metric, FID. Please refer to Sec. 4.2 for empirical comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.481 + ], + "angle": 0, + "content": "Text-to-video generation. Similar to story visualization, text-to-video generation also generates multiple frames from a given text. Since the pioneering work of Sync-DRAW [25] in text-to-video generation, [9, 18, 27] utilize generative adversarial network for high-quality image generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.893, + 0.723 + ], + "angle": 0, + "content": "Recently, high-quality video generation models are proposed, including GODIVA [46] and NUWA [47]. Furthermore, recent studies generate high-resolution videos, making sequential frames in high-quality [11, 38]. However, most of state-of-the-art text-to-video model [11, 38] generates a video from a single sentence, mostly having consistent backgrounds. Very recently, there is a method proposed to generate a video from a long paragraph [45]. Although they generate the video in a long time horizon, they require both a mega-scale model trained with huge data and a detailed paragraph where each sentence is describing the scene that are close in time. In contrast, story visualization requires generating frames arbitrarily distant in time (i.e., so-called 'key-frames') corresponding to different sentences, requiring to generate an image sequence that have contextually convincing background." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Text-to-image generation. Text-to-image generation is a sub-problem of story visualization, with literature focusing on semantic relevance and resolution improvements. Recently, text-based image synthesis has been greatly improved with the help of a vast amount of training data with a hyper-scale model including DALL-E [32] and its successor DALL-E2, CogView [5] and Make-A-Scene [8] using a sketch input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although text-to-image generation models generates very high-quality images, it may lack encoding the context, metaphoric sentences spread acorss multiple sentences. In" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3126" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "addition, naively using state-of-the-art text-to-image generation models is computationally prohibited. For example, diffusion-based models [31, 34] have hyper-scale model size, (e.g., Imagen [34] parameter count of 2-B, DALL-E2 [31] parameter count of 3.5-B) making it non-trivial for applying it in a wide range of inference scenarios that may not have the sufficient computing resource. Here, we consider relatively light architectures as our base model for computational efficiency. More discussion are in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.242, + 0.473, + 0.409 + ], + "angle": 0, + "content": "Transformer using memory. To mitigate context fragmentation issue [4], i.e., losing long-term dependency over a data stream in context, there are efforts to encode long-term contexts in generating an image sequence. [4, 15] adopt a recurrent path into transformer architecture. Specifically, the modeling of new data segments is conditioned on historical hidden states produced in the previous time step and uses highly summarized memory states. Unlike the conventional memory architectures, we propose a novel memory that has a dense connection from the past with attentive weighting schemes for their better usage (Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.473, + 0.588 + ], + "angle": 0, + "content": "Online augmentation. While offline data augmentation that prepares data outside of learning process is prevalent in many literature [1, 12, 19], online-augmentation that depends on training is seldom explored especially in story visualization task [40]. [35, 40] propose an online augmentation for image classification and VQA task. [26] uses bi-level optimization in image classification. In medical domain, [48] investigate the online augmentation for personalized image based Histopathology diagnosis [48]. Note that we show the benefit of online augmentation over the offline augmentation in the SV for the first time in this literature." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.603, + 0.188, + 0.621 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.472, + 0.765 + ], + "angle": 0, + "content": "There are multiple challenges in story visualization, including (1) generating semantically natural images without artifacts from a description, (2) encoding the context (e.g., consistent background of a scene - seasonal elements, environmental objects such as table and location, and consistent characters) spread across the sentences in a given paragraph [17] and (3) addressing the linguistic variation at the inference time (i.e., a given text description may not be in different writing style)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.902 + ], + "angle": 0, + "content": "For photo-realistic image generation from a language description, we recently witness unprecedented improvements in the quality of generated images by the help of large scale model [3,5-7, 31, 43]. To leverage the large model's benefits for the story visualization task, we may use it to generate an image sequence by gradually concatenating sentences to visualize the story (i.e., use the first sentence to generate the first image, use a concatenated sentence of the first and the second to generate the second, and so on). Although this" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.087, + 0.695, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.087, + 0.897, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.266, + 0.895, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Bi-directional (between text and image modality) Transformer for text or image 'sequence' generation with the proposed memory. Our bi-directional 'base' model can simultaneously generate a sequence of both text and image from \\( t = 1 \\ldots 5 \\) with the proposed memory (Sec. 3.1). We use a segment embedding to indicate input modality and enable output of any modality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.892, + 0.42 + ], + "angle": 0, + "content": "may generate a single photo-realistic image, it is not able to capture the context sparsely spread across the sentences as story progresses (see supplement for more discussion)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.51 + ], + "angle": 0, + "content": "A more involved way of using the large model for SV is to fine-tune it for the downstream SV or using a video generation models. However, as the most high-performing models are huge in size and the codes and the pre-trained models are not publicly available, the computational cost and reproducibility reduces practicality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.895, + 0.662 + ], + "angle": 0, + "content": "For a given dataset without using large extra training data, we propose a new memory module to better encode past information with a Transformer by an attentively weighted densely connected architecture (Sec. 3.1). We further propose to generate pseudo-texts during the learning process to augment (online-augmentation) for better linguistic generalization without requiring large external data by learning the bi-directional Transformer [14, 30, 33] in both directions of generating images from texts and vice versa as illustrated in Fig. 2 (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.795 + ], + "angle": 0, + "content": "Base model. As shown in Fig. 2, we use bi-directional (i.e., multi-modal) transformer that iteratively generates images and texts in both ways. Similar to [32], the image tokens are sequentially predicted from the input text sequences by the Transformer. Then, the decoder of the VQ-VAE [42] translates the predicted image tokens into image sequence. The text tokens are also sequentially predicted from the input image token sequence by the same Transformer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Particularly, for the bi-directional multi-modal generation, i.e., generating simultaneously text and image from the unified architecture, we add two embeddings; a positional embedding for absolute position between tokens and a segment embedding for distinguishing source and target. Tokens of a text \\(\\left(\\{t_{1},\\ldots ,t_{m}\\}\\right)\\) and an image \\(\\left(\\{z_{1},\\ldots ,z_{n}\\}\\right)\\) \\((m,n:\\#\\) of tokens for text and image) are fed into the Trans" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3127" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.473, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.221, + 0.473, + 0.36 + ], + "angle": 0, + "content": "Figure 3. Proposed memory module for the bi-directional Transformer. Memory module updates current memory \\( M_{t} \\) by utilizing memory attention and GRU operation with propagated memory, \\( M_{t-1} \\), and previously calculated hidden state, \\( H_{t}^{l} \\). To produce image, we propose a novel memory attention (Attn.) mask denoted by blue squares to select only text tokens for less bias to previously generated image tokens. Within the memory attention mask, white-colored squares filter out the content, i.e., image (img) token, whereas the blue squares allow the text (txt) token to propagate through for memory interaction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.415 + ], + "angle": 0, + "content": "former to predict tokens in the other modality in multiple epochs with the following objective function written as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.426, + 0.422, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {j, t 2 i} = \\sum_ {k = 1} ^ {n} - \\ln p _ {j} (z _ {k} | t _ {1}, \\dots , t _ {m}, z _ {1}, \\dots , z _ {k - 1}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.468, + 0.469, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {j, i 2 t} = \\sum_ {k = 1} ^ {m} - \\ln p _ {j} \\left(t _ {k} \\mid z _ {1}, \\dots , z _ {n}, t _ {1}, \\dots , t _ {k - 1}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.51, + 0.285, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {j} = \\mathcal {L} _ {j, t 2 i} + \\lambda_ {1} \\mathcal {L} _ {j, i 2 t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.612 + ], + "angle": 0, + "content": "where \\( p_j(\\cdot) \\) is a likelihood of \\( j \\)-th generated tokens in one modality given the other modality, \\( L_{j,t2i} \\) is a loss corresponding to \\( j \\)-th text-to-image generation (i.e., negative log likelihood), \\( L_{j,i2t} \\) is vice versa. \\( \\lambda_1 \\) is a balancing hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.644 + ], + "angle": 0, + "content": "To train the model, we iteratively train the generation model in each modality multiple times per each epoch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.652, + 0.245, + 0.669 + ], + "angle": 0, + "content": "3.1. Context Memory" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.825 + ], + "angle": 0, + "content": "To encode the context and calculate the propagated memory, as depicted in Fig. 3, we first apply cross attention between the current hidden state, \\( H_{t}^{l} \\in \\mathcal{R}^{T_{c} \\times d} \\) calculated from \\( (l - 1) \\)-th transformer layer and memory state at time \\( (t - 1) \\), \\( M_{t - 1} \\in \\mathcal{R}^{T_{M} \\times d} \\) (\\( T_{c} \\): # tokens, \\( T_{M} \\): # memory states, \\( d \\): hidden state dimension). We then obtain \\( S_{t} = \\text{Attn}(M_{t - 1}, H_{t}, H_{t}) \\), where \\( \\text{Attn}(Q, K, V) \\coloneqq \\text{Softmax}\\left(\\frac{QK^{T}}{\\sqrt{d_{q}}}\\right)V \\) and \\( d_{q} \\) is a query dimension (\\( Q \\)), with the memory attention mask depicted in the blue box in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In particular, we apply a memory mask in the attention operation to select text tokens as memory content (depicted as blue-shaded grid cells in the right of Fig. 3) because including image content as memory could be strong constraint for text-to-image generation (see empirical studies of the" + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.089, + 0.882, + 0.113 + ], + "angle": 0, + "content": "(a) Conventional memory connection (Dai et al., ACL'19/Lei et al., ACL'20) Layer #" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.104, + 0.887, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.239, + 0.697, + 0.252 + ], + "angle": 0, + "content": "(b) Proposed memory connection" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.252, + 0.886, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.389, + 0.894, + 0.5 + ], + "angle": 0, + "content": "Figure 4. Comparison of our memory connection scheme to the conventional one [4,15]. We depict the memory connections in the multi-layer Transformer architecture per each sentence input (unfolded in a row). Unlike (a) the previous memory architecture having a single memory path from the immediate past in all layers, (b) ours has a dense connection from all the pasts to be attentively weighted (denoted here by \\(\\bar{M}\\)'s) in only the last layer. We discuss its empirical benefit in Sec. 4.2.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.893, + 0.587 + ], + "angle": 0, + "content": "proposed mask in the supplementary material). Then we feed the \\( S_{t} \\) and \\( M_{t-1} \\) to the GRU in order to compute the information to be stored in the memory, \\( M_{t} \\), propagated to \\( (t+1) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.895, + 0.77 + ], + "angle": 0, + "content": "Memory connection. Following [4, 15], we stack multiple Transformer's layers, depicted in Fig. 4-(b). Compared to conventional memory updating architectures with serial hidden state connections in all layers (Fig. 4-(a)), our approach only has memory connections in the last layer, which we call it as partial-level memory augmented (PMA). This is because the later layers (closer to the last layer) achieve better representations with higher-level features, abstract and structured representations [28, 29, 41] with improved computational efficiency, which is unlike prior works (see Sec. 4.2.2 and supp for further discussion)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.771, + 0.895, + 0.832 + ], + "angle": 0, + "content": "Additionally, not all historic information is equally important for generating an image at a time step. Similar to the masked self-attention [43], we attentively weight the past information for better modeling of sparse context as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.844, + 0.848, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\bar {M} _ {1: (t - 1)} = A t t n (M _ {(t - 1)}, M _ {[ 1: (t - 2) ]}, M _ {[ 1: (t - 2) ]}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.866, + 0.891, + 0.883 + ], + "angle": 0, + "content": "\\[\n\\tilde {M} _ {(t - 1)} = \\left[ M _ {t - 1}; \\bar {M} _ {1: (t - 1)} \\right], \\quad (3 \\leq t \\leq 5), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.886, + 0.856, + 0.905 + ], + "angle": 0, + "content": "\\[\nH _ {t} ^ {l} = \\operatorname {A t t n} \\left(H _ {t} ^ {l}, \\left[ H _ {t} ^ {l}; \\tilde {M} _ {(t - 1)} \\right], \\left[ H _ {t} ^ {l}; \\tilde {M} _ {(t - 1)} \\right]\\right),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3128" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.184 + ], + "angle": 0, + "content": "where \\( M_{(t - 1)} \\) is a memory at time \\( (t - 1) \\) and \\( [1:(t - 2)] \\) refers to the concatenation from time 1 to \\( (t - 2) \\) as depicted in Fig. 4-(b). By doing so, we fuse the contextual information into current hidden state \\( H_{t} \\). At \\( t = 2 \\), we use \\( M_{1} \\), instead of \\( \\tilde{M}_{1} \\) as the \\( M_{1} \\) is only available. We call it as attentively weighted memory (AWM)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.192, + 0.32, + 0.208 + ], + "angle": 0, + "content": "3.2. Online Text-Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.471, + 0.321 + ], + "angle": 0, + "content": "Vedamtam et al. argue that multiple descriptions for an image help generalization as they address language variations presented in descriptions [44]. Hence, a number of image captioning datasets [10, 20, 37] provide multiple natural language directives, obtained by multiple human annotators, in both training and evaluation splits. But the SV benchmark datasets [17] provide only a single sentence per an image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.471, + 0.457 + ], + "angle": 0, + "content": "To address the linguistic variations of a text input at inference process, we first consider to generate a pseudo text by a well-trained image-to-text generation model, which we refer it as offline-augmentation [16]. But the offline augmentation generates only a single sentence, which may not provide sufficient diversity. Instead, we propose to generate multiple pseudo-texts and augment them in an online fashion when training our model to increase the diversity. We call it online text augmentation, depicted in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.624 + ], + "angle": 0, + "content": "Thanks to our bi-directional multi-modal architecture, we can naturally integrate the process of generating pseudotexts to the process of learning image-to-text model and the text-to-image generation model as depicted in Fig. 5. In the early epochs, less meaningful sentences are generated, but as training progresses, more meaningful sentences are generated (see orange box in Fig. 5). As a side-product, by supervising the model learning with intermediate goals at each time step, we expect to expedite the convergence of learning. When we use the online text augmentation, we can rewrite the objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.634, + 0.47, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {j, p t 2 i} = \\sum_ {k = 1} ^ {n} - \\ln p _ {j} \\left(z _ {k} \\mid \\hat {t} _ {1}, \\dots , \\hat {t} _ {m}, z _ {1}, \\dots , z _ {k - 1}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.677, + 0.37, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {j} = \\mathcal {L} _ {j, t 2 i} + \\lambda_ {1} \\mathcal {L} _ {j, i 2 t} + \\lambda_ {2} \\mathcal {L} _ {j, p t 2 i},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{j,pt2i}\\) is the additional loss with the augmented pseudo-texts and \\(\\mathcal{L}_{j,t2i},\\mathcal{L}_{j,i2t}\\) are defined in Eq. 1 and \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are balancing hyper-parameters. \\(\\hat{t}\\) means the pseudotext token, predicted during online augmentation without gradient flow. Fig. 6 shows pseudo-texts generated by our method. Detailed training procedure and more examples are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.823, + 0.21, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.268, + 0.865 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Datasets. We use two popular benchmark datasets for evaluating the task of story visualization. Following [23], we use" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.891, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.22, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 5. Online text augmentation. We iteratively augment the pseudo-text for better linguistic generalization. (1) generating a pseudo-text at each epoch for each ground-truth image, without gradient flow. (2) generating an image for every epoch. (3,4) learn the text-to-image model with \\( L_{CE} \\). In early epochs, it generates less meaningful pseudo-text (thus discard them), but produces pseudotexts matching with the input image as the training progresses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.894, + 0.392 + ], + "angle": 0, + "content": "Pororo-SV dataset for story visualization task without data overlap. Following prior work [22], we use Flintstones-SV which was originally exploited in the text-to-video synthesis task [9]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.895, + 0.527 + ], + "angle": 0, + "content": "Metrics. For the evaluation metrics, we use Fréchet Inception Distance (FID) (used in [22, 39]) for visual quality of the generated images, character classification score (Char. F1, Frm. Acc.) [17] for character consistency, and global semantic matching (B-2/3, R-Prec.) [22, 23] between generated image sequence and story paragraph.. We elaborate on the details of the datasets and implementation in the supplementary material for the sake of space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.895, + 0.646 + ], + "angle": 0, + "content": "Baselines. We compare our method to a number of prior arts including state-of-the-art story visualization methods such as StoryGAN [17], CPCSV [39], DuCo-StoryGAN [23], VLC-StoryGAN [22] and VP-CSV [2] for both quantitative and qualitative analysis in the story visualization setting. Moreover, we compare [24] in the story continuation setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Implementation details. To train our model, we first tokenize the text and image inputs. Particularly, the image inputs are encoded by the encoder of the VQ-VAE [42] into image tokens and decoded by the decoder of the VQ-VAE for generating image sequence. Once the inputs are tokenized, we append two special embedding tokens to the text and image tokens. These tokens signify the beginning of each modality, denoted as 'SOS' (start of sentence) and 'SOI' (start of image). Subsequently, we add two embeddings: (1) a positional embedding to indicate absolute position of each token and (2) a segment embedding to distinguish source and target modality, e.g., for text-to-image generation task, the source corresponds to the text while the target corresponds to the image. Through the utilization of segment embeddings to distinguish between source and target, we enable the model to generate target data from source data" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3129" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.088, + 0.852, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.297, + 0.894, + 0.325 + ], + "angle": 0, + "content": "Figure 6. Generated pseudo-texts by the proposed method during training. We depict the generated pseudo-texts in training procedure. We utilize them as supplementary supervisions to address linguistic diversity of a sentence at inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.61 + ], + "angle": 0, + "content": "within a unified architecture, as illustrated in Fig. 2. Here, we set the token length for each modality as \\( T_{text} = 80 \\) and \\( T_{image} = 256 \\) (i.e., \\( 16 \\times 16 \\)), the hidden dimension size to 512, the number of transformer layer to 6, and the number of attention heads to 16, thereafter the number of trainable parameters approximately 93.7-M as shown in Tab. 1. Particularly, we generate an image per each sentence without any sampling process. We train the CMOTA using AdamW [21] optimizer with \\( \\beta_{1} = 0.9 \\), \\( \\beta_{2} = 0.95 \\), \\( \\epsilon = 1e - 8 \\), weight decay multiplier with \\( 1e - 2 \\) and learning rate with \\( 4.5e - 6 \\) multiplied by batch size. Moreover, we set the hyper-parameters for loss balancing in Eq. 3 as \\( \\lambda_{1} = 1.0 \\) and \\( \\lambda_{2} = 0.5 \\) and the number of memory state, \\( T_{M} \\) as 1. Finally, to generate a high-resolution image, we modified the VQ-VAE with the additional trainable parameters (i.e., 7.1-M shown in Tab. 1). This modification allows for the generation of high-resolution images, increasing the image resolution from \\( 64 \\times 64 \\) to \\( 128 \\times 128 \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.623, + 0.281, + 0.639 + ], + "angle": 0, + "content": "4.2. Quantitative Analysis" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.396, + 0.663 + ], + "angle": 0, + "content": "4.2.1 Comparison with the State of the Arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.779 + ], + "angle": 0, + "content": "In Table 1, we summarize the performance of the prior arts and the proposed CMOTA in Pororo-SV [17] and Flintstones-SV [9] dataset in the various metrics. By default, all methods generate \\(64 \\times 64\\) images along with the higher resolution \\(128 \\times 128\\) images. CMOTA outperforms existing methods by a large margin in both benchmarks in most of metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The very recent method VP-CSV [2] performs on par to our CMOTA while ours still outperforms in FID, BLEU and R-precision, implying that CMOTA generates high quality images maintaining global semantic matching between story paragraph and images better than the VP-CSV. The lower performance than the VP-CSV in Char. F1 score and Frm. Acc. is attributed to the specialized character-centric module that only focuses the model to generate accurate characters" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.895, + 0.398 + ], + "angle": 0, + "content": "for each image at the expense of the other performance metrics. In addition, our CMOTA-HR model even outperforms the VP-CSV [2] by large margin without the specialized character-centric module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.581 + ], + "angle": 0, + "content": "Low resolution to high resolution. Not surprisingly, CMOTA-HR outperforms CMOTA with default (low-) resolution (i.e., \\(64 \\times 64\\)) in almost all metrics. Interestingly, however, VLC-StoryGAN-HR [22] shows similar performance to \\(64 \\times 64\\) model in Frame accuracy, BLEU, R-precision, and drops performance in all other metrics. For the increased FID scores by the VLC-StoryGAN when goes to high resolution, we hypothesize that the constituency parse trees of input sentences in the VLC-StoryGAN increases the difficulty in generating visual details as the images contain more visual details in high resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.895, + 0.794 + ], + "angle": 0, + "content": "In addition, we observe that the performance drop of FID in both our CMOTA and VLC, when goes to high resolution. It is because as high-resolution image requires the model to depict fine-grained details, generation task would be more difficult compared to the low-resolution image generation. In particular, the drop of FID in Flintstones-SV is larger than that in Pororo-SV. It is because there are more complicated visual details in the Flintstones-SV than the Pororo-SV dataset. Note that the FID drops by our method from low-resolution to high resolution is less than the VLC in Pororo-SV. Furthermore, in terms of computational cost, our CMOTA gains performance boost by using high-resolution effectively with just adding relatively small computational cost on top of original \\(64 \\times 64\\) model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Character consistency. Although our method does not explicitly enforce the character semantic preservation, the proposed memory delivers the semantic across the sentence implicitly. We here investigate the character context preservation performance by comparing our method to [22, 23] on per-character classification F1-score on test split of Pororo" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3130" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.872, + 0.3 + ], + "angle": 0, + "content": "
DatasetResolutionMethods# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
Pororo-SV64 × 64StoryGAN [17]-158.0618.599.343.24 / 1.221.51 ± 0.15
CP-CSV [39]-140.2421.7810.033.25 / 1.221.76 ± 0.04
DuCo-StoryGAN [23]101M96.5138.0113.973.68 / 1.343.56 ± 0.04
VLC-StoryGAN [22]100M84.9643.0217.363.80 / 1.443.28 ± 0.00
VP-CSV [2]-65.5156.8425.874.45 / 1.806.95 ± 0.00
CMOTA (Ours)96.6M52.1353.2524.724.58 / 1.907.34 ± 0.03
128 × 128VLC-StoryGAN-HR†[22]102.6M97.0840.3617.173.89 / 1.583.47 ± 0.03
CMOTA-HR (Ours)103.7M52.7758.8628.895.45 / 2.3416.36 ± 0.05
Flinstones-SV64 × 64StoryGAN [17]-127.1946.2032.9613.87 / 7.831.72 ± 0.18
DuCo-StoryGAN [23]101M78.0254.9236.3415.48 / 9.172.64 ± 0.17
VLC-StoryGAN‡[22]100M72.8758.8139.18--
CMOTA (Ours)96.6M36.7179.7466.0119.85 / 12.9810.50 ± 0.35
128 × 128CMOTA-HR (Ours)103.7M54.8186.4474.0622.18 / 15.1725.71 ± 0.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.304, + 0.895, + 0.377 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with the state of the arts. On the test split of Pororo-SV and Flintstones-SV. # Params. refers to the number of trainable parameters. Char. F1 refers to character F1 score. Frm. Acc. refers to frame accuracy. R-Prec. refers to R-Precision. ↓ indicates 'lower the better' and ↑ indicates 'higher the better'. Experiments are done in both \\(64 \\times 64\\) and \\(128 \\times 128\\) resolutions. 'HR' refers to its high resolution \\((128 \\times 128)\\) version. † indicates our results with author's implementation (https://github.com/adymaharana/VLCStoryGan).‡ indicates the absolute values that we computed by the given relative values in [22]." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.391, + 0.47, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.547, + 0.472, + 0.645 + ], + "angle": 0, + "content": "Figure 7. Per-character Classification F1-score on the Test Split of Pororo-SV Dataset. Normalized character (Char.) frequency is shown in red line. It represents the relative number of characters compared to the count of Pororo in training dataset. While DuCo/VLC-StoryGAN shows degraded performance as the character frequency decreases, CMOTA outperforms them across all the characters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.701 + ], + "angle": 0, + "content": "SV dataset and normalized character frequency in train split dataset as illustrated in Fig. 7. Especially, it is interesting to compare with [2] as it is explicitly proposed for that purpose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.471, + 0.793 + ], + "angle": 0, + "content": "As shown in the figure, our method outperforms prior arts [22, 23] by large margins. Although we did not have any special module for character information preserving, unlike [2], our model maintains such information over multiple sentences, even outperforming the image quality denoted as FID by \\(-13.38\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.248, + 0.83 + ], + "angle": 0, + "content": "4.2.2 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Benefit of the proposed architectural components. To investigate the benefit of each proposed architectural component, we build up the full model with the components from the vanilla, single-directional transformer without memory" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.393, + 0.667, + 0.406 + ], + "angle": 0, + "content": "as summarized in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.896, + 0.635 + ], + "angle": 0, + "content": "Using the proposed partial-level connected memory architecture only with the vanilla transformer (Tr.), the generation performance improves in all metrics compared to the Tr. Also, if we compare it with all-level connected memory architecture [15], PMA shows better performance in all metrics as in Tab. 3 (see supplementary material for more experimental analysis for various memory connection schemes). Employing attentional weighted memory (AWM) improves performance overall with slight degradation in FID (+0.12). The bi-directional learning scheme (Bi.) improves performance in all metrics except the character classification related metrics, i.e., Char.F1 and Frm. Acc. It is because there are captions that does not contain characters' name, just mentioning characters as 'friends', making the model difficult to generate specific characters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.897, + 0.829 + ], + "angle": 0, + "content": "Benefit of the text augmentation. By augmenting the generated text in offline manner (i.e., using an image captioning model pretrained on the dataset of interest), we observe overall performance increase in all metrics. We hypothesize its reason as that the pseudo-text may convey contextual information such as characters' existence or description of background, thereby making the generation model to be robust to language variation as shown in a first row of Tab. 4. By the online augmentation of pseudo texts (i.e., gradually updating text generator, thus generating a diverse and gradually varying set of pseudo-texts), the performance further improves, as shown in Tab. 2 and Tab. 4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.847, + 0.819, + 0.862 + ], + "angle": 0, + "content": "4.2.3 Comparison with Large-Scale Models" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Despite the unfairness of comparison to the large-scale models due to the model size, we compare CMOTA to larger" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3131" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.847, + 0.205 + ], + "angle": 0, + "content": "
Methods# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
VLC-StoryGAN [22]100M84.9643.0217.363.28 / 1.443.28
Single Directional Transformer (Tr.)93.7M63.8845.4818.444.18 / 1.695.67
+ Partial-level Memory Connection (PMA)95.8M59.0549.7221.794.41 / 1.776.28
+ Attentional Weighted Memory (AWM)96.6M56.9850.1522.014.50 / 1.916.79
+ Bi-directional Training (Bi.)96.6M54.6951.2722.154.52 / 1.847.12
+ Online Augmentation (Full model)96.6M52.1353.2524.724.58 / 1.907.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.21, + 0.895, + 0.266 + ], + "angle": 0, + "content": "Table 2. Benefit of the proposed components. On the test split of Pororo-SV Dataset. As a baseline, we use a single directional transformer-based text-to-image generation model. We gradually add each of proposed components to investigate its effect for improving performance, i.e., partial-level memory connection (Sec. 3.1), attentively weighted memory (Sec. 3.1), bi-directional training (Sec 3.2), and online-augmentation (Sec. 3.2)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.281, + 0.468, + 0.317 + ], + "angle": 0, + "content": "
Memory Architecture# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
All-Level Connection [15]118M61.2347.2119.214.21 / 1.726.08
Partial-level Connection (Ours)95.8M59.0549.7221.794.41 / 1.776.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.325, + 0.472, + 0.382 + ], + "angle": 0, + "content": "Table 3. Benefit of the proposed memory connection scheme. Proposed scheme outperforms conventional memory connection that uses all-level connections [15] (Fig. 4-(a)) with less number of parameters (test split of Pororo-SV)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.393, + 0.468, + 0.434 + ], + "angle": 0, + "content": "
Augmentation# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
Offline96.6M54.5151.3222.314.50 / 1.907.09
Online (Ours)96.6M52.1353.2524.724.58 / 1.907.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.443, + 0.472, + 0.486 + ], + "angle": 0, + "content": "Table 4. Benefit of the online augmentation. On the Pororo-SV test split. The proposed online augmentation outperforms the offline augmentation using a pretrained captioner [15]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.569 + ], + "angle": 0, + "content": "models [24] in Tab. 5. it is surprising to see that CMOTA outperforms on an evaluation metric of character classification (i.e., Char.F1 and Frm. Acc.), regardless of using prompt-tuning or fine-tuning of StoryDALL-E." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.472, + 0.661 + ], + "angle": 0, + "content": "But as expected, the image quality denoted as FID is worse than the StoryDALL-E thanks to the huge size of the model and training dataset. Nevertheless, FID score is slightly better (Pororo-SV) and on-par (Flintstones-SV) compared to 'prompt-tune' that updates \\(30\\%\\) of parameters [24]. We discuss more about this in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.677, + 0.308, + 0.693 + ], + "angle": 0, + "content": "4.3. Human Preference Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.838 + ], + "angle": 0, + "content": "We conduct a larger scale (than the prior arts) human survey using the Amazon Mechanical Turk platform to qualitatively compare the generated image sequence by our method over VLC-StoryGAN [22] on three criteria, i.e., visual quality, temporal consistency, and semantic relevancy between generated images and descriptions, following [17, 22, 23, 39]. Unlike the previous arts that employ 2 to 9 human judges [2, 17, 22, 23, 39], we employ 100 judges for statistically more reliable results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As shown in the top table in Tab. 6, it is clear that human subjects prefer the generated images by our model over those from VLC-StoryGAN [22] on winning ratio(\\%). We discuss more details about this study in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.281, + 0.89, + 0.394 + ], + "angle": 0, + "content": "
Methods# Params.FID↓Char. F1↑Frm. Acc↑
Porre-SVStoryDALL-E (prompt-tune)1.3B61.2329.6811.65
StoryDALL-E (fine-tune)1.3B25.9036.9717.26
MEGA-StoryDALL-E (fine-tune)2.8B23.4839.9118.01
CMOTA (Ours)96.6M55.2651.4822.73
Flinstones-SVStoryDALL-E (prompt-tune)1.3B53.7142.3832.54
StoryDALL-E (fine-tune)1.3B26.4973.4355.19
MEGA-StoryDALL-E (fine-tune)2.8B23.5874.2654.68
CMOTA (Ours)96.6M58.5979.7562.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.403, + 0.895, + 0.528 + ], + "angle": 0, + "content": "Table 5. Quantitative comparisons to large scale models. On the test split of Pororo-SV and Flintstones-SV dataset. All models are implemented on top of the large pretrained transformer (i.e., StoryDALL-E [24] pretrained on 14 million and MEGAStoryDALL-E [24] pretrained on 15 million from Conceptual Caption dataset [36]). In [24], the 'prompt-tune' update \\(30\\%\\) of model parameters compared to full fine-tune'. Without pretraining, we use a much small model compared to the proir arts, i.e., 96.6M (CMOTA) vs. 1.3B (StoryDALL-E)." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.541, + 0.89, + 0.643 + ], + "angle": 0, + "content": "
ResolutionAttributeVLC-SG [22]TieOurs
64 × 64 (Low-Res.)Visual Quality27.8%8.6%63.6%
Temproal Consistency24.3%8.3%59.0%
Semantic Relevance31.2%10.8%57.9%
128 × 128 (High-Res.)Visual Quality21.9%1.5%76.6%
Temporal Consistency21.8%2.5%75.7%
Semantic Relevance23.6%1.7%74.6%
" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.652, + 0.889, + 0.692 + ], + "angle": 0, + "content": "
CMOTA w/o Mem.TieCMOTA
Temporal Consistency34.5%4.3%61.2%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.828 + ], + "angle": 0, + "content": "Table 6. Human preference studies. (Top) With 100 judges in the Amazon Mechanical Turk, on Pororo-SV test split dataset. Win \\((\\%)\\) refers to the \\(\\%\\) times one model is preferred over the others. \"Tie\" refers to the same. Ours are clearly preferred over the VLC-SG (refers to VLC-StoryGAN) [22]. (Bottom) For the model using our memory (Mem.) module, contrasting results with and without the memory module that considers the contexts spread across the sentences, our method significantly improves human preference in temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Human preference for memory usage. We further investigate the benefit of using the proposed memory module (Sec. 3.1) by the human study with 100 annotators in Amazon Mechanical Turk. For the comparative analysis between" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3132" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "CMOTA and CMOTA-w/o-Memory, we use 30 randomly sampled image sequences on Pororo-SV test set to conduct A/B test following [39]. Particularly, the annotators need to consider the temporal consistency with semantic relevance between generated image sequence and captions. As depicted in the bottom table in Tab. 6, we observe that \\(61.2\\%\\) of annotators prefer our CMOTA with memory module over the one without it. We can conclude that the memory module has a great deal in creating temporally consistent image sequence with respect to semantic relevancy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.254, + 0.27, + 0.27 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.471, + 0.444 + ], + "angle": 0, + "content": "We now qualitatively analyze the quality of generated image sequences in Fig. 8 on the Pororo-SV dataset's test split. The top row shows the image sequence of ground truth, the two rows (2-3) contain prior works [17,23] and the final row is the image sequence generated by CMOTA and its high resolution version (CMOTA-HR). CMOTA generates a semantically more plausible image sequence with better visual quality, compared to prior works [22, 23]. Particularly, the CMOTA-HR demonstrates the ability to generate visual details more effectively even as the size of the image increases, while maintaining similar semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.474, + 0.61 + ], + "angle": 0, + "content": "Furthermore, we qualitatively investigate the advantage of using memory architecture and summarize the results in Fig. 9. As shown in the figure, the proposed memory architecture generates a semantically more plausible image sequence with proper context, e.g., background; the CMOTA without memory fails to capture proper background context (shown in dotted red box) since a single sentence could be interpreted in many ways. In contrast, CMOTA (with the context memory) generates an image sequence with plausible background with characters preserved without an explicit character-centric model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.625, + 0.196, + 0.641 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.472, + 0.788 + ], + "angle": 0, + "content": "We propose to better encode semantic context, e.g., plausible background and characters, for story visualization using a new memory architecture in a multi-modal bi-directional Transformer. We further propose an online text augmentation training scheme to generate pseudo-text descriptions as an intermediate supervision for addressing linguistic diversity in the texts at inference. The proposed method generates a temporally coherent and semantically relevant image sequence for each sentence in the given text paragraph." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.472, + 0.909 + ], + "angle": 0, + "content": "The proposed method outperforms prior works by a large margin on various metrics on the two popular SV benchmark datasets, and also outperforms some of hyper-scale models in multiple semantic understanding metrics. Although computationally prohibited and the pre-trained models are not publicly available, it would be intriguing to apply the proposed memory module and the online augmentation scheme to a large model." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.091, + 0.882, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.416, + 0.896, + 0.5 + ], + "angle": 0, + "content": "Figure 8. Qualitative results compared to prior arts. On the Pororo-SV's test split. We compare our CMOTA to the prior arts including DuCo-StoryGAN and VLC-StoryGAN. Ours generates a semantically more plausible and temporally coherent image sequence compared to the prior arts. All images except CMOTA-HR \\((128\\times 128)\\) are generated with resolution of \\(64\\times 64\\) for comparison." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.522, + 0.885, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.746, + 0.895, + 0.817 + ], + "angle": 0, + "content": "Figure 9. Benefit of memory module in encoding background context. Our CMOTA generates both semantically and visually plausible image sequences. Without memory module, inconsistent images (i.e., abrupt background change) are generated as indicated by the red dotted box." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Acknowledgement. This work is partly supported by the NRF grant (No.2022R1A2C4002300) \\(25\\%\\) , IITP grants (No.2020-0-01361, AI GS Program (Yonsei University) \\(5 \\%\\) , No.2021-0-02068, AI Innovation Hub \\(5 \\%\\) 2022-0-00077 \\(15\\%\\) , 2022-0-00113 \\(15\\%\\) , 2022-0-00959 \\(15\\%\\) , 2022-0-00871 \\(10\\%\\) , 2022-0-00951 \\(10\\%\\) ) funded by the Korea government (MSIT)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3133" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.17 + ], + "angle": 0, + "content": "[1] Alexander Buslaev, Vladimir I. Iglovikov, Eugene Khvedchenya, Alex Parinov, Mikhail Druzhinin, and Alexandr A. Kalinin. Albumimations: Fast and flexible image augmentations. Information, 11(2), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Hong Chen, Rujun Han, Te-Lin Wu, Hideki Nakayama, and Nanyun Peng. Character-centric story visualization via visual planning and token alignment. ArXiv, 2022. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[3] Mark Chen, Alec Radford, Jeff Wu, Heewoo Jun, Prafulla Dhariwal, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.472, + 0.31 + ], + "angle": 0, + "content": "[4] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov. Transformer-XL: Attentive language models beyond a fixed-length context. In ACL, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.472, + 0.381 + ], + "angle": 0, + "content": "[5] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. In Advances in Neural Information Processing Systems, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.472, + 0.423 + ], + "angle": 0, + "content": "[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. ArXiv, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.472, + 0.478 + ], + "angle": 0, + "content": "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.472, + 0.534 + ], + "angle": 0, + "content": "[8] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.472, + 0.577 + ], + "angle": 0, + "content": "[9] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.578, + 0.472, + 0.647 + ], + "angle": 0, + "content": "[10] Danna Gurari, Qing Li, Chi Lin, Yinan Zhao, Anhong Guo, Abigale Stangl, and Jeffrey P. Bigham. Vizwiz-priv: A dataset for recognizing the presence and purpose of private visual information in images taken by blind people. In CVPR, June 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.649, + 0.472, + 0.69 + ], + "angle": 0, + "content": "[11] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. ArXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.691, + 0.472, + 0.746 + ], + "angle": 0, + "content": "[12] Philip T. G. Jackson, Amir Atapour-Abarghouei, Stephen Bonner, T. Breckon, and Boguslaw Obara. Style augmentation: Data augmentation via style randomization. In CVPR Workshops, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.472, + 0.788 + ], + "angle": 0, + "content": "[13] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. IEEE Access, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.472, + 0.845 + ], + "angle": 0, + "content": "[14] Taehoon Kim, Gwangmo Song, Sihaeng Lee, Sangyun Kim, Yewon Seo, Soonyoung Lee, Seung Hwan Kim, Honglak Lee, and Kyunghoon Bae. L-verse: Bidirectional generation between image and text. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[15] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L Berg, and Mohit Bansal. Mart: Memory-augmented recurrent transformer for coherent video paragraph captioning. In ACL, 2020. 3, 4, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[17] Yitong Li, Zhe Gan, Yelong Shen, Jingjing Liu, Yu Cheng, Yuexin Wu, Lawrence Carin, David Carlson, and Jianfeng Gao. Storygan: A sequential conditional gan for story visualization. In CVPR, 2019. 1, 2, 3, 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.235 + ], + "angle": 0, + "content": "[18] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.894, + 0.307 + ], + "angle": 0, + "content": "[19] Swee Kiat Lim, Yi Loo, Ngoc-Trung Tran, Ngai-Man Cheung, Gemma Roig, and Yuval Elovici. Doping: Generative data augmentation for unsupervised anomaly detection with gan. 2018 IEEE International Conference on Data Mining (ICDM), pages 1122–1127, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.894, + 0.363 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.365, + 0.893, + 0.394 + ], + "angle": 0, + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.395, + 0.894, + 0.437 + ], + "angle": 0, + "content": "[22] Adyasha Maharana and Mohit Bansal. Integrating visuospatial, linguistic and commonsense structure into story visualization. In EMNLP, 2021. 2, 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.438, + 0.894, + 0.493 + ], + "angle": 0, + "content": "[23] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Improving generation and evaluation of visual stories via semantic consistency. In NAACL-HLT, 2021. 2, 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.496, + 0.893, + 0.538 + ], + "angle": 0, + "content": "[24] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Storydall-e: Adapting pretrained text-to-image transformers for story continuation. ArXiv, 2022. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.894, + 0.594 + ], + "angle": 0, + "content": "[25] Gaurav Mittal, Tanya Marwah, and Vineeth N. Balasubramanian. Sync-draw: Automatic video generation using deep recurrent attentive architectures. In ACM Multimedia, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.597, + 0.894, + 0.652 + ], + "angle": 0, + "content": "[26] Saypraseuth Mounsaveng, Issam Laradji, Ismail Ben Ayed, David Vazquez, and Marco Pedersoli. Learning data augmentation with online bilevel optimization for image classification. arXiv, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.654, + 0.893, + 0.696 + ], + "angle": 0, + "content": "[27] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.698, + 0.893, + 0.727 + ], + "angle": 0, + "content": "[28] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.893, + 0.769 + ], + "angle": 0, + "content": "[29] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.771, + 0.893, + 0.814 + ], + "angle": 0, + "content": "[30] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Learn, imagine and create: Text-to-image generation from prior knowledge. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3134" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "[33] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[35] Meet Shah, Xinlei Chen, Marcus Rohrbach, and Devi Parikh. Cycle-consistency for robust visual question answering. CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.262, + 0.472, + 0.315 + ], + "angle": 0, + "content": "[36] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.318, + 0.472, + 0.358 + ], + "angle": 0, + "content": "[37] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.472, + 0.428 + ], + "angle": 0, + "content": "[38] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.431, + 0.472, + 0.47 + ], + "angle": 0, + "content": "[39] Yun-Zhu Song, Zhi Rui Tam, Hung-Jen Chen, Huiao-Han Lu, and Hong-Han Shuai. Character-preserving coherent story visualization. In ECCV, 2020. 2, 5, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.473, + 0.472, + 0.526 + ], + "angle": 0, + "content": "[40] Zhiqiang Tang, Yunhe Gao, Leonid Karlinsky, Prasanna Sattigeri, Rogerio Feris, and Dimitris Metaxas. Onlineaugment: Online data augmentation with less domain knowledge. arXiv, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.529, + 0.472, + 0.556 + ], + "angle": 0, + "content": "[41] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.558, + 0.472, + 0.585 + ], + "angle": 0, + "content": "[42] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NIPS, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.472, + 0.641 + ], + "angle": 0, + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.472, + 0.682 + ], + "angle": 0, + "content": "[44] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, June 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.685, + 0.472, + 0.753 + ], + "angle": 0, + "content": "[45] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.756, + 0.472, + 0.808 + ], + "angle": 0, + "content": "[46] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.811, + 0.472, + 0.851 + ], + "angle": 0, + "content": "[47] Chenfei Wu, Jian Liang, Lei Ji, F. Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pre-training for neural visual world creation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.853, + 0.472, + 0.894 + ], + "angle": 0, + "content": "[48] Zhiyue Wu, Yijie Wang, Haibo Mi, Hongzuo Xu, Wei Zhang, and Lanlan Feng. Oada: An online data augmentation method for raw histopathology images. ICONIP, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "3135" + } + ] +] \ No newline at end of file diff --git a/2023/Story Visualization by Online Text Augmentation with Context Memory/b1180507-958c-45ce-9ffe-139cc0187a6e_origin.pdf b/2023/Story Visualization by Online Text Augmentation with Context Memory/b1180507-958c-45ce-9ffe-139cc0187a6e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0161e3f264b80f3beda2c03c39f6f0c515c0dfdf --- /dev/null +++ b/2023/Story Visualization by Online Text Augmentation with Context Memory/b1180507-958c-45ce-9ffe-139cc0187a6e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:765fead399615727f823f7040e3176ababadd3e241478d03e73c8dabbc6d12be +size 9080754 diff --git a/2023/Story Visualization by Online Text Augmentation with Context Memory/full.md b/2023/Story Visualization by Online Text Augmentation with Context Memory/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b51145887891f13d8173e8f98373fe4334b9c4d8 --- /dev/null +++ b/2023/Story Visualization by Online Text Augmentation with Context Memory/full.md @@ -0,0 +1,323 @@ +# Story Visualization by Online Text Augmentation with Context Memory + +Daechul Ahn $^{1,\S}$ Daneul Kim $^{2}$ Gwangmo Song $^{3}$ Seung Hwan Kim $^{3}$ Honglak Lee $^{3,4}$ Dongyeop Kang $^{5}$ Jonghyun Choi $^{1,\dagger}$ + +$^{1}$ Yonsei University $^{2}$ GIST $^{3}$ LG AI Research $^{4}$ University of Michigan $^{5}$ University of Minnesota $\{\mathrm{dcahn, jc}\} @$ yonsei.ac.kr flytodk98@gm.gist.ac.kr gwangmo.song@lgresearch.ai skcruise@gmail.com honglak@eecs.umich.edu dongyeop@umn.edu + +# Abstract + +Story visualization (SV) is a challenging text-to-image generation task for the difficulty of not only rendering visual details from the text descriptions but also encoding a long-term context across multiple sentences. While prior efforts mostly focus on generating a semantically relevant image for each sentence, encoding a context spread across the given paragraph to generate contextually convincing images (e.g., with a correct character or with a proper background of the scene) remains a challenge. To this end, we propose a novel memory architecture for the Bi-directional Transformer framework with an online text augmentation that generates multiple pseudo-descriptions as supplementary supervision during training for better generalization to the language variation at inference. In extensive experiments on the two popular SV benchmarks, i.e., the Pororo-SV and Flintstones-SV, the proposed method significantly outperforms the state of the arts in various metrics including FID, character F1, frame accuracy, BLEU-2/3, and R-precision with similar or less computational complexity. + +# 1. Introduction + +Story visualization (SV) [17] is a task of generating a sequence of images from a paragraph, i.e., a sequence of natural language sentences. It is challenging for the requirement of rendering the visual details in images with convincing background of a scene – seasonal elements, environmental objects such as table, location, and the proper character appearing, which here we refer to as context, spread across the given text sentences. Specifically, it needs to encode implicit context presented in the given sentences since each one often omit visual details (i.e., they may be spread over the sentences) necessary to generate a semantically correct im + +![](images/8fd9a836d27890f7568aea7b4d5486157e940b4cdceec2ecfcc90611105eeaf7.jpg) + +![](images/1df1c4aca7b4b3bf57e58ddf623eb289f08b0c7e3d4bc98353703be65c9d7dfe.jpg) + +![](images/7ae9701ae8dcc4ac7bd78f0737dcd051f2db78d57e94f325ff2dbd2a0f922b43.jpg) +Figure 1. Linguistic variations in story visualization and the overview of the proposed method. (a) An example of a story data with various linguistic variations (Var.#) for each image. (b) Modeling temporal context spread across sentences by our context memory (Sec. 3.1). (c) Addressing linguistic variations by online text augmentation from each image for every epoch (Sec. 3.2). (d) Benefit of the proposed context memory (left) and online text augmentation (right). + +age. For example, we can think of Fig. 1-(d), where Sent.#1 (i.e., sentence 1) and Sent.#2 (i.e., sentence 2) are given as sequences. After Sent.#1 is given, generated image by Sent.#2 often exhibits a background that is not semantically + +correct [17,22,23]. However, if we use contextual information given in Sent.#1, "The woods are covered with snow", it leads to correct image with matching background. + +In addition, widely used benchmark datasets for story visualization task [9, 17] provide a single text-image pair for training and inference, mostly due to the annotation budget constraint. This prevents the model from learning language variations, and thus harms the linguistic generalization performance of the model. + +To address the aforementioned challenges without requiring large scale data and models, we propose a new memory scheme in bi-directional Transformer for encoding the context which generates pseudo-texts in an online fashion to address linguistic variations at inference. We call our model as Context Memory and Online Text Augmentation or CMOTA for short. We empirically validate that our model outperforms the state-of-the-art SV methods by large margins on various metrics evaluated with widely used benchmarks in the literature, i.e., Pororo-SV and Flintstones-SV. + +Note that while large pre-trained models [3,5-7, 31, 43] have shown great success in synthesizing an image or a video from a language description [13, 32, 38], huge computational complexity and large training data makes the models prohibited. Moreover, although we propose and evaluate the model for the standard benchmark datasets without large pretraining data trained with a large model, it would be interesting to apply and evaluate the proposal in large models for further improvement. + +We summarize our contributions as follows: + +- We propose a new memory architecture for Transformer to selectively make use of contexts in a story paragraph. +- For better generalization of linguistic variations in the given paragraph at inference, we generate pseudo-texts and augment them in an online fashion for richer linguistic supervision. +- Our model significantly outperforms prior arts (even some hyper-scale models) by large margins in five evaluation metrics with similar or less computational complexity. + +# 2. Related Work + +Story visualization. StoryGAN [17] is one of the recent methods that utilized a story-level discriminator to improve global consistency in generated images. CP-CSV [39] disentangles figure and background information to enhance character consistency. In order to improve global semantic matching between paragraph and generated image sequence, DuCo-StoryGAN [23] presents a pre-trained video captioner as an auxiliary loss along with other design improvements on top of StoryGAN. More recently, VLC-StoryGAN [22] utilizes constituency parse-trees and common sense knowledge to improve consistency and an object-level feedback loop to + +improve image quality. Another recent work VP-CSV [2] is a two-stage approach, i.e., 1) character generation and 2) background completion, using Transformer model to address this task. We discuss these in more detail in the supplementary material for the space sake. But the prior arts largely neglect to encode the story narrative, which is our primary contribution here. + +Very recently, Maharana et al. propose a new task setup of story continuation; using first image as a condition. They fine-tune the large model DALL-E [32] for the SV, which they call StoryDALL-E [24]. Ours differs from it in several aspects as follows. We have explicit memory connections between adjacent image generators using the context memory to globally encode the sentences. In contrast, [24] utilizes global story embedding as additional input for understanding context. Further, while it is huge in size (1.3B parameters), trained with 14 million text-image pair, ours are much larger (97M parameters) and outperforms it in multiple metrics by large margins; even in the image quality metric, FID. Please refer to Sec. 4.2 for empirical comparisons. + +Text-to-video generation. Similar to story visualization, text-to-video generation also generates multiple frames from a given text. Since the pioneering work of Sync-DRAW [25] in text-to-video generation, [9, 18, 27] utilize generative adversarial network for high-quality image generation. + +Recently, high-quality video generation models are proposed, including GODIVA [46] and NUWA [47]. Furthermore, recent studies generate high-resolution videos, making sequential frames in high-quality [11, 38]. However, most of state-of-the-art text-to-video model [11, 38] generates a video from a single sentence, mostly having consistent backgrounds. Very recently, there is a method proposed to generate a video from a long paragraph [45]. Although they generate the video in a long time horizon, they require both a mega-scale model trained with huge data and a detailed paragraph where each sentence is describing the scene that are close in time. In contrast, story visualization requires generating frames arbitrarily distant in time (i.e., so-called 'key-frames') corresponding to different sentences, requiring to generate an image sequence that have contextually convincing background. + +Text-to-image generation. Text-to-image generation is a sub-problem of story visualization, with literature focusing on semantic relevance and resolution improvements. Recently, text-based image synthesis has been greatly improved with the help of a vast amount of training data with a hyper-scale model including DALL-E [32] and its successor DALL-E2, CogView [5] and Make-A-Scene [8] using a sketch input. + +Although text-to-image generation models generates very high-quality images, it may lack encoding the context, metaphoric sentences spread acorss multiple sentences. In + +addition, naively using state-of-the-art text-to-image generation models is computationally prohibited. For example, diffusion-based models [31, 34] have hyper-scale model size, (e.g., Imagen [34] parameter count of 2-B, DALL-E2 [31] parameter count of 3.5-B) making it non-trivial for applying it in a wide range of inference scenarios that may not have the sufficient computing resource. Here, we consider relatively light architectures as our base model for computational efficiency. More discussion are in the supplement. + +Transformer using memory. To mitigate context fragmentation issue [4], i.e., losing long-term dependency over a data stream in context, there are efforts to encode long-term contexts in generating an image sequence. [4, 15] adopt a recurrent path into transformer architecture. Specifically, the modeling of new data segments is conditioned on historical hidden states produced in the previous time step and uses highly summarized memory states. Unlike the conventional memory architectures, we propose a novel memory that has a dense connection from the past with attentive weighting schemes for their better usage (Sec. 3.1). + +Online augmentation. While offline data augmentation that prepares data outside of learning process is prevalent in many literature [1, 12, 19], online-augmentation that depends on training is seldom explored especially in story visualization task [40]. [35, 40] propose an online augmentation for image classification and VQA task. [26] uses bi-level optimization in image classification. In medical domain, [48] investigate the online augmentation for personalized image based Histopathology diagnosis [48]. Note that we show the benefit of online augmentation over the offline augmentation in the SV for the first time in this literature. + +# 3. Approach + +There are multiple challenges in story visualization, including (1) generating semantically natural images without artifacts from a description, (2) encoding the context (e.g., consistent background of a scene - seasonal elements, environmental objects such as table and location, and consistent characters) spread across the sentences in a given paragraph [17] and (3) addressing the linguistic variation at the inference time (i.e., a given text description may not be in different writing style). + +For photo-realistic image generation from a language description, we recently witness unprecedented improvements in the quality of generated images by the help of large scale model [3,5-7, 31, 43]. To leverage the large model's benefits for the story visualization task, we may use it to generate an image sequence by gradually concatenating sentences to visualize the story (i.e., use the first sentence to generate the first image, use a concatenated sentence of the first and the second to generate the second, and so on). Although this + +![](images/3955a06e84f176a10a7ea75926fdecc9626cee52112529d11c1c239d5cc11879.jpg) +Figure 2. Bi-directional (between text and image modality) Transformer for text or image 'sequence' generation with the proposed memory. Our bi-directional 'base' model can simultaneously generate a sequence of both text and image from $t = 1 \ldots 5$ with the proposed memory (Sec. 3.1). We use a segment embedding to indicate input modality and enable output of any modality. + +![](images/3b70e431c038215fce69a5c897e9a5815e323660565cd9c0737be7af8c8f5b4d.jpg) + +may generate a single photo-realistic image, it is not able to capture the context sparsely spread across the sentences as story progresses (see supplement for more discussion). + +A more involved way of using the large model for SV is to fine-tune it for the downstream SV or using a video generation models. However, as the most high-performing models are huge in size and the codes and the pre-trained models are not publicly available, the computational cost and reproducibility reduces practicality. + +For a given dataset without using large extra training data, we propose a new memory module to better encode past information with a Transformer by an attentively weighted densely connected architecture (Sec. 3.1). We further propose to generate pseudo-texts during the learning process to augment (online-augmentation) for better linguistic generalization without requiring large external data by learning the bi-directional Transformer [14, 30, 33] in both directions of generating images from texts and vice versa as illustrated in Fig. 2 (Sec. 3.2). + +Base model. As shown in Fig. 2, we use bi-directional (i.e., multi-modal) transformer that iteratively generates images and texts in both ways. Similar to [32], the image tokens are sequentially predicted from the input text sequences by the Transformer. Then, the decoder of the VQ-VAE [42] translates the predicted image tokens into image sequence. The text tokens are also sequentially predicted from the input image token sequence by the same Transformer. + +Particularly, for the bi-directional multi-modal generation, i.e., generating simultaneously text and image from the unified architecture, we add two embeddings; a positional embedding for absolute position between tokens and a segment embedding for distinguishing source and target. Tokens of a text $\left(\{t_{1},\ldots ,t_{m}\}\right)$ and an image $\left(\{z_{1},\ldots ,z_{n}\}\right)$ $(m,n:\#$ of tokens for text and image) are fed into the Trans + +![](images/ad8c26c496bb10c7838063ee34de59e890db959c6be088b6191e309a03c38613.jpg) +Figure 3. Proposed memory module for the bi-directional Transformer. Memory module updates current memory $M_{t}$ by utilizing memory attention and GRU operation with propagated memory, $M_{t-1}$ , and previously calculated hidden state, $H_{t}^{l}$ . To produce image, we propose a novel memory attention (Attn.) mask denoted by blue squares to select only text tokens for less bias to previously generated image tokens. Within the memory attention mask, white-colored squares filter out the content, i.e., image (img) token, whereas the blue squares allow the text (txt) token to propagate through for memory interaction. + +former to predict tokens in the other modality in multiple epochs with the following objective function written as: + +$$ +\mathcal {L} _ {j, t 2 i} = \sum_ {k = 1} ^ {n} - \ln p _ {j} (z _ {k} | t _ {1}, \dots , t _ {m}, z _ {1}, \dots , z _ {k - 1}), +$$ + +$$ +\mathcal {L} _ {j, i 2 t} = \sum_ {k = 1} ^ {m} - \ln p _ {j} \left(t _ {k} \mid z _ {1}, \dots , z _ {n}, t _ {1}, \dots , t _ {k - 1}\right), \tag {1} +$$ + +$$ +\mathcal {L} _ {j} = \mathcal {L} _ {j, t 2 i} + \lambda_ {1} \mathcal {L} _ {j, i 2 t}, +$$ + +where $p_j(\cdot)$ is a likelihood of $j$ -th generated tokens in one modality given the other modality, $L_{j,t2i}$ is a loss corresponding to $j$ -th text-to-image generation (i.e., negative log likelihood), $L_{j,i2t}$ is vice versa. $\lambda_1$ is a balancing hyperparameter. + +To train the model, we iteratively train the generation model in each modality multiple times per each epoch. + +# 3.1. Context Memory + +To encode the context and calculate the propagated memory, as depicted in Fig. 3, we first apply cross attention between the current hidden state, $H_{t}^{l} \in \mathcal{R}^{T_{c} \times d}$ calculated from $(l - 1)$ -th transformer layer and memory state at time $(t - 1)$ , $M_{t - 1} \in \mathcal{R}^{T_{M} \times d}$ ( $T_{c}$ : # tokens, $T_{M}$ : # memory states, $d$ : hidden state dimension). We then obtain $S_{t} = \text{Attn}(M_{t - 1}, H_{t}, H_{t})$ , where $\text{Attn}(Q, K, V) \coloneqq \text{Softmax}\left(\frac{QK^{T}}{\sqrt{d_{q}}}\right)V$ and $d_{q}$ is a query dimension ( $Q$ ), with the memory attention mask depicted in the blue box in Fig. 3. + +In particular, we apply a memory mask in the attention operation to select text tokens as memory content (depicted as blue-shaded grid cells in the right of Fig. 3) because including image content as memory could be strong constraint for text-to-image generation (see empirical studies of the + +![](images/896d69dc3fe965aaaea659183cc0f3e253e467e30dd31e1970137e125d9668c9.jpg) +(a) Conventional memory connection (Dai et al., ACL'19/Lei et al., ACL'20) Layer # + +![](images/bed5f34fa3f0b433bf9403ca37139612e6068ddffadc1f83c8bd2baec37415b7.jpg) +(b) Proposed memory connection +Figure 4. Comparison of our memory connection scheme to the conventional one [4,15]. We depict the memory connections in the multi-layer Transformer architecture per each sentence input (unfolded in a row). Unlike (a) the previous memory architecture having a single memory path from the immediate past in all layers, (b) ours has a dense connection from all the pasts to be attentively weighted (denoted here by $\bar{M}$ 's) in only the last layer. We discuss its empirical benefit in Sec. 4.2.2. + +proposed mask in the supplementary material). Then we feed the $S_{t}$ and $M_{t-1}$ to the GRU in order to compute the information to be stored in the memory, $M_{t}$ , propagated to $(t+1)$ . + +Memory connection. Following [4, 15], we stack multiple Transformer's layers, depicted in Fig. 4-(b). Compared to conventional memory updating architectures with serial hidden state connections in all layers (Fig. 4-(a)), our approach only has memory connections in the last layer, which we call it as partial-level memory augmented (PMA). This is because the later layers (closer to the last layer) achieve better representations with higher-level features, abstract and structured representations [28, 29, 41] with improved computational efficiency, which is unlike prior works (see Sec. 4.2.2 and supp for further discussion). + +Additionally, not all historic information is equally important for generating an image at a time step. Similar to the masked self-attention [43], we attentively weight the past information for better modeling of sparse context as: + +$$ +\bar {M} _ {1: (t - 1)} = A t t n (M _ {(t - 1)}, M _ {[ 1: (t - 2) ]}, M _ {[ 1: (t - 2) ]}), +$$ + +$$ +\tilde {M} _ {(t - 1)} = \left[ M _ {t - 1}; \bar {M} _ {1: (t - 1)} \right], \quad (3 \leq t \leq 5), \tag {2} +$$ + +$$ +H _ {t} ^ {l} = \operatorname {A t t n} \left(H _ {t} ^ {l}, \left[ H _ {t} ^ {l}; \tilde {M} _ {(t - 1)} \right], \left[ H _ {t} ^ {l}; \tilde {M} _ {(t - 1)} \right]\right), +$$ + +where $M_{(t - 1)}$ is a memory at time $(t - 1)$ and $[1:(t - 2)]$ refers to the concatenation from time 1 to $(t - 2)$ as depicted in Fig. 4-(b). By doing so, we fuse the contextual information into current hidden state $H_{t}$ . At $t = 2$ , we use $M_{1}$ , instead of $\tilde{M}_{1}$ as the $M_{1}$ is only available. We call it as attentively weighted memory (AWM). + +# 3.2. Online Text-Augmentation + +Vedamtam et al. argue that multiple descriptions for an image help generalization as they address language variations presented in descriptions [44]. Hence, a number of image captioning datasets [10, 20, 37] provide multiple natural language directives, obtained by multiple human annotators, in both training and evaluation splits. But the SV benchmark datasets [17] provide only a single sentence per an image. + +To address the linguistic variations of a text input at inference process, we first consider to generate a pseudo text by a well-trained image-to-text generation model, which we refer it as offline-augmentation [16]. But the offline augmentation generates only a single sentence, which may not provide sufficient diversity. Instead, we propose to generate multiple pseudo-texts and augment them in an online fashion when training our model to increase the diversity. We call it online text augmentation, depicted in Fig. 5. + +Thanks to our bi-directional multi-modal architecture, we can naturally integrate the process of generating pseudotexts to the process of learning image-to-text model and the text-to-image generation model as depicted in Fig. 5. In the early epochs, less meaningful sentences are generated, but as training progresses, more meaningful sentences are generated (see orange box in Fig. 5). As a side-product, by supervising the model learning with intermediate goals at each time step, we expect to expedite the convergence of learning. When we use the online text augmentation, we can rewrite the objective as: + +$$ +\mathcal {L} _ {j, p t 2 i} = \sum_ {k = 1} ^ {n} - \ln p _ {j} \left(z _ {k} \mid \hat {t} _ {1}, \dots , \hat {t} _ {m}, z _ {1}, \dots , z _ {k - 1}\right), \tag {3} +$$ + +$$ +\mathcal {L} _ {j} = \mathcal {L} _ {j, t 2 i} + \lambda_ {1} \mathcal {L} _ {j, i 2 t} + \lambda_ {2} \mathcal {L} _ {j, p t 2 i}, +$$ + +where $\mathcal{L}_{j,pt2i}$ is the additional loss with the augmented pseudo-texts and $\mathcal{L}_{j,t2i},\mathcal{L}_{j,i2t}$ are defined in Eq. 1 and $\lambda_{1}$ and $\lambda_{2}$ are balancing hyper-parameters. $\hat{t}$ means the pseudotext token, predicted during online augmentation without gradient flow. Fig. 6 shows pseudo-texts generated by our method. Detailed training procedure and more examples are in the supplementary material. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We use two popular benchmark datasets for evaluating the task of story visualization. Following [23], we use + +![](images/04d01c9788f45c18214ccf229c06ae564bd6c3ce9a6ce0020c44d05e099441e7.jpg) +Figure 5. Online text augmentation. We iteratively augment the pseudo-text for better linguistic generalization. (1) generating a pseudo-text at each epoch for each ground-truth image, without gradient flow. (2) generating an image for every epoch. (3,4) learn the text-to-image model with $L_{CE}$ . In early epochs, it generates less meaningful pseudo-text (thus discard them), but produces pseudotexts matching with the input image as the training progresses. + +Pororo-SV dataset for story visualization task without data overlap. Following prior work [22], we use Flintstones-SV which was originally exploited in the text-to-video synthesis task [9]. + +Metrics. For the evaluation metrics, we use Fréchet Inception Distance (FID) (used in [22, 39]) for visual quality of the generated images, character classification score (Char. F1, Frm. Acc.) [17] for character consistency, and global semantic matching (B-2/3, R-Prec.) [22, 23] between generated image sequence and story paragraph.. We elaborate on the details of the datasets and implementation in the supplementary material for the sake of space. + +Baselines. We compare our method to a number of prior arts including state-of-the-art story visualization methods such as StoryGAN [17], CPCSV [39], DuCo-StoryGAN [23], VLC-StoryGAN [22] and VP-CSV [2] for both quantitative and qualitative analysis in the story visualization setting. Moreover, we compare [24] in the story continuation setting. + +Implementation details. To train our model, we first tokenize the text and image inputs. Particularly, the image inputs are encoded by the encoder of the VQ-VAE [42] into image tokens and decoded by the decoder of the VQ-VAE for generating image sequence. Once the inputs are tokenized, we append two special embedding tokens to the text and image tokens. These tokens signify the beginning of each modality, denoted as 'SOS' (start of sentence) and 'SOI' (start of image). Subsequently, we add two embeddings: (1) a positional embedding to indicate absolute position of each token and (2) a segment embedding to distinguish source and target modality, e.g., for text-to-image generation task, the source corresponds to the text while the target corresponds to the image. Through the utilization of segment embeddings to distinguish between source and target, we enable the model to generate target data from source data + +![](images/50f14c46cbb54df636f9d29a4eaf74fcb6fb20a0475a951b141c0b65cbd9d59b.jpg) +Figure 6. Generated pseudo-texts by the proposed method during training. We depict the generated pseudo-texts in training procedure. We utilize them as supplementary supervisions to address linguistic diversity of a sentence at inference. + +within a unified architecture, as illustrated in Fig. 2. Here, we set the token length for each modality as $T_{text} = 80$ and $T_{image} = 256$ (i.e., $16 \times 16$ ), the hidden dimension size to 512, the number of transformer layer to 6, and the number of attention heads to 16, thereafter the number of trainable parameters approximately 93.7-M as shown in Tab. 1. Particularly, we generate an image per each sentence without any sampling process. We train the CMOTA using AdamW [21] optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.95$ , $\epsilon = 1e - 8$ , weight decay multiplier with $1e - 2$ and learning rate with $4.5e - 6$ multiplied by batch size. Moreover, we set the hyper-parameters for loss balancing in Eq. 3 as $\lambda_{1} = 1.0$ and $\lambda_{2} = 0.5$ and the number of memory state, $T_{M}$ as 1. Finally, to generate a high-resolution image, we modified the VQ-VAE with the additional trainable parameters (i.e., 7.1-M shown in Tab. 1). This modification allows for the generation of high-resolution images, increasing the image resolution from $64 \times 64$ to $128 \times 128$ . + +# 4.2. Quantitative Analysis + +# 4.2.1 Comparison with the State of the Arts + +In Table 1, we summarize the performance of the prior arts and the proposed CMOTA in Pororo-SV [17] and Flintstones-SV [9] dataset in the various metrics. By default, all methods generate $64 \times 64$ images along with the higher resolution $128 \times 128$ images. CMOTA outperforms existing methods by a large margin in both benchmarks in most of metrics. + +The very recent method VP-CSV [2] performs on par to our CMOTA while ours still outperforms in FID, BLEU and R-precision, implying that CMOTA generates high quality images maintaining global semantic matching between story paragraph and images better than the VP-CSV. The lower performance than the VP-CSV in Char. F1 score and Frm. Acc. is attributed to the specialized character-centric module that only focuses the model to generate accurate characters + +for each image at the expense of the other performance metrics. In addition, our CMOTA-HR model even outperforms the VP-CSV [2] by large margin without the specialized character-centric module. + +Low resolution to high resolution. Not surprisingly, CMOTA-HR outperforms CMOTA with default (low-) resolution (i.e., $64 \times 64$ ) in almost all metrics. Interestingly, however, VLC-StoryGAN-HR [22] shows similar performance to $64 \times 64$ model in Frame accuracy, BLEU, R-precision, and drops performance in all other metrics. For the increased FID scores by the VLC-StoryGAN when goes to high resolution, we hypothesize that the constituency parse trees of input sentences in the VLC-StoryGAN increases the difficulty in generating visual details as the images contain more visual details in high resolution. + +In addition, we observe that the performance drop of FID in both our CMOTA and VLC, when goes to high resolution. It is because as high-resolution image requires the model to depict fine-grained details, generation task would be more difficult compared to the low-resolution image generation. In particular, the drop of FID in Flintstones-SV is larger than that in Pororo-SV. It is because there are more complicated visual details in the Flintstones-SV than the Pororo-SV dataset. Note that the FID drops by our method from low-resolution to high resolution is less than the VLC in Pororo-SV. Furthermore, in terms of computational cost, our CMOTA gains performance boost by using high-resolution effectively with just adding relatively small computational cost on top of original $64 \times 64$ model. + +Character consistency. Although our method does not explicitly enforce the character semantic preservation, the proposed memory delivers the semantic across the sentence implicitly. We here investigate the character context preservation performance by comparing our method to [22, 23] on per-character classification F1-score on test split of Pororo + +
DatasetResolutionMethods# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
Pororo-SV64 × 64StoryGAN [17]-158.0618.599.343.24 / 1.221.51 ± 0.15
CP-CSV [39]-140.2421.7810.033.25 / 1.221.76 ± 0.04
DuCo-StoryGAN [23]101M96.5138.0113.973.68 / 1.343.56 ± 0.04
VLC-StoryGAN [22]100M84.9643.0217.363.80 / 1.443.28 ± 0.00
VP-CSV [2]-65.5156.8425.874.45 / 1.806.95 ± 0.00
CMOTA (Ours)96.6M52.1353.2524.724.58 / 1.907.34 ± 0.03
128 × 128VLC-StoryGAN-HR†[22]102.6M97.0840.3617.173.89 / 1.583.47 ± 0.03
CMOTA-HR (Ours)103.7M52.7758.8628.895.45 / 2.3416.36 ± 0.05
Flinstones-SV64 × 64StoryGAN [17]-127.1946.2032.9613.87 / 7.831.72 ± 0.18
DuCo-StoryGAN [23]101M78.0254.9236.3415.48 / 9.172.64 ± 0.17
VLC-StoryGAN‡[22]100M72.8758.8139.18--
CMOTA (Ours)96.6M36.7179.7466.0119.85 / 12.9810.50 ± 0.35
128 × 128CMOTA-HR (Ours)103.7M54.8186.4474.0622.18 / 15.1725.71 ± 0.70
+ +Table 1. Quantitative comparison with the state of the arts. On the test split of Pororo-SV and Flintstones-SV. # Params. refers to the number of trainable parameters. Char. F1 refers to character F1 score. Frm. Acc. refers to frame accuracy. R-Prec. refers to R-Precision. ↓ indicates 'lower the better' and ↑ indicates 'higher the better'. Experiments are done in both $64 \times 64$ and $128 \times 128$ resolutions. 'HR' refers to its high resolution $(128 \times 128)$ version. † indicates our results with author's implementation (https://github.com/adymaharana/VLCStoryGan).‡ indicates the absolute values that we computed by the given relative values in [22]. + +![](images/2e8a7fb8bd7ea9252edc4ce8f6894437c847fffec2e2857fa70e8a47d1aa3b63.jpg) +Figure 7. Per-character Classification F1-score on the Test Split of Pororo-SV Dataset. Normalized character (Char.) frequency is shown in red line. It represents the relative number of characters compared to the count of Pororo in training dataset. While DuCo/VLC-StoryGAN shows degraded performance as the character frequency decreases, CMOTA outperforms them across all the characters. + +SV dataset and normalized character frequency in train split dataset as illustrated in Fig. 7. Especially, it is interesting to compare with [2] as it is explicitly proposed for that purpose. + +As shown in the figure, our method outperforms prior arts [22, 23] by large margins. Although we did not have any special module for character information preserving, unlike [2], our model maintains such information over multiple sentences, even outperforming the image quality denoted as FID by $-13.38$ . + +# 4.2.2 Ablation Studies + +Benefit of the proposed architectural components. To investigate the benefit of each proposed architectural component, we build up the full model with the components from the vanilla, single-directional transformer without memory + +as summarized in Tab. 2. + +Using the proposed partial-level connected memory architecture only with the vanilla transformer (Tr.), the generation performance improves in all metrics compared to the Tr. Also, if we compare it with all-level connected memory architecture [15], PMA shows better performance in all metrics as in Tab. 3 (see supplementary material for more experimental analysis for various memory connection schemes). Employing attentional weighted memory (AWM) improves performance overall with slight degradation in FID (+0.12). The bi-directional learning scheme (Bi.) improves performance in all metrics except the character classification related metrics, i.e., Char.F1 and Frm. Acc. It is because there are captions that does not contain characters' name, just mentioning characters as 'friends', making the model difficult to generate specific characters. + +Benefit of the text augmentation. By augmenting the generated text in offline manner (i.e., using an image captioning model pretrained on the dataset of interest), we observe overall performance increase in all metrics. We hypothesize its reason as that the pseudo-text may convey contextual information such as characters' existence or description of background, thereby making the generation model to be robust to language variation as shown in a first row of Tab. 4. By the online augmentation of pseudo texts (i.e., gradually updating text generator, thus generating a diverse and gradually varying set of pseudo-texts), the performance further improves, as shown in Tab. 2 and Tab. 4. + +# 4.2.3 Comparison with Large-Scale Models + +Despite the unfairness of comparison to the large-scale models due to the model size, we compare CMOTA to larger + +
Methods# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
VLC-StoryGAN [22]100M84.9643.0217.363.28 / 1.443.28
Single Directional Transformer (Tr.)93.7M63.8845.4818.444.18 / 1.695.67
+ Partial-level Memory Connection (PMA)95.8M59.0549.7221.794.41 / 1.776.28
+ Attentional Weighted Memory (AWM)96.6M56.9850.1522.014.50 / 1.916.79
+ Bi-directional Training (Bi.)96.6M54.6951.2722.154.52 / 1.847.12
+ Online Augmentation (Full model)96.6M52.1353.2524.724.58 / 1.907.34
+ +Table 2. Benefit of the proposed components. On the test split of Pororo-SV Dataset. As a baseline, we use a single directional transformer-based text-to-image generation model. We gradually add each of proposed components to investigate its effect for improving performance, i.e., partial-level memory connection (Sec. 3.1), attentively weighted memory (Sec. 3.1), bi-directional training (Sec 3.2), and online-augmentation (Sec. 3.2). + +
Memory Architecture# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
All-Level Connection [15]118M61.2347.2119.214.21 / 1.726.08
Partial-level Connection (Ours)95.8M59.0549.7221.794.41 / 1.776.28
+ +Table 3. Benefit of the proposed memory connection scheme. Proposed scheme outperforms conventional memory connection that uses all-level connections [15] (Fig. 4-(a)) with less number of parameters (test split of Pororo-SV). + +
Augmentation# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
Offline96.6M54.5151.3222.314.50 / 1.907.09
Online (Ours)96.6M52.1353.2524.724.58 / 1.907.34
+ +models [24] in Tab. 5. it is surprising to see that CMOTA outperforms on an evaluation metric of character classification (i.e., Char.F1 and Frm. Acc.), regardless of using prompt-tuning or fine-tuning of StoryDALL-E. + +But as expected, the image quality denoted as FID is worse than the StoryDALL-E thanks to the huge size of the model and training dataset. Nevertheless, FID score is slightly better (Pororo-SV) and on-par (Flintstones-SV) compared to 'prompt-tune' that updates $30\%$ of parameters [24]. We discuss more about this in the supplementary material. + +# 4.3. Human Preference Study + +We conduct a larger scale (than the prior arts) human survey using the Amazon Mechanical Turk platform to qualitatively compare the generated image sequence by our method over VLC-StoryGAN [22] on three criteria, i.e., visual quality, temporal consistency, and semantic relevancy between generated images and descriptions, following [17, 22, 23, 39]. Unlike the previous arts that employ 2 to 9 human judges [2, 17, 22, 23, 39], we employ 100 judges for statistically more reliable results. + +As shown in the top table in Tab. 6, it is clear that human subjects prefer the generated images by our model over those from VLC-StoryGAN [22] on winning ratio(\%). We discuss more details about this study in the supplementary material. + +Table 4. Benefit of the online augmentation. On the Pororo-SV test split. The proposed online augmentation outperforms the offline augmentation using a pretrained captioner [15]. + +
Methods# Params.FID↓Char. F1↑Frm. Acc↑
Porre-SVStoryDALL-E (prompt-tune)1.3B61.2329.6811.65
StoryDALL-E (fine-tune)1.3B25.9036.9717.26
MEGA-StoryDALL-E (fine-tune)2.8B23.4839.9118.01
CMOTA (Ours)96.6M55.2651.4822.73
Flinstones-SVStoryDALL-E (prompt-tune)1.3B53.7142.3832.54
StoryDALL-E (fine-tune)1.3B26.4973.4355.19
MEGA-StoryDALL-E (fine-tune)2.8B23.5874.2654.68
CMOTA (Ours)96.6M58.5979.7562.98
+ +Table 5. Quantitative comparisons to large scale models. On the test split of Pororo-SV and Flintstones-SV dataset. All models are implemented on top of the large pretrained transformer (i.e., StoryDALL-E [24] pretrained on 14 million and MEGAStoryDALL-E [24] pretrained on 15 million from Conceptual Caption dataset [36]). In [24], the 'prompt-tune' update $30\%$ of model parameters compared to full fine-tune'. Without pretraining, we use a much small model compared to the proir arts, i.e., 96.6M (CMOTA) vs. 1.3B (StoryDALL-E). + +
ResolutionAttributeVLC-SG [22]TieOurs
64 × 64 (Low-Res.)Visual Quality27.8%8.6%63.6%
Temproal Consistency24.3%8.3%59.0%
Semantic Relevance31.2%10.8%57.9%
128 × 128 (High-Res.)Visual Quality21.9%1.5%76.6%
Temporal Consistency21.8%2.5%75.7%
Semantic Relevance23.6%1.7%74.6%
+ +
CMOTA w/o Mem.TieCMOTA
Temporal Consistency34.5%4.3%61.2%
+ +Table 6. Human preference studies. (Top) With 100 judges in the Amazon Mechanical Turk, on Pororo-SV test split dataset. Win $(\%)$ refers to the $\%$ times one model is preferred over the others. "Tie" refers to the same. Ours are clearly preferred over the VLC-SG (refers to VLC-StoryGAN) [22]. (Bottom) For the model using our memory (Mem.) module, contrasting results with and without the memory module that considers the contexts spread across the sentences, our method significantly improves human preference in temporal consistency. + +Human preference for memory usage. We further investigate the benefit of using the proposed memory module (Sec. 3.1) by the human study with 100 annotators in Amazon Mechanical Turk. For the comparative analysis between + +CMOTA and CMOTA-w/o-Memory, we use 30 randomly sampled image sequences on Pororo-SV test set to conduct A/B test following [39]. Particularly, the annotators need to consider the temporal consistency with semantic relevance between generated image sequence and captions. As depicted in the bottom table in Tab. 6, we observe that $61.2\%$ of annotators prefer our CMOTA with memory module over the one without it. We can conclude that the memory module has a great deal in creating temporally consistent image sequence with respect to semantic relevancy. + +# 4.4. Qualitative Analysis + +We now qualitatively analyze the quality of generated image sequences in Fig. 8 on the Pororo-SV dataset's test split. The top row shows the image sequence of ground truth, the two rows (2-3) contain prior works [17,23] and the final row is the image sequence generated by CMOTA and its high resolution version (CMOTA-HR). CMOTA generates a semantically more plausible image sequence with better visual quality, compared to prior works [22, 23]. Particularly, the CMOTA-HR demonstrates the ability to generate visual details more effectively even as the size of the image increases, while maintaining similar semantics. + +Furthermore, we qualitatively investigate the advantage of using memory architecture and summarize the results in Fig. 9. As shown in the figure, the proposed memory architecture generates a semantically more plausible image sequence with proper context, e.g., background; the CMOTA without memory fails to capture proper background context (shown in dotted red box) since a single sentence could be interpreted in many ways. In contrast, CMOTA (with the context memory) generates an image sequence with plausible background with characters preserved without an explicit character-centric model. + +# 5. Conclusion + +We propose to better encode semantic context, e.g., plausible background and characters, for story visualization using a new memory architecture in a multi-modal bi-directional Transformer. We further propose an online text augmentation training scheme to generate pseudo-text descriptions as an intermediate supervision for addressing linguistic diversity in the texts at inference. The proposed method generates a temporally coherent and semantically relevant image sequence for each sentence in the given text paragraph. + +The proposed method outperforms prior works by a large margin on various metrics on the two popular SV benchmark datasets, and also outperforms some of hyper-scale models in multiple semantic understanding metrics. Although computationally prohibited and the pre-trained models are not publicly available, it would be intriguing to apply the proposed memory module and the online augmentation scheme to a large model. + +![](images/707bec45ef572e62b2f3e99daa00148fa57ea05e04ad390e3f96aebc6e8b8ff2.jpg) +Figure 8. Qualitative results compared to prior arts. On the Pororo-SV's test split. We compare our CMOTA to the prior arts including DuCo-StoryGAN and VLC-StoryGAN. Ours generates a semantically more plausible and temporally coherent image sequence compared to the prior arts. All images except CMOTA-HR $(128\times 128)$ are generated with resolution of $64\times 64$ for comparison. + +![](images/37b9e587e9dd43b4f188deca6e4b7bfb988c0d7b4421954911bb2ffc4745a3c3.jpg) +Figure 9. Benefit of memory module in encoding background context. Our CMOTA generates both semantically and visually plausible image sequences. Without memory module, inconsistent images (i.e., abrupt background change) are generated as indicated by the red dotted box. + +Acknowledgement. This work is partly supported by the NRF grant (No.2022R1A2C4002300) $25\%$ , IITP grants (No.2020-0-01361, AI GS Program (Yonsei University) $5 \%$ , No.2021-0-02068, AI Innovation Hub $5 \%$ 2022-0-00077 $15\%$ , 2022-0-00113 $15\%$ , 2022-0-00959 $15\%$ , 2022-0-00871 $10\%$ , 2022-0-00951 $10\%$ ) funded by the Korea government (MSIT). + +# References + +[1] Alexander Buslaev, Vladimir I. Iglovikov, Eugene Khvedchenya, Alex Parinov, Mikhail Druzhinin, and Alexandr A. Kalinin. Albumimations: Fast and flexible image augmentations. Information, 11(2), 2020. 3 +[2] Hong Chen, Rujun Han, Te-Lin Wu, Hideki Nakayama, and Nanyun Peng. Character-centric story visualization via visual planning and token alignment. ArXiv, 2022. 2, 5, 6, 7, 8 +[3] Mark Chen, Alec Radford, Jeff Wu, Heewoo Jun, Prafulla Dhariwal, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 2, 3 +[4] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov. Transformer-XL: Attentive language models beyond a fixed-length context. In ACL, 2019. 3, 4 +[5] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. In Advances in Neural Information Processing Systems, 2021. 2, 3 +[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. ArXiv, 2022. 2, 3 +[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv, 2022. 2, 3 +[8] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, 2022. 2 +[9] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 2, 5, 6 +[10] Danna Gurari, Qing Li, Chi Lin, Yinan Zhao, Anhong Guo, Abigale Stangl, and Jeffrey P. Bigham. Vizwiz-priv: A dataset for recognizing the presence and purpose of private visual information in images taken by blind people. In CVPR, June 2019. 5 +[11] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. ArXiv, 2022. 2 +[12] Philip T. G. Jackson, Amir Atapour-Abarghouei, Stephen Bonner, T. Breckon, and Boguslaw Obara. Style augmentation: Data augmentation via style randomization. In CVPR Workshops, 2019. 3 +[13] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. IEEE Access, 2020. 2 +[14] Taehoon Kim, Gwangmo Song, Sihaeng Lee, Sangyun Kim, Yewon Seo, Soonyoung Lee, Seung Hwan Kim, Honglak Lee, and Kyunghoon Bae. L-verse: Bidirectional generation between image and text. In CVPR, 2022. 3 +[15] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L Berg, and Mohit Bansal. Mart: Memory-augmented recurrent transformer for coherent video paragraph captioning. In ACL, 2020. 3, 4, 7, 8 + +[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5 +[17] Yitong Li, Zhe Gan, Yelong Shen, Jingjing Liu, Yu Cheng, Yuexin Wu, Lawrence Carin, David Carlson, and Jianfeng Gao. Storygan: A sequential conditional gan for story visualization. In CVPR, 2019. 1, 2, 3, 5, 6, 7, 8, 9 +[18] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 2 +[19] Swee Kiat Lim, Yi Loo, Ngoc-Trung Tran, Ngai-Man Cheung, Gemma Roig, and Yuval Elovici. Doping: Generative data augmentation for unsupervised anomaly detection with gan. 2018 IEEE International Conference on Data Mining (ICDM), pages 1122–1127, 2018. 3 +[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5 +[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6 +[22] Adyasha Maharana and Mohit Bansal. Integrating visuospatial, linguistic and commonsense structure into story visualization. In EMNLP, 2021. 2, 5, 6, 7, 8, 9 +[23] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Improving generation and evaluation of visual stories via semantic consistency. In NAACL-HLT, 2021. 2, 5, 6, 7, 8, 9 +[24] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Storydall-e: Adapting pretrained text-to-image transformers for story continuation. ArXiv, 2022. 2, 5, 8 +[25] Gaurav Mittal, Tanya Marwah, and Vineeth N. Balasubramanian. Sync-draw: Automatic video generation using deep recurrent attentive architectures. In ACM Multimedia, 2017. 2 +[26] Saypraseuth Mounsaveng, Issam Laradji, Ismail Ben Ayed, David Vazquez, and Marco Pedersoli. Learning data augmentation with online bilevel optimization for image classification. arXiv, 2020. 3 +[27] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 2 +[28] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, 2019. 4 +[29] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 4 +[30] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Learn, imagine and create: Text-to-image generation from prior knowledge. In NeurIPS, 2019. 3 +[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv, 2022. 2, 3 +[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2, 3 + +[33] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 3 +[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. 2022. 3 +[35] Meet Shah, Xinlei Chen, Marcus Rohrbach, and Devi Parikh. Cycle-consistency for robust visual question answering. CVPR, 2019. 3 +[36] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 8 +[37] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, 2020. 5 +[38] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, 2022. 2 +[39] Yun-Zhu Song, Zhi Rui Tam, Hung-Jen Chen, Huiao-Han Lu, and Hong-Han Shuai. Character-preserving coherent story visualization. In ECCV, 2020. 2, 5, 7, 8, 9 +[40] Zhiqiang Tang, Yunhe Gao, Leonid Karlinsky, Prasanna Sattigeri, Rogerio Feris, and Dimitris Metaxas. Onlineaugment: Online data augmentation with less domain knowledge. arXiv, 2020. 3 +[41] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 4 +[42] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NIPS, 2017. 3, 5 +[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2, 3, 4 +[44] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, June 2015. 5 +[45] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv, 2022. 2 +[46] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions, 2021. 2 +[47] Chenfei Wu, Jian Liang, Lei Ji, F. Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pre-training for neural visual world creation. In ECCV, 2022. 2 +[48] Zhiyue Wu, Yijie Wang, Haibo Mi, Hongzuo Xu, Wei Zhang, and Lanlan Feng. Oada: An online data augmentation method for raw histopathology images. ICONIP, 2021. 3 \ No newline at end of file diff --git a/2023/Story Visualization by Online Text Augmentation with Context Memory/images.zip b/2023/Story Visualization by Online Text Augmentation with Context Memory/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..746ec8b4500286f2f1e7d2d474f2575e603805ae --- /dev/null +++ b/2023/Story Visualization by Online Text Augmentation with Context Memory/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e631cd16308dd37d82aa2357185d98ac31d2a1e79eeae208ea3eafd970960bd4 +size 893127 diff --git a/2023/Story Visualization by Online Text Augmentation with Context Memory/layout.json b/2023/Story Visualization by Online Text Augmentation with Context Memory/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8f86c30ad9b8e8ade95381e621588306be134b8b --- /dev/null +++ b/2023/Story Visualization by Online Text Augmentation with Context Memory/layout.json @@ -0,0 +1,8822 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 519, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 519, + 121 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 519, + 121 + ], + "type": "text", + "content": "Story Visualization by Online Text Augmentation with Context Memory" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "spans": [ + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "text", + "content": "Daechul Ahn" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "inline_equation", + "content": "^{1,\\S}" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "text", + "content": " Daneul Kim" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "text", + "content": " Gwangmo Song" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "text", + "content": " Seung Hwan Kim" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "text", + "content": " Honglak Lee" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "text", + "content": " Dongyeop Kang" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "text", + "content": " Jonghyun Choi" + }, + { + "bbox": [ + 108, + 142, + 484, + 174 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "spans": [ + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "text", + "content": "Yonsei University " + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "text", + "content": "GIST " + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "text", + "content": "LG AI Research " + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "text", + "content": "University of Michigan " + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "text", + "content": "University of Minnesota " + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{dcahn, jc}\\} @" + }, + { + "bbox": [ + 63, + 180, + 533, + 229 + ], + "type": "text", + "content": "yonsei.ac.kr flytodk98@gm.gist.ac.kr gwangmo.song@lgresearch.ai skcruise@gmail.com honglak@eecs.umich.edu dongyeop@umn.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 256, + 192, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 256, + 192, + 269 + ], + "spans": [ + { + "bbox": [ + 143, + 256, + 192, + 269 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 282, + 290, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 290, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 290, + 510 + ], + "type": "text", + "content": "Story visualization (SV) is a challenging text-to-image generation task for the difficulty of not only rendering visual details from the text descriptions but also encoding a long-term context across multiple sentences. While prior efforts mostly focus on generating a semantically relevant image for each sentence, encoding a context spread across the given paragraph to generate contextually convincing images (e.g., with a correct character or with a proper background of the scene) remains a challenge. To this end, we propose a novel memory architecture for the Bi-directional Transformer framework with an online text augmentation that generates multiple pseudo-descriptions as supplementary supervision during training for better generalization to the language variation at inference. In extensive experiments on the two popular SV benchmarks, i.e., the Pororo-SV and Flintstones-SV, the proposed method significantly outperforms the state of the arts in various metrics including FID, character F1, frame accuracy, BLEU-2/3, and R-precision with similar or less computational complexity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 533, + 128, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 128, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 128, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 553, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 288, + 685 + ], + "type": "text", + "content": "Story visualization (SV) [17] is a task of generating a sequence of images from a paragraph, i.e., a sequence of natural language sentences. It is challenging for the requirement of rendering the visual details in images with convincing background of a scene – seasonal elements, environmental objects such as table, location, and the proper character appearing, which here we refer to as context, spread across the given text sentences. Specifically, it needs to encode implicit context presented in the given sentences since each one often omit visual details (i.e., they may be spread over the sentences) necessary to generate a semantically correct im" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 256, + 544, + 359 + ], + "blocks": [ + { + "bbox": [ + 310, + 256, + 544, + 359 + ], + "lines": [ + { + "bbox": [ + 310, + 256, + 544, + 359 + ], + "spans": [ + { + "bbox": [ + 310, + 256, + 544, + 359 + ], + "type": "image", + "image_path": "8fd9a836d27890f7568aea7b4d5486157e940b4cdceec2ecfcc90611105eeaf7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 365, + 543, + 469 + ], + "blocks": [ + { + "bbox": [ + 310, + 365, + 543, + 469 + ], + "lines": [ + { + "bbox": [ + 310, + 365, + 543, + 469 + ], + "spans": [ + { + "bbox": [ + 310, + 365, + 543, + 469 + ], + "type": "image", + "image_path": "1df1c4aca7b4b3bf57e58ddf623eb289f08b0c7e3d4bc98353703be65c9d7dfe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 472, + 545, + 565 + ], + "blocks": [ + { + "bbox": [ + 308, + 472, + 545, + 565 + ], + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 565 + ], + "type": "image", + "image_path": "7ae9701ae8dcc4ac7bd78f0737dcd051f2db78d57e94f325ff2dbd2a0f922b43.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 567, + 547, + 655 + ], + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 655 + ], + "type": "text", + "content": "Figure 1. Linguistic variations in story visualization and the overview of the proposed method. (a) An example of a story data with various linguistic variations (Var.#) for each image. (b) Modeling temporal context spread across sentences by our context memory (Sec. 3.1). (c) Addressing linguistic variations by online text augmentation from each image for every epoch (Sec. 3.2). (d) Benefit of the proposed context memory (left) and online text augmentation (right)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "age. For example, we can think of Fig. 1-(d), where Sent.#1 (i.e., sentence 1) and Sent.#2 (i.e., sentence 2) are given as sequences. After Sent.#1 is given, generated image by Sent.#2 often exhibits a background that is not semantically" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 693, + 288, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 693, + 288, + 703 + ], + "spans": [ + { + "bbox": [ + 45, + 693, + 288, + 703 + ], + "type": "text", + "content": "§: work done while interning at LG AI Research. †: corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 703, + 235, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 703, + 235, + 712 + ], + "spans": [ + { + "bbox": [ + 45, + 703, + 235, + 712 + ], + "type": "text", + "content": "Code: https://github.com/yonseivnl/cmota" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3125" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "content": "correct [17,22,23]. However, if we use contextual information given in Sent.#1, \"The woods are covered with snow\", it leads to correct image with matching background." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 287, + 180 + ], + "type": "text", + "content": "In addition, widely used benchmark datasets for story visualization task [9, 17] provide a single text-image pair for training and inference, mostly due to the annotation budget constraint. This prevents the model from learning language variations, and thus harms the linguistic generalization performance of the model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 181, + 288, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 288, + 300 + ], + "type": "text", + "content": "To address the aforementioned challenges without requiring large scale data and models, we propose a new memory scheme in bi-directional Transformer for encoding the context which generates pseudo-texts in an online fashion to address linguistic variations at inference. We call our model as Context Memory and Online Text Augmentation or CMOTA for short. We empirically validate that our model outperforms the state-of-the-art SV methods by large margins on various metrics evaluated with widely used benchmarks in the literature, i.e., Pororo-SV and Flintstones-SV." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 301, + 288, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 288, + 408 + ], + "type": "text", + "content": "Note that while large pre-trained models [3,5-7, 31, 43] have shown great success in synthesizing an image or a video from a language description [13, 32, 38], huge computational complexity and large training data makes the models prohibited. Moreover, although we propose and evaluate the model for the standard benchmark datasets without large pretraining data trained with a large model, it would be interesting to apply and evaluate the proposal in large models for further improvement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 409, + 237, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 409, + 237, + 420 + ], + "spans": [ + { + "bbox": [ + 59, + 409, + 237, + 420 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 426, + 287, + 548 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 59, + 426, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 426, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 59, + 426, + 287, + 449 + ], + "type": "text", + "content": "- We propose a new memory architecture for Transformer to selectively make use of contexts in a story paragraph." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 451, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 451, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 59, + 451, + 287, + 498 + ], + "type": "text", + "content": "- For better generalization of linguistic variations in the given paragraph at inference, we generate pseudo-texts and augment them in an online fashion for richer linguistic supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 501, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 501, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 59, + 501, + 287, + 548 + ], + "type": "text", + "content": "- Our model significantly outperforms prior arts (even some hyper-scale models) by large margins in five evaluation metrics with similar or less computational complexity." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 561, + 134, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 134, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 134, + 573 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Story visualization. StoryGAN [17] is one of the recent methods that utilized a story-level discriminator to improve global consistency in generated images. CP-CSV [39] disentangles figure and background information to enhance character consistency. In order to improve global semantic matching between paragraph and generated image sequence, DuCo-StoryGAN [23] presents a pre-trained video captioner as an auxiliary loss along with other design improvements on top of StoryGAN. More recently, VLC-StoryGAN [22] utilizes constituency parse-trees and common sense knowledge to improve consistency and an object-level feedback loop to" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "type": "text", + "content": "improve image quality. Another recent work VP-CSV [2] is a two-stage approach, i.e., 1) character generation and 2) background completion, using Transformer model to address this task. We discuss these in more detail in the supplementary material for the space sake. But the prior arts largely neglect to encode the story narrative, which is our primary contribution here." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 156, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 312 + ], + "type": "text", + "content": "Very recently, Maharana et al. propose a new task setup of story continuation; using first image as a condition. They fine-tune the large model DALL-E [32] for the SV, which they call StoryDALL-E [24]. Ours differs from it in several aspects as follows. We have explicit memory connections between adjacent image generators using the context memory to globally encode the sentences. In contrast, [24] utilizes global story embedding as additional input for understanding context. Further, while it is huge in size (1.3B parameters), trained with 14 million text-image pair, ours are much larger (97M parameters) and outperforms it in multiple metrics by large margins; even in the image quality metric, FID. Please refer to Sec. 4.2 for empirical comparisons." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 321, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 380 + ], + "type": "text", + "content": "Text-to-video generation. Similar to story visualization, text-to-video generation also generates multiple frames from a given text. Since the pioneering work of Sync-DRAW [25] in text-to-video generation, [9, 18, 27] utilize generative adversarial network for high-quality image generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 381, + 546, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 546, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 546, + 572 + ], + "type": "text", + "content": "Recently, high-quality video generation models are proposed, including GODIVA [46] and NUWA [47]. Furthermore, recent studies generate high-resolution videos, making sequential frames in high-quality [11, 38]. However, most of state-of-the-art text-to-video model [11, 38] generates a video from a single sentence, mostly having consistent backgrounds. Very recently, there is a method proposed to generate a video from a long paragraph [45]. Although they generate the video in a long time horizon, they require both a mega-scale model trained with huge data and a detailed paragraph where each sentence is describing the scene that are close in time. In contrast, story visualization requires generating frames arbitrarily distant in time (i.e., so-called 'key-frames') corresponding to different sentences, requiring to generate an image sequence that have contextually convincing background." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": "Text-to-image generation. Text-to-image generation is a sub-problem of story visualization, with literature focusing on semantic relevance and resolution improvements. Recently, text-based image synthesis has been greatly improved with the help of a vast amount of training data with a hyper-scale model including DALL-E [32] and its successor DALL-E2, CogView [5] and Make-A-Scene [8] using a sketch input." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Although text-to-image generation models generates very high-quality images, it may lack encoding the context, metaphoric sentences spread acorss multiple sentences. In" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3126" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "addition, naively using state-of-the-art text-to-image generation models is computationally prohibited. For example, diffusion-based models [31, 34] have hyper-scale model size, (e.g., Imagen [34] parameter count of 2-B, DALL-E2 [31] parameter count of 3.5-B) making it non-trivial for applying it in a wide range of inference scenarios that may not have the sufficient computing resource. Here, we consider relatively light architectures as our base model for computational efficiency. More discussion are in the supplement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 191, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 191, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 289, + 323 + ], + "type": "text", + "content": "Transformer using memory. To mitigate context fragmentation issue [4], i.e., losing long-term dependency over a data stream in context, there are efforts to encode long-term contexts in generating an image sequence. [4, 15] adopt a recurrent path into transformer architecture. Specifically, the modeling of new data segments is conditioned on historical hidden states produced in the previous time step and uses highly summarized memory states. Unlike the conventional memory architectures, we propose a novel memory that has a dense connection from the past with attentive weighting schemes for their better usage (Sec. 3.1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 334, + 289, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 289, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 289, + 465 + ], + "type": "text", + "content": "Online augmentation. While offline data augmentation that prepares data outside of learning process is prevalent in many literature [1, 12, 19], online-augmentation that depends on training is seldom explored especially in story visualization task [40]. [35, 40] propose an online augmentation for image classification and VQA task. [26] uses bi-level optimization in image classification. In medical domain, [48] investigate the online augmentation for personalized image based Histopathology diagnosis [48]. Note that we show the benefit of online augmentation over the offline augmentation in the SV for the first time in this literature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 477, + 115, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 115, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 115, + 491 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 605 + ], + "type": "text", + "content": "There are multiple challenges in story visualization, including (1) generating semantically natural images without artifacts from a description, (2) encoding the context (e.g., consistent background of a scene - seasonal elements, environmental objects such as table and location, and consistent characters) spread across the sentences in a given paragraph [17] and (3) addressing the linguistic variation at the inference time (i.e., a given text description may not be in different writing style)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "For photo-realistic image generation from a language description, we recently witness unprecedented improvements in the quality of generated images by the help of large scale model [3,5-7, 31, 43]. To leverage the large model's benefits for the story visualization task, we may use it to generate an image sequence by gradually concatenating sentences to visualize the story (i.e., use the first sentence to generate the first image, use a concatenated sentence of the first and the second to generate the second, and so on). Although this" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 68, + 425, + 208 + ], + "blocks": [ + { + "bbox": [ + 309, + 68, + 425, + 208 + ], + "lines": [ + { + "bbox": [ + 309, + 68, + 425, + 208 + ], + "spans": [ + { + "bbox": [ + 309, + 68, + 425, + 208 + ], + "type": "image", + "image_path": "3955a06e84f176a10a7ea75926fdecc9626cee52112529d11c1c239d5cc11879.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 210, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 304, + 210, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 547, + 277 + ], + "type": "text", + "content": "Figure 2. Bi-directional (between text and image modality) Transformer for text or image 'sequence' generation with the proposed memory. Our bi-directional 'base' model can simultaneously generate a sequence of both text and image from " + }, + { + "bbox": [ + 304, + 210, + 547, + 277 + ], + "type": "inline_equation", + "content": "t = 1 \\ldots 5" + }, + { + "bbox": [ + 304, + 210, + 547, + 277 + ], + "type": "text", + "content": " with the proposed memory (Sec. 3.1). We use a segment embedding to indicate input modality and enable output of any modality." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 428, + 68, + 548, + 209 + ], + "blocks": [ + { + "bbox": [ + 428, + 68, + 548, + 209 + ], + "lines": [ + { + "bbox": [ + 428, + 68, + 548, + 209 + ], + "spans": [ + { + "bbox": [ + 428, + 68, + 548, + 209 + ], + "type": "image", + "image_path": "3b70e431c038215fce69a5c897e9a5815e323660565cd9c0737be7af8c8f5b4d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 297, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 332 + ], + "type": "text", + "content": "may generate a single photo-realistic image, it is not able to capture the context sparsely spread across the sentences as story progresses (see supplement for more discussion)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "text", + "content": "A more involved way of using the large model for SV is to fine-tune it for the downstream SV or using a video generation models. However, as the most high-performing models are huge in size and the codes and the pre-trained models are not publicly available, the computational cost and reproducibility reduces practicality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 404, + 547, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 547, + 524 + ], + "type": "text", + "content": "For a given dataset without using large extra training data, we propose a new memory module to better encode past information with a Transformer by an attentively weighted densely connected architecture (Sec. 3.1). We further propose to generate pseudo-texts during the learning process to augment (online-augmentation) for better linguistic generalization without requiring large external data by learning the bi-directional Transformer [14, 30, 33] in both directions of generating images from texts and vice versa as illustrated in Fig. 2 (Sec. 3.2)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 629 + ], + "type": "text", + "content": "Base model. As shown in Fig. 2, we use bi-directional (i.e., multi-modal) transformer that iteratively generates images and texts in both ways. Similar to [32], the image tokens are sequentially predicted from the input text sequences by the Transformer. Then, the decoder of the VQ-VAE [42] translates the predicted image tokens into image sequence. The text tokens are also sequentially predicted from the input image token sequence by the same Transformer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "text", + "content": "Particularly, for the bi-directional multi-modal generation, i.e., generating simultaneously text and image from the unified architecture, we add two embeddings; a positional embedding for absolute position between tokens and a segment embedding for distinguishing source and target. Tokens of a text " + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\left(\\{t_{1},\\ldots ,t_{m}\\}\\right)" + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "text", + "content": " and an image " + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\left(\\{z_{1},\\ldots ,z_{n}\\}\\right)" + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "inline_equation", + "content": "(m,n:\\#" + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "text", + "content": " of tokens for text and image) are fed into the Trans" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3127" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 289, + 171 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 289, + 171 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 289, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 289, + 171 + ], + "type": "image", + "image_path": "ad8c26c496bb10c7838063ee34de59e890db959c6be088b6191e309a03c38613.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 289, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 289, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 289, + 285 + ], + "type": "text", + "content": "Figure 3. Proposed memory module for the bi-directional Transformer. Memory module updates current memory " + }, + { + "bbox": [ + 46, + 175, + 289, + 285 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 46, + 175, + 289, + 285 + ], + "type": "text", + "content": " by utilizing memory attention and GRU operation with propagated memory, " + }, + { + "bbox": [ + 46, + 175, + 289, + 285 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 46, + 175, + 289, + 285 + ], + "type": "text", + "content": ", and previously calculated hidden state, " + }, + { + "bbox": [ + 46, + 175, + 289, + 285 + ], + "type": "inline_equation", + "content": "H_{t}^{l}" + }, + { + "bbox": [ + 46, + 175, + 289, + 285 + ], + "type": "text", + "content": ". To produce image, we propose a novel memory attention (Attn.) mask denoted by blue squares to select only text tokens for less bias to previously generated image tokens. Within the memory attention mask, white-colored squares filter out the content, i.e., image (img) token, whereas the blue squares allow the text (txt) token to propagate through for memory interaction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 287, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 328 + ], + "type": "text", + "content": "former to predict tokens in the other modality in multiple epochs with the following objective function written as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 337, + 258, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 337, + 258, + 368 + ], + "spans": [ + { + "bbox": [ + 63, + 337, + 258, + 368 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {j, t 2 i} = \\sum_ {k = 1} ^ {n} - \\ln p _ {j} (z _ {k} | t _ {1}, \\dots , t _ {m}, z _ {1}, \\dots , z _ {k - 1}),", + "image_path": "4fb1400fab5737b0e2efe0d1a308a19c4ad6171ff6efbe08465dd777c2c293d8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 370, + 287, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 370, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 63, + 370, + 287, + 401 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {j, i 2 t} = \\sum_ {k = 1} ^ {m} - \\ln p _ {j} \\left(t _ {k} \\mid z _ {1}, \\dots , z _ {n}, t _ {1}, \\dots , t _ {k - 1}\\right), \\tag {1}", + "image_path": "182093e84a0b3003986268672f0e99fbda38f2458b8ba497dd679e8542bc4c67.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 403, + 174, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 403, + 174, + 417 + ], + "spans": [ + { + "bbox": [ + 77, + 403, + 174, + 417 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {j} = \\mathcal {L} _ {j, t 2 i} + \\lambda_ {1} \\mathcal {L} _ {j, i 2 t},", + "image_path": "eb1db1d180330e42195a722b7d9dcd0cd459c920247b8c6ce002c2e86344d4d5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "inline_equation", + "content": "p_j(\\cdot)" + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": " is a likelihood of " + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": "-th generated tokens in one modality given the other modality, " + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "inline_equation", + "content": "L_{j,t2i}" + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": " is a loss corresponding to " + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": "-th text-to-image generation (i.e., negative log likelihood), " + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "inline_equation", + "content": "L_{j,i2t}" + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": " is vice versa. " + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": " is a balancing hyperparameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 510 + ], + "type": "text", + "content": "To train the model, we iteratively train the generation model in each modality multiple times per each epoch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 516, + 149, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 516, + 149, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 149, + 529 + ], + "type": "text", + "content": "3.1. Context Memory" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": "To encode the context and calculate the propagated memory, as depicted in Fig. 3, we first apply cross attention between the current hidden state, " + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "H_{t}^{l} \\in \\mathcal{R}^{T_{c} \\times d}" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": " calculated from " + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "(l - 1)" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": "-th transformer layer and memory state at time " + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "M_{t - 1} \\in \\mathcal{R}^{T_{M} \\times d}" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": ": # tokens, " + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "T_{M}" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": ": # memory states, " + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": ": hidden state dimension). We then obtain " + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "S_{t} = \\text{Attn}(M_{t - 1}, H_{t}, H_{t})" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\text{Attn}(Q, K, V) \\coloneqq \\text{Softmax}\\left(\\frac{QK^{T}}{\\sqrt{d_{q}}}\\right)V" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "d_{q}" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": " is a query dimension (" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 534, + 289, + 653 + ], + "type": "text", + "content": "), with the memory attention mask depicted in the blue box in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "type": "text", + "content": "In particular, we apply a memory mask in the attention operation to select text tokens as memory content (depicted as blue-shaded grid cells in the right of Fig. 3) because including image content as memory could be strong constraint for text-to-image generation (see empirical studies of the" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 82, + 542, + 187 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 539, + 89 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 539, + 89 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 539, + 89 + ], + "type": "text", + "content": "(a) Conventional memory connection (Dai et al., ACL'19/Lei et al., ACL'20) Layer #" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 82, + 542, + 187 + ], + "lines": [ + { + "bbox": [ + 306, + 82, + 542, + 187 + ], + "spans": [ + { + "bbox": [ + 306, + 82, + 542, + 187 + ], + "type": "image", + "image_path": "896d69dc3fe965aaaea659183cc0f3e253e467e30dd31e1970137e125d9668c9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 199, + 542, + 305 + ], + "blocks": [ + { + "bbox": [ + 306, + 189, + 426, + 199 + ], + "lines": [ + { + "bbox": [ + 306, + 189, + 426, + 199 + ], + "spans": [ + { + "bbox": [ + 306, + 189, + 426, + 199 + ], + "type": "text", + "content": "(b) Proposed memory connection" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 199, + 542, + 305 + ], + "lines": [ + { + "bbox": [ + 306, + 199, + 542, + 305 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 542, + 305 + ], + "type": "image", + "image_path": "bed5f34fa3f0b433bf9403ca37139612e6068ddffadc1f83c8bd2baec37415b7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 308, + 547, + 396 + ], + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 396 + ], + "type": "text", + "content": "Figure 4. Comparison of our memory connection scheme to the conventional one [4,15]. We depict the memory connections in the multi-layer Transformer architecture per each sentence input (unfolded in a row). Unlike (a) the previous memory architecture having a single memory path from the immediate past in all layers, (b) ours has a dense connection from all the pasts to be attentively weighted (denoted here by " + }, + { + "bbox": [ + 304, + 308, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\bar{M}" + }, + { + "bbox": [ + 304, + 308, + 547, + 396 + ], + "type": "text", + "content": "'s) in only the last layer. We discuss its empirical benefit in Sec. 4.2.2." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "type": "text", + "content": "proposed mask in the supplementary material). Then we feed the " + }, + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "type": "text", + "content": " to the GRU in order to compute the information to be stored in the memory, " + }, + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "type": "text", + "content": ", propagated to " + }, + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "type": "inline_equation", + "content": "(t+1)" + }, + { + "bbox": [ + 304, + 416, + 546, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 478, + 547, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 547, + 609 + ], + "type": "text", + "content": "Memory connection. Following [4, 15], we stack multiple Transformer's layers, depicted in Fig. 4-(b). Compared to conventional memory updating architectures with serial hidden state connections in all layers (Fig. 4-(a)), our approach only has memory connections in the last layer, which we call it as partial-level memory augmented (PMA). This is because the later layers (closer to the last layer) achieve better representations with higher-level features, abstract and structured representations [28, 29, 41] with improved computational efficiency, which is unlike prior works (see Sec. 4.2.2 and supp for further discussion)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 610, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 610, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 610, + 547, + 658 + ], + "type": "text", + "content": "Additionally, not all historic information is equally important for generating an image at a time step. Similar to the masked self-attention [43], we attentively weight the past information for better modeling of sparse context as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 668, + 518, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 668, + 518, + 683 + ], + "spans": [ + { + "bbox": [ + 315, + 668, + 518, + 683 + ], + "type": "interline_equation", + "content": "\\bar {M} _ {1: (t - 1)} = A t t n (M _ {(t - 1)}, M _ {[ 1: (t - 2) ]}, M _ {[ 1: (t - 2) ]}),", + "image_path": "614cd41960405aaec855ec1ac6990e9896558657f85646e4d49f3bf0ac692578.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 323, + 685, + 545, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 685, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 323, + 685, + 545, + 699 + ], + "type": "interline_equation", + "content": "\\tilde {M} _ {(t - 1)} = \\left[ M _ {t - 1}; \\bar {M} _ {1: (t - 1)} \\right], \\quad (3 \\leq t \\leq 5), \\tag {2}", + "image_path": "11e0e4a57c40449c1894124da892a12036ec8804f5e1d44cf2e92008032b91db.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 340, + 701, + 523, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 701, + 523, + 716 + ], + "spans": [ + { + "bbox": [ + 340, + 701, + 523, + 716 + ], + "type": "interline_equation", + "content": "H _ {t} ^ {l} = \\operatorname {A t t n} \\left(H _ {t} ^ {l}, \\left[ H _ {t} ^ {l}; \\tilde {M} _ {(t - 1)} \\right], \\left[ H _ {t} ^ {l}; \\tilde {M} _ {(t - 1)} \\right]\\right),", + "image_path": "9a0d7f0a6436ad63a7f5ff913259d700f6e0d9c4d2d998b570f4e244dfdf12d9.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3128" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "M_{(t - 1)}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " is a memory at time " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "[1:(t - 2)]" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " refers to the concatenation from time 1 to " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "(t - 2)" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " as depicted in Fig. 4-(b). By doing so, we fuse the contextual information into current hidden state " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "H_{t}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": ". At " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "t = 2" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": ", instead of " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " is only available. We call it as attentively weighted memory (AWM)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 195, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 195, + 164 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 195, + 164 + ], + "type": "text", + "content": "3.2. Online Text-Augmentation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 170, + 288, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 288, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 288, + 254 + ], + "type": "text", + "content": "Vedamtam et al. argue that multiple descriptions for an image help generalization as they address language variations presented in descriptions [44]. Hence, a number of image captioning datasets [10, 20, 37] provide multiple natural language directives, obtained by multiple human annotators, in both training and evaluation splits. But the SV benchmark datasets [17] provide only a single sentence per an image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 254, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 288, + 361 + ], + "type": "text", + "content": "To address the linguistic variations of a text input at inference process, we first consider to generate a pseudo text by a well-trained image-to-text generation model, which we refer it as offline-augmentation [16]. But the offline augmentation generates only a single sentence, which may not provide sufficient diversity. Instead, we propose to generate multiple pseudo-texts and augment them in an online fashion when training our model to increase the diversity. We call it online text augmentation, depicted in Fig. 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 289, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 494 + ], + "type": "text", + "content": "Thanks to our bi-directional multi-modal architecture, we can naturally integrate the process of generating pseudotexts to the process of learning image-to-text model and the text-to-image generation model as depicted in Fig. 5. In the early epochs, less meaningful sentences are generated, but as training progresses, more meaningful sentences are generated (see orange box in Fig. 5). As a side-product, by supervising the model learning with intermediate goals at each time step, we expect to expedite the convergence of learning. When we use the online text augmentation, we can rewrite the objective as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 502, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 502, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 61, + 502, + 287, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {j, p t 2 i} = \\sum_ {k = 1} ^ {n} - \\ln p _ {j} \\left(z _ {k} \\mid \\hat {t} _ {1}, \\dots , \\hat {t} _ {m}, z _ {1}, \\dots , z _ {k - 1}\\right), \\tag {3}", + "image_path": "af27cc7fa89d115bbbc393607874f453b189e01a5c69286e6e7a56639007f9e4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 536, + 226, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 536, + 226, + 550 + ], + "spans": [ + { + "bbox": [ + 78, + 536, + 226, + 550 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {j} = \\mathcal {L} _ {j, t 2 i} + \\lambda_ {1} \\mathcal {L} _ {j, i 2 t} + \\lambda_ {2} \\mathcal {L} _ {j, p t 2 i},", + "image_path": "3a4cb3865181c5a8b66baf1ed3d1009e82ee82ab0a7109929ed3cad85c13bb14.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{j,pt2i}" + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "text", + "content": " is the additional loss with the augmented pseudo-texts and " + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{j,t2i},\\mathcal{L}_{j,i2t}" + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "text", + "content": " are defined in Eq. 1 and " + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "text", + "content": " are balancing hyper-parameters. " + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "inline_equation", + "content": "\\hat{t}" + }, + { + "bbox": [ + 46, + 557, + 288, + 641 + ], + "type": "text", + "content": " means the pseudotext token, predicted during online augmentation without gradient flow. Fig. 6 shows pseudo-texts generated by our method. Detailed training procedure and more examples are in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 164, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 164, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 164, + 685 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 714 + ], + "type": "text", + "content": "Datasets. We use two popular benchmark datasets for evaluating the task of story visualization. Following [23], we use" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 172 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 172 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 172 + ], + "type": "image", + "image_path": "04d01c9788f45c18214ccf229c06ae564bd6c3ce9a6ce0020c44d05e099441e7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 174, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 304, + 174, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 547, + 251 + ], + "type": "text", + "content": "Figure 5. Online text augmentation. We iteratively augment the pseudo-text for better linguistic generalization. (1) generating a pseudo-text at each epoch for each ground-truth image, without gradient flow. (2) generating an image for every epoch. (3,4) learn the text-to-image model with " + }, + { + "bbox": [ + 304, + 174, + 547, + 251 + ], + "type": "inline_equation", + "content": "L_{CE}" + }, + { + "bbox": [ + 304, + 174, + 547, + 251 + ], + "type": "text", + "content": ". In early epochs, it generates less meaningful pseudo-text (thus discard them), but produces pseudotexts matching with the input image as the training progresses." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 261, + 547, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 547, + 310 + ], + "type": "text", + "content": "Pororo-SV dataset for story visualization task without data overlap. Following prior work [22], we use Flintstones-SV which was originally exploited in the text-to-video synthesis task [9]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 320, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 547, + 417 + ], + "type": "text", + "content": "Metrics. For the evaluation metrics, we use Fréchet Inception Distance (FID) (used in [22, 39]) for visual quality of the generated images, character classification score (Char. F1, Frm. Acc.) [17] for character consistency, and global semantic matching (B-2/3, R-Prec.) [22, 23] between generated image sequence and story paragraph.. We elaborate on the details of the datasets and implementation in the supplementary material for the sake of space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 547, + 511 + ], + "type": "text", + "content": "Baselines. We compare our method to a number of prior arts including state-of-the-art story visualization methods such as StoryGAN [17], CPCSV [39], DuCo-StoryGAN [23], VLC-StoryGAN [22] and VP-CSV [2] for both quantitative and qualitative analysis in the story visualization setting. Moreover, we compare [24] in the story continuation setting." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 521, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 548, + 714 + ], + "type": "text", + "content": "Implementation details. To train our model, we first tokenize the text and image inputs. Particularly, the image inputs are encoded by the encoder of the VQ-VAE [42] into image tokens and decoded by the decoder of the VQ-VAE for generating image sequence. Once the inputs are tokenized, we append two special embedding tokens to the text and image tokens. These tokens signify the beginning of each modality, denoted as 'SOS' (start of sentence) and 'SOI' (start of image). Subsequently, we add two embeddings: (1) a positional embedding to indicate absolute position of each token and (2) a segment embedding to distinguish source and target modality, e.g., for text-to-image generation task, the source corresponds to the text while the target corresponds to the image. Through the utilization of segment embeddings to distinguish between source and target, we enable the model to generate target data from source data" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3129" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 521, + 232 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 521, + 232 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 521, + 232 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 521, + 232 + ], + "type": "image", + "image_path": "50f14c46cbb54df636f9d29a4eaf74fcb6fb20a0475a951b141c0b65cbd9d59b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 235, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 235, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 547, + 257 + ], + "type": "text", + "content": "Figure 6. Generated pseudo-texts by the proposed method during training. We depict the generated pseudo-texts in training procedure. We utilize them as supplementary supervisions to address linguistic diversity of a sentence at inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": "within a unified architecture, as illustrated in Fig. 2. Here, we set the token length for each modality as " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "T_{text} = 80" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "T_{image} = 256" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": "), the hidden dimension size to 512, the number of transformer layer to 6, and the number of attention heads to 16, thereafter the number of trainable parameters approximately 93.7-M as shown in Tab. 1. Particularly, we generate an image per each sentence without any sampling process. We train the CMOTA using AdamW [21] optimizer with " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.95" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon = 1e - 8" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": ", weight decay multiplier with " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "1e - 2" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": " and learning rate with " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "4.5e - 6" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": " multiplied by batch size. Moreover, we set the hyper-parameters for loss balancing in Eq. 3 as " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 1.0" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.5" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": " and the number of memory state, " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "T_{M}" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": " as 1. Finally, to generate a high-resolution image, we modified the VQ-VAE with the additional trainable parameters (i.e., 7.1-M shown in Tab. 1). This modification allows for the generation of high-resolution images, increasing the image resolution from " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 493, + 171, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 493, + 171, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 171, + 506 + ], + "type": "text", + "content": "4.2. Quantitative Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 513, + 242, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 242, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 242, + 525 + ], + "type": "text", + "content": "4.2.1 Comparison with the State of the Arts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": "In Table 1, we summarize the performance of the prior arts and the proposed CMOTA in Pororo-SV [17] and Flintstones-SV [9] dataset in the various metrics. By default, all methods generate " + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": " images along with the higher resolution " + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 533, + 287, + 616 + ], + "type": "text", + "content": " images. CMOTA outperforms existing methods by a large margin in both benchmarks in most of metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "The very recent method VP-CSV [2] performs on par to our CMOTA while ours still outperforms in FID, BLEU and R-precision, implying that CMOTA generates high quality images maintaining global semantic matching between story paragraph and images better than the VP-CSV. The lower performance than the VP-CSV in Char. F1 score and Frm. Acc. is attributed to the specialized character-centric module that only focuses the model to generate accurate characters" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 266, + 547, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 315 + ], + "type": "text", + "content": "for each image at the expense of the other performance metrics. In addition, our CMOTA-HR model even outperforms the VP-CSV [2] by large margin without the specialized character-centric module." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 328, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 460 + ], + "type": "text", + "content": "Low resolution to high resolution. Not surprisingly, CMOTA-HR outperforms CMOTA with default (low-) resolution (i.e., " + }, + { + "bbox": [ + 304, + 328, + 547, + 460 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 328, + 547, + 460 + ], + "type": "text", + "content": ") in almost all metrics. Interestingly, however, VLC-StoryGAN-HR [22] shows similar performance to " + }, + { + "bbox": [ + 304, + 328, + 547, + 460 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 328, + 547, + 460 + ], + "type": "text", + "content": " model in Frame accuracy, BLEU, R-precision, and drops performance in all other metrics. For the increased FID scores by the VLC-StoryGAN when goes to high resolution, we hypothesize that the constituency parse trees of input sentences in the VLC-StoryGAN increases the difficulty in generating visual details as the images contain more visual details in high resolution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "content": "In addition, we observe that the performance drop of FID in both our CMOTA and VLC, when goes to high resolution. It is because as high-resolution image requires the model to depict fine-grained details, generation task would be more difficult compared to the low-resolution image generation. In particular, the drop of FID in Flintstones-SV is larger than that in Pororo-SV. It is because there are more complicated visual details in the Flintstones-SV than the Pororo-SV dataset. Note that the FID drops by our method from low-resolution to high resolution is less than the VLC in Pororo-SV. Furthermore, in terms of computational cost, our CMOTA gains performance boost by using high-resolution effectively with just adding relatively small computational cost on top of original " + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 460, + 547, + 628 + ], + "type": "text", + "content": " model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "type": "text", + "content": "Character consistency. Although our method does not explicitly enforce the character semantic preservation, the proposed memory delivers the semantic across the sentence implicitly. We here investigate the character context preservation performance by comparing our method to [22, 23] on per-character classification F1-score on test split of Pororo" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3130" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 533, + 237 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 533, + 237 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 533, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 533, + 237 + ], + "type": "table", + "html": "
DatasetResolutionMethods# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
Pororo-SV64 × 64StoryGAN [17]-158.0618.599.343.24 / 1.221.51 ± 0.15
CP-CSV [39]-140.2421.7810.033.25 / 1.221.76 ± 0.04
DuCo-StoryGAN [23]101M96.5138.0113.973.68 / 1.343.56 ± 0.04
VLC-StoryGAN [22]100M84.9643.0217.363.80 / 1.443.28 ± 0.00
VP-CSV [2]-65.5156.8425.874.45 / 1.806.95 ± 0.00
CMOTA (Ours)96.6M52.1353.2524.724.58 / 1.907.34 ± 0.03
128 × 128VLC-StoryGAN-HR†[22]102.6M97.0840.3617.173.89 / 1.583.47 ± 0.03
CMOTA-HR (Ours)103.7M52.7758.8628.895.45 / 2.3416.36 ± 0.05
Flinstones-SV64 × 64StoryGAN [17]-127.1946.2032.9613.87 / 7.831.72 ± 0.18
DuCo-StoryGAN [23]101M78.0254.9236.3415.48 / 9.172.64 ± 0.17
VLC-StoryGAN‡[22]100M72.8758.8139.18--
CMOTA (Ours)96.6M36.7179.7466.0119.85 / 12.9810.50 ± 0.35
128 × 128CMOTA-HR (Ours)103.7M54.8186.4474.0622.18 / 15.1725.71 ± 0.70
", + "image_path": "d70b33be9327ff9e0c7bbf0048cec972c3d50cf94273b94683246068411cd1e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 240, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 547, + 298 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with the state of the arts. On the test split of Pororo-SV and Flintstones-SV. # Params. refers to the number of trainable parameters. Char. F1 refers to character F1 score. Frm. Acc. refers to frame accuracy. R-Prec. refers to R-Precision. ↓ indicates 'lower the better' and ↑ indicates 'higher the better'. Experiments are done in both " + }, + { + "bbox": [ + 46, + 240, + 547, + 298 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 240, + 547, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 240, + 547, + 298 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 240, + 547, + 298 + ], + "type": "text", + "content": " resolutions. 'HR' refers to its high resolution " + }, + { + "bbox": [ + 46, + 240, + 547, + 298 + ], + "type": "inline_equation", + "content": "(128 \\times 128)" + }, + { + "bbox": [ + 46, + 240, + 547, + 298 + ], + "type": "text", + "content": " version. † indicates our results with author's implementation (https://github.com/adymaharana/VLCStoryGan).‡ indicates the absolute values that we computed by the given relative values in [22]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 309, + 287, + 430 + ], + "blocks": [ + { + "bbox": [ + 48, + 309, + 287, + 430 + ], + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 430 + ], + "type": "image", + "image_path": "2e8a7fb8bd7ea9252edc4ce8f6894437c847fffec2e2857fa70e8a47d1aa3b63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 433, + 288, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 433, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 288, + 510 + ], + "type": "text", + "content": "Figure 7. Per-character Classification F1-score on the Test Split of Pororo-SV Dataset. Normalized character (Char.) frequency is shown in red line. It represents the relative number of characters compared to the count of Pororo in training dataset. While DuCo/VLC-StoryGAN shows degraded performance as the character frequency decreases, CMOTA outperforms them across all the characters." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 519, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 555 + ], + "type": "text", + "content": "SV dataset and normalized character frequency in train split dataset as illustrated in Fig. 7. Especially, it is interesting to compare with [2] as it is explicitly proposed for that purpose." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 555, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 288, + 628 + ], + "type": "text", + "content": "As shown in the figure, our method outperforms prior arts [22, 23] by large margins. Although we did not have any special module for character information preserving, unlike [2], our model maintains such information over multiple sentences, even outperforming the image quality denoted as FID by " + }, + { + "bbox": [ + 46, + 555, + 288, + 628 + ], + "type": "inline_equation", + "content": "-13.38" + }, + { + "bbox": [ + 46, + 555, + 288, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 646, + 151, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 151, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 151, + 657 + ], + "type": "text", + "content": "4.2.2 Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "Benefit of the proposed architectural components. To investigate the benefit of each proposed architectural component, we build up the full model with the components from the vanilla, single-directional transformer without memory" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 311, + 408, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 311, + 408, + 321 + ], + "spans": [ + { + "bbox": [ + 305, + 311, + 408, + 321 + ], + "type": "text", + "content": "as summarized in Tab. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 323, + 548, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 548, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 548, + 502 + ], + "type": "text", + "content": "Using the proposed partial-level connected memory architecture only with the vanilla transformer (Tr.), the generation performance improves in all metrics compared to the Tr. Also, if we compare it with all-level connected memory architecture [15], PMA shows better performance in all metrics as in Tab. 3 (see supplementary material for more experimental analysis for various memory connection schemes). Employing attentional weighted memory (AWM) improves performance overall with slight degradation in FID (+0.12). The bi-directional learning scheme (Bi.) improves performance in all metrics except the character classification related metrics, i.e., Char.F1 and Frm. Acc. It is because there are captions that does not contain characters' name, just mentioning characters as 'friends', making the model difficult to generate specific characters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 512, + 548, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 548, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 548, + 656 + ], + "type": "text", + "content": "Benefit of the text augmentation. By augmenting the generated text in offline manner (i.e., using an image captioning model pretrained on the dataset of interest), we observe overall performance increase in all metrics. We hypothesize its reason as that the pseudo-text may convey contextual information such as characters' existence or description of background, thereby making the generation model to be robust to language variation as shown in a first row of Tab. 4. By the online augmentation of pseudo texts (i.e., gradually updating text generator, thus generating a diverse and gradually varying set of pseudo-texts), the performance further improves, as shown in Tab. 2 and Tab. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 670, + 501, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 670, + 501, + 682 + ], + "spans": [ + { + "bbox": [ + 305, + 670, + 501, + 682 + ], + "type": "text", + "content": "4.2.3 Comparison with Large-Scale Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Despite the unfairness of comparison to the large-scale models due to the model size, we compare CMOTA to larger" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3131" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 518, + 162 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 518, + 162 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 518, + 162 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 518, + 162 + ], + "type": "table", + "html": "
Methods# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
VLC-StoryGAN [22]100M84.9643.0217.363.28 / 1.443.28
Single Directional Transformer (Tr.)93.7M63.8845.4818.444.18 / 1.695.67
+ Partial-level Memory Connection (PMA)95.8M59.0549.7221.794.41 / 1.776.28
+ Attentional Weighted Memory (AWM)96.6M56.9850.1522.014.50 / 1.916.79
+ Bi-directional Training (Bi.)96.6M54.6951.2722.154.52 / 1.847.12
+ Online Augmentation (Full model)96.6M52.1353.2524.724.58 / 1.907.34
", + "image_path": "db2291dadd07fd827f99f5dd46b03f11167a0f062e87c60e0a0ba3141f9384dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 222, + 286, + 251 + ], + "blocks": [ + { + "bbox": [ + 46, + 166, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 547, + 210 + ], + "type": "text", + "content": "Table 2. Benefit of the proposed components. On the test split of Pororo-SV Dataset. As a baseline, we use a single directional transformer-based text-to-image generation model. We gradually add each of proposed components to investigate its effect for improving performance, i.e., partial-level memory connection (Sec. 3.1), attentively weighted memory (Sec. 3.1), bi-directional training (Sec 3.2), and online-augmentation (Sec. 3.2)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 222, + 286, + 251 + ], + "lines": [ + { + "bbox": [ + 48, + 222, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 286, + 251 + ], + "type": "table", + "html": "
Memory Architecture# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
All-Level Connection [15]118M61.2347.2119.214.21 / 1.726.08
Partial-level Connection (Ours)95.8M59.0549.7221.794.41 / 1.776.28
", + "image_path": "7ccea10ab79183a8f57bc2047a8707d02a0b417812a509d14e887a2e04e12a88.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 311, + 286, + 343 + ], + "blocks": [ + { + "bbox": [ + 46, + 257, + 288, + 302 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 288, + 302 + ], + "type": "text", + "content": "Table 3. Benefit of the proposed memory connection scheme. Proposed scheme outperforms conventional memory connection that uses all-level connections [15] (Fig. 4-(a)) with less number of parameters (test split of Pororo-SV)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 311, + 286, + 343 + ], + "lines": [ + { + "bbox": [ + 48, + 311, + 286, + 343 + ], + "spans": [ + { + "bbox": [ + 48, + 311, + 286, + 343 + ], + "type": "table", + "html": "
Augmentation# Params.FID↓Char. F1↑Frm. Acc.↑BLEU-2/3↑R-Prec.↑
Offline96.6M54.5151.3222.314.50 / 1.907.09
Online (Ours)96.6M52.1353.2524.724.58 / 1.907.34
", + "image_path": "707442da1c3ea159cf1074d80e9bcef2d2e7f93220a5f70c3ad02f227fca328a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "text", + "content": "models [24] in Tab. 5. it is surprising to see that CMOTA outperforms on an evaluation metric of character classification (i.e., Char.F1 and Frm. Acc.), regardless of using prompt-tuning or fine-tuning of StoryDALL-E." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 451, + 288, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 288, + 523 + ], + "type": "text", + "content": "But as expected, the image quality denoted as FID is worse than the StoryDALL-E thanks to the huge size of the model and training dataset. Nevertheless, FID score is slightly better (Pororo-SV) and on-par (Flintstones-SV) compared to 'prompt-tune' that updates " + }, + { + "bbox": [ + 46, + 451, + 288, + 523 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 451, + 288, + 523 + ], + "type": "text", + "content": " of parameters [24]. We discuss more about this in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 536, + 188, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 188, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 188, + 548 + ], + "type": "text", + "content": "4.3. Human Preference Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 556, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 663 + ], + "type": "text", + "content": "We conduct a larger scale (than the prior arts) human survey using the Amazon Mechanical Turk platform to qualitatively compare the generated image sequence by our method over VLC-StoryGAN [22] on three criteria, i.e., visual quality, temporal consistency, and semantic relevancy between generated images and descriptions, following [17, 22, 23, 39]. Unlike the previous arts that employ 2 to 9 human judges [2, 17, 22, 23, 39], we employ 100 judges for statistically more reliable results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "As shown in the top table in Tab. 6, it is clear that human subjects prefer the generated images by our model over those from VLC-StoryGAN [22] on winning ratio(\\%). We discuss more details about this study in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 222, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 46, + 350, + 288, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 288, + 384 + ], + "type": "text", + "content": "Table 4. Benefit of the online augmentation. On the Pororo-SV test split. The proposed online augmentation outperforms the offline augmentation using a pretrained captioner [15]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 222, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 309, + 222, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 309, + 222, + 544, + 312 + ], + "type": "table", + "html": "
Methods# Params.FID↓Char. F1↑Frm. Acc↑
Porre-SVStoryDALL-E (prompt-tune)1.3B61.2329.6811.65
StoryDALL-E (fine-tune)1.3B25.9036.9717.26
MEGA-StoryDALL-E (fine-tune)2.8B23.4839.9118.01
CMOTA (Ours)96.6M55.2651.4822.73
Flinstones-SVStoryDALL-E (prompt-tune)1.3B53.7142.3832.54
StoryDALL-E (fine-tune)1.3B26.4973.4355.19
MEGA-StoryDALL-E (fine-tune)2.8B23.5874.2654.68
CMOTA (Ours)96.6M58.5979.7562.98
", + "image_path": "a5dc334913ec539608d0fb890250f80b34043a7cb60e8a282df43eddd0307997.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 308, + 428, + 544, + 509 + ], + "blocks": [ + { + "bbox": [ + 304, + 319, + 547, + 418 + ], + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 418 + ], + "type": "text", + "content": "Table 5. Quantitative comparisons to large scale models. On the test split of Pororo-SV and Flintstones-SV dataset. All models are implemented on top of the large pretrained transformer (i.e., StoryDALL-E [24] pretrained on 14 million and MEGAStoryDALL-E [24] pretrained on 15 million from Conceptual Caption dataset [36]). In [24], the 'prompt-tune' update " + }, + { + "bbox": [ + 304, + 319, + 547, + 418 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 319, + 547, + 418 + ], + "type": "text", + "content": " of model parameters compared to full fine-tune'. Without pretraining, we use a much small model compared to the proir arts, i.e., 96.6M (CMOTA) vs. 1.3B (StoryDALL-E)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 428, + 544, + 509 + ], + "lines": [ + { + "bbox": [ + 308, + 428, + 544, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 544, + 509 + ], + "type": "table", + "html": "
ResolutionAttributeVLC-SG [22]TieOurs
64 × 64 (Low-Res.)Visual Quality27.8%8.6%63.6%
Temproal Consistency24.3%8.3%59.0%
Semantic Relevance31.2%10.8%57.9%
128 × 128 (High-Res.)Visual Quality21.9%1.5%76.6%
Temporal Consistency21.8%2.5%75.7%
Semantic Relevance23.6%1.7%74.6%
", + "image_path": "c54a5ca024a0a5c644ae22248274be7ec0410ee936b004ed3a833a402ae2ab9f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 308, + 516, + 544, + 548 + ], + "blocks": [ + { + "bbox": [ + 308, + 516, + 544, + 548 + ], + "lines": [ + { + "bbox": [ + 308, + 516, + 544, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 516, + 544, + 548 + ], + "type": "table", + "html": "
CMOTA w/o Mem.TieCMOTA
Temporal Consistency34.5%4.3%61.2%
", + "image_path": "0e463c1abe7075372df2f4de2590212a883f82bfa7ab0570fa2a968d4c1856bd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 555, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 655 + ], + "type": "text", + "content": "Table 6. Human preference studies. (Top) With 100 judges in the Amazon Mechanical Turk, on Pororo-SV test split dataset. Win " + }, + { + "bbox": [ + 304, + 555, + 545, + 655 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 304, + 555, + 545, + 655 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 304, + 555, + 545, + 655 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 304, + 555, + 545, + 655 + ], + "type": "text", + "content": " times one model is preferred over the others. \"Tie\" refers to the same. Ours are clearly preferred over the VLC-SG (refers to VLC-StoryGAN) [22]. (Bottom) For the model using our memory (Mem.) module, contrasting results with and without the memory module that considers the contexts spread across the sentences, our method significantly improves human preference in temporal consistency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Human preference for memory usage. We further investigate the benefit of using the proposed memory module (Sec. 3.1) by the human study with 100 annotators in Amazon Mechanical Turk. For the comparative analysis between" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3132" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "CMOTA and CMOTA-w/o-Memory, we use 30 randomly sampled image sequences on Pororo-SV test set to conduct A/B test following [39]. Particularly, the annotators need to consider the temporal consistency with semantic relevance between generated image sequence and captions. As depicted in the bottom table in Tab. 6, we observe that " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "61.2\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " of annotators prefer our CMOTA with memory module over the one without it. We can conclude that the memory module has a great deal in creating temporally consistent image sequence with respect to semantic relevancy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 201, + 165, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 165, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 165, + 213 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 219, + 288, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 288, + 351 + ], + "type": "text", + "content": "We now qualitatively analyze the quality of generated image sequences in Fig. 8 on the Pororo-SV dataset's test split. The top row shows the image sequence of ground truth, the two rows (2-3) contain prior works [17,23] and the final row is the image sequence generated by CMOTA and its high resolution version (CMOTA-HR). CMOTA generates a semantically more plausible image sequence with better visual quality, compared to prior works [22, 23]. Particularly, the CMOTA-HR demonstrates the ability to generate visual details more effectively even as the size of the image increases, while maintaining similar semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 351, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 290, + 483 + ], + "type": "text", + "content": "Furthermore, we qualitatively investigate the advantage of using memory architecture and summarize the results in Fig. 9. As shown in the figure, the proposed memory architecture generates a semantically more plausible image sequence with proper context, e.g., background; the CMOTA without memory fails to capture proper background context (shown in dotted red box) since a single sentence could be interpreted in many ways. In contrast, CMOTA (with the context memory) generates an image sequence with plausible background with characters preserved without an explicit character-centric model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 495, + 119, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 119, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 119, + 507 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 515, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 288, + 624 + ], + "type": "text", + "content": "We propose to better encode semantic context, e.g., plausible background and characters, for story visualization using a new memory architecture in a multi-modal bi-directional Transformer. We further propose an online text augmentation training scheme to generate pseudo-text descriptions as an intermediate supervision for addressing linguistic diversity in the texts at inference. The proposed method generates a temporally coherent and semantically relevant image sequence for each sentence in the given text paragraph." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 624, + 288, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 288, + 719 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 288, + 719 + ], + "type": "text", + "content": "The proposed method outperforms prior works by a large margin on various metrics on the two popular SV benchmark datasets, and also outperforms some of hyper-scale models in multiple semantic understanding metrics. Although computationally prohibited and the pre-trained models are not publicly available, it would be intriguing to apply the proposed memory module and the online augmentation scheme to a large model." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 539, + 328 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 539, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 539, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 539, + 328 + ], + "type": "image", + "image_path": "707bec45ef572e62b2f3e99daa00148fa57ea05e04ad390e3f96aebc6e8b8ff2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 329, + 548, + 396 + ], + "lines": [ + { + "bbox": [ + 305, + 329, + 548, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 548, + 396 + ], + "type": "text", + "content": "Figure 8. Qualitative results compared to prior arts. On the Pororo-SV's test split. We compare our CMOTA to the prior arts including DuCo-StoryGAN and VLC-StoryGAN. Ours generates a semantically more plausible and temporally coherent image sequence compared to the prior arts. All images except CMOTA-HR " + }, + { + "bbox": [ + 305, + 329, + 548, + 396 + ], + "type": "inline_equation", + "content": "(128\\times 128)" + }, + { + "bbox": [ + 305, + 329, + 548, + 396 + ], + "type": "text", + "content": " are generated with resolution of " + }, + { + "bbox": [ + 305, + 329, + 548, + 396 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 305, + 329, + 548, + 396 + ], + "type": "text", + "content": " for comparison." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 413, + 541, + 589 + ], + "blocks": [ + { + "bbox": [ + 313, + 413, + 541, + 589 + ], + "lines": [ + { + "bbox": [ + 313, + 413, + 541, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 541, + 589 + ], + "type": "image", + "image_path": "37b9e587e9dd43b4f188deca6e4b7bfb988c0d7b4421954911bb2ffc4745a3c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 590, + 547, + 647 + ], + "lines": [ + { + "bbox": [ + 305, + 590, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 590, + 547, + 647 + ], + "type": "text", + "content": "Figure 9. Benefit of memory module in encoding background context. Our CMOTA generates both semantically and visually plausible image sequences. Without memory module, inconsistent images (i.e., abrupt background change) are generated as indicated by the red dotted box." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgement. This work is partly supported by the NRF grant (No.2022R1A2C4002300) " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " , IITP grants (No.2020-0-01361, AI GS Program (Yonsei University) " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "5 \\%" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " , No.2021-0-02068, AI Innovation Hub " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "5 \\%" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " 2022-0-00077 " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " , 2022-0-00113 " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " , 2022-0-00959 " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " , 2022-0-00871 " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " , 2022-0-00951 " + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": " ) funded by the Korea government (MSIT)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3133" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 134 + ], + "type": "text", + "content": "[1] Alexander Buslaev, Vladimir I. Iglovikov, Eugene Khvedchenya, Alex Parinov, Mikhail Druzhinin, and Alexandr A. Kalinin. Albumimations: Fast and flexible image augmentations. Information, 11(2), 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 168 + ], + "type": "text", + "content": "[2] Hong Chen, Rujun Han, Te-Lin Wu, Hideki Nakayama, and Nanyun Peng. Character-centric story visualization via visual planning and token alignment. ArXiv, 2022. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 202 + ], + "type": "text", + "content": "[3] Mark Chen, Alec Radford, Jeff Wu, Heewoo Jun, Prafulla Dhariwal, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "type": "text", + "content": "[4] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov. Transformer-XL: Attentive language models beyond a fixed-length context. In ACL, 2019. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 246, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 301 + ], + "type": "text", + "content": "[5] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. In Advances in Neural Information Processing Systems, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 301, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 288, + 335 + ], + "type": "text", + "content": "[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. ArXiv, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "type": "text", + "content": "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 422 + ], + "type": "text", + "content": "[8] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 424, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 456 + ], + "type": "text", + "content": "[9] Tanmay Gupta, Dustin Schwenk, Ali Farhadi, Derek Hoiem, and Aniruddha Kembhavi. Imagine this! scripts to compositions to videos. In ECCV, 2018. 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 457, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 457, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 288, + 512 + ], + "type": "text", + "content": "[10] Danna Gurari, Qing Li, Chi Lin, Yinan Zhao, Anhong Guo, Abigale Stangl, and Jeffrey P. Bigham. Vizwiz-priv: A dataset for recognizing the presence and purpose of private visual information in images taken by blind people. In CVPR, June 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 546 + ], + "type": "text", + "content": "[11] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. ArXiv, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 590 + ], + "type": "text", + "content": "[12] Philip T. G. Jackson, Amir Atapour-Abarghouei, Stephen Bonner, T. Breckon, and Boguslaw Obara. Style augmentation: Data augmentation via style randomization. In CVPR Workshops, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "type": "text", + "content": "[13] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. IEEE Access, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 669 + ], + "type": "text", + "content": "[14] Taehoon Kim, Gwangmo Song, Sihaeng Lee, Sangyun Kim, Yewon Seo, Soonyoung Lee, Seung Hwan Kim, Honglak Lee, and Kyunghoon Bae. L-verse: Bidirectional generation between image and text. In CVPR, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[15] Jie Lei, Liwei Wang, Yelong Shen, Dong Yu, Tamara L Berg, and Mohit Bansal. Mart: Memory-augmented recurrent transformer for coherent video paragraph captioning. In ACL, 2020. 3, 4, 7, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 152 + ], + "type": "text", + "content": "[17] Yitong Li, Zhe Gan, Yelong Shen, Jingjing Liu, Yu Cheng, Yuexin Wu, Lawrence Carin, David Carlson, and Jianfeng Gao. Storygan: A sequential conditional gan for story visualization. In CVPR, 2019. 1, 2, 3, 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 547, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 186 + ], + "type": "text", + "content": "[18] Yitong Li, Martin Renqiang Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In AAAI, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 243 + ], + "type": "text", + "content": "[19] Swee Kiat Lim, Yi Loo, Ngoc-Trung Tran, Ngai-Man Cheung, Gemma Roig, and Yuval Elovici. Doping: Generative data augmentation for unsupervised anomaly detection with gan. 2018 IEEE International Conference on Data Mining (ICDM), pages 1122–1127, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 243, + 547, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 547, + 287 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 289, + 546, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 546, + 312 + ], + "type": "text", + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 312, + 547, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 547, + 346 + ], + "type": "text", + "content": "[22] Adyasha Maharana and Mohit Bansal. Integrating visuospatial, linguistic and commonsense structure into story visualization. In EMNLP, 2021. 2, 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 346, + 547, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 346, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 346, + 547, + 390 + ], + "type": "text", + "content": "[23] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Improving generation and evaluation of visual stories via semantic consistency. In NAACL-HLT, 2021. 2, 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 392, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 392, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 392, + 546, + 426 + ], + "type": "text", + "content": "[24] Adyasha Maharana, Darryl Hannan, and Mohit Bansal. Storydall-e: Adapting pretrained text-to-image transformers for story continuation. ArXiv, 2022. 2, 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 426, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 547, + 470 + ], + "type": "text", + "content": "[25] Gaurav Mittal, Tanya Marwah, and Vineeth N. Balasubramanian. Sync-draw: Automatic video generation using deep recurrent attentive architectures. In ACM Multimedia, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 472, + 547, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 547, + 516 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 547, + 516 + ], + "type": "text", + "content": "[26] Saypraseuth Mounsaveng, Issam Laradji, Ismail Ben Ayed, David Vazquez, and Marco Pedersoli. Learning data augmentation with online bilevel optimization for image classification. arXiv, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 517, + 546, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 517, + 546, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 517, + 546, + 551 + ], + "type": "text", + "content": "[27] Yingwei Pan, Zhaofan Qiu, Ting Yao, Houqiang Li, and Tao Mei. To create what you tell: Generating videos from captions. In ACM Multimedia, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 552, + 546, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 552, + 546, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 552, + 546, + 575 + ], + "type": "text", + "content": "[28] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, 2019. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 576, + 546, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 546, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 546, + 609 + ], + "type": "text", + "content": "[29] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 610, + 546, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 610, + 546, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 610, + 546, + 644 + ], + "type": "text", + "content": "[30] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Learn, imagine and create: Text-to-image generation from prior knowledge. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 646, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 546, + 678 + ], + "type": "text", + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv, 2022. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 679, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 546, + 713 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2, 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3134" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 708 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "text", + "content": "[33] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "type": "text", + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 205 + ], + "type": "text", + "content": "[35] Meet Shah, Xinlei Chen, Marcus Rohrbach, and Devi Parikh. Cycle-consistency for robust visual question answering. CVPR, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 207, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 207, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 207, + 288, + 249 + ], + "type": "text", + "content": "[36] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 251, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 288, + 283 + ], + "type": "text", + "content": "[37] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 285, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 288, + 338 + ], + "type": "text", + "content": "[38] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 341, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 341, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 48, + 341, + 288, + 372 + ], + "type": "text", + "content": "[39] Yun-Zhu Song, Zhi Rui Tam, Hung-Jen Chen, Huiao-Han Lu, and Hong-Han Shuai. Character-preserving coherent story visualization. In ECCV, 2020. 2, 5, 7, 8, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 374, + 288, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 374, + 288, + 416 + ], + "spans": [ + { + "bbox": [ + 48, + 374, + 288, + 416 + ], + "type": "text", + "content": "[40] Zhiqiang Tang, Yunhe Gao, Leonid Karlinsky, Prasanna Sattigeri, Rogerio Feris, and Dimitris Metaxas. Onlineaugment: Online data augmentation with less domain knowledge. arXiv, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 418, + 288, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 288, + 440 + ], + "type": "text", + "content": "[41] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 441, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 441, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 288, + 463 + ], + "type": "text", + "content": "[42] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In NIPS, 2017. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 464, + 288, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 288, + 507 + ], + "type": "text", + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 509, + 288, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 288, + 540 + ], + "type": "text", + "content": "[44] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, June 2015. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 542, + 288, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 542, + 288, + 596 + ], + "spans": [ + { + "bbox": [ + 48, + 542, + 288, + 596 + ], + "type": "text", + "content": "[45] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 598, + 288, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 598, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 48, + 598, + 288, + 639 + ], + "type": "text", + "content": "[46] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Generating open-domain videos from natural descriptions, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 642, + 288, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 642, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 48, + 642, + 288, + 673 + ], + "type": "text", + "content": "[47] Chenfei Wu, Jian Liang, Lei Ji, F. Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pre-training for neural visual world creation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 675, + 288, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 675, + 288, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 675, + 288, + 708 + ], + "type": "text", + "content": "[48] Zhiyue Wu, Yijie Wang, Haibo Mi, Hongzuo Xu, Wei Zhang, and Lanlan Feng. Oada: An online data augmentation method for raw histopathology images. ICONIP, 2021. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3135" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/246a74f2-93a7-4d68-8843-6238c9108476_content_list.json b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/246a74f2-93a7-4d68-8843-6238c9108476_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0e701c35e9ce4ed2677994be8d004044c1c5e1e7 --- /dev/null +++ b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/246a74f2-93a7-4d68-8843-6238c9108476_content_list.json @@ -0,0 +1,1837 @@ +[ + { + "type": "text", + "text": "Strata-NeRF : Neural Radiance Fields for Stratified Scenes", + "text_level": 1, + "bbox": [ + 186, + 130, + 782, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ankit Dhiman $^{1,2}$ R Srinath $^{1}$ Harsh Rangwani $^{1}$ Rishubh Parihar $^{1}$", + "bbox": [ + 218, + 179, + 779, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lokesh R Boregowda2 Srinath Sridhar3 R Venkatesh Babu1", + "bbox": [ + 243, + 198, + 751, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Vision and AI Lab, IISc Bangalore $^{2}$ Samsung R & D Institute India - Bangalore $^{3}$ Brown University", + "bbox": [ + 86, + 215, + 913, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Field (NeRF) approaches learn the underlying 3D representation of a scene and generate photorealistic novel views with high fidelity. However, most proposed settings concentrate on modelling a single object or a single level of a scene. However, in the real world, we may capture a scene at multiple levels, resulting in a layered capture. For example, tourists usually capture a monument's exterior structure before capturing the inner structure. Modelling such scenes in 3D with seamless switching between levels can drastically improve immersive experiences. However, most existing techniques struggle in modelling such scenes. We propose Strata-NeRF, a single neural radiance field that implicitly captures a scene with multiple levels. Strata-NeRF achieves this by conditioning the NeRFs on Vector Quantized (VQ) latent representations which allow sudden changes in scene structure. We evaluate the effectiveness of our approach in multi-layered synthetic dataset comprising diverse scenes and then further validate its generalization on the real-world RealEstate10K dataset. We find that Strata-NeRF effectively captures stratified scenes, minimizes artifacts, and synthesizes high-fidelity views compared to existing approaches. https://ankitatiisc.github.io/Strata-NeRF/", + "bbox": [ + 75, + 300, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 678, + 209, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis is an ill-posed problem widely encountered in various areas such as augmented reality [24, 28], virtual reality [11], etc. A paradigm change for solving these kinds of problems was brought by the introduction of Neural Radiance Fields (NeRF) [34]. NeRFs are neural networks that take in the spatial coordinates and camera parameters as input and output the corresponding radiance field. Earlier version of NeRFs enable the generation of high-fidelity novel views for bounded scenes, significantly improving over existing techniques like Structure From Motion [47]. Further, the capability of NeRFs have been recently extended to model unbounded scenes by Mip-NeRF 360 [2]. This enabled NeRFs to model complex real-world", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa84ea35264191c360c3f75145a50d663499eca8f61cd710e513babfd5d20007.jpg", + "image_caption": [ + "Figure 1. Top, wireframe view of a multi-layered stratified scene with three levels (monkey head inside sphere inside a cube). The camera colors indicate views of a specific level. Strata-NeRF enables high-quality reconstruction of such stratified scenes using a single neural network." + ], + "image_footnote": [], + "bbox": [ + 522, + 271, + 870, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scenes, where the scene content can exist at any distance from the camera.", + "bbox": [ + 496, + 643, + 890, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, similar to unboundedness in scenes, hierarchies in scenes are also natural. For example, images captured in a house can be categorized into images captured outside and inside across various rooms. Modelling such hierarchical scenes jointly for all levels through a NeRF could be particularly useful in cases of Virtual Reality applications. As it would not require switching to a different NeRF for each level, reducing memory requirement and latency in switching. Further, as the different hierarchies of a scene usually share texture and architectural commonalities, it could lead to effective knowledge sharing and reduce the requirement of training independent models. For tackling the above novel objective, we introduce a paradigm of scenes that can be deconstructed into several tiers, termed \"Stratified Scenes\". A \"stratified\" scene has several levels or", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17603", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a1547f58edfe6ef92fa832071ab103ef30211b20adf9dc42edb6a8d75c33ca3.jpg", + "image_caption": [ + "Figure 2. Novel views for stratified scene in Figure 1, from Mip-NeRF 360 [2] (left) and our method \"Strata-NeRF\" (right). Existing methods struggle to capture stratified scenes with a single network while ours produces sharp results." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 467, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "groupings of structure (Figure 1). In our work, we first propose a synthetic dataset of stratified scenes, i.e. scenes having multiple levels. This dataset comprises scenes from two categories: (i) Simpler geometry, such as spheres, cubes, or tetrahedron meshes, and (ii) Complex geometry, which closely emulates a real-world setup.", + "bbox": [ + 75, + 321, + 468, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On such datasets, we find methods such as Mip-NeRF 360 perform well for each level of the hierarchy independently, but produce unsatisfactory results when images from all hierarchical levels are used together for training (Figure 3). This can be attributed to the continuous nature of NeRFs, which is unsuitable for modelling the sudden changes in scenes with shifts in hierarchical levels. Hence, in this work, we introduce Strata-NeRF that explicitly aims to model the hierarchies by conditioning [22, 38, 39, 66, 43] the NeRF on Vector Quantized (VQ) latents. The VQ latents enable the modelling of discontinuities and sudden changes in the scene, as they are discrete and less correlated with others [56]. In practice, the VQ conditioning is achieved by introducing two lightweight modules: the \"Latent Generator\" module that compresses the implicit information in encoded 3D positions to generate VQ latent code, which is directed through the \"Latent Routing\" module to condition various layers of radiance field. The additional parameters introduced through these modules are significantly less than training an independent NeRF model for each level, leading to a significant reduction in memory.", + "bbox": [ + 75, + 414, + 470, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For evaluating the proposed Strata-NeRF we first test on the proposed synthetic Stratified Scenes dataset, where we find that Strata-NeRF learns the structure in scenes across all levels. In contrast, other baselines produce cloudy and sub-optimal novel views (Figure 2). Further, to test the generalizability of the proposed method on real-world scenes, we utilize the high-resolution RealEstate10K dataset. We find that Strata-NeRF significantly outperforms other baselines and produces high-fidelity novel views without artifacts compared to baselines. This is also observed quantitatively through improvement in metrics, where it establishes", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a new state-of-the-art. In summary,", + "bbox": [ + 500, + 90, + 735, + 107 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We first introduce the task of implicit representation for 3D stratified (hierarchical) scenes using a single radiance field network. For this, we introduce a novel synthetic dataset comprising of scenes ranging from simple to complex geometries.", + "- For implicit modelling of the stratified scenes, we propose Strata-NeRF, which conditions the radiance field based on discrete Vector-Quantized (VQ) latents to model the sudden changes in scenes due to change in hierarchical level (i.e. strata).", + "- Strata-NeRF significantly outperforms the baselines across the synthetic dataset and generalizes well on the real-world scene dataset of RealState10k." + ], + "bbox": [ + 517, + 114, + 893, + 320 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 330, + 640, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating photo-realistic novel views from densely sampled images is a classical problem. Earlier methods solved this issue using light-field-based interpolation techniques [10, 18, 27]. These techniques interpreted the input images as 2D slices of a 4D function - the light field. The only caveat in these methods is their overreliance on dense views. Another popular technique is Structure From Motion (SFM) which reconstructs 3D structure of a scene or an object by using a sequence of 2D images. We suggest readers to read survey papers [47, 37] to understand SFM methods in detail. Shum et al. [49] also provides an excellent review on traditional image based rendering techniques.", + "bbox": [ + 496, + 357, + 890, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Volume Reconstruction. NeRF [34] has shown remarkable results in encoding the 3D geometry of a scene implicitly using the multi-layer perceptron (MLP). Specifically, it trains an MLP, which takes 3D position and a viewing direction to predict colour and occupancy. Many papers have extended this idea to solve different scenarios such as dynamic scenes, low-light scenes, synthesis from fewer views, accelerating the performance etc. Mip-Nerf [1] mitigates the problem of aliasing when a novel view is generated at a different resolution. MVSNeRF [7] generalizes across all the scenes and optimizes the geometry and radiance field using only a few views. NerfingMVS [61] utilizes conventional SFM reconstruction and learning-based priors to predict the radiance field. UNISURF [36] combines implicit surface models and radiance fields to render both surface and volume rendering.", + "bbox": [ + 496, + 537, + 892, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "AR-NeRF [24] replaced pin-hole based camera ray-tracing with aperture camera based ray-tracing. DiVeR [62] uses a voxel based representation to learn the radiance field, Mip-NeRF 360 [2] improves view synthesis on the unbounded scenes and also proposed an online distillation scheme which significantly reduced the training and inference time. Neural Rays [31] solves the occlusion problem by predicting the visibility of the 3D points in their", + "bbox": [ + 496, + 780, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17604", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "representation. Scene Representation Transformers [46] uses Vision Transformers [13] to infer latent representations to render the novel views. Further, many methods [30, 17, 44, 65, 51, 21, 58] have been proposed to improve the slow training and inference time for neural radiance field based methods. Despite many works, no work has focused on modelling the stratified scenes.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF Extensions. Relighting discusses how to model different types of light and then using this model to relight a scene [32, 3, 50, 57, 20]. Breaking the myth that radiance field can only be used in small and bounded scenes, recent methods [52, 55, 45] have scaled it to large-scale city scenes. Another line of work focuses on modelling the dynamic scenes with presence of moving objects [38, 63, 29, 41, 14, 54, 16, 39] through NeRFs.", + "bbox": [ + 75, + 196, + 470, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields and Latents. Recently, a lot of methods have made use of the latents to bring generative capabilities to neural radiance fields. GRAF [48] uses disentangled shape and appearance latent codes to generalize on an object category. For viewpoint invariance, they used typical GAN based training. Pi-GAN [5] uses volumetric rendering equations for consistent 3D views in a generative framework. Pixel-NeRF [66] learns a scene prior to generalize across different scenes. GSN [12] decomposes the radiance field of a scene into local radiance fields by conditioning on a 2D grid of latent codes. Code-NeRF [22] learns the variation of object shapes and textures across by learning separate latent embeddings. LOLNeRF [43] uses a shared latent space which conditions a neural radinace field to model shape and appearance of a single class. PixNerF [4] extends Pi-GAN [5] and maps images to a latent manifold allowing object-centric novel views given a single image of an object. NeRF-W [33] optimizes latent codes to model the scene variations to produce temporally consistent novel view renderings. In contrast to these methods, we propose conditioning NeRF on learnable Vector Quantized latents.", + "bbox": [ + 75, + 316, + 470, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vector Quantized Variational Autoencoders (VQ-VAE) [56]: VQ-VAE uses vector quantization to represent a discrete latent distribution. VQ-VAE has shown applications in Image Generation [42, 40], speech and audio processing [19, 59]. Further, it's extension like VQ-VAE2 [42] uses hierarchical latent space for high-quality generation.", + "bbox": [ + 76, + 648, + 468, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 753, + 215, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF represents a scene as an implicit function $f: (X, d) \\to (c, \\sigma)$ which maps a 3D position $X = (x, y, z)$ and $d = (\\theta, \\phi)$ to a color $c = (r, g, b)$ and occupancy density $\\sigma$ . An MLP parametrizes this implicit function $f$ . Before sending the inputs $X$ and $d$ through the network, a positional encoding is used to project them in a high dimensional space [53]. Finally, the volume rendering [23] procedure enables NeRF to represent scenes with photo-realistic", + "bbox": [ + 76, + 779, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b9854a1c6399d6dd8bc8df78c224f813bfd07a8756cd7494c667923fec50ead5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 95, + 883, + 170 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d6a9596238b154f66487c5f6de6d25364118e8870ecb4c4b4c54065c7a7c95cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 186, + 616, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/49bfdba3fe4114185e2110448895684aa537079a13db16091e8459f04c4edbce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 186, + 733, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5c7bf12b3d5137ba0c8d9cd2238b449574259832c763ddae7a2dfad346d2a123.jpg", + "image_caption": [ + "Level 0" + ], + "image_footnote": [], + "bbox": [ + 740, + 186, + 852, + 244 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f1180bdc06949be623b7d97c2c831c0d2627052aa4b09d0b26e82c0dff063219.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 864, + 181, + 883, + 202 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2a45ebf2d9b025f5284e9b6ed34e21eeb9833855b41958b33e68fdc45fd9cac8.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 506, + 255, + 617, + 334 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75c95094e4fdd136e8abb46fce225d7da6f355a9ff71d2933a2536f064c152bf.jpg", + "image_caption": [ + "mip360 (level-wise)" + ], + "image_footnote": [], + "bbox": [ + 625, + 255, + 738, + 334 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4c638ba9307783c8a7f14efca3e9c245677b18b8e9d9177d1359f06d4a987301.jpg", + "image_caption": [ + "mip360", + "Figure 3. Analysis on \"Dragon in pyramid\" scene. The top row shows the layout of the levels in 3D scene. Observe that baseline works fine on the scenes when trained individually. Artefacts occur when the baseline is trained on views from the entire scene." + ], + "image_footnote": [], + "bbox": [ + 746, + 255, + 857, + 334 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/efb310eb32c6c7d39c19763434e4960798e8a2250a249151dcc6504f7ae86cac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 864, + 266, + 883, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rendering from novel camera viewpoints.", + "bbox": [ + 500, + 431, + 772, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Volume Rendering. At the crux of NeRF lies the volume rendering equation. A ray $r(t) = o + td$ is cast from the camera center $o$ through the pixel along direction $d$ . The pixel's color value is estimated by integrating along the ray $r(t)$ as described in Eq. 1", + "bbox": [ + 498, + 448, + 890, + 523 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc (r) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (r (t)) c (r (t), d) d t \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 537, + 890, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where transmittance $T(t) = \\exp(-\\int_{t_n}^t \\sigma(r(s)) ds)$ is the probability that a ray passes unhindered from the near plane $(t_n)$ to plane $(t)$ and use this probability to integrate till far plane $(t_f)$ . In Mip-NeRF [1], a ray $r(t)$ is divided into intervals $T_i = [t_i, t_{i+1})$ which corresponds to a conical frustum. For each interval $T_i$ , it computes the mean and variance $(\\mu, \\Sigma)$ and uses it for integrated position encoding as illustrated in Eq. 2.", + "bbox": [ + 498, + 587, + 890, + 709 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma (\\mu , \\Sigma) = \\left\\{\\left[ \\begin{array}{l} \\sin (2 ^ {l} \\mu) \\exp (- 2 ^ {2 l - 1} \\operatorname {d i a g} (\\Sigma)) \\\\ \\cos (2 ^ {l} \\mu) \\exp (- 2 ^ {2 l - 1} \\operatorname {d i a g} (\\Sigma)) \\end{array} \\right] \\right\\} _ {0} ^ {L - 1} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 722, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This solves the aliasing issue in the original NeRF. Mip-NeRF 360 [2] proposed coarse-to-fine online distillation for proposal sampling, which efficiently reduces the training time as the proposed MLP only predicts density. They also proposed ray parametrization and regularisation techniques to alleviate hanging artifacts in unbounded scenes. We'll refer Mip-NeRF 360 [2] as mip360 in all our discussions. We choose mip360 [2] as the baseline for all our experiments.", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17605", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/ea3c8479e7d5e1dcbc0eff2276525f37083499d707aed10c56d0f8f95ce3db6c.jpg", + "table_caption": [ + "Table 1. A quantitative comparison of mip360 (level-wise) and mip360 (all views) on \"Dragon in pyramid\" scene." + ], + "table_footnote": [], + "table_body": "
Level 0Level 1
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
mip360 (level-wise)31.53900.91810.130429.85600.81330.3484
mip36030.88470.90060.136724.38760.70550.5163
", + "bbox": [ + 76, + 114, + 468, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Motivation", + "text_level": 1, + "bbox": [ + 76, + 194, + 194, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The majority of real-world scenarios are stratified with multiple levels. For example, a commodity store has exterior and interior structures. This work addresses an essential question for such stratified scenes: Can a single radiance field learn such hierarchical scenes? This section introduces and discusses our observations on one such stratified scene: \"Dragon in Pyramid\", as illustrated in Figure 3. The outer structure of \"Dragon in Pyramid\" is a Mayan pyramid that has a dragon inside it. To validate our claim, we first train the baseline model on each level, i.e., on outer pyramid views and inner views (focusing dragon) independently. We refer to these separately trained models as mip360 (levelwise). Then, we train a single mip360 model using the outer and inner views for the scene. The term \"level\" in our work refers to each level in a stratified scene. In the scene depicted in Figure 3, level 0 denotes the pyramid's outer construction, while level 1 denotes the pyramid's interior structure, which contains a dragon.", + "bbox": [ + 73, + 220, + 468, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 shows that the baseline model performs remarkably well when trained separately on each level. In comparison, the metric values for the baseline model trained jointly on both levels of stratified scene declines. PSNR at level 1 is $24.39dB$ , a $5.47dB$ reduction compared to mip360 (level-wise). Similarly, performance in level 0 has declined, but less dramatically than in the inner level. This pattern is observed across all metrics. Furthermore, the qualitative results illustrated in Figure 3 backs up the quantitative study's findings. Figure 3 indicates that mip360 (level-wise) generates novel views on par with the ground truth. However, shown in Figure 3, the jointly trained model has white artifacts on the pyramid's outer structure and haziness in front of the dragon inside the pyramid. This demonstrates that current radiance field networks have issues while learning a 3D representation of a stratified scene. We perform a similar experiment for a RealEstate10K scene in Appendix E.1 in the supplementary material.", + "bbox": [ + 75, + 493, + 470, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Method", + "text_level": 1, + "bbox": [ + 76, + 782, + 166, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section describes our method: Strata-NeRF for stratified scenes. We generate latent codes with the latent generator described in Section 5.1. This latent code is fed into the radiance field architecture through the latent router, described in Section 5.2. Figure 4 depicts the overall architecture of Strata-NeRF. We adopt the base neural radiance", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "field architecture proposed in mip360 [2].", + "bbox": [ + 500, + 90, + 779, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Latent Generator", + "text_level": 1, + "bbox": [ + 500, + 114, + 671, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A latent space reflects the scene's \"compressed\" representation. It has been shown in various works that this space has rich properties. VQ-VAE [56] learns a codebook to model the discrete distribution of the latent space of a variational-autoencoder. The encoder's output is compared to all of the vectors in the codebook. The nearest vector is fed into the decoder as input. Since most data in the world is discrete, VQ based models have been highly successful in image generation [15], speech encoding [56], and other applications. In a stratified scene, the definition of level is also discrete. Hence, our method employs VQ-VAE as a latent generator because of their proven success in representing discrete distributions.", + "bbox": [ + 496, + 138, + 890, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use Integrated Positional Encoded (IPE) [2] $\\gamma (\\mathbf{x})$ as input to our latent generator. We encode $\\gamma (\\mathbf{x})$ and then search the codebook for the closest vector. After that, the closest vector from the codebook is used to condition the radiance field network. Specifically, $\\gamma (\\mathbf{x})$ is passed through a set of two hidden layers to generate an encoded input $\\mathbf{z}$ . The encoded latent code $\\mathbf{z}$ is then passed through the quantizer bottleneck to determine the quantized latent code $\\mathbf{z}_{\\mathrm{e}}$ where $\\mathbf{z}_{\\mathrm{e}}\\in E$ ; where $E\\in R^{N\\times D}$ is the codebook; $N$ is the number of vectors in the codebook, and $D$ is the dimension of the latent space. $\\mathbf{z}_{\\mathrm{e}}$ is then supplied into the decoder network, which consists of two hidden layers, to yield $\\mathbf{y}$ as the reconstructed output of $\\gamma (\\mathbf{x})$ . The quantized latent $\\mathbf{z}_{\\mathrm{e}}$ is also sent into the radiance field network through the \"Latent Router\" block. Loss for this variational autoencoder (VAE) block is defined as follows:", + "bbox": [ + 496, + 333, + 892, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v q} = \\left| \\left| \\gamma (\\mathbf {x}) - \\mathbf {y} \\right| \\right| _ {2} ^ {2} + \\left| \\left| s g \\left(\\mathbf {z} _ {\\mathbf {e}}\\right) - \\mathbf {z} \\right| \\right| _ {2} ^ {2} + \\beta \\left| \\left| \\mathbf {z} _ {\\mathbf {e}} - s g (\\mathbf {z}) \\right| \\right| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 584, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The \"Latent Generator\" module based on VAE is jointly trained with the NeRF through backpropagation.", + "bbox": [ + 496, + 609, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2. Latent Router", + "text_level": 1, + "bbox": [ + 500, + 648, + 645, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Latent Router block is inspired by the CodeNeRF architecture [22], in which shape and texture latent codes are sent to the NeRF MLP through a residual connection. In our architecture, the quantized latent codes $z_{e}$ that are generated in the \"Latent Generator\" block are input to the Radiance field after passing through an MLP layer in the Latent Router as shown in Figure 4.", + "bbox": [ + 496, + 672, + 890, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.3. Training Strata-NeRF", + "text_level": 1, + "bbox": [ + 500, + 786, + 712, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For training Strata-NeRF, we utilize the losses suggested by mip360 [2] as we use a similar radiance field design. $\\mathcal{L}_{\\text{recon}}(c(r,t),c^{*}(r))$ denotes the reconstruction loss between the estimated colour along a ray and the actual colour value. $\\mathcal{L}_{\\text{dist}}(s,w)$ is the distortion loss where $s$ is the normalized ray distances and $w$ is the weight vector. Note that", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17606", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/effc3552e5db8befdd6ace230bb6508ca83ce1e803fbadd50c9e854b5f9e2427.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [ + "Radiance field Latent Router LatentGenerator VectorCodebook 3D point" + ], + "bbox": [ + 91, + 88, + 658, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f82905e3db6a328b7c99eb4b1856abb6cc7d7c7bfe12007b50242160bd91306b.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 666, + 93, + 893, + 257 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/858f690860d1d1badfde3ab2b0f564ae58e6608801ec5f1fb358e02834c24755.jpg", + "table_caption": [ + "Table 2. Characteristic Comparison of the proposed methods" + ], + "table_footnote": [], + "table_body": "
MethodDiscrete RepresentationPhotometric LossesVAE loss
NeRF [34]X✓X
mip360 [2]X✓X
Plenoxel[64]✓✓X
Instant-NGP[35]✓✓X
TensoRF[6]✓✓X
Ours✓✓✓
", + "bbox": [ + 78, + 378, + 468, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we don't alter anything in the proposal MLP. More details are provided in mip360 [2]. The total loss for Strata-NeRF is given as:", + "bbox": [ + 75, + 518, + 468, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c o n}} (c (r, t), c ^ {*} (r)) + \\lambda_ {1} \\mathcal {L} _ {\\text {d i s t}} (s, w) + \\lambda_ {2} \\mathcal {L} _ {v q} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 573, + 468, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $\\lambda_{1} = 0.01$ , $\\lambda_{2} = 0.1$ and $\\beta = 1.0$ across all our experiments, as they work robustly [2] for Strata-NeRF.", + "bbox": [ + 75, + 604, + 468, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 647, + 209, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We discuss implementation details in Section 6.1. Section 6.2 discusses the dataset used for evaluating our method with other baselines. In Section 6.3, we present quantitative and qualitative comparison with the baseline methods. Additionally, we discuss the ablations for the proposed method.", + "bbox": [ + 75, + 672, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 771, + 294, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method builds on mip360 [2] as the base radiance field. We use a latent generator network which consists of an encoder-decoder architecture and a vector-codebook. The encoder has two linear layers of hidden size 48, and the decoder has one linear layer of hidden size 96. The output dimension of our decoder matches the output from Integrated Positional Encoding (IPE) block. The size of", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/771cc9a9ac616270b2333e7d5f4876af7c98ab201f37d60ae097ed00f791970a.jpg", + "image_caption": [ + "Figure 4. For each $3D$ point along the projected ray, we generate a latent code using our \"Latent generator\" module. The generated latent code is routed to the MLP using \"Latent Router\". Vector Codebooks learn the discrete distribution of positionally encoded $3D$ points. (a) Our model's end-to-end architecture; (b) components of the \"Latent Generator\" and \"Latent Router\" blocks." + ], + "image_footnote": [], + "bbox": [ + 506, + 375, + 622, + 489 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1cc12f92d5e5142b4798012e2865807b9c640e4b7fd26d1dd4562ee495335ed7.jpg", + "image_caption": [ + "Figure 5. Skeleton mesh of the stratified scenes : Bhutanese House and Coffee Shop. More details are in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 624, + 380, + 890, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "our codebook is 1024, and the dimension of each vector in the codebook is 48. We condition the neural radiance field through the latent generated after the quantization step in the latent generator. We use a Latent routing module consisting of two linear layers of hidden-size 256. As illustrated in Figure 4, the output of the linear layer in the routing module conditions the first two layers of the radiance field network. We employ the losses outlined in Section 5. On each scene, we train our approach for $150k$ iterations. We use Adam [25] optimizer with a learning rate of $1e^{-6}$ . Further details are provided in supplementary material.", + "bbox": [ + 496, + 547, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.2. Evaluation Dataset", + "text_level": 1, + "bbox": [ + 500, + 724, + 681, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Most of the radiance field methods evaluate their results on the synthetic (Blender) and real-world (LLFF) datasets proposed in NeRF [34]. These scenes either include a solitary object on a white background or a frontal view of a natural scene. According to our description of stratified scenes, these datasets has only one level. Even large-scale reconstruction datasets like TanksandTemples [26] are not representative of our setting as they only have views either inside or outside of the structure. Similarly, Scannet [9] a dataset for real-world interior scenes, lacks the characteristics of", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17607", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/468413363cb4cf1cb4176e0715ac6764594be26d58afd49114d787ebb3f5e675.jpg", + "table_caption": [ + "Table 3. Quantitative evaluation on test-set against baselines discussed in Section 6.1. Each column is depicts the best and second best." + ], + "table_footnote": [], + "table_body": "
Cube-Sphere-Monkey
Level 0Level 1Level2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]28.33140.93830.103418.18060.49760.498122.11780.59950.382522.87660.67840.3280
mip360 [2]28.31490.92980.115619.04430.53430.493024.91360.73260.324524.09090.73220.3110
Plenoxels [64]25.35470.91690.123813.11480.33200.689521.55680.65230.380320.00870.63370.3979
Instant-NGP [35]28.21040.91680.112314.36480.18300.721617.69140.27440.599720.08890.45810.4779
TensoRF [6]32.00770.95320.069213.74870.15370.710613.00750.24960.688619.58800.45210.4894
Ours26.93350.92980.125525.70880.77380.295926.19120.81720.254926.27780.84030.2254
Bhutanese House
Level 0Level 1Level 2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]11.44780.69170.371117.12090.58860.707818.39180.69520.659115.65350.65850.5793
mip360 [2]26.62400.90020.206224.59460.72960.473929.42250.85770.415626.88040.82910.3652
Plenoxels [64]15.22050.77520.305213.03860.46700.670319.30500.58190.588615.85470.60800.5214
Instant-NGP [35]23.97910.92170.150024.73160.70090.423727.66170.81360.378625.45750.81210.3174
TensoRF [6]13.88800.76070.314217.02440.48560.642116.81700.63060.633215.90980.62560.5298
Ours27.68420.90460.204524.91800.73710.461629.46460.85750.417227.35560.83310.3611
", + "bbox": [ + 76, + 101, + 893, + 340 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/195f1fd62d3baa16954f87e6e48794556fd36d36714a35f36cfca66db935cd54.jpg", + "table_caption": [ + "Table 4. Quantitative evaluation on test-set against baselines discussed in Section 6.1. Each column is depicts the best and second best." + ], + "table_footnote": [], + "table_body": "
Coffee Shop
Level 0Level 1Level2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]06.74460.61970.469816.13980.49150.798212.88890.42130.815811.92440.51080.6946
mip360 [2]26.20730.88250.186727.05000.80860.378534.20230.93620.195029.15320.87570.2534
Plenoxels [64]19.32040.79680.257912.38710.40440.690422.43250.68560.458518.04670.62890.4689
Instant-NGP [35]29.94250.93240.099228.10400.81930.345229.65740.86800.262129.23470.87320.2355
TensoRF [6]33.03370.94350.069219.31150.53310.658021.18520.71690.459424.51020.73120.3955
Ours26.44990.88020.193928.63920.84030.345033.26920.92540.224329.45280.88190.2544
Dragon In Pyramid
Level 0Level 1Level 2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]14.64050.65950.380020.83680.60520.6856---17.73860.63230.5328
mip360 [2]30.87580.90060.136724.38900.70540.5163---27.63240.80300.3265
Plenoxels [64]13.06670.62470.421714.51260.35720.6498---13.78960.491005358
Instant-NGP [35]23.90540.90100.094924.73890.65940.4664---24.32220.78020.2807
TensoRF [6]35.30150.96320.041419.55730.52210.6809---27.42940.74270.3611
Ours29.47730.87000.169926.17220.74890.4573---27.82480.80950.3136
", + "bbox": [ + 76, + 364, + 890, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a stratified dataset. Because of the direct unavailability of stratified scenes, we built our own dataset that replicates the intended \"stratified\" scenario. We create a synthetic scene dataset using a mesh-editing software Blender [8] and real scene dataset by altering RealEstate10K dataset which was proposed for the camera localization task.", + "bbox": [ + 75, + 628, + 468, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed synthetic dataset has two important variations based on: (a) the number of stratified levels and (b) the geometric complexity. We classify based on the geometry's complexity as follows: (a) Simple Scenes: Stratified scenes using geometric components such as the sphere, cube, and so on; and (b) Complex Scenes: Stratified scenes that mimic real-world scenes. For Simple Scenes, we leverage models and textures provided by Blender [8]. We utilized publicly available graphical models and composited them to create a real-world configuration for Complex scenes. For example, to design the \"Coffee shop\" scene, we selected a building", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "structure for the outer level and walls and glasses for the intermediate level structure. For the core level, we composited elements such as a cash register, coffee cups, and so on to simulate a real-world coffee-shop scene. To avoid photo-metric changes, we use fixed illumination. For each stratified level, the camera settings: field of vision and focal length are fixed. Each scene is rendered at $200 \\times 200$ resolution. The camera viewpoint are sampled evenly from the curved surface of a hemisphere and then randomly divided into train, validation, and test sets. Inner objects in Simple Scenes are rendered from the surface of a sphere. Figure 5 depicts the proposed dataset's skeletal meshes. Further information on dataset is present in Appendix B in the supplementary material.", + "bbox": [ + 496, + 628, + 890, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RealEstate10K dataset. We extracted four scenes \"Spanish Colonial Retreat in Scottsdale Arizona\", \"139 Barton Avenue Toronto Ontario\", \"31 Brian Dr Rochester NY\"", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17608", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b06b5d6c90a66e7dfd9bfef3c88ab39e11f9df4d8c9a90404484986078f1202d.jpg", + "image_caption": [ + "Figure 6. (From top to bottom) Qualitative results on the proposed synthetic datasets (Figure 5). Each row represents a novel view from a level of the stratified scene. The ground-truth (GT) is shown in Column 1. Compared to baselines (Column 2-4), our method's (Column 5) renderings are more consistent to GT." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 468, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and “7 Rutledge Ave Highland Mills” from RealEstate10K dataset. We manually inspected and removed regions which had dynamic components in them. More details about converting RealEstaet10k dataset for our stratified setting is provided in Appendix C the supplementary material.", + "bbox": [ + 75, + 532, + 468, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 619, + 197, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present quantitative and qualitative analysis of Strata-NeRF on the datasets described in Section 6.2.", + "bbox": [ + 75, + 643, + 468, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our model with NeRF [34], mip360 [2], Instant-NGP [35], TensoRF [6] and Plenoxels [64]. We chose Plenoxels [64] for comparison because it uses sparse-voxel representation which already discretizes the continuous 3D space, which can be useful in stratified scenes. It is worth noting that the sizes of the synthetic scenes in our dataset differ. As a consequence, the authors' recommended configuration file did not produce the optimal results. As a result, we modified the configuration files for unbounded scenes released by the creators of mip360 [2] to improve performance. For Instant-NGP [35], TensoRF [6] and Plenoxels [64], we change the hyperparameters like bound and scale as suggested in the official implementatinos. More information is in Appendix D in the supplementary material. Table 2 provides an overview", + "bbox": [ + 75, + 672, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c4bba7a03aca2b49eb795ae930ebed8f0664a8add23eb34043105c273a4f7b3.jpg", + "table_caption": [ + "Table 5. Quantitative comparison of our model and baseline on \"139 Barton Avenue\" scene of RealEstate10K dataset." + ], + "table_footnote": [], + "table_body": "
MetricsLevel 0Level 1Level 2Level 3Level 4Level 5
mip360[2]PSNR ↑18.08616.49624.45920.86217.47910.999
SSIM ↑0.6180.5950.7710.7020.5840.409
OursPSNR ↑23.16421.66525.23624.15622.87925.409
SSIM ↑0.8260.7570.7890.7910.7530.782
", + "bbox": [ + 501, + 114, + 890, + 178 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b09331081213b13702f15ea35dbcb69581155dea9c29a40fedbea0ba26dcbcef.jpg", + "table_caption": [ + "Table 6. Quantitative comparison of our model and mip360 baseline on Six Layer Scene." + ], + "table_footnote": [], + "table_body": "
DatasetLevelsmip360 [2]Oursmip360 [2]Ours
Spanish Colonial Retreat520.10622.5140.6220.685
31 Brian Dr Rochester423.27328.0260.7150.835
139 Barton Avenue618.99123.4330.6420.780
7 Rutledge Ave719.62125.0400.5660.791
", + "bbox": [ + 500, + 218, + 890, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of baselines.", + "bbox": [ + 498, + 310, + 586, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Table 3 & 4 shows the average PSNR, SSIM [60] and LPIPS [67] for each stratified level in unseen test views. We find that our method surpasses other methods across all metrics most of the time. The baseline mip360 [2] works fine for the exterior structure but fails for the inner layers in the \"Cube-Sphere-Monkey\" scene. Strata-NeRF, on the other hand, offers superior metrics at all stratified levels. The baseline models do well in the outer scene but perform sub-optimally in the inner levels, especially in level 1. These outcomes demonstrate that our method outperforms the baseline models significantly.", + "bbox": [ + 496, + 325, + 890, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6 shows the summary of average PSNR and SSIM for all the levels in a scene for RealEstate10K dataset. In this case, we only compare our method with mip360 as it is the best performing one among others on the synthetic dataset. We observe that our method outperforms the baseline method in all scenarios. Further, we present level-wise result for a specific scene in Table 5. We observe that for real datasets with increase in number of levels, the magnitude of performance improvement increases, which demonstrates the effectiveness of the proposed approach. Further, we also compare Instant-NGP [35] and TensoRF [6] on a RealEstate10K scene in Appendix E.2 in the supplementary material.", + "bbox": [ + 496, + 491, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 6 & 8 depicts the qualitative results for the synthetic dataset scenes described in Section 6.2. We observe that NeRF [34] performs poorly regardless in majority of scenarios. The generated novel views for \"Coffee Shop\" are poor. It only works well in level 0 of \"Cube-Sphere-Monkey\" dataset. mip360 [2] outperforms NeRF but falters in level 1. Furthermore, in level 0 of the \"Cube-Sphere-Monkey\" dataset, mip360 only generates a white patch with no visible structure. For RealEstate10K dataset, it can be observed in Figure 7 that mip360 generates blurry results compared to our approach. Further, we find that our approach generates consistent and structurally salient novel views throughout all levels and scenes. We show qualitative results for Instant-NGP [35] and Ten", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17609", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2335e406acbed83df77332ccf297d06596559f5bb93dac890cb5f3e10746c33e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 346, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f426afd44ec0bf60187004a390613f1001ee1c09fdb430a56739cc36d84a93a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 148, + 346, + 209 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e4656c1cfd6f9dfe597ceeef25b4fd8f5f64a0a9bbbc0f3a30d3085f1a4c1fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 210, + 346, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ae91a27972f11a0b86516b95e70e7d57aac74d12e4014096eb94e889e4a4c7c.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison on Scenes from RealEstate10K dataset between mip360 (left image) and our method Strata-NeRF (right image) in a pair. Each row represents a scene in RealEstate10K and each pair represents a level in that scene. Our method outperforms and produces good quality novel views compared to mip360." + ], + "image_footnote": [], + "bbox": [ + 80, + 271, + 346, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2baa77f5b93102d0f175d88813483383a65b80626d21fe9f5a44280b9782b7a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 89, + 617, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ecf8f8534a7b07fb158e7934c144407d555ddb097fed468452b42547a9f3d549.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 148, + 617, + 209 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32c82bd0a4851286daf03538d9f8698e363965b7cfe8af18fc9ecd1775fc41c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 210, + 617, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6127d0447afda68b414c7a13dcc74ce31a0bf66885585e1da7a2e869b0e1a25a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 271, + 617, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abf5469b15af913778d4be15a7e44a40a2054607abfb8fc03d0d932f9d862174.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 90, + 888, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6f10cb0051e30406cb83fe1d8216ea67940c0538558953ee11a9d3d62346aaa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 148, + 888, + 209 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/716157bd223321246c710b06174958f422890355552f41e8f843d26776d52d08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 210, + 888, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/142e4544b3e2651b67d58c8ab97caff3dd544e2bf77d79272bc02e907c5dc69a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 271, + 888, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33db4324fb50753a7c500b03ea06727ff5089286259a789363ddac3577b06c8f.jpg", + "image_caption": [ + "Figure 8. (From top to bottom) Qualitative results on the proposed synthetic datasets. Each row represents a novel view from each level of the stratified scene. The ground-truth view is shown in Column 1. Compared to prior works (Column 2-4) our method's (Column 5) renderings are more similar to the ground-truth." + ], + "image_footnote": [], + "bbox": [ + 81, + 402, + 468, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "soRF [6] in Appendix E.2 in the supplementary material.", + "bbox": [ + 75, + 625, + 460, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Worst Case Analysis. When comparing different methods, average metrics are often insufficient to determine which method is superior to the others. As we have observed in Figure 9 that the baseline method fails on some of test images, hence we also compare the methods in worst care scenarios. The worst-case analysis describes a method's worst performance on the dataset. The worst case analysis is particularly useful to detect the shortcomings of the methods. We present analysis in two categories: (a) histogram distribution for each metric on the test set, and (b) qualitative comparison of the worst-case scenario for our method on PSNR metric.", + "bbox": [ + 75, + 642, + 467, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 9 compares PSNR histogram plots on test-set views for the \"Cube-Sphere-Monkey\" scene. We can see that the mip360 approach performs poorly on PSNR and ranks low on practically all stratification levels. This supports our argument that the mip360 approach produces ar", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/24e3e768dfcc5b33ab923fe9b86f2bd66b6ce0fe9f27d539c776850633a18f69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 402, + 883, + 500 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da75ae1f9def0ad2b00312b8fac541c3c3c93e9cf7237eb18794aa914fd2955c.jpg", + "image_caption": [ + "Figure 9. (Top Row) Comparison of histogram plots for the test-set for PSNR on \"Cube-Sphere-Monkey\". Note how distribution of our our method is always towards the right compared to other methods. $x - axis$ denote metric value and $y - axis$ denotes the frequency. A qualitative comparison of our method's worst-case PSNR results. PSNR is present at the bottom of the result image." + ], + "image_footnote": [], + "bbox": [ + 506, + 507, + 883, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tifacts in such stratified scenes. For our method, the PSNR distributions are on the right. This implies that the novel views on test-set from our method will not be having serious artifacts in most cases, demonstrating its reliability.", + "bbox": [ + 496, + 731, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Images in Figure 9 depict the qualitative results for the worst-case PSNR instances. All methods perform well in level 0. Hence, we are discussing interior levels which are level 1 and level 2. Other approaches fail in the worst-case scenario for our method at level 1. The outputs from NeRF, mip360 and Plenoxel are visually impaired. At level 2, our method has less blur compared to other approaches. These", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17610", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/631f443465fa0af41048ac563ee11f3e596718971ad8f0c9ac4c1783c8588d57.jpg", + "image_caption": [ + "Figure 10. Novel-views from different levels of 'Real Estate Video Tour 7 Rutledge Ave Highland Mills NY 10930 Orange County NY' scene in Real Estate 10K dataset. The two rows are from two-different view-points." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 348, + 219 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/85acfe2cb876e987d62c3b129cd1ead3ac06a733cf0722afe41f04726714f496.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 88, + 617, + 219 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/79208b58ad8e7ad9ed3112b3f81cd95dc5e6988c81d18c94161352fde2a8c63e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 89, + 887, + 219 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/509a0168e30e0cd2688cc1ab20db3c901526dd588650c45410230b37bc6408f2.jpg", + "table_caption": [ + "Table 7. Quantitative comparison of our model and baseline on Synthetic Six Layer Scene." + ], + "table_footnote": [], + "table_body": "
MetricsLevel 0Level 1Level 2Level 3Level 4Level 5
mip360 [2]PSNR ↑22.21516.18315.08412.01221.81321.539
SSIM ↑0.7770.4420.5100.3440.8170.647
OursPSNR ↑23.88921.44921.45624.09528.28321.898
SSIM ↑0.8330.6810.6850.7220.8830.686
", + "bbox": [ + 78, + 296, + 468, + 363 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/31aebb8b14c59cfadc4e06a15aa2d4207d8883b9ec8c4bf66dee0faf531acbcd.jpg", + "image_caption": [ + "Figure 11. Comparisons of different codebook size on \"Dragon in Pyramid\" scene for different vector-codebook sizes. Note at size=1024 we achieve the best results with less artifacts." + ], + "image_footnote": [], + "bbox": [ + 81, + 378, + 460, + 522 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "findings demonstrate that our method is better suited to represent stratified scenes than others.", + "bbox": [ + 75, + 582, + 467, + 611 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ablation Studies. To analyse our proposed method, we present an ablation on the size of the vector codebook in our latent generator. Table 8 shows the ablation for the size of the vector codebook on the \"Coffee Shop\" dataset. We trialed with codebook sizes of 512, 1024 and 4096. We found that size 1024 provides optimal performance. As shown in Figure 11, increasing the codebook size induces haziness in the generated novel views, while decreasing the size creates white artifacts in level 0. As a result, we fix the size 1024 for all of our synthetic experiments. Whereas for RealEstate10K dataset we find that codebook size of 4096 produces the optimal tradeoff of results across levels, as it contains more number of levels and details. We further discuss the key architectural design choices for Latent Generator and Latent Router modules in Appendix E.5.", + "bbox": [ + 75, + 613, + 467, + 838 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "No. of levels: To further test the efficacy of our method on higher number of levels, we created a \"Simple Geometry\" scene consisting of primitive geometry shapes like cube and spheres. More details are in the supplementary", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b9229ae9a4e6cc049fb7c699e94815f228daf981357edf53cd5d2e95e9172b48.jpg", + "table_caption": [ + "Table 8. Quantitative results on \"Cube-Sphere-Monkey\" scene for ablation on size of the vector codebook in Latent Generator." + ], + "table_footnote": [], + "table_body": "
SizePSNR ↑SSIM ↑LPIPS ↓
Level 0Level 1Level 0Level 1Level 0Level 1
51229.545826.34970.87430.73950.16750.4899
102429.483426.17150.87010.74890.13670.5163
409628.460927.82740.86280.73420.17760.5027
", + "bbox": [ + 501, + 295, + 890, + 359 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "material. Table 7 displays the results for both the baseline and our approach across a six levels stratified scene. The average PSNR/SSIM for the mip360 baseline is $15.35 / 0.487$ , while our method achieved PSNR/SSIM of $23.54 / 0.754$ which improves PSNR and SSIM by $53.35\\%$ and $54.83\\%$ respectively. This shows that our method performs better on increasing number of levels when compared with the baseline method. These observations also hold true for scenes in the RealEstate10K dataset as shown in Table 5.", + "bbox": [ + 496, + 386, + 890, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 537, + 617, + 553 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we focus on the problem of modelling the 3D representation of a stratified and hierarchical scene, implicitly through a single neural field. For this, we propose Strata-NeRF, which models scenes with stratified structures by introducing a VQ-VAE-based latent generator to implicitly learn the distribution of latent space of input 3D locations and condition the neural radiance field with the latent code generated from this distribution. We also introduce a new synthetic dataset with stratified-level scenes and use it to analyse various existing approaches. Through quantitative, qualitative, and worst-case analysis on this dataset, we show that Strata-NeRF has a more stable 3D representation than the other methods. Further, the improvements due to Strata-NeRF also generalize to real-world RealEstate10K dataset, where it outperforms baselines by a significant margin establishing a new state-of-the-art. We believe designing a new volume rendering equation for modelling complex stratified scenes is a good direction for future work.", + "bbox": [ + 496, + 564, + 890, + 835 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by Samsung R&D Institute India, Bangalore, PMRF and Kotak IISc AI-ML Centre (KIAC). Srinath Sridhar was partly supported by NSF grant CNS-2038897", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17611", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 9", + "[3] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 3", + "[4] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision (ECCV). Springer, 2022. 5, 6, 7, 8", + "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[8] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6", + "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5", + "[10] Abe Davis, Marc Levoy, and Fredo Durand. Unstructured light fields. In Computer Graphics Forum, volume 31. Wiley Online Library, 2012. 2", + "[11] Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. Fov-nerf: Foveated neural radiance fields for virtual reality. IEEE Transactions on Visualization and Computer Graphics, 28(11), 2022. 1", + "[12] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W Taylor, and Joshua M Susskind. Unconstrained scene generation with locally conditioned radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[14] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (ICCV). IEEE Computer Society, 2021. 3", + "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4", + "[16] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[17] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[18] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2", + "[19] Karol Gregor, George Papamakarios, Frederic Besse, Lars Buesing, and Theophane Weber. Temporal difference variational auto-encoder. arXiv preprint arXiv:1806.03107, 2018. 3", + "[20] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[21] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 4", + "[23] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3), 1984. 3", + "[24] Takuhiro Kaneko. Ar-nerf: Unsupervised learning of depth and defocus effects from natural images with aperture rendering neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2", + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5", + "[26] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 5" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17612", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2", + "[28] Chaojian Li, Sixu Li, Yang Zhao, Wenbo Zhu, and Yingyan Lin. Rt-nerf: Real-time on-device neural radiance fields towards immersive ar/vr rendering. In Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design, 2022. 1", + "[29] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[30] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3", + "[31] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[33] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (CVPR), June 2021. 3", + "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision (ECCV). Springer, 2020. 1, 2, 5, 6, 7", + "[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4), 2022. 5, 6, 7", + "[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[37] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26, 2017. 2", + "[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3", + "[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Brualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 3", + "[40] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3", + "[43] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagon, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3", + "[44] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[45] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[46] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2", + "[48] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3", + "[49] Harry Shum and Sing Bing Kang. Review of image-based rendering techniques. In Visual Communications and Image Processing 2000, volume 4067. SPIE, 2000. 2", + "[50] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[51] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17613", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[52] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[53] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3", + "[54] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[55] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[56] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 3, 4", + "[57] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (CVPR). IEEE, 2022. 3", + "[58] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[59] Xin Wang, Shinji Takaki, Junichi Yamagishi, Simon King, and Keiichi Tokuda. A vector quantized variational autoencoder (vq-vae) autoregressive neural $f_{-}0$ model for statistical parametric speech synthesis. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28, 2019. 3", + "[60] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4), 2004. 7", + "[61] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[62] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[64] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 5, 6, 7", + "[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[66] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3", + "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7" + ], + "bbox": [ + 501, + 92, + 890, + 401 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17614", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/246a74f2-93a7-4d68-8843-6238c9108476_model.json b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/246a74f2-93a7-4d68-8843-6238c9108476_model.json new file mode 100644 index 0000000000000000000000000000000000000000..61c310808af8ac7ce11fc42a8af6c9bc2ac27f51 --- /dev/null +++ b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/246a74f2-93a7-4d68-8843-6238c9108476_model.json @@ -0,0 +1,2688 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.783, + 0.151 + ], + "angle": 0, + "content": "Strata-NeRF : Neural Radiance Fields for Stratified Scenes" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.18, + 0.78, + 0.199 + ], + "angle": 0, + "content": "Ankit Dhiman\\(^{1,2}\\) R Srinath\\(^{1}\\) Harsh Rangwani\\(^{1}\\) Rishubh Parihar\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.199, + 0.753, + 0.216 + ], + "angle": 0, + "content": "Lokesh R Boregowda2 Srinath Sridhar3 R Venkatesh Babu1" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.216, + 0.914, + 0.236 + ], + "angle": 0, + "content": "\\(^{1}\\)Vision and AI Lab, IISc Bangalore \\(^{2}\\)Samsung R & D Institute India - Bangalore \\(^{3}\\)Brown University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.65 + ], + "angle": 0, + "content": "Neural Radiance Field (NeRF) approaches learn the underlying 3D representation of a scene and generate photorealistic novel views with high fidelity. However, most proposed settings concentrate on modelling a single object or a single level of a scene. However, in the real world, we may capture a scene at multiple levels, resulting in a layered capture. For example, tourists usually capture a monument's exterior structure before capturing the inner structure. Modelling such scenes in 3D with seamless switching between levels can drastically improve immersive experiences. However, most existing techniques struggle in modelling such scenes. We propose Strata-NeRF, a single neural radiance field that implicitly captures a scene with multiple levels. Strata-NeRF achieves this by conditioning the NeRFs on Vector Quantized (VQ) latent representations which allow sudden changes in scene structure. We evaluate the effectiveness of our approach in multi-layered synthetic dataset comprising diverse scenes and then further validate its generalization on the real-world RealEstate10K dataset. We find that Strata-NeRF effectively captures stratified scenes, minimizes artifacts, and synthesizes high-fidelity views compared to existing approaches. https://ankitatiisc.github.io/Strata-NeRF/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.679, + 0.21, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Novel view synthesis is an ill-posed problem widely encountered in various areas such as augmented reality [24, 28], virtual reality [11], etc. A paradigm change for solving these kinds of problems was brought by the introduction of Neural Radiance Fields (NeRF) [34]. NeRFs are neural networks that take in the spatial coordinates and camera parameters as input and output the corresponding radiance field. Earlier version of NeRFs enable the generation of high-fidelity novel views for bounded scenes, significantly improving over existing techniques like Structure From Motion [47]. Further, the capability of NeRFs have been recently extended to model unbounded scenes by Mip-NeRF 360 [2]. This enabled NeRFs to model complex real-world" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.272, + 0.872, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.549, + 0.895, + 0.62 + ], + "angle": 0, + "content": "Figure 1. Top, wireframe view of a multi-layered stratified scene with three levels (monkey head inside sphere inside a cube). The camera colors indicate views of a specific level. Strata-NeRF enables high-quality reconstruction of such stratified scenes using a single neural network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.673 + ], + "angle": 0, + "content": "scenes, where the scene content can exist at any distance from the camera." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "However, similar to unboundedness in scenes, hierarchies in scenes are also natural. For example, images captured in a house can be categorized into images captured outside and inside across various rooms. Modelling such hierarchical scenes jointly for all levels through a NeRF could be particularly useful in cases of Virtual Reality applications. As it would not require switching to a different NeRF for each level, reducing memory requirement and latency in switching. Further, as the different hierarchies of a scene usually share texture and architectural commonalities, it could lead to effective knowledge sharing and reduce the requirement of training independent models. For tackling the above novel objective, we introduce a paradigm of scenes that can be deconstructed into several tiers, termed \"Stratified Scenes\". A \"stratified\" scene has several levels or" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17603" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.468, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.237, + 0.472, + 0.295 + ], + "angle": 0, + "content": "Figure 2. Novel views for stratified scene in Figure 1, from Mip-NeRF 360 [2] (left) and our method \"Strata-NeRF\" (right). Existing methods struggle to capture stratified scenes with a single network while ours produces sharp results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.47, + 0.414 + ], + "angle": 0, + "content": "groupings of structure (Figure 1). In our work, we first propose a synthetic dataset of stratified scenes, i.e. scenes having multiple levels. This dataset comprises scenes from two categories: (i) Simpler geometry, such as spheres, cubes, or tetrahedron meshes, and (ii) Complex geometry, which closely emulates a real-world setup." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.471, + 0.734 + ], + "angle": 0, + "content": "On such datasets, we find methods such as Mip-NeRF 360 perform well for each level of the hierarchy independently, but produce unsatisfactory results when images from all hierarchical levels are used together for training (Figure 3). This can be attributed to the continuous nature of NeRFs, which is unsuitable for modelling the sudden changes in scenes with shifts in hierarchical levels. Hence, in this work, we introduce Strata-NeRF that explicitly aims to model the hierarchies by conditioning [22, 38, 39, 66, 43] the NeRF on Vector Quantized (VQ) latents. The VQ latents enable the modelling of discontinuities and sudden changes in the scene, as they are discrete and less correlated with others [56]. In practice, the VQ conditioning is achieved by introducing two lightweight modules: the \"Latent Generator\" module that compresses the implicit information in encoded 3D positions to generate VQ latent code, which is directed through the \"Latent Routing\" module to condition various layers of radiance field. The additional parameters introduced through these modules are significantly less than training an independent NeRF model for each level, leading to a significant reduction in memory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.902 + ], + "angle": 0, + "content": "For evaluating the proposed Strata-NeRF we first test on the proposed synthetic Stratified Scenes dataset, where we find that Strata-NeRF learns the structure in scenes across all levels. In contrast, other baselines produce cloudy and sub-optimal novel views (Figure 2). Further, to test the generalizability of the proposed method on real-world scenes, we utilize the high-resolution RealEstate10K dataset. We find that Strata-NeRF significantly outperforms other baselines and produces high-fidelity novel views without artifacts compared to baselines. This is also observed quantitatively through improvement in metrics, where it establishes" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.736, + 0.108 + ], + "angle": 0, + "content": "a new state-of-the-art. In summary," + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.116, + 0.892, + 0.193 + ], + "angle": 0, + "content": "- We first introduce the task of implicit representation for 3D stratified (hierarchical) scenes using a single radiance field network. For this, we introduce a novel synthetic dataset comprising of scenes ranging from simple to complex geometries." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.196, + 0.894, + 0.272 + ], + "angle": 0, + "content": "- For implicit modelling of the stratified scenes, we propose Strata-NeRF, which conditions the radiance field based on discrete Vector-Quantized (VQ) latents to model the sudden changes in scenes due to change in hierarchical level (i.e. strata)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.277, + 0.892, + 0.321 + ], + "angle": 0, + "content": "- Strata-NeRF significantly outperforms the baselines across the synthetic dataset and generalizes well on the real-world scene dataset of RealState10k." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.116, + 0.894, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.642, + 0.348 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.538 + ], + "angle": 0, + "content": "Generating photo-realistic novel views from densely sampled images is a classical problem. Earlier methods solved this issue using light-field-based interpolation techniques [10, 18, 27]. These techniques interpreted the input images as 2D slices of a 4D function - the light field. The only caveat in these methods is their overreliance on dense views. Another popular technique is Structure From Motion (SFM) which reconstructs 3D structure of a scene or an object by using a sequence of 2D images. We suggest readers to read survey papers [47, 37] to understand SFM methods in detail. Shum et al. [49] also provides an excellent review on traditional image based rendering techniques." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.893, + 0.78 + ], + "angle": 0, + "content": "Neural Volume Reconstruction. NeRF [34] has shown remarkable results in encoding the 3D geometry of a scene implicitly using the multi-layer perceptron (MLP). Specifically, it trains an MLP, which takes 3D position and a viewing direction to predict colour and occupancy. Many papers have extended this idea to solve different scenarios such as dynamic scenes, low-light scenes, synthesis from fewer views, accelerating the performance etc. Mip-Nerf [1] mitigates the problem of aliasing when a novel view is generated at a different resolution. MVSNeRF [7] generalizes across all the scenes and optimizes the geometry and radiance field using only a few views. NerfingMVS [61] utilizes conventional SFM reconstruction and learning-based priors to predict the radiance field. UNISURF [36] combines implicit surface models and radiance fields to render both surface and volume rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.902 + ], + "angle": 0, + "content": "AR-NeRF [24] replaced pin-hole based camera ray-tracing with aperture camera based ray-tracing. DiVeR [62] uses a voxel based representation to learn the radiance field, Mip-NeRF 360 [2] improves view synthesis on the unbounded scenes and also proposed an online distillation scheme which significantly reduced the training and inference time. Neural Rays [31] solves the occlusion problem by predicting the visibility of the 3D points in their" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17604" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.197 + ], + "angle": 0, + "content": "representation. Scene Representation Transformers [46] uses Vision Transformers [13] to infer latent representations to render the novel views. Further, many methods [30, 17, 44, 65, 51, 21, 58] have been proposed to improve the slow training and inference time for neural radiance field based methods. Despite many works, no work has focused on modelling the stratified scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.471, + 0.318 + ], + "angle": 0, + "content": "NeRF Extensions. Relighting discusses how to model different types of light and then using this model to relight a scene [32, 3, 50, 57, 20]. Breaking the myth that radiance field can only be used in small and bounded scenes, recent methods [52, 55, 45] have scaled it to large-scale city scenes. Another line of work focuses on modelling the dynamic scenes with presence of moving objects [38, 63, 29, 41, 14, 54, 16, 39] through NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.649 + ], + "angle": 0, + "content": "Neural Radiance Fields and Latents. Recently, a lot of methods have made use of the latents to bring generative capabilities to neural radiance fields. GRAF [48] uses disentangled shape and appearance latent codes to generalize on an object category. For viewpoint invariance, they used typical GAN based training. Pi-GAN [5] uses volumetric rendering equations for consistent 3D views in a generative framework. Pixel-NeRF [66] learns a scene prior to generalize across different scenes. GSN [12] decomposes the radiance field of a scene into local radiance fields by conditioning on a 2D grid of latent codes. Code-NeRF [22] learns the variation of object shapes and textures across by learning separate latent embeddings. LOLNeRF [43] uses a shared latent space which conditions a neural radinace field to model shape and appearance of a single class. PixNerF [4] extends Pi-GAN [5] and maps images to a latent manifold allowing object-centric novel views given a single image of an object. NeRF-W [33] optimizes latent codes to model the scene variations to produce temporally consistent novel view renderings. In contrast to these methods, we propose conditioning NeRF on learnable Vector Quantized latents." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.65, + 0.47, + 0.742 + ], + "angle": 0, + "content": "Vector Quantized Variational Autoencoders (VQ-VAE) [56]: VQ-VAE uses vector quantization to represent a discrete latent distribution. VQ-VAE has shown applications in Image Generation [42, 40], speech and audio processing [19, 59]. Further, it's extension like VQ-VAE2 [42] uses hierarchical latent space for high-quality generation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.216, + 0.77 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "NeRF represents a scene as an implicit function \\( f: (X, d) \\to (c, \\sigma) \\) which maps a 3D position \\( X = (x, y, z) \\) and \\( d = (\\theta, \\phi) \\) to a color \\( c = (r, g, b) \\) and occupancy density \\( \\sigma \\). An MLP parametrizes this implicit function \\( f \\). Before sending the inputs \\( X \\) and \\( d \\) through the network, a positional encoding is used to project them in a high dimensional space [53]. Finally, the volume rendering [23] procedure enables NeRF to represent scenes with photo-realistic" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.096, + 0.885, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.187, + 0.617, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.188, + 0.735, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.187, + 0.854, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.865, + 0.182, + 0.884, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.867, + 0.21, + 0.882, + 0.248 + ], + "angle": 0, + "content": "Level 0" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.256, + 0.619, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.337, + 0.58, + 0.346 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.256, + 0.739, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.337, + 0.735, + 0.346 + ], + "angle": 0, + "content": "mip360 (level-wise)" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.256, + 0.859, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.337, + 0.828, + 0.346 + ], + "angle": 0, + "content": "mip360" + }, + { + "type": "image", + "bbox": [ + 0.865, + 0.267, + 0.884, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.351, + 0.892, + 0.406 + ], + "angle": 0, + "content": "Figure 3. Analysis on \"Dragon in pyramid\" scene. The top row shows the layout of the levels in 3D scene. Observe that baseline works fine on the scenes when trained individually. Artefacts occur when the baseline is trained on views from the entire scene." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.433, + 0.774, + 0.448 + ], + "angle": 0, + "content": "rendering from novel camera viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.449, + 0.892, + 0.525 + ], + "angle": 0, + "content": "Volume Rendering. At the crux of NeRF lies the volume rendering equation. A ray \\( r(t) = o + td \\) is cast from the camera center \\( o \\) through the pixel along direction \\( d \\). The pixel's color value is estimated by integrating along the ray \\( r(t) \\) as described in Eq. 1" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.538, + 0.892, + 0.573 + ], + "angle": 0, + "content": "\\[\nc (r) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (r (t)) c (r (t), d) d t \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.588, + 0.892, + 0.71 + ], + "angle": 0, + "content": "where transmittance \\( T(t) = \\exp(-\\int_{t_n}^t \\sigma(r(s)) ds) \\) is the probability that a ray passes unhindered from the near plane \\( (t_n) \\) to plane \\( (t) \\) and use this probability to integrate till far plane \\( (t_f) \\). In Mip-NeRF [1], a ray \\( r(t) \\) is divided into intervals \\( T_i = [t_i, t_{i+1}) \\) which corresponds to a conical frustum. For each interval \\( T_i \\), it computes the mean and variance \\( (\\mu, \\Sigma) \\) and uses it for integrated position encoding as illustrated in Eq. 2." + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.723, + 0.892, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\gamma (\\mu , \\Sigma) = \\left\\{\\left[ \\begin{array}{l} \\sin (2 ^ {l} \\mu) \\exp (- 2 ^ {2 l - 1} \\operatorname {d i a g} (\\Sigma)) \\\\ \\cos (2 ^ {l} \\mu) \\exp (- 2 ^ {2 l - 1} \\operatorname {d i a g} (\\Sigma)) \\end{array} \\right] \\right\\} _ {0} ^ {L - 1} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This solves the aliasing issue in the original NeRF. Mip-NeRF 360 [2] proposed coarse-to-fine online distillation for proposal sampling, which efficiently reduces the training time as the proposed MLP only predicts density. They also proposed ray parametrization and regularisation techniques to alleviate hanging artifacts in unbounded scenes. We'll refer Mip-NeRF 360 [2] as mip360 in all our discussions. We choose mip360 [2] as the baseline for all our experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17605" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.115 + ], + "angle": 0, + "content": "Table 1. A quantitative comparison of mip360 (level-wise) and mip360 (all views) on \"Dragon in pyramid\" scene." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "
Level 0Level 1
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
mip360 (level-wise)31.53900.91810.130429.85600.81330.3484
mip36030.88470.90060.136724.38760.70550.5163
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.195, + 0.195, + 0.21 + ], + "angle": 0, + "content": "4. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.222, + 0.47, + 0.493 + ], + "angle": 0, + "content": "The majority of real-world scenarios are stratified with multiple levels. For example, a commodity store has exterior and interior structures. This work addresses an essential question for such stratified scenes: Can a single radiance field learn such hierarchical scenes? This section introduces and discusses our observations on one such stratified scene: \"Dragon in Pyramid\", as illustrated in Figure 3. The outer structure of \"Dragon in Pyramid\" is a Mayan pyramid that has a dragon inside it. To validate our claim, we first train the baseline model on each level, i.e., on outer pyramid views and inner views (focusing dragon) independently. We refer to these separately trained models as mip360 (levelwise). Then, we train a single mip360 model using the outer and inner views for the scene. The term \"level\" in our work refers to each level in a stratified scene. In the scene depicted in Figure 3, level 0 denotes the pyramid's outer construction, while level 1 denotes the pyramid's interior structure, which contains a dragon." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.471, + 0.768 + ], + "angle": 0, + "content": "Table 1 shows that the baseline model performs remarkably well when trained separately on each level. In comparison, the metric values for the baseline model trained jointly on both levels of stratified scene declines. PSNR at level 1 is \\(24.39dB\\), a \\(5.47dB\\) reduction compared to mip360 (level-wise). Similarly, performance in level 0 has declined, but less dramatically than in the inner level. This pattern is observed across all metrics. Furthermore, the qualitative results illustrated in Figure 3 backs up the quantitative study's findings. Figure 3 indicates that mip360 (level-wise) generates novel views on par with the ground truth. However, shown in Figure 3, the jointly trained model has white artifacts on the pyramid's outer structure and haziness in front of the dragon inside the pyramid. This demonstrates that current radiance field networks have issues while learning a 3D representation of a stratified scene. We perform a similar experiment for a RealEstate10K scene in Appendix E.1 in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.168, + 0.8 + ], + "angle": 0, + "content": "5. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "This section describes our method: Strata-NeRF for stratified scenes. We generate latent codes with the latent generator described in Section 5.1. This latent code is fed into the radiance field architecture through the latent router, described in Section 5.2. Figure 4 depicts the overall architecture of Strata-NeRF. We adopt the base neural radiance" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.78, + 0.108 + ], + "angle": 0, + "content": "field architecture proposed in mip360 [2]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.673, + 0.13 + ], + "angle": 0, + "content": "5.1. Latent Generator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.333 + ], + "angle": 0, + "content": "A latent space reflects the scene's \"compressed\" representation. It has been shown in various works that this space has rich properties. VQ-VAE [56] learns a codebook to model the discrete distribution of the latent space of a variational-autoencoder. The encoder's output is compared to all of the vectors in the codebook. The nearest vector is fed into the decoder as input. Since most data in the world is discrete, VQ based models have been highly successful in image generation [15], speech encoding [56], and other applications. In a stratified scene, the definition of level is also discrete. Hence, our method employs VQ-VAE as a latent generator because of their proven success in representing discrete distributions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.893, + 0.576 + ], + "angle": 0, + "content": "We use Integrated Positional Encoded (IPE) [2] \\(\\gamma (\\mathbf{x})\\) as input to our latent generator. We encode \\(\\gamma (\\mathbf{x})\\) and then search the codebook for the closest vector. After that, the closest vector from the codebook is used to condition the radiance field network. Specifically, \\(\\gamma (\\mathbf{x})\\) is passed through a set of two hidden layers to generate an encoded input \\(\\mathbf{z}\\). The encoded latent code \\(\\mathbf{z}\\) is then passed through the quantizer bottleneck to determine the quantized latent code \\(\\mathbf{z}_{\\mathrm{e}}\\) where \\(\\mathbf{z}_{\\mathrm{e}}\\in E\\); where \\(E\\in R^{N\\times D}\\) is the codebook; \\(N\\) is the number of vectors in the codebook, and \\(D\\) is the dimension of the latent space. \\(\\mathbf{z}_{\\mathrm{e}}\\) is then supplied into the decoder network, which consists of two hidden layers, to yield \\(\\mathbf{y}\\) as the reconstructed output of \\(\\gamma (\\mathbf{x})\\). The quantized latent \\(\\mathbf{z}_{\\mathrm{e}}\\) is also sent into the radiance field network through the \"Latent Router\" block. Loss for this variational autoencoder (VAE) block is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.585, + 0.892, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v q} = \\left| \\left| \\gamma (\\mathbf {x}) - \\mathbf {y} \\right| \\right| _ {2} ^ {2} + \\left| \\left| s g \\left(\\mathbf {z} _ {\\mathbf {e}}\\right) - \\mathbf {z} \\right| \\right| _ {2} ^ {2} + \\beta \\left| \\left| \\mathbf {z} _ {\\mathbf {e}} - s g (\\mathbf {z}) \\right| \\right| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.643 + ], + "angle": 0, + "content": "The \"Latent Generator\" module based on VAE is jointly trained with the NeRF through backpropagation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.65, + 0.646, + 0.665 + ], + "angle": 0, + "content": "5.2. Latent Router" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.78 + ], + "angle": 0, + "content": "The Latent Router block is inspired by the CodeNeRF architecture [22], in which shape and texture latent codes are sent to the NeRF MLP through a residual connection. In our architecture, the quantized latent codes \\( z_{e} \\) that are generated in the \"Latent Generator\" block are input to the Radiance field after passing through an MLP layer in the Latent Router as shown in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.714, + 0.803 + ], + "angle": 0, + "content": "5.3. Training Strata-NeRF" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "For training Strata-NeRF, we utilize the losses suggested by mip360 [2] as we use a similar radiance field design. \\(\\mathcal{L}_{\\text{recon}}(c(r,t),c^{*}(r))\\) denotes the reconstruction loss between the estimated colour along a ray and the actual colour value. \\(\\mathcal{L}_{\\text{dist}}(s,w)\\) is the distortion loss where \\(s\\) is the normalized ray distances and \\(w\\) is the weight vector. Note that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17606" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.659, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.263, + 0.385, + 0.277 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.094, + 0.895, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.263, + 0.81, + 0.277 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.088, + 0.279, + 0.877, + 0.301 + ], + "angle": 0, + "content": "Radiance field Latent Router LatentGenerator VectorCodebook 3D point" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.893, + 0.348 + ], + "angle": 0, + "content": "Figure 4. For each \\(3D\\) point along the projected ray, we generate a latent code using our \"Latent generator\" module. The generated latent code is routed to the MLP using \"Latent Router\". Vector Codebooks learn the discrete distribution of positionally encoded \\(3D\\) points. (a) Our model's end-to-end architecture; (b) components of the \"Latent Generator\" and \"Latent Router\" blocks." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.367, + 0.439, + 0.38 + ], + "angle": 0, + "content": "Table 2. Characteristic Comparison of the proposed methods" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.38, + 0.47, + 0.497 + ], + "angle": 0, + "content": "
MethodDiscrete RepresentationPhotometric LossesVAE loss
NeRF [34]X✓X
mip360 [2]X✓X
Plenoxel[64]✓✓X
Instant-NGP[35]✓✓X
TensoRF[6]✓✓X
Ours✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.566 + ], + "angle": 0, + "content": "we don't alter anything in the proposal MLP. More details are provided in mip360 [2]. The total loss for Strata-NeRF is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.574, + 0.469, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c o n}} (c (r, t), c ^ {*} (r)) + \\lambda_ {1} \\mathcal {L} _ {\\text {d i s t}} (s, w) + \\lambda_ {2} \\mathcal {L} _ {v q} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.636 + ], + "angle": 0, + "content": "We use \\(\\lambda_{1} = 0.01\\), \\(\\lambda_{2} = 0.1\\) and \\(\\beta = 1.0\\) across all our experiments, as they work robustly [2] for Strata-NeRF." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.21, + 0.665 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.764 + ], + "angle": 0, + "content": "We discuss implementation details in Section 6.1. Section 6.2 discusses the dataset used for evaluating our method with other baselines. In Section 6.3, we present quantitative and qualitative comparison with the baseline methods. Additionally, we discuss the ablations for the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.295, + 0.789 + ], + "angle": 0, + "content": "6.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our method builds on mip360 [2] as the base radiance field. We use a latent generator network which consists of an encoder-decoder architecture and a vector-codebook. The encoder has two linear layers of hidden size 48, and the decoder has one linear layer of hidden size 96. The output dimension of our decoder matches the output from Integrated Positional Encoding (IPE) block. The size of" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.375, + 0.623, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.381, + 0.892, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.495, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Figure 5. Skeleton mesh of the stratified scenes : Bhutanese House and Coffee Shop. More details are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.716 + ], + "angle": 0, + "content": "our codebook is 1024, and the dimension of each vector in the codebook is 48. We condition the neural radiance field through the latent generated after the quantization step in the latent generator. We use a Latent routing module consisting of two linear layers of hidden-size 256. As illustrated in Figure 4, the output of the linear layer in the routing module conditions the first two layers of the radiance field network. We employ the losses outlined in Section 5. On each scene, we train our approach for \\(150k\\) iterations. We use Adam [25] optimizer with a learning rate of \\(1e^{-6}\\). Further details are provided in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.683, + 0.74 + ], + "angle": 0, + "content": "6.2. Evaluation Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Most of the radiance field methods evaluate their results on the synthetic (Blender) and real-world (LLFF) datasets proposed in NeRF [34]. These scenes either include a solitary object on a white background or a frontal view of a natural scene. According to our description of stratified scenes, these datasets has only one level. Even large-scale reconstruction datasets like TanksandTemples [26] are not representative of our setting as they only have views either inside or outside of the structure. Similarly, Scannet [9] a dataset for real-world interior scenes, lacks the characteristics of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17607" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.088, + 0.873, + 0.102 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation on test-set against baselines discussed in Section 6.1. Each column is depicts the best and second best." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.102, + 0.895, + 0.342 + ], + "angle": 0, + "content": "
Cube-Sphere-Monkey
Level 0Level 1Level2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]28.33140.93830.103418.18060.49760.498122.11780.59950.382522.87660.67840.3280
mip360 [2]28.31490.92980.115619.04430.53430.493024.91360.73260.324524.09090.73220.3110
Plenoxels [64]25.35470.91690.123813.11480.33200.689521.55680.65230.380320.00870.63370.3979
Instant-NGP [35]28.21040.91680.112314.36480.18300.721617.69140.27440.599720.08890.45810.4779
TensoRF [6]32.00770.95320.069213.74870.15370.710613.00750.24960.688619.58800.45210.4894
Ours26.93350.92980.125525.70880.77380.295926.19120.81720.254926.27780.84030.2254
Bhutanese House
Level 0Level 1Level 2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]11.44780.69170.371117.12090.58860.707818.39180.69520.659115.65350.65850.5793
mip360 [2]26.62400.90020.206224.59460.72960.473929.42250.85770.415626.88040.82910.3652
Plenoxels [64]15.22050.77520.305213.03860.46700.670319.30500.58190.588615.85470.60800.5214
Instant-NGP [35]23.97910.92170.150024.73160.70090.423727.66170.81360.378625.45750.81210.3174
TensoRF [6]13.88800.76070.314217.02440.48560.642116.81700.63060.633215.90980.62560.5298
Ours27.68420.90460.204524.91800.73710.461629.46460.85750.417227.35560.83310.3611
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.352, + 0.873, + 0.366 + ], + "angle": 0, + "content": "Table 4. Quantitative evaluation on test-set against baselines discussed in Section 6.1. Each column is depicts the best and second best." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.366, + 0.892, + 0.605 + ], + "angle": 0, + "content": "
Coffee Shop
Level 0Level 1Level2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]06.74460.61970.469816.13980.49150.798212.88890.42130.815811.92440.51080.6946
mip360 [2]26.20730.88250.186727.05000.80860.378534.20230.93620.195029.15320.87570.2534
Plenoxels [64]19.32040.79680.257912.38710.40440.690422.43250.68560.458518.04670.62890.4689
Instant-NGP [35]29.94250.93240.099228.10400.81930.345229.65740.86800.262129.23470.87320.2355
TensoRF [6]33.03370.94350.069219.31150.53310.658021.18520.71690.459424.51020.73120.3955
Ours26.44990.88020.193928.63920.84030.345033.26920.92540.224329.45280.88190.2544
Dragon In Pyramid
Level 0Level 1Level 2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]14.64050.65950.380020.83680.60520.6856---17.73860.63230.5328
mip360 [2]30.87580.90060.136724.38900.70540.5163---27.63240.80300.3265
Plenoxels [64]13.06670.62470.421714.51260.35720.6498---13.78960.491005358
Instant-NGP [35]23.90540.90100.094924.73890.65940.4664---24.32220.78020.2807
TensoRF [6]35.30150.96320.041419.55730.52210.6809---27.42940.74270.3611
Ours29.47730.87000.169926.17220.74890.4573---27.82480.80950.3136
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.72 + ], + "angle": 0, + "content": "a stratified dataset. Because of the direct unavailability of stratified scenes, we built our own dataset that replicates the intended \"stratified\" scenario. We create a synthetic scene dataset using a mesh-editing software Blender [8] and real scene dataset by altering RealEstate10K dataset which was proposed for the camera localization task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The proposed synthetic dataset has two important variations based on: (a) the number of stratified levels and (b) the geometric complexity. We classify based on the geometry's complexity as follows: (a) Simple Scenes: Stratified scenes using geometric components such as the sphere, cube, and so on; and (b) Complex Scenes: Stratified scenes that mimic real-world scenes. For Simple Scenes, we leverage models and textures provided by Blender [8]. We utilized publicly available graphical models and composited them to create a real-world configuration for Complex scenes. For example, to design the \"Coffee shop\" scene, we selected a building" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.842 + ], + "angle": 0, + "content": "structure for the outer level and walls and glasses for the intermediate level structure. For the core level, we composited elements such as a cash register, coffee cups, and so on to simulate a real-world coffee-shop scene. To avoid photo-metric changes, we use fixed illumination. For each stratified level, the camera settings: field of vision and focal length are fixed. Each scene is rendered at \\(200 \\times 200\\) resolution. The camera viewpoint are sampled evenly from the curved surface of a hemisphere and then randomly divided into train, validation, and test sets. Inner objects in Simple Scenes are rendered from the surface of a sphere. Figure 5 depicts the proposed dataset's skeletal meshes. Further information on dataset is present in Appendix B in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "RealEstate10K dataset. We extracted four scenes \"Spanish Colonial Retreat in Scottsdale Arizona\", \"139 Barton Avenue Toronto Ontario\", \"31 Brian Dr Rochester NY\"" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.47, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.507 + ], + "angle": 0, + "content": "Figure 6. (From top to bottom) Qualitative results on the proposed synthetic datasets (Figure 5). Each row represents a novel view from a level of the stratified scene. The ground-truth (GT) is shown in Column 1. Compared to baselines (Column 2-4), our method's (Column 5) renderings are more consistent to GT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.61 + ], + "angle": 0, + "content": "and “7 Rutledge Ave Highland Mills” from RealEstate10K dataset. We manually inspected and removed regions which had dynamic components in them. More details about converting RealEstaet10k dataset for our stratified setting is provided in Appendix C the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.199, + 0.635 + ], + "angle": 0, + "content": "6.3. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.673 + ], + "angle": 0, + "content": "We present quantitative and qualitative analysis of Strata-NeRF on the datasets described in Section 6.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare our model with NeRF [34], mip360 [2], Instant-NGP [35], TensoRF [6] and Plenoxels [64]. We chose Plenoxels [64] for comparison because it uses sparse-voxel representation which already discretizes the continuous 3D space, which can be useful in stratified scenes. It is worth noting that the sizes of the synthetic scenes in our dataset differ. As a consequence, the authors' recommended configuration file did not produce the optimal results. As a result, we modified the configuration files for unbounded scenes released by the creators of mip360 [2] to improve performance. For Instant-NGP [35], TensoRF [6] and Plenoxels [64], we change the hyperparameters like bound and scale as suggested in the official implementatinos. More information is in Appendix D in the supplementary material. Table 2 provides an overview" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.088, + 0.892, + 0.114 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of our model and baseline on \"139 Barton Avenue\" scene of RealEstate10K dataset." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.115, + 0.892, + 0.179 + ], + "angle": 0, + "content": "
MetricsLevel 0Level 1Level 2Level 3Level 4Level 5
mip360[2]PSNR ↑18.08616.49624.45920.86217.47910.999
SSIM ↑0.6180.5950.7710.7020.5840.409
OursPSNR ↑23.16421.66525.23624.15622.87925.409
SSIM ↑0.8260.7570.7890.7910.7530.782
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.192, + 0.892, + 0.219 + ], + "angle": 0, + "content": "Table 6. Quantitative comparison of our model and mip360 baseline on Six Layer Scene." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.219, + 0.892, + 0.284 + ], + "angle": 0, + "content": "
DatasetLevelsmip360 [2]Oursmip360 [2]Ours
Spanish Colonial Retreat520.10622.5140.6220.685
31 Brian Dr Rochester423.27328.0260.7150.835
139 Barton Avenue618.99123.4330.6420.780
7 Rutledge Ave719.62125.0400.5660.791
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.311, + 0.587, + 0.324 + ], + "angle": 0, + "content": "of baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Quantitative Results. Table 3 & 4 shows the average PSNR, SSIM [60] and LPIPS [67] for each stratified level in unseen test views. We find that our method surpasses other methods across all metrics most of the time. The baseline mip360 [2] works fine for the exterior structure but fails for the inner layers in the \"Cube-Sphere-Monkey\" scene. Strata-NeRF, on the other hand, offers superior metrics at all stratified levels. The baseline models do well in the outer scene but perform sub-optimally in the inner levels, especially in level 1. These outcomes demonstrate that our method outperforms the baseline models significantly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Table 6 shows the summary of average PSNR and SSIM for all the levels in a scene for RealEstate10K dataset. In this case, we only compare our method with mip360 as it is the best performing one among others on the synthetic dataset. We observe that our method outperforms the baseline method in all scenarios. Further, we present level-wise result for a specific scene in Table 5. We observe that for real datasets with increase in number of levels, the magnitude of performance improvement increases, which demonstrates the effectiveness of the proposed approach. Further, we also compare Instant-NGP [35] and TensoRF [6] on a RealEstate10K scene in Appendix E.2 in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results. Figure 6 & 8 depicts the qualitative results for the synthetic dataset scenes described in Section 6.2. We observe that NeRF [34] performs poorly regardless in majority of scenarios. The generated novel views for \"Coffee Shop\" are poor. It only works well in level 0 of \"Cube-Sphere-Monkey\" dataset. mip360 [2] outperforms NeRF but falters in level 1. Furthermore, in level 0 of the \"Cube-Sphere-Monkey\" dataset, mip360 only generates a white patch with no visible structure. For RealEstate10K dataset, it can be observed in Figure 7 that mip360 generates blurry results compared to our approach. Further, we find that our approach generates consistent and structurally salient novel views throughout all levels and scenes. We show qualitative results for Instant-NGP [35] and Ten" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17609" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.348, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.15, + 0.348, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.211, + 0.348, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.272, + 0.348, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.09, + 0.619, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.15, + 0.619, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.211, + 0.619, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.272, + 0.619, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.091, + 0.889, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.15, + 0.889, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.211, + 0.889, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.272, + 0.889, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.336, + 0.893, + 0.379 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison on Scenes from RealEstate10K dataset between mip360 (left image) and our method Strata-NeRF (right image) in a pair. Each row represents a scene in RealEstate10K and each pair represents a level in that scene. Our method outperforms and produces good quality novel views compared to mip360." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.403, + 0.47, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.597 + ], + "angle": 0, + "content": "Figure 8. (From top to bottom) Qualitative results on the proposed synthetic datasets. Each row represents a novel view from each level of the stratified scene. The ground-truth view is shown in Column 1. Compared to prior works (Column 2-4) our method's (Column 5) renderings are more similar to the ground-truth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.462, + 0.641 + ], + "angle": 0, + "content": "soRF [6] in Appendix E.2 in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.468, + 0.822 + ], + "angle": 0, + "content": "Worst Case Analysis. When comparing different methods, average metrics are often insufficient to determine which method is superior to the others. As we have observed in Figure 9 that the baseline method fails on some of test images, hence we also compare the methods in worst care scenarios. The worst-case analysis describes a method's worst performance on the dataset. The worst case analysis is particularly useful to detect the shortcomings of the methods. We present analysis in two categories: (a) histogram distribution for each metric on the test set, and (b) qualitative comparison of the worst-case scenario for our method on PSNR metric." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Figure 9 compares PSNR histogram plots on test-set views for the \"Cube-Sphere-Monkey\" scene. We can see that the mip360 approach performs poorly on PSNR and ranks low on practically all stratification levels. This supports our argument that the mip360 approach produces ar" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.403, + 0.885, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.508, + 0.885, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.634, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Figure 9. (Top Row) Comparison of histogram plots for the test-set for PSNR on \"Cube-Sphere-Monkey\". Note how distribution of our our method is always towards the right compared to other methods. \\( x - axis \\) denote metric value and \\( y - axis \\) denotes the frequency. A qualitative comparison of our method's worst-case PSNR results. PSNR is present at the bottom of the result image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.794 + ], + "angle": 0, + "content": "tifacts in such stratified scenes. For our method, the PSNR distributions are on the right. This implies that the novel views on test-set from our method will not be having serious artifacts in most cases, demonstrating its reliability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Images in Figure 9 depict the qualitative results for the worst-case PSNR instances. All methods perform well in level 0. Hence, we are discussing interior levels which are level 1 and level 2. Other approaches fail in the worst-case scenario for our method at level 1. The outputs from NeRF, mip360 and Plenoxel are visually impaired. At level 2, our method has less blur compared to other approaches. These" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17610" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.349, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.089, + 0.619, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.09, + 0.888, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.221, + 0.89, + 0.25 + ], + "angle": 0, + "content": "Figure 10. Novel-views from different levels of 'Real Estate Video Tour 7 Rutledge Ave Highland Mills NY 10930 Orange County NY' scene in Real Estate 10K dataset. The two rows are from two-different view-points." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.27, + 0.468, + 0.297 + ], + "angle": 0, + "content": "Table 7. Quantitative comparison of our model and baseline on Synthetic Six Layer Scene." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.297, + 0.47, + 0.364 + ], + "angle": 0, + "content": "
MetricsLevel 0Level 1Level 2Level 3Level 4Level 5
mip360 [2]PSNR ↑22.21516.18315.08412.01221.81321.539
SSIM ↑0.7770.4420.5100.3440.8170.647
OursPSNR ↑23.88921.44921.45624.09528.28321.898
SSIM ↑0.8330.6810.6850.7220.8830.686
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.379, + 0.462, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Figure 11. Comparisons of different codebook size on \"Dragon in Pyramid\" scene for different vector-codebook sizes. Note at size=1024 we achieve the best results with less artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.468, + 0.612 + ], + "angle": 0, + "content": "findings demonstrate that our method is better suited to represent stratified scenes than others." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Ablation Studies. To analyse our proposed method, we present an ablation on the size of the vector codebook in our latent generator. Table 8 shows the ablation for the size of the vector codebook on the \"Coffee Shop\" dataset. We trialed with codebook sizes of 512, 1024 and 4096. We found that size 1024 provides optimal performance. As shown in Figure 11, increasing the codebook size induces haziness in the generated novel views, while decreasing the size creates white artifacts in level 0. As a result, we fix the size 1024 for all of our synthetic experiments. Whereas for RealEstate10K dataset we find that codebook size of 4096 produces the optimal tradeoff of results across levels, as it contains more number of levels and details. We further discuss the key architectural design choices for Latent Generator and Latent Router modules in Appendix E.5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "No. of levels: To further test the efficacy of our method on higher number of levels, we created a \"Simple Geometry\" scene consisting of primitive geometry shapes like cube and spheres. More details are in the supplementary" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.27, + 0.892, + 0.296 + ], + "angle": 0, + "content": "Table 8. Quantitative results on \"Cube-Sphere-Monkey\" scene for ablation on size of the vector codebook in Latent Generator." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.296, + 0.892, + 0.36 + ], + "angle": 0, + "content": "
SizePSNR ↑SSIM ↑LPIPS ↓
Level 0Level 1Level 0Level 1Level 0Level 1
51229.545826.34970.87430.73950.16750.4899
102429.483426.17150.87010.74890.13670.5163
409628.460927.82740.86280.73420.17760.5027
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.522 + ], + "angle": 0, + "content": "material. Table 7 displays the results for both the baseline and our approach across a six levels stratified scene. The average PSNR/SSIM for the mip360 baseline is \\(15.35 / 0.487\\), while our method achieved PSNR/SSIM of \\(23.54 / 0.754\\) which improves PSNR and SSIM by \\(53.35\\%\\) and \\(54.83\\%\\) respectively. This shows that our method performs better on increasing number of levels when compared with the baseline method. These observations also hold true for scenes in the RealEstate10K dataset as shown in Table 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.539, + 0.618, + 0.554 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.837 + ], + "angle": 0, + "content": "In this work, we focus on the problem of modelling the 3D representation of a stratified and hierarchical scene, implicitly through a single neural field. For this, we propose Strata-NeRF, which models scenes with stratified structures by introducing a VQ-VAE-based latent generator to implicitly learn the distribution of latent space of input 3D locations and condition the neural radiance field with the latent code generated from this distribution. We also introduce a new synthetic dataset with stratified-level scenes and use it to analyse various existing approaches. Through quantitative, qualitative, and worst-case analysis on this dataset, we show that Strata-NeRF has a more stable 3D representation than the other methods. Further, the improvements due to Strata-NeRF also generalize to real-world RealEstate10K dataset, where it outperforms baselines by a significant margin establishing a new state-of-the-art. We believe designing a new volume rendering equation for modelling complex stratified scenes is a good direction for future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by Samsung R&D Institute India, Bangalore, PMRF and Kotak IISc AI-ML Centre (KIAC). Srinath Sridhar was partly supported by NSF grant CNS-2038897" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17611" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[3] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[4] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.479 + ], + "angle": 0, + "content": "[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision (ECCV). Springer, 2022. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[8] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.468, + 0.718 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.761 + ], + "angle": 0, + "content": "[10] Abe Davis, Marc Levoy, and Fredo Durand. Unstructured light fields. In Computer Graphics Forum, volume 31. Wiley Online Library, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.468, + 0.831 + ], + "angle": 0, + "content": "[11] Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. Fov-nerf: Foveated neural radiance fields for virtual reality. IEEE Transactions on Visualization and Computer Graphics, 28(11), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[12] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W Taylor, and Joshua M Susskind. Unconstrained scene generation with locally conditioned radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[14] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (ICCV). IEEE Computer Society, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[16] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[17] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[18] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[19] Karol Gregor, George Papamakarios, Frederic Besse, Lars Buesing, and Theophane Weber. Temporal difference variational auto-encoder. arXiv preprint arXiv:1806.03107, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[21] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[23] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3), 1984. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.895, + 0.803 + ], + "angle": 0, + "content": "[24] Takuhiro Kaneko. Ar-nerf: Unsupervised learning of depth and defocus effects from natural images with aperture rendering neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17612" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[27] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.471, + 0.205 + ], + "angle": 0, + "content": "[28] Chaojian Li, Sixu Li, Yang Zhao, Wenbo Zhu, and Yingyan Lin. Rt-nerf: Real-time on-device neural radiance fields towards immersive ar/vr rendering. In Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[29] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[30] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[31] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[33] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (CVPR), June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision (ECCV). Springer, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.471, + 0.691 + ], + "angle": 0, + "content": "[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4), 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[37] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Brualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[40] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[43] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagon, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[44] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[45] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[46] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[48] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[49] Harry Shum and Sing Bing Kang. Review of image-based rendering techniques. In Visual Communications and Image Processing 2000, volume 4067. SPIE, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[50] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17613" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.118 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.203 + ], + "angle": 0, + "content": "[52] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[53] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[54] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[55] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[56] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[57] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (CVPR). IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[58] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[59] Xin Wang, Shinji Takaki, Junichi Yamagishi, Simon King, and Keiichi Tokuda. A vector quantized variational autoencoder (vq-vae) autoregressive neural \\( f_{-}0 \\) model for statistical parametric speech synthesis. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[60] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4), 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[61] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[62] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[63] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[64] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.276, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[66] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.332, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17614" + } + ] +] \ No newline at end of file diff --git a/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/246a74f2-93a7-4d68-8843-6238c9108476_origin.pdf b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/246a74f2-93a7-4d68-8843-6238c9108476_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f61fec4d88756c1b3ffb9c6248f39f18756ea44b --- /dev/null +++ b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/246a74f2-93a7-4d68-8843-6238c9108476_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb550d6b93860a6c8af25126248540725400827b5b9528f39236a1282606236b +size 2844719 diff --git a/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/full.md b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..582b9ec2bce8ca456ace03961cb57f6a7f9f79a8 --- /dev/null +++ b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/full.md @@ -0,0 +1,366 @@ +# Strata-NeRF : Neural Radiance Fields for Stratified Scenes + +Ankit Dhiman $^{1,2}$ R Srinath $^{1}$ Harsh Rangwani $^{1}$ Rishubh Parihar $^{1}$ + +Lokesh R Boregowda2 Srinath Sridhar3 R Venkatesh Babu1 + +$^{1}$ Vision and AI Lab, IISc Bangalore $^{2}$ Samsung R & D Institute India - Bangalore $^{3}$ Brown University + +# Abstract + +Neural Radiance Field (NeRF) approaches learn the underlying 3D representation of a scene and generate photorealistic novel views with high fidelity. However, most proposed settings concentrate on modelling a single object or a single level of a scene. However, in the real world, we may capture a scene at multiple levels, resulting in a layered capture. For example, tourists usually capture a monument's exterior structure before capturing the inner structure. Modelling such scenes in 3D with seamless switching between levels can drastically improve immersive experiences. However, most existing techniques struggle in modelling such scenes. We propose Strata-NeRF, a single neural radiance field that implicitly captures a scene with multiple levels. Strata-NeRF achieves this by conditioning the NeRFs on Vector Quantized (VQ) latent representations which allow sudden changes in scene structure. We evaluate the effectiveness of our approach in multi-layered synthetic dataset comprising diverse scenes and then further validate its generalization on the real-world RealEstate10K dataset. We find that Strata-NeRF effectively captures stratified scenes, minimizes artifacts, and synthesizes high-fidelity views compared to existing approaches. https://ankitatiisc.github.io/Strata-NeRF/ + +# 1. Introduction + +Novel view synthesis is an ill-posed problem widely encountered in various areas such as augmented reality [24, 28], virtual reality [11], etc. A paradigm change for solving these kinds of problems was brought by the introduction of Neural Radiance Fields (NeRF) [34]. NeRFs are neural networks that take in the spatial coordinates and camera parameters as input and output the corresponding radiance field. Earlier version of NeRFs enable the generation of high-fidelity novel views for bounded scenes, significantly improving over existing techniques like Structure From Motion [47]. Further, the capability of NeRFs have been recently extended to model unbounded scenes by Mip-NeRF 360 [2]. This enabled NeRFs to model complex real-world + +![](images/aa84ea35264191c360c3f75145a50d663499eca8f61cd710e513babfd5d20007.jpg) +Figure 1. Top, wireframe view of a multi-layered stratified scene with three levels (monkey head inside sphere inside a cube). The camera colors indicate views of a specific level. Strata-NeRF enables high-quality reconstruction of such stratified scenes using a single neural network. + +scenes, where the scene content can exist at any distance from the camera. + +However, similar to unboundedness in scenes, hierarchies in scenes are also natural. For example, images captured in a house can be categorized into images captured outside and inside across various rooms. Modelling such hierarchical scenes jointly for all levels through a NeRF could be particularly useful in cases of Virtual Reality applications. As it would not require switching to a different NeRF for each level, reducing memory requirement and latency in switching. Further, as the different hierarchies of a scene usually share texture and architectural commonalities, it could lead to effective knowledge sharing and reduce the requirement of training independent models. For tackling the above novel objective, we introduce a paradigm of scenes that can be deconstructed into several tiers, termed "Stratified Scenes". A "stratified" scene has several levels or + +![](images/2a1547f58edfe6ef92fa832071ab103ef30211b20adf9dc42edb6a8d75c33ca3.jpg) +Figure 2. Novel views for stratified scene in Figure 1, from Mip-NeRF 360 [2] (left) and our method "Strata-NeRF" (right). Existing methods struggle to capture stratified scenes with a single network while ours produces sharp results. + +groupings of structure (Figure 1). In our work, we first propose a synthetic dataset of stratified scenes, i.e. scenes having multiple levels. This dataset comprises scenes from two categories: (i) Simpler geometry, such as spheres, cubes, or tetrahedron meshes, and (ii) Complex geometry, which closely emulates a real-world setup. + +On such datasets, we find methods such as Mip-NeRF 360 perform well for each level of the hierarchy independently, but produce unsatisfactory results when images from all hierarchical levels are used together for training (Figure 3). This can be attributed to the continuous nature of NeRFs, which is unsuitable for modelling the sudden changes in scenes with shifts in hierarchical levels. Hence, in this work, we introduce Strata-NeRF that explicitly aims to model the hierarchies by conditioning [22, 38, 39, 66, 43] the NeRF on Vector Quantized (VQ) latents. The VQ latents enable the modelling of discontinuities and sudden changes in the scene, as they are discrete and less correlated with others [56]. In practice, the VQ conditioning is achieved by introducing two lightweight modules: the "Latent Generator" module that compresses the implicit information in encoded 3D positions to generate VQ latent code, which is directed through the "Latent Routing" module to condition various layers of radiance field. The additional parameters introduced through these modules are significantly less than training an independent NeRF model for each level, leading to a significant reduction in memory. + +For evaluating the proposed Strata-NeRF we first test on the proposed synthetic Stratified Scenes dataset, where we find that Strata-NeRF learns the structure in scenes across all levels. In contrast, other baselines produce cloudy and sub-optimal novel views (Figure 2). Further, to test the generalizability of the proposed method on real-world scenes, we utilize the high-resolution RealEstate10K dataset. We find that Strata-NeRF significantly outperforms other baselines and produces high-fidelity novel views without artifacts compared to baselines. This is also observed quantitatively through improvement in metrics, where it establishes + +a new state-of-the-art. In summary, + +- We first introduce the task of implicit representation for 3D stratified (hierarchical) scenes using a single radiance field network. For this, we introduce a novel synthetic dataset comprising of scenes ranging from simple to complex geometries. +- For implicit modelling of the stratified scenes, we propose Strata-NeRF, which conditions the radiance field based on discrete Vector-Quantized (VQ) latents to model the sudden changes in scenes due to change in hierarchical level (i.e. strata). +- Strata-NeRF significantly outperforms the baselines across the synthetic dataset and generalizes well on the real-world scene dataset of RealState10k. + +# 2. Related Work + +Generating photo-realistic novel views from densely sampled images is a classical problem. Earlier methods solved this issue using light-field-based interpolation techniques [10, 18, 27]. These techniques interpreted the input images as 2D slices of a 4D function - the light field. The only caveat in these methods is their overreliance on dense views. Another popular technique is Structure From Motion (SFM) which reconstructs 3D structure of a scene or an object by using a sequence of 2D images. We suggest readers to read survey papers [47, 37] to understand SFM methods in detail. Shum et al. [49] also provides an excellent review on traditional image based rendering techniques. + +Neural Volume Reconstruction. NeRF [34] has shown remarkable results in encoding the 3D geometry of a scene implicitly using the multi-layer perceptron (MLP). Specifically, it trains an MLP, which takes 3D position and a viewing direction to predict colour and occupancy. Many papers have extended this idea to solve different scenarios such as dynamic scenes, low-light scenes, synthesis from fewer views, accelerating the performance etc. Mip-Nerf [1] mitigates the problem of aliasing when a novel view is generated at a different resolution. MVSNeRF [7] generalizes across all the scenes and optimizes the geometry and radiance field using only a few views. NerfingMVS [61] utilizes conventional SFM reconstruction and learning-based priors to predict the radiance field. UNISURF [36] combines implicit surface models and radiance fields to render both surface and volume rendering. + +AR-NeRF [24] replaced pin-hole based camera ray-tracing with aperture camera based ray-tracing. DiVeR [62] uses a voxel based representation to learn the radiance field, Mip-NeRF 360 [2] improves view synthesis on the unbounded scenes and also proposed an online distillation scheme which significantly reduced the training and inference time. Neural Rays [31] solves the occlusion problem by predicting the visibility of the 3D points in their + +representation. Scene Representation Transformers [46] uses Vision Transformers [13] to infer latent representations to render the novel views. Further, many methods [30, 17, 44, 65, 51, 21, 58] have been proposed to improve the slow training and inference time for neural radiance field based methods. Despite many works, no work has focused on modelling the stratified scenes. + +NeRF Extensions. Relighting discusses how to model different types of light and then using this model to relight a scene [32, 3, 50, 57, 20]. Breaking the myth that radiance field can only be used in small and bounded scenes, recent methods [52, 55, 45] have scaled it to large-scale city scenes. Another line of work focuses on modelling the dynamic scenes with presence of moving objects [38, 63, 29, 41, 14, 54, 16, 39] through NeRFs. + +Neural Radiance Fields and Latents. Recently, a lot of methods have made use of the latents to bring generative capabilities to neural radiance fields. GRAF [48] uses disentangled shape and appearance latent codes to generalize on an object category. For viewpoint invariance, they used typical GAN based training. Pi-GAN [5] uses volumetric rendering equations for consistent 3D views in a generative framework. Pixel-NeRF [66] learns a scene prior to generalize across different scenes. GSN [12] decomposes the radiance field of a scene into local radiance fields by conditioning on a 2D grid of latent codes. Code-NeRF [22] learns the variation of object shapes and textures across by learning separate latent embeddings. LOLNeRF [43] uses a shared latent space which conditions a neural radinace field to model shape and appearance of a single class. PixNerF [4] extends Pi-GAN [5] and maps images to a latent manifold allowing object-centric novel views given a single image of an object. NeRF-W [33] optimizes latent codes to model the scene variations to produce temporally consistent novel view renderings. In contrast to these methods, we propose conditioning NeRF on learnable Vector Quantized latents. + +Vector Quantized Variational Autoencoders (VQ-VAE) [56]: VQ-VAE uses vector quantization to represent a discrete latent distribution. VQ-VAE has shown applications in Image Generation [42, 40], speech and audio processing [19, 59]. Further, it's extension like VQ-VAE2 [42] uses hierarchical latent space for high-quality generation. + +# 3. Preliminaries + +NeRF represents a scene as an implicit function $f: (X, d) \to (c, \sigma)$ which maps a 3D position $X = (x, y, z)$ and $d = (\theta, \phi)$ to a color $c = (r, g, b)$ and occupancy density $\sigma$ . An MLP parametrizes this implicit function $f$ . Before sending the inputs $X$ and $d$ through the network, a positional encoding is used to project them in a high dimensional space [53]. Finally, the volume rendering [23] procedure enables NeRF to represent scenes with photo-realistic + +![](images/b9854a1c6399d6dd8bc8df78c224f813bfd07a8756cd7494c667923fec50ead5.jpg) + +![](images/d6a9596238b154f66487c5f6de6d25364118e8870ecb4c4b4c54065c7a7c95cf.jpg) + +![](images/49bfdba3fe4114185e2110448895684aa537079a13db16091e8459f04c4edbce.jpg) + +![](images/5c7bf12b3d5137ba0c8d9cd2238b449574259832c763ddae7a2dfad346d2a123.jpg) +Level 0 + +![](images/f1180bdc06949be623b7d97c2c831c0d2627052aa4b09d0b26e82c0dff063219.jpg) + +![](images/2a45ebf2d9b025f5284e9b6ed34e21eeb9833855b41958b33e68fdc45fd9cac8.jpg) +GT + +![](images/75c95094e4fdd136e8abb46fce225d7da6f355a9ff71d2933a2536f064c152bf.jpg) +mip360 (level-wise) + +![](images/4c638ba9307783c8a7f14efca3e9c245677b18b8e9d9177d1359f06d4a987301.jpg) +mip360 +Figure 3. Analysis on "Dragon in pyramid" scene. The top row shows the layout of the levels in 3D scene. Observe that baseline works fine on the scenes when trained individually. Artefacts occur when the baseline is trained on views from the entire scene. + +![](images/efb310eb32c6c7d39c19763434e4960798e8a2250a249151dcc6504f7ae86cac.jpg) + +rendering from novel camera viewpoints. + +Volume Rendering. At the crux of NeRF lies the volume rendering equation. A ray $r(t) = o + td$ is cast from the camera center $o$ through the pixel along direction $d$ . The pixel's color value is estimated by integrating along the ray $r(t)$ as described in Eq. 1 + +$$ +c (r) = \int_ {t _ {n}} ^ {t _ {f}} T (t) \sigma (r (t)) c (r (t), d) d t \tag {1} +$$ + +where transmittance $T(t) = \exp(-\int_{t_n}^t \sigma(r(s)) ds)$ is the probability that a ray passes unhindered from the near plane $(t_n)$ to plane $(t)$ and use this probability to integrate till far plane $(t_f)$ . In Mip-NeRF [1], a ray $r(t)$ is divided into intervals $T_i = [t_i, t_{i+1})$ which corresponds to a conical frustum. For each interval $T_i$ , it computes the mean and variance $(\mu, \Sigma)$ and uses it for integrated position encoding as illustrated in Eq. 2. + +$$ +\gamma (\mu , \Sigma) = \left\{\left[ \begin{array}{l} \sin (2 ^ {l} \mu) \exp (- 2 ^ {2 l - 1} \operatorname {d i a g} (\Sigma)) \\ \cos (2 ^ {l} \mu) \exp (- 2 ^ {2 l - 1} \operatorname {d i a g} (\Sigma)) \end{array} \right] \right\} _ {0} ^ {L - 1} \tag {2} +$$ + +This solves the aliasing issue in the original NeRF. Mip-NeRF 360 [2] proposed coarse-to-fine online distillation for proposal sampling, which efficiently reduces the training time as the proposed MLP only predicts density. They also proposed ray parametrization and regularisation techniques to alleviate hanging artifacts in unbounded scenes. We'll refer Mip-NeRF 360 [2] as mip360 in all our discussions. We choose mip360 [2] as the baseline for all our experiments. + +Table 1. A quantitative comparison of mip360 (level-wise) and mip360 (all views) on "Dragon in pyramid" scene. + +
Level 0Level 1
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
mip360 (level-wise)31.53900.91810.130429.85600.81330.3484
mip36030.88470.90060.136724.38760.70550.5163
+ +# 4. Motivation + +The majority of real-world scenarios are stratified with multiple levels. For example, a commodity store has exterior and interior structures. This work addresses an essential question for such stratified scenes: Can a single radiance field learn such hierarchical scenes? This section introduces and discusses our observations on one such stratified scene: "Dragon in Pyramid", as illustrated in Figure 3. The outer structure of "Dragon in Pyramid" is a Mayan pyramid that has a dragon inside it. To validate our claim, we first train the baseline model on each level, i.e., on outer pyramid views and inner views (focusing dragon) independently. We refer to these separately trained models as mip360 (levelwise). Then, we train a single mip360 model using the outer and inner views for the scene. The term "level" in our work refers to each level in a stratified scene. In the scene depicted in Figure 3, level 0 denotes the pyramid's outer construction, while level 1 denotes the pyramid's interior structure, which contains a dragon. + +Table 1 shows that the baseline model performs remarkably well when trained separately on each level. In comparison, the metric values for the baseline model trained jointly on both levels of stratified scene declines. PSNR at level 1 is $24.39dB$ , a $5.47dB$ reduction compared to mip360 (level-wise). Similarly, performance in level 0 has declined, but less dramatically than in the inner level. This pattern is observed across all metrics. Furthermore, the qualitative results illustrated in Figure 3 backs up the quantitative study's findings. Figure 3 indicates that mip360 (level-wise) generates novel views on par with the ground truth. However, shown in Figure 3, the jointly trained model has white artifacts on the pyramid's outer structure and haziness in front of the dragon inside the pyramid. This demonstrates that current radiance field networks have issues while learning a 3D representation of a stratified scene. We perform a similar experiment for a RealEstate10K scene in Appendix E.1 in the supplementary material. + +# 5. Method + +This section describes our method: Strata-NeRF for stratified scenes. We generate latent codes with the latent generator described in Section 5.1. This latent code is fed into the radiance field architecture through the latent router, described in Section 5.2. Figure 4 depicts the overall architecture of Strata-NeRF. We adopt the base neural radiance + +field architecture proposed in mip360 [2]. + +# 5.1. Latent Generator + +A latent space reflects the scene's "compressed" representation. It has been shown in various works that this space has rich properties. VQ-VAE [56] learns a codebook to model the discrete distribution of the latent space of a variational-autoencoder. The encoder's output is compared to all of the vectors in the codebook. The nearest vector is fed into the decoder as input. Since most data in the world is discrete, VQ based models have been highly successful in image generation [15], speech encoding [56], and other applications. In a stratified scene, the definition of level is also discrete. Hence, our method employs VQ-VAE as a latent generator because of their proven success in representing discrete distributions. + +We use Integrated Positional Encoded (IPE) [2] $\gamma (\mathbf{x})$ as input to our latent generator. We encode $\gamma (\mathbf{x})$ and then search the codebook for the closest vector. After that, the closest vector from the codebook is used to condition the radiance field network. Specifically, $\gamma (\mathbf{x})$ is passed through a set of two hidden layers to generate an encoded input $\mathbf{z}$ . The encoded latent code $\mathbf{z}$ is then passed through the quantizer bottleneck to determine the quantized latent code $\mathbf{z}_{\mathrm{e}}$ where $\mathbf{z}_{\mathrm{e}}\in E$ ; where $E\in R^{N\times D}$ is the codebook; $N$ is the number of vectors in the codebook, and $D$ is the dimension of the latent space. $\mathbf{z}_{\mathrm{e}}$ is then supplied into the decoder network, which consists of two hidden layers, to yield $\mathbf{y}$ as the reconstructed output of $\gamma (\mathbf{x})$ . The quantized latent $\mathbf{z}_{\mathrm{e}}$ is also sent into the radiance field network through the "Latent Router" block. Loss for this variational autoencoder (VAE) block is defined as follows: + +$$ +\mathcal {L} _ {v q} = \left| \left| \gamma (\mathbf {x}) - \mathbf {y} \right| \right| _ {2} ^ {2} + \left| \left| s g \left(\mathbf {z} _ {\mathbf {e}}\right) - \mathbf {z} \right| \right| _ {2} ^ {2} + \beta \left| \left| \mathbf {z} _ {\mathbf {e}} - s g (\mathbf {z}) \right| \right| _ {2} ^ {2} \tag {3} +$$ + +The "Latent Generator" module based on VAE is jointly trained with the NeRF through backpropagation. + +# 5.2. Latent Router + +The Latent Router block is inspired by the CodeNeRF architecture [22], in which shape and texture latent codes are sent to the NeRF MLP through a residual connection. In our architecture, the quantized latent codes $z_{e}$ that are generated in the "Latent Generator" block are input to the Radiance field after passing through an MLP layer in the Latent Router as shown in Figure 4. + +# 5.3. Training Strata-NeRF + +For training Strata-NeRF, we utilize the losses suggested by mip360 [2] as we use a similar radiance field design. $\mathcal{L}_{\text{recon}}(c(r,t),c^{*}(r))$ denotes the reconstruction loss between the estimated colour along a ray and the actual colour value. $\mathcal{L}_{\text{dist}}(s,w)$ is the distortion loss where $s$ is the normalized ray distances and $w$ is the weight vector. Note that + +(a) +![](images/effc3552e5db8befdd6ace230bb6508ca83ce1e803fbadd50c9e854b5f9e2427.jpg) +Radiance field Latent Router LatentGenerator VectorCodebook 3D point + +![](images/f82905e3db6a328b7c99eb4b1856abb6cc7d7c7bfe12007b50242160bd91306b.jpg) +(b) + +Table 2. Characteristic Comparison of the proposed methods + +
MethodDiscrete RepresentationPhotometric LossesVAE loss
NeRF [34]X✓X
mip360 [2]X✓X
Plenoxel[64]✓✓X
Instant-NGP[35]✓✓X
TensoRF[6]✓✓X
Ours✓✓✓
+ +we don't alter anything in the proposal MLP. More details are provided in mip360 [2]. The total loss for Strata-NeRF is given as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {r e c o n}} (c (r, t), c ^ {*} (r)) + \lambda_ {1} \mathcal {L} _ {\text {d i s t}} (s, w) + \lambda_ {2} \mathcal {L} _ {v q} \tag {4} +$$ + +We use $\lambda_{1} = 0.01$ , $\lambda_{2} = 0.1$ and $\beta = 1.0$ across all our experiments, as they work robustly [2] for Strata-NeRF. + +# 6. Experiments + +We discuss implementation details in Section 6.1. Section 6.2 discusses the dataset used for evaluating our method with other baselines. In Section 6.3, we present quantitative and qualitative comparison with the baseline methods. Additionally, we discuss the ablations for the proposed method. + +# 6.1. Implementation Details + +Our method builds on mip360 [2] as the base radiance field. We use a latent generator network which consists of an encoder-decoder architecture and a vector-codebook. The encoder has two linear layers of hidden size 48, and the decoder has one linear layer of hidden size 96. The output dimension of our decoder matches the output from Integrated Positional Encoding (IPE) block. The size of + +![](images/771cc9a9ac616270b2333e7d5f4876af7c98ab201f37d60ae097ed00f791970a.jpg) +Figure 4. For each $3D$ point along the projected ray, we generate a latent code using our "Latent generator" module. The generated latent code is routed to the MLP using "Latent Router". Vector Codebooks learn the discrete distribution of positionally encoded $3D$ points. (a) Our model's end-to-end architecture; (b) components of the "Latent Generator" and "Latent Router" blocks. + +![](images/1cc12f92d5e5142b4798012e2865807b9c640e4b7fd26d1dd4562ee495335ed7.jpg) +Figure 5. Skeleton mesh of the stratified scenes : Bhutanese House and Coffee Shop. More details are in the supplementary material. + +our codebook is 1024, and the dimension of each vector in the codebook is 48. We condition the neural radiance field through the latent generated after the quantization step in the latent generator. We use a Latent routing module consisting of two linear layers of hidden-size 256. As illustrated in Figure 4, the output of the linear layer in the routing module conditions the first two layers of the radiance field network. We employ the losses outlined in Section 5. On each scene, we train our approach for $150k$ iterations. We use Adam [25] optimizer with a learning rate of $1e^{-6}$ . Further details are provided in supplementary material. + +# 6.2. Evaluation Dataset + +Most of the radiance field methods evaluate their results on the synthetic (Blender) and real-world (LLFF) datasets proposed in NeRF [34]. These scenes either include a solitary object on a white background or a frontal view of a natural scene. According to our description of stratified scenes, these datasets has only one level. Even large-scale reconstruction datasets like TanksandTemples [26] are not representative of our setting as they only have views either inside or outside of the structure. Similarly, Scannet [9] a dataset for real-world interior scenes, lacks the characteristics of + +Table 3. Quantitative evaluation on test-set against baselines discussed in Section 6.1. Each column is depicts the best and second best. + +
Cube-Sphere-Monkey
Level 0Level 1Level2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]28.33140.93830.103418.18060.49760.498122.11780.59950.382522.87660.67840.3280
mip360 [2]28.31490.92980.115619.04430.53430.493024.91360.73260.324524.09090.73220.3110
Plenoxels [64]25.35470.91690.123813.11480.33200.689521.55680.65230.380320.00870.63370.3979
Instant-NGP [35]28.21040.91680.112314.36480.18300.721617.69140.27440.599720.08890.45810.4779
TensoRF [6]32.00770.95320.069213.74870.15370.710613.00750.24960.688619.58800.45210.4894
Ours26.93350.92980.125525.70880.77380.295926.19120.81720.254926.27780.84030.2254
Bhutanese House
Level 0Level 1Level 2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]11.44780.69170.371117.12090.58860.707818.39180.69520.659115.65350.65850.5793
mip360 [2]26.62400.90020.206224.59460.72960.473929.42250.85770.415626.88040.82910.3652
Plenoxels [64]15.22050.77520.305213.03860.46700.670319.30500.58190.588615.85470.60800.5214
Instant-NGP [35]23.97910.92170.150024.73160.70090.423727.66170.81360.378625.45750.81210.3174
TensoRF [6]13.88800.76070.314217.02440.48560.642116.81700.63060.633215.90980.62560.5298
Ours27.68420.90460.204524.91800.73710.461629.46460.85750.417227.35560.83310.3611
+ +Table 4. Quantitative evaluation on test-set against baselines discussed in Section 6.1. Each column is depicts the best and second best. + +
Coffee Shop
Level 0Level 1Level2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]06.74460.61970.469816.13980.49150.798212.88890.42130.815811.92440.51080.6946
mip360 [2]26.20730.88250.186727.05000.80860.378534.20230.93620.195029.15320.87570.2534
Plenoxels [64]19.32040.79680.257912.38710.40440.690422.43250.68560.458518.04670.62890.4689
Instant-NGP [35]29.94250.93240.099228.10400.81930.345229.65740.86800.262129.23470.87320.2355
TensoRF [6]33.03370.94350.069219.31150.53310.658021.18520.71690.459424.51020.73120.3955
Ours26.44990.88020.193928.63920.84030.345033.26920.92540.224329.45280.88190.2544
Dragon In Pyramid
Level 0Level 1Level 2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]14.64050.65950.380020.83680.60520.6856---17.73860.63230.5328
mip360 [2]30.87580.90060.136724.38900.70540.5163---27.63240.80300.3265
Plenoxels [64]13.06670.62470.421714.51260.35720.6498---13.78960.491005358
Instant-NGP [35]23.90540.90100.094924.73890.65940.4664---24.32220.78020.2807
TensoRF [6]35.30150.96320.041419.55730.52210.6809---27.42940.74270.3611
Ours29.47730.87000.169926.17220.74890.4573---27.82480.80950.3136
+ +a stratified dataset. Because of the direct unavailability of stratified scenes, we built our own dataset that replicates the intended "stratified" scenario. We create a synthetic scene dataset using a mesh-editing software Blender [8] and real scene dataset by altering RealEstate10K dataset which was proposed for the camera localization task. + +The proposed synthetic dataset has two important variations based on: (a) the number of stratified levels and (b) the geometric complexity. We classify based on the geometry's complexity as follows: (a) Simple Scenes: Stratified scenes using geometric components such as the sphere, cube, and so on; and (b) Complex Scenes: Stratified scenes that mimic real-world scenes. For Simple Scenes, we leverage models and textures provided by Blender [8]. We utilized publicly available graphical models and composited them to create a real-world configuration for Complex scenes. For example, to design the "Coffee shop" scene, we selected a building + +structure for the outer level and walls and glasses for the intermediate level structure. For the core level, we composited elements such as a cash register, coffee cups, and so on to simulate a real-world coffee-shop scene. To avoid photo-metric changes, we use fixed illumination. For each stratified level, the camera settings: field of vision and focal length are fixed. Each scene is rendered at $200 \times 200$ resolution. The camera viewpoint are sampled evenly from the curved surface of a hemisphere and then randomly divided into train, validation, and test sets. Inner objects in Simple Scenes are rendered from the surface of a sphere. Figure 5 depicts the proposed dataset's skeletal meshes. Further information on dataset is present in Appendix B in the supplementary material. + +RealEstate10K dataset. We extracted four scenes "Spanish Colonial Retreat in Scottsdale Arizona", "139 Barton Avenue Toronto Ontario", "31 Brian Dr Rochester NY" + +![](images/b06b5d6c90a66e7dfd9bfef3c88ab39e11f9df4d8c9a90404484986078f1202d.jpg) +Figure 6. (From top to bottom) Qualitative results on the proposed synthetic datasets (Figure 5). Each row represents a novel view from a level of the stratified scene. The ground-truth (GT) is shown in Column 1. Compared to baselines (Column 2-4), our method's (Column 5) renderings are more consistent to GT. + +and “7 Rutledge Ave Highland Mills” from RealEstate10K dataset. We manually inspected and removed regions which had dynamic components in them. More details about converting RealEstaet10k dataset for our stratified setting is provided in Appendix C the supplementary material. + +# 6.3. Evaluation + +We present quantitative and qualitative analysis of Strata-NeRF on the datasets described in Section 6.2. + +Baselines. We compare our model with NeRF [34], mip360 [2], Instant-NGP [35], TensoRF [6] and Plenoxels [64]. We chose Plenoxels [64] for comparison because it uses sparse-voxel representation which already discretizes the continuous 3D space, which can be useful in stratified scenes. It is worth noting that the sizes of the synthetic scenes in our dataset differ. As a consequence, the authors' recommended configuration file did not produce the optimal results. As a result, we modified the configuration files for unbounded scenes released by the creators of mip360 [2] to improve performance. For Instant-NGP [35], TensoRF [6] and Plenoxels [64], we change the hyperparameters like bound and scale as suggested in the official implementatinos. More information is in Appendix D in the supplementary material. Table 2 provides an overview + +Table 5. Quantitative comparison of our model and baseline on "139 Barton Avenue" scene of RealEstate10K dataset. + +
MetricsLevel 0Level 1Level 2Level 3Level 4Level 5
mip360[2]PSNR ↑18.08616.49624.45920.86217.47910.999
SSIM ↑0.6180.5950.7710.7020.5840.409
OursPSNR ↑23.16421.66525.23624.15622.87925.409
SSIM ↑0.8260.7570.7890.7910.7530.782
+ +Table 6. Quantitative comparison of our model and mip360 baseline on Six Layer Scene. + +
DatasetLevelsmip360 [2]Oursmip360 [2]Ours
Spanish Colonial Retreat520.10622.5140.6220.685
31 Brian Dr Rochester423.27328.0260.7150.835
139 Barton Avenue618.99123.4330.6420.780
7 Rutledge Ave719.62125.0400.5660.791
+ +of baselines. + +Quantitative Results. Table 3 & 4 shows the average PSNR, SSIM [60] and LPIPS [67] for each stratified level in unseen test views. We find that our method surpasses other methods across all metrics most of the time. The baseline mip360 [2] works fine for the exterior structure but fails for the inner layers in the "Cube-Sphere-Monkey" scene. Strata-NeRF, on the other hand, offers superior metrics at all stratified levels. The baseline models do well in the outer scene but perform sub-optimally in the inner levels, especially in level 1. These outcomes demonstrate that our method outperforms the baseline models significantly. + +Table 6 shows the summary of average PSNR and SSIM for all the levels in a scene for RealEstate10K dataset. In this case, we only compare our method with mip360 as it is the best performing one among others on the synthetic dataset. We observe that our method outperforms the baseline method in all scenarios. Further, we present level-wise result for a specific scene in Table 5. We observe that for real datasets with increase in number of levels, the magnitude of performance improvement increases, which demonstrates the effectiveness of the proposed approach. Further, we also compare Instant-NGP [35] and TensoRF [6] on a RealEstate10K scene in Appendix E.2 in the supplementary material. + +Qualitative Results. Figure 6 & 8 depicts the qualitative results for the synthetic dataset scenes described in Section 6.2. We observe that NeRF [34] performs poorly regardless in majority of scenarios. The generated novel views for "Coffee Shop" are poor. It only works well in level 0 of "Cube-Sphere-Monkey" dataset. mip360 [2] outperforms NeRF but falters in level 1. Furthermore, in level 0 of the "Cube-Sphere-Monkey" dataset, mip360 only generates a white patch with no visible structure. For RealEstate10K dataset, it can be observed in Figure 7 that mip360 generates blurry results compared to our approach. Further, we find that our approach generates consistent and structurally salient novel views throughout all levels and scenes. We show qualitative results for Instant-NGP [35] and Ten + +![](images/2335e406acbed83df77332ccf297d06596559f5bb93dac890cb5f3e10746c33e.jpg) + +![](images/f426afd44ec0bf60187004a390613f1001ee1c09fdb430a56739cc36d84a93a3.jpg) + +![](images/5e4656c1cfd6f9dfe597ceeef25b4fd8f5f64a0a9bbbc0f3a30d3085f1a4c1fb.jpg) + +![](images/4ae91a27972f11a0b86516b95e70e7d57aac74d12e4014096eb94e889e4a4c7c.jpg) +Figure 7. Qualitative comparison on Scenes from RealEstate10K dataset between mip360 (left image) and our method Strata-NeRF (right image) in a pair. Each row represents a scene in RealEstate10K and each pair represents a level in that scene. Our method outperforms and produces good quality novel views compared to mip360. + +![](images/2baa77f5b93102d0f175d88813483383a65b80626d21fe9f5a44280b9782b7a4.jpg) + +![](images/ecf8f8534a7b07fb158e7934c144407d555ddb097fed468452b42547a9f3d549.jpg) + +![](images/32c82bd0a4851286daf03538d9f8698e363965b7cfe8af18fc9ecd1775fc41c7.jpg) + +![](images/6127d0447afda68b414c7a13dcc74ce31a0bf66885585e1da7a2e869b0e1a25a.jpg) + +![](images/abf5469b15af913778d4be15a7e44a40a2054607abfb8fc03d0d932f9d862174.jpg) + +![](images/f6f10cb0051e30406cb83fe1d8216ea67940c0538558953ee11a9d3d62346aaa.jpg) + +![](images/716157bd223321246c710b06174958f422890355552f41e8f843d26776d52d08.jpg) + +![](images/142e4544b3e2651b67d58c8ab97caff3dd544e2bf77d79272bc02e907c5dc69a.jpg) + +![](images/33db4324fb50753a7c500b03ea06727ff5089286259a789363ddac3577b06c8f.jpg) +Figure 8. (From top to bottom) Qualitative results on the proposed synthetic datasets. Each row represents a novel view from each level of the stratified scene. The ground-truth view is shown in Column 1. Compared to prior works (Column 2-4) our method's (Column 5) renderings are more similar to the ground-truth. + +soRF [6] in Appendix E.2 in the supplementary material. + +Worst Case Analysis. When comparing different methods, average metrics are often insufficient to determine which method is superior to the others. As we have observed in Figure 9 that the baseline method fails on some of test images, hence we also compare the methods in worst care scenarios. The worst-case analysis describes a method's worst performance on the dataset. The worst case analysis is particularly useful to detect the shortcomings of the methods. We present analysis in two categories: (a) histogram distribution for each metric on the test set, and (b) qualitative comparison of the worst-case scenario for our method on PSNR metric. + +Figure 9 compares PSNR histogram plots on test-set views for the "Cube-Sphere-Monkey" scene. We can see that the mip360 approach performs poorly on PSNR and ranks low on practically all stratification levels. This supports our argument that the mip360 approach produces ar + +![](images/24e3e768dfcc5b33ab923fe9b86f2bd66b6ce0fe9f27d539c776850633a18f69.jpg) + +![](images/da75ae1f9def0ad2b00312b8fac541c3c3c93e9cf7237eb18794aa914fd2955c.jpg) +Figure 9. (Top Row) Comparison of histogram plots for the test-set for PSNR on "Cube-Sphere-Monkey". Note how distribution of our our method is always towards the right compared to other methods. $x - axis$ denote metric value and $y - axis$ denotes the frequency. A qualitative comparison of our method's worst-case PSNR results. PSNR is present at the bottom of the result image. + +tifacts in such stratified scenes. For our method, the PSNR distributions are on the right. This implies that the novel views on test-set from our method will not be having serious artifacts in most cases, demonstrating its reliability. + +Images in Figure 9 depict the qualitative results for the worst-case PSNR instances. All methods perform well in level 0. Hence, we are discussing interior levels which are level 1 and level 2. Other approaches fail in the worst-case scenario for our method at level 1. The outputs from NeRF, mip360 and Plenoxel are visually impaired. At level 2, our method has less blur compared to other approaches. These + +![](images/631f443465fa0af41048ac563ee11f3e596718971ad8f0c9ac4c1783c8588d57.jpg) +Figure 10. Novel-views from different levels of 'Real Estate Video Tour 7 Rutledge Ave Highland Mills NY 10930 Orange County NY' scene in Real Estate 10K dataset. The two rows are from two-different view-points. + +![](images/85acfe2cb876e987d62c3b129cd1ead3ac06a733cf0722afe41f04726714f496.jpg) + +![](images/79208b58ad8e7ad9ed3112b3f81cd95dc5e6988c81d18c94161352fde2a8c63e.jpg) + +Table 7. Quantitative comparison of our model and baseline on Synthetic Six Layer Scene. + +
MetricsLevel 0Level 1Level 2Level 3Level 4Level 5
mip360 [2]PSNR ↑22.21516.18315.08412.01221.81321.539
SSIM ↑0.7770.4420.5100.3440.8170.647
OursPSNR ↑23.88921.44921.45624.09528.28321.898
SSIM ↑0.8330.6810.6850.7220.8830.686
+ +![](images/31aebb8b14c59cfadc4e06a15aa2d4207d8883b9ec8c4bf66dee0faf531acbcd.jpg) +Figure 11. Comparisons of different codebook size on "Dragon in Pyramid" scene for different vector-codebook sizes. Note at size=1024 we achieve the best results with less artifacts. + +findings demonstrate that our method is better suited to represent stratified scenes than others. + +Ablation Studies. To analyse our proposed method, we present an ablation on the size of the vector codebook in our latent generator. Table 8 shows the ablation for the size of the vector codebook on the "Coffee Shop" dataset. We trialed with codebook sizes of 512, 1024 and 4096. We found that size 1024 provides optimal performance. As shown in Figure 11, increasing the codebook size induces haziness in the generated novel views, while decreasing the size creates white artifacts in level 0. As a result, we fix the size 1024 for all of our synthetic experiments. Whereas for RealEstate10K dataset we find that codebook size of 4096 produces the optimal tradeoff of results across levels, as it contains more number of levels and details. We further discuss the key architectural design choices for Latent Generator and Latent Router modules in Appendix E.5. + +No. of levels: To further test the efficacy of our method on higher number of levels, we created a "Simple Geometry" scene consisting of primitive geometry shapes like cube and spheres. More details are in the supplementary + +Table 8. Quantitative results on "Cube-Sphere-Monkey" scene for ablation on size of the vector codebook in Latent Generator. + +
SizePSNR ↑SSIM ↑LPIPS ↓
Level 0Level 1Level 0Level 1Level 0Level 1
51229.545826.34970.87430.73950.16750.4899
102429.483426.17150.87010.74890.13670.5163
409628.460927.82740.86280.73420.17760.5027
+ +material. Table 7 displays the results for both the baseline and our approach across a six levels stratified scene. The average PSNR/SSIM for the mip360 baseline is $15.35 / 0.487$ , while our method achieved PSNR/SSIM of $23.54 / 0.754$ which improves PSNR and SSIM by $53.35\%$ and $54.83\%$ respectively. This shows that our method performs better on increasing number of levels when compared with the baseline method. These observations also hold true for scenes in the RealEstate10K dataset as shown in Table 5. + +# 7. Conclusion + +In this work, we focus on the problem of modelling the 3D representation of a stratified and hierarchical scene, implicitly through a single neural field. For this, we propose Strata-NeRF, which models scenes with stratified structures by introducing a VQ-VAE-based latent generator to implicitly learn the distribution of latent space of input 3D locations and condition the neural radiance field with the latent code generated from this distribution. We also introduce a new synthetic dataset with stratified-level scenes and use it to analyse various existing approaches. Through quantitative, qualitative, and worst-case analysis on this dataset, we show that Strata-NeRF has a more stable 3D representation than the other methods. Further, the improvements due to Strata-NeRF also generalize to real-world RealEstate10K dataset, where it outperforms baselines by a significant margin establishing a new state-of-the-art. We believe designing a new volume rendering equation for modelling complex stratified scenes is a good direction for future work. + +Acknowledgement. This work was supported by Samsung R&D Institute India, Bangalore, PMRF and Kotak IISc AI-ML Centre (KIAC). Srinath Sridhar was partly supported by NSF grant CNS-2038897 + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 9 +[3] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 3 +[4] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision (ECCV). Springer, 2022. 5, 6, 7, 8 +[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[8] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6 +[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5 +[10] Abe Davis, Marc Levoy, and Fredo Durand. Unstructured light fields. In Computer Graphics Forum, volume 31. Wiley Online Library, 2012. 2 +[11] Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. Fov-nerf: Foveated neural radiance fields for virtual reality. IEEE Transactions on Visualization and Computer Graphics, 28(11), 2022. 1 +[12] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W Taylor, and Joshua M Susskind. Unconstrained scene generation with locally conditioned radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 + +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[14] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (ICCV). IEEE Computer Society, 2021. 3 +[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4 +[16] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[17] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[18] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2 +[19] Karol Gregor, George Papamakarios, Frederic Besse, Lars Buesing, and Theophane Weber. Temporal difference variational auto-encoder. arXiv preprint arXiv:1806.03107, 2018. 3 +[20] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[21] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 4 +[23] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3), 1984. 3 +[24] Takuhiro Kaneko. Ar-nerf: Unsupervised learning of depth and defocus effects from natural images with aperture rendering neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5 +[26] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 5 + +[27] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2 +[28] Chaojian Li, Sixu Li, Yang Zhao, Wenbo Zhu, and Yingyan Lin. Rt-nerf: Real-time on-device neural radiance fields towards immersive ar/vr rendering. In Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design, 2022. 1 +[29] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[30] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3 +[31] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[33] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (CVPR), June 2021. 3 +[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision (ECCV). Springer, 2020. 1, 2, 5, 6, 7 +[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4), 2022. 5, 6, 7 +[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[37] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26, 2017. 2 +[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3 +[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin- + +Brualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 3 +[40] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3 +[43] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagon, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[44] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[45] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[46] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2 +[48] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3 +[49] Harry Shum and Sing Bing Kang. Review of image-based rendering techniques. In Visual Communications and Image Processing 2000, volume 4067. SPIE, 2000. 2 +[50] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[51] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[52] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[53] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3 +[54] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[55] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[56] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 3, 4 +[57] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (CVPR). IEEE, 2022. 3 +[58] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[59] Xin Wang, Shinji Takaki, Junichi Yamagishi, Simon King, and Keiichi Tokuda. A vector quantized variational autoencoder (vq-vae) autoregressive neural $f_{-}0$ model for statistical parametric speech synthesis. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28, 2019. 3 +[60] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4), 2004. 7 +[61] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[62] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 + +[63] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[64] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 5, 6, 7 +[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[66] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3 +[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7 \ No newline at end of file diff --git a/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/images.zip b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0e87ce46a794800f1f3b9ff2c683d64a7b95b6e7 --- /dev/null +++ b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7851548b6eac7d4aaddc891d934f5b98cc05141e963e54e1202ce0bbf53e4f0b +size 985691 diff --git a/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/layout.json b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e9364527346e93ea90dadc5df7c1756642c03ef9 --- /dev/null +++ b/2023/Strata-NeRF _ Neural Radiance Fields for Stratified Scenes/layout.json @@ -0,0 +1,9759 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 479, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 479, + 119 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 479, + 119 + ], + "type": "text", + "content": "Strata-NeRF : Neural Radiance Fields for Stratified Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 142, + 477, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 142, + 477, + 157 + ], + "spans": [ + { + "bbox": [ + 134, + 142, + 477, + 157 + ], + "type": "text", + "content": "Ankit Dhiman" + }, + { + "bbox": [ + 134, + 142, + 477, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 134, + 142, + 477, + 157 + ], + "type": "text", + "content": " R Srinath" + }, + { + "bbox": [ + 134, + 142, + 477, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 142, + 477, + 157 + ], + "type": "text", + "content": " Harsh Rangwani" + }, + { + "bbox": [ + 134, + 142, + 477, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 142, + 477, + 157 + ], + "type": "text", + "content": " Rishubh Parihar" + }, + { + "bbox": [ + 134, + 142, + 477, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 157, + 460, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 157, + 460, + 171 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 460, + 171 + ], + "type": "text", + "content": "Lokesh R Boregowda2 Srinath Sridhar3 R Venkatesh Babu1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 171, + 559, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 559, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 171, + 559, + 186 + ], + "type": "text", + "content": "Vision and AI Lab, IISc Bangalore " + }, + { + "bbox": [ + 53, + 171, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 171, + 559, + 186 + ], + "type": "text", + "content": "Samsung R & D Institute India - Bangalore " + }, + { + "bbox": [ + 53, + 171, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 171, + 559, + 186 + ], + "type": "text", + "content": "Brown University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 238, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 514 + ], + "type": "text", + "content": "Neural Radiance Field (NeRF) approaches learn the underlying 3D representation of a scene and generate photorealistic novel views with high fidelity. However, most proposed settings concentrate on modelling a single object or a single level of a scene. However, in the real world, we may capture a scene at multiple levels, resulting in a layered capture. For example, tourists usually capture a monument's exterior structure before capturing the inner structure. Modelling such scenes in 3D with seamless switching between levels can drastically improve immersive experiences. However, most existing techniques struggle in modelling such scenes. We propose Strata-NeRF, a single neural radiance field that implicitly captures a scene with multiple levels. Strata-NeRF achieves this by conditioning the NeRFs on Vector Quantized (VQ) latent representations which allow sudden changes in scene structure. We evaluate the effectiveness of our approach in multi-layered synthetic dataset comprising diverse scenes and then further validate its generalization on the real-world RealEstate10K dataset. We find that Strata-NeRF effectively captures stratified scenes, minimizes artifacts, and synthesizes high-fidelity views compared to existing approaches. https://ankitatiisc.github.io/Strata-NeRF/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 537, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 128, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "Novel view synthesis is an ill-posed problem widely encountered in various areas such as augmented reality [24, 28], virtual reality [11], etc. A paradigm change for solving these kinds of problems was brought by the introduction of Neural Radiance Fields (NeRF) [34]. NeRFs are neural networks that take in the spatial coordinates and camera parameters as input and output the corresponding radiance field. Earlier version of NeRFs enable the generation of high-fidelity novel views for bounded scenes, significantly improving over existing techniques like Structure From Motion [47]. Further, the capability of NeRFs have been recently extended to model unbounded scenes by Mip-NeRF 360 [2]. This enabled NeRFs to model complex real-world" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 215, + 533, + 430 + ], + "blocks": [ + { + "bbox": [ + 320, + 215, + 533, + 430 + ], + "lines": [ + { + "bbox": [ + 320, + 215, + 533, + 430 + ], + "spans": [ + { + "bbox": [ + 320, + 215, + 533, + 430 + ], + "type": "image", + "image_path": "aa84ea35264191c360c3f75145a50d663499eca8f61cd710e513babfd5d20007.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 434, + 547, + 491 + ], + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 491 + ], + "type": "text", + "content": "Figure 1. Top, wireframe view of a multi-layered stratified scene with three levels (monkey head inside sphere inside a cube). The camera colors indicate views of a specific level. Strata-NeRF enables high-quality reconstruction of such stratified scenes using a single neural network." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "type": "text", + "content": "scenes, where the scene content can exist at any distance from the camera." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "However, similar to unboundedness in scenes, hierarchies in scenes are also natural. For example, images captured in a house can be categorized into images captured outside and inside across various rooms. Modelling such hierarchical scenes jointly for all levels through a NeRF could be particularly useful in cases of Virtual Reality applications. As it would not require switching to a different NeRF for each level, reducing memory requirement and latency in switching. Further, as the different hierarchies of a scene usually share texture and architectural commonalities, it could lead to effective knowledge sharing and reduce the requirement of training independent models. For tackling the above novel objective, we introduce a paradigm of scenes that can be deconstructed into several tiers, termed \"Stratified Scenes\". A \"stratified\" scene has several levels or" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17603" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 286, + 185 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 185 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 185 + ], + "type": "image", + "image_path": "2a1547f58edfe6ef92fa832071ab103ef30211b20adf9dc42edb6a8d75c33ca3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 187, + 288, + 233 + ], + "lines": [ + { + "bbox": [ + 47, + 187, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 187, + 288, + 233 + ], + "type": "text", + "content": "Figure 2. Novel views for stratified scene in Figure 1, from Mip-NeRF 360 [2] (left) and our method \"Strata-NeRF\" (right). Existing methods struggle to capture stratified scenes with a single network while ours produces sharp results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 327 + ], + "type": "text", + "content": "groupings of structure (Figure 1). In our work, we first propose a synthetic dataset of stratified scenes, i.e. scenes having multiple levels. This dataset comprises scenes from two categories: (i) Simpler geometry, such as spheres, cubes, or tetrahedron meshes, and (ii) Complex geometry, which closely emulates a real-world setup." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 328, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 288, + 581 + ], + "type": "text", + "content": "On such datasets, we find methods such as Mip-NeRF 360 perform well for each level of the hierarchy independently, but produce unsatisfactory results when images from all hierarchical levels are used together for training (Figure 3). This can be attributed to the continuous nature of NeRFs, which is unsuitable for modelling the sudden changes in scenes with shifts in hierarchical levels. Hence, in this work, we introduce Strata-NeRF that explicitly aims to model the hierarchies by conditioning [22, 38, 39, 66, 43] the NeRF on Vector Quantized (VQ) latents. The VQ latents enable the modelling of discontinuities and sudden changes in the scene, as they are discrete and less correlated with others [56]. In practice, the VQ conditioning is achieved by introducing two lightweight modules: the \"Latent Generator\" module that compresses the implicit information in encoded 3D positions to generate VQ latent code, which is directed through the \"Latent Routing\" module to condition various layers of radiance field. The additional parameters introduced through these modules are significantly less than training an independent NeRF model for each level, leading to a significant reduction in memory." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "For evaluating the proposed Strata-NeRF we first test on the proposed synthetic Stratified Scenes dataset, where we find that Strata-NeRF learns the structure in scenes across all levels. In contrast, other baselines produce cloudy and sub-optimal novel views (Figure 2). Further, to test the generalizability of the proposed method on real-world scenes, we utilize the high-resolution RealEstate10K dataset. We find that Strata-NeRF significantly outperforms other baselines and produces high-fidelity novel views without artifacts compared to baselines. This is also observed quantitatively through improvement in metrics, where it establishes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 450, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 450, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 450, + 85 + ], + "type": "text", + "content": "a new state-of-the-art. In summary," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 91, + 547, + 254 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 91, + 545, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 545, + 152 + ], + "type": "text", + "content": "- We first introduce the task of implicit representation for 3D stratified (hierarchical) scenes using a single radiance field network. For this, we introduce a novel synthetic dataset comprising of scenes ranging from simple to complex geometries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 155, + 547, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 155, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 155, + 547, + 215 + ], + "type": "text", + "content": "- For implicit modelling of the stratified scenes, we propose Strata-NeRF, which conditions the radiance field based on discrete Vector-Quantized (VQ) latents to model the sudden changes in scenes due to change in hierarchical level (i.e. strata)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 219, + 545, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 219, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 317, + 219, + 545, + 254 + ], + "type": "text", + "content": "- Strata-NeRF significantly outperforms the baselines across the synthetic dataset and generalizes well on the real-world scene dataset of RealState10k." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 392, + 275 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 283, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 426 + ], + "type": "text", + "content": "Generating photo-realistic novel views from densely sampled images is a classical problem. Earlier methods solved this issue using light-field-based interpolation techniques [10, 18, 27]. These techniques interpreted the input images as 2D slices of a 4D function - the light field. The only caveat in these methods is their overreliance on dense views. Another popular technique is Structure From Motion (SFM) which reconstructs 3D structure of a scene or an object by using a sequence of 2D images. We suggest readers to read survey papers [47, 37] to understand SFM methods in detail. Shum et al. [49] also provides an excellent review on traditional image based rendering techniques." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 426, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 617 + ], + "type": "text", + "content": "Neural Volume Reconstruction. NeRF [34] has shown remarkable results in encoding the 3D geometry of a scene implicitly using the multi-layer perceptron (MLP). Specifically, it trains an MLP, which takes 3D position and a viewing direction to predict colour and occupancy. Many papers have extended this idea to solve different scenarios such as dynamic scenes, low-light scenes, synthesis from fewer views, accelerating the performance etc. Mip-Nerf [1] mitigates the problem of aliasing when a novel view is generated at a different resolution. MVSNeRF [7] generalizes across all the scenes and optimizes the geometry and radiance field using only a few views. NerfingMVS [61] utilizes conventional SFM reconstruction and learning-based priors to predict the radiance field. UNISURF [36] combines implicit surface models and radiance fields to render both surface and volume rendering." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "content": "AR-NeRF [24] replaced pin-hole based camera ray-tracing with aperture camera based ray-tracing. DiVeR [62] uses a voxel based representation to learn the radiance field, Mip-NeRF 360 [2] improves view synthesis on the unbounded scenes and also proposed an online distillation scheme which significantly reduced the training and inference time. Neural Rays [31] solves the occlusion problem by predicting the visibility of the 3D points in their" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17604" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "representation. Scene Representation Transformers [46] uses Vision Transformers [13] to infer latent representations to render the novel views. Further, many methods [30, 17, 44, 65, 51, 21, 58] have been proposed to improve the slow training and inference time for neural radiance field based methods. Despite many works, no work has focused on modelling the stratified scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 251 + ], + "type": "text", + "content": "NeRF Extensions. Relighting discusses how to model different types of light and then using this model to relight a scene [32, 3, 50, 57, 20]. Breaking the myth that radiance field can only be used in small and bounded scenes, recent methods [52, 55, 45] have scaled it to large-scale city scenes. Another line of work focuses on modelling the dynamic scenes with presence of moving objects [38, 63, 29, 41, 14, 54, 16, 39] through NeRFs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 288, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 514 + ], + "type": "text", + "content": "Neural Radiance Fields and Latents. Recently, a lot of methods have made use of the latents to bring generative capabilities to neural radiance fields. GRAF [48] uses disentangled shape and appearance latent codes to generalize on an object category. For viewpoint invariance, they used typical GAN based training. Pi-GAN [5] uses volumetric rendering equations for consistent 3D views in a generative framework. Pixel-NeRF [66] learns a scene prior to generalize across different scenes. GSN [12] decomposes the radiance field of a scene into local radiance fields by conditioning on a 2D grid of latent codes. Code-NeRF [22] learns the variation of object shapes and textures across by learning separate latent embeddings. LOLNeRF [43] uses a shared latent space which conditions a neural radinace field to model shape and appearance of a single class. PixNerF [4] extends Pi-GAN [5] and maps images to a latent manifold allowing object-centric novel views given a single image of an object. NeRF-W [33] optimizes latent codes to model the scene variations to produce temporally consistent novel view renderings. In contrast to these methods, we propose conditioning NeRF on learnable Vector Quantized latents." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 514, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 587 + ], + "type": "text", + "content": "Vector Quantized Variational Autoencoders (VQ-VAE) [56]: VQ-VAE uses vector quantization to represent a discrete latent distribution. VQ-VAE has shown applications in Image Generation [42, 40], speech and audio processing [19, 59]. Further, it's extension like VQ-VAE2 [42] uses hierarchical latent space for high-quality generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 132, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 132, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 132, + 609 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "content": "NeRF represents a scene as an implicit function " + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "f: (X, d) \\to (c, \\sigma)" + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "content": " which maps a 3D position " + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "X = (x, y, z)" + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "d = (\\theta, \\phi)" + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "content": " to a color " + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "c = (r, g, b)" + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "content": " and occupancy density " + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "content": ". An MLP parametrizes this implicit function " + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "content": ". Before sending the inputs " + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 617, + 288, + 713 + ], + "type": "text", + "content": " through the network, a positional encoding is used to project them in a high dimensional space [53]. Finally, the volume rendering [23] procedure enables NeRF to represent scenes with photo-realistic" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 76, + 541, + 135 + ], + "blocks": [ + { + "bbox": [ + 310, + 76, + 541, + 135 + ], + "lines": [ + { + "bbox": [ + 310, + 76, + 541, + 135 + ], + "spans": [ + { + "bbox": [ + 310, + 76, + 541, + 135 + ], + "type": "image", + "image_path": "b9854a1c6399d6dd8bc8df78c224f813bfd07a8756cd7494c667923fec50ead5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 148, + 377, + 194 + ], + "blocks": [ + { + "bbox": [ + 310, + 148, + 377, + 194 + ], + "lines": [ + { + "bbox": [ + 310, + 148, + 377, + 194 + ], + "spans": [ + { + "bbox": [ + 310, + 148, + 377, + 194 + ], + "type": "image", + "image_path": "d6a9596238b154f66487c5f6de6d25364118e8870ecb4c4b4c54065c7a7c95cf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 381, + 148, + 449, + 194 + ], + "blocks": [ + { + "bbox": [ + 381, + 148, + 449, + 194 + ], + "lines": [ + { + "bbox": [ + 381, + 148, + 449, + 194 + ], + "spans": [ + { + "bbox": [ + 381, + 148, + 449, + 194 + ], + "type": "image", + "image_path": "49bfdba3fe4114185e2110448895684aa537079a13db16091e8459f04c4edbce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 453, + 148, + 522, + 194 + ], + "blocks": [ + { + "bbox": [ + 453, + 148, + 522, + 194 + ], + "lines": [ + { + "bbox": [ + 453, + 148, + 522, + 194 + ], + "spans": [ + { + "bbox": [ + 453, + 148, + 522, + 194 + ], + "type": "image", + "image_path": "5c7bf12b3d5137ba0c8d9cd2238b449574259832c763ddae7a2dfad346d2a123.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 530, + 166, + 539, + 196 + ], + "lines": [ + { + "bbox": [ + 530, + 166, + 539, + 196 + ], + "spans": [ + { + "bbox": [ + 530, + 166, + 539, + 196 + ], + "type": "text", + "content": "Level 0" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 529, + 144, + 541, + 160 + ], + "blocks": [ + { + "bbox": [ + 529, + 144, + 541, + 160 + ], + "lines": [ + { + "bbox": [ + 529, + 144, + 541, + 160 + ], + "spans": [ + { + "bbox": [ + 529, + 144, + 541, + 160 + ], + "type": "image", + "image_path": "f1180bdc06949be623b7d97c2c831c0d2627052aa4b09d0b26e82c0dff063219.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 202, + 378, + 265 + ], + "blocks": [ + { + "bbox": [ + 310, + 202, + 378, + 265 + ], + "lines": [ + { + "bbox": [ + 310, + 202, + 378, + 265 + ], + "spans": [ + { + "bbox": [ + 310, + 202, + 378, + 265 + ], + "type": "image", + "image_path": "2a45ebf2d9b025f5284e9b6ed34e21eeb9833855b41958b33e68fdc45fd9cac8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 266, + 354, + 274 + ], + "lines": [ + { + "bbox": [ + 343, + 266, + 354, + 274 + ], + "spans": [ + { + "bbox": [ + 343, + 266, + 354, + 274 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 383, + 202, + 452, + 265 + ], + "blocks": [ + { + "bbox": [ + 383, + 202, + 452, + 265 + ], + "lines": [ + { + "bbox": [ + 383, + 202, + 452, + 265 + ], + "spans": [ + { + "bbox": [ + 383, + 202, + 452, + 265 + ], + "type": "image", + "image_path": "75c95094e4fdd136e8abb46fce225d7da6f355a9ff71d2933a2536f064c152bf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 266, + 449, + 274 + ], + "lines": [ + { + "bbox": [ + 392, + 266, + 449, + 274 + ], + "spans": [ + { + "bbox": [ + 392, + 266, + 449, + 274 + ], + "type": "text", + "content": "mip360 (level-wise)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 457, + 202, + 525, + 265 + ], + "blocks": [ + { + "bbox": [ + 457, + 202, + 525, + 265 + ], + "lines": [ + { + "bbox": [ + 457, + 202, + 525, + 265 + ], + "spans": [ + { + "bbox": [ + 457, + 202, + 525, + 265 + ], + "type": "image", + "image_path": "4c638ba9307783c8a7f14efca3e9c245677b18b8e9d9177d1359f06d4a987301.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 266, + 506, + 274 + ], + "lines": [ + { + "bbox": [ + 482, + 266, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 482, + 266, + 506, + 274 + ], + "type": "text", + "content": "mip360" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 277, + 545, + 321 + ], + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 321 + ], + "type": "text", + "content": "Figure 3. Analysis on \"Dragon in pyramid\" scene. The top row shows the layout of the levels in 3D scene. Observe that baseline works fine on the scenes when trained individually. Artefacts occur when the baseline is trained on views from the entire scene." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 529, + 211, + 541, + 259 + ], + "blocks": [ + { + "bbox": [ + 529, + 211, + 541, + 259 + ], + "lines": [ + { + "bbox": [ + 529, + 211, + 541, + 259 + ], + "spans": [ + { + "bbox": [ + 529, + 211, + 541, + 259 + ], + "type": "image", + "image_path": "efb310eb32c6c7d39c19763434e4960798e8a2250a249151dcc6504f7ae86cac.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 342, + 473, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 473, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 473, + 354 + ], + "type": "text", + "content": "rendering from novel camera viewpoints." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "type": "text", + "content": "Volume Rendering. At the crux of NeRF lies the volume rendering equation. A ray " + }, + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "type": "inline_equation", + "content": "r(t) = o + td" + }, + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "type": "text", + "content": " is cast from the camera center " + }, + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "type": "text", + "content": " through the pixel along direction " + }, + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "type": "text", + "content": ". The pixel's color value is estimated by integrating along the ray " + }, + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "type": "inline_equation", + "content": "r(t)" + }, + { + "bbox": [ + 305, + 355, + 545, + 415 + ], + "type": "text", + "content": " as described in Eq. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 349, + 426, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 426, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 349, + 426, + 545, + 453 + ], + "type": "interline_equation", + "content": "c (r) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (r (t)) c (r (t), d) d t \\tag {1}", + "image_path": "a28d7d4ef2d57024d3f359d8e1e323ef1e8b5e30fd2eb1748cbd664c04ca366c.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "text", + "content": "where transmittance " + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "inline_equation", + "content": "T(t) = \\exp(-\\int_{t_n}^t \\sigma(r(s)) ds)" + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "text", + "content": " is the probability that a ray passes unhindered from the near plane " + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "inline_equation", + "content": "(t_n)" + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "text", + "content": " to plane " + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "inline_equation", + "content": "(t)" + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "text", + "content": " and use this probability to integrate till far plane " + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "inline_equation", + "content": "(t_f)" + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "text", + "content": ". In Mip-NeRF [1], a ray " + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "inline_equation", + "content": "r(t)" + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "text", + "content": " is divided into intervals " + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "inline_equation", + "content": "T_i = [t_i, t_{i+1})" + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "text", + "content": " which corresponds to a conical frustum. For each interval " + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "text", + "content": ", it computes the mean and variance " + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "inline_equation", + "content": "(\\mu, \\Sigma)" + }, + { + "bbox": [ + 305, + 465, + 545, + 562 + ], + "type": "text", + "content": " and uses it for integrated position encoding as illustrated in Eq. 2." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 572, + 545, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 572, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 315, + 572, + 545, + 617 + ], + "type": "interline_equation", + "content": "\\gamma (\\mu , \\Sigma) = \\left\\{\\left[ \\begin{array}{l} \\sin (2 ^ {l} \\mu) \\exp (- 2 ^ {2 l - 1} \\operatorname {d i a g} (\\Sigma)) \\\\ \\cos (2 ^ {l} \\mu) \\exp (- 2 ^ {2 l - 1} \\operatorname {d i a g} (\\Sigma)) \\end{array} \\right] \\right\\} _ {0} ^ {L - 1} \\tag {2}", + "image_path": "35552c1a5ff590dc85d50a67baad2156d5c4a1b617b5b256ad1d3a0f94c24f5e.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "This solves the aliasing issue in the original NeRF. Mip-NeRF 360 [2] proposed coarse-to-fine online distillation for proposal sampling, which efficiently reduces the training time as the proposed MLP only predicts density. They also proposed ray parametrization and regularisation techniques to alleviate hanging artifacts in unbounded scenes. We'll refer Mip-NeRF 360 [2] as mip360 in all our discussions. We choose mip360 [2] as the baseline for all our experiments." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17605" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 91, + 287, + 133 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 91 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 91 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 91 + ], + "type": "text", + "content": "Table 1. A quantitative comparison of mip360 (level-wise) and mip360 (all views) on \"Dragon in pyramid\" scene." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 91, + 287, + 133 + ], + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 133 + ], + "type": "table", + "html": "
Level 0Level 1
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
mip360 (level-wise)31.53900.91810.130429.85600.81330.3484
mip36030.88470.90060.136724.38760.70550.5163
", + "image_path": "ea3c8479e7d5e1dcbc0eff2276525f37083499d707aed10c56d0f8f95ce3db6c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 154, + 119, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 154, + 119, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 119, + 166 + ], + "type": "text", + "content": "4. Motivation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 175, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 175, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 45, + 175, + 287, + 390 + ], + "type": "text", + "content": "The majority of real-world scenarios are stratified with multiple levels. For example, a commodity store has exterior and interior structures. This work addresses an essential question for such stratified scenes: Can a single radiance field learn such hierarchical scenes? This section introduces and discusses our observations on one such stratified scene: \"Dragon in Pyramid\", as illustrated in Figure 3. The outer structure of \"Dragon in Pyramid\" is a Mayan pyramid that has a dragon inside it. To validate our claim, we first train the baseline model on each level, i.e., on outer pyramid views and inner views (focusing dragon) independently. We refer to these separately trained models as mip360 (levelwise). Then, we train a single mip360 model using the outer and inner views for the scene. The term \"level\" in our work refers to each level in a stratified scene. In the scene depicted in Figure 3, level 0 denotes the pyramid's outer construction, while level 1 denotes the pyramid's interior structure, which contains a dragon." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 391, + 288, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 288, + 608 + ], + "type": "text", + "content": "Table 1 shows that the baseline model performs remarkably well when trained separately on each level. In comparison, the metric values for the baseline model trained jointly on both levels of stratified scene declines. PSNR at level 1 is " + }, + { + "bbox": [ + 46, + 391, + 288, + 608 + ], + "type": "inline_equation", + "content": "24.39dB" + }, + { + "bbox": [ + 46, + 391, + 288, + 608 + ], + "type": "text", + "content": ", a " + }, + { + "bbox": [ + 46, + 391, + 288, + 608 + ], + "type": "inline_equation", + "content": "5.47dB" + }, + { + "bbox": [ + 46, + 391, + 288, + 608 + ], + "type": "text", + "content": " reduction compared to mip360 (level-wise). Similarly, performance in level 0 has declined, but less dramatically than in the inner level. This pattern is observed across all metrics. Furthermore, the qualitative results illustrated in Figure 3 backs up the quantitative study's findings. Figure 3 indicates that mip360 (level-wise) generates novel views on par with the ground truth. However, shown in Figure 3, the jointly trained model has white artifacts on the pyramid's outer structure and haziness in front of the dragon inside the pyramid. This demonstrates that current radiance field networks have issues while learning a 3D representation of a stratified scene. We perform a similar experiment for a RealEstate10K scene in Appendix E.1 in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 620, + 102, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 102, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 102, + 633 + ], + "type": "text", + "content": "5. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "This section describes our method: Strata-NeRF for stratified scenes. We generate latent codes with the latent generator described in Section 5.1. This latent code is fed into the radiance field architecture through the latent router, described in Section 5.2. Figure 4 depicts the overall architecture of Strata-NeRF. We adopt the base neural radiance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 477, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 477, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 477, + 85 + ], + "type": "text", + "content": "field architecture proposed in mip360 [2]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 91, + 411, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 411, + 102 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 411, + 102 + ], + "type": "text", + "content": "5.1. Latent Generator" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 110, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 263 + ], + "type": "text", + "content": "A latent space reflects the scene's \"compressed\" representation. It has been shown in various works that this space has rich properties. VQ-VAE [56] learns a codebook to model the discrete distribution of the latent space of a variational-autoencoder. The encoder's output is compared to all of the vectors in the codebook. The nearest vector is fed into the decoder as input. Since most data in the world is discrete, VQ based models have been highly successful in image generation [15], speech encoding [56], and other applications. In a stratified scene, the definition of level is also discrete. Hence, our method employs VQ-VAE as a latent generator because of their proven success in representing discrete distributions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": "We use Integrated Positional Encoded (IPE) [2] " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x})" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " as input to our latent generator. We encode " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x})" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " and then search the codebook for the closest vector. After that, the closest vector from the codebook is used to condition the radiance field network. Specifically, " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x})" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " is passed through a set of two hidden layers to generate an encoded input " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": ". The encoded latent code " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " is then passed through the quantizer bottleneck to determine the quantized latent code " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{e}}" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{e}}\\in E" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": "; where " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "E\\in R^{N\\times D}" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " is the codebook; " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " is the number of vectors in the codebook, and " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " is the dimension of the latent space. " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{e}}" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " is then supplied into the decoder network, which consists of two hidden layers, to yield " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " as the reconstructed output of " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x})" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": ". The quantized latent " + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{e}}" + }, + { + "bbox": [ + 304, + 264, + 546, + 456 + ], + "type": "text", + "content": " is also sent into the radiance field network through the \"Latent Router\" block. Loss for this variational autoencoder (VAE) block is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 463, + 545, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 463, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 310, + 463, + 545, + 478 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v q} = \\left| \\left| \\gamma (\\mathbf {x}) - \\mathbf {y} \\right| \\right| _ {2} ^ {2} + \\left| \\left| s g \\left(\\mathbf {z} _ {\\mathbf {e}}\\right) - \\mathbf {z} \\right| \\right| _ {2} ^ {2} + \\beta \\left| \\left| \\mathbf {z} _ {\\mathbf {e}} - s g (\\mathbf {z}) \\right| \\right| _ {2} ^ {2} \\tag {3}", + "image_path": "1320822cc5fd30cb0c1787a2e61201f6f665ebda31fd4ab43a4cfda31e894925.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 483, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 509 + ], + "type": "text", + "content": "The \"Latent Generator\" module based on VAE is jointly trained with the NeRF through backpropagation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 514, + 395, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 395, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 395, + 526 + ], + "type": "text", + "content": "5.2. Latent Router" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": "The Latent Router block is inspired by the CodeNeRF architecture [22], in which shape and texture latent codes are sent to the NeRF MLP through a residual connection. In our architecture, the quantized latent codes " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": " that are generated in the \"Latent Generator\" block are input to the Radiance field after passing through an MLP layer in the Latent Router as shown in Figure 4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 623, + 436, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 436, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 436, + 635 + ], + "type": "text", + "content": "5.3. Training Strata-NeRF" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "For training Strata-NeRF, we utilize the losses suggested by mip360 [2] as we use a similar radiance field design. " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{recon}}(c(r,t),c^{*}(r))" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " denotes the reconstruction loss between the estimated colour along a ray and the actual colour value. " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{dist}}(s,w)" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is the distortion loss where " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is the normalized ray distances and " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is the weight vector. Note that" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17606" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 403, + 205 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 403, + 205 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 403, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 403, + 205 + ], + "type": "image", + "image_path": "effc3552e5db8befdd6ace230bb6508ca83ce1e803fbadd50c9e854b5f9e2427.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 208, + 235, + 219 + ], + "lines": [ + { + "bbox": [ + 223, + 208, + 235, + 219 + ], + "spans": [ + { + "bbox": [ + 223, + 208, + 235, + 219 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 220, + 536, + 238 + ], + "lines": [ + { + "bbox": [ + 53, + 220, + 536, + 238 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 536, + 238 + ], + "type": "text", + "content": "Radiance field Latent Router LatentGenerator VectorCodebook 3D point" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 408, + 74, + 547, + 204 + ], + "blocks": [ + { + "bbox": [ + 408, + 74, + 547, + 204 + ], + "lines": [ + { + "bbox": [ + 408, + 74, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 408, + 74, + 547, + 204 + ], + "type": "image", + "image_path": "f82905e3db6a328b7c99eb4b1856abb6cc7d7c7bfe12007b50242160bd91306b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 208, + 495, + 219 + ], + "lines": [ + { + "bbox": [ + 482, + 208, + 495, + 219 + ], + "spans": [ + { + "bbox": [ + 482, + 208, + 495, + 219 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 300, + 287, + 393 + ], + "blocks": [ + { + "bbox": [ + 47, + 290, + 268, + 300 + ], + "lines": [ + { + "bbox": [ + 47, + 290, + 268, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 290, + 268, + 300 + ], + "type": "text", + "content": "Table 2. Characteristic Comparison of the proposed methods" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 300, + 287, + 393 + ], + "lines": [ + { + "bbox": [ + 48, + 300, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 287, + 393 + ], + "type": "table", + "html": "
MethodDiscrete RepresentationPhotometric LossesVAE loss
NeRF [34]X✓X
mip360 [2]X✓X
Plenoxel[64]✓✓X
Instant-NGP[35]✓✓X
TensoRF[6]✓✓X
Ours✓✓✓
", + "image_path": "858f690860d1d1badfde3ab2b0f564ae58e6608801ec5f1fb358e02834c24755.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 411, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 448 + ], + "type": "text", + "content": "we don't alter anything in the proposal MLP. More details are provided in mip360 [2]. The total loss for Strata-NeRF is given as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 454, + 287, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 454, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 51, + 454, + 287, + 478 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {r e c o n}} (c (r, t), c ^ {*} (r)) + \\lambda_ {1} \\mathcal {L} _ {\\text {d i s t}} (s, w) + \\lambda_ {2} \\mathcal {L} _ {v q} \\tag {4}", + "image_path": "9eb84d88c072f5547328a4d67b6db92761ec1f191ad2da444ea3a2686a2fad74.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.01" + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.1" + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\beta = 1.0" + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "text", + "content": " across all our experiments, as they work robustly [2] for Strata-NeRF." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 513, + 128, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 128, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 128, + 526 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": "We discuss implementation details in Section 6.1. Section 6.2 discusses the dataset used for evaluating our method with other baselines. In Section 6.3, we present quantitative and qualitative comparison with the baseline methods. Additionally, we discuss the ablations for the proposed method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 611, + 180, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 180, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 180, + 624 + ], + "type": "text", + "content": "6.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "Our method builds on mip360 [2] as the base radiance field. We use a latent generator network which consists of an encoder-decoder architecture and a vector-codebook. The encoder has two linear layers of hidden size 48, and the decoder has one linear layer of hidden size 96. The output dimension of our decoder matches the output from Integrated Positional Encoding (IPE) block. The size of" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 310, + 297, + 381, + 388 + ], + "blocks": [ + { + "bbox": [ + 46, + 240, + 546, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 546, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 546, + 275 + ], + "type": "text", + "content": "Figure 4. For each " + }, + { + "bbox": [ + 46, + 240, + 546, + 275 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 46, + 240, + 546, + 275 + ], + "type": "text", + "content": " point along the projected ray, we generate a latent code using our \"Latent generator\" module. The generated latent code is routed to the MLP using \"Latent Router\". Vector Codebooks learn the discrete distribution of positionally encoded " + }, + { + "bbox": [ + 46, + 240, + 546, + 275 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 46, + 240, + 546, + 275 + ], + "type": "text", + "content": " points. (a) Our model's end-to-end architecture; (b) components of the \"Latent Generator\" and \"Latent Router\" blocks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 297, + 381, + 388 + ], + "lines": [ + { + "bbox": [ + 310, + 297, + 381, + 388 + ], + "spans": [ + { + "bbox": [ + 310, + 297, + 381, + 388 + ], + "type": "image", + "image_path": "771cc9a9ac616270b2333e7d5f4876af7c98ab201f37d60ae097ed00f791970a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 382, + 301, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 382, + 301, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 382, + 301, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 382, + 301, + 545, + 388 + ], + "type": "image", + "image_path": "1cc12f92d5e5142b4798012e2865807b9c640e4b7fd26d1dd4562ee495335ed7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 392, + 545, + 414 + ], + "lines": [ + { + "bbox": [ + 305, + 392, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 545, + 414 + ], + "type": "text", + "content": "Figure 5. Skeleton mesh of the stratified scenes : Bhutanese House and Coffee Shop. More details are in the supplementary material." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": "our codebook is 1024, and the dimension of each vector in the codebook is 48. We condition the neural radiance field through the latent generated after the quantization step in the latent generator. We use a Latent routing module consisting of two linear layers of hidden-size 256. As illustrated in Figure 4, the output of the linear layer in the routing module conditions the first two layers of the radiance field network. We employ the losses outlined in Section 5. On each scene, we train our approach for " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "150k" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": " iterations. We use Adam [25] optimizer with a learning rate of " + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "inline_equation", + "content": "1e^{-6}" + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": ". Further details are provided in supplementary material." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 574, + 417, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 417, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 417, + 586 + ], + "type": "text", + "content": "6.2. Evaluation Dataset" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Most of the radiance field methods evaluate their results on the synthetic (Blender) and real-world (LLFF) datasets proposed in NeRF [34]. These scenes either include a solitary object on a white background or a frontal view of a natural scene. According to our description of stratified scenes, these datasets has only one level. Even large-scale reconstruction datasets like TanksandTemples [26] are not representative of our setting as they only have views either inside or outside of the structure. Similarly, Scannet [9] a dataset for real-world interior scenes, lacks the characteristics of" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17607" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 80, + 547, + 270 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 534, + 80 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 534, + 80 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 534, + 80 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation on test-set against baselines discussed in Section 6.1. Each column is depicts the best and second best." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 80, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 47, + 80, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 80, + 547, + 270 + ], + "type": "table", + "html": "
Cube-Sphere-Monkey
Level 0Level 1Level2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]28.33140.93830.103418.18060.49760.498122.11780.59950.382522.87660.67840.3280
mip360 [2]28.31490.92980.115619.04430.53430.493024.91360.73260.324524.09090.73220.3110
Plenoxels [64]25.35470.91690.123813.11480.33200.689521.55680.65230.380320.00870.63370.3979
Instant-NGP [35]28.21040.91680.112314.36480.18300.721617.69140.27440.599720.08890.45810.4779
TensoRF [6]32.00770.95320.069213.74870.15370.710613.00750.24960.688619.58800.45210.4894
Ours26.93350.92980.125525.70880.77380.295926.19120.81720.254926.27780.84030.2254
Bhutanese House
Level 0Level 1Level 2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]11.44780.69170.371117.12090.58860.707818.39180.69520.659115.65350.65850.5793
mip360 [2]26.62400.90020.206224.59460.72960.473929.42250.85770.415626.88040.82910.3652
Plenoxels [64]15.22050.77520.305213.03860.46700.670319.30500.58190.588615.85470.60800.5214
Instant-NGP [35]23.97910.92170.150024.73160.70090.423727.66170.81360.378625.45750.81210.3174
TensoRF [6]13.88800.76070.314217.02440.48560.642116.81700.63060.633215.90980.62560.5298
Ours27.68420.90460.204524.91800.73710.461629.46460.85750.417227.35560.83310.3611
", + "image_path": "468413363cb4cf1cb4176e0715ac6764594be26d58afd49114d787ebb3f5e675.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 289, + 545, + 479 + ], + "blocks": [ + { + "bbox": [ + 47, + 278, + 534, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 278, + 534, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 534, + 289 + ], + "type": "text", + "content": "Table 4. Quantitative evaluation on test-set against baselines discussed in Section 6.1. Each column is depicts the best and second best." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 289, + 545, + 479 + ], + "lines": [ + { + "bbox": [ + 47, + 289, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 545, + 479 + ], + "type": "table", + "html": "
Coffee Shop
Level 0Level 1Level2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]06.74460.61970.469816.13980.49150.798212.88890.42130.815811.92440.51080.6946
mip360 [2]26.20730.88250.186727.05000.80860.378534.20230.93620.195029.15320.87570.2534
Plenoxels [64]19.32040.79680.257912.38710.40440.690422.43250.68560.458518.04670.62890.4689
Instant-NGP [35]29.94250.93240.099228.10400.81930.345229.65740.86800.262129.23470.87320.2355
TensoRF [6]33.03370.94350.069219.31150.53310.658021.18520.71690.459424.51020.73120.3955
Ours26.44990.88020.193928.63920.84030.345033.26920.92540.224329.45280.88190.2544
Dragon In Pyramid
Level 0Level 1Level 2Total
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Nerf [34]14.64050.65950.380020.83680.60520.6856---17.73860.63230.5328
mip360 [2]30.87580.90060.136724.38900.70540.5163---27.63240.80300.3265
Plenoxels [64]13.06670.62470.421714.51260.35720.6498---13.78960.491005358
Instant-NGP [35]23.90540.90100.094924.73890.65940.4664---24.32220.78020.2807
TensoRF [6]35.30150.96320.041419.55730.52210.6809---27.42940.74270.3611
Ours29.47730.87000.169926.17220.74890.4573---27.82480.80950.3136
", + "image_path": "195f1fd62d3baa16954f87e6e48794556fd36d36714a35f36cfca66db935cd54.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 570 + ], + "type": "text", + "content": "a stratified dataset. Because of the direct unavailability of stratified scenes, we built our own dataset that replicates the intended \"stratified\" scenario. We create a synthetic scene dataset using a mesh-editing software Blender [8] and real scene dataset by altering RealEstate10K dataset which was proposed for the camera localization task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "The proposed synthetic dataset has two important variations based on: (a) the number of stratified levels and (b) the geometric complexity. We classify based on the geometry's complexity as follows: (a) Simple Scenes: Stratified scenes using geometric components such as the sphere, cube, and so on; and (b) Complex Scenes: Stratified scenes that mimic real-world scenes. For Simple Scenes, we leverage models and textures provided by Blender [8]. We utilized publicly available graphical models and composited them to create a real-world configuration for Complex scenes. For example, to design the \"Coffee shop\" scene, we selected a building" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 498, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 666 + ], + "type": "text", + "content": "structure for the outer level and walls and glasses for the intermediate level structure. For the core level, we composited elements such as a cash register, coffee cups, and so on to simulate a real-world coffee-shop scene. To avoid photo-metric changes, we use fixed illumination. For each stratified level, the camera settings: field of vision and focal length are fixed. Each scene is rendered at " + }, + { + "bbox": [ + 304, + 498, + 545, + 666 + ], + "type": "inline_equation", + "content": "200 \\times 200" + }, + { + "bbox": [ + 304, + 498, + 545, + 666 + ], + "type": "text", + "content": " resolution. The camera viewpoint are sampled evenly from the curved surface of a hemisphere and then randomly divided into train, validation, and test sets. Inner objects in Simple Scenes are rendered from the surface of a sphere. Figure 5 depicts the proposed dataset's skeletal meshes. Further information on dataset is present in Appendix B in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "RealEstate10K dataset. We extracted four scenes \"Spanish Colonial Retreat in Scottsdale Arizona\", \"139 Barton Avenue Toronto Ontario\", \"31 Brian Dr Rochester NY\"" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17608" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 287, + 344 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 287, + 344 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 287, + 344 + ], + "type": "image", + "image_path": "b06b5d6c90a66e7dfd9bfef3c88ab39e11f9df4d8c9a90404484986078f1202d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 347, + 287, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 401 + ], + "type": "text", + "content": "Figure 6. (From top to bottom) Qualitative results on the proposed synthetic datasets (Figure 5). Each row represents a novel view from a level of the stratified scene. The ground-truth (GT) is shown in Column 1. Compared to baselines (Column 2-4), our method's (Column 5) renderings are more consistent to GT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 422, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 483 + ], + "type": "text", + "content": "and “7 Rutledge Ave Highland Mills” from RealEstate10K dataset. We manually inspected and removed regions which had dynamic components in them. More details about converting RealEstaet10k dataset for our stratified setting is provided in Appendix C the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 491, + 121, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 121, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 121, + 502 + ], + "type": "text", + "content": "6.3. Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 533 + ], + "type": "text", + "content": "We present quantitative and qualitative analysis of Strata-NeRF on the datasets described in Section 6.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": "Baselines. We compare our model with NeRF [34], mip360 [2], Instant-NGP [35], TensoRF [6] and Plenoxels [64]. We chose Plenoxels [64] for comparison because it uses sparse-voxel representation which already discretizes the continuous 3D space, which can be useful in stratified scenes. It is worth noting that the sizes of the synthetic scenes in our dataset differ. As a consequence, the authors' recommended configuration file did not produce the optimal results. As a result, we modified the configuration files for unbounded scenes released by the creators of mip360 [2] to improve performance. For Instant-NGP [35], TensoRF [6] and Plenoxels [64], we change the hyperparameters like bound and scale as suggested in the official implementatinos. More information is in Appendix D in the supplementary material. Table 2 provides an overview" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 91, + 545, + 141 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 545, + 90 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 545, + 90 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 545, + 90 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of our model and baseline on \"139 Barton Avenue\" scene of RealEstate10K dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 91, + 545, + 141 + ], + "lines": [ + { + "bbox": [ + 307, + 91, + 545, + 141 + ], + "spans": [ + { + "bbox": [ + 307, + 91, + 545, + 141 + ], + "type": "table", + "html": "
MetricsLevel 0Level 1Level 2Level 3Level 4Level 5
mip360[2]PSNR ↑18.08616.49624.45920.86217.47910.999
SSIM ↑0.6180.5950.7710.7020.5840.409
OursPSNR ↑23.16421.66525.23624.15622.87925.409
SSIM ↑0.8260.7570.7890.7910.7530.782
", + "image_path": "4c4bba7a03aca2b49eb795ae930ebed8f0664a8add23eb34043105c273a4f7b3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 173, + 545, + 224 + ], + "blocks": [ + { + "bbox": [ + 305, + 152, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 305, + 152, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 545, + 173 + ], + "type": "text", + "content": "Table 6. Quantitative comparison of our model and mip360 baseline on Six Layer Scene." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 173, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 306, + 173, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 173, + 545, + 224 + ], + "type": "table", + "html": "
DatasetLevelsmip360 [2]Oursmip360 [2]Ours
Spanish Colonial Retreat520.10622.5140.6220.685
31 Brian Dr Rochester423.27328.0260.7150.835
139 Barton Avenue618.99123.4330.6420.780
7 Rutledge Ave719.62125.0400.5660.791
", + "image_path": "b09331081213b13702f15ea35dbcb69581155dea9c29a40fedbea0ba26dcbcef.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 246, + 359, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 246, + 359, + 256 + ], + "spans": [ + { + "bbox": [ + 305, + 246, + 359, + 256 + ], + "type": "text", + "content": "of baselines." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 389 + ], + "type": "text", + "content": "Quantitative Results. Table 3 & 4 shows the average PSNR, SSIM [60] and LPIPS [67] for each stratified level in unseen test views. We find that our method surpasses other methods across all metrics most of the time. The baseline mip360 [2] works fine for the exterior structure but fails for the inner layers in the \"Cube-Sphere-Monkey\" scene. Strata-NeRF, on the other hand, offers superior metrics at all stratified levels. The baseline models do well in the outer scene but perform sub-optimally in the inner levels, especially in level 1. These outcomes demonstrate that our method outperforms the baseline models significantly." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 389, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 544 + ], + "type": "text", + "content": "Table 6 shows the summary of average PSNR and SSIM for all the levels in a scene for RealEstate10K dataset. In this case, we only compare our method with mip360 as it is the best performing one among others on the synthetic dataset. We observe that our method outperforms the baseline method in all scenarios. Further, we present level-wise result for a specific scene in Table 5. We observe that for real datasets with increase in number of levels, the magnitude of performance improvement increases, which demonstrates the effectiveness of the proposed approach. Further, we also compare Instant-NGP [35] and TensoRF [6] on a RealEstate10K scene in Appendix E.2 in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Qualitative Results. Figure 6 & 8 depicts the qualitative results for the synthetic dataset scenes described in Section 6.2. We observe that NeRF [34] performs poorly regardless in majority of scenarios. The generated novel views for \"Coffee Shop\" are poor. It only works well in level 0 of \"Cube-Sphere-Monkey\" dataset. mip360 [2] outperforms NeRF but falters in level 1. Furthermore, in level 0 of the \"Cube-Sphere-Monkey\" dataset, mip360 only generates a white patch with no visible structure. For RealEstate10K dataset, it can be observed in Figure 7 that mip360 generates blurry results compared to our approach. Further, we find that our approach generates consistent and structurally salient novel views throughout all levels and scenes. We show qualitative results for Instant-NGP [35] and Ten" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17609" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 212, + 118 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 212, + 118 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 212, + 118 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 212, + 118 + ], + "type": "image", + "image_path": "2335e406acbed83df77332ccf297d06596559f5bb93dac890cb5f3e10746c33e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 118, + 212, + 166 + ], + "blocks": [ + { + "bbox": [ + 49, + 118, + 212, + 166 + ], + "lines": [ + { + "bbox": [ + 49, + 118, + 212, + 166 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 212, + 166 + ], + "type": "image", + "image_path": "f426afd44ec0bf60187004a390613f1001ee1c09fdb430a56739cc36d84a93a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 167, + 212, + 215 + ], + "blocks": [ + { + "bbox": [ + 49, + 167, + 212, + 215 + ], + "lines": [ + { + "bbox": [ + 49, + 167, + 212, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 167, + 212, + 215 + ], + "type": "image", + "image_path": "5e4656c1cfd6f9dfe597ceeef25b4fd8f5f64a0a9bbbc0f3a30d3085f1a4c1fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 215, + 212, + 262 + ], + "blocks": [ + { + "bbox": [ + 49, + 215, + 212, + 262 + ], + "lines": [ + { + "bbox": [ + 49, + 215, + 212, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 215, + 212, + 262 + ], + "type": "image", + "image_path": "4ae91a27972f11a0b86516b95e70e7d57aac74d12e4014096eb94e889e4a4c7c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 546, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 546, + 300 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison on Scenes from RealEstate10K dataset between mip360 (left image) and our method Strata-NeRF (right image) in a pair. Each row represents a scene in RealEstate10K and each pair represents a level in that scene. Our method outperforms and produces good quality novel views compared to mip360." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 214, + 71, + 378, + 118 + ], + "blocks": [ + { + "bbox": [ + 214, + 71, + 378, + 118 + ], + "lines": [ + { + "bbox": [ + 214, + 71, + 378, + 118 + ], + "spans": [ + { + "bbox": [ + 214, + 71, + 378, + 118 + ], + "type": "image", + "image_path": "2baa77f5b93102d0f175d88813483383a65b80626d21fe9f5a44280b9782b7a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 214, + 118, + 378, + 166 + ], + "blocks": [ + { + "bbox": [ + 214, + 118, + 378, + 166 + ], + "lines": [ + { + "bbox": [ + 214, + 118, + 378, + 166 + ], + "spans": [ + { + "bbox": [ + 214, + 118, + 378, + 166 + ], + "type": "image", + "image_path": "ecf8f8534a7b07fb158e7934c144407d555ddb097fed468452b42547a9f3d549.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 214, + 167, + 378, + 215 + ], + "blocks": [ + { + "bbox": [ + 214, + 167, + 378, + 215 + ], + "lines": [ + { + "bbox": [ + 214, + 167, + 378, + 215 + ], + "spans": [ + { + "bbox": [ + 214, + 167, + 378, + 215 + ], + "type": "image", + "image_path": "32c82bd0a4851286daf03538d9f8698e363965b7cfe8af18fc9ecd1775fc41c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 214, + 215, + 378, + 262 + ], + "blocks": [ + { + "bbox": [ + 214, + 215, + 378, + 262 + ], + "lines": [ + { + "bbox": [ + 214, + 215, + 378, + 262 + ], + "spans": [ + { + "bbox": [ + 214, + 215, + 378, + 262 + ], + "type": "image", + "image_path": "6127d0447afda68b414c7a13dcc74ce31a0bf66885585e1da7a2e869b0e1a25a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 380, + 72, + 544, + 118 + ], + "blocks": [ + { + "bbox": [ + 380, + 72, + 544, + 118 + ], + "lines": [ + { + "bbox": [ + 380, + 72, + 544, + 118 + ], + "spans": [ + { + "bbox": [ + 380, + 72, + 544, + 118 + ], + "type": "image", + "image_path": "abf5469b15af913778d4be15a7e44a40a2054607abfb8fc03d0d932f9d862174.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 380, + 118, + 544, + 166 + ], + "blocks": [ + { + "bbox": [ + 380, + 118, + 544, + 166 + ], + "lines": [ + { + "bbox": [ + 380, + 118, + 544, + 166 + ], + "spans": [ + { + "bbox": [ + 380, + 118, + 544, + 166 + ], + "type": "image", + "image_path": "f6f10cb0051e30406cb83fe1d8216ea67940c0538558953ee11a9d3d62346aaa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 380, + 167, + 544, + 215 + ], + "blocks": [ + { + "bbox": [ + 380, + 167, + 544, + 215 + ], + "lines": [ + { + "bbox": [ + 380, + 167, + 544, + 215 + ], + "spans": [ + { + "bbox": [ + 380, + 167, + 544, + 215 + ], + "type": "image", + "image_path": "716157bd223321246c710b06174958f422890355552f41e8f843d26776d52d08.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 380, + 215, + 544, + 262 + ], + "blocks": [ + { + "bbox": [ + 380, + 215, + 544, + 262 + ], + "lines": [ + { + "bbox": [ + 380, + 215, + 544, + 262 + ], + "spans": [ + { + "bbox": [ + 380, + 215, + 544, + 262 + ], + "type": "image", + "image_path": "142e4544b3e2651b67d58c8ab97caff3dd544e2bf77d79272bc02e907c5dc69a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 50, + 319, + 287, + 415 + ], + "blocks": [ + { + "bbox": [ + 50, + 319, + 287, + 415 + ], + "lines": [ + { + "bbox": [ + 50, + 319, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 50, + 319, + 287, + 415 + ], + "type": "image", + "image_path": "33db4324fb50753a7c500b03ea06727ff5089286259a789363ddac3577b06c8f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 418, + 287, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 472 + ], + "type": "text", + "content": "Figure 8. (From top to bottom) Qualitative results on the proposed synthetic datasets. Each row represents a novel view from each level of the stratified scene. The ground-truth view is shown in Column 1. Compared to prior works (Column 2-4) our method's (Column 5) renderings are more similar to the ground-truth." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 495, + 282, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 282, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 282, + 507 + ], + "type": "text", + "content": "soRF [6] in Appendix E.2 in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 509, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 286, + 651 + ], + "type": "text", + "content": "Worst Case Analysis. When comparing different methods, average metrics are often insufficient to determine which method is superior to the others. As we have observed in Figure 9 that the baseline method fails on some of test images, hence we also compare the methods in worst care scenarios. The worst-case analysis describes a method's worst performance on the dataset. The worst case analysis is particularly useful to detect the shortcomings of the methods. We present analysis in two categories: (a) histogram distribution for each metric on the test set, and (b) qualitative comparison of the worst-case scenario for our method on PSNR metric." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "Figure 9 compares PSNR histogram plots on test-set views for the \"Cube-Sphere-Monkey\" scene. We can see that the mip360 approach performs poorly on PSNR and ranks low on practically all stratification levels. This supports our argument that the mip360 approach produces ar" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 310, + 319, + 541, + 396 + ], + "blocks": [ + { + "bbox": [ + 310, + 319, + 541, + 396 + ], + "lines": [ + { + "bbox": [ + 310, + 319, + 541, + 396 + ], + "spans": [ + { + "bbox": [ + 310, + 319, + 541, + 396 + ], + "type": "image", + "image_path": "24e3e768dfcc5b33ab923fe9b86f2bd66b6ce0fe9f27d539c776850633a18f69.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 310, + 402, + 541, + 498 + ], + "blocks": [ + { + "bbox": [ + 310, + 402, + 541, + 498 + ], + "lines": [ + { + "bbox": [ + 310, + 402, + 541, + 498 + ], + "spans": [ + { + "bbox": [ + 310, + 402, + 541, + 498 + ], + "type": "image", + "image_path": "da75ae1f9def0ad2b00312b8fac541c3c3c93e9cf7237eb18794aa914fd2955c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 502, + 545, + 568 + ], + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 568 + ], + "type": "text", + "content": "Figure 9. (Top Row) Comparison of histogram plots for the test-set for PSNR on \"Cube-Sphere-Monkey\". Note how distribution of our our method is always towards the right compared to other methods. " + }, + { + "bbox": [ + 304, + 502, + 545, + 568 + ], + "type": "inline_equation", + "content": "x - axis" + }, + { + "bbox": [ + 304, + 502, + 545, + 568 + ], + "type": "text", + "content": " denote metric value and " + }, + { + "bbox": [ + 304, + 502, + 545, + 568 + ], + "type": "inline_equation", + "content": "y - axis" + }, + { + "bbox": [ + 304, + 502, + 545, + 568 + ], + "type": "text", + "content": " denotes the frequency. A qualitative comparison of our method's worst-case PSNR results. PSNR is present at the bottom of the result image." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "content": "tifacts in such stratified scenes. For our method, the PSNR distributions are on the right. This implies that the novel views on test-set from our method will not be having serious artifacts in most cases, demonstrating its reliability." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Images in Figure 9 depict the qualitative results for the worst-case PSNR instances. All methods perform well in level 0. Hence, we are discussing interior levels which are level 1 and level 2. Other approaches fail in the worst-case scenario for our method at level 1. The outputs from NeRF, mip360 and Plenoxel are visually impaired. At level 2, our method has less blur compared to other approaches. These" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17610" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 213, + 174 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 213, + 174 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 213, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 213, + 174 + ], + "type": "image", + "image_path": "631f443465fa0af41048ac563ee11f3e596718971ad8f0c9ac4c1783c8588d57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 544, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 544, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 544, + 198 + ], + "type": "text", + "content": "Figure 10. Novel-views from different levels of 'Real Estate Video Tour 7 Rutledge Ave Highland Mills NY 10930 Orange County NY' scene in Real Estate 10K dataset. The two rows are from two-different view-points." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 214, + 70, + 378, + 174 + ], + "blocks": [ + { + "bbox": [ + 214, + 70, + 378, + 174 + ], + "lines": [ + { + "bbox": [ + 214, + 70, + 378, + 174 + ], + "spans": [ + { + "bbox": [ + 214, + 70, + 378, + 174 + ], + "type": "image", + "image_path": "85acfe2cb876e987d62c3b129cd1ead3ac06a733cf0722afe41f04726714f496.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 71, + 543, + 174 + ], + "blocks": [ + { + "bbox": [ + 379, + 71, + 543, + 174 + ], + "lines": [ + { + "bbox": [ + 379, + 71, + 543, + 174 + ], + "spans": [ + { + "bbox": [ + 379, + 71, + 543, + 174 + ], + "type": "image", + "image_path": "79208b58ad8e7ad9ed3112b3f81cd95dc5e6988c81d18c94161352fde2a8c63e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 235, + 287, + 288 + ], + "blocks": [ + { + "bbox": [ + 46, + 213, + 286, + 235 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 286, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 286, + 235 + ], + "type": "text", + "content": "Table 7. Quantitative comparison of our model and baseline on Synthetic Six Layer Scene." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 235, + 287, + 288 + ], + "lines": [ + { + "bbox": [ + 48, + 235, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 287, + 288 + ], + "type": "table", + "html": "
MetricsLevel 0Level 1Level 2Level 3Level 4Level 5
mip360 [2]PSNR ↑22.21516.18315.08412.01221.81321.539
SSIM ↑0.7770.4420.5100.3440.8170.647
OursPSNR ↑23.88921.44921.45624.09528.28321.898
SSIM ↑0.8330.6810.6850.7220.8830.686
", + "image_path": "509a0168e30e0cd2688cc1ab20db3c901526dd588650c45410230b37bc6408f2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 300, + 282, + 414 + ], + "blocks": [ + { + "bbox": [ + 50, + 300, + 282, + 414 + ], + "lines": [ + { + "bbox": [ + 50, + 300, + 282, + 414 + ], + "spans": [ + { + "bbox": [ + 50, + 300, + 282, + 414 + ], + "type": "image", + "image_path": "31aebb8b14c59cfadc4e06a15aa2d4207d8883b9ec8c4bf66dee0faf531acbcd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 421, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 453 + ], + "type": "text", + "content": "Figure 11. Comparisons of different codebook size on \"Dragon in Pyramid\" scene for different vector-codebook sizes. Note at size=1024 we achieve the best results with less artifacts." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 461, + 286, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 286, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 286, + 484 + ], + "type": "text", + "content": "findings demonstrate that our method is better suited to represent stratified scenes than others." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 664 + ], + "type": "text", + "content": "Ablation Studies. To analyse our proposed method, we present an ablation on the size of the vector codebook in our latent generator. Table 8 shows the ablation for the size of the vector codebook on the \"Coffee Shop\" dataset. We trialed with codebook sizes of 512, 1024 and 4096. We found that size 1024 provides optimal performance. As shown in Figure 11, increasing the codebook size induces haziness in the generated novel views, while decreasing the size creates white artifacts in level 0. As a result, we fix the size 1024 for all of our synthetic experiments. Whereas for RealEstate10K dataset we find that codebook size of 4096 produces the optimal tradeoff of results across levels, as it contains more number of levels and details. We further discuss the key architectural design choices for Latent Generator and Latent Router modules in Appendix E.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "No. of levels: To further test the efficacy of our method on higher number of levels, we created a \"Simple Geometry\" scene consisting of primitive geometry shapes like cube and spheres. More details are in the supplementary" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 234, + 545, + 285 + ], + "blocks": [ + { + "bbox": [ + 305, + 213, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 234 + ], + "type": "text", + "content": "Table 8. Quantitative results on \"Cube-Sphere-Monkey\" scene for ablation on size of the vector codebook in Latent Generator." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 234, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 285 + ], + "type": "table", + "html": "
SizePSNR ↑SSIM ↑LPIPS ↓
Level 0Level 1Level 0Level 1Level 0Level 1
51229.545826.34970.87430.73950.16750.4899
102429.483426.17150.87010.74890.13670.5163
409628.460927.82740.86280.73420.17760.5027
", + "image_path": "b9229ae9a4e6cc049fb7c699e94815f228daf981357edf53cd5d2e95e9172b48.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "type": "text", + "content": "material. Table 7 displays the results for both the baseline and our approach across a six levels stratified scene. The average PSNR/SSIM for the mip360 baseline is " + }, + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "type": "inline_equation", + "content": "15.35 / 0.487" + }, + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "type": "text", + "content": ", while our method achieved PSNR/SSIM of " + }, + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "type": "inline_equation", + "content": "23.54 / 0.754" + }, + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "type": "text", + "content": " which improves PSNR and SSIM by " + }, + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "type": "inline_equation", + "content": "53.35\\%" + }, + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "type": "inline_equation", + "content": "54.83\\%" + }, + { + "bbox": [ + 304, + 306, + 545, + 413 + ], + "type": "text", + "content": " respectively. This shows that our method performs better on increasing number of levels when compared with the baseline method. These observations also hold true for scenes in the RealEstate10K dataset as shown in Table 5." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "text", + "content": "In this work, we focus on the problem of modelling the 3D representation of a stratified and hierarchical scene, implicitly through a single neural field. For this, we propose Strata-NeRF, which models scenes with stratified structures by introducing a VQ-VAE-based latent generator to implicitly learn the distribution of latent space of input 3D locations and condition the neural radiance field with the latent code generated from this distribution. We also introduce a new synthetic dataset with stratified-level scenes and use it to analyse various existing approaches. Through quantitative, qualitative, and worst-case analysis on this dataset, we show that Strata-NeRF has a more stable 3D representation than the other methods. Further, the improvements due to Strata-NeRF also generalize to real-world RealEstate10K dataset, where it outperforms baselines by a significant margin establishing a new state-of-the-art. We believe designing a new volume rendering equation for modelling complex stratified scenes is a good direction for future work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by Samsung R&D Institute India, Bangalore, PMRF and Kotak IISc AI-ML Centre (KIAC). Srinath Sridhar was partly supported by NSF grant CNS-2038897" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17611" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 212 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 267 + ], + "type": "text", + "content": "[3] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "text", + "content": "[4] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 379 + ], + "type": "text", + "content": "[5] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 423 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision (ECCV). Springer, 2022. 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 425, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 479 + ], + "type": "text", + "content": "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 480, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 513 + ], + "type": "text", + "content": "[8] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 286, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 286, + 568 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 286, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 602 + ], + "type": "text", + "content": "[10] Abe Davis, Marc Levoy, and Fredo Durand. Unstructured light fields. In Computer Graphics Forum, volume 31. Wiley Online Library, 2012. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 286, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 286, + 658 + ], + "type": "text", + "content": "[11] Nianchen Deng, Zhenyi He, Jiannan Ye, Budmonde Duinkharjav, Praneeth Chakravarthula, Xubo Yang, and Qi Sun. Fov-nerf: Foveated neural radiance fields for virtual reality. IEEE Transactions on Visualization and Computer Graphics, 28(11), 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 713 + ], + "type": "text", + "content": "[12] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W Taylor, and Joshua M Susskind. Unconstrained scene generation with locally conditioned radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "text", + "content": "[14] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (ICCV). IEEE Computer Society, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "text", + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[16] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "text", + "content": "[17] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 371 + ], + "type": "text", + "content": "[18] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "type": "text", + "content": "[19] Karol Gregor, George Papamakarios, Frederic Besse, Lars Buesing, and Theophane Weber. Temporal difference variational auto-encoder. arXiv preprint arXiv:1806.03107, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 459 + ], + "type": "text", + "content": "[20] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[21] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "type": "text", + "content": "[22] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 580 + ], + "type": "text", + "content": "[23] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3), 1984. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 582, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 547, + 635 + ], + "type": "text", + "content": "[24] Takuhiro Kaneko. Ar-nerf: Unsupervised learning of depth and defocus effects from natural images with aperture rendering neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[26] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17612" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[27] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 162 + ], + "type": "text", + "content": "[28] Chaojian Li, Sixu Li, Yang Zhao, Wenbo Zhu, and Yingyan Lin. Rt-nerf: Real-time on-device neural radiance fields towards immersive ar/vr rendering. In Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "type": "text", + "content": "[29] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 259 + ], + "type": "text", + "content": "[30] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 316 + ], + "type": "text", + "content": "[31] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 381 + ], + "type": "text", + "content": "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 448 + ], + "type": "text", + "content": "[33] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (CVPR), June 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision (ECCV). Springer, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 504, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 288, + 547 + ], + "type": "text", + "content": "[35] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4), 2022. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "type": "text", + "content": "[36] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 635 + ], + "type": "text", + "content": "[37] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "text", + "content": "[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Brualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 152 + ], + "type": "text", + "content": "[40] Jialun Peng, Dong Liu, Songcen Xu, and Houqiang Li. Generating diverse structure for image inpainting with hierarchical vq-vae. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[42] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 297 + ], + "type": "text", + "content": "[43] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagon, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 298, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 353 + ], + "type": "text", + "content": "[44] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 354, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 410 + ], + "type": "text", + "content": "[45] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 411, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 487 + ], + "type": "text", + "content": "[46] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "type": "text", + "content": "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "text", + "content": "[48] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 578, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 612 + ], + "type": "text", + "content": "[49] Harry Shum and Sing Bing Kang. Review of image-based rendering techniques. In Visual Communications and Image Processing 2000, volume 4067. SPIE, 2000. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "type": "text", + "content": "[50] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "type": "text", + "content": "[51] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17613" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 93 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 93 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 93 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 160 + ], + "type": "text", + "content": "[52] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 227 + ], + "type": "text", + "content": "[53] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 293 + ], + "type": "text", + "content": "[54] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "type": "text", + "content": "[55] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "text", + "content": "[56] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "type": "text", + "content": "[57] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (CVPR). IEEE, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "type": "text", + "content": "[58] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 503, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 558 + ], + "type": "text", + "content": "[59] Xin Wang, Shinji Takaki, Junichi Yamagishi, Simon King, and Keiichi Tokuda. A vector quantized variational autoencoder (vq-vae) autoregressive neural " + }, + { + "bbox": [ + 48, + 503, + 287, + 558 + ], + "type": "inline_equation", + "content": "f_{-}0" + }, + { + "bbox": [ + 48, + 503, + 287, + 558 + ], + "type": "text", + "content": " model for statistical parametric speech synthesis. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "text", + "content": "[60] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4), 2004. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 658 + ], + "type": "text", + "content": "[61] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[62] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. Diver: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 318 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[63] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[64] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "text", + "content": "[65] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 218, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 218, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 218, + 545, + 262 + ], + "type": "text", + "content": "[66] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 262, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 318 + ], + "type": "text", + "content": "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17614" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/5c4682dc-11cf-43c9-b82f-059171805e73_content_list.json b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/5c4682dc-11cf-43c9-b82f-059171805e73_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef1f69731db30bda25bf89de0b06ef14d59caad7 --- /dev/null +++ b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/5c4682dc-11cf-43c9-b82f-059171805e73_content_list.json @@ -0,0 +1,1710 @@ +[ + { + "type": "text", + "text": "Strip-MLP: Efficient Token Interaction for Vision MLP", + "text_level": 1, + "bbox": [ + 202, + 130, + 767, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guiping Cao $^{1,2}$ Shengda Luo $^{1}$ Wenjian Huang $^{1}$ Xiangyuan Lan $^{2,*}$ Dongmei Jiang $^{2}$ Yaowei Wang $^{2}$ Jianguo Zhang $^{1,2,*}$", + "bbox": [ + 192, + 179, + 769, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Southern University of Science and Technology, Shenzhen, China $^{2}$ Peng Cheng Laboratory, Shenzhen, China", + "bbox": [ + 222, + 217, + 750, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "12131099@mail.sustech.edu.cn, {luosd, huangwj, zhangjg}@sustech.edu.cn, {lanxy, jiangdm, wangyw}@pcl.ac.cn", + "bbox": [ + 169, + 253, + 795, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Token interaction operation is one of the core modules in MLP-based models to exchange and aggregate information between different spatial locations. However, the power of token interaction on the spatial dimension is highly dependent on the spatial resolution of the feature maps, which limits the model's expressive ability, especially in deep layers where the feature are down-sampled to a small spatial size. To address this issue, we present a novel method called Strip-MLP to enrich the token interaction power in three ways. Firstly, we introduce a new MLP paradigm called Strip MLP layer that allows the token to interact with other tokens in a cross-strip manner, enabling the tokens in a row (or column) to contribute to the information aggregations in adjacent but different strips of rows (or columns). Secondly, a Cascade Group Strip Mixing Module (CGSMM) is proposed to overcome the performance degradation caused by small spatial feature size. The module allows tokens to interact more effectively in the manners of within-patch and cross-patch, which is independent to the feature spatial size. Finally, based on the Strip MLP layer, we propose a novel Local Strip Mixing Module (LSMM) to boost the token interaction power in the local region. Extensive experiments demonstrate that Strip-MLP significantly improves the performance of MLP-based models on small datasets and obtains comparable or even better results on ImageNet. In particular, Strip-MLP models achieve higher average Top-1 accuracy than existing MLP-based models by $+2.44\\%$ on Caltech-101 and $+2.16\\%$ on CIFAR-100. The source codes will be available at https://github.com/MedProcess/Strip_MLP.", + "bbox": [ + 76, + 353, + 473, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 818, + 207, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In computer vision, Convolutional Neural Networks (CNNs) are one of the most popular network backbones, which have made a series of breakthroughs [8]. Inspired by", + "bbox": [ + 76, + 832, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4491c943a6c738d7f2000e08d7a09d1c31030a5bd3e4ddb4b67c136e9d374b3a.jpg", + "image_caption": [ + "Figure 1. Token interaction of the MLP layer on the down-sampled image feature. (a) The process of which the feature resolution will be down-sampled from $H \\times W \\times C$ into a small spatial size of $\\frac{H}{8} \\times \\frac{W}{8} \\times 8C$ . (b) Token interaction on different feature resolutions." + ], + "image_footnote": [], + "bbox": [ + 560, + 320, + 833, + 410 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the great success of self-attention-based architectures [32] in natural language processing (NLP), the Transformer models [5, 19, 34] are introduced into the field of computer vision, and have achieved comparable results with state-of-the-art (SOTA) CNNs. Although ViT [5] and its variants outperform traditional CNNs, the models introduce high computational complexity to construct attention maps. Recently, some studies [20, 26] in the vision community suggest that the attention mechanism is not necessary and simpler model architectures are proposed.", + "bbox": [ + 496, + 489, + 893, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MLP-based models, like MLP-Mixer [28], gMLP [16] and ViP [10] process the data with the Multilayer Perceptrons (MLP), showing great potential to improve the performance of vision models [29]. As the first visual deep MLP network, MLP-Mixer [28] introduces two types of MLP layers: Channel-Mixing MLPs (CMM) and Token-Mixing MLPs (TMM). For CMM, the module mainly mixes the information between different channels of each token. For TMM, it allows each spatial token to interact with all other tokens (the whole image) in a single MLP layer. However, this design also introduces a much larger number of parameters and higher computational complexity prone to overfitting. To address this problem, Sparse MLP (SMLP) [26] and Vision Permutator (ViP) [10] propose a similar layer of parallel structure, which applies one dimension MLP along the axial directions, and parameters are shared among rows or columns, respectively. Therefore, it reduces the number", + "bbox": [ + 496, + 643, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1494", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of model parameters and computational complexity, avoiding the common over-fitting problem.", + "bbox": [ + 78, + 90, + 465, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although SMLP and ViP alleviate some deficiencies of MLP-Mixer [28], both methods bring the challenge that the token interaction power is highly dependent on the feature spatial size when interacting tokens on spatial rows (or columns). As shown in Fig. 1, the spatial feature resolution is down-sampled to a small size but with more channels, which means the feature pattern of each token is mainly concentrated on the channel dimension rather than the spatial one. Interacting tokens along the spatial dimension by sharing the weights among all channels would seriously ignore the feature pattern differences among different channels, which may degrade the token interaction power, especially in deep layers with small spatial feature resolution. Here, we mark this problem as the Token's interaction dilemma. Taking SMLP as an example, we analyze the feature resolution and complexity of the model in different stages in detail (seen in Sec. 3.4). We find that as the spatial feature size decreases by down-sampling stage by stage, the token interaction layer also becomes smaller and smaller, which makes the token interaction power degraded rapidly.", + "bbox": [ + 80, + 123, + 467, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a new efficient Strip MLP model, dubbed Strip-MLP, to enrich the power of the token interaction layer in three ways. For the level of a single MLP layer, inspired by the cross-block normalization schemes of HOG [3] and the sparse connections between the biological neurons, we design a Strip MLP layer to allow the token to interact with other tokens in a crossstrip manner, enabling each row or column of the tokens to contribute differently to other rows or columns. For the token interaction module level, we develop channel-wise group mixing of CGSMM, enabling the tokens in a row (or column) to contribute to the information aggregations in adjacent but different strips of rows (or columns). to tackle the problem that the token interaction power decreases in deep layers with the spatial feature size significantly reduced but with multiplying channels. Considering the existing methods [10, 26, 28] interact the tokens mainly in the long range of row (or column), which may not aggregate tokens well in the local region, we propose a new Local Strip Mixing Module (LSMM) with a small Strip MLP unit to strengthen the token interaction power on local interactions.", + "bbox": [ + 80, + 426, + 467, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed Strip-MLP model significantly boosts the token interaction power, and the main contributions are:", + "bbox": [ + 80, + 743, + 465, + 772 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A new MLP paradigm for vision MLP: Strip MLP layer, which aggregates the adjacent tokens in a cross strip manner and enables each row or column of the tokens to contribute differently to other rows or columns, interacting tokens more efficiently.", + "- Designing a Cascade Group Strip Mixing module and a Local Strip Mixing module, which effectively improves the model's token interaction power and boosts" + ], + "bbox": [ + 93, + 780, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the tokens aggregation in the local region, respectively;", + "bbox": [ + 532, + 90, + 888, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments show that Strip-MLP remarkably improves the performances of the MLP-based models. Strip-MLP achieves higher average Top-1 accuracy by $+2.44\\%$ on Caltech-101 and $+2.16\\%$ on CIFAR-100 over the existing MLP-based models. In addition, our models achieve comparable or even better performances on ImageNet-1K compared with traditional MLP-based models, and other popular CNNs and transformer-based models.", + "bbox": [ + 517, + 107, + 888, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 260, + 638, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep neural networks for vision recognition can be mainly divided into three categories: Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Multi-Layer Perceptron-based models (MLPs) [10].", + "bbox": [ + 503, + 280, + 888, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CNNs-Based Models. CNNs are the de-facto standard deep learning network models for vision tasks and have been intensely studied in the vision community. AlexNet [15] is a symbolically significant model that won the ILSVRC 2012 contest with far higher performance than others. Since then, CNNs-based models have attracted more attentions, and lots of deeper and more effective architectures [8, 11, 23-25, 33, 39, 40] have been proposed. With the convolution and pooling layer, CNNs aggregate the feature in a local region but not well in the long-term dependencies, which are optimized by the new vision model like Transformer [32] and MLP [28] models.", + "bbox": [ + 503, + 340, + 888, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer-Based Models. Transformer [32] is introduced for machine translation and becomes the reference model for all-natural language processing (NLP) tasks. Inspired by the great success of Transformer in NLP, ViTs [5] first applies a standard Transformer to images, which attains excellent results compared to the SOTA CNNs model. DeiT [30] introduces several training strategies and distillation methods to make ViTs more effective on the smaller ImageNet-1K dataset. By proposing a hierarchical Transformer with shifted windows, Swin Transformer [19] achieves SOTA accuracy on ImageNet-1K, bringing greater efficiency by self-attention in local window and cross-window connection. For these models, self-attention is the core module but with heavy computational burdens to obtain an attention map.", + "bbox": [ + 503, + 523, + 888, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MLP-Based Models. Without the convolutions and self-attention mechanism, MLP-Mixer [28] builds the architecture that only uses the MLP layer and achieves competitive performance on image classification benchmarks. Since then, the researchers have developed many MLP-like variants [1, 7, 16, 26, 27, 29, 31, 35, 38] models. The work [17] gives a comprehensive survey about visual deep MLP models and compares the intrinsic connections and differences between convolution, self-attention mechanism, and Token-mixing MLP in detail. Sparse MLP [26] introduced a sparse", + "bbox": [ + 503, + 750, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1495", + "bbox": [ + 485, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c5a7848b3d15c4f0fd4fee5d97f45e64feb6afbe1e482af05689489606d22fe6.jpg", + "image_caption": [ + "Figure 2. The overall and components architecture of Strip-MLP. (a) Strip-MLP has four stages, and $T1$ to $T4$ means the repeated times of the block in each stage. (b) The Strip Mixing Block splits the head (channel dimension) into two heads, and they are fed into the two parallel branches. (c) Channel Mixing Block architecture." + ], + "image_footnote": [], + "bbox": [ + 151, + 89, + 821, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "operation to separately aggregate information along axial directions, avoiding the quadratic computational complexity of conventional MLP. Hire-MLP [7] presents a novel variant of MLP-based architecture via hierarchically rearranging tokens to aggregate local and global spatial information. Wave-MLP [27] represents each token as a wave function with two parts of amplitude and phase, which has the ability to model varying contents from different input images. Above all, there is still the limitation of token interaction power degraded significantly especially when the spatial feature resolution becomes small, which has been overlooked by previous studies. In this paper, we aim to enrich the token interaction power on both the single MLP layer and token interaction module to advance the performance of MLP-based models.", + "bbox": [ + 75, + 392, + 472, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 632, + 169, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first present the overall architecture of Strip-MLP. Then, we show the key proposed components: Strip MLP Layer, Cascade Group Strip Mixing Module (CGSMM) and Local Strip Mixing Module (LSMM) of the model in detail, and present the comparison analysis between the Strip MLP and traditional MLP model in terms of parameters and complexity. Finally, we define four kinds of architecture variants to compare the performance of the models in different sizes.", + "bbox": [ + 75, + 652, + 468, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 76, + 792, + 272, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of the Strip-MLP model is depicted in Fig. 2 (a). We design the Strip-MLP as a hierarchical structure of a multi-stages model. Given the input image $I$ , Strip-MLP model makes patch embedding and models the image feature in four stages.", + "bbox": [ + 75, + 809, + 468, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Patch Embedding. Strip-MLP firstly splits the input", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ (H, W: image height and width) into a sequence of image patches (also referred as tokens) $I_{p} \\in \\mathbb{R}^{hw \\times c}$ (c: the number of channels), where the patch size is $p \\times p$ and patch number is $hw\\left(h = \\frac{H}{p}, w = \\frac{W}{p}\\right)$ , and then all patches are linearly projected into a desired higher dimension $C$ of the feature $(X \\in \\mathbb{R}^{hw \\times C})$ .", + "bbox": [ + 496, + 391, + 890, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mixing Block. The purpose of mixing block is to boost the interactions between the features of diverse spatial locations and channels. To effectively aggregate spatial and channel information, we design two sub-blocks of the Strip Mixing Block and Channel Mixing Block. The Strip Mixing Block is comprised of CGSMM and LSMM, which blend and aggregate spatial information more efficiently at the global and local level, respectively. Patch merging module aims to merge the feature in which the spatial dimension will be reduced by $2 \\times 2$ , and the channel dimension increases by 2 times so that the model down-samples the feature from $\\frac{H}{p} \\times \\frac{W}{p}$ into $\\frac{H}{8p} \\times \\frac{W}{8p}$ stage by stage. To obtain multi-scale features, we apply a single convolutional layer to the output features of stages 1 and 2. Then, we add the resulting features to the input features of stages 3 and 4.", + "bbox": [ + 496, + 483, + 892, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Head Layer. The features extracted by the multiple blocks are fed into a global average pooling (GAP) layer to reduce the feature dimension. Finally, the feature will be fed into a fully-connected head layer for classification.", + "bbox": [ + 496, + 710, + 890, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Strip Mixing Block and Channel Mixing Block", + "text_level": 1, + "bbox": [ + 498, + 776, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Strip Mixing Block. To improve the token interaction power, we design the block to aggregate both long and short range interactions in a parallel manner. As illustrated in Fig. 2 (b), we split the feature in channel dimension, with one-half of the channel feature being fed into CGSMM to model the long-range interactions, and the remaining feature being fed into LSMM to aggregate the local interac", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1496", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4ab71b6bb1c028c6a1606c9715c1f4989d2f035242aa57032a5b61195168bdef.jpg", + "image_caption": [ + "(a) Cascade Group Strip Mixing Module" + ], + "image_footnote": [], + "bbox": [ + 150, + 89, + 645, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6400613916a9863fc20833d1ff07defb126b953838f96da8bed118121060456d.jpg", + "image_caption": [ + "(b) Local Strip Mixing Module", + "Figure 3. (a) The architecture of CGSMM. It contains the Strip MLP layer, the feature splitting, restoring, and permuting operation in a cascade mode. (b) The structure of the Local Strip Mixing module. The dark green block composes of a re-weight module." + ], + "image_footnote": [], + "bbox": [ + 650, + 89, + 816, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions. Given the input feature $X\\in \\mathbb{R}^{H\\times W\\times C}$ , the block can be formulated as:", + "bbox": [ + 75, + 318, + 468, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {m} = M L P \\left(\\left(D W S C (X)\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 354, + 468, + 369 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY = F C \\left(C a t \\left(C G S M M \\left(X _ {m _ {1}} ^ {\\frac {C}{2}}\\right), L S M M \\left(X _ {m _ {2}} ^ {\\frac {C}{2}}\\right)\\right)\\right) + X \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 373, + 468, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X_{m}$ and $Y$ are the intermediate and output feature of the block. $X_{m}$ is split into $X_{m_1}$ and $X_{m_2}$ with half channel, respectively. DWSC means depth-wise convolution [2], and the kernel size is $3\\times 3$ . MLP is a serial connection of fullyconnection (FC), batch normalization [12] and GELU [9] activation layers. $Cat(*)$ denotes concatenation operation.", + "bbox": [ + 76, + 395, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Channel Mixing Block. The block aims to aggregate the information between channels of the token, and the basic structure is shown in Fig. 2 (c). Following the inverted bottleneck in [20] and global response normalization (GRN) in [36], we design the Channel Mixing block to increase the contrast and selectivity of channels.", + "bbox": [ + 75, + 484, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. CGSMM and LSMM", + "text_level": 1, + "bbox": [ + 76, + 580, + 277, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first introduce the Strip MLP Layer. Based on this layer, we design CGSMM using a simple but effective strategy that splits the feature into patches along channel dimension and interacts tokens more effectively in the manners of within-patch and cross-patch, which is independent to the feature spatial size. In addition, the design of existing methods [10, 26, 28] only allows tokens to interact in a long range of rows (or columns) with sharing weight; thus the resulting model may struggle with effectively aggregating both global and local information simultaneously. Therefore, we design LSMM to capture the local interactions more efficiently.", + "bbox": [ + 75, + 598, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Strip MLP Layer. In MLP-based models, the majority of MLP layers treat each row and column of the data in isolation, as formulated in Eq. (3), which may lead to the inefficiency of token interaction. Inspired by the cross-block normalization schemes of HOG [3], which overlaps the blocks and enables each scalar cell response to contribute into different blocks, we propose the Strip MLP layer. The proposed layer applies MLP on the \"strip\" data of adjacent", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rows or columns in the spatial direction in order to aggregate the feature in a cross-strip manner. Given the input $X$ , we formulated the Strip MLP layer (setting strip width as 3 for example) in Eqs. (4) and (5):", + "bbox": [ + 498, + 320, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {*, j} ^ {h} = W _ {1} X _ {*, j}; \\quad X _ {i, *} ^ {w} = W _ {2} X _ {i, *} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 378, + 890, + 397 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {*, j} ^ {h} = W _ {3} \\operatorname {C a t} \\left(X _ {*, j - 1}, X _ {*, j}, X _ {*, j + 1}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 398, + 890, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {i, *} ^ {w} = W _ {4} \\operatorname {C a t} \\left(X _ {i - 1, *}, X _ {i, *}, X _ {i + 1, *}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 422, + 890, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{1}\\sim W_{4}$ are the weights of the MLP layer, $i$ and $j$ are the token index in row and column. The superiorities of the Strip MLP layer are mainly in two aspects: on one hand, Strip MLP layer enables the token to interact with other tokens in both short and long spatial ranges simultaneously. On the other hand, similar to the HOG cross-block normalization process, each row (or column) not only serves for the current row (or column) tokens aggregation but also makes contribution to the adjacent rows (or columns) feature aggregation. For example, in Eq. (4), $X_{*,j}$ makes different contributions for aggregating the processed feature of $X_{*,j - 1}^{h},X_{*,j}^{h}$ and $X_{*,j + 1}^{h}$ , making tokens interacting more efficiently in a cross-strip manner.", + "bbox": [ + 496, + 441, + 890, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cascade Group Strip Mixing Module. The architecture is shown in Fig. 3 (a). The module applies the Strip MLP layer in row and column direction in a cascade mode. As the operation of the Strip MLP layer on row is similar to that on column, we take one of them to show the method.", + "bbox": [ + 496, + 638, + 890, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Patch Splitting of the Data along Channel Dimension. Given the input feature $X_{m_1} \\in \\mathbb{R}^{H \\times W \\times \\frac{C}{2}}$ , the module firstly permutes the feature into $X_{m_1}^{H \\times \\frac{C}{2} \\times W}$ , and splits the feature into $P$ patches in the channel dimension and concatenated along the row $(X^{H \\times \\frac{C}{2P} \\times PW})$ dimension.", + "bbox": [ + 496, + 714, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Group Strip MLP Layer. To improve the token interaction power, we propose a Group Strip MLP Layer (GSML) to interact tokens in the manners of within-patch and cross-patch. In particular, we apply the unshared weights of Strip MLP on different patches, and the tokens within the same patch share the weights for interaction. Then, we restore the feature to original shape and concatenate it with the input", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1497", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feature. To interact the tokens crossing patches, we apply a channel fully-connected (Channel FC) layer to interact tokens between different patches.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local Strip Mixing Module. The module is illustrated in Fig. 3 (b). In order to better aggregate the local interactions on spatial dimension, we define a small Strip MLP unit, where the strip width and length are 3 and 7, respectively. Given the input feature $X_{m_2}^{H \\times W \\times \\frac{C}{2}}$ , we aggregate the local interactions in rows and columns direction simultaneously. Following [1, 10, 27], we sum all branches with a re-weight module.", + "bbox": [ + 75, + 137, + 468, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Parameter and Complexity Analysis", + "text_level": 1, + "bbox": [ + 76, + 275, + 390, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In multi-stage processing architecture, the Token's interaction dilemma becomes more serious, as the feature spatial resolution will be down-sampled in deep layers, and the token interaction power will be weakened. In this section, we present the comparison analysis of parameters and complexity to show the effectiveness of the proposed CGSMM. Considering the majority MLP-Based models have the similar model structure of token interaction, we analyze the parameters and complexities with Sparse MLP [26], which is a popular model with good performance on various datasets.", + "bbox": [ + 75, + 294, + 468, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sparse MLP [26] firstly applies MLP on the columns and rows of $X$ to map $\\mathbb{R}^{CW\\times H}$ to $\\mathbb{R}^{CW\\times H}$ separately. Then, the model concatenates the two processed features with the input feature $X$ and uses a linear layer on the channel dimension to fuse the feature from $\\mathbb{R}^{HW\\times 3C}$ to $\\mathbb{R}^{HW\\times C}$ . The number of parameters and FLOPs of the first step are $W^2 +H^2$ and $CHW(H + W)$ ; they are $3C^2$ and $3HWC^2$ for the fusion step. Comparatively, the number of parameters and FLOPs of Strip MLP layer for token interaction are $3P(H^{2} + W^{2})$ and $3CHW(H + W)$ , respectively. The fusion step has $4C^2$ parameters and $4HWC^2$ FLOPs. In particular, for $H_{1} = W_{1} = 56$ , $H_{4} = W_{4} = 7$ , $C_1 = 112$ , $C_4 = 112\\times 2^3 = 896$ , and $P = \\frac{C}{4}$ , we calculate the number of parameters and FLOPs of both two layers. As shown in Tab. 1, for the Sparse MLP block, the number of parameters for token interaction in stage 4 (only $0.10k$ ) decreases by 62.70 times than in stage 1. In addition, most of the number of the block's parameters and the FLOPs concentrate in the fusion step. For example, only $0.01\\%$ parameters ( $0.52\\%$ FLOPs) are used for the token interaction step in stage 4. Based on the above analysis, to improve the token interaction power, it would be better to redesign the block to balance the number of parameters and FLOPs.", + "bbox": [ + 75, + 446, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In CGSMM, we use a simple but effective strategy that splits the feature along channel dimension into patches and interacts the tokens in the manners of within-patch and cross-patch. No matter how the spatial resolution decreases, the module can still interact the tokens with channel-wise specificity in different patches. In Tab. 1, we improve the Proportion of the number of parameters in the token inter", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/32ce2c452b21075bcd8e235260096dd25e16821ccb5a12a76e0e35c5884d439a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ItemsSparse MLPStrip MLP
ParamsFLOPsParamsFLOPs
Stage 16.27k39.34M526.85k118.01M
Stage 40.10k0.62M65.86k1.84M
Changes↓62.70↓63.45↓8.00↓64.14
Stage 4 Fusion2.41M118.01M3.21M157.35M
Proportion0.01%0.52%2.01%1.16%
", + "bbox": [ + 506, + 89, + 885, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. The model parameters and complexity comparison between Sparse MLP and Strip MLP layer in different stages. Stage 1 and Stage 4 show the parameters (Params) and floating-point operations (FLOPs) for token interaction in one layer between different stages. Stage 4 Fusion indicates the Params and FLOPs in the feature fusion step of the token interaction module. Proportion means the ratio between the token interaction step and other steps in the token interaction module of stage 4.", + "bbox": [ + 498, + 188, + 890, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "action layer from $0.01\\%$ to $2.01\\%$ in stage 4. Although our GSML brings more parameters and computations when just considering one token interaction layer, the total number of parameters and overall computational complexity of the model are decreased as the token interaction power is improved and we set small model configurations of $T_{1} \\sim T_{4}$ . Additionally, the experiments in Sec. 4 show that our models achieve better performance with fewer parameters and FLOPs than other SOTA models.", + "bbox": [ + 496, + 313, + 890, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Architecture Variants", + "text_level": 1, + "bbox": [ + 500, + 454, + 702, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We developed four variants of our Strip-MLP network: Strip-MLP-T\\* (light tiny), Strip-MLP-T (tiny), Strip-MLPSmall), Strip-MLP-B (base), which have similar or even smaller model sizes, compared with MLP-baed models [7, 26, 27] and Swin Transformer [18], respectively. The hyperparameters of the four variants of the models are as follows:", + "bbox": [ + 496, + 473, + 890, + 564 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Strip-MLP-T*: $C = 80, \\{T_1 \\sim T_4\\} = \\{2, 2, 6, 2\\}$ ;", + "- Strip-MLP-T: $C = 80, \\{T_1 \\sim T_4\\} = \\{2, 2, 12, 2\\}$ ;", + "- Strip-MLP-S: $C = 96, \\{T_1 \\sim T_4\\} = \\{2, 2, 18, 2\\}$ ;", + "- Strip-MLP-B: $C = 112, \\{T_1 \\sim T_4\\} = \\{2, 2, 18, 2\\}$ ;" + ], + "bbox": [ + 514, + 574, + 872, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ means the channel number of the hidden layers in the first stage, and $T_{1} \\sim T_{4}$ is the number of times repeated for Strip Mixing Block and Channel Mixing Block in each of the four stages, as shown in Fig. 2.", + "bbox": [ + 496, + 642, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 712, + 632, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our Strip-MLP with three main categories of networks: CNNs-based, Transformer-based, and MLP-based models. We first show the experimental settings of the datasets and implementation details. Then, the proposed method is compared with other popular methods on three datasets. Finally, we ablate the critical design components of Strip-MLP network.", + "bbox": [ + 496, + 732, + 890, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 849, + 705, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Methods are evaluated in four datasets which vary significantly in the number of training images:", + "bbox": [ + 500, + 873, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1498", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5649ca10ddc69344cbbac0bd7c69d8add112355e48239a6e6facf78dd17d0458.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Image classification results of our Strip-MLP and other models on Caltech-101. Top-1 denotes the Top-1 accuracy. Here, the patch number of CGSMM is C/2." + ], + "table_body": "
MethodParamsFLOPsTop-1(%)
CNNs-Based
ResNet50 [8]23.71M4.12G89.61
ResNet101 [8]42.71M7.85G88.56
ResNet152 [8]58.35M11.58G89.04
Transformer-Based
ViT-B/16 [5]85.85M16.86G53.96
Swin-T [19]27.60M4.36G80.40
Swin-S [19]48.91M8.52G79.83
Swin-B [19]86.85M15.14G78.42
MLP-Based
Wave-MLP-T [27]16.73M2.48G86.89
Hire-MLP-T [7]17.58M2.16G87.85
Strip-MLP-T*(ours)17.67M2.53G90.11
Wave-MLP-S [27]30.25M4.55G87.43
Hire-MLP-S [7]32.65M4.26G88.48
Sparse-MLP-T [26]23.55M5.02G90.54
Strip-MLP-T(ours)24.33M3.67G90.93
Wave-MLP-M [27]43.59M7.93G88.45
Hire-MLP-B [7]57.77M8.17G88.62
Sparse-MLP-S [26]47.87M10.36G91.07
Strip-MLP-S(ours)43.66M6.83G92.09
Wave-MLP-B [27]62.90M10.27G88.81
Hire-MLP-L [7]95.03M13.50G88.19
Sparse-MLP-B [26]65.07M14.04G91.67
Strip-MLP-B(ours)58.49M9.22G92.26
", + "bbox": [ + 83, + 88, + 460, + 433 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Caltech-101 [6]: it is a benchmark dataset that includes 101 classes, totally around $9k$ images. We randomly take $80\\%$ of each class of the data as the training set and use the remaining data as the testing set.", + "- CIFAR-10 [14]: it consists of $60k \\times 32$ images in 10 classes, and there are $50k$ images", + "- CIFAR-100 [14]: it has 100 classes containing 600 images of each class (totally $60k$ ), and there are 500 training images and 100 testing images per class.", + "- ImageNet-1K [4]: it is an image classification benchmark containing $1.28M$ training images and $50k$ testing images of 1000 classes." + ], + "bbox": [ + 93, + 492, + 467, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. In all of our experiments, the input image size is $224 \\times 224$ , and the input patch size is $4 \\times 4$ . Based on the deep-learning framework of PyTorch [21], we train our models from scratch without any extra data for pre-training except for the transfer learning experiments. We employ an AdamW [13] optimizer with 300 epochs for all training datasets and apply a cosine decay learning rate scheduler and 30 epochs of linear warm-up. In addition, we adopt the same augmentation and regularization strategies as to Swin-Transformer [19].", + "bbox": [ + 75, + 678, + 467, + 827 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Image Classification on Small Datasets", + "text_level": 1, + "bbox": [ + 76, + 837, + 408, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on Small Datasets of Caltech-101, CIFAR-10 and CIFAR-100. The results consistently indicate that our Strip-MLP efficiently improves the token", + "bbox": [ + 76, + 854, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "interaction power, and advances the MLP-based models' performance on the small dataset than other models.", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on Caltech-101. Tab. 2 presents the image classification result of three types models on Caltech-101 dataset. Compared to ResNet [8], all the four variants of our Strip-MLP models achieve higher Top-1 accuracy with fewer parameters and FLOPs, demonstrating an average increase of $+2.69\\%$ over three widely-used ResNet models.", + "bbox": [ + 496, + 123, + 890, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the Transformer-based models, e.g. Swin Transformer [19], Strip-MLP models noticeably surpass all variants models: $+10.53\\% / 12.26\\% / 13.84\\%$ over SwinTransformer (T/S/B) models. ViT-B/16 [5] only gets an accuracy of $53.96\\%$ , indicating that transformer-based models depend more on the scale of the datasets.", + "bbox": [ + 496, + 214, + 890, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When compared to the MLP-based models, like Sparse MLP [26], Wave-MLP [27] and Hire-MLP [7], our models still achieve the best accuracy with fewer model parameters and FLOPs. For example, with fewer FLOPs $(-4.28\\mathrm{G})$ Strip-MLP-B $(92.26\\%)$ has an increase of $+3.64\\%$ Top-1 accuracy than Hire-MLP-B $(88.62\\%)$ .", + "bbox": [ + 496, + 306, + 890, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on CIFAR-10. Tab. 3 shows the classification results of performance comparisons. Compared with CNNs-based models, all four variants of our Strip-MLP models achieve higher Top-1 accuracy with fewer parameters and FLOPs, with an average increase of $+1.14\\%$ (98.06% vs. 96.92%). When compared with transformer-based models, our Strip-MLP achieves better performance by $+3.06\\%$ (98.06% vs. 95.00%) with fewer parameters (34.49M vs. 54.37M) and FLOPs (5.56G vs. 9.34G). Compared with MLP-based models, all of our Strip-MLP models obtain the best Top-1 accuracy with fewer parameters and FLOPs. For example, the average Top-1 accuracy of Strip-MLP is higher than Hire-MLP/Wave-MLP/Sparse-MLP by $+1.45\\% / 0.69\\% / 0.19\\%$ (98.06% vs. 96.61%/97.37%/97.87%) with fewer average parameters (34.49M vs. 50.71M/38.32M/45.43M) and fewer average FLOPs (5.56G vs. 9.36G/6.31G/9.81G)", + "bbox": [ + 496, + 398, + 890, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In particular, with only 16.76M parameters and 2.53G FLOPs, our Strip-MLP- $T^{*}$ noticeably surpasses all variants of ResNet [8] and Swin-Transformer [19] models, and Strip-MLP- $T^{*}$ outperforms the transfer learning models of ViT-S/16 [5]/TST-14 [37] by $+0.69\\% / +0.29\\%$ , consistently demonstrating our Strip-MLP has significant superiorities in performance and complexity.", + "bbox": [ + 496, + 657, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on CIFAR-100. Tab. 4 compares the image classification results on CIFAR-100 with three kinds of architecture models. Compared with CNNs-based models [8], Strip-MLP obtains average higher accuracy of $+2.05\\% \\sim +2.43\\%$ with a smaller number of parameters and FLOPs. When compared to the self-attention based model, e.g. Swin Transformer [19], our model shows great superiority that Strip-MLP-B achieves $86.40\\%$ Top-1 accuracy, $+9.64\\%$ than the Swin-B model with fewer pa", + "bbox": [ + 496, + 763, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1499", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fcf25a67eea67df98c6dc63189609979264ae75d55d06b9078b6a0de08642513.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Image classification results of different models on CIFAR-10. The patch number of CGSMM is C/4. All models are trained from scratch without any extra data." + ], + "table_body": "
MethodParamsFLOPsTop-1 (%)
CNNs-Based
ResNet50 [8]23.53M4.12G97.06
ResNet101 [8]42.52M7.85G96.70
ResNet152 [8]58.16M11.58G96.99
Transformer-Based
Swin-T [19]27.53M4.36G95.25
Swin-S [19]48.84M8.52G94.87
Swin-B [19]86.75M15.14G94.87
MLP-Based
Wave-MLP-T [27]16.69M2.48G96.91
Hire-MLP-T [7]17.53M2.16G96.20
Strip-MLP-T*(ours)16.76M2.53G97.79
Wave-MLP-S [27]30.20M4.55G97.27
Hire-MLP-S [7]32.61M4.26G96.70
Sparse-MLP-T [26]23.49M5.02G97.67
Strip-MLP-T(ours)23.07M3.67G97.94
Wave-MLP-M [27]43.55M7.93G97.56
Hire-MLP-B [7]57.72M8.17G96.88
Sparse-MLP-S [26]47.80M10.36G97.99
Strip-MLP-S(ours)41.81M6.83G98.23
Wave-MLP-B [27]62.83M10.27G97.72
Hire-MLP-L [7]94.96M13.50G96.64
Sparse-MLP-B [26]64.99M14.04G97.95
Strip-MLP-B(ours)56.33M9.22G98.28
", + "bbox": [ + 80, + 88, + 467, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rameters (58.49M vs. 86.85M) and lower FLOPs (9.22G vs. 15.14G). In addition, our models keep the superiority over the other MLP-based models in that Strip-MLP achieves the best Top-1 accuracy (with a higher range of $+1.04\\% \\sim +3.01\\%$ ) compared to the MLP-based models, e.g. Wave-MLP [27], Hire-MLP [7], and Sparse MLP [26], with fewer parameters and FLOPs.", + "bbox": [ + 76, + 479, + 468, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Transfer Learning on Small Datasets", + "text_level": 1, + "bbox": [ + 76, + 595, + 397, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the transfer learning experiment of DynaMixer [35], we show the advantages of our models on the transfer learning tasks. With the pre-trained Strip-MLP $\\mathrm{T}^*$ model on the ImageNet-1K, we transfer the model to downstream datasets of CIFAR-10 and CIFAR-100. Compared to the previous models, such as ViT [5], ViP [10] and DynaMixer [35] models, our Strip-MLP $\\mathrm{T}^*$ models achieve the best accuracies of $98.8\\%$ and $89.4\\%$ on two datasets, as shown in Tab. 5, $+1.7\\% / 0.6\\%$ than ViT-S/16 and DynaMixer-S on CIFAR-10, $+2.3\\% / 0.8\\%$ over ViT-S/16 and DynaMixer-S on CIFAR-100. More importantly, the number of parameters of our model is only 18M, which is far less than other models, indicating that our model is more efficient and effective.", + "bbox": [ + 75, + 614, + 470, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Image Classification on ImageNet-1K", + "text_level": 1, + "bbox": [ + 76, + 837, + 401, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments on ImageNet-1K to show the effectiveness of the method. Tab. 6 shows the performance together with the complexity (in terms of FLOPS and the", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b0c30cdbb39d414a3839593c637eb11a448b2104ae25a22f3a0728e9010dd24.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParamsFLOPsTop-1(%)
CNNs-Based
ResNet50 [8]23.71M4.12G83.03
ResNet101 [8]42.71M7.85G83.75
ResNet152 [8]58.35M11.58G84.35
Transformer-Based
Swin-T [19]27.60M4.36G78.56
Swin-S [19]48.91M8.52G78.10
Swin-B [19]86.85M15.14G76.76
MLP-Based
Wave-MLP-T [27]16.73M2.48G82.77
Hire-MLP-T [7]17.58M2.16G82.44
Strip-MLP-T*(ours)17.67M2.53G85.10
Wave-MLP-S [27]30.25M4.55G83.34
Hire-MLP-S [7]32.65M4.26G82.94
Sparse-MLP-T [26]23.55M5.02G84.19
Strip-MLP-T(ours)24.33M3.67G85.23
Wave-MLP-M [27]43.59M7.93G83.95
Hire-MLP-B [7]57.77M8.17G83.17
Sparse-MLP-S [26]47.87M10.36G84.27
Strip-MLP-S(ours)43.66M6.83G86.18
Wave-MLP-B [27]62.89M10.27G84.23
Hire-MLP-L [7]95.03M13.50G83.53
Sparse-MLP-B [26]65.07M14.04G84.46
Strip-MLP-B(ours)58.49M9.22G86.40
", + "bbox": [ + 506, + 88, + 885, + 428 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5996c2853e1c8bba2eee66618d88b7a874f9c6707a1c8fb9c1e86679ba9b0488.jpg", + "table_caption": [ + "Table 4. Image classification results of different models on CIFAR-100. The patch number of CGSMM is C/2." + ], + "table_footnote": [ + "Table 5. The transfer learning results of the model pre-trained on ImageNet-1K and fine-tuned to CIFAR-10 and CIFAR-100." + ], + "table_body": "
ModelDatasetParamsTop-1(%)
ViT-S/16 [5]49M97.1
TST-14 [37]22M97.5
ViP-S/7 [10]CIFAR-1025M98.0
DynaMixer-S [35]26M98.2
Strip-MLP-T*(ours)18M98.8
ViT-S/16 [5]49M87.1
TST-14 [37]22M88.4
ViP-S/7 [10]CIFAR-10025M88.4
DynaMixer-S [35]26M88.6
Strip-MLP-T*(ours)18M89.4
", + "bbox": [ + 506, + 453, + 885, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "number of parameters) of three types of models. Strip-MLP outperforms the CNNs-based model: $+0.5\\%$ with half the number of parameters (25M vs. 39M) and FLOPs (3.7G vs. 8.0G) for Strip-MLP-T over RegNetY-8G [22]. Compared to Transformer-based models, e.g. Swin Transformer [19], Strip-MLP achieves comparable results but with obvious superiority in terms of the number of parameters and FLOPs, such as Strip-MLP-B achieves similar performance to Swin-B but with fewer parameters (57M vs. 88M) and FLOPs (9.2G vs. 15.4G). Compared to the MLP-based models, Strip-MLP still shows the same efficiency in the various models. The Strip-MLP- $\\mathrm{T}^*$ achieves average higher accuracy by $+2.12\\%$ than other MLP-based models (81.2% vs. 79.08%). Both Strip-MLP-B and Wave-MLP-B [27] models get the same accuracy of $83.6\\%$ . However, Strip-MLP-B uses fewer parameters (57M vs. 63M) and FLOPs (9.2G vs. 10.2G), indicating that the token-interaction power of our model is more efficient than other", + "bbox": [ + 496, + 628, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1500", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/44287b98728f96476dbeec2c03b3fd2075e323c115ea4c920a297164b721dd61.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Classification results of three kinds of models on ImageNet-1K without extra data. The group number of CGSMM is C/4. Throughput is tested on a single V100 GPU following [19]." + ], + "table_body": "
MethodParamsFLOPsThroughput (image/s)Top-1 (%)
CNNs-Based
ResNet50 [8]26M4.1G-78.5
ResNet101 [8]45M7.9G-79.8
RegNetY-4G [22]21M4.0G115780.0
RegNetY-8G [22]39M8.0G59281.7
RegNetY-16G [22]84M16.0G33582.9
Transformer-Based
DeiT-S [30]22M4.6G94079.8
DeiT-B [30]86M17.5G29281.8
Swin-T [19]29M4.5G75581.3
Swin-S [19]50M8.7G43783.0
Swin-B [19]88M15.4G27883.5
MLP-Based
ResMLP-S12 [29]15M3.0G141576.6
gMLP-S [16]20M4.5G-79.6
CycleMLP-B1 [1]15M2.1G103878.9
Hire-MLP-Tiny [7]18M2.1G156279.7
Wave-MLP-T [27]17M2.4G120880.6
Strip-MLP-T*(ours)18M2.5G81481.2
ResMLP-S24 [29]30M6.0G71579.4
Sparse-MLP-T [26]24M5.0G63981.9
CycleMLP-B2 [1]27M4.0G64181.6
Hire-MLP-Small [7]33M4.2G80882.1
Wave-MLP-S [27]30M4.5G72082.6
Strip-MLP-T(ours)25M3.7G59782.2
CycleMLP-B4 [1]52M10.1G32183.0
Sparse-MLP-S [26]49M10.3G36183.1
Wave-MLP-M [27]44M7.9G41383.4
Strip-MLP-S(ours)43M6.8G38183.3
ResMLP-B24 [29]116M23.0G23181.0
gMLP-B [16]73M15.8G-81.6
CycleMLP-B5 [1]76M12.3G24783.2
Sparse-MLP-B [26]66M14.0G27883.4
Hire-MLP-Base [7]58M8.1G44183.2
Wave-MLP-B [27]63M10.2G34183.6
Strip-MLP-B(ours)57M9.2G30083.6
", + "bbox": [ + 76, + 87, + 472, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MLP-based models. In addition, we apply the Strip MLP layer into Wave-MLP-T (namely Wave-Strip-MLP-T) on ImageNet-1K, and we set the strip width as 3. The Wave-Strip-MLP-T gets better results than the original Wave-MLP-T model: $81.2\\%$ vs. $80.6\\%$ . The performance improvement by $+0.6\\%$ shows that our Strip MLP makes token interaction more efficient and can easily be used in other MLP models.", + "bbox": [ + 75, + 627, + 468, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, as shown in Fig. 4, our Strip-MLP has significant superiority in performance and complexity over other state-of-the-art methods on different datasets, demonstrating our Strip-MLP is remarkably efficient and effective to enrich the token interaction power for vision MLP.", + "bbox": [ + 75, + 750, + 468, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 835, + 241, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To better show the effectiveness of the proposed method, we ablate the key components of the model design. Due to the limited GPU resources, we do ablation studies on the", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c088981642af2026df2d272c7aee33f19deb036828e817a08e2e91f9d538d058.jpg", + "table_caption": [], + "table_footnote": [ + "Table 7. Ablation study on the strip width of Strip MLP. We randomly select the model of Strip-MLP-S for Caltech-101 and Strip-MLP-S for CIFAR-100." + ], + "table_body": "
Strip WidthDatasetParamsFLOPsTop-1(%)
141.33M6.71G91.38
3Caltech-10143.66M6.83G92.09
545.99M6.96G91.72
748.33M7.08G91.86
140.71M6.71G86.09
3CIFAR-10041.88M6.83G86.50
543.04M6.96G86.39
744.21M7.08G86.35
", + "bbox": [ + 501, + 88, + 890, + 204 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7663e7ce023765890bb68c9d02f23eb87b4b538fe610b92a88c3845969584287.jpg", + "table_caption": [], + "table_footnote": [ + "Table 8. Ablation study about the patch number of the CGSMM. The base model is Strip-MLP-S. $C/1$ means splitting the feature $C$ into C patches in channel dimension, namely $P = C/1 = C$ ." + ], + "table_body": "
DatasetPatchesParamsFLOPsTop-1(%)
Caltech-101C/147.22M6.83G91.81
C/243.66M6.83G92.09
C/441.88M6.83G91.38
C/840.99M6.83G91.33
140.16M6.83G90.93
CIFAR-100C/147.22M6.83G86.20
C/243.66M6.83G86.18
C/441.88M6.83G86.50
C/840.99M6.83G85.98
140.16M6.83G85.03
", + "bbox": [ + 501, + 244, + 890, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "datasets of both Caltech-101 and CIFAR-100.", + "bbox": [ + 500, + 431, + 803, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of Strip Width in Strip MLP. The strip width affects the token interaction ranges and determines the field of each row or column's contribution to adjacent tokens. We further conduct experiments to verify its effect by varying the width from 1 to 7 with a step at 2. Tab. 7 shows the performances of our models on Caltech-101 [6] and CIFAR-100 [14], respectively. When the width becomes larger, the performance increases and tends to saturate, indicating that Strip MLP layer improves the token interaction power. In all of our comparison experiments in previous sections, we consistently set the strip width as 3 according to the ablation experiment results.", + "bbox": [ + 496, + 446, + 890, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of Patch Number in CGSMM. Different patch number brings varying degrees of token interaction power improvement and affects the model performance. In Tab. 8, we design five decreased patch numbers from C into 1 to show the effectiveness of CGSMM. Without the patch splitting $(P = 1)$ and GSML operation, the model performance decreases by $1.16\\%$ on Caltech-101 $(92.09\\% \\rightarrow 90.93\\%)$ and decreases by $1.47\\%$ on CIFAR-100 $(86.50\\% \\rightarrow 85.03\\%)$ , which consistently proves the effectiveness of CGSMM in improving the token interaction power. From the experimental results, we can observe that the optimal patch number is different on the datasets with scale differences, so the optimal patch number should be determined through validation experiments. We consistently set the patch number as C/4 in other ablation studies.", + "bbox": [ + 496, + 628, + 890, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cascade vs. Parallel Architecture of GSML. Applying the GSML in a cascade architecture enables the token interacts with other tokens across the entire 2D space in just one", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1501", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da2a372e6dd3a92b01b72bb3ae18d45f3cffe811b555ccc3b34dc7672c0eb6f5.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 96, + 97, + 287, + 215 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c6ee3e3a0618b95346da772b4215bffd339d7a08937fc9b6fdb4dbb75e7f2365.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 289, + 95, + 480, + 215 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c3f1c9808432c7e150dc2405f6bfff2ce1f15998572a6fcc8260daa8ee2cbb44.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 483, + 95, + 676, + 215 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/afdfca1c88a0a22a5aa7e35e65849535c686b829c3d08eb76305d1d5c7816207.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 679, + 95, + 874, + 215 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fcdc215a8d20a99fc4b69d5a13d78fbff2f79a03d09bcd8fd6899bd9ff484147.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 94, + 250, + 290, + 366 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2acffc65d5fb5f26ae4119e7494d70c001a34684da9aa66b1cd924dd772912c6.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 292, + 250, + 486, + 366 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5586cd5f719801abf8b86ece3cb56df4636167654ed88d3b79555aff15605919.jpg", + "image_caption": [ + "(g)", + "Figure 4. Performance comparison on four datasets of ImageNet-1K [4], CIFAR-100 [14], CIFAR-10 [14], and Caltech-101 [6]. The Top-1 accuracy, number of parameters and FLOPs are reported with networks of CNN-based, Transformer-based, and MLP-based models." + ], + "image_footnote": [], + "bbox": [ + 488, + 247, + 681, + 366 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ac8ff3de25819cc0eb6500cda587101794a58c3d13e72c45115709614a9b15a2.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 681, + 247, + 872, + 366 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1f3769d8ab40f0ea65c19ed2a3e65e02f96aa341c29f7cf9137bb31505c91711.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDatasetParamsFLOPsTop-1(%)
Parallel CascadeCaltech-10142.67M6.65G91.61
43.66M6.83G92.09
Parallel CascadeCIFAR-10040.90M6.66G85.89
41.88M6.83G86.50
", + "bbox": [ + 78, + 430, + 470, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 9. Ablation study on the cascade and parallel structure of the Strip MLP in Group Strip MLP. Parallel means applying the Strip MLP on the row and column direction in parallel. Cascade means the proposed CGSMM method in this paper.", + "bbox": [ + 76, + 503, + 468, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "module, which needs two modules for the parallel structure so that may decrease the efficiency of token interaction. In Tab. 9, we test the effects between the cascade structure of CGSMM and the parallel structure of Parallel Group Strip Mixing Module (PGSMM). Our experiments show that the cascade structure gets a higher accuracy compared to the parallel structure, with an increase of $+0.48\\%$ and $+0.61\\%$ on the Caltech-101 and CIFAR-100 datasets, respectively.", + "bbox": [ + 75, + 569, + 468, + 689 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Roles of CGSMM & LSMM. To verify the significance of two complementary branches, we conduct experiments where we only keep one of the branches of CGSMM or LSMM. Tab. 10 displays the outcomes of our trial on Caltech-101 and CIFAR-100. Notably, we found that removing any module from either of the two branches significantly reduced the model's performance. These results highlight the essential role of both CGSMM and LSMM in enriching the tokens' interaction power.", + "bbox": [ + 75, + 690, + 468, + 827 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper presents an efficient and effective token interaction MLP model of Strip-MLP for vision MLP. For the", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c45a9ea44b4206243170035fe9f110306809664613d6132de528ed4cab9e798c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDatasetParamsFLOPsTop-1(%)
CGSMM OnlyCaltech-10143.40M7.55G91.10
LSMM Only43.96M6.00G90.31
CGSMM & LSMM41.88M6.83G91.38
CGSMM OnlyCIFAR-10043.40M7.55G85.20
LSMM Only43.96M6.00G86.13
CGSMM & LSMM41.88M6.83G86.50
", + "bbox": [ + 500, + 430, + 895, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 10. Ablation study on the roles of two branches of CGSMM and LSMM. The base model is Strip-MLP-S. When any module branch has been removed, all input features would be fed into the remaining branch module.", + "bbox": [ + 500, + 527, + 890, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Token's interaction dilemma problem, we design the Strip Mixing Block to enrich the token interaction power by three strategies. Our experimental analysis reveals that the StripMLP remarkably improves the performances of the MLP-based models on small datasets and keeps comparable or even better results on ImageNet with great superiorities on the number of parameters and FLOPs. These observations clearly demonstrate that our models make token interaction more efficient among MLP-based models in image classification tasks. It is expected that Strip MLP layer has the potential to be a standard layer in the future and have promising performance on diverse vision tasks.", + "bbox": [ + 496, + 616, + 890, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement: This work is supported by National Key Research and Development Program of China (2021YFF1200800); Peng Cheng Laboratory Research Project (PCL2023AS6-1); and Guangdong Basic and Applied Basic Research Foundation under Grant 2022A1515110573.", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1502", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shoufa Chen, Enze Xie, Chongjian Ge, Ding Liang, and Ping Luo. Cycle.mp: A mlp-like architecture for dense prediction. In International Conference on Learning Representation (ICLR), Oral, 2022.", + "[2] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017.", + "[3] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005.", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations.", + "[6] Li Fei-Fei, Robert Fergus, and Pietro Perona. One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence, 28(4):594-611, 2006.", + "[7] Jianyuan Guo, Yehui Tang, Kai Han, Xinghao Chen, Han Wu, Chao Xu, Chang Xu, and Yunhe Wang. Hire-mlp: Vision mlp via hierarchical rearrangement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 826-836, 2022.", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[9] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016.", + "[10] Qibin Hou, Zihang Jiang, Li Yuan, Ming-Ming Cheng, Shuicheng Yan, and Jiashi Feng. Vision permutator: A permutable mlp-like architecture for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.", + "[11] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.", + "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015.", + "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.", + "[14] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009.", + "[15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Hanxiao Liu, Zihang Dai, David So, and Quoc V Le. Pay attention to mlps. Advances in Neural Information Processing Systems, 34:9204-9215, 2021.", + "[17] Ruiyang Liu, Yinghui Li, Linmi Tao, Dun Liang, and Haitao Zheng. Are we ready for a new paradigm shift? a survey on visual deep mlp. Patterns, 3(7):100520, 2022.", + "[18] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022.", + "[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021.", + "[20] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022.", + "[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019.", + "[22] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020.", + "[23] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.", + "[24] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015.", + "[25] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019.", + "[26] Chuanxin Tang, Yucheng Zhao, Guangting Wang, Chong Luo, Wenxuan Xie, and Wenjun Zeng. Sparse mlp for image recognition: is self-attention really necessary? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2344–2351, 2022.", + "[27] Yehui Tang, Kai Han, Jianyuan Guo, Chang Xu, Yanxi Li, Chao Xu, and Yunhe Wang. An image patch is a wave: Phase-aware vision mlp. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10935-10944, 2022.", + "[28] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261-24272, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1503", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Gautier Izacard, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, et al. Resmlp: Feedforward networks for image classification with data-efficient training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.", + "[30] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021.", + "[31] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022.", + "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[33] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020.", + "[34] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021.", + "[35] Ziyu Wang, Wenhao Jiang, Yiming M Zhu, Li Yuan, Yibing Song, and Wei Liu. Dynamixer: a vision mlp architecture with dynamic mixing. In International Conference on Machine Learning, pages 22691-22701. PMLR, 2022.", + "[36] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023.", + "[37] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021.", + "[38] David Junhao Zhang, Kunchang Li, Yunpeng Chen, Yali Wang, Shashwat Chandra, Yu Qiao, Luoqi Liu, and Mike Zheng Shou. Morphmlp: A self-attention free, mlp-like backbone for image and video. arXiv preprint arXiv:2111.12527, 2021.", + "[39] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018.", + "[40] Yipeng Zhang, Zahra M Aghajan, Matias Ison, Qiujing Lu, Hanlin Tang, Guldamla Kalender, Tonmoy Monsoor, Jie Zheng, Gabriel Kreiman, Vwani Roychowdhury, et al. De" + ], + "bbox": [ + 78, + 90, + 470, + 895 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "coding of human identity by computer vision and neuronal vision. Scientific reports, 13(1):651, 2023.", + "bbox": [ + 532, + 90, + 890, + 119 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1504", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/5c4682dc-11cf-43c9-b82f-059171805e73_model.json b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/5c4682dc-11cf-43c9-b82f-059171805e73_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e818c04896dab8cc14fd321cde0247ba1c94c10e --- /dev/null +++ b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/5c4682dc-11cf-43c9-b82f-059171805e73_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.204, + 0.131, + 0.768, + 0.154 + ], + "angle": 0, + "content": "Strip-MLP: Efficient Token Interaction for Vision MLP" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.18, + 0.77, + 0.217 + ], + "angle": 0, + "content": "Guiping Cao\\(^{1,2}\\) Shengda Luo\\(^{1}\\) Wenjian Huang\\(^{1}\\) Xiangyuan Lan\\(^{2,*}\\) Dongmei Jiang\\(^{2}\\) Yaowei Wang\\(^{2}\\) Jianguo Zhang\\(^{1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.218, + 0.75, + 0.252 + ], + "angle": 0, + "content": "\\(^{1}\\)Southern University of Science and Technology, Shenzhen, China \\(^{2}\\)Peng Cheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.255, + 0.796, + 0.288 + ], + "angle": 0, + "content": "12131099@mail.sustech.edu.cn, {luosd, huangwj, zhangjg}@sustech.edu.cn, {lanxy, jiangdm, wangyw}@pcl.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.354, + 0.474, + 0.807 + ], + "angle": 0, + "content": "Token interaction operation is one of the core modules in MLP-based models to exchange and aggregate information between different spatial locations. However, the power of token interaction on the spatial dimension is highly dependent on the spatial resolution of the feature maps, which limits the model's expressive ability, especially in deep layers where the feature are down-sampled to a small spatial size. To address this issue, we present a novel method called Strip-MLP to enrich the token interaction power in three ways. Firstly, we introduce a new MLP paradigm called Strip MLP layer that allows the token to interact with other tokens in a cross-strip manner, enabling the tokens in a row (or column) to contribute to the information aggregations in adjacent but different strips of rows (or columns). Secondly, a Cascade Group Strip Mixing Module (CGSMM) is proposed to overcome the performance degradation caused by small spatial feature size. The module allows tokens to interact more effectively in the manners of within-patch and cross-patch, which is independent to the feature spatial size. Finally, based on the Strip MLP layer, we propose a novel Local Strip Mixing Module (LSMM) to boost the token interaction power in the local region. Extensive experiments demonstrate that Strip-MLP significantly improves the performance of MLP-based models on small datasets and obtains comparable or even better results on ImageNet. In particular, Strip-MLP models achieve higher average Top-1 accuracy than existing MLP-based models by \\(+2.44\\%\\) on Caltech-101 and \\(+2.16\\%\\) on CIFAR-100. The source codes will be available at https://github.com/MedProcess/Strip_MLP." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.819, + 0.209, + 0.833 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.833, + 0.47, + 0.879 + ], + "angle": 0, + "content": "In computer vision, Convolutional Neural Networks (CNNs) are one of the most popular network backbones, which have made a series of breakthroughs [8]. Inspired by" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.321, + 0.834, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.413, + 0.895, + 0.471 + ], + "angle": 0, + "content": "Figure 1. Token interaction of the MLP layer on the down-sampled image feature. (a) The process of which the feature resolution will be down-sampled from \\( H \\times W \\times C \\) into a small spatial size of \\( \\frac{H}{8} \\times \\frac{W}{8} \\times 8C \\). (b) Token interaction on different feature resolutions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.895, + 0.642 + ], + "angle": 0, + "content": "the great success of self-attention-based architectures [32] in natural language processing (NLP), the Transformer models [5, 19, 34] are introduced into the field of computer vision, and have achieved comparable results with state-of-the-art (SOTA) CNNs. Although ViT [5] and its variants outperform traditional CNNs, the models introduce high computational complexity to construct attention maps. Recently, some studies [20, 26] in the vision community suggest that the attention mechanism is not necessary and simpler model architectures are proposed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.896, + 0.902 + ], + "angle": 0, + "content": "MLP-based models, like MLP-Mixer [28], gMLP [16] and ViP [10] process the data with the Multilayer Perceptrons (MLP), showing great potential to improve the performance of vision models [29]. As the first visual deep MLP network, MLP-Mixer [28] introduces two types of MLP layers: Channel-Mixing MLPs (CMM) and Token-Mixing MLPs (TMM). For CMM, the module mainly mixes the information between different channels of each token. For TMM, it allows each spatial token to interact with all other tokens (the whole image) in a single MLP layer. However, this design also introduces a much larger number of parameters and higher computational complexity prone to overfitting. To address this problem, Sparse MLP (SMLP) [26] and Vision Permutator (ViP) [10] propose a similar layer of parallel structure, which applies one dimension MLP along the axial directions, and parameters are shared among rows or columns, respectively. Therefore, it reduces the number" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1494" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.467, + 0.121 + ], + "angle": 0, + "content": "of model parameters and computational complexity, avoiding the common over-fitting problem." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.124, + 0.468, + 0.424 + ], + "angle": 0, + "content": "Although SMLP and ViP alleviate some deficiencies of MLP-Mixer [28], both methods bring the challenge that the token interaction power is highly dependent on the feature spatial size when interacting tokens on spatial rows (or columns). As shown in Fig. 1, the spatial feature resolution is down-sampled to a small size but with more channels, which means the feature pattern of each token is mainly concentrated on the channel dimension rather than the spatial one. Interacting tokens along the spatial dimension by sharing the weights among all channels would seriously ignore the feature pattern differences among different channels, which may degrade the token interaction power, especially in deep layers with small spatial feature resolution. Here, we mark this problem as the Token's interaction dilemma. Taking SMLP as an example, we analyze the feature resolution and complexity of the model in different stages in detail (seen in Sec. 3.4). We find that as the spatial feature size decreases by down-sampling stage by stage, the token interaction layer also becomes smaller and smaller, which makes the token interaction power degraded rapidly." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.427, + 0.468, + 0.742 + ], + "angle": 0, + "content": "To address these challenges, we propose a new efficient Strip MLP model, dubbed Strip-MLP, to enrich the power of the token interaction layer in three ways. For the level of a single MLP layer, inspired by the cross-block normalization schemes of HOG [3] and the sparse connections between the biological neurons, we design a Strip MLP layer to allow the token to interact with other tokens in a crossstrip manner, enabling each row or column of the tokens to contribute differently to other rows or columns. For the token interaction module level, we develop channel-wise group mixing of CGSMM, enabling the tokens in a row (or column) to contribute to the information aggregations in adjacent but different strips of rows (or columns). to tackle the problem that the token interaction power decreases in deep layers with the spatial feature size significantly reduced but with multiplying channels. Considering the existing methods [10, 26, 28] interact the tokens mainly in the long range of row (or column), which may not aggregate tokens well in the local region, we propose a new Local Strip Mixing Module (LSMM) with a small Strip MLP unit to strengthen the token interaction power on local interactions." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.744, + 0.467, + 0.773 + ], + "angle": 0, + "content": "The proposed Strip-MLP model significantly boosts the token interaction power, and the main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.781, + 0.468, + 0.854 + ], + "angle": 0, + "content": "- A new MLP paradigm for vision MLP: Strip MLP layer, which aggregates the adjacent tokens in a cross strip manner and enables each row or column of the tokens to contribute differently to other rows or columns, interacting tokens more efficiently." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- Designing a Cascade Group Strip Mixing module and a Local Strip Mixing module, which effectively improves the model's token interaction power and boosts" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.781, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.89, + 0.107 + ], + "angle": 0, + "content": "the tokens aggregation in the local region, respectively;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.108, + 0.89, + 0.241 + ], + "angle": 0, + "content": "- Extensive experiments show that Strip-MLP remarkably improves the performances of the MLP-based models. Strip-MLP achieves higher average Top-1 accuracy by \\(+2.44\\%\\) on Caltech-101 and \\(+2.16\\%\\) on CIFAR-100 over the existing MLP-based models. In addition, our models achieve comparable or even better performances on ImageNet-1K compared with traditional MLP-based models, and other popular CNNs and transformer-based models." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.261, + 0.64, + 0.276 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.281, + 0.89, + 0.34 + ], + "angle": 0, + "content": "Deep neural networks for vision recognition can be mainly divided into three categories: Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Multi-Layer Perceptron-based models (MLPs) [10]." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.342, + 0.89, + 0.522 + ], + "angle": 0, + "content": "CNNs-Based Models. CNNs are the de-facto standard deep learning network models for vision tasks and have been intensely studied in the vision community. AlexNet [15] is a symbolically significant model that won the ILSVRC 2012 contest with far higher performance than others. Since then, CNNs-based models have attracted more attentions, and lots of deeper and more effective architectures [8, 11, 23-25, 33, 39, 40] have been proposed. With the convolution and pooling layer, CNNs aggregate the feature in a local region but not well in the long-term dependencies, which are optimized by the new vision model like Transformer [32] and MLP [28] models." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.524, + 0.89, + 0.749 + ], + "angle": 0, + "content": "Transformer-Based Models. Transformer [32] is introduced for machine translation and becomes the reference model for all-natural language processing (NLP) tasks. Inspired by the great success of Transformer in NLP, ViTs [5] first applies a standard Transformer to images, which attains excellent results compared to the SOTA CNNs model. DeiT [30] introduces several training strategies and distillation methods to make ViTs more effective on the smaller ImageNet-1K dataset. By proposing a hierarchical Transformer with shifted windows, Swin Transformer [19] achieves SOTA accuracy on ImageNet-1K, bringing greater efficiency by self-attention in local window and cross-window connection. For these models, self-attention is the core module but with heavy computational burdens to obtain an attention map." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.75, + 0.89, + 0.9 + ], + "angle": 0, + "content": "MLP-Based Models. Without the convolutions and self-attention mechanism, MLP-Mixer [28] builds the architecture that only uses the MLP layer and achieves competitive performance on image classification benchmarks. Since then, the researchers have developed many MLP-like variants [1, 7, 16, 26, 27, 29, 31, 35, 38] models. The work [17] gives a comprehensive survey about visual deep MLP models and compares the intrinsic connections and differences between convolution, self-attention mechanism, and Token-mixing MLP in detail. Sparse MLP [26] introduced a sparse" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.947, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1495" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.09, + 0.822, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.337, + 0.894, + 0.38 + ], + "angle": 0, + "content": "Figure 2. The overall and components architecture of Strip-MLP. (a) Strip-MLP has four stages, and \\( T1 \\) to \\( T4 \\) means the repeated times of the block in each stage. (b) The Strip Mixing Block splits the head (channel dimension) into two heads, and they are fed into the two parallel branches. (c) Channel Mixing Block architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.473, + 0.62 + ], + "angle": 0, + "content": "operation to separately aggregate information along axial directions, avoiding the quadratic computational complexity of conventional MLP. Hire-MLP [7] presents a novel variant of MLP-based architecture via hierarchically rearranging tokens to aggregate local and global spatial information. Wave-MLP [27] represents each token as a wave function with two parts of amplitude and phase, which has the ability to model varying contents from different input images. Above all, there is still the limitation of token interaction power degraded significantly especially when the spatial feature resolution becomes small, which has been overlooked by previous studies. In this paper, we aim to enrich the token interaction power on both the single MLP layer and token interaction module to advance the performance of MLP-based models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.633, + 0.17, + 0.648 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.789 + ], + "angle": 0, + "content": "In this section, we first present the overall architecture of Strip-MLP. Then, we show the key proposed components: Strip MLP Layer, Cascade Group Strip Mixing Module (CGSMM) and Local Strip Mixing Module (LSMM) of the model in detail, and present the comparison analysis between the Strip MLP and traditional MLP model in terms of parameters and complexity. Finally, we define four kinds of architecture variants to compare the performance of the models in different sizes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.793, + 0.274, + 0.807 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.885 + ], + "angle": 0, + "content": "An overview of the Strip-MLP model is depicted in Fig. 2 (a). We design the Strip-MLP as a hierarchical structure of a multi-stages model. Given the input image \\(I\\), Strip-MLP model makes patch embedding and models the image feature in four stages." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Patch Embedding. Strip-MLP firstly splits the input" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.892, + 0.483 + ], + "angle": 0, + "content": "image \\(I \\in \\mathbb{R}^{H \\times W \\times 3}\\) (H, W: image height and width) into a sequence of image patches (also referred as tokens) \\(I_{p} \\in \\mathbb{R}^{hw \\times c}\\) (c: the number of channels), where the patch size is \\(p \\times p\\) and patch number is \\(hw\\left(h = \\frac{H}{p}, w = \\frac{W}{p}\\right)\\), and then all patches are linearly projected into a desired higher dimension \\(C\\) of the feature \\((X \\in \\mathbb{R}^{hw \\times C})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.893, + 0.711 + ], + "angle": 0, + "content": "Mixing Block. The purpose of mixing block is to boost the interactions between the features of diverse spatial locations and channels. To effectively aggregate spatial and channel information, we design two sub-blocks of the Strip Mixing Block and Channel Mixing Block. The Strip Mixing Block is comprised of CGSMM and LSMM, which blend and aggregate spatial information more efficiently at the global and local level, respectively. Patch merging module aims to merge the feature in which the spatial dimension will be reduced by \\(2 \\times 2\\), and the channel dimension increases by 2 times so that the model down-samples the feature from \\(\\frac{H}{p} \\times \\frac{W}{p}\\) into \\(\\frac{H}{8p} \\times \\frac{W}{8p}\\) stage by stage. To obtain multi-scale features, we apply a single convolutional layer to the output features of stages 1 and 2. Then, we add the resulting features to the input features of stages 3 and 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.711, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Head Layer. The features extracted by the multiple blocks are fed into a global average pooling (GAP) layer to reduce the feature dimension. Finally, the feature will be fed into a fully-connected head layer for classification." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.777, + 0.892, + 0.793 + ], + "angle": 0, + "content": "3.2. Strip Mixing Block and Channel Mixing Block" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Strip Mixing Block. To improve the token interaction power, we design the block to aggregate both long and short range interactions in a parallel manner. As illustrated in Fig. 2 (b), we split the feature in channel dimension, with one-half of the channel feature being fed into CGSMM to model the long-range interactions, and the remaining feature being fed into LSMM to aggregate the local interac" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1496" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.09, + 0.647, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.252, + 0.496, + 0.264 + ], + "angle": 0, + "content": "(a) Cascade Group Strip Mixing Module" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.09, + 0.818, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.252, + 0.808, + 0.264 + ], + "angle": 0, + "content": "(b) Local Strip Mixing Module" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.268, + 0.894, + 0.296 + ], + "angle": 0, + "content": "Figure 3. (a) The architecture of CGSMM. It contains the Strip MLP layer, the feature splitting, restoring, and permuting operation in a cascade mode. (b) The structure of the Local Strip Mixing module. The dark green block composes of a re-weight module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.47, + 0.35 + ], + "angle": 0, + "content": "tions. Given the input feature \\(X\\in \\mathbb{R}^{H\\times W\\times C}\\), the block can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.355, + 0.469, + 0.371 + ], + "angle": 0, + "content": "\\[\nX _ {m} = M L P \\left(\\left(D W S C (X)\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.374, + 0.469, + 0.396 + ], + "angle": 0, + "content": "\\[\nY = F C \\left(C a t \\left(C G S M M \\left(X _ {m _ {1}} ^ {\\frac {C}{2}}\\right), L S M M \\left(X _ {m _ {2}} ^ {\\frac {C}{2}}\\right)\\right)\\right) + X \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.396, + 0.469, + 0.486 + ], + "angle": 0, + "content": "where \\(X_{m}\\) and \\(Y\\) are the intermediate and output feature of the block. \\(X_{m}\\) is split into \\(X_{m_1}\\) and \\(X_{m_2}\\) with half channel, respectively. DWSC means depth-wise convolution [2], and the kernel size is \\(3\\times 3\\). MLP is a serial connection of fullyconnection (FC), batch normalization [12] and GELU [9] activation layers. \\(Cat(*)\\) denotes concatenation operation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.469, + 0.577 + ], + "angle": 0, + "content": "Channel Mixing Block. The block aims to aggregate the information between channels of the token, and the basic structure is shown in Fig. 2 (c). Following the inverted bottleneck in [20] and global response normalization (GRN) in [36], we design the Channel Mixing block to increase the contrast and selectivity of channels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.581, + 0.279, + 0.595 + ], + "angle": 0, + "content": "3.3. CGSMM and LSMM" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.78 + ], + "angle": 0, + "content": "We first introduce the Strip MLP Layer. Based on this layer, we design CGSMM using a simple but effective strategy that splits the feature into patches along channel dimension and interacts tokens more effectively in the manners of within-patch and cross-patch, which is independent to the feature spatial size. In addition, the design of existing methods [10, 26, 28] only allows tokens to interact in a long range of rows (or columns) with sharing weight; thus the resulting model may struggle with effectively aggregating both global and local information simultaneously. Therefore, we design LSMM to capture the local interactions more efficiently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Strip MLP Layer. In MLP-based models, the majority of MLP layers treat each row and column of the data in isolation, as formulated in Eq. (3), which may lead to the inefficiency of token interaction. Inspired by the cross-block normalization schemes of HOG [3], which overlaps the blocks and enables each scalar cell response to contribute into different blocks, we propose the Strip MLP layer. The proposed layer applies MLP on the \"strip\" data of adjacent" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.321, + 0.892, + 0.379 + ], + "angle": 0, + "content": "rows or columns in the spatial direction in order to aggregate the feature in a cross-strip manner. Given the input \\( X \\), we formulated the Strip MLP layer (setting strip width as 3 for example) in Eqs. (4) and (5):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.379, + 0.891, + 0.398 + ], + "angle": 0, + "content": "\\[\nX _ {*, j} ^ {h} = W _ {1} X _ {*, j}; \\quad X _ {i, *} ^ {w} = W _ {2} X _ {i, *} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.4, + 0.891, + 0.418 + ], + "angle": 0, + "content": "\\[\nX _ {*, j} ^ {h} = W _ {3} \\operatorname {C a t} \\left(X _ {*, j - 1}, X _ {*, j}, X _ {*, j + 1}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.424, + 0.891, + 0.441 + ], + "angle": 0, + "content": "\\[\nX _ {i, *} ^ {w} = W _ {4} \\operatorname {C a t} \\left(X _ {i - 1, *}, X _ {i, *}, X _ {i + 1, *}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.638 + ], + "angle": 0, + "content": "where \\(W_{1}\\sim W_{4}\\) are the weights of the MLP layer, \\(i\\) and \\(j\\) are the token index in row and column. The superiorities of the Strip MLP layer are mainly in two aspects: on one hand, Strip MLP layer enables the token to interact with other tokens in both short and long spatial ranges simultaneously. On the other hand, similar to the HOG cross-block normalization process, each row (or column) not only serves for the current row (or column) tokens aggregation but also makes contribution to the adjacent rows (or columns) feature aggregation. For example, in Eq. (4), \\(X_{*,j}\\) makes different contributions for aggregating the processed feature of \\(X_{*,j - 1}^{h},X_{*,j}^{h}\\) and \\(X_{*,j + 1}^{h}\\), making tokens interacting more efficiently in a cross-strip manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.714 + ], + "angle": 0, + "content": "Cascade Group Strip Mixing Module. The architecture is shown in Fig. 3 (a). The module applies the Strip MLP layer in row and column direction in a cascade mode. As the operation of the Strip MLP layer on row is similar to that on column, we take one of them to show the method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Patch Splitting of the Data along Channel Dimension. Given the input feature \\(X_{m_1} \\in \\mathbb{R}^{H \\times W \\times \\frac{C}{2}}\\), the module firstly permutes the feature into \\(X_{m_1}^{H \\times \\frac{C}{2} \\times W}\\), and splits the feature into \\(P\\) patches in the channel dimension and concatenated along the row \\((X^{H \\times \\frac{C}{2P} \\times PW})\\) dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Group Strip MLP Layer. To improve the token interaction power, we propose a Group Strip MLP Layer (GSML) to interact tokens in the manners of within-patch and cross-patch. In particular, we apply the unshared weights of Strip MLP on different patches, and the tokens within the same patch share the weights for interaction. Then, we restore the feature to original shape and concatenate it with the input" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1497" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "feature. To interact the tokens crossing patches, we apply a channel fully-connected (Channel FC) layer to interact tokens between different patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.47, + 0.262 + ], + "angle": 0, + "content": "Local Strip Mixing Module. The module is illustrated in Fig. 3 (b). In order to better aggregate the local interactions on spatial dimension, we define a small Strip MLP unit, where the strip width and length are 3 and 7, respectively. Given the input feature \\( X_{m_2}^{H \\times W \\times \\frac{C}{2}} \\), we aggregate the local interactions in rows and columns direction simultaneously. Following [1, 10, 27], we sum all branches with a re-weight module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.276, + 0.391, + 0.292 + ], + "angle": 0, + "content": "3.4. Parameter and Complexity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.469, + 0.446 + ], + "angle": 0, + "content": "In multi-stage processing architecture, the Token's interaction dilemma becomes more serious, as the feature spatial resolution will be down-sampled in deep layers, and the token interaction power will be weakened. In this section, we present the comparison analysis of parameters and complexity to show the effectiveness of the proposed CGSMM. Considering the majority MLP-Based models have the similar model structure of token interaction, we analyze the parameters and complexities with Sparse MLP [26], which is a popular model with good performance on various datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Sparse MLP [26] firstly applies MLP on the columns and rows of \\(X\\) to map \\(\\mathbb{R}^{CW\\times H}\\) to \\(\\mathbb{R}^{CW\\times H}\\) separately. Then, the model concatenates the two processed features with the input feature \\(X\\) and uses a linear layer on the channel dimension to fuse the feature from \\(\\mathbb{R}^{HW\\times 3C}\\) to \\(\\mathbb{R}^{HW\\times C}\\). The number of parameters and FLOPs of the first step are \\(W^2 +H^2\\) and \\(CHW(H + W)\\); they are \\(3C^2\\) and \\(3HWC^2\\) for the fusion step. Comparatively, the number of parameters and FLOPs of Strip MLP layer for token interaction are \\(3P(H^{2} + W^{2})\\) and \\(3CHW(H + W)\\), respectively. The fusion step has \\(4C^2\\) parameters and \\(4HWC^2\\) FLOPs. In particular, for \\(H_{1} = W_{1} = 56\\), \\(H_{4} = W_{4} = 7\\), \\(C_1 = 112\\), \\(C_4 = 112\\times 2^3 = 896\\), and \\(P = \\frac{C}{4}\\), we calculate the number of parameters and FLOPs of both two layers. As shown in Tab. 1, for the Sparse MLP block, the number of parameters for token interaction in stage 4 (only \\(0.10k\\)) decreases by 62.70 times than in stage 1. In addition, most of the number of the block's parameters and the FLOPs concentrate in the fusion step. For example, only \\(0.01\\%\\) parameters (\\(0.52\\%\\) FLOPs) are used for the token interaction step in stage 4. Based on the above analysis, to improve the token interaction power, it would be better to redesign the block to balance the number of parameters and FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In CGSMM, we use a simple but effective strategy that splits the feature along channel dimension into patches and interacts the tokens in the manners of within-patch and cross-patch. No matter how the spatial resolution decreases, the module can still interact the tokens with channel-wise specificity in different patches. In Tab. 1, we improve the Proportion of the number of parameters in the token inter" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.09, + 0.887, + 0.188 + ], + "angle": 0, + "content": "
ItemsSparse MLPStrip MLP
ParamsFLOPsParamsFLOPs
Stage 16.27k39.34M526.85k118.01M
Stage 40.10k0.62M65.86k1.84M
Changes↓62.70↓63.45↓8.00↓64.14
Stage 4 Fusion2.41M118.01M3.21M157.35M
Proportion0.01%0.52%2.01%1.16%
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.189, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Table 1. The model parameters and complexity comparison between Sparse MLP and Strip MLP layer in different stages. Stage 1 and Stage 4 show the parameters (Params) and floating-point operations (FLOPs) for token interaction in one layer between different stages. Stage 4 Fusion indicates the Params and FLOPs in the feature fusion step of the token interaction module. Proportion means the ratio between the token interaction step and other steps in the token interaction module of stage 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.449 + ], + "angle": 0, + "content": "action layer from \\(0.01\\%\\) to \\(2.01\\%\\) in stage 4. Although our GSML brings more parameters and computations when just considering one token interaction layer, the total number of parameters and overall computational complexity of the model are decreased as the token interaction power is improved and we set small model configurations of \\(T_{1} \\sim T_{4}\\). Additionally, the experiments in Sec. 4 show that our models achieve better performance with fewer parameters and FLOPs than other SOTA models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.704, + 0.469 + ], + "angle": 0, + "content": "3.5. Architecture Variants" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.565 + ], + "angle": 0, + "content": "We developed four variants of our Strip-MLP network: Strip-MLP-T\\* (light tiny), Strip-MLP-T (tiny), Strip-MLPSmall), Strip-MLP-B (base), which have similar or even smaller model sizes, compared with MLP-baed models [7, 26, 27] and Swin Transformer [18], respectively. The hyperparameters of the four variants of the models are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.575, + 0.865, + 0.591 + ], + "angle": 0, + "content": "- Strip-MLP-T*: \\( C = 80, \\{T_1 \\sim T_4\\} = \\{2, 2, 6, 2\\} \\);" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.591, + 0.865, + 0.606 + ], + "angle": 0, + "content": "- Strip-MLP-T: \\(C = 80, \\{T_1 \\sim T_4\\} = \\{2, 2, 12, 2\\}\\);" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.606, + 0.865, + 0.621 + ], + "angle": 0, + "content": "- Strip-MLP-S: \\(C = 96, \\{T_1 \\sim T_4\\} = \\{2, 2, 18, 2\\}\\);" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.621, + 0.874, + 0.637 + ], + "angle": 0, + "content": "- Strip-MLP-B: \\(C = 112, \\{T_1 \\sim T_4\\} = \\{2, 2, 18, 2\\}\\);" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.575, + 0.874, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.703 + ], + "angle": 0, + "content": "where \\( C \\) means the channel number of the hidden layers in the first stage, and \\( T_{1} \\sim T_{4} \\) is the number of times repeated for Strip Mixing Block and Channel Mixing Block in each of the four stages, as shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.713, + 0.633, + 0.729 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.839 + ], + "angle": 0, + "content": "We compare our Strip-MLP with three main categories of networks: CNNs-based, Transformer-based, and MLP-based models. We first show the experimental settings of the datasets and implementation details. Then, the proposed method is compared with other popular methods on three datasets. Finally, we ablate the critical design components of Strip-MLP network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.85, + 0.707, + 0.866 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.874, + 0.892, + 0.904 + ], + "angle": 0, + "content": "Datasets. Methods are evaluated in four datasets which vary significantly in the number of training images:" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1498" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.462, + 0.434 + ], + "angle": 0, + "content": "
MethodParamsFLOPsTop-1(%)
CNNs-Based
ResNet50 [8]23.71M4.12G89.61
ResNet101 [8]42.71M7.85G88.56
ResNet152 [8]58.35M11.58G89.04
Transformer-Based
ViT-B/16 [5]85.85M16.86G53.96
Swin-T [19]27.60M4.36G80.40
Swin-S [19]48.91M8.52G79.83
Swin-B [19]86.85M15.14G78.42
MLP-Based
Wave-MLP-T [27]16.73M2.48G86.89
Hire-MLP-T [7]17.58M2.16G87.85
Strip-MLP-T*(ours)17.67M2.53G90.11
Wave-MLP-S [27]30.25M4.55G87.43
Hire-MLP-S [7]32.65M4.26G88.48
Sparse-MLP-T [26]23.55M5.02G90.54
Strip-MLP-T(ours)24.33M3.67G90.93
Wave-MLP-M [27]43.59M7.93G88.45
Hire-MLP-B [7]57.77M8.17G88.62
Sparse-MLP-S [26]47.87M10.36G91.07
Strip-MLP-S(ours)43.66M6.83G92.09
Wave-MLP-B [27]62.90M10.27G88.81
Hire-MLP-L [7]95.03M13.50G88.19
Sparse-MLP-B [26]65.07M14.04G91.67
Strip-MLP-B(ours)58.49M9.22G92.26
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.435, + 0.47, + 0.475 + ], + "angle": 0, + "content": "Table 2. Image classification results of our Strip-MLP and other models on Caltech-101. Top-1 denotes the Top-1 accuracy. Here, the patch number of CGSMM is C/2." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.493, + 0.468, + 0.551 + ], + "angle": 0, + "content": "- Caltech-101 [6]: it is a benchmark dataset that includes 101 classes, totally around \\(9k\\) images. We randomly take \\(80\\%\\) of each class of the data as the training set and use the remaining data as the testing set." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.553, + 0.468, + 0.581 + ], + "angle": 0, + "content": "- CIFAR-10 [14]: it consists of \\(60k \\times 32\\) images in 10 classes, and there are \\(50k\\) images" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.584, + 0.468, + 0.626 + ], + "angle": 0, + "content": "- CIFAR-100 [14]: it has 100 classes containing 600 images of each class (totally \\(60k\\)), and there are 500 training images and 100 testing images per class." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.629, + 0.468, + 0.672 + ], + "angle": 0, + "content": "- ImageNet-1K [4]: it is an image classification benchmark containing \\( 1.28M \\) training images and \\( 50k \\) testing images of 1000 classes." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.493, + 0.468, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.468, + 0.828 + ], + "angle": 0, + "content": "Training Details. In all of our experiments, the input image size is \\(224 \\times 224\\), and the input patch size is \\(4 \\times 4\\). Based on the deep-learning framework of PyTorch [21], we train our models from scratch without any extra data for pre-training except for the transfer learning experiments. We employ an AdamW [13] optimizer with 300 epochs for all training datasets and apply a cosine decay learning rate scheduler and 30 epochs of linear warm-up. In addition, we adopt the same augmentation and regularization strategies as to Swin-Transformer [19]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.838, + 0.41, + 0.853 + ], + "angle": 0, + "content": "4.2. Image Classification on Small Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.899 + ], + "angle": 0, + "content": "We conduct experiments on Small Datasets of Caltech-101, CIFAR-10 and CIFAR-100. The results consistently indicate that our Strip-MLP efficiently improves the token" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "interaction power, and advances the MLP-based models' performance on the small dataset than other models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.124, + 0.891, + 0.213 + ], + "angle": 0, + "content": "Results on Caltech-101. Tab. 2 presents the image classification result of three types models on Caltech-101 dataset. Compared to ResNet [8], all the four variants of our Strip-MLP models achieve higher Top-1 accuracy with fewer parameters and FLOPs, demonstrating an average increase of \\(+2.69\\%\\) over three widely-used ResNet models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.891, + 0.305 + ], + "angle": 0, + "content": "For the Transformer-based models, e.g. Swin Transformer [19], Strip-MLP models noticeably surpass all variants models: \\(+10.53\\% / 12.26\\% / 13.84\\%\\) over SwinTransformer (T/S/B) models. ViT-B/16 [5] only gets an accuracy of \\(53.96\\%\\), indicating that transformer-based models depend more on the scale of the datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.891, + 0.397 + ], + "angle": 0, + "content": "When compared to the MLP-based models, like Sparse MLP [26], Wave-MLP [27] and Hire-MLP [7], our models still achieve the best accuracy with fewer model parameters and FLOPs. For example, with fewer FLOPs \\((-4.28\\mathrm{G})\\) Strip-MLP-B \\((92.26\\%)\\) has an increase of \\(+3.64\\%\\) Top-1 accuracy than Hire-MLP-B \\((88.62\\%)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Results on CIFAR-10. Tab. 3 shows the classification results of performance comparisons. Compared with CNNs-based models, all four variants of our Strip-MLP models achieve higher Top-1 accuracy with fewer parameters and FLOPs, with an average increase of \\(+1.14\\%\\) (98.06% vs. 96.92%). When compared with transformer-based models, our Strip-MLP achieves better performance by \\(+3.06\\%\\) (98.06% vs. 95.00%) with fewer parameters (34.49M vs. 54.37M) and FLOPs (5.56G vs. 9.34G). Compared with MLP-based models, all of our Strip-MLP models obtain the best Top-1 accuracy with fewer parameters and FLOPs. For example, the average Top-1 accuracy of Strip-MLP is higher than Hire-MLP/Wave-MLP/Sparse-MLP by \\(+1.45\\% / 0.69\\% / 0.19\\%\\) (98.06% vs. 96.61%/97.37%/97.87%) with fewer average parameters (34.49M vs. 50.71M/38.32M/45.43M) and fewer average FLOPs (5.56G vs. 9.36G/6.31G/9.81G)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.763 + ], + "angle": 0, + "content": "In particular, with only 16.76M parameters and 2.53G FLOPs, our Strip-MLP-\\(T^{*}\\) noticeably surpasses all variants of ResNet [8] and Swin-Transformer [19] models, and Strip-MLP-\\(T^{*}\\) outperforms the transfer learning models of ViT-S/16 [5]/TST-14 [37] by \\(+0.69\\% / +0.29\\%\\), consistently demonstrating our Strip-MLP has significant superiorities in performance and complexity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Results on CIFAR-100. Tab. 4 compares the image classification results on CIFAR-100 with three kinds of architecture models. Compared with CNNs-based models [8], Strip-MLP obtains average higher accuracy of \\(+2.05\\% \\sim +2.43\\%\\) with a smaller number of parameters and FLOPs. When compared to the self-attention based model, e.g. Swin Transformer [19], our model shows great superiority that Strip-MLP-B achieves \\(86.40\\%\\) Top-1 accuracy, \\(+9.64\\%\\) than the Swin-B model with fewer pa" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1499" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.429 + ], + "angle": 0, + "content": "
MethodParamsFLOPsTop-1 (%)
CNNs-Based
ResNet50 [8]23.53M4.12G97.06
ResNet101 [8]42.52M7.85G96.70
ResNet152 [8]58.16M11.58G96.99
Transformer-Based
Swin-T [19]27.53M4.36G95.25
Swin-S [19]48.84M8.52G94.87
Swin-B [19]86.75M15.14G94.87
MLP-Based
Wave-MLP-T [27]16.69M2.48G96.91
Hire-MLP-T [7]17.53M2.16G96.20
Strip-MLP-T*(ours)16.76M2.53G97.79
Wave-MLP-S [27]30.20M4.55G97.27
Hire-MLP-S [7]32.61M4.26G96.70
Sparse-MLP-T [26]23.49M5.02G97.67
Strip-MLP-T(ours)23.07M3.67G97.94
Wave-MLP-M [27]43.55M7.93G97.56
Hire-MLP-B [7]57.72M8.17G96.88
Sparse-MLP-S [26]47.80M10.36G97.99
Strip-MLP-S(ours)41.81M6.83G98.23
Wave-MLP-B [27]62.83M10.27G97.72
Hire-MLP-L [7]94.96M13.50G96.64
Sparse-MLP-B [26]64.99M14.04G97.95
Strip-MLP-B(ours)56.33M9.22G98.28
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.078, + 0.43, + 0.47, + 0.471 + ], + "angle": 0, + "content": "Table 3. Image classification results of different models on CIFAR-10. The patch number of CGSMM is C/4. All models are trained from scratch without any extra data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.48, + 0.47, + 0.587 + ], + "angle": 0, + "content": "rameters (58.49M vs. 86.85M) and lower FLOPs (9.22G vs. 15.14G). In addition, our models keep the superiority over the other MLP-based models in that Strip-MLP achieves the best Top-1 accuracy (with a higher range of \\(+1.04\\% \\sim +3.01\\%\\)) compared to the MLP-based models, e.g. Wave-MLP [27], Hire-MLP [7], and Sparse MLP [26], with fewer parameters and FLOPs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.597, + 0.398, + 0.613 + ], + "angle": 0, + "content": "4.3. Transfer Learning on Small Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.471, + 0.827 + ], + "angle": 0, + "content": "Following the transfer learning experiment of DynaMixer [35], we show the advantages of our models on the transfer learning tasks. With the pre-trained Strip-MLP\\(\\mathrm{T}^*\\) model on the ImageNet-1K, we transfer the model to downstream datasets of CIFAR-10 and CIFAR-100. Compared to the previous models, such as ViT [5], ViP [10] and DynaMixer [35] models, our Strip-MLP\\(\\mathrm{T}^*\\) models achieve the best accuracies of \\(98.8\\%\\) and \\(89.4\\%\\) on two datasets, as shown in Tab. 5, \\(+1.7\\% / 0.6\\%\\) than ViT-S/16 and DynaMixer-S on CIFAR-10, \\(+2.3\\% / 0.8\\%\\) over ViT-S/16 and DynaMixer-S on CIFAR-100. More importantly, the number of parameters of our model is only 18M, which is far less than other models, indicating that our model is more efficient and effective." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.838, + 0.402, + 0.854 + ], + "angle": 0, + "content": "4.4. Image Classification on ImageNet-1K" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We conduct experiments on ImageNet-1K to show the effectiveness of the method. Tab. 6 shows the performance together with the complexity (in terms of FLOPS and the" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.089, + 0.887, + 0.429 + ], + "angle": 0, + "content": "
MethodParamsFLOPsTop-1(%)
CNNs-Based
ResNet50 [8]23.71M4.12G83.03
ResNet101 [8]42.71M7.85G83.75
ResNet152 [8]58.35M11.58G84.35
Transformer-Based
Swin-T [19]27.60M4.36G78.56
Swin-S [19]48.91M8.52G78.10
Swin-B [19]86.85M15.14G76.76
MLP-Based
Wave-MLP-T [27]16.73M2.48G82.77
Hire-MLP-T [7]17.58M2.16G82.44
Strip-MLP-T*(ours)17.67M2.53G85.10
Wave-MLP-S [27]30.25M4.55G83.34
Hire-MLP-S [7]32.65M4.26G82.94
Sparse-MLP-T [26]23.55M5.02G84.19
Strip-MLP-T(ours)24.33M3.67G85.23
Wave-MLP-M [27]43.59M7.93G83.95
Hire-MLP-B [7]57.77M8.17G83.17
Sparse-MLP-S [26]47.87M10.36G84.27
Strip-MLP-S(ours)43.66M6.83G86.18
Wave-MLP-B [27]62.89M10.27G84.23
Hire-MLP-L [7]95.03M13.50G83.53
Sparse-MLP-B [26]65.07M14.04G84.46
Strip-MLP-B(ours)58.49M9.22G86.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.502, + 0.43, + 0.89, + 0.454 + ], + "angle": 0, + "content": "Table 4. Image classification results of different models on CIFAR-100. The patch number of CGSMM is C/2." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.454, + 0.887, + 0.594 + ], + "angle": 0, + "content": "
ModelDatasetParamsTop-1(%)
ViT-S/16 [5]49M97.1
TST-14 [37]22M97.5
ViP-S/7 [10]CIFAR-1025M98.0
DynaMixer-S [35]26M98.2
Strip-MLP-T*(ours)18M98.8
ViT-S/16 [5]49M87.1
TST-14 [37]22M88.4
ViP-S/7 [10]CIFAR-10025M88.4
DynaMixer-S [35]26M88.6
Strip-MLP-T*(ours)18M89.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.594, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Table 5. The transfer learning results of the model pre-trained on ImageNet-1K and fine-tuned to CIFAR-10 and CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.901 + ], + "angle": 0, + "content": "number of parameters) of three types of models. Strip-MLP outperforms the CNNs-based model: \\(+0.5\\%\\) with half the number of parameters (25M vs. 39M) and FLOPs (3.7G vs. 8.0G) for Strip-MLP-T over RegNetY-8G [22]. Compared to Transformer-based models, e.g. Swin Transformer [19], Strip-MLP achieves comparable results but with obvious superiority in terms of the number of parameters and FLOPs, such as Strip-MLP-B achieves similar performance to Swin-B but with fewer parameters (57M vs. 88M) and FLOPs (9.2G vs. 15.4G). Compared to the MLP-based models, Strip-MLP still shows the same efficiency in the various models. The Strip-MLP-\\(\\mathrm{T}^*\\) achieves average higher accuracy by \\(+2.12\\%\\) than other MLP-based models (81.2% vs. 79.08%). Both Strip-MLP-B and Wave-MLP-B [27] models get the same accuracy of \\(83.6\\%\\). However, Strip-MLP-B uses fewer parameters (57M vs. 63M) and FLOPs (9.2G vs. 10.2G), indicating that the token-interaction power of our model is more efficient than other" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1500" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.088, + 0.473, + 0.577 + ], + "angle": 0, + "content": "
MethodParamsFLOPsThroughput (image/s)Top-1 (%)
CNNs-Based
ResNet50 [8]26M4.1G-78.5
ResNet101 [8]45M7.9G-79.8
RegNetY-4G [22]21M4.0G115780.0
RegNetY-8G [22]39M8.0G59281.7
RegNetY-16G [22]84M16.0G33582.9
Transformer-Based
DeiT-S [30]22M4.6G94079.8
DeiT-B [30]86M17.5G29281.8
Swin-T [19]29M4.5G75581.3
Swin-S [19]50M8.7G43783.0
Swin-B [19]88M15.4G27883.5
MLP-Based
ResMLP-S12 [29]15M3.0G141576.6
gMLP-S [16]20M4.5G-79.6
CycleMLP-B1 [1]15M2.1G103878.9
Hire-MLP-Tiny [7]18M2.1G156279.7
Wave-MLP-T [27]17M2.4G120880.6
Strip-MLP-T*(ours)18M2.5G81481.2
ResMLP-S24 [29]30M6.0G71579.4
Sparse-MLP-T [26]24M5.0G63981.9
CycleMLP-B2 [1]27M4.0G64181.6
Hire-MLP-Small [7]33M4.2G80882.1
Wave-MLP-S [27]30M4.5G72082.6
Strip-MLP-T(ours)25M3.7G59782.2
CycleMLP-B4 [1]52M10.1G32183.0
Sparse-MLP-S [26]49M10.3G36183.1
Wave-MLP-M [27]44M7.9G41383.4
Strip-MLP-S(ours)43M6.8G38183.3
ResMLP-B24 [29]116M23.0G23181.0
gMLP-B [16]73M15.8G-81.6
CycleMLP-B5 [1]76M12.3G24783.2
Sparse-MLP-B [26]66M14.0G27883.4
Hire-MLP-Base [7]58M8.1G44183.2
Wave-MLP-B [27]63M10.2G34183.6
Strip-MLP-B(ours)57M9.2G30083.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.577, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Table 6. Classification results of three kinds of models on ImageNet-1K without extra data. The group number of CGSMM is C/4. Throughput is tested on a single V100 GPU following [19]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.748 + ], + "angle": 0, + "content": "MLP-based models. In addition, we apply the Strip MLP layer into Wave-MLP-T (namely Wave-Strip-MLP-T) on ImageNet-1K, and we set the strip width as 3. The Wave-Strip-MLP-T gets better results than the original Wave-MLP-T model: \\(81.2\\%\\) vs. \\(80.6\\%\\). The performance improvement by \\(+0.6\\%\\) shows that our Strip MLP makes token interaction more efficient and can easily be used in other MLP models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.826 + ], + "angle": 0, + "content": "In conclusion, as shown in Fig. 4, our Strip-MLP has significant superiority in performance and complexity over other state-of-the-art methods on different datasets, demonstrating our Strip-MLP is remarkably efficient and effective to enrich the token interaction power for vision MLP." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.837, + 0.242, + 0.852 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To better show the effectiveness of the proposed method, we ablate the key components of the model design. Due to the limited GPU resources, we do ablation studies on the" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.205 + ], + "angle": 0, + "content": "
Strip WidthDatasetParamsFLOPsTop-1(%)
141.33M6.71G91.38
3Caltech-10143.66M6.83G92.09
545.99M6.96G91.72
748.33M7.08G91.86
140.71M6.71G86.09
3CIFAR-10041.88M6.83G86.50
543.04M6.96G86.39
744.21M7.08G86.35
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.205, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Table 7. Ablation study on the strip width of Strip MLP. We randomly select the model of Strip-MLP-S for Caltech-101 and Strip-MLP-S for CIFAR-100." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.246, + 0.892, + 0.381 + ], + "angle": 0, + "content": "
DatasetPatchesParamsFLOPsTop-1(%)
Caltech-101C/147.22M6.83G91.81
C/243.66M6.83G92.09
C/441.88M6.83G91.38
C/840.99M6.83G91.33
140.16M6.83G90.93
CIFAR-100C/147.22M6.83G86.20
C/243.66M6.83G86.18
C/441.88M6.83G86.50
C/840.99M6.83G85.98
140.16M6.83G85.03
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.381, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Table 8. Ablation study about the patch number of the CGSMM. The base model is Strip-MLP-S. \\(C/1\\) means splitting the feature \\(C\\) into C patches in channel dimension, namely \\(P = C/1 = C\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.432, + 0.804, + 0.446 + ], + "angle": 0, + "content": "datasets of both Caltech-101 and CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.628 + ], + "angle": 0, + "content": "The effect of Strip Width in Strip MLP. The strip width affects the token interaction ranges and determines the field of each row or column's contribution to adjacent tokens. We further conduct experiments to verify its effect by varying the width from 1 to 7 with a step at 2. Tab. 7 shows the performances of our models on Caltech-101 [6] and CIFAR-100 [14], respectively. When the width becomes larger, the performance increases and tends to saturate, indicating that Strip MLP layer improves the token interaction power. In all of our comparison experiments in previous sections, we consistently set the strip width as 3 according to the ablation experiment results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Effects of Patch Number in CGSMM. Different patch number brings varying degrees of token interaction power improvement and affects the model performance. In Tab. 8, we design five decreased patch numbers from C into 1 to show the effectiveness of CGSMM. Without the patch splitting \\((P = 1)\\) and GSML operation, the model performance decreases by \\(1.16\\%\\) on Caltech-101 \\((92.09\\% \\rightarrow 90.93\\%)\\) and decreases by \\(1.47\\%\\) on CIFAR-100 \\((86.50\\% \\rightarrow 85.03\\%)\\), which consistently proves the effectiveness of CGSMM in improving the token interaction power. From the experimental results, we can observe that the optimal patch number is different on the datasets with scale differences, so the optimal patch number should be determined through validation experiments. We consistently set the patch number as C/4 in other ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Cascade vs. Parallel Architecture of GSML. Applying the GSML in a cascade architecture enables the token interacts with other tokens across the entire 2D space in just one" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1501" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.098, + 0.289, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.225, + 0.195, + 0.235 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.097, + 0.482, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.224, + 0.387, + 0.235 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.097, + 0.678, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.224, + 0.582, + 0.235 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.097, + 0.875, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.224, + 0.784, + 0.235 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.251, + 0.291, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.376, + 0.195, + 0.387 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.251, + 0.487, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.376, + 0.389, + 0.387 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.248, + 0.682, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.376, + 0.586, + 0.387 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.248, + 0.874, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.376, + 0.783, + 0.387 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.391, + 0.89, + 0.419 + ], + "angle": 0, + "content": "Figure 4. Performance comparison on four datasets of ImageNet-1K [4], CIFAR-100 [14], CIFAR-10 [14], and Caltech-101 [6]. The Top-1 accuracy, number of parameters and FLOPs are reported with networks of CNN-based, Transformer-based, and MLP-based models." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.431, + 0.472, + 0.504 + ], + "angle": 0, + "content": "
MethodDatasetParamsFLOPsTop-1(%)
Parallel CascadeCaltech-10142.67M6.65G91.61
43.66M6.83G92.09
Parallel CascadeCIFAR-10040.90M6.66G85.89
41.88M6.83G86.50
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.504, + 0.47, + 0.56 + ], + "angle": 0, + "content": "Table 9. Ablation study on the cascade and parallel structure of the Strip MLP in Group Strip MLP. Parallel means applying the Strip MLP on the row and column direction in parallel. Cascade means the proposed CGSMM method in this paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.469, + 0.69 + ], + "angle": 0, + "content": "module, which needs two modules for the parallel structure so that may decrease the efficiency of token interaction. In Tab. 9, we test the effects between the cascade structure of CGSMM and the parallel structure of Parallel Group Strip Mixing Module (PGSMM). Our experiments show that the cascade structure gets a higher accuracy compared to the parallel structure, with an increase of \\(+0.48\\%\\) and \\(+0.61\\%\\) on the Caltech-101 and CIFAR-100 datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.469, + 0.828 + ], + "angle": 0, + "content": "Roles of CGSMM & LSMM. To verify the significance of two complementary branches, we conduct experiments where we only keep one of the branches of CGSMM or LSMM. Tab. 10 displays the outcomes of our trial on Caltech-101 and CIFAR-100. Notably, we found that removing any module from either of the two branches significantly reduced the model's performance. These results highlight the essential role of both CGSMM and LSMM in enriching the tokens' interaction power." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.195, + 0.86 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This paper presents an efficient and effective token interaction MLP model of Strip-MLP for vision MLP. For the" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.431, + 0.896, + 0.529 + ], + "angle": 0, + "content": "
MethodDatasetParamsFLOPsTop-1(%)
CGSMM OnlyCaltech-10143.40M7.55G91.10
LSMM Only43.96M6.00G90.31
CGSMM & LSMM41.88M6.83G91.38
CGSMM OnlyCIFAR-10043.40M7.55G85.20
LSMM Only43.96M6.00G86.13
CGSMM & LSMM41.88M6.83G86.50
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.529, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Table 10. Ablation study on the roles of two branches of CGSMM and LSMM. The base model is Strip-MLP-S. When any module branch has been removed, all input features would be fed into the remaining branch module." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.892, + 0.798 + ], + "angle": 0, + "content": "Token's interaction dilemma problem, we design the Strip Mixing Block to enrich the token interaction power by three strategies. Our experimental analysis reveals that the StripMLP remarkably improves the performances of the MLP-based models on small datasets and keeps comparable or even better results on ImageNet with great superiorities on the number of parameters and FLOPs. These observations clearly demonstrate that our models make token interaction more efficient among MLP-based models in image classification tasks. It is expected that Strip MLP layer has the potential to be a standard layer in the future and have promising performance on diverse vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement: This work is supported by National Key Research and Development Program of China (2021YFF1200800); Peng Cheng Laboratory Research Project (PCL2023AS6-1); and Guangdong Basic and Applied Basic Research Foundation under Grant 2022A1515110573." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1502" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Shoufa Chen, Enze Xie, Chongjian Ge, Ding Liang, and Ping Luo. Cycle.mp: A mlp-like architecture for dense prediction. In International Conference on Learning Representation (ICLR), Oral, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.17, + 0.47, + 0.224 + ], + "angle": 0, + "content": "[2] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.224, + 0.471, + 0.278 + ], + "angle": 0, + "content": "[3] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.471, + 0.415 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[6] Li Fei-Fei, Robert Fergus, and Pietro Perona. One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence, 28(4):594-611, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.455, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[7] Jianyuan Guo, Yehui Tang, Kai Han, Xinghao Chen, Han Wu, Chao Xu, Chang Xu, and Yunhe Wang. Hire-mlp: Vision mlp via hierarchical rearrangement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 826-836, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.577, + 0.471, + 0.603 + ], + "angle": 0, + "content": "[9] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.47, + 0.669 + ], + "angle": 0, + "content": "[10] Qibin Hou, Zihang Jiang, Li Yuan, Ming-Ming Cheng, Shuicheng Yan, and Jiashi Feng. Vision permutator: A permutable mlp-like architecture for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.47, + 0.739 + ], + "angle": 0, + "content": "[11] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.739, + 0.47, + 0.793 + ], + "angle": 0, + "content": "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.793, + 0.47, + 0.833 + ], + "angle": 0, + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.833, + 0.47, + 0.86 + ], + "angle": 0, + "content": "[14] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Hanxiao Liu, Zihang Dai, David So, and Quoc V Le. Pay attention to mlps. Advances in Neural Information Processing Systems, 34:9204-9215, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[17] Ruiyang Liu, Yinghui Li, Linmi Tao, Dun Liang, and Haitao Zheng. Are we ready for a new paradigm shift? a survey on visual deep mlp. Patterns, 3(7):100520, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.174, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[18] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.242, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[20] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[22] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[23] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[24] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[25] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[26] Chuanxin Tang, Yucheng Zhao, Guangting Wang, Chong Luo, Wenxuan Xie, and Wenjun Zeng. Sparse mlp for image recognition: is self-attention really necessary? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2344–2351, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.812 + ], + "angle": 0, + "content": "[27] Yehui Tang, Kai Han, Jianyuan Guo, Chang Xu, Yanxi Li, Chao Xu, and Yunhe Wang. An image patch is a wave: Phase-aware vision mlp. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10935-10944, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.812, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[28] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261-24272, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1503" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[29] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Gautier Izacard, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, et al. Resmlp: Feedforward networks for image classification with data-efficient training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.175, + 0.47, + 0.243 + ], + "angle": 0, + "content": "[30] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.243, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[31] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.311, + 0.47, + 0.366 + ], + "angle": 0, + "content": "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.366, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[33] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.47, + 0.529 + ], + "angle": 0, + "content": "[34] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.529, + 0.47, + 0.583 + ], + "angle": 0, + "content": "[35] Ziyu Wang, Wenhao Jiang, Yiming M Zhu, Li Yuan, Yibing Song, and Wei Liu. Dynamixer: a vision mlp architecture with dynamic mixing. In International Conference on Machine Learning, pages 22691-22701. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.583, + 0.47, + 0.638 + ], + "angle": 0, + "content": "[36] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.638, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[37] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[38] David Junhao Zhang, Kunchang Li, Yunpeng Chen, Yali Wang, Shashwat Chandra, Yu Qiao, Luoqi Liu, and Mike Zheng Shou. Morphmlp: A self-attention free, mlp-like backbone for image and video. arXiv preprint arXiv:2111.12527, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.47, + 0.853 + ], + "angle": 0, + "content": "[39] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.853, + 0.47, + 0.896 + ], + "angle": 0, + "content": "[40] Yipeng Zhang, Zahra M Aghajan, Matias Ison, Qiujing Lu, Hanlin Tang, Guldamla Kalender, Tonmoy Monsoor, Jie Zheng, Gabriel Kreiman, Vwani Roychowdhury, et al. De" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "coding of human identity by computer vision and neuronal vision. Scientific reports, 13(1):651, 2023." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1504" + } + ] +] \ No newline at end of file diff --git a/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/5c4682dc-11cf-43c9-b82f-059171805e73_origin.pdf b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/5c4682dc-11cf-43c9-b82f-059171805e73_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aeafa59fca6ad5f08b856ff135f605ee3aa573af --- /dev/null +++ b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/5c4682dc-11cf-43c9-b82f-059171805e73_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:507b3aa2380e657fe586ddb397f709b8f337029ad5441f5bd94188b9f7833b72 +size 1853646 diff --git a/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/full.md b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3e57de71e9df9e883c015972a794c3f3b09ec270 --- /dev/null +++ b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/full.md @@ -0,0 +1,331 @@ +# Strip-MLP: Efficient Token Interaction for Vision MLP + +Guiping Cao $^{1,2}$ Shengda Luo $^{1}$ Wenjian Huang $^{1}$ Xiangyuan Lan $^{2,*}$ Dongmei Jiang $^{2}$ Yaowei Wang $^{2}$ Jianguo Zhang $^{1,2,*}$ + +$^{1}$ Southern University of Science and Technology, Shenzhen, China $^{2}$ Peng Cheng Laboratory, Shenzhen, China + +12131099@mail.sustech.edu.cn, {luosd, huangwj, zhangjg}@sustech.edu.cn, {lanxy, jiangdm, wangyw}@pcl.ac.cn + +# Abstract + +Token interaction operation is one of the core modules in MLP-based models to exchange and aggregate information between different spatial locations. However, the power of token interaction on the spatial dimension is highly dependent on the spatial resolution of the feature maps, which limits the model's expressive ability, especially in deep layers where the feature are down-sampled to a small spatial size. To address this issue, we present a novel method called Strip-MLP to enrich the token interaction power in three ways. Firstly, we introduce a new MLP paradigm called Strip MLP layer that allows the token to interact with other tokens in a cross-strip manner, enabling the tokens in a row (or column) to contribute to the information aggregations in adjacent but different strips of rows (or columns). Secondly, a Cascade Group Strip Mixing Module (CGSMM) is proposed to overcome the performance degradation caused by small spatial feature size. The module allows tokens to interact more effectively in the manners of within-patch and cross-patch, which is independent to the feature spatial size. Finally, based on the Strip MLP layer, we propose a novel Local Strip Mixing Module (LSMM) to boost the token interaction power in the local region. Extensive experiments demonstrate that Strip-MLP significantly improves the performance of MLP-based models on small datasets and obtains comparable or even better results on ImageNet. In particular, Strip-MLP models achieve higher average Top-1 accuracy than existing MLP-based models by $+2.44\%$ on Caltech-101 and $+2.16\%$ on CIFAR-100. The source codes will be available at https://github.com/MedProcess/Strip_MLP. + +# 1. Introduction + +In computer vision, Convolutional Neural Networks (CNNs) are one of the most popular network backbones, which have made a series of breakthroughs [8]. Inspired by + +![](images/4491c943a6c738d7f2000e08d7a09d1c31030a5bd3e4ddb4b67c136e9d374b3a.jpg) +Figure 1. Token interaction of the MLP layer on the down-sampled image feature. (a) The process of which the feature resolution will be down-sampled from $H \times W \times C$ into a small spatial size of $\frac{H}{8} \times \frac{W}{8} \times 8C$ . (b) Token interaction on different feature resolutions. + +the great success of self-attention-based architectures [32] in natural language processing (NLP), the Transformer models [5, 19, 34] are introduced into the field of computer vision, and have achieved comparable results with state-of-the-art (SOTA) CNNs. Although ViT [5] and its variants outperform traditional CNNs, the models introduce high computational complexity to construct attention maps. Recently, some studies [20, 26] in the vision community suggest that the attention mechanism is not necessary and simpler model architectures are proposed. + +MLP-based models, like MLP-Mixer [28], gMLP [16] and ViP [10] process the data with the Multilayer Perceptrons (MLP), showing great potential to improve the performance of vision models [29]. As the first visual deep MLP network, MLP-Mixer [28] introduces two types of MLP layers: Channel-Mixing MLPs (CMM) and Token-Mixing MLPs (TMM). For CMM, the module mainly mixes the information between different channels of each token. For TMM, it allows each spatial token to interact with all other tokens (the whole image) in a single MLP layer. However, this design also introduces a much larger number of parameters and higher computational complexity prone to overfitting. To address this problem, Sparse MLP (SMLP) [26] and Vision Permutator (ViP) [10] propose a similar layer of parallel structure, which applies one dimension MLP along the axial directions, and parameters are shared among rows or columns, respectively. Therefore, it reduces the number + +of model parameters and computational complexity, avoiding the common over-fitting problem. + +Although SMLP and ViP alleviate some deficiencies of MLP-Mixer [28], both methods bring the challenge that the token interaction power is highly dependent on the feature spatial size when interacting tokens on spatial rows (or columns). As shown in Fig. 1, the spatial feature resolution is down-sampled to a small size but with more channels, which means the feature pattern of each token is mainly concentrated on the channel dimension rather than the spatial one. Interacting tokens along the spatial dimension by sharing the weights among all channels would seriously ignore the feature pattern differences among different channels, which may degrade the token interaction power, especially in deep layers with small spatial feature resolution. Here, we mark this problem as the Token's interaction dilemma. Taking SMLP as an example, we analyze the feature resolution and complexity of the model in different stages in detail (seen in Sec. 3.4). We find that as the spatial feature size decreases by down-sampling stage by stage, the token interaction layer also becomes smaller and smaller, which makes the token interaction power degraded rapidly. + +To address these challenges, we propose a new efficient Strip MLP model, dubbed Strip-MLP, to enrich the power of the token interaction layer in three ways. For the level of a single MLP layer, inspired by the cross-block normalization schemes of HOG [3] and the sparse connections between the biological neurons, we design a Strip MLP layer to allow the token to interact with other tokens in a crossstrip manner, enabling each row or column of the tokens to contribute differently to other rows or columns. For the token interaction module level, we develop channel-wise group mixing of CGSMM, enabling the tokens in a row (or column) to contribute to the information aggregations in adjacent but different strips of rows (or columns). to tackle the problem that the token interaction power decreases in deep layers with the spatial feature size significantly reduced but with multiplying channels. Considering the existing methods [10, 26, 28] interact the tokens mainly in the long range of row (or column), which may not aggregate tokens well in the local region, we propose a new Local Strip Mixing Module (LSMM) with a small Strip MLP unit to strengthen the token interaction power on local interactions. + +The proposed Strip-MLP model significantly boosts the token interaction power, and the main contributions are: + +- A new MLP paradigm for vision MLP: Strip MLP layer, which aggregates the adjacent tokens in a cross strip manner and enables each row or column of the tokens to contribute differently to other rows or columns, interacting tokens more efficiently. +- Designing a Cascade Group Strip Mixing module and a Local Strip Mixing module, which effectively improves the model's token interaction power and boosts + +the tokens aggregation in the local region, respectively; + +- Extensive experiments show that Strip-MLP remarkably improves the performances of the MLP-based models. Strip-MLP achieves higher average Top-1 accuracy by $+2.44\%$ on Caltech-101 and $+2.16\%$ on CIFAR-100 over the existing MLP-based models. In addition, our models achieve comparable or even better performances on ImageNet-1K compared with traditional MLP-based models, and other popular CNNs and transformer-based models. + +# 2. Related Work + +Deep neural networks for vision recognition can be mainly divided into three categories: Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Multi-Layer Perceptron-based models (MLPs) [10]. + +CNNs-Based Models. CNNs are the de-facto standard deep learning network models for vision tasks and have been intensely studied in the vision community. AlexNet [15] is a symbolically significant model that won the ILSVRC 2012 contest with far higher performance than others. Since then, CNNs-based models have attracted more attentions, and lots of deeper and more effective architectures [8, 11, 23-25, 33, 39, 40] have been proposed. With the convolution and pooling layer, CNNs aggregate the feature in a local region but not well in the long-term dependencies, which are optimized by the new vision model like Transformer [32] and MLP [28] models. + +Transformer-Based Models. Transformer [32] is introduced for machine translation and becomes the reference model for all-natural language processing (NLP) tasks. Inspired by the great success of Transformer in NLP, ViTs [5] first applies a standard Transformer to images, which attains excellent results compared to the SOTA CNNs model. DeiT [30] introduces several training strategies and distillation methods to make ViTs more effective on the smaller ImageNet-1K dataset. By proposing a hierarchical Transformer with shifted windows, Swin Transformer [19] achieves SOTA accuracy on ImageNet-1K, bringing greater efficiency by self-attention in local window and cross-window connection. For these models, self-attention is the core module but with heavy computational burdens to obtain an attention map. + +MLP-Based Models. Without the convolutions and self-attention mechanism, MLP-Mixer [28] builds the architecture that only uses the MLP layer and achieves competitive performance on image classification benchmarks. Since then, the researchers have developed many MLP-like variants [1, 7, 16, 26, 27, 29, 31, 35, 38] models. The work [17] gives a comprehensive survey about visual deep MLP models and compares the intrinsic connections and differences between convolution, self-attention mechanism, and Token-mixing MLP in detail. Sparse MLP [26] introduced a sparse + +![](images/c5a7848b3d15c4f0fd4fee5d97f45e64feb6afbe1e482af05689489606d22fe6.jpg) +Figure 2. The overall and components architecture of Strip-MLP. (a) Strip-MLP has four stages, and $T1$ to $T4$ means the repeated times of the block in each stage. (b) The Strip Mixing Block splits the head (channel dimension) into two heads, and they are fed into the two parallel branches. (c) Channel Mixing Block architecture. + +operation to separately aggregate information along axial directions, avoiding the quadratic computational complexity of conventional MLP. Hire-MLP [7] presents a novel variant of MLP-based architecture via hierarchically rearranging tokens to aggregate local and global spatial information. Wave-MLP [27] represents each token as a wave function with two parts of amplitude and phase, which has the ability to model varying contents from different input images. Above all, there is still the limitation of token interaction power degraded significantly especially when the spatial feature resolution becomes small, which has been overlooked by previous studies. In this paper, we aim to enrich the token interaction power on both the single MLP layer and token interaction module to advance the performance of MLP-based models. + +# 3. Method + +In this section, we first present the overall architecture of Strip-MLP. Then, we show the key proposed components: Strip MLP Layer, Cascade Group Strip Mixing Module (CGSMM) and Local Strip Mixing Module (LSMM) of the model in detail, and present the comparison analysis between the Strip MLP and traditional MLP model in terms of parameters and complexity. Finally, we define four kinds of architecture variants to compare the performance of the models in different sizes. + +# 3.1. Overall Architecture + +An overview of the Strip-MLP model is depicted in Fig. 2 (a). We design the Strip-MLP as a hierarchical structure of a multi-stages model. Given the input image $I$ , Strip-MLP model makes patch embedding and models the image feature in four stages. + +Patch Embedding. Strip-MLP firstly splits the input + +image $I \in \mathbb{R}^{H \times W \times 3}$ (H, W: image height and width) into a sequence of image patches (also referred as tokens) $I_{p} \in \mathbb{R}^{hw \times c}$ (c: the number of channels), where the patch size is $p \times p$ and patch number is $hw\left(h = \frac{H}{p}, w = \frac{W}{p}\right)$ , and then all patches are linearly projected into a desired higher dimension $C$ of the feature $(X \in \mathbb{R}^{hw \times C})$ . + +Mixing Block. The purpose of mixing block is to boost the interactions between the features of diverse spatial locations and channels. To effectively aggregate spatial and channel information, we design two sub-blocks of the Strip Mixing Block and Channel Mixing Block. The Strip Mixing Block is comprised of CGSMM and LSMM, which blend and aggregate spatial information more efficiently at the global and local level, respectively. Patch merging module aims to merge the feature in which the spatial dimension will be reduced by $2 \times 2$ , and the channel dimension increases by 2 times so that the model down-samples the feature from $\frac{H}{p} \times \frac{W}{p}$ into $\frac{H}{8p} \times \frac{W}{8p}$ stage by stage. To obtain multi-scale features, we apply a single convolutional layer to the output features of stages 1 and 2. Then, we add the resulting features to the input features of stages 3 and 4. + +Head Layer. The features extracted by the multiple blocks are fed into a global average pooling (GAP) layer to reduce the feature dimension. Finally, the feature will be fed into a fully-connected head layer for classification. + +# 3.2. Strip Mixing Block and Channel Mixing Block + +Strip Mixing Block. To improve the token interaction power, we design the block to aggregate both long and short range interactions in a parallel manner. As illustrated in Fig. 2 (b), we split the feature in channel dimension, with one-half of the channel feature being fed into CGSMM to model the long-range interactions, and the remaining feature being fed into LSMM to aggregate the local interac + +![](images/4ab71b6bb1c028c6a1606c9715c1f4989d2f035242aa57032a5b61195168bdef.jpg) +(a) Cascade Group Strip Mixing Module + +![](images/6400613916a9863fc20833d1ff07defb126b953838f96da8bed118121060456d.jpg) +(b) Local Strip Mixing Module +Figure 3. (a) The architecture of CGSMM. It contains the Strip MLP layer, the feature splitting, restoring, and permuting operation in a cascade mode. (b) The structure of the Local Strip Mixing module. The dark green block composes of a re-weight module. + +tions. Given the input feature $X\in \mathbb{R}^{H\times W\times C}$ , the block can be formulated as: + +$$ +X _ {m} = M L P \left(\left(D W S C (X)\right)\right) \tag {1} +$$ + +$$ +Y = F C \left(C a t \left(C G S M M \left(X _ {m _ {1}} ^ {\frac {C}{2}}\right), L S M M \left(X _ {m _ {2}} ^ {\frac {C}{2}}\right)\right)\right) + X \tag {2} +$$ + +where $X_{m}$ and $Y$ are the intermediate and output feature of the block. $X_{m}$ is split into $X_{m_1}$ and $X_{m_2}$ with half channel, respectively. DWSC means depth-wise convolution [2], and the kernel size is $3\times 3$ . MLP is a serial connection of fullyconnection (FC), batch normalization [12] and GELU [9] activation layers. $Cat(*)$ denotes concatenation operation. + +Channel Mixing Block. The block aims to aggregate the information between channels of the token, and the basic structure is shown in Fig. 2 (c). Following the inverted bottleneck in [20] and global response normalization (GRN) in [36], we design the Channel Mixing block to increase the contrast and selectivity of channels. + +# 3.3. CGSMM and LSMM + +We first introduce the Strip MLP Layer. Based on this layer, we design CGSMM using a simple but effective strategy that splits the feature into patches along channel dimension and interacts tokens more effectively in the manners of within-patch and cross-patch, which is independent to the feature spatial size. In addition, the design of existing methods [10, 26, 28] only allows tokens to interact in a long range of rows (or columns) with sharing weight; thus the resulting model may struggle with effectively aggregating both global and local information simultaneously. Therefore, we design LSMM to capture the local interactions more efficiently. + +Strip MLP Layer. In MLP-based models, the majority of MLP layers treat each row and column of the data in isolation, as formulated in Eq. (3), which may lead to the inefficiency of token interaction. Inspired by the cross-block normalization schemes of HOG [3], which overlaps the blocks and enables each scalar cell response to contribute into different blocks, we propose the Strip MLP layer. The proposed layer applies MLP on the "strip" data of adjacent + +rows or columns in the spatial direction in order to aggregate the feature in a cross-strip manner. Given the input $X$ , we formulated the Strip MLP layer (setting strip width as 3 for example) in Eqs. (4) and (5): + +$$ +X _ {*, j} ^ {h} = W _ {1} X _ {*, j}; \quad X _ {i, *} ^ {w} = W _ {2} X _ {i, *} \tag {3} +$$ + +$$ +X _ {*, j} ^ {h} = W _ {3} \operatorname {C a t} \left(X _ {*, j - 1}, X _ {*, j}, X _ {*, j + 1}\right) \tag {4} +$$ + +$$ +X _ {i, *} ^ {w} = W _ {4} \operatorname {C a t} \left(X _ {i - 1, *}, X _ {i, *}, X _ {i + 1, *}\right) \tag {5} +$$ + +where $W_{1}\sim W_{4}$ are the weights of the MLP layer, $i$ and $j$ are the token index in row and column. The superiorities of the Strip MLP layer are mainly in two aspects: on one hand, Strip MLP layer enables the token to interact with other tokens in both short and long spatial ranges simultaneously. On the other hand, similar to the HOG cross-block normalization process, each row (or column) not only serves for the current row (or column) tokens aggregation but also makes contribution to the adjacent rows (or columns) feature aggregation. For example, in Eq. (4), $X_{*,j}$ makes different contributions for aggregating the processed feature of $X_{*,j - 1}^{h},X_{*,j}^{h}$ and $X_{*,j + 1}^{h}$ , making tokens interacting more efficiently in a cross-strip manner. + +Cascade Group Strip Mixing Module. The architecture is shown in Fig. 3 (a). The module applies the Strip MLP layer in row and column direction in a cascade mode. As the operation of the Strip MLP layer on row is similar to that on column, we take one of them to show the method. + +Patch Splitting of the Data along Channel Dimension. Given the input feature $X_{m_1} \in \mathbb{R}^{H \times W \times \frac{C}{2}}$ , the module firstly permutes the feature into $X_{m_1}^{H \times \frac{C}{2} \times W}$ , and splits the feature into $P$ patches in the channel dimension and concatenated along the row $(X^{H \times \frac{C}{2P} \times PW})$ dimension. + +Group Strip MLP Layer. To improve the token interaction power, we propose a Group Strip MLP Layer (GSML) to interact tokens in the manners of within-patch and cross-patch. In particular, we apply the unshared weights of Strip MLP on different patches, and the tokens within the same patch share the weights for interaction. Then, we restore the feature to original shape and concatenate it with the input + +feature. To interact the tokens crossing patches, we apply a channel fully-connected (Channel FC) layer to interact tokens between different patches. + +Local Strip Mixing Module. The module is illustrated in Fig. 3 (b). In order to better aggregate the local interactions on spatial dimension, we define a small Strip MLP unit, where the strip width and length are 3 and 7, respectively. Given the input feature $X_{m_2}^{H \times W \times \frac{C}{2}}$ , we aggregate the local interactions in rows and columns direction simultaneously. Following [1, 10, 27], we sum all branches with a re-weight module. + +# 3.4. Parameter and Complexity Analysis + +In multi-stage processing architecture, the Token's interaction dilemma becomes more serious, as the feature spatial resolution will be down-sampled in deep layers, and the token interaction power will be weakened. In this section, we present the comparison analysis of parameters and complexity to show the effectiveness of the proposed CGSMM. Considering the majority MLP-Based models have the similar model structure of token interaction, we analyze the parameters and complexities with Sparse MLP [26], which is a popular model with good performance on various datasets. + +Sparse MLP [26] firstly applies MLP on the columns and rows of $X$ to map $\mathbb{R}^{CW\times H}$ to $\mathbb{R}^{CW\times H}$ separately. Then, the model concatenates the two processed features with the input feature $X$ and uses a linear layer on the channel dimension to fuse the feature from $\mathbb{R}^{HW\times 3C}$ to $\mathbb{R}^{HW\times C}$ . The number of parameters and FLOPs of the first step are $W^2 +H^2$ and $CHW(H + W)$ ; they are $3C^2$ and $3HWC^2$ for the fusion step. Comparatively, the number of parameters and FLOPs of Strip MLP layer for token interaction are $3P(H^{2} + W^{2})$ and $3CHW(H + W)$ , respectively. The fusion step has $4C^2$ parameters and $4HWC^2$ FLOPs. In particular, for $H_{1} = W_{1} = 56$ , $H_{4} = W_{4} = 7$ , $C_1 = 112$ , $C_4 = 112\times 2^3 = 896$ , and $P = \frac{C}{4}$ , we calculate the number of parameters and FLOPs of both two layers. As shown in Tab. 1, for the Sparse MLP block, the number of parameters for token interaction in stage 4 (only $0.10k$ ) decreases by 62.70 times than in stage 1. In addition, most of the number of the block's parameters and the FLOPs concentrate in the fusion step. For example, only $0.01\%$ parameters ( $0.52\%$ FLOPs) are used for the token interaction step in stage 4. Based on the above analysis, to improve the token interaction power, it would be better to redesign the block to balance the number of parameters and FLOPs. + +In CGSMM, we use a simple but effective strategy that splits the feature along channel dimension into patches and interacts the tokens in the manners of within-patch and cross-patch. No matter how the spatial resolution decreases, the module can still interact the tokens with channel-wise specificity in different patches. In Tab. 1, we improve the Proportion of the number of parameters in the token inter + +
ItemsSparse MLPStrip MLP
ParamsFLOPsParamsFLOPs
Stage 16.27k39.34M526.85k118.01M
Stage 40.10k0.62M65.86k1.84M
Changes↓62.70↓63.45↓8.00↓64.14
Stage 4 Fusion2.41M118.01M3.21M157.35M
Proportion0.01%0.52%2.01%1.16%
+ +Table 1. The model parameters and complexity comparison between Sparse MLP and Strip MLP layer in different stages. Stage 1 and Stage 4 show the parameters (Params) and floating-point operations (FLOPs) for token interaction in one layer between different stages. Stage 4 Fusion indicates the Params and FLOPs in the feature fusion step of the token interaction module. Proportion means the ratio between the token interaction step and other steps in the token interaction module of stage 4. + +action layer from $0.01\%$ to $2.01\%$ in stage 4. Although our GSML brings more parameters and computations when just considering one token interaction layer, the total number of parameters and overall computational complexity of the model are decreased as the token interaction power is improved and we set small model configurations of $T_{1} \sim T_{4}$ . Additionally, the experiments in Sec. 4 show that our models achieve better performance with fewer parameters and FLOPs than other SOTA models. + +# 3.5. Architecture Variants + +We developed four variants of our Strip-MLP network: Strip-MLP-T\* (light tiny), Strip-MLP-T (tiny), Strip-MLPSmall), Strip-MLP-B (base), which have similar or even smaller model sizes, compared with MLP-baed models [7, 26, 27] and Swin Transformer [18], respectively. The hyperparameters of the four variants of the models are as follows: + +- Strip-MLP-T*: $C = 80, \{T_1 \sim T_4\} = \{2, 2, 6, 2\}$ ; +- Strip-MLP-T: $C = 80, \{T_1 \sim T_4\} = \{2, 2, 12, 2\}$ ; +- Strip-MLP-S: $C = 96, \{T_1 \sim T_4\} = \{2, 2, 18, 2\}$ ; +- Strip-MLP-B: $C = 112, \{T_1 \sim T_4\} = \{2, 2, 18, 2\}$ ; + +where $C$ means the channel number of the hidden layers in the first stage, and $T_{1} \sim T_{4}$ is the number of times repeated for Strip Mixing Block and Channel Mixing Block in each of the four stages, as shown in Fig. 2. + +# 4. Experiments + +We compare our Strip-MLP with three main categories of networks: CNNs-based, Transformer-based, and MLP-based models. We first show the experimental settings of the datasets and implementation details. Then, the proposed method is compared with other popular methods on three datasets. Finally, we ablate the critical design components of Strip-MLP network. + +# 4.1. Experimental Settings + +Datasets. Methods are evaluated in four datasets which vary significantly in the number of training images: + +
MethodParamsFLOPsTop-1(%)
CNNs-Based
ResNet50 [8]23.71M4.12G89.61
ResNet101 [8]42.71M7.85G88.56
ResNet152 [8]58.35M11.58G89.04
Transformer-Based
ViT-B/16 [5]85.85M16.86G53.96
Swin-T [19]27.60M4.36G80.40
Swin-S [19]48.91M8.52G79.83
Swin-B [19]86.85M15.14G78.42
MLP-Based
Wave-MLP-T [27]16.73M2.48G86.89
Hire-MLP-T [7]17.58M2.16G87.85
Strip-MLP-T*(ours)17.67M2.53G90.11
Wave-MLP-S [27]30.25M4.55G87.43
Hire-MLP-S [7]32.65M4.26G88.48
Sparse-MLP-T [26]23.55M5.02G90.54
Strip-MLP-T(ours)24.33M3.67G90.93
Wave-MLP-M [27]43.59M7.93G88.45
Hire-MLP-B [7]57.77M8.17G88.62
Sparse-MLP-S [26]47.87M10.36G91.07
Strip-MLP-S(ours)43.66M6.83G92.09
Wave-MLP-B [27]62.90M10.27G88.81
Hire-MLP-L [7]95.03M13.50G88.19
Sparse-MLP-B [26]65.07M14.04G91.67
Strip-MLP-B(ours)58.49M9.22G92.26
+ +Table 2. Image classification results of our Strip-MLP and other models on Caltech-101. Top-1 denotes the Top-1 accuracy. Here, the patch number of CGSMM is C/2. + +- Caltech-101 [6]: it is a benchmark dataset that includes 101 classes, totally around $9k$ images. We randomly take $80\%$ of each class of the data as the training set and use the remaining data as the testing set. +- CIFAR-10 [14]: it consists of $60k \times 32$ images in 10 classes, and there are $50k$ images +- CIFAR-100 [14]: it has 100 classes containing 600 images of each class (totally $60k$ ), and there are 500 training images and 100 testing images per class. +- ImageNet-1K [4]: it is an image classification benchmark containing $1.28M$ training images and $50k$ testing images of 1000 classes. + +Training Details. In all of our experiments, the input image size is $224 \times 224$ , and the input patch size is $4 \times 4$ . Based on the deep-learning framework of PyTorch [21], we train our models from scratch without any extra data for pre-training except for the transfer learning experiments. We employ an AdamW [13] optimizer with 300 epochs for all training datasets and apply a cosine decay learning rate scheduler and 30 epochs of linear warm-up. In addition, we adopt the same augmentation and regularization strategies as to Swin-Transformer [19]. + +# 4.2. Image Classification on Small Datasets + +We conduct experiments on Small Datasets of Caltech-101, CIFAR-10 and CIFAR-100. The results consistently indicate that our Strip-MLP efficiently improves the token + +interaction power, and advances the MLP-based models' performance on the small dataset than other models. + +Results on Caltech-101. Tab. 2 presents the image classification result of three types models on Caltech-101 dataset. Compared to ResNet [8], all the four variants of our Strip-MLP models achieve higher Top-1 accuracy with fewer parameters and FLOPs, demonstrating an average increase of $+2.69\%$ over three widely-used ResNet models. + +For the Transformer-based models, e.g. Swin Transformer [19], Strip-MLP models noticeably surpass all variants models: $+10.53\% / 12.26\% / 13.84\%$ over SwinTransformer (T/S/B) models. ViT-B/16 [5] only gets an accuracy of $53.96\%$ , indicating that transformer-based models depend more on the scale of the datasets. + +When compared to the MLP-based models, like Sparse MLP [26], Wave-MLP [27] and Hire-MLP [7], our models still achieve the best accuracy with fewer model parameters and FLOPs. For example, with fewer FLOPs $(-4.28\mathrm{G})$ Strip-MLP-B $(92.26\%)$ has an increase of $+3.64\%$ Top-1 accuracy than Hire-MLP-B $(88.62\%)$ . + +Results on CIFAR-10. Tab. 3 shows the classification results of performance comparisons. Compared with CNNs-based models, all four variants of our Strip-MLP models achieve higher Top-1 accuracy with fewer parameters and FLOPs, with an average increase of $+1.14\%$ (98.06% vs. 96.92%). When compared with transformer-based models, our Strip-MLP achieves better performance by $+3.06\%$ (98.06% vs. 95.00%) with fewer parameters (34.49M vs. 54.37M) and FLOPs (5.56G vs. 9.34G). Compared with MLP-based models, all of our Strip-MLP models obtain the best Top-1 accuracy with fewer parameters and FLOPs. For example, the average Top-1 accuracy of Strip-MLP is higher than Hire-MLP/Wave-MLP/Sparse-MLP by $+1.45\% / 0.69\% / 0.19\%$ (98.06% vs. 96.61%/97.37%/97.87%) with fewer average parameters (34.49M vs. 50.71M/38.32M/45.43M) and fewer average FLOPs (5.56G vs. 9.36G/6.31G/9.81G) + +In particular, with only 16.76M parameters and 2.53G FLOPs, our Strip-MLP- $T^{*}$ noticeably surpasses all variants of ResNet [8] and Swin-Transformer [19] models, and Strip-MLP- $T^{*}$ outperforms the transfer learning models of ViT-S/16 [5]/TST-14 [37] by $+0.69\% / +0.29\%$ , consistently demonstrating our Strip-MLP has significant superiorities in performance and complexity. + +Results on CIFAR-100. Tab. 4 compares the image classification results on CIFAR-100 with three kinds of architecture models. Compared with CNNs-based models [8], Strip-MLP obtains average higher accuracy of $+2.05\% \sim +2.43\%$ with a smaller number of parameters and FLOPs. When compared to the self-attention based model, e.g. Swin Transformer [19], our model shows great superiority that Strip-MLP-B achieves $86.40\%$ Top-1 accuracy, $+9.64\%$ than the Swin-B model with fewer pa + +
MethodParamsFLOPsTop-1 (%)
CNNs-Based
ResNet50 [8]23.53M4.12G97.06
ResNet101 [8]42.52M7.85G96.70
ResNet152 [8]58.16M11.58G96.99
Transformer-Based
Swin-T [19]27.53M4.36G95.25
Swin-S [19]48.84M8.52G94.87
Swin-B [19]86.75M15.14G94.87
MLP-Based
Wave-MLP-T [27]16.69M2.48G96.91
Hire-MLP-T [7]17.53M2.16G96.20
Strip-MLP-T*(ours)16.76M2.53G97.79
Wave-MLP-S [27]30.20M4.55G97.27
Hire-MLP-S [7]32.61M4.26G96.70
Sparse-MLP-T [26]23.49M5.02G97.67
Strip-MLP-T(ours)23.07M3.67G97.94
Wave-MLP-M [27]43.55M7.93G97.56
Hire-MLP-B [7]57.72M8.17G96.88
Sparse-MLP-S [26]47.80M10.36G97.99
Strip-MLP-S(ours)41.81M6.83G98.23
Wave-MLP-B [27]62.83M10.27G97.72
Hire-MLP-L [7]94.96M13.50G96.64
Sparse-MLP-B [26]64.99M14.04G97.95
Strip-MLP-B(ours)56.33M9.22G98.28
+ +Table 3. Image classification results of different models on CIFAR-10. The patch number of CGSMM is C/4. All models are trained from scratch without any extra data. + +rameters (58.49M vs. 86.85M) and lower FLOPs (9.22G vs. 15.14G). In addition, our models keep the superiority over the other MLP-based models in that Strip-MLP achieves the best Top-1 accuracy (with a higher range of $+1.04\% \sim +3.01\%$ ) compared to the MLP-based models, e.g. Wave-MLP [27], Hire-MLP [7], and Sparse MLP [26], with fewer parameters and FLOPs. + +# 4.3. Transfer Learning on Small Datasets + +Following the transfer learning experiment of DynaMixer [35], we show the advantages of our models on the transfer learning tasks. With the pre-trained Strip-MLP $\mathrm{T}^*$ model on the ImageNet-1K, we transfer the model to downstream datasets of CIFAR-10 and CIFAR-100. Compared to the previous models, such as ViT [5], ViP [10] and DynaMixer [35] models, our Strip-MLP $\mathrm{T}^*$ models achieve the best accuracies of $98.8\%$ and $89.4\%$ on two datasets, as shown in Tab. 5, $+1.7\% / 0.6\%$ than ViT-S/16 and DynaMixer-S on CIFAR-10, $+2.3\% / 0.8\%$ over ViT-S/16 and DynaMixer-S on CIFAR-100. More importantly, the number of parameters of our model is only 18M, which is far less than other models, indicating that our model is more efficient and effective. + +# 4.4. Image Classification on ImageNet-1K + +We conduct experiments on ImageNet-1K to show the effectiveness of the method. Tab. 6 shows the performance together with the complexity (in terms of FLOPS and the + +
MethodParamsFLOPsTop-1(%)
CNNs-Based
ResNet50 [8]23.71M4.12G83.03
ResNet101 [8]42.71M7.85G83.75
ResNet152 [8]58.35M11.58G84.35
Transformer-Based
Swin-T [19]27.60M4.36G78.56
Swin-S [19]48.91M8.52G78.10
Swin-B [19]86.85M15.14G76.76
MLP-Based
Wave-MLP-T [27]16.73M2.48G82.77
Hire-MLP-T [7]17.58M2.16G82.44
Strip-MLP-T*(ours)17.67M2.53G85.10
Wave-MLP-S [27]30.25M4.55G83.34
Hire-MLP-S [7]32.65M4.26G82.94
Sparse-MLP-T [26]23.55M5.02G84.19
Strip-MLP-T(ours)24.33M3.67G85.23
Wave-MLP-M [27]43.59M7.93G83.95
Hire-MLP-B [7]57.77M8.17G83.17
Sparse-MLP-S [26]47.87M10.36G84.27
Strip-MLP-S(ours)43.66M6.83G86.18
Wave-MLP-B [27]62.89M10.27G84.23
Hire-MLP-L [7]95.03M13.50G83.53
Sparse-MLP-B [26]65.07M14.04G84.46
Strip-MLP-B(ours)58.49M9.22G86.40
+ +Table 4. Image classification results of different models on CIFAR-100. The patch number of CGSMM is C/2. + +
ModelDatasetParamsTop-1(%)
ViT-S/16 [5]49M97.1
TST-14 [37]22M97.5
ViP-S/7 [10]CIFAR-1025M98.0
DynaMixer-S [35]26M98.2
Strip-MLP-T*(ours)18M98.8
ViT-S/16 [5]49M87.1
TST-14 [37]22M88.4
ViP-S/7 [10]CIFAR-10025M88.4
DynaMixer-S [35]26M88.6
Strip-MLP-T*(ours)18M89.4
+ +Table 5. The transfer learning results of the model pre-trained on ImageNet-1K and fine-tuned to CIFAR-10 and CIFAR-100. + +number of parameters) of three types of models. Strip-MLP outperforms the CNNs-based model: $+0.5\%$ with half the number of parameters (25M vs. 39M) and FLOPs (3.7G vs. 8.0G) for Strip-MLP-T over RegNetY-8G [22]. Compared to Transformer-based models, e.g. Swin Transformer [19], Strip-MLP achieves comparable results but with obvious superiority in terms of the number of parameters and FLOPs, such as Strip-MLP-B achieves similar performance to Swin-B but with fewer parameters (57M vs. 88M) and FLOPs (9.2G vs. 15.4G). Compared to the MLP-based models, Strip-MLP still shows the same efficiency in the various models. The Strip-MLP- $\mathrm{T}^*$ achieves average higher accuracy by $+2.12\%$ than other MLP-based models (81.2% vs. 79.08%). Both Strip-MLP-B and Wave-MLP-B [27] models get the same accuracy of $83.6\%$ . However, Strip-MLP-B uses fewer parameters (57M vs. 63M) and FLOPs (9.2G vs. 10.2G), indicating that the token-interaction power of our model is more efficient than other + +
MethodParamsFLOPsThroughput (image/s)Top-1 (%)
CNNs-Based
ResNet50 [8]26M4.1G-78.5
ResNet101 [8]45M7.9G-79.8
RegNetY-4G [22]21M4.0G115780.0
RegNetY-8G [22]39M8.0G59281.7
RegNetY-16G [22]84M16.0G33582.9
Transformer-Based
DeiT-S [30]22M4.6G94079.8
DeiT-B [30]86M17.5G29281.8
Swin-T [19]29M4.5G75581.3
Swin-S [19]50M8.7G43783.0
Swin-B [19]88M15.4G27883.5
MLP-Based
ResMLP-S12 [29]15M3.0G141576.6
gMLP-S [16]20M4.5G-79.6
CycleMLP-B1 [1]15M2.1G103878.9
Hire-MLP-Tiny [7]18M2.1G156279.7
Wave-MLP-T [27]17M2.4G120880.6
Strip-MLP-T*(ours)18M2.5G81481.2
ResMLP-S24 [29]30M6.0G71579.4
Sparse-MLP-T [26]24M5.0G63981.9
CycleMLP-B2 [1]27M4.0G64181.6
Hire-MLP-Small [7]33M4.2G80882.1
Wave-MLP-S [27]30M4.5G72082.6
Strip-MLP-T(ours)25M3.7G59782.2
CycleMLP-B4 [1]52M10.1G32183.0
Sparse-MLP-S [26]49M10.3G36183.1
Wave-MLP-M [27]44M7.9G41383.4
Strip-MLP-S(ours)43M6.8G38183.3
ResMLP-B24 [29]116M23.0G23181.0
gMLP-B [16]73M15.8G-81.6
CycleMLP-B5 [1]76M12.3G24783.2
Sparse-MLP-B [26]66M14.0G27883.4
Hire-MLP-Base [7]58M8.1G44183.2
Wave-MLP-B [27]63M10.2G34183.6
Strip-MLP-B(ours)57M9.2G30083.6
+ +Table 6. Classification results of three kinds of models on ImageNet-1K without extra data. The group number of CGSMM is C/4. Throughput is tested on a single V100 GPU following [19]. + +MLP-based models. In addition, we apply the Strip MLP layer into Wave-MLP-T (namely Wave-Strip-MLP-T) on ImageNet-1K, and we set the strip width as 3. The Wave-Strip-MLP-T gets better results than the original Wave-MLP-T model: $81.2\%$ vs. $80.6\%$ . The performance improvement by $+0.6\%$ shows that our Strip MLP makes token interaction more efficient and can easily be used in other MLP models. + +In conclusion, as shown in Fig. 4, our Strip-MLP has significant superiority in performance and complexity over other state-of-the-art methods on different datasets, demonstrating our Strip-MLP is remarkably efficient and effective to enrich the token interaction power for vision MLP. + +# 4.5. Ablation Studies + +To better show the effectiveness of the proposed method, we ablate the key components of the model design. Due to the limited GPU resources, we do ablation studies on the + +
Strip WidthDatasetParamsFLOPsTop-1(%)
141.33M6.71G91.38
3Caltech-10143.66M6.83G92.09
545.99M6.96G91.72
748.33M7.08G91.86
140.71M6.71G86.09
3CIFAR-10041.88M6.83G86.50
543.04M6.96G86.39
744.21M7.08G86.35
+ +Table 7. Ablation study on the strip width of Strip MLP. We randomly select the model of Strip-MLP-S for Caltech-101 and Strip-MLP-S for CIFAR-100. + +
DatasetPatchesParamsFLOPsTop-1(%)
Caltech-101C/147.22M6.83G91.81
C/243.66M6.83G92.09
C/441.88M6.83G91.38
C/840.99M6.83G91.33
140.16M6.83G90.93
CIFAR-100C/147.22M6.83G86.20
C/243.66M6.83G86.18
C/441.88M6.83G86.50
C/840.99M6.83G85.98
140.16M6.83G85.03
+ +Table 8. Ablation study about the patch number of the CGSMM. The base model is Strip-MLP-S. $C/1$ means splitting the feature $C$ into C patches in channel dimension, namely $P = C/1 = C$ . + +datasets of both Caltech-101 and CIFAR-100. + +The effect of Strip Width in Strip MLP. The strip width affects the token interaction ranges and determines the field of each row or column's contribution to adjacent tokens. We further conduct experiments to verify its effect by varying the width from 1 to 7 with a step at 2. Tab. 7 shows the performances of our models on Caltech-101 [6] and CIFAR-100 [14], respectively. When the width becomes larger, the performance increases and tends to saturate, indicating that Strip MLP layer improves the token interaction power. In all of our comparison experiments in previous sections, we consistently set the strip width as 3 according to the ablation experiment results. + +Effects of Patch Number in CGSMM. Different patch number brings varying degrees of token interaction power improvement and affects the model performance. In Tab. 8, we design five decreased patch numbers from C into 1 to show the effectiveness of CGSMM. Without the patch splitting $(P = 1)$ and GSML operation, the model performance decreases by $1.16\%$ on Caltech-101 $(92.09\% \rightarrow 90.93\%)$ and decreases by $1.47\%$ on CIFAR-100 $(86.50\% \rightarrow 85.03\%)$ , which consistently proves the effectiveness of CGSMM in improving the token interaction power. From the experimental results, we can observe that the optimal patch number is different on the datasets with scale differences, so the optimal patch number should be determined through validation experiments. We consistently set the patch number as C/4 in other ablation studies. + +Cascade vs. Parallel Architecture of GSML. Applying the GSML in a cascade architecture enables the token interacts with other tokens across the entire 2D space in just one + +![](images/da2a372e6dd3a92b01b72bb3ae18d45f3cffe811b555ccc3b34dc7672c0eb6f5.jpg) +(a) + +![](images/c6ee3e3a0618b95346da772b4215bffd339d7a08937fc9b6fdb4dbb75e7f2365.jpg) +(b) + +![](images/c3f1c9808432c7e150dc2405f6bfff2ce1f15998572a6fcc8260daa8ee2cbb44.jpg) +(c) + +![](images/afdfca1c88a0a22a5aa7e35e65849535c686b829c3d08eb76305d1d5c7816207.jpg) +(d) + +![](images/fcdc215a8d20a99fc4b69d5a13d78fbff2f79a03d09bcd8fd6899bd9ff484147.jpg) +(e) + +![](images/2acffc65d5fb5f26ae4119e7494d70c001a34684da9aa66b1cd924dd772912c6.jpg) +(f) + +![](images/5586cd5f719801abf8b86ece3cb56df4636167654ed88d3b79555aff15605919.jpg) +(g) +Figure 4. Performance comparison on four datasets of ImageNet-1K [4], CIFAR-100 [14], CIFAR-10 [14], and Caltech-101 [6]. The Top-1 accuracy, number of parameters and FLOPs are reported with networks of CNN-based, Transformer-based, and MLP-based models. + +![](images/ac8ff3de25819cc0eb6500cda587101794a58c3d13e72c45115709614a9b15a2.jpg) +(h) + +
MethodDatasetParamsFLOPsTop-1(%)
Parallel CascadeCaltech-10142.67M6.65G91.61
43.66M6.83G92.09
Parallel CascadeCIFAR-10040.90M6.66G85.89
41.88M6.83G86.50
+ +Table 9. Ablation study on the cascade and parallel structure of the Strip MLP in Group Strip MLP. Parallel means applying the Strip MLP on the row and column direction in parallel. Cascade means the proposed CGSMM method in this paper. + +module, which needs two modules for the parallel structure so that may decrease the efficiency of token interaction. In Tab. 9, we test the effects between the cascade structure of CGSMM and the parallel structure of Parallel Group Strip Mixing Module (PGSMM). Our experiments show that the cascade structure gets a higher accuracy compared to the parallel structure, with an increase of $+0.48\%$ and $+0.61\%$ on the Caltech-101 and CIFAR-100 datasets, respectively. + +Roles of CGSMM & LSMM. To verify the significance of two complementary branches, we conduct experiments where we only keep one of the branches of CGSMM or LSMM. Tab. 10 displays the outcomes of our trial on Caltech-101 and CIFAR-100. Notably, we found that removing any module from either of the two branches significantly reduced the model's performance. These results highlight the essential role of both CGSMM and LSMM in enriching the tokens' interaction power. + +# 5. Conclusion + +This paper presents an efficient and effective token interaction MLP model of Strip-MLP for vision MLP. For the + +
MethodDatasetParamsFLOPsTop-1(%)
CGSMM OnlyCaltech-10143.40M7.55G91.10
LSMM Only43.96M6.00G90.31
CGSMM & LSMM41.88M6.83G91.38
CGSMM OnlyCIFAR-10043.40M7.55G85.20
LSMM Only43.96M6.00G86.13
CGSMM & LSMM41.88M6.83G86.50
+ +Table 10. Ablation study on the roles of two branches of CGSMM and LSMM. The base model is Strip-MLP-S. When any module branch has been removed, all input features would be fed into the remaining branch module. + +Token's interaction dilemma problem, we design the Strip Mixing Block to enrich the token interaction power by three strategies. Our experimental analysis reveals that the StripMLP remarkably improves the performances of the MLP-based models on small datasets and keeps comparable or even better results on ImageNet with great superiorities on the number of parameters and FLOPs. These observations clearly demonstrate that our models make token interaction more efficient among MLP-based models in image classification tasks. It is expected that Strip MLP layer has the potential to be a standard layer in the future and have promising performance on diverse vision tasks. + +Acknowledgement: This work is supported by National Key Research and Development Program of China (2021YFF1200800); Peng Cheng Laboratory Research Project (PCL2023AS6-1); and Guangdong Basic and Applied Basic Research Foundation under Grant 2022A1515110573. + +# References + +[1] Shoufa Chen, Enze Xie, Chongjian Ge, Ding Liang, and Ping Luo. Cycle.mp: A mlp-like architecture for dense prediction. In International Conference on Learning Representation (ICLR), Oral, 2022. +[2] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. +[3] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005. +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations. +[6] Li Fei-Fei, Robert Fergus, and Pietro Perona. One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence, 28(4):594-611, 2006. +[7] Jianyuan Guo, Yehui Tang, Kai Han, Xinghao Chen, Han Wu, Chao Xu, Chang Xu, and Yunhe Wang. Hire-mlp: Vision mlp via hierarchical rearrangement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 826-836, 2022. +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[9] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. +[10] Qibin Hou, Zihang Jiang, Li Yuan, Ming-Ming Cheng, Shuicheng Yan, and Jiashi Feng. Vision permutator: A permutable mlp-like architecture for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. +[11] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. +[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. +[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[14] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. +[15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. + +[16] Hanxiao Liu, Zihang Dai, David So, and Quoc V Le. Pay attention to mlps. Advances in Neural Information Processing Systems, 34:9204-9215, 2021. +[17] Ruiyang Liu, Yinghui Li, Linmi Tao, Dun Liang, and Haitao Zheng. Are we ready for a new paradigm shift? a survey on visual deep mlp. Patterns, 3(7):100520, 2022. +[18] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. +[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. +[20] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. +[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. +[22] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. +[23] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. +[24] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. +[25] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. +[26] Chuanxin Tang, Yucheng Zhao, Guangting Wang, Chong Luo, Wenxuan Xie, and Wenjun Zeng. Sparse mlp for image recognition: is self-attention really necessary? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2344–2351, 2022. +[27] Yehui Tang, Kai Han, Jianyuan Guo, Chang Xu, Yanxi Li, Chao Xu, and Yunhe Wang. An image patch is a wave: Phase-aware vision mlp. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10935-10944, 2022. +[28] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261-24272, 2021. + +[29] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Gautier Izacard, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, et al. Resmlp: Feedforward networks for image classification with data-efficient training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. +[30] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. +[31] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022. +[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[33] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. +[34] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. +[35] Ziyu Wang, Wenhao Jiang, Yiming M Zhu, Li Yuan, Yibing Song, and Wei Liu. Dynamixer: a vision mlp architecture with dynamic mixing. In International Conference on Machine Learning, pages 22691-22701. PMLR, 2022. +[36] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023. +[37] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021. +[38] David Junhao Zhang, Kunchang Li, Yunpeng Chen, Yali Wang, Shashwat Chandra, Yu Qiao, Luoqi Liu, and Mike Zheng Shou. Morphmlp: A self-attention free, mlp-like backbone for image and video. arXiv preprint arXiv:2111.12527, 2021. +[39] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. +[40] Yipeng Zhang, Zahra M Aghajan, Matias Ison, Qiujing Lu, Hanlin Tang, Guldamla Kalender, Tonmoy Monsoor, Jie Zheng, Gabriel Kreiman, Vwani Roychowdhury, et al. De + +coding of human identity by computer vision and neuronal vision. Scientific reports, 13(1):651, 2023. \ No newline at end of file diff --git a/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/images.zip b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8dd909b3d3da95a3c83733de43c2d9ebbfdb1204 --- /dev/null +++ b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:176a84490b9af54f62940ab4e4911bbd36f7df652902eb9b765d03e06add396e +size 1006044 diff --git a/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/layout.json b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..71e1d64b8b66ffe05971840ded232427a525c439 --- /dev/null +++ b/2023/Strip-MLP_ Efficient Token Interaction for Vision MLP/layout.json @@ -0,0 +1,9742 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 470, + 121 + ], + "type": "text", + "content": "Strip-MLP: Efficient Token Interaction for Vision MLP" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "spans": [ + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "text", + "content": "Guiping Cao" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "text", + "content": " Shengda Luo" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "text", + "content": " Wenjian Huang" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "text", + "content": " Xiangyuan Lan" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "text", + "content": " Dongmei Jiang" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "text", + "content": " Yaowei Wang" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "text", + "content": " Jianguo Zhang" + }, + { + "bbox": [ + 118, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 172, + 459, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 172, + 459, + 199 + ], + "spans": [ + { + "bbox": [ + 136, + 172, + 459, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 172, + 459, + 199 + ], + "type": "text", + "content": "Southern University of Science and Technology, Shenzhen, China " + }, + { + "bbox": [ + 136, + 172, + 459, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 172, + 459, + 199 + ], + "type": "text", + "content": "Peng Cheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 201, + 487, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 201, + 487, + 228 + ], + "spans": [ + { + "bbox": [ + 104, + 201, + 487, + 228 + ], + "type": "text", + "content": "12131099@mail.sustech.edu.cn, {luosd, huangwj, zhangjg}@sustech.edu.cn, {lanxy, jiangdm, wangyw}@pcl.ac.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 280, + 290, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 280, + 290, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 290, + 639 + ], + "type": "text", + "content": "Token interaction operation is one of the core modules in MLP-based models to exchange and aggregate information between different spatial locations. However, the power of token interaction on the spatial dimension is highly dependent on the spatial resolution of the feature maps, which limits the model's expressive ability, especially in deep layers where the feature are down-sampled to a small spatial size. To address this issue, we present a novel method called Strip-MLP to enrich the token interaction power in three ways. Firstly, we introduce a new MLP paradigm called Strip MLP layer that allows the token to interact with other tokens in a cross-strip manner, enabling the tokens in a row (or column) to contribute to the information aggregations in adjacent but different strips of rows (or columns). Secondly, a Cascade Group Strip Mixing Module (CGSMM) is proposed to overcome the performance degradation caused by small spatial feature size. The module allows tokens to interact more effectively in the manners of within-patch and cross-patch, which is independent to the feature spatial size. Finally, based on the Strip MLP layer, we propose a novel Local Strip Mixing Module (LSMM) to boost the token interaction power in the local region. Extensive experiments demonstrate that Strip-MLP significantly improves the performance of MLP-based models on small datasets and obtains comparable or even better results on ImageNet. In particular, Strip-MLP models achieve higher average Top-1 accuracy than existing MLP-based models by " + }, + { + "bbox": [ + 47, + 280, + 290, + 639 + ], + "type": "inline_equation", + "content": "+2.44\\%" + }, + { + "bbox": [ + 47, + 280, + 290, + 639 + ], + "type": "text", + "content": " on Caltech-101 and " + }, + { + "bbox": [ + 47, + 280, + 290, + 639 + ], + "type": "inline_equation", + "content": "+2.16\\%" + }, + { + "bbox": [ + 47, + 280, + 290, + 639 + ], + "type": "text", + "content": " on CIFAR-100. The source codes will be available at https://github.com/MedProcess/Strip_MLP." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 648, + 127, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 648, + 127, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 648, + 127, + 659 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 659, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 287, + 696 + ], + "type": "text", + "content": "In computer vision, Convolutional Neural Networks (CNNs) are one of the most popular network backbones, which have made a series of breakthroughs [8]. Inspired by" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 343, + 254, + 510, + 325 + ], + "blocks": [ + { + "bbox": [ + 343, + 254, + 510, + 325 + ], + "lines": [ + { + "bbox": [ + 343, + 254, + 510, + 325 + ], + "spans": [ + { + "bbox": [ + 343, + 254, + 510, + 325 + ], + "type": "image", + "image_path": "4491c943a6c738d7f2000e08d7a09d1c31030a5bd3e4ddb4b67c136e9d374b3a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 327, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 305, + 327, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 547, + 373 + ], + "type": "text", + "content": "Figure 1. Token interaction of the MLP layer on the down-sampled image feature. (a) The process of which the feature resolution will be down-sampled from " + }, + { + "bbox": [ + 305, + 327, + 547, + 373 + ], + "type": "inline_equation", + "content": "H \\times W \\times C" + }, + { + "bbox": [ + 305, + 327, + 547, + 373 + ], + "type": "text", + "content": " into a small spatial size of " + }, + { + "bbox": [ + 305, + 327, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\frac{H}{8} \\times \\frac{W}{8} \\times 8C" + }, + { + "bbox": [ + 305, + 327, + 547, + 373 + ], + "type": "text", + "content": ". (b) Token interaction on different feature resolutions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 388, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 508 + ], + "type": "text", + "content": "the great success of self-attention-based architectures [32] in natural language processing (NLP), the Transformer models [5, 19, 34] are introduced into the field of computer vision, and have achieved comparable results with state-of-the-art (SOTA) CNNs. Although ViT [5] and its variants outperform traditional CNNs, the models introduce high computational complexity to construct attention maps. Recently, some studies [20, 26] in the vision community suggest that the attention mechanism is not necessary and simpler model architectures are proposed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 714 + ], + "type": "text", + "content": "MLP-based models, like MLP-Mixer [28], gMLP [16] and ViP [10] process the data with the Multilayer Perceptrons (MLP), showing great potential to improve the performance of vision models [29]. As the first visual deep MLP network, MLP-Mixer [28] introduces two types of MLP layers: Channel-Mixing MLPs (CMM) and Token-Mixing MLPs (TMM). For CMM, the module mainly mixes the information between different channels of each token. For TMM, it allows each spatial token to interact with all other tokens (the whole image) in a single MLP layer. However, this design also introduces a much larger number of parameters and higher computational complexity prone to overfitting. To address this problem, Sparse MLP (SMLP) [26] and Vision Permutator (ViP) [10] propose a similar layer of parallel structure, which applies one dimension MLP along the axial directions, and parameters are shared among rows or columns, respectively. Therefore, it reduces the number" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1494" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 285, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 285, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 285, + 95 + ], + "type": "text", + "content": "of model parameters and computational complexity, avoiding the common over-fitting problem." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 98, + 286, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 98, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 49, + 98, + 286, + 335 + ], + "type": "text", + "content": "Although SMLP and ViP alleviate some deficiencies of MLP-Mixer [28], both methods bring the challenge that the token interaction power is highly dependent on the feature spatial size when interacting tokens on spatial rows (or columns). As shown in Fig. 1, the spatial feature resolution is down-sampled to a small size but with more channels, which means the feature pattern of each token is mainly concentrated on the channel dimension rather than the spatial one. Interacting tokens along the spatial dimension by sharing the weights among all channels would seriously ignore the feature pattern differences among different channels, which may degrade the token interaction power, especially in deep layers with small spatial feature resolution. Here, we mark this problem as the Token's interaction dilemma. Taking SMLP as an example, we analyze the feature resolution and complexity of the model in different stages in detail (seen in Sec. 3.4). We find that as the spatial feature size decreases by down-sampling stage by stage, the token interaction layer also becomes smaller and smaller, which makes the token interaction power degraded rapidly." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 338, + 286, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 286, + 587 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 286, + 587 + ], + "type": "text", + "content": "To address these challenges, we propose a new efficient Strip MLP model, dubbed Strip-MLP, to enrich the power of the token interaction layer in three ways. For the level of a single MLP layer, inspired by the cross-block normalization schemes of HOG [3] and the sparse connections between the biological neurons, we design a Strip MLP layer to allow the token to interact with other tokens in a crossstrip manner, enabling each row or column of the tokens to contribute differently to other rows or columns. For the token interaction module level, we develop channel-wise group mixing of CGSMM, enabling the tokens in a row (or column) to contribute to the information aggregations in adjacent but different strips of rows (or columns). to tackle the problem that the token interaction power decreases in deep layers with the spatial feature size significantly reduced but with multiplying channels. Considering the existing methods [10, 26, 28] interact the tokens mainly in the long range of row (or column), which may not aggregate tokens well in the local region, we propose a new Local Strip Mixing Module (LSMM) with a small Strip MLP unit to strengthen the token interaction power on local interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 589, + 285, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 589, + 285, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 589, + 285, + 612 + ], + "type": "text", + "content": "The proposed Strip-MLP model significantly boosts the token interaction power, and the main contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 618, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 57, + 618, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 618, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 57, + 618, + 286, + 676 + ], + "type": "text", + "content": "- A new MLP paradigm for vision MLP: Strip MLP layer, which aggregates the adjacent tokens in a cross strip manner and enables each row or column of the tokens to contribute differently to other rows or columns, interacting tokens more efficiently." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 677, + 286, + 712 + ], + "type": "text", + "content": "- Designing a Cascade Group Strip Mixing module and a Local Strip Mixing module, which effectively improves the model's token interaction power and boosts" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 72, + 544, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 544, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 544, + 84 + ], + "type": "text", + "content": "the tokens aggregation in the local region, respectively;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 85, + 544, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 85, + 544, + 190 + ], + "spans": [ + { + "bbox": [ + 317, + 85, + 544, + 190 + ], + "type": "text", + "content": "- Extensive experiments show that Strip-MLP remarkably improves the performances of the MLP-based models. Strip-MLP achieves higher average Top-1 accuracy by " + }, + { + "bbox": [ + 317, + 85, + 544, + 190 + ], + "type": "inline_equation", + "content": "+2.44\\%" + }, + { + "bbox": [ + 317, + 85, + 544, + 190 + ], + "type": "text", + "content": " on Caltech-101 and " + }, + { + "bbox": [ + 317, + 85, + 544, + 190 + ], + "type": "inline_equation", + "content": "+2.16\\%" + }, + { + "bbox": [ + 317, + 85, + 544, + 190 + ], + "type": "text", + "content": " on CIFAR-100 over the existing MLP-based models. In addition, our models achieve comparable or even better performances on ImageNet-1K compared with traditional MLP-based models, and other popular CNNs and transformer-based models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 206, + 391, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 391, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 391, + 218 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 222, + 544, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 222, + 544, + 269 + ], + "spans": [ + { + "bbox": [ + 308, + 222, + 544, + 269 + ], + "type": "text", + "content": "Deep neural networks for vision recognition can be mainly divided into three categories: Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and Multi-Layer Perceptron-based models (MLPs) [10]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 270, + 544, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 270, + 544, + 413 + ], + "spans": [ + { + "bbox": [ + 308, + 270, + 544, + 413 + ], + "type": "text", + "content": "CNNs-Based Models. CNNs are the de-facto standard deep learning network models for vision tasks and have been intensely studied in the vision community. AlexNet [15] is a symbolically significant model that won the ILSVRC 2012 contest with far higher performance than others. Since then, CNNs-based models have attracted more attentions, and lots of deeper and more effective architectures [8, 11, 23-25, 33, 39, 40] have been proposed. With the convolution and pooling layer, CNNs aggregate the feature in a local region but not well in the long-term dependencies, which are optimized by the new vision model like Transformer [32] and MLP [28] models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 415, + 544, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 544, + 593 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 544, + 593 + ], + "type": "text", + "content": "Transformer-Based Models. Transformer [32] is introduced for machine translation and becomes the reference model for all-natural language processing (NLP) tasks. Inspired by the great success of Transformer in NLP, ViTs [5] first applies a standard Transformer to images, which attains excellent results compared to the SOTA CNNs model. DeiT [30] introduces several training strategies and distillation methods to make ViTs more effective on the smaller ImageNet-1K dataset. By proposing a hierarchical Transformer with shifted windows, Swin Transformer [19] achieves SOTA accuracy on ImageNet-1K, bringing greater efficiency by self-attention in local window and cross-window connection. For these models, self-attention is the core module but with heavy computational burdens to obtain an attention map." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 594, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 594, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 594, + 544, + 712 + ], + "type": "text", + "content": "MLP-Based Models. Without the convolutions and self-attention mechanism, MLP-Mixer [28] builds the architecture that only uses the MLP layer and achieves competitive performance on image classification benchmarks. Since then, the researchers have developed many MLP-like variants [1, 7, 16, 26, 27, 29, 31, 35, 38] models. The work [17] gives a comprehensive survey about visual deep MLP models and compares the intrinsic connections and differences between convolution, self-attention mechanism, and Token-mixing MLP in detail. Sparse MLP [26] introduced a sparse" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 750, + 314, + 757 + ], + "type": "text", + "content": "1495" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 71, + 503, + 264 + ], + "blocks": [ + { + "bbox": [ + 93, + 71, + 503, + 264 + ], + "lines": [ + { + "bbox": [ + 93, + 71, + 503, + 264 + ], + "spans": [ + { + "bbox": [ + 93, + 71, + 503, + 264 + ], + "type": "image", + "image_path": "c5a7848b3d15c4f0fd4fee5d97f45e64feb6afbe1e482af05689489606d22fe6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 547, + 300 + ], + "type": "text", + "content": "Figure 2. The overall and components architecture of Strip-MLP. (a) Strip-MLP has four stages, and " + }, + { + "bbox": [ + 46, + 266, + 547, + 300 + ], + "type": "inline_equation", + "content": "T1" + }, + { + "bbox": [ + 46, + 266, + 547, + 300 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 266, + 547, + 300 + ], + "type": "inline_equation", + "content": "T4" + }, + { + "bbox": [ + 46, + 266, + 547, + 300 + ], + "type": "text", + "content": " means the repeated times of the block in each stage. (b) The Strip Mixing Block splits the head (channel dimension) into two heads, and they are fed into the two parallel branches. (c) Channel Mixing Block architecture." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 311, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 289, + 491 + ], + "type": "text", + "content": "operation to separately aggregate information along axial directions, avoiding the quadratic computational complexity of conventional MLP. Hire-MLP [7] presents a novel variant of MLP-based architecture via hierarchically rearranging tokens to aggregate local and global spatial information. Wave-MLP [27] represents each token as a wave function with two parts of amplitude and phase, which has the ability to model varying contents from different input images. Above all, there is still the limitation of token interaction power degraded significantly especially when the spatial feature resolution becomes small, which has been overlooked by previous studies. In this paper, we aim to enrich the token interaction power on both the single MLP layer and token interaction module to advance the performance of MLP-based models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 501, + 104, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 104, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 104, + 513 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": "In this section, we first present the overall architecture of Strip-MLP. Then, we show the key proposed components: Strip MLP Layer, Cascade Group Strip Mixing Module (CGSMM) and Local Strip Mixing Module (LSMM) of the model in detail, and present the comparison analysis between the Strip MLP and traditional MLP model in terms of parameters and complexity. Finally, we define four kinds of architecture variants to compare the performance of the models in different sizes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 628, + 167, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 167, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 167, + 639 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "type": "text", + "content": "An overview of the Strip-MLP model is depicted in Fig. 2 (a). We design the Strip-MLP as a hierarchical structure of a multi-stages model. Given the input image " + }, + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 641, + 287, + 700 + ], + "type": "text", + "content": ", Strip-MLP model makes patch embedding and models the image feature in four stages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Patch Embedding. Strip-MLP firstly splits the input" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "text", + "content": "image " + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "text", + "content": " (H, W: image height and width) into a sequence of image patches (also referred as tokens) " + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "inline_equation", + "content": "I_{p} \\in \\mathbb{R}^{hw \\times c}" + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "text", + "content": " (c: the number of channels), where the patch size is " + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "inline_equation", + "content": "p \\times p" + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "text", + "content": " and patch number is " + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "inline_equation", + "content": "hw\\left(h = \\frac{H}{p}, w = \\frac{W}{p}\\right)" + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "text", + "content": ", and then all patches are linearly projected into a desired higher dimension " + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "text", + "content": " of the feature " + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "inline_equation", + "content": "(X \\in \\mathbb{R}^{hw \\times C})" + }, + { + "bbox": [ + 304, + 310, + 545, + 382 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 383, + 546, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 546, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 546, + 563 + ], + "type": "text", + "content": "Mixing Block. The purpose of mixing block is to boost the interactions between the features of diverse spatial locations and channels. To effectively aggregate spatial and channel information, we design two sub-blocks of the Strip Mixing Block and Channel Mixing Block. The Strip Mixing Block is comprised of CGSMM and LSMM, which blend and aggregate spatial information more efficiently at the global and local level, respectively. Patch merging module aims to merge the feature in which the spatial dimension will be reduced by " + }, + { + "bbox": [ + 304, + 383, + 546, + 563 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 383, + 546, + 563 + ], + "type": "text", + "content": ", and the channel dimension increases by 2 times so that the model down-samples the feature from " + }, + { + "bbox": [ + 304, + 383, + 546, + 563 + ], + "type": "inline_equation", + "content": "\\frac{H}{p} \\times \\frac{W}{p}" + }, + { + "bbox": [ + 304, + 383, + 546, + 563 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 383, + 546, + 563 + ], + "type": "inline_equation", + "content": "\\frac{H}{8p} \\times \\frac{W}{8p}" + }, + { + "bbox": [ + 304, + 383, + 546, + 563 + ], + "type": "text", + "content": " stage by stage. To obtain multi-scale features, we apply a single convolutional layer to the output features of stages 1 and 2. Then, we add the resulting features to the input features of stages 3 and 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 563, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 611 + ], + "type": "text", + "content": "Head Layer. The features extracted by the multiple blocks are fed into a global average pooling (GAP) layer to reduce the feature dimension. Finally, the feature will be fed into a fully-connected head layer for classification." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 615, + 545, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 615, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 615, + 545, + 628 + ], + "type": "text", + "content": "3.2. Strip Mixing Block and Channel Mixing Block" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Strip Mixing Block. To improve the token interaction power, we design the block to aggregate both long and short range interactions in a parallel manner. As illustrated in Fig. 2 (b), we split the feature in channel dimension, with one-half of the channel feature being fed into CGSMM to model the long-range interactions, and the remaining feature being fed into LSMM to aggregate the local interac" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1496" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 71, + 395, + 198 + ], + "blocks": [ + { + "bbox": [ + 92, + 71, + 395, + 198 + ], + "lines": [ + { + "bbox": [ + 92, + 71, + 395, + 198 + ], + "spans": [ + { + "bbox": [ + 92, + 71, + 395, + 198 + ], + "type": "image", + "image_path": "4ab71b6bb1c028c6a1606c9715c1f4989d2f035242aa57032a5b61195168bdef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 199, + 303, + 209 + ], + "lines": [ + { + "bbox": [ + 184, + 199, + 303, + 209 + ], + "spans": [ + { + "bbox": [ + 184, + 199, + 303, + 209 + ], + "type": "text", + "content": "(a) Cascade Group Strip Mixing Module" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 398, + 71, + 500, + 198 + ], + "blocks": [ + { + "bbox": [ + 398, + 71, + 500, + 198 + ], + "lines": [ + { + "bbox": [ + 398, + 71, + 500, + 198 + ], + "spans": [ + { + "bbox": [ + 398, + 71, + 500, + 198 + ], + "type": "image", + "image_path": "6400613916a9863fc20833d1ff07defb126b953838f96da8bed118121060456d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 199, + 494, + 209 + ], + "lines": [ + { + "bbox": [ + 403, + 199, + 494, + 209 + ], + "spans": [ + { + "bbox": [ + 403, + 199, + 494, + 209 + ], + "type": "text", + "content": "(b) Local Strip Mixing Module" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 212, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 547, + 234 + ], + "type": "text", + "content": "Figure 3. (a) The architecture of CGSMM. It contains the Strip MLP layer, the feature splitting, restoring, and permuting operation in a cascade mode. (b) The structure of the Local Strip Mixing module. The dark green block composes of a re-weight module." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 252, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 287, + 277 + ], + "type": "text", + "content": "tions. Given the input feature " + }, + { + "bbox": [ + 46, + 252, + 287, + 277 + ], + "type": "inline_equation", + "content": "X\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 46, + 252, + 287, + 277 + ], + "type": "text", + "content": ", the block can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 281, + 287, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 281, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 287, + 293 + ], + "type": "interline_equation", + "content": "X _ {m} = M L P \\left(\\left(D W S C (X)\\right)\\right) \\tag {1}", + "image_path": "c72a4c4685358946d28813d13c365f17a86b8170ad104ba0f369afc411eb08c4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 296, + 287, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 296, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 287, + 313 + ], + "type": "interline_equation", + "content": "Y = F C \\left(C a t \\left(C G S M M \\left(X _ {m _ {1}} ^ {\\frac {C}{2}}\\right), L S M M \\left(X _ {m _ {2}} ^ {\\frac {C}{2}}\\right)\\right)\\right) + X \\tag {2}", + "image_path": "5748207f3e575b9f3ff8cc2ee6b9f10d1b14a26f9c48e4f108bc1a6fd33a6948.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "X_{m}" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": " are the intermediate and output feature of the block. " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "X_{m}" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": " is split into " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "X_{m_1}" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "X_{m_2}" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": " with half channel, respectively. DWSC means depth-wise convolution [2], and the kernel size is " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": ". MLP is a serial connection of fullyconnection (FC), batch normalization [12] and GELU [9] activation layers. " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "Cat(*)" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": " denotes concatenation operation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 456 + ], + "type": "text", + "content": "Channel Mixing Block. The block aims to aggregate the information between channels of the token, and the basic structure is shown in Fig. 2 (c). Following the inverted bottleneck in [20] and global response normalization (GRN) in [36], we design the Channel Mixing block to increase the contrast and selectivity of channels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 460, + 170, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 170, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 170, + 471 + ], + "type": "text", + "content": "3.3. CGSMM and LSMM" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 474, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 617 + ], + "type": "text", + "content": "We first introduce the Strip MLP Layer. Based on this layer, we design CGSMM using a simple but effective strategy that splits the feature into patches along channel dimension and interacts tokens more effectively in the manners of within-patch and cross-patch, which is independent to the feature spatial size. In addition, the design of existing methods [10, 26, 28] only allows tokens to interact in a long range of rows (or columns) with sharing weight; thus the resulting model may struggle with effectively aggregating both global and local information simultaneously. Therefore, we design LSMM to capture the local interactions more efficiently." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "Strip MLP Layer. In MLP-based models, the majority of MLP layers treat each row and column of the data in isolation, as formulated in Eq. (3), which may lead to the inefficiency of token interaction. Inspired by the cross-block normalization schemes of HOG [3], which overlaps the blocks and enables each scalar cell response to contribute into different blocks, we propose the Strip MLP layer. The proposed layer applies MLP on the \"strip\" data of adjacent" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 254, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 254, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 545, + 300 + ], + "type": "text", + "content": "rows or columns in the spatial direction in order to aggregate the feature in a cross-strip manner. Given the input " + }, + { + "bbox": [ + 305, + 254, + 545, + 300 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 254, + 545, + 300 + ], + "type": "text", + "content": ", we formulated the Strip MLP layer (setting strip width as 3 for example) in Eqs. (4) and (5):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 300, + 545, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 300, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 353, + 300, + 545, + 315 + ], + "type": "interline_equation", + "content": "X _ {*, j} ^ {h} = W _ {1} X _ {*, j}; \\quad X _ {i, *} ^ {w} = W _ {2} X _ {i, *} \\tag {3}", + "image_path": "bf59f28ab1353a914863a28f1f9b32988c1169612cfb0265b3ea211d3df7bb15.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 316, + 545, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 316, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 347, + 316, + 545, + 331 + ], + "type": "interline_equation", + "content": "X _ {*, j} ^ {h} = W _ {3} \\operatorname {C a t} \\left(X _ {*, j - 1}, X _ {*, j}, X _ {*, j + 1}\\right) \\tag {4}", + "image_path": "3f9a1b2d081f2ed59727448b374a96ee99ee79ba99ea62d2d8919cabaab0429f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 348, + 335, + 545, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 335, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 348, + 335, + 545, + 349 + ], + "type": "interline_equation", + "content": "X _ {i, *} ^ {w} = W _ {4} \\operatorname {C a t} \\left(X _ {i - 1, *}, X _ {i, *}, X _ {i + 1, *}\\right) \\tag {5}", + "image_path": "5bff3c070889bcc4a07fb3236e502b0ad49ec1dac51e89167484f1b2ea5aebb6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "inline_equation", + "content": "W_{1}\\sim W_{4}" + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "content": " are the weights of the MLP layer, " + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "content": " are the token index in row and column. The superiorities of the Strip MLP layer are mainly in two aspects: on one hand, Strip MLP layer enables the token to interact with other tokens in both short and long spatial ranges simultaneously. On the other hand, similar to the HOG cross-block normalization process, each row (or column) not only serves for the current row (or column) tokens aggregation but also makes contribution to the adjacent rows (or columns) feature aggregation. For example, in Eq. (4), " + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "inline_equation", + "content": "X_{*,j}" + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "content": " makes different contributions for aggregating the processed feature of " + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "inline_equation", + "content": "X_{*,j - 1}^{h},X_{*,j}^{h}" + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "inline_equation", + "content": "X_{*,j + 1}^{h}" + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "content": ", making tokens interacting more efficiently in a cross-strip manner." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 506, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 565 + ], + "type": "text", + "content": "Cascade Group Strip Mixing Module. The architecture is shown in Fig. 3 (a). The module applies the Strip MLP layer in row and column direction in a cascade mode. As the operation of the Strip MLP layer on row is similar to that on column, we take one of them to show the method." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "type": "text", + "content": "Patch Splitting of the Data along Channel Dimension. Given the input feature " + }, + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "type": "inline_equation", + "content": "X_{m_1} \\in \\mathbb{R}^{H \\times W \\times \\frac{C}{2}}" + }, + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "type": "text", + "content": ", the module firstly permutes the feature into " + }, + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "type": "inline_equation", + "content": "X_{m_1}^{H \\times \\frac{C}{2} \\times W}" + }, + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "type": "text", + "content": ", and splits the feature into " + }, + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "type": "text", + "content": " patches in the channel dimension and concatenated along the row " + }, + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "type": "inline_equation", + "content": "(X^{H \\times \\frac{C}{2P} \\times PW})" + }, + { + "bbox": [ + 304, + 566, + 545, + 629 + ], + "type": "text", + "content": " dimension." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Group Strip MLP Layer. To improve the token interaction power, we propose a Group Strip MLP Layer (GSML) to interact tokens in the manners of within-patch and cross-patch. In particular, we apply the unshared weights of Strip MLP on different patches, and the tokens within the same patch share the weights for interaction. Then, we restore the feature to original shape and concatenate it with the input" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1497" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "feature. To interact the tokens crossing patches, we apply a channel fully-connected (Channel FC) layer to interact tokens between different patches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 287, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 287, + 207 + ], + "type": "text", + "content": "Local Strip Mixing Module. The module is illustrated in Fig. 3 (b). In order to better aggregate the local interactions on spatial dimension, we define a small Strip MLP unit, where the strip width and length are 3 and 7, respectively. Given the input feature " + }, + { + "bbox": [ + 46, + 109, + 287, + 207 + ], + "type": "inline_equation", + "content": "X_{m_2}^{H \\times W \\times \\frac{C}{2}}" + }, + { + "bbox": [ + 46, + 109, + 287, + 207 + ], + "type": "text", + "content": ", we aggregate the local interactions in rows and columns direction simultaneously. Following [1, 10, 27], we sum all branches with a re-weight module." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 218, + 239, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 239, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 239, + 231 + ], + "type": "text", + "content": "3.4. Parameter and Complexity Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 233, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 353 + ], + "type": "text", + "content": "In multi-stage processing architecture, the Token's interaction dilemma becomes more serious, as the feature spatial resolution will be down-sampled in deep layers, and the token interaction power will be weakened. In this section, we present the comparison analysis of parameters and complexity to show the effectiveness of the proposed CGSMM. Considering the majority MLP-Based models have the similar model structure of token interaction, we analyze the parameters and complexities with Sparse MLP [26], which is a popular model with good performance on various datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": "Sparse MLP [26] firstly applies MLP on the columns and rows of " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " to map " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{CW\\times H}" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{CW\\times H}" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " separately. Then, the model concatenates the two processed features with the input feature " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " and uses a linear layer on the channel dimension to fuse the feature from " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{HW\\times 3C}" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{HW\\times C}" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": ". The number of parameters and FLOPs of the first step are " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "W^2 +H^2" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "CHW(H + W)" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": "; they are " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "3C^2" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "3HWC^2" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " for the fusion step. Comparatively, the number of parameters and FLOPs of Strip MLP layer for token interaction are " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "3P(H^{2} + W^{2})" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "3CHW(H + W)" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": ", respectively. The fusion step has " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "4C^2" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " parameters and " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "4HWC^2" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " FLOPs. In particular, for " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "H_{1} = W_{1} = 56" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "H_{4} = W_{4} = 7" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "C_1 = 112" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "C_4 = 112\\times 2^3 = 896" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "P = \\frac{C}{4}" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": ", we calculate the number of parameters and FLOPs of both two layers. As shown in Tab. 1, for the Sparse MLP block, the number of parameters for token interaction in stage 4 (only " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "0.10k" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": ") decreases by 62.70 times than in stage 1. In addition, most of the number of the block's parameters and the FLOPs concentrate in the fusion step. For example, only " + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "0.01\\%" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " parameters (" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "inline_equation", + "content": "0.52\\%" + }, + { + "bbox": [ + 46, + 354, + 287, + 628 + ], + "type": "text", + "content": " FLOPs) are used for the token interaction step in stage 4. Based on the above analysis, to improve the token interaction power, it would be better to redesign the block to balance the number of parameters and FLOPs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "In CGSMM, we use a simple but effective strategy that splits the feature along channel dimension into patches and interacts the tokens in the manners of within-patch and cross-patch. No matter how the spatial resolution decreases, the module can still interact the tokens with channel-wise specificity in different patches. In Tab. 1, we improve the Proportion of the number of parameters in the token inter" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 542, + 148 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 542, + 148 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 542, + 148 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 542, + 148 + ], + "type": "table", + "html": "
ItemsSparse MLPStrip MLP
ParamsFLOPsParamsFLOPs
Stage 16.27k39.34M526.85k118.01M
Stage 40.10k0.62M65.86k1.84M
Changes↓62.70↓63.45↓8.00↓64.14
Stage 4 Fusion2.41M118.01M3.21M157.35M
Proportion0.01%0.52%2.01%1.16%
", + "image_path": "32ce2c452b21075bcd8e235260096dd25e16821ccb5a12a76e0e35c5884d439a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 149, + 545, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 149, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 149, + 545, + 237 + ], + "type": "text", + "content": "Table 1. The model parameters and complexity comparison between Sparse MLP and Strip MLP layer in different stages. Stage 1 and Stage 4 show the parameters (Params) and floating-point operations (FLOPs) for token interaction in one layer between different stages. Stage 4 Fusion indicates the Params and FLOPs in the feature fusion step of the token interaction module. Proportion means the ratio between the token interaction step and other steps in the token interaction module of stage 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 248, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 545, + 355 + ], + "type": "text", + "content": "action layer from " + }, + { + "bbox": [ + 304, + 248, + 545, + 355 + ], + "type": "inline_equation", + "content": "0.01\\%" + }, + { + "bbox": [ + 304, + 248, + 545, + 355 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 248, + 545, + 355 + ], + "type": "inline_equation", + "content": "2.01\\%" + }, + { + "bbox": [ + 304, + 248, + 545, + 355 + ], + "type": "text", + "content": " in stage 4. Although our GSML brings more parameters and computations when just considering one token interaction layer, the total number of parameters and overall computational complexity of the model are decreased as the token interaction power is improved and we set small model configurations of " + }, + { + "bbox": [ + 304, + 248, + 545, + 355 + ], + "type": "inline_equation", + "content": "T_{1} \\sim T_{4}" + }, + { + "bbox": [ + 304, + 248, + 545, + 355 + ], + "type": "text", + "content": ". Additionally, the experiments in Sec. 4 show that our models achieve better performance with fewer parameters and FLOPs than other SOTA models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 360, + 430, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 430, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 430, + 371 + ], + "type": "text", + "content": "3.5. Architecture Variants" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 375, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 447 + ], + "type": "text", + "content": "We developed four variants of our Strip-MLP network: Strip-MLP-T\\* (light tiny), Strip-MLP-T (tiny), Strip-MLPSmall), Strip-MLP-B (base), which have similar or even smaller model sizes, compared with MLP-baed models [7, 26, 27] and Swin Transformer [18], respectively. The hyperparameters of the four variants of the models are as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 455, + 534, + 504 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 315, + 455, + 529, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 455, + 529, + 468 + ], + "spans": [ + { + "bbox": [ + 315, + 455, + 529, + 468 + ], + "type": "text", + "content": "- Strip-MLP-T*: " + }, + { + "bbox": [ + 315, + 455, + 529, + 468 + ], + "type": "inline_equation", + "content": "C = 80, \\{T_1 \\sim T_4\\} = \\{2, 2, 6, 2\\}" + }, + { + "bbox": [ + 315, + 455, + 529, + 468 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 468, + 529, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 468, + 529, + 479 + ], + "spans": [ + { + "bbox": [ + 315, + 468, + 529, + 479 + ], + "type": "text", + "content": "- Strip-MLP-T: " + }, + { + "bbox": [ + 315, + 468, + 529, + 479 + ], + "type": "inline_equation", + "content": "C = 80, \\{T_1 \\sim T_4\\} = \\{2, 2, 12, 2\\}" + }, + { + "bbox": [ + 315, + 468, + 529, + 479 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 479, + 529, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 479, + 529, + 491 + ], + "spans": [ + { + "bbox": [ + 315, + 479, + 529, + 491 + ], + "type": "text", + "content": "- Strip-MLP-S: " + }, + { + "bbox": [ + 315, + 479, + 529, + 491 + ], + "type": "inline_equation", + "content": "C = 96, \\{T_1 \\sim T_4\\} = \\{2, 2, 18, 2\\}" + }, + { + "bbox": [ + 315, + 479, + 529, + 491 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 491, + 534, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 491, + 534, + 504 + ], + "spans": [ + { + "bbox": [ + 315, + 491, + 534, + 504 + ], + "type": "text", + "content": "- Strip-MLP-B: " + }, + { + "bbox": [ + 315, + 491, + 534, + 504 + ], + "type": "inline_equation", + "content": "C = 112, \\{T_1 \\sim T_4\\} = \\{2, 2, 18, 2\\}" + }, + { + "bbox": [ + 315, + 491, + 534, + 504 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "text", + "content": " means the channel number of the hidden layers in the first stage, and " + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "inline_equation", + "content": "T_{1} \\sim T_{4}" + }, + { + "bbox": [ + 304, + 509, + 545, + 556 + ], + "type": "text", + "content": " is the number of times repeated for Strip Mixing Block and Channel Mixing Block in each of the four stages, as shown in Fig. 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 564, + 387, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 564, + 387, + 577 + ], + "spans": [ + { + "bbox": [ + 306, + 564, + 387, + 577 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": "We compare our Strip-MLP with three main categories of networks: CNNs-based, Transformer-based, and MLP-based models. We first show the experimental settings of the datasets and implementation details. Then, the proposed method is compared with other popular methods on three datasets. Finally, we ablate the critical design components of Strip-MLP network." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 673, + 432, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 432, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 432, + 685 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 692, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 692, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 692, + 545, + 715 + ], + "type": "text", + "content": "Datasets. Methods are evaluated in four datasets which vary significantly in the number of training images:" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1498" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 282, + 343 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 282, + 343 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 282, + 343 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 282, + 343 + ], + "type": "table", + "html": "
MethodParamsFLOPsTop-1(%)
CNNs-Based
ResNet50 [8]23.71M4.12G89.61
ResNet101 [8]42.71M7.85G88.56
ResNet152 [8]58.35M11.58G89.04
Transformer-Based
ViT-B/16 [5]85.85M16.86G53.96
Swin-T [19]27.60M4.36G80.40
Swin-S [19]48.91M8.52G79.83
Swin-B [19]86.85M15.14G78.42
MLP-Based
Wave-MLP-T [27]16.73M2.48G86.89
Hire-MLP-T [7]17.58M2.16G87.85
Strip-MLP-T*(ours)17.67M2.53G90.11
Wave-MLP-S [27]30.25M4.55G87.43
Hire-MLP-S [7]32.65M4.26G88.48
Sparse-MLP-T [26]23.55M5.02G90.54
Strip-MLP-T(ours)24.33M3.67G90.93
Wave-MLP-M [27]43.59M7.93G88.45
Hire-MLP-B [7]57.77M8.17G88.62
Sparse-MLP-S [26]47.87M10.36G91.07
Strip-MLP-S(ours)43.66M6.83G92.09
Wave-MLP-B [27]62.90M10.27G88.81
Hire-MLP-L [7]95.03M13.50G88.19
Sparse-MLP-B [26]65.07M14.04G91.67
Strip-MLP-B(ours)58.49M9.22G92.26
", + "image_path": "5649ca10ddc69344cbbac0bd7c69d8add112355e48239a6e6facf78dd17d0458.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 344, + 287, + 376 + ], + "lines": [ + { + "bbox": [ + 47, + 344, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 344, + 287, + 376 + ], + "type": "text", + "content": "Table 2. Image classification results of our Strip-MLP and other models on Caltech-101. Top-1 denotes the Top-1 accuracy. Here, the patch number of CGSMM is C/2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 390, + 286, + 532 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 57, + 390, + 286, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 390, + 286, + 436 + ], + "spans": [ + { + "bbox": [ + 57, + 390, + 286, + 436 + ], + "type": "text", + "content": "- Caltech-101 [6]: it is a benchmark dataset that includes 101 classes, totally around " + }, + { + "bbox": [ + 57, + 390, + 286, + 436 + ], + "type": "inline_equation", + "content": "9k" + }, + { + "bbox": [ + 57, + 390, + 286, + 436 + ], + "type": "text", + "content": " images. We randomly take " + }, + { + "bbox": [ + 57, + 390, + 286, + 436 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 57, + 390, + 286, + 436 + ], + "type": "text", + "content": " of each class of the data as the training set and use the remaining data as the testing set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 437, + 286, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 437, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 57, + 437, + 286, + 460 + ], + "type": "text", + "content": "- CIFAR-10 [14]: it consists of " + }, + { + "bbox": [ + 57, + 437, + 286, + 460 + ], + "type": "inline_equation", + "content": "60k \\times 32" + }, + { + "bbox": [ + 57, + 437, + 286, + 460 + ], + "type": "text", + "content": " images in 10 classes, and there are " + }, + { + "bbox": [ + 57, + 437, + 286, + 460 + ], + "type": "inline_equation", + "content": "50k" + }, + { + "bbox": [ + 57, + 437, + 286, + 460 + ], + "type": "text", + "content": " images" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 462, + 286, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 462, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 57, + 462, + 286, + 495 + ], + "type": "text", + "content": "- CIFAR-100 [14]: it has 100 classes containing 600 images of each class (totally " + }, + { + "bbox": [ + 57, + 462, + 286, + 495 + ], + "type": "inline_equation", + "content": "60k" + }, + { + "bbox": [ + 57, + 462, + 286, + 495 + ], + "type": "text", + "content": "), and there are 500 training images and 100 testing images per class." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 498, + 286, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 498, + 286, + 532 + ], + "spans": [ + { + "bbox": [ + 57, + 498, + 286, + 532 + ], + "type": "text", + "content": "- ImageNet-1K [4]: it is an image classification benchmark containing " + }, + { + "bbox": [ + 57, + 498, + 286, + 532 + ], + "type": "inline_equation", + "content": "1.28M" + }, + { + "bbox": [ + 57, + 498, + 286, + 532 + ], + "type": "text", + "content": " training images and " + }, + { + "bbox": [ + 57, + 498, + 286, + 532 + ], + "type": "inline_equation", + "content": "50k" + }, + { + "bbox": [ + 57, + 498, + 286, + 532 + ], + "type": "text", + "content": " testing images of 1000 classes." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 537, + 286, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 286, + 655 + ], + "type": "text", + "content": "Training Details. In all of our experiments, the input image size is " + }, + { + "bbox": [ + 46, + 537, + 286, + 655 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 537, + 286, + 655 + ], + "type": "text", + "content": ", and the input patch size is " + }, + { + "bbox": [ + 46, + 537, + 286, + 655 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 537, + 286, + 655 + ], + "type": "text", + "content": ". Based on the deep-learning framework of PyTorch [21], we train our models from scratch without any extra data for pre-training except for the transfer learning experiments. We employ an AdamW [13] optimizer with 300 epochs for all training datasets and apply a cosine decay learning rate scheduler and 30 epochs of linear warm-up. In addition, we adopt the same augmentation and regularization strategies as to Swin-Transformer [19]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 663, + 250, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 250, + 675 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 250, + 675 + ], + "type": "text", + "content": "4.2. Image Classification on Small Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": "We conduct experiments on Small Datasets of Caltech-101, CIFAR-10 and CIFAR-100. The results consistently indicate that our Strip-MLP efficiently improves the token" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "interaction power, and advances the MLP-based models' performance on the small dataset than other models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 98, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 98, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 545, + 168 + ], + "type": "text", + "content": "Results on Caltech-101. Tab. 2 presents the image classification result of three types models on Caltech-101 dataset. Compared to ResNet [8], all the four variants of our Strip-MLP models achieve higher Top-1 accuracy with fewer parameters and FLOPs, demonstrating an average increase of " + }, + { + "bbox": [ + 304, + 98, + 545, + 168 + ], + "type": "inline_equation", + "content": "+2.69\\%" + }, + { + "bbox": [ + 304, + 98, + 545, + 168 + ], + "type": "text", + "content": " over three widely-used ResNet models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 170, + 545, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 241 + ], + "type": "text", + "content": "For the Transformer-based models, e.g. Swin Transformer [19], Strip-MLP models noticeably surpass all variants models: " + }, + { + "bbox": [ + 304, + 170, + 545, + 241 + ], + "type": "inline_equation", + "content": "+10.53\\% / 12.26\\% / 13.84\\%" + }, + { + "bbox": [ + 304, + 170, + 545, + 241 + ], + "type": "text", + "content": " over SwinTransformer (T/S/B) models. ViT-B/16 [5] only gets an accuracy of " + }, + { + "bbox": [ + 304, + 170, + 545, + 241 + ], + "type": "inline_equation", + "content": "53.96\\%" + }, + { + "bbox": [ + 304, + 170, + 545, + 241 + ], + "type": "text", + "content": ", indicating that transformer-based models depend more on the scale of the datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "type": "text", + "content": "When compared to the MLP-based models, like Sparse MLP [26], Wave-MLP [27] and Hire-MLP [7], our models still achieve the best accuracy with fewer model parameters and FLOPs. For example, with fewer FLOPs " + }, + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "type": "inline_equation", + "content": "(-4.28\\mathrm{G})" + }, + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "type": "text", + "content": " Strip-MLP-B " + }, + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "type": "inline_equation", + "content": "(92.26\\%)" + }, + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "type": "text", + "content": " has an increase of " + }, + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "type": "inline_equation", + "content": "+3.64\\%" + }, + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "type": "text", + "content": " Top-1 accuracy than Hire-MLP-B " + }, + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "type": "inline_equation", + "content": "(88.62\\%)" + }, + { + "bbox": [ + 304, + 243, + 545, + 314 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "type": "text", + "content": "Results on CIFAR-10. Tab. 3 shows the classification results of performance comparisons. Compared with CNNs-based models, all four variants of our Strip-MLP models achieve higher Top-1 accuracy with fewer parameters and FLOPs, with an average increase of " + }, + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "type": "inline_equation", + "content": "+1.14\\%" + }, + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "type": "text", + "content": " (98.06% vs. 96.92%). When compared with transformer-based models, our Strip-MLP achieves better performance by " + }, + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "type": "inline_equation", + "content": "+3.06\\%" + }, + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "type": "text", + "content": " (98.06% vs. 95.00%) with fewer parameters (34.49M vs. 54.37M) and FLOPs (5.56G vs. 9.34G). Compared with MLP-based models, all of our Strip-MLP models obtain the best Top-1 accuracy with fewer parameters and FLOPs. For example, the average Top-1 accuracy of Strip-MLP is higher than Hire-MLP/Wave-MLP/Sparse-MLP by " + }, + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "type": "inline_equation", + "content": "+1.45\\% / 0.69\\% / 0.19\\%" + }, + { + "bbox": [ + 304, + 316, + 545, + 518 + ], + "type": "text", + "content": " (98.06% vs. 96.61%/97.37%/97.87%) with fewer average parameters (34.49M vs. 50.71M/38.32M/45.43M) and fewer average FLOPs (5.56G vs. 9.36G/6.31G/9.81G)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "type": "text", + "content": "In particular, with only 16.76M parameters and 2.53G FLOPs, our Strip-MLP-" + }, + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "type": "text", + "content": " noticeably surpasses all variants of ResNet [8] and Swin-Transformer [19] models, and Strip-MLP-" + }, + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "type": "text", + "content": " outperforms the transfer learning models of ViT-S/16 [5]/TST-14 [37] by " + }, + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "type": "inline_equation", + "content": "+0.69\\% / +0.29\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "type": "text", + "content": ", consistently demonstrating our Strip-MLP has significant superiorities in performance and complexity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": "Results on CIFAR-100. Tab. 4 compares the image classification results on CIFAR-100 with three kinds of architecture models. Compared with CNNs-based models [8], Strip-MLP obtains average higher accuracy of " + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "+2.05\\% \\sim +2.43\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": " with a smaller number of parameters and FLOPs. When compared to the self-attention based model, e.g. Swin Transformer [19], our model shows great superiority that Strip-MLP-B achieves " + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "86.40\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": " Top-1 accuracy, " + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "+9.64\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": " than the Swin-B model with fewer pa" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1499" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 286, + 339 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 339 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 339 + ], + "type": "table", + "html": "
MethodParamsFLOPsTop-1 (%)
CNNs-Based
ResNet50 [8]23.53M4.12G97.06
ResNet101 [8]42.52M7.85G96.70
ResNet152 [8]58.16M11.58G96.99
Transformer-Based
Swin-T [19]27.53M4.36G95.25
Swin-S [19]48.84M8.52G94.87
Swin-B [19]86.75M15.14G94.87
MLP-Based
Wave-MLP-T [27]16.69M2.48G96.91
Hire-MLP-T [7]17.53M2.16G96.20
Strip-MLP-T*(ours)16.76M2.53G97.79
Wave-MLP-S [27]30.20M4.55G97.27
Hire-MLP-S [7]32.61M4.26G96.70
Sparse-MLP-T [26]23.49M5.02G97.67
Strip-MLP-T(ours)23.07M3.67G97.94
Wave-MLP-M [27]43.55M7.93G97.56
Hire-MLP-B [7]57.72M8.17G96.88
Sparse-MLP-S [26]47.80M10.36G97.99
Strip-MLP-S(ours)41.81M6.83G98.23
Wave-MLP-B [27]62.83M10.27G97.72
Hire-MLP-L [7]94.96M13.50G96.64
Sparse-MLP-B [26]64.99M14.04G97.95
Strip-MLP-B(ours)56.33M9.22G98.28
", + "image_path": "fcf25a67eea67df98c6dc63189609979264ae75d55d06b9078b6a0de08642513.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 340, + 287, + 373 + ], + "lines": [ + { + "bbox": [ + 47, + 340, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 340, + 287, + 373 + ], + "type": "text", + "content": "Table 3. Image classification results of different models on CIFAR-10. The patch number of CGSMM is C/4. All models are trained from scratch without any extra data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 380, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 287, + 464 + ], + "type": "text", + "content": "rameters (58.49M vs. 86.85M) and lower FLOPs (9.22G vs. 15.14G). In addition, our models keep the superiority over the other MLP-based models in that Strip-MLP achieves the best Top-1 accuracy (with a higher range of " + }, + { + "bbox": [ + 47, + 380, + 287, + 464 + ], + "type": "inline_equation", + "content": "+1.04\\% \\sim +3.01\\%" + }, + { + "bbox": [ + 47, + 380, + 287, + 464 + ], + "type": "text", + "content": ") compared to the MLP-based models, e.g. Wave-MLP [27], Hire-MLP [7], and Sparse MLP [26], with fewer parameters and FLOPs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 472, + 243, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 243, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 243, + 485 + ], + "type": "text", + "content": "4.3. Transfer Learning on Small Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "text", + "content": "Following the transfer learning experiment of DynaMixer [35], we show the advantages of our models on the transfer learning tasks. With the pre-trained Strip-MLP" + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^*" + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "text", + "content": " model on the ImageNet-1K, we transfer the model to downstream datasets of CIFAR-10 and CIFAR-100. Compared to the previous models, such as ViT [5], ViP [10] and DynaMixer [35] models, our Strip-MLP" + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^*" + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "text", + "content": " models achieve the best accuracies of " + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "inline_equation", + "content": "98.8\\%" + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "inline_equation", + "content": "89.4\\%" + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "text", + "content": " on two datasets, as shown in Tab. 5, " + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "inline_equation", + "content": "+1.7\\% / 0.6\\%" + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "text", + "content": " than ViT-S/16 and DynaMixer-S on CIFAR-10, " + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "inline_equation", + "content": "+2.3\\% / 0.8\\%" + }, + { + "bbox": [ + 46, + 487, + 288, + 654 + ], + "type": "text", + "content": " over ViT-S/16 and DynaMixer-S on CIFAR-100. More importantly, the number of parameters of our model is only 18M, which is far less than other models, indicating that our model is more efficient and effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 663, + 246, + 676 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 246, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 246, + 676 + ], + "type": "text", + "content": "4.4. Image Classification on ImageNet-1K" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "We conduct experiments on ImageNet-1K to show the effectiveness of the method. Tab. 6 shows the performance together with the complexity (in terms of FLOPS and the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 542, + 339 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 542, + 339 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 542, + 339 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 542, + 339 + ], + "type": "table", + "html": "
MethodParamsFLOPsTop-1(%)
CNNs-Based
ResNet50 [8]23.71M4.12G83.03
ResNet101 [8]42.71M7.85G83.75
ResNet152 [8]58.35M11.58G84.35
Transformer-Based
Swin-T [19]27.60M4.36G78.56
Swin-S [19]48.91M8.52G78.10
Swin-B [19]86.85M15.14G76.76
MLP-Based
Wave-MLP-T [27]16.73M2.48G82.77
Hire-MLP-T [7]17.58M2.16G82.44
Strip-MLP-T*(ours)17.67M2.53G85.10
Wave-MLP-S [27]30.25M4.55G83.34
Hire-MLP-S [7]32.65M4.26G82.94
Sparse-MLP-T [26]23.55M5.02G84.19
Strip-MLP-T(ours)24.33M3.67G85.23
Wave-MLP-M [27]43.59M7.93G83.95
Hire-MLP-B [7]57.77M8.17G83.17
Sparse-MLP-S [26]47.87M10.36G84.27
Strip-MLP-S(ours)43.66M6.83G86.18
Wave-MLP-B [27]62.89M10.27G84.23
Hire-MLP-L [7]95.03M13.50G83.53
Sparse-MLP-B [26]65.07M14.04G84.46
Strip-MLP-B(ours)58.49M9.22G86.40
", + "image_path": "6b0c30cdbb39d414a3839593c637eb11a448b2104ae25a22f3a0728e9010dd24.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 359, + 542, + 470 + ], + "blocks": [ + { + "bbox": [ + 307, + 340, + 544, + 359 + ], + "lines": [ + { + "bbox": [ + 307, + 340, + 544, + 359 + ], + "spans": [ + { + "bbox": [ + 307, + 340, + 544, + 359 + ], + "type": "text", + "content": "Table 4. Image classification results of different models on CIFAR-100. The patch number of CGSMM is C/2." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 359, + 542, + 470 + ], + "lines": [ + { + "bbox": [ + 310, + 359, + 542, + 470 + ], + "spans": [ + { + "bbox": [ + 310, + 359, + 542, + 470 + ], + "type": "table", + "html": "
ModelDatasetParamsTop-1(%)
ViT-S/16 [5]49M97.1
TST-14 [37]22M97.5
ViP-S/7 [10]CIFAR-1025M98.0
DynaMixer-S [35]26M98.2
Strip-MLP-T*(ours)18M98.8
ViT-S/16 [5]49M87.1
TST-14 [37]22M88.4
ViP-S/7 [10]CIFAR-10025M88.4
DynaMixer-S [35]26M88.6
Strip-MLP-T*(ours)18M89.4
", + "image_path": "5996c2853e1c8bba2eee66618d88b7a874f9c6707a1c8fb9c1e86679ba9b0488.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 470, + 545, + 491 + ], + "lines": [ + { + "bbox": [ + 306, + 470, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 545, + 491 + ], + "type": "text", + "content": "Table 5. The transfer learning results of the model pre-trained on ImageNet-1K and fine-tuned to CIFAR-10 and CIFAR-100." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": "number of parameters) of three types of models. Strip-MLP outperforms the CNNs-based model: " + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "inline_equation", + "content": "+0.5\\%" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": " with half the number of parameters (25M vs. 39M) and FLOPs (3.7G vs. 8.0G) for Strip-MLP-T over RegNetY-8G [22]. Compared to Transformer-based models, e.g. Swin Transformer [19], Strip-MLP achieves comparable results but with obvious superiority in terms of the number of parameters and FLOPs, such as Strip-MLP-B achieves similar performance to Swin-B but with fewer parameters (57M vs. 88M) and FLOPs (9.2G vs. 15.4G). Compared to the MLP-based models, Strip-MLP still shows the same efficiency in the various models. The Strip-MLP-" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^*" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": " achieves average higher accuracy by " + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "inline_equation", + "content": "+2.12\\%" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": " than other MLP-based models (81.2% vs. 79.08%). Both Strip-MLP-B and Wave-MLP-B [27] models get the same accuracy of " + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "inline_equation", + "content": "83.6\\%" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": ". However, Strip-MLP-B uses fewer parameters (57M vs. 63M) and FLOPs (9.2G vs. 10.2G), indicating that the token-interaction power of our model is more efficient than other" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1500" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 69, + 289, + 456 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 289, + 456 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 289, + 456 + ], + "type": "table", + "html": "
MethodParamsFLOPsThroughput (image/s)Top-1 (%)
CNNs-Based
ResNet50 [8]26M4.1G-78.5
ResNet101 [8]45M7.9G-79.8
RegNetY-4G [22]21M4.0G115780.0
RegNetY-8G [22]39M8.0G59281.7
RegNetY-16G [22]84M16.0G33582.9
Transformer-Based
DeiT-S [30]22M4.6G94079.8
DeiT-B [30]86M17.5G29281.8
Swin-T [19]29M4.5G75581.3
Swin-S [19]50M8.7G43783.0
Swin-B [19]88M15.4G27883.5
MLP-Based
ResMLP-S12 [29]15M3.0G141576.6
gMLP-S [16]20M4.5G-79.6
CycleMLP-B1 [1]15M2.1G103878.9
Hire-MLP-Tiny [7]18M2.1G156279.7
Wave-MLP-T [27]17M2.4G120880.6
Strip-MLP-T*(ours)18M2.5G81481.2
ResMLP-S24 [29]30M6.0G71579.4
Sparse-MLP-T [26]24M5.0G63981.9
CycleMLP-B2 [1]27M4.0G64181.6
Hire-MLP-Small [7]33M4.2G80882.1
Wave-MLP-S [27]30M4.5G72082.6
Strip-MLP-T(ours)25M3.7G59782.2
CycleMLP-B4 [1]52M10.1G32183.0
Sparse-MLP-S [26]49M10.3G36183.1
Wave-MLP-M [27]44M7.9G41383.4
Strip-MLP-S(ours)43M6.8G38183.3
ResMLP-B24 [29]116M23.0G23181.0
gMLP-B [16]73M15.8G-81.6
CycleMLP-B5 [1]76M12.3G24783.2
Sparse-MLP-B [26]66M14.0G27883.4
Hire-MLP-Base [7]58M8.1G44183.2
Wave-MLP-B [27]63M10.2G34183.6
Strip-MLP-B(ours)57M9.2G30083.6
", + "image_path": "44287b98728f96476dbeec2c03b3fd2075e323c115ea4c920a297164b721dd61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 456, + 287, + 489 + ], + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 489 + ], + "type": "text", + "content": "Table 6. Classification results of three kinds of models on ImageNet-1K without extra data. The group number of CGSMM is C/4. Throughput is tested on a single V100 GPU following [19]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 497, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 592 + ], + "type": "text", + "content": "MLP-based models. In addition, we apply the Strip MLP layer into Wave-MLP-T (namely Wave-Strip-MLP-T) on ImageNet-1K, and we set the strip width as 3. The Wave-Strip-MLP-T gets better results than the original Wave-MLP-T model: " + }, + { + "bbox": [ + 46, + 497, + 287, + 592 + ], + "type": "inline_equation", + "content": "81.2\\%" + }, + { + "bbox": [ + 46, + 497, + 287, + 592 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 497, + 287, + 592 + ], + "type": "inline_equation", + "content": "80.6\\%" + }, + { + "bbox": [ + 46, + 497, + 287, + 592 + ], + "type": "text", + "content": ". The performance improvement by " + }, + { + "bbox": [ + 46, + 497, + 287, + 592 + ], + "type": "inline_equation", + "content": "+0.6\\%" + }, + { + "bbox": [ + 46, + 497, + 287, + 592 + ], + "type": "text", + "content": " shows that our Strip MLP makes token interaction more efficient and can easily be used in other MLP models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 654 + ], + "type": "text", + "content": "In conclusion, as shown in Fig. 4, our Strip-MLP has significant superiority in performance and complexity over other state-of-the-art methods on different datasets, demonstrating our Strip-MLP is remarkably efficient and effective to enrich the token interaction power for vision MLP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 662, + 148, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 148, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 148, + 674 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "To better show the effectiveness of the proposed method, we ablate the key components of the model design. Due to the limited GPU resources, we do ablation studies on the" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 162 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 162 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 162 + ], + "type": "table", + "html": "
Strip WidthDatasetParamsFLOPsTop-1(%)
141.33M6.71G91.38
3Caltech-10143.66M6.83G92.09
545.99M6.96G91.72
748.33M7.08G91.86
140.71M6.71G86.09
3CIFAR-10041.88M6.83G86.50
543.04M6.96G86.39
744.21M7.08G86.35
", + "image_path": "c088981642af2026df2d272c7aee33f19deb036828e817a08e2e91f9d538d058.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 162, + 545, + 193 + ], + "lines": [ + { + "bbox": [ + 306, + 162, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 162, + 545, + 193 + ], + "type": "text", + "content": "Table 7. Ablation study on the strip width of Strip MLP. We randomly select the model of Strip-MLP-S for Caltech-101 and Strip-MLP-S for CIFAR-100." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 194, + 545, + 301 + ], + "blocks": [ + { + "bbox": [ + 307, + 194, + 545, + 301 + ], + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 301 + ], + "type": "table", + "html": "
DatasetPatchesParamsFLOPsTop-1(%)
Caltech-101C/147.22M6.83G91.81
C/243.66M6.83G92.09
C/441.88M6.83G91.38
C/840.99M6.83G91.33
140.16M6.83G90.93
CIFAR-100C/147.22M6.83G86.20
C/243.66M6.83G86.18
C/441.88M6.83G86.50
C/840.99M6.83G85.98
140.16M6.83G85.03
", + "image_path": "7663e7ce023765890bb68c9d02f23eb87b4b538fe610b92a88c3845969584287.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 301, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 306, + 301, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 545, + 334 + ], + "type": "text", + "content": "Table 8. Ablation study about the patch number of the CGSMM. The base model is Strip-MLP-S. " + }, + { + "bbox": [ + 306, + 301, + 545, + 334 + ], + "type": "inline_equation", + "content": "C/1" + }, + { + "bbox": [ + 306, + 301, + 545, + 334 + ], + "type": "text", + "content": " means splitting the feature " + }, + { + "bbox": [ + 306, + 301, + 545, + 334 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 306, + 301, + 545, + 334 + ], + "type": "text", + "content": " into C patches in channel dimension, namely " + }, + { + "bbox": [ + 306, + 301, + 545, + 334 + ], + "type": "inline_equation", + "content": "P = C/1 = C" + }, + { + "bbox": [ + 306, + 301, + 545, + 334 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 342, + 492, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 492, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 492, + 353 + ], + "type": "text", + "content": "datasets of both Caltech-101 and CIFAR-100." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 497 + ], + "type": "text", + "content": "The effect of Strip Width in Strip MLP. The strip width affects the token interaction ranges and determines the field of each row or column's contribution to adjacent tokens. We further conduct experiments to verify its effect by varying the width from 1 to 7 with a step at 2. Tab. 7 shows the performances of our models on Caltech-101 [6] and CIFAR-100 [14], respectively. When the width becomes larger, the performance increases and tends to saturate, indicating that Strip MLP layer improves the token interaction power. In all of our comparison experiments in previous sections, we consistently set the strip width as 3 according to the ablation experiment results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "content": "Effects of Patch Number in CGSMM. Different patch number brings varying degrees of token interaction power improvement and affects the model performance. In Tab. 8, we design five decreased patch numbers from C into 1 to show the effectiveness of CGSMM. Without the patch splitting " + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "inline_equation", + "content": "(P = 1)" + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "content": " and GSML operation, the model performance decreases by " + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "inline_equation", + "content": "1.16\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "content": " on Caltech-101 " + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "inline_equation", + "content": "(92.09\\% \\rightarrow 90.93\\%)" + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "content": " and decreases by " + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "inline_equation", + "content": "1.47\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "content": " on CIFAR-100 " + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "inline_equation", + "content": "(86.50\\% \\rightarrow 85.03\\%)" + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "content": ", which consistently proves the effectiveness of CGSMM in improving the token interaction power. From the experimental results, we can observe that the optimal patch number is different on the datasets with scale differences, so the optimal patch number should be determined through validation experiments. We consistently set the patch number as C/4 in other ablation studies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Cascade vs. Parallel Architecture of GSML. Applying the GSML in a cascade architecture enables the token interacts with other tokens across the entire 2D space in just one" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1501" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 77, + 176, + 171 + ], + "blocks": [ + { + "bbox": [ + 59, + 77, + 176, + 171 + ], + "lines": [ + { + "bbox": [ + 59, + 77, + 176, + 171 + ], + "spans": [ + { + "bbox": [ + 59, + 77, + 176, + 171 + ], + "type": "image", + "image_path": "da2a372e6dd3a92b01b72bb3ae18d45f3cffe811b555ccc3b34dc7672c0eb6f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 178, + 119, + 186 + ], + "lines": [ + { + "bbox": [ + 108, + 178, + 119, + 186 + ], + "spans": [ + { + "bbox": [ + 108, + 178, + 119, + 186 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 76, + 294, + 171 + ], + "blocks": [ + { + "bbox": [ + 177, + 76, + 294, + 171 + ], + "lines": [ + { + "bbox": [ + 177, + 76, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 177, + 76, + 294, + 171 + ], + "type": "image", + "image_path": "c6ee3e3a0618b95346da772b4215bffd339d7a08937fc9b6fdb4dbb75e7f2365.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 177, + 236, + 186 + ], + "lines": [ + { + "bbox": [ + 225, + 177, + 236, + 186 + ], + "spans": [ + { + "bbox": [ + 225, + 177, + 236, + 186 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 296, + 76, + 414, + 171 + ], + "blocks": [ + { + "bbox": [ + 296, + 76, + 414, + 171 + ], + "lines": [ + { + "bbox": [ + 296, + 76, + 414, + 171 + ], + "spans": [ + { + "bbox": [ + 296, + 76, + 414, + 171 + ], + "type": "image", + "image_path": "c3f1c9808432c7e150dc2405f6bfff2ce1f15998572a6fcc8260daa8ee2cbb44.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 177, + 356, + 186 + ], + "lines": [ + { + "bbox": [ + 345, + 177, + 356, + 186 + ], + "spans": [ + { + "bbox": [ + 345, + 177, + 356, + 186 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 416, + 76, + 535, + 171 + ], + "blocks": [ + { + "bbox": [ + 416, + 76, + 535, + 171 + ], + "lines": [ + { + "bbox": [ + 416, + 76, + 535, + 171 + ], + "spans": [ + { + "bbox": [ + 416, + 76, + 535, + 171 + ], + "type": "image", + "image_path": "afdfca1c88a0a22a5aa7e35e65849535c686b829c3d08eb76305d1d5c7816207.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 177, + 479, + 186 + ], + "lines": [ + { + "bbox": [ + 468, + 177, + 479, + 186 + ], + "spans": [ + { + "bbox": [ + 468, + 177, + 479, + 186 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 198, + 178, + 290 + ], + "blocks": [ + { + "bbox": [ + 58, + 198, + 178, + 290 + ], + "lines": [ + { + "bbox": [ + 58, + 198, + 178, + 290 + ], + "spans": [ + { + "bbox": [ + 58, + 198, + 178, + 290 + ], + "type": "image", + "image_path": "fcdc215a8d20a99fc4b69d5a13d78fbff2f79a03d09bcd8fd6899bd9ff484147.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 297, + 119, + 306 + ], + "lines": [ + { + "bbox": [ + 108, + 297, + 119, + 306 + ], + "spans": [ + { + "bbox": [ + 108, + 297, + 119, + 306 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 179, + 198, + 298, + 290 + ], + "blocks": [ + { + "bbox": [ + 179, + 198, + 298, + 290 + ], + "lines": [ + { + "bbox": [ + 179, + 198, + 298, + 290 + ], + "spans": [ + { + "bbox": [ + 179, + 198, + 298, + 290 + ], + "type": "image", + "image_path": "2acffc65d5fb5f26ae4119e7494d70c001a34684da9aa66b1cd924dd772912c6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 297, + 238, + 306 + ], + "lines": [ + { + "bbox": [ + 228, + 297, + 238, + 306 + ], + "spans": [ + { + "bbox": [ + 228, + 297, + 238, + 306 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 299, + 196, + 417, + 290 + ], + "blocks": [ + { + "bbox": [ + 299, + 196, + 417, + 290 + ], + "lines": [ + { + "bbox": [ + 299, + 196, + 417, + 290 + ], + "spans": [ + { + "bbox": [ + 299, + 196, + 417, + 290 + ], + "type": "image", + "image_path": "5586cd5f719801abf8b86ece3cb56df4636167654ed88d3b79555aff15605919.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 297, + 358, + 306 + ], + "lines": [ + { + "bbox": [ + 348, + 297, + 358, + 306 + ], + "spans": [ + { + "bbox": [ + 348, + 297, + 358, + 306 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 309, + 544, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 309, + 544, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 544, + 331 + ], + "type": "text", + "content": "Figure 4. Performance comparison on four datasets of ImageNet-1K [4], CIFAR-100 [14], CIFAR-10 [14], and Caltech-101 [6]. The Top-1 accuracy, number of parameters and FLOPs are reported with networks of CNN-based, Transformer-based, and MLP-based models." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 417, + 196, + 534, + 290 + ], + "blocks": [ + { + "bbox": [ + 417, + 196, + 534, + 290 + ], + "lines": [ + { + "bbox": [ + 417, + 196, + 534, + 290 + ], + "spans": [ + { + "bbox": [ + 417, + 196, + 534, + 290 + ], + "type": "image", + "image_path": "ac8ff3de25819cc0eb6500cda587101794a58c3d13e72c45115709614a9b15a2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 297, + 479, + 306 + ], + "lines": [ + { + "bbox": [ + 468, + 297, + 479, + 306 + ], + "spans": [ + { + "bbox": [ + 468, + 297, + 479, + 306 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 48, + 341, + 288, + 399 + ], + "blocks": [ + { + "bbox": [ + 48, + 341, + 288, + 399 + ], + "lines": [ + { + "bbox": [ + 48, + 341, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 341, + 288, + 399 + ], + "type": "table", + "html": "
MethodDatasetParamsFLOPsTop-1(%)
Parallel CascadeCaltech-10142.67M6.65G91.61
43.66M6.83G92.09
Parallel CascadeCIFAR-10040.90M6.66G85.89
41.88M6.83G86.50
", + "image_path": "1f3769d8ab40f0ea65c19ed2a3e65e02f96aa341c29f7cf9137bb31505c91711.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 399, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 287, + 443 + ], + "type": "text", + "content": "Table 9. Ablation study on the cascade and parallel structure of the Strip MLP in Group Strip MLP. Parallel means applying the Strip MLP on the row and column direction in parallel. Cascade means the proposed CGSMM method in this paper." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 451, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 546 + ], + "type": "text", + "content": "module, which needs two modules for the parallel structure so that may decrease the efficiency of token interaction. In Tab. 9, we test the effects between the cascade structure of CGSMM and the parallel structure of Parallel Group Strip Mixing Module (PGSMM). Our experiments show that the cascade structure gets a higher accuracy compared to the parallel structure, with an increase of " + }, + { + "bbox": [ + 46, + 451, + 287, + 546 + ], + "type": "inline_equation", + "content": "+0.48\\%" + }, + { + "bbox": [ + 46, + 451, + 287, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 451, + 287, + 546 + ], + "type": "inline_equation", + "content": "+0.61\\%" + }, + { + "bbox": [ + 46, + 451, + 287, + 546 + ], + "type": "text", + "content": " on the Caltech-101 and CIFAR-100 datasets, respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 547, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 287, + 655 + ], + "type": "text", + "content": "Roles of CGSMM & LSMM. To verify the significance of two complementary branches, we conduct experiments where we only keep one of the branches of CGSMM or LSMM. Tab. 10 displays the outcomes of our trial on Caltech-101 and CIFAR-100. Notably, we found that removing any module from either of the two branches significantly reduced the model's performance. These results highlight the essential role of both CGSMM and LSMM in enriching the tokens' interaction power." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "This paper presents an efficient and effective token interaction MLP model of Strip-MLP for vision MLP. For the" + } + ] + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 306, + 341, + 548, + 418 + ], + "blocks": [ + { + "bbox": [ + 306, + 341, + 548, + 418 + ], + "lines": [ + { + "bbox": [ + 306, + 341, + 548, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 548, + 418 + ], + "type": "table", + "html": "
MethodDatasetParamsFLOPsTop-1(%)
CGSMM OnlyCaltech-10143.40M7.55G91.10
LSMM Only43.96M6.00G90.31
CGSMM & LSMM41.88M6.83G91.38
CGSMM OnlyCIFAR-10043.40M7.55G85.20
LSMM Only43.96M6.00G86.13
CGSMM & LSMM41.88M6.83G86.50
", + "image_path": "c45a9ea44b4206243170035fe9f110306809664613d6132de528ed4cab9e798c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 418, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 545, + 461 + ], + "type": "text", + "content": "Table 10. Ablation study on the roles of two branches of CGSMM and LSMM. The base model is Strip-MLP-S. When any module branch has been removed, all input features would be fed into the remaining branch module." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 545, + 632 + ], + "type": "text", + "content": "Token's interaction dilemma problem, we design the Strip Mixing Block to enrich the token interaction power by three strategies. Our experimental analysis reveals that the StripMLP remarkably improves the performances of the MLP-based models on small datasets and keeps comparable or even better results on ImageNet with great superiorities on the number of parameters and FLOPs. These observations clearly demonstrate that our models make token interaction more efficient among MLP-based models in image classification tasks. It is expected that Strip MLP layer has the potential to be a standard layer in the future and have promising performance on diverse vision tasks." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement: This work is supported by National Key Research and Development Program of China (2021YFF1200800); Peng Cheng Laboratory Research Project (PCL2023AS6-1); and Guangdong Basic and Applied Basic Research Foundation under Grant 2022A1515110573." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1502" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Shoufa Chen, Enze Xie, Chongjian Ge, Ding Liang, and Ping Luo. Cycle.mp: A mlp-like architecture for dense prediction. In International Conference on Learning Representation (ICLR), Oral, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 134, + 287, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 134, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 134, + 287, + 177 + ], + "type": "text", + "content": "[2] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 177, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 177, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 177, + 288, + 220 + ], + "type": "text", + "content": "[3] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), volume 1, pages 886-893. IEEE, 2005." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 262 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 328 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 328, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 288, + 360 + ], + "type": "text", + "content": "[6] Li Fei-Fei, Robert Fergus, and Pietro Perona. One-shot learning of object categories. IEEE transactions on pattern analysis and machine intelligence, 28(4):594-611, 2006." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 360, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 288, + 414 + ], + "type": "text", + "content": "[7] Jianyuan Guo, Yehui Tang, Kai Han, Xinghao Chen, Han Wu, Chao Xu, Chang Xu, and Yunhe Wang. Hire-mlp: Vision mlp via hierarchical rearrangement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 826-836, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 456, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 288, + 477 + ], + "type": "text", + "content": "[9] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 477, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 529 + ], + "type": "text", + "content": "[10] Qibin Hou, Zihang Jiang, Li Yuan, Ming-Ming Cheng, Shuicheng Yan, and Jiashi Feng. Vision permutator: A permutable mlp-like architecture for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 531, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 585 + ], + "type": "text", + "content": "[11] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 585, + 287, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 585, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 585, + 287, + 628 + ], + "type": "text", + "content": "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 628, + 287, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 628, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 48, + 628, + 287, + 659 + ], + "type": "text", + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 659, + 287, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 659, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 48, + 659, + 287, + 681 + ], + "type": "text", + "content": "[14] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[15] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Hanxiao Liu, Zihang Dai, David So, and Quoc V Le. Pay attention to mlps. Advances in Neural Information Processing Systems, 34:9204-9215, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 106, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 137 + ], + "type": "text", + "content": "[17] Ruiyang Liu, Yinghui Li, Linmi Tao, Dun Liang, and Haitao Zheng. Are we ready for a new paradigm shift? a survey on visual deep mlp. Patterns, 3(7):100520, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 137, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 137, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 545, + 191 + ], + "type": "text", + "content": "[18] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 191, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 191, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 545, + 245 + ], + "type": "text", + "content": "[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 245, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 298 + ], + "type": "text", + "content": "[20] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 299, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 354 + ], + "type": "text", + "content": "[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 354, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 406 + ], + "type": "text", + "content": "[22] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 406, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 438 + ], + "type": "text", + "content": "[23] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 438, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 492 + ], + "type": "text", + "content": "[24] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 492, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 534 + ], + "type": "text", + "content": "[25] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 535, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 589 + ], + "type": "text", + "content": "[26] Chuanxin Tang, Yucheng Zhao, Guangting Wang, Chong Luo, Wenxuan Xie, and Wenjun Zeng. Sparse mlp for image recognition: is self-attention really necessary? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2344–2351, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "type": "text", + "content": "[27] Yehui Tang, Kai Han, Jianyuan Guo, Chang Xu, Yanxi Li, Chao Xu, and Yunhe Wang. An image patch is a wave: Phase-aware vision mlp. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10935-10944, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 643, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 643, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 307, + 643, + 545, + 707 + ], + "type": "text", + "content": "[28] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in Neural Information Processing Systems, 34:24261-24272, 2021." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1503" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 709 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[29] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Gautier Izacard, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, et al. Resmlp: Feedforward networks for image classification with data-efficient training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 138, + 287, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 138, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 48, + 138, + 287, + 192 + ], + "type": "text", + "content": "[30] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 192, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 192, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 48, + 192, + 288, + 246 + ], + "type": "text", + "content": "[31] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5769-5780, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 246, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 246, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 246, + 287, + 289 + ], + "type": "text", + "content": "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 289, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 289, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 287, + 353 + ], + "type": "text", + "content": "[33] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 353, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 418 + ], + "type": "text", + "content": "[34] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 418, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 287, + 461 + ], + "type": "text", + "content": "[35] Ziyu Wang, Wenhao Jiang, Yiming M Zhu, Li Yuan, Yibing Song, and Wei Liu. Dynamixer: a vision mlp architecture with dynamic mixing. In International Conference on Machine Learning, pages 22691-22701. PMLR, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 461, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 461, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 461, + 287, + 505 + ], + "type": "text", + "content": "[36] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 505, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 505, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 505, + 287, + 568 + ], + "type": "text", + "content": "[37] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "text", + "content": "[38] David Junhao Zhang, Kunchang Li, Yunpeng Chen, Yali Wang, Shashwat Chandra, Yu Qiao, Luoqi Liu, and Mike Zheng Shou. Morphmlp: A self-attention free, mlp-like backbone for image and video. arXiv preprint arXiv:2111.12527, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 622, + 287, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 675 + ], + "type": "text", + "content": "[39] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 675, + 287, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 675, + 287, + 709 + ], + "spans": [ + { + "bbox": [ + 48, + 675, + 287, + 709 + ], + "type": "text", + "content": "[40] Yipeng Zhang, Zahra M Aghajan, Matias Ison, Qiujing Lu, Hanlin Tang, Guldamla Kalender, Tonmoy Monsoor, Jie Zheng, Gabriel Kreiman, Vwani Roychowdhury, et al. De" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "content": "coding of human identity by computer vision and neuronal vision. Scientific reports, 13(1):651, 2023." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1504" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Strivec_ Sparse Tri-Vector Radiance Fields/35b59dcb-1ce5-40a4-83d2-07192fb62870_content_list.json b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/35b59dcb-1ce5-40a4-83d2-07192fb62870_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cb54057b90c1c4e14fc9a519d5b4874e695dcb8b --- /dev/null +++ b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/35b59dcb-1ce5-40a4-83d2-07192fb62870_content_list.json @@ -0,0 +1,1825 @@ +[ + { + "type": "text", + "text": "Strivec: Sparse Tri-Vector Radiance Fields", + "text_level": 1, + "bbox": [ + 267, + 130, + 702, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Quankai Gao\\*1", + "bbox": [ + 122, + 179, + 243, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiangeng Xu*1", + "bbox": [ + 285, + 180, + 406, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Southern California", + "bbox": [ + 127, + 198, + 401, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Su²", + "bbox": [ + 449, + 180, + 519, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ulrich Neumann1", + "bbox": [ + 560, + 180, + 699, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ UC San Diego", + "bbox": [ + 491, + 198, + 616, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zexiang $\\mathbf{X}\\mathbf{u}^3$", + "bbox": [ + 741, + 180, + 844, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Adobe Research", + "bbox": [ + 700, + 198, + 841, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{quankaig, qiangenx, uneumann}@usc.edu", + "bbox": [ + 174, + 219, + 491, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "haosu@ucsd.edu", + "bbox": [ + 516, + 219, + 643, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zexu@adobe.com", + "bbox": [ + 669, + 219, + 795, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose Strivec, a novel neural representation that models a 3D scene as a radiance field with sparsely distributed and compactly factorized local tensor feature grids. Our approach leverages tensor decomposition, following the recent work TensoRF [6], to model the tensor grids. In contrast to TensoRF which uses a global tensor and focuses on their vector-matrix decomposition, we propose to utilize a cloud of local tensors and apply the classic CANDE-COMP/PARAFAC (CP) decomposition [4] to factorize each tensor into triple vectors that express local feature distributions along spatial axes and compactly encode a local neural field. We also apply multi-scale tensor grids to discover the geometry and appearance commonalities and exploit spatial coherence with the tri-vector factorization at multiple local scales. The final radiance field properties are regressed by aggregating neural features from multiple local tensors across all scales. Our tri-vector tensors are sparsely distributed around the actual scene surface, discovered by a fast coarse reconstruction, leveraging the sparsity of a 3D scene. We demonstrate that our model can achieve better rendering quality while using significantly fewer parameters than previous methods, including TensoRF and Instant-NGP [23].", + "bbox": [ + 75, + 300, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 675, + 209, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Representing 3D scenes as radiance fields [22] has enabled photo-realistic rendering quality and emerged as a popular design choice in 3D vision and graphics applications. While many methods [27, 42, 3] (following NeRF [22]) purely use MLPs to represent neural fields, recent works, like TensorRF [6] and Instant-NGP [23], have demonstrated the advantages of using shared global feature encoding for radiance field modeling, in terms of speed, compactness, and quality. However, these methods share and assign neural features uniformly in a scene (with tensor factors or hash tables), assuming the scene content is", + "bbox": [ + 75, + 700, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/641a53b3096cce5fbd019327af205a915e6d635ed962cb19b3599d951e341efc.jpg", + "image_caption": [ + "TensoRF-VM-192" + ], + "image_footnote": [], + "bbox": [ + 509, + 266, + 599, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9248d0b165cc49ce9fd12f35ab0f38afb3a64f213adf6a2acffc1ed73378d83.jpg", + "image_caption": [ + "INGP" + ], + "image_footnote": [], + "bbox": [ + 604, + 267, + 696, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9874ff302c4eaddfa71b9f908f3d0f669e9a9f69efe39d6d861d2a5e281b7859.jpg", + "image_caption": [ + "Ours-48" + ], + "image_footnote": [], + "bbox": [ + 700, + 267, + 790, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21c797888d26fac8ebf3e801526d618fcfb0186bac9b60e86fd494054c85ae9d.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 795, + 267, + 885, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c828c2bae543459a9bb5ef72640bff7345c2444a151c9a411940f55a7ed31f71.jpg", + "image_caption": [ + "Figure 1: We compare with previous methods in terms of rendering quality (PSNR) and model capacity (number of parameters) on the NeRF Synthetic dataset on the bottom. Our method and TensorRF are shown with different model sizes. Our approach consistently achieve better rendering quality with fewer model parameters than TensorRF, as well as other methods like iNGP. On the top, we show one example of visual comparisons of the mic scene that has challenging fine-grained geometric structures, where our approach captures most of the details and is the closest to the reference. Note that the results of NeRF and Point-NeRF use 200k optimization steps while the rest use only 30k steps." + ], + "image_footnote": [], + "bbox": [ + 504, + 367, + 890, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "equally complex over the entire space, which can be inefficient (requiring high model capacity) to accurately model intricate local scene details (see Fig.1).", + "bbox": [ + 496, + 775, + 892, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We aim to accurately and compactly model a 3D scene and reproduce the complex local details. To this end, we propose Strivec, a novel neural scene representation that utilizes sparsely distributed and compactly factorized local tensor grids to model a volumetric radiance field for high-", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 98, + 875, + 212, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Code and results: https://github.com/Zerg-Overmind/Strivec", + "bbox": [ + 91, + 887, + 408, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17569", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quality novel view synthesis. As shown in Fig.1, our approach is able to accurately model the complex scene structures that are not recovered well by previous methods. More importantly, our superior rendering quality is achieved with much less model capacity.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, we base our model on TensoRF [6], a recent approach that leverages tensor factorization in radiance field modeling. It is fast, compact, and of high rendering quality. TensoRF applies CP and vector-matrix (VM) decomposition techniques to factorize a field into vectors and matrices and model the entire scene as a global factorized tensor. Instead of a single global tensor, we leverage a sparse set of multiple small local tensors distributed around the scene surface for more efficient scene modeling. Specifically, each of our tensors represents a local radiance field inside its local bounding box and is compactly modeled with factorized triple vectors based on the CP decomposition.", + "bbox": [ + 75, + 169, + 468, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that the global CP decomposition in TensoRF has led to a highly compact model but cannot achieve comparable rendering quality to their VM decomposition. This is because a tri-vector CP component is rank-one, while a global feature grid of an entire 3D scene is often complex and of high rank, requiring a large (impractical) number of CP components for high accuracy. TensoRF addresses this by introducing matrix factors in their VM decomposition, essentially increasing the rank of each tensor component. Our model instead consists of multiple small tensor grids, exploiting local spatial commonalities in a scene. Compared to a global tensor, our local tensor is less complex and of much lower rank, thus effectively reducing the required number of CP components (per tensor) and enabling practical high-quality radiance field reconstruction with highly compact tri-vector factors. Our local tri-vector tensors can lead to superior rendering quality and compactness over TensoRF's VM model (see Fig. 1). We also observe that our local tensors are generally more robust than a global tensor against the orientation of spatial axes (which can affect the rank of a tensor and thus affects the quality; see Fig. 2).", + "bbox": [ + 75, + 353, + 470, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Importantly, adopting local tensors (instead of a global one) also brings us the flexibility to allocate neural features according to the actual scene distribution, enabling more efficient scene modeling and better usage of model parameters than a global representation. To do so, we pre-acquire coarse scene geometry – that can be easily achieved via a fast RGB $\\sigma$ volume reconstruction (like DVGO [32]) or multi-view stereo (like Point-NeRF [39]) – to directly distribute local tensors around the actual scene surface, leading to a sparse scene representation that avoids unnecessarily modeling the empty scene space. Note that while previous methods have also leveraged sparse representations (with voxels [19, 41] or points [39]) of radiance fields, their local features are modeled and optimized independently. Our model instead correlates a group of local features inside a", + "bbox": [ + 75, + 674, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "local box and compactly express them with triple vectors, uniquely exploiting the local spatial coherence along axes and imposing local low-rank priors in the feature encoding via tensor factorization. Moreover, unlike previous sparse representations that only use a single-scale feature grid or point cloud, we distribute multi-scale local tensors to effectively model the scene geometry and appearance at multiple scales in a hierarchical manner. In particular, for an arbitrary 3D location, we aggregate the neural features from its neighboring tri-vector components at all scales and decode the volume density and view-dependent color from the aggregated features for radiance field rendering.", + "bbox": [ + 496, + 90, + 890, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach takes the best of previous local and global radiance field representations. Compared with global representations like TensoRF and Instant-NGP, our model takes advantage of the sparsity of a scene more directly; compared with local representations like Plenoxels and Point-NeRF, our model makes use of the local smoothness and coherence of scene geometry and appearance. As shown in our experimental results on both synthetic and real datasets, our model is able to achieve state-of-the-art rendering quality on these datasets, outperforming previous methods, including TensoRF and Instant-NGP, while using significantly fewer model parameters, demonstrating the superior representational power of our model.", + "bbox": [ + 496, + 272, + 890, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 479, + 640, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene representations. To represent a 3D scene, traditional and learning-based methods have studied various representations, such as depth map [13, 18], mesh [15, 36, 30], point cloud [28, 1, 35] and implicit function [7, 21, 24, 40]. In recent years, continuous neural field representations stand out in various 3D tasks such as single-view 3D reconstruction [38, 11], surface completion [8, 26], multi-view reconstruction [24] and novel view synthesis [22, 20]. Compared with traditional discrete representations, a continuous field have no limitation on spatial resolution, e.g., volume resolution or the number of points. It can also naturally be represented by neural networks, such as an MLP, which are known for approximating complex functions well.", + "bbox": [ + 496, + 505, + 890, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural field representations. Specifically, NeRF [22] represents a 3D scene as a radiance field with a global coordinate MLP, which models geometry, lighting and texture information jointly, leading to photo-realistic rendering quality in novel view synthesis. Apart from its advantage, purely MLP-based NeRF models [3, 34] in general suffer from inefficiency [2] when modeling highly complex or large-scale scenes, due to limited model capacity, slow optimization speed, and the cost of modeling empty space.", + "bbox": [ + 496, + 719, + 890, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To model radiance fields more efficiently, recent works have explored combining neural fields with various traditional 3D representations, including voxels [19, 41, 32, 43]", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17570", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and points [39]. Low-rank representations such as triplane [5, 10] and tensor decomposition [6, 25] have also been studied. In particular, DVGO [32] and Plenoxels [41] respectively use dense and sparse voxels with neural features for radiance field modeling. While being efficient to optimize, these localized feature grid-based representations lead to a large model size and can face overfitting issues when the features are of very high resolution. Consequently, DVGO can also work with a low-resolution grid and Plenoxels requires additional spatial regularization terms. On the other hand, recent works have adopted global feature encoding to express a high-resolution feature grid, including Instant-NGP [23] that hashes spatial features into multi-scale hash tables and TensorRF [6] that factorizes a feature grid into vector and matrix factors. These global feature encoding methods exploit the spatial correlation across the entire scene space, leading to fast and compact reconstruction and surpassing previous MLP-based or grid-based representations on rendering quality. However, similar to NeRF, such global representation can also be limited by its model capacity when representing highly complex or large-scale content.", + "bbox": [ + 75, + 90, + 472, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach instead combines local and global representations. Our tri-vector fields are sparsely distributed in the scene, similar to local representations (like plenoxels and Point-NeRF); meanwhile, features in each field are represented by tri-vector components shared across the local region as done in TensoRF, exploiting spatial feature commonalities. Our model leverages both spatial sparsity and coherence, leading to much higher compactness and better reconstruction quality than previous local and global representations (see Tab. 1).", + "bbox": [ + 75, + 429, + 468, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relevant to our work, previous methods, such as KiloNeRF [29] and BlockNeRF [33] have also utilized multiple local MLPs to represent a scene. Specifically, KiloNeRF focuses and speeding up NeRF and their rendering quality is sacrificed; BlockNeRF essentially uses multiple NeRFs to increase the total model capacity. Instead of pure MLPs, our work is built upon tensor factorization-based feature encoding as done in TensoRF [6], and we in fact achieve superior rendering quality while decreasing the model capacity.", + "bbox": [ + 75, + 587, + 470, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Sparse Tri-Vector Field Representation", + "text_level": 1, + "bbox": [ + 76, + 756, + 429, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now present our novel radiance field representation. In essence, our model consists of a cloud of small local trivector tensors at multiple scales, designed to leverage both sparsity and multi-scale spatial coherence (see Fig. 2).", + "bbox": [ + 75, + 787, + 468, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{T} = \\{\\tau_n|n = 1,\\dots,N\\}$ denote a cloud of tri-vector tensors. Each local tensor $\\tau$ is located at $p$ , covering a local cuboid space $\\omega$ with an edge length of $l$ . This cloud of tri", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vector tensors represents a radiance field for the 3D space:", + "bbox": [ + 500, + 90, + 883, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega = \\bigcup_ {n = 1} ^ {N} \\omega_ {n}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 117, + 890, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, each tensor $\\tau$ encodes a local multi-channel feature grid that includes a (single-channel) density grid $A_{\\sigma}$ and a (multi-channel) appearance grid $A_{c}$ , similar to the tensor grid in TensoRF [6]. In contrast to using a single global tensor in TensoRF [6], we model the volume density and view-dependent colors with multiple local tensors. In particular, for an arbitrary location $\\chi \\in \\Omega$ , we select $M$ nearest tensors that cover $\\chi$ . Across the selected tensors, we aggregate the extracted density and appearance features recovered by their tri-vector factors for radiance field property regression, where the volume density $\\sigma$ is directly obtained after the aggregation and the view-dependent color $c$ is regressed by a small MLP $\\psi$ along with the viewing direction $\\mathbf{d}$ . The continuous radiance field can be expressed as:", + "bbox": [ + 496, + 169, + 890, + 381 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {\\chi}, c _ {\\chi} = A _ {\\sigma} \\left(\\left\\{\\mathcal {G} ^ {\\sigma} (\\chi) \\right\\}\\right), \\psi \\left(A _ {c} \\left(\\left\\{\\mathcal {G} ^ {c} (\\chi) \\right\\}\\right), \\mathbf {d}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 393, + 890, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Local tri-vector tensors.", + "text_level": 1, + "bbox": [ + 500, + 421, + 718, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We apply the classic Canonical polyadic (CP) decomposition [4] to model our local tensors with tri-vector components.", + "bbox": [ + 496, + 444, + 890, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CP decomposition. CP decomposition factorizes a $M$ dimension tensor $\\tau \\in \\mathbb{R}^{I_1\\times I_2\\times \\ldots \\times I_M}$ into a linear combination of $R$ rank-1 tensors:", + "bbox": [ + 496, + 508, + 890, + 553 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tau = \\sum_ {r = 1} ^ {R} \\lambda_ {r} \\mathbf {v} _ {r} ^ {0} \\otimes \\mathbf {v} _ {r} ^ {1} \\otimes \\dots , \\otimes \\mathbf {v} _ {r} ^ {M}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 563, + 890, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes outer product; the weighting factor $\\lambda_r$ can be absorbed into vectors $\\{\\mathbf{v}_r^0,\\dots ,\\mathbf{v}_r^M\\}$ .", + "bbox": [ + 496, + 614, + 890, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Density and appearance tensors. In our case of modeling a 3D radiance field, we set the geometry grid $\\mathcal{G}^{\\sigma} \\in \\mathbb{R}^{I \\times J \\times K}$ as a 3D tensor. And the multi-channel appearance grid $\\mathcal{G}^c \\in \\mathbb{R}^{I \\times J \\times K \\times P}$ corresponds to a 4D tensor. The fourth appearance mode is of lower dimension (compared with the spatial modes), representing the final dimension of the features sent to the MLP decoder network.", + "bbox": [ + 496, + 664, + 890, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to Eqn.3, we factorize each tensor's feature grids, $\\mathcal{G}^{\\sigma}$ and $\\mathcal{G}^{c}$ , by CP decomposition:", + "bbox": [ + 496, + 770, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} ^ {\\sigma} = \\sum_ {r = 1} ^ {R _ {\\sigma}} \\mathcal {A} _ {\\sigma , r} = \\sum_ {r = 1} ^ {R _ {\\sigma}} \\mathbf {v} _ {\\sigma , r} ^ {X} \\otimes \\mathbf {v} _ {\\sigma , r} ^ {Y} \\otimes \\mathbf {v} _ {\\sigma , r} ^ {Z}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 811, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} ^ {c} = \\sum_ {r = 1} ^ {R _ {c}} \\mathcal {A} _ {c, r} \\otimes \\mathbf {b} _ {r} = \\sum_ {r = 1} ^ {R _ {c}} \\mathbf {v} _ {c, r} ^ {X} \\otimes \\mathbf {v} _ {c, r} ^ {Y} \\otimes \\mathbf {v} _ {c, r} ^ {Z} \\otimes \\mathbf {b} _ {r}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 856, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17571", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/14f89b20e162f2eae18ac0949108c925d2dc223ca3550182ad5e2a226d8d28f7.jpg", + "image_caption": [ + "Figure 2: Overview of our Sparse Tri-Vector Radiance Fields. We distribute our local tensors based on a coarse geometry estimated by a fast $\\mathrm{RGB}\\sigma$ volume reconstruction as done in DVGO [32]. Here, we show our model running under $S = 2$ scales. Each local tensor is factorized as axis-aligned triple based on CP decomposition. For any shading point $\\chi$ , we extract and evaluate features in each local tensor, according to the factorization (Sec. 3.1). Then, we aggregate these features among nearby tensors (Sec. 3.2) and across different scales (Sec. 3.3). Finally, the density and color are decoded (Sec. 3.4) and used by volume rendering (Sec.4)." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 893, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $R_{\\sigma}$ and $R_{c}$ denote numbers of component; $\\mathcal{A}_{\\sigma,r}$ and $\\mathcal{A}_{c,r}$ are the component tensors that are factorized spatially; $\\mathbf{v}_{\\sigma,r}^{X}, \\dots, \\mathbf{v}_{c,r}^{X}, \\dots$ are the 1D vectors with resolution $I, J, K$ , modeling scene geometry and appearance along $X, Y, Z$ axis; $R_{\\sigma}$ and $R_{c}$ are the component numbers; $\\mathbf{b}_r$ expresses the feature dimension.", + "bbox": [ + 76, + 428, + 468, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As done in TensoRF [6], we stack all feature-mode vectors $\\mathbf{b}_r$ as columns together, which ends up a $P\\times R_{c}$ appearance matrix $\\mathbf{B}$ . This matrix models the appearance feature variations of the tensor and functions like a appearance dictionary. Note that naively following CP decomposition like TensoRF will assign a different appearance matrix for every local tensor. Instead, we propose to utilize a global appearance matrix $\\mathbf{B}^c$ shared across the entire cloud of local tensors, leading to a global appearance dictionary that explains the color correlations across scene. This further improves both the computational efficiency and model compactness of our model.", + "bbox": [ + 75, + 518, + 468, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, each of our local tensors is represented by their unique local tri-vector factors $\\mathbf{v}_r^X$ , $\\mathbf{v}_r^Y$ , $\\mathbf{v}_r^Z$ .", + "bbox": [ + 76, + 700, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature evaluation. To achieve a continuous field, we consider trilinear interpolation when evaluating the tensor grid features. For a location $\\chi$ , we first compute its relative position $\\tilde{\\chi}$ to the selected tensor located at $p$ :", + "bbox": [ + 75, + 752, + 468, + 814 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x}, \\tilde {y}, \\tilde {z} = x - p _ {x}, y - p _ {y}, z - p _ {z}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 825, + 468, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, for example, to get $\\mathcal{A}_{\\sigma,r}$ at $(\\tilde{x},\\tilde{y},\\tilde{z})$ , we can compute and trilinearly interpolate eight $\\mathcal{A}_{\\sigma,r}$ on the corners. As mentioned in [6], applying linear interpolation on each", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vector first is mathematically equivalent and can reduce the computation cost. Under the rule of outer product, we have $\\mathcal{A}_{r,i,j,k} = \\mathbf{v}_{r,i}^{X}\\mathbf{v}_{r,j}^{Y}\\mathbf{v}_{r,k}^{Z}$ , then the interpolated density features at location $\\chi$ are:", + "bbox": [ + 498, + 428, + 890, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} ^ {\\sigma} (\\chi) = \\sum_ {r} \\mathbf {v} _ {\\sigma , r} ^ {X} (\\tilde {x}) \\mathbf {v} _ {\\sigma , r} ^ {Y} (\\tilde {y}) \\mathbf {v} _ {\\sigma , r} ^ {Z} (\\tilde {z}) = \\sum_ {r} \\mathcal {A} _ {\\sigma , r} (\\tilde {\\chi}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 498, + 890, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_{\\sigma,r}^{X}(\\tilde{x})$ is $\\mathbf{v}_{\\sigma,r}^{X}$ 's linearly interpolated value at $(\\tilde{x})$ along its $X$ axis. Here, $\\mathcal{G}^{\\sigma}(\\chi)$ is a scalar.", + "bbox": [ + 498, + 539, + 890, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, the interpolated appearance features can be computed as:", + "bbox": [ + 498, + 570, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {G} ^ {c} (\\chi) = \\sum_ {r} \\mathbf {v} _ {c, r} ^ {X} (\\tilde {x}) \\mathbf {v} _ {c, r} ^ {Y} (\\tilde {y}) \\mathbf {v} _ {c, r} ^ {Z} (\\tilde {z}) \\mathbf {b} _ {r} (8) \\\\ = \\sum_ {r} \\mathcal {A} _ {c, r} (\\tilde {\\chi}) \\mathbf {b} _ {r} (9) \\\\ = \\mathbf {B} \\cdot \\left(\\oplus \\left[ \\mathcal {A} _ {c, r} \\right] _ {r}\\right), (10) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 609, + 890, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes concatenation, $\\cdot$ denotes dot product. The appearance feature $\\mathcal{G}^c (\\chi)\\in \\mathbb{R}^P$ is a vector.", + "bbox": [ + 498, + 704, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Feature aggregation.", + "text_level": 1, + "bbox": [ + 500, + 743, + 694, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose to aggregate the features from $M$ neighboring tensors to jointly model the volume density and appearance for each 3D location $\\chi$ . In particular, inspired by Point-NeRF, we leverage an inverse distance-based weighting function to directly aggregate the multi-tensor features. Specifically, this weight can be expressed by", + "bbox": [ + 496, + 766, + 890, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {m} = \\frac {1}{\\| p _ {m} - \\chi \\|}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 864, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17572", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this weight function, we directly obtain the density feature via the weighted sum:", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {\\sigma} (\\chi) = \\frac {1}{\\sum w _ {m}} \\sum_ {m = 1} ^ {M} w _ {m} \\mathcal {G} _ {m} ^ {\\sigma} (\\chi). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 133, + 468, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, the appearance feature aggregation can be expressed in a similar way, while using the shared appearance matrix (as described in Sec. 3.1) across local tensors:", + "bbox": [ + 76, + 186, + 468, + 232 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f ^ {c} (\\chi) = \\frac {1}{\\sum w _ {m}} \\sum_ {m = 1} ^ {M} w _ {m} \\mathcal {G} _ {m} ^ {c} (\\chi) (13) \\\\ = \\frac {1}{\\sum w _ {m}} \\sum_ {m = 1} ^ {M} w _ {m} \\mathbf {B} ^ {c} \\cdot \\left(\\oplus \\left[ \\mathcal {A} _ {c, r} \\right] _ {r}\\right) (14) \\\\ = \\frac {1}{\\sum w _ {m}} \\mathbf {B} ^ {c} \\cdot \\left(\\sum_ {m = 1} ^ {M} w _ {m} \\left(\\oplus \\left[ \\mathcal {A} _ {c, r} \\right] _ {r}\\right)\\right). (15) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 244, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that owing to sharing the appearance matrix across tensors, we reduce the computational complexity from $O(M \\cdot P \\cdot R_c)$ in Eqn.14, to $O((M + P) \\cdot R_c)$ in Eqn.15.", + "bbox": [ + 76, + 386, + 468, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Multi-scale tri-vector fields.", + "text_level": 1, + "bbox": [ + 76, + 441, + 325, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Complex 3D scenes often contain multi-frequency geometry and appearance details. This motivates us to build multi-scale tensor clouds to discover the local geometry and appearance commonalities at multiple scales. Our final radiance field is modeled by multiple tri-vector tensor clouds at $S$ different scales. Different clouds consist of tensors with different resolutions.", + "bbox": [ + 76, + 465, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To regress the density and appearance at a location $\\chi$ , we gather the density and appearance features from a set of tensor clouds that cover $\\chi$ , $\\{\\mathcal{T}_s|1\\leq s\\leq S,\\chi \\in \\Omega_s\\}$ . Please note that tensor clouds of certain scales might not cover the location, so that $\\| \\{\\mathcal{T}_s\\} \\| \\leq S$ . We simply compute the mean features across these scales:", + "bbox": [ + 76, + 573, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {\\sigma} (\\chi) = \\frac {1}{\\| \\left\\{\\mathcal {T} _ {s} \\right\\} \\|} \\sum_ {s} f _ {s} ^ {\\sigma} (\\chi), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 672, + 468, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {c} (\\chi) = \\frac {1}{\\| \\left\\{\\mathcal {T} _ {s} \\right\\} \\|} \\sum_ {s} f _ {s} ^ {c} (\\chi). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 710, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $f^{\\sigma}(\\chi)$ and $f^{c}(\\chi)$ are the final density and appearance features we aggregate across multiple scales and multiple neighboring tensors.", + "bbox": [ + 76, + 758, + 468, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.Decoding.", + "text_level": 1, + "bbox": [ + 76, + 816, + 189, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply softplus activation on the density feature $f^{\\sigma}(\\chi)$ to obtain the final volume density and regress the view-dependent color by sending the appearance feature $f^{c}(\\chi)$ and the viewing direction $\\mathbf{d}$ to the MLP decoder $\\psi$ .", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Rendering and Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 90, + 781, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Volume Rendering We evaluates each pixel's color with physically-based volume rendering via differentiable ray marching. Following NeRF [22], we sample $Q$ shading points at $\\{\\chi_q\\mid q = 1,\\dots,Q\\}$ along the ray, and accumulate radiance by density:", + "bbox": [ + 498, + 116, + 890, + 191 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc = \\sum_ {q = 1} ^ {Q} T _ {q} \\left(1 - \\exp \\left(- \\sigma_ {q} \\delta_ {q}\\right)\\right) c _ {q}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 207, + 890, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {q} = \\exp (- \\sum_ {t = 1} ^ {q - 1} \\sigma_ {t} \\delta_ {t}).\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 256, + 736, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\sigma_{q}$ and $c_{q}$ are the density and radiance at shading points; $\\delta_t$ is the marching distance per step; $T$ denotes transmittance.", + "bbox": [ + 500, + 309, + 890, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distributing local tensors. First of all, to better leverage the sparsity of a scene, we first obtain a geometric prior that roughly covers the scene geometry. The geometric prior can be in any commonly-used form, e.g., point cloud, occupancy grid, octree, or mesh vertices. Then we can uniformly distribute tensors in the spatial neighborhood of the geometry. For a multi-scale model, each of the scale will be distributed independently. For most of our results, we quickly optimize a coarse RGBA volume from the multiview images and use the optimized occupancy grid as the geometry prior, as done in DVGO [32], which finishes in seconds.", + "bbox": [ + 498, + 368, + 890, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To maintain training stability and speed, each tensor $\\tau$ 's position $p$ and coverage $\\omega$ is fixed once determined. We also initialize the $3(R_{\\sigma} + R_c)$ vectors $(\\mathbf{v}_{\\sigma,r}^{X}, \\dots, \\mathbf{v}_{c,r}^{X}, \\dots)$ of each tensor by normal distribution. For each scale $s$ , a $P \\times R_{c}$ appearance matrix $\\mathbf{B}_s^c$ is shared by all tri-vector tensors of that scale. Specifically, “ $\\mathbf{B}^c \\cdot ()$ ” in Eqn.15 can be efficiently implemented as a fully-connected neural layer. Therefore, $\\mathbf{B}^c$ for each scale and a global appearance MLP $\\psi$ will be implemented as neural networks and initialized by default methods [12].", + "bbox": [ + 498, + 551, + 890, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimization and objectives. Given a set of multi-view RGB images with camera poses, the sparse tri-vector radiance field is per-scene optimized to reconstruct the radiance fields, under the supervision of the ground truth pixel colors. Following the volume rendering equation 18, the L2 rendering loss can be past back to the global MLP and aggregated features, then, all the way to the appearance matrices and the feature vectors of local tensors.", + "bbox": [ + 498, + 732, + 890, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply a rendering loss to supervise the reconstruction and also apply an $L1$ regularization loss on density feature vectors $\\mathbf{v}_{\\sigma,r}$ to promote geometry sparsity and to avoid", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17573", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bcc09d4c20522b19d5a8132730c099cacf248ff298e2006aaaa3d2a87fec56a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBatchSizeStepsTime↓#Param.(M)↓PSNR↑SSIM↑LPIPSVgg↓LPIPSAlex↓
NeRF[20]4096300k35.0h1.2531.010.9470.081-
Plenoxels[41]5000128k11.4m194.5031.710.9580.049-
DVGO[32]500030k15.0m153.0031.950.9600.0530.035
Point-NeRF200k[39]4096200k5.5h27.7433.310.962*0.0490.027
InstantNGP[23]10k-85k30k3.9m11.6432.590.960--
TensoRF-CP[6]409630k25.2m0.9831.560.9490.0760.041
TensoRF-VM[6]409630k17.4m17.9533.140.9630.0470.027
Ours-24409630k34.3m7.0733.240.9630.0460.026
Ours-48409630k35.7m13.5233.550.9650.0440.025
", + "bbox": [ + 91, + 87, + 877, + 231 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Comparisons of our method with other radiance-based models [20, 37, 19, 39, 6, 23] on the Synthetic-NeRF dataset [20]. Ours-24 is the one with 24 components while Ours-48 is the one with 48 components. We report the corresponding rendering quality (PSNR, SSIM, and LPIPS), model capacity (number of parameters), and training time, batch size and steps. Our model achieves the best rendering quality with a compact model size. We report PointNeRF's updated SSIM.", + "bbox": [ + 75, + 234, + 893, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "overfitting as done in TensoRF [6]:", + "bbox": [ + 76, + 325, + 312, + 340 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\| C - \\tilde {C} \\| _ {2} ^ {2}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 351, + 468, + 369 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L 1} = \\frac {1}{N} \\sum_ {r} ^ {R _ {\\sigma}} \\| \\mathbf {v} _ {\\sigma , r} \\|, \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 373, + 468, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tilde{C}$ is the ground truth pixel color, $\\| \\mathbf{v}_{\\sigma ,r}\\|$ is the sum of absolute values of elements on density vectors, and $N$ is the total number of elements. The total loss is:", + "bbox": [ + 76, + 426, + 468, + 472 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {r} + \\alpha \\mathcal {L} _ {L 1}. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 486, + 468, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We set the weight of the sparsity term $\\alpha$ as $1e^{-5}$ by default.", + "bbox": [ + 76, + 513, + 464, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Implementation", + "text_level": 1, + "bbox": [ + 76, + 542, + 235, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain coarse scene geometry, we modify the coarse density estimation introduced in [32] and get a $100^{3}$ occupancy volume in 30 seconds. We can skip this step if there exists available geometric data, e.g., the meshes in ScanNet [9], or point clouds from multiview stereo. According to the experiments, our method is not very sensitive to the initial geometry. Please refer to our supplementary material for more details. We set the default number of scales to 3. In a scene box of [-1,1], we rasterize the scene geometry (occupied voxels centers or points) onto 3 grids with different voxel sizes, e.g., $0.4^{3}, 0.2^{3}, 0.1^{3}$ . For each grid, we distribute tri-vector tensors at the center of its occupied voxels. The tensor spatial coverage of these 3 scales is $0.6^{3}, 0.3^{3}, 0.15^{3}$ , respectively. For each scale, we query $M = 4$ nearby tensors. Following [32], our feature decoding network $\\psi$ is a 2-layer MLP with 128 channels. For each scale, its appearance matrix $\\mathbf{B}^c$ is implemented by a single linear layer of 27 channels.", + "bbox": [ + 75, + 568, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement the framework in PyTorch [14] with customized CUDA operations. During optimization, we adopt the coarse to fine strategy in [6], linearly up-sample the vectors' dimension $(I,J,K)$ from 29 to 121 for scale one,", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "15 to 61 for scale two, and 7 to 31 for scale three. The up-sampling happens at step 2000, 3000, 4000, 5500, and 7000. We use the Adam optimizer [16] with initial learning rates of 0.02 for vectors and 0.001 for networks. On a single 3090 RTX GPU, we train each model for 30000 steps while each batch contains 4096 rays. Please find more details in the supplemental materials.", + "bbox": [ + 496, + 325, + 890, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 498, + 445, + 633, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Evaluation on the NeRF Synthetic Dataset.", + "text_level": 1, + "bbox": [ + 498, + 469, + 864, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate our method on the Synthetic-NeRF dataset [20] and the quantitative results compared with other methods are reported in Tab.1, including NeRF [22], Plenoxels [41], DVGO [32], Point-NeRF [39], iNGP [23], and TensoRF [6]. We report our models of two different model sizes with different numbers of components; both settings are with the same 3 scales of local tensors.", + "bbox": [ + 496, + 492, + 890, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach achieves the best averaged PSNR, $\\mathrm{LPIPS}_{Vgg}$ and $\\mathrm{LPIPS}_{Alex}$ in all the methods, leading to superior visual quality as shown in Fig. 3 Meanwhile, our high rendering quality is achieved with a compact model size. When compared with local voxel-based representations, such as Plenoxels and DVGO, our approach are significantly better.", + "bbox": [ + 496, + 598, + 890, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the other hand, global featuring encoding-based methods, like iNGP and TensorRF, are known for their high compactness and can achieve higher rendering quality than local voxel-based methods. Nonetheless, our method can still outperform them. Note that, even our smaller model (Ours-24) leads to better rendering quality than both iNGP and TensorRF that leverage global feature encoding, while our model uses significantly fewer parameters (about $60\\%$ and $40\\%$ of the size of iNGP and TensorRF). This clearly demonstrates the high visual quality and high compatness of our model with our sparsely distributed tri-vector tensors.", + "bbox": [ + 496, + 704, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In all the baseline methods, Point-NeRF is able to achieve relatively higher rendering quality than others.", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17574", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/557987c07a49e8537a71ac83d5031db513debc16f6efd48474ab5095d1bceb33.jpg", + "image_caption": [ + "Figure 3: Qualitative comparisons on the NeRF Synthetic dataset [22]." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 872, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, this is enabled by optimizing their model for 300k steps with a long period of 5.5 hours. In contrast, our method achieves higher quality with significantly fewer optimization steps (only 30k) and optimization time (about 36 min). As expected, our model is slower to optimize than TensoRF due to the additional costs of multi-tensor aggregation. However, though speed is not our focus in this work, our model can still converge quickly and lead to significantly faster reconstruction than MLP-based methods, such as NeRF, as well as Point-NeRF that is point-based.", + "bbox": [ + 75, + 383, + 472, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance w.r.t. rotation. We observe that tensor factorization-based methods can be sensitive to the orientation of the coordinate frames, since axis-aligned features are used; in essence, this is because the rank of a sparse tensor is sensitive to rotation, as shown in Fig. 4. Especially, this can benefit reconstruction on synthetic synthetic scenes where the objects are perfectly aligned with the axes, e.g. the lego scene in the NeRF synthetic data. However, we find that our method based on local tensors are more robust against the orientation of the axes than a global TensoRF. In particular, we compare our method with TensoRF in Tab.2 with different degrees of rotation around the Z axis on two scenes, lego (which is strongly aligned with axes) and chair (which is less aligned and thus less affected). As shown in the table, while both methods are affected by the rotation, our method has much smaller drops of PSNRs.", + "bbox": [ + 75, + 542, + 472, + 785 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c3970cc2762a279ac752922c262a38174e3a3363baaa68af38d85ac2e906c17.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
chair / legorot 0°rot 5°rot 45°
TensoRF-CP33.60 / 34.5032.90 / 29.7932.50 / 28.57
TensoRF-VM35.76 / 36.4634.91 /32.5334.55 / 32.31
Ours-4835.88 / 36.5235.72 / 35.3735.64 / 34.97
", + "bbox": [ + 78, + 804, + 473, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Comparisons of our method with TensoRF [6] on the chair and lego scenes of Synthetic-NeRF dataset [20] when considering rotation of different angles around $z$ -axis.", + "bbox": [ + 75, + 867, + 470, + 912 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c26d2db69956d3a5b1b4a95ac446048372848aee999a53e74b76ce3c70884e1f.jpg", + "image_caption": [ + "Figure 4: A toy example to illustrate the TensoRF-CP with global decomposition in (left) axis-aligned and (right) non-axis-aligned situations. The bottom shows the grid values. In axis-aligned case, only 1 component is needed to represent the scene (vector bases recover grid values by outer product). In non-axis-aligned case, however, 3 components are needed because the rank of matrix changes from 1 to 3 after scene rotation. While our design with local low-rank tensors can alleviate this issue." + ], + "image_footnote": [], + "bbox": [ + 504, + 378, + 686, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6048dfe4579b808ce8a9e6f269bbe6f02625dc5ef5d516011b79cdcdd8bf9fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 378, + 890, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Evaluation on the real datasets.", + "text_level": 1, + "bbox": [ + 500, + 723, + 774, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The ScanNet dataset. We evaluate our method on the real dataset, ScanNet [9] with the two scenes selected by NSVF [19], and compare with other methods. We follow the same experiment setting as done in NSVF [19], using the provided mesh to distribute our tensors, and optimize our model, TensoRF for the same 100k steps for fair comparisons. Please note that Point-NeRF uses all scanned depth images as initial geometry instead of meshes. Therefore, we also obtain the results of Point-NeRF 100k steps from the authors, using the provided mesh for fairness. We find the", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17575", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b464a67390f1ad6718da847a507ad8400deef103299bf826d2da4b8e922115ca.jpg", + "image_caption": [ + "Figure 5: Qualitative comparisons on the ScanNet dataset." + ], + "image_footnote": [], + "bbox": [ + 117, + 75, + 854, + 306 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/28e5432a11371fc7c5e9d6df2d8ced63eaa5e45b0a23b4a7a96a2b396033a7ce.jpg", + "table_caption": [ + "Average over Scene 101 and Scene 241" + ], + "table_footnote": [ + "Table 3: Quantitative comparison on two scenes in the ScanNet dataset [9]. Point-NeRF, TensoRF-CP, TensoRFVM and Ours-48 are optimized for $100\\mathrm{k}$ steps." + ], + "table_body": "
PSNR ↑SSIM ↑LPIPSAlex↓#Param.(M) ↓
SRN [31]18.250.5920.586-
NeRF [20]22.990.6200.369-
NSVF [19]25.480.6880.301-
Point-NeRF[39]25.920.8910.273159.01
TensoRF-CP[6]27.540.7510.3280.97
TensoRF-VM[6]28.610.7870.29049.92
Ours-4829.050.7920.24312.82
", + "bbox": [ + 76, + 354, + 478, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scannet data has many holes in the provided mesh geometry, while methods, such as NSVF and Point-NeRF, require accurate initial geometry; though Point-NeRF can potentially fix them with it point growing technique as shown in their original paper, it is not able to fully address them in 100k step optimization and lead to holes in their final rendering. Our approach instead does not require very accurate coarse geometry, since our local tensors cover relatively large regions. We show the quantitative results in Tab. 3 and qualitative results in Fig. 5. Note that our method can also perform well on real scenes, achieving the highest performance in terms of PSNR and $\\mathrm{LPIPS}_{Alex}$ , while using the second smallest model size (the smallest one is TensoRF $\\mathrm{CP}_{100k}$ ). Our visual quality is also higher than the comparison methods.", + "bbox": [ + 75, + 542, + 468, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Tanks and Temples dataset. We also evaluate our method on another real dataset, Tanks and Temples [9] with the 5 object scenes selected by NSVF [19]. We using the very coarse geometries estimated by DVGO[32] to distribute our tensors. We follow the same experiment setting as done in TensoRF [6], optimizing our model for the same $30\\mathrm{k}$ steps for fair comparisons. As is shown in", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/027ceafb4023c990e286e1efe9bc2695d32fa5b39c07bcd0089e5d0711caf86d.jpg", + "table_caption": [ + "Tab. 3, our method outperforms other methods in terms of PSNR, SSIM and LPIPS ${}_{Alex}$ ,while using the second smallest model size." + ], + "table_footnote": [ + "Table 4: Ablation under different scale settings on Synthetic-NeRF dataset. We select 3 scales of tensors with cube sizes of 0.6, 0.3, and 0.15." + ], + "table_body": "
ScalePSNR ↑SSIM ↑#Param.(M) ↓Time ↓
Single(0.6)32.220.9571.7518.22m
Single(0.3)32.730.9614.1521.31m
Single(0.15)31.960.95210.2028.55m
Multi(0.6, 0.3)33.110.9636.2030.12m
Multi(0.6, 0.3, 0.15)33.550.96513.5235.70m
", + "bbox": [ + 500, + 398, + 897, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 560, + 650, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze our model in terms of different scales in Table.4, while keeping the number of components the same (here we use 48). The scale here is the size of our local tensors of each axis. We considered 3 different scales, i.e., 0.6, 0.3, and 0.15 respectively as single-scale settings and some of their combinations as multi-scale settings. Note that even with a single scale (0.3), the performance of our method can be comparable with some other methods such as iNGP [23] while ours have less than half of the model size. When increasing the number of scales or decreasing the size of local tensors, our model size will also increase. In general, there is a trade-off of our method between scales and computational consumption (time and size).", + "bbox": [ + 496, + 583, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Usually, a smaller scale can lead to better performance, though our method with a scale of 0.15 does not strictly follow because we don't have high-quality input geometry to place these local tensors with a very small size. In fact, according to our per-scene breakdown results on the Synthetic-NeRF dataset (please refer to our supplemental material), single-scale(0.075) can achieve higher performance than single-scale(0.3) and single-scale(0.15) on most", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17576", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "scenes, except for ship because it has many thin structures that our coarse reconstruction does not cover.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/865a71a5b395d8bd45d63fcea042d6f5128947806641bc6b03a7eda4c6d4fa3e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPSAlex ↓#Param.(M) ↓
NeRF [22]25.780.8640.198-
NSVF [19]28.400.9000.153-
TensoRF-CP30k[6]27.590.8970.1440.97
TensoRF-VM30k[6]28.560.9200.12549.92
Ours-4830k28.700.9220.11314.11
", + "bbox": [ + 78, + 130, + 478, + 218 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5: Quantitative comparison on scenes in the Tanks and Temples dataset [17] selected in NSVF [19]. TensoRF-CP, TensoRF-VM and Ours-48 are optimized for 30k steps.", + "bbox": [ + 76, + 220, + 468, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also compare our method with a variant that uses vector-matrix (VM) decomposition [6] in each local tensor instead of CP decomposition. Please refer to Appendix.A. for more details. Also, we can achieve a higher training and inference speed without a significant performance drop, which we refer to the supplementary material.", + "bbox": [ + 76, + 273, + 468, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 376, + 194, + 391 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we have presented a novel approach for high-quality neural scene reconstruction and photorealistic novel view synthesis. We propose a novel tensor factorization-based scene representation, which leverages CP decomposition to compactly model a 3D scene as a sparse set of multi-scale tri-vector tensors that express local radiance fields. Our representation leverages both sparsity and spatial local coherence, and leads to accurate and efficient modeling of complex scene geometry and appearance. We demonstrate that the sparse tri-vector radiance fields can achieve superior rendering quality than previous state-of-the-art neural representations, including TensoRF and iNGP, while using significantly fewer parameters.", + "bbox": [ + 75, + 402, + 468, + 599 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 609, + 171, + 626 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In International Conference on Machine Learning (ICML), pages 40-49, 2018. 2", + "[2] Relja Arandjelović and Andrew Zisserman. Nerf in detail: Learning to sample for view synthesis. arXiv preprint arXiv:2106.05264, 2021. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 1, 2", + "[4] J Douglas Carroll and Jih-Jie Chang. Analysis of individual differences in multidimensional scaling via an n-way generalization of \"eckart-young\" decomposition. Psychometrika, 35(3):283-319, 1970. 1, 3", + "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J" + ], + "bbox": [ + 86, + 635, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022. 3", + "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), pages 333-350. Springer, 2022. 1, 2, 3, 4, 6, 7, 8, 9", + "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6970-6981, 2020. 2", + "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5828-5839, 2017. 6, 7, 8", + "[10] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479–12488, 2023. 3", + "[11] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4857-4866, 2020. 2", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 1026-1034, 2015. 5", + "[13] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2821-2830, 2018. 2", + "[14] Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan. Pytorch. Programming with TensorFlow: Solution for Edge Computing Applications, pages 87-104, 2021. 6", + "[15] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In European Conference on Computer Vision (ECCV), 2018. 2", + "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene" + ], + "bbox": [ + 503, + 90, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17577", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "reconstruction. ACM Transactions on Graphics, 36(4), 2017. 9", + "[18] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2016. 2", + "[19] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. arXiv preprint arXiv:2007.11571, 2020. 2, 6, 7, 8, 9", + "[20] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7210-7219, 2021. 2, 6, 7, 8", + "[21] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 1, 2, 5, 6, 7, 9", + "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):1-15, 2022. 1, 3, 6, 8", + "[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[25] Anton Obukhov, Mikhail Usvyatsov, Christos Sakaridis, Konrad Schindler, and Luc Van Gool. Tt-nf: Tensor train neural fields. arXiv preprint arXiv:2209.15529, 2022. 3", + "[26] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2", + "[27] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1", + "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[29] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. arXiv preprint arXiv:2103.13744, 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Jonathan Richard Shewchuk. Tetrahedral mesh generation by delaunay refinement. In Proceedings of the fourteenth annual symposium on Computational geometry, pages 86-95, 1998. 2", + "[31] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. arXiv preprint arXiv:1906.01618, 2019. 8", + "[32] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 2, 3, 4, 5, 6, 8", + "[33] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 3", + "[34] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2", + "[35] Jinglu Wang, Bo Sun, and Yan Lu. MVPnet: Multi-view point regression networks for 3D object reconstruction from a single image. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2", + "[36] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single RGB images. In European Conference on Computer Vision (ECCV), 2018. 2", + "[37] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6", + "[38] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 2", + "[39] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5438-5448, 2022. 2, 3, 6, 8", + "[40] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Proc. NeurIPS, 2020. 2", + "[41] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 2, 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17578", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1", + "[43] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5449-5458, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17579", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Strivec_ Sparse Tri-Vector Radiance Fields/35b59dcb-1ce5-40a4-83d2-07192fb62870_model.json b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/35b59dcb-1ce5-40a4-83d2-07192fb62870_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4ffc4d840c445f426ddefed3e3130f840f832eaa --- /dev/null +++ b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/35b59dcb-1ce5-40a4-83d2-07192fb62870_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.268, + 0.131, + 0.703, + 0.154 + ], + "angle": 0, + "content": "Strivec: Sparse Tri-Vector Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.18, + 0.245, + 0.198 + ], + "angle": 0, + "content": "Quankai Gao\\*1" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.181, + 0.408, + 0.199 + ], + "angle": 0, + "content": "Qiangeng Xu*1" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.199, + 0.403, + 0.216 + ], + "angle": 0, + "content": "1University of Southern California" + }, + { + "type": "text", + "bbox": [ + 0.45, + 0.181, + 0.52, + 0.198 + ], + "angle": 0, + "content": "Hao Su²" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.181, + 0.7, + 0.198 + ], + "angle": 0, + "content": "Ulrich Neumann1" + }, + { + "type": "text", + "bbox": [ + 0.492, + 0.199, + 0.617, + 0.217 + ], + "angle": 0, + "content": "\\(^{2}\\)UC San Diego" + }, + { + "type": "text", + "bbox": [ + 0.742, + 0.181, + 0.845, + 0.198 + ], + "angle": 0, + "content": "Zexiang \\(\\mathbf{X}\\mathbf{u}^3\\)" + }, + { + "type": "text", + "bbox": [ + 0.702, + 0.199, + 0.842, + 0.216 + ], + "angle": 0, + "content": "\\(^{3}\\)Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.22, + 0.493, + 0.234 + ], + "angle": 0, + "content": "{quankaig, qiangenx, uneumann}@usc.edu" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.22, + 0.644, + 0.233 + ], + "angle": 0, + "content": "haosu@ucsd.edu" + }, + { + "type": "text", + "bbox": [ + 0.67, + 0.22, + 0.796, + 0.233 + ], + "angle": 0, + "content": "zexu@adobe.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.65 + ], + "angle": 0, + "content": "We propose Strivec, a novel neural representation that models a 3D scene as a radiance field with sparsely distributed and compactly factorized local tensor feature grids. Our approach leverages tensor decomposition, following the recent work TensoRF [6], to model the tensor grids. In contrast to TensoRF which uses a global tensor and focuses on their vector-matrix decomposition, we propose to utilize a cloud of local tensors and apply the classic CANDE-COMP/PARAFAC (CP) decomposition [4] to factorize each tensor into triple vectors that express local feature distributions along spatial axes and compactly encode a local neural field. We also apply multi-scale tensor grids to discover the geometry and appearance commonalities and exploit spatial coherence with the tri-vector factorization at multiple local scales. The final radiance field properties are regressed by aggregating neural features from multiple local tensors across all scales. Our tri-vector tensors are sparsely distributed around the actual scene surface, discovered by a fast coarse reconstruction, leveraging the sparsity of a 3D scene. We demonstrate that our model can achieve better rendering quality while using significantly fewer parameters than previous methods, including TensoRF and Instant-NGP [23]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.676, + 0.21, + 0.691 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Representing 3D scenes as radiance fields [22] has enabled photo-realistic rendering quality and emerged as a popular design choice in 3D vision and graphics applications. While many methods [27, 42, 3] (following NeRF [22]) purely use MLPs to represent neural fields, recent works, like TensorRF [6] and Instant-NGP [23], have demonstrated the advantages of using shared global feature encoding for radiance field modeling, in terms of speed, compactness, and quality. However, these methods share and assign neural features uniformly in a scene (with tensor factors or hash tables), assuming the scene content is" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.267, + 0.6, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.353, + 0.596, + 0.363 + ], + "angle": 0, + "content": "TensoRF-VM-192" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.268, + 0.697, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.354, + 0.661, + 0.363 + ], + "angle": 0, + "content": "INGP" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.268, + 0.791, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.354, + 0.765, + 0.363 + ], + "angle": 0, + "content": "Ours-48" + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.268, + 0.886, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.82, + 0.353, + 0.866, + 0.363 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.368, + 0.891, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.749 + ], + "angle": 0, + "content": "Figure 1: We compare with previous methods in terms of rendering quality (PSNR) and model capacity (number of parameters) on the NeRF Synthetic dataset on the bottom. Our method and TensorRF are shown with different model sizes. Our approach consistently achieve better rendering quality with fewer model parameters than TensorRF, as well as other methods like iNGP. On the top, we show one example of visual comparisons of the mic scene that has challenging fine-grained geometric structures, where our approach captures most of the details and is the closest to the reference. Note that the results of NeRF and Point-NeRF use 200k optimization steps while the rest use only 30k steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.776, + 0.893, + 0.822 + ], + "angle": 0, + "content": "equally complex over the entire space, which can be inefficient (requiring high model capacity) to accurately model intricate local scene details (see Fig.1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We aim to accurately and compactly model a 3D scene and reproduce the complex local details. To this end, we propose Strivec, a novel neural scene representation that utilizes sparsely distributed and compactly factorized local tensor grids to model a volumetric radiance field for high-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.876, + 0.214, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.888, + 0.41, + 0.901 + ], + "angle": 0, + "content": "Code and results: https://github.com/Zerg-Overmind/Strivec" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.876, + 0.41, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17569" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "quality novel view synthesis. As shown in Fig.1, our approach is able to accurately model the complex scene structures that are not recovered well by previous methods. More importantly, our superior rendering quality is achieved with much less model capacity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.47, + 0.352 + ], + "angle": 0, + "content": "In particular, we base our model on TensoRF [6], a recent approach that leverages tensor factorization in radiance field modeling. It is fast, compact, and of high rendering quality. TensoRF applies CP and vector-matrix (VM) decomposition techniques to factorize a field into vectors and matrices and model the entire scene as a global factorized tensor. Instead of a single global tensor, we leverage a sparse set of multiple small local tensors distributed around the scene surface for more efficient scene modeling. Specifically, each of our tensors represents a local radiance field inside its local bounding box and is compactly modeled with factorized triple vectors based on the CP decomposition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.471, + 0.672 + ], + "angle": 0, + "content": "Note that the global CP decomposition in TensoRF has led to a highly compact model but cannot achieve comparable rendering quality to their VM decomposition. This is because a tri-vector CP component is rank-one, while a global feature grid of an entire 3D scene is often complex and of high rank, requiring a large (impractical) number of CP components for high accuracy. TensoRF addresses this by introducing matrix factors in their VM decomposition, essentially increasing the rank of each tensor component. Our model instead consists of multiple small tensor grids, exploiting local spatial commonalities in a scene. Compared to a global tensor, our local tensor is less complex and of much lower rank, thus effectively reducing the required number of CP components (per tensor) and enabling practical high-quality radiance field reconstruction with highly compact tri-vector factors. Our local tri-vector tensors can lead to superior rendering quality and compactness over TensoRF's VM model (see Fig. 1). We also observe that our local tensors are generally more robust than a global tensor against the orientation of spatial axes (which can affect the rank of a tensor and thus affects the quality; see Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Importantly, adopting local tensors (instead of a global one) also brings us the flexibility to allocate neural features according to the actual scene distribution, enabling more efficient scene modeling and better usage of model parameters than a global representation. To do so, we pre-acquire coarse scene geometry – that can be easily achieved via a fast RGB\\(\\sigma\\) volume reconstruction (like DVGO [32]) or multi-view stereo (like Point-NeRF [39]) – to directly distribute local tensors around the actual scene surface, leading to a sparse scene representation that avoids unnecessarily modeling the empty scene space. Note that while previous methods have also leveraged sparse representations (with voxels [19, 41] or points [39]) of radiance fields, their local features are modeled and optimized independently. Our model instead correlates a group of local features inside a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.272 + ], + "angle": 0, + "content": "local box and compactly express them with triple vectors, uniquely exploiting the local spatial coherence along axes and imposing local low-rank priors in the feature encoding via tensor factorization. Moreover, unlike previous sparse representations that only use a single-scale feature grid or point cloud, we distribute multi-scale local tensors to effectively model the scene geometry and appearance at multiple scales in a hierarchical manner. In particular, for an arbitrary 3D location, we aggregate the neural features from its neighboring tri-vector components at all scales and decode the volume density and view-dependent color from the aggregated features for radiance field rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Our approach takes the best of previous local and global radiance field representations. Compared with global representations like TensoRF and Instant-NGP, our model takes advantage of the sparsity of a scene more directly; compared with local representations like Plenoxels and Point-NeRF, our model makes use of the local smoothness and coherence of scene geometry and appearance. As shown in our experimental results on both synthetic and real datasets, our model is able to achieve state-of-the-art rendering quality on these datasets, outperforming previous methods, including TensoRF and Instant-NGP, while using significantly fewer model parameters, demonstrating the superior representational power of our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.481, + 0.642, + 0.496 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Scene representations. To represent a 3D scene, traditional and learning-based methods have studied various representations, such as depth map [13, 18], mesh [15, 36, 30], point cloud [28, 1, 35] and implicit function [7, 21, 24, 40]. In recent years, continuous neural field representations stand out in various 3D tasks such as single-view 3D reconstruction [38, 11], surface completion [8, 26], multi-view reconstruction [24] and novel view synthesis [22, 20]. Compared with traditional discrete representations, a continuous field have no limitation on spatial resolution, e.g., volume resolution or the number of points. It can also naturally be represented by neural networks, such as an MLP, which are known for approximating complex functions well." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Neural field representations. Specifically, NeRF [22] represents a 3D scene as a radiance field with a global coordinate MLP, which models geometry, lighting and texture information jointly, leading to photo-realistic rendering quality in novel view synthesis. Apart from its advantage, purely MLP-based NeRF models [3, 34] in general suffer from inefficiency [2] when modeling highly complex or large-scale scenes, due to limited model capacity, slow optimization speed, and the cost of modeling empty space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To model radiance fields more efficiently, recent works have explored combining neural fields with various traditional 3D representations, including voxels [19, 41, 32, 43]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17570" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "and points [39]. Low-rank representations such as triplane [5, 10] and tensor decomposition [6, 25] have also been studied. In particular, DVGO [32] and Plenoxels [41] respectively use dense and sparse voxels with neural features for radiance field modeling. While being efficient to optimize, these localized feature grid-based representations lead to a large model size and can face overfitting issues when the features are of very high resolution. Consequently, DVGO can also work with a low-resolution grid and Plenoxels requires additional spatial regularization terms. On the other hand, recent works have adopted global feature encoding to express a high-resolution feature grid, including Instant-NGP [23] that hashes spatial features into multi-scale hash tables and TensorRF [6] that factorizes a feature grid into vector and matrix factors. These global feature encoding methods exploit the spatial correlation across the entire scene space, leading to fast and compact reconstruction and surpassing previous MLP-based or grid-based representations on rendering quality. However, similar to NeRF, such global representation can also be limited by its model capacity when representing highly complex or large-scale content." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.583 + ], + "angle": 0, + "content": "Our approach instead combines local and global representations. Our tri-vector fields are sparsely distributed in the scene, similar to local representations (like plenoxels and Point-NeRF); meanwhile, features in each field are represented by tri-vector components shared across the local region as done in TensoRF, exploiting spatial feature commonalities. Our model leverages both spatial sparsity and coherence, leading to much higher compactness and better reconstruction quality than previous local and global representations (see Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.588, + 0.471, + 0.725 + ], + "angle": 0, + "content": "Relevant to our work, previous methods, such as KiloNeRF [29] and BlockNeRF [33] have also utilized multiple local MLPs to represent a scene. Specifically, KiloNeRF focuses and speeding up NeRF and their rendering quality is sacrificed; BlockNeRF essentially uses multiple NeRFs to increase the total model capacity. Instead of pure MLPs, our work is built upon tensor factorization-based feature encoding as done in TensoRF [6], and we in fact achieve superior rendering quality while decreasing the model capacity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.43, + 0.775 + ], + "angle": 0, + "content": "3. Sparse Tri-Vector Field Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.789, + 0.469, + 0.85 + ], + "angle": 0, + "content": "We now present our novel radiance field representation. In essence, our model consists of a cloud of small local trivector tensors at multiple scales, designed to leverage both sparsity and multi-scale spatial coherence (see Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Let \\(\\mathcal{T} = \\{\\tau_n|n = 1,\\dots,N\\}\\) denote a cloud of tri-vector tensors. Each local tensor \\(\\tau\\) is located at \\(p\\), covering a local cuboid space \\(\\omega\\) with an edge length of \\(l\\). This cloud of tri" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.885, + 0.108 + ], + "angle": 0, + "content": "vector tensors represents a radiance field for the 3D space:" + }, + { + "type": "equation", + "bbox": [ + 0.65, + 0.118, + 0.892, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\Omega = \\bigcup_ {n = 1} ^ {N} \\omega_ {n}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.17, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Here, each tensor \\(\\tau\\) encodes a local multi-channel feature grid that includes a (single-channel) density grid \\(A_{\\sigma}\\) and a (multi-channel) appearance grid \\(A_{c}\\), similar to the tensor grid in TensoRF [6]. In contrast to using a single global tensor in TensoRF [6], we model the volume density and view-dependent colors with multiple local tensors. In particular, for an arbitrary location \\(\\chi \\in \\Omega\\), we select \\(M\\) nearest tensors that cover \\(\\chi\\). Across the selected tensors, we aggregate the extracted density and appearance features recovered by their tri-vector factors for radiance field property regression, where the volume density \\(\\sigma\\) is directly obtained after the aggregation and the view-dependent color \\(c\\) is regressed by a small MLP \\(\\psi\\) along with the viewing direction \\(\\mathbf{d}\\). The continuous radiance field can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.394, + 0.892, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {\\chi}, c _ {\\chi} = A _ {\\sigma} \\left(\\left\\{\\mathcal {G} ^ {\\sigma} (\\chi) \\right\\}\\right), \\psi \\left(A _ {c} \\left(\\left\\{\\mathcal {G} ^ {c} (\\chi) \\right\\}\\right), \\mathbf {d}\\right). \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.422, + 0.72, + 0.436 + ], + "angle": 0, + "content": "3.1. Local tri-vector tensors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.49 + ], + "angle": 0, + "content": "We apply the classic Canonical polyadic (CP) decomposition [4] to model our local tensors with tri-vector components." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.554 + ], + "angle": 0, + "content": "CP decomposition. CP decomposition factorizes a \\(M\\) dimension tensor \\(\\tau \\in \\mathbb{R}^{I_1\\times I_2\\times \\ldots \\times I_M}\\) into a linear combination of \\(R\\) rank-1 tensors:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.564, + 0.892, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\tau = \\sum_ {r = 1} ^ {R} \\lambda_ {r} \\mathbf {v} _ {r} ^ {0} \\otimes \\mathbf {v} _ {r} ^ {1} \\otimes \\dots , \\otimes \\mathbf {v} _ {r} ^ {M}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.892, + 0.647 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes outer product; the weighting factor \\(\\lambda_r\\) can be absorbed into vectors \\(\\{\\mathbf{v}_r^0,\\dots ,\\mathbf{v}_r^M\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Density and appearance tensors. In our case of modeling a 3D radiance field, we set the geometry grid \\(\\mathcal{G}^{\\sigma} \\in \\mathbb{R}^{I \\times J \\times K}\\) as a 3D tensor. And the multi-channel appearance grid \\(\\mathcal{G}^c \\in \\mathbb{R}^{I \\times J \\times K \\times P}\\) corresponds to a 4D tensor. The fourth appearance mode is of lower dimension (compared with the spatial modes), representing the final dimension of the features sent to the MLP decoder network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.771, + 0.892, + 0.802 + ], + "angle": 0, + "content": "According to Eqn.3, we factorize each tensor's feature grids, \\(\\mathcal{G}^{\\sigma}\\) and \\(\\mathcal{G}^{c}\\), by CP decomposition:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.812, + 0.892, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} ^ {\\sigma} = \\sum_ {r = 1} ^ {R _ {\\sigma}} \\mathcal {A} _ {\\sigma , r} = \\sum_ {r = 1} ^ {R _ {\\sigma}} \\mathbf {v} _ {\\sigma , r} ^ {X} \\otimes \\mathbf {v} _ {\\sigma , r} ^ {Y} \\otimes \\mathbf {v} _ {\\sigma , r} ^ {Z}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.857, + 0.892, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} ^ {c} = \\sum_ {r = 1} ^ {R _ {c}} \\mathcal {A} _ {c, r} \\otimes \\mathbf {b} _ {r} = \\sum_ {r = 1} ^ {R _ {c}} \\mathbf {v} _ {c, r} ^ {X} \\otimes \\mathbf {v} _ {c, r} ^ {Y} \\otimes \\mathbf {v} _ {c, r} ^ {Z} \\otimes \\mathbf {b} _ {r}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17571" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.895, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.895, + 0.41 + ], + "angle": 0, + "content": "Figure 2: Overview of our Sparse Tri-Vector Radiance Fields. We distribute our local tensors based on a coarse geometry estimated by a fast \\(\\mathrm{RGB}\\sigma\\) volume reconstruction as done in DVGO [32]. Here, we show our model running under \\(S = 2\\) scales. Each local tensor is factorized as axis-aligned triple based on CP decomposition. For any shading point \\(\\chi\\), we extract and evaluate features in each local tensor, according to the factorization (Sec. 3.1). Then, we aggregate these features among nearby tensors (Sec. 3.2) and across different scales (Sec. 3.3). Finally, the density and color are decoded (Sec. 3.4) and used by volume rendering (Sec.4)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.429, + 0.47, + 0.519 + ], + "angle": 0, + "content": "Here \\(R_{\\sigma}\\) and \\(R_{c}\\) denote numbers of component; \\(\\mathcal{A}_{\\sigma,r}\\) and \\(\\mathcal{A}_{c,r}\\) are the component tensors that are factorized spatially; \\(\\mathbf{v}_{\\sigma,r}^{X}, \\dots, \\mathbf{v}_{c,r}^{X}, \\dots\\) are the 1D vectors with resolution \\(I, J, K\\), modeling scene geometry and appearance along \\(X, Y, Z\\) axis; \\(R_{\\sigma}\\) and \\(R_{c}\\) are the component numbers; \\(\\mathbf{b}_r\\) expresses the feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.7 + ], + "angle": 0, + "content": "As done in TensoRF [6], we stack all feature-mode vectors \\(\\mathbf{b}_r\\) as columns together, which ends up a \\(P\\times R_{c}\\) appearance matrix \\(\\mathbf{B}\\). This matrix models the appearance feature variations of the tensor and functions like a appearance dictionary. Note that naively following CP decomposition like TensoRF will assign a different appearance matrix for every local tensor. Instead, we propose to utilize a global appearance matrix \\(\\mathbf{B}^c\\) shared across the entire cloud of local tensors, leading to a global appearance dictionary that explains the color correlations across scene. This further improves both the computational efficiency and model compactness of our model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.702, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Therefore, each of our local tensors is represented by their unique local tri-vector factors \\(\\mathbf{v}_r^X\\), \\(\\mathbf{v}_r^Y\\), \\(\\mathbf{v}_r^Z\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.47, + 0.815 + ], + "angle": 0, + "content": "Feature evaluation. To achieve a continuous field, we consider trilinear interpolation when evaluating the tensor grid features. For a location \\(\\chi\\), we first compute its relative position \\(\\tilde{\\chi}\\) to the selected tensor located at \\(p\\):" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.827, + 0.47, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\tilde {x}, \\tilde {y}, \\tilde {z} = x - p _ {x}, y - p _ {y}, z - p _ {z}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Then, for example, to get \\(\\mathcal{A}_{\\sigma,r}\\) at \\((\\tilde{x},\\tilde{y},\\tilde{z})\\), we can compute and trilinearly interpolate eight \\(\\mathcal{A}_{\\sigma,r}\\) on the corners. As mentioned in [6], applying linear interpolation on each" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.429, + 0.892, + 0.49 + ], + "angle": 0, + "content": "vector first is mathematically equivalent and can reduce the computation cost. Under the rule of outer product, we have \\(\\mathcal{A}_{r,i,j,k} = \\mathbf{v}_{r,i}^{X}\\mathbf{v}_{r,j}^{Y}\\mathbf{v}_{r,k}^{Z}\\), then the interpolated density features at location \\(\\chi\\) are:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.499, + 0.892, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} ^ {\\sigma} (\\chi) = \\sum_ {r} \\mathbf {v} _ {\\sigma , r} ^ {X} (\\tilde {x}) \\mathbf {v} _ {\\sigma , r} ^ {Y} (\\tilde {y}) \\mathbf {v} _ {\\sigma , r} ^ {Z} (\\tilde {z}) = \\sum_ {r} \\mathcal {A} _ {\\sigma , r} (\\tilde {\\chi}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.54, + 0.892, + 0.571 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_{\\sigma,r}^{X}(\\tilde{x})\\) is \\(\\mathbf{v}_{\\sigma,r}^{X}\\)'s linearly interpolated value at \\((\\tilde{x})\\) along its \\(X\\) axis. Here, \\(\\mathcal{G}^{\\sigma}(\\chi)\\) is a scalar." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.571, + 0.892, + 0.601 + ], + "angle": 0, + "content": "Similarly, the interpolated appearance features can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.61, + 0.891, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {G} ^ {c} (\\chi) = \\sum_ {r} \\mathbf {v} _ {c, r} ^ {X} (\\tilde {x}) \\mathbf {v} _ {c, r} ^ {Y} (\\tilde {y}) \\mathbf {v} _ {c, r} ^ {Z} (\\tilde {z}) \\mathbf {b} _ {r} (8) \\\\ = \\sum_ {r} \\mathcal {A} _ {c, r} (\\tilde {\\chi}) \\mathbf {b} _ {r} (9) \\\\ = \\mathbf {B} \\cdot \\left(\\oplus \\left[ \\mathcal {A} _ {c, r} \\right] _ {r}\\right), (10) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.892, + 0.736 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes concatenation, \\(\\cdot\\) denotes dot product. The appearance feature \\(\\mathcal{G}^c (\\chi)\\in \\mathbb{R}^P\\) is a vector." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.744, + 0.695, + 0.761 + ], + "angle": 0, + "content": "3.2. Feature aggregation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.767, + 0.892, + 0.859 + ], + "angle": 0, + "content": "We propose to aggregate the features from \\(M\\) neighboring tensors to jointly model the volume density and appearance for each 3D location \\(\\chi\\). In particular, inspired by Point-NeRF, we leverage an inverse distance-based weighting function to directly aggregate the multi-tensor features. Specifically, this weight can be expressed by" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.866, + 0.892, + 0.9 + ], + "angle": 0, + "content": "\\[\nw _ {m} = \\frac {1}{\\| p _ {m} - \\chi \\|}. \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17572" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "With this weight function, we directly obtain the density feature via the weighted sum:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.135, + 0.47, + 0.176 + ], + "angle": 0, + "content": "\\[\nf ^ {\\sigma} (\\chi) = \\frac {1}{\\sum w _ {m}} \\sum_ {m = 1} ^ {M} w _ {m} \\mathcal {G} _ {m} ^ {\\sigma} (\\chi). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.188, + 0.469, + 0.233 + ], + "angle": 0, + "content": "Similarly, the appearance feature aggregation can be expressed in a similar way, while using the shared appearance matrix (as described in Sec. 3.1) across local tensors:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.245, + 0.469, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f ^ {c} (\\chi) = \\frac {1}{\\sum w _ {m}} \\sum_ {m = 1} ^ {M} w _ {m} \\mathcal {G} _ {m} ^ {c} (\\chi) (13) \\\\ = \\frac {1}{\\sum w _ {m}} \\sum_ {m = 1} ^ {M} w _ {m} \\mathbf {B} ^ {c} \\cdot \\left(\\oplus \\left[ \\mathcal {A} _ {c, r} \\right] _ {r}\\right) (14) \\\\ = \\frac {1}{\\sum w _ {m}} \\mathbf {B} ^ {c} \\cdot \\left(\\sum_ {m = 1} ^ {M} w _ {m} \\left(\\oplus \\left[ \\mathcal {A} _ {c, r} \\right] _ {r}\\right)\\right). (15) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.387, + 0.469, + 0.434 + ], + "angle": 0, + "content": "Note that owing to sharing the appearance matrix across tensors, we reduce the computational complexity from \\( O(M \\cdot P \\cdot R_c) \\) in Eqn.14, to \\( O((M + P) \\cdot R_c) \\) in Eqn.15." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.443, + 0.326, + 0.458 + ], + "angle": 0, + "content": "3.3. Multi-scale tri-vector fields." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.467, + 0.469, + 0.572 + ], + "angle": 0, + "content": "Complex 3D scenes often contain multi-frequency geometry and appearance details. This motivates us to build multi-scale tensor clouds to discover the local geometry and appearance commonalities at multiple scales. Our final radiance field is modeled by multiple tri-vector tensor clouds at \\( S \\) different scales. Different clouds consist of tensors with different resolutions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.574, + 0.469, + 0.663 + ], + "angle": 0, + "content": "To regress the density and appearance at a location \\(\\chi\\), we gather the density and appearance features from a set of tensor clouds that cover \\(\\chi\\), \\(\\{\\mathcal{T}_s|1\\leq s\\leq S,\\chi \\in \\Omega_s\\}\\). Please note that tensor clouds of certain scales might not cover the location, so that \\(\\| \\{\\mathcal{T}_s\\} \\| \\leq S\\). We simply compute the mean features across these scales:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.673, + 0.469, + 0.709 + ], + "angle": 0, + "content": "\\[\nf ^ {\\sigma} (\\chi) = \\frac {1}{\\| \\left\\{\\mathcal {T} _ {s} \\right\\} \\|} \\sum_ {s} f _ {s} ^ {\\sigma} (\\chi), \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.711, + 0.469, + 0.746 + ], + "angle": 0, + "content": "\\[\nf ^ {c} (\\chi) = \\frac {1}{\\| \\left\\{\\mathcal {T} _ {s} \\right\\} \\|} \\sum_ {s} f _ {s} ^ {c} (\\chi). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.759, + 0.469, + 0.806 + ], + "angle": 0, + "content": "Note that \\( f^{\\sigma}(\\chi) \\) and \\( f^{c}(\\chi) \\) are the final density and appearance features we aggregate across multiple scales and multiple neighboring tensors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.817, + 0.19, + 0.833 + ], + "angle": 0, + "content": "3.4.Decoding." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We apply softplus activation on the density feature \\( f^{\\sigma}(\\chi) \\) to obtain the final volume density and regress the view-dependent color by sending the appearance feature \\( f^{c}(\\chi) \\) and the viewing direction \\( \\mathbf{d} \\) to the MLP decoder \\( \\psi \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.782, + 0.108 + ], + "angle": 0, + "content": "4. Rendering and Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.117, + 0.892, + 0.193 + ], + "angle": 0, + "content": "Volume Rendering We evaluates each pixel's color with physically-based volume rendering via differentiable ray marching. Following NeRF [22], we sample \\(Q\\) shading points at \\(\\{\\chi_q\\mid q = 1,\\dots,Q\\}\\) along the ray, and accumulate radiance by density:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.208, + 0.892, + 0.258 + ], + "angle": 0, + "content": "\\[\nc = \\sum_ {q = 1} ^ {Q} T _ {q} \\left(1 - \\exp \\left(- \\sigma_ {q} \\delta_ {q}\\right)\\right) c _ {q}, \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.257, + 0.737, + 0.296 + ], + "angle": 0, + "content": "\\[\nT _ {q} = \\exp (- \\sum_ {t = 1} ^ {q - 1} \\sigma_ {t} \\delta_ {t}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.31, + 0.891, + 0.342 + ], + "angle": 0, + "content": "\\(\\sigma_{q}\\) and \\(c_{q}\\) are the density and radiance at shading points; \\(\\delta_t\\) is the marching distance per step; \\(T\\) denotes transmittance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.369, + 0.892, + 0.55 + ], + "angle": 0, + "content": "Distributing local tensors. First of all, to better leverage the sparsity of a scene, we first obtain a geometric prior that roughly covers the scene geometry. The geometric prior can be in any commonly-used form, e.g., point cloud, occupancy grid, octree, or mesh vertices. Then we can uniformly distribute tensors in the spatial neighborhood of the geometry. For a multi-scale model, each of the scale will be distributed independently. For most of our results, we quickly optimize a coarse RGBA volume from the multiview images and use the optimized occupancy grid as the geometry prior, as done in DVGO [32], which finishes in seconds." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.553, + 0.892, + 0.705 + ], + "angle": 0, + "content": "To maintain training stability and speed, each tensor \\(\\tau\\)'s position \\(p\\) and coverage \\(\\omega\\) is fixed once determined. We also initialize the \\(3(R_{\\sigma} + R_c)\\) vectors \\((\\mathbf{v}_{\\sigma,r}^{X}, \\dots, \\mathbf{v}_{c,r}^{X}, \\dots)\\) of each tensor by normal distribution. For each scale \\(s\\), a \\(P \\times R_{c}\\) appearance matrix \\(\\mathbf{B}_s^c\\) is shared by all tri-vector tensors of that scale. Specifically, “\\(\\mathbf{B}^c \\cdot ()\\)” in Eqn.15 can be efficiently implemented as a fully-connected neural layer. Therefore, \\(\\mathbf{B}^c\\) for each scale and a global appearance MLP \\(\\psi\\) will be implemented as neural networks and initialized by default methods [12]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.733, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Optimization and objectives. Given a set of multi-view RGB images with camera poses, the sparse tri-vector radiance field is per-scene optimized to reconstruct the radiance fields, under the supervision of the ground truth pixel colors. Following the volume rendering equation 18, the L2 rendering loss can be past back to the global MLP and aggregated features, then, all the way to the appearance matrices and the feature vectors of local tensors." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We apply a rendering loss to supervise the reconstruction and also apply an \\(L1\\) regularization loss on density feature vectors \\(\\mathbf{v}_{\\sigma,r}\\) to promote geometry sparsity and to avoid" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17573" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.088, + 0.878, + 0.232 + ], + "angle": 0, + "content": "
MethodBatchSizeStepsTime↓#Param.(M)↓PSNR↑SSIM↑LPIPSVgg↓LPIPSAlex↓
NeRF[20]4096300k35.0h1.2531.010.9470.081-
Plenoxels[41]5000128k11.4m194.5031.710.9580.049-
DVGO[32]500030k15.0m153.0031.950.9600.0530.035
Point-NeRF200k[39]4096200k5.5h27.7433.310.962*0.0490.027
InstantNGP[23]10k-85k30k3.9m11.6432.590.960--
TensoRF-CP[6]409630k25.2m0.9831.560.9490.0760.041
TensoRF-VM[6]409630k17.4m17.9533.140.9630.0470.027
Ours-24409630k34.3m7.0733.240.9630.0460.026
Ours-48409630k35.7m13.5233.550.9650.0440.025
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.236, + 0.894, + 0.3 + ], + "angle": 0, + "content": "Table 1: Comparisons of our method with other radiance-based models [20, 37, 19, 39, 6, 23] on the Synthetic-NeRF dataset [20]. Ours-24 is the one with 24 components while Ours-48 is the one with 48 components. We report the corresponding rendering quality (PSNR, SSIM, and LPIPS), model capacity (number of parameters), and training time, batch size and steps. Our model achieves the best rendering quality with a compact model size. We report PointNeRF's updated SSIM." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.327, + 0.313, + 0.342 + ], + "angle": 0, + "content": "overfitting as done in TensoRF [6]:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.352, + 0.469, + 0.371 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\| C - \\tilde {C} \\| _ {2} ^ {2}, \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.374, + 0.469, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L 1} = \\frac {1}{N} \\sum_ {r} ^ {R _ {\\sigma}} \\| \\mathbf {v} _ {\\sigma , r} \\|, \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.428, + 0.47, + 0.473 + ], + "angle": 0, + "content": "where \\(\\tilde{C}\\) is the ground truth pixel color, \\(\\| \\mathbf{v}_{\\sigma ,r}\\|\\) is the sum of absolute values of elements on density vectors, and \\(N\\) is the total number of elements. The total loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.487, + 0.469, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {r} + \\alpha \\mathcal {L} _ {L 1}. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.514, + 0.465, + 0.53 + ], + "angle": 0, + "content": "We set the weight of the sparsity term \\(\\alpha\\) as \\(1e^{-5}\\) by default." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.236, + 0.56 + ], + "angle": 0, + "content": "5. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.84 + ], + "angle": 0, + "content": "To obtain coarse scene geometry, we modify the coarse density estimation introduced in [32] and get a \\(100^{3}\\) occupancy volume in 30 seconds. We can skip this step if there exists available geometric data, e.g., the meshes in ScanNet [9], or point clouds from multiview stereo. According to the experiments, our method is not very sensitive to the initial geometry. Please refer to our supplementary material for more details. We set the default number of scales to 3. In a scene box of [-1,1], we rasterize the scene geometry (occupied voxels centers or points) onto 3 grids with different voxel sizes, e.g., \\(0.4^{3}, 0.2^{3}, 0.1^{3}\\). For each grid, we distribute tri-vector tensors at the center of its occupied voxels. The tensor spatial coverage of these 3 scales is \\(0.6^{3}, 0.3^{3}, 0.15^{3}\\), respectively. For each scale, we query \\(M = 4\\) nearby tensors. Following [32], our feature decoding network \\(\\psi\\) is a 2-layer MLP with 128 channels. For each scale, its appearance matrix \\(\\mathbf{B}^c\\) is implemented by a single linear layer of 27 channels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We implement the framework in PyTorch [14] with customized CUDA operations. During optimization, we adopt the coarse to fine strategy in [6], linearly up-sample the vectors' dimension \\((I,J,K)\\) from 29 to 121 for scale one," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.432 + ], + "angle": 0, + "content": "15 to 61 for scale two, and 7 to 31 for scale three. The up-sampling happens at step 2000, 3000, 4000, 5500, and 7000. We use the Adam optimizer [16] with initial learning rates of 0.02 for vectors and 0.001 for networks. On a single 3090 RTX GPU, we train each model for 30000 steps while each batch contains 4096 rays. Please find more details in the supplemental materials." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.446, + 0.634, + 0.463 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.47, + 0.866, + 0.487 + ], + "angle": 0, + "content": "6.1. Evaluation on the NeRF Synthetic Dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.599 + ], + "angle": 0, + "content": "We first evaluate our method on the Synthetic-NeRF dataset [20] and the quantitative results compared with other methods are reported in Tab.1, including NeRF [22], Plenoxels [41], DVGO [32], Point-NeRF [39], iNGP [23], and TensoRF [6]. We report our models of two different model sizes with different numbers of components; both settings are with the same 3 scales of local tensors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Our approach achieves the best averaged PSNR, \\(\\mathrm{LPIPS}_{Vgg}\\) and \\(\\mathrm{LPIPS}_{Alex}\\) in all the methods, leading to superior visual quality as shown in Fig. 3 Meanwhile, our high rendering quality is achieved with a compact model size. When compared with local voxel-based representations, such as Plenoxels and DVGO, our approach are significantly better." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.87 + ], + "angle": 0, + "content": "On the other hand, global featuring encoding-based methods, like iNGP and TensorRF, are known for their high compactness and can achieve higher rendering quality than local voxel-based methods. Nonetheless, our method can still outperform them. Note that, even our smaller model (Ours-24) leads to better rendering quality than both iNGP and TensorRF that leverage global feature encoding, while our model uses significantly fewer parameters (about \\(60\\%\\) and \\(40\\%\\) of the size of iNGP and TensorRF). This clearly demonstrates the high visual quality and high compatness of our model with our sparsely distributed tri-vector tensors." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In all the baseline methods, Point-NeRF is able to achieve relatively higher rendering quality than others." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17574" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.873, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.348, + 0.719, + 0.364 + ], + "angle": 0, + "content": "Figure 3: Qualitative comparisons on the NeRF Synthetic dataset [22]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.473, + 0.536 + ], + "angle": 0, + "content": "However, this is enabled by optimizing their model for 300k steps with a long period of 5.5 hours. In contrast, our method achieves higher quality with significantly fewer optimization steps (only 30k) and optimization time (about 36 min). As expected, our model is slower to optimize than TensoRF due to the additional costs of multi-tensor aggregation. However, though speed is not our focus in this work, our model can still converge quickly and lead to significantly faster reconstruction than MLP-based methods, such as NeRF, as well as Point-NeRF that is point-based." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.473, + 0.786 + ], + "angle": 0, + "content": "Performance w.r.t. rotation. We observe that tensor factorization-based methods can be sensitive to the orientation of the coordinate frames, since axis-aligned features are used; in essence, this is because the rank of a sparse tensor is sensitive to rotation, as shown in Fig. 4. Especially, this can benefit reconstruction on synthetic synthetic scenes where the objects are perfectly aligned with the axes, e.g. the lego scene in the NeRF synthetic data. However, we find that our method based on local tensors are more robust against the orientation of the axes than a global TensoRF. In particular, we compare our method with TensoRF in Tab.2 with different degrees of rotation around the Z axis on two scenes, lego (which is strongly aligned with axes) and chair (which is less aligned and thus less affected). As shown in the table, while both methods are affected by the rotation, our method has much smaller drops of PSNRs." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.805, + 0.475, + 0.864 + ], + "angle": 0, + "content": "
chair / legorot 0°rot 5°rot 45°
TensoRF-CP33.60 / 34.5032.90 / 29.7932.50 / 28.57
TensoRF-VM35.76 / 36.4634.91 /32.5334.55 / 32.31
Ours-4835.88 / 36.5235.72 / 35.3735.64 / 34.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.868, + 0.471, + 0.914 + ], + "angle": 0, + "content": "Table 2: Comparisons of our method with TensoRF [6] on the chair and lego scenes of Synthetic-NeRF dataset [20] when considering rotation of different angles around \\(z\\)-axis." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.38, + 0.687, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.38, + 0.891, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.698 + ], + "angle": 0, + "content": "Figure 4: A toy example to illustrate the TensoRF-CP with global decomposition in (left) axis-aligned and (right) non-axis-aligned situations. The bottom shows the grid values. In axis-aligned case, only 1 component is needed to represent the scene (vector bases recover grid values by outer product). In non-axis-aligned case, however, 3 components are needed because the rank of matrix changes from 1 to 3 after scene rotation. While our design with local low-rank tensors can alleviate this issue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.776, + 0.739 + ], + "angle": 0, + "content": "6.2. Evaluation on the real datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The ScanNet dataset. We evaluate our method on the real dataset, ScanNet [9] with the two scenes selected by NSVF [19], and compare with other methods. We follow the same experiment setting as done in NSVF [19], using the provided mesh to distribute our tensors, and optimize our model, TensoRF for the same 100k steps for fair comparisons. Please note that Point-NeRF uses all scanned depth images as initial geometry instead of meshes. Therefore, we also obtain the results of Point-NeRF 100k steps from the authors, using the provided mesh for fairness. We find the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17575" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.076, + 0.855, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.313, + 0.678, + 0.328 + ], + "angle": 0, + "content": "Figure 5: Qualitative comparisons on the ScanNet dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.209, + 0.343, + 0.451, + 0.355 + ], + "angle": 0, + "content": "Average over Scene 101 and Scene 241" + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.355, + 0.48, + 0.47 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPSAlex↓#Param.(M) ↓
SRN [31]18.250.5920.586-
NeRF [20]22.990.6200.369-
NSVF [19]25.480.6880.301-
Point-NeRF[39]25.920.8910.273159.01
TensoRF-CP[6]27.540.7510.3280.97
TensoRF-VM[6]28.610.7870.29049.92
Ours-4829.050.7920.24312.82
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.473, + 0.471, + 0.519 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison on two scenes in the ScanNet dataset [9]. Point-NeRF, TensoRF-CP, TensoRFVM and Ours-48 are optimized for \\(100\\mathrm{k}\\) steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.771 + ], + "angle": 0, + "content": "Scannet data has many holes in the provided mesh geometry, while methods, such as NSVF and Point-NeRF, require accurate initial geometry; though Point-NeRF can potentially fix them with it point growing technique as shown in their original paper, it is not able to fully address them in 100k step optimization and lead to holes in their final rendering. Our approach instead does not require very accurate coarse geometry, since our local tensors cover relatively large regions. We show the quantitative results in Tab. 3 and qualitative results in Fig. 5. Note that our method can also perform well on real scenes, achieving the highest performance in terms of PSNR and \\(\\mathrm{LPIPS}_{Alex}\\), while using the second smallest model size (the smallest one is TensoRF\\(\\mathrm{CP}_{100k}\\)). Our visual quality is also higher than the comparison methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The Tanks and Temples dataset. We also evaluate our method on another real dataset, Tanks and Temples [9] with the 5 object scenes selected by NSVF [19]. We using the very coarse geometries estimated by DVGO[32] to distribute our tensors. We follow the same experiment setting as done in TensoRF [6], optimizing our model for the same \\(30\\mathrm{k}\\) steps for fair comparisons. As is shown in" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.343, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Tab. 3, our method outperforms other methods in terms of PSNR, SSIM and LPIPS \\( {}_{Alex} \\) ,while using the second smallest model size." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.399, + 0.898, + 0.487 + ], + "angle": 0, + "content": "
ScalePSNR ↑SSIM ↑#Param.(M) ↓Time ↓
Single(0.6)32.220.9571.7518.22m
Single(0.3)32.730.9614.1521.31m
Single(0.15)31.960.95210.2028.55m
Multi(0.6, 0.3)33.110.9636.2030.12m
Multi(0.6, 0.3, 0.15)33.550.96513.5235.70m
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.49, + 0.895, + 0.534 + ], + "angle": 0, + "content": "Table 4: Ablation under different scale settings on Synthetic-NeRF dataset. We select 3 scales of tensors with cube sizes of 0.6, 0.3, and 0.15." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.651, + 0.577 + ], + "angle": 0, + "content": "6.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.78 + ], + "angle": 0, + "content": "We analyze our model in terms of different scales in Table.4, while keeping the number of components the same (here we use 48). The scale here is the size of our local tensors of each axis. We considered 3 different scales, i.e., 0.6, 0.3, and 0.15 respectively as single-scale settings and some of their combinations as multi-scale settings. Note that even with a single scale (0.3), the performance of our method can be comparable with some other methods such as iNGP [23] while ours have less than half of the model size. When increasing the number of scales or decreasing the size of local tensors, our model size will also increase. In general, there is a trade-off of our method between scales and computational consumption (time and size)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Usually, a smaller scale can lead to better performance, though our method with a scale of 0.15 does not strictly follow because we don't have high-quality input geometry to place these local tensors with a very small size. In fact, according to our per-scene breakdown results on the Synthetic-NeRF dataset (please refer to our supplemental material), single-scale(0.075) can achieve higher performance than single-scale(0.3) and single-scale(0.15) on most" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17576" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "scenes, except for ship because it has many thin structures that our coarse reconstruction does not cover." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.131, + 0.48, + 0.219 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPSAlex ↓#Param.(M) ↓
NeRF [22]25.780.8640.198-
NSVF [19]28.400.9000.153-
TensoRF-CP30k[6]27.590.8970.1440.97
TensoRF-VM30k[6]28.560.9200.12549.92
Ours-4830k28.700.9220.11314.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.222, + 0.47, + 0.268 + ], + "angle": 0, + "content": "Table 5: Quantitative comparison on scenes in the Tanks and Temples dataset [17] selected in NSVF [19]. TensoRF-CP, TensoRF-VM and Ours-48 are optimized for 30k steps." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.275, + 0.469, + 0.366 + ], + "angle": 0, + "content": "We also compare our method with a variant that uses vector-matrix (VM) decomposition [6] in each local tensor instead of CP decomposition. Please refer to Appendix.A. for more details. Also, we can achieve a higher training and inference speed without a significant performance drop, which we refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.377, + 0.195, + 0.392 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.6 + ], + "angle": 0, + "content": "In this work, we have presented a novel approach for high-quality neural scene reconstruction and photorealistic novel view synthesis. We propose a novel tensor factorization-based scene representation, which leverages CP decomposition to compactly model a 3D scene as a sparse set of multi-scale tri-vector tensors that express local radiance fields. Our representation leverages both sparsity and spatial local coherence, and leads to accurate and efficient modeling of complex scene geometry and appearance. We demonstrate that the sparse tri-vector radiance fields can achieve superior rendering quality than previous state-of-the-art neural representations, including TensoRF and iNGP, while using significantly fewer parameters." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.611, + 0.173, + 0.627 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.636, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In International Conference on Machine Learning (ICML), pages 40-49, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.693, + 0.472, + 0.733 + ], + "angle": 0, + "content": "[2] Relja Arandjelović and Andrew Zisserman. Nerf in detail: Learning to sample for view synthesis. arXiv preprint arXiv:2106.05264, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.734, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.818, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[4] J Douglas Carroll and Jih-Jie Chang. Analysis of individual differences in multidimensional scaling via an n-way generalization of \"eckart-young\" decomposition. Psychometrika, 35(3):283-319, 1970. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.636, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.16 + ], + "angle": 0, + "content": "Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), pages 333-350. Springer, 2022. 1, 2, 3, 4, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6970-6981, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.344, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5828-5839, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[10] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479–12488, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[11] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4857-4866, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 1026-1034, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[13] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2821-2830, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[14] Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan. Pytorch. Programming with TensorFlow: Solution for Edge Computing Applications, pages 87-104, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[15] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17577" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.118 + ], + "angle": 0, + "content": "reconstruction. ACM Transactions on Graphics, 36(4), 2017. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.188 + ], + "angle": 0, + "content": "[18] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[19] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. arXiv preprint arXiv:2007.11571, 2020. 2, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[20] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7210-7219, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[21] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 1, 2, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):1-15, 2022. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[25] Anton Obukhov, Mikhail Usvyatsov, Christos Sakaridis, Konrad Schindler, and Luc Van Gool. Tt-nf: Tensor train neural fields. arXiv preprint arXiv:2209.15529, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[26] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[27] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[29] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. arXiv preprint arXiv:2103.13744, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[30] Jonathan Richard Shewchuk. Tetrahedral mesh generation by delaunay refinement. In Proceedings of the fourteenth annual symposium on Computational geometry, pages 86-95, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[31] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. arXiv preprint arXiv:1906.01618, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[32] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[33] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[34] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[35] Jinglu Wang, Bo Sun, and Yan Lu. MVPnet: Multi-view point regression networks for 3D object reconstruction from a single image. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[36] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single RGB images. In European Conference on Computer Vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[37] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[38] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[39] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5438-5448, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[40] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Proc. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[41] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17578" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[42] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[43] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5449-5458, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17579" + } + ] +] \ No newline at end of file diff --git a/2023/Strivec_ Sparse Tri-Vector Radiance Fields/35b59dcb-1ce5-40a4-83d2-07192fb62870_origin.pdf b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/35b59dcb-1ce5-40a4-83d2-07192fb62870_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..954ef59da9518fb3100b36802232bcc690900a8e --- /dev/null +++ b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/35b59dcb-1ce5-40a4-83d2-07192fb62870_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab96227fdb5525b66bd0be1a113f4e4899bc01c88beca6a3a4df9d3b59d08a85 +size 1456663 diff --git a/2023/Strivec_ Sparse Tri-Vector Radiance Fields/full.md b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..269eedab3e47396df0eac5c7214c60ff69c9eb76 --- /dev/null +++ b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/full.md @@ -0,0 +1,366 @@ +# Strivec: Sparse Tri-Vector Radiance Fields + +Quankai Gao\*1 + +Qiangeng Xu*1 + +1University of Southern California + +Hao Su² + +Ulrich Neumann1 + +$^{2}$ UC San Diego + +Zexiang $\mathbf{X}\mathbf{u}^3$ + +$^{3}$ Adobe Research + +{quankaig, qiangenx, uneumann}@usc.edu + +haosu@ucsd.edu + +zexu@adobe.com + +# Abstract + +We propose Strivec, a novel neural representation that models a 3D scene as a radiance field with sparsely distributed and compactly factorized local tensor feature grids. Our approach leverages tensor decomposition, following the recent work TensoRF [6], to model the tensor grids. In contrast to TensoRF which uses a global tensor and focuses on their vector-matrix decomposition, we propose to utilize a cloud of local tensors and apply the classic CANDE-COMP/PARAFAC (CP) decomposition [4] to factorize each tensor into triple vectors that express local feature distributions along spatial axes and compactly encode a local neural field. We also apply multi-scale tensor grids to discover the geometry and appearance commonalities and exploit spatial coherence with the tri-vector factorization at multiple local scales. The final radiance field properties are regressed by aggregating neural features from multiple local tensors across all scales. Our tri-vector tensors are sparsely distributed around the actual scene surface, discovered by a fast coarse reconstruction, leveraging the sparsity of a 3D scene. We demonstrate that our model can achieve better rendering quality while using significantly fewer parameters than previous methods, including TensoRF and Instant-NGP [23]. + +# 1. Introduction + +Representing 3D scenes as radiance fields [22] has enabled photo-realistic rendering quality and emerged as a popular design choice in 3D vision and graphics applications. While many methods [27, 42, 3] (following NeRF [22]) purely use MLPs to represent neural fields, recent works, like TensorRF [6] and Instant-NGP [23], have demonstrated the advantages of using shared global feature encoding for radiance field modeling, in terms of speed, compactness, and quality. However, these methods share and assign neural features uniformly in a scene (with tensor factors or hash tables), assuming the scene content is + +![](images/641a53b3096cce5fbd019327af205a915e6d635ed962cb19b3599d951e341efc.jpg) +TensoRF-VM-192 + +![](images/c9248d0b165cc49ce9fd12f35ab0f38afb3a64f213adf6a2acffc1ed73378d83.jpg) +INGP + +![](images/9874ff302c4eaddfa71b9f908f3d0f669e9a9f69efe39d6d861d2a5e281b7859.jpg) +Ours-48 + +![](images/21c797888d26fac8ebf3e801526d618fcfb0186bac9b60e86fd494054c85ae9d.jpg) +Reference + +![](images/c828c2bae543459a9bb5ef72640bff7345c2444a151c9a411940f55a7ed31f71.jpg) +Figure 1: We compare with previous methods in terms of rendering quality (PSNR) and model capacity (number of parameters) on the NeRF Synthetic dataset on the bottom. Our method and TensorRF are shown with different model sizes. Our approach consistently achieve better rendering quality with fewer model parameters than TensorRF, as well as other methods like iNGP. On the top, we show one example of visual comparisons of the mic scene that has challenging fine-grained geometric structures, where our approach captures most of the details and is the closest to the reference. Note that the results of NeRF and Point-NeRF use 200k optimization steps while the rest use only 30k steps. + +equally complex over the entire space, which can be inefficient (requiring high model capacity) to accurately model intricate local scene details (see Fig.1). + +We aim to accurately and compactly model a 3D scene and reproduce the complex local details. To this end, we propose Strivec, a novel neural scene representation that utilizes sparsely distributed and compactly factorized local tensor grids to model a volumetric radiance field for high- + +quality novel view synthesis. As shown in Fig.1, our approach is able to accurately model the complex scene structures that are not recovered well by previous methods. More importantly, our superior rendering quality is achieved with much less model capacity. + +In particular, we base our model on TensoRF [6], a recent approach that leverages tensor factorization in radiance field modeling. It is fast, compact, and of high rendering quality. TensoRF applies CP and vector-matrix (VM) decomposition techniques to factorize a field into vectors and matrices and model the entire scene as a global factorized tensor. Instead of a single global tensor, we leverage a sparse set of multiple small local tensors distributed around the scene surface for more efficient scene modeling. Specifically, each of our tensors represents a local radiance field inside its local bounding box and is compactly modeled with factorized triple vectors based on the CP decomposition. + +Note that the global CP decomposition in TensoRF has led to a highly compact model but cannot achieve comparable rendering quality to their VM decomposition. This is because a tri-vector CP component is rank-one, while a global feature grid of an entire 3D scene is often complex and of high rank, requiring a large (impractical) number of CP components for high accuracy. TensoRF addresses this by introducing matrix factors in their VM decomposition, essentially increasing the rank of each tensor component. Our model instead consists of multiple small tensor grids, exploiting local spatial commonalities in a scene. Compared to a global tensor, our local tensor is less complex and of much lower rank, thus effectively reducing the required number of CP components (per tensor) and enabling practical high-quality radiance field reconstruction with highly compact tri-vector factors. Our local tri-vector tensors can lead to superior rendering quality and compactness over TensoRF's VM model (see Fig. 1). We also observe that our local tensors are generally more robust than a global tensor against the orientation of spatial axes (which can affect the rank of a tensor and thus affects the quality; see Fig. 2). + +Importantly, adopting local tensors (instead of a global one) also brings us the flexibility to allocate neural features according to the actual scene distribution, enabling more efficient scene modeling and better usage of model parameters than a global representation. To do so, we pre-acquire coarse scene geometry – that can be easily achieved via a fast RGB $\sigma$ volume reconstruction (like DVGO [32]) or multi-view stereo (like Point-NeRF [39]) – to directly distribute local tensors around the actual scene surface, leading to a sparse scene representation that avoids unnecessarily modeling the empty scene space. Note that while previous methods have also leveraged sparse representations (with voxels [19, 41] or points [39]) of radiance fields, their local features are modeled and optimized independently. Our model instead correlates a group of local features inside a + +local box and compactly express them with triple vectors, uniquely exploiting the local spatial coherence along axes and imposing local low-rank priors in the feature encoding via tensor factorization. Moreover, unlike previous sparse representations that only use a single-scale feature grid or point cloud, we distribute multi-scale local tensors to effectively model the scene geometry and appearance at multiple scales in a hierarchical manner. In particular, for an arbitrary 3D location, we aggregate the neural features from its neighboring tri-vector components at all scales and decode the volume density and view-dependent color from the aggregated features for radiance field rendering. + +Our approach takes the best of previous local and global radiance field representations. Compared with global representations like TensoRF and Instant-NGP, our model takes advantage of the sparsity of a scene more directly; compared with local representations like Plenoxels and Point-NeRF, our model makes use of the local smoothness and coherence of scene geometry and appearance. As shown in our experimental results on both synthetic and real datasets, our model is able to achieve state-of-the-art rendering quality on these datasets, outperforming previous methods, including TensoRF and Instant-NGP, while using significantly fewer model parameters, demonstrating the superior representational power of our model. + +# 2. Related Work + +Scene representations. To represent a 3D scene, traditional and learning-based methods have studied various representations, such as depth map [13, 18], mesh [15, 36, 30], point cloud [28, 1, 35] and implicit function [7, 21, 24, 40]. In recent years, continuous neural field representations stand out in various 3D tasks such as single-view 3D reconstruction [38, 11], surface completion [8, 26], multi-view reconstruction [24] and novel view synthesis [22, 20]. Compared with traditional discrete representations, a continuous field have no limitation on spatial resolution, e.g., volume resolution or the number of points. It can also naturally be represented by neural networks, such as an MLP, which are known for approximating complex functions well. + +Neural field representations. Specifically, NeRF [22] represents a 3D scene as a radiance field with a global coordinate MLP, which models geometry, lighting and texture information jointly, leading to photo-realistic rendering quality in novel view synthesis. Apart from its advantage, purely MLP-based NeRF models [3, 34] in general suffer from inefficiency [2] when modeling highly complex or large-scale scenes, due to limited model capacity, slow optimization speed, and the cost of modeling empty space. + +To model radiance fields more efficiently, recent works have explored combining neural fields with various traditional 3D representations, including voxels [19, 41, 32, 43] + +and points [39]. Low-rank representations such as triplane [5, 10] and tensor decomposition [6, 25] have also been studied. In particular, DVGO [32] and Plenoxels [41] respectively use dense and sparse voxels with neural features for radiance field modeling. While being efficient to optimize, these localized feature grid-based representations lead to a large model size and can face overfitting issues when the features are of very high resolution. Consequently, DVGO can also work with a low-resolution grid and Plenoxels requires additional spatial regularization terms. On the other hand, recent works have adopted global feature encoding to express a high-resolution feature grid, including Instant-NGP [23] that hashes spatial features into multi-scale hash tables and TensorRF [6] that factorizes a feature grid into vector and matrix factors. These global feature encoding methods exploit the spatial correlation across the entire scene space, leading to fast and compact reconstruction and surpassing previous MLP-based or grid-based representations on rendering quality. However, similar to NeRF, such global representation can also be limited by its model capacity when representing highly complex or large-scale content. + +Our approach instead combines local and global representations. Our tri-vector fields are sparsely distributed in the scene, similar to local representations (like plenoxels and Point-NeRF); meanwhile, features in each field are represented by tri-vector components shared across the local region as done in TensoRF, exploiting spatial feature commonalities. Our model leverages both spatial sparsity and coherence, leading to much higher compactness and better reconstruction quality than previous local and global representations (see Tab. 1). + +Relevant to our work, previous methods, such as KiloNeRF [29] and BlockNeRF [33] have also utilized multiple local MLPs to represent a scene. Specifically, KiloNeRF focuses and speeding up NeRF and their rendering quality is sacrificed; BlockNeRF essentially uses multiple NeRFs to increase the total model capacity. Instead of pure MLPs, our work is built upon tensor factorization-based feature encoding as done in TensoRF [6], and we in fact achieve superior rendering quality while decreasing the model capacity. + +# 3. Sparse Tri-Vector Field Representation + +We now present our novel radiance field representation. In essence, our model consists of a cloud of small local trivector tensors at multiple scales, designed to leverage both sparsity and multi-scale spatial coherence (see Fig. 2). + +Let $\mathcal{T} = \{\tau_n|n = 1,\dots,N\}$ denote a cloud of tri-vector tensors. Each local tensor $\tau$ is located at $p$ , covering a local cuboid space $\omega$ with an edge length of $l$ . This cloud of tri + +vector tensors represents a radiance field for the 3D space: + +$$ +\Omega = \bigcup_ {n = 1} ^ {N} \omega_ {n}. \tag {1} +$$ + +Here, each tensor $\tau$ encodes a local multi-channel feature grid that includes a (single-channel) density grid $A_{\sigma}$ and a (multi-channel) appearance grid $A_{c}$ , similar to the tensor grid in TensoRF [6]. In contrast to using a single global tensor in TensoRF [6], we model the volume density and view-dependent colors with multiple local tensors. In particular, for an arbitrary location $\chi \in \Omega$ , we select $M$ nearest tensors that cover $\chi$ . Across the selected tensors, we aggregate the extracted density and appearance features recovered by their tri-vector factors for radiance field property regression, where the volume density $\sigma$ is directly obtained after the aggregation and the view-dependent color $c$ is regressed by a small MLP $\psi$ along with the viewing direction $\mathbf{d}$ . The continuous radiance field can be expressed as: + +$$ +\sigma_ {\chi}, c _ {\chi} = A _ {\sigma} \left(\left\{\mathcal {G} ^ {\sigma} (\chi) \right\}\right), \psi \left(A _ {c} \left(\left\{\mathcal {G} ^ {c} (\chi) \right\}\right), \mathbf {d}\right). \tag {2} +$$ + +# 3.1. Local tri-vector tensors. + +We apply the classic Canonical polyadic (CP) decomposition [4] to model our local tensors with tri-vector components. + +CP decomposition. CP decomposition factorizes a $M$ dimension tensor $\tau \in \mathbb{R}^{I_1\times I_2\times \ldots \times I_M}$ into a linear combination of $R$ rank-1 tensors: + +$$ +\tau = \sum_ {r = 1} ^ {R} \lambda_ {r} \mathbf {v} _ {r} ^ {0} \otimes \mathbf {v} _ {r} ^ {1} \otimes \dots , \otimes \mathbf {v} _ {r} ^ {M}, \tag {3} +$$ + +where $\otimes$ denotes outer product; the weighting factor $\lambda_r$ can be absorbed into vectors $\{\mathbf{v}_r^0,\dots ,\mathbf{v}_r^M\}$ . + +Density and appearance tensors. In our case of modeling a 3D radiance field, we set the geometry grid $\mathcal{G}^{\sigma} \in \mathbb{R}^{I \times J \times K}$ as a 3D tensor. And the multi-channel appearance grid $\mathcal{G}^c \in \mathbb{R}^{I \times J \times K \times P}$ corresponds to a 4D tensor. The fourth appearance mode is of lower dimension (compared with the spatial modes), representing the final dimension of the features sent to the MLP decoder network. + +According to Eqn.3, we factorize each tensor's feature grids, $\mathcal{G}^{\sigma}$ and $\mathcal{G}^{c}$ , by CP decomposition: + +$$ +\mathcal {G} ^ {\sigma} = \sum_ {r = 1} ^ {R _ {\sigma}} \mathcal {A} _ {\sigma , r} = \sum_ {r = 1} ^ {R _ {\sigma}} \mathbf {v} _ {\sigma , r} ^ {X} \otimes \mathbf {v} _ {\sigma , r} ^ {Y} \otimes \mathbf {v} _ {\sigma , r} ^ {Z}, \tag {4} +$$ + +$$ +\mathcal {G} ^ {c} = \sum_ {r = 1} ^ {R _ {c}} \mathcal {A} _ {c, r} \otimes \mathbf {b} _ {r} = \sum_ {r = 1} ^ {R _ {c}} \mathbf {v} _ {c, r} ^ {X} \otimes \mathbf {v} _ {c, r} ^ {Y} \otimes \mathbf {v} _ {c, r} ^ {Z} \otimes \mathbf {b} _ {r}, \tag {5} +$$ + +![](images/14f89b20e162f2eae18ac0949108c925d2dc223ca3550182ad5e2a226d8d28f7.jpg) +Figure 2: Overview of our Sparse Tri-Vector Radiance Fields. We distribute our local tensors based on a coarse geometry estimated by a fast $\mathrm{RGB}\sigma$ volume reconstruction as done in DVGO [32]. Here, we show our model running under $S = 2$ scales. Each local tensor is factorized as axis-aligned triple based on CP decomposition. For any shading point $\chi$ , we extract and evaluate features in each local tensor, according to the factorization (Sec. 3.1). Then, we aggregate these features among nearby tensors (Sec. 3.2) and across different scales (Sec. 3.3). Finally, the density and color are decoded (Sec. 3.4) and used by volume rendering (Sec.4). + +Here $R_{\sigma}$ and $R_{c}$ denote numbers of component; $\mathcal{A}_{\sigma,r}$ and $\mathcal{A}_{c,r}$ are the component tensors that are factorized spatially; $\mathbf{v}_{\sigma,r}^{X}, \dots, \mathbf{v}_{c,r}^{X}, \dots$ are the 1D vectors with resolution $I, J, K$ , modeling scene geometry and appearance along $X, Y, Z$ axis; $R_{\sigma}$ and $R_{c}$ are the component numbers; $\mathbf{b}_r$ expresses the feature dimension. + +As done in TensoRF [6], we stack all feature-mode vectors $\mathbf{b}_r$ as columns together, which ends up a $P\times R_{c}$ appearance matrix $\mathbf{B}$ . This matrix models the appearance feature variations of the tensor and functions like a appearance dictionary. Note that naively following CP decomposition like TensoRF will assign a different appearance matrix for every local tensor. Instead, we propose to utilize a global appearance matrix $\mathbf{B}^c$ shared across the entire cloud of local tensors, leading to a global appearance dictionary that explains the color correlations across scene. This further improves both the computational efficiency and model compactness of our model. + +Therefore, each of our local tensors is represented by their unique local tri-vector factors $\mathbf{v}_r^X$ , $\mathbf{v}_r^Y$ , $\mathbf{v}_r^Z$ . + +Feature evaluation. To achieve a continuous field, we consider trilinear interpolation when evaluating the tensor grid features. For a location $\chi$ , we first compute its relative position $\tilde{\chi}$ to the selected tensor located at $p$ : + +$$ +\tilde {x}, \tilde {y}, \tilde {z} = x - p _ {x}, y - p _ {y}, z - p _ {z}. \tag {6} +$$ + +Then, for example, to get $\mathcal{A}_{\sigma,r}$ at $(\tilde{x},\tilde{y},\tilde{z})$ , we can compute and trilinearly interpolate eight $\mathcal{A}_{\sigma,r}$ on the corners. As mentioned in [6], applying linear interpolation on each + +vector first is mathematically equivalent and can reduce the computation cost. Under the rule of outer product, we have $\mathcal{A}_{r,i,j,k} = \mathbf{v}_{r,i}^{X}\mathbf{v}_{r,j}^{Y}\mathbf{v}_{r,k}^{Z}$ , then the interpolated density features at location $\chi$ are: + +$$ +\mathcal {G} ^ {\sigma} (\chi) = \sum_ {r} \mathbf {v} _ {\sigma , r} ^ {X} (\tilde {x}) \mathbf {v} _ {\sigma , r} ^ {Y} (\tilde {y}) \mathbf {v} _ {\sigma , r} ^ {Z} (\tilde {z}) = \sum_ {r} \mathcal {A} _ {\sigma , r} (\tilde {\chi}), \tag {7} +$$ + +where $\mathbf{v}_{\sigma,r}^{X}(\tilde{x})$ is $\mathbf{v}_{\sigma,r}^{X}$ 's linearly interpolated value at $(\tilde{x})$ along its $X$ axis. Here, $\mathcal{G}^{\sigma}(\chi)$ is a scalar. + +Similarly, the interpolated appearance features can be computed as: + +$$ +\begin{array}{l} \mathcal {G} ^ {c} (\chi) = \sum_ {r} \mathbf {v} _ {c, r} ^ {X} (\tilde {x}) \mathbf {v} _ {c, r} ^ {Y} (\tilde {y}) \mathbf {v} _ {c, r} ^ {Z} (\tilde {z}) \mathbf {b} _ {r} (8) \\ = \sum_ {r} \mathcal {A} _ {c, r} (\tilde {\chi}) \mathbf {b} _ {r} (9) \\ = \mathbf {B} \cdot \left(\oplus \left[ \mathcal {A} _ {c, r} \right] _ {r}\right), (10) \\ \end{array} +$$ + +where $\oplus$ denotes concatenation, $\cdot$ denotes dot product. The appearance feature $\mathcal{G}^c (\chi)\in \mathbb{R}^P$ is a vector. + +# 3.2. Feature aggregation. + +We propose to aggregate the features from $M$ neighboring tensors to jointly model the volume density and appearance for each 3D location $\chi$ . In particular, inspired by Point-NeRF, we leverage an inverse distance-based weighting function to directly aggregate the multi-tensor features. Specifically, this weight can be expressed by + +$$ +w _ {m} = \frac {1}{\| p _ {m} - \chi \|}. \tag {11} +$$ + +With this weight function, we directly obtain the density feature via the weighted sum: + +$$ +f ^ {\sigma} (\chi) = \frac {1}{\sum w _ {m}} \sum_ {m = 1} ^ {M} w _ {m} \mathcal {G} _ {m} ^ {\sigma} (\chi). \tag {12} +$$ + +Similarly, the appearance feature aggregation can be expressed in a similar way, while using the shared appearance matrix (as described in Sec. 3.1) across local tensors: + +$$ +\begin{array}{l} f ^ {c} (\chi) = \frac {1}{\sum w _ {m}} \sum_ {m = 1} ^ {M} w _ {m} \mathcal {G} _ {m} ^ {c} (\chi) (13) \\ = \frac {1}{\sum w _ {m}} \sum_ {m = 1} ^ {M} w _ {m} \mathbf {B} ^ {c} \cdot \left(\oplus \left[ \mathcal {A} _ {c, r} \right] _ {r}\right) (14) \\ = \frac {1}{\sum w _ {m}} \mathbf {B} ^ {c} \cdot \left(\sum_ {m = 1} ^ {M} w _ {m} \left(\oplus \left[ \mathcal {A} _ {c, r} \right] _ {r}\right)\right). (15) \\ \end{array} +$$ + +Note that owing to sharing the appearance matrix across tensors, we reduce the computational complexity from $O(M \cdot P \cdot R_c)$ in Eqn.14, to $O((M + P) \cdot R_c)$ in Eqn.15. + +# 3.3. Multi-scale tri-vector fields. + +Complex 3D scenes often contain multi-frequency geometry and appearance details. This motivates us to build multi-scale tensor clouds to discover the local geometry and appearance commonalities at multiple scales. Our final radiance field is modeled by multiple tri-vector tensor clouds at $S$ different scales. Different clouds consist of tensors with different resolutions. + +To regress the density and appearance at a location $\chi$ , we gather the density and appearance features from a set of tensor clouds that cover $\chi$ , $\{\mathcal{T}_s|1\leq s\leq S,\chi \in \Omega_s\}$ . Please note that tensor clouds of certain scales might not cover the location, so that $\| \{\mathcal{T}_s\} \| \leq S$ . We simply compute the mean features across these scales: + +$$ +f ^ {\sigma} (\chi) = \frac {1}{\| \left\{\mathcal {T} _ {s} \right\} \|} \sum_ {s} f _ {s} ^ {\sigma} (\chi), \tag {16} +$$ + +$$ +f ^ {c} (\chi) = \frac {1}{\| \left\{\mathcal {T} _ {s} \right\} \|} \sum_ {s} f _ {s} ^ {c} (\chi). \tag {17} +$$ + +Note that $f^{\sigma}(\chi)$ and $f^{c}(\chi)$ are the final density and appearance features we aggregate across multiple scales and multiple neighboring tensors. + +# 3.4.Decoding. + +We apply softplus activation on the density feature $f^{\sigma}(\chi)$ to obtain the final volume density and regress the view-dependent color by sending the appearance feature $f^{c}(\chi)$ and the viewing direction $\mathbf{d}$ to the MLP decoder $\psi$ . + +# 4. Rendering and Reconstruction + +Volume Rendering We evaluates each pixel's color with physically-based volume rendering via differentiable ray marching. Following NeRF [22], we sample $Q$ shading points at $\{\chi_q\mid q = 1,\dots,Q\}$ along the ray, and accumulate radiance by density: + +$$ +c = \sum_ {q = 1} ^ {Q} T _ {q} \left(1 - \exp \left(- \sigma_ {q} \delta_ {q}\right)\right) c _ {q}, \tag {18} +$$ + +$$ +T _ {q} = \exp (- \sum_ {t = 1} ^ {q - 1} \sigma_ {t} \delta_ {t}). +$$ + +$\sigma_{q}$ and $c_{q}$ are the density and radiance at shading points; $\delta_t$ is the marching distance per step; $T$ denotes transmittance. + +Distributing local tensors. First of all, to better leverage the sparsity of a scene, we first obtain a geometric prior that roughly covers the scene geometry. The geometric prior can be in any commonly-used form, e.g., point cloud, occupancy grid, octree, or mesh vertices. Then we can uniformly distribute tensors in the spatial neighborhood of the geometry. For a multi-scale model, each of the scale will be distributed independently. For most of our results, we quickly optimize a coarse RGBA volume from the multiview images and use the optimized occupancy grid as the geometry prior, as done in DVGO [32], which finishes in seconds. + +To maintain training stability and speed, each tensor $\tau$ 's position $p$ and coverage $\omega$ is fixed once determined. We also initialize the $3(R_{\sigma} + R_c)$ vectors $(\mathbf{v}_{\sigma,r}^{X}, \dots, \mathbf{v}_{c,r}^{X}, \dots)$ of each tensor by normal distribution. For each scale $s$ , a $P \times R_{c}$ appearance matrix $\mathbf{B}_s^c$ is shared by all tri-vector tensors of that scale. Specifically, “ $\mathbf{B}^c \cdot ()$ ” in Eqn.15 can be efficiently implemented as a fully-connected neural layer. Therefore, $\mathbf{B}^c$ for each scale and a global appearance MLP $\psi$ will be implemented as neural networks and initialized by default methods [12]. + +Optimization and objectives. Given a set of multi-view RGB images with camera poses, the sparse tri-vector radiance field is per-scene optimized to reconstruct the radiance fields, under the supervision of the ground truth pixel colors. Following the volume rendering equation 18, the L2 rendering loss can be past back to the global MLP and aggregated features, then, all the way to the appearance matrices and the feature vectors of local tensors. + +We apply a rendering loss to supervise the reconstruction and also apply an $L1$ regularization loss on density feature vectors $\mathbf{v}_{\sigma,r}$ to promote geometry sparsity and to avoid + +
MethodBatchSizeStepsTime↓#Param.(M)↓PSNR↑SSIM↑LPIPSVgg↓LPIPSAlex↓
NeRF[20]4096300k35.0h1.2531.010.9470.081-
Plenoxels[41]5000128k11.4m194.5031.710.9580.049-
DVGO[32]500030k15.0m153.0031.950.9600.0530.035
Point-NeRF200k[39]4096200k5.5h27.7433.310.962*0.0490.027
InstantNGP[23]10k-85k30k3.9m11.6432.590.960--
TensoRF-CP[6]409630k25.2m0.9831.560.9490.0760.041
TensoRF-VM[6]409630k17.4m17.9533.140.9630.0470.027
Ours-24409630k34.3m7.0733.240.9630.0460.026
Ours-48409630k35.7m13.5233.550.9650.0440.025
+ +Table 1: Comparisons of our method with other radiance-based models [20, 37, 19, 39, 6, 23] on the Synthetic-NeRF dataset [20]. Ours-24 is the one with 24 components while Ours-48 is the one with 48 components. We report the corresponding rendering quality (PSNR, SSIM, and LPIPS), model capacity (number of parameters), and training time, batch size and steps. Our model achieves the best rendering quality with a compact model size. We report PointNeRF's updated SSIM. + +overfitting as done in TensoRF [6]: + +$$ +\mathcal {L} _ {r} = \| C - \tilde {C} \| _ {2} ^ {2}, \tag {19} +$$ + +$$ +\mathcal {L} _ {L 1} = \frac {1}{N} \sum_ {r} ^ {R _ {\sigma}} \| \mathbf {v} _ {\sigma , r} \|, \tag {20} +$$ + +where $\tilde{C}$ is the ground truth pixel color, $\| \mathbf{v}_{\sigma ,r}\|$ is the sum of absolute values of elements on density vectors, and $N$ is the total number of elements. The total loss is: + +$$ +\mathcal {L} = \mathcal {L} _ {r} + \alpha \mathcal {L} _ {L 1}. \tag {21} +$$ + +We set the weight of the sparsity term $\alpha$ as $1e^{-5}$ by default. + +# 5. Implementation + +To obtain coarse scene geometry, we modify the coarse density estimation introduced in [32] and get a $100^{3}$ occupancy volume in 30 seconds. We can skip this step if there exists available geometric data, e.g., the meshes in ScanNet [9], or point clouds from multiview stereo. According to the experiments, our method is not very sensitive to the initial geometry. Please refer to our supplementary material for more details. We set the default number of scales to 3. In a scene box of [-1,1], we rasterize the scene geometry (occupied voxels centers or points) onto 3 grids with different voxel sizes, e.g., $0.4^{3}, 0.2^{3}, 0.1^{3}$ . For each grid, we distribute tri-vector tensors at the center of its occupied voxels. The tensor spatial coverage of these 3 scales is $0.6^{3}, 0.3^{3}, 0.15^{3}$ , respectively. For each scale, we query $M = 4$ nearby tensors. Following [32], our feature decoding network $\psi$ is a 2-layer MLP with 128 channels. For each scale, its appearance matrix $\mathbf{B}^c$ is implemented by a single linear layer of 27 channels. + +We implement the framework in PyTorch [14] with customized CUDA operations. During optimization, we adopt the coarse to fine strategy in [6], linearly up-sample the vectors' dimension $(I,J,K)$ from 29 to 121 for scale one, + +15 to 61 for scale two, and 7 to 31 for scale three. The up-sampling happens at step 2000, 3000, 4000, 5500, and 7000. We use the Adam optimizer [16] with initial learning rates of 0.02 for vectors and 0.001 for networks. On a single 3090 RTX GPU, we train each model for 30000 steps while each batch contains 4096 rays. Please find more details in the supplemental materials. + +# 6. Experiments + +# 6.1. Evaluation on the NeRF Synthetic Dataset. + +We first evaluate our method on the Synthetic-NeRF dataset [20] and the quantitative results compared with other methods are reported in Tab.1, including NeRF [22], Plenoxels [41], DVGO [32], Point-NeRF [39], iNGP [23], and TensoRF [6]. We report our models of two different model sizes with different numbers of components; both settings are with the same 3 scales of local tensors. + +Our approach achieves the best averaged PSNR, $\mathrm{LPIPS}_{Vgg}$ and $\mathrm{LPIPS}_{Alex}$ in all the methods, leading to superior visual quality as shown in Fig. 3 Meanwhile, our high rendering quality is achieved with a compact model size. When compared with local voxel-based representations, such as Plenoxels and DVGO, our approach are significantly better. + +On the other hand, global featuring encoding-based methods, like iNGP and TensorRF, are known for their high compactness and can achieve higher rendering quality than local voxel-based methods. Nonetheless, our method can still outperform them. Note that, even our smaller model (Ours-24) leads to better rendering quality than both iNGP and TensorRF that leverage global feature encoding, while our model uses significantly fewer parameters (about $60\%$ and $40\%$ of the size of iNGP and TensorRF). This clearly demonstrates the high visual quality and high compatness of our model with our sparsely distributed tri-vector tensors. + +In all the baseline methods, Point-NeRF is able to achieve relatively higher rendering quality than others. + +![](images/557987c07a49e8537a71ac83d5031db513debc16f6efd48474ab5095d1bceb33.jpg) +Figure 3: Qualitative comparisons on the NeRF Synthetic dataset [22]. + +However, this is enabled by optimizing their model for 300k steps with a long period of 5.5 hours. In contrast, our method achieves higher quality with significantly fewer optimization steps (only 30k) and optimization time (about 36 min). As expected, our model is slower to optimize than TensoRF due to the additional costs of multi-tensor aggregation. However, though speed is not our focus in this work, our model can still converge quickly and lead to significantly faster reconstruction than MLP-based methods, such as NeRF, as well as Point-NeRF that is point-based. + +Performance w.r.t. rotation. We observe that tensor factorization-based methods can be sensitive to the orientation of the coordinate frames, since axis-aligned features are used; in essence, this is because the rank of a sparse tensor is sensitive to rotation, as shown in Fig. 4. Especially, this can benefit reconstruction on synthetic synthetic scenes where the objects are perfectly aligned with the axes, e.g. the lego scene in the NeRF synthetic data. However, we find that our method based on local tensors are more robust against the orientation of the axes than a global TensoRF. In particular, we compare our method with TensoRF in Tab.2 with different degrees of rotation around the Z axis on two scenes, lego (which is strongly aligned with axes) and chair (which is less aligned and thus less affected). As shown in the table, while both methods are affected by the rotation, our method has much smaller drops of PSNRs. + +
chair / legorot 0°rot 5°rot 45°
TensoRF-CP33.60 / 34.5032.90 / 29.7932.50 / 28.57
TensoRF-VM35.76 / 36.4634.91 /32.5334.55 / 32.31
Ours-4835.88 / 36.5235.72 / 35.3735.64 / 34.97
+ +Table 2: Comparisons of our method with TensoRF [6] on the chair and lego scenes of Synthetic-NeRF dataset [20] when considering rotation of different angles around $z$ -axis. + +![](images/c26d2db69956d3a5b1b4a95ac446048372848aee999a53e74b76ce3c70884e1f.jpg) +Figure 4: A toy example to illustrate the TensoRF-CP with global decomposition in (left) axis-aligned and (right) non-axis-aligned situations. The bottom shows the grid values. In axis-aligned case, only 1 component is needed to represent the scene (vector bases recover grid values by outer product). In non-axis-aligned case, however, 3 components are needed because the rank of matrix changes from 1 to 3 after scene rotation. While our design with local low-rank tensors can alleviate this issue. + +![](images/a6048dfe4579b808ce8a9e6f269bbe6f02625dc5ef5d516011b79cdcdd8bf9fc.jpg) + +# 6.2. Evaluation on the real datasets. + +The ScanNet dataset. We evaluate our method on the real dataset, ScanNet [9] with the two scenes selected by NSVF [19], and compare with other methods. We follow the same experiment setting as done in NSVF [19], using the provided mesh to distribute our tensors, and optimize our model, TensoRF for the same 100k steps for fair comparisons. Please note that Point-NeRF uses all scanned depth images as initial geometry instead of meshes. Therefore, we also obtain the results of Point-NeRF 100k steps from the authors, using the provided mesh for fairness. We find the + +![](images/b464a67390f1ad6718da847a507ad8400deef103299bf826d2da4b8e922115ca.jpg) +Figure 5: Qualitative comparisons on the ScanNet dataset. + +Average over Scene 101 and Scene 241 + +
PSNR ↑SSIM ↑LPIPSAlex↓#Param.(M) ↓
SRN [31]18.250.5920.586-
NeRF [20]22.990.6200.369-
NSVF [19]25.480.6880.301-
Point-NeRF[39]25.920.8910.273159.01
TensoRF-CP[6]27.540.7510.3280.97
TensoRF-VM[6]28.610.7870.29049.92
Ours-4829.050.7920.24312.82
+ +Table 3: Quantitative comparison on two scenes in the ScanNet dataset [9]. Point-NeRF, TensoRF-CP, TensoRFVM and Ours-48 are optimized for $100\mathrm{k}$ steps. + +Scannet data has many holes in the provided mesh geometry, while methods, such as NSVF and Point-NeRF, require accurate initial geometry; though Point-NeRF can potentially fix them with it point growing technique as shown in their original paper, it is not able to fully address them in 100k step optimization and lead to holes in their final rendering. Our approach instead does not require very accurate coarse geometry, since our local tensors cover relatively large regions. We show the quantitative results in Tab. 3 and qualitative results in Fig. 5. Note that our method can also perform well on real scenes, achieving the highest performance in terms of PSNR and $\mathrm{LPIPS}_{Alex}$ , while using the second smallest model size (the smallest one is TensoRF $\mathrm{CP}_{100k}$ ). Our visual quality is also higher than the comparison methods. + +The Tanks and Temples dataset. We also evaluate our method on another real dataset, Tanks and Temples [9] with the 5 object scenes selected by NSVF [19]. We using the very coarse geometries estimated by DVGO[32] to distribute our tensors. We follow the same experiment setting as done in TensoRF [6], optimizing our model for the same $30\mathrm{k}$ steps for fair comparisons. As is shown in + +Tab. 3, our method outperforms other methods in terms of PSNR, SSIM and LPIPS ${}_{Alex}$ ,while using the second smallest model size. + +
ScalePSNR ↑SSIM ↑#Param.(M) ↓Time ↓
Single(0.6)32.220.9571.7518.22m
Single(0.3)32.730.9614.1521.31m
Single(0.15)31.960.95210.2028.55m
Multi(0.6, 0.3)33.110.9636.2030.12m
Multi(0.6, 0.3, 0.15)33.550.96513.5235.70m
+ +Table 4: Ablation under different scale settings on Synthetic-NeRF dataset. We select 3 scales of tensors with cube sizes of 0.6, 0.3, and 0.15. + +# 6.3. Ablation study + +We analyze our model in terms of different scales in Table.4, while keeping the number of components the same (here we use 48). The scale here is the size of our local tensors of each axis. We considered 3 different scales, i.e., 0.6, 0.3, and 0.15 respectively as single-scale settings and some of their combinations as multi-scale settings. Note that even with a single scale (0.3), the performance of our method can be comparable with some other methods such as iNGP [23] while ours have less than half of the model size. When increasing the number of scales or decreasing the size of local tensors, our model size will also increase. In general, there is a trade-off of our method between scales and computational consumption (time and size). + +Usually, a smaller scale can lead to better performance, though our method with a scale of 0.15 does not strictly follow because we don't have high-quality input geometry to place these local tensors with a very small size. In fact, according to our per-scene breakdown results on the Synthetic-NeRF dataset (please refer to our supplemental material), single-scale(0.075) can achieve higher performance than single-scale(0.3) and single-scale(0.15) on most + +scenes, except for ship because it has many thin structures that our coarse reconstruction does not cover. + +
PSNR ↑SSIM ↑LPIPSAlex ↓#Param.(M) ↓
NeRF [22]25.780.8640.198-
NSVF [19]28.400.9000.153-
TensoRF-CP30k[6]27.590.8970.1440.97
TensoRF-VM30k[6]28.560.9200.12549.92
Ours-4830k28.700.9220.11314.11
+ +Table 5: Quantitative comparison on scenes in the Tanks and Temples dataset [17] selected in NSVF [19]. TensoRF-CP, TensoRF-VM and Ours-48 are optimized for 30k steps. + +We also compare our method with a variant that uses vector-matrix (VM) decomposition [6] in each local tensor instead of CP decomposition. Please refer to Appendix.A. for more details. Also, we can achieve a higher training and inference speed without a significant performance drop, which we refer to the supplementary material. + +# 7. Conclusion + +In this work, we have presented a novel approach for high-quality neural scene reconstruction and photorealistic novel view synthesis. We propose a novel tensor factorization-based scene representation, which leverages CP decomposition to compactly model a 3D scene as a sparse set of multi-scale tri-vector tensors that express local radiance fields. Our representation leverages both sparsity and spatial local coherence, and leads to accurate and efficient modeling of complex scene geometry and appearance. We demonstrate that the sparse tri-vector radiance fields can achieve superior rendering quality than previous state-of-the-art neural representations, including TensoRF and iNGP, while using significantly fewer parameters. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In International Conference on Machine Learning (ICML), pages 40-49, 2018. 2 +[2] Relja Arandjelović and Andrew Zisserman. Nerf in detail: Learning to sample for view synthesis. arXiv preprint arXiv:2106.05264, 2021. 2 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 1, 2 +[4] J Douglas Carroll and Jih-Jie Chang. Analysis of individual differences in multidimensional scaling via an n-way generalization of "eckart-young" decomposition. Psychometrika, 35(3):283-319, 1970. 1, 3 +[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J + +Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022. 3 +[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), pages 333-350. Springer, 2022. 1, 2, 3, 4, 6, 7, 8, 9 +[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6970-6981, 2020. 2 +[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5828-5839, 2017. 6, 7, 8 +[10] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479–12488, 2023. 3 +[11] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4857-4866, 2020. 2 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 1026-1034, 2015. 5 +[13] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2821-2830, 2018. 2 +[14] Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan. Pytorch. Programming with TensorFlow: Solution for Edge Computing Applications, pages 87-104, 2021. 6 +[15] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In European Conference on Computer Vision (ECCV), 2018. 2 +[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene + +reconstruction. ACM Transactions on Graphics, 36(4), 2017. 9 +[18] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2016. 2 +[19] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. arXiv preprint arXiv:2007.11571, 2020. 2, 6, 7, 8, 9 +[20] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7210-7219, 2021. 2, 6, 7, 8 +[21] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 1, 2, 5, 6, 7, 9 +[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):1-15, 2022. 1, 3, 6, 8 +[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[25] Anton Obukhov, Mikhail Usvyatsov, Christos Sakaridis, Konrad Schindler, and Luc Van Gool. Tt-nf: Tensor train neural fields. arXiv preprint arXiv:2209.15529, 2022. 3 +[26] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2 +[27] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1 +[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[29] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. arXiv preprint arXiv:2103.13744, 2021. 3 + +[30] Jonathan Richard Shewchuk. Tetrahedral mesh generation by delaunay refinement. In Proceedings of the fourteenth annual symposium on Computational geometry, pages 86-95, 1998. 2 +[31] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. arXiv preprint arXiv:1906.01618, 2019. 8 +[32] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 2, 3, 4, 5, 6, 8 +[33] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 3 +[34] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2 +[35] Jinglu Wang, Bo Sun, and Yan Lu. MVPnet: Multi-view point regression networks for 3D object reconstruction from a single image. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2 +[36] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single RGB images. In European Conference on Computer Vision (ECCV), 2018. 2 +[37] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6 +[38] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 2 +[39] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5438-5448, 2022. 2, 3, 6, 8 +[40] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Proc. NeurIPS, 2020. 2 +[41] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 2, 3, 6 + +[42] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1 +[43] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5449-5458, 2022. 2 \ No newline at end of file diff --git a/2023/Strivec_ Sparse Tri-Vector Radiance Fields/images.zip b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..416ec5a95684d60530e1917335ae2872ef4e691e --- /dev/null +++ b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a7ea33761dd8e835309cff13a07c432d4fdad4fa8514d9b74a92da8cdc45317 +size 558447 diff --git a/2023/Strivec_ Sparse Tri-Vector Radiance Fields/layout.json b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..03ea40e4a240c859767a7865d95400bf21f2f43f --- /dev/null +++ b/2023/Strivec_ Sparse Tri-Vector Radiance Fields/layout.json @@ -0,0 +1,9545 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 164, + 103, + 430, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 103, + 430, + 121 + ], + "spans": [ + { + "bbox": [ + 164, + 103, + 430, + 121 + ], + "type": "text", + "content": "Strivec: Sparse Tri-Vector Radiance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 149, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 149, + 156 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 149, + 156 + ], + "type": "text", + "content": "Quankai Gao\\*1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 143, + 249, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 143, + 249, + 157 + ], + "spans": [ + { + "bbox": [ + 175, + 143, + 249, + 157 + ], + "type": "text", + "content": "Qiangeng Xu*1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 157, + 246, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 157, + 246, + 171 + ], + "spans": [ + { + "bbox": [ + 78, + 157, + 246, + 171 + ], + "type": "text", + "content": "1University of Southern California" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 275, + 143, + 318, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 143, + 318, + 156 + ], + "spans": [ + { + "bbox": [ + 275, + 143, + 318, + 156 + ], + "type": "text", + "content": "Hao Su²" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 343, + 143, + 428, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 143, + 428, + 156 + ], + "spans": [ + { + "bbox": [ + 343, + 143, + 428, + 156 + ], + "type": "text", + "content": "Ulrich Neumann1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 301, + 157, + 377, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 157, + 377, + 171 + ], + "spans": [ + { + "bbox": [ + 301, + 157, + 377, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 301, + 157, + 377, + 171 + ], + "type": "text", + "content": "UC San Diego" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 454, + 143, + 517, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 143, + 517, + 156 + ], + "spans": [ + { + "bbox": [ + 454, + 143, + 517, + 156 + ], + "type": "text", + "content": "Zexiang " + }, + { + "bbox": [ + 454, + 143, + 517, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\mathbf{u}^3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 429, + 157, + 515, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 429, + 157, + 515, + 171 + ], + "spans": [ + { + "bbox": [ + 429, + 157, + 515, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 429, + 157, + 515, + 171 + ], + "type": "text", + "content": "Adobe Research" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 174, + 301, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 174, + 301, + 185 + ], + "spans": [ + { + "bbox": [ + 107, + 174, + 301, + 185 + ], + "type": "text", + "content": "{quankaig, qiangenx, uneumann}@usc.edu" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 174, + 394, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 394, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 394, + 184 + ], + "type": "text", + "content": "haosu@ucsd.edu" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 410, + 174, + 487, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 174, + 487, + 184 + ], + "spans": [ + { + "bbox": [ + 410, + 174, + 487, + 184 + ], + "type": "text", + "content": "zexu@adobe.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 238, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 514 + ], + "type": "text", + "content": "We propose Strivec, a novel neural representation that models a 3D scene as a radiance field with sparsely distributed and compactly factorized local tensor feature grids. Our approach leverages tensor decomposition, following the recent work TensoRF [6], to model the tensor grids. In contrast to TensoRF which uses a global tensor and focuses on their vector-matrix decomposition, we propose to utilize a cloud of local tensors and apply the classic CANDE-COMP/PARAFAC (CP) decomposition [4] to factorize each tensor into triple vectors that express local feature distributions along spatial axes and compactly encode a local neural field. We also apply multi-scale tensor grids to discover the geometry and appearance commonalities and exploit spatial coherence with the tri-vector factorization at multiple local scales. The final radiance field properties are regressed by aggregating neural features from multiple local tensors across all scales. Our tri-vector tensors are sparsely distributed around the actual scene surface, discovered by a fast coarse reconstruction, leveraging the sparsity of a 3D scene. We demonstrate that our model can achieve better rendering quality while using significantly fewer parameters than previous methods, including TensoRF and Instant-NGP [23]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 535, + 128, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 128, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 128, + 547 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": "Representing 3D scenes as radiance fields [22] has enabled photo-realistic rendering quality and emerged as a popular design choice in 3D vision and graphics applications. While many methods [27, 42, 3] (following NeRF [22]) purely use MLPs to represent neural fields, recent works, like TensorRF [6] and Instant-NGP [23], have demonstrated the advantages of using shared global feature encoding for radiance field modeling, in terms of speed, compactness, and quality. However, these methods share and assign neural features uniformly in a scene (with tensor factors or hash tables), assuming the scene content is" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 312, + 211, + 367, + 279 + ], + "blocks": [ + { + "bbox": [ + 312, + 211, + 367, + 279 + ], + "lines": [ + { + "bbox": [ + 312, + 211, + 367, + 279 + ], + "spans": [ + { + "bbox": [ + 312, + 211, + 367, + 279 + ], + "type": "image", + "image_path": "641a53b3096cce5fbd019327af205a915e6d635ed962cb19b3599d951e341efc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 279, + 364, + 287 + ], + "lines": [ + { + "bbox": [ + 314, + 279, + 364, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 279, + 364, + 287 + ], + "type": "text", + "content": "TensoRF-VM-192" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 370, + 212, + 426, + 279 + ], + "blocks": [ + { + "bbox": [ + 370, + 212, + 426, + 279 + ], + "lines": [ + { + "bbox": [ + 370, + 212, + 426, + 279 + ], + "spans": [ + { + "bbox": [ + 370, + 212, + 426, + 279 + ], + "type": "image", + "image_path": "c9248d0b165cc49ce9fd12f35ab0f38afb3a64f213adf6a2acffc1ed73378d83.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 280, + 404, + 287 + ], + "lines": [ + { + "bbox": [ + 389, + 280, + 404, + 287 + ], + "spans": [ + { + "bbox": [ + 389, + 280, + 404, + 287 + ], + "type": "text", + "content": "INGP" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 429, + 212, + 484, + 278 + ], + "blocks": [ + { + "bbox": [ + 429, + 212, + 484, + 278 + ], + "lines": [ + { + "bbox": [ + 429, + 212, + 484, + 278 + ], + "spans": [ + { + "bbox": [ + 429, + 212, + 484, + 278 + ], + "type": "image", + "image_path": "9874ff302c4eaddfa71b9f908f3d0f669e9a9f69efe39d6d861d2a5e281b7859.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 280, + 468, + 287 + ], + "lines": [ + { + "bbox": [ + 443, + 280, + 468, + 287 + ], + "spans": [ + { + "bbox": [ + 443, + 280, + 468, + 287 + ], + "type": "text", + "content": "Ours-48" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 487, + 212, + 542, + 278 + ], + "blocks": [ + { + "bbox": [ + 487, + 212, + 542, + 278 + ], + "lines": [ + { + "bbox": [ + 487, + 212, + 542, + 278 + ], + "spans": [ + { + "bbox": [ + 487, + 212, + 542, + 278 + ], + "type": "image", + "image_path": "21c797888d26fac8ebf3e801526d618fcfb0186bac9b60e86fd494054c85ae9d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 279, + 529, + 287 + ], + "lines": [ + { + "bbox": [ + 501, + 279, + 529, + 287 + ], + "spans": [ + { + "bbox": [ + 501, + 279, + 529, + 287 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 309, + 291, + 545, + 420 + ], + "blocks": [ + { + "bbox": [ + 309, + 291, + 545, + 420 + ], + "lines": [ + { + "bbox": [ + 309, + 291, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 309, + 291, + 545, + 420 + ], + "type": "image", + "image_path": "c828c2bae543459a9bb5ef72640bff7345c2444a151c9a411940f55a7ed31f71.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 593 + ], + "type": "text", + "content": "Figure 1: We compare with previous methods in terms of rendering quality (PSNR) and model capacity (number of parameters) on the NeRF Synthetic dataset on the bottom. Our method and TensorRF are shown with different model sizes. Our approach consistently achieve better rendering quality with fewer model parameters than TensorRF, as well as other methods like iNGP. On the top, we show one example of visual comparisons of the mic scene that has challenging fine-grained geometric structures, where our approach captures most of the details and is the closest to the reference. Note that the results of NeRF and Point-NeRF use 200k optimization steps while the rest use only 30k steps." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 614, + 546, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 614, + 546, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 614, + 546, + 651 + ], + "type": "text", + "content": "equally complex over the entire space, which can be inefficient (requiring high model capacity) to accurately model intricate local scene details (see Fig.1)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "We aim to accurately and compactly model a 3D scene and reproduce the complex local details. To this end, we propose Strivec, a novel neural scene representation that utilizes sparsely distributed and compactly factorized local tensor grids to model a volumetric radiance field for high-" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 60, + 693, + 130, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 693, + 130, + 703 + ], + "spans": [ + { + "bbox": [ + 60, + 693, + 130, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 56, + 703, + 250, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 703, + 250, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 703, + 250, + 713 + ], + "type": "text", + "content": "Code and results: https://github.com/Zerg-Overmind/Strivec" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17569" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "quality novel view synthesis. As shown in Fig.1, our approach is able to accurately model the complex scene structures that are not recovered well by previous methods. More importantly, our superior rendering quality is achieved with much less model capacity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 278 + ], + "type": "text", + "content": "In particular, we base our model on TensoRF [6], a recent approach that leverages tensor factorization in radiance field modeling. It is fast, compact, and of high rendering quality. TensoRF applies CP and vector-matrix (VM) decomposition techniques to factorize a field into vectors and matrices and model the entire scene as a global factorized tensor. Instead of a single global tensor, we leverage a sparse set of multiple small local tensors distributed around the scene surface for more efficient scene modeling. Specifically, each of our tensors represents a local radiance field inside its local bounding box and is compactly modeled with factorized triple vectors based on the CP decomposition." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 280, + 288, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 288, + 532 + ], + "type": "text", + "content": "Note that the global CP decomposition in TensoRF has led to a highly compact model but cannot achieve comparable rendering quality to their VM decomposition. This is because a tri-vector CP component is rank-one, while a global feature grid of an entire 3D scene is often complex and of high rank, requiring a large (impractical) number of CP components for high accuracy. TensoRF addresses this by introducing matrix factors in their VM decomposition, essentially increasing the rank of each tensor component. Our model instead consists of multiple small tensor grids, exploiting local spatial commonalities in a scene. Compared to a global tensor, our local tensor is less complex and of much lower rank, thus effectively reducing the required number of CP components (per tensor) and enabling practical high-quality radiance field reconstruction with highly compact tri-vector factors. Our local tri-vector tensors can lead to superior rendering quality and compactness over TensoRF's VM model (see Fig. 1). We also observe that our local tensors are generally more robust than a global tensor against the orientation of spatial axes (which can affect the rank of a tensor and thus affects the quality; see Fig. 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": "Importantly, adopting local tensors (instead of a global one) also brings us the flexibility to allocate neural features according to the actual scene distribution, enabling more efficient scene modeling and better usage of model parameters than a global representation. To do so, we pre-acquire coarse scene geometry – that can be easily achieved via a fast RGB" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " volume reconstruction (like DVGO [32]) or multi-view stereo (like Point-NeRF [39]) – to directly distribute local tensors around the actual scene surface, leading to a sparse scene representation that avoids unnecessarily modeling the empty scene space. Note that while previous methods have also leveraged sparse representations (with voxels [19, 41] or points [39]) of radiance fields, their local features are modeled and optimized independently. Our model instead correlates a group of local features inside a" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": "local box and compactly express them with triple vectors, uniquely exploiting the local spatial coherence along axes and imposing local low-rank priors in the feature encoding via tensor factorization. Moreover, unlike previous sparse representations that only use a single-scale feature grid or point cloud, we distribute multi-scale local tensors to effectively model the scene geometry and appearance at multiple scales in a hierarchical manner. In particular, for an arbitrary 3D location, we aggregate the neural features from its neighboring tri-vector components at all scales and decode the volume density and view-dependent color from the aggregated features for radiance field rendering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 216, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 371 + ], + "type": "text", + "content": "Our approach takes the best of previous local and global radiance field representations. Compared with global representations like TensoRF and Instant-NGP, our model takes advantage of the sparsity of a scene more directly; compared with local representations like Plenoxels and Point-NeRF, our model makes use of the local smoothness and coherence of scene geometry and appearance. As shown in our experimental results on both synthetic and real datasets, our model is able to achieve state-of-the-art rendering quality on these datasets, outperforming previous methods, including TensoRF and Instant-NGP, while using significantly fewer model parameters, demonstrating the superior representational power of our model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 380, + 392, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 392, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 392, + 392 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 400, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 556 + ], + "type": "text", + "content": "Scene representations. To represent a 3D scene, traditional and learning-based methods have studied various representations, such as depth map [13, 18], mesh [15, 36, 30], point cloud [28, 1, 35] and implicit function [7, 21, 24, 40]. In recent years, continuous neural field representations stand out in various 3D tasks such as single-view 3D reconstruction [38, 11], surface completion [8, 26], multi-view reconstruction [24] and novel view synthesis [22, 20]. Compared with traditional discrete representations, a continuous field have no limitation on spatial resolution, e.g., volume resolution or the number of points. It can also naturally be represented by neural networks, such as an MLP, which are known for approximating complex functions well." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "text", + "content": "Neural field representations. Specifically, NeRF [22] represents a 3D scene as a radiance field with a global coordinate MLP, which models geometry, lighting and texture information jointly, leading to photo-realistic rendering quality in novel view synthesis. Apart from its advantage, purely MLP-based NeRF models [3, 34] in general suffer from inefficiency [2] when modeling highly complex or large-scale scenes, due to limited model capacity, slow optimization speed, and the cost of modeling empty space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "To model radiance fields more efficiently, recent works have explored combining neural fields with various traditional 3D representations, including voxels [19, 41, 32, 43]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17570" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 335 + ], + "type": "text", + "content": "and points [39]. Low-rank representations such as triplane [5, 10] and tensor decomposition [6, 25] have also been studied. In particular, DVGO [32] and Plenoxels [41] respectively use dense and sparse voxels with neural features for radiance field modeling. While being efficient to optimize, these localized feature grid-based representations lead to a large model size and can face overfitting issues when the features are of very high resolution. Consequently, DVGO can also work with a low-resolution grid and Plenoxels requires additional spatial regularization terms. On the other hand, recent works have adopted global feature encoding to express a high-resolution feature grid, including Instant-NGP [23] that hashes spatial features into multi-scale hash tables and TensorRF [6] that factorizes a feature grid into vector and matrix factors. These global feature encoding methods exploit the spatial correlation across the entire scene space, leading to fast and compact reconstruction and surpassing previous MLP-based or grid-based representations on rendering quality. However, similar to NeRF, such global representation can also be limited by its model capacity when representing highly complex or large-scale content." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 340, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 461 + ], + "type": "text", + "content": "Our approach instead combines local and global representations. Our tri-vector fields are sparsely distributed in the scene, similar to local representations (like plenoxels and Point-NeRF); meanwhile, features in each field are represented by tri-vector components shared across the local region as done in TensoRF, exploiting spatial feature commonalities. Our model leverages both spatial sparsity and coherence, leading to much higher compactness and better reconstruction quality than previous local and global representations (see Tab. 1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 465, + 288, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 465, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 288, + 574 + ], + "type": "text", + "content": "Relevant to our work, previous methods, such as KiloNeRF [29] and BlockNeRF [33] have also utilized multiple local MLPs to represent a scene. Specifically, KiloNeRF focuses and speeding up NeRF and their rendering quality is sacrificed; BlockNeRF essentially uses multiple NeRFs to increase the total model capacity. Instead of pure MLPs, our work is built upon tensor factorization-based feature encoding as done in TensoRF [6], and we in fact achieve superior rendering quality while decreasing the model capacity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 599, + 263, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 263, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 263, + 613 + ], + "type": "text", + "content": "3. Sparse Tri-Vector Field Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 624, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 673 + ], + "type": "text", + "content": "We now present our novel radiance field representation. In essence, our model consists of a cloud of small local trivector tensors at multiple scales, designed to leverage both sparsity and multi-scale spatial coherence (see Fig. 2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\tau_n|n = 1,\\dots,N\\}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " denote a cloud of tri-vector tensors. Each local tensor " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " is located at " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": ", covering a local cuboid space " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " with an edge length of " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": ". This cloud of tri" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 541, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 541, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 541, + 85 + ], + "type": "text", + "content": "vector tensors represents a radiance field for the 3D space:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 397, + 93, + 545, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 93, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 397, + 93, + 545, + 127 + ], + "type": "interline_equation", + "content": "\\Omega = \\bigcup_ {n = 1} ^ {N} \\omega_ {n}. \\tag {1}", + "image_path": "82f0e268c26681600e266637f4a84eb23d5794f49e5a446b6d41a72fe6a08af9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": "Here, each tensor " + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": " encodes a local multi-channel feature grid that includes a (single-channel) density grid " + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "inline_equation", + "content": "A_{\\sigma}" + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": " and a (multi-channel) appearance grid " + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "inline_equation", + "content": "A_{c}" + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": ", similar to the tensor grid in TensoRF [6]. In contrast to using a single global tensor in TensoRF [6], we model the volume density and view-dependent colors with multiple local tensors. In particular, for an arbitrary location " + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\chi \\in \\Omega" + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": ", we select " + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": " nearest tensors that cover " + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": ". Across the selected tensors, we aggregate the extracted density and appearance features recovered by their tri-vector factors for radiance field property regression, where the volume density " + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": " is directly obtained after the aggregation and the view-dependent color " + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": " is regressed by a small MLP " + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": " along with the viewing direction " + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 134, + 545, + 302 + ], + "type": "text", + "content": ". The continuous radiance field can be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 332, + 312, + 545, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 312, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 332, + 312, + 545, + 326 + ], + "type": "interline_equation", + "content": "\\sigma_ {\\chi}, c _ {\\chi} = A _ {\\sigma} \\left(\\left\\{\\mathcal {G} ^ {\\sigma} (\\chi) \\right\\}\\right), \\psi \\left(A _ {c} \\left(\\left\\{\\mathcal {G} ^ {c} (\\chi) \\right\\}\\right), \\mathbf {d}\\right). \\tag {2}", + "image_path": "83a1c82de614cefd4b50e046cf4059cb9eb622d16a334598d156a95e01c78aa0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 334, + 440, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 440, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 440, + 345 + ], + "type": "text", + "content": "3.1. Local tri-vector tensors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 352, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 388 + ], + "type": "text", + "content": "We apply the classic Canonical polyadic (CP) decomposition [4] to model our local tensors with tri-vector components." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "text", + "content": "CP decomposition. CP decomposition factorizes a " + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "text", + "content": " dimension tensor " + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\tau \\in \\mathbb{R}^{I_1\\times I_2\\times \\ldots \\times I_M}" + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "text", + "content": " into a linear combination of " + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 403, + 545, + 438 + ], + "type": "text", + "content": " rank-1 tensors:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 446, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 446, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 354, + 446, + 545, + 479 + ], + "type": "interline_equation", + "content": "\\tau = \\sum_ {r = 1} ^ {R} \\lambda_ {r} \\mathbf {v} _ {r} ^ {0} \\otimes \\mathbf {v} _ {r} ^ {1} \\otimes \\dots , \\otimes \\mathbf {v} _ {r} ^ {M}, \\tag {3}", + "image_path": "da8ee5fafffc38e93f1e80a03de00dcd83c1fed33a66535f41020383595502c6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 487, + 545, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 487, + 545, + 512 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 304, + 487, + 545, + 512 + ], + "type": "text", + "content": " denotes outer product; the weighting factor " + }, + { + "bbox": [ + 304, + 487, + 545, + 512 + ], + "type": "inline_equation", + "content": "\\lambda_r" + }, + { + "bbox": [ + 304, + 487, + 545, + 512 + ], + "type": "text", + "content": " can be absorbed into vectors " + }, + { + "bbox": [ + 304, + 487, + 545, + 512 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_r^0,\\dots ,\\mathbf{v}_r^M\\}" + }, + { + "bbox": [ + 304, + 487, + 545, + 512 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "text", + "content": "Density and appearance tensors. In our case of modeling a 3D radiance field, we set the geometry grid " + }, + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\sigma} \\in \\mathbb{R}^{I \\times J \\times K}" + }, + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "text", + "content": " as a 3D tensor. And the multi-channel appearance grid " + }, + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^c \\in \\mathbb{R}^{I \\times J \\times K \\times P}" + }, + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "text", + "content": " corresponds to a 4D tensor. The fourth appearance mode is of lower dimension (compared with the spatial modes), representing the final dimension of the features sent to the MLP decoder network." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 610, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 610, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 610, + 545, + 635 + ], + "type": "text", + "content": "According to Eqn.3, we factorize each tensor's feature grids, " + }, + { + "bbox": [ + 304, + 610, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\sigma}" + }, + { + "bbox": [ + 304, + 610, + 545, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 610, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{c}" + }, + { + "bbox": [ + 304, + 610, + 545, + 635 + ], + "type": "text", + "content": ", by CP decomposition:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 643, + 545, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 643, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 643, + 545, + 676 + ], + "type": "interline_equation", + "content": "\\mathcal {G} ^ {\\sigma} = \\sum_ {r = 1} ^ {R _ {\\sigma}} \\mathcal {A} _ {\\sigma , r} = \\sum_ {r = 1} ^ {R _ {\\sigma}} \\mathbf {v} _ {\\sigma , r} ^ {X} \\otimes \\mathbf {v} _ {\\sigma , r} ^ {Y} \\otimes \\mathbf {v} _ {\\sigma , r} ^ {Z}, \\tag {4}", + "image_path": "4a35d380060f7339a010bc4e111298cd3350d04c6aebd928b9558746a22ce195.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 678, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {G} ^ {c} = \\sum_ {r = 1} ^ {R _ {c}} \\mathcal {A} _ {c, r} \\otimes \\mathbf {b} _ {r} = \\sum_ {r = 1} ^ {R _ {c}} \\mathbf {v} _ {c, r} ^ {X} \\otimes \\mathbf {v} _ {c, r} ^ {Y} \\otimes \\mathbf {v} _ {c, r} ^ {Z} \\otimes \\mathbf {b} _ {r}, \\tag {5}", + "image_path": "5da8292f45dd0ebb783d309309620ff73727fe72abca4bbb63a486081c99f7b4.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17571" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 547, + 242 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 547, + 242 + ], + "type": "image", + "image_path": "14f89b20e162f2eae18ac0949108c925d2dc223ca3550182ad5e2a226d8d28f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 324 + ], + "type": "text", + "content": "Figure 2: Overview of our Sparse Tri-Vector Radiance Fields. We distribute our local tensors based on a coarse geometry estimated by a fast " + }, + { + "bbox": [ + 46, + 251, + 547, + 324 + ], + "type": "inline_equation", + "content": "\\mathrm{RGB}\\sigma" + }, + { + "bbox": [ + 46, + 251, + 547, + 324 + ], + "type": "text", + "content": " volume reconstruction as done in DVGO [32]. Here, we show our model running under " + }, + { + "bbox": [ + 46, + 251, + 547, + 324 + ], + "type": "inline_equation", + "content": "S = 2" + }, + { + "bbox": [ + 46, + 251, + 547, + 324 + ], + "type": "text", + "content": " scales. Each local tensor is factorized as axis-aligned triple based on CP decomposition. For any shading point " + }, + { + "bbox": [ + 46, + 251, + 547, + 324 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 46, + 251, + 547, + 324 + ], + "type": "text", + "content": ", we extract and evaluate features in each local tensor, according to the factorization (Sec. 3.1). Then, we aggregate these features among nearby tensors (Sec. 3.2) and across different scales (Sec. 3.3). Finally, the density and color are decoded (Sec. 3.4) and used by volume rendering (Sec.4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "inline_equation", + "content": "R_{\\sigma}" + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "inline_equation", + "content": "R_{c}" + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": " denote numbers of component; " + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\sigma,r}" + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{c,r}" + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": " are the component tensors that are factorized spatially; " + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\sigma,r}^{X}, \\dots, \\mathbf{v}_{c,r}^{X}, \\dots" + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": " are the 1D vectors with resolution " + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "inline_equation", + "content": "I, J, K" + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": ", modeling scene geometry and appearance along " + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "inline_equation", + "content": "X, Y, Z" + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": " axis; " + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "inline_equation", + "content": "R_{\\sigma}" + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "inline_equation", + "content": "R_{c}" + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": " are the component numbers; " + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_r" + }, + { + "bbox": [ + 47, + 339, + 287, + 411 + ], + "type": "text", + "content": " expresses the feature dimension." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": "As done in TensoRF [6], we stack all feature-mode vectors " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_r" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": " as columns together, which ends up a " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "P\\times R_{c}" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": " appearance matrix " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": ". This matrix models the appearance feature variations of the tensor and functions like a appearance dictionary. Note that naively following CP decomposition like TensoRF will assign a different appearance matrix for every local tensor. Instead, we propose to utilize a global appearance matrix " + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^c" + }, + { + "bbox": [ + 46, + 411, + 287, + 554 + ], + "type": "text", + "content": " shared across the entire cloud of local tensors, leading to a global appearance dictionary that explains the color correlations across scene. This further improves both the computational efficiency and model compactness of our model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 555, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 287, + 581 + ], + "type": "text", + "content": "Therefore, each of our local tensors is represented by their unique local tri-vector factors " + }, + { + "bbox": [ + 47, + 555, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_r^X" + }, + { + "bbox": [ + 47, + 555, + 287, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 555, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_r^Y" + }, + { + "bbox": [ + 47, + 555, + 287, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 555, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_r^Z" + }, + { + "bbox": [ + 47, + 555, + 287, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 596, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 287, + 645 + ], + "type": "text", + "content": "Feature evaluation. To achieve a continuous field, we consider trilinear interpolation when evaluating the tensor grid features. For a location " + }, + { + "bbox": [ + 46, + 596, + 287, + 645 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 46, + 596, + 287, + 645 + ], + "type": "text", + "content": ", we first compute its relative position " + }, + { + "bbox": [ + 46, + 596, + 287, + 645 + ], + "type": "inline_equation", + "content": "\\tilde{\\chi}" + }, + { + "bbox": [ + 46, + 596, + 287, + 645 + ], + "type": "text", + "content": " to the selected tensor located at " + }, + { + "bbox": [ + 46, + 596, + 287, + 645 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 596, + 287, + 645 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 654, + 287, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 654, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 99, + 654, + 287, + 668 + ], + "type": "interline_equation", + "content": "\\tilde {x}, \\tilde {y}, \\tilde {z} = x - p _ {x}, y - p _ {y}, z - p _ {z}. \\tag {6}", + "image_path": "0d7a48bc2486abe1520ec739b2dc8bd6c4f32e92e0e7549b9be697aee78b0736.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "Then, for example, to get " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\sigma,r}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\tilde{x},\\tilde{y},\\tilde{z})" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ", we can compute and trilinearly interpolate eight " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\sigma,r}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " on the corners. As mentioned in [6], applying linear interpolation on each" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 339, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 339, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 339, + 545, + 388 + ], + "type": "text", + "content": "vector first is mathematically equivalent and can reduce the computation cost. Under the rule of outer product, we have " + }, + { + "bbox": [ + 305, + 339, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{r,i,j,k} = \\mathbf{v}_{r,i}^{X}\\mathbf{v}_{r,j}^{Y}\\mathbf{v}_{r,k}^{Z}" + }, + { + "bbox": [ + 305, + 339, + 545, + 388 + ], + "type": "text", + "content": ", then the interpolated density features at location " + }, + { + "bbox": [ + 305, + 339, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 305, + 339, + 545, + 388 + ], + "type": "text", + "content": " are:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 395, + 545, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 395, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 312, + 395, + 545, + 420 + ], + "type": "interline_equation", + "content": "\\mathcal {G} ^ {\\sigma} (\\chi) = \\sum_ {r} \\mathbf {v} _ {\\sigma , r} ^ {X} (\\tilde {x}) \\mathbf {v} _ {\\sigma , r} ^ {Y} (\\tilde {y}) \\mathbf {v} _ {\\sigma , r} ^ {Z} (\\tilde {z}) = \\sum_ {r} \\mathcal {A} _ {\\sigma , r} (\\tilde {\\chi}), \\tag {7}", + "image_path": "961bba7a87bc4f8a978fed4b63e0566130f4e02ef51e6ab482648b0d53db2d41.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\sigma,r}^{X}(\\tilde{x})" + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\sigma,r}^{X}" + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "text", + "content": "'s linearly interpolated value at " + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "inline_equation", + "content": "(\\tilde{x})" + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "text", + "content": " along its " + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "text", + "content": " axis. Here, " + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{\\sigma}(\\chi)" + }, + { + "bbox": [ + 305, + 427, + 545, + 452 + ], + "type": "text", + "content": " is a scalar." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 452, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 452, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 452, + 545, + 475 + ], + "type": "text", + "content": "Similarly, the interpolated appearance features can be computed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 483, + 545, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 483, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 347, + 483, + 545, + 550 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {G} ^ {c} (\\chi) = \\sum_ {r} \\mathbf {v} _ {c, r} ^ {X} (\\tilde {x}) \\mathbf {v} _ {c, r} ^ {Y} (\\tilde {y}) \\mathbf {v} _ {c, r} ^ {Z} (\\tilde {z}) \\mathbf {b} _ {r} (8) \\\\ = \\sum_ {r} \\mathcal {A} _ {c, r} (\\tilde {\\chi}) \\mathbf {b} _ {r} (9) \\\\ = \\mathbf {B} \\cdot \\left(\\oplus \\left[ \\mathcal {A} _ {c, r} \\right] _ {r}\\right), (10) \\\\ \\end{array}", + "image_path": "9736a47e980719323346d600d342f5627b72a1a643ee34f1cec4038fde982b7b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 558, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 545, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 558, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 305, + 558, + 545, + 582 + ], + "type": "text", + "content": " denotes concatenation, " + }, + { + "bbox": [ + 305, + 558, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 305, + 558, + 545, + 582 + ], + "type": "text", + "content": " denotes dot product. The appearance feature " + }, + { + "bbox": [ + 305, + 558, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^c (\\chi)\\in \\mathbb{R}^P" + }, + { + "bbox": [ + 305, + 558, + 545, + 582 + ], + "type": "text", + "content": " is a vector." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 589, + 425, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 589, + 425, + 602 + ], + "spans": [ + { + "bbox": [ + 306, + 589, + 425, + 602 + ], + "type": "text", + "content": "3.2. Feature aggregation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 607, + 545, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 607, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 304, + 607, + 545, + 680 + ], + "type": "text", + "content": "We propose to aggregate the features from " + }, + { + "bbox": [ + 304, + 607, + 545, + 680 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 607, + 545, + 680 + ], + "type": "text", + "content": " neighboring tensors to jointly model the volume density and appearance for each 3D location " + }, + { + "bbox": [ + 304, + 607, + 545, + 680 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 304, + 607, + 545, + 680 + ], + "type": "text", + "content": ". In particular, inspired by Point-NeRF, we leverage an inverse distance-based weighting function to directly aggregate the multi-tensor features. Specifically, this weight can be expressed by" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 387, + 685, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 685, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 387, + 685, + 545, + 712 + ], + "type": "interline_equation", + "content": "w _ {m} = \\frac {1}{\\| p _ {m} - \\chi \\|}. \\tag {11}", + "image_path": "8571c82eeba5e6ec9cf61e780b6d839a5929b008ba344601a39799dd20322196.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17572" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "With this weight function, we directly obtain the density feature via the weighted sum:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 98, + 106, + 287, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 98, + 106, + 287, + 139 + ], + "type": "interline_equation", + "content": "f ^ {\\sigma} (\\chi) = \\frac {1}{\\sum w _ {m}} \\sum_ {m = 1} ^ {M} w _ {m} \\mathcal {G} _ {m} ^ {\\sigma} (\\chi). \\tag {12}", + "image_path": "0928c9afaef6326fc5f5e678f18673314db4f565d498055a52667255953dc328.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 287, + 184 + ], + "type": "text", + "content": "Similarly, the appearance feature aggregation can be expressed in a similar way, while using the shared appearance matrix (as described in Sec. 3.1) across local tensors:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 194, + 287, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 194, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 70, + 194, + 287, + 297 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f ^ {c} (\\chi) = \\frac {1}{\\sum w _ {m}} \\sum_ {m = 1} ^ {M} w _ {m} \\mathcal {G} _ {m} ^ {c} (\\chi) (13) \\\\ = \\frac {1}{\\sum w _ {m}} \\sum_ {m = 1} ^ {M} w _ {m} \\mathbf {B} ^ {c} \\cdot \\left(\\oplus \\left[ \\mathcal {A} _ {c, r} \\right] _ {r}\\right) (14) \\\\ = \\frac {1}{\\sum w _ {m}} \\mathbf {B} ^ {c} \\cdot \\left(\\sum_ {m = 1} ^ {M} w _ {m} \\left(\\oplus \\left[ \\mathcal {A} _ {c, r} \\right] _ {r}\\right)\\right). (15) \\\\ \\end{array}", + "image_path": "ade5305d866119255f0ed1ed827371009197837fe354466cb5692ccc23718b06.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 306, + 287, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 287, + 343 + ], + "type": "text", + "content": "Note that owing to sharing the appearance matrix across tensors, we reduce the computational complexity from " + }, + { + "bbox": [ + 47, + 306, + 287, + 343 + ], + "type": "inline_equation", + "content": "O(M \\cdot P \\cdot R_c)" + }, + { + "bbox": [ + 47, + 306, + 287, + 343 + ], + "type": "text", + "content": " in Eqn.14, to " + }, + { + "bbox": [ + 47, + 306, + 287, + 343 + ], + "type": "inline_equation", + "content": "O((M + P) \\cdot R_c)" + }, + { + "bbox": [ + 47, + 306, + 287, + 343 + ], + "type": "text", + "content": " in Eqn.15." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 350, + 199, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 199, + 362 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 199, + 362 + ], + "type": "text", + "content": "3.3. Multi-scale tri-vector fields." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 369, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 287, + 453 + ], + "type": "text", + "content": "Complex 3D scenes often contain multi-frequency geometry and appearance details. This motivates us to build multi-scale tensor clouds to discover the local geometry and appearance commonalities at multiple scales. Our final radiance field is modeled by multiple tri-vector tensor clouds at " + }, + { + "bbox": [ + 47, + 369, + 287, + 453 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 369, + 287, + 453 + ], + "type": "text", + "content": " different scales. Different clouds consist of tensors with different resolutions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "type": "text", + "content": "To regress the density and appearance at a location " + }, + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "type": "text", + "content": ", we gather the density and appearance features from a set of tensor clouds that cover " + }, + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_s|1\\leq s\\leq S,\\chi \\in \\Omega_s\\}" + }, + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "type": "text", + "content": ". Please note that tensor clouds of certain scales might not cover the location, so that " + }, + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\| \\{\\mathcal{T}_s\\} \\| \\leq S" + }, + { + "bbox": [ + 47, + 454, + 287, + 525 + ], + "type": "text", + "content": ". We simply compute the mean features across these scales:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 533, + 287, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 533, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 107, + 533, + 287, + 561 + ], + "type": "interline_equation", + "content": "f ^ {\\sigma} (\\chi) = \\frac {1}{\\| \\left\\{\\mathcal {T} _ {s} \\right\\} \\|} \\sum_ {s} f _ {s} ^ {\\sigma} (\\chi), \\tag {16}", + "image_path": "0bef98f24c93058d3f70813062c2b136d6fb28a9f7df62a24565be5070ed223b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 563, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 563, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 110, + 563, + 287, + 590 + ], + "type": "interline_equation", + "content": "f ^ {c} (\\chi) = \\frac {1}{\\| \\left\\{\\mathcal {T} _ {s} \\right\\} \\|} \\sum_ {s} f _ {s} ^ {c} (\\chi). \\tag {17}", + "image_path": "f77cd1b2805f059b36eebdbdbbed2e8997c7e3fc1889de880b84e2da83a4a17d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 601, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 287, + 638 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 47, + 601, + 287, + 638 + ], + "type": "inline_equation", + "content": "f^{\\sigma}(\\chi)" + }, + { + "bbox": [ + 47, + 601, + 287, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 601, + 287, + 638 + ], + "type": "inline_equation", + "content": "f^{c}(\\chi)" + }, + { + "bbox": [ + 47, + 601, + 287, + 638 + ], + "type": "text", + "content": " are the final density and appearance features we aggregate across multiple scales and multiple neighboring tensors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 647, + 116, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 116, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 116, + 659 + ], + "type": "text", + "content": "3.4.Decoding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "We apply softplus activation on the density feature " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "f^{\\sigma}(\\chi)" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " to obtain the final volume density and regress the view-dependent color by sending the appearance feature " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "f^{c}(\\chi)" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " and the viewing direction " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " to the MLP decoder " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 72, + 478, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 478, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 478, + 85 + ], + "type": "text", + "content": "4. Rendering and Reconstruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 92, + 545, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 92, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 92, + 545, + 152 + ], + "type": "text", + "content": "Volume Rendering We evaluates each pixel's color with physically-based volume rendering via differentiable ray marching. Following NeRF [22], we sample " + }, + { + "bbox": [ + 305, + 92, + 545, + 152 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 305, + 92, + 545, + 152 + ], + "type": "text", + "content": " shading points at " + }, + { + "bbox": [ + 305, + 92, + 545, + 152 + ], + "type": "inline_equation", + "content": "\\{\\chi_q\\mid q = 1,\\dots,Q\\}" + }, + { + "bbox": [ + 305, + 92, + 545, + 152 + ], + "type": "text", + "content": " along the ray, and accumulate radiance by density:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 164, + 545, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 164, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 361, + 164, + 545, + 204 + ], + "type": "interline_equation", + "content": "c = \\sum_ {q = 1} ^ {Q} T _ {q} \\left(1 - \\exp \\left(- \\sigma_ {q} \\delta_ {q}\\right)\\right) c _ {q}, \\tag {18}", + "image_path": "0402af36233bdb0ef0193e1742205332e56e8eb621d792ba84bf743a25eed75d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 357, + 203, + 451, + 234 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 203, + 451, + 234 + ], + "spans": [ + { + "bbox": [ + 357, + 203, + 451, + 234 + ], + "type": "interline_equation", + "content": "T _ {q} = \\exp (- \\sum_ {t = 1} ^ {q - 1} \\sigma_ {t} \\delta_ {t}).", + "image_path": "481cad42a65e31445ab88560c912235207e0688b32e70a32c2545aa6afbd8290.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 245, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 245, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 245, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\sigma_{q}" + }, + { + "bbox": [ + 306, + 245, + 545, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 245, + 545, + 270 + ], + "type": "inline_equation", + "content": "c_{q}" + }, + { + "bbox": [ + 306, + 245, + 545, + 270 + ], + "type": "text", + "content": " are the density and radiance at shading points; " + }, + { + "bbox": [ + 306, + 245, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 306, + 245, + 545, + 270 + ], + "type": "text", + "content": " is the marching distance per step; " + }, + { + "bbox": [ + 306, + 245, + 545, + 270 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 245, + 545, + 270 + ], + "type": "text", + "content": " denotes transmittance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 292, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 292, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 545, + 435 + ], + "type": "text", + "content": "Distributing local tensors. First of all, to better leverage the sparsity of a scene, we first obtain a geometric prior that roughly covers the scene geometry. The geometric prior can be in any commonly-used form, e.g., point cloud, occupancy grid, octree, or mesh vertices. Then we can uniformly distribute tensors in the spatial neighborhood of the geometry. For a multi-scale model, each of the scale will be distributed independently. For most of our results, we quickly optimize a coarse RGBA volume from the multiview images and use the optimized occupancy grid as the geometry prior, as done in DVGO [32], which finishes in seconds." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": "To maintain training stability and speed, each tensor " + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": "'s position " + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": " and coverage " + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": " is fixed once determined. We also initialize the " + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "3(R_{\\sigma} + R_c)" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": " vectors " + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "(\\mathbf{v}_{\\sigma,r}^{X}, \\dots, \\mathbf{v}_{c,r}^{X}, \\dots)" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": " of each tensor by normal distribution. For each scale " + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": ", a " + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "P \\times R_{c}" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": " appearance matrix " + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_s^c" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": " is shared by all tri-vector tensors of that scale. Specifically, “" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^c \\cdot ()" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": "” in Eqn.15 can be efficiently implemented as a fully-connected neural layer. Therefore, " + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^c" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": " for each scale and a global appearance MLP " + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 305, + 437, + 545, + 558 + ], + "type": "text", + "content": " will be implemented as neural networks and initialized by default methods [12]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 580, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 580, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 305, + 580, + 545, + 675 + ], + "type": "text", + "content": "Optimization and objectives. Given a set of multi-view RGB images with camera poses, the sparse tri-vector radiance field is per-scene optimized to reconstruct the radiance fields, under the supervision of the ground truth pixel colors. Following the volume rendering equation 18, the L2 rendering loss can be past back to the global MLP and aggregated features, then, all the way to the appearance matrices and the feature vectors of local tensors." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "We apply a rendering loss to supervise the reconstruction and also apply an " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " regularization loss on density feature vectors " + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\sigma,r}" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": " to promote geometry sparsity and to avoid" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17573" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 537, + 183 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 537, + 183 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 537, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 537, + 183 + ], + "type": "table", + "html": "
MethodBatchSizeStepsTime↓#Param.(M)↓PSNR↑SSIM↑LPIPSVgg↓LPIPSAlex↓
NeRF[20]4096300k35.0h1.2531.010.9470.081-
Plenoxels[41]5000128k11.4m194.5031.710.9580.049-
DVGO[32]500030k15.0m153.0031.950.9600.0530.035
Point-NeRF200k[39]4096200k5.5h27.7433.310.962*0.0490.027
InstantNGP[23]10k-85k30k3.9m11.6432.590.960--
TensoRF-CP[6]409630k25.2m0.9831.560.9490.0760.041
TensoRF-VM[6]409630k17.4m17.9533.140.9630.0470.027
Ours-24409630k34.3m7.0733.240.9630.0460.026
Ours-48409630k35.7m13.5233.550.9650.0440.025
", + "image_path": "bcc09d4c20522b19d5a8132730c099cacf248ff298e2006aaaa3d2a87fec56a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 186, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 547, + 237 + ], + "type": "text", + "content": "Table 1: Comparisons of our method with other radiance-based models [20, 37, 19, 39, 6, 23] on the Synthetic-NeRF dataset [20]. Ours-24 is the one with 24 components while Ours-48 is the one with 48 components. We report the corresponding rendering quality (PSNR, SSIM, and LPIPS), model capacity (number of parameters), and training time, batch size and steps. Our model achieves the best rendering quality with a compact model size. We report PointNeRF's updated SSIM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 258, + 191, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 191, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 191, + 270 + ], + "type": "text", + "content": "overfitting as done in TensoRF [6]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 278, + 287, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 278, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 139, + 278, + 287, + 293 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\| C - \\tilde {C} \\| _ {2} ^ {2}, \\tag {19}", + "image_path": "8a49a2322806833e2a7338913e00933058a0bd4ed7a57d98ce1434b685d2726a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 296, + 287, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 296, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 121, + 296, + 287, + 327 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L 1} = \\frac {1}{N} \\sum_ {r} ^ {R _ {\\sigma}} \\| \\mathbf {v} _ {\\sigma , r} \\|, \\tag {20}", + "image_path": "e1d72216903fc37c66c5df618adffc5530612689ed042c7e1da248063888a75e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 338, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 287, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 338, + 287, + 374 + ], + "type": "inline_equation", + "content": "\\tilde{C}" + }, + { + "bbox": [ + 47, + 338, + 287, + 374 + ], + "type": "text", + "content": " is the ground truth pixel color, " + }, + { + "bbox": [ + 47, + 338, + 287, + 374 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{v}_{\\sigma ,r}\\|" + }, + { + "bbox": [ + 47, + 338, + 287, + 374 + ], + "type": "text", + "content": " is the sum of absolute values of elements on density vectors, and " + }, + { + "bbox": [ + 47, + 338, + 287, + 374 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 338, + 287, + 374 + ], + "type": "text", + "content": " is the total number of elements. The total loss is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 385, + 287, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 385, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 130, + 385, + 287, + 397 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {r} + \\alpha \\mathcal {L} _ {L 1}. \\tag {21}", + "image_path": "af10cfdb8736eccbcd6fecce1428036e59844000cd459424b234fb0199a7635b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 407, + 284, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 284, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 284, + 419 + ], + "type": "text", + "content": "We set the weight of the sparsity term " + }, + { + "bbox": [ + 47, + 407, + 284, + 419 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 407, + 284, + 419 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 407, + 284, + 419 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 47, + 407, + 284, + 419 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 430, + 144, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 144, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 144, + 443 + ], + "type": "text", + "content": "5. Implementation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "text", + "content": "To obtain coarse scene geometry, we modify the coarse density estimation introduced in [32] and get a " + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "inline_equation", + "content": "100^{3}" + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "text", + "content": " occupancy volume in 30 seconds. We can skip this step if there exists available geometric data, e.g., the meshes in ScanNet [9], or point clouds from multiview stereo. According to the experiments, our method is not very sensitive to the initial geometry. Please refer to our supplementary material for more details. We set the default number of scales to 3. In a scene box of [-1,1], we rasterize the scene geometry (occupied voxels centers or points) onto 3 grids with different voxel sizes, e.g., " + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "inline_equation", + "content": "0.4^{3}, 0.2^{3}, 0.1^{3}" + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "text", + "content": ". For each grid, we distribute tri-vector tensors at the center of its occupied voxels. The tensor spatial coverage of these 3 scales is " + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "inline_equation", + "content": "0.6^{3}, 0.3^{3}, 0.15^{3}" + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "text", + "content": ", respectively. For each scale, we query " + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "text", + "content": " nearby tensors. Following [32], our feature decoding network " + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "text", + "content": " is a 2-layer MLP with 128 channels. For each scale, its appearance matrix " + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^c" + }, + { + "bbox": [ + 46, + 450, + 287, + 665 + ], + "type": "text", + "content": " is implemented by a single linear layer of 27 channels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "We implement the framework in PyTorch [14] with customized CUDA operations. During optimization, we adopt the coarse to fine strategy in [6], linearly up-sample the vectors' dimension " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "(I,J,K)" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " from 29 to 121 for scale one," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 342 + ], + "type": "text", + "content": "15 to 61 for scale two, and 7 to 31 for scale three. The up-sampling happens at step 2000, 3000, 4000, 5500, and 7000. We use the Adam optimizer [16] with initial learning rates of 0.02 for vectors and 0.001 for networks. On a single 3090 RTX GPU, we train each model for 30000 steps while each batch contains 4096 rays. Please find more details in the supplemental materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 353, + 388, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 353, + 388, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 388, + 366 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 372, + 529, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 529, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 529, + 385 + ], + "type": "text", + "content": "6.1. Evaluation on the NeRF Synthetic Dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 390, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 474 + ], + "type": "text", + "content": "We first evaluate our method on the Synthetic-NeRF dataset [20] and the quantitative results compared with other methods are reported in Tab.1, including NeRF [22], Plenoxels [41], DVGO [32], Point-NeRF [39], iNGP [23], and TensoRF [6]. We report our models of two different model sizes with different numbers of components; both settings are with the same 3 scales of local tensors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "content": "Our approach achieves the best averaged PSNR, " + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS}_{Vgg}" + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS}_{Alex}" + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "content": " in all the methods, leading to superior visual quality as shown in Fig. 3 Meanwhile, our high rendering quality is achieved with a compact model size. When compared with local voxel-based representations, such as Plenoxels and DVGO, our approach are significantly better." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": "On the other hand, global featuring encoding-based methods, like iNGP and TensorRF, are known for their high compactness and can achieve higher rendering quality than local voxel-based methods. Nonetheless, our method can still outperform them. Note that, even our smaller model (Ours-24) leads to better rendering quality than both iNGP and TensorRF that leverage global feature encoding, while our model uses significantly fewer parameters (about " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 689 + ], + "type": "text", + "content": " of the size of iNGP and TensorRF). This clearly demonstrates the high visual quality and high compatness of our model with our sparsely distributed tri-vector tensors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "In all the baseline methods, Point-NeRF is able to achieve relatively higher rendering quality than others." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17574" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 534, + 266 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 534, + 266 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 534, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 534, + 266 + ], + "type": "image", + "image_path": "557987c07a49e8537a71ac83d5031db513debc16f6efd48474ab5095d1bceb33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 275, + 440, + 288 + ], + "lines": [ + { + "bbox": [ + 153, + 275, + 440, + 288 + ], + "spans": [ + { + "bbox": [ + 153, + 275, + 440, + 288 + ], + "type": "text", + "content": "Figure 3: Qualitative comparisons on the NeRF Synthetic dataset [22]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 289, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 424 + ], + "type": "text", + "content": "However, this is enabled by optimizing their model for 300k steps with a long period of 5.5 hours. In contrast, our method achieves higher quality with significantly fewer optimization steps (only 30k) and optimization time (about 36 min). As expected, our model is slower to optimize than TensoRF due to the additional costs of multi-tensor aggregation. However, though speed is not our focus in this work, our model can still converge quickly and lead to significantly faster reconstruction than MLP-based methods, such as NeRF, as well as Point-NeRF that is point-based." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 289, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 289, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 289, + 622 + ], + "type": "text", + "content": "Performance w.r.t. rotation. We observe that tensor factorization-based methods can be sensitive to the orientation of the coordinate frames, since axis-aligned features are used; in essence, this is because the rank of a sparse tensor is sensitive to rotation, as shown in Fig. 4. Especially, this can benefit reconstruction on synthetic synthetic scenes where the objects are perfectly aligned with the axes, e.g. the lego scene in the NeRF synthetic data. However, we find that our method based on local tensors are more robust against the orientation of the axes than a global TensoRF. In particular, we compare our method with TensoRF in Tab.2 with different degrees of rotation around the Z axis on two scenes, lego (which is strongly aligned with axes) and chair (which is less aligned and thus less affected). As shown in the table, while both methods are affected by the rotation, our method has much smaller drops of PSNRs." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 637, + 290, + 684 + ], + "blocks": [ + { + "bbox": [ + 48, + 637, + 290, + 684 + ], + "lines": [ + { + "bbox": [ + 48, + 637, + 290, + 684 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 290, + 684 + ], + "type": "table", + "html": "
chair / legorot 0°rot 5°rot 45°
TensoRF-CP33.60 / 34.5032.90 / 29.7932.50 / 28.57
TensoRF-VM35.76 / 36.4634.91 /32.5334.55 / 32.31
Ours-4835.88 / 36.5235.72 / 35.3735.64 / 34.97
", + "image_path": "4c3970cc2762a279ac752922c262a38174e3a3363baaa68af38d85ac2e906c17.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 687, + 288, + 723 + ], + "lines": [ + { + "bbox": [ + 46, + 687, + 288, + 723 + ], + "spans": [ + { + "bbox": [ + 46, + 687, + 288, + 723 + ], + "type": "text", + "content": "Table 2: Comparisons of our method with TensoRF [6] on the chair and lego scenes of Synthetic-NeRF dataset [20] when considering rotation of different angles around " + }, + { + "bbox": [ + 46, + 687, + 288, + 723 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 687, + 288, + 723 + ], + "type": "text", + "content": "-axis." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 309, + 300, + 420, + 430 + ], + "blocks": [ + { + "bbox": [ + 309, + 300, + 420, + 430 + ], + "lines": [ + { + "bbox": [ + 309, + 300, + 420, + 430 + ], + "spans": [ + { + "bbox": [ + 309, + 300, + 420, + 430 + ], + "type": "image", + "image_path": "c26d2db69956d3a5b1b4a95ac446048372848aee999a53e74b76ce3c70884e1f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 445, + 545, + 552 + ], + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 552 + ], + "type": "text", + "content": "Figure 4: A toy example to illustrate the TensoRF-CP with global decomposition in (left) axis-aligned and (right) non-axis-aligned situations. The bottom shows the grid values. In axis-aligned case, only 1 component is needed to represent the scene (vector bases recover grid values by outer product). In non-axis-aligned case, however, 3 components are needed because the rank of matrix changes from 1 to 3 after scene rotation. While our design with local low-rank tensors can alleviate this issue." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 425, + 300, + 545, + 435 + ], + "blocks": [ + { + "bbox": [ + 425, + 300, + 545, + 435 + ], + "lines": [ + { + "bbox": [ + 425, + 300, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 425, + 300, + 545, + 435 + ], + "type": "image", + "image_path": "a6048dfe4579b808ce8a9e6f269bbe6f02625dc5ef5d516011b79cdcdd8bf9fc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 573, + 474, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 474, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 474, + 585 + ], + "type": "text", + "content": "6.2. Evaluation on the real datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "The ScanNet dataset. We evaluate our method on the real dataset, ScanNet [9] with the two scenes selected by NSVF [19], and compare with other methods. We follow the same experiment setting as done in NSVF [19], using the provided mesh to distribute our tensors, and optimize our model, TensoRF for the same 100k steps for fair comparisons. Please note that Point-NeRF uses all scanned depth images as initial geometry instead of meshes. Therefore, we also obtain the results of Point-NeRF 100k steps from the authors, using the provided mesh for fairness. We find the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17575" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 60, + 523, + 243 + ], + "blocks": [ + { + "bbox": [ + 72, + 60, + 523, + 243 + ], + "lines": [ + { + "bbox": [ + 72, + 60, + 523, + 243 + ], + "spans": [ + { + "bbox": [ + 72, + 60, + 523, + 243 + ], + "type": "image", + "image_path": "b464a67390f1ad6718da847a507ad8400deef103299bf826d2da4b8e922115ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 247, + 414, + 259 + ], + "lines": [ + { + "bbox": [ + 178, + 247, + 414, + 259 + ], + "spans": [ + { + "bbox": [ + 178, + 247, + 414, + 259 + ], + "type": "text", + "content": "Figure 5: Qualitative comparisons on the ScanNet dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 281, + 293, + 372 + ], + "blocks": [ + { + "bbox": [ + 127, + 271, + 276, + 281 + ], + "lines": [ + { + "bbox": [ + 127, + 271, + 276, + 281 + ], + "spans": [ + { + "bbox": [ + 127, + 271, + 276, + 281 + ], + "type": "text", + "content": "Average over Scene 101 and Scene 241" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 281, + 293, + 372 + ], + "lines": [ + { + "bbox": [ + 47, + 281, + 293, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 293, + 372 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPSAlex↓#Param.(M) ↓
SRN [31]18.250.5920.586-
NeRF [20]22.990.6200.369-
NSVF [19]25.480.6880.301-
Point-NeRF[39]25.920.8910.273159.01
TensoRF-CP[6]27.540.7510.3280.97
TensoRF-VM[6]28.610.7870.29049.92
Ours-4829.050.7920.24312.82
", + "image_path": "28e5432a11371fc7c5e9d6df2d8ced63eaa5e45b0a23b4a7a96a2b396033a7ce.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 374, + 288, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 374, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 288, + 411 + ], + "type": "text", + "content": "Table 3: Quantitative comparison on two scenes in the ScanNet dataset [9]. Point-NeRF, TensoRF-CP, TensoRFVM and Ours-48 are optimized for " + }, + { + "bbox": [ + 46, + 374, + 288, + 411 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 46, + 374, + 288, + 411 + ], + "type": "text", + "content": " steps." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 430, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 610 + ], + "type": "text", + "content": "Scannet data has many holes in the provided mesh geometry, while methods, such as NSVF and Point-NeRF, require accurate initial geometry; though Point-NeRF can potentially fix them with it point growing technique as shown in their original paper, it is not able to fully address them in 100k step optimization and lead to holes in their final rendering. Our approach instead does not require very accurate coarse geometry, since our local tensors cover relatively large regions. We show the quantitative results in Tab. 3 and qualitative results in Fig. 5. Note that our method can also perform well on real scenes, achieving the highest performance in terms of PSNR and " + }, + { + "bbox": [ + 46, + 430, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS}_{Alex}" + }, + { + "bbox": [ + 46, + 430, + 287, + 610 + ], + "type": "text", + "content": ", while using the second smallest model size (the smallest one is TensoRF" + }, + { + "bbox": [ + 46, + 430, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathrm{CP}_{100k}" + }, + { + "bbox": [ + 46, + 430, + 287, + 610 + ], + "type": "text", + "content": "). Our visual quality is also higher than the comparison methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "The Tanks and Temples dataset. We also evaluate our method on another real dataset, Tanks and Temples [9] with the 5 object scenes selected by NSVF [19]. We using the very coarse geometries estimated by DVGO[32] to distribute our tensors. We follow the same experiment setting as done in TensoRF [6], optimizing our model for the same " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "30\\mathrm{k}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " steps for fair comparisons. As is shown in" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 316, + 549, + 385 + ], + "blocks": [ + { + "bbox": [ + 305, + 271, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 305, + 271, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 547, + 307 + ], + "type": "text", + "content": "Tab. 3, our method outperforms other methods in terms of PSNR, SSIM and LPIPS " + }, + { + "bbox": [ + 305, + 271, + 547, + 307 + ], + "type": "inline_equation", + "content": "{}_{Alex}" + }, + { + "bbox": [ + 305, + 271, + 547, + 307 + ], + "type": "text", + "content": " ,while using the second smallest model size." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 316, + 549, + 385 + ], + "lines": [ + { + "bbox": [ + 306, + 316, + 549, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 549, + 385 + ], + "type": "table", + "html": "
ScalePSNR ↑SSIM ↑#Param.(M) ↓Time ↓
Single(0.6)32.220.9571.7518.22m
Single(0.3)32.730.9614.1521.31m
Single(0.15)31.960.95210.2028.55m
Multi(0.6, 0.3)33.110.9636.2030.12m
Multi(0.6, 0.3, 0.15)33.550.96513.5235.70m
", + "image_path": "027ceafb4023c990e286e1efe9bc2695d32fa5b39c07bcd0089e5d0711caf86d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 388, + 547, + 422 + ], + "lines": [ + { + "bbox": [ + 305, + 388, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 547, + 422 + ], + "type": "text", + "content": "Table 4: Ablation under different scale settings on Synthetic-NeRF dataset. We select 3 scales of tensors with cube sizes of 0.6, 0.3, and 0.15." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 444, + 398, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 398, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 398, + 456 + ], + "type": "text", + "content": "6.3. Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 617 + ], + "type": "text", + "content": "We analyze our model in terms of different scales in Table.4, while keeping the number of components the same (here we use 48). The scale here is the size of our local tensors of each axis. We considered 3 different scales, i.e., 0.6, 0.3, and 0.15 respectively as single-scale settings and some of their combinations as multi-scale settings. Note that even with a single scale (0.3), the performance of our method can be comparable with some other methods such as iNGP [23] while ours have less than half of the model size. When increasing the number of scales or decreasing the size of local tensors, our model size will also increase. In general, there is a trade-off of our method between scales and computational consumption (time and size)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "Usually, a smaller scale can lead to better performance, though our method with a scale of 0.15 does not strictly follow because we don't have high-quality input geometry to place these local tensors with a very small size. In fact, according to our per-scene breakdown results on the Synthetic-NeRF dataset (please refer to our supplemental material), single-scale(0.075) can achieve higher performance than single-scale(0.3) and single-scale(0.15) on most" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17576" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "scenes, except for ship because it has many thin structures that our coarse reconstruction does not cover." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 103, + 293, + 173 + ], + "blocks": [ + { + "bbox": [ + 48, + 103, + 293, + 173 + ], + "lines": [ + { + "bbox": [ + 48, + 103, + 293, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 293, + 173 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPSAlex ↓#Param.(M) ↓
NeRF [22]25.780.8640.198-
NSVF [19]28.400.9000.153-
TensoRF-CP30k[6]27.590.8970.1440.97
TensoRF-VM30k[6]28.560.9200.12549.92
Ours-4830k28.700.9220.11314.11
", + "image_path": "865a71a5b395d8bd45d63fcea042d6f5128947806641bc6b03a7eda4c6d4fa3e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 175, + 287, + 212 + ], + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 212 + ], + "type": "text", + "content": "Table 5: Quantitative comparison on scenes in the Tanks and Temples dataset [17] selected in NSVF [19]. TensoRF-CP, TensoRF-VM and Ours-48 are optimized for 30k steps." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 217, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 287, + 289 + ], + "type": "text", + "content": "We also compare our method with a variant that uses vector-matrix (VM) decomposition [6] in each local tensor instead of CP decomposition. Please refer to Appendix.A. for more details. Also, we can achieve a higher training and inference speed without a significant performance drop, which we refer to the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 298, + 119, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 119, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 119, + 310 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 319, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 475 + ], + "type": "text", + "content": "In this work, we have presented a novel approach for high-quality neural scene reconstruction and photorealistic novel view synthesis. We propose a novel tensor factorization-based scene representation, which leverages CP decomposition to compactly model a 3D scene as a sparse set of multi-scale tri-vector tensors that express local radiance fields. Our representation leverages both sparsity and spatial local coherence, and leads to accurate and efficient modeling of complex scene geometry and appearance. We demonstrate that the sparse tri-vector radiance fields can achieve superior rendering quality than previous state-of-the-art neural representations, including TensoRF and iNGP, while using significantly fewer parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 483, + 105, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 105, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 105, + 496 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 503, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 503, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 503, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 53, + 503, + 287, + 548 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In International Conference on Machine Learning (ICML), pages 40-49, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 548, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 288, + 580 + ], + "type": "text", + "content": "[2] Relja Arandjelović and Andrew Zisserman. Nerf in detail: Learning to sample for view synthesis. arXiv preprint arXiv:2106.05264, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 581, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 581, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 581, + 288, + 645 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 690 + ], + "type": "text", + "content": "[4] J Douglas Carroll and Jih-Jie Chang. Analysis of individual differences in multidimensional scaling via an n-way generalization of \"eckart-young\" decomposition. Psychometrika, 35(3):283-319, 1970. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 712 + ], + "type": "text", + "content": "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 72, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 126 + ], + "type": "text", + "content": "Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 545, + 172 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), pages 333-350. Springer, 2022. 1, 2, 3, 4, 6, 7, 8, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 545, + 215 + ], + "type": "text", + "content": "[7] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 545, + 271 + ], + "type": "text", + "content": "[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6970-6981, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 272, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 272, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 312, + 272, + 545, + 337 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5828-5839, 2017. 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 403 + ], + "type": "text", + "content": "[10] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12479–12488, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "type": "text", + "content": "[11] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4857-4866, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 1026-1034, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 568 + ], + "type": "text", + "content": "[13] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2821-2830, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 570, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 612 + ], + "type": "text", + "content": "[14] Sagar Imambi, Kolla Bhanu Prakash, and GR Kanagachidambaresan. Pytorch. Programming with TensorFlow: Solution for Edge Computing Applications, pages 87-104, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[15] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17577" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 93 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 93 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 93 + ], + "type": "text", + "content": "reconstruction. ACM Transactions on Graphics, 36(4), 2017. 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "text", + "content": "[18] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[19] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. arXiv preprint arXiv:2007.11571, 2020. 2, 6, 7, 8, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 249 + ], + "type": "text", + "content": "[20] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7210-7219, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 304 + ], + "type": "text", + "content": "[21] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "text", + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), pages 405-421. Springer, 2020. 1, 2, 5, 6, 7, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "type": "text", + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):1-15, 2022. 1, 3, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "text", + "content": "[25] Anton Obukhov, Mikhail Usvyatsov, Christos Sakaridis, Konrad Schindler, and Luc Van Gool. Tt-nf: Tensor train neural fields. arXiv preprint arXiv:2209.15529, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 493, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 557 + ], + "type": "text", + "content": "[26] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "text", + "content": "[27] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5865-5874, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 667 + ], + "type": "text", + "content": "[28] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[29] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. arXiv preprint arXiv:2103.13744, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[30] Jonathan Richard Shewchuk. Tetrahedral mesh generation by delaunay refinement. In Proceedings of the fourteenth annual symposium on Computational geometry, pages 86-95, 1998. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[31] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. arXiv preprint arXiv:1906.01618, 2019. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[32] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "type": "text", + "content": "[33] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 286, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 352 + ], + "type": "text", + "content": "[34] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "text", + "content": "[35] Jinglu Wang, Bo Sun, and Yan Lu. MVPnet: Multi-view point regression networks for 3D object reconstruction from a single image. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[36] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single RGB images. In European Conference on Computer Vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 444, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 510 + ], + "type": "text", + "content": "[37] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 512, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 566 + ], + "type": "text", + "content": "[38] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 567, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 623 + ], + "type": "text", + "content": "[39] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5438-5448, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[40] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Proc. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[41] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 2, 3, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17578" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[42] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[43] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5449-5458, 2022. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17579" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Structural Alignment for Network Pruning through Partial Regularization/bbe97c12-8d15-479d-81ec-fff8ed1209f4_content_list.json b/2023/Structural Alignment for Network Pruning through Partial Regularization/bbe97c12-8d15-479d-81ec-fff8ed1209f4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f47c1afd7221c2bda5505ac95a07310797b4a90c --- /dev/null +++ b/2023/Structural Alignment for Network Pruning through Partial Regularization/bbe97c12-8d15-479d-81ec-fff8ed1209f4_content_list.json @@ -0,0 +1,1496 @@ +[ + { + "type": "text", + "text": "Structural Alignment for Network Pruning through Partial Regularization", + "text_level": 1, + "bbox": [ + 107, + 128, + 861, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shangqian Gao $^{1}$ , Zeyu Zhang $^{2}$ , Yanfu Zhang $^{1}$ , Feihu Huang $^{1}$ , and Heng Huang $^{*3}$", + "bbox": [ + 163, + 191, + 803, + 212 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ Department of Electrical and Computer Engineering, University of Pittsburgh $^{2}$ School of Information, University of Arizona", + "$^{3}$ Department of Computer Science, University of Maryland at College Park" + ], + "bbox": [ + 174, + 222, + 792, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shg84, yaz91}@pitt.edu, zeyuzhang@email.arizona.edu, huangfeihu2018@gmail.com, heng@umd.edu", + "bbox": [ + 138, + 276, + 831, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 347, + 313, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel channel pruning method to reduce the computational and storage costs of Convolutional Neural Networks (CNNs). Many existing one-shot pruning methods directly remove redundant structures, which brings a huge gap between the model before and after network pruning. This gap will no doubt result in performance loss for network pruning. To mitigate this gap, we first learn a target sub-network during the model training process, and then we use this sub-network to guide the learning of model weights through partial regularization. The target sub-network is learned and produced by using an architecture generator, and it can be optimized efficiently. In addition, we also derive the proximal gradient for our proposed partial regularization to facilitate the structural alignment process. With these designs, the gap between the pruned model and the sub-network is reduced, thus improving the pruning performance. Empirical results also suggest that the sub-network found by our method has a much higher performance than the one-shot pruning setting. Extensive experiments show that our method can achieve state-of-the-art performances on CIFAR-10 and ImageNet with ResNets and MobileNet-V2.", + "bbox": [ + 75, + 380, + 473, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 739, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Convolutional Neural Networks (CNNs) have achieved many successes in different computer vision tasks [27, 46, 47, 50, 2, 6, 11]. To tackle real-world challenges, recent CNNs [27, 51, 14] become larger and larger regarding width, depth, etc. With such capacities, CNNs can obtain better performance on different benchmarks at the cost", + "bbox": [ + 75, + 763, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Corresponding author. This work was partially supported by NSF IIS 1838627, 1837956, 1956002, 2211492, CNS 2213701, CCF 2217003, DBI 2225775.", + "bbox": [ + 75, + 862, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of computational and storage burdens. At the same time, with the recent developments of mobile and embedded devices, the demand for deploying CNNs on these devices has increased dramatically. However, there is a natural conflict between the size of CNNs and the hardware capability of these devices. To overcome these challenges, many works [13, 12] try to reduce the size of CNNs, and make them possible to be deployed on edge devices.", + "bbox": [ + 496, + 349, + 890, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are many directions to reduce the size of CNNs. Among them, weight pruning and structural pruning are two popular topics. Structural pruning, especially channel pruning, is more friendly to hardware than weight pruning since no post-processing steps are required to acquire savings in computational and storage costs. Thus, our paper focuses on channel pruning for CNNs. Many existing one-shot pruning works [44, 56, 9, 42, 17, 8] prune trained models directly. No matter what method is used, there will be a significant gap between the selected sub-network and the pruned model. Such a gap creates difficulties in regaining performance during the fine-tuning process. On the other hand, soft pruning methods [16, 23] softly remove structures during the training process, which can produce good results with a shorter fine-tuning process. However, soft pruning methods generally perform worse than typical one-shot pruning methods, probably because the weight space is restricted during the training process because of soft pruning.", + "bbox": [ + 496, + 473, + 892, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle the above problems, we introduce a novel partial regularization technique to align model weights and the discovered sub-network during the training process, which can produce a high performance sub-network and reduce the gap between the sub-network and the original model. In addition, unlike soft-pruning methods, all model structures are used for training. The partial regularization term contains a partial group lasso regularization on selected weights, and other weights remain intact without modifica", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17402", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions. An architecture generator is trained to select which weights should be aligned, and it is also updated during the training process. Our partial regularization formulation is related to partial regularization in lasso [40]. Inspired by the nonmonotone proximal gradient (NPG) method used in [40], we also use a proximal gradient method to solve the partial regularization problem in our setting. Note that our method dynamically changes which channels should be put in the partial regularization. As a result, we add a scalar to balance the regularization strength for different layers because the number of pruned channels is different for them. To maximally keep the capacity of the original model, we insert the partial regularization in the middle of the training process. This is because weights are vulnerable to pruning at the early training stage, and the FLOPs regularization will dominate updates of the architecture generator, which can create bad sub-networks and mislead the training process. Finally, we update model weights and the architecture generator periodically, and they are connected by the partial regularization term during training. To maintain similar training efficiency as the original model, we only use a small portion of samples to train the architecture generator. Thus, there is only a small overhead compared to the original training process. Our method successfully finds performant sub-networks from the original model with these techniques. In summary, the contributions of this paper can be summarized as follows:", + "bbox": [ + 76, + 90, + 472, + 498 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We propose to align the sub-network in the original model with the final pruned model through partial regularization. By structural alignment, the gap between the selected sub-network and the pruned model is largely reduced, which naturally improves the performance of the pruned model.", + "2) We use an architecture generator parameterized by neural networks to select the proper sub-network structure and guide the partial regularization. Inspired by partial regularization in lasso [40], we propose to solve the partial regularization problem via proximal gradients, which facilitate the alignment process.", + "3) Empirical results show that the sub-network discovered by our method performs much better than the one-shot pruning setting. Extensive experiments on CIFAR-10 and ImageNet show that our method outperforms existing channel pruning methods on ResNets and MobileNet-V2." + ], + "bbox": [ + 86, + 506, + 468, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 800, + 227, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weight-Level Pruning. Weight pruning removes redundant connections based on individual weights. High compression rates can be achieved by weight pruning, but they can not directly achieve acceleration, and specially designed sparse matrix libraries are required. One of the early", + "bbox": [ + 76, + 825, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "works [13] proposes to measure the importance of weights with their $L_{1}$ or $L_{2}$ magnitude, and unimportant weights are removed. Instead of magnitude, SNIP [28] updates the importance of each weight by using gradients from the loss function. SNIP can be used at the initialization time. The assumption of the Lottery ticket hypothesis [7] suggests that there exist sparse sub-networks (winning tickets) that can achieve the performance of the full model. In addition, with repeated training and fine-tuning, it can achieve better results. On the other hand, rethinking network pruning [39] argues that the learned topology from pruning algorithms is the key to achieving better performance. In addition, weight rewinding [48] shows that resetting weights to values from the middle training process can also produce good results. Similar to weight rewinding, our method does not modify weight training at the beginning, the partial regularization is inserted in the middle training process.", + "bbox": [ + 496, + 90, + 893, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Structural-Level Pruning. Different from weight pruning, structural pruning is more friendly to hardware since it requires little or no post-processing steps to achieve acceleration. Similar to weight pruning, one of the early structural pruning methods [29] measure the importance of filters by using the sum of the absolute value of kernel weights. Besides using the magnitude to measure channel or filter importance, other methods utilize the scaling factor of batchnorm [21] to indicate which channels are important because batchnorm [21] is popular for the design of recent CNNs [14, 49]. To prune channels, Liu et al. [37] apply the sparse regularization to the scaling factors of batchnorm, and the channel will be pruned if the corresponding scaling factor is small. Structure sparse selection [20] introduces scaling factor to specific structures, such as neurons, groups, or residual blocks, and the sparsity regularization is applied to these structures. Structures with small values will also be removed. Another line of research formulates channel pruning as a constrained optimization problem [24, 56, 9, 23, 59], and learnable parameters are used to control each channel. These parameters are end-to-end differentiable, which is amenable to gradient based optimization methods. Our method is also related to these researches. Different from these methods, the learning of sub-networks accomplishes the training of model weights in our method. In addition, we use partial regularization to promote the selected sub-network, which reduces the gap between the pruned model and the sub-network. Besides using gates, Automatic Model Compression (AMC) [17] uses policy gradient to update the policy network. This policy network is then used to decide the left width of each layer. Collaborative channel pruning [44] prunes channels by exploiting inter-channel dependency. Greedy forward selection [55] starts from an empty network and greedily adds Important channels from the full model. MetaPruning [38, 32] uses a hypernet to generate parameters for sub", + "bbox": [ + 496, + 356, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17403", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6df72f0c9c5234eb531cda79203f02f34d04ab942b85e39f4903a5e347d60045.jpg", + "image_caption": [ + "Figure 1: Overview of the proposed method. The AGN generates the sub-network to guide the partial regularization during the training process of model weights. The AGN is trained by evaluating sub-networks on a sub-set from the whole dataset. Both model weights and AGN are trained in an end-to-end differentiable manner." + ], + "image_footnote": [], + "bbox": [ + 166, + 88, + 808, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "networks, and evolutionary algorithms are utilized to find the best sub-networks. MetaPruning shows that pruning should accomplish trained model weights.", + "bbox": [ + 75, + 441, + 468, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides these pruning methods, regularization based pruning methods can also be applied to structural pruning. Previous works [54, 31] use group sparsity to prune different structures. In addition to structural sparsity, structural sharing is considered in other works [58, 10]. Our method relates to regularization based methods; however, the formulation of our partial regularization only aligns selected channels dynamically, and other weights are intact.", + "bbox": [ + 75, + 487, + 468, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other Related Works. Besides the above-mentioned methods, there are works from other perspectives, including bayesian pruning [41, 43], weight quantization [4, 45], and knowledge distillation [19].", + "bbox": [ + 75, + 608, + 468, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 684, + 250, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 710, + 187, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before introducing our method, we first describe notations and provide an overview of our method. In a CNN, the feature map of $l$ th layer can be represented by $\\mathcal{F}_l \\in \\Re^{C_l \\times W_l \\times H_l}$ , $l = 1, \\ldots, L$ , where $C_l$ is the number of channels, $H_l$ and $W_l$ are height and width of the current feature map, $L$ is the number of layers. Similarly, the weights of $l$ th layer can be written as $\\mathcal{W}_l \\in \\Re^{C_l \\times C_{l-1} \\times k_l \\times k_l}$ , and $k_l$ is the kernel size of this layer. The mini-batch dimension of feature maps is ignored to simplify notations.", + "bbox": [ + 75, + 733, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core motivation of our proposed method is to reduce the gap between the selected sub-network from the origi", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nal model and the pruned model. To achieve this goal, we need two processes. First, we need to choose the desired sub-network from the original model, and this sub-network should also be updated when model weights are trained during the learning process. Secondly, we use the sub-network to guide the partial regularization term, which is used to decide which weights should be regularized. In addition, partial regularization should not be fixed to accommodate the changes in the selected sub-network.", + "bbox": [ + 496, + 441, + 890, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Learning the Sub-network", + "text_level": 1, + "bbox": [ + 500, + 585, + 740, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use an Architecture Generator Network (AGN) to generate the desired sub-network architecture $\\mathbf{v} \\in \\{0,1\\}^N$ , where 0 or 1 is used to depict the removal or keep of a channel, and $N$ is the total number of channels from all layers. The large parameterization space of AGN can facilitate the learning of sub-network structures. To generate $\\mathbf{v}$ , the following equation is used:", + "bbox": [ + 496, + 609, + 890, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} = \\operatorname {A G N} (\\Theta), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 726, + 890, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and AGN is composed of gated recurrent unit (GRU) [3] and dense layers. In addition, Gumbel-Sigmoid [22] with STE [1] are used to produce the final binary vector $\\mathbf{v}$ , and they are placed after the output of dense layers. More details of AGN are provided in the supplementary materials.", + "bbox": [ + 496, + 753, + 890, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once we have $\\mathbf{v}$ , we can apply it to the feature maps to produce a sub-network. The feature map of the $l$ th layer is then modified as follows:", + "bbox": [ + 496, + 829, + 890, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {F}} _ {l} = \\mathbf {v} _ {l} \\odot \\mathcal {F} _ {l}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 882, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17404", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ is element-wise multiplication, $\\mathbf{v}_l$ is the architecture vector of $l$ th layer, and $\\mathbf{v}_l$ is resized to have the same size of $\\mathcal{F}_l$ . The feature map $\\mathcal{F}_l$ is from the output of the activation function. The overall loss function for generating the desired sub-network is as follows:", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\Theta} \\mathcal {I} _ {\\theta} (\\Theta) := \\mathcal {L} \\left(f (x; \\mathcal {W}, \\mathbf {v}), y\\right) + \\lambda \\mathcal {R} _ {\\text {F L O P s}} (T (\\mathbf {v}), p T _ {\\text {t o t a l}}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 175, + 467, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T(\\mathbf{v})$ is the current FLOPs decided by the vector $\\mathbf{v}$ , $T_{\\mathrm{total}}$ is the total FLOPs of the original model, $p \\in (0,1]$ is a hyperparameter deciding the remaining fraction of FLOPs, $\\lambda$ is the hyper-parameter controlling the strength of FLOPs regularization, $f(x; \\mathcal{W}, \\mathbf{v})$ is a CNN parameterized by $\\mathcal{W}$ and the sub-network structure is determined by the architecture vector $\\mathbf{v}$ , $\\mathcal{L}$ is the task loss function and $\\mathcal{R}_{\\mathrm{FLOPs}}$ is the regularization term for FLOPs. The regularization term $\\mathcal{R}_{\\mathrm{FLOPs}}$ is $\\mathcal{R}_{\\mathrm{FLOPs}}(x,y) = \\log (\\max (x,y) / y)$ . With Eq. 3, we can find promising sub-networks when training the AGN and model weights periodically.", + "bbox": [ + 75, + 210, + 470, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Partial Regularization", + "text_level": 1, + "bbox": [ + 76, + 386, + 284, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a sub-network $\\mathbf{v}$ obtained from AGN, we can then reduce the gap between the sub-network and the pruned model and thus increase its performance. We can formulate the optimization problem with partial regularization as follows:", + "bbox": [ + 75, + 409, + 468, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {W}} \\mathcal {J} _ {w} (\\mathcal {W}) := \\mathcal {L} (f (x; \\mathcal {W}), y) + \\gamma \\mathcal {R} _ {w} (\\mathcal {W}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 496, + 468, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{R}_w$ is the partial regularization term, and $\\gamma$ controls the strength of the partial regularization. $\\mathcal{R}_w$ has the following form:", + "bbox": [ + 75, + 529, + 468, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {w} (\\mathcal {W}) = \\sum_ {l = 1} ^ {L} \\sum_ {i \\in S _ {l}} \\frac {\\hat {N} _ {l}}{\\hat {N}} \\| \\mathcal {W} _ {[ i,:,:,: ]} \\| _ {\\mathrm {G L}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 585, + 468, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{N}_l = \\sum 1 - \\mathbf{v}_l$ , $\\hat{N} = \\sum 1 - \\mathbf{v}$ and $S_{l} = \\{i\\mid \\mathrm{if}\\mathbf{v}_{l[i]} = 0\\}$ . $S_{l}$ contains the indices of pruned channels which are decided by AGN. $\\frac{\\hat{N}_l}{\\hat{N}}$ is a scalar to adjust the regularization strength given different layers. The numerator $\\hat{N}_l$ is the number of pruned channels of the $l$ th layer, and the denominator $\\hat{N}$ is the number of pruned channels from all layers. It is easy to see that $\\sum_{l=1}^{L} \\frac{\\hat{N}_l}{\\hat{N}} = 1$ . $\\|x\\|_{\\mathrm{GL}}$ is the norm of grouped weights for Group Lasso, and $\\|x\\|_{\\mathrm{GL}} = \\sqrt{\\sum_{i=1}^{|x|} x_i^2}$ where $|x|$ represents the number of elements in $x$ . In Eq. 5, we assume the corresponding layer is pruned across the output dimension. If it is pruned across the input dimension, we have $\\sum_{i \\in S_l} \\|W_{l[:i,::]}\\|_{\\mathrm{GL}}$ .", + "bbox": [ + 75, + 638, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of using Group Lasso for our partial regularization is to reduce the distance between the dense model and the pruned model, other suitable functions may also be useful, but we found that the partial regularization with Group", + "bbox": [ + 75, + 840, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Structural Alignment for Network Pruning" + ], + "code_body": "Input: $D,D_{\\mathrm{AGN}},p,\\lambda ,\\gamma ,E,E_{\\mathrm{start}}$ \nInitialization: initialize $\\mathcal{W}$ and $\\theta$ \nfor $e:=1$ to $E$ do \n/\\* Optimizing model weights. Freeze $\\Theta$ of the AGN. /\\* for a mini-batch $(x,y)$ in $D$ do 1. calculate the gradients w.r.t model weights: $\\nabla_{\\mathcal{W}}\\mathcal{L}$ 2. update model weights using any stochastic optimizer. 3.if if e $\\geq E_{\\mathrm{start}}$ then generate v from the AGN by using Eq. 1. apply the proximal gradient step following Eq.8. \nend \n/\\* Optimizing $\\Theta$ of the AGN. Freeze model weights W. /\\* if if e $\\geq E_{\\mathrm{start}}$ then for a mini-batch $(x,y)$ in $D_{\\mathrm{AGN}}$ do 1. generate v from the AGN by using Eq. 1 and apply it to the model. 2. calculate gradients w.r.t to $\\mathcal{J}$ in Eq. 3: $\\nabla_{\\Theta}\\mathcal{J}_{\\theta}$ 3.update the AGN with ADAM end \nend \nPruning the model with resulting v, and fine-tuning it.", + "bbox": [ + 519, + 127, + 867, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lasso already produces good results. Another benefit of the partial regularization $\\mathcal{R}_w$ is that it will not penalize weights that are not pruned. By doing so, we can avoid the problem of overpenalizing all weights' magnitude. In addition, $\\mathbf{v}$ is updated during the training process, and $\\mathcal{R}_w$ will dynamically regularize weights. As a result, $\\mathbf{v}$ can flexibly change instead of falling into a fixed sub-network during the optimization process.", + "bbox": [ + 496, + 527, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Proximal Gradients for Partial Regularization", + "text_level": 1, + "bbox": [ + 500, + 657, + 888, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. 5 has a similar formulation of the partial regularization of lasso [40]. In [40], the related optimization problem is solved via a nonmonotone proximal gradient (NPG) method. However, NPG requires frequent evaluation of the loss function to ensure the loss value after proximal gradients is less or equal to the loss value before the update. With CNNs, the costs of NPG are too large due to frequent loss evaluations. As a result, we use a one-step proximal gradient update to solve the problem defined in Eq. 4.", + "bbox": [ + 496, + 680, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proximal operator of $\\mathcal{R}_w$ is defined as:", + "bbox": [ + 517, + 816, + 805, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {p r o x} _ {\\gamma \\mathcal {R} _ {w}} (x) = \\underset {y} {\\arg \\min } \\gamma \\mathcal {R} _ {w} (y) + \\frac {1}{2} \\| x - y \\| ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 840, + 890, + 873 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq. 6, $\\| x - y\\| ^2$ defines the sum of the square difference", + "bbox": [ + 500, + 883, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17405", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "between $x$ and $y$ . We denote model weights after $t$ th update as $\\mathcal{W}^t$ , and $\\mathcal{W}^{t + 1}$ can be obtained by:", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} ^ {t + 1} = \\operatorname {p r o x} _ {\\alpha^ {t + 1} \\gamma \\mathcal {R} _ {w}} (u (\\mathcal {W} ^ {t}, \\alpha^ {t + 1})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 128, + 468, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $u(\\mathcal{W}^t, \\alpha^{t+1})$ is an update rule that can be applied for many algorithms, and $\\alpha^{t+1}$ is the learning rate at the current step. Take SGD as an example, $u(\\mathcal{W}^t, \\alpha^{t+1}) = \\mathcal{W}^t - \\alpha^{t+1}\\nabla_{\\mathcal{W}^t}\\mathcal{L}$ . With Eq. 6 and Eq. 7, model weights $\\mathcal{W}$ can then be updated with proximal gradients.", + "bbox": [ + 76, + 156, + 468, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now present the analytical solution to Eq. 6, so that model weights can be efficiently updated. From the structure of $\\mathcal{R}_w$ , we know that channels with indices $i \\in S_l$ will be regularized. This is equivalent to not regularizing channels if $i \\notin S_l$ . Consequently, we have the following form of the proximal gradient operator:", + "bbox": [ + 75, + 233, + 468, + 324 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {p r o x} _ {\\alpha \\gamma \\mathcal {R} _ {w}} (\\mathcal {W} _ {l}) = \\left\\{ \\begin{array}{c} \\frac {\\mathcal {W} _ {l [ i , ; ; ; ]}}{\\| (1 - \\mathbf {v} _ {l}) \\odot \\mathcal {W} _ {l} \\| _ {2}} \\max \\left(0, - \\frac {\\hat {N} _ {l}}{\\hat {N}} \\alpha \\gamma \\right. \\\\ \\left. + \\| (1 - \\mathbf {v} _ {l}) \\odot \\mathcal {W} _ {l} \\| _ {2}\\right), \\text {i f} i \\in S _ {l}, \\\\ \\mathcal {W} _ {l [ i, ; ; ; ; ]}, \\text {i f} i \\notin S _ {l}. \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 332, + 468, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq. 8, we omit the step notation $t$ to simplify the notations, and we still assume $W_{l}$ is pruned along the output dimension. In this equation, it is easier to see that $\\frac{\\hat{N}_l}{N}$ can balance the regularization strength given different layers. Due to the property of our proposed partial regularization, the term $\\| (1 - \\mathbf{v}_l)\\odot \\mathcal{W}_l\\| _2$ also changes dynamically, since $\\hat{N}_l$ is changed after updates of the AGN. As a result, if no adjustment is applied, no matter how large or small $\\hat{N}_l$ is, only a constant value $\\frac{1}{L}\\alpha \\gamma$ will be used for the soft-thresholding, $\\max (0,\\| (1 - \\mathbf{v}_l)\\odot \\mathcal{W}_l\\| _2 - \\frac{1}{L}\\alpha \\gamma)$ , in all circumstances, which is not reasonable. To accompany the changes of $v_{l}$ and consequently $\\hat{N}_l$ , we use $\\frac{\\hat{N}_l}{N}$ to dynamically balance the soft-thresholding parameter between different layers. With Eq. 8, we can efficiently update $\\mathcal{W}$ with our proposed partial regularization.", + "bbox": [ + 75, + 411, + 468, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Network Pruning via Structural Alignment", + "text_level": 1, + "bbox": [ + 76, + 650, + 444, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present the algorithm of our method in Algorithm 1. In Algorithm 1, $D$ is the training dataset; $D_{\\mathrm{AGN}}$ is a sub-dataset within $D$ and it is used to train AGN; $p$ decides how much FLOPs is preserved and it described in section 3.2; $\\gamma$ and $\\lambda$ are hyperparameters to control the strength of $\\mathcal{R}_{\\mathrm{FLOPs}}$ and $\\mathcal{R}_w$ ; $E$ is the total number of epochs; $E_{\\mathrm{start}}$ is the start epoch to train AGN and apply $\\mathcal{R}_w$ when optimizing model weights. Note that, to reduce the overhead brought by training AGN, we only use a small sub-set sampled from $D$ . As we discussed in section 1, we need to set a start epoch for AGN and $\\mathcal{R}_w$ . If we apply $\\mathcal{R}_w$ when training starts $(E_{\\mathrm{start}} = 0)$ , it will largely restrict the final performance of the whole model before pruning. Instead, we can have deserved results if we start partial regularization in the middle of the training process. The reason for this problem can", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "come from several perspectives. For example, at the beginning of the training, the classification loss can not produce accurate guidance for pruning since weights are not trained properly. Consequently, it will produce a bad sub-network, and following this sub-network only gives even worse results.", + "bbox": [ + 496, + 90, + 890, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We summarize our method in Fig. 1. The inference path when training model weights and the AGN are different, and they are connected by partial regularization, which is different from current one-shot pruning and soft pruning methods. With this design, model weights are not directly affected by the sub-network architecture, which creates a smooth transition for the selected sub-network before and after pruning.", + "bbox": [ + 496, + 181, + 892, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 308, + 630, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 500, + 332, + 596, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use CIFAR-10 [26] and ImageNet [5] to evaluate the performance of our method. Our method requires one hyper-parameter $p$ to control the FLOPs budget. The detailed choices of $p$ are listed in supplementary materials. We choose ResNets [14] and MobileNet-V2 [49] for comparison. For CIFAR-10, we compare our method with other methods on ResNet-56 and MobileNetV2. For ImageNet, we select ResNet-34, ResNet-50, ResNet-101, and MobileNetV2 as our target models.", + "bbox": [ + 496, + 356, + 890, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We set $\\lambda$ in Eq. 3 to 4.0 for all models and datasets. Similarly, we set $\\gamma$ to 0.0005 for all settings. We set $E_{\\mathrm{start}}$ at $20\\%$ of the total training epochs. Detailed numbers of $E_{\\mathrm{start}}$ are listed in supplementary materials. The range of $E_{\\mathrm{start}}$ is quite large, which is hard to be explored thoroughly. The current setting already provides good results, but better settings may also exist. To reduce the training costs of the AGN, we random sample $5\\%$ of samples from the original dataset for constructing $D_{\\mathrm{AGN}}$ . With this setup, the additional costs are less than $5\\%$ of the original training costs. We train the parameters $\\Theta$ of the AGN using ADAM [25] with a start learning rate 0.001. Besides the training of the AGN, we follow the standard training recipe of ResNets for both CIFAR-10 and ImageNet. For MobileNet-V2, We follow the training settings in their original paper [49]. After training, we prune the model by using the sub-network generated from the AGN. The fine-tuning settings are similar to the training setting. Due to space limitations, details of training and fine-tuning are also presented in supplementary materials. In the experimental section, our method is abbreviated as SANP: Structural Alignment for Network Pruning.", + "bbox": [ + 496, + 492, + 892, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. CIFAR-10 Results", + "text_level": 1, + "bbox": [ + 500, + 832, + 674, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results of CIFAR-10 are presented in Tab. 1. For ResNet-56, our method achieves the best performance (in terms of $\\Delta$ -Acc) compared to other methods. Specifically,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17406", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5f0c7ce49a5171b0d5bff126f0ecda9c9e964ac61b0d6f7bed186765404b9d3c.jpg", + "table_caption": [ + "Table 1: Comparison of results on CIFAR-10. $\\Delta$ -Acc represents the performance changes relative to the baseline, and $+ / -$ indicates an increase/decrease, respectively." + ], + "table_footnote": [], + "table_body": "
ArchitectureMethodBaseline AccPruned AccΔ-AccPruned FLOPs
ResNet-56DCP-Adapt [60]93.80%93.81%+0.01%47.0%
SCP [23]93.69%93.23%-0.46%51.5%
FPGM [18]93.59%92.93%-0.66%52.6%
SFP [16]93.59%92.26%-1.33%52.6%
FPC [15]93.59%93.24%-0.25%52.9%
HRank [35]93.26%92.17%-0.09%50.0%
DMC [9]93.62%92.69%+0.07%50.0%
GNN-RL [57]93.49%93.59%+0.10%54.0%
SANP (ours)93.49%93.81%+0.32%52.0%
MobileNetV2Uniform [60]94.47%94.17%-0.30%26.0%
DCP [60]94.47%94.69%+0.22%26.0%
DMC [9]94.23%94.49%+0.26%40.0%
SCOP [53]94.48%94.24%-0.24%40.3%
SANP (ours)94.52%94.97%+0.45%46.0%
", + "bbox": [ + 189, + 130, + 777, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a69ecfd6f65995fcb7b7a6c90c2cb5ebe89d2669a23950b0e09b6a684353793b.jpg", + "image_caption": [ + "(a) ResNet-56." + ], + "image_footnote": [], + "bbox": [ + 81, + 352, + 472, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44889f264491c4ad361fd83bc38ac42dff4cb89bd70997869a23378856d8663f.jpg", + "image_caption": [ + "(b) MoibleNet-V2." + ], + "image_footnote": [], + "bbox": [ + 493, + 352, + 883, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our method outperforms the second best method GNN-RL by $0.22\\%$ regarding $\\Delta$ -Acc (SANP $+0.32\\%$ vs. GNN-RL $+0.10\\%$ when pruning similar FLOPs (SANP $52.0\\%$ vs. GNN-RL $54.0\\%$ ). Our method also outperforms HRank and DMC by $0.25\\%$ and $0.41\\%$ separately (DMC $+0.07\\%$ and HRank $-0.09\\%$ ). The gap between other methods and our method is even larger. For MobileNet-V2, our method prunes most FLOPs $(46.0\\%)$ and achieves the best performance ( $\\Delta$ -Acc: $+0.45\\%$ ). Compared to the second best method, DMC, our method prunes $6\\%$ more FLOPs and outperforms it by $0.19\\%$ . SCOP also prunes $40\\%$ FLOPs, and the gap between our method and SCOP is even larger $(0.69\\%$ better in terms of $\\Delta$ -Acc). The uniform setting and DCP prunes around $26\\%$ FLOPs, but the performance is still worse than our method.", + "bbox": [ + 75, + 542, + 473, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. ImageNet Results", + "text_level": 1, + "bbox": [ + 76, + 782, + 250, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All results for the ImageNet dataset are shown in Tab. 2. ResNet-34. Our method achieves $73.43\\%$ Top-1 accuracy and $91.48\\%$ Top-5 accuracy, which is better than the baseline by $0.19\\%$ and $0.16\\%$ for Top-1/Top-5 accuracy separately. At the same time, our method removes $44.1\\%$ FLOPs, which is on par with other methods. It is obvious", + "bbox": [ + 75, + 806, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86694253368960dfd7c748f962eb19a7c490cb11809763b991738ab32460be87.jpg", + "image_caption": [ + "(a)", + "Figure 3: Top-1 and Top-5 accuracy after pruning given different settings." + ], + "image_footnote": [], + "bbox": [ + 501, + 539, + 691, + 652 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/824a7051fd2f86530a986497fab1772f7b059443360c694b93b27c11738777ee.jpg", + "image_caption": [ + "Figure 2: (a) and (b): the average norm of channels with or without partial regularization for ResNet-56 and MobileNet-V2.", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 699, + 537, + 888, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that the advantage of our method is clear compared to other methods. Our method prunes similar FLOPs to SCOP and DMC. However, the $\\Delta$ Top-1 Acc of our method is $0.88\\%$ and $0.92\\%$ better than SCOP and DMC, respectively, and we have similar observations for $\\Delta$ Top-5 Acc (0.60% and $0.47\\%$ better than SCOP and DMC). Taylor has the second best $\\Delta$ Top-1 Acc, but the pruned FLOPs is much lower than our method (ours: $44.1\\%$ vs. Taylor: $24.2\\%$ ). The advantage of our method compared to FPGM is more obvious.", + "bbox": [ + 496, + 732, + 890, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ResNet-50. Our method achieves $76.47\\%$ Top-1 accuracy and $93.00\\%$ Top-5 accuracy, which is also better than the", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17407", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/26b799c5748001e05f74af0d6f1d83ef5b28e4d12fa1df3f57d7494325e0251e.jpg", + "table_caption": [ + "Table 2: Comparison results on ImageNet with ResNet-34/50/101 and MobileNet-V2." + ], + "table_footnote": [], + "table_body": "
ArchitectureMethodBaseline Top-1 AccBaseline Top-5 AccΔ Top-1 AccΔ Top-5 AccPruned FLOPs
ResNet-34FPGM [18]73.92%91.62%-1.29%-0.54%41.1%
Taylor [42]73.31%--0.48%-24.2%
DMC [9]73.30%91.42%-0.73%-0.31%43.4%
SCOP [53]73.31%91.42%-0.69%-0.44%44.8%
SANP (ours)73.24%91.32%+0.19%+0.16%44.1%
ResNet-50DCP [60]76.01%92.93%-1.06%-0.61%55.6%
CCP [44]76.15%92.87%-0.94%-0.45%54.1%
FPGM [18]76.15%92.87%-1.32%-0.55%53.5%
ABCP [36]76.01%92.96%-2.15%-1.27%54.3%
DMC [9]76.15%92.87%-0.80%-0.38%55.0%
SCOP [53]76.15%92.87%-0.89%-0.34%54.6%
PFP [34]76.13%92.86%-0.92%-0.45%44.0%
CHIP [52]76.15%92.87%+0.00%+0.04%48.7%
NPPM [8]76.15%92.87%-0.19%+0.12%56.0%
Random-Pruning [30]75.83%92.92%-0.75%-0.40%51.0%
GNN-RL [57]76.10%--1.82%-53.0%
SANP (ours)76.06%92.86%+0.41%+0.17%56.2%
ResNet-101FPGM [18]77.37%93.56%-0.05%0.00%41.1%
Taylor [42]77.37%--0.02%-39.8%
DMC [9]77.37%93.56%+0.04%+0.03%56.0%
PFP [34]77.37%93.56%-0.94%-0.44%45.1%
SANP (ours)77.53%93.71%+0.61%+0.29%55.4%
MobileNet-V2Uniform [49]71.80%91.00%-2.00%-1.40%30.0%
AMC [17]71.80%--1.00%-30.0%
CC [33]71.88%--0.97%-28.3%
MetaPruning [38]72.00%--0.80%-30.7%
Random-Pruning [30]71.88%--1.01%-29.1%
SANP (ours)71.91%90.30%+0.14%+0.07%29.1%
", + "bbox": [ + 138, + 114, + 826, + 452 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/def7b7fd1b5f65fa5b6d64c6b364d2bbfc780322729256b4a66e21483374f6b6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingsArchitectureBaseline Top-1 AccΔ Top-1 AccPruned FLOPs
One-Shot SANPResNet-3473.31%-0.50%44.0%
73.24%+0.19%44.1%
One-Shot SANPResNet-5076.13%-0.57%56.0%
76.06%+0.41%55.2%
One-Shot SANPMobileNetV271.88%-0.42%29.3%
71.91%+0.07%29.1%
", + "bbox": [ + 80, + 467, + 465, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Performance of pruned models given different pruning settings on ImageNet.", + "bbox": [ + 76, + 554, + 468, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "baseline Top-1/Top-5 accuracy. The second best method, CHIP, removes $48.7\\%$ FLOPs while maintaining the original performance. Our method outperforms CHIP by $0.41\\%$ in terms of the $\\Delta$ Top-1 accuracy while pruning near $8\\%$ more FLOPs. NPPM is a strong baseline for ResNet-50, our method outperforms by $0.60\\%$ in $\\Delta$ -Top-1 Acc. GNN-RL and Random-Pruning are two recent pruning works. Our method outperforms them by $2.23\\%$ and $1.16\\%$ separately, while our method prunes more FLOPs. PHP, DCP, CCP, and DMC have similar $\\Delta$ -Top 1 accuracy. The gap between our method and these methods ranges from $1.21\\%$ to $1.47\\%$ regarding $\\Delta$ -Top 1 accuracy. The advantage of our method compared to the rest methods is more apparent.", + "bbox": [ + 75, + 608, + 468, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ResNet-101. Our method achieves $78.14\\%$ Top-1 accuracy and $94.00\\%$ Top-5 accuracy, which is $0.61\\%$ and $0.29\\%$ better than the baseline Top-1 and Top-5 accuracy. Although DMC prunes a little bit more FLOPs, it only achieves $77.41\\%$ Top-1 accuracy after pruning which is $0.73\\%$ lower than our method. FPGA, Taylor, and PFP remove less than", + "bbox": [ + 76, + 809, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "50% FLOPs. Our method prunes at least 10% more FLOPs and still has an advantage in terms of $\\Delta$ -Top 1 accuracy (from 0.66% to 1.55%).", + "bbox": [ + 498, + 469, + 890, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MobileNet-V2. MobileNet-V2 is generally harder to prune compared to ResNets. All comparison methods on MobileNet-V2 remove around $30\\%$ FLOPs. Our method achieves $72.05\\%$ Top-1 accuracy and $90.37\\%$ Top-5 accuracy, which is $0.14\\%$ and $0.07\\%$ better than the baseline Top-1 and Top-5 accuracy. Given the similar pruning rate, our method is $1.15\\%$ , $0.94\\%$ , $1.11\\%$ and $1.14\\%$ higher than Random-Pruning, MetaPruning CC, and AMC separately regarding $\\Delta$ -Top-1 Acc. MetaPruning prunes most FLOPs, but the performance is much lower than our method. In short, our method can also be applied to lightweight CNNs, like MobileNet-V2.", + "bbox": [ + 496, + 515, + 892, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis of Our Method", + "text_level": 1, + "bbox": [ + 500, + 705, + 720, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effectiveness of partial regularization. To verify whether our proposed partial regularization is effective, we plot the average channel norm of different groups within each block for ResNet-56 and MobileNet-V2 on CIFAR-10. The results are shown in Fig. 2. The average channel norm for the group with partial regularization and without partial regularization are obtained by $\\frac{1}{\\tilde{N}_l}\\sum_{i\\in S_l}\\| \\mathcal{W}_{l[i,:;:]} \\|_{\\mathrm{GL}}$ and $\\frac{1}{C_l - \\tilde{N}_l}\\sum_{i\\notin S_l}\\| \\mathcal{W}_{l[i,:;:]} \\|_{\\mathrm{GL}}$ separately. Weights with partial regularization are effectively aligned. On the contrary, weights without partial regularization are lightly affected, which justifies the strength of our proposed partial regular", + "bbox": [ + 496, + 727, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17408", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/281bc250104f5325c56a506ac5bfa79fdea6ceacdbf80cfd963b2b98953b7825.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 85, + 254, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e29c52c97e5da8fdfcdc649c285d170d024999e8b8f3b8dcdb660622505e323c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 87, + 464, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a4ccd9bd6001d512932246f5ed7b4192731416cd37dfbf2f1c1c7effb7c2ead.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 671, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03ece7e4f2f553b57c01f487e585c92c27a064a44bc23463885b5fa8a8edc4da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 87, + 883, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4addfc2a4679d8270777b550b2b9a49a82b5adb6e486c44e85482612dc02af7.jpg", + "image_caption": [ + "(a)", + "(e)", + "Figure 4: (a, e): the impact of $\\lambda$ in $\\mathcal{R}_{\\mathrm{FLOPs}}$ . (b, f): the effect of the architecture of AGN. (c, g): the effect of $E_{\\mathrm{start}}$ . (d, h): the effect of different setups. Experiments are conducted on CIFAR-10 with ResNet-56 and $p = 0.5$ (a,b,c,d,e) and $p = 0.35$ (f,g,h)." + ], + "image_footnote": [], + "bbox": [ + 76, + 217, + 251, + 324 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0ddbd8370737cf69b68333a1cde4cd37120b0cc67bad76bf7a3cb729dc146f8.jpg", + "image_caption": [ + "(b)", + "(f)" + ], + "image_footnote": [], + "bbox": [ + 285, + 215, + 464, + 324 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a8e34b812542773adfbf228da27dc0c06c62eea1146a08ef07c23af33365447.jpg", + "image_caption": [ + "(c)", + "(g)" + ], + "image_footnote": [], + "bbox": [ + 496, + 217, + 673, + 324 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67bd6b1df5e318a5e01a1c01a1ed1147e8a891764e311eb5c0c47dbe41c7418f.jpg", + "image_caption": [ + "(d)", + "(h)" + ], + "image_footnote": [], + "bbox": [ + 707, + 215, + 883, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ization.", + "text_level": 1, + "bbox": [ + 76, + 422, + 130, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The impact of $\\lambda$ . We study the impact of $\\lambda$ when training the AGN, and we plot the test accuracy and $\\mathcal{R}_{\\mathrm{FLOPs}}$ in Fig. 4a and Fig. 4e. From the figures, we can see that if $\\lambda$ is too large, it will have negative impacts on learned subnetworks. Otherwise, our method is robust to $\\lambda$ ,", + "bbox": [ + 75, + 439, + 468, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of AGN. In Fig. 4b and Fig. 4f, we plot the test accuracy when learning the AGN given different pruning rates. We construct a Simple baseline, which parameterizes each channel by using one learnable parameter. We can see that when the parameterization space shrinks, the performance of learned sub-networks will be affected severely. In addition, the learning is slower, and the best sub-network performance is also much worse than using AGN.", + "bbox": [ + 75, + 517, + 468, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of $\\mathbf{E}_{\\mathrm{start}}$ . In the method section, we argue that inserting partial regularization in the middle training process is beneficial. We plot the results of $E_{\\mathrm{start}} = 0$ and $E_{\\mathrm{start}} = 40$ in Fig. 4c and Fig. 4g. In general, they can find sub-networks with similar performance, but $E_{\\mathrm{start}} = 0$ is worse than $E_{\\mathrm{start}} = 40$ when the pruning rate is large ( $p = 0.35$ ). More importantly, the full model accuracy of $E_{\\mathrm{start}} = 0$ is $92.96\\%$ (average across different runs), which is around $0.50\\%$ worse than $E_{\\mathrm{start}} = 40$ , which suggests that using partial regularization at the beginning will limit the capacity of the full model.", + "bbox": [ + 75, + 641, + 470, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of different setups. We construct two additional baselines to see how partial regularization helps to produce a better sub-network within the full model. The One-Shot baseline directly trains the AGN on the pre-trained model. The w/o Partial Regularization baseline set $\\gamma = 0$ , and the rest settings are the same as our method. The related", + "bbox": [ + 75, + 809, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results are shown in Fig. 4d and Fig. 4h. From these figures, we can see that partial regularization always produces the best sub-network from the full model. In addition, 'w/o partial regularization' is worse than the one-shot setting. This is probably because it is hard to capture the changes of model weights without partial regularization, which makes the training of AGN much harder than the rest settings. More comparisons on the ImageNet dataset. To further show how our method improves the one-shot setting, we present the sub-network performance before and after fine-tuning for one-shot and partial regularization settings in Fig. 3 and Tab. 3. From Fig. 3, we can see that partial regularization still produces better sub-networks on ImageNet, and it is around $10\\%$ better in terms of Top-1 accuracy than the one-shot setting for different models. The advantage of partial regularization naturally extends to results after fine-tuning, as shown in Tab. 3.", + "bbox": [ + 496, + 422, + 893, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 704, + 617, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we investigate how partial regularization helps to produce a better sub-network for network pruning. Specifically, our method uses AGN to guide partial regularization across the training process. We further provide an efficient way to update model weights through proximal gradients. With these designs, partial regularization effectively reduces the gap between the sub-network within the full model and the pruned model. Our method then starts from a better sub-network, thus resulting in a better final pruned model. Extensive experimental results on CIFAR-10 and ImageNet show the effectiveness of our method.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17409", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3", + "[2] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016. 1", + "[3] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 3", + "[4] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in neural information processing systems, pages 3123-3131, 2015. 3", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 248-255. IEEE, 2009. 5", + "[6] Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, and H.T. Kung. Neural mean discrepancy for efficient out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19217-19227, June 2022. 1", + "[7] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2019. 2", + "[8] Shangqian Gao, Feihu Huang, Weidong Cai, and Heng Huang. Network pruning via performance maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9270-9280, 2021. 1, 7", + "[9] Shangqian Gao, Feihu Huang, Jian Pei, and Heng Huang. Discrete model compression with resource constraint for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1899-1908, 2020. 1, 2, 6, 7", + "[10] Shangqian Gao, Burak Uzkent, Yilin Shen, Heng Huang, and Hongxia Jin. Learning to jointly share and prune weights for grounding based vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[11] Junfeng Guo, Yiming Li, Xun Chen, Hanqing Guo, Lichao Sun, and Cong Liu. SCALE-UP: An efficient black-box input-level backdoor detection via analyzing scaled prediction consistency. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[12] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Advances in neural information processing systems, pages 1135-1143, 2015. 1, 2", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 5", + "[15] Yang He, Yuhang Ding, Ping Liu, Linchao Zhu, Hanwang Zhang, and Yi Yang. Learning filter pruning criteria for deep convolutional neural networks acceleration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2009-2018, 2020. 6", + "[16] Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating deep convolutional neural networks. In International Joint Conference on Artificial Intelligence (IJCAI), pages 2234-2240, 2018. 1, 6", + "[17] Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, and Song Han. Amc: Automl for model compression and acceleration on mobile devices. In Proceedings of the European Conference on Computer Vision (ECCV), pages 784-800, 2018. 1, 2, 7", + "[18] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2019. 6, 7", + "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[20] Zehao Huang and Naiyan Wang. Data-driven sparse structure selection for deep neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 304-320, 2018. 2", + "[21] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32Nd International Conference on International Conference on Machine Learning - Volume 37, ICML, pages 448-456. JMLR.org, 2015. 2", + "[22] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 3", + "[23] Minsoo Kang and Bohyung Han. Operation-aware soft channel pruning using differentiable masks. In International Conference on Machine Learning, pages 5122-5131. PMLR, 2020. 1, 2, 6", + "[24] Jaedeok Kim, Chiyoun Park, Hyun-Joo Jung, and Yoonsuck Choe. Plug-in, trainable gate for streamlining arbitrary neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 2", + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5", + "[26] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, Cite-seer, 2009. 5" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17410", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097-1105, 2012. 1", + "[28] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. *ICLR*, 2019. 2", + "[29] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. *ICLR*, 2017. 2", + "[30] Yawei Li, Kamil Adamczewski, Wen Li, Shuhang Gu, Radu Timofte, and Luc Van Gool. Revisiting random channel pruning for neural network compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 191-201, 2022. 7", + "[31] Yawei Li, Shuhang Gu, Christoph Mayer, Luc Van Gool, and Radu Timofte. Group sparsity: The hinge between filter pruning and decomposition for network compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8018-8027, 2020. 3", + "[32] Yawei Li, Shuhang Gu, Kai Zhang, Luc Van Gool, and Radu Timofte. Dhp: Differentiable meta pruning via hypernetworks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 608-624. Springer, 2020. 2", + "[33] Yuchao Li, Shaohui Lin, Jianzhuang Liu, Qixiang Ye, Mengdi Wang, Fei Chao, Fan Yang, Jincheng Ma, Qi Tian, and Rongrong Ji. Towards compact cnns via collaborative compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6438-6447, 2021. 7", + "[34] Lucas Liebenwein, Cenk Baykal, Harry Lang, Dan Feldman, and Daniela Rus. Provable filter pruning for efficient neural networks. In International Conference on Learning Representations, 2020. 7", + "[35] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. Hrank: Filter pruning using high-rank feature map. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6", + "[36] Mingbao Lin, Rongrong Ji, Yuxin Zhang, Baochang Zhang, Yongjian Wu, and Yonghong Tian. Channel pruning via automatic structure search. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages 673 - 679, 2020. 7", + "[37] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 2", + "[38] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In Proceedings of the IEEE International Conference on Computer Vision, pages 3296-3305, 2019. 2, 7", + "[39] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. In International Conference on Learning Representations, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Zhaosong Lu and Xiaorui Li. Sparse recovery via partial regularization: Models, theory, and algorithms. Mathematics of Operations Research, 43(4):1290-1316, 2018. 2, 4", + "[41] Dmitry Molchanov, Arsenii Ashukha, and Dmitry Vetrov. Variational dropout sparsifies deep neural networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 2498-2507. JMLR.org, 2017. 3", + "[42] Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11264–11272, 2019. 1, 7", + "[43] Kirill Neklyudov, Dmitry Molchanov, Arsenii Ashukha, and Dmitry P Vetrov. Structured bayesian pruning via log-normal multiplicative noise. In Advances in Neural Information Processing Systems, pages 6775-6784, 2017. 3", + "[44] Hanyu Peng, Jiaxiang Wu, Shifeng Chen, and Junzhou Huang. Collaborative channel pruning for deep networks. In International Conference on Machine Learning, pages 5113-5122, 2019. 1, 2, 7", + "[45] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In European Conference on Computer Vision, pages 525-542. Springer, 2016. 3", + "[46] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1", + "[47] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91-99, 2015. 1", + "[48] Alex Renda, Jonathan Frankle, and Michael Carbin. Comparing rewinding and fine-tuning in neural network pruning. In International Conference on Learning Representations, 2020. 2", + "[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 2, 5, 7", + "[50] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in neural information processing systems, pages 568-576, 2014. 1", + "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1", + "[52] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. Chip: Channel independence-based pruning for compact neural networks. Advances in Neural Information Processing Systems, 34, 2021. 7", + "[53] Yehui Tang, Yunhe Wang, Yixing Xu, Dacheng Tao, Chunjing Xu, Chao Xu, and Chang Xu. Scop: Scientific control for reliable neural network pruning. Advances in Neural Information Processing Systems, 33, 2020. 6, 7", + "[54] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17411", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Advances in neural information processing systems, pages 2074-2082, 2016.3", + "[55] Mao Ye, Chengyue Gong, Lizhen Nie, Denny Zhou, Adam Klivans, and Qiang Liu. Good subnetworks provably exist: Pruning via greedy forward selection. In International Conference on Machine Learning, pages 10820-10830. PMLR, 2020. 2", + "[56] Zhonghui You, Kun Yan, Jinmian Ye, Meng Ma, and Ping Wang. Gate decorator: Global filter pruning method for accelerating deep convolutional neural networks. In Advances in Neural Information Processing Systems, pages 2130-2141, 2019. 1, 2", + "[57] Sixing Yu, Arya Mazaheri, and Ali Jannesari. Topology-aware network pruning using multi-stage graph embedding and reinforcement learning. In International Conference on Machine Learning, pages 25656-25667. PMLR, 2022. 6, 7", + "[58] Dejiao Zhang, Haozhu Wang, Mario Figueiredo, and Laura Balzano. Learning to share: Simultaneous parameter tying and sparsification in deep learning. 2018. 3", + "[59] Yanfu Zhang, Shangqian Gao, and Heng Huang. Exploration and estimation for model compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 487-496, 2021. 2", + "[60] Zhuangwei Zhuang, Mingkui Tan, Bohan Zhuang, Jing Liu, Yong Guo, Qingyao Wu, Junzhou Huang, and Jinhui Zhu. Discrimination-aware channel pruning for deep neural networks. In Advances in Neural Information Processing Systems, pages 875-886, 2018. 6, 7" + ], + "bbox": [ + 78, + 90, + 467, + 486 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17412", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Structural Alignment for Network Pruning through Partial Regularization/bbe97c12-8d15-479d-81ec-fff8ed1209f4_model.json b/2023/Structural Alignment for Network Pruning through Partial Regularization/bbe97c12-8d15-479d-81ec-fff8ed1209f4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eb928b698651ee278c2668885fb95a5e45fa0a4a --- /dev/null +++ b/2023/Structural Alignment for Network Pruning through Partial Regularization/bbe97c12-8d15-479d-81ec-fff8ed1209f4_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.13, + 0.862, + 0.154 + ], + "angle": 0, + "content": "Structural Alignment for Network Pruning through Partial Regularization" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.193, + 0.804, + 0.213 + ], + "angle": 0, + "content": "Shangqian Gao\\(^{1}\\), Zeyu Zhang\\(^{2}\\), Yanfu Zhang\\(^{1}\\), Feihu Huang\\(^{1}\\), and Heng Huang\\(^{*3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.223, + 0.794, + 0.259 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Electrical and Computer Engineering, University of Pittsburgh \\(^{2}\\)School of Information, University of Arizona" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.259, + 0.781, + 0.277 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Computer Science, University of Maryland at College Park" + }, + { + "type": "list", + "bbox": [ + 0.175, + 0.223, + 0.794, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.277, + 0.832, + 0.304 + ], + "angle": 0, + "content": "{shg84, yaz91}@pitt.edu, zeyuzhang@email.arizona.edu, huangfeihu2018@gmail.com, heng@umd.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.348, + 0.314, + 0.365 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.474, + 0.712 + ], + "angle": 0, + "content": "In this paper, we propose a novel channel pruning method to reduce the computational and storage costs of Convolutional Neural Networks (CNNs). Many existing one-shot pruning methods directly remove redundant structures, which brings a huge gap between the model before and after network pruning. This gap will no doubt result in performance loss for network pruning. To mitigate this gap, we first learn a target sub-network during the model training process, and then we use this sub-network to guide the learning of model weights through partial regularization. The target sub-network is learned and produced by using an architecture generator, and it can be optimized efficiently. In addition, we also derive the proximal gradient for our proposed partial regularization to facilitate the structural alignment process. With these designs, the gap between the pruned model and the sub-network is reduced, thus improving the pruning performance. Empirical results also suggest that the sub-network found by our method has a much higher performance than the one-shot pruning setting. Extensive experiments show that our method can achieve state-of-the-art performances on CIFAR-10 and ImageNet with ResNets and MobileNet-V2." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Convolutional Neural Networks (CNNs) have achieved many successes in different computer vision tasks [27, 46, 47, 50, 2, 6, 11]. To tackle real-world challenges, recent CNNs [27, 51, 14] become larger and larger regarding width, depth, etc. With such capacities, CNNs can obtain better performance on different benchmarks at the cost" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author. This work was partially supported by NSF IIS 1838627, 1837956, 1956002, 2211492, CNS 2213701, CCF 2217003, DBI 2225775." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.472 + ], + "angle": 0, + "content": "of computational and storage burdens. At the same time, with the recent developments of mobile and embedded devices, the demand for deploying CNNs on these devices has increased dramatically. However, there is a natural conflict between the size of CNNs and the hardware capability of these devices. To overcome these challenges, many works [13, 12] try to reduce the size of CNNs, and make them possible to be deployed on edge devices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.893, + 0.761 + ], + "angle": 0, + "content": "There are many directions to reduce the size of CNNs. Among them, weight pruning and structural pruning are two popular topics. Structural pruning, especially channel pruning, is more friendly to hardware than weight pruning since no post-processing steps are required to acquire savings in computational and storage costs. Thus, our paper focuses on channel pruning for CNNs. Many existing one-shot pruning works [44, 56, 9, 42, 17, 8] prune trained models directly. No matter what method is used, there will be a significant gap between the selected sub-network and the pruned model. Such a gap creates difficulties in regaining performance during the fine-tuning process. On the other hand, soft pruning methods [16, 23] softly remove structures during the training process, which can produce good results with a shorter fine-tuning process. However, soft pruning methods generally perform worse than typical one-shot pruning methods, probably because the weight space is restricted during the training process because of soft pruning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To tackle the above problems, we introduce a novel partial regularization technique to align model weights and the discovered sub-network during the training process, which can produce a high performance sub-network and reduce the gap between the sub-network and the original model. In addition, unlike soft-pruning methods, all model structures are used for training. The partial regularization term contains a partial group lasso regularization on selected weights, and other weights remain intact without modifica" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "17402" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.499 + ], + "angle": 0, + "content": "tions. An architecture generator is trained to select which weights should be aligned, and it is also updated during the training process. Our partial regularization formulation is related to partial regularization in lasso [40]. Inspired by the nonmonotone proximal gradient (NPG) method used in [40], we also use a proximal gradient method to solve the partial regularization problem in our setting. Note that our method dynamically changes which channels should be put in the partial regularization. As a result, we add a scalar to balance the regularization strength for different layers because the number of pruned channels is different for them. To maximally keep the capacity of the original model, we insert the partial regularization in the middle of the training process. This is because weights are vulnerable to pruning at the early training stage, and the FLOPs regularization will dominate updates of the architecture generator, which can create bad sub-networks and mislead the training process. Finally, we update model weights and the architecture generator periodically, and they are connected by the partial regularization term during training. To maintain similar training efficiency as the original model, we only use a small portion of samples to train the architecture generator. Thus, there is only a small overhead compared to the original training process. Our method successfully finds performant sub-networks from the original model with these techniques. In summary, the contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.507, + 0.47, + 0.598 + ], + "angle": 0, + "content": "1) We propose to align the sub-network in the original model with the final pruned model through partial regularization. By structural alignment, the gap between the selected sub-network and the pruned model is largely reduced, which naturally improves the performance of the pruned model." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.603, + 0.47, + 0.693 + ], + "angle": 0, + "content": "2) We use an architecture generator parameterized by neural networks to select the proper sub-network structure and guide the partial regularization. Inspired by partial regularization in lasso [40], we propose to solve the partial regularization problem via proximal gradients, which facilitate the alignment process." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.697, + 0.47, + 0.788 + ], + "angle": 0, + "content": "3) Empirical results show that the sub-network discovered by our method performs much better than the one-shot pruning setting. Extensive experiments on CIFAR-10 and ImageNet show that our method outperforms existing channel pruning methods on ResNets and MobileNet-V2." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.507, + 0.47, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.228, + 0.817 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Weight-Level Pruning. Weight pruning removes redundant connections based on individual weights. High compression rates can be achieved by weight pruning, but they can not directly achieve acceleration, and specially designed sparse matrix libraries are required. One of the early" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.351 + ], + "angle": 0, + "content": "works [13] proposes to measure the importance of weights with their \\( L_{1} \\) or \\( L_{2} \\) magnitude, and unimportant weights are removed. Instead of magnitude, SNIP [28] updates the importance of each weight by using gradients from the loss function. SNIP can be used at the initialization time. The assumption of the Lottery ticket hypothesis [7] suggests that there exist sparse sub-networks (winning tickets) that can achieve the performance of the full model. In addition, with repeated training and fine-tuning, it can achieve better results. On the other hand, rethinking network pruning [39] argues that the learned topology from pruning algorithms is the key to achieving better performance. In addition, weight rewinding [48] shows that resetting weights to values from the middle training process can also produce good results. Similar to weight rewinding, our method does not modify weight training at the beginning, the partial regularization is inserted in the middle training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.895, + 0.904 + ], + "angle": 0, + "content": "Structural-Level Pruning. Different from weight pruning, structural pruning is more friendly to hardware since it requires little or no post-processing steps to achieve acceleration. Similar to weight pruning, one of the early structural pruning methods [29] measure the importance of filters by using the sum of the absolute value of kernel weights. Besides using the magnitude to measure channel or filter importance, other methods utilize the scaling factor of batchnorm [21] to indicate which channels are important because batchnorm [21] is popular for the design of recent CNNs [14, 49]. To prune channels, Liu et al. [37] apply the sparse regularization to the scaling factors of batchnorm, and the channel will be pruned if the corresponding scaling factor is small. Structure sparse selection [20] introduces scaling factor to specific structures, such as neurons, groups, or residual blocks, and the sparsity regularization is applied to these structures. Structures with small values will also be removed. Another line of research formulates channel pruning as a constrained optimization problem [24, 56, 9, 23, 59], and learnable parameters are used to control each channel. These parameters are end-to-end differentiable, which is amenable to gradient based optimization methods. Our method is also related to these researches. Different from these methods, the learning of sub-networks accomplishes the training of model weights in our method. In addition, we use partial regularization to promote the selected sub-network, which reduces the gap between the pruned model and the sub-network. Besides using gates, Automatic Model Compression (AMC) [17] uses policy gradient to update the policy network. This policy network is then used to decide the left width of each layer. Collaborative channel pruning [44] prunes channels by exploiting inter-channel dependency. Greedy forward selection [55] starts from an empty network and greedily adds Important channels from the full model. MetaPruning [38, 32] uses a hypernet to generate parameters for sub" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17403" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.089, + 0.81, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.895, + 0.416 + ], + "angle": 0, + "content": "Figure 1: Overview of the proposed method. The AGN generates the sub-network to guide the partial regularization during the training process of model weights. The AGN is trained by evaluating sub-networks on a sub-set from the whole dataset. Both model weights and AGN are trained in an end-to-end differentiable manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.488 + ], + "angle": 0, + "content": "networks, and evolutionary algorithms are utilized to find the best sub-networks. MetaPruning shows that pruning should accomplish trained model weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.608 + ], + "angle": 0, + "content": "Besides these pruning methods, regularization based pruning methods can also be applied to structural pruning. Previous works [54, 31] use group sparsity to prune different structures. In addition to structural sparsity, structural sharing is considered in other works [58, 10]. Our method relates to regularization based methods; however, the formulation of our partial regularization only aligns selected channels dynamically, and other weights are intact." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Other Related Works. Besides the above-mentioned methods, there are works from other perspectives, including bayesian pruning [41, 43], weight quantization [4, 45], and knowledge distillation [19]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.685, + 0.251, + 0.703 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.711, + 0.188, + 0.725 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Before introducing our method, we first describe notations and provide an overview of our method. In a CNN, the feature map of \\( l \\)th layer can be represented by \\( \\mathcal{F}_l \\in \\Re^{C_l \\times W_l \\times H_l} \\), \\( l = 1, \\ldots, L \\), where \\( C_l \\) is the number of channels, \\( H_l \\) and \\( W_l \\) are height and width of the current feature map, \\( L \\) is the number of layers. Similarly, the weights of \\( l \\)th layer can be written as \\( \\mathcal{W}_l \\in \\Re^{C_l \\times C_{l-1} \\times k_l \\times k_l} \\), and \\( k_l \\) is the kernel size of this layer. The mini-batch dimension of feature maps is ignored to simplify notations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The core motivation of our proposed method is to reduce the gap between the selected sub-network from the origi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.579 + ], + "angle": 0, + "content": "nal model and the pruned model. To achieve this goal, we need two processes. First, we need to choose the desired sub-network from the original model, and this sub-network should also be updated when model weights are trained during the learning process. Secondly, we use the sub-network to guide the partial regularization term, which is used to decide which weights should be regularized. In addition, partial regularization should not be fixed to accommodate the changes in the selected sub-network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.741, + 0.603 + ], + "angle": 0, + "content": "3.2. Learning the Sub-network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.716 + ], + "angle": 0, + "content": "We use an Architecture Generator Network (AGN) to generate the desired sub-network architecture \\(\\mathbf{v} \\in \\{0,1\\}^N\\), where 0 or 1 is used to depict the removal or keep of a channel, and \\(N\\) is the total number of channels from all layers. The large parameterization space of AGN can facilitate the learning of sub-network structures. To generate \\(\\mathbf{v}\\), the following equation is used:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.727, + 0.892, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} = \\operatorname {A G N} (\\Theta), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.754, + 0.892, + 0.829 + ], + "angle": 0, + "content": "and AGN is composed of gated recurrent unit (GRU) [3] and dense layers. In addition, Gumbel-Sigmoid [22] with STE [1] are used to produce the final binary vector \\(\\mathbf{v}\\), and they are placed after the output of dense layers. More details of AGN are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.83, + 0.892, + 0.873 + ], + "angle": 0, + "content": "Once we have \\(\\mathbf{v}\\), we can apply it to the feature maps to produce a sub-network. The feature map of the \\(l\\)th layer is then modified as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.883, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {F}} _ {l} = \\mathbf {v} _ {l} \\odot \\mathcal {F} _ {l}, \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17404" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "where \\(\\odot\\) is element-wise multiplication, \\(\\mathbf{v}_l\\) is the architecture vector of \\(l\\)th layer, and \\(\\mathbf{v}_l\\) is resized to have the same size of \\(\\mathcal{F}_l\\). The feature map \\(\\mathcal{F}_l\\) is from the output of the activation function. The overall loss function for generating the desired sub-network is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.176, + 0.468, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\Theta} \\mathcal {I} _ {\\theta} (\\Theta) := \\mathcal {L} \\left(f (x; \\mathcal {W}, \\mathbf {v}), y\\right) + \\lambda \\mathcal {R} _ {\\text {F L O P s}} (T (\\mathbf {v}), p T _ {\\text {t o t a l}}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.211, + 0.471, + 0.378 + ], + "angle": 0, + "content": "where \\( T(\\mathbf{v}) \\) is the current FLOPs decided by the vector \\( \\mathbf{v} \\), \\( T_{\\mathrm{total}} \\) is the total FLOPs of the original model, \\( p \\in (0,1] \\) is a hyperparameter deciding the remaining fraction of FLOPs, \\( \\lambda \\) is the hyper-parameter controlling the strength of FLOPs regularization, \\( f(x; \\mathcal{W}, \\mathbf{v}) \\) is a CNN parameterized by \\( \\mathcal{W} \\) and the sub-network structure is determined by the architecture vector \\( \\mathbf{v} \\), \\( \\mathcal{L} \\) is the task loss function and \\( \\mathcal{R}_{\\mathrm{FLOPs}} \\) is the regularization term for FLOPs. The regularization term \\( \\mathcal{R}_{\\mathrm{FLOPs}} \\) is \\( \\mathcal{R}_{\\mathrm{FLOPs}}(x,y) = \\log (\\max (x,y) / y) \\). With Eq. 3, we can find promising sub-networks when training the AGN and model weights periodically." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.387, + 0.285, + 0.403 + ], + "angle": 0, + "content": "3.3. Partial Regularization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.484 + ], + "angle": 0, + "content": "Given a sub-network \\(\\mathbf{v}\\) obtained from AGN, we can then reduce the gap between the sub-network and the pruned model and thus increase its performance. We can formulate the optimization problem with partial regularization as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.497, + 0.469, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {W}} \\mathcal {J} _ {w} (\\mathcal {W}) := \\mathcal {L} (f (x; \\mathcal {W}), y) + \\gamma \\mathcal {R} _ {w} (\\mathcal {W}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.575 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}_w\\) is the partial regularization term, and \\(\\gamma\\) controls the strength of the partial regularization. \\(\\mathcal{R}_w\\) has the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.586, + 0.47, + 0.628 + ], + "angle": 0, + "content": "\\[\nR _ {w} (\\mathcal {W}) = \\sum_ {l = 1} ^ {L} \\sum_ {i \\in S _ {l}} \\frac {\\hat {N} _ {l}}{\\hat {N}} \\| \\mathcal {W} _ {[ i,:,:,: ]} \\| _ {\\mathrm {G L}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.841 + ], + "angle": 0, + "content": "where \\(\\hat{N}_l = \\sum 1 - \\mathbf{v}_l\\), \\(\\hat{N} = \\sum 1 - \\mathbf{v}\\) and \\(S_{l} = \\{i\\mid \\mathrm{if}\\mathbf{v}_{l[i]} = 0\\}\\). \\(S_{l}\\) contains the indices of pruned channels which are decided by AGN. \\(\\frac{\\hat{N}_l}{\\hat{N}}\\) is a scalar to adjust the regularization strength given different layers. The numerator \\(\\hat{N}_l\\) is the number of pruned channels of the \\(l\\)th layer, and the denominator \\(\\hat{N}\\) is the number of pruned channels from all layers. It is easy to see that \\(\\sum_{l=1}^{L} \\frac{\\hat{N}_l}{\\hat{N}} = 1\\). \\(\\|x\\|_{\\mathrm{GL}}\\) is the norm of grouped weights for Group Lasso, and \\(\\|x\\|_{\\mathrm{GL}} = \\sqrt{\\sum_{i=1}^{|x|} x_i^2}\\) where \\(|x|\\) represents the number of elements in \\(x\\). In Eq. 5, we assume the corresponding layer is pruned across the output dimension. If it is pruned across the input dimension, we have \\(\\sum_{i \\in S_l} \\|W_{l[:i,::]}\\|_{\\mathrm{GL}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The goal of using Group Lasso for our partial regularization is to reduce the distance between the dense model and the pruned model, other suitable functions may also be useful, but we found that the partial regularization with Group" + }, + { + "type": "code_caption", + "bbox": [ + 0.509, + 0.095, + 0.848, + 0.126 + ], + "angle": 0, + "content": "Algorithm 1: Structural Alignment for Network Pruning" + }, + { + "type": "algorithm", + "bbox": [ + 0.52, + 0.128, + 0.868, + 0.497 + ], + "angle": 0, + "content": "Input: \\(D,D_{\\mathrm{AGN}},p,\\lambda ,\\gamma ,E,E_{\\mathrm{start}}\\) \nInitialization: initialize \\(\\mathcal{W}\\) and \\(\\theta\\) \nfor \\(e:=1\\) to \\(E\\) do \n/\\* Optimizing model weights. Freeze \\(\\Theta\\) of the AGN. /\\* for a mini-batch \\((x,y)\\) in \\(D\\) do 1. calculate the gradients w.r.t model weights: \\(\\nabla_{\\mathcal{W}}\\mathcal{L}\\) 2. update model weights using any stochastic optimizer. 3.if if e \\(\\geq E_{\\mathrm{start}}\\) then generate v from the AGN by using Eq. 1. apply the proximal gradient step following Eq.8. \nend \n/\\* Optimizing \\(\\Theta\\) of the AGN. Freeze model weights W. /\\* if if e \\(\\geq E_{\\mathrm{start}}\\) then for a mini-batch \\((x,y)\\) in \\(D_{\\mathrm{AGN}}\\) do 1. generate v from the AGN by using Eq. 1 and apply it to the model. 2. calculate gradients w.r.t to \\(\\mathcal{J}\\) in Eq. 3: \\(\\nabla_{\\Theta}\\mathcal{J}_{\\theta}\\) 3.update the AGN with ADAM end \nend \nPruning the model with resulting v, and fine-tuning it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.649 + ], + "angle": 0, + "content": "Lasso already produces good results. Another benefit of the partial regularization \\(\\mathcal{R}_w\\) is that it will not penalize weights that are not pruned. By doing so, we can avoid the problem of overpenalizing all weights' magnitude. In addition, \\(\\mathbf{v}\\) is updated during the training process, and \\(\\mathcal{R}_w\\) will dynamically regularize weights. As a result, \\(\\mathbf{v}\\) can flexibly change instead of falling into a fixed sub-network during the optimization process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.658, + 0.89, + 0.674 + ], + "angle": 0, + "content": "3.4. Proximal Gradients for Partial Regularization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Eq. 5 has a similar formulation of the partial regularization of lasso [40]. In [40], the related optimization problem is solved via a nonmonotone proximal gradient (NPG) method. However, NPG requires frequent evaluation of the loss function to ensure the loss value after proximal gradients is less or equal to the loss value before the update. With CNNs, the costs of NPG are too large due to frequent loss evaluations. As a result, we use a one-step proximal gradient update to solve the problem defined in Eq. 4." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.817, + 0.806, + 0.832 + ], + "angle": 0, + "content": "The proximal operator of \\(\\mathcal{R}_w\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.842, + 0.892, + 0.874 + ], + "angle": 0, + "content": "\\[\n\\operatorname {p r o x} _ {\\gamma \\mathcal {R} _ {w}} (x) = \\underset {y} {\\arg \\min } \\gamma \\mathcal {R} _ {w} (y) + \\frac {1}{2} \\| x - y \\| ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.885, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In Eq. 6, \\( \\| x - y\\| ^2 \\) defines the sum of the square difference" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17405" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "between \\(x\\) and \\(y\\). We denote model weights after \\(t\\)th update as \\(\\mathcal{W}^t\\), and \\(\\mathcal{W}^{t + 1}\\) can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.13, + 0.469, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} ^ {t + 1} = \\operatorname {p r o x} _ {\\alpha^ {t + 1} \\gamma \\mathcal {R} _ {w}} (u (\\mathcal {W} ^ {t}, \\alpha^ {t + 1})), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.157, + 0.469, + 0.233 + ], + "angle": 0, + "content": "where \\( u(\\mathcal{W}^t, \\alpha^{t+1}) \\) is an update rule that can be applied for many algorithms, and \\( \\alpha^{t+1} \\) is the learning rate at the current step. Take SGD as an example, \\( u(\\mathcal{W}^t, \\alpha^{t+1}) = \\mathcal{W}^t - \\alpha^{t+1}\\nabla_{\\mathcal{W}^t}\\mathcal{L} \\). With Eq. 6 and Eq. 7, model weights \\( \\mathcal{W} \\) can then be updated with proximal gradients." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.325 + ], + "angle": 0, + "content": "We now present the analytical solution to Eq. 6, so that model weights can be efficiently updated. From the structure of \\(\\mathcal{R}_w\\), we know that channels with indices \\(i \\in S_l\\) will be regularized. This is equivalent to not regularizing channels if \\(i \\notin S_l\\). Consequently, we have the following form of the proximal gradient operator:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\operatorname {p r o x} _ {\\alpha \\gamma \\mathcal {R} _ {w}} (\\mathcal {W} _ {l}) = \\left\\{ \\begin{array}{c} \\frac {\\mathcal {W} _ {l [ i , ; ; ; ]}}{\\| (1 - \\mathbf {v} _ {l}) \\odot \\mathcal {W} _ {l} \\| _ {2}} \\max \\left(0, - \\frac {\\hat {N} _ {l}}{\\hat {N}} \\alpha \\gamma \\right. \\\\ \\left. + \\| (1 - \\mathbf {v} _ {l}) \\odot \\mathcal {W} _ {l} \\| _ {2}\\right), \\text {i f} i \\in S _ {l}, \\\\ \\mathcal {W} _ {l [ i, ; ; ; ; ]}, \\text {i f} i \\notin S _ {l}. \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.469, + 0.643 + ], + "angle": 0, + "content": "In Eq. 8, we omit the step notation \\(t\\) to simplify the notations, and we still assume \\(W_{l}\\) is pruned along the output dimension. In this equation, it is easier to see that \\(\\frac{\\hat{N}_l}{N}\\) can balance the regularization strength given different layers. Due to the property of our proposed partial regularization, the term \\(\\| (1 - \\mathbf{v}_l)\\odot \\mathcal{W}_l\\| _2\\) also changes dynamically, since \\(\\hat{N}_l\\) is changed after updates of the AGN. As a result, if no adjustment is applied, no matter how large or small \\(\\hat{N}_l\\) is, only a constant value \\(\\frac{1}{L}\\alpha \\gamma\\) will be used for the soft-thresholding, \\(\\max (0,\\| (1 - \\mathbf{v}_l)\\odot \\mathcal{W}_l\\| _2 - \\frac{1}{L}\\alpha \\gamma)\\), in all circumstances, which is not reasonable. To accompany the changes of \\(v_{l}\\) and consequently \\(\\hat{N}_l\\), we use \\(\\frac{\\hat{N}_l}{N}\\) to dynamically balance the soft-thresholding parameter between different layers. With Eq. 8, we can efficiently update \\(\\mathcal{W}\\) with our proposed partial regularization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.445, + 0.668 + ], + "angle": 0, + "content": "3.5. Network Pruning via Structural Alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We present the algorithm of our method in Algorithm 1. In Algorithm 1, \\( D \\) is the training dataset; \\( D_{\\mathrm{AGN}} \\) is a sub-dataset within \\( D \\) and it is used to train AGN; \\( p \\) decides how much FLOPs is preserved and it described in section 3.2; \\( \\gamma \\) and \\( \\lambda \\) are hyperparameters to control the strength of \\( \\mathcal{R}_{\\mathrm{FLOPs}} \\) and \\( \\mathcal{R}_w \\); \\( E \\) is the total number of epochs; \\( E_{\\mathrm{start}} \\) is the start epoch to train AGN and apply \\( \\mathcal{R}_w \\) when optimizing model weights. Note that, to reduce the overhead brought by training AGN, we only use a small sub-set sampled from \\( D \\). As we discussed in section 1, we need to set a start epoch for AGN and \\( \\mathcal{R}_w \\). If we apply \\( \\mathcal{R}_w \\) when training starts \\( (E_{\\mathrm{start}} = 0) \\), it will largely restrict the final performance of the whole model before pruning. Instead, we can have deserved results if we start partial regularization in the middle of the training process. The reason for this problem can" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.18 + ], + "angle": 0, + "content": "come from several perspectives. For example, at the beginning of the training, the classification loss can not produce accurate guidance for pruning since weights are not trained properly. Consequently, it will produce a bad sub-network, and following this sub-network only gives even worse results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.893, + 0.304 + ], + "angle": 0, + "content": "We summarize our method in Fig. 1. The inference path when training model weights and the AGN are different, and they are connected by partial regularization, which is different from current one-shot pruning and soft pruning methods. With this design, model weights are not directly affected by the sub-network architecture, which creates a smooth transition for the selected sub-network before and after pruning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.309, + 0.632, + 0.326 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.333, + 0.598, + 0.35 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.492 + ], + "angle": 0, + "content": "We use CIFAR-10 [26] and ImageNet [5] to evaluate the performance of our method. Our method requires one hyper-parameter \\( p \\) to control the FLOPs budget. The detailed choices of \\( p \\) are listed in supplementary materials. We choose ResNets [14] and MobileNet-V2 [49] for comparison. For CIFAR-10, we compare our method with other methods on ResNet-56 and MobileNetV2. For ImageNet, we select ResNet-34, ResNet-50, ResNet-101, and MobileNetV2 as our target models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.893, + 0.826 + ], + "angle": 0, + "content": "We set \\(\\lambda\\) in Eq. 3 to 4.0 for all models and datasets. Similarly, we set \\(\\gamma\\) to 0.0005 for all settings. We set \\(E_{\\mathrm{start}}\\) at \\(20\\%\\) of the total training epochs. Detailed numbers of \\(E_{\\mathrm{start}}\\) are listed in supplementary materials. The range of \\(E_{\\mathrm{start}}\\) is quite large, which is hard to be explored thoroughly. The current setting already provides good results, but better settings may also exist. To reduce the training costs of the AGN, we random sample \\(5\\%\\) of samples from the original dataset for constructing \\(D_{\\mathrm{AGN}}\\). With this setup, the additional costs are less than \\(5\\%\\) of the original training costs. We train the parameters \\(\\Theta\\) of the AGN using ADAM [25] with a start learning rate 0.001. Besides the training of the AGN, we follow the standard training recipe of ResNets for both CIFAR-10 and ImageNet. For MobileNet-V2, We follow the training settings in their original paper [49]. After training, we prune the model by using the sub-network generated from the AGN. The fine-tuning settings are similar to the training setting. Due to space limitations, details of training and fine-tuning are also presented in supplementary materials. In the experimental section, our method is abbreviated as SANP: Structural Alignment for Network Pruning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.676, + 0.848 + ], + "angle": 0, + "content": "4.2. CIFAR-10 Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The results of CIFAR-10 are presented in Tab. 1. For ResNet-56, our method achieves the best performance (in terms of \\(\\Delta\\)-Acc) compared to other methods. Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17406" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Table 1: Comparison of results on CIFAR-10. \\(\\Delta\\)-Acc represents the performance changes relative to the baseline, and \\(+ / -\\) indicates an increase/decrease, respectively." + }, + { + "type": "table", + "bbox": [ + 0.191, + 0.131, + 0.778, + 0.342 + ], + "angle": 0, + "content": "
ArchitectureMethodBaseline AccPruned AccΔ-AccPruned FLOPs
ResNet-56DCP-Adapt [60]93.80%93.81%+0.01%47.0%
SCP [23]93.69%93.23%-0.46%51.5%
FPGM [18]93.59%92.93%-0.66%52.6%
SFP [16]93.59%92.26%-1.33%52.6%
FPC [15]93.59%93.24%-0.25%52.9%
HRank [35]93.26%92.17%-0.09%50.0%
DMC [9]93.62%92.69%+0.07%50.0%
GNN-RL [57]93.49%93.59%+0.10%54.0%
SANP (ours)93.49%93.81%+0.32%52.0%
MobileNetV2Uniform [60]94.47%94.17%-0.30%26.0%
DCP [60]94.47%94.69%+0.22%26.0%
DMC [9]94.23%94.49%+0.26%40.0%
SCOP [53]94.48%94.24%-0.24%40.3%
SANP (ours)94.52%94.97%+0.45%46.0%
" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.353, + 0.473, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.488, + 0.323, + 0.501 + ], + "angle": 0, + "content": "(a) ResNet-56." + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.353, + 0.885, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.488, + 0.748, + 0.5 + ], + "angle": 0, + "content": "(b) MoibleNet-V2." + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.513, + 0.89, + 0.529 + ], + "angle": 0, + "content": "Figure 2: (a) and (b): the average norm of channels with or without partial regularization for ResNet-56 and MobileNet-V2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.474, + 0.771 + ], + "angle": 0, + "content": "our method outperforms the second best method GNN-RL by \\(0.22\\%\\) regarding \\(\\Delta\\)-Acc (SANP \\(+0.32\\%\\) vs. GNN-RL \\(+0.10\\%\\) when pruning similar FLOPs (SANP \\(52.0\\%\\) vs. GNN-RL \\(54.0\\%\\)). Our method also outperforms HRank and DMC by \\(0.25\\%\\) and \\(0.41\\%\\) separately (DMC \\(+0.07\\%\\) and HRank \\(-0.09\\%\\)). The gap between other methods and our method is even larger. For MobileNet-V2, our method prunes most FLOPs \\((46.0\\%)\\) and achieves the best performance (\\(\\Delta\\)-Acc: \\(+0.45\\%\\)). Compared to the second best method, DMC, our method prunes \\(6\\%\\) more FLOPs and outperforms it by \\(0.19\\%\\). SCOP also prunes \\(40\\%\\) FLOPs, and the gap between our method and SCOP is even larger \\((0.69\\%\\) better in terms of \\(\\Delta\\)-Acc). The uniform setting and DCP prunes around \\(26\\%\\) FLOPs, but the performance is still worse than our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.25, + 0.8 + ], + "angle": 0, + "content": "4.3. ImageNet Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.47, + 0.901 + ], + "angle": 0, + "content": "All results for the ImageNet dataset are shown in Tab. 2. ResNet-34. Our method achieves \\(73.43\\%\\) Top-1 accuracy and \\(91.48\\%\\) Top-5 accuracy, which is better than the baseline by \\(0.19\\%\\) and \\(0.16\\%\\) for Top-1/Top-5 accuracy separately. At the same time, our method removes \\(44.1\\%\\) FLOPs, which is on par with other methods. It is obvious" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.54, + 0.692, + 0.654 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.659, + 0.608, + 0.672 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.539, + 0.889, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.659, + 0.805, + 0.671 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.685, + 0.89, + 0.714 + ], + "angle": 0, + "content": "Figure 3: Top-1 and Top-5 accuracy after pruning given different settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.869 + ], + "angle": 0, + "content": "that the advantage of our method is clear compared to other methods. Our method prunes similar FLOPs to SCOP and DMC. However, the \\(\\Delta\\) Top-1 Acc of our method is \\(0.88\\%\\) and \\(0.92\\%\\) better than SCOP and DMC, respectively, and we have similar observations for \\(\\Delta\\) Top-5 Acc (0.60% and \\(0.47\\%\\) better than SCOP and DMC). Taylor has the second best \\(\\Delta\\) Top-1 Acc, but the pruned FLOPs is much lower than our method (ours: \\(44.1\\%\\) vs. Taylor: \\(24.2\\%\\)). The advantage of our method compared to FPGM is more obvious." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ResNet-50. Our method achieves \\(76.47\\%\\) Top-1 accuracy and \\(93.00\\%\\) Top-5 accuracy, which is also better than the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17407" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.202, + 0.089, + 0.768, + 0.105 + ], + "angle": 0, + "content": "Table 2: Comparison results on ImageNet with ResNet-34/50/101 and MobileNet-V2." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.116, + 0.828, + 0.453 + ], + "angle": 0, + "content": "
ArchitectureMethodBaseline Top-1 AccBaseline Top-5 AccΔ Top-1 AccΔ Top-5 AccPruned FLOPs
ResNet-34FPGM [18]73.92%91.62%-1.29%-0.54%41.1%
Taylor [42]73.31%--0.48%-24.2%
DMC [9]73.30%91.42%-0.73%-0.31%43.4%
SCOP [53]73.31%91.42%-0.69%-0.44%44.8%
SANP (ours)73.24%91.32%+0.19%+0.16%44.1%
ResNet-50DCP [60]76.01%92.93%-1.06%-0.61%55.6%
CCP [44]76.15%92.87%-0.94%-0.45%54.1%
FPGM [18]76.15%92.87%-1.32%-0.55%53.5%
ABCP [36]76.01%92.96%-2.15%-1.27%54.3%
DMC [9]76.15%92.87%-0.80%-0.38%55.0%
SCOP [53]76.15%92.87%-0.89%-0.34%54.6%
PFP [34]76.13%92.86%-0.92%-0.45%44.0%
CHIP [52]76.15%92.87%+0.00%+0.04%48.7%
NPPM [8]76.15%92.87%-0.19%+0.12%56.0%
Random-Pruning [30]75.83%92.92%-0.75%-0.40%51.0%
GNN-RL [57]76.10%--1.82%-53.0%
SANP (ours)76.06%92.86%+0.41%+0.17%56.2%
ResNet-101FPGM [18]77.37%93.56%-0.05%0.00%41.1%
Taylor [42]77.37%--0.02%-39.8%
DMC [9]77.37%93.56%+0.04%+0.03%56.0%
PFP [34]77.37%93.56%-0.94%-0.44%45.1%
SANP (ours)77.53%93.71%+0.61%+0.29%55.4%
MobileNet-V2Uniform [49]71.80%91.00%-2.00%-1.40%30.0%
AMC [17]71.80%--1.00%-30.0%
CC [33]71.88%--0.97%-28.3%
MetaPruning [38]72.00%--0.80%-30.7%
Random-Pruning [30]71.88%--1.01%-29.1%
SANP (ours)71.91%90.30%+0.14%+0.07%29.1%
" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.468, + 0.467, + 0.544 + ], + "angle": 0, + "content": "
SettingsArchitectureBaseline Top-1 AccΔ Top-1 AccPruned FLOPs
One-Shot SANPResNet-3473.31%-0.50%44.0%
73.24%+0.19%44.1%
One-Shot SANPResNet-5076.13%-0.57%56.0%
76.06%+0.41%55.2%
One-Shot SANPMobileNetV271.88%-0.42%29.3%
71.91%+0.07%29.1%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.555, + 0.47, + 0.585 + ], + "angle": 0, + "content": "Table 3: Performance of pruned models given different pruning settings on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.806 + ], + "angle": 0, + "content": "baseline Top-1/Top-5 accuracy. The second best method, CHIP, removes \\(48.7\\%\\) FLOPs while maintaining the original performance. Our method outperforms CHIP by \\(0.41\\%\\) in terms of the \\(\\Delta\\) Top-1 accuracy while pruning near \\(8\\%\\) more FLOPs. NPPM is a strong baseline for ResNet-50, our method outperforms by \\(0.60\\%\\) in \\(\\Delta\\) -Top-1 Acc. GNN-RL and Random-Pruning are two recent pruning works. Our method outperforms them by \\(2.23\\%\\) and \\(1.16\\%\\) separately, while our method prunes more FLOPs. PHP, DCP, CCP, and DMC have similar \\(\\Delta\\) -Top 1 accuracy. The gap between our method and these methods ranges from \\(1.21\\%\\) to \\(1.47\\%\\) regarding \\(\\Delta\\) -Top 1 accuracy. The advantage of our method compared to the rest methods is more apparent." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "ResNet-101. Our method achieves \\(78.14\\%\\) Top-1 accuracy and \\(94.00\\%\\) Top-5 accuracy, which is \\(0.61\\%\\) and \\(0.29\\%\\) better than the baseline Top-1 and Top-5 accuracy. Although DMC prunes a little bit more FLOPs, it only achieves \\(77.41\\%\\) Top-1 accuracy after pruning which is \\(0.73\\%\\) lower than our method. FPGA, Taylor, and PFP remove less than" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.514 + ], + "angle": 0, + "content": "50% FLOPs. Our method prunes at least 10% more FLOPs and still has an advantage in terms of \\( \\Delta \\)-Top 1 accuracy (from 0.66% to 1.55%)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.893, + 0.697 + ], + "angle": 0, + "content": "MobileNet-V2. MobileNet-V2 is generally harder to prune compared to ResNets. All comparison methods on MobileNet-V2 remove around \\(30\\%\\) FLOPs. Our method achieves \\(72.05\\%\\) Top-1 accuracy and \\(90.37\\%\\) Top-5 accuracy, which is \\(0.14\\%\\) and \\(0.07\\%\\) better than the baseline Top-1 and Top-5 accuracy. Given the similar pruning rate, our method is \\(1.15\\%\\), \\(0.94\\%\\), \\(1.11\\%\\) and \\(1.14\\%\\) higher than Random-Pruning, MetaPruning CC, and AMC separately regarding \\(\\Delta\\)-Top-1 Acc. MetaPruning prunes most FLOPs, but the performance is much lower than our method. In short, our method can also be applied to lightweight CNNs, like MobileNet-V2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.706, + 0.722, + 0.722 + ], + "angle": 0, + "content": "4.4. Analysis of Our Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The effectiveness of partial regularization. To verify whether our proposed partial regularization is effective, we plot the average channel norm of different groups within each block for ResNet-56 and MobileNet-V2 on CIFAR-10. The results are shown in Fig. 2. The average channel norm for the group with partial regularization and without partial regularization are obtained by \\(\\frac{1}{\\tilde{N}_l}\\sum_{i\\in S_l}\\| \\mathcal{W}_{l[i,:;:]} \\|_{\\mathrm{GL}}\\) and \\(\\frac{1}{C_l - \\tilde{N}_l}\\sum_{i\\notin S_l}\\| \\mathcal{W}_{l[i,:;:]} \\|_{\\mathrm{GL}}\\) separately. Weights with partial regularization are effectively aligned. On the contrary, weights without partial regularization are lightly affected, which justifies the strength of our proposed partial regular" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17408" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.255, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.204, + 0.18, + 0.216 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.088, + 0.465, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.204, + 0.39, + 0.216 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.673, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.204, + 0.6, + 0.216 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.088, + 0.885, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.203, + 0.811, + 0.215 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.218, + 0.253, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.161, + 0.332, + 0.18, + 0.344 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.217, + 0.465, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.332, + 0.389, + 0.344 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.218, + 0.674, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.332, + 0.601, + 0.345 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.217, + 0.885, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.332, + 0.811, + 0.344 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.357, + 0.894, + 0.403 + ], + "angle": 0, + "content": "Figure 4: (a, e): the impact of \\(\\lambda\\) in \\(\\mathcal{R}_{\\mathrm{FLOPs}}\\). (b, f): the effect of the architecture of AGN. (c, g): the effect of \\(E_{\\mathrm{start}}\\). (d, h): the effect of different setups. Experiments are conducted on CIFAR-10 with ResNet-56 and \\(p = 0.5\\) (a,b,c,d,e) and \\(p = 0.35\\) (f,g,h)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.424, + 0.131, + 0.437 + ], + "angle": 0, + "content": "ization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.516 + ], + "angle": 0, + "content": "The impact of \\(\\lambda\\). We study the impact of \\(\\lambda\\) when training the AGN, and we plot the test accuracy and \\(\\mathcal{R}_{\\mathrm{FLOPs}}\\) in Fig. 4a and Fig. 4e. From the figures, we can see that if \\(\\lambda\\) is too large, it will have negative impacts on learned subnetworks. Otherwise, our method is robust to \\(\\lambda\\)," + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.639 + ], + "angle": 0, + "content": "The effect of AGN. In Fig. 4b and Fig. 4f, we plot the test accuracy when learning the AGN given different pruning rates. We construct a Simple baseline, which parameterizes each channel by using one learnable parameter. We can see that when the parameterization space shrinks, the performance of learned sub-networks will be affected severely. In addition, the learning is slower, and the best sub-network performance is also much worse than using AGN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.471, + 0.807 + ], + "angle": 0, + "content": "The effect of \\(\\mathbf{E}_{\\mathrm{start}}\\). In the method section, we argue that inserting partial regularization in the middle training process is beneficial. We plot the results of \\(E_{\\mathrm{start}} = 0\\) and \\(E_{\\mathrm{start}} = 40\\) in Fig. 4c and Fig. 4g. In general, they can find sub-networks with similar performance, but \\(E_{\\mathrm{start}} = 0\\) is worse than \\(E_{\\mathrm{start}} = 40\\) when the pruning rate is large (\\(p = 0.35\\)). More importantly, the full model accuracy of \\(E_{\\mathrm{start}} = 0\\) is \\(92.96\\%\\) (average across different runs), which is around \\(0.50\\%\\) worse than \\(E_{\\mathrm{start}} = 40\\), which suggests that using partial regularization at the beginning will limit the capacity of the full model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The effect of different setups. We construct two additional baselines to see how partial regularization helps to produce a better sub-network within the full model. The One-Shot baseline directly trains the AGN on the pre-trained model. The w/o Partial Regularization baseline set \\(\\gamma = 0\\), and the rest settings are the same as our method. The related" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.895, + 0.68 + ], + "angle": 0, + "content": "results are shown in Fig. 4d and Fig. 4h. From these figures, we can see that partial regularization always produces the best sub-network from the full model. In addition, 'w/o partial regularization' is worse than the one-shot setting. This is probably because it is hard to capture the changes of model weights without partial regularization, which makes the training of AGN much harder than the rest settings. More comparisons on the ImageNet dataset. To further show how our method improves the one-shot setting, we present the sub-network performance before and after fine-tuning for one-shot and partial regularization settings in Fig. 3 and Tab. 3. From Fig. 3, we can see that partial regularization still produces better sub-networks on ImageNet, and it is around \\(10\\%\\) better in terms of Top-1 accuracy than the one-shot setting for different models. The advantage of partial regularization naturally extends to results after fine-tuning, as shown in Tab. 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.705, + 0.619, + 0.721 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this paper, we investigate how partial regularization helps to produce a better sub-network for network pruning. Specifically, our method uses AGN to guide partial regularization across the training process. We further provide an efficient way to update model weights through proximal gradients. With these designs, partial regularization effectively reduces the gap between the sub-network within the full model and the pruned model. Our method then starts from a better sub-network, thus resulting in a better final pruned model. Extensive experimental results on CIFAR-10 and ImageNet show the effectiveness of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17409" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[3] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.358 + ], + "angle": 0, + "content": "[4] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in neural information processing systems, pages 3123-3131, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.472, + 0.429 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.472, + 0.5 + ], + "angle": 0, + "content": "[6] Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, and H.T. Kung. Neural mean discrepancy for efficient out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19217-19227, June 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.472, + 0.557 + ], + "angle": 0, + "content": "[7] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.472, + 0.628 + ], + "angle": 0, + "content": "[8] Shangqian Gao, Feihu Huang, Weidong Cai, and Heng Huang. Network pruning via performance maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9270-9280, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.472, + 0.7 + ], + "angle": 0, + "content": "[9] Shangqian Gao, Feihu Huang, Jian Pei, and Heng Huang. Discrete model compression with resource constraint for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1899-1908, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[10] Shangqian Gao, Burak Uzkent, Yilin Shen, Heng Huang, and Hongxia Jin. Learning to jointly share and prune weights for grounding based vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[11] Junfeng Guo, Yiming Li, Xun Chen, Hanqing Guo, Lichao Sun, and Cong Liu. SCALE-UP: An efficient black-box input-level backdoor detection via analyzing scaled prediction consistency. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[13] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Advances in neural information processing systems, pages 1135-1143, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[15] Yang He, Yuhang Ding, Ping Liu, Linchao Zhu, Hanwang Zhang, and Yi Yang. Learning filter pruning criteria for deep convolutional neural networks acceleration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2009-2018, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[16] Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating deep convolutional neural networks. In International Joint Conference on Artificial Intelligence (IJCAI), pages 2234-2240, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.339, + 0.892, + 0.407 + ], + "angle": 0, + "content": "[17] Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, and Song Han. Amc: Automl for model compression and acceleration on mobile devices. In Proceedings of the European Conference on Computer Vision (ECCV), pages 784-800, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.41, + 0.892, + 0.479 + ], + "angle": 0, + "content": "[18] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.482, + 0.892, + 0.522 + ], + "angle": 0, + "content": "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[20] Zehao Huang and Naiyan Wang. Data-driven sparse structure selection for deep neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 304-320, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.584, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[21] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32Nd International Conference on International Conference on Machine Learning - Volume 37, ICML, pages 448-456. JMLR.org, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.656, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[22] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[23] Minsoo Kang and Bohyung Han. Operation-aware soft channel pruning using differentiable masks. In International Conference on Machine Learning, pages 5122-5131. PMLR, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[24] Jaedeok Kim, Chiyoun Park, Hyun-Joo Jung, and Yoonsuck Choe. Plug-in, trainable gate for streamlining arbitrary neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, Cite-seer, 2009. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17410" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097-1105, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[28] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. *ICLR*, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.231 + ], + "angle": 0, + "content": "[29] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. *ICLR*, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.233, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[30] Yawei Li, Kamil Adamczewski, Wen Li, Shuhang Gu, Radu Timofte, and Luc Van Gool. Revisiting random channel pruning for neural network compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 191-201, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.303, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[31] Yawei Li, Shuhang Gu, Christoph Mayer, Luc Van Gool, and Radu Timofte. Group sparsity: The hinge between filter pruning and decomposition for network compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8018-8027, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.373, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[32] Yawei Li, Shuhang Gu, Kai Zhang, Luc Van Gool, and Radu Timofte. Dhp: Differentiable meta pruning via hypernetworks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 608-624. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[33] Yuchao Li, Shaohui Lin, Jianzhuang Liu, Qixiang Ye, Mengdi Wang, Fei Chao, Fan Yang, Jincheng Ma, Qi Tian, and Rongrong Ji. Towards compact cnns via collaborative compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6438-6447, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.525, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[34] Lucas Liebenwein, Cenk Baykal, Harry Lang, Dan Feldman, and Daniela Rus. Provable filter pruning for efficient neural networks. In International Conference on Learning Representations, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.581, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[35] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. Hrank: Filter pruning using high-rank feature map. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.651, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[36] Mingbao Lin, Rongrong Ji, Yuxin Zhang, Baochang Zhang, Yongjian Wu, and Yonghong Tian. Channel pruning via automatic structure search. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages 673 - 679, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.721, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[37] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[38] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In Proceedings of the IEEE International Conference on Computer Vision, pages 3296-3305, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[39] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[40] Zhaosong Lu and Xiaorui Li. Sparse recovery via partial regularization: Models, theory, and algorithms. Mathematics of Operations Research, 43(4):1290-1316, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[41] Dmitry Molchanov, Arsenii Ashukha, and Dmitry Vetrov. Variational dropout sparsifies deep neural networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 2498-2507. JMLR.org, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[42] Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11264–11272, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[43] Kirill Neklyudov, Dmitry Molchanov, Arsenii Ashukha, and Dmitry P Vetrov. Structured bayesian pruning via log-normal multiplicative noise. In Advances in Neural Information Processing Systems, pages 6775-6784, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[44] Hanyu Peng, Jiaxiang Wu, Shifeng Chen, and Junzhou Huang. Collaborative channel pruning for deep networks. In International Conference on Machine Learning, pages 5113-5122, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[45] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In European Conference on Computer Vision, pages 525-542. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[46] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[47] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91-99, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.592 + ], + "angle": 0, + "content": "[48] Alex Renda, Jonathan Frankle, and Michael Carbin. Comparing rewinding and fine-tuning in neural network pruning. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[50] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in neural information processing systems, pages 568-576, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[52] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. Chip: Channel independence-based pruning for compact neural networks. Advances in Neural Information Processing Systems, 34, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[53] Yehui Tang, Yunhe Wang, Yixing Xu, Dacheng Tao, Chunjing Xu, Chao Xu, and Chang Xu. Scop: Scientific control for reliable neural network pruning. Advances in Neural Information Processing Systems, 33, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17411" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "In Advances in neural information processing systems, pages 2074-2082, 2016.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[55] Mao Ye, Chengyue Gong, Lizhen Nie, Denny Zhou, Adam Klivans, and Qiang Liu. Good subnetworks provably exist: Pruning via greedy forward selection. In International Conference on Machine Learning, pages 10820-10830. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[56] Zhonghui You, Kun Yan, Jinmian Ye, Meng Ma, and Ping Wang. Gate decorator: Global filter pruning method for accelerating deep convolutional neural networks. In Advances in Neural Information Processing Systems, pages 2130-2141, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.468, + 0.319 + ], + "angle": 0, + "content": "[57] Sixing Yu, Arya Mazaheri, and Ali Jannesari. Topology-aware network pruning using multi-stage graph embedding and reinforcement learning. In International Conference on Machine Learning, pages 25656-25667. PMLR, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.361 + ], + "angle": 0, + "content": "[58] Dejiao Zhang, Haozhu Wang, Mario Figueiredo, and Laura Balzano. Learning to share: Simultaneous parameter tying and sparsification in deep learning. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[59] Yanfu Zhang, Shangqian Gao, and Heng Huang. Exploration and estimation for model compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 487-496, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[60] Zhuangwei Zhuang, Mingkui Tan, Bohan Zhuang, Jing Liu, Yong Guo, Qingyao Wu, Junzhou Huang, and Jinhui Zhu. Discrimination-aware channel pruning for deep neural networks. In Advances in Neural Information Processing Systems, pages 875-886, 2018. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17412" + } + ] +] \ No newline at end of file diff --git a/2023/Structural Alignment for Network Pruning through Partial Regularization/bbe97c12-8d15-479d-81ec-fff8ed1209f4_origin.pdf b/2023/Structural Alignment for Network Pruning through Partial Regularization/bbe97c12-8d15-479d-81ec-fff8ed1209f4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f597f395ee40b975ee2e3282fd9f664451608eb3 --- /dev/null +++ b/2023/Structural Alignment for Network Pruning through Partial Regularization/bbe97c12-8d15-479d-81ec-fff8ed1209f4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8cf25d2b079a0b115d5fc5e1fcdb0b6a4a9c728bc7e880694afb762913f09a7 +size 3325422 diff --git a/2023/Structural Alignment for Network Pruning through Partial Regularization/full.md b/2023/Structural Alignment for Network Pruning through Partial Regularization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d39787ece53f6eee265ab889deb72d9cdc469e77 --- /dev/null +++ b/2023/Structural Alignment for Network Pruning through Partial Regularization/full.md @@ -0,0 +1,316 @@ +# Structural Alignment for Network Pruning through Partial Regularization + +Shangqian Gao $^{1}$ , Zeyu Zhang $^{2}$ , Yanfu Zhang $^{1}$ , Feihu Huang $^{1}$ , and Heng Huang $^{*3}$ + +$^{1}$ Department of Electrical and Computer Engineering, University of Pittsburgh $^{2}$ School of Information, University of Arizona +$^{3}$ Department of Computer Science, University of Maryland at College Park + +{shg84, yaz91}@pitt.edu, zeyuzhang@email.arizona.edu, huangfeihu2018@gmail.com, heng@umd.edu + +# Abstract + +In this paper, we propose a novel channel pruning method to reduce the computational and storage costs of Convolutional Neural Networks (CNNs). Many existing one-shot pruning methods directly remove redundant structures, which brings a huge gap between the model before and after network pruning. This gap will no doubt result in performance loss for network pruning. To mitigate this gap, we first learn a target sub-network during the model training process, and then we use this sub-network to guide the learning of model weights through partial regularization. The target sub-network is learned and produced by using an architecture generator, and it can be optimized efficiently. In addition, we also derive the proximal gradient for our proposed partial regularization to facilitate the structural alignment process. With these designs, the gap between the pruned model and the sub-network is reduced, thus improving the pruning performance. Empirical results also suggest that the sub-network found by our method has a much higher performance than the one-shot pruning setting. Extensive experiments show that our method can achieve state-of-the-art performances on CIFAR-10 and ImageNet with ResNets and MobileNet-V2. + +# 1. Introduction + +Convolutional Neural Networks (CNNs) have achieved many successes in different computer vision tasks [27, 46, 47, 50, 2, 6, 11]. To tackle real-world challenges, recent CNNs [27, 51, 14] become larger and larger regarding width, depth, etc. With such capacities, CNNs can obtain better performance on different benchmarks at the cost + +*Corresponding author. This work was partially supported by NSF IIS 1838627, 1837956, 1956002, 2211492, CNS 2213701, CCF 2217003, DBI 2225775. + +of computational and storage burdens. At the same time, with the recent developments of mobile and embedded devices, the demand for deploying CNNs on these devices has increased dramatically. However, there is a natural conflict between the size of CNNs and the hardware capability of these devices. To overcome these challenges, many works [13, 12] try to reduce the size of CNNs, and make them possible to be deployed on edge devices. + +There are many directions to reduce the size of CNNs. Among them, weight pruning and structural pruning are two popular topics. Structural pruning, especially channel pruning, is more friendly to hardware than weight pruning since no post-processing steps are required to acquire savings in computational and storage costs. Thus, our paper focuses on channel pruning for CNNs. Many existing one-shot pruning works [44, 56, 9, 42, 17, 8] prune trained models directly. No matter what method is used, there will be a significant gap between the selected sub-network and the pruned model. Such a gap creates difficulties in regaining performance during the fine-tuning process. On the other hand, soft pruning methods [16, 23] softly remove structures during the training process, which can produce good results with a shorter fine-tuning process. However, soft pruning methods generally perform worse than typical one-shot pruning methods, probably because the weight space is restricted during the training process because of soft pruning. + +To tackle the above problems, we introduce a novel partial regularization technique to align model weights and the discovered sub-network during the training process, which can produce a high performance sub-network and reduce the gap between the sub-network and the original model. In addition, unlike soft-pruning methods, all model structures are used for training. The partial regularization term contains a partial group lasso regularization on selected weights, and other weights remain intact without modifica + +tions. An architecture generator is trained to select which weights should be aligned, and it is also updated during the training process. Our partial regularization formulation is related to partial regularization in lasso [40]. Inspired by the nonmonotone proximal gradient (NPG) method used in [40], we also use a proximal gradient method to solve the partial regularization problem in our setting. Note that our method dynamically changes which channels should be put in the partial regularization. As a result, we add a scalar to balance the regularization strength for different layers because the number of pruned channels is different for them. To maximally keep the capacity of the original model, we insert the partial regularization in the middle of the training process. This is because weights are vulnerable to pruning at the early training stage, and the FLOPs regularization will dominate updates of the architecture generator, which can create bad sub-networks and mislead the training process. Finally, we update model weights and the architecture generator periodically, and they are connected by the partial regularization term during training. To maintain similar training efficiency as the original model, we only use a small portion of samples to train the architecture generator. Thus, there is only a small overhead compared to the original training process. Our method successfully finds performant sub-networks from the original model with these techniques. In summary, the contributions of this paper can be summarized as follows: + +1) We propose to align the sub-network in the original model with the final pruned model through partial regularization. By structural alignment, the gap between the selected sub-network and the pruned model is largely reduced, which naturally improves the performance of the pruned model. +2) We use an architecture generator parameterized by neural networks to select the proper sub-network structure and guide the partial regularization. Inspired by partial regularization in lasso [40], we propose to solve the partial regularization problem via proximal gradients, which facilitate the alignment process. +3) Empirical results show that the sub-network discovered by our method performs much better than the one-shot pruning setting. Extensive experiments on CIFAR-10 and ImageNet show that our method outperforms existing channel pruning methods on ResNets and MobileNet-V2. + +# 2. Related Works + +Weight-Level Pruning. Weight pruning removes redundant connections based on individual weights. High compression rates can be achieved by weight pruning, but they can not directly achieve acceleration, and specially designed sparse matrix libraries are required. One of the early + +works [13] proposes to measure the importance of weights with their $L_{1}$ or $L_{2}$ magnitude, and unimportant weights are removed. Instead of magnitude, SNIP [28] updates the importance of each weight by using gradients from the loss function. SNIP can be used at the initialization time. The assumption of the Lottery ticket hypothesis [7] suggests that there exist sparse sub-networks (winning tickets) that can achieve the performance of the full model. In addition, with repeated training and fine-tuning, it can achieve better results. On the other hand, rethinking network pruning [39] argues that the learned topology from pruning algorithms is the key to achieving better performance. In addition, weight rewinding [48] shows that resetting weights to values from the middle training process can also produce good results. Similar to weight rewinding, our method does not modify weight training at the beginning, the partial regularization is inserted in the middle training process. + +Structural-Level Pruning. Different from weight pruning, structural pruning is more friendly to hardware since it requires little or no post-processing steps to achieve acceleration. Similar to weight pruning, one of the early structural pruning methods [29] measure the importance of filters by using the sum of the absolute value of kernel weights. Besides using the magnitude to measure channel or filter importance, other methods utilize the scaling factor of batchnorm [21] to indicate which channels are important because batchnorm [21] is popular for the design of recent CNNs [14, 49]. To prune channels, Liu et al. [37] apply the sparse regularization to the scaling factors of batchnorm, and the channel will be pruned if the corresponding scaling factor is small. Structure sparse selection [20] introduces scaling factor to specific structures, such as neurons, groups, or residual blocks, and the sparsity regularization is applied to these structures. Structures with small values will also be removed. Another line of research formulates channel pruning as a constrained optimization problem [24, 56, 9, 23, 59], and learnable parameters are used to control each channel. These parameters are end-to-end differentiable, which is amenable to gradient based optimization methods. Our method is also related to these researches. Different from these methods, the learning of sub-networks accomplishes the training of model weights in our method. In addition, we use partial regularization to promote the selected sub-network, which reduces the gap between the pruned model and the sub-network. Besides using gates, Automatic Model Compression (AMC) [17] uses policy gradient to update the policy network. This policy network is then used to decide the left width of each layer. Collaborative channel pruning [44] prunes channels by exploiting inter-channel dependency. Greedy forward selection [55] starts from an empty network and greedily adds Important channels from the full model. MetaPruning [38, 32] uses a hypernet to generate parameters for sub + +![](images/6df72f0c9c5234eb531cda79203f02f34d04ab942b85e39f4903a5e347d60045.jpg) +Figure 1: Overview of the proposed method. The AGN generates the sub-network to guide the partial regularization during the training process of model weights. The AGN is trained by evaluating sub-networks on a sub-set from the whole dataset. Both model weights and AGN are trained in an end-to-end differentiable manner. + +networks, and evolutionary algorithms are utilized to find the best sub-networks. MetaPruning shows that pruning should accomplish trained model weights. + +Besides these pruning methods, regularization based pruning methods can also be applied to structural pruning. Previous works [54, 31] use group sparsity to prune different structures. In addition to structural sparsity, structural sharing is considered in other works [58, 10]. Our method relates to regularization based methods; however, the formulation of our partial regularization only aligns selected channels dynamically, and other weights are intact. + +Other Related Works. Besides the above-mentioned methods, there are works from other perspectives, including bayesian pruning [41, 43], weight quantization [4, 45], and knowledge distillation [19]. + +# 3. Proposed Method + +# 3.1. Overview + +Before introducing our method, we first describe notations and provide an overview of our method. In a CNN, the feature map of $l$ th layer can be represented by $\mathcal{F}_l \in \Re^{C_l \times W_l \times H_l}$ , $l = 1, \ldots, L$ , where $C_l$ is the number of channels, $H_l$ and $W_l$ are height and width of the current feature map, $L$ is the number of layers. Similarly, the weights of $l$ th layer can be written as $\mathcal{W}_l \in \Re^{C_l \times C_{l-1} \times k_l \times k_l}$ , and $k_l$ is the kernel size of this layer. The mini-batch dimension of feature maps is ignored to simplify notations. + +The core motivation of our proposed method is to reduce the gap between the selected sub-network from the origi + +nal model and the pruned model. To achieve this goal, we need two processes. First, we need to choose the desired sub-network from the original model, and this sub-network should also be updated when model weights are trained during the learning process. Secondly, we use the sub-network to guide the partial regularization term, which is used to decide which weights should be regularized. In addition, partial regularization should not be fixed to accommodate the changes in the selected sub-network. + +# 3.2. Learning the Sub-network + +We use an Architecture Generator Network (AGN) to generate the desired sub-network architecture $\mathbf{v} \in \{0,1\}^N$ , where 0 or 1 is used to depict the removal or keep of a channel, and $N$ is the total number of channels from all layers. The large parameterization space of AGN can facilitate the learning of sub-network structures. To generate $\mathbf{v}$ , the following equation is used: + +$$ +\mathbf {v} = \operatorname {A G N} (\Theta), \tag {1} +$$ + +and AGN is composed of gated recurrent unit (GRU) [3] and dense layers. In addition, Gumbel-Sigmoid [22] with STE [1] are used to produce the final binary vector $\mathbf{v}$ , and they are placed after the output of dense layers. More details of AGN are provided in the supplementary materials. + +Once we have $\mathbf{v}$ , we can apply it to the feature maps to produce a sub-network. The feature map of the $l$ th layer is then modified as follows: + +$$ +\widehat {\mathcal {F}} _ {l} = \mathbf {v} _ {l} \odot \mathcal {F} _ {l}, \tag {2} +$$ + +where $\odot$ is element-wise multiplication, $\mathbf{v}_l$ is the architecture vector of $l$ th layer, and $\mathbf{v}_l$ is resized to have the same size of $\mathcal{F}_l$ . The feature map $\mathcal{F}_l$ is from the output of the activation function. The overall loss function for generating the desired sub-network is as follows: + +$$ +\min _ {\Theta} \mathcal {I} _ {\theta} (\Theta) := \mathcal {L} \left(f (x; \mathcal {W}, \mathbf {v}), y\right) + \lambda \mathcal {R} _ {\text {F L O P s}} (T (\mathbf {v}), p T _ {\text {t o t a l}}) \tag {3} +$$ + +where $T(\mathbf{v})$ is the current FLOPs decided by the vector $\mathbf{v}$ , $T_{\mathrm{total}}$ is the total FLOPs of the original model, $p \in (0,1]$ is a hyperparameter deciding the remaining fraction of FLOPs, $\lambda$ is the hyper-parameter controlling the strength of FLOPs regularization, $f(x; \mathcal{W}, \mathbf{v})$ is a CNN parameterized by $\mathcal{W}$ and the sub-network structure is determined by the architecture vector $\mathbf{v}$ , $\mathcal{L}$ is the task loss function and $\mathcal{R}_{\mathrm{FLOPs}}$ is the regularization term for FLOPs. The regularization term $\mathcal{R}_{\mathrm{FLOPs}}$ is $\mathcal{R}_{\mathrm{FLOPs}}(x,y) = \log (\max (x,y) / y)$ . With Eq. 3, we can find promising sub-networks when training the AGN and model weights periodically. + +# 3.3. Partial Regularization + +Given a sub-network $\mathbf{v}$ obtained from AGN, we can then reduce the gap between the sub-network and the pruned model and thus increase its performance. We can formulate the optimization problem with partial regularization as follows: + +$$ +\min _ {\mathcal {W}} \mathcal {J} _ {w} (\mathcal {W}) := \mathcal {L} (f (x; \mathcal {W}), y) + \gamma \mathcal {R} _ {w} (\mathcal {W}), \tag {4} +$$ + +where $\mathcal{R}_w$ is the partial regularization term, and $\gamma$ controls the strength of the partial regularization. $\mathcal{R}_w$ has the following form: + +$$ +R _ {w} (\mathcal {W}) = \sum_ {l = 1} ^ {L} \sum_ {i \in S _ {l}} \frac {\hat {N} _ {l}}{\hat {N}} \| \mathcal {W} _ {[ i,:,:,: ]} \| _ {\mathrm {G L}}, \tag {5} +$$ + +where $\hat{N}_l = \sum 1 - \mathbf{v}_l$ , $\hat{N} = \sum 1 - \mathbf{v}$ and $S_{l} = \{i\mid \mathrm{if}\mathbf{v}_{l[i]} = 0\}$ . $S_{l}$ contains the indices of pruned channels which are decided by AGN. $\frac{\hat{N}_l}{\hat{N}}$ is a scalar to adjust the regularization strength given different layers. The numerator $\hat{N}_l$ is the number of pruned channels of the $l$ th layer, and the denominator $\hat{N}$ is the number of pruned channels from all layers. It is easy to see that $\sum_{l=1}^{L} \frac{\hat{N}_l}{\hat{N}} = 1$ . $\|x\|_{\mathrm{GL}}$ is the norm of grouped weights for Group Lasso, and $\|x\|_{\mathrm{GL}} = \sqrt{\sum_{i=1}^{|x|} x_i^2}$ where $|x|$ represents the number of elements in $x$ . In Eq. 5, we assume the corresponding layer is pruned across the output dimension. If it is pruned across the input dimension, we have $\sum_{i \in S_l} \|W_{l[:i,::]}\|_{\mathrm{GL}}$ . + +The goal of using Group Lasso for our partial regularization is to reduce the distance between the dense model and the pruned model, other suitable functions may also be useful, but we found that the partial regularization with Group + +Algorithm 1: Structural Alignment for Network Pruning +Input: $D,D_{\mathrm{AGN}},p,\lambda ,\gamma ,E,E_{\mathrm{start}}$ +Initialization: initialize $\mathcal{W}$ and $\theta$ +for $e:=1$ to $E$ do +/\* Optimizing model weights. Freeze $\Theta$ of the AGN. /\* for a mini-batch $(x,y)$ in $D$ do 1. calculate the gradients w.r.t model weights: $\nabla_{\mathcal{W}}\mathcal{L}$ 2. update model weights using any stochastic optimizer. 3.if if e $\geq E_{\mathrm{start}}$ then generate v from the AGN by using Eq. 1. apply the proximal gradient step following Eq.8. +end +/\* Optimizing $\Theta$ of the AGN. Freeze model weights W. /\* if if e $\geq E_{\mathrm{start}}$ then for a mini-batch $(x,y)$ in $D_{\mathrm{AGN}}$ do 1. generate v from the AGN by using Eq. 1 and apply it to the model. 2. calculate gradients w.r.t to $\mathcal{J}$ in Eq. 3: $\nabla_{\Theta}\mathcal{J}_{\theta}$ 3.update the AGN with ADAM end +end +Pruning the model with resulting v, and fine-tuning it. + +Lasso already produces good results. Another benefit of the partial regularization $\mathcal{R}_w$ is that it will not penalize weights that are not pruned. By doing so, we can avoid the problem of overpenalizing all weights' magnitude. In addition, $\mathbf{v}$ is updated during the training process, and $\mathcal{R}_w$ will dynamically regularize weights. As a result, $\mathbf{v}$ can flexibly change instead of falling into a fixed sub-network during the optimization process. + +# 3.4. Proximal Gradients for Partial Regularization + +Eq. 5 has a similar formulation of the partial regularization of lasso [40]. In [40], the related optimization problem is solved via a nonmonotone proximal gradient (NPG) method. However, NPG requires frequent evaluation of the loss function to ensure the loss value after proximal gradients is less or equal to the loss value before the update. With CNNs, the costs of NPG are too large due to frequent loss evaluations. As a result, we use a one-step proximal gradient update to solve the problem defined in Eq. 4. + +The proximal operator of $\mathcal{R}_w$ is defined as: + +$$ +\operatorname {p r o x} _ {\gamma \mathcal {R} _ {w}} (x) = \underset {y} {\arg \min } \gamma \mathcal {R} _ {w} (y) + \frac {1}{2} \| x - y \| ^ {2}. \tag {6} +$$ + +In Eq. 6, $\| x - y\| ^2$ defines the sum of the square difference + +between $x$ and $y$ . We denote model weights after $t$ th update as $\mathcal{W}^t$ , and $\mathcal{W}^{t + 1}$ can be obtained by: + +$$ +\mathcal {W} ^ {t + 1} = \operatorname {p r o x} _ {\alpha^ {t + 1} \gamma \mathcal {R} _ {w}} (u (\mathcal {W} ^ {t}, \alpha^ {t + 1})), \tag {7} +$$ + +where $u(\mathcal{W}^t, \alpha^{t+1})$ is an update rule that can be applied for many algorithms, and $\alpha^{t+1}$ is the learning rate at the current step. Take SGD as an example, $u(\mathcal{W}^t, \alpha^{t+1}) = \mathcal{W}^t - \alpha^{t+1}\nabla_{\mathcal{W}^t}\mathcal{L}$ . With Eq. 6 and Eq. 7, model weights $\mathcal{W}$ can then be updated with proximal gradients. + +We now present the analytical solution to Eq. 6, so that model weights can be efficiently updated. From the structure of $\mathcal{R}_w$ , we know that channels with indices $i \in S_l$ will be regularized. This is equivalent to not regularizing channels if $i \notin S_l$ . Consequently, we have the following form of the proximal gradient operator: + +$$ +\operatorname {p r o x} _ {\alpha \gamma \mathcal {R} _ {w}} (\mathcal {W} _ {l}) = \left\{ \begin{array}{c} \frac {\mathcal {W} _ {l [ i , ; ; ; ]}}{\| (1 - \mathbf {v} _ {l}) \odot \mathcal {W} _ {l} \| _ {2}} \max \left(0, - \frac {\hat {N} _ {l}}{\hat {N}} \alpha \gamma \right. \\ \left. + \| (1 - \mathbf {v} _ {l}) \odot \mathcal {W} _ {l} \| _ {2}\right), \text {i f} i \in S _ {l}, \\ \mathcal {W} _ {l [ i, ; ; ; ; ]}, \text {i f} i \notin S _ {l}. \end{array} \right. \tag {8} +$$ + +In Eq. 8, we omit the step notation $t$ to simplify the notations, and we still assume $W_{l}$ is pruned along the output dimension. In this equation, it is easier to see that $\frac{\hat{N}_l}{N}$ can balance the regularization strength given different layers. Due to the property of our proposed partial regularization, the term $\| (1 - \mathbf{v}_l)\odot \mathcal{W}_l\| _2$ also changes dynamically, since $\hat{N}_l$ is changed after updates of the AGN. As a result, if no adjustment is applied, no matter how large or small $\hat{N}_l$ is, only a constant value $\frac{1}{L}\alpha \gamma$ will be used for the soft-thresholding, $\max (0,\| (1 - \mathbf{v}_l)\odot \mathcal{W}_l\| _2 - \frac{1}{L}\alpha \gamma)$ , in all circumstances, which is not reasonable. To accompany the changes of $v_{l}$ and consequently $\hat{N}_l$ , we use $\frac{\hat{N}_l}{N}$ to dynamically balance the soft-thresholding parameter between different layers. With Eq. 8, we can efficiently update $\mathcal{W}$ with our proposed partial regularization. + +# 3.5. Network Pruning via Structural Alignment + +We present the algorithm of our method in Algorithm 1. In Algorithm 1, $D$ is the training dataset; $D_{\mathrm{AGN}}$ is a sub-dataset within $D$ and it is used to train AGN; $p$ decides how much FLOPs is preserved and it described in section 3.2; $\gamma$ and $\lambda$ are hyperparameters to control the strength of $\mathcal{R}_{\mathrm{FLOPs}}$ and $\mathcal{R}_w$ ; $E$ is the total number of epochs; $E_{\mathrm{start}}$ is the start epoch to train AGN and apply $\mathcal{R}_w$ when optimizing model weights. Note that, to reduce the overhead brought by training AGN, we only use a small sub-set sampled from $D$ . As we discussed in section 1, we need to set a start epoch for AGN and $\mathcal{R}_w$ . If we apply $\mathcal{R}_w$ when training starts $(E_{\mathrm{start}} = 0)$ , it will largely restrict the final performance of the whole model before pruning. Instead, we can have deserved results if we start partial regularization in the middle of the training process. The reason for this problem can + +come from several perspectives. For example, at the beginning of the training, the classification loss can not produce accurate guidance for pruning since weights are not trained properly. Consequently, it will produce a bad sub-network, and following this sub-network only gives even worse results. + +We summarize our method in Fig. 1. The inference path when training model weights and the AGN are different, and they are connected by partial regularization, which is different from current one-shot pruning and soft pruning methods. With this design, model weights are not directly affected by the sub-network architecture, which creates a smooth transition for the selected sub-network before and after pruning. + +# 4. Experiments + +# 4.1. Settings + +We use CIFAR-10 [26] and ImageNet [5] to evaluate the performance of our method. Our method requires one hyper-parameter $p$ to control the FLOPs budget. The detailed choices of $p$ are listed in supplementary materials. We choose ResNets [14] and MobileNet-V2 [49] for comparison. For CIFAR-10, we compare our method with other methods on ResNet-56 and MobileNetV2. For ImageNet, we select ResNet-34, ResNet-50, ResNet-101, and MobileNetV2 as our target models. + +We set $\lambda$ in Eq. 3 to 4.0 for all models and datasets. Similarly, we set $\gamma$ to 0.0005 for all settings. We set $E_{\mathrm{start}}$ at $20\%$ of the total training epochs. Detailed numbers of $E_{\mathrm{start}}$ are listed in supplementary materials. The range of $E_{\mathrm{start}}$ is quite large, which is hard to be explored thoroughly. The current setting already provides good results, but better settings may also exist. To reduce the training costs of the AGN, we random sample $5\%$ of samples from the original dataset for constructing $D_{\mathrm{AGN}}$ . With this setup, the additional costs are less than $5\%$ of the original training costs. We train the parameters $\Theta$ of the AGN using ADAM [25] with a start learning rate 0.001. Besides the training of the AGN, we follow the standard training recipe of ResNets for both CIFAR-10 and ImageNet. For MobileNet-V2, We follow the training settings in their original paper [49]. After training, we prune the model by using the sub-network generated from the AGN. The fine-tuning settings are similar to the training setting. Due to space limitations, details of training and fine-tuning are also presented in supplementary materials. In the experimental section, our method is abbreviated as SANP: Structural Alignment for Network Pruning. + +# 4.2. CIFAR-10 Results + +The results of CIFAR-10 are presented in Tab. 1. For ResNet-56, our method achieves the best performance (in terms of $\Delta$ -Acc) compared to other methods. Specifically, + +Table 1: Comparison of results on CIFAR-10. $\Delta$ -Acc represents the performance changes relative to the baseline, and $+ / -$ indicates an increase/decrease, respectively. + +
ArchitectureMethodBaseline AccPruned AccΔ-AccPruned FLOPs
ResNet-56DCP-Adapt [60]93.80%93.81%+0.01%47.0%
SCP [23]93.69%93.23%-0.46%51.5%
FPGM [18]93.59%92.93%-0.66%52.6%
SFP [16]93.59%92.26%-1.33%52.6%
FPC [15]93.59%93.24%-0.25%52.9%
HRank [35]93.26%92.17%-0.09%50.0%
DMC [9]93.62%92.69%+0.07%50.0%
GNN-RL [57]93.49%93.59%+0.10%54.0%
SANP (ours)93.49%93.81%+0.32%52.0%
MobileNetV2Uniform [60]94.47%94.17%-0.30%26.0%
DCP [60]94.47%94.69%+0.22%26.0%
DMC [9]94.23%94.49%+0.26%40.0%
SCOP [53]94.48%94.24%-0.24%40.3%
SANP (ours)94.52%94.97%+0.45%46.0%
+ +![](images/a69ecfd6f65995fcb7b7a6c90c2cb5ebe89d2669a23950b0e09b6a684353793b.jpg) +(a) ResNet-56. + +![](images/44889f264491c4ad361fd83bc38ac42dff4cb89bd70997869a23378856d8663f.jpg) +(b) MoibleNet-V2. + +our method outperforms the second best method GNN-RL by $0.22\%$ regarding $\Delta$ -Acc (SANP $+0.32\%$ vs. GNN-RL $+0.10\%$ when pruning similar FLOPs (SANP $52.0\%$ vs. GNN-RL $54.0\%$ ). Our method also outperforms HRank and DMC by $0.25\%$ and $0.41\%$ separately (DMC $+0.07\%$ and HRank $-0.09\%$ ). The gap between other methods and our method is even larger. For MobileNet-V2, our method prunes most FLOPs $(46.0\%)$ and achieves the best performance ( $\Delta$ -Acc: $+0.45\%$ ). Compared to the second best method, DMC, our method prunes $6\%$ more FLOPs and outperforms it by $0.19\%$ . SCOP also prunes $40\%$ FLOPs, and the gap between our method and SCOP is even larger $(0.69\%$ better in terms of $\Delta$ -Acc). The uniform setting and DCP prunes around $26\%$ FLOPs, but the performance is still worse than our method. + +# 4.3. ImageNet Results + +All results for the ImageNet dataset are shown in Tab. 2. ResNet-34. Our method achieves $73.43\%$ Top-1 accuracy and $91.48\%$ Top-5 accuracy, which is better than the baseline by $0.19\%$ and $0.16\%$ for Top-1/Top-5 accuracy separately. At the same time, our method removes $44.1\%$ FLOPs, which is on par with other methods. It is obvious + +![](images/86694253368960dfd7c748f962eb19a7c490cb11809763b991738ab32460be87.jpg) +(a) +Figure 3: Top-1 and Top-5 accuracy after pruning given different settings. + +![](images/824a7051fd2f86530a986497fab1772f7b059443360c694b93b27c11738777ee.jpg) +Figure 2: (a) and (b): the average norm of channels with or without partial regularization for ResNet-56 and MobileNet-V2. +(b) + +that the advantage of our method is clear compared to other methods. Our method prunes similar FLOPs to SCOP and DMC. However, the $\Delta$ Top-1 Acc of our method is $0.88\%$ and $0.92\%$ better than SCOP and DMC, respectively, and we have similar observations for $\Delta$ Top-5 Acc (0.60% and $0.47\%$ better than SCOP and DMC). Taylor has the second best $\Delta$ Top-1 Acc, but the pruned FLOPs is much lower than our method (ours: $44.1\%$ vs. Taylor: $24.2\%$ ). The advantage of our method compared to FPGM is more obvious. + +ResNet-50. Our method achieves $76.47\%$ Top-1 accuracy and $93.00\%$ Top-5 accuracy, which is also better than the + +Table 2: Comparison results on ImageNet with ResNet-34/50/101 and MobileNet-V2. + +
ArchitectureMethodBaseline Top-1 AccBaseline Top-5 AccΔ Top-1 AccΔ Top-5 AccPruned FLOPs
ResNet-34FPGM [18]73.92%91.62%-1.29%-0.54%41.1%
Taylor [42]73.31%--0.48%-24.2%
DMC [9]73.30%91.42%-0.73%-0.31%43.4%
SCOP [53]73.31%91.42%-0.69%-0.44%44.8%
SANP (ours)73.24%91.32%+0.19%+0.16%44.1%
ResNet-50DCP [60]76.01%92.93%-1.06%-0.61%55.6%
CCP [44]76.15%92.87%-0.94%-0.45%54.1%
FPGM [18]76.15%92.87%-1.32%-0.55%53.5%
ABCP [36]76.01%92.96%-2.15%-1.27%54.3%
DMC [9]76.15%92.87%-0.80%-0.38%55.0%
SCOP [53]76.15%92.87%-0.89%-0.34%54.6%
PFP [34]76.13%92.86%-0.92%-0.45%44.0%
CHIP [52]76.15%92.87%+0.00%+0.04%48.7%
NPPM [8]76.15%92.87%-0.19%+0.12%56.0%
Random-Pruning [30]75.83%92.92%-0.75%-0.40%51.0%
GNN-RL [57]76.10%--1.82%-53.0%
SANP (ours)76.06%92.86%+0.41%+0.17%56.2%
ResNet-101FPGM [18]77.37%93.56%-0.05%0.00%41.1%
Taylor [42]77.37%--0.02%-39.8%
DMC [9]77.37%93.56%+0.04%+0.03%56.0%
PFP [34]77.37%93.56%-0.94%-0.44%45.1%
SANP (ours)77.53%93.71%+0.61%+0.29%55.4%
MobileNet-V2Uniform [49]71.80%91.00%-2.00%-1.40%30.0%
AMC [17]71.80%--1.00%-30.0%
CC [33]71.88%--0.97%-28.3%
MetaPruning [38]72.00%--0.80%-30.7%
Random-Pruning [30]71.88%--1.01%-29.1%
SANP (ours)71.91%90.30%+0.14%+0.07%29.1%
+ +
SettingsArchitectureBaseline Top-1 AccΔ Top-1 AccPruned FLOPs
One-Shot SANPResNet-3473.31%-0.50%44.0%
73.24%+0.19%44.1%
One-Shot SANPResNet-5076.13%-0.57%56.0%
76.06%+0.41%55.2%
One-Shot SANPMobileNetV271.88%-0.42%29.3%
71.91%+0.07%29.1%
+ +Table 3: Performance of pruned models given different pruning settings on ImageNet. + +baseline Top-1/Top-5 accuracy. The second best method, CHIP, removes $48.7\%$ FLOPs while maintaining the original performance. Our method outperforms CHIP by $0.41\%$ in terms of the $\Delta$ Top-1 accuracy while pruning near $8\%$ more FLOPs. NPPM is a strong baseline for ResNet-50, our method outperforms by $0.60\%$ in $\Delta$ -Top-1 Acc. GNN-RL and Random-Pruning are two recent pruning works. Our method outperforms them by $2.23\%$ and $1.16\%$ separately, while our method prunes more FLOPs. PHP, DCP, CCP, and DMC have similar $\Delta$ -Top 1 accuracy. The gap between our method and these methods ranges from $1.21\%$ to $1.47\%$ regarding $\Delta$ -Top 1 accuracy. The advantage of our method compared to the rest methods is more apparent. + +ResNet-101. Our method achieves $78.14\%$ Top-1 accuracy and $94.00\%$ Top-5 accuracy, which is $0.61\%$ and $0.29\%$ better than the baseline Top-1 and Top-5 accuracy. Although DMC prunes a little bit more FLOPs, it only achieves $77.41\%$ Top-1 accuracy after pruning which is $0.73\%$ lower than our method. FPGA, Taylor, and PFP remove less than + +50% FLOPs. Our method prunes at least 10% more FLOPs and still has an advantage in terms of $\Delta$ -Top 1 accuracy (from 0.66% to 1.55%). + +MobileNet-V2. MobileNet-V2 is generally harder to prune compared to ResNets. All comparison methods on MobileNet-V2 remove around $30\%$ FLOPs. Our method achieves $72.05\%$ Top-1 accuracy and $90.37\%$ Top-5 accuracy, which is $0.14\%$ and $0.07\%$ better than the baseline Top-1 and Top-5 accuracy. Given the similar pruning rate, our method is $1.15\%$ , $0.94\%$ , $1.11\%$ and $1.14\%$ higher than Random-Pruning, MetaPruning CC, and AMC separately regarding $\Delta$ -Top-1 Acc. MetaPruning prunes most FLOPs, but the performance is much lower than our method. In short, our method can also be applied to lightweight CNNs, like MobileNet-V2. + +# 4.4. Analysis of Our Method + +The effectiveness of partial regularization. To verify whether our proposed partial regularization is effective, we plot the average channel norm of different groups within each block for ResNet-56 and MobileNet-V2 on CIFAR-10. The results are shown in Fig. 2. The average channel norm for the group with partial regularization and without partial regularization are obtained by $\frac{1}{\tilde{N}_l}\sum_{i\in S_l}\| \mathcal{W}_{l[i,:;:]} \|_{\mathrm{GL}}$ and $\frac{1}{C_l - \tilde{N}_l}\sum_{i\notin S_l}\| \mathcal{W}_{l[i,:;:]} \|_{\mathrm{GL}}$ separately. Weights with partial regularization are effectively aligned. On the contrary, weights without partial regularization are lightly affected, which justifies the strength of our proposed partial regular + +![](images/281bc250104f5325c56a506ac5bfa79fdea6ceacdbf80cfd963b2b98953b7825.jpg) + +![](images/e29c52c97e5da8fdfcdc649c285d170d024999e8b8f3b8dcdb660622505e323c.jpg) + +![](images/6a4ccd9bd6001d512932246f5ed7b4192731416cd37dfbf2f1c1c7effb7c2ead.jpg) + +![](images/03ece7e4f2f553b57c01f487e585c92c27a064a44bc23463885b5fa8a8edc4da.jpg) + +![](images/e4addfc2a4679d8270777b550b2b9a49a82b5adb6e486c44e85482612dc02af7.jpg) +(a) +(e) +Figure 4: (a, e): the impact of $\lambda$ in $\mathcal{R}_{\mathrm{FLOPs}}$ . (b, f): the effect of the architecture of AGN. (c, g): the effect of $E_{\mathrm{start}}$ . (d, h): the effect of different setups. Experiments are conducted on CIFAR-10 with ResNet-56 and $p = 0.5$ (a,b,c,d,e) and $p = 0.35$ (f,g,h). + +![](images/c0ddbd8370737cf69b68333a1cde4cd37120b0cc67bad76bf7a3cb729dc146f8.jpg) +(b) +(f) + +![](images/6a8e34b812542773adfbf228da27dc0c06c62eea1146a08ef07c23af33365447.jpg) +(c) +(g) + +![](images/67bd6b1df5e318a5e01a1c01a1ed1147e8a891764e311eb5c0c47dbe41c7418f.jpg) +(d) +(h) + +# ization. + +The impact of $\lambda$ . We study the impact of $\lambda$ when training the AGN, and we plot the test accuracy and $\mathcal{R}_{\mathrm{FLOPs}}$ in Fig. 4a and Fig. 4e. From the figures, we can see that if $\lambda$ is too large, it will have negative impacts on learned subnetworks. Otherwise, our method is robust to $\lambda$ , + +The effect of AGN. In Fig. 4b and Fig. 4f, we plot the test accuracy when learning the AGN given different pruning rates. We construct a Simple baseline, which parameterizes each channel by using one learnable parameter. We can see that when the parameterization space shrinks, the performance of learned sub-networks will be affected severely. In addition, the learning is slower, and the best sub-network performance is also much worse than using AGN. + +The effect of $\mathbf{E}_{\mathrm{start}}$ . In the method section, we argue that inserting partial regularization in the middle training process is beneficial. We plot the results of $E_{\mathrm{start}} = 0$ and $E_{\mathrm{start}} = 40$ in Fig. 4c and Fig. 4g. In general, they can find sub-networks with similar performance, but $E_{\mathrm{start}} = 0$ is worse than $E_{\mathrm{start}} = 40$ when the pruning rate is large ( $p = 0.35$ ). More importantly, the full model accuracy of $E_{\mathrm{start}} = 0$ is $92.96\%$ (average across different runs), which is around $0.50\%$ worse than $E_{\mathrm{start}} = 40$ , which suggests that using partial regularization at the beginning will limit the capacity of the full model. + +The effect of different setups. We construct two additional baselines to see how partial regularization helps to produce a better sub-network within the full model. The One-Shot baseline directly trains the AGN on the pre-trained model. The w/o Partial Regularization baseline set $\gamma = 0$ , and the rest settings are the same as our method. The related + +results are shown in Fig. 4d and Fig. 4h. From these figures, we can see that partial regularization always produces the best sub-network from the full model. In addition, 'w/o partial regularization' is worse than the one-shot setting. This is probably because it is hard to capture the changes of model weights without partial regularization, which makes the training of AGN much harder than the rest settings. More comparisons on the ImageNet dataset. To further show how our method improves the one-shot setting, we present the sub-network performance before and after fine-tuning for one-shot and partial regularization settings in Fig. 3 and Tab. 3. From Fig. 3, we can see that partial regularization still produces better sub-networks on ImageNet, and it is around $10\%$ better in terms of Top-1 accuracy than the one-shot setting for different models. The advantage of partial regularization naturally extends to results after fine-tuning, as shown in Tab. 3. + +# 5. Conclusion + +In this paper, we investigate how partial regularization helps to produce a better sub-network for network pruning. Specifically, our method uses AGN to guide partial regularization across the training process. We further provide an efficient way to update model weights through proximal gradients. With these designs, partial regularization effectively reduces the gap between the sub-network within the full model and the pruned model. Our method then starts from a better sub-network, thus resulting in a better final pruned model. Extensive experimental results on CIFAR-10 and ImageNet show the effectiveness of our method. + +# References + +[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3 +[2] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016. 1 +[3] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 3 +[4] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in neural information processing systems, pages 3123-3131, 2015. 3 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 248-255. IEEE, 2009. 5 +[6] Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, and H.T. Kung. Neural mean discrepancy for efficient out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19217-19227, June 2022. 1 +[7] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2019. 2 +[8] Shangqian Gao, Feihu Huang, Weidong Cai, and Heng Huang. Network pruning via performance maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9270-9280, 2021. 1, 7 +[9] Shangqian Gao, Feihu Huang, Jian Pei, and Heng Huang. Discrete model compression with resource constraint for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1899-1908, 2020. 1, 2, 6, 7 +[10] Shangqian Gao, Burak Uzkent, Yilin Shen, Heng Huang, and Hongxia Jin. Learning to jointly share and prune weights for grounding based vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3 +[11] Junfeng Guo, Yiming Li, Xun Chen, Hanqing Guo, Lichao Sun, and Cong Liu. SCALE-UP: An efficient black-box input-level backdoor detection via analyzing scaled prediction consistency. In The Eleventh International Conference on Learning Representations, 2023. 1 +[12] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1 + +[13] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Advances in neural information processing systems, pages 1135-1143, 2015. 1, 2 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 5 +[15] Yang He, Yuhang Ding, Ping Liu, Linchao Zhu, Hanwang Zhang, and Yi Yang. Learning filter pruning criteria for deep convolutional neural networks acceleration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2009-2018, 2020. 6 +[16] Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating deep convolutional neural networks. In International Joint Conference on Artificial Intelligence (IJCAI), pages 2234-2240, 2018. 1, 6 +[17] Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, and Song Han. Amc: Automl for model compression and acceleration on mobile devices. In Proceedings of the European Conference on Computer Vision (ECCV), pages 784-800, 2018. 1, 2, 7 +[18] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2019. 6, 7 +[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[20] Zehao Huang and Naiyan Wang. Data-driven sparse structure selection for deep neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 304-320, 2018. 2 +[21] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32Nd International Conference on International Conference on Machine Learning - Volume 37, ICML, pages 448-456. JMLR.org, 2015. 2 +[22] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 3 +[23] Minsoo Kang and Bohyung Han. Operation-aware soft channel pruning using differentiable masks. In International Conference on Machine Learning, pages 5122-5131. PMLR, 2020. 1, 2, 6 +[24] Jaedeok Kim, Chiyoun Park, Hyun-Joo Jung, and Yoonsuck Choe. Plug-in, trainable gate for streamlining arbitrary neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 2 +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5 +[26] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, Cite-seer, 2009. 5 + +[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097-1105, 2012. 1 +[28] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. *ICLR*, 2019. 2 +[29] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. *ICLR*, 2017. 2 +[30] Yawei Li, Kamil Adamczewski, Wen Li, Shuhang Gu, Radu Timofte, and Luc Van Gool. Revisiting random channel pruning for neural network compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 191-201, 2022. 7 +[31] Yawei Li, Shuhang Gu, Christoph Mayer, Luc Van Gool, and Radu Timofte. Group sparsity: The hinge between filter pruning and decomposition for network compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8018-8027, 2020. 3 +[32] Yawei Li, Shuhang Gu, Kai Zhang, Luc Van Gool, and Radu Timofte. Dhp: Differentiable meta pruning via hypernetworks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 608-624. Springer, 2020. 2 +[33] Yuchao Li, Shaohui Lin, Jianzhuang Liu, Qixiang Ye, Mengdi Wang, Fei Chao, Fan Yang, Jincheng Ma, Qi Tian, and Rongrong Ji. Towards compact cnns via collaborative compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6438-6447, 2021. 7 +[34] Lucas Liebenwein, Cenk Baykal, Harry Lang, Dan Feldman, and Daniela Rus. Provable filter pruning for efficient neural networks. In International Conference on Learning Representations, 2020. 7 +[35] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. Hrank: Filter pruning using high-rank feature map. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6 +[36] Mingbao Lin, Rongrong Ji, Yuxin Zhang, Baochang Zhang, Yongjian Wu, and Yonghong Tian. Channel pruning via automatic structure search. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages 673 - 679, 2020. 7 +[37] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 2 +[38] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In Proceedings of the IEEE International Conference on Computer Vision, pages 3296-3305, 2019. 2, 7 +[39] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. In International Conference on Learning Representations, 2019. 2 + +[40] Zhaosong Lu and Xiaorui Li. Sparse recovery via partial regularization: Models, theory, and algorithms. Mathematics of Operations Research, 43(4):1290-1316, 2018. 2, 4 +[41] Dmitry Molchanov, Arsenii Ashukha, and Dmitry Vetrov. Variational dropout sparsifies deep neural networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 2498-2507. JMLR.org, 2017. 3 +[42] Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11264–11272, 2019. 1, 7 +[43] Kirill Neklyudov, Dmitry Molchanov, Arsenii Ashukha, and Dmitry P Vetrov. Structured bayesian pruning via log-normal multiplicative noise. In Advances in Neural Information Processing Systems, pages 6775-6784, 2017. 3 +[44] Hanyu Peng, Jiaxiang Wu, Shifeng Chen, and Junzhou Huang. Collaborative channel pruning for deep networks. In International Conference on Machine Learning, pages 5113-5122, 2019. 1, 2, 7 +[45] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In European Conference on Computer Vision, pages 525-542. Springer, 2016. 3 +[46] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1 +[47] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91-99, 2015. 1 +[48] Alex Renda, Jonathan Frankle, and Michael Carbin. Comparing rewinding and fine-tuning in neural network pruning. In International Conference on Learning Representations, 2020. 2 +[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 2, 5, 7 +[50] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in neural information processing systems, pages 568-576, 2014. 1 +[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1 +[52] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. Chip: Channel independence-based pruning for compact neural networks. Advances in Neural Information Processing Systems, 34, 2021. 7 +[53] Yehui Tang, Yunhe Wang, Yixing Xu, Dacheng Tao, Chunjing Xu, Chao Xu, and Chang Xu. Scop: Scientific control for reliable neural network pruning. Advances in Neural Information Processing Systems, 33, 2020. 6, 7 +[54] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. + +In Advances in neural information processing systems, pages 2074-2082, 2016.3 +[55] Mao Ye, Chengyue Gong, Lizhen Nie, Denny Zhou, Adam Klivans, and Qiang Liu. Good subnetworks provably exist: Pruning via greedy forward selection. In International Conference on Machine Learning, pages 10820-10830. PMLR, 2020. 2 +[56] Zhonghui You, Kun Yan, Jinmian Ye, Meng Ma, and Ping Wang. Gate decorator: Global filter pruning method for accelerating deep convolutional neural networks. In Advances in Neural Information Processing Systems, pages 2130-2141, 2019. 1, 2 +[57] Sixing Yu, Arya Mazaheri, and Ali Jannesari. Topology-aware network pruning using multi-stage graph embedding and reinforcement learning. In International Conference on Machine Learning, pages 25656-25667. PMLR, 2022. 6, 7 +[58] Dejiao Zhang, Haozhu Wang, Mario Figueiredo, and Laura Balzano. Learning to share: Simultaneous parameter tying and sparsification in deep learning. 2018. 3 +[59] Yanfu Zhang, Shangqian Gao, and Heng Huang. Exploration and estimation for model compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 487-496, 2021. 2 +[60] Zhuangwei Zhuang, Mingkui Tan, Bohan Zhuang, Jing Liu, Yong Guo, Qingyao Wu, Junzhou Huang, and Jinhui Zhu. Discrimination-aware channel pruning for deep neural networks. In Advances in Neural Information Processing Systems, pages 875-886, 2018. 6, 7 \ No newline at end of file diff --git a/2023/Structural Alignment for Network Pruning through Partial Regularization/images.zip b/2023/Structural Alignment for Network Pruning through Partial Regularization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9e4b84c5a1e18b331a1a2fe7dccc867414f4ee97 --- /dev/null +++ b/2023/Structural Alignment for Network Pruning through Partial Regularization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83f3d070adfd8d3afb2ad852fda63001159d6a0403ee69ceb6d5a934cfe5debd +size 504474 diff --git a/2023/Structural Alignment for Network Pruning through Partial Regularization/layout.json b/2023/Structural Alignment for Network Pruning through Partial Regularization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9c90c3c07150f769dca5c72c1c42c2188971fbb4 --- /dev/null +++ b/2023/Structural Alignment for Network Pruning through Partial Regularization/layout.json @@ -0,0 +1,11445 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 527, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 527, + 121 + ], + "type": "text", + "content": "Structural Alignment for Network Pruning through Partial Regularization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "spans": [ + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "text", + "content": "Shangqian Gao" + }, + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "text", + "content": ", Zeyu Zhang" + }, + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "text", + "content": ", Yanfu Zhang" + }, + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "text", + "content": ", Feihu Huang" + }, + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "text", + "content": ", and Heng Huang" + }, + { + "bbox": [ + 100, + 152, + 492, + 168 + ], + "type": "inline_equation", + "content": "^{*3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 176, + 485, + 219 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 107, + 176, + 485, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 176, + 485, + 205 + ], + "spans": [ + { + "bbox": [ + 107, + 176, + 485, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 176, + 485, + 205 + ], + "type": "text", + "content": "Department of Electrical and Computer Engineering, University of Pittsburgh " + }, + { + "bbox": [ + 107, + 176, + 485, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 176, + 485, + 205 + ], + "type": "text", + "content": "School of Information, University of Arizona" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 205, + 477, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 205, + 477, + 219 + ], + "spans": [ + { + "bbox": [ + 110, + 205, + 477, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 110, + 205, + 477, + 219 + ], + "type": "text", + "content": "Department of Computer Science, University of Maryland at College Park" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 85, + 219, + 509, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 219, + 509, + 240 + ], + "spans": [ + { + "bbox": [ + 85, + 219, + 509, + 240 + ], + "type": "text", + "content": "{shg84, yaz91}@pitt.edu, zeyuzhang@email.arizona.edu, huangfeihu2018@gmail.com, heng@umd.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 275, + 192, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 275, + 192, + 289 + ], + "spans": [ + { + "bbox": [ + 143, + 275, + 192, + 289 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 301, + 290, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 290, + 563 + ], + "type": "text", + "content": "In this paper, we propose a novel channel pruning method to reduce the computational and storage costs of Convolutional Neural Networks (CNNs). Many existing one-shot pruning methods directly remove redundant structures, which brings a huge gap between the model before and after network pruning. This gap will no doubt result in performance loss for network pruning. To mitigate this gap, we first learn a target sub-network during the model training process, and then we use this sub-network to guide the learning of model weights through partial regularization. The target sub-network is learned and produced by using an architecture generator, and it can be optimized efficiently. In addition, we also derive the proximal gradient for our proposed partial regularization to facilitate the structural alignment process. With these designs, the gap between the pruned model and the sub-network is reduced, thus improving the pruning performance. Empirical results also suggest that the sub-network found by our method has a much higher performance than the one-shot pruning setting. Extensive experiments show that our method can achieve state-of-the-art performances on CIFAR-10 and ImageNet with ResNets and MobileNet-V2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "content": "Convolutional Neural Networks (CNNs) have achieved many successes in different computer vision tasks [27, 46, 47, 50, 2, 6, 11]. To tackle real-world challenges, recent CNNs [27, 51, 14] become larger and larger regarding width, depth, etc. With such capacities, CNNs can obtain better performance on different benchmarks at the cost" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "content": "*Corresponding author. This work was partially supported by NSF IIS 1838627, 1837956, 1956002, 2211492, CNS 2213701, CCF 2217003, DBI 2225775." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 277, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 373 + ], + "type": "text", + "content": "of computational and storage burdens. At the same time, with the recent developments of mobile and embedded devices, the demand for deploying CNNs on these devices has increased dramatically. However, there is a natural conflict between the size of CNNs and the hardware capability of these devices. To overcome these challenges, many works [13, 12] try to reduce the size of CNNs, and make them possible to be deployed on edge devices." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 375, + 546, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 546, + 602 + ], + "type": "text", + "content": "There are many directions to reduce the size of CNNs. Among them, weight pruning and structural pruning are two popular topics. Structural pruning, especially channel pruning, is more friendly to hardware than weight pruning since no post-processing steps are required to acquire savings in computational and storage costs. Thus, our paper focuses on channel pruning for CNNs. Many existing one-shot pruning works [44, 56, 9, 42, 17, 8] prune trained models directly. No matter what method is used, there will be a significant gap between the selected sub-network and the pruned model. Such a gap creates difficulties in regaining performance during the fine-tuning process. On the other hand, soft pruning methods [16, 23] softly remove structures during the training process, which can produce good results with a shorter fine-tuning process. However, soft pruning methods generally perform worse than typical one-shot pruning methods, probably because the weight space is restricted during the training process because of soft pruning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "To tackle the above problems, we introduce a novel partial regularization technique to align model weights and the discovered sub-network during the training process, which can produce a high performance sub-network and reduce the gap between the sub-network and the original model. In addition, unlike soft-pruning methods, all model structures are used for training. The partial regularization term contains a partial group lasso regularization on selected weights, and other weights remain intact without modifica" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "17402" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 395 + ], + "type": "text", + "content": "tions. An architecture generator is trained to select which weights should be aligned, and it is also updated during the training process. Our partial regularization formulation is related to partial regularization in lasso [40]. Inspired by the nonmonotone proximal gradient (NPG) method used in [40], we also use a proximal gradient method to solve the partial regularization problem in our setting. Note that our method dynamically changes which channels should be put in the partial regularization. As a result, we add a scalar to balance the regularization strength for different layers because the number of pruned channels is different for them. To maximally keep the capacity of the original model, we insert the partial regularization in the middle of the training process. This is because weights are vulnerable to pruning at the early training stage, and the FLOPs regularization will dominate updates of the architecture generator, which can create bad sub-networks and mislead the training process. Finally, we update model weights and the architecture generator periodically, and they are connected by the partial regularization term during training. To maintain similar training efficiency as the original model, we only use a small portion of samples to train the architecture generator. Thus, there is only a small overhead compared to the original training process. Our method successfully finds performant sub-networks from the original model with these techniques. In summary, the contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 401, + 287, + 624 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 54, + 401, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 401, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 401, + 287, + 473 + ], + "type": "text", + "content": "1) We propose to align the sub-network in the original model with the final pruned model through partial regularization. By structural alignment, the gap between the selected sub-network and the pruned model is largely reduced, which naturally improves the performance of the pruned model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 477, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 287, + 548 + ], + "type": "text", + "content": "2) We use an architecture generator parameterized by neural networks to select the proper sub-network structure and guide the partial regularization. Inspired by partial regularization in lasso [40], we propose to solve the partial regularization problem via proximal gradients, which facilitate the alignment process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 552, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 552, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 552, + 287, + 624 + ], + "type": "text", + "content": "3) Empirical results show that the sub-network discovered by our method performs much better than the one-shot pruning setting. Extensive experiments on CIFAR-10 and ImageNet show that our method outperforms existing channel pruning methods on ResNets and MobileNet-V2." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 634, + 139, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 139, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 139, + 647 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 715 + ], + "type": "text", + "content": "Weight-Level Pruning. Weight pruning removes redundant connections based on individual weights. High compression rates can be achieved by weight pruning, but they can not directly achieve acceleration, and specially designed sparse matrix libraries are required. One of the early" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "content": "works [13] proposes to measure the importance of weights with their " + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "content": " magnitude, and unimportant weights are removed. Instead of magnitude, SNIP [28] updates the importance of each weight by using gradients from the loss function. SNIP can be used at the initialization time. The assumption of the Lottery ticket hypothesis [7] suggests that there exist sparse sub-networks (winning tickets) that can achieve the performance of the full model. In addition, with repeated training and fine-tuning, it can achieve better results. On the other hand, rethinking network pruning [39] argues that the learned topology from pruning algorithms is the key to achieving better performance. In addition, weight rewinding [48] shows that resetting weights to values from the middle training process can also produce good results. Similar to weight rewinding, our method does not modify weight training at the beginning, the partial regularization is inserted in the middle training process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 282, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 547, + 715 + ], + "type": "text", + "content": "Structural-Level Pruning. Different from weight pruning, structural pruning is more friendly to hardware since it requires little or no post-processing steps to achieve acceleration. Similar to weight pruning, one of the early structural pruning methods [29] measure the importance of filters by using the sum of the absolute value of kernel weights. Besides using the magnitude to measure channel or filter importance, other methods utilize the scaling factor of batchnorm [21] to indicate which channels are important because batchnorm [21] is popular for the design of recent CNNs [14, 49]. To prune channels, Liu et al. [37] apply the sparse regularization to the scaling factors of batchnorm, and the channel will be pruned if the corresponding scaling factor is small. Structure sparse selection [20] introduces scaling factor to specific structures, such as neurons, groups, or residual blocks, and the sparsity regularization is applied to these structures. Structures with small values will also be removed. Another line of research formulates channel pruning as a constrained optimization problem [24, 56, 9, 23, 59], and learnable parameters are used to control each channel. These parameters are end-to-end differentiable, which is amenable to gradient based optimization methods. Our method is also related to these researches. Different from these methods, the learning of sub-networks accomplishes the training of model weights in our method. In addition, we use partial regularization to promote the selected sub-network, which reduces the gap between the pruned model and the sub-network. Besides using gates, Automatic Model Compression (AMC) [17] uses policy gradient to update the policy network. This policy network is then used to decide the left width of each layer. Collaborative channel pruning [44] prunes channels by exploiting inter-channel dependency. Greedy forward selection [55] starts from an empty network and greedily adds Important channels from the full model. MetaPruning [38, 32] uses a hypernet to generate parameters for sub" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17403" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 70, + 495, + 281 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 495, + 281 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 495, + 281 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 495, + 281 + ], + "type": "image", + "image_path": "6df72f0c9c5234eb531cda79203f02f34d04ab942b85e39f4903a5e347d60045.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 547, + 329 + ], + "type": "text", + "content": "Figure 1: Overview of the proposed method. The AGN generates the sub-network to guide the partial regularization during the training process of model weights. The AGN is trained by evaluating sub-networks on a sub-set from the whole dataset. Both model weights and AGN are trained in an end-to-end differentiable manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 350, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 386 + ], + "type": "text", + "content": "networks, and evolutionary algorithms are utilized to find the best sub-networks. MetaPruning shows that pruning should accomplish trained model weights." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 386, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 481 + ], + "type": "text", + "content": "Besides these pruning methods, regularization based pruning methods can also be applied to structural pruning. Previous works [54, 31] use group sparsity to prune different structures. In addition to structural sparsity, structural sharing is considered in other works [58, 10]. Our method relates to regularization based methods; however, the formulation of our partial regularization only aligns selected channels dynamically, and other weights are intact." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 482, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 531 + ], + "type": "text", + "content": "Other Related Works. Besides the above-mentioned methods, there are works from other perspectives, including bayesian pruning [41, 43], weight quantization [4, 45], and knowledge distillation [19]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 542, + 153, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 153, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 153, + 556 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 563, + 115, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 115, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 115, + 574 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": "Before introducing our method, we first describe notations and provide an overview of our method. In a CNN, the feature map of " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": "th layer can be represented by " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_l \\in \\Re^{C_l \\times W_l \\times H_l}" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "l = 1, \\ldots, L" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": " is the number of channels, " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "H_l" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "W_l" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": " are height and width of the current feature map, " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": " is the number of layers. Similarly, the weights of " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": "th layer can be written as " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_l \\in \\Re^{C_l \\times C_{l-1} \\times k_l \\times k_l}" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "inline_equation", + "content": "k_l" + }, + { + "bbox": [ + 46, + 581, + 287, + 689 + ], + "type": "text", + "content": " is the kernel size of this layer. The mini-batch dimension of feature maps is ignored to simplify notations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "The core motivation of our proposed method is to reduce the gap between the selected sub-network from the origi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 350, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 458 + ], + "type": "text", + "content": "nal model and the pruned model. To achieve this goal, we need two processes. First, we need to choose the desired sub-network from the original model, and this sub-network should also be updated when model weights are trained during the learning process. Secondly, we use the sub-network to guide the partial regularization term, which is used to decide which weights should be regularized. In addition, partial regularization should not be fixed to accommodate the changes in the selected sub-network." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 464, + 453, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 453, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 453, + 477 + ], + "type": "text", + "content": "3.2. Learning the Sub-network" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": "We use an Architecture Generator Network (AGN) to generate the desired sub-network architecture " + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\{0,1\\}^N" + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": ", where 0 or 1 is used to depict the removal or keep of a channel, and " + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": " is the total number of channels from all layers. The large parameterization space of AGN can facilitate the learning of sub-network structures. To generate " + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": ", the following equation is used:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 389, + 575, + 545, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 575, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 389, + 575, + 545, + 588 + ], + "type": "interline_equation", + "content": "\\mathbf {v} = \\operatorname {A G N} (\\Theta), \\tag {1}", + "image_path": "9ef0b2f85b7b2740b7bfb9762a45a765ef259aab1f515f349afb48c0d1c63163.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 597, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 656 + ], + "type": "text", + "content": "and AGN is composed of gated recurrent unit (GRU) [3] and dense layers. In addition, Gumbel-Sigmoid [22] with STE [1] are used to produce the final binary vector " + }, + { + "bbox": [ + 304, + 597, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 597, + 545, + 656 + ], + "type": "text", + "content": ", and they are placed after the output of dense layers. More details of AGN are provided in the supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 657, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 657, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 545, + 691 + ], + "type": "text", + "content": "Once we have " + }, + { + "bbox": [ + 304, + 657, + 545, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 657, + 545, + 691 + ], + "type": "text", + "content": ", we can apply it to the feature maps to produce a sub-network. The feature map of the " + }, + { + "bbox": [ + 304, + 657, + 545, + 691 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 657, + 545, + 691 + ], + "type": "text", + "content": "th layer is then modified as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 394, + 699, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 699, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 394, + 699, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {F}} _ {l} = \\mathbf {v} _ {l} \\odot \\mathcal {F} _ {l}, \\tag {2}", + "image_path": "f21d8756d0f0df4344451d8074d2b02fa53e8d6a5414d0d0c58e38610cbddf4e.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17404" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": " is element-wise multiplication, " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_l" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": " is the architecture vector of " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "th layer, and " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_l" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": " is resized to have the same size of " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_l" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": ". The feature map " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_l" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": " is from the output of the activation function. The overall loss function for generating the desired sub-network is as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 139, + 286, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 139, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 286, + 166 + ], + "type": "interline_equation", + "content": "\\min _ {\\Theta} \\mathcal {I} _ {\\theta} (\\Theta) := \\mathcal {L} \\left(f (x; \\mathcal {W}, \\mathbf {v}), y\\right) + \\lambda \\mathcal {R} _ {\\text {F L O P s}} (T (\\mathbf {v}), p T _ {\\text {t o t a l}}) \\tag {3}", + "image_path": "e80f2a63185d3ef58b7b0296c14346d15cb4972dfcaa78703378d3f8e8739bdc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "T(\\mathbf{v})" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": " is the current FLOPs decided by the vector " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{total}}" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": " is the total FLOPs of the original model, " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "p \\in (0,1]" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": " is a hyperparameter deciding the remaining fraction of FLOPs, " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": " is the hyper-parameter controlling the strength of FLOPs regularization, " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "f(x; \\mathcal{W}, \\mathbf{v})" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": " is a CNN parameterized by " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": " and the sub-network structure is determined by the architecture vector " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": " is the task loss function and " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{FLOPs}}" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": " is the regularization term for FLOPs. The regularization term " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{FLOPs}}" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{FLOPs}}(x,y) = \\log (\\max (x,y) / y)" + }, + { + "bbox": [ + 46, + 167, + 288, + 299 + ], + "type": "text", + "content": ". With Eq. 3, we can find promising sub-networks when training the AGN and model weights periodically." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 306, + 174, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 174, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 174, + 319 + ], + "type": "text", + "content": "3.3. Partial Regularization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 324, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 383 + ], + "type": "text", + "content": "Given a sub-network " + }, + { + "bbox": [ + 46, + 324, + 287, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 324, + 287, + 383 + ], + "type": "text", + "content": " obtained from AGN, we can then reduce the gap between the sub-network and the pruned model and thus increase its performance. We can formulate the optimization problem with partial regularization as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 393, + 287, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 393, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 75, + 393, + 287, + 410 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {W}} \\mathcal {J} _ {w} (\\mathcal {W}) := \\mathcal {L} (f (x; \\mathcal {W}), y) + \\gamma \\mathcal {R} _ {w} (\\mathcal {W}), \\tag {4}", + "image_path": "8af89aada4eb51e7683354533ac67cc74e77e3c7b335c8145b6a14354916a229.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 419, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 419, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_w" + }, + { + "bbox": [ + 46, + 419, + 287, + 455 + ], + "type": "text", + "content": " is the partial regularization term, and " + }, + { + "bbox": [ + 46, + 419, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 419, + 287, + 455 + ], + "type": "text", + "content": " controls the strength of the partial regularization. " + }, + { + "bbox": [ + 46, + 419, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_w" + }, + { + "bbox": [ + 46, + 419, + 287, + 455 + ], + "type": "text", + "content": " has the following form:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 464, + 287, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 464, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 91, + 464, + 287, + 497 + ], + "type": "interline_equation", + "content": "R _ {w} (\\mathcal {W}) = \\sum_ {l = 1} ^ {L} \\sum_ {i \\in S _ {l}} \\frac {\\hat {N} _ {l}}{\\hat {N}} \\| \\mathcal {W} _ {[ i,:,:,: ]} \\| _ {\\mathrm {G L}}, \\tag {5}", + "image_path": "39e90c1d15a2a758d77277614491778fc0fb62cf5961796275567ed631006322.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{N}_l = \\sum 1 - \\mathbf{v}_l" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{N} = \\sum 1 - \\mathbf{v}" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "S_{l} = \\{i\\mid \\mathrm{if}\\mathbf{v}_{l[i]} = 0\\}" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": " contains the indices of pruned channels which are decided by AGN. " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\frac{\\hat{N}_l}{\\hat{N}}" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": " is a scalar to adjust the regularization strength given different layers. The numerator " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{N}_l" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": " is the number of pruned channels of the " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": "th layer, and the denominator " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{N}" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": " is the number of pruned channels from all layers. It is easy to see that " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\sum_{l=1}^{L} \\frac{\\hat{N}_l}{\\hat{N}} = 1" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\|x\\|_{\\mathrm{GL}}" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": " is the norm of grouped weights for Group Lasso, and " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\|x\\|_{\\mathrm{GL}} = \\sqrt{\\sum_{i=1}^{|x|} x_i^2}" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "|x|" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": " represents the number of elements in " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": ". In Eq. 5, we assume the corresponding layer is pruned across the output dimension. If it is pruned across the input dimension, we have " + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\sum_{i \\in S_l} \\|W_{l[:i,::]}\\|_{\\mathrm{GL}}" + }, + { + "bbox": [ + 46, + 506, + 287, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": "The goal of using Group Lasso for our partial regularization is to reduce the distance between the dense model and the pruned model, other suitable functions may also be useful, but we found that the partial regularization with Group" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 318, + 101, + 531, + 393 + ], + "blocks": [ + { + "bbox": [ + 311, + 75, + 518, + 99 + ], + "lines": [ + { + "bbox": [ + 311, + 75, + 518, + 99 + ], + "spans": [ + { + "bbox": [ + 311, + 75, + 518, + 99 + ], + "type": "text", + "content": "Algorithm 1: Structural Alignment for Network Pruning" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "lines": [ + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "spans": [ + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "D,D_{\\mathrm{AGN}},p,\\lambda ,\\gamma ,E,E_{\\mathrm{start}}" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " \nInitialization: initialize " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "e:=1" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " do \n/\\* Optimizing model weights. Freeze " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " of the AGN. /\\* for a mini-batch " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " do 1. calculate the gradients w.r.t model weights: " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathcal{W}}\\mathcal{L}" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " 2. update model weights using any stochastic optimizer. 3.if if e " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "\\geq E_{\\mathrm{start}}" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " then generate v from the AGN by using Eq. 1. apply the proximal gradient step following Eq.8. \nend \n/\\* Optimizing " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " of the AGN. Freeze model weights W. /\\* if if e " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "\\geq E_{\\mathrm{start}}" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " then for a mini-batch " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{AGN}}" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " do 1. generate v from the AGN by using Eq. 1 and apply it to the model. 2. calculate gradients w.r.t to " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " in Eq. 3: " + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "inline_equation", + "content": "\\nabla_{\\Theta}\\mathcal{J}_{\\theta}" + }, + { + "bbox": [ + 318, + 101, + 531, + 393 + ], + "type": "text", + "content": " 3.update the AGN with ADAM end \nend \nPruning the model with resulting v, and fine-tuning it." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": "Lasso already produces good results. Another benefit of the partial regularization " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_w" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": " is that it will not penalize weights that are not pruned. By doing so, we can avoid the problem of overpenalizing all weights' magnitude. In addition, " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": " is updated during the training process, and " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_w" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": " will dynamically regularize weights. As a result, " + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 418, + 545, + 514 + ], + "type": "text", + "content": " can flexibly change instead of falling into a fixed sub-network during the optimization process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 521, + 544, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 544, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 544, + 533 + ], + "type": "text", + "content": "3.4. Proximal Gradients for Partial Regularization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 539, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 647 + ], + "type": "text", + "content": "Eq. 5 has a similar formulation of the partial regularization of lasso [40]. In [40], the related optimization problem is solved via a nonmonotone proximal gradient (NPG) method. However, NPG requires frequent evaluation of the loss function to ensure the loss value after proximal gradients is less or equal to the loss value before the update. With CNNs, the costs of NPG are too large due to frequent loss evaluations. As a result, we use a one-step proximal gradient update to solve the problem defined in Eq. 4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 647, + 493, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 493, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 493, + 658 + ], + "type": "text", + "content": "The proximal operator of " + }, + { + "bbox": [ + 317, + 647, + 493, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_w" + }, + { + "bbox": [ + 317, + 647, + 493, + 658 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 323, + 666, + 545, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 666, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 323, + 666, + 545, + 692 + ], + "type": "interline_equation", + "content": "\\operatorname {p r o x} _ {\\gamma \\mathcal {R} _ {w}} (x) = \\underset {y} {\\arg \\min } \\gamma \\mathcal {R} _ {w} (y) + \\frac {1}{2} \\| x - y \\| ^ {2}. \\tag {6}", + "image_path": "dd52b1dc1946bdec68cd0635e65705bd810b7fab137ee7ba258302f207d0bfbf.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "type": "text", + "content": "In Eq. 6, " + }, + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\| x - y\\| ^2" + }, + { + "bbox": [ + 306, + 700, + 545, + 714 + ], + "type": "text", + "content": " defines the sum of the square difference" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17405" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "between " + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": ". We denote model weights after " + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "th update as " + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^t" + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^{t + 1}" + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": " can be obtained by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 102, + 287, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 102, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 89, + 102, + 287, + 118 + ], + "type": "interline_equation", + "content": "\\mathcal {W} ^ {t + 1} = \\operatorname {p r o x} _ {\\alpha^ {t + 1} \\gamma \\mathcal {R} _ {w}} (u (\\mathcal {W} ^ {t}, \\alpha^ {t + 1})), \\tag {7}", + "image_path": "0786d599f8778e4e0d12da9aa6a8c12aed3cc57b4604e0f484a6159714ae171f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "type": "inline_equation", + "content": "u(\\mathcal{W}^t, \\alpha^{t+1})" + }, + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "type": "text", + "content": " is an update rule that can be applied for many algorithms, and " + }, + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "type": "inline_equation", + "content": "\\alpha^{t+1}" + }, + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "type": "text", + "content": " is the learning rate at the current step. Take SGD as an example, " + }, + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "type": "inline_equation", + "content": "u(\\mathcal{W}^t, \\alpha^{t+1}) = \\mathcal{W}^t - \\alpha^{t+1}\\nabla_{\\mathcal{W}^t}\\mathcal{L}" + }, + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "type": "text", + "content": ". With Eq. 6 and Eq. 7, model weights " + }, + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 47, + 124, + 287, + 184 + ], + "type": "text", + "content": " can then be updated with proximal gradients." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 185, + 287, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 257 + ], + "type": "text", + "content": "We now present the analytical solution to Eq. 6, so that model weights can be efficiently updated. From the structure of " + }, + { + "bbox": [ + 46, + 185, + 287, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_w" + }, + { + "bbox": [ + 46, + 185, + 287, + 257 + ], + "type": "text", + "content": ", we know that channels with indices " + }, + { + "bbox": [ + 46, + 185, + 287, + 257 + ], + "type": "inline_equation", + "content": "i \\in S_l" + }, + { + "bbox": [ + 46, + 185, + 287, + 257 + ], + "type": "text", + "content": " will be regularized. This is equivalent to not regularizing channels if " + }, + { + "bbox": [ + 46, + 185, + 287, + 257 + ], + "type": "inline_equation", + "content": "i \\notin S_l" + }, + { + "bbox": [ + 46, + 185, + 287, + 257 + ], + "type": "text", + "content": ". Consequently, we have the following form of the proximal gradient operator:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 324 + ], + "type": "interline_equation", + "content": "\\operatorname {p r o x} _ {\\alpha \\gamma \\mathcal {R} _ {w}} (\\mathcal {W} _ {l}) = \\left\\{ \\begin{array}{c} \\frac {\\mathcal {W} _ {l [ i , ; ; ; ]}}{\\| (1 - \\mathbf {v} _ {l}) \\odot \\mathcal {W} _ {l} \\| _ {2}} \\max \\left(0, - \\frac {\\hat {N} _ {l}}{\\hat {N}} \\alpha \\gamma \\right. \\\\ \\left. + \\| (1 - \\mathbf {v} _ {l}) \\odot \\mathcal {W} _ {l} \\| _ {2}\\right), \\text {i f} i \\in S _ {l}, \\\\ \\mathcal {W} _ {l [ i, ; ; ; ; ]}, \\text {i f} i \\notin S _ {l}. \\end{array} \\right. \\tag {8}", + "image_path": "b0c074e4657d368eebfdafeb481a1098e7303957f172ecef6b1f202b0fd89cba.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": "In Eq. 8, we omit the step notation " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": " to simplify the notations, and we still assume " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "W_{l}" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": " is pruned along the output dimension. In this equation, it is easier to see that " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\frac{\\hat{N}_l}{N}" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": " can balance the regularization strength given different layers. Due to the property of our proposed partial regularization, the term " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\| (1 - \\mathbf{v}_l)\\odot \\mathcal{W}_l\\| _2" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": " also changes dynamically, since " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\hat{N}_l" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": " is changed after updates of the AGN. As a result, if no adjustment is applied, no matter how large or small " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\hat{N}_l" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": " is, only a constant value " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\frac{1}{L}\\alpha \\gamma" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": " will be used for the soft-thresholding, " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\max (0,\\| (1 - \\mathbf{v}_l)\\odot \\mathcal{W}_l\\| _2 - \\frac{1}{L}\\alpha \\gamma)" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": ", in all circumstances, which is not reasonable. To accompany the changes of " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": " and consequently " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\hat{N}_l" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\frac{\\hat{N}_l}{N}" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": " to dynamically balance the soft-thresholding parameter between different layers. With Eq. 8, we can efficiently update " + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 326, + 287, + 509 + ], + "type": "text", + "content": " with our proposed partial regularization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 515, + 272, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 272, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 272, + 529 + ], + "type": "text", + "content": "3.5. Network Pruning via Structural Alignment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "We present the algorithm of our method in Algorithm 1. In Algorithm 1, " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is the training dataset; " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{AGN}}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is a sub-dataset within " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " and it is used to train AGN; " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " decides how much FLOPs is preserved and it described in section 3.2; " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " are hyperparameters to control the strength of " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{FLOPs}}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_w" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is the total number of epochs; " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is the start epoch to train AGN and apply " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_w" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " when optimizing model weights. Note that, to reduce the overhead brought by training AGN, we only use a small sub-set sampled from " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": ". As we discussed in section 1, we need to set a start epoch for AGN and " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_w" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": ". If we apply " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_w" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " when training starts " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "(E_{\\mathrm{start}} = 0)" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": ", it will largely restrict the final performance of the whole model before pruning. Instead, we can have deserved results if we start partial regularization in the middle of the training process. The reason for this problem can" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "type": "text", + "content": "come from several perspectives. For example, at the beginning of the training, the classification loss can not produce accurate guidance for pruning since weights are not trained properly. Consequently, it will produce a bad sub-network, and following this sub-network only gives even worse results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 144, + 546, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 240 + ], + "type": "text", + "content": "We summarize our method in Fig. 1. The inference path when training model weights and the AGN are different, and they are connected by partial regularization, which is different from current one-shot pruning and soft pruning methods. With this design, model weights are not directly affected by the sub-network architecture, which creates a smooth transition for the selected sub-network before and after pruning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 244, + 386, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 244, + 386, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 244, + 386, + 258 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 263, + 365, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 263, + 365, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 263, + 365, + 277 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 282, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 389 + ], + "type": "text", + "content": "We use CIFAR-10 [26] and ImageNet [5] to evaluate the performance of our method. Our method requires one hyper-parameter " + }, + { + "bbox": [ + 304, + 282, + 545, + 389 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 282, + 545, + 389 + ], + "type": "text", + "content": " to control the FLOPs budget. The detailed choices of " + }, + { + "bbox": [ + 304, + 282, + 545, + 389 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 282, + 545, + 389 + ], + "type": "text", + "content": " are listed in supplementary materials. We choose ResNets [14] and MobileNet-V2 [49] for comparison. For CIFAR-10, we compare our method with other methods on ResNet-56 and MobileNetV2. For ImageNet, we select ResNet-34, ResNet-50, ResNet-101, and MobileNetV2 as our target models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": " in Eq. 3 to 4.0 for all models and datasets. Similarly, we set " + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": " to 0.0005 for all settings. We set " + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}}" + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": " of the total training epochs. Detailed numbers of " + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}}" + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": " are listed in supplementary materials. The range of " + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}}" + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": " is quite large, which is hard to be explored thoroughly. The current setting already provides good results, but better settings may also exist. To reduce the training costs of the AGN, we random sample " + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": " of samples from the original dataset for constructing " + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{AGN}}" + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": ". With this setup, the additional costs are less than " + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": " of the original training costs. We train the parameters " + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 390, + 546, + 654 + ], + "type": "text", + "content": " of the AGN using ADAM [25] with a start learning rate 0.001. Besides the training of the AGN, we follow the standard training recipe of ResNets for both CIFAR-10 and ImageNet. For MobileNet-V2, We follow the training settings in their original paper [49]. After training, we prune the model by using the sub-network generated from the AGN. The fine-tuning settings are similar to the training setting. Due to space limitations, details of training and fine-tuning are also presented in supplementary materials. In the experimental section, our method is abbreviated as SANP: Structural Alignment for Network Pruning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 659, + 413, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 413, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 413, + 671 + ], + "type": "text", + "content": "4.2. CIFAR-10 Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "The results of CIFAR-10 are presented in Tab. 1. For ResNet-56, our method achieves the best performance (in terms of " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "-Acc) compared to other methods. Specifically," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17406" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 103, + 476, + 270 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 95 + ], + "type": "text", + "content": "Table 1: Comparison of results on CIFAR-10. " + }, + { + "bbox": [ + 46, + 70, + 545, + 95 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 70, + 545, + 95 + ], + "type": "text", + "content": "-Acc represents the performance changes relative to the baseline, and " + }, + { + "bbox": [ + 46, + 70, + 545, + 95 + ], + "type": "inline_equation", + "content": "+ / -" + }, + { + "bbox": [ + 46, + 70, + 545, + 95 + ], + "type": "text", + "content": " indicates an increase/decrease, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 116, + 103, + 476, + 270 + ], + "lines": [ + { + "bbox": [ + 116, + 103, + 476, + 270 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 476, + 270 + ], + "type": "table", + "html": "
ArchitectureMethodBaseline AccPruned AccΔ-AccPruned FLOPs
ResNet-56DCP-Adapt [60]93.80%93.81%+0.01%47.0%
SCP [23]93.69%93.23%-0.46%51.5%
FPGM [18]93.59%92.93%-0.66%52.6%
SFP [16]93.59%92.26%-1.33%52.6%
FPC [15]93.59%93.24%-0.25%52.9%
HRank [35]93.26%92.17%-0.09%50.0%
DMC [9]93.62%92.69%+0.07%50.0%
GNN-RL [57]93.49%93.59%+0.10%54.0%
SANP (ours)93.49%93.81%+0.32%52.0%
MobileNetV2Uniform [60]94.47%94.17%-0.30%26.0%
DCP [60]94.47%94.69%+0.22%26.0%
DMC [9]94.23%94.49%+0.26%40.0%
SCOP [53]94.48%94.24%-0.24%40.3%
SANP (ours)94.52%94.97%+0.45%46.0%
", + "image_path": "5f0c7ce49a5171b0d5bff126f0ecda9c9e964ac61b0d6f7bed186765404b9d3c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 279, + 289, + 381 + ], + "blocks": [ + { + "bbox": [ + 50, + 279, + 289, + 381 + ], + "lines": [ + { + "bbox": [ + 50, + 279, + 289, + 381 + ], + "spans": [ + { + "bbox": [ + 50, + 279, + 289, + 381 + ], + "type": "image", + "image_path": "a69ecfd6f65995fcb7b7a6c90c2cb5ebe89d2669a23950b0e09b6a684353793b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 386, + 197, + 396 + ], + "lines": [ + { + "bbox": [ + 143, + 386, + 197, + 396 + ], + "spans": [ + { + "bbox": [ + 143, + 386, + 197, + 396 + ], + "type": "text", + "content": "(a) ResNet-56." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 279, + 541, + 382 + ], + "blocks": [ + { + "bbox": [ + 302, + 279, + 541, + 382 + ], + "lines": [ + { + "bbox": [ + 302, + 279, + 541, + 382 + ], + "spans": [ + { + "bbox": [ + 302, + 279, + 541, + 382 + ], + "type": "image", + "image_path": "44889f264491c4ad361fd83bc38ac42dff4cb89bd70997869a23378856d8663f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 386, + 457, + 396 + ], + "lines": [ + { + "bbox": [ + 388, + 386, + 457, + 396 + ], + "spans": [ + { + "bbox": [ + 388, + 386, + 457, + 396 + ], + "type": "text", + "content": "(b) MoibleNet-V2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": "our method outperforms the second best method GNN-RL by " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "0.22\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " regarding " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": "-Acc (SANP " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "+0.32\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " vs. GNN-RL " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "+0.10\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " when pruning similar FLOPs (SANP " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "52.0\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " vs. GNN-RL " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "54.0\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": "). Our method also outperforms HRank and DMC by " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "0.25\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "0.41\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " separately (DMC " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "+0.07\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " and HRank " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "-0.09\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": "). The gap between other methods and our method is even larger. For MobileNet-V2, our method prunes most FLOPs " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "(46.0\\%)" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " and achieves the best performance (" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": "-Acc: " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "+0.45\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": "). Compared to the second best method, DMC, our method prunes " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " more FLOPs and outperforms it by " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "0.19\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": ". SCOP also prunes " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " FLOPs, and the gap between our method and SCOP is even larger " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "(0.69\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " better in terms of " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": "-Acc). The uniform setting and DCP prunes around " + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 46, + 430, + 290, + 610 + ], + "type": "text", + "content": " FLOPs, but the performance is still worse than our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 620, + 153, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 153, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 153, + 633 + ], + "type": "text", + "content": "4.3. ImageNet Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "text", + "content": "All results for the ImageNet dataset are shown in Tab. 2. ResNet-34. Our method achieves " + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "inline_equation", + "content": "73.43\\%" + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "text", + "content": " Top-1 accuracy and " + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "inline_equation", + "content": "91.48\\%" + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "text", + "content": " Top-5 accuracy, which is better than the baseline by " + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.19\\%" + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.16\\%" + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "text", + "content": " for Top-1/Top-5 accuracy separately. At the same time, our method removes " + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "inline_equation", + "content": "44.1\\%" + }, + { + "bbox": [ + 46, + 639, + 287, + 713 + ], + "type": "text", + "content": " FLOPs, which is on par with other methods. It is obvious" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 427, + 423, + 517 + ], + "blocks": [ + { + "bbox": [ + 307, + 427, + 423, + 517 + ], + "lines": [ + { + "bbox": [ + 307, + 427, + 423, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 427, + 423, + 517 + ], + "type": "image", + "image_path": "86694253368960dfd7c748f962eb19a7c490cb11809763b991738ab32460be87.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 521, + 372, + 532 + ], + "lines": [ + { + "bbox": [ + 361, + 521, + 372, + 532 + ], + "spans": [ + { + "bbox": [ + 361, + 521, + 372, + 532 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 542, + 544, + 565 + ], + "lines": [ + { + "bbox": [ + 306, + 542, + 544, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 542, + 544, + 565 + ], + "type": "text", + "content": "Figure 3: Top-1 and Top-5 accuracy after pruning given different settings." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 428, + 426, + 544, + 517 + ], + "blocks": [ + { + "bbox": [ + 47, + 406, + 544, + 418 + ], + "lines": [ + { + "bbox": [ + 47, + 406, + 544, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 544, + 418 + ], + "type": "text", + "content": "Figure 2: (a) and (b): the average norm of channels with or without partial regularization for ResNet-56 and MobileNet-V2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 426, + 544, + 517 + ], + "lines": [ + { + "bbox": [ + 428, + 426, + 544, + 517 + ], + "spans": [ + { + "bbox": [ + 428, + 426, + 544, + 517 + ], + "type": "image", + "image_path": "824a7051fd2f86530a986497fab1772f7b059443360c694b93b27c11738777ee.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 521, + 492, + 531 + ], + "lines": [ + { + "bbox": [ + 481, + 521, + 492, + 531 + ], + "spans": [ + { + "bbox": [ + 481, + 521, + 492, + 531 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": "that the advantage of our method is clear compared to other methods. Our method prunes similar FLOPs to SCOP and DMC. However, the " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": " Top-1 Acc of our method is " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "0.88\\%" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "0.92\\%" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": " better than SCOP and DMC, respectively, and we have similar observations for " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": " Top-5 Acc (0.60% and " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "0.47\\%" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": " better than SCOP and DMC). Taylor has the second best " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": " Top-1 Acc, but the pruned FLOPs is much lower than our method (ours: " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "44.1\\%" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": " vs. Taylor: " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "24.2\\%" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": "). The advantage of our method compared to FPGM is more obvious." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "ResNet-50. Our method achieves " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "76.47\\%" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " Top-1 accuracy and " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "93.00\\%" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " Top-5 accuracy, which is also better than the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17407" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 91, + 506, + 358 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 470, + 83 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 470, + 83 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 470, + 83 + ], + "type": "text", + "content": "Table 2: Comparison results on ImageNet with ResNet-34/50/101 and MobileNet-V2." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 91, + 506, + 358 + ], + "lines": [ + { + "bbox": [ + 85, + 91, + 506, + 358 + ], + "spans": [ + { + "bbox": [ + 85, + 91, + 506, + 358 + ], + "type": "table", + "html": "
ArchitectureMethodBaseline Top-1 AccBaseline Top-5 AccΔ Top-1 AccΔ Top-5 AccPruned FLOPs
ResNet-34FPGM [18]73.92%91.62%-1.29%-0.54%41.1%
Taylor [42]73.31%--0.48%-24.2%
DMC [9]73.30%91.42%-0.73%-0.31%43.4%
SCOP [53]73.31%91.42%-0.69%-0.44%44.8%
SANP (ours)73.24%91.32%+0.19%+0.16%44.1%
ResNet-50DCP [60]76.01%92.93%-1.06%-0.61%55.6%
CCP [44]76.15%92.87%-0.94%-0.45%54.1%
FPGM [18]76.15%92.87%-1.32%-0.55%53.5%
ABCP [36]76.01%92.96%-2.15%-1.27%54.3%
DMC [9]76.15%92.87%-0.80%-0.38%55.0%
SCOP [53]76.15%92.87%-0.89%-0.34%54.6%
PFP [34]76.13%92.86%-0.92%-0.45%44.0%
CHIP [52]76.15%92.87%+0.00%+0.04%48.7%
NPPM [8]76.15%92.87%-0.19%+0.12%56.0%
Random-Pruning [30]75.83%92.92%-0.75%-0.40%51.0%
GNN-RL [57]76.10%--1.82%-53.0%
SANP (ours)76.06%92.86%+0.41%+0.17%56.2%
ResNet-101FPGM [18]77.37%93.56%-0.05%0.00%41.1%
Taylor [42]77.37%--0.02%-39.8%
DMC [9]77.37%93.56%+0.04%+0.03%56.0%
PFP [34]77.37%93.56%-0.94%-0.44%45.1%
SANP (ours)77.53%93.71%+0.61%+0.29%55.4%
MobileNet-V2Uniform [49]71.80%91.00%-2.00%-1.40%30.0%
AMC [17]71.80%--1.00%-30.0%
CC [33]71.88%--0.97%-28.3%
MetaPruning [38]72.00%--0.80%-30.7%
Random-Pruning [30]71.88%--1.01%-29.1%
SANP (ours)71.91%90.30%+0.14%+0.07%29.1%
", + "image_path": "26b799c5748001e05f74af0d6f1d83ef5b28e4d12fa1df3f57d7494325e0251e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 370, + 285, + 430 + ], + "blocks": [ + { + "bbox": [ + 49, + 370, + 285, + 430 + ], + "lines": [ + { + "bbox": [ + 49, + 370, + 285, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 370, + 285, + 430 + ], + "type": "table", + "html": "
SettingsArchitectureBaseline Top-1 AccΔ Top-1 AccPruned FLOPs
One-Shot SANPResNet-3473.31%-0.50%44.0%
73.24%+0.19%44.1%
One-Shot SANPResNet-5076.13%-0.57%56.0%
76.06%+0.41%55.2%
One-Shot SANPMobileNetV271.88%-0.42%29.3%
71.91%+0.07%29.1%
", + "image_path": "def7b7fd1b5f65fa5b6d64c6b364d2bbfc780322729256b4a66e21483374f6b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 439, + 287, + 463 + ], + "lines": [ + { + "bbox": [ + 47, + 439, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 287, + 463 + ], + "type": "text", + "content": "Table 3: Performance of pruned models given different pruning settings on ImageNet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": "baseline Top-1/Top-5 accuracy. The second best method, CHIP, removes " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "48.7\\%" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " FLOPs while maintaining the original performance. Our method outperforms CHIP by " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "0.41\\%" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " in terms of the " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " Top-1 accuracy while pruning near " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " more FLOPs. NPPM is a strong baseline for ResNet-50, our method outperforms by " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "0.60\\%" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " -Top-1 Acc. GNN-RL and Random-Pruning are two recent pruning works. Our method outperforms them by " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "2.23\\%" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "1.16\\%" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " separately, while our method prunes more FLOPs. PHP, DCP, CCP, and DMC have similar " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " -Top 1 accuracy. The gap between our method and these methods ranges from " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "1.21\\%" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "1.47\\%" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " regarding " + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 482, + 287, + 638 + ], + "type": "text", + "content": " -Top 1 accuracy. The advantage of our method compared to the rest methods is more apparent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": "ResNet-101. Our method achieves " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "78.14\\%" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " Top-1 accuracy and " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "94.00\\%" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " Top-5 accuracy, which is " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.61\\%" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.29\\%" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " better than the baseline Top-1 and Top-5 accuracy. Although DMC prunes a little bit more FLOPs, it only achieves " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "77.41\\%" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " Top-1 accuracy after pruning which is " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.73\\%" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " lower than our method. FPGA, Taylor, and PFP remove less than" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "type": "text", + "content": "50% FLOPs. Our method prunes at least 10% more FLOPs and still has an advantage in terms of " + }, + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 305, + 372, + 545, + 407 + ], + "type": "text", + "content": "-Top 1 accuracy (from 0.66% to 1.55%)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": "MobileNet-V2. MobileNet-V2 is generally harder to prune compared to ResNets. All comparison methods on MobileNet-V2 remove around " + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": " FLOPs. Our method achieves " + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "inline_equation", + "content": "72.05\\%" + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": " Top-1 accuracy and " + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "inline_equation", + "content": "90.37\\%" + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": " Top-5 accuracy, which is " + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "inline_equation", + "content": "0.14\\%" + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "inline_equation", + "content": "0.07\\%" + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": " better than the baseline Top-1 and Top-5 accuracy. Given the similar pruning rate, our method is " + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "inline_equation", + "content": "1.15\\%" + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "inline_equation", + "content": "0.94\\%" + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "inline_equation", + "content": "1.11\\%" + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "inline_equation", + "content": "1.14\\%" + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": " higher than Random-Pruning, MetaPruning CC, and AMC separately regarding " + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 408, + 546, + 552 + ], + "type": "text", + "content": "-Top-1 Acc. MetaPruning prunes most FLOPs, but the performance is much lower than our method. In short, our method can also be applied to lightweight CNNs, like MobileNet-V2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 559, + 441, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 441, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 441, + 571 + ], + "type": "text", + "content": "4.4. Analysis of Our Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": "The effectiveness of partial regularization. To verify whether our proposed partial regularization is effective, we plot the average channel norm of different groups within each block for ResNet-56 and MobileNet-V2 on CIFAR-10. The results are shown in Fig. 2. The average channel norm for the group with partial regularization and without partial regularization are obtained by " + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\tilde{N}_l}\\sum_{i\\in S_l}\\| \\mathcal{W}_{l[i,:;:]} \\|_{\\mathrm{GL}}" + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\frac{1}{C_l - \\tilde{N}_l}\\sum_{i\\notin S_l}\\| \\mathcal{W}_{l[i,:;:]} \\|_{\\mathrm{GL}}" + }, + { + "bbox": [ + 304, + 576, + 547, + 715 + ], + "type": "text", + "content": " separately. Weights with partial regularization are effectively aligned. On the contrary, weights without partial regularization are lightly affected, which justifies the strength of our proposed partial regular" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17408" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 156, + 156 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 156, + 156 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 156, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 156, + 156 + ], + "type": "image", + "image_path": "281bc250104f5325c56a506ac5bfa79fdea6ceacdbf80cfd963b2b98953b7825.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 69, + 284, + 156 + ], + "blocks": [ + { + "bbox": [ + 175, + 69, + 284, + 156 + ], + "lines": [ + { + "bbox": [ + 175, + 69, + 284, + 156 + ], + "spans": [ + { + "bbox": [ + 175, + 69, + 284, + 156 + ], + "type": "image", + "image_path": "e29c52c97e5da8fdfcdc649c285d170d024999e8b8f3b8dcdb660622505e323c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 411, + 156 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 411, + 156 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 411, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 411, + 156 + ], + "type": "image", + "image_path": "6a4ccd9bd6001d512932246f5ed7b4192731416cd37dfbf2f1c1c7effb7c2ead.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 433, + 69, + 541, + 156 + ], + "blocks": [ + { + "bbox": [ + 433, + 69, + 541, + 156 + ], + "lines": [ + { + "bbox": [ + 433, + 69, + 541, + 156 + ], + "spans": [ + { + "bbox": [ + 433, + 69, + 541, + 156 + ], + "type": "image", + "image_path": "03ece7e4f2f553b57c01f487e585c92c27a064a44bc23463885b5fa8a8edc4da.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 47, + 172, + 154, + 257 + ], + "blocks": [ + { + "bbox": [ + 97, + 161, + 110, + 171 + ], + "lines": [ + { + "bbox": [ + 97, + 161, + 110, + 171 + ], + "spans": [ + { + "bbox": [ + 97, + 161, + 110, + 171 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 172, + 154, + 257 + ], + "lines": [ + { + "bbox": [ + 47, + 172, + 154, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 172, + 154, + 257 + ], + "type": "image", + "image_path": "e4addfc2a4679d8270777b550b2b9a49a82b5adb6e486c44e85482612dc02af7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 98, + 262, + 110, + 272 + ], + "lines": [ + { + "bbox": [ + 98, + 262, + 110, + 272 + ], + "spans": [ + { + "bbox": [ + 98, + 262, + 110, + 272 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "text", + "content": "Figure 4: (a, e): the impact of " + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{FLOPs}}" + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "text", + "content": ". (b, f): the effect of the architecture of AGN. (c, g): the effect of " + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}}" + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "text", + "content": ". (d, h): the effect of different setups. Experiments are conducted on CIFAR-10 with ResNet-56 and " + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "inline_equation", + "content": "p = 0.5" + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "text", + "content": " (a,b,c,d,e) and " + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "inline_equation", + "content": "p = 0.35" + }, + { + "bbox": [ + 46, + 282, + 547, + 319 + ], + "type": "text", + "content": " (f,g,h)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 175, + 171, + 284, + 257 + ], + "blocks": [ + { + "bbox": [ + 227, + 161, + 238, + 171 + ], + "lines": [ + { + "bbox": [ + 227, + 161, + 238, + 171 + ], + "spans": [ + { + "bbox": [ + 227, + 161, + 238, + 171 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 171, + 284, + 257 + ], + "lines": [ + { + "bbox": [ + 175, + 171, + 284, + 257 + ], + "spans": [ + { + "bbox": [ + 175, + 171, + 284, + 257 + ], + "type": "image", + "image_path": "c0ddbd8370737cf69b68333a1cde4cd37120b0cc67bad76bf7a3cb729dc146f8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 262, + 238, + 272 + ], + "lines": [ + { + "bbox": [ + 227, + 262, + 238, + 272 + ], + "spans": [ + { + "bbox": [ + 227, + 262, + 238, + 272 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 304, + 172, + 412, + 257 + ], + "blocks": [ + { + "bbox": [ + 355, + 161, + 367, + 171 + ], + "lines": [ + { + "bbox": [ + 355, + 161, + 367, + 171 + ], + "spans": [ + { + "bbox": [ + 355, + 161, + 367, + 171 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 172, + 412, + 257 + ], + "lines": [ + { + "bbox": [ + 304, + 172, + 412, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 412, + 257 + ], + "type": "image", + "image_path": "6a8e34b812542773adfbf228da27dc0c06c62eea1146a08ef07c23af33365447.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 262, + 367, + 273 + ], + "lines": [ + { + "bbox": [ + 355, + 262, + 367, + 273 + ], + "spans": [ + { + "bbox": [ + 355, + 262, + 367, + 273 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 433, + 171, + 541, + 257 + ], + "blocks": [ + { + "bbox": [ + 484, + 160, + 496, + 170 + ], + "lines": [ + { + "bbox": [ + 484, + 160, + 496, + 170 + ], + "spans": [ + { + "bbox": [ + 484, + 160, + 496, + 170 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 433, + 171, + 541, + 257 + ], + "lines": [ + { + "bbox": [ + 433, + 171, + 541, + 257 + ], + "spans": [ + { + "bbox": [ + 433, + 171, + 541, + 257 + ], + "type": "image", + "image_path": "67bd6b1df5e318a5e01a1c01a1ed1147e8a891764e311eb5c0c47dbe41c7418f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 262, + 496, + 272 + ], + "lines": [ + { + "bbox": [ + 484, + 262, + 496, + 272 + ], + "spans": [ + { + "bbox": [ + 484, + 262, + 496, + 272 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 335, + 80, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 80, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 80, + 346 + ], + "type": "text", + "content": "ization." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "text", + "content": "The impact of " + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "text", + "content": ". We study the impact of " + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "text", + "content": " when training the AGN, and we plot the test accuracy and " + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{FLOPs}}" + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "text", + "content": " in Fig. 4a and Fig. 4e. From the figures, we can see that if " + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "text", + "content": " is too large, it will have negative impacts on learned subnetworks. Otherwise, our method is robust to " + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 348, + 287, + 408 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 410, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 506 + ], + "type": "text", + "content": "The effect of AGN. In Fig. 4b and Fig. 4f, we plot the test accuracy when learning the AGN given different pruning rates. We construct a Simple baseline, which parameterizes each channel by using one learnable parameter. We can see that when the parameterization space shrinks, the performance of learned sub-networks will be affected severely. In addition, the learning is slower, and the best sub-network performance is also much worse than using AGN." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": "The effect of " + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{start}}" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": ". In the method section, we argue that inserting partial regularization in the middle training process is beneficial. We plot the results of " + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}} = 0" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}} = 40" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": " in Fig. 4c and Fig. 4g. In general, they can find sub-networks with similar performance, but " + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}} = 0" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": " is worse than " + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}} = 40" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": " when the pruning rate is large (" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "inline_equation", + "content": "p = 0.35" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": "). More importantly, the full model accuracy of " + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}} = 0" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "inline_equation", + "content": "92.96\\%" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": " (average across different runs), which is around " + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "inline_equation", + "content": "0.50\\%" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": " worse than " + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{start}} = 40" + }, + { + "bbox": [ + 46, + 508, + 288, + 639 + ], + "type": "text", + "content": ", which suggests that using partial regularization at the beginning will limit the capacity of the full model." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": "The effect of different setups. We construct two additional baselines to see how partial regularization helps to produce a better sub-network within the full model. The One-Shot baseline directly trains the AGN on the pre-trained model. The w/o Partial Regularization baseline set " + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\gamma = 0" + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": ", and the rest settings are the same as our method. The related" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 335, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 547, + 538 + ], + "type": "text", + "content": "results are shown in Fig. 4d and Fig. 4h. From these figures, we can see that partial regularization always produces the best sub-network from the full model. In addition, 'w/o partial regularization' is worse than the one-shot setting. This is probably because it is hard to capture the changes of model weights without partial regularization, which makes the training of AGN much harder than the rest settings. More comparisons on the ImageNet dataset. To further show how our method improves the one-shot setting, we present the sub-network performance before and after fine-tuning for one-shot and partial regularization settings in Fig. 3 and Tab. 3. From Fig. 3, we can see that partial regularization still produces better sub-networks on ImageNet, and it is around " + }, + { + "bbox": [ + 304, + 335, + 547, + 538 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 335, + 547, + 538 + ], + "type": "text", + "content": " better in terms of Top-1 accuracy than the one-shot setting for different models. The advantage of partial regularization naturally extends to results after fine-tuning, as shown in Tab. 3." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 558, + 378, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 378, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 378, + 571 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we investigate how partial regularization helps to produce a better sub-network for network pruning. Specifically, our method uses AGN to guide partial regularization across the training process. We further provide an efficient way to update model weights through proximal gradients. With these designs, partial regularization effectively reduces the gap between the sub-network within the full model and the pruned model. Our method then starts from a better sub-network, thus resulting in a better final pruned model. Extensive experimental results on CIFAR-10 and ImageNet show the effectiveness of our method." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17409" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 283 + ], + "type": "text", + "content": "[4] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in neural information processing systems, pages 3123-3131, 2015. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 339 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 396 + ], + "type": "text", + "content": "[6] Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, and H.T. Kung. Neural mean discrepancy for efficient out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19217-19227, June 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "text", + "content": "[7] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 443, + 288, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 288, + 497 + ], + "type": "text", + "content": "[8] Shangqian Gao, Feihu Huang, Weidong Cai, and Heng Huang. Network pruning via performance maximization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9270-9280, 2021. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 499, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 288, + 554 + ], + "type": "text", + "content": "[9] Shangqian Gao, Feihu Huang, Jian Pei, and Heng Huang. Discrete model compression with resource constraint for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1899-1908, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 610 + ], + "type": "text", + "content": "[10] Shangqian Gao, Burak Uzkent, Yilin Shen, Heng Huang, and Hongxia Jin. Learning to jointly share and prune weights for grounding based vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[11] Junfeng Guo, Yiming Li, Xun Chen, Hanqing Guo, Lichao Sun, and Cong Liu. SCALE-UP: An efficient black-box input-level backdoor detection via analyzing scaled prediction consistency. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[12] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[13] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Advances in neural information processing systems, pages 1135-1143, 2015. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 165, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 545, + 219 + ], + "type": "text", + "content": "[15] Yang He, Yuhang Ding, Ping Liu, Linchao Zhu, Hanwang Zhang, and Yi Yang. Learning filter pruning criteria for deep convolutional neural networks acceleration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2009-2018, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 266 + ], + "type": "text", + "content": "[16] Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating deep convolutional neural networks. In International Joint Conference on Artificial Intelligence (IJCAI), pages 2234-2240, 2018. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 268, + 545, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 268, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 268, + 545, + 322 + ], + "type": "text", + "content": "[17] Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, and Song Han. Amc: Automl for model compression and acceleration on mobile devices. In Proceedings of the European Conference on Computer Vision (ECCV), pages 784-800, 2018. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 324, + 545, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 324, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 308, + 324, + 545, + 379 + ], + "type": "text", + "content": "[18] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2019. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 381, + 545, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 381, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 308, + 381, + 545, + 413 + ], + "type": "text", + "content": "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 416, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 547, + 459 + ], + "type": "text", + "content": "[20] Zehao Huang and Naiyan Wang. Data-driven sparse structure selection for deep neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 304-320, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 462, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 462, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 462, + 545, + 517 + ], + "type": "text", + "content": "[21] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32Nd International Conference on International Conference on Machine Learning - Volume 37, ICML, pages 448-456. JMLR.org, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 519, + 545, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 519, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 308, + 519, + 545, + 551 + ], + "type": "text", + "content": "[22] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 554, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 545, + 597 + ], + "type": "text", + "content": "[23] Minsoo Kang and Bohyung Han. Operation-aware soft channel pruning using differentiable masks. In International Conference on Machine Learning, pages 5122-5131. PMLR, 2020. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 643 + ], + "type": "text", + "content": "[24] Jaedeok Kim, Chiyoun Park, Hyun-Joo Jung, and Yoonsuck Choe. Plug-in, trainable gate for streamlining arbitrary neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 646, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 646, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 646, + 545, + 677 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[26] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, Cite-seer, 2009. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17410" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097-1105, 2012. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "text", + "content": "[28] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. *ICLR*, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 151, + 287, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 182 + ], + "type": "text", + "content": "[29] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. *ICLR*, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 184, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 184, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 184, + 287, + 239 + ], + "type": "text", + "content": "[30] Yawei Li, Kamil Adamczewski, Wen Li, Shuhang Gu, Radu Timofte, and Luc Van Gool. Revisiting random channel pruning for neural network compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 191-201, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 287, + 293 + ], + "type": "text", + "content": "[31] Yawei Li, Shuhang Gu, Christoph Mayer, Luc Van Gool, and Radu Timofte. Group sparsity: The hinge between filter pruning and decomposition for network compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8018-8027, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 295, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 295, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 295, + 287, + 348 + ], + "type": "text", + "content": "[32] Yawei Li, Shuhang Gu, Kai Zhang, Luc Van Gool, and Radu Timofte. Dhp: Differentiable meta pruning via hypernetworks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 608-624. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 350, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 287, + 414 + ], + "type": "text", + "content": "[33] Yuchao Li, Shaohui Lin, Jianzhuang Liu, Qixiang Ye, Mengdi Wang, Fei Chao, Fan Yang, Jincheng Ma, Qi Tian, and Rongrong Ji. Towards compact cnns via collaborative compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6438-6447, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 415, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 415, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 49, + 415, + 287, + 458 + ], + "type": "text", + "content": "[34] Lucas Liebenwein, Cenk Baykal, Harry Lang, Dan Feldman, and Daniela Rus. Provable filter pruning for efficient neural networks. In International Conference on Learning Representations, 2020. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 287, + 514 + ], + "type": "text", + "content": "[35] Mingbao Lin, Rongrong Ji, Yan Wang, Yichen Zhang, Baochang Zhang, Yonghong Tian, and Ling Shao. Hrank: Filter pruning using high-rank feature map. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 515, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 515, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 49, + 515, + 287, + 568 + ], + "type": "text", + "content": "[36] Mingbao Lin, Rongrong Ji, Yuxin Zhang, Baochang Zhang, Yongjian Wu, and Yonghong Tian. Channel pruning via automatic structure search. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), pages 673 - 679, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 571, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 571, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 571, + 287, + 613 + ], + "type": "text", + "content": "[37] Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In ICCV, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 668 + ], + "type": "text", + "content": "[38] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In Proceedings of the IEEE International Conference on Computer Vision, pages 3296-3305, 2019. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[39] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[40] Zhaosong Lu and Xiaorui Li. Sparse recovery via partial regularization: Models, theory, and algorithms. Mathematics of Operations Research, 43(4):1290-1316, 2018. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[41] Dmitry Molchanov, Arsenii Ashukha, and Dmitry Vetrov. Variational dropout sparsifies deep neural networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 2498-2507. JMLR.org, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 204 + ], + "type": "text", + "content": "[42] Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11264–11272, 2019. 1, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[43] Kirill Neklyudov, Dmitry Molchanov, Arsenii Ashukha, and Dmitry P Vetrov. Structured bayesian pruning via log-normal multiplicative noise. In Advances in Neural Information Processing Systems, pages 6775-6784, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "text", + "content": "[44] Hanyu Peng, Jiaxiang Wu, Shifeng Chen, and Junzhou Huang. Collaborative channel pruning for deep networks. In International Conference on Machine Learning, pages 5113-5122, 2019. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "text", + "content": "[45] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet classification using binary convolutional neural networks. In European Conference on Computer Vision, pages 525-542. Springer, 2016. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[46] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "text", + "content": "[47] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91-99, 2015. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 468 + ], + "type": "text", + "content": "[48] Alex Renda, Jonathan Frankle, and Michael Carbin. Comparing rewinding and fine-tuning in neural network pruning. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 525 + ], + "type": "text", + "content": "[49] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 2, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "text", + "content": "[50] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in neural information processing systems, pages 568-576, 2014. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "type": "text", + "content": "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 646 + ], + "type": "text", + "content": "[52] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. Chip: Channel independence-based pruning for compact neural networks. Advances in Neural Information Processing Systems, 34, 2021. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[53] Yehui Tang, Yunhe Wang, Yixing Xu, Dacheng Tao, Chunjing Xu, Chao Xu, and Chang Xu. Scop: Scientific control for reliable neural network pruning. Advances in Neural Information Processing Systems, 33, 2020. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[54] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17411" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 385 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "In Advances in neural information processing systems, pages 2074-2082, 2016.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 149 + ], + "type": "text", + "content": "[55] Mao Ye, Chengyue Gong, Lizhen Nie, Denny Zhou, Adam Klivans, and Qiang Liu. Good subnetworks provably exist: Pruning via greedy forward selection. In International Conference on Machine Learning, pages 10820-10830. PMLR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 206 + ], + "type": "text", + "content": "[56] Zhonghui You, Kun Yan, Jinmian Ye, Meng Ma, and Ping Wang. Gate decorator: Global filter pruning method for accelerating deep convolutional neural networks. In Advances in Neural Information Processing Systems, pages 2130-2141, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 286, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 286, + 252 + ], + "type": "text", + "content": "[57] Sixing Yu, Arya Mazaheri, and Ali Jannesari. Topology-aware network pruning using multi-stage graph embedding and reinforcement learning. In International Conference on Machine Learning, pages 25656-25667. PMLR, 2022. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 285 + ], + "type": "text", + "content": "[58] Dejiao Zhang, Haozhu Wang, Mario Figueiredo, and Laura Balzano. Learning to share: Simultaneous parameter tying and sparsification in deep learning. 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 286, + 330 + ], + "type": "text", + "content": "[59] Yanfu Zhang, Shangqian Gao, and Heng Huang. Exploration and estimation for model compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 487-496, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 286, + 385 + ], + "type": "text", + "content": "[60] Zhuangwei Zhuang, Mingkui Tan, Bohan Zhuang, Jing Liu, Yong Guo, Qingyao Wu, Junzhou Huang, and Jinhui Zhu. Discrimination-aware channel pruning for deep neural networks. In Advances in Neural Information Processing Systems, pages 875-886, 2018. 6, 7" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17412" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Structure Invariant Transformation for better Adversarial Transferability/303ed349-48a8-4393-8e2e-88000a7d20c9_content_list.json b/2023/Structure Invariant Transformation for better Adversarial Transferability/303ed349-48a8-4393-8e2e-88000a7d20c9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..08b32143ccb43ac128ac98b22805409d2aeec634 --- /dev/null +++ b/2023/Structure Invariant Transformation for better Adversarial Transferability/303ed349-48a8-4393-8e2e-88000a7d20c9_content_list.json @@ -0,0 +1,1954 @@ +[ + { + "type": "text", + "text": "Structure Invariant Transformation for better Adversarial Transferability", + "text_level": 1, + "bbox": [ + 109, + 130, + 856, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaosen Wang \nHuawei Singularity Security Lab", + "bbox": [ + 99, + 181, + 362, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xiaosen@hust.edu.cn", + "bbox": [ + 143, + 220, + 316, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeliang Zhang HUST", + "bbox": [ + 418, + 181, + 539, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hust0426@gmail.com", + "bbox": [ + 398, + 220, + 560, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianping Zhang Chinese University of Hong Kong", + "bbox": [ + 598, + 181, + 867, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jpzhang@cse.cuhk.edu.hk", + "bbox": [ + 632, + 220, + 836, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given the severe vulnerability of Deep Neural Networks (DNNs) against adversarial examples, there is an urgent need for an effective adversarial attack to identify the deficiencies of DNNs in security-sensitive applications. As one of the prevalent black-box adversarial attacks, the existing transfer-based attacks still cannot achieve comparable performance with the white-box attacks. Among these, input transformation based attacks have shown remarkable effectiveness in boosting transferability. In this work, we find that the existing input transformation based attacks transform the input image globally, resulting in limited diversity of the transformed images. We postulate that the more diverse transformed images result in better transferability. Thus, we investigate how to locally apply various transformations onto the input image to improve such diversity while preserving the structure of image. To this end, we propose a novel input transformation based attack, called Structure Invariant Transformation (SIA), which applies a random image transformation onto each image block to craft a set of diverse images for gradient calculation. Extensive experiments on the standard ImageNet dataset demonstrate that SIA exhibits much better transferability than the existing SOTA input transformation based attacks on CNN-based and transformer-based models, showing its generality and superiority in boosting transferability. Code is available at https://github.com/xiaosen-wang/SIT.", + "bbox": [ + 75, + 300, + 473, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 739, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the unprecedented progress of Deep Neural Networks (DNNs) [25, 20, 23, 50, 12], they have been deployed in many security-sensitive applications, such as face recognition [41, 51, 43, 44], autonomous driving [14, 31], etc. On the other hand, recent works have found that DNNs are vulnerable to adversarial examples [47, 15], which mislead the deep models with imperceptible perturbations. This brings a huge threat to the real-world applications [42, 13, 45, 57, 71, 78] and makes it imperative for an", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af123078fe2f1c8acaac5afc6622d7ea8abf93cff6c1319afb1260a5c7c00e41.jpg", + "image_caption": [ + "Raw Image" + ], + "image_footnote": [], + "bbox": [ + 514, + 301, + 596, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca181908d16984ee20827d0d4bd15ed460b59041280cf6de4dfd54f53a97fc51.jpg", + "image_caption": [ + "DIM" + ], + "image_footnote": [], + "bbox": [ + 607, + 268, + 692, + 333 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f1fc384044917aac920f5630552beba2b0d1f571134445f3b0d1632c3e5883e.jpg", + "image_caption": [ + "Admix" + ], + "image_footnote": [], + "bbox": [ + 607, + 352, + 692, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/832227d53516e0694060b525fa38b28240182fae977294e494752ed3bb079bdb.jpg", + "image_caption": [ + "TIM" + ], + "image_footnote": [], + "bbox": [ + 700, + 268, + 785, + 333 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a08afce371b4e17f65b89af169699fe7de41538ea190d48475a62e164d0a7c57.jpg", + "image_caption": [ + "SSA", + "Figure 1: The raw image and its transformed images by DIM (resize factor of 0.8) [64], TIM (translated by 15 pixels) [11], SIM (scale factor of 0.5), Admix (admix strength of 0.2 and scale factor of 0.5) [55], SSA (turning factor of 0.5) [36] and our proposed SIA $(3\\times 3$ blocks)." + ], + "image_footnote": [], + "bbox": [ + 700, + 352, + 785, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5530125076220063c78af0a9df28ab9cc4112e97291a9315c8250c58c5384086.jpg", + "image_caption": [ + "SIM" + ], + "image_footnote": [], + "bbox": [ + 795, + 268, + 879, + 333 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dc2dddd99efc9d96aace3011f725c70e0ae0898deed9d4f361dbc4b4cb077e95.jpg", + "image_caption": [ + "SIA (Ours)" + ], + "image_footnote": [], + "bbox": [ + 795, + 352, + 879, + 416 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "effective attack to identify the deficiencies of DNNs when we robustify the deep models or deploy them for commercial applications.", + "bbox": [ + 496, + 551, + 890, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing adversarial attacks usually fall into two categories: white-box attacks [15, 38, 37, 26, 54] can fetch any information of the target model, including (hyper-)parameters, gradient, architecture, while black-box attacks [24, 7, 4, 27, 33, 77] are only allowed limited access to the target model. One of the significant properties of adversarial examples is their transferability [64, 10, 53, 76], in which the adversarial examples generated on one model can still mislead other models, making it possible to attack the real-world applications in the black-box setting. However, existing adversarial attacks [26, 37] often exhibit superior white-box attack performance but poor transferability.", + "bbox": [ + 496, + 597, + 892, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To craft more transferable adversarial examples, various techniques have been proposed, such as momentum-based methods [10, 32, 53, 56], input transformations [64, 11, 32, 68, 72], ensemble attacks [33, 66, 36], advanced objective functions [61, 80] and model-specific approaches [28, 60]. Among which, input transformations (e.g., random resizing and padding [64], translation [11], scale [32] admix [55], etc.) that transform the image before gradient calculation,", + "bbox": [ + 496, + 780, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4607", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "have achieved superior transferability and attracted broad attention. Nevertheless, we find that the existing input transformation based attacks, even the SOTA attack Admix [55] and SSA [36], transform the image globally without changing the local relationship among objects in the input image. We argue that the diversity of such transformed images is still not enough, leading to limited transferability.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we postulate and empirically validate that the more diverse transformed images lead to better transferability. Based on this observation, instead of applying a single transformation on the input image, we apply different transformations locally on different parts of the image to enhance the diversity of transformed images. As shown in Fig. 1, such transformation can bring much more difference to the generated image compared with the raw image but still preserve the global structure of the object. Based on this transformation, we propose a novel input transformation based attack called structure invariant attack (SIA), which utilizes the gradient of these transformed images to update the adversarial examples for better transferability.", + "bbox": [ + 75, + 198, + 470, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 96, + 393, + 398, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We empirically validate that the high diversity of transformed images is beneficial to improve transferability, which sheds new light on how to design new input transformations for more transferable adversarial examples.", + "- We design a new image transformation method, which locally applies different transformations on different parts of the image to generate more diverse images but preserve its global structure.", + "- Based on the proposed image transformation, we devise a new input transformation based attack, called structure invariant attack (SIA), to generate more transferable adversarial examples.", + "- Extensive experiments on ImageNet dataset demonstrate that SIA outperforms the baselines with a clear margin on CNN-based as well as transformer-based models, showing its superiority and generality." + ], + "bbox": [ + 76, + 411, + 467, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 664, + 217, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial examples [47] have brought an impressive threat to the DNN-enabled applications, such as computer vision [15, 58, 3], natural language processing [1, 52, 69], speech recognition [5, 70], etc. To identify the vulnerability of DNNs, various attack methods have been proposed recently, such as gradient-based attacks [15, 38, 26, 37], score-based attacks [24, 49, 17, 29, 2], decision-based attacks [4, 6, 27, 57] and transfer-based attacks [33, 10, 64, 32, 55]. Among these, transfer-based attacks do not access any information of the target model, making it applicable to attack any model in the physical world. In this work, we focus on generating more transferable adversarial examples and briefly introduce the existing transfer-based attacks, the corresponding defense methods, and data augmentations.", + "bbox": [ + 75, + 688, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Adversarial Attack", + "text_level": 1, + "bbox": [ + 500, + 90, + 683, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the first gradient-based attack, Fast Gradient Sign Method (FGSM) [15] adds the perturbation in the gradient direction to the benign sample, leading to high attack efficiency but limited performance. Later, I-FGSM [26] extends FGSM into an iterative version, which achieves much better white-box attack performance but poor transferability. Given the high efficiency and effectiveness of I-FGSM, numerous transfer-based attacks are proposed to boost transferability based on I-FGSM to attack the deep model in the black-box setting.", + "bbox": [ + 496, + 113, + 890, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Momentum-based methods. MI-FGSM [10] introduces momentum into I-FGSM to stabilize the optimization direction and escape local maxima. NI-FGSM [32] adopts Nesterov Accelerated Gradient to accumulate the momentum, which achieves better transferability. Variance tuning [53] adopts the gradient variance of the previous iteration to tune the current gradient in MI-FGSM and NI-FGSM, significantly improving transferability. EMIFGSM [56] enhances the momentum by accumulating several samples' gradients in the previous gradient's direction to further stabilize the optimization direction.", + "bbox": [ + 496, + 266, + 890, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Input transformation based attacks. DIM [64] is the first input transformation based attack, which adds padding to a randomly resized image for fixed size before gradient calculation. TIM [11] optimizes the perturbation over an ensemble of translated images, which is further approximated by convolving the gradient at the untranslated image with a pre-defined kernel. SIM [32] scales the images with different scale factors for gradient calculation. DEM [81] averages the gradient on several diverse transformed images similar to DIM with various resizing factors. Admix [55] calculates the gradient on the input image admixed with a small portion of each add-in image from other categories while using the original label of the input. Wu et al. [62] train an adversarial transformation network to destroy the adversarial perturbation and require the synthesized adversarial examples resistant to such transformations. SSA [36] adds Gaussian noise and randomly masks the image in the frequency domain to transform the input image.", + "bbox": [ + 496, + 431, + 890, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ensemble attacks. Liu et al. [33] first found that the adversarial examples generated on multiple models denoted as ensemble attack, are more transferable. Recently, Xiong et al. [66] reduce the gradient variance between various models to boost the ensemble attack.", + "bbox": [ + 496, + 704, + 890, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advanced objective functions. The above attacks often take the cross-entropy loss as the objective function. Researchers also find that some regularizers are beneficial to boost transferability. For instance, Zhou et al. [80] additionally maximize the difference of the intermediate feature maps between the benign sample and adversarial example. Wu et al. [61] adopt a regularizer about the distance of the attention maps between these samples.", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4608", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model-specific approaches. Some works utilize the surrogate model's architecture to improve transferability. Li et al. [28] densely add dropout [46] after each layer to create several ghost networks for better transferability. SGM [60] adopts more gradient from the skip connection in ResNets [20] to boost adversarial transferability. LinBP [19] replaces the zeros with ones in the derivative of ReLU to make the model more linear, leading to improved transferability.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Adversarial Defense", + "text_level": 1, + "bbox": [ + 76, + 234, + 269, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Numerous adversarial defenses have been proposed to mitigate the threat of adversarial attacks. Adversarial training [15, 48], which adopts the adversarial examples during the training process, has been shown as one of the most effective methods [3] but taking huge computation cost. Recently, Wong et al. [59] find that single-step adversarial examples can bring satisfiable robustness, making it possible to be applied on large-scale datasets (e.g., ImageNet [25]). Pre-processing the input samples before the model is shown to be another effective way. Liao et al. [30] design a high-level representation guided denoiser (HGD) to eliminate the adversarial perturbation. Xie et al. [63] find that random resizing and padding on the input image can mitigate the adversarial threat. Naseer et al. [39] train a neural representation purifier (NRP) by a self-supervised adversarial training mechanism to purify the input sample, which exhibits superior effectiveness against transfer-based adversarial examples. On the other hand, certified defense methods aim to provide provable defense in a given radius [16, 74, 8]. For instance, randomized smoothing (RS) trains a robust ImageNet classifier with a tight robustness guarantee [8].", + "bbox": [ + 75, + 258, + 472, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Data Augmentations", + "text_level": 1, + "bbox": [ + 76, + 583, + 272, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data augmentations often transform (e.g., flipping, rotation, cropping, etc.) the image during the training process for better generalization. Mixup [75] interpolates two images and their labels to generate virtual samples for training, which also inspires Admix to enhance transferability. Cutmix [73] pastes an image patch to the original patch and mixes the labels accordingly. AutoAugment [9] automatically searches for improved data augmentation policies (operations and parameters) on the dataset for better generalization, which has been widely adopted in deep learning. Unlike these data augmentation strategies, we aim to construct a set of diverse images by transforming the image block using various transformations, which can be used for gradient calculation to achieve better transferability.", + "bbox": [ + 75, + 606, + 468, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 830, + 210, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our motivation, provide a detailed description of the proposed SIA, and highlight the difference between SIA and AutoAugment.", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/4da5e54a36f24390a471c5ebc5025009613a13999e90785ebf00e3631ce6de52.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TIMDIMSIMSSAAdmix
Transferability57.477.679.380.683.6
LPIPS0.250.430.480.540.73
", + "bbox": [ + 506, + 88, + 885, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: The transferability of TIM, DIM, SIM, Admix, SSA, and similarity between 1,000 images and the transformed images evaluated by LPIPS. The transferability is evaluated by the attack success rate of Inception-v3 on the adversarial examples generated on ResNet-18.", + "bbox": [ + 496, + 160, + 890, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 500, + 260, + 622, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since Xie et al. [64] found that transforming the image by random resizing and padding before gradient calculation can generate more transferable adversarial examples, various input transformation based attacks are proposed to further improve transferability. As shown in Fig. 1, the input transformation (e.g., DIM [64] vs. Admix [55]) with better transferability tends to bring more obvious visual change to the input image. This inspires us with a new assumption:", + "bbox": [ + 496, + 282, + 890, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assumption 1 Without harming the semantic information, the more diverse the transformed image is, the better transferability the adversarial examples have.", + "bbox": [ + 496, + 414, + 890, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To validate this hypothesis, we utilize the learned perceptual image patch similarity (LPIPS) [79] to evaluate the semantic similarity between the benign samples and the transformed images:", + "bbox": [ + 496, + 468, + 890, + 529 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L P I P S} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) = \\frac {1}{H \\times W} \\sum_ {l} \\sum_ {h, w} \\| \\boldsymbol {z} _ {h, w} ^ {l} - \\hat {\\boldsymbol {z}} _ {h, w} ^ {l} \\| _ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 536, + 890, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z^l$ and $\\hat{z}^l$ are the extracted feature from $l$ -th layer of SqueezeNet [22] with $x$ and $\\hat{x}$ as input, respectively. A smaller LPIPS value indicates better similarity between the two images. The semantic similarity between the raw images and transformed images by DIM, TIM, SIM and Admix are summarized in Tab. 1. As we can see, the similarity decreases when the transferability of attack increases because the transformed images are more diverse. Moreover, as shown in Fig. 1, such diverse images are significantly different from the raw images, which introduces instability when calculating the gradient. Hence, the more powerful input transformation based attack needs to calculate the gradient on multiple transformed images to eliminate the instability. For instance, DIM utilizes single image, SIM adopts 5 images, while the SOTA Admix/SSA takes 15/20 images.", + "bbox": [ + 496, + 583, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The relation between the diversity of transformed images and transferability inspires us to generate more diverse images for gradient calculation so that we can craft more transferable adversarial examples. In this work, we apply various input transformations on different blocks of a single input image to obtain more diverse images, detailed in Sec. 3.2.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4609", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/6d9fdbc49ba45c9030bf689a8068adeb9fbc21f2530851fe76d962ce9072431d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RawVShiftHShiftVFlipHFlipRotateScaleAdd NoiseResizeDCTDropout
", + "bbox": [ + 78, + 89, + 898, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2: The raw and transformed images using various transformations adopted by SIA (See details in Appendix A).", + "bbox": [ + 98, + 176, + 867, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bb320fe443122e66f26d3abbe023f06e06ffd0d2a970384ad4953a260648d04a.jpg", + "image_caption": [ + "Raw Image" + ], + "image_footnote": [], + "bbox": [ + 88, + 215, + 205, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8af069c8090e5d480041e98e6fc262359e226a77e2aebe686a78fc26f34af958.jpg", + "image_caption": [ + "Scale" + ], + "image_footnote": [], + "bbox": [ + 215, + 215, + 331, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1afb90fdd68ced872ce7e7d2f8b370557058966d66e97bdb215a22f0830e0488.jpg", + "image_caption": [ + "Block Scale", + "Figure 2: The randomly sampled raw image and its transformed images by scaling on the full image and $3 \\times 3$ blocks." + ], + "image_footnote": [], + "bbox": [ + 341, + 215, + 457, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Structure Invariant Attack", + "text_level": 1, + "bbox": [ + 76, + 392, + 318, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the limited number of the existing image transformations, we first investigate that how to generate a more diverse image using a single transformation? Without loss of generality, we take scale as an example, which is the base operation of SIM [32] (also a particular case of Admix [55]). As shown in Fig. 2, scaling the image changes its intensity uniformly on all the pixels, resulting in limited diverse images, which is also validated in Tab. 1. To improve the diversity of scaled images, we scale the image with different factors for different parts. Specifically, we randomly split the image into several image blocks and independently scale each image block with different scaled factors. As illustrated in Fig. 2, scaling the image blocks can bring a much more diverse transformed image while humans can still catch the visual information from the image. To explain why such image does not confuse humans, we define the structure of image as follows:", + "bbox": [ + 75, + 415, + 470, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1 (Structure of Image). Given an image $\\mathbf{x}$ which is randomly split into $s \\times s$ blocks, the relative relation between each anchor point is the structure of image, where the anchor point is the center of the image block.", + "bbox": [ + 76, + 680, + 468, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We argue that the structure of image depicts important semantic information for human recognition. For instance, the dog's body should be between its head and tail, while its legs should be under its body. Scaling the image blocks with various factors does not change the structure of image so that the generated image can be correctly recognized by humans as well as deep models.", + "bbox": [ + 75, + 750, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, transforming the image blocks does not harm the recognition but crafts more diverse images, which is of great benefit to improve transferability. To further", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: Structure Invariant Attack", + "text_level": 1, + "bbox": [ + 509, + 222, + 777, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Classifier $\\pmb{f}(\\cdot)$ with the loss function $J$ ; The benign sample $\\pmb{x}$ with ground-truth label $y$ ; The maximum perturbation $\\epsilon$ , number of iterations $T$ and decay factor $\\mu$ ; Splitting number $s$ ; Number of transformed images $N$", + "bbox": [ + 519, + 239, + 866, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: An adversarial example.", + "bbox": [ + 521, + 316, + 743, + 330 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 $\\alpha = \\epsilon /T$ $\\pmb {g}_0 = 0$ $x_0^{adv} = x$", + "2 for $t = 0\\to T - 1$ do" + ], + "bbox": [ + 506, + 330, + 715, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Constructing a set $\\mathcal{X}$ of $N$ transformed images using SIT", + "bbox": [ + 506, + 362, + 854, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Calculating the average gradient on $\\mathcal{X}$", + "bbox": [ + 508, + 392, + 803, + 406 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {g}} _ {t + 1} = \\frac {1}{N} \\sum_ {\\boldsymbol {x} _ {i} \\in \\mathcal {X}} \\nabla_ {\\boldsymbol {x}} J (\\boldsymbol {x} _ {i}, y) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 417, + 867, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 | Updating the momentum:", + "bbox": [ + 506, + 469, + 715, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {t + 1} = \\mu \\boldsymbol {g} _ {t} + \\frac {\\bar {\\boldsymbol {g}} _ {t + 1}}{\\| \\bar {\\boldsymbol {g}} _ {t + 1} \\| _ {1}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 494, + 866, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6 | Updating the adversarial example:", + "bbox": [ + 506, + 539, + 772, + 554 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t + 1} ^ {a d v} = \\operatorname {C l i p} \\left(\\boldsymbol {x} _ {t} ^ {a d v} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\boldsymbol {g} _ {t + 1}\\right), 0, 1\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 566, + 866, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "7 return $x_{T}^{adv}$", + "bbox": [ + 506, + 599, + 604, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "boost the diversity of transformed images, we apply various image transformations to different image blocks, where the adopted transformations are summarized in Tab. 2. To avoid information loss, we add the constraints on some transformations. For instance, the rotation can only rotate the image $180^{\\circ}$ to avoid dropping some pixels. We denote such transformation as structure invariant transformation (SIT).", + "bbox": [ + 496, + 655, + 890, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since SIT can generate more diverse images while maintaining the critical semantic information, we adopt SIT as the input transformation to conduct the adversarial attack, denoted as structure invariant attack (SIA). Instead of directly calculating the gradient on the input image, SIA calculates the gradient on several images transformed by SIT. Note that SIA is generally applicable to any gradient-based attacks. Due to the limited space, we integrate SIA into MIFGSM [10] and summarize the algorithm in Algorithm 1.", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4610", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/20fe0ecb2ad38d5b5fe1ca29c6ec19511b055c4796e46c3c123469f92f574f12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 93, + 669, + 108 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fc8162f0fa6e6f4a85aa9463d728b401433b12fb6cace461ad6b3940560cce77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 113, + 279, + 213 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5cbbdeb4e015eb7e17752df75e066f79f97b196c2cdd173dd34c0bee4ab99cae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 113, + 480, + 213 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59f646129b033fb3d6dbe7bd5ab2ac10b5b314c44699b213dd4fd3ee0eedb616.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 113, + 684, + 214 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dcc9afcb05052c71dc4744e7f72be33a388d4dddf8cef728aa8cce6bfba38f1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 114, + 885, + 213 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d67246c6c6eab4dd9d1b519b6cf39af249e52d27426117538de9e66b0e8bd6e6.jpg", + "image_caption": [ + "Figure 3: Attack success rates $(\\%)$ of eight deep models on the adversarial examples crafted on each model by TIM, DIM, DEM, Admix, SSA, and SIA." + ], + "image_footnote": [], + "bbox": [ + 84, + 220, + 279, + 319 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f54a3fe1438516f6c9521f3a412c32892ae56a0e459311a138fca1f5a321df98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 220, + 480, + 319 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/122fafbe6e052d31c12ec2f3c98b9b7154f783fa392057d75f3c94d88b15b420.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 220, + 684, + 320 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1614697564ff397200f7b4c57ae65a540dfaf361dc40d2567d5a34ba6e9a5ba0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 220, + 885, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. SIT vs. AutoAugment", + "text_level": 1, + "bbox": [ + 76, + 392, + 276, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both SIT and AutoAugment [9] leverage several image transformations to pre-process the images. We summarize the difference as follows:", + "bbox": [ + 75, + 415, + 468, + 459 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SIT improves the adversarial transferability when attacking the model while AutoAugment boosts the model generalization during the training process.", + "- SIT randomly samples the transformation while AutoAugment needs to search for a good policy for each transformation on each dataset.", + "- SIT applies various transformations on different image blocks locally but preserves the global structure. On the contrary, AutoAugment applies two transformations sequentially on the image.", + "- The fine-grained transformations by SIT generate more diverse images than AutoAugment." + ], + "bbox": [ + 76, + 463, + 468, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 659, + 200, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct extensive evaluations on ImageNet dataset to validate the effectiveness of SIA.", + "bbox": [ + 75, + 685, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 76, + 724, + 277, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. To align with previous works [10, 32, 53, 55], we randomly sample 1,000 images pertaining to 1,000 categories from the ILSVRC 2012 validation set [40], which are correctly classified by the adopted models.", + "bbox": [ + 75, + 750, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our proposed SIA with five competitive input transformation based attacks, namely DIM [64], TIM [11], DEM [81], Admix [55], and SSA [36], which are integrated into MI-FGSM [10]. We also integrate the baseline methods with two model-specific approaches, namely LinBP [19] and SGM [60].", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Victim Models. We evaluate the attack performance on two popular model architectures, namely Convolutional Network Works, i.e., ResNet-18 [20], ResNet-101 [20], ResNext-50 [65], DenseNet-121 [23], MobileNet [21], and Transformers, i.e., Vision Transformer (ViT) [12] and Swin Transformer (Swin) [34]. Furthermore, we study several SOTA defense methods, including one adversarial training method, i.e., ensemble adversarially trained model (Incv3_ens) [48], the top-3 submissions in NIPS 2017 defense competition, i.e., high-level representation guided denoiser (HGD) [30], random resizing and padding (R&P) [63] and NIPS-r3 $^1$ , three input pre-processing based defenses, namely FD [35], JPEG [18] and Bit-Red [67], a certified defense, i.e., randomized smoothing (RS) [8] and a deep denoiser, i.e., neural representation purifier (NRP) [39].", + "bbox": [ + 496, + 392, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Settings. We follow MI-FGSM [10] with the perturbation budget $\\epsilon = 16$ , number of iteration $T = 10$ , step size $\\alpha = \\epsilon / T = 1.6$ and decay factor $\\mu = 1$ . DIM [64] adopts the transformation probability of 0.5 and TIM [11] utilizes the Gaussian kernel with the size of $7 \\times 7$ . DEM [81] uses the resize ratios: [1.14, 1.27, 1.4, 1.53, 1.66]. Admix admixes 3 images from other categories with the strength of 0.2 and 5 scaled images for each admixed image. SSA [36] sets the turning factor as 0.5 and the standard deviation as $\\epsilon$ . SIA sets the splitting number $s = 3$ and number of transformed images for gradient calculation $N = 20$ .", + "bbox": [ + 496, + 619, + 890, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Attacking a Single Model", + "text_level": 1, + "bbox": [ + 500, + 795, + 732, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To validate the effectiveness of the proposed SIA, we first compare SIA with five SOTA input transformation based attacks, namely DIM, TIM, DEM, Admix and SSA.", + "bbox": [ + 496, + 819, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://github.com/anlthms/nips-2017/tree/master/mmd", + "bbox": [ + 500, + 875, + 857, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4611", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5a96ed0444eb2a1cd0d72c197d5df10483bbf76d3b5ae1e31882ce320d0f3513.jpg", + "image_caption": [ + "Figure 4: Attack success rates $(\\%)$ of eight models on the adversarial examples generated on ResNet-18 when integrating TIM, DIM, DEM, Admix, SSA, and SIA into SGM, respectively." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 460, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We generate the adversarial examples on a single model and test them on the other models. The attack success rates, i.e., the misclassification rates of the victim model on the crafted adversarial examples, are summarized in Fig. 3.", + "bbox": [ + 75, + 428, + 468, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From the figure, we can observe that all the attackers can achieve the attack success rate of $100.0\\%$ or near $100.0\\%$ , showing that the input transformation based attacks do not degrade the white-box attack performance. As for the black-box performance, TIM exhibits the poorest transferability on these normally trained models. Admix consistently achieves better transferability than DIM and DEM on CNN-based models. Surprisingly, DIM achieves even better transferability than Admix when generating adversarial examples on transformer-based models. SSA exhibits the best transferability among the baselines in most cases. Also, even for the transformer-based models, the adversarial examples generated on Swin show the poorest transferability on ViT than other CNN-based models. Hence, we argue that it is necessary to evaluate the effectiveness of transfer-based attacks on both CNN-based and transformer-based models. We need an in-depth analysis of the transferability among the emerging transformer-based models. Compared with the baselines, SIA consistently performs much better than the best baselines on all eight models with different architectures. In particular, SIA outperforms the winner-up method with a clear margin of $14.3\\%$ on average and achieves an attack success rate higher than the best baseline of at least $2.8\\%$ on all the models. Such consistent and superior performance demonstrates that the proposed SIA is general to various model architectures (either CNN or transformer) to boost transferability effectively.", + "bbox": [ + 75, + 493, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc97c92d2c10f857ec943002cdca9f8a15070a36900b2413f607548f164c8cde.jpg", + "image_caption": [ + "Figure 5: Attack success rates $(\\%)$ of eight models on the adversarial examples generated on ResNet-18 when integrating TIM, DIM, DEM, Admix, SSA, and SIA into LinBP, respectively." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 883, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Integration to model-specific approaches", + "text_level": 1, + "bbox": [ + 500, + 421, + 846, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further verify the scalability of the proposed SIA, we integrate existing input transformation-based attacks into two model-specific approaches, namely SGM and LinBP. The adversarial examples are generated on ResNet-18 and test them on the other models.", + "bbox": [ + 496, + 444, + 890, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As depicted in Fig. 4 and Fig. 5, SGM and LinBP can significantly boost the adversarial transferability of vallina input transformation-based attacks. Compared with the baselines, SIA still achieves much better transferability, which has been boosted by an average margin of $3.3\\%$ and $3.6\\%$ when integrated into SGM and LinBP, respectively. These superior results validate the generality of SIA to various transfer-based attacks and underscore the potential of SIA in augmenting adversarial transferability through the fusion of different strategies.", + "bbox": [ + 496, + 520, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Attacking Ensemble Models", + "text_level": 1, + "bbox": [ + 500, + 680, + 751, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Liu et al. [33] have shown that attacking ensemble models can effectively improve transferability. As shown in Sec. 4.2, the adversarial examples often exhibit poorer transferability across the CNN-based models and transformer-based models than being transferred among the CNN-based models. Hence, we generate adversarial examples on ensemble CNN-based and ensemble transformer-based models and test them on the remaining models to evaluate SIA when attacking ensemble models.", + "bbox": [ + 496, + 704, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 6, when attacking ensemble models, the adversarial examples generated by all attacks exhibit better transferability than that crafted on a single model, showing the excellent compatibility of all the input trans", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4612", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/37dd8323c287535f2d89323967e6331264bca7bdfce575d7a8c3ac27a928c6a5.jpg", + "image_caption": [ + "Figure 6: Attack success rates (\\%) of eight models on the adversarial examples crafted on ensemble models by TIM, DIM, DEM, Admix, SSA, and SIA, respectively. We generate adversarial examples on the CNN-based models and test them on the transformer-based models, and vice versa." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 460, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formation based attacks with such a setting. When generating the adversarial examples on ViT and Swin, DIM and DEM achieve better transferability than Admix and SSA, highlighting the difference between crafting adversarial examples on different architectures and the necessity to evaluate the transferability on these models. On all eight models, SIA achieves the attack success rate of at least $88.3\\%$ . It outperforms the winner-up approach with a clear margin of $6.5\\%$ , showing its superior effectiveness in generating transferable adversarial examples. In particular, SIA achieves $94.9\\%$ and $99.3\\%$ attack success rates on ViT and Swin, respectively, when the adversarial examples are generated on CNN-based models without access to the attention module in the transformer. This further supports our motivation that improving the diversity of transformed images can significantly boost transferability, even on models with completely different architectures.", + "bbox": [ + 75, + 439, + 468, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Attacking Defense Methods", + "text_level": 1, + "bbox": [ + 76, + 709, + 326, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SIA has achieved superior attack performance on eight normally trained models with different architectures when attacking single model as well as ensemble models. Recently, several defenses have been proposed to mitigate the threat of adversarial examples on ImageNet dataset. To validate the effectiveness of these defenses, we adopt the adversarial examples generated on these eight models simultaneously to attack the defense models, including Inc-v3ens, HGD, R&P, NIPS-r3, FD, JPEG, Bit-Red, RS and NRP.", + "bbox": [ + 75, + 732, + 467, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The attack results on these defense methods are summarized in Fig. 7. Overall, DIM is on par with TIM on", + "bbox": [ + 76, + 869, + 467, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c00f382c99a1449f97cc2353af5ef232d0f76e7c9a0e397eee1ab0e5e7b8e1fc.jpg", + "image_caption": [ + "Figure 7: Attack success rates $(\\%)$ of various defense methods on the adversarial examples generated by TIM, DIM, DEM, Admix, SSA, and SIA under ensemble model setting. The adversarial examples are generated on the eight models simultaneously." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 883, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "these defense methods while Admix can consistently outperform the other four baselines. SIA can consistently achieve better transferability than the baselines. In particular, SIA achieves the attack success rate of $78.2\\%$ on the certified defense method (i.e., RS), and SIA can achieve the attack success rate of at least $89.9\\%$ on the other eight defense methods, including the powerful denoising method NRP. Such high attack performance indicates the insufficiency of existing defense methods and raises a new security issue to designing more robust deep learning models.", + "bbox": [ + 496, + 439, + 890, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 603, + 663, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further gain insights on the superior attack performance achieved by SIA, we conduct a series of ablation studies to validate that all the transformations and splitting the image into blocks can help improve transferability. We generate the adversarial examples on ResNet-18 and test them on the other seven models for all the experiments.", + "bbox": [ + 496, + 627, + 890, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Are all the transformations beneficial for boosting the transerability? There are ten different input transformations that SIA might apply to each image block. To investigate whether each transformation is of benefit to generate more transferable adversarial examples, we first adopt one or two transformations to implement SIA. The results are summarized in Fig. 8. The average attack success rate of MI-FGSM is $59.7\\%$ . On the diagonal line of Fig. 8, we only adopt a single transformation for each block and the lowest average attack success rate is $67.3\\%$ by adopting DCT, which outperforms MI-FGSM with a margin of $7.6\\%$ . This demonstrates the high effectiveness of SIA and", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4613", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a37f2683e2b6b054e435ad8bf3a6a7fc8c903ac120fd644e56c9277b62fafa6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SIA-VShift-HShift-VFlip-HFlip-Rotate-Sclae-Add Noise-Resize-DCT-Dropout
92.189.790.190.188.488.390.190.690.290.190.7
", + "bbox": [ + 112, + 88, + 857, + 135 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: The average attack success rates of adversarial examples crafted by SIA and SIA without a single transformation. The adversaries are generated on ResNet-18 and tested on the other seven deep models. - indicates removing such transformation.", + "bbox": [ + 75, + 145, + 893, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19e5b9ac816739d615e5847af4b383f2ee4ef7eb81f0ab022e00b030afb6650e.jpg", + "image_caption": [ + "Figure 8: The average attack success rates $(\\%)$ on the adversarial examples generated by SIA using one (diagonal) or two (other) transformations." + ], + "image_footnote": [], + "bbox": [ + 84, + 204, + 467, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "its generality to various transformations. We can also observe that the transferability can be further improved when we combine any two transformations, showing the advantage of combining these transformations. To further validate the necessity of each transformation, we generate adversarial examples by SIA without each transformation and summarize the results in Tab. 3. Since removing a single transformation does not significantly decrease the diversity of the transformation, these attacks achieve similar attack performance. However, no matter which transformation is removed, the transferability will be decreased, supporting that each transformation is of benefit to generate more transferable adversarial examples.", + "bbox": [ + 75, + 520, + 468, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Are the image blocks beneficial for boosting transferability? We have shown that the abundant transformations can effectively improve transferability. Here we further explore whether it is useful to apply these transformations to the image block instead of the raw image. For comparison, we randomly apply the transformation to the raw image $s \\times s$ times, denoted as $\\mathrm{SIA_r}$ . The results are reported in Fig. 9. We can observe that $\\mathrm{SIA_r}$ exhibits better transferability than MI-FGSM, which also validates our motivation that improving the diversity of transformed images can generate more transferable adversarial examples. Our SIA consistently performs better than $\\mathrm{SIA_r}$ , indicating that applying", + "bbox": [ + 75, + 719, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa9c8322747df196024fc159650aaad0713635eaa7a62acf81b178ded62941fc.jpg", + "image_caption": [ + "Figure 9: The average attack success rates $(\\%)$ of adversarial examples generated by SIA and $\\mathrm{SIA_r}$ (randomly transform the image without the block partition)." + ], + "image_footnote": [], + "bbox": [ + 506, + 204, + 885, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the transformation onto each image block is of great benefit to improve transferability.", + "bbox": [ + 498, + 517, + 890, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Parameter Studies", + "text_level": 1, + "bbox": [ + 500, + 559, + 679, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this subsection, we conduct parameter studies to explore the impact of two hyper-parameters, namely the number of blocks $s$ and the number of transformed images for gradient calculation $N$ . All the adversarial examples are generated on ResNet-18 and tested on the other seven models in the black-box setting.", + "bbox": [ + 496, + 582, + 890, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the number of blocks $s$ . The number of blocks determines how diverse the transformed images are, which can influence the attack performance. To find a good value for $s$ , we conduct SIA with $s$ from 1 to 5 and summarize the attack results in Fig. 10. When $s \\leq 3$ , increasing the value of $s$ leads to better diversity, which can improve the attack performance. However, when we continually increase the value of $s$ , the diversity is increased, but it also introduces more variance to the gradient, decaying the attack performance slightly. Hence, we adopt $s = 3$ to balance the diversity of images and variance of the gradient for better attack performance.", + "bbox": [ + 496, + 672, + 892, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the number of images for gradient calculation $N$ . SIA calculates the average gradient on $N$ images to eliminate the variance introduced by the transformation. To de", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4614", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f0f3974934d2ffa05e95e725d521c3807bf3446061ac766eeef25ab5cb64921.jpg", + "image_caption": [ + "Figure 10: Attack success rates $(\\%)$ of adversarial examples generated by SIA with various number of blocks $s$" + ], + "image_footnote": [], + "bbox": [ + 86, + 94, + 467, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "termine a good value for $N$ , we evaluate SIA with $N$ from 1 to 30 and report the attack success rates in Fig. 11. As we can observe, when $N = 1$ , SIA introduces massive variance into the gradient and achieves the lowest performance on all models. When we increase the value of $N$ , the variance among the gradients can be eliminated, and SIA achieves better transferability before $N = 20$ . When $N > 20$ , increasing $N$ can only introduce computation cost without performance improvement. To balance the attack performance and computation cost, we adopt $N = 20$ in our experiments.", + "bbox": [ + 75, + 404, + 472, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 609, + 194, + 626 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we find that the existing input transformation based attack with better transferability often generates more diverse transformed images. Based on this finding, we design a new image transformation, called structure invariant transformation (SIT), which splits the image into several blocks and randomly transforms each image block independently. With such image transformation, we propose a novel input transformation based attack, dubbed structure invariant attack (SIA), which calculates the average gradient on several transformed images by SIT to update the perturbation. Extensive evaluations demonstrate that SIA can achieve remarkably better transferability than the existing SOTA attacks. In our opinion, SIA provides a new direction by applying the transformation onto the image block to effectively boost transferability, which sheds new light on generating more transferable adversarial examples with fine-grained transformations.", + "bbox": [ + 75, + 643, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/750e2b03d21a661c6217aa6a5326ca38cfd9f1cf0709e405b1b53368430649ef.jpg", + "image_caption": [ + "Figure 11: Attack success rates $(\\%)$ of adversarial examples generated by SIA with various number of images $N$" + ], + "image_footnote": [], + "bbox": [ + 509, + 94, + 883, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 388, + 596, + 404 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Moustafa Alzantot, Yash Sharma, Ahmed Elgohary, BoJhang Ho, Mani B. Srivastava, and Kai-Wei Chang. Generating Natural Language Adversarial Examples. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 2890-2896, 2018. 2", + "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square Attack: A Query-Efficient Black-Box Adversarial Attack via Random Search. In Proceedings of the European Conference on Computer Vision, pages 484–501, 2020. 2", + "[3] Anish Athalye, Nicholas Carlini, and David A. Wagner. Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. In Proceedings of the International Conference on Machine Learning, pages 274–283, 2018. 2, 3", + "[4] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models. In Proceedings of the International Conference on Learning Representations, 2018. 1, 2", + "[5] Nicholas Carlini and David A. Wagner. Audio Adversarial Examples: Targeted Attacks on Speech-to-Text. In Proceedings of the IEEE Symposium on Security and Privacy (Workshops), pages 1-7, 2018. 2", + "[6] Minhao Cheng, Thong Le, Pin-Yu Chen, Huan Zhang, Jinfeng Yi, and Cho-Jui Hsieh. Query-Efficient Hard-label Black-box Attack: An Optimization-based Approach. In Proceedings of the International Conference on Learning Representations, 2019. 2", + "[7] Shuyu Cheng, Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Improving Black-box Adversarial Attacks with a Transfer-based Prior. In Proceedings of the Advances in Neural Information Processing Systems, pages 10932-10942, 2019. 1" + ], + "bbox": [ + 508, + 415, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4615", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[8] Jeremy M. Cohen, Elan Rosenfeld, and J. Zico Kolter. Certified Adversarial Robustness via Randomized Smoothing. In Proceedings of the International Conference on Machine Learning, pages 1310-1320, 2019. 3, 5", + "[9] Ekin D. Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, and Quoc V. Le. AutoAugment: Learning Augmentation Strategies From Data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 3, 5", + "[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting Adversarial Attacks With Momentum. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 1, 2, 4, 5", + "[11] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading Defenses to Transferable Adversarial Examples by Translation-Invariant Attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 1, 2, 5", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In Proceedings of the International Conference on Learning Representations, 2021. 1, 5", + "[13] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust Physical-World Attacks on Deep Learning Visual Classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1625–1634, 2018. 1", + "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we Ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1", + "[15] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and Harnessing Adversarial Examples. In Proceedings of the International Conference on Learning Representations, 2015. 1, 2, 3", + "[16] Sven Gowal, Krishnamurthy Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timothy Arthur Mann, and Pushmeet Kohli. Scalable Verified Training for Provably Robust Image Classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4841-4850, 2019. 3", + "[17] Chuan Guo, Jacob R. Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Q. Weinberger. Simple Black-box Adversarial Attacks. In Proceedings of the International Conference on Machine Learning, pages 2484-2493, 2019. 2", + "[18] Chuan Guo, Mayank Rana, Moustapha Cissé, and Laurens van der Maaten. Countering Adversarial Images Using Input Transformations. In Proceedings of the International Conference on Learning Representations, 2018. 5", + "[19] Yiwen Guo, Qizhang Li, and Hao Chen. Backpropagating linearly improves transferability of adversarial examples." + ], + "bbox": [ + 80, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Advances in neural information processing systems, 33:85-95, 2020. 3, 5", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1, 3, 5", + "[21] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv preprint arXiv:1704.04861, 2017. 5", + "[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-Eccitation Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 3", + "[23] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely Connected Convolutional Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2261-2269, 2017. 1, 5", + "[24] Andrew Ilyas, Logan Engstrom, Anish Athalye, and Jessy Lin. Black-box Adversarial Attacks with Limited Queries and Information. In Proceedings of the International Conference on Machine Learning, pages 2142-2151, 2018. 1, 2", + "[25] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Proceedings of the Advances in Neural Information Processing Systems, pages 1106-1114, 2012. 1, 3", + "[26] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial Examples in the Physical World. In Proceedings of the International Conference on Learning Representations (Workshops), 2017. 1, 2", + "[27] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. QEBA: Query-Efficient Boundary-Based Blackbox Attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1218-1227, 2020. 1, 2", + "[28] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan L. Yuille. Learning Transferable Adversarial Examples via Ghost Networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11458-11465, 2020. 1, 3", + "[29] Yandong Li, Lijun Li, Liqiang Wang, Tong Zhang, and Boqing Gong. NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks. In Proceedings of the International Conference on Machine Learning, pages 3866–3876, 2019. 2", + "[30] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense Against Adversarial Attacks Using High-Level Representation Guided Denoiser. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2018. 3, 5", + "[31] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4616", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Daan Wierstra. Continuous Control with Deep Reinforcement Learning. In Proceedings of the International Conference on Learning Representations, 2016. 1", + "[32] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov Accelerated Gradient and Scale Invariance for Adversarial Attacks. In Proceedings of the International Conference on Learning Representations, 2020. 1, 2, 4, 5", + "[33] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into Transferable Adversarial Examples and Blackbox Attacks. In Proceedings of the International Conference on Learning Representations, 2017. 1, 2, 6", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9992-10002, 2021. 5", + "[35] Zihao Liu, Qi Liu, Tao Liu, Nuo Xu, Xue Lin, Yanzhi Wang, and Wujie Wen. Feature Distillation: DNN-Oriented JPEG Compression Against Adversarial Examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 860-868, 2019. 5", + "[36] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xi-anglong Liu, Jian Zhang, and Jingkuan Song. Frequency Domain Model Augmentation for Adversarial Attack. In Proceedings of the European Conference on Computer Vision, pages 549–566, 2022. 1, 2, 5", + "[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards Deep Learning Models Resistant to Adversarial Attacks. In Proceedings of the International Conference on Learning Representations, 2018. 1, 2", + "[38] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2574-2582, 2016. 1, 2", + "[39] Muzammal Naseer, Salman H. Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A Self-supervised Approach for Adversarial Robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 259-268, 2020. 3, 5", + "[40] Russakovsky Olga, Deng Jia, Su Hao, Krause Jonathan, Satheesh Sanjeev, Ma Sean, hUANG Zhiheng, Karpathy Andrej, Khosla Adiya, and Bernstein Michael et al. Imagenet large scale visual recognition challenge. In International Journal of Computer Vision, pages 211-252, 2015. 5", + "[41] Florian Schroff, Dmitry Kalenichenko, and James Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 1", + "[42] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K. Reiter. Accessorize to a Crime: Real and Stealthy Attacks on State-of-the-Art Face Recognition. In Proceedings of the ACM Conference on Computer and Communications Security, pages 1528-1540, 2016. 1" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Luchuan Song, Zheng Fang, Xiaodan Li, Xiaoyi Dong, Zhenchao Jin, Yuefeng Chen, and Siwei Lyu. Adaptive face forgery detection in cross domain. In European Conference on Computer Vision, pages 467-484. Springer, 2022. 1", + "[44] Luchuan Song, Xiaodan Li, Zheng Fang, Zhenchao Jin, YueFeng Chen, and Chenliang Xu. Face forgery detection via symmetric transformer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4102-4111, 2022. 1", + "[45] Luchuan Song, Bin Liu, Guojun Yin, Xiaoyi Dong, Yufei Zhang, and Jia-Xuan Bai. Tacr-net: editing on deep video and voice portraits. In Proceedings of the 29th ACM International Conference on Multimedia, pages 478-486, 2021. 1", + "[46] Nitish Srivastava, Geoffrey E. Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a Simple Way to prevent Neural Networks from Overfitting. The Journal of Machine Learning Research, 15(1):1929-1958, 2014. 3", + "[47] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing Properties of Neural Networks. In Proceedings of the International Conference on Learning Representations, 2014. 1, 2", + "[48] Florian Tramèr, Alexey Kurakin, Nicolas Papernot, Ian J. Goodfellow, Dan Boneh, and Patrick D. McDaniel. Ensemble Adversarial Training: Attacks and Defenses. In Proceedings of the International Conference on Learning Representations, 2018. 3, 5", + "[49] Jonathan Uesato, Brendan O'Donoghue, Pushmeet Kohli, and Aïron van den Oord. Adversarial Risk and the Dangers of Evaluating Against Weak Attacks. In Proceedings of the International Conference on Machine Learning, pages 5032-5041, 2018. 2", + "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Proceedings of the Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 1", + "[51] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. CosFace: Large Margin Cosine Loss for Deep Face Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5265-5274, 2018. 1", + "[52] Xiaosen Wang, Jin Hao, Yichen Yang, and Kun He. Natural Language Adversarial Defense through Synonym Encoding. In Proceedings of the Conference on Uncertainty in Artificial Intelligence, pages 823-833, 2021. 2", + "[53] Xiaosen Wang and Kun He. Enhancing the Transferability of Adversarial Attacks Through Variance Tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 1, 2, 5", + "[54] Xiaosen Wang, Kun He, Chuanbiao Song, Liwei Wang, and John E Hopcroft. AT-GAN: An Adversarial Generator Model for Non-constrained Adversarial Examples. arXiv preprint arXiv:1904.07793, 2019. 1", + "[55] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the Transferability of Adversarial At" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4617", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16138-16147, 2021. 1, 2, 3, 4, 5", + "[56] Xiaosen Wang, Jiadong Lin, Han Hu, Jingdong Wang, and Kun He. Boosting Adversarial Transferability through Enhanced Momentum. In Proceedings of the British Machine Vision Conference, page 272, 2021. 1, 2", + "[57] Xiaosen Wang, Zeliang Zhang, Kangheng Tong, Dihong Gong, Kun He, Zhifeng Li, and Wei Liu. Triangle Attack: A Query-Efficient Decision-Based Adversarial Attack. In Proceedings of the European Conference on Computer Vision, pages 156-174, 2022. 1, 2", + "[58] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable Adversarial Attacks for Image and Video Object Detection. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 954-960, 2019. 2", + "[59] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is Better than Free: Revisiting Adversarial Training. In Proceedings of the International Conference on Learning Representations, 2020. 3", + "[60] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip Connections Matter: On the Transferability of Adversarial Examples Generated with ResNets. In Proceedings of the International Conference on Learning Representations, 2020. 1, 3, 5", + "[61] Weibin Wu, Yuxin Su, Xixian Chen, Shenglin Zhao, Irwin King, Michael R. Lyu, and Yu-Wing Tai. Boosting the Transferability of Adversarial Samples via Attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1158-1167, 2020. 1, 2", + "[62] Weibin Wu, Yuxin Su, Michael R Lyu, and Irwin King. Improving the Transferability of Adversarial Samples with Adversarial Transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9024-9033, 2021. 2", + "[63] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan L. Yuille. Mitigating Adversarial Effects Through Randomization. In Proceedings of the International Conference on Learning Representations, 2018. 3, 5", + "[64] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving Transferability of Adversarial Examples With Input Diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 1, 2, 3, 5", + "[65] Saining Xie, Ross B. Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated Residual Transformations for Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 5", + "[66] Yifeng Xiong, Jiadong Lin, Min Zhang, John E. Hopcroft, and Kun He. Stochastic Variance Reduced Ensemble Adversarial Attack for Boosting the Adversarial Transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14972, 2022. 1, 2", + "[67] Weilin Xu, David Evans, and Yanjun Qi. Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the Network and Distributed System Security Symposium, 2018. 5", + "[68] Bo Yang, Hengwei Zhang, Zheming Li, Yuchen Zhang, Kaiyong Xu, and Jindong Wang. Adversarial example generation with adbelief optimizer and crop invariance. Applied Intelligence, 53(2):2332-2347, 2023. 1", + "[69] Yichen Yang, Xiaosen Wang, and Kun He. Robust Textual Embedding against Word-level Adversarial Attacks. pages 2214-2224, 2022. 2", + "[70] Zhuolin Yang, Bo Li, Pin-Yu Chen, and Dawn Song. Characterizing Audio Adversarial Examples Using Temporal Dependency. In Proceedings of the International Conference on Learning Representations, 2019. 2", + "[71] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural Color Fool: Towards Boosting Black-box Unrestricted Attacks. In Proceedings of the Advances in Neural Information Processing Systems, 2022. 1", + "[72] Zheng Yuan, Jie Zhang, and Shiguang Shan. Adaptive image transformations for transfer-based adversarial attack. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1", + "[73] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6022-6031, 2019. 3", + "[74] Huan Zhang, Hongge Chen, Chaowei Xiao, Sven Gowal, Robert Stanforth, Bo Li, Duane S. Boning, and Cho-Jui Hsieh. Towards Stable and Efficient Training of Verifiably Robust Neural Networks. In Proceedings of the International Conference on Learning Representations, 2020. 3", + "[75] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. Mixup: Beyond Empirical Risk Minimization. In Proceedings of the International Conference on Learning Representations, 2018. 3", + "[76] Jianping Zhang, Jen-tse Huang, Wenxuan Wang, Yichen Li, Weibin Wu, Xiaosen Wang, Yuxin Su, and Michael R. Lyu. Improving the Transferability of Adversarial Samples by Path-Augmented Method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8173-8182, 2023. 1", + "[77] Qilong Zhang, Xiaodan Li, Yuefeng Chen, Jingkuan Song, Lianli Gao, Yuan He, and Hui Xue. Beyond ImageNet Attack: Towards Crafting Adversarial Examples for Black-box Domains. In Proceedings of the International Conference on Learning Representations, 2022. 1", + "[78] Qilong Zhang, Chaoning Zhang, Chaoqun Li, Jingkuan Song, Lianli Gao, and Heng Tao Shen. Practical No-box Adversarial Attacks with Training-free Hybrid Image Transformation. arXiv preprint arXiv:2203.04607, 2022. 1", + "[79] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 3", + "[80] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable Ad" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4618", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "versarial Perturbations. In Proceedings of the European Conference on Computer Vision, pages 471-486, 2018. 1, 2", + "[81] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the Transferability of Adversarial Examples with Resized-Diverse-Inputs, Diversity-Ensemble and Region Fitting. In Proceedings of the European Conference on Computer Vision, pages 563–579, 2020. 2, 5" + ], + "bbox": [ + 78, + 92, + 467, + 191 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "4619", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/Structure Invariant Transformation for better Adversarial Transferability/303ed349-48a8-4393-8e2e-88000a7d20c9_model.json b/2023/Structure Invariant Transformation for better Adversarial Transferability/303ed349-48a8-4393-8e2e-88000a7d20c9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bf1bdb911d6041560b6211ea1604f85db436a70e --- /dev/null +++ b/2023/Structure Invariant Transformation for better Adversarial Transferability/303ed349-48a8-4393-8e2e-88000a7d20c9_model.json @@ -0,0 +1,3020 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.131, + 0.857, + 0.154 + ], + "angle": 0, + "content": "Structure Invariant Transformation for better Adversarial Transferability" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.182, + 0.363, + 0.217 + ], + "angle": 0, + "content": "Xiaosen Wang \nHuawei Singularity Security Lab" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.221, + 0.317, + 0.233 + ], + "angle": 0, + "content": "xiaosen@hust.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.182, + 0.54, + 0.216 + ], + "angle": 0, + "content": "Zeliang Zhang HUST" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.221, + 0.562, + 0.234 + ], + "angle": 0, + "content": "hust0426@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.599, + 0.182, + 0.869, + 0.217 + ], + "angle": 0, + "content": "Jianping Zhang Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.633, + 0.221, + 0.837, + 0.234 + ], + "angle": 0, + "content": "jpzhang@cse.cuhk.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.71 + ], + "angle": 0, + "content": "Given the severe vulnerability of Deep Neural Networks (DNNs) against adversarial examples, there is an urgent need for an effective adversarial attack to identify the deficiencies of DNNs in security-sensitive applications. As one of the prevalent black-box adversarial attacks, the existing transfer-based attacks still cannot achieve comparable performance with the white-box attacks. Among these, input transformation based attacks have shown remarkable effectiveness in boosting transferability. In this work, we find that the existing input transformation based attacks transform the input image globally, resulting in limited diversity of the transformed images. We postulate that the more diverse transformed images result in better transferability. Thus, we investigate how to locally apply various transformations onto the input image to improve such diversity while preserving the structure of image. To this end, we propose a novel input transformation based attack, called Structure Invariant Transformation (SIA), which applies a random image transformation onto each image block to craft a set of diverse images for gradient calculation. Extensive experiments on the standard ImageNet dataset demonstrate that SIA exhibits much better transferability than the existing SOTA input transformation based attacks on CNN-based and transformer-based models, showing its generality and superiority in boosting transferability. Code is available at https://github.com/xiaosen-wang/SIT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.74, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "With the unprecedented progress of Deep Neural Networks (DNNs) [25, 20, 23, 50, 12], they have been deployed in many security-sensitive applications, such as face recognition [41, 51, 43, 44], autonomous driving [14, 31], etc. On the other hand, recent works have found that DNNs are vulnerable to adversarial examples [47, 15], which mislead the deep models with imperceptible perturbations. This brings a huge threat to the real-world applications [42, 13, 45, 57, 71, 78] and makes it imperative for an" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.303, + 0.598, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.373, + 0.592, + 0.387 + ], + "angle": 0, + "content": "Raw Image" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.269, + 0.693, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.339, + 0.665, + 0.35 + ], + "angle": 0, + "content": "DIM" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.353, + 0.693, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.422, + 0.672, + 0.434 + ], + "angle": 0, + "content": "Admix" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.269, + 0.786, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.338, + 0.758, + 0.349 + ], + "angle": 0, + "content": "TIM" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.353, + 0.786, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.422, + 0.758, + 0.434 + ], + "angle": 0, + "content": "SSA" + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.269, + 0.88, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.339, + 0.851, + 0.349 + ], + "angle": 0, + "content": "SIM" + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.353, + 0.88, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.422, + 0.871, + 0.435 + ], + "angle": 0, + "content": "SIA (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.525 + ], + "angle": 0, + "content": "Figure 1: The raw image and its transformed images by DIM (resize factor of 0.8) [64], TIM (translated by 15 pixels) [11], SIM (scale factor of 0.5), Admix (admix strength of 0.2 and scale factor of 0.5) [55], SSA (turning factor of 0.5) [36] and our proposed SIA \\((3\\times 3\\) blocks)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.598 + ], + "angle": 0, + "content": "effective attack to identify the deficiencies of DNNs when we robustify the deep models or deploy them for commercial applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.893, + 0.78 + ], + "angle": 0, + "content": "Existing adversarial attacks usually fall into two categories: white-box attacks [15, 38, 37, 26, 54] can fetch any information of the target model, including (hyper-)parameters, gradient, architecture, while black-box attacks [24, 7, 4, 27, 33, 77] are only allowed limited access to the target model. One of the significant properties of adversarial examples is their transferability [64, 10, 53, 76], in which the adversarial examples generated on one model can still mislead other models, making it possible to attack the real-world applications in the black-box setting. However, existing adversarial attacks [26, 37] often exhibit superior white-box attack performance but poor transferability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To craft more transferable adversarial examples, various techniques have been proposed, such as momentum-based methods [10, 32, 53, 56], input transformations [64, 11, 32, 68, 72], ensemble attacks [33, 66, 36], advanced objective functions [61, 80] and model-specific approaches [28, 60]. Among which, input transformations (e.g., random resizing and padding [64], translation [11], scale [32] admix [55], etc.) that transform the image before gradient calculation," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4607" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "have achieved superior transferability and attracted broad attention. Nevertheless, we find that the existing input transformation based attacks, even the SOTA attack Admix [55] and SSA [36], transform the image globally without changing the local relationship among objects in the input image. We argue that the diversity of such transformed images is still not enough, leading to limited transferability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.394 + ], + "angle": 0, + "content": "In this work, we postulate and empirically validate that the more diverse transformed images lead to better transferability. Based on this observation, instead of applying a single transformation on the input image, we apply different transformations locally on different parts of the image to enhance the diversity of transformed images. As shown in Fig. 1, such transformation can bring much more difference to the generated image compared with the raw image but still preserve the global structure of the object. Based on this transformation, we propose a novel input transformation based attack called structure invariant attack (SIA), which utilizes the gradient of these transformed images to update the adversarial examples for better transferability." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.395, + 0.4, + 0.408 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.412, + 0.468, + 0.47 + ], + "angle": 0, + "content": "- We empirically validate that the high diversity of transformed images is beneficial to improve transferability, which sheds new light on how to design new input transformations for more transferable adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.472, + 0.468, + 0.531 + ], + "angle": 0, + "content": "- We design a new image transformation method, which locally applies different transformations on different parts of the image to generate more diverse images but preserve its global structure." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.533, + 0.468, + 0.591 + ], + "angle": 0, + "content": "- Based on the proposed image transformation, we devise a new input transformation based attack, called structure invariant attack (SIA), to generate more transferable adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.593, + 0.468, + 0.653 + ], + "angle": 0, + "content": "- Extensive experiments on ImageNet dataset demonstrate that SIA outperforms the baselines with a clear margin on CNN-based as well as transformer-based models, showing its superiority and generality." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.412, + 0.468, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.665, + 0.218, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Adversarial examples [47] have brought an impressive threat to the DNN-enabled applications, such as computer vision [15, 58, 3], natural language processing [1, 52, 69], speech recognition [5, 70], etc. To identify the vulnerability of DNNs, various attack methods have been proposed recently, such as gradient-based attacks [15, 38, 26, 37], score-based attacks [24, 49, 17, 29, 2], decision-based attacks [4, 6, 27, 57] and transfer-based attacks [33, 10, 64, 32, 55]. Among these, transfer-based attacks do not access any information of the target model, making it applicable to attack any model in the physical world. In this work, we focus on generating more transferable adversarial examples and briefly introduce the existing transfer-based attacks, the corresponding defense methods, and data augmentations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.684, + 0.105 + ], + "angle": 0, + "content": "2.1. Adversarial Attack" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.265 + ], + "angle": 0, + "content": "As the first gradient-based attack, Fast Gradient Sign Method (FGSM) [15] adds the perturbation in the gradient direction to the benign sample, leading to high attack efficiency but limited performance. Later, I-FGSM [26] extends FGSM into an iterative version, which achieves much better white-box attack performance but poor transferability. Given the high efficiency and effectiveness of I-FGSM, numerous transfer-based attacks are proposed to boost transferability based on I-FGSM to attack the deep model in the black-box setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Momentum-based methods. MI-FGSM [10] introduces momentum into I-FGSM to stabilize the optimization direction and escape local maxima. NI-FGSM [32] adopts Nesterov Accelerated Gradient to accumulate the momentum, which achieves better transferability. Variance tuning [53] adopts the gradient variance of the previous iteration to tune the current gradient in MI-FGSM and NI-FGSM, significantly improving transferability. EMIFGSM [56] enhances the momentum by accumulating several samples' gradients in the previous gradient's direction to further stabilize the optimization direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Input transformation based attacks. DIM [64] is the first input transformation based attack, which adds padding to a randomly resized image for fixed size before gradient calculation. TIM [11] optimizes the perturbation over an ensemble of translated images, which is further approximated by convolving the gradient at the untranslated image with a pre-defined kernel. SIM [32] scales the images with different scale factors for gradient calculation. DEM [81] averages the gradient on several diverse transformed images similar to DIM with various resizing factors. Admix [55] calculates the gradient on the input image admixed with a small portion of each add-in image from other categories while using the original label of the input. Wu et al. [62] train an adversarial transformation network to destroy the adversarial perturbation and require the synthesized adversarial examples resistant to such transformations. SSA [36] adds Gaussian noise and randomly masks the image in the frequency domain to transform the input image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Ensemble attacks. Liu et al. [33] first found that the adversarial examples generated on multiple models denoted as ensemble attack, are more transferable. Recently, Xiong et al. [66] reduce the gradient variance between various models to boost the ensemble attack." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Advanced objective functions. The above attacks often take the cross-entropy loss as the objective function. Researchers also find that some regularizers are beneficial to boost transferability. For instance, Zhou et al. [80] additionally maximize the difference of the intermediate feature maps between the benign sample and adversarial example. Wu et al. [61] adopt a regularizer about the distance of the attention maps between these samples." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4608" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "Model-specific approaches. Some works utilize the surrogate model's architecture to improve transferability. Li et al. [28] densely add dropout [46] after each layer to create several ghost networks for better transferability. SGM [60] adopts more gradient from the skip connection in ResNets [20] to boost adversarial transferability. LinBP [19] replaces the zeros with ones in the derivative of ReLU to make the model more linear, leading to improved transferability." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.235, + 0.27, + 0.25 + ], + "angle": 0, + "content": "2.2. Adversarial Defense" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.577 + ], + "angle": 0, + "content": "Numerous adversarial defenses have been proposed to mitigate the threat of adversarial attacks. Adversarial training [15, 48], which adopts the adversarial examples during the training process, has been shown as one of the most effective methods [3] but taking huge computation cost. Recently, Wong et al. [59] find that single-step adversarial examples can bring satisfiable robustness, making it possible to be applied on large-scale datasets (e.g., ImageNet [25]). Pre-processing the input samples before the model is shown to be another effective way. Liao et al. [30] design a high-level representation guided denoiser (HGD) to eliminate the adversarial perturbation. Xie et al. [63] find that random resizing and padding on the input image can mitigate the adversarial threat. Naseer et al. [39] train a neural representation purifier (NRP) by a self-supervised adversarial training mechanism to purify the input sample, which exhibits superior effectiveness against transfer-based adversarial examples. On the other hand, certified defense methods aim to provide provable defense in a given radius [16, 74, 8]. For instance, randomized smoothing (RS) trains a robust ImageNet classifier with a tight robustness guarantee [8]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.584, + 0.273, + 0.599 + ], + "angle": 0, + "content": "2.3. Data Augmentations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.47, + 0.819 + ], + "angle": 0, + "content": "Data augmentations often transform (e.g., flipping, rotation, cropping, etc.) the image during the training process for better generalization. Mixup [75] interpolates two images and their labels to generate virtual samples for training, which also inspires Admix to enhance transferability. Cutmix [73] pastes an image patch to the original patch and mixes the labels accordingly. AutoAugment [9] automatically searches for improved data augmentation policies (operations and parameters) on the dataset for better generalization, which has been widely adopted in deep learning. Unlike these data augmentation strategies, we aim to construct a set of diverse images by transforming the image block using various transformations, which can be used for gradient calculation to achieve better transferability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.212, + 0.848 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce our motivation, provide a detailed description of the proposed SIA, and highlight the difference between SIA and AutoAugment." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.089, + 0.887, + 0.151 + ], + "angle": 0, + "content": "
TIMDIMSIMSSAAdmix
Transferability57.477.679.380.683.6
LPIPS0.250.430.480.540.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.161, + 0.892, + 0.238 + ], + "angle": 0, + "content": "Table 1: The transferability of TIM, DIM, SIM, Admix, SSA, and similarity between 1,000 images and the transformed images evaluated by LPIPS. The transferability is evaluated by the attack success rate of Inception-v3 on the adversarial examples generated on ResNet-18." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.261, + 0.623, + 0.276 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.892, + 0.406 + ], + "angle": 0, + "content": "Since Xie et al. [64] found that transforming the image by random resizing and padding before gradient calculation can generate more transferable adversarial examples, various input transformation based attacks are proposed to further improve transferability. As shown in Fig. 1, the input transformation (e.g., DIM [64] vs. Admix [55]) with better transferability tends to bring more obvious visual change to the input image. This inspires us with a new assumption:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Assumption 1 Without harming the semantic information, the more diverse the transformed image is, the better transferability the adversarial examples have." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.53 + ], + "angle": 0, + "content": "To validate this hypothesis, we utilize the learned perceptual image patch similarity (LPIPS) [79] to evaluate the semantic similarity between the benign samples and the transformed images:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.537, + 0.892, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L P I P S} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) = \\frac {1}{H \\times W} \\sum_ {l} \\sum_ {h, w} \\| \\boldsymbol {z} _ {h, w} ^ {l} - \\hat {\\boldsymbol {z}} _ {h, w} ^ {l} \\| _ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.81 + ], + "angle": 0, + "content": "where \\( z^l \\) and \\( \\hat{z}^l \\) are the extracted feature from \\( l \\)-th layer of SqueezeNet [22] with \\( x \\) and \\( \\hat{x} \\) as input, respectively. A smaller LPIPS value indicates better similarity between the two images. The semantic similarity between the raw images and transformed images by DIM, TIM, SIM and Admix are summarized in Tab. 1. As we can see, the similarity decreases when the transferability of attack increases because the transformed images are more diverse. Moreover, as shown in Fig. 1, such diverse images are significantly different from the raw images, which introduces instability when calculating the gradient. Hence, the more powerful input transformation based attack needs to calculate the gradient on multiple transformed images to eliminate the instability. For instance, DIM utilizes single image, SIM adopts 5 images, while the SOTA Admix/SSA takes 15/20 images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The relation between the diversity of transformed images and transferability inspires us to generate more diverse images for gradient calculation so that we can craft more transferable adversarial examples. In this work, we apply various input transformations on different blocks of a single input image to obtain more diverse images, detailed in Sec. 3.2." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4609" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.899, + 0.166 + ], + "angle": 0, + "content": "
RawVShiftHShiftVFlipHFlipRotateScaleAdd NoiseResizeDCTDropout
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.177, + 0.869, + 0.193 + ], + "angle": 0, + "content": "Table 2: The raw and transformed images using various transformations adopted by SIA (See details in Appendix A)." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.216, + 0.207, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.312, + 0.183, + 0.326 + ], + "angle": 0, + "content": "Raw Image" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.216, + 0.332, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.312, + 0.292, + 0.323 + ], + "angle": 0, + "content": "Scale" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.216, + 0.458, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.312, + 0.436, + 0.324 + ], + "angle": 0, + "content": "Block Scale" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.338, + 0.47, + 0.369 + ], + "angle": 0, + "content": "Figure 2: The randomly sampled raw image and its transformed images by scaling on the full image and \\(3 \\times 3\\) blocks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.393, + 0.32, + 0.407 + ], + "angle": 0, + "content": "3.2. Structure Invariant Attack" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.471, + 0.675 + ], + "angle": 0, + "content": "Considering the limited number of the existing image transformations, we first investigate that how to generate a more diverse image using a single transformation? Without loss of generality, we take scale as an example, which is the base operation of SIM [32] (also a particular case of Admix [55]). As shown in Fig. 2, scaling the image changes its intensity uniformly on all the pixels, resulting in limited diverse images, which is also validated in Tab. 1. To improve the diversity of scaled images, we scale the image with different factors for different parts. Specifically, we randomly split the image into several image blocks and independently scale each image block with different scaled factors. As illustrated in Fig. 2, scaling the image blocks can bring a much more diverse transformed image while humans can still catch the visual information from the image. To explain why such image does not confuse humans, we define the structure of image as follows:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.681, + 0.469, + 0.743 + ], + "angle": 0, + "content": "Definition 1 (Structure of Image). Given an image \\( \\mathbf{x} \\) which is randomly split into \\( s \\times s \\) blocks, the relative relation between each anchor point is the structure of image, where the anchor point is the center of the image block." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.856 + ], + "angle": 0, + "content": "We argue that the structure of image depicts important semantic information for human recognition. For instance, the dog's body should be between its head and tail, while its legs should be under its body. Scaling the image blocks with various factors does not change the structure of image so that the generated image can be correctly recognized by humans as well as deep models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In summary, transforming the image blocks does not harm the recognition but crafts more diverse images, which is of great benefit to improve transferability. To further" + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.223, + 0.779, + 0.237 + ], + "angle": 0, + "content": "Algorithm 1: Structure Invariant Attack" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.241, + 0.867, + 0.316 + ], + "angle": 0, + "content": "Input: Classifier \\( \\pmb{f}(\\cdot) \\) with the loss function \\( J \\); The benign sample \\( \\pmb{x} \\) with ground-truth label \\( y \\); The maximum perturbation \\( \\epsilon \\), number of iterations \\( T \\) and decay factor \\( \\mu \\); Splitting number \\( s \\); Number of transformed images \\( N \\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.317, + 0.744, + 0.332 + ], + "angle": 0, + "content": "Output: An adversarial example." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.332, + 0.717, + 0.347 + ], + "angle": 0, + "content": "1 \\(\\alpha = \\epsilon /T\\) \\(\\pmb {g}_0 = 0\\) \\(x_0^{adv} = x\\)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.348, + 0.672, + 0.36 + ], + "angle": 0, + "content": "2 for \\(t = 0\\to T - 1\\) do" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.332, + 0.717, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.363, + 0.855, + 0.392 + ], + "angle": 0, + "content": "Constructing a set \\(\\mathcal{X}\\) of \\(N\\) transformed images using SIT" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.393, + 0.804, + 0.407 + ], + "angle": 0, + "content": "4 Calculating the average gradient on \\(\\mathcal{X}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.419, + 0.868, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {g}} _ {t + 1} = \\frac {1}{N} \\sum_ {\\boldsymbol {x} _ {i} \\in \\mathcal {X}} \\nabla_ {\\boldsymbol {x}} J (\\boldsymbol {x} _ {i}, y) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.47, + 0.717, + 0.484 + ], + "angle": 0, + "content": "5 | Updating the momentum:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.496, + 0.867, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {t + 1} = \\mu \\boldsymbol {g} _ {t} + \\frac {\\bar {\\boldsymbol {g}} _ {t + 1}}{\\| \\bar {\\boldsymbol {g}} _ {t + 1} \\| _ {1}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.54, + 0.773, + 0.555 + ], + "angle": 0, + "content": "6 | Updating the adversarial example:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.567, + 0.867, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t + 1} ^ {a d v} = \\operatorname {C l i p} \\left(\\boldsymbol {x} _ {t} ^ {a d v} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\boldsymbol {g} _ {t + 1}\\right), 0, 1\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.601, + 0.605, + 0.618 + ], + "angle": 0, + "content": "7 return \\(x_{T}^{adv}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.762 + ], + "angle": 0, + "content": "boost the diversity of transformed images, we apply various image transformations to different image blocks, where the adopted transformations are summarized in Tab. 2. To avoid information loss, we add the constraints on some transformations. For instance, the rotation can only rotate the image \\(180^{\\circ}\\) to avoid dropping some pixels. We denote such transformation as structure invariant transformation (SIT)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Since SIT can generate more diverse images while maintaining the critical semantic information, we adopt SIT as the input transformation to conduct the adversarial attack, denoted as structure invariant attack (SIA). Instead of directly calculating the gradient on the input image, SIA calculates the gradient on several images transformed by SIT. Note that SIA is generally applicable to any gradient-based attacks. Due to the limited space, we integrate SIA into MIFGSM [10] and summarize the algorithm in Algorithm 1." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4610" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.298, + 0.094, + 0.671, + 0.109 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.114, + 0.28, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.114, + 0.482, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.114, + 0.685, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.115, + 0.887, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.221, + 0.28, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.222, + 0.482, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.222, + 0.685, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.222, + 0.887, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Figure 3: Attack success rates \\((\\%)\\) of eight deep models on the adversarial examples crafted on each model by TIM, DIM, DEM, Admix, SSA, and SIA." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.277, + 0.409 + ], + "angle": 0, + "content": "3.3. SIT vs. AutoAugment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.46 + ], + "angle": 0, + "content": "Both SIT and AutoAugment [9] leverage several image transformations to pre-process the images. We summarize the difference as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.464, + 0.47, + 0.509 + ], + "angle": 0, + "content": "- SIT improves the adversarial transferability when attacking the model while AutoAugment boosts the model generalization during the training process." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.51, + 0.47, + 0.554 + ], + "angle": 0, + "content": "- SIT randomly samples the transformation while AutoAugment needs to search for a good policy for each transformation on each dataset." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.555, + 0.47, + 0.615 + ], + "angle": 0, + "content": "- SIT applies various transformations on different image blocks locally but preserves the global structure. On the contrary, AutoAugment applies two transformations sequentially on the image." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.616, + 0.47, + 0.646 + ], + "angle": 0, + "content": "- The fine-grained transformations by SIT generate more diverse images than AutoAugment." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.464, + 0.47, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.66, + 0.202, + 0.677 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.469, + 0.716 + ], + "angle": 0, + "content": "In this section, we conduct extensive evaluations on ImageNet dataset to validate the effectiveness of SIA." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.279, + 0.743 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Dataset. To align with previous works [10, 32, 53, 55], we randomly sample 1,000 images pertaining to 1,000 categories from the ILSVRC 2012 validation set [40], which are correctly classified by the adopted models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Baselines. We compare our proposed SIA with five competitive input transformation based attacks, namely DIM [64], TIM [11], DEM [81], Admix [55], and SSA [36], which are integrated into MI-FGSM [10]. We also integrate the baseline methods with two model-specific approaches, namely LinBP [19] and SGM [60]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.62 + ], + "angle": 0, + "content": "Victim Models. We evaluate the attack performance on two popular model architectures, namely Convolutional Network Works, i.e., ResNet-18 [20], ResNet-101 [20], ResNext-50 [65], DenseNet-121 [23], MobileNet [21], and Transformers, i.e., Vision Transformer (ViT) [12] and Swin Transformer (Swin) [34]. Furthermore, we study several SOTA defense methods, including one adversarial training method, i.e., ensemble adversarially trained model (Incv3_ens) [48], the top-3 submissions in NIPS 2017 defense competition, i.e., high-level representation guided denoiser (HGD) [30], random resizing and padding (R&P) [63] and NIPS-r3\\(^1\\), three input pre-processing based defenses, namely FD [35], JPEG [18] and Bit-Red [67], a certified defense, i.e., randomized smoothing (RS) [8] and a deep denoiser, i.e., neural representation purifier (NRP) [39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.786 + ], + "angle": 0, + "content": "Evaluation Settings. We follow MI-FGSM [10] with the perturbation budget \\(\\epsilon = 16\\), number of iteration \\(T = 10\\), step size \\(\\alpha = \\epsilon / T = 1.6\\) and decay factor \\(\\mu = 1\\). DIM [64] adopts the transformation probability of 0.5 and TIM [11] utilizes the Gaussian kernel with the size of \\(7 \\times 7\\). DEM [81] uses the resize ratios: [1.14, 1.27, 1.4, 1.53, 1.66]. Admix admixes 3 images from other categories with the strength of 0.2 and 5 scaled images for each admixed image. SSA [36] sets the turning factor as 0.5 and the standard deviation as \\(\\epsilon\\). SIA sets the splitting number \\(s = 3\\) and number of transformed images for gradient calculation \\(N = 20\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.796, + 0.733, + 0.813 + ], + "angle": 0, + "content": "4.2. Attacking a Single Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.82, + 0.892, + 0.865 + ], + "angle": 0, + "content": "To validate the effectiveness of the proposed SIA, we first compare SIA with five SOTA input transformation based attacks, namely DIM, TIM, DEM, Admix and SSA." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.858, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/anlthms/nips-2017/tree/master/mmd" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4611" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.462, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.47, + 0.396 + ], + "angle": 0, + "content": "Figure 4: Attack success rates \\((\\%)\\) of eight models on the adversarial examples generated on ResNet-18 when integrating TIM, DIM, DEM, Admix, SSA, and SIA into SGM, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.492 + ], + "angle": 0, + "content": "We generate the adversarial examples on a single model and test them on the other models. The attack success rates, i.e., the misclassification rates of the victim model on the crafted adversarial examples, are summarized in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.471, + 0.903 + ], + "angle": 0, + "content": "From the figure, we can observe that all the attackers can achieve the attack success rate of \\(100.0\\%\\) or near \\(100.0\\%\\), showing that the input transformation based attacks do not degrade the white-box attack performance. As for the black-box performance, TIM exhibits the poorest transferability on these normally trained models. Admix consistently achieves better transferability than DIM and DEM on CNN-based models. Surprisingly, DIM achieves even better transferability than Admix when generating adversarial examples on transformer-based models. SSA exhibits the best transferability among the baselines in most cases. Also, even for the transformer-based models, the adversarial examples generated on Swin show the poorest transferability on ViT than other CNN-based models. Hence, we argue that it is necessary to evaluate the effectiveness of transfer-based attacks on both CNN-based and transformer-based models. We need an in-depth analysis of the transferability among the emerging transformer-based models. Compared with the baselines, SIA consistently performs much better than the best baselines on all eight models with different architectures. In particular, SIA outperforms the winner-up method with a clear margin of \\(14.3\\%\\) on average and achieves an attack success rate higher than the best baseline of at least \\(2.8\\%\\) on all the models. Such consistent and superior performance demonstrates that the proposed SIA is general to various model architectures (either CNN or transformer) to boost transferability effectively." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.885, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.396 + ], + "angle": 0, + "content": "Figure 5: Attack success rates \\((\\%)\\) of eight models on the adversarial examples generated on ResNet-18 when integrating TIM, DIM, DEM, Admix, SSA, and SIA into LinBP, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.422, + 0.848, + 0.439 + ], + "angle": 0, + "content": "4.3. Integration to model-specific approaches" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.52 + ], + "angle": 0, + "content": "To further verify the scalability of the proposed SIA, we integrate existing input transformation-based attacks into two model-specific approaches, namely SGM and LinBP. The adversarial examples are generated on ResNet-18 and test them on the other models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.673 + ], + "angle": 0, + "content": "As depicted in Fig. 4 and Fig. 5, SGM and LinBP can significantly boost the adversarial transferability of vallina input transformation-based attacks. Compared with the baselines, SIA still achieves much better transferability, which has been boosted by an average margin of \\(3.3\\%\\) and \\(3.6\\%\\) when integrated into SGM and LinBP, respectively. These superior results validate the generality of SIA to various transfer-based attacks and underscore the potential of SIA in augmenting adversarial transferability through the fusion of different strategies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.753, + 0.697 + ], + "angle": 0, + "content": "4.4. Attacking Ensemble Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Liu et al. [33] have shown that attacking ensemble models can effectively improve transferability. As shown in Sec. 4.2, the adversarial examples often exhibit poorer transferability across the CNN-based models and transformer-based models than being transferred among the CNN-based models. Hence, we generate adversarial examples on ensemble CNN-based and ensemble transformer-based models and test them on the remaining models to evaluate SIA when attacking ensemble models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 6, when attacking ensemble models, the adversarial examples generated by all attacks exhibit better transferability than that crafted on a single model, showing the excellent compatibility of all the input trans" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4612" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.462, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.47, + 0.412 + ], + "angle": 0, + "content": "Figure 6: Attack success rates (\\%) of eight models on the adversarial examples crafted on ensemble models by TIM, DIM, DEM, Admix, SSA, and SIA, respectively. We generate adversarial examples on the CNN-based models and test them on the transformer-based models, and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.697 + ], + "angle": 0, + "content": "formation based attacks with such a setting. When generating the adversarial examples on ViT and Swin, DIM and DEM achieve better transferability than Admix and SSA, highlighting the difference between crafting adversarial examples on different architectures and the necessity to evaluate the transferability on these models. On all eight models, SIA achieves the attack success rate of at least \\(88.3\\%\\). It outperforms the winner-up approach with a clear margin of \\(6.5\\%\\), showing its superior effectiveness in generating transferable adversarial examples. In particular, SIA achieves \\(94.9\\%\\) and \\(99.3\\%\\) attack success rates on ViT and Swin, respectively, when the adversarial examples are generated on CNN-based models without access to the attention module in the transformer. This further supports our motivation that improving the diversity of transformed images can significantly boost transferability, even on models with completely different architectures." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.327, + 0.725 + ], + "angle": 0, + "content": "4.5. Attacking Defense Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.468, + 0.869 + ], + "angle": 0, + "content": "SIA has achieved superior attack performance on eight normally trained models with different architectures when attacking single model as well as ensemble models. Recently, several defenses have been proposed to mitigate the threat of adversarial examples on ImageNet dataset. To validate the effectiveness of these defenses, we adopt the adversarial examples generated on these eight models simultaneously to attack the defense models, including Inc-v3ens, HGD, R&P, NIPS-r3, FD, JPEG, Bit-Red, RS and NRP." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.902 + ], + "angle": 0, + "content": "The attack results on these defense methods are summarized in Fig. 7. Overall, DIM is on par with TIM on" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.885, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.335, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Figure 7: Attack success rates \\((\\%)\\) of various defense methods on the adversarial examples generated by TIM, DIM, DEM, Admix, SSA, and SIA under ensemble model setting. The adversarial examples are generated on the eight models simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.591 + ], + "angle": 0, + "content": "these defense methods while Admix can consistently outperform the other four baselines. SIA can consistently achieve better transferability than the baselines. In particular, SIA achieves the attack success rate of \\(78.2\\%\\) on the certified defense method (i.e., RS), and SIA can achieve the attack success rate of at least \\(89.9\\%\\) on the other eight defense methods, including the powerful denoising method NRP. Such high attack performance indicates the insufficiency of existing defense methods and raises a new security issue to designing more robust deep learning models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.665, + 0.619 + ], + "angle": 0, + "content": "4.6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.719 + ], + "angle": 0, + "content": "To further gain insights on the superior attack performance achieved by SIA, we conduct a series of ablation studies to validate that all the transformations and splitting the image into blocks can help improve transferability. We generate the adversarial examples on ResNet-18 and test them on the other seven models for all the experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Are all the transformations beneficial for boosting the transerability? There are ten different input transformations that SIA might apply to each image block. To investigate whether each transformation is of benefit to generate more transferable adversarial examples, we first adopt one or two transformations to implement SIA. The results are summarized in Fig. 8. The average attack success rate of MI-FGSM is \\(59.7\\%\\). On the diagonal line of Fig. 8, we only adopt a single transformation for each block and the lowest average attack success rate is \\(67.3\\%\\) by adopting DCT, which outperforms MI-FGSM with a margin of \\(7.6\\%\\). This demonstrates the high effectiveness of SIA and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4613" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.858, + 0.136 + ], + "angle": 0, + "content": "
SIA-VShift-HShift-VFlip-HFlip-Rotate-Sclae-Add Noise-Resize-DCT-Dropout
92.189.790.190.188.488.390.190.690.290.190.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.146, + 0.894, + 0.178 + ], + "angle": 0, + "content": "Table 3: The average attack success rates of adversarial examples crafted by SIA and SIA without a single transformation. The adversaries are generated on ResNet-18 and tested on the other seven deep models. - indicates removing such transformation." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.205, + 0.468, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.492 + ], + "angle": 0, + "content": "Figure 8: The average attack success rates \\((\\%)\\) on the adversarial examples generated by SIA using one (diagonal) or two (other) transformations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.718 + ], + "angle": 0, + "content": "its generality to various transformations. We can also observe that the transferability can be further improved when we combine any two transformations, showing the advantage of combining these transformations. To further validate the necessity of each transformation, we generate adversarial examples by SIA without each transformation and summarize the results in Tab. 3. Since removing a single transformation does not significantly decrease the diversity of the transformation, these attacks achieve similar attack performance. However, no matter which transformation is removed, the transferability will be decreased, supporting that each transformation is of benefit to generate more transferable adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.904 + ], + "angle": 0, + "content": "Are the image blocks beneficial for boosting transferability? We have shown that the abundant transformations can effectively improve transferability. Here we further explore whether it is useful to apply these transformations to the image block instead of the raw image. For comparison, we randomly apply the transformation to the raw image \\( s \\times s \\) times, denoted as \\( \\mathrm{SIA_r} \\). The results are reported in Fig. 9. We can observe that \\( \\mathrm{SIA_r} \\) exhibits better transferability than MI-FGSM, which also validates our motivation that improving the diversity of transformed images can generate more transferable adversarial examples. Our SIA consistently performs better than \\( \\mathrm{SIA_r} \\), indicating that applying" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.205, + 0.887, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.445, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Figure 9: The average attack success rates \\((\\%)\\) of adversarial examples generated by SIA and \\(\\mathrm{SIA_r}\\) (randomly transform the image without the block partition)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.551 + ], + "angle": 0, + "content": "the transformation onto each image block is of great benefit to improve transferability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.681, + 0.575 + ], + "angle": 0, + "content": "4.7. Parameter Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.674 + ], + "angle": 0, + "content": "In this subsection, we conduct parameter studies to explore the impact of two hyper-parameters, namely the number of blocks \\( s \\) and the number of transformed images for gradient calculation \\( N \\). All the adversarial examples are generated on ResNet-18 and tested on the other seven models in the black-box setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.854 + ], + "angle": 0, + "content": "On the number of blocks \\( s \\). The number of blocks determines how diverse the transformed images are, which can influence the attack performance. To find a good value for \\( s \\), we conduct SIA with \\( s \\) from 1 to 5 and summarize the attack results in Fig. 10. When \\( s \\leq 3 \\), increasing the value of \\( s \\) leads to better diversity, which can improve the attack performance. However, when we continually increase the value of \\( s \\), the diversity is increased, but it also introduces more variance to the gradient, decaying the attack performance slightly. Hence, we adopt \\( s = 3 \\) to balance the diversity of images and variance of the gradient for better attack performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "On the number of images for gradient calculation \\(N\\). SIA calculates the average gradient on \\(N\\) images to eliminate the variance introduced by the transformation. To de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4614" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.095, + 0.468, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.362 + ], + "angle": 0, + "content": "Figure 10: Attack success rates \\((\\%)\\) of adversarial examples generated by SIA with various number of blocks \\(s\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.473, + 0.572 + ], + "angle": 0, + "content": "termine a good value for \\( N \\), we evaluate SIA with \\( N \\) from 1 to 30 and report the attack success rates in Fig. 11. As we can observe, when \\( N = 1 \\), SIA introduces massive variance into the gradient and achieves the lowest performance on all models. When we increase the value of \\( N \\), the variance among the gradients can be eliminated, and SIA achieves better transferability before \\( N = 20 \\). When \\( N > 20 \\), increasing \\( N \\) can only introduce computation cost without performance improvement. To balance the attack performance and computation cost, we adopt \\( N = 20 \\) in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.611, + 0.196, + 0.627 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.901 + ], + "angle": 0, + "content": "In this work, we find that the existing input transformation based attack with better transferability often generates more diverse transformed images. Based on this finding, we design a new image transformation, called structure invariant transformation (SIT), which splits the image into several blocks and randomly transforms each image block independently. With such image transformation, we propose a novel input transformation based attack, dubbed structure invariant attack (SIA), which calculates the average gradient on several transformed images by SIT to update the perturbation. Extensive evaluations demonstrate that SIA can achieve remarkably better transferability than the existing SOTA attacks. In our opinion, SIA provides a new direction by applying the transformation onto the image block to effectively boost transferability, which sheds new light on generating more transferable adversarial examples with fine-grained transformations." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.095, + 0.885, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.332, + 0.892, + 0.362 + ], + "angle": 0, + "content": "Figure 11: Attack success rates \\((\\%)\\) of adversarial examples generated by SIA with various number of images \\(N\\)" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.39, + 0.597, + 0.405 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.416, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[1] Moustafa Alzantot, Yash Sharma, Ahmed Elgohary, BoJhang Ho, Mani B. Srivastava, and Kai-Wei Chang. Generating Natural Language Adversarial Examples. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 2890-2896, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.488, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square Attack: A Query-Efficient Black-Box Adversarial Attack via Random Search. In Proceedings of the European Conference on Computer Vision, pages 484–501, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.559, + 0.892, + 0.627 + ], + "angle": 0, + "content": "[3] Anish Athalye, Nicholas Carlini, and David A. Wagner. Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. In Proceedings of the International Conference on Machine Learning, pages 274–283, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.63, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[4] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models. In Proceedings of the International Conference on Learning Representations, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.702, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[5] Nicholas Carlini and David A. Wagner. Audio Adversarial Examples: Targeted Attacks on Speech-to-Text. In Proceedings of the IEEE Symposium on Security and Privacy (Workshops), pages 1-7, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.76, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[6] Minhao Cheng, Thong Le, Pin-Yu Chen, Huan Zhang, Jinfeng Yi, and Cho-Jui Hsieh. Query-Efficient Hard-label Black-box Attack: An Optimization-based Approach. In Proceedings of the International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.831, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[7] Shuyu Cheng, Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Improving Black-box Adversarial Attacks with a Transfer-based Prior. In Proceedings of the Advances in Neural Information Processing Systems, pages 10932-10942, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.416, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4615" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[8] Jeremy M. Cohen, Elan Rosenfeld, and J. Zico Kolter. Certified Adversarial Robustness via Randomized Smoothing. In Proceedings of the International Conference on Machine Learning, pages 1310-1320, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.15, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[9] Ekin D. Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, and Quoc V. Le. AutoAugment: Learning Augmentation Strategies From Data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting Adversarial Attacks With Momentum. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[11] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading Defenses to Transferable Adversarial Examples by Translation-Invariant Attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In Proceedings of the International Conference on Learning Representations, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.47, + 0.545 + ], + "angle": 0, + "content": "[13] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust Physical-World Attacks on Deep Learning Visual Classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1625–1634, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.547, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we Ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.617, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[15] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and Harnessing Adversarial Examples. In Proceedings of the International Conference on Learning Representations, 2015. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.675, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[16] Sven Gowal, Krishnamurthy Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timothy Arthur Mann, and Pushmeet Kohli. Scalable Verified Training for Provably Robust Image Classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4841-4850, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.759, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[17] Chuan Guo, Jacob R. Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Q. Weinberger. Simple Black-box Adversarial Attacks. In Proceedings of the International Conference on Machine Learning, pages 2484-2493, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[18] Chuan Guo, Mayank Rana, Moustapha Cissé, and Laurens van der Maaten. Countering Adversarial Images Using Input Transformations. In Proceedings of the International Conference on Learning Representations, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[19] Yiwen Guo, Qizhang Li, and Hao Chen. Backpropagating linearly improves transferability of adversarial examples." + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Advances in neural information processing systems, 33:85-95, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[21] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv preprint arXiv:1704.04861, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-Eccitation Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[23] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely Connected Convolutional Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2261-2269, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[24] Andrew Ilyas, Logan Engstrom, Anish Athalye, and Jessy Lin. Black-box Adversarial Attacks with Limited Queries and Information. In Proceedings of the International Conference on Machine Learning, pages 2142-2151, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[25] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Proceedings of the Advances in Neural Information Processing Systems, pages 1106-1114, 2012. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[26] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial Examples in the Physical World. In Proceedings of the International Conference on Learning Representations (Workshops), 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[27] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. QEBA: Query-Efficient Boundary-Based Blackbox Attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1218-1227, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[28] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan L. Yuille. Learning Transferable Adversarial Examples via Ghost Networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11458-11465, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[29] Yandong Li, Lijun Li, Liqiang Wang, Tong Zhang, and Boqing Gong. NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks. In Proceedings of the International Conference on Machine Learning, pages 3866–3876, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[30] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense Against Adversarial Attacks Using High-Level Representation Guided Denoiser. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[31] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4616" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Daan Wierstra. Continuous Control with Deep Reinforcement Learning. In Proceedings of the International Conference on Learning Representations, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[32] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov Accelerated Gradient and Scale Invariance for Adversarial Attacks. In Proceedings of the International Conference on Learning Representations, 2020. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[33] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into Transferable Adversarial Examples and Blackbox Attacks. In Proceedings of the International Conference on Learning Representations, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.333 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9992-10002, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[35] Zihao Liu, Qi Liu, Tao Liu, Nuo Xu, Xue Lin, Yanzhi Wang, and Wujie Wen. Feature Distillation: DNN-Oriented JPEG Compression Against Adversarial Examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 860-868, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[36] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xi-anglong Liu, Jian Zhang, and Jingkuan Song. Frequency Domain Model Augmentation for Adversarial Attack. In Proceedings of the European Conference on Computer Vision, pages 549–566, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards Deep Learning Models Resistant to Adversarial Attacks. In Proceedings of the International Conference on Learning Representations, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.617 + ], + "angle": 0, + "content": "[38] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2574-2582, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.47, + 0.688 + ], + "angle": 0, + "content": "[39] Muzammal Naseer, Salman H. Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A Self-supervised Approach for Adversarial Robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 259-268, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[40] Russakovsky Olga, Deng Jia, Su Hao, Krause Jonathan, Satheesh Sanjeev, Ma Sean, hUANG Zhiheng, Karpathy Andrej, Khosla Adiya, and Bernstein Michael et al. Imagenet large scale visual recognition challenge. In International Journal of Computer Vision, pages 211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[41] Florian Schroff, Dmitry Kalenichenko, and James Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[42] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K. Reiter. Accessorize to a Crime: Real and Stealthy Attacks on State-of-the-Art Face Recognition. In Proceedings of the ACM Conference on Computer and Communications Security, pages 1528-1540, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[43] Luchuan Song, Zheng Fang, Xiaodan Li, Xiaoyi Dong, Zhenchao Jin, Yuefeng Chen, and Siwei Lyu. Adaptive face forgery detection in cross domain. In European Conference on Computer Vision, pages 467-484. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[44] Luchuan Song, Xiaodan Li, Zheng Fang, Zhenchao Jin, YueFeng Chen, and Chenliang Xu. Face forgery detection via symmetric transformer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4102-4111, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.285 + ], + "angle": 0, + "content": "[45] Luchuan Song, Bin Liu, Guojun Yin, Xiaoyi Dong, Yufei Zhang, and Jia-Xuan Bai. Tacr-net: editing on deep video and voice portraits. In Proceedings of the 29th ACM International Conference on Multimedia, pages 478-486, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.354 + ], + "angle": 0, + "content": "[46] Nitish Srivastava, Geoffrey E. Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a Simple Way to prevent Neural Networks from Overfitting. The Journal of Machine Learning Research, 15(1):1929-1958, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[47] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing Properties of Neural Networks. In Proceedings of the International Conference on Learning Representations, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[48] Florian Tramèr, Alexey Kurakin, Nicolas Papernot, Ian J. Goodfellow, Dan Boneh, and Patrick D. McDaniel. Ensemble Adversarial Training: Attacks and Defenses. In Proceedings of the International Conference on Learning Representations, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[49] Jonathan Uesato, Brendan O'Donoghue, Pushmeet Kohli, and Aïron van den Oord. Adversarial Risk and the Dangers of Evaluating Against Weak Attacks. In Proceedings of the International Conference on Machine Learning, pages 5032-5041, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Proceedings of the Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[51] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. CosFace: Large Margin Cosine Loss for Deep Face Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5265-5274, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[52] Xiaosen Wang, Jin Hao, Yichen Yang, and Kun He. Natural Language Adversarial Defense through Synonym Encoding. In Proceedings of the Conference on Uncertainty in Artificial Intelligence, pages 823-833, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[53] Xiaosen Wang and Kun He. Enhancing the Transferability of Adversarial Attacks Through Variance Tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[54] Xiaosen Wang, Kun He, Chuanbiao Song, Liwei Wang, and John E Hopcroft. AT-GAN: An Adversarial Generator Model for Non-constrained Adversarial Examples. arXiv preprint arXiv:1904.07793, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[55] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the Transferability of Adversarial At" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4617" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "tacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16138-16147, 2021. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[56] Xiaosen Wang, Jiadong Lin, Han Hu, Jingdong Wang, and Kun He. Boosting Adversarial Transferability through Enhanced Momentum. In Proceedings of the British Machine Vision Conference, page 272, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[57] Xiaosen Wang, Zeliang Zhang, Kangheng Tong, Dihong Gong, Kun He, Zhifeng Li, and Wei Liu. Triangle Attack: A Query-Efficient Decision-Based Adversarial Attack. In Proceedings of the European Conference on Computer Vision, pages 156-174, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[58] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable Adversarial Attacks for Image and Video Object Detection. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 954-960, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.329, + 0.47, + 0.383 + ], + "angle": 0, + "content": "[59] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is Better than Free: Revisiting Adversarial Training. In Proceedings of the International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[60] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip Connections Matter: On the Transferability of Adversarial Examples Generated with ResNets. In Proceedings of the International Conference on Learning Representations, 2020. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[61] Weibin Wu, Yuxin Su, Xixian Chen, Shenglin Zhao, Irwin King, Michael R. Lyu, and Yu-Wing Tai. Boosting the Transferability of Adversarial Samples via Attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1158-1167, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[62] Weibin Wu, Yuxin Su, Michael R Lyu, and Irwin King. Improving the Transferability of Adversarial Samples with Adversarial Transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9024-9033, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[63] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan L. Yuille. Mitigating Adversarial Effects Through Randomization. In Proceedings of the International Conference on Learning Representations, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[64] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving Transferability of Adversarial Examples With Input Diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[65] Saining Xie, Ross B. Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated Residual Transformations for Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[66] Yifeng Xiong, Jiadong Lin, Min Zhang, John E. Hopcroft, and Kun He. Stochastic Variance Reduced Ensemble Adversarial Attack for Boosting the Adversarial Transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14972, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[67] Weilin Xu, David Evans, and Yanjun Qi. Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the Network and Distributed System Security Symposium, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[68] Bo Yang, Hengwei Zhang, Zheming Li, Yuchen Zhang, Kaiyong Xu, and Jindong Wang. Adversarial example generation with adbelief optimizer and crop invariance. Applied Intelligence, 53(2):2332-2347, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[69] Yichen Yang, Xiaosen Wang, and Kun He. Robust Textual Embedding against Word-level Adversarial Attacks. pages 2214-2224, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[70] Zhuolin Yang, Bo Li, Pin-Yu Chen, and Dawn Song. Characterizing Audio Adversarial Examples Using Temporal Dependency. In Proceedings of the International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[71] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural Color Fool: Towards Boosting Black-box Unrestricted Attacks. In Proceedings of the Advances in Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[72] Zheng Yuan, Jie Zhang, and Shiguang Shan. Adaptive image transformations for transfer-based adversarial attack. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[73] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6022-6031, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[74] Huan Zhang, Hongge Chen, Chaowei Xiao, Sven Gowal, Robert Stanforth, Bo Li, Duane S. Boning, and Cho-Jui Hsieh. Towards Stable and Efficient Training of Verifiably Robust Neural Networks. In Proceedings of the International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[75] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. Mixup: Beyond Empirical Risk Minimization. In Proceedings of the International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[76] Jianping Zhang, Jen-tse Huang, Wenxuan Wang, Yichen Li, Weibin Wu, Xiaosen Wang, Yuxin Su, and Michael R. Lyu. Improving the Transferability of Adversarial Samples by Path-Augmented Method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8173-8182, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[77] Qilong Zhang, Xiaodan Li, Yuefeng Chen, Jingkuan Song, Lianli Gao, Yuan He, and Hui Xue. Beyond ImageNet Attack: Towards Crafting Adversarial Examples for Black-box Domains. In Proceedings of the International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[78] Qilong Zhang, Chaoning Zhang, Chaoqun Li, Jingkuan Song, Lianli Gao, and Heng Tao Shen. Practical No-box Adversarial Attacks with Training-free Hybrid Image Transformation. arXiv preprint arXiv:2203.04607, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[79] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[80] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable Ad" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4618" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "versarial Perturbations. In Proceedings of the European Conference on Computer Vision, pages 471-486, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.192 + ], + "angle": 0, + "content": "[81] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the Transferability of Adversarial Examples with Resized-Diverse-Inputs, Diversity-Ensemble and Region Fitting. In Proceedings of the European Conference on Computer Vision, pages 563–579, 2020. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4619" + } + ] +] \ No newline at end of file diff --git a/2023/Structure Invariant Transformation for better Adversarial Transferability/303ed349-48a8-4393-8e2e-88000a7d20c9_origin.pdf b/2023/Structure Invariant Transformation for better Adversarial Transferability/303ed349-48a8-4393-8e2e-88000a7d20c9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..75cf348f92689ce7dc0c9bccd1734f6f57ecbfc1 --- /dev/null +++ b/2023/Structure Invariant Transformation for better Adversarial Transferability/303ed349-48a8-4393-8e2e-88000a7d20c9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77cafdc35dee0c796046fa26f66a7c761040ce898c97d65caccee129a0fd26bf +size 3437662 diff --git a/2023/Structure Invariant Transformation for better Adversarial Transferability/full.md b/2023/Structure Invariant Transformation for better Adversarial Transferability/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4925e1de50bac0cd77e154867ab29508da1d782d --- /dev/null +++ b/2023/Structure Invariant Transformation for better Adversarial Transferability/full.md @@ -0,0 +1,392 @@ +# Structure Invariant Transformation for better Adversarial Transferability + +Xiaosen Wang +Huawei Singularity Security Lab + +xiaosen@hust.edu.cn + +Zeliang Zhang HUST + +hust0426@gmail.com + +Jianping Zhang Chinese University of Hong Kong + +jpzhang@cse.cuhk.edu.hk + +# Abstract + +Given the severe vulnerability of Deep Neural Networks (DNNs) against adversarial examples, there is an urgent need for an effective adversarial attack to identify the deficiencies of DNNs in security-sensitive applications. As one of the prevalent black-box adversarial attacks, the existing transfer-based attacks still cannot achieve comparable performance with the white-box attacks. Among these, input transformation based attacks have shown remarkable effectiveness in boosting transferability. In this work, we find that the existing input transformation based attacks transform the input image globally, resulting in limited diversity of the transformed images. We postulate that the more diverse transformed images result in better transferability. Thus, we investigate how to locally apply various transformations onto the input image to improve such diversity while preserving the structure of image. To this end, we propose a novel input transformation based attack, called Structure Invariant Transformation (SIA), which applies a random image transformation onto each image block to craft a set of diverse images for gradient calculation. Extensive experiments on the standard ImageNet dataset demonstrate that SIA exhibits much better transferability than the existing SOTA input transformation based attacks on CNN-based and transformer-based models, showing its generality and superiority in boosting transferability. Code is available at https://github.com/xiaosen-wang/SIT. + +# 1. Introduction + +With the unprecedented progress of Deep Neural Networks (DNNs) [25, 20, 23, 50, 12], they have been deployed in many security-sensitive applications, such as face recognition [41, 51, 43, 44], autonomous driving [14, 31], etc. On the other hand, recent works have found that DNNs are vulnerable to adversarial examples [47, 15], which mislead the deep models with imperceptible perturbations. This brings a huge threat to the real-world applications [42, 13, 45, 57, 71, 78] and makes it imperative for an + +![](images/af123078fe2f1c8acaac5afc6622d7ea8abf93cff6c1319afb1260a5c7c00e41.jpg) +Raw Image + +![](images/ca181908d16984ee20827d0d4bd15ed460b59041280cf6de4dfd54f53a97fc51.jpg) +DIM + +![](images/0f1fc384044917aac920f5630552beba2b0d1f571134445f3b0d1632c3e5883e.jpg) +Admix + +![](images/832227d53516e0694060b525fa38b28240182fae977294e494752ed3bb079bdb.jpg) +TIM + +![](images/a08afce371b4e17f65b89af169699fe7de41538ea190d48475a62e164d0a7c57.jpg) +SSA +Figure 1: The raw image and its transformed images by DIM (resize factor of 0.8) [64], TIM (translated by 15 pixels) [11], SIM (scale factor of 0.5), Admix (admix strength of 0.2 and scale factor of 0.5) [55], SSA (turning factor of 0.5) [36] and our proposed SIA $(3\times 3$ blocks). + +![](images/5530125076220063c78af0a9df28ab9cc4112e97291a9315c8250c58c5384086.jpg) +SIM + +![](images/dc2dddd99efc9d96aace3011f725c70e0ae0898deed9d4f361dbc4b4cb077e95.jpg) +SIA (Ours) + +effective attack to identify the deficiencies of DNNs when we robustify the deep models or deploy them for commercial applications. + +Existing adversarial attacks usually fall into two categories: white-box attacks [15, 38, 37, 26, 54] can fetch any information of the target model, including (hyper-)parameters, gradient, architecture, while black-box attacks [24, 7, 4, 27, 33, 77] are only allowed limited access to the target model. One of the significant properties of adversarial examples is their transferability [64, 10, 53, 76], in which the adversarial examples generated on one model can still mislead other models, making it possible to attack the real-world applications in the black-box setting. However, existing adversarial attacks [26, 37] often exhibit superior white-box attack performance but poor transferability. + +To craft more transferable adversarial examples, various techniques have been proposed, such as momentum-based methods [10, 32, 53, 56], input transformations [64, 11, 32, 68, 72], ensemble attacks [33, 66, 36], advanced objective functions [61, 80] and model-specific approaches [28, 60]. Among which, input transformations (e.g., random resizing and padding [64], translation [11], scale [32] admix [55], etc.) that transform the image before gradient calculation, + +have achieved superior transferability and attracted broad attention. Nevertheless, we find that the existing input transformation based attacks, even the SOTA attack Admix [55] and SSA [36], transform the image globally without changing the local relationship among objects in the input image. We argue that the diversity of such transformed images is still not enough, leading to limited transferability. + +In this work, we postulate and empirically validate that the more diverse transformed images lead to better transferability. Based on this observation, instead of applying a single transformation on the input image, we apply different transformations locally on different parts of the image to enhance the diversity of transformed images. As shown in Fig. 1, such transformation can bring much more difference to the generated image compared with the raw image but still preserve the global structure of the object. Based on this transformation, we propose a novel input transformation based attack called structure invariant attack (SIA), which utilizes the gradient of these transformed images to update the adversarial examples for better transferability. + +Our contributions are summarized as follows: + +- We empirically validate that the high diversity of transformed images is beneficial to improve transferability, which sheds new light on how to design new input transformations for more transferable adversarial examples. +- We design a new image transformation method, which locally applies different transformations on different parts of the image to generate more diverse images but preserve its global structure. +- Based on the proposed image transformation, we devise a new input transformation based attack, called structure invariant attack (SIA), to generate more transferable adversarial examples. +- Extensive experiments on ImageNet dataset demonstrate that SIA outperforms the baselines with a clear margin on CNN-based as well as transformer-based models, showing its superiority and generality. + +# 2. Related Work + +Adversarial examples [47] have brought an impressive threat to the DNN-enabled applications, such as computer vision [15, 58, 3], natural language processing [1, 52, 69], speech recognition [5, 70], etc. To identify the vulnerability of DNNs, various attack methods have been proposed recently, such as gradient-based attacks [15, 38, 26, 37], score-based attacks [24, 49, 17, 29, 2], decision-based attacks [4, 6, 27, 57] and transfer-based attacks [33, 10, 64, 32, 55]. Among these, transfer-based attacks do not access any information of the target model, making it applicable to attack any model in the physical world. In this work, we focus on generating more transferable adversarial examples and briefly introduce the existing transfer-based attacks, the corresponding defense methods, and data augmentations. + +# 2.1. Adversarial Attack + +As the first gradient-based attack, Fast Gradient Sign Method (FGSM) [15] adds the perturbation in the gradient direction to the benign sample, leading to high attack efficiency but limited performance. Later, I-FGSM [26] extends FGSM into an iterative version, which achieves much better white-box attack performance but poor transferability. Given the high efficiency and effectiveness of I-FGSM, numerous transfer-based attacks are proposed to boost transferability based on I-FGSM to attack the deep model in the black-box setting. + +Momentum-based methods. MI-FGSM [10] introduces momentum into I-FGSM to stabilize the optimization direction and escape local maxima. NI-FGSM [32] adopts Nesterov Accelerated Gradient to accumulate the momentum, which achieves better transferability. Variance tuning [53] adopts the gradient variance of the previous iteration to tune the current gradient in MI-FGSM and NI-FGSM, significantly improving transferability. EMIFGSM [56] enhances the momentum by accumulating several samples' gradients in the previous gradient's direction to further stabilize the optimization direction. + +Input transformation based attacks. DIM [64] is the first input transformation based attack, which adds padding to a randomly resized image for fixed size before gradient calculation. TIM [11] optimizes the perturbation over an ensemble of translated images, which is further approximated by convolving the gradient at the untranslated image with a pre-defined kernel. SIM [32] scales the images with different scale factors for gradient calculation. DEM [81] averages the gradient on several diverse transformed images similar to DIM with various resizing factors. Admix [55] calculates the gradient on the input image admixed with a small portion of each add-in image from other categories while using the original label of the input. Wu et al. [62] train an adversarial transformation network to destroy the adversarial perturbation and require the synthesized adversarial examples resistant to such transformations. SSA [36] adds Gaussian noise and randomly masks the image in the frequency domain to transform the input image. + +Ensemble attacks. Liu et al. [33] first found that the adversarial examples generated on multiple models denoted as ensemble attack, are more transferable. Recently, Xiong et al. [66] reduce the gradient variance between various models to boost the ensemble attack. + +Advanced objective functions. The above attacks often take the cross-entropy loss as the objective function. Researchers also find that some regularizers are beneficial to boost transferability. For instance, Zhou et al. [80] additionally maximize the difference of the intermediate feature maps between the benign sample and adversarial example. Wu et al. [61] adopt a regularizer about the distance of the attention maps between these samples. + +Model-specific approaches. Some works utilize the surrogate model's architecture to improve transferability. Li et al. [28] densely add dropout [46] after each layer to create several ghost networks for better transferability. SGM [60] adopts more gradient from the skip connection in ResNets [20] to boost adversarial transferability. LinBP [19] replaces the zeros with ones in the derivative of ReLU to make the model more linear, leading to improved transferability. + +# 2.2. Adversarial Defense + +Numerous adversarial defenses have been proposed to mitigate the threat of adversarial attacks. Adversarial training [15, 48], which adopts the adversarial examples during the training process, has been shown as one of the most effective methods [3] but taking huge computation cost. Recently, Wong et al. [59] find that single-step adversarial examples can bring satisfiable robustness, making it possible to be applied on large-scale datasets (e.g., ImageNet [25]). Pre-processing the input samples before the model is shown to be another effective way. Liao et al. [30] design a high-level representation guided denoiser (HGD) to eliminate the adversarial perturbation. Xie et al. [63] find that random resizing and padding on the input image can mitigate the adversarial threat. Naseer et al. [39] train a neural representation purifier (NRP) by a self-supervised adversarial training mechanism to purify the input sample, which exhibits superior effectiveness against transfer-based adversarial examples. On the other hand, certified defense methods aim to provide provable defense in a given radius [16, 74, 8]. For instance, randomized smoothing (RS) trains a robust ImageNet classifier with a tight robustness guarantee [8]. + +# 2.3. Data Augmentations + +Data augmentations often transform (e.g., flipping, rotation, cropping, etc.) the image during the training process for better generalization. Mixup [75] interpolates two images and their labels to generate virtual samples for training, which also inspires Admix to enhance transferability. Cutmix [73] pastes an image patch to the original patch and mixes the labels accordingly. AutoAugment [9] automatically searches for improved data augmentation policies (operations and parameters) on the dataset for better generalization, which has been widely adopted in deep learning. Unlike these data augmentation strategies, we aim to construct a set of diverse images by transforming the image block using various transformations, which can be used for gradient calculation to achieve better transferability. + +# 3. Methodology + +In this section, we introduce our motivation, provide a detailed description of the proposed SIA, and highlight the difference between SIA and AutoAugment. + +
TIMDIMSIMSSAAdmix
Transferability57.477.679.380.683.6
LPIPS0.250.430.480.540.73
+ +Table 1: The transferability of TIM, DIM, SIM, Admix, SSA, and similarity between 1,000 images and the transformed images evaluated by LPIPS. The transferability is evaluated by the attack success rate of Inception-v3 on the adversarial examples generated on ResNet-18. + +# 3.1. Motivation + +Since Xie et al. [64] found that transforming the image by random resizing and padding before gradient calculation can generate more transferable adversarial examples, various input transformation based attacks are proposed to further improve transferability. As shown in Fig. 1, the input transformation (e.g., DIM [64] vs. Admix [55]) with better transferability tends to bring more obvious visual change to the input image. This inspires us with a new assumption: + +Assumption 1 Without harming the semantic information, the more diverse the transformed image is, the better transferability the adversarial examples have. + +To validate this hypothesis, we utilize the learned perceptual image patch similarity (LPIPS) [79] to evaluate the semantic similarity between the benign samples and the transformed images: + +$$ +\operatorname {L P I P S} (\boldsymbol {x}, \hat {\boldsymbol {x}}) = \frac {1}{H \times W} \sum_ {l} \sum_ {h, w} \| \boldsymbol {z} _ {h, w} ^ {l} - \hat {\boldsymbol {z}} _ {h, w} ^ {l} \| _ {2} \tag {1} +$$ + +where $z^l$ and $\hat{z}^l$ are the extracted feature from $l$ -th layer of SqueezeNet [22] with $x$ and $\hat{x}$ as input, respectively. A smaller LPIPS value indicates better similarity between the two images. The semantic similarity between the raw images and transformed images by DIM, TIM, SIM and Admix are summarized in Tab. 1. As we can see, the similarity decreases when the transferability of attack increases because the transformed images are more diverse. Moreover, as shown in Fig. 1, such diverse images are significantly different from the raw images, which introduces instability when calculating the gradient. Hence, the more powerful input transformation based attack needs to calculate the gradient on multiple transformed images to eliminate the instability. For instance, DIM utilizes single image, SIM adopts 5 images, while the SOTA Admix/SSA takes 15/20 images. + +The relation between the diversity of transformed images and transferability inspires us to generate more diverse images for gradient calculation so that we can craft more transferable adversarial examples. In this work, we apply various input transformations on different blocks of a single input image to obtain more diverse images, detailed in Sec. 3.2. + +
RawVShiftHShiftVFlipHFlipRotateScaleAdd NoiseResizeDCTDropout
+ +Table 2: The raw and transformed images using various transformations adopted by SIA (See details in Appendix A). + +![](images/bb320fe443122e66f26d3abbe023f06e06ffd0d2a970384ad4953a260648d04a.jpg) +Raw Image + +![](images/8af069c8090e5d480041e98e6fc262359e226a77e2aebe686a78fc26f34af958.jpg) +Scale + +![](images/1afb90fdd68ced872ce7e7d2f8b370557058966d66e97bdb215a22f0830e0488.jpg) +Block Scale +Figure 2: The randomly sampled raw image and its transformed images by scaling on the full image and $3 \times 3$ blocks. + +# 3.2. Structure Invariant Attack + +Considering the limited number of the existing image transformations, we first investigate that how to generate a more diverse image using a single transformation? Without loss of generality, we take scale as an example, which is the base operation of SIM [32] (also a particular case of Admix [55]). As shown in Fig. 2, scaling the image changes its intensity uniformly on all the pixels, resulting in limited diverse images, which is also validated in Tab. 1. To improve the diversity of scaled images, we scale the image with different factors for different parts. Specifically, we randomly split the image into several image blocks and independently scale each image block with different scaled factors. As illustrated in Fig. 2, scaling the image blocks can bring a much more diverse transformed image while humans can still catch the visual information from the image. To explain why such image does not confuse humans, we define the structure of image as follows: + +Definition 1 (Structure of Image). Given an image $\mathbf{x}$ which is randomly split into $s \times s$ blocks, the relative relation between each anchor point is the structure of image, where the anchor point is the center of the image block. + +We argue that the structure of image depicts important semantic information for human recognition. For instance, the dog's body should be between its head and tail, while its legs should be under its body. Scaling the image blocks with various factors does not change the structure of image so that the generated image can be correctly recognized by humans as well as deep models. + +In summary, transforming the image blocks does not harm the recognition but crafts more diverse images, which is of great benefit to improve transferability. To further + +# Algorithm 1: Structure Invariant Attack + +Input: Classifier $\pmb{f}(\cdot)$ with the loss function $J$ ; The benign sample $\pmb{x}$ with ground-truth label $y$ ; The maximum perturbation $\epsilon$ , number of iterations $T$ and decay factor $\mu$ ; Splitting number $s$ ; Number of transformed images $N$ + +Output: An adversarial example. + +1 $\alpha = \epsilon /T$ $\pmb {g}_0 = 0$ $x_0^{adv} = x$ +2 for $t = 0\to T - 1$ do + +Constructing a set $\mathcal{X}$ of $N$ transformed images using SIT + +4 Calculating the average gradient on $\mathcal{X}$ + +$$ +\bar {\boldsymbol {g}} _ {t + 1} = \frac {1}{N} \sum_ {\boldsymbol {x} _ {i} \in \mathcal {X}} \nabla_ {\boldsymbol {x}} J (\boldsymbol {x} _ {i}, y) \tag {2} +$$ + +5 | Updating the momentum: + +$$ +\boldsymbol {g} _ {t + 1} = \mu \boldsymbol {g} _ {t} + \frac {\bar {\boldsymbol {g}} _ {t + 1}}{\| \bar {\boldsymbol {g}} _ {t + 1} \| _ {1}} \tag {3} +$$ + +6 | Updating the adversarial example: + +$$ +\boldsymbol {x} _ {t + 1} ^ {a d v} = \operatorname {C l i p} \left(\boldsymbol {x} _ {t} ^ {a d v} + \alpha \cdot \operatorname {s i g n} \left(\boldsymbol {g} _ {t + 1}\right), 0, 1\right) \tag {4} +$$ + +7 return $x_{T}^{adv}$ + +boost the diversity of transformed images, we apply various image transformations to different image blocks, where the adopted transformations are summarized in Tab. 2. To avoid information loss, we add the constraints on some transformations. For instance, the rotation can only rotate the image $180^{\circ}$ to avoid dropping some pixels. We denote such transformation as structure invariant transformation (SIT). + +Since SIT can generate more diverse images while maintaining the critical semantic information, we adopt SIT as the input transformation to conduct the adversarial attack, denoted as structure invariant attack (SIA). Instead of directly calculating the gradient on the input image, SIA calculates the gradient on several images transformed by SIT. Note that SIA is generally applicable to any gradient-based attacks. Due to the limited space, we integrate SIA into MIFGSM [10] and summarize the algorithm in Algorithm 1. + +![](images/20fe0ecb2ad38d5b5fe1ca29c6ec19511b055c4796e46c3c123469f92f574f12.jpg) + +![](images/fc8162f0fa6e6f4a85aa9463d728b401433b12fb6cace461ad6b3940560cce77.jpg) + +![](images/5cbbdeb4e015eb7e17752df75e066f79f97b196c2cdd173dd34c0bee4ab99cae.jpg) + +![](images/59f646129b033fb3d6dbe7bd5ab2ac10b5b314c44699b213dd4fd3ee0eedb616.jpg) + +![](images/dcc9afcb05052c71dc4744e7f72be33a388d4dddf8cef728aa8cce6bfba38f1d.jpg) + +![](images/d67246c6c6eab4dd9d1b519b6cf39af249e52d27426117538de9e66b0e8bd6e6.jpg) +Figure 3: Attack success rates $(\%)$ of eight deep models on the adversarial examples crafted on each model by TIM, DIM, DEM, Admix, SSA, and SIA. + +![](images/f54a3fe1438516f6c9521f3a412c32892ae56a0e459311a138fca1f5a321df98.jpg) + +![](images/122fafbe6e052d31c12ec2f3c98b9b7154f783fa392057d75f3c94d88b15b420.jpg) + +![](images/1614697564ff397200f7b4c57ae65a540dfaf361dc40d2567d5a34ba6e9a5ba0.jpg) + +# 3.3. SIT vs. AutoAugment + +Both SIT and AutoAugment [9] leverage several image transformations to pre-process the images. We summarize the difference as follows: + +- SIT improves the adversarial transferability when attacking the model while AutoAugment boosts the model generalization during the training process. +- SIT randomly samples the transformation while AutoAugment needs to search for a good policy for each transformation on each dataset. +- SIT applies various transformations on different image blocks locally but preserves the global structure. On the contrary, AutoAugment applies two transformations sequentially on the image. +- The fine-grained transformations by SIT generate more diverse images than AutoAugment. + +# 4. Experiment + +In this section, we conduct extensive evaluations on ImageNet dataset to validate the effectiveness of SIA. + +# 4.1. Experimental Setting + +Dataset. To align with previous works [10, 32, 53, 55], we randomly sample 1,000 images pertaining to 1,000 categories from the ILSVRC 2012 validation set [40], which are correctly classified by the adopted models. + +Baselines. We compare our proposed SIA with five competitive input transformation based attacks, namely DIM [64], TIM [11], DEM [81], Admix [55], and SSA [36], which are integrated into MI-FGSM [10]. We also integrate the baseline methods with two model-specific approaches, namely LinBP [19] and SGM [60]. + +Victim Models. We evaluate the attack performance on two popular model architectures, namely Convolutional Network Works, i.e., ResNet-18 [20], ResNet-101 [20], ResNext-50 [65], DenseNet-121 [23], MobileNet [21], and Transformers, i.e., Vision Transformer (ViT) [12] and Swin Transformer (Swin) [34]. Furthermore, we study several SOTA defense methods, including one adversarial training method, i.e., ensemble adversarially trained model (Incv3_ens) [48], the top-3 submissions in NIPS 2017 defense competition, i.e., high-level representation guided denoiser (HGD) [30], random resizing and padding (R&P) [63] and NIPS-r3 $^1$ , three input pre-processing based defenses, namely FD [35], JPEG [18] and Bit-Red [67], a certified defense, i.e., randomized smoothing (RS) [8] and a deep denoiser, i.e., neural representation purifier (NRP) [39]. + +Evaluation Settings. We follow MI-FGSM [10] with the perturbation budget $\epsilon = 16$ , number of iteration $T = 10$ , step size $\alpha = \epsilon / T = 1.6$ and decay factor $\mu = 1$ . DIM [64] adopts the transformation probability of 0.5 and TIM [11] utilizes the Gaussian kernel with the size of $7 \times 7$ . DEM [81] uses the resize ratios: [1.14, 1.27, 1.4, 1.53, 1.66]. Admix admixes 3 images from other categories with the strength of 0.2 and 5 scaled images for each admixed image. SSA [36] sets the turning factor as 0.5 and the standard deviation as $\epsilon$ . SIA sets the splitting number $s = 3$ and number of transformed images for gradient calculation $N = 20$ . + +# 4.2. Attacking a Single Model + +To validate the effectiveness of the proposed SIA, we first compare SIA with five SOTA input transformation based attacks, namely DIM, TIM, DEM, Admix and SSA. + +![](images/5a96ed0444eb2a1cd0d72c197d5df10483bbf76d3b5ae1e31882ce320d0f3513.jpg) +Figure 4: Attack success rates $(\%)$ of eight models on the adversarial examples generated on ResNet-18 when integrating TIM, DIM, DEM, Admix, SSA, and SIA into SGM, respectively. + +We generate the adversarial examples on a single model and test them on the other models. The attack success rates, i.e., the misclassification rates of the victim model on the crafted adversarial examples, are summarized in Fig. 3. + +From the figure, we can observe that all the attackers can achieve the attack success rate of $100.0\%$ or near $100.0\%$ , showing that the input transformation based attacks do not degrade the white-box attack performance. As for the black-box performance, TIM exhibits the poorest transferability on these normally trained models. Admix consistently achieves better transferability than DIM and DEM on CNN-based models. Surprisingly, DIM achieves even better transferability than Admix when generating adversarial examples on transformer-based models. SSA exhibits the best transferability among the baselines in most cases. Also, even for the transformer-based models, the adversarial examples generated on Swin show the poorest transferability on ViT than other CNN-based models. Hence, we argue that it is necessary to evaluate the effectiveness of transfer-based attacks on both CNN-based and transformer-based models. We need an in-depth analysis of the transferability among the emerging transformer-based models. Compared with the baselines, SIA consistently performs much better than the best baselines on all eight models with different architectures. In particular, SIA outperforms the winner-up method with a clear margin of $14.3\%$ on average and achieves an attack success rate higher than the best baseline of at least $2.8\%$ on all the models. Such consistent and superior performance demonstrates that the proposed SIA is general to various model architectures (either CNN or transformer) to boost transferability effectively. + +![](images/cc97c92d2c10f857ec943002cdca9f8a15070a36900b2413f607548f164c8cde.jpg) +Figure 5: Attack success rates $(\%)$ of eight models on the adversarial examples generated on ResNet-18 when integrating TIM, DIM, DEM, Admix, SSA, and SIA into LinBP, respectively. + +# 4.3. Integration to model-specific approaches + +To further verify the scalability of the proposed SIA, we integrate existing input transformation-based attacks into two model-specific approaches, namely SGM and LinBP. The adversarial examples are generated on ResNet-18 and test them on the other models. + +As depicted in Fig. 4 and Fig. 5, SGM and LinBP can significantly boost the adversarial transferability of vallina input transformation-based attacks. Compared with the baselines, SIA still achieves much better transferability, which has been boosted by an average margin of $3.3\%$ and $3.6\%$ when integrated into SGM and LinBP, respectively. These superior results validate the generality of SIA to various transfer-based attacks and underscore the potential of SIA in augmenting adversarial transferability through the fusion of different strategies. + +# 4.4. Attacking Ensemble Models + +Liu et al. [33] have shown that attacking ensemble models can effectively improve transferability. As shown in Sec. 4.2, the adversarial examples often exhibit poorer transferability across the CNN-based models and transformer-based models than being transferred among the CNN-based models. Hence, we generate adversarial examples on ensemble CNN-based and ensemble transformer-based models and test them on the remaining models to evaluate SIA when attacking ensemble models. + +As shown in Fig. 6, when attacking ensemble models, the adversarial examples generated by all attacks exhibit better transferability than that crafted on a single model, showing the excellent compatibility of all the input trans + +![](images/37dd8323c287535f2d89323967e6331264bca7bdfce575d7a8c3ac27a928c6a5.jpg) +Figure 6: Attack success rates (\%) of eight models on the adversarial examples crafted on ensemble models by TIM, DIM, DEM, Admix, SSA, and SIA, respectively. We generate adversarial examples on the CNN-based models and test them on the transformer-based models, and vice versa. + +formation based attacks with such a setting. When generating the adversarial examples on ViT and Swin, DIM and DEM achieve better transferability than Admix and SSA, highlighting the difference between crafting adversarial examples on different architectures and the necessity to evaluate the transferability on these models. On all eight models, SIA achieves the attack success rate of at least $88.3\%$ . It outperforms the winner-up approach with a clear margin of $6.5\%$ , showing its superior effectiveness in generating transferable adversarial examples. In particular, SIA achieves $94.9\%$ and $99.3\%$ attack success rates on ViT and Swin, respectively, when the adversarial examples are generated on CNN-based models without access to the attention module in the transformer. This further supports our motivation that improving the diversity of transformed images can significantly boost transferability, even on models with completely different architectures. + +# 4.5. Attacking Defense Methods + +SIA has achieved superior attack performance on eight normally trained models with different architectures when attacking single model as well as ensemble models. Recently, several defenses have been proposed to mitigate the threat of adversarial examples on ImageNet dataset. To validate the effectiveness of these defenses, we adopt the adversarial examples generated on these eight models simultaneously to attack the defense models, including Inc-v3ens, HGD, R&P, NIPS-r3, FD, JPEG, Bit-Red, RS and NRP. + +The attack results on these defense methods are summarized in Fig. 7. Overall, DIM is on par with TIM on + +![](images/c00f382c99a1449f97cc2353af5ef232d0f76e7c9a0e397eee1ab0e5e7b8e1fc.jpg) +Figure 7: Attack success rates $(\%)$ of various defense methods on the adversarial examples generated by TIM, DIM, DEM, Admix, SSA, and SIA under ensemble model setting. The adversarial examples are generated on the eight models simultaneously. + +these defense methods while Admix can consistently outperform the other four baselines. SIA can consistently achieve better transferability than the baselines. In particular, SIA achieves the attack success rate of $78.2\%$ on the certified defense method (i.e., RS), and SIA can achieve the attack success rate of at least $89.9\%$ on the other eight defense methods, including the powerful denoising method NRP. Such high attack performance indicates the insufficiency of existing defense methods and raises a new security issue to designing more robust deep learning models. + +# 4.6. Ablation Studies + +To further gain insights on the superior attack performance achieved by SIA, we conduct a series of ablation studies to validate that all the transformations and splitting the image into blocks can help improve transferability. We generate the adversarial examples on ResNet-18 and test them on the other seven models for all the experiments. + +Are all the transformations beneficial for boosting the transerability? There are ten different input transformations that SIA might apply to each image block. To investigate whether each transformation is of benefit to generate more transferable adversarial examples, we first adopt one or two transformations to implement SIA. The results are summarized in Fig. 8. The average attack success rate of MI-FGSM is $59.7\%$ . On the diagonal line of Fig. 8, we only adopt a single transformation for each block and the lowest average attack success rate is $67.3\%$ by adopting DCT, which outperforms MI-FGSM with a margin of $7.6\%$ . This demonstrates the high effectiveness of SIA and + +
SIA-VShift-HShift-VFlip-HFlip-Rotate-Sclae-Add Noise-Resize-DCT-Dropout
92.189.790.190.188.488.390.190.690.290.190.7
+ +Table 3: The average attack success rates of adversarial examples crafted by SIA and SIA without a single transformation. The adversaries are generated on ResNet-18 and tested on the other seven deep models. - indicates removing such transformation. + +![](images/19e5b9ac816739d615e5847af4b383f2ee4ef7eb81f0ab022e00b030afb6650e.jpg) +Figure 8: The average attack success rates $(\%)$ on the adversarial examples generated by SIA using one (diagonal) or two (other) transformations. + +its generality to various transformations. We can also observe that the transferability can be further improved when we combine any two transformations, showing the advantage of combining these transformations. To further validate the necessity of each transformation, we generate adversarial examples by SIA without each transformation and summarize the results in Tab. 3. Since removing a single transformation does not significantly decrease the diversity of the transformation, these attacks achieve similar attack performance. However, no matter which transformation is removed, the transferability will be decreased, supporting that each transformation is of benefit to generate more transferable adversarial examples. + +Are the image blocks beneficial for boosting transferability? We have shown that the abundant transformations can effectively improve transferability. Here we further explore whether it is useful to apply these transformations to the image block instead of the raw image. For comparison, we randomly apply the transformation to the raw image $s \times s$ times, denoted as $\mathrm{SIA_r}$ . The results are reported in Fig. 9. We can observe that $\mathrm{SIA_r}$ exhibits better transferability than MI-FGSM, which also validates our motivation that improving the diversity of transformed images can generate more transferable adversarial examples. Our SIA consistently performs better than $\mathrm{SIA_r}$ , indicating that applying + +![](images/aa9c8322747df196024fc159650aaad0713635eaa7a62acf81b178ded62941fc.jpg) +Figure 9: The average attack success rates $(\%)$ of adversarial examples generated by SIA and $\mathrm{SIA_r}$ (randomly transform the image without the block partition). + +the transformation onto each image block is of great benefit to improve transferability. + +# 4.7. Parameter Studies + +In this subsection, we conduct parameter studies to explore the impact of two hyper-parameters, namely the number of blocks $s$ and the number of transformed images for gradient calculation $N$ . All the adversarial examples are generated on ResNet-18 and tested on the other seven models in the black-box setting. + +On the number of blocks $s$ . The number of blocks determines how diverse the transformed images are, which can influence the attack performance. To find a good value for $s$ , we conduct SIA with $s$ from 1 to 5 and summarize the attack results in Fig. 10. When $s \leq 3$ , increasing the value of $s$ leads to better diversity, which can improve the attack performance. However, when we continually increase the value of $s$ , the diversity is increased, but it also introduces more variance to the gradient, decaying the attack performance slightly. Hence, we adopt $s = 3$ to balance the diversity of images and variance of the gradient for better attack performance. + +On the number of images for gradient calculation $N$ . SIA calculates the average gradient on $N$ images to eliminate the variance introduced by the transformation. To de + +![](images/5f0f3974934d2ffa05e95e725d521c3807bf3446061ac766eeef25ab5cb64921.jpg) +Figure 10: Attack success rates $(\%)$ of adversarial examples generated by SIA with various number of blocks $s$ + +termine a good value for $N$ , we evaluate SIA with $N$ from 1 to 30 and report the attack success rates in Fig. 11. As we can observe, when $N = 1$ , SIA introduces massive variance into the gradient and achieves the lowest performance on all models. When we increase the value of $N$ , the variance among the gradients can be eliminated, and SIA achieves better transferability before $N = 20$ . When $N > 20$ , increasing $N$ can only introduce computation cost without performance improvement. To balance the attack performance and computation cost, we adopt $N = 20$ in our experiments. + +# 5. Conclusion + +In this work, we find that the existing input transformation based attack with better transferability often generates more diverse transformed images. Based on this finding, we design a new image transformation, called structure invariant transformation (SIT), which splits the image into several blocks and randomly transforms each image block independently. With such image transformation, we propose a novel input transformation based attack, dubbed structure invariant attack (SIA), which calculates the average gradient on several transformed images by SIT to update the perturbation. Extensive evaluations demonstrate that SIA can achieve remarkably better transferability than the existing SOTA attacks. In our opinion, SIA provides a new direction by applying the transformation onto the image block to effectively boost transferability, which sheds new light on generating more transferable adversarial examples with fine-grained transformations. + +![](images/750e2b03d21a661c6217aa6a5326ca38cfd9f1cf0709e405b1b53368430649ef.jpg) +Figure 11: Attack success rates $(\%)$ of adversarial examples generated by SIA with various number of images $N$ + +# References + +[1] Moustafa Alzantot, Yash Sharma, Ahmed Elgohary, BoJhang Ho, Mani B. Srivastava, and Kai-Wei Chang. Generating Natural Language Adversarial Examples. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 2890-2896, 2018. 2 +[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square Attack: A Query-Efficient Black-Box Adversarial Attack via Random Search. In Proceedings of the European Conference on Computer Vision, pages 484–501, 2020. 2 +[3] Anish Athalye, Nicholas Carlini, and David A. Wagner. Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. In Proceedings of the International Conference on Machine Learning, pages 274–283, 2018. 2, 3 +[4] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models. In Proceedings of the International Conference on Learning Representations, 2018. 1, 2 +[5] Nicholas Carlini and David A. Wagner. Audio Adversarial Examples: Targeted Attacks on Speech-to-Text. In Proceedings of the IEEE Symposium on Security and Privacy (Workshops), pages 1-7, 2018. 2 +[6] Minhao Cheng, Thong Le, Pin-Yu Chen, Huan Zhang, Jinfeng Yi, and Cho-Jui Hsieh. Query-Efficient Hard-label Black-box Attack: An Optimization-based Approach. In Proceedings of the International Conference on Learning Representations, 2019. 2 +[7] Shuyu Cheng, Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Improving Black-box Adversarial Attacks with a Transfer-based Prior. In Proceedings of the Advances in Neural Information Processing Systems, pages 10932-10942, 2019. 1 + +[8] Jeremy M. Cohen, Elan Rosenfeld, and J. Zico Kolter. Certified Adversarial Robustness via Randomized Smoothing. In Proceedings of the International Conference on Machine Learning, pages 1310-1320, 2019. 3, 5 +[9] Ekin D. Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, and Quoc V. Le. AutoAugment: Learning Augmentation Strategies From Data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 3, 5 +[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting Adversarial Attacks With Momentum. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 1, 2, 4, 5 +[11] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading Defenses to Transferable Adversarial Examples by Translation-Invariant Attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 1, 2, 5 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In Proceedings of the International Conference on Learning Representations, 2021. 1, 5 +[13] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust Physical-World Attacks on Deep Learning Visual Classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1625–1634, 2018. 1 +[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we Ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1 +[15] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and Harnessing Adversarial Examples. In Proceedings of the International Conference on Learning Representations, 2015. 1, 2, 3 +[16] Sven Gowal, Krishnamurthy Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timothy Arthur Mann, and Pushmeet Kohli. Scalable Verified Training for Provably Robust Image Classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4841-4850, 2019. 3 +[17] Chuan Guo, Jacob R. Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Q. Weinberger. Simple Black-box Adversarial Attacks. In Proceedings of the International Conference on Machine Learning, pages 2484-2493, 2019. 2 +[18] Chuan Guo, Mayank Rana, Moustapha Cissé, and Laurens van der Maaten. Countering Adversarial Images Using Input Transformations. In Proceedings of the International Conference on Learning Representations, 2018. 5 +[19] Yiwen Guo, Qizhang Li, and Hao Chen. Backpropagating linearly improves transferability of adversarial examples. + +Advances in neural information processing systems, 33:85-95, 2020. 3, 5 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1, 3, 5 +[21] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv preprint arXiv:1704.04861, 2017. 5 +[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-Eccitation Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 3 +[23] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely Connected Convolutional Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2261-2269, 2017. 1, 5 +[24] Andrew Ilyas, Logan Engstrom, Anish Athalye, and Jessy Lin. Black-box Adversarial Attacks with Limited Queries and Information. In Proceedings of the International Conference on Machine Learning, pages 2142-2151, 2018. 1, 2 +[25] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Proceedings of the Advances in Neural Information Processing Systems, pages 1106-1114, 2012. 1, 3 +[26] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial Examples in the Physical World. In Proceedings of the International Conference on Learning Representations (Workshops), 2017. 1, 2 +[27] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. QEBA: Query-Efficient Boundary-Based Blackbox Attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1218-1227, 2020. 1, 2 +[28] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan L. Yuille. Learning Transferable Adversarial Examples via Ghost Networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11458-11465, 2020. 1, 3 +[29] Yandong Li, Lijun Li, Liqiang Wang, Tong Zhang, and Boqing Gong. NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks. In Proceedings of the International Conference on Machine Learning, pages 3866–3876, 2019. 2 +[30] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense Against Adversarial Attacks Using High-Level Representation Guided Denoiser. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2018. 3, 5 +[31] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and + +Daan Wierstra. Continuous Control with Deep Reinforcement Learning. In Proceedings of the International Conference on Learning Representations, 2016. 1 +[32] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov Accelerated Gradient and Scale Invariance for Adversarial Attacks. In Proceedings of the International Conference on Learning Representations, 2020. 1, 2, 4, 5 +[33] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into Transferable Adversarial Examples and Blackbox Attacks. In Proceedings of the International Conference on Learning Representations, 2017. 1, 2, 6 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9992-10002, 2021. 5 +[35] Zihao Liu, Qi Liu, Tao Liu, Nuo Xu, Xue Lin, Yanzhi Wang, and Wujie Wen. Feature Distillation: DNN-Oriented JPEG Compression Against Adversarial Examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 860-868, 2019. 5 +[36] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xi-anglong Liu, Jian Zhang, and Jingkuan Song. Frequency Domain Model Augmentation for Adversarial Attack. In Proceedings of the European Conference on Computer Vision, pages 549–566, 2022. 1, 2, 5 +[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards Deep Learning Models Resistant to Adversarial Attacks. In Proceedings of the International Conference on Learning Representations, 2018. 1, 2 +[38] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2574-2582, 2016. 1, 2 +[39] Muzammal Naseer, Salman H. Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A Self-supervised Approach for Adversarial Robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 259-268, 2020. 3, 5 +[40] Russakovsky Olga, Deng Jia, Su Hao, Krause Jonathan, Satheesh Sanjeev, Ma Sean, hUANG Zhiheng, Karpathy Andrej, Khosla Adiya, and Bernstein Michael et al. Imagenet large scale visual recognition challenge. In International Journal of Computer Vision, pages 211-252, 2015. 5 +[41] Florian Schroff, Dmitry Kalenichenko, and James Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 1 +[42] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K. Reiter. Accessorize to a Crime: Real and Stealthy Attacks on State-of-the-Art Face Recognition. In Proceedings of the ACM Conference on Computer and Communications Security, pages 1528-1540, 2016. 1 + +[43] Luchuan Song, Zheng Fang, Xiaodan Li, Xiaoyi Dong, Zhenchao Jin, Yuefeng Chen, and Siwei Lyu. Adaptive face forgery detection in cross domain. In European Conference on Computer Vision, pages 467-484. Springer, 2022. 1 +[44] Luchuan Song, Xiaodan Li, Zheng Fang, Zhenchao Jin, YueFeng Chen, and Chenliang Xu. Face forgery detection via symmetric transformer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4102-4111, 2022. 1 +[45] Luchuan Song, Bin Liu, Guojun Yin, Xiaoyi Dong, Yufei Zhang, and Jia-Xuan Bai. Tacr-net: editing on deep video and voice portraits. In Proceedings of the 29th ACM International Conference on Multimedia, pages 478-486, 2021. 1 +[46] Nitish Srivastava, Geoffrey E. Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a Simple Way to prevent Neural Networks from Overfitting. The Journal of Machine Learning Research, 15(1):1929-1958, 2014. 3 +[47] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing Properties of Neural Networks. In Proceedings of the International Conference on Learning Representations, 2014. 1, 2 +[48] Florian Tramèr, Alexey Kurakin, Nicolas Papernot, Ian J. Goodfellow, Dan Boneh, and Patrick D. McDaniel. Ensemble Adversarial Training: Attacks and Defenses. In Proceedings of the International Conference on Learning Representations, 2018. 3, 5 +[49] Jonathan Uesato, Brendan O'Donoghue, Pushmeet Kohli, and Aïron van den Oord. Adversarial Risk and the Dangers of Evaluating Against Weak Attacks. In Proceedings of the International Conference on Machine Learning, pages 5032-5041, 2018. 2 +[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Proceedings of the Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 1 +[51] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. CosFace: Large Margin Cosine Loss for Deep Face Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5265-5274, 2018. 1 +[52] Xiaosen Wang, Jin Hao, Yichen Yang, and Kun He. Natural Language Adversarial Defense through Synonym Encoding. In Proceedings of the Conference on Uncertainty in Artificial Intelligence, pages 823-833, 2021. 2 +[53] Xiaosen Wang and Kun He. Enhancing the Transferability of Adversarial Attacks Through Variance Tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 1, 2, 5 +[54] Xiaosen Wang, Kun He, Chuanbiao Song, Liwei Wang, and John E Hopcroft. AT-GAN: An Adversarial Generator Model for Non-constrained Adversarial Examples. arXiv preprint arXiv:1904.07793, 2019. 1 +[55] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the Transferability of Adversarial At + +tacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16138-16147, 2021. 1, 2, 3, 4, 5 +[56] Xiaosen Wang, Jiadong Lin, Han Hu, Jingdong Wang, and Kun He. Boosting Adversarial Transferability through Enhanced Momentum. In Proceedings of the British Machine Vision Conference, page 272, 2021. 1, 2 +[57] Xiaosen Wang, Zeliang Zhang, Kangheng Tong, Dihong Gong, Kun He, Zhifeng Li, and Wei Liu. Triangle Attack: A Query-Efficient Decision-Based Adversarial Attack. In Proceedings of the European Conference on Computer Vision, pages 156-174, 2022. 1, 2 +[58] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable Adversarial Attacks for Image and Video Object Detection. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 954-960, 2019. 2 +[59] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is Better than Free: Revisiting Adversarial Training. In Proceedings of the International Conference on Learning Representations, 2020. 3 +[60] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip Connections Matter: On the Transferability of Adversarial Examples Generated with ResNets. In Proceedings of the International Conference on Learning Representations, 2020. 1, 3, 5 +[61] Weibin Wu, Yuxin Su, Xixian Chen, Shenglin Zhao, Irwin King, Michael R. Lyu, and Yu-Wing Tai. Boosting the Transferability of Adversarial Samples via Attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1158-1167, 2020. 1, 2 +[62] Weibin Wu, Yuxin Su, Michael R Lyu, and Irwin King. Improving the Transferability of Adversarial Samples with Adversarial Transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9024-9033, 2021. 2 +[63] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan L. Yuille. Mitigating Adversarial Effects Through Randomization. In Proceedings of the International Conference on Learning Representations, 2018. 3, 5 +[64] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving Transferability of Adversarial Examples With Input Diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 1, 2, 3, 5 +[65] Saining Xie, Ross B. Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated Residual Transformations for Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 5 +[66] Yifeng Xiong, Jiadong Lin, Min Zhang, John E. Hopcroft, and Kun He. Stochastic Variance Reduced Ensemble Adversarial Attack for Boosting the Adversarial Transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14972, 2022. 1, 2 +[67] Weilin Xu, David Evans, and Yanjun Qi. Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks. + +In Proceedings of the Network and Distributed System Security Symposium, 2018. 5 +[68] Bo Yang, Hengwei Zhang, Zheming Li, Yuchen Zhang, Kaiyong Xu, and Jindong Wang. Adversarial example generation with adbelief optimizer and crop invariance. Applied Intelligence, 53(2):2332-2347, 2023. 1 +[69] Yichen Yang, Xiaosen Wang, and Kun He. Robust Textual Embedding against Word-level Adversarial Attacks. pages 2214-2224, 2022. 2 +[70] Zhuolin Yang, Bo Li, Pin-Yu Chen, and Dawn Song. Characterizing Audio Adversarial Examples Using Temporal Dependency. In Proceedings of the International Conference on Learning Representations, 2019. 2 +[71] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural Color Fool: Towards Boosting Black-box Unrestricted Attacks. In Proceedings of the Advances in Neural Information Processing Systems, 2022. 1 +[72] Zheng Yuan, Jie Zhang, and Shiguang Shan. Adaptive image transformations for transfer-based adversarial attack. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1 +[73] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6022-6031, 2019. 3 +[74] Huan Zhang, Hongge Chen, Chaowei Xiao, Sven Gowal, Robert Stanforth, Bo Li, Duane S. Boning, and Cho-Jui Hsieh. Towards Stable and Efficient Training of Verifiably Robust Neural Networks. In Proceedings of the International Conference on Learning Representations, 2020. 3 +[75] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. Mixup: Beyond Empirical Risk Minimization. In Proceedings of the International Conference on Learning Representations, 2018. 3 +[76] Jianping Zhang, Jen-tse Huang, Wenxuan Wang, Yichen Li, Weibin Wu, Xiaosen Wang, Yuxin Su, and Michael R. Lyu. Improving the Transferability of Adversarial Samples by Path-Augmented Method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8173-8182, 2023. 1 +[77] Qilong Zhang, Xiaodan Li, Yuefeng Chen, Jingkuan Song, Lianli Gao, Yuan He, and Hui Xue. Beyond ImageNet Attack: Towards Crafting Adversarial Examples for Black-box Domains. In Proceedings of the International Conference on Learning Representations, 2022. 1 +[78] Qilong Zhang, Chaoning Zhang, Chaoqun Li, Jingkuan Song, Lianli Gao, and Heng Tao Shen. Practical No-box Adversarial Attacks with Training-free Hybrid Image Transformation. arXiv preprint arXiv:2203.04607, 2022. 1 +[79] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 3 +[80] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable Ad + +versarial Perturbations. In Proceedings of the European Conference on Computer Vision, pages 471-486, 2018. 1, 2 +[81] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the Transferability of Adversarial Examples with Resized-Diverse-Inputs, Diversity-Ensemble and Region Fitting. In Proceedings of the European Conference on Computer Vision, pages 563–579, 2020. 2, 5 \ No newline at end of file diff --git a/2023/Structure Invariant Transformation for better Adversarial Transferability/images.zip b/2023/Structure Invariant Transformation for better Adversarial Transferability/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eb387383f3273efcb226220b891e52762a5f93ec --- /dev/null +++ b/2023/Structure Invariant Transformation for better Adversarial Transferability/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8d6cb50e2c9b99d25d33d1a69f979b30c38a1233a71d1109c73eecda209652f +size 501021 diff --git a/2023/Structure Invariant Transformation for better Adversarial Transferability/layout.json b/2023/Structure Invariant Transformation for better Adversarial Transferability/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..597bfc96142ddf572923bdcaac503eca5a34a39e --- /dev/null +++ b/2023/Structure Invariant Transformation for better Adversarial Transferability/layout.json @@ -0,0 +1,10903 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 524, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 524, + 121 + ], + "type": "text", + "content": "Structure Invariant Transformation for better Adversarial Transferability" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 144, + 222, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 144, + 222, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 144, + 222, + 171 + ], + "type": "text", + "content": "Xiaosen Wang \nHuawei Singularity Security Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 175, + 194, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 175, + 194, + 184 + ], + "spans": [ + { + "bbox": [ + 88, + 175, + 194, + 184 + ], + "type": "text", + "content": "xiaosen@hust.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 256, + 144, + 330, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 144, + 330, + 171 + ], + "spans": [ + { + "bbox": [ + 256, + 144, + 330, + 171 + ], + "type": "text", + "content": "Zeliang Zhang HUST" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 244, + 175, + 343, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 175, + 343, + 185 + ], + "spans": [ + { + "bbox": [ + 244, + 175, + 343, + 185 + ], + "type": "text", + "content": "hust0426@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 366, + 144, + 531, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 144, + 531, + 171 + ], + "spans": [ + { + "bbox": [ + 366, + 144, + 531, + 171 + ], + "type": "text", + "content": "Jianping Zhang Chinese University of Hong Kong" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 387, + 175, + 512, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 175, + 512, + 185 + ], + "spans": [ + { + "bbox": [ + 387, + 175, + 512, + 185 + ], + "type": "text", + "content": "jpzhang@cse.cuhk.edu.hk" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 238, + 290, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 562 + ], + "type": "text", + "content": "Given the severe vulnerability of Deep Neural Networks (DNNs) against adversarial examples, there is an urgent need for an effective adversarial attack to identify the deficiencies of DNNs in security-sensitive applications. As one of the prevalent black-box adversarial attacks, the existing transfer-based attacks still cannot achieve comparable performance with the white-box attacks. Among these, input transformation based attacks have shown remarkable effectiveness in boosting transferability. In this work, we find that the existing input transformation based attacks transform the input image globally, resulting in limited diversity of the transformed images. We postulate that the more diverse transformed images result in better transferability. Thus, we investigate how to locally apply various transformations onto the input image to improve such diversity while preserving the structure of image. To this end, we propose a novel input transformation based attack, called Structure Invariant Transformation (SIA), which applies a random image transformation onto each image block to craft a set of diverse images for gradient calculation. Extensive experiments on the standard ImageNet dataset demonstrate that SIA exhibits much better transferability than the existing SOTA input transformation based attacks on CNN-based and transformer-based models, showing its generality and superiority in boosting transferability. Code is available at https://github.com/xiaosen-wang/SIT." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "With the unprecedented progress of Deep Neural Networks (DNNs) [25, 20, 23, 50, 12], they have been deployed in many security-sensitive applications, such as face recognition [41, 51, 43, 44], autonomous driving [14, 31], etc. On the other hand, recent works have found that DNNs are vulnerable to adversarial examples [47, 15], which mislead the deep models with imperceptible perturbations. This brings a huge threat to the real-world applications [42, 13, 45, 57, 71, 78] and makes it imperative for an" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 239, + 365, + 291 + ], + "blocks": [ + { + "bbox": [ + 315, + 239, + 365, + 291 + ], + "lines": [ + { + "bbox": [ + 315, + 239, + 365, + 291 + ], + "spans": [ + { + "bbox": [ + 315, + 239, + 365, + 291 + ], + "type": "image", + "image_path": "af123078fe2f1c8acaac5afc6622d7ea8abf93cff6c1319afb1260a5c7c00e41.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 295, + 362, + 306 + ], + "lines": [ + { + "bbox": [ + 318, + 295, + 362, + 306 + ], + "spans": [ + { + "bbox": [ + 318, + 295, + 362, + 306 + ], + "type": "text", + "content": "Raw Image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 372, + 213, + 424, + 264 + ], + "blocks": [ + { + "bbox": [ + 372, + 213, + 424, + 264 + ], + "lines": [ + { + "bbox": [ + 372, + 213, + 424, + 264 + ], + "spans": [ + { + "bbox": [ + 372, + 213, + 424, + 264 + ], + "type": "image", + "image_path": "ca181908d16984ee20827d0d4bd15ed460b59041280cf6de4dfd54f53a97fc51.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 268, + 406, + 277 + ], + "lines": [ + { + "bbox": [ + 388, + 268, + 406, + 277 + ], + "spans": [ + { + "bbox": [ + 388, + 268, + 406, + 277 + ], + "type": "text", + "content": "DIM" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 372, + 279, + 424, + 330 + ], + "blocks": [ + { + "bbox": [ + 372, + 279, + 424, + 330 + ], + "lines": [ + { + "bbox": [ + 372, + 279, + 424, + 330 + ], + "spans": [ + { + "bbox": [ + 372, + 279, + 424, + 330 + ], + "type": "image", + "image_path": "0f1fc384044917aac920f5630552beba2b0d1f571134445f3b0d1632c3e5883e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 334, + 411, + 343 + ], + "lines": [ + { + "bbox": [ + 384, + 334, + 411, + 343 + ], + "spans": [ + { + "bbox": [ + 384, + 334, + 411, + 343 + ], + "type": "text", + "content": "Admix" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 429, + 213, + 481, + 264 + ], + "blocks": [ + { + "bbox": [ + 429, + 213, + 481, + 264 + ], + "lines": [ + { + "bbox": [ + 429, + 213, + 481, + 264 + ], + "spans": [ + { + "bbox": [ + 429, + 213, + 481, + 264 + ], + "type": "image", + "image_path": "832227d53516e0694060b525fa38b28240182fae977294e494752ed3bb079bdb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 267, + 463, + 276 + ], + "lines": [ + { + "bbox": [ + 446, + 267, + 463, + 276 + ], + "spans": [ + { + "bbox": [ + 446, + 267, + 463, + 276 + ], + "type": "text", + "content": "TIM" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 429, + 279, + 481, + 330 + ], + "blocks": [ + { + "bbox": [ + 429, + 279, + 481, + 330 + ], + "lines": [ + { + "bbox": [ + 429, + 279, + 481, + 330 + ], + "spans": [ + { + "bbox": [ + 429, + 279, + 481, + 330 + ], + "type": "image", + "image_path": "a08afce371b4e17f65b89af169699fe7de41538ea190d48475a62e164d0a7c57.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 334, + 463, + 343 + ], + "lines": [ + { + "bbox": [ + 446, + 334, + 463, + 343 + ], + "spans": [ + { + "bbox": [ + 446, + 334, + 463, + 343 + ], + "type": "text", + "content": "SSA" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "text", + "content": "Figure 1: The raw image and its transformed images by DIM (resize factor of 0.8) [64], TIM (translated by 15 pixels) [11], SIM (scale factor of 0.5), Admix (admix strength of 0.2 and scale factor of 0.5) [55], SSA (turning factor of 0.5) [36] and our proposed SIA " + }, + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "inline_equation", + "content": "(3\\times 3" + }, + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "text", + "content": " blocks)." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 487, + 213, + 538, + 264 + ], + "blocks": [ + { + "bbox": [ + 487, + 213, + 538, + 264 + ], + "lines": [ + { + "bbox": [ + 487, + 213, + 538, + 264 + ], + "spans": [ + { + "bbox": [ + 487, + 213, + 538, + 264 + ], + "type": "image", + "image_path": "5530125076220063c78af0a9df28ab9cc4112e97291a9315c8250c58c5384086.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 268, + 520, + 276 + ], + "lines": [ + { + "bbox": [ + 504, + 268, + 520, + 276 + ], + "spans": [ + { + "bbox": [ + 504, + 268, + 520, + 276 + ], + "type": "text", + "content": "SIM" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 487, + 279, + 538, + 330 + ], + "blocks": [ + { + "bbox": [ + 487, + 279, + 538, + 330 + ], + "lines": [ + { + "bbox": [ + 487, + 279, + 538, + 330 + ], + "spans": [ + { + "bbox": [ + 487, + 279, + 538, + 330 + ], + "type": "image", + "image_path": "dc2dddd99efc9d96aace3011f725c70e0ae0898deed9d4f361dbc4b4cb077e95.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 334, + 533, + 344 + ], + "lines": [ + { + "bbox": [ + 491, + 334, + 533, + 344 + ], + "spans": [ + { + "bbox": [ + 491, + 334, + 533, + 344 + ], + "type": "text", + "content": "SIA (Ours)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 437, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 473 + ], + "type": "text", + "content": "effective attack to identify the deficiencies of DNNs when we robustify the deep models or deploy them for commercial applications." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 473, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 546, + 617 + ], + "type": "text", + "content": "Existing adversarial attacks usually fall into two categories: white-box attacks [15, 38, 37, 26, 54] can fetch any information of the target model, including (hyper-)parameters, gradient, architecture, while black-box attacks [24, 7, 4, 27, 33, 77] are only allowed limited access to the target model. One of the significant properties of adversarial examples is their transferability [64, 10, 53, 76], in which the adversarial examples generated on one model can still mislead other models, making it possible to attack the real-world applications in the black-box setting. However, existing adversarial attacks [26, 37] often exhibit superior white-box attack performance but poor transferability." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": "To craft more transferable adversarial examples, various techniques have been proposed, such as momentum-based methods [10, 32, 53, 56], input transformations [64, 11, 32, 68, 72], ensemble attacks [33, 66, 36], advanced objective functions [61, 80] and model-specific approaches [28, 60]. Among which, input transformations (e.g., random resizing and padding [64], translation [11], scale [32] admix [55], etc.) that transform the image before gradient calculation," + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4607" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "have achieved superior transferability and attracted broad attention. Nevertheless, we find that the existing input transformation based attacks, even the SOTA attack Admix [55] and SSA [36], transform the image globally without changing the local relationship among objects in the input image. We argue that the diversity of such transformed images is still not enough, leading to limited transferability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 312 + ], + "type": "text", + "content": "In this work, we postulate and empirically validate that the more diverse transformed images lead to better transferability. Based on this observation, instead of applying a single transformation on the input image, we apply different transformations locally on different parts of the image to enhance the diversity of transformed images. As shown in Fig. 1, such transformation can bring much more difference to the generated image compared with the raw image but still preserve the global structure of the object. Based on this transformation, we propose a novel input transformation based attack called structure invariant attack (SIA), which utilizes the gradient of these transformed images to update the adversarial examples for better transferability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 312, + 244, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 312, + 244, + 323 + ], + "spans": [ + { + "bbox": [ + 59, + 312, + 244, + 323 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 326, + 286, + 517 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 326, + 286, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 286, + 372 + ], + "type": "text", + "content": "- We empirically validate that the high diversity of transformed images is beneficial to improve transferability, which sheds new light on how to design new input transformations for more transferable adversarial examples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 373, + 286, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 286, + 420 + ], + "type": "text", + "content": "- We design a new image transformation method, which locally applies different transformations on different parts of the image to generate more diverse images but preserve its global structure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 422, + 286, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 286, + 468 + ], + "type": "text", + "content": "- Based on the proposed image transformation, we devise a new input transformation based attack, called structure invariant attack (SIA), to generate more transferable adversarial examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 469, + 286, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 286, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 286, + 517 + ], + "type": "text", + "content": "- Extensive experiments on ImageNet dataset demonstrate that SIA outperforms the baselines with a clear margin on CNN-based as well as transformer-based models, showing its superiority and generality." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 526, + 133, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 133, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 133, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": "Adversarial examples [47] have brought an impressive threat to the DNN-enabled applications, such as computer vision [15, 58, 3], natural language processing [1, 52, 69], speech recognition [5, 70], etc. To identify the vulnerability of DNNs, various attack methods have been proposed recently, such as gradient-based attacks [15, 38, 26, 37], score-based attacks [24, 49, 17, 29, 2], decision-based attacks [4, 6, 27, 57] and transfer-based attacks [33, 10, 64, 32, 55]. Among these, transfer-based attacks do not access any information of the target model, making it applicable to attack any model in the physical world. In this work, we focus on generating more transferable adversarial examples and briefly introduce the existing transfer-based attacks, the corresponding defense methods, and data augmentations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 418, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 418, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 418, + 83 + ], + "type": "text", + "content": "2.1. Adversarial Attack" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 209 + ], + "type": "text", + "content": "As the first gradient-based attack, Fast Gradient Sign Method (FGSM) [15] adds the perturbation in the gradient direction to the benign sample, leading to high attack efficiency but limited performance. Later, I-FGSM [26] extends FGSM into an iterative version, which achieves much better white-box attack performance but poor transferability. Given the high efficiency and effectiveness of I-FGSM, numerous transfer-based attacks are proposed to boost transferability based on I-FGSM to attack the deep model in the black-box setting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 211, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 545, + 342 + ], + "type": "text", + "content": "Momentum-based methods. MI-FGSM [10] introduces momentum into I-FGSM to stabilize the optimization direction and escape local maxima. NI-FGSM [32] adopts Nesterov Accelerated Gradient to accumulate the momentum, which achieves better transferability. Variance tuning [53] adopts the gradient variance of the previous iteration to tune the current gradient in MI-FGSM and NI-FGSM, significantly improving transferability. EMIFGSM [56] enhances the momentum by accumulating several samples' gradients in the previous gradient's direction to further stabilize the optimization direction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 342, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 557 + ], + "type": "text", + "content": "Input transformation based attacks. DIM [64] is the first input transformation based attack, which adds padding to a randomly resized image for fixed size before gradient calculation. TIM [11] optimizes the perturbation over an ensemble of translated images, which is further approximated by convolving the gradient at the untranslated image with a pre-defined kernel. SIM [32] scales the images with different scale factors for gradient calculation. DEM [81] averages the gradient on several diverse transformed images similar to DIM with various resizing factors. Admix [55] calculates the gradient on the input image admixed with a small portion of each add-in image from other categories while using the original label of the input. Wu et al. [62] train an adversarial transformation network to destroy the adversarial perturbation and require the synthesized adversarial examples resistant to such transformations. SSA [36] adds Gaussian noise and randomly masks the image in the frequency domain to transform the input image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "text", + "content": "Ensemble attacks. Liu et al. [33] first found that the adversarial examples generated on multiple models denoted as ensemble attack, are more transferable. Recently, Xiong et al. [66] reduce the gradient variance between various models to boost the ensemble attack." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Advanced objective functions. The above attacks often take the cross-entropy loss as the objective function. Researchers also find that some regularizers are beneficial to boost transferability. For instance, Zhou et al. [80] additionally maximize the difference of the intermediate feature maps between the benign sample and adversarial example. Wu et al. [61] adopt a regularizer about the distance of the attention maps between these samples." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4608" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "Model-specific approaches. Some works utilize the surrogate model's architecture to improve transferability. Li et al. [28] densely add dropout [46] after each layer to create several ghost networks for better transferability. SGM [60] adopts more gradient from the skip connection in ResNets [20] to boost adversarial transferability. LinBP [19] replaces the zeros with ones in the derivative of ReLU to make the model more linear, leading to improved transferability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 186, + 165, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 165, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 165, + 198 + ], + "type": "text", + "content": "2.2. Adversarial Defense" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 205, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 456 + ], + "type": "text", + "content": "Numerous adversarial defenses have been proposed to mitigate the threat of adversarial attacks. Adversarial training [15, 48], which adopts the adversarial examples during the training process, has been shown as one of the most effective methods [3] but taking huge computation cost. Recently, Wong et al. [59] find that single-step adversarial examples can bring satisfiable robustness, making it possible to be applied on large-scale datasets (e.g., ImageNet [25]). Pre-processing the input samples before the model is shown to be another effective way. Liao et al. [30] design a high-level representation guided denoiser (HGD) to eliminate the adversarial perturbation. Xie et al. [63] find that random resizing and padding on the input image can mitigate the adversarial threat. Naseer et al. [39] train a neural representation purifier (NRP) by a self-supervised adversarial training mechanism to purify the input sample, which exhibits superior effectiveness against transfer-based adversarial examples. On the other hand, certified defense methods aim to provide provable defense in a given radius [16, 74, 8]. For instance, randomized smoothing (RS) trains a robust ImageNet classifier with a tight robustness guarantee [8]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 462, + 167, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 167, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 167, + 474 + ], + "type": "text", + "content": "2.3. Data Augmentations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 480, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 648 + ], + "type": "text", + "content": "Data augmentations often transform (e.g., flipping, rotation, cropping, etc.) the image during the training process for better generalization. Mixup [75] interpolates two images and their labels to generate virtual samples for training, which also inspires Admix to enhance transferability. Cutmix [73] pastes an image patch to the original patch and mixes the labels accordingly. AutoAugment [9] automatically searches for improved data augmentation policies (operations and parameters) on the dataset for better generalization, which has been widely adopted in deep learning. Unlike these data augmentation strategies, we aim to construct a set of diverse images by transforming the image block using various transformations, which can be used for gradient calculation to achieve better transferability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 129, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 129, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 129, + 671 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "In this section, we introduce our motivation, provide a detailed description of the proposed SIA, and highlight the difference between SIA and AutoAugment." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 542, + 119 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 542, + 119 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 542, + 119 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 542, + 119 + ], + "type": "table", + "html": "
TIMDIMSIMSSAAdmix
Transferability57.477.679.380.683.6
LPIPS0.250.430.480.540.73
", + "image_path": "4da5e54a36f24390a471c5ebc5025009613a13999e90785ebf00e3631ce6de52.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 127, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 304, + 127, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 545, + 188 + ], + "type": "text", + "content": "Table 1: The transferability of TIM, DIM, SIM, Admix, SSA, and similarity between 1,000 images and the transformed images evaluated by LPIPS. The transferability is evaluated by the attack success rate of Inception-v3 on the adversarial examples generated on ResNet-18." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 206, + 381, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 206, + 381, + 218 + ], + "spans": [ + { + "bbox": [ + 306, + 206, + 381, + 218 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 224, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 321 + ], + "type": "text", + "content": "Since Xie et al. [64] found that transforming the image by random resizing and padding before gradient calculation can generate more transferable adversarial examples, various input transformation based attacks are proposed to further improve transferability. As shown in Fig. 1, the input transformation (e.g., DIM [64] vs. Admix [55]) with better transferability tends to bring more obvious visual change to the input image. This inspires us with a new assumption:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 328, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 364 + ], + "type": "text", + "content": "Assumption 1 Without harming the semantic information, the more diverse the transformed image is, the better transferability the adversarial examples have." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 371, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 419 + ], + "type": "text", + "content": "To validate this hypothesis, we utilize the learned perceptual image patch similarity (LPIPS) [79] to evaluate the semantic similarity between the benign samples and the transformed images:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 425, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 425, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 318, + 425, + 545, + 453 + ], + "type": "interline_equation", + "content": "\\operatorname {L P I P S} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}}) = \\frac {1}{H \\times W} \\sum_ {l} \\sum_ {h, w} \\| \\boldsymbol {z} _ {h, w} ^ {l} - \\hat {\\boldsymbol {z}} _ {h, w} ^ {l} \\| _ {2} \\tag {1}", + "image_path": "27899dbe2423a96dec3d452d3e518cc3f530dfeb04b7920025d14bbf03b42c78.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "inline_equation", + "content": "z^l" + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\hat{z}^l" + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "text", + "content": " are the extracted feature from " + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "text", + "content": "-th layer of SqueezeNet [22] with " + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 462, + 545, + 641 + ], + "type": "text", + "content": " as input, respectively. A smaller LPIPS value indicates better similarity between the two images. The semantic similarity between the raw images and transformed images by DIM, TIM, SIM and Admix are summarized in Tab. 1. As we can see, the similarity decreases when the transferability of attack increases because the transformed images are more diverse. Moreover, as shown in Fig. 1, such diverse images are significantly different from the raw images, which introduces instability when calculating the gradient. Hence, the more powerful input transformation based attack needs to calculate the gradient on multiple transformed images to eliminate the instability. For instance, DIM utilizes single image, SIM adopts 5 images, while the SOTA Admix/SSA takes 15/20 images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "The relation between the diversity of transformed images and transferability inspires us to generate more diverse images for gradient calculation so that we can craft more transferable adversarial examples. In this work, we apply various input transformations on different blocks of a single input image to obtain more diverse images, detailed in Sec. 3.2." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4609" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 550, + 131 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 550, + 131 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 550, + 131 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 550, + 131 + ], + "type": "table", + "html": "
RawVShiftHShiftVFlipHFlipRotateScaleAdd NoiseResizeDCTDropout
", + "image_path": "6d9fdbc49ba45c9030bf689a8068adeb9fbc21f2530851fe76d962ce9072431d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 60, + 140, + 531, + 152 + ], + "lines": [ + { + "bbox": [ + 60, + 140, + 531, + 152 + ], + "spans": [ + { + "bbox": [ + 60, + 140, + 531, + 152 + ], + "type": "text", + "content": "Table 2: The raw and transformed images using various transformations adopted by SIA (See details in Appendix A)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 54, + 171, + 126, + 243 + ], + "blocks": [ + { + "bbox": [ + 54, + 171, + 126, + 243 + ], + "lines": [ + { + "bbox": [ + 54, + 171, + 126, + 243 + ], + "spans": [ + { + "bbox": [ + 54, + 171, + 126, + 243 + ], + "type": "image", + "image_path": "bb320fe443122e66f26d3abbe023f06e06ffd0d2a970384ad4953a260648d04a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 247, + 111, + 258 + ], + "lines": [ + { + "bbox": [ + 69, + 247, + 111, + 258 + ], + "spans": [ + { + "bbox": [ + 69, + 247, + 111, + 258 + ], + "type": "text", + "content": "Raw Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 132, + 171, + 203, + 243 + ], + "blocks": [ + { + "bbox": [ + 132, + 171, + 203, + 243 + ], + "lines": [ + { + "bbox": [ + 132, + 171, + 203, + 243 + ], + "spans": [ + { + "bbox": [ + 132, + 171, + 203, + 243 + ], + "type": "image", + "image_path": "8af069c8090e5d480041e98e6fc262359e226a77e2aebe686a78fc26f34af958.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 247, + 178, + 255 + ], + "lines": [ + { + "bbox": [ + 157, + 247, + 178, + 255 + ], + "spans": [ + { + "bbox": [ + 157, + 247, + 178, + 255 + ], + "type": "text", + "content": "Scale" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 209, + 171, + 280, + 243 + ], + "blocks": [ + { + "bbox": [ + 209, + 171, + 280, + 243 + ], + "lines": [ + { + "bbox": [ + 209, + 171, + 280, + 243 + ], + "spans": [ + { + "bbox": [ + 209, + 171, + 280, + 243 + ], + "type": "image", + "image_path": "1afb90fdd68ced872ce7e7d2f8b370557058966d66e97bdb215a22f0830e0488.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 247, + 266, + 256 + ], + "lines": [ + { + "bbox": [ + 222, + 247, + 266, + 256 + ], + "spans": [ + { + "bbox": [ + 222, + 247, + 266, + 256 + ], + "type": "text", + "content": "Block Scale" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 267, + 287, + 292 + ], + "lines": [ + { + "bbox": [ + 47, + 267, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 267, + 287, + 292 + ], + "type": "text", + "content": "Figure 2: The randomly sampled raw image and its transformed images by scaling on the full image and " + }, + { + "bbox": [ + 47, + 267, + 287, + 292 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 267, + 287, + 292 + ], + "type": "text", + "content": " blocks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 311, + 195, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 195, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 195, + 322 + ], + "type": "text", + "content": "3.2. Structure Invariant Attack" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 329, + 288, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 288, + 534 + ], + "type": "text", + "content": "Considering the limited number of the existing image transformations, we first investigate that how to generate a more diverse image using a single transformation? Without loss of generality, we take scale as an example, which is the base operation of SIM [32] (also a particular case of Admix [55]). As shown in Fig. 2, scaling the image changes its intensity uniformly on all the pixels, resulting in limited diverse images, which is also validated in Tab. 1. To improve the diversity of scaled images, we scale the image with different factors for different parts. Specifically, we randomly split the image into several image blocks and independently scale each image block with different scaled factors. As illustrated in Fig. 2, scaling the image blocks can bring a much more diverse transformed image while humans can still catch the visual information from the image. To explain why such image does not confuse humans, we define the structure of image as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "type": "text", + "content": "Definition 1 (Structure of Image). Given an image " + }, + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "type": "text", + "content": " which is randomly split into " + }, + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "s \\times s" + }, + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "type": "text", + "content": " blocks, the relative relation between each anchor point is the structure of image, where the anchor point is the center of the image block." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "content": "We argue that the structure of image depicts important semantic information for human recognition. For instance, the dog's body should be between its head and tail, while its legs should be under its body. Scaling the image blocks with various factors does not change the structure of image so that the generated image can be correctly recognized by humans as well as deep models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "In summary, transforming the image blocks does not harm the recognition but crafts more diverse images, which is of great benefit to improve transferability. To further" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 176, + 476, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 176, + 476, + 187 + ], + "spans": [ + { + "bbox": [ + 312, + 176, + 476, + 187 + ], + "type": "text", + "content": "Algorithm 1: Structure Invariant Attack" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "spans": [ + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "text", + "content": "Input: Classifier " + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "inline_equation", + "content": "\\pmb{f}(\\cdot)" + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "text", + "content": " with the loss function " + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "text", + "content": "; The benign sample " + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "text", + "content": " with ground-truth label " + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "text", + "content": "; The maximum perturbation " + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "text", + "content": ", number of iterations " + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "text", + "content": " and decay factor " + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "text", + "content": "; Splitting number " + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "text", + "content": "; Number of transformed images " + }, + { + "bbox": [ + 318, + 190, + 530, + 250 + ], + "type": "inline_equation", + "content": "N" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 251, + 455, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 251, + 455, + 262 + ], + "spans": [ + { + "bbox": [ + 319, + 251, + 455, + 262 + ], + "type": "text", + "content": "Output: An adversarial example." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 262, + 438, + 285 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 310, + 262, + 438, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 262, + 438, + 274 + ], + "spans": [ + { + "bbox": [ + 310, + 262, + 438, + 274 + ], + "type": "text", + "content": "1 " + }, + { + "bbox": [ + 310, + 262, + 438, + 274 + ], + "type": "inline_equation", + "content": "\\alpha = \\epsilon /T" + }, + { + "bbox": [ + 310, + 262, + 438, + 274 + ], + "type": "inline_equation", + "content": "\\pmb {g}_0 = 0" + }, + { + "bbox": [ + 310, + 262, + 438, + 274 + ], + "type": "inline_equation", + "content": "x_0^{adv} = x" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 275, + 411, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 275, + 411, + 285 + ], + "spans": [ + { + "bbox": [ + 310, + 275, + 411, + 285 + ], + "type": "text", + "content": "2 for " + }, + { + "bbox": [ + 310, + 275, + 411, + 285 + ], + "type": "inline_equation", + "content": "t = 0\\to T - 1" + }, + { + "bbox": [ + 310, + 275, + 411, + 285 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 310, + 287, + 523, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 287, + 523, + 310 + ], + "spans": [ + { + "bbox": [ + 310, + 287, + 523, + 310 + ], + "type": "text", + "content": "Constructing a set " + }, + { + "bbox": [ + 310, + 287, + 523, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 310, + 287, + 523, + 310 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 310, + 287, + 523, + 310 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 310, + 287, + 523, + 310 + ], + "type": "text", + "content": " transformed images using SIT" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 311, + 492, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 311, + 492, + 322 + ], + "spans": [ + { + "bbox": [ + 311, + 311, + 492, + 322 + ], + "type": "text", + "content": "4 Calculating the average gradient on " + }, + { + "bbox": [ + 311, + 311, + 492, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 374, + 331, + 531, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 331, + 531, + 361 + ], + "spans": [ + { + "bbox": [ + 374, + 331, + 531, + 361 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {g}} _ {t + 1} = \\frac {1}{N} \\sum_ {\\boldsymbol {x} _ {i} \\in \\mathcal {X}} \\nabla_ {\\boldsymbol {x}} J (\\boldsymbol {x} _ {i}, y) \\tag {2}", + "image_path": "7dda62cdb19be5f8ed87ed605f47eaca6cd615e2e5648045d8500e0334108ffb.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 310, + 372, + 438, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 372, + 438, + 383 + ], + "spans": [ + { + "bbox": [ + 310, + 372, + 438, + 383 + ], + "type": "text", + "content": "5 | Updating the momentum:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 384, + 392, + 530, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 392, + 530, + 418 + ], + "spans": [ + { + "bbox": [ + 384, + 392, + 530, + 418 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {t + 1} = \\mu \\boldsymbol {g} _ {t} + \\frac {\\bar {\\boldsymbol {g}} _ {t + 1}}{\\| \\bar {\\boldsymbol {g}} _ {t + 1} \\| _ {1}} \\tag {3}", + "image_path": "76ca17629410c225d4f27c2d8170cc4026edf6726f8dc0b4cb5d6202d0b6c90c.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 310, + 427, + 473, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 427, + 473, + 439 + ], + "spans": [ + { + "bbox": [ + 310, + 427, + 473, + 439 + ], + "type": "text", + "content": "6 | Updating the adversarial example:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 343, + 449, + 530, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 449, + 530, + 464 + ], + "spans": [ + { + "bbox": [ + 343, + 449, + 530, + 464 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t + 1} ^ {a d v} = \\operatorname {C l i p} \\left(\\boldsymbol {x} _ {t} ^ {a d v} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\boldsymbol {g} _ {t + 1}\\right), 0, 1\\right) \\tag {4}", + "image_path": "e6addbb938c47089168f4b43d3a1895bca8ece39bb483b5f780ce581cbe9bb57.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 310, + 475, + 370, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 475, + 370, + 489 + ], + "spans": [ + { + "bbox": [ + 310, + 475, + 370, + 489 + ], + "type": "text", + "content": "7 return " + }, + { + "bbox": [ + 310, + 475, + 370, + 489 + ], + "type": "inline_equation", + "content": "x_{T}^{adv}" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 519, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 603 + ], + "type": "text", + "content": "boost the diversity of transformed images, we apply various image transformations to different image blocks, where the adopted transformations are summarized in Tab. 2. To avoid information loss, we add the constraints on some transformations. For instance, the rotation can only rotate the image " + }, + { + "bbox": [ + 304, + 519, + 545, + 603 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 304, + 519, + 545, + 603 + ], + "type": "text", + "content": " to avoid dropping some pixels. We denote such transformation as structure invariant transformation (SIT)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "Since SIT can generate more diverse images while maintaining the critical semantic information, we adopt SIT as the input transformation to conduct the adversarial attack, denoted as structure invariant attack (SIA). Instead of directly calculating the gradient on the input image, SIA calculates the gradient on several images transformed by SIT. Note that SIA is generally applicable to any gradient-based attacks. Due to the limited space, we integrate SIA into MIFGSM [10] and summarize the algorithm in Algorithm 1." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4610" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 182, + 74, + 410, + 86 + ], + "blocks": [ + { + "bbox": [ + 182, + 74, + 410, + 86 + ], + "lines": [ + { + "bbox": [ + 182, + 74, + 410, + 86 + ], + "spans": [ + { + "bbox": [ + 182, + 74, + 410, + 86 + ], + "type": "image", + "image_path": "20fe0ecb2ad38d5b5fe1ca29c6ec19511b055c4796e46c3c123469f92f574f12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 90, + 171, + 169 + ], + "blocks": [ + { + "bbox": [ + 52, + 90, + 171, + 169 + ], + "lines": [ + { + "bbox": [ + 52, + 90, + 171, + 169 + ], + "spans": [ + { + "bbox": [ + 52, + 90, + 171, + 169 + ], + "type": "image", + "image_path": "fc8162f0fa6e6f4a85aa9463d728b401433b12fb6cace461ad6b3940560cce77.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 176, + 90, + 294, + 169 + ], + "blocks": [ + { + "bbox": [ + 176, + 90, + 294, + 169 + ], + "lines": [ + { + "bbox": [ + 176, + 90, + 294, + 169 + ], + "spans": [ + { + "bbox": [ + 176, + 90, + 294, + 169 + ], + "type": "image", + "image_path": "5cbbdeb4e015eb7e17752df75e066f79f97b196c2cdd173dd34c0bee4ab99cae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 90, + 419, + 170 + ], + "blocks": [ + { + "bbox": [ + 299, + 90, + 419, + 170 + ], + "lines": [ + { + "bbox": [ + 299, + 90, + 419, + 170 + ], + "spans": [ + { + "bbox": [ + 299, + 90, + 419, + 170 + ], + "type": "image", + "image_path": "59f646129b033fb3d6dbe7bd5ab2ac10b5b314c44699b213dd4fd3ee0eedb616.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 424, + 91, + 542, + 169 + ], + "blocks": [ + { + "bbox": [ + 424, + 91, + 542, + 169 + ], + "lines": [ + { + "bbox": [ + 424, + 91, + 542, + 169 + ], + "spans": [ + { + "bbox": [ + 424, + 91, + 542, + 169 + ], + "type": "image", + "image_path": "dcc9afcb05052c71dc4744e7f72be33a388d4dddf8cef728aa8cce6bfba38f1d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 175, + 171, + 253 + ], + "blocks": [ + { + "bbox": [ + 52, + 175, + 171, + 253 + ], + "lines": [ + { + "bbox": [ + 52, + 175, + 171, + 253 + ], + "spans": [ + { + "bbox": [ + 52, + 175, + 171, + 253 + ], + "type": "image", + "image_path": "d67246c6c6eab4dd9d1b519b6cf39af249e52d27426117538de9e66b0e8bd6e6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 265, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 545, + 289 + ], + "type": "text", + "content": "Figure 3: Attack success rates " + }, + { + "bbox": [ + 46, + 265, + 545, + 289 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 265, + 545, + 289 + ], + "type": "text", + "content": " of eight deep models on the adversarial examples crafted on each model by TIM, DIM, DEM, Admix, SSA, and SIA." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 176, + 175, + 294, + 253 + ], + "blocks": [ + { + "bbox": [ + 176, + 175, + 294, + 253 + ], + "lines": [ + { + "bbox": [ + 176, + 175, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 176, + 175, + 294, + 253 + ], + "type": "image", + "image_path": "f54a3fe1438516f6c9521f3a412c32892ae56a0e459311a138fca1f5a321df98.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 300, + 175, + 419, + 254 + ], + "blocks": [ + { + "bbox": [ + 300, + 175, + 419, + 254 + ], + "lines": [ + { + "bbox": [ + 300, + 175, + 419, + 254 + ], + "spans": [ + { + "bbox": [ + 300, + 175, + 419, + 254 + ], + "type": "image", + "image_path": "122fafbe6e052d31c12ec2f3c98b9b7154f783fa392057d75f3c94d88b15b420.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 424, + 175, + 542, + 253 + ], + "blocks": [ + { + "bbox": [ + 424, + 175, + 542, + 253 + ], + "lines": [ + { + "bbox": [ + 424, + 175, + 542, + 253 + ], + "spans": [ + { + "bbox": [ + 424, + 175, + 542, + 253 + ], + "type": "image", + "image_path": "1614697564ff397200f7b4c57ae65a540dfaf361dc40d2567d5a34ba6e9a5ba0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 311, + 169, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 169, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 169, + 323 + ], + "type": "text", + "content": "3.3. SIT vs. AutoAugment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 329, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 364 + ], + "type": "text", + "content": "Both SIT and AutoAugment [9] leverage several image transformations to pre-process the images. We summarize the difference as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 367, + 287, + 511 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 287, + 403 + ], + "type": "text", + "content": "- SIT improves the adversarial transferability when attacking the model while AutoAugment boosts the model generalization during the training process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 403, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 438 + ], + "type": "text", + "content": "- SIT randomly samples the transformation while AutoAugment needs to search for a good policy for each transformation on each dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 439, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 287, + 487 + ], + "type": "text", + "content": "- SIT applies various transformations on different image blocks locally but preserves the global structure. On the contrary, AutoAugment applies two transformations sequentially on the image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 487, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 287, + 511 + ], + "type": "text", + "content": "- The fine-grained transformations by SIT generate more diverse images than AutoAugment." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 522, + 123, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 123, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 123, + 536 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 543, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 567 + ], + "type": "text", + "content": "In this section, we conduct extensive evaluations on ImageNet dataset to validate the effectiveness of SIA." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 574, + 170, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 170, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 170, + 588 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 641 + ], + "type": "text", + "content": "Dataset. To align with previous works [10, 32, 53, 55], we randomly sample 1,000 images pertaining to 1,000 categories from the ILSVRC 2012 validation set [40], which are correctly classified by the adopted models." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Baselines. We compare our proposed SIA with five competitive input transformation based attacks, namely DIM [64], TIM [11], DEM [81], Admix [55], and SSA [36], which are integrated into MI-FGSM [10]. We also integrate the baseline methods with two model-specific approaches, namely LinBP [19] and SGM [60]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 311, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 491 + ], + "type": "text", + "content": "Victim Models. We evaluate the attack performance on two popular model architectures, namely Convolutional Network Works, i.e., ResNet-18 [20], ResNet-101 [20], ResNext-50 [65], DenseNet-121 [23], MobileNet [21], and Transformers, i.e., Vision Transformer (ViT) [12] and Swin Transformer (Swin) [34]. Furthermore, we study several SOTA defense methods, including one adversarial training method, i.e., ensemble adversarially trained model (Incv3_ens) [48], the top-3 submissions in NIPS 2017 defense competition, i.e., high-level representation guided denoiser (HGD) [30], random resizing and padding (R&P) [63] and NIPS-r3" + }, + { + "bbox": [ + 304, + 311, + 545, + 491 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 304, + 311, + 545, + 491 + ], + "type": "text", + "content": ", three input pre-processing based defenses, namely FD [35], JPEG [18] and Bit-Red [67], a certified defense, i.e., randomized smoothing (RS) [8] and a deep denoiser, i.e., neural representation purifier (NRP) [39]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "text", + "content": "Evaluation Settings. We follow MI-FGSM [10] with the perturbation budget " + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\epsilon = 16" + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "text", + "content": ", number of iteration " + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "inline_equation", + "content": "T = 10" + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "text", + "content": ", step size " + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\alpha = \\epsilon / T = 1.6" + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "text", + "content": " and decay factor " + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\mu = 1" + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "text", + "content": ". DIM [64] adopts the transformation probability of 0.5 and TIM [11] utilizes the Gaussian kernel with the size of " + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "text", + "content": ". DEM [81] uses the resize ratios: [1.14, 1.27, 1.4, 1.53, 1.66]. Admix admixes 3 images from other categories with the strength of 0.2 and 5 scaled images for each admixed image. SSA [36] sets the turning factor as 0.5 and the standard deviation as " + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "text", + "content": ". SIA sets the splitting number " + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "inline_equation", + "content": "s = 3" + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "text", + "content": " and number of transformed images for gradient calculation " + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "inline_equation", + "content": "N = 20" + }, + { + "bbox": [ + 304, + 491, + 545, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 630, + 448, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 630, + 448, + 643 + ], + "spans": [ + { + "bbox": [ + 306, + 630, + 448, + 643 + ], + "type": "text", + "content": "4.2. Attacking a Single Model" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 649, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 649, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 649, + 545, + 685 + ], + "type": "text", + "content": "To validate the effectiveness of the proposed SIA, we first compare SIA with five SOTA input transformation based attacks, namely DIM, TIM, DEM, Admix and SSA." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 525, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 525, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 525, + 712 + ], + "type": "text", + "content": "1https://github.com/anlthms/nips-2017/tree/master/mmd" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4611" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 282, + 252 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 282, + 252 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 282, + 252 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 282, + 252 + ], + "type": "image", + "image_path": "5a96ed0444eb2a1cd0d72c197d5df10483bbf76d3b5ae1e31882ce320d0f3513.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 265, + 287, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 313 + ], + "type": "text", + "content": "Figure 4: Attack success rates " + }, + { + "bbox": [ + 46, + 265, + 287, + 313 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 265, + 287, + 313 + ], + "type": "text", + "content": " of eight models on the adversarial examples generated on ResNet-18 when integrating TIM, DIM, DEM, Admix, SSA, and SIA into SGM, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "type": "text", + "content": "We generate the adversarial examples on a single model and test them on the other models. The attack success rates, i.e., the misclassification rates of the victim model on the crafted adversarial examples, are summarized in Fig. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "type": "text", + "content": "From the figure, we can observe that all the attackers can achieve the attack success rate of " + }, + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "type": "inline_equation", + "content": "100.0\\%" + }, + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "type": "text", + "content": " or near " + }, + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "type": "inline_equation", + "content": "100.0\\%" + }, + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "type": "text", + "content": ", showing that the input transformation based attacks do not degrade the white-box attack performance. As for the black-box performance, TIM exhibits the poorest transferability on these normally trained models. Admix consistently achieves better transferability than DIM and DEM on CNN-based models. Surprisingly, DIM achieves even better transferability than Admix when generating adversarial examples on transformer-based models. SSA exhibits the best transferability among the baselines in most cases. Also, even for the transformer-based models, the adversarial examples generated on Swin show the poorest transferability on ViT than other CNN-based models. Hence, we argue that it is necessary to evaluate the effectiveness of transfer-based attacks on both CNN-based and transformer-based models. We need an in-depth analysis of the transferability among the emerging transformer-based models. Compared with the baselines, SIA consistently performs much better than the best baselines on all eight models with different architectures. In particular, SIA outperforms the winner-up method with a clear margin of " + }, + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "type": "inline_equation", + "content": "14.3\\%" + }, + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "type": "text", + "content": " on average and achieves an attack success rate higher than the best baseline of at least " + }, + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 46, + 391, + 288, + 715 + ], + "type": "text", + "content": " on all the models. Such consistent and superior performance demonstrates that the proposed SIA is general to various model architectures (either CNN or transformer) to boost transferability effectively." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 541, + 254 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 541, + 254 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 541, + 254 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 541, + 254 + ], + "type": "image", + "image_path": "cc97c92d2c10f857ec943002cdca9f8a15070a36900b2413f607548f164c8cde.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 265, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 313 + ], + "type": "text", + "content": "Figure 5: Attack success rates " + }, + { + "bbox": [ + 304, + 265, + 545, + 313 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 304, + 265, + 545, + 313 + ], + "type": "text", + "content": " of eight models on the adversarial examples generated on ResNet-18 when integrating TIM, DIM, DEM, Admix, SSA, and SIA into LinBP, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 334, + 518, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 518, + 347 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 518, + 347 + ], + "type": "text", + "content": "4.3. Integration to model-specific approaches" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 352, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 411 + ], + "type": "text", + "content": "To further verify the scalability of the proposed SIA, we integrate existing input transformation-based attacks into two model-specific approaches, namely SGM and LinBP. The adversarial examples are generated on ResNet-18 and test them on the other models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "content": "As depicted in Fig. 4 and Fig. 5, SGM and LinBP can significantly boost the adversarial transferability of vallina input transformation-based attacks. Compared with the baselines, SIA still achieves much better transferability, which has been boosted by an average margin of " + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 304, + 412, + 545, + 533 + ], + "type": "text", + "content": " when integrated into SGM and LinBP, respectively. These superior results validate the generality of SIA to various transfer-based attacks and underscore the potential of SIA in augmenting adversarial transferability through the fusion of different strategies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 539, + 460, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 460, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 460, + 552 + ], + "type": "text", + "content": "4.4. Attacking Ensemble Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "content": "Liu et al. [33] have shown that attacking ensemble models can effectively improve transferability. As shown in Sec. 4.2, the adversarial examples often exhibit poorer transferability across the CNN-based models and transformer-based models than being transferred among the CNN-based models. Hence, we generate adversarial examples on ensemble CNN-based and ensemble transformer-based models and test them on the remaining models to evaluate SIA when attacking ensemble models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "As shown in Fig. 6, when attacking ensemble models, the adversarial examples generated by all attacks exhibit better transferability than that crafted on a single model, showing the excellent compatibility of all the input trans" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4612" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 282, + 252 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 282, + 252 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 282, + 252 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 282, + 252 + ], + "type": "image", + "image_path": "37dd8323c287535f2d89323967e6331264bca7bdfce575d7a8c3ac27a928c6a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 265, + 287, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 326 + ], + "type": "text", + "content": "Figure 6: Attack success rates (\\%) of eight models on the adversarial examples crafted on ensemble models by TIM, DIM, DEM, Admix, SSA, and SIA, respectively. We generate adversarial examples on the CNN-based models and test them on the transformer-based models, and vice versa." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "type": "text", + "content": "formation based attacks with such a setting. When generating the adversarial examples on ViT and Swin, DIM and DEM achieve better transferability than Admix and SSA, highlighting the difference between crafting adversarial examples on different architectures and the necessity to evaluate the transferability on these models. On all eight models, SIA achieves the attack success rate of at least " + }, + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "type": "inline_equation", + "content": "88.3\\%" + }, + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "type": "text", + "content": ". It outperforms the winner-up approach with a clear margin of " + }, + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "type": "text", + "content": ", showing its superior effectiveness in generating transferable adversarial examples. In particular, SIA achieves " + }, + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "type": "inline_equation", + "content": "94.9\\%" + }, + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "type": "inline_equation", + "content": "99.3\\%" + }, + { + "bbox": [ + 46, + 348, + 287, + 552 + ], + "type": "text", + "content": " attack success rates on ViT and Swin, respectively, when the adversarial examples are generated on CNN-based models without access to the attention module in the transformer. This further supports our motivation that improving the diversity of transformed images can significantly boost transferability, even on models with completely different architectures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 562, + 200, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 200, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 200, + 574 + ], + "type": "text", + "content": "4.5. Attacking Defense Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 580, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 286, + 688 + ], + "type": "text", + "content": "SIA has achieved superior attack performance on eight normally trained models with different architectures when attacking single model as well as ensemble models. Recently, several defenses have been proposed to mitigate the threat of adversarial examples on ImageNet dataset. To validate the effectiveness of these defenses, we adopt the adversarial examples generated on these eight models simultaneously to attack the defense models, including Inc-v3ens, HGD, R&P, NIPS-r3, FD, JPEG, Bit-Red, RS and NRP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 714 + ], + "type": "text", + "content": "The attack results on these defense methods are summarized in Fig. 7. Overall, DIM is on par with TIM on" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 541, + 252 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 541, + 252 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 541, + 252 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 541, + 252 + ], + "type": "image", + "image_path": "c00f382c99a1449f97cc2353af5ef232d0f76e7c9a0e397eee1ab0e5e7b8e1fc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 265, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 305, + 265, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 545, + 326 + ], + "type": "text", + "content": "Figure 7: Attack success rates " + }, + { + "bbox": [ + 305, + 265, + 545, + 326 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 265, + 545, + 326 + ], + "type": "text", + "content": " of various defense methods on the adversarial examples generated by TIM, DIM, DEM, Admix, SSA, and SIA under ensemble model setting. The adversarial examples are generated on the eight models simultaneously." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 348, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 468 + ], + "type": "text", + "content": "these defense methods while Admix can consistently outperform the other four baselines. SIA can consistently achieve better transferability than the baselines. In particular, SIA achieves the attack success rate of " + }, + { + "bbox": [ + 304, + 348, + 545, + 468 + ], + "type": "inline_equation", + "content": "78.2\\%" + }, + { + "bbox": [ + 304, + 348, + 545, + 468 + ], + "type": "text", + "content": " on the certified defense method (i.e., RS), and SIA can achieve the attack success rate of at least " + }, + { + "bbox": [ + 304, + 348, + 545, + 468 + ], + "type": "inline_equation", + "content": "89.9\\%" + }, + { + "bbox": [ + 304, + 348, + 545, + 468 + ], + "type": "text", + "content": " on the other eight defense methods, including the powerful denoising method NRP. Such high attack performance indicates the insufficiency of existing defense methods and raises a new security issue to designing more robust deep learning models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 478, + 406, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 406, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 406, + 490 + ], + "type": "text", + "content": "4.6. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 497, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 569 + ], + "type": "text", + "content": "To further gain insights on the superior attack performance achieved by SIA, we conduct a series of ablation studies to validate that all the transformations and splitting the image into blocks can help improve transferability. We generate the adversarial examples on ResNet-18 and test them on the other seven models for all the experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Are all the transformations beneficial for boosting the transerability? There are ten different input transformations that SIA might apply to each image block. To investigate whether each transformation is of benefit to generate more transferable adversarial examples, we first adopt one or two transformations to implement SIA. The results are summarized in Fig. 8. The average attack success rate of MI-FGSM is " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "59.7\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". On the diagonal line of Fig. 8, we only adopt a single transformation for each block and the lowest average attack success rate is " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "67.3\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " by adopting DCT, which outperforms MI-FGSM with a margin of " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". This demonstrates the high effectiveness of SIA and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4613" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 525, + 107 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 525, + 107 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 525, + 107 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 525, + 107 + ], + "type": "table", + "html": "
SIA-VShift-HShift-VFlip-HFlip-Rotate-Sclae-Add Noise-Resize-DCT-Dropout
92.189.790.190.188.488.390.190.690.290.190.7
", + "image_path": "8a37f2683e2b6b054e435ad8bf3a6a7fc8c903ac120fd644e56c9277b62fafa6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 115, + 547, + 140 + ], + "lines": [ + { + "bbox": [ + 46, + 115, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 46, + 115, + 547, + 140 + ], + "type": "text", + "content": "Table 3: The average attack success rates of adversarial examples crafted by SIA and SIA without a single transformation. The adversaries are generated on ResNet-18 and tested on the other seven deep models. - indicates removing such transformation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 52, + 162, + 286, + 339 + ], + "blocks": [ + { + "bbox": [ + 52, + 162, + 286, + 339 + ], + "lines": [ + { + "bbox": [ + 52, + 162, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 52, + 162, + 286, + 339 + ], + "type": "image", + "image_path": "19e5b9ac816739d615e5847af4b383f2ee4ef7eb81f0ab022e00b030afb6650e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "text", + "content": "Figure 8: The average attack success rates " + }, + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 352, + 287, + 389 + ], + "type": "text", + "content": " on the adversarial examples generated by SIA using one (diagonal) or two (other) transformations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 412, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 568 + ], + "type": "text", + "content": "its generality to various transformations. We can also observe that the transferability can be further improved when we combine any two transformations, showing the advantage of combining these transformations. To further validate the necessity of each transformation, we generate adversarial examples by SIA without each transformation and summarize the results in Tab. 3. Since removing a single transformation does not significantly decrease the diversity of the transformation, these attacks achieve similar attack performance. However, no matter which transformation is removed, the transferability will be decreased, supporting that each transformation is of benefit to generate more transferable adversarial examples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": "Are the image blocks beneficial for boosting transferability? We have shown that the abundant transformations can effectively improve transferability. Here we further explore whether it is useful to apply these transformations to the image block instead of the raw image. For comparison, we randomly apply the transformation to the raw image " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "s \\times s" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": " times, denoted as " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SIA_r}" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": ". The results are reported in Fig. 9. We can observe that " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SIA_r}" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": " exhibits better transferability than MI-FGSM, which also validates our motivation that improving the diversity of transformed images can generate more transferable adversarial examples. Our SIA consistently performs better than " + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SIA_r}" + }, + { + "bbox": [ + 46, + 570, + 288, + 715 + ], + "type": "text", + "content": ", indicating that applying" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 162, + 542, + 339 + ], + "blocks": [ + { + "bbox": [ + 310, + 162, + 542, + 339 + ], + "lines": [ + { + "bbox": [ + 310, + 162, + 542, + 339 + ], + "spans": [ + { + "bbox": [ + 310, + 162, + 542, + 339 + ], + "type": "image", + "image_path": "aa9c8322747df196024fc159650aaad0713635eaa7a62acf81b178ded62941fc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 352, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 305, + 352, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 545, + 389 + ], + "type": "text", + "content": "Figure 9: The average attack success rates " + }, + { + "bbox": [ + 305, + 352, + 545, + 389 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 352, + 545, + 389 + ], + "type": "text", + "content": " of adversarial examples generated by SIA and " + }, + { + "bbox": [ + 305, + 352, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\mathrm{SIA_r}" + }, + { + "bbox": [ + 305, + 352, + 545, + 389 + ], + "type": "text", + "content": " (randomly transform the image without the block partition)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 410, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 545, + 436 + ], + "type": "text", + "content": "the transformation onto each image block is of great benefit to improve transferability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 443, + 416, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 416, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 416, + 455 + ], + "type": "text", + "content": "4.7. Parameter Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 461, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 533 + ], + "type": "text", + "content": "In this subsection, we conduct parameter studies to explore the impact of two hyper-parameters, namely the number of blocks " + }, + { + "bbox": [ + 304, + 461, + 545, + 533 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 461, + 545, + 533 + ], + "type": "text", + "content": " and the number of transformed images for gradient calculation " + }, + { + "bbox": [ + 304, + 461, + 545, + 533 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 461, + 545, + 533 + ], + "type": "text", + "content": ". All the adversarial examples are generated on ResNet-18 and tested on the other seven models in the black-box setting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "content": "On the number of blocks " + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "content": ". The number of blocks determines how diverse the transformed images are, which can influence the attack performance. To find a good value for " + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "content": ", we conduct SIA with " + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "content": " from 1 to 5 and summarize the attack results in Fig. 10. When " + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "inline_equation", + "content": "s \\leq 3" + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "content": ", increasing the value of " + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "content": " leads to better diversity, which can improve the attack performance. However, when we continually increase the value of " + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "content": ", the diversity is increased, but it also introduces more variance to the gradient, decaying the attack performance slightly. Hence, we adopt " + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "inline_equation", + "content": "s = 3" + }, + { + "bbox": [ + 304, + 533, + 546, + 676 + ], + "type": "text", + "content": " to balance the diversity of images and variance of the gradient for better attack performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "On the number of images for gradient calculation " + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": ". SIA calculates the average gradient on " + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": " images to eliminate the variance introduced by the transformation. To de" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4614" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 75, + 286, + 248 + ], + "blocks": [ + { + "bbox": [ + 53, + 75, + 286, + 248 + ], + "lines": [ + { + "bbox": [ + 53, + 75, + 286, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 286, + 248 + ], + "type": "image", + "image_path": "5f0f3974934d2ffa05e95e725d521c3807bf3446061ac766eeef25ab5cb64921.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 287, + 286 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 286 + ], + "type": "text", + "content": "Figure 10: Attack success rates " + }, + { + "bbox": [ + 46, + 262, + 287, + 286 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 262, + 287, + 286 + ], + "type": "text", + "content": " of adversarial examples generated by SIA with various number of blocks " + }, + { + "bbox": [ + 46, + 262, + 287, + 286 + ], + "type": "inline_equation", + "content": "s" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "text", + "content": "termine a good value for " + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "text", + "content": ", we evaluate SIA with " + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "text", + "content": " from 1 to 30 and report the attack success rates in Fig. 11. As we can observe, when " + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "text", + "content": ", SIA introduces massive variance into the gradient and achieves the lowest performance on all models. When we increase the value of " + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "text", + "content": ", the variance among the gradients can be eliminated, and SIA achieves better transferability before " + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "inline_equation", + "content": "N = 20" + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "inline_equation", + "content": "N > 20" + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "text", + "content": ", increasing " + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "text", + "content": " can only introduce computation cost without performance improvement. To balance the attack performance and computation cost, we adopt " + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "inline_equation", + "content": "N = 20" + }, + { + "bbox": [ + 46, + 320, + 289, + 453 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 483, + 119, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 119, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 119, + 496 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": "In this work, we find that the existing input transformation based attack with better transferability often generates more diverse transformed images. Based on this finding, we design a new image transformation, called structure invariant transformation (SIT), which splits the image into several blocks and randomly transforms each image block independently. With such image transformation, we propose a novel input transformation based attack, dubbed structure invariant attack (SIA), which calculates the average gradient on several transformed images by SIT to update the perturbation. Extensive evaluations demonstrate that SIA can achieve remarkably better transferability than the existing SOTA attacks. In our opinion, SIA provides a new direction by applying the transformation onto the image block to effectively boost transferability, which sheds new light on generating more transferable adversarial examples with fine-grained transformations." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 312, + 75, + 541, + 248 + ], + "blocks": [ + { + "bbox": [ + 312, + 75, + 541, + 248 + ], + "lines": [ + { + "bbox": [ + 312, + 75, + 541, + 248 + ], + "spans": [ + { + "bbox": [ + 312, + 75, + 541, + 248 + ], + "type": "image", + "image_path": "750e2b03d21a661c6217aa6a5326ca38cfd9f1cf0709e405b1b53368430649ef.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "text", + "content": "Figure 11: Attack success rates " + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "text", + "content": " of adversarial examples generated by SIA with various number of images " + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "inline_equation", + "content": "N" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 308, + 365, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 365, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 365, + 320 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 311, + 329, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 311, + 329, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 329, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 311, + 329, + 545, + 384 + ], + "type": "text", + "content": "[1] Moustafa Alzantot, Yash Sharma, Ahmed Elgohary, BoJhang Ho, Mani B. Srivastava, and Kai-Wei Chang. Generating Natural Language Adversarial Examples. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 2890-2896, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 386, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 386, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 311, + 386, + 545, + 441 + ], + "type": "text", + "content": "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square Attack: A Query-Efficient Black-Box Adversarial Attack via Random Search. In Proceedings of the European Conference on Computer Vision, pages 484–501, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 442, + 545, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 442, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 311, + 442, + 545, + 496 + ], + "type": "text", + "content": "[3] Anish Athalye, Nicholas Carlini, and David A. Wagner. Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. In Proceedings of the International Conference on Machine Learning, pages 274–283, 2018. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 498, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 498, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 311, + 498, + 545, + 553 + ], + "type": "text", + "content": "[4] Wieland Brendel, Jonas Rauber, and Matthias Bethge. Decision-Based Adversarial Attacks: Reliable Attacks Against Black-Box Machine Learning Models. In Proceedings of the International Conference on Learning Representations, 2018. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 555, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 555, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 311, + 555, + 545, + 600 + ], + "type": "text", + "content": "[5] Nicholas Carlini and David A. Wagner. Audio Adversarial Examples: Targeted Attacks on Speech-to-Text. In Proceedings of the IEEE Symposium on Security and Privacy (Workshops), pages 1-7, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 311, + 601, + 545, + 656 + ], + "type": "text", + "content": "[6] Minhao Cheng, Thong Le, Pin-Yu Chen, Huan Zhang, Jinfeng Yi, and Cho-Jui Hsieh. Query-Efficient Hard-label Black-box Attack: An Optimization-based Approach. In Proceedings of the International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 311, + 658, + 545, + 712 + ], + "type": "text", + "content": "[7] Shuyu Cheng, Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Improving Black-box Adversarial Attacks with a Transfer-based Prior. In Proceedings of the Advances in Neural Information Processing Systems, pages 10932-10942, 2019. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4615" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "type": "text", + "content": "[8] Jeremy M. Cohen, Elan Rosenfeld, and J. Zico Kolter. Certified Adversarial Robustness via Randomized Smoothing. In Proceedings of the International Conference on Machine Learning, pages 1310-1320, 2019. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 173 + ], + "type": "text", + "content": "[9] Ekin D. Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, and Quoc V. Le. AutoAugment: Learning Augmentation Strategies From Data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 113-123, 2019. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 229 + ], + "type": "text", + "content": "[10] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting Adversarial Attacks With Momentum. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. 1, 2, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 287, + 285 + ], + "type": "text", + "content": "[11] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading Defenses to Transferable Adversarial Examples by Translation-Invariant Attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 1, 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 287, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 287, + 363 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In Proceedings of the International Conference on Learning Representations, 2021. 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 365, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 287, + 431 + ], + "type": "text", + "content": "[13] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi Kohno, and Dawn Song. Robust Physical-World Attacks on Deep Learning Visual Classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1625–1634, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 433, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 287, + 486 + ], + "type": "text", + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we Ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 488, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 488, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 49, + 488, + 287, + 531 + ], + "type": "text", + "content": "[15] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and Harnessing Adversarial Examples. In Proceedings of the International Conference on Learning Representations, 2015. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 534, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 534, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 287, + 599 + ], + "type": "text", + "content": "[16] Sven Gowal, Krishnamurthy Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timothy Arthur Mann, and Pushmeet Kohli. Scalable Verified Training for Provably Robust Image Classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4841-4850, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "type": "text", + "content": "[17] Chuan Guo, Jacob R. Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Q. Weinberger. Simple Black-box Adversarial Attacks. In Proceedings of the International Conference on Machine Learning, pages 2484-2493, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[18] Chuan Guo, Mayank Rana, Moustapha Cissé, and Laurens van der Maaten. Countering Adversarial Images Using Input Transformations. In Proceedings of the International Conference on Learning Representations, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 713 + ], + "type": "text", + "content": "[19] Yiwen Guo, Qizhang Li, and Hao Chen. Backpropagating linearly improves transferability of adversarial examples." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "Advances in neural information processing systems, 33:85-95, 2020. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "text", + "content": "[21] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv preprint arXiv:1704.04861, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "text", + "content": "[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-Eccitation Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[23] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely Connected Convolutional Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2261-2269, 2017. 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 295, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 346 + ], + "type": "text", + "content": "[24] Andrew Ilyas, Logan Engstrom, Anish Athalye, and Jessy Lin. Black-box Adversarial Attacks with Limited Queries and Information. In Proceedings of the International Conference on Machine Learning, pages 2142-2151, 2018. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 350, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 402 + ], + "type": "text", + "content": "[25] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. In Proceedings of the Advances in Neural Information Processing Systems, pages 1106-1114, 2012. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 405, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 448 + ], + "type": "text", + "content": "[26] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial Examples in the Physical World. In Proceedings of the International Conference on Learning Representations (Workshops), 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[27] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. QEBA: Query-Efficient Boundary-Based Blackbox Attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1218-1227, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "text", + "content": "[28] Yingwei Li, Song Bai, Yuyin Zhou, Cihang Xie, Zhishuai Zhang, and Alan L. Yuille. Learning Transferable Adversarial Examples via Ghost Networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11458-11465, 2020. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 559, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 623 + ], + "type": "text", + "content": "[29] Yandong Li, Lijun Li, Liqiang Wang, Tong Zhang, and Boqing Gong. NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks. In Proceedings of the International Conference on Machine Learning, pages 3866–3876, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 625, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 689 + ], + "type": "text", + "content": "[30] Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense Against Adversarial Attacks Using High-Level Representation Guided Denoiser. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2018. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "text", + "content": "[31] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4616" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Daan Wierstra. Continuous Control with Deep Reinforcement Learning. In Proceedings of the International Conference on Learning Representations, 2016. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[32] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E. Hopcroft. Nesterov Accelerated Gradient and Scale Invariance for Adversarial Attacks. In Proceedings of the International Conference on Learning Representations, 2020. 1, 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "text", + "content": "[33] Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into Transferable Adversarial Examples and Blackbox Attacks. In Proceedings of the International Conference on Learning Representations, 2017. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 263 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9992-10002, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 265, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 319 + ], + "type": "text", + "content": "[35] Zihao Liu, Qi Liu, Tao Liu, Nuo Xu, Xue Lin, Yanzhi Wang, and Wujie Wen. Feature Distillation: DNN-Oriented JPEG Compression Against Adversarial Examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 860-868, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "text", + "content": "[36] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xi-anglong Liu, Jian Zhang, and Jingkuan Song. Frequency Domain Model Augmentation for Adversarial Attack. In Proceedings of the European Conference on Computer Vision, pages 549–566, 2022. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "type": "text", + "content": "[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards Deep Learning Models Resistant to Adversarial Attacks. In Proceedings of the International Conference on Learning Representations, 2018. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "type": "text", + "content": "[38] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2574-2582, 2016. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[39] Muzammal Naseer, Salman H. Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A Self-supervised Approach for Adversarial Robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 259-268, 2020. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "text", + "content": "[40] Russakovsky Olga, Deng Jia, Su Hao, Krause Jonathan, Satheesh Sanjeev, Ma Sean, hUANG Zhiheng, Karpathy Andrej, Khosla Adiya, and Bernstein Michael et al. Imagenet large scale visual recognition challenge. In International Journal of Computer Vision, pages 211-252, 2015. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "text", + "content": "[41] Florian Schroff, Dmitry Kalenichenko, and James Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815-823, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[42] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K. Reiter. Accessorize to a Crime: Real and Stealthy Attacks on State-of-the-Art Face Recognition. In Proceedings of the ACM Conference on Computer and Communications Security, pages 1528-1540, 2016. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[43] Luchuan Song, Zheng Fang, Xiaodan Li, Xiaoyi Dong, Zhenchao Jin, Yuefeng Chen, and Siwei Lyu. Adaptive face forgery detection in cross domain. In European Conference on Computer Vision, pages 467-484. Springer, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[44] Luchuan Song, Xiaodan Li, Zheng Fang, Zhenchao Jin, YueFeng Chen, and Chenliang Xu. Face forgery detection via symmetric transformer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4102-4111, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 173, + 545, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 225 + ], + "type": "text", + "content": "[45] Luchuan Song, Bin Liu, Guojun Yin, Xiaoyi Dong, Yufei Zhang, and Jia-Xuan Bai. Tacr-net: editing on deep video and voice portraits. In Proceedings of the 29th ACM International Conference on Multimedia, pages 478-486, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 228, + 545, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 280 + ], + "type": "text", + "content": "[46] Nitish Srivastava, Geoffrey E. Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a Simple Way to prevent Neural Networks from Overfitting. The Journal of Machine Learning Research, 15(1):1929-1958, 2014. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 282, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 545, + 336 + ], + "type": "text", + "content": "[47] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing Properties of Neural Networks. In Proceedings of the International Conference on Learning Representations, 2014. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "text", + "content": "[48] Florian Tramèr, Alexey Kurakin, Nicolas Papernot, Ian J. Goodfellow, Dan Boneh, and Patrick D. McDaniel. Ensemble Adversarial Training: Attacks and Defenses. In Proceedings of the International Conference on Learning Representations, 2018. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 393, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 446 + ], + "type": "text", + "content": "[49] Jonathan Uesato, Brendan O'Donoghue, Pushmeet Kohli, and Aïron van den Oord. Adversarial Risk and the Dangers of Evaluating Against Weak Attacks. In Proceedings of the International Conference on Machine Learning, pages 5032-5041, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 502 + ], + "type": "text", + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Proceedings of the Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "type": "text", + "content": "[51] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. CosFace: Large Margin Cosine Loss for Deep Face Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5265-5274, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "text", + "content": "[52] Xiaosen Wang, Jin Hao, Yichen Yang, and Kun He. Natural Language Adversarial Defense through Synonym Encoding. In Proceedings of the Conference on Uncertainty in Artificial Intelligence, pages 823-833, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 603, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 646 + ], + "type": "text", + "content": "[53] Xiaosen Wang and Kun He. Enhancing the Transferability of Adversarial Attacks Through Variance Tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1924-1933, 2021. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[54] Xiaosen Wang, Kun He, Chuanbiao Song, Liwei Wang, and John E Hopcroft. AT-GAN: An Adversarial Generator Model for Non-constrained Adversarial Examples. arXiv preprint arXiv:1904.07793, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[55] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the Transferability of Adversarial At" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4617" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "tacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16138-16147, 2021. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[56] Xiaosen Wang, Jiadong Lin, Han Hu, Jingdong Wang, and Kun He. Boosting Adversarial Transferability through Enhanced Momentum. In Proceedings of the British Machine Vision Conference, page 272, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "text", + "content": "[57] Xiaosen Wang, Zeliang Zhang, Kangheng Tong, Dihong Gong, Kun He, Zhifeng Li, and Wei Liu. Triangle Attack: A Query-Efficient Decision-Based Adversarial Attack. In Proceedings of the European Conference on Computer Vision, pages 156-174, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 258 + ], + "type": "text", + "content": "[58] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable Adversarial Attacks for Image and Video Object Detection. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 954-960, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 260, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 260, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 260, + 287, + 303 + ], + "type": "text", + "content": "[59] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is Better than Free: Revisiting Adversarial Training. In Proceedings of the International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 358 + ], + "type": "text", + "content": "[60] Dongxian Wu, Yisen Wang, Shu-Tao Xia, James Bailey, and Xingjun Ma. Skip Connections Matter: On the Transferability of Adversarial Examples Generated with ResNets. In Proceedings of the International Conference on Learning Representations, 2020. 1, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "type": "text", + "content": "[61] Weibin Wu, Yuxin Su, Xixian Chen, Shenglin Zhao, Irwin King, Michael R. Lyu, and Yu-Wing Tai. Boosting the Transferability of Adversarial Samples via Attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1158-1167, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "text", + "content": "[62] Weibin Wu, Yuxin Su, Michael R Lyu, and Irwin King. Improving the Transferability of Adversarial Samples with Adversarial Transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9024-9033, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[63] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan L. Yuille. Mitigating Adversarial Effects Through Randomization. In Proceedings of the International Conference on Learning Representations, 2018. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 569 + ], + "type": "text", + "content": "[64] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L. Yuille. Improving Transferability of Adversarial Examples With Input Diversity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2730-2739, 2019. 1, 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "text", + "content": "[65] Saining Xie, Ross B. Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated Residual Transformations for Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987-5995, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "text", + "content": "[66] Yifeng Xiong, Jiadong Lin, Min Zhang, John E. Hopcroft, and Kun He. Stochastic Variance Reduced Ensemble Adversarial Attack for Boosting the Adversarial Transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14972, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "text", + "content": "[67] Weilin Xu, David Evans, and Yanjun Qi. Feature Squeezing: Detecting Adversarial Examples in Deep Neural Networks." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the Network and Distributed System Security Symposium, 2018. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[68] Bo Yang, Hengwei Zhang, Zheming Li, Yuchen Zhang, Kaiyong Xu, and Jindong Wang. Adversarial example generation with adbelief optimizer and crop invariance. Applied Intelligence, 53(2):2332-2347, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[69] Yichen Yang, Xiaosen Wang, and Kun He. Robust Textual Embedding against Word-level Adversarial Attacks. pages 2214-2224, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "text", + "content": "[70] Zhuolin Yang, Bo Li, Pin-Yu Chen, and Dawn Song. Characterizing Audio Adversarial Examples Using Temporal Dependency. In Proceedings of the International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "text", + "content": "[71] Shengming Yuan, Qilong Zhang, Lianli Gao, Yaya Cheng, and Jingkuan Song. Natural Color Fool: Towards Boosting Black-box Unrestricted Attacks. In Proceedings of the Advances in Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 308 + ], + "type": "text", + "content": "[72] Zheng Yuan, Jie Zhang, and Shiguang Shan. Adaptive image transformations for transfer-based adversarial attack. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 310, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 365 + ], + "type": "text", + "content": "[73] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6022-6031, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 366, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 421 + ], + "type": "text", + "content": "[74] Huan Zhang, Hongge Chen, Chaowei Xiao, Sven Gowal, Robert Stanforth, Bo Li, Duane S. Boning, and Cho-Jui Hsieh. Towards Stable and Efficient Training of Verifiably Robust Neural Networks. In Proceedings of the International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 466 + ], + "type": "text", + "content": "[75] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. Mixup: Beyond Empirical Risk Minimization. In Proceedings of the International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 468, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 533 + ], + "type": "text", + "content": "[76] Jianping Zhang, Jen-tse Huang, Wenxuan Wang, Yichen Li, Weibin Wu, Xiaosen Wang, Yuxin Su, and Michael R. Lyu. Improving the Transferability of Adversarial Samples by Path-Augmented Method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8173-8182, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "type": "text", + "content": "[77] Qilong Zhang, Xiaodan Li, Yuefeng Chen, Jingkuan Song, Lianli Gao, Yuan He, and Hui Xue. Beyond ImageNet Attack: Towards Crafting Adversarial Examples for Black-box Domains. In Proceedings of the International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 634 + ], + "type": "text", + "content": "[78] Qilong Zhang, Chaoning Zhang, Chaoqun Li, Jingkuan Song, Lianli Gao, and Heng Tao Shen. Practical No-box Adversarial Attacks with Training-free Hybrid Image Transformation. arXiv preprint arXiv:2203.04607, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[79] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[80] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable Ad" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4618" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 152 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "text", + "content": "versarial Perturbations. In Proceedings of the European Conference on Computer Vision, pages 471-486, 2018. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 152 + ], + "type": "text", + "content": "[81] Junhua Zou, Zhisong Pan, Junyang Qiu, Xin Liu, Ting Rui, and Wei Li. Improving the Transferability of Adversarial Examples with Resized-Diverse-Inputs, Diversity-Ensemble and Region Fitting. In Proceedings of the European Conference on Computer Vision, pages 563–579, 2020. 2, 5" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "4619" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/4fea04cc-819a-4990-99d9-0ccc1f6c300d_content_list.json b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/4fea04cc-819a-4990-99d9-0ccc1f6c300d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..87549d913b3db4214fcbb8d641f34c06163dc901 --- /dev/null +++ b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/4fea04cc-819a-4990-99d9-0ccc1f6c300d_content_list.json @@ -0,0 +1,1617 @@ +[ + { + "type": "text", + "text": "Structure and Content-Guided Video Synthesis with Diffusion Models", + "text_level": 1, + "bbox": [ + 132, + 130, + 836, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Patrick Esser Johnathan Chiu Parmida Atighehchian Jonathan Granskog Anastasis Germanidis Runway https://research.runwayml.com/gen1", + "bbox": [ + 199, + 180, + 769, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f1f4460496ae43009365495e3372b877ee4e114a30ba3847e5b92b434710795.jpg", + "image_caption": [ + "Figure 1. Guided Video Synthesis We present an approach based on latent video diffusion models that synthesizes videos (top and bottom) guided by content described through text (top) or images (bottom) while keeping the structure of an input video (middle)." + ], + "image_footnote": [], + "bbox": [ + 91, + 300, + 890, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 568, + 313, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-guided generative diffusion models unlock powerful image creation and editing tools. Recent approaches that edit the content of footage while retaining structure require expensive re-training for every input or rely on error-prone propagation of image edits across frames.", + "bbox": [ + 75, + 614, + 470, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we present a structure and content-guided video diffusion model that edits videos based on descriptions of the desired output. Conflicts between user-provided content edits and structure representations occur due to insufficient disentanglement between the two aspects. As a solution, we show that training on monocular depth estimates with varying levels of detail provides control over structure and content fidelity. A novel guidance method, enabled by joint video and image training, exposes explicit control over temporal consistency. Our experiments demonstrate a wide variety of successes; fine-grained control over output characteristics, customization based on a few reference images, and a strong user preference towards results by our model.", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 616, + 630, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Demand for more intuitive and performant video editing tools has increased as video-centric platforms have been popularized. But editing in the format is still complex and time-consuming due the temporal nature of video data. State-of-the-art machine learning models have shown great promise in improving editing workflows.", + "bbox": [ + 496, + 642, + 890, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative approaches for image synthesis recently experienced a rapid surge in quality and popularity due to the introduction of powerful diffusion models trained on largescale datasets. Text-conditioned models, such as DALLE 2 [34] and Stable Diffusion [38], enable novice users to generate detailed imagery given only a text prompt as input. Latent diffusion models especially enable efficient methods for producing imagery via synthesis in a perceptually compressed space.", + "bbox": [ + 496, + 734, + 890, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by this progress, we investigate generative models suited for interactive applications in video editing.", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7346", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current methods repurpose existing image models by either propagating edits with approaches that compute explicit correspondences [5] or by finetuning on each individual video [63]. We aim to circumvent expensive per-video training and correspondence calculation to achieve fast inference for arbitrary videos.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a controllable structure and content-aware latent video diffusion model trained on a large-scale dataset of uncaptioned videos and images. We opt to represent structure with monocular depth estimates, and content with embeddings predicted by a pre-trained neural network. Our approach offers several powerful modes of control. First, we train our model such that the content of inferred videos, e.g. their appearance or style, match user-provided images or text prompts (Fig. 1). Second, we vary the fidelity of the structure representation during training to allow selecting the strength of the structure preservation at test-time. Finally, we also adjust the inference process via a custom guidance method, inspired by classifier-free guidance, to enable control over temporal consistency.", + "bbox": [ + 75, + 181, + 470, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we present the following contributions:", + "bbox": [ + 96, + 393, + 441, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We extend latent diffusion models to video generation by introducing temporal layers into a pre-trained image model and by joint training on images and videos.", + "- We present a structure and content-aware model that edits videos given example images or text. Our method does not require per-video training or pre-processing.", + "- We demonstrate full control over temporal, content and structure consistency. We show for the first time that joint image-video training enables control over temporal stability. And, training on varying levels of detail in the structure representation allows choosing the desired level of preservation during inference.", + "- We show that our approach is preferred over several other approaches in a user study. We further improve the accuracy of previously unseen content by finetuning on a small set of images of the desired subject." + ], + "bbox": [ + 96, + 417, + 468, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 709, + 218, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Controllable video editing and media synthesis is an active area of research. In this section, we review prior work in related areas and connect our method to these approaches.", + "bbox": [ + 75, + 734, + 468, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unconditional video generation Generative adversarial networks (GANs) [12] can learn to synthesize videos based on specific training data [59, 45, 1, 56]. These methods often struggle with stability during optimization, and produce fixed-length videos [59, 45] or longer videos where artifacts accumulate over time [50]. [6] synthesize longer videos using a GAN with a better encoding of the time axis. Autoregressive transformers have also been proposed for uncon-", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ditional video generation [11, 64]. Our focus is on providing user control over the synthesis process whereas these approaches are limited to sampling random content resembling their training distribution.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models for image synthesis Diffusion models (DMs) [51, 53] can synthesize detailed media in many formats, such as images [34, 38], 3d shapes [66] and animations [54]. Many works improve diffusion-based image synthesis by changing the parameterization [14, 27, 46], introducing advanced sampling methods [52, 24, 22, 47, 20], designing more powerful architectures [3, 15, 57, 30], or conditioning on additional information [25]. Text-conditioning, based on embeddings from CLIP [32] or T5 [33], has become a particularly powerful approach for providing artistic control over model output [44, 28, 34, 3, 65, 10]. Latent diffusion models (LDMs) [38] perform diffusion in a compressed latent space reducing memory requirements and runtime. Our video model is an LDM trained simultaneously on videos and images.", + "bbox": [ + 496, + 152, + 892, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models for video synthesis Recently, diffusion models, masked generative models and autoregressive models have been applied to text-conditioned video synthesis [17, 13, 58, 67, 18, 49]. Similar to [17] and [49], we extend image synthesis diffusion models to video generation by introducing temporal connections into a pre-existing image model. Our model edits videos rather than synthesizing them from scratch. We demonstrate through a user study that our model with explicit conditioning over structure is preferred over other related approaches.", + "bbox": [ + 496, + 382, + 890, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video translation and propagation Image-to-image translation models, such as pix2pix [19, 62], can modify each frame in a video individually. This produces temporal inconsistencies as the time axis is ignored. Accounting for temporal or geometric information, such as flow, can increase consistency across frames when repurposing image synthesis models [42, 9]. We can extract such structural information to aid our spatio-temporal LDM in text- and image-guided video synthesis. Many generative adversarial methods, such as vid2vid [61, 60], leverage this type of input to guide synthesis.", + "bbox": [ + 496, + 536, + 890, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video style transfer takes a reference style image and statistically applies its style to an input video [40, 8, 55]. In contrast, our method edits both style and content while preserving the structure of a video instead of matching feature statistics only. Text2Live [5] allows editing input videos using text prompts by decomposing a video into neural layers [21]. Once available, a layered video representation [37] provides consistent propagation across frames. SinFusion [29] and Tune-a-Video [63] use diffusion models to edit videos but require per-video training. This limits the practicality of the approaches in creative tools. We opt to instead train our model on a large-scale dataset permitting inference on any video without individual training.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7347", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/49a6b142c1d07da6f025c6c46f1ccbb1a3ecf0c53b66bbd7274985a943e959d0.jpg", + "image_caption": [ + "Figure 2. Overview: During training (left), input videos $x$ are encoded to $z_{0}$ with a fixed encoder $\\mathcal{E}$ and diffused to $z_{t}$ . We extract a structure representation $s$ by encoding depth maps obtained with MiDaS, and a content representation $c$ by encoding one of the frames with CLIP. The model then learns to reverse the diffusion process in the latent space, with the help of $s$ , which gets concatenated to $z_{t}$ , as well as $c$ , which is provided via cross-attention blocks. During inference (right), the structure $s$ of an input video is provided in the same manner. To specify content via text, we convert CLIP text embeddings to image embeddings via a prior." + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 459, + 241 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/07748d2211bea29d7fb16cd61303243e0afac993e41a17118137d1da77979aac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 468, + 90, + 887, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 339, + 168, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For our purposes, it will be helpful to think of a video in terms of its content and structure. By structure, we refer to characteristics describing geometry and dynamics, e.g. shapes and locations of subjects as well as their temporal changes. We define content as features describing the appearance and semantics of the video, such as the colors and styles of objects and the lighting. The goal of our model is to edit the content of a video while retaining its structure.", + "bbox": [ + 75, + 372, + 468, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve this, we learn a generative model $p(x|s,c)$ of videos $x$ conditioned on representations of structure $s$ and content $c$ . We infer the shape representation $s$ from an input video, and modify it based on a text prompt $c$ describing the edit. First, we describe our realization of the generative model as a conditional latent video diffusion model and, then, we describe our choices for shape and content representations. Finally, we discuss the optimization process of our model. See Fig. 2 for an overview.", + "bbox": [ + 75, + 501, + 472, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Latent diffusion models", + "text_level": 1, + "bbox": [ + 76, + 669, + 294, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models Diffusion models [51] learn to reverse a fixed forward diffusion process, which is defined as", + "bbox": [ + 75, + 700, + 468, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) := \\mathcal {N} \\left(x _ {t}, \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathcal {I}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 757, + 468, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Normally-distributed noise is slowly added to each sample $x_{t-1}$ to obtain $x_t$ . The forward process models a fixed Markov chain and the noise is dependent on a variance schedule $\\beta_t$ where $t \\in \\{1, \\dots, T\\}$ , with $T$ being the total number of steps in our diffusion chain, and $x_0 \\coloneqq x$ .", + "bbox": [ + 75, + 801, + 468, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning to Denoise The reverse process is defined accord", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing to the following equation with parameters $\\theta$", + "bbox": [ + 498, + 340, + 815, + 356 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {0}\\right) := \\int p _ {\\theta} \\left(x _ {0: T}\\right) d x _ {1: T} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 364, + 890, + 396 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {0: T}\\right) = p \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 398, + 890, + 440 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) := \\mathcal {N} \\left(x _ {t - 1}, \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 443, + 890, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using a fixed variance $\\Sigma_{\\theta}(x_t,t)$ , we are left learning the means of the reverse process $\\mu_{\\theta}(x_t,t)$ . Training is typically performed via a reweighted variational bound on the maximum likelihood objective, resulting in a loss", + "bbox": [ + 496, + 470, + 890, + 531 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL := \\mathbb {E} _ {t, q} \\lambda_ {t} \\| \\mu_ {t} \\left(x _ {t}, x _ {0}\\right) - \\mu_ {\\theta} \\left(x _ {t}, t\\right) \\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 540, + 890, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_t(x_t, x_0)$ is the mean of the forward process posterior $q(x_{t-1} | x_t, x_0)$ , which is available in closed form [14].", + "bbox": [ + 496, + 568, + 890, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameterization The mean $\\mu_{\\theta}(x_t,t)$ is then predicted by a UNet architecture [39] that receives the noisy input $x_{t}$ and the diffusion timestep $t$ as inputs. Other parameterizations and weightings, such as $\\epsilon$ [14] and $v$ -parameterizations [46], can significantly improve sample quality compared to directly predicting the mean. Similar to [13], we found that $v$ -parameterization improves color consistency thus all our experiments use it (see supp. material for more details).", + "bbox": [ + 496, + 599, + 890, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent diffusion Latent diffusion models [38] (LDMs) take the diffusion process into the latent space. This provides improved separation between compressive and generative learning phases of the model. Specifically, LDMs use an autoencoder where an encoder $\\mathcal{E}$ maps input data $x$ to a lower dimensional latent code according to $z = \\mathcal{E}(x)$ while a decoder $\\mathcal{D}$ converts latent codes back to the input space such that perceptually $x \\approx \\mathcal{D}(\\mathcal{E}(x))$ .", + "bbox": [ + 496, + 719, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our encoder downsamples RGB-images $x\\in \\mathbb{R}^{3\\times H\\times W}$ by a factor of eight and outputs four channels, resulting in a latent code $z\\in \\mathbb{R}^{4\\times H / 8\\times W / 8}$ . Thus, the diffusion UNet operates on a much smaller representation which significantly", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7348", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1ce833b0bd6217de8e60b9a56f025de8b65431fc08cc7c2b25aec479669c2868.jpg", + "image_caption": [ + "Figure 3. Temporal Extension: We extend an image-based UNet architecture to videos, by adding temporal layers in its building blocks. We add a 1D temporal convolution after each 2D spatial convolution in its residual blocks (left), and we add a 1D temporal attention block after each of its 2D spatial attention blocks (right)." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 269, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a1d8c064045f53420d1d4f077e731bf29ffffca4f42880d83270ffcfc2690ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 90, + 467, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "improves runtime and memory efficiency. The latter is crucial for video modeling where the additional time axis increases memory costs.", + "bbox": [ + 75, + 271, + 468, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Spatio-temporal Latent Diffusion", + "text_level": 1, + "bbox": [ + 76, + 324, + 369, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To correctly model a distribution over video frames, the architecture must account for temporal relationships. We also want to jointly learn an image model with shared parameters to benefit from better generalization obtained by training on large-scale image datasets.", + "bbox": [ + 75, + 347, + 468, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this, we extend an image architecture by introducing temporal layers, which are only active for video inputs. All other layers are shared between the image and video model. The autoencoder remains fixed and processes each frame in a video independently.", + "bbox": [ + 75, + 422, + 468, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The UNet consists of two main building blocks: Residual blocks and transformer blocks (see Fig. 3). Similar to [17, 49], we extend them to videos by adding both 1D convolutions across time and 1D self-attention across time. In each residual block, we introduce one temporal convolution after each 2D convolution. Similarly, after each spatial 2D transformer block, we also include one temporal 1D transformer block, which mimics its spatial counterpart along the time axis. We also input learnable positional encodings of the frame index into temporal transformer blocks.", + "bbox": [ + 75, + 498, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Representing Content and Structure", + "text_level": 1, + "bbox": [ + 76, + 657, + 393, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditional Diffusion Models Diffusion models are well-suited to modeling conditional distributions such as $p(x|s, c)$ . The forward process $q$ remains unchanged while the conditioning variables $s, c$ become additional inputs to the model.", + "bbox": [ + 75, + 680, + 468, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to edit a video based on a text prompt describing the desired output. We choose to train on uncaptioned video data due to the lack of large-scale paired video-text datasets of similar quality as image datasets like [48]. Therefore, during training, we must derive structure and content representations from the training video $x$ itself, i.e. $s = s(x)$ and $c = c(x)$ , resulting in a per-example loss of", + "bbox": [ + 75, + 756, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {t} \\left\\| \\mu_ {t} \\left(\\mathcal {E} (x) _ {t}, \\mathcal {E} (x) _ {0}\\right) - \\mu_ {\\theta} \\left(\\mathcal {E} (x) _ {t}, t, s (x), c (x)\\right) \\right\\| ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, during inference, structure $s$ and content $c$ are derived from an input video $y$ and from a text prompt $t$ respectively. An edited version $x$ of $y$ is obtained by sampling the generative model conditioned on $s(y)$ and $c(t)$ :", + "bbox": [ + 496, + 90, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz \\sim p _ {\\theta} (z | s (y), c (t)), x = \\mathcal {D} (z). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 162, + 890, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Content Representation We utilize CLIP [32] to infer a content representation from both text inputs $t$ and video inputs $x$ similar to previous works [35, 3]. CLIP embeddings are a promising content representation as they are more sensitive to semantic and stylistic properties while being more invariant towards geometric attributes [34]. During training, we encode a random frame in each input video with CLIP. To support text-based editing at inference, we train a prior model that allows sampling image embeddings from text embeddings [35, 49].", + "bbox": [ + 496, + 190, + 890, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structure Representation We need a representation that provides adequate separation between structure and content. We find that depth estimates extracted from input video frames provide the desired properties as they encode significantly less content information compared to simpler structure representations, such as edge filters which also encode textural properties. Still, depth maps reveal the silhouettes of objects which can prevent content edits involving changes in object shape.", + "bbox": [ + 496, + 340, + 890, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To offer control over the amount of structure to preserve, we propose to train a model on structure representations with varying amounts of information. In particular, we blur depth estimates given a parameter $t_s$ . During training, $t_s$ is randomly sampled between 0 and $T_s$ . The parameter can then be controlled at inference to achieve different editing effects (see Fig. 10).", + "bbox": [ + 496, + 478, + 890, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While depths map work well for our usecase, our approach generalizes to other geometric guidance features or combinations of features that might be more helpful for other specific applications. For example, models focusing on human video synthesis might benefit from estimated poses or face landmarks.", + "bbox": [ + 496, + 584, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditioning Mechanisms We account for the different characteristics of our content and structure representations with two different conditioning mechanisms. Since structure represents a significant portion of the spatial information of video frames, we use concatenation for conditioning to make effective use of this information. In contrast, attributes described by the content representation are not tied to particular locations. Hence, we leverage cross-attention which can effectively transport this information to any position.", + "bbox": [ + 496, + 672, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the spatial transformer blocks of the UNet architecture for cross-attention conditioning. Each contains two attention operations, where the first one performs a spatial self-attention and the second one a cross attention with keys and values computed from the CLIP image embedding.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7349", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d1517ada36b78af12d70bb7b303f9901bde5e96dcacaf865c363dd08c565ad71.jpg", + "image_caption": [ + "Figure 4. Temporal Control: By training image and video models jointly, we obtain explicit control over the temporal consistency of edited videos via a temporal guidance scale $\\omega_{t}$ . On the left, frame consistency measured via CLIP cosine similarity of consecutive frames increases monotonically with $\\omega_{t}$ , while mean squared error between frames warped with optical flow decreases monotonically. On the right, lower scales (0.5 in the middle row) achieve edits with a \"hand-drawn\" look, whereas higher scales (1.5 in the bottom row) result in smoother results. Top row shows the original input video, the two edits use the prompt \"pencil sketch of a man looking at the camera\"." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 405, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/31ee90a69d58b6e62cc7bd826b5b2059575e76ab0ffdb225891fbd5ff8b13359.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prompt", + "text_level": 1, + "bbox": [ + 122, + 299, + 176, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Driving Video (top) and Result (bottom)", + "text_level": 1, + "bbox": [ + 401, + 299, + 668, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a man using a laptop inside a train, anime style", + "bbox": [ + 106, + 351, + 191, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a woman and man take selfies while walking down the street, claymation", + "bbox": [ + 106, + 455, + 191, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "kite-surfer in the ocean at sunset", + "bbox": [ + 106, + 614, + 189, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "alien expplorer hiking in the mountains", + "bbox": [ + 106, + 727, + 191, + 800 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ce111ac29162328e1eaf09914a94e114951db0b7d4109df2605d4b4293350891.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 319, + 859, + 441 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e42bb00d6644197b0e66ec95016b9adf507f5e4b8b075ee34bd03f8360ebfbd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 446, + 859, + 569 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fda5ce90649ab45342be472ae5f4aa0bf1a728a2736ce4f19c10a255d36e4b16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 575, + 859, + 696 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/85be900d6f5fd989d4f2db2b2f03b2f5e450916997cf48be6ef634e217f8efa5.jpg", + "image_caption": [ + "Figure 5. Our approach enables a wide range of video edits, including changes to animation styles such as anime or claymation, changes of environment such as time of day, and changing characters such as humans to aliens." + ], + "image_footnote": [], + "bbox": [ + 207, + 700, + 857, + 823 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7350", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2ca212a548bf5cddbe32e0509d4c948c14f09dd9609dbf091c24bfb251e6e1ef.jpg", + "image_caption": [ + "Figure 6. Prompt-vs-frame consistency: Image models such as SD-Depth achieve good prompt consistency but fail to produce consistent edits across frames. Propagation based approaches such as IVS and Text2Live increase frame consistency but fail to provide edits reflecting the prompt accurately. Our method achieves the best combination of frame and prompt consistency." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 472, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To condition on structure, we first estimate depth maps for all input frames using the MiDaS DPT-Large model [36]. We then apply $t_s$ iterations of blurring and downsampling to the depth maps, where $t_s$ controls the amount of structure to preserve. We resample the perturbed depth map to the resolution of the RGB-frames and encode it using $\\mathcal{E}$ . This latent representation of structure is concatenated with the input $z_t$ given to the UNet. We also input four channels containing a sinusoidal embedding of $t_s$ .", + "bbox": [ + 75, + 345, + 468, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sampling While Eq. (2) provides a direct way to sample from the trained model, many other sampling methods [52, 24, 22] require only a fraction of the number of diffusion timesteps to achieve good sample quality. We use DDIM [52] throughout our experiments. Furthermore, classifier-free diffusion guidance [16] significantly improves sample quality. For a conditional model $\\mu_{\\theta}(x_t,t,c)$ , this is achieved by training the model to also perform unconditional predictions $\\mu_{\\theta}(x_t,t,\\emptyset)$ and then adjusting predictions during sampling according to", + "bbox": [ + 75, + 481, + 470, + 633 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mu} _ {\\theta} (x _ {t}, t, c) = \\mu_ {\\theta} (x _ {t}, t, \\emptyset) + \\omega (\\mu_ {\\theta} (x _ {t}, t, c) - \\mu_ {\\theta} (x _ {t}, t, \\emptyset))\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 643, + 465, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\omega$ is the guidance scale that controls the strength. Based on the intuition that $\\omega$ extrapolates the direction between an unconditional and a conditional model, we apply this idea to control temporal consistency of our model. Specifically, since we are training both an image and a video model with shared parameters, we can consider predictions by both models for the same input. Let $\\mu_{\\theta}(z_t,t,c,s)$ denote the prediction of our video model, and let $\\mu_{\\theta}^{\\pi}(z_t,t,c,s)$ denote the prediction of the image model applied to each frame individually. Taking classifier-free guidance for $c$ into account, we then adjust our prediction according to", + "bbox": [ + 75, + 671, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\mu} _ {\\theta} (z _ {t}, t, c, s) = \\mu_ {\\theta} ^ {\\pi} (z _ {t}, t, \\emptyset , s) \\\\ + \\omega_ {t} \\left(\\mu_ {\\theta} \\left(x _ {t}, t, \\emptyset , s\\right) - \\mu_ {\\theta} ^ {\\pi} \\left(x _ {t}, t, \\emptyset , s\\right)\\right) \\tag {8} \\\\ + \\omega (\\mu_ {\\theta} (x _ {t}, t, c, s) - \\mu_ {\\theta} (x _ {t}, t, \\emptyset , s)) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 848, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc2e31a43f198dfd5c9c37db17a2722366044e9ca9ae6494c9f9f94cbac41a27.jpg", + "image_caption": [ + "Figure 7. User Preferences: Based on our user study, the results from our model are preferred over the baseline models." + ], + "image_footnote": [ + "ours SDEdit (50% strength) IVS (50% strength) Depth-SDours (-, 50% SDEdit) Deformu SDEdit (75% strength) IVS (75% strength) Text2LIVEours (-, 75% SDEdit)" + ], + "bbox": [ + 501, + 113, + 890, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiments demonstrate that this approach controls temporal consistency in the outputs, see Fig. 4.", + "bbox": [ + 498, + 300, + 890, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Optimization", + "text_level": 1, + "bbox": [ + 500, + 340, + 638, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train on an internal dataset of 240M images and 6.4M video clips. We use image batches of size 9216 with resolutions of $320 \\times 320$ , $384 \\times 320$ and $448 \\times 256$ , as well as the same resolutions with flipped aspect ratios. We sample image batches with a probability of $12.5\\%$ . For videos, we use batch size 1152 and 8 frames from each video sampled four frames apart with a resolution of $448 \\times 256$ .", + "bbox": [ + 496, + 364, + 890, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model in multiple stages. First, we initialize model weights based on a pretrained text-conditional latent diffusion model $[38]^1$ . We change the conditioning from CLIP text embeddings to CLIP image embeddings and fine-tune for 15k steps on images only. Afterwards, we introduce temporal connections as described in Sec. 3.2 and train jointly on images and videos for 75k steps. We then add conditioning on structure $s$ with $t_s \\equiv 0$ fixed and train for 25k steps. Finally, we resume training with $t_s$ sampled uniformly between 0 and 7 for another 10k steps.", + "bbox": [ + 496, + 470, + 892, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 636, + 586, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate our approach, we use videos from DAVIS [31] and various stock footage. To automatically create edit prompts, we first run a captioning model [23] to obtain a description of the original video content. We then use GPT-3 [7] to generate edited prompts.", + "bbox": [ + 498, + 661, + 890, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 747, + 681, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate that our approach performs well on a number of diverse inputs (see Fig. 5). Our method handles a large variety of footage, such as landscapes and close-ups, and diverse camera motion without any explicit tracking of the input. Our depth-based structure representation combined with large-scale image-video joint training enable strong generalization and powerful editing capabilities.", + "bbox": [ + 496, + 770, + 890, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://github.com/runwayml/stable-diffusion", + "bbox": [ + 517, + 887, + 872, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7351", + "bbox": [ + 482, + 924, + 513, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/896f9d2efe6f6ac540f6af75a3640a0bab9f6c20ebcec5d086b90e2566d5a444.jpg", + "image_caption": [ + "Input", + "Mask", + "A snowboarder in a snow park on the mountain", + "Figure 8. Background Editing: Masking the denoising process allows us to restrict edits to backgrounds for more control over results." + ], + "image_footnote": [], + "bbox": [ + 207, + 88, + 859, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ties. For example, we can produce various animation styles, changes in time of day, and more complex changes of subject, such as turning a hiker into an alien (see Fig. 5). Please see the supplementary material for more results.", + "bbox": [ + 75, + 309, + 467, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using CLIP image embeddings as the content representation allows users to specify content through images. As an example application, we demonstrate character replacement in Fig. 9. For every video in a set of six videos, we resynthesize it five times, each time providing a single content image taken from another video in the set. We can retain content characteristics with $t_s = 3$ despite large differences in their pose and shape.", + "bbox": [ + 75, + 369, + 467, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, we illustrate the use of masked video editing in Fig. 8, where the model predicts everything outside the masked area(s) while retaining the original content inside the masked area. Notably, this technique resembles inpainting with diffusion models [43, 25].", + "bbox": [ + 75, + 489, + 467, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. User Study", + "text_level": 1, + "bbox": [ + 76, + 575, + 199, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We benchmark against Text2Live [5], a recent approach for text-guided video editing that employs layered neural atlases [21]. As a baseline, we compare against SDEdit [26] in two ways; per-frame generated results and a first-frame result propagated by a few-shot video stylization method [55] (IVS). We also include two depth-based versions of Stable Diffusion; one trained with depth-conditioning [2] and one that retains past results based on depth estimates [9]. We also include an ablation: applying SDEdit to our video model trained without conditioning on a structure representation (ours, $\\sim s$ ).", + "bbox": [ + 75, + 598, + 467, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We judge the success of our method qualitatively based on a user study. We run the user study using Amazon Mechanical Turk (AMT) on an evaluation set of 35 representative video editing prompts. For each example, we ask 5 annotators to compare faithfulness to the video editing prompt (\"Which video better represents the provided edited caption?\") between a baseline and our method, presented in random order, and use a majority vote for the final result.", + "bbox": [ + 75, + 763, + 467, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results can be found in Fig. 7. Across all compared", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods, results from our approach are preferred roughly 3 out of 4 times. A visual comparison among the methods can be found in the supplementary. We observe that SDEdit is sensitive to the editing strength. Low values fail to achieve the desired editing effect whereas high values change the structure of the input. Even with a fixed seed, both style and structure can change in unnatural ways between frames as their relationship is ignored by image-based approaches. Propagation of SDEdit outputs (IVS) leads to more consistent results but often introduces propagation artifacts especially with large motion. Depth-SD produces accurate, structure-preserving edits for individual frames but frames are inconsistent across time. The outputs of Text2Live tend to be temporally smooth due to its reliance on Layered Neural Atlases [21], but it often produces edits that represent the edit prompt inaccurately. A direct comparison with Text2Live is difficult as it requires input masks and separate edit prompts for foreground and background. In addition, computing a neural atlas takes about 10 hours whereas our approach requires approximately a minute.", + "bbox": [ + 496, + 309, + 890, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 619, + 720, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We quantify trade-offs between frame consistency and prompt consistency with the following two metrics.", + "bbox": [ + 496, + 643, + 890, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Frame consistency We compute CLIP image embeddings on all frames of output videos and report the average cosine similarity between all pairs of consecutive frames.", + "bbox": [ + 496, + 674, + 890, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt consistency We compute CLIP image embeddings on all frames of output videos and the CLIP text embedding of the edit prompt. We report average cosine similarity between text and image embedding over all frames.", + "bbox": [ + 496, + 719, + 890, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 shows the results of each model using our frame consistency and prompt consistency metrics. Our model tends to outperform the baseline models in both aspects (placed higher in the upper-right quadrant of the graph). We also notice a slight tradeoff with increasing the strength parameters in the baseline models: larger strength scales implies higher prompt consistency at the cost of lower frame consistency. Increasing the temporal scale $(\\omega_{t})$ of our model", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7352", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f6488259aa3253a32713eb6e4cf4dd0e6787e91732143fe84c6231696b88c09.jpg", + "image_caption": [ + "Figure 9. Image Prompting: We combine the structure of a driving video (first column) with content from other videos (first row)." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 468, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results in higher frame consistency but lower prompt consistency. We also observe that an increased structure scale $(t_{s})$ results in higher prompt consistency as the content becomes less determined by the input structure.", + "bbox": [ + 75, + 599, + 468, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.Customization", + "text_level": 1, + "bbox": [ + 76, + 676, + 225, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Customization of image models enables generation of previously unseen content, such as specific people or styles, based on a small dataset used for finetuning [41]. We fine-tune our depth-conditioned latent video diffusion model on a set of 15-30 images and produce videos containing the desired subject. Half of the batch elements are of the subject and the other half belong to the original training dataset.", + "bbox": [ + 75, + 702, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 10 shows an example with different numbers of customization steps as well as different levels of structure preservation $t_{s}$ . Customization improves fidelity to the style and appearance of the character. In combination with higher $t_{s}$ values, accurate animations are possible despite using a driving video of a person with different characteristics.", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79edf299909923c61304836278cf75d17787db030dc4005eb0536b31bffe848c.jpg", + "image_caption": [ + "Figure 10. Controlling Fidelity: We obtain control over structure and appearance-fidelity. Each cell shows three frames produced with decreasing structure-fidelity $t_s$ (left-to-right) and increasing number of customization training steps (top-to-bottom). The bottom shows examples of images used for customization (red border) and the input image (blue border). Same driving video as in Fig. 1." + ], + "image_footnote": [], + "bbox": [ + 500, + 71, + 890, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 601, + 617, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our latent video diffusion model synthesizes new videos given structure and content information. We ensure structural consistency by conditioning on depth estimates while content is controlled with images or natural language. Temporal layers and joint image-video training achieve stable results across frames. A novel guidance method, inspired by classifier-free guidance, allows for control over temporal consistency. By training on depth maps with varying degrees of detail, we can adjust the level of structure preservation. This, together with model customization, improves content fidelity. Our quantitative evaluation and user study show that our method is preferred over related approaches. Future works should investigate other conditioning data, such as facial landmarks and pose estimates, and additional 3d-priors to improve generated results. Our model is intended for creative applications in content creation, but we realize the risks of dual-use and hope that further work will be aimed at combating abuse of generative models.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7353", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards high resolution video generation with progressive growing of sliced Wasserstein gans. arXiv preprint arXiv:1810.02419, 2018. 2", + "[2] Stability AI. Stable diffusion depth. https://github.com/Stability-AI/stablediffusion, 2022. 7", + "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 4", + "[4] Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie S. Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise, 2023.", + "[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European Conference on Computer Vision, pages 707-723. Springer, 2022. 2, 7", + "[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A Efros, and Tero Karras. Generating long videos of dynamic scenes. 2022. 2", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc., 2020. 6", + "[8] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1105-1114, 2017. 2", + "[9] deforum. Deforum stable diffusion. https://github.com/deforum/stable-diffusion, 2022. 2, 7", + "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 19822-19835. Curran Associates, Inc., 2021. 2", + "[11] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh." + ], + "bbox": [ + 78, + 116, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Long video generation with time-agnostic vqgan and time-sensitive transformer. arXiv preprint arXiv:2204.03638, 2022. 2", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2", + "[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3", + "[15] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47-1, 2022. 2", + "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 6", + "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2, 4", + "[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2", + "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Proc. NeurIPS, 2022. 2", + "[21] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 2, 7", + "[22] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 2, 6", + "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 6", + "[24] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 6", + "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7354", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 7", + "[26] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. CoRR, abs/2108.01073, 2021. 7", + "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2", + "[28] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 2", + "[29] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2", + "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers, 2022. 2", + "[31] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 6", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4", + "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1), jun 2022. 2", + "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 1, 2, 4", + "[35] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8821-8831. PMLR, 18-24 Jul 2021. 4", + "[36] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:1623-1637, 2019. 6", + "[37] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew William Fitzgibbon. Unwrap mosaics: a new repre" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sentation for video editing. ACM SIGGRAPH 2008 papers, 2008. 2", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2, 3, 6", + "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3", + "[40] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In Bodo Rosenhahn and Bjoern Andres, editors, Pattern Recognition, pages 26-36, Cham, 2016. Springer International Publishing. 2", + "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 8", + "[42] Alexander S. Disco diffusion v5.2 - warp fusion. https://github.com/Sxela/DiscoDiffusion-Warp, 2022. 2", + "[43] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models, 2021. 7", + "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2", + "[45] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Proceedings of the IEEE international conference on computer vision, pages 2830-2839, 2017. 2", + "[46] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 2, 3", + "[47] Robin San-Roman, Eliya Nachmani, and Lior Wolf. Noise estimation for generative diffusion models. arXiv preprint arXiv:2104.02600, 2021. 2", + "[48] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 4", + "[49] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 2, 4", + "[50] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2, 2021. 2", + "[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7355", + "bbox": [ + 482, + 924, + 514, + 935 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of Machine Learning Research, pages 2256-2265, Lille, France, 07-09 Jul 2015. PMLR. 2, 3", + "[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 6", + "[53] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[54] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2", + "[55] Ondrej Texler, David Futschik, Michal Kucera, Ondrej Jamriška, Šárka Sochorová, Mengei Chai, Sergey Tulyakov, and Daniel Sýkora. Interactive video stylization using few-shot patch-based training. ACM Transactions on Graphics, 39(4):73, 2020. 2, 7", + "[56] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2", + "[57] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. Advances in Neural Information Processing Systems, 34:11287-11302, 2021. 2", + "[58] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description, 2022. 2", + "[59] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2", + "[60] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2", + "[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 2", + "[62] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2", + "[63] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2", + "[64] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2", + "[65] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation, 2022. 2", + "[66] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[67] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 233 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7356", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/4fea04cc-819a-4990-99d9-0ccc1f6c300d_model.json b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/4fea04cc-819a-4990-99d9-0ccc1f6c300d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..955dfd677544cd5c785f9e44edb512d0805f8966 --- /dev/null +++ b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/4fea04cc-819a-4990-99d9-0ccc1f6c300d_model.json @@ -0,0 +1,2455 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.133, + 0.131, + 0.838, + 0.152 + ], + "angle": 0, + "content": "Structure and Content-Guided Video Synthesis with Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.181, + 0.771, + 0.253 + ], + "angle": 0, + "content": "Patrick Esser Johnathan Chiu Parmida Atighehchian Jonathan Granskog Anastasis Germanidis Runway https://research.runwayml.com/gen1" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.301, + 0.891, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.505, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Figure 1. Guided Video Synthesis We present an approach based on latent video diffusion models that synthesizes videos (top and bottom) guided by content described through text (top) or images (bottom) while keeping the structure of an input video (middle)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.569, + 0.314, + 0.585 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.471, + 0.692 + ], + "angle": 0, + "content": "Text-guided generative diffusion models unlock powerful image creation and editing tools. Recent approaches that edit the content of footage while retaining structure require expensive re-training for every input or rely on error-prone propagation of image edits across frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this work, we present a structure and content-guided video diffusion model that edits videos based on descriptions of the desired output. Conflicts between user-provided content edits and structure representations occur due to insufficient disentanglement between the two aspects. As a solution, we show that training on monocular depth estimates with varying levels of detail provides control over structure and content fidelity. A novel guidance method, enabled by joint video and image training, exposes explicit control over temporal consistency. Our experiments demonstrate a wide variety of successes; fine-grained control over output characteristics, customization based on a few reference images, and a strong user preference towards results by our model." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.617, + 0.631, + 0.632 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Demand for more intuitive and performant video editing tools has increased as video-centric platforms have been popularized. But editing in the format is still complex and time-consuming due the temporal nature of video data. State-of-the-art machine learning models have shown great promise in improving editing workflows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Generative approaches for image synthesis recently experienced a rapid surge in quality and popularity due to the introduction of powerful diffusion models trained on largescale datasets. Text-conditioned models, such as DALLE 2 [34] and Stable Diffusion [38], enable novice users to generate detailed imagery given only a text prompt as input. Latent diffusion models especially enable efficient methods for producing imagery via synthesis in a perceptually compressed space." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Motivated by this progress, we investigate generative models suited for interactive applications in video editing." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7346" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "Current methods repurpose existing image models by either propagating edits with approaches that compute explicit correspondences [5] or by finetuning on each individual video [63]. We aim to circumvent expensive per-video training and correspondence calculation to achieve fast inference for arbitrary videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.394 + ], + "angle": 0, + "content": "We propose a controllable structure and content-aware latent video diffusion model trained on a large-scale dataset of uncaptioned videos and images. We opt to represent structure with monocular depth estimates, and content with embeddings predicted by a pre-trained neural network. Our approach offers several powerful modes of control. First, we train our model such that the content of inferred videos, e.g. their appearance or style, match user-provided images or text prompts (Fig. 1). Second, we vary the fidelity of the structure representation during training to allow selecting the strength of the structure preservation at test-time. Finally, we also adjust the inference process via a custom guidance method, inspired by classifier-free guidance, to enable control over temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.394, + 0.442, + 0.409 + ], + "angle": 0, + "content": "In summary, we present the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.418, + 0.468, + 0.463 + ], + "angle": 0, + "content": "- We extend latent diffusion models to video generation by introducing temporal layers into a pre-trained image model and by joint training on images and videos." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.473, + 0.469, + 0.519 + ], + "angle": 0, + "content": "- We present a structure and content-aware model that edits videos given example images or text. Our method does not require per-video training or pre-processing." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.528, + 0.469, + 0.618 + ], + "angle": 0, + "content": "- We demonstrate full control over temporal, content and structure consistency. We show for the first time that joint image-video training enables control over temporal stability. And, training on varying levels of detail in the structure representation allows choosing the desired level of preservation during inference." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.628, + 0.469, + 0.688 + ], + "angle": 0, + "content": "- We show that our approach is preferred over several other approaches in a user study. We further improve the accuracy of previously unseen content by finetuning on a small set of images of the desired subject." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.418, + 0.469, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.71, + 0.219, + 0.725 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Controllable video editing and media synthesis is an active area of research. In this section, we review prior work in related areas and connect our method to these approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Unconditional video generation Generative adversarial networks (GANs) [12] can learn to synthesize videos based on specific training data [59, 45, 1, 56]. These methods often struggle with stability during optimization, and produce fixed-length videos [59, 45] or longer videos where artifacts accumulate over time [50]. [6] synthesize longer videos using a GAN with a better encoding of the time axis. Autoregressive transformers have also been proposed for uncon-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "ditional video generation [11, 64]. Our focus is on providing user control over the synthesis process whereas these approaches are limited to sampling random content resembling their training distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.893, + 0.381 + ], + "angle": 0, + "content": "Diffusion models for image synthesis Diffusion models (DMs) [51, 53] can synthesize detailed media in many formats, such as images [34, 38], 3d shapes [66] and animations [54]. Many works improve diffusion-based image synthesis by changing the parameterization [14, 27, 46], introducing advanced sampling methods [52, 24, 22, 47, 20], designing more powerful architectures [3, 15, 57, 30], or conditioning on additional information [25]. Text-conditioning, based on embeddings from CLIP [32] or T5 [33], has become a particularly powerful approach for providing artistic control over model output [44, 28, 34, 3, 65, 10]. Latent diffusion models (LDMs) [38] perform diffusion in a compressed latent space reducing memory requirements and runtime. Our video model is an LDM trained simultaneously on videos and images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Diffusion models for video synthesis Recently, diffusion models, masked generative models and autoregressive models have been applied to text-conditioned video synthesis [17, 13, 58, 67, 18, 49]. Similar to [17] and [49], we extend image synthesis diffusion models to video generation by introducing temporal connections into a pre-existing image model. Our model edits videos rather than synthesizing them from scratch. We demonstrate through a user study that our model with explicit conditioning over structure is preferred over other related approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Video translation and propagation Image-to-image translation models, such as pix2pix [19, 62], can modify each frame in a video individually. This produces temporal inconsistencies as the time axis is ignored. Accounting for temporal or geometric information, such as flow, can increase consistency across frames when repurposing image synthesis models [42, 9]. We can extract such structural information to aid our spatio-temporal LDM in text- and image-guided video synthesis. Many generative adversarial methods, such as vid2vid [61, 60], leverage this type of input to guide synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video style transfer takes a reference style image and statistically applies its style to an input video [40, 8, 55]. In contrast, our method edits both style and content while preserving the structure of a video instead of matching feature statistics only. Text2Live [5] allows editing input videos using text prompts by decomposing a video into neural layers [21]. Once available, a layered video representation [37] provides consistent propagation across frames. SinFusion [29] and Tune-a-Video [63] use diffusion models to edit videos but require per-video training. This limits the practicality of the approaches in creative tools. We opt to instead train our model on a large-scale dataset permitting inference on any video without individual training." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7347" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.092, + 0.46, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.469, + 0.092, + 0.888, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.247, + 0.894, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Overview: During training (left), input videos \\( x \\) are encoded to \\( z_{0} \\) with a fixed encoder \\( \\mathcal{E} \\) and diffused to \\( z_{t} \\). We extract a structure representation \\( s \\) by encoding depth maps obtained with MiDaS, and a content representation \\( c \\) by encoding one of the frames with CLIP. The model then learns to reverse the diffusion process in the latent space, with the help of \\( s \\), which gets concatenated to \\( z_{t} \\), as well as \\( c \\), which is provided via cross-attention blocks. During inference (right), the structure \\( s \\) of an input video is provided in the same manner. To specify content via text, we convert CLIP text embeddings to image embeddings via a prior." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.34, + 0.169, + 0.357 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.495 + ], + "angle": 0, + "content": "For our purposes, it will be helpful to think of a video in terms of its content and structure. By structure, we refer to characteristics describing geometry and dynamics, e.g. shapes and locations of subjects as well as their temporal changes. We define content as features describing the appearance and semantics of the video, such as the colors and styles of objects and the lighting. The goal of our model is to edit the content of a video while retaining its structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.473, + 0.639 + ], + "angle": 0, + "content": "To achieve this, we learn a generative model \\( p(x|s,c) \\) of videos \\( x \\) conditioned on representations of structure \\( s \\) and content \\( c \\). We infer the shape representation \\( s \\) from an input video, and modify it based on a text prompt \\( c \\) describing the edit. First, we describe our realization of the generative model as a conditional latent video diffusion model and, then, we describe our choices for shape and content representations. Finally, we discuss the optimization process of our model. See Fig. 2 for an overview." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.67, + 0.295, + 0.685 + ], + "angle": 0, + "content": "3.1. Latent diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.469, + 0.732 + ], + "angle": 0, + "content": "Diffusion models Diffusion models [51] learn to reverse a fixed forward diffusion process, which is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.758, + 0.47, + 0.777 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) := \\mathcal {N} \\left(x _ {t}, \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathcal {I}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.469, + 0.879 + ], + "angle": 0, + "content": "Normally-distributed noise is slowly added to each sample \\( x_{t-1} \\) to obtain \\( x_t \\). The forward process models a fixed Markov chain and the noise is dependent on a variance schedule \\( \\beta_t \\) where \\( t \\in \\{1, \\dots, T\\} \\), with \\( T \\) being the total number of steps in our diffusion chain, and \\( x_0 \\coloneqq x \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Learning to Denoise The reverse process is defined accord" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.342, + 0.816, + 0.357 + ], + "angle": 0, + "content": "ing to the following equation with parameters \\(\\theta\\)" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.366, + 0.892, + 0.397 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {0}\\right) := \\int p _ {\\theta} \\left(x _ {0: T}\\right) d x _ {1: T} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.4, + 0.892, + 0.441 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {0: T}\\right) = p \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.444, + 0.892, + 0.461 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) := \\mathcal {N} \\left(x _ {t - 1}, \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Using a fixed variance \\(\\Sigma_{\\theta}(x_t,t)\\), we are left learning the means of the reverse process \\(\\mu_{\\theta}(x_t,t)\\). Training is typically performed via a reweighted variational bound on the maximum likelihood objective, resulting in a loss" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.541, + 0.892, + 0.56 + ], + "angle": 0, + "content": "\\[\nL := \\mathbb {E} _ {t, q} \\lambda_ {t} \\| \\mu_ {t} \\left(x _ {t}, x _ {0}\\right) - \\mu_ {\\theta} \\left(x _ {t}, t\\right) \\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.599 + ], + "angle": 0, + "content": "where \\(\\mu_t(x_t, x_0)\\) is the mean of the forward process posterior \\(q(x_{t-1} | x_t, x_0)\\), which is available in closed form [14]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Parameterization The mean \\(\\mu_{\\theta}(x_t,t)\\) is then predicted by a UNet architecture [39] that receives the noisy input \\(x_{t}\\) and the diffusion timestep \\(t\\) as inputs. Other parameterizations and weightings, such as \\(\\epsilon\\) [14] and \\(v\\)-parameterizations [46], can significantly improve sample quality compared to directly predicting the mean. Similar to [13], we found that \\(v\\)-parameterization improves color consistency thus all our experiments use it (see supp. material for more details)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Latent diffusion Latent diffusion models [38] (LDMs) take the diffusion process into the latent space. This provides improved separation between compressive and generative learning phases of the model. Specifically, LDMs use an autoencoder where an encoder \\(\\mathcal{E}\\) maps input data \\(x\\) to a lower dimensional latent code according to \\(z = \\mathcal{E}(x)\\) while a decoder \\(\\mathcal{D}\\) converts latent codes back to the input space such that perceptually \\(x \\approx \\mathcal{D}(\\mathcal{E}(x))\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our encoder downsamples RGB-images \\(x\\in \\mathbb{R}^{3\\times H\\times W}\\) by a factor of eight and outputs four channels, resulting in a latent code \\(z\\in \\mathbb{R}^{4\\times H / 8\\times W / 8}\\). Thus, the diffusion UNet operates on a much smaller representation which significantly" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7348" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.091, + 0.27, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.091, + 0.468, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.18, + 0.47, + 0.251 + ], + "angle": 0, + "content": "Figure 3. Temporal Extension: We extend an image-based UNet architecture to videos, by adding temporal layers in its building blocks. We add a 1D temporal convolution after each 2D spatial convolution in its residual blocks (left), and we add a 1D temporal attention block after each of its 2D spatial attention blocks (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.469, + 0.318 + ], + "angle": 0, + "content": "improves runtime and memory efficiency. The latter is crucial for video modeling where the additional time axis increases memory costs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.325, + 0.37, + 0.341 + ], + "angle": 0, + "content": "3.2. Spatio-temporal Latent Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.469, + 0.424 + ], + "angle": 0, + "content": "To correctly model a distribution over video frames, the architecture must account for temporal relationships. We also want to jointly learn an image model with shared parameters to benefit from better generalization obtained by training on large-scale image datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.469, + 0.499 + ], + "angle": 0, + "content": "To achieve this, we extend an image architecture by introducing temporal layers, which are only active for video inputs. All other layers are shared between the image and video model. The autoencoder remains fixed and processes each frame in a video independently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.499, + 0.469, + 0.651 + ], + "angle": 0, + "content": "The UNet consists of two main building blocks: Residual blocks and transformer blocks (see Fig. 3). Similar to [17, 49], we extend them to videos by adding both 1D convolutions across time and 1D self-attention across time. In each residual block, we introduce one temporal convolution after each 2D convolution. Similarly, after each spatial 2D transformer block, we also include one temporal 1D transformer block, which mimics its spatial counterpart along the time axis. We also input learnable positional encodings of the frame index into temporal transformer blocks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.658, + 0.394, + 0.674 + ], + "angle": 0, + "content": "3.3. Representing Content and Structure" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.469, + 0.756 + ], + "angle": 0, + "content": "Conditional Diffusion Models Diffusion models are well-suited to modeling conditional distributions such as \\( p(x|s, c) \\). The forward process \\( q \\) remains unchanged while the conditioning variables \\( s, c \\) become additional inputs to the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.469, + 0.876 + ], + "angle": 0, + "content": "Our goal is to edit a video based on a text prompt describing the desired output. We choose to train on uncaptioned video data due to the lack of large-scale paired video-text datasets of similar quality as image datasets like [48]. Therefore, during training, we must derive structure and content representations from the training video \\( x \\) itself, i.e. \\( s = s(x) \\) and \\( c = c(x) \\), resulting in a per-example loss of" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.884, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {t} \\left\\| \\mu_ {t} \\left(\\mathcal {E} (x) _ {t}, \\mathcal {E} (x) _ {0}\\right) - \\mu_ {\\theta} \\left(\\mathcal {E} (x) _ {t}, t, s (x), c (x)\\right) \\right\\| ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.153 + ], + "angle": 0, + "content": "In contrast, during inference, structure \\( s \\) and content \\( c \\) are derived from an input video \\( y \\) and from a text prompt \\( t \\) respectively. An edited version \\( x \\) of \\( y \\) is obtained by sampling the generative model conditioned on \\( s(y) \\) and \\( c(t) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.164, + 0.892, + 0.18 + ], + "angle": 0, + "content": "\\[\nz \\sim p _ {\\theta} (z | s (y), c (t)), x = \\mathcal {D} (z). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.191, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Content Representation We utilize CLIP [32] to infer a content representation from both text inputs \\( t \\) and video inputs \\( x \\) similar to previous works [35, 3]. CLIP embeddings are a promising content representation as they are more sensitive to semantic and stylistic properties while being more invariant towards geometric attributes [34]. During training, we encode a random frame in each input video with CLIP. To support text-based editing at inference, we train a prior model that allows sampling image embeddings from text embeddings [35, 49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Structure Representation We need a representation that provides adequate separation between structure and content. We find that depth estimates extracted from input video frames provide the desired properties as they encode significantly less content information compared to simpler structure representations, such as edge filters which also encode textural properties. Still, depth maps reveal the silhouettes of objects which can prevent content edits involving changes in object shape." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.584 + ], + "angle": 0, + "content": "To offer control over the amount of structure to preserve, we propose to train a model on structure representations with varying amounts of information. In particular, we blur depth estimates given a parameter \\( t_s \\). During training, \\( t_s \\) is randomly sampled between 0 and \\( T_s \\). The parameter can then be controlled at inference to achieve different editing effects (see Fig. 10)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.674 + ], + "angle": 0, + "content": "While depths map work well for our usecase, our approach generalizes to other geometric guidance features or combinations of features that might be more helpful for other specific applications. For example, models focusing on human video synthesis might benefit from estimated poses or face landmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Conditioning Mechanisms We account for the different characteristics of our content and structure representations with two different conditioning mechanisms. Since structure represents a significant portion of the spatial information of video frames, we use concatenation for conditioning to make effective use of this information. In contrast, attributes described by the content representation are not tied to particular locations. Hence, we leverage cross-attention which can effectively transport this information to any position." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We use the spatial transformer blocks of the UNet architecture for cross-attention conditioning. Each contains two attention operations, where the first one performs a spatial self-attention and the second one a cross attention with keys and values computed from the CLIP image embedding." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7349" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.406, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.215, + 0.894, + 0.285 + ], + "angle": 0, + "content": "Figure 4. Temporal Control: By training image and video models jointly, we obtain explicit control over the temporal consistency of edited videos via a temporal guidance scale \\(\\omega_{t}\\). On the left, frame consistency measured via CLIP cosine similarity of consecutive frames increases monotonically with \\(\\omega_{t}\\), while mean squared error between frames warped with optical flow decreases monotonically. On the right, lower scales (0.5 in the middle row) achieve edits with a \"hand-drawn\" look, whereas higher scales (1.5 in the bottom row) result in smoother results. Top row shows the original input video, the two edits use the prompt \"pencil sketch of a man looking at the camera\"." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.3, + 0.177, + 0.315 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "title", + "bbox": [ + 0.402, + 0.3, + 0.669, + 0.314 + ], + "angle": 0, + "content": "Driving Video (top) and Result (bottom)" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.352, + 0.192, + 0.411 + ], + "angle": 0, + "content": "a man using a laptop inside a train, anime style" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.456, + 0.192, + 0.56 + ], + "angle": 0, + "content": "a woman and man take selfies while walking down the street, claymation" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.615, + 0.191, + 0.658 + ], + "angle": 0, + "content": "kite-surfer in the ocean at sunset" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.728, + 0.192, + 0.801 + ], + "angle": 0, + "content": "alien expplorer hiking in the mountains" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.32, + 0.861, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.447, + 0.861, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.576, + 0.861, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.702, + 0.859, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.827, + 0.893, + 0.854 + ], + "angle": 0, + "content": "Figure 5. Our approach enables a wide range of video edits, including changes to animation styles such as anime or claymation, changes of environment such as time of day, and changing characters such as humans to aliens." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7350" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.473, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.473, + 0.324 + ], + "angle": 0, + "content": "Figure 6. Prompt-vs-frame consistency: Image models such as SD-Depth achieve good prompt consistency but fail to produce consistent edits across frames. Propagation based approaches such as IVS and Text2Live increase frame consistency but fail to provide edits reflecting the prompt accurately. Our method achieves the best combination of frame and prompt consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.47, + 0.481 + ], + "angle": 0, + "content": "To condition on structure, we first estimate depth maps for all input frames using the MiDaS DPT-Large model [36]. We then apply \\( t_s \\) iterations of blurring and downsampling to the depth maps, where \\( t_s \\) controls the amount of structure to preserve. We resample the perturbed depth map to the resolution of the RGB-frames and encode it using \\( \\mathcal{E} \\). This latent representation of structure is concatenated with the input \\( z_t \\) given to the UNet. We also input four channels containing a sinusoidal embedding of \\( t_s \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.471, + 0.634 + ], + "angle": 0, + "content": "Sampling While Eq. (2) provides a direct way to sample from the trained model, many other sampling methods [52, 24, 22] require only a fraction of the number of diffusion timesteps to achieve good sample quality. We use DDIM [52] throughout our experiments. Furthermore, classifier-free diffusion guidance [16] significantly improves sample quality. For a conditional model \\(\\mu_{\\theta}(x_t,t,c)\\), this is achieved by training the model to also perform unconditional predictions \\(\\mu_{\\theta}(x_t,t,\\emptyset)\\) and then adjusting predictions during sampling according to" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.645, + 0.466, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mu} _ {\\theta} (x _ {t}, t, c) = \\mu_ {\\theta} (x _ {t}, t, \\emptyset) + \\omega (\\mu_ {\\theta} (x _ {t}, t, c) - \\mu_ {\\theta} (x _ {t}, t, \\emptyset))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.839 + ], + "angle": 0, + "content": "where \\(\\omega\\) is the guidance scale that controls the strength. Based on the intuition that \\(\\omega\\) extrapolates the direction between an unconditional and a conditional model, we apply this idea to control temporal consistency of our model. Specifically, since we are training both an image and a video model with shared parameters, we can consider predictions by both models for the same input. Let \\(\\mu_{\\theta}(z_t,t,c,s)\\) denote the prediction of our video model, and let \\(\\mu_{\\theta}^{\\pi}(z_t,t,c,s)\\) denote the prediction of the image model applied to each frame individually. Taking classifier-free guidance for \\(c\\) into account, we then adjust our prediction according to" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.849, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\mu} _ {\\theta} (z _ {t}, t, c, s) = \\mu_ {\\theta} ^ {\\pi} (z _ {t}, t, \\emptyset , s) \\\\ + \\omega_ {t} \\left(\\mu_ {\\theta} \\left(x _ {t}, t, \\emptyset , s\\right) - \\mu_ {\\theta} ^ {\\pi} \\left(x _ {t}, t, \\emptyset , s\\right)\\right) \\tag {8} \\\\ + \\omega (\\mu_ {\\theta} (x _ {t}, t, c, s) - \\mu_ {\\theta} (x _ {t}, t, \\emptyset , s)) \\\\ \\end{array}\n\\]" + }, + { + "type": "image_footnote", + "bbox": [ + 0.526, + 0.091, + 0.86, + 0.106 + ], + "angle": 0, + "content": "ours SDEdit (50% strength) IVS (50% strength) Depth-SDours (-, 50% SDEdit) Deformu SDEdit (75% strength) IVS (75% strength) Text2LIVEours (-, 75% SDEdit)" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.114, + 0.892, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.248, + 0.892, + 0.276 + ], + "angle": 0, + "content": "Figure 7. User Preferences: Based on our user study, the results from our model are preferred over the baseline models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.301, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Our experiments demonstrate that this approach controls temporal consistency in the outputs, see Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.342, + 0.64, + 0.358 + ], + "angle": 0, + "content": "3.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.47 + ], + "angle": 0, + "content": "We train on an internal dataset of 240M images and 6.4M video clips. We use image batches of size 9216 with resolutions of \\(320 \\times 320\\), \\(384 \\times 320\\) and \\(448 \\times 256\\), as well as the same resolutions with flipped aspect ratios. We sample image batches with a probability of \\(12.5\\%\\). For videos, we use batch size 1152 and 8 frames from each video sampled four frames apart with a resolution of \\(448 \\times 256\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.893, + 0.623 + ], + "angle": 0, + "content": "We train our model in multiple stages. First, we initialize model weights based on a pretrained text-conditional latent diffusion model \\([38]^1\\). We change the conditioning from CLIP text embeddings to CLIP image embeddings and fine-tune for 15k steps on images only. Afterwards, we introduce temporal connections as described in Sec. 3.2 and train jointly on images and videos for 75k steps. We then add conditioning on structure \\(s\\) with \\(t_s \\equiv 0\\) fixed and train for 25k steps. Finally, we resume training with \\(t_s\\) sampled uniformly between 0 and 7 for another 10k steps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.587, + 0.652 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.662, + 0.892, + 0.738 + ], + "angle": 0, + "content": "To evaluate our approach, we use videos from DAVIS [31] and various stock footage. To automatically create edit prompts, we first run a captioning model [23] to obtain a description of the original video content. We then use GPT-3 [7] to generate edited prompts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.748, + 0.683, + 0.764 + ], + "angle": 0, + "content": "4.1. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.771, + 0.892, + 0.879 + ], + "angle": 0, + "content": "We demonstrate that our approach performs well on a number of diverse inputs (see Fig. 5). Our method handles a large variety of footage, such as landscapes and close-ups, and diverse camera motion without any explicit tracking of the input. Our depth-based structure representation combined with large-scale image-video joint training enable strong generalization and powerful editing capabilities." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.874, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/runwayml/stable-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.514, + 0.937 + ], + "angle": 0, + "content": "7351" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.113, + 0.141, + 0.125 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.172, + 0.141, + 0.184 + ], + "angle": 0, + "content": "Mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.213, + 0.193, + 0.261 + ], + "angle": 0, + "content": "A snowboarder in a snow park on the mountain" + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.089, + 0.86, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.271, + 0.872, + 0.285 + ], + "angle": 0, + "content": "Figure 8. Background Editing: Masking the denoising process allows us to restrict edits to backgrounds for more control over results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.468, + 0.37 + ], + "angle": 0, + "content": "ties. For example, we can produce various animation styles, changes in time of day, and more complex changes of subject, such as turning a hiker into an alien (see Fig. 5). Please see the supplementary material for more results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.468, + 0.491 + ], + "angle": 0, + "content": "Using CLIP image embeddings as the content representation allows users to specify content through images. As an example application, we demonstrate character replacement in Fig. 9. For every video in a set of six videos, we resynthesize it five times, each time providing a single content image taken from another video in the set. We can retain content characteristics with \\( t_s = 3 \\) despite large differences in their pose and shape." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.468, + 0.567 + ], + "angle": 0, + "content": "Lastly, we illustrate the use of masked video editing in Fig. 8, where the model predicts everything outside the masked area(s) while retaining the original content inside the masked area. Notably, this technique resembles inpainting with diffusion models [43, 25]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.2, + 0.592 + ], + "angle": 0, + "content": "4.2. User Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.468, + 0.764 + ], + "angle": 0, + "content": "We benchmark against Text2Live [5], a recent approach for text-guided video editing that employs layered neural atlases [21]. As a baseline, we compare against SDEdit [26] in two ways; per-frame generated results and a first-frame result propagated by a few-shot video stylization method [55] (IVS). We also include two depth-based versions of Stable Diffusion; one trained with depth-conditioning [2] and one that retains past results based on depth estimates [9]. We also include an ablation: applying SDEdit to our video model trained without conditioning on a structure representation (ours, \\(\\sim s\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.885 + ], + "angle": 0, + "content": "We judge the success of our method qualitatively based on a user study. We run the user study using Amazon Mechanical Turk (AMT) on an evaluation set of 35 representative video editing prompts. For each example, we ask 5 annotators to compare faithfulness to the video editing prompt (\"Which video better represents the provided edited caption?\") between a baseline and our method, presented in random order, and use a majority vote for the final result." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The results can be found in Fig. 7. Across all compared" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.612 + ], + "angle": 0, + "content": "methods, results from our approach are preferred roughly 3 out of 4 times. A visual comparison among the methods can be found in the supplementary. We observe that SDEdit is sensitive to the editing strength. Low values fail to achieve the desired editing effect whereas high values change the structure of the input. Even with a fixed seed, both style and structure can change in unnatural ways between frames as their relationship is ignored by image-based approaches. Propagation of SDEdit outputs (IVS) leads to more consistent results but often introduces propagation artifacts especially with large motion. Depth-SD produces accurate, structure-preserving edits for individual frames but frames are inconsistent across time. The outputs of Text2Live tend to be temporally smooth due to its reliance on Layered Neural Atlases [21], but it often produces edits that represent the edit prompt inaccurately. A direct comparison with Text2Live is difficult as it requires input masks and separate edit prompts for foreground and background. In addition, computing a neural atlas takes about 10 hours whereas our approach requires approximately a minute." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.722, + 0.637 + ], + "angle": 0, + "content": "4.3.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.674 + ], + "angle": 0, + "content": "We quantify trade-offs between frame consistency and prompt consistency with the following two metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Frame consistency We compute CLIP image embeddings on all frames of output videos and report the average cosine similarity between all pairs of consecutive frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Prompt consistency We compute CLIP image embeddings on all frames of output videos and the CLIP text embedding of the edit prompt. We report average cosine similarity between text and image embedding over all frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 6 shows the results of each model using our frame consistency and prompt consistency metrics. Our model tends to outperform the baseline models in both aspects (placed higher in the upper-right quadrant of the graph). We also notice a slight tradeoff with increasing the strength parameters in the baseline models: larger strength scales implies higher prompt consistency at the cost of lower frame consistency. Increasing the temporal scale \\((\\omega_{t})\\) of our model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7352" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.47, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.555, + 0.47, + 0.584 + ], + "angle": 0, + "content": "Figure 9. Image Prompting: We combine the structure of a driving video (first column) with content from other videos (first row)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.47, + 0.661 + ], + "angle": 0, + "content": "results in higher frame consistency but lower prompt consistency. We also observe that an increased structure scale \\((t_{s})\\) results in higher prompt consistency as the content becomes less determined by the input structure." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.677, + 0.226, + 0.691 + ], + "angle": 0, + "content": "4.4.Customization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Customization of image models enables generation of previously unseen content, such as specific people or styles, based on a small dataset used for finetuning [41]. We fine-tune our depth-conditioned latent video diffusion model on a set of 15-30 images and produce videos containing the desired subject. Half of the batch elements are of the subject and the other half belong to the original training dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Fig. 10 shows an example with different numbers of customization steps as well as different levels of structure preservation \\( t_{s} \\). Customization improves fidelity to the style and appearance of the character. In combination with higher \\( t_{s} \\) values, accurate animations are possible despite using a driving video of a person with different characteristics." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.072, + 0.892, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.501, + 0.892, + 0.586 + ], + "angle": 0, + "content": "Figure 10. Controlling Fidelity: We obtain control over structure and appearance-fidelity. Each cell shows three frames produced with decreasing structure-fidelity \\( t_s \\) (left-to-right) and increasing number of customization training steps (top-to-bottom). The bottom shows examples of images used for customization (red border) and the input image (blue border). Same driving video as in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.619, + 0.617 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our latent video diffusion model synthesizes new videos given structure and content information. We ensure structural consistency by conditioning on depth estimates while content is controlled with images or natural language. Temporal layers and joint image-video training achieve stable results across frames. A novel guidance method, inspired by classifier-free guidance, allows for control over temporal consistency. By training on depth maps with varying degrees of detail, we can adjust the level of structure preservation. This, together with model customization, improves content fidelity. Our quantitative evaluation and user study show that our method is preferred over related approaches. Future works should investigate other conditioning data, such as facial landmarks and pose estimates, and additional 3d-priors to improve generated results. Our model is intended for creative applications in content creation, but we realize the risks of dual-use and hope that further work will be aimed at combating abuse of generative models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7353" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards high resolution video generation with progressive growing of sliced Wasserstein gans. arXiv preprint arXiv:1810.02419, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.174, + 0.468, + 0.214 + ], + "angle": 0, + "content": "[2] Stability AI. Stable diffusion depth. https://github.com/Stability-AI/stablediffusion, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.468, + 0.299 + ], + "angle": 0, + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.468, + 0.358 + ], + "angle": 0, + "content": "[4] Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie S. Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.361, + 0.468, + 0.416 + ], + "angle": 0, + "content": "[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European Conference on Computer Vision, pages 707-723. Springer, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.419, + 0.468, + 0.473 + ], + "angle": 0, + "content": "[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A Efros, and Tero Karras. Generating long videos of dynamic scenes. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.477, + 0.468, + 0.655 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc., 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.468, + 0.714 + ], + "angle": 0, + "content": "[8] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1105-1114, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.716, + 0.468, + 0.755 + ], + "angle": 0, + "content": "[9] deforum. Deforum stable diffusion. https://github.com/deforum/stable-diffusion, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.468, + 0.869 + ], + "angle": 0, + "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 19822-19835. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[11] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Long video generation with time-agnostic vqgan and time-sensitive transformer. arXiv preprint arXiv:2204.03638, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.357, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47-1, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.606 + ], + "angle": 0, + "content": "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Proc. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[21] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[22] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7354" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[26] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. CoRR, abs/2108.01073, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[28] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[29] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[31] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1), jun 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[35] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8821-8831. PMLR, 18-24 Jul 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[36] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:1623-1637, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew William Fitzgibbon. Unwrap mosaics: a new repre" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "sentation for video editing. ACM SIGGRAPH 2008 papers, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[40] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In Bodo Rosenhahn and Bjoern Andres, editors, Pattern Recognition, pages 26-36, Cham, 2016. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[42] Alexander S. Disco diffusion v5.2 - warp fusion. https://github.com/Sxela/DiscoDiffusion-Warp, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[43] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[45] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Proceedings of the IEEE international conference on computer vision, pages 2830-2839, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[46] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.61, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[47] Robin San-Roman, Eliya Nachmani, and Lior Wolf. Noise estimation for generative diffusion models. arXiv preprint arXiv:2104.02600, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[48] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[49] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.832 + ], + "angle": 0, + "content": "[50] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.515, + 0.936 + ], + "angle": 0, + "content": "7355" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "of Machine Learning Research, pages 2256-2265, Lille, France, 07-09 Jul 2015. PMLR. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.163 + ], + "angle": 0, + "content": "[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[53] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[54] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[55] Ondrej Texler, David Futschik, Michal Kucera, Ondrej Jamriška, Šárka Sochorová, Mengei Chai, Sergey Tulyakov, and Daniel Sýkora. Interactive video stylization using few-shot patch-based training. ACM Transactions on Graphics, 39(4):73, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[56] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[57] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. Advances in Neural Information Processing Systems, 34:11287-11302, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[58] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[59] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[60] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[62] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[63] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[64] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[65] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[66] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[67] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "7356" + } + ] +] \ No newline at end of file diff --git a/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/4fea04cc-819a-4990-99d9-0ccc1f6c300d_origin.pdf b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/4fea04cc-819a-4990-99d9-0ccc1f6c300d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..854b656d76ae6fb323085dbcecfb370144b32750 --- /dev/null +++ b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/4fea04cc-819a-4990-99d9-0ccc1f6c300d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ff6097c8a2fa11512393f5e9c70847b8957cd63f29996462752d8d8fdf3af98 +size 10288548 diff --git a/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/full.md b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6ce271d2ebdc352353ad59a975a5db83c81fd972 --- /dev/null +++ b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/full.md @@ -0,0 +1,334 @@ +# Structure and Content-Guided Video Synthesis with Diffusion Models + +Patrick Esser Johnathan Chiu Parmida Atighehchian Jonathan Granskog Anastasis Germanidis Runway https://research.runwayml.com/gen1 + +![](images/6f1f4460496ae43009365495e3372b877ee4e114a30ba3847e5b92b434710795.jpg) +Figure 1. Guided Video Synthesis We present an approach based on latent video diffusion models that synthesizes videos (top and bottom) guided by content described through text (top) or images (bottom) while keeping the structure of an input video (middle). + +# Abstract + +Text-guided generative diffusion models unlock powerful image creation and editing tools. Recent approaches that edit the content of footage while retaining structure require expensive re-training for every input or rely on error-prone propagation of image edits across frames. + +In this work, we present a structure and content-guided video diffusion model that edits videos based on descriptions of the desired output. Conflicts between user-provided content edits and structure representations occur due to insufficient disentanglement between the two aspects. As a solution, we show that training on monocular depth estimates with varying levels of detail provides control over structure and content fidelity. A novel guidance method, enabled by joint video and image training, exposes explicit control over temporal consistency. Our experiments demonstrate a wide variety of successes; fine-grained control over output characteristics, customization based on a few reference images, and a strong user preference towards results by our model. + +# 1. Introduction + +Demand for more intuitive and performant video editing tools has increased as video-centric platforms have been popularized. But editing in the format is still complex and time-consuming due the temporal nature of video data. State-of-the-art machine learning models have shown great promise in improving editing workflows. + +Generative approaches for image synthesis recently experienced a rapid surge in quality and popularity due to the introduction of powerful diffusion models trained on largescale datasets. Text-conditioned models, such as DALLE 2 [34] and Stable Diffusion [38], enable novice users to generate detailed imagery given only a text prompt as input. Latent diffusion models especially enable efficient methods for producing imagery via synthesis in a perceptually compressed space. + +Motivated by this progress, we investigate generative models suited for interactive applications in video editing. + +Current methods repurpose existing image models by either propagating edits with approaches that compute explicit correspondences [5] or by finetuning on each individual video [63]. We aim to circumvent expensive per-video training and correspondence calculation to achieve fast inference for arbitrary videos. + +We propose a controllable structure and content-aware latent video diffusion model trained on a large-scale dataset of uncaptioned videos and images. We opt to represent structure with monocular depth estimates, and content with embeddings predicted by a pre-trained neural network. Our approach offers several powerful modes of control. First, we train our model such that the content of inferred videos, e.g. their appearance or style, match user-provided images or text prompts (Fig. 1). Second, we vary the fidelity of the structure representation during training to allow selecting the strength of the structure preservation at test-time. Finally, we also adjust the inference process via a custom guidance method, inspired by classifier-free guidance, to enable control over temporal consistency. + +In summary, we present the following contributions: + +- We extend latent diffusion models to video generation by introducing temporal layers into a pre-trained image model and by joint training on images and videos. +- We present a structure and content-aware model that edits videos given example images or text. Our method does not require per-video training or pre-processing. +- We demonstrate full control over temporal, content and structure consistency. We show for the first time that joint image-video training enables control over temporal stability. And, training on varying levels of detail in the structure representation allows choosing the desired level of preservation during inference. +- We show that our approach is preferred over several other approaches in a user study. We further improve the accuracy of previously unseen content by finetuning on a small set of images of the desired subject. + +# 2. Related Work + +Controllable video editing and media synthesis is an active area of research. In this section, we review prior work in related areas and connect our method to these approaches. + +Unconditional video generation Generative adversarial networks (GANs) [12] can learn to synthesize videos based on specific training data [59, 45, 1, 56]. These methods often struggle with stability during optimization, and produce fixed-length videos [59, 45] or longer videos where artifacts accumulate over time [50]. [6] synthesize longer videos using a GAN with a better encoding of the time axis. Autoregressive transformers have also been proposed for uncon- + +ditional video generation [11, 64]. Our focus is on providing user control over the synthesis process whereas these approaches are limited to sampling random content resembling their training distribution. + +Diffusion models for image synthesis Diffusion models (DMs) [51, 53] can synthesize detailed media in many formats, such as images [34, 38], 3d shapes [66] and animations [54]. Many works improve diffusion-based image synthesis by changing the parameterization [14, 27, 46], introducing advanced sampling methods [52, 24, 22, 47, 20], designing more powerful architectures [3, 15, 57, 30], or conditioning on additional information [25]. Text-conditioning, based on embeddings from CLIP [32] or T5 [33], has become a particularly powerful approach for providing artistic control over model output [44, 28, 34, 3, 65, 10]. Latent diffusion models (LDMs) [38] perform diffusion in a compressed latent space reducing memory requirements and runtime. Our video model is an LDM trained simultaneously on videos and images. + +Diffusion models for video synthesis Recently, diffusion models, masked generative models and autoregressive models have been applied to text-conditioned video synthesis [17, 13, 58, 67, 18, 49]. Similar to [17] and [49], we extend image synthesis diffusion models to video generation by introducing temporal connections into a pre-existing image model. Our model edits videos rather than synthesizing them from scratch. We demonstrate through a user study that our model with explicit conditioning over structure is preferred over other related approaches. + +Video translation and propagation Image-to-image translation models, such as pix2pix [19, 62], can modify each frame in a video individually. This produces temporal inconsistencies as the time axis is ignored. Accounting for temporal or geometric information, such as flow, can increase consistency across frames when repurposing image synthesis models [42, 9]. We can extract such structural information to aid our spatio-temporal LDM in text- and image-guided video synthesis. Many generative adversarial methods, such as vid2vid [61, 60], leverage this type of input to guide synthesis. + +Video style transfer takes a reference style image and statistically applies its style to an input video [40, 8, 55]. In contrast, our method edits both style and content while preserving the structure of a video instead of matching feature statistics only. Text2Live [5] allows editing input videos using text prompts by decomposing a video into neural layers [21]. Once available, a layered video representation [37] provides consistent propagation across frames. SinFusion [29] and Tune-a-Video [63] use diffusion models to edit videos but require per-video training. This limits the practicality of the approaches in creative tools. We opt to instead train our model on a large-scale dataset permitting inference on any video without individual training. + +![](images/49a6b142c1d07da6f025c6c46f1ccbb1a3ecf0c53b66bbd7274985a943e959d0.jpg) +Figure 2. Overview: During training (left), input videos $x$ are encoded to $z_{0}$ with a fixed encoder $\mathcal{E}$ and diffused to $z_{t}$ . We extract a structure representation $s$ by encoding depth maps obtained with MiDaS, and a content representation $c$ by encoding one of the frames with CLIP. The model then learns to reverse the diffusion process in the latent space, with the help of $s$ , which gets concatenated to $z_{t}$ , as well as $c$ , which is provided via cross-attention blocks. During inference (right), the structure $s$ of an input video is provided in the same manner. To specify content via text, we convert CLIP text embeddings to image embeddings via a prior. + +![](images/07748d2211bea29d7fb16cd61303243e0afac993e41a17118137d1da77979aac.jpg) + +# 3. Method + +For our purposes, it will be helpful to think of a video in terms of its content and structure. By structure, we refer to characteristics describing geometry and dynamics, e.g. shapes and locations of subjects as well as their temporal changes. We define content as features describing the appearance and semantics of the video, such as the colors and styles of objects and the lighting. The goal of our model is to edit the content of a video while retaining its structure. + +To achieve this, we learn a generative model $p(x|s,c)$ of videos $x$ conditioned on representations of structure $s$ and content $c$ . We infer the shape representation $s$ from an input video, and modify it based on a text prompt $c$ describing the edit. First, we describe our realization of the generative model as a conditional latent video diffusion model and, then, we describe our choices for shape and content representations. Finally, we discuss the optimization process of our model. See Fig. 2 for an overview. + +# 3.1. Latent diffusion models + +Diffusion models Diffusion models [51] learn to reverse a fixed forward diffusion process, which is defined as + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) := \mathcal {N} \left(x _ {t}, \sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} \mathcal {I}\right). \tag {1} +$$ + +Normally-distributed noise is slowly added to each sample $x_{t-1}$ to obtain $x_t$ . The forward process models a fixed Markov chain and the noise is dependent on a variance schedule $\beta_t$ where $t \in \{1, \dots, T\}$ , with $T$ being the total number of steps in our diffusion chain, and $x_0 \coloneqq x$ . + +Learning to Denoise The reverse process is defined accord + +ing to the following equation with parameters $\theta$ + +$$ +p _ {\theta} \left(x _ {0}\right) := \int p _ {\theta} \left(x _ {0: T}\right) d x _ {1: T} \tag {2} +$$ + +$$ +p _ {\theta} \left(x _ {0: T}\right) = p \left(x _ {T}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right), \tag {3} +$$ + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) := \mathcal {N} \left(x _ {t - 1}, \mu_ {\theta} \left(x _ {t}, t\right), \Sigma_ {\theta} \left(x _ {t}, t\right)\right). \tag {4} +$$ + +Using a fixed variance $\Sigma_{\theta}(x_t,t)$ , we are left learning the means of the reverse process $\mu_{\theta}(x_t,t)$ . Training is typically performed via a reweighted variational bound on the maximum likelihood objective, resulting in a loss + +$$ +L := \mathbb {E} _ {t, q} \lambda_ {t} \| \mu_ {t} \left(x _ {t}, x _ {0}\right) - \mu_ {\theta} \left(x _ {t}, t\right) \| ^ {2}, \tag {5} +$$ + +where $\mu_t(x_t, x_0)$ is the mean of the forward process posterior $q(x_{t-1} | x_t, x_0)$ , which is available in closed form [14]. + +Parameterization The mean $\mu_{\theta}(x_t,t)$ is then predicted by a UNet architecture [39] that receives the noisy input $x_{t}$ and the diffusion timestep $t$ as inputs. Other parameterizations and weightings, such as $\epsilon$ [14] and $v$ -parameterizations [46], can significantly improve sample quality compared to directly predicting the mean. Similar to [13], we found that $v$ -parameterization improves color consistency thus all our experiments use it (see supp. material for more details). + +Latent diffusion Latent diffusion models [38] (LDMs) take the diffusion process into the latent space. This provides improved separation between compressive and generative learning phases of the model. Specifically, LDMs use an autoencoder where an encoder $\mathcal{E}$ maps input data $x$ to a lower dimensional latent code according to $z = \mathcal{E}(x)$ while a decoder $\mathcal{D}$ converts latent codes back to the input space such that perceptually $x \approx \mathcal{D}(\mathcal{E}(x))$ . + +Our encoder downsamples RGB-images $x\in \mathbb{R}^{3\times H\times W}$ by a factor of eight and outputs four channels, resulting in a latent code $z\in \mathbb{R}^{4\times H / 8\times W / 8}$ . Thus, the diffusion UNet operates on a much smaller representation which significantly + +![](images/1ce833b0bd6217de8e60b9a56f025de8b65431fc08cc7c2b25aec479669c2868.jpg) +Figure 3. Temporal Extension: We extend an image-based UNet architecture to videos, by adding temporal layers in its building blocks. We add a 1D temporal convolution after each 2D spatial convolution in its residual blocks (left), and we add a 1D temporal attention block after each of its 2D spatial attention blocks (right). + +![](images/6a1d8c064045f53420d1d4f077e731bf29ffffca4f42880d83270ffcfc2690ea.jpg) + +improves runtime and memory efficiency. The latter is crucial for video modeling where the additional time axis increases memory costs. + +# 3.2. Spatio-temporal Latent Diffusion + +To correctly model a distribution over video frames, the architecture must account for temporal relationships. We also want to jointly learn an image model with shared parameters to benefit from better generalization obtained by training on large-scale image datasets. + +To achieve this, we extend an image architecture by introducing temporal layers, which are only active for video inputs. All other layers are shared between the image and video model. The autoencoder remains fixed and processes each frame in a video independently. + +The UNet consists of two main building blocks: Residual blocks and transformer blocks (see Fig. 3). Similar to [17, 49], we extend them to videos by adding both 1D convolutions across time and 1D self-attention across time. In each residual block, we introduce one temporal convolution after each 2D convolution. Similarly, after each spatial 2D transformer block, we also include one temporal 1D transformer block, which mimics its spatial counterpart along the time axis. We also input learnable positional encodings of the frame index into temporal transformer blocks. + +# 3.3. Representing Content and Structure + +Conditional Diffusion Models Diffusion models are well-suited to modeling conditional distributions such as $p(x|s, c)$ . The forward process $q$ remains unchanged while the conditioning variables $s, c$ become additional inputs to the model. + +Our goal is to edit a video based on a text prompt describing the desired output. We choose to train on uncaptioned video data due to the lack of large-scale paired video-text datasets of similar quality as image datasets like [48]. Therefore, during training, we must derive structure and content representations from the training video $x$ itself, i.e. $s = s(x)$ and $c = c(x)$ , resulting in a per-example loss of + +$$ +\lambda_ {t} \left\| \mu_ {t} \left(\mathcal {E} (x) _ {t}, \mathcal {E} (x) _ {0}\right) - \mu_ {\theta} \left(\mathcal {E} (x) _ {t}, t, s (x), c (x)\right) \right\| ^ {2}. \tag {6} +$$ + +In contrast, during inference, structure $s$ and content $c$ are derived from an input video $y$ and from a text prompt $t$ respectively. An edited version $x$ of $y$ is obtained by sampling the generative model conditioned on $s(y)$ and $c(t)$ : + +$$ +z \sim p _ {\theta} (z | s (y), c (t)), x = \mathcal {D} (z). \tag {7} +$$ + +Content Representation We utilize CLIP [32] to infer a content representation from both text inputs $t$ and video inputs $x$ similar to previous works [35, 3]. CLIP embeddings are a promising content representation as they are more sensitive to semantic and stylistic properties while being more invariant towards geometric attributes [34]. During training, we encode a random frame in each input video with CLIP. To support text-based editing at inference, we train a prior model that allows sampling image embeddings from text embeddings [35, 49]. + +Structure Representation We need a representation that provides adequate separation between structure and content. We find that depth estimates extracted from input video frames provide the desired properties as they encode significantly less content information compared to simpler structure representations, such as edge filters which also encode textural properties. Still, depth maps reveal the silhouettes of objects which can prevent content edits involving changes in object shape. + +To offer control over the amount of structure to preserve, we propose to train a model on structure representations with varying amounts of information. In particular, we blur depth estimates given a parameter $t_s$ . During training, $t_s$ is randomly sampled between 0 and $T_s$ . The parameter can then be controlled at inference to achieve different editing effects (see Fig. 10). + +While depths map work well for our usecase, our approach generalizes to other geometric guidance features or combinations of features that might be more helpful for other specific applications. For example, models focusing on human video synthesis might benefit from estimated poses or face landmarks. + +Conditioning Mechanisms We account for the different characteristics of our content and structure representations with two different conditioning mechanisms. Since structure represents a significant portion of the spatial information of video frames, we use concatenation for conditioning to make effective use of this information. In contrast, attributes described by the content representation are not tied to particular locations. Hence, we leverage cross-attention which can effectively transport this information to any position. + +We use the spatial transformer blocks of the UNet architecture for cross-attention conditioning. Each contains two attention operations, where the first one performs a spatial self-attention and the second one a cross attention with keys and values computed from the CLIP image embedding. + +![](images/d1517ada36b78af12d70bb7b303f9901bde5e96dcacaf865c363dd08c565ad71.jpg) +Figure 4. Temporal Control: By training image and video models jointly, we obtain explicit control over the temporal consistency of edited videos via a temporal guidance scale $\omega_{t}$ . On the left, frame consistency measured via CLIP cosine similarity of consecutive frames increases monotonically with $\omega_{t}$ , while mean squared error between frames warped with optical flow decreases monotonically. On the right, lower scales (0.5 in the middle row) achieve edits with a "hand-drawn" look, whereas higher scales (1.5 in the bottom row) result in smoother results. Top row shows the original input video, the two edits use the prompt "pencil sketch of a man looking at the camera". + +![](images/31ee90a69d58b6e62cc7bd826b5b2059575e76ab0ffdb225891fbd5ff8b13359.jpg) + +# Prompt + +# Driving Video (top) and Result (bottom) + +a man using a laptop inside a train, anime style + +a woman and man take selfies while walking down the street, claymation + +kite-surfer in the ocean at sunset + +alien expplorer hiking in the mountains + +![](images/ce111ac29162328e1eaf09914a94e114951db0b7d4109df2605d4b4293350891.jpg) + +![](images/e42bb00d6644197b0e66ec95016b9adf507f5e4b8b075ee34bd03f8360ebfbd6.jpg) + +![](images/fda5ce90649ab45342be472ae5f4aa0bf1a728a2736ce4f19c10a255d36e4b16.jpg) + +![](images/85be900d6f5fd989d4f2db2b2f03b2f5e450916997cf48be6ef634e217f8efa5.jpg) +Figure 5. Our approach enables a wide range of video edits, including changes to animation styles such as anime or claymation, changes of environment such as time of day, and changing characters such as humans to aliens. + +![](images/2ca212a548bf5cddbe32e0509d4c948c14f09dd9609dbf091c24bfb251e6e1ef.jpg) +Figure 6. Prompt-vs-frame consistency: Image models such as SD-Depth achieve good prompt consistency but fail to produce consistent edits across frames. Propagation based approaches such as IVS and Text2Live increase frame consistency but fail to provide edits reflecting the prompt accurately. Our method achieves the best combination of frame and prompt consistency. + +To condition on structure, we first estimate depth maps for all input frames using the MiDaS DPT-Large model [36]. We then apply $t_s$ iterations of blurring and downsampling to the depth maps, where $t_s$ controls the amount of structure to preserve. We resample the perturbed depth map to the resolution of the RGB-frames and encode it using $\mathcal{E}$ . This latent representation of structure is concatenated with the input $z_t$ given to the UNet. We also input four channels containing a sinusoidal embedding of $t_s$ . + +Sampling While Eq. (2) provides a direct way to sample from the trained model, many other sampling methods [52, 24, 22] require only a fraction of the number of diffusion timesteps to achieve good sample quality. We use DDIM [52] throughout our experiments. Furthermore, classifier-free diffusion guidance [16] significantly improves sample quality. For a conditional model $\mu_{\theta}(x_t,t,c)$ , this is achieved by training the model to also perform unconditional predictions $\mu_{\theta}(x_t,t,\emptyset)$ and then adjusting predictions during sampling according to + +$$ +\tilde {\mu} _ {\theta} (x _ {t}, t, c) = \mu_ {\theta} (x _ {t}, t, \emptyset) + \omega (\mu_ {\theta} (x _ {t}, t, c) - \mu_ {\theta} (x _ {t}, t, \emptyset)) +$$ + +where $\omega$ is the guidance scale that controls the strength. Based on the intuition that $\omega$ extrapolates the direction between an unconditional and a conditional model, we apply this idea to control temporal consistency of our model. Specifically, since we are training both an image and a video model with shared parameters, we can consider predictions by both models for the same input. Let $\mu_{\theta}(z_t,t,c,s)$ denote the prediction of our video model, and let $\mu_{\theta}^{\pi}(z_t,t,c,s)$ denote the prediction of the image model applied to each frame individually. Taking classifier-free guidance for $c$ into account, we then adjust our prediction according to + +$$ +\begin{array}{l} \tilde {\mu} _ {\theta} (z _ {t}, t, c, s) = \mu_ {\theta} ^ {\pi} (z _ {t}, t, \emptyset , s) \\ + \omega_ {t} \left(\mu_ {\theta} \left(x _ {t}, t, \emptyset , s\right) - \mu_ {\theta} ^ {\pi} \left(x _ {t}, t, \emptyset , s\right)\right) \tag {8} \\ + \omega (\mu_ {\theta} (x _ {t}, t, c, s) - \mu_ {\theta} (x _ {t}, t, \emptyset , s)) \\ \end{array} +$$ + +Figure 7. User Preferences: Based on our user study, the results from our model are preferred over the baseline models. +![](images/fc2e31a43f198dfd5c9c37db17a2722366044e9ca9ae6494c9f9f94cbac41a27.jpg) +ours SDEdit (50% strength) IVS (50% strength) Depth-SDours (-, 50% SDEdit) Deformu SDEdit (75% strength) IVS (75% strength) Text2LIVEours (-, 75% SDEdit) + +Our experiments demonstrate that this approach controls temporal consistency in the outputs, see Fig. 4. + +# 3.4. Optimization + +We train on an internal dataset of 240M images and 6.4M video clips. We use image batches of size 9216 with resolutions of $320 \times 320$ , $384 \times 320$ and $448 \times 256$ , as well as the same resolutions with flipped aspect ratios. We sample image batches with a probability of $12.5\%$ . For videos, we use batch size 1152 and 8 frames from each video sampled four frames apart with a resolution of $448 \times 256$ . + +We train our model in multiple stages. First, we initialize model weights based on a pretrained text-conditional latent diffusion model $[38]^1$ . We change the conditioning from CLIP text embeddings to CLIP image embeddings and fine-tune for 15k steps on images only. Afterwards, we introduce temporal connections as described in Sec. 3.2 and train jointly on images and videos for 75k steps. We then add conditioning on structure $s$ with $t_s \equiv 0$ fixed and train for 25k steps. Finally, we resume training with $t_s$ sampled uniformly between 0 and 7 for another 10k steps. + +# 4. Results + +To evaluate our approach, we use videos from DAVIS [31] and various stock footage. To automatically create edit prompts, we first run a captioning model [23] to obtain a description of the original video content. We then use GPT-3 [7] to generate edited prompts. + +# 4.1. Qualitative Results + +We demonstrate that our approach performs well on a number of diverse inputs (see Fig. 5). Our method handles a large variety of footage, such as landscapes and close-ups, and diverse camera motion without any explicit tracking of the input. Our depth-based structure representation combined with large-scale image-video joint training enable strong generalization and powerful editing capabilities. + +![](images/896f9d2efe6f6ac540f6af75a3640a0bab9f6c20ebcec5d086b90e2566d5a444.jpg) +Input +Mask +A snowboarder in a snow park on the mountain +Figure 8. Background Editing: Masking the denoising process allows us to restrict edits to backgrounds for more control over results. + +ties. For example, we can produce various animation styles, changes in time of day, and more complex changes of subject, such as turning a hiker into an alien (see Fig. 5). Please see the supplementary material for more results. + +Using CLIP image embeddings as the content representation allows users to specify content through images. As an example application, we demonstrate character replacement in Fig. 9. For every video in a set of six videos, we resynthesize it five times, each time providing a single content image taken from another video in the set. We can retain content characteristics with $t_s = 3$ despite large differences in their pose and shape. + +Lastly, we illustrate the use of masked video editing in Fig. 8, where the model predicts everything outside the masked area(s) while retaining the original content inside the masked area. Notably, this technique resembles inpainting with diffusion models [43, 25]. + +# 4.2. User Study + +We benchmark against Text2Live [5], a recent approach for text-guided video editing that employs layered neural atlases [21]. As a baseline, we compare against SDEdit [26] in two ways; per-frame generated results and a first-frame result propagated by a few-shot video stylization method [55] (IVS). We also include two depth-based versions of Stable Diffusion; one trained with depth-conditioning [2] and one that retains past results based on depth estimates [9]. We also include an ablation: applying SDEdit to our video model trained without conditioning on a structure representation (ours, $\sim s$ ). + +We judge the success of our method qualitatively based on a user study. We run the user study using Amazon Mechanical Turk (AMT) on an evaluation set of 35 representative video editing prompts. For each example, we ask 5 annotators to compare faithfulness to the video editing prompt ("Which video better represents the provided edited caption?") between a baseline and our method, presented in random order, and use a majority vote for the final result. + +The results can be found in Fig. 7. Across all compared + +methods, results from our approach are preferred roughly 3 out of 4 times. A visual comparison among the methods can be found in the supplementary. We observe that SDEdit is sensitive to the editing strength. Low values fail to achieve the desired editing effect whereas high values change the structure of the input. Even with a fixed seed, both style and structure can change in unnatural ways between frames as their relationship is ignored by image-based approaches. Propagation of SDEdit outputs (IVS) leads to more consistent results but often introduces propagation artifacts especially with large motion. Depth-SD produces accurate, structure-preserving edits for individual frames but frames are inconsistent across time. The outputs of Text2Live tend to be temporally smooth due to its reliance on Layered Neural Atlases [21], but it often produces edits that represent the edit prompt inaccurately. A direct comparison with Text2Live is difficult as it requires input masks and separate edit prompts for foreground and background. In addition, computing a neural atlas takes about 10 hours whereas our approach requires approximately a minute. + +# 4.3.Quantitative Evaluation + +We quantify trade-offs between frame consistency and prompt consistency with the following two metrics. + +Frame consistency We compute CLIP image embeddings on all frames of output videos and report the average cosine similarity between all pairs of consecutive frames. + +Prompt consistency We compute CLIP image embeddings on all frames of output videos and the CLIP text embedding of the edit prompt. We report average cosine similarity between text and image embedding over all frames. + +Fig. 6 shows the results of each model using our frame consistency and prompt consistency metrics. Our model tends to outperform the baseline models in both aspects (placed higher in the upper-right quadrant of the graph). We also notice a slight tradeoff with increasing the strength parameters in the baseline models: larger strength scales implies higher prompt consistency at the cost of lower frame consistency. Increasing the temporal scale $(\omega_{t})$ of our model + +![](images/8f6488259aa3253a32713eb6e4cf4dd0e6787e91732143fe84c6231696b88c09.jpg) +Figure 9. Image Prompting: We combine the structure of a driving video (first column) with content from other videos (first row). + +results in higher frame consistency but lower prompt consistency. We also observe that an increased structure scale $(t_{s})$ results in higher prompt consistency as the content becomes less determined by the input structure. + +# 4.4.Customization + +Customization of image models enables generation of previously unseen content, such as specific people or styles, based on a small dataset used for finetuning [41]. We fine-tune our depth-conditioned latent video diffusion model on a set of 15-30 images and produce videos containing the desired subject. Half of the batch elements are of the subject and the other half belong to the original training dataset. + +Fig. 10 shows an example with different numbers of customization steps as well as different levels of structure preservation $t_{s}$ . Customization improves fidelity to the style and appearance of the character. In combination with higher $t_{s}$ values, accurate animations are possible despite using a driving video of a person with different characteristics. + +![](images/79edf299909923c61304836278cf75d17787db030dc4005eb0536b31bffe848c.jpg) +Figure 10. Controlling Fidelity: We obtain control over structure and appearance-fidelity. Each cell shows three frames produced with decreasing structure-fidelity $t_s$ (left-to-right) and increasing number of customization training steps (top-to-bottom). The bottom shows examples of images used for customization (red border) and the input image (blue border). Same driving video as in Fig. 1. + +# 5. Conclusion + +Our latent video diffusion model synthesizes new videos given structure and content information. We ensure structural consistency by conditioning on depth estimates while content is controlled with images or natural language. Temporal layers and joint image-video training achieve stable results across frames. A novel guidance method, inspired by classifier-free guidance, allows for control over temporal consistency. By training on depth maps with varying degrees of detail, we can adjust the level of structure preservation. This, together with model customization, improves content fidelity. Our quantitative evaluation and user study show that our method is preferred over related approaches. Future works should investigate other conditioning data, such as facial landmarks and pose estimates, and additional 3d-priors to improve generated results. Our model is intended for creative applications in content creation, but we realize the risks of dual-use and hope that further work will be aimed at combating abuse of generative models. + +# References + +[1] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards high resolution video generation with progressive growing of sliced Wasserstein gans. arXiv preprint arXiv:1810.02419, 2018. 2 +[2] Stability AI. Stable diffusion depth. https://github.com/Stability-AI/stablediffusion, 2022. 7 +[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 4 +[4] Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie S. Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise, 2023. +[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European Conference on Computer Vision, pages 707-723. Springer, 2022. 2, 7 +[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A Efros, and Tero Karras. Generating long videos of dynamic scenes. 2022. 2 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc., 2020. 6 +[8] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1105-1114, 2017. 2 +[9] deforum. Deforum stable diffusion. https://github.com/deforum/stable-diffusion, 2022. 2, 7 +[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 19822-19835. Curran Associates, Inc., 2021. 2 +[11] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. + +Long video generation with time-agnostic vqgan and time-sensitive transformer. arXiv preprint arXiv:2204.03638, 2022. 2 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2 +[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3 +[15] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47-1, 2022. 2 +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 6 +[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2, 4 +[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Proc. NeurIPS, 2022. 2 +[21] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 2, 7 +[22] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 2, 6 +[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 6 +[24] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 6 +[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting + +using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 7 +[26] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. CoRR, abs/2108.01073, 2021. 7 +[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2 +[28] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 2 +[29] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2 +[30] William Peebles and Saining Xie. Scalable diffusion models with transformers, 2022. 2 +[31] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 6 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4 +[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1), jun 2022. 2 +[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 1, 2, 4 +[35] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8821-8831. PMLR, 18-24 Jul 2021. 4 +[36] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:1623-1637, 2019. 6 +[37] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew William Fitzgibbon. Unwrap mosaics: a new repre + +sentation for video editing. ACM SIGGRAPH 2008 papers, 2008. 2 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2, 3, 6 +[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3 +[40] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In Bodo Rosenhahn and Bjoern Andres, editors, Pattern Recognition, pages 26-36, Cham, 2016. Springer International Publishing. 2 +[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 8 +[42] Alexander S. Disco diffusion v5.2 - warp fusion. https://github.com/Sxela/DiscoDiffusion-Warp, 2022. 2 +[43] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models, 2021. 7 +[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2 +[45] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Proceedings of the IEEE international conference on computer vision, pages 2830-2839, 2017. 2 +[46] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 2, 3 +[47] Robin San-Roman, Eliya Nachmani, and Lior Wolf. Noise estimation for generative diffusion models. arXiv preprint arXiv:2104.02600, 2021. 2 +[48] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 4 +[49] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 2, 4 +[50] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2, 2021. 2 +[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings + +of Machine Learning Research, pages 2256-2265, Lille, France, 07-09 Jul 2015. PMLR. 2, 3 +[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 6 +[53] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[54] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2 +[55] Ondrej Texler, David Futschik, Michal Kucera, Ondrej Jamriška, Šárka Sochorová, Mengei Chai, Sergey Tulyakov, and Daniel Sýkora. Interactive video stylization using few-shot patch-based training. ACM Transactions on Graphics, 39(4):73, 2020. 2, 7 +[56] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2 +[57] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. Advances in Neural Information Processing Systems, 34:11287-11302, 2021. 2 +[58] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description, 2022. 2 +[59] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2 +[60] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2 +[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 2 +[62] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2 +[63] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2 +[64] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2 +[65] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and + +Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation, 2022. 2 +[66] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[67] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models, 2022. 2 \ No newline at end of file diff --git a/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/images.zip b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..79be66ffdd7f0200067ffee79d93d8529821c08a --- /dev/null +++ b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:821a344e6ce19b96f5140aebd7126ab8f35a190f257fef9916d42365cd63c608 +size 1050872 diff --git a/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/layout.json b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d8801ec534be3797592b6dd3a176698c9d937d74 --- /dev/null +++ b/2023/Structure and Content-Guided Video Synthesis with Diffusion Models/layout.json @@ -0,0 +1,9258 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 81, + 103, + 512, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 103, + 512, + 120 + ], + "spans": [ + { + "bbox": [ + 81, + 103, + 512, + 120 + ], + "type": "text", + "content": "Structure and Content-Guided Video Synthesis with Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 143, + 471, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 143, + 471, + 200 + ], + "spans": [ + { + "bbox": [ + 122, + 143, + 471, + 200 + ], + "type": "text", + "content": "Patrick Esser Johnathan Chiu Parmida Atighehchian Jonathan Granskog Anastasis Germanidis Runway https://research.runwayml.com/gen1" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 238, + 545, + 396 + ], + "blocks": [ + { + "bbox": [ + 56, + 238, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 56, + 238, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 545, + 396 + ], + "type": "image", + "image_path": "6f1f4460496ae43009365495e3372b877ee4e114a30ba3847e5b92b434710795.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 399, + 546, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 399, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 546, + 422 + ], + "type": "text", + "content": "Figure 1. Guided Video Synthesis We present an approach based on latent video diffusion models that synthesizes videos (top and bottom) guided by content described through text (top) or images (bottom) while keeping the structure of an input video (middle)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 450, + 192, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 450, + 192, + 463 + ], + "spans": [ + { + "bbox": [ + 143, + 450, + 192, + 463 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 487, + 288, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 288, + 548 + ], + "type": "text", + "content": "Text-guided generative diffusion models unlock powerful image creation and editing tools. Recent approaches that edit the content of footage while retaining structure require expensive re-training for every input or rely on error-prone propagation of image edits across frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "In this work, we present a structure and content-guided video diffusion model that edits videos based on descriptions of the desired output. Conflicts between user-provided content edits and structure representations occur due to insufficient disentanglement between the two aspects. As a solution, we show that training on monocular depth estimates with varying levels of detail provides control over structure and content fidelity. A novel guidance method, enabled by joint video and image training, exposes explicit control over temporal consistency. Our experiments demonstrate a wide variety of successes; fine-grained control over output characteristics, customization based on a few reference images, and a strong user preference towards results by our model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 488, + 386, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 488, + 386, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 488, + 386, + 500 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 509, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 581 + ], + "type": "text", + "content": "Demand for more intuitive and performant video editing tools has increased as video-centric platforms have been popularized. But editing in the format is still complex and time-consuming due the temporal nature of video data. State-of-the-art machine learning models have shown great promise in improving editing workflows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": "Generative approaches for image synthesis recently experienced a rapid surge in quality and popularity due to the introduction of powerful diffusion models trained on largescale datasets. Text-conditioned models, such as DALLE 2 [34] and Stable Diffusion [38], enable novice users to generate detailed imagery given only a text prompt as input. Latent diffusion models especially enable efficient methods for producing imagery via synthesis in a perceptually compressed space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Motivated by this progress, we investigate generative models suited for interactive applications in video editing." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7346" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "Current methods repurpose existing image models by either propagating edits with approaches that compute explicit correspondences [5] or by finetuning on each individual video [63]. We aim to circumvent expensive per-video training and correspondence calculation to achieve fast inference for arbitrary videos." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 312 + ], + "type": "text", + "content": "We propose a controllable structure and content-aware latent video diffusion model trained on a large-scale dataset of uncaptioned videos and images. We opt to represent structure with monocular depth estimates, and content with embeddings predicted by a pre-trained neural network. Our approach offers several powerful modes of control. First, we train our model such that the content of inferred videos, e.g. their appearance or style, match user-provided images or text prompts (Fig. 1). Second, we vary the fidelity of the structure representation during training to allow selecting the strength of the structure preservation at test-time. Finally, we also adjust the inference process via a custom guidance method, inspired by classifier-free guidance, to enable control over temporal consistency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 312, + 270, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 312, + 270, + 323 + ], + "spans": [ + { + "bbox": [ + 59, + 312, + 270, + 323 + ], + "type": "text", + "content": "In summary, we present the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 331, + 287, + 544 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 331, + 286, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 331, + 286, + 366 + ], + "spans": [ + { + "bbox": [ + 59, + 331, + 286, + 366 + ], + "type": "text", + "content": "- We extend latent diffusion models to video generation by introducing temporal layers into a pre-trained image model and by joint training on images and videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 374, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 374, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 59, + 374, + 287, + 411 + ], + "type": "text", + "content": "- We present a structure and content-aware model that edits videos given example images or text. Our method does not require per-video training or pre-processing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 418, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 418, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 59, + 418, + 287, + 489 + ], + "type": "text", + "content": "- We demonstrate full control over temporal, content and structure consistency. We show for the first time that joint image-video training enables control over temporal stability. And, training on varying levels of detail in the structure representation allows choosing the desired level of preservation during inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 497, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 497, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 59, + 497, + 287, + 544 + ], + "type": "text", + "content": "- We show that our approach is preferred over several other approaches in a user study. We further improve the accuracy of previously unseen content by finetuning on a small set of images of the desired subject." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 562, + 134, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 134, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 134, + 574 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 617 + ], + "type": "text", + "content": "Controllable video editing and media synthesis is an active area of research. In this section, we review prior work in related areas and connect our method to these approaches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "Unconditional video generation Generative adversarial networks (GANs) [12] can learn to synthesize videos based on specific training data [59, 45, 1, 56]. These methods often struggle with stability during optimization, and produce fixed-length videos [59, 45] or longer videos where artifacts accumulate over time [50]. [6] synthesize longer videos using a GAN with a better encoding of the time axis. Autoregressive transformers have also been proposed for uncon-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "ditional video generation [11, 64]. Our focus is on providing user control over the synthesis process whereas these approaches are limited to sampling random content resembling their training distribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 121, + 546, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 301 + ], + "type": "text", + "content": "Diffusion models for image synthesis Diffusion models (DMs) [51, 53] can synthesize detailed media in many formats, such as images [34, 38], 3d shapes [66] and animations [54]. Many works improve diffusion-based image synthesis by changing the parameterization [14, 27, 46], introducing advanced sampling methods [52, 24, 22, 47, 20], designing more powerful architectures [3, 15, 57, 30], or conditioning on additional information [25]. Text-conditioning, based on embeddings from CLIP [32] or T5 [33], has become a particularly powerful approach for providing artistic control over model output [44, 28, 34, 3, 65, 10]. Latent diffusion models (LDMs) [38] perform diffusion in a compressed latent space reducing memory requirements and runtime. Our video model is an LDM trained simultaneously on videos and images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 303, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 422 + ], + "type": "text", + "content": "Diffusion models for video synthesis Recently, diffusion models, masked generative models and autoregressive models have been applied to text-conditioned video synthesis [17, 13, 58, 67, 18, 49]. Similar to [17] and [49], we extend image synthesis diffusion models to video generation by introducing temporal connections into a pre-existing image model. Our model edits videos rather than synthesizing them from scratch. We demonstrate through a user study that our model with explicit conditioning over structure is preferred over other related approaches." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 425, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 555 + ], + "type": "text", + "content": "Video translation and propagation Image-to-image translation models, such as pix2pix [19, 62], can modify each frame in a video individually. This produces temporal inconsistencies as the time axis is ignored. Accounting for temporal or geometric information, such as flow, can increase consistency across frames when repurposing image synthesis models [42, 9]. We can extract such structural information to aid our spatio-temporal LDM in text- and image-guided video synthesis. Many generative adversarial methods, such as vid2vid [61, 60], leverage this type of input to guide synthesis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Video style transfer takes a reference style image and statistically applies its style to an input video [40, 8, 55]. In contrast, our method edits both style and content while preserving the structure of a video instead of matching feature statistics only. Text2Live [5] allows editing input videos using text prompts by decomposing a video into neural layers [21]. Once available, a layered video representation [37] provides consistent propagation across frames. SinFusion [29] and Tune-a-Video [63] use diffusion models to edit videos but require per-video training. This limits the practicality of the approaches in creative tools. We opt to instead train our model on a large-scale dataset permitting inference on any video without individual training." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7347" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 281, + 191 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 281, + 191 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 281, + 191 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 281, + 191 + ], + "type": "image", + "image_path": "49a6b142c1d07da6f025c6c46f1ccbb1a3ecf0c53b66bbd7274985a943e959d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": "Figure 2. Overview: During training (left), input videos " + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": " are encoded to " + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": " with a fixed encoder " + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": " and diffused to " + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": ". We extract a structure representation " + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": " by encoding depth maps obtained with MiDaS, and a content representation " + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": " by encoding one of the frames with CLIP. The model then learns to reverse the diffusion process in the latent space, with the help of " + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": ", which gets concatenated to " + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": ", which is provided via cross-attention blocks. During inference (right), the structure " + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 195, + 547, + 251 + ], + "type": "text", + "content": " of an input video is provided in the same manner. To specify content via text, we convert CLIP text embeddings to image embeddings via a prior." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 287, + 72, + 543, + 191 + ], + "blocks": [ + { + "bbox": [ + 287, + 72, + 543, + 191 + ], + "lines": [ + { + "bbox": [ + 287, + 72, + 543, + 191 + ], + "spans": [ + { + "bbox": [ + 287, + 72, + 543, + 191 + ], + "type": "image", + "image_path": "07748d2211bea29d7fb16cd61303243e0afac993e41a17118137d1da77979aac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 269, + 103, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 103, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 103, + 282 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 295, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 392 + ], + "type": "text", + "content": "For our purposes, it will be helpful to think of a video in terms of its content and structure. By structure, we refer to characteristics describing geometry and dynamics, e.g. shapes and locations of subjects as well as their temporal changes. We define content as features describing the appearance and semantics of the video, such as the colors and styles of objects and the lighting. The goal of our model is to edit the content of a video while retaining its structure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "text", + "content": "To achieve this, we learn a generative model " + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "inline_equation", + "content": "p(x|s,c)" + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "text", + "content": " of videos " + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "text", + "content": " conditioned on representations of structure " + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "text", + "content": " and content " + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "text", + "content": ". We infer the shape representation " + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "text", + "content": " from an input video, and modify it based on a text prompt " + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 397, + 289, + 506 + ], + "type": "text", + "content": " describing the edit. First, we describe our realization of the generative model as a conditional latent video diffusion model and, then, we describe our choices for shape and content representations. Finally, we discuss the optimization process of our model. See Fig. 2 for an overview." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 530, + 180, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 180, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 180, + 542 + ], + "type": "text", + "content": "3.1. Latent diffusion models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 555, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 579 + ], + "type": "text", + "content": "Diffusion models Diffusion models [51] learn to reverse a fixed forward diffusion process, which is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 600, + 287, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 600, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 79, + 600, + 287, + 615 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) := \\mathcal {N} \\left(x _ {t}, \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathcal {I}\\right). \\tag {1}", + "image_path": "45343357fbcc56ed6d4b9c793aee77275f840b583bdc526be22289704e1b2662.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": "Normally-distributed noise is slowly added to each sample " + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": ". The forward process models a fixed Markov chain and the noise is dependent on a variance schedule " + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": " being the total number of steps in our diffusion chain, and " + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "x_0 \\coloneqq x" + }, + { + "bbox": [ + 46, + 635, + 287, + 696 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Learning to Denoise The reverse process is defined accord" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 270, + 499, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 499, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 499, + 282 + ], + "type": "text", + "content": "ing to the following equation with parameters " + }, + { + "bbox": [ + 305, + 270, + 499, + 282 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 366, + 289, + 545, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 289, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 366, + 289, + 545, + 314 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {0}\\right) := \\int p _ {\\theta} \\left(x _ {0: T}\\right) d x _ {1: T} \\tag {2}", + "image_path": "090f9f7882d8c5e54dd366ae94914d951825ea1bd6d18f9e788aeda533b88155.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 316, + 545, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 316, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 353, + 316, + 545, + 349 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {0: T}\\right) = p \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right), \\tag {3}", + "image_path": "291714c4416ee003d004abb2b9a1d3ff1a66dddcab22a8a655accf4033891a02.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 351, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 351, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 330, + 351, + 545, + 365 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) := \\mathcal {N} \\left(x _ {t - 1}, \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right). \\tag {4}", + "image_path": "d12e97ef123f7f30e9b3810d2ac9cb2704a0dc194b30681229810821ef1cebe4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "content": "Using a fixed variance " + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "content": ", we are left learning the means of the reverse process " + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "content": ". Training is typically performed via a reweighted variational bound on the maximum likelihood objective, resulting in a loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 428, + 545, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 428, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 345, + 428, + 545, + 443 + ], + "type": "interline_equation", + "content": "L := \\mathbb {E} _ {t, q} \\lambda_ {t} \\| \\mu_ {t} \\left(x _ {t}, x _ {0}\\right) - \\mu_ {\\theta} \\left(x _ {t}, t\\right) \\| ^ {2}, \\tag {5}", + "image_path": "49989949c3c1b65d0461b5e0619ed6f956d6d69f08fc85588073f9040cc8a352.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 450, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 450, + 545, + 474 + ], + "type": "inline_equation", + "content": "\\mu_t(x_t, x_0)" + }, + { + "bbox": [ + 304, + 450, + 545, + 474 + ], + "type": "text", + "content": " is the mean of the forward process posterior " + }, + { + "bbox": [ + 304, + 450, + 545, + 474 + ], + "type": "inline_equation", + "content": "q(x_{t-1} | x_t, x_0)" + }, + { + "bbox": [ + 304, + 450, + 545, + 474 + ], + "type": "text", + "content": ", which is available in closed form [14]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "text", + "content": "Parameterization The mean " + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "text", + "content": " is then predicted by a UNet architecture [39] that receives the noisy input " + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "text", + "content": " and the diffusion timestep " + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "text", + "content": " as inputs. Other parameterizations and weightings, such as " + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "text", + "content": " [14] and " + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "text", + "content": "-parameterizations [46], can significantly improve sample quality compared to directly predicting the mean. Similar to [13], we found that " + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 475, + 545, + 569 + ], + "type": "text", + "content": "-parameterization improves color consistency thus all our experiments use it (see supp. material for more details)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": "Latent diffusion Latent diffusion models [38] (LDMs) take the diffusion process into the latent space. This provides improved separation between compressive and generative learning phases of the model. Specifically, LDMs use an autoencoder where an encoder " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": " maps input data " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": " to a lower dimensional latent code according to " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x)" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": " while a decoder " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": " converts latent codes back to the input space such that perceptually " + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "inline_equation", + "content": "x \\approx \\mathcal{D}(\\mathcal{E}(x))" + }, + { + "bbox": [ + 304, + 570, + 545, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Our encoder downsamples RGB-images " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "x\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " by a factor of eight and outputs four channels, resulting in a latent code " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "z\\in \\mathbb{R}^{4\\times H / 8\\times W / 8}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ". Thus, the diffusion UNet operates on a much smaller representation which significantly" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "7348" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 165, + 137 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 165, + 137 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 165, + 137 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 165, + 137 + ], + "type": "image", + "image_path": "1ce833b0bd6217de8e60b9a56f025de8b65431fc08cc7c2b25aec479669c2868.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 142, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 142, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 142, + 287, + 198 + ], + "type": "text", + "content": "Figure 3. Temporal Extension: We extend an image-based UNet architecture to videos, by adding temporal layers in its building blocks. We add a 1D temporal convolution after each 2D spatial convolution in its residual blocks (left), and we add a 1D temporal attention block after each of its 2D spatial attention blocks (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 72, + 286, + 137 + ], + "blocks": [ + { + "bbox": [ + 170, + 72, + 286, + 137 + ], + "lines": [ + { + "bbox": [ + 170, + 72, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 170, + 72, + 286, + 137 + ], + "type": "image", + "image_path": "6a1d8c064045f53420d1d4f077e731bf29ffffca4f42880d83270ffcfc2690ea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 215, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 287, + 251 + ], + "type": "text", + "content": "improves runtime and memory efficiency. The latter is crucial for video modeling where the additional time axis increases memory costs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 257, + 226, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 226, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 226, + 270 + ], + "type": "text", + "content": "3.2. Spatio-temporal Latent Diffusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "type": "text", + "content": "To correctly model a distribution over video frames, the architecture must account for temporal relationships. We also want to jointly learn an image model with shared parameters to benefit from better generalization obtained by training on large-scale image datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 335, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 395 + ], + "type": "text", + "content": "To achieve this, we extend an image architecture by introducing temporal layers, which are only active for video inputs. All other layers are shared between the image and video model. The autoencoder remains fixed and processes each frame in a video independently." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 395, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 395, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 287, + 515 + ], + "type": "text", + "content": "The UNet consists of two main building blocks: Residual blocks and transformer blocks (see Fig. 3). Similar to [17, 49], we extend them to videos by adding both 1D convolutions across time and 1D self-attention across time. In each residual block, we introduce one temporal convolution after each 2D convolution. Similarly, after each spatial 2D transformer block, we also include one temporal 1D transformer block, which mimics its spatial counterpart along the time axis. We also input learnable positional encodings of the frame index into temporal transformer blocks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 521, + 241, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 241, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 241, + 533 + ], + "type": "text", + "content": "3.3. Representing Content and Structure" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 539, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 598 + ], + "type": "text", + "content": "Conditional Diffusion Models Diffusion models are well-suited to modeling conditional distributions such as " + }, + { + "bbox": [ + 46, + 539, + 287, + 598 + ], + "type": "inline_equation", + "content": "p(x|s, c)" + }, + { + "bbox": [ + 46, + 539, + 287, + 598 + ], + "type": "text", + "content": ". The forward process " + }, + { + "bbox": [ + 46, + 539, + 287, + 598 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 539, + 287, + 598 + ], + "type": "text", + "content": " remains unchanged while the conditioning variables " + }, + { + "bbox": [ + 46, + 539, + 287, + 598 + ], + "type": "inline_equation", + "content": "s, c" + }, + { + "bbox": [ + 46, + 539, + 287, + 598 + ], + "type": "text", + "content": " become additional inputs to the model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 599, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 693 + ], + "type": "text", + "content": "Our goal is to edit a video based on a text prompt describing the desired output. We choose to train on uncaptioned video data due to the lack of large-scale paired video-text datasets of similar quality as image datasets like [48]. Therefore, during training, we must derive structure and content representations from the training video " + }, + { + "bbox": [ + 46, + 599, + 287, + 693 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 599, + 287, + 693 + ], + "type": "text", + "content": " itself, i.e. " + }, + { + "bbox": [ + 46, + 599, + 287, + 693 + ], + "type": "inline_equation", + "content": "s = s(x)" + }, + { + "bbox": [ + 46, + 599, + 287, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 599, + 287, + 693 + ], + "type": "inline_equation", + "content": "c = c(x)" + }, + { + "bbox": [ + 46, + 599, + 287, + 693 + ], + "type": "text", + "content": ", resulting in a per-example loss of" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\lambda_ {t} \\left\\| \\mu_ {t} \\left(\\mathcal {E} (x) _ {t}, \\mathcal {E} (x) _ {0}\\right) - \\mu_ {\\theta} \\left(\\mathcal {E} (x) _ {t}, t, s (x), c (x)\\right) \\right\\| ^ {2}. \\tag {6}", + "image_path": "eb10087f73af68479fb617f856d5d2fe1d7f157fc124a8da4fc31ac750747e91.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": "In contrast, during inference, structure " + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": " and content " + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": " are derived from an input video " + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": " and from a text prompt " + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": " respectively. An edited version " + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": " is obtained by sampling the generative model conditioned on " + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "s(y)" + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "c(t)" + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 129, + 545, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 129, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 356, + 129, + 545, + 142 + ], + "type": "interline_equation", + "content": "z \\sim p _ {\\theta} (z | s (y), c (t)), x = \\mathcal {D} (z). \\tag {7}", + "image_path": "4ef096fd545bfc4d820f35010ce101b9ab7e88cf4aaf49e2cea0367f64556660.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "content": "Content Representation We utilize CLIP [32] to infer a content representation from both text inputs " + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "content": " and video inputs " + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 151, + 545, + 270 + ], + "type": "text", + "content": " similar to previous works [35, 3]. CLIP embeddings are a promising content representation as they are more sensitive to semantic and stylistic properties while being more invariant towards geometric attributes [34]. During training, we encode a random frame in each input video with CLIP. To support text-based editing at inference, we train a prior model that allows sampling image embeddings from text embeddings [35, 49]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "text", + "content": "Structure Representation We need a representation that provides adequate separation between structure and content. We find that depth estimates extracted from input video frames provide the desired properties as they encode significantly less content information compared to simpler structure representations, such as edge filters which also encode textural properties. Still, depth maps reveal the silhouettes of objects which can prevent content edits involving changes in object shape." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "type": "text", + "content": "To offer control over the amount of structure to preserve, we propose to train a model on structure representations with varying amounts of information. In particular, we blur depth estimates given a parameter " + }, + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "type": "text", + "content": ". During training, " + }, + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "type": "text", + "content": " is randomly sampled between 0 and " + }, + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "type": "inline_equation", + "content": "T_s" + }, + { + "bbox": [ + 304, + 379, + 545, + 462 + ], + "type": "text", + "content": ". The parameter can then be controlled at inference to achieve different editing effects (see Fig. 10)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 463, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 533 + ], + "type": "text", + "content": "While depths map work well for our usecase, our approach generalizes to other geometric guidance features or combinations of features that might be more helpful for other specific applications. For example, models focusing on human video synthesis might benefit from estimated poses or face landmarks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 652 + ], + "type": "text", + "content": "Conditioning Mechanisms We account for the different characteristics of our content and structure representations with two different conditioning mechanisms. Since structure represents a significant portion of the spatial information of video frames, we use concatenation for conditioning to make effective use of this information. In contrast, attributes described by the content representation are not tied to particular locations. Hence, we leverage cross-attention which can effectively transport this information to any position." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "We use the spatial transformer blocks of the UNet architecture for cross-attention conditioning. Each contains two attention operations, where the first one performs a spatial self-attention and the second one a cross attention with keys and values computed from the CLIP image embedding." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7349" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 248, + 167 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 248, + 167 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 248, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 248, + 167 + ], + "type": "image", + "image_path": "d1517ada36b78af12d70bb7b303f9901bde5e96dcacaf865c363dd08c565ad71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 170, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 547, + 225 + ], + "type": "text", + "content": "Figure 4. Temporal Control: By training image and video models jointly, we obtain explicit control over the temporal consistency of edited videos via a temporal guidance scale " + }, + { + "bbox": [ + 46, + 170, + 547, + 225 + ], + "type": "inline_equation", + "content": "\\omega_{t}" + }, + { + "bbox": [ + 46, + 170, + 547, + 225 + ], + "type": "text", + "content": ". On the left, frame consistency measured via CLIP cosine similarity of consecutive frames increases monotonically with " + }, + { + "bbox": [ + 46, + 170, + 547, + 225 + ], + "type": "inline_equation", + "content": "\\omega_{t}" + }, + { + "bbox": [ + 46, + 170, + 547, + 225 + ], + "type": "text", + "content": ", while mean squared error between frames warped with optical flow decreases monotonically. On the right, lower scales (0.5 in the middle row) achieve edits with a \"hand-drawn\" look, whereas higher scales (1.5 in the bottom row) result in smoother results. Top row shows the original input video, the two edits use the prompt \"pencil sketch of a man looking at the camera\"." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 271, + 72, + 545, + 168 + ], + "blocks": [ + { + "bbox": [ + 271, + 72, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 271, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 271, + 72, + 545, + 168 + ], + "type": "image", + "image_path": "31ee90a69d58b6e62cc7bd826b5b2059575e76ab0ffdb225891fbd5ff8b13359.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 237, + 108, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 237, + 108, + 249 + ], + "spans": [ + { + "bbox": [ + 75, + 237, + 108, + 249 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 237, + 409, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 237, + 409, + 248 + ], + "spans": [ + { + "bbox": [ + 246, + 237, + 409, + 248 + ], + "type": "text", + "content": "Driving Video (top) and Result (bottom)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 278, + 117, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 278, + 117, + 325 + ], + "spans": [ + { + "bbox": [ + 65, + 278, + 117, + 325 + ], + "type": "text", + "content": "a man using a laptop inside a train, anime style" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 361, + 117, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 361, + 117, + 443 + ], + "spans": [ + { + "bbox": [ + 65, + 361, + 117, + 443 + ], + "type": "text", + "content": "a woman and man take selfies while walking down the street, claymation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 487, + 116, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 487, + 116, + 521 + ], + "spans": [ + { + "bbox": [ + 65, + 487, + 116, + 521 + ], + "type": "text", + "content": "kite-surfer in the ocean at sunset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 576, + 117, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 576, + 117, + 634 + ], + "spans": [ + { + "bbox": [ + 65, + 576, + 117, + 634 + ], + "type": "text", + "content": "alien expplorer hiking in the mountains" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 129, + 253, + 526, + 350 + ], + "blocks": [ + { + "bbox": [ + 129, + 253, + 526, + 350 + ], + "lines": [ + { + "bbox": [ + 129, + 253, + 526, + 350 + ], + "spans": [ + { + "bbox": [ + 129, + 253, + 526, + 350 + ], + "type": "image", + "image_path": "ce111ac29162328e1eaf09914a94e114951db0b7d4109df2605d4b4293350891.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 129, + 354, + 526, + 451 + ], + "blocks": [ + { + "bbox": [ + 129, + 354, + 526, + 451 + ], + "lines": [ + { + "bbox": [ + 129, + 354, + 526, + 451 + ], + "spans": [ + { + "bbox": [ + 129, + 354, + 526, + 451 + ], + "type": "image", + "image_path": "e42bb00d6644197b0e66ec95016b9adf507f5e4b8b075ee34bd03f8360ebfbd6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 129, + 456, + 526, + 552 + ], + "blocks": [ + { + "bbox": [ + 129, + 456, + 526, + 552 + ], + "lines": [ + { + "bbox": [ + 129, + 456, + 526, + 552 + ], + "spans": [ + { + "bbox": [ + 129, + 456, + 526, + 552 + ], + "type": "image", + "image_path": "fda5ce90649ab45342be472ae5f4aa0bf1a728a2736ce4f19c10a255d36e4b16.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 127, + 555, + 525, + 652 + ], + "blocks": [ + { + "bbox": [ + 127, + 555, + 525, + 652 + ], + "lines": [ + { + "bbox": [ + 127, + 555, + 525, + 652 + ], + "spans": [ + { + "bbox": [ + 127, + 555, + 525, + 652 + ], + "type": "image", + "image_path": "85be900d6f5fd989d4f2db2b2f03b2f5e450916997cf48be6ef634e217f8efa5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 654, + 546, + 676 + ], + "lines": [ + { + "bbox": [ + 46, + 654, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 546, + 676 + ], + "type": "text", + "content": "Figure 5. Our approach enables a wide range of video edits, including changes to animation styles such as anime or claymation, changes of environment such as time of day, and changing characters such as humans to aliens." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7350" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 289, + 186 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 186 + ], + "type": "image", + "image_path": "2ca212a548bf5cddbe32e0509d4c948c14f09dd9609dbf091c24bfb251e6e1ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 289, + 256 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 289, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 289, + 256 + ], + "type": "text", + "content": "Figure 6. Prompt-vs-frame consistency: Image models such as SD-Depth achieve good prompt consistency but fail to produce consistent edits across frames. Propagation based approaches such as IVS and Text2Live increase frame consistency but fail to provide edits reflecting the prompt accurately. Our method achieves the best combination of frame and prompt consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "text", + "content": "To condition on structure, we first estimate depth maps for all input frames using the MiDaS DPT-Large model [36]. We then apply " + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "text", + "content": " iterations of blurring and downsampling to the depth maps, where " + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "text", + "content": " controls the amount of structure to preserve. We resample the perturbed depth map to the resolution of the RGB-frames and encode it using " + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "text", + "content": ". This latent representation of structure is concatenated with the input " + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "text", + "content": " given to the UNet. We also input four channels containing a sinusoidal embedding of " + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 46, + 274, + 287, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 381, + 288, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 288, + 502 + ], + "type": "text", + "content": "Sampling While Eq. (2) provides a direct way to sample from the trained model, many other sampling methods [52, 24, 22] require only a fraction of the number of diffusion timesteps to achieve good sample quality. We use DDIM [52] throughout our experiments. Furthermore, classifier-free diffusion guidance [16] significantly improves sample quality. For a conditional model " + }, + { + "bbox": [ + 46, + 381, + 288, + 502 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t,t,c)" + }, + { + "bbox": [ + 46, + 381, + 288, + 502 + ], + "type": "text", + "content": ", this is achieved by training the model to also perform unconditional predictions " + }, + { + "bbox": [ + 46, + 381, + 288, + 502 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t,t,\\emptyset)" + }, + { + "bbox": [ + 46, + 381, + 288, + 502 + ], + "type": "text", + "content": " and then adjusting predictions during sampling according to" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 510, + 285, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 510, + 285, + 525 + ], + "spans": [ + { + "bbox": [ + 49, + 510, + 285, + 525 + ], + "type": "interline_equation", + "content": "\\tilde {\\mu} _ {\\theta} (x _ {t}, t, c) = \\mu_ {\\theta} (x _ {t}, t, \\emptyset) + \\omega (\\mu_ {\\theta} (x _ {t}, t, c) - \\mu_ {\\theta} (x _ {t}, t, \\emptyset))", + "image_path": "8631d55a1f98d1b8dd732d448f9414c1b82d35609c52985eca3aea003a9d3612.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "text", + "content": " is the guidance scale that controls the strength. Based on the intuition that " + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "text", + "content": " extrapolates the direction between an unconditional and a conditional model, we apply this idea to control temporal consistency of our model. Specifically, since we are training both an image and a video model with shared parameters, we can consider predictions by both models for the same input. Let " + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(z_t,t,c,s)" + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "text", + "content": " denote the prediction of our video model, and let " + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}^{\\pi}(z_t,t,c,s)" + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "text", + "content": " denote the prediction of the image model applied to each frame individually. Taking classifier-free guidance for " + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 532, + 287, + 664 + ], + "type": "text", + "content": " into account, we then adjust our prediction according to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 672, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 672, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 57, + 672, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\mu} _ {\\theta} (z _ {t}, t, c, s) = \\mu_ {\\theta} ^ {\\pi} (z _ {t}, t, \\emptyset , s) \\\\ + \\omega_ {t} \\left(\\mu_ {\\theta} \\left(x _ {t}, t, \\emptyset , s\\right) - \\mu_ {\\theta} ^ {\\pi} \\left(x _ {t}, t, \\emptyset , s\\right)\\right) \\tag {8} \\\\ + \\omega (\\mu_ {\\theta} (x _ {t}, t, c, s) - \\mu_ {\\theta} (x _ {t}, t, \\emptyset , s)) \\\\ \\end{array}", + "image_path": "bf1533b22df987a47fa109404d470068aaade74c067309b07f96ed7c2475f718.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 90, + 545, + 192 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 526, + 83 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 526, + 83 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 526, + 83 + ], + "type": "text", + "content": "ours SDEdit (50% strength) IVS (50% strength) Depth-SDours (-, 50% SDEdit) Deformu SDEdit (75% strength) IVS (75% strength) Text2LIVEours (-, 75% SDEdit)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 307, + 90, + 545, + 192 + ], + "lines": [ + { + "bbox": [ + 307, + 90, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 307, + 90, + 545, + 192 + ], + "type": "image", + "image_path": "fc2e31a43f198dfd5c9c37db17a2722366044e9ca9ae6494c9f9f94cbac41a27.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 196, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 306, + 196, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 306, + 196, + 545, + 218 + ], + "type": "text", + "content": "Figure 7. User Preferences: Based on our user study, the results from our model are preferred over the baseline models." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 238, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 238, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 545, + 262 + ], + "type": "text", + "content": "Our experiments demonstrate that this approach controls temporal consistency in the outputs, see Fig. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 270, + 391, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 391, + 283 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 391, + 283 + ], + "type": "text", + "content": "3.4. Optimization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": "We train on an internal dataset of 240M images and 6.4M video clips. We use image batches of size 9216 with resolutions of " + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "inline_equation", + "content": "320 \\times 320" + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "inline_equation", + "content": "384 \\times 320" + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "inline_equation", + "content": "448 \\times 256" + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": ", as well as the same resolutions with flipped aspect ratios. We sample image batches with a probability of " + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "inline_equation", + "content": "12.5\\%" + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": ". For videos, we use batch size 1152 and 8 frames from each video sampled four frames apart with a resolution of " + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "inline_equation", + "content": "448 \\times 256" + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "type": "text", + "content": "We train our model in multiple stages. First, we initialize model weights based on a pretrained text-conditional latent diffusion model " + }, + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "type": "inline_equation", + "content": "[38]^1" + }, + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "type": "text", + "content": ". We change the conditioning from CLIP text embeddings to CLIP image embeddings and fine-tune for 15k steps on images only. Afterwards, we introduce temporal connections as described in Sec. 3.2 and train jointly on images and videos for 75k steps. We then add conditioning on structure " + }, + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "type": "inline_equation", + "content": "t_s \\equiv 0" + }, + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "type": "text", + "content": " fixed and train for 25k steps. Finally, we resume training with " + }, + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 304, + 373, + 546, + 493 + ], + "type": "text", + "content": " sampled uniformly between 0 and 7 for another 10k steps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 504, + 359, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 359, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 359, + 516 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 524, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 524, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 305, + 524, + 545, + 584 + ], + "type": "text", + "content": "To evaluate our approach, we use videos from DAVIS [31] and various stock footage. To automatically create edit prompts, we first run a captioning model [23] to obtain a description of the original video content. We then use GPT-3 [7] to generate edited prompts." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 592, + 417, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 592, + 417, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 592, + 417, + 605 + ], + "type": "text", + "content": "4.1. Qualitative Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 610, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 610, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 610, + 545, + 696 + ], + "type": "text", + "content": "We demonstrate that our approach performs well on a number of diverse inputs (see Fig. 5). Our method handles a large variety of footage, such as landscapes and close-ups, and diverse camera motion without any explicit tracking of the input. Our depth-based structure representation combined with large-scale image-video joint training enable strong generalization and powerful editing capabilities." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 703, + 534, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 534, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 534, + 713 + ], + "type": "text", + "content": "1https://github.com/runwayml/stable-diffusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 732, + 314, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 314, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 314, + 742 + ], + "type": "text", + "content": "7351" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 70, + 526, + 211 + ], + "blocks": [ + { + "bbox": [ + 65, + 89, + 86, + 99 + ], + "lines": [ + { + "bbox": [ + 65, + 89, + 86, + 99 + ], + "spans": [ + { + "bbox": [ + 65, + 89, + 86, + 99 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 136, + 86, + 145 + ], + "lines": [ + { + "bbox": [ + 65, + 136, + 86, + 145 + ], + "spans": [ + { + "bbox": [ + 65, + 136, + 86, + 145 + ], + "type": "text", + "content": "Mask" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 168, + 118, + 206 + ], + "lines": [ + { + "bbox": [ + 65, + 168, + 118, + 206 + ], + "spans": [ + { + "bbox": [ + 65, + 168, + 118, + 206 + ], + "type": "text", + "content": "A snowboarder in a snow park on the mountain" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 127, + 70, + 526, + 211 + ], + "lines": [ + { + "bbox": [ + 127, + 70, + 526, + 211 + ], + "spans": [ + { + "bbox": [ + 127, + 70, + 526, + 211 + ], + "type": "image", + "image_path": "896f9d2efe6f6ac540f6af75a3640a0bab9f6c20ebcec5d086b90e2566d5a444.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 214, + 533, + 225 + ], + "lines": [ + { + "bbox": [ + 47, + 214, + 533, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 533, + 225 + ], + "type": "text", + "content": "Figure 8. Background Editing: Masking the denoising process allows us to restrict edits to backgrounds for more control over results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 245, + 286, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 286, + 293 + ], + "type": "text", + "content": "ties. For example, we can produce various animation styles, changes in time of day, and more complex changes of subject, such as turning a hiker into an alien (see Fig. 5). Please see the supplementary material for more results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 293, + 286, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 286, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 286, + 388 + ], + "type": "text", + "content": "Using CLIP image embeddings as the content representation allows users to specify content through images. As an example application, we demonstrate character replacement in Fig. 9. For every video in a set of six videos, we resynthesize it five times, each time providing a single content image taken from another video in the set. We can retain content characteristics with " + }, + { + "bbox": [ + 46, + 293, + 286, + 388 + ], + "type": "inline_equation", + "content": "t_s = 3" + }, + { + "bbox": [ + 46, + 293, + 286, + 388 + ], + "type": "text", + "content": " despite large differences in their pose and shape." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 388, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 449 + ], + "type": "text", + "content": "Lastly, we illustrate the use of masked video editing in Fig. 8, where the model predicts everything outside the masked area(s) while retaining the original content inside the masked area. Notably, this technique resembles inpainting with diffusion models [43, 25]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 456, + 122, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 122, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 122, + 468 + ], + "type": "text", + "content": "4.2. User Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 474, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 286, + 605 + ], + "type": "text", + "content": "We benchmark against Text2Live [5], a recent approach for text-guided video editing that employs layered neural atlases [21]. As a baseline, we compare against SDEdit [26] in two ways; per-frame generated results and a first-frame result propagated by a few-shot video stylization method [55] (IVS). We also include two depth-based versions of Stable Diffusion; one trained with depth-conditioning [2] and one that retains past results based on depth estimates [9]. We also include an ablation: applying SDEdit to our video model trained without conditioning on a structure representation (ours, " + }, + { + "bbox": [ + 46, + 474, + 286, + 605 + ], + "type": "inline_equation", + "content": "\\sim s" + }, + { + "bbox": [ + 46, + 474, + 286, + 605 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 700 + ], + "type": "text", + "content": "We judge the success of our method qualitatively based on a user study. We run the user study using Amazon Mechanical Turk (AMT) on an evaluation set of 35 representative video editing prompts. For each example, we ask 5 annotators to compare faithfulness to the video editing prompt (\"Which video better represents the provided edited caption?\") between a baseline and our method, presented in random order, and use a majority vote for the final result." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "The results can be found in Fig. 7. Across all compared" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 245, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 484 + ], + "type": "text", + "content": "methods, results from our approach are preferred roughly 3 out of 4 times. A visual comparison among the methods can be found in the supplementary. We observe that SDEdit is sensitive to the editing strength. Low values fail to achieve the desired editing effect whereas high values change the structure of the input. Even with a fixed seed, both style and structure can change in unnatural ways between frames as their relationship is ignored by image-based approaches. Propagation of SDEdit outputs (IVS) leads to more consistent results but often introduces propagation artifacts especially with large motion. Depth-SD produces accurate, structure-preserving edits for individual frames but frames are inconsistent across time. The outputs of Text2Live tend to be temporally smooth due to its reliance on Layered Neural Atlases [21], but it often produces edits that represent the edit prompt inaccurately. A direct comparison with Text2Live is difficult as it requires input masks and separate edit prompts for foreground and background. In addition, computing a neural atlas takes about 10 hours whereas our approach requires approximately a minute." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 491, + 441, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 441, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 441, + 504 + ], + "type": "text", + "content": "4.3.Quantitative Evaluation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "type": "text", + "content": "We quantify trade-offs between frame consistency and prompt consistency with the following two metrics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 569 + ], + "type": "text", + "content": "Frame consistency We compute CLIP image embeddings on all frames of output videos and report the average cosine similarity between all pairs of consecutive frames." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "content": "Prompt consistency We compute CLIP image embeddings on all frames of output videos and the CLIP text embedding of the edit prompt. We report average cosine similarity between text and image embedding over all frames." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Fig. 6 shows the results of each model using our frame consistency and prompt consistency metrics. Our model tends to outperform the baseline models in both aspects (placed higher in the upper-right quadrant of the graph). We also notice a slight tradeoff with increasing the strength parameters in the baseline models: larger strength scales implies higher prompt consistency at the cost of lower frame consistency. Increasing the temporal scale " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\omega_{t})" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " of our model" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7352" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 287, + 437 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 287, + 437 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 287, + 437 + ], + "type": "image", + "image_path": "8f6488259aa3253a32713eb6e4cf4dd0e6787e91732143fe84c6231696b88c09.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 439, + 287, + 462 + ], + "lines": [ + { + "bbox": [ + 47, + 439, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 287, + 462 + ], + "type": "text", + "content": "Figure 9. Image Prompting: We combine the structure of a driving video (first column) with content from other videos (first row)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "text", + "content": "results in higher frame consistency but lower prompt consistency. We also observe that an increased structure scale " + }, + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "inline_equation", + "content": "(t_{s})" + }, + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "text", + "content": " results in higher prompt consistency as the content becomes less determined by the input structure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 536, + 138, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 138, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 138, + 547 + ], + "type": "text", + "content": "4.4.Customization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": "Customization of image models enables generation of previously unseen content, such as specific people or styles, based on a small dataset used for finetuning [41]. We fine-tune our depth-conditioned latent video diffusion model on a set of 15-30 images and produce videos containing the desired subject. Half of the batch elements are of the subject and the other half belong to the original training dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Fig. 10 shows an example with different numbers of customization steps as well as different levels of structure preservation " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "t_{s}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": ". Customization improves fidelity to the style and appearance of the character. In combination with higher " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "t_{s}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " values, accurate animations are possible despite using a driving video of a person with different characteristics." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 57, + 545, + 393 + ], + "blocks": [ + { + "bbox": [ + 306, + 57, + 545, + 393 + ], + "lines": [ + { + "bbox": [ + 306, + 57, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 57, + 545, + 393 + ], + "type": "image", + "image_path": "79edf299909923c61304836278cf75d17787db030dc4005eb0536b31bffe848c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 396, + 545, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 464 + ], + "type": "text", + "content": "Figure 10. Controlling Fidelity: We obtain control over structure and appearance-fidelity. Each cell shows three frames produced with decreasing structure-fidelity " + }, + { + "bbox": [ + 305, + 396, + 545, + 464 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 305, + 396, + 545, + 464 + ], + "type": "text", + "content": " (left-to-right) and increasing number of customization training steps (top-to-bottom). The bottom shows examples of images used for customization (red border) and the input image (blue border). Same driving video as in Fig. 1." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 476, + 378, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 378, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 378, + 488 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Our latent video diffusion model synthesizes new videos given structure and content information. We ensure structural consistency by conditioning on depth estimates while content is controlled with images or natural language. Temporal layers and joint image-video training achieve stable results across frames. A novel guidance method, inspired by classifier-free guidance, allows for control over temporal consistency. By training on depth maps with varying degrees of detail, we can adjust the level of structure preservation. This, together with model customization, improves content fidelity. Our quantitative evaluation and user study show that our method is preferred over related approaches. Future works should investigate other conditioning data, such as facial landmarks and pose estimates, and additional 3d-priors to improve generated results. Our model is intended for creative applications in content creation, but we realize the risks of dual-use and hope that further work will be aimed at combating abuse of generative models." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7353" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "text", + "content": "[1] Dinesh Acharya, Zhiwu Huang, Danda Pani Paudel, and Luc Van Gool. Towards high resolution video generation with progressive growing of sliced Wasserstein gans. arXiv preprint arXiv:1810.02419, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 286, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 286, + 169 + ], + "type": "text", + "content": "[2] Stability AI. Stable diffusion depth. https://github.com/Stability-AI/stablediffusion, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 172, + 286, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 286, + 236 + ], + "type": "text", + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 286, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 286, + 283 + ], + "type": "text", + "content": "[4] Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie S. Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 286, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 286, + 329 + ], + "type": "text", + "content": "[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European Conference on Computer Vision, pages 707-723. Springer, 2022. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 331, + 286, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 286, + 374 + ], + "type": "text", + "content": "[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A Efros, and Tero Karras. Generating long videos of dynamic scenes. 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 377, + 286, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 377, + 286, + 518 + ], + "spans": [ + { + "bbox": [ + 53, + 377, + 286, + 518 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc., 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 521, + 286, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 286, + 565 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 286, + 565 + ], + "type": "text", + "content": "[8] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1105-1114, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 567, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 286, + 597 + ], + "type": "text", + "content": "[9] deforum. Deforum stable diffusion. https://github.com/deforum/stable-diffusion, 2022. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 286, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 286, + 688 + ], + "type": "text", + "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. Cogview: Mastering text-to-image generation via transformers. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 19822-19835. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[11] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "Long video generation with time-agnostic vqgan and time-sensitive transformer. arXiv preprint arXiv:2204.03638, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 27. Curran Associates, Inc., 2014. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 173, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 227 + ], + "type": "text", + "content": "[13] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 282, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 282, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 282, + 545, + 326 + ], + "type": "text", + "content": "[15] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47-1, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 327, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 348 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 380 + ], + "type": "text", + "content": "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 382, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 425 + ], + "type": "text", + "content": "[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 426, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 479 + ], + "type": "text", + "content": "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 513 + ], + "type": "text", + "content": "[20] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Proc. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "text", + "content": "[21] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 579 + ], + "type": "text", + "content": "[22] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 689 + ], + "type": "text", + "content": "[24] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7354" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 106 + ], + "type": "text", + "content": "using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[26] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. CoRR, abs/2108.01073, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 297 + ], + "type": "text", + "content": "[28] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 16784-16804. PMLR, 17-23 Jul 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 331 + ], + "type": "text", + "content": "[29] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 353 + ], + "type": "text", + "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "text", + "content": "[31] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 400, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 466 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "text", + "content": "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1), jun 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "text", + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 1, 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 633 + ], + "type": "text", + "content": "[35] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8821-8831. PMLR, 18-24 Jul 2021. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[36] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:1623-1637, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[37] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew William Fitzgibbon. Unwrap mosaics: a new repre" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "sentation for video editing. ACM SIGGRAPH 2008 papers, 2008. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "type": "text", + "content": "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "text", + "content": "[40] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In Bodo Rosenhahn and Bjoern Andres, editors, Pattern Recognition, pages 26-36, Cham, 2016. Springer International Publishing. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 272 + ], + "type": "text", + "content": "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 273, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 294 + ], + "type": "text", + "content": "[42] Alexander S. Disco diffusion v5.2 - warp fusion. https://github.com/Sxela/DiscoDiffusion-Warp, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 295, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 545, + 337 + ], + "type": "text", + "content": "[43] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 404 + ], + "type": "text", + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 449 + ], + "type": "text", + "content": "[45] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Proceedings of the IEEE international conference on computer vision, pages 2830-2839, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "text", + "content": "[46] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 483, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 514 + ], + "type": "text", + "content": "[47] Robin San-Roman, Eliya Nachmani, and Lior Wolf. Noise estimation for generative diffusion models. arXiv preprint arXiv:2104.02600, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 570 + ], + "type": "text", + "content": "[48] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "text", + "content": "[49] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 658 + ], + "type": "text", + "content": "[50] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 659, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 713 + ], + "type": "text", + "content": "[51] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 741 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 741 + ], + "type": "text", + "content": "7355" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "of Machine Learning Research, pages 2256-2265, Lille, France, 07-09 Jul 2015. PMLR. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "text", + "content": "[53] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "type": "text", + "content": "[54] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[55] Ondrej Texler, David Futschik, Michal Kucera, Ondrej Jamriška, Šárka Sochorová, Mengei Chai, Sergey Tulyakov, and Daniel Sýkora. Interactive video stylization using few-shot patch-based training. ACM Transactions on Graphics, 39(4):73, 2020. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[56] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1526-1535, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 342 + ], + "type": "text", + "content": "[57] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. Advances in Neural Information Processing Systems, 34:11287-11302, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 344, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 398 + ], + "type": "text", + "content": "[58] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 433 + ], + "type": "text", + "content": "[59] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "text", + "content": "[60] Ting-Chun Wang, Ming-Yu Liu, Andrew Tao, Guilin Liu, Jan Kautz, and Bryan Catanzaro. Few-shot video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "text", + "content": "[61] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[62] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[63] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[64] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[65] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 185 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "type": "text", + "content": "[66] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[67] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models, 2022. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "7356" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/4d750dca-9a8b-44f9-b884-62a2d6314ac8_content_list.json b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/4d750dca-9a8b-44f9-b884-62a2d6314ac8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..750067c2df4963d129a47022338851e3c597165b --- /dev/null +++ b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/4d750dca-9a8b-44f9-b884-62a2d6314ac8_content_list.json @@ -0,0 +1,2315 @@ +[ + { + "type": "text", + "text": "Structure-Aware Surface Reconstruction via Primitive Assembly", + "text_level": 1, + "bbox": [ + 156, + 130, + 813, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingen Jiang $^{1,2,3,4^*}$ Mingyang Zhao $^{1,5^*}$ Shiqing Xin $^{4}$ Yanchao Yang $^{6}$", + "bbox": [ + 192, + 186, + 776, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanxiao Wang $^{1,3}$ Xiaohong Jia $^{2,3}$ Dong-Ming Yan $^{1,3\\dagger}$", + "bbox": [ + 256, + 210, + 720, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MAIS & NLPR, CASIA $^{2}$ KLMM, AMSS, CAS", + "bbox": [ + 297, + 229, + 679, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ UCAS $^{4}$ Shandong University $^{5}$ BAAI $^{6}$ The University of Hong Kong", + "bbox": [ + 186, + 247, + 781, + 266 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7aa06ecab35991e406759e0990b1bbfce592b2d4f6ebad5f072439fb9b300141.jpg", + "image_caption": [ + "Figure 1: Surface reconstruction for measurement point clouds. In contrast to piecewise approximations (KSR), our method not only delivers manifold and watertight surfaces but also retains the primitive structures (planes, cylinders, spheres, cones, and tori) faithfully. Also, our method enables the recovery of the sharp geometry of objects (e.g., feature lines) as a byproduct." + ], + "image_footnote": [], + "bbox": [ + 127, + 267, + 833, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 564, + 313, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel and efficient method for reconstructing manifold surfaces from point clouds. Unlike previous approaches that use dense implicit reconstructions or piecewise approximations and overlook inherent structures like quadrics in CAD models, our method faithfully preserves these quadric structures by assembling primitives. To achieve high-quality primitive extraction, we use a variational shape approximation, followed by a mesh arrangement for space partitioning and candidate primitive patches generation. We then introduce an effective pruning mechanism to classify candidate primitive patches as active or inactive, and further prune inactive patches to reduce the search space and speed up surface extraction significantly. Finally, the optimal active patches are computed by a binary linear programming and assembled as manifold and watertight surfaces. We perform extensive experiments on a wide range of CAD objects to validate its effectiveness.", + "bbox": [ + 75, + 583, + 472, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 563, + 630, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The problem of converting unorganized point clouds into surface meshes, also known as surface reconstruction, has various applications in 3D vision, computer graphics, and architecture geometry. In recent decades, although many powerful algorithms have been devoted to pursuing faithful and accurate reconstructions [19, 21, 34, 28, 27, 49], it remains a challenging task due to the inherent ill-posed nature [11, 36, 5].", + "bbox": [ + 496, + 593, + 893, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The popular Poisson reconstruction methods [24, 25] implicitly generate dense smooth surfaces for point clouds, while others mainly focus on giving a compact approximation to objects by piecewise planar shapes or polygonal meshes [7, 13, 3, 35]. However, man-made objects in real-world scenarios are typically composed of non-planar quadratic surfaces or quadric structures, such as the Computer-Aided Design (CAD) models (see Fig. 1 as an example). Therefore, reconstructing manifold surface meshes from measurement data meanwhile retaining their intrinsic primitive structures in principle shall generate higher-quality approximation results than solely relying on poly-", + "bbox": [ + 496, + 719, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The authors contribute equally to this work.", + "bbox": [ + 94, + 875, + 334, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author (yandongming@gmail.com).", + "bbox": [ + 94, + 887, + 372, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14171", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gons.", + "bbox": [ + 76, + 93, + 116, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We argue that a method that solves the aforesaid problem ought to satisfy the following objectives: 1) Structure-awareness. The reconstruction surfaces should faithfully keep the structural features of the input point cloud, for instance, the common primitives in CAD, meanwhile correctly recovering the detailed geometry, and giving a high-fidelity approximation (e.g., the minimal Hausdorff distance) to the target object. 2) Manifold guarantees. The output surface meshes should be oriented 2-manifold and watertight [35]. 3) Efficiency. The developed algorithm should be scalable to large-scale point clouds with an acceptable computational complexity [3].", + "bbox": [ + 76, + 106, + 467, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we develop a novel and efficient framework that takes the underlying geometric structures of point clouds into account when performing shape reconstruction. Our method ensures high-quality outputs and aims to achieve the above objectives. The core of our algorithm lies in a faithful reconstruction of intrinsic primitive structures presented in CAD models, as opposed to the direct surface approximation based on piecewise planar primitives. We consider five representative types of primitives commonly found in modern CAD systems, including four quadrics (planes, cylinders, cones, spheres) and tori. We adopt a variational shape approximation to optimize the extracted primitives followed by mesh arrangement techniques to attain primitive patches. A patch is a manifold that consists of a set of connected triangle meshes. As the size of patches is typically smaller than that of facets, our method enables a more stable and efficient reconstruction process than plane-based ones. Moreover, we propose a patch pruning mechanism to further speed up the consecutive surface extraction process. This mechanism distinguishes patches as either active or inactive, allowing us to perform fast local collision detection instead of a time-consuming global processing. Lastly, we select an optimal set of patches from the active candidate patches to assemble a reconstruction surface that respects the underlying structure of the object. To this end, we optimize a Binary Linear Programming (BLP) under hard constraints, as done in [35], to guarantee that the resulting surface is both manifold and watertight.", + "bbox": [ + 76, + 287, + 467, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We perform extensive experiments to validate the effectiveness of our algorithm and compare it with state-of-the-art approaches using a variety of CAD models. Experimental results evidence that our method outperforms competitors in terms of both approximation precision and reconstruction efficiency, while also preserving the intrinsic primitive structures of the input point clouds with much better fidelity. To summarize, the main technical contributions include:", + "bbox": [ + 76, + 710, + 467, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- An efficient structure-aware reconstruction method that produces accurate surface models while preserving the underlying intrinsic primitive structures of the", + "bbox": [ + 94, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "input point clouds.", + "bbox": [ + 532, + 90, + 656, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A patch pruning technique that avoids the time-consuming global collision detection and accelerates the surface extraction by a large margin.", + "- A patch-induced binary linear programming for assembling manifold and watertight surfaces from candidate active patches, which significantly reduces the search space and helps achieve efficient and accurate surface reconstruction." + ], + "bbox": [ + 517, + 114, + 888, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 256, + 640, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we briefly review the works that are most relevant to ours. Please refer to [5, 23] for a more comprehensive discussion.", + "bbox": [ + 498, + 281, + 890, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Geometric primitive extraction. Segmenting 3D point cloud models into a collection of geometric primitives such as planes and cylinders is helpful in most reconstruction approaches. Popular extraction frameworks include region growing [31, 39] and RANSAC [43, 52]. The former strategy takes the neighboring structure of models into consideration and propagates candidate primitives on their $k$ -nearest neighbor graph until the fitting error exceeds a preset tolerance. Instead, RANSAC adopts a hypothesis and verification mechanism to recover primitives that embrace the most inliers. In order to make full use of the a priori structure of the input model, Li et al. [29] started with primitive extraction by RANSAC, followed by the refinement subject to spatial relationships such as parallelism and orthogonality. This idea was further extended to extract planar primitives simultaneously [32] or sequentially [37]. Recently, Yu et al. [49] proposed a new algorithm for extracting planar primitives, which has powerful potential for plane-assembly reconstruction. In our work, we leverage an efficient region growing strategy [39] for primitive extraction followed by a refinement via variational shape optimization.", + "bbox": [ + 498, + 325, + 892, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assembly-based reconstruction. The core idea of assembly-based methods is to integrate all extracted primitives into a hybrid mesh. Connectivity analysis and slicing are two principal assembly manners. Connectivity strategies [8, 42, 46] consider neighboring relationships between primitives and encode them into an adjacency graph, which is adapted to regularize output meshes. Slicing strategies [6, 7, 3, 33, 35, 13] extend the extracted primitives and slice them to construct a spatial partition for candidate facets. Connectivity strategies are typically efficient but suffer from linkage errors of adjacent graphs generated by incomplete inputs. Arikan et al. [1] remedied the incomplete part by a user-guided interactive method that is labor-intensive especially for large-scale measurement data, whereas Lafarge et al. [26] provided an automatic solution that first consolidated the point clouds and then executed a min-cut optimization to extract dense triangle meshes.", + "bbox": [ + 498, + 643, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14172", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dfc11adea0621a6c7e6ad414e238db5da038e88ec50b03180d7c8595eb623d4a.jpg", + "image_caption": [ + "Figure 2: Overview of the proposed method. (a) Point cloud after primitive extraction; (b) Accomplished space partitioning; (c) Active primitive patches after patch pruning; (d) The reconstructed surface." + ], + "image_footnote": [], + "bbox": [ + 86, + 77, + 890, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alternatively, slicing strategies deal with incomplete models by means of extending primitives. Nan et al. [35] performed an exhaustive partition to compensate for all possible missing and thus guaranteed a manifold output via BLP. Nevertheless, this manner typically leads to high computational complexity for the subsequent surface extraction. To circumvent this shortcoming, Fang et al. [13] proposed a hybrid method that combines the connectivity analysis with the slicing process to make a limited partition, while Bauchet et al. [3] adopted a kinetic data structure [16, 15] to monitor the growing of convex piecewise planar polygons until they collide, which also results in a limited partition. Unlike previous approaches, we propose a simper partition by considering five types of primitives simultaneously rather than merely planar shapes. Moreover, we develop a propagation strategy inspired by Bauchet et al. [3] to further simplify and speed up the surface extraction.", + "bbox": [ + 75, + 263, + 472, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based reconstruction. With the advancements in deep learning, there have been several notable works that utilize learning-based techniques to reconstruct meshes. Voxel-inspired methods, such as those presented in references [38] and [10], train a network of models to convert voxels into implicit surface representations. To attain more accurate representations, SIREN [45] and DIGS [4] used signed distance functions as implicit surface representations, and then extracted meshes via marching cubes [30]. Points2Surf [12] employed a patch-based learning framework to directly attain meshes from point clouds without the need for normals, while DSE [40] adopted 2D Delaunay triangulations to ensure manifold meshes and BSP-Net [9] output compact polygonal meshes with arbitrary topology. However, most of these methods depend on implicit surface representations, which may not explicitly capture the underlying structures such as quadric primitives of objects. In addition, it is hard to perform model editing tasks for surfaces generated by implicit reconstructions.", + "bbox": [ + 75, + 542, + 472, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 844, + 210, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed method consists of three major steps as illustrated in Fig. 2. Given an input point cloud $\\mathcal{P} = \\{p_j\\in$", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbb{R}^3\\}_{j=1}^N$ (Fig. 2(a)), we first partition $\\mathcal{P}$ by a set of proxy meshes defined by the extracted geometric primitives and then apply a mesh arrangement operation to get candidate manifold patches (Fig. 2(b)) in Section 3.1. Later, a fast propagation mechanism is presented in Section 3.2 is used to prune patches that are inactive (Fig. 2(c)) to accelerate the subsequent surface extraction procedure. Finally, we deliver a manifold and watertight output surface (Fig. 2(d)) by formulating a patch-induced optimization under the BLP framework in Section 3.3.", + "bbox": [ + 496, + 263, + 893, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Space Partitioning", + "text_level": 1, + "bbox": [ + 500, + 426, + 679, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of this step is to generate a finite space partitioning by arranging proxy meshes of geometry primitives extracted from the input point cloud $\\mathcal{P}$ .", + "bbox": [ + 496, + 450, + 890, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Primitive extraction. We first utilize an efficient region growing strategy introduced in [39] to attain the rough segmentation components $\\mathcal{S} = \\{s_i\\subset \\mathbb{R}^3\\}_{i = 1}^M$ and the corresponding primitives $\\mathcal{Q} = \\{q_i\\subset \\mathbb{R}^3\\}_{i = 1}^M$ from $\\mathcal{P}$ , where $q_{i}$ is the supporting primitive of $s_i$ inversely, each point $p_j\\in s_i$ is the supporting point of $q_{i}$ . Let $\\mathcal{L} = (\\mathcal{S},\\mathcal{Q})$ be the initial layout of the extracted primitives, we further refine $\\mathcal{L}$ via the variational shape approximation based on the $\\ell^2$ and the $\\ell^{2,1}$ metrics [11, 47], which aims to optimize the following expression", + "bbox": [ + 496, + 515, + 890, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {L}} \\mathcal {F} (\\mathcal {L}) = \\sum_ {i = 1} ^ {M} \\sum_ {p _ {j} \\in s _ {i}} \\operatorname {d i s t} \\left(p _ {j}, q _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 678, + 890, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $dist(p_j, q_i)$ stands for the hybrid metric that measures the cost of the point $p_j \\in s_i$ with respect to the corresponding primitive $q_i \\in \\mathcal{Q}$ in terms of both the spatial $\\ell^2$ distance and the normal $\\ell^{2,1}$ deviation. Fig. 2(a) presents the primitive extraction result based on five types of primitives that we are currently considering. Noted that the proposed framework is not constrained by the above detected primitive types. Instead, it can be integrated with powerful detection algorithms that support a wider variety of primitive types, enabling more complex reconstructions with better accuracy.", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14173", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6ac4b1d6dc82edafb4c43521b167161c1151b6a4d8ac1901070154f791e74c6a.jpg", + "image_caption": [ + "Figure 3: Visual description of patches and (non-) boundary curves. (a) $\\mathcal{C}_{ij}$ is the non-boundary curve generated by patches $m_{i}$ and $m_{j}$ , while $\\bar{\\mathcal{C}}_i$ and $\\bar{\\mathcal{C}}_j$ are the boundary curves of $m_{i}$ and $m_j$ , separately. (b) For each non-boundary curve, there are four patches around it." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 457, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proxy mesh arrangement. After the optimization of the extracted primitives, we generate the discrete approximation to all primitives via proxy meshes. These meshes are trimmed by the bounding box of $\\mathcal{P}$ and then all meshes are processed by a mesh arrangement [50]. The mesh arrangement results are a set of candidate primitive patches $\\mathcal{M} = \\{m_i\\subset \\mathbb{R}^3\\}_{i = 1}^K$ as illustrated in Fig. 2(b), which partition the current working space into a series of cells that form a closed manifold without self-intersections. Here each patch $m_{i}$ represents a subset of the input meshes and is the maximal connected set of triangle meshes as illustrated in Fig. 3. Due to the connectivity of patches, each $m_{i}$ may embrace several non-boundary curves $\\mathcal{C}_{ij}$ (colored red in Fig. 3(a)) that are shared with other primitive patches, which are defined as", + "bbox": [ + 75, + 276, + 472, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {i j} = m _ {i} \\bigcap m _ {j} (i \\neq j), m _ {i}, m _ {j} \\in \\mathcal {M}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 508, + 468, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "whereas at most one boundary curve $\\bar{C}_i$ (colored green in Fig. 3(a)). For example, there are in total of four neighboring patches that share the same $C_{ij}$ in Fig. 3(b). Since we perform space partitioning based on primitive patches rather than discretizing each of them as multiple piecewise planes, the developed algorithm indeed is several orders of magnitude faster than previous plane-induced approaches. The efficiency superiority of our method is also quantitatively validated by the following experiments.", + "bbox": [ + 75, + 537, + 472, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pruning of Candidate Primitive Patches", + "text_level": 1, + "bbox": [ + 76, + 680, + 419, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is known that the computational complexity of the ultimate surface extraction in shape reconstruction regarding the size of patches is exponential [44]. To expedite the subsequent optimization process, we further propose a pruning mechanism which effectively circumvents the time-consuming and memory-intensive global collision detection to classify current primitive patches as active and inactive, from which we prune inactive ones to reduce the search space for surface extraction. As illustrated in Fig. 2, the significant time gap between the two routes(black and red) demonstrates the effectiveness of our pruning mechanism. The proposed procedure is composed of two steps, initialization and pruning, as presented subsequently.", + "bbox": [ + 75, + 703, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: Pseudo-code of the propagation process", + "text_level": 1, + "bbox": [ + 509, + 95, + 843, + 109 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Priority queue $Q$ of representative triangles.", + "bbox": [ + 519, + 112, + 859, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Active primitive patches $\\tilde{\\mathcal{M}}$", + "bbox": [ + 521, + 127, + 772, + 141 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1 Initialize propagation cost of each triangle as $c = 0$ .", + "bbox": [ + 506, + 142, + 862, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2 while $Q\\neq \\emptyset$ do", + "bbox": [ + 506, + 157, + 635, + 171 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 $\\mid (t,c)\\gets Q.pop()$ and $\\hat{t}\\leftarrow t$", + "bbox": [ + 506, + 172, + 735, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 for each triangle edge $e\\in \\hat{t}$ do", + "bbox": [ + 506, + 189, + 754, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5 for triangle $t_e\\cap t = e$ do", + "bbox": [ + 506, + 205, + 743, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6 | | if $\\text{prim}(t_e) == \\text{prim}(t) \\& t_e$ is", + "bbox": [ + 506, + 220, + 818, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "inactive & $t_e$ is not blocked then", + "bbox": [ + 506, + 234, + 823, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "7 Compute the height $h_e$ of $t_e$ on $e$ .", + "bbox": [ + 506, + 251, + 844, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "8 $Q$ .push((te,c+he)).", + "bbox": [ + 506, + 266, + 769, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initialization. To describe the point distribution of the input point cloud as fidelity as possible, we first select a set of representative triangle meshes denoted as $\\mathcal{I}$ from $\\mathcal{M}$ by means of the triangle coverage $Cover(t)$ which is defined as", + "bbox": [ + 496, + 330, + 893, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C o v e r} (t) = \\frac {\\operatorname {V a l i d} (t)}{\\operatorname {A r e a} (t)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 402, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\text{Area}(t)$ stands for the area of a triangle mesh $t$ and $\\text{Valid}(t)$ is formulated as", + "bbox": [ + 498, + 439, + 893, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV a l i d (t) = \\sum_ {p \\in \\mathcal {P} _ {t}} 1 (\\| p - p ^ {\\prime} \\| _ {2} < \\epsilon) a _ {p}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 478, + 890, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_t\\subset \\mathcal{P}$ encompasses the measurement points that their projections are attached to $t$ $\\| p - p^{\\prime}\\|_{2}$ denotes the $\\ell^2$ Euclidean distance from $p$ to its projective point $p^\\prime$ on $t$", + "bbox": [ + 498, + 520, + 890, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1 is the indicator function, and $a_{p}$ is the corresponding geodesic Voronoi area of $p$ , which is calculated as in the inset: $p$ and its $k$ -nearest neighbors (i.e., hard red points) are first projected to the supporting plane $\\Omega_t$", + "bbox": [ + 498, + 566, + 694, + 686 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/15deff45d25e82bb7cc5e77f6ad8d2ee4b50c41adac1751201a339de7ca3b510.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 566, + 862, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where its attached triangle mesh $t$ lies on, then $a_{p}$ is equivalent to the area of the 2D Voronoi cell [2]. $t$ is termed representative if its Cover $(t)$ surpasses a user-specified value. We populate all representative meshes of $\\mathcal{I}$ into a priority queue $Q$ with an ascending order.", + "bbox": [ + 496, + 686, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pruning. Next, representative triangle meshes are propagated to enable a limited partition to speedup the final surface extraction. To this end, all triangles of $Q$ are initially labeled as inactive prior to entering the loop of the priority queue. For each round during the loop, we first pop the top element, for instance, $(t_i, c_i)$ , out of the priority queue and prompt $t_i$ to be an active state $\\hat{t}_i$ . The propagation cost $c_i$ of a triangle mesh $t_i$ is defined as the height", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14174", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$h_i$ from its vertex to the base edge shared with another adjacent triangle. Sequentially, we propagate all the triangles that share both the same edge and the same supporting primitive with $t_i$ and pop the update elements out of $Q$ . For instance, for edges that are not located on non-boundary curves, there are at most three update elements $(t_{i1},c_i + h_1)$ , $(t_{i2},c_i + h_2)$ , $(t_{i3},c_i + h_3)$ , whereas, for edges that coincide non-boundary curves, we perform a fast local collision detection to check whether there exist active meshes, if yes then we evoke barricades and stop the propagation process. The above propagation proceeds until $Q$ becomes empty. We summarize the proposed propagation process in Algorithm 1 to make it more understandable. Consequently, a primitive patch is active only if its triangles are all active, otherwise, it is still inactive. We use $\\tilde{\\mathcal{M}}$ to represent the set of active patches. The patches in Fig. 2(b) after pruning are shown in Fig. 2(c). As observed, the size of candidate patches has been significantly reduced compared with Fig. 2(b).", + "bbox": [ + 75, + 90, + 472, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Like [3], we also allow for the propagation crossing primitive patches $\\mathcal{G}$ times (with $\\mathcal{G} = 2$ in default) at most. This can deal with small parts that may have been missed during the primitive detection. However, in contrast to previous global collision detection manners [3], our local strategy preserves much better the propagation efficiency.", + "bbox": [ + 75, + 378, + 470, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Surface Extraction", + "text_level": 1, + "bbox": [ + 76, + 479, + 261, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we adopt a Binary Linear Programming (BLP) analogous to [35] to extract the surface model from $\\tilde{\\mathcal{M}}$ . Specifically, the state of patches in $\\tilde{\\mathcal{M}}$ is encoded as $\\mathcal{X}_{\\tilde{M}} = \\{x_{\\tilde{m}}|\\tilde{m}\\in \\tilde{\\mathcal{M}},x_{\\tilde{m}}\\in \\{0,1\\} \\}$ , where the binary discrete variable $x_{\\tilde{m}}$ indicates whether the patch $\\tilde{m}$ is selected $(x_{\\tilde{m}} = 1)$ or not $(x_{\\tilde{m}} = 0)$ . To ensure the resulting surface manifold and watertight, we constrain the curves of the selected patches to be all non-boundary, and thus the two patches sharing the same non-boundary curve are either selected at the same time or both not selected. Consequently, our objective function is defined as", + "bbox": [ + 75, + 502, + 472, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\mathcal {X} _ {\\tilde {\\mathcal {M}}}} \\mathbf {U} (\\mathcal {X} _ {\\tilde {\\mathcal {M}}}) + \\lambda \\mathbf {B} (\\mathcal {X} _ {\\tilde {\\mathcal {M}}}), \\\\ s. t. \\left\\{ \\begin{array}{l} \\sum_ {\\tilde {m} _ {i} \\cap \\tilde {m} _ {j} \\neq \\emptyset , i \\neq j} \\left(x _ {\\tilde {m} _ {i}} + x _ {\\tilde {m} _ {j}}\\right) = 0 o r 2, \\tilde {m} _ {i}, \\tilde {m} _ {j} \\in \\tilde {\\mathcal {M}} \\\\ x _ {\\tilde {m} _ {i}} \\in \\{0, 1 \\}, \\quad i = 1, 2, 3 \\dots | \\tilde {\\mathcal {M}} |. \\end{array} \\right. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 681, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{U}(\\mathcal{X}_{\\tilde{\\mathcal{M}}})$ and $\\mathbf{B}(\\mathcal{X}_{\\tilde{\\mathcal{M}}})$ are penalty terms regarding the patches and the neighboring patch pairs, respectively, and $\\lambda$ is a balance coefficient. Moreover $\\mathbf{U}(\\mathcal{X}_{\\tilde{\\mathcal{M}}})$ encourages the selection of active patches that are with high fidelity in the form of", + "bbox": [ + 75, + 779, + 470, + 853 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} \\left(\\mathcal {X} _ {\\tilde {M}}\\right) = \\sum_ {\\tilde {m} \\in \\tilde {\\mathcal {M}}} x _ {\\tilde {m}} \\cdot \\left(\\frac {\\mathcal {A} _ {\\tilde {m}} - \\bar {\\mathcal {A}} _ {\\tilde {m}}}{\\mathcal {A}} - \\frac {| \\mathcal {P} _ {\\tilde {m}} |}{| \\mathcal {P} |}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 862, + 468, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\mathcal{A}_{\\tilde{m}}$ and $\\bar{\\mathcal{A}}_{\\tilde{m}}$ denote the area and the coverage area of the patch $\\tilde{m}$ , respectively. The coverage area of $\\tilde{m}$ is the summation of the coverage area of all its triangles. $\\mathcal{A}$ is a normalization factor equivalent to the sum of the area of all patches. $\\mathcal{P}_{\\tilde{m}}$ is the set of points whose projections are locating in $\\tilde{m}$ . On the other hand, $\\mathcal{B}(\\tilde{\\mathcal{X}})$ is designed to control the model complexity. Specifically, it put penalties to the non-boundary curve length of two patches:", + "bbox": [ + 496, + 90, + 893, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} \\left(\\mathcal {X} _ {\\tilde {\\mathcal {M}}}\\right) = \\frac {1}{L} \\sum_ {i, j} \\left| \\mathcal {C} _ {i j} \\right| \\cdot x _ {\\tilde {m} _ {i}} \\cdot x _ {\\tilde {m} _ {j}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 220, + 890, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|\\mathcal{C}_{ij}|$ denotes the length of the curve shared by two active patches $\\tilde{m}_i$ and $\\tilde{m}_j$ with different supporting primitives and $L$ is the sum of the length of all curves. Fig. 2(d) presents the final reconstruction surface by BLP, which not only generates a manifold and watertight shape but also successfully preserves the intrinsic primitive structures.", + "bbox": [ + 496, + 268, + 893, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Evaluations and Discussions", + "text_level": 1, + "bbox": [ + 498, + 373, + 879, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details and metrics. In this section, we execute extensive experiments to validate the advantages of the developed method in terms of both reconstruction precision and efficiency and make various comparisons with state-of-the-art approaches. Like [3], the precision is quantitatively evaluated by the symmetric mean Hausdorff distance (SMH) between the measurement point clouds and the reconstruction meshes (values are normalized by the diagonal length of the models' bounding box and multiplied by a factor of $10^{2}$ ), while the efficiency is assessed by the running time. Our algorithm is implemented in $\\mathrm{C}++$ using the geometry processing library libigl [22]. We adopt the mathematical programming solver Gurobi [18] to optimize Eq. 5. All reconstruction results are rendered by Blender [14] for better visualization. Our experiments are conducted on a desktop with Intel(R) Core i7-7700K CPU@4.2GHz and 48GB RAM.", + "bbox": [ + 496, + 398, + 893, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction effect. Our algorithm has demonstrated good reconstruction results from point clouds to manifold meshes on various CAD objects as illustrated in Fig. 4. The input point clouds presented in Fig. 4(a)-(h) are either sampled from real-world CAD models or courtesy by publicly available benchmark datasets [41, 51]. As observed, there are multiple explicit intrinsic primitive structures in each model. Instead of approximating them directly by piecewise planar primitives, our algorithm is capable of reconstructing the intrinsic primitives faithfully even for quite challenging or complicated models. For instance, the rich spherical and conic structures existing in the Minarets model of Fig. 4(b) require careful treatments, yet our method is still able to achieve highly accurate reconstructions while preserving all structural details successfully. At the same time,", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14175", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f85af81be8a05e864169d8870b1cdc1767ece2c194aa1ae05a6250ad1a66fc70.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 143, + 84, + 230, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f373b54946d17a8c3aa6de1cd72d62257d78d03eb262a5013d9d7febdc1c7f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 83, + 392, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08daddc0b9859a06af6152337881d56f1d858844efd2f5493644e5cc840026e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 84, + 542, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa0d88f40c136778575bf627483cf752b758c205d64cd5f76eb2ae23e575ba9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 84, + 700, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/88cbd7dd0e889fd8b501f761eb9ba5c5ba36a5c012f3a3c036d6cca41db981e1.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 137, + 162, + 230, + 227 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f8019e35731c67bcde3cbb1b0d030957d555c91a1550d07038a036b6caca6531.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 299, + 166, + 393, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa73e1fb03136d1e332835c4433b54b3d69fdfdccc95592b71872df16d17d139.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 166, + 547, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2967a618cca8641a9a671c62701495475ea478c6cf13309a5ac2fe73926b236d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 171, + 705, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7e9e70d971760a48c435f4320ec6f96082a8e59725a74e52b7543f7c28c329b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 85, + 866, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/58932b0e464b7fcf3484769d1081aff2693565d0118a5f2875dbd0b7ea8dc4a9.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 130, + 244, + 235, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/182f55ad874a7419534daea35c046eae52a9f86c14ef3068f5c954011027bf58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 244, + 397, + 296 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9bfba2f70dcbda172919956545615a66128738877a130fc1a740842026df6908.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 244, + 550, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae38cdbf35efdaf43afbb707680612d62c99a5f93b040f2912a10443e94b82d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 247, + 712, + 299 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3c329c267ec47063e992bd864771e3465e1aa6a8df30eedda3086e47ed88930.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 171, + 867, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d41c99358da7634afc561c82f3d35bad3672589582364e469b29bed4ee5100f0.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 135, + 310, + 236, + 373 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55e5bf41ae7a5bf3fcb59e7da57f8f2565ccd0c7a12d8442fb45adc51c4cb425.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 310, + 395, + 373 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b877af64b5cf1e89128997cc67c3d9b04f45b6e0c5c6721b107c41e3f1a7fa2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 310, + 555, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bbba14f044722479a8c9f85e463b731a4e77664879af5d56a192a188485b5353.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 309, + 715, + 373 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6d24df649ec78854124ab95822f905269b0edeb2fd56d1ee27473da28aba25b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 248, + 869, + 299 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/789af587d81aa0e8acd3875b3a72498c8546ee6609f09e567fd03df8c67ec8cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 309, + 870, + 372 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb71a968918d026e27e4d439b334cf9f711c9b7f82a4ea4b7e967c931a2930df.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 138, + 392, + 241, + 465 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2670d38ec9e351c5fa434d4ed125ba73169ddfb1e708e7d34eb6d46300584b92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 398, + 395, + 470 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8ac5bba0b0cde9dbd34fd7404f1990a973222237f5e4f11c73d806e83739a1f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 397, + 560, + 470 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5eac263144508524a4b1ccad2874299b45c7bd5bd3370ba8c34c8b1b0521b97d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 398, + 727, + 473 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95fe49212056e029bd1108c91c4a2072d1e36bf6d2205edc1702f33ea129d969.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 398, + 872, + 473 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a03926641a3d0832715b029625d1d4592fdd2c8e55e7126c81938df299e8fef3.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 143, + 479, + 233, + 556 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a61610d427f31e424114c0c76f3bb01bf67e4529026fb108b01b901db0e7ab4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 481, + 385, + 556 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f7d4e3ff452e045f7f6588caa4dc6a6d03466b11111c6bdb7a309839d624a696.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 483, + 547, + 559 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6fbb3642d226ad5c1b315b86ca1958524619e8c2c8e03efc01e37deee24bd3ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 484, + 720, + 561 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b231689cc0488831320d8e5555754c1f19b89ad49c20cc851eeab05f8428c507.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 484, + 866, + 560 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa9df84afdc972e8e5a3f6861e6bdf37f5df4c62fc5d57a45893efe226b96885.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 142, + 574, + 230, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/93779c0fca2d7581e66860f73101053460b52a9f5fa58c1d5cc83867622bdb0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 574, + 383, + 666 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f0b6540f330e6f9043af10d05208d0436df583f5e6472784079f4fe218b9435.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 573, + 544, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1a2123c3fd365fe3e9bd4dc2a81c4f995abf803da08c3808feb79365aedca25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 573, + 720, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ebc44a8c908f54c0a7a4f1f55da6f582985410a9d1484708827731343727235.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 573, + 870, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/edcebd04c2c6e64671b61ed83c46a8ff1a92207e917aebc255e689f9991b7ff9.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 142, + 695, + 251, + 732 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e9640d3f5591aef29042483bcaefdd2752c5ed22091eb3affcc66b03deb37ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 691, + 400, + 729 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82627de130599f1c3fdb5149160845fb9c732551b22bf41aaa525932716c96e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 691, + 560, + 729 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a2b238be5790533745a1c14ee1f21b69d2a1a17226afba8eee161e699824a390.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 688, + 728, + 726 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c5a75d12d6f8e95d46384ec27de0fe94019fe0b0897b5a88124fa29c459d83c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 684, + 883, + 720 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f117fa0e6923e2a02ea3430a2447ef94d4f34190683cb23ec6b64bae8813e50e.jpg", + "image_caption": [ + "Figure 4: Reconstruction results by the proposed method. From left to right: Input measurement point clouds, extracted geometric primitives (different colors stand for different primitives), output reconstruction models, recovered sharp geometry along with the original point clouds, and colored models showing the error distribution regarding the Hausdorff distance. Our method produces high-quality reconstructions and preserves the intrinsic primitive structures as well as the sharp features well." + ], + "image_footnote": [], + "bbox": [ + 782, + 727, + 880, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our method produces high-quality sharp feature lines attained by patch intersections, as shown in the forth column of Fig. 4. This is an important advantage for many practi", + "bbox": [ + 75, + 848, + 468, + 893 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cal applications such as sketch-style rendering, as it ensures that the reconstructed object retains its original shape and does not appear distorted or blurry. The fifth column of", + "bbox": [ + 498, + 848, + 893, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14176", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/974fcc9e3b019c9e70f5a3cc51179a7771d7913be20cf1ed89fa759b35259810.jpg", + "table_caption": [ + "Table 1: Statistical results of the quantity and the runtime (in second) during reconstruction of the CAD models in Fig. 4." + ], + "table_footnote": [], + "table_body": "
Models Metrics(a)(b)(c)(d)(e)(f)(g)(h)
# Points of the input point cloud10k593k10k10k74k7k16k14k
# Extracted primitives2853212034322927
# Generated candidate patches36466024103072349704177501092
# Active patches after pruning224533697414172586342472667
Patch decrease (%)38.42644.07328.05842.32447.96817.98637.06738.919
# Determined active patches of the output model605915191935339986114
Mesh arrangement (s)7.85321.7522.4631.7275.2362.9355.4311.421
Patch pruning (s)0.92715.4210.7920.4763.0430.4881.1310.726
Surface extraction (s)1.2882.0860.6590.0911.1120.1570.3140.137
", + "bbox": [ + 91, + 94, + 874, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fa71908fe31c1fc14a0f0b0e2716f8dba99fccb06c04cd220df91c4e9d9c12e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 270, + 169, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aea85b175c3958133c9f8ab856700231ba14dca5949a652fa66afa353d2f86fd.jpg", + "image_caption": [ + "$\\sigma = 0.01(0.82m)$ $\\mathrm{SMH} = 0.11\\%$" + ], + "image_footnote": [], + "bbox": [ + 84, + 325, + 169, + 377 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9976b643d3d0b5c3f807bf2a9af81e9c65abd30aec58072e9bbdd3faea7d0b5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 187, + 270, + 272, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8cf39710a210f9d567ee3cb24177cc45abd17a4083783cc255105801ac8120be.jpg", + "image_caption": [ + "$\\sigma = 0.02$ $(1.64\\mathrm{m})$ $\\mathrm{SMH} = 0.11\\%$" + ], + "image_footnote": [], + "bbox": [ + 187, + 324, + 267, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/73ac6ed6fee6cbff3a9e0c9fb855f1dd847eb0178e1f765678777f0be2da5f34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 270, + 377, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe7226a44429187dbb1cc7c4e190a4e6d323731e387c6f4c40e0ffa80cb62fd8.jpg", + "image_caption": [ + "Figure 5: Robustness test against noise disturbance. Top: Input point clouds with increasing noise. Bottom: Our reconstruction results along with the noise level $\\sigma$ and the reconstruction error SMH." + ], + "image_footnote": [], + "bbox": [ + 294, + 324, + 377, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f0f924ea69fc10392ed3f4224354d9bced3241286753adbacffcb9d902bd949.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 268, + 480, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa2f83ddf29487e253687ad986a039ea7c395d57eca64814dd67e8a7218c560c.jpg", + "image_caption": [ + "$\\sigma = 0.04(3.28m)$ $\\mathrm{SMH} = 0.12\\%$", + "$\\sigma = 0.08(6.56m)$ $\\mathrm{SMH} = 46\\%$" + ], + "image_footnote": [], + "bbox": [ + 395, + 325, + 478, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Statistical results. Table 1 summarizes the quantitative results of our proposed method for the models in Fig. 4. Our algorithm significantly reduces candidate patches, especially for complex models like Fig. 4(b),(d),(e), where the patch decrease ratio exceeds $40\\%$ . In terms of efficiency, patch pruning and surface extraction outperform mesh arrangement, and further improvements can be made by exploring a more efficient intersection mechanism.", + "bbox": [ + 75, + 563, + 468, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness assessment. We also assess the ability of the proposed algorithm against noise disturbance and occlusion data. As depicted in Fig. 5, we increase the standard deviation $\\sigma$ of Gaussian noise from $0.01(0.82m)$ to $0.04(3.28m)$ , where values in parentheses are real intensity relative to the bounding sphere's radius of the model. Our method consistently generates accurate reconstructions in terms of both topological structures and approximation distance (SMH), suggesting its high resilience to noise. Even when $\\sigma = 0.08(6.56m)$ , which has caused significant topological alteration of the original point cloud, our method still maintains the overall structures of the object. Additionally, we verify the robustness of our method to occlusion", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe77783fb2c125a78cfeb45ea320f2e49a9a7419540611125ca4edd63f3e197b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 270, + 578, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56f2a99d447e3865ab2fe4a231085d6259bc7c207e801d8d8b6df1262ed1b405.jpg", + "image_caption": [ + "Fig. 4 visualizes the reconstruction error of our algorithm with respect to the Hausdorff distance, where our method consistently produces accurate approximations for all objects, demonstrating its stability and effectiveness.", + "Figure 6: Robustness test against occlusion data. Top: Input point clouds with increasing occlusions. Bottom: Our reconstruction results along with the reconstruction error." + ], + "image_footnote": [], + "bbox": [ + 511, + 337, + 578, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c7edd5ef763cd9cbd1e758fa3c09cf01fc1eddeb94e0ec0a78411df7ad33d17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 270, + 671, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56b8ea4ef2dd450b6413795466611f6277dfa45a64b253611883ce0dda4b10fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 337, + 673, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d999a2334889471094c890a426e2776c6ab8cc66af128e1ac2dcf34fa99effc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 270, + 772, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/160dd7fbaaa67577e88f48d686662c89e0fa3018345e1bea6c1ff9bbcec27368.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 334, + 776, + 406 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f562a790c01adba344ce0c9f25dbb7ff8244f1057c1c2b8f272634c9093f976.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 270, + 872, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd5e33066618cdcfb154a6629f2b116b6be67e3537510ab2ad7a912269160573.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 334, + 877, + 406 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a86d988ba0bcd37f211a107147f0e7a89974730dbc35224d598ad9b9535099d4.jpg", + "table_caption": [ + "Table 2: Comparisons with state-of-the-art approaches, where \"w\" and \"w/o\" indicate with and without the patch pruning, separately." + ], + "table_footnote": [], + "table_body": "
MetricsiPSR[20]SIREN[45]DSE[40]KSR[3]Ours
ww/o
SMH (%)4.2341.793.15169.442.452.45
Time (s)201.5891208.124303.865570.46845.556.777
", + "bbox": [ + 535, + 518, + 854, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sion data. Fig. 6 reports our reconstruction results after progressively erasing point clouds. As observed, our method shows remarkable resistance to occlusions and has successfully compensated the missing structures. Conversely, this is also quite helpful to complete the imperfect point clouds by sampling points in the reconstruction models.", + "bbox": [ + 496, + 577, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons. Subsequently, we conduct a comparative analysis of our method with representative state-of-the-art approaches, including densely implicit reconstruction methods iPSR [20], DSE [40], and SIREN [45], as well as the plane-assembly approach KSR [3]. The quantitative comparison results in terms of reconstruction accuracy and efficiency on the Cheese model is presented in Table 2. The results demonstrate that our method achieves competitive reconstruction accuracy with implicit approaches iPSR and DSE, and outperforms SIREN and KSR by a large margin. Additionally, our algorithm performs significantly faster than competitors, even without patch pruning. The qualitative reconstruction results presented in Fig. 7 show that benefiting from the introduced primitive assembly, our", + "bbox": [ + 496, + 689, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14177", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa05c046cb730132e2746dd8aaa5698e2d6ba0543347c305a725933bada9fdde.jpg", + "image_caption": [ + "Figure 7: Qualitative comparisons with state-of-the-art approaches. Our method enables the shape reconstruction faithfully and successfully recovers the sharp geometry (e.g., feature lines and corners) from unorganized point clouds." + ], + "image_footnote": [], + "bbox": [ + 101, + 75, + 205, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e125e2ff80a78fcad8357fde7b939c0d75b2578658b2e5636d226f53b4fdb894.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 75, + 339, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e30e0715d4535abacf4766dabea16958dec36614113714d703520e56b961ec08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 75, + 472, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ff753b28d8961edc3e10ec28c5a722c86809e864e443e4df95e37e8f16b0ea5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 75, + 604, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6f528e3486add7164707d83501c092377ac2a1d3b2c7a095445b4310010408b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 75, + 728, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d305bfbd766ef046c7b2e150e11aa61e6e9a8d3b232cd17b767ed0b0b8adb2ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 75, + 867, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3a78679cd70373e09992b1063403b4dd171bd812edc2f7d279da315844945fe8.jpg", + "table_caption": [ + "Table 3: Quantitative comparisons on the ABC300 dataset." + ], + "table_footnote": [], + "table_body": "
Methods MetricsiPSR[20]SIREN[45]DSE[40]KSR[3]ComplexGen[17]CAPRI-Net[48]Ours
SMH (%)2.7124.894.0626.4324.2415.874.81
Time (s)24.0641565.21732.8004.208553.089183.8443.728
", + "bbox": [ + 80, + 303, + 468, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method successfully recovers sharp features from unorganized point clouds, which is a salient advantage compared to implicit methods.", + "bbox": [ + 75, + 354, + 468, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also conduct quantitative comparisons with several predecessors (additionally including CAPRI-Net [48] and ComplexGen [17]) on 300 CAD models from the ABC dataset (abbreviated as ABC300). Table 3 demonstrates that our method achieves consistently competitive reconstruction accuracy with implicit approaches such as iPSR and DSE (while also providing advantages for sharp feature recovery and model editing), and significantly surpasses KSR, CAPRI-Net, and ComplexGen. Furthermore, our method is considerably faster than its competitors.", + "bbox": [ + 75, + 398, + 470, + 550 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57819ac51dffa0a3feb27fb698e57b63453f2902e65d4a29efa06c8dd692a768.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 583, + 158, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/440759cec00bc591d814faf9e4ee642d43fcd342b2701449ebd78f4e3e90166a.jpg", + "image_caption": [ + "Figure 8: Model editing based on our structure-aware reconstruction algorithm. (a) The original reconstruction object. (b)-(d) The newly edited results." + ], + "image_footnote": [], + "bbox": [ + 179, + 565, + 256, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/187111fa2dadba5ad3b66c877006ebfc6b0f0060d7828af5bb1df6f0533aaced.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 563, + 361, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e579ee14a8f29a2453a7ac2bf5980b79cca03347e6145953e8a4331160d3c12e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 551, + 468, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Balance parameter. We investigate the impact of $\\lambda$ on surface extraction by varying its values to different settings (see Fig. 9(a)). It can be seen that increasing $\\lambda$ brings the reduction of the complexity of the output model. However, this also degrades the reconstruction accuracy. For example, when $\\lambda$ is set to 0.7, the inner hollows of the input point clouds disappear.", + "bbox": [ + 75, + 700, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Application to model editing. Our reconstruction method can be employed for flexible model editing by taking advantage of the structure-aware merits. As shown in Fig. 8, since each quadratic structure of the object is preserved along with their detailed geometric parameters, one", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4df7b8eafaa7e27a489eeaaa56cd3e2e193162ae3329ba3a0911fe85537a998e.jpg", + "image_caption": [ + "(a)", + "Input", + "Figure 9: (a) Impact of parameter $\\lambda$ . As $\\lambda$ is increased, the complexity of the output model decreases. (b) In sparse input scenarios, small primitives can be easily lost, leading to failures." + ], + "image_footnote": [], + "bbox": [ + 509, + 306, + 588, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0569f070d0f6812f0c8582ba61da016e8eae8366895c3e07ad94a5656a4cb85e.jpg", + "image_caption": [ + "$\\lambda = 0.2$" + ], + "image_footnote": [], + "bbox": [ + 604, + 309, + 676, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/05474ab7c9720fa731cd3b47c9a8d6f53ad8eb32e3e4c51809f4e45808322281.jpg", + "image_caption": [ + "$\\lambda = 0.7$" + ], + "image_footnote": [], + "bbox": [ + 696, + 309, + 767, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5267e588ffc774d6484e2e637d88e441f6b8adc46dc6f3dbaea560c34f7a991d.jpg", + "image_caption": [ + "#Point=10k" + ], + "image_footnote": [], + "bbox": [ + 781, + 303, + 906, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "can modify the reconstruction shape by simply specifying various parameters. For instance, by re-defining the radius or height of the cylindrical primitives in Fig. 8(a), we can resize them to our desirable shapes (Fig. 8(b)-(d)).", + "bbox": [ + 496, + 441, + 890, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Our approach primarily focuses on the rapid and accurate reconstruction of representative primitive structures in CAD models. However, as shown in Fig 9(b), our reconstruction quality may deteriorate for inputs lacking explicit primitives. Additionally, inadequate primitive extraction due to severe occlusion or noise can degrade the quality of the reconstruction results.", + "bbox": [ + 496, + 523, + 890, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 647, + 767, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose an accurate and fast shape reconstruction pipeline for converting disordered 3D point clouds into manifold, watertight surfaces. Our method preserves representative intrinsic primitive structures in CAD systems through primitive assembly, avoiding dense implicit reconstructions and extending piecewise planar approximations. We also develop an effective patch pruning mechanism, greatly expediting surface extraction by reducing patch intersections. Additionally, our method is highly resilient to noise and can reliably recover sharp object geometry. Extensive experiments show our algorithm's superiority over competing methods.", + "bbox": [ + 496, + 672, + 890, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our future work, we plan to expand the current primitive types to general primitives, and even spline surfaces, in order to efficiently process more complex CAD models.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14178", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 78, + 90, + 243, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially funded by the National Key Research and Development Program (2021YFB1715900), the CAS Project for Young Scientists in Basic Research (YSBR-034), the National Natural Science Foundation of China (62172415, 62272277, 12022117), and the HKU-100 Research Award.", + "bbox": [ + 76, + 114, + 468, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 220, + 171, + 234 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Murat Arikan, Michael Schwärzler, Simon Flöry, Michael Wimmer, and Stefan Maierhofer. O-snap: Optimization-based snapping for modeling architecture. ACM Trans. Graph., 32(1):6:1-6:15, 2013. 2", + "[2] Gavin Barill, Neil G Dickson, Ryan Schmidt, David IW Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. ACM Trans. Graph., 37(4):43:1-43:12, 2018. 4", + "[3] Jean-Philippe Bauchet and Florent Lafarge. Kinetic shape reconstruction. ACM Trans. Graph., 39(5):156:1-156:14, 2020. 1, 2, 3, 5, 7, 8", + "[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3", + "[5] Matthew Berger, Andrea Tagliasacchi, Lee M Seversky, Pierre Alliez, Gael Guennebaud, Joshua A Levine, Andrei Sharf, and Claudio T Silva. A survey of surface reconstruction from point clouds. Comput. Graph. Forum, 36(1):301-329, 2017. 1, 2", + "[6] Alexandre Boulch, Martin de La Gorce, and Renaud Marlet. Piecewise-planar 3d reconstruction with edge and corner regularization. Comput. Graph. Forum, 33(5):55-64, 2014. 2", + "[7] Anne-Laure Chauve, Patrick Labatut, and Jean-Philippe Pons. Robust piecewise-planar 3d reconstruction and completion from large-scale unstructured point data. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1261-1268, 2010. 1, 2", + "[8] Jie Chen and Baoquan Chen. Architectural modeling from sparsely scanned range data. Int. J. Comput. Vis., 78(2):223-236, 2008. 2", + "[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 3", + "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5939-5948, 2019. 3", + "[11] David Cohen-Steiner, Pierre Alliez, and Mathieu Desbrun. Variational shape approximation. ACM Trans. Graph., 23(3):905-914, 2004. 1, 3", + "[12] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In Eur. Conf. Comput. Vis., pages 108-124, 2020. 3", + "[13] Hao Fang and Florent Lafarge. Connect-and-slice: an hybrid approach for reconstructing 3d objects. In IEEE Conf." + ], + "bbox": [ + 78, + 244, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Comput. Vis. Pattern Recog., pages 13490-13498, 2020. 1, 2, 3", + "[14] Blender Foundation. Blender - a 3D modelling and rendering package. http://www.blender.org.5", + "[15] Leonidas Guibas. Kinetic data structures. In Handbook of Data Structures and Applications, pages 377-388. 2018. 3", + "[16] Leonidas Guibas, Lyle Ramshaw, and Jorge Stolfi. A kinetic framework for computational geometry. In IEEE Annual Symposium on Foundations of Computer Science, pages 100-111, 1983. 3", + "[17] Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. Complexgen: Cad reconstruction by b-rep chain complex generation. ACM Trans. Graph., 41(4):129:1-129:18, 2022. 8", + "[18] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual. https://www.gurobi.com.5", + "[19] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. In Proc. ACM SIGGRAPH, pages 71-78, 1992. 1", + "[20] Fei Hou, Chiyu Wang, Wencheng Wang, Hong Qin, Chen Qian, and Ying He. Iterative poisson surface reconstruction (ipsr) for unoriented points. ACM Trans. Graph., 41(4):128:1-128:13, 2022. 7, 8", + "[21] Hui Huang, Dan Li, Hao Zhang, Uri Ascher, and Daniel Cohen-Or. Consolidation of unorganized point clouds for surface reconstruction. ACM Trans. Graph., 28(5):1-7, 2009. 1", + "[22] Alec Jacobson and Daniele Panozzo. Libigl: Prototyping geometry processing research in c++. In SIGGRAPH Asia 2017 courses, pages 1-172. 2017. 5", + "[23] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3d data. Comput. Graph. Forum, 38(1):167-196, 2019. 2", + "[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Comput. Graph. Forum (Proceedings of SGP), page 61-70, 2006. 1", + "[25] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. Graph., 32(3):29:1-29:13, 2013. 1", + "[26] Florent Lafarge and Pierre Alliez. Surface reconstruction through point set structuring. Comput. Graph. Forum, 32(2pt2):225-234, 2013. 2", + "[27] Minglei Li, Liangliang Nan, Neil Smith, and Peter Wonka. Reconstructing building mass models from UAV images. Comput. Graph., 54:84-93, 2016. 1", + "[28] Minglei Li, Peter Wonka, and Liangliang Nan. Manhattanworld urban reconstruction from point clouds. In Eur. Conf. Comput. Vis., pages 54-69, 2016. 1", + "[29] Yangyan Li, Xiaokun Wu, Yiorgos Chrysathou, Andrei Sharf, Daniel Cohen-Or, and Niloy J Mitra. Globfit: Consistently fitting primitives by discovering global relations. ACM Trans. Graph., 30(4):52:1-52:12, 2011. 2", + "[30] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proc. ACM SIGGRAPH, volume 21, pages 163-169. 3" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14179", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] David Marshall, Gabor Lukacs, and Ralph Martin. Robust segmentation of primitives from range data in the presence of geometric degeneracy. IEEE Trans. Pattern Anal. Mach. Intell., 23(3):304-314, 2001. 2", + "[32] Aron Monszpart, Nicolas Mellado, Gabriel J Brostow, and Niloy J Mitra. Rapter: rebuilding man-made scenes with regular arrangements of planes. ACM Trans. Graph., 34(4):103:1-103:12, 2015. 2", + "[33] Claudio Mura, Oliver Mattausch, and Renato Pajarola. Piecewise-planar reconstruction of multi-room interiors with arbitrary wall arrangements. Comput. Graph. Forum, 35(7):179-188, 2016. 2", + "[34] Liangliang Nan, Andrei Sharf, Hao Zhang, Daniel Cohen-Or, and Baoquan Chen. Smartboxes for interactive urban reconstruction. ACM Trans. Graph., 29(4):93:1–93:10, 2010. 1", + "[35] Liangliang Nan and Peter Wonka. Polyfit: Polygonal surface reconstruction from point clouds. In IEEE Int. Conf. Comput. Vis., pages 2353-2361, 2017. 1, 2, 3, 5", + "[36] Anh Nguyen and Bac Le. 3d point cloud segmentation: A survey. In IEEE Int. Conf. Robot. Autom. Mechatron, pages 225-230, 2013. 1", + "[37] Sven Oesau, Florent Lafarge, and Pierre Alliez. Planar shape detection and regularization in tandem. Comput. Graph. Forum, 35(1):203-215, 2016. 2", + "[38] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 165-174, 2019. 3", + "[39] Tahir Rabbani, Frank Van Den Heuvel, and George Vosselmann. Segmentation of point clouds using smoothness constraint. ISPRS J. PhotoGramm., 36(5):248-253, 2006. 2, 3", + "[40] Marie-Julie Rakotosaona, Paul Guerrero, Noam Aigerman, Niloy J. Mitra, and Maks Ovsjanikov. Learning delaunay surface elements for mesh reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22-31, June 2021. 3, 7, 8", + "[41] Chiara Romanengo, Andrea Raffo, Yifan Qie, Nabil Anwer, and Bianca Falcidieno. Fit4cad: A point cloud benchmark for fitting simple geometric primitives in cad objects. Comput. Graph., 102:133-143, 2022. 5", + "[42] Falko Schindler, Wolfgang Wörstner, and Jan-Michael Frahm. Classification and reconstruction of surfaces from point clouds of man-made objects. In IEEE Int. Conf. Comput. Vis. Worksh., pages 257-263, 2011. 2", + "[43] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient ransac for point-cloud shape detection. Comput. Graph. Forum, 26(2):214-226, 2007. 2", + "[44] Gerard Sierksma and Yori Zwols. Linear and integer optimization: theory and practice. 2015. 4", + "[45] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Int. Conf. Neur. Info. Process. Sys., volume 33, pages 7462-7473, 2020. 3, 7, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Marc Van Kreveld, Thijs Van Lankveld, and Remco C. Veltkamp. On the shape of a set of points and lines in the plane. Comput. Graph. Forum, 30(5):1553-1562, 2011. 2", + "[47] Dong-Ming Yan, Wenping Wang, Yang Liu, and Zhouwang Yang. Variational mesh segmentation via quadric surface fitting. Computer-Aided Design, 44(11):1072-1082, 2012. 3", + "[48] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. Capri-net: Learning compact cad shapes with adaptive primitive assembly. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11768-11778, 2022. 8", + "[49] Mulin Yu and Florent Lafarge. Finding good configurations of planar primitives in unorganized point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6367-6376, 2022, 1, 2", + "[50] Qingnan Zhou, Eitan Grinspun, Denis Zorin, and Alec Jacobson. Mesh arrangements for solid geometry. ACM Trans. Graph., 35(4):39:1-39:15, 2016. 4", + "[51] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016. 5", + "[52] Marco Zuliani, Charles S Kenney, and BS Manjunath. The multiransac algorithm and its application to detect planar homographies. In IEEE Int. Conf. Image Process., volume 3, pages III-153, 2005. 2" + ], + "bbox": [ + 501, + 92, + 890, + 445 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14180", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/4d750dca-9a8b-44f9-b884-62a2d6314ac8_model.json b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/4d750dca-9a8b-44f9-b884-62a2d6314ac8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b2be673334bb52596ebf35643bb3a8eee6fb2719 --- /dev/null +++ b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/4d750dca-9a8b-44f9-b884-62a2d6314ac8_model.json @@ -0,0 +1,2981 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.815, + 0.154 + ], + "angle": 0, + "content": "Structure-Aware Surface Reconstruction via Primitive Assembly" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.187, + 0.777, + 0.21 + ], + "angle": 0, + "content": "Jingen Jiang\\(^{1,2,3,4^*}\\) Mingyang Zhao\\(^{1,5^*}\\) Shiqing Xin\\(^{4}\\) Yanchao Yang\\(^{6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.211, + 0.721, + 0.231 + ], + "angle": 0, + "content": "Hanxiao Wang\\(^{1,3}\\) Xiaohong Jia\\(^{2,3}\\) Dong-Ming Yan\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.231, + 0.68, + 0.247 + ], + "angle": 0, + "content": "\\(^{1}\\)MAIS & NLPR, CASIA \\(^{2}\\)KLMM, AMSS, CAS" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.248, + 0.782, + 0.267 + ], + "angle": 0, + "content": "\\(^{3}\\)UCAS \\(^{4}\\)Shandong University \\(^{5}\\)BAAI \\(^{6}\\)The University of Hong Kong" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.268, + 0.834, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.506, + 0.895, + 0.552 + ], + "angle": 0, + "content": "Figure 1: Surface reconstruction for measurement point clouds. In contrast to piecewise approximations (KSR), our method not only delivers manifold and watertight surfaces but also retains the primitive structures (planes, cylinders, spheres, cones, and tori) faithfully. Also, our method enables the recovery of the sharp geometry of objects (e.g., feature lines) as a byproduct." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.565, + 0.314, + 0.58 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.844 + ], + "angle": 0, + "content": "We propose a novel and efficient method for reconstructing manifold surfaces from point clouds. Unlike previous approaches that use dense implicit reconstructions or piecewise approximations and overlook inherent structures like quadrics in CAD models, our method faithfully preserves these quadric structures by assembling primitives. To achieve high-quality primitive extraction, we use a variational shape approximation, followed by a mesh arrangement for space partitioning and candidate primitive patches generation. We then introduce an effective pruning mechanism to classify candidate primitive patches as active or inactive, and further prune inactive patches to reduce the search space and speed up surface extraction significantly. Finally, the optimal active patches are computed by a binary linear programming and assembled as manifold and watertight surfaces. We perform extensive experiments on a wide range of CAD objects to validate its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.564, + 0.631, + 0.58 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.895, + 0.716 + ], + "angle": 0, + "content": "The problem of converting unorganized point clouds into surface meshes, also known as surface reconstruction, has various applications in 3D vision, computer graphics, and architecture geometry. In recent decades, although many powerful algorithms have been devoted to pursuing faithful and accurate reconstructions [19, 21, 34, 28, 27, 49], it remains a challenging task due to the inherent ill-posed nature [11, 36, 5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.903 + ], + "angle": 0, + "content": "The popular Poisson reconstruction methods [24, 25] implicitly generate dense smooth surfaces for point clouds, while others mainly focus on giving a compact approximation to objects by piecewise planar shapes or polygonal meshes [7, 13, 3, 35]. However, man-made objects in real-world scenarios are typically composed of non-planar quadratic surfaces or quadric structures, such as the Computer-Aided Design (CAD) models (see Fig. 1 as an example). Therefore, reconstructing manifold surface meshes from measurement data meanwhile retaining their intrinsic primitive structures in principle shall generate higher-quality approximation results than solely relying on poly-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.335, + 0.888 + ], + "angle": 0, + "content": "*The authors contribute equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.888, + 0.374, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author (yandongming@gmail.com)." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.374, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14171" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.094, + 0.117, + 0.106 + ], + "angle": 0, + "content": "gons." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.107, + 0.468, + 0.287 + ], + "angle": 0, + "content": "We argue that a method that solves the aforesaid problem ought to satisfy the following objectives: 1) Structure-awareness. The reconstruction surfaces should faithfully keep the structural features of the input point cloud, for instance, the common primitives in CAD, meanwhile correctly recovering the detailed geometry, and giving a high-fidelity approximation (e.g., the minimal Hausdorff distance) to the target object. 2) Manifold guarantees. The output surface meshes should be oriented 2-manifold and watertight [35]. 3) Efficiency. The developed algorithm should be scalable to large-scale point clouds with an acceptable computational complexity [3]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.288, + 0.468, + 0.71 + ], + "angle": 0, + "content": "In this work, we develop a novel and efficient framework that takes the underlying geometric structures of point clouds into account when performing shape reconstruction. Our method ensures high-quality outputs and aims to achieve the above objectives. The core of our algorithm lies in a faithful reconstruction of intrinsic primitive structures presented in CAD models, as opposed to the direct surface approximation based on piecewise planar primitives. We consider five representative types of primitives commonly found in modern CAD systems, including four quadrics (planes, cylinders, cones, spheres) and tori. We adopt a variational shape approximation to optimize the extracted primitives followed by mesh arrangement techniques to attain primitive patches. A patch is a manifold that consists of a set of connected triangle meshes. As the size of patches is typically smaller than that of facets, our method enables a more stable and efficient reconstruction process than plane-based ones. Moreover, we propose a patch pruning mechanism to further speed up the consecutive surface extraction process. This mechanism distinguishes patches as either active or inactive, allowing us to perform fast local collision detection instead of a time-consuming global processing. Lastly, we select an optimal set of patches from the active candidate patches to assemble a reconstruction surface that respects the underlying structure of the object. To this end, we optimize a Binary Linear Programming (BLP) under hard constraints, as done in [35], to guarantee that the resulting surface is both manifold and watertight." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.711, + 0.468, + 0.845 + ], + "angle": 0, + "content": "We perform extensive experiments to validate the effectiveness of our algorithm and compare it with state-of-the-art approaches using a variety of CAD models. Experimental results evidence that our method outperforms competitors in terms of both approximation precision and reconstruction efficiency, while also preserving the intrinsic primitive structures of the input point clouds with much better fidelity. To summarize, the main technical contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- An efficient structure-aware reconstruction method that produces accurate surface models while preserving the underlying intrinsic primitive structures of the" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.658, + 0.107 + ], + "angle": 0, + "content": "input point clouds." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.116, + 0.89, + 0.16 + ], + "angle": 0, + "content": "- A patch pruning technique that avoids the time-consuming global collision detection and accelerates the surface extraction by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.169, + 0.89, + 0.243 + ], + "angle": 0, + "content": "- A patch-induced binary linear programming for assembling manifold and watertight surfaces from candidate active patches, which significantly reduces the search space and helps achieve efficient and accurate surface reconstruction." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.116, + 0.89, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.257, + 0.642, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.282, + 0.892, + 0.326 + ], + "angle": 0, + "content": "In this section, we briefly review the works that are most relevant to ours. Please refer to [5, 23] for a more comprehensive discussion." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.327, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Geometric primitive extraction. Segmenting 3D point cloud models into a collection of geometric primitives such as planes and cylinders is helpful in most reconstruction approaches. Popular extraction frameworks include region growing [31, 39] and RANSAC [43, 52]. The former strategy takes the neighboring structure of models into consideration and propagates candidate primitives on their \\(k\\)-nearest neighbor graph until the fitting error exceeds a preset tolerance. Instead, RANSAC adopts a hypothesis and verification mechanism to recover primitives that embrace the most inliers. In order to make full use of the a priori structure of the input model, Li et al. [29] started with primitive extraction by RANSAC, followed by the refinement subject to spatial relationships such as parallelism and orthogonality. This idea was further extended to extract planar primitives simultaneously [32] or sequentially [37]. Recently, Yu et al. [49] proposed a new algorithm for extracting planar primitives, which has powerful potential for plane-assembly reconstruction. In our work, we leverage an efficient region growing strategy [39] for primitive extraction followed by a refinement via variational shape optimization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.645, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Assembly-based reconstruction. The core idea of assembly-based methods is to integrate all extracted primitives into a hybrid mesh. Connectivity analysis and slicing are two principal assembly manners. Connectivity strategies [8, 42, 46] consider neighboring relationships between primitives and encode them into an adjacency graph, which is adapted to regularize output meshes. Slicing strategies [6, 7, 3, 33, 35, 13] extend the extracted primitives and slice them to construct a spatial partition for candidate facets. Connectivity strategies are typically efficient but suffer from linkage errors of adjacent graphs generated by incomplete inputs. Arikan et al. [1] remedied the incomplete part by a user-guided interactive method that is labor-intensive especially for large-scale measurement data, whereas Lafarge et al. [26] provided an automatic solution that first consolidated the point clouds and then executed a min-cut optimization to extract dense triangle meshes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14172" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.078, + 0.892, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.22, + 0.894, + 0.251 + ], + "angle": 0, + "content": "Figure 2: Overview of the proposed method. (a) Point cloud after primitive extraction; (b) Accomplished space partitioning; (c) Active primitive patches after patch pruning; (d) The reconstructed surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.265, + 0.473, + 0.522 + ], + "angle": 0, + "content": "Alternatively, slicing strategies deal with incomplete models by means of extending primitives. Nan et al. [35] performed an exhaustive partition to compensate for all possible missing and thus guaranteed a manifold output via BLP. Nevertheless, this manner typically leads to high computational complexity for the subsequent surface extraction. To circumvent this shortcoming, Fang et al. [13] proposed a hybrid method that combines the connectivity analysis with the slicing process to make a limited partition, while Bauchet et al. [3] adopted a kinetic data structure [16, 15] to monitor the growing of convex piecewise planar polygons until they collide, which also results in a limited partition. Unlike previous approaches, we propose a simper partition by considering five types of primitives simultaneously rather than merely planar shapes. Moreover, we develop a propagation strategy inspired by Bauchet et al. [3] to further simplify and speed up the surface extraction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.473, + 0.83 + ], + "angle": 0, + "content": "Learning-based reconstruction. With the advancements in deep learning, there have been several notable works that utilize learning-based techniques to reconstruct meshes. Voxel-inspired methods, such as those presented in references [38] and [10], train a network of models to convert voxels into implicit surface representations. To attain more accurate representations, SIREN [45] and DIGS [4] used signed distance functions as implicit surface representations, and then extracted meshes via marching cubes [30]. Points2Surf [12] employed a patch-based learning framework to directly attain meshes from point clouds without the need for normals, while DSE [40] adopted 2D Delaunay triangulations to ensure manifold meshes and BSP-Net [9] output compact polygonal meshes with arbitrary topology. However, most of these methods depend on implicit surface representations, which may not explicitly capture the underlying structures such as quadric primitives of objects. In addition, it is hard to perform model editing tasks for surfaces generated by implicit reconstructions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.212, + 0.862 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The proposed method consists of three major steps as illustrated in Fig. 2. Given an input point cloud \\(\\mathcal{P} = \\{p_j\\in\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.895, + 0.416 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^3\\}_{j=1}^N\\) (Fig. 2(a)), we first partition \\(\\mathcal{P}\\) by a set of proxy meshes defined by the extracted geometric primitives and then apply a mesh arrangement operation to get candidate manifold patches (Fig. 2(b)) in Section 3.1. Later, a fast propagation mechanism is presented in Section 3.2 is used to prune patches that are inactive (Fig. 2(c)) to accelerate the subsequent surface extraction procedure. Finally, we deliver a manifold and watertight output surface (Fig. 2(d)) by formulating a patch-induced optimization under the BLP framework in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.427, + 0.68, + 0.443 + ], + "angle": 0, + "content": "3.1. Space Partitioning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.892, + 0.496 + ], + "angle": 0, + "content": "The goal of this step is to generate a finite space partitioning by arranging proxy meshes of geometry primitives extracted from the input point cloud \\(\\mathcal{P}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Primitive extraction. We first utilize an efficient region growing strategy introduced in [39] to attain the rough segmentation components \\(\\mathcal{S} = \\{s_i\\subset \\mathbb{R}^3\\}_{i = 1}^M\\) and the corresponding primitives \\(\\mathcal{Q} = \\{q_i\\subset \\mathbb{R}^3\\}_{i = 1}^M\\) from \\(\\mathcal{P}\\), where \\(q_{i}\\) is the supporting primitive of \\(s_i\\) inversely, each point \\(p_j\\in s_i\\) is the supporting point of \\(q_{i}\\). Let \\(\\mathcal{L} = (\\mathcal{S},\\mathcal{Q})\\) be the initial layout of the extracted primitives, we further refine \\(\\mathcal{L}\\) via the variational shape approximation based on the \\(\\ell^2\\) and the \\(\\ell^{2,1}\\) metrics [11, 47], which aims to optimize the following expression" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.679, + 0.892, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {L}} \\mathcal {F} (\\mathcal {L}) = \\sum_ {i = 1} ^ {M} \\sum_ {p _ {j} \\in s _ {i}} \\operatorname {d i s t} \\left(p _ {j}, q _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "where \\(dist(p_j, q_i)\\) stands for the hybrid metric that measures the cost of the point \\(p_j \\in s_i\\) with respect to the corresponding primitive \\(q_i \\in \\mathcal{Q}\\) in terms of both the spatial \\(\\ell^2\\) distance and the normal \\(\\ell^{2,1}\\) deviation. Fig. 2(a) presents the primitive extraction result based on five types of primitives that we are currently considering. Noted that the proposed framework is not constrained by the above detected primitive types. Instead, it can be integrated with powerful detection algorithms that support a wider variety of primitive types, enabling more complex reconstructions with better accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14173" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.458, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.197, + 0.471, + 0.273 + ], + "angle": 0, + "content": "Figure 3: Visual description of patches and (non-) boundary curves. (a) \\(\\mathcal{C}_{ij}\\) is the non-boundary curve generated by patches \\(m_{i}\\) and \\(m_{j}\\), while \\(\\bar{\\mathcal{C}}_i\\) and \\(\\bar{\\mathcal{C}}_j\\) are the boundary curves of \\(m_{i}\\) and \\(m_j\\), separately. (b) For each non-boundary curve, there are four patches around it." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.473, + 0.504 + ], + "angle": 0, + "content": "Proxy mesh arrangement. After the optimization of the extracted primitives, we generate the discrete approximation to all primitives via proxy meshes. These meshes are trimmed by the bounding box of \\(\\mathcal{P}\\) and then all meshes are processed by a mesh arrangement [50]. The mesh arrangement results are a set of candidate primitive patches \\(\\mathcal{M} = \\{m_i\\subset \\mathbb{R}^3\\}_{i = 1}^K\\) as illustrated in Fig. 2(b), which partition the current working space into a series of cells that form a closed manifold without self-intersections. Here each patch \\(m_{i}\\) represents a subset of the input meshes and is the maximal connected set of triangle meshes as illustrated in Fig. 3. Due to the connectivity of patches, each \\(m_{i}\\) may embrace several non-boundary curves \\(\\mathcal{C}_{ij}\\) (colored red in Fig. 3(a)) that are shared with other primitive patches, which are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.509, + 0.47, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {i j} = m _ {i} \\bigcap m _ {j} (i \\neq j), m _ {i}, m _ {j} \\in \\mathcal {M}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.473, + 0.675 + ], + "angle": 0, + "content": "whereas at most one boundary curve \\(\\bar{C}_i\\) (colored green in Fig. 3(a)). For example, there are in total of four neighboring patches that share the same \\(C_{ij}\\) in Fig. 3(b). Since we perform space partitioning based on primitive patches rather than discretizing each of them as multiple piecewise planes, the developed algorithm indeed is several orders of magnitude faster than previous plane-induced approaches. The efficiency superiority of our method is also quantitatively validated by the following experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.421, + 0.698 + ], + "angle": 0, + "content": "3.2. Pruning of Candidate Primitive Patches" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.903 + ], + "angle": 0, + "content": "It is known that the computational complexity of the ultimate surface extraction in shape reconstruction regarding the size of patches is exponential [44]. To expedite the subsequent optimization process, we further propose a pruning mechanism which effectively circumvents the time-consuming and memory-intensive global collision detection to classify current primitive patches as active and inactive, from which we prune inactive ones to reduce the search space for surface extraction. As illustrated in Fig. 2, the significant time gap between the two routes(black and red) demonstrates the effectiveness of our pruning mechanism. The proposed procedure is composed of two steps, initialization and pruning, as presented subsequently." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.096, + 0.844, + 0.111 + ], + "angle": 0, + "content": "Algorithm 1: Pseudo-code of the propagation process" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.113, + 0.86, + 0.127 + ], + "angle": 0, + "content": "Input: Priority queue \\(Q\\) of representative triangles." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.128, + 0.773, + 0.142 + ], + "angle": 0, + "content": "Output: Active primitive patches \\(\\tilde{\\mathcal{M}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.143, + 0.863, + 0.158 + ], + "angle": 0, + "content": "1 Initialize propagation cost of each triangle as \\( c = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.159, + 0.636, + 0.172 + ], + "angle": 0, + "content": "2 while \\(Q\\neq \\emptyset\\) do" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.173, + 0.736, + 0.189 + ], + "angle": 0, + "content": "3 \\(\\mid (t,c)\\gets Q.pop()\\) and \\(\\hat{t}\\leftarrow t\\)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.19, + 0.755, + 0.205 + ], + "angle": 0, + "content": "4 for each triangle edge \\(e\\in \\hat{t}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.206, + 0.744, + 0.22 + ], + "angle": 0, + "content": "5 for triangle \\(t_e\\cap t = e\\) do" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.221, + 0.819, + 0.235 + ], + "angle": 0, + "content": "6 | | if \\( \\text{prim}(t_e) == \\text{prim}(t) \\& t_e \\) is" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.236, + 0.824, + 0.25 + ], + "angle": 0, + "content": "inactive & \\(t_e\\) is not blocked then" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.252, + 0.845, + 0.266 + ], + "angle": 0, + "content": "7 Compute the height \\(h_e\\) of \\(t_e\\) on \\(e\\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.267, + 0.77, + 0.281 + ], + "angle": 0, + "content": "8 \\(Q\\) .push((te,c+he))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Initialization. To describe the point distribution of the input point cloud as fidelity as possible, we first select a set of representative triangle meshes denoted as \\(\\mathcal{I}\\) from \\(\\mathcal{M}\\) by means of the triangle coverage \\(Cover(t)\\) which is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.403, + 0.892, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C o v e r} (t) = \\frac {\\operatorname {V a l i d} (t)}{\\operatorname {A r e a} (t)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.44, + 0.894, + 0.47 + ], + "angle": 0, + "content": "where \\( \\text{Area}(t) \\) stands for the area of a triangle mesh \\( t \\) and \\( \\text{Valid}(t) \\) is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.479, + 0.892, + 0.513 + ], + "angle": 0, + "content": "\\[\nV a l i d (t) = \\sum_ {p \\in \\mathcal {P} _ {t}} 1 (\\| p - p ^ {\\prime} \\| _ {2} < \\epsilon) a _ {p}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.892, + 0.566 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_t\\subset \\mathcal{P}\\) encompasses the measurement points that their projections are attached to \\(t\\) \\(\\| p - p^{\\prime}\\|_{2}\\) denotes the \\(\\ell^2\\) Euclidean distance from \\(p\\) to its projective point \\(p^\\prime\\) on \\(t\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.567, + 0.695, + 0.687 + ], + "angle": 0, + "content": "1 is the indicator function, and \\(a_{p}\\) is the corresponding geodesic Voronoi area of \\(p\\), which is calculated as in the inset: \\(p\\) and its \\(k\\)-nearest neighbors (i.e., hard red points) are first projected to the supporting plane \\(\\Omega_t\\)" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.567, + 0.864, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.763 + ], + "angle": 0, + "content": "where its attached triangle mesh \\(t\\) lies on, then \\(a_{p}\\) is equivalent to the area of the 2D Voronoi cell [2]. \\(t\\) is termed representative if its Cover \\((t)\\) surpasses a user-specified value. We populate all representative meshes of \\(\\mathcal{I}\\) into a priority queue \\(Q\\) with an ascending order." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Pruning. Next, representative triangle meshes are propagated to enable a limited partition to speedup the final surface extraction. To this end, all triangles of \\( Q \\) are initially labeled as inactive prior to entering the loop of the priority queue. For each round during the loop, we first pop the top element, for instance, \\( (t_i, c_i) \\), out of the priority queue and prompt \\( t_i \\) to be an active state \\( \\hat{t}_i \\). The propagation cost \\( c_i \\) of a triangle mesh \\( t_i \\) is defined as the height" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14174" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.378 + ], + "angle": 0, + "content": "\\(h_i\\) from its vertex to the base edge shared with another adjacent triangle. Sequentially, we propagate all the triangles that share both the same edge and the same supporting primitive with \\(t_i\\) and pop the update elements out of \\(Q\\). For instance, for edges that are not located on non-boundary curves, there are at most three update elements \\((t_{i1},c_i + h_1)\\), \\((t_{i2},c_i + h_2)\\), \\((t_{i3},c_i + h_3)\\), whereas, for edges that coincide non-boundary curves, we perform a fast local collision detection to check whether there exist active meshes, if yes then we evoke barricades and stop the propagation process. The above propagation proceeds until \\(Q\\) becomes empty. We summarize the proposed propagation process in Algorithm 1 to make it more understandable. Consequently, a primitive patch is active only if its triangles are all active, otherwise, it is still inactive. We use \\(\\tilde{\\mathcal{M}}\\) to represent the set of active patches. The patches in Fig. 2(b) after pruning are shown in Fig. 2(c). As observed, the size of candidate patches has been significantly reduced compared with Fig. 2(b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.472, + 0.471 + ], + "angle": 0, + "content": "Like [3], we also allow for the propagation crossing primitive patches \\(\\mathcal{G}\\) times (with \\(\\mathcal{G} = 2\\) in default) at most. This can deal with small parts that may have been missed during the primitive detection. However, in contrast to previous global collision detection manners [3], our local strategy preserves much better the propagation efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.48, + 0.262, + 0.495 + ], + "angle": 0, + "content": "3.3. Surface Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.473, + 0.673 + ], + "angle": 0, + "content": "Finally, we adopt a Binary Linear Programming (BLP) analogous to [35] to extract the surface model from \\(\\tilde{\\mathcal{M}}\\). Specifically, the state of patches in \\(\\tilde{\\mathcal{M}}\\) is encoded as \\(\\mathcal{X}_{\\tilde{M}} = \\{x_{\\tilde{m}}|\\tilde{m}\\in \\tilde{\\mathcal{M}},x_{\\tilde{m}}\\in \\{0,1\\} \\}\\), where the binary discrete variable \\(x_{\\tilde{m}}\\) indicates whether the patch \\(\\tilde{m}\\) is selected \\((x_{\\tilde{m}} = 1)\\) or not \\((x_{\\tilde{m}} = 0)\\). To ensure the resulting surface manifold and watertight, we constrain the curves of the selected patches to be all non-boundary, and thus the two patches sharing the same non-boundary curve are either selected at the same time or both not selected. Consequently, our objective function is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.683, + 0.47, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\mathcal {X} _ {\\tilde {\\mathcal {M}}}} \\mathbf {U} (\\mathcal {X} _ {\\tilde {\\mathcal {M}}}) + \\lambda \\mathbf {B} (\\mathcal {X} _ {\\tilde {\\mathcal {M}}}), \\\\ s. t. \\left\\{ \\begin{array}{l} \\sum_ {\\tilde {m} _ {i} \\cap \\tilde {m} _ {j} \\neq \\emptyset , i \\neq j} \\left(x _ {\\tilde {m} _ {i}} + x _ {\\tilde {m} _ {j}}\\right) = 0 o r 2, \\tilde {m} _ {i}, \\tilde {m} _ {j} \\in \\tilde {\\mathcal {M}} \\\\ x _ {\\tilde {m} _ {i}} \\in \\{0, 1 \\}, \\quad i = 1, 2, 3 \\dots | \\tilde {\\mathcal {M}} |. \\end{array} \\right. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.854 + ], + "angle": 0, + "content": "where \\(\\mathbf{U}(\\mathcal{X}_{\\tilde{\\mathcal{M}}})\\) and \\(\\mathbf{B}(\\mathcal{X}_{\\tilde{\\mathcal{M}}})\\) are penalty terms regarding the patches and the neighboring patch pairs, respectively, and \\(\\lambda\\) is a balance coefficient. Moreover \\(\\mathbf{U}(\\mathcal{X}_{\\tilde{\\mathcal{M}}})\\) encourages the selection of active patches that are with high fidelity in the form of" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.863, + 0.47, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} \\left(\\mathcal {X} _ {\\tilde {M}}\\right) = \\sum_ {\\tilde {m} \\in \\tilde {\\mathcal {M}}} x _ {\\tilde {m}} \\cdot \\left(\\frac {\\mathcal {A} _ {\\tilde {m}} - \\bar {\\mathcal {A}} _ {\\tilde {m}}}{\\mathcal {A}} - \\frac {| \\mathcal {P} _ {\\tilde {m}} |}{| \\mathcal {P} |}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.214 + ], + "angle": 0, + "content": "Here \\(\\mathcal{A}_{\\tilde{m}}\\) and \\(\\bar{\\mathcal{A}}_{\\tilde{m}}\\) denote the area and the coverage area of the patch \\(\\tilde{m}\\), respectively. The coverage area of \\(\\tilde{m}\\) is the summation of the coverage area of all its triangles. \\(\\mathcal{A}\\) is a normalization factor equivalent to the sum of the area of all patches. \\(\\mathcal{P}_{\\tilde{m}}\\) is the set of points whose projections are locating in \\(\\tilde{m}\\). On the other hand, \\(\\mathcal{B}(\\tilde{\\mathcal{X}})\\) is designed to control the model complexity. Specifically, it put penalties to the non-boundary curve length of two patches:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.222, + 0.892, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} \\left(\\mathcal {X} _ {\\tilde {\\mathcal {M}}}\\right) = \\frac {1}{L} \\sum_ {i, j} \\left| \\mathcal {C} _ {i j} \\right| \\cdot x _ {\\tilde {m} _ {i}} \\cdot x _ {\\tilde {m} _ {j}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.362 + ], + "angle": 0, + "content": "where \\( |\\mathcal{C}_{ij}| \\) denotes the length of the curve shared by two active patches \\( \\tilde{m}_i \\) and \\( \\tilde{m}_j \\) with different supporting primitives and \\( L \\) is the sum of the length of all curves. Fig. 2(d) presents the final reconstruction surface by BLP, which not only generates a manifold and watertight shape but also successfully preserves the intrinsic primitive structures." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.374, + 0.88, + 0.391 + ], + "angle": 0, + "content": "4. Experimental Evaluations and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Implementation details and metrics. In this section, we execute extensive experiments to validate the advantages of the developed method in terms of both reconstruction precision and efficiency and make various comparisons with state-of-the-art approaches. Like [3], the precision is quantitatively evaluated by the symmetric mean Hausdorff distance (SMH) between the measurement point clouds and the reconstruction meshes (values are normalized by the diagonal length of the models' bounding box and multiplied by a factor of \\(10^{2}\\)), while the efficiency is assessed by the running time. Our algorithm is implemented in \\(\\mathrm{C}++\\) using the geometry processing library libigl [22]. We adopt the mathematical programming solver Gurobi [18] to optimize Eq. 5. All reconstruction results are rendered by Blender [14] for better visualization. Our experiments are conducted on a desktop with Intel(R) Core i7-7700K CPU@4.2GHz and 48GB RAM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Reconstruction effect. Our algorithm has demonstrated good reconstruction results from point clouds to manifold meshes on various CAD objects as illustrated in Fig. 4. The input point clouds presented in Fig. 4(a)-(h) are either sampled from real-world CAD models or courtesy by publicly available benchmark datasets [41, 51]. As observed, there are multiple explicit intrinsic primitive structures in each model. Instead of approximating them directly by piecewise planar primitives, our algorithm is capable of reconstructing the intrinsic primitives faithfully even for quite challenging or complicated models. For instance, the rich spherical and conic structures existing in the Minarets model of Fig. 4(b) require careful treatments, yet our method is still able to achieve highly accurate reconstructions while preserving all structural details successfully. At the same time," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14175" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.113, + 0.111, + 0.129 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.085, + 0.231, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.084, + 0.393, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.085, + 0.543, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.085, + 0.702, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.186, + 0.111, + 0.201 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.163, + 0.232, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.167, + 0.395, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.167, + 0.548, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.172, + 0.706, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.086, + 0.867, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.111, + 0.275 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.245, + 0.236, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.246, + 0.398, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.246, + 0.552, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.248, + 0.714, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.172, + 0.869, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.339, + 0.111, + 0.353 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.311, + 0.237, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.311, + 0.397, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.311, + 0.556, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.31, + 0.717, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.249, + 0.87, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.31, + 0.871, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.416, + 0.111, + 0.431 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.393, + 0.243, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.399, + 0.397, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.398, + 0.561, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.399, + 0.728, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.4, + 0.873, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.503, + 0.11, + 0.518 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.481, + 0.234, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.482, + 0.387, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.484, + 0.548, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.485, + 0.721, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.485, + 0.867, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.604, + 0.111, + 0.62 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.575, + 0.231, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.575, + 0.384, + 0.667 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.574, + 0.545, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.574, + 0.721, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.574, + 0.871, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.702, + 0.111, + 0.717 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.696, + 0.252, + 0.733 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.693, + 0.401, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.693, + 0.561, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.689, + 0.73, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.685, + 0.885, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.728, + 0.882, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.759, + 0.892, + 0.833 + ], + "angle": 0, + "content": "Figure 4: Reconstruction results by the proposed method. From left to right: Input measurement point clouds, extracted geometric primitives (different colors stand for different primitives), output reconstruction models, recovered sharp geometry along with the original point clouds, and colored models showing the error distribution regarding the Hausdorff distance. Our method produces high-quality reconstructions and preserves the intrinsic primitive structures as well as the sharp features well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.849, + 0.47, + 0.895 + ], + "angle": 0, + "content": "our method produces high-quality sharp feature lines attained by patch intersections, as shown in the forth column of Fig. 4. This is an important advantage for many practi" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.849, + 0.894, + 0.895 + ], + "angle": 0, + "content": "cal applications such as sketch-style rendering, as it ensures that the reconstructed object retains its original shape and does not appear distorted or blurry. The fifth column of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14176" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.088, + 0.079, + 0.882, + 0.094 + ], + "angle": 0, + "content": "Table 1: Statistical results of the quantity and the runtime (in second) during reconstruction of the CAD models in Fig. 4." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.095, + 0.875, + 0.257 + ], + "angle": 0, + "content": "
Models Metrics(a)(b)(c)(d)(e)(f)(g)(h)
# Points of the input point cloud10k593k10k10k74k7k16k14k
# Extracted primitives2853212034322927
# Generated candidate patches36466024103072349704177501092
# Active patches after pruning224533697414172586342472667
Patch decrease (%)38.42644.07328.05842.32447.96817.98637.06738.919
# Determined active patches of the output model605915191935339986114
Mesh arrangement (s)7.85321.7522.4631.7275.2362.9355.4311.421
Patch pruning (s)0.92715.4210.7920.4763.0430.4881.1310.726
Surface extraction (s)1.2882.0860.6590.0911.1120.1570.3140.137
" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.271, + 0.171, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.326, + 0.17, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.385, + 0.16, + 0.401 + ], + "angle": 0, + "content": "\\(\\sigma = 0.01(0.82m)\\) \\(\\mathrm{SMH} = 0.11\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.271, + 0.273, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.325, + 0.269, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.386, + 0.261, + 0.401 + ], + "angle": 0, + "content": "\\(\\sigma = 0.02\\) \\((1.64\\mathrm{m})\\) \\(\\mathrm{SMH} = 0.11\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.271, + 0.379, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.325, + 0.378, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.386, + 0.368, + 0.401 + ], + "angle": 0, + "content": "\\(\\sigma = 0.04(3.28m)\\) \\(\\mathrm{SMH} = 0.12\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.269, + 0.482, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.326, + 0.479, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.386, + 0.47, + 0.401 + ], + "angle": 0, + "content": "\\(\\sigma = 0.08(6.56m)\\) \\(\\mathrm{SMH} = 46\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.407, + 0.47, + 0.467 + ], + "angle": 0, + "content": "Figure 5: Robustness test against noise disturbance. Top: Input point clouds with increasing noise. Bottom: Our reconstruction results along with the noise level \\(\\sigma\\) and the reconstruction error SMH." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.484, + 0.47, + 0.545 + ], + "angle": 0, + "content": "Fig. 4 visualizes the reconstruction error of our algorithm with respect to the Hausdorff distance, where our method consistently produces accurate approximations for all objects, demonstrating its stability and effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.686 + ], + "angle": 0, + "content": "Statistical results. Table 1 summarizes the quantitative results of our proposed method for the models in Fig. 4. Our algorithm significantly reduces candidate patches, especially for complex models like Fig. 4(b),(d),(e), where the patch decrease ratio exceeds \\(40\\%\\). In terms of efficiency, patch pruning and surface extraction outperform mesh arrangement, and further improvements can be made by exploring a more efficient intersection mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Robustness assessment. We also assess the ability of the proposed algorithm against noise disturbance and occlusion data. As depicted in Fig. 5, we increase the standard deviation \\(\\sigma\\) of Gaussian noise from \\(0.01(0.82m)\\) to \\(0.04(3.28m)\\), where values in parentheses are real intensity relative to the bounding sphere's radius of the model. Our method consistently generates accurate reconstructions in terms of both topological structures and approximation distance (SMH), suggesting its high resilience to noise. Even when \\(\\sigma = 0.08(6.56m)\\), which has caused significant topological alteration of the original point cloud, our method still maintains the overall structures of the object. Additionally, we verify the robustness of our method to occlusion" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.271, + 0.58, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.338, + 0.58, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.271, + 0.672, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.338, + 0.674, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.271, + 0.773, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.335, + 0.777, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.271, + 0.873, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.335, + 0.879, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Figure 6: Robustness test against occlusion data. Top: Input point clouds with increasing occlusions. Bottom: Our reconstruction results along with the reconstruction error." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.472, + 0.892, + 0.517 + ], + "angle": 0, + "content": "Table 2: Comparisons with state-of-the-art approaches, where \"w\" and \"w/o\" indicate with and without the patch pruning, separately." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.519, + 0.855, + 0.566 + ], + "angle": 0, + "content": "
MetricsiPSR[20]SIREN[45]DSE[40]KSR[3]Ours
ww/o
SMH (%)4.2341.793.15169.442.452.45
Time (s)201.5891208.124303.865570.46845.556.777
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.669 + ], + "angle": 0, + "content": "sion data. Fig. 6 reports our reconstruction results after progressively erasing point clouds. As observed, our method shows remarkable resistance to occlusions and has successfully compensated the missing structures. Conversely, this is also quite helpful to complete the imperfect point clouds by sampling points in the reconstruction models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Comparisons. Subsequently, we conduct a comparative analysis of our method with representative state-of-the-art approaches, including densely implicit reconstruction methods iPSR [20], DSE [40], and SIREN [45], as well as the plane-assembly approach KSR [3]. The quantitative comparison results in terms of reconstruction accuracy and efficiency on the Cheese model is presented in Table 2. The results demonstrate that our method achieves competitive reconstruction accuracy with implicit approaches iPSR and DSE, and outperforms SIREN and KSR by a large margin. Additionally, our algorithm performs significantly faster than competitors, even without patch pruning. The qualitative reconstruction results presented in Fig. 7 show that benefiting from the introduced primitive assembly, our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14177" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.077, + 0.207, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.077, + 0.34, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.077, + 0.473, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.077, + 0.605, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.077, + 0.73, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.077, + 0.868, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.892, + 0.269 + ], + "angle": 0, + "content": "Figure 7: Qualitative comparisons with state-of-the-art approaches. Our method enables the shape reconstruction faithfully and successfully recovers the sharp geometry (e.g., feature lines and corners) from unorganized point clouds." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.288, + 0.468, + 0.302 + ], + "angle": 0, + "content": "Table 3: Quantitative comparisons on the ABC300 dataset." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.304, + 0.47, + 0.343 + ], + "angle": 0, + "content": "
Methods MetricsiPSR[20]SIREN[45]DSE[40]KSR[3]ComplexGen[17]CAPRI-Net[48]Ours
SMH (%)2.7124.894.0626.4324.2415.874.81
Time (s)24.0641565.21732.8004.208553.089183.8443.728
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.399 + ], + "angle": 0, + "content": "method successfully recovers sharp features from unorganized point clouds, which is a salient advantage compared to implicit methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.551 + ], + "angle": 0, + "content": "We also conduct quantitative comparisons with several predecessors (additionally including CAPRI-Net [48] and ComplexGen [17]) on 300 CAD models from the ABC dataset (abbreviated as ABC300). Table 3 demonstrates that our method achieves consistently competitive reconstruction accuracy with implicit approaches such as iPSR and DSE (while also providing advantages for sharp feature recovery and model editing), and significantly surpasses KSR, CAPRI-Net, and ComplexGen. Furthermore, our method is considerably faster than its competitors." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.584, + 0.16, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.566, + 0.257, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.564, + 0.362, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.553, + 0.469, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.688 + ], + "angle": 0, + "content": "Figure 8: Model editing based on our structure-aware reconstruction algorithm. (a) The original reconstruction object. (b)-(d) The newly edited results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Balance parameter. We investigate the impact of \\(\\lambda\\) on surface extraction by varying its values to different settings (see Fig. 9(a)). It can be seen that increasing \\(\\lambda\\) brings the reduction of the complexity of the output model. However, this also degrades the reconstruction accuracy. For example, when \\(\\lambda\\) is set to 0.7, the inner hollows of the input point clouds disappear." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Application to model editing. Our reconstruction method can be employed for flexible model editing by taking advantage of the structure-aware merits. As shown in Fig. 8, since each quadratic structure of the object is preserved along with their detailed geometric parameters, one" + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.353, + 0.521, + 0.365 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.294, + 0.566, + 0.306 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.307, + 0.589, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.294, + 0.664, + 0.305 + ], + "angle": 0, + "content": "\\(\\lambda = 0.2\\)" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.31, + 0.678, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.294, + 0.751, + 0.304 + ], + "angle": 0, + "content": "\\(\\lambda = 0.7\\)" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.31, + 0.768, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.294, + 0.836, + 0.304 + ], + "angle": 0, + "content": "#Point=10k" + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.304, + 0.908, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.434 + ], + "angle": 0, + "content": "Figure 9: (a) Impact of parameter \\(\\lambda\\). As \\(\\lambda\\) is increased, the complexity of the output model decreases. (b) In sparse input scenarios, small primitives can be easily lost, leading to failures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.503 + ], + "angle": 0, + "content": "can modify the reconstruction shape by simply specifying various parameters. For instance, by re-defining the radius or height of the cylindrical primitives in Fig. 8(a), we can resize them to our desirable shapes (Fig. 8(b)-(d))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.892, + 0.631 + ], + "angle": 0, + "content": "Limitations. Our approach primarily focuses on the rapid and accurate reconstruction of representative primitive structures in CAD models. However, as shown in Fig 9(b), our reconstruction quality may deteriorate for inputs lacking explicit primitives. Additionally, inadequate primitive extraction due to severe occlusion or noise can degrade the quality of the reconstruction results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.648, + 0.769, + 0.664 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.854 + ], + "angle": 0, + "content": "We propose an accurate and fast shape reconstruction pipeline for converting disordered 3D point clouds into manifold, watertight surfaces. Our method preserves representative intrinsic primitive structures in CAD systems through primitive assembly, avoiding dense implicit reconstructions and extending piecewise planar approximations. We also develop an effective patch pruning mechanism, greatly expediting surface extraction by reducing patch intersections. Additionally, our method is highly resilient to noise and can reliably recover sharp object geometry. Extensive experiments show our algorithm's superiority over competing methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In our future work, we plan to expand the current primitive types to general primitives, and even spline surfaces, in order to efficiently process more complex CAD models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14178" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.091, + 0.245, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.116, + 0.47, + 0.206 + ], + "angle": 0, + "content": "This work was partially funded by the National Key Research and Development Program (2021YFB1715900), the CAS Project for Young Scientists in Basic Research (YSBR-034), the National Natural Science Foundation of China (62172415, 62272277, 12022117), and the HKU-100 Research Award." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.221, + 0.173, + 0.236 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[1] Murat Arikan, Michael Schwärzler, Simon Flöry, Michael Wimmer, and Stefan Maierhofer. O-snap: Optimization-based snapping for modeling architecture. ACM Trans. Graph., 32(1):6:1-6:15, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.472, + 0.345 + ], + "angle": 0, + "content": "[2] Gavin Barill, Neil G Dickson, Ryan Schmidt, David IW Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. ACM Trans. Graph., 37(4):43:1-43:12, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[3] Jean-Philippe Bauchet and Florent Lafarge. Kinetic shape reconstruction. ACM Trans. Graph., 39(5):156:1-156:14, 2020. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.389, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.446, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[5] Matthew Berger, Andrea Tagliasacchi, Lee M Seversky, Pierre Alliez, Gael Guennebaud, Joshua A Levine, Andrei Sharf, and Claudio T Silva. A survey of surface reconstruction from point clouds. Comput. Graph. Forum, 36(1):301-329, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.517, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[6] Alexandre Boulch, Martin de La Gorce, and Renaud Marlet. Piecewise-planar 3d reconstruction with edge and corner regularization. Comput. Graph. Forum, 33(5):55-64, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.574, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[7] Anne-Laure Chauve, Patrick Labatut, and Jean-Philippe Pons. Robust piecewise-planar 3d reconstruction and completion from large-scale unstructured point data. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1261-1268, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.645, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[8] Jie Chen and Baoquan Chen. Architectural modeling from sparsely scanned range data. Int. J. Comput. Vis., 78(2):223-236, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.688, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5939-5948, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[11] David Cohen-Steiner, Pierre Alliez, and Mathieu Desbrun. Variational shape approximation. ACM Trans. Graph., 23(3):905-914, 2004. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In Eur. Conf. Comput. Vis., pages 108-124, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Hao Fang and Florent Lafarge. Connect-and-slice: an hybrid approach for reconstructing 3d objects. In IEEE Conf." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.246, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Comput. Vis. Pattern Recog., pages 13490-13498, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.15 + ], + "angle": 0, + "content": "[14] Blender Foundation. Blender - a 3D modelling and rendering package. http://www.blender.org.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.151, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[15] Leonidas Guibas. Kinetic data structures. In Handbook of Data Structures and Applications, pages 377-388. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.18, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[16] Leonidas Guibas, Lyle Ramshaw, and Jorge Stolfi. A kinetic framework for computational geometry. In IEEE Annual Symposium on Foundations of Computer Science, pages 100-111, 1983. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[17] Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. Complexgen: Cad reconstruction by b-rep chain complex generation. ACM Trans. Graph., 41(4):129:1-129:18, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[18] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual. https://www.gurobi.com.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[19] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. In Proc. ACM SIGGRAPH, pages 71-78, 1992. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[20] Fei Hou, Chiyu Wang, Wencheng Wang, Hong Qin, Chen Qian, and Ying He. Iterative poisson surface reconstruction (ipsr) for unoriented points. ACM Trans. Graph., 41(4):128:1-128:13, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[21] Hui Huang, Dan Li, Hao Zhang, Uri Ascher, and Daniel Cohen-Or. Consolidation of unorganized point clouds for surface reconstruction. ACM Trans. Graph., 28(5):1-7, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[22] Alec Jacobson and Daniele Panozzo. Libigl: Prototyping geometry processing research in c++. In SIGGRAPH Asia 2017 courses, pages 1-172. 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[23] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3d data. Comput. Graph. Forum, 38(1):167-196, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Comput. Graph. Forum (Proceedings of SGP), page 61-70, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[25] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. Graph., 32(3):29:1-29:13, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[26] Florent Lafarge and Pierre Alliez. Surface reconstruction through point set structuring. Comput. Graph. Forum, 32(2pt2):225-234, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[27] Minglei Li, Liangliang Nan, Neil Smith, and Peter Wonka. Reconstructing building mass models from UAV images. Comput. Graph., 54:84-93, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[28] Minglei Li, Peter Wonka, and Liangliang Nan. Manhattanworld urban reconstruction from point clouds. In Eur. Conf. Comput. Vis., pages 54-69, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[29] Yangyan Li, Xiaokun Wu, Yiorgos Chrysathou, Andrei Sharf, Daniel Cohen-Or, and Niloy J Mitra. Globfit: Consistently fitting primitives by discovering global relations. ACM Trans. Graph., 30(4):52:1-52:12, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proc. ACM SIGGRAPH, volume 21, pages 163-169. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14179" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] David Marshall, Gabor Lukacs, and Ralph Martin. Robust segmentation of primitives from range data in the presence of geometric degeneracy. IEEE Trans. Pattern Anal. Mach. Intell., 23(3):304-314, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[32] Aron Monszpart, Nicolas Mellado, Gabriel J Brostow, and Niloy J Mitra. Rapter: rebuilding man-made scenes with regular arrangements of planes. ACM Trans. Graph., 34(4):103:1-103:12, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[33] Claudio Mura, Oliver Mattausch, and Renato Pajarola. Piecewise-planar reconstruction of multi-room interiors with arbitrary wall arrangements. Comput. Graph. Forum, 35(7):179-188, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.317 + ], + "angle": 0, + "content": "[34] Liangliang Nan, Andrei Sharf, Hao Zhang, Daniel Cohen-Or, and Baoquan Chen. Smartboxes for interactive urban reconstruction. ACM Trans. Graph., 29(4):93:1–93:10, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[35] Liangliang Nan and Peter Wonka. Polyfit: Polygonal surface reconstruction from point clouds. In IEEE Int. Conf. Comput. Vis., pages 2353-2361, 2017. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[36] Anh Nguyen and Bac Le. 3d point cloud segmentation: A survey. In IEEE Int. Conf. Robot. Autom. Mechatron, pages 225-230, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[37] Sven Oesau, Florent Lafarge, and Pierre Alliez. Planar shape detection and regularization in tandem. Comput. Graph. Forum, 35(1):203-215, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[38] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 165-174, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[39] Tahir Rabbani, Frank Van Den Heuvel, and George Vosselmann. Segmentation of point clouds using smoothness constraint. ISPRS J. PhotoGramm., 36(5):248-253, 2006. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[40] Marie-Julie Rakotosaona, Paul Guerrero, Noam Aigerman, Niloy J. Mitra, and Maks Ovsjanikov. Learning delaunay surface elements for mesh reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22-31, June 2021. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[41] Chiara Romanengo, Andrea Raffo, Yifan Qie, Nabil Anwer, and Bianca Falcidieno. Fit4cad: A point cloud benchmark for fitting simple geometric primitives in cad objects. Comput. Graph., 102:133-143, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[42] Falko Schindler, Wolfgang Wörstner, and Jan-Michael Frahm. Classification and reconstruction of surfaces from point clouds of man-made objects. In IEEE Int. Conf. Comput. Vis. Worksh., pages 257-263, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[43] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient ransac for point-cloud shape detection. Comput. Graph. Forum, 26(2):214-226, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[44] Gerard Sierksma and Yori Zwols. Linear and integer optimization: theory and practice. 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[45] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Int. Conf. Neur. Info. Process. Sys., volume 33, pages 7462-7473, 2020. 3, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[46] Marc Van Kreveld, Thijs Van Lankveld, and Remco C. Veltkamp. On the shape of a set of points and lines in the plane. Comput. Graph. Forum, 30(5):1553-1562, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[47] Dong-Ming Yan, Wenping Wang, Yang Liu, and Zhouwang Yang. Variational mesh segmentation via quadric surface fitting. Computer-Aided Design, 44(11):1072-1082, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[48] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. Capri-net: Learning compact cad shapes with adaptive primitive assembly. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11768-11778, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[49] Mulin Yu and Florent Lafarge. Finding good configurations of planar primitives in unorganized point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6367-6376, 2022, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.306, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[50] Qingnan Zhou, Eitan Grinspun, Denis Zorin, and Alec Jacobson. Mesh arrangements for solid geometry. ACM Trans. Graph., 35(4):39:1-39:15, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[51] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[52] Marco Zuliani, Charles S Kenney, and BS Manjunath. The multiransac algorithm and its application to detect planar homographies. In IEEE Int. Conf. Image Process., volume 3, pages III-153, 2005. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14180" + } + ] +] \ No newline at end of file diff --git a/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/4d750dca-9a8b-44f9-b884-62a2d6314ac8_origin.pdf b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/4d750dca-9a8b-44f9-b884-62a2d6314ac8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dd2abfbdf8ef30f7c7f72a93340342f5b0124b22 --- /dev/null +++ b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/4d750dca-9a8b-44f9-b884-62a2d6314ac8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b48227286307e27fb29b5b24861377b0f1dc3d0b9ff9a21bc68ddfdaa0b9ca8 +size 3568314 diff --git a/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/full.md b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d15067a8fac30e47f75a7c67a53929e36f9dcd0b --- /dev/null +++ b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/full.md @@ -0,0 +1,434 @@ +# Structure-Aware Surface Reconstruction via Primitive Assembly + +Jingen Jiang $^{1,2,3,4^*}$ Mingyang Zhao $^{1,5^*}$ Shiqing Xin $^{4}$ Yanchao Yang $^{6}$ + +Hanxiao Wang $^{1,3}$ Xiaohong Jia $^{2,3}$ Dong-Ming Yan $^{1,3\dagger}$ + +$^{1}$ MAIS & NLPR, CASIA $^{2}$ KLMM, AMSS, CAS + +$^{3}$ UCAS $^{4}$ Shandong University $^{5}$ BAAI $^{6}$ The University of Hong Kong + +![](images/7aa06ecab35991e406759e0990b1bbfce592b2d4f6ebad5f072439fb9b300141.jpg) +Figure 1: Surface reconstruction for measurement point clouds. In contrast to piecewise approximations (KSR), our method not only delivers manifold and watertight surfaces but also retains the primitive structures (planes, cylinders, spheres, cones, and tori) faithfully. Also, our method enables the recovery of the sharp geometry of objects (e.g., feature lines) as a byproduct. + +# Abstract + +We propose a novel and efficient method for reconstructing manifold surfaces from point clouds. Unlike previous approaches that use dense implicit reconstructions or piecewise approximations and overlook inherent structures like quadrics in CAD models, our method faithfully preserves these quadric structures by assembling primitives. To achieve high-quality primitive extraction, we use a variational shape approximation, followed by a mesh arrangement for space partitioning and candidate primitive patches generation. We then introduce an effective pruning mechanism to classify candidate primitive patches as active or inactive, and further prune inactive patches to reduce the search space and speed up surface extraction significantly. Finally, the optimal active patches are computed by a binary linear programming and assembled as manifold and watertight surfaces. We perform extensive experiments on a wide range of CAD objects to validate its effectiveness. + +# 1. Introduction + +The problem of converting unorganized point clouds into surface meshes, also known as surface reconstruction, has various applications in 3D vision, computer graphics, and architecture geometry. In recent decades, although many powerful algorithms have been devoted to pursuing faithful and accurate reconstructions [19, 21, 34, 28, 27, 49], it remains a challenging task due to the inherent ill-posed nature [11, 36, 5]. + +The popular Poisson reconstruction methods [24, 25] implicitly generate dense smooth surfaces for point clouds, while others mainly focus on giving a compact approximation to objects by piecewise planar shapes or polygonal meshes [7, 13, 3, 35]. However, man-made objects in real-world scenarios are typically composed of non-planar quadratic surfaces or quadric structures, such as the Computer-Aided Design (CAD) models (see Fig. 1 as an example). Therefore, reconstructing manifold surface meshes from measurement data meanwhile retaining their intrinsic primitive structures in principle shall generate higher-quality approximation results than solely relying on poly- + +gons. + +We argue that a method that solves the aforesaid problem ought to satisfy the following objectives: 1) Structure-awareness. The reconstruction surfaces should faithfully keep the structural features of the input point cloud, for instance, the common primitives in CAD, meanwhile correctly recovering the detailed geometry, and giving a high-fidelity approximation (e.g., the minimal Hausdorff distance) to the target object. 2) Manifold guarantees. The output surface meshes should be oriented 2-manifold and watertight [35]. 3) Efficiency. The developed algorithm should be scalable to large-scale point clouds with an acceptable computational complexity [3]. + +In this work, we develop a novel and efficient framework that takes the underlying geometric structures of point clouds into account when performing shape reconstruction. Our method ensures high-quality outputs and aims to achieve the above objectives. The core of our algorithm lies in a faithful reconstruction of intrinsic primitive structures presented in CAD models, as opposed to the direct surface approximation based on piecewise planar primitives. We consider five representative types of primitives commonly found in modern CAD systems, including four quadrics (planes, cylinders, cones, spheres) and tori. We adopt a variational shape approximation to optimize the extracted primitives followed by mesh arrangement techniques to attain primitive patches. A patch is a manifold that consists of a set of connected triangle meshes. As the size of patches is typically smaller than that of facets, our method enables a more stable and efficient reconstruction process than plane-based ones. Moreover, we propose a patch pruning mechanism to further speed up the consecutive surface extraction process. This mechanism distinguishes patches as either active or inactive, allowing us to perform fast local collision detection instead of a time-consuming global processing. Lastly, we select an optimal set of patches from the active candidate patches to assemble a reconstruction surface that respects the underlying structure of the object. To this end, we optimize a Binary Linear Programming (BLP) under hard constraints, as done in [35], to guarantee that the resulting surface is both manifold and watertight. + +We perform extensive experiments to validate the effectiveness of our algorithm and compare it with state-of-the-art approaches using a variety of CAD models. Experimental results evidence that our method outperforms competitors in terms of both approximation precision and reconstruction efficiency, while also preserving the intrinsic primitive structures of the input point clouds with much better fidelity. To summarize, the main technical contributions include: + +- An efficient structure-aware reconstruction method that produces accurate surface models while preserving the underlying intrinsic primitive structures of the + +input point clouds. + +- A patch pruning technique that avoids the time-consuming global collision detection and accelerates the surface extraction by a large margin. +- A patch-induced binary linear programming for assembling manifold and watertight surfaces from candidate active patches, which significantly reduces the search space and helps achieve efficient and accurate surface reconstruction. + +# 2. Related Work + +In this section, we briefly review the works that are most relevant to ours. Please refer to [5, 23] for a more comprehensive discussion. + +Geometric primitive extraction. Segmenting 3D point cloud models into a collection of geometric primitives such as planes and cylinders is helpful in most reconstruction approaches. Popular extraction frameworks include region growing [31, 39] and RANSAC [43, 52]. The former strategy takes the neighboring structure of models into consideration and propagates candidate primitives on their $k$ -nearest neighbor graph until the fitting error exceeds a preset tolerance. Instead, RANSAC adopts a hypothesis and verification mechanism to recover primitives that embrace the most inliers. In order to make full use of the a priori structure of the input model, Li et al. [29] started with primitive extraction by RANSAC, followed by the refinement subject to spatial relationships such as parallelism and orthogonality. This idea was further extended to extract planar primitives simultaneously [32] or sequentially [37]. Recently, Yu et al. [49] proposed a new algorithm for extracting planar primitives, which has powerful potential for plane-assembly reconstruction. In our work, we leverage an efficient region growing strategy [39] for primitive extraction followed by a refinement via variational shape optimization. + +Assembly-based reconstruction. The core idea of assembly-based methods is to integrate all extracted primitives into a hybrid mesh. Connectivity analysis and slicing are two principal assembly manners. Connectivity strategies [8, 42, 46] consider neighboring relationships between primitives and encode them into an adjacency graph, which is adapted to regularize output meshes. Slicing strategies [6, 7, 3, 33, 35, 13] extend the extracted primitives and slice them to construct a spatial partition for candidate facets. Connectivity strategies are typically efficient but suffer from linkage errors of adjacent graphs generated by incomplete inputs. Arikan et al. [1] remedied the incomplete part by a user-guided interactive method that is labor-intensive especially for large-scale measurement data, whereas Lafarge et al. [26] provided an automatic solution that first consolidated the point clouds and then executed a min-cut optimization to extract dense triangle meshes. + +![](images/dfc11adea0621a6c7e6ad414e238db5da038e88ec50b03180d7c8595eb623d4a.jpg) +Figure 2: Overview of the proposed method. (a) Point cloud after primitive extraction; (b) Accomplished space partitioning; (c) Active primitive patches after patch pruning; (d) The reconstructed surface. + +Alternatively, slicing strategies deal with incomplete models by means of extending primitives. Nan et al. [35] performed an exhaustive partition to compensate for all possible missing and thus guaranteed a manifold output via BLP. Nevertheless, this manner typically leads to high computational complexity for the subsequent surface extraction. To circumvent this shortcoming, Fang et al. [13] proposed a hybrid method that combines the connectivity analysis with the slicing process to make a limited partition, while Bauchet et al. [3] adopted a kinetic data structure [16, 15] to monitor the growing of convex piecewise planar polygons until they collide, which also results in a limited partition. Unlike previous approaches, we propose a simper partition by considering five types of primitives simultaneously rather than merely planar shapes. Moreover, we develop a propagation strategy inspired by Bauchet et al. [3] to further simplify and speed up the surface extraction. + +Learning-based reconstruction. With the advancements in deep learning, there have been several notable works that utilize learning-based techniques to reconstruct meshes. Voxel-inspired methods, such as those presented in references [38] and [10], train a network of models to convert voxels into implicit surface representations. To attain more accurate representations, SIREN [45] and DIGS [4] used signed distance functions as implicit surface representations, and then extracted meshes via marching cubes [30]. Points2Surf [12] employed a patch-based learning framework to directly attain meshes from point clouds without the need for normals, while DSE [40] adopted 2D Delaunay triangulations to ensure manifold meshes and BSP-Net [9] output compact polygonal meshes with arbitrary topology. However, most of these methods depend on implicit surface representations, which may not explicitly capture the underlying structures such as quadric primitives of objects. In addition, it is hard to perform model editing tasks for surfaces generated by implicit reconstructions. + +# 3. Methodology + +The proposed method consists of three major steps as illustrated in Fig. 2. Given an input point cloud $\mathcal{P} = \{p_j\in$ + +$\mathbb{R}^3\}_{j=1}^N$ (Fig. 2(a)), we first partition $\mathcal{P}$ by a set of proxy meshes defined by the extracted geometric primitives and then apply a mesh arrangement operation to get candidate manifold patches (Fig. 2(b)) in Section 3.1. Later, a fast propagation mechanism is presented in Section 3.2 is used to prune patches that are inactive (Fig. 2(c)) to accelerate the subsequent surface extraction procedure. Finally, we deliver a manifold and watertight output surface (Fig. 2(d)) by formulating a patch-induced optimization under the BLP framework in Section 3.3. + +# 3.1. Space Partitioning + +The goal of this step is to generate a finite space partitioning by arranging proxy meshes of geometry primitives extracted from the input point cloud $\mathcal{P}$ . + +Primitive extraction. We first utilize an efficient region growing strategy introduced in [39] to attain the rough segmentation components $\mathcal{S} = \{s_i\subset \mathbb{R}^3\}_{i = 1}^M$ and the corresponding primitives $\mathcal{Q} = \{q_i\subset \mathbb{R}^3\}_{i = 1}^M$ from $\mathcal{P}$ , where $q_{i}$ is the supporting primitive of $s_i$ inversely, each point $p_j\in s_i$ is the supporting point of $q_{i}$ . Let $\mathcal{L} = (\mathcal{S},\mathcal{Q})$ be the initial layout of the extracted primitives, we further refine $\mathcal{L}$ via the variational shape approximation based on the $\ell^2$ and the $\ell^{2,1}$ metrics [11, 47], which aims to optimize the following expression + +$$ +\min _ {\mathcal {L}} \mathcal {F} (\mathcal {L}) = \sum_ {i = 1} ^ {M} \sum_ {p _ {j} \in s _ {i}} \operatorname {d i s t} \left(p _ {j}, q _ {i}\right), \tag {1} +$$ + +where $dist(p_j, q_i)$ stands for the hybrid metric that measures the cost of the point $p_j \in s_i$ with respect to the corresponding primitive $q_i \in \mathcal{Q}$ in terms of both the spatial $\ell^2$ distance and the normal $\ell^{2,1}$ deviation. Fig. 2(a) presents the primitive extraction result based on five types of primitives that we are currently considering. Noted that the proposed framework is not constrained by the above detected primitive types. Instead, it can be integrated with powerful detection algorithms that support a wider variety of primitive types, enabling more complex reconstructions with better accuracy. + +![](images/6ac4b1d6dc82edafb4c43521b167161c1151b6a4d8ac1901070154f791e74c6a.jpg) +Figure 3: Visual description of patches and (non-) boundary curves. (a) $\mathcal{C}_{ij}$ is the non-boundary curve generated by patches $m_{i}$ and $m_{j}$ , while $\bar{\mathcal{C}}_i$ and $\bar{\mathcal{C}}_j$ are the boundary curves of $m_{i}$ and $m_j$ , separately. (b) For each non-boundary curve, there are four patches around it. + +Proxy mesh arrangement. After the optimization of the extracted primitives, we generate the discrete approximation to all primitives via proxy meshes. These meshes are trimmed by the bounding box of $\mathcal{P}$ and then all meshes are processed by a mesh arrangement [50]. The mesh arrangement results are a set of candidate primitive patches $\mathcal{M} = \{m_i\subset \mathbb{R}^3\}_{i = 1}^K$ as illustrated in Fig. 2(b), which partition the current working space into a series of cells that form a closed manifold without self-intersections. Here each patch $m_{i}$ represents a subset of the input meshes and is the maximal connected set of triangle meshes as illustrated in Fig. 3. Due to the connectivity of patches, each $m_{i}$ may embrace several non-boundary curves $\mathcal{C}_{ij}$ (colored red in Fig. 3(a)) that are shared with other primitive patches, which are defined as + +$$ +\mathcal {C} _ {i j} = m _ {i} \bigcap m _ {j} (i \neq j), m _ {i}, m _ {j} \in \mathcal {M}, \tag {2} +$$ + +whereas at most one boundary curve $\bar{C}_i$ (colored green in Fig. 3(a)). For example, there are in total of four neighboring patches that share the same $C_{ij}$ in Fig. 3(b). Since we perform space partitioning based on primitive patches rather than discretizing each of them as multiple piecewise planes, the developed algorithm indeed is several orders of magnitude faster than previous plane-induced approaches. The efficiency superiority of our method is also quantitatively validated by the following experiments. + +# 3.2. Pruning of Candidate Primitive Patches + +It is known that the computational complexity of the ultimate surface extraction in shape reconstruction regarding the size of patches is exponential [44]. To expedite the subsequent optimization process, we further propose a pruning mechanism which effectively circumvents the time-consuming and memory-intensive global collision detection to classify current primitive patches as active and inactive, from which we prune inactive ones to reduce the search space for surface extraction. As illustrated in Fig. 2, the significant time gap between the two routes(black and red) demonstrates the effectiveness of our pruning mechanism. The proposed procedure is composed of two steps, initialization and pruning, as presented subsequently. + +# Algorithm 1: Pseudo-code of the propagation process + +Input: Priority queue $Q$ of representative triangles. + +Output: Active primitive patches $\tilde{\mathcal{M}}$ + +1 Initialize propagation cost of each triangle as $c = 0$ . + +2 while $Q\neq \emptyset$ do + +3 $\mid (t,c)\gets Q.pop()$ and $\hat{t}\leftarrow t$ + +4 for each triangle edge $e\in \hat{t}$ do + +5 for triangle $t_e\cap t = e$ do + +6 | | if $\text{prim}(t_e) == \text{prim}(t) \& t_e$ is + +inactive & $t_e$ is not blocked then + +7 Compute the height $h_e$ of $t_e$ on $e$ . + +8 $Q$ .push((te,c+he)). + +Initialization. To describe the point distribution of the input point cloud as fidelity as possible, we first select a set of representative triangle meshes denoted as $\mathcal{I}$ from $\mathcal{M}$ by means of the triangle coverage $Cover(t)$ which is defined as + +$$ +\operatorname {C o v e r} (t) = \frac {\operatorname {V a l i d} (t)}{\operatorname {A r e a} (t)}, \tag {3} +$$ + +where $\text{Area}(t)$ stands for the area of a triangle mesh $t$ and $\text{Valid}(t)$ is formulated as + +$$ +V a l i d (t) = \sum_ {p \in \mathcal {P} _ {t}} 1 (\| p - p ^ {\prime} \| _ {2} < \epsilon) a _ {p}, \tag {4} +$$ + +where $\mathcal{P}_t\subset \mathcal{P}$ encompasses the measurement points that their projections are attached to $t$ $\| p - p^{\prime}\|_{2}$ denotes the $\ell^2$ Euclidean distance from $p$ to its projective point $p^\prime$ on $t$ + +1 is the indicator function, and $a_{p}$ is the corresponding geodesic Voronoi area of $p$ , which is calculated as in the inset: $p$ and its $k$ -nearest neighbors (i.e., hard red points) are first projected to the supporting plane $\Omega_t$ + +![](images/15deff45d25e82bb7cc5e77f6ad8d2ee4b50c41adac1751201a339de7ca3b510.jpg) + +where its attached triangle mesh $t$ lies on, then $a_{p}$ is equivalent to the area of the 2D Voronoi cell [2]. $t$ is termed representative if its Cover $(t)$ surpasses a user-specified value. We populate all representative meshes of $\mathcal{I}$ into a priority queue $Q$ with an ascending order. + +Pruning. Next, representative triangle meshes are propagated to enable a limited partition to speedup the final surface extraction. To this end, all triangles of $Q$ are initially labeled as inactive prior to entering the loop of the priority queue. For each round during the loop, we first pop the top element, for instance, $(t_i, c_i)$ , out of the priority queue and prompt $t_i$ to be an active state $\hat{t}_i$ . The propagation cost $c_i$ of a triangle mesh $t_i$ is defined as the height + +$h_i$ from its vertex to the base edge shared with another adjacent triangle. Sequentially, we propagate all the triangles that share both the same edge and the same supporting primitive with $t_i$ and pop the update elements out of $Q$ . For instance, for edges that are not located on non-boundary curves, there are at most three update elements $(t_{i1},c_i + h_1)$ , $(t_{i2},c_i + h_2)$ , $(t_{i3},c_i + h_3)$ , whereas, for edges that coincide non-boundary curves, we perform a fast local collision detection to check whether there exist active meshes, if yes then we evoke barricades and stop the propagation process. The above propagation proceeds until $Q$ becomes empty. We summarize the proposed propagation process in Algorithm 1 to make it more understandable. Consequently, a primitive patch is active only if its triangles are all active, otherwise, it is still inactive. We use $\tilde{\mathcal{M}}$ to represent the set of active patches. The patches in Fig. 2(b) after pruning are shown in Fig. 2(c). As observed, the size of candidate patches has been significantly reduced compared with Fig. 2(b). + +Like [3], we also allow for the propagation crossing primitive patches $\mathcal{G}$ times (with $\mathcal{G} = 2$ in default) at most. This can deal with small parts that may have been missed during the primitive detection. However, in contrast to previous global collision detection manners [3], our local strategy preserves much better the propagation efficiency. + +# 3.3. Surface Extraction + +Finally, we adopt a Binary Linear Programming (BLP) analogous to [35] to extract the surface model from $\tilde{\mathcal{M}}$ . Specifically, the state of patches in $\tilde{\mathcal{M}}$ is encoded as $\mathcal{X}_{\tilde{M}} = \{x_{\tilde{m}}|\tilde{m}\in \tilde{\mathcal{M}},x_{\tilde{m}}\in \{0,1\} \}$ , where the binary discrete variable $x_{\tilde{m}}$ indicates whether the patch $\tilde{m}$ is selected $(x_{\tilde{m}} = 1)$ or not $(x_{\tilde{m}} = 0)$ . To ensure the resulting surface manifold and watertight, we constrain the curves of the selected patches to be all non-boundary, and thus the two patches sharing the same non-boundary curve are either selected at the same time or both not selected. Consequently, our objective function is defined as + +$$ +\begin{array}{l} \min _ {\mathcal {X} _ {\tilde {\mathcal {M}}}} \mathbf {U} (\mathcal {X} _ {\tilde {\mathcal {M}}}) + \lambda \mathbf {B} (\mathcal {X} _ {\tilde {\mathcal {M}}}), \\ s. t. \left\{ \begin{array}{l} \sum_ {\tilde {m} _ {i} \cap \tilde {m} _ {j} \neq \emptyset , i \neq j} \left(x _ {\tilde {m} _ {i}} + x _ {\tilde {m} _ {j}}\right) = 0 o r 2, \tilde {m} _ {i}, \tilde {m} _ {j} \in \tilde {\mathcal {M}} \\ x _ {\tilde {m} _ {i}} \in \{0, 1 \}, \quad i = 1, 2, 3 \dots | \tilde {\mathcal {M}} |. \end{array} \right. \tag {5} \\ \end{array} +$$ + +where $\mathbf{U}(\mathcal{X}_{\tilde{\mathcal{M}}})$ and $\mathbf{B}(\mathcal{X}_{\tilde{\mathcal{M}}})$ are penalty terms regarding the patches and the neighboring patch pairs, respectively, and $\lambda$ is a balance coefficient. Moreover $\mathbf{U}(\mathcal{X}_{\tilde{\mathcal{M}}})$ encourages the selection of active patches that are with high fidelity in the form of + +$$ +\mathbf {U} \left(\mathcal {X} _ {\tilde {M}}\right) = \sum_ {\tilde {m} \in \tilde {\mathcal {M}}} x _ {\tilde {m}} \cdot \left(\frac {\mathcal {A} _ {\tilde {m}} - \bar {\mathcal {A}} _ {\tilde {m}}}{\mathcal {A}} - \frac {| \mathcal {P} _ {\tilde {m}} |}{| \mathcal {P} |}\right). \tag {6} +$$ + +Here $\mathcal{A}_{\tilde{m}}$ and $\bar{\mathcal{A}}_{\tilde{m}}$ denote the area and the coverage area of the patch $\tilde{m}$ , respectively. The coverage area of $\tilde{m}$ is the summation of the coverage area of all its triangles. $\mathcal{A}$ is a normalization factor equivalent to the sum of the area of all patches. $\mathcal{P}_{\tilde{m}}$ is the set of points whose projections are locating in $\tilde{m}$ . On the other hand, $\mathcal{B}(\tilde{\mathcal{X}})$ is designed to control the model complexity. Specifically, it put penalties to the non-boundary curve length of two patches: + +$$ +\mathcal {B} \left(\mathcal {X} _ {\tilde {\mathcal {M}}}\right) = \frac {1}{L} \sum_ {i, j} \left| \mathcal {C} _ {i j} \right| \cdot x _ {\tilde {m} _ {i}} \cdot x _ {\tilde {m} _ {j}}, \tag {7} +$$ + +where $|\mathcal{C}_{ij}|$ denotes the length of the curve shared by two active patches $\tilde{m}_i$ and $\tilde{m}_j$ with different supporting primitives and $L$ is the sum of the length of all curves. Fig. 2(d) presents the final reconstruction surface by BLP, which not only generates a manifold and watertight shape but also successfully preserves the intrinsic primitive structures. + +# 4. Experimental Evaluations and Discussions + +Implementation details and metrics. In this section, we execute extensive experiments to validate the advantages of the developed method in terms of both reconstruction precision and efficiency and make various comparisons with state-of-the-art approaches. Like [3], the precision is quantitatively evaluated by the symmetric mean Hausdorff distance (SMH) between the measurement point clouds and the reconstruction meshes (values are normalized by the diagonal length of the models' bounding box and multiplied by a factor of $10^{2}$ ), while the efficiency is assessed by the running time. Our algorithm is implemented in $\mathrm{C}++$ using the geometry processing library libigl [22]. We adopt the mathematical programming solver Gurobi [18] to optimize Eq. 5. All reconstruction results are rendered by Blender [14] for better visualization. Our experiments are conducted on a desktop with Intel(R) Core i7-7700K CPU@4.2GHz and 48GB RAM. + +Reconstruction effect. Our algorithm has demonstrated good reconstruction results from point clouds to manifold meshes on various CAD objects as illustrated in Fig. 4. The input point clouds presented in Fig. 4(a)-(h) are either sampled from real-world CAD models or courtesy by publicly available benchmark datasets [41, 51]. As observed, there are multiple explicit intrinsic primitive structures in each model. Instead of approximating them directly by piecewise planar primitives, our algorithm is capable of reconstructing the intrinsic primitives faithfully even for quite challenging or complicated models. For instance, the rich spherical and conic structures existing in the Minarets model of Fig. 4(b) require careful treatments, yet our method is still able to achieve highly accurate reconstructions while preserving all structural details successfully. At the same time, + +![](images/f85af81be8a05e864169d8870b1cdc1767ece2c194aa1ae05a6250ad1a66fc70.jpg) +(a) + +![](images/1f373b54946d17a8c3aa6de1cd72d62257d78d03eb262a5013d9d7febdc1c7f8.jpg) + +![](images/08daddc0b9859a06af6152337881d56f1d858844efd2f5493644e5cc840026e8.jpg) + +![](images/fa0d88f40c136778575bf627483cf752b758c205d64cd5f76eb2ae23e575ba9f.jpg) + +![](images/88cbd7dd0e889fd8b501f761eb9ba5c5ba36a5c012f3a3c036d6cca41db981e1.jpg) +(b) + +![](images/f8019e35731c67bcde3cbb1b0d030957d555c91a1550d07038a036b6caca6531.jpg) + +![](images/aa73e1fb03136d1e332835c4433b54b3d69fdfdccc95592b71872df16d17d139.jpg) + +![](images/2967a618cca8641a9a671c62701495475ea478c6cf13309a5ac2fe73926b236d.jpg) + +![](images/7e9e70d971760a48c435f4320ec6f96082a8e59725a74e52b7543f7c28c329b4.jpg) + +![](images/58932b0e464b7fcf3484769d1081aff2693565d0118a5f2875dbd0b7ea8dc4a9.jpg) +(c) + +![](images/182f55ad874a7419534daea35c046eae52a9f86c14ef3068f5c954011027bf58.jpg) + +![](images/9bfba2f70dcbda172919956545615a66128738877a130fc1a740842026df6908.jpg) + +![](images/ae38cdbf35efdaf43afbb707680612d62c99a5f93b040f2912a10443e94b82d3.jpg) + +![](images/e3c329c267ec47063e992bd864771e3465e1aa6a8df30eedda3086e47ed88930.jpg) + +![](images/d41c99358da7634afc561c82f3d35bad3672589582364e469b29bed4ee5100f0.jpg) +(d) + +![](images/55e5bf41ae7a5bf3fcb59e7da57f8f2565ccd0c7a12d8442fb45adc51c4cb425.jpg) + +![](images/4b877af64b5cf1e89128997cc67c3d9b04f45b6e0c5c6721b107c41e3f1a7fa2.jpg) + +![](images/bbba14f044722479a8c9f85e463b731a4e77664879af5d56a192a188485b5353.jpg) + +![](images/6d24df649ec78854124ab95822f905269b0edeb2fd56d1ee27473da28aba25b7.jpg) + +![](images/789af587d81aa0e8acd3875b3a72498c8546ee6609f09e567fd03df8c67ec8cd.jpg) + +![](images/eb71a968918d026e27e4d439b334cf9f711c9b7f82a4ea4b7e967c931a2930df.jpg) +(e) + +![](images/2670d38ec9e351c5fa434d4ed125ba73169ddfb1e708e7d34eb6d46300584b92.jpg) + +![](images/8ac5bba0b0cde9dbd34fd7404f1990a973222237f5e4f11c73d806e83739a1f3.jpg) + +![](images/5eac263144508524a4b1ccad2874299b45c7bd5bd3370ba8c34c8b1b0521b97d.jpg) + +![](images/95fe49212056e029bd1108c91c4a2072d1e36bf6d2205edc1702f33ea129d969.jpg) + +![](images/a03926641a3d0832715b029625d1d4592fdd2c8e55e7126c81938df299e8fef3.jpg) +(f) + +![](images/a61610d427f31e424114c0c76f3bb01bf67e4529026fb108b01b901db0e7ab4b.jpg) + +![](images/f7d4e3ff452e045f7f6588caa4dc6a6d03466b11111c6bdb7a309839d624a696.jpg) + +![](images/6fbb3642d226ad5c1b315b86ca1958524619e8c2c8e03efc01e37deee24bd3ff.jpg) + +![](images/b231689cc0488831320d8e5555754c1f19b89ad49c20cc851eeab05f8428c507.jpg) + +![](images/fa9df84afdc972e8e5a3f6861e6bdf37f5df4c62fc5d57a45893efe226b96885.jpg) +(g) + +![](images/93779c0fca2d7581e66860f73101053460b52a9f5fa58c1d5cc83867622bdb0a.jpg) + +![](images/2f0b6540f330e6f9043af10d05208d0436df583f5e6472784079f4fe218b9435.jpg) + +![](images/b1a2123c3fd365fe3e9bd4dc2a81c4f995abf803da08c3808feb79365aedca25.jpg) + +![](images/3ebc44a8c908f54c0a7a4f1f55da6f582985410a9d1484708827731343727235.jpg) + +![](images/edcebd04c2c6e64671b61ed83c46a8ff1a92207e917aebc255e689f9991b7ff9.jpg) +(h) + +![](images/6e9640d3f5591aef29042483bcaefdd2752c5ed22091eb3affcc66b03deb37ee.jpg) + +![](images/82627de130599f1c3fdb5149160845fb9c732551b22bf41aaa525932716c96e1.jpg) + +![](images/a2b238be5790533745a1c14ee1f21b69d2a1a17226afba8eee161e699824a390.jpg) + +![](images/c5a75d12d6f8e95d46384ec27de0fe94019fe0b0897b5a88124fa29c459d83c8.jpg) + +![](images/f117fa0e6923e2a02ea3430a2447ef94d4f34190683cb23ec6b64bae8813e50e.jpg) +Figure 4: Reconstruction results by the proposed method. From left to right: Input measurement point clouds, extracted geometric primitives (different colors stand for different primitives), output reconstruction models, recovered sharp geometry along with the original point clouds, and colored models showing the error distribution regarding the Hausdorff distance. Our method produces high-quality reconstructions and preserves the intrinsic primitive structures as well as the sharp features well. + +our method produces high-quality sharp feature lines attained by patch intersections, as shown in the forth column of Fig. 4. This is an important advantage for many practi + +cal applications such as sketch-style rendering, as it ensures that the reconstructed object retains its original shape and does not appear distorted or blurry. The fifth column of + +Table 1: Statistical results of the quantity and the runtime (in second) during reconstruction of the CAD models in Fig. 4. + +
Models Metrics(a)(b)(c)(d)(e)(f)(g)(h)
# Points of the input point cloud10k593k10k10k74k7k16k14k
# Extracted primitives2853212034322927
# Generated candidate patches36466024103072349704177501092
# Active patches after pruning224533697414172586342472667
Patch decrease (%)38.42644.07328.05842.32447.96817.98637.06738.919
# Determined active patches of the output model605915191935339986114
Mesh arrangement (s)7.85321.7522.4631.7275.2362.9355.4311.421
Patch pruning (s)0.92715.4210.7920.4763.0430.4881.1310.726
Surface extraction (s)1.2882.0860.6590.0911.1120.1570.3140.137
+ +![](images/0fa71908fe31c1fc14a0f0b0e2716f8dba99fccb06c04cd220df91c4e9d9c12e.jpg) + +![](images/aea85b175c3958133c9f8ab856700231ba14dca5949a652fa66afa353d2f86fd.jpg) +$\sigma = 0.01(0.82m)$ $\mathrm{SMH} = 0.11\%$ + +![](images/9976b643d3d0b5c3f807bf2a9af81e9c65abd30aec58072e9bbdd3faea7d0b5f.jpg) + +![](images/8cf39710a210f9d567ee3cb24177cc45abd17a4083783cc255105801ac8120be.jpg) +$\sigma = 0.02$ $(1.64\mathrm{m})$ $\mathrm{SMH} = 0.11\%$ + +![](images/73ac6ed6fee6cbff3a9e0c9fb855f1dd847eb0178e1f765678777f0be2da5f34.jpg) + +![](images/fe7226a44429187dbb1cc7c4e190a4e6d323731e387c6f4c40e0ffa80cb62fd8.jpg) +Figure 5: Robustness test against noise disturbance. Top: Input point clouds with increasing noise. Bottom: Our reconstruction results along with the noise level $\sigma$ and the reconstruction error SMH. + +![](images/6f0f924ea69fc10392ed3f4224354d9bced3241286753adbacffcb9d902bd949.jpg) + +![](images/aa2f83ddf29487e253687ad986a039ea7c395d57eca64814dd67e8a7218c560c.jpg) +$\sigma = 0.04(3.28m)$ $\mathrm{SMH} = 0.12\%$ +$\sigma = 0.08(6.56m)$ $\mathrm{SMH} = 46\%$ + +Statistical results. Table 1 summarizes the quantitative results of our proposed method for the models in Fig. 4. Our algorithm significantly reduces candidate patches, especially for complex models like Fig. 4(b),(d),(e), where the patch decrease ratio exceeds $40\%$ . In terms of efficiency, patch pruning and surface extraction outperform mesh arrangement, and further improvements can be made by exploring a more efficient intersection mechanism. + +Robustness assessment. We also assess the ability of the proposed algorithm against noise disturbance and occlusion data. As depicted in Fig. 5, we increase the standard deviation $\sigma$ of Gaussian noise from $0.01(0.82m)$ to $0.04(3.28m)$ , where values in parentheses are real intensity relative to the bounding sphere's radius of the model. Our method consistently generates accurate reconstructions in terms of both topological structures and approximation distance (SMH), suggesting its high resilience to noise. Even when $\sigma = 0.08(6.56m)$ , which has caused significant topological alteration of the original point cloud, our method still maintains the overall structures of the object. Additionally, we verify the robustness of our method to occlusion + +![](images/fe77783fb2c125a78cfeb45ea320f2e49a9a7419540611125ca4edd63f3e197b.jpg) + +![](images/56f2a99d447e3865ab2fe4a231085d6259bc7c207e801d8d8b6df1262ed1b405.jpg) +Fig. 4 visualizes the reconstruction error of our algorithm with respect to the Hausdorff distance, where our method consistently produces accurate approximations for all objects, demonstrating its stability and effectiveness. +Figure 6: Robustness test against occlusion data. Top: Input point clouds with increasing occlusions. Bottom: Our reconstruction results along with the reconstruction error. + +![](images/6c7edd5ef763cd9cbd1e758fa3c09cf01fc1eddeb94e0ec0a78411df7ad33d17.jpg) + +![](images/56b8ea4ef2dd450b6413795466611f6277dfa45a64b253611883ce0dda4b10fa.jpg) + +![](images/9d999a2334889471094c890a426e2776c6ab8cc66af128e1ac2dcf34fa99effc.jpg) + +![](images/160dd7fbaaa67577e88f48d686662c89e0fa3018345e1bea6c1ff9bbcec27368.jpg) + +![](images/6f562a790c01adba344ce0c9f25dbb7ff8244f1057c1c2b8f272634c9093f976.jpg) + +![](images/cd5e33066618cdcfb154a6629f2b116b6be67e3537510ab2ad7a912269160573.jpg) + +Table 2: Comparisons with state-of-the-art approaches, where "w" and "w/o" indicate with and without the patch pruning, separately. + +
MetricsiPSR[20]SIREN[45]DSE[40]KSR[3]Ours
ww/o
SMH (%)4.2341.793.15169.442.452.45
Time (s)201.5891208.124303.865570.46845.556.777
+ +sion data. Fig. 6 reports our reconstruction results after progressively erasing point clouds. As observed, our method shows remarkable resistance to occlusions and has successfully compensated the missing structures. Conversely, this is also quite helpful to complete the imperfect point clouds by sampling points in the reconstruction models. + +Comparisons. Subsequently, we conduct a comparative analysis of our method with representative state-of-the-art approaches, including densely implicit reconstruction methods iPSR [20], DSE [40], and SIREN [45], as well as the plane-assembly approach KSR [3]. The quantitative comparison results in terms of reconstruction accuracy and efficiency on the Cheese model is presented in Table 2. The results demonstrate that our method achieves competitive reconstruction accuracy with implicit approaches iPSR and DSE, and outperforms SIREN and KSR by a large margin. Additionally, our algorithm performs significantly faster than competitors, even without patch pruning. The qualitative reconstruction results presented in Fig. 7 show that benefiting from the introduced primitive assembly, our + +![](images/fa05c046cb730132e2746dd8aaa5698e2d6ba0543347c305a725933bada9fdde.jpg) +Figure 7: Qualitative comparisons with state-of-the-art approaches. Our method enables the shape reconstruction faithfully and successfully recovers the sharp geometry (e.g., feature lines and corners) from unorganized point clouds. + +![](images/e125e2ff80a78fcad8357fde7b939c0d75b2578658b2e5636d226f53b4fdb894.jpg) + +![](images/e30e0715d4535abacf4766dabea16958dec36614113714d703520e56b961ec08.jpg) + +![](images/6ff753b28d8961edc3e10ec28c5a722c86809e864e443e4df95e37e8f16b0ea5.jpg) + +![](images/d6f528e3486add7164707d83501c092377ac2a1d3b2c7a095445b4310010408b.jpg) + +![](images/d305bfbd766ef046c7b2e150e11aa61e6e9a8d3b232cd17b767ed0b0b8adb2ed.jpg) + +Table 3: Quantitative comparisons on the ABC300 dataset. + +
Methods MetricsiPSR[20]SIREN[45]DSE[40]KSR[3]ComplexGen[17]CAPRI-Net[48]Ours
SMH (%)2.7124.894.0626.4324.2415.874.81
Time (s)24.0641565.21732.8004.208553.089183.8443.728
+ +method successfully recovers sharp features from unorganized point clouds, which is a salient advantage compared to implicit methods. + +We also conduct quantitative comparisons with several predecessors (additionally including CAPRI-Net [48] and ComplexGen [17]) on 300 CAD models from the ABC dataset (abbreviated as ABC300). Table 3 demonstrates that our method achieves consistently competitive reconstruction accuracy with implicit approaches such as iPSR and DSE (while also providing advantages for sharp feature recovery and model editing), and significantly surpasses KSR, CAPRI-Net, and ComplexGen. Furthermore, our method is considerably faster than its competitors. + +![](images/57819ac51dffa0a3feb27fb698e57b63453f2902e65d4a29efa06c8dd692a768.jpg) + +![](images/440759cec00bc591d814faf9e4ee642d43fcd342b2701449ebd78f4e3e90166a.jpg) +Figure 8: Model editing based on our structure-aware reconstruction algorithm. (a) The original reconstruction object. (b)-(d) The newly edited results. + +![](images/187111fa2dadba5ad3b66c877006ebfc6b0f0060d7828af5bb1df6f0533aaced.jpg) + +![](images/e579ee14a8f29a2453a7ac2bf5980b79cca03347e6145953e8a4331160d3c12e.jpg) + +Balance parameter. We investigate the impact of $\lambda$ on surface extraction by varying its values to different settings (see Fig. 9(a)). It can be seen that increasing $\lambda$ brings the reduction of the complexity of the output model. However, this also degrades the reconstruction accuracy. For example, when $\lambda$ is set to 0.7, the inner hollows of the input point clouds disappear. + +Application to model editing. Our reconstruction method can be employed for flexible model editing by taking advantage of the structure-aware merits. As shown in Fig. 8, since each quadratic structure of the object is preserved along with their detailed geometric parameters, one + +![](images/4df7b8eafaa7e27a489eeaaa56cd3e2e193162ae3329ba3a0911fe85537a998e.jpg) +(a) +Input +Figure 9: (a) Impact of parameter $\lambda$ . As $\lambda$ is increased, the complexity of the output model decreases. (b) In sparse input scenarios, small primitives can be easily lost, leading to failures. + +![](images/0569f070d0f6812f0c8582ba61da016e8eae8366895c3e07ad94a5656a4cb85e.jpg) +$\lambda = 0.2$ + +![](images/05474ab7c9720fa731cd3b47c9a8d6f53ad8eb32e3e4c51809f4e45808322281.jpg) +$\lambda = 0.7$ + +![](images/5267e588ffc774d6484e2e637d88e441f6b8adc46dc6f3dbaea560c34f7a991d.jpg) +#Point=10k + +can modify the reconstruction shape by simply specifying various parameters. For instance, by re-defining the radius or height of the cylindrical primitives in Fig. 8(a), we can resize them to our desirable shapes (Fig. 8(b)-(d)). + +Limitations. Our approach primarily focuses on the rapid and accurate reconstruction of representative primitive structures in CAD models. However, as shown in Fig 9(b), our reconstruction quality may deteriorate for inputs lacking explicit primitives. Additionally, inadequate primitive extraction due to severe occlusion or noise can degrade the quality of the reconstruction results. + +# 5. Conclusion and Future Work + +We propose an accurate and fast shape reconstruction pipeline for converting disordered 3D point clouds into manifold, watertight surfaces. Our method preserves representative intrinsic primitive structures in CAD systems through primitive assembly, avoiding dense implicit reconstructions and extending piecewise planar approximations. We also develop an effective patch pruning mechanism, greatly expediting surface extraction by reducing patch intersections. Additionally, our method is highly resilient to noise and can reliably recover sharp object geometry. Extensive experiments show our algorithm's superiority over competing methods. + +In our future work, we plan to expand the current primitive types to general primitives, and even spline surfaces, in order to efficiently process more complex CAD models. + +# Acknowledgements + +This work was partially funded by the National Key Research and Development Program (2021YFB1715900), the CAS Project for Young Scientists in Basic Research (YSBR-034), the National Natural Science Foundation of China (62172415, 62272277, 12022117), and the HKU-100 Research Award. + +# References + +[1] Murat Arikan, Michael Schwärzler, Simon Flöry, Michael Wimmer, and Stefan Maierhofer. O-snap: Optimization-based snapping for modeling architecture. ACM Trans. Graph., 32(1):6:1-6:15, 2013. 2 +[2] Gavin Barill, Neil G Dickson, Ryan Schmidt, David IW Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. ACM Trans. Graph., 37(4):43:1-43:12, 2018. 4 +[3] Jean-Philippe Bauchet and Florent Lafarge. Kinetic shape reconstruction. ACM Trans. Graph., 39(5):156:1-156:14, 2020. 1, 2, 3, 5, 7, 8 +[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3 +[5] Matthew Berger, Andrea Tagliasacchi, Lee M Seversky, Pierre Alliez, Gael Guennebaud, Joshua A Levine, Andrei Sharf, and Claudio T Silva. A survey of surface reconstruction from point clouds. Comput. Graph. Forum, 36(1):301-329, 2017. 1, 2 +[6] Alexandre Boulch, Martin de La Gorce, and Renaud Marlet. Piecewise-planar 3d reconstruction with edge and corner regularization. Comput. Graph. Forum, 33(5):55-64, 2014. 2 +[7] Anne-Laure Chauve, Patrick Labatut, and Jean-Philippe Pons. Robust piecewise-planar 3d reconstruction and completion from large-scale unstructured point data. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1261-1268, 2010. 1, 2 +[8] Jie Chen and Baoquan Chen. Architectural modeling from sparsely scanned range data. Int. J. Comput. Vis., 78(2):223-236, 2008. 2 +[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 3 +[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5939-5948, 2019. 3 +[11] David Cohen-Steiner, Pierre Alliez, and Mathieu Desbrun. Variational shape approximation. ACM Trans. Graph., 23(3):905-914, 2004. 1, 3 +[12] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In Eur. Conf. Comput. Vis., pages 108-124, 2020. 3 +[13] Hao Fang and Florent Lafarge. Connect-and-slice: an hybrid approach for reconstructing 3d objects. In IEEE Conf. + +Comput. Vis. Pattern Recog., pages 13490-13498, 2020. 1, 2, 3 +[14] Blender Foundation. Blender - a 3D modelling and rendering package. http://www.blender.org.5 +[15] Leonidas Guibas. Kinetic data structures. In Handbook of Data Structures and Applications, pages 377-388. 2018. 3 +[16] Leonidas Guibas, Lyle Ramshaw, and Jorge Stolfi. A kinetic framework for computational geometry. In IEEE Annual Symposium on Foundations of Computer Science, pages 100-111, 1983. 3 +[17] Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. Complexgen: Cad reconstruction by b-rep chain complex generation. ACM Trans. Graph., 41(4):129:1-129:18, 2022. 8 +[18] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual. https://www.gurobi.com.5 +[19] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. In Proc. ACM SIGGRAPH, pages 71-78, 1992. 1 +[20] Fei Hou, Chiyu Wang, Wencheng Wang, Hong Qin, Chen Qian, and Ying He. Iterative poisson surface reconstruction (ipsr) for unoriented points. ACM Trans. Graph., 41(4):128:1-128:13, 2022. 7, 8 +[21] Hui Huang, Dan Li, Hao Zhang, Uri Ascher, and Daniel Cohen-Or. Consolidation of unorganized point clouds for surface reconstruction. ACM Trans. Graph., 28(5):1-7, 2009. 1 +[22] Alec Jacobson and Daniele Panozzo. Libigl: Prototyping geometry processing research in c++. In SIGGRAPH Asia 2017 courses, pages 1-172. 2017. 5 +[23] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3d data. Comput. Graph. Forum, 38(1):167-196, 2019. 2 +[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Comput. Graph. Forum (Proceedings of SGP), page 61-70, 2006. 1 +[25] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. Graph., 32(3):29:1-29:13, 2013. 1 +[26] Florent Lafarge and Pierre Alliez. Surface reconstruction through point set structuring. Comput. Graph. Forum, 32(2pt2):225-234, 2013. 2 +[27] Minglei Li, Liangliang Nan, Neil Smith, and Peter Wonka. Reconstructing building mass models from UAV images. Comput. Graph., 54:84-93, 2016. 1 +[28] Minglei Li, Peter Wonka, and Liangliang Nan. Manhattanworld urban reconstruction from point clouds. In Eur. Conf. Comput. Vis., pages 54-69, 2016. 1 +[29] Yangyan Li, Xiaokun Wu, Yiorgos Chrysathou, Andrei Sharf, Daniel Cohen-Or, and Niloy J Mitra. Globfit: Consistently fitting primitives by discovering global relations. ACM Trans. Graph., 30(4):52:1-52:12, 2011. 2 +[30] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proc. ACM SIGGRAPH, volume 21, pages 163-169. 3 + +[31] David Marshall, Gabor Lukacs, and Ralph Martin. Robust segmentation of primitives from range data in the presence of geometric degeneracy. IEEE Trans. Pattern Anal. Mach. Intell., 23(3):304-314, 2001. 2 +[32] Aron Monszpart, Nicolas Mellado, Gabriel J Brostow, and Niloy J Mitra. Rapter: rebuilding man-made scenes with regular arrangements of planes. ACM Trans. Graph., 34(4):103:1-103:12, 2015. 2 +[33] Claudio Mura, Oliver Mattausch, and Renato Pajarola. Piecewise-planar reconstruction of multi-room interiors with arbitrary wall arrangements. Comput. Graph. Forum, 35(7):179-188, 2016. 2 +[34] Liangliang Nan, Andrei Sharf, Hao Zhang, Daniel Cohen-Or, and Baoquan Chen. Smartboxes for interactive urban reconstruction. ACM Trans. Graph., 29(4):93:1–93:10, 2010. 1 +[35] Liangliang Nan and Peter Wonka. Polyfit: Polygonal surface reconstruction from point clouds. In IEEE Int. Conf. Comput. Vis., pages 2353-2361, 2017. 1, 2, 3, 5 +[36] Anh Nguyen and Bac Le. 3d point cloud segmentation: A survey. In IEEE Int. Conf. Robot. Autom. Mechatron, pages 225-230, 2013. 1 +[37] Sven Oesau, Florent Lafarge, and Pierre Alliez. Planar shape detection and regularization in tandem. Comput. Graph. Forum, 35(1):203-215, 2016. 2 +[38] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 165-174, 2019. 3 +[39] Tahir Rabbani, Frank Van Den Heuvel, and George Vosselmann. Segmentation of point clouds using smoothness constraint. ISPRS J. PhotoGramm., 36(5):248-253, 2006. 2, 3 +[40] Marie-Julie Rakotosaona, Paul Guerrero, Noam Aigerman, Niloy J. Mitra, and Maks Ovsjanikov. Learning delaunay surface elements for mesh reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22-31, June 2021. 3, 7, 8 +[41] Chiara Romanengo, Andrea Raffo, Yifan Qie, Nabil Anwer, and Bianca Falcidieno. Fit4cad: A point cloud benchmark for fitting simple geometric primitives in cad objects. Comput. Graph., 102:133-143, 2022. 5 +[42] Falko Schindler, Wolfgang Wörstner, and Jan-Michael Frahm. Classification and reconstruction of surfaces from point clouds of man-made objects. In IEEE Int. Conf. Comput. Vis. Worksh., pages 257-263, 2011. 2 +[43] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient ransac for point-cloud shape detection. Comput. Graph. Forum, 26(2):214-226, 2007. 2 +[44] Gerard Sierksma and Yori Zwols. Linear and integer optimization: theory and practice. 2015. 4 +[45] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Int. Conf. Neur. Info. Process. Sys., volume 33, pages 7462-7473, 2020. 3, 7, 8 + +[46] Marc Van Kreveld, Thijs Van Lankveld, and Remco C. Veltkamp. On the shape of a set of points and lines in the plane. Comput. Graph. Forum, 30(5):1553-1562, 2011. 2 +[47] Dong-Ming Yan, Wenping Wang, Yang Liu, and Zhouwang Yang. Variational mesh segmentation via quadric surface fitting. Computer-Aided Design, 44(11):1072-1082, 2012. 3 +[48] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. Capri-net: Learning compact cad shapes with adaptive primitive assembly. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11768-11778, 2022. 8 +[49] Mulin Yu and Florent Lafarge. Finding good configurations of planar primitives in unorganized point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6367-6376, 2022, 1, 2 +[50] Qingnan Zhou, Eitan Grinspun, Denis Zorin, and Alec Jacobson. Mesh arrangements for solid geometry. ACM Trans. Graph., 35(4):39:1-39:15, 2016. 4 +[51] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016. 5 +[52] Marco Zuliani, Charles S Kenney, and BS Manjunath. The multiransac algorithm and its application to detect planar homographies. In IEEE Int. Conf. Image Process., volume 3, pages III-153, 2005. 2 \ No newline at end of file diff --git a/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/images.zip b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..41fb20efc6ce12f4f72646cc4460528ecb439a28 --- /dev/null +++ b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1068c4a3d0ea2cba25a1b0a78bb223ed375cc0599686c57e3ad661909533c1fe +size 627492 diff --git a/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/layout.json b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fdf46d80ed528db6129f2c67dfeb1cca986d5cd5 --- /dev/null +++ b/2023/Structure-Aware Surface Reconstruction via Primitive Assembly/layout.json @@ -0,0 +1,12501 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "type": "text", + "content": "Structure-Aware Surface Reconstruction via Primitive Assembly" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 148, + 475, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 148, + 475, + 166 + ], + "spans": [ + { + "bbox": [ + 118, + 148, + 475, + 166 + ], + "type": "text", + "content": "Jingen Jiang" + }, + { + "bbox": [ + 118, + 148, + 475, + 166 + ], + "type": "inline_equation", + "content": "^{1,2,3,4^*}" + }, + { + "bbox": [ + 118, + 148, + 475, + 166 + ], + "type": "text", + "content": " Mingyang Zhao" + }, + { + "bbox": [ + 118, + 148, + 475, + 166 + ], + "type": "inline_equation", + "content": "^{1,5^*}" + }, + { + "bbox": [ + 118, + 148, + 475, + 166 + ], + "type": "text", + "content": " Shiqing Xin" + }, + { + "bbox": [ + 118, + 148, + 475, + 166 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 118, + 148, + 475, + 166 + ], + "type": "text", + "content": " Yanchao Yang" + }, + { + "bbox": [ + 118, + 148, + 475, + 166 + ], + "type": "inline_equation", + "content": "^{6}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 167, + 441, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 167, + 441, + 182 + ], + "spans": [ + { + "bbox": [ + 157, + 167, + 441, + 182 + ], + "type": "text", + "content": "Hanxiao Wang" + }, + { + "bbox": [ + 157, + 167, + 441, + 182 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 157, + 167, + 441, + 182 + ], + "type": "text", + "content": " Xiaohong Jia" + }, + { + "bbox": [ + 157, + 167, + 441, + 182 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 157, + 167, + 441, + 182 + ], + "type": "text", + "content": " Dong-Ming Yan" + }, + { + "bbox": [ + 157, + 167, + 441, + 182 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 182, + 182, + 416, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 182, + 416, + 195 + ], + "spans": [ + { + "bbox": [ + 182, + 182, + 416, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 182, + 182, + 416, + 195 + ], + "type": "text", + "content": "MAIS & NLPR, CASIA " + }, + { + "bbox": [ + 182, + 182, + 416, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 182, + 182, + 416, + 195 + ], + "type": "text", + "content": "KLMM, AMSS, CAS" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 196, + 478, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 196, + 478, + 211 + ], + "spans": [ + { + "bbox": [ + 114, + 196, + 478, + 211 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 114, + 196, + 478, + 211 + ], + "type": "text", + "content": "UCAS " + }, + { + "bbox": [ + 114, + 196, + 478, + 211 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 114, + 196, + 478, + 211 + ], + "type": "text", + "content": "Shandong University " + }, + { + "bbox": [ + 114, + 196, + 478, + 211 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 114, + 196, + 478, + 211 + ], + "type": "text", + "content": "BAAI " + }, + { + "bbox": [ + 114, + 196, + 478, + 211 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 114, + 196, + 478, + 211 + ], + "type": "text", + "content": "The University of Hong Kong" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 78, + 212, + 510, + 392 + ], + "blocks": [ + { + "bbox": [ + 78, + 212, + 510, + 392 + ], + "lines": [ + { + "bbox": [ + 78, + 212, + 510, + 392 + ], + "spans": [ + { + "bbox": [ + 78, + 212, + 510, + 392 + ], + "type": "image", + "image_path": "7aa06ecab35991e406759e0990b1bbfce592b2d4f6ebad5f072439fb9b300141.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 400, + 547, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 400, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 547, + 437 + ], + "type": "text", + "content": "Figure 1: Surface reconstruction for measurement point clouds. In contrast to piecewise approximations (KSR), our method not only delivers manifold and watertight surfaces but also retains the primitive structures (planes, cylinders, spheres, cones, and tori) faithfully. Also, our method enables the recovery of the sharp geometry of objects (e.g., feature lines) as a byproduct." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 447, + 192, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 447, + 192, + 459 + ], + "spans": [ + { + "bbox": [ + 143, + 447, + 192, + 459 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "content": "We propose a novel and efficient method for reconstructing manifold surfaces from point clouds. Unlike previous approaches that use dense implicit reconstructions or piecewise approximations and overlook inherent structures like quadrics in CAD models, our method faithfully preserves these quadric structures by assembling primitives. To achieve high-quality primitive extraction, we use a variational shape approximation, followed by a mesh arrangement for space partitioning and candidate primitive patches generation. We then introduce an effective pruning mechanism to classify candidate primitive patches as active or inactive, and further prune inactive patches to reduce the search space and speed up surface extraction significantly. Finally, the optimal active patches are computed by a binary linear programming and assembled as manifold and watertight surfaces. We perform extensive experiments on a wide range of CAD objects to validate its effectiveness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 446, + 386, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 446, + 386, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 386, + 459 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 470, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 547, + 567 + ], + "type": "text", + "content": "The problem of converting unorganized point clouds into surface meshes, also known as surface reconstruction, has various applications in 3D vision, computer graphics, and architecture geometry. In recent decades, although many powerful algorithms have been devoted to pursuing faithful and accurate reconstructions [19, 21, 34, 28, 27, 49], it remains a challenging task due to the inherent ill-posed nature [11, 36, 5]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": "The popular Poisson reconstruction methods [24, 25] implicitly generate dense smooth surfaces for point clouds, while others mainly focus on giving a compact approximation to objects by piecewise planar shapes or polygonal meshes [7, 13, 3, 35]. However, man-made objects in real-world scenarios are typically composed of non-planar quadratic surfaces or quadric structures, such as the Computer-Aided Design (CAD) models (see Fig. 1 as an example). Therefore, reconstructing manifold surface meshes from measurement data meanwhile retaining their intrinsic primitive structures in principle shall generate higher-quality approximation results than solely relying on poly-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 205, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 205, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 205, + 703 + ], + "type": "text", + "content": "*The authors contribute equally to this work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 58, + 703, + 228, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 228, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 228, + 713 + ], + "type": "text", + "content": "† Corresponding author (yandongming@gmail.com)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14171" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 74, + 71, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 74, + 71, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 74, + 71, + 83 + ], + "type": "text", + "content": "gons." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 84, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 84, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 84, + 286, + 227 + ], + "type": "text", + "content": "We argue that a method that solves the aforesaid problem ought to satisfy the following objectives: 1) Structure-awareness. The reconstruction surfaces should faithfully keep the structural features of the input point cloud, for instance, the common primitives in CAD, meanwhile correctly recovering the detailed geometry, and giving a high-fidelity approximation (e.g., the minimal Hausdorff distance) to the target object. 2) Manifold guarantees. The output surface meshes should be oriented 2-manifold and watertight [35]. 3) Efficiency. The developed algorithm should be scalable to large-scale point clouds with an acceptable computational complexity [3]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 228, + 286, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 286, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 286, + 562 + ], + "type": "text", + "content": "In this work, we develop a novel and efficient framework that takes the underlying geometric structures of point clouds into account when performing shape reconstruction. Our method ensures high-quality outputs and aims to achieve the above objectives. The core of our algorithm lies in a faithful reconstruction of intrinsic primitive structures presented in CAD models, as opposed to the direct surface approximation based on piecewise planar primitives. We consider five representative types of primitives commonly found in modern CAD systems, including four quadrics (planes, cylinders, cones, spheres) and tori. We adopt a variational shape approximation to optimize the extracted primitives followed by mesh arrangement techniques to attain primitive patches. A patch is a manifold that consists of a set of connected triangle meshes. As the size of patches is typically smaller than that of facets, our method enables a more stable and efficient reconstruction process than plane-based ones. Moreover, we propose a patch pruning mechanism to further speed up the consecutive surface extraction process. This mechanism distinguishes patches as either active or inactive, allowing us to perform fast local collision detection instead of a time-consuming global processing. Lastly, we select an optimal set of patches from the active candidate patches to assemble a reconstruction surface that respects the underlying structure of the object. To this end, we optimize a Binary Linear Programming (BLP) under hard constraints, as done in [35], to guarantee that the resulting surface is both manifold and watertight." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 563, + 286, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 286, + 669 + ], + "type": "text", + "content": "We perform extensive experiments to validate the effectiveness of our algorithm and compare it with state-of-the-art approaches using a variety of CAD models. Experimental results evidence that our method outperforms competitors in terms of both approximation precision and reconstruction efficiency, while also preserving the intrinsic primitive structures of the input point clouds with much better fidelity. To summarize, the main technical contributions include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 286, + 713 + ], + "type": "text", + "content": "- An efficient structure-aware reconstruction method that produces accurate surface models while preserving the underlying intrinsic primitive structures of the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 72, + 402, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 402, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 402, + 84 + ], + "type": "text", + "content": "input point clouds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 91, + 544, + 192 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 91, + 544, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 544, + 126 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 544, + 126 + ], + "type": "text", + "content": "- A patch pruning technique that avoids the time-consuming global collision detection and accelerates the surface extraction by a large margin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 133, + 544, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 133, + 544, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 133, + 544, + 192 + ], + "type": "text", + "content": "- A patch-induced binary linear programming for assembling manifold and watertight surfaces from candidate active patches, which significantly reduces the search space and helps achieve efficient and accurate surface reconstruction." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 223, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 223, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 545, + 258 + ], + "type": "text", + "content": "In this section, we briefly review the works that are most relevant to ours. Please refer to [5, 23] for a more comprehensive discussion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 258, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 546, + 510 + ], + "type": "text", + "content": "Geometric primitive extraction. Segmenting 3D point cloud models into a collection of geometric primitives such as planes and cylinders is helpful in most reconstruction approaches. Popular extraction frameworks include region growing [31, 39] and RANSAC [43, 52]. The former strategy takes the neighboring structure of models into consideration and propagates candidate primitives on their " + }, + { + "bbox": [ + 305, + 258, + 546, + 510 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 258, + 546, + 510 + ], + "type": "text", + "content": "-nearest neighbor graph until the fitting error exceeds a preset tolerance. Instead, RANSAC adopts a hypothesis and verification mechanism to recover primitives that embrace the most inliers. In order to make full use of the a priori structure of the input model, Li et al. [29] started with primitive extraction by RANSAC, followed by the refinement subject to spatial relationships such as parallelism and orthogonality. This idea was further extended to extract planar primitives simultaneously [32] or sequentially [37]. Recently, Yu et al. [49] proposed a new algorithm for extracting planar primitives, which has powerful potential for plane-assembly reconstruction. In our work, we leverage an efficient region growing strategy [39] for primitive extraction followed by a refinement via variational shape optimization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 546, + 713 + ], + "type": "text", + "content": "Assembly-based reconstruction. The core idea of assembly-based methods is to integrate all extracted primitives into a hybrid mesh. Connectivity analysis and slicing are two principal assembly manners. Connectivity strategies [8, 42, 46] consider neighboring relationships between primitives and encode them into an adjacency graph, which is adapted to regularize output meshes. Slicing strategies [6, 7, 3, 33, 35, 13] extend the extracted primitives and slice them to construct a spatial partition for candidate facets. Connectivity strategies are typically efficient but suffer from linkage errors of adjacent graphs generated by incomplete inputs. Arikan et al. [1] remedied the incomplete part by a user-guided interactive method that is labor-intensive especially for large-scale measurement data, whereas Lafarge et al. [26] provided an automatic solution that first consolidated the point clouds and then executed a min-cut optimization to extract dense triangle meshes." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14172" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 61, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 53, + 61, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 53, + 61, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 61, + 545, + 171 + ], + "type": "image", + "image_path": "dfc11adea0621a6c7e6ad414e238db5da038e88ec50b03180d7c8595eb623d4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 174, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 547, + 198 + ], + "type": "text", + "content": "Figure 2: Overview of the proposed method. (a) Point cloud after primitive extraction; (b) Accomplished space partitioning; (c) Active primitive patches after patch pruning; (d) The reconstructed surface." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 209, + 289, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 289, + 413 + ], + "type": "text", + "content": "Alternatively, slicing strategies deal with incomplete models by means of extending primitives. Nan et al. [35] performed an exhaustive partition to compensate for all possible missing and thus guaranteed a manifold output via BLP. Nevertheless, this manner typically leads to high computational complexity for the subsequent surface extraction. To circumvent this shortcoming, Fang et al. [13] proposed a hybrid method that combines the connectivity analysis with the slicing process to make a limited partition, while Bauchet et al. [3] adopted a kinetic data structure [16, 15] to monitor the growing of convex piecewise planar polygons until they collide, which also results in a limited partition. Unlike previous approaches, we propose a simper partition by considering five types of primitives simultaneously rather than merely planar shapes. Moreover, we develop a propagation strategy inspired by Bauchet et al. [3] to further simplify and speed up the surface extraction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 289, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 289, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 289, + 657 + ], + "type": "text", + "content": "Learning-based reconstruction. With the advancements in deep learning, there have been several notable works that utilize learning-based techniques to reconstruct meshes. Voxel-inspired methods, such as those presented in references [38] and [10], train a network of models to convert voxels into implicit surface representations. To attain more accurate representations, SIREN [45] and DIGS [4] used signed distance functions as implicit surface representations, and then extracted meshes via marching cubes [30]. Points2Surf [12] employed a patch-based learning framework to directly attain meshes from point clouds without the need for normals, while DSE [40] adopted 2D Delaunay triangulations to ensure manifold meshes and BSP-Net [9] output compact polygonal meshes with arbitrary topology. However, most of these methods depend on implicit surface representations, which may not explicitly capture the underlying structures such as quadric primitives of objects. In addition, it is hard to perform model editing tasks for surfaces generated by implicit reconstructions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 669, + 129, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 129, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 129, + 682 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "The proposed method consists of three major steps as illustrated in Fig. 2. Given an input point cloud " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{p_j\\in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 209, + 547, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 547, + 329 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3\\}_{j=1}^N" + }, + { + "bbox": [ + 304, + 209, + 547, + 329 + ], + "type": "text", + "content": " (Fig. 2(a)), we first partition " + }, + { + "bbox": [ + 304, + 209, + 547, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 209, + 547, + 329 + ], + "type": "text", + "content": " by a set of proxy meshes defined by the extracted geometric primitives and then apply a mesh arrangement operation to get candidate manifold patches (Fig. 2(b)) in Section 3.1. Later, a fast propagation mechanism is presented in Section 3.2 is used to prune patches that are inactive (Fig. 2(c)) to accelerate the subsequent surface extraction procedure. Finally, we deliver a manifold and watertight output surface (Fig. 2(d)) by formulating a patch-induced optimization under the BLP framework in Section 3.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 338, + 416, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 416, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 416, + 350 + ], + "type": "text", + "content": "3.1. Space Partitioning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 357, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 392 + ], + "type": "text", + "content": "The goal of this step is to generate a finite space partitioning by arranging proxy meshes of geometry primitives extracted from the input point cloud " + }, + { + "bbox": [ + 304, + 357, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 357, + 545, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": "Primitive extraction. We first utilize an efficient region growing strategy introduced in [39] to attain the rough segmentation components " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{s_i\\subset \\mathbb{R}^3\\}_{i = 1}^M" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": " and the corresponding primitives " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{q_i\\subset \\mathbb{R}^3\\}_{i = 1}^M" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": " is the supporting primitive of " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": " inversely, each point " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "p_j\\in s_i" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": " is the supporting point of " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = (\\mathcal{S},\\mathcal{Q})" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": " be the initial layout of the extracted primitives, we further refine " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": " via the variational shape approximation based on the " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\ell^2" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\ell^{2,1}" + }, + { + "bbox": [ + 304, + 408, + 545, + 529 + ], + "type": "text", + "content": " metrics [11, 47], which aims to optimize the following expression" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 537, + 545, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 537, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 353, + 537, + 545, + 572 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {L}} \\mathcal {F} (\\mathcal {L}) = \\sum_ {i = 1} ^ {M} \\sum_ {p _ {j} \\in s _ {i}} \\operatorname {d i s t} \\left(p _ {j}, q _ {i}\\right), \\tag {1}", + "image_path": "d1b024911d9fdf38a4707bdb853db0f94cfb6deda33c79b1f2fa946b09d49586.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "dist(p_j, q_i)" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " stands for the hybrid metric that measures the cost of the point " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "p_j \\in s_i" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " with respect to the corresponding primitive " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "q_i \\in \\mathcal{Q}" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " in terms of both the spatial " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\ell^2" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " distance and the normal " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\ell^{2,1}" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " deviation. Fig. 2(a) presents the primitive extraction result based on five types of primitives that we are currently considering. Noted that the proposed framework is not constrained by the above detected primitive types. Instead, it can be integrated with powerful detection algorithms that support a wider variety of primitive types, enabling more complex reconstructions with better accuracy." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14173" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 280, + 153 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 280, + 153 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 280, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 280, + 153 + ], + "type": "image", + "image_path": "6ac4b1d6dc82edafb4c43521b167161c1151b6a4d8ac1901070154f791e74c6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "text", + "content": "Figure 3: Visual description of patches and (non-) boundary curves. (a) " + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{ij}" + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "text", + "content": " is the non-boundary curve generated by patches " + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "inline_equation", + "content": "m_{j}" + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{C}}_i" + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{C}}_j" + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "text", + "content": " are the boundary curves of " + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 46, + 156, + 288, + 216 + ], + "type": "text", + "content": ", separately. (b) For each non-boundary curve, there are four patches around it." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "text", + "content": "Proxy mesh arrangement. After the optimization of the extracted primitives, we generate the discrete approximation to all primitives via proxy meshes. These meshes are trimmed by the bounding box of " + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "text", + "content": " and then all meshes are processed by a mesh arrangement [50]. The mesh arrangement results are a set of candidate primitive patches " + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{m_i\\subset \\mathbb{R}^3\\}_{i = 1}^K" + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "text", + "content": " as illustrated in Fig. 2(b), which partition the current working space into a series of cells that form a closed manifold without self-intersections. Here each patch " + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "text", + "content": " represents a subset of the input meshes and is the maximal connected set of triangle meshes as illustrated in Fig. 3. Due to the connectivity of patches, each " + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "text", + "content": " may embrace several non-boundary curves " + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{ij}" + }, + { + "bbox": [ + 46, + 219, + 289, + 399 + ], + "type": "text", + "content": " (colored red in Fig. 3(a)) that are shared with other primitive patches, which are defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 403, + 287, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 403, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 85, + 403, + 287, + 420 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {i j} = m _ {i} \\bigcap m _ {j} (i \\neq j), m _ {i}, m _ {j} \\in \\mathcal {M}, \\tag {2}", + "image_path": "ce7223bde0e98c646dc486f319cf37f0dddaae71c0d924db8b663f855da1f71c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 534 + ], + "type": "text", + "content": "whereas at most one boundary curve " + }, + { + "bbox": [ + 46, + 426, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\bar{C}_i" + }, + { + "bbox": [ + 46, + 426, + 289, + 534 + ], + "type": "text", + "content": " (colored green in Fig. 3(a)). For example, there are in total of four neighboring patches that share the same " + }, + { + "bbox": [ + 46, + 426, + 289, + 534 + ], + "type": "inline_equation", + "content": "C_{ij}" + }, + { + "bbox": [ + 46, + 426, + 289, + 534 + ], + "type": "text", + "content": " in Fig. 3(b). Since we perform space partitioning based on primitive patches rather than discretizing each of them as multiple piecewise planes, the developed algorithm indeed is several orders of magnitude faster than previous plane-induced approaches. The efficiency superiority of our method is also quantitatively validated by the following experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 539, + 257, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 257, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 257, + 552 + ], + "type": "text", + "content": "3.2. Pruning of Candidate Primitive Patches" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": "It is known that the computational complexity of the ultimate surface extraction in shape reconstruction regarding the size of patches is exponential [44]. To expedite the subsequent optimization process, we further propose a pruning mechanism which effectively circumvents the time-consuming and memory-intensive global collision detection to classify current primitive patches as active and inactive, from which we prune inactive ones to reduce the search space for surface extraction. As illustrated in Fig. 2, the significant time gap between the two routes(black and red) demonstrates the effectiveness of our pruning mechanism. The proposed procedure is composed of two steps, initialization and pruning, as presented subsequently." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 76, + 516, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 76, + 516, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 516, + 87 + ], + "type": "text", + "content": "Algorithm 1: Pseudo-code of the propagation process" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 89, + 526, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 89, + 526, + 100 + ], + "spans": [ + { + "bbox": [ + 318, + 89, + 526, + 100 + ], + "type": "text", + "content": "Input: Priority queue " + }, + { + "bbox": [ + 318, + 89, + 526, + 100 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 318, + 89, + 526, + 100 + ], + "type": "text", + "content": " of representative triangles." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 101, + 473, + 112 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 101, + 473, + 112 + ], + "spans": [ + { + "bbox": [ + 319, + 101, + 473, + 112 + ], + "type": "text", + "content": "Output: Active primitive patches " + }, + { + "bbox": [ + 319, + 101, + 473, + 112 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{M}}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 113, + 528, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 113, + 528, + 125 + ], + "spans": [ + { + "bbox": [ + 310, + 113, + 528, + 125 + ], + "type": "text", + "content": "1 Initialize propagation cost of each triangle as " + }, + { + "bbox": [ + 310, + 113, + 528, + 125 + ], + "type": "inline_equation", + "content": "c = 0" + }, + { + "bbox": [ + 310, + 113, + 528, + 125 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 125, + 389, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 125, + 389, + 136 + ], + "spans": [ + { + "bbox": [ + 310, + 125, + 389, + 136 + ], + "type": "text", + "content": "2 while " + }, + { + "bbox": [ + 310, + 125, + 389, + 136 + ], + "type": "inline_equation", + "content": "Q\\neq \\emptyset" + }, + { + "bbox": [ + 310, + 125, + 389, + 136 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 137, + 450, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 137, + 450, + 149 + ], + "spans": [ + { + "bbox": [ + 310, + 137, + 450, + 149 + ], + "type": "text", + "content": "3 " + }, + { + "bbox": [ + 310, + 137, + 450, + 149 + ], + "type": "inline_equation", + "content": "\\mid (t,c)\\gets Q.pop()" + }, + { + "bbox": [ + 310, + 137, + 450, + 149 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 310, + 137, + 450, + 149 + ], + "type": "inline_equation", + "content": "\\hat{t}\\leftarrow t" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 150, + 462, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 150, + 462, + 162 + ], + "spans": [ + { + "bbox": [ + 310, + 150, + 462, + 162 + ], + "type": "text", + "content": "4 for each triangle edge " + }, + { + "bbox": [ + 310, + 150, + 462, + 162 + ], + "type": "inline_equation", + "content": "e\\in \\hat{t}" + }, + { + "bbox": [ + 310, + 150, + 462, + 162 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 163, + 455, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 163, + 455, + 174 + ], + "spans": [ + { + "bbox": [ + 310, + 163, + 455, + 174 + ], + "type": "text", + "content": "5 for triangle " + }, + { + "bbox": [ + 310, + 163, + 455, + 174 + ], + "type": "inline_equation", + "content": "t_e\\cap t = e" + }, + { + "bbox": [ + 310, + 163, + 455, + 174 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 175, + 501, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 175, + 501, + 186 + ], + "spans": [ + { + "bbox": [ + 310, + 175, + 501, + 186 + ], + "type": "text", + "content": "6 | | if " + }, + { + "bbox": [ + 310, + 175, + 501, + 186 + ], + "type": "inline_equation", + "content": "\\text{prim}(t_e) == \\text{prim}(t) \\& t_e" + }, + { + "bbox": [ + 310, + 175, + 501, + 186 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 186, + 504, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 186, + 504, + 198 + ], + "spans": [ + { + "bbox": [ + 310, + 186, + 504, + 198 + ], + "type": "text", + "content": "inactive & " + }, + { + "bbox": [ + 310, + 186, + 504, + 198 + ], + "type": "inline_equation", + "content": "t_e" + }, + { + "bbox": [ + 310, + 186, + 504, + 198 + ], + "type": "text", + "content": " is not blocked then" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 199, + 517, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 199, + 517, + 210 + ], + "spans": [ + { + "bbox": [ + 310, + 199, + 517, + 210 + ], + "type": "text", + "content": "7 Compute the height " + }, + { + "bbox": [ + 310, + 199, + 517, + 210 + ], + "type": "inline_equation", + "content": "h_e" + }, + { + "bbox": [ + 310, + 199, + 517, + 210 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 310, + 199, + 517, + 210 + ], + "type": "inline_equation", + "content": "t_e" + }, + { + "bbox": [ + 310, + 199, + 517, + 210 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 310, + 199, + 517, + 210 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 310, + 199, + 517, + 210 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 211, + 471, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 211, + 471, + 222 + ], + "spans": [ + { + "bbox": [ + 310, + 211, + 471, + 222 + ], + "type": "text", + "content": "8 " + }, + { + "bbox": [ + 310, + 211, + 471, + 222 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 310, + 211, + 471, + 222 + ], + "type": "text", + "content": " .push((te,c+he))." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 262, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 312 + ], + "type": "text", + "content": "Initialization. To describe the point distribution of the input point cloud as fidelity as possible, we first select a set of representative triangle meshes denoted as " + }, + { + "bbox": [ + 304, + 262, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 262, + 547, + 312 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 262, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 262, + 547, + 312 + ], + "type": "text", + "content": " by means of the triangle coverage " + }, + { + "bbox": [ + 304, + 262, + 547, + 312 + ], + "type": "inline_equation", + "content": "Cover(t)" + }, + { + "bbox": [ + 304, + 262, + 547, + 312 + ], + "type": "text", + "content": " which is defined as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 376, + 319, + 545, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 319, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 376, + 319, + 545, + 346 + ], + "type": "interline_equation", + "content": "\\operatorname {C o v e r} (t) = \\frac {\\operatorname {V a l i d} (t)}{\\operatorname {A r e a} (t)}, \\tag {3}", + "image_path": "57d53df0cac6d46c332fcac83696a896e33cfc07ec4937d426ff1ed4b987931b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 348, + 547, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 547, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 348, + 547, + 372 + ], + "type": "inline_equation", + "content": "\\text{Area}(t)" + }, + { + "bbox": [ + 305, + 348, + 547, + 372 + ], + "type": "text", + "content": " stands for the area of a triangle mesh " + }, + { + "bbox": [ + 305, + 348, + 547, + 372 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 348, + 547, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 348, + 547, + 372 + ], + "type": "inline_equation", + "content": "\\text{Valid}(t)" + }, + { + "bbox": [ + 305, + 348, + 547, + 372 + ], + "type": "text", + "content": " is formulated as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 348, + 379, + 545, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 379, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 348, + 379, + 545, + 406 + ], + "type": "interline_equation", + "content": "V a l i d (t) = \\sum_ {p \\in \\mathcal {P} _ {t}} 1 (\\| p - p ^ {\\prime} \\| _ {2} < \\epsilon) a _ {p}, \\tag {4}", + "image_path": "daa3bc4c28f32da2199646e9afc9047fced6d1a230896eda68565623e5dac1b1.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t\\subset \\mathcal{P}" + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "text", + "content": " encompasses the measurement points that their projections are attached to " + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\| p - p^{\\prime}\\|_{2}" + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\ell^2" + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "text", + "content": " Euclidean distance from " + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "text", + "content": " to its projective point " + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "inline_equation", + "content": "p^\\prime" + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 412, + 545, + 448 + ], + "type": "inline_equation", + "content": "t" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "text", + "content": "1 is the indicator function, and " + }, + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "inline_equation", + "content": "a_{p}" + }, + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "text", + "content": " is the corresponding geodesic Voronoi area of " + }, + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "text", + "content": ", which is calculated as in the inset: " + }, + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "text", + "content": " and its " + }, + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "text", + "content": "-nearest neighbors (i.e., hard red points) are first projected to the supporting plane " + }, + { + "bbox": [ + 305, + 449, + 425, + 544 + ], + "type": "inline_equation", + "content": "\\Omega_t" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 436, + 449, + 528, + 538 + ], + "blocks": [ + { + "bbox": [ + 436, + 449, + 528, + 538 + ], + "lines": [ + { + "bbox": [ + 436, + 449, + 528, + 538 + ], + "spans": [ + { + "bbox": [ + 436, + 449, + 528, + 538 + ], + "type": "image", + "image_path": "15deff45d25e82bb7cc5e77f6ad8d2ee4b50c41adac1751201a339de7ca3b510.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": "where its attached triangle mesh " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " lies on, then " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "a_{p}" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " is equivalent to the area of the 2D Voronoi cell [2]. " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " is termed representative if its Cover " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "(t)" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " surpasses a user-specified value. We populate all representative meshes of " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " into a priority queue " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " with an ascending order." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Pruning. Next, representative triangle meshes are propagated to enable a limited partition to speedup the final surface extraction. To this end, all triangles of " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " are initially labeled as inactive prior to entering the loop of the priority queue. For each round during the loop, we first pop the top element, for instance, " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "(t_i, c_i)" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": ", out of the priority queue and prompt " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " to be an active state " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{t}_i" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": ". The propagation cost " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " of a triangle mesh " + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": " is defined as the height" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14174" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": " from its vertex to the base edge shared with another adjacent triangle. Sequentially, we propagate all the triangles that share both the same edge and the same supporting primitive with " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": " and pop the update elements out of " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": ". For instance, for edges that are not located on non-boundary curves, there are at most three update elements " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "(t_{i1},c_i + h_1)" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "(t_{i2},c_i + h_2)" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "(t_{i3},c_i + h_3)" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": ", whereas, for edges that coincide non-boundary curves, we perform a fast local collision detection to check whether there exist active meshes, if yes then we evoke barricades and stop the propagation process. The above propagation proceeds until " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": " becomes empty. We summarize the proposed propagation process in Algorithm 1 to make it more understandable. Consequently, a primitive patch is active only if its triangles are all active, otherwise, it is still inactive. We use " + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 299 + ], + "type": "text", + "content": " to represent the set of active patches. The patches in Fig. 2(b) after pruning are shown in Fig. 2(c). As observed, the size of candidate patches has been significantly reduced compared with Fig. 2(b)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "text", + "content": "Like [3], we also allow for the propagation crossing primitive patches " + }, + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "text", + "content": " times (with " + }, + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = 2" + }, + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "text", + "content": " in default) at most. This can deal with small parts that may have been missed during the primitive detection. However, in contrast to previous global collision detection manners [3], our local strategy preserves much better the propagation efficiency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 380, + 160, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 160, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 160, + 392 + ], + "type": "text", + "content": "3.3. Surface Extraction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "text", + "content": "Finally, we adopt a Binary Linear Programming (BLP) analogous to [35] to extract the surface model from " + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "text", + "content": ". Specifically, the state of patches in " + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{M}}" + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "text", + "content": " is encoded as " + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{\\tilde{M}} = \\{x_{\\tilde{m}}|\\tilde{m}\\in \\tilde{\\mathcal{M}},x_{\\tilde{m}}\\in \\{0,1\\} \\}" + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "text", + "content": ", where the binary discrete variable " + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "inline_equation", + "content": "x_{\\tilde{m}}" + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "text", + "content": " indicates whether the patch " + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "inline_equation", + "content": "\\tilde{m}" + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "text", + "content": " is selected " + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "inline_equation", + "content": "(x_{\\tilde{m}} = 1)" + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "text", + "content": " or not " + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "inline_equation", + "content": "(x_{\\tilde{m}} = 0)" + }, + { + "bbox": [ + 46, + 398, + 289, + 533 + ], + "type": "text", + "content": ". To ensure the resulting surface manifold and watertight, we constrain the curves of the selected patches to be all non-boundary, and thus the two patches sharing the same non-boundary curve are either selected at the same time or both not selected. Consequently, our objective function is defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 540, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 540, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 52, + 540, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\mathcal {X} _ {\\tilde {\\mathcal {M}}}} \\mathbf {U} (\\mathcal {X} _ {\\tilde {\\mathcal {M}}}) + \\lambda \\mathbf {B} (\\mathcal {X} _ {\\tilde {\\mathcal {M}}}), \\\\ s. t. \\left\\{ \\begin{array}{l} \\sum_ {\\tilde {m} _ {i} \\cap \\tilde {m} _ {j} \\neq \\emptyset , i \\neq j} \\left(x _ {\\tilde {m} _ {i}} + x _ {\\tilde {m} _ {j}}\\right) = 0 o r 2, \\tilde {m} _ {i}, \\tilde {m} _ {j} \\in \\tilde {\\mathcal {M}} \\\\ x _ {\\tilde {m} _ {i}} \\in \\{0, 1 \\}, \\quad i = 1, 2, 3 \\dots | \\tilde {\\mathcal {M}} |. \\end{array} \\right. \\tag {5} \\\\ \\end{array}", + "image_path": "e44dfd5749a8fb547b4fe50008ef0f8aa15b39708e94a4b7a21e5fe2fb7da364.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{U}(\\mathcal{X}_{\\tilde{\\mathcal{M}}})" + }, + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{B}(\\mathcal{X}_{\\tilde{\\mathcal{M}}})" + }, + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "type": "text", + "content": " are penalty terms regarding the patches and the neighboring patch pairs, respectively, and " + }, + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "type": "text", + "content": " is a balance coefficient. Moreover " + }, + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{U}(\\mathcal{X}_{\\tilde{\\mathcal{M}}})" + }, + { + "bbox": [ + 46, + 617, + 288, + 676 + ], + "type": "text", + "content": " encourages the selection of active patches that are with high fidelity in the form of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 683, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 683, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 70, + 683, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\mathbf {U} \\left(\\mathcal {X} _ {\\tilde {M}}\\right) = \\sum_ {\\tilde {m} \\in \\tilde {\\mathcal {M}}} x _ {\\tilde {m}} \\cdot \\left(\\frac {\\mathcal {A} _ {\\tilde {m}} - \\bar {\\mathcal {A}} _ {\\tilde {m}}}{\\mathcal {A}} - \\frac {| \\mathcal {P} _ {\\tilde {m}} |}{| \\mathcal {P} |}\\right). \\tag {6}", + "image_path": "62d941d6b69e8644ef38a6a3a9b57c64b4a3eea349bbb83273935e7cf9e2db8f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\tilde{m}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{A}}_{\\tilde{m}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " denote the area and the coverage area of the patch " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "\\tilde{m}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": ", respectively. The coverage area of " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "\\tilde{m}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " is the summation of the coverage area of all its triangles. " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " is a normalization factor equivalent to the sum of the area of all patches. " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\tilde{m}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " is the set of points whose projections are locating in " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "\\tilde{m}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": ". On the other hand, " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\tilde{\\mathcal{X}})" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " is designed to control the model complexity. Specifically, it put penalties to the non-boundary curve length of two patches:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 353, + 175, + 545, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 175, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 353, + 175, + 545, + 206 + ], + "type": "interline_equation", + "content": "\\mathcal {B} \\left(\\mathcal {X} _ {\\tilde {\\mathcal {M}}}\\right) = \\frac {1}{L} \\sum_ {i, j} \\left| \\mathcal {C} _ {i j} \\right| \\cdot x _ {\\tilde {m} _ {i}} \\cdot x _ {\\tilde {m} _ {j}}, \\tag {7}", + "image_path": "81fdee11981f4c802b92ac28d460ccd235328510f87b7152cf83a6d71d7945a4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}_{ij}|" + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "text", + "content": " denotes the length of the curve shared by two active patches " + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "inline_equation", + "content": "\\tilde{m}_i" + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "inline_equation", + "content": "\\tilde{m}_j" + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "text", + "content": " with different supporting primitives and " + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "text", + "content": " is the sum of the length of all curves. Fig. 2(d) presents the final reconstruction surface by BLP, which not only generates a manifold and watertight shape but also successfully preserves the intrinsic primitive structures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 296, + 538, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 296, + 538, + 309 + ], + "spans": [ + { + "bbox": [ + 305, + 296, + 538, + 309 + ], + "type": "text", + "content": "4. Experimental Evaluations and Discussions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 316, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 519 + ], + "type": "text", + "content": "Implementation details and metrics. In this section, we execute extensive experiments to validate the advantages of the developed method in terms of both reconstruction precision and efficiency and make various comparisons with state-of-the-art approaches. Like [3], the precision is quantitatively evaluated by the symmetric mean Hausdorff distance (SMH) between the measurement point clouds and the reconstruction meshes (values are normalized by the diagonal length of the models' bounding box and multiplied by a factor of " + }, + { + "bbox": [ + 304, + 316, + 547, + 519 + ], + "type": "inline_equation", + "content": "10^{2}" + }, + { + "bbox": [ + 304, + 316, + 547, + 519 + ], + "type": "text", + "content": "), while the efficiency is assessed by the running time. Our algorithm is implemented in " + }, + { + "bbox": [ + 304, + 316, + 547, + 519 + ], + "type": "inline_equation", + "content": "\\mathrm{C}++" + }, + { + "bbox": [ + 304, + 316, + 547, + 519 + ], + "type": "text", + "content": " using the geometry processing library libigl [22]. We adopt the mathematical programming solver Gurobi [18] to optimize Eq. 5. All reconstruction results are rendered by Blender [14] for better visualization. Our experiments are conducted on a desktop with Intel(R) Core i7-7700K CPU@4.2GHz and 48GB RAM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "Reconstruction effect. Our algorithm has demonstrated good reconstruction results from point clouds to manifold meshes on various CAD objects as illustrated in Fig. 4. The input point clouds presented in Fig. 4(a)-(h) are either sampled from real-world CAD models or courtesy by publicly available benchmark datasets [41, 51]. As observed, there are multiple explicit intrinsic primitive structures in each model. Instead of approximating them directly by piecewise planar primitives, our algorithm is capable of reconstructing the intrinsic primitives faithfully even for quite challenging or complicated models. For instance, the rich spherical and conic structures existing in the Minarets model of Fig. 4(b) require careful treatments, yet our method is still able to achieve highly accurate reconstructions while preserving all structural details successfully. At the same time," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "14175" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 67, + 141, + 127 + ], + "blocks": [ + { + "bbox": [ + 52, + 89, + 67, + 102 + ], + "lines": [ + { + "bbox": [ + 52, + 89, + 67, + 102 + ], + "spans": [ + { + "bbox": [ + 52, + 89, + 67, + 102 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 67, + 141, + 127 + ], + "lines": [ + { + "bbox": [ + 88, + 67, + 141, + 127 + ], + "spans": [ + { + "bbox": [ + 88, + 67, + 141, + 127 + ], + "type": "image", + "image_path": "f85af81be8a05e864169d8870b1cdc1767ece2c194aa1ae05a6250ad1a66fc70.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 186, + 66, + 240, + 126 + ], + "blocks": [ + { + "bbox": [ + 186, + 66, + 240, + 126 + ], + "lines": [ + { + "bbox": [ + 186, + 66, + 240, + 126 + ], + "spans": [ + { + "bbox": [ + 186, + 66, + 240, + 126 + ], + "type": "image", + "image_path": "1f373b54946d17a8c3aa6de1cd72d62257d78d03eb262a5013d9d7febdc1c7f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 278, + 67, + 332, + 126 + ], + "blocks": [ + { + "bbox": [ + 278, + 67, + 332, + 126 + ], + "lines": [ + { + "bbox": [ + 278, + 67, + 332, + 126 + ], + "spans": [ + { + "bbox": [ + 278, + 67, + 332, + 126 + ], + "type": "image", + "image_path": "08daddc0b9859a06af6152337881d56f1d858844efd2f5493644e5cc840026e8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 376, + 67, + 429, + 126 + ], + "blocks": [ + { + "bbox": [ + 376, + 67, + 429, + 126 + ], + "lines": [ + { + "bbox": [ + 376, + 67, + 429, + 126 + ], + "spans": [ + { + "bbox": [ + 376, + 67, + 429, + 126 + ], + "type": "image", + "image_path": "fa0d88f40c136778575bf627483cf752b758c205d64cd5f76eb2ae23e575ba9f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 84, + 129, + 141, + 180 + ], + "blocks": [ + { + "bbox": [ + 53, + 147, + 67, + 159 + ], + "lines": [ + { + "bbox": [ + 53, + 147, + 67, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 67, + 159 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 84, + 129, + 141, + 180 + ], + "lines": [ + { + "bbox": [ + 84, + 129, + 141, + 180 + ], + "spans": [ + { + "bbox": [ + 84, + 129, + 141, + 180 + ], + "type": "image", + "image_path": "88cbd7dd0e889fd8b501f761eb9ba5c5ba36a5c012f3a3c036d6cca41db981e1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 183, + 132, + 241, + 182 + ], + "blocks": [ + { + "bbox": [ + 183, + 132, + 241, + 182 + ], + "lines": [ + { + "bbox": [ + 183, + 132, + 241, + 182 + ], + "spans": [ + { + "bbox": [ + 183, + 132, + 241, + 182 + ], + "type": "image", + "image_path": "f8019e35731c67bcde3cbb1b0d030957d555c91a1550d07038a036b6caca6531.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 276, + 132, + 335, + 182 + ], + "blocks": [ + { + "bbox": [ + 276, + 132, + 335, + 182 + ], + "lines": [ + { + "bbox": [ + 276, + 132, + 335, + 182 + ], + "spans": [ + { + "bbox": [ + 276, + 132, + 335, + 182 + ], + "type": "image", + "image_path": "aa73e1fb03136d1e332835c4433b54b3d69fdfdccc95592b71872df16d17d139.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 374, + 136, + 432, + 187 + ], + "blocks": [ + { + "bbox": [ + 374, + 136, + 432, + 187 + ], + "lines": [ + { + "bbox": [ + 374, + 136, + 432, + 187 + ], + "spans": [ + { + "bbox": [ + 374, + 136, + 432, + 187 + ], + "type": "image", + "image_path": "2967a618cca8641a9a671c62701495475ea478c6cf13309a5ac2fe73926b236d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 477, + 68, + 530, + 126 + ], + "blocks": [ + { + "bbox": [ + 477, + 68, + 530, + 126 + ], + "lines": [ + { + "bbox": [ + 477, + 68, + 530, + 126 + ], + "spans": [ + { + "bbox": [ + 477, + 68, + 530, + 126 + ], + "type": "image", + "image_path": "7e9e70d971760a48c435f4320ec6f96082a8e59725a74e52b7543f7c28c329b4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 80, + 194, + 144, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 67, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 67, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 67, + 217 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 194, + 144, + 235 + ], + "lines": [ + { + "bbox": [ + 80, + 194, + 144, + 235 + ], + "spans": [ + { + "bbox": [ + 80, + 194, + 144, + 235 + ], + "type": "image", + "image_path": "58932b0e464b7fcf3484769d1081aff2693565d0118a5f2875dbd0b7ea8dc4a9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 178, + 194, + 243, + 235 + ], + "blocks": [ + { + "bbox": [ + 178, + 194, + 243, + 235 + ], + "lines": [ + { + "bbox": [ + 178, + 194, + 243, + 235 + ], + "spans": [ + { + "bbox": [ + 178, + 194, + 243, + 235 + ], + "type": "image", + "image_path": "182f55ad874a7419534daea35c046eae52a9f86c14ef3068f5c954011027bf58.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 272, + 194, + 337, + 236 + ], + "blocks": [ + { + "bbox": [ + 272, + 194, + 337, + 236 + ], + "lines": [ + { + "bbox": [ + 272, + 194, + 337, + 236 + ], + "spans": [ + { + "bbox": [ + 272, + 194, + 337, + 236 + ], + "type": "image", + "image_path": "9bfba2f70dcbda172919956545615a66128738877a130fc1a740842026df6908.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 373, + 196, + 436, + 237 + ], + "blocks": [ + { + "bbox": [ + 373, + 196, + 436, + 237 + ], + "lines": [ + { + "bbox": [ + 373, + 196, + 436, + 237 + ], + "spans": [ + { + "bbox": [ + 373, + 196, + 436, + 237 + ], + "type": "image", + "image_path": "ae38cdbf35efdaf43afbb707680612d62c99a5f93b040f2912a10443e94b82d3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 473, + 136, + 531, + 186 + ], + "blocks": [ + { + "bbox": [ + 473, + 136, + 531, + 186 + ], + "lines": [ + { + "bbox": [ + 473, + 136, + 531, + 186 + ], + "spans": [ + { + "bbox": [ + 473, + 136, + 531, + 186 + ], + "type": "image", + "image_path": "e3c329c267ec47063e992bd864771e3465e1aa6a8df30eedda3086e47ed88930.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 83, + 246, + 145, + 296 + ], + "blocks": [ + { + "bbox": [ + 54, + 268, + 67, + 279 + ], + "lines": [ + { + "bbox": [ + 54, + 268, + 67, + 279 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 67, + 279 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 246, + 145, + 296 + ], + "lines": [ + { + "bbox": [ + 83, + 246, + 145, + 296 + ], + "spans": [ + { + "bbox": [ + 83, + 246, + 145, + 296 + ], + "type": "image", + "image_path": "d41c99358da7634afc561c82f3d35bad3672589582364e469b29bed4ee5100f0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 181, + 246, + 242, + 296 + ], + "blocks": [ + { + "bbox": [ + 181, + 246, + 242, + 296 + ], + "lines": [ + { + "bbox": [ + 181, + 246, + 242, + 296 + ], + "spans": [ + { + "bbox": [ + 181, + 246, + 242, + 296 + ], + "type": "image", + "image_path": "55e5bf41ae7a5bf3fcb59e7da57f8f2565ccd0c7a12d8442fb45adc51c4cb425.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 278, + 246, + 340, + 297 + ], + "blocks": [ + { + "bbox": [ + 278, + 246, + 340, + 297 + ], + "lines": [ + { + "bbox": [ + 278, + 246, + 340, + 297 + ], + "spans": [ + { + "bbox": [ + 278, + 246, + 340, + 297 + ], + "type": "image", + "image_path": "4b877af64b5cf1e89128997cc67c3d9b04f45b6e0c5c6721b107c41e3f1a7fa2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 377, + 245, + 438, + 296 + ], + "blocks": [ + { + "bbox": [ + 377, + 245, + 438, + 296 + ], + "lines": [ + { + "bbox": [ + 377, + 245, + 438, + 296 + ], + "spans": [ + { + "bbox": [ + 377, + 245, + 438, + 296 + ], + "type": "image", + "image_path": "bbba14f044722479a8c9f85e463b731a4e77664879af5d56a192a188485b5353.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 468, + 197, + 532, + 237 + ], + "blocks": [ + { + "bbox": [ + 468, + 197, + 532, + 237 + ], + "lines": [ + { + "bbox": [ + 468, + 197, + 532, + 237 + ], + "spans": [ + { + "bbox": [ + 468, + 197, + 532, + 237 + ], + "type": "image", + "image_path": "6d24df649ec78854124ab95822f905269b0edeb2fd56d1ee27473da28aba25b7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 472, + 245, + 533, + 295 + ], + "blocks": [ + { + "bbox": [ + 472, + 245, + 533, + 295 + ], + "lines": [ + { + "bbox": [ + 472, + 245, + 533, + 295 + ], + "spans": [ + { + "bbox": [ + 472, + 245, + 533, + 295 + ], + "type": "image", + "image_path": "789af587d81aa0e8acd3875b3a72498c8546ee6609f09e567fd03df8c67ec8cd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 85, + 311, + 148, + 369 + ], + "blocks": [ + { + "bbox": [ + 54, + 329, + 67, + 341 + ], + "lines": [ + { + "bbox": [ + 54, + 329, + 67, + 341 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 67, + 341 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 311, + 148, + 369 + ], + "lines": [ + { + "bbox": [ + 85, + 311, + 148, + 369 + ], + "spans": [ + { + "bbox": [ + 85, + 311, + 148, + 369 + ], + "type": "image", + "image_path": "eb71a968918d026e27e4d439b334cf9f711c9b7f82a4ea4b7e967c931a2930df.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 178, + 316, + 242, + 373 + ], + "blocks": [ + { + "bbox": [ + 178, + 316, + 242, + 373 + ], + "lines": [ + { + "bbox": [ + 178, + 316, + 242, + 373 + ], + "spans": [ + { + "bbox": [ + 178, + 316, + 242, + 373 + ], + "type": "image", + "image_path": "2670d38ec9e351c5fa434d4ed125ba73169ddfb1e708e7d34eb6d46300584b92.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 277, + 315, + 343, + 373 + ], + "blocks": [ + { + "bbox": [ + 277, + 315, + 343, + 373 + ], + "lines": [ + { + "bbox": [ + 277, + 315, + 343, + 373 + ], + "spans": [ + { + "bbox": [ + 277, + 315, + 343, + 373 + ], + "type": "image", + "image_path": "8ac5bba0b0cde9dbd34fd7404f1990a973222237f5e4f11c73d806e83739a1f3.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 380, + 316, + 445, + 375 + ], + "blocks": [ + { + "bbox": [ + 380, + 316, + 445, + 375 + ], + "lines": [ + { + "bbox": [ + 380, + 316, + 445, + 375 + ], + "spans": [ + { + "bbox": [ + 380, + 316, + 445, + 375 + ], + "type": "image", + "image_path": "5eac263144508524a4b1ccad2874299b45c7bd5bd3370ba8c34c8b1b0521b97d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 469, + 316, + 534, + 375 + ], + "blocks": [ + { + "bbox": [ + 469, + 316, + 534, + 375 + ], + "lines": [ + { + "bbox": [ + 469, + 316, + 534, + 375 + ], + "spans": [ + { + "bbox": [ + 469, + 316, + 534, + 375 + ], + "type": "image", + "image_path": "95fe49212056e029bd1108c91c4a2072d1e36bf6d2205edc1702f33ea129d969.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 88, + 380, + 143, + 441 + ], + "blocks": [ + { + "bbox": [ + 55, + 398, + 67, + 410 + ], + "lines": [ + { + "bbox": [ + 55, + 398, + 67, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 67, + 410 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 380, + 143, + 441 + ], + "lines": [ + { + "bbox": [ + 88, + 380, + 143, + 441 + ], + "spans": [ + { + "bbox": [ + 88, + 380, + 143, + 441 + ], + "type": "image", + "image_path": "a03926641a3d0832715b029625d1d4592fdd2c8e55e7126c81938df299e8fef3.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 181, + 381, + 236, + 441 + ], + "blocks": [ + { + "bbox": [ + 181, + 381, + 236, + 441 + ], + "lines": [ + { + "bbox": [ + 181, + 381, + 236, + 441 + ], + "spans": [ + { + "bbox": [ + 181, + 381, + 236, + 441 + ], + "type": "image", + "image_path": "a61610d427f31e424114c0c76f3bb01bf67e4529026fb108b01b901db0e7ab4b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 279, + 383, + 335, + 443 + ], + "blocks": [ + { + "bbox": [ + 279, + 383, + 335, + 443 + ], + "lines": [ + { + "bbox": [ + 279, + 383, + 335, + 443 + ], + "spans": [ + { + "bbox": [ + 279, + 383, + 335, + 443 + ], + "type": "image", + "image_path": "f7d4e3ff452e045f7f6588caa4dc6a6d03466b11111c6bdb7a309839d624a696.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 384, + 384, + 441, + 445 + ], + "blocks": [ + { + "bbox": [ + 384, + 384, + 441, + 445 + ], + "lines": [ + { + "bbox": [ + 384, + 384, + 441, + 445 + ], + "spans": [ + { + "bbox": [ + 384, + 384, + 441, + 445 + ], + "type": "image", + "image_path": "6fbb3642d226ad5c1b315b86ca1958524619e8c2c8e03efc01e37deee24bd3ff.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 475, + 384, + 530, + 444 + ], + "blocks": [ + { + "bbox": [ + 475, + 384, + 530, + 444 + ], + "lines": [ + { + "bbox": [ + 475, + 384, + 530, + 444 + ], + "spans": [ + { + "bbox": [ + 475, + 384, + 530, + 444 + ], + "type": "image", + "image_path": "b231689cc0488831320d8e5555754c1f19b89ad49c20cc851eeab05f8428c507.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 87, + 455, + 141, + 529 + ], + "blocks": [ + { + "bbox": [ + 54, + 478, + 67, + 491 + ], + "lines": [ + { + "bbox": [ + 54, + 478, + 67, + 491 + ], + "spans": [ + { + "bbox": [ + 54, + 478, + 67, + 491 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 455, + 141, + 529 + ], + "lines": [ + { + "bbox": [ + 87, + 455, + 141, + 529 + ], + "spans": [ + { + "bbox": [ + 87, + 455, + 141, + 529 + ], + "type": "image", + "image_path": "fa9df84afdc972e8e5a3f6861e6bdf37f5df4c62fc5d57a45893efe226b96885.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 179, + 455, + 235, + 528 + ], + "blocks": [ + { + "bbox": [ + 179, + 455, + 235, + 528 + ], + "lines": [ + { + "bbox": [ + 179, + 455, + 235, + 528 + ], + "spans": [ + { + "bbox": [ + 179, + 455, + 235, + 528 + ], + "type": "image", + "image_path": "93779c0fca2d7581e66860f73101053460b52a9f5fa58c1d5cc83867622bdb0a.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 278, + 454, + 333, + 529 + ], + "blocks": [ + { + "bbox": [ + 278, + 454, + 333, + 529 + ], + "lines": [ + { + "bbox": [ + 278, + 454, + 333, + 529 + ], + "spans": [ + { + "bbox": [ + 278, + 454, + 333, + 529 + ], + "type": "image", + "image_path": "2f0b6540f330e6f9043af10d05208d0436df583f5e6472784079f4fe218b9435.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 386, + 454, + 441, + 529 + ], + "blocks": [ + { + "bbox": [ + 386, + 454, + 441, + 529 + ], + "lines": [ + { + "bbox": [ + 386, + 454, + 441, + 529 + ], + "spans": [ + { + "bbox": [ + 386, + 454, + 441, + 529 + ], + "type": "image", + "image_path": "b1a2123c3fd365fe3e9bd4dc2a81c4f995abf803da08c3808feb79365aedca25.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 477, + 454, + 533, + 529 + ], + "blocks": [ + { + "bbox": [ + 477, + 454, + 533, + 529 + ], + "lines": [ + { + "bbox": [ + 477, + 454, + 533, + 529 + ], + "spans": [ + { + "bbox": [ + 477, + 454, + 533, + 529 + ], + "type": "image", + "image_path": "3ebc44a8c908f54c0a7a4f1f55da6f582985410a9d1484708827731343727235.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 87, + 551, + 154, + 580 + ], + "blocks": [ + { + "bbox": [ + 54, + 555, + 67, + 567 + ], + "lines": [ + { + "bbox": [ + 54, + 555, + 67, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 555, + 67, + 567 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 551, + 154, + 580 + ], + "lines": [ + { + "bbox": [ + 87, + 551, + 154, + 580 + ], + "spans": [ + { + "bbox": [ + 87, + 551, + 154, + 580 + ], + "type": "image", + "image_path": "edcebd04c2c6e64671b61ed83c46a8ff1a92207e917aebc255e689f9991b7ff9.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 178, + 548, + 245, + 578 + ], + "blocks": [ + { + "bbox": [ + 178, + 548, + 245, + 578 + ], + "lines": [ + { + "bbox": [ + 178, + 548, + 245, + 578 + ], + "spans": [ + { + "bbox": [ + 178, + 548, + 245, + 578 + ], + "type": "image", + "image_path": "6e9640d3f5591aef29042483bcaefdd2752c5ed22091eb3affcc66b03deb37ee.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 276, + 548, + 343, + 578 + ], + "blocks": [ + { + "bbox": [ + 276, + 548, + 343, + 578 + ], + "lines": [ + { + "bbox": [ + 276, + 548, + 343, + 578 + ], + "spans": [ + { + "bbox": [ + 276, + 548, + 343, + 578 + ], + "type": "image", + "image_path": "82627de130599f1c3fdb5149160845fb9c732551b22bf41aaa525932716c96e1.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 380, + 545, + 446, + 575 + ], + "blocks": [ + { + "bbox": [ + 380, + 545, + 446, + 575 + ], + "lines": [ + { + "bbox": [ + 380, + 545, + 446, + 575 + ], + "spans": [ + { + "bbox": [ + 380, + 545, + 446, + 575 + ], + "type": "image", + "image_path": "a2b238be5790533745a1c14ee1f21b69d2a1a17226afba8eee161e699824a390.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 475, + 542, + 541, + 571 + ], + "blocks": [ + { + "bbox": [ + 475, + 542, + 541, + 571 + ], + "lines": [ + { + "bbox": [ + 475, + 542, + 541, + 571 + ], + "spans": [ + { + "bbox": [ + 475, + 542, + 541, + 571 + ], + "type": "image", + "image_path": "c5a75d12d6f8e95d46384ec27de0fe94019fe0b0897b5a88124fa29c459d83c8.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 479, + 576, + 539, + 598 + ], + "blocks": [ + { + "bbox": [ + 479, + 576, + 539, + 598 + ], + "lines": [ + { + "bbox": [ + 479, + 576, + 539, + 598 + ], + "spans": [ + { + "bbox": [ + 479, + 576, + 539, + 598 + ], + "type": "image", + "image_path": "f117fa0e6923e2a02ea3430a2447ef94d4f34190683cb23ec6b64bae8813e50e.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 601, + 545, + 659 + ], + "lines": [ + { + "bbox": [ + 45, + 601, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 45, + 601, + 545, + 659 + ], + "type": "text", + "content": "Figure 4: Reconstruction results by the proposed method. From left to right: Input measurement point clouds, extracted geometric primitives (different colors stand for different primitives), output reconstruction models, recovered sharp geometry along with the original point clouds, and colored models showing the error distribution regarding the Hausdorff distance. Our method produces high-quality reconstructions and preserves the intrinsic primitive structures as well as the sharp features well." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "bbox": [ + 46, + 672, + 287, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 672, + 287, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 672, + 287, + 708 + ], + "type": "text", + "content": "our method produces high-quality sharp feature lines attained by patch intersections, as shown in the forth column of Fig. 4. This is an important advantage for many practi" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 305, + 672, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 672, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 305, + 672, + 547, + 708 + ], + "type": "text", + "content": "cal applications such as sketch-style rendering, as it ensures that the reconstructed object retains its original shape and does not appear distorted or blurry. The fifth column of" + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14176" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 75, + 535, + 203 + ], + "blocks": [ + { + "bbox": [ + 53, + 62, + 539, + 74 + ], + "lines": [ + { + "bbox": [ + 53, + 62, + 539, + 74 + ], + "spans": [ + { + "bbox": [ + 53, + 62, + 539, + 74 + ], + "type": "text", + "content": "Table 1: Statistical results of the quantity and the runtime (in second) during reconstruction of the CAD models in Fig. 4." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 75, + 535, + 203 + ], + "lines": [ + { + "bbox": [ + 56, + 75, + 535, + 203 + ], + "spans": [ + { + "bbox": [ + 56, + 75, + 535, + 203 + ], + "type": "table", + "html": "
Models Metrics(a)(b)(c)(d)(e)(f)(g)(h)
# Points of the input point cloud10k593k10k10k74k7k16k14k
# Extracted primitives2853212034322927
# Generated candidate patches36466024103072349704177501092
# Active patches after pruning224533697414172586342472667
Patch decrease (%)38.42644.07328.05842.32447.96817.98637.06738.919
# Determined active patches of the output model605915191935339986114
Mesh arrangement (s)7.85321.7522.4631.7275.2362.9355.4311.421
Patch pruning (s)0.92715.4210.7920.4763.0430.4881.1310.726
Surface extraction (s)1.2882.0860.6590.0911.1120.1570.3140.137
", + "image_path": "974fcc9e3b019c9e70f5a3cc51179a7771d7913be20cf1ed89fa759b35259810.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 214, + 104, + 256 + ], + "blocks": [ + { + "bbox": [ + 52, + 214, + 104, + 256 + ], + "lines": [ + { + "bbox": [ + 52, + 214, + 104, + 256 + ], + "spans": [ + { + "bbox": [ + 52, + 214, + 104, + 256 + ], + "type": "image", + "image_path": "0fa71908fe31c1fc14a0f0b0e2716f8dba99fccb06c04cd220df91c4e9d9c12e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 258, + 104, + 299 + ], + "blocks": [ + { + "bbox": [ + 52, + 258, + 104, + 299 + ], + "lines": [ + { + "bbox": [ + 52, + 258, + 104, + 299 + ], + "spans": [ + { + "bbox": [ + 52, + 258, + 104, + 299 + ], + "type": "image", + "image_path": "aea85b175c3958133c9f8ab856700231ba14dca5949a652fa66afa353d2f86fd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 304, + 97, + 317 + ], + "lines": [ + { + "bbox": [ + 56, + 304, + 97, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 97, + 317 + ], + "type": "inline_equation", + "content": "\\sigma = 0.01(0.82m)" + }, + { + "bbox": [ + 56, + 304, + 97, + 317 + ], + "type": "inline_equation", + "content": "\\mathrm{SMH} = 0.11\\%" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 115, + 214, + 167, + 256 + ], + "blocks": [ + { + "bbox": [ + 115, + 214, + 167, + 256 + ], + "lines": [ + { + "bbox": [ + 115, + 214, + 167, + 256 + ], + "spans": [ + { + "bbox": [ + 115, + 214, + 167, + 256 + ], + "type": "image", + "image_path": "9976b643d3d0b5c3f807bf2a9af81e9c65abd30aec58072e9bbdd3faea7d0b5f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 115, + 257, + 164, + 298 + ], + "blocks": [ + { + "bbox": [ + 115, + 257, + 164, + 298 + ], + "lines": [ + { + "bbox": [ + 115, + 257, + 164, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 257, + 164, + 298 + ], + "type": "image", + "image_path": "8cf39710a210f9d567ee3cb24177cc45abd17a4083783cc255105801ac8120be.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 305, + 159, + 317 + ], + "lines": [ + { + "bbox": [ + 117, + 305, + 159, + 317 + ], + "spans": [ + { + "bbox": [ + 117, + 305, + 159, + 317 + ], + "type": "inline_equation", + "content": "\\sigma = 0.02" + }, + { + "bbox": [ + 117, + 305, + 159, + 317 + ], + "type": "inline_equation", + "content": "(1.64\\mathrm{m})" + }, + { + "bbox": [ + 117, + 305, + 159, + 317 + ], + "type": "inline_equation", + "content": "\\mathrm{SMH} = 0.11\\%" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 179, + 214, + 231, + 256 + ], + "blocks": [ + { + "bbox": [ + 179, + 214, + 231, + 256 + ], + "lines": [ + { + "bbox": [ + 179, + 214, + 231, + 256 + ], + "spans": [ + { + "bbox": [ + 179, + 214, + 231, + 256 + ], + "type": "image", + "image_path": "73ac6ed6fee6cbff3a9e0c9fb855f1dd847eb0178e1f765678777f0be2da5f34.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 180, + 257, + 231, + 298 + ], + "blocks": [ + { + "bbox": [ + 180, + 257, + 231, + 298 + ], + "lines": [ + { + "bbox": [ + 180, + 257, + 231, + 298 + ], + "spans": [ + { + "bbox": [ + 180, + 257, + 231, + 298 + ], + "type": "image", + "image_path": "fe7226a44429187dbb1cc7c4e190a4e6d323731e387c6f4c40e0ffa80cb62fd8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "text", + "content": "Figure 5: Robustness test against noise disturbance. Top: Input point clouds with increasing noise. Bottom: Our reconstruction results along with the noise level " + }, + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 322, + 287, + 369 + ], + "type": "text", + "content": " and the reconstruction error SMH." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 239, + 213, + 294, + 257 + ], + "blocks": [ + { + "bbox": [ + 239, + 213, + 294, + 257 + ], + "lines": [ + { + "bbox": [ + 239, + 213, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 239, + 213, + 294, + 257 + ], + "type": "image", + "image_path": "6f0f924ea69fc10392ed3f4224354d9bced3241286753adbacffcb9d902bd949.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 242, + 258, + 293, + 298 + ], + "blocks": [ + { + "bbox": [ + 184, + 305, + 225, + 317 + ], + "lines": [ + { + "bbox": [ + 184, + 305, + 225, + 317 + ], + "spans": [ + { + "bbox": [ + 184, + 305, + 225, + 317 + ], + "type": "inline_equation", + "content": "\\sigma = 0.04(3.28m)" + }, + { + "bbox": [ + 184, + 305, + 225, + 317 + ], + "type": "inline_equation", + "content": "\\mathrm{SMH} = 0.12\\%" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 242, + 258, + 293, + 298 + ], + "lines": [ + { + "bbox": [ + 242, + 258, + 293, + 298 + ], + "spans": [ + { + "bbox": [ + 242, + 258, + 293, + 298 + ], + "type": "image", + "image_path": "aa2f83ddf29487e253687ad986a039ea7c395d57eca64814dd67e8a7218c560c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 305, + 287, + 317 + ], + "lines": [ + { + "bbox": [ + 246, + 305, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 246, + 305, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\sigma = 0.08(6.56m)" + }, + { + "bbox": [ + 246, + 305, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\mathrm{SMH} = 46\\%" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 446, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 543 + ], + "type": "text", + "content": "Statistical results. Table 1 summarizes the quantitative results of our proposed method for the models in Fig. 4. Our algorithm significantly reduces candidate patches, especially for complex models like Fig. 4(b),(d),(e), where the patch decrease ratio exceeds " + }, + { + "bbox": [ + 46, + 446, + 287, + 543 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 446, + 287, + 543 + ], + "type": "text", + "content": ". In terms of efficiency, patch pruning and surface extraction outperform mesh arrangement, and further improvements can be made by exploring a more efficient intersection mechanism." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "Robustness assessment. We also assess the ability of the proposed algorithm against noise disturbance and occlusion data. As depicted in Fig. 5, we increase the standard deviation " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " of Gaussian noise from " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "0.01(0.82m)" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "0.04(3.28m)" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ", where values in parentheses are real intensity relative to the bounding sphere's radius of the model. Our method consistently generates accurate reconstructions in terms of both topological structures and approximation distance (SMH), suggesting its high resilience to noise. Even when " + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\sigma = 0.08(6.56m)" + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": ", which has caused significant topological alteration of the original point cloud, our method still maintains the overall structures of the object. Additionally, we verify the robustness of our method to occlusion" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 313, + 214, + 354, + 264 + ], + "blocks": [ + { + "bbox": [ + 313, + 214, + 354, + 264 + ], + "lines": [ + { + "bbox": [ + 313, + 214, + 354, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 354, + 264 + ], + "type": "image", + "image_path": "fe77783fb2c125a78cfeb45ea320f2e49a9a7419540611125ca4edd63f3e197b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 313, + 267, + 354, + 322 + ], + "blocks": [ + { + "bbox": [ + 46, + 383, + 287, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 383, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 287, + 431 + ], + "type": "text", + "content": "Fig. 4 visualizes the reconstruction error of our algorithm with respect to the Hausdorff distance, where our method consistently produces accurate approximations for all objects, demonstrating its stability and effectiveness." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 267, + 354, + 322 + ], + "lines": [ + { + "bbox": [ + 313, + 267, + 354, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 354, + 322 + ], + "type": "image", + "image_path": "56f2a99d447e3865ab2fe4a231085d6259bc7c207e801d8d8b6df1262ed1b405.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 361 + ], + "type": "text", + "content": "Figure 6: Robustness test against occlusion data. Top: Input point clouds with increasing occlusions. Bottom: Our reconstruction results along with the reconstruction error." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 371, + 214, + 411, + 264 + ], + "blocks": [ + { + "bbox": [ + 371, + 214, + 411, + 264 + ], + "lines": [ + { + "bbox": [ + 371, + 214, + 411, + 264 + ], + "spans": [ + { + "bbox": [ + 371, + 214, + 411, + 264 + ], + "type": "image", + "image_path": "6c7edd5ef763cd9cbd1e758fa3c09cf01fc1eddeb94e0ec0a78411df7ad33d17.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 372, + 267, + 412, + 322 + ], + "blocks": [ + { + "bbox": [ + 372, + 267, + 412, + 322 + ], + "lines": [ + { + "bbox": [ + 372, + 267, + 412, + 322 + ], + "spans": [ + { + "bbox": [ + 372, + 267, + 412, + 322 + ], + "type": "image", + "image_path": "56b8ea4ef2dd450b6413795466611f6277dfa45a64b253611883ce0dda4b10fa.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 432, + 214, + 473, + 263 + ], + "blocks": [ + { + "bbox": [ + 432, + 214, + 473, + 263 + ], + "lines": [ + { + "bbox": [ + 432, + 214, + 473, + 263 + ], + "spans": [ + { + "bbox": [ + 432, + 214, + 473, + 263 + ], + "type": "image", + "image_path": "9d999a2334889471094c890a426e2776c6ab8cc66af128e1ac2dcf34fa99effc.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 435, + 265, + 475, + 322 + ], + "blocks": [ + { + "bbox": [ + 435, + 265, + 475, + 322 + ], + "lines": [ + { + "bbox": [ + 435, + 265, + 475, + 322 + ], + "spans": [ + { + "bbox": [ + 435, + 265, + 475, + 322 + ], + "type": "image", + "image_path": "160dd7fbaaa67577e88f48d686662c89e0fa3018345e1bea6c1ff9bbcec27368.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 495, + 214, + 534, + 263 + ], + "blocks": [ + { + "bbox": [ + 495, + 214, + 534, + 263 + ], + "lines": [ + { + "bbox": [ + 495, + 214, + 534, + 263 + ], + "spans": [ + { + "bbox": [ + 495, + 214, + 534, + 263 + ], + "type": "image", + "image_path": "6f562a790c01adba344ce0c9f25dbb7ff8244f1057c1c2b8f272634c9093f976.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 497, + 265, + 537, + 322 + ], + "blocks": [ + { + "bbox": [ + 497, + 265, + 537, + 322 + ], + "lines": [ + { + "bbox": [ + 497, + 265, + 537, + 322 + ], + "spans": [ + { + "bbox": [ + 497, + 265, + 537, + 322 + ], + "type": "image", + "image_path": "cd5e33066618cdcfb154a6629f2b116b6be67e3537510ab2ad7a912269160573.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 328, + 411, + 523, + 448 + ], + "blocks": [ + { + "bbox": [ + 305, + 373, + 545, + 409 + ], + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 409 + ], + "type": "text", + "content": "Table 2: Comparisons with state-of-the-art approaches, where \"w\" and \"w/o\" indicate with and without the patch pruning, separately." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 411, + 523, + 448 + ], + "lines": [ + { + "bbox": [ + 328, + 411, + 523, + 448 + ], + "spans": [ + { + "bbox": [ + 328, + 411, + 523, + 448 + ], + "type": "table", + "html": "
MetricsiPSR[20]SIREN[45]DSE[40]KSR[3]Ours
ww/o
SMH (%)4.2341.793.15169.442.452.45
Time (s)201.5891208.124303.865570.46845.556.777
", + "image_path": "a86d988ba0bcd37f211a107147f0e7a89974730dbc35224d598ad9b9535099d4.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 457, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 529 + ], + "type": "text", + "content": "sion data. Fig. 6 reports our reconstruction results after progressively erasing point clouds. As observed, our method shows remarkable resistance to occlusions and has successfully compensated the missing structures. Conversely, this is also quite helpful to complete the imperfect point clouds by sampling points in the reconstruction models." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "Comparisons. Subsequently, we conduct a comparative analysis of our method with representative state-of-the-art approaches, including densely implicit reconstruction methods iPSR [20], DSE [40], and SIREN [45], as well as the plane-assembly approach KSR [3]. The quantitative comparison results in terms of reconstruction accuracy and efficiency on the Cheese model is presented in Table 2. The results demonstrate that our method achieves competitive reconstruction accuracy with implicit approaches iPSR and DSE, and outperforms SIREN and KSR by a large margin. Additionally, our algorithm performs significantly faster than competitors, even without patch pruning. The qualitative reconstruction results presented in Fig. 7 show that benefiting from the introduced primitive assembly, our" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14177" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 60, + 126, + 184 + ], + "blocks": [ + { + "bbox": [ + 62, + 60, + 126, + 184 + ], + "lines": [ + { + "bbox": [ + 62, + 60, + 126, + 184 + ], + "spans": [ + { + "bbox": [ + 62, + 60, + 126, + 184 + ], + "type": "image", + "image_path": "fa05c046cb730132e2746dd8aaa5698e2d6ba0543347c305a725933bada9fdde.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 187, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 545, + 213 + ], + "type": "text", + "content": "Figure 7: Qualitative comparisons with state-of-the-art approaches. Our method enables the shape reconstruction faithfully and successfully recovers the sharp geometry (e.g., feature lines and corners) from unorganized point clouds." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 142, + 60, + 208, + 184 + ], + "blocks": [ + { + "bbox": [ + 142, + 60, + 208, + 184 + ], + "lines": [ + { + "bbox": [ + 142, + 60, + 208, + 184 + ], + "spans": [ + { + "bbox": [ + 142, + 60, + 208, + 184 + ], + "type": "image", + "image_path": "e125e2ff80a78fcad8357fde7b939c0d75b2578658b2e5636d226f53b4fdb894.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 224, + 60, + 289, + 183 + ], + "blocks": [ + { + "bbox": [ + 224, + 60, + 289, + 183 + ], + "lines": [ + { + "bbox": [ + 224, + 60, + 289, + 183 + ], + "spans": [ + { + "bbox": [ + 224, + 60, + 289, + 183 + ], + "type": "image", + "image_path": "e30e0715d4535abacf4766dabea16958dec36614113714d703520e56b961ec08.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 60, + 370, + 182 + ], + "blocks": [ + { + "bbox": [ + 309, + 60, + 370, + 182 + ], + "lines": [ + { + "bbox": [ + 309, + 60, + 370, + 182 + ], + "spans": [ + { + "bbox": [ + 309, + 60, + 370, + 182 + ], + "type": "image", + "image_path": "6ff753b28d8961edc3e10ec28c5a722c86809e864e443e4df95e37e8f16b0ea5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 388, + 60, + 446, + 182 + ], + "blocks": [ + { + "bbox": [ + 388, + 60, + 446, + 182 + ], + "lines": [ + { + "bbox": [ + 388, + 60, + 446, + 182 + ], + "spans": [ + { + "bbox": [ + 388, + 60, + 446, + 182 + ], + "type": "image", + "image_path": "d6f528e3486add7164707d83501c092377ac2a1d3b2c7a095445b4310010408b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 472, + 60, + 531, + 182 + ], + "blocks": [ + { + "bbox": [ + 472, + 60, + 531, + 182 + ], + "lines": [ + { + "bbox": [ + 472, + 60, + 531, + 182 + ], + "spans": [ + { + "bbox": [ + 472, + 60, + 531, + 182 + ], + "type": "image", + "image_path": "d305bfbd766ef046c7b2e150e11aa61e6e9a8d3b232cd17b767ed0b0b8adb2ed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 49, + 240, + 287, + 271 + ], + "blocks": [ + { + "bbox": [ + 47, + 228, + 286, + 239 + ], + "lines": [ + { + "bbox": [ + 47, + 228, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 286, + 239 + ], + "type": "text", + "content": "Table 3: Quantitative comparisons on the ABC300 dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 240, + 287, + 271 + ], + "lines": [ + { + "bbox": [ + 49, + 240, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 49, + 240, + 287, + 271 + ], + "type": "table", + "html": "
Methods MetricsiPSR[20]SIREN[45]DSE[40]KSR[3]ComplexGen[17]CAPRI-Net[48]Ours
SMH (%)2.7124.894.0626.4324.2415.874.81
Time (s)24.0641565.21732.8004.208553.089183.8443.728
", + "image_path": "3a78679cd70373e09992b1063403b4dd171bd812edc2f7d279da315844945fe8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 316 + ], + "type": "text", + "content": "method successfully recovers sharp features from unorganized point clouds, which is a salient advantage compared to implicit methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 316, + 288, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 436 + ], + "type": "text", + "content": "We also conduct quantitative comparisons with several predecessors (additionally including CAPRI-Net [48] and ComplexGen [17]) on 300 CAD models from the ABC dataset (abbreviated as ABC300). Table 3 demonstrates that our method achieves consistently competitive reconstruction accuracy with implicit approaches such as iPSR and DSE (while also providing advantages for sharp feature recovery and model editing), and significantly surpasses KSR, CAPRI-Net, and ComplexGen. Furthermore, our method is considerably faster than its competitors." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 52, + 462, + 97, + 503 + ], + "blocks": [ + { + "bbox": [ + 52, + 462, + 97, + 503 + ], + "lines": [ + { + "bbox": [ + 52, + 462, + 97, + 503 + ], + "spans": [ + { + "bbox": [ + 52, + 462, + 97, + 503 + ], + "type": "image", + "image_path": "57819ac51dffa0a3feb27fb698e57b63453f2902e65d4a29efa06c8dd692a768.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 110, + 448, + 157, + 506 + ], + "blocks": [ + { + "bbox": [ + 110, + 448, + 157, + 506 + ], + "lines": [ + { + "bbox": [ + 110, + 448, + 157, + 506 + ], + "spans": [ + { + "bbox": [ + 110, + 448, + 157, + 506 + ], + "type": "image", + "image_path": "440759cec00bc591d814faf9e4ee642d43fcd342b2701449ebd78f4e3e90166a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 509, + 287, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 544 + ], + "type": "text", + "content": "Figure 8: Model editing based on our structure-aware reconstruction algorithm. (a) The original reconstruction object. (b)-(d) The newly edited results." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 165, + 446, + 221, + 506 + ], + "blocks": [ + { + "bbox": [ + 165, + 446, + 221, + 506 + ], + "lines": [ + { + "bbox": [ + 165, + 446, + 221, + 506 + ], + "spans": [ + { + "bbox": [ + 165, + 446, + 221, + 506 + ], + "type": "image", + "image_path": "187111fa2dadba5ad3b66c877006ebfc6b0f0060d7828af5bb1df6f0533aaced.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 230, + 437, + 287, + 505 + ], + "blocks": [ + { + "bbox": [ + 230, + 437, + 287, + 505 + ], + "lines": [ + { + "bbox": [ + 230, + 437, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 230, + 437, + 287, + 505 + ], + "type": "image", + "image_path": "e579ee14a8f29a2453a7ac2bf5980b79cca03347e6145953e8a4331160d3c12e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "content": "Balance parameter. We investigate the impact of " + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "content": " on surface extraction by varying its values to different settings (see Fig. 9(a)). It can be seen that increasing " + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "content": " brings the reduction of the complexity of the output model. However, this also degrades the reconstruction accuracy. For example, when " + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 555, + 287, + 640 + ], + "type": "text", + "content": " is set to 0.7, the inner hollows of the input point clouds disappear." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "Application to model editing. Our reconstruction method can be employed for flexible model editing by taking advantage of the structure-aware merits. As shown in Fig. 8, since each quadratic structure of the object is preserved along with their detailed geometric parameters, one" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 312, + 243, + 360, + 291 + ], + "blocks": [ + { + "bbox": [ + 308, + 279, + 318, + 289 + ], + "lines": [ + { + "bbox": [ + 308, + 279, + 318, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 279, + 318, + 289 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 232, + 346, + 242 + ], + "lines": [ + { + "bbox": [ + 326, + 232, + 346, + 242 + ], + "spans": [ + { + "bbox": [ + 326, + 232, + 346, + 242 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 243, + 360, + 291 + ], + "lines": [ + { + "bbox": [ + 312, + 243, + 360, + 291 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 360, + 291 + ], + "type": "image", + "image_path": "4df7b8eafaa7e27a489eeaaa56cd3e2e193162ae3329ba3a0911fe85537a998e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 296, + 545, + 343 + ], + "lines": [ + { + "bbox": [ + 304, + 296, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 545, + 343 + ], + "type": "text", + "content": "Figure 9: (a) Impact of parameter " + }, + { + "bbox": [ + 304, + 296, + 545, + 343 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 296, + 545, + 343 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 304, + 296, + 545, + 343 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 296, + 545, + 343 + ], + "type": "text", + "content": " is increased, the complexity of the output model decreases. (b) In sparse input scenarios, small primitives can be easily lost, leading to failures." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 370, + 245, + 414, + 289 + ], + "blocks": [ + { + "bbox": [ + 378, + 232, + 406, + 241 + ], + "lines": [ + { + "bbox": [ + 378, + 232, + 406, + 241 + ], + "spans": [ + { + "bbox": [ + 378, + 232, + 406, + 241 + ], + "type": "inline_equation", + "content": "\\lambda = 0.2" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 245, + 414, + 289 + ], + "lines": [ + { + "bbox": [ + 370, + 245, + 414, + 289 + ], + "spans": [ + { + "bbox": [ + 370, + 245, + 414, + 289 + ], + "type": "image", + "image_path": "0569f070d0f6812f0c8582ba61da016e8eae8366895c3e07ad94a5656a4cb85e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 426, + 245, + 470, + 289 + ], + "blocks": [ + { + "bbox": [ + 432, + 232, + 459, + 240 + ], + "lines": [ + { + "bbox": [ + 432, + 232, + 459, + 240 + ], + "spans": [ + { + "bbox": [ + 432, + 232, + 459, + 240 + ], + "type": "inline_equation", + "content": "\\lambda = 0.7" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 245, + 470, + 289 + ], + "lines": [ + { + "bbox": [ + 426, + 245, + 470, + 289 + ], + "spans": [ + { + "bbox": [ + 426, + 245, + 470, + 289 + ], + "type": "image", + "image_path": "05474ab7c9720fa731cd3b47c9a8d6f53ad8eb32e3e4c51809f4e45808322281.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 478, + 240, + 555, + 289 + ], + "blocks": [ + { + "bbox": [ + 479, + 232, + 511, + 240 + ], + "lines": [ + { + "bbox": [ + 479, + 232, + 511, + 240 + ], + "spans": [ + { + "bbox": [ + 479, + 232, + 511, + 240 + ], + "type": "text", + "content": "#Point=10k" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 478, + 240, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 478, + 240, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 478, + 240, + 555, + 289 + ], + "type": "image", + "image_path": "5267e588ffc774d6484e2e637d88e441f6b8adc46dc6f3dbaea560c34f7a991d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 398 + ], + "type": "text", + "content": "can modify the reconstruction shape by simply specifying various parameters. For instance, by re-defining the radius or height of the cylindrical primitives in Fig. 8(a), we can resize them to our desirable shapes (Fig. 8(b)-(d))." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 415, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 499 + ], + "type": "text", + "content": "Limitations. Our approach primarily focuses on the rapid and accurate reconstruction of representative primitive structures in CAD models. However, as shown in Fig 9(b), our reconstruction quality may deteriorate for inputs lacking explicit primitives. Additionally, inadequate primitive extraction due to severe occlusion or noise can degrade the quality of the reconstruction results." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 513, + 470, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 470, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 470, + 525 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 676 + ], + "type": "text", + "content": "We propose an accurate and fast shape reconstruction pipeline for converting disordered 3D point clouds into manifold, watertight surfaces. Our method preserves representative intrinsic primitive structures in CAD systems through primitive assembly, avoiding dense implicit reconstructions and extending piecewise planar approximations. We also develop an effective patch pruning mechanism, greatly expediting surface extraction by reducing patch intersections. Additionally, our method is highly resilient to noise and can reliably recover sharp object geometry. Extensive experiments show our algorithm's superiority over competing methods." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "In our future work, we plan to expand the current primitive types to general primitives, and even spline surfaces, in order to efficiently process more complex CAD models." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14178" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 149, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 163 + ], + "type": "text", + "content": "This work was partially funded by the National Key Research and Development Program (2021YFB1715900), the CAS Project for Young Scientists in Basic Research (YSBR-034), the National Natural Science Foundation of China (62172415, 62272277, 12022117), and the HKU-100 Research Award." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 105, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 105, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 105, + 186 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 194, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 239 + ], + "type": "text", + "content": "[1] Murat Arikan, Michael Schwärzler, Simon Flöry, Michael Wimmer, and Stefan Maierhofer. O-snap: Optimization-based snapping for modeling architecture. ACM Trans. Graph., 32(1):6:1-6:15, 2013. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 273 + ], + "type": "text", + "content": "[2] Gavin Barill, Neil G Dickson, Ryan Schmidt, David IW Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. ACM Trans. Graph., 37(4):43:1-43:12, 2018. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 287, + 306 + ], + "type": "text", + "content": "[3] Jean-Philippe Bauchet and Florent Lafarge. Kinetic shape reconstruction. ACM Trans. Graph., 39(5):156:1-156:14, 2020. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 287, + 352 + ], + "type": "text", + "content": "[4] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 353, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 287, + 407 + ], + "type": "text", + "content": "[5] Matthew Berger, Andrea Tagliasacchi, Lee M Seversky, Pierre Alliez, Gael Guennebaud, Joshua A Levine, Andrei Sharf, and Claudio T Silva. A survey of surface reconstruction from point clouds. Comput. Graph. Forum, 36(1):301-329, 2017. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 453 + ], + "type": "text", + "content": "[6] Alexandre Boulch, Martin de La Gorce, and Renaud Marlet. Piecewise-planar 3d reconstruction with edge and corner regularization. Comput. Graph. Forum, 33(5):55-64, 2014. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 509 + ], + "type": "text", + "content": "[7] Anne-Laure Chauve, Patrick Labatut, and Jean-Philippe Pons. Robust piecewise-planar 3d reconstruction and completion from large-scale unstructured point data. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1261-1268, 2010. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 510, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 542 + ], + "type": "text", + "content": "[8] Jie Chen and Baoquan Chen. Architectural modeling from sparsely scanned range data. Int. J. Comput. Vis., 78(2):223-236, 2008. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 577 + ], + "type": "text", + "content": "[9] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "text", + "content": "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5939-5948, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 644 + ], + "type": "text", + "content": "[11] David Cohen-Steiner, Pierre Alliez, and Mathieu Desbrun. Variational shape approximation. ACM Trans. Graph., 23(3):905-914, 2004. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[12] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In Eur. Conf. Comput. Vis., pages 108-124, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[13] Hao Fang and Florent Lafarge. Connect-and-slice: an hybrid approach for reconstructing 3d objects. In IEEE Conf." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Comput. Vis. Pattern Recog., pages 13490-13498, 2020. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "type": "text", + "content": "[14] Blender Foundation. Blender - a 3D modelling and rendering package. http://www.blender.org.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 119, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 119, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 547, + 140 + ], + "type": "text", + "content": "[15] Leonidas Guibas. Kinetic data structures. In Handbook of Data Structures and Applications, pages 377-388. 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 142, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 142, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 142, + 545, + 185 + ], + "type": "text", + "content": "[16] Leonidas Guibas, Lyle Ramshaw, and Jorge Stolfi. A kinetic framework for computational geometry. In IEEE Annual Symposium on Foundations of Computer Science, pages 100-111, 1983. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "type": "text", + "content": "[17] Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. Complexgen: Cad reconstruction by b-rep chain complex generation. ACM Trans. Graph., 41(4):129:1-129:18, 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 231, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 253 + ], + "type": "text", + "content": "[18] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual. https://www.gurobi.com.5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 255, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 545, + 297 + ], + "type": "text", + "content": "[19] Hugues Hoppe, Tony DeRose, Tom Duchamp, John McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. In Proc. ACM SIGGRAPH, pages 71-78, 1992. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "text", + "content": "[20] Fei Hou, Chiyu Wang, Wencheng Wang, Hong Qin, Chen Qian, and Ying He. Iterative poisson surface reconstruction (ipsr) for unoriented points. ACM Trans. Graph., 41(4):128:1-128:13, 2022. 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 344, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 387 + ], + "type": "text", + "content": "[21] Hui Huang, Dan Li, Hao Zhang, Uri Ascher, and Daniel Cohen-Or. Consolidation of unorganized point clouds for surface reconstruction. ACM Trans. Graph., 28(5):1-7, 2009. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 421 + ], + "type": "text", + "content": "[22] Alec Jacobson and Daniele Panozzo. Libigl: Prototyping geometry processing research in c++. In SIGGRAPH Asia 2017 courses, pages 1-172. 2017. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "text", + "content": "[23] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3d data. Comput. Graph. Forum, 38(1):167-196, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 468, + 545, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 500 + ], + "type": "text", + "content": "[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Comput. Graph. Forum (Proceedings of SGP), page 61-70, 2006. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 501, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 532 + ], + "type": "text", + "content": "[25] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. Graph., 32(3):29:1-29:13, 2013. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "text", + "content": "[26] Florent Lafarge and Pierre Alliez. Surface reconstruction through point set structuring. Comput. Graph. Forum, 32(2pt2):225-234, 2013. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "type": "text", + "content": "[27] Minglei Li, Liangliang Nan, Neil Smith, and Peter Wonka. Reconstructing building mass models from UAV images. Comput. Graph., 54:84-93, 2016. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 602, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 635 + ], + "type": "text", + "content": "[28] Minglei Li, Peter Wonka, and Liangliang Nan. Manhattanworld urban reconstruction from point clouds. In Eur. Conf. Comput. Vis., pages 54-69, 2016. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "type": "text", + "content": "[29] Yangyan Li, Xiaokun Wu, Yiorgos Chrysathou, Andrei Sharf, Daniel Cohen-Or, and Niloy J Mitra. Globfit: Consistently fitting primitives by discovering global relations. ACM Trans. Graph., 30(4):52:1-52:12, 2011. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[30] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proc. ACM SIGGRAPH, volume 21, pages 163-169. 3" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14179" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] David Marshall, Gabor Lukacs, and Ralph Martin. Robust segmentation of primitives from range data in the presence of geometric degeneracy. IEEE Trans. Pattern Anal. Mach. Intell., 23(3):304-314, 2001. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "text", + "content": "[32] Aron Monszpart, Nicolas Mellado, Gabriel J Brostow, and Niloy J Mitra. Rapter: rebuilding man-made scenes with regular arrangements of planes. ACM Trans. Graph., 34(4):103:1-103:12, 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[33] Claudio Mura, Oliver Mattausch, and Renato Pajarola. Piecewise-planar reconstruction of multi-room interiors with arbitrary wall arrangements. Comput. Graph. Forum, 35(7):179-188, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 251 + ], + "type": "text", + "content": "[34] Liangliang Nan, Andrei Sharf, Hao Zhang, Daniel Cohen-Or, and Baoquan Chen. Smartboxes for interactive urban reconstruction. ACM Trans. Graph., 29(4):93:1–93:10, 2010. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "type": "text", + "content": "[35] Liangliang Nan and Peter Wonka. Polyfit: Polygonal surface reconstruction from point clouds. In IEEE Int. Conf. Comput. Vis., pages 2353-2361, 2017. 1, 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "text", + "content": "[36] Anh Nguyen and Bac Le. 3d point cloud segmentation: A survey. In IEEE Int. Conf. Robot. Autom. Mechatron, pages 225-230, 2013. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 353 + ], + "type": "text", + "content": "[37] Sven Oesau, Florent Lafarge, and Pierre Alliez. Planar shape detection and regularization in tandem. Comput. Graph. Forum, 35(1):203-215, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 408 + ], + "type": "text", + "content": "[38] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 165-174, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "text", + "content": "[39] Tahir Rabbani, Frank Van Den Heuvel, and George Vosselmann. Segmentation of point clouds using smoothness constraint. ISPRS J. PhotoGramm., 36(5):248-253, 2006. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 509 + ], + "type": "text", + "content": "[40] Marie-Julie Rakotosaona, Paul Guerrero, Noam Aigerman, Niloy J. Mitra, and Maks Ovsjanikov. Learning delaunay surface elements for mesh reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22-31, June 2021. 3, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 555 + ], + "type": "text", + "content": "[41] Chiara Romanengo, Andrea Raffo, Yifan Qie, Nabil Anwer, and Bianca Falcidieno. Fit4cad: A point cloud benchmark for fitting simple geometric primitives in cad objects. Comput. Graph., 102:133-143, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 600 + ], + "type": "text", + "content": "[42] Falko Schindler, Wolfgang Wörstner, and Jan-Michael Frahm. Classification and reconstruction of surfaces from point clouds of man-made objects. In IEEE Int. Conf. Comput. Vis. Worksh., pages 257-263, 2011. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 633 + ], + "type": "text", + "content": "[43] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient ransac for point-cloud shape detection. Comput. Graph. Forum, 26(2):214-226, 2007. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 657 + ], + "type": "text", + "content": "[44] Gerard Sierksma and Yori Zwols. Linear and integer optimization: theory and practice. 2015. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[45] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Int. Conf. Neur. Info. Process. Sys., volume 33, pages 7462-7473, 2020. 3, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 353 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[46] Marc Van Kreveld, Thijs Van Lankveld, and Remco C. Veltkamp. On the shape of a set of points and lines in the plane. Comput. Graph. Forum, 30(5):1553-1562, 2011. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[47] Dong-Ming Yan, Wenping Wang, Yang Liu, and Zhouwang Yang. Variational mesh segmentation via quadric surface fitting. Computer-Aided Design, 44(11):1072-1082, 2012. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "text", + "content": "[48] Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. Capri-net: Learning compact cad shapes with adaptive primitive assembly. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11768-11778, 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 239 + ], + "type": "text", + "content": "[49] Mulin Yu and Florent Lafarge. Finding good configurations of planar primitives in unorganized point clouds. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6367-6376, 2022, 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 242, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 545, + 274 + ], + "type": "text", + "content": "[50] Qingnan Zhou, Eitan Grinspun, Denis Zorin, and Alec Jacobson. Mesh arrangements for solid geometry. ACM Trans. Graph., 35(4):39:1-39:15, 2016. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 307 + ], + "type": "text", + "content": "[51] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 353 + ], + "type": "text", + "content": "[52] Marco Zuliani, Charles S Kenney, and BS Manjunath. The multiransac algorithm and its application to detect planar homographies. In IEEE Int. Conf. Image Process., volume 3, pages III-153, 2005. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14180" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/88a53c18-99d0-4b4c-b0b2-16a2556a29ae_content_list.json b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/88a53c18-99d0-4b4c-b0b2-16a2556a29ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0ef7b0a74c2b478971993daf378ef2893a76dbf0 --- /dev/null +++ b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/88a53c18-99d0-4b4c-b0b2-16a2556a29ae_content_list.json @@ -0,0 +1,1957 @@ +[ + { + "type": "text", + "text": "Studying How to Efficiently and Effectively Guide Models with Explanations", + "text_level": 1, + "bbox": [ + 98, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sukrut Rao*, Moritz Böhle*, Amin Parchami-Araghi, Bernt Schiele \nMax Planck Institute for Informatics, Saarland Informatics Campus, Saarbrücken, Germany", + "bbox": [ + 122, + 181, + 844, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sukrut.rao,mboehle,mparcham,schiele}@mpi-inf.mpg.de", + "bbox": [ + 254, + 218, + 712, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite being highly performant, deep neural networks might base their decisions on features that spuriously correlate with the provided labels, thus hurting generalization. To mitigate this, 'model guidance' has recently gained popularity, i.e. the idea of regularizing the models' explanations to ensure that they are \"right for the right reasons\" [49]. While various techniques to achieve such model guidance have been proposed, experimental validation of these approaches has thus far been limited to relatively simple and / or synthetic datasets. To better understand the effectiveness of the various design choices that have been explored in the context of model guidance, in this work we conduct an in-depth evaluation across various loss functions, attribution methods, models, and 'guidance depths' on the PASCAL VOC 2007 and MS COCO 2014 datasets. As annotation costs for model guidance can limit its applicability, we also place a particular focus on efficiency. Specifically, we guide the models via bounding box annotations, which are much cheaper to obtain than the commonly used segmentation masks, and evaluate the robustness of model guidance under limited (e.g. with only $1\\%$ of annotated images) or overly coarse annotations. Further, we propose using the EPG score as an additional evaluation metric and loss function ('Energy loss'). We show that optimizing for the Energy loss leads to models that exhibit a distinct focus on object-specific features, despite only using bounding box annotations that also include background regions. Lastly, we show that such model guidance can improve generalization under distribution shifts. Code available at: https://github.com/sukrutrao/Model-Guidance", + "bbox": [ + 75, + 300, + 473, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 773, + 209, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) excel at learning predictive features that allow them to correctly classify a set of training images with ease. The features learnt on the training set, however, do not necessarily transfer to unseen images: i.e., instead of learning the actual class-relevant fea", + "bbox": [ + 75, + 800, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/06af46ab131b75d0fd79d635858da8b202be74af1806d7c84ba16998e820671e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 268, + 869, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46ae3bb8b102d1418ad1f9744ad1bebe4dd1ffcf4c52f20827d5ed7e307137f8.jpg", + "image_caption": [ + "Fig. 1: (a) Model guidance increases object focus. Models may rely on irrelevant background features or spurious correlations (e.g. presence of person provides positive evidence for bicycle, center row, col. 1). Guiding the model via bounding box annotations can mitigate this and consistently increases the focus on object features (bottom row). (b) Model guidance can improve accuracy. In the presence of spurious correlations in the training data, non-guided models might focus on the wrong features. In the example image in (b), the waterbird is incorrectly classified to be a landbird due to the background (col. 3). Guiding the model via bounding box annotation (as shown in col. 2), the model can be guided to focus on the bird features for classification (col. 4)." + ], + "image_footnote": [], + "bbox": [ + 522, + 426, + 575, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b9d6a75277e78fcfdc7194a485eede0fb298b87da971f1522c1d156a35ac17eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 428, + 648, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c76c9e34f23c307f6b917e14d755c7c4d169d55763d83ba0c6aeac1c91e3fdd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 417, + 727, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/43a4783bb4bbe5f99ae90e6e5b08395ed5cdd3d143d363a78bee7e472b4c3797.jpg", + "image_caption": [ + "Landbird Confidence: $97\\%$" + ], + "image_footnote": [], + "bbox": [ + 736, + 417, + 807, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f26565be58284a21fe4aa0090c377bd3239e308462cce93e1d4617623e1213d6.jpg", + "image_caption": [ + "Waterbird Confidence: $87\\%$" + ], + "image_footnote": [], + "bbox": [ + 813, + 417, + 885, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tures, DNNs might memorize individual images (cf. [18]) or exploit spurious correlations in the training data (cf. [68]). For example, if bikes are highly correlated with people in the training data, a model might learn to associate the presence of a person in an image as positive evidence for a bike (e.g. Fig. 1a, col. 1, rows 1-2), which can limit how well it generalizes. Similarly, a bird classifier might rely on background features from the bird's habitat, and fail to correctly classify in a different habitat (cf. Fig. 1b cols. 1-3 and [42]).", + "bbox": [ + 496, + 681, + 892, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To detect such behaviour, recent advances in model interpretability have provided attribution methods (e.g. [53, 62, 57, 6]) to understand a model's reasoning. These methods typically provide attention maps that highlight regions of importance in an input to explain the model's decisions", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1922", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de598a16f0e7a21eb615b23f22952d798467a3441debf5c3d2cc06b5552a47e2.jpg", + "image_caption": [ + "Fig. 2: Qualitative results of model guidance. We show model-inherent B-cos explanations (input layer) of a B-cos ResNet-50 and GradCAM explanations (final layer) of a conventional ResNet-50 before ('Standard') and after optimization ('Guided') for images from the VOC test set, using our proposed Energy loss (Eq. (6)). Guiding the model via bounding box annotations consistently increases the focus on object features for both methods. Specifically, we find that background attributions are consistently suppressed in both cases." + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 870, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and can help identify incorrect reasoning such as reliance on spurious or irrelevant features, see for example Fig. 1b.", + "bbox": [ + 75, + 381, + 468, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As many attribution methods are in fact themselves differentiable (e.g. [57, 62, 53, 6]), recent work [49, 56, 24, 23, 66, 64] has explored the idea of using them to guide the models to make them \"right for the right reasons\" [49]. Specifically, models can be guided by jointly optimizing for correct classification as well as for attributing importance to regions deemed relevant by humans. This can help the model focus on the relevant features of a class, and correct errors in reasoning (Fig. 1b, col. 4). Such guidance has the added benefit of providing well-localized explanations that are thus easier to understand for end users (e.g. Fig. 2).", + "bbox": [ + 75, + 411, + 468, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While model guidance has shown promising results, a detailed study of how to do this most effectively is crucially missing. In particular, model guidance has so far been studied for a limited set of attribution methods and models and usually on relatively simple and/or synthetic datasets; further, the evaluation settings between approaches can significantly differ, which makes a fair comparison difficult.", + "bbox": [ + 75, + 578, + 468, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this work, we perform an in-depth evaluation of model guidance on large scale, real-world datasets, to better understand the effectiveness of a variety of design choices. Specifically, we evaluate model guidance along the following dimensions: the model architecture, the guidance depth1, the attribution method, and the loss function. In this context, we propose using the EPG score [67]—an evaluation metric that has thus far been used to evaluate the quality of attribution methods—as an additional loss function (which we call the Energy loss) as it is fully differentiable.", + "bbox": [ + 75, + 684, + 468, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Further, as annotation costs can be a major hurdle for making model guidance practical, we place a particular fo", + "bbox": [ + 76, + 835, + 468, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cus on efficient guidance. Specifically, we use bounding boxes instead of semantic segmentation masks, and evaluate the robustness of guidance techniques under limited or overly coarse annotations to reduce data collection costs.", + "bbox": [ + 500, + 381, + 890, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We find that our Energy loss lends itself well to those settings. On the one hand, it exhibits a high degree of robustness to limited or noisy bounding box annotations (cf. Figs. 10 and 12). On the other hand, despite the coarseness of bounding box guidance, it maintains a clear focus on object-specific features inside the bounding boxes, see Fig. 1a, row 3. In contrast, prior approaches often regularize for a uniform distribution of the attribution values inside the annotation masks, and thus tend to exhibit much lower attribution granularity (cf. Fig. 9).", + "bbox": [ + 496, + 444, + 892, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. (1) We perform an in-depth evaluation of model guidance on challenging large scale, multi-label classification datasets (PASCAL VOC 2007 [16], MS COCO 2014 [34]), assessing the impact of attribution methods, model architectures, guidance depths, and loss functions. Further, we show that, despite being relatively coarse, bounding box supervision can provide sufficient guidance to the models whilst being much cheaper to obtain than semantic segmentation masks. (2) We propose using the Energy Pointing Game (EPG) score [67] as an alternative to the IoU metric for evaluating the effectiveness of such guidance and show that the EPG score constitutes a good loss function for model guidance, particularly when using bounding boxes. (3) We show that model guidance can be performed cost-effectively by using annotation masks that are noisy or are available for only a small fraction (e.g. $1\\%$ ) of the training data. (4) We show through experiments on the Waterbirds-100 dataset [51, 42] that model guidance with a small number of annotations suffices to improve the model's generalization under distribution shifts at test time.", + "bbox": [ + 496, + 598, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1The layer at which guidance is applied, e.g. typically at the last convolutional layer for GradCAM [53] or the first layer for IxG [57].", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1923", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attribution Methods [58, 60, 62, 57, 53, 67, 13, 30, 9, 43, 20, 70, 47, 12, 4] are often used to explain black-box models by generating heatmaps that highlight input regions important to the model's decision. However, such methods are often not faithful to the model [1, 46, 31, 72, 2] and risk misleading users. Recent work proposes inherently interpretable models [8, 6] that address this by providing modelfaithful explanations by design. In our work, we use both popular post-hoc and model-inherent attribution methods to guide models and discuss their effectiveness.", + "bbox": [ + 75, + 114, + 468, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attribution Priors: Several approaches have been proposed for training better models by enforcing desirable properties on their attributions. These include enforcing consistency against augmentations [45, 44, 25], smoothness [15, 37, 32], separation of classes [71, 44, 61, 39, 59], or constraining the model's attention [22, 3]. In contrast, in this work, we focus on providing explicit human guidance to the model using bounding box annotations. This constitutes more explicit guidance but allows fine-grained control over the model's reasoning even with few annotations.", + "bbox": [ + 75, + 266, + 470, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model Guidance: In contrast to the indirect regularization effect achieved by attribution priors, various approaches have been proposed (cf. [21, 65]) to actively guide models by regularizing their attributions, for tasks such as classification [49, 24, 23, 48, 42, 26, 63, 36, 66, 64, 52, 55, 35, 56, 69, 17], segmentation [33], VQA [54, 63], and knowledge distillation [19]. The goal of such approaches is not only to improve performance, but also make sure that the model is \"right for the right reasons\" [49]. For classifiers, this typically involves jointly optimizing both for classification performance and localization to object features. While various benefits of model guidance have been reported, most prior work evaluate on simple datasets [49, 55, 24, 23] and, thus far, no common evaluation setting has emerged. Recently, [11] has extended model guidance to ImageNet, showing that its benefits can scale to large scale problems. In contrast to [11], who investigated one particular attribution method [10], our focus lies on a better understanding of the impact of the different design choices for model guidance.", + "bbox": [ + 75, + 416, + 470, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To distill the most effective techniques for model guidance, in this work, we conduct an in-depth evaluation on challenging, commonly used real-world multi-label classification datasets (PASCAL VOC 2007, MS COCO 2014). Specifically, we perform a comprehensive comparison across multiple dimensions of interest: the loss function, the model architecture, the guidance depth, and the attribution method. For this, we evaluate the localization losses introduced in the closest related work, i.e. RRR [49], HAICS [56], and GRADIA [24]; additionally, we propose using the EPG metric [67] as a loss function and show that it has various desirable properties, in particular when guiding models via bounding box annotations.", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0b0d9ba08868ff0b01f0a00e2862efa1810cab23cbb1ded746e94d3dcf66b357.jpg", + "image_caption": [ + "Fig. 3: Model guidance overview. We jointly optimize for classification $(\\mathcal{L}_{\\mathrm{class}})$ and localization of attributions to human-annotated bounding boxes $(\\mathcal{L}_{\\mathrm{loc}})$ , to guide the model to focus on object features. Various localization loss functions can be used, see Sec. 3.4." + ], + "image_footnote": [], + "bbox": [ + 498, + 87, + 893, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, model guidance has also been used to mitigate reliance on spurious features using language guidance [42], and we show that using a small number of coarse bounding box annotations can be similarly effective.", + "bbox": [ + 496, + 255, + 890, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluating Model Guidance: The benefits of model guidance have typically been shown via improvements in classification performance (e.g. [49, 48]) or an increase in IoU between object masks and attribution maps (e.g. [23, 33]). In addition to these metrics, we also evaluate on the EPG metric [67], which has thus far only been used to evaluate the quality of the attribution methods themselves. We further show that it lends itself well to being used as a guidance loss, as it places only minor constraints on the model, and, in contrast to the IoU metric, it is fully differentiable.", + "bbox": [ + 496, + 316, + 892, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Guiding Models Using Attributions", + "text_level": 1, + "bbox": [ + 498, + 478, + 818, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide an overview of the model guidance approach that jointly optimizes for classification and localization (Sec. 3.1). Specifically, we describe the attribution methods (Sec. 3.2), metrics (Sec. 3.3), and localization loss functions (Sec. 3.4) that we evaluate in Sec. 5. In Sec. 3.5 we discuss our strategy to train for localization in the presence of multiple ground truth classes.", + "bbox": [ + 496, + 503, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notation: We consider a multi-label classification problem with $K$ classes with $X \\in \\mathbb{R}^{C \\times H \\times W}$ the input image and $y \\in \\{0,1\\}^K$ the one-hot encoding of the image labels. With $A_k \\in \\mathbb{R}^{H \\times W}$ we denote an attribution map for a class $k$ for $X$ using a classifier $f$ ; $A_k^+$ denotes the positive component of the attributions, $\\hat{A}_k = \\frac{A_k}{\\max(\\mathrm{abs}(A_k))}$ normalized attributions, and $\\hat{A}_k^+ = \\frac{A_k^+}{\\max(A_k^+)}$ normalized positive attributions. Finally, $M_k \\in \\{0,1\\}^{H \\times W}$ denotes the binary mask for class $k$ , which is given by the union of bounding boxes of all occurrences of class $k$ in $X$ .", + "bbox": [ + 496, + 609, + 890, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Guidance Procedure", + "text_level": 1, + "bbox": [ + 500, + 782, + 746, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following prior work (e.g. [49, 56, 24, 23]), the model is trained jointly for classification and localization (cf. Fig. 3):", + "bbox": [ + 498, + 806, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c l a s s}} + \\lambda_ {\\text {l o c}} \\mathcal {L} _ {\\text {l o c}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 849, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "I.e., the loss consists of a classification loss $(\\mathcal{L}_{\\mathrm{class}})$ , for which we use binary cross-entropy, and a localization loss", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1924", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$(\\mathcal{L}_{\\mathrm{loc}})$ , which we discuss in Sec. 3.4; here, the hyperparameter $\\lambda_{\\mathrm{loc}}$ controls the weight given to each of the objectives.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Attribution Methods", + "text_level": 1, + "bbox": [ + 76, + 133, + 272, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to prior work that typically use GradCAM [53] attributions, we perform an evaluation over a selection of popularly used differentiable $^2$ attribution methods which have been shown to localize well [46]: IxG [57], IntGrad [62], and GradCAM [53]. We further evaluate model-inherent explanations of the recently proposed B-cos models [6]. To ensure comparability across attribution methods [46], we evaluate all attribution methods at the input, various intermediate, and the final spatial layer.", + "bbox": [ + 75, + 157, + 468, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "IxG [57] computes the element-wise product $\\odot$ of the input and the gradients of the $k$ -th output w.r.t. the input, i.e. $X\\odot \\nabla_{X}f_{k}(X)$ . For piece-wise linear models such as DNNs with ReLU activations [38], this faithfully computes the linear contributions of a given input pixel to the model output.", + "bbox": [ + 75, + 294, + 468, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GradCAM [53] computes importance attributions as a ReLU-thresholded, gradient-weighted sum of activation maps. In detail, it is given by $\\mathrm{ReLU}(\\sum_{c}\\alpha_{c}^{k}\\odot U_{c})$ with $c$ denoting the channel dimension, and $\\alpha^k$ the average-pooled gradients of the output for class $k$ with respect to the activations $U$ of the last convolutional layer in the model.", + "bbox": [ + 75, + 369, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "IntGrad [62] takes an axiomatic approach and is formulated as the integral of gradients over a straight line path from a baseline input to the given input $X$ . Approximating this integral requires several gradient computations, making it computationally expensive for use in model guidance. To alleviate this, when optimizing with IntGrad, we use the recently proposed $\\mathcal{X}$ -DNN models [28] that allow for an exact computation of IntGrad in a single backward pass.", + "bbox": [ + 75, + 462, + 468, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "B-cos [6] attributions are generated using the inherently-interpretable B-cos networks, which promote alignment between the input $\\mathbf{x}$ and a dynamic weight matrix $\\mathbf{W}(\\mathbf{x})$ during optimization. In our experiments, we use the contribution maps given by the element-wise product of the dynamic weights with the input $(\\mathbf{W}_k^T (\\mathbf{x})\\odot \\mathbf{x})$ , which faithfully represent the contribution of each pixel to class $k$ . To be able to guide B-cos models, we developed a differentiable implementation of B-cos explanations, see supplement.", + "bbox": [ + 75, + 583, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 731, + 261, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate the models' performance on both our training objectives: classification and localization. For classification, we use the F1 score and mean average precision (mAP). We discuss the localization metrics below.", + "bbox": [ + 75, + 755, + 468, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intersection over Union (IoU) is a commonly used metric (cf. [23]) that computes the intersection between the ground truth annotation masks and the binarized attribution maps,", + "bbox": [ + 76, + 816, + 468, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "normalized by their union; for binarization, a threshold parameter needs to be chosen. In this work, the ground truth masks are taken to be the union of all bounding boxes of a class in the image and, following prior work [20], the threshold parameter is selected via a heldout set.", + "bbox": [ + 500, + 90, + 890, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Energy-based Pointing Game (EPG) [67] measures the concentration of attribution energy within the mask, i.e. the fraction of positive attributions inside the bounding boxes:", + "bbox": [ + 500, + 167, + 890, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {E P G} _ {k} = \\frac {\\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} M _ {k , h w} A _ {k , h w} ^ {+}}{\\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} A _ {k , h w} ^ {+}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 219, + 890, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to IoU, EPG more faithfully takes into account the relative importance given to each input region, since it does not binarize the attributions. Like IoU, the scores lie in [0, 1], with higher scores indicating better localization.", + "bbox": [ + 498, + 268, + 890, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Localization Losses", + "text_level": 1, + "bbox": [ + 500, + 337, + 684, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate the most commonly used localization losses $(\\mathcal{L}_{\\mathrm{loc}}$ in Eq. (1)) from prior work. We describe these losses as applied on attribution maps of an image for a single class $k$ , as well as the proposed EPG-derived Energy loss.", + "bbox": [ + 498, + 359, + 890, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$L_{1}$ loss ([24, 23], Eq. (3)) minimizes the $L_{1}$ distance between annotation masks and normalized positive attributions $\\hat{A}_k^+$ , guiding the model towards uniform attributions inside the mask and suppressing attributions outside of it.", + "bbox": [ + 500, + 421, + 890, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l o c}, k} = \\frac {1}{H \\times W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\| M _ {k, h w} - \\hat {A} _ {k, h w} ^ {+} \\| _ {1} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 487, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Per-pixel cross entropy (PPCE) loss ([56], Eq. (4)) applies a binary cross entropy loss between the mask and the normalized positive annotations $\\hat{A}_k^+$ , thus guiding the model to maximize the attributions inside the mask:", + "bbox": [ + 498, + 515, + 890, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l o c}, k} = - \\frac {1}{\\| M _ {k} \\| _ {1}} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} M _ {k, h w} \\log \\left(\\hat {A} _ {k, h w} ^ {+}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 580, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As PPCE does not constrain attributions outside the mask, there is no explicit pressure to avoid spurious features.", + "bbox": [ + 500, + 609, + 888, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RRR\\* loss ([49], Eq. (5)). [49] introduced the RRR loss to regularize the normalized input gradients $\\hat{A}_{k,hw}$ as", + "bbox": [ + 500, + 640, + 888, + 671 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l o c}, k} = \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left(1 - M _ {k, h w}\\right) \\hat {A} _ {k, h w} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 679, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extend it to our setting, we take $\\hat{A}_{k,hw}$ to be given by an arbitrary attribution method (e.g. IntGrad); we denote this generalized version by RRR*. In contrast to the PPCE loss, RRR* only regularizes attributions outside the ground truth masks. While it thus does not introduce a uniformity prior similar to the $L_{1}$ loss, it also does not explicitly promote high importance attributions inside the masks.", + "bbox": [ + 498, + 709, + 890, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Energy Loss. In addition to the losses described in prior work, we propose to also evaluate using the EPG score ([67], Eq. (2)) as a loss function for model guidance, as it is fully differentiable. In particular, we simply define it as", + "bbox": [ + 498, + 816, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l o c}, k} = - \\operatorname {E P G} _ {k}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 885, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2. Differentiability is necessary for optimizing attributions via gradient descent, so non-differentiable methods (e.g. [47, 43]) are not considered.", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1925", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7c1905b038578c97d9c819dc9c64c3fced972e9568c5b314f49e47545c77d741.jpg", + "image_caption": [ + "Fig. 4: Selecting models for evaluation. For each configuration, we evaluate every model at every checkpoint and measure its performance across various metrics (F1, EPG, IoU) on the validation set; i.e. every point in the left graph corresponds to one model (for B-cos models optimized via the Energy loss at the input layer). Instead of evaluating a single model on the test set, we evaluate all Pareto-dominant models, as indicated in the center and right plot." + ], + "image_footnote": [], + "bbox": [ + 71, + 89, + 467, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike existing localization losses that either (i) do not constrain attributions across the entire input $(\\mathsf{RRR}^*,\\mathsf{PPCE})$ or (ii) force the model to attribute uniformly within the mask even if it includes irrelevant background regions $(L_{1},$ PPCE), maximizing the EPG score jointly optimizes for higher attribution energy within the mask and lower attribution energy outside the mask. By not enforcing a uniformity prior, we find that the Energy loss is able to provide effective guidance while allowing the model to learn freely what to focus on within the bounding boxes (Sec. 5).", + "bbox": [ + 75, + 290, + 468, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Efficient Optimization", + "text_level": 1, + "bbox": [ + 76, + 454, + 284, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast to prior work [49, 56, 24, 23], we perform model guidance on a multi-label classification setting, and consequently there are multiple ground truth classes whose attribution localization could be optimized. Computing and optimizing for several attributions within an image would add a significant overhead to the computational cost of training (multiple backward passes). Hence, for efficiency, we sample one ground truth class $k$ per image at random for every batch and only optimize for localization of that class, i.e., $\\mathcal{L}_{\\mathrm{loc}} = \\mathcal{L}_{\\mathrm{loc},k}$ . We find that this still provides effective model guidance while keeping the training cost tractable.", + "bbox": [ + 75, + 478, + 468, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 661, + 269, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe our experimental setup and how we select the best models across metrics; for full details, see supplement. We evaluate across all possible choices for each category, and discuss our results in Sec. 5. Datasets: We evaluate on PASCAL VOC 2007 [16] and MS COCO 2014 [34] for multi-label image classification. In Sec. 5.5, to understand the effectiveness of model guidance in mitigating spurious correlations, we also evaluate on the synthetically constructed Waterbirds-100 dataset [51, 42], where landbirds are perfectly correlated with land backgrounds on the training and validation sets, but are equally likely to occur on land or water in the test set (similar for waterbirds and water). With this dataset, we evaluate model guidance for suppressing undesired features.", + "bbox": [ + 75, + 688, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attribution Methods and Architectures: As described in Sec. 3.2, we evaluate with IxG [57], IntGrad [62], B-cos [6, 7], and GradCAM [53] using models with a ResNet-50 [27] backbone. For IntGrad, we use an $\\mathcal{X}$ -DNN ResNet-50 [28] to reduce the computational cost, and a B-cos ResNet-50 for the B-cos attributions. To emphasize that the results generalize across different backbones, we further provide results for a B-cos ViT-S [14, 7] and a B-cos DenseNet-121 [29, 7]. We evaluate optimizing the attributions at different network layers, such as at the input image and the last convolutional layers' output3, as well as at multiple intermediate layers. Within the main paper, we highlight some of the most representative and insightful results, the full set of results can be found in the supplement. All models were pretrained on ImageNet [50], and model guidance was applied when fine-tuning the models on the target dataset.", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Localization Losses: As described in Sec. 3.4, we compare four localization losses in our evaluation: (i) Energy, (ii) $L_{1}$ [24, 23], (iii) PPCE [56], and (iv) RRR* (cf. Sec. 3.4, [49]).", + "bbox": [ + 496, + 333, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics: As discussed in Sec. 3.3, we evaluate both for classification and localization performance of the models. For classification, we report the F1 scores, similar results with mAP scores can be found in the supplement. For localization, we evaluate using the EPG and IoU scores.", + "bbox": [ + 496, + 380, + 890, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Selecting the best models: As we evaluate for two distinct objectives (classification + localization), it is not trivial to decide which models perform 'the best', e.g. a model that provides the best classification performance might provide significantly worse localization than a model that provides only slightly lower classification performance. Finding the right balance and deciding which of those models in fact constitutes the 'better' model depends on the preference of the end user. Hence, instead of selecting models based on a single metric, we select the set of Pareto-dominant models [40, 41, 5] across three metrics—F1, EPG, and IoU—for each training configuration, as defined by a combination of attribution method, layer, and loss. Specifically, as shown in Fig. 4, we train each configuration using three different choices of $\\lambda_{\\mathrm{loc}}$ , and select the set of Pareto-dominant models among all checkpoints (epochs and $\\lambda_{\\mathrm{loc}}$ ). This provides a more holistic view of the general trends on the effectiveness of model guidance for each configuration.", + "bbox": [ + 496, + 455, + 890, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 742, + 705, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we discuss our experimental findings. In particular, in Sec. 5.1, we first discuss the impact of the loss functions on the EPG and IoU scores of the models; in Sec. 5.2, we then analyze the impact of the models and attribution methods; further in Sec. 5.3, we show that guiding the models via their explanations can lead to improved classification accuracy. In Sec. 5.4, we present additional", + "bbox": [ + 496, + 768, + 890, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3As typically used in IxG (input) and GradCAM (final) respectively.", + "bbox": [ + 514, + 886, + 877, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1926", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/764f59edc7188ec75dcec414c7552a78107ec05f498b12eedbbbb5130475fa2c.jpg", + "image_caption": [ + "(a) PASCAL VOC results for EPG vs. F1." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 387, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4aca215a1dcdf93ee79856e08a888240bcaae524424608497686c3e34aca2b8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 90, + 627, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a7d99e6d0282dec595c3f093439d3f16fc412916ea77ca7651947db7d9e5d30b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 90, + 867, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/98957feb031bab97836e5557cf46b9bfda752eac90e54fe772656344f125f10d.jpg", + "image_caption": [ + "(b) MS COCO results for EPG vs. F1." + ], + "image_footnote": [], + "bbox": [ + 101, + 285, + 387, + 452 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bbcccf4741a0ea918ecf24ff859dc2edefc52dc5869a87b289281a21bb5d03be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 285, + 627, + 452 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/30a1afe1bf695ecd1a3cf233243c3111e88bc837e5321126b6c31ced676b4e29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 286, + 867, + 452 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf6798f67f3cdb993070819b8ad63e26b8f3bdcd2e1631b4917bbe88405df895.jpg", + "image_caption": [ + "Fig. 5: EPG vs. F1, for different datasets ((a): VOC; (b): COCO), losses (markers) and models (columns), optimized at different layers (rows); additionally, we show the performance of the baseline model before fine-tuning and demarcate regions that strictly dominate (are strictly dominated by) the baseline performance in green (grey). For each configuration, we show the Pareto fronts (cf. Fig. 4) across regularization strengths $\\lambda_{\\mathrm{loc}}$ and epochs (cf. Sec. 5 and Fig. 4). We find the Energy loss to give the best trade-off between EPG and F1." + ], + "image_footnote": [], + "bbox": [ + 101, + 546, + 387, + 654 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ed3e743196a6931b8b142e0773581defda3439b3d94332aa17af2b6c745f881.jpg", + "image_caption": [ + "Fig. 6: IoU vs. F1, for different losses (markers) and models (columns) for VOC; results for COCO are in the supplement. Additionally, we show the performance of the baseline model before fine-tuning and demarcate regions that strictly dominate (are strictly dominated by) the baseline model in green (grey). For each configuration, we show the Pareto fronts (Fig. 4) across regularization strengths $\\lambda_{\\mathrm{loc}}$ and all epochs; for details, see Secs. 4 and 5. Across all configurations, we find the $L_{1}$ loss to provide the largest gains in IoU at the lowest cost." + ], + "image_footnote": [], + "bbox": [ + 393, + 547, + 627, + 654 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/961ad1912b1b2924ee63112b372d0ad9667d12f307e7cec569605538a6cfaa0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 547, + 867, + 654 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9fb3b15b6d409434a3aaf2d1988a0266eaf4b85308a2b1f66504a6c022f1e917.jpg", + "image_caption": [ + "Fig. 7: EPG vs. F1 on VOC. We observe the same trends as in Fig. 5a for different backbone architectures, specifically a B-cos DenseNet-121 and a B-cos ViT-S. For IoU results, see supplement." + ], + "image_footnote": [], + "bbox": [ + 81, + 729, + 465, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "studies in which we evaluate and discuss the cost of model guidance approaches: in particular, we study model guidance with limited additional labels, with increasingly coarse", + "bbox": [ + 75, + 852, + 470, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bounding boxes, and at deep layers in the network. Finally, in Sec. 5.5, we show the utility of model guidance in improving accuracy in the presence of distribution shifts. For easier reference, we label our individual findings as R1-R9. Note. To draw conclusive insights and highlight general and reliable trends in the experiments, we compare the Pareto curves (see Fig. 4) of individual configurations. If the Pareto curve of a specific loss (e.g. Energy in Fig. 5) consistently Pareto-dominates the Pareto curves of all other losses, we can confidently conclude that for the combination of evaluated metrics (e.g. EPG vs. F1), this loss is the best choice.", + "bbox": [ + 496, + 732, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1927", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57dacd8004bef9b1b5471ea5f84371e0d5cd90b1fe77101f40e107e02c8858ad.jpg", + "image_caption": [ + "Fig. 8: Faster training by guiding at later layers. While input-level attributions tend to be more detailed (cf. Fig. 2), they are costlier to compute than attributions at later layers. However, we find that guidance at later layers (e.g. @Mid3) also significantly improves input-level attributions, yielding similar EPG results as input-level guidance (@Input) at up to twice the training speed; for IoU results, see supplement." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 310, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68f77076c0fc0a2854aceeb2cae1d34d01a892c4ae8764f80201ea7bb1a06057.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 88, + 496, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/462d3d36f2f0a59e351b2e1613c75efd7065699303ff68a1520a4d679bc1c385.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 681, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64bc2fd43df95645b1076b0b3e789287ee399fbfe3b8d4805c50ecfb31b00a18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 88, + 867, + 194 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Comparing loss functions for model guidance", + "text_level": 1, + "bbox": [ + 76, + 270, + 460, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the following, we highlight the main insights gained from the quantitative evaluations. For a qualitative comparison between the losses, please see Fig. 9; note that we show examples for a B-cos model as the differences become clearest; full results can be found in the supplement.", + "bbox": [ + 75, + 294, + 468, + 369 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "R1 The Energy loss yields the best EPG scores. In Fig. 5, we plot the Pareto curves for EPG vs. F1 scores for a wide range of configurations (see Sec. 4) on VOC (a) and COCO (b); specifically, we group the results by model type (Vanilla, $\\mathcal{X}$ -DNN, B-cos), the layer depths at which the attribution was regularized (Input / Final), and the loss used during optimization (Energy, $L_{1}$ , PPCE, RRR*). From these results it becomes apparent that the optimization with the Energy loss yields the best trade-off between accuracy (F1) and the EPG score: e.g., when looking at the upper right plot in Fig. 5a we can see that the Energy loss (red dots) improves over the baseline B-cos model (white cross) by improving the localization in terms of EPG score with only a minor cost in classification performance (i.e. F1 score). Further trading off F1 scores yields even higher EPG scores. Importantly, the Energy loss Pareto-dominates all the other losses (RRR*: blue diamonds; $L_{1}$ : green triangles; PPCE: yellow pentagons). This is is also true for the other network types (Vanilla ResNet-50, Fig. 5a (top left), and $\\mathcal{X}$ -DNN, Fig. 5a (top center)) and at the final layer (bottom row), and generalizes across backbone architectures (Fig. 7). When comparing Fig. 5a and Fig. 5b, we also find these results to be highly consistent between datasets.", + "R2 The $L_{1}$ loss yields the best IoU performance. Similarly, in Fig. 6, we plot the Pareto curves of IoU vs. F1 scores for various configurations at the final layer; for the IoU results at the input layer and on the COCO dataset, please see the supplement. For IoU, the $L_{1}$ loss provides the best trade-off and, with few exceptions, $L_{1}$ -guided models Pareto-dominate all other models in all configurations.", + "R3 The Energy loss focuses best on on-object features. By not forcing the models to highlight the entire bounding boxes (see Sec. 3.4), we find that the Energy loss also suppresses background features within the bounding boxes, thus better preserving fine details of the explanations" + ], + "bbox": [ + 75, + 369, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15644f355a0560a56d95c049e06dea1248bd183d6c7311aa4f89b6c5bd84c400.jpg", + "image_caption": [ + "Fig. 9: Loss comparison for input attributions (atts.) of a B-cos model. We showatts. before (baseline, col. 2) and after guidance (cols. 3-6) for a specific image (col. 1) and its bounding box annotation. We find that Energy and RRR* yield sparseatts, whereas $L_{1}$ yields smootheratts, as it is optimized to fill the entire bounding box. For PPCE we observe only a minor effect on theatts." + ], + "image_footnote": [], + "bbox": [ + 504, + 268, + 887, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(cf. Figs. 9 and 11). To quantify this, we evaluate the distribution of Energy (Eq. (2)) just within the bounding boxes. For this, we take advantage of the segmentation mask annotations available for a subset of the VOC test set. Specifically, we measure the Energy contained in the segmentation masks versus the entire bounding box, which indicates how much of the attributions actually highlight on-object features. We find that the Energy loss outperforms $L_{1}$ across all models and configurations; see supplement for details.", + "bbox": [ + 496, + 420, + 890, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In short, we find that the Energy loss works best for improving the EPG metric, whereas the $L_{1}$ loss yields the highest gains in terms of IoU; depending on the use case, either of these losses could thus be recommendable. However, we find that the Energy loss is more robust to annotation errors (R8, Sec. 5.4), and, as discussed in R3, the Energy loss more reliably focuses on object-specific features.", + "bbox": [ + 496, + 558, + 890, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparing models and attribution methods", + "text_level": 1, + "bbox": [ + 500, + 678, + 872, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the following, we highlight our findings regarding different attribution methods and models. Given the similarity of the results between GradCAM and IxG, and since B-cos attributions performed better than GradCAM for B-cos models, we show GradCAM results in the supplement.", + "bbox": [ + 496, + 702, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "R4 At the input layer, B-cos explanations perform best. We find that the B-cos models not only achieve the highest EPG/IoU performance before applying model guidance, ('baseslines') but also obtain the highest gains in EPG and IoU and thus the highest overall performance (for EPG see Fig. 5, right; for IoU, see supplement): e.g., an Energy-based B-cos model achieves an EPG score of $71.7@79.4\\%$ F1, thus significantly outperforming the best EPG scores of", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1928", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both other model types at a much lower cost in F1 (Vanilla: 55.8 @ $69.0\\%$ , $\\mathcal{X}$ -DNN: 62.3 @ $68.9\\%$ ). This is also observed qualitatively, as we show in the supplement.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "R5 Regularizing at the final layer yields consistent gains. As can be seen in Fig. 5 (bottom) and Fig. 6, all models can be guided well via regularization at the final layer, i.e. all models show improvements in IoU and EPG score.", + "bbox": [ + 75, + 137, + 467, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In short, we find model guidance to work well across all tested models when optimizing at the final layer (R5), highlighting its wide applicability. However, to obtain highly detailed and well-localized attributions at the input layer, the model-inherent explanations of the B-cos models seem to lend themselves much better to such guidance (R4).", + "bbox": [ + 75, + 196, + 467, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Improving accuracy with model guidance", + "text_level": 1, + "bbox": [ + 76, + 297, + 429, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "R6 Model guidance can improve accuracy. For both the Vanilla models (final layer) and the $\\mathcal{X}$ -DNNs (input+final), we found models that improve the localization metrics and the F1 score. These improvements are particularly pronounced for the $\\mathcal{X}$ -DNN: e.g., we find models that improve the EPG and F1 scores by $\\Delta = 7.2$ p.p. and $\\Delta = 1.4$ p.p. respectively (Fig. 5, center top), or the IoU and F1 scores by $\\Delta = 11.9$ p.p. and $\\Delta = 1.4$ p.p. (Fig. 6, center).", + "bbox": [ + 75, + 321, + 467, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, overall we observe a trade-off between localization and accuracy (Figs. 5 and 6). Given the similarity of the training and test distributions, focusing on the object need not improve classification performance, as spurious features are also present at test time. Further, the guided model is discouraged from relying on contextual features, making the classification more challenging. In Sec. 5.5, we show that guidance can significantly improve performance when there is a distribution shift between training and test.", + "bbox": [ + 75, + 443, + 467, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Efficiency and robustness considerations", + "text_level": 1, + "bbox": [ + 76, + 589, + 421, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While bounding boxes decrease the data collection cost with respect to segmentation masks, they can nonetheless be expensive to obtain, especially when expert knowledge is required. To further reduce those costs, in this section, we assess the robustness of guiding the model with a limited number (R7) or increasingly coarse annotations (R8). Apart from data efficiency, we further explore how training efficiency can be improved for fine-grained (i.e. input-level) explanations (R9), as explanations at early layers are more costly to obtain than those at later layers.", + "bbox": [ + 75, + 613, + 467, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "R7 Model guidance requires only few add. annotations. In Fig. 12, we show that the EPG score can be significantly improved with a very limited number of annotations; for IoU results, see supplement. Specifically, we find that when using only $1\\%$ of the training data (25 annotated images) for VOC, improvements of up to $\\Delta = 23.0$ p.p. ( $\\Delta = 1.4$ ) in EPG (IoU) can be obtained, at a minor drop in F1 ( $\\Delta = 0.3$ p.p. and $\\Delta = 2.5$ p.p. respectively). When annotating up to $10\\%$ of the images, very similar results can be achieved as", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with full annotation (see e.g. cols. 2+3 in Fig. 12).", + "bbox": [ + 498, + 90, + 831, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "R8 The Energy loss is highly robust to annotation errors. As discussed in Sec. 3.4, the Energy loss only directs the model on which features not to use and does not impose a uniform prior on the attributions within the bounding boxes. As a result, we find it to be much more stable to annotation errors: e.g., in Fig. 10, we visualize how the EPG (top) and IoU (bottom) scores of the best performing models under the Energy (left) and $L_{1}$ loss (right) evolve when using coarser bounding boxes; for this, we simply dilate the bounding box size by $p \\in \\{10,25,50\\} \\%$ during training, see Fig. 11. While the models optimized via the $L_{1}$ loss achieve increasingly worse results (right), the Energy-optimized models are essentially unaffected by the coarseness of the annotations.", + "bbox": [ + 496, + 107, + 890, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f385551e04751c4ed6c0b068e6b01fe2ff7332ef175fa5fbb71be2568634703c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 328, + 687, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3eb12969c3771dbff17478fb55232e4d736c41504d1530c037e666481a91a7cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 328, + 880, + 406 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/acedbb47fc5a14fd412f4776d2b3ad4129b40abe204f5564486250c8cf229612.jpg", + "image_caption": [ + "Fig. 10: Quantitative results for dilated bounding boxes for a B-cos model at the input layer. We show EPG and IoU (top and bottom) results for models trained with various amounts of annotation errors (increasingly large bounding boxes, see Fig. 11). The Energy loss yields highly consistent results despite training with heavily dilated bounding boxes (left), whereas the results of the $L_{1}$ loss (right) worsen markedly; best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 491, + 415, + 687, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b57e7c7fc7253a8c736dc3f1bbaba077b59a9dae3ec20d35a382f47dd59edc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 415, + 879, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76750dff31b25685457afbe6b1c59b05ebd692428814b23b9cece6fc7073e0dd.jpg", + "image_caption": [ + "Fig. 11: Qualitative results for dilated bounding boxes for a B-cos model at input. Examples for attributions (rows $2 + 3$ ) of models trained with dilated bounding boxes (row 1). In contrast to $L_{1}$ , models trained with Energy show significant gains in object focus even with significant noise (e.g. 'Baseline' vs. $50\\%$ )." + ], + "image_footnote": [], + "bbox": [ + 506, + 633, + 696, + 773 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53d85424fbc3d08bf14d5d6ac9eb76b301d08444ebe4265b9c93b3fabd20ac34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 633, + 887, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In short, we find that the models can be guided effectively at a low cost in terms of annotation effort, as only few", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1929", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "annotations (e.g. 25 for VOC) are required (cf. R7), and, especially for the Energy loss, these annotations can be very coarse and do not have to be 'pixel-perfect' (cf. R8).", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "R9 Guidance at deep layers can be effective. While guided input-level explanations of B-cos networks exhibit a high degree of detail, regularizing those explanations comes at an added training cost. In particular, optimizing at the input layer requires backpropagating through the entire network to compute the attributions. In an effort to reduce training costs whilst maintaining the benefits of fine-grained explanations at input resolution, we evaluate if input-level attributions benefit from an optimization at deeper layers.", + "bbox": [ + 75, + 136, + 468, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Specifically, we regularize B-cos attributions at the final and at three intermediate layers (Mid{1,2,3}), and evaluate the localization of attributions at the input. We find (Fig. 8) that training at a deeper layer can provide significant speed-ups in training time with often a negligible cost in localization performance. E.g., since we do not have to compute a full backward pass through the entire model during training, optimizing at Mid2 (col. 2 in Fig. 8) provides similar gains in localization but with a $1.7\\mathrm{x}$ speed-up in training time.", + "bbox": [ + 75, + 273, + 468, + 409 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/958e439fecad0c533c1d88ff82b917fa6451906acf38fd1f5d3d7b3ee3521e6d.jpg", + "image_caption": [ + "Fig. 12: EPG results with limited annotations for a B-cos model at the input layer, optimized with the Energy and the $L_{1}$ loss. Using bounding box annotations for as little as $1\\%$ (left) of the images yields significant improvements in EPG, and with $10\\%$ (center) similar gains as in the fully annotated setting (right) are obtained." + ], + "image_footnote": [], + "bbox": [ + 68, + 422, + 464, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.5. Effectiveness against spurious correlations", + "text_level": 1, + "bbox": [ + 76, + 619, + 439, + 636 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To evaluate the potential for mitigating spurious correlations, we evaluate model guidance with the Energy and $L_{1}$ losses on the synthetically constructed Waterbirds-100 dataset [51, 42]. We perform model guidance under two settings: (1) the conventional setting to classify between landbirds and waterbirds, using the region within the bounding box as the mask; and (2) the reversed setting [42] to classify the background, i.e., land vs. water, using the region outside the bounding box as the mask. To simulate a limited annotation budget, we only use bounding boxes for a random $1\\%$ of the training set, and report results averaged over four runs. We show the results for the worst-group accuracy (i.e., images containing a waterbird on land) and the overall accuracy using B-cos models in Tab. 1; full results for all attributions and models can be found in the supplement.", + "bbox": [ + 75, + 643, + 468, + 869 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Both losses consistently and significantly improve the accuracy in the conventional and the reversed settings by", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c1ca9fe9ad24bf88749f50181dd215a50a76af682ee79a0dc622f5ded838f8ba.jpg", + "image_caption": [ + "Fig. 13: Qualitative Waterbirds-100 results. Without guidance, a model might focus on the background to classify birds (baseline) and thus misclassify waterbirds on land (col. 2). Guided models can correct such errors and focus on the desired feature: in cols. $3 + 4$ $(5 + 6)$ the model is guided to classify by using the bird (background) features and arrives at the desired prediction. Model predictions and confidence scores are indicated below the images." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 172 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/97fa313dbd7cca7938b99692246a08597eebb687b39ae2eacdd420aabd5a6282.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConventionalReversed
ModelWorstOverallWorstOverall
Baseline43.4 (±2.4)68.7 (±0.2)56.6 (±2.4)80.1 (±0.2)
Energy56.1 (±4.0)71.2 (±0.1)62.8 (±2.1)83.6 (±1.1)
L151.1 (±1.9)69.5 (±0.2)58.8 (±5.0)82.2 (±0.9)
", + "bbox": [ + 504, + 277, + 887, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 1: Waterbirds-100 results. We find that model guidance is effective in improving both worst-group ('Waterbird on Land') and overall accuracy in the conventional (Landbird vs. Waterbird) and reversed (Land vs. Water) settings; full results in the supplement.", + "bbox": [ + 496, + 358, + 890, + 414 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "guiding the model to select the 'right' features, i.e. birds (conventional) or background (reversed). This guidance can also be observed qualitatively (cf. Fig. 13).", + "bbox": [ + 496, + 422, + 890, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Discussion And Conclusion", + "text_level": 1, + "bbox": [ + 500, + 500, + 751, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we comprehensively evaluated various models, attribution methods, and loss functions for their utility in guiding models to be \"right for the right reasons\".", + "bbox": [ + 496, + 532, + 890, + 578 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In summary, we find that guiding models via bounding boxes can significantly improve EPG and IoU performance of the optimized attribution method, with the Energy loss working best to improve the EPG score (R1) and the $L_{1}$ loss yielding the highest gains in IoU scores (R2). While the B-cos models achieve the best results in IoU and EPG score at the input layer (R4), all tested model types (Vanilla, $\\mathcal{X}$ -DNN, B-cos) lend themselves well to being optimized at the final layer (R5), which can even improve attribution maps at early layers (R9). Further, we find that regularizing the explanations of the models and thereby 'telling them where to look' can increase the object recognition performance (mAP/accuracy) of some models (R6), especially when strong spurious correlations are present (Sec. 5.5). Interestingly, those gains (EPG, IoU), can be achieved with relatively little additional annotation (R7). Lastly, we find that by not assuming a uniform prior over the attributions within the annotated bounding boxes, training with the energy loss is more robust to annotation errors (R8) and results in models that produce attribution maps that are more focused on class-specific features (R3).", + "bbox": [ + 496, + 583, + 892, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1930", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity Checks for Saliency Maps. In NeurIPS, 2018. 3", + "[2] Julius Adebayo, Michael Muelly, Harold Abelson, and Been Kim. Post hoc Explanations may be Ineffective for Detecting Unknown Spurious Correlation. In ICLR, 2022. 3", + "[3] Saeid Asgari, Aliasghar Khani, Fereshte Khani, Ali Gholami, Linh Tran, Ali Mahdavi-Amiri, and Ghassan Hamarneh. MaskTune: Mitigating Spurious Correlations by Forcing to Explore. In NeurIPS, 2022. 3", + "[4] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation. PLOS One, 10(7):e0130140, 2015. 3", + "[5] Jürgen Backhaus. The Pareto Principle. Analyse & Kritik, 2(2):146-171, 1980. 5", + "[6] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-Cos Networks: Alignment is All We Need for Interpretability. In CVPR, pages 10329-10338, 2022. 1, 2, 3, 4, 5", + "[7] Moritz Böhle, Navdeepal Singh, Mario Fritz, and Bernt Schiele. B-cos Alignment for Inherently Interpretable CNNs and Vision Transformers. arXiv preprint arXiv:2306.10898, 2023.5", + "[8] Moritz Böhle, Mario Fritz, and Bernt Schiele. Convolutional Dynamic Alignment Networks for Interpretable Classifications. In CVPR, pages 10029-10038, 2021. 3", + "[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks. In WACV, pages 839-847, 2018. 3", + "[10] Hila Chefer, Shir Gur, and Lior Wolf. Transformer Interpretability Beyond Attention Visualization. In CVPR, pages 782-791, 2021. 3", + "[11] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing Relevance Maps of Vision Transformers Improves Robustness. In NeurIPS, 2022. 3", + "[12] Piotr Dabkowski and Yarin Gal. Real Time Image Saliency for Black Box Classifiers. In NeurIPS, 2017. 3", + "[13] Saurabh Desai and Harish G. Ramaswamy. Ablation-CAM: Visual Explanations for Deep Convolutional Network via Gradient-free Localization. In WACV, pages 983-991, 2020. 3", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5", + "[15] Gabriel Erion, Joseph D Janizek, Pascal Sturmfels, Scott M Lundberg, and Su-In Lee. Improving Performance of Deep Learning Models with Axiomatic Attribution Priors and Expected Gradients. Nature Machine Intelligence, 3(7):620-631, 2021. 3", + "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. IJCV, 88:303-308, 2009. 2, 5", + "[17] Thomas Fel, Ivan F Rodriguez Rodriguez, Drew Linsley, and" + ], + "bbox": [ + 80, + 114, + 468, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Thomas Serre. Harmonizing the Object Recognition Strategies of Deep Neural Networks with Humans. In NeurIPS, 2022. 3", + "[18] Vitaly Feldman and Chiyuan Zhang. What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation. In NeurIPS, pages 2881-2891, 2020. 1", + "[19] Patrick Fernandes, Marcos Treviso, Danish Pruthi, André FT Martins, and Graham Neubig. Learning to Scaffold: Optimizing Model Explanations for Teaching. In NeurIPS, 2022. 3", + "[20] Ruth C Fong and Andrea Vedaldi. Interpretable Explanations of Black Boxes by Meaningful Perturbation. In ICCV, pages 3429-3437, 2017. 3, 4", + "[21] Felix Friedrich, Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. A Typology to Explore and Guide Explanatory Interactive Machine Learning. arXiv preprint arXiv:2203.03668, 2022. 3", + "[22] Hiroshi Fukui, Tsubasa Hirakawa, Takayoshi Yamashita, and Hironobu Fujiyoshi. Attention Branch Network: Learning of Attention Mechanism for Visual Explanation. In CVPR, pages 10705-10714, 2019. 3", + "[23] Yuyang Gao, Tong Steven Sun, Guangji Bai, Siyi Gu, Sungsoo Ray Hong, and Zhao Liang. RES: A Robust Framework for Guiding Visual Explanation. In KDD, pages 432-442, 2022. 2, 3, 4, 5", + "[24] Yuyang Gao, Tong Steven Sun, Liang Zhao, and Sungsoo Ray Hong. Aligning Eyes between Humans and Deep Neural Network through Interactive Attention Alignment. ACM HCI, 6(CSCW2):1-28, 2022. 2, 3, 4, 5", + "[25] Hao Guo, Kang Zheng, Xiaochuan Fan, Hongkai Yu, and Song Wang. Visual Attention Consistency under Image Transforms for Multi-Label Image Classification. In CVPR, pages 729–739, 2019. 3", + "[26] Misgina Tsighe Hagos, Kathleen M Curran, and Brian Mac Namee. Identifying Spurious Correlations and Correcting them with an Explanation-based Learning. arXiv preprint arXiv:2211.08285, 2022. 3", + "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, pages 770-778, 2016. 5", + "[28] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Fast Axiomatic Attribution for Neural Networks. In NeurIPS, pages 19513-19524, 2021. 4, 5", + "[29] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely Connected Convolutional Networks. In CVPR, pages 4700-4708, 2017. 5", + "[30] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. LayerCAM: Exploring Hierarchical Class Activation Maps for Localization. IEEE TIP, 30:5875-5888, 2021. 3", + "[31] Joon Sik Kim, Gregory Plumb, and Ameet Talwalkar. Sanity Simulations for Saliency Methods. In ICML, 2022. 3", + "[32] Keisuke Kiritoshi, Ryosuke Tanno, and Tomonori Izumitani. L1-Norm Gradient Penalty for Noise Reduction of Attribution Maps. In CVPRW, pages 118-121, 2019. 3", + "[33] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell Me Where to Look: Guided Attention Inference Network. In CVPR, pages 9215-9223, 2018. 3", + "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence" + ], + "bbox": [ + 503, + 93, + 890, + 886 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1931", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, pages 740-755. Springer, 2014. 2, 5", + "[35] Drew Linsley, Dan Shiebler, Sven Eberhardt, and Thomas Serre. Learning what and where to attend. In ICLR, 2019. 3", + "[36] Masahiro Mitsuhara, Hiroshi Fukui, Yusuke Sakashita, Takanori Ogata, Tsubasa Hirakawa, Takayoshi Yamashita, and Hironobu Fujiyoshi. Embedding Human Knowledge into Deep Neural Network via Attention Map. In VISIGRAPP, pages 626-636, 2021. 3", + "[37] Ofir Moshe, Gil Fidel, Ron Bitton, and Asaf Shabtai. Improving Interpretability via Regularization of Neural Activation Sensitivity. arXiv preprint arXiv:2211.08686, 2022. 3", + "[38] Vinod Nair and Geoffrey E Hinton. Rectified Linear Units Improve Restricted Boltzmann Machines. In ICML, pages 807-814, 2010. 4", + "[39] Krishna Kanth Nakka and Mathieu Salzmann. Towards Robust Fine-Grained Recognition by Maximal Separation of Discriminative Features. In ACCV, 2020. 3", + "[40] Vilfredo Pareto. Il massimo di utilità dato alla libera concorrenza. Giornale degli economisti, pages 48-66, 1894. 5", + "[41] Vilfredo Pareto. The Maximum of Utility given by Free Competition. Giornale degli Economisti e Annali di Economia, 67(3):387-403, 2008. 5", + "[42] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On Guiding Visual Attention with Language Specification. In CVPR, pages 18092-18102, 2022. 1, 2, 3, 5, 9", + "[43] Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized Input Sampling for Explanation of Black-box Models. In BMVC, 2018. 3, 4", + "[44] Vipin Pillai, Soroush Abbasi Koohpayegani, Ashley Ouligian, Dennis Fong, and Hamed Piriavash. Consistent Explanations by Contrastive Learning. In CVPR, pages 10213-10222, 2022. 3", + "[45] Vipin Pillai and Hamed Piri siavash. Explainable Models with Consistent Interpretations. In AAAI, 2021. 3", + "[46] Sukrut Rao, Moritz Böhle, and Bernt Schiele. Towards Better Understanding Attribution Methods. In CVPR, pages 10223-10232, 2022. 3, 4", + "[47] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"Why Should I Trust You?: Explaining the Predictions of Any Classifier. In KDD, pages 1135-1144, 2016. 3, 4", + "[48] Laura Rieger, Chandan Singh, William Murdoch, and Bin Yu. Interpretations are Useful: Penalizing Explanations to Align Neural Networks with Prior Knowledge. In ICML, pages 8116-8126, 2020. 3", + "[49] Andrew Slavin Ross, Michael C Hughes, and Finale Doshi-Velez. Right for the Right Reasons: Training Differentiable Models by Constraining their Explanations. In *IJCAI*, pages 2662–2670, 2017. 1, 2, 3, 4, 5", + "[50] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211–252, 2015. 5", + "[51] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization. In ICLR, 2020. 2, 5, 9", + "[52] Patrick Schramowski, Wolfgang Stammer, Stefano Teso," + ], + "bbox": [ + 78, + 90, + 467, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Anna Brugger, Franziska Herbert, Xiaoting Shao, Hans-Georg Luigs, Anne-Katrin Mahlein, and Kristian Kersting. Making Deep Neural Networks Right for the Right Scientific Reasons by Interacting with their Explanations. Nature Machine Intelligence, 2(8):476-486, 2020. 3", + "[53] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. In ICCV, pages 618-626, 2017. 1, 2, 3, 4, 5", + "[54] Ramprasaath R Selvaraju, Stefan Lee, Yilin Shen, Hongxia Jin, Shalini Ghosh, Larry Heck, Dhruv Batra, and Devi Parikh. Taking a HINT: Leveraging Explanations to Make Vision and Language Models More Grounded. In ICCV, pages 2591-2600, 2019. 3", + "[55] Xiaoting Shao, Arseny Skryagin, Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. Right for Better Reasons: Training Differentiable Models by Constraining their Influence Functions. In AAAI, volume 35, pages 9533-9540, 2021. 3", + "[56] Haifeng Shen, Kewen Liao, Zhibin Liao, Job Doornberg, Maoying Qiao, Anton Van Den Hengel, and Johan W Verjans. Human-AI Interactive and Continuous Sensemaking: A Case Study of Image Classification using Scribble Attention Maps. In Extended Abstracts of CHI, pages 1-8, 2021. 2, 3, 4, 5", + "[57] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning Important Features Through Propagating Activation Differences. In ICML, pages 3145-3153, 2017. 1, 2, 3, 4, 5", + "[58] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps. In ICLRw, 2014. 3", + "[59] Krishna Kumar Singh, Dhruv Mahajan, Kristen Grauman, Yong Jae Lee, Matt Feiszli, and Deepti Ghadiyaram. Don't Judge an Object by its Context: Learning to Overcome Contextual Bias. In CVPR, pages 11070-11078, 2020. 3", + "[60] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for Simplicity: The All Convolutional Net. In ICLRw, 2015. 3", + "[61] Guolei Sun, Salman Khan, Wen Li, Hisham Cholakkal, Fahad Shahbaz Khan, and Luc Van Gool. Fixing Localization Errors to Improve Image Classification. In ECCV, pages 271-287. Springer, 2020. 3", + "[62] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic Attribution for Deep Networks. In ICML, pages 3319-3328, 2017. 1, 2, 3, 4, 5", + "[63] Damien Teney, Ehsan Abbasnedjad, and Anton van den Hengel. Learning What Makes a Difference from Counterfactual Examples and Gradient Supervision. In ECCV, pages 580-599. Springer, 2020. 3", + "[64] Stefano Teso. Toward Faithful Explanatory Active Learning with Self-Explainable Neural Nets. In Workshop on IAL, pages 4–16. CEUR Workshop Proceedings, 2019. 2, 3", + "[65] Stefano Teso, Öznur Alkan, Wolfgang Stammer, and Elizabeth Daly. Leveraging Explanations in Interactive Machine Learning: An Overview. Frontiers in Artificial Intelligence, 6:1066049, 2023. 3", + "[66] Stefano Teso and Kristian Kersting. Explanatory Interactive Machine Learning. In AIES, pages 239-245, 2019. 2, 3", + "[67] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian" + ], + "bbox": [ + 503, + 90, + 890, + 890 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1932", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks. In CVPRW, pages 111-119, 2020. 2, 3, 4", + "[68] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or Signal: The Role of Image Backgrounds in Object Recognition. In *ICLR*, 2021. 1", + "[69] Ziyan Yang, Kushal Kafle, Franck Dernoncourt, and Vicente Ordonez. Improving Visual Grounding by Encouraging Consistent Gradient-based Explanations. In CVPR, pages 19165–19174, 2023. 3", + "[70] Matthew D Zeiler and Rob Fergus. Visualizing and Understanding Convolutional Networks. In ECCV, pages 818-833, 2014. 3", + "[71] Michael Zhang, Nimit Sharad Sohoni, Hongyang R. Zhang, Chelsea Finn, and Christopher Ré. Correct-N-Contrast: a Contrastive Approach for Improving Robustness to Spurious Correlations. In ICML, pages 26484–26516, 2022. 3", + "[72] Yilun Zhou, Serena Booth, Marco Tulio Ribeiro, and Julie Shah. Do Feature Attribution Methods Correctly Attribute Features? In AAAI, volume 36, pages 9623-9633, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 363 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1933", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/88a53c18-99d0-4b4c-b0b2-16a2556a29ae_model.json b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/88a53c18-99d0-4b4c-b0b2-16a2556a29ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e680663e3a908965563578218be3b5119efe06f6 --- /dev/null +++ b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/88a53c18-99d0-4b4c-b0b2-16a2556a29ae_model.json @@ -0,0 +1,2820 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.872, + 0.154 + ], + "angle": 0, + "content": "Studying How to Efficiently and Effectively Guide Models with Explanations" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.182, + 0.846, + 0.217 + ], + "angle": 0, + "content": "Sukrut Rao*, Moritz Böhle*, Amin Parchami-Araghi, Bernt Schiele \nMax Planck Institute for Informatics, Saarland Informatics Campus, Saarbrücken, Germany" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.219, + 0.714, + 0.234 + ], + "angle": 0, + "content": "{sukrut.rao,mboehle,mparcham,schiele}@mpi-inf.mpg.de" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.755 + ], + "angle": 0, + "content": "Despite being highly performant, deep neural networks might base their decisions on features that spuriously correlate with the provided labels, thus hurting generalization. To mitigate this, 'model guidance' has recently gained popularity, i.e. the idea of regularizing the models' explanations to ensure that they are \"right for the right reasons\" [49]. While various techniques to achieve such model guidance have been proposed, experimental validation of these approaches has thus far been limited to relatively simple and / or synthetic datasets. To better understand the effectiveness of the various design choices that have been explored in the context of model guidance, in this work we conduct an in-depth evaluation across various loss functions, attribution methods, models, and 'guidance depths' on the PASCAL VOC 2007 and MS COCO 2014 datasets. As annotation costs for model guidance can limit its applicability, we also place a particular focus on efficiency. Specifically, we guide the models via bounding box annotations, which are much cheaper to obtain than the commonly used segmentation masks, and evaluate the robustness of model guidance under limited (e.g. with only \\(1\\%\\) of annotated images) or overly coarse annotations. Further, we propose using the EPG score as an additional evaluation metric and loss function ('Energy loss'). We show that optimizing for the Energy loss leads to models that exhibit a distinct focus on object-specific features, despite only using bounding box annotations that also include background regions. Lastly, we show that such model guidance can improve generalization under distribution shifts. Code available at: https://github.com/sukrutrao/Model-Guidance" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.775, + 0.21, + 0.79 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.801, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) excel at learning predictive features that allow them to correctly classify a set of training images with ease. The features learnt on the training set, however, do not necessarily transfer to unseen images: i.e., instead of learning the actual class-relevant fea" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.27, + 0.87, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.427, + 0.576, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.429, + 0.65, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.419, + 0.728, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.419, + 0.808, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.492, + 0.806, + 0.505 + ], + "angle": 0, + "content": "Landbird Confidence: \\(97\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.419, + 0.886, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.492, + 0.885, + 0.505 + ], + "angle": 0, + "content": "Waterbird Confidence: \\(87\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Fig. 1: (a) Model guidance increases object focus. Models may rely on irrelevant background features or spurious correlations (e.g. presence of person provides positive evidence for bicycle, center row, col. 1). Guiding the model via bounding box annotations can mitigate this and consistently increases the focus on object features (bottom row). (b) Model guidance can improve accuracy. In the presence of spurious correlations in the training data, non-guided models might focus on the wrong features. In the example image in (b), the waterbird is incorrectly classified to be a landbird due to the background (col. 3). Guiding the model via bounding box annotation (as shown in col. 2), the model can be guided to focus on the bird features for classification (col. 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.893, + 0.82 + ], + "angle": 0, + "content": "tures, DNNs might memorize individual images (cf. [18]) or exploit spurious correlations in the training data (cf. [68]). For example, if bikes are highly correlated with people in the training data, a model might learn to associate the presence of a person in an image as positive evidence for a bike (e.g. Fig. 1a, col. 1, rows 1-2), which can limit how well it generalizes. Similarly, a bird classifier might rely on background features from the bird's habitat, and fail to correctly classify in a different habitat (cf. Fig. 1b cols. 1-3 and [42])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To detect such behaviour, recent advances in model interpretability have provided attribution methods (e.g. [53, 62, 57, 6]) to understand a model's reasoning. These methods typically provide attention maps that highlight regions of importance in an input to explain the model's decisions" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1922" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.088, + 0.872, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.309, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Fig. 2: Qualitative results of model guidance. We show model-inherent B-cos explanations (input layer) of a B-cos ResNet-50 and GradCAM explanations (final layer) of a conventional ResNet-50 before ('Standard') and after optimization ('Guided') for images from the VOC test set, using our proposed Energy loss (Eq. (6)). Guiding the model via bounding box annotations consistently increases the focus on object features for both methods. Specifically, we find that background attributions are consistently suppressed in both cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.469, + 0.412 + ], + "angle": 0, + "content": "and can help identify incorrect reasoning such as reliance on spurious or irrelevant features, see for example Fig. 1b." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.469, + 0.578 + ], + "angle": 0, + "content": "As many attribution methods are in fact themselves differentiable (e.g. [57, 62, 53, 6]), recent work [49, 56, 24, 23, 66, 64] has explored the idea of using them to guide the models to make them \"right for the right reasons\" [49]. Specifically, models can be guided by jointly optimizing for correct classification as well as for attributing importance to regions deemed relevant by humans. This can help the model focus on the relevant features of a class, and correct errors in reasoning (Fig. 1b, col. 4). Such guidance has the added benefit of providing well-localized explanations that are thus easier to understand for end users (e.g. Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.469, + 0.684 + ], + "angle": 0, + "content": "While model guidance has shown promising results, a detailed study of how to do this most effectively is crucially missing. In particular, model guidance has so far been studied for a limited set of attribution methods and models and usually on relatively simple and/or synthetic datasets; further, the evaluation settings between approaches can significantly differ, which makes a fair comparison difficult." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.469, + 0.835 + ], + "angle": 0, + "content": "Therefore, in this work, we perform an in-depth evaluation of model guidance on large scale, real-world datasets, to better understand the effectiveness of a variety of design choices. Specifically, we evaluate model guidance along the following dimensions: the model architecture, the guidance depth1, the attribution method, and the loss function. In this context, we propose using the EPG score [67]—an evaluation metric that has thus far been used to evaluate the quality of attribution methods—as an additional loss function (which we call the Energy loss) as it is fully differentiable." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.836, + 0.469, + 0.866 + ], + "angle": 0, + "content": "Further, as annotation costs can be a major hurdle for making model guidance practical, we place a particular fo" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.382, + 0.892, + 0.442 + ], + "angle": 0, + "content": "cus on efficient guidance. Specifically, we use bounding boxes instead of semantic segmentation masks, and evaluate the robustness of guidance techniques under limited or overly coarse annotations to reduce data collection costs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.893, + 0.597 + ], + "angle": 0, + "content": "We find that our Energy loss lends itself well to those settings. On the one hand, it exhibits a high degree of robustness to limited or noisy bounding box annotations (cf. Figs. 10 and 12). On the other hand, despite the coarseness of bounding box guidance, it maintains a clear focus on object-specific features inside the bounding boxes, see Fig. 1a, row 3. In contrast, prior approaches often regularize for a uniform distribution of the attribution values inside the annotation masks, and thus tend to exhibit much lower attribution granularity (cf. Fig. 9)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Contributions. (1) We perform an in-depth evaluation of model guidance on challenging large scale, multi-label classification datasets (PASCAL VOC 2007 [16], MS COCO 2014 [34]), assessing the impact of attribution methods, model architectures, guidance depths, and loss functions. Further, we show that, despite being relatively coarse, bounding box supervision can provide sufficient guidance to the models whilst being much cheaper to obtain than semantic segmentation masks. (2) We propose using the Energy Pointing Game (EPG) score [67] as an alternative to the IoU metric for evaluating the effectiveness of such guidance and show that the EPG score constitutes a good loss function for model guidance, particularly when using bounding boxes. (3) We show that model guidance can be performed cost-effectively by using annotation masks that are noisy or are available for only a small fraction (e.g. \\(1\\%\\)) of the training data. (4) We show through experiments on the Waterbirds-100 dataset [51, 42] that model guidance with a small number of annotations suffices to improve the model's generalization under distribution shifts at test time." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "1The layer at which guidance is applied, e.g. typically at the last convolutional layer for GradCAM [53] or the first layer for IxG [57]." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1923" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.265 + ], + "angle": 0, + "content": "Attribution Methods [58, 60, 62, 57, 53, 67, 13, 30, 9, 43, 20, 70, 47, 12, 4] are often used to explain black-box models by generating heatmaps that highlight input regions important to the model's decision. However, such methods are often not faithful to the model [1, 46, 31, 72, 2] and risk misleading users. Recent work proposes inherently interpretable models [8, 6] that address this by providing modelfaithful explanations by design. In our work, we use both popular post-hoc and model-inherent attribution methods to guide models and discuss their effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.471, + 0.417 + ], + "angle": 0, + "content": "Attribution Priors: Several approaches have been proposed for training better models by enforcing desirable properties on their attributions. These include enforcing consistency against augmentations [45, 44, 25], smoothness [15, 37, 32], separation of classes [71, 44, 61, 39, 59], or constraining the model's attention [22, 3]. In contrast, in this work, we focus on providing explicit human guidance to the model using bounding box annotations. This constitutes more explicit guidance but allows fine-grained control over the model's reasoning even with few annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.471, + 0.705 + ], + "angle": 0, + "content": "Model Guidance: In contrast to the indirect regularization effect achieved by attribution priors, various approaches have been proposed (cf. [21, 65]) to actively guide models by regularizing their attributions, for tasks such as classification [49, 24, 23, 48, 42, 26, 63, 36, 66, 64, 52, 55, 35, 56, 69, 17], segmentation [33], VQA [54, 63], and knowledge distillation [19]. The goal of such approaches is not only to improve performance, but also make sure that the model is \"right for the right reasons\" [49]. For classifiers, this typically involves jointly optimizing both for classification performance and localization to object features. While various benefits of model guidance have been reported, most prior work evaluate on simple datasets [49, 55, 24, 23] and, thus far, no common evaluation setting has emerged. Recently, [11] has extended model guidance to ImageNet, showing that its benefits can scale to large scale problems. In contrast to [11], who investigated one particular attribution method [10], our focus lies on a better understanding of the impact of the different design choices for model guidance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To distill the most effective techniques for model guidance, in this work, we conduct an in-depth evaluation on challenging, commonly used real-world multi-label classification datasets (PASCAL VOC 2007, MS COCO 2014). Specifically, we perform a comprehensive comparison across multiple dimensions of interest: the loss function, the model architecture, the guidance depth, and the attribution method. For this, we evaluate the localization losses introduced in the closest related work, i.e. RRR [49], HAICS [56], and GRADIA [24]; additionally, we propose using the EPG metric [67] as a loss function and show that it has various desirable properties, in particular when guiding models via bounding box annotations." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.088, + 0.895, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Fig. 3: Model guidance overview. We jointly optimize for classification \\((\\mathcal{L}_{\\mathrm{class}})\\) and localization of attributions to human-annotated bounding boxes \\((\\mathcal{L}_{\\mathrm{loc}})\\), to guide the model to focus on object features. Various localization loss functions can be used, see Sec. 3.4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Finally, model guidance has also been used to mitigate reliance on spurious features using language guidance [42], and we show that using a small number of coarse bounding box annotations can be similarly effective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.893, + 0.468 + ], + "angle": 0, + "content": "Evaluating Model Guidance: The benefits of model guidance have typically been shown via improvements in classification performance (e.g. [49, 48]) or an increase in IoU between object masks and attribution maps (e.g. [23, 33]). In addition to these metrics, we also evaluate on the EPG metric [67], which has thus far only been used to evaluate the quality of the attribution methods themselves. We further show that it lends itself well to being used as a guidance loss, as it places only minor constraints on the model, and, in contrast to the IoU metric, it is fully differentiable." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.479, + 0.82, + 0.496 + ], + "angle": 0, + "content": "3. Guiding Models Using Attributions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.61 + ], + "angle": 0, + "content": "In this section, we provide an overview of the model guidance approach that jointly optimizes for classification and localization (Sec. 3.1). Specifically, we describe the attribution methods (Sec. 3.2), metrics (Sec. 3.3), and localization loss functions (Sec. 3.4) that we evaluate in Sec. 5. In Sec. 3.5 we discuss our strategy to train for localization in the presence of multiple ground truth classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.776 + ], + "angle": 0, + "content": "Notation: We consider a multi-label classification problem with \\(K\\) classes with \\(X \\in \\mathbb{R}^{C \\times H \\times W}\\) the input image and \\(y \\in \\{0,1\\}^K\\) the one-hot encoding of the image labels. With \\(A_k \\in \\mathbb{R}^{H \\times W}\\) we denote an attribution map for a class \\(k\\) for \\(X\\) using a classifier \\(f\\); \\(A_k^+\\) denotes the positive component of the attributions, \\(\\hat{A}_k = \\frac{A_k}{\\max(\\mathrm{abs}(A_k))}\\) normalized attributions, and \\(\\hat{A}_k^+ = \\frac{A_k^+}{\\max(A_k^+)}\\) normalized positive attributions. Finally, \\(M_k \\in \\{0,1\\}^{H \\times W}\\) denotes the binary mask for class \\(k\\), which is given by the union of bounding boxes of all occurrences of class \\(k\\) in \\(X\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.748, + 0.798 + ], + "angle": 0, + "content": "3.1. Model Guidance Procedure" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.807, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Following prior work (e.g. [49, 56, 24, 23]), the model is trained jointly for classification and localization (cf. Fig. 3):" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.851, + 0.891, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c l a s s}} + \\lambda_ {\\text {l o c}} \\mathcal {L} _ {\\text {l o c}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "I.e., the loss consists of a classification loss \\((\\mathcal{L}_{\\mathrm{class}})\\), for which we use binary cross-entropy, and a localization loss" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1924" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "\\((\\mathcal{L}_{\\mathrm{loc}})\\), which we discuss in Sec. 3.4; here, the hyperparameter \\(\\lambda_{\\mathrm{loc}}\\) controls the weight given to each of the objectives." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.134, + 0.273, + 0.149 + ], + "angle": 0, + "content": "3.2. Attribution Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.158, + 0.469, + 0.293 + ], + "angle": 0, + "content": "In contrast to prior work that typically use GradCAM [53] attributions, we perform an evaluation over a selection of popularly used differentiable\\(^2\\) attribution methods which have been shown to localize well [46]: IxG [57], IntGrad [62], and GradCAM [53]. We further evaluate model-inherent explanations of the recently proposed B-cos models [6]. To ensure comparability across attribution methods [46], we evaluate all attribution methods at the input, various intermediate, and the final spatial layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.469, + 0.369 + ], + "angle": 0, + "content": "IxG [57] computes the element-wise product \\(\\odot\\) of the input and the gradients of the \\(k\\)-th output w.r.t. the input, i.e. \\(X\\odot \\nabla_{X}f_{k}(X)\\). For piece-wise linear models such as DNNs with ReLU activations [38], this faithfully computes the linear contributions of a given input pixel to the model output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.469, + 0.461 + ], + "angle": 0, + "content": "GradCAM [53] computes importance attributions as a ReLU-thresholded, gradient-weighted sum of activation maps. In detail, it is given by \\(\\mathrm{ReLU}(\\sum_{c}\\alpha_{c}^{k}\\odot U_{c})\\) with \\(c\\) denoting the channel dimension, and \\(\\alpha^k\\) the average-pooled gradients of the output for class \\(k\\) with respect to the activations \\(U\\) of the last convolutional layer in the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.583 + ], + "angle": 0, + "content": "IntGrad [62] takes an axiomatic approach and is formulated as the integral of gradients over a straight line path from a baseline input to the given input \\( X \\). Approximating this integral requires several gradient computations, making it computationally expensive for use in model guidance. To alleviate this, when optimizing with IntGrad, we use the recently proposed \\( \\mathcal{X} \\)-DNN models [28] that allow for an exact computation of IntGrad in a single backward pass." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.469, + 0.72 + ], + "angle": 0, + "content": "B-cos [6] attributions are generated using the inherently-interpretable B-cos networks, which promote alignment between the input \\(\\mathbf{x}\\) and a dynamic weight matrix \\(\\mathbf{W}(\\mathbf{x})\\) during optimization. In our experiments, we use the contribution maps given by the element-wise product of the dynamic weights with the input \\((\\mathbf{W}_k^T (\\mathbf{x})\\odot \\mathbf{x})\\), which faithfully represent the contribution of each pixel to class \\(k\\). To be able to guide B-cos models, we developed a differentiable implementation of B-cos explanations, see supplement." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.732, + 0.262, + 0.747 + ], + "angle": 0, + "content": "3.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.469, + 0.816 + ], + "angle": 0, + "content": "We evaluate the models' performance on both our training objectives: classification and localization. For classification, we use the F1 score and mean average precision (mAP). We discuss the localization metrics below." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.818, + 0.469, + 0.863 + ], + "angle": 0, + "content": "Intersection over Union (IoU) is a commonly used metric (cf. [23]) that computes the intersection between the ground truth annotation masks and the binarized attribution maps," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "normalized by their union; for binarization, a threshold parameter needs to be chosen. In this work, the ground truth masks are taken to be the union of all bounding boxes of a class in the image and, following prior work [20], the threshold parameter is selected via a heldout set." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.168, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Energy-based Pointing Game (EPG) [67] measures the concentration of attribution energy within the mask, i.e. the fraction of positive attributions inside the bounding boxes:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.22, + 0.891, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\mathrm {E P G} _ {k} = \\frac {\\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} M _ {k , h w} A _ {k , h w} ^ {+}}{\\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} A _ {k , h w} ^ {+}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.269, + 0.892, + 0.33 + ], + "angle": 0, + "content": "In contrast to IoU, EPG more faithfully takes into account the relative importance given to each input region, since it does not binarize the attributions. Like IoU, the scores lie in [0, 1], with higher scores indicating better localization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.338, + 0.686, + 0.352 + ], + "angle": 0, + "content": "3.4. Localization Losses" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.361, + 0.892, + 0.421 + ], + "angle": 0, + "content": "We evaluate the most commonly used localization losses \\((\\mathcal{L}_{\\mathrm{loc}}\\) in Eq. (1)) from prior work. We describe these losses as applied on attribution maps of an image for a single class \\(k\\), as well as the proposed EPG-derived Energy loss." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.422, + 0.892, + 0.481 + ], + "angle": 0, + "content": "\\(L_{1}\\) loss ([24, 23], Eq. (3)) minimizes the \\(L_{1}\\) distance between annotation masks and normalized positive attributions \\(\\hat{A}_k^+\\), guiding the model towards uniform attributions inside the mask and suppressing attributions outside of it." + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.488, + 0.891, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {l o c}, k} = \\frac {1}{H \\times W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\| M _ {k, h w} - \\hat {A} _ {k, h w} ^ {+} \\| _ {1} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.516, + 0.892, + 0.575 + ], + "angle": 0, + "content": "Per-pixel cross entropy (PPCE) loss ([56], Eq. (4)) applies a binary cross entropy loss between the mask and the normalized positive annotations \\(\\hat{A}_k^+\\), thus guiding the model to maximize the attributions inside the mask:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.582, + 0.891, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l o c}, k} = - \\frac {1}{\\| M _ {k} \\| _ {1}} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} M _ {k, h w} \\log \\left(\\hat {A} _ {k, h w} ^ {+}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.611, + 0.89, + 0.64 + ], + "angle": 0, + "content": "As PPCE does not constrain attributions outside the mask, there is no explicit pressure to avoid spurious features." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.641, + 0.89, + 0.672 + ], + "angle": 0, + "content": "RRR\\* loss ([49], Eq. (5)). [49] introduced the RRR loss to regularize the normalized input gradients \\(\\hat{A}_{k,hw}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.68, + 0.891, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {l o c}, k} = \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left(1 - M _ {k, h w}\\right) \\hat {A} _ {k, h w} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.71, + 0.892, + 0.815 + ], + "angle": 0, + "content": "To extend it to our setting, we take \\(\\hat{A}_{k,hw}\\) to be given by an arbitrary attribution method (e.g. IntGrad); we denote this generalized version by RRR*. In contrast to the PPCE loss, RRR* only regularizes attributions outside the ground truth masks. While it thus does not introduce a uniformity prior similar to the \\(L_{1}\\) loss, it also does not explicitly promote high importance attributions inside the masks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.817, + 0.892, + 0.876 + ], + "angle": 0, + "content": "Energy Loss. In addition to the losses described in prior work, we propose to also evaluate using the EPG score ([67], Eq. (2)) as a loss function for model guidance, as it is fully differentiable. In particular, we simply define it as" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l o c}, k} = - \\operatorname {E P G} _ {k}. \\tag {6}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "2. Differentiability is necessary for optimizing attributions via gradient descent, so non-differentiable methods (e.g. [47, 43]) are not considered." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "1925" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.072, + 0.09, + 0.468, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.187, + 0.47, + 0.284 + ], + "angle": 0, + "content": "Fig. 4: Selecting models for evaluation. For each configuration, we evaluate every model at every checkpoint and measure its performance across various metrics (F1, EPG, IoU) on the validation set; i.e. every point in the left graph corresponds to one model (for B-cos models optimized via the Energy loss at the input layer). Instead of evaluating a single model on the test set, we evaluate all Pareto-dominant models, as indicated in the center and right plot." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.47, + 0.442 + ], + "angle": 0, + "content": "Unlike existing localization losses that either (i) do not constrain attributions across the entire input \\((\\mathsf{RRR}^*,\\mathsf{PPCE})\\) or (ii) force the model to attribute uniformly within the mask even if it includes irrelevant background regions \\((L_{1},\\) PPCE), maximizing the EPG score jointly optimizes for higher attribution energy within the mask and lower attribution energy outside the mask. By not enforcing a uniformity prior, we find that the Energy loss is able to provide effective guidance while allowing the model to learn freely what to focus on within the bounding boxes (Sec. 5)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.285, + 0.47 + ], + "angle": 0, + "content": "3.5. Efficient Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.646 + ], + "angle": 0, + "content": "In contrast to prior work [49, 56, 24, 23], we perform model guidance on a multi-label classification setting, and consequently there are multiple ground truth classes whose attribution localization could be optimized. Computing and optimizing for several attributions within an image would add a significant overhead to the computational cost of training (multiple backward passes). Hence, for efficiency, we sample one ground truth class \\( k \\) per image at random for every batch and only optimize for localization of that class, i.e., \\( \\mathcal{L}_{\\mathrm{loc}} = \\mathcal{L}_{\\mathrm{loc},k} \\). We find that this still provides effective model guidance while keeping the training cost tractable." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.662, + 0.27, + 0.68 + ], + "angle": 0, + "content": "4. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this section, we describe our experimental setup and how we select the best models across metrics; for full details, see supplement. We evaluate across all possible choices for each category, and discuss our results in Sec. 5. Datasets: We evaluate on PASCAL VOC 2007 [16] and MS COCO 2014 [34] for multi-label image classification. In Sec. 5.5, to understand the effectiveness of model guidance in mitigating spurious correlations, we also evaluate on the synthetically constructed Waterbirds-100 dataset [51, 42], where landbirds are perfectly correlated with land backgrounds on the training and validation sets, but are equally likely to occur on land or water in the test set (similar for waterbirds and water). With this dataset, we evaluate model guidance for suppressing undesired features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Attribution Methods and Architectures: As described in Sec. 3.2, we evaluate with IxG [57], IntGrad [62], B-cos [6, 7], and GradCAM [53] using models with a ResNet-50 [27] backbone. For IntGrad, we use an \\(\\mathcal{X}\\)-DNN ResNet-50 [28] to reduce the computational cost, and a B-cos ResNet-50 for the B-cos attributions. To emphasize that the results generalize across different backbones, we further provide results for a B-cos ViT-S [14, 7] and a B-cos DenseNet-121 [29, 7]. We evaluate optimizing the attributions at different network layers, such as at the input image and the last convolutional layers' output3, as well as at multiple intermediate layers. Within the main paper, we highlight some of the most representative and insightful results, the full set of results can be found in the supplement. All models were pretrained on ImageNet [50], and model guidance was applied when fine-tuning the models on the target dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Localization Losses: As described in Sec. 3.4, we compare four localization losses in our evaluation: (i) Energy, (ii) \\(L_{1}\\) [24, 23], (iii) PPCE [56], and (iv) RRR* (cf. Sec. 3.4, [49])." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.455 + ], + "angle": 0, + "content": "Evaluation Metrics: As discussed in Sec. 3.3, we evaluate both for classification and localization performance of the models. For classification, we report the F1 scores, similar results with mAP scores can be found in the supplement. For localization, we evaluate using the EPG and IoU scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Selecting the best models: As we evaluate for two distinct objectives (classification + localization), it is not trivial to decide which models perform 'the best', e.g. a model that provides the best classification performance might provide significantly worse localization than a model that provides only slightly lower classification performance. Finding the right balance and deciding which of those models in fact constitutes the 'better' model depends on the preference of the end user. Hence, instead of selecting models based on a single metric, we select the set of Pareto-dominant models [40, 41, 5] across three metrics—F1, EPG, and IoU—for each training configuration, as defined by a combination of attribution method, layer, and loss. Specifically, as shown in Fig. 4, we train each configuration using three different choices of \\(\\lambda_{\\mathrm{loc}}\\), and select the set of Pareto-dominant models among all checkpoints (epochs and \\(\\lambda_{\\mathrm{loc}}\\)). This provides a more holistic view of the general trends on the effectiveness of model guidance for each configuration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.743, + 0.706, + 0.76 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.892, + 0.876 + ], + "angle": 0, + "content": "In this section, we discuss our experimental findings. In particular, in Sec. 5.1, we first discuss the impact of the loss functions on the EPG and IoU scores of the models; in Sec. 5.2, we then analyze the impact of the models and attribution methods; further in Sec. 5.3, we show that guiding the models via their explanations can lead to improved classification accuracy. In Sec. 5.4, we present additional" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.879, + 0.901 + ], + "angle": 0, + "content": "3As typically used in IxG (input) and GradCAM (final) respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1926" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.388, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.091, + 0.628, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.092, + 0.868, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.268, + 0.602, + 0.279 + ], + "angle": 0, + "content": "(a) PASCAL VOC results for EPG vs. F1." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.286, + 0.388, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.286, + 0.628, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.287, + 0.868, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.463, + 0.591, + 0.474 + ], + "angle": 0, + "content": "(b) MS COCO results for EPG vs. F1." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.481, + 0.893, + 0.539 + ], + "angle": 0, + "content": "Fig. 5: EPG vs. F1, for different datasets ((a): VOC; (b): COCO), losses (markers) and models (columns), optimized at different layers (rows); additionally, we show the performance of the baseline model before fine-tuning and demarcate regions that strictly dominate (are strictly dominated by) the baseline performance in green (grey). For each configuration, we show the Pareto fronts (cf. Fig. 4) across regularization strengths \\(\\lambda_{\\mathrm{loc}}\\) and epochs (cf. Sec. 5 and Fig. 4). We find the Energy loss to give the best trade-off between EPG and F1." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.547, + 0.388, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.548, + 0.628, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.548, + 0.868, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.66, + 0.893, + 0.716 + ], + "angle": 0, + "content": "Fig. 6: IoU vs. F1, for different losses (markers) and models (columns) for VOC; results for COCO are in the supplement. Additionally, we show the performance of the baseline model before fine-tuning and demarcate regions that strictly dominate (are strictly dominated by) the baseline model in green (grey). For each configuration, we show the Pareto fronts (Fig. 4) across regularization strengths \\(\\lambda_{\\mathrm{loc}}\\) and all epochs; for details, see Secs. 4 and 5. Across all configurations, we find the \\(L_{1}\\) loss to provide the largest gains in IoU at the lowest cost." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.731, + 0.466, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.803, + 0.471, + 0.847 + ], + "angle": 0, + "content": "Fig. 7: EPG vs. F1 on VOC. We observe the same trends as in Fig. 5a for different backbone architectures, specifically a B-cos DenseNet-121 and a B-cos ViT-S. For IoU results, see supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.853, + 0.471, + 0.899 + ], + "angle": 0, + "content": "studies in which we evaluate and discuss the cost of model guidance approaches: in particular, we study model guidance with limited additional labels, with increasingly coarse" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.901 + ], + "angle": 0, + "content": "bounding boxes, and at deep layers in the network. Finally, in Sec. 5.5, we show the utility of model guidance in improving accuracy in the presence of distribution shifts. For easier reference, we label our individual findings as R1-R9. Note. To draw conclusive insights and highlight general and reliable trends in the experiments, we compare the Pareto curves (see Fig. 4) of individual configurations. If the Pareto curve of a specific loss (e.g. Energy in Fig. 5) consistently Pareto-dominates the Pareto curves of all other losses, we can confidently conclude that for the combination of evaluated metrics (e.g. EPG vs. F1), this loss is the best choice." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1927" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.312, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.089, + 0.498, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.683, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.089, + 0.868, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.2, + 0.894, + 0.242 + ], + "angle": 0, + "content": "Fig. 8: Faster training by guiding at later layers. While input-level attributions tend to be more detailed (cf. Fig. 2), they are costlier to compute than attributions at later layers. However, we find that guidance at later layers (e.g. @Mid3) also significantly improves input-level attributions, yielding similar EPG results as input-level guidance (@Input) at up to twice the training speed; for IoU results, see supplement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.271, + 0.461, + 0.287 + ], + "angle": 0, + "content": "5.1. Comparing loss functions for model guidance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.47, + 0.37 + ], + "angle": 0, + "content": "In the following, we highlight the main insights gained from the quantitative evaluations. For a qualitative comparison between the losses, please see Fig. 9; note that we show examples for a B-cos model as the differences become clearest; full results can be found in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.717 + ], + "angle": 0, + "content": "R1 The Energy loss yields the best EPG scores. In Fig. 5, we plot the Pareto curves for EPG vs. F1 scores for a wide range of configurations (see Sec. 4) on VOC (a) and COCO (b); specifically, we group the results by model type (Vanilla, \\(\\mathcal{X}\\)-DNN, B-cos), the layer depths at which the attribution was regularized (Input / Final), and the loss used during optimization (Energy, \\(L_{1}\\), PPCE, RRR*). From these results it becomes apparent that the optimization with the Energy loss yields the best trade-off between accuracy (F1) and the EPG score: e.g., when looking at the upper right plot in Fig. 5a we can see that the Energy loss (red dots) improves over the baseline B-cos model (white cross) by improving the localization in terms of EPG score with only a minor cost in classification performance (i.e. F1 score). Further trading off F1 scores yields even higher EPG scores. Importantly, the Energy loss Pareto-dominates all the other losses (RRR*: blue diamonds; \\(L_{1}\\): green triangles; PPCE: yellow pentagons). This is is also true for the other network types (Vanilla ResNet-50, Fig. 5a (top left), and \\(\\mathcal{X}\\)-DNN, Fig. 5a (top center)) and at the final layer (bottom row), and generalizes across backbone architectures (Fig. 7). When comparing Fig. 5a and Fig. 5b, we also find these results to be highly consistent between datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.824 + ], + "angle": 0, + "content": "R2 The \\( L_{1} \\) loss yields the best IoU performance. Similarly, in Fig. 6, we plot the Pareto curves of IoU vs. F1 scores for various configurations at the final layer; for the IoU results at the input layer and on the COCO dataset, please see the supplement. For IoU, the \\( L_{1} \\) loss provides the best trade-off and, with few exceptions, \\( L_{1} \\)-guided models Pareto-dominate all other models in all configurations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "R3 The Energy loss focuses best on on-object features. By not forcing the models to highlight the entire bounding boxes (see Sec. 3.4), we find that the Energy loss also suppresses background features within the bounding boxes, thus better preserving fine details of the explanations" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.27, + 0.888, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.33, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Fig. 9: Loss comparison for input attributions (atts.) of a B-cos model. We showatts. before (baseline, col. 2) and after guidance (cols. 3-6) for a specific image (col. 1) and its bounding box annotation. We find that Energy and RRR* yield sparseatts, whereas \\( L_{1} \\) yields smootheratts, as it is optimized to fill the entire bounding box. For PPCE we observe only a minor effect on theatts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.558 + ], + "angle": 0, + "content": "(cf. Figs. 9 and 11). To quantify this, we evaluate the distribution of Energy (Eq. (2)) just within the bounding boxes. For this, we take advantage of the segmentation mask annotations available for a subset of the VOC test set. Specifically, we measure the Energy contained in the segmentation masks versus the entire bounding box, which indicates how much of the attributions actually highlight on-object features. We find that the Energy loss outperforms \\( L_{1} \\) across all models and configurations; see supplement for details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.559, + 0.892, + 0.665 + ], + "angle": 0, + "content": "In short, we find that the Energy loss works best for improving the EPG metric, whereas the \\(L_{1}\\) loss yields the highest gains in terms of IoU; depending on the use case, either of these losses could thus be recommendable. However, we find that the Energy loss is more robust to annotation errors (R8, Sec. 5.4), and, as discussed in R3, the Energy loss more reliably focuses on object-specific features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.873, + 0.695 + ], + "angle": 0, + "content": "5.2. Comparing models and attribution methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.779 + ], + "angle": 0, + "content": "In the following, we highlight our findings regarding different attribution methods and models. Given the similarity of the results between GradCAM and IxG, and since B-cos attributions performed better than GradCAM for B-cos models, we show GradCAM results in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "R4 At the input layer, B-cos explanations perform best. We find that the B-cos models not only achieve the highest EPG/IoU performance before applying model guidance, ('baseslines') but also obtain the highest gains in EPG and IoU and thus the highest overall performance (for EPG see Fig. 5, right; for IoU, see supplement): e.g., an Energy-based B-cos model achieves an EPG score of \\(71.7@79.4\\%\\) F1, thus significantly outperforming the best EPG scores of" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1928" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "both other model types at a much lower cost in F1 (Vanilla: 55.8 @ \\(69.0\\%\\), \\(\\mathcal{X}\\)-DNN: 62.3 @ \\(68.9\\%\\)). This is also observed qualitatively, as we show in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.197 + ], + "angle": 0, + "content": "R5 Regularizing at the final layer yields consistent gains. As can be seen in Fig. 5 (bottom) and Fig. 6, all models can be guided well via regularization at the final layer, i.e. all models show improvements in IoU and EPG score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.468, + 0.289 + ], + "angle": 0, + "content": "In short, we find model guidance to work well across all tested models when optimizing at the final layer (R5), highlighting its wide applicability. However, to obtain highly detailed and well-localized attributions at the input layer, the model-inherent explanations of the B-cos models seem to lend themselves much better to such guidance (R4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.299, + 0.431, + 0.315 + ], + "angle": 0, + "content": "5.3. Improving accuracy with model guidance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.468, + 0.444 + ], + "angle": 0, + "content": "R6 Model guidance can improve accuracy. For both the Vanilla models (final layer) and the \\(\\mathcal{X}\\)-DNNs (input+final), we found models that improve the localization metrics and the F1 score. These improvements are particularly pronounced for the \\(\\mathcal{X}\\)-DNN: e.g., we find models that improve the EPG and F1 scores by \\(\\Delta = 7.2\\) p.p. and \\(\\Delta = 1.4\\) p.p. respectively (Fig. 5, center top), or the IoU and F1 scores by \\(\\Delta = 11.9\\) p.p. and \\(\\Delta = 1.4\\) p.p. (Fig. 6, center)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.468, + 0.58 + ], + "angle": 0, + "content": "However, overall we observe a trade-off between localization and accuracy (Figs. 5 and 6). Given the similarity of the training and test distributions, focusing on the object need not improve classification performance, as spurious features are also present at test time. Further, the guided model is discouraged from relying on contextual features, making the classification more challenging. In Sec. 5.5, we show that guidance can significantly improve performance when there is a distribution shift between training and test." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.423, + 0.606 + ], + "angle": 0, + "content": "5.4. Efficiency and robustness considerations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.764 + ], + "angle": 0, + "content": "While bounding boxes decrease the data collection cost with respect to segmentation masks, they can nonetheless be expensive to obtain, especially when expert knowledge is required. To further reduce those costs, in this section, we assess the robustness of guiding the model with a limited number (R7) or increasingly coarse annotations (R8). Apart from data efficiency, we further explore how training efficiency can be improved for fine-grained (i.e. input-level) explanations (R9), as explanations at early layers are more costly to obtain than those at later layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "R7 Model guidance requires only few add. annotations. In Fig. 12, we show that the EPG score can be significantly improved with a very limited number of annotations; for IoU results, see supplement. Specifically, we find that when using only \\(1\\%\\) of the training data (25 annotated images) for VOC, improvements of up to \\(\\Delta = 23.0\\) p.p. (\\(\\Delta = 1.4\\)) in EPG (IoU) can be obtained, at a minor drop in F1 (\\(\\Delta = 0.3\\) p.p. and \\(\\Delta = 2.5\\) p.p. respectively). When annotating up to \\(10\\%\\) of the images, very similar results can be achieved as" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.832, + 0.107 + ], + "angle": 0, + "content": "with full annotation (see e.g. cols. 2+3 in Fig. 12)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.892, + 0.317 + ], + "angle": 0, + "content": "R8 The Energy loss is highly robust to annotation errors. As discussed in Sec. 3.4, the Energy loss only directs the model on which features not to use and does not impose a uniform prior on the attributions within the bounding boxes. As a result, we find it to be much more stable to annotation errors: e.g., in Fig. 10, we visualize how the EPG (top) and IoU (bottom) scores of the best performing models under the Energy (left) and \\( L_{1} \\) loss (right) evolve when using coarser bounding boxes; for this, we simply dilate the bounding box size by \\( p \\in \\{10,25,50\\} \\% \\) during training, see Fig. 11. While the models optimized via the \\( L_{1} \\) loss achieve increasingly worse results (right), the Energy-optimized models are essentially unaffected by the coarseness of the annotations." + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.329, + 0.688, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.329, + 0.882, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.416, + 0.688, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.416, + 0.88, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.6 + ], + "angle": 0, + "content": "Fig. 10: Quantitative results for dilated bounding boxes for a B-cos model at the input layer. We show EPG and IoU (top and bottom) results for models trained with various amounts of annotation errors (increasingly large bounding boxes, see Fig. 11). The Energy loss yields highly consistent results despite training with heavily dilated bounding boxes (left), whereas the results of the \\( L_{1} \\) loss (right) worsen markedly; best viewed in color." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.634, + 0.697, + 0.774 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.634, + 0.888, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.849 + ], + "angle": 0, + "content": "Fig. 11: Qualitative results for dilated bounding boxes for a B-cos model at input. Examples for attributions (rows \\(2 + 3\\)) of models trained with dilated bounding boxes (row 1). In contrast to \\(L_{1}\\), models trained with Energy show significant gains in object focus even with significant noise (e.g. 'Baseline' vs. \\(50\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "In short, we find that the models can be guided effectively at a low cost in terms of annotation effort, as only few" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1929" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "annotations (e.g. 25 for VOC) are required (cf. R7), and, especially for the Energy loss, these annotations can be very coarse and do not have to be 'pixel-perfect' (cf. R8)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.469, + 0.273 + ], + "angle": 0, + "content": "R9 Guidance at deep layers can be effective. While guided input-level explanations of B-cos networks exhibit a high degree of detail, regularizing those explanations comes at an added training cost. In particular, optimizing at the input layer requires backpropagating through the entire network to compute the attributions. In an effort to reduce training costs whilst maintaining the benefits of fine-grained explanations at input resolution, we evaluate if input-level attributions benefit from an optimization at deeper layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.47, + 0.41 + ], + "angle": 0, + "content": "Specifically, we regularize B-cos attributions at the final and at three intermediate layers (Mid{1,2,3}), and evaluate the localization of attributions at the input. We find (Fig. 8) that training at a deeper layer can provide significant speed-ups in training time with often a negligible cost in localization performance. E.g., since we do not have to compute a full backward pass through the entire model during training, optimizing at Mid2 (col. 2 in Fig. 8) provides similar gains in localization but with a \\(1.7\\mathrm{x}\\) speed-up in training time." + }, + { + "type": "image", + "bbox": [ + 0.069, + 0.423, + 0.465, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.592 + ], + "angle": 0, + "content": "Fig. 12: EPG results with limited annotations for a B-cos model at the input layer, optimized with the Energy and the \\(L_{1}\\) loss. Using bounding box annotations for as little as \\(1\\%\\) (left) of the images yields significant improvements in EPG, and with \\(10\\%\\) (center) similar gains as in the fully annotated setting (right) are obtained." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.44, + 0.637 + ], + "angle": 0, + "content": "5.5. Effectiveness against spurious correlations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.87 + ], + "angle": 0, + "content": "To evaluate the potential for mitigating spurious correlations, we evaluate model guidance with the Energy and \\( L_{1} \\) losses on the synthetically constructed Waterbirds-100 dataset [51, 42]. We perform model guidance under two settings: (1) the conventional setting to classify between landbirds and waterbirds, using the region within the bounding box as the mask; and (2) the reversed setting [42] to classify the background, i.e., land vs. water, using the region outside the bounding box as the mask. To simulate a limited annotation budget, we only use bounding boxes for a random \\( 1\\% \\) of the training set, and report results averaged over four runs. We show the results for the worst-group accuracy (i.e., images containing a waterbird on land) and the overall accuracy using B-cos models in Tab. 1; full results for all attributions and models can be found in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Both losses consistently and significantly improve the accuracy in the conventional and the reversed settings by" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.177, + 0.892, + 0.275 + ], + "angle": 0, + "content": "Fig. 13: Qualitative Waterbirds-100 results. Without guidance, a model might focus on the background to classify birds (baseline) and thus misclassify waterbirds on land (col. 2). Guided models can correct such errors and focus on the desired feature: in cols. \\(3 + 4\\) \\((5 + 6)\\) the model is guided to classify by using the bird (background) features and arrives at the desired prediction. Model predictions and confidence scores are indicated below the images." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.279, + 0.888, + 0.357 + ], + "angle": 0, + "content": "
ConventionalReversed
ModelWorstOverallWorstOverall
Baseline43.4 (±2.4)68.7 (±0.2)56.6 (±2.4)80.1 (±0.2)
Energy56.1 (±4.0)71.2 (±0.1)62.8 (±2.1)83.6 (±1.1)
L151.1 (±1.9)69.5 (±0.2)58.8 (±5.0)82.2 (±0.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.415 + ], + "angle": 0, + "content": "Table 1: Waterbirds-100 results. We find that model guidance is effective in improving both worst-group ('Waterbird on Land') and overall accuracy in the conventional (Landbird vs. Waterbird) and reversed (Land vs. Water) settings; full results in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.892, + 0.469 + ], + "angle": 0, + "content": "guiding the model to select the 'right' features, i.e. birds (conventional) or background (reversed). This guidance can also be observed qualitatively (cf. Fig. 13)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.501, + 0.753, + 0.517 + ], + "angle": 0, + "content": "6. Discussion And Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.579 + ], + "angle": 0, + "content": "In this work, we comprehensively evaluated various models, attribution methods, and loss functions for their utility in guiding models to be \"right for the right reasons\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In summary, we find that guiding models via bounding boxes can significantly improve EPG and IoU performance of the optimized attribution method, with the Energy loss working best to improve the EPG score (R1) and the \\(L_{1}\\) loss yielding the highest gains in IoU scores (R2). While the B-cos models achieve the best results in IoU and EPG score at the input layer (R4), all tested model types (Vanilla, \\(\\mathcal{X}\\)-DNN, B-cos) lend themselves well to being optimized at the final layer (R5), which can even improve attribution maps at early layers (R9). Further, we find that regularizing the explanations of the models and thereby 'telling them where to look' can increase the object recognition performance (mAP/accuracy) of some models (R6), especially when strong spurious correlations are present (Sec. 5.5). Interestingly, those gains (EPG, IoU), can be achieved with relatively little additional annotation (R7). Lastly, we find that by not assuming a uniform prior over the attributions within the annotated bounding boxes, training with the energy loss is more robust to annotation errors (R8) and results in models that produce attribution maps that are more focused on class-specific features (R3)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1930" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity Checks for Saliency Maps. In NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.157, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[2] Julius Adebayo, Michael Muelly, Harold Abelson, and Been Kim. Post hoc Explanations may be Ineffective for Detecting Unknown Spurious Correlation. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.198, + 0.469, + 0.25 + ], + "angle": 0, + "content": "[3] Saeid Asgari, Aliasghar Khani, Fereshte Khani, Ali Gholami, Linh Tran, Ali Mahdavi-Amiri, and Ghassan Hamarneh. MaskTune: Mitigating Spurious Correlations by Forcing to Explore. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.251, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[4] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation. PLOS One, 10(7):e0130140, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.319, + 0.468, + 0.345 + ], + "angle": 0, + "content": "[5] Jürgen Backhaus. The Pareto Principle. Analyse & Kritik, 2(2):146-171, 1980. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.346, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[6] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-Cos Networks: Alignment is All We Need for Interpretability. In CVPR, pages 10329-10338, 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.386, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[7] Moritz Böhle, Navdeepal Singh, Mario Fritz, and Bernt Schiele. B-cos Alignment for Inherently Interpretable CNNs and Vision Transformers. arXiv preprint arXiv:2306.10898, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.439, + 0.469, + 0.48 + ], + "angle": 0, + "content": "[8] Moritz Böhle, Mario Fritz, and Bernt Schiele. Convolutional Dynamic Alignment Networks for Interpretable Classifications. In CVPR, pages 10029-10038, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.48, + 0.469, + 0.534 + ], + "angle": 0, + "content": "[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks. In WACV, pages 839-847, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.535, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[10] Hila Chefer, Shir Gur, and Lior Wolf. Transformer Interpretability Beyond Attention Visualization. In CVPR, pages 782-791, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.574, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[11] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing Relevance Maps of Vision Transformers Improves Robustness. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.614, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[12] Piotr Dabkowski and Yarin Gal. Real Time Image Saliency for Black Box Classifiers. In NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.641, + 0.469, + 0.693 + ], + "angle": 0, + "content": "[13] Saurabh Desai and Harish G. Ramaswamy. Ablation-CAM: Visual Explanations for Deep Convolutional Network via Gradient-free Localization. In WACV, pages 983-991, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.694, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[15] Gabriel Erion, Joseph D Janizek, Pascal Sturmfels, Scott M Lundberg, and Su-In Lee. Improving Performance of Deep Learning Models with Axiomatic Attribution Priors and Expected Gradients. Nature Machine Intelligence, 3(7):620-631, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.845, + 0.469, + 0.884 + ], + "angle": 0, + "content": "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. IJCV, 88:303-308, 2009. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.885, + 0.469, + 0.898 + ], + "angle": 0, + "content": "[17] Thomas Fel, Ivan F Rodriguez Rodriguez, Drew Linsley, and" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.116, + 0.47, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.094, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Thomas Serre. Harmonizing the Object Recognition Strategies of Deep Neural Networks with Humans. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.134, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[18] Vitaly Feldman and Chiyuan Zhang. What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation. In NeurIPS, pages 2881-2891, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.175, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[19] Patrick Fernandes, Marcos Treviso, Danish Pruthi, André FT Martins, and Graham Neubig. Learning to Scaffold: Optimizing Model Explanations for Teaching. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.215, + 0.892, + 0.253 + ], + "angle": 0, + "content": "[20] Ruth C Fong and Andrea Vedaldi. Interpretable Explanations of Black Boxes by Meaningful Perturbation. In ICCV, pages 3429-3437, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.255, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[21] Felix Friedrich, Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. A Typology to Explore and Guide Explanatory Interactive Machine Learning. arXiv preprint arXiv:2203.03668, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.309, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[22] Hiroshi Fukui, Tsubasa Hirakawa, Takayoshi Yamashita, and Hironobu Fujiyoshi. Attention Branch Network: Learning of Attention Mechanism for Visual Explanation. In CVPR, pages 10705-10714, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[23] Yuyang Gao, Tong Steven Sun, Guangji Bai, Siyi Gu, Sungsoo Ray Hong, and Zhao Liang. RES: A Robust Framework for Guiding Visual Explanation. In KDD, pages 432-442, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.417, + 0.892, + 0.47 + ], + "angle": 0, + "content": "[24] Yuyang Gao, Tong Steven Sun, Liang Zhao, and Sungsoo Ray Hong. Aligning Eyes between Humans and Deep Neural Network through Interactive Attention Alignment. ACM HCI, 6(CSCW2):1-28, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.471, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[25] Hao Guo, Kang Zheng, Xiaochuan Fan, Hongkai Yu, and Song Wang. Visual Attention Consistency under Image Transforms for Multi-Label Image Classification. In CVPR, pages 729–739, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.578 + ], + "angle": 0, + "content": "[26] Misgina Tsighe Hagos, Kathleen M Curran, and Brian Mac Namee. Identifying Spurious Correlations and Correcting them with an Explanation-based Learning. arXiv preprint arXiv:2211.08285, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.579, + 0.892, + 0.619 + ], + "angle": 0, + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.62, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[28] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Fast Axiomatic Attribution for Neural Networks. In NeurIPS, pages 19513-19524, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[29] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely Connected Convolutional Networks. In CVPR, pages 4700-4708, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.892, + 0.753 + ], + "angle": 0, + "content": "[30] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. LayerCAM: Exploring Hierarchical Class Activation Maps for Localization. IEEE TIP, 30:5875-5888, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.755, + 0.892, + 0.779 + ], + "angle": 0, + "content": "[31] Joon Sik Kim, Gregory Plumb, and Ameet Talwalkar. Sanity Simulations for Saliency Methods. In ICML, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.781, + 0.892, + 0.821 + ], + "angle": 0, + "content": "[32] Keisuke Kiritoshi, Ryosuke Tanno, and Tomonori Izumitani. L1-Norm Gradient Penalty for Noise Reduction of Attribution Maps. In CVPRW, pages 118-121, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.821, + 0.892, + 0.861 + ], + "angle": 0, + "content": "[33] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell Me Where to Look: Guided Attention Inference Network. In CVPR, pages 9215-9223, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.862, + 0.892, + 0.887 + ], + "angle": 0, + "content": "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1931" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, pages 740-755. Springer, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.12, + 0.468, + 0.146 + ], + "angle": 0, + "content": "[35] Drew Linsley, Dan Shiebler, Sven Eberhardt, and Thomas Serre. Learning what and where to attend. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.146, + 0.468, + 0.214 + ], + "angle": 0, + "content": "[36] Masahiro Mitsuhara, Hiroshi Fukui, Yusuke Sakashita, Takanori Ogata, Tsubasa Hirakawa, Takayoshi Yamashita, and Hironobu Fujiyoshi. Embedding Human Knowledge into Deep Neural Network via Attention Map. In VISIGRAPP, pages 626-636, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.214, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[37] Ofir Moshe, Gil Fidel, Ron Bitton, and Asaf Shabtai. Improving Interpretability via Regularization of Neural Activation Sensitivity. arXiv preprint arXiv:2211.08686, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.255, + 0.468, + 0.294 + ], + "angle": 0, + "content": "[38] Vinod Nair and Geoffrey E Hinton. Rectified Linear Units Improve Restricted Boltzmann Machines. In ICML, pages 807-814, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.468, + 0.334 + ], + "angle": 0, + "content": "[39] Krishna Kanth Nakka and Mathieu Salzmann. Towards Robust Fine-Grained Recognition by Maximal Separation of Discriminative Features. In ACCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.468, + 0.362 + ], + "angle": 0, + "content": "[40] Vilfredo Pareto. Il massimo di utilità dato alla libera concorrenza. Giornale degli economisti, pages 48-66, 1894. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[41] Vilfredo Pareto. The Maximum of Utility given by Free Competition. Giornale degli Economisti e Annali di Economia, 67(3):387-403, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.402, + 0.468, + 0.455 + ], + "angle": 0, + "content": "[42] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On Guiding Visual Attention with Language Specification. In CVPR, pages 18092-18102, 2022. 1, 2, 3, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.468, + 0.495 + ], + "angle": 0, + "content": "[43] Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized Input Sampling for Explanation of Black-box Models. In BMVC, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.496, + 0.468, + 0.549 + ], + "angle": 0, + "content": "[44] Vipin Pillai, Soroush Abbasi Koohpayegani, Ashley Ouligian, Dennis Fong, and Hamed Piriavash. Consistent Explanations by Contrastive Learning. In CVPR, pages 10213-10222, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.549, + 0.468, + 0.576 + ], + "angle": 0, + "content": "[45] Vipin Pillai and Hamed Piri siavash. Explainable Models with Consistent Interpretations. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.468, + 0.616 + ], + "angle": 0, + "content": "[46] Sukrut Rao, Moritz Böhle, and Bernt Schiele. Towards Better Understanding Attribution Methods. In CVPR, pages 10223-10232, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[47] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"Why Should I Trust You?: Explaining the Predictions of Any Classifier. In KDD, pages 1135-1144, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.468, + 0.711 + ], + "angle": 0, + "content": "[48] Laura Rieger, Chandan Singh, William Murdoch, and Bin Yu. Interpretations are Useful: Penalizing Explanations to Align Neural Networks with Prior Knowledge. In ICML, pages 8116-8126, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.711, + 0.468, + 0.764 + ], + "angle": 0, + "content": "[49] Andrew Slavin Ross, Michael C Hughes, and Finale Doshi-Velez. Right for the Right Reasons: Training Differentiable Models by Constraining their Explanations. In *IJCAI*, pages 2662–2670, 2017. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.764, + 0.468, + 0.833 + ], + "angle": 0, + "content": "[50] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211–252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.833, + 0.468, + 0.887 + ], + "angle": 0, + "content": "[51] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization. In ICLR, 2020. 2, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[52] Patrick Schramowski, Wolfgang Stammer, Stefano Teso," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.527, + 0.092, + 0.892, + 0.161 + ], + "angle": 0, + "content": "Anna Brugger, Franziska Herbert, Xiaoting Shao, Hans-Georg Luigs, Anne-Katrin Mahlein, and Kristian Kersting. Making Deep Neural Networks Right for the Right Scientific Reasons by Interacting with their Explanations. Nature Machine Intelligence, 2(8):476-486, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.161, + 0.892, + 0.229 + ], + "angle": 0, + "content": "[53] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. In ICCV, pages 618-626, 2017. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.229, + 0.892, + 0.296 + ], + "angle": 0, + "content": "[54] Ramprasaath R Selvaraju, Stefan Lee, Yilin Shen, Hongxia Jin, Shalini Ghosh, Larry Heck, Dhruv Batra, and Devi Parikh. Taking a HINT: Leveraging Explanations to Make Vision and Language Models More Grounded. In ICCV, pages 2591-2600, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.296, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[55] Xiaoting Shao, Arseny Skryagin, Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. Right for Better Reasons: Training Differentiable Models by Constraining their Influence Functions. In AAAI, volume 35, pages 9533-9540, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[56] Haifeng Shen, Kewen Liao, Zhibin Liao, Job Doornberg, Maoying Qiao, Anton Van Den Hengel, and Johan W Verjans. Human-AI Interactive and Continuous Sensemaking: A Case Study of Image Classification using Scribble Attention Maps. In Extended Abstracts of CHI, pages 1-8, 2021. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[57] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning Important Features Through Propagating Activation Differences. In ICML, pages 3145-3153, 2017. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.473, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[58] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps. In ICLRw, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.514, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[59] Krishna Kumar Singh, Dhruv Mahajan, Kristen Grauman, Yong Jae Lee, Matt Feiszli, and Deepti Ghadiyaram. Don't Judge an Object by its Context: Learning to Overcome Contextual Bias. In CVPR, pages 11070-11078, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[60] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for Simplicity: The All Convolutional Net. In ICLRw, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[61] Guolei Sun, Salman Khan, Wen Li, Hisham Cholakkal, Fahad Shahbaz Khan, and Luc Van Gool. Fixing Localization Errors to Improve Image Classification. In ECCV, pages 271-287. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[62] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic Attribution for Deep Networks. In ICML, pages 3319-3328, 2017. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[63] Damien Teney, Ehsan Abbasnedjad, and Anton van den Hengel. Learning What Makes a Difference from Counterfactual Examples and Gradient Supervision. In ECCV, pages 580-599. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.757, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[64] Stefano Teso. Toward Faithful Explanatory Active Learning with Self-Explainable Neural Nets. In Workshop on IAL, pages 4–16. CEUR Workshop Proceedings, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.797, + 0.892, + 0.85 + ], + "angle": 0, + "content": "[65] Stefano Teso, Öznur Alkan, Wolfgang Stammer, and Elizabeth Daly. Leveraging Explanations in Interactive Machine Learning: An Overview. Frontiers in Artificial Intelligence, 6:1066049, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.85, + 0.892, + 0.878 + ], + "angle": 0, + "content": "[66] Stefano Teso and Kristian Kersting. Explanatory Interactive Machine Learning. In AIES, pages 239-245, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.878, + 0.892, + 0.891 + ], + "angle": 0, + "content": "[67] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.092, + 0.892, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1932" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks. In CVPRW, pages 111-119, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.174 + ], + "angle": 0, + "content": "[68] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or Signal: The Role of Image Backgrounds in Object Recognition. In *ICLR*, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.174, + 0.469, + 0.227 + ], + "angle": 0, + "content": "[69] Ziyan Yang, Kushal Kafle, Franck Dernoncourt, and Vicente Ordonez. Improving Visual Grounding by Encouraging Consistent Gradient-based Explanations. In CVPR, pages 19165–19174, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.228, + 0.469, + 0.268 + ], + "angle": 0, + "content": "[70] Matthew D Zeiler and Rob Fergus. Visualizing and Understanding Convolutional Networks. In ECCV, pages 818-833, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[71] Michael Zhang, Nimit Sharad Sohoni, Hongyang R. Zhang, Chelsea Finn, and Christopher Ré. Correct-N-Contrast: a Contrastive Approach for Improving Robustness to Spurious Correlations. In ICML, pages 26484–26516, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.469, + 0.364 + ], + "angle": 0, + "content": "[72] Yilun Zhou, Serena Booth, Marco Tulio Ribeiro, and Julie Shah. Do Feature Attribution Methods Correctly Attribute Features? In AAAI, volume 36, pages 9623-9633, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1933" + } + ] +] \ No newline at end of file diff --git a/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/88a53c18-99d0-4b4c-b0b2-16a2556a29ae_origin.pdf b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/88a53c18-99d0-4b4c-b0b2-16a2556a29ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..235aa05c4cfe3cd7d68c1c92f88e1eea8a71a0bc --- /dev/null +++ b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/88a53c18-99d0-4b4c-b0b2-16a2556a29ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e389d3d897671537d40ccbf34dc1a6b5ff00410ad000143ed85b63c6cf2a3c73 +size 3915146 diff --git a/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/full.md b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aeccba89af31261f81a6b8b91d5e800e92128669 --- /dev/null +++ b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/full.md @@ -0,0 +1,382 @@ +# Studying How to Efficiently and Effectively Guide Models with Explanations + +Sukrut Rao*, Moritz Böhle*, Amin Parchami-Araghi, Bernt Schiele +Max Planck Institute for Informatics, Saarland Informatics Campus, Saarbrücken, Germany + +{sukrut.rao,mboehle,mparcham,schiele}@mpi-inf.mpg.de + +# Abstract + +Despite being highly performant, deep neural networks might base their decisions on features that spuriously correlate with the provided labels, thus hurting generalization. To mitigate this, 'model guidance' has recently gained popularity, i.e. the idea of regularizing the models' explanations to ensure that they are "right for the right reasons" [49]. While various techniques to achieve such model guidance have been proposed, experimental validation of these approaches has thus far been limited to relatively simple and / or synthetic datasets. To better understand the effectiveness of the various design choices that have been explored in the context of model guidance, in this work we conduct an in-depth evaluation across various loss functions, attribution methods, models, and 'guidance depths' on the PASCAL VOC 2007 and MS COCO 2014 datasets. As annotation costs for model guidance can limit its applicability, we also place a particular focus on efficiency. Specifically, we guide the models via bounding box annotations, which are much cheaper to obtain than the commonly used segmentation masks, and evaluate the robustness of model guidance under limited (e.g. with only $1\%$ of annotated images) or overly coarse annotations. Further, we propose using the EPG score as an additional evaluation metric and loss function ('Energy loss'). We show that optimizing for the Energy loss leads to models that exhibit a distinct focus on object-specific features, despite only using bounding box annotations that also include background regions. Lastly, we show that such model guidance can improve generalization under distribution shifts. Code available at: https://github.com/sukrutrao/Model-Guidance + +# 1. Introduction + +Deep neural networks (DNNs) excel at learning predictive features that allow them to correctly classify a set of training images with ease. The features learnt on the training set, however, do not necessarily transfer to unseen images: i.e., instead of learning the actual class-relevant fea + +![](images/06af46ab131b75d0fd79d635858da8b202be74af1806d7c84ba16998e820671e.jpg) + +![](images/46ae3bb8b102d1418ad1f9744ad1bebe4dd1ffcf4c52f20827d5ed7e307137f8.jpg) +Fig. 1: (a) Model guidance increases object focus. Models may rely on irrelevant background features or spurious correlations (e.g. presence of person provides positive evidence for bicycle, center row, col. 1). Guiding the model via bounding box annotations can mitigate this and consistently increases the focus on object features (bottom row). (b) Model guidance can improve accuracy. In the presence of spurious correlations in the training data, non-guided models might focus on the wrong features. In the example image in (b), the waterbird is incorrectly classified to be a landbird due to the background (col. 3). Guiding the model via bounding box annotation (as shown in col. 2), the model can be guided to focus on the bird features for classification (col. 4). + +![](images/b9d6a75277e78fcfdc7194a485eede0fb298b87da971f1522c1d156a35ac17eb.jpg) + +![](images/c76c9e34f23c307f6b917e14d755c7c4d169d55763d83ba0c6aeac1c91e3fdd7.jpg) + +![](images/43a4783bb4bbe5f99ae90e6e5b08395ed5cdd3d143d363a78bee7e472b4c3797.jpg) +Landbird Confidence: $97\%$ + +![](images/f26565be58284a21fe4aa0090c377bd3239e308462cce93e1d4617623e1213d6.jpg) +Waterbird Confidence: $87\%$ + +tures, DNNs might memorize individual images (cf. [18]) or exploit spurious correlations in the training data (cf. [68]). For example, if bikes are highly correlated with people in the training data, a model might learn to associate the presence of a person in an image as positive evidence for a bike (e.g. Fig. 1a, col. 1, rows 1-2), which can limit how well it generalizes. Similarly, a bird classifier might rely on background features from the bird's habitat, and fail to correctly classify in a different habitat (cf. Fig. 1b cols. 1-3 and [42]). + +To detect such behaviour, recent advances in model interpretability have provided attribution methods (e.g. [53, 62, 57, 6]) to understand a model's reasoning. These methods typically provide attention maps that highlight regions of importance in an input to explain the model's decisions + +![](images/de598a16f0e7a21eb615b23f22952d798467a3441debf5c3d2cc06b5552a47e2.jpg) +Fig. 2: Qualitative results of model guidance. We show model-inherent B-cos explanations (input layer) of a B-cos ResNet-50 and GradCAM explanations (final layer) of a conventional ResNet-50 before ('Standard') and after optimization ('Guided') for images from the VOC test set, using our proposed Energy loss (Eq. (6)). Guiding the model via bounding box annotations consistently increases the focus on object features for both methods. Specifically, we find that background attributions are consistently suppressed in both cases. + +and can help identify incorrect reasoning such as reliance on spurious or irrelevant features, see for example Fig. 1b. + +As many attribution methods are in fact themselves differentiable (e.g. [57, 62, 53, 6]), recent work [49, 56, 24, 23, 66, 64] has explored the idea of using them to guide the models to make them "right for the right reasons" [49]. Specifically, models can be guided by jointly optimizing for correct classification as well as for attributing importance to regions deemed relevant by humans. This can help the model focus on the relevant features of a class, and correct errors in reasoning (Fig. 1b, col. 4). Such guidance has the added benefit of providing well-localized explanations that are thus easier to understand for end users (e.g. Fig. 2). + +While model guidance has shown promising results, a detailed study of how to do this most effectively is crucially missing. In particular, model guidance has so far been studied for a limited set of attribution methods and models and usually on relatively simple and/or synthetic datasets; further, the evaluation settings between approaches can significantly differ, which makes a fair comparison difficult. + +Therefore, in this work, we perform an in-depth evaluation of model guidance on large scale, real-world datasets, to better understand the effectiveness of a variety of design choices. Specifically, we evaluate model guidance along the following dimensions: the model architecture, the guidance depth1, the attribution method, and the loss function. In this context, we propose using the EPG score [67]—an evaluation metric that has thus far been used to evaluate the quality of attribution methods—as an additional loss function (which we call the Energy loss) as it is fully differentiable. + +Further, as annotation costs can be a major hurdle for making model guidance practical, we place a particular fo + +cus on efficient guidance. Specifically, we use bounding boxes instead of semantic segmentation masks, and evaluate the robustness of guidance techniques under limited or overly coarse annotations to reduce data collection costs. + +We find that our Energy loss lends itself well to those settings. On the one hand, it exhibits a high degree of robustness to limited or noisy bounding box annotations (cf. Figs. 10 and 12). On the other hand, despite the coarseness of bounding box guidance, it maintains a clear focus on object-specific features inside the bounding boxes, see Fig. 1a, row 3. In contrast, prior approaches often regularize for a uniform distribution of the attribution values inside the annotation masks, and thus tend to exhibit much lower attribution granularity (cf. Fig. 9). + +Contributions. (1) We perform an in-depth evaluation of model guidance on challenging large scale, multi-label classification datasets (PASCAL VOC 2007 [16], MS COCO 2014 [34]), assessing the impact of attribution methods, model architectures, guidance depths, and loss functions. Further, we show that, despite being relatively coarse, bounding box supervision can provide sufficient guidance to the models whilst being much cheaper to obtain than semantic segmentation masks. (2) We propose using the Energy Pointing Game (EPG) score [67] as an alternative to the IoU metric for evaluating the effectiveness of such guidance and show that the EPG score constitutes a good loss function for model guidance, particularly when using bounding boxes. (3) We show that model guidance can be performed cost-effectively by using annotation masks that are noisy or are available for only a small fraction (e.g. $1\%$ ) of the training data. (4) We show through experiments on the Waterbirds-100 dataset [51, 42] that model guidance with a small number of annotations suffices to improve the model's generalization under distribution shifts at test time. + +# 2. Related Work + +Attribution Methods [58, 60, 62, 57, 53, 67, 13, 30, 9, 43, 20, 70, 47, 12, 4] are often used to explain black-box models by generating heatmaps that highlight input regions important to the model's decision. However, such methods are often not faithful to the model [1, 46, 31, 72, 2] and risk misleading users. Recent work proposes inherently interpretable models [8, 6] that address this by providing modelfaithful explanations by design. In our work, we use both popular post-hoc and model-inherent attribution methods to guide models and discuss their effectiveness. + +Attribution Priors: Several approaches have been proposed for training better models by enforcing desirable properties on their attributions. These include enforcing consistency against augmentations [45, 44, 25], smoothness [15, 37, 32], separation of classes [71, 44, 61, 39, 59], or constraining the model's attention [22, 3]. In contrast, in this work, we focus on providing explicit human guidance to the model using bounding box annotations. This constitutes more explicit guidance but allows fine-grained control over the model's reasoning even with few annotations. + +Model Guidance: In contrast to the indirect regularization effect achieved by attribution priors, various approaches have been proposed (cf. [21, 65]) to actively guide models by regularizing their attributions, for tasks such as classification [49, 24, 23, 48, 42, 26, 63, 36, 66, 64, 52, 55, 35, 56, 69, 17], segmentation [33], VQA [54, 63], and knowledge distillation [19]. The goal of such approaches is not only to improve performance, but also make sure that the model is "right for the right reasons" [49]. For classifiers, this typically involves jointly optimizing both for classification performance and localization to object features. While various benefits of model guidance have been reported, most prior work evaluate on simple datasets [49, 55, 24, 23] and, thus far, no common evaluation setting has emerged. Recently, [11] has extended model guidance to ImageNet, showing that its benefits can scale to large scale problems. In contrast to [11], who investigated one particular attribution method [10], our focus lies on a better understanding of the impact of the different design choices for model guidance. + +To distill the most effective techniques for model guidance, in this work, we conduct an in-depth evaluation on challenging, commonly used real-world multi-label classification datasets (PASCAL VOC 2007, MS COCO 2014). Specifically, we perform a comprehensive comparison across multiple dimensions of interest: the loss function, the model architecture, the guidance depth, and the attribution method. For this, we evaluate the localization losses introduced in the closest related work, i.e. RRR [49], HAICS [56], and GRADIA [24]; additionally, we propose using the EPG metric [67] as a loss function and show that it has various desirable properties, in particular when guiding models via bounding box annotations. + +![](images/0b0d9ba08868ff0b01f0a00e2862efa1810cab23cbb1ded746e94d3dcf66b357.jpg) +Fig. 3: Model guidance overview. We jointly optimize for classification $(\mathcal{L}_{\mathrm{class}})$ and localization of attributions to human-annotated bounding boxes $(\mathcal{L}_{\mathrm{loc}})$ , to guide the model to focus on object features. Various localization loss functions can be used, see Sec. 3.4. + +Finally, model guidance has also been used to mitigate reliance on spurious features using language guidance [42], and we show that using a small number of coarse bounding box annotations can be similarly effective. + +Evaluating Model Guidance: The benefits of model guidance have typically been shown via improvements in classification performance (e.g. [49, 48]) or an increase in IoU between object masks and attribution maps (e.g. [23, 33]). In addition to these metrics, we also evaluate on the EPG metric [67], which has thus far only been used to evaluate the quality of the attribution methods themselves. We further show that it lends itself well to being used as a guidance loss, as it places only minor constraints on the model, and, in contrast to the IoU metric, it is fully differentiable. + +# 3. Guiding Models Using Attributions + +In this section, we provide an overview of the model guidance approach that jointly optimizes for classification and localization (Sec. 3.1). Specifically, we describe the attribution methods (Sec. 3.2), metrics (Sec. 3.3), and localization loss functions (Sec. 3.4) that we evaluate in Sec. 5. In Sec. 3.5 we discuss our strategy to train for localization in the presence of multiple ground truth classes. + +Notation: We consider a multi-label classification problem with $K$ classes with $X \in \mathbb{R}^{C \times H \times W}$ the input image and $y \in \{0,1\}^K$ the one-hot encoding of the image labels. With $A_k \in \mathbb{R}^{H \times W}$ we denote an attribution map for a class $k$ for $X$ using a classifier $f$ ; $A_k^+$ denotes the positive component of the attributions, $\hat{A}_k = \frac{A_k}{\max(\mathrm{abs}(A_k))}$ normalized attributions, and $\hat{A}_k^+ = \frac{A_k^+}{\max(A_k^+)}$ normalized positive attributions. Finally, $M_k \in \{0,1\}^{H \times W}$ denotes the binary mask for class $k$ , which is given by the union of bounding boxes of all occurrences of class $k$ in $X$ . + +# 3.1. Model Guidance Procedure + +Following prior work (e.g. [49, 56, 24, 23]), the model is trained jointly for classification and localization (cf. Fig. 3): + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c l a s s}} + \lambda_ {\text {l o c}} \mathcal {L} _ {\text {l o c}}. \tag {1} +$$ + +I.e., the loss consists of a classification loss $(\mathcal{L}_{\mathrm{class}})$ , for which we use binary cross-entropy, and a localization loss + +$(\mathcal{L}_{\mathrm{loc}})$ , which we discuss in Sec. 3.4; here, the hyperparameter $\lambda_{\mathrm{loc}}$ controls the weight given to each of the objectives. + +# 3.2. Attribution Methods + +In contrast to prior work that typically use GradCAM [53] attributions, we perform an evaluation over a selection of popularly used differentiable $^2$ attribution methods which have been shown to localize well [46]: IxG [57], IntGrad [62], and GradCAM [53]. We further evaluate model-inherent explanations of the recently proposed B-cos models [6]. To ensure comparability across attribution methods [46], we evaluate all attribution methods at the input, various intermediate, and the final spatial layer. + +IxG [57] computes the element-wise product $\odot$ of the input and the gradients of the $k$ -th output w.r.t. the input, i.e. $X\odot \nabla_{X}f_{k}(X)$ . For piece-wise linear models such as DNNs with ReLU activations [38], this faithfully computes the linear contributions of a given input pixel to the model output. + +GradCAM [53] computes importance attributions as a ReLU-thresholded, gradient-weighted sum of activation maps. In detail, it is given by $\mathrm{ReLU}(\sum_{c}\alpha_{c}^{k}\odot U_{c})$ with $c$ denoting the channel dimension, and $\alpha^k$ the average-pooled gradients of the output for class $k$ with respect to the activations $U$ of the last convolutional layer in the model. + +IntGrad [62] takes an axiomatic approach and is formulated as the integral of gradients over a straight line path from a baseline input to the given input $X$ . Approximating this integral requires several gradient computations, making it computationally expensive for use in model guidance. To alleviate this, when optimizing with IntGrad, we use the recently proposed $\mathcal{X}$ -DNN models [28] that allow for an exact computation of IntGrad in a single backward pass. + +B-cos [6] attributions are generated using the inherently-interpretable B-cos networks, which promote alignment between the input $\mathbf{x}$ and a dynamic weight matrix $\mathbf{W}(\mathbf{x})$ during optimization. In our experiments, we use the contribution maps given by the element-wise product of the dynamic weights with the input $(\mathbf{W}_k^T (\mathbf{x})\odot \mathbf{x})$ , which faithfully represent the contribution of each pixel to class $k$ . To be able to guide B-cos models, we developed a differentiable implementation of B-cos explanations, see supplement. + +# 3.3. Evaluation Metrics + +We evaluate the models' performance on both our training objectives: classification and localization. For classification, we use the F1 score and mean average precision (mAP). We discuss the localization metrics below. + +Intersection over Union (IoU) is a commonly used metric (cf. [23]) that computes the intersection between the ground truth annotation masks and the binarized attribution maps, + +normalized by their union; for binarization, a threshold parameter needs to be chosen. In this work, the ground truth masks are taken to be the union of all bounding boxes of a class in the image and, following prior work [20], the threshold parameter is selected via a heldout set. + +Energy-based Pointing Game (EPG) [67] measures the concentration of attribution energy within the mask, i.e. the fraction of positive attributions inside the bounding boxes: + +$$ +\mathrm {E P G} _ {k} = \frac {\sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} M _ {k , h w} A _ {k , h w} ^ {+}}{\sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} A _ {k , h w} ^ {+}}. \tag {2} +$$ + +In contrast to IoU, EPG more faithfully takes into account the relative importance given to each input region, since it does not binarize the attributions. Like IoU, the scores lie in [0, 1], with higher scores indicating better localization. + +# 3.4. Localization Losses + +We evaluate the most commonly used localization losses $(\mathcal{L}_{\mathrm{loc}}$ in Eq. (1)) from prior work. We describe these losses as applied on attribution maps of an image for a single class $k$ , as well as the proposed EPG-derived Energy loss. + +$L_{1}$ loss ([24, 23], Eq. (3)) minimizes the $L_{1}$ distance between annotation masks and normalized positive attributions $\hat{A}_k^+$ , guiding the model towards uniform attributions inside the mask and suppressing attributions outside of it. + +$$ +\mathcal {L} _ {\mathrm {l o c}, k} = \frac {1}{H \times W} \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} \| M _ {k, h w} - \hat {A} _ {k, h w} ^ {+} \| _ {1} \tag {3} +$$ + +Per-pixel cross entropy (PPCE) loss ([56], Eq. (4)) applies a binary cross entropy loss between the mask and the normalized positive annotations $\hat{A}_k^+$ , thus guiding the model to maximize the attributions inside the mask: + +$$ +\mathcal {L} _ {\text {l o c}, k} = - \frac {1}{\| M _ {k} \| _ {1}} \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} M _ {k, h w} \log \left(\hat {A} _ {k, h w} ^ {+}\right). \tag {4} +$$ + +As PPCE does not constrain attributions outside the mask, there is no explicit pressure to avoid spurious features. + +RRR\* loss ([49], Eq. (5)). [49] introduced the RRR loss to regularize the normalized input gradients $\hat{A}_{k,hw}$ as + +$$ +\mathcal {L} _ {\mathrm {l o c}, k} = \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} \left(1 - M _ {k, h w}\right) \hat {A} _ {k, h w} ^ {2}. \tag {5} +$$ + +To extend it to our setting, we take $\hat{A}_{k,hw}$ to be given by an arbitrary attribution method (e.g. IntGrad); we denote this generalized version by RRR*. In contrast to the PPCE loss, RRR* only regularizes attributions outside the ground truth masks. While it thus does not introduce a uniformity prior similar to the $L_{1}$ loss, it also does not explicitly promote high importance attributions inside the masks. + +Energy Loss. In addition to the losses described in prior work, we propose to also evaluate using the EPG score ([67], Eq. (2)) as a loss function for model guidance, as it is fully differentiable. In particular, we simply define it as + +$$ +\mathcal {L} _ {\text {l o c}, k} = - \operatorname {E P G} _ {k}. \tag {6} +$$ + +![](images/7c1905b038578c97d9c819dc9c64c3fced972e9568c5b314f49e47545c77d741.jpg) +Fig. 4: Selecting models for evaluation. For each configuration, we evaluate every model at every checkpoint and measure its performance across various metrics (F1, EPG, IoU) on the validation set; i.e. every point in the left graph corresponds to one model (for B-cos models optimized via the Energy loss at the input layer). Instead of evaluating a single model on the test set, we evaluate all Pareto-dominant models, as indicated in the center and right plot. + +Unlike existing localization losses that either (i) do not constrain attributions across the entire input $(\mathsf{RRR}^*,\mathsf{PPCE})$ or (ii) force the model to attribute uniformly within the mask even if it includes irrelevant background regions $(L_{1},$ PPCE), maximizing the EPG score jointly optimizes for higher attribution energy within the mask and lower attribution energy outside the mask. By not enforcing a uniformity prior, we find that the Energy loss is able to provide effective guidance while allowing the model to learn freely what to focus on within the bounding boxes (Sec. 5). + +# 3.5. Efficient Optimization + +In contrast to prior work [49, 56, 24, 23], we perform model guidance on a multi-label classification setting, and consequently there are multiple ground truth classes whose attribution localization could be optimized. Computing and optimizing for several attributions within an image would add a significant overhead to the computational cost of training (multiple backward passes). Hence, for efficiency, we sample one ground truth class $k$ per image at random for every batch and only optimize for localization of that class, i.e., $\mathcal{L}_{\mathrm{loc}} = \mathcal{L}_{\mathrm{loc},k}$ . We find that this still provides effective model guidance while keeping the training cost tractable. + +# 4. Experimental Setup + +In this section, we describe our experimental setup and how we select the best models across metrics; for full details, see supplement. We evaluate across all possible choices for each category, and discuss our results in Sec. 5. Datasets: We evaluate on PASCAL VOC 2007 [16] and MS COCO 2014 [34] for multi-label image classification. In Sec. 5.5, to understand the effectiveness of model guidance in mitigating spurious correlations, we also evaluate on the synthetically constructed Waterbirds-100 dataset [51, 42], where landbirds are perfectly correlated with land backgrounds on the training and validation sets, but are equally likely to occur on land or water in the test set (similar for waterbirds and water). With this dataset, we evaluate model guidance for suppressing undesired features. + +Attribution Methods and Architectures: As described in Sec. 3.2, we evaluate with IxG [57], IntGrad [62], B-cos [6, 7], and GradCAM [53] using models with a ResNet-50 [27] backbone. For IntGrad, we use an $\mathcal{X}$ -DNN ResNet-50 [28] to reduce the computational cost, and a B-cos ResNet-50 for the B-cos attributions. To emphasize that the results generalize across different backbones, we further provide results for a B-cos ViT-S [14, 7] and a B-cos DenseNet-121 [29, 7]. We evaluate optimizing the attributions at different network layers, such as at the input image and the last convolutional layers' output3, as well as at multiple intermediate layers. Within the main paper, we highlight some of the most representative and insightful results, the full set of results can be found in the supplement. All models were pretrained on ImageNet [50], and model guidance was applied when fine-tuning the models on the target dataset. + +Localization Losses: As described in Sec. 3.4, we compare four localization losses in our evaluation: (i) Energy, (ii) $L_{1}$ [24, 23], (iii) PPCE [56], and (iv) RRR* (cf. Sec. 3.4, [49]). + +Evaluation Metrics: As discussed in Sec. 3.3, we evaluate both for classification and localization performance of the models. For classification, we report the F1 scores, similar results with mAP scores can be found in the supplement. For localization, we evaluate using the EPG and IoU scores. + +Selecting the best models: As we evaluate for two distinct objectives (classification + localization), it is not trivial to decide which models perform 'the best', e.g. a model that provides the best classification performance might provide significantly worse localization than a model that provides only slightly lower classification performance. Finding the right balance and deciding which of those models in fact constitutes the 'better' model depends on the preference of the end user. Hence, instead of selecting models based on a single metric, we select the set of Pareto-dominant models [40, 41, 5] across three metrics—F1, EPG, and IoU—for each training configuration, as defined by a combination of attribution method, layer, and loss. Specifically, as shown in Fig. 4, we train each configuration using three different choices of $\lambda_{\mathrm{loc}}$ , and select the set of Pareto-dominant models among all checkpoints (epochs and $\lambda_{\mathrm{loc}}$ ). This provides a more holistic view of the general trends on the effectiveness of model guidance for each configuration. + +# 5. Experimental Results + +In this section, we discuss our experimental findings. In particular, in Sec. 5.1, we first discuss the impact of the loss functions on the EPG and IoU scores of the models; in Sec. 5.2, we then analyze the impact of the models and attribution methods; further in Sec. 5.3, we show that guiding the models via their explanations can lead to improved classification accuracy. In Sec. 5.4, we present additional + +![](images/764f59edc7188ec75dcec414c7552a78107ec05f498b12eedbbbb5130475fa2c.jpg) +(a) PASCAL VOC results for EPG vs. F1. + +![](images/4aca215a1dcdf93ee79856e08a888240bcaae524424608497686c3e34aca2b8f.jpg) + +![](images/a7d99e6d0282dec595c3f093439d3f16fc412916ea77ca7651947db7d9e5d30b.jpg) + +![](images/98957feb031bab97836e5557cf46b9bfda752eac90e54fe772656344f125f10d.jpg) +(b) MS COCO results for EPG vs. F1. + +![](images/bbcccf4741a0ea918ecf24ff859dc2edefc52dc5869a87b289281a21bb5d03be.jpg) + +![](images/30a1afe1bf695ecd1a3cf233243c3111e88bc837e5321126b6c31ced676b4e29.jpg) + +![](images/bf6798f67f3cdb993070819b8ad63e26b8f3bdcd2e1631b4917bbe88405df895.jpg) +Fig. 5: EPG vs. F1, for different datasets ((a): VOC; (b): COCO), losses (markers) and models (columns), optimized at different layers (rows); additionally, we show the performance of the baseline model before fine-tuning and demarcate regions that strictly dominate (are strictly dominated by) the baseline performance in green (grey). For each configuration, we show the Pareto fronts (cf. Fig. 4) across regularization strengths $\lambda_{\mathrm{loc}}$ and epochs (cf. Sec. 5 and Fig. 4). We find the Energy loss to give the best trade-off between EPG and F1. + +![](images/9ed3e743196a6931b8b142e0773581defda3439b3d94332aa17af2b6c745f881.jpg) +Fig. 6: IoU vs. F1, for different losses (markers) and models (columns) for VOC; results for COCO are in the supplement. Additionally, we show the performance of the baseline model before fine-tuning and demarcate regions that strictly dominate (are strictly dominated by) the baseline model in green (grey). For each configuration, we show the Pareto fronts (Fig. 4) across regularization strengths $\lambda_{\mathrm{loc}}$ and all epochs; for details, see Secs. 4 and 5. Across all configurations, we find the $L_{1}$ loss to provide the largest gains in IoU at the lowest cost. + +![](images/961ad1912b1b2924ee63112b372d0ad9667d12f307e7cec569605538a6cfaa0d.jpg) + +![](images/9fb3b15b6d409434a3aaf2d1988a0266eaf4b85308a2b1f66504a6c022f1e917.jpg) +Fig. 7: EPG vs. F1 on VOC. We observe the same trends as in Fig. 5a for different backbone architectures, specifically a B-cos DenseNet-121 and a B-cos ViT-S. For IoU results, see supplement. + +studies in which we evaluate and discuss the cost of model guidance approaches: in particular, we study model guidance with limited additional labels, with increasingly coarse + +bounding boxes, and at deep layers in the network. Finally, in Sec. 5.5, we show the utility of model guidance in improving accuracy in the presence of distribution shifts. For easier reference, we label our individual findings as R1-R9. Note. To draw conclusive insights and highlight general and reliable trends in the experiments, we compare the Pareto curves (see Fig. 4) of individual configurations. If the Pareto curve of a specific loss (e.g. Energy in Fig. 5) consistently Pareto-dominates the Pareto curves of all other losses, we can confidently conclude that for the combination of evaluated metrics (e.g. EPG vs. F1), this loss is the best choice. + +![](images/57dacd8004bef9b1b5471ea5f84371e0d5cd90b1fe77101f40e107e02c8858ad.jpg) +Fig. 8: Faster training by guiding at later layers. While input-level attributions tend to be more detailed (cf. Fig. 2), they are costlier to compute than attributions at later layers. However, we find that guidance at later layers (e.g. @Mid3) also significantly improves input-level attributions, yielding similar EPG results as input-level guidance (@Input) at up to twice the training speed; for IoU results, see supplement. + +![](images/68f77076c0fc0a2854aceeb2cae1d34d01a892c4ae8764f80201ea7bb1a06057.jpg) + +![](images/462d3d36f2f0a59e351b2e1613c75efd7065699303ff68a1520a4d679bc1c385.jpg) + +![](images/64bc2fd43df95645b1076b0b3e789287ee399fbfe3b8d4805c50ecfb31b00a18.jpg) + +# 5.1. Comparing loss functions for model guidance + +In the following, we highlight the main insights gained from the quantitative evaluations. For a qualitative comparison between the losses, please see Fig. 9; note that we show examples for a B-cos model as the differences become clearest; full results can be found in the supplement. + +R1 The Energy loss yields the best EPG scores. In Fig. 5, we plot the Pareto curves for EPG vs. F1 scores for a wide range of configurations (see Sec. 4) on VOC (a) and COCO (b); specifically, we group the results by model type (Vanilla, $\mathcal{X}$ -DNN, B-cos), the layer depths at which the attribution was regularized (Input / Final), and the loss used during optimization (Energy, $L_{1}$ , PPCE, RRR*). From these results it becomes apparent that the optimization with the Energy loss yields the best trade-off between accuracy (F1) and the EPG score: e.g., when looking at the upper right plot in Fig. 5a we can see that the Energy loss (red dots) improves over the baseline B-cos model (white cross) by improving the localization in terms of EPG score with only a minor cost in classification performance (i.e. F1 score). Further trading off F1 scores yields even higher EPG scores. Importantly, the Energy loss Pareto-dominates all the other losses (RRR*: blue diamonds; $L_{1}$ : green triangles; PPCE: yellow pentagons). This is is also true for the other network types (Vanilla ResNet-50, Fig. 5a (top left), and $\mathcal{X}$ -DNN, Fig. 5a (top center)) and at the final layer (bottom row), and generalizes across backbone architectures (Fig. 7). When comparing Fig. 5a and Fig. 5b, we also find these results to be highly consistent between datasets. +R2 The $L_{1}$ loss yields the best IoU performance. Similarly, in Fig. 6, we plot the Pareto curves of IoU vs. F1 scores for various configurations at the final layer; for the IoU results at the input layer and on the COCO dataset, please see the supplement. For IoU, the $L_{1}$ loss provides the best trade-off and, with few exceptions, $L_{1}$ -guided models Pareto-dominate all other models in all configurations. +R3 The Energy loss focuses best on on-object features. By not forcing the models to highlight the entire bounding boxes (see Sec. 3.4), we find that the Energy loss also suppresses background features within the bounding boxes, thus better preserving fine details of the explanations + +![](images/15644f355a0560a56d95c049e06dea1248bd183d6c7311aa4f89b6c5bd84c400.jpg) +Fig. 9: Loss comparison for input attributions (atts.) of a B-cos model. We showatts. before (baseline, col. 2) and after guidance (cols. 3-6) for a specific image (col. 1) and its bounding box annotation. We find that Energy and RRR* yield sparseatts, whereas $L_{1}$ yields smootheratts, as it is optimized to fill the entire bounding box. For PPCE we observe only a minor effect on theatts. + +(cf. Figs. 9 and 11). To quantify this, we evaluate the distribution of Energy (Eq. (2)) just within the bounding boxes. For this, we take advantage of the segmentation mask annotations available for a subset of the VOC test set. Specifically, we measure the Energy contained in the segmentation masks versus the entire bounding box, which indicates how much of the attributions actually highlight on-object features. We find that the Energy loss outperforms $L_{1}$ across all models and configurations; see supplement for details. + +In short, we find that the Energy loss works best for improving the EPG metric, whereas the $L_{1}$ loss yields the highest gains in terms of IoU; depending on the use case, either of these losses could thus be recommendable. However, we find that the Energy loss is more robust to annotation errors (R8, Sec. 5.4), and, as discussed in R3, the Energy loss more reliably focuses on object-specific features. + +# 5.2. Comparing models and attribution methods + +In the following, we highlight our findings regarding different attribution methods and models. Given the similarity of the results between GradCAM and IxG, and since B-cos attributions performed better than GradCAM for B-cos models, we show GradCAM results in the supplement. + +R4 At the input layer, B-cos explanations perform best. We find that the B-cos models not only achieve the highest EPG/IoU performance before applying model guidance, ('baseslines') but also obtain the highest gains in EPG and IoU and thus the highest overall performance (for EPG see Fig. 5, right; for IoU, see supplement): e.g., an Energy-based B-cos model achieves an EPG score of $71.7@79.4\%$ F1, thus significantly outperforming the best EPG scores of + +both other model types at a much lower cost in F1 (Vanilla: 55.8 @ $69.0\%$ , $\mathcal{X}$ -DNN: 62.3 @ $68.9\%$ ). This is also observed qualitatively, as we show in the supplement. + +R5 Regularizing at the final layer yields consistent gains. As can be seen in Fig. 5 (bottom) and Fig. 6, all models can be guided well via regularization at the final layer, i.e. all models show improvements in IoU and EPG score. + +In short, we find model guidance to work well across all tested models when optimizing at the final layer (R5), highlighting its wide applicability. However, to obtain highly detailed and well-localized attributions at the input layer, the model-inherent explanations of the B-cos models seem to lend themselves much better to such guidance (R4). + +# 5.3. Improving accuracy with model guidance + +R6 Model guidance can improve accuracy. For both the Vanilla models (final layer) and the $\mathcal{X}$ -DNNs (input+final), we found models that improve the localization metrics and the F1 score. These improvements are particularly pronounced for the $\mathcal{X}$ -DNN: e.g., we find models that improve the EPG and F1 scores by $\Delta = 7.2$ p.p. and $\Delta = 1.4$ p.p. respectively (Fig. 5, center top), or the IoU and F1 scores by $\Delta = 11.9$ p.p. and $\Delta = 1.4$ p.p. (Fig. 6, center). + +However, overall we observe a trade-off between localization and accuracy (Figs. 5 and 6). Given the similarity of the training and test distributions, focusing on the object need not improve classification performance, as spurious features are also present at test time. Further, the guided model is discouraged from relying on contextual features, making the classification more challenging. In Sec. 5.5, we show that guidance can significantly improve performance when there is a distribution shift between training and test. + +# 5.4. Efficiency and robustness considerations + +While bounding boxes decrease the data collection cost with respect to segmentation masks, they can nonetheless be expensive to obtain, especially when expert knowledge is required. To further reduce those costs, in this section, we assess the robustness of guiding the model with a limited number (R7) or increasingly coarse annotations (R8). Apart from data efficiency, we further explore how training efficiency can be improved for fine-grained (i.e. input-level) explanations (R9), as explanations at early layers are more costly to obtain than those at later layers. + +R7 Model guidance requires only few add. annotations. In Fig. 12, we show that the EPG score can be significantly improved with a very limited number of annotations; for IoU results, see supplement. Specifically, we find that when using only $1\%$ of the training data (25 annotated images) for VOC, improvements of up to $\Delta = 23.0$ p.p. ( $\Delta = 1.4$ ) in EPG (IoU) can be obtained, at a minor drop in F1 ( $\Delta = 0.3$ p.p. and $\Delta = 2.5$ p.p. respectively). When annotating up to $10\%$ of the images, very similar results can be achieved as + +with full annotation (see e.g. cols. 2+3 in Fig. 12). + +R8 The Energy loss is highly robust to annotation errors. As discussed in Sec. 3.4, the Energy loss only directs the model on which features not to use and does not impose a uniform prior on the attributions within the bounding boxes. As a result, we find it to be much more stable to annotation errors: e.g., in Fig. 10, we visualize how the EPG (top) and IoU (bottom) scores of the best performing models under the Energy (left) and $L_{1}$ loss (right) evolve when using coarser bounding boxes; for this, we simply dilate the bounding box size by $p \in \{10,25,50\} \%$ during training, see Fig. 11. While the models optimized via the $L_{1}$ loss achieve increasingly worse results (right), the Energy-optimized models are essentially unaffected by the coarseness of the annotations. + +![](images/f385551e04751c4ed6c0b068e6b01fe2ff7332ef175fa5fbb71be2568634703c.jpg) + +![](images/3eb12969c3771dbff17478fb55232e4d736c41504d1530c037e666481a91a7cf.jpg) + +![](images/acedbb47fc5a14fd412f4776d2b3ad4129b40abe204f5564486250c8cf229612.jpg) +Fig. 10: Quantitative results for dilated bounding boxes for a B-cos model at the input layer. We show EPG and IoU (top and bottom) results for models trained with various amounts of annotation errors (increasingly large bounding boxes, see Fig. 11). The Energy loss yields highly consistent results despite training with heavily dilated bounding boxes (left), whereas the results of the $L_{1}$ loss (right) worsen markedly; best viewed in color. + +![](images/0b57e7c7fc7253a8c736dc3f1bbaba077b59a9dae3ec20d35a382f47dd59edc2.jpg) + +![](images/76750dff31b25685457afbe6b1c59b05ebd692428814b23b9cece6fc7073e0dd.jpg) +Fig. 11: Qualitative results for dilated bounding boxes for a B-cos model at input. Examples for attributions (rows $2 + 3$ ) of models trained with dilated bounding boxes (row 1). In contrast to $L_{1}$ , models trained with Energy show significant gains in object focus even with significant noise (e.g. 'Baseline' vs. $50\%$ ). + +![](images/53d85424fbc3d08bf14d5d6ac9eb76b301d08444ebe4265b9c93b3fabd20ac34.jpg) + +In short, we find that the models can be guided effectively at a low cost in terms of annotation effort, as only few + +annotations (e.g. 25 for VOC) are required (cf. R7), and, especially for the Energy loss, these annotations can be very coarse and do not have to be 'pixel-perfect' (cf. R8). + +R9 Guidance at deep layers can be effective. While guided input-level explanations of B-cos networks exhibit a high degree of detail, regularizing those explanations comes at an added training cost. In particular, optimizing at the input layer requires backpropagating through the entire network to compute the attributions. In an effort to reduce training costs whilst maintaining the benefits of fine-grained explanations at input resolution, we evaluate if input-level attributions benefit from an optimization at deeper layers. + +Specifically, we regularize B-cos attributions at the final and at three intermediate layers (Mid{1,2,3}), and evaluate the localization of attributions at the input. We find (Fig. 8) that training at a deeper layer can provide significant speed-ups in training time with often a negligible cost in localization performance. E.g., since we do not have to compute a full backward pass through the entire model during training, optimizing at Mid2 (col. 2 in Fig. 8) provides similar gains in localization but with a $1.7\mathrm{x}$ speed-up in training time. + +![](images/958e439fecad0c533c1d88ff82b917fa6451906acf38fd1f5d3d7b3ee3521e6d.jpg) +Fig. 12: EPG results with limited annotations for a B-cos model at the input layer, optimized with the Energy and the $L_{1}$ loss. Using bounding box annotations for as little as $1\%$ (left) of the images yields significant improvements in EPG, and with $10\%$ (center) similar gains as in the fully annotated setting (right) are obtained. + +# 5.5. Effectiveness against spurious correlations + +To evaluate the potential for mitigating spurious correlations, we evaluate model guidance with the Energy and $L_{1}$ losses on the synthetically constructed Waterbirds-100 dataset [51, 42]. We perform model guidance under two settings: (1) the conventional setting to classify between landbirds and waterbirds, using the region within the bounding box as the mask; and (2) the reversed setting [42] to classify the background, i.e., land vs. water, using the region outside the bounding box as the mask. To simulate a limited annotation budget, we only use bounding boxes for a random $1\%$ of the training set, and report results averaged over four runs. We show the results for the worst-group accuracy (i.e., images containing a waterbird on land) and the overall accuracy using B-cos models in Tab. 1; full results for all attributions and models can be found in the supplement. + +Both losses consistently and significantly improve the accuracy in the conventional and the reversed settings by + +![](images/c1ca9fe9ad24bf88749f50181dd215a50a76af682ee79a0dc622f5ded838f8ba.jpg) +Fig. 13: Qualitative Waterbirds-100 results. Without guidance, a model might focus on the background to classify birds (baseline) and thus misclassify waterbirds on land (col. 2). Guided models can correct such errors and focus on the desired feature: in cols. $3 + 4$ $(5 + 6)$ the model is guided to classify by using the bird (background) features and arrives at the desired prediction. Model predictions and confidence scores are indicated below the images. + +
ConventionalReversed
ModelWorstOverallWorstOverall
Baseline43.4 (±2.4)68.7 (±0.2)56.6 (±2.4)80.1 (±0.2)
Energy56.1 (±4.0)71.2 (±0.1)62.8 (±2.1)83.6 (±1.1)
L151.1 (±1.9)69.5 (±0.2)58.8 (±5.0)82.2 (±0.9)
+ +Table 1: Waterbirds-100 results. We find that model guidance is effective in improving both worst-group ('Waterbird on Land') and overall accuracy in the conventional (Landbird vs. Waterbird) and reversed (Land vs. Water) settings; full results in the supplement. + +guiding the model to select the 'right' features, i.e. birds (conventional) or background (reversed). This guidance can also be observed qualitatively (cf. Fig. 13). + +# 6. Discussion And Conclusion + +In this work, we comprehensively evaluated various models, attribution methods, and loss functions for their utility in guiding models to be "right for the right reasons". + +In summary, we find that guiding models via bounding boxes can significantly improve EPG and IoU performance of the optimized attribution method, with the Energy loss working best to improve the EPG score (R1) and the $L_{1}$ loss yielding the highest gains in IoU scores (R2). While the B-cos models achieve the best results in IoU and EPG score at the input layer (R4), all tested model types (Vanilla, $\mathcal{X}$ -DNN, B-cos) lend themselves well to being optimized at the final layer (R5), which can even improve attribution maps at early layers (R9). Further, we find that regularizing the explanations of the models and thereby 'telling them where to look' can increase the object recognition performance (mAP/accuracy) of some models (R6), especially when strong spurious correlations are present (Sec. 5.5). Interestingly, those gains (EPG, IoU), can be achieved with relatively little additional annotation (R7). Lastly, we find that by not assuming a uniform prior over the attributions within the annotated bounding boxes, training with the energy loss is more robust to annotation errors (R8) and results in models that produce attribution maps that are more focused on class-specific features (R3). + +# References + +[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity Checks for Saliency Maps. In NeurIPS, 2018. 3 +[2] Julius Adebayo, Michael Muelly, Harold Abelson, and Been Kim. Post hoc Explanations may be Ineffective for Detecting Unknown Spurious Correlation. In ICLR, 2022. 3 +[3] Saeid Asgari, Aliasghar Khani, Fereshte Khani, Ali Gholami, Linh Tran, Ali Mahdavi-Amiri, and Ghassan Hamarneh. MaskTune: Mitigating Spurious Correlations by Forcing to Explore. In NeurIPS, 2022. 3 +[4] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation. PLOS One, 10(7):e0130140, 2015. 3 +[5] Jürgen Backhaus. The Pareto Principle. Analyse & Kritik, 2(2):146-171, 1980. 5 +[6] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-Cos Networks: Alignment is All We Need for Interpretability. In CVPR, pages 10329-10338, 2022. 1, 2, 3, 4, 5 +[7] Moritz Böhle, Navdeepal Singh, Mario Fritz, and Bernt Schiele. B-cos Alignment for Inherently Interpretable CNNs and Vision Transformers. arXiv preprint arXiv:2306.10898, 2023.5 +[8] Moritz Böhle, Mario Fritz, and Bernt Schiele. Convolutional Dynamic Alignment Networks for Interpretable Classifications. In CVPR, pages 10029-10038, 2021. 3 +[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks. In WACV, pages 839-847, 2018. 3 +[10] Hila Chefer, Shir Gur, and Lior Wolf. Transformer Interpretability Beyond Attention Visualization. In CVPR, pages 782-791, 2021. 3 +[11] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing Relevance Maps of Vision Transformers Improves Robustness. In NeurIPS, 2022. 3 +[12] Piotr Dabkowski and Yarin Gal. Real Time Image Saliency for Black Box Classifiers. In NeurIPS, 2017. 3 +[13] Saurabh Desai and Harish G. Ramaswamy. Ablation-CAM: Visual Explanations for Deep Convolutional Network via Gradient-free Localization. In WACV, pages 983-991, 2020. 3 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5 +[15] Gabriel Erion, Joseph D Janizek, Pascal Sturmfels, Scott M Lundberg, and Su-In Lee. Improving Performance of Deep Learning Models with Axiomatic Attribution Priors and Expected Gradients. Nature Machine Intelligence, 3(7):620-631, 2021. 3 +[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. IJCV, 88:303-308, 2009. 2, 5 +[17] Thomas Fel, Ivan F Rodriguez Rodriguez, Drew Linsley, and + +Thomas Serre. Harmonizing the Object Recognition Strategies of Deep Neural Networks with Humans. In NeurIPS, 2022. 3 +[18] Vitaly Feldman and Chiyuan Zhang. What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation. In NeurIPS, pages 2881-2891, 2020. 1 +[19] Patrick Fernandes, Marcos Treviso, Danish Pruthi, André FT Martins, and Graham Neubig. Learning to Scaffold: Optimizing Model Explanations for Teaching. In NeurIPS, 2022. 3 +[20] Ruth C Fong and Andrea Vedaldi. Interpretable Explanations of Black Boxes by Meaningful Perturbation. In ICCV, pages 3429-3437, 2017. 3, 4 +[21] Felix Friedrich, Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. A Typology to Explore and Guide Explanatory Interactive Machine Learning. arXiv preprint arXiv:2203.03668, 2022. 3 +[22] Hiroshi Fukui, Tsubasa Hirakawa, Takayoshi Yamashita, and Hironobu Fujiyoshi. Attention Branch Network: Learning of Attention Mechanism for Visual Explanation. In CVPR, pages 10705-10714, 2019. 3 +[23] Yuyang Gao, Tong Steven Sun, Guangji Bai, Siyi Gu, Sungsoo Ray Hong, and Zhao Liang. RES: A Robust Framework for Guiding Visual Explanation. In KDD, pages 432-442, 2022. 2, 3, 4, 5 +[24] Yuyang Gao, Tong Steven Sun, Liang Zhao, and Sungsoo Ray Hong. Aligning Eyes between Humans and Deep Neural Network through Interactive Attention Alignment. ACM HCI, 6(CSCW2):1-28, 2022. 2, 3, 4, 5 +[25] Hao Guo, Kang Zheng, Xiaochuan Fan, Hongkai Yu, and Song Wang. Visual Attention Consistency under Image Transforms for Multi-Label Image Classification. In CVPR, pages 729–739, 2019. 3 +[26] Misgina Tsighe Hagos, Kathleen M Curran, and Brian Mac Namee. Identifying Spurious Correlations and Correcting them with an Explanation-based Learning. arXiv preprint arXiv:2211.08285, 2022. 3 +[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, pages 770-778, 2016. 5 +[28] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Fast Axiomatic Attribution for Neural Networks. In NeurIPS, pages 19513-19524, 2021. 4, 5 +[29] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely Connected Convolutional Networks. In CVPR, pages 4700-4708, 2017. 5 +[30] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. LayerCAM: Exploring Hierarchical Class Activation Maps for Localization. IEEE TIP, 30:5875-5888, 2021. 3 +[31] Joon Sik Kim, Gregory Plumb, and Ameet Talwalkar. Sanity Simulations for Saliency Methods. In ICML, 2022. 3 +[32] Keisuke Kiritoshi, Ryosuke Tanno, and Tomonori Izumitani. L1-Norm Gradient Penalty for Noise Reduction of Attribution Maps. In CVPRW, pages 118-121, 2019. 3 +[33] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell Me Where to Look: Guided Attention Inference Network. In CVPR, pages 9215-9223, 2018. 3 +[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence + +Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, pages 740-755. Springer, 2014. 2, 5 +[35] Drew Linsley, Dan Shiebler, Sven Eberhardt, and Thomas Serre. Learning what and where to attend. In ICLR, 2019. 3 +[36] Masahiro Mitsuhara, Hiroshi Fukui, Yusuke Sakashita, Takanori Ogata, Tsubasa Hirakawa, Takayoshi Yamashita, and Hironobu Fujiyoshi. Embedding Human Knowledge into Deep Neural Network via Attention Map. In VISIGRAPP, pages 626-636, 2021. 3 +[37] Ofir Moshe, Gil Fidel, Ron Bitton, and Asaf Shabtai. Improving Interpretability via Regularization of Neural Activation Sensitivity. arXiv preprint arXiv:2211.08686, 2022. 3 +[38] Vinod Nair and Geoffrey E Hinton. Rectified Linear Units Improve Restricted Boltzmann Machines. In ICML, pages 807-814, 2010. 4 +[39] Krishna Kanth Nakka and Mathieu Salzmann. Towards Robust Fine-Grained Recognition by Maximal Separation of Discriminative Features. In ACCV, 2020. 3 +[40] Vilfredo Pareto. Il massimo di utilità dato alla libera concorrenza. Giornale degli economisti, pages 48-66, 1894. 5 +[41] Vilfredo Pareto. The Maximum of Utility given by Free Competition. Giornale degli Economisti e Annali di Economia, 67(3):387-403, 2008. 5 +[42] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On Guiding Visual Attention with Language Specification. In CVPR, pages 18092-18102, 2022. 1, 2, 3, 5, 9 +[43] Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized Input Sampling for Explanation of Black-box Models. In BMVC, 2018. 3, 4 +[44] Vipin Pillai, Soroush Abbasi Koohpayegani, Ashley Ouligian, Dennis Fong, and Hamed Piriavash. Consistent Explanations by Contrastive Learning. In CVPR, pages 10213-10222, 2022. 3 +[45] Vipin Pillai and Hamed Piri siavash. Explainable Models with Consistent Interpretations. In AAAI, 2021. 3 +[46] Sukrut Rao, Moritz Böhle, and Bernt Schiele. Towards Better Understanding Attribution Methods. In CVPR, pages 10223-10232, 2022. 3, 4 +[47] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. "Why Should I Trust You?: Explaining the Predictions of Any Classifier. In KDD, pages 1135-1144, 2016. 3, 4 +[48] Laura Rieger, Chandan Singh, William Murdoch, and Bin Yu. Interpretations are Useful: Penalizing Explanations to Align Neural Networks with Prior Knowledge. In ICML, pages 8116-8126, 2020. 3 +[49] Andrew Slavin Ross, Michael C Hughes, and Finale Doshi-Velez. Right for the Right Reasons: Training Differentiable Models by Constraining their Explanations. In *IJCAI*, pages 2662–2670, 2017. 1, 2, 3, 4, 5 +[50] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211–252, 2015. 5 +[51] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization. In ICLR, 2020. 2, 5, 9 +[52] Patrick Schramowski, Wolfgang Stammer, Stefano Teso, + +Anna Brugger, Franziska Herbert, Xiaoting Shao, Hans-Georg Luigs, Anne-Katrin Mahlein, and Kristian Kersting. Making Deep Neural Networks Right for the Right Scientific Reasons by Interacting with their Explanations. Nature Machine Intelligence, 2(8):476-486, 2020. 3 +[53] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. In ICCV, pages 618-626, 2017. 1, 2, 3, 4, 5 +[54] Ramprasaath R Selvaraju, Stefan Lee, Yilin Shen, Hongxia Jin, Shalini Ghosh, Larry Heck, Dhruv Batra, and Devi Parikh. Taking a HINT: Leveraging Explanations to Make Vision and Language Models More Grounded. In ICCV, pages 2591-2600, 2019. 3 +[55] Xiaoting Shao, Arseny Skryagin, Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. Right for Better Reasons: Training Differentiable Models by Constraining their Influence Functions. In AAAI, volume 35, pages 9533-9540, 2021. 3 +[56] Haifeng Shen, Kewen Liao, Zhibin Liao, Job Doornberg, Maoying Qiao, Anton Van Den Hengel, and Johan W Verjans. Human-AI Interactive and Continuous Sensemaking: A Case Study of Image Classification using Scribble Attention Maps. In Extended Abstracts of CHI, pages 1-8, 2021. 2, 3, 4, 5 +[57] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning Important Features Through Propagating Activation Differences. In ICML, pages 3145-3153, 2017. 1, 2, 3, 4, 5 +[58] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps. In ICLRw, 2014. 3 +[59] Krishna Kumar Singh, Dhruv Mahajan, Kristen Grauman, Yong Jae Lee, Matt Feiszli, and Deepti Ghadiyaram. Don't Judge an Object by its Context: Learning to Overcome Contextual Bias. In CVPR, pages 11070-11078, 2020. 3 +[60] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for Simplicity: The All Convolutional Net. In ICLRw, 2015. 3 +[61] Guolei Sun, Salman Khan, Wen Li, Hisham Cholakkal, Fahad Shahbaz Khan, and Luc Van Gool. Fixing Localization Errors to Improve Image Classification. In ECCV, pages 271-287. Springer, 2020. 3 +[62] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic Attribution for Deep Networks. In ICML, pages 3319-3328, 2017. 1, 2, 3, 4, 5 +[63] Damien Teney, Ehsan Abbasnedjad, and Anton van den Hengel. Learning What Makes a Difference from Counterfactual Examples and Gradient Supervision. In ECCV, pages 580-599. Springer, 2020. 3 +[64] Stefano Teso. Toward Faithful Explanatory Active Learning with Self-Explainable Neural Nets. In Workshop on IAL, pages 4–16. CEUR Workshop Proceedings, 2019. 2, 3 +[65] Stefano Teso, Öznur Alkan, Wolfgang Stammer, and Elizabeth Daly. Leveraging Explanations in Interactive Machine Learning: An Overview. Frontiers in Artificial Intelligence, 6:1066049, 2023. 3 +[66] Stefano Teso and Kristian Kersting. Explanatory Interactive Machine Learning. In AIES, pages 239-245, 2019. 2, 3 +[67] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian + +Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks. In CVPRW, pages 111-119, 2020. 2, 3, 4 +[68] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or Signal: The Role of Image Backgrounds in Object Recognition. In *ICLR*, 2021. 1 +[69] Ziyan Yang, Kushal Kafle, Franck Dernoncourt, and Vicente Ordonez. Improving Visual Grounding by Encouraging Consistent Gradient-based Explanations. In CVPR, pages 19165–19174, 2023. 3 +[70] Matthew D Zeiler and Rob Fergus. Visualizing and Understanding Convolutional Networks. In ECCV, pages 818-833, 2014. 3 +[71] Michael Zhang, Nimit Sharad Sohoni, Hongyang R. Zhang, Chelsea Finn, and Christopher Ré. Correct-N-Contrast: a Contrastive Approach for Improving Robustness to Spurious Correlations. In ICML, pages 26484–26516, 2022. 3 +[72] Yilun Zhou, Serena Booth, Marco Tulio Ribeiro, and Julie Shah. Do Feature Attribution Methods Correctly Attribute Features? In AAAI, volume 36, pages 9623-9633, 2022. 3 \ No newline at end of file diff --git a/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/images.zip b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..24356c57f6e444390fcc274c2960bfa7db295e61 --- /dev/null +++ b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b4e9160534167189e5d4942f21773440377707c3c01bf76bf9d8207aac4c70b +size 644820 diff --git a/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/layout.json b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1ade302f0f3b878f9fff58edb483f27e8e03e303 --- /dev/null +++ b/2023/Studying How to Efficiently and Effectively Guide Models with Explanations/layout.json @@ -0,0 +1,10779 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "text", + "content": "Studying How to Efficiently and Effectively Guide Models with Explanations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 144, + 517, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 144, + 517, + 171 + ], + "spans": [ + { + "bbox": [ + 75, + 144, + 517, + 171 + ], + "type": "text", + "content": "Sukrut Rao*, Moritz Böhle*, Amin Parchami-Araghi, Bernt Schiele \nMax Planck Institute for Informatics, Saarland Informatics Campus, Saarbrücken, Germany" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 173, + 436, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 173, + 436, + 185 + ], + "spans": [ + { + "bbox": [ + 156, + 173, + 436, + 185 + ], + "type": "text", + "content": "{sukrut.rao,mboehle,mparcham,schiele}@mpi-inf.mpg.de" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 597 + ], + "type": "text", + "content": "Despite being highly performant, deep neural networks might base their decisions on features that spuriously correlate with the provided labels, thus hurting generalization. To mitigate this, 'model guidance' has recently gained popularity, i.e. the idea of regularizing the models' explanations to ensure that they are \"right for the right reasons\" [49]. While various techniques to achieve such model guidance have been proposed, experimental validation of these approaches has thus far been limited to relatively simple and / or synthetic datasets. To better understand the effectiveness of the various design choices that have been explored in the context of model guidance, in this work we conduct an in-depth evaluation across various loss functions, attribution methods, models, and 'guidance depths' on the PASCAL VOC 2007 and MS COCO 2014 datasets. As annotation costs for model guidance can limit its applicability, we also place a particular focus on efficiency. Specifically, we guide the models via bounding box annotations, which are much cheaper to obtain than the commonly used segmentation masks, and evaluate the robustness of model guidance under limited (e.g. with only " + }, + { + "bbox": [ + 46, + 238, + 290, + 597 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 597 + ], + "type": "text", + "content": " of annotated images) or overly coarse annotations. Further, we propose using the EPG score as an additional evaluation metric and loss function ('Energy loss'). We show that optimizing for the Energy loss leads to models that exhibit a distinct focus on object-specific features, despite only using bounding box annotations that also include background regions. Lastly, we show that such model guidance can improve generalization under distribution shifts. Code available at: https://github.com/sukrutrao/Model-Guidance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 613, + 128, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 128, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 128, + 625 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 634, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 634, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 634, + 287, + 694 + ], + "type": "text", + "content": "Deep neural networks (DNNs) excel at learning predictive features that allow them to correctly classify a set of training images with ease. The features learnt on the training set, however, do not necessarily transfer to unseen images: i.e., instead of learning the actual class-relevant fea" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 213, + 532, + 327 + ], + "blocks": [ + { + "bbox": [ + 318, + 213, + 532, + 327 + ], + "lines": [ + { + "bbox": [ + 318, + 213, + 532, + 327 + ], + "spans": [ + { + "bbox": [ + 318, + 213, + 532, + 327 + ], + "type": "image", + "image_path": "06af46ab131b75d0fd79d635858da8b202be74af1806d7c84ba16998e820671e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 338, + 352, + 389 + ], + "blocks": [ + { + "bbox": [ + 320, + 338, + 352, + 389 + ], + "lines": [ + { + "bbox": [ + 320, + 338, + 352, + 389 + ], + "spans": [ + { + "bbox": [ + 320, + 338, + 352, + 389 + ], + "type": "image", + "image_path": "46ae3bb8b102d1418ad1f9744ad1bebe4dd1ffcf4c52f20827d5ed7e307137f8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 403, + 545, + 535 + ], + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 535 + ], + "type": "text", + "content": "Fig. 1: (a) Model guidance increases object focus. Models may rely on irrelevant background features or spurious correlations (e.g. presence of person provides positive evidence for bicycle, center row, col. 1). Guiding the model via bounding box annotations can mitigate this and consistently increases the focus on object features (bottom row). (b) Model guidance can improve accuracy. In the presence of spurious correlations in the training data, non-guided models might focus on the wrong features. In the example image in (b), the waterbird is incorrectly classified to be a landbird due to the background (col. 3). Guiding the model via bounding box annotation (as shown in col. 2), the model can be guided to focus on the bird features for classification (col. 4)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 354, + 339, + 397, + 389 + ], + "blocks": [ + { + "bbox": [ + 354, + 339, + 397, + 389 + ], + "lines": [ + { + "bbox": [ + 354, + 339, + 397, + 389 + ], + "spans": [ + { + "bbox": [ + 354, + 339, + 397, + 389 + ], + "type": "image", + "image_path": "b9d6a75277e78fcfdc7194a485eede0fb298b87da971f1522c1d156a35ac17eb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 402, + 331, + 445, + 389 + ], + "blocks": [ + { + "bbox": [ + 402, + 331, + 445, + 389 + ], + "lines": [ + { + "bbox": [ + 402, + 331, + 445, + 389 + ], + "spans": [ + { + "bbox": [ + 402, + 331, + 445, + 389 + ], + "type": "image", + "image_path": "c76c9e34f23c307f6b917e14d755c7c4d169d55763d83ba0c6aeac1c91e3fdd7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 451, + 331, + 494, + 389 + ], + "blocks": [ + { + "bbox": [ + 451, + 331, + 494, + 389 + ], + "lines": [ + { + "bbox": [ + 451, + 331, + 494, + 389 + ], + "spans": [ + { + "bbox": [ + 451, + 331, + 494, + 389 + ], + "type": "image", + "image_path": "43a4783bb4bbe5f99ae90e6e5b08395ed5cdd3d143d363a78bee7e472b4c3797.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 389, + 493, + 399 + ], + "lines": [ + { + "bbox": [ + 451, + 389, + 493, + 399 + ], + "spans": [ + { + "bbox": [ + 451, + 389, + 493, + 399 + ], + "type": "text", + "content": "Landbird Confidence: " + }, + { + "bbox": [ + 451, + 389, + 493, + 399 + ], + "type": "inline_equation", + "content": "97\\%" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 498, + 331, + 542, + 388 + ], + "blocks": [ + { + "bbox": [ + 498, + 331, + 542, + 388 + ], + "lines": [ + { + "bbox": [ + 498, + 331, + 542, + 388 + ], + "spans": [ + { + "bbox": [ + 498, + 331, + 542, + 388 + ], + "type": "image", + "image_path": "f26565be58284a21fe4aa0090c377bd3239e308462cce93e1d4617623e1213d6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 389, + 541, + 399 + ], + "lines": [ + { + "bbox": [ + 499, + 389, + 541, + 399 + ], + "spans": [ + { + "bbox": [ + 499, + 389, + 541, + 399 + ], + "type": "text", + "content": "Waterbird Confidence: " + }, + { + "bbox": [ + 499, + 389, + 541, + 399 + ], + "type": "inline_equation", + "content": "87\\%" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 540, + 546, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 546, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 546, + 649 + ], + "type": "text", + "content": "tures, DNNs might memorize individual images (cf. [18]) or exploit spurious correlations in the training data (cf. [68]). For example, if bikes are highly correlated with people in the training data, a model might learn to associate the presence of a person in an image as positive evidence for a bike (e.g. Fig. 1a, col. 1, rows 1-2), which can limit how well it generalizes. Similarly, a bird classifier might rely on background features from the bird's habitat, and fail to correctly classify in a different habitat (cf. Fig. 1b cols. 1-3 and [42])." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "To detect such behaviour, recent advances in model interpretability have provided attribution methods (e.g. [53, 62, 57, 6]) to understand a model's reasoning. These methods typically provide attention maps that highlight regions of importance in an input to explain the model's decisions" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1922" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 533, + 242 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 533, + 242 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 533, + 242 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 533, + 242 + ], + "type": "image", + "image_path": "de598a16f0e7a21eb615b23f22952d798467a3441debf5c3d2cc06b5552a47e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 244, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 244, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 547, + 289 + ], + "type": "text", + "content": "Fig. 2: Qualitative results of model guidance. We show model-inherent B-cos explanations (input layer) of a B-cos ResNet-50 and GradCAM explanations (final layer) of a conventional ResNet-50 before ('Standard') and after optimization ('Guided') for images from the VOC test set, using our proposed Energy loss (Eq. (6)). Guiding the model via bounding box annotations consistently increases the focus on object features for both methods. Specifically, we find that background attributions are consistently suppressed in both cases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 302, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 326 + ], + "type": "text", + "content": "and can help identify incorrect reasoning such as reliance on spurious or irrelevant features, see for example Fig. 1b." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 326, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 457 + ], + "type": "text", + "content": "As many attribution methods are in fact themselves differentiable (e.g. [57, 62, 53, 6]), recent work [49, 56, 24, 23, 66, 64] has explored the idea of using them to guide the models to make them \"right for the right reasons\" [49]. Specifically, models can be guided by jointly optimizing for correct classification as well as for attributing importance to regions deemed relevant by humans. This can help the model focus on the relevant features of a class, and correct errors in reasoning (Fig. 1b, col. 4). Such guidance has the added benefit of providing well-localized explanations that are thus easier to understand for end users (e.g. Fig. 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 458, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 541 + ], + "type": "text", + "content": "While model guidance has shown promising results, a detailed study of how to do this most effectively is crucially missing. In particular, model guidance has so far been studied for a limited set of attribution methods and models and usually on relatively simple and/or synthetic datasets; further, the evaluation settings between approaches can significantly differ, which makes a fair comparison difficult." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 542, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 661 + ], + "type": "text", + "content": "Therefore, in this work, we perform an in-depth evaluation of model guidance on large scale, real-world datasets, to better understand the effectiveness of a variety of design choices. Specifically, we evaluate model guidance along the following dimensions: the model architecture, the guidance depth1, the attribution method, and the loss function. In this context, we propose using the EPG score [67]—an evaluation metric that has thus far been used to evaluate the quality of attribution methods—as an additional loss function (which we call the Energy loss) as it is fully differentiable." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 662, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 287, + 685 + ], + "type": "text", + "content": "Further, as annotation costs can be a major hurdle for making model guidance practical, we place a particular fo" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 302, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 302, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 302, + 545, + 350 + ], + "type": "text", + "content": "cus on efficient guidance. Specifically, we use bounding boxes instead of semantic segmentation masks, and evaluate the robustness of guidance techniques under limited or overly coarse annotations to reduce data collection costs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 352, + 546, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 546, + 472 + ], + "type": "text", + "content": "We find that our Energy loss lends itself well to those settings. On the one hand, it exhibits a high degree of robustness to limited or noisy bounding box annotations (cf. Figs. 10 and 12). On the other hand, despite the coarseness of bounding box guidance, it maintains a clear focus on object-specific features inside the bounding boxes, see Fig. 1a, row 3. In contrast, prior approaches often regularize for a uniform distribution of the attribution values inside the annotation masks, and thus tend to exhibit much lower attribution granularity (cf. Fig. 9)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": "Contributions. (1) We perform an in-depth evaluation of model guidance on challenging large scale, multi-label classification datasets (PASCAL VOC 2007 [16], MS COCO 2014 [34]), assessing the impact of attribution methods, model architectures, guidance depths, and loss functions. Further, we show that, despite being relatively coarse, bounding box supervision can provide sufficient guidance to the models whilst being much cheaper to obtain than semantic segmentation masks. (2) We propose using the Energy Pointing Game (EPG) score [67] as an alternative to the IoU metric for evaluating the effectiveness of such guidance and show that the EPG score constitutes a good loss function for model guidance, particularly when using bounding boxes. (3) We show that model guidance can be performed cost-effectively by using annotation masks that are noisy or are available for only a small fraction (e.g. " + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": ") of the training data. (4) We show through experiments on the Waterbirds-100 dataset [51, 42] that model guidance with a small number of annotations suffices to improve the model's generalization under distribution shifts at test time." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "1The layer at which guidance is applied, e.g. typically at the last convolutional layer for GradCAM [53] or the first layer for IxG [57]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1923" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 209 + ], + "type": "text", + "content": "Attribution Methods [58, 60, 62, 57, 53, 67, 13, 30, 9, 43, 20, 70, 47, 12, 4] are often used to explain black-box models by generating heatmaps that highlight input regions important to the model's decision. However, such methods are often not faithful to the model [1, 46, 31, 72, 2] and risk misleading users. Recent work proposes inherently interpretable models [8, 6] that address this by providing modelfaithful explanations by design. In our work, we use both popular post-hoc and model-inherent attribution methods to guide models and discuss their effectiveness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 211, + 288, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 288, + 330 + ], + "type": "text", + "content": "Attribution Priors: Several approaches have been proposed for training better models by enforcing desirable properties on their attributions. These include enforcing consistency against augmentations [45, 44, 25], smoothness [15, 37, 32], separation of classes [71, 44, 61, 39, 59], or constraining the model's attention [22, 3]. In contrast, in this work, we focus on providing explicit human guidance to the model using bounding box annotations. This constitutes more explicit guidance but allows fine-grained control over the model's reasoning even with few annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 330, + 288, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 288, + 558 + ], + "type": "text", + "content": "Model Guidance: In contrast to the indirect regularization effect achieved by attribution priors, various approaches have been proposed (cf. [21, 65]) to actively guide models by regularizing their attributions, for tasks such as classification [49, 24, 23, 48, 42, 26, 63, 36, 66, 64, 52, 55, 35, 56, 69, 17], segmentation [33], VQA [54, 63], and knowledge distillation [19]. The goal of such approaches is not only to improve performance, but also make sure that the model is \"right for the right reasons\" [49]. For classifiers, this typically involves jointly optimizing both for classification performance and localization to object features. While various benefits of model guidance have been reported, most prior work evaluate on simple datasets [49, 55, 24, 23] and, thus far, no common evaluation setting has emerged. Recently, [11] has extended model guidance to ImageNet, showing that its benefits can scale to large scale problems. In contrast to [11], who investigated one particular attribution method [10], our focus lies on a better understanding of the impact of the different design choices for model guidance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "To distill the most effective techniques for model guidance, in this work, we conduct an in-depth evaluation on challenging, commonly used real-world multi-label classification datasets (PASCAL VOC 2007, MS COCO 2014). Specifically, we perform a comprehensive comparison across multiple dimensions of interest: the loss function, the model architecture, the guidance depth, and the attribution method. For this, we evaluate the localization losses introduced in the closest related work, i.e. RRR [49], HAICS [56], and GRADIA [24]; additionally, we propose using the EPG metric [67] as a loss function and show that it has various desirable properties, in particular when guiding models via bounding box annotations." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 305, + 69, + 547, + 151 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 547, + 151 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 547, + 151 + ], + "type": "image", + "image_path": "0b0d9ba08868ff0b01f0a00e2862efa1810cab23cbb1ded746e94d3dcf66b357.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "type": "text", + "content": "Fig. 3: Model guidance overview. We jointly optimize for classification " + }, + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{class}})" + }, + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "type": "text", + "content": " and localization of attributions to human-annotated bounding boxes " + }, + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{loc}})" + }, + { + "bbox": [ + 304, + 152, + 545, + 196 + ], + "type": "text", + "content": ", to guide the model to focus on object features. Various localization loss functions can be used, see Sec. 3.4." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 202, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 251 + ], + "type": "text", + "content": "Finally, model guidance has also been used to mitigate reliance on spurious features using language guidance [42], and we show that using a small number of coarse bounding box annotations can be similarly effective." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 251, + 546, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 546, + 370 + ], + "type": "text", + "content": "Evaluating Model Guidance: The benefits of model guidance have typically been shown via improvements in classification performance (e.g. [49, 48]) or an increase in IoU between object masks and attribution maps (e.g. [23, 33]). In addition to these metrics, we also evaluate on the EPG metric [67], which has thus far only been used to evaluate the quality of the attribution methods themselves. We further show that it lends itself well to being used as a guidance loss, as it places only minor constraints on the model, and, in contrast to the IoU metric, it is fully differentiable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 379, + 501, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 379, + 501, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 501, + 392 + ], + "type": "text", + "content": "3. Guiding Models Using Attributions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 399, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 483 + ], + "type": "text", + "content": "In this section, we provide an overview of the model guidance approach that jointly optimizes for classification and localization (Sec. 3.1). Specifically, we describe the attribution methods (Sec. 3.2), metrics (Sec. 3.3), and localization loss functions (Sec. 3.4) that we evaluate in Sec. 5. In Sec. 3.5 we discuss our strategy to train for localization in the presence of multiple ground truth classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": "Notation: We consider a multi-label classification problem with " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " classes with " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " the input image and " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}^K" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " the one-hot encoding of the image labels. With " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "A_k \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " we denote an attribution map for a class " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " using a classifier " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "A_k^+" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " denotes the positive component of the attributions, " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\hat{A}_k = \\frac{A_k}{\\max(\\mathrm{abs}(A_k))}" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " normalized attributions, and " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\hat{A}_k^+ = \\frac{A_k^+}{\\max(A_k^+)}" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " normalized positive attributions. Finally, " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "M_k \\in \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " denotes the binary mask for class " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": ", which is given by the union of bounding boxes of all occurrences of class " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 620, + 457, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 457, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 457, + 632 + ], + "type": "text", + "content": "3.1. Model Guidance Procedure" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 639, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 639, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 305, + 639, + 545, + 663 + ], + "type": "text", + "content": "Following prior work (e.g. [49, 56, 24, 23]), the model is trained jointly for classification and localization (cf. Fig. 3):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 379, + 673, + 545, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 673, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 379, + 673, + 545, + 687 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c l a s s}} + \\lambda_ {\\text {l o c}} \\mathcal {L} _ {\\text {l o c}}. \\tag {1}", + "image_path": "da9fe893c4ad292fc262a7bafdf92abad1252deabad7fb81905b93e87ac6e0fe.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "I.e., the loss consists of a classification loss " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{class}})" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": ", for which we use binary cross-entropy, and a localization loss" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1924" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{loc}})" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": ", which we discuss in Sec. 3.4; here, the hyperparameter " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{loc}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " controls the weight given to each of the objectives." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 106, + 167, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 167, + 118 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 167, + 118 + ], + "type": "text", + "content": "3.2. Attribution Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 125, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 125, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 287, + 232 + ], + "type": "text", + "content": "In contrast to prior work that typically use GradCAM [53] attributions, we perform an evaluation over a selection of popularly used differentiable" + }, + { + "bbox": [ + 46, + 125, + 287, + 232 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 125, + 287, + 232 + ], + "type": "text", + "content": " attribution methods which have been shown to localize well [46]: IxG [57], IntGrad [62], and GradCAM [53]. We further evaluate model-inherent explanations of the recently proposed B-cos models [6]. To ensure comparability across attribution methods [46], we evaluate all attribution methods at the input, various intermediate, and the final spatial layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 233, + 287, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 292 + ], + "type": "text", + "content": "IxG [57] computes the element-wise product " + }, + { + "bbox": [ + 46, + 233, + 287, + 292 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 233, + 287, + 292 + ], + "type": "text", + "content": " of the input and the gradients of the " + }, + { + "bbox": [ + 46, + 233, + 287, + 292 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 233, + 287, + 292 + ], + "type": "text", + "content": "-th output w.r.t. the input, i.e. " + }, + { + "bbox": [ + 46, + 233, + 287, + 292 + ], + "type": "inline_equation", + "content": "X\\odot \\nabla_{X}f_{k}(X)" + }, + { + "bbox": [ + 46, + 233, + 287, + 292 + ], + "type": "text", + "content": ". For piece-wise linear models such as DNNs with ReLU activations [38], this faithfully computes the linear contributions of a given input pixel to the model output." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "content": "GradCAM [53] computes importance attributions as a ReLU-thresholded, gradient-weighted sum of activation maps. In detail, it is given by " + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\mathrm{ReLU}(\\sum_{c}\\alpha_{c}^{k}\\odot U_{c})" + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "content": " denoting the channel dimension, and " + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\alpha^k" + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "content": " the average-pooled gradients of the output for class " + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "content": " with respect to the activations " + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 293, + 287, + 365 + ], + "type": "text", + "content": " of the last convolutional layer in the model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "type": "text", + "content": "IntGrad [62] takes an axiomatic approach and is formulated as the integral of gradients over a straight line path from a baseline input to the given input " + }, + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "type": "text", + "content": ". Approximating this integral requires several gradient computations, making it computationally expensive for use in model guidance. To alleviate this, when optimizing with IntGrad, we use the recently proposed " + }, + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 366, + 287, + 461 + ], + "type": "text", + "content": "-DNN models [28] that allow for an exact computation of IntGrad in a single backward pass." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "type": "text", + "content": "B-cos [6] attributions are generated using the inherently-interpretable B-cos networks, which promote alignment between the input " + }, + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "type": "text", + "content": " and a dynamic weight matrix " + }, + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{W}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "type": "text", + "content": " during optimization. In our experiments, we use the contribution maps given by the element-wise product of the dynamic weights with the input " + }, + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}_k^T (\\mathbf{x})\\odot \\mathbf{x})" + }, + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "type": "text", + "content": ", which faithfully represent the contribution of each pixel to class " + }, + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 462, + 287, + 570 + ], + "type": "text", + "content": ". To be able to guide B-cos models, we developed a differentiable implementation of B-cos explanations, see supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 579, + 160, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 160, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 160, + 591 + ], + "type": "text", + "content": "3.3. Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 598, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 598, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 598, + 287, + 646 + ], + "type": "text", + "content": "We evaluate the models' performance on both our training objectives: classification and localization. For classification, we use the F1 score and mean average precision (mAP). We discuss the localization metrics below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 647, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 683 + ], + "type": "text", + "content": "Intersection over Union (IoU) is a commonly used metric (cf. [23]) that computes the intersection between the ground truth annotation masks and the binarized attribution maps," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 131 + ], + "type": "text", + "content": "normalized by their union; for binarization, a threshold parameter needs to be chosen. In this work, the ground truth masks are taken to be the union of all bounding boxes of a class in the image and, following prior work [20], the threshold parameter is selected via a heldout set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 133, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 133, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 545, + 168 + ], + "type": "text", + "content": "Energy-based Pointing Game (EPG) [67] measures the concentration of attribution energy within the mask, i.e. the fraction of positive attributions inside the bounding boxes:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 174, + 545, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 174, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 347, + 174, + 545, + 208 + ], + "type": "interline_equation", + "content": "\\mathrm {E P G} _ {k} = \\frac {\\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} M _ {k , h w} A _ {k , h w} ^ {+}}{\\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} A _ {k , h w} ^ {+}}. \\tag {2}", + "image_path": "0cd12c479b9d656ea4b95e6c542a0cba8fcb13f2d8c56f64d2cacebfe31865d6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 213, + 545, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 261 + ], + "type": "text", + "content": "In contrast to IoU, EPG more faithfully takes into account the relative importance given to each input region, since it does not binarize the attributions. Like IoU, the scores lie in [0, 1], with higher scores indicating better localization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 267, + 419, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 267, + 419, + 278 + ], + "spans": [ + { + "bbox": [ + 306, + 267, + 419, + 278 + ], + "type": "text", + "content": "3.4. Localization Losses" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 285, + 545, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 333 + ], + "type": "text", + "content": "We evaluate the most commonly used localization losses " + }, + { + "bbox": [ + 305, + 285, + 545, + 333 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{loc}}" + }, + { + "bbox": [ + 305, + 285, + 545, + 333 + ], + "type": "text", + "content": " in Eq. (1)) from prior work. We describe these losses as applied on attribution maps of an image for a single class " + }, + { + "bbox": [ + 305, + 285, + 545, + 333 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 285, + 545, + 333 + ], + "type": "text", + "content": ", as well as the proposed EPG-derived Energy loss." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 334, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 545, + 380 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 306, + 334, + 545, + 380 + ], + "type": "text", + "content": " loss ([24, 23], Eq. (3)) minimizes the " + }, + { + "bbox": [ + 306, + 334, + 545, + 380 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 306, + 334, + 545, + 380 + ], + "type": "text", + "content": " distance between annotation masks and normalized positive attributions " + }, + { + "bbox": [ + 306, + 334, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\hat{A}_k^+" + }, + { + "bbox": [ + 306, + 334, + 545, + 380 + ], + "type": "text", + "content": ", guiding the model towards uniform attributions inside the mask and suppressing attributions outside of it." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 386, + 545, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 386, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 320, + 386, + 545, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {l o c}, k} = \\frac {1}{H \\times W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\| M _ {k, h w} - \\hat {A} _ {k, h w} ^ {+} \\| _ {1} \\tag {3}", + "image_path": "ff2d1df3cea69d121b85b22982b7ad6944a254b1741d6ae311ba54057e2238ef.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 408, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 408, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 545, + 455 + ], + "type": "text", + "content": "Per-pixel cross entropy (PPCE) loss ([56], Eq. (4)) applies a binary cross entropy loss between the mask and the normalized positive annotations " + }, + { + "bbox": [ + 305, + 408, + 545, + 455 + ], + "type": "inline_equation", + "content": "\\hat{A}_k^+" + }, + { + "bbox": [ + 305, + 408, + 545, + 455 + ], + "type": "text", + "content": ", thus guiding the model to maximize the attributions inside the mask:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 460, + 545, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 545, + 478 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l o c}, k} = - \\frac {1}{\\| M _ {k} \\| _ {1}} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} M _ {k, h w} \\log \\left(\\hat {A} _ {k, h w} ^ {+}\\right). \\tag {4}", + "image_path": "2661a0412e20f1e8e1af094b6cd413629afb50e29e3d1f241eb04e5337ceffaf.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 483, + 544, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 544, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 544, + 506 + ], + "type": "text", + "content": "As PPCE does not constrain attributions outside the mask, there is no explicit pressure to avoid spurious features." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 507, + 544, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 507, + 544, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 507, + 544, + 532 + ], + "type": "text", + "content": "RRR\\* loss ([49], Eq. (5)). [49] introduced the RRR loss to regularize the normalized input gradients " + }, + { + "bbox": [ + 306, + 507, + 544, + 532 + ], + "type": "inline_equation", + "content": "\\hat{A}_{k,hw}" + }, + { + "bbox": [ + 306, + 507, + 544, + 532 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 337, + 538, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 538, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 337, + 538, + 545, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {l o c}, k} = \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\left(1 - M _ {k, h w}\\right) \\hat {A} _ {k, h w} ^ {2}. \\tag {5}", + "image_path": "73927c96024ea2d53adfa6682de5a436404317679b1fe812dd399fd86138504f.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 562, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 545, + 645 + ], + "type": "text", + "content": "To extend it to our setting, we take " + }, + { + "bbox": [ + 305, + 562, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\hat{A}_{k,hw}" + }, + { + "bbox": [ + 305, + 562, + 545, + 645 + ], + "type": "text", + "content": " to be given by an arbitrary attribution method (e.g. IntGrad); we denote this generalized version by RRR*. In contrast to the PPCE loss, RRR* only regularizes attributions outside the ground truth masks. While it thus does not introduce a uniformity prior similar to the " + }, + { + "bbox": [ + 305, + 562, + 545, + 645 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 305, + 562, + 545, + 645 + ], + "type": "text", + "content": " loss, it also does not explicitly promote high importance attributions inside the masks." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 647, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 545, + 693 + ], + "type": "text", + "content": "Energy Loss. In addition to the losses described in prior work, we propose to also evaluate using the EPG score ([67], Eq. (2)) as a loss function for model guidance, as it is fully differentiable. In particular, we simply define it as" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 389, + 701, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 389, + 701, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l o c}, k} = - \\operatorname {E P G} _ {k}. \\tag {6}", + "image_path": "fa0cb43236ddc28fe4041a9131bd3bd6543b6adeb71e447fd068ea39ca060707.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "2. Differentiability is necessary for optimizing attributions via gradient descent, so non-differentiable methods (e.g. [47, 43]) are not considered." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1925" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 44, + 71, + 286, + 144 + ], + "blocks": [ + { + "bbox": [ + 44, + 71, + 286, + 144 + ], + "lines": [ + { + "bbox": [ + 44, + 71, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 44, + 71, + 286, + 144 + ], + "type": "image", + "image_path": "7c1905b038578c97d9c819dc9c64c3fced972e9568c5b314f49e47545c77d741.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 148, + 287, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 224 + ], + "type": "text", + "content": "Fig. 4: Selecting models for evaluation. For each configuration, we evaluate every model at every checkpoint and measure its performance across various metrics (F1, EPG, IoU) on the validation set; i.e. every point in the left graph corresponds to one model (for B-cos models optimized via the Energy loss at the input layer). Instead of evaluating a single model on the test set, we evaluate all Pareto-dominant models, as indicated in the center and right plot." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 230, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 287, + 350 + ], + "type": "text", + "content": "Unlike existing localization losses that either (i) do not constrain attributions across the entire input " + }, + { + "bbox": [ + 46, + 230, + 287, + 350 + ], + "type": "inline_equation", + "content": "(\\mathsf{RRR}^*,\\mathsf{PPCE})" + }, + { + "bbox": [ + 46, + 230, + 287, + 350 + ], + "type": "text", + "content": " or (ii) force the model to attribute uniformly within the mask even if it includes irrelevant background regions " + }, + { + "bbox": [ + 46, + 230, + 287, + 350 + ], + "type": "inline_equation", + "content": "(L_{1}," + }, + { + "bbox": [ + 46, + 230, + 287, + 350 + ], + "type": "text", + "content": " PPCE), maximizing the EPG score jointly optimizes for higher attribution energy within the mask and lower attribution energy outside the mask. By not enforcing a uniformity prior, we find that the Energy loss is able to provide effective guidance while allowing the model to learn freely what to focus on within the bounding boxes (Sec. 5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 360, + 174, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 174, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 174, + 372 + ], + "type": "text", + "content": "3.5. Efficient Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 379, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 511 + ], + "type": "text", + "content": "In contrast to prior work [49, 56, 24, 23], we perform model guidance on a multi-label classification setting, and consequently there are multiple ground truth classes whose attribution localization could be optimized. Computing and optimizing for several attributions within an image would add a significant overhead to the computational cost of training (multiple backward passes). Hence, for efficiency, we sample one ground truth class " + }, + { + "bbox": [ + 46, + 379, + 287, + 511 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 379, + 287, + 511 + ], + "type": "text", + "content": " per image at random for every batch and only optimize for localization of that class, i.e., " + }, + { + "bbox": [ + 46, + 379, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{loc}} = \\mathcal{L}_{\\mathrm{loc},k}" + }, + { + "bbox": [ + 46, + 379, + 287, + 511 + ], + "type": "text", + "content": ". We find that this still provides effective model guidance while keeping the training cost tractable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 524, + 165, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 165, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 165, + 538 + ], + "type": "text", + "content": "4. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": "In this section, we describe our experimental setup and how we select the best models across metrics; for full details, see supplement. We evaluate across all possible choices for each category, and discuss our results in Sec. 5. Datasets: We evaluate on PASCAL VOC 2007 [16] and MS COCO 2014 [34] for multi-label image classification. In Sec. 5.5, to understand the effectiveness of model guidance in mitigating spurious correlations, we also evaluate on the synthetically constructed Waterbirds-100 dataset [51, 42], where landbirds are perfectly correlated with land backgrounds on the training and validation sets, but are equally likely to occur on land or water in the test set (similar for waterbirds and water). With this dataset, we evaluate model guidance for suppressing undesired features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "Attribution Methods and Architectures: As described in Sec. 3.2, we evaluate with IxG [57], IntGrad [62], B-cos [6, 7], and GradCAM [53] using models with a ResNet-50 [27] backbone. For IntGrad, we use an " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "-DNN ResNet-50 [28] to reduce the computational cost, and a B-cos ResNet-50 for the B-cos attributions. To emphasize that the results generalize across different backbones, we further provide results for a B-cos ViT-S [14, 7] and a B-cos DenseNet-121 [29, 7]. We evaluate optimizing the attributions at different network layers, such as at the input image and the last convolutional layers' output3, as well as at multiple intermediate layers. Within the main paper, we highlight some of the most representative and insightful results, the full set of results can be found in the supplement. All models were pretrained on ImageNet [50], and model guidance was applied when fine-tuning the models on the target dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 264, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 300 + ], + "type": "text", + "content": "Localization Losses: As described in Sec. 3.4, we compare four localization losses in our evaluation: (i) Energy, (ii) " + }, + { + "bbox": [ + 304, + 264, + 545, + 300 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 264, + 545, + 300 + ], + "type": "text", + "content": " [24, 23], (iii) PPCE [56], and (iv) RRR* (cf. Sec. 3.4, [49])." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 301, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 360 + ], + "type": "text", + "content": "Evaluation Metrics: As discussed in Sec. 3.3, we evaluate both for classification and localization performance of the models. For classification, we report the F1 scores, similar results with mAP scores can be found in the supplement. For localization, we evaluate using the EPG and IoU scores." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 361, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 576 + ], + "type": "text", + "content": "Selecting the best models: As we evaluate for two distinct objectives (classification + localization), it is not trivial to decide which models perform 'the best', e.g. a model that provides the best classification performance might provide significantly worse localization than a model that provides only slightly lower classification performance. Finding the right balance and deciding which of those models in fact constitutes the 'better' model depends on the preference of the end user. Hence, instead of selecting models based on a single metric, we select the set of Pareto-dominant models [40, 41, 5] across three metrics—F1, EPG, and IoU—for each training configuration, as defined by a combination of attribution method, layer, and loss. Specifically, as shown in Fig. 4, we train each configuration using three different choices of " + }, + { + "bbox": [ + 304, + 361, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{loc}}" + }, + { + "bbox": [ + 304, + 361, + 545, + 576 + ], + "type": "text", + "content": ", and select the set of Pareto-dominant models among all checkpoints (epochs and " + }, + { + "bbox": [ + 304, + 361, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{loc}}" + }, + { + "bbox": [ + 304, + 361, + 545, + 576 + ], + "type": "text", + "content": "). This provides a more holistic view of the general trends on the effectiveness of model guidance for each configuration." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 588, + 432, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 588, + 432, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 588, + 432, + 601 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 609, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 545, + 693 + ], + "type": "text", + "content": "In this section, we discuss our experimental findings. In particular, in Sec. 5.1, we first discuss the impact of the loss functions on the EPG and IoU scores of the models; in Sec. 5.2, we then analyze the impact of the models and attribution methods; further in Sec. 5.3, we show that guiding the models via their explanations can lead to improved classification accuracy. In Sec. 5.4, we present additional" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 537, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 537, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 537, + 713 + ], + "type": "text", + "content": "3As typically used in IxG (input) and GradCAM (final) respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1926" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 237, + 205 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 237, + 205 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 237, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 237, + 205 + ], + "type": "image", + "image_path": "764f59edc7188ec75dcec414c7552a78107ec05f498b12eedbbbb5130475fa2c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 212, + 368, + 220 + ], + "lines": [ + { + "bbox": [ + 226, + 212, + 368, + 220 + ], + "spans": [ + { + "bbox": [ + 226, + 212, + 368, + 220 + ], + "type": "text", + "content": "(a) PASCAL VOC results for EPG vs. F1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 241, + 72, + 384, + 205 + ], + "blocks": [ + { + "bbox": [ + 241, + 72, + 384, + 205 + ], + "lines": [ + { + "bbox": [ + 241, + 72, + 384, + 205 + ], + "spans": [ + { + "bbox": [ + 241, + 72, + 384, + 205 + ], + "type": "image", + "image_path": "4aca215a1dcdf93ee79856e08a888240bcaae524424608497686c3e34aca2b8f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 386, + 72, + 531, + 205 + ], + "blocks": [ + { + "bbox": [ + 386, + 72, + 531, + 205 + ], + "lines": [ + { + "bbox": [ + 386, + 72, + 531, + 205 + ], + "spans": [ + { + "bbox": [ + 386, + 72, + 531, + 205 + ], + "type": "image", + "image_path": "a7d99e6d0282dec595c3f093439d3f16fc412916ea77ca7651947db7d9e5d30b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 226, + 237, + 358 + ], + "blocks": [ + { + "bbox": [ + 62, + 226, + 237, + 358 + ], + "lines": [ + { + "bbox": [ + 62, + 226, + 237, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 226, + 237, + 358 + ], + "type": "image", + "image_path": "98957feb031bab97836e5557cf46b9bfda752eac90e54fe772656344f125f10d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 366, + 361, + 375 + ], + "lines": [ + { + "bbox": [ + 231, + 366, + 361, + 375 + ], + "spans": [ + { + "bbox": [ + 231, + 366, + 361, + 375 + ], + "type": "text", + "content": "(b) MS COCO results for EPG vs. F1." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 241, + 226, + 384, + 358 + ], + "blocks": [ + { + "bbox": [ + 241, + 226, + 384, + 358 + ], + "lines": [ + { + "bbox": [ + 241, + 226, + 384, + 358 + ], + "spans": [ + { + "bbox": [ + 241, + 226, + 384, + 358 + ], + "type": "image", + "image_path": "bbcccf4741a0ea918ecf24ff859dc2edefc52dc5869a87b289281a21bb5d03be.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 386, + 227, + 531, + 358 + ], + "blocks": [ + { + "bbox": [ + 386, + 227, + 531, + 358 + ], + "lines": [ + { + "bbox": [ + 386, + 227, + 531, + 358 + ], + "spans": [ + { + "bbox": [ + 386, + 227, + 531, + 358 + ], + "type": "image", + "image_path": "30a1afe1bf695ecd1a3cf233243c3111e88bc837e5321126b6c31ced676b4e29.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 62, + 433, + 237, + 518 + ], + "blocks": [ + { + "bbox": [ + 45, + 380, + 546, + 426 + ], + "lines": [ + { + "bbox": [ + 45, + 380, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 45, + 380, + 546, + 426 + ], + "type": "text", + "content": "Fig. 5: EPG vs. F1, for different datasets ((a): VOC; (b): COCO), losses (markers) and models (columns), optimized at different layers (rows); additionally, we show the performance of the baseline model before fine-tuning and demarcate regions that strictly dominate (are strictly dominated by) the baseline performance in green (grey). For each configuration, we show the Pareto fronts (cf. Fig. 4) across regularization strengths " + }, + { + "bbox": [ + 45, + 380, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{loc}}" + }, + { + "bbox": [ + 45, + 380, + 546, + 426 + ], + "type": "text", + "content": " and epochs (cf. Sec. 5 and Fig. 4). We find the Energy loss to give the best trade-off between EPG and F1." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 433, + 237, + 518 + ], + "lines": [ + { + "bbox": [ + 62, + 433, + 237, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 237, + 518 + ], + "type": "image", + "image_path": "bf6798f67f3cdb993070819b8ad63e26b8f3bdcd2e1631b4917bbe88405df895.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 241, + 434, + 384, + 518 + ], + "blocks": [ + { + "bbox": [ + 241, + 434, + 384, + 518 + ], + "lines": [ + { + "bbox": [ + 241, + 434, + 384, + 518 + ], + "spans": [ + { + "bbox": [ + 241, + 434, + 384, + 518 + ], + "type": "image", + "image_path": "9ed3e743196a6931b8b142e0773581defda3439b3d94332aa17af2b6c745f881.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 522, + 546, + 567 + ], + "lines": [ + { + "bbox": [ + 45, + 522, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 45, + 522, + 546, + 567 + ], + "type": "text", + "content": "Fig. 6: IoU vs. F1, for different losses (markers) and models (columns) for VOC; results for COCO are in the supplement. Additionally, we show the performance of the baseline model before fine-tuning and demarcate regions that strictly dominate (are strictly dominated by) the baseline model in green (grey). For each configuration, we show the Pareto fronts (Fig. 4) across regularization strengths " + }, + { + "bbox": [ + 45, + 522, + 546, + 567 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{loc}}" + }, + { + "bbox": [ + 45, + 522, + 546, + 567 + ], + "type": "text", + "content": " and all epochs; for details, see Secs. 4 and 5. Across all configurations, we find the " + }, + { + "bbox": [ + 45, + 522, + 546, + 567 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 45, + 522, + 546, + 567 + ], + "type": "text", + "content": " loss to provide the largest gains in IoU at the lowest cost." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 386, + 434, + 531, + 518 + ], + "blocks": [ + { + "bbox": [ + 386, + 434, + 531, + 518 + ], + "lines": [ + { + "bbox": [ + 386, + 434, + 531, + 518 + ], + "spans": [ + { + "bbox": [ + 386, + 434, + 531, + 518 + ], + "type": "image", + "image_path": "961ad1912b1b2924ee63112b372d0ad9667d12f307e7cec569605538a6cfaa0d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 50, + 578, + 285, + 632 + ], + "blocks": [ + { + "bbox": [ + 50, + 578, + 285, + 632 + ], + "lines": [ + { + "bbox": [ + 50, + 578, + 285, + 632 + ], + "spans": [ + { + "bbox": [ + 50, + 578, + 285, + 632 + ], + "type": "image", + "image_path": "9fb3b15b6d409434a3aaf2d1988a0266eaf4b85308a2b1f66504a6c022f1e917.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 635, + 288, + 670 + ], + "lines": [ + { + "bbox": [ + 46, + 635, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 288, + 670 + ], + "type": "text", + "content": "Fig. 7: EPG vs. F1 on VOC. We observe the same trends as in Fig. 5a for different backbone architectures, specifically a B-cos DenseNet-121 and a B-cos ViT-S. For IoU results, see supplement." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 675, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 675, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 675, + 288, + 712 + ], + "type": "text", + "content": "studies in which we evaluate and discuss the cost of model guidance approaches: in particular, we study model guidance with limited additional labels, with increasingly coarse" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "content": "bounding boxes, and at deep layers in the network. Finally, in Sec. 5.5, we show the utility of model guidance in improving accuracy in the presence of distribution shifts. For easier reference, we label our individual findings as R1-R9. Note. To draw conclusive insights and highlight general and reliable trends in the experiments, we compare the Pareto curves (see Fig. 4) of individual configurations. If the Pareto curve of a specific loss (e.g. Energy in Fig. 5) consistently Pareto-dominates the Pareto curves of all other losses, we can confidently conclude that for the combination of evaluated metrics (e.g. EPG vs. F1), this loss is the best choice." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1927" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 190, + 155 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 190, + 155 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 190, + 155 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 190, + 155 + ], + "type": "image", + "image_path": "57dacd8004bef9b1b5471ea5f84371e0d5cd90b1fe77101f40e107e02c8858ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 158, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 158, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 547, + 191 + ], + "type": "text", + "content": "Fig. 8: Faster training by guiding at later layers. While input-level attributions tend to be more detailed (cf. Fig. 2), they are costlier to compute than attributions at later layers. However, we find that guidance at later layers (e.g. @Mid3) also significantly improves input-level attributions, yielding similar EPG results as input-level guidance (@Input) at up to twice the training speed; for IoU results, see supplement." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 191, + 70, + 304, + 154 + ], + "blocks": [ + { + "bbox": [ + 191, + 70, + 304, + 154 + ], + "lines": [ + { + "bbox": [ + 191, + 70, + 304, + 154 + ], + "spans": [ + { + "bbox": [ + 191, + 70, + 304, + 154 + ], + "type": "image", + "image_path": "68f77076c0fc0a2854aceeb2cae1d34d01a892c4ae8764f80201ea7bb1a06057.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 417, + 154 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 417, + 154 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 417, + 154 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 417, + 154 + ], + "type": "image", + "image_path": "462d3d36f2f0a59e351b2e1613c75efd7065699303ff68a1520a4d679bc1c385.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 417, + 70, + 531, + 154 + ], + "blocks": [ + { + "bbox": [ + 417, + 70, + 531, + 154 + ], + "lines": [ + { + "bbox": [ + 417, + 70, + 531, + 154 + ], + "spans": [ + { + "bbox": [ + 417, + 70, + 531, + 154 + ], + "type": "image", + "image_path": "64bc2fd43df95645b1076b0b3e789287ee399fbfe3b8d4805c50ecfb31b00a18.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 214, + 282, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 282, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 282, + 227 + ], + "type": "text", + "content": "5.1. Comparing loss functions for model guidance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 293 + ], + "type": "text", + "content": "In the following, we highlight the main insights gained from the quantitative evaluations. For a qualitative comparison between the losses, please see Fig. 9; note that we show examples for a B-cos model as the differences become clearest; full results can be found in the supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 293, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "type": "text", + "content": "R1 The Energy loss yields the best EPG scores. In Fig. 5, we plot the Pareto curves for EPG vs. F1 scores for a wide range of configurations (see Sec. 4) on VOC (a) and COCO (b); specifically, we group the results by model type (Vanilla, " + }, + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "type": "text", + "content": "-DNN, B-cos), the layer depths at which the attribution was regularized (Input / Final), and the loss used during optimization (Energy, " + }, + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "type": "text", + "content": ", PPCE, RRR*). From these results it becomes apparent that the optimization with the Energy loss yields the best trade-off between accuracy (F1) and the EPG score: e.g., when looking at the upper right plot in Fig. 5a we can see that the Energy loss (red dots) improves over the baseline B-cos model (white cross) by improving the localization in terms of EPG score with only a minor cost in classification performance (i.e. F1 score). Further trading off F1 scores yields even higher EPG scores. Importantly, the Energy loss Pareto-dominates all the other losses (RRR*: blue diamonds; " + }, + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "type": "text", + "content": ": green triangles; PPCE: yellow pentagons). This is is also true for the other network types (Vanilla ResNet-50, Fig. 5a (top left), and " + }, + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 293, + 287, + 567 + ], + "type": "text", + "content": "-DNN, Fig. 5a (top center)) and at the final layer (bottom row), and generalizes across backbone architectures (Fig. 7). When comparing Fig. 5a and Fig. 5b, we also find these results to be highly consistent between datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "text", + "content": "R2 The " + }, + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "text", + "content": " loss yields the best IoU performance. Similarly, in Fig. 6, we plot the Pareto curves of IoU vs. F1 scores for various configurations at the final layer; for the IoU results at the input layer and on the COCO dataset, please see the supplement. For IoU, the " + }, + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "text", + "content": " loss provides the best trade-off and, with few exceptions, " + }, + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 569, + 287, + 652 + ], + "type": "text", + "content": "-guided models Pareto-dominate all other models in all configurations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "R3 The Energy loss focuses best on on-object features. By not forcing the models to highlight the entire bounding boxes (see Sec. 3.4), we find that the Energy loss also suppresses background features within the bounding boxes, thus better preserving fine details of the explanations" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 309, + 213, + 543, + 257 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 543, + 257 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 543, + 257 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 543, + 257 + ], + "type": "image", + "image_path": "15644f355a0560a56d95c049e06dea1248bd183d6c7311aa4f89b6c5bd84c400.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 261, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 305, + 261, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 545, + 327 + ], + "type": "text", + "content": "Fig. 9: Loss comparison for input attributions (atts.) of a B-cos model. We showatts. before (baseline, col. 2) and after guidance (cols. 3-6) for a specific image (col. 1) and its bounding box annotation. We find that Energy and RRR* yield sparseatts, whereas " + }, + { + "bbox": [ + 305, + 261, + 545, + 327 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 305, + 261, + 545, + 327 + ], + "type": "text", + "content": " yields smootheratts, as it is optimized to fill the entire bounding box. For PPCE we observe only a minor effect on theatts." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 333, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 441 + ], + "type": "text", + "content": "(cf. Figs. 9 and 11). To quantify this, we evaluate the distribution of Energy (Eq. (2)) just within the bounding boxes. For this, we take advantage of the segmentation mask annotations available for a subset of the VOC test set. Specifically, we measure the Energy contained in the segmentation masks versus the entire bounding box, which indicates how much of the attributions actually highlight on-object features. We find that the Energy loss outperforms " + }, + { + "bbox": [ + 304, + 333, + 545, + 441 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 333, + 545, + 441 + ], + "type": "text", + "content": " across all models and configurations; see supplement for details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 442, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 442, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 545, + 526 + ], + "type": "text", + "content": "In short, we find that the Energy loss works best for improving the EPG metric, whereas the " + }, + { + "bbox": [ + 304, + 442, + 545, + 526 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 442, + 545, + 526 + ], + "type": "text", + "content": " loss yields the highest gains in terms of IoU; depending on the use case, either of these losses could thus be recommendable. However, we find that the Energy loss is more robust to annotation errors (R8, Sec. 5.4), and, as discussed in R3, the Energy loss more reliably focuses on object-specific features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 537, + 534, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 534, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 534, + 550 + ], + "type": "text", + "content": "5.2. Comparing models and attribution methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 556, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 616 + ], + "type": "text", + "content": "In the following, we highlight our findings regarding different attribution methods and models. Given the similarity of the results between GradCAM and IxG, and since B-cos attributions performed better than GradCAM for B-cos models, we show GradCAM results in the supplement." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "R4 At the input layer, B-cos explanations perform best. We find that the B-cos models not only achieve the highest EPG/IoU performance before applying model guidance, ('baseslines') but also obtain the highest gains in EPG and IoU and thus the highest overall performance (for EPG see Fig. 5, right; for IoU, see supplement): e.g., an Energy-based B-cos model achieves an EPG score of " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "71.7@79.4\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " F1, thus significantly outperforming the best EPG scores of" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1928" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "both other model types at a much lower cost in F1 (Vanilla: 55.8 @ " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "69.0\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "-DNN: 62.3 @ " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "68.9\\%" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "). This is also observed qualitatively, as we show in the supplement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 156 + ], + "type": "text", + "content": "R5 Regularizing at the final layer yields consistent gains. As can be seen in Fig. 5 (bottom) and Fig. 6, all models can be guided well via regularization at the final layer, i.e. all models show improvements in IoU and EPG score." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 286, + 228 + ], + "type": "text", + "content": "In short, we find model guidance to work well across all tested models when optimizing at the final layer (R5), highlighting its wide applicability. However, to obtain highly detailed and well-localized attributions at the input layer, the model-inherent explanations of the B-cos models seem to lend themselves much better to such guidance (R4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 236, + 263, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 263, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 263, + 249 + ], + "type": "text", + "content": "5.3. Improving accuracy with model guidance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "text", + "content": "R6 Model guidance can improve accuracy. For both the Vanilla models (final layer) and the " + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "text", + "content": "-DNNs (input+final), we found models that improve the localization metrics and the F1 score. These improvements are particularly pronounced for the " + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "text", + "content": "-DNN: e.g., we find models that improve the EPG and F1 scores by " + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "inline_equation", + "content": "\\Delta = 7.2" + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "text", + "content": " p.p. and " + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "inline_equation", + "content": "\\Delta = 1.4" + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "text", + "content": " p.p. respectively (Fig. 5, center top), or the IoU and F1 scores by " + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "inline_equation", + "content": "\\Delta = 11.9" + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "text", + "content": " p.p. and " + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "inline_equation", + "content": "\\Delta = 1.4" + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "text", + "content": " p.p. (Fig. 6, center)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 351, + 286, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 286, + 459 + ], + "type": "text", + "content": "However, overall we observe a trade-off between localization and accuracy (Figs. 5 and 6). Given the similarity of the training and test distributions, focusing on the object need not improve classification performance, as spurious features are also present at test time. Further, the guided model is discouraged from relying on contextual features, making the classification more challenging. In Sec. 5.5, we show that guidance can significantly improve performance when there is a distribution shift between training and test." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 467, + 258, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 258, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 258, + 479 + ], + "type": "text", + "content": "5.4. Efficiency and robustness considerations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 605 + ], + "type": "text", + "content": "While bounding boxes decrease the data collection cost with respect to segmentation masks, they can nonetheless be expensive to obtain, especially when expert knowledge is required. To further reduce those costs, in this section, we assess the robustness of guiding the model with a limited number (R7) or increasingly coarse annotations (R8). Apart from data efficiency, we further explore how training efficiency can be improved for fine-grained (i.e. input-level) explanations (R9), as explanations at early layers are more costly to obtain than those at later layers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "R7 Model guidance requires only few add. annotations. In Fig. 12, we show that the EPG score can be significantly improved with a very limited number of annotations; for IoU results, see supplement. Specifically, we find that when using only " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": " of the training data (25 annotated images) for VOC, improvements of up to " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\Delta = 23.0" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": " p.p. (" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\Delta = 1.4" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": ") in EPG (IoU) can be obtained, at a minor drop in F1 (" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\Delta = 0.3" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": " p.p. and " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\Delta = 2.5" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": " p.p. respectively). When annotating up to " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": " of the images, very similar results can be achieved as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 509, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 509, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 509, + 84 + ], + "type": "text", + "content": "with full annotation (see e.g. cols. 2+3 in Fig. 12)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 85, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 545, + 251 + ], + "type": "text", + "content": "R8 The Energy loss is highly robust to annotation errors. As discussed in Sec. 3.4, the Energy loss only directs the model on which features not to use and does not impose a uniform prior on the attributions within the bounding boxes. As a result, we find it to be much more stable to annotation errors: e.g., in Fig. 10, we visualize how the EPG (top) and IoU (bottom) scores of the best performing models under the Energy (left) and " + }, + { + "bbox": [ + 304, + 85, + 545, + 251 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 85, + 545, + 251 + ], + "type": "text", + "content": " loss (right) evolve when using coarser bounding boxes; for this, we simply dilate the bounding box size by " + }, + { + "bbox": [ + 304, + 85, + 545, + 251 + ], + "type": "inline_equation", + "content": "p \\in \\{10,25,50\\} \\%" + }, + { + "bbox": [ + 304, + 85, + 545, + 251 + ], + "type": "text", + "content": " during training, see Fig. 11. While the models optimized via the " + }, + { + "bbox": [ + 304, + 85, + 545, + 251 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 85, + 545, + 251 + ], + "type": "text", + "content": " loss achieve increasingly worse results (right), the Energy-optimized models are essentially unaffected by the coarseness of the annotations." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 301, + 260, + 421, + 323 + ], + "blocks": [ + { + "bbox": [ + 301, + 260, + 421, + 323 + ], + "lines": [ + { + "bbox": [ + 301, + 260, + 421, + 323 + ], + "spans": [ + { + "bbox": [ + 301, + 260, + 421, + 323 + ], + "type": "image", + "image_path": "f385551e04751c4ed6c0b068e6b01fe2ff7332ef175fa5fbb71be2568634703c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 430, + 260, + 539, + 322 + ], + "blocks": [ + { + "bbox": [ + 430, + 260, + 539, + 322 + ], + "lines": [ + { + "bbox": [ + 430, + 260, + 539, + 322 + ], + "spans": [ + { + "bbox": [ + 430, + 260, + 539, + 322 + ], + "type": "image", + "image_path": "3eb12969c3771dbff17478fb55232e4d736c41504d1530c037e666481a91a7cf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 301, + 329, + 421, + 393 + ], + "blocks": [ + { + "bbox": [ + 301, + 329, + 421, + 393 + ], + "lines": [ + { + "bbox": [ + 301, + 329, + 421, + 393 + ], + "spans": [ + { + "bbox": [ + 301, + 329, + 421, + 393 + ], + "type": "image", + "image_path": "acedbb47fc5a14fd412f4776d2b3ad4129b40abe204f5564486250c8cf229612.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 397, + 545, + 475 + ], + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 475 + ], + "type": "text", + "content": "Fig. 10: Quantitative results for dilated bounding boxes for a B-cos model at the input layer. We show EPG and IoU (top and bottom) results for models trained with various amounts of annotation errors (increasingly large bounding boxes, see Fig. 11). The Energy loss yields highly consistent results despite training with heavily dilated bounding boxes (left), whereas the results of the " + }, + { + "bbox": [ + 304, + 397, + 545, + 475 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 397, + 545, + 475 + ], + "type": "text", + "content": " loss (right) worsen markedly; best viewed in color." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 430, + 329, + 538, + 393 + ], + "blocks": [ + { + "bbox": [ + 430, + 329, + 538, + 393 + ], + "lines": [ + { + "bbox": [ + 430, + 329, + 538, + 393 + ], + "spans": [ + { + "bbox": [ + 430, + 329, + 538, + 393 + ], + "type": "image", + "image_path": "0b57e7c7fc7253a8c736dc3f1bbaba077b59a9dae3ec20d35a382f47dd59edc2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 310, + 502, + 426, + 613 + ], + "blocks": [ + { + "bbox": [ + 310, + 502, + 426, + 613 + ], + "lines": [ + { + "bbox": [ + 310, + 502, + 426, + 613 + ], + "spans": [ + { + "bbox": [ + 310, + 502, + 426, + 613 + ], + "type": "image", + "image_path": "76750dff31b25685457afbe6b1c59b05ebd692428814b23b9cece6fc7073e0dd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 617, + 545, + 672 + ], + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 672 + ], + "type": "text", + "content": "Fig. 11: Qualitative results for dilated bounding boxes for a B-cos model at input. Examples for attributions (rows " + }, + { + "bbox": [ + 304, + 617, + 545, + 672 + ], + "type": "inline_equation", + "content": "2 + 3" + }, + { + "bbox": [ + 304, + 617, + 545, + 672 + ], + "type": "text", + "content": ") of models trained with dilated bounding boxes (row 1). In contrast to " + }, + { + "bbox": [ + 304, + 617, + 545, + 672 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 617, + 545, + 672 + ], + "type": "text", + "content": ", models trained with Energy show significant gains in object focus even with significant noise (e.g. 'Baseline' vs. " + }, + { + "bbox": [ + 304, + 617, + 545, + 672 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 672 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 429, + 502, + 543, + 613 + ], + "blocks": [ + { + "bbox": [ + 429, + 502, + 543, + 613 + ], + "lines": [ + { + "bbox": [ + 429, + 502, + 543, + 613 + ], + "spans": [ + { + "bbox": [ + 429, + 502, + 543, + 613 + ], + "type": "image", + "image_path": "53d85424fbc3d08bf14d5d6ac9eb76b301d08444ebe4265b9c93b3fabd20ac34.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In short, we find that the models can be guided effectively at a low cost in terms of annotation effort, as only few" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1929" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "annotations (e.g. 25 for VOC) are required (cf. R7), and, especially for the Energy loss, these annotations can be very coarse and do not have to be 'pixel-perfect' (cf. R8)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 216 + ], + "type": "text", + "content": "R9 Guidance at deep layers can be effective. While guided input-level explanations of B-cos networks exhibit a high degree of detail, regularizing those explanations comes at an added training cost. In particular, optimizing at the input layer requires backpropagating through the entire network to compute the attributions. In an effort to reduce training costs whilst maintaining the benefits of fine-grained explanations at input resolution, we evaluate if input-level attributions benefit from an optimization at deeper layers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 217, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 324 + ], + "type": "text", + "content": "Specifically, we regularize B-cos attributions at the final and at three intermediate layers (Mid{1,2,3}), and evaluate the localization of attributions at the input. We find (Fig. 8) that training at a deeper layer can provide significant speed-ups in training time with often a negligible cost in localization performance. E.g., since we do not have to compute a full backward pass through the entire model during training, optimizing at Mid2 (col. 2 in Fig. 8) provides similar gains in localization but with a " + }, + { + "bbox": [ + 46, + 217, + 287, + 324 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{x}" + }, + { + "bbox": [ + 46, + 217, + 287, + 324 + ], + "type": "text", + "content": " speed-up in training time." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 42, + 335, + 284, + 409 + ], + "blocks": [ + { + "bbox": [ + 42, + 335, + 284, + 409 + ], + "lines": [ + { + "bbox": [ + 42, + 335, + 284, + 409 + ], + "spans": [ + { + "bbox": [ + 42, + 335, + 284, + 409 + ], + "type": "image", + "image_path": "958e439fecad0c533c1d88ff82b917fa6451906acf38fd1f5d3d7b3ee3521e6d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 412, + 287, + 468 + ], + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 468 + ], + "type": "text", + "content": "Fig. 12: EPG results with limited annotations for a B-cos model at the input layer, optimized with the Energy and the " + }, + { + "bbox": [ + 46, + 412, + 287, + 468 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 412, + 287, + 468 + ], + "type": "text", + "content": " loss. Using bounding box annotations for as little as " + }, + { + "bbox": [ + 46, + 412, + 287, + 468 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 412, + 287, + 468 + ], + "type": "text", + "content": " (left) of the images yields significant improvements in EPG, and with " + }, + { + "bbox": [ + 46, + 412, + 287, + 468 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 412, + 287, + 468 + ], + "type": "text", + "content": " (center) similar gains as in the fully annotated setting (right) are obtained." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 491, + 269, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 269, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 269, + 504 + ], + "type": "text", + "content": "5.5. Effectiveness against spurious correlations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": "To evaluate the potential for mitigating spurious correlations, we evaluate model guidance with the Energy and " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": " losses on the synthetically constructed Waterbirds-100 dataset [51, 42]. We perform model guidance under two settings: (1) the conventional setting to classify between landbirds and waterbirds, using the region within the bounding box as the mask; and (2) the reversed setting [42] to classify the background, i.e., land vs. water, using the region outside the bounding box as the mask. To simulate a limited annotation budget, we only use bounding boxes for a random " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": " of the training set, and report results averaged over four runs. We show the results for the worst-group accuracy (i.e., images containing a waterbird on land) and the overall accuracy using B-cos models in Tab. 1; full results for all attributions and models can be found in the supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Both losses consistently and significantly improve the accuracy in the conventional and the reversed settings by" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 137 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 137 + ], + "type": "image", + "image_path": "c1ca9fe9ad24bf88749f50181dd215a50a76af682ee79a0dc622f5ded838f8ba.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 140, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 304, + 140, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 545, + 217 + ], + "type": "text", + "content": "Fig. 13: Qualitative Waterbirds-100 results. Without guidance, a model might focus on the background to classify birds (baseline) and thus misclassify waterbirds on land (col. 2). Guided models can correct such errors and focus on the desired feature: in cols. " + }, + { + "bbox": [ + 304, + 140, + 545, + 217 + ], + "type": "inline_equation", + "content": "3 + 4" + }, + { + "bbox": [ + 304, + 140, + 545, + 217 + ], + "type": "inline_equation", + "content": "(5 + 6)" + }, + { + "bbox": [ + 304, + 140, + 545, + 217 + ], + "type": "text", + "content": " the model is guided to classify by using the bird (background) features and arrives at the desired prediction. Model predictions and confidence scores are indicated below the images." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 220, + 543, + 282 + ], + "blocks": [ + { + "bbox": [ + 309, + 220, + 543, + 282 + ], + "lines": [ + { + "bbox": [ + 309, + 220, + 543, + 282 + ], + "spans": [ + { + "bbox": [ + 309, + 220, + 543, + 282 + ], + "type": "table", + "html": "
ConventionalReversed
ModelWorstOverallWorstOverall
Baseline43.4 (±2.4)68.7 (±0.2)56.6 (±2.4)80.1 (±0.2)
Energy56.1 (±4.0)71.2 (±0.1)62.8 (±2.1)83.6 (±1.1)
L151.1 (±1.9)69.5 (±0.2)58.8 (±5.0)82.2 (±0.9)
", + "image_path": "97fa313dbd7cca7938b99692246a08597eebb687b39ae2eacdd420aabd5a6282.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 284, + 545, + 328 + ], + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 328 + ], + "type": "text", + "content": "Table 1: Waterbirds-100 results. We find that model guidance is effective in improving both worst-group ('Waterbird on Land') and overall accuracy in the conventional (Landbird vs. Waterbird) and reversed (Land vs. Water) settings; full results in the supplement." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 335, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 371 + ], + "type": "text", + "content": "guiding the model to select the 'right' features, i.e. birds (conventional) or background (reversed). This guidance can also be observed qualitatively (cf. Fig. 13)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 396, + 460, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 460, + 409 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 460, + 409 + ], + "type": "text", + "content": "6. Discussion And Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 458 + ], + "type": "text", + "content": "In this work, we comprehensively evaluated various models, attribution methods, and loss functions for their utility in guiding models to be \"right for the right reasons\"." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": "In summary, we find that guiding models via bounding boxes can significantly improve EPG and IoU performance of the optimized attribution method, with the Energy loss working best to improve the EPG score (R1) and the " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " loss yielding the highest gains in IoU scores (R2). While the B-cos models achieve the best results in IoU and EPG score at the input layer (R4), all tested model types (Vanilla, " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": "-DNN, B-cos) lend themselves well to being optimized at the final layer (R5), which can even improve attribution maps at early layers (R9). Further, we find that regularizing the explanations of the models and thereby 'telling them where to look' can increase the object recognition performance (mAP/accuracy) of some models (R6), especially when strong spurious correlations are present (Sec. 5.5). Interestingly, those gains (EPG, IoU), can be achieved with relatively little additional annotation (R7). Lastly, we find that by not assuming a uniform prior over the attributions within the annotated bounding boxes, training with the energy loss is more robust to annotation errors (R8) and results in models that produce attribution maps that are more focused on class-specific features (R3)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1930" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 91, + 287, + 711 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 49, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity Checks for Saliency Maps. In NeurIPS, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 124, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 124, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 49, + 124, + 287, + 156 + ], + "type": "text", + "content": "[2] Julius Adebayo, Michael Muelly, Harold Abelson, and Been Kim. Post hoc Explanations may be Ineffective for Detecting Unknown Spurious Correlation. In ICLR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 156, + 287, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 156, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 156, + 287, + 198 + ], + "type": "text", + "content": "[3] Saeid Asgari, Aliasghar Khani, Fereshte Khani, Ali Gholami, Linh Tran, Ali Mahdavi-Amiri, and Ghassan Hamarneh. MaskTune: Mitigating Spurious Correlations by Forcing to Explore. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 198, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 287, + 251 + ], + "type": "text", + "content": "[4] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation. PLOS One, 10(7):e0130140, 2015. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 252, + 286, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 252, + 286, + 273 + ], + "spans": [ + { + "bbox": [ + 49, + 252, + 286, + 273 + ], + "type": "text", + "content": "[5] Jürgen Backhaus. The Pareto Principle. Analyse & Kritik, 2(2):146-171, 1980. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 274, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 274, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 274, + 287, + 304 + ], + "type": "text", + "content": "[6] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-Cos Networks: Alignment is All We Need for Interpretability. In CVPR, pages 10329-10338, 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 305, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 305, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 49, + 305, + 287, + 346 + ], + "type": "text", + "content": "[7] Moritz Böhle, Navdeepal Singh, Mario Fritz, and Bernt Schiele. B-cos Alignment for Inherently Interpretable CNNs and Vision Transformers. arXiv preprint arXiv:2306.10898, 2023.5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 347, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 347, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 347, + 287, + 380 + ], + "type": "text", + "content": "[8] Moritz Böhle, Mario Fritz, and Bernt Schiele. Convolutional Dynamic Alignment Networks for Interpretable Classifications. In CVPR, pages 10029-10038, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 380, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 380, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 49, + 380, + 287, + 422 + ], + "type": "text", + "content": "[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-CAM++: Improved Visual Explanations for Deep Convolutional Networks. In WACV, pages 839-847, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 423, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 423, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 423, + 287, + 453 + ], + "type": "text", + "content": "[10] Hila Chefer, Shir Gur, and Lior Wolf. Transformer Interpretability Beyond Attention Visualization. In CVPR, pages 782-791, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 454, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 454, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 49, + 454, + 287, + 485 + ], + "type": "text", + "content": "[11] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing Relevance Maps of Vision Transformers Improves Robustness. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 486, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 486, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 49, + 486, + 287, + 506 + ], + "type": "text", + "content": "[12] Piotr Dabkowski and Yarin Gal. Real Time Image Saliency for Black Box Classifiers. In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 507, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 507, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 49, + 507, + 287, + 548 + ], + "type": "text", + "content": "[13] Saurabh Desai and Harish G. Ramaswamy. Ablation-CAM: Visual Explanations for Deep Convolutional Network via Gradient-free Localization. In WACV, pages 983-991, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 549, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 549, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 549, + 287, + 613 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In ICLR, 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 614, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 667 + ], + "type": "text", + "content": "[15] Gabriel Erion, Joseph D Janizek, Pascal Sturmfels, Scott M Lundberg, and Su-In Lee. Improving Performance of Deep Learning Models with Axiomatic Attribution Priors and Expected Gradients. Nature Machine Intelligence, 3(7):620-631, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 669, + 287, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 669, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 49, + 669, + 287, + 700 + ], + "type": "text", + "content": "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. IJCV, 88:303-308, 2009. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 49, + 700, + 287, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 700, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 49, + 700, + 287, + 711 + ], + "type": "text", + "content": "[17] Thomas Fel, Ivan F Rodriguez Rodriguez, Drew Linsley, and" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 702 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 323, + 74, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 74, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 323, + 74, + 545, + 105 + ], + "type": "text", + "content": "Thomas Serre. Harmonizing the Object Recognition Strategies of Deep Neural Networks with Humans. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 106, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 137 + ], + "type": "text", + "content": "[18] Vitaly Feldman and Chiyuan Zhang. What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation. In NeurIPS, pages 2881-2891, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 138, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 138, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 308, + 138, + 545, + 169 + ], + "type": "text", + "content": "[19] Patrick Fernandes, Marcos Treviso, Danish Pruthi, André FT Martins, and Graham Neubig. Learning to Scaffold: Optimizing Model Explanations for Teaching. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 170, + 545, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 170, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 308, + 170, + 545, + 200 + ], + "type": "text", + "content": "[20] Ruth C Fong and Andrea Vedaldi. Interpretable Explanations of Black Boxes by Meaningful Perturbation. In ICCV, pages 3429-3437, 2017. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "text", + "content": "[21] Felix Friedrich, Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. A Typology to Explore and Guide Explanatory Interactive Machine Learning. arXiv preprint arXiv:2203.03668, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 244, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 244, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 244, + 545, + 285 + ], + "type": "text", + "content": "[22] Hiroshi Fukui, Tsubasa Hirakawa, Takayoshi Yamashita, and Hironobu Fujiyoshi. Attention Branch Network: Learning of Attention Mechanism for Visual Explanation. In CVPR, pages 10705-10714, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 287, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 329 + ], + "type": "text", + "content": "[23] Yuyang Gao, Tong Steven Sun, Guangji Bai, Siyi Gu, Sungsoo Ray Hong, and Zhao Liang. RES: A Robust Framework for Guiding Visual Explanation. In KDD, pages 432-442, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 330, + 545, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 330, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 308, + 330, + 545, + 372 + ], + "type": "text", + "content": "[24] Yuyang Gao, Tong Steven Sun, Liang Zhao, and Sungsoo Ray Hong. Aligning Eyes between Humans and Deep Neural Network through Interactive Attention Alignment. ACM HCI, 6(CSCW2):1-28, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 373, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 415 + ], + "type": "text", + "content": "[25] Hao Guo, Kang Zheng, Xiaochuan Fan, Hongkai Yu, and Song Wang. Visual Attention Consistency under Image Transforms for Multi-Label Image Classification. In CVPR, pages 729–739, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 415, + 545, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 457 + ], + "type": "text", + "content": "[26] Misgina Tsighe Hagos, Kathleen M Curran, and Brian Mac Namee. Identifying Spurious Correlations and Correcting them with an Explanation-based Learning. arXiv preprint arXiv:2211.08285, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 458, + 545, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 458, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 308, + 458, + 545, + 490 + ], + "type": "text", + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 491, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 491, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 491, + 545, + 521 + ], + "type": "text", + "content": "[28] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Fast Axiomatic Attribution for Neural Networks. In NeurIPS, pages 19513-19524, 2021. 4, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 522, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 553 + ], + "type": "text", + "content": "[29] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely Connected Convolutional Networks. In CVPR, pages 4700-4708, 2017. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 554, + 545, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 545, + 596 + ], + "type": "text", + "content": "[30] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. LayerCAM: Exploring Hierarchical Class Activation Maps for Localization. IEEE TIP, 30:5875-5888, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 597, + 545, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 597, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 308, + 597, + 545, + 616 + ], + "type": "text", + "content": "[31] Joon Sik Kim, Gregory Plumb, and Ameet Talwalkar. Sanity Simulations for Saliency Methods. In ICML, 2022. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 618, + 545, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 618, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 308, + 618, + 545, + 650 + ], + "type": "text", + "content": "[32] Keisuke Kiritoshi, Ryosuke Tanno, and Tomonori Izumitani. L1-Norm Gradient Penalty for Noise Reduction of Attribution Maps. In CVPRW, pages 118-121, 2019. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 650, + 545, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 650, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 308, + 650, + 545, + 681 + ], + "type": "text", + "content": "[33] Kunpeng Li, Ziyan Wu, Kuan-Chuan Peng, Jan Ernst, and Yun Fu. Tell Me Where to Look: Guided Attention Inference Network. In CVPR, pages 9215-9223, 2018. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 682, + 545, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 682, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 308, + 682, + 545, + 702 + ], + "type": "text", + "content": "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1931" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 63, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 286, + 95 + ], + "type": "text", + "content": "Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, pages 740-755. Springer, 2014. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 286, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 286, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 286, + 115 + ], + "type": "text", + "content": "[35] Drew Linsley, Dan Shiebler, Sven Eberhardt, and Thomas Serre. Learning what and where to attend. In ICLR, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 115, + 286, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 115, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 48, + 115, + 286, + 169 + ], + "type": "text", + "content": "[36] Masahiro Mitsuhara, Hiroshi Fukui, Yusuke Sakashita, Takanori Ogata, Tsubasa Hirakawa, Takayoshi Yamashita, and Hironobu Fujiyoshi. Embedding Human Knowledge into Deep Neural Network via Attention Map. In VISIGRAPP, pages 626-636, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 169, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 169, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 169, + 286, + 201 + ], + "type": "text", + "content": "[37] Ofir Moshe, Gil Fidel, Ron Bitton, and Asaf Shabtai. Improving Interpretability via Regularization of Neural Activation Sensitivity. arXiv preprint arXiv:2211.08686, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 201, + 286, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 286, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 286, + 232 + ], + "type": "text", + "content": "[38] Vinod Nair and Geoffrey E Hinton. Rectified Linear Units Improve Restricted Boltzmann Machines. In ICML, pages 807-814, 2010. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 233, + 286, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 286, + 264 + ], + "type": "text", + "content": "[39] Krishna Kanth Nakka and Mathieu Salzmann. Towards Robust Fine-Grained Recognition by Maximal Separation of Discriminative Features. In ACCV, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 264, + 286, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 286, + 286 + ], + "type": "text", + "content": "[40] Vilfredo Pareto. Il massimo di utilità dato alla libera concorrenza. Giornale degli economisti, pages 48-66, 1894. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 286, + 286, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 286, + 318 + ], + "type": "text", + "content": "[41] Vilfredo Pareto. The Maximum of Utility given by Free Competition. Giornale degli Economisti e Annali di Economia, 67(3):387-403, 2008. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 318, + 286, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 318, + 286, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 318, + 286, + 360 + ], + "type": "text", + "content": "[42] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On Guiding Visual Attention with Language Specification. In CVPR, pages 18092-18102, 2022. 1, 2, 3, 5, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 360, + 286, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 286, + 392 + ], + "type": "text", + "content": "[43] Vitali Petsiuk, Abir Das, and Kate Saenko. RISE: Randomized Input Sampling for Explanation of Black-box Models. In BMVC, 2018. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 392, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 286, + 434 + ], + "type": "text", + "content": "[44] Vipin Pillai, Soroush Abbasi Koohpayegani, Ashley Ouligian, Dennis Fong, and Hamed Piriavash. Consistent Explanations by Contrastive Learning. In CVPR, pages 10213-10222, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 434, + 286, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 286, + 456 + ], + "type": "text", + "content": "[45] Vipin Pillai and Hamed Piri siavash. Explainable Models with Consistent Interpretations. In AAAI, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 456, + 286, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 286, + 487 + ], + "type": "text", + "content": "[46] Sukrut Rao, Moritz Böhle, and Bernt Schiele. Towards Better Understanding Attribution Methods. In CVPR, pages 10223-10232, 2022. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "type": "text", + "content": "[47] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"Why Should I Trust You?: Explaining the Predictions of Any Classifier. In KDD, pages 1135-1144, 2016. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 520, + 286, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 286, + 563 + ], + "type": "text", + "content": "[48] Laura Rieger, Chandan Singh, William Murdoch, and Bin Yu. Interpretations are Useful: Penalizing Explanations to Align Neural Networks with Prior Knowledge. In ICML, pages 8116-8126, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 563, + 286, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 563, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 286, + 605 + ], + "type": "text", + "content": "[49] Andrew Slavin Ross, Michael C Hughes, and Finale Doshi-Velez. Right for the Right Reasons: Training Differentiable Models by Constraining their Explanations. In *IJCAI*, pages 2662–2670, 2017. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 605, + 286, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 605, + 286, + 659 + ], + "spans": [ + { + "bbox": [ + 48, + 605, + 286, + 659 + ], + "type": "text", + "content": "[50] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211–252, 2015. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 659, + 286, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 659, + 286, + 702 + ], + "spans": [ + { + "bbox": [ + 48, + 659, + 286, + 702 + ], + "type": "text", + "content": "[51] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization. In ICLR, 2020. 2, 5, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 286, + 713 + ], + "type": "text", + "content": "[52] Patrick Schramowski, Wolfgang Stammer, Stefano Teso," + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 72, + 545, + 705 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 322, + 72, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 545, + 127 + ], + "type": "text", + "content": "Anna Brugger, Franziska Herbert, Xiaoting Shao, Hans-Georg Luigs, Anne-Katrin Mahlein, and Kristian Kersting. Making Deep Neural Networks Right for the Right Scientific Reasons by Interacting with their Explanations. Nature Machine Intelligence, 2(8):476-486, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 127, + 545, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 127, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 308, + 127, + 545, + 181 + ], + "type": "text", + "content": "[53] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. In ICCV, pages 618-626, 2017. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 181, + 545, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 181, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 308, + 181, + 545, + 234 + ], + "type": "text", + "content": "[54] Ramprasaath R Selvaraju, Stefan Lee, Yilin Shen, Hongxia Jin, Shalini Ghosh, Larry Heck, Dhruv Batra, and Devi Parikh. Taking a HINT: Leveraging Explanations to Make Vision and Language Models More Grounded. In ICCV, pages 2591-2600, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 234, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 234, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 308, + 234, + 545, + 288 + ], + "type": "text", + "content": "[55] Xiaoting Shao, Arseny Skryagin, Wolfgang Stammer, Patrick Schramowski, and Kristian Kersting. Right for Better Reasons: Training Differentiable Models by Constraining their Influence Functions. In AAAI, volume 35, pages 9533-9540, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 288, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 342 + ], + "type": "text", + "content": "[56] Haifeng Shen, Kewen Liao, Zhibin Liao, Job Doornberg, Maoying Qiao, Anton Van Den Hengel, and Johan W Verjans. Human-AI Interactive and Continuous Sensemaking: A Case Study of Image Classification using Scribble Attention Maps. In Extended Abstracts of CHI, pages 1-8, 2021. 2, 3, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 342, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 374 + ], + "type": "text", + "content": "[57] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning Important Features Through Propagating Activation Differences. In ICML, pages 3145-3153, 2017. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 374, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 374, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 374, + 545, + 407 + ], + "type": "text", + "content": "[58] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps. In ICLRw, 2014. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 407, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 407, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 407, + 545, + 449 + ], + "type": "text", + "content": "[59] Krishna Kumar Singh, Dhruv Mahajan, Kristen Grauman, Yong Jae Lee, Matt Feiszli, and Deepti Ghadiyaram. Don't Judge an Object by its Context: Learning to Overcome Contextual Bias. In CVPR, pages 11070-11078, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "text", + "content": "[60] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for Simplicity: The All Convolutional Net. In ICLRw, 2015. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 481, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 545, + 524 + ], + "type": "text", + "content": "[61] Guolei Sun, Salman Khan, Wen Li, Hisham Cholakkal, Fahad Shahbaz Khan, and Luc Van Gool. Fixing Localization Errors to Improve Image Classification. In ECCV, pages 271-287. Springer, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 524, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 545, + 555 + ], + "type": "text", + "content": "[62] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic Attribution for Deep Networks. In ICML, pages 3319-3328, 2017. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 599 + ], + "type": "text", + "content": "[63] Damien Teney, Ehsan Abbasnedjad, and Anton van den Hengel. Learning What Makes a Difference from Counterfactual Examples and Gradient Supervision. In ECCV, pages 580-599. Springer, 2020. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 599, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 599, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 308, + 599, + 545, + 631 + ], + "type": "text", + "content": "[64] Stefano Teso. Toward Faithful Explanatory Active Learning with Self-Explainable Neural Nets. In Workshop on IAL, pages 4–16. CEUR Workshop Proceedings, 2019. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 631, + 545, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 631, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 308, + 631, + 545, + 673 + ], + "type": "text", + "content": "[65] Stefano Teso, Öznur Alkan, Wolfgang Stammer, and Elizabeth Daly. Leveraging Explanations in Interactive Machine Learning: An Overview. Frontiers in Artificial Intelligence, 6:1066049, 2023. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 673, + 545, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 673, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 308, + 673, + 545, + 695 + ], + "type": "text", + "content": "[66] Stefano Teso and Kristian Kersting. Explanatory Interactive Machine Learning. In AIES, pages 239-245, 2019. 2, 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 695, + 545, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 695, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 308, + 695, + 545, + 705 + ], + "type": "text", + "content": "[67] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1932" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 288 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 63, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 287, + 106 + ], + "type": "text", + "content": "Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks. In CVPRW, pages 111-119, 2020. 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 137 + ], + "type": "text", + "content": "[68] Kai Yuanqing Xiao, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. Noise or Signal: The Role of Image Backgrounds in Object Recognition. In *ICLR*, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 137, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 287, + 179 + ], + "type": "text", + "content": "[69] Ziyan Yang, Kushal Kafle, Franck Dernoncourt, and Vicente Ordonez. Improving Visual Grounding by Encouraging Consistent Gradient-based Explanations. In CVPR, pages 19165–19174, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 180, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 180, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 48, + 180, + 287, + 212 + ], + "type": "text", + "content": "[70] Matthew D Zeiler and Rob Fergus. Visualizing and Understanding Convolutional Networks. In ECCV, pages 818-833, 2014. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 212, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 287, + 255 + ], + "type": "text", + "content": "[71] Michael Zhang, Nimit Sharad Sohoni, Hongyang R. Zhang, Chelsea Finn, and Christopher Ré. Correct-N-Contrast: a Contrastive Approach for Improving Robustness to Spurious Correlations. In ICML, pages 26484–26516, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 287, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 288 + ], + "type": "text", + "content": "[72] Yilun Zhou, Serena Booth, Marco Tulio Ribeiro, and Julie Shah. Do Feature Attribution Methods Correctly Attribute Features? In AAAI, volume 36, pages 9623-9633, 2022. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1933" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/fea588fd-ec6f-4543-ad58-5dd9ea412b01_content_list.json b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/fea588fd-ec6f-4543-ad58-5dd9ea412b01_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8bb45e36a07824a062043eedc01b2a3fc7773d3a --- /dev/null +++ b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/fea588fd-ec6f-4543-ad58-5dd9ea412b01_content_list.json @@ -0,0 +1,1878 @@ +[ + { + "type": "text", + "text": "StyleDiffusion: Controllable Disentangled Style Transfer via Diffusion Models", + "text_level": 1, + "bbox": [ + 91, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhizhong Wang*, Lei Zhao† Wei Xing \nCollege of Computer Science and Technology, Zhejiang University \n{endywon, czhl, WXING}@zju.edu.cn", + "bbox": [ + 218, + 180, + 750, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Content and style (C-S) disentanglement is a fundamental problem and critical challenge of style transfer. Existing approaches based on explicit definitions (e.g., Gram matrix) or implicit learning (e.g., GANs) are neither interpretable nor easy to control, resulting in entangled representations and less satisfying results. In this paper, we propose a new C-S disentangled framework for style transfer without using previous assumptions. The key insight is to explicitly extract the content information and implicitly learn the complementary style information, yielding interpretable and controllable C-S disentanglement and style transfer. A simple yet effective CLIP-based style disentanglement loss coordinated with a style reconstruction prior is introduced to disentangle C-S in the CLIP image space. By further leveraging the powerful style removal and generative ability of diffusion models, our framework achieves superior results than state of the art and flexible C-S disentanglement and trade-off control. Our work provides new insights into the C-S disentanglement in style transfer and demonstrates the potential of diffusion models for learning well-disentangled C-S characteristics.", + "bbox": [ + 75, + 300, + 473, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 645, + 209, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given a reference style image, e.g., Starry Night by Vincent Van Gogh, style transfer aims to transfer its artistic style, such as colors and brushstrokes, to an arbitrary content target. To achieve such a goal, it must first properly separate the style from the content and then transfer it to another content. This raises two fundamental challenges: (1) \"how to disentangle content and style (C-S)\" and (2) \"how to transfer style to another content\".", + "bbox": [ + 75, + 670, + 468, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To resolve these challenges, valuable efforts have been devoted. Gatys et al. [18] proposed A Neural Algorithm of Artistic Style to achieve style transfer, which explicitly defines the high-level features extracted from a pre-trained Convolutional Neural Network (CNN) (e.g., VGG [71]) as", + "bbox": [ + 75, + 791, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "content, and the feature correlations (i.e., Gram matrix) as style. This approach acquires visually stunning results and inspires a large number of successors [32, 27, 50, 1, 90]. Despite the successes, by diving into the essence of style transfer, we observed three problems with these approaches: (1) The C-S are not completely disentangled. Theoretically, the C-S representations are intertwined. For example, matching the content representation of an image may also match its Gram matrix, and vice versa. (2) What CNN learned is a black box rugged to interpret [92], which makes the C-S definitions [18] uninterpretable and hard to control. (3) The transfer process is modeled as a separate optimization of content loss and style loss [18], so there lacks a deep understanding of the relationship between C-S. These problems usually lead to unbalanced stylizations and disharmonious artifacts [6], as will be shown in later Fig. 3.", + "bbox": [ + 500, + 268, + 893, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, disentangled representation learning [24] provides other ideas to implicitly disentangle C-S, either supervised [44, 34] or unsupervised [9, 93]. For style transfer, Kotovenko et al. [42] utilized fixpoint triplet style loss and disentanglement loss to enforce a GAN [20]-based framework to learn separate C-S representations in an unsupervised manner. Similarly, TPFR [74] learned to disentangle C-S in latent space via metric learning and two-stage peer-regularization, producing high-quality images even in the zero-shot setting. While these approaches successfully enforce properties \"encouraged\" by the corresponding losses, they still have three main problems: (1) Well-disentangled models seemingly cannot be identified without supervision [54, 64], which means the unsupervised learning [42, 74] may not achieve truly disentangled C-S, as will be shown in later Fig. 3. (2) These approaches are all based on GANs and thus often confined to the GAN predefined domains, e.g., a specific artist's style domain [69]. (3) The implicitly learned C-S representations are still black boxes that are hard to interpret and control [54].", + "bbox": [ + 498, + 517, + 893, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Facing the challenges above, in this paper, we propose a new C-S disentangled framework for style transfer without using previous assumptions such as Gram matrix [18] or GANs [42]. Our key insight stems from the fact that the definition of an image's style is much more complex", + "bbox": [ + 498, + 825, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when Zhizhong Wang was an intern at Huawei.", + "bbox": [ + 91, + 875, + 462, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7677", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "than its content, e.g., we can easily identify the content of a painting by its structures, semantics, or shapes, but it is intractable to define the style [61, 21, 35, 82]. Therefore, we can bypass such a dilemma by explicitly extracting the content information and implicitly learning its complementary style information. Since we strictly constrain style as the complement of content, the C-S can be completely disentangled, and the control of disentanglement has been transformed into the control of content extraction. It achieves both controllability and interpretability.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, achieving plausible and controllable content extraction is also non-trivial because the contents extracted from the content images and style images should share the same content domain, and the details of the extracted contents should be easy to control. To this end, we resort to recent developed diffusion models [25, 73] and introduce a diffusion-based style removal module to smoothly dispel the style information of the content and style images, extracting the domain-aligned content information. Moreover, owing to the strong generative capability of diffusion models, we also introduce a diffusion-based style transfer module to better learn the disentangled style information of the style image and transfer it to the content image. The style disentanglement and transfer are encouraged via a simple yet effective CLIP [62]-based style disentanglement loss, which induces the transfer mapping of the content image's content to its stylization (i.e., the stylized result) to be aligned with that of the style image's content to its stylization (i.e., the style image itself) in the CLIP image space. By further coordinating with a style reconstruction prior, it achieves both generalized and faithful style transfer. We conduct comprehensive comparisons and ablation study to demonstrate the effectiveness and superiority of our framework. With the well-disentangled C-S, it achieves very promising stylizations with fine style details, well-preserved contents, and a deep understanding of the relationship between C-S.", + "bbox": [ + 75, + 242, + 472, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are threefold:", + "bbox": [ + 96, + 635, + 390, + 648 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel C-S disentangled framework for style transfer, which achieves more interpretable and controllable C-S disentanglement and higher-quality stylized results.", + "- We introduce diffusion models to our framework and demonstrate their effectiveness and superiority in controllable style removal and learning well-disentangled C-S characteristics.", + "- A new CLIP-based style disentanglement loss coordinated with a style reconstruction prior is introduced to disentangle C-S in the CLIP image space." + ], + "bbox": [ + 93, + 659, + 468, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 844, + 218, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Style Transfer (NST). The pioneering work of Gatys et al. [18] has opened the era of NST [31]. Since", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "then, this task has experienced tremendous progress, including efficiency [32, 49, 85], quality [22, 70, 84, 52, 10, 7, 1, 43, 78, 53, 6, 87, 29, 94, 12, 91, 79], generality [5, 27, 50, 59, 13, 30, 26, 80, 90, 55, 88], and diversity [75, 81, 83]. Despite these successes, the essence of these approaches is mostly based on the explicitly defined C-S representations, such as Gram matrix [18], which have several limitations as discussed in Sec. 1. In our work, we propose new disentangled C-S representations explicitly extracted or implicitly learned by diffusion models, achieving more effective style transfer and higher-quality results.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Disentangled Representation Learning (DRL). The task of DRL [24] aims at modeling the factors of data variations [48]. Earlier works used labeled data to factorize representations in a supervised manner [34]. Recently, unsupervised settings have been largely explored [39], especially for disentangling style from content [93, 28, 48, 37, 86, 42, 60, 64, 8, 45]. However, due to the dependence on GANs [20], their C-S disentanglement is usually restricted in the GAN pre-defined domains (e.g., Van Gogh's style domain). Besides, disentanglement cannot be effectively achieved without providing sufficient data [54]. In contrast, our framework learns the disentangled style from a single style image, and the disentanglement can be easily achieved by providing only a few ( $\\sim$ 50) content images for training.", + "bbox": [ + 496, + 257, + 892, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models [72] such as denoising diffusion probabilistic models (DDPMs) [25, 57] have recently shown great success in image generation [73, 14, 16], image manipulation [56, 2, 38], and text-conditional synthesis [58, 68, 63, 65, 23, 4, 51]. These works have demonstrated the power of diffusion models to achieve higher-quality results than other generative models like VAEs [76], auto-regressive models [15], flows [41], and GANs [36]. Inspired by them, we introduce a diffusion-based style removal module and a style transfer module in our framework. These modules can smoothly remove the style information of images and better learn the recovery of it to achieve higher-quality style transfer results. To the best of our knowledge, our work is the first to introduce diffusion models to the field of neural style transfer.", + "bbox": [ + 496, + 469, + 893, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 500, + 705, + 627, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Denoising diffusion probabilistic models (DDPMs) [72, 25] are latent variable models that consist of two diffusion processes, i.e., a forward diffusion process and a reverse diffusion process. The forward process is a fixed Markov Chain that sequentially produces a series of latents $x_{1}, \\ldots, x_{T}$ by gradually adding Gaussian noises at each timestep $t \\in [1, T]$ :", + "bbox": [ + 496, + 731, + 890, + 838 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) := \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 845, + 890, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\beta_{t} \\in (0,1)$ is a fixed variance schedule. An important property of the forward process is that given clean data", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7678", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eefdcc0c02e6e3253a5130b9b1f512b64ec36b470b7a294df09b936ff28bea86.jpg", + "image_caption": [ + "Figure 1. Overview of our proposed StyleDiffusion. The content image $I_{c}$ and style image $I_{s}$ are first fed into a diffusion-based style removal module to explicitly extract the domain-aligned content information. Then, the content of $I_{c}$ is fed into a diffusion-based style transfer module to obtain the stylized result $I_{cs}$ . During training, we fine-tune the style transfer module via a CLIP-based style disentanglement loss $\\mathcal{L}_{SD}$ coordinated with a style reconstruction prior (see details in Sec. 4.3, we omit it here for brevity) to implicitly learn the disentangled style information of $I_{s}$ ." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 872, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$x_0, x_t$ can be directly sampled as:", + "bbox": [ + 76, + 443, + 305, + 458 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {0}\\right) := \\mathcal {N} (\\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 468, + 468, + 494 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} := \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon ,\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 489, + 375, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t} := 1 - \\beta_{t}$ and $\\bar{\\alpha}_{t} := \\prod_{s=0}^{t} \\alpha_{s}$ . Noise $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ has the same dimensionality as data $x_{0}$ and latent $x_{t}$ .", + "bbox": [ + 75, + 518, + 468, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The reverse process generates a reverse sequence by sampling the posteriors $q(x_{t - 1}|x_t)$ , starting from a Gaussian noise sample $x_{T} \\sim \\mathcal{N}(0,\\mathbf{I})$ . However, since $q(x_{t - 1}|x_t)$ is intractable, DDPMs learn parameterized Gaussian transitions $p_{\\theta}(x_{t - 1}|x_t)$ with a learned mean $\\mu_{\\theta}(x_t,t)$ and a fixed variance $\\sigma_t^2\\mathbf{I}$ [25]:", + "bbox": [ + 75, + 551, + 468, + 642 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (x _ {t - 1} | x _ {t}) := \\mathcal {N} (\\mu_ {\\theta} (x _ {t}, t), \\sigma_ {t} ^ {2} \\mathbf {I}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 652, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}(x_t,t)$ is the function of a noise approximator $\\epsilon_{\\theta}(x_t,t)$ . Then, the reverse process can be expressed as:", + "bbox": [ + 76, + 681, + 468, + 712 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} := \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right) + \\sigma_ {t} \\mathbf {z}, \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 720, + 468, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})$ is a standard Gaussian noise independent of $x_{t}$ . $\\epsilon_{\\theta}(x_{t}, t)$ is learned by a deep neural network [66] through optimizing the following loss:", + "bbox": [ + 76, + 765, + 468, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\| \\epsilon_ {\\theta} (x _ {t}, t) - \\epsilon \\| ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 821, + 468, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Later, instead of using the fixed variances, Nichol and Dhariwal [57] presented a strategy for learning the variances. Song et al. [73] proposed DDIM, which formulates", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an alternative non-Markovian noising process that has the same forward marginals as DDPM but allows a different reverse process:", + "bbox": [ + 498, + 443, + 890, + 488 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} := \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (x _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} (x _ {t}, t) + \\sigma_ {t} \\mathbf {z}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 500, + 890, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{\\theta}(x_t, t)$ is the predicted $x_0$ at timestep $t$ given $x_t$ and $\\epsilon_{\\theta}(x_t, t)$ :", + "bbox": [ + 498, + 537, + 890, + 568 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} (x _ {t}, t) := \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (x _ {t} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 577, + 890, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Changing the choice of $\\sigma_{t}$ values in Eq. (6) can achieve different reverse processes. Especially when $\\sigma_{t} = 0$ , which is called DDIM [73], the reverse process becomes a deterministic mapping from latents to images, which enables nearly perfect inversion [38]. Besides, it can also accelerate the reverse process with much fewer sampling steps [14, 38].", + "bbox": [ + 496, + 619, + 890, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 724, + 589, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our task can be described as follows: given a style image $I_{s}$ and an arbitrary content image $I_{c}$ , we want to first disentangle the content and style of them and then transfer the style of $I_{s}$ to the content of $I_{c}$ . To achieve so, as stated in Sec. 1, our key idea is to explicitly extract the content information and then implicitly learn the complementary style information. Since our framework is built upon diffusion models [25, 73], we dub it StyleDiffusion.", + "bbox": [ + 496, + 750, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 1 shows the overview of our StyleDiffusion, which consists of three key ingredients: I) a diffusion-based style", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7679", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "removal module, II) a diffusion-based style transfer module, and III) a CLIP-based style disentanglement loss coordinated with a style reconstruction prior. In the following subsections, we will introduce each of them in detail.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Style Removal Module", + "text_level": 1, + "bbox": [ + 76, + 161, + 287, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The style removal module aims at removing the style information of the content and style images, explicitly extracting the domain-aligned content information. Any reasonable content extraction operation can be used, depending on how the users define the content. For instance, users may want to use the structural outline as the content, so they can extract the outlines [33, 89] here. However, as discussed in Sec. 1, one challenge is controllability since the control of C-S disentanglement has been transformed into the control of content extraction. To this end, we introduce a diffusion-based style removal module to achieve both plausible and controllable content extraction.", + "bbox": [ + 75, + 184, + 468, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input image, e.g., the style image $I_{s}$ , since the color is an integral part of style [47], our style removal module first removes its color by a commonly used ITU-R 601-2 luma transform [19]. The obtained grayscale image is denoted as $I_{s}^{\\prime}$ . Then, we leverage a pre-trained diffusion model [14] $\\epsilon_{\\theta}$ to remove the style details such as brushstrokes and textures of $I_{s}^{\\prime}$ , extracting the content $I_{s}^{c}$ . The insight is that the pre-trained diffusion model can help eliminate the domain-specific characteristics of input images and align them to the pre-trained domain [11, 38]. We assume that images with different styles belong to different domains, but their contents should share the same domain. Therefore, we can pre-train the diffusion model on a surrogate domain, e.g., the photograph domain, and then use this domain to construct the contents of images. After pretraining, the diffusion model can convert the input images from diverse domains to the latents $x$ via the forward process and then inverse them to the photograph domain via the reverse process. In this way, the style characteristics can be ideally dispelled, leaving only the contents of the images.", + "bbox": [ + 75, + 366, + 470, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, in order to obtain the results with fewer sampling steps and ensure that the content structures of the input images can be well preserved, we adopt the deterministic DDIM [73] sampling as the reverse process (Eq. (8)), and the ODE approximation of its reversal [38] as the forward process (Eq. (9)):", + "bbox": [ + 75, + 667, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (x _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\epsilon_ {\\theta} (x _ {t}, t), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 768, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t + 1} = \\sqrt {\\bar {\\alpha} _ {t + 1}} f _ {\\theta} (x _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t + 1}} \\epsilon_ {\\theta} (x _ {t}, t), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 800, + 468, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\theta}(x_t,t)$ is defined in Eq. (7). The forward and reverse diffusion processes enable us to easily control the intensity of style removal by adjusting the number of return step $T_{remov}$ (see details in later Sec. 5.1). With the increase of $T_{remov}$ , more style characteristics will be removed, and", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the main content structures are retained, as will be shown in later Sec. 5.3. Note that for content images that are photographs, the diffusion processes are optional since they are already within the pre-trained domain, and there is almost no style except the colors to be dispelled. The superiority of diffusion-based style removal against other operations, such as Auto-Encoder (AE) [50]-based style removal, can be found in supplementary material (SM).", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Style Transfer Module", + "text_level": 1, + "bbox": [ + 498, + 220, + 709, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The style transfer module aims to learn the disentangled style information of the style image and transfer it to the content image. A common generative model like AEs [27] can be used here. However, inspired by the recent great success of diffusion models [14, 38], we introduce a diffusion-based style transfer module, which can better learn the disentangled style information in our framework and achieve higher-quality and more flexible stylizations (see Sec. 5.3).", + "bbox": [ + 496, + 244, + 890, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a content image $I_{c}$ , denote $I_{c}^{c}$ is the content of $I_{c}$ extracted by the style removal module (Sec. 4.1). We first convert it to the latent $x$ using a pre-trained diffusion model $\\epsilon_{\\theta}$ . Then, guided by a CLIP-based style disentanglement loss coordinated with a style reconstruction prior (Sec. 4.3), the reverse process of the diffusion model is fine-tuned $(\\epsilon_{\\theta} \\rightarrow \\epsilon_{\\hat{\\theta}})$ to generate the stylized result $I_{cs}$ referenced by the style image $I_{s}$ . Once the fine-tuning is completed, any content image can be manipulated into the stylized result with the disentangled style of the style image $I_{s}$ . To make the training easier and more stable, we adopt the deterministic DDIM forward and reverse processes in Eq. (8) and Eq. (9) during the fine-tuning. However, at inference, the stochastic DDPM [25] forward process (Eq. (2)) can also be used directly to help obtain diverse results [81] (Sec. 5.3).", + "bbox": [ + 496, + 366, + 892, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Loss Functions and Fine-tuning", + "text_level": 1, + "bbox": [ + 498, + 602, + 779, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Enforcing the style transfer module (Sec. 4.2) to learn and transfer the disentangled style information should address two key questions: (1) \"how to regularize the learned style is disentangled\" and (2) \"how to aptly transfer it to other contents\". To answer these questions, we introduce a novel CLIP-based style disentanglement loss coordinated with a style reconstruction prior to train the networks.", + "bbox": [ + 496, + 625, + 890, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP-based Style Disentanglement Loss. Denote $I_{c}^{c}$ and $I_{s}^{c}$ are the respective contents of the content image $I_{c}$ and the style image $I_{s}$ extracted by the style removal module (Sec. 4.1). We aim to learn the disentangled style information of the style image $I_{s}$ complementary to its content $I_{s}^{c}$ . Therefore, a straightforward way to obtain the disentangled style information is a direct subtraction:", + "bbox": [ + 496, + 731, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {s} ^ {p x} = I _ {s} - I _ {s} ^ {c}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 849, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 Unless otherwise specified, we do not use the diffusion processes for content images in order to better maintain the content structures.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7680", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/43f281c92aa3a206fb934e88e358668a72ab7ec73a0254bb4084c0c2e37e7971.jpg", + "image_caption": [ + "(a) L1 loss" + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 344, + 213 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5ea1ad39b632f0621e39e7da76d120caf1fca0843fe9087d7e933b298a28251f.jpg", + "image_caption": [ + "(b) Direction loss", + "Figure 2. Illustration of different loss functions to transfer the disentangled style information. (a) L1 loss cannot guarantee the stylized results are within the style domain and may suffer from a collapse problem. (b) Direction loss aligns the disentangled directions but cannot realize accurate mappings. (c) Combining L1 loss and direction loss is able to achieve accurate one-to-one mappings from the content domain to the style domain." + ], + "image_footnote": [], + "bbox": [ + 351, + 87, + 616, + 214 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f34e880991651244fe8ad29e70daa1ac748236e07afaecd5c59ef722e061d78.jpg", + "image_caption": [ + "(c) Direction + L1 loss" + ], + "image_footnote": [], + "bbox": [ + 622, + 88, + 890, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the simple pixel differences do not contain meaningful semantic information, thus cannot achieve plausible results [18, 42]. To address this problem, we can formulate the disentanglement in a latent semantic space:", + "bbox": [ + 75, + 323, + 468, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {s} = E \\left(I _ {s}\\right) - E \\left(I _ {s} ^ {c}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 393, + 468, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E$ is a well-pre-trained projector. Specifically, since $I_{s}$ and $I_{s}^{c}$ have similar contents but with different styles, the projector $E$ must have the ability to distinguish them in terms of the style characteristics. In other words, as we define that images with different styles belong to different domains, the projector $E$ should be able to distinguish the domains of $I_{s}$ and $I_{s}^{c}$ . Fortunately, inspired by the recent vision-language model CLIP [62] that encapsulates knowledgeable semantic information of not only the photograph domain but also the artistic domain [17, 63, 46], we can use its image encoder as our projector $E$ off the shelf. The open-domain CLIP space here serves as a good metric space to measure the \"style distance\" between content and its stylized result. This \"style distance\" thus can be interpreted as the disentangled style information. Note that here the style is implicitly defined as the complement of content, which is fundamentally different from the Gram matrix [18] that is an explicit style definition independent of content (see comparisons in Sec. 5.3). The comparisons between CLIP space and other possible spaces can be found in SM.", + "bbox": [ + 75, + 420, + 468, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the disentangled style information $D_{s}$ , the next question is how to properly transfer it to other contents. A possible solution is directly optimizing the L1 loss:", + "bbox": [ + 75, + 723, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {c s} = E \\left(I _ {c s}\\right) - E \\left(I _ {c} ^ {c}\\right), \\\\ \\text {2} I _ {1} = \\text {＂} \\text {D} _ {1} = \\text {＂} \\text {D} _ {2} = \\text {＂} \\end{array} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 777, + 468, + 803 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S D} ^ {L 1} = \\parallel D _ {c s} - D _ {s} \\parallel ,\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 799, + 336, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{cs}$ is the stylized result, $D_{cs}$ is the disentangled style information of $I_{cs}$ . However, as illustrated in Fig. 2 (a) and further validated in later Sec. 5.3, minimizing the L1 loss cannot guarantee the stylized result $I_{cs}$ is within the style domain of the style image $I_{s}$ . It is because L1 loss only", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "minimizes the absolute pixel difference (i.e., Manhattan distance); thus, it may produce stylized images that satisfy the Manhattan distance but deviate from the target style domain in the transfer direction. Besides, it may also lead to a collapse problem where a stylized output meets the same Manhattan distance with different contents in the latent space.", + "bbox": [ + 496, + 323, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these problems, we can further constrain the disentangled directions as follows:", + "bbox": [ + 496, + 414, + 890, + 444 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S D} ^ {d i r} = 1 - \\frac {D _ {c s} \\cdot D _ {s}}{\\| D _ {c s} \\| \\| D _ {s} \\|.} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 454, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This direction loss aligns the transfer direction of the content image's content to its stylization (i.e., the stylized result) with the direction of the style image's content to its stylization (i.e., the style image itself), as illustrated in Fig. 2 (b). Collaborated with this loss, the L1 loss $\\mathcal{L}_{SD}^{L1}$ thus can achieve accurate one-to-one mappings from contents in the content domain to their stylizations in the style domain, as illustrated in Fig. 2 (c).", + "bbox": [ + 496, + 497, + 890, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, our style disentanglement loss is defined as a compound of $\\mathcal{L}_{SD}^{L1}$ and $\\mathcal{L}_{SD}^{dir}$ :", + "bbox": [ + 496, + 618, + 890, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S D} = \\lambda_ {L 1} \\mathcal {L} _ {S D} ^ {L 1} + \\lambda_ {d i r} \\mathcal {L} _ {S D} ^ {d i r}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 659, + 890, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{L1}$ and $\\lambda_{dir}$ are hyper-parameters set to 10 and 1 in our experiments. Since our style information is induced by the difference between content and its stylized result, we can deeply understand the relationship between C-S through learning. As a result, the style can be naturally and harmoniously transferred to the content, leading to better stylized images, as will be shown in later Fig. 3.", + "bbox": [ + 496, + 689, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Style Reconstruction Prior. To fully use the prior information provided by the style image and further elevate the stylization effects, we integrate a style reconstruction prior into the fine-tuning of the style transfer module. Intuitively, given the content $I_{s}^{c}$ of the style image $I_{s}$ , the style transfer module should be capable of recovering it to the original style image as much as possible. Therefore, we can define", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7681", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9f812457c75c9843b7d0b042d9757d602fee9d29caf8180363d673706d56a6ca.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons with state of the art. Zoom-in for better comparison. Please see more in SM." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 893, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a style reconstruction loss as follows:", + "bbox": [ + 76, + 571, + 326, + 585 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S R} = \\left\\| I _ {s s} - I _ {s} \\right\\|, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 599, + 468, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I_{ss}$ is the stylized result given $I_s^c$ as content. We optimize it separately before optimizing the style disentanglement loss $\\mathcal{L}_{SD}$ . The detailed fine-tuning procedure can be found in SM. The style reconstruction prior helps our model recover the style information more sufficiently. It also provides a good initialization for the optimization of $\\mathcal{L}_{SD}$ , which helps the latter give full play to its ability, thus producing higher-quality results (see later Sec. 5.3).", + "bbox": [ + 75, + 627, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 761, + 282, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 786, + 292, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use ADM diffusion model [14] pre-trained on ImageNet [67] and adopt a fast sampling strategy [38]. Specifically, instead of sequentially conducting the diffusion processes until the last timestep $T$ (e.g., 1000), we accelerate them by performing up to $T_{\\{\\cdot\\}} < T$ (which is called return step), i.e., $T_{\\text{remov}} = 601$ for style removal and $T_{\\text{trans}} =$", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "301 for style transfer. Moreover, as suggested by [38], we further accelerate the forward and reverse processes with fewer discretization steps, i.e., $(S_{for}, S_{rev}) = (40, 40)$ ( $S_{for}$ for forward process and $S_{rev}$ for reverse process) for style removal, and $(S_{for}, S_{rev}) = (40, 6)$ for style transfer. When fine-tuning or inference, we can adjust $T_{remov}$ or $T_{trans}$ to flexibly control the degree of style removal and C-S disentanglement, as will be shown in Sec. 5.3. To fine-tune the model for a target style image, we randomly sample 50 images from ImageNet as the content images. We use Adam optimizer [40] with an initial learning rate of 4e-6 and increase it linearly by 1.2 per epoch. All models are fine-tuned with 5 epochs. See more details in SM.", + "bbox": [ + 496, + 571, + 890, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparisons with Prior Arts", + "text_level": 1, + "bbox": [ + 500, + 782, + 759, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our StyleDiffusion against ten state-of-the-art (SOTA) methods [18, 90, 12, 1, 53, 6, 13, 74, 32, 52]. For fair comparisons, all these methods are fine-tuned or trained on the target styles similar to our approach.", + "bbox": [ + 496, + 808, + 890, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. As can be observed in Fig. 3, due to the entangling of C-S representations, Gatys [18]", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7682", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c725bbdaaef55b1f5995fbbc3f387bbabdbb212ea06cabd629e8356fdc882b66.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OursGatysEFDMStyTr2ArtFlowAdaAttNIECASTMASTTPFRJohnsonLapStyle
SSIM ↑0.6720.3110.3160.5370.5010.5420.3650.3920.5360.6340.657
CLIP Score ↑0.7410.6770.6070.5310.5460.5770.6460.5900.6440.5370.595
Style Loss ↓0.8370.1110.1780.2160.2580.3100.2840.2290.9890.3640.274
UserStyle-43.1%41.2%39.3%36.4%37.2%33.8%39.1%14.5%42.8%47.3%
StudyOverall-26.0%38.1%44.0%34.2%43.9%32.7%32.2%22.6%43.4%46.2%
Training Time/h~0.4-~3~4~3~3~3~3~10~1~3
Testing Time/s5.61210.1650.0280.1680.2040.0760.0340.0660.3020.0150.008
", + "bbox": [ + 81, + 88, + 883, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons with state of the art. The training/testing time is measured with an Nvidia Tesla A100 GPU, and the testing time is averaged on images of size $512 \\times 512$ pixels. $\\uparrow$ : Higher is better. $\\downarrow$ : Lower is better.", + "bbox": [ + 75, + 203, + 892, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and EFDM [90] often produce unsatisfying results with distorted contents (e.g., rows 1-3) and messy textures (e.g., rows 4-8). StyTr² [12] and ArtFlow [1] improve the results by adopting more advanced networks [77, 41], but they may still produce inferior results with halo boundaries (e.g., rows 2-3) or dirty artifacts (e.g., rows 4-6). AdaAttN [53] performs per-point attentive normalization to preserve the content structures better, but the stylization effects may be degraded in some cases (e.g., rows 1, 2, 4, and 5). IECAST [6] utilizes contrastive learning and external learning for style transfer, so fine-tuning it on a single style image would result in degraded results. MAST [13] uses multi-adaptation networks to disentangle C-S. However, since it still relies on the C-S representations of [18], the results usually exhibit messy textures and conspicuous artifacts. TPFR [74] is a GAN-based framework that learns to disentangle C-S in latent space. As the results show, it cannot recover correct style details and often generates deviated stylizations, which signifies that it may not learn truly disentangled C-S representations [54]. Like our method, Johnson [32] and LapStyle [52] also train separate models for each style. However, due to the trade-off between C-S losses of [18], they may produce less-stylized results or introduce unnatural patterns (e.g., rows 1-6).", + "bbox": [ + 73, + 257, + 472, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By contrast, our StyleDiffusion completely disentangles C-S based on diffusion models. Therefore, it can generate high-quality results with sufficient style details (e.g., rows 1-4) and well-preserved contents (e.g., rows 5-8). Compared with the previous methods that tend to produce mixed results of content and style, our approach can better consider the relationship between them. Thus, the stylizations are more natural and harmonious, especially for challenging styles such as cubism (e.g., row 2) and oil painting (e.g., rows 1, 3, 4, and 5).", + "bbox": [ + 75, + 623, + 468, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparisons. We also resort to quantitative metrics to better evaluate our method, as shown in Tab. 1. We collect 32 content and 12 style images to synthesize 384 stylized results and compute the average Structural Similarity Index (SSIM) [1] to assess the content similarity. To evaluate the style similarity, we calculate the CLIP image similarity score [62] and Style Loss [18, 27] between the style images and the corresponding stylized results. As", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shown in Tab. 1, our method obtains the highest SSIM and CLIP Score while the Style Loss is relatively higher than other methods. It is because these methods are directly trained to optimize Style Loss. Nevertheless, the Style Loss achieved by our method is still comparable and lower than the GAN-based TPFR [74]. Furthermore, it is noteworthy that our method can also incorporate Style Loss to enhance the performance in this regard (see later Sec. 5.3).", + "bbox": [ + 496, + 257, + 893, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. As style transfer is highly subjective and CLIP Score and Style Loss are biased to the training objective, we additionally resort to user study to evaluate the style similarity and overall stylization quality. We randomly select 50 C-S pairs for each user. Given each C-S pair, we show the stylized results generated by our method and a randomly selected SOTA method side by side in random order. The users are asked to choose (1) which result transfers the style patterns better and (2) which result has overall better stylization effects. We obtain 1000 votes for each question from 20 users and show the percentage of votes that existing methods are preferred to ours in Tab. 1. The lower numbers indicate our method is more preferred than the competitors. As the results show, our method is superior to others in both style consistency and overall quality.", + "bbox": [ + 496, + 378, + 895, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency. As shown in the bottom two rows of Tab. 1, our approach requires less training time than others as it is fine-tuned on only a few ( $\\sim 50$ ) content images. When testing, our approach is faster than the optimization-based method Gatys [18], albeit slower than the remaining feedforward methods due to the utilization of diffusion models. We discuss it in later Sec. 6, and more timing and resource details can be found in SM.", + "bbox": [ + 496, + 606, + 893, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 739, + 653, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Control of C-S Disentanglement. A prominent advantage of our StyleDiffusion is that we can flexibly control the C-S disentanglement by adjusting the content extraction of the style removal module (Sec. 4.1). Fig. 4 demonstrates the continuous control achieved by adjusting the return step $T_{\\text{remov}}$ of the style removal module. As shown in the top row, with the increase of $T_{\\text{remov}}$ , more style characteristics are dispelled, and the main content structures are retained. Correspondingly, when more style is removed in the", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7683", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e7d8272db81e048072bbe46ba61b8de4c2bd0ba6d9c5d1a519a02fb83c09a8a.jpg", + "image_caption": [ + "Figure 4. Control of C-S disentanglement by adjusting the return step $T_{\\text{remov}}$ of the style removal module. The top row shows the extracted contents of the style image. The bottom row shows the corresponding stylized results. * denotes our default setting. Zoom-in for better comparison. See SM for quantitative analyses." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "top row, it will be aptly transferred to the stylized results in the bottom row, e.g., the twisted brushstrokes and the star patterns. It validates that our method successfully separates style from content in a controllable manner and properly transfers it to other contents. Moreover, the flexible C-S disentanglement also makes our StyleDiffusion versatile for other tasks, such as photo-realistic style transfer (see SM).", + "bbox": [ + 75, + 324, + 467, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Superiority of Diffusion-based Style Transfer. Although our style transfer module is not limited to the diffusion model, using it offers three main advantages: (1) Flexible C-S trade-off control. As shown in Fig. 5, we can flexibly control the C-S trade-off at both the training stage (top row) and the testing stage (bottom row) by adjusting the return step $T_{trans}$ of the diffusion model. With the increase of $T_{trans}$ , more style characteristics are transferred, yet the content structures may be ruined (e.g., the last column). When proper $T_{trans}$ is adopted, e.g., $T_{trans} = 301$ , the sweet spot can be well achieved. Interestingly, as shown in the last two columns of the bottom row, though the model is trained on $T_{trans} = 301$ , we can extrapolate the style by using larger $T_{trans}$ (e.g., 401) at the testing stage (but the results may be degraded when using too large $T_{trans}$ , e.g., 601). It provides a very flexible way for users to adjust the results according to their preferences. This property, however, cannot be simply achieved by using other models, e.g., the widely used AEs [27, 50], since our framework does not involve any feature transforms [27, 50] or C-S losses trade-off [3]. (2) Higher-quality stylizations. Owing to the strong generative ability of the diffusion model, it can achieve higher-quality stylizations than other models. For comparison, we use the pre-trained VGG-AE [27, 46] as the style transfer module and fine-tune its decoder network for each style. As shown in column (b) of Fig. 6, though the results are still acceptable, they may produce distorted contents and inferior textures, clearly worse than the results generated by the diffusion model in column (a). This is also validated by the bottom quantitative scores. It signifies that the diffusion model can better learn the disentangled", + "bbox": [ + 75, + 431, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6397c032e8816519cced80362c26afb7c20fc6b9c75b61afce656a5be8fa0cf.jpg", + "image_caption": [ + "Figure 5. Control of C-S trade-off by adjusting the return step $T_{trans}$ of the style transfer module. The top row shows adjusting $T_{trans}$ at the training stage while fixing $T_{trans} = 301$ at the testing stage. The bottom row shows adjusting $T_{trans}$ at the testing stage while fixing $T_{trans} = 301$ at the training stage. * denotes our default setting. Zoom-in for better comparison. See SM for quantitative analyses." + ], + "image_footnote": [], + "bbox": [ + 503, + 88, + 890, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de144b610ab417f1fdbba4a40504f3455bc799d6718e8da644e93de0354cf834.jpg", + "image_caption": [ + "Figure 6. Diffusion-based vs. AE-based style transfer." + ], + "image_footnote": [], + "bbox": [ + 503, + 338, + 694, + 515 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ed65172626828e48354751e2d7c738a8585ea16633fc9543a789e0057e9741b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 338, + 890, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "content and style characteristics in our framework, helping produce better style transfer results. (3) Diversified style transfer. As mentioned in Sec. 4.2, during inference, we can directly adopt the stochastic DDPM [25] forward process (Eq. (2)) to obtain diverse results (see SM). The diverse results can give users endless choices to obtain more satisfactory results. However, using other models like AEs in our framework cannot easily achieve it [81].", + "bbox": [ + 496, + 566, + 890, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss Analyses. To verify the effectiveness of each loss term used for fine-tuning our StyleDiffusion, we present ablation study results in Fig. 7 (a-d). (1) Using L1 loss $\\mathcal{L}_{SD}^{L1}$ successfully transfers the cubism style like the blocky patterns in the top row, but the colors stray from the style images, especially in the bottom row. It is consistent with our earlier analyses in Sec. 4.3 that the L1 loss is prone to produce implausible results outside the style domain. (2) Adding direction loss $\\mathcal{L}_{SD}^{dir}$ helps pull the results closer to the style domain. The textures are enhanced in the top row, and the colors are more plausible in the top and bottom rows. (3) By further coordinating with the style reconstruction prior $\\mathcal{L}_{SR}$ , the stylization effects are significantly elevated where the style information is recovered more sup", + "bbox": [ + 496, + 689, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7684", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f52741d8174e9d4f012e5b04328b8f0f80d528a3b5b9ee92eaf7d4f4d9f0cfaa.jpg", + "image_caption": [ + "Figure 7. Ablation study on loss functions. * denotes our full model. Zoom-in for better comparison." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 279, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/290b33325c8f5254e246f04302a417f9b7d45eac542a7530f3b2edcda07bbd0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 88, + 387, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3aeab92b7b3c4a9375ce0fade7133235f5164b55e04c3c5ba1afa693f81bc3a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 89, + 486, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3c44f8bc9590fa95abe69a8bcf01cdf08115889609fc9967eafe80b815a43aae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 90, + 586, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/74f0f139e37338b03d1763dfcd660f2d625447ca0b0f9964d858e9f0b402af57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 90, + 686, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c1e0a8238a89ea23aebe0dac3a707991c8dcbbac643edae9c529e593c9041d60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 89, + 792, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/76cf3ba0e08d752ca646da0a3b51039468eb169f5119c2d672e4b58bce884899.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 90, + 892, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ficiently. It may be because it provides a good initialization for the optimization of $\\mathcal{L}_{SD}^{L1}$ and $\\mathcal{L}_{SD}^{dir}$ , which helps them give full play to their abilities. As verified in Fig. 7 (d), using the style reconstruction alone cannot learn meaningful style patterns except for basic colors. All the above analyses are also supported by the bottom quantitative scores.", + "bbox": [ + 75, + 319, + 467, + 409 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparison with Gram Loss. To further verify the superiority of our proposed losses, we replace them with the widely used Gram Loss [18, 27] in Fig. 7 (e-f). As can be observed, Gram Loss destroys the content structures severely, e.g., the zebra head in the top row and the enlarged area in the bottom row. This is because it does not disentangle C-S and only matches the global statistics without considering the relationship between C-S. In contrast, our losses focus on learning the disentangled style information apart from the content, which is induced by the difference between content and its stylized result. Therefore, they can better understand the relationship between C-S, achieving more satisfactory results with fine style details and better-preserved contents, as validated by Fig. 7 (c) and the bottom quantitative scores. Furthermore, we also conduct comparisons between our proposed losses and Gram Loss [18, 27] on the AE baseline [27, 46] to eliminate the impact of diffusion models. As shown in Fig. 8 (a-b), our losses can achieve more satisfactory results than Gram Loss, which is consistent with the results in Fig. 7. Moreover, as shown in Fig. 8 (c), they can also be combined with Gram Loss to improve the performance on the Style Loss metric. However, it may affect the full disentanglement of C-S in our framework, which strays from our target and decreases the content preservation (see SSIM score in Fig. 8 (c)). Therefore, we do not incorporate Gram Loss in our framework by default.", + "bbox": [ + 75, + 410, + 467, + 816 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 76, + 830, + 323, + 845 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we present a new framework for more interpretable and controllable C-S disentanglement and style transfer. Our framework, termed StyleDiffusion, leverages", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/39eefed754ea71adc992013a9465c3efc10cfaa82c74b273da2b5bcdadf0ab85.jpg", + "image_caption": [ + "Figure 8. More loss function ablation study on the AE baseline." + ], + "image_footnote": [], + "bbox": [ + 501, + 316, + 658, + 492 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0a6ea978a6ec74a27740154edef2bed2cabdb04889ae7ec647d485548beaf9b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 316, + 736, + 492 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/de7a1aac3d645bb0bba6fdff43f6a046fc2dad5446037e41a1ece42899b0a3f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 316, + 813, + 491 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/91ca077552027d6c6f07d7947471b6b15a217c01a601c2cffbaa33cca39b735b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 815, + 316, + 892, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "diffusion models to explicitly extract the content information and implicitly learn the complementary style information. A novel CLIP-based style disentanglement loss coordinated with a style reconstruction prior is also introduced to encourage the disentanglement and style transfer. Our method yields very encouraging stylizations, especially for challenging styles, and the experimental results verify its effectiveness and superiority against state of the art.", + "bbox": [ + 496, + 539, + 890, + 659 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Currently, the framework still suffers from several limitations: (1) The model needs to be fine-tuned for each style, and arbitrary style transfer is left to our future work. (2) The efficiency is not fast enough due to the use of diffusion models. Further research in accelerating diffusion sampling would be helpful. (3) There are some failure cases analyzed in SM, which may help inspire future improvements. Moreover, our framework may also be applied to other image translation [28] or manipulation [60] tasks, and we would like to explore them in our future work.", + "bbox": [ + 496, + 661, + 890, + 811 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "Acknowledgments. We thank Zeyi Huang and Xiaoting Zhang for their insightful discussions and suggestions. This work was supported in part by the National Program of China (2020YFC1523201, 62172365, 19ZDA197), Zhejiang Elite Program (2022C01222), and Key Technologies and Product Research and Development Projects for Cultural Relics Protection and Trading Circulation.", + "bbox": [ + 500, + 827, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7685", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 862-871, 2021. 1, 2, 6, 7", + "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 2", + "[3] Mohammad Babaeizadeh and Golnaz Ghiasi. Adjustable real-time style transfer. In International Conference on Learning Representations (ICLR), 2019. 8", + "[4] Andreas Blattmann, Robin Rombach, Kaan Oktay, Jonas Müller, and Björn Ommer. Retrieval-augmented diffusion models. Advances in Neural Information Processing Systems, 35:15309-15324, 2022. 2", + "[5] Dongdong Chen, Lu Yuan, Jing Liao, Nenghai Yu, and Gang Hua. Stylebank: An explicit representation for neural image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1897-1906, 2017. 2", + "[6] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems (NeurIPS), 34:26561-26573, 2021. 1, 2, 6, 7", + "[7] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dualist: Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 872-881, 2021. 2", + "[8] Haibo Chen, Lei Zhao, Huiming Zhang, Zhizhong Wang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Diverse image style transfer via invertible cross-space mapping. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14860-14869. IEEE Computer Society, 2021. 2", + "[9] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in Neural Information Processing Systems (NeurIPS), 29, 2016. 1", + "[10] Jiaxin Cheng, Ayush Jaiswal, Yue Wu, Pradeep Natarajan, and Prem Natarajan. Style-aware normalized loss for improving arbitrary style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 134-143, 2021. 2", + "[11] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14347-14356. IEEE, 2021. 4", + "[12] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Im" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "age style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11326-11336, 2022. 2, 6, 7", + "[13] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM International Conference on Multimedia (ACM MM), pages 2719-2727, 2020. 2, 6, 7", + "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021. 2, 3, 4, 6", + "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2", + "[16] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 579-587, 2023. 2", + "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 5", + "[18] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 1, 2, 5, 6, 7, 9", + "[19] Rafael C Gonzalez. Digital image processing. Pearson education india, 2009. 4", + "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 1, 2", + "[21] Daniel J Graham, James M Hughes, Helmut Leder, and Daniel N Rockmore. Statistics, vision, and the analysis of artistic style. Wiley Interdisciplinary Reviews: Computational Statistics, 4(2):115-123, 2012. 2", + "[22] Shuyang Gu, Congliang Chen, Jing Liao, and Lu Yuan. Arbitrary style transfer with deep feature reshuffle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8222-8231, 2018. 2", + "[23] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2", + "[24] Irina Higgins, David Amos, David Pfau, Sebastien Racaniere, Loic Matthew, Danilo Rezende, and Alexander Lerchner. Towards a definition of disentangled representations. arXiv preprint arXiv:1812.02230, 2018. 1, 2", + "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information" + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7686", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Processing Systems (NeurIPS), 33:6840-6851, 2020. 2, 3, 4, 8", + "[26] Kibeom Hong, Seogkyu Jeon, Huan Yang, Jianlong Fu, and Hyeran Byun. Domain-aware universal style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[27] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1501-1510, 2017. 1, 2, 4, 7, 8, 9", + "[28] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 172-189, 2018. 2, 9", + "[29] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14861-14869, 2021. 2", + "[30] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4369-4376, 2020. 2", + "[31] Yongcheng Jing, Yezhou Yang, Zunlei Feng, Jingwen Ye, Yizhou Yu, and Mingli Song. Neural style transfer: A review. IEEE Transactions on Visualization and Computer Graphics (TVCG), 26(11):3365-3385, 2019. 2", + "[32] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711. Springer, 2016. 1, 2, 6, 7", + "[33] Henry Kang, Seungyong Lee, and Charles K Chui. Coherent line drawing. In Proceedings of the 5th international symposium on Non-photorealistic animation and rendering, pages 43–50, 2007. 4", + "[34] Theofanis Karaletsos, Serge Belongie, and Gunnar Ratsch. Bayesian representation learning with oracle constraints. arXiv preprint arXiv:1506.05011, 2015. 1, 2", + "[35] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, and Holger Winnemoeller. Recognizing image style. arXiv preprint arXiv:1311.3715, 2013. 2", + "[36] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019. 2", + "[37] Hadi Kazemi, Seyed Mehdi Iranmanesh, and Nasser Nasrabadi. Style and content disentanglement in generative adversarial networks. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 848-856. IEEE, 2019. 2", + "[38] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition (CVPR), pages 2426-2435, 2022. 2, 3, 4, 6", + "[39] Hyunjik Kim and Andriy Mnih. Disentangling by factorising. In International Conference on Machine Learning (ICML), pages 2649-2658. PMLR, 2018. 2", + "[40] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6", + "[41] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 2, 7", + "[42] Dmytro Kotovenko, Artsiom Sanakoyeu, Sabine Lang, and Bjorn Ommer. Content and style disentanglement for artistic style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4422-4431, 2019. 1, 2, 5", + "[43] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Bjorn Ommer. Rethinking style transfer: From pixels to parameterized brushstrokes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12196-12205, 2021. 2", + "[44] Tejas D Kulkarni, William F Whitney, Pushmeet Kohli, and Josh Tenenbaum. Deep convolutional inverse graphics network. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 1", + "[45] Gihyun Kwon and Jong Chul Ye. Diagonal attention and style-based gan for content-style disentanglement in image generation and translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13980-13989, 2021. 2", + "[46] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18062-18071, 2022. 5, 8, 9", + "[47] Berel Lang. The concept of style. Cornell University Press, 1987. 4", + "[48] Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Maneesh Singh, and Ming-Hsuan Yang. Diverse image-to-image translation via disentangled representations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 35-51, 2018. 2", + "[49] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[50] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. In Advances in Neural Information Processing Systems (NeurIPS), pages 386-396, 2017. 1, 2, 4, 8", + "[51] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2", + "[52] Tianwei Lin, Zhuoqi Ma, Fu Li, Dongliang He, Xin Li, Errui Ding, Nannan Wang, Jie Li, and Xinbo Gao. Drafting and" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7687", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "revision: Laplacian pyramid network for fast high-quality artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5141-5150, 2021. 2, 6, 7", + "[53] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6649-6658, 2021. 2, 6, 7", + "[54] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In International Conference on Machine Learning (ICML), pages 4114-4124. PMLR, 2019. 1, 2, 7", + "[55] Haofei Lu and Zhizhong Wang. Universal video style transfer via crystallization, separation, and blending. In Proceedings of the International Joint Conferences on Artificial Intelligence Organization (IJCAI), Vienna, Austria, pages 23-29, 2022. 2", + "[56] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2", + "[57] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171. PMLR, 2021. 2, 3", + "[58] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), pages 16784-16804. PMLR, 2022. 2", + "[59] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5880–5888, 2019. 2", + "[60] Taesung Park, Jun-Yan Zhu, Oliver Wang, Jingwan Lu, Eli Shechtman, Alexei Efros, and Richard Zhang. Swapping autoencoder for deep image manipulation. Advances in Neural Information Processing Systems (NeurIPS), 33:7198-7211, 2020. 2, 9", + "[61] DM Parker and Jan B Deregowski. Perception and artistic style. Elsevier, 1991. 2", + "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 5, 7", + "[63] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Xuanchi Ren, Tao Yang, Yuwang Wang, and Wenjun Zeng. Rethinking content and style: Exploring bias for unsupervised disentanglement. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1823-1832. IEEE, 2021. 1, 2", + "[65] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2", + "[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-assisted Intervention, pages 234-241. Springer, 2015. 3", + "[67] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 6", + "[68] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[69] Artsiom Sanakoyeu, Dmytro Kotovenko, Sabine Lang, and Bjorn Ommer. A style-aware content loss for real-time hd style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), pages 698-714, 2018. 1", + "[70] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8242-8250, 2018. 2", + "[71] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1", + "[72] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICML), pages 2256-2265. PMLR, 2015. 2", + "[73] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2020. 2, 3, 4", + "[74] Jan Svoboda, Asha Anoosheh, Christian Osendorfer, and Jonathan Masci. Two-stage peer-regularized feature recombination for arbitrary image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13816–13825, 2020. 1, 6, 7", + "[75] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6924-6932, 2017. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7688", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2", + "[77] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 7", + "[78] Pei Wang, Yijun Li, and Nuno Vasconcelos. Rethinking and improving the robustness of image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 124-133, 2021. 2", + "[79] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: Towards aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia (ACM MM), pages 1095-1106, 2022. 2", + "[80] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2624-2632, 2022. 2", + "[81] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7789-7798, 2020. 2, 4, 8", + "[82] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Evaluate and improve the quality of neural style transfer. Computer Vision and Image Understanding (CVIU), 207:103203, 2021. 2", + "[83] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. DivSwapper: Towards diversified patch-based arbitrary style transfer. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), pages 4980–4987, 2022. 2", + "[84] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Glstylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8):575-586, 2020. 2", + "[85] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2742-2750, 2023. 2", + "[86] Wayne Wu, Kaidi Cao, Cheng Li, Chen Qian, and Chen Change Loy. Disentangling content and style via unsupervised geometry distillation. arXiv preprint arXiv:1905.04538, 2019. 2", + "[87] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14618-14627, 2021. 2", + "[88] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[89] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1395-1403, 2015. 4", + "[90] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8035-8045, 2022. 1, 2, 6, 7", + "[91] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 2", + "[92] Yu Zhang, Peter Tino, Ales Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1", + "[93] Yexun Zhang, Ya Zhang, and Wenbin Cai. Separating style and content for generalized style transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8447-8455, 2018. 1, 2", + "[94] Zhiwen Zuo, Lei Zhao, Shuobin Lian, Haibo Chen, Zhizhong Wang, Ailin Li, Wei Xing, and Dongming Lu. Style fader generative adversarial networks for style degree controllable artistic style transfer. In Proc. Int. Joint Conf. on Artif. Intell.(IJCAI), pages 5002-5009, 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 458 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "7689", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/fea588fd-ec6f-4543-ad58-5dd9ea412b01_model.json b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/fea588fd-ec6f-4543-ad58-5dd9ea412b01_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e747b4e8318bde233202961fb3208dd4841b6808 --- /dev/null +++ b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/fea588fd-ec6f-4543-ad58-5dd9ea412b01_model.json @@ -0,0 +1,2932 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.88, + 0.153 + ], + "angle": 0, + "content": "StyleDiffusion: Controllable Disentangled Style Transfer via Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.181, + 0.75, + 0.234 + ], + "angle": 0, + "content": "Zhizhong Wang*, Lei Zhao† Wei Xing \nCollege of Computer Science and Technology, Zhejiang University \n{endywon, czhl, WXING}@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.618 + ], + "angle": 0, + "content": "Content and style (C-S) disentanglement is a fundamental problem and critical challenge of style transfer. Existing approaches based on explicit definitions (e.g., Gram matrix) or implicit learning (e.g., GANs) are neither interpretable nor easy to control, resulting in entangled representations and less satisfying results. In this paper, we propose a new C-S disentangled framework for style transfer without using previous assumptions. The key insight is to explicitly extract the content information and implicitly learn the complementary style information, yielding interpretable and controllable C-S disentanglement and style transfer. A simple yet effective CLIP-based style disentanglement loss coordinated with a style reconstruction prior is introduced to disentangle C-S in the CLIP image space. By further leveraging the powerful style removal and generative ability of diffusion models, our framework achieves superior results than state of the art and flexible C-S disentanglement and trade-off control. Our work provides new insights into the C-S disentanglement in style transfer and demonstrates the potential of diffusion models for learning well-disentangled C-S characteristics." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.646, + 0.21, + 0.661 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.791 + ], + "angle": 0, + "content": "Given a reference style image, e.g., Starry Night by Vincent Van Gogh, style transfer aims to transfer its artistic style, such as colors and brushstrokes, to an arbitrary content target. To achieve such a goal, it must first properly separate the style from the content and then transfer it to another content. This raises two fundamental challenges: (1) \"how to disentangle content and style (C-S)\" and (2) \"how to transfer style to another content\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.47, + 0.868 + ], + "angle": 0, + "content": "To resolve these challenges, valuable efforts have been devoted. Gatys et al. [18] proposed A Neural Algorithm of Artistic Style to achieve style transfer, which explicitly defines the high-level features extracted from a pre-trained Convolutional Neural Network (CNN) (e.g., VGG [71]) as" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.27, + 0.895, + 0.513 + ], + "angle": 0, + "content": "content, and the feature correlations (i.e., Gram matrix) as style. This approach acquires visually stunning results and inspires a large number of successors [32, 27, 50, 1, 90]. Despite the successes, by diving into the essence of style transfer, we observed three problems with these approaches: (1) The C-S are not completely disentangled. Theoretically, the C-S representations are intertwined. For example, matching the content representation of an image may also match its Gram matrix, and vice versa. (2) What CNN learned is a black box rugged to interpret [92], which makes the C-S definitions [18] uninterpretable and hard to control. (3) The transfer process is modeled as a separate optimization of content loss and style loss [18], so there lacks a deep understanding of the relationship between C-S. These problems usually lead to unbalanced stylizations and disharmonious artifacts [6], as will be shown in later Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.895, + 0.821 + ], + "angle": 0, + "content": "On the other hand, disentangled representation learning [24] provides other ideas to implicitly disentangle C-S, either supervised [44, 34] or unsupervised [9, 93]. For style transfer, Kotovenko et al. [42] utilized fixpoint triplet style loss and disentanglement loss to enforce a GAN [20]-based framework to learn separate C-S representations in an unsupervised manner. Similarly, TPFR [74] learned to disentangle C-S in latent space via metric learning and two-stage peer-regularization, producing high-quality images even in the zero-shot setting. While these approaches successfully enforce properties \"encouraged\" by the corresponding losses, they still have three main problems: (1) Well-disentangled models seemingly cannot be identified without supervision [54, 64], which means the unsupervised learning [42, 74] may not achieve truly disentangled C-S, as will be shown in later Fig. 3. (2) These approaches are all based on GANs and thus often confined to the GAN predefined domains, e.g., a specific artist's style domain [69]. (3) The implicitly learned C-S representations are still black boxes that are hard to interpret and control [54]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Facing the challenges above, in this paper, we propose a new C-S disentangled framework for style transfer without using previous assumptions such as Gram matrix [18] or GANs [42]. Our key insight stems from the fact that the definition of an image's style is much more complex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.875, + 0.463, + 0.888 + ], + "angle": 0, + "content": "*This work was done when Zhizhong Wang was an intern at Huawei." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.875, + 0.463, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7677" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "than its content, e.g., we can easily identify the content of a painting by its structures, semantics, or shapes, but it is intractable to define the style [61, 21, 35, 82]. Therefore, we can bypass such a dilemma by explicitly extracting the content information and implicitly learning its complementary style information. Since we strictly constrain style as the complement of content, the C-S can be completely disentangled, and the control of disentanglement has been transformed into the control of content extraction. It achieves both controllability and interpretability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.635 + ], + "angle": 0, + "content": "However, achieving plausible and controllable content extraction is also non-trivial because the contents extracted from the content images and style images should share the same content domain, and the details of the extracted contents should be easy to control. To this end, we resort to recent developed diffusion models [25, 73] and introduce a diffusion-based style removal module to smoothly dispel the style information of the content and style images, extracting the domain-aligned content information. Moreover, owing to the strong generative capability of diffusion models, we also introduce a diffusion-based style transfer module to better learn the disentangled style information of the style image and transfer it to the content image. The style disentanglement and transfer are encouraged via a simple yet effective CLIP [62]-based style disentanglement loss, which induces the transfer mapping of the content image's content to its stylization (i.e., the stylized result) to be aligned with that of the style image's content to its stylization (i.e., the style image itself) in the CLIP image space. By further coordinating with a style reconstruction prior, it achieves both generalized and faithful style transfer. We conduct comprehensive comparisons and ablation study to demonstrate the effectiveness and superiority of our framework. With the well-disentangled C-S, it achieves very promising stylizations with fine style details, well-preserved contents, and a deep understanding of the relationship between C-S." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.636, + 0.392, + 0.65 + ], + "angle": 0, + "content": "In summary, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.66, + 0.469, + 0.72 + ], + "angle": 0, + "content": "- We propose a novel C-S disentangled framework for style transfer, which achieves more interpretable and controllable C-S disentanglement and higher-quality stylized results." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.724, + 0.469, + 0.782 + ], + "angle": 0, + "content": "- We introduce diffusion models to our framework and demonstrate their effectiveness and superiority in controllable style removal and learning well-disentangled C-S characteristics." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.786, + 0.469, + 0.832 + ], + "angle": 0, + "content": "- A new CLIP-based style disentanglement loss coordinated with a style reconstruction prior is introduced to disentangle C-S in the CLIP image space." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.66, + 0.469, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.22, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Neural Style Transfer (NST). The pioneering work of Gatys et al. [18] has opened the era of NST [31]. Since" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "then, this task has experienced tremendous progress, including efficiency [32, 49, 85], quality [22, 70, 84, 52, 10, 7, 1, 43, 78, 53, 6, 87, 29, 94, 12, 91, 79], generality [5, 27, 50, 59, 13, 30, 26, 80, 90, 55, 88], and diversity [75, 81, 83]. Despite these successes, the essence of these approaches is mostly based on the explicitly defined C-S representations, such as Gram matrix [18], which have several limitations as discussed in Sec. 1. In our work, we propose new disentangled C-S representations explicitly extracted or implicitly learned by diffusion models, achieving more effective style transfer and higher-quality results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.893, + 0.469 + ], + "angle": 0, + "content": "Disentangled Representation Learning (DRL). The task of DRL [24] aims at modeling the factors of data variations [48]. Earlier works used labeled data to factorize representations in a supervised manner [34]. Recently, unsupervised settings have been largely explored [39], especially for disentangling style from content [93, 28, 48, 37, 86, 42, 60, 64, 8, 45]. However, due to the dependence on GANs [20], their C-S disentanglement is usually restricted in the GAN pre-defined domains (e.g., Van Gogh's style domain). Besides, disentanglement cannot be effectively achieved without providing sufficient data [54]. In contrast, our framework learns the disentangled style from a single style image, and the disentanglement can be easily achieved by providing only a few (\\(\\sim\\)50) content images for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.894, + 0.697 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models [72] such as denoising diffusion probabilistic models (DDPMs) [25, 57] have recently shown great success in image generation [73, 14, 16], image manipulation [56, 2, 38], and text-conditional synthesis [58, 68, 63, 65, 23, 4, 51]. These works have demonstrated the power of diffusion models to achieve higher-quality results than other generative models like VAEs [76], auto-regressive models [15], flows [41], and GANs [36]. Inspired by them, we introduce a diffusion-based style removal module and a style transfer module in our framework. These modules can smoothly remove the style information of images and better learn the recovery of it to achieve higher-quality style transfer results. To the best of our knowledge, our work is the first to introduce diffusion models to the field of neural style transfer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.707, + 0.629, + 0.724 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Denoising diffusion probabilistic models (DDPMs) [72, 25] are latent variable models that consist of two diffusion processes, i.e., a forward diffusion process and a reverse diffusion process. The forward process is a fixed Markov Chain that sequentially produces a series of latents \\( x_{1}, \\ldots, x_{T} \\) by gradually adding Gaussian noises at each timestep \\( t \\in [1, T] \\):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.846, + 0.892, + 0.865 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) := \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\beta_{t} \\in (0,1)\\) is a fixed variance schedule. An important property of the forward process is that given clean data" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7678" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.873, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.349, + 0.895, + 0.42 + ], + "angle": 0, + "content": "Figure 1. Overview of our proposed StyleDiffusion. The content image \\( I_{c} \\) and style image \\( I_{s} \\) are first fed into a diffusion-based style removal module to explicitly extract the domain-aligned content information. Then, the content of \\( I_{c} \\) is fed into a diffusion-based style transfer module to obtain the stylized result \\( I_{cs} \\). During training, we fine-tune the style transfer module via a CLIP-based style disentanglement loss \\( \\mathcal{L}_{SD} \\) coordinated with a style reconstruction prior (see details in Sec. 4.3, we omit it here for brevity) to implicitly learn the disentangled style information of \\( I_{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.444, + 0.306, + 0.459 + ], + "angle": 0, + "content": "\\(x_0, x_t\\) can be directly sampled as:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.469, + 0.469, + 0.496 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {0}\\right) := \\mathcal {N} (\\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.491, + 0.376, + 0.507 + ], + "angle": 0, + "content": "\\[\nx _ {t} := \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.469, + 0.551 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} := 1 - \\beta_{t}\\) and \\(\\bar{\\alpha}_{t} := \\prod_{s=0}^{t} \\alpha_{s}\\). Noise \\(\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})\\) has the same dimensionality as data \\(x_{0}\\) and latent \\(x_{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.643 + ], + "angle": 0, + "content": "The reverse process generates a reverse sequence by sampling the posteriors \\( q(x_{t - 1}|x_t) \\), starting from a Gaussian noise sample \\( x_{T} \\sim \\mathcal{N}(0,\\mathbf{I}) \\). However, since \\( q(x_{t - 1}|x_t) \\) is intractable, DDPMs learn parameterized Gaussian transitions \\( p_{\\theta}(x_{t - 1}|x_t) \\) with a learned mean \\( \\mu_{\\theta}(x_t,t) \\) and a fixed variance \\( \\sigma_t^2\\mathbf{I} \\) [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.653, + 0.47, + 0.673 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (x _ {t - 1} | x _ {t}) := \\mathcal {N} (\\mu_ {\\theta} (x _ {t}, t), \\sigma_ {t} ^ {2} \\mathbf {I}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.682, + 0.47, + 0.713 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}(x_t,t)\\) is the function of a noise approximator \\(\\epsilon_{\\theta}(x_t,t)\\). Then, the reverse process can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.722, + 0.47, + 0.756 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} := \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right) + \\sigma_ {t} \\mathbf {z}, \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.766, + 0.47, + 0.813 + ], + "angle": 0, + "content": "where \\(\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})\\) is a standard Gaussian noise independent of \\(x_{t}\\). \\(\\epsilon_{\\theta}(x_{t}, t)\\) is learned by a deep neural network [66] through optimizing the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.822, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\| \\epsilon_ {\\theta} (x _ {t}, t) - \\epsilon \\| ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Later, instead of using the fixed variances, Nichol and Dhariwal [57] presented a strategy for learning the variances. Song et al. [73] proposed DDIM, which formulates" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.444, + 0.892, + 0.489 + ], + "angle": 0, + "content": "an alternative non-Markovian noising process that has the same forward marginals as DDPM but allows a different reverse process:" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.5, + 0.892, + 0.537 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} := \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (x _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} (x _ {t}, t) + \\sigma_ {t} \\mathbf {z}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.538, + 0.892, + 0.569 + ], + "angle": 0, + "content": "where \\( f_{\\theta}(x_t, t) \\) is the predicted \\( x_0 \\) at timestep \\( t \\) given \\( x_t \\) and \\( \\epsilon_{\\theta}(x_t, t) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.578, + 0.892, + 0.613 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} (x _ {t}, t) := \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (x _ {t} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Changing the choice of \\(\\sigma_{t}\\) values in Eq. (6) can achieve different reverse processes. Especially when \\(\\sigma_{t} = 0\\), which is called DDIM [73], the reverse process becomes a deterministic mapping from latents to images, which enables nearly perfect inversion [38]. Besides, it can also accelerate the reverse process with much fewer sampling steps [14, 38]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.591, + 0.74 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Our task can be described as follows: given a style image \\( I_{s} \\) and an arbitrary content image \\( I_{c} \\), we want to first disentangle the content and style of them and then transfer the style of \\( I_{s} \\) to the content of \\( I_{c} \\). To achieve so, as stated in Sec. 1, our key idea is to explicitly extract the content information and then implicitly learn the complementary style information. Since our framework is built upon diffusion models [25, 73], we dub it StyleDiffusion." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Fig. 1 shows the overview of our StyleDiffusion, which consists of three key ingredients: I) a diffusion-based style" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7679" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "removal module, II) a diffusion-based style transfer module, and III) a CLIP-based style disentanglement loss coordinated with a style reconstruction prior. In the following subsections, we will introduce each of them in detail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.162, + 0.288, + 0.178 + ], + "angle": 0, + "content": "4.1. Style Removal Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.47, + 0.365 + ], + "angle": 0, + "content": "The style removal module aims at removing the style information of the content and style images, explicitly extracting the domain-aligned content information. Any reasonable content extraction operation can be used, depending on how the users define the content. For instance, users may want to use the structural outline as the content, so they can extract the outlines [33, 89] here. However, as discussed in Sec. 1, one challenge is controllability since the control of C-S disentanglement has been transformed into the control of content extraction. To this end, we introduce a diffusion-based style removal module to achieve both plausible and controllable content extraction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.471, + 0.669 + ], + "angle": 0, + "content": "Given an input image, e.g., the style image \\( I_{s} \\), since the color is an integral part of style [47], our style removal module first removes its color by a commonly used ITU-R 601-2 luma transform [19]. The obtained grayscale image is denoted as \\( I_{s}^{\\prime} \\). Then, we leverage a pre-trained diffusion model [14] \\( \\epsilon_{\\theta} \\) to remove the style details such as brushstrokes and textures of \\( I_{s}^{\\prime} \\), extracting the content \\( I_{s}^{c} \\). The insight is that the pre-trained diffusion model can help eliminate the domain-specific characteristics of input images and align them to the pre-trained domain [11, 38]. We assume that images with different styles belong to different domains, but their contents should share the same domain. Therefore, we can pre-train the diffusion model on a surrogate domain, e.g., the photograph domain, and then use this domain to construct the contents of images. After pretraining, the diffusion model can convert the input images from diverse domains to the latents \\( x \\) via the forward process and then inverse them to the photograph domain via the reverse process. In this way, the style characteristics can be ideally dispelled, leaving only the contents of the images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.47, + 0.76 + ], + "angle": 0, + "content": "Specifically, in order to obtain the results with fewer sampling steps and ensure that the content structures of the input images can be well preserved, we adopt the deterministic DDIM [73] sampling as the reverse process (Eq. (8)), and the ODE approximation of its reversal [38] as the forward process (Eq. (9)):" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.77, + 0.469, + 0.789 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (x _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\epsilon_ {\\theta} (x _ {t}, t), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.801, + 0.469, + 0.819 + ], + "angle": 0, + "content": "\\[\nx _ {t + 1} = \\sqrt {\\bar {\\alpha} _ {t + 1}} f _ {\\theta} (x _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t + 1}} \\epsilon_ {\\theta} (x _ {t}, t), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( f_{\\theta}(x_t,t) \\) is defined in Eq. (7). The forward and reverse diffusion processes enable us to easily control the intensity of style removal by adjusting the number of return step \\( T_{remov} \\) (see details in later Sec. 5.1). With the increase of \\( T_{remov} \\), more style characteristics will be removed, and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "the main content structures are retained, as will be shown in later Sec. 5.3. Note that for content images that are photographs, the diffusion processes are optional since they are already within the pre-trained domain, and there is almost no style except the colors to be dispelled. The superiority of diffusion-based style removal against other operations, such as Auto-Encoder (AE) [50]-based style removal, can be found in supplementary material (SM)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.222, + 0.71, + 0.238 + ], + "angle": 0, + "content": "4.2. Style Transfer Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.367 + ], + "angle": 0, + "content": "The style transfer module aims to learn the disentangled style information of the style image and transfer it to the content image. A common generative model like AEs [27] can be used here. However, inspired by the recent great success of diffusion models [14, 38], we introduce a diffusion-based style transfer module, which can better learn the disentangled style information in our framework and achieve higher-quality and more flexible stylizations (see Sec. 5.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.893, + 0.593 + ], + "angle": 0, + "content": "Given a content image \\( I_{c} \\), denote \\( I_{c}^{c} \\) is the content of \\( I_{c} \\) extracted by the style removal module (Sec. 4.1). We first convert it to the latent \\( x \\) using a pre-trained diffusion model \\( \\epsilon_{\\theta} \\). Then, guided by a CLIP-based style disentanglement loss coordinated with a style reconstruction prior (Sec. 4.3), the reverse process of the diffusion model is fine-tuned \\( (\\epsilon_{\\theta} \\rightarrow \\epsilon_{\\hat{\\theta}}) \\) to generate the stylized result \\( I_{cs} \\) referenced by the style image \\( I_{s} \\). Once the fine-tuning is completed, any content image can be manipulated into the stylized result with the disentangled style of the style image \\( I_{s} \\). To make the training easier and more stable, we adopt the deterministic DDIM forward and reverse processes in Eq. (8) and Eq. (9) during the fine-tuning. However, at inference, the stochastic DDPM [25] forward process (Eq. (2)) can also be used directly to help obtain diverse results [81] (Sec. 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.603, + 0.78, + 0.619 + ], + "angle": 0, + "content": "4.3. Loss Functions and Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Enforcing the style transfer module (Sec. 4.2) to learn and transfer the disentangled style information should address two key questions: (1) \"how to regularize the learned style is disentangled\" and (2) \"how to aptly transfer it to other contents\". To answer these questions, we introduce a novel CLIP-based style disentanglement loss coordinated with a style reconstruction prior to train the networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.837 + ], + "angle": 0, + "content": "CLIP-based Style Disentanglement Loss. Denote \\( I_{c}^{c} \\) and \\( I_{s}^{c} \\) are the respective contents of the content image \\( I_{c} \\) and the style image \\( I_{s} \\) extracted by the style removal module (Sec. 4.1). We aim to learn the disentangled style information of the style image \\( I_{s} \\) complementary to its content \\( I_{s}^{c} \\). Therefore, a straightforward way to obtain the disentangled style information is a direct subtraction:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.85, + 0.891, + 0.867 + ], + "angle": 0, + "content": "\\[\nD _ {s} ^ {p x} = I _ {s} - I _ {s} ^ {c}. \\tag {10}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1 Unless otherwise specified, we do not use the diffusion processes for content images in order to better maintain the content structures." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7680" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.346, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.218, + 0.25, + 0.233 + ], + "angle": 0, + "content": "(a) L1 loss" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.088, + 0.617, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.218, + 0.551, + 0.233 + ], + "angle": 0, + "content": "(b) Direction loss" + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.089, + 0.891, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.218, + 0.838, + 0.233 + ], + "angle": 0, + "content": "(c) Direction + L1 loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.894, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Illustration of different loss functions to transfer the disentangled style information. (a) L1 loss cannot guarantee the stylized results are within the style domain and may suffer from a collapse problem. (b) Direction loss aligns the disentangled directions but cannot realize accurate mappings. (c) Combining L1 loss and direction loss is able to achieve accurate one-to-one mappings from the content domain to the style domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.384 + ], + "angle": 0, + "content": "However, the simple pixel differences do not contain meaningful semantic information, thus cannot achieve plausible results [18, 42]. To address this problem, we can formulate the disentanglement in a latent semantic space:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.395, + 0.469, + 0.412 + ], + "angle": 0, + "content": "\\[\nD _ {s} = E \\left(I _ {s}\\right) - E \\left(I _ {s} ^ {c}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.47, + 0.723 + ], + "angle": 0, + "content": "where \\( E \\) is a well-pre-trained projector. Specifically, since \\( I_{s} \\) and \\( I_{s}^{c} \\) have similar contents but with different styles, the projector \\( E \\) must have the ability to distinguish them in terms of the style characteristics. In other words, as we define that images with different styles belong to different domains, the projector \\( E \\) should be able to distinguish the domains of \\( I_{s} \\) and \\( I_{s}^{c} \\). Fortunately, inspired by the recent vision-language model CLIP [62] that encapsulates knowledgeable semantic information of not only the photograph domain but also the artistic domain [17, 63, 46], we can use its image encoder as our projector \\( E \\) off the shelf. The open-domain CLIP space here serves as a good metric space to measure the \"style distance\" between content and its stylized result. This \"style distance\" thus can be interpreted as the disentangled style information. Note that here the style is implicitly defined as the complement of content, which is fundamentally different from the Gram matrix [18] that is an explicit style definition independent of content (see comparisons in Sec. 5.3). The comparisons between CLIP space and other possible spaces can be found in SM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.469, + 0.769 + ], + "angle": 0, + "content": "After obtaining the disentangled style information \\( D_{s} \\), the next question is how to properly transfer it to other contents. A possible solution is directly optimizing the L1 loss:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.779, + 0.469, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {c s} = E \\left(I _ {c s}\\right) - E \\left(I _ {c} ^ {c}\\right), \\\\ \\text {2} I _ {1} = \\text {＂} \\text {D} _ {1} = \\text {＂} \\text {D} _ {2} = \\text {＂} \\end{array} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.8, + 0.338, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S D} ^ {L 1} = \\parallel D _ {c s} - D _ {s} \\parallel ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\( I_{cs} \\) is the stylized result, \\( D_{cs} \\) is the disentangled style information of \\( I_{cs} \\). However, as illustrated in Fig. 2 (a) and further validated in later Sec. 5.3, minimizing the L1 loss cannot guarantee the stylized result \\( I_{cs} \\) is within the style domain of the style image \\( I_{s} \\). It is because L1 loss only" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.891, + 0.414 + ], + "angle": 0, + "content": "minimizes the absolute pixel difference (i.e., Manhattan distance); thus, it may produce stylized images that satisfy the Manhattan distance but deviate from the target style domain in the transfer direction. Besides, it may also lead to a collapse problem where a stylized output meets the same Manhattan distance with different contents in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.891, + 0.445 + ], + "angle": 0, + "content": "To address these problems, we can further constrain the disentangled directions as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.455, + 0.891, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S D} ^ {d i r} = 1 - \\frac {D _ {c s} \\cdot D _ {s}}{\\| D _ {c s} \\| \\| D _ {s} \\|.} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.619 + ], + "angle": 0, + "content": "This direction loss aligns the transfer direction of the content image's content to its stylization (i.e., the stylized result) with the direction of the style image's content to its stylization (i.e., the style image itself), as illustrated in Fig. 2 (b). Collaborated with this loss, the L1 loss \\(\\mathcal{L}_{SD}^{L1}\\) thus can achieve accurate one-to-one mappings from contents in the content domain to their stylizations in the style domain, as illustrated in Fig. 2 (c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.619, + 0.892, + 0.65 + ], + "angle": 0, + "content": "Finally, our style disentanglement loss is defined as a compound of \\(\\mathcal{L}_{SD}^{L1}\\) and \\(\\mathcal{L}_{SD}^{dir}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.66, + 0.891, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S D} = \\lambda_ {L 1} \\mathcal {L} _ {S D} ^ {L 1} + \\lambda_ {d i r} \\mathcal {L} _ {S D} ^ {d i r}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.795 + ], + "angle": 0, + "content": "where \\(\\lambda_{L1}\\) and \\(\\lambda_{dir}\\) are hyper-parameters set to 10 and 1 in our experiments. Since our style information is induced by the difference between content and its stylized result, we can deeply understand the relationship between C-S through learning. As a result, the style can be naturally and harmoniously transferred to the content, leading to better stylized images, as will be shown in later Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Style Reconstruction Prior. To fully use the prior information provided by the style image and further elevate the stylization effects, we integrate a style reconstruction prior into the fine-tuning of the style transfer module. Intuitively, given the content \\( I_{s}^{c} \\) of the style image \\( I_{s} \\), the style transfer module should be capable of recovering it to the original style image as much as possible. Therefore, we can define" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7681" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.894, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.53, + 0.818, + 0.547 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons with state of the art. Zoom-in for better comparison. Please see more in SM." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.573, + 0.327, + 0.587 + ], + "angle": 0, + "content": "a style reconstruction loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.6, + 0.469, + 0.616 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S R} = \\left\\| I _ {s s} - I _ {s} \\right\\|, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.749 + ], + "angle": 0, + "content": "where \\( I_{ss} \\) is the stylized result given \\( I_s^c \\) as content. We optimize it separately before optimizing the style disentanglement loss \\( \\mathcal{L}_{SD} \\). The detailed fine-tuning procedure can be found in SM. The style reconstruction prior helps our model recover the style information more sufficiently. It also provides a good initialization for the optimization of \\( \\mathcal{L}_{SD} \\), which helps the latter give full play to its ability, thus producing higher-quality results (see later Sec. 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.762, + 0.283, + 0.779 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.294, + 0.803 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We use ADM diffusion model [14] pre-trained on ImageNet [67] and adopt a fast sampling strategy [38]. Specifically, instead of sequentially conducting the diffusion processes until the last timestep \\( T \\) (e.g., 1000), we accelerate them by performing up to \\( T_{\\{\\cdot\\}} < T \\) (which is called return step), i.e., \\( T_{\\text{remov}} = 601 \\) for style removal and \\( T_{\\text{trans}} = \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.768 + ], + "angle": 0, + "content": "301 for style transfer. Moreover, as suggested by [38], we further accelerate the forward and reverse processes with fewer discretization steps, i.e., \\((S_{for}, S_{rev}) = (40, 40)\\) (\\(S_{for}\\) for forward process and \\(S_{rev}\\) for reverse process) for style removal, and \\((S_{for}, S_{rev}) = (40, 6)\\) for style transfer. When fine-tuning or inference, we can adjust \\(T_{remov}\\) or \\(T_{trans}\\) to flexibly control the degree of style removal and C-S disentanglement, as will be shown in Sec. 5.3. To fine-tune the model for a target style image, we randomly sample 50 images from ImageNet as the content images. We use Adam optimizer [40] with an initial learning rate of 4e-6 and increase it linearly by 1.2 per epoch. All models are fine-tuned with 5 epochs. See more details in SM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.784, + 0.761, + 0.8 + ], + "angle": 0, + "content": "5.2. Comparisons with Prior Arts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.809, + 0.892, + 0.869 + ], + "angle": 0, + "content": "We compare our StyleDiffusion against ten state-of-the-art (SOTA) methods [18, 90, 12, 1, 53, 6, 13, 74, 32, 52]. For fair comparisons, all these methods are fine-tuned or trained on the target styles similar to our approach." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitative Comparisons. As can be observed in Fig. 3, due to the entangling of C-S representations, Gatys [18]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7682" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.885, + 0.201 + ], + "angle": 0, + "content": "
OursGatysEFDMStyTr2ArtFlowAdaAttNIECASTMASTTPFRJohnsonLapStyle
SSIM ↑0.6720.3110.3160.5370.5010.5420.3650.3920.5360.6340.657
CLIP Score ↑0.7410.6770.6070.5310.5460.5770.6460.5900.6440.5370.595
Style Loss ↓0.8370.1110.1780.2160.2580.3100.2840.2290.9890.3640.274
UserStyle-43.1%41.2%39.3%36.4%37.2%33.8%39.1%14.5%42.8%47.3%
StudyOverall-26.0%38.1%44.0%34.2%43.9%32.7%32.2%22.6%43.4%46.2%
Training Time/h~0.4-~3~4~3~3~3~3~10~1~3
Testing Time/s5.61210.1650.0280.1680.2040.0760.0340.0660.3020.0150.008
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.204, + 0.893, + 0.234 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with state of the art. The training/testing time is measured with an Nvidia Tesla A100 GPU, and the testing time is averaged on images of size \\(512 \\times 512\\) pixels. \\(\\uparrow\\): Higher is better. \\(\\downarrow\\): Lower is better." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.258, + 0.473, + 0.621 + ], + "angle": 0, + "content": "and EFDM [90] often produce unsatisfying results with distorted contents (e.g., rows 1-3) and messy textures (e.g., rows 4-8). StyTr² [12] and ArtFlow [1] improve the results by adopting more advanced networks [77, 41], but they may still produce inferior results with halo boundaries (e.g., rows 2-3) or dirty artifacts (e.g., rows 4-6). AdaAttN [53] performs per-point attentive normalization to preserve the content structures better, but the stylization effects may be degraded in some cases (e.g., rows 1, 2, 4, and 5). IECAST [6] utilizes contrastive learning and external learning for style transfer, so fine-tuning it on a single style image would result in degraded results. MAST [13] uses multi-adaptation networks to disentangle C-S. However, since it still relies on the C-S representations of [18], the results usually exhibit messy textures and conspicuous artifacts. TPFR [74] is a GAN-based framework that learns to disentangle C-S in latent space. As the results show, it cannot recover correct style details and often generates deviated stylizations, which signifies that it may not learn truly disentangled C-S representations [54]. Like our method, Johnson [32] and LapStyle [52] also train separate models for each style. However, due to the trade-off between C-S losses of [18], they may produce less-stylized results or introduce unnatural patterns (e.g., rows 1-6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.777 + ], + "angle": 0, + "content": "By contrast, our StyleDiffusion completely disentangles C-S based on diffusion models. Therefore, it can generate high-quality results with sufficient style details (e.g., rows 1-4) and well-preserved contents (e.g., rows 5-8). Compared with the previous methods that tend to produce mixed results of content and style, our approach can better consider the relationship between them. Thus, the stylizations are more natural and harmonious, especially for challenging styles such as cubism (e.g., row 2) and oil painting (e.g., rows 1, 3, 4, and 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Quantitative Comparisons. We also resort to quantitative metrics to better evaluate our method, as shown in Tab. 1. We collect 32 content and 12 style images to synthesize 384 stylized results and compute the average Structural Similarity Index (SSIM) [1] to assess the content similarity. To evaluate the style similarity, we calculate the CLIP image similarity score [62] and Style Loss [18, 27] between the style images and the corresponding stylized results. As" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.895, + 0.379 + ], + "angle": 0, + "content": "shown in Tab. 1, our method obtains the highest SSIM and CLIP Score while the Style Loss is relatively higher than other methods. It is because these methods are directly trained to optimize Style Loss. Nevertheless, the Style Loss achieved by our method is still comparable and lower than the GAN-based TPFR [74]. Furthermore, it is noteworthy that our method can also incorporate Style Loss to enhance the performance in this regard (see later Sec. 5.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.896, + 0.606 + ], + "angle": 0, + "content": "User Study. As style transfer is highly subjective and CLIP Score and Style Loss are biased to the training objective, we additionally resort to user study to evaluate the style similarity and overall stylization quality. We randomly select 50 C-S pairs for each user. Given each C-S pair, we show the stylized results generated by our method and a randomly selected SOTA method side by side in random order. The users are asked to choose (1) which result transfers the style patterns better and (2) which result has overall better stylization effects. We obtain 1000 votes for each question from 20 users and show the percentage of votes that existing methods are preferred to ours in Tab. 1. The lower numbers indicate our method is more preferred than the competitors. As the results show, our method is superior to others in both style consistency and overall quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.895, + 0.729 + ], + "angle": 0, + "content": "Efficiency. As shown in the bottom two rows of Tab. 1, our approach requires less training time than others as it is fine-tuned on only a few (\\(\\sim 50\\)) content images. When testing, our approach is faster than the optimization-based method Gatys [18], albeit slower than the remaining feedforward methods due to the utilization of diffusion models. We discuss it in later Sec. 6, and more timing and resource details can be found in SM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.655, + 0.758 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Control of C-S Disentanglement. A prominent advantage of our StyleDiffusion is that we can flexibly control the C-S disentanglement by adjusting the content extraction of the style removal module (Sec. 4.1). Fig. 4 demonstrates the continuous control achieved by adjusting the return step \\( T_{\\text{remov}} \\) of the style removal module. As shown in the top row, with the increase of \\( T_{\\text{remov}} \\), more style characteristics are dispelled, and the main content structures are retained. Correspondingly, when more style is removed in the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7683" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.227, + 0.47, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Control of C-S disentanglement by adjusting the return step \\( T_{\\text{remov}} \\) of the style removal module. The top row shows the extracted contents of the style image. The bottom row shows the corresponding stylized results. * denotes our default setting. Zoom-in for better comparison. See SM for quantitative analyses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.468, + 0.431 + ], + "angle": 0, + "content": "top row, it will be aptly transferred to the stylized results in the bottom row, e.g., the twisted brushstrokes and the star patterns. It validates that our method successfully separates style from content in a controllable manner and properly transfers it to other contents. Moreover, the flexible C-S disentanglement also makes our StyleDiffusion versatile for other tasks, such as photo-realistic style transfer (see SM)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Superiority of Diffusion-based Style Transfer. Although our style transfer module is not limited to the diffusion model, using it offers three main advantages: (1) Flexible C-S trade-off control. As shown in Fig. 5, we can flexibly control the C-S trade-off at both the training stage (top row) and the testing stage (bottom row) by adjusting the return step \\( T_{trans} \\) of the diffusion model. With the increase of \\( T_{trans} \\), more style characteristics are transferred, yet the content structures may be ruined (e.g., the last column). When proper \\( T_{trans} \\) is adopted, e.g., \\( T_{trans} = 301 \\), the sweet spot can be well achieved. Interestingly, as shown in the last two columns of the bottom row, though the model is trained on \\( T_{trans} = 301 \\), we can extrapolate the style by using larger \\( T_{trans} \\) (e.g., 401) at the testing stage (but the results may be degraded when using too large \\( T_{trans} \\), e.g., 601). It provides a very flexible way for users to adjust the results according to their preferences. This property, however, cannot be simply achieved by using other models, e.g., the widely used AEs [27, 50], since our framework does not involve any feature transforms [27, 50] or C-S losses trade-off [3]. (2) Higher-quality stylizations. Owing to the strong generative ability of the diffusion model, it can achieve higher-quality stylizations than other models. For comparison, we use the pre-trained VGG-AE [27, 46] as the style transfer module and fine-tune its decoder network for each style. As shown in column (b) of Fig. 6, though the results are still acceptable, they may produce distorted contents and inferior textures, clearly worse than the results generated by the diffusion model in column (a). This is also validated by the bottom quantitative scores. It signifies that the diffusion model can better learn the disentangled" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.089, + 0.891, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.226, + 0.893, + 0.324 + ], + "angle": 0, + "content": "Figure 5. Control of C-S trade-off by adjusting the return step \\( T_{trans} \\) of the style transfer module. The top row shows adjusting \\( T_{trans} \\) at the training stage while fixing \\( T_{trans} = 301 \\) at the testing stage. The bottom row shows adjusting \\( T_{trans} \\) at the testing stage while fixing \\( T_{trans} = 301 \\) at the training stage. * denotes our default setting. Zoom-in for better comparison. See SM for quantitative analyses." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.339, + 0.695, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.339, + 0.892, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.526, + 0.866, + 0.54 + ], + "angle": 0, + "content": "Figure 6. Diffusion-based vs. AE-based style transfer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.688 + ], + "angle": 0, + "content": "content and style characteristics in our framework, helping produce better style transfer results. (3) Diversified style transfer. As mentioned in Sec. 4.2, during inference, we can directly adopt the stochastic DDPM [25] forward process (Eq. (2)) to obtain diverse results (see SM). The diverse results can give users endless choices to obtain more satisfactory results. However, using other models like AEs in our framework cannot easily achieve it [81]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Loss Analyses. To verify the effectiveness of each loss term used for fine-tuning our StyleDiffusion, we present ablation study results in Fig. 7 (a-d). (1) Using L1 loss \\(\\mathcal{L}_{SD}^{L1}\\) successfully transfers the cubism style like the blocky patterns in the top row, but the colors stray from the style images, especially in the bottom row. It is consistent with our earlier analyses in Sec. 4.3 that the L1 loss is prone to produce implausible results outside the style domain. (2) Adding direction loss \\(\\mathcal{L}_{SD}^{dir}\\) helps pull the results closer to the style domain. The textures are enhanced in the top row, and the colors are more plausible in the top and bottom rows. (3) By further coordinating with the style reconstruction prior \\(\\mathcal{L}_{SR}\\), the stylization effects are significantly elevated where the style information is recovered more sup" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7684" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.28, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.089, + 0.388, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.09, + 0.488, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.091, + 0.588, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.091, + 0.687, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.09, + 0.793, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.091, + 0.893, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.281, + 0.79, + 0.295 + ], + "angle": 0, + "content": "Figure 7. Ablation study on loss functions. * denotes our full model. Zoom-in for better comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.468, + 0.41 + ], + "angle": 0, + "content": "ficiently. It may be because it provides a good initialization for the optimization of \\(\\mathcal{L}_{SD}^{L1}\\) and \\(\\mathcal{L}_{SD}^{dir}\\), which helps them give full play to their abilities. As verified in Fig. 7 (d), using the style reconstruction alone cannot learn meaningful style patterns except for basic colors. All the above analyses are also supported by the bottom quantitative scores." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.468, + 0.818 + ], + "angle": 0, + "content": "Comparison with Gram Loss. To further verify the superiority of our proposed losses, we replace them with the widely used Gram Loss [18, 27] in Fig. 7 (e-f). As can be observed, Gram Loss destroys the content structures severely, e.g., the zebra head in the top row and the enlarged area in the bottom row. This is because it does not disentangle C-S and only matches the global statistics without considering the relationship between C-S. In contrast, our losses focus on learning the disentangled style information apart from the content, which is induced by the difference between content and its stylized result. Therefore, they can better understand the relationship between C-S, achieving more satisfactory results with fine style details and better-preserved contents, as validated by Fig. 7 (c) and the bottom quantitative scores. Furthermore, we also conduct comparisons between our proposed losses and Gram Loss [18, 27] on the AE baseline [27, 46] to eliminate the impact of diffusion models. As shown in Fig. 8 (a-b), our losses can achieve more satisfactory results than Gram Loss, which is consistent with the results in Fig. 7. Moreover, as shown in Fig. 8 (c), they can also be combined with Gram Loss to improve the performance on the Style Loss metric. However, it may affect the full disentanglement of C-S in our framework, which strays from our target and decreases the content preservation (see SSIM score in Fig. 8 (c)). Therefore, we do not incorporate Gram Loss in our framework by default." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.325, + 0.846 + ], + "angle": 0, + "content": "6. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this work, we present a new framework for more interpretable and controllable C-S disentanglement and style transfer. Our framework, termed StyleDiffusion, leverages" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.317, + 0.659, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.318, + 0.738, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.318, + 0.815, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.318, + 0.893, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.498, + 0.89, + 0.512 + ], + "angle": 0, + "content": "Figure 8. More loss function ablation study on the AE baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.66 + ], + "angle": 0, + "content": "diffusion models to explicitly extract the content information and implicitly learn the complementary style information. A novel CLIP-based style disentanglement loss coordinated with a style reconstruction prior is also introduced to encourage the disentanglement and style transfer. Our method yields very encouraging stylizations, especially for challenging styles, and the experimental results verify its effectiveness and superiority against state of the art." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.892, + 0.812 + ], + "angle": 0, + "content": "Currently, the framework still suffers from several limitations: (1) The model needs to be fine-tuned for each style, and arbitrary style transfer is left to our future work. (2) The efficiency is not fast enough due to the use of diffusion models. Further research in accelerating diffusion sampling would be helpful. (3) There are some failure cases analyzed in SM, which may help inspire future improvements. Moreover, our framework may also be applied to other image translation [28] or manipulation [60] tasks, and we would like to explore them in our future work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.828, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments. We thank Zeyi Huang and Xiaoting Zhang for their insightful discussions and suggestions. This work was supported in part by the National Program of China (2020YFC1523201, 62172365, 19ZDA197), Zhejiang Elite Program (2022C01222), and Key Technologies and Product Research and Development Projects for Cultural Relics Protection and Trading Circulation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7685" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 862-871, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.253 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Mohammad Babaeizadeh and Golnaz Ghiasi. Adjustable real-time style transfer. In International Conference on Learning Representations (ICLR), 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Robin Rombach, Kaan Oktay, Jonas Müller, and Björn Ommer. Retrieval-augmented diffusion models. Advances in Neural Information Processing Systems, 35:15309-15324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[5] Dongdong Chen, Lu Yuan, Jing Liao, Nenghai Yu, and Gang Hua. Stylebank: An explicit representation for neural image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1897-1906, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[6] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems (NeurIPS), 34:26561-26573, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[7] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dualist: Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 872-881, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[8] Haibo Chen, Lei Zhao, Huiming Zhang, Zhizhong Wang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Diverse image style transfer via invertible cross-space mapping. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14860-14869. IEEE Computer Society, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.471, + 0.732 + ], + "angle": 0, + "content": "[9] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in Neural Information Processing Systems (NeurIPS), 29, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.733, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[10] Jiaxin Cheng, Ayush Jaiswal, Yue Wu, Pradeep Natarajan, and Prem Natarajan. Style-aware normalized loss for improving arbitrary style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 134-143, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.804, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[11] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14347-14356. IEEE, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[12] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Im" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.895, + 0.135 + ], + "angle": 0, + "content": "age style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11326-11336, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[13] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM International Conference on Multimedia (ACM MM), pages 2719-2727, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.894, + 0.331 + ], + "angle": 0, + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[16] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 579-587, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.894, + 0.461 + ], + "angle": 0, + "content": "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.894, + 0.531 + ], + "angle": 0, + "content": "[18] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 1, 2, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[19] Rafael C Gonzalez. Digital image processing. Pearson education india, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.894, + 0.628 + ], + "angle": 0, + "content": "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[21] Daniel J Graham, James M Hughes, Helmut Leder, and Daniel N Rockmore. Statistics, vision, and the analysis of artistic style. Wiley Interdisciplinary Reviews: Computational Statistics, 4(2):115-123, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.894, + 0.744 + ], + "angle": 0, + "content": "[22] Shuyang Gu, Congliang Chen, Jing Liao, and Lu Yuan. Arbitrary style transfer with deep feature reshuffle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8222-8231, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[23] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[24] Irina Higgins, David Amos, David Pfau, Sebastien Racaniere, Loic Matthew, Danilo Rezende, and Alexander Lerchner. Towards a definition of disentangled representations. arXiv preprint arXiv:1812.02230, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.903 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7686" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "Processing Systems (NeurIPS), 33:6840-6851, 2020. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[26] Kibeom Hong, Seogkyu Jeon, Huan Yang, Jianlong Fu, and Hyeran Byun. Domain-aware universal style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[27] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1501-1510, 2017. 1, 2, 4, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[28] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 172-189, 2018. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[29] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14861-14869, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[30] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4369-4376, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[31] Yongcheng Jing, Yezhou Yang, Zunlei Feng, Jingwen Ye, Yizhou Yu, and Mingli Song. Neural style transfer: A review. IEEE Transactions on Visualization and Computer Graphics (TVCG), 26(11):3365-3385, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[32] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711. Springer, 2016. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[33] Henry Kang, Seungyong Lee, and Charles K Chui. Coherent line drawing. In Proceedings of the 5th international symposium on Non-photorealistic animation and rendering, pages 43–50, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[34] Theofanis Karaletsos, Serge Belongie, and Gunnar Ratsch. Bayesian representation learning with oracle constraints. arXiv preprint arXiv:1506.05011, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[35] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, and Holger Winnemoeller. Recognizing image style. arXiv preprint arXiv:1311.3715, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[36] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[37] Hadi Kazemi, Seyed Mehdi Iranmanesh, and Nasser Nasrabadi. Style and content disentanglement in generative adversarial networks. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 848-856. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 2426-2435, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[39] Hyunjik Kim and Andriy Mnih. Disentangling by factorising. In International Conference on Machine Learning (ICML), pages 2649-2658. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[40] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.249 + ], + "angle": 0, + "content": "[41] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[42] Dmytro Kotovenko, Artsiom Sanakoyeu, Sabine Lang, and Bjorn Ommer. Content and style disentanglement for artistic style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4422-4431, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[43] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Bjorn Ommer. Rethinking style transfer: From pixels to parameterized brushstrokes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12196-12205, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[44] Tejas D Kulkarni, William F Whitney, Pushmeet Kohli, and Josh Tenenbaum. Deep convolutional inverse graphics network. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[45] Gihyun Kwon and Jong Chul Ye. Diagonal attention and style-based gan for content-style disentanglement in image generation and translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13980-13989, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[46] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18062-18071, 2022. 5, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[47] Berel Lang. The concept of style. Cornell University Press, 1987. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[48] Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Maneesh Singh, and Ming-Hsuan Yang. Diverse image-to-image translation via disentangled representations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 35-51, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[49] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[50] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. In Advances in Neural Information Processing Systems (NeurIPS), pages 386-396, 2017. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[51] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Tianwei Lin, Zhuoqi Ma, Fu Li, Dongliang He, Xin Li, Errui Ding, Nannan Wang, Jie Li, and Xinbo Gao. Drafting and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7687" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "revision: Laplacian pyramid network for fast high-quality artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5141-5150, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.471, + 0.233 + ], + "angle": 0, + "content": "[53] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6649-6658, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.317 + ], + "angle": 0, + "content": "[54] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In International Conference on Machine Learning (ICML), pages 4114-4124. PMLR, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.319, + 0.471, + 0.387 + ], + "angle": 0, + "content": "[55] Haofei Lu and Zhizhong Wang. Universal video style transfer via crystallization, separation, and blending. In Proceedings of the International Joint Conferences on Artificial Intelligence Organization (IJCAI), Vienna, Austria, pages 23-29, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[56] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.471, + 0.516 + ], + "angle": 0, + "content": "[57] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.471, + 0.602 + ], + "angle": 0, + "content": "[58] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), pages 16784-16804. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[59] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5880–5888, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.729 + ], + "angle": 0, + "content": "[60] Taesung Park, Jun-Yan Zhu, Oliver Wang, Jingwan Lu, Eli Shechtman, Alexei Efros, and Richard Zhang. Swapping autoencoder for deep image manipulation. Advances in Neural Information Processing Systems (NeurIPS), 33:7198-7211, 2020. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[61] DM Parker and Jan B Deregowski. Perception and artistic style. Elsevier, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[63] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[64] Xuanchi Ren, Tao Yang, Yuwang Wang, and Wenjun Zeng. Rethinking content and style: Exploring bias for unsupervised disentanglement. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1823-1832. IEEE, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.894, + 0.235 + ], + "angle": 0, + "content": "[65] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.237, + 0.893, + 0.307 + ], + "angle": 0, + "content": "[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-assisted Intervention, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.894, + 0.379 + ], + "angle": 0, + "content": "[67] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.381, + 0.893, + 0.465 + ], + "angle": 0, + "content": "[68] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.468, + 0.893, + 0.523 + ], + "angle": 0, + "content": "[69] Artsiom Sanakoyeu, Dmytro Kotovenko, Sabine Lang, and Bjorn Ommer. A style-aware content loss for real-time hd style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), pages 698-714, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.893, + 0.595 + ], + "angle": 0, + "content": "[70] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8242-8250, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.599, + 0.893, + 0.64 + ], + "angle": 0, + "content": "[71] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.643, + 0.893, + 0.711 + ], + "angle": 0, + "content": "[72] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICML), pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.715, + 0.893, + 0.757 + ], + "angle": 0, + "content": "[73] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.894, + 0.829 + ], + "angle": 0, + "content": "[74] Jan Svoboda, Asha Anoosheh, Christian Osendorfer, and Jonathan Masci. Two-stage peer-regularized feature recombination for arbitrary image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13816–13825, 2020. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[75] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6924-6932, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7688" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[76] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[77] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.19, + 0.469, + 0.246 + ], + "angle": 0, + "content": "[78] Pei Wang, Yijun Li, and Nuno Vasconcelos. Rethinking and improving the robustness of image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 124-133, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.246, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[79] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: Towards aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia (ACM MM), pages 1095-1106, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[80] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2624-2632, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.385, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[81] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7789-7798, 2020. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.469, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[82] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Evaluate and improve the quality of neural style transfer. Computer Vision and Image Understanding (CVIU), 207:103203, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.525, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[83] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. DivSwapper: Towards diversified patch-based arbitrary style transfer. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), pages 4980–4987, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.595, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[84] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Glstylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8):575-586, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.65, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[85] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2742-2750, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.72, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[86] Wayne Wu, Kaidi Cao, Cheng Li, Chen Qian, and Chen Change Loy. Disentangling content and style via unsupervised geometry distillation. arXiv preprint arXiv:1905.04538, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.775, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[87] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14618-14627, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[88] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[89] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1395-1403, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[90] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8035-8045, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[91] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[92] Yu Zhang, Peter Tino, Ales Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.332, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[93] Yexun Zhang, Ya Zhang, and Wenbin Cai. Separating style and content for generalized style transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8447-8455, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[94] Zhiwen Zuo, Lei Zhao, Shuobin Lian, Haibo Chen, Zhizhong Wang, Ailin Li, Wei Xing, and Dongming Lu. Style fader generative adversarial networks for style degree controllable artistic style transfer. In Proc. Int. Joint Conf. on Artif. Intell.(IJCAI), pages 5002-5009, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7689" + } + ] +] \ No newline at end of file diff --git a/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/fea588fd-ec6f-4543-ad58-5dd9ea412b01_origin.pdf b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/fea588fd-ec6f-4543-ad58-5dd9ea412b01_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..598c81b95f169fb7b1037c9917875a87d336ac28 --- /dev/null +++ b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/fea588fd-ec6f-4543-ad58-5dd9ea412b01_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28ae033a7bcced81f8cb104fa32e257c28494677004bedfe6b14d646c3cd0695 +size 8915732 diff --git a/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/full.md b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..991e3a2d2f20cf1a81bd5b5dd577281c8d9d1ccf --- /dev/null +++ b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/full.md @@ -0,0 +1,402 @@ +# StyleDiffusion: Controllable Disentangled Style Transfer via Diffusion Models + +Zhizhong Wang*, Lei Zhao† Wei Xing +College of Computer Science and Technology, Zhejiang University +{endywon, czhl, WXING}@zju.edu.cn + +# Abstract + +Content and style (C-S) disentanglement is a fundamental problem and critical challenge of style transfer. Existing approaches based on explicit definitions (e.g., Gram matrix) or implicit learning (e.g., GANs) are neither interpretable nor easy to control, resulting in entangled representations and less satisfying results. In this paper, we propose a new C-S disentangled framework for style transfer without using previous assumptions. The key insight is to explicitly extract the content information and implicitly learn the complementary style information, yielding interpretable and controllable C-S disentanglement and style transfer. A simple yet effective CLIP-based style disentanglement loss coordinated with a style reconstruction prior is introduced to disentangle C-S in the CLIP image space. By further leveraging the powerful style removal and generative ability of diffusion models, our framework achieves superior results than state of the art and flexible C-S disentanglement and trade-off control. Our work provides new insights into the C-S disentanglement in style transfer and demonstrates the potential of diffusion models for learning well-disentangled C-S characteristics. + +# 1. Introduction + +Given a reference style image, e.g., Starry Night by Vincent Van Gogh, style transfer aims to transfer its artistic style, such as colors and brushstrokes, to an arbitrary content target. To achieve such a goal, it must first properly separate the style from the content and then transfer it to another content. This raises two fundamental challenges: (1) "how to disentangle content and style (C-S)" and (2) "how to transfer style to another content". + +To resolve these challenges, valuable efforts have been devoted. Gatys et al. [18] proposed A Neural Algorithm of Artistic Style to achieve style transfer, which explicitly defines the high-level features extracted from a pre-trained Convolutional Neural Network (CNN) (e.g., VGG [71]) as + +content, and the feature correlations (i.e., Gram matrix) as style. This approach acquires visually stunning results and inspires a large number of successors [32, 27, 50, 1, 90]. Despite the successes, by diving into the essence of style transfer, we observed three problems with these approaches: (1) The C-S are not completely disentangled. Theoretically, the C-S representations are intertwined. For example, matching the content representation of an image may also match its Gram matrix, and vice versa. (2) What CNN learned is a black box rugged to interpret [92], which makes the C-S definitions [18] uninterpretable and hard to control. (3) The transfer process is modeled as a separate optimization of content loss and style loss [18], so there lacks a deep understanding of the relationship between C-S. These problems usually lead to unbalanced stylizations and disharmonious artifacts [6], as will be shown in later Fig. 3. + +On the other hand, disentangled representation learning [24] provides other ideas to implicitly disentangle C-S, either supervised [44, 34] or unsupervised [9, 93]. For style transfer, Kotovenko et al. [42] utilized fixpoint triplet style loss and disentanglement loss to enforce a GAN [20]-based framework to learn separate C-S representations in an unsupervised manner. Similarly, TPFR [74] learned to disentangle C-S in latent space via metric learning and two-stage peer-regularization, producing high-quality images even in the zero-shot setting. While these approaches successfully enforce properties "encouraged" by the corresponding losses, they still have three main problems: (1) Well-disentangled models seemingly cannot be identified without supervision [54, 64], which means the unsupervised learning [42, 74] may not achieve truly disentangled C-S, as will be shown in later Fig. 3. (2) These approaches are all based on GANs and thus often confined to the GAN predefined domains, e.g., a specific artist's style domain [69]. (3) The implicitly learned C-S representations are still black boxes that are hard to interpret and control [54]. + +Facing the challenges above, in this paper, we propose a new C-S disentangled framework for style transfer without using previous assumptions such as Gram matrix [18] or GANs [42]. Our key insight stems from the fact that the definition of an image's style is much more complex + +than its content, e.g., we can easily identify the content of a painting by its structures, semantics, or shapes, but it is intractable to define the style [61, 21, 35, 82]. Therefore, we can bypass such a dilemma by explicitly extracting the content information and implicitly learning its complementary style information. Since we strictly constrain style as the complement of content, the C-S can be completely disentangled, and the control of disentanglement has been transformed into the control of content extraction. It achieves both controllability and interpretability. + +However, achieving plausible and controllable content extraction is also non-trivial because the contents extracted from the content images and style images should share the same content domain, and the details of the extracted contents should be easy to control. To this end, we resort to recent developed diffusion models [25, 73] and introduce a diffusion-based style removal module to smoothly dispel the style information of the content and style images, extracting the domain-aligned content information. Moreover, owing to the strong generative capability of diffusion models, we also introduce a diffusion-based style transfer module to better learn the disentangled style information of the style image and transfer it to the content image. The style disentanglement and transfer are encouraged via a simple yet effective CLIP [62]-based style disentanglement loss, which induces the transfer mapping of the content image's content to its stylization (i.e., the stylized result) to be aligned with that of the style image's content to its stylization (i.e., the style image itself) in the CLIP image space. By further coordinating with a style reconstruction prior, it achieves both generalized and faithful style transfer. We conduct comprehensive comparisons and ablation study to demonstrate the effectiveness and superiority of our framework. With the well-disentangled C-S, it achieves very promising stylizations with fine style details, well-preserved contents, and a deep understanding of the relationship between C-S. + +In summary, our contributions are threefold: + +- We propose a novel C-S disentangled framework for style transfer, which achieves more interpretable and controllable C-S disentanglement and higher-quality stylized results. +- We introduce diffusion models to our framework and demonstrate their effectiveness and superiority in controllable style removal and learning well-disentangled C-S characteristics. +- A new CLIP-based style disentanglement loss coordinated with a style reconstruction prior is introduced to disentangle C-S in the CLIP image space. + +# 2. Related Work + +Neural Style Transfer (NST). The pioneering work of Gatys et al. [18] has opened the era of NST [31]. Since + +then, this task has experienced tremendous progress, including efficiency [32, 49, 85], quality [22, 70, 84, 52, 10, 7, 1, 43, 78, 53, 6, 87, 29, 94, 12, 91, 79], generality [5, 27, 50, 59, 13, 30, 26, 80, 90, 55, 88], and diversity [75, 81, 83]. Despite these successes, the essence of these approaches is mostly based on the explicitly defined C-S representations, such as Gram matrix [18], which have several limitations as discussed in Sec. 1. In our work, we propose new disentangled C-S representations explicitly extracted or implicitly learned by diffusion models, achieving more effective style transfer and higher-quality results. + +Disentangled Representation Learning (DRL). The task of DRL [24] aims at modeling the factors of data variations [48]. Earlier works used labeled data to factorize representations in a supervised manner [34]. Recently, unsupervised settings have been largely explored [39], especially for disentangling style from content [93, 28, 48, 37, 86, 42, 60, 64, 8, 45]. However, due to the dependence on GANs [20], their C-S disentanglement is usually restricted in the GAN pre-defined domains (e.g., Van Gogh's style domain). Besides, disentanglement cannot be effectively achieved without providing sufficient data [54]. In contrast, our framework learns the disentangled style from a single style image, and the disentanglement can be easily achieved by providing only a few ( $\sim$ 50) content images for training. + +Diffusion Models. Diffusion models [72] such as denoising diffusion probabilistic models (DDPMs) [25, 57] have recently shown great success in image generation [73, 14, 16], image manipulation [56, 2, 38], and text-conditional synthesis [58, 68, 63, 65, 23, 4, 51]. These works have demonstrated the power of diffusion models to achieve higher-quality results than other generative models like VAEs [76], auto-regressive models [15], flows [41], and GANs [36]. Inspired by them, we introduce a diffusion-based style removal module and a style transfer module in our framework. These modules can smoothly remove the style information of images and better learn the recovery of it to achieve higher-quality style transfer results. To the best of our knowledge, our work is the first to introduce diffusion models to the field of neural style transfer. + +# 3. Background + +Denoising diffusion probabilistic models (DDPMs) [72, 25] are latent variable models that consist of two diffusion processes, i.e., a forward diffusion process and a reverse diffusion process. The forward process is a fixed Markov Chain that sequentially produces a series of latents $x_{1}, \ldots, x_{T}$ by gradually adding Gaussian noises at each timestep $t \in [1, T]$ : + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) := \mathcal {N} \left(\sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where $\beta_{t} \in (0,1)$ is a fixed variance schedule. An important property of the forward process is that given clean data + +![](images/eefdcc0c02e6e3253a5130b9b1f512b64ec36b470b7a294df09b936ff28bea86.jpg) +Figure 1. Overview of our proposed StyleDiffusion. The content image $I_{c}$ and style image $I_{s}$ are first fed into a diffusion-based style removal module to explicitly extract the domain-aligned content information. Then, the content of $I_{c}$ is fed into a diffusion-based style transfer module to obtain the stylized result $I_{cs}$ . During training, we fine-tune the style transfer module via a CLIP-based style disentanglement loss $\mathcal{L}_{SD}$ coordinated with a style reconstruction prior (see details in Sec. 4.3, we omit it here for brevity) to implicitly learn the disentangled style information of $I_{s}$ . + +$x_0, x_t$ can be directly sampled as: + +$$ +q \left(x _ {t} \mid x _ {0}\right) := \mathcal {N} (\sqrt {\bar {\alpha} _ {t}} x _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}), \tag {2} +$$ + +$$ +x _ {t} := \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , +$$ + +where $\alpha_{t} := 1 - \beta_{t}$ and $\bar{\alpha}_{t} := \prod_{s=0}^{t} \alpha_{s}$ . Noise $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ has the same dimensionality as data $x_{0}$ and latent $x_{t}$ . + +The reverse process generates a reverse sequence by sampling the posteriors $q(x_{t - 1}|x_t)$ , starting from a Gaussian noise sample $x_{T} \sim \mathcal{N}(0,\mathbf{I})$ . However, since $q(x_{t - 1}|x_t)$ is intractable, DDPMs learn parameterized Gaussian transitions $p_{\theta}(x_{t - 1}|x_t)$ with a learned mean $\mu_{\theta}(x_t,t)$ and a fixed variance $\sigma_t^2\mathbf{I}$ [25]: + +$$ +p _ {\theta} (x _ {t - 1} | x _ {t}) := \mathcal {N} (\mu_ {\theta} (x _ {t}, t), \sigma_ {t} ^ {2} \mathbf {I}), \tag {3} +$$ + +where $\mu_{\theta}(x_t,t)$ is the function of a noise approximator $\epsilon_{\theta}(x_t,t)$ . Then, the reverse process can be expressed as: + +$$ +x _ {t - 1} := \frac {1}{\sqrt {\alpha_ {t}}} \left(x _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} \left(x _ {t}, t\right)\right) + \sigma_ {t} \mathbf {z}, \quad (4) +$$ + +where $\mathbf{z} \sim \mathcal{N}(0, \mathbf{I})$ is a standard Gaussian noise independent of $x_{t}$ . $\epsilon_{\theta}(x_{t}, t)$ is learned by a deep neural network [66] through optimizing the following loss: + +$$ +\min _ {\theta} \| \epsilon_ {\theta} (x _ {t}, t) - \epsilon \| ^ {2}. \tag {5} +$$ + +Later, instead of using the fixed variances, Nichol and Dhariwal [57] presented a strategy for learning the variances. Song et al. [73] proposed DDIM, which formulates + +an alternative non-Markovian noising process that has the same forward marginals as DDPM but allows a different reverse process: + +$$ +x _ {t - 1} := \sqrt {\bar {\alpha} _ {t - 1}} f _ {\theta} (x _ {t}, t) + \sqrt {1 - \bar {\alpha} _ {t - 1} - \sigma_ {t} ^ {2}} \epsilon_ {\theta} (x _ {t}, t) + \sigma_ {t} \mathbf {z}, \tag {6} +$$ + +where $f_{\theta}(x_t, t)$ is the predicted $x_0$ at timestep $t$ given $x_t$ and $\epsilon_{\theta}(x_t, t)$ : + +$$ +f _ {\theta} (x _ {t}, t) := \frac {x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} (x _ {t} , t)}{\sqrt {\bar {\alpha} _ {t}}}. \tag {7} +$$ + +Changing the choice of $\sigma_{t}$ values in Eq. (6) can achieve different reverse processes. Especially when $\sigma_{t} = 0$ , which is called DDIM [73], the reverse process becomes a deterministic mapping from latents to images, which enables nearly perfect inversion [38]. Besides, it can also accelerate the reverse process with much fewer sampling steps [14, 38]. + +# 4. Method + +Our task can be described as follows: given a style image $I_{s}$ and an arbitrary content image $I_{c}$ , we want to first disentangle the content and style of them and then transfer the style of $I_{s}$ to the content of $I_{c}$ . To achieve so, as stated in Sec. 1, our key idea is to explicitly extract the content information and then implicitly learn the complementary style information. Since our framework is built upon diffusion models [25, 73], we dub it StyleDiffusion. + +Fig. 1 shows the overview of our StyleDiffusion, which consists of three key ingredients: I) a diffusion-based style + +removal module, II) a diffusion-based style transfer module, and III) a CLIP-based style disentanglement loss coordinated with a style reconstruction prior. In the following subsections, we will introduce each of them in detail. + +# 4.1. Style Removal Module + +The style removal module aims at removing the style information of the content and style images, explicitly extracting the domain-aligned content information. Any reasonable content extraction operation can be used, depending on how the users define the content. For instance, users may want to use the structural outline as the content, so they can extract the outlines [33, 89] here. However, as discussed in Sec. 1, one challenge is controllability since the control of C-S disentanglement has been transformed into the control of content extraction. To this end, we introduce a diffusion-based style removal module to achieve both plausible and controllable content extraction. + +Given an input image, e.g., the style image $I_{s}$ , since the color is an integral part of style [47], our style removal module first removes its color by a commonly used ITU-R 601-2 luma transform [19]. The obtained grayscale image is denoted as $I_{s}^{\prime}$ . Then, we leverage a pre-trained diffusion model [14] $\epsilon_{\theta}$ to remove the style details such as brushstrokes and textures of $I_{s}^{\prime}$ , extracting the content $I_{s}^{c}$ . The insight is that the pre-trained diffusion model can help eliminate the domain-specific characteristics of input images and align them to the pre-trained domain [11, 38]. We assume that images with different styles belong to different domains, but their contents should share the same domain. Therefore, we can pre-train the diffusion model on a surrogate domain, e.g., the photograph domain, and then use this domain to construct the contents of images. After pretraining, the diffusion model can convert the input images from diverse domains to the latents $x$ via the forward process and then inverse them to the photograph domain via the reverse process. In this way, the style characteristics can be ideally dispelled, leaving only the contents of the images. + +Specifically, in order to obtain the results with fewer sampling steps and ensure that the content structures of the input images can be well preserved, we adopt the deterministic DDIM [73] sampling as the reverse process (Eq. (8)), and the ODE approximation of its reversal [38] as the forward process (Eq. (9)): + +$$ +x _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} f _ {\theta} (x _ {t}, t) + \sqrt {1 - \bar {\alpha} _ {t - 1}} \epsilon_ {\theta} (x _ {t}, t), \tag {8} +$$ + +$$ +x _ {t + 1} = \sqrt {\bar {\alpha} _ {t + 1}} f _ {\theta} (x _ {t}, t) + \sqrt {1 - \bar {\alpha} _ {t + 1}} \epsilon_ {\theta} (x _ {t}, t), \tag {9} +$$ + +where $f_{\theta}(x_t,t)$ is defined in Eq. (7). The forward and reverse diffusion processes enable us to easily control the intensity of style removal by adjusting the number of return step $T_{remov}$ (see details in later Sec. 5.1). With the increase of $T_{remov}$ , more style characteristics will be removed, and + +the main content structures are retained, as will be shown in later Sec. 5.3. Note that for content images that are photographs, the diffusion processes are optional since they are already within the pre-trained domain, and there is almost no style except the colors to be dispelled. The superiority of diffusion-based style removal against other operations, such as Auto-Encoder (AE) [50]-based style removal, can be found in supplementary material (SM). + +# 4.2. Style Transfer Module + +The style transfer module aims to learn the disentangled style information of the style image and transfer it to the content image. A common generative model like AEs [27] can be used here. However, inspired by the recent great success of diffusion models [14, 38], we introduce a diffusion-based style transfer module, which can better learn the disentangled style information in our framework and achieve higher-quality and more flexible stylizations (see Sec. 5.3). + +Given a content image $I_{c}$ , denote $I_{c}^{c}$ is the content of $I_{c}$ extracted by the style removal module (Sec. 4.1). We first convert it to the latent $x$ using a pre-trained diffusion model $\epsilon_{\theta}$ . Then, guided by a CLIP-based style disentanglement loss coordinated with a style reconstruction prior (Sec. 4.3), the reverse process of the diffusion model is fine-tuned $(\epsilon_{\theta} \rightarrow \epsilon_{\hat{\theta}})$ to generate the stylized result $I_{cs}$ referenced by the style image $I_{s}$ . Once the fine-tuning is completed, any content image can be manipulated into the stylized result with the disentangled style of the style image $I_{s}$ . To make the training easier and more stable, we adopt the deterministic DDIM forward and reverse processes in Eq. (8) and Eq. (9) during the fine-tuning. However, at inference, the stochastic DDPM [25] forward process (Eq. (2)) can also be used directly to help obtain diverse results [81] (Sec. 5.3). + +# 4.3. Loss Functions and Fine-tuning + +Enforcing the style transfer module (Sec. 4.2) to learn and transfer the disentangled style information should address two key questions: (1) "how to regularize the learned style is disentangled" and (2) "how to aptly transfer it to other contents". To answer these questions, we introduce a novel CLIP-based style disentanglement loss coordinated with a style reconstruction prior to train the networks. + +CLIP-based Style Disentanglement Loss. Denote $I_{c}^{c}$ and $I_{s}^{c}$ are the respective contents of the content image $I_{c}$ and the style image $I_{s}$ extracted by the style removal module (Sec. 4.1). We aim to learn the disentangled style information of the style image $I_{s}$ complementary to its content $I_{s}^{c}$ . Therefore, a straightforward way to obtain the disentangled style information is a direct subtraction: + +$$ +D _ {s} ^ {p x} = I _ {s} - I _ {s} ^ {c}. \tag {10} +$$ + +![](images/43f281c92aa3a206fb934e88e358668a72ab7ec73a0254bb4084c0c2e37e7971.jpg) +(a) L1 loss + +![](images/5ea1ad39b632f0621e39e7da76d120caf1fca0843fe9087d7e933b298a28251f.jpg) +(b) Direction loss +Figure 2. Illustration of different loss functions to transfer the disentangled style information. (a) L1 loss cannot guarantee the stylized results are within the style domain and may suffer from a collapse problem. (b) Direction loss aligns the disentangled directions but cannot realize accurate mappings. (c) Combining L1 loss and direction loss is able to achieve accurate one-to-one mappings from the content domain to the style domain. + +![](images/6f34e880991651244fe8ad29e70daa1ac748236e07afaecd5c59ef722e061d78.jpg) +(c) Direction + L1 loss + +However, the simple pixel differences do not contain meaningful semantic information, thus cannot achieve plausible results [18, 42]. To address this problem, we can formulate the disentanglement in a latent semantic space: + +$$ +D _ {s} = E \left(I _ {s}\right) - E \left(I _ {s} ^ {c}\right), \tag {11} +$$ + +where $E$ is a well-pre-trained projector. Specifically, since $I_{s}$ and $I_{s}^{c}$ have similar contents but with different styles, the projector $E$ must have the ability to distinguish them in terms of the style characteristics. In other words, as we define that images with different styles belong to different domains, the projector $E$ should be able to distinguish the domains of $I_{s}$ and $I_{s}^{c}$ . Fortunately, inspired by the recent vision-language model CLIP [62] that encapsulates knowledgeable semantic information of not only the photograph domain but also the artistic domain [17, 63, 46], we can use its image encoder as our projector $E$ off the shelf. The open-domain CLIP space here serves as a good metric space to measure the "style distance" between content and its stylized result. This "style distance" thus can be interpreted as the disentangled style information. Note that here the style is implicitly defined as the complement of content, which is fundamentally different from the Gram matrix [18] that is an explicit style definition independent of content (see comparisons in Sec. 5.3). The comparisons between CLIP space and other possible spaces can be found in SM. + +After obtaining the disentangled style information $D_{s}$ , the next question is how to properly transfer it to other contents. A possible solution is directly optimizing the L1 loss: + +$$ +\begin{array}{l} D _ {c s} = E \left(I _ {c s}\right) - E \left(I _ {c} ^ {c}\right), \\ \text {2} I _ {1} = \text {＂} \text {D} _ {1} = \text {＂} \text {D} _ {2} = \text {＂} \end{array} \tag {12} +$$ + +$$ +\mathcal {L} _ {S D} ^ {L 1} = \parallel D _ {c s} - D _ {s} \parallel , +$$ + +where $I_{cs}$ is the stylized result, $D_{cs}$ is the disentangled style information of $I_{cs}$ . However, as illustrated in Fig. 2 (a) and further validated in later Sec. 5.3, minimizing the L1 loss cannot guarantee the stylized result $I_{cs}$ is within the style domain of the style image $I_{s}$ . It is because L1 loss only + +minimizes the absolute pixel difference (i.e., Manhattan distance); thus, it may produce stylized images that satisfy the Manhattan distance but deviate from the target style domain in the transfer direction. Besides, it may also lead to a collapse problem where a stylized output meets the same Manhattan distance with different contents in the latent space. + +To address these problems, we can further constrain the disentangled directions as follows: + +$$ +\mathcal {L} _ {S D} ^ {d i r} = 1 - \frac {D _ {c s} \cdot D _ {s}}{\| D _ {c s} \| \| D _ {s} \|.} \tag {13} +$$ + +This direction loss aligns the transfer direction of the content image's content to its stylization (i.e., the stylized result) with the direction of the style image's content to its stylization (i.e., the style image itself), as illustrated in Fig. 2 (b). Collaborated with this loss, the L1 loss $\mathcal{L}_{SD}^{L1}$ thus can achieve accurate one-to-one mappings from contents in the content domain to their stylizations in the style domain, as illustrated in Fig. 2 (c). + +Finally, our style disentanglement loss is defined as a compound of $\mathcal{L}_{SD}^{L1}$ and $\mathcal{L}_{SD}^{dir}$ : + +$$ +\mathcal {L} _ {S D} = \lambda_ {L 1} \mathcal {L} _ {S D} ^ {L 1} + \lambda_ {d i r} \mathcal {L} _ {S D} ^ {d i r}, \tag {14} +$$ + +where $\lambda_{L1}$ and $\lambda_{dir}$ are hyper-parameters set to 10 and 1 in our experiments. Since our style information is induced by the difference between content and its stylized result, we can deeply understand the relationship between C-S through learning. As a result, the style can be naturally and harmoniously transferred to the content, leading to better stylized images, as will be shown in later Fig. 3. + +Style Reconstruction Prior. To fully use the prior information provided by the style image and further elevate the stylization effects, we integrate a style reconstruction prior into the fine-tuning of the style transfer module. Intuitively, given the content $I_{s}^{c}$ of the style image $I_{s}$ , the style transfer module should be capable of recovering it to the original style image as much as possible. Therefore, we can define + +![](images/9f812457c75c9843b7d0b042d9757d602fee9d29caf8180363d673706d56a6ca.jpg) +Figure 3. Qualitative comparisons with state of the art. Zoom-in for better comparison. Please see more in SM. + +a style reconstruction loss as follows: + +$$ +\mathcal {L} _ {S R} = \left\| I _ {s s} - I _ {s} \right\|, \tag {15} +$$ + +where $I_{ss}$ is the stylized result given $I_s^c$ as content. We optimize it separately before optimizing the style disentanglement loss $\mathcal{L}_{SD}$ . The detailed fine-tuning procedure can be found in SM. The style reconstruction prior helps our model recover the style information more sufficiently. It also provides a good initialization for the optimization of $\mathcal{L}_{SD}$ , which helps the latter give full play to its ability, thus producing higher-quality results (see later Sec. 5.3). + +# 5. Experimental Results + +# 5.1. Implementation Details + +We use ADM diffusion model [14] pre-trained on ImageNet [67] and adopt a fast sampling strategy [38]. Specifically, instead of sequentially conducting the diffusion processes until the last timestep $T$ (e.g., 1000), we accelerate them by performing up to $T_{\{\cdot\}} < T$ (which is called return step), i.e., $T_{\text{remov}} = 601$ for style removal and $T_{\text{trans}} =$ + +301 for style transfer. Moreover, as suggested by [38], we further accelerate the forward and reverse processes with fewer discretization steps, i.e., $(S_{for}, S_{rev}) = (40, 40)$ ( $S_{for}$ for forward process and $S_{rev}$ for reverse process) for style removal, and $(S_{for}, S_{rev}) = (40, 6)$ for style transfer. When fine-tuning or inference, we can adjust $T_{remov}$ or $T_{trans}$ to flexibly control the degree of style removal and C-S disentanglement, as will be shown in Sec. 5.3. To fine-tune the model for a target style image, we randomly sample 50 images from ImageNet as the content images. We use Adam optimizer [40] with an initial learning rate of 4e-6 and increase it linearly by 1.2 per epoch. All models are fine-tuned with 5 epochs. See more details in SM. + +# 5.2. Comparisons with Prior Arts + +We compare our StyleDiffusion against ten state-of-the-art (SOTA) methods [18, 90, 12, 1, 53, 6, 13, 74, 32, 52]. For fair comparisons, all these methods are fine-tuned or trained on the target styles similar to our approach. + +Qualitative Comparisons. As can be observed in Fig. 3, due to the entangling of C-S representations, Gatys [18] + +
OursGatysEFDMStyTr2ArtFlowAdaAttNIECASTMASTTPFRJohnsonLapStyle
SSIM ↑0.6720.3110.3160.5370.5010.5420.3650.3920.5360.6340.657
CLIP Score ↑0.7410.6770.6070.5310.5460.5770.6460.5900.6440.5370.595
Style Loss ↓0.8370.1110.1780.2160.2580.3100.2840.2290.9890.3640.274
UserStyle-43.1%41.2%39.3%36.4%37.2%33.8%39.1%14.5%42.8%47.3%
StudyOverall-26.0%38.1%44.0%34.2%43.9%32.7%32.2%22.6%43.4%46.2%
Training Time/h~0.4-~3~4~3~3~3~3~10~1~3
Testing Time/s5.61210.1650.0280.1680.2040.0760.0340.0660.3020.0150.008
+ +Table 1. Quantitative comparisons with state of the art. The training/testing time is measured with an Nvidia Tesla A100 GPU, and the testing time is averaged on images of size $512 \times 512$ pixels. $\uparrow$ : Higher is better. $\downarrow$ : Lower is better. + +and EFDM [90] often produce unsatisfying results with distorted contents (e.g., rows 1-3) and messy textures (e.g., rows 4-8). StyTr² [12] and ArtFlow [1] improve the results by adopting more advanced networks [77, 41], but they may still produce inferior results with halo boundaries (e.g., rows 2-3) or dirty artifacts (e.g., rows 4-6). AdaAttN [53] performs per-point attentive normalization to preserve the content structures better, but the stylization effects may be degraded in some cases (e.g., rows 1, 2, 4, and 5). IECAST [6] utilizes contrastive learning and external learning for style transfer, so fine-tuning it on a single style image would result in degraded results. MAST [13] uses multi-adaptation networks to disentangle C-S. However, since it still relies on the C-S representations of [18], the results usually exhibit messy textures and conspicuous artifacts. TPFR [74] is a GAN-based framework that learns to disentangle C-S in latent space. As the results show, it cannot recover correct style details and often generates deviated stylizations, which signifies that it may not learn truly disentangled C-S representations [54]. Like our method, Johnson [32] and LapStyle [52] also train separate models for each style. However, due to the trade-off between C-S losses of [18], they may produce less-stylized results or introduce unnatural patterns (e.g., rows 1-6). + +By contrast, our StyleDiffusion completely disentangles C-S based on diffusion models. Therefore, it can generate high-quality results with sufficient style details (e.g., rows 1-4) and well-preserved contents (e.g., rows 5-8). Compared with the previous methods that tend to produce mixed results of content and style, our approach can better consider the relationship between them. Thus, the stylizations are more natural and harmonious, especially for challenging styles such as cubism (e.g., row 2) and oil painting (e.g., rows 1, 3, 4, and 5). + +Quantitative Comparisons. We also resort to quantitative metrics to better evaluate our method, as shown in Tab. 1. We collect 32 content and 12 style images to synthesize 384 stylized results and compute the average Structural Similarity Index (SSIM) [1] to assess the content similarity. To evaluate the style similarity, we calculate the CLIP image similarity score [62] and Style Loss [18, 27] between the style images and the corresponding stylized results. As + +shown in Tab. 1, our method obtains the highest SSIM and CLIP Score while the Style Loss is relatively higher than other methods. It is because these methods are directly trained to optimize Style Loss. Nevertheless, the Style Loss achieved by our method is still comparable and lower than the GAN-based TPFR [74]. Furthermore, it is noteworthy that our method can also incorporate Style Loss to enhance the performance in this regard (see later Sec. 5.3). + +User Study. As style transfer is highly subjective and CLIP Score and Style Loss are biased to the training objective, we additionally resort to user study to evaluate the style similarity and overall stylization quality. We randomly select 50 C-S pairs for each user. Given each C-S pair, we show the stylized results generated by our method and a randomly selected SOTA method side by side in random order. The users are asked to choose (1) which result transfers the style patterns better and (2) which result has overall better stylization effects. We obtain 1000 votes for each question from 20 users and show the percentage of votes that existing methods are preferred to ours in Tab. 1. The lower numbers indicate our method is more preferred than the competitors. As the results show, our method is superior to others in both style consistency and overall quality. + +Efficiency. As shown in the bottom two rows of Tab. 1, our approach requires less training time than others as it is fine-tuned on only a few ( $\sim 50$ ) content images. When testing, our approach is faster than the optimization-based method Gatys [18], albeit slower than the remaining feedforward methods due to the utilization of diffusion models. We discuss it in later Sec. 6, and more timing and resource details can be found in SM. + +# 5.3. Ablation Study + +Control of C-S Disentanglement. A prominent advantage of our StyleDiffusion is that we can flexibly control the C-S disentanglement by adjusting the content extraction of the style removal module (Sec. 4.1). Fig. 4 demonstrates the continuous control achieved by adjusting the return step $T_{\text{remov}}$ of the style removal module. As shown in the top row, with the increase of $T_{\text{remov}}$ , more style characteristics are dispelled, and the main content structures are retained. Correspondingly, when more style is removed in the + +![](images/9e7d8272db81e048072bbe46ba61b8de4c2bd0ba6d9c5d1a519a02fb83c09a8a.jpg) +Figure 4. Control of C-S disentanglement by adjusting the return step $T_{\text{remov}}$ of the style removal module. The top row shows the extracted contents of the style image. The bottom row shows the corresponding stylized results. * denotes our default setting. Zoom-in for better comparison. See SM for quantitative analyses. + +top row, it will be aptly transferred to the stylized results in the bottom row, e.g., the twisted brushstrokes and the star patterns. It validates that our method successfully separates style from content in a controllable manner and properly transfers it to other contents. Moreover, the flexible C-S disentanglement also makes our StyleDiffusion versatile for other tasks, such as photo-realistic style transfer (see SM). + +Superiority of Diffusion-based Style Transfer. Although our style transfer module is not limited to the diffusion model, using it offers three main advantages: (1) Flexible C-S trade-off control. As shown in Fig. 5, we can flexibly control the C-S trade-off at both the training stage (top row) and the testing stage (bottom row) by adjusting the return step $T_{trans}$ of the diffusion model. With the increase of $T_{trans}$ , more style characteristics are transferred, yet the content structures may be ruined (e.g., the last column). When proper $T_{trans}$ is adopted, e.g., $T_{trans} = 301$ , the sweet spot can be well achieved. Interestingly, as shown in the last two columns of the bottom row, though the model is trained on $T_{trans} = 301$ , we can extrapolate the style by using larger $T_{trans}$ (e.g., 401) at the testing stage (but the results may be degraded when using too large $T_{trans}$ , e.g., 601). It provides a very flexible way for users to adjust the results according to their preferences. This property, however, cannot be simply achieved by using other models, e.g., the widely used AEs [27, 50], since our framework does not involve any feature transforms [27, 50] or C-S losses trade-off [3]. (2) Higher-quality stylizations. Owing to the strong generative ability of the diffusion model, it can achieve higher-quality stylizations than other models. For comparison, we use the pre-trained VGG-AE [27, 46] as the style transfer module and fine-tune its decoder network for each style. As shown in column (b) of Fig. 6, though the results are still acceptable, they may produce distorted contents and inferior textures, clearly worse than the results generated by the diffusion model in column (a). This is also validated by the bottom quantitative scores. It signifies that the diffusion model can better learn the disentangled + +![](images/d6397c032e8816519cced80362c26afb7c20fc6b9c75b61afce656a5be8fa0cf.jpg) +Figure 5. Control of C-S trade-off by adjusting the return step $T_{trans}$ of the style transfer module. The top row shows adjusting $T_{trans}$ at the training stage while fixing $T_{trans} = 301$ at the testing stage. The bottom row shows adjusting $T_{trans}$ at the testing stage while fixing $T_{trans} = 301$ at the training stage. * denotes our default setting. Zoom-in for better comparison. See SM for quantitative analyses. + +![](images/de144b610ab417f1fdbba4a40504f3455bc799d6718e8da644e93de0354cf834.jpg) +Figure 6. Diffusion-based vs. AE-based style transfer. + +![](images/7ed65172626828e48354751e2d7c738a8585ea16633fc9543a789e0057e9741b.jpg) + +content and style characteristics in our framework, helping produce better style transfer results. (3) Diversified style transfer. As mentioned in Sec. 4.2, during inference, we can directly adopt the stochastic DDPM [25] forward process (Eq. (2)) to obtain diverse results (see SM). The diverse results can give users endless choices to obtain more satisfactory results. However, using other models like AEs in our framework cannot easily achieve it [81]. + +Loss Analyses. To verify the effectiveness of each loss term used for fine-tuning our StyleDiffusion, we present ablation study results in Fig. 7 (a-d). (1) Using L1 loss $\mathcal{L}_{SD}^{L1}$ successfully transfers the cubism style like the blocky patterns in the top row, but the colors stray from the style images, especially in the bottom row. It is consistent with our earlier analyses in Sec. 4.3 that the L1 loss is prone to produce implausible results outside the style domain. (2) Adding direction loss $\mathcal{L}_{SD}^{dir}$ helps pull the results closer to the style domain. The textures are enhanced in the top row, and the colors are more plausible in the top and bottom rows. (3) By further coordinating with the style reconstruction prior $\mathcal{L}_{SR}$ , the stylization effects are significantly elevated where the style information is recovered more sup + +![](images/f52741d8174e9d4f012e5b04328b8f0f80d528a3b5b9ee92eaf7d4f4d9f0cfaa.jpg) +Figure 7. Ablation study on loss functions. * denotes our full model. Zoom-in for better comparison. + +![](images/290b33325c8f5254e246f04302a417f9b7d45eac542a7530f3b2edcda07bbd0e.jpg) + +![](images/3aeab92b7b3c4a9375ce0fade7133235f5164b55e04c3c5ba1afa693f81bc3a3.jpg) + +![](images/3c44f8bc9590fa95abe69a8bcf01cdf08115889609fc9967eafe80b815a43aae.jpg) + +![](images/74f0f139e37338b03d1763dfcd660f2d625447ca0b0f9964d858e9f0b402af57.jpg) + +![](images/c1e0a8238a89ea23aebe0dac3a707991c8dcbbac643edae9c529e593c9041d60.jpg) + +![](images/76cf3ba0e08d752ca646da0a3b51039468eb169f5119c2d672e4b58bce884899.jpg) + +ficiently. It may be because it provides a good initialization for the optimization of $\mathcal{L}_{SD}^{L1}$ and $\mathcal{L}_{SD}^{dir}$ , which helps them give full play to their abilities. As verified in Fig. 7 (d), using the style reconstruction alone cannot learn meaningful style patterns except for basic colors. All the above analyses are also supported by the bottom quantitative scores. + +Comparison with Gram Loss. To further verify the superiority of our proposed losses, we replace them with the widely used Gram Loss [18, 27] in Fig. 7 (e-f). As can be observed, Gram Loss destroys the content structures severely, e.g., the zebra head in the top row and the enlarged area in the bottom row. This is because it does not disentangle C-S and only matches the global statistics without considering the relationship between C-S. In contrast, our losses focus on learning the disentangled style information apart from the content, which is induced by the difference between content and its stylized result. Therefore, they can better understand the relationship between C-S, achieving more satisfactory results with fine style details and better-preserved contents, as validated by Fig. 7 (c) and the bottom quantitative scores. Furthermore, we also conduct comparisons between our proposed losses and Gram Loss [18, 27] on the AE baseline [27, 46] to eliminate the impact of diffusion models. As shown in Fig. 8 (a-b), our losses can achieve more satisfactory results than Gram Loss, which is consistent with the results in Fig. 7. Moreover, as shown in Fig. 8 (c), they can also be combined with Gram Loss to improve the performance on the Style Loss metric. However, it may affect the full disentanglement of C-S in our framework, which strays from our target and decreases the content preservation (see SSIM score in Fig. 8 (c)). Therefore, we do not incorporate Gram Loss in our framework by default. + +# 6. Conclusion and Limitation + +In this work, we present a new framework for more interpretable and controllable C-S disentanglement and style transfer. Our framework, termed StyleDiffusion, leverages + +![](images/39eefed754ea71adc992013a9465c3efc10cfaa82c74b273da2b5bcdadf0ab85.jpg) +Figure 8. More loss function ablation study on the AE baseline. + +![](images/0a6ea978a6ec74a27740154edef2bed2cabdb04889ae7ec647d485548beaf9b1.jpg) + +![](images/de7a1aac3d645bb0bba6fdff43f6a046fc2dad5446037e41a1ece42899b0a3f5.jpg) + +![](images/91ca077552027d6c6f07d7947471b6b15a217c01a601c2cffbaa33cca39b735b.jpg) + +diffusion models to explicitly extract the content information and implicitly learn the complementary style information. A novel CLIP-based style disentanglement loss coordinated with a style reconstruction prior is also introduced to encourage the disentanglement and style transfer. Our method yields very encouraging stylizations, especially for challenging styles, and the experimental results verify its effectiveness and superiority against state of the art. + +Currently, the framework still suffers from several limitations: (1) The model needs to be fine-tuned for each style, and arbitrary style transfer is left to our future work. (2) The efficiency is not fast enough due to the use of diffusion models. Further research in accelerating diffusion sampling would be helpful. (3) There are some failure cases analyzed in SM, which may help inspire future improvements. Moreover, our framework may also be applied to other image translation [28] or manipulation [60] tasks, and we would like to explore them in our future work. + +# References + +[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 862-871, 2021. 1, 2, 6, 7 +[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 2 +[3] Mohammad Babaeizadeh and Golnaz Ghiasi. Adjustable real-time style transfer. In International Conference on Learning Representations (ICLR), 2019. 8 +[4] Andreas Blattmann, Robin Rombach, Kaan Oktay, Jonas Müller, and Björn Ommer. Retrieval-augmented diffusion models. Advances in Neural Information Processing Systems, 35:15309-15324, 2022. 2 +[5] Dongdong Chen, Lu Yuan, Jing Liao, Nenghai Yu, and Gang Hua. Stylebank: An explicit representation for neural image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1897-1906, 2017. 2 +[6] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems (NeurIPS), 34:26561-26573, 2021. 1, 2, 6, 7 +[7] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dualist: Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 872-881, 2021. 2 +[8] Haibo Chen, Lei Zhao, Huiming Zhang, Zhizhong Wang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Diverse image style transfer via invertible cross-space mapping. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14860-14869. IEEE Computer Society, 2021. 2 +[9] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in Neural Information Processing Systems (NeurIPS), 29, 2016. 1 +[10] Jiaxin Cheng, Ayush Jaiswal, Yue Wu, Pradeep Natarajan, and Prem Natarajan. Style-aware normalized loss for improving arbitrary style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 134-143, 2021. 2 +[11] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14347-14356. IEEE, 2021. 4 +[12] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Im + +age style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11326-11336, 2022. 2, 6, 7 +[13] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM International Conference on Multimedia (ACM MM), pages 2719-2727, 2020. 2, 6, 7 +[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021. 2, 3, 4, 6 +[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2 +[16] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 579-587, 2023. 2 +[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 5 +[18] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 1, 2, 5, 6, 7, 9 +[19] Rafael C Gonzalez. Digital image processing. Pearson education india, 2009. 4 +[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 1, 2 +[21] Daniel J Graham, James M Hughes, Helmut Leder, and Daniel N Rockmore. Statistics, vision, and the analysis of artistic style. Wiley Interdisciplinary Reviews: Computational Statistics, 4(2):115-123, 2012. 2 +[22] Shuyang Gu, Congliang Chen, Jing Liao, and Lu Yuan. Arbitrary style transfer with deep feature reshuffle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8222-8231, 2018. 2 +[23] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2 +[24] Irina Higgins, David Amos, David Pfau, Sebastien Racaniere, Loic Matthew, Danilo Rezende, and Alexander Lerchner. Towards a definition of disentangled representations. arXiv preprint arXiv:1812.02230, 2018. 1, 2 +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information + +Processing Systems (NeurIPS), 33:6840-6851, 2020. 2, 3, 4, 8 +[26] Kibeom Hong, Seogkyu Jeon, Huan Yang, Jianlong Fu, and Hyeran Byun. Domain-aware universal style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[27] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1501-1510, 2017. 1, 2, 4, 7, 8, 9 +[28] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 172-189, 2018. 2, 9 +[29] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14861-14869, 2021. 2 +[30] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4369-4376, 2020. 2 +[31] Yongcheng Jing, Yezhou Yang, Zunlei Feng, Jingwen Ye, Yizhou Yu, and Mingli Song. Neural style transfer: A review. IEEE Transactions on Visualization and Computer Graphics (TVCG), 26(11):3365-3385, 2019. 2 +[32] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711. Springer, 2016. 1, 2, 6, 7 +[33] Henry Kang, Seungyong Lee, and Charles K Chui. Coherent line drawing. In Proceedings of the 5th international symposium on Non-photorealistic animation and rendering, pages 43–50, 2007. 4 +[34] Theofanis Karaletsos, Serge Belongie, and Gunnar Ratsch. Bayesian representation learning with oracle constraints. arXiv preprint arXiv:1506.05011, 2015. 1, 2 +[35] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, and Holger Winnemoeller. Recognizing image style. arXiv preprint arXiv:1311.3715, 2013. 2 +[36] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019. 2 +[37] Hadi Kazemi, Seyed Mehdi Iranmanesh, and Nasser Nasrabadi. Style and content disentanglement in generative adversarial networks. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 848-856. IEEE, 2019. 2 +[38] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition (CVPR), pages 2426-2435, 2022. 2, 3, 4, 6 +[39] Hyunjik Kim and Andriy Mnih. Disentangling by factorising. In International Conference on Machine Learning (ICML), pages 2649-2658. PMLR, 2018. 2 +[40] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6 +[41] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 2, 7 +[42] Dmytro Kotovenko, Artsiom Sanakoyeu, Sabine Lang, and Bjorn Ommer. Content and style disentanglement for artistic style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4422-4431, 2019. 1, 2, 5 +[43] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Bjorn Ommer. Rethinking style transfer: From pixels to parameterized brushstrokes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12196-12205, 2021. 2 +[44] Tejas D Kulkarni, William F Whitney, Pushmeet Kohli, and Josh Tenenbaum. Deep convolutional inverse graphics network. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 1 +[45] Gihyun Kwon and Jong Chul Ye. Diagonal attention and style-based gan for content-style disentanglement in image generation and translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13980-13989, 2021. 2 +[46] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18062-18071, 2022. 5, 8, 9 +[47] Berel Lang. The concept of style. Cornell University Press, 1987. 4 +[48] Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Maneesh Singh, and Ming-Hsuan Yang. Diverse image-to-image translation via disentangled representations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 35-51, 2018. 2 +[49] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[50] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. In Advances in Neural Information Processing Systems (NeurIPS), pages 386-396, 2017. 1, 2, 4, 8 +[51] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2 +[52] Tianwei Lin, Zhuoqi Ma, Fu Li, Dongliang He, Xin Li, Errui Ding, Nannan Wang, Jie Li, and Xinbo Gao. Drafting and + +revision: Laplacian pyramid network for fast high-quality artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5141-5150, 2021. 2, 6, 7 +[53] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6649-6658, 2021. 2, 6, 7 +[54] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In International Conference on Machine Learning (ICML), pages 4114-4124. PMLR, 2019. 1, 2, 7 +[55] Haofei Lu and Zhizhong Wang. Universal video style transfer via crystallization, separation, and blending. In Proceedings of the International Joint Conferences on Artificial Intelligence Organization (IJCAI), Vienna, Austria, pages 23-29, 2022. 2 +[56] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2 +[57] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171. PMLR, 2021. 2, 3 +[58] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), pages 16784-16804. PMLR, 2022. 2 +[59] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5880–5888, 2019. 2 +[60] Taesung Park, Jun-Yan Zhu, Oliver Wang, Jingwan Lu, Eli Shechtman, Alexei Efros, and Richard Zhang. Swapping autoencoder for deep image manipulation. Advances in Neural Information Processing Systems (NeurIPS), 33:7198-7211, 2020. 2, 9 +[61] DM Parker and Jan B Deregowski. Perception and artistic style. Elsevier, 1991. 2 +[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 5, 7 +[63] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 5 + +[64] Xuanchi Ren, Tao Yang, Yuwang Wang, and Wenjun Zeng. Rethinking content and style: Exploring bias for unsupervised disentanglement. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1823-1832. IEEE, 2021. 1, 2 +[65] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2 +[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-assisted Intervention, pages 234-241. Springer, 2015. 3 +[67] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 6 +[68] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[69] Artsiom Sanakoyeu, Dmytro Kotovenko, Sabine Lang, and Bjorn Ommer. A style-aware content loss for real-time hd style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), pages 698-714, 2018. 1 +[70] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8242-8250, 2018. 2 +[71] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1 +[72] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICML), pages 2256-2265. PMLR, 2015. 2 +[73] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2020. 2, 3, 4 +[74] Jan Svoboda, Asha Anoosheh, Christian Osendorfer, and Jonathan Masci. Two-stage peer-regularized feature recombination for arbitrary image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13816–13825, 2020. 1, 6, 7 +[75] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6924-6932, 2017. 2 + +[76] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2 +[77] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 7 +[78] Pei Wang, Yijun Li, and Nuno Vasconcelos. Rethinking and improving the robustness of image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 124-133, 2021. 2 +[79] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: Towards aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia (ACM MM), pages 1095-1106, 2022. 2 +[80] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2624-2632, 2022. 2 +[81] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7789-7798, 2020. 2, 4, 8 +[82] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Evaluate and improve the quality of neural style transfer. Computer Vision and Image Understanding (CVIU), 207:103203, 2021. 2 +[83] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. DivSwapper: Towards diversified patch-based arbitrary style transfer. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), pages 4980–4987, 2022. 2 +[84] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Glstylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8):575-586, 2020. 2 +[85] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2742-2750, 2023. 2 +[86] Wayne Wu, Kaidi Cao, Cheng Li, Chen Qian, and Chen Change Loy. Disentangling content and style via unsupervised geometry distillation. arXiv preprint arXiv:1905.04538, 2019. 2 +[87] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14618-14627, 2021. 2 +[88] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 + +[89] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1395-1403, 2015. 4 +[90] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8035-8045, 2022. 1, 2, 6, 7 +[91] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 2 +[92] Yu Zhang, Peter Tino, Ales Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1 +[93] Yexun Zhang, Ya Zhang, and Wenbin Cai. Separating style and content for generalized style transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8447-8455, 2018. 1, 2 +[94] Zhiwen Zuo, Lei Zhao, Shuobin Lian, Haibo Chen, Zhizhong Wang, Ailin Li, Wei Xing, and Dongming Lu. Style fader generative adversarial networks for style degree controllable artistic style transfer. In Proc. Int. Joint Conf. on Artif. Intell.(IJCAI), pages 5002-5009, 2022. 2 \ No newline at end of file diff --git a/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/images.zip b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fa3e62c33b3f390223ae6a3a2f20727c859dacc3 --- /dev/null +++ b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf83cb1cb8a7a588de61410d2b23a971c9427d93ba69cedc5daa10866354ec4b +size 1065567 diff --git a/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/layout.json b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..789d216b498115b28f7bd3ebbaa247793b9e0ce3 --- /dev/null +++ b/2023/StyleDiffusion_ Controllable Disentangled Style Transfer via Diffusion Models/layout.json @@ -0,0 +1,11327 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 538, + 121 + ], + "type": "text", + "content": "StyleDiffusion: Controllable Disentangled Style Transfer via Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 143, + 459, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 143, + 459, + 185 + ], + "spans": [ + { + "bbox": [ + 134, + 143, + 459, + 185 + ], + "type": "text", + "content": "Zhizhong Wang*, Lei Zhao† Wei Xing \nCollege of Computer Science and Technology, Zhejiang University \n{endywon, czhl, WXING}@zju.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 489 + ], + "type": "text", + "content": "Content and style (C-S) disentanglement is a fundamental problem and critical challenge of style transfer. Existing approaches based on explicit definitions (e.g., Gram matrix) or implicit learning (e.g., GANs) are neither interpretable nor easy to control, resulting in entangled representations and less satisfying results. In this paper, we propose a new C-S disentangled framework for style transfer without using previous assumptions. The key insight is to explicitly extract the content information and implicitly learn the complementary style information, yielding interpretable and controllable C-S disentanglement and style transfer. A simple yet effective CLIP-based style disentanglement loss coordinated with a style reconstruction prior is introduced to disentangle C-S in the CLIP image space. By further leveraging the powerful style removal and generative ability of diffusion models, our framework achieves superior results than state of the art and flexible C-S disentanglement and trade-off control. Our work provides new insights into the C-S disentanglement in style transfer and demonstrates the potential of diffusion models for learning well-disentangled C-S characteristics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 511, + 128, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 128, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 128, + 523 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 531, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 626 + ], + "type": "text", + "content": "Given a reference style image, e.g., Starry Night by Vincent Van Gogh, style transfer aims to transfer its artistic style, such as colors and brushstrokes, to an arbitrary content target. To achieve such a goal, it must first properly separate the style from the content and then transfer it to another content. This raises two fundamental challenges: (1) \"how to disentangle content and style (C-S)\" and (2) \"how to transfer style to another content\"." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "type": "text", + "content": "To resolve these challenges, valuable efforts have been devoted. Gatys et al. [18] proposed A Neural Algorithm of Artistic Style to achieve style transfer, which explicitly defines the high-level features extracted from a pre-trained Convolutional Neural Network (CNN) (e.g., VGG [71]) as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 213, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 547, + 406 + ], + "type": "text", + "content": "content, and the feature correlations (i.e., Gram matrix) as style. This approach acquires visually stunning results and inspires a large number of successors [32, 27, 50, 1, 90]. Despite the successes, by diving into the essence of style transfer, we observed three problems with these approaches: (1) The C-S are not completely disentangled. Theoretically, the C-S representations are intertwined. For example, matching the content representation of an image may also match its Gram matrix, and vice versa. (2) What CNN learned is a black box rugged to interpret [92], which makes the C-S definitions [18] uninterpretable and hard to control. (3) The transfer process is modeled as a separate optimization of content loss and style loss [18], so there lacks a deep understanding of the relationship between C-S. These problems usually lead to unbalanced stylizations and disharmonious artifacts [6], as will be shown in later Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 410, + 547, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 547, + 650 + ], + "type": "text", + "content": "On the other hand, disentangled representation learning [24] provides other ideas to implicitly disentangle C-S, either supervised [44, 34] or unsupervised [9, 93]. For style transfer, Kotovenko et al. [42] utilized fixpoint triplet style loss and disentanglement loss to enforce a GAN [20]-based framework to learn separate C-S representations in an unsupervised manner. Similarly, TPFR [74] learned to disentangle C-S in latent space via metric learning and two-stage peer-regularization, producing high-quality images even in the zero-shot setting. While these approaches successfully enforce properties \"encouraged\" by the corresponding losses, they still have three main problems: (1) Well-disentangled models seemingly cannot be identified without supervision [54, 64], which means the unsupervised learning [42, 74] may not achieve truly disentangled C-S, as will be shown in later Fig. 3. (2) These approaches are all based on GANs and thus often confined to the GAN predefined domains, e.g., a specific artist's style domain [69]. (3) The implicitly learned C-S representations are still black boxes that are hard to interpret and control [54]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 547, + 715 + ], + "type": "text", + "content": "Facing the challenges above, in this paper, we propose a new C-S disentangled framework for style transfer without using previous assumptions such as Gram matrix [18] or GANs [42]. Our key insight stems from the fact that the definition of an image's style is much more complex" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 693, + 283, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 283, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 283, + 703 + ], + "type": "text", + "content": "*This work was done when Zhizhong Wang was an intern at Huawei." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7677" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "than its content, e.g., we can easily identify the content of a painting by its structures, semantics, or shapes, but it is intractable to define the style [61, 21, 35, 82]. Therefore, we can bypass such a dilemma by explicitly extracting the content information and implicitly learning its complementary style information. Since we strictly constrain style as the complement of content, the C-S can be completely disentangled, and the control of disentanglement has been transformed into the control of content extraction. It achieves both controllability and interpretability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 502 + ], + "type": "text", + "content": "However, achieving plausible and controllable content extraction is also non-trivial because the contents extracted from the content images and style images should share the same content domain, and the details of the extracted contents should be easy to control. To this end, we resort to recent developed diffusion models [25, 73] and introduce a diffusion-based style removal module to smoothly dispel the style information of the content and style images, extracting the domain-aligned content information. Moreover, owing to the strong generative capability of diffusion models, we also introduce a diffusion-based style transfer module to better learn the disentangled style information of the style image and transfer it to the content image. The style disentanglement and transfer are encouraged via a simple yet effective CLIP [62]-based style disentanglement loss, which induces the transfer mapping of the content image's content to its stylization (i.e., the stylized result) to be aligned with that of the style image's content to its stylization (i.e., the style image itself) in the CLIP image space. By further coordinating with a style reconstruction prior, it achieves both generalized and faithful style transfer. We conduct comprehensive comparisons and ablation study to demonstrate the effectiveness and superiority of our framework. With the well-disentangled C-S, it achieves very promising stylizations with fine style details, well-preserved contents, and a deep understanding of the relationship between C-S." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 503, + 239, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 503, + 239, + 514 + ], + "spans": [ + { + "bbox": [ + 59, + 503, + 239, + 514 + ], + "type": "text", + "content": "In summary, our contributions are threefold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 522, + 287, + 658 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 57, + 522, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 287, + 570 + ], + "type": "text", + "content": "- We propose a novel C-S disentangled framework for style transfer, which achieves more interpretable and controllable C-S disentanglement and higher-quality stylized results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 573, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 573, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 57, + 573, + 287, + 619 + ], + "type": "text", + "content": "- We introduce diffusion models to our framework and demonstrate their effectiveness and superiority in controllable style removal and learning well-disentangled C-S characteristics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 622, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 622, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 622, + 287, + 658 + ], + "type": "text", + "content": "- A new CLIP-based style disentanglement loss coordinated with a style reconstruction prior is introduced to disentangle C-S in the CLIP image space." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 134, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Neural Style Transfer (NST). The pioneering work of Gatys et al. [18] has opened the era of NST [31]. Since" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "then, this task has experienced tremendous progress, including efficiency [32, 49, 85], quality [22, 70, 84, 52, 10, 7, 1, 43, 78, 53, 6, 87, 29, 94, 12, 91, 79], generality [5, 27, 50, 59, 13, 30, 26, 80, 90, 55, 88], and diversity [75, 81, 83]. Despite these successes, the essence of these approaches is mostly based on the explicitly defined C-S representations, such as Gram matrix [18], which have several limitations as discussed in Sec. 1. In our work, we propose new disentangled C-S representations explicitly extracted or implicitly learned by diffusion models, achieving more effective style transfer and higher-quality results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 204, + 546, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 546, + 371 + ], + "type": "text", + "content": "Disentangled Representation Learning (DRL). The task of DRL [24] aims at modeling the factors of data variations [48]. Earlier works used labeled data to factorize representations in a supervised manner [34]. Recently, unsupervised settings have been largely explored [39], especially for disentangling style from content [93, 28, 48, 37, 86, 42, 60, 64, 8, 45]. However, due to the dependence on GANs [20], their C-S disentanglement is usually restricted in the GAN pre-defined domains (e.g., Van Gogh's style domain). Besides, disentanglement cannot be effectively achieved without providing sufficient data [54]. In contrast, our framework learns the disentangled style from a single style image, and the disentanglement can be easily achieved by providing only a few (" + }, + { + "bbox": [ + 304, + 204, + 546, + 371 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 304, + 204, + 546, + 371 + ], + "type": "text", + "content": "50) content images for training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 372, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 547, + 552 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models [72] such as denoising diffusion probabilistic models (DDPMs) [25, 57] have recently shown great success in image generation [73, 14, 16], image manipulation [56, 2, 38], and text-conditional synthesis [58, 68, 63, 65, 23, 4, 51]. These works have demonstrated the power of diffusion models to achieve higher-quality results than other generative models like VAEs [76], auto-regressive models [15], flows [41], and GANs [36]. Inspired by them, we introduce a diffusion-based style removal module and a style transfer module in our framework. These modules can smoothly remove the style information of images and better learn the recovery of it to achieve higher-quality style transfer results. To the best of our knowledge, our work is the first to introduce diffusion models to the field of neural style transfer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 559, + 384, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 384, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 384, + 573 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "type": "text", + "content": "Denoising diffusion probabilistic models (DDPMs) [72, 25] are latent variable models that consist of two diffusion processes, i.e., a forward diffusion process and a reverse diffusion process. The forward process is a fixed Markov Chain that sequentially produces a series of latents " + }, + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "type": "inline_equation", + "content": "x_{1}, \\ldots, x_{T}" + }, + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "type": "text", + "content": " by gradually adding Gaussian noises at each timestep " + }, + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "type": "inline_equation", + "content": "t \\in [1, T]" + }, + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 670, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 670, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 347, + 670, + 545, + 685 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) := \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "1120ade23c50f96f20301cbcbdcc2ee9294aa8b9e892033b0b98ef33b5f8b6c6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{t} \\in (0,1)" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is a fixed variance schedule. An important property of the forward process is that given clean data" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7678" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 534, + 270 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 534, + 270 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 534, + 270 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 534, + 270 + ], + "type": "image", + "image_path": "eefdcc0c02e6e3253a5130b9b1f512b64ec36b470b7a294df09b936ff28bea86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "text", + "content": "Figure 1. Overview of our proposed StyleDiffusion. The content image " + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "text", + "content": " and style image " + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "text", + "content": " are first fed into a diffusion-based style removal module to explicitly extract the domain-aligned content information. Then, the content of " + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "text", + "content": " is fed into a diffusion-based style transfer module to obtain the stylized result " + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "text", + "content": ". During training, we fine-tune the style transfer module via a CLIP-based style disentanglement loss " + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SD}" + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "text", + "content": " coordinated with a style reconstruction prior (see details in Sec. 4.3, we omit it here for brevity) to implicitly learn the disentangled style information of " + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 276, + 547, + 332 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 351, + 187, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 187, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 187, + 363 + ], + "type": "inline_equation", + "content": "x_0, x_t" + }, + { + "bbox": [ + 47, + 351, + 187, + 363 + ], + "type": "text", + "content": " can be directly sampled as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 371, + 287, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 371, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 93, + 371, + 287, + 392 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {0}\\right) := \\mathcal {N} (\\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}), \\tag {2}", + "image_path": "201ef3cd7243dbbb324a29058cc6363a8ba7928dda3ba846253af754f2985041.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 388, + 230, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 388, + 230, + 401 + ], + "spans": [ + { + "bbox": [ + 121, + 388, + 230, + 401 + ], + "type": "interline_equation", + "content": "x _ {t} := \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon ,", + "image_path": "57bdd7492b3aff86fe0546d1a98e073e14d6d474e1aa37b4ae9084c132bb1721.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\alpha_{t} := 1 - \\beta_{t}" + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} := \\prod_{s=0}^{t} \\alpha_{s}" + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "text", + "content": ". Noise " + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "text", + "content": " has the same dimensionality as data " + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "text", + "content": " and latent " + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 411, + 287, + 436 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "text", + "content": "The reverse process generates a reverse sequence by sampling the posteriors " + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "inline_equation", + "content": "q(x_{t - 1}|x_t)" + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "text", + "content": ", starting from a Gaussian noise sample " + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "text", + "content": ". However, since " + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "inline_equation", + "content": "q(x_{t - 1}|x_t)" + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "text", + "content": " is intractable, DDPMs learn parameterized Gaussian transitions " + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t)" + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "text", + "content": " with a learned mean " + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "text", + "content": " and a fixed variance " + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\sigma_t^2\\mathbf{I}" + }, + { + "bbox": [ + 46, + 437, + 287, + 509 + ], + "type": "text", + "content": " [25]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 517, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 517, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 94, + 517, + 287, + 533 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (x _ {t - 1} | x _ {t}) := \\mathcal {N} (\\mu_ {\\theta} (x _ {t}, t), \\sigma_ {t} ^ {2} \\mathbf {I}), \\tag {3}", + "image_path": "bd20fa649013fb5aa1a944013cc5a5dbfa26bbcf99282527097de3f1904cdc65.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 540, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 287, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 540, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 47, + 540, + 287, + 564 + ], + "type": "text", + "content": " is the function of a noise approximator " + }, + { + "bbox": [ + 47, + 540, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 47, + 540, + 287, + 564 + ], + "type": "text", + "content": ". Then, the reverse process can be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 72, + 571, + 287, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 571, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 72, + 571, + 287, + 598 + ], + "type": "interline_equation", + "content": "x _ {t - 1} := \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right) + \\sigma_ {t} \\mathbf {z}, \\quad (4)", + "image_path": "82014a03db9cc549eb4574fae7d77400810e698950f362086c9ae6ce2380ca00.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 606, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 606, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 606, + 287, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 606, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 47, + 606, + 287, + 643 + ], + "type": "text", + "content": " is a standard Gaussian noise independent of " + }, + { + "bbox": [ + 47, + 606, + 287, + 643 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 606, + 287, + 643 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 606, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_{t}, t)" + }, + { + "bbox": [ + 47, + 606, + 287, + 643 + ], + "type": "text", + "content": " is learned by a deep neural network [66] through optimizing the following loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 118, + 651, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 651, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 118, + 651, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\| \\epsilon_ {\\theta} (x _ {t}, t) - \\epsilon \\| ^ {2}. \\tag {5}", + "image_path": "f7e8614cbf5834d67f581230e9cd62cf4f505f26d01bf565acf36b9bd2ab8eeb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "Later, instead of using the fixed variances, Nichol and Dhariwal [57] presented a strategy for learning the variances. Song et al. [73] proposed DDIM, which formulates" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 351, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 351, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 545, + 387 + ], + "type": "text", + "content": "an alternative non-Markovian noising process that has the same forward marginals as DDPM but allows a different reverse process:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 396, + 545, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 425 + ], + "type": "interline_equation", + "content": "x _ {t - 1} := \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (x _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} (x _ {t}, t) + \\sigma_ {t} \\mathbf {z}, \\tag {6}", + "image_path": "ea97121fb1472b9fdd467e47bcad9e3a15d62cce630f66ef833fdc85f35b69eb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x_t, t)" + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "text", + "content": " is the predicted " + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, t)" + }, + { + "bbox": [ + 305, + 426, + 545, + 450 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 457, + 545, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 457, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 351, + 457, + 545, + 485 + ], + "type": "interline_equation", + "content": "f _ {\\theta} (x _ {t}, t) := \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (x _ {t} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {7}", + "image_path": "0ffd725bdfc4711089752107d378ea0776a941aceb6930d571bf9637f211a977.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": "Changing the choice of " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": " values in Eq. (6) can achieve different reverse processes. Especially when " + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\sigma_{t} = 0" + }, + { + "bbox": [ + 304, + 491, + 545, + 564 + ], + "type": "text", + "content": ", which is called DDIM [73], the reverse process becomes a deterministic mapping from latents to images, which enables nearly perfect inversion [38]. Besides, it can also accelerate the reverse process with much fewer sampling steps [14, 38]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 574, + 361, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 361, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 361, + 586 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "Our task can be described as follows: given a style image " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " and an arbitrary content image " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": ", we want to first disentangle the content and style of them and then transfer the style of " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " to the content of " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": ". To achieve so, as stated in Sec. 1, our key idea is to explicitly extract the content information and then implicitly learn the complementary style information. Since our framework is built upon diffusion models [25, 73], we dub it StyleDiffusion." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Fig. 1 shows the overview of our StyleDiffusion, which consists of three key ingredients: I) a diffusion-based style" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7679" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "removal module, II) a diffusion-based style transfer module, and III) a CLIP-based style disentanglement loss coordinated with a style reconstruction prior. In the following subsections, we will introduce each of them in detail." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 128, + 176, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 128, + 176, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 176, + 140 + ], + "type": "text", + "content": "4.1. Style Removal Module" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 146, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 287, + 289 + ], + "type": "text", + "content": "The style removal module aims at removing the style information of the content and style images, explicitly extracting the domain-aligned content information. Any reasonable content extraction operation can be used, depending on how the users define the content. For instance, users may want to use the structural outline as the content, so they can extract the outlines [33, 89] here. However, as discussed in Sec. 1, one challenge is controllability since the control of C-S disentanglement has been transformed into the control of content extraction. To this end, we introduce a diffusion-based style removal module to achieve both plausible and controllable content extraction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "text", + "content": "Given an input image, e.g., the style image " + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "text", + "content": ", since the color is an integral part of style [47], our style removal module first removes its color by a commonly used ITU-R 601-2 luma transform [19]. The obtained grayscale image is denoted as " + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "inline_equation", + "content": "I_{s}^{\\prime}" + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "text", + "content": ". Then, we leverage a pre-trained diffusion model [14] " + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "text", + "content": " to remove the style details such as brushstrokes and textures of " + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "inline_equation", + "content": "I_{s}^{\\prime}" + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "text", + "content": ", extracting the content " + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "inline_equation", + "content": "I_{s}^{c}" + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "text", + "content": ". The insight is that the pre-trained diffusion model can help eliminate the domain-specific characteristics of input images and align them to the pre-trained domain [11, 38]. We assume that images with different styles belong to different domains, but their contents should share the same domain. Therefore, we can pre-train the diffusion model on a surrogate domain, e.g., the photograph domain, and then use this domain to construct the contents of images. After pretraining, the diffusion model can convert the input images from diverse domains to the latents " + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 290, + 288, + 529 + ], + "type": "text", + "content": " via the forward process and then inverse them to the photograph domain via the reverse process. In this way, the style characteristics can be ideally dispelled, leaving only the contents of the images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": "Specifically, in order to obtain the results with fewer sampling steps and ensure that the content structures of the input images can be well preserved, we adopt the deterministic DDIM [73] sampling as the reverse process (Eq. (8)), and the ODE approximation of its reversal [38] as the forward process (Eq. (9)):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 609, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 609, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 63, + 609, + 287, + 624 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} f _ {\\theta} (x _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\epsilon_ {\\theta} (x _ {t}, t), \\tag {8}", + "image_path": "ef99ab99c1e5a6ca0a0bdf3c918a1a6d0ca39b150cc4638720ce75da348c2dad.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 634, + 287, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 634, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 63, + 634, + 287, + 648 + ], + "type": "interline_equation", + "content": "x _ {t + 1} = \\sqrt {\\bar {\\alpha} _ {t + 1}} f _ {\\theta} (x _ {t}, t) + \\sqrt {1 - \\bar {\\alpha} _ {t + 1}} \\epsilon_ {\\theta} (x _ {t}, t), \\tag {9}", + "image_path": "fb431d7739a135c47dbaa80c30ba85eeeeae2c7b418b8bab5398b39024df870f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " is defined in Eq. (7). The forward and reverse diffusion processes enable us to easily control the intensity of style removal by adjusting the number of return step " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "T_{remov}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " (see details in later Sec. 5.1). With the increase of " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "T_{remov}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": ", more style characteristics will be removed, and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "the main content structures are retained, as will be shown in later Sec. 5.3. Note that for content images that are photographs, the diffusion processes are optional since they are already within the pre-trained domain, and there is almost no style except the colors to be dispelled. The superiority of diffusion-based style removal against other operations, such as Auto-Encoder (AE) [50]-based style removal, can be found in supplementary material (SM)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 175, + 434, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 175, + 434, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 434, + 188 + ], + "type": "text", + "content": "4.2. Style Transfer Module" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 194, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 290 + ], + "type": "text", + "content": "The style transfer module aims to learn the disentangled style information of the style image and transfer it to the content image. A common generative model like AEs [27] can be used here. However, inspired by the recent great success of diffusion models [14, 38], we introduce a diffusion-based style transfer module, which can better learn the disentangled style information in our framework and achieve higher-quality and more flexible stylizations (see Sec. 5.3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": "Given a content image " + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": ", denote " + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "inline_equation", + "content": "I_{c}^{c}" + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": " is the content of " + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": " extracted by the style removal module (Sec. 4.1). We first convert it to the latent " + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": " using a pre-trained diffusion model " + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": ". Then, guided by a CLIP-based style disentanglement loss coordinated with a style reconstruction prior (Sec. 4.3), the reverse process of the diffusion model is fine-tuned " + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "inline_equation", + "content": "(\\epsilon_{\\theta} \\rightarrow \\epsilon_{\\hat{\\theta}})" + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": " to generate the stylized result " + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": " referenced by the style image " + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": ". Once the fine-tuning is completed, any content image can be manipulated into the stylized result with the disentangled style of the style image " + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 290, + 546, + 469 + ], + "type": "text", + "content": ". To make the training easier and more stable, we adopt the deterministic DDIM forward and reverse processes in Eq. (8) and Eq. (9) during the fine-tuning. However, at inference, the stochastic DDPM [25] forward process (Eq. (2)) can also be used directly to help obtain diverse results [81] (Sec. 5.3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 477, + 477, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 477, + 477, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 477, + 490 + ], + "type": "text", + "content": "4.3. Loss Functions and Fine-tuning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 495, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 579 + ], + "type": "text", + "content": "Enforcing the style transfer module (Sec. 4.2) to learn and transfer the disentangled style information should address two key questions: (1) \"how to regularize the learned style is disentangled\" and (2) \"how to aptly transfer it to other contents\". To answer these questions, we introduce a novel CLIP-based style disentanglement loss coordinated with a style reconstruction prior to train the networks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "text", + "content": "CLIP-based Style Disentanglement Loss. Denote " + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "inline_equation", + "content": "I_{c}^{c}" + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "inline_equation", + "content": "I_{s}^{c}" + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "text", + "content": " are the respective contents of the content image " + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "text", + "content": " and the style image " + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "text", + "content": " extracted by the style removal module (Sec. 4.1). We aim to learn the disentangled style information of the style image " + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "text", + "content": " complementary to its content " + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "inline_equation", + "content": "I_{s}^{c}" + }, + { + "bbox": [ + 304, + 579, + 545, + 662 + ], + "type": "text", + "content": ". Therefore, a straightforward way to obtain the disentangled style information is a direct subtraction:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 392, + 673, + 545, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 673, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 392, + 673, + 545, + 686 + ], + "type": "interline_equation", + "content": "D _ {s} ^ {p x} = I _ {s} - I _ {s} ^ {c}. \\tag {10}", + "image_path": "27f0d02a859665349a3adb6fdbd0873e4a09c05f877001fdc96bf36cc5d460d3.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1 Unless otherwise specified, we do not use the diffusion processes for content images in order to better maintain the content structures." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7680" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 211, + 169 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 211, + 169 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 211, + 169 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 211, + 169 + ], + "type": "image", + "image_path": "43f281c92aa3a206fb934e88e358668a72ab7ec73a0254bb4084c0c2e37e7971.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 172, + 153, + 184 + ], + "lines": [ + { + "bbox": [ + 107, + 172, + 153, + 184 + ], + "spans": [ + { + "bbox": [ + 107, + 172, + 153, + 184 + ], + "type": "text", + "content": "(a) L1 loss" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 69, + 377, + 170 + ], + "blocks": [ + { + "bbox": [ + 215, + 69, + 377, + 170 + ], + "lines": [ + { + "bbox": [ + 215, + 69, + 377, + 170 + ], + "spans": [ + { + "bbox": [ + 215, + 69, + 377, + 170 + ], + "type": "image", + "image_path": "5ea1ad39b632f0621e39e7da76d120caf1fca0843fe9087d7e933b298a28251f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 172, + 337, + 184 + ], + "lines": [ + { + "bbox": [ + 264, + 172, + 337, + 184 + ], + "spans": [ + { + "bbox": [ + 264, + 172, + 337, + 184 + ], + "type": "text", + "content": "(b) Direction loss" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 192, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 547, + 237 + ], + "type": "text", + "content": "Figure 2. Illustration of different loss functions to transfer the disentangled style information. (a) L1 loss cannot guarantee the stylized results are within the style domain and may suffer from a collapse problem. (b) Direction loss aligns the disentangled directions but cannot realize accurate mappings. (c) Combining L1 loss and direction loss is able to achieve accurate one-to-one mappings from the content domain to the style domain." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 545, + 170 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 545, + 170 + ], + "type": "image", + "image_path": "6f34e880991651244fe8ad29e70daa1ac748236e07afaecd5c59ef722e061d78.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 172, + 512, + 184 + ], + "lines": [ + { + "bbox": [ + 417, + 172, + 512, + 184 + ], + "spans": [ + { + "bbox": [ + 417, + 172, + 512, + 184 + ], + "type": "text", + "content": "(c) Direction + L1 loss" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 256, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 304 + ], + "type": "text", + "content": "However, the simple pixel differences do not contain meaningful semantic information, thus cannot achieve plausible results [18, 42]. To address this problem, we can formulate the disentanglement in a latent semantic space:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 120, + 312, + 287, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 312, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 120, + 312, + 287, + 326 + ], + "type": "interline_equation", + "content": "D _ {s} = E \\left(I _ {s}\\right) - E \\left(I _ {s} ^ {c}\\right), \\tag {11}", + "image_path": "99815a1e5f91131fb3bb861fd19f46bd51a50f3d076d550467e8c38f87d7f458.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "text", + "content": " is a well-pre-trained projector. Specifically, since " + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "inline_equation", + "content": "I_{s}^{c}" + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "text", + "content": " have similar contents but with different styles, the projector " + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "text", + "content": " must have the ability to distinguish them in terms of the style characteristics. In other words, as we define that images with different styles belong to different domains, the projector " + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "text", + "content": " should be able to distinguish the domains of " + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "inline_equation", + "content": "I_{s}^{c}" + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "text", + "content": ". Fortunately, inspired by the recent vision-language model CLIP [62] that encapsulates knowledgeable semantic information of not only the photograph domain but also the artistic domain [17, 63, 46], we can use its image encoder as our projector " + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 333, + 287, + 572 + ], + "type": "text", + "content": " off the shelf. The open-domain CLIP space here serves as a good metric space to measure the \"style distance\" between content and its stylized result. This \"style distance\" thus can be interpreted as the disentangled style information. Note that here the style is implicitly defined as the complement of content, which is fundamentally different from the Gram matrix [18] that is an explicit style definition independent of content (see comparisons in Sec. 5.3). The comparisons between CLIP space and other possible spaces can be found in SM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 573, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 609 + ], + "type": "text", + "content": "After obtaining the disentangled style information " + }, + { + "bbox": [ + 46, + 573, + 287, + 609 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 46, + 573, + 287, + 609 + ], + "type": "text", + "content": ", the next question is how to properly transfer it to other contents. A possible solution is directly optimizing the L1 loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 119, + 616, + 287, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 616, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 119, + 616, + 287, + 636 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {c s} = E \\left(I _ {c s}\\right) - E \\left(I _ {c} ^ {c}\\right), \\\\ \\text {2} I _ {1} = \\text {＂} \\text {D} _ {1} = \\text {＂} \\text {D} _ {2} = \\text {＂} \\end{array} \\tag {12}", + "image_path": "3bece4aadd805dc612378ca40b293b965adb007073230e188f97713f998ee02f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 117, + 633, + 206, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 633, + 206, + 646 + ], + "spans": [ + { + "bbox": [ + 117, + 633, + 206, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S D} ^ {L 1} = \\parallel D _ {c s} - D _ {s} \\parallel ,", + "image_path": "756054d0624e033596c9e1836e17dc0bfc6ba82d5b25caaefd8d572b5831bde5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " is the stylized result, " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "D_{cs}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " is the disentangled style information of " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": ". However, as illustrated in Fig. 2 (a) and further validated in later Sec. 5.3, minimizing the L1 loss cannot guarantee the stylized result " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " is within the style domain of the style image " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": ". It is because L1 loss only" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 256, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 327 + ], + "type": "text", + "content": "minimizes the absolute pixel difference (i.e., Manhattan distance); thus, it may produce stylized images that satisfy the Manhattan distance but deviate from the target style domain in the transfer direction. Besides, it may also lead to a collapse problem where a stylized output meets the same Manhattan distance with different contents in the latent space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 328, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 352 + ], + "type": "text", + "content": "To address these problems, we can further constrain the disentangled directions as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 366, + 360, + 545, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 360, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 366, + 360, + 545, + 387 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S D} ^ {d i r} = 1 - \\frac {D _ {c s} \\cdot D _ {s}}{\\| D _ {c s} \\| \\| D _ {s} \\|.} \\tag {13}", + "image_path": "ff658f618c3a3817d4b0c25d05d2dbca80e5e7bb124304d0a3a9fc0f7b3b9208.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "text", + "content": "This direction loss aligns the transfer direction of the content image's content to its stylization (i.e., the stylized result) with the direction of the style image's content to its stylization (i.e., the style image itself), as illustrated in Fig. 2 (b). Collaborated with this loss, the L1 loss " + }, + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SD}^{L1}" + }, + { + "bbox": [ + 304, + 394, + 545, + 490 + ], + "type": "text", + "content": " thus can achieve accurate one-to-one mappings from contents in the content domain to their stylizations in the style domain, as illustrated in Fig. 2 (c)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 490, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 490, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 545, + 514 + ], + "type": "text", + "content": "Finally, our style disentanglement loss is defined as a compound of " + }, + { + "bbox": [ + 304, + 490, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SD}^{L1}" + }, + { + "bbox": [ + 304, + 490, + 545, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 490, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SD}^{dir}" + }, + { + "bbox": [ + 304, + 490, + 545, + 514 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 364, + 522, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 522, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 364, + 522, + 545, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S D} = \\lambda_ {L 1} \\mathcal {L} _ {S D} ^ {L 1} + \\lambda_ {d i r} \\mathcal {L} _ {S D} ^ {d i r}, \\tag {14}", + "image_path": "9aa3fceb3119121254d15f8f91b4466e523d80ee7d53d8ec919875f92f3da171.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\lambda_{L1}" + }, + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\lambda_{dir}" + }, + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "type": "text", + "content": " are hyper-parameters set to 10 and 1 in our experiments. Since our style information is induced by the difference between content and its stylized result, we can deeply understand the relationship between C-S through learning. As a result, the style can be naturally and harmoniously transferred to the content, leading to better stylized images, as will be shown in later Fig. 3." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Style Reconstruction Prior. To fully use the prior information provided by the style image and further elevate the stylization effects, we integrate a style reconstruction prior into the fine-tuning of the style transfer module. Intuitively, given the content " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{s}^{c}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " of the style image " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", the style transfer module should be capable of recovering it to the original style image as much as possible. Therefore, we can define" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7681" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 547, + 418 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 418 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 418 + ], + "type": "image", + "image_path": "9f812457c75c9843b7d0b042d9757d602fee9d29caf8180363d673706d56a6ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 419, + 500, + 433 + ], + "lines": [ + { + "bbox": [ + 91, + 419, + 500, + 433 + ], + "spans": [ + { + "bbox": [ + 91, + 419, + 500, + 433 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons with state of the art. Zoom-in for better comparison. Please see more in SM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 453, + 200, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 200, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 200, + 464 + ], + "type": "text", + "content": "a style reconstruction loss as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 475, + 287, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 475, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 125, + 475, + 287, + 487 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S R} = \\left\\| I _ {s s} - I _ {s} \\right\\|, \\tag {15}", + "image_path": "4192a42bb2ea4a8e6afc03ad4f1bcbfbfddaef2144b2d1b1001c58fbf1c4adea.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "inline_equation", + "content": "I_{ss}" + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "content": " is the stylized result given " + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "inline_equation", + "content": "I_s^c" + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "content": " as content. We optimize it separately before optimizing the style disentanglement loss " + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SD}" + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "content": ". The detailed fine-tuning procedure can be found in SM. The style reconstruction prior helps our model recover the style information more sufficiently. It also provides a good initialization for the optimization of " + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SD}" + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "content": ", which helps the latter give full play to its ability, thus producing higher-quality results (see later Sec. 5.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 603, + 173, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 173, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 173, + 616 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 179, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 179, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 179, + 635 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "We use ADM diffusion model [14] pre-trained on ImageNet [67] and adopt a fast sampling strategy [38]. Specifically, instead of sequentially conducting the diffusion processes until the last timestep " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " (e.g., 1000), we accelerate them by performing up to " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "T_{\\{\\cdot\\}} < T" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " (which is called return step), i.e., " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "T_{\\text{remov}} = 601" + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": " for style removal and " + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "inline_equation", + "content": "T_{\\text{trans}} =" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "text", + "content": "301 for style transfer. Moreover, as suggested by [38], we further accelerate the forward and reverse processes with fewer discretization steps, i.e., " + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "inline_equation", + "content": "(S_{for}, S_{rev}) = (40, 40)" + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "inline_equation", + "content": "S_{for}" + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "text", + "content": " for forward process and " + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "inline_equation", + "content": "S_{rev}" + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "text", + "content": " for reverse process) for style removal, and " + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "inline_equation", + "content": "(S_{for}, S_{rev}) = (40, 6)" + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "text", + "content": " for style transfer. When fine-tuning or inference, we can adjust " + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "inline_equation", + "content": "T_{remov}" + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "inline_equation", + "content": "T_{trans}" + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "text", + "content": " to flexibly control the degree of style removal and C-S disentanglement, as will be shown in Sec. 5.3. To fine-tune the model for a target style image, we randomly sample 50 images from ImageNet as the content images. We use Adam optimizer [40] with an initial learning rate of 4e-6 and increase it linearly by 1.2 per epoch. All models are fine-tuned with 5 epochs. See more details in SM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 620, + 465, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 465, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 465, + 633 + ], + "type": "text", + "content": "5.2. Comparisons with Prior Arts" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 640, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 545, + 688 + ], + "type": "text", + "content": "We compare our StyleDiffusion against ten state-of-the-art (SOTA) methods [18, 90, 12, 1, 53, 6, 13, 74, 32, 52]. For fair comparisons, all these methods are fine-tuned or trained on the target styles similar to our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Qualitative Comparisons. As can be observed in Fig. 3, due to the entangling of C-S representations, Gatys [18]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7682" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 541, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 541, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 541, + 159 + ], + "type": "table", + "html": "
OursGatysEFDMStyTr2ArtFlowAdaAttNIECASTMASTTPFRJohnsonLapStyle
SSIM ↑0.6720.3110.3160.5370.5010.5420.3650.3920.5360.6340.657
CLIP Score ↑0.7410.6770.6070.5310.5460.5770.6460.5900.6440.5370.595
Style Loss ↓0.8370.1110.1780.2160.2580.3100.2840.2290.9890.3640.274
UserStyle-43.1%41.2%39.3%36.4%37.2%33.8%39.1%14.5%42.8%47.3%
StudyOverall-26.0%38.1%44.0%34.2%43.9%32.7%32.2%22.6%43.4%46.2%
Training Time/h~0.4-~3~4~3~3~3~3~10~1~3
Testing Time/s5.61210.1650.0280.1680.2040.0760.0340.0660.3020.0150.008
", + "image_path": "c725bbdaaef55b1f5995fbbc3f387bbabdbb212ea06cabd629e8356fdc882b66.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 161, + 546, + 185 + ], + "lines": [ + { + "bbox": [ + 46, + 161, + 546, + 185 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 546, + 185 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with state of the art. The training/testing time is measured with an Nvidia Tesla A100 GPU, and the testing time is averaged on images of size " + }, + { + "bbox": [ + 46, + 161, + 546, + 185 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 161, + 546, + 185 + ], + "type": "text", + "content": " pixels. " + }, + { + "bbox": [ + 46, + 161, + 546, + 185 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 161, + 546, + 185 + ], + "type": "text", + "content": ": Higher is better. " + }, + { + "bbox": [ + 46, + 161, + 546, + 185 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 161, + 546, + 185 + ], + "type": "text", + "content": ": Lower is better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 45, + 204, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 204, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 45, + 204, + 289, + 491 + ], + "type": "text", + "content": "and EFDM [90] often produce unsatisfying results with distorted contents (e.g., rows 1-3) and messy textures (e.g., rows 4-8). StyTr² [12] and ArtFlow [1] improve the results by adopting more advanced networks [77, 41], but they may still produce inferior results with halo boundaries (e.g., rows 2-3) or dirty artifacts (e.g., rows 4-6). AdaAttN [53] performs per-point attentive normalization to preserve the content structures better, but the stylization effects may be degraded in some cases (e.g., rows 1, 2, 4, and 5). IECAST [6] utilizes contrastive learning and external learning for style transfer, so fine-tuning it on a single style image would result in degraded results. MAST [13] uses multi-adaptation networks to disentangle C-S. However, since it still relies on the C-S representations of [18], the results usually exhibit messy textures and conspicuous artifacts. TPFR [74] is a GAN-based framework that learns to disentangle C-S in latent space. As the results show, it cannot recover correct style details and often generates deviated stylizations, which signifies that it may not learn truly disentangled C-S representations [54]. Like our method, Johnson [32] and LapStyle [52] also train separate models for each style. However, due to the trade-off between C-S losses of [18], they may produce less-stylized results or introduce unnatural patterns (e.g., rows 1-6)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 494, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 615 + ], + "type": "text", + "content": "By contrast, our StyleDiffusion completely disentangles C-S based on diffusion models. Therefore, it can generate high-quality results with sufficient style details (e.g., rows 1-4) and well-preserved contents (e.g., rows 5-8). Compared with the previous methods that tend to produce mixed results of content and style, our approach can better consider the relationship between them. Thus, the stylizations are more natural and harmonious, especially for challenging styles such as cubism (e.g., row 2) and oil painting (e.g., rows 1, 3, 4, and 5)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Quantitative Comparisons. We also resort to quantitative metrics to better evaluate our method, as shown in Tab. 1. We collect 32 content and 12 style images to synthesize 384 stylized results and compute the average Structural Similarity Index (SSIM) [1] to assess the content similarity. To evaluate the style similarity, we calculate the CLIP image similarity score [62] and Style Loss [18, 27] between the style images and the corresponding stylized results. As" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 547, + 300 + ], + "type": "text", + "content": "shown in Tab. 1, our method obtains the highest SSIM and CLIP Score while the Style Loss is relatively higher than other methods. It is because these methods are directly trained to optimize Style Loss. Nevertheless, the Style Loss achieved by our method is still comparable and lower than the GAN-based TPFR [74]. Furthermore, it is noteworthy that our method can also incorporate Style Loss to enhance the performance in this regard (see later Sec. 5.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 300, + 548, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 548, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 548, + 479 + ], + "type": "text", + "content": "User Study. As style transfer is highly subjective and CLIP Score and Style Loss are biased to the training objective, we additionally resort to user study to evaluate the style similarity and overall stylization quality. We randomly select 50 C-S pairs for each user. Given each C-S pair, we show the stylized results generated by our method and a randomly selected SOTA method side by side in random order. The users are asked to choose (1) which result transfers the style patterns better and (2) which result has overall better stylization effects. We obtain 1000 votes for each question from 20 users and show the percentage of votes that existing methods are preferred to ours in Tab. 1. The lower numbers indicate our method is more preferred than the competitors. As the results show, our method is superior to others in both style consistency and overall quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 480, + 547, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 547, + 577 + ], + "type": "text", + "content": "Efficiency. As shown in the bottom two rows of Tab. 1, our approach requires less training time than others as it is fine-tuned on only a few (" + }, + { + "bbox": [ + 304, + 480, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\sim 50" + }, + { + "bbox": [ + 304, + 480, + 547, + 577 + ], + "type": "text", + "content": ") content images. When testing, our approach is faster than the optimization-based method Gatys [18], albeit slower than the remaining feedforward methods due to the utilization of diffusion models. We discuss it in later Sec. 6, and more timing and resource details can be found in SM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 586, + 400, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 400, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 400, + 600 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Control of C-S Disentanglement. A prominent advantage of our StyleDiffusion is that we can flexibly control the C-S disentanglement by adjusting the content extraction of the style removal module (Sec. 4.1). Fig. 4 demonstrates the continuous control achieved by adjusting the return step " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "T_{\\text{remov}}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " of the style removal module. As shown in the top row, with the increase of " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "T_{\\text{remov}}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", more style characteristics are dispelled, and the main content structures are retained. Correspondingly, when more style is removed in the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7683" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 177 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 177 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 177 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 177 + ], + "type": "image", + "image_path": "9e7d8272db81e048072bbe46ba61b8de4c2bd0ba6d9c5d1a519a02fb83c09a8a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 179, + 287, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 287, + 236 + ], + "type": "text", + "content": "Figure 4. Control of C-S disentanglement by adjusting the return step " + }, + { + "bbox": [ + 46, + 179, + 287, + 236 + ], + "type": "inline_equation", + "content": "T_{\\text{remov}}" + }, + { + "bbox": [ + 46, + 179, + 287, + 236 + ], + "type": "text", + "content": " of the style removal module. The top row shows the extracted contents of the style image. The bottom row shows the corresponding stylized results. * denotes our default setting. Zoom-in for better comparison. See SM for quantitative analyses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 257, + 286, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 286, + 341 + ], + "type": "text", + "content": "top row, it will be aptly transferred to the stylized results in the bottom row, e.g., the twisted brushstrokes and the star patterns. It validates that our method successfully separates style from content in a controllable manner and properly transfers it to other contents. Moreover, the flexible C-S disentanglement also makes our StyleDiffusion versatile for other tasks, such as photo-realistic style transfer (see SM)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "text", + "content": "Superiority of Diffusion-based Style Transfer. Although our style transfer module is not limited to the diffusion model, using it offers three main advantages: (1) Flexible C-S trade-off control. As shown in Fig. 5, we can flexibly control the C-S trade-off at both the training stage (top row) and the testing stage (bottom row) by adjusting the return step " + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{trans}" + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "text", + "content": " of the diffusion model. With the increase of " + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{trans}" + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "text", + "content": ", more style characteristics are transferred, yet the content structures may be ruined (e.g., the last column). When proper " + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{trans}" + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "text", + "content": " is adopted, e.g., " + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{trans} = 301" + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "text", + "content": ", the sweet spot can be well achieved. Interestingly, as shown in the last two columns of the bottom row, though the model is trained on " + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{trans} = 301" + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "text", + "content": ", we can extrapolate the style by using larger " + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{trans}" + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "text", + "content": " (e.g., 401) at the testing stage (but the results may be degraded when using too large " + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "inline_equation", + "content": "T_{trans}" + }, + { + "bbox": [ + 46, + 342, + 287, + 713 + ], + "type": "text", + "content": ", e.g., 601). It provides a very flexible way for users to adjust the results according to their preferences. This property, however, cannot be simply achieved by using other models, e.g., the widely used AEs [27, 50], since our framework does not involve any feature transforms [27, 50] or C-S losses trade-off [3]. (2) Higher-quality stylizations. Owing to the strong generative ability of the diffusion model, it can achieve higher-quality stylizations than other models. For comparison, we use the pre-trained VGG-AE [27, 46] as the style transfer module and fine-tune its decoder network for each style. As shown in column (b) of Fig. 6, though the results are still acceptable, they may produce distorted contents and inferior textures, clearly worse than the results generated by the diffusion model in column (a). This is also validated by the bottom quantitative scores. It signifies that the diffusion model can better learn the disentangled" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 70, + 545, + 176 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 545, + 176 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 545, + 176 + ], + "type": "image", + "image_path": "d6397c032e8816519cced80362c26afb7c20fc6b9c75b61afce656a5be8fa0cf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "text", + "content": "Figure 5. Control of C-S trade-off by adjusting the return step " + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "inline_equation", + "content": "T_{trans}" + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "text", + "content": " of the style transfer module. The top row shows adjusting " + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "inline_equation", + "content": "T_{trans}" + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "text", + "content": " at the training stage while fixing " + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "inline_equation", + "content": "T_{trans} = 301" + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "text", + "content": " at the testing stage. The bottom row shows adjusting " + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "inline_equation", + "content": "T_{trans}" + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "text", + "content": " at the testing stage while fixing " + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "inline_equation", + "content": "T_{trans} = 301" + }, + { + "bbox": [ + 305, + 178, + 546, + 256 + ], + "type": "text", + "content": " at the training stage. * denotes our default setting. Zoom-in for better comparison. See SM for quantitative analyses." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 268, + 425, + 408 + ], + "blocks": [ + { + "bbox": [ + 308, + 268, + 425, + 408 + ], + "lines": [ + { + "bbox": [ + 308, + 268, + 425, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 268, + 425, + 408 + ], + "type": "image", + "image_path": "de144b610ab417f1fdbba4a40504f3455bc799d6718e8da644e93de0354cf834.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 416, + 529, + 427 + ], + "lines": [ + { + "bbox": [ + 322, + 416, + 529, + 427 + ], + "spans": [ + { + "bbox": [ + 322, + 416, + 529, + 427 + ], + "type": "text", + "content": "Figure 6. Diffusion-based vs. AE-based style transfer." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 428, + 268, + 545, + 408 + ], + "blocks": [ + { + "bbox": [ + 428, + 268, + 545, + 408 + ], + "lines": [ + { + "bbox": [ + 428, + 268, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 428, + 268, + 545, + 408 + ], + "type": "image", + "image_path": "7ed65172626828e48354751e2d7c738a8585ea16633fc9543a789e0057e9741b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": "content and style characteristics in our framework, helping produce better style transfer results. (3) Diversified style transfer. As mentioned in Sec. 4.2, during inference, we can directly adopt the stochastic DDPM [25] forward process (Eq. (2)) to obtain diverse results (see SM). The diverse results can give users endless choices to obtain more satisfactory results. However, using other models like AEs in our framework cannot easily achieve it [81]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "Loss Analyses. To verify the effectiveness of each loss term used for fine-tuning our StyleDiffusion, we present ablation study results in Fig. 7 (a-d). (1) Using L1 loss " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SD}^{L1}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " successfully transfers the cubism style like the blocky patterns in the top row, but the colors stray from the style images, especially in the bottom row. It is consistent with our earlier analyses in Sec. 4.3 that the L1 loss is prone to produce implausible results outside the style domain. (2) Adding direction loss " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SD}^{dir}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " helps pull the results closer to the style domain. The textures are enhanced in the top row, and the colors are more plausible in the top and bottom rows. (3) By further coordinating with the style reconstruction prior " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SR}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", the stylization effects are significantly elevated where the style information is recovered more sup" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7684" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 171, + 217 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 171, + 217 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 171, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 171, + 217 + ], + "type": "image", + "image_path": "f52741d8174e9d4f012e5b04328b8f0f80d528a3b5b9ee92eaf7d4f4d9f0cfaa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 222, + 483, + 233 + ], + "lines": [ + { + "bbox": [ + 108, + 222, + 483, + 233 + ], + "spans": [ + { + "bbox": [ + 108, + 222, + 483, + 233 + ], + "type": "text", + "content": "Figure 7. Ablation study on loss functions. * denotes our full model. Zoom-in for better comparison." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 237, + 217 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 237, + 217 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 237, + 217 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 237, + 217 + ], + "type": "image", + "image_path": "290b33325c8f5254e246f04302a417f9b7d45eac542a7530f3b2edcda07bbd0e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 238, + 71, + 298, + 217 + ], + "blocks": [ + { + "bbox": [ + 238, + 71, + 298, + 217 + ], + "lines": [ + { + "bbox": [ + 238, + 71, + 298, + 217 + ], + "spans": [ + { + "bbox": [ + 238, + 71, + 298, + 217 + ], + "type": "image", + "image_path": "3aeab92b7b3c4a9375ce0fade7133235f5164b55e04c3c5ba1afa693f81bc3a3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 359, + 217 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 359, + 217 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 359, + 217 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 359, + 217 + ], + "type": "image", + "image_path": "3c44f8bc9590fa95abe69a8bcf01cdf08115889609fc9967eafe80b815a43aae.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 360, + 72, + 420, + 217 + ], + "blocks": [ + { + "bbox": [ + 360, + 72, + 420, + 217 + ], + "lines": [ + { + "bbox": [ + 360, + 72, + 420, + 217 + ], + "spans": [ + { + "bbox": [ + 360, + 72, + 420, + 217 + ], + "type": "image", + "image_path": "74f0f139e37338b03d1763dfcd660f2d625447ca0b0f9964d858e9f0b402af57.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 425, + 71, + 485, + 217 + ], + "blocks": [ + { + "bbox": [ + 425, + 71, + 485, + 217 + ], + "lines": [ + { + "bbox": [ + 425, + 71, + 485, + 217 + ], + "spans": [ + { + "bbox": [ + 425, + 71, + 485, + 217 + ], + "type": "image", + "image_path": "c1e0a8238a89ea23aebe0dac3a707991c8dcbbac643edae9c529e593c9041d60.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 486, + 72, + 546, + 217 + ], + "blocks": [ + { + "bbox": [ + 486, + 72, + 546, + 217 + ], + "lines": [ + { + "bbox": [ + 486, + 72, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 486, + 72, + 546, + 217 + ], + "type": "image", + "image_path": "76cf3ba0e08d752ca646da0a3b51039468eb169f5119c2d672e4b58bce884899.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 253, + 286, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 286, + 324 + ], + "type": "text", + "content": "ficiently. It may be because it provides a good initialization for the optimization of " + }, + { + "bbox": [ + 46, + 253, + 286, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SD}^{L1}" + }, + { + "bbox": [ + 46, + 253, + 286, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 253, + 286, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SD}^{dir}" + }, + { + "bbox": [ + 46, + 253, + 286, + 324 + ], + "type": "text", + "content": ", which helps them give full play to their abilities. As verified in Fig. 7 (d), using the style reconstruction alone cannot learn meaningful style patterns except for basic colors. All the above analyses are also supported by the bottom quantitative scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 325, + 286, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 286, + 647 + ], + "type": "text", + "content": "Comparison with Gram Loss. To further verify the superiority of our proposed losses, we replace them with the widely used Gram Loss [18, 27] in Fig. 7 (e-f). As can be observed, Gram Loss destroys the content structures severely, e.g., the zebra head in the top row and the enlarged area in the bottom row. This is because it does not disentangle C-S and only matches the global statistics without considering the relationship between C-S. In contrast, our losses focus on learning the disentangled style information apart from the content, which is induced by the difference between content and its stylized result. Therefore, they can better understand the relationship between C-S, achieving more satisfactory results with fine style details and better-preserved contents, as validated by Fig. 7 (c) and the bottom quantitative scores. Furthermore, we also conduct comparisons between our proposed losses and Gram Loss [18, 27] on the AE baseline [27, 46] to eliminate the impact of diffusion models. As shown in Fig. 8 (a-b), our losses can achieve more satisfactory results than Gram Loss, which is consistent with the results in Fig. 7. Moreover, as shown in Fig. 8 (c), they can also be combined with Gram Loss to improve the performance on the Style Loss metric. However, it may affect the full disentanglement of C-S in our framework, which strays from our target and decreases the content preservation (see SSIM score in Fig. 8 (c)). Therefore, we do not incorporate Gram Loss in our framework by default." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 658, + 198, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 198, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 198, + 670 + ], + "type": "text", + "content": "6. Conclusion and Limitation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "In this work, we present a new framework for more interpretable and controllable C-S disentanglement and style transfer. Our framework, termed StyleDiffusion, leverages" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 251, + 403, + 390 + ], + "blocks": [ + { + "bbox": [ + 307, + 251, + 403, + 390 + ], + "lines": [ + { + "bbox": [ + 307, + 251, + 403, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 403, + 390 + ], + "type": "image", + "image_path": "39eefed754ea71adc992013a9465c3efc10cfaa82c74b273da2b5bcdadf0ab85.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 394, + 544, + 405 + ], + "lines": [ + { + "bbox": [ + 306, + 394, + 544, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 544, + 405 + ], + "type": "text", + "content": "Figure 8. More loss function ablation study on the AE baseline." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 403, + 251, + 451, + 390 + ], + "blocks": [ + { + "bbox": [ + 403, + 251, + 451, + 390 + ], + "lines": [ + { + "bbox": [ + 403, + 251, + 451, + 390 + ], + "spans": [ + { + "bbox": [ + 403, + 251, + 451, + 390 + ], + "type": "image", + "image_path": "0a6ea978a6ec74a27740154edef2bed2cabdb04889ae7ec647d485548beaf9b1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 452, + 251, + 498, + 389 + ], + "blocks": [ + { + "bbox": [ + 452, + 251, + 498, + 389 + ], + "lines": [ + { + "bbox": [ + 452, + 251, + 498, + 389 + ], + "spans": [ + { + "bbox": [ + 452, + 251, + 498, + 389 + ], + "type": "image", + "image_path": "de7a1aac3d645bb0bba6fdff43f6a046fc2dad5446037e41a1ece42899b0a3f5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 499, + 251, + 546, + 389 + ], + "blocks": [ + { + "bbox": [ + 499, + 251, + 546, + 389 + ], + "lines": [ + { + "bbox": [ + 499, + 251, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 499, + 251, + 546, + 389 + ], + "type": "image", + "image_path": "91ca077552027d6c6f07d7947471b6b15a217c01a601c2cffbaa33cca39b735b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 427, + 545, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 522 + ], + "type": "text", + "content": "diffusion models to explicitly extract the content information and implicitly learn the complementary style information. A novel CLIP-based style disentanglement loss coordinated with a style reconstruction prior is also introduced to encourage the disentanglement and style transfer. Our method yields very encouraging stylizations, especially for challenging styles, and the experimental results verify its effectiveness and superiority against state of the art." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 524, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 524, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 545, + 643 + ], + "type": "text", + "content": "Currently, the framework still suffers from several limitations: (1) The model needs to be fine-tuned for each style, and arbitrary style transfer is left to our future work. (2) The efficiency is not fast enough due to the use of diffusion models. Further research in accelerating diffusion sampling would be helpful. (3) There are some failure cases analyzed in SM, which may help inspire future improvements. Moreover, our framework may also be applied to other image translation [28] or manipulation [60] tasks, and we would like to explore them in our future work." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 655, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 655, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 655, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgments. We thank Zeyi Huang and Xiaoting Zhang for their insightful discussions and suggestions. This work was supported in part by the National Program of China (2020YFC1523201, 62172365, 19ZDA197), Zhejiang Elite Program (2022C01222), and Key Technologies and Product Research and Development Projects for Cultural Relics Protection and Trading Circulation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7685" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 862-871, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "type": "text", + "content": "[2] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 235 + ], + "type": "text", + "content": "[3] Mohammad Babaeizadeh and Golnaz Ghiasi. Adjustable real-time style transfer. In International Conference on Learning Representations (ICLR), 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Robin Rombach, Kaan Oktay, Jonas Müller, and Björn Ommer. Retrieval-augmented diffusion models. Advances in Neural Information Processing Systems, 35:15309-15324, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 288, + 334 + ], + "type": "text", + "content": "[5] Dongdong Chen, Lu Yuan, Jing Liao, Nenghai Yu, and Gang Hua. Stylebank: An explicit representation for neural image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1897-1906, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 390 + ], + "type": "text", + "content": "[6] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems (NeurIPS), 34:26561-26573, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 456 + ], + "type": "text", + "content": "[7] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dualist: Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 872-881, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 523 + ], + "type": "text", + "content": "[8] Haibo Chen, Lei Zhao, Huiming Zhang, Zhizhong Wang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Diverse image style transfer via invertible cross-space mapping. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14860-14869. IEEE Computer Society, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 288, + 579 + ], + "type": "text", + "content": "[9] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in Neural Information Processing Systems (NeurIPS), 29, 2016. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 635 + ], + "type": "text", + "content": "[10] Jiaxin Cheng, Ayush Jaiswal, Yue Wu, Pradeep Natarajan, and Prem Natarajan. Style-aware normalized loss for improving arbitrary style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 134-143, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "text", + "content": "[11] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14347-14356. IEEE, 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[12] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Im" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "text", + "content": "age style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11326-11336, 2022. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "type": "text", + "content": "[13] Yingying Deng, Fan Tang, Weiming Dong, Wen Sun, Feiyue Huang, and Changsheng Xu. Arbitrary style transfer via multi-adaptation network. In Proceedings of the 28th ACM International Conference on Multimedia (ACM MM), pages 2719-2727, 2020. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "text", + "content": "[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 34:8780-8794, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 547, + 262 + ], + "type": "text", + "content": "[15] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 263, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 547, + 319 + ], + "type": "text", + "content": "[16] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 579-587, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 320, + 547, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 547, + 365 + ], + "type": "text", + "content": "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 547, + 420 + ], + "type": "text", + "content": "[18] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 1, 2, 5, 6, 7, 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 421, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 547, + 442 + ], + "type": "text", + "content": "[19] Rafael C Gonzalez. Digital image processing. Pearson education india, 2009. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 444, + 547, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 547, + 497 + ], + "type": "text", + "content": "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 500, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 547, + 544 + ], + "type": "text", + "content": "[21] Daniel J Graham, James M Hughes, Helmut Leder, and Daniel N Rockmore. Statistics, vision, and the analysis of artistic style. Wiley Interdisciplinary Reviews: Computational Statistics, 4(2):115-123, 2012. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 545, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 547, + 589 + ], + "type": "text", + "content": "[22] Shuyang Gu, Congliang Chen, Jing Liao, and Lu Yuan. Arbitrary style transfer with deep feature reshuffle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8222-8231, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 645 + ], + "type": "text", + "content": "[23] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "type": "text", + "content": "[24] Irina Higgins, David Amos, David Pfau, Sebastien Racaniere, Loic Matthew, Danilo Rezende, and Alexander Lerchner. Towards a definition of disentangled representations. arXiv preprint arXiv:1812.02230, 2018. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 547, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 715 + ], + "type": "text", + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7686" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "Processing Systems (NeurIPS), 33:6840-6851, 2020. 2, 3, 4, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[26] Kibeom Hong, Seogkyu Jeon, Huan Yang, Jianlong Fu, and Hyeran Byun. Domain-aware universal style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 195 + ], + "type": "text", + "content": "[27] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1501-1510, 2017. 1, 2, 4, 7, 8, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[28] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 172-189, 2018. 2, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "text", + "content": "[29] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14861-14869, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "type": "text", + "content": "[30] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4369-4376, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "text", + "content": "[31] Yongcheng Jing, Yezhou Yang, Zunlei Feng, Jingwen Ye, Yizhou Yu, and Mingli Song. Neural style transfer: A review. IEEE Transactions on Visualization and Computer Graphics (TVCG), 26(11):3365-3385, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 443 + ], + "type": "text", + "content": "[32] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pages 694–711. Springer, 2016. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "text", + "content": "[33] Henry Kang, Seungyong Lee, and Charles K Chui. Coherent line drawing. In Proceedings of the 5th international symposium on Non-photorealistic animation and rendering, pages 43–50, 2007. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 521 + ], + "type": "text", + "content": "[34] Theofanis Karaletsos, Serge Belongie, and Gunnar Ratsch. Bayesian representation learning with oracle constraints. arXiv preprint arXiv:1506.05011, 2015. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 524, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 287, + 566 + ], + "type": "text", + "content": "[35] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, and Holger Winnemoeller. Recognizing image style. arXiv preprint arXiv:1311.3715, 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "text", + "content": "[36] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[37] Hadi Kazemi, Seyed Mehdi Iranmanesh, and Nasser Nasrabadi. Style and content disentanglement in generative adversarial networks. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 848-856. IEEE, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[38] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 2426-2435, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[39] Hyunjik Kim and Andriy Mnih. Disentangling by factorising. In International Conference on Machine Learning (ICML), pages 2649-2658. PMLR, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[40] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 164, + 545, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 197 + ], + "type": "text", + "content": "[41] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 251 + ], + "type": "text", + "content": "[42] Dmytro Kotovenko, Artsiom Sanakoyeu, Sabine Lang, and Bjorn Ommer. Content and style disentanglement for artistic style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4422-4431, 2019. 1, 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[43] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Bjorn Ommer. Rethinking style transfer: From pixels to parameterized brushstrokes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12196-12205, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "text", + "content": "[44] Tejas D Kulkarni, William F Whitney, Pushmeet Kohli, and Josh Tenenbaum. Deep convolutional inverse graphics network. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 355, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 408 + ], + "type": "text", + "content": "[45] Gihyun Kwon and Jong Chul Ye. Diagonal attention and style-based gan for content-style disentanglement in image generation and translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13980-13989, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "text", + "content": "[46] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18062-18071, 2022. 5, 8, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 456, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 477 + ], + "type": "text", + "content": "[47] Berel Lang. The concept of style. Cornell University Press, 1987. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 479, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 533 + ], + "type": "text", + "content": "[48] Hsin-Ying Lee, Hung-Yu Tseng, Jia-Bin Huang, Maneesh Singh, and Ming-Hsuan Yang. Diverse image-to-image translation via disentangled representations. In Proceedings of the European Conference on Computer Vision (ECCV), pages 35-51, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 535, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 588 + ], + "type": "text", + "content": "[49] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 634 + ], + "type": "text", + "content": "[50] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. In Advances in Neural Information Processing Systems (NeurIPS), pages 386-396, 2017. 1, 2, 4, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[51] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[52] Tianwei Lin, Zhuoqi Ma, Fu Li, Dongliang He, Xin Li, Errui Ding, Nannan Wang, Jie Li, and Xinbo Gao. Drafting and" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7687" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 117 + ], + "type": "text", + "content": "revision: Laplacian pyramid network for fast high-quality artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5141-5150, 2021. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 184 + ], + "type": "text", + "content": "[53] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6649-6658, 2021. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 251 + ], + "type": "text", + "content": "[54] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In International Conference on Machine Learning (ICML), pages 4114-4124. PMLR, 2019. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 252, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 252, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 288, + 306 + ], + "type": "text", + "content": "[55] Haofei Lu and Zhizhong Wang. Universal video style transfer via crystallization, separation, and blending. In Proceedings of the International Joint Conferences on Artificial Intelligence Organization (IJCAI), Vienna, Austria, pages 23-29, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "type": "text", + "content": "[56] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "type": "text", + "content": "[57] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), pages 8162-8171. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 411, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 476 + ], + "type": "text", + "content": "[58] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), pages 16784-16804. PMLR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 521 + ], + "type": "text", + "content": "[59] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5880–5888, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 523, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 577 + ], + "type": "text", + "content": "[60] Taesung Park, Jun-Yan Zhu, Oliver Wang, Jingwan Lu, Eli Shechtman, Alexei Efros, and Richard Zhang. Swapping autoencoder for deep image manipulation. Advances in Neural Information Processing Systems (NeurIPS), 33:7198-7211, 2020. 2, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 579, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 601 + ], + "type": "text", + "content": "[61] DM Parker and Jan B Deregowski. Perception and artistic style. Elsevier, 1991. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 668 + ], + "type": "text", + "content": "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[63] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[64] Xuanchi Ren, Tao Yang, Yuwang Wang, and Wenjun Zeng. Rethinking content and style: Exploring bias for unsupervised disentanglement. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1823-1832. IEEE, 2021. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 131, + 547, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 547, + 186 + ], + "type": "text", + "content": "[65] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 187, + 546, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 187, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 187, + 546, + 243 + ], + "type": "text", + "content": "[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-assisted Intervention, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 245, + 547, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 547, + 300 + ], + "type": "text", + "content": "[67] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 301, + 546, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 301, + 546, + 368 + ], + "spans": [ + { + "bbox": [ + 308, + 301, + 546, + 368 + ], + "type": "text", + "content": "[68] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 370, + 546, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 370, + 546, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 370, + 546, + 414 + ], + "type": "text", + "content": "[69] Artsiom Sanakoyeu, Dmytro Kotovenko, Sabine Lang, and Bjorn Ommer. A style-aware content loss for real-time hd style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), pages 698-714, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 416, + 546, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 546, + 471 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 546, + 471 + ], + "type": "text", + "content": "[70] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8242-8250, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 474, + 546, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 474, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 474, + 546, + 506 + ], + "type": "text", + "content": "[71] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 509, + 546, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 509, + 546, + 563 + ], + "spans": [ + { + "bbox": [ + 308, + 509, + 546, + 563 + ], + "type": "text", + "content": "[72] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning (ICML), pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 566, + 546, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 566, + 546, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 566, + 546, + 599 + ], + "type": "text", + "content": "[73] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2020. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 601, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 547, + 656 + ], + "type": "text", + "content": "[74] Jan Svoboda, Asha Anoosheh, Christian Osendorfer, and Jonathan Masci. Two-stage peer-regularized feature recombination for arbitrary image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13816–13825, 2020. 1, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "type": "text", + "content": "[75] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6924-6932, 2017. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7688" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 286, + 106 + ], + "type": "text", + "content": "[76] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 287, + 150 + ], + "type": "text", + "content": "[77] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 150, + 287, + 194 + ], + "type": "text", + "content": "[78] Pei Wang, Yijun Li, and Nuno Vasconcelos. Rethinking and improving the robustness of image style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 124-133, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 194, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 194, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 287, + 250 + ], + "type": "text", + "content": "[79] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: Towards aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia (ACM MM), pages 1095-1106, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 250, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 287, + 304 + ], + "type": "text", + "content": "[80] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2624-2632, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 304, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 304, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 49, + 304, + 287, + 369 + ], + "type": "text", + "content": "[81] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7789-7798, 2020. 2, 4, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 371, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 371, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 49, + 371, + 287, + 415 + ], + "type": "text", + "content": "[82] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Evaluate and improve the quality of neural style transfer. Computer Vision and Image Understanding (CVIU), 207:103203, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 415, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 415, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 49, + 415, + 287, + 470 + ], + "type": "text", + "content": "[83] Zhizhong Wang, Lei Zhao, Haibo Chen, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. DivSwapper: Towards diversified patch-based arbitrary style transfer. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), pages 4980–4987, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 49, + 471, + 287, + 514 + ], + "type": "text", + "content": "[84] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Glstylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8):575-586, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 514, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 514, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 49, + 514, + 287, + 569 + ], + "type": "text", + "content": "[85] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2742-2750, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 570, + 287, + 613 + ], + "type": "text", + "content": "[86] Wayne Wu, Kaidi Cao, Cheng Li, Chen Qian, and Chen Change Loy. Disentangling content and style via unsupervised geometry distillation. arXiv preprint arXiv:1905.04538, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 287, + 668 + ], + "type": "text", + "content": "[87] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14618-14627, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 669, + 287, + 712 + ], + "type": "text", + "content": "[88] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 363 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 308, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 308, + 73, + 545, + 115 + ], + "type": "text", + "content": "[89] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1395-1403, 2015. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 173 + ], + "type": "text", + "content": "[90] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8035-8045, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 173, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 228 + ], + "type": "text", + "content": "[91] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. Domain enhanced arbitrary image style transfer via contrastive learning. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 229, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 545, + 262 + ], + "type": "text", + "content": "[92] Yu Zhang, Peter Tino, Ales Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 262, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 307 + ], + "type": "text", + "content": "[93] Yexun Zhang, Ya Zhang, and Wenbin Cai. Separating style and content for generalized style transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8447-8455, 2018. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 308, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 363 + ], + "type": "text", + "content": "[94] Zhiwen Zuo, Lei Zhao, Shuobin Lian, Haibo Chen, Zhizhong Wang, Ailin Li, Wei Xing, and Dongming Lu. Style fader generative adversarial networks for style degree controllable artistic style transfer. In Proc. Int. Joint Conf. on Artif. Intell.(IJCAI), pages 5002-5009, 2022. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7689" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/b30ac484-13c8-4c42-8711-cf47c43f4550_content_list.json b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/b30ac484-13c8-4c42-8711-cf47c43f4550_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..35f85db699105d7a8cd6faa04ba421e7006ef180 --- /dev/null +++ b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/b30ac484-13c8-4c42-8711-cf47c43f4550_content_list.json @@ -0,0 +1,1457 @@ +[ + { + "type": "text", + "text": "StyleDomain: Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation", + "text_level": 1, + "bbox": [ + 104, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aibek Alanov $^{2,1,*}$ , Vadim Titov $^{2,*}$ , Maksim Nakhodnov $^{2,3*}$ , Dmitry Vetrov $^{1,2}$", + "bbox": [ + 187, + 202, + 781, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ HSE University $^{2}$ AIRI $^{3}$ Lomonosov Moscow State University", + "bbox": [ + 284, + 223, + 684, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "aalanov@hse.ru, titov@2a2i.org, nakhodnov@2a2i.org, dvetrov@hse.ru", + "bbox": [ + 191, + 241, + 777, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/703fc3be0760b37e479d897a4a507ca0ff9c997fe1607ba5cbf88db2032d4c9b.jpg", + "image_caption": [ + "Figure 1: Smooth cross-domain image morphing. Morphing in the space of generator weights could be successfully combined with morphing using StyleDomain directions, i.e. directions in the StyleSpace that can adapt the generator to new domains (see Section 4)." + ], + "image_footnote": [], + "bbox": [ + 76, + 282, + 483, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/295ca5be4e5e5a987fdcfb3a2a6f80cd846f89d49c53320019adac59146680a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 282, + 890, + 443 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 521, + 313, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain adaptation of GANs is a problem of fine-tuning GAN models pretrained on a large dataset (e.g. StyleGAN) to a specific domain with few samples (e.g. painting faces, sketches, etc.). While there are many methods that tackle this problem in different ways, there are still many important questions that remain unanswered. In this paper, we provide a systematic and in-depth analysis of the domain adaptation problem of GANs, focusing on the StyleGAN model. We perform a detailed exploration of the most important parts of StyleGAN that are responsible for adapting the generator to a new domain depending on the similarity between the source and target domains. As a result of this study, we propose new efficient and lightweight parameterizations of StyleGAN for domain adaptation. Particularly, we show that there exist directions in StyleSpace (StyleDomain directions) that are sufficient for adapting to similar domains. For dissimilar domains, we propose Affine+ and AffineLight+ parameterizations that allows us to outperform existing baselines in few-shot adaptation while having significantly less training parameters. Finally, we examine StyleDomain directions and discover their many sur", + "bbox": [ + 73, + 555, + 472, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prising properties that we apply for domain mixing and cross-domain image morphing. Source code can be found at https://github.com/AIRI-Institute/StyleDomain.", + "bbox": [ + 498, + 523, + 893, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 601, + 632, + 617 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years GANs [12, 20, 21, 19, 5] have shown impressive results in image synthesis and offered many ways to control the generated data. In particular, the state-of-the-art StyleGAN models [20, 21, 19] have many practical applications such as image enhancement [46, 24, 6, 41], image editing [17, 35, 14, 37, 1, 43, 29], image-to-image translation [31, 16, 36, 11] thanks to their high-quality image generation and their latent representation that has rich semantics and disentangled controls for localized meaningful image manipulations. However, it comes at a price, as the training of StyleGAN requires a large, high-quality dataset that significantly limits its applicability because many real-world domains are represented by few images. The standard approach to deal with this problem is transfer learning, i.e. fine-tuning the model pretrained on the source domain $A$ to the target domain $B$ .", + "bbox": [ + 496, + 628, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are many domain adaptation methods for StyleGAN [18, 39, 50, 52, 23, 45, 28, 31, 4, 11, 55, 44] that", + "bbox": [ + 500, + 871, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 888, + 204, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2184", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tackle this problem in different ways depending on the number of available images (e.g., one-shot/few-shot) from the target domain $B$ and the similarity between the source $A$ and target $B$ domains (e.g., faces $\\rightarrow$ sketches, artistic portraits, or faces $\\rightarrow$ dogs, cats). Most of these works implicitly assume that StyleGAN can be adapted to a new domain only if we fine-tune almost all its weights, even for similar domains. However, this common wisdom is poorly investigated and verified and there is a lack of analysis of which parts of StyleGAN are important depending on different data regimes and the similarity between domains.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to provide a systematic and comprehensive analysis of this question. Our investigation of the properties of the aligned StyleGAN models consists of several parts. First, in Section 3, we identify what parts of the StyleGAN are sufficient for its adaptation depending on the similarity between the source $A$ and target $B$ domains. We discover that fine-tuning the whole synthesis convolutional network is not always necessary. In the case of similar $A$ and $B$ domains, the affine layers are sufficient for the adaptation. For more distant domains, we should optimize more parameters, however not the whole network. It suggests investigating new more efficient and lightweight parameterizations of StyleGAN to utilize them for domain adaptation.", + "bbox": [ + 75, + 260, + 472, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the second part of our analysis, we propose two new parameterizations of StyleGAN. For similar domains, we consider the latent space that is formed by the output of affine layers, i.e. StyleSpace [43]. We show that we can directly optimize directions in this space that can adapt to similar target domains with the same quality as fine-tuning all weights of StyleGAN (we call such directions as Style-Domain directions). Further, we explore that we can zero out $80\\%$ of StyleDomain direction coordinates without a quality degradation that gives even more lightweight parameterization (StyleSpaceSparse). For more distant domains, we propose a new parameterization Affine+ that consists of affine layers and only one convolutional block from the synthesis network. It reduces the number of trainable parameters by 6 times and achieves the same quality. Then, we further improve Affine+ parameterization by utilizing low-rank decomposition for weights of affine layers and obtain AffineLight+ parameterization. It allows us to optimize by two orders less parameters compared to training the whole StyleGAN. These parameterizations show the state-of-the-art performance for few-shot adaptation for dissimilar domains outperforming more complicated and expensive baselines.", + "bbox": [ + 75, + 474, + 472, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, in Section 4, we inspect StyleDomain directions and discover their surprising properties. The first one is mixability, i.e. we can sum up these directions to obtain a new mixed domain (e.g., see Figure 6 as a mix of the Joker style, Pixar style and the style from the image). The", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "second impressive property is transferability, i.e. the same StyleDomain directions can be applied to StyleGAN models that were fine-tuned to other domains (e.g., see Figure 5 where we apply directions found for faces to dogs, cats and churches). We apply these findings to standard computer vision tasks such as image-to-image translation and cross-domain morphing.", + "bbox": [ + 496, + 90, + 893, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 210, + 640, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Latent Spaces of StyleGAN. After recent remarkable success of GANs [12, 20, 21, 19, 5] in image synthesis, many works appeared that explore their latent representation for controllable image manipulation. In particular, the latent space of StyleGAN [20, 21, 19] has attracted considerable attention. It consists of three levels: (i) the first latent space, $\\mathcal{Z}$ , is raw random noise (typically Gaussian); (ii) the intermediate latent spaces $\\mathcal{W}, \\mathcal{W} + [1]$ are formed by the output of the mapping network; (iii) the last level is StyleSpace, $\\mathcal{S}$ , [43] that is spanned by the channel-wise style parameters after affine layers. It has been shown that these latent spaces have rich semantics [20, 21], and especially the StyleSpace that demonstrates the most disentangled and localized semantical directions [43]. In recent years many works have proposed to utilize such appealing properties of the StyleGAN latent spaces for image editing tasks [17, 35, 14, 37, 1, 43, 29]. To apply these methods for real images it is necessary to inverse them into one of the latent space of StyleGAN which is another task that draws significant attention [1, 21, 13, 30, 33, 38, 54, 53]. We should note that all mentioned methods for image manipulation by controlling StyleGAN latent space allow only in-domain editing. In this paper, we show that in StyleSpace there exist such directions that can change the domain of images.", + "bbox": [ + 496, + 234, + 893, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain Adaptation of StyleGAN. Recent years the problem of fine-tuning StyleGAN has generated a great deal of interest as it allows training the state-of-the-art generative model for a domain with few samples. There have appeared many works that tackle this problem in different ways depending on how similar the target domain is to the source one. Roughly, these methods can be divided into two groups. The first one deals with the case when the target and source domains are dissimilar (e.g. faces $\\rightarrow$ cats, churches, etc.). In contrast, the second group considers the setting of similar domains (e.g. faces $\\rightarrow$ stylized faces, painting faces, sketches, etc.). Methods from the first group typically require hundreds or thousands of samples from a new domain to adapt faithfully and they leverage data augmentations [18, 39, 50, 52], or freeze some layers of the discriminator to prevent overfitting [25], or train the discriminator with auxiliary losses to match the data more accurately [23, 45].", + "bbox": [ + 496, + 613, + 895, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the setting of the second group it is sufficient to have", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2185", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b94d569f9f4895fc11e5443215419a22e84d554bb362d8d6924632304271b512.jpg", + "image_caption": [ + "Figure 2: StyleGAN2 architecture. We introduce new latent space $S +$ for the for domain adaptation that combines StyleSpace and weight offsets for one block from the synthesis network." + ], + "image_footnote": [ + "Synthesis Network", + "Affine Layers", + "Mapping Network", + "Affine + ConvBlock", + "Weight Offsets" + ], + "bbox": [ + 76, + 89, + 669, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "only several samples (up to dozens) from the target domain for the successful adaptation and such approaches utilize another techniques. In particular, they introduce additional regularization terms [40, 22], preserve pairwise distances between instances in the source domain via cross-domain consistency loss [28], mix weights of the fine-tuned and the base generators [31], use an auxiliary small network to enhance the training [42]. More advanced methods [11, 55, 4] utilize the pretrained CLIP model [32] as the vision-language supervision for the text-based adaptation [11, 4] or the one-shot image-based adaptation [7, 55, 9, 49, 48, 4].", + "bbox": [ + 75, + 393, + 468, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent work [51] analyzes the performance of methods from the second group for dissimilar domains in the low data regime. It was shown that those methods have a significant quality degradation in the few-shot regime. However, it can be partially mitigated by constraining model parameters. There are also approaches that introduce more lightweight parameterizations for domain adaptation [27, 34, 7, 4], however they work only for similar domains. In our paper, we propose efficient and highly lightweight parameterizations for both similar and dissimilar domains and show that they can achieve results on par with the state-of-the-art methods that optimize all weights of StyleGAN.", + "bbox": [ + 75, + 564, + 468, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are few works that analyze the domain adaptation process of StyleGAN thoroughly. The paper [44] is the first attempt to perform such an in-depth study. In particular, they explore which parts of StyleGAN are mostly changed during fine-tuning and the transferability of the latent semantics after domain adaptation. However, this work does not analyze which parts of StyleGAN are sufficient for adapting depending on the similarity between source and target domains. In our paper, we provide a more systematic and comprehensive analysis that completes and improves", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the results from [44].", + "bbox": [ + 500, + 393, + 640, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Importance of Each Part of the StyleGAN", + "text_level": 1, + "bbox": [ + 500, + 420, + 872, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, our goal is to analyze what parts of StyleGAN are important for domain adaptation. Similarly to previous works, we specifically focus on the state-of-the-art GAN architecture, StyleGAN2 [21]. For the source domain, we consider FFHQ [20] as it is the large high-quality dataset that is suitable for training StyleGAN2 from scratch. For the target one, we test a wide range of different domains that we will describe further.", + "bbox": [ + 496, + 445, + 890, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StyleGAN2 structure and its main components. We provide a diagram description of the StyleGAN2 architecture in Figure 2. It consists of three parts:", + "bbox": [ + 496, + 566, + 890, + 612 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- mapping network $f_{M}$ that transforms the input noise $z \\in \\mathcal{Z}$ (typically Gaussian) to the intermediate latent vector $w \\in \\mathcal{W}$ ;", + "- affine layers $f_1^A, \\ldots, f_N^A$ , each of them takes as input the vector $w$ and maps it to corresponding style vector $s_1 = f_1^A(w), \\ldots, s_N = f_N^A(w)$ . The concatenation of these vectors form the vector from the StyleSpace: $s = (s_1, \\ldots, s_N) \\in S$ ;", + "- synthesis network that is a composition of modulated convolutions. The weights of each convolution are modulated by the input style vector $s_i$ and applied to the input feature maps. The synthesis network also has tRGB convolutional layers that transform feature maps to RGB images and they also are modulated by corresponding style vectors." + ], + "bbox": [ + 517, + 619, + 890, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Accordingly, the StyleGAN2 $G_{\\theta}$ generates from the input noise $z$ the output image $I = G_{\\theta}(s(z))$ where $\\theta$ are", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2186", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "weights of the synthesis network, $s(z) = f^{A}(f_{M}(z))$ $f^{A} = \\{f_{1}^{A},\\dots ,f_{N}^{A}\\}$", + "bbox": [ + 76, + 90, + 467, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We will analyze these three components of StyleGAN2 and their impact on the domain adaptation process. It is common wisdom that the most important part for adaptation is the synthesis network, while the mapping network and affine layers are mostly responsible for the semantic manipulations within the source domain [44]. We aim to verify whether such a conception is correct.", + "bbox": [ + 75, + 121, + 467, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our experiments, we additionally consider the impact of the combination of affine layers and one convolutional block from the synthesis network on domain adaptation. It is a way to probe the intermediate case between affine layers and the synthesis network.", + "bbox": [ + 75, + 227, + 467, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Method to analyze the impact of each component. The paper [44] has proposed to analyze the impact of each component by resetting its weights in the fine-tuned generator to their pretrained values and assessing the quality of the generated images. In our work, we propose another approach: to directly fine-tune only one of these components to explore which ones are sufficient for domain adaptation.", + "bbox": [ + 75, + 301, + 467, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us describe our method in more detail. The optimization process for the problem of domain adaptation is the following", + "bbox": [ + 75, + 407, + 467, + 452 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta} \\left(s \\left(z _ {i}\\right)\\right)\\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 463, + 467, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_D$ is domain adaptation loss that depends on the domain $D$ (we discuss it further) and the samples from the generator $\\{G_{\\theta}(s(z_i))\\}_{i = 1}^{K}$ , $z_{1},\\ldots ,z_{K}\\in \\mathcal{Z}$ are sampled noises.", + "bbox": [ + 75, + 497, + 467, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typically the generator $G_{\\theta}$ is optimized with respect to all components, i.e.", + "bbox": [ + 75, + 556, + 467, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta} \\left(s \\left(z _ {i}\\right)\\right)\\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min _ {\\theta , f ^ {A}, f _ {M}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 597, + 467, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose to investigate settings when we optimize with respect to only one these components. We denote each parameter space as: $\\{\\theta\\} - \\text{SymConv}$ parameterization, $\\{f^A\\} - \\text{Affine}$ parameterization, $\\{f_M\\} - \\text{Mapping}$ parameterization. The case we fine-tune all components of the StyleGAN2 we call Full parameterization.", + "bbox": [ + 76, + 635, + 467, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Domain adaptation settings. In our study, we consider two settings: one-shot and few-shot. For each data regime, we use different domains depending on their similarity to the source domain of realistic faces from FFHQ:", + "bbox": [ + 75, + 724, + 467, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- one-shot domains: for this setting, we consider only similar domains. It is the case when the target domain has the same geometry of faces and it preserves the identity of the person. It alters only the style of the image. In this regime, we consider not only one-shot image-based adaptation with the reference stylized face but additionally text-based adaptation with", + "bbox": [ + 94, + 794, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the text description of the target style (e.g., \"Photo in the style of anime (pixar, sketch, etc.)\"). See examples of such domains in Figure 3.", + "bbox": [ + 529, + 90, + 890, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- few-shot domains: for this regime, we examine more distant domains that have a face-like form but change the face geometry and identity in a stronger manner. As examples, we consider AFHQ dogs faces and cats faces [8].", + "bbox": [ + 517, + 147, + 890, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depending on the data regime, we use different domain loss function $\\mathcal{L}_D$ . For one-shot domain adaptation, we apply the optimization loss from StyleGAN-NADA [11] in the case of text-based adaptation and another loss from DiFa [48] for one-shot image-based adaptation. For few-shot domain adaptation, we utilize the fine-tuning procedure from StyleGAN-ADA [18]. For more details about domain adaptation loss functions see Appendix A.1.", + "bbox": [ + 496, + 233, + 890, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain quantitative comparisons in the case of a one-shot setting, we use Quality and Diversity metrics that were proposed in the HyperDomainNet paper [4]. For few-shot adaptation, we compute FID metric [15] using the standard protocol from [18].", + "bbox": [ + 496, + 354, + 890, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis for one-shot domains. For the analysis, we choose different text-based and one-shot image-based domains (see Appendix A.2 for the full list and more details). In experiments, we consider the four parameterizations (Full, SyntConv, Affine, Mapping) we discussed above.", + "bbox": [ + 496, + 430, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We provide qualitative results in Figure 3 with quantitative ones in Table 1. More results see in Appendix A.2.", + "bbox": [ + 496, + 521, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that all three parameterizations, Full, Synt-Conv and Affine, show comparable performance in terms of both visual quality and objective metrics. The fact that the synthesis network is sufficient for similar domains was clear from the previous work [44]. However, our finding that the affine part is also sufficient is a new and surprising result. It means that we can change the domain of generated images without fine-tuning the synthesis network but just passing the modified style vector that comes from the affine part. Also, we observe that the mapping network shows poor vi", + "bbox": [ + 496, + 551, + 890, + 702 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/70b2688821fb80c1baef86f054e04713612815f378a842658d551c365cdfbceb.jpg", + "image_caption": [ + "Figure 3: Text-based and image-based adaptation for different parameterizations. Affine, StyleSpace and StyleSpaceSparse parameterizations yield performance comparable with Full one. This style image is called \"Disney\"." + ], + "image_footnote": [], + "bbox": [ + 504, + 727, + 906, + 823 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2187", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/dd4e903860fd019df2f5431ac17e7cc8d77284a3184946a721c5a9fe5290c7d5.jpg", + "table_caption": [ + "Table 1: Quality and Diversity metrics [4] for text-based and one-shot image-based domain adaptations with different parameterizations. Affine, StyleSpace and StyleSpaceSparse parameterizations achieve results comparable with the Full one." + ], + "table_footnote": [], + "table_body": "
Parameter SpaceSizeBoteroSketchDisney (image)Titan Erwin (image)
QualityDiversityQualityDiversityQualityDiversityQualityDiversity
Full30.3M0.3120.2280.2080.2960.7130.2470.7600.194
SyntConv23.6M0.3110.2240.1910.2920.7110.2590.7410.217
Affine4.6M0.2980.2210.1940.2960.5650.3590.6500.314
Mapping2.1M0.2260.1150.1820.1430.7170.0800.6450.102
StyleSpace6.0K0.3090.230.1930.3060.6270.3080.6720.296
StyleSpaceSparse1.2K0.3220.2130.2010.2690.6170.3040.6590.303
", + "bbox": [ + 76, + 130, + 893, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sual quality and low diversity in the generated images when considering Diversity metric. It indicates that for successful adaptation it is important to update the style vector from $S$ rather than the intermediate latent vector from $\\mathcal{W}$ space.", + "bbox": [ + 75, + 311, + 468, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis for few-shot domains. For this setting, as we discussed above, we consider two datasets, AFHQ Dogs and Cats [8], and the results are provided in Figure 4 and in Table 2. See more results in Appendix A.3.", + "bbox": [ + 75, + 385, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that results for Dogs and Cats are different compared to similar domains. In particular, we see that the Affine parameterization does not demonstrate the same quality as the Full parameterization. It can be seen from the degraded visual quality and visible gap in FID metric. However, it is still surprising that the generated images after adaptation have an adequate visual appearance considering that we do not fine-tune the synthesis network at all. Synt-Conv expectedly achieves results comparable with Full parameterization, and Mapping conversely shows poor quality on all datasets.", + "bbox": [ + 75, + 459, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/08f2b42e7a34e19e87b3c07701cc48340e44300a3c2deddf8c39640c4a3d28f0.jpg", + "table_caption": [ + "Table 2: FID scores for domain adaptation with different parameterizations. We observe a significant gap between Affine and Full parametrizations that, however, can be drastically reduced by introducing Affine+ parameterization." + ], + "table_footnote": [], + "table_body": "
Parameter SpaceSizeDomains
DogCat
Full30.3M20.37.1
SyntConv23.6M19.77.2
Affine4.6M70.127.6
Mapping2.1M208.2226.1
Affine+5.1M18.67.0
AffineLight+0.6M20.68.9
StyleSpace6.0K75.822.0
", + "bbox": [ + 117, + 744, + 426, + 916 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Efficient and lightweight parameterizations of StyleGAN", + "text_level": 1, + "bbox": [ + 500, + 310, + 890, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "StyleSpace and StyleSpaceSparse. Our findings from the previous section suggest that we can change the domain of generated images by modifying the style vector in the StyleSpace $S$ . To check this hypothesis, we will adapt StyleGAN2 by directly optimizing the direction in $S$ , i.e. during fine-tuning of StyleGAN we will optimize only $\\Delta s^{D}$ :", + "bbox": [ + 496, + 353, + 890, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta} \\left(s \\left(z _ {i}\\right) + \\Delta s\\right)\\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min _ {\\Delta s}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 464, + 890, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta s = (\\Delta s_1, \\ldots, \\Delta s_N) \\in S$ is the optimized direction in the $S$ for adapting the generator $G_{\\theta}$ to the domain $D$ . We call such directions $\\Delta s$ as StyleDomain directions.", + "bbox": [ + 496, + 500, + 890, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further, we explore that we can zero out most of the coordinates of StyleDomain directions without quality degradation. We use standard pruning technique when we leave $20\\%$ of the largest absolute values in the StyleDomain vector and set the rest to zero. We call such parameterization as StyleSpaceSparse. We examine different pruning rates and its performance in Appendix A.4.", + "bbox": [ + 496, + 545, + 890, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply these parameterizations to one-shot and few-shot domains and obtain the following results.", + "bbox": [ + 496, + 651, + 890, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For one-shot adaptation, we provide results in Figure 3 and in Table 1 (see more results in Appendix A.4). We observe that optimizing the StyleDomain direction $\\Delta s^{D}$", + "bbox": [ + 496, + 680, + 890, + 726 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e375af9591b6bf33ac2871c650d23e5a2a6d0ca730a95a892e1cca6b72543254.jpg", + "image_caption": [ + "Figure 4: Domain adaptation for dissimilar domains. Affine+ parameterization produces results on par with the Full one." + ], + "image_footnote": [], + "bbox": [ + 500, + 746, + 911, + 840 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2188", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "achieves the same results both visually and quantitatively as the Full parameterization. It is new and important observation that StyleSpace allows not only image editing within domain but also generating samples from out-of-domain of realistic human faces.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For few-shot domains, we observe that StyleSpace is not sufficient, which is expressed by the same significant quality degradation as for Affine parameterization. Further, we introduce a new parameterization that is efficient for more distant domains.", + "bbox": [ + 75, + 167, + 468, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Affine+ and AffineLight+. We aim to improve Affine parameterization for successfully adapting to dissimilar domains such as Dogs and Cats. We propose to extend it by adding one block from synthesis layer with specified spatial resolution. Such block has two convolutional layers with weights $\\theta_{1},\\theta_{2}\\in \\mathbb{R}^{512\\times 512\\times 3\\times 3}$ . Instead of fine-tuning all these weights we introduce more compact parameterization as offsets $\\Delta \\theta_{1},\\Delta \\theta_{2}$ to these weights that are same across spatial dimensions, i.e. $\\Delta \\theta_{1},\\Delta \\theta_{2}\\in \\mathbb{R}^{512\\times 512\\times 1\\times 1}$ (we observe that such reduction in size does affect the performance). In addition to $\\Delta \\theta_{1},\\Delta \\theta_{2}$ we introduce the offsets to the weights $\\theta^{tRGB}\\in \\mathbb{R}^{3\\times 512\\times 3\\times 3}$ of the tRGB convolutional layer of the same block with similar parameterization $\\Delta \\theta^{tRGB}\\in \\mathbb{R}^{3\\times 512\\times 1\\times 1}$ . Further, we omit such detail about tRGB part in the sake of brevity. So, for this parameterization the optimization procedure has the following form:", + "bbox": [ + 75, + 243, + 468, + 498 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta , \\Delta \\theta_ {1}, \\Delta \\theta_ {2}} (s (z _ {i})) \\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min _ {\\Delta \\theta_ {1}, \\Delta \\theta_ {2}, f ^ {A}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 507, + 468, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $G_{\\Delta \\theta_1, \\Delta \\theta_2}$ is the generator with weight offsets $\\Delta \\theta_1, \\Delta \\theta_2$ for the one block from the synthesis network.", + "bbox": [ + 76, + 547, + 468, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We call such parameter space as Affine+. We examine all blocks of the synthesis network to choose for this parameterization. We end up with the block with $64 \\times 64$ resolution", + "bbox": [ + 75, + 579, + 468, + 625 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/270e298312ff9af55cc641e5b8ce4538179684e2365643b64795f4a0059c91c2.jpg", + "image_caption": [ + "Figure 5: StyleSpace directions transfer from text-based and image-based domain adaptation to other fine-tuned models. We can successfully transfer style while preserving image content." + ], + "image_footnote": [], + "bbox": [ + 76, + 647, + 486, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as it shows the best performance (see results of this analysis in Appendix A.4).", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While Affine+ has already had 6 times less parameters than Full parameterization, we further reduce its size by applying low-rank decomposition to the weights of affine layers. We denote this parameterization as AffineLight+. It gives us by two orders less parameters than Full parameterization with insignificant degradation in quality. We will show in Section 5 that in low data regimes, this parameterization achieves comparable performance as Affine+ and outperforms other baselines. More details about applied low-rank decomposition can be found in Appendix A.4.", + "bbox": [ + 496, + 122, + 892, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply these two parameterizations to few-shot domains and obtain the following results.", + "bbox": [ + 498, + 275, + 890, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We provide results in Figure 4 and in Table 2 (see more results in Appendix A.4). We see that Affine+ parameterization allows us removing the performance gap from the Full parameterization both qualitatively and quantitatively. While the number of parameters for the additional block in Affine+ accounts for only $2\\%$ of the synthesis network size. It shows that the style vector allows adapting the generator even to more distant domains if we just add a small part of the synthesis network.", + "bbox": [ + 496, + 306, + 890, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that AffineLight+ still shows adequate performance, except it has 100 times less parameters than the Full parameterization. We should notice that this parameterization is more suitable for low data regimes than those that we consider in Section 5.", + "bbox": [ + 496, + 443, + 890, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Properties of the StyleDomain directions. We investigate StyleDomain directions that adapt the generator to similar domains and explore surprising properties. The first one is mixability, i.e. we discover that StyleDomain directions can be combined with each other. In particular, we can consider several directions that correspond to different similar domains and take their sum. The resulting direction will adapt the generator to the semantically mixed domain. We provide different examples of such combinations in Figure 6", + "bbox": [ + 496, + 518, + 892, + 656 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/324ff89f310e7558ab38a9dd703d6d34e4c02a7157dee8ef664326322ba33bdf.jpg", + "image_caption": [ + "Figure 6: Example of mixing StyleDomain directions. We can combine different directions in order to perform adaptation into a semantically mixed domain." + ], + "image_footnote": [], + "bbox": [ + 506, + 695, + 898, + 835 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2189", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a9d49fe13fa9cf1e4d34b2091a72926f452d3b3052213005f98683637b34bd5.jpg", + "image_caption": [ + "Figure 7: Comparison with baselines for one-shot image-based domain adaptation. StyleSpace and StyleSpaceSparse parameterizations achieve comparable quality as other methods while having much less trainable parameters." + ], + "image_footnote": [], + "bbox": [ + 91, + 116, + 875, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(see more results in Appendix A.5).", + "bbox": [ + 76, + 345, + 316, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The second property is transferability, i.e. we can transfer StyleDomain directions between different StyleGAN2 models. In particular, let us consider the base generator $G_{\\theta}$ pretrained on realistic faces and the generator $G_{\\theta}^{Dog}$ fine-tuned to domain of dogs in the Full parameterization. We verify that we can apply StyleDomain directions to $G_{\\theta}^{Dog}$ which were optimized for $G_{\\theta}$ . Specifically, we take Style-Domain directions that were optimized for $G_{\\theta}$ to adapt it to different text-based and one-shot image-based domains (e.g. Pixar, Disney, etc.). Next, we apply these directions to generators that were fine-tuned from $G_{\\theta}$ to other domains (e.g. Dogs, Cats). We provide results of this experiment in Figure 5 (see more results in Appendix A.5).", + "bbox": [ + 75, + 369, + 472, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we check that StyleDomain directions can be successfully combined with latent controls for image editing. For a more detailed exploration of this property, see in Appendix A.5.", + "bbox": [ + 75, + 573, + 470, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 345, + 633, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "One-shot domain adaptation. For image-based one-shot adaptation, we consider the main baselines: Target-CLIP [7], JoJoGAN [9], MTG [55], GOSA [49], DiFa [48], DomMod [4]. We apply our parameterizations StyleSpace and StyleSpaceSparse to DiFa model because this configuration achieves the highest performance (experiments with other base models can be found in Appendix A.6). Similarly, we apply DomMod parameterization to the DiFa model. We use StyleGAN2 as the GAN architecture and FFHQ as the source domain. For fair comparison, we strictly follow baseline default configurations in the choice of source-target adaptation setups and hyper-parameters. As target domain we use a wide range of different style images from baseline works. The full list of target images can be found in Appendix A.6.", + "bbox": [ + 496, + 373, + 893, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide both quantitative and qualitative results in Table 3 and Figure 7, correspondingly. More results and", + "bbox": [ + 498, + 603, + 893, + 635 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ac21db5242508d2a5bc07cc553e242fe3aee00d677007f31028c9bee89bad22e.jpg", + "table_caption": [ + "Table 3: Quality and Diversity metrics [4] for one-shot image-based domain adaptations with different methods. Memory denotes the memory needed for keeping adapted generators for all 12 domains for each method. StyleSpace and StyleSpaceSparse parameterizations achieve results comparable to other baselines while having much less trainable parameters." + ], + "table_footnote": [], + "table_body": "
MethodSizeMemoryTitan ErwinDisneyAcross 12 domains
QualityDiversityQualityDiversityQualityDiversity
JoJoGAN [9]30M1.80GB0.5720.2920.5910.2600.590 ± 0.0480.257 ± 0.025
MTG [55]30M1.80GB0.6070.2690.5090.2340.586 ± 0.0540.263 ± 0.028
GOSA [49]30M1.80GB0.5470.2830.6170.2160.584 ± 0.0340.252 ± 0.030
DiFa [48]30M1.80GB0.7190.2260.6990.2630.734 ± 0.0470.215 ± 0.038
TargetCLIP [7]9.0K420KB0.4740.3060.5020.3330.491 ± 0.0430.322 ± 0.015
DomMod (DiFa) [4]6.0K280KB0.7050.2500.6250.2940.679 ± 0.0490.253 ± 0.037
StyleSpace (DiFa)6.0K280KB0.6720.2960.6270.3080.644 ± 0.0410.298 ± 0.025
StyleSpaceSparse (DiFa)1.2K56.4KB0.6590.3030.6170.3040.638 ± 0.0460.299 ± 0.026
", + "bbox": [ + 78, + 710, + 928, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2190", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "samples can be found in Appendix A.6. We observe that the DiFa model achieves the best Quality metric but has low Diversity compared to other methods. Our parameterizations applied to the DiFa model balance its performance in terms of these metrics. So, StyleSpace (DiFa) and StyleSpaceSparse (DiFa) achieves uniformly better results than JoJOGAN, MTG, GOSA baselines. DomMod parameterization also show good performance, but it is comparable with our StyleSpaceSparse parameterization that has 5 times less parameters. We also report the overall memory needed for keeping adapted generators for 12 domains. We observe that StyleSpaceSparse requires significantly less space than other models. It will be especially important if we scale the number of target domains to thousands and more. We note that TargetCLIP that also has small number of trainable parameters show very poor results visually and in terms of Quality metric. It has a high Diversity only because its generated images are very close to the original ones. To provide more comprehensive comparison we additionally conduct user studies and present results in Appendix A.6.4.", + "bbox": [ + 75, + 90, + 472, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Few-shot domain adaptation. As main baselines for this task, we consider the vanilla StyleGAN-ADA [18] (we will denote it as ADA), CDC [28] and recent SOTA method AdAM [51]. We compare our parameterizations Affine+ and AffineLight+ applied to ADA with these baselines on two datasets: Dogs and Cats [8]. We study the efficiency of these methods with different numbers of available target samples. For fair comparison, we rigorously follow the training setups of all baselines. We note that we use the same number of training iterations and the same batch size for all methods. We observe that in the AdAM paper [51] the number of training iterations is set to a small value (less than 10K) and it causes underfitting of the vanilla ADA method. Therefore, we increase this number for all methods to 50K and use the same batch size of 4. Then, for each method, we report the best FID value that it obtains during", + "bbox": [ + 75, + 397, + 472, + 640 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f5bea334163ffac3287ba57ac13af8234dba941b9aef2c153ab3feff2861680a.jpg", + "image_caption": [ + "Figure 8: Few-shot training results for different number of shots. Proposed ADA (Affine+) and ADA (AffineLight+) show uniformly better performance than baselines." + ], + "image_footnote": [], + "bbox": [ + 76, + 667, + 488, + 840 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9acdee676117753a6968a50b024ed3f30e8b894126514cf63b8fa9b40fe361dc.jpg", + "table_caption": [ + "Table 4: Results for few-shot training with 10-shots. Proposed ADA (Affine+) and ADA (AffineLight+) achieve better performance." + ], + "table_footnote": [], + "table_body": "
MethodSizeDomains (10-shots)
CatDog
CDC [28]30M66.24184.56
AdAM [51]19M47.05119.61
ADA (Full) [18]30M51.38100.25
ADA (Affine+)5.1M38.4096.38
ADA (AffineLight+)0.6M43.91101.31
", + "bbox": [ + 506, + 145, + 883, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training. See more details about training in Appendix A.7.", + "bbox": [ + 498, + 296, + 883, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report results for few-shot training for different numbers of shots in Figure 8 and separately for 10 shots in Table 4. More results and samples can be found in Appendix A.7. Firstly, we observe that the AdAM method shows results not better than the vanilla ADA (Full) when trained for sufficient number of iterations. Secondly, we see that ADA (Affine+) and ADA (AffineLight+) based on our parameterizations achieve better results uniformly for all numbers of shots. It shows that these parameterizations is especially suitable for low data regimes.", + "bbox": [ + 496, + 311, + 890, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-domain image translation. We consider two setups of standard image-to-image problem. In the first one, we translate images from the source domain to the target domain unconditionally. In the second setup, we perform a reference-based translation, where the resulting image combines the pose and the shape of the source image with the style from a reference image. Details of these experiments can be found in Appendix A.9.", + "bbox": [ + 496, + 464, + 890, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-domain image morphing. Cross-domain morphing is a smooth transition between two images from different domains. This task is known as challenging [3, 10] and it is successfully tackled in the work [44] using aligned StyleGAN2 models. The idea is to interpolate between aligned generator weights to obtain a smooth transition between domains. We propose more complex image morphing by utilizing the transferability of StyleDomain directions. For example, we can apply a direction that stands for Sketch or Pixar style to Dogs domain to obtain a smooth transition between sketchy and pixar-like dog (see Figure 1). See many examples of such complex cross-domain morphing in Appendix A.10.", + "bbox": [ + 496, + 585, + 892, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 799, + 617, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we provide an extensive analysis of the process of StyleGAN domain adaptation. We reveal the sufficient components of the generator for successful adaptation depending on the similarity between the source and target domains. We discover the ability of StyleSpace to change", + "bbox": [ + 496, + 824, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2191", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the domain of generated images by StyleDomain directions. We also propose new efficient parameterizations Affine+ and AffineLight+ for few-shot adaptation that outperform existing baselines. Further, we explore and leverage the properties of StyleDomain directions. We believe that our investigation can attract more attention to the exploration of new and interesting properties of StyleSpace.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 76, + 212, + 256, + 229 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The results of sections 1, 2, 3 were obtained by Aibek Alanov and Dmitry Vetrov with the support of the grant for research centers in the field of AI provided by the Analytical Center for the Government of the Russian Federation (ACRF) in accordance with the agreement on the provision of subsidies (identifier of the agreement 000000D730321P5Q0002) and the agreement with HSE University No. 70-2021-00139. This research was supported in part through computational resources of HPC facilities at HSE University.", + "bbox": [ + 75, + 238, + 472, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 404, + 174, + 420 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 1, 2", + "[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 33", + "[3] Kfir Aberman, Jing Liao, Mingyi Shi, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Neural best-buddies: Sparse cross-domain correspondence. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 8", + "[4] Aibek Alanov, Vadim Titov, and Dmitry Vetrov. Hyperdomainnet: Universal domain adaptation for generative adversarial networks. arXiv preprint arXiv:2210.08884, 2022. 1, 3, 4, 5, 7, 13, 18, 23, 24, 34, 35, 36, 37", + "[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1, 2", + "[6] Kelvin CK Chan, Xintao Wang, Xiangyu Xu, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for large-factor image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14245-14254, 2021. 1", + "[7] Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf. Image-based clip-guided essence transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 695-711. Springer, 2022. 3, 7, 34, 35", + "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vi" + ], + "bbox": [ + 86, + 429, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion and pattern recognition, pages 8188-8197, 2020. 4, 5, 8, 40", + "[9] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 128-152. Springer, 2022. 3, 7, 34, 35", + "[10] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, volume 39, pages 303–313. Wiley Online Library, 2020. 8", + "[11] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 1, 3, 4, 12", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2", + "[13] Shanyan Guan, Ying Tai, Bingbing Ni, Feida Zhu, Feiyue Huang, and Xiaokang Yang. Collaborative learning for faster stylegan embedding. arXiv preprint arXiv:2007.01758, 2020.2", + "[14] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020. 1, 2", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4", + "[16] Jialu Huang, Jing Liao, and Sam Kwong. Unsupervised image-to-image translation via pre-trained stylegan2 network. IEEE Transactions on Multimedia, 24:1435-1448, 2021. 1", + "[17] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the\" steerability\" of generative adversarial networks. arXiv preprint arXiv:1907.07171, 2019. 1, 2", + "[18] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 1, 2, 4, 8, 12, 19, 40", + "[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2", + "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3", + "[21] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1, 2, 3, 12" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2192", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3", + "[23] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. In International Conference on Learning Representations, 2020. 1, 2", + "[24] Xuan Luo, Xuaner Zhang, Paul Yoo, Ricardo Martin-Brualla, Jason Lawrence, and Steven M Seitz. Time-travel rephotography. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 1", + "[25] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze the discriminator: a simple baseline for fine-tuning gans. arXiv preprint arXiv:2002.10964, 2020. 2", + "[26] M-E Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1447-1454. IEEE, 2006. 19", + "[27] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 3", + "[28] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 1, 3, 8, 40", + "[29] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 1, 2", + "[30] Stanislav Pidhorskyi, Donald A Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14104-14113, 2020. 2", + "[31] Justin NM Pinkney and Doron Adler. Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334, 2020. 1, 3, 19", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 12", + "[33] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2", + "[34] Esther Robb, Wen-Sheng Chu, Abhishek Kumar, and Jia-Bin Huang. Few-shot adaptation of generative adversarial networks. arXiv preprint arXiv:2010.11943, 2020. 3", + "[35] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vi" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion and pattern recognition, pages 9243-9252, 2020. 1, 2, 31", + "[36] Guoxian Song, Linjie Luo, Jing Liu, Wan-Chun Ma, Chunpong Lai, Chuanxia Zheng, and Tat-Jen Cham. Agilegan: stylizing portraits by inversion-consistent transfer learning. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 1", + "[37] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 1, 2", + "[38] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 2", + "[39] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021. 1, 2, 12", + "[40] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7921-7931, 2021. 3", + "[41] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9168-9178, 2021. 1", + "[42] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. Minegan: effective knowledge transfer from gans to target domains with few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9332-9341, 2020. 3", + "[43] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2, 31, 33", + "[44] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 1, 3, 4, 8, 47", + "[45] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. Advances in Neural Information Processing Systems, 34:9378-9390, 2021. 1, 2", + "[46] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021. 1", + "[47] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 19, 40" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2193", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Yabo Zhang, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, Wangmeng Zuo, et al. Towards diverse and faithful one-shot adaptation of generative adversarial networks. In Advances in Neural Information Processing Systems, 2022. 3, 4, 7, 12, 34, 35", + "[49] Zicheng Zhang, Yinglu Liu, Congying Han, Tiande Guo, Ting Yao, and Tao Mei. Generalized one-shot domain adaption of generative adversarial networks. arXiv preprint arXiv:2209.03665, 2022. 3, 7, 34, 35", + "[50] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in Neural Information Processing Systems, 33:7559-7570, 2020. 1, 2, 12", + "[51] Yunqing Zhao, Keshigeyan Chandrasegaran, Milad Abdol-lahzadeh, and Ngai-Man Cheung. Few-shot image generation via adaptation-aware kernel modulation. arXiv preprint arXiv:2210.16559, 2022. 3, 8, 40", + "[52] Zhengli Zhao, Zizhao Zhang, Ting Chen, Sameer Singh, and Han Zhang. Image augmentations for gan training. arXiv preprint arXiv:2006.02595, 2020. 1, 2, 12", + "[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2", + "[54] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In European conference on computer vision, pages 597-613. Springer, 2016. 2", + "[55] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. arXiv preprint arXiv:2110.08398, 2021. 1, 3, 7, 34, 35" + ], + "bbox": [ + 78, + 90, + 468, + 544 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2194", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/b30ac484-13c8-4c42-8711-cf47c43f4550_model.json b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/b30ac484-13c8-4c42-8711-cf47c43f4550_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b43e459917b0645105330aba1836e7df7f67120e --- /dev/null +++ b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/b30ac484-13c8-4c42-8711-cf47c43f4550_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "StyleDomain: Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.203, + 0.782, + 0.222 + ], + "angle": 0, + "content": "Aibek Alanov\\(^{2,1,*}\\), Vadim Titov\\(^{2,*}\\), Maksim Nakhodnov\\(^{2,3*}\\), Dmitry Vetrov\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.224, + 0.686, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)HSE University \\(^{2}\\)AIRI \\(^{3}\\)Lomonosov Moscow State University" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.242, + 0.778, + 0.257 + ], + "angle": 0, + "content": "aalanov@hse.ru, titov@2a2i.org, nakhodnov@2a2i.org, dvetrov@hse.ru" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.283, + 0.484, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.283, + 0.892, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.456, + 0.894, + 0.503 + ], + "angle": 0, + "content": "Figure 1: Smooth cross-domain image morphing. Morphing in the space of generator weights could be successfully combined with morphing using StyleDomain directions, i.e. directions in the StyleSpace that can adapt the generator to new domains (see Section 4)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.522, + 0.314, + 0.538 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.556, + 0.473, + 0.875 + ], + "angle": 0, + "content": "Domain adaptation of GANs is a problem of fine-tuning GAN models pretrained on a large dataset (e.g. StyleGAN) to a specific domain with few samples (e.g. painting faces, sketches, etc.). While there are many methods that tackle this problem in different ways, there are still many important questions that remain unanswered. In this paper, we provide a systematic and in-depth analysis of the domain adaptation problem of GANs, focusing on the StyleGAN model. We perform a detailed exploration of the most important parts of StyleGAN that are responsible for adapting the generator to a new domain depending on the similarity between the source and target domains. As a result of this study, we propose new efficient and lightweight parameterizations of StyleGAN for domain adaptation. Particularly, we show that there exist directions in StyleSpace (StyleDomain directions) that are sufficient for adapting to similar domains. For dissimilar domains, we propose Affine+ and AffineLight+ parameterizations that allows us to outperform existing baselines in few-shot adaptation while having significantly less training parameters. Finally, we examine StyleDomain directions and discover their many sur" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.524, + 0.895, + 0.57 + ], + "angle": 0, + "content": "prising properties that we apply for domain mixing and cross-domain image morphing. Source code can be found at https://github.com/AIRI-Institute/StyleDomain." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.602, + 0.633, + 0.618 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Recent years GANs [12, 20, 21, 19, 5] have shown impressive results in image synthesis and offered many ways to control the generated data. In particular, the state-of-the-art StyleGAN models [20, 21, 19] have many practical applications such as image enhancement [46, 24, 6, 41], image editing [17, 35, 14, 37, 1, 43, 29], image-to-image translation [31, 16, 36, 11] thanks to their high-quality image generation and their latent representation that has rich semantics and disentangled controls for localized meaningful image manipulations. However, it comes at a price, as the training of StyleGAN requires a large, high-quality dataset that significantly limits its applicability because many real-world domains are represented by few images. The standard approach to deal with this problem is transfer learning, i.e. fine-tuning the model pretrained on the source domain \\( A \\) to the target domain \\( B \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.872, + 0.895, + 0.903 + ], + "angle": 0, + "content": "There are many domain adaptation methods for StyleGAN [18, 39, 50, 52, 23, 45, 28, 31, 4, 11, 55, 44] that" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.889, + 0.205, + 0.902 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2184" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "tackle this problem in different ways depending on the number of available images (e.g., one-shot/few-shot) from the target domain \\( B \\) and the similarity between the source \\( A \\) and target \\( B \\) domains (e.g., faces \\( \\rightarrow \\) sketches, artistic portraits, or faces \\( \\rightarrow \\) dogs, cats). Most of these works implicitly assume that StyleGAN can be adapted to a new domain only if we fine-tune almost all its weights, even for similar domains. However, this common wisdom is poorly investigated and verified and there is a lack of analysis of which parts of StyleGAN are important depending on different data regimes and the similarity between domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.473 + ], + "angle": 0, + "content": "In this work, we aim to provide a systematic and comprehensive analysis of this question. Our investigation of the properties of the aligned StyleGAN models consists of several parts. First, in Section 3, we identify what parts of the StyleGAN are sufficient for its adaptation depending on the similarity between the source \\( A \\) and target \\( B \\) domains. We discover that fine-tuning the whole synthesis convolutional network is not always necessary. In the case of similar \\( A \\) and \\( B \\) domains, the affine layers are sufficient for the adaptation. For more distant domains, we should optimize more parameters, however not the whole network. It suggests investigating new more efficient and lightweight parameterizations of StyleGAN to utilize them for domain adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.822 + ], + "angle": 0, + "content": "In the second part of our analysis, we propose two new parameterizations of StyleGAN. For similar domains, we consider the latent space that is formed by the output of affine layers, i.e. StyleSpace [43]. We show that we can directly optimize directions in this space that can adapt to similar target domains with the same quality as fine-tuning all weights of StyleGAN (we call such directions as Style-Domain directions). Further, we explore that we can zero out \\(80\\%\\) of StyleDomain direction coordinates without a quality degradation that gives even more lightweight parameterization (StyleSpaceSparse). For more distant domains, we propose a new parameterization Affine+ that consists of affine layers and only one convolutional block from the synthesis network. It reduces the number of trainable parameters by 6 times and achieves the same quality. Then, we further improve Affine+ parameterization by utilizing low-rank decomposition for weights of affine layers and obtain AffineLight+ parameterization. It allows us to optimize by two orders less parameters compared to training the whole StyleGAN. These parameterizations show the state-of-the-art performance for few-shot adaptation for dissimilar domains outperforming more complicated and expensive baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Additionally, in Section 4, we inspect StyleDomain directions and discover their surprising properties. The first one is mixability, i.e. we can sum up these directions to obtain a new mixed domain (e.g., see Figure 6 as a mix of the Joker style, Pixar style and the style from the image). The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.199 + ], + "angle": 0, + "content": "second impressive property is transferability, i.e. the same StyleDomain directions can be applied to StyleGAN models that were fine-tuned to other domains (e.g., see Figure 5 where we apply directions found for faces to dogs, cats and churches). We apply these findings to standard computer vision tasks such as image-to-image translation and cross-domain morphing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.211, + 0.642, + 0.227 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.895, + 0.613 + ], + "angle": 0, + "content": "Latent Spaces of StyleGAN. After recent remarkable success of GANs [12, 20, 21, 19, 5] in image synthesis, many works appeared that explore their latent representation for controllable image manipulation. In particular, the latent space of StyleGAN [20, 21, 19] has attracted considerable attention. It consists of three levels: (i) the first latent space, \\(\\mathcal{Z}\\), is raw random noise (typically Gaussian); (ii) the intermediate latent spaces \\(\\mathcal{W}, \\mathcal{W} + [1]\\) are formed by the output of the mapping network; (iii) the last level is StyleSpace, \\(\\mathcal{S}\\), [43] that is spanned by the channel-wise style parameters after affine layers. It has been shown that these latent spaces have rich semantics [20, 21], and especially the StyleSpace that demonstrates the most disentangled and localized semantical directions [43]. In recent years many works have proposed to utilize such appealing properties of the StyleGAN latent spaces for image editing tasks [17, 35, 14, 37, 1, 43, 29]. To apply these methods for real images it is necessary to inverse them into one of the latent space of StyleGAN which is another task that draws significant attention [1, 21, 13, 30, 33, 38, 54, 53]. We should note that all mentioned methods for image manipulation by controlling StyleGAN latent space allow only in-domain editing. In this paper, we show that in StyleSpace there exist such directions that can change the domain of images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.896, + 0.885 + ], + "angle": 0, + "content": "Domain Adaptation of StyleGAN. Recent years the problem of fine-tuning StyleGAN has generated a great deal of interest as it allows training the state-of-the-art generative model for a domain with few samples. There have appeared many works that tackle this problem in different ways depending on how similar the target domain is to the source one. Roughly, these methods can be divided into two groups. The first one deals with the case when the target and source domains are dissimilar (e.g. faces \\(\\rightarrow\\) cats, churches, etc.). In contrast, the second group considers the setting of similar domains (e.g. faces \\(\\rightarrow\\) stylized faces, painting faces, sketches, etc.). Methods from the first group typically require hundreds or thousands of samples from a new domain to adapt faithfully and they leverage data augmentations [18, 39, 50, 52], or freeze some layers of the discriminator to prevent overfitting [25], or train the discriminator with auxiliary losses to match the data more accurately [23, 45]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In the setting of the second group it is sufficient to have" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2185" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.67, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.695, + 0.097, + 0.848, + 0.115 + ], + "angle": 0, + "content": "Synthesis Network" + }, + { + "type": "image_footnote", + "bbox": [ + 0.719, + 0.128, + 0.824, + 0.144 + ], + "angle": 0, + "content": "Affine Layers" + }, + { + "type": "image_footnote", + "bbox": [ + 0.702, + 0.158, + 0.841, + 0.173 + ], + "angle": 0, + "content": "Mapping Network" + }, + { + "type": "image_footnote", + "bbox": [ + 0.692, + 0.185, + 0.844, + 0.199 + ], + "angle": 0, + "content": "Affine + ConvBlock" + }, + { + "type": "image_footnote", + "bbox": [ + 0.713, + 0.213, + 0.83, + 0.228 + ], + "angle": 0, + "content": "Weight Offsets" + }, + { + "type": "list", + "bbox": [ + 0.692, + 0.097, + 0.848, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.354, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Figure 2: StyleGAN2 architecture. We introduce new latent space \\( S + \\) for the for domain adaptation that combines StyleSpace and weight offsets for one block from the synthesis network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.47, + 0.56 + ], + "angle": 0, + "content": "only several samples (up to dozens) from the target domain for the successful adaptation and such approaches utilize another techniques. In particular, they introduce additional regularization terms [40, 22], preserve pairwise distances between instances in the source domain via cross-domain consistency loss [28], mix weights of the fine-tuned and the base generators [31], use an auxiliary small network to enhance the training [42]. More advanced methods [11, 55, 4] utilize the pretrained CLIP model [32] as the vision-language supervision for the text-based adaptation [11, 4] or the one-shot image-based adaptation [7, 55, 9, 49, 48, 4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.746 + ], + "angle": 0, + "content": "Recent work [51] analyzes the performance of methods from the second group for dissimilar domains in the low data regime. It was shown that those methods have a significant quality degradation in the few-shot regime. However, it can be partially mitigated by constraining model parameters. There are also approaches that introduce more lightweight parameterizations for domain adaptation [27, 34, 7, 4], however they work only for similar domains. In our paper, we propose efficient and highly lightweight parameterizations for both similar and dissimilar domains and show that they can achieve results on par with the state-of-the-art methods that optimize all weights of StyleGAN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "There are few works that analyze the domain adaptation process of StyleGAN thoroughly. The paper [44] is the first attempt to perform such an in-depth study. In particular, they explore which parts of StyleGAN are mostly changed during fine-tuning and the transferability of the latent semantics after domain adaptation. However, this work does not analyze which parts of StyleGAN are sufficient for adapting depending on the similarity between source and target domains. In our paper, we provide a more systematic and comprehensive analysis that completes and improves" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.394, + 0.641, + 0.409 + ], + "angle": 0, + "content": "the results from [44]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.874, + 0.438 + ], + "angle": 0, + "content": "3. Importance of Each Part of the StyleGAN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.566 + ], + "angle": 0, + "content": "In this section, our goal is to analyze what parts of StyleGAN are important for domain adaptation. Similarly to previous works, we specifically focus on the state-of-the-art GAN architecture, StyleGAN2 [21]. For the source domain, we consider FFHQ [20] as it is the large high-quality dataset that is suitable for training StyleGAN2 from scratch. For the target one, we test a wide range of different domains that we will describe further." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.613 + ], + "angle": 0, + "content": "StyleGAN2 structure and its main components. We provide a diagram description of the StyleGAN2 architecture in Figure 2. It consists of three parts:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.62, + 0.892, + 0.665 + ], + "angle": 0, + "content": "- mapping network \\( f_{M} \\) that transforms the input noise \\( z \\in \\mathcal{Z} \\) (typically Gaussian) to the intermediate latent vector \\( w \\in \\mathcal{W} \\);" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.673, + 0.892, + 0.75 + ], + "angle": 0, + "content": "- affine layers \\( f_1^A, \\ldots, f_N^A \\), each of them takes as input the vector \\( w \\) and maps it to corresponding style vector \\( s_1 = f_1^A(w), \\ldots, s_N = f_N^A(w) \\). The concatenation of these vectors form the vector from the StyleSpace: \\( s = (s_1, \\ldots, s_N) \\in S \\);" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.758, + 0.892, + 0.864 + ], + "angle": 0, + "content": "- synthesis network that is a composition of modulated convolutions. The weights of each convolution are modulated by the input style vector \\( s_i \\) and applied to the input feature maps. The synthesis network also has tRGB convolutional layers that transform feature maps to RGB images and they also are modulated by corresponding style vectors." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.62, + 0.892, + 0.864 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Accordingly, the StyleGAN2 \\( G_{\\theta} \\) generates from the input noise \\( z \\) the output image \\( I = G_{\\theta}(s(z)) \\) where \\( \\theta \\) are" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2186" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.468, + 0.122 + ], + "angle": 0, + "content": "weights of the synthesis network, \\(s(z) = f^{A}(f_{M}(z))\\) \\(f^{A} = \\{f_{1}^{A},\\dots ,f_{N}^{A}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.227 + ], + "angle": 0, + "content": "We will analyze these three components of StyleGAN2 and their impact on the domain adaptation process. It is common wisdom that the most important part for adaptation is the synthesis network, while the mapping network and affine layers are mostly responsible for the semantic manipulations within the source domain [44]. We aim to verify whether such a conception is correct." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.468, + 0.302 + ], + "angle": 0, + "content": "In our experiments, we additionally consider the impact of the combination of affine layers and one convolutional block from the synthesis network on domain adaptation. It is a way to probe the intermediate case between affine layers and the synthesis network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.468, + 0.408 + ], + "angle": 0, + "content": "Method to analyze the impact of each component. The paper [44] has proposed to analyze the impact of each component by resetting its weights in the fine-tuned generator to their pretrained values and assessing the quality of the generated images. In our work, we propose another approach: to directly fine-tune only one of these components to explore which ones are sufficient for domain adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.468, + 0.453 + ], + "angle": 0, + "content": "Let us describe our method in more detail. The optimization process for the problem of domain adaptation is the following" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.464, + 0.468, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta} \\left(s \\left(z _ {i}\\right)\\right)\\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.468, + 0.557 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_D\\) is domain adaptation loss that depends on the domain \\(D\\) (we discuss it further) and the samples from the generator \\(\\{G_{\\theta}(s(z_i))\\}_{i = 1}^{K}\\), \\(z_{1},\\ldots ,z_{K}\\in \\mathcal{Z}\\) are sampled noises." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.468, + 0.588 + ], + "angle": 0, + "content": "Typically the generator \\( G_{\\theta} \\) is optimized with respect to all components, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.598, + 0.468, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta} \\left(s \\left(z _ {i}\\right)\\right)\\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min _ {\\theta , f ^ {A}, f _ {M}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.636, + 0.468, + 0.726 + ], + "angle": 0, + "content": "We propose to investigate settings when we optimize with respect to only one these components. We denote each parameter space as: \\(\\{\\theta\\} - \\text{SymConv}\\) parameterization, \\(\\{f^A\\} - \\text{Affine}\\) parameterization, \\(\\{f_M\\} - \\text{Mapping}\\) parameterization. The case we fine-tune all components of the StyleGAN2 we call Full parameterization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.468, + 0.786 + ], + "angle": 0, + "content": "Domain adaptation settings. In our study, we consider two settings: one-shot and few-shot. For each data regime, we use different domains depending on their similarity to the source domain of realistic faces from FFHQ:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- one-shot domains: for this setting, we consider only similar domains. It is the case when the target domain has the same geometry of faces and it preserves the identity of the person. It alters only the style of the image. In this regime, we consider not only one-shot image-based adaptation with the reference stylized face but additionally text-based adaptation with" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "the text description of the target style (e.g., \"Photo in the style of anime (pixar, sketch, etc.)\"). See examples of such domains in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.148, + 0.892, + 0.223 + ], + "angle": 0, + "content": "- few-shot domains: for this regime, we examine more distant domains that have a face-like form but change the face geometry and identity in a stronger manner. As examples, we consider AFHQ dogs faces and cats faces [8]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Depending on the data regime, we use different domain loss function \\(\\mathcal{L}_D\\). For one-shot domain adaptation, we apply the optimization loss from StyleGAN-NADA [11] in the case of text-based adaptation and another loss from DiFa [48] for one-shot image-based adaptation. For few-shot domain adaptation, we utilize the fine-tuning procedure from StyleGAN-ADA [18]. For more details about domain adaptation loss functions see Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.431 + ], + "angle": 0, + "content": "To obtain quantitative comparisons in the case of a one-shot setting, we use Quality and Diversity metrics that were proposed in the HyperDomainNet paper [4]. For few-shot adaptation, we compute FID metric [15] using the standard protocol from [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Analysis for one-shot domains. For the analysis, we choose different text-based and one-shot image-based domains (see Appendix A.2 for the full list and more details). In experiments, we consider the four parameterizations (Full, SyntConv, Affine, Mapping) we discussed above." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.552 + ], + "angle": 0, + "content": "We provide qualitative results in Figure 3 with quantitative ones in Table 1. More results see in Appendix A.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.703 + ], + "angle": 0, + "content": "We observe that all three parameterizations, Full, Synt-Conv and Affine, show comparable performance in terms of both visual quality and objective metrics. The fact that the synthesis network is sufficient for similar domains was clear from the previous work [44]. However, our finding that the affine part is also sufficient is a new and surprising result. It means that we can change the domain of generated images without fine-tuning the synthesis network but just passing the modified style vector that comes from the affine part. Also, we observe that the mapping network shows poor vi" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.728, + 0.907, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.836, + 0.892, + 0.897 + ], + "angle": 0, + "content": "Figure 3: Text-based and image-based adaptation for different parameterizations. Affine, StyleSpace and StyleSpaceSparse parameterizations yield performance comparable with Full one. This style image is called \"Disney\"." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2187" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Table 1: Quality and Diversity metrics [4] for text-based and one-shot image-based domain adaptations with different parameterizations. Affine, StyleSpace and StyleSpaceSparse parameterizations achieve results comparable with the Full one." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.131, + 0.895, + 0.289 + ], + "angle": 0, + "content": "
Parameter SpaceSizeBoteroSketchDisney (image)Titan Erwin (image)
QualityDiversityQualityDiversityQualityDiversityQualityDiversity
Full30.3M0.3120.2280.2080.2960.7130.2470.7600.194
SyntConv23.6M0.3110.2240.1910.2920.7110.2590.7410.217
Affine4.6M0.2980.2210.1940.2960.5650.3590.6500.314
Mapping2.1M0.2260.1150.1820.1430.7170.0800.6450.102
StyleSpace6.0K0.3090.230.1930.3060.6270.3080.6720.296
StyleSpaceSparse1.2K0.3220.2130.2010.2690.6170.3040.6590.303
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.47, + 0.373 + ], + "angle": 0, + "content": "sual quality and low diversity in the generated images when considering Diversity metric. It indicates that for successful adaptation it is important to update the style vector from \\(S\\) rather than the intermediate latent vector from \\(\\mathcal{W}\\) space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.47, + 0.448 + ], + "angle": 0, + "content": "Analysis for few-shot domains. For this setting, as we discussed above, we consider two datasets, AFHQ Dogs and Cats [8], and the results are provided in Figure 4 and in Table 2. See more results in Appendix A.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.47, + 0.626 + ], + "angle": 0, + "content": "We observe that results for Dogs and Cats are different compared to similar domains. In particular, we see that the Affine parameterization does not demonstrate the same quality as the Full parameterization. It can be seen from the degraded visual quality and visible gap in FID metric. However, it is still surprising that the generated images after adaptation have an adequate visual appearance considering that we do not fine-tune the synthesis network at all. Synt-Conv expectedly achieves results comparable with Full parameterization, and Mapping conversely shows poor quality on all datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Table 2: FID scores for domain adaptation with different parameterizations. We observe a significant gap between Affine and Full parametrizations that, however, can be drastically reduced by introducing Affine+ parameterization." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.745, + 0.428, + 0.917 + ], + "angle": 0, + "content": "
Parameter SpaceSizeDomains
DogCat
Full30.3M20.37.1
SyntConv23.6M19.77.2
Affine4.6M70.127.6
Mapping2.1M208.2226.1
Affine+5.1M18.67.0
AffineLight+0.6M20.68.9
StyleSpace6.0K75.822.0
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.311, + 0.892, + 0.346 + ], + "angle": 0, + "content": "4. Efficient and lightweight parameterizations of StyleGAN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.459 + ], + "angle": 0, + "content": "StyleSpace and StyleSpaceSparse. Our findings from the previous section suggest that we can change the domain of generated images by modifying the style vector in the StyleSpace \\( S \\). To check this hypothesis, we will adapt StyleGAN2 by directly optimizing the direction in \\( S \\), i.e. during fine-tuning of StyleGAN we will optimize only \\( \\Delta s^{D} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.465, + 0.892, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta} \\left(s \\left(z _ {i}\\right) + \\Delta s\\right)\\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min _ {\\Delta s}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.545 + ], + "angle": 0, + "content": "where \\(\\Delta s = (\\Delta s_1, \\ldots, \\Delta s_N) \\in S\\) is the optimized direction in the \\(S\\) for adapting the generator \\(G_{\\theta}\\) to the domain \\(D\\). We call such directions \\(\\Delta s\\) as StyleDomain directions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.652 + ], + "angle": 0, + "content": "Further, we explore that we can zero out most of the coordinates of StyleDomain directions without quality degradation. We use standard pruning technique when we leave \\(20\\%\\) of the largest absolute values in the StyleDomain vector and set the rest to zero. We call such parameterization as StyleSpaceSparse. We examine different pruning rates and its performance in Appendix A.4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.681 + ], + "angle": 0, + "content": "We apply these parameterizations to one-shot and few-shot domains and obtain the following results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.727 + ], + "angle": 0, + "content": "For one-shot adaptation, we provide results in Figure 3 and in Table 1 (see more results in Appendix A.4). We observe that optimizing the StyleDomain direction \\(\\Delta s^{D}\\)" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.747, + 0.912, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.851, + 0.892, + 0.896 + ], + "angle": 0, + "content": "Figure 4: Domain adaptation for dissimilar domains. Affine+ parameterization produces results on par with the Full one." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2188" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "achieves the same results both visually and quantitatively as the Full parameterization. It is new and important observation that StyleSpace allows not only image editing within domain but also generating samples from out-of-domain of realistic human faces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.469, + 0.243 + ], + "angle": 0, + "content": "For few-shot domains, we observe that StyleSpace is not sufficient, which is expressed by the same significant quality degradation as for Affine parameterization. Further, we introduce a new parameterization that is efficient for more distant domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.47, + 0.499 + ], + "angle": 0, + "content": "Affine+ and AffineLight+. We aim to improve Affine parameterization for successfully adapting to dissimilar domains such as Dogs and Cats. We propose to extend it by adding one block from synthesis layer with specified spatial resolution. Such block has two convolutional layers with weights \\(\\theta_{1},\\theta_{2}\\in \\mathbb{R}^{512\\times 512\\times 3\\times 3}\\). Instead of fine-tuning all these weights we introduce more compact parameterization as offsets \\(\\Delta \\theta_{1},\\Delta \\theta_{2}\\) to these weights that are same across spatial dimensions, i.e. \\(\\Delta \\theta_{1},\\Delta \\theta_{2}\\in \\mathbb{R}^{512\\times 512\\times 1\\times 1}\\) (we observe that such reduction in size does affect the performance). In addition to \\(\\Delta \\theta_{1},\\Delta \\theta_{2}\\) we introduce the offsets to the weights \\(\\theta^{tRGB}\\in \\mathbb{R}^{3\\times 512\\times 3\\times 3}\\) of the tRGB convolutional layer of the same block with similar parameterization \\(\\Delta \\theta^{tRGB}\\in \\mathbb{R}^{3\\times 512\\times 1\\times 1}\\). Further, we omit such detail about tRGB part in the sake of brevity. So, for this parameterization the optimization procedure has the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.508, + 0.469, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta , \\Delta \\theta_ {1}, \\Delta \\theta_ {2}} (s (z _ {i})) \\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min _ {\\Delta \\theta_ {1}, \\Delta \\theta_ {2}, f ^ {A}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.549, + 0.469, + 0.579 + ], + "angle": 0, + "content": "where \\( G_{\\Delta \\theta_1, \\Delta \\theta_2} \\) is the generator with weight offsets \\( \\Delta \\theta_1, \\Delta \\theta_2 \\) for the one block from the synthesis network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.625 + ], + "angle": 0, + "content": "We call such parameter space as Affine+. We examine all blocks of the synthesis network to choose for this parameterization. We end up with the block with \\(64 \\times 64\\) resolution" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.648, + 0.487, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.836, + 0.47, + 0.897 + ], + "angle": 0, + "content": "Figure 5: StyleSpace directions transfer from text-based and image-based domain adaptation to other fine-tuned models. We can successfully transfer style while preserving image content." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "as it shows the best performance (see results of this analysis in Appendix A.4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.893, + 0.274 + ], + "angle": 0, + "content": "While Affine+ has already had 6 times less parameters than Full parameterization, we further reduce its size by applying low-rank decomposition to the weights of affine layers. We denote this parameterization as AffineLight+. It gives us by two orders less parameters than Full parameterization with insignificant degradation in quality. We will show in Section 5 that in low data regimes, this parameterization achieves comparable performance as Affine+ and outperforms other baselines. More details about applied low-rank decomposition can be found in Appendix A.4." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.276, + 0.892, + 0.306 + ], + "angle": 0, + "content": "We apply these two parameterizations to few-shot domains and obtain the following results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.442 + ], + "angle": 0, + "content": "We provide results in Figure 4 and in Table 2 (see more results in Appendix A.4). We see that Affine+ parameterization allows us removing the performance gap from the Full parameterization both qualitatively and quantitatively. While the number of parameters for the additional block in Affine+ accounts for only \\(2\\%\\) of the synthesis network size. It shows that the style vector allows adapting the generator even to more distant domains if we just add a small part of the synthesis network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.518 + ], + "angle": 0, + "content": "We observe that AffineLight+ still shows adequate performance, except it has 100 times less parameters than the Full parameterization. We should notice that this parameterization is more suitable for low data regimes than those that we consider in Section 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.893, + 0.657 + ], + "angle": 0, + "content": "Properties of the StyleDomain directions. We investigate StyleDomain directions that adapt the generator to similar domains and explore surprising properties. The first one is mixability, i.e. we discover that StyleDomain directions can be combined with each other. In particular, we can consider several directions that correspond to different similar domains and take their sum. The resulting direction will adapt the generator to the semantically mixed domain. We provide different examples of such combinations in Figure 6" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.696, + 0.9, + 0.837 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.851, + 0.892, + 0.897 + ], + "angle": 0, + "content": "Figure 6: Example of mixing StyleDomain directions. We can combine different directions in order to perform adaptation into a semantically mixed domain." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2189" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.117, + 0.877, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 7: Comparison with baselines for one-shot image-based domain adaptation. StyleSpace and StyleSpaceSparse parameterizations achieve comparable quality as other methods while having much less trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.347, + 0.318, + 0.364 + ], + "angle": 0, + "content": "(see more results in Appendix A.5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.473, + 0.568 + ], + "angle": 0, + "content": "The second property is transferability, i.e. we can transfer StyleDomain directions between different StyleGAN2 models. In particular, let us consider the base generator \\( G_{\\theta} \\) pretrained on realistic faces and the generator \\( G_{\\theta}^{Dog} \\) fine-tuned to domain of dogs in the Full parameterization. We verify that we can apply StyleDomain directions to \\( G_{\\theta}^{Dog} \\) which were optimized for \\( G_{\\theta} \\). Specifically, we take Style-Domain directions that were optimized for \\( G_{\\theta} \\) to adapt it to different text-based and one-shot image-based domains (e.g. Pixar, Disney, etc.). Next, we apply these directions to generators that were fine-tuned from \\( G_{\\theta} \\) to other domains (e.g. Dogs, Cats). We provide results of this experiment in Figure 5 (see more results in Appendix A.5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.471, + 0.635 + ], + "angle": 0, + "content": "Finally, we check that StyleDomain directions can be successfully combined with latent controls for image editing. For a more detailed exploration of this property, see in Appendix A.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.346, + 0.634, + 0.364 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.895, + 0.602 + ], + "angle": 0, + "content": "One-shot domain adaptation. For image-based one-shot adaptation, we consider the main baselines: Target-CLIP [7], JoJoGAN [9], MTG [55], GOSA [49], DiFa [48], DomMod [4]. We apply our parameterizations StyleSpace and StyleSpaceSparse to DiFa model because this configuration achieves the highest performance (experiments with other base models can be found in Appendix A.6). Similarly, we apply DomMod parameterization to the DiFa model. We use StyleGAN2 as the GAN architecture and FFHQ as the source domain. For fair comparison, we strictly follow baseline default configurations in the choice of source-target adaptation setups and hyper-parameters. As target domain we use a wide range of different style images from baseline works. The full list of target images can be found in Appendix A.6." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.604, + 0.895, + 0.636 + ], + "angle": 0, + "content": "We provide both quantitative and qualitative results in Table 3 and Figure 7, correspondingly. More results and" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.655, + 0.895, + 0.703 + ], + "angle": 0, + "content": "Table 3: Quality and Diversity metrics [4] for one-shot image-based domain adaptations with different methods. Memory denotes the memory needed for keeping adapted generators for all 12 domains for each method. StyleSpace and StyleSpaceSparse parameterizations achieve results comparable to other baselines while having much less trainable parameters." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.711, + 0.929, + 0.898 + ], + "angle": 0, + "content": "
MethodSizeMemoryTitan ErwinDisneyAcross 12 domains
QualityDiversityQualityDiversityQualityDiversity
JoJoGAN [9]30M1.80GB0.5720.2920.5910.2600.590 ± 0.0480.257 ± 0.025
MTG [55]30M1.80GB0.6070.2690.5090.2340.586 ± 0.0540.263 ± 0.028
GOSA [49]30M1.80GB0.5470.2830.6170.2160.584 ± 0.0340.252 ± 0.030
DiFa [48]30M1.80GB0.7190.2260.6990.2630.734 ± 0.0470.215 ± 0.038
TargetCLIP [7]9.0K420KB0.4740.3060.5020.3330.491 ± 0.0430.322 ± 0.015
DomMod (DiFa) [4]6.0K280KB0.7050.2500.6250.2940.679 ± 0.0490.253 ± 0.037
StyleSpace (DiFa)6.0K280KB0.6720.2960.6270.3080.644 ± 0.0410.298 ± 0.025
StyleSpaceSparse (DiFa)1.2K56.4KB0.6590.3030.6170.3040.638 ± 0.0460.299 ± 0.026
" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "2190" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.395 + ], + "angle": 0, + "content": "samples can be found in Appendix A.6. We observe that the DiFa model achieves the best Quality metric but has low Diversity compared to other methods. Our parameterizations applied to the DiFa model balance its performance in terms of these metrics. So, StyleSpace (DiFa) and StyleSpaceSparse (DiFa) achieves uniformly better results than JoJOGAN, MTG, GOSA baselines. DomMod parameterization also show good performance, but it is comparable with our StyleSpaceSparse parameterization that has 5 times less parameters. We also report the overall memory needed for keeping adapted generators for 12 domains. We observe that StyleSpaceSparse requires significantly less space than other models. It will be especially important if we scale the number of target domains to thousands and more. We note that TargetCLIP that also has small number of trainable parameters show very poor results visually and in terms of Quality metric. It has a high Diversity only because its generated images are very close to the original ones. To provide more comprehensive comparison we additionally conduct user studies and present results in Appendix A.6.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.641 + ], + "angle": 0, + "content": "Few-shot domain adaptation. As main baselines for this task, we consider the vanilla StyleGAN-ADA [18] (we will denote it as ADA), CDC [28] and recent SOTA method AdAM [51]. We compare our parameterizations Affine+ and AffineLight+ applied to ADA with these baselines on two datasets: Dogs and Cats [8]. We study the efficiency of these methods with different numbers of available target samples. For fair comparison, we rigorously follow the training setups of all baselines. We note that we use the same number of training iterations and the same batch size for all methods. We observe that in the AdAM paper [51] the number of training iterations is set to a small value (less than 10K) and it causes underfitting of the vanilla ADA method. Therefore, we increase this number for all methods to 50K and use the same batch size of 4. Then, for each method, we report the best FID value that it obtains during" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.669, + 0.489, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.851, + 0.472, + 0.898 + ], + "angle": 0, + "content": "Figure 8: Few-shot training results for different number of shots. Proposed ADA (Affine+) and ADA (AffineLight+) show uniformly better performance than baselines." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Table 4: Results for few-shot training with 10-shots. Proposed ADA (Affine+) and ADA (AffineLight+) achieve better performance." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.146, + 0.885, + 0.288 + ], + "angle": 0, + "content": "
MethodSizeDomains (10-shots)
CatDog
CDC [28]30M66.24184.56
AdAM [51]19M47.05119.61
ADA (Full) [18]30M51.38100.25
ADA (Affine+)5.1M38.4096.38
ADA (AffineLight+)0.6M43.91101.31
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.297, + 0.885, + 0.312 + ], + "angle": 0, + "content": "training. See more details about training in Appendix A.7." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.464 + ], + "angle": 0, + "content": "We report results for few-shot training for different numbers of shots in Figure 8 and separately for 10 shots in Table 4. More results and samples can be found in Appendix A.7. Firstly, we observe that the AdAM method shows results not better than the vanilla ADA (Full) when trained for sufficient number of iterations. Secondly, we see that ADA (Affine+) and ADA (AffineLight+) based on our parameterizations achieve better results uniformly for all numbers of shots. It shows that these parameterizations is especially suitable for low data regimes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.892, + 0.585 + ], + "angle": 0, + "content": "Cross-domain image translation. We consider two setups of standard image-to-image problem. In the first one, we translate images from the source domain to the target domain unconditionally. In the second setup, we perform a reference-based translation, where the resulting image combines the pose and the shape of the source image with the style from a reference image. Details of these experiments can be found in Appendix A.9." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.893, + 0.783 + ], + "angle": 0, + "content": "Cross-domain image morphing. Cross-domain morphing is a smooth transition between two images from different domains. This task is known as challenging [3, 10] and it is successfully tackled in the work [44] using aligned StyleGAN2 models. The idea is to interpolate between aligned generator weights to obtain a smooth transition between domains. We propose more complex image morphing by utilizing the transferability of StyleDomain directions. For example, we can apply a direction that stands for Sketch or Pixar style to Dogs domain to obtain a smooth transition between sketchy and pixar-like dog (see Figure 1). See many examples of such complex cross-domain morphing in Appendix A.10." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.619, + 0.815 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this paper, we provide an extensive analysis of the process of StyleGAN domain adaptation. We reveal the sufficient components of the generator for successful adaptation depending on the similarity between the source and target domains. We discover the ability of StyleSpace to change" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2191" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "the domain of generated images by StyleDomain directions. We also propose new efficient parameterizations Affine+ and AffineLight+ for few-shot adaptation that outperform existing baselines. Further, we explore and leverage the properties of StyleDomain directions. We believe that our investigation can attract more attention to the exploration of new and interesting properties of StyleSpace." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.213, + 0.258, + 0.23 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.473, + 0.39 + ], + "angle": 0, + "content": "The results of sections 1, 2, 3 were obtained by Aibek Alanov and Dmitry Vetrov with the support of the grant for research centers in the field of AI provided by the Analytical Center for the Government of the Russian Federation (ACRF) in accordance with the agreement on the provision of subsidies (identifier of the agreement 000000D730321P5Q0002) and the agreement with HSE University No. 70-2021-00139. This research was supported in part through computational resources of HPC facilities at HSE University." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.405, + 0.175, + 0.421 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.47, + 0.487 + ], + "angle": 0, + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.472, + 0.557 + ], + "angle": 0, + "content": "[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 33" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.615 + ], + "angle": 0, + "content": "[3] Kfir Aberman, Jing Liao, Mingyi Shi, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Neural best-buddies: Sparse cross-domain correspondence. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.47, + 0.672 + ], + "angle": 0, + "content": "[4] Aibek Alanov, Vadim Titov, and Dmitry Vetrov. Hyperdomainnet: Universal domain adaptation for generative adversarial networks. arXiv preprint arXiv:2210.08884, 2022. 1, 3, 4, 5, 7, 13, 18, 23, 24, 34, 35, 36, 37" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.674, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.717, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[6] Kelvin CK Chan, Xintao Wang, Xiangyu Xu, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for large-factor image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14245-14254, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.788, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[7] Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf. Image-based clip-guided essence transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 695-711. Springer, 2022. 3, 7, 34, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vi" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.43, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "sion and pattern recognition, pages 8188-8197, 2020. 4, 5, 8, 40" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.121, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[9] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 128-152. Springer, 2022. 3, 7, 34, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[10] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, volume 39, pages 303–313. Wiley Online Library, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[11] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 1, 3, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[13] Shanyan Guan, Ying Tai, Bingbing Ni, Feida Zhu, Feiyue Huang, and Xiaokang Yang. Collaborative learning for faster stylegan embedding. arXiv preprint arXiv:2007.01758, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[14] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[16] Jialu Huang, Jing Liao, and Sam Kwong. Unsupervised image-to-image translation via pre-trained stylegan2 network. IEEE Transactions on Multimedia, 24:1435-1448, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[17] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the\" steerability\" of generative adversarial networks. arXiv preprint arXiv:1907.07171, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[18] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 1, 2, 4, 8, 12, 19, 40" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[21] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1, 2, 3, 12" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2192" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[22] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[23] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. In International Conference on Learning Representations, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[24] Xuan Luo, Xuaner Zhang, Paul Yoo, Ricardo Martin-Brualla, Jason Lawrence, and Steven M Seitz. Time-travel rephotography. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[25] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze the discriminator: a simple baseline for fine-tuning gans. arXiv preprint arXiv:2002.10964, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[26] M-E Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1447-1454. IEEE, 2006. 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.405 + ], + "angle": 0, + "content": "[27] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[28] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 1, 3, 8, 40" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[29] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.47, + 0.617 + ], + "angle": 0, + "content": "[30] Stanislav Pidhorskyi, Donald A Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14104-14113, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[31] Justin NM Pinkney and Doron Adler. Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334, 2020. 1, 3, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[33] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[34] Esther Robb, Wen-Sheng Chu, Abhishek Kumar, and Jia-Bin Huang. Few-shot adaptation of generative adversarial networks. arXiv preprint arXiv:2010.11943, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vi" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "sion and pattern recognition, pages 9243-9252, 2020. 1, 2, 31" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[36] Guoxian Song, Linjie Luo, Jing Liu, Wan-Chun Ma, Chunpong Lai, Chuanxia Zheng, and Tat-Jen Cham. Agilegan: stylizing portraits by inversion-consistent transfer learning. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[37] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[38] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[39] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021. 1, 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[40] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7921-7931, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[41] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9168-9178, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[42] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. Minegan: effective knowledge transfer from gans to target domains with few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9332-9341, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[43] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2, 31, 33" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[44] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 1, 3, 4, 8, 47" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[45] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. Advances in Neural Information Processing Systems, 34:9378-9390, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[46] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[47] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 19, 40" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "2193" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[48] Yabo Zhang, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, Wangmeng Zuo, et al. Towards diverse and faithful one-shot adaptation of generative adversarial networks. In Advances in Neural Information Processing Systems, 2022. 3, 4, 7, 12, 34, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[49] Zicheng Zhang, Yinglu Liu, Congying Han, Tiande Guo, Ting Yao, and Tao Mei. Generalized one-shot domain adaption of generative adversarial networks. arXiv preprint arXiv:2209.03665, 2022. 3, 7, 34, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[50] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in Neural Information Processing Systems, 33:7559-7570, 2020. 1, 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.331 + ], + "angle": 0, + "content": "[51] Yunqing Zhao, Keshigeyan Chandrasegaran, Milad Abdol-lahzadeh, and Ngai-Man Cheung. Few-shot image generation via adaptation-aware kernel modulation. arXiv preprint arXiv:2210.16559, 2022. 3, 8, 40" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[52] Zhengli Zhao, Zizhao Zhang, Ting Chen, Sameer Singh, and Han Zhang. Image augmentations for gan training. arXiv preprint arXiv:2006.02595, 2020. 1, 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.47, + 0.431 + ], + "angle": 0, + "content": "[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.47, + 0.488 + ], + "angle": 0, + "content": "[54] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In European conference on computer vision, pages 597-613. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.47, + 0.545 + ], + "angle": 0, + "content": "[55] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. arXiv preprint arXiv:2110.08398, 2021. 1, 3, 7, 34, 35" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2194" + } + ] +] \ No newline at end of file diff --git a/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/b30ac484-13c8-4c42-8711-cf47c43f4550_origin.pdf b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/b30ac484-13c8-4c42-8711-cf47c43f4550_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..75d1ab10a85167d3e5e48a1783df2b1373556bcc --- /dev/null +++ b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/b30ac484-13c8-4c42-8711-cf47c43f4550_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3638facd8e43999e6298466095cecdfd0302aae714be4c35ff3cfe55e599b880 +size 8799665 diff --git a/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/full.md b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e451219f78f2d6e13c58b251ca8b8deaf4f827f3 --- /dev/null +++ b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/full.md @@ -0,0 +1,292 @@ +# StyleDomain: Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation + +Aibek Alanov $^{2,1,*}$ , Vadim Titov $^{2,*}$ , Maksim Nakhodnov $^{2,3*}$ , Dmitry Vetrov $^{1,2}$ + +$^{1}$ HSE University $^{2}$ AIRI $^{3}$ Lomonosov Moscow State University + +aalanov@hse.ru, titov@2a2i.org, nakhodnov@2a2i.org, dvetrov@hse.ru + +![](images/703fc3be0760b37e479d897a4a507ca0ff9c997fe1607ba5cbf88db2032d4c9b.jpg) +Figure 1: Smooth cross-domain image morphing. Morphing in the space of generator weights could be successfully combined with morphing using StyleDomain directions, i.e. directions in the StyleSpace that can adapt the generator to new domains (see Section 4). + +![](images/295ca5be4e5e5a987fdcfb3a2a6f80cd846f89d49c53320019adac59146680a2.jpg) + +# Abstract + +Domain adaptation of GANs is a problem of fine-tuning GAN models pretrained on a large dataset (e.g. StyleGAN) to a specific domain with few samples (e.g. painting faces, sketches, etc.). While there are many methods that tackle this problem in different ways, there are still many important questions that remain unanswered. In this paper, we provide a systematic and in-depth analysis of the domain adaptation problem of GANs, focusing on the StyleGAN model. We perform a detailed exploration of the most important parts of StyleGAN that are responsible for adapting the generator to a new domain depending on the similarity between the source and target domains. As a result of this study, we propose new efficient and lightweight parameterizations of StyleGAN for domain adaptation. Particularly, we show that there exist directions in StyleSpace (StyleDomain directions) that are sufficient for adapting to similar domains. For dissimilar domains, we propose Affine+ and AffineLight+ parameterizations that allows us to outperform existing baselines in few-shot adaptation while having significantly less training parameters. Finally, we examine StyleDomain directions and discover their many sur + +prising properties that we apply for domain mixing and cross-domain image morphing. Source code can be found at https://github.com/AIRI-Institute/StyleDomain. + +# 1. Introduction + +Recent years GANs [12, 20, 21, 19, 5] have shown impressive results in image synthesis and offered many ways to control the generated data. In particular, the state-of-the-art StyleGAN models [20, 21, 19] have many practical applications such as image enhancement [46, 24, 6, 41], image editing [17, 35, 14, 37, 1, 43, 29], image-to-image translation [31, 16, 36, 11] thanks to their high-quality image generation and their latent representation that has rich semantics and disentangled controls for localized meaningful image manipulations. However, it comes at a price, as the training of StyleGAN requires a large, high-quality dataset that significantly limits its applicability because many real-world domains are represented by few images. The standard approach to deal with this problem is transfer learning, i.e. fine-tuning the model pretrained on the source domain $A$ to the target domain $B$ . + +There are many domain adaptation methods for StyleGAN [18, 39, 50, 52, 23, 45, 28, 31, 4, 11, 55, 44] that + +tackle this problem in different ways depending on the number of available images (e.g., one-shot/few-shot) from the target domain $B$ and the similarity between the source $A$ and target $B$ domains (e.g., faces $\rightarrow$ sketches, artistic portraits, or faces $\rightarrow$ dogs, cats). Most of these works implicitly assume that StyleGAN can be adapted to a new domain only if we fine-tune almost all its weights, even for similar domains. However, this common wisdom is poorly investigated and verified and there is a lack of analysis of which parts of StyleGAN are important depending on different data regimes and the similarity between domains. + +In this work, we aim to provide a systematic and comprehensive analysis of this question. Our investigation of the properties of the aligned StyleGAN models consists of several parts. First, in Section 3, we identify what parts of the StyleGAN are sufficient for its adaptation depending on the similarity between the source $A$ and target $B$ domains. We discover that fine-tuning the whole synthesis convolutional network is not always necessary. In the case of similar $A$ and $B$ domains, the affine layers are sufficient for the adaptation. For more distant domains, we should optimize more parameters, however not the whole network. It suggests investigating new more efficient and lightweight parameterizations of StyleGAN to utilize them for domain adaptation. + +In the second part of our analysis, we propose two new parameterizations of StyleGAN. For similar domains, we consider the latent space that is formed by the output of affine layers, i.e. StyleSpace [43]. We show that we can directly optimize directions in this space that can adapt to similar target domains with the same quality as fine-tuning all weights of StyleGAN (we call such directions as Style-Domain directions). Further, we explore that we can zero out $80\%$ of StyleDomain direction coordinates without a quality degradation that gives even more lightweight parameterization (StyleSpaceSparse). For more distant domains, we propose a new parameterization Affine+ that consists of affine layers and only one convolutional block from the synthesis network. It reduces the number of trainable parameters by 6 times and achieves the same quality. Then, we further improve Affine+ parameterization by utilizing low-rank decomposition for weights of affine layers and obtain AffineLight+ parameterization. It allows us to optimize by two orders less parameters compared to training the whole StyleGAN. These parameterizations show the state-of-the-art performance for few-shot adaptation for dissimilar domains outperforming more complicated and expensive baselines. + +Additionally, in Section 4, we inspect StyleDomain directions and discover their surprising properties. The first one is mixability, i.e. we can sum up these directions to obtain a new mixed domain (e.g., see Figure 6 as a mix of the Joker style, Pixar style and the style from the image). The + +second impressive property is transferability, i.e. the same StyleDomain directions can be applied to StyleGAN models that were fine-tuned to other domains (e.g., see Figure 5 where we apply directions found for faces to dogs, cats and churches). We apply these findings to standard computer vision tasks such as image-to-image translation and cross-domain morphing. + +# 2. Related Work + +Latent Spaces of StyleGAN. After recent remarkable success of GANs [12, 20, 21, 19, 5] in image synthesis, many works appeared that explore their latent representation for controllable image manipulation. In particular, the latent space of StyleGAN [20, 21, 19] has attracted considerable attention. It consists of three levels: (i) the first latent space, $\mathcal{Z}$ , is raw random noise (typically Gaussian); (ii) the intermediate latent spaces $\mathcal{W}, \mathcal{W} + [1]$ are formed by the output of the mapping network; (iii) the last level is StyleSpace, $\mathcal{S}$ , [43] that is spanned by the channel-wise style parameters after affine layers. It has been shown that these latent spaces have rich semantics [20, 21], and especially the StyleSpace that demonstrates the most disentangled and localized semantical directions [43]. In recent years many works have proposed to utilize such appealing properties of the StyleGAN latent spaces for image editing tasks [17, 35, 14, 37, 1, 43, 29]. To apply these methods for real images it is necessary to inverse them into one of the latent space of StyleGAN which is another task that draws significant attention [1, 21, 13, 30, 33, 38, 54, 53]. We should note that all mentioned methods for image manipulation by controlling StyleGAN latent space allow only in-domain editing. In this paper, we show that in StyleSpace there exist such directions that can change the domain of images. + +Domain Adaptation of StyleGAN. Recent years the problem of fine-tuning StyleGAN has generated a great deal of interest as it allows training the state-of-the-art generative model for a domain with few samples. There have appeared many works that tackle this problem in different ways depending on how similar the target domain is to the source one. Roughly, these methods can be divided into two groups. The first one deals with the case when the target and source domains are dissimilar (e.g. faces $\rightarrow$ cats, churches, etc.). In contrast, the second group considers the setting of similar domains (e.g. faces $\rightarrow$ stylized faces, painting faces, sketches, etc.). Methods from the first group typically require hundreds or thousands of samples from a new domain to adapt faithfully and they leverage data augmentations [18, 39, 50, 52], or freeze some layers of the discriminator to prevent overfitting [25], or train the discriminator with auxiliary losses to match the data more accurately [23, 45]. + +In the setting of the second group it is sufficient to have + +Figure 2: StyleGAN2 architecture. We introduce new latent space $S +$ for the for domain adaptation that combines StyleSpace and weight offsets for one block from the synthesis network. +![](images/b94d569f9f4895fc11e5443215419a22e84d554bb362d8d6924632304271b512.jpg) +Synthesis Network +Affine Layers +Mapping Network +Affine + ConvBlock +Weight Offsets + +only several samples (up to dozens) from the target domain for the successful adaptation and such approaches utilize another techniques. In particular, they introduce additional regularization terms [40, 22], preserve pairwise distances between instances in the source domain via cross-domain consistency loss [28], mix weights of the fine-tuned and the base generators [31], use an auxiliary small network to enhance the training [42]. More advanced methods [11, 55, 4] utilize the pretrained CLIP model [32] as the vision-language supervision for the text-based adaptation [11, 4] or the one-shot image-based adaptation [7, 55, 9, 49, 48, 4]. + +Recent work [51] analyzes the performance of methods from the second group for dissimilar domains in the low data regime. It was shown that those methods have a significant quality degradation in the few-shot regime. However, it can be partially mitigated by constraining model parameters. There are also approaches that introduce more lightweight parameterizations for domain adaptation [27, 34, 7, 4], however they work only for similar domains. In our paper, we propose efficient and highly lightweight parameterizations for both similar and dissimilar domains and show that they can achieve results on par with the state-of-the-art methods that optimize all weights of StyleGAN. + +There are few works that analyze the domain adaptation process of StyleGAN thoroughly. The paper [44] is the first attempt to perform such an in-depth study. In particular, they explore which parts of StyleGAN are mostly changed during fine-tuning and the transferability of the latent semantics after domain adaptation. However, this work does not analyze which parts of StyleGAN are sufficient for adapting depending on the similarity between source and target domains. In our paper, we provide a more systematic and comprehensive analysis that completes and improves + +the results from [44]. + +# 3. Importance of Each Part of the StyleGAN + +In this section, our goal is to analyze what parts of StyleGAN are important for domain adaptation. Similarly to previous works, we specifically focus on the state-of-the-art GAN architecture, StyleGAN2 [21]. For the source domain, we consider FFHQ [20] as it is the large high-quality dataset that is suitable for training StyleGAN2 from scratch. For the target one, we test a wide range of different domains that we will describe further. + +StyleGAN2 structure and its main components. We provide a diagram description of the StyleGAN2 architecture in Figure 2. It consists of three parts: + +- mapping network $f_{M}$ that transforms the input noise $z \in \mathcal{Z}$ (typically Gaussian) to the intermediate latent vector $w \in \mathcal{W}$ ; +- affine layers $f_1^A, \ldots, f_N^A$ , each of them takes as input the vector $w$ and maps it to corresponding style vector $s_1 = f_1^A(w), \ldots, s_N = f_N^A(w)$ . The concatenation of these vectors form the vector from the StyleSpace: $s = (s_1, \ldots, s_N) \in S$ ; +- synthesis network that is a composition of modulated convolutions. The weights of each convolution are modulated by the input style vector $s_i$ and applied to the input feature maps. The synthesis network also has tRGB convolutional layers that transform feature maps to RGB images and they also are modulated by corresponding style vectors. + +Accordingly, the StyleGAN2 $G_{\theta}$ generates from the input noise $z$ the output image $I = G_{\theta}(s(z))$ where $\theta$ are + +weights of the synthesis network, $s(z) = f^{A}(f_{M}(z))$ $f^{A} = \{f_{1}^{A},\dots ,f_{N}^{A}\}$ + +We will analyze these three components of StyleGAN2 and their impact on the domain adaptation process. It is common wisdom that the most important part for adaptation is the synthesis network, while the mapping network and affine layers are mostly responsible for the semantic manipulations within the source domain [44]. We aim to verify whether such a conception is correct. + +In our experiments, we additionally consider the impact of the combination of affine layers and one convolutional block from the synthesis network on domain adaptation. It is a way to probe the intermediate case between affine layers and the synthesis network. + +Method to analyze the impact of each component. The paper [44] has proposed to analyze the impact of each component by resetting its weights in the fine-tuned generator to their pretrained values and assessing the quality of the generated images. In our work, we propose another approach: to directly fine-tune only one of these components to explore which ones are sufficient for domain adaptation. + +Let us describe our method in more detail. The optimization process for the problem of domain adaptation is the following + +$$ +\mathcal {L} _ {D} \left(\left\{G _ {\theta} \left(s \left(z _ {i}\right)\right)\right\} _ {i = 1} ^ {K}\right)\rightarrow \min , \tag {1} +$$ + +where $\mathcal{L}_D$ is domain adaptation loss that depends on the domain $D$ (we discuss it further) and the samples from the generator $\{G_{\theta}(s(z_i))\}_{i = 1}^{K}$ , $z_{1},\ldots ,z_{K}\in \mathcal{Z}$ are sampled noises. + +Typically the generator $G_{\theta}$ is optimized with respect to all components, i.e. + +$$ +\mathcal {L} _ {D} \left(\left\{G _ {\theta} \left(s \left(z _ {i}\right)\right)\right\} _ {i = 1} ^ {K}\right)\rightarrow \min _ {\theta , f ^ {A}, f _ {M}}. \tag {2} +$$ + +We propose to investigate settings when we optimize with respect to only one these components. We denote each parameter space as: $\{\theta\} - \text{SymConv}$ parameterization, $\{f^A\} - \text{Affine}$ parameterization, $\{f_M\} - \text{Mapping}$ parameterization. The case we fine-tune all components of the StyleGAN2 we call Full parameterization. + +Domain adaptation settings. In our study, we consider two settings: one-shot and few-shot. For each data regime, we use different domains depending on their similarity to the source domain of realistic faces from FFHQ: + +- one-shot domains: for this setting, we consider only similar domains. It is the case when the target domain has the same geometry of faces and it preserves the identity of the person. It alters only the style of the image. In this regime, we consider not only one-shot image-based adaptation with the reference stylized face but additionally text-based adaptation with + +the text description of the target style (e.g., "Photo in the style of anime (pixar, sketch, etc.)"). See examples of such domains in Figure 3. + +- few-shot domains: for this regime, we examine more distant domains that have a face-like form but change the face geometry and identity in a stronger manner. As examples, we consider AFHQ dogs faces and cats faces [8]. + +Depending on the data regime, we use different domain loss function $\mathcal{L}_D$ . For one-shot domain adaptation, we apply the optimization loss from StyleGAN-NADA [11] in the case of text-based adaptation and another loss from DiFa [48] for one-shot image-based adaptation. For few-shot domain adaptation, we utilize the fine-tuning procedure from StyleGAN-ADA [18]. For more details about domain adaptation loss functions see Appendix A.1. + +To obtain quantitative comparisons in the case of a one-shot setting, we use Quality and Diversity metrics that were proposed in the HyperDomainNet paper [4]. For few-shot adaptation, we compute FID metric [15] using the standard protocol from [18]. + +Analysis for one-shot domains. For the analysis, we choose different text-based and one-shot image-based domains (see Appendix A.2 for the full list and more details). In experiments, we consider the four parameterizations (Full, SyntConv, Affine, Mapping) we discussed above. + +We provide qualitative results in Figure 3 with quantitative ones in Table 1. More results see in Appendix A.2. + +We observe that all three parameterizations, Full, Synt-Conv and Affine, show comparable performance in terms of both visual quality and objective metrics. The fact that the synthesis network is sufficient for similar domains was clear from the previous work [44]. However, our finding that the affine part is also sufficient is a new and surprising result. It means that we can change the domain of generated images without fine-tuning the synthesis network but just passing the modified style vector that comes from the affine part. Also, we observe that the mapping network shows poor vi + +![](images/70b2688821fb80c1baef86f054e04713612815f378a842658d551c365cdfbceb.jpg) +Figure 3: Text-based and image-based adaptation for different parameterizations. Affine, StyleSpace and StyleSpaceSparse parameterizations yield performance comparable with Full one. This style image is called "Disney". + +Table 1: Quality and Diversity metrics [4] for text-based and one-shot image-based domain adaptations with different parameterizations. Affine, StyleSpace and StyleSpaceSparse parameterizations achieve results comparable with the Full one. + +
Parameter SpaceSizeBoteroSketchDisney (image)Titan Erwin (image)
QualityDiversityQualityDiversityQualityDiversityQualityDiversity
Full30.3M0.3120.2280.2080.2960.7130.2470.7600.194
SyntConv23.6M0.3110.2240.1910.2920.7110.2590.7410.217
Affine4.6M0.2980.2210.1940.2960.5650.3590.6500.314
Mapping2.1M0.2260.1150.1820.1430.7170.0800.6450.102
StyleSpace6.0K0.3090.230.1930.3060.6270.3080.6720.296
StyleSpaceSparse1.2K0.3220.2130.2010.2690.6170.3040.6590.303
+ +sual quality and low diversity in the generated images when considering Diversity metric. It indicates that for successful adaptation it is important to update the style vector from $S$ rather than the intermediate latent vector from $\mathcal{W}$ space. + +Analysis for few-shot domains. For this setting, as we discussed above, we consider two datasets, AFHQ Dogs and Cats [8], and the results are provided in Figure 4 and in Table 2. See more results in Appendix A.3. + +We observe that results for Dogs and Cats are different compared to similar domains. In particular, we see that the Affine parameterization does not demonstrate the same quality as the Full parameterization. It can be seen from the degraded visual quality and visible gap in FID metric. However, it is still surprising that the generated images after adaptation have an adequate visual appearance considering that we do not fine-tune the synthesis network at all. Synt-Conv expectedly achieves results comparable with Full parameterization, and Mapping conversely shows poor quality on all datasets. + +Table 2: FID scores for domain adaptation with different parameterizations. We observe a significant gap between Affine and Full parametrizations that, however, can be drastically reduced by introducing Affine+ parameterization. + +
Parameter SpaceSizeDomains
DogCat
Full30.3M20.37.1
SyntConv23.6M19.77.2
Affine4.6M70.127.6
Mapping2.1M208.2226.1
Affine+5.1M18.67.0
AffineLight+0.6M20.68.9
StyleSpace6.0K75.822.0
+ +# 4. Efficient and lightweight parameterizations of StyleGAN + +StyleSpace and StyleSpaceSparse. Our findings from the previous section suggest that we can change the domain of generated images by modifying the style vector in the StyleSpace $S$ . To check this hypothesis, we will adapt StyleGAN2 by directly optimizing the direction in $S$ , i.e. during fine-tuning of StyleGAN we will optimize only $\Delta s^{D}$ : + +$$ +\mathcal {L} _ {D} \left(\left\{G _ {\theta} \left(s \left(z _ {i}\right) + \Delta s\right)\right\} _ {i = 1} ^ {K}\right)\rightarrow \min _ {\Delta s}, \tag {3} +$$ + +where $\Delta s = (\Delta s_1, \ldots, \Delta s_N) \in S$ is the optimized direction in the $S$ for adapting the generator $G_{\theta}$ to the domain $D$ . We call such directions $\Delta s$ as StyleDomain directions. + +Further, we explore that we can zero out most of the coordinates of StyleDomain directions without quality degradation. We use standard pruning technique when we leave $20\%$ of the largest absolute values in the StyleDomain vector and set the rest to zero. We call such parameterization as StyleSpaceSparse. We examine different pruning rates and its performance in Appendix A.4. + +We apply these parameterizations to one-shot and few-shot domains and obtain the following results. + +For one-shot adaptation, we provide results in Figure 3 and in Table 1 (see more results in Appendix A.4). We observe that optimizing the StyleDomain direction $\Delta s^{D}$ + +![](images/e375af9591b6bf33ac2871c650d23e5a2a6d0ca730a95a892e1cca6b72543254.jpg) +Figure 4: Domain adaptation for dissimilar domains. Affine+ parameterization produces results on par with the Full one. + +achieves the same results both visually and quantitatively as the Full parameterization. It is new and important observation that StyleSpace allows not only image editing within domain but also generating samples from out-of-domain of realistic human faces. + +For few-shot domains, we observe that StyleSpace is not sufficient, which is expressed by the same significant quality degradation as for Affine parameterization. Further, we introduce a new parameterization that is efficient for more distant domains. + +Affine+ and AffineLight+. We aim to improve Affine parameterization for successfully adapting to dissimilar domains such as Dogs and Cats. We propose to extend it by adding one block from synthesis layer with specified spatial resolution. Such block has two convolutional layers with weights $\theta_{1},\theta_{2}\in \mathbb{R}^{512\times 512\times 3\times 3}$ . Instead of fine-tuning all these weights we introduce more compact parameterization as offsets $\Delta \theta_{1},\Delta \theta_{2}$ to these weights that are same across spatial dimensions, i.e. $\Delta \theta_{1},\Delta \theta_{2}\in \mathbb{R}^{512\times 512\times 1\times 1}$ (we observe that such reduction in size does affect the performance). In addition to $\Delta \theta_{1},\Delta \theta_{2}$ we introduce the offsets to the weights $\theta^{tRGB}\in \mathbb{R}^{3\times 512\times 3\times 3}$ of the tRGB convolutional layer of the same block with similar parameterization $\Delta \theta^{tRGB}\in \mathbb{R}^{3\times 512\times 1\times 1}$ . Further, we omit such detail about tRGB part in the sake of brevity. So, for this parameterization the optimization procedure has the following form: + +$$ +\mathcal {L} _ {D} \left(\left\{G _ {\theta , \Delta \theta_ {1}, \Delta \theta_ {2}} (s (z _ {i})) \right\} _ {i = 1} ^ {K}\right)\rightarrow \min _ {\Delta \theta_ {1}, \Delta \theta_ {2}, f ^ {A}}, \tag {4} +$$ + +where $G_{\Delta \theta_1, \Delta \theta_2}$ is the generator with weight offsets $\Delta \theta_1, \Delta \theta_2$ for the one block from the synthesis network. + +We call such parameter space as Affine+. We examine all blocks of the synthesis network to choose for this parameterization. We end up with the block with $64 \times 64$ resolution + +![](images/270e298312ff9af55cc641e5b8ce4538179684e2365643b64795f4a0059c91c2.jpg) +Figure 5: StyleSpace directions transfer from text-based and image-based domain adaptation to other fine-tuned models. We can successfully transfer style while preserving image content. + +as it shows the best performance (see results of this analysis in Appendix A.4). + +While Affine+ has already had 6 times less parameters than Full parameterization, we further reduce its size by applying low-rank decomposition to the weights of affine layers. We denote this parameterization as AffineLight+. It gives us by two orders less parameters than Full parameterization with insignificant degradation in quality. We will show in Section 5 that in low data regimes, this parameterization achieves comparable performance as Affine+ and outperforms other baselines. More details about applied low-rank decomposition can be found in Appendix A.4. + +We apply these two parameterizations to few-shot domains and obtain the following results. + +We provide results in Figure 4 and in Table 2 (see more results in Appendix A.4). We see that Affine+ parameterization allows us removing the performance gap from the Full parameterization both qualitatively and quantitatively. While the number of parameters for the additional block in Affine+ accounts for only $2\%$ of the synthesis network size. It shows that the style vector allows adapting the generator even to more distant domains if we just add a small part of the synthesis network. + +We observe that AffineLight+ still shows adequate performance, except it has 100 times less parameters than the Full parameterization. We should notice that this parameterization is more suitable for low data regimes than those that we consider in Section 5. + +Properties of the StyleDomain directions. We investigate StyleDomain directions that adapt the generator to similar domains and explore surprising properties. The first one is mixability, i.e. we discover that StyleDomain directions can be combined with each other. In particular, we can consider several directions that correspond to different similar domains and take their sum. The resulting direction will adapt the generator to the semantically mixed domain. We provide different examples of such combinations in Figure 6 + +![](images/324ff89f310e7558ab38a9dd703d6d34e4c02a7157dee8ef664326322ba33bdf.jpg) +Figure 6: Example of mixing StyleDomain directions. We can combine different directions in order to perform adaptation into a semantically mixed domain. + +![](images/1a9d49fe13fa9cf1e4d34b2091a72926f452d3b3052213005f98683637b34bd5.jpg) +Figure 7: Comparison with baselines for one-shot image-based domain adaptation. StyleSpace and StyleSpaceSparse parameterizations achieve comparable quality as other methods while having much less trainable parameters. + +(see more results in Appendix A.5). + +The second property is transferability, i.e. we can transfer StyleDomain directions between different StyleGAN2 models. In particular, let us consider the base generator $G_{\theta}$ pretrained on realistic faces and the generator $G_{\theta}^{Dog}$ fine-tuned to domain of dogs in the Full parameterization. We verify that we can apply StyleDomain directions to $G_{\theta}^{Dog}$ which were optimized for $G_{\theta}$ . Specifically, we take Style-Domain directions that were optimized for $G_{\theta}$ to adapt it to different text-based and one-shot image-based domains (e.g. Pixar, Disney, etc.). Next, we apply these directions to generators that were fine-tuned from $G_{\theta}$ to other domains (e.g. Dogs, Cats). We provide results of this experiment in Figure 5 (see more results in Appendix A.5). + +Finally, we check that StyleDomain directions can be successfully combined with latent controls for image editing. For a more detailed exploration of this property, see in Appendix A.5. + +# 5. Experiments + +One-shot domain adaptation. For image-based one-shot adaptation, we consider the main baselines: Target-CLIP [7], JoJoGAN [9], MTG [55], GOSA [49], DiFa [48], DomMod [4]. We apply our parameterizations StyleSpace and StyleSpaceSparse to DiFa model because this configuration achieves the highest performance (experiments with other base models can be found in Appendix A.6). Similarly, we apply DomMod parameterization to the DiFa model. We use StyleGAN2 as the GAN architecture and FFHQ as the source domain. For fair comparison, we strictly follow baseline default configurations in the choice of source-target adaptation setups and hyper-parameters. As target domain we use a wide range of different style images from baseline works. The full list of target images can be found in Appendix A.6. + +We provide both quantitative and qualitative results in Table 3 and Figure 7, correspondingly. More results and + +Table 3: Quality and Diversity metrics [4] for one-shot image-based domain adaptations with different methods. Memory denotes the memory needed for keeping adapted generators for all 12 domains for each method. StyleSpace and StyleSpaceSparse parameterizations achieve results comparable to other baselines while having much less trainable parameters. + +
MethodSizeMemoryTitan ErwinDisneyAcross 12 domains
QualityDiversityQualityDiversityQualityDiversity
JoJoGAN [9]30M1.80GB0.5720.2920.5910.2600.590 ± 0.0480.257 ± 0.025
MTG [55]30M1.80GB0.6070.2690.5090.2340.586 ± 0.0540.263 ± 0.028
GOSA [49]30M1.80GB0.5470.2830.6170.2160.584 ± 0.0340.252 ± 0.030
DiFa [48]30M1.80GB0.7190.2260.6990.2630.734 ± 0.0470.215 ± 0.038
TargetCLIP [7]9.0K420KB0.4740.3060.5020.3330.491 ± 0.0430.322 ± 0.015
DomMod (DiFa) [4]6.0K280KB0.7050.2500.6250.2940.679 ± 0.0490.253 ± 0.037
StyleSpace (DiFa)6.0K280KB0.6720.2960.6270.3080.644 ± 0.0410.298 ± 0.025
StyleSpaceSparse (DiFa)1.2K56.4KB0.6590.3030.6170.3040.638 ± 0.0460.299 ± 0.026
+ +samples can be found in Appendix A.6. We observe that the DiFa model achieves the best Quality metric but has low Diversity compared to other methods. Our parameterizations applied to the DiFa model balance its performance in terms of these metrics. So, StyleSpace (DiFa) and StyleSpaceSparse (DiFa) achieves uniformly better results than JoJOGAN, MTG, GOSA baselines. DomMod parameterization also show good performance, but it is comparable with our StyleSpaceSparse parameterization that has 5 times less parameters. We also report the overall memory needed for keeping adapted generators for 12 domains. We observe that StyleSpaceSparse requires significantly less space than other models. It will be especially important if we scale the number of target domains to thousands and more. We note that TargetCLIP that also has small number of trainable parameters show very poor results visually and in terms of Quality metric. It has a high Diversity only because its generated images are very close to the original ones. To provide more comprehensive comparison we additionally conduct user studies and present results in Appendix A.6.4. + +Few-shot domain adaptation. As main baselines for this task, we consider the vanilla StyleGAN-ADA [18] (we will denote it as ADA), CDC [28] and recent SOTA method AdAM [51]. We compare our parameterizations Affine+ and AffineLight+ applied to ADA with these baselines on two datasets: Dogs and Cats [8]. We study the efficiency of these methods with different numbers of available target samples. For fair comparison, we rigorously follow the training setups of all baselines. We note that we use the same number of training iterations and the same batch size for all methods. We observe that in the AdAM paper [51] the number of training iterations is set to a small value (less than 10K) and it causes underfitting of the vanilla ADA method. Therefore, we increase this number for all methods to 50K and use the same batch size of 4. Then, for each method, we report the best FID value that it obtains during + +![](images/f5bea334163ffac3287ba57ac13af8234dba941b9aef2c153ab3feff2861680a.jpg) +Figure 8: Few-shot training results for different number of shots. Proposed ADA (Affine+) and ADA (AffineLight+) show uniformly better performance than baselines. + +Table 4: Results for few-shot training with 10-shots. Proposed ADA (Affine+) and ADA (AffineLight+) achieve better performance. + +
MethodSizeDomains (10-shots)
CatDog
CDC [28]30M66.24184.56
AdAM [51]19M47.05119.61
ADA (Full) [18]30M51.38100.25
ADA (Affine+)5.1M38.4096.38
ADA (AffineLight+)0.6M43.91101.31
+ +training. See more details about training in Appendix A.7. + +We report results for few-shot training for different numbers of shots in Figure 8 and separately for 10 shots in Table 4. More results and samples can be found in Appendix A.7. Firstly, we observe that the AdAM method shows results not better than the vanilla ADA (Full) when trained for sufficient number of iterations. Secondly, we see that ADA (Affine+) and ADA (AffineLight+) based on our parameterizations achieve better results uniformly for all numbers of shots. It shows that these parameterizations is especially suitable for low data regimes. + +Cross-domain image translation. We consider two setups of standard image-to-image problem. In the first one, we translate images from the source domain to the target domain unconditionally. In the second setup, we perform a reference-based translation, where the resulting image combines the pose and the shape of the source image with the style from a reference image. Details of these experiments can be found in Appendix A.9. + +Cross-domain image morphing. Cross-domain morphing is a smooth transition between two images from different domains. This task is known as challenging [3, 10] and it is successfully tackled in the work [44] using aligned StyleGAN2 models. The idea is to interpolate between aligned generator weights to obtain a smooth transition between domains. We propose more complex image morphing by utilizing the transferability of StyleDomain directions. For example, we can apply a direction that stands for Sketch or Pixar style to Dogs domain to obtain a smooth transition between sketchy and pixar-like dog (see Figure 1). See many examples of such complex cross-domain morphing in Appendix A.10. + +# 6. Conclusion + +In this paper, we provide an extensive analysis of the process of StyleGAN domain adaptation. We reveal the sufficient components of the generator for successful adaptation depending on the similarity between the source and target domains. We discover the ability of StyleSpace to change + +the domain of generated images by StyleDomain directions. We also propose new efficient parameterizations Affine+ and AffineLight+ for few-shot adaptation that outperform existing baselines. Further, we explore and leverage the properties of StyleDomain directions. We believe that our investigation can attract more attention to the exploration of new and interesting properties of StyleSpace. + +# 7. Acknowledgments + +The results of sections 1, 2, 3 were obtained by Aibek Alanov and Dmitry Vetrov with the support of the grant for research centers in the field of AI provided by the Analytical Center for the Government of the Russian Federation (ACRF) in accordance with the agreement on the provision of subsidies (identifier of the agreement 000000D730321P5Q0002) and the agreement with HSE University No. 70-2021-00139. This research was supported in part through computational resources of HPC facilities at HSE University. + +# References + +[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 1, 2 +[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 33 +[3] Kfir Aberman, Jing Liao, Mingyi Shi, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Neural best-buddies: Sparse cross-domain correspondence. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 8 +[4] Aibek Alanov, Vadim Titov, and Dmitry Vetrov. Hyperdomainnet: Universal domain adaptation for generative adversarial networks. arXiv preprint arXiv:2210.08884, 2022. 1, 3, 4, 5, 7, 13, 18, 23, 24, 34, 35, 36, 37 +[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1, 2 +[6] Kelvin CK Chan, Xintao Wang, Xiangyu Xu, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for large-factor image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14245-14254, 2021. 1 +[7] Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf. Image-based clip-guided essence transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 695-711. Springer, 2022. 3, 7, 34, 35 +[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vi + +sion and pattern recognition, pages 8188-8197, 2020. 4, 5, 8, 40 +[9] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 128-152. Springer, 2022. 3, 7, 34, 35 +[10] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, volume 39, pages 303–313. Wiley Online Library, 2020. 8 +[11] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 1, 3, 4, 12 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2 +[13] Shanyan Guan, Ying Tai, Bingbing Ni, Feida Zhu, Feiyue Huang, and Xiaokang Yang. Collaborative learning for faster stylegan embedding. arXiv preprint arXiv:2007.01758, 2020.2 +[14] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020. 1, 2 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4 +[16] Jialu Huang, Jing Liao, and Sam Kwong. Unsupervised image-to-image translation via pre-trained stylegan2 network. IEEE Transactions on Multimedia, 24:1435-1448, 2021. 1 +[17] Ali Jahanian, Lucy Chai, and Phillip Isola. "On the" steerability" of generative adversarial networks. arXiv preprint arXiv:1907.07171, 2019. 1, 2 +[18] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 1, 2, 4, 8, 12, 19, 40 +[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2 +[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3 +[21] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1, 2, 3, 12 + +[22] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3 +[23] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. In International Conference on Learning Representations, 2020. 1, 2 +[24] Xuan Luo, Xuaner Zhang, Paul Yoo, Ricardo Martin-Brualla, Jason Lawrence, and Steven M Seitz. Time-travel rephotography. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 1 +[25] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze the discriminator: a simple baseline for fine-tuning gans. arXiv preprint arXiv:2002.10964, 2020. 2 +[26] M-E Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1447-1454. IEEE, 2006. 19 +[27] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 3 +[28] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 1, 3, 8, 40 +[29] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 1, 2 +[30] Stanislav Pidhorskyi, Donald A Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14104-14113, 2020. 2 +[31] Justin NM Pinkney and Doron Adler. Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334, 2020. 1, 3, 19 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 12 +[33] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2 +[34] Esther Robb, Wen-Sheng Chu, Abhishek Kumar, and Jia-Bin Huang. Few-shot adaptation of generative adversarial networks. arXiv preprint arXiv:2010.11943, 2020. 3 +[35] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vi + +sion and pattern recognition, pages 9243-9252, 2020. 1, 2, 31 +[36] Guoxian Song, Linjie Luo, Jing Liu, Wan-Chun Ma, Chunpong Lai, Chuanxia Zheng, and Tat-Jen Cham. Agilegan: stylizing portraits by inversion-consistent transfer learning. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 1 +[37] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 1, 2 +[38] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 2 +[39] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021. 1, 2, 12 +[40] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7921-7931, 2021. 3 +[41] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9168-9178, 2021. 1 +[42] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. Minegan: effective knowledge transfer from gans to target domains with few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9332-9341, 2020. 3 +[43] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2, 31, 33 +[44] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 1, 3, 4, 8, 47 +[45] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. Advances in Neural Information Processing Systems, 34:9378-9390, 2021. 1, 2 +[46] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021. 1 +[47] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 19, 40 + +[48] Yabo Zhang, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, Wangmeng Zuo, et al. Towards diverse and faithful one-shot adaptation of generative adversarial networks. In Advances in Neural Information Processing Systems, 2022. 3, 4, 7, 12, 34, 35 +[49] Zicheng Zhang, Yinglu Liu, Congying Han, Tiande Guo, Ting Yao, and Tao Mei. Generalized one-shot domain adaption of generative adversarial networks. arXiv preprint arXiv:2209.03665, 2022. 3, 7, 34, 35 +[50] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in Neural Information Processing Systems, 33:7559-7570, 2020. 1, 2, 12 +[51] Yunqing Zhao, Keshigeyan Chandrasegaran, Milad Abdol-lahzadeh, and Ngai-Man Cheung. Few-shot image generation via adaptation-aware kernel modulation. arXiv preprint arXiv:2210.16559, 2022. 3, 8, 40 +[52] Zhengli Zhao, Zizhao Zhang, Ting Chen, Sameer Singh, and Han Zhang. Image augmentations for gan training. arXiv preprint arXiv:2006.02595, 2020. 1, 2, 12 +[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2 +[54] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In European conference on computer vision, pages 597-613. Springer, 2016. 2 +[55] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. arXiv preprint arXiv:2110.08398, 2021. 1, 3, 7, 34, 35 \ No newline at end of file diff --git a/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/images.zip b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b96e5c335aea176dcf460087de27c17d3955dba4 --- /dev/null +++ b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6a0b45ed7d0bf193d210d58f18ede4c6239494d0adec3b28d5b72b858050843 +size 655748 diff --git a/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/layout.json b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..30c33707e5d10bec609e731cc689dbead889d882 --- /dev/null +++ b/2023/StyleDomain_ Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation/layout.json @@ -0,0 +1,8115 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "text", + "content": "StyleDomain: Efficient and Lightweight Parameterizations of StyleGAN for One-shot and Few-shot Domain Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 160, + 478, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 160, + 478, + 175 + ], + "spans": [ + { + "bbox": [ + 115, + 160, + 478, + 175 + ], + "type": "text", + "content": "Aibek Alanov" + }, + { + "bbox": [ + 115, + 160, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{2,1,*}" + }, + { + "bbox": [ + 115, + 160, + 478, + 175 + ], + "type": "text", + "content": ", Vadim Titov" + }, + { + "bbox": [ + 115, + 160, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 115, + 160, + 478, + 175 + ], + "type": "text", + "content": ", Maksim Nakhodnov" + }, + { + "bbox": [ + 115, + 160, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 115, + 160, + 478, + 175 + ], + "type": "text", + "content": ", Dmitry Vetrov" + }, + { + "bbox": [ + 115, + 160, + 478, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 174, + 177, + 419, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 177, + 419, + 190 + ], + "spans": [ + { + "bbox": [ + 174, + 177, + 419, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 174, + 177, + 419, + 190 + ], + "type": "text", + "content": "HSE University " + }, + { + "bbox": [ + 174, + 177, + 419, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 174, + 177, + 419, + 190 + ], + "type": "text", + "content": "AIRI " + }, + { + "bbox": [ + 174, + 177, + 419, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 174, + 177, + 419, + 190 + ], + "type": "text", + "content": "Lomonosov Moscow State University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 191, + 476, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 191, + 476, + 203 + ], + "spans": [ + { + "bbox": [ + 117, + 191, + 476, + 203 + ], + "type": "text", + "content": "aalanov@hse.ru, titov@2a2i.org, nakhodnov@2a2i.org, dvetrov@hse.ru" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 47, + 224, + 296, + 352 + ], + "blocks": [ + { + "bbox": [ + 47, + 224, + 296, + 352 + ], + "lines": [ + { + "bbox": [ + 47, + 224, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 296, + 352 + ], + "type": "image", + "image_path": "703fc3be0760b37e479d897a4a507ca0ff9c997fe1607ba5cbf88db2032d4c9b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 361, + 547, + 398 + ], + "lines": [ + { + "bbox": [ + 45, + 361, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 45, + 361, + 547, + 398 + ], + "type": "text", + "content": "Figure 1: Smooth cross-domain image morphing. Morphing in the space of generator weights could be successfully combined with morphing using StyleDomain directions, i.e. directions in the StyleSpace that can adapt the generator to new domains (see Section 4)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 296, + 224, + 545, + 351 + ], + "blocks": [ + { + "bbox": [ + 296, + 224, + 545, + 351 + ], + "lines": [ + { + "bbox": [ + 296, + 224, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 296, + 224, + 545, + 351 + ], + "type": "image", + "image_path": "295ca5be4e5e5a987fdcfb3a2a6f80cd846f89d49c53320019adac59146680a2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 413, + 192, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 413, + 192, + 426 + ], + "spans": [ + { + "bbox": [ + 143, + 413, + 192, + 426 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 440, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 440, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 440, + 289, + 693 + ], + "type": "text", + "content": "Domain adaptation of GANs is a problem of fine-tuning GAN models pretrained on a large dataset (e.g. StyleGAN) to a specific domain with few samples (e.g. painting faces, sketches, etc.). While there are many methods that tackle this problem in different ways, there are still many important questions that remain unanswered. In this paper, we provide a systematic and in-depth analysis of the domain adaptation problem of GANs, focusing on the StyleGAN model. We perform a detailed exploration of the most important parts of StyleGAN that are responsible for adapting the generator to a new domain depending on the similarity between the source and target domains. As a result of this study, we propose new efficient and lightweight parameterizations of StyleGAN for domain adaptation. Particularly, we show that there exist directions in StyleSpace (StyleDomain directions) that are sufficient for adapting to similar domains. For dissimilar domains, we propose Affine+ and AffineLight+ parameterizations that allows us to outperform existing baselines in few-shot adaptation while having significantly less training parameters. Finally, we examine StyleDomain directions and discover their many sur" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 415, + 547, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 547, + 451 + ], + "type": "text", + "content": "prising properties that we apply for domain mixing and cross-domain image morphing. Source code can be found at https://github.com/AIRI-Institute/StyleDomain." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 476, + 387, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 387, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 387, + 489 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "type": "text", + "content": "Recent years GANs [12, 20, 21, 19, 5] have shown impressive results in image synthesis and offered many ways to control the generated data. In particular, the state-of-the-art StyleGAN models [20, 21, 19] have many practical applications such as image enhancement [46, 24, 6, 41], image editing [17, 35, 14, 37, 1, 43, 29], image-to-image translation [31, 16, 36, 11] thanks to their high-quality image generation and their latent representation that has rich semantics and disentangled controls for localized meaningful image manipulations. However, it comes at a price, as the training of StyleGAN requires a large, high-quality dataset that significantly limits its applicability because many real-world domains are represented by few images. The standard approach to deal with this problem is transfer learning, i.e. fine-tuning the model pretrained on the source domain " + }, + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "type": "text", + "content": " to the target domain " + }, + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 498, + 547, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 690, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 690, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 690, + 547, + 715 + ], + "type": "text", + "content": "There are many domain adaptation methods for StyleGAN [18, 39, 50, 52, 23, 45, 28, 31, 4, 11, 55, 44] that" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 704, + 125, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 704, + 125, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 704, + 125, + 714 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2184" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "tackle this problem in different ways depending on the number of available images (e.g., one-shot/few-shot) from the target domain " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " and the similarity between the source " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " domains (e.g., faces " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " sketches, artistic portraits, or faces " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " dogs, cats). Most of these works implicitly assume that StyleGAN can be adapted to a new domain only if we fine-tune almost all its weights, even for similar domains. However, this common wisdom is poorly investigated and verified and there is a lack of analysis of which parts of StyleGAN are important depending on different data regimes and the similarity between domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "type": "text", + "content": "In this work, we aim to provide a systematic and comprehensive analysis of this question. Our investigation of the properties of the aligned StyleGAN models consists of several parts. First, in Section 3, we identify what parts of the StyleGAN are sufficient for its adaptation depending on the similarity between the source " + }, + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "type": "text", + "content": " domains. We discover that fine-tuning the whole synthesis convolutional network is not always necessary. In the case of similar " + }, + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 206, + 289, + 374 + ], + "type": "text", + "content": " domains, the affine layers are sufficient for the adaptation. For more distant domains, we should optimize more parameters, however not the whole network. It suggests investigating new more efficient and lightweight parameterizations of StyleGAN to utilize them for domain adaptation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 376, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 651 + ], + "type": "text", + "content": "In the second part of our analysis, we propose two new parameterizations of StyleGAN. For similar domains, we consider the latent space that is formed by the output of affine layers, i.e. StyleSpace [43]. We show that we can directly optimize directions in this space that can adapt to similar target domains with the same quality as fine-tuning all weights of StyleGAN (we call such directions as Style-Domain directions). Further, we explore that we can zero out " + }, + { + "bbox": [ + 46, + 376, + 289, + 651 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 376, + 289, + 651 + ], + "type": "text", + "content": " of StyleDomain direction coordinates without a quality degradation that gives even more lightweight parameterization (StyleSpaceSparse). For more distant domains, we propose a new parameterization Affine+ that consists of affine layers and only one convolutional block from the synthesis network. It reduces the number of trainable parameters by 6 times and achieves the same quality. Then, we further improve Affine+ parameterization by utilizing low-rank decomposition for weights of affine layers and obtain AffineLight+ parameterization. It allows us to optimize by two orders less parameters compared to training the whole StyleGAN. These parameterizations show the state-of-the-art performance for few-shot adaptation for dissimilar domains outperforming more complicated and expensive baselines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "Additionally, in Section 4, we inspect StyleDomain directions and discover their surprising properties. The first one is mixability, i.e. we can sum up these directions to obtain a new mixed domain (e.g., see Figure 6 as a mix of the Joker style, Pixar style and the style from the image). The" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 157 + ], + "type": "text", + "content": "second impressive property is transferability, i.e. the same StyleDomain directions can be applied to StyleGAN models that were fine-tuned to other domains (e.g., see Figure 5 where we apply directions found for faces to dogs, cats and churches). We apply these findings to standard computer vision tasks such as image-to-image translation and cross-domain morphing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 167, + 392, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 392, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 392, + 179 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 186, + 547, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 547, + 485 + ], + "type": "text", + "content": "Latent Spaces of StyleGAN. After recent remarkable success of GANs [12, 20, 21, 19, 5] in image synthesis, many works appeared that explore their latent representation for controllable image manipulation. In particular, the latent space of StyleGAN [20, 21, 19] has attracted considerable attention. It consists of three levels: (i) the first latent space, " + }, + { + "bbox": [ + 304, + 186, + 547, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 186, + 547, + 485 + ], + "type": "text", + "content": ", is raw random noise (typically Gaussian); (ii) the intermediate latent spaces " + }, + { + "bbox": [ + 304, + 186, + 547, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{W}, \\mathcal{W} + [1]" + }, + { + "bbox": [ + 304, + 186, + 547, + 485 + ], + "type": "text", + "content": " are formed by the output of the mapping network; (iii) the last level is StyleSpace, " + }, + { + "bbox": [ + 304, + 186, + 547, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 186, + 547, + 485 + ], + "type": "text", + "content": ", [43] that is spanned by the channel-wise style parameters after affine layers. It has been shown that these latent spaces have rich semantics [20, 21], and especially the StyleSpace that demonstrates the most disentangled and localized semantical directions [43]. In recent years many works have proposed to utilize such appealing properties of the StyleGAN latent spaces for image editing tasks [17, 35, 14, 37, 1, 43, 29]. To apply these methods for real images it is necessary to inverse them into one of the latent space of StyleGAN which is another task that draws significant attention [1, 21, 13, 30, 33, 38, 54, 53]. We should note that all mentioned methods for image manipulation by controlling StyleGAN latent space allow only in-domain editing. In this paper, we show that in StyleSpace there exist such directions that can change the domain of images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 486, + 548, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 548, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 548, + 700 + ], + "type": "text", + "content": "Domain Adaptation of StyleGAN. Recent years the problem of fine-tuning StyleGAN has generated a great deal of interest as it allows training the state-of-the-art generative model for a domain with few samples. There have appeared many works that tackle this problem in different ways depending on how similar the target domain is to the source one. Roughly, these methods can be divided into two groups. The first one deals with the case when the target and source domains are dissimilar (e.g. faces " + }, + { + "bbox": [ + 304, + 486, + 548, + 700 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 486, + 548, + 700 + ], + "type": "text", + "content": " cats, churches, etc.). In contrast, the second group considers the setting of similar domains (e.g. faces " + }, + { + "bbox": [ + 304, + 486, + 548, + 700 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 486, + 548, + 700 + ], + "type": "text", + "content": " stylized faces, painting faces, sketches, etc.). Methods from the first group typically require hundreds or thousands of samples from a new domain to adapt faithfully and they leverage data augmentations [18, 39, 50, 52], or freeze some layers of the discriminator to prevent overfitting [25], or train the discriminator with auxiliary losses to match the data more accurately [23, 45]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "In the setting of the second group it is sufficient to have" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2185" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 410, + 270 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 410, + 270 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 410, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 410, + 270 + ], + "type": "image", + "image_path": "b94d569f9f4895fc11e5443215419a22e84d554bb362d8d6924632304271b512.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 76, + 518, + 91 + ], + "lines": [ + { + "bbox": [ + 425, + 76, + 518, + 91 + ], + "spans": [ + { + "bbox": [ + 425, + 76, + 518, + 91 + ], + "type": "text", + "content": "Synthesis Network" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 440, + 101, + 504, + 114 + ], + "lines": [ + { + "bbox": [ + 440, + 101, + 504, + 114 + ], + "spans": [ + { + "bbox": [ + 440, + 101, + 504, + 114 + ], + "type": "text", + "content": "Affine Layers" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 429, + 125, + 514, + 137 + ], + "lines": [ + { + "bbox": [ + 429, + 125, + 514, + 137 + ], + "spans": [ + { + "bbox": [ + 429, + 125, + 514, + 137 + ], + "type": "text", + "content": "Mapping Network" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 423, + 146, + 516, + 157 + ], + "lines": [ + { + "bbox": [ + 423, + 146, + 516, + 157 + ], + "spans": [ + { + "bbox": [ + 423, + 146, + 516, + 157 + ], + "type": "text", + "content": "Affine + ConvBlock" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 436, + 168, + 507, + 180 + ], + "lines": [ + { + "bbox": [ + 436, + 168, + 507, + 180 + ], + "spans": [ + { + "bbox": [ + 436, + 168, + 507, + 180 + ], + "type": "text", + "content": "Weight Offsets" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 280, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 280, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 545, + 304 + ], + "type": "text", + "content": "Figure 2: StyleGAN2 architecture. We introduce new latent space " + }, + { + "bbox": [ + 46, + 280, + 545, + 304 + ], + "type": "inline_equation", + "content": "S +" + }, + { + "bbox": [ + 46, + 280, + 545, + 304 + ], + "type": "text", + "content": " for the for domain adaptation that combines StyleSpace and weight offsets for one block from the synthesis network." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 443 + ], + "type": "text", + "content": "only several samples (up to dozens) from the target domain for the successful adaptation and such approaches utilize another techniques. In particular, they introduce additional regularization terms [40, 22], preserve pairwise distances between instances in the source domain via cross-domain consistency loss [28], mix weights of the fine-tuned and the base generators [31], use an auxiliary small network to enhance the training [42]. More advanced methods [11, 55, 4] utilize the pretrained CLIP model [32] as the vision-language supervision for the text-based adaptation [11, 4] or the one-shot image-based adaptation [7, 55, 9, 49, 48, 4]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": "Recent work [51] analyzes the performance of methods from the second group for dissimilar domains in the low data regime. It was shown that those methods have a significant quality degradation in the few-shot regime. However, it can be partially mitigated by constraining model parameters. There are also approaches that introduce more lightweight parameterizations for domain adaptation [27, 34, 7, 4], however they work only for similar domains. In our paper, we propose efficient and highly lightweight parameterizations for both similar and dissimilar domains and show that they can achieve results on par with the state-of-the-art methods that optimize all weights of StyleGAN." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "There are few works that analyze the domain adaptation process of StyleGAN thoroughly. The paper [44] is the first attempt to perform such an in-depth study. In particular, they explore which parts of StyleGAN are mostly changed during fine-tuning and the transferability of the latent semantics after domain adaptation. However, this work does not analyze which parts of StyleGAN are sufficient for adapting depending on the similarity between source and target domains. In our paper, we provide a more systematic and comprehensive analysis that completes and improves" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 312, + 392, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 392, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 392, + 323 + ], + "type": "text", + "content": "the results from [44]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 333, + 534, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 534, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 534, + 346 + ], + "type": "text", + "content": "3. Importance of Each Part of the StyleGAN" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 353, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 448 + ], + "type": "text", + "content": "In this section, our goal is to analyze what parts of StyleGAN are important for domain adaptation. Similarly to previous works, we specifically focus on the state-of-the-art GAN architecture, StyleGAN2 [21]. For the source domain, we consider FFHQ [20] as it is the large high-quality dataset that is suitable for training StyleGAN2 from scratch. For the target one, we test a wide range of different domains that we will describe further." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 485 + ], + "type": "text", + "content": "StyleGAN2 structure and its main components. We provide a diagram description of the StyleGAN2 architecture in Figure 2. It consists of three parts:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 491, + 545, + 684 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 317, + 491, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 545, + 526 + ], + "type": "text", + "content": "- mapping network " + }, + { + "bbox": [ + 317, + 491, + 545, + 526 + ], + "type": "inline_equation", + "content": "f_{M}" + }, + { + "bbox": [ + 317, + 491, + 545, + 526 + ], + "type": "text", + "content": " that transforms the input noise " + }, + { + "bbox": [ + 317, + 491, + 545, + 526 + ], + "type": "inline_equation", + "content": "z \\in \\mathcal{Z}" + }, + { + "bbox": [ + 317, + 491, + 545, + 526 + ], + "type": "text", + "content": " (typically Gaussian) to the intermediate latent vector " + }, + { + "bbox": [ + 317, + 491, + 545, + 526 + ], + "type": "inline_equation", + "content": "w \\in \\mathcal{W}" + }, + { + "bbox": [ + 317, + 491, + 545, + 526 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "type": "text", + "content": "- affine layers " + }, + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "type": "inline_equation", + "content": "f_1^A, \\ldots, f_N^A" + }, + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "type": "text", + "content": ", each of them takes as input the vector " + }, + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "type": "text", + "content": " and maps it to corresponding style vector " + }, + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "type": "inline_equation", + "content": "s_1 = f_1^A(w), \\ldots, s_N = f_N^A(w)" + }, + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "type": "text", + "content": ". The concatenation of these vectors form the vector from the StyleSpace: " + }, + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "type": "inline_equation", + "content": "s = (s_1, \\ldots, s_N) \\in S" + }, + { + "bbox": [ + 317, + 533, + 545, + 594 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 600, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 600, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 317, + 600, + 545, + 684 + ], + "type": "text", + "content": "- synthesis network that is a composition of modulated convolutions. The weights of each convolution are modulated by the input style vector " + }, + { + "bbox": [ + 317, + 600, + 545, + 684 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 317, + 600, + 545, + 684 + ], + "type": "text", + "content": " and applied to the input feature maps. The synthesis network also has tRGB convolutional layers that transform feature maps to RGB images and they also are modulated by corresponding style vectors." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "Accordingly, the StyleGAN2 " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " generates from the input noise " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " the output image " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "I = G_{\\theta}(s(z))" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " are" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2186" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": "weights of the synthesis network, " + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "inline_equation", + "content": "s(z) = f^{A}(f_{M}(z))" + }, + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "inline_equation", + "content": "f^{A} = \\{f_{1}^{A},\\dots ,f_{N}^{A}\\}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 179 + ], + "type": "text", + "content": "We will analyze these three components of StyleGAN2 and their impact on the domain adaptation process. It is common wisdom that the most important part for adaptation is the synthesis network, while the mapping network and affine layers are mostly responsible for the semantic manipulations within the source domain [44]. We aim to verify whether such a conception is correct." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 286, + 239 + ], + "type": "text", + "content": "In our experiments, we additionally consider the impact of the combination of affine layers and one convolutional block from the synthesis network on domain adaptation. It is a way to probe the intermediate case between affine layers and the synthesis network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 239, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 323 + ], + "type": "text", + "content": "Method to analyze the impact of each component. The paper [44] has proposed to analyze the impact of each component by resetting its weights in the fine-tuned generator to their pretrained values and assessing the quality of the generated images. In our work, we propose another approach: to directly fine-tune only one of these components to explore which ones are sufficient for domain adaptation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 323, + 286, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 286, + 358 + ], + "type": "text", + "content": "Let us describe our method in more detail. The optimization process for the problem of domain adaptation is the following" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 367, + 286, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 367, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 102, + 367, + 286, + 387 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta} \\left(s \\left(z _ {i}\\right)\\right)\\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min , \\tag {1}", + "image_path": "974d065aa0b5fdba8af85eade829c8332210dc5f28a93541fa5471e81e8bef00.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "type": "text", + "content": " is domain adaptation loss that depends on the domain " + }, + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "type": "text", + "content": " (we discuss it further) and the samples from the generator " + }, + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "type": "inline_equation", + "content": "\\{G_{\\theta}(s(z_i))\\}_{i = 1}^{K}" + }, + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "type": "inline_equation", + "content": "z_{1},\\ldots ,z_{K}\\in \\mathcal{Z}" + }, + { + "bbox": [ + 46, + 394, + 286, + 441 + ], + "type": "text", + "content": " are sampled noises." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 441, + 286, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 286, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 286, + 465 + ], + "type": "text", + "content": "Typically the generator " + }, + { + "bbox": [ + 46, + 441, + 286, + 465 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 441, + 286, + 465 + ], + "type": "text", + "content": " is optimized with respect to all components, i.e." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 473, + 286, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 473, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 94, + 473, + 286, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta} \\left(s \\left(z _ {i}\\right)\\right)\\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min _ {\\theta , f ^ {A}, f _ {M}}. \\tag {2}", + "image_path": "91daf58edfc8820b6259b1c0ec680d91ab42a0a208e96d3eaf7287fe04cc06ef.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 503, + 286, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 286, + 574 + ], + "type": "text", + "content": "We propose to investigate settings when we optimize with respect to only one these components. We denote each parameter space as: " + }, + { + "bbox": [ + 47, + 503, + 286, + 574 + ], + "type": "inline_equation", + "content": "\\{\\theta\\} - \\text{SymConv}" + }, + { + "bbox": [ + 47, + 503, + 286, + 574 + ], + "type": "text", + "content": " parameterization, " + }, + { + "bbox": [ + 47, + 503, + 286, + 574 + ], + "type": "inline_equation", + "content": "\\{f^A\\} - \\text{Affine}" + }, + { + "bbox": [ + 47, + 503, + 286, + 574 + ], + "type": "text", + "content": " parameterization, " + }, + { + "bbox": [ + 47, + 503, + 286, + 574 + ], + "type": "inline_equation", + "content": "\\{f_M\\} - \\text{Mapping}" + }, + { + "bbox": [ + 47, + 503, + 286, + 574 + ], + "type": "text", + "content": " parameterization. The case we fine-tune all components of the StyleGAN2 we call Full parameterization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 574, + 286, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 286, + 622 + ], + "type": "text", + "content": "Domain adaptation settings. In our study, we consider two settings: one-shot and few-shot. For each data regime, we use different domains depending on their similarity to the source domain of realistic faces from FFHQ:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 629, + 286, + 713 + ], + "type": "text", + "content": "- one-shot domains: for this setting, we consider only similar domains. It is the case when the target domain has the same geometry of faces and it preserves the identity of the person. It alters only the style of the image. In this regime, we consider not only one-shot image-based adaptation with the reference stylized face but additionally text-based adaptation with" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 324, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 108 + ], + "type": "text", + "content": "the text description of the target style (e.g., \"Photo in the style of anime (pixar, sketch, etc.)\"). See examples of such domains in Figure 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 117, + 545, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 117, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 317, + 117, + 545, + 176 + ], + "type": "text", + "content": "- few-shot domains: for this regime, we examine more distant domains that have a face-like form but change the face geometry and identity in a stronger manner. As examples, we consider AFHQ dogs faces and cats faces [8]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 185, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 281 + ], + "type": "text", + "content": "Depending on the data regime, we use different domain loss function " + }, + { + "bbox": [ + 304, + 185, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 304, + 185, + 545, + 281 + ], + "type": "text", + "content": ". For one-shot domain adaptation, we apply the optimization loss from StyleGAN-NADA [11] in the case of text-based adaptation and another loss from DiFa [48] for one-shot image-based adaptation. For few-shot domain adaptation, we utilize the fine-tuning procedure from StyleGAN-ADA [18]. For more details about domain adaptation loss functions see Appendix A.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "type": "text", + "content": "To obtain quantitative comparisons in the case of a one-shot setting, we use Quality and Diversity metrics that were proposed in the HyperDomainNet paper [4]. For few-shot adaptation, we compute FID metric [15] using the standard protocol from [18]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 341, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 411 + ], + "type": "text", + "content": "Analysis for one-shot domains. For the analysis, we choose different text-based and one-shot image-based domains (see Appendix A.2 for the full list and more details). In experiments, we consider the four parameterizations (Full, SyntConv, Affine, Mapping) we discussed above." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 413, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 437 + ], + "type": "text", + "content": "We provide qualitative results in Figure 3 with quantitative ones in Table 1. More results see in Appendix A.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": "We observe that all three parameterizations, Full, Synt-Conv and Affine, show comparable performance in terms of both visual quality and objective metrics. The fact that the synthesis network is sufficient for similar domains was clear from the previous work [44]. However, our finding that the affine part is also sufficient is a new and surprising result. It means that we can change the domain of generated images without fine-tuning the synthesis network but just passing the modified style vector that comes from the affine part. Also, we observe that the mapping network shows poor vi" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 309, + 576, + 555, + 652 + ], + "blocks": [ + { + "bbox": [ + 309, + 576, + 555, + 652 + ], + "lines": [ + { + "bbox": [ + 309, + 576, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 309, + 576, + 555, + 652 + ], + "type": "image", + "image_path": "70b2688821fb80c1baef86f054e04713612815f378a842658d551c365cdfbceb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 662, + 545, + 710 + ], + "lines": [ + { + "bbox": [ + 304, + 662, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 545, + 710 + ], + "type": "text", + "content": "Figure 3: Text-based and image-based adaptation for different parameterizations. Affine, StyleSpace and StyleSpaceSparse parameterizations yield performance comparable with Full one. This style image is called \"Disney\"." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2187" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 103, + 547, + 228 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 95 + ], + "type": "text", + "content": "Table 1: Quality and Diversity metrics [4] for text-based and one-shot image-based domain adaptations with different parameterizations. Affine, StyleSpace and StyleSpaceSparse parameterizations achieve results comparable with the Full one." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 103, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 103, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 547, + 228 + ], + "type": "table", + "html": "
Parameter SpaceSizeBoteroSketchDisney (image)Titan Erwin (image)
QualityDiversityQualityDiversityQualityDiversityQualityDiversity
Full30.3M0.3120.2280.2080.2960.7130.2470.7600.194
SyntConv23.6M0.3110.2240.1910.2920.7110.2590.7410.217
Affine4.6M0.2980.2210.1940.2960.5650.3590.6500.314
Mapping2.1M0.2260.1150.1820.1430.7170.0800.6450.102
StyleSpace6.0K0.3090.230.1930.3060.6270.3080.6720.296
StyleSpaceSparse1.2K0.3220.2130.2010.2690.6170.3040.6590.303
", + "image_path": "dd4e903860fd019df2f5431ac17e7cc8d77284a3184946a721c5a9fe5290c7d5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 247, + 287, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 295 + ], + "type": "text", + "content": "sual quality and low diversity in the generated images when considering Diversity metric. It indicates that for successful adaptation it is important to update the style vector from " + }, + { + "bbox": [ + 46, + 247, + 287, + 295 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 247, + 287, + 295 + ], + "type": "text", + "content": " rather than the intermediate latent vector from " + }, + { + "bbox": [ + 46, + 247, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 247, + 287, + 295 + ], + "type": "text", + "content": " space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 354 + ], + "type": "text", + "content": "Analysis for few-shot domains. For this setting, as we discussed above, we consider two datasets, AFHQ Dogs and Cats [8], and the results are provided in Figure 4 and in Table 2. See more results in Appendix A.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 364, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 495 + ], + "type": "text", + "content": "We observe that results for Dogs and Cats are different compared to similar domains. In particular, we see that the Affine parameterization does not demonstrate the same quality as the Full parameterization. It can be seen from the degraded visual quality and visible gap in FID metric. However, it is still surprising that the generated images after adaptation have an adequate visual appearance considering that we do not fine-tune the synthesis network at all. Synt-Conv expectedly achieves results comparable with Full parameterization, and Mapping conversely shows poor quality on all datasets." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 72, + 590, + 261, + 726 + ], + "blocks": [ + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 581 + ], + "type": "text", + "content": "Table 2: FID scores for domain adaptation with different parameterizations. We observe a significant gap between Affine and Full parametrizations that, however, can be drastically reduced by introducing Affine+ parameterization." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 590, + 261, + 726 + ], + "lines": [ + { + "bbox": [ + 72, + 590, + 261, + 726 + ], + "spans": [ + { + "bbox": [ + 72, + 590, + 261, + 726 + ], + "type": "table", + "html": "
Parameter SpaceSizeDomains
DogCat
Full30.3M20.37.1
SyntConv23.6M19.77.2
Affine4.6M70.127.6
Mapping2.1M208.2226.1
Affine+5.1M18.67.0
AffineLight+0.6M20.68.9
StyleSpace6.0K75.822.0
", + "image_path": "08f2b42e7a34e19e87b3c07701cc48340e44300a3c2deddf8c39640c4a3d28f0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 246, + 545, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 545, + 274 + ], + "type": "text", + "content": "4. Efficient and lightweight parameterizations of StyleGAN" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 280, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 363 + ], + "type": "text", + "content": "StyleSpace and StyleSpaceSparse. Our findings from the previous section suggest that we can change the domain of generated images by modifying the style vector in the StyleSpace " + }, + { + "bbox": [ + 304, + 280, + 545, + 363 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 280, + 545, + 363 + ], + "type": "text", + "content": ". To check this hypothesis, we will adapt StyleGAN2 by directly optimizing the direction in " + }, + { + "bbox": [ + 304, + 280, + 545, + 363 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 280, + 545, + 363 + ], + "type": "text", + "content": ", i.e. during fine-tuning of StyleGAN we will optimize only " + }, + { + "bbox": [ + 304, + 280, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\Delta s^{D}" + }, + { + "bbox": [ + 304, + 280, + 545, + 363 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 368, + 545, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 368, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 348, + 368, + 545, + 389 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta} \\left(s \\left(z _ {i}\\right) + \\Delta s\\right)\\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min _ {\\Delta s}, \\tag {3}", + "image_path": "43397762a6073fb268b7db193d15dc23e9aacf8d2f513ccc13e5ffd20fe4be78.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\Delta s = (\\Delta s_1, \\ldots, \\Delta s_N) \\in S" + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": " is the optimized direction in the " + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": " for adapting the generator " + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": " to the domain " + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": ". We call such directions " + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": " as StyleDomain directions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 432, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 516 + ], + "type": "text", + "content": "Further, we explore that we can zero out most of the coordinates of StyleDomain directions without quality degradation. We use standard pruning technique when we leave " + }, + { + "bbox": [ + 304, + 432, + 545, + 516 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 432, + 545, + 516 + ], + "type": "text", + "content": " of the largest absolute values in the StyleDomain vector and set the rest to zero. We call such parameterization as StyleSpaceSparse. We examine different pruning rates and its performance in Appendix A.4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 516, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 539 + ], + "type": "text", + "content": "We apply these parameterizations to one-shot and few-shot domains and obtain the following results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "text", + "content": "For one-shot adaptation, we provide results in Figure 3 and in Table 1 (see more results in Appendix A.4). We observe that optimizing the StyleDomain direction " + }, + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\Delta s^{D}" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 591, + 558, + 666 + ], + "blocks": [ + { + "bbox": [ + 306, + 591, + 558, + 666 + ], + "lines": [ + { + "bbox": [ + 306, + 591, + 558, + 666 + ], + "spans": [ + { + "bbox": [ + 306, + 591, + 558, + 666 + ], + "type": "image", + "image_path": "e375af9591b6bf33ac2871c650d23e5a2a6d0ca730a95a892e1cca6b72543254.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 673, + 545, + 709 + ], + "lines": [ + { + "bbox": [ + 304, + 673, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 673, + 545, + 709 + ], + "type": "text", + "content": "Figure 4: Domain adaptation for dissimilar domains. Affine+ parameterization produces results on par with the Full one." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2188" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "achieves the same results both visually and quantitatively as the Full parameterization. It is new and important observation that StyleSpace allows not only image editing within domain but also generating samples from out-of-domain of realistic human faces." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "text", + "content": "For few-shot domains, we observe that StyleSpace is not sufficient, which is expressed by the same significant quality degradation as for Affine parameterization. Further, we introduce a new parameterization that is efficient for more distant domains." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "text", + "content": "Affine+ and AffineLight+. We aim to improve Affine parameterization for successfully adapting to dissimilar domains such as Dogs and Cats. We propose to extend it by adding one block from synthesis layer with specified spatial resolution. Such block has two convolutional layers with weights " + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\theta_{1},\\theta_{2}\\in \\mathbb{R}^{512\\times 512\\times 3\\times 3}" + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "text", + "content": ". Instead of fine-tuning all these weights we introduce more compact parameterization as offsets " + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{1},\\Delta \\theta_{2}" + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "text", + "content": " to these weights that are same across spatial dimensions, i.e. " + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{1},\\Delta \\theta_{2}\\in \\mathbb{R}^{512\\times 512\\times 1\\times 1}" + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "text", + "content": " (we observe that such reduction in size does affect the performance). In addition to " + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{1},\\Delta \\theta_{2}" + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "text", + "content": " we introduce the offsets to the weights " + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\theta^{tRGB}\\in \\mathbb{R}^{3\\times 512\\times 3\\times 3}" + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "text", + "content": " of the tRGB convolutional layer of the same block with similar parameterization " + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "inline_equation", + "content": "\\Delta \\theta^{tRGB}\\in \\mathbb{R}^{3\\times 512\\times 1\\times 1}" + }, + { + "bbox": [ + 46, + 193, + 287, + 395 + ], + "type": "text", + "content": ". Further, we omit such detail about tRGB part in the sake of brevity. So, for this parameterization the optimization procedure has the following form:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 402, + 287, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 402, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 72, + 402, + 287, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} \\left(\\left\\{G _ {\\theta , \\Delta \\theta_ {1}, \\Delta \\theta_ {2}} (s (z _ {i})) \\right\\} _ {i = 1} ^ {K}\\right)\\rightarrow \\min _ {\\Delta \\theta_ {1}, \\Delta \\theta_ {2}, f ^ {A}}, \\tag {4}", + "image_path": "2ff6e7f220bce63c6d9aaaf4f9c1bcae594787ab6927c5b2b3dcadcf9461d378.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 434, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 434, + 287, + 458 + ], + "type": "inline_equation", + "content": "G_{\\Delta \\theta_1, \\Delta \\theta_2}" + }, + { + "bbox": [ + 47, + 434, + 287, + 458 + ], + "type": "text", + "content": " is the generator with weight offsets " + }, + { + "bbox": [ + 47, + 434, + 287, + 458 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_1, \\Delta \\theta_2" + }, + { + "bbox": [ + 47, + 434, + 287, + 458 + ], + "type": "text", + "content": " for the one block from the synthesis network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 459, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 495 + ], + "type": "text", + "content": "We call such parameter space as Affine+. We examine all blocks of the synthesis network to choose for this parameterization. We end up with the block with " + }, + { + "bbox": [ + 46, + 459, + 287, + 495 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 459, + 287, + 495 + ], + "type": "text", + "content": " resolution" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 47, + 513, + 298, + 654 + ], + "blocks": [ + { + "bbox": [ + 47, + 513, + 298, + 654 + ], + "lines": [ + { + "bbox": [ + 47, + 513, + 298, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 298, + 654 + ], + "type": "image", + "image_path": "270e298312ff9af55cc641e5b8ce4538179684e2365643b64795f4a0059c91c2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 662, + 287, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 662, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 287, + 710 + ], + "type": "text", + "content": "Figure 5: StyleSpace directions transfer from text-based and image-based domain adaptation to other fine-tuned models. We can successfully transfer style while preserving image content." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "as it shows the best performance (see results of this analysis in Appendix A.4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 97, + 546, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 546, + 217 + ], + "type": "text", + "content": "While Affine+ has already had 6 times less parameters than Full parameterization, we further reduce its size by applying low-rank decomposition to the weights of affine layers. We denote this parameterization as AffineLight+. It gives us by two orders less parameters than Full parameterization with insignificant degradation in quality. We will show in Section 5 that in low data regimes, this parameterization achieves comparable performance as Affine+ and outperforms other baselines. More details about applied low-rank decomposition can be found in Appendix A.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 218, + 545, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 218, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 305, + 218, + 545, + 242 + ], + "type": "text", + "content": "We apply these two parameterizations to few-shot domains and obtain the following results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "content": "We provide results in Figure 4 and in Table 2 (see more results in Appendix A.4). We see that Affine+ parameterization allows us removing the performance gap from the Full parameterization both qualitatively and quantitatively. While the number of parameters for the additional block in Affine+ accounts for only " + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 243, + 545, + 350 + ], + "type": "text", + "content": " of the synthesis network size. It shows that the style vector allows adapting the generator even to more distant domains if we just add a small part of the synthesis network." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 351, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 410 + ], + "type": "text", + "content": "We observe that AffineLight+ still shows adequate performance, except it has 100 times less parameters than the Full parameterization. We should notice that this parameterization is more suitable for low data regimes than those that we consider in Section 5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 411, + 546, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 546, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 546, + 520 + ], + "type": "text", + "content": "Properties of the StyleDomain directions. We investigate StyleDomain directions that adapt the generator to similar domains and explore surprising properties. The first one is mixability, i.e. we discover that StyleDomain directions can be combined with each other. In particular, we can consider several directions that correspond to different similar domains and take their sum. The resulting direction will adapt the generator to the semantically mixed domain. We provide different examples of such combinations in Figure 6" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 551, + 550, + 662 + ], + "blocks": [ + { + "bbox": [ + 310, + 551, + 550, + 662 + ], + "lines": [ + { + "bbox": [ + 310, + 551, + 550, + 662 + ], + "spans": [ + { + "bbox": [ + 310, + 551, + 550, + 662 + ], + "type": "image", + "image_path": "324ff89f310e7558ab38a9dd703d6d34e4c02a7157dee8ef664326322ba33bdf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 673, + 545, + 710 + ], + "lines": [ + { + "bbox": [ + 304, + 673, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 673, + 545, + 710 + ], + "type": "text", + "content": "Figure 6: Example of mixing StyleDomain directions. We can combine different directions in order to perform adaptation into a semantically mixed domain." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2189" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 92, + 536, + 218 + ], + "blocks": [ + { + "bbox": [ + 56, + 92, + 536, + 218 + ], + "lines": [ + { + "bbox": [ + 56, + 92, + 536, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 92, + 536, + 218 + ], + "type": "image", + "image_path": "1a9d49fe13fa9cf1e4d34b2091a72926f452d3b3052213005f98683637b34bd5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 255 + ], + "type": "text", + "content": "Figure 7: Comparison with baselines for one-shot image-based domain adaptation. StyleSpace and StyleSpaceSparse parameterizations achieve comparable quality as other methods while having much less trainable parameters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 274, + 194, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 194, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 194, + 288 + ], + "type": "text", + "content": "(see more results in Appendix A.5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "content": "The second property is transferability, i.e. we can transfer StyleDomain directions between different StyleGAN2 models. In particular, let us consider the base generator " + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "content": " pretrained on realistic faces and the generator " + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "inline_equation", + "content": "G_{\\theta}^{Dog}" + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "content": " fine-tuned to domain of dogs in the Full parameterization. We verify that we can apply StyleDomain directions to " + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "inline_equation", + "content": "G_{\\theta}^{Dog}" + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "content": " which were optimized for " + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "content": ". Specifically, we take Style-Domain directions that were optimized for " + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "content": " to adapt it to different text-based and one-shot image-based domains (e.g. Pixar, Disney, etc.). Next, we apply these directions to generators that were fine-tuned from " + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 46, + 293, + 289, + 449 + ], + "type": "text", + "content": " to other domains (e.g. Dogs, Cats). We provide results of this experiment in Figure 5 (see more results in Appendix A.5)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 454, + 288, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 288, + 502 + ], + "type": "text", + "content": "Finally, we check that StyleDomain directions can be successfully combined with latent controls for image editing. For a more detailed exploration of this property, see in Appendix A.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 274, + 388, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 388, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 388, + 288 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 296, + 547, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 547, + 476 + ], + "type": "text", + "content": "One-shot domain adaptation. For image-based one-shot adaptation, we consider the main baselines: Target-CLIP [7], JoJoGAN [9], MTG [55], GOSA [49], DiFa [48], DomMod [4]. We apply our parameterizations StyleSpace and StyleSpaceSparse to DiFa model because this configuration achieves the highest performance (experiments with other base models can be found in Appendix A.6). Similarly, we apply DomMod parameterization to the DiFa model. We use StyleGAN2 as the GAN architecture and FFHQ as the source domain. For fair comparison, we strictly follow baseline default configurations in the choice of source-target adaptation setups and hyper-parameters. As target domain we use a wide range of different style images from baseline works. The full list of target images can be found in Appendix A.6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 478, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 478, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 478, + 547, + 503 + ], + "type": "text", + "content": "We provide both quantitative and qualitative results in Table 3 and Figure 7, correspondingly. More results and" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 48, + 563, + 568, + 711 + ], + "blocks": [ + { + "bbox": [ + 46, + 518, + 547, + 556 + ], + "lines": [ + { + "bbox": [ + 46, + 518, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 547, + 556 + ], + "type": "text", + "content": "Table 3: Quality and Diversity metrics [4] for one-shot image-based domain adaptations with different methods. Memory denotes the memory needed for keeping adapted generators for all 12 domains for each method. StyleSpace and StyleSpaceSparse parameterizations achieve results comparable to other baselines while having much less trainable parameters." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 563, + 568, + 711 + ], + "lines": [ + { + "bbox": [ + 48, + 563, + 568, + 711 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 568, + 711 + ], + "type": "table", + "html": "
MethodSizeMemoryTitan ErwinDisneyAcross 12 domains
QualityDiversityQualityDiversityQualityDiversity
JoJoGAN [9]30M1.80GB0.5720.2920.5910.2600.590 ± 0.0480.257 ± 0.025
MTG [55]30M1.80GB0.6070.2690.5090.2340.586 ± 0.0540.263 ± 0.028
GOSA [49]30M1.80GB0.5470.2830.6170.2160.584 ± 0.0340.252 ± 0.030
DiFa [48]30M1.80GB0.7190.2260.6990.2630.734 ± 0.0470.215 ± 0.038
TargetCLIP [7]9.0K420KB0.4740.3060.5020.3330.491 ± 0.0430.322 ± 0.015
DomMod (DiFa) [4]6.0K280KB0.7050.2500.6250.2940.679 ± 0.0490.253 ± 0.037
StyleSpace (DiFa)6.0K280KB0.6720.2960.6270.3080.644 ± 0.0410.298 ± 0.025
StyleSpaceSparse (DiFa)1.2K56.4KB0.6590.3030.6170.3040.638 ± 0.0460.299 ± 0.026
", + "image_path": "ac21db5242508d2a5bc07cc553e242fe3aee00d677007f31028c9bee89bad22e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "2190" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "samples can be found in Appendix A.6. We observe that the DiFa model achieves the best Quality metric but has low Diversity compared to other methods. Our parameterizations applied to the DiFa model balance its performance in terms of these metrics. So, StyleSpace (DiFa) and StyleSpaceSparse (DiFa) achieves uniformly better results than JoJOGAN, MTG, GOSA baselines. DomMod parameterization also show good performance, but it is comparable with our StyleSpaceSparse parameterization that has 5 times less parameters. We also report the overall memory needed for keeping adapted generators for 12 domains. We observe that StyleSpaceSparse requires significantly less space than other models. It will be especially important if we scale the number of target domains to thousands and more. We note that TargetCLIP that also has small number of trainable parameters show very poor results visually and in terms of Quality metric. It has a high Diversity only because its generated images are very close to the original ones. To provide more comprehensive comparison we additionally conduct user studies and present results in Appendix A.6.4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 289, + 507 + ], + "type": "text", + "content": "Few-shot domain adaptation. As main baselines for this task, we consider the vanilla StyleGAN-ADA [18] (we will denote it as ADA), CDC [28] and recent SOTA method AdAM [51]. We compare our parameterizations Affine+ and AffineLight+ applied to ADA with these baselines on two datasets: Dogs and Cats [8]. We study the efficiency of these methods with different numbers of available target samples. For fair comparison, we rigorously follow the training setups of all baselines. We note that we use the same number of training iterations and the same batch size for all methods. We observe that in the AdAM paper [51] the number of training iterations is set to a small value (less than 10K) and it causes underfitting of the vanilla ADA method. Therefore, we increase this number for all methods to 50K and use the same batch size of 4. Then, for each method, we report the best FID value that it obtains during" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 47, + 529, + 299, + 666 + ], + "blocks": [ + { + "bbox": [ + 47, + 529, + 299, + 666 + ], + "lines": [ + { + "bbox": [ + 47, + 529, + 299, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 299, + 666 + ], + "type": "image", + "image_path": "f5bea334163ffac3287ba57ac13af8234dba941b9aef2c153ab3feff2861680a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 673, + 288, + 711 + ], + "lines": [ + { + "bbox": [ + 46, + 673, + 288, + 711 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 288, + 711 + ], + "type": "text", + "content": "Figure 8: Few-shot training results for different number of shots. Proposed ADA (Affine+) and ADA (AffineLight+) show uniformly better performance than baselines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 310, + 115, + 541, + 228 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 106 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 106 + ], + "type": "text", + "content": "Table 4: Results for few-shot training with 10-shots. Proposed ADA (Affine+) and ADA (AffineLight+) achieve better performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 115, + 541, + 228 + ], + "lines": [ + { + "bbox": [ + 310, + 115, + 541, + 228 + ], + "spans": [ + { + "bbox": [ + 310, + 115, + 541, + 228 + ], + "type": "table", + "html": "
MethodSizeDomains (10-shots)
CatDog
CDC [28]30M66.24184.56
AdAM [51]19M47.05119.61
ADA (Full) [18]30M51.38100.25
ADA (Affine+)5.1M38.4096.38
ADA (AffineLight+)0.6M43.91101.31
", + "image_path": "9acdee676117753a6968a50b024ed3f30e8b894126514cf63b8fa9b40fe361dc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 235, + 541, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 235, + 541, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 235, + 541, + 247 + ], + "type": "text", + "content": "training. See more details about training in Appendix A.7." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 247, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 367 + ], + "type": "text", + "content": "We report results for few-shot training for different numbers of shots in Figure 8 and separately for 10 shots in Table 4. More results and samples can be found in Appendix A.7. Firstly, we observe that the AdAM method shows results not better than the vanilla ADA (Full) when trained for sufficient number of iterations. Secondly, we see that ADA (Affine+) and ADA (AffineLight+) based on our parameterizations achieve better results uniformly for all numbers of shots. It shows that these parameterizations is especially suitable for low data regimes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 368, + 545, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 545, + 463 + ], + "type": "text", + "content": "Cross-domain image translation. We consider two setups of standard image-to-image problem. In the first one, we translate images from the source domain to the target domain unconditionally. In the second setup, we perform a reference-based translation, where the resulting image combines the pose and the shape of the source image with the style from a reference image. Details of these experiments can be found in Appendix A.9." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 464, + 546, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 546, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 546, + 620 + ], + "type": "text", + "content": "Cross-domain image morphing. Cross-domain morphing is a smooth transition between two images from different domains. This task is known as challenging [3, 10] and it is successfully tackled in the work [44] using aligned StyleGAN2 models. The idea is to interpolate between aligned generator weights to obtain a smooth transition between domains. We propose more complex image morphing by utilizing the transferability of StyleDomain directions. For example, we can apply a direction that stands for Sketch or Pixar style to Dogs domain to obtain a smooth transition between sketchy and pixar-like dog (see Figure 1). See many examples of such complex cross-domain morphing in Appendix A.10." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 633, + 378, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 378, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 378, + 645 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 715 + ], + "type": "text", + "content": "In this paper, we provide an extensive analysis of the process of StyleGAN domain adaptation. We reveal the sufficient components of the generator for successful adaptation depending on the similarity between the source and target domains. We discover the ability of StyleSpace to change" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "2191" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "the domain of generated images by StyleDomain directions. We also propose new efficient parameterizations Affine+ and AffineLight+ for few-shot adaptation that outperform existing baselines. Further, we explore and leverage the properties of StyleDomain directions. We believe that our investigation can attract more attention to the exploration of new and interesting properties of StyleSpace." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 168, + 157, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 157, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 157, + 182 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 189, + 289, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 289, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 289, + 308 + ], + "type": "text", + "content": "The results of sections 1, 2, 3 were obtained by Aibek Alanov and Dmitry Vetrov with the support of the grant for research centers in the field of AI provided by the Analytical Center for the Government of the Russian Federation (ACRF) in accordance with the agreement on the provision of subsidies (identifier of the agreement 000000D730321P5Q0002) and the agreement with HSE University No. 70-2021-00139. This research was supported in part through computational resources of HPC facilities at HSE University." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 320, + 107, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 107, + 333 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 107, + 333 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 340, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 340, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 385 + ], + "type": "text", + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "type": "text", + "content": "[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 33" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 442, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 487 + ], + "type": "text", + "content": "[3] Kfir Aberman, Jing Liao, Mingyi Shi, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Neural best-buddies: Sparse cross-domain correspondence. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 488, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 287, + 532 + ], + "type": "text", + "content": "[4] Aibek Alanov, Vadim Titov, and Dmitry Vetrov. Hyperdomainnet: Universal domain adaptation for generative adversarial networks. arXiv preprint arXiv:2210.08884, 2022. 1, 3, 4, 5, 7, 13, 18, 23, 24, 34, 35, 36, 37" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 533, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 287, + 566 + ], + "type": "text", + "content": "[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 567, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 567, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 567, + 288, + 622 + ], + "type": "text", + "content": "[6] Kelvin CK Chan, Xintao Wang, Xiangyu Xu, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for large-factor image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14245-14254, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 678 + ], + "type": "text", + "content": "[7] Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf. Image-based clip-guided essence transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 695-711. Springer, 2022. 3, 7, 34, 35" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 287, + 713 + ], + "type": "text", + "content": "[8] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vi" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "sion and pattern recognition, pages 8188-8197, 2020. 4, 5, 8, 40" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 95, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 95, + 545, + 140 + ], + "type": "text", + "content": "[9] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 128-152. Springer, 2022. 3, 7, 34, 35" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 193 + ], + "type": "text", + "content": "[10] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, volume 39, pages 303–313. Wiley Online Library, 2020. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "type": "text", + "content": "[11] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 1, 3, 4, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 282 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 284, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 545, + 326 + ], + "type": "text", + "content": "[13] Shanyan Guan, Ying Tai, Bingbing Ni, Feida Zhu, Feiyue Huang, and Xiaokang Yang. Collaborative learning for faster stylegan embedding. arXiv preprint arXiv:2007.01758, 2020.2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 370 + ], + "type": "text", + "content": "[14] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 425 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "text", + "content": "[16] Jialu Huang, Jing Liao, and Sam Kwong. Unsupervised image-to-image translation via pre-trained stylegan2 network. IEEE Transactions on Multimedia, 24:1435-1448, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 471, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 503 + ], + "type": "text", + "content": "[17] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the\" steerability\" of generative adversarial networks. arXiv preprint arXiv:1907.07171, 2019. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "type": "text", + "content": "[18] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 1, 2, 4, 8, 12, 19, 40" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 602 + ], + "type": "text", + "content": "[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 603, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 657 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "text", + "content": "[21] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1, 2, 3, 12" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2192" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[22] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[23] Bingchen Liu, Yizhe Zhu, Kunpeng Song, and Ahmed Elgammal. Towards faster and stabilized gan training for high-fidelity few-shot image synthesis. In International Conference on Learning Representations, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "text", + "content": "[24] Xuan Luo, Xuaner Zhang, Paul Yoo, Ricardo Martin-Brualla, Jason Lawrence, and Steven M Seitz. Time-travel rephotography. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 230 + ], + "type": "text", + "content": "[25] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze the discriminator: a simple baseline for fine-tuning gans. arXiv preprint arXiv:2002.10964, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 275 + ], + "type": "text", + "content": "[26] M-E Nilsback and Andrew Zisserman. A visual vocabulary for flower classification. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), volume 2, pages 1447-1454. IEEE, 2006. 19" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 320 + ], + "type": "text", + "content": "[27] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 386 + ], + "type": "text", + "content": "[28] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 1, 3, 8, 40" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "text", + "content": "[29] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 488 + ], + "type": "text", + "content": "[30] Stanislav Pidhorskyi, Donald A Adjeroh, and Gianfranco Doretto. Adversarial latent autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14104-14113, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "text", + "content": "[31] Justin NM Pinkney and Doron Adler. Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334, 2020. 1, 3, 19" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[33] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[34] Esther Robb, Wen-Sheng Chu, Abhishek Kumar, and Jia-Bin Huang. Few-shot adaptation of generative adversarial networks. arXiv preprint arXiv:2010.11943, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[35] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF conference on computer vi" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "sion and pattern recognition, pages 9243-9252, 2020. 1, 2, 31" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "text", + "content": "[36] Guoxian Song, Linjie Luo, Jing Liu, Wan-Chun Ma, Chunpong Lai, Chuanxia Zheng, and Tat-Jen Cham. Agilegan: stylizing portraits by inversion-consistent transfer learning. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 205 + ], + "type": "text", + "content": "[37] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6142-6151, 2020. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "type": "text", + "content": "[38] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "type": "text", + "content": "[39] Ngoc-Trung Tran, Viet-Hung Tran, Ngoc-Bao Nguyen, Trung-Kien Nguyen, and Ngai-Man Cheung. On data augmentation for gan training. IEEE Transactions on Image Processing, 30:1882-1897, 2021. 1, 2, 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "text", + "content": "[40] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7921-7931, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[41] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9168-9178, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 470 + ], + "type": "text", + "content": "[42] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. Minegan: effective knowledge transfer from gans to target domains with few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9332-9341, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "type": "text", + "content": "[43] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 1, 2, 31, 33" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 568 + ], + "type": "text", + "content": "[44] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. Stylealign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 1, 3, 4, 8, 47" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 613 + ], + "type": "text", + "content": "[45] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. Advances in Neural Information Processing Systems, 34:9378-9390, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "text", + "content": "[46] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[47] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 19, 40" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2193" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 431 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[48] Yabo Zhang, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, Wangmeng Zuo, et al. Towards diverse and faithful one-shot adaptation of generative adversarial networks. In Advances in Neural Information Processing Systems, 2022. 3, 4, 7, 12, 34, 35" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[49] Zicheng Zhang, Yinglu Liu, Congying Han, Tiande Guo, Ting Yao, and Tao Mei. Generalized one-shot domain adaption of generative adversarial networks. arXiv preprint arXiv:2209.03665, 2022. 3, 7, 34, 35" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "text", + "content": "[50] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in Neural Information Processing Systems, 33:7559-7570, 2020. 1, 2, 12" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[51] Yunqing Zhao, Keshigeyan Chandrasegaran, Milad Abdol-lahzadeh, and Ngai-Man Cheung. Few-shot image generation via adaptation-aware kernel modulation. arXiv preprint arXiv:2210.16559, 2022. 3, 8, 40" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 297 + ], + "type": "text", + "content": "[52] Zhengli Zhao, Zizhao Zhang, Ting Chen, Sameer Singh, and Han Zhang. Image augmentations for gan training. arXiv preprint arXiv:2006.02595, 2020. 1, 2, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "text", + "content": "[53] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 386 + ], + "type": "text", + "content": "[54] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In European conference on computer vision, pages 597-613. Springer, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 387, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 431 + ], + "type": "text", + "content": "[55] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. arXiv preprint arXiv:2110.08398, 2021. 1, 3, 7, 34, 35" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2194" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/71e23834-fb4f-4c96-b627-797882a6ea6f_content_list.json b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/71e23834-fb4f-4c96-b627-797882a6ea6f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d441a08d53c714e2a627f9f955f5661375a5bc7d --- /dev/null +++ b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/71e23834-fb4f-4c96-b627-797882a6ea6f_content_list.json @@ -0,0 +1,2163 @@ +[ + { + "type": "text", + "text": "StyleGANEX: StyleGAN-Based Manipulation Beyond Cropped Aligned Faces", + "text_level": 1, + "bbox": [ + 91, + 130, + 879, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuai Yang Liming Jiang Ziwei Liu Chen Change Loy S-Lab, Nanyang Technological University", + "bbox": [ + 220, + 176, + 750, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shuai.yang, liming002, ziwei.liu, ccloy}@ntu.edu.sg", + "bbox": [ + 254, + 215, + 714, + 231 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/24138a5cf244f6b66f18abea521c020c4e2094f264edfe0ec5904e992b1dc660.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 247, + 679, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7bbe1fc312a04d55ea14d1b6bc2371a42c134998921f9b81c60e1279a2842e3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 247, + 888, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a02fde41a725b936f700d5f0efbc0abfab6eb6b32fc265ab506ddaf142dec22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 351, + 478, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec4c3f7113165c993a297da4c4cffb69a314dc846787ec58ef3ead13b3ac105c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 349, + 888, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4303ceac55161479649b24e326d7a7084fabb25da09bafd8eb88d9a8ac5641fa.jpg", + "image_caption": [ + "Figure 1: We expand StyleGAN to encompass a diverse set of tasks that go beyond the constraints of cropped aligned faces." + ], + "image_footnote": [], + "bbox": [ + 81, + 446, + 888, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 585, + 313, + 601 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in face manipulation using StyleGAN have produced impressive results. However, StyleGAN is inherently limited to cropped aligned faces at a fixed image resolution it is pre-trained on. In this paper, we propose a simple and effective solution to this limitation by using dilated convolutions to rescale the receptive fields of shallow layers in StyleGAN, without altering any model parameters. This allows fixed-size small features at shallow layers to be extended into larger ones that can accommodate variable resolutions, making them more robust in characterizing unaligned faces. To enable real face inversion and manipulation, we introduce a corresponding encoder that provides the first-layer feature of the extended StyleGAN in addition to the latent style code. We validate the effectiveness of our method using unaligned face inputs of various resolutions in a diverse set of face manipulation tasks, including facial attribute editing, super-resolution, sketch/mask-to-face translation, and face toonification. Project page https:", + "bbox": [ + 75, + 609, + 473, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "//www.mmlab-ntu.com/project/styleganex", + "bbox": [ + 78, + 883, + 457, + 898 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 585, + 632, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "StyleGAN [17, 18] has emerged as one of the most successful models for generating high-quality faces. Building upon StyleGAN, researchers have developed a range of face manipulation models [1, 23, 29, 26, 10, 28, 40, 36]. These models typically map real face images or other face-related inputs to the latent space of StyleGAN, perform semantic editing in the latent space, and then map the edited latent code back to the image space. This approach enables a variety of tasks, including facial attribute editing, face restoration, sketch-to-face translation, and face toonification. As the manipulated faces remain within the generative space of StyleGAN, the quality of the image output is guaranteed.", + "bbox": [ + 496, + 612, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite its ability to ensure high-quality image output, the generative space of StyleGAN is limited by a fixed-crop constraint that restricts image resolution and face layout. As a result, existing face manipulation models based on StyleGAN can only handle cropped and aligned face images. In such images with a limited field of view (FoV), the face typically dominates the image, leaving little room for back-", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21000", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ground and clothing, and often resulting in partially cropped hair. However, in everyday portrait photos such as selfies, faces occupy a smaller proportion of the image, allowing for a complete hairstyle and upper body seen. Portrait videos such as those from live streaming require an even larger background area to accommodate face movement. To process these types of inputs, which we refer to as normal FoV face images and videos, existing manipulation models need to align, crop, and edit the face before pasting the result back onto the original image [4, 19, 31]. This approach often results in discontinuity near the seams, e.g., only editing the hair color inside the cropped area.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While StyleGAN3 [16] was introduced to address unaligned faces, a recent study [4] found that even StyleGAN3 requires face realignment before effectively projecting to its latent space. Moreover, StyleGAN3 is still constrained by a fixed image resolution. Motivated by the translation equivariance of convolutions, VToonify [37] addresses the fixed-crop limitation of StyleGAN by removing its shallow layers to accept input features of any resolution. However, these shallow layers are crucial for capturing high-level features of the face, such as pose, hairstyle, and face shape. By removing these layers, the network loses its ability to perform latent editing on these important features, which is a distinctive capability of StyleGAN. Therefore, the challenge remains in overcoming the fixed-crop limitation of StyleGAN while preserving its original style manipulation abilities, which is a valuable research problem to solve.", + "bbox": [ + 75, + 275, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a simple yet effective approach for refactoring StyleGAN to overcome the fixed-crop limitation. In particular, we refactor its shallow layers instead of removing them, allowing the first layer to accept input features of any resolution. This simple change expands StyleGAN's style latent space into a more powerful joint style latent and first-layer feature space $(W^{+} - F$ space), extending the generative space beyond cropped aligned faces. Furthermore, our refactoring only changes the receptive field of shallow-layer convolutions, leaving all pre-trained model parameters intact. Hence, the refactored StyleGAN (StyleGANEX) can directly load the original StyleGAN parameters, fully compatible with the generative space of StyleGAN, and retains its style representation and editing ability. This means that the StyleGAN editing vectors found in previous studies [26, 10, 28] can be directly applied to StyleGANEX for normal FoV face editing, e.g., changing the face pose, as shown in Fig. 1(a).", + "bbox": [ + 75, + 518, + 472, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on StyleGANEX, we further design a corresponding encoder that projects normal FoV face images to the $W^{+} - F$ space for real face inversion and manipulation. Our encoder builds upon pSp encoder [23] and aggregates its multi-layer features to predict the first-layer feature of StyleGANEX. The encoder and StyleGANEX form a fully convolutional encoder-decoder framework. With the first-", + "bbox": [ + 75, + 795, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "layer feature as the bottleneck layer, whose resolution is $1/32$ of the output image, our framework can handle images and videos of various resolutions, as long as their side lengths are divisible by 32. Depending on the input and output types, our framework can perform a wide range of face manipulation tasks. In this paper, we select several representative tasks, as shown in Fig. 1, including facial attribute editing, face super-resolution, sketch/mask-to-face translation and video face toonification. While the focus of these tasks in the past is limited to cropped aligned faces, our framework can handle normal FoV faces, showing significant advantages over previous StyleGAN-based approaches. To summarize, our main contributions are:", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel StyleGANEX architecture with extended $W^{+} - F$ space, which overcomes the fixed-crop limitation of StyleGAN.", + "- An effective encoder that is able to project normal FoV face images into the $W^{+} - F$ domain.", + "- A generic and versatile fully convolutional framework for face manipulation beyond cropped aligned faces." + ], + "bbox": [ + 514, + 290, + 890, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 420, + 640, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "StyleGAN inversion. StyleGAN inversion aims at projecting real face images into the latent space of StyleGAN for further manipulation. Image2StyleGAN [1] analyzes the latent space and proposes $W^{+}$ space to reconstruct real faces with latent code optimization. PIE [29] and IDinvert [41] further consider the editability of the latent code during optimization. To speed up inversion, pSp [23] and e4e [30] train an encoder to directly project the target face to its corresponding latent code, which is however hard to reconstruct fine details and handle occlusions. To solve this issue, Restyle [3] and HFGI [33] predict the residue of latent codes or mid-layer features to reduce errors, respectively. Instead of focusing on the latent code, PTI [24] optimizes StyleGAN itself to fit a target face, which is accelerated by HyperInverter [8] and HyperStyle [5] to predict offsets of network parameters with hyper networks. The above methods are limited to cropped aligned faces for valid face editing. With the extended $W^{+} - F$ space, our framework is able to perform inversion on normal FoV face images.", + "bbox": [ + 496, + 446, + 893, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "StyleGAN-based face manipulation. An intuitive way of StyleGAN-based face manipulation is to optimize the latent code online to achieve certain objectives such as a pixel-level constraints [1], sketch-based structure constraints [19] or text-guided semantic constraints [21, 9]. Another way is to search for offline editing vectors to add to the latent code for manipulation. Supervised methods identify meaningful editing vectors based on attribute labels or pretrained classifiers [26, 27, 13, 2]. On the other hand, unsupervised methods statistically analyze the StyleGAN latent space to discover semantically significant editing direc", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21001", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions by principal component analysis [10], low-rank factorization [40] and closed-form factorization [28]. Meanwhile, face manipulation can be directly realized by image-to-image translation frameworks, where StyleGAN is used to generate paired training data [32, 37] or to build the decoder [23, 6, 37]. To achieve spatial editing, methods [42, 20] are proposed to manipulate mid-layer features in addition to the latent code, similar to our $W^{+} - F$ space. Moreover, BDInvert [14] and StyleHEAT [38] introduces feature transformations for unaligned face editing. However, the above methods, as well as HFGI [33], follow StyleGAN features' original fixed resolution, thus still suffering the crop limitation. Differently, our method predicts the first-layer feature that can have various resolutions in StyleGANEX. Moreover, as we will show later in Sec. 5.1, compared to VToonify [37], StyleGANEX retains the complete style manipulation abilities of the shallow layers of StyleGAN, with a jointly trained latent code and feature extractor, preserving vivid details and supporting more diverse manipulation tasks.", + "bbox": [ + 76, + 90, + 472, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. StyleGANEX", + "text_level": 1, + "bbox": [ + 76, + 410, + 215, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Analysis of the Fixed-Crop Limitation", + "text_level": 1, + "bbox": [ + 76, + 436, + 405, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StyleGAN has great potential for handling normal FoV face images. The generator of StyleGAN is a fully convolutional architecture that can naturally handle different feature resolutions, as convolution operations and the style modulation of StyleGAN are independent of the input resolution. Additionally, the translation equivariance of convolution operations naturally supports feature translation. As analyzed in VToonify [37], if we translate or rotate the feature of the 7th layer of StyleGAN, the resulting face will also be shifted or rotated, as shown in Figs. 2(a)(b)(d).", + "bbox": [ + 76, + 460, + 468, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The limiting factor originates from StyleGAN's constant first-layer feature. First, the first-layer feature has a fixed resolution of $4 \\times 4$ , limiting the output to $1024 \\times 1024$ resolution. Second, $4 \\times 4$ resolution is inadequate to characterize the spatial information of unaligned faces. We have taken a step further than the analysis in VToonify [37] to investigate translation and rotation on the first-layer feature. As shown in Fig. 2(c), sub-pixel translation fuses adjacent feature values severely, resulting in a blurry face due to the small number of elements (only 16) in the first-layer feature. In Fig. 2(e), the first-layer feature fails to provide enough spatial information for a valid rotation. In comparison, the 7-th layer has a higher resolution ( $32 \\times 32$ ), making it better suited for capturing spatial information. However, only a single layer alone provides limited style control, as the full facial structural styles are hierarchically modeled by seven shallow layers. Simply ignoring low-resolution layers, as in VToonify, disables flexible latent editing over the styles of these layers, such as pose, age, and face shape.", + "bbox": [ + 76, + 613, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4dd2f7b042a90d2a10d4f3c078d29d1a52f4fef4d160a1993f7dff28e2959671.jpg", + "image_caption": [ + "(a) random face" + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 581, + 148 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ee0a5d11c10510ab1c005852e1af646a2c024584a33bab0164c9d9ec7b54fdda.jpg", + "image_caption": [ + "(b) shift layer#7" + ], + "image_footnote": [], + "bbox": [ + 583, + 90, + 656, + 148 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/061f085f73cb3728542f5d6c99e17c0023609d63814507960f215398d6169b7f.jpg", + "image_caption": [ + "(c) shift layer#1" + ], + "image_footnote": [], + "bbox": [ + 658, + 90, + 733, + 148 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6476d53f8cdbeb96d51bd3412c8ce99ed95c5f3a1bc629f71ce4c40a66060e20.jpg", + "image_caption": [ + "(d) rotate layer#7" + ], + "image_footnote": [], + "bbox": [ + 735, + 90, + 810, + 148 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d4bc1f3feb8f3bb0efae3c3d5cfdbceb0452a94d7d4bb04c743b9db9e226bfd2.jpg", + "image_caption": [ + "(e) rotate layer#1" + ], + "image_footnote": [], + "bbox": [ + 815, + 90, + 890, + 148 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e9e946eeb589cb141531e090236900e04ad9959a2c1d23adee6eb206fc75a298.jpg", + "image_caption": [ + "Figure 2: Analysis of StyleGAN in generating unaligned faces. (a) Generated face. (b)(c) Face generated by translating the feature maps of the 7-th and 1-st layers of StyleGAN respectively to shift the face by 150 pixels. (d)(e) Face generated by rotating the feature maps of the 7-th and 1-st layers of StyleGAN by 10 degrees, respectively.", + "(a) real face photo", + "Figure 3: Analysis of StyleGANEX in generating unaligned faces. (a) A real face photo. (b) Reconstructed face by projecting (a) into the $W^{+} - F$ space of StyleGANEX. (c) Face generated by translating the first-layer feature map to shift the face by 300 pixels. (d) Face generated by rotating the first-layer feature map by 10 degrees." + ], + "image_footnote": [], + "bbox": [ + 503, + 267, + 598, + 335 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8b0baa2f0e2280d354352224c86f6fa41813a265bcbae6172a37037c66cdb512.jpg", + "image_caption": [ + "(b) projected face" + ], + "image_footnote": [], + "bbox": [ + 601, + 268, + 696, + 335 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/32e1b5de8500028b7611a25ef1cbcaebd62ff1864d7ef0afa23ded9ee02094cc.jpg", + "image_caption": [ + "(c) shift layer#1" + ], + "image_footnote": [], + "bbox": [ + 697, + 268, + 792, + 335 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1cf618f374e035e114bf8f014572bef5422d516786723eccc01b70f578c1e7f0.jpg", + "image_caption": [ + "(d) rotate layer#1" + ], + "image_footnote": [], + "bbox": [ + 795, + 268, + 890, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Expanding the shallow layers of StyleGAN to have the same $32 \\times 32$ resolution as the 7th layer, or more generally, $H / 32 \\times W / 32$ resolution for an $H \\times W$ image, would provide enough structure and layout information to combine the style controllability of shallow layers with support for normal FoV faces. This is precisely the key idea of StyleGANEX, and we will introduce our simple solution in Section 3.2. As a preview of the performance of the expanded layers in enabling face manipulation beyond cropped and aligned faces, Fig. 3 shows that for a $1472 \\times 1600$ normal FoV face photo, we can obtain its latent code and an additional $46 \\times 50$ first-layer feature as the input to StyleGANEX (inversion method explained in Section 4.2). Face translation and rotation can be realized by shifting or rotating the first-layer feature. Additionally, the face can be effectively edited by applying style mixing [17] or InterFaceGAN editing vectors [26] to the latent code, as shown in Fig. 1(a).", + "bbox": [ + 498, + 470, + 892, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. From StyleGAN to StyleGANEX", + "text_level": 1, + "bbox": [ + 500, + 739, + 787, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 4 illustrates the generator architectures of StyleGAN and StyleGANEX. Compared to StyleGAN, we first replace the constant $4 \\times 4$ first-layer feature with a variable feature whose resolution is $1/32$ of the output image. Then, we remove the upsample operations before the 8-th layer, allowing features in the seven shallow layers to share the same resolution as the 7-th layer. However, the convolution kernels or reception fields of these layers do not match their input features with the enlarged resolution. To solve", + "bbox": [ + 498, + 763, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21002", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5bd3a5e9b589ed01fc97536ed7581464501c396357edf0264444fc3b98203b13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 261, + 314 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a283461e5c0acbaa9c835876077a2b31484bfab1e735c0556542563932f5a3cb.jpg", + "image_caption": [ + "Figure 4: Refactor StyleGAN to StyleGANEX. For simplicity, learned weights, biases and noises are omitted." + ], + "image_footnote": [], + "bbox": [ + 277, + 88, + 467, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this problem, we enlarge the reception fields by modifying the convolutions to their dilated versions. For example, the first layer only needs to change the dilation factor from 1 to 8. With this simple modification, StyleGAN is refactored to StyleGANEX. Since the first layer becomes variable, the original $W^{+}$ latent space is extended to a joint $W^{+} - F$ space, where the latent code $w^{+} \\in W^{+}$ provides style cues, and the first-layer feature $f \\in F$ mainly encodes spatial information.", + "bbox": [ + 75, + 415, + 468, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The refactoring of StyleGAN to StyleGANEX has three key advantages. 1) Support for unaligned faces. The resolution enlargement and variable first-layer features of StyleGANEX overcome the fixed-crop limitation. 2) Compatibility. No model parameters are altered during refactoring, meaning that StyleGANEX can directly load pretrained StyleGAN parameters without retraining. In fact, if we upsample the StyleGAN's constant input feature by $8 \\times$ with nearest neighbor interpolation to serve as $f$ of StyleGANEX, StyleGANEX degrades exactly to StyleGAN with the same $1024 \\times 1024$ generative space. The computational cost of the refactoring is also minimal, with generating an image taking 0.026s and 0.028s for StyleGAN and StyleGANEX, respectively. 3) Flexible manipulation. StyleGANEX retains the style representation and editing ability of StyleGAN, meaning that abundant StyleGAN-based face manipulation techniques can be applied to StyleGANEX.", + "bbox": [ + 75, + 551, + 470, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Face Manipulation with StyleGANEX", + "text_level": 1, + "bbox": [ + 76, + 821, + 418, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. StyleGANEX Encoder", + "text_level": 1, + "bbox": [ + 76, + 847, + 285, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section introduces our StyleGANEX encoder $E$ which is used to project real face images into the $W^{+} - F$", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/143eae5cf6a3f2e6efba528a654ba7cb1192618df0df4e3509be087d8f5745b2.jpg", + "image_caption": [ + "Figure 5: Details of StyleGANEX Encoder." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 890, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "space of StyleGANEX $G$ . The encoder builds upon the pSp encoder [23], as depicted in Fig. 5. Specifically, for $F$ space, we concatenate pSp features in the middle layers and add a convolution layer to map the concatenated features to the first-layer input feature $f$ of $G$ . For $W^{+}$ space, the original pSp encoder takes a $256 \\times 256$ image as input and convolves it to eighteen $1 \\times 1 \\times 512$ features to map to a latent code $w^{+} \\in \\mathbb{R}^{18 \\times 512}$ . To make $E$ accept more general $H \\times W$ images, we add global average pooling to resize all features to $1 \\times 1 \\times 512$ before mapping to latent codes. To support various face manipulation tasks flexibly, we can extract $f$ and $w^{+}$ from different sources. Let $E_{F}$ and $E_{W}$ be the operation of $E$ to extract the first-layer feature and the latent code, respectively. We have", + "bbox": [ + 496, + 337, + 890, + 549 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf, w ^ {+} = E _ {F} \\left(x _ {1}\\right), E _ {W} \\left(x _ {2}\\right) := E \\left(x _ {1}, x _ {2}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 559, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{1}$ and $x_{2}$ are the source inputs for face layout and face style, respectively. Then, a general form of image generation by $G$ from $x_{1}$ and $x_{2}$ is $G(E_{F}(x_{1}), E_{W}(x_{2}))$ . In some face manipulation tasks like super-resolution [6] and toonification [37], passing encoder features to the generator via skip connections helps preserve the details of the input image. We thus introduce a scalar parameter $\\ell$ to the generation process, indicating the $\\ell$ shallow layers of $G$ receive the encoder features $(\\ell = 0$ means no skip connections):", + "bbox": [ + 496, + 587, + 890, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = G \\left(E _ {F} \\left(x _ {1}, \\ell\\right), E _ {W} \\left(x _ {2}\\right)\\right) := G \\left(E \\left(x _ {1}, x _ {2}, \\ell\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 734, + 890, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_{F}(x_{1},\\ell)$ provides both $f$ and the skipped encoder features. For $x_{1}$ of $H\\times W$ resolution, $f$ and the generated image $\\hat{x}$ will be of $H / 8\\times W / 8$ and $4H\\times 4W$ , respectively. The resolution of $x_{2}$ can be independent of $x_{1}$ and $\\hat{x}$", + "bbox": [ + 496, + 762, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. StyleGANEX Inversion and Editing", + "text_level": 1, + "bbox": [ + 500, + 832, + 808, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To find appropriate $\\hat{f}$ and $\\hat{w}^+$ that precisely reconstruct a target image $x$ , we perform a two-step StyleGANEX inversion. Step I projects $x$ to initial $f$ and $w^+$ with $E$ . Step", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21003", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d5c2194e9f68e66133808db72ab55b3dbf92a60fba7124c8e47f48d7b79cbf21.jpg", + "image_caption": [ + "(a) face photo" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 176, + 162 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4f16ec3423c7f14004d8ad934ce3800017349edcbdd61d2339d7d7b39cc2acd7.jpg", + "image_caption": [ + "(b) encoded face" + ], + "image_footnote": [], + "bbox": [ + 176, + 90, + 272, + 162 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b514d10c07b6b37a66d09e4d1f77307de3b189e39a61e433c3fea75bb3006e3b.jpg", + "image_caption": [ + "(c) (b)+optimize", + "(d) domain transfer" + ], + "image_footnote": [], + "bbox": [ + 272, + 90, + 367, + 162 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3a57a0d2749731b37067e1c9326cef7d50b8cadad51d6ca018fee25782a508fc.jpg", + "image_caption": [ + "Figure 6: StyleGANEX inversion." + ], + "image_footnote": [], + "bbox": [ + 369, + 90, + 464, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "II optimizes $f$ and $w^{+}$ to further reduce the reconstruction error. The training of $E$ follows pSp with reconstruction losses and a regularization loss [23]:", + "bbox": [ + 75, + 210, + 468, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} (\\hat {x}, x) + \\lambda_ {1} \\mathcal {L} _ {\\text {r e g}} \\left(E _ {W} (\\tilde {x})\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 268, + 468, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{x} = G(E(x, \\tilde{x}, 0))$ and $\\tilde{x}$ is the cropped aligned face region of $x$ . We empirically find that using $\\tilde{x}$ instead of $x$ predicts more accurate $w^{+}$ since StyleGAN is originally trained on cropped aligned faces. $\\mathcal{L}_{\\mathrm{reg}}$ encourages the predicted $w^{+}$ closer to the average latent code to improve the image quality. $\\mathcal{L}_{\\mathrm{rec}}$ measures the distance between the reconstructed $\\hat{x}$ and the target $x$ in terms of pixel similarity, perceptual similarity, and identity preservation:", + "bbox": [ + 75, + 295, + 468, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e c}} (\\hat {x}, x) = \\lambda_ {2} \\mathcal {L} _ {2} (\\hat {x}, x) + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {L P I P S}} (\\hat {x}, x) + \\lambda_ {4} \\mathcal {L} _ {\\mathrm {I D}} (\\hat {x}, x).\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 428, + 465, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 6(b), $\\hat{x}$ largely approximates $x$ . But the background details and clothings are still hard to reconstruct. Therefore, we further optimize $f$ and $w^{+}$", + "bbox": [ + 75, + 455, + 468, + 501 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f}, \\hat {w} ^ {+} = \\underset {f, w ^ {+}} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\mathrm {L P I P S}} (G (f, w ^ {+}), x), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 510, + 468, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f$ and $w^{+}$ are initialized by $E(x,\\tilde{x},0)$ . The optimized $G(\\hat{f},\\hat{w}^{+})$ in Fig. 6(c) well reconstructs $x$ .", + "bbox": [ + 75, + 549, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After inversion, we can perform flexible editing over $x$ as in StyleGAN. Figure 1(a) shows two examples: we can exchange the last 11 elements of $\\hat{w}^{+}$ with random samples, to mix the color and texture styles; we can add InterFaceGAN editing vectors [26] to $\\hat{w}^{+}$ to make a young face. Moreover, as shown in Fig. 6(d), if we load $G$ a pre-trained StyleGAN-NADA Disney Princess model [9] (let $G'$ denote the new $G$ ), we can obtain $G'(\\hat{f}, \\hat{w}^{+})$ , a cartoon version of $x$ .", + "bbox": [ + 75, + 580, + 468, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. StyleGANEX-Based Translation", + "text_level": 1, + "bbox": [ + 76, + 710, + 362, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The encoder and StyleGANEX form an end-to-end image-to-image translation framework in Fig. 5. Depending on the type of paired training data, it can be trained to efficiently realize different face manipulation tasks. As with pSp [23], we will fix StyleGANEX generator and only train the encoder on the given task.", + "bbox": [ + 75, + 734, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face super-resolution. Given low-resolution and high-resolution training image pairs $(x,y)$ , we can train $E$ to recover $y$ from $x$ to learn face super-resolution with the loss", + "bbox": [ + 75, + 825, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} (\\hat {y}, y) + \\lambda_ {5} \\mathcal {L} _ {\\text {a d v}} (\\hat {y}, y), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{y} = G(E(x_{\\uparrow},\\tilde{x}_{\\uparrow},7))$ $\\uparrow$ is the upsample operation to make $x$ match the input resolution of $E$ $\\tilde{x}$ is the cropped aligned face region of $x$ . We add an adversarial loss $\\mathcal{L}_{\\mathrm{adv}}$ to improve the realism of the generated image.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sketch/mask-to-face translation. Given a real face $y$ as target and its sketch or parsing mask $x$ as source, we can train $E$ to translate $x$ to $y$ with Eq. (3) as objectives. In this task, we add a trainable light-weight translation network $T$ to map $x$ to an intermediate domain where $E$ can more easily extract features. For the style condition, $G$ 's first 7 layers use the latent code extracted from $x$ to provide structural styles, while its last 11 layers use the latent code from $\\tilde{y}$ to provide color and texture styles to simplify reconstruction. Therefore, $\\hat{y} = G(E_F(T(x),\\ell),E_W^{1:7}(T(x))\\oplus E_W^{8:18}(\\tilde{y}))$ , where $\\oplus$ is concatenation operation, and the superscript of $E_{W}$ means taking the $1\\sim 7$ or $8\\sim 18$ elements of $w^{+}$ . $\\ell = 1$ for sketch inputs and $\\ell = 3$ for mask inputs.", + "bbox": [ + 496, + 152, + 892, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video face editing. Given paired original face, edited face, and its editing vector $(x,y,v)$ (we can simply generate $x = G_0(w^+)$ and $y = G_0(w^+ +v)$ from random latent code $w^{+}$ with StyleGAN $G_{0})$ , we train $E$ for face editing with", + "bbox": [ + 496, + 347, + 890, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} (\\hat {y}, y) + \\lambda_ {5} \\mathcal {L} _ {\\text {a d v}} (\\hat {y}, y) + \\lambda_ {6} \\mathcal {L} _ {\\text {t m p}} (\\hat {y}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 416, + 890, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{y} = G(E_F(x,13),E_W(\\tilde{x}) + v)$ . $\\mathcal{L}_{\\mathrm{tmp}}$ is the flicker suppression loss [37] to improve temporal consistency.", + "bbox": [ + 496, + 439, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video toonification. For video face toonification, we have paired original face and toonified face $(x,y)$ . They can be generated as $x = G_0(w^+)$ and $y = G_0'(w^+)$ from random latent code $w^+$ following Toonify [22], where $G_0'$ is the StyleGAN fine-tuned on cartoon images. Let $G'$ denote StyleGANEX loaded with $G_0'$ , then we train $E$ using the objectives of Eq. (6) with $\\hat{y} = G'(E(x,\\tilde{x},13))$ .", + "bbox": [ + 496, + 470, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that compared to face editing based on StyleGANEX inversion in Sec. 4.2, the solution in this section does not require time-consuming latent optimization, and temporal consistency is enforced by the flicker suppression loss. Therefore, this solution is more suitable for efficient and coherent video face manipulation, which is a unique feature of the proposed framework. On the other hand, the solution based on StyleGANEX inversion is more flexible. There is no need to train new $E$ for every editing vector $\\mathcal{V}$ or fine-funed StyleGAN $G_0^{\\prime}$ .", + "bbox": [ + 496, + 577, + 890, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 739, + 705, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We follow pSp [23] to set $\\lambda_{2} = 1$ and $\\lambda_{3} = 0.8$ for all tasks, $\\lambda_{4} = 0.1$ for inversion task and 0 for other tasks. We set $\\lambda_{1}$ to 0.0001, 0.005 and 0 for inversion, sketch/mask-to-face and other tasks, respectively. We set $\\lambda_{5} = 0.1$ and $\\lambda_{6} = 30$ empirically. The translation network $T$ consists of two downsampling convolutional layers, two ResBlocks [11] and two upsampling convolutional layers, with small channel number 16. All experiments are performed using a single NVIDIA Tesla V100 GPU.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21004", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f83d168f36c42383e6e3745da39e5f7464610ef7b2e9158d33fd2dd6a8d3bd84.jpg", + "image_caption": [ + "Figure 7: StyleGANEX inversion and facial attribute/style editing." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 893, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We process FFHQ [17] to obtain 70,000 aligned training images of $1280 \\times 1280$ resolution for all tasks except two video-related tasks that use StyleGAN generated data. We use BiSeNet [39] to extract parsing masks and follow pSp [23] to extract sketches from face images. We augment all training data with random geometric transformations [15] like scaling, translation and rotation to make faces unaligned. We use images and videos from FaceForensics++ [25], Unsplash and Pexels as our testing dataset.", + "bbox": [ + 75, + 359, + 472, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Face Manipulation", + "text_level": 1, + "bbox": [ + 76, + 512, + 259, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Face editing. Figure 7 provides an overview of the performance of face inversion and attribute editing on StyleGANEX. We apply inversion to normal FoV face photos/paintings and use various editing vectors from InterFaceGAN [26] and LowRankGAN [40], and the pre-trained StyleGAN-NADA Ukiyo-e model [9], to edit the facial attributes or styles. As shown, these StyleGAN editing techniques work well on StyleGANEX. We also compare with pSp [23], HyperStyle [5] and BDInvert [14] in Fig. 8. Since these baselines are designed for cropped faces, we paste and blend their edited results back into the original image. For a fair comparison, we apply the same optimization method used in our approach to pSp for precise inversion (BDInvert is an optimization-based method and HyperStyle already uses extra hyper networks to simulate optimization). For editing that alters structures or colors, even precise inversion and blending cannot eliminate the obvious discontinuity along the seams, as indicated by the red arrows. In contrast, our approach processes the entire image as a whole and avoids such issues. Remarkably, our method successfully turns the whole hair into black in Fig.8(b), transfers the exemplar blond hairstyle onto the target face in Fig.8(c), and renders the full background with the StyleGAN-NADA Edvard Munch style in Fig.8(d).", + "bbox": [ + 75, + 537, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6669ec1c77690680d482c38522b323886d380ffc459d451c8af3f2f1ee13ec76.jpg", + "image_caption": [ + "Figure 8: Comparison on face editing." + ], + "image_footnote": [], + "bbox": [ + 503, + 359, + 890, + 577 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f730d692533d28b4b70deb9726abaf36ca981c41e1b3d1bfb3a2393acad16bc.jpg", + "image_caption": [ + "Figure 9: Comparison on super-resolution." + ], + "image_footnote": [], + "bbox": [ + 503, + 601, + 888, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Face super resolution. We show our $32 \\times$ super-resolution results in Fig. 9(d), where both the face and non-face regions are reasonably restored. We further follow pSp to train a single mode on multiple rescaling factors ( $4 \\sim 64$ ) with $\\ell = 3$ to make a fair comparison. In pSp's results, the non-face region is super-resolved by Real-ESRGAN [35]. As in Fig. 9(b)(c), our method surpasses pSp in precise de", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21005", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1616fe9e97e02c103eccd6dfe08e80a19631492f752484d6b753c32042a058ad.jpg", + "image_caption": [ + "(a) input" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 155, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e456360602ea4dcccad86764a72dbcfec5d1a006fcf0803d3ca526f3bb78dd78.jpg", + "image_caption": [ + "(b) pix2pixHD" + ], + "image_footnote": [], + "bbox": [ + 156, + 88, + 235, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb6c650148c34e52000471dce32ac06b8c14f093e85aec0170f7b884cf600fb0.jpg", + "image_caption": [ + "(c)TSIT" + ], + "image_footnote": [], + "bbox": [ + 235, + 88, + 310, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f9dd9dc017622ba2d31a410ae7df0dc138f3601f5b77218069c564d1a673caa.jpg", + "image_caption": [ + "(d) pSp" + ], + "image_footnote": [], + "bbox": [ + 313, + 88, + 387, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/829cfa69c6100131be56f5c7cd2d0bde87c3dec524c8c85dc83dc800fdafb168.jpg", + "image_caption": [ + "(e) ours" + ], + "image_footnote": [], + "bbox": [ + 388, + 89, + 464, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e895e6ba0d0312dd2829448cfeefb3879cbb9a316cc827ee5708dcea8cdc016.jpg", + "image_caption": [ + "Figure 10: Comparison on sketch-to-face translation.", + "(a) input", + "Figure 11: Comparison on mask-to-face translation." + ], + "image_footnote": [], + "bbox": [ + 81, + 282, + 156, + 448 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e78721fe8a8cc41b5f9d2744129cd9684a7a2bc145d7c33222cbb365d3335ce1.jpg", + "image_caption": [ + "(b) pix2pixHD" + ], + "image_footnote": [], + "bbox": [ + 156, + 282, + 233, + 448 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7bfc16a583e8c849ad292d1638e14283c18f49c3934f9915f2d6457c541594b.jpg", + "image_caption": [ + "(c)TSIT" + ], + "image_footnote": [], + "bbox": [ + 235, + 282, + 310, + 448 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa83c66ee39dcb666865ad262a9426322752ac73687eff840e099091123e34ef.jpg", + "image_caption": [ + "(d) pSp" + ], + "image_footnote": [], + "bbox": [ + 312, + 282, + 387, + 448 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d311c8b28bd2ba176ac7c5091a503effecf0043cba77c4e2ce14122ccd79c900.jpg", + "image_caption": [ + "(e) ours" + ], + "image_footnote": [], + "bbox": [ + 388, + 282, + 464, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tail restoration (e.g., glasses) and uniform super-resolution without discontinuity between face and non-face regions.", + "bbox": [ + 76, + 493, + 467, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sketch/mask to face translation. We compare our method with image-to-image translation models pix2pixHD [34] and TSIT [12], and StyleGAN-based pSp in Figs. 10-11. Pix2pixHD's results have many artifacts and monotonous colors. TSIT requires the inputs' side lengths to be divisible by 128. We find padding the input leads to failed translation. Therefore, we show its results on centrally cropped inputs, which are blurry. PSp generates realistic results, which are however less similar to the input sketch/mask. By comparison, our method can translate whole images and achieve realism and structural consistency to the inputs. For quantitative evaluation, we conduct a user study, where 30 subjects are invited to select what they consider to be the best results from the four methods. Each task uses eight results for evaluation. Table 1 summarizes the preference scores, where our method receives the best score.", + "bbox": [ + 76, + 523, + 467, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video face editing. We compare with pSp, HyperStyle, StyleHEAT [38] and STIT [31]. StyleHEAT uses features of a fixed $64 \\times 64$ resolution for unaligned but still cropped $1024 \\times 1024$ face reenactment. Specifically, it generates videos based on warping the features of the first frame, which however limits its inversion accuracy. STIT extends PTI [24] for full video processing by stitching. STIT cannot well preserve the complex hair details (Fig. 12, yellow box). As with image face editing, all four baselines are limited", + "bbox": [ + 76, + 763, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddbca3467ddba59a9c594f65c49979dca63a9010faf87bc7c1cfb6e6c41cf7ea.jpg", + "image_caption": [ + "black hair" + ], + "image_footnote": [], + "bbox": [ + 503, + 104, + 678, + 419 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49437caeddab0a76672e1813ff87ab0fc2ff0ae12244e4d92a095a97cf1aa20e.jpg", + "image_caption": [ + "young" + ], + "image_footnote": [], + "bbox": [ + 683, + 104, + 861, + 419 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1222658c3a436b757f348c1e635bda5771e68318ec976e81b3ed191c74d4d927.jpg", + "image_caption": [ + "editing control", + "Figure 12: Comparison on video face editing." + ], + "image_footnote": [], + "bbox": [ + 506, + 431, + 885, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/693498dde616d7c6f6af9058d117bb1c858d2e5029d5b866a09b2c59379f8732.jpg", + "image_caption": [ + "Figure 13: Comparison on video face toonification." + ], + "image_footnote": [], + "bbox": [ + 506, + 500, + 887, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to editing cropped regions, leading to discontinuity along the stitching seams. By comparison, our method uses the first-layer feature and skipped mid-layer features to provide spatial information, which achieves more coherent results. Moreover, we can randomly scale the editing vector $v$ (by multiplying a scale factor) instead of using a fixed $v$ during training. Then during testing, our method can flexibly adjust the editing degree by scaling $v$ for users to select as in Fig. 12(e).", + "bbox": [ + 496, + 732, + 890, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video face toonification. Compared with VToonify-T [37], our method preserves more details of the non-face region", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21006", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8b144bf1f3c778cd7693491167595294d5b7555ec55d9122e006c72bf168c7cd.jpg", + "table_caption": [ + "Table 1: User preference scores. Best scores are in bold." + ], + "table_footnote": [], + "table_body": "
Taskpix2pixHD [34]TSIT [12]pSp [23]ours
sketch-to-face0.4000.1210.0290.450
mask-to-face0.1080.0750.1210.696
TaskVToonify-T [37]ours
video toonify0.0830.917
", + "bbox": [ + 81, + 109, + 467, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/210fe0e65a5141b0967a19f8385e1d21ee85531e6beb7fadf6ef05e958367eb2.jpg", + "image_caption": [ + "(a) input" + ], + "image_footnote": [], + "bbox": [ + 81, + 203, + 156, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6efefed11ec12bea56cc5fbd6ad726f4ac81da1b0aa68c32059a4cbf56f52865.jpg", + "image_caption": [ + "(b) two steps" + ], + "image_footnote": [], + "bbox": [ + 158, + 203, + 232, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fae8f11de126cb43c57cb92dc05f35ecfe08c8d30c32da49821e4cd9ecf45c35.jpg", + "image_caption": [ + "(c) transferred (b)" + ], + "image_footnote": [], + "bbox": [ + 233, + 203, + 310, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cc38d86679334248b0dec497d9f6e71f236f8fcc889eb4aad7aa8b38a444e49.jpg", + "image_caption": [ + "(d) only opt (e)" + ], + "image_footnote": [], + "bbox": [ + 312, + 203, + 388, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee108ec916865ae682756eb3b9fece01c1674b6679fb40767cec0bd96bae4e29.jpg", + "image_caption": [ + "e) transferred (d)" + ], + "image_footnote": [], + "bbox": [ + 390, + 203, + 465, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc65cb43ba151b11045ed81f8781018dae063c9c77f382c540281d74af2c7f71.jpg", + "image_caption": [ + "Figure 14: Effect of encoder in StyleGANEX inversion.", + "(a) input $x_{\\uparrow}$", + "Figure 15: Input choice to provide valid style information." + ], + "image_footnote": [], + "bbox": [ + 81, + 301, + 176, + 375 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f68a7ad0ac727ff24559025487b3b406698d46ff9c6e4801ad5ee6d42320b89b.jpg", + "image_caption": [ + "(b) $G(E(x_{1},\\tilde{x}_{1},7))$" + ], + "image_footnote": [], + "bbox": [ + 179, + 301, + 272, + 375 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac8d38edda084a257454478ba5d355b2047f7639b85fe4a4012d53ed57020d9c.jpg", + "image_caption": [ + "(c) $G(E(x_{\\uparrow},x_{\\uparrow},7))$" + ], + "image_footnote": [], + "bbox": [ + 274, + 301, + 370, + 375 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b721664c3714187b6575c4df0b76b5f47b490481737bb9e49c947e87073dc619.jpg", + "image_caption": [ + "(d) $G(E(x_{\\uparrow},\\ddot{x}_{\\uparrow},7))$" + ], + "image_footnote": [], + "bbox": [ + 372, + 301, + 467, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and generates shaper faces in Fig. 13. The reason is that VToonify-T uses a fixed latent code extractor while our method trains a joint latent code and feature extractor, thus our method is more powerful for reconstructing the details. Moreover, our method retains StyleGAN's shallow layers, which helps provide key facial features to make the stylized face more vivid. Table 1 shows a quantitative comparison on ten results, and our method obtains the best user score.", + "bbox": [ + 75, + 424, + 468, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 558, + 228, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of Step II of our two-step inversion is verified in Fig. 6. We further study the effect of Step I in Fig. 14. With Step I providing a good prediction of $w^{+}$ and $f$ , Step II only needs 500-iteration optimization for precise reconstruction (Fig. 14(b)) and valid domain transfer to Disney Princess (Fig. 14(c)). However, if we directly optimize a mean $w^{+}$ and a random $f$ , the result is poor even with 2,000 iterations (indicated by a red arrow in Fig. 14(d)) and the optimized $w^{+}$ and $f$ overfit the input, which is not suitable for editing like domain transfer (Fig. 14(e)).", + "bbox": [ + 75, + 582, + 468, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 15, we study the input choice to extract $w^{+}$ . The cropped aligned faces are the default choice. If we instead use the whole image to extract $w^{+}$ , the background will affect the facial style, leading to poor restoration in Fig. 15(c). We further find reasonable results (Fig. 15(d)) can still be obtained by cropping the input to decrease the background proportion. Note that the face is not aligned in the cropped image $(\\ddot{x}_{\\uparrow})$ , which is useful for applications like super-resolution where extremely low-resolution faces are hard to detect and align. Users can simply manually crop the face region to provide valid style information.", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f092e80676832d6190f65952c4573c65f2b53b258515bf7fabcf531b563c76a.jpg", + "image_caption": [ + "(a) input" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 598, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc4b9ea92ef0092cc60c86678a90e29a14846df855fb7216b51f8ac46e8522df.jpg", + "image_caption": [ + "(b) $\\ell = 0$" + ], + "image_footnote": [], + "bbox": [ + 598, + 90, + 694, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d8c35664d3f2b23439b2792e4a39adbdf274cdefd6e5abdd8bc42d384a85ab79.jpg", + "image_caption": [ + "(c) $\\ell = 3$" + ], + "image_footnote": [], + "bbox": [ + 697, + 90, + 790, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9baacb02cca2dda7e70a2d8ccd85083315fba85c16200a4e2e1fcea822f05b3c.jpg", + "image_caption": [ + "(d) $\\ell = 7$" + ], + "image_footnote": [], + "bbox": [ + 792, + 90, + 888, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd53acac17a62b70bb19661bc4c53a13f6004012db5bdfaed976842ef321946e.jpg", + "image_caption": [ + "Figure 16: Effect of skip connections.", + "(a) input" + ], + "image_footnote": [], + "bbox": [ + 503, + 200, + 566, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f047f5d509967c9d03c4f6b32f918b2834de9073944e27bf76f5c316feb352c7.jpg", + "image_caption": [ + "(b) $\\ell = 1$" + ], + "image_footnote": [], + "bbox": [ + 568, + 200, + 632, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd48b4c7194423db2a9d78901a09ed522666921c19a5eb58719cd0f821986719.jpg", + "image_caption": [ + "(c) $\\ell = 0$" + ], + "image_footnote": [], + "bbox": [ + 635, + 200, + 689, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d13d7f6fa122bdef8a63cbe05bcf9a2656956ec19840216a0a5b78527881002.jpg", + "image_caption": [ + "(d) input" + ], + "image_footnote": [], + "bbox": [ + 694, + 200, + 759, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7b22fee3dbfdc405d85ced28e1a9c0e422fd2a99fb86cd90b46104fd44aadec.jpg", + "image_caption": [ + "(e) $\\ell = 1$" + ], + "image_footnote": [], + "bbox": [ + 764, + 200, + 825, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c1539a8d7a99bae4d56f199870a06f9b58121754f9643206271a6abd7997dc7.jpg", + "image_caption": [ + "(f) $\\ell = 0$" + ], + "image_footnote": [], + "bbox": [ + 826, + 200, + 890, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a39666f4cd49bc3378b174e838d9b1b5fd21e16c893141d32f814e4bf6fb8f2c.jpg", + "image_caption": [ + "Figure 17: Performance on low-quality sketches.", + "input" + ], + "image_footnote": [], + "bbox": [ + 504, + 291, + 596, + 400 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bea45d4ca4152e69942dd8b1186ed07fa0af5bc4ff0501df2e9562b11ef1cc4c.jpg", + "image_caption": [ + "inversion" + ], + "image_footnote": [], + "bbox": [ + 601, + 292, + 694, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3d09525ba003773acf8646011e375b0ca5950b4ab8a157568f70c7830ac61d5.jpg", + "image_caption": [ + "blocky shape" + ], + "image_footnote": [], + "bbox": [ + 697, + 292, + 790, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/427ecd39e97c8dff0b94308ad0ea88dbeb3be7bbff0e0b692b1d29a531471d5c.jpg", + "image_caption": [ + "add grass" + ], + "image_footnote": [], + "bbox": [ + 794, + 292, + 888, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b966d440e3da220da2920b84aa67f6a3c7d7fc2606024cdfb02fd69e912a4afb.jpg", + "image_caption": [ + "redness", + "Figure 18: Performance on car." + ], + "image_footnote": [], + "bbox": [ + 601, + 420, + 694, + 529 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c4defc0a8ba0803d5b2f98a4b685dd6a6897496da9f8d0990b91a6c79bf166e.jpg", + "image_caption": [ + "style mixing" + ], + "image_footnote": [], + "bbox": [ + 697, + 420, + 792, + 529 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a498d13ebcd693cb8de3eaa147743abc9723ace36bc5cdf625ba70da4509aac4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 420, + 888, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 16, we study the effect of skip connection. Without it $(\\ell = 0)$ , the glasses cannot be generated. Skip connection provides mid-layer features to preserve the details of the input. However, too many extra features will alter the feature distribution of StyleGAN, e.g., always generating sunglasses as the input has no segmentation of eyes. Thus, we use $\\ell = 3$ to balance between input-output consistency and diversity. Inversely, we can use a small $\\ell$ to enhance the model robustness to low-quality inputs. For example, we can generate more realistic faces with $\\ell = 0$ on DeepFace-Drawing low-quality sketches [7] as in Fig. 17.", + "bbox": [ + 496, + 590, + 890, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Results on Non-Facial Dataset", + "text_level": 1, + "bbox": [ + 500, + 770, + 766, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed refactoring is domain-agnostic, thus it can be applied to StyleGAN pre-trained on other domains like cars. An example with a wide vertical field of view is shown in Fig. 18, where we perform optimization-based StyleGAN inversion, image editing with the editing vectors found by GANSpace [10] and style mixing on the refactored model. The resulting manipulated cars look plausible.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21007", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d1d5cbd21f6323168486be213ea2f578478608e78722b2310251bdbd0e23a5c.jpg", + "image_caption": [ + "(c) severely rotated face", + "Figure 19: Limitations." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 218 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.4. Limitations", + "text_level": 1, + "bbox": [ + 76, + 265, + 202, + 279 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "First, our framework currently relies on an inefficient optimization process for precise reconstruction. While future work can explore more efficient inversion methods (e.g., iterative residue prediction and hyper networks), it is important to note that this paper focuses on overcoming the fixed-crop limitation of StyleGAN, rather than the specific topic of GAN inversion. Second, StyleGANEX is limited by the feature representation of StyleGAN. While it shows great potential in handling normal FoV face images, out-of-distribution features such as complex clothing and human bodies may not be well handled as in Fig. 19(a)(b). As in Fig. 19(c), while our method can handle faces rotated 20 degrees, it still struggles with handling large rotation angles. However, this can be easily resolved by rough alignment of the input image in the middle as our method does not require accurate alignment. Finally, after translation, we can undo the alignment to obtain a plausible result as in right of Fig. 19(c). Third, StyleGANEX, like StyleGAN, focuses on face manipulation and may not support out-of-distribution semantical editing of non-facial regions. Last, StyleGANEX may inherit the model bias of StyleGAN. Applying it to tasks with severe data imbalance might lead to unsatisfactory results on under-represented data.", + "bbox": [ + 76, + 287, + 473, + 636 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 648, + 194, + 664 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have presented an effective approach to refactor StyleGAN to overcome its fixed-crop limitation while retaining its style control abilities. The refactored model, called StyleGANEX, fully inherits the parameters of the pre-trained StyleGAN without retraining, and is thus fully compatible with the generative space of StyleGAN. We further introduced a StyleGANEX encoder to project normal FoV face images to the joint $W^{+} - F$ space of StyleGANEX for real face inversion and manipulation. Our approach offers an effective solution to the general issue of StyleGAN and extends its capability beyond fixed-resolution data.", + "bbox": [ + 75, + 674, + 468, + 839 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is al", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "so supported by MOE AcRF Tier 1 (2021-T1-001-088) and NTU NAP grant.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 138, + 596, + 154 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proc. Int'l Conf. Computer Vision, pages 4432-4441, 2019. 1, 2", + "[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2", + "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinements. In Proc. Int'l Conf. Computer Vision, pages 6711–6720, 2021. 2", + "[4] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3, 2022. 2", + "[5] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 2, 6", + "[6] Kelvin CK Chan, Xintao Wang, Xiangyu Xu, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for large-factor image super-resolution. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2021. 3, 4", + "[7] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. ACM Transactions on Graphics, 39(4):72-1, 2020. 8", + "[8] Tan M Dinh, Anh Tuan Tran, Rang Nguyen, and Binh-Son Hua. Hyperinverter: Improving stylegan inversion via hypernetwork. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11389-11398, 2022. 2", + "[9] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clipguided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 5, 6", + "[10] Erik Härkönen, Aaron Hertzman, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Advances in Neural Information Processing Systems, 2020. 1, 2, 3, 8", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 770-778, 2016. 5", + "[12] Liming Jiang, Changxu Zhang, Mingyang Huang, Chunxiao Liu, Jianping Shi, and Chen Change Loy. Tset: A simple and versatile framework for image-to-image translation. In Proc. European Conf. Computer Vision, pages 206-222. Springer, 2020. 7, 8" + ], + "bbox": [ + 501, + 165, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21008", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In Proc. Int'l Conf. Computer Vision, 2021. 2", + "[14] Kyoungkook Kang, Seongtae Kim, and Sunghyun Cho. Gan inversion for out-of-range images with geometric transformations. In Proc. Int'l Conf. Computer Vision, pages 13941-13949, 2021. 3, 6", + "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems, 2020. 6", + "[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 3, 6", + "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1", + "[19] Feng-Lin Liu, Shu-Yu Chen, Yukun Lai, YueRen Jiang, Hongbo Fu, and Lin Gao. Deepfacevideoediting: Sketch-based deep editing of face videos. ACM Transactions on Graphics, 41(4):167, 2022. 2", + "[20] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, JunYan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11399-11409, 2022. 3", + "[21] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 2085–2094, October 2021. 2", + "[22] Justin NM Pinkney and Doron Adler. Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334, 2020. 5", + "[23] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 8", + "[24] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics, 2022. 2, 7", + "[25] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. In Proc. Int'l Conf. Computer Vision, 2019. 6", + "[26] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 1, 2, 3, 5, 6" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2", + "[28] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 2, 3", + "[29] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics, 39(6):1-14, 2020. 1, 2", + "[30] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 2", + "[31] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia, pages 1–9, 2022. 2, 7", + "[32] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. Stylegan2 distillation for feed-forward image manipulation. In Proc. European Conf. Computer Vision, pages 170-186. Springer, 2020. 3", + "[33] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 2, 3", + "[34] Ting Chun Wang, Ming Yu Liu, Jun Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2018. 7, 8", + "[35] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proc. Int'l Conf. Computer Vision, pages 1905-1914, 2021. 6", + "[36] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. StyleAlign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 1", + "[37] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Vtoonify: Controllable high-resolution portrait video style transfer. ACM Transactions on Graphics, 41(6):1-15, 2022. 2, 3, 4, 5, 7, 8", + "[38] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In Proc. European Conf. Computer Vision, pages 85-101. Springer, 2022. 3, 7", + "[39] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proc. European Conf. Computer Vision, pages 334-349. Springer, 2018. 6", + "[40] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zheng-Jun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21009", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in gans. In Advances in Neural Information Processing Systems, volume 34, 2021. 1, 3, 6", + "[41] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proc. European Conf. Computer Vision, pages 592-608. Springer, 2020. 2", + "[42] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Transactions on Graphics, 40(6):1-13, 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21010", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/71e23834-fb4f-4c96-b627-797882a6ea6f_model.json b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/71e23834-fb4f-4c96-b627-797882a6ea6f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..57d590cf3d8ce67c8c4d64797dea6b1352c71c4e --- /dev/null +++ b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/71e23834-fb4f-4c96-b627-797882a6ea6f_model.json @@ -0,0 +1,3137 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.88, + 0.155 + ], + "angle": 0, + "content": "StyleGANEX: StyleGAN-Based Manipulation Beyond Cropped Aligned Faces" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.177, + 0.751, + 0.214 + ], + "angle": 0, + "content": "Shuai Yang Liming Jiang Ziwei Liu Chen Change Loy S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.216, + 0.715, + 0.232 + ], + "angle": 0, + "content": "{shuai.yang, liming002, ziwei.liu, ccloy}@ntu.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.248, + 0.681, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.248, + 0.889, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.352, + 0.48, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.351, + 0.89, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.448, + 0.89, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.558, + 0.89, + 0.575 + ], + "angle": 0, + "content": "Figure 1: We expand StyleGAN to encompass a diverse set of tasks that go beyond the constraints of cropped aligned faces." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.587, + 0.314, + 0.602 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.474, + 0.883 + ], + "angle": 0, + "content": "Recent advances in face manipulation using StyleGAN have produced impressive results. However, StyleGAN is inherently limited to cropped aligned faces at a fixed image resolution it is pre-trained on. In this paper, we propose a simple and effective solution to this limitation by using dilated convolutions to rescale the receptive fields of shallow layers in StyleGAN, without altering any model parameters. This allows fixed-size small features at shallow layers to be extended into larger ones that can accommodate variable resolutions, making them more robust in characterizing unaligned faces. To enable real face inversion and manipulation, we introduce a corresponding encoder that provides the first-layer feature of the extended StyleGAN in addition to the latent style code. We validate the effectiveness of our method using unaligned face inputs of various resolutions in a diverse set of face manipulation tasks, including facial attribute editing, super-resolution, sketch/mask-to-face translation, and face toonification. Project page https:" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.884, + 0.458, + 0.899 + ], + "angle": 0, + "content": "//www.mmlab-ntu.com/project/styleganex" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.587, + 0.633, + 0.603 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.894, + 0.794 + ], + "angle": 0, + "content": "StyleGAN [17, 18] has emerged as one of the most successful models for generating high-quality faces. Building upon StyleGAN, researchers have developed a range of face manipulation models [1, 23, 29, 26, 10, 28, 40, 36]. These models typically map real face images or other face-related inputs to the latent space of StyleGAN, perform semantic editing in the latent space, and then map the edited latent code back to the image space. This approach enables a variety of tasks, including facial attribute editing, face restoration, sketch-to-face translation, and face toonification. As the manipulated faces remain within the generative space of StyleGAN, the quality of the image output is guaranteed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Despite its ability to ensure high-quality image output, the generative space of StyleGAN is limited by a fixed-crop constraint that restricts image resolution and face layout. As a result, existing face manipulation models based on StyleGAN can only handle cropped and aligned face images. In such images with a limited field of view (FoV), the face typically dominates the image, leaving little room for back-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21000" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "ground and clothing, and often resulting in partially cropped hair. However, in everyday portrait photos such as selfies, faces occupy a smaller proportion of the image, allowing for a complete hairstyle and upper body seen. Portrait videos such as those from live streaming require an even larger background area to accommodate face movement. To process these types of inputs, which we refer to as normal FoV face images and videos, existing manipulation models need to align, crop, and edit the face before pasting the result back onto the original image [4, 19, 31]. This approach often results in discontinuity near the seams, e.g., only editing the hair color inside the cropped area." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.518 + ], + "angle": 0, + "content": "While StyleGAN3 [16] was introduced to address unaligned faces, a recent study [4] found that even StyleGAN3 requires face realignment before effectively projecting to its latent space. Moreover, StyleGAN3 is still constrained by a fixed image resolution. Motivated by the translation equivariance of convolutions, VToonify [37] addresses the fixed-crop limitation of StyleGAN by removing its shallow layers to accept input features of any resolution. However, these shallow layers are crucial for capturing high-level features of the face, such as pose, hairstyle, and face shape. By removing these layers, the network loses its ability to perform latent editing on these important features, which is a distinctive capability of StyleGAN. Therefore, the challenge remains in overcoming the fixed-crop limitation of StyleGAN while preserving its original style manipulation abilities, which is a valuable research problem to solve." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.793 + ], + "angle": 0, + "content": "In this paper, we propose a simple yet effective approach for refactoring StyleGAN to overcome the fixed-crop limitation. In particular, we refactor its shallow layers instead of removing them, allowing the first layer to accept input features of any resolution. This simple change expands StyleGAN's style latent space into a more powerful joint style latent and first-layer feature space \\((W^{+} - F\\) space), extending the generative space beyond cropped aligned faces. Furthermore, our refactoring only changes the receptive field of shallow-layer convolutions, leaving all pre-trained model parameters intact. Hence, the refactored StyleGAN (StyleGANEX) can directly load the original StyleGAN parameters, fully compatible with the generative space of StyleGAN, and retains its style representation and editing ability. This means that the StyleGAN editing vectors found in previous studies [26, 10, 28] can be directly applied to StyleGANEX for normal FoV face editing, e.g., changing the face pose, as shown in Fig. 1(a)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Based on StyleGANEX, we further design a corresponding encoder that projects normal FoV face images to the \\( W^{+} - F \\) space for real face inversion and manipulation. Our encoder builds upon pSp encoder [23] and aggregates its multi-layer features to predict the first-layer feature of StyleGANEX. The encoder and StyleGANEX form a fully convolutional encoder-decoder framework. With the first-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "layer feature as the bottleneck layer, whose resolution is \\(1/32\\) of the output image, our framework can handle images and videos of various resolutions, as long as their side lengths are divisible by 32. Depending on the input and output types, our framework can perform a wide range of face manipulation tasks. In this paper, we select several representative tasks, as shown in Fig. 1, including facial attribute editing, face super-resolution, sketch/mask-to-face translation and video face toonification. While the focus of these tasks in the past is limited to cropped aligned faces, our framework can handle normal FoV faces, showing significant advantages over previous StyleGAN-based approaches. To summarize, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.291, + 0.892, + 0.337 + ], + "angle": 0, + "content": "- A novel StyleGANEX architecture with extended \\( W^{+} - F \\) space, which overcomes the fixed-crop limitation of StyleGAN." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.34, + 0.892, + 0.371 + ], + "angle": 0, + "content": "- An effective encoder that is able to project normal FoV face images into the \\( W^{+} - F \\) domain." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.374, + 0.892, + 0.405 + ], + "angle": 0, + "content": "- A generic and versatile fully convolutional framework for face manipulation beyond cropped aligned faces." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.291, + 0.892, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.642, + 0.436 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.895, + 0.735 + ], + "angle": 0, + "content": "StyleGAN inversion. StyleGAN inversion aims at projecting real face images into the latent space of StyleGAN for further manipulation. Image2StyleGAN [1] analyzes the latent space and proposes \\( W^{+} \\) space to reconstruct real faces with latent code optimization. PIE [29] and IDinvert [41] further consider the editability of the latent code during optimization. To speed up inversion, pSp [23] and e4e [30] train an encoder to directly project the target face to its corresponding latent code, which is however hard to reconstruct fine details and handle occlusions. To solve this issue, Restyle [3] and HFGI [33] predict the residue of latent codes or mid-layer features to reduce errors, respectively. Instead of focusing on the latent code, PTI [24] optimizes StyleGAN itself to fit a target face, which is accelerated by HyperInverter [8] and HyperStyle [5] to predict offsets of network parameters with hyper networks. The above methods are limited to cropped aligned faces for valid face editing. With the extended \\( W^{+} - F \\) space, our framework is able to perform inversion on normal FoV face images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "StyleGAN-based face manipulation. An intuitive way of StyleGAN-based face manipulation is to optimize the latent code online to achieve certain objectives such as a pixel-level constraints [1], sketch-based structure constraints [19] or text-guided semantic constraints [21, 9]. Another way is to search for offline editing vectors to add to the latent code for manipulation. Supervised methods identify meaningful editing vectors based on attribute labels or pretrained classifiers [26, 27, 13, 2]. On the other hand, unsupervised methods statistically analyze the StyleGAN latent space to discover semantically significant editing direc" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "21001" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.394 + ], + "angle": 0, + "content": "tions by principal component analysis [10], low-rank factorization [40] and closed-form factorization [28]. Meanwhile, face manipulation can be directly realized by image-to-image translation frameworks, where StyleGAN is used to generate paired training data [32, 37] or to build the decoder [23, 6, 37]. To achieve spatial editing, methods [42, 20] are proposed to manipulate mid-layer features in addition to the latent code, similar to our \\( W^{+} - F \\) space. Moreover, BDInvert [14] and StyleHEAT [38] introduces feature transformations for unaligned face editing. However, the above methods, as well as HFGI [33], follow StyleGAN features' original fixed resolution, thus still suffering the crop limitation. Differently, our method predicts the first-layer feature that can have various resolutions in StyleGANEX. Moreover, as we will show later in Sec. 5.1, compared to VToonify [37], StyleGANEX retains the complete style manipulation abilities of the shallow layers of StyleGAN, with a jointly trained latent code and feature extractor, preserving vivid details and supporting more diverse manipulation tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.411, + 0.216, + 0.427 + ], + "angle": 0, + "content": "3. StyleGANEX" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.437, + 0.406, + 0.453 + ], + "angle": 0, + "content": "3.1. Analysis of the Fixed-Crop Limitation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.462, + 0.47, + 0.612 + ], + "angle": 0, + "content": "StyleGAN has great potential for handling normal FoV face images. The generator of StyleGAN is a fully convolutional architecture that can naturally handle different feature resolutions, as convolution operations and the style modulation of StyleGAN are independent of the input resolution. Additionally, the translation equivariance of convolution operations naturally supports feature translation. As analyzed in VToonify [37], if we translate or rotate the feature of the 7th layer of StyleGAN, the resulting face will also be shifted or rotated, as shown in Figs. 2(a)(b)(d)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "The limiting factor originates from StyleGAN's constant first-layer feature. First, the first-layer feature has a fixed resolution of \\(4 \\times 4\\), limiting the output to \\(1024 \\times 1024\\) resolution. Second, \\(4 \\times 4\\) resolution is inadequate to characterize the spatial information of unaligned faces. We have taken a step further than the analysis in VToonify [37] to investigate translation and rotation on the first-layer feature. As shown in Fig. 2(c), sub-pixel translation fuses adjacent feature values severely, resulting in a blurry face due to the small number of elements (only 16) in the first-layer feature. In Fig. 2(e), the first-layer feature fails to provide enough spatial information for a valid rotation. In comparison, the 7-th layer has a higher resolution (\\(32 \\times 32\\)), making it better suited for capturing spatial information. However, only a single layer alone provides limited style control, as the full facial structural styles are hierarchically modeled by seven shallow layers. Simply ignoring low-resolution layers, as in VToonify, disables flexible latent editing over the styles of these layers, such as pose, age, and face shape." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.583, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.15, + 0.574, + 0.161 + ], + "angle": 0, + "content": "(a) random face" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.091, + 0.658, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.15, + 0.653, + 0.161 + ], + "angle": 0, + "content": "(b) shift layer#7" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.091, + 0.734, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.15, + 0.728, + 0.161 + ], + "angle": 0, + "content": "(c) shift layer#1" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.091, + 0.812, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.15, + 0.81, + 0.161 + ], + "angle": 0, + "content": "(d) rotate layer#7" + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.091, + 0.891, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.15, + 0.887, + 0.161 + ], + "angle": 0, + "content": "(e) rotate layer#1" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.17, + 0.894, + 0.261 + ], + "angle": 0, + "content": "Figure 2: Analysis of StyleGAN in generating unaligned faces. (a) Generated face. (b)(c) Face generated by translating the feature maps of the 7-th and 1-st layers of StyleGAN respectively to shift the face by 150 pixels. (d)(e) Face generated by rotating the feature maps of the 7-th and 1-st layers of StyleGAN by 10 degrees, respectively." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.268, + 0.599, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.338, + 0.598, + 0.35 + ], + "angle": 0, + "content": "(a) real face photo" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.269, + 0.697, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.338, + 0.692, + 0.35 + ], + "angle": 0, + "content": "(b) projected face" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.269, + 0.794, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.338, + 0.783, + 0.35 + ], + "angle": 0, + "content": "(c) shift layer#1" + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.269, + 0.891, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.338, + 0.885, + 0.35 + ], + "angle": 0, + "content": "(d) rotate layer#1" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.359, + 0.893, + 0.45 + ], + "angle": 0, + "content": "Figure 3: Analysis of StyleGANEX in generating unaligned faces. (a) A real face photo. (b) Reconstructed face by projecting (a) into the \\( W^{+} - F \\) space of StyleGANEX. (c) Face generated by translating the first-layer feature map to shift the face by 300 pixels. (d) Face generated by rotating the first-layer feature map by 10 degrees." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.471, + 0.893, + 0.729 + ], + "angle": 0, + "content": "Expanding the shallow layers of StyleGAN to have the same \\(32 \\times 32\\) resolution as the 7th layer, or more generally, \\(H / 32 \\times W / 32\\) resolution for an \\(H \\times W\\) image, would provide enough structure and layout information to combine the style controllability of shallow layers with support for normal FoV faces. This is precisely the key idea of StyleGANEX, and we will introduce our simple solution in Section 3.2. As a preview of the performance of the expanded layers in enabling face manipulation beyond cropped and aligned faces, Fig. 3 shows that for a \\(1472 \\times 1600\\) normal FoV face photo, we can obtain its latent code and an additional \\(46 \\times 50\\) first-layer feature as the input to StyleGANEX (inversion method explained in Section 4.2). Face translation and rotation can be realized by shifting or rotating the first-layer feature. Additionally, the face can be effectively edited by applying style mixing [17] or InterFaceGAN editing vectors [26] to the latent code, as shown in Fig. 1(a)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.788, + 0.757 + ], + "angle": 0, + "content": "3.2. From StyleGAN to StyleGANEX" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Figure 4 illustrates the generator architectures of StyleGAN and StyleGANEX. Compared to StyleGAN, we first replace the constant \\(4 \\times 4\\) first-layer feature with a variable feature whose resolution is \\(1/32\\) of the output image. Then, we remove the upsample operations before the 8-th layer, allowing features in the seven shallow layers to share the same resolution as the 7-th layer. However, the convolution kernels or reception fields of these layers do not match their input features with the enlarged resolution. To solve" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21002" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.262, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.089, + 0.468, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.365, + 0.47, + 0.397 + ], + "angle": 0, + "content": "Figure 4: Refactor StyleGAN to StyleGANEX. For simplicity, learned weights, biases and noises are omitted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.55 + ], + "angle": 0, + "content": "this problem, we enlarge the reception fields by modifying the convolutions to their dilated versions. For example, the first layer only needs to change the dilation factor from 1 to 8. With this simple modification, StyleGAN is refactored to StyleGANEX. Since the first layer becomes variable, the original \\( W^{+} \\) latent space is extended to a joint \\( W^{+} - F \\) space, where the latent code \\( w^{+} \\in W^{+} \\) provides style cues, and the first-layer feature \\( f \\in F \\) mainly encodes spatial information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.471, + 0.81 + ], + "angle": 0, + "content": "The refactoring of StyleGAN to StyleGANEX has three key advantages. 1) Support for unaligned faces. The resolution enlargement and variable first-layer features of StyleGANEX overcome the fixed-crop limitation. 2) Compatibility. No model parameters are altered during refactoring, meaning that StyleGANEX can directly load pretrained StyleGAN parameters without retraining. In fact, if we upsample the StyleGAN's constant input feature by \\(8 \\times\\) with nearest neighbor interpolation to serve as \\(f\\) of StyleGANEX, StyleGANEX degrades exactly to StyleGAN with the same \\(1024 \\times 1024\\) generative space. The computational cost of the refactoring is also minimal, with generating an image taking 0.026s and 0.028s for StyleGAN and StyleGANEX, respectively. 3) Flexible manipulation. StyleGANEX retains the style representation and editing ability of StyleGAN, meaning that abundant StyleGAN-based face manipulation techniques can be applied to StyleGANEX." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.419, + 0.839 + ], + "angle": 0, + "content": "4. Face Manipulation with StyleGANEX" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.287, + 0.863 + ], + "angle": 0, + "content": "4.1. StyleGANEX Encoder" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This section introduces our StyleGANEX encoder \\(E\\) which is used to project real face images into the \\(W^{+} - F\\)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.315, + 0.845, + 0.33 + ], + "angle": 0, + "content": "Figure 5: Details of StyleGANEX Encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.55 + ], + "angle": 0, + "content": "space of StyleGANEX \\( G \\). The encoder builds upon the pSp encoder [23], as depicted in Fig. 5. Specifically, for \\( F \\) space, we concatenate pSp features in the middle layers and add a convolution layer to map the concatenated features to the first-layer input feature \\( f \\) of \\( G \\). For \\( W^{+} \\) space, the original pSp encoder takes a \\( 256 \\times 256 \\) image as input and convolves it to eighteen \\( 1 \\times 1 \\times 512 \\) features to map to a latent code \\( w^{+} \\in \\mathbb{R}^{18 \\times 512} \\). To make \\( E \\) accept more general \\( H \\times W \\) images, we add global average pooling to resize all features to \\( 1 \\times 1 \\times 512 \\) before mapping to latent codes. To support various face manipulation tasks flexibly, we can extract \\( f \\) and \\( w^{+} \\) from different sources. Let \\( E_{F} \\) and \\( E_{W} \\) be the operation of \\( E \\) to extract the first-layer feature and the latent code, respectively. We have" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.56, + 0.891, + 0.577 + ], + "angle": 0, + "content": "\\[\nf, w ^ {+} = E _ {F} \\left(x _ {1}\\right), E _ {W} \\left(x _ {2}\\right) := E \\left(x _ {1}, x _ {2}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.725 + ], + "angle": 0, + "content": "where \\( x_{1} \\) and \\( x_{2} \\) are the source inputs for face layout and face style, respectively. Then, a general form of image generation by \\( G \\) from \\( x_{1} \\) and \\( x_{2} \\) is \\( G(E_{F}(x_{1}), E_{W}(x_{2})) \\). In some face manipulation tasks like super-resolution [6] and toonification [37], passing encoder features to the generator via skip connections helps preserve the details of the input image. We thus introduce a scalar parameter \\( \\ell \\) to the generation process, indicating the \\( \\ell \\) shallow layers of \\( G \\) receive the encoder features \\( (\\ell = 0 \\) means no skip connections):" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.735, + 0.891, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = G \\left(E _ {F} \\left(x _ {1}, \\ell\\right), E _ {W} \\left(x _ {2}\\right)\\right) := G \\left(E \\left(x _ {1}, x _ {2}, \\ell\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.891, + 0.824 + ], + "angle": 0, + "content": "where \\(E_{F}(x_{1},\\ell)\\) provides both \\(f\\) and the skipped encoder features. For \\(x_{1}\\) of \\(H\\times W\\) resolution, \\(f\\) and the generated image \\(\\hat{x}\\) will be of \\(H / 8\\times W / 8\\) and \\(4H\\times 4W\\) , respectively. The resolution of \\(x_{2}\\) can be independent of \\(x_{1}\\) and \\(\\hat{x}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.81, + 0.849 + ], + "angle": 0, + "content": "4.2. StyleGANEX Inversion and Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To find appropriate \\(\\hat{f}\\) and \\(\\hat{w}^+\\) that precisely reconstruct a target image \\(x\\), we perform a two-step StyleGANEX inversion. Step I projects \\(x\\) to initial \\(f\\) and \\(w^+\\) with \\(E\\). Step" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21003" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.177, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.166, + 0.165, + 0.179 + ], + "angle": 0, + "content": "(a) face photo" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.091, + 0.273, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.166, + 0.268, + 0.179 + ], + "angle": 0, + "content": "(b) encoded face" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.091, + 0.369, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.166, + 0.36, + 0.179 + ], + "angle": 0, + "content": "(c) (b)+optimize" + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.091, + 0.465, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.166, + 0.468, + 0.179 + ], + "angle": 0, + "content": "(d) domain transfer" + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.184, + 0.392, + 0.2 + ], + "angle": 0, + "content": "Figure 6: StyleGANEX inversion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.47, + 0.259 + ], + "angle": 0, + "content": "II optimizes \\( f \\) and \\( w^{+} \\) to further reduce the reconstruction error. The training of \\( E \\) follows pSp with reconstruction losses and a regularization loss [23]:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.27, + 0.469, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} (\\hat {x}, x) + \\lambda_ {1} \\mathcal {L} _ {\\text {r e g}} \\left(E _ {W} (\\tilde {x})\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.47, + 0.419 + ], + "angle": 0, + "content": "where \\(\\hat{x} = G(E(x, \\tilde{x}, 0))\\) and \\(\\tilde{x}\\) is the cropped aligned face region of \\(x\\). We empirically find that using \\(\\tilde{x}\\) instead of \\(x\\) predicts more accurate \\(w^{+}\\) since StyleGAN is originally trained on cropped aligned faces. \\(\\mathcal{L}_{\\mathrm{reg}}\\) encourages the predicted \\(w^{+}\\) closer to the average latent code to improve the image quality. \\(\\mathcal{L}_{\\mathrm{rec}}\\) measures the distance between the reconstructed \\(\\hat{x}\\) and the target \\(x\\) in terms of pixel similarity, perceptual similarity, and identity preservation:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.429, + 0.466, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e c}} (\\hat {x}, x) = \\lambda_ {2} \\mathcal {L} _ {2} (\\hat {x}, x) + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {L P I P S}} (\\hat {x}, x) + \\lambda_ {4} \\mathcal {L} _ {\\mathrm {I D}} (\\hat {x}, x).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.502 + ], + "angle": 0, + "content": "As shown in Fig. 6(b), \\(\\hat{x}\\) largely approximates \\(x\\). But the background details and clothings are still hard to reconstruct. Therefore, we further optimize \\(f\\) and \\(w^{+}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.511, + 0.469, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\hat {f}, \\hat {w} ^ {+} = \\underset {f, w ^ {+}} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\mathrm {L P I P S}} (G (f, w ^ {+}), x), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.469, + 0.582 + ], + "angle": 0, + "content": "where \\(f\\) and \\(w^{+}\\) are initialized by \\(E(x,\\tilde{x},0)\\). The optimized \\(G(\\hat{f},\\hat{w}^{+})\\) in Fig. 6(c) well reconstructs \\(x\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.703 + ], + "angle": 0, + "content": "After inversion, we can perform flexible editing over \\( x \\) as in StyleGAN. Figure 1(a) shows two examples: we can exchange the last 11 elements of \\( \\hat{w}^{+} \\) with random samples, to mix the color and texture styles; we can add InterFaceGAN editing vectors [26] to \\( \\hat{w}^{+} \\) to make a young face. Moreover, as shown in Fig. 6(d), if we load \\( G \\) a pre-trained StyleGAN-NADA Disney Princess model [9] (let \\( G' \\) denote the new \\( G \\)), we can obtain \\( G'(\\hat{f}, \\hat{w}^{+}) \\), a cartoon version of \\( x \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.363, + 0.727 + ], + "angle": 0, + "content": "4.3. StyleGANEX-Based Translation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.825 + ], + "angle": 0, + "content": "The encoder and StyleGANEX form an end-to-end image-to-image translation framework in Fig. 5. Depending on the type of paired training data, it can be trained to efficiently realize different face manipulation tasks. As with pSp [23], we will fix StyleGANEX generator and only train the encoder on the given task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Face super-resolution. Given low-resolution and high-resolution training image pairs \\((x,y)\\), we can train \\(E\\) to recover \\(y\\) from \\(x\\) to learn face super-resolution with the loss" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} (\\hat {y}, y) + \\lambda_ {5} \\mathcal {L} _ {\\text {a d v}} (\\hat {y}, y), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "where \\(\\hat{y} = G(E(x_{\\uparrow},\\tilde{x}_{\\uparrow},7))\\) \\(\\uparrow\\) is the upsample operation to make \\(x\\) match the input resolution of \\(E\\) \\(\\tilde{x}\\) is the cropped aligned face region of \\(x\\) . We add an adversarial loss \\(\\mathcal{L}_{\\mathrm{adv}}\\) to improve the realism of the generated image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.348 + ], + "angle": 0, + "content": "Sketch/mask-to-face translation. Given a real face \\(y\\) as target and its sketch or parsing mask \\(x\\) as source, we can train \\(E\\) to translate \\(x\\) to \\(y\\) with Eq. (3) as objectives. In this task, we add a trainable light-weight translation network \\(T\\) to map \\(x\\) to an intermediate domain where \\(E\\) can more easily extract features. For the style condition, \\(G\\)'s first 7 layers use the latent code extracted from \\(x\\) to provide structural styles, while its last 11 layers use the latent code from \\(\\tilde{y}\\) to provide color and texture styles to simplify reconstruction. Therefore, \\(\\hat{y} = G(E_F(T(x),\\ell),E_W^{1:7}(T(x))\\oplus E_W^{8:18}(\\tilde{y}))\\), where \\(\\oplus\\) is concatenation operation, and the superscript of \\(E_{W}\\) means taking the \\(1\\sim 7\\) or \\(8\\sim 18\\) elements of \\(w^{+}\\). \\(\\ell = 1\\) for sketch inputs and \\(\\ell = 3\\) for mask inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.409 + ], + "angle": 0, + "content": "Video face editing. Given paired original face, edited face, and its editing vector \\((x,y,v)\\) (we can simply generate \\(x = G_0(w^+)\\) and \\(y = G_0(w^+ +v)\\) from random latent code \\(w^{+}\\) with StyleGAN \\(G_{0})\\) , we train \\(E\\) for face editing with" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.417, + 0.892, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} (\\hat {y}, y) + \\lambda_ {5} \\mathcal {L} _ {\\text {a d v}} (\\hat {y}, y) + \\lambda_ {6} \\mathcal {L} _ {\\text {t m p}} (\\hat {y}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.471 + ], + "angle": 0, + "content": "where \\(\\hat{y} = G(E_F(x,13),E_W(\\tilde{x}) + v)\\). \\(\\mathcal{L}_{\\mathrm{tmp}}\\) is the flicker suppression loss [37] to improve temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.577 + ], + "angle": 0, + "content": "Video toonification. For video face toonification, we have paired original face and toonified face \\((x,y)\\). They can be generated as \\( x = G_0(w^+) \\) and \\( y = G_0'(w^+) \\) from random latent code \\( w^+ \\) following Toonify [22], where \\( G_0' \\) is the StyleGAN fine-tuned on cartoon images. Let \\( G' \\) denote StyleGANEX loaded with \\( G_0' \\), then we train \\( E \\) using the objectives of Eq. (6) with \\( \\hat{y} = G'(E(x,\\tilde{x},13)) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.729 + ], + "angle": 0, + "content": "Note that compared to face editing based on StyleGANEX inversion in Sec. 4.2, the solution in this section does not require time-consuming latent optimization, and temporal consistency is enforced by the flicker suppression loss. Therefore, this solution is more suitable for efficient and coherent video face manipulation, which is a unique feature of the proposed framework. On the other hand, the solution based on StyleGANEX inversion is more flexible. There is no need to train new \\(E\\) for every editing vector \\(\\mathcal{V}\\) or fine-funed StyleGAN \\(G_0^{\\prime}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.706, + 0.757 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation details. We follow pSp [23] to set \\(\\lambda_{2} = 1\\) and \\(\\lambda_{3} = 0.8\\) for all tasks, \\(\\lambda_{4} = 0.1\\) for inversion task and 0 for other tasks. We set \\(\\lambda_{1}\\) to 0.0001, 0.005 and 0 for inversion, sketch/mask-to-face and other tasks, respectively. We set \\(\\lambda_{5} = 0.1\\) and \\(\\lambda_{6} = 30\\) empirically. The translation network \\(T\\) consists of two downsampling convolutional layers, two ResBlocks [11] and two upsampling convolutional layers, with small channel number 16. All experiments are performed using a single NVIDIA Tesla V100 GPU." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21004" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.895, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.334, + 0.715, + 0.35 + ], + "angle": 0, + "content": "Figure 7: StyleGANEX inversion and facial attribute/style editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.473, + 0.499 + ], + "angle": 0, + "content": "Datasets. We process FFHQ [17] to obtain 70,000 aligned training images of \\(1280 \\times 1280\\) resolution for all tasks except two video-related tasks that use StyleGAN generated data. We use BiSeNet [39] to extract parsing masks and follow pSp [23] to extract sketches from face images. We augment all training data with random geometric transformations [15] like scaling, translation and rotation to make faces unaligned. We use images and videos from FaceForensics++ [25], Unsplash and Pexels as our testing dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.26, + 0.53 + ], + "angle": 0, + "content": "5.1. Face Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Face editing. Figure 7 provides an overview of the performance of face inversion and attribute editing on StyleGANEX. We apply inversion to normal FoV face photos/paintings and use various editing vectors from InterFaceGAN [26] and LowRankGAN [40], and the pre-trained StyleGAN-NADA Ukiyo-e model [9], to edit the facial attributes or styles. As shown, these StyleGAN editing techniques work well on StyleGANEX. We also compare with pSp [23], HyperStyle [5] and BDInvert [14] in Fig. 8. Since these baselines are designed for cropped faces, we paste and blend their edited results back into the original image. For a fair comparison, we apply the same optimization method used in our approach to pSp for precise inversion (BDInvert is an optimization-based method and HyperStyle already uses extra hyper networks to simulate optimization). For editing that alters structures or colors, even precise inversion and blending cannot eliminate the obvious discontinuity along the seams, as indicated by the red arrows. In contrast, our approach processes the entire image as a whole and avoids such issues. Remarkably, our method successfully turns the whole hair into black in Fig.8(b), transfers the exemplar blond hairstyle onto the target face in Fig.8(c), and renders the full background with the StyleGAN-NADA Edvard Munch style in Fig.8(d)." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.581, + 0.828, + 0.597 + ], + "angle": 0, + "content": "Figure 8: Comparison on face editing." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.602, + 0.889, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.765, + 0.846, + 0.781 + ], + "angle": 0, + "content": "Figure 9: Comparison on super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Face super resolution. We show our \\(32 \\times\\) super-resolution results in Fig. 9(d), where both the face and non-face regions are reasonably restored. We further follow pSp to train a single mode on multiple rescaling factors (\\(4 \\sim 64\\)) with \\(\\ell = 3\\) to make a fair comparison. In pSp's results, the non-face region is super-resolved by Real-ESRGAN [35]. As in Fig. 9(b)(c), our method surpasses pSp in precise de" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "21005" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.156, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.244, + 0.143, + 0.256 + ], + "angle": 0, + "content": "(a) input" + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.089, + 0.236, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.244, + 0.234, + 0.256 + ], + "angle": 0, + "content": "(b) pix2pixHD" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.089, + 0.312, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.245, + 0.293, + 0.256 + ], + "angle": 0, + "content": "(c)TSIT" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.089, + 0.388, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.245, + 0.368, + 0.256 + ], + "angle": 0, + "content": "(d) pSp" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.09, + 0.465, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.245, + 0.446, + 0.256 + ], + "angle": 0, + "content": "(e) ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.263, + 0.457, + 0.278 + ], + "angle": 0, + "content": "Figure 10: Comparison on sketch-to-face translation." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.283, + 0.157, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.45, + 0.142, + 0.462 + ], + "angle": 0, + "content": "(a) input" + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.283, + 0.235, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.45, + 0.234, + 0.462 + ], + "angle": 0, + "content": "(b) pix2pixHD" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.283, + 0.311, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.45, + 0.295, + 0.461 + ], + "angle": 0, + "content": "(c)TSIT" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.283, + 0.388, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.45, + 0.365, + 0.462 + ], + "angle": 0, + "content": "(d) pSp" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.283, + 0.465, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.45, + 0.447, + 0.461 + ], + "angle": 0, + "content": "(e) ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.467, + 0.454, + 0.482 + ], + "angle": 0, + "content": "Figure 11: Comparison on mask-to-face translation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.494, + 0.468, + 0.524 + ], + "angle": 0, + "content": "tail restoration (e.g., glasses) and uniform super-resolution without discontinuity between face and non-face regions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.524, + 0.468, + 0.763 + ], + "angle": 0, + "content": "Sketch/mask to face translation. We compare our method with image-to-image translation models pix2pixHD [34] and TSIT [12], and StyleGAN-based pSp in Figs. 10-11. Pix2pixHD's results have many artifacts and monotonous colors. TSIT requires the inputs' side lengths to be divisible by 128. We find padding the input leads to failed translation. Therefore, we show its results on centrally cropped inputs, which are blurry. PSp generates realistic results, which are however less similar to the input sketch/mask. By comparison, our method can translate whole images and achieve realism and structural consistency to the inputs. For quantitative evaluation, we conduct a user study, where 30 subjects are invited to select what they consider to be the best results from the four methods. Each task uses eight results for evaluation. Table 1 summarizes the preference scores, where our method receives the best score." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Video face editing. We compare with pSp, HyperStyle, StyleHEAT [38] and STIT [31]. StyleHEAT uses features of a fixed \\(64 \\times 64\\) resolution for unaligned but still cropped \\(1024 \\times 1024\\) face reenactment. Specifically, it generates videos based on warping the features of the first frame, which however limits its inversion accuracy. STIT extends PTI [24] for full video processing by stitching. STIT cannot well preserve the complex hair details (Fig. 12, yellow box). As with image face editing, all four baselines are limited" + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.094, + 0.642, + 0.105 + ], + "angle": 0, + "content": "black hair" + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.095, + 0.786, + 0.106 + ], + "angle": 0, + "content": "young" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.106, + 0.679, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.106, + 0.862, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.422, + 0.748, + 0.434 + ], + "angle": 0, + "content": "editing control" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.433, + 0.887, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.476, + 0.851, + 0.492 + ], + "angle": 0, + "content": "Figure 12: Comparison on video face editing." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.5, + 0.888, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.706, + 0.871, + 0.722 + ], + "angle": 0, + "content": "Figure 13: Comparison on video face toonification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.869 + ], + "angle": 0, + "content": "to editing cropped regions, leading to discontinuity along the stitching seams. By comparison, our method uses the first-layer feature and skipped mid-layer features to provide spatial information, which achieves more coherent results. Moreover, we can randomly scale the editing vector \\( v \\) (by multiplying a scale factor) instead of using a fixed \\( v \\) during training. Then during testing, our method can flexibly adjust the editing degree by scaling \\( v \\) for users to select as in Fig. 12(e)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video face toonification. Compared with VToonify-T [37], our method preserves more details of the non-face region" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21006" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.09, + 0.462, + 0.104 + ], + "angle": 0, + "content": "Table 1: User preference scores. Best scores are in bold." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.11, + 0.468, + 0.2 + ], + "angle": 0, + "content": "
Taskpix2pixHD [34]TSIT [12]pSp [23]ours
sketch-to-face0.4000.1210.0290.450
mask-to-face0.1080.0750.1210.696
TaskVToonify-T [37]ours
video toonify0.0830.917
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.204, + 0.158, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.265, + 0.14, + 0.275 + ], + "angle": 0, + "content": "(a) input" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.204, + 0.233, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.265, + 0.226, + 0.276 + ], + "angle": 0, + "content": "(b) two steps" + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.204, + 0.312, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.265, + 0.313, + 0.276 + ], + "angle": 0, + "content": "(c) transferred (b)" + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.204, + 0.39, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.265, + 0.39, + 0.276 + ], + "angle": 0, + "content": "(d) only opt (e)" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.204, + 0.467, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.265, + 0.468, + 0.276 + ], + "angle": 0, + "content": "e) transferred (d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.281, + 0.46, + 0.296 + ], + "angle": 0, + "content": "Figure 14: Effect of encoder in StyleGANEX inversion." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.302, + 0.177, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.377, + 0.155, + 0.39 + ], + "angle": 0, + "content": "(a) input \\(x_{\\uparrow}\\)" + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.302, + 0.273, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.377, + 0.271, + 0.39 + ], + "angle": 0, + "content": "(b) \\(G(E(x_{1},\\tilde{x}_{1},7))\\)" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.302, + 0.371, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.377, + 0.367, + 0.39 + ], + "angle": 0, + "content": "(c) \\(G(E(x_{\\uparrow},x_{\\uparrow},7))\\)" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.302, + 0.468, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.377, + 0.466, + 0.39 + ], + "angle": 0, + "content": "(d) \\(G(E(x_{\\uparrow},\\ddot{x}_{\\uparrow},7))\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.397, + 0.468, + 0.412 + ], + "angle": 0, + "content": "Figure 15: Input choice to provide valid style information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.469, + 0.545 + ], + "angle": 0, + "content": "and generates shaper faces in Fig. 13. The reason is that VToonify-T uses a fixed latent code extractor while our method trains a joint latent code and feature extractor, thus our method is more powerful for reconstructing the details. Moreover, our method retains StyleGAN's shallow layers, which helps provide key facial features to make the stylized face more vivid. Table 1 shows a quantitative comparison on ten results, and our method obtains the best user score." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.23, + 0.575 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.734 + ], + "angle": 0, + "content": "The effect of Step II of our two-step inversion is verified in Fig. 6. We further study the effect of Step I in Fig. 14. With Step I providing a good prediction of \\( w^{+} \\) and \\( f \\), Step II only needs 500-iteration optimization for precise reconstruction (Fig. 14(b)) and valid domain transfer to Disney Princess (Fig. 14(c)). However, if we directly optimize a mean \\( w^{+} \\) and a random \\( f \\), the result is poor even with 2,000 iterations (indicated by a red arrow in Fig. 14(d)) and the optimized \\( w^{+} \\) and \\( f \\) overfit the input, which is not suitable for editing like domain transfer (Fig. 14(e))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In Fig. 15, we study the input choice to extract \\( w^{+} \\). The cropped aligned faces are the default choice. If we instead use the whole image to extract \\( w^{+} \\), the background will affect the facial style, leading to poor restoration in Fig. 15(c). We further find reasonable results (Fig. 15(d)) can still be obtained by cropping the input to decrease the background proportion. Note that the face is not aligned in the cropped image \\( (\\ddot{x}_{\\uparrow}) \\), which is useful for applications like super-resolution where extremely low-resolution faces are hard to detect and align. Users can simply manually crop the face region to provide valid style information." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.599, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.158, + 0.574, + 0.171 + ], + "angle": 0, + "content": "(a) input" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.091, + 0.696, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.158, + 0.667, + 0.17 + ], + "angle": 0, + "content": "(b) \\(\\ell = 0\\)" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.091, + 0.791, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.158, + 0.762, + 0.17 + ], + "angle": 0, + "content": "(c) \\(\\ell = 3\\)" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.091, + 0.889, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.157, + 0.863, + 0.17 + ], + "angle": 0, + "content": "(d) \\(\\ell = 7\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.177, + 0.824, + 0.192 + ], + "angle": 0, + "content": "Figure 16: Effect of skip connections." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.201, + 0.568, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.253, + 0.556, + 0.264 + ], + "angle": 0, + "content": "(a) input" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.201, + 0.633, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.253, + 0.622, + 0.263 + ], + "angle": 0, + "content": "(b) \\(\\ell = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.201, + 0.69, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.253, + 0.684, + 0.263 + ], + "angle": 0, + "content": "(c) \\(\\ell = 0\\)" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.201, + 0.761, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.253, + 0.748, + 0.263 + ], + "angle": 0, + "content": "(d) input" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.201, + 0.826, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.253, + 0.812, + 0.263 + ], + "angle": 0, + "content": "(e) \\(\\ell = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.201, + 0.892, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.843, + 0.253, + 0.879, + 0.263 + ], + "angle": 0, + "content": "(f) \\(\\ell = 0\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.271, + 0.863, + 0.286 + ], + "angle": 0, + "content": "Figure 17: Performance on low-quality sketches." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.292, + 0.598, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.405, + 0.569, + 0.418 + ], + "angle": 0, + "content": "input" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.293, + 0.696, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.404, + 0.681, + 0.417 + ], + "angle": 0, + "content": "inversion" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.293, + 0.792, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.404, + 0.788, + 0.418 + ], + "angle": 0, + "content": "blocky shape" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.293, + 0.89, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.404, + 0.873, + 0.418 + ], + "angle": 0, + "content": "add grass" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.421, + 0.696, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.533, + 0.673, + 0.545 + ], + "angle": 0, + "content": "redness" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.421, + 0.793, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.533, + 0.834, + 0.547 + ], + "angle": 0, + "content": "style mixing" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.421, + 0.89, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.554, + 0.804, + 0.569 + ], + "angle": 0, + "content": "Figure 18: Performance on car." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.892, + 0.758 + ], + "angle": 0, + "content": "In Fig. 16, we study the effect of skip connection. Without it \\((\\ell = 0)\\), the glasses cannot be generated. Skip connection provides mid-layer features to preserve the details of the input. However, too many extra features will alter the feature distribution of StyleGAN, e.g., always generating sunglasses as the input has no segmentation of eyes. Thus, we use \\(\\ell = 3\\) to balance between input-output consistency and diversity. Inversely, we can use a small \\(\\ell\\) to enhance the model robustness to low-quality inputs. For example, we can generate more realistic faces with \\(\\ell = 0\\) on DeepFace-Drawing low-quality sketches [7] as in Fig. 17." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.767, + 0.785 + ], + "angle": 0, + "content": "5.3. Results on Non-Facial Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The proposed refactoring is domain-agnostic, thus it can be applied to StyleGAN pre-trained on other domains like cars. An example with a wide vertical field of view is shown in Fig. 18, where we perform optimization-based StyleGAN inversion, image editing with the editing vectors found by GANSpace [10] and style mixing on the refactored model. The resulting manipulated cars look plausible." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21007" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.221, + 0.338, + 0.234 + ], + "angle": 0, + "content": "(c) severely rotated face" + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.238, + 0.353, + 0.252 + ], + "angle": 0, + "content": "Figure 19: Limitations." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.266, + 0.204, + 0.28 + ], + "angle": 0, + "content": "5.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.289, + 0.474, + 0.637 + ], + "angle": 0, + "content": "First, our framework currently relies on an inefficient optimization process for precise reconstruction. While future work can explore more efficient inversion methods (e.g., iterative residue prediction and hyper networks), it is important to note that this paper focuses on overcoming the fixed-crop limitation of StyleGAN, rather than the specific topic of GAN inversion. Second, StyleGANEX is limited by the feature representation of StyleGAN. While it shows great potential in handling normal FoV face images, out-of-distribution features such as complex clothing and human bodies may not be well handled as in Fig. 19(a)(b). As in Fig. 19(c), while our method can handle faces rotated 20 degrees, it still struggles with handling large rotation angles. However, this can be easily resolved by rough alignment of the input image in the middle as our method does not require accurate alignment. Finally, after translation, we can undo the alignment to obtain a plausible result as in right of Fig. 19(c). Third, StyleGANEX, like StyleGAN, focuses on face manipulation and may not support out-of-distribution semantical editing of non-facial regions. Last, StyleGANEX may inherit the model bias of StyleGAN. Applying it to tasks with severe data imbalance might lead to unsatisfactory results on under-represented data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.196, + 0.665 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.84 + ], + "angle": 0, + "content": "We have presented an effective approach to refactor StyleGAN to overcome its fixed-crop limitation while retaining its style control abilities. The refactored model, called StyleGANEX, fully inherits the parameters of the pre-trained StyleGAN without retraining, and is thus fully compatible with the generative space of StyleGAN. We further introduced a StyleGANEX encoder to project normal FoV face images to the joint \\( W^{+} - F \\) space of StyleGANEX for real face inversion and manipulation. Our approach offers an effective solution to the general issue of StyleGAN and extends its capability beyond fixed-resolution data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Acknowledgments. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is al" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "so supported by MOE AcRF Tier 1 (2021-T1-001-088) and NTU NAP grant." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.14, + 0.597, + 0.155 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.166, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proc. Int'l Conf. Computer Vision, pages 4432-4441, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.224, + 0.894, + 0.293 + ], + "angle": 0, + "content": "[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.296, + 0.895, + 0.35 + ], + "angle": 0, + "content": "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinements. In Proc. Int'l Conf. Computer Vision, pages 6711–6720, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.354, + 0.895, + 0.407 + ], + "angle": 0, + "content": "[4] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.412, + 0.895, + 0.48 + ], + "angle": 0, + "content": "[5] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.484, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[6] Kelvin CK Chan, Xintao Wang, Xiangyu Xu, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for large-factor image super-resolution. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.541, + 0.892, + 0.596 + ], + "angle": 0, + "content": "[7] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. ACM Transactions on Graphics, 39(4):72-1, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.6, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[8] Tan M Dinh, Anh Tuan Tran, Rang Nguyen, and Binh-Son Hua. Hyperinverter: Improving stylegan inversion via hypernetwork. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11389-11398, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.658, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[9] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clipguided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.716, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[10] Erik Härkönen, Aaron Hertzman, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Advances in Neural Information Processing Systems, 2020. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.774, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[12] Liming Jiang, Changxu Zhang, Mingyang Huang, Chunxiao Liu, Jianping Shi, and Chen Change Loy. Tset: A simple and versatile framework for image-to-image translation. In Proc. European Conf. Computer Vision, pages 206-222. Springer, 2020. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.166, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21008" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.135 + ], + "angle": 0, + "content": "[13] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In Proc. Int'l Conf. Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.137, + 0.469, + 0.192 + ], + "angle": 0, + "content": "[14] Kyoungkook Kang, Seongtae Kim, and Sunghyun Cho. Gan inversion for out-of-range images with geometric transformations. In Proc. Int'l Conf. Computer Vision, pages 13941-13949, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.25 + ], + "angle": 0, + "content": "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.252, + 0.469, + 0.308 + ], + "angle": 0, + "content": "[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.31, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.44, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[19] Feng-Lin Liu, Shu-Yu Chen, Yukun Lai, YueRen Jiang, Hongbo Fu, and Lin Gao. Deepfacevideoediting: Sketch-based deep editing of face videos. ACM Transactions on Graphics, 41(4):167, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[20] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, JunYan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11399-11409, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.625 + ], + "angle": 0, + "content": "[21] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 2085–2094, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[22] Justin NM Pinkney and Doron Adler. Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[23] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[24] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[25] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. In Proc. Int'l Conf. Computer Vision, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[26] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[27] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[28] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[29] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics, 39(6):1-14, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[30] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[31] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia, pages 1–9, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[32] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. Stylegan2 distillation for feed-forward image manipulation. In Proc. European Conf. Computer Vision, pages 170-186. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[33] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[34] Ting Chun Wang, Ming Yu Liu, Jun Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[35] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proc. Int'l Conf. Computer Vision, pages 1905-1914, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[36] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. StyleAlign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[37] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Vtoonify: Controllable high-resolution portrait video style transfer. ACM Transactions on Graphics, 41(6):1-15, 2022. 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[38] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In Proc. European Conf. Computer Vision, pages 85-101. Springer, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[39] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proc. European Conf. Computer Vision, pages 334-349. Springer, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[40] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zheng-Jun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21009" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "in gans. In Advances in Neural Information Processing Systems, volume 34, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[41] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proc. European Conf. Computer Vision, pages 592-608. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.468, + 0.232 + ], + "angle": 0, + "content": "[42] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Transactions on Graphics, 40(6):1-13, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21010" + } + ] +] \ No newline at end of file diff --git a/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/71e23834-fb4f-4c96-b627-797882a6ea6f_origin.pdf b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/71e23834-fb4f-4c96-b627-797882a6ea6f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c9e5059c14b2527f024fbf6afe793619091ca97a --- /dev/null +++ b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/71e23834-fb4f-4c96-b627-797882a6ea6f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddcd41f1319e6a9b65f1aa2a74daffe9d90381beb806de68a680b24154e4edd0 +size 4158676 diff --git a/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/full.md b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/full.md new file mode 100644 index 0000000000000000000000000000000000000000..40d79dc9543f0be8a3c94fd77fbeff5fac1bbe60 --- /dev/null +++ b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/full.md @@ -0,0 +1,431 @@ +# StyleGANEX: StyleGAN-Based Manipulation Beyond Cropped Aligned Faces + +Shuai Yang Liming Jiang Ziwei Liu Chen Change Loy S-Lab, Nanyang Technological University + +{shuai.yang, liming002, ziwei.liu, ccloy}@ntu.edu.sg + +![](images/24138a5cf244f6b66f18abea521c020c4e2094f264edfe0ec5904e992b1dc660.jpg) + +![](images/7bbe1fc312a04d55ea14d1b6bc2371a42c134998921f9b81c60e1279a2842e3e.jpg) + +![](images/8a02fde41a725b936f700d5f0efbc0abfab6eb6b32fc265ab506ddaf142dec22.jpg) + +![](images/ec4c3f7113165c993a297da4c4cffb69a314dc846787ec58ef3ead13b3ac105c.jpg) + +![](images/4303ceac55161479649b24e326d7a7084fabb25da09bafd8eb88d9a8ac5641fa.jpg) +Figure 1: We expand StyleGAN to encompass a diverse set of tasks that go beyond the constraints of cropped aligned faces. + +# Abstract + +Recent advances in face manipulation using StyleGAN have produced impressive results. However, StyleGAN is inherently limited to cropped aligned faces at a fixed image resolution it is pre-trained on. In this paper, we propose a simple and effective solution to this limitation by using dilated convolutions to rescale the receptive fields of shallow layers in StyleGAN, without altering any model parameters. This allows fixed-size small features at shallow layers to be extended into larger ones that can accommodate variable resolutions, making them more robust in characterizing unaligned faces. To enable real face inversion and manipulation, we introduce a corresponding encoder that provides the first-layer feature of the extended StyleGAN in addition to the latent style code. We validate the effectiveness of our method using unaligned face inputs of various resolutions in a diverse set of face manipulation tasks, including facial attribute editing, super-resolution, sketch/mask-to-face translation, and face toonification. Project page https: + +//www.mmlab-ntu.com/project/styleganex + +# 1. Introduction + +StyleGAN [17, 18] has emerged as one of the most successful models for generating high-quality faces. Building upon StyleGAN, researchers have developed a range of face manipulation models [1, 23, 29, 26, 10, 28, 40, 36]. These models typically map real face images or other face-related inputs to the latent space of StyleGAN, perform semantic editing in the latent space, and then map the edited latent code back to the image space. This approach enables a variety of tasks, including facial attribute editing, face restoration, sketch-to-face translation, and face toonification. As the manipulated faces remain within the generative space of StyleGAN, the quality of the image output is guaranteed. + +Despite its ability to ensure high-quality image output, the generative space of StyleGAN is limited by a fixed-crop constraint that restricts image resolution and face layout. As a result, existing face manipulation models based on StyleGAN can only handle cropped and aligned face images. In such images with a limited field of view (FoV), the face typically dominates the image, leaving little room for back- + +ground and clothing, and often resulting in partially cropped hair. However, in everyday portrait photos such as selfies, faces occupy a smaller proportion of the image, allowing for a complete hairstyle and upper body seen. Portrait videos such as those from live streaming require an even larger background area to accommodate face movement. To process these types of inputs, which we refer to as normal FoV face images and videos, existing manipulation models need to align, crop, and edit the face before pasting the result back onto the original image [4, 19, 31]. This approach often results in discontinuity near the seams, e.g., only editing the hair color inside the cropped area. + +While StyleGAN3 [16] was introduced to address unaligned faces, a recent study [4] found that even StyleGAN3 requires face realignment before effectively projecting to its latent space. Moreover, StyleGAN3 is still constrained by a fixed image resolution. Motivated by the translation equivariance of convolutions, VToonify [37] addresses the fixed-crop limitation of StyleGAN by removing its shallow layers to accept input features of any resolution. However, these shallow layers are crucial for capturing high-level features of the face, such as pose, hairstyle, and face shape. By removing these layers, the network loses its ability to perform latent editing on these important features, which is a distinctive capability of StyleGAN. Therefore, the challenge remains in overcoming the fixed-crop limitation of StyleGAN while preserving its original style manipulation abilities, which is a valuable research problem to solve. + +In this paper, we propose a simple yet effective approach for refactoring StyleGAN to overcome the fixed-crop limitation. In particular, we refactor its shallow layers instead of removing them, allowing the first layer to accept input features of any resolution. This simple change expands StyleGAN's style latent space into a more powerful joint style latent and first-layer feature space $(W^{+} - F$ space), extending the generative space beyond cropped aligned faces. Furthermore, our refactoring only changes the receptive field of shallow-layer convolutions, leaving all pre-trained model parameters intact. Hence, the refactored StyleGAN (StyleGANEX) can directly load the original StyleGAN parameters, fully compatible with the generative space of StyleGAN, and retains its style representation and editing ability. This means that the StyleGAN editing vectors found in previous studies [26, 10, 28] can be directly applied to StyleGANEX for normal FoV face editing, e.g., changing the face pose, as shown in Fig. 1(a). + +Based on StyleGANEX, we further design a corresponding encoder that projects normal FoV face images to the $W^{+} - F$ space for real face inversion and manipulation. Our encoder builds upon pSp encoder [23] and aggregates its multi-layer features to predict the first-layer feature of StyleGANEX. The encoder and StyleGANEX form a fully convolutional encoder-decoder framework. With the first- + +layer feature as the bottleneck layer, whose resolution is $1/32$ of the output image, our framework can handle images and videos of various resolutions, as long as their side lengths are divisible by 32. Depending on the input and output types, our framework can perform a wide range of face manipulation tasks. In this paper, we select several representative tasks, as shown in Fig. 1, including facial attribute editing, face super-resolution, sketch/mask-to-face translation and video face toonification. While the focus of these tasks in the past is limited to cropped aligned faces, our framework can handle normal FoV faces, showing significant advantages over previous StyleGAN-based approaches. To summarize, our main contributions are: + +- A novel StyleGANEX architecture with extended $W^{+} - F$ space, which overcomes the fixed-crop limitation of StyleGAN. +- An effective encoder that is able to project normal FoV face images into the $W^{+} - F$ domain. +- A generic and versatile fully convolutional framework for face manipulation beyond cropped aligned faces. + +# 2. Related Work + +StyleGAN inversion. StyleGAN inversion aims at projecting real face images into the latent space of StyleGAN for further manipulation. Image2StyleGAN [1] analyzes the latent space and proposes $W^{+}$ space to reconstruct real faces with latent code optimization. PIE [29] and IDinvert [41] further consider the editability of the latent code during optimization. To speed up inversion, pSp [23] and e4e [30] train an encoder to directly project the target face to its corresponding latent code, which is however hard to reconstruct fine details and handle occlusions. To solve this issue, Restyle [3] and HFGI [33] predict the residue of latent codes or mid-layer features to reduce errors, respectively. Instead of focusing on the latent code, PTI [24] optimizes StyleGAN itself to fit a target face, which is accelerated by HyperInverter [8] and HyperStyle [5] to predict offsets of network parameters with hyper networks. The above methods are limited to cropped aligned faces for valid face editing. With the extended $W^{+} - F$ space, our framework is able to perform inversion on normal FoV face images. + +StyleGAN-based face manipulation. An intuitive way of StyleGAN-based face manipulation is to optimize the latent code online to achieve certain objectives such as a pixel-level constraints [1], sketch-based structure constraints [19] or text-guided semantic constraints [21, 9]. Another way is to search for offline editing vectors to add to the latent code for manipulation. Supervised methods identify meaningful editing vectors based on attribute labels or pretrained classifiers [26, 27, 13, 2]. On the other hand, unsupervised methods statistically analyze the StyleGAN latent space to discover semantically significant editing direc + +tions by principal component analysis [10], low-rank factorization [40] and closed-form factorization [28]. Meanwhile, face manipulation can be directly realized by image-to-image translation frameworks, where StyleGAN is used to generate paired training data [32, 37] or to build the decoder [23, 6, 37]. To achieve spatial editing, methods [42, 20] are proposed to manipulate mid-layer features in addition to the latent code, similar to our $W^{+} - F$ space. Moreover, BDInvert [14] and StyleHEAT [38] introduces feature transformations for unaligned face editing. However, the above methods, as well as HFGI [33], follow StyleGAN features' original fixed resolution, thus still suffering the crop limitation. Differently, our method predicts the first-layer feature that can have various resolutions in StyleGANEX. Moreover, as we will show later in Sec. 5.1, compared to VToonify [37], StyleGANEX retains the complete style manipulation abilities of the shallow layers of StyleGAN, with a jointly trained latent code and feature extractor, preserving vivid details and supporting more diverse manipulation tasks. + +# 3. StyleGANEX + +# 3.1. Analysis of the Fixed-Crop Limitation + +StyleGAN has great potential for handling normal FoV face images. The generator of StyleGAN is a fully convolutional architecture that can naturally handle different feature resolutions, as convolution operations and the style modulation of StyleGAN are independent of the input resolution. Additionally, the translation equivariance of convolution operations naturally supports feature translation. As analyzed in VToonify [37], if we translate or rotate the feature of the 7th layer of StyleGAN, the resulting face will also be shifted or rotated, as shown in Figs. 2(a)(b)(d). + +The limiting factor originates from StyleGAN's constant first-layer feature. First, the first-layer feature has a fixed resolution of $4 \times 4$ , limiting the output to $1024 \times 1024$ resolution. Second, $4 \times 4$ resolution is inadequate to characterize the spatial information of unaligned faces. We have taken a step further than the analysis in VToonify [37] to investigate translation and rotation on the first-layer feature. As shown in Fig. 2(c), sub-pixel translation fuses adjacent feature values severely, resulting in a blurry face due to the small number of elements (only 16) in the first-layer feature. In Fig. 2(e), the first-layer feature fails to provide enough spatial information for a valid rotation. In comparison, the 7-th layer has a higher resolution ( $32 \times 32$ ), making it better suited for capturing spatial information. However, only a single layer alone provides limited style control, as the full facial structural styles are hierarchically modeled by seven shallow layers. Simply ignoring low-resolution layers, as in VToonify, disables flexible latent editing over the styles of these layers, such as pose, age, and face shape. + +![](images/4dd2f7b042a90d2a10d4f3c078d29d1a52f4fef4d160a1993f7dff28e2959671.jpg) +(a) random face + +![](images/ee0a5d11c10510ab1c005852e1af646a2c024584a33bab0164c9d9ec7b54fdda.jpg) +(b) shift layer#7 + +![](images/061f085f73cb3728542f5d6c99e17c0023609d63814507960f215398d6169b7f.jpg) +(c) shift layer#1 + +![](images/6476d53f8cdbeb96d51bd3412c8ce99ed95c5f3a1bc629f71ce4c40a66060e20.jpg) +(d) rotate layer#7 + +![](images/d4bc1f3feb8f3bb0efae3c3d5cfdbceb0452a94d7d4bb04c743b9db9e226bfd2.jpg) +(e) rotate layer#1 + +![](images/e9e946eeb589cb141531e090236900e04ad9959a2c1d23adee6eb206fc75a298.jpg) +Figure 2: Analysis of StyleGAN in generating unaligned faces. (a) Generated face. (b)(c) Face generated by translating the feature maps of the 7-th and 1-st layers of StyleGAN respectively to shift the face by 150 pixels. (d)(e) Face generated by rotating the feature maps of the 7-th and 1-st layers of StyleGAN by 10 degrees, respectively. +(a) real face photo +Figure 3: Analysis of StyleGANEX in generating unaligned faces. (a) A real face photo. (b) Reconstructed face by projecting (a) into the $W^{+} - F$ space of StyleGANEX. (c) Face generated by translating the first-layer feature map to shift the face by 300 pixels. (d) Face generated by rotating the first-layer feature map by 10 degrees. + +![](images/8b0baa2f0e2280d354352224c86f6fa41813a265bcbae6172a37037c66cdb512.jpg) +(b) projected face + +![](images/32e1b5de8500028b7611a25ef1cbcaebd62ff1864d7ef0afa23ded9ee02094cc.jpg) +(c) shift layer#1 + +![](images/1cf618f374e035e114bf8f014572bef5422d516786723eccc01b70f578c1e7f0.jpg) +(d) rotate layer#1 + +Expanding the shallow layers of StyleGAN to have the same $32 \times 32$ resolution as the 7th layer, or more generally, $H / 32 \times W / 32$ resolution for an $H \times W$ image, would provide enough structure and layout information to combine the style controllability of shallow layers with support for normal FoV faces. This is precisely the key idea of StyleGANEX, and we will introduce our simple solution in Section 3.2. As a preview of the performance of the expanded layers in enabling face manipulation beyond cropped and aligned faces, Fig. 3 shows that for a $1472 \times 1600$ normal FoV face photo, we can obtain its latent code and an additional $46 \times 50$ first-layer feature as the input to StyleGANEX (inversion method explained in Section 4.2). Face translation and rotation can be realized by shifting or rotating the first-layer feature. Additionally, the face can be effectively edited by applying style mixing [17] or InterFaceGAN editing vectors [26] to the latent code, as shown in Fig. 1(a). + +# 3.2. From StyleGAN to StyleGANEX + +Figure 4 illustrates the generator architectures of StyleGAN and StyleGANEX. Compared to StyleGAN, we first replace the constant $4 \times 4$ first-layer feature with a variable feature whose resolution is $1/32$ of the output image. Then, we remove the upsample operations before the 8-th layer, allowing features in the seven shallow layers to share the same resolution as the 7-th layer. However, the convolution kernels or reception fields of these layers do not match their input features with the enlarged resolution. To solve + +![](images/5bd3a5e9b589ed01fc97536ed7581464501c396357edf0264444fc3b98203b13.jpg) + +![](images/a283461e5c0acbaa9c835876077a2b31484bfab1e735c0556542563932f5a3cb.jpg) +Figure 4: Refactor StyleGAN to StyleGANEX. For simplicity, learned weights, biases and noises are omitted. + +this problem, we enlarge the reception fields by modifying the convolutions to their dilated versions. For example, the first layer only needs to change the dilation factor from 1 to 8. With this simple modification, StyleGAN is refactored to StyleGANEX. Since the first layer becomes variable, the original $W^{+}$ latent space is extended to a joint $W^{+} - F$ space, where the latent code $w^{+} \in W^{+}$ provides style cues, and the first-layer feature $f \in F$ mainly encodes spatial information. + +The refactoring of StyleGAN to StyleGANEX has three key advantages. 1) Support for unaligned faces. The resolution enlargement and variable first-layer features of StyleGANEX overcome the fixed-crop limitation. 2) Compatibility. No model parameters are altered during refactoring, meaning that StyleGANEX can directly load pretrained StyleGAN parameters without retraining. In fact, if we upsample the StyleGAN's constant input feature by $8 \times$ with nearest neighbor interpolation to serve as $f$ of StyleGANEX, StyleGANEX degrades exactly to StyleGAN with the same $1024 \times 1024$ generative space. The computational cost of the refactoring is also minimal, with generating an image taking 0.026s and 0.028s for StyleGAN and StyleGANEX, respectively. 3) Flexible manipulation. StyleGANEX retains the style representation and editing ability of StyleGAN, meaning that abundant StyleGAN-based face manipulation techniques can be applied to StyleGANEX. + +# 4. Face Manipulation with StyleGANEX + +# 4.1. StyleGANEX Encoder + +This section introduces our StyleGANEX encoder $E$ which is used to project real face images into the $W^{+} - F$ + +![](images/143eae5cf6a3f2e6efba528a654ba7cb1192618df0df4e3509be087d8f5745b2.jpg) +Figure 5: Details of StyleGANEX Encoder. + +space of StyleGANEX $G$ . The encoder builds upon the pSp encoder [23], as depicted in Fig. 5. Specifically, for $F$ space, we concatenate pSp features in the middle layers and add a convolution layer to map the concatenated features to the first-layer input feature $f$ of $G$ . For $W^{+}$ space, the original pSp encoder takes a $256 \times 256$ image as input and convolves it to eighteen $1 \times 1 \times 512$ features to map to a latent code $w^{+} \in \mathbb{R}^{18 \times 512}$ . To make $E$ accept more general $H \times W$ images, we add global average pooling to resize all features to $1 \times 1 \times 512$ before mapping to latent codes. To support various face manipulation tasks flexibly, we can extract $f$ and $w^{+}$ from different sources. Let $E_{F}$ and $E_{W}$ be the operation of $E$ to extract the first-layer feature and the latent code, respectively. We have + +$$ +f, w ^ {+} = E _ {F} \left(x _ {1}\right), E _ {W} \left(x _ {2}\right) := E \left(x _ {1}, x _ {2}\right), \tag {1} +$$ + +where $x_{1}$ and $x_{2}$ are the source inputs for face layout and face style, respectively. Then, a general form of image generation by $G$ from $x_{1}$ and $x_{2}$ is $G(E_{F}(x_{1}), E_{W}(x_{2}))$ . In some face manipulation tasks like super-resolution [6] and toonification [37], passing encoder features to the generator via skip connections helps preserve the details of the input image. We thus introduce a scalar parameter $\ell$ to the generation process, indicating the $\ell$ shallow layers of $G$ receive the encoder features $(\ell = 0$ means no skip connections): + +$$ +\hat {x} = G \left(E _ {F} \left(x _ {1}, \ell\right), E _ {W} \left(x _ {2}\right)\right) := G \left(E \left(x _ {1}, x _ {2}, \ell\right)\right), \tag {2} +$$ + +where $E_{F}(x_{1},\ell)$ provides both $f$ and the skipped encoder features. For $x_{1}$ of $H\times W$ resolution, $f$ and the generated image $\hat{x}$ will be of $H / 8\times W / 8$ and $4H\times 4W$ , respectively. The resolution of $x_{2}$ can be independent of $x_{1}$ and $\hat{x}$ + +# 4.2. StyleGANEX Inversion and Editing + +To find appropriate $\hat{f}$ and $\hat{w}^+$ that precisely reconstruct a target image $x$ , we perform a two-step StyleGANEX inversion. Step I projects $x$ to initial $f$ and $w^+$ with $E$ . Step + +![](images/d5c2194e9f68e66133808db72ab55b3dbf92a60fba7124c8e47f48d7b79cbf21.jpg) +(a) face photo + +![](images/4f16ec3423c7f14004d8ad934ce3800017349edcbdd61d2339d7d7b39cc2acd7.jpg) +(b) encoded face + +![](images/b514d10c07b6b37a66d09e4d1f77307de3b189e39a61e433c3fea75bb3006e3b.jpg) +(c) (b)+optimize +(d) domain transfer + +![](images/3a57a0d2749731b37067e1c9326cef7d50b8cadad51d6ca018fee25782a508fc.jpg) +Figure 6: StyleGANEX inversion. + +II optimizes $f$ and $w^{+}$ to further reduce the reconstruction error. The training of $E$ follows pSp with reconstruction losses and a regularization loss [23]: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} (\hat {x}, x) + \lambda_ {1} \mathcal {L} _ {\text {r e g}} \left(E _ {W} (\tilde {x})\right), \tag {3} +$$ + +where $\hat{x} = G(E(x, \tilde{x}, 0))$ and $\tilde{x}$ is the cropped aligned face region of $x$ . We empirically find that using $\tilde{x}$ instead of $x$ predicts more accurate $w^{+}$ since StyleGAN is originally trained on cropped aligned faces. $\mathcal{L}_{\mathrm{reg}}$ encourages the predicted $w^{+}$ closer to the average latent code to improve the image quality. $\mathcal{L}_{\mathrm{rec}}$ measures the distance between the reconstructed $\hat{x}$ and the target $x$ in terms of pixel similarity, perceptual similarity, and identity preservation: + +$$ +\mathcal {L} _ {\mathrm {r e c}} (\hat {x}, x) = \lambda_ {2} \mathcal {L} _ {2} (\hat {x}, x) + \lambda_ {3} \mathcal {L} _ {\mathrm {L P I P S}} (\hat {x}, x) + \lambda_ {4} \mathcal {L} _ {\mathrm {I D}} (\hat {x}, x). +$$ + +As shown in Fig. 6(b), $\hat{x}$ largely approximates $x$ . But the background details and clothings are still hard to reconstruct. Therefore, we further optimize $f$ and $w^{+}$ + +$$ +\hat {f}, \hat {w} ^ {+} = \underset {f, w ^ {+}} {\operatorname {a r g m i n}} \mathcal {L} _ {\mathrm {L P I P S}} (G (f, w ^ {+}), x), \tag {4} +$$ + +where $f$ and $w^{+}$ are initialized by $E(x,\tilde{x},0)$ . The optimized $G(\hat{f},\hat{w}^{+})$ in Fig. 6(c) well reconstructs $x$ . + +After inversion, we can perform flexible editing over $x$ as in StyleGAN. Figure 1(a) shows two examples: we can exchange the last 11 elements of $\hat{w}^{+}$ with random samples, to mix the color and texture styles; we can add InterFaceGAN editing vectors [26] to $\hat{w}^{+}$ to make a young face. Moreover, as shown in Fig. 6(d), if we load $G$ a pre-trained StyleGAN-NADA Disney Princess model [9] (let $G'$ denote the new $G$ ), we can obtain $G'(\hat{f}, \hat{w}^{+})$ , a cartoon version of $x$ . + +# 4.3. StyleGANEX-Based Translation + +The encoder and StyleGANEX form an end-to-end image-to-image translation framework in Fig. 5. Depending on the type of paired training data, it can be trained to efficiently realize different face manipulation tasks. As with pSp [23], we will fix StyleGANEX generator and only train the encoder on the given task. + +Face super-resolution. Given low-resolution and high-resolution training image pairs $(x,y)$ , we can train $E$ to recover $y$ from $x$ to learn face super-resolution with the loss + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} (\hat {y}, y) + \lambda_ {5} \mathcal {L} _ {\text {a d v}} (\hat {y}, y), \tag {5} +$$ + +where $\hat{y} = G(E(x_{\uparrow},\tilde{x}_{\uparrow},7))$ $\uparrow$ is the upsample operation to make $x$ match the input resolution of $E$ $\tilde{x}$ is the cropped aligned face region of $x$ . We add an adversarial loss $\mathcal{L}_{\mathrm{adv}}$ to improve the realism of the generated image. + +Sketch/mask-to-face translation. Given a real face $y$ as target and its sketch or parsing mask $x$ as source, we can train $E$ to translate $x$ to $y$ with Eq. (3) as objectives. In this task, we add a trainable light-weight translation network $T$ to map $x$ to an intermediate domain where $E$ can more easily extract features. For the style condition, $G$ 's first 7 layers use the latent code extracted from $x$ to provide structural styles, while its last 11 layers use the latent code from $\tilde{y}$ to provide color and texture styles to simplify reconstruction. Therefore, $\hat{y} = G(E_F(T(x),\ell),E_W^{1:7}(T(x))\oplus E_W^{8:18}(\tilde{y}))$ , where $\oplus$ is concatenation operation, and the superscript of $E_{W}$ means taking the $1\sim 7$ or $8\sim 18$ elements of $w^{+}$ . $\ell = 1$ for sketch inputs and $\ell = 3$ for mask inputs. + +Video face editing. Given paired original face, edited face, and its editing vector $(x,y,v)$ (we can simply generate $x = G_0(w^+)$ and $y = G_0(w^+ +v)$ from random latent code $w^{+}$ with StyleGAN $G_{0})$ , we train $E$ for face editing with + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} (\hat {y}, y) + \lambda_ {5} \mathcal {L} _ {\text {a d v}} (\hat {y}, y) + \lambda_ {6} \mathcal {L} _ {\text {t m p}} (\hat {y}), \tag {6} +$$ + +where $\hat{y} = G(E_F(x,13),E_W(\tilde{x}) + v)$ . $\mathcal{L}_{\mathrm{tmp}}$ is the flicker suppression loss [37] to improve temporal consistency. + +Video toonification. For video face toonification, we have paired original face and toonified face $(x,y)$ . They can be generated as $x = G_0(w^+)$ and $y = G_0'(w^+)$ from random latent code $w^+$ following Toonify [22], where $G_0'$ is the StyleGAN fine-tuned on cartoon images. Let $G'$ denote StyleGANEX loaded with $G_0'$ , then we train $E$ using the objectives of Eq. (6) with $\hat{y} = G'(E(x,\tilde{x},13))$ . + +Note that compared to face editing based on StyleGANEX inversion in Sec. 4.2, the solution in this section does not require time-consuming latent optimization, and temporal consistency is enforced by the flicker suppression loss. Therefore, this solution is more suitable for efficient and coherent video face manipulation, which is a unique feature of the proposed framework. On the other hand, the solution based on StyleGANEX inversion is more flexible. There is no need to train new $E$ for every editing vector $\mathcal{V}$ or fine-funed StyleGAN $G_0^{\prime}$ . + +# 5. Experimental Results + +Implementation details. We follow pSp [23] to set $\lambda_{2} = 1$ and $\lambda_{3} = 0.8$ for all tasks, $\lambda_{4} = 0.1$ for inversion task and 0 for other tasks. We set $\lambda_{1}$ to 0.0001, 0.005 and 0 for inversion, sketch/mask-to-face and other tasks, respectively. We set $\lambda_{5} = 0.1$ and $\lambda_{6} = 30$ empirically. The translation network $T$ consists of two downsampling convolutional layers, two ResBlocks [11] and two upsampling convolutional layers, with small channel number 16. All experiments are performed using a single NVIDIA Tesla V100 GPU. + +![](images/f83d168f36c42383e6e3745da39e5f7464610ef7b2e9158d33fd2dd6a8d3bd84.jpg) +Figure 7: StyleGANEX inversion and facial attribute/style editing. + +Datasets. We process FFHQ [17] to obtain 70,000 aligned training images of $1280 \times 1280$ resolution for all tasks except two video-related tasks that use StyleGAN generated data. We use BiSeNet [39] to extract parsing masks and follow pSp [23] to extract sketches from face images. We augment all training data with random geometric transformations [15] like scaling, translation and rotation to make faces unaligned. We use images and videos from FaceForensics++ [25], Unsplash and Pexels as our testing dataset. + +# 5.1. Face Manipulation + +Face editing. Figure 7 provides an overview of the performance of face inversion and attribute editing on StyleGANEX. We apply inversion to normal FoV face photos/paintings and use various editing vectors from InterFaceGAN [26] and LowRankGAN [40], and the pre-trained StyleGAN-NADA Ukiyo-e model [9], to edit the facial attributes or styles. As shown, these StyleGAN editing techniques work well on StyleGANEX. We also compare with pSp [23], HyperStyle [5] and BDInvert [14] in Fig. 8. Since these baselines are designed for cropped faces, we paste and blend their edited results back into the original image. For a fair comparison, we apply the same optimization method used in our approach to pSp for precise inversion (BDInvert is an optimization-based method and HyperStyle already uses extra hyper networks to simulate optimization). For editing that alters structures or colors, even precise inversion and blending cannot eliminate the obvious discontinuity along the seams, as indicated by the red arrows. In contrast, our approach processes the entire image as a whole and avoids such issues. Remarkably, our method successfully turns the whole hair into black in Fig.8(b), transfers the exemplar blond hairstyle onto the target face in Fig.8(c), and renders the full background with the StyleGAN-NADA Edvard Munch style in Fig.8(d). + +![](images/6669ec1c77690680d482c38522b323886d380ffc459d451c8af3f2f1ee13ec76.jpg) +Figure 8: Comparison on face editing. + +![](images/0f730d692533d28b4b70deb9726abaf36ca981c41e1b3d1bfb3a2393acad16bc.jpg) +Figure 9: Comparison on super-resolution. + +Face super resolution. We show our $32 \times$ super-resolution results in Fig. 9(d), where both the face and non-face regions are reasonably restored. We further follow pSp to train a single mode on multiple rescaling factors ( $4 \sim 64$ ) with $\ell = 3$ to make a fair comparison. In pSp's results, the non-face region is super-resolved by Real-ESRGAN [35]. As in Fig. 9(b)(c), our method surpasses pSp in precise de + +![](images/1616fe9e97e02c103eccd6dfe08e80a19631492f752484d6b753c32042a058ad.jpg) +(a) input + +![](images/e456360602ea4dcccad86764a72dbcfec5d1a006fcf0803d3ca526f3bb78dd78.jpg) +(b) pix2pixHD + +![](images/cb6c650148c34e52000471dce32ac06b8c14f093e85aec0170f7b884cf600fb0.jpg) +(c)TSIT + +![](images/0f9dd9dc017622ba2d31a410ae7df0dc138f3601f5b77218069c564d1a673caa.jpg) +(d) pSp + +![](images/829cfa69c6100131be56f5c7cd2d0bde87c3dec524c8c85dc83dc800fdafb168.jpg) +(e) ours + +![](images/8e895e6ba0d0312dd2829448cfeefb3879cbb9a316cc827ee5708dcea8cdc016.jpg) +Figure 10: Comparison on sketch-to-face translation. +(a) input +Figure 11: Comparison on mask-to-face translation. + +![](images/e78721fe8a8cc41b5f9d2744129cd9684a7a2bc145d7c33222cbb365d3335ce1.jpg) +(b) pix2pixHD + +![](images/c7bfc16a583e8c849ad292d1638e14283c18f49c3934f9915f2d6457c541594b.jpg) +(c)TSIT + +![](images/fa83c66ee39dcb666865ad262a9426322752ac73687eff840e099091123e34ef.jpg) +(d) pSp + +![](images/d311c8b28bd2ba176ac7c5091a503effecf0043cba77c4e2ce14122ccd79c900.jpg) +(e) ours + +tail restoration (e.g., glasses) and uniform super-resolution without discontinuity between face and non-face regions. + +Sketch/mask to face translation. We compare our method with image-to-image translation models pix2pixHD [34] and TSIT [12], and StyleGAN-based pSp in Figs. 10-11. Pix2pixHD's results have many artifacts and monotonous colors. TSIT requires the inputs' side lengths to be divisible by 128. We find padding the input leads to failed translation. Therefore, we show its results on centrally cropped inputs, which are blurry. PSp generates realistic results, which are however less similar to the input sketch/mask. By comparison, our method can translate whole images and achieve realism and structural consistency to the inputs. For quantitative evaluation, we conduct a user study, where 30 subjects are invited to select what they consider to be the best results from the four methods. Each task uses eight results for evaluation. Table 1 summarizes the preference scores, where our method receives the best score. + +Video face editing. We compare with pSp, HyperStyle, StyleHEAT [38] and STIT [31]. StyleHEAT uses features of a fixed $64 \times 64$ resolution for unaligned but still cropped $1024 \times 1024$ face reenactment. Specifically, it generates videos based on warping the features of the first frame, which however limits its inversion accuracy. STIT extends PTI [24] for full video processing by stitching. STIT cannot well preserve the complex hair details (Fig. 12, yellow box). As with image face editing, all four baselines are limited + +![](images/ddbca3467ddba59a9c594f65c49979dca63a9010faf87bc7c1cfb6e6c41cf7ea.jpg) +black hair + +![](images/49437caeddab0a76672e1813ff87ab0fc2ff0ae12244e4d92a095a97cf1aa20e.jpg) +young + +![](images/1222658c3a436b757f348c1e635bda5771e68318ec976e81b3ed191c74d4d927.jpg) +editing control +Figure 12: Comparison on video face editing. + +![](images/693498dde616d7c6f6af9058d117bb1c858d2e5029d5b866a09b2c59379f8732.jpg) +Figure 13: Comparison on video face toonification. + +to editing cropped regions, leading to discontinuity along the stitching seams. By comparison, our method uses the first-layer feature and skipped mid-layer features to provide spatial information, which achieves more coherent results. Moreover, we can randomly scale the editing vector $v$ (by multiplying a scale factor) instead of using a fixed $v$ during training. Then during testing, our method can flexibly adjust the editing degree by scaling $v$ for users to select as in Fig. 12(e). + +Video face toonification. Compared with VToonify-T [37], our method preserves more details of the non-face region + +Table 1: User preference scores. Best scores are in bold. + +
Taskpix2pixHD [34]TSIT [12]pSp [23]ours
sketch-to-face0.4000.1210.0290.450
mask-to-face0.1080.0750.1210.696
TaskVToonify-T [37]ours
video toonify0.0830.917
+ +![](images/210fe0e65a5141b0967a19f8385e1d21ee85531e6beb7fadf6ef05e958367eb2.jpg) +(a) input + +![](images/6efefed11ec12bea56cc5fbd6ad726f4ac81da1b0aa68c32059a4cbf56f52865.jpg) +(b) two steps + +![](images/fae8f11de126cb43c57cb92dc05f35ecfe08c8d30c32da49821e4cd9ecf45c35.jpg) +(c) transferred (b) + +![](images/4cc38d86679334248b0dec497d9f6e71f236f8fcc889eb4aad7aa8b38a444e49.jpg) +(d) only opt (e) + +![](images/ee108ec916865ae682756eb3b9fece01c1674b6679fb40767cec0bd96bae4e29.jpg) +e) transferred (d) + +![](images/cc65cb43ba151b11045ed81f8781018dae063c9c77f382c540281d74af2c7f71.jpg) +Figure 14: Effect of encoder in StyleGANEX inversion. +(a) input $x_{\uparrow}$ +Figure 15: Input choice to provide valid style information. + +![](images/f68a7ad0ac727ff24559025487b3b406698d46ff9c6e4801ad5ee6d42320b89b.jpg) +(b) $G(E(x_{1},\tilde{x}_{1},7))$ + +![](images/ac8d38edda084a257454478ba5d355b2047f7639b85fe4a4012d53ed57020d9c.jpg) +(c) $G(E(x_{\uparrow},x_{\uparrow},7))$ + +![](images/b721664c3714187b6575c4df0b76b5f47b490481737bb9e49c947e87073dc619.jpg) +(d) $G(E(x_{\uparrow},\ddot{x}_{\uparrow},7))$ + +and generates shaper faces in Fig. 13. The reason is that VToonify-T uses a fixed latent code extractor while our method trains a joint latent code and feature extractor, thus our method is more powerful for reconstructing the details. Moreover, our method retains StyleGAN's shallow layers, which helps provide key facial features to make the stylized face more vivid. Table 1 shows a quantitative comparison on ten results, and our method obtains the best user score. + +# 5.2. Ablation Study + +The effect of Step II of our two-step inversion is verified in Fig. 6. We further study the effect of Step I in Fig. 14. With Step I providing a good prediction of $w^{+}$ and $f$ , Step II only needs 500-iteration optimization for precise reconstruction (Fig. 14(b)) and valid domain transfer to Disney Princess (Fig. 14(c)). However, if we directly optimize a mean $w^{+}$ and a random $f$ , the result is poor even with 2,000 iterations (indicated by a red arrow in Fig. 14(d)) and the optimized $w^{+}$ and $f$ overfit the input, which is not suitable for editing like domain transfer (Fig. 14(e)). + +In Fig. 15, we study the input choice to extract $w^{+}$ . The cropped aligned faces are the default choice. If we instead use the whole image to extract $w^{+}$ , the background will affect the facial style, leading to poor restoration in Fig. 15(c). We further find reasonable results (Fig. 15(d)) can still be obtained by cropping the input to decrease the background proportion. Note that the face is not aligned in the cropped image $(\ddot{x}_{\uparrow})$ , which is useful for applications like super-resolution where extremely low-resolution faces are hard to detect and align. Users can simply manually crop the face region to provide valid style information. + +![](images/5f092e80676832d6190f65952c4573c65f2b53b258515bf7fabcf531b563c76a.jpg) +(a) input + +![](images/cc4b9ea92ef0092cc60c86678a90e29a14846df855fb7216b51f8ac46e8522df.jpg) +(b) $\ell = 0$ + +![](images/d8c35664d3f2b23439b2792e4a39adbdf274cdefd6e5abdd8bc42d384a85ab79.jpg) +(c) $\ell = 3$ + +![](images/9baacb02cca2dda7e70a2d8ccd85083315fba85c16200a4e2e1fcea822f05b3c.jpg) +(d) $\ell = 7$ + +![](images/cd53acac17a62b70bb19661bc4c53a13f6004012db5bdfaed976842ef321946e.jpg) +Figure 16: Effect of skip connections. +(a) input + +![](images/f047f5d509967c9d03c4f6b32f918b2834de9073944e27bf76f5c316feb352c7.jpg) +(b) $\ell = 1$ + +![](images/bd48b4c7194423db2a9d78901a09ed522666921c19a5eb58719cd0f821986719.jpg) +(c) $\ell = 0$ + +![](images/9d13d7f6fa122bdef8a63cbe05bcf9a2656956ec19840216a0a5b78527881002.jpg) +(d) input + +![](images/b7b22fee3dbfdc405d85ced28e1a9c0e422fd2a99fb86cd90b46104fd44aadec.jpg) +(e) $\ell = 1$ + +![](images/4c1539a8d7a99bae4d56f199870a06f9b58121754f9643206271a6abd7997dc7.jpg) +(f) $\ell = 0$ + +![](images/a39666f4cd49bc3378b174e838d9b1b5fd21e16c893141d32f814e4bf6fb8f2c.jpg) +Figure 17: Performance on low-quality sketches. +input + +![](images/bea45d4ca4152e69942dd8b1186ed07fa0af5bc4ff0501df2e9562b11ef1cc4c.jpg) +inversion + +![](images/f3d09525ba003773acf8646011e375b0ca5950b4ab8a157568f70c7830ac61d5.jpg) +blocky shape + +![](images/427ecd39e97c8dff0b94308ad0ea88dbeb3be7bbff0e0b692b1d29a531471d5c.jpg) +add grass + +![](images/b966d440e3da220da2920b84aa67f6a3c7d7fc2606024cdfb02fd69e912a4afb.jpg) +redness +Figure 18: Performance on car. + +![](images/3c4defc0a8ba0803d5b2f98a4b685dd6a6897496da9f8d0990b91a6c79bf166e.jpg) +style mixing + +![](images/a498d13ebcd693cb8de3eaa147743abc9723ace36bc5cdf625ba70da4509aac4.jpg) + +In Fig. 16, we study the effect of skip connection. Without it $(\ell = 0)$ , the glasses cannot be generated. Skip connection provides mid-layer features to preserve the details of the input. However, too many extra features will alter the feature distribution of StyleGAN, e.g., always generating sunglasses as the input has no segmentation of eyes. Thus, we use $\ell = 3$ to balance between input-output consistency and diversity. Inversely, we can use a small $\ell$ to enhance the model robustness to low-quality inputs. For example, we can generate more realistic faces with $\ell = 0$ on DeepFace-Drawing low-quality sketches [7] as in Fig. 17. + +# 5.3. Results on Non-Facial Dataset + +The proposed refactoring is domain-agnostic, thus it can be applied to StyleGAN pre-trained on other domains like cars. An example with a wide vertical field of view is shown in Fig. 18, where we perform optimization-based StyleGAN inversion, image editing with the editing vectors found by GANSpace [10] and style mixing on the refactored model. The resulting manipulated cars look plausible. + +![](images/2d1d5cbd21f6323168486be213ea2f578478608e78722b2310251bdbd0e23a5c.jpg) +(c) severely rotated face +Figure 19: Limitations. + +# 5.4. Limitations + +First, our framework currently relies on an inefficient optimization process for precise reconstruction. While future work can explore more efficient inversion methods (e.g., iterative residue prediction and hyper networks), it is important to note that this paper focuses on overcoming the fixed-crop limitation of StyleGAN, rather than the specific topic of GAN inversion. Second, StyleGANEX is limited by the feature representation of StyleGAN. While it shows great potential in handling normal FoV face images, out-of-distribution features such as complex clothing and human bodies may not be well handled as in Fig. 19(a)(b). As in Fig. 19(c), while our method can handle faces rotated 20 degrees, it still struggles with handling large rotation angles. However, this can be easily resolved by rough alignment of the input image in the middle as our method does not require accurate alignment. Finally, after translation, we can undo the alignment to obtain a plausible result as in right of Fig. 19(c). Third, StyleGANEX, like StyleGAN, focuses on face manipulation and may not support out-of-distribution semantical editing of non-facial regions. Last, StyleGANEX may inherit the model bias of StyleGAN. Applying it to tasks with severe data imbalance might lead to unsatisfactory results on under-represented data. + +# 6. Conclusion + +We have presented an effective approach to refactor StyleGAN to overcome its fixed-crop limitation while retaining its style control abilities. The refactored model, called StyleGANEX, fully inherits the parameters of the pre-trained StyleGAN without retraining, and is thus fully compatible with the generative space of StyleGAN. We further introduced a StyleGANEX encoder to project normal FoV face images to the joint $W^{+} - F$ space of StyleGANEX for real face inversion and manipulation. Our approach offers an effective solution to the general issue of StyleGAN and extends its capability beyond fixed-resolution data. + +Acknowledgments. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is al + +so supported by MOE AcRF Tier 1 (2021-T1-001-088) and NTU NAP grant. + +# References + +[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proc. Int'l Conf. Computer Vision, pages 4432-4441, 2019. 1, 2 +[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2 +[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinements. In Proc. Int'l Conf. Computer Vision, pages 6711–6720, 2021. 2 +[4] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3, 2022. 2 +[5] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 2, 6 +[6] Kelvin CK Chan, Xintao Wang, Xiangyu Xu, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for large-factor image super-resolution. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2021. 3, 4 +[7] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. ACM Transactions on Graphics, 39(4):72-1, 2020. 8 +[8] Tan M Dinh, Anh Tuan Tran, Rang Nguyen, and Binh-Son Hua. Hyperinverter: Improving stylegan inversion via hypernetwork. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11389-11398, 2022. 2 +[9] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clipguided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 5, 6 +[10] Erik Härkönen, Aaron Hertzman, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Advances in Neural Information Processing Systems, 2020. 1, 2, 3, 8 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 770-778, 2016. 5 +[12] Liming Jiang, Changxu Zhang, Mingyang Huang, Chunxiao Liu, Jianping Shi, and Chen Change Loy. Tset: A simple and versatile framework for image-to-image translation. In Proc. European Conf. Computer Vision, pages 206-222. Springer, 2020. 7, 8 + +[13] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In Proc. Int'l Conf. Computer Vision, 2021. 2 +[14] Kyoungkook Kang, Seongtae Kim, and Sunghyun Cho. Gan inversion for out-of-range images with geometric transformations. In Proc. Int'l Conf. Computer Vision, pages 13941-13949, 2021. 3, 6 +[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems, 2020. 6 +[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 2 +[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 3, 6 +[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1 +[19] Feng-Lin Liu, Shu-Yu Chen, Yukun Lai, YueRen Jiang, Hongbo Fu, and Lin Gao. Deepfacevideoediting: Sketch-based deep editing of face videos. ACM Transactions on Graphics, 41(4):167, 2022. 2 +[20] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, JunYan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11399-11409, 2022. 3 +[21] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 2085–2094, October 2021. 2 +[22] Justin NM Pinkney and Doron Adler. Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334, 2020. 5 +[23] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 8 +[24] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics, 2022. 2, 7 +[25] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. In Proc. Int'l Conf. Computer Vision, 2019. 6 +[26] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 1, 2, 3, 5, 6 + +[27] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2 +[28] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 2, 3 +[29] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics, 39(6):1-14, 2020. 1, 2 +[30] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 2 +[31] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia, pages 1–9, 2022. 2, 7 +[32] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. Stylegan2 distillation for feed-forward image manipulation. In Proc. European Conf. Computer Vision, pages 170-186. Springer, 2020. 3 +[33] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 2, 3 +[34] Ting Chun Wang, Ming Yu Liu, Jun Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2018. 7, 8 +[35] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proc. Int'l Conf. Computer Vision, pages 1905-1914, 2021. 6 +[36] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. StyleAlign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 1 +[37] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Vtoonify: Controllable high-resolution portrait video style transfer. ACM Transactions on Graphics, 41(6):1-15, 2022. 2, 3, 4, 5, 7, 8 +[38] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In Proc. European Conf. Computer Vision, pages 85-101. Springer, 2022. 3, 7 +[39] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proc. European Conf. Computer Vision, pages 334-349. Springer, 2018. 6 +[40] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zheng-Jun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces + +in gans. In Advances in Neural Information Processing Systems, volume 34, 2021. 1, 3, 6 +[41] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proc. European Conf. Computer Vision, pages 592-608. Springer, 2020. 2 +[42] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Transactions on Graphics, 40(6):1-13, 2021. 3 \ No newline at end of file diff --git a/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/images.zip b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ec2b2fbabc9220af5b75baff52cb2cece369fb1d --- /dev/null +++ b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f43b6f5baa3acf8eea4fa39cf066912f3004b8874c19a02587c047f9b97db65 +size 964071 diff --git a/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/layout.json b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7ec578886776629d9642aba6cb3bb62105d7a02a --- /dev/null +++ b/2023/StyleGANEX_ StyleGAN-Based Manipulation Beyond Cropped Aligned Faces/layout.json @@ -0,0 +1,13911 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 538, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 538, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 538, + 122 + ], + "type": "text", + "content": "StyleGANEX: StyleGAN-Based Manipulation Beyond Cropped Aligned Faces" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 140, + 459, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 140, + 459, + 169 + ], + "spans": [ + { + "bbox": [ + 135, + 140, + 459, + 169 + ], + "type": "text", + "content": "Shuai Yang Liming Jiang Ziwei Liu Chen Change Loy S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 171, + 437, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 171, + 437, + 183 + ], + "spans": [ + { + "bbox": [ + 156, + 171, + 437, + 183 + ], + "type": "text", + "content": "{shuai.yang, liming002, ziwei.liu, ccloy}@ntu.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 196, + 416, + 273 + ], + "blocks": [ + { + "bbox": [ + 50, + 196, + 416, + 273 + ], + "lines": [ + { + "bbox": [ + 50, + 196, + 416, + 273 + ], + "spans": [ + { + "bbox": [ + 50, + 196, + 416, + 273 + ], + "type": "image", + "image_path": "24138a5cf244f6b66f18abea521c020c4e2094f264edfe0ec5904e992b1dc660.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 425, + 196, + 544, + 273 + ], + "blocks": [ + { + "bbox": [ + 425, + 196, + 544, + 273 + ], + "lines": [ + { + "bbox": [ + 425, + 196, + 544, + 273 + ], + "spans": [ + { + "bbox": [ + 425, + 196, + 544, + 273 + ], + "type": "image", + "image_path": "7bbe1fc312a04d55ea14d1b6bc2371a42c134998921f9b81c60e1279a2842e3e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 278, + 293, + 350 + ], + "blocks": [ + { + "bbox": [ + 49, + 278, + 293, + 350 + ], + "lines": [ + { + "bbox": [ + 49, + 278, + 293, + 350 + ], + "spans": [ + { + "bbox": [ + 49, + 278, + 293, + 350 + ], + "type": "image", + "image_path": "8a02fde41a725b936f700d5f0efbc0abfab6eb6b32fc265ab506ddaf142dec22.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 298, + 277, + 544, + 350 + ], + "blocks": [ + { + "bbox": [ + 298, + 277, + 544, + 350 + ], + "lines": [ + { + "bbox": [ + 298, + 277, + 544, + 350 + ], + "spans": [ + { + "bbox": [ + 298, + 277, + 544, + 350 + ], + "type": "image", + "image_path": "ec4c3f7113165c993a297da4c4cffb69a314dc846787ec58ef3ead13b3ac105c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 354, + 544, + 435 + ], + "blocks": [ + { + "bbox": [ + 50, + 354, + 544, + 435 + ], + "lines": [ + { + "bbox": [ + 50, + 354, + 544, + 435 + ], + "spans": [ + { + "bbox": [ + 50, + 354, + 544, + 435 + ], + "type": "image", + "image_path": "4303ceac55161479649b24e326d7a7084fabb25da09bafd8eb88d9a8ac5641fa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 441, + 544, + 455 + ], + "lines": [ + { + "bbox": [ + 47, + 441, + 544, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 544, + 455 + ], + "type": "text", + "content": "Figure 1: We expand StyleGAN to encompass a diverse set of tasks that go beyond the constraints of cropped aligned faces." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 464, + 192, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 464, + 192, + 476 + ], + "spans": [ + { + "bbox": [ + 143, + 464, + 192, + 476 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 483, + 290, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 290, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 290, + 699 + ], + "type": "text", + "content": "Recent advances in face manipulation using StyleGAN have produced impressive results. However, StyleGAN is inherently limited to cropped aligned faces at a fixed image resolution it is pre-trained on. In this paper, we propose a simple and effective solution to this limitation by using dilated convolutions to rescale the receptive fields of shallow layers in StyleGAN, without altering any model parameters. This allows fixed-size small features at shallow layers to be extended into larger ones that can accommodate variable resolutions, making them more robust in characterizing unaligned faces. To enable real face inversion and manipulation, we introduce a corresponding encoder that provides the first-layer feature of the extended StyleGAN in addition to the latent style code. We validate the effectiveness of our method using unaligned face inputs of various resolutions in a diverse set of face manipulation tasks, including facial attribute editing, super-resolution, sketch/mask-to-face translation, and face toonification. Project page https:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 700, + 280, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 700, + 280, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 700, + 280, + 712 + ], + "type": "text", + "content": "//www.mmlab-ntu.com/project/styleganex" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 464, + 387, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 387, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 387, + 477 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 485, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 628 + ], + "type": "text", + "content": "StyleGAN [17, 18] has emerged as one of the most successful models for generating high-quality faces. Building upon StyleGAN, researchers have developed a range of face manipulation models [1, 23, 29, 26, 10, 28, 40, 36]. These models typically map real face images or other face-related inputs to the latent space of StyleGAN, perform semantic editing in the latent space, and then map the edited latent code back to the image space. This approach enables a variety of tasks, including facial attribute editing, face restoration, sketch-to-face translation, and face toonification. As the manipulated faces remain within the generative space of StyleGAN, the quality of the image output is guaranteed." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Despite its ability to ensure high-quality image output, the generative space of StyleGAN is limited by a fixed-crop constraint that restricts image resolution and face layout. As a result, existing face manipulation models based on StyleGAN can only handle cropped and aligned face images. In such images with a limited field of view (FoV), the face typically dominates the image, leaving little room for back-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21000" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "ground and clothing, and often resulting in partially cropped hair. However, in everyday portrait photos such as selfies, faces occupy a smaller proportion of the image, allowing for a complete hairstyle and upper body seen. Portrait videos such as those from live streaming require an even larger background area to accommodate face movement. To process these types of inputs, which we refer to as normal FoV face images and videos, existing manipulation models need to align, crop, and edit the face before pasting the result back onto the original image [4, 19, 31]. This approach often results in discontinuity near the seams, e.g., only editing the hair color inside the cropped area." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 410 + ], + "type": "text", + "content": "While StyleGAN3 [16] was introduced to address unaligned faces, a recent study [4] found that even StyleGAN3 requires face realignment before effectively projecting to its latent space. Moreover, StyleGAN3 is still constrained by a fixed image resolution. Motivated by the translation equivariance of convolutions, VToonify [37] addresses the fixed-crop limitation of StyleGAN by removing its shallow layers to accept input features of any resolution. However, these shallow layers are crucial for capturing high-level features of the face, such as pose, hairstyle, and face shape. By removing these layers, the network loses its ability to perform latent editing on these important features, which is a distinctive capability of StyleGAN. Therefore, the challenge remains in overcoming the fixed-crop limitation of StyleGAN while preserving its original style manipulation abilities, which is a valuable research problem to solve." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 628 + ], + "type": "text", + "content": "In this paper, we propose a simple yet effective approach for refactoring StyleGAN to overcome the fixed-crop limitation. In particular, we refactor its shallow layers instead of removing them, allowing the first layer to accept input features of any resolution. This simple change expands StyleGAN's style latent space into a more powerful joint style latent and first-layer feature space " + }, + { + "bbox": [ + 46, + 411, + 289, + 628 + ], + "type": "inline_equation", + "content": "(W^{+} - F" + }, + { + "bbox": [ + 46, + 411, + 289, + 628 + ], + "type": "text", + "content": " space), extending the generative space beyond cropped aligned faces. Furthermore, our refactoring only changes the receptive field of shallow-layer convolutions, leaving all pre-trained model parameters intact. Hence, the refactored StyleGAN (StyleGANEX) can directly load the original StyleGAN parameters, fully compatible with the generative space of StyleGAN, and retains its style representation and editing ability. This means that the StyleGAN editing vectors found in previous studies [26, 10, 28] can be directly applied to StyleGANEX for normal FoV face editing, e.g., changing the face pose, as shown in Fig. 1(a)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "content": "Based on StyleGANEX, we further design a corresponding encoder that projects normal FoV face images to the " + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "inline_equation", + "content": "W^{+} - F" + }, + { + "bbox": [ + 46, + 630, + 289, + 715 + ], + "type": "text", + "content": " space for real face inversion and manipulation. Our encoder builds upon pSp encoder [23] and aggregates its multi-layer features to predict the first-layer feature of StyleGANEX. The encoder and StyleGANEX form a fully convolutional encoder-decoder framework. With the first-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "layer feature as the bottleneck layer, whose resolution is " + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "inline_equation", + "content": "1/32" + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": " of the output image, our framework can handle images and videos of various resolutions, as long as their side lengths are divisible by 32. Depending on the input and output types, our framework can perform a wide range of face manipulation tasks. In this paper, we select several representative tasks, as shown in Fig. 1, including facial attribute editing, face super-resolution, sketch/mask-to-face translation and video face toonification. While the focus of these tasks in the past is limited to cropped aligned faces, our framework can handle normal FoV faces, showing significant advantages over previous StyleGAN-based approaches. To summarize, our main contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 230, + 545, + 320 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 315, + 230, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 230, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 545, + 266 + ], + "type": "text", + "content": "- A novel StyleGANEX architecture with extended " + }, + { + "bbox": [ + 315, + 230, + 545, + 266 + ], + "type": "inline_equation", + "content": "W^{+} - F" + }, + { + "bbox": [ + 315, + 230, + 545, + 266 + ], + "type": "text", + "content": " space, which overcomes the fixed-crop limitation of StyleGAN." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 269, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 269, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 315, + 269, + 545, + 293 + ], + "type": "text", + "content": "- An effective encoder that is able to project normal FoV face images into the " + }, + { + "bbox": [ + 315, + 269, + 545, + 293 + ], + "type": "inline_equation", + "content": "W^{+} - F" + }, + { + "bbox": [ + 315, + 269, + 545, + 293 + ], + "type": "text", + "content": " domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 296, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 296, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 315, + 296, + 545, + 320 + ], + "type": "text", + "content": "- A generic and versatile fully convolutional framework for face manipulation beyond cropped aligned faces." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 333, + 392, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 392, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 392, + 345 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": "StyleGAN inversion. StyleGAN inversion aims at projecting real face images into the latent space of StyleGAN for further manipulation. Image2StyleGAN [1] analyzes the latent space and proposes " + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": " space to reconstruct real faces with latent code optimization. PIE [29] and IDinvert [41] further consider the editability of the latent code during optimization. To speed up inversion, pSp [23] and e4e [30] train an encoder to directly project the target face to its corresponding latent code, which is however hard to reconstruct fine details and handle occlusions. To solve this issue, Restyle [3] and HFGI [33] predict the residue of latent codes or mid-layer features to reduce errors, respectively. Instead of focusing on the latent code, PTI [24] optimizes StyleGAN itself to fit a target face, which is accelerated by HyperInverter [8] and HyperStyle [5] to predict offsets of network parameters with hyper networks. The above methods are limited to cropped aligned faces for valid face editing. With the extended " + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "inline_equation", + "content": "W^{+} - F" + }, + { + "bbox": [ + 304, + 354, + 547, + 582 + ], + "type": "text", + "content": " space, our framework is able to perform inversion on normal FoV face images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "StyleGAN-based face manipulation. An intuitive way of StyleGAN-based face manipulation is to optimize the latent code online to achieve certain objectives such as a pixel-level constraints [1], sketch-based structure constraints [19] or text-guided semantic constraints [21, 9]. Another way is to search for offline editing vectors to add to the latent code for manipulation. Supervised methods identify meaningful editing vectors based on attribute labels or pretrained classifiers [26, 27, 13, 2]. On the other hand, unsupervised methods statistically analyze the StyleGAN latent space to discover semantically significant editing direc" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "21001" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": "tions by principal component analysis [10], low-rank factorization [40] and closed-form factorization [28]. Meanwhile, face manipulation can be directly realized by image-to-image translation frameworks, where StyleGAN is used to generate paired training data [32, 37] or to build the decoder [23, 6, 37]. To achieve spatial editing, methods [42, 20] are proposed to manipulate mid-layer features in addition to the latent code, similar to our " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "W^{+} - F" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": " space. Moreover, BDInvert [14] and StyleHEAT [38] introduces feature transformations for unaligned face editing. However, the above methods, as well as HFGI [33], follow StyleGAN features' original fixed resolution, thus still suffering the crop limitation. Differently, our method predicts the first-layer feature that can have various resolutions in StyleGANEX. Moreover, as we will show later in Sec. 5.1, compared to VToonify [37], StyleGANEX retains the complete style manipulation abilities of the shallow layers of StyleGAN, with a jointly trained latent code and feature extractor, preserving vivid details and supporting more diverse manipulation tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 325, + 132, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 132, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 132, + 338 + ], + "type": "text", + "content": "3. StyleGANEX" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 346, + 248, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 248, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 248, + 358 + ], + "type": "text", + "content": "3.1. Analysis of the Fixed-Crop Limitation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 365, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 287, + 484 + ], + "type": "text", + "content": "StyleGAN has great potential for handling normal FoV face images. The generator of StyleGAN is a fully convolutional architecture that can naturally handle different feature resolutions, as convolution operations and the style modulation of StyleGAN are independent of the input resolution. Additionally, the translation equivariance of convolution operations naturally supports feature translation. As analyzed in VToonify [37], if we translate or rotate the feature of the 7th layer of StyleGAN, the resulting face will also be shifted or rotated, as shown in Figs. 2(a)(b)(d)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "text", + "content": "The limiting factor originates from StyleGAN's constant first-layer feature. First, the first-layer feature has a fixed resolution of " + }, + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "text", + "content": ", limiting the output to " + }, + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "text", + "content": " resolution. Second, " + }, + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "text", + "content": " resolution is inadequate to characterize the spatial information of unaligned faces. We have taken a step further than the analysis in VToonify [37] to investigate translation and rotation on the first-layer feature. As shown in Fig. 2(c), sub-pixel translation fuses adjacent feature values severely, resulting in a blurry face due to the small number of elements (only 16) in the first-layer feature. In Fig. 2(e), the first-layer feature fails to provide enough spatial information for a valid rotation. In comparison, the 7-th layer has a higher resolution (" + }, + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 47, + 486, + 289, + 714 + ], + "type": "text", + "content": "), making it better suited for capturing spatial information. However, only a single layer alone provides limited style control, as the full facial structural styles are hierarchically modeled by seven shallow layers. Simply ignoring low-resolution layers, as in VToonify, disables flexible latent editing over the styles of these layers, such as pose, age, and face shape." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 356, + 118 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 356, + 118 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 356, + 118 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 356, + 118 + ], + "type": "image", + "image_path": "4dd2f7b042a90d2a10d4f3c078d29d1a52f4fef4d160a1993f7dff28e2959671.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 118, + 351, + 127 + ], + "lines": [ + { + "bbox": [ + 309, + 118, + 351, + 127 + ], + "spans": [ + { + "bbox": [ + 309, + 118, + 351, + 127 + ], + "type": "text", + "content": "(a) random face" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 357, + 72, + 402, + 118 + ], + "blocks": [ + { + "bbox": [ + 357, + 72, + 402, + 118 + ], + "lines": [ + { + "bbox": [ + 357, + 72, + 402, + 118 + ], + "spans": [ + { + "bbox": [ + 357, + 72, + 402, + 118 + ], + "type": "image", + "image_path": "ee0a5d11c10510ab1c005852e1af646a2c024584a33bab0164c9d9ec7b54fdda.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 118, + 399, + 127 + ], + "lines": [ + { + "bbox": [ + 358, + 118, + 399, + 127 + ], + "spans": [ + { + "bbox": [ + 358, + 118, + 399, + 127 + ], + "type": "text", + "content": "(b) shift layer#7" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 449, + 118 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 449, + 118 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 449, + 118 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 449, + 118 + ], + "type": "image", + "image_path": "061f085f73cb3728542f5d6c99e17c0023609d63814507960f215398d6169b7f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 118, + 445, + 127 + ], + "lines": [ + { + "bbox": [ + 405, + 118, + 445, + 127 + ], + "spans": [ + { + "bbox": [ + 405, + 118, + 445, + 127 + ], + "type": "text", + "content": "(c) shift layer#1" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 450, + 72, + 496, + 118 + ], + "blocks": [ + { + "bbox": [ + 450, + 72, + 496, + 118 + ], + "lines": [ + { + "bbox": [ + 450, + 72, + 496, + 118 + ], + "spans": [ + { + "bbox": [ + 450, + 72, + 496, + 118 + ], + "type": "image", + "image_path": "6476d53f8cdbeb96d51bd3412c8ce99ed95c5f3a1bc629f71ce4c40a66060e20.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 118, + 495, + 127 + ], + "lines": [ + { + "bbox": [ + 451, + 118, + 495, + 127 + ], + "spans": [ + { + "bbox": [ + 451, + 118, + 495, + 127 + ], + "type": "text", + "content": "(d) rotate layer#7" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 499, + 72, + 545, + 118 + ], + "blocks": [ + { + "bbox": [ + 499, + 72, + 545, + 118 + ], + "lines": [ + { + "bbox": [ + 499, + 72, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 499, + 72, + 545, + 118 + ], + "type": "image", + "image_path": "d4bc1f3feb8f3bb0efae3c3d5cfdbceb0452a94d7d4bb04c743b9db9e226bfd2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 118, + 542, + 127 + ], + "lines": [ + { + "bbox": [ + 499, + 118, + 542, + 127 + ], + "spans": [ + { + "bbox": [ + 499, + 118, + 542, + 127 + ], + "type": "text", + "content": "(e) rotate layer#1" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 308, + 212, + 366, + 266 + ], + "blocks": [ + { + "bbox": [ + 305, + 134, + 547, + 206 + ], + "lines": [ + { + "bbox": [ + 305, + 134, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 134, + 547, + 206 + ], + "type": "text", + "content": "Figure 2: Analysis of StyleGAN in generating unaligned faces. (a) Generated face. (b)(c) Face generated by translating the feature maps of the 7-th and 1-st layers of StyleGAN respectively to shift the face by 150 pixels. (d)(e) Face generated by rotating the feature maps of the 7-th and 1-st layers of StyleGAN by 10 degrees, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 212, + 366, + 266 + ], + "lines": [ + { + "bbox": [ + 308, + 212, + 366, + 266 + ], + "spans": [ + { + "bbox": [ + 308, + 212, + 366, + 266 + ], + "type": "image", + "image_path": "e9e946eeb589cb141531e090236900e04ad9959a2c1d23adee6eb206fc75a298.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 267, + 365, + 277 + ], + "lines": [ + { + "bbox": [ + 310, + 267, + 365, + 277 + ], + "spans": [ + { + "bbox": [ + 310, + 267, + 365, + 277 + ], + "type": "text", + "content": "(a) real face photo" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 284, + 546, + 356 + ], + "lines": [ + { + "bbox": [ + 305, + 284, + 546, + 356 + ], + "spans": [ + { + "bbox": [ + 305, + 284, + 546, + 356 + ], + "type": "text", + "content": "Figure 3: Analysis of StyleGANEX in generating unaligned faces. (a) A real face photo. (b) Reconstructed face by projecting (a) into the " + }, + { + "bbox": [ + 305, + 284, + 546, + 356 + ], + "type": "inline_equation", + "content": "W^{+} - F" + }, + { + "bbox": [ + 305, + 284, + 546, + 356 + ], + "type": "text", + "content": " space of StyleGANEX. (c) Face generated by translating the first-layer feature map to shift the face by 300 pixels. (d) Face generated by rotating the first-layer feature map by 10 degrees." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 368, + 213, + 426, + 266 + ], + "blocks": [ + { + "bbox": [ + 368, + 213, + 426, + 266 + ], + "lines": [ + { + "bbox": [ + 368, + 213, + 426, + 266 + ], + "spans": [ + { + "bbox": [ + 368, + 213, + 426, + 266 + ], + "type": "image", + "image_path": "8b0baa2f0e2280d354352224c86f6fa41813a265bcbae6172a37037c66cdb512.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 267, + 423, + 277 + ], + "lines": [ + { + "bbox": [ + 370, + 267, + 423, + 277 + ], + "spans": [ + { + "bbox": [ + 370, + 267, + 423, + 277 + ], + "type": "text", + "content": "(b) projected face" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 427, + 213, + 485, + 266 + ], + "blocks": [ + { + "bbox": [ + 427, + 213, + 485, + 266 + ], + "lines": [ + { + "bbox": [ + 427, + 213, + 485, + 266 + ], + "spans": [ + { + "bbox": [ + 427, + 213, + 485, + 266 + ], + "type": "image", + "image_path": "32e1b5de8500028b7611a25ef1cbcaebd62ff1864d7ef0afa23ded9ee02094cc.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 267, + 479, + 277 + ], + "lines": [ + { + "bbox": [ + 432, + 267, + 479, + 277 + ], + "spans": [ + { + "bbox": [ + 432, + 267, + 479, + 277 + ], + "type": "text", + "content": "(c) shift layer#1" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 487, + 213, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 487, + 213, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 487, + 213, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 487, + 213, + 545, + 266 + ], + "type": "image", + "image_path": "1cf618f374e035e114bf8f014572bef5422d516786723eccc01b70f578c1e7f0.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 267, + 541, + 277 + ], + "lines": [ + { + "bbox": [ + 489, + 267, + 541, + 277 + ], + "spans": [ + { + "bbox": [ + 489, + 267, + 541, + 277 + ], + "type": "text", + "content": "(d) rotate layer#1" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "text", + "content": "Expanding the shallow layers of StyleGAN to have the same " + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "text", + "content": " resolution as the 7th layer, or more generally, " + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "inline_equation", + "content": "H / 32 \\times W / 32" + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "text", + "content": " resolution for an " + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "text", + "content": " image, would provide enough structure and layout information to combine the style controllability of shallow layers with support for normal FoV faces. This is precisely the key idea of StyleGANEX, and we will introduce our simple solution in Section 3.2. As a preview of the performance of the expanded layers in enabling face manipulation beyond cropped and aligned faces, Fig. 3 shows that for a " + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "inline_equation", + "content": "1472 \\times 1600" + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "text", + "content": " normal FoV face photo, we can obtain its latent code and an additional " + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "inline_equation", + "content": "46 \\times 50" + }, + { + "bbox": [ + 305, + 373, + 546, + 577 + ], + "type": "text", + "content": " first-layer feature as the input to StyleGANEX (inversion method explained in Section 4.2). Face translation and rotation can be realized by shifting or rotating the first-layer feature. Additionally, the face can be effectively edited by applying style mixing [17] or InterFaceGAN editing vectors [26] to the latent code, as shown in Fig. 1(a)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 586, + 482, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 482, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 482, + 599 + ], + "type": "text", + "content": "3.2. From StyleGAN to StyleGANEX" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 547, + 713 + ], + "type": "text", + "content": "Figure 4 illustrates the generator architectures of StyleGAN and StyleGANEX. Compared to StyleGAN, we first replace the constant " + }, + { + "bbox": [ + 305, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 305, + 605, + 547, + 713 + ], + "type": "text", + "content": " first-layer feature with a variable feature whose resolution is " + }, + { + "bbox": [ + 305, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "1/32" + }, + { + "bbox": [ + 305, + 605, + 547, + 713 + ], + "type": "text", + "content": " of the output image. Then, we remove the upsample operations before the 8-th layer, allowing features in the seven shallow layers to share the same resolution as the 7-th layer. However, the convolution kernels or reception fields of these layers do not match their input features with the enlarged resolution. To solve" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21002" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 160, + 249 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 160, + 249 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 160, + 249 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 160, + 249 + ], + "type": "image", + "image_path": "5bd3a5e9b589ed01fc97536ed7581464501c396357edf0264444fc3b98203b13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 70, + 286, + 281 + ], + "blocks": [ + { + "bbox": [ + 170, + 70, + 286, + 281 + ], + "lines": [ + { + "bbox": [ + 170, + 70, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 170, + 70, + 286, + 281 + ], + "type": "image", + "image_path": "a283461e5c0acbaa9c835876077a2b31484bfab1e735c0556542563932f5a3cb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 289, + 287, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 314 + ], + "type": "text", + "content": "Figure 4: Refactor StyleGAN to StyleGANEX. For simplicity, learned weights, biases and noises are omitted." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "type": "text", + "content": "this problem, we enlarge the reception fields by modifying the convolutions to their dilated versions. For example, the first layer only needs to change the dilation factor from 1 to 8. With this simple modification, StyleGAN is refactored to StyleGANEX. Since the first layer becomes variable, the original " + }, + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "type": "text", + "content": " latent space is extended to a joint " + }, + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "type": "inline_equation", + "content": "W^{+} - F" + }, + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "type": "text", + "content": " space, where the latent code " + }, + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "type": "inline_equation", + "content": "w^{+} \\in W^{+}" + }, + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "type": "text", + "content": " provides style cues, and the first-layer feature " + }, + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "type": "inline_equation", + "content": "f \\in F" + }, + { + "bbox": [ + 46, + 329, + 287, + 435 + ], + "type": "text", + "content": " mainly encodes spatial information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 288, + 641 + ], + "type": "text", + "content": "The refactoring of StyleGAN to StyleGANEX has three key advantages. 1) Support for unaligned faces. The resolution enlargement and variable first-layer features of StyleGANEX overcome the fixed-crop limitation. 2) Compatibility. No model parameters are altered during refactoring, meaning that StyleGANEX can directly load pretrained StyleGAN parameters without retraining. In fact, if we upsample the StyleGAN's constant input feature by " + }, + { + "bbox": [ + 46, + 437, + 288, + 641 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 46, + 437, + 288, + 641 + ], + "type": "text", + "content": " with nearest neighbor interpolation to serve as " + }, + { + "bbox": [ + 46, + 437, + 288, + 641 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 437, + 288, + 641 + ], + "type": "text", + "content": " of StyleGANEX, StyleGANEX degrades exactly to StyleGAN with the same " + }, + { + "bbox": [ + 46, + 437, + 288, + 641 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 437, + 288, + 641 + ], + "type": "text", + "content": " generative space. The computational cost of the refactoring is also minimal, with generating an image taking 0.026s and 0.028s for StyleGAN and StyleGANEX, respectively. 3) Flexible manipulation. StyleGANEX retains the style representation and editing ability of StyleGAN, meaning that abundant StyleGAN-based face manipulation techniques can be applied to StyleGANEX." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 651, + 256, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 256, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 256, + 664 + ], + "type": "text", + "content": "4. Face Manipulation with StyleGANEX" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 175, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 175, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 175, + 683 + ], + "type": "text", + "content": "4.1. StyleGANEX Encoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "This section introduces our StyleGANEX encoder " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " which is used to project real face images into the " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "W^{+} - F" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 545, + 239 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 239 + ], + "type": "image", + "image_path": "143eae5cf6a3f2e6efba528a654ba7cb1192618df0df4e3509be087d8f5745b2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 249, + 517, + 261 + ], + "lines": [ + { + "bbox": [ + 333, + 249, + 517, + 261 + ], + "spans": [ + { + "bbox": [ + 333, + 249, + 517, + 261 + ], + "type": "text", + "content": "Figure 5: Details of StyleGANEX Encoder." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": "space of StyleGANEX " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": ". The encoder builds upon the pSp encoder [23], as depicted in Fig. 5. Specifically, for " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " space, we concatenate pSp features in the middle layers and add a convolution layer to map the concatenated features to the first-layer input feature " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "W^{+}" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " space, the original pSp encoder takes a " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " image as input and convolves it to eighteen " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "1 \\times 1 \\times 512" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " features to map to a latent code " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "w^{+} \\in \\mathbb{R}^{18 \\times 512}" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": ". To make " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " accept more general " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " images, we add global average pooling to resize all features to " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "1 \\times 1 \\times 512" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " before mapping to latent codes. To support various face manipulation tasks flexibly, we can extract " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " from different sources. Let " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "E_{F}" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "E_{W}" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " be the operation of " + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 267, + 545, + 435 + ], + "type": "text", + "content": " to extract the first-layer feature and the latent code, respectively. We have" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 340, + 443, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 443, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 340, + 443, + 545, + 456 + ], + "type": "interline_equation", + "content": "f, w ^ {+} = E _ {F} \\left(x _ {1}\\right), E _ {W} \\left(x _ {2}\\right) := E \\left(x _ {1}, x _ {2}\\right), \\tag {1}", + "image_path": "8f786d59860afaf19aeba4cc2a6e2e0e15680b9b33124250031378bb295110c7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": " are the source inputs for face layout and face style, respectively. Then, a general form of image generation by " + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "inline_equation", + "content": "G(E_{F}(x_{1}), E_{W}(x_{2}))" + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": ". In some face manipulation tasks like super-resolution [6] and toonification [37], passing encoder features to the generator via skip connections helps preserve the details of the input image. We thus introduce a scalar parameter " + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": " to the generation process, indicating the " + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": " shallow layers of " + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": " receive the encoder features " + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "inline_equation", + "content": "(\\ell = 0" + }, + { + "bbox": [ + 304, + 465, + 545, + 574 + ], + "type": "text", + "content": " means no skip connections):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 582, + 545, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 582, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 318, + 582, + 545, + 595 + ], + "type": "interline_equation", + "content": "\\hat {x} = G \\left(E _ {F} \\left(x _ {1}, \\ell\\right), E _ {W} \\left(x _ {2}\\right)\\right) := G \\left(E \\left(x _ {1}, x _ {2}, \\ell\\right)\\right), \\tag {2}", + "image_path": "e4cc8aff49defa8f4bd217cafb6cfa5efbaca82fbb34618dad78ef0afadd76f2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "E_{F}(x_{1},\\ell)" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " provides both " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " and the skipped encoder features. For " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " resolution, " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " and the generated image " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " will be of " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "H / 8\\times W / 8" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "4H\\times 4W" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " , respectively. The resolution of " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " can be independent of " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\hat{x}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 659, + 495, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 495, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 495, + 672 + ], + "type": "text", + "content": "4.2. StyleGANEX Inversion and Editing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "To find appropriate " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{w}^+" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " that precisely reconstruct a target image " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", we perform a two-step StyleGANEX inversion. Step I projects " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to initial " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "w^+" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ". Step" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21003" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 108, + 129 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 108, + 129 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 108, + 129 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 108, + 129 + ], + "type": "image", + "image_path": "d5c2194e9f68e66133808db72ab55b3dbf92a60fba7124c8e47f48d7b79cbf21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 131, + 100, + 141 + ], + "lines": [ + { + "bbox": [ + 54, + 131, + 100, + 141 + ], + "spans": [ + { + "bbox": [ + 54, + 131, + 100, + 141 + ], + "type": "text", + "content": "(a) face photo" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 72, + 167, + 129 + ], + "blocks": [ + { + "bbox": [ + 108, + 72, + 167, + 129 + ], + "lines": [ + { + "bbox": [ + 108, + 72, + 167, + 129 + ], + "spans": [ + { + "bbox": [ + 108, + 72, + 167, + 129 + ], + "type": "image", + "image_path": "4f16ec3423c7f14004d8ad934ce3800017349edcbdd61d2339d7d7b39cc2acd7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 131, + 164, + 141 + ], + "lines": [ + { + "bbox": [ + 110, + 131, + 164, + 141 + ], + "spans": [ + { + "bbox": [ + 110, + 131, + 164, + 141 + ], + "type": "text", + "content": "(b) encoded face" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 72, + 225, + 129 + ], + "blocks": [ + { + "bbox": [ + 167, + 72, + 225, + 129 + ], + "lines": [ + { + "bbox": [ + 167, + 72, + 225, + 129 + ], + "spans": [ + { + "bbox": [ + 167, + 72, + 225, + 129 + ], + "type": "image", + "image_path": "b514d10c07b6b37a66d09e4d1f77307de3b189e39a61e433c3fea75bb3006e3b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 131, + 220, + 141 + ], + "lines": [ + { + "bbox": [ + 168, + 131, + 220, + 141 + ], + "spans": [ + { + "bbox": [ + 168, + 131, + 220, + 141 + ], + "type": "text", + "content": "(c) (b)+optimize" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 222, + 131, + 286, + 141 + ], + "lines": [ + { + "bbox": [ + 222, + 131, + 286, + 141 + ], + "spans": [ + { + "bbox": [ + 222, + 131, + 286, + 141 + ], + "type": "text", + "content": "(d) domain transfer" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 226, + 72, + 284, + 129 + ], + "blocks": [ + { + "bbox": [ + 226, + 72, + 284, + 129 + ], + "lines": [ + { + "bbox": [ + 226, + 72, + 284, + 129 + ], + "spans": [ + { + "bbox": [ + 226, + 72, + 284, + 129 + ], + "type": "image", + "image_path": "3a57a0d2749731b37067e1c9326cef7d50b8cadad51d6ca018fee25782a508fc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 145, + 239, + 158 + ], + "lines": [ + { + "bbox": [ + 95, + 145, + 239, + 158 + ], + "spans": [ + { + "bbox": [ + 95, + 145, + 239, + 158 + ], + "type": "text", + "content": "Figure 6: StyleGANEX inversion." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 167, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 287, + 205 + ], + "type": "text", + "content": "II optimizes " + }, + { + "bbox": [ + 46, + 167, + 287, + 205 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 167, + 287, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 167, + 287, + 205 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 167, + 287, + 205 + ], + "type": "text", + "content": " to further reduce the reconstruction error. The training of " + }, + { + "bbox": [ + 46, + 167, + 287, + 205 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 167, + 287, + 205 + ], + "type": "text", + "content": " follows pSp with reconstruction losses and a regularization loss [23]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 96, + 213, + 287, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 213, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 96, + 213, + 287, + 227 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} (\\hat {x}, x) + \\lambda_ {1} \\mathcal {L} _ {\\text {r e g}} \\left(E _ {W} (\\tilde {x})\\right), \\tag {3}", + "image_path": "3ebeac738f929577e7c347f0a2980c3c84773111afd92009bd6fe9dc5a87670f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\hat{x} = G(E(x, \\tilde{x}, 0))" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": " is the cropped aligned face region of " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": ". We empirically find that using " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": " predicts more accurate " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": " since StyleGAN is originally trained on cropped aligned faces. " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": " encourages the predicted " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": " closer to the average latent code to improve the image quality. " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": " measures the distance between the reconstructed " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": " and the target " + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 234, + 287, + 331 + ], + "type": "text", + "content": " in terms of pixel similarity, perceptual similarity, and identity preservation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 339, + 285, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 339, + 285, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 339, + 285, + 353 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e c}} (\\hat {x}, x) = \\lambda_ {2} \\mathcal {L} _ {2} (\\hat {x}, x) + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {L P I P S}} (\\hat {x}, x) + \\lambda_ {4} \\mathcal {L} _ {\\mathrm {I D}} (\\hat {x}, x).", + "image_path": "acebb9c33da8a5149afc6e73d489ab063ca060947ce55252c7ab489e9e9c86ee.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 361, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 397 + ], + "type": "text", + "content": "As shown in Fig. 6(b), " + }, + { + "bbox": [ + 46, + 361, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 46, + 361, + 287, + 397 + ], + "type": "text", + "content": " largely approximates " + }, + { + "bbox": [ + 46, + 361, + 287, + 397 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 361, + 287, + 397 + ], + "type": "text", + "content": ". But the background details and clothings are still hard to reconstruct. Therefore, we further optimize " + }, + { + "bbox": [ + 46, + 361, + 287, + 397 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 361, + 287, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 287, + 397 + ], + "type": "inline_equation", + "content": "w^{+}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 87, + 404, + 287, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 404, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 87, + 404, + 287, + 426 + ], + "type": "interline_equation", + "content": "\\hat {f}, \\hat {w} ^ {+} = \\underset {f, w ^ {+}} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\mathrm {L P I P S}} (G (f, w ^ {+}), x), \\tag {4}", + "image_path": "9e5120a50304c134928ccf0f8702766d87a17a782d6361ee21983e36d272c5fb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "text", + "content": " are initialized by " + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "inline_equation", + "content": "E(x,\\tilde{x},0)" + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "text", + "content": ". The optimized " + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "inline_equation", + "content": "G(\\hat{f},\\hat{w}^{+})" + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "text", + "content": " in Fig. 6(c) well reconstructs " + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 435, + 287, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": "After inversion, we can perform flexible editing over " + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": " as in StyleGAN. Figure 1(a) shows two examples: we can exchange the last 11 elements of " + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\hat{w}^{+}" + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": " with random samples, to mix the color and texture styles; we can add InterFaceGAN editing vectors [26] to " + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\hat{w}^{+}" + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": " to make a young face. Moreover, as shown in Fig. 6(d), if we load " + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": " a pre-trained StyleGAN-NADA Disney Princess model [9] (let " + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "G'" + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": " denote the new " + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": "), we can obtain " + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "G'(\\hat{f}, \\hat{w}^{+})" + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": ", a cartoon version of " + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 563, + 222, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 222, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 222, + 575 + ], + "type": "text", + "content": "4.3. StyleGANEX-Based Translation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": "The encoder and StyleGANEX form an end-to-end image-to-image translation framework in Fig. 5. Depending on the type of paired training data, it can be trained to efficiently realize different face manipulation tasks. As with pSp [23], we will fix StyleGANEX generator and only train the encoder on the given task." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": "Face super-resolution. Given low-resolution and high-resolution training image pairs " + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": ", we can train " + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": " to recover " + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": " to learn face super-resolution with the loss" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 103, + 701, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 103, + 701, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} (\\hat {y}, y) + \\lambda_ {5} \\mathcal {L} _ {\\text {a d v}} (\\hat {y}, y), \\tag {5}", + "image_path": "fbf2aa9a63b5fd46ed209f80fbfa78bb85602271de545d46c5d9de96d7a993ff.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\hat{y} = G(E(x_{\\uparrow},\\tilde{x}_{\\uparrow},7))" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the upsample operation to make " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " match the input resolution of " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the cropped aligned face region of " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " . We add an adversarial loss " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " to improve the realism of the generated image." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": "Sketch/mask-to-face translation. Given a real face " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " as target and its sketch or parsing mask " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " as source, we can train " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " to translate " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " with Eq. (3) as objectives. In this task, we add a trainable light-weight translation network " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " to map " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " to an intermediate domain where " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " can more easily extract features. For the style condition, " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": "'s first 7 layers use the latent code extracted from " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " to provide structural styles, while its last 11 layers use the latent code from " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " to provide color and texture styles to simplify reconstruction. Therefore, " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "\\hat{y} = G(E_F(T(x),\\ell),E_W^{1:7}(T(x))\\oplus E_W^{8:18}(\\tilde{y}))" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " is concatenation operation, and the superscript of " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "E_{W}" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " means taking the " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "1\\sim 7" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "8\\sim 18" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " elements of " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "\\ell = 1" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " for sketch inputs and " + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "inline_equation", + "content": "\\ell = 3" + }, + { + "bbox": [ + 304, + 121, + 546, + 275 + ], + "type": "text", + "content": " for mask inputs." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "text", + "content": "Video face editing. Given paired original face, edited face, and its editing vector " + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "inline_equation", + "content": "(x,y,v)" + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "text", + "content": " (we can simply generate " + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "inline_equation", + "content": "x = G_0(w^+)" + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "inline_equation", + "content": "y = G_0(w^+ +v)" + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "text", + "content": " from random latent code " + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "text", + "content": " with StyleGAN " + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "inline_equation", + "content": "G_{0})" + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "text", + "content": " , we train " + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 275, + 545, + 323 + ], + "type": "text", + "content": " for face editing with" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 335, + 330, + 545, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 330, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 335, + 330, + 545, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} (\\hat {y}, y) + \\lambda_ {5} \\mathcal {L} _ {\\text {a d v}} (\\hat {y}, y) + \\lambda_ {6} \\mathcal {L} _ {\\text {t m p}} (\\hat {y}), \\tag {6}", + "image_path": "f3f40a774ed23ae65d8192f5cc8879d0ac14b91a9b659ef095a5e1577a00313b.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 348, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 348, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\hat{y} = G(E_F(x,13),E_W(\\tilde{x}) + v)" + }, + { + "bbox": [ + 304, + 348, + 545, + 373 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 348, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tmp}}" + }, + { + "bbox": [ + 304, + 348, + 545, + 373 + ], + "type": "text", + "content": " is the flicker suppression loss [37] to improve temporal consistency." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": "Video toonification. For video face toonification, we have paired original face and toonified face " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": ". They can be generated as " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "x = G_0(w^+)" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "y = G_0'(w^+)" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " from random latent code " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "w^+" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " following Toonify [22], where " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "G_0'" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " is the StyleGAN fine-tuned on cartoon images. Let " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "G'" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " denote StyleGANEX loaded with " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "G_0'" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": ", then we train " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": " using the objectives of Eq. (6) with " + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\hat{y} = G'(E(x,\\tilde{x},13))" + }, + { + "bbox": [ + 304, + 373, + 545, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 457, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 577 + ], + "type": "text", + "content": "Note that compared to face editing based on StyleGANEX inversion in Sec. 4.2, the solution in this section does not require time-consuming latent optimization, and temporal consistency is enforced by the flicker suppression loss. Therefore, this solution is more suitable for efficient and coherent video face manipulation, which is a unique feature of the proposed framework. On the other hand, the solution based on StyleGANEX inversion is more flexible. There is no need to train new " + }, + { + "bbox": [ + 304, + 457, + 545, + 577 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 457, + 545, + 577 + ], + "type": "text", + "content": " for every editing vector " + }, + { + "bbox": [ + 304, + 457, + 545, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 457, + 545, + 577 + ], + "type": "text", + "content": " or fine-funed StyleGAN " + }, + { + "bbox": [ + 304, + 457, + 545, + 577 + ], + "type": "inline_equation", + "content": "G_0^{\\prime}" + }, + { + "bbox": [ + 304, + 457, + 545, + 577 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 586, + 432, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 432, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 432, + 599 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Implementation details. We follow pSp [23] to set " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 1" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{3} = 0.8" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " for all tasks, " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{4} = 0.1" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " for inversion task and 0 for other tasks. We set " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " to 0.0001, 0.005 and 0 for inversion, sketch/mask-to-face and other tasks, respectively. We set " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{5} = 0.1" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{6} = 30" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " empirically. The translation network " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " consists of two downsampling convolutional layers, two ResBlocks [11] and two upsampling convolutional layers, with small channel number 16. All experiments are performed using a single NVIDIA Tesla V100 GPU." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21004" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 547, + 262 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 547, + 262 + ], + "type": "image", + "image_path": "f83d168f36c42383e6e3745da39e5f7464610ef7b2e9158d33fd2dd6a8d3bd84.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 264, + 437, + 277 + ], + "lines": [ + { + "bbox": [ + 154, + 264, + 437, + 277 + ], + "spans": [ + { + "bbox": [ + 154, + 264, + 437, + 277 + ], + "type": "text", + "content": "Figure 7: StyleGANEX inversion and facial attribute/style editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 285, + 289, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 289, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 289, + 395 + ], + "type": "text", + "content": "Datasets. We process FFHQ [17] to obtain 70,000 aligned training images of " + }, + { + "bbox": [ + 46, + 285, + 289, + 395 + ], + "type": "inline_equation", + "content": "1280 \\times 1280" + }, + { + "bbox": [ + 46, + 285, + 289, + 395 + ], + "type": "text", + "content": " resolution for all tasks except two video-related tasks that use StyleGAN generated data. We use BiSeNet [39] to extract parsing masks and follow pSp [23] to extract sketches from face images. We augment all training data with random geometric transformations [15] like scaling, translation and rotation to make faces unaligned. We use images and videos from FaceForensics++ [25], Unsplash and Pexels as our testing dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 406, + 159, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 159, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 159, + 419 + ], + "type": "text", + "content": "5.1. Face Manipulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 715 + ], + "type": "text", + "content": "Face editing. Figure 7 provides an overview of the performance of face inversion and attribute editing on StyleGANEX. We apply inversion to normal FoV face photos/paintings and use various editing vectors from InterFaceGAN [26] and LowRankGAN [40], and the pre-trained StyleGAN-NADA Ukiyo-e model [9], to edit the facial attributes or styles. As shown, these StyleGAN editing techniques work well on StyleGANEX. We also compare with pSp [23], HyperStyle [5] and BDInvert [14] in Fig. 8. Since these baselines are designed for cropped faces, we paste and blend their edited results back into the original image. For a fair comparison, we apply the same optimization method used in our approach to pSp for precise inversion (BDInvert is an optimization-based method and HyperStyle already uses extra hyper networks to simulate optimization). For editing that alters structures or colors, even precise inversion and blending cannot eliminate the obvious discontinuity along the seams, as indicated by the red arrows. In contrast, our approach processes the entire image as a whole and avoids such issues. Remarkably, our method successfully turns the whole hair into black in Fig.8(b), transfers the exemplar blond hairstyle onto the target face in Fig.8(c), and renders the full background with the StyleGAN-NADA Edvard Munch style in Fig.8(d)." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 285, + 545, + 457 + ], + "blocks": [ + { + "bbox": [ + 308, + 285, + 545, + 457 + ], + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 457 + ], + "type": "image", + "image_path": "6669ec1c77690680d482c38522b323886d380ffc459d451c8af3f2f1ee13ec76.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 460, + 506, + 472 + ], + "lines": [ + { + "bbox": [ + 345, + 460, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 345, + 460, + 506, + 472 + ], + "type": "text", + "content": "Figure 8: Comparison on face editing." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 476, + 544, + 602 + ], + "blocks": [ + { + "bbox": [ + 308, + 476, + 544, + 602 + ], + "lines": [ + { + "bbox": [ + 308, + 476, + 544, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 476, + 544, + 602 + ], + "type": "image", + "image_path": "0f730d692533d28b4b70deb9726abaf36ca981c41e1b3d1bfb3a2393acad16bc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 605, + 517, + 618 + ], + "lines": [ + { + "bbox": [ + 334, + 605, + 517, + 618 + ], + "spans": [ + { + "bbox": [ + 334, + 605, + 517, + 618 + ], + "type": "text", + "content": "Figure 9: Comparison on super-resolution." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Face super resolution. We show our " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " super-resolution results in Fig. 9(d), where both the face and non-face regions are reasonably restored. We further follow pSp to train a single mode on multiple rescaling factors (" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "4 \\sim 64" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ") with " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\ell = 3" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " to make a fair comparison. In pSp's results, the non-face region is super-resolved by Real-ESRGAN [35]. As in Fig. 9(b)(c), our method surpasses pSp in precise de" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21005" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 95, + 192 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 95, + 192 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 95, + 192 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 95, + 192 + ], + "type": "image", + "image_path": "1616fe9e97e02c103eccd6dfe08e80a19631492f752484d6b753c32042a058ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 193, + 87, + 202 + ], + "lines": [ + { + "bbox": [ + 59, + 193, + 87, + 202 + ], + "spans": [ + { + "bbox": [ + 59, + 193, + 87, + 202 + ], + "type": "text", + "content": "(a) input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 96, + 70, + 144, + 192 + ], + "blocks": [ + { + "bbox": [ + 96, + 70, + 144, + 192 + ], + "lines": [ + { + "bbox": [ + 96, + 70, + 144, + 192 + ], + "spans": [ + { + "bbox": [ + 96, + 70, + 144, + 192 + ], + "type": "image", + "image_path": "e456360602ea4dcccad86764a72dbcfec5d1a006fcf0803d3ca526f3bb78dd78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 193, + 143, + 202 + ], + "lines": [ + { + "bbox": [ + 97, + 193, + 143, + 202 + ], + "spans": [ + { + "bbox": [ + 97, + 193, + 143, + 202 + ], + "type": "text", + "content": "(b) pix2pixHD" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 144, + 70, + 190, + 192 + ], + "blocks": [ + { + "bbox": [ + 144, + 70, + 190, + 192 + ], + "lines": [ + { + "bbox": [ + 144, + 70, + 190, + 192 + ], + "spans": [ + { + "bbox": [ + 144, + 70, + 190, + 192 + ], + "type": "image", + "image_path": "cb6c650148c34e52000471dce32ac06b8c14f093e85aec0170f7b884cf600fb0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 194, + 179, + 202 + ], + "lines": [ + { + "bbox": [ + 153, + 194, + 179, + 202 + ], + "spans": [ + { + "bbox": [ + 153, + 194, + 179, + 202 + ], + "type": "text", + "content": "(c)TSIT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 192, + 70, + 237, + 192 + ], + "blocks": [ + { + "bbox": [ + 192, + 70, + 237, + 192 + ], + "lines": [ + { + "bbox": [ + 192, + 70, + 237, + 192 + ], + "spans": [ + { + "bbox": [ + 192, + 70, + 237, + 192 + ], + "type": "image", + "image_path": "0f9dd9dc017622ba2d31a410ae7df0dc138f3601f5b77218069c564d1a673caa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 194, + 225, + 202 + ], + "lines": [ + { + "bbox": [ + 202, + 194, + 225, + 202 + ], + "spans": [ + { + "bbox": [ + 202, + 194, + 225, + 202 + ], + "type": "text", + "content": "(d) pSp" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 238, + 71, + 284, + 192 + ], + "blocks": [ + { + "bbox": [ + 238, + 71, + 284, + 192 + ], + "lines": [ + { + "bbox": [ + 238, + 71, + 284, + 192 + ], + "spans": [ + { + "bbox": [ + 238, + 71, + 284, + 192 + ], + "type": "image", + "image_path": "829cfa69c6100131be56f5c7cd2d0bde87c3dec524c8c85dc83dc800fdafb168.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 194, + 272, + 202 + ], + "lines": [ + { + "bbox": [ + 247, + 194, + 272, + 202 + ], + "spans": [ + { + "bbox": [ + 247, + 194, + 272, + 202 + ], + "type": "text", + "content": "(e) ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 224, + 96, + 355 + ], + "blocks": [ + { + "bbox": [ + 54, + 208, + 279, + 220 + ], + "lines": [ + { + "bbox": [ + 54, + 208, + 279, + 220 + ], + "spans": [ + { + "bbox": [ + 54, + 208, + 279, + 220 + ], + "type": "text", + "content": "Figure 10: Comparison on sketch-to-face translation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 224, + 96, + 355 + ], + "lines": [ + { + "bbox": [ + 50, + 224, + 96, + 355 + ], + "spans": [ + { + "bbox": [ + 50, + 224, + 96, + 355 + ], + "type": "image", + "image_path": "8e895e6ba0d0312dd2829448cfeefb3879cbb9a316cc827ee5708dcea8cdc016.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 356, + 86, + 365 + ], + "lines": [ + { + "bbox": [ + 58, + 356, + 86, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 356, + 86, + 365 + ], + "type": "text", + "content": "(a) input" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 369, + 277, + 381 + ], + "lines": [ + { + "bbox": [ + 56, + 369, + 277, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 277, + 381 + ], + "type": "text", + "content": "Figure 11: Comparison on mask-to-face translation." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 96, + 224, + 143, + 355 + ], + "blocks": [ + { + "bbox": [ + 96, + 224, + 143, + 355 + ], + "lines": [ + { + "bbox": [ + 96, + 224, + 143, + 355 + ], + "spans": [ + { + "bbox": [ + 96, + 224, + 143, + 355 + ], + "type": "image", + "image_path": "e78721fe8a8cc41b5f9d2744129cd9684a7a2bc145d7c33222cbb365d3335ce1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 356, + 143, + 365 + ], + "lines": [ + { + "bbox": [ + 97, + 356, + 143, + 365 + ], + "spans": [ + { + "bbox": [ + 97, + 356, + 143, + 365 + ], + "type": "text", + "content": "(b) pix2pixHD" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 144, + 224, + 190, + 355 + ], + "blocks": [ + { + "bbox": [ + 144, + 224, + 190, + 355 + ], + "lines": [ + { + "bbox": [ + 144, + 224, + 190, + 355 + ], + "spans": [ + { + "bbox": [ + 144, + 224, + 190, + 355 + ], + "type": "image", + "image_path": "c7bfc16a583e8c849ad292d1638e14283c18f49c3934f9915f2d6457c541594b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 356, + 180, + 365 + ], + "lines": [ + { + "bbox": [ + 154, + 356, + 180, + 365 + ], + "spans": [ + { + "bbox": [ + 154, + 356, + 180, + 365 + ], + "type": "text", + "content": "(c)TSIT" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 191, + 224, + 237, + 355 + ], + "blocks": [ + { + "bbox": [ + 191, + 224, + 237, + 355 + ], + "lines": [ + { + "bbox": [ + 191, + 224, + 237, + 355 + ], + "spans": [ + { + "bbox": [ + 191, + 224, + 237, + 355 + ], + "type": "image", + "image_path": "fa83c66ee39dcb666865ad262a9426322752ac73687eff840e099091123e34ef.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 356, + 223, + 365 + ], + "lines": [ + { + "bbox": [ + 200, + 356, + 223, + 365 + ], + "spans": [ + { + "bbox": [ + 200, + 356, + 223, + 365 + ], + "type": "text", + "content": "(d) pSp" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 238, + 224, + 284, + 355 + ], + "blocks": [ + { + "bbox": [ + 238, + 224, + 284, + 355 + ], + "lines": [ + { + "bbox": [ + 238, + 224, + 284, + 355 + ], + "spans": [ + { + "bbox": [ + 238, + 224, + 284, + 355 + ], + "type": "image", + "image_path": "d311c8b28bd2ba176ac7c5091a503effecf0043cba77c4e2ce14122ccd79c900.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 356, + 273, + 365 + ], + "lines": [ + { + "bbox": [ + 247, + 356, + 273, + 365 + ], + "spans": [ + { + "bbox": [ + 247, + 356, + 273, + 365 + ], + "type": "text", + "content": "(e) ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 391, + 286, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 286, + 415 + ], + "type": "text", + "content": "tail restoration (e.g., glasses) and uniform super-resolution without discontinuity between face and non-face regions." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 415, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 286, + 604 + ], + "type": "text", + "content": "Sketch/mask to face translation. We compare our method with image-to-image translation models pix2pixHD [34] and TSIT [12], and StyleGAN-based pSp in Figs. 10-11. Pix2pixHD's results have many artifacts and monotonous colors. TSIT requires the inputs' side lengths to be divisible by 128. We find padding the input leads to failed translation. Therefore, we show its results on centrally cropped inputs, which are blurry. PSp generates realistic results, which are however less similar to the input sketch/mask. By comparison, our method can translate whole images and achieve realism and structural consistency to the inputs. For quantitative evaluation, we conduct a user study, where 30 subjects are invited to select what they consider to be the best results from the four methods. Each task uses eight results for evaluation. Table 1 summarizes the preference scores, where our method receives the best score." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": "Video face editing. We compare with pSp, HyperStyle, StyleHEAT [38] and STIT [31]. StyleHEAT uses features of a fixed " + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": " resolution for unaligned but still cropped " + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": " face reenactment. Specifically, it generates videos based on warping the features of the first frame, which however limits its inversion accuracy. STIT extends PTI [24] for full video processing by stitching. STIT cannot well preserve the complex hair details (Fig. 12, yellow box). As with image face editing, all four baselines are limited" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 308, + 83, + 415, + 332 + ], + "blocks": [ + { + "bbox": [ + 357, + 74, + 392, + 83 + ], + "lines": [ + { + "bbox": [ + 357, + 74, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 357, + 74, + 392, + 83 + ], + "type": "text", + "content": "black hair" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 83, + 415, + 332 + ], + "lines": [ + { + "bbox": [ + 308, + 83, + 415, + 332 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 415, + 332 + ], + "type": "image", + "image_path": "ddbca3467ddba59a9c594f65c49979dca63a9010faf87bc7c1cfb6e6c41cf7ea.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 418, + 83, + 527, + 332 + ], + "blocks": [ + { + "bbox": [ + 457, + 75, + 481, + 83 + ], + "lines": [ + { + "bbox": [ + 457, + 75, + 481, + 83 + ], + "spans": [ + { + "bbox": [ + 457, + 75, + 481, + 83 + ], + "type": "text", + "content": "young" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 418, + 83, + 527, + 332 + ], + "lines": [ + { + "bbox": [ + 418, + 83, + 527, + 332 + ], + "spans": [ + { + "bbox": [ + 418, + 83, + 527, + 332 + ], + "type": "image", + "image_path": "49437caeddab0a76672e1813ff87ab0fc2ff0ae12244e4d92a095a97cf1aa20e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 310, + 342, + 542, + 371 + ], + "blocks": [ + { + "bbox": [ + 406, + 334, + 457, + 343 + ], + "lines": [ + { + "bbox": [ + 406, + 334, + 457, + 343 + ], + "spans": [ + { + "bbox": [ + 406, + 334, + 457, + 343 + ], + "type": "text", + "content": "editing control" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 342, + 542, + 371 + ], + "lines": [ + { + "bbox": [ + 310, + 342, + 542, + 371 + ], + "spans": [ + { + "bbox": [ + 310, + 342, + 542, + 371 + ], + "type": "image", + "image_path": "1222658c3a436b757f348c1e635bda5771e68318ec976e81b3ed191c74d4d927.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 376, + 520, + 389 + ], + "lines": [ + { + "bbox": [ + 329, + 376, + 520, + 389 + ], + "spans": [ + { + "bbox": [ + 329, + 376, + 520, + 389 + ], + "type": "text", + "content": "Figure 12: Comparison on video face editing." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 310, + 396, + 543, + 553 + ], + "blocks": [ + { + "bbox": [ + 310, + 396, + 543, + 553 + ], + "lines": [ + { + "bbox": [ + 310, + 396, + 543, + 553 + ], + "spans": [ + { + "bbox": [ + 310, + 396, + 543, + 553 + ], + "type": "image", + "image_path": "693498dde616d7c6f6af9058d117bb1c858d2e5029d5b866a09b2c59379f8732.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 559, + 533, + 571 + ], + "lines": [ + { + "bbox": [ + 317, + 559, + 533, + 571 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 533, + 571 + ], + "type": "text", + "content": "Figure 13: Comparison on video face toonification." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": "to editing cropped regions, leading to discontinuity along the stitching seams. By comparison, our method uses the first-layer feature and skipped mid-layer features to provide spatial information, which achieves more coherent results. Moreover, we can randomly scale the editing vector " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": " (by multiplying a scale factor) instead of using a fixed " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": " during training. Then during testing, our method can flexibly adjust the editing degree by scaling " + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 580, + 545, + 688 + ], + "type": "text", + "content": " for users to select as in Fig. 12(e)." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Video face toonification. Compared with VToonify-T [37], our method preserves more details of the non-face region" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21006" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 87, + 286, + 158 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 282, + 82 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 282, + 82 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 282, + 82 + ], + "type": "text", + "content": "Table 1: User preference scores. Best scores are in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 87, + 286, + 158 + ], + "lines": [ + { + "bbox": [ + 50, + 87, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 50, + 87, + 286, + 158 + ], + "type": "table", + "html": "
Taskpix2pixHD [34]TSIT [12]pSp [23]ours
sketch-to-face0.4000.1210.0290.450
mask-to-face0.1080.0750.1210.696
TaskVToonify-T [37]ours
video toonify0.0830.917
", + "image_path": "8b144bf1f3c778cd7693491167595294d5b7555ec55d9122e006c72bf168c7cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 161, + 96, + 208 + ], + "blocks": [ + { + "bbox": [ + 50, + 161, + 96, + 208 + ], + "lines": [ + { + "bbox": [ + 50, + 161, + 96, + 208 + ], + "spans": [ + { + "bbox": [ + 50, + 161, + 96, + 208 + ], + "type": "image", + "image_path": "210fe0e65a5141b0967a19f8385e1d21ee85531e6beb7fadf6ef05e958367eb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 209, + 85, + 217 + ], + "lines": [ + { + "bbox": [ + 57, + 209, + 85, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 85, + 217 + ], + "type": "text", + "content": "(a) input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 97, + 161, + 142, + 208 + ], + "blocks": [ + { + "bbox": [ + 97, + 161, + 142, + 208 + ], + "lines": [ + { + "bbox": [ + 97, + 161, + 142, + 208 + ], + "spans": [ + { + "bbox": [ + 97, + 161, + 142, + 208 + ], + "type": "image", + "image_path": "6efefed11ec12bea56cc5fbd6ad726f4ac81da1b0aa68c32059a4cbf56f52865.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 209, + 138, + 218 + ], + "lines": [ + { + "bbox": [ + 97, + 209, + 138, + 218 + ], + "spans": [ + { + "bbox": [ + 97, + 209, + 138, + 218 + ], + "type": "text", + "content": "(b) two steps" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 143, + 161, + 190, + 208 + ], + "blocks": [ + { + "bbox": [ + 143, + 161, + 190, + 208 + ], + "lines": [ + { + "bbox": [ + 143, + 161, + 190, + 208 + ], + "spans": [ + { + "bbox": [ + 143, + 161, + 190, + 208 + ], + "type": "image", + "image_path": "fae8f11de126cb43c57cb92dc05f35ecfe08c8d30c32da49821e4cd9ecf45c35.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 209, + 191, + 218 + ], + "lines": [ + { + "bbox": [ + 142, + 209, + 191, + 218 + ], + "spans": [ + { + "bbox": [ + 142, + 209, + 191, + 218 + ], + "type": "text", + "content": "(c) transferred (b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 191, + 161, + 238, + 208 + ], + "blocks": [ + { + "bbox": [ + 191, + 161, + 238, + 208 + ], + "lines": [ + { + "bbox": [ + 191, + 161, + 238, + 208 + ], + "spans": [ + { + "bbox": [ + 191, + 161, + 238, + 208 + ], + "type": "image", + "image_path": "4cc38d86679334248b0dec497d9f6e71f236f8fcc889eb4aad7aa8b38a444e49.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 209, + 238, + 218 + ], + "lines": [ + { + "bbox": [ + 191, + 209, + 238, + 218 + ], + "spans": [ + { + "bbox": [ + 191, + 209, + 238, + 218 + ], + "type": "text", + "content": "(d) only opt (e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 239, + 161, + 285, + 208 + ], + "blocks": [ + { + "bbox": [ + 239, + 161, + 285, + 208 + ], + "lines": [ + { + "bbox": [ + 239, + 161, + 285, + 208 + ], + "spans": [ + { + "bbox": [ + 239, + 161, + 285, + 208 + ], + "type": "image", + "image_path": "ee108ec916865ae682756eb3b9fece01c1674b6679fb40767cec0bd96bae4e29.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 209, + 286, + 218 + ], + "lines": [ + { + "bbox": [ + 239, + 209, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 239, + 209, + 286, + 218 + ], + "type": "text", + "content": "e) transferred (d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 50, + 239, + 108, + 297 + ], + "blocks": [ + { + "bbox": [ + 52, + 222, + 281, + 234 + ], + "lines": [ + { + "bbox": [ + 52, + 222, + 281, + 234 + ], + "spans": [ + { + "bbox": [ + 52, + 222, + 281, + 234 + ], + "type": "text", + "content": "Figure 14: Effect of encoder in StyleGANEX inversion." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 239, + 108, + 297 + ], + "lines": [ + { + "bbox": [ + 50, + 239, + 108, + 297 + ], + "spans": [ + { + "bbox": [ + 50, + 239, + 108, + 297 + ], + "type": "image", + "image_path": "cc65cb43ba151b11045ed81f8781018dae063c9c77f382c540281d74af2c7f71.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 298, + 94, + 308 + ], + "lines": [ + { + "bbox": [ + 59, + 298, + 94, + 308 + ], + "spans": [ + { + "bbox": [ + 59, + 298, + 94, + 308 + ], + "type": "text", + "content": "(a) input " + }, + { + "bbox": [ + 59, + 298, + 94, + 308 + ], + "type": "inline_equation", + "content": "x_{\\uparrow}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 314, + 286, + 326 + ], + "lines": [ + { + "bbox": [ + 47, + 314, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 286, + 326 + ], + "type": "text", + "content": "Figure 15: Input choice to provide valid style information." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 110, + 239, + 167, + 297 + ], + "blocks": [ + { + "bbox": [ + 110, + 239, + 167, + 297 + ], + "lines": [ + { + "bbox": [ + 110, + 239, + 167, + 297 + ], + "spans": [ + { + "bbox": [ + 110, + 239, + 167, + 297 + ], + "type": "image", + "image_path": "f68a7ad0ac727ff24559025487b3b406698d46ff9c6e4801ad5ee6d42320b89b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 298, + 165, + 308 + ], + "lines": [ + { + "bbox": [ + 111, + 298, + 165, + 308 + ], + "spans": [ + { + "bbox": [ + 111, + 298, + 165, + 308 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 111, + 298, + 165, + 308 + ], + "type": "inline_equation", + "content": "G(E(x_{1},\\tilde{x}_{1},7))" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 168, + 239, + 227, + 297 + ], + "blocks": [ + { + "bbox": [ + 168, + 239, + 227, + 297 + ], + "lines": [ + { + "bbox": [ + 168, + 239, + 227, + 297 + ], + "spans": [ + { + "bbox": [ + 168, + 239, + 227, + 297 + ], + "type": "image", + "image_path": "ac8d38edda084a257454478ba5d355b2047f7639b85fe4a4012d53ed57020d9c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 298, + 224, + 308 + ], + "lines": [ + { + "bbox": [ + 169, + 298, + 224, + 308 + ], + "spans": [ + { + "bbox": [ + 169, + 298, + 224, + 308 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 169, + 298, + 224, + 308 + ], + "type": "inline_equation", + "content": "G(E(x_{\\uparrow},x_{\\uparrow},7))" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 228, + 239, + 286, + 297 + ], + "blocks": [ + { + "bbox": [ + 228, + 239, + 286, + 297 + ], + "lines": [ + { + "bbox": [ + 228, + 239, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 228, + 239, + 286, + 297 + ], + "type": "image", + "image_path": "b721664c3714187b6575c4df0b76b5f47b490481737bb9e49c947e87073dc619.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 298, + 285, + 308 + ], + "lines": [ + { + "bbox": [ + 229, + 298, + 285, + 308 + ], + "spans": [ + { + "bbox": [ + 229, + 298, + 285, + 308 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 229, + 298, + 285, + 308 + ], + "type": "inline_equation", + "content": "G(E(x_{\\uparrow},\\ddot{x}_{\\uparrow},7))" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 336, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 431 + ], + "type": "text", + "content": "and generates shaper faces in Fig. 13. The reason is that VToonify-T uses a fixed latent code extractor while our method trains a joint latent code and feature extractor, thus our method is more powerful for reconstructing the details. Moreover, our method retains StyleGAN's shallow layers, which helps provide key facial features to make the stylized face more vivid. Table 1 shows a quantitative comparison on ten results, and our method obtains the best user score." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 442, + 140, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 140, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 140, + 455 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": "The effect of Step II of our two-step inversion is verified in Fig. 6. We further study the effect of Step I in Fig. 14. With Step I providing a good prediction of " + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": ", Step II only needs 500-iteration optimization for precise reconstruction (Fig. 14(b)) and valid domain transfer to Disney Princess (Fig. 14(c)). However, if we directly optimize a mean " + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": " and a random " + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": ", the result is poor even with 2,000 iterations (indicated by a red arrow in Fig. 14(d)) and the optimized " + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 461, + 287, + 581 + ], + "type": "text", + "content": " overfit the input, which is not suitable for editing like domain transfer (Fig. 14(e))." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "In Fig. 15, we study the input choice to extract " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": ". The cropped aligned faces are the default choice. If we instead use the whole image to extract " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "w^{+}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": ", the background will affect the facial style, leading to poor restoration in Fig. 15(c). We further find reasonable results (Fig. 15(d)) can still be obtained by cropping the input to decrease the background proportion. Note that the face is not aligned in the cropped image " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\ddot{x}_{\\uparrow})" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": ", which is useful for applications like super-resolution where extremely low-resolution faces are hard to detect and align. Users can simply manually crop the face region to provide valid style information." + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 366, + 123 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 366, + 123 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 366, + 123 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 366, + 123 + ], + "type": "image", + "image_path": "5f092e80676832d6190f65952c4573c65f2b53b258515bf7fabcf531b563c76a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 125, + 351, + 135 + ], + "lines": [ + { + "bbox": [ + 321, + 125, + 351, + 135 + ], + "spans": [ + { + "bbox": [ + 321, + 125, + 351, + 135 + ], + "type": "text", + "content": "(a) input" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 366, + 72, + 425, + 123 + ], + "blocks": [ + { + "bbox": [ + 366, + 72, + 425, + 123 + ], + "lines": [ + { + "bbox": [ + 366, + 72, + 425, + 123 + ], + "spans": [ + { + "bbox": [ + 366, + 72, + 425, + 123 + ], + "type": "image", + "image_path": "cc4b9ea92ef0092cc60c86678a90e29a14846df855fb7216b51f8ac46e8522df.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 125, + 408, + 134 + ], + "lines": [ + { + "bbox": [ + 380, + 125, + 408, + 134 + ], + "spans": [ + { + "bbox": [ + 380, + 125, + 408, + 134 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 380, + 125, + 408, + 134 + ], + "type": "inline_equation", + "content": "\\ell = 0" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 484, + 123 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 484, + 123 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 484, + 123 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 484, + 123 + ], + "type": "image", + "image_path": "d8c35664d3f2b23439b2792e4a39adbdf274cdefd6e5abdd8bc42d384a85ab79.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 125, + 466, + 134 + ], + "lines": [ + { + "bbox": [ + 440, + 125, + 466, + 134 + ], + "spans": [ + { + "bbox": [ + 440, + 125, + 466, + 134 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 440, + 125, + 466, + 134 + ], + "type": "inline_equation", + "content": "\\ell = 3" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 485, + 72, + 544, + 122 + ], + "blocks": [ + { + "bbox": [ + 485, + 72, + 544, + 122 + ], + "lines": [ + { + "bbox": [ + 485, + 72, + 544, + 122 + ], + "spans": [ + { + "bbox": [ + 485, + 72, + 544, + 122 + ], + "type": "image", + "image_path": "9baacb02cca2dda7e70a2d8ccd85083315fba85c16200a4e2e1fcea822f05b3c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 124, + 528, + 134 + ], + "lines": [ + { + "bbox": [ + 501, + 124, + 528, + 134 + ], + "spans": [ + { + "bbox": [ + 501, + 124, + 528, + 134 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 501, + 124, + 528, + 134 + ], + "type": "inline_equation", + "content": "\\ell = 7" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 308, + 159, + 347, + 198 + ], + "blocks": [ + { + "bbox": [ + 347, + 140, + 504, + 152 + ], + "lines": [ + { + "bbox": [ + 347, + 140, + 504, + 152 + ], + "spans": [ + { + "bbox": [ + 347, + 140, + 504, + 152 + ], + "type": "text", + "content": "Figure 16: Effect of skip connections." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 159, + 347, + 198 + ], + "lines": [ + { + "bbox": [ + 308, + 159, + 347, + 198 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 347, + 198 + ], + "type": "image", + "image_path": "cd53acac17a62b70bb19661bc4c53a13f6004012db5bdfaed976842ef321946e.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 200, + 340, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 200, + 340, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 340, + 209 + ], + "type": "text", + "content": "(a) input" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 348, + 159, + 387, + 198 + ], + "blocks": [ + { + "bbox": [ + 348, + 159, + 387, + 198 + ], + "lines": [ + { + "bbox": [ + 348, + 159, + 387, + 198 + ], + "spans": [ + { + "bbox": [ + 348, + 159, + 387, + 198 + ], + "type": "image", + "image_path": "f047f5d509967c9d03c4f6b32f918b2834de9073944e27bf76f5c316feb352c7.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 200, + 380, + 208 + ], + "lines": [ + { + "bbox": [ + 358, + 200, + 380, + 208 + ], + "spans": [ + { + "bbox": [ + 358, + 200, + 380, + 208 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 358, + 200, + 380, + 208 + ], + "type": "inline_equation", + "content": "\\ell = 1" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 389, + 159, + 422, + 198 + ], + "blocks": [ + { + "bbox": [ + 389, + 159, + 422, + 198 + ], + "lines": [ + { + "bbox": [ + 389, + 159, + 422, + 198 + ], + "spans": [ + { + "bbox": [ + 389, + 159, + 422, + 198 + ], + "type": "image", + "image_path": "bd48b4c7194423db2a9d78901a09ed522666921c19a5eb58719cd0f821986719.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 200, + 418, + 208 + ], + "lines": [ + { + "bbox": [ + 395, + 200, + 418, + 208 + ], + "spans": [ + { + "bbox": [ + 395, + 200, + 418, + 208 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 395, + 200, + 418, + 208 + ], + "type": "inline_equation", + "content": "\\ell = 0" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 425, + 159, + 465, + 198 + ], + "blocks": [ + { + "bbox": [ + 425, + 159, + 465, + 198 + ], + "lines": [ + { + "bbox": [ + 425, + 159, + 465, + 198 + ], + "spans": [ + { + "bbox": [ + 425, + 159, + 465, + 198 + ], + "type": "image", + "image_path": "9d13d7f6fa122bdef8a63cbe05bcf9a2656956ec19840216a0a5b78527881002.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 200, + 457, + 208 + ], + "lines": [ + { + "bbox": [ + 430, + 200, + 457, + 208 + ], + "spans": [ + { + "bbox": [ + 430, + 200, + 457, + 208 + ], + "type": "text", + "content": "(d) input" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 468, + 159, + 505, + 198 + ], + "blocks": [ + { + "bbox": [ + 468, + 159, + 505, + 198 + ], + "lines": [ + { + "bbox": [ + 468, + 159, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 468, + 159, + 505, + 198 + ], + "type": "image", + "image_path": "b7b22fee3dbfdc405d85ced28e1a9c0e422fd2a99fb86cd90b46104fd44aadec.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 200, + 496, + 208 + ], + "lines": [ + { + "bbox": [ + 474, + 200, + 496, + 208 + ], + "spans": [ + { + "bbox": [ + 474, + 200, + 496, + 208 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 474, + 200, + 496, + 208 + ], + "type": "inline_equation", + "content": "\\ell = 1" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 506, + 159, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 506, + 159, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 506, + 159, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 506, + 159, + 545, + 198 + ], + "type": "image", + "image_path": "4c1539a8d7a99bae4d56f199870a06f9b58121754f9643206271a6abd7997dc7.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 515, + 200, + 537, + 208 + ], + "lines": [ + { + "bbox": [ + 515, + 200, + 537, + 208 + ], + "spans": [ + { + "bbox": [ + 515, + 200, + 537, + 208 + ], + "type": "text", + "content": "(f) " + }, + { + "bbox": [ + 515, + 200, + 537, + 208 + ], + "type": "inline_equation", + "content": "\\ell = 0" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 309, + 231, + 365, + 317 + ], + "blocks": [ + { + "bbox": [ + 323, + 214, + 528, + 226 + ], + "lines": [ + { + "bbox": [ + 323, + 214, + 528, + 226 + ], + "spans": [ + { + "bbox": [ + 323, + 214, + 528, + 226 + ], + "type": "text", + "content": "Figure 17: Performance on low-quality sketches." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 231, + 365, + 317 + ], + "lines": [ + { + "bbox": [ + 309, + 231, + 365, + 317 + ], + "spans": [ + { + "bbox": [ + 309, + 231, + 365, + 317 + ], + "type": "image", + "image_path": "a39666f4cd49bc3378b174e838d9b1b5fd21e16c893141d32f814e4bf6fb8f2c.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 320, + 348, + 331 + ], + "lines": [ + { + "bbox": [ + 324, + 320, + 348, + 331 + ], + "spans": [ + { + "bbox": [ + 324, + 320, + 348, + 331 + ], + "type": "text", + "content": "input" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 368, + 232, + 425, + 318 + ], + "blocks": [ + { + "bbox": [ + 368, + 232, + 425, + 318 + ], + "lines": [ + { + "bbox": [ + 368, + 232, + 425, + 318 + ], + "spans": [ + { + "bbox": [ + 368, + 232, + 425, + 318 + ], + "type": "image", + "image_path": "bea45d4ca4152e69942dd8b1186ed07fa0af5bc4ff0501df2e9562b11ef1cc4c.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 319, + 416, + 330 + ], + "lines": [ + { + "bbox": [ + 376, + 319, + 416, + 330 + ], + "spans": [ + { + "bbox": [ + 376, + 319, + 416, + 330 + ], + "type": "text", + "content": "inversion" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 427, + 232, + 484, + 318 + ], + "blocks": [ + { + "bbox": [ + 427, + 232, + 484, + 318 + ], + "lines": [ + { + "bbox": [ + 427, + 232, + 484, + 318 + ], + "spans": [ + { + "bbox": [ + 427, + 232, + 484, + 318 + ], + "type": "image", + "image_path": "f3d09525ba003773acf8646011e375b0ca5950b4ab8a157568f70c7830ac61d5.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 319, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 427, + 319, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 427, + 319, + 482, + 331 + ], + "type": "text", + "content": "blocky shape" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 486, + 232, + 544, + 318 + ], + "blocks": [ + { + "bbox": [ + 486, + 232, + 544, + 318 + ], + "lines": [ + { + "bbox": [ + 486, + 232, + 544, + 318 + ], + "spans": [ + { + "bbox": [ + 486, + 232, + 544, + 318 + ], + "type": "image", + "image_path": "427ecd39e97c8dff0b94308ad0ea88dbeb3be7bbff0e0b692b1d29a531471d5c.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 319, + 534, + 331 + ], + "lines": [ + { + "bbox": [ + 493, + 319, + 534, + 331 + ], + "spans": [ + { + "bbox": [ + 493, + 319, + 534, + 331 + ], + "type": "text", + "content": "add grass" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 368, + 333, + 425, + 419 + ], + "blocks": [ + { + "bbox": [ + 368, + 333, + 425, + 419 + ], + "lines": [ + { + "bbox": [ + 368, + 333, + 425, + 419 + ], + "spans": [ + { + "bbox": [ + 368, + 333, + 425, + 419 + ], + "type": "image", + "image_path": "b966d440e3da220da2920b84aa67f6a3c7d7fc2606024cdfb02fd69e912a4afb.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 422, + 411, + 431 + ], + "lines": [ + { + "bbox": [ + 379, + 422, + 411, + 431 + ], + "spans": [ + { + "bbox": [ + 379, + 422, + 411, + 431 + ], + "type": "text", + "content": "redness" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 438, + 492, + 450 + ], + "lines": [ + { + "bbox": [ + 359, + 438, + 492, + 450 + ], + "spans": [ + { + "bbox": [ + 359, + 438, + 492, + 450 + ], + "type": "text", + "content": "Figure 18: Performance on car." + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 427, + 333, + 485, + 419 + ], + "blocks": [ + { + "bbox": [ + 427, + 333, + 485, + 419 + ], + "lines": [ + { + "bbox": [ + 427, + 333, + 485, + 419 + ], + "spans": [ + { + "bbox": [ + 427, + 333, + 485, + 419 + ], + "type": "image", + "image_path": "3c4defc0a8ba0803d5b2f98a4b685dd6a6897496da9f8d0990b91a6c79bf166e.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 422, + 510, + 433 + ], + "lines": [ + { + "bbox": [ + 458, + 422, + 510, + 433 + ], + "spans": [ + { + "bbox": [ + 458, + 422, + 510, + 433 + ], + "type": "text", + "content": "style mixing" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 486, + 333, + 544, + 419 + ], + "blocks": [ + { + "bbox": [ + 486, + 333, + 544, + 419 + ], + "lines": [ + { + "bbox": [ + 486, + 333, + 544, + 419 + ], + "spans": [ + { + "bbox": [ + 486, + 333, + 544, + 419 + ], + "type": "image", + "image_path": "a498d13ebcd693cb8de3eaa147743abc9723ace36bc5cdf625ba70da4509aac4.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "text", + "content": "In Fig. 16, we study the effect of skip connection. Without it " + }, + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "inline_equation", + "content": "(\\ell = 0)" + }, + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "text", + "content": ", the glasses cannot be generated. Skip connection provides mid-layer features to preserve the details of the input. However, too many extra features will alter the feature distribution of StyleGAN, e.g., always generating sunglasses as the input has no segmentation of eyes. Thus, we use " + }, + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\ell = 3" + }, + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "text", + "content": " to balance between input-output consistency and diversity. Inversely, we can use a small " + }, + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "text", + "content": " to enhance the model robustness to low-quality inputs. For example, we can generate more realistic faces with " + }, + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\ell = 0" + }, + { + "bbox": [ + 304, + 468, + 545, + 600 + ], + "type": "text", + "content": " on DeepFace-Drawing low-quality sketches [7] as in Fig. 17." + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 306, + 610, + 469, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 469, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 469, + 621 + ], + "type": "text", + "content": "5.3. Results on Non-Facial Dataset" + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "The proposed refactoring is domain-agnostic, thus it can be applied to StyleGAN pre-trained on other domains like cars. An example with a wide vertical field of view is shown in Fig. 18, where we perform optimization-based StyleGAN inversion, image editing with the editing vectors found by GANSpace [10] and style mixing on the refactored model. The resulting manipulated cars look plausible." + } + ] + } + ], + "index": 64 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21007" + } + ] + } + ], + "index": 65 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 173 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 173 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 173 + ], + "type": "image", + "image_path": "2d1d5cbd21f6323168486be213ea2f578478608e78722b2310251bdbd0e23a5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 175, + 206, + 185 + ], + "lines": [ + { + "bbox": [ + 127, + 175, + 206, + 185 + ], + "spans": [ + { + "bbox": [ + 127, + 175, + 206, + 185 + ], + "type": "text", + "content": "(c) severely rotated face" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 117, + 188, + 216, + 199 + ], + "lines": [ + { + "bbox": [ + 117, + 188, + 216, + 199 + ], + "spans": [ + { + "bbox": [ + 117, + 188, + 216, + 199 + ], + "type": "text", + "content": "Figure 19: Limitations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 210, + 124, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 210, + 124, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 124, + 221 + ], + "type": "text", + "content": "5.4. Limitations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 228, + 290, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 290, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 290, + 504 + ], + "type": "text", + "content": "First, our framework currently relies on an inefficient optimization process for precise reconstruction. While future work can explore more efficient inversion methods (e.g., iterative residue prediction and hyper networks), it is important to note that this paper focuses on overcoming the fixed-crop limitation of StyleGAN, rather than the specific topic of GAN inversion. Second, StyleGANEX is limited by the feature representation of StyleGAN. While it shows great potential in handling normal FoV face images, out-of-distribution features such as complex clothing and human bodies may not be well handled as in Fig. 19(a)(b). As in Fig. 19(c), while our method can handle faces rotated 20 degrees, it still struggles with handling large rotation angles. However, this can be easily resolved by rough alignment of the input image in the middle as our method does not require accurate alignment. Finally, after translation, we can undo the alignment to obtain a plausible result as in right of Fig. 19(c). Third, StyleGANEX, like StyleGAN, focuses on face manipulation and may not support out-of-distribution semantical editing of non-facial regions. Last, StyleGANEX may inherit the model bias of StyleGAN. Applying it to tasks with severe data imbalance might lead to unsatisfactory results on under-represented data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 119, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 119, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 119, + 526 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 665 + ], + "type": "text", + "content": "We have presented an effective approach to refactor StyleGAN to overcome its fixed-crop limitation while retaining its style control abilities. The refactored model, called StyleGANEX, fully inherits the parameters of the pre-trained StyleGAN without retraining, and is thus fully compatible with the generative space of StyleGAN. We further introduced a StyleGANEX encoder to project normal FoV face images to the joint " + }, + { + "bbox": [ + 46, + 534, + 287, + 665 + ], + "type": "inline_equation", + "content": "W^{+} - F" + }, + { + "bbox": [ + 46, + 534, + 287, + 665 + ], + "type": "text", + "content": " space of StyleGANEX for real face inversion and manipulation. Our approach offers an effective solution to the general issue of StyleGAN and extends its capability beyond fixed-resolution data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Acknowledgments. This study is supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). It is al" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "so supported by MOE AcRF Tier 1 (2021-T1-001-088) and NTU NAP grant." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 110, + 365, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 110, + 365, + 122 + ], + "spans": [ + { + "bbox": [ + 307, + 110, + 365, + 122 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 131, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 311, + 131, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 131, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 311, + 131, + 545, + 175 + ], + "type": "text", + "content": "[1] Rameen Abdal, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proc. Int'l Conf. Computer Vision, pages 4432-4441, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 177, + 547, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 177, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 311, + 177, + 547, + 232 + ], + "type": "text", + "content": "[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 234, + 547, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 234, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 311, + 234, + 547, + 277 + ], + "type": "text", + "content": "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinements. In Proc. Int'l Conf. Computer Vision, pages 6711–6720, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 280, + 547, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 280, + 547, + 322 + ], + "spans": [ + { + "bbox": [ + 311, + 280, + 547, + 322 + ], + "type": "text", + "content": "[4] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 326, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 326, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 311, + 326, + 547, + 380 + ], + "type": "text", + "content": "[5] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 18511-18521, 2022. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 311, + 383, + 545, + 426 + ], + "type": "text", + "content": "[6] Kelvin CK Chan, Xintao Wang, Xiangyu Xu, Jinwei Gu, and Chen Change Loy. Glean: Generative latent bank for large-factor image super-resolution. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2021. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 428, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 428, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 311, + 428, + 545, + 472 + ], + "type": "text", + "content": "[7] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. ACM Transactions on Graphics, 39(4):72-1, 2020. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 475, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 475, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 311, + 475, + 545, + 518 + ], + "type": "text", + "content": "[8] Tan M Dinh, Anh Tuan Tran, Rang Nguyen, and Binh-Son Hua. Hyperinverter: Improving stylegan inversion via hypernetwork. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11389-11398, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 521, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 521, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 311, + 521, + 545, + 564 + ], + "type": "text", + "content": "[9] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clipguided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "type": "text", + "content": "[10] Erik Härkönen, Aaron Hertzman, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In Advances in Neural Information Processing Systems, 2020. 1, 2, 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[12] Liming Jiang, Changxu Zhang, Mingyang Huang, Chunxiao Liu, Jianping Shi, and Chen Change Loy. Tset: A simple and versatile framework for image-to-image translation. In Proc. European Conf. Computer Vision, pages 206-222. Springer, 2020. 7, 8" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21008" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[13] Yuming Jiang, Ziqi Huang, Xingang Pan, Chen Change Loy, and Ziwei Liu. Talk-to-edit: Fine-grained facial editing via dialog. In Proc. Int'l Conf. Computer Vision, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 108, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 108, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 108, + 287, + 152 + ], + "type": "text", + "content": "[14] Kyoungkook Kang, Seongtae Kim, and Sunghyun Cho. Gan inversion for out-of-range images with geometric transformations. In Proc. Int'l Conf. Computer Vision, pages 13941-13949, 2021. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 198 + ], + "type": "text", + "content": "[15] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Advances in Neural Information Processing Systems, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 199, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 199, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 287, + 243 + ], + "type": "text", + "content": "[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 245, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 245, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 245, + 287, + 289 + ], + "type": "text", + "content": "[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 291, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 287, + 346 + ], + "type": "text", + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 348, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 348, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 348, + 287, + 392 + ], + "type": "text", + "content": "[19] Feng-Lin Liu, Shu-Yu Chen, Yukun Lai, YueRen Jiang, Hongbo Fu, and Lin Gao. Deepfacevideoediting: Sketch-based deep editing of face videos. ACM Transactions on Graphics, 41(4):167, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "type": "text", + "content": "[20] Gaurav Parmar, Yijun Li, Jingwan Lu, Richard Zhang, JunYan Zhu, and Krishna Kumar Singh. Spatially-adaptive multilayer selection for gan inversion and editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11399-11409, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 450, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 495 + ], + "type": "text", + "content": "[21] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 2085–2094, October 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 497, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 287, + 529 + ], + "type": "text", + "content": "[22] Justin NM Pinkney and Doron Adler. Resolution dependent gan interpolation for controllable image synthesis between domains. arXiv preprint arXiv:2010.05334, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "type": "text", + "content": "[23] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2021. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 588, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 621 + ], + "type": "text", + "content": "[24] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics, 2022. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[25] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. In Proc. Int'l Conf. Computer Vision, 2019. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[26] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 9243-9252, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[27] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[28] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 1532-1540, 2021. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "text", + "content": "[29] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. ACM Transactions on Graphics, 39(6):1-14, 2020. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 258 + ], + "type": "text", + "content": "[30] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics, 40(4):1-14, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[31] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. Stitch it in time: Gan-based facial editing of real videos. In SIGGRAPH Asia, pages 1–9, 2022. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "type": "text", + "content": "[32] Yuri Viazovetskyi, Vladimir Ivashkin, and Evgeny Kashin. Stylegan2 distillation for feed-forward image manipulation. In Proc. European Conf. Computer Vision, pages 170-186. Springer, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[33] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 448 + ], + "type": "text", + "content": "[34] Ting Chun Wang, Ming Yu Liu, Jun Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proc. IEEE Int'l Conf. Computer Vision and Pattern Recognition, 2018. 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[35] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proc. Int'l Conf. Computer Vision, pages 1905-1914, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "type": "text", + "content": "[36] Zongze Wu, Yotam Nitzan, Eli Shechtman, and Dani Lischinski. StyleAlign: Analysis and applications of aligned stylegan models. arXiv preprint arXiv:2110.11323, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "text", + "content": "[37] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Vtoonify: Controllable high-resolution portrait video style transfer. ACM Transactions on Graphics, 41(6):1-15, 2022. 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 570, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 634 + ], + "type": "text", + "content": "[38] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In Proc. European Conf. Computer Vision, pages 85-101. Springer, 2022. 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "text", + "content": "[39] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proc. European Conf. Computer Vision, pages 334-349. Springer, 2018. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[40] Jiapeng Zhu, Ruili Feng, Yujun Shen, Deli Zhao, Zheng-Jun Zha, Jingren Zhou, and Qifeng Chen. Low-rank subspaces" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21009" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 183 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "in gans. In Advances in Neural Information Processing Systems, volume 34, 2021. 1, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "text", + "content": "[41] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In Proc. European Conf. Computer Vision, pages 592-608. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 286, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 286, + 183 + ], + "type": "text", + "content": "[42] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Barbershop: Gan-based image compositing using segmentation masks. ACM Transactions on Graphics, 40(6):1-13, 2021. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21010" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/ee18bf56-6e38-4711-9114-1ec06f896ae1_content_list.json b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/ee18bf56-6e38-4711-9114-1ec06f896ae1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a9c0e41eb8b8804d6e1461d77cf790b6be6f3001 --- /dev/null +++ b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/ee18bf56-6e38-4711-9114-1ec06f896ae1_content_list.json @@ -0,0 +1,1661 @@ +[ + { + "type": "text", + "text": "StyleInV: A Temporal Style Modulated Inversion Network for Unconditional Video Generation", + "text_level": 1, + "bbox": [ + 173, + 119, + 797, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhan Wang, Liming Jiang, Chen Change Loy \nS-Lab, Nanyang Technological University \n{yuhan004, liming002, ccloy}@ntu.edu.sg", + "bbox": [ + 299, + 185, + 671, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unconditional video generation is a challenging task that involves synthesizing high-quality videos that are both coherent and of extended duration. To address this challenge, researchers have used pretrained StyleGAN image generators for high-quality frame synthesis and focused on motion generator design. The motion generator is trained in an autoregressive manner using heavy 3D convolutional discriminators to ensure motion coherence during video generation. In this paper, we introduce a novel motion generator design that uses a learning-based inversion network for GAN. The encoder in our method captures rich and smooth priors from encoding images to latents, and given the latent of an initially generated frame as guidance, our method can generate smooth future latent by modulating the inversion encoder temporally. Our method enjoys the advantage of sparse training and naturally constrains the generation space of our motion generator with the inversion network guided by the initial frame, eliminating the need for heavy discriminators. Moreover, our method supports style transfer with simple fine-tuning when the encoder is paired with a pretrained StyleGAN generator. Extensive experiments conducted on various benchmarks demonstrate the superiority of our method in generating long and high-resolution videos with decent single-frame quality and temporal consistency. Code is available at https://github.com/johannwyh/StyleInV.", + "bbox": [ + 75, + 305, + 473, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 723, + 209, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unconditional video generation aims at learning a generative model to create novel videos from latent vectors. Despite extensive studies [42, 32, 33, 38, 11, 50] in addressing this problem, it remains challenging to generate high-resolution videos with both favorable quality and motion coherence over a long-term duration. The core difficulties in this task lie in modeling consistent motion and managing the high memory consumption introduced by the addition of the temporal dimension.", + "bbox": [ + 75, + 750, + 468, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To ensure high single-frame resolution and quality, many", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/277a5ac1fb41b9074a00430b0e4b98da665a53bd83c976dfde04a4d851567ecc.jpg", + "image_caption": [ + "(a) Autoregressive Framework (e.g., MoCoGAN-HD)" + ], + "image_footnote": [], + "bbox": [ + 527, + 273, + 864, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f1dae7e75cc5c72e52c718ff9325e806423d2651a326c0f784991ffd34f9157.jpg", + "image_caption": [ + "(b) Non-autoregressive Framework (Ours)", + "Figure 1: A comparison between autoregressive and non-autoregressive pipeline: (a) Previous autoregressive motion generators require generating the whole clip for a 3D-convolution-based discriminator. (b) Our non-autoregressive motion generator, StyleInV, is an inversion network modulated by temporal style (as a random function of $t$ ), which enjoys sparse training using a 2D-convolution-based discriminator." + ], + "image_footnote": [], + "bbox": [ + 535, + 436, + 861, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "existing studies, such as MoCoGAN-HD [36], employ a powerful image generator such as StyleGAN [22] as a backbone to serve as a strong generative prior. This approach shifts the focus towards developing a robust motion generator that can capture temporally coherent motion. Most of these methods model motion in an auto-regressive manner, where the next latent is sampled conditioned on the previous one (see Fig. 1). However, this design has two main drawbacks. First, while good performance requires seeing", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22851", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b14e68c39c107fbc4ea11e36709f5f1f825d7a28b201b9499b6f00d3e38f1705.jpg", + "image_caption": [ + "Figure 2: Inverted latent space visualization and modulated inversion process: When the StyleGAN generator is trained with video frame data, $\\mathcal{W}$ space is well clustered by human identities and provides promising inversion results. Thus, the modulated inversion process can easily find the target latent corresponding to the same identity (shifted to the next motion) as the source one." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 444, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a long sequence of images, the use of heavy 3D discriminators limits its ability to be trained with longer videos. Second, the autoregressive motion generator can lead to motion collapse when extrapolating to generate longer videos.", + "bbox": [ + 75, + 486, + 468, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we present an effective framework for non-autoregressive motion generation that is capable of generating long and high-resolution videos. Our approach leverages learning-based Generative Adversarial Network (GAN) inversion, which learns the inverse mapping of GANs via an inversion network that consists of an encoder and a decoder1. To generate long and coherent videos, we exploit the unique characteristic of the inversion encoder, which captures a rich and smooth manifold between the mapping of images and latent. As illustrated in Fig. 1, to generate a sequence of smooth motion latents, we just need to provide the initial latent code and modulate the inversion encoder with temporal style codes, which are encodings of timestamps with randomness. The motion latents can then be mapped by a StyleGAN decoder to generate a video.", + "bbox": [ + 75, + 547, + 470, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed framework offers several advantages in a single unified framework. First, the use of an inversion network naturally constrains the generation space to stay consistent with the desired appearance, which is de", + "bbox": [ + 75, + 776, + 470, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fined by the initial latent code. As demonstrated in Fig. 2, this leads to a significant benefit. Second, thanks to the flexibility of the inversion network in accepting temporal styles of arbitrary timestamps, the framework allows non-autoregressive generation and sparse training [52, 35]. These merits help alleviate the need for heavy discriminators to ensure temporal consistency, as is required in existing approaches. In our implementation, we only need to use a 2D convolutional discriminator instead of a 3D discriminator like MoCoGAN-HD. Third, Unlike existing state-of-the-art methods [52, 35, 7] that couple content and motion decoding in one synthesis network, our framework can naturally support content decoder fine-tuning on different image datasets. Specifically, after fine-tuning the decoder (e.g., StyleGAN2) on another image dataset with the mapping layers and low-resolution synthesis layers fixed, given the same sequence of synthesized motion latents, the generated video can possess the new style of the fine-tuning dataset while preserving the motion patterns of the video generated by the parent content decoder.", + "bbox": [ + 500, + 90, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contribution of this work is a novel motion generator that modulates a GAN inversion network. This is the first attempt to build such a generator, and it offers several advantages in a unified framework over existing approaches. These advantages include consistent generation, sparse training, and flexibility in supporting style transfer with simple fine-tuning. We additionally contribute a reformulation to the conventional sparse training, through first-frame-aware acyclic positional encoding (FFA-APE) and first-frame-aware sparse training (FFA-ST), to ensure that our motion generator can faithfully reconstruct the initial frame and that the generated video is smooth and continuous. Extensive experiments on DeeperForensics [17], FaceForensics [31], SkyTimelapse [47] and Tai-Chi-HD [34] datasets show that our model is comparable to or even better than state-of-the-art unconditional video generation methods [36, 52, 35] both qualitatively and quantitatively.", + "bbox": [ + 498, + 393, + 892, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 662, + 640, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GAN inversion. The goal of GAN inversion is to find the corresponding vector in the latent space of a pretrained GAN [22, 23] to reconstruct the input image. Existing methods can be classified into three categories [46]: (1) learning-based methods [8, 37, 49, 3, 2, 44, 29], which leverage an encoder network to directly map an image into a latent vector; (2) optimization-based methods [43, 1, 45, 48, 56, 57], which iteratively find the latent vector that best reconstructs the input image using gradient descent; and (3) hybrid models [6, 5, 9, 55], which initialize the iteration process with the result of an encoder network. The design of our motion generator follows the learning-based approach. Therefore, our method is trainable, efficient for single-image inference, and suitable for hierarchical mod-", + "bbox": [ + 498, + 688, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1In many contexts, the decoder is a StyleGAN, and the encoder learns to encode a given image to meaningful latent vectors in the StyleGAN space. There is a variety of image manipulation applications [46, 29, 4] developed based upon such an inversion framework.", + "bbox": [ + 75, + 849, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22852", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ulation. We devise the motion generator on the $\\mathcal{W}$ space and use the StyleGAN generated latent as the initial content code to guide the modulated inversion process (see Fig. 2).", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unconditional video generation. Unconditional video generation aims to model the distribution of real videos in a training dataset and generate videos from sampled noise vectors. Many recent studies on this topic are inspired by the success of GANs in image generation. VGAN [42] applies 3D convolutions in both the generator and discriminator, while TGAN [32] optimizes this design by decomposing the generator into an image generator, which is shared by the generation of each frame, and a motion generator2. This framework has been followed by most subsequent studies, such as MoCoGAN [38], which applies a content-motion decomposition. Some approaches [33, 11, 19] have focused on reducing the computational cost of the video discriminator, but the cost is still proportional to the video duration and resolution. Some recent methods have applied more advanced generative frameworks and techniques to unconditional video generation. For example, VideoGPT [50] uses VQ-VAE [28] and GPT [10] to formulate a non-GAN-based video generation approach. Recent studies have also explored unconditional video generation with higher resolution and longer duration. For example, Long-Video-GAN [7] develops a two-phase model that focuses on improving the long-term temporal dynamics of video generation. MoCoGAN-HD [36] and StyleVideoGAN [13] study the generation of latent trajectories in the latent space of a pretrained StyleGAN2 generator. Our approach is inspired by these studies, but differs in the design of the motion generator. Our motion generator is non-autoregressive, thus alleviating the use of heavy discriminators, and it is unique since it obtains the motion latent via modulating a GAN inversion network. This design allows us to attain better motion consistency and semantics.", + "bbox": [ + 75, + 138, + 470, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works [52, 35] explored neural representation-based generators and trained them sparsely as an image GAN. StyleSV [53] improves this framework by introducing StyleGAN3 [21] architecture and several temporal designs. In our work, we extend the idea of sparse training to first-frame-aware sparse training, allowing it to be applied to a generation pipeline conditioned on the initial latent.", + "bbox": [ + 75, + 625, + 468, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-based video generation. The diffusion models [15, 30], a new paradigm for image generation tasks, have also achieved significant progress in the task of unconditional video generation [16, 24, 41, 51]. Despite their success, temporal consistency is still an open problem for diffusion models, and GAN-based models exhibit a clear advantage in terms of inference speed.", + "bbox": [ + 75, + 734, + 468, + 840 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2cf7ca05a5f0e6b5c34507ab42726cc8feeacf35c1acadb12f52db5aff4cea89.jpg", + "image_caption": [ + "Figure 3: From a typical inversion encoder to StyleInV: We add AdaIN layers at the end of each residual block to inject the temporal style, which is a combination of time positional encoding and the first frame latent code. Here \"A\" stands for a learned affine transform [22]." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 887, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 532, + 635, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries of Inversion Encoder", + "text_level": 1, + "bbox": [ + 500, + 558, + 805, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An inversion encoder maps an input image to a vector in the $\\mathcal{W}$ or $\\mathcal{W}+$ latent space of a pretrained StyleGAN2 generator. The generated image that corresponds to this vector should faithfully reconstruct the details of the input image. Therefore, when based on $\\mathcal{W}$ latent space, given an input image $\\mathbf{x}$ , the reconstruction process can be defined on top of the inversion network Inv as:", + "bbox": [ + 496, + 580, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} := G (\\operatorname {I n v} (\\mathbf {x})) := G (E (\\mathbf {x}) + \\overline {{\\mathbf {w}}}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 695, + 890, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $E$ and $G$ denote the inversion encoder and StyleGAN generator, respectively. $\\overline{\\mathbf{w}}\\in \\mathbb{R}^{512}$ denotes the average latent vector of the generator in the $\\mathcal{W}$ latent space. In our implementation, the encoder $E$ is a convolutional network backbone that outputs a 512-dimensional vector from the last layer embedding, as shown in Fig. 3(left). We build the encoder on the $\\mathcal{W}$ latent space, which eases the design of temporal modulation.", + "bbox": [ + 496, + 718, + 893, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Temporal Style Modulated Inversion Encoder", + "text_level": 1, + "bbox": [ + 500, + 847, + 885, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We observe that the latent space of a StyleGAN trained on a video dataset is typically well-clustered by its content", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2In the original paper of TGAN [32], the authors called this module temporal generator, which is equivalent to the motion generator used in subsequent studies [38, 36, 52] and in our paper.", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22853", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "subject. Figure 2 shows an example of human face videos, where we depict the results of inverting video clips of different identities into the $\\mathcal{W}$ space and visualizing them with t-SNE [40]. It can be observed that the latent space is grouped by human identities. We also observe the same property in video datasets that follow other distributions. This phenomenon suggests that the inversion network inherits some important temporal priors that we could leverage to maintain motion consistency in generated videos.", + "bbox": [ + 75, + 90, + 468, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by this observation, we propose StyleInV, in which the motion latent is generated by modulating a GAN inversion network with temporal styles. Figure 3(right) illustrates the pipeline of our framework. The temporal style $s_t$ of a timestamp $t$ consists of two parts: the motion code $v_t$ and the latent code of the initial frame $w_0$ . Inspired by [35], we use an acyclic positional encoding module to compute a dynamic embedding of the timestamp $t$ . However, unlike [35], we make the embedding of the zero timestamp fixed, so this module becomes first-frame-aware. We provide more details in Section 3.3. The latent code $w_0$ of the initial frame is concatenated with the motion code for content-adaptive affine transform.", + "bbox": [ + 75, + 226, + 470, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The temporal style is injected into the inversion encoder through AdaIN layers at the end of each convolution block. With this design, the encoder $E$ of StyleInV becomes a function of the initial latent code $w_{0}$ and timestamp $t$ . The modulated inversion process can be defined as:", + "bbox": [ + 76, + 422, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {t} := G \\left(\\operatorname {S t y l e I n V} \\left(\\boldsymbol {w} _ {0}, t\\right)\\right) := G \\left(E \\left(G \\left(\\boldsymbol {w} _ {0}\\right), \\boldsymbol {s} _ {t}\\right) + \\boldsymbol {w} _ {0}\\right). \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 508, + 468, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, the output of $E$ serves as the residual w.r.t. $w_0$ , instead of $\\overline{w}$ . This modification provides more explicit content information guidance for the inversion encoder.", + "bbox": [ + 75, + 535, + 468, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, we first train a raw inversion encoder following Eq. (1) on all video frames. Then, we use this network to initialize the weights of all convolution layers in the StyleInV encoder. Other parameters (e.g., FFA-APE and Affine Transforms) are randomly initialized. Finally, the entire StyleInV encoder is trained end-to-end.", + "bbox": [ + 75, + 580, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. FFA-APE", + "text_level": 1, + "bbox": [ + 76, + 680, + 191, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The original implementation of acyclic positional encoding (APE) [35] samples a series of noise vectors $\\boldsymbol{z}_{t_0}^{\\mathsf{m}},\\dots ,\\boldsymbol{z}_{t_n}^{\\mathsf{m}}\\sim \\mathcal{N}(\\mathbf{0},\\boldsymbol {I})$ where $t_i = i\\cdot \\delta^z$ . We call these temporal points anchor points. Here $\\delta^z$ is a set constant distance between adjacent anchor points. Then, the noise vectors are mapped to tokens $\\boldsymbol{u}_{t_0},\\dots ,\\boldsymbol{u}_{t_n}$ by a paddingless conv1d-based motion mapping network. The computation of the acyclic positional encoding $\\boldsymbol{v}_t$ of arbitrary timestamp $t$ is achieved by a scalable and learnable interpolation between the tokens of two adjacent anchor points that cover $t$ . The computation pipeline is shown in Fig. 4.", + "bbox": [ + 75, + 703, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our non-autoregressive generation pipeline, the modulated inversion encoder needs to faithfully reconstruct the", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a8a92dc3843daf5f14c6b568af60c48db73ba8826ba843afb7b689d0728245fc.jpg", + "image_caption": [ + "Figure 4: FFA-APE: A simplified case to compute $v_{16}$ for demonstrating the original acyclic positional encoding and our FFA-APE when $\\delta^z = 32$ and convld kernel size is 3. In FFA-APE, the encoding of zero timestamp $(u_0)$ only depends on constant paddings and a constant noise vector, thus is fixed for any sampled noise vector sequence." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 890, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "initial frame when the input timestamp is zero, making it necessary to fix the computation of the APE for the zero timestamp $v_{0}$ . The original APE computation for $v_{0}$ is dynamic and depends on randomly sampled noise vectors, which can lead to dynamic output that is not desired. To address this, we devise a first-frame-aware acyclic positional encoding (FFA-APE) method that fixes $v_{0}$ while maintaining the smoothness of APE (see Fig. 4). We achieve this by replacing the noise vector for the first anchor point with a learnable constant vector $c_{0}^{\\mathrm{m}}$ , and using left-sided conv1d layers with constant padding instead of the padding-less conv1d layers. This way, the value of $v_{0}$ only depends on the constant vector $c_{0}^{\\mathrm{m}}$ and the left-padded vectors, which are also constant. As a result, $v_{0}$ is naturally fixed without affecting the continuity of positional encoding.", + "bbox": [ + 496, + 337, + 892, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. FFA-ST", + "text_level": 1, + "bbox": [ + 500, + 574, + 599, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the first-frame-aware sparse training specially designed for our framework. Recent non-autoregressive video generation approaches [52, 35] use a discriminator design that only considers $k$ frames $\\pmb{x}_{t_1},\\dots ,\\pmb{x}_{t_k}$ for each video, distinguishing the realness of the input conditioned on the time difference of input frames $\\delta_{i} = t_{i + 1} - t_{i}$ . This training scheme is called sparse training. StyleGAN-V [35] has analyzed the choice of $k$ and found that $k = 3$ is ideal for most datasets. The discriminator is defined as $D(\\pmb{x}_{t_{1,2,3}},\\delta_{1,2})$ .", + "bbox": [ + 496, + 598, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow this training scheme to make full use of our non-autoregressive framework. Nonetheless, using only three randomly sampled timestamps to train the generator and discriminator can result in sharp transitions at the beginning of the generated video, where the generated $\\pmb{x}_0$ and $\\pmb{x}_1$ usually diverge too much, and sometimes even switch to another identity and never return. This happens because although we define the generation process of a video as a modulated inversion process of the start frame, the discriminator is unaware of it. The discriminator only focuses on", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22854", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the smoothness of generated latent trajectories, failing to ensure the motion generator produces frames that share the identity with the start frame.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve this problem, we introduce the initial frame into the discriminator to enhance content consistency and motion smoothness. The adversarial loss for the first-frame-aware discriminator (FFA-D) can be written as:", + "bbox": [ + 75, + 136, + 468, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {y} _ {t _ {0, 1, 2, 3}} = G \\left(\\text {S t y l e I n V} \\left(w _ {0}, t _ {0, 1, 2, 3}\\right)\\right), \\\\ \\mathcal {L} _ {a d v} = \\mathbb {E} _ {\\boldsymbol {x} \\sim p _ {v}} \\left[ \\log D \\left(\\boldsymbol {x} _ {t _ {0, 1, 2, 3}}, \\delta_ {0, 1, 2}\\right) \\right] \\tag {3} \\\\ + \\mathbb {E} _ {w _ {0} \\sim p _ {\\mathcal {W}}} \\left[ \\log (1 - D (\\boldsymbol {y} _ {t _ {0, 1, 2, 3}}, \\delta_ {0, 1, 2}) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 205, + 468, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we specify $t_0 = 0$ . Here, $p_v$ and $p_{\\mathcal{W}}$ denote the real data distribution and $\\mathcal{W}$ latent space distribution, respectively. To explicitly enforce initial frame reconstruction, we use a $L_2$ loss for the generated $y_{t_0}$ :", + "bbox": [ + 75, + 273, + 468, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L _ {2}} = \\left\\| G \\left(w _ {0}\\right) - G \\left(\\text {S t y l e I n V} \\left(w _ {0}, 0\\right)\\right) \\right\\| _ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 345, + 468, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we apply latent regularization [29, 26] to the encoder's output, so as to enhance content consistency:", + "bbox": [ + 76, + 372, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\sum_ {i = 0} ^ {3} | | E (G (w _ {0}), t _ {i}) | | _ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 412, + 468, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss function for training our motion generator and the discriminator is defined as:", + "bbox": [ + 76, + 448, + 468, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {E} \\max _ {D} \\mathcal {L} _ {a d v} + \\min _ {E} \\left(\\lambda_ {L _ {2}} \\mathcal {L} _ {L _ {2}} + \\lambda_ {r e g} \\mathcal {L} _ {r e g}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 489, + 468, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\lambda_{L_2}$ and $\\lambda_{reg}$ are the loss hyperparameters. We also apply discriminator adaptive augmentation [20, 35] and $r1$ regularization [22, 35] to further improve the training stability and generation quality.", + "bbox": [ + 75, + 521, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Finetuning-based Style Transfer", + "text_level": 1, + "bbox": [ + 76, + 590, + 361, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our 'inversion encoder+decoder' framework can naturally take a pretrained StyleGAN model as the generator. And such a configuration allows the generator to be finetuned for different styles, and yet still able to use the motion generator for generating new video with styles. The capability is not possible with existing non-autoregressive video generation methods [52, 35] because they cannot be finetuned under an image GAN training scheme.", + "bbox": [ + 75, + 613, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve style transfer, as illustrated in Fig. 5, we fine-tune the pre-trained StyleGAN model using an image dataset, such as MegaCartoon [27], while keeping the mapping network and low-resolution $(\\leqslant 32^2$ , coarse and middle layers in [22]) synthesis blocks fixed. This configuration maintains the distribution of the $\\mathcal{W}$ space during finetuning. To improve identity preservation and reduce artifacts, we apply both a perceptual loss [18] and an identity loss [12] between the images generated by the original and fine-tuned StyleGAN. We show some visual results in Fig. 8. The style-transferred video maintains the", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/555f9ea1ab613a2e8e335f4d9eb0bc8af194509ffc3946ec7326c60b3e51f314.jpg", + "image_caption": [ + "Figure 5: Finetuning-based style transfer: our framework allows easy fine-tuning of decoder (a pretrained StyleGAN generator) to a new domain by freezing mapping and low-resolution $(\\leqslant 32^{2})$ layers. Standard identity loss and perceptual loss are applied to improve identity preservation and reduce artifacts. The style-transferred videos can then be generated by incorporating the StyleInV motion generator." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 890, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "same motion pattern as the video generated by the parent model, while adopting a new style from the fine-tuning image dataset. It is noteworthy that the finetuning process is independent of the video generation training. It means that the finetuning-based style transfer is \"plug-and-play\" as the fine-tuned image generator can be used on any StyleInV models. There is no additional computational cost in the inference either.", + "bbox": [ + 498, + 438, + 890, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 571, + 632, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use four video datasets in our main experiments: DeeperForensics $256^{2}$ [17], FaceForensics $256^{2}$ [31], SkyTimelapse $256^{2}$ [47] and TaiChi $256^{2}$ [34]. The cropping strategy for DeeperForensics [17] and FaceForensics [31] is different. For DeeperForensics, we use a stabilized FFHQ [22] cropping strategy [25], while we follow the strategy firstly adopted by TGAN-V2 [33] for FaceForensics. Please refer to the supplementary material for a detailed discussion.", + "bbox": [ + 498, + 598, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We explore four state-of-the-art methods for comparison: MoCoGAN-HD [36], DIGAN [52], StyleGAN-V [35] and Long-Video-GAN [7]. Among these methods, MoCoGAN-HD and DIGAN require an explicit setting of the training clip length. We follow the default setting of their paper to set the clip length as 16 for both methods. This setting is identical to StyleGAN-V [35].", + "bbox": [ + 498, + 734, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, on DeeperForensics, we explore an optimized setting on DIGAN and MoCoGAN-HD for a more fair comparison. For DIGAN, we directly increase the clip length to 128 frames. For MoCoGAN-HD, we apply the", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22855", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "first-frame-aware sparse training to train its motion generator, so as to avoid using a heavy 3D discriminator, allowing it to be trained with 128-frame clips as other methods.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We use Fréchet Inception Distance (FID) [14] and Fréchet Video Distance (FVD) [39] to evaluate all models quantitatively. In practice, we follow the metric calculation framework provided by StyleGAN-V [35] to first generate a fake video dataset with 2,048 synthesized clips, each of 128 frames. For FID, we sample 50k frames from real and fake video datasets to compute the result. For FVD, we compute $\\mathrm{FVD}_{16}$ and $\\mathrm{FVD}_{128}$ with the first 16 frames and all 128 frames of each clip, respectively. We use FID results to show the single-frame image quality of each method.", + "bbox": [ + 75, + 137, + 467, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure a fair comparison, we re-benchmark the quantitative results of every method on every dataset. We retrain all the baselines using the official paper setting, except for MoCoGAN-HD on SkyTimelapse, where an officially released checkpoint is available. For more implementation details, please refer to the supplementary material.", + "bbox": [ + 75, + 289, + 467, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 76, + 391, + 215, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. Table 1 summarizes the quantitative results of our method compared to other baselines. Our method achieves competitive quantitative results on all the benchmarks. Notably, although MoCoGAN-HD and DIGAN are trained with clips of 16 frames, we still outperform them in terms of $\\mathrm{FVD}_{16}$ metrics on all four datasets.", + "bbox": [ + 75, + 415, + 467, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. Figure 6 shows the qualitative comparison between our method and the baselines on all four datasets. MoCoGAN-HD and DIGAN both suffer from motion collapse, resulting in a degraded generation quality over time. StyleGAN-V shows an impressive visual performance on FaceForensics and SkyTimelapse, but it sometimes fails to maintain the identity and accessories on DeeperForensics and lacks diversity and magnitude of motion over a long time span on TaiChi (the subject gradually fixes at one state). Long-Video-GAN is exceptionally good at SkyTimelapse, but it cannot achieve similar performance on other datasets. It fails to maintain the identity on DeeperForensics, and its single-frame content on TaiChi lacks details and is inferior to other methods. The generated videos by Long-Video-GAN collapse on FaceForensics.", + "bbox": [ + 75, + 506, + 467, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast to existing methods, our method demonstrates stable results on all four datasets, particularly with superior identity preservation on human-face video and long-term generation quality on TaiChi. Although our method outperforms existing methods in terms of content quality, continuity, and quantitative results, the motion semantics of our generated videos on SkyTimelapse are inferior to those on other datasets. This could be one of the limitations of our work and an area for future improvement. Extended experiments. We present more in-depth comparisons in Table 2 and Fig. 7 by introducing training im", + "bbox": [ + 75, + 734, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: FID, $\\mathrm{FVD}_{16}$ and $\\mathrm{FVD}_{128}$ results of video generation methods on (a) DeeperForensics $256^{2}$ , (b) FaceForensics $256^{2}$ , (c) TaiChi $256^{2}$ , and (d) SkyTimelapse $256^{2}$ . **Bolds** indicate best and **underline** indicate the second best.", + "bbox": [ + 498, + 88, + 890, + 148 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a4bf78635fb69b58b0a3af5273817b40503d6d3b1c5cace12b96428dee08c88a.jpg", + "table_caption": [ + "(a) DeeperForensics $256^{2}$" + ], + "table_footnote": [], + "table_body": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD135.30101.07610.30
DIGAN191.9946.691060.27
StyleGAN-V59.5939.3368.81
Long-Video-GAN56.5474.77169.45
StyleInV (ours)54.0541.5853.93
", + "bbox": [ + 511, + 176, + 877, + 277 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8ffcfff41b90f82c21429be436830ed26686478724dfedc866b8283c909345bd.jpg", + "table_caption": [ + "(b) FaceForensics $256^{2}$" + ], + "table_footnote": [], + "table_body": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD24.45112.67486.69
DIGAN151.53146.621993.20
StyleGAN-V8.6452.92108.86
Long-Video-GAN40.40233.26567.78
StyleInV (ours)12.0647.88103.63
", + "bbox": [ + 511, + 301, + 877, + 402 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/485826393b94d8ba8e005d3e58b4e3253cde9cbb63fc897d58c19904d458faa4.jpg", + "table_caption": [ + "(c) TaiChi $256^{2}$" + ], + "table_footnote": [], + "table_body": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD73.61315.03622.95
DIGAN67.24196.77954.93
StyleGAN-V35.68254.74477.78
Long-Video-GAN43.90248.55502.65
StyleInV (ours)41.55185.72328.90
", + "bbox": [ + 511, + 426, + 877, + 526 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3f5a460fa021cb8189d1b4a370cc8f9436959b2f5ef3911485d47aacaafc2d07.jpg", + "table_caption": [ + "(d) SkyTimelapse $256^{2}$" + ], + "table_footnote": [], + "table_body": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD251.81696.584116.03
DIGAN32.83148.08269.43
StyleGAN-V16.9581.32197.83
Long-Video-GAN25.41116.50152.70
StyleInV (ours)14.3277.04194.25
", + "bbox": [ + 514, + 550, + 877, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: FID, $\\mathrm{FVD}_{16}$ and $\\mathrm{FVD}_{128}$ results of extended experiments on DeeperForensics $256^{2}$ . We apply sparse training to MoCoGAN-HD [36] (#1) and change the preset clip length of DIGAN [52] to 128 (#2). **Bolds** indicate best. (-) indicates a smaller (better) quantitative result, while (+) indicates a larger (worse) one, compared with Table 1a.", + "bbox": [ + 498, + 664, + 890, + 755 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7f7e9f5f15910feea0330622064310a3409b11c157667c6abf7b3fdbc0f4708d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#MethodFID (↓)FVD16 (↓)FVD128 (↓)
1[36] + Sparse Training55.84 (-)54.58 (-)129.13 (-)
2[52] + Clip 12874.80 (-)87.42 (+)95.80 (-)
3StyleInV (ours)54.0541.5853.93
", + "bbox": [ + 504, + 763, + 888, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "provements to baselines. Increasing the clip length generally improves the results of MoCoGAN-HD and DIGAN, but they are still inferior to our method. Notably, training", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22856", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56532472b55d24987e00ab4a2a7ea17bcf8e592b0c366abea6ba417212b5770b.jpg", + "image_caption": [ + "(a) DeeperForensics $256^{2}$" + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 475, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa0537a460bfc8d84f3504c1cd466626fb3b59f20d7125d0aa03596485022036.jpg", + "image_caption": [ + "(b) FaceForensics $256^{2}$" + ], + "image_footnote": [], + "bbox": [ + 491, + 88, + 883, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43ec2d0d21d64d2c460706b674201465506293d1307c2c8012a1adbfe48e9723.jpg", + "image_caption": [ + "(c) TaiChi $256^{2}$" + ], + "image_footnote": [], + "bbox": [ + 84, + 309, + 475, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/534c12848b3af10c60bc89e47d58713f7b31efbb9efafdd5d49edf17dab4a01c.jpg", + "image_caption": [ + "(d) SkyTimelapse $256^{2}$" + ], + "image_footnote": [], + "bbox": [ + 496, + 309, + 883, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea122f05fbb5a5bfe1ae4aea506e0b126b977781c803f94e5b79e2384b369130.jpg", + "image_caption": [ + "Figure 6: Uncurated samples from the existing methods on DeeperForensics $256^{2}$ , FaceForensics $256^{2}$ , TaiChi $256^{2}$ and SkyTimelapse $256^{2}$ , respectively. We sample a 128-frame video and display every 16 frames, starting from $t = 0$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 585, + 478, + 705 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93fd0e7287387afd330d39731987c3ce921e7bcac9a1a197ccde13494cb39b12.jpg", + "image_caption": [ + "Figure 7: Qualitative comparison of extended experiments. “M+ST” and “D-C128” correspond to Table 2 (#1) and (#2), respectively. Each row shows the first and last 128 frames of a 2056-frame (68.5s) video, displayed every 16 frames." + ], + "image_footnote": [], + "bbox": [ + 480, + 585, + 888, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with longer clips harms the short-term $\\mathrm{FVD}_{16}$ result of DIGAN, which indicates its tradeoff between duration length and local temporal quality. Qualitatively, for both methods, the generated content is evidently improved within 128 frames, although the sparsely trained MoCoGAN-HD exhibits issues with identity switching. Motion collapse is still observed when MoCoGAN-HD and DIGAN generates long videos. In contrast, our method can stably generate ex", + "bbox": [ + 75, + 775, + 468, + 896 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tremely long videos without motion collapse. Our method outperforms the sparsely trained MoCoGAN-HD, demonstrating the superiority of our motion generator design.", + "bbox": [ + 498, + 775, + 890, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Properties", + "text_level": 1, + "bbox": [ + 500, + 830, + 614, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As discussed in Section 3.5, our method has the unique advantage over state-of-the-art methods, such as StyleGAN-V and Long-Video-GAN, on its high compati", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22857", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fffce52942f24b84dde75e8bad4ec7faad5f6fb379d3d890f26be0c5e61ee3e0.jpg", + "image_caption": [ + "Figure 8: Finetuning-based style transfer result. The 1st row is generated by the parent model trained on CelebV-HQ [54]. The 2nd, 3rd, and 4th row uses the StyleGAN generator fintuned on Cartoon, Arcane, and MetFace, respectively." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 491, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ecdfb14ff0a42adaca4cec5a778eea248f2b918b169ecf1fa81d11a57d08d8b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 90, + 887, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f09d88d1e8513870048be291257c1dd67a2f7f0d77367991c2b0ded16e7b516e.jpg", + "image_caption": [ + "Figure 9: Initial-frame conditioned generation and style transferred results." + ], + "image_footnote": [], + "bbox": [ + 81, + 306, + 467, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bility with StyleGAN-based downstream techniques.", + "bbox": [ + 75, + 640, + 424, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finetuning-based style transfer. We train the parent model (motion generator and StyleGAN) on CelebV-HQ [54], as its rich identity makes it more suitable for transfer learning. To perform style transfer, we fine-tune the StyleGAN on the Cartoon [27], MetFace [20], and Arcane datasets following the procedure outlined in Section 3.5. In Fig. 8, we show examples where the same StyleInV-generated latent sequence is decoded by different but aligned StyleGAN generators. Our method achieves satisfactory results in terms of smooth video style transfer with well-aligned face structure, identity, and expression, demonstrating its desirable properties and potential for various applications. More results can be found in the supplementary video.", + "bbox": [ + 75, + 657, + 468, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initial-frame conditioned generation. Our network supports generating a series of content given a real-world im", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/57ce44e6928ac730e35f9cc8344300da4566e605a1aad58f6426a678d4db0dad.jpg", + "table_caption": [ + "Table 3: Ablation result on the DeeperForensics dataset." + ], + "table_footnote": [], + "table_body": "
#MethodFID (↓)FVD16 (↓)FVD128 (↓)
1w/o inversion encoder54.3559.49152.82
2w/o FFA-APE55.2688.98144.52
3w/o Eq.(4) & Eq.(3)52.5567.4358.88
4w/o Eq.(3)53.9586.3259.76
5Ours54.0541.5853.93
", + "bbox": [ + 504, + 335, + 888, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "age as the initial frame. We first inverse the image into the StyleGAN2 latent space with a pSp [29] encoder, which is trained to initialize the weights of StyleInV. We treat it as the 512-dimensional initial frame latent $w_{0}$ , then use it to generate a video with our StyleInV. The generated latent sequence can be also applied to a finetuned image generator to synthesize a style-transferred animation video. Through this pipeline, the real image is reconstructed twice, the first time is during the inversion process, while the second time is when synthesizing $G(\\text{StyleInV}(w_{0},0))$ .", + "bbox": [ + 496, + 470, + 890, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When the real images are sampled from the training dataset (see Fig. 9 first two rows), $G(\\text{StyleInV}(w_0, 0))$ can faithfully reconstruct the raw image and generate high-quality videos. We then test the generation quality for real images sampled out of the training set (see Fig. 9 last two rows, where we select Benedict Cumberbatch and Anne Hathaway). We use the StyleGAN2 generator and StyleInV model trained on CelebV-HQ [54] dataset as it is richer in its identities. The results show that our StyleInV network can still generate meaningful videos while reconstructing the initial frame decently, and the style transfer results are smooth and well-aligned. Please refer to the supplementary material for more results.", + "bbox": [ + 496, + 623, + 892, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 830, + 663, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion generator design. We explore two alternative motion generator designs. The first is the autoregressive MoCoGAN-HD design, which has been discussed in Sec-", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22858", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5305fab939e19ec3b5aafaad54263943881abfb05bcfe9219646bba506746dc1.jpg", + "image_caption": [ + "Figure 10: Ablating FFA-APE. Generate the first frame with a fixed $w_0$ but different temporal noise sequences. The model without FFA-APE fails to reconstruct the initial frame and generates the first frame with randomness." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 467, + 176 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b7dd93cd4ea4567fc4badbf1360ab4d02d8d8ebe048a67c9cf13adabe7d94316.jpg", + "image_caption": [ + "Figure 11: Ablating FFA-ST. The $L_{2}$ loss ensures the initial frame reconstruction. But without the initial frame included in the discriminator, we still cannot preserve the identity." + ], + "image_footnote": [], + "bbox": [ + 81, + 267, + 465, + 353 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tion 4.1. For the second design, we remove all Convs, AdaIN and affine transform layers in Fig. 3 and let the mapped temporal style be the output of inversion encoder, i.e., the residual w.r.t. initial frame latent $w_0$ . It largely harms identity preservation. Both $\\mathrm{FVD}_{16}$ and $\\mathrm{FVD}_{128}$ degrade significantly as is shown in Table 3(#1).", + "bbox": [ + 75, + 441, + 467, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "FFA-APE. We evaluate the importance of our first-frame-aware acyclic positional encoding (FFA-APE) module by replacing it with the original design proposed by [35]. As shown in Fig. 10, the dynamic embedding of the zero timestamp prevents the network from faithfully reconstructing the initial frame. In contrast, our full method can stably realize reconstruction. In addition, the L2 loss in Eq. (4) fails to converge for the ablation method, and its gradient further harms the learning of the positional encoding module, leading to a much worse quantitative result shown in Table 3 (#2).", + "bbox": [ + 75, + 535, + 467, + 700 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "FFA-ST. We conduct two ablation experiments for the FFA-ST modules. In the first experiment, we remove the initial frame from the discriminator and remove the reconstruction loss (Eq.(4)) (Table 3(#3)). In the second experiment, we only remove the initial frame from the discriminator while keeping the reconstruction loss (Table 3(#4)). As shown in Fig. 11, without the reconstruction loss, our model cannot reconstruct the initial frame accurately. With the $L_{2}$ loss, the initial frame is reconstructed, but there is a sudden transition between the first two frames, and sometimes, the identity also changes, leading to a worse $\\mathrm{FVD}_{16}$ result. These experiments demonstrate the importance of our first-frame-aware discriminator (FFA-D).", + "bbox": [ + 75, + 704, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 90, + 617, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have presented a novel approach for unconditional video generation by employing a pretrained StyleGAN image generator. The proposed StyleInV motion generator generates latents in the StyleGAN2 latent space by modulating a learning-based inversion network, and thus capable of inheriting its informative priors of the initial latent. Our network features non-autoregressive training and uniquely supports fine-tuning based style transfer. Extensive experiments demonstrate the superiority of our method in generating long and high-resolution videos, outperforming state-of-the-art baselines. Here we also briefly discuss our limitations and broader impacts.", + "bbox": [ + 496, + 114, + 890, + 296 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.1. Limitations", + "text_level": 1, + "bbox": [ + 500, + 306, + 624, + 321 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inferior motion semantics on SkyTimelapse. Our motion semantics on SkyTimelapse [47] are inferior to those on other datasets. This could be due to different dataset characteristics, as videos in SkyTimelapse are not subject-centric and typically driven by global motions, which does not align perfectly with our model nature.", + "bbox": [ + 496, + 330, + 890, + 420 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The impact of dataset identity richness. When the scale of facial identities in the video dataset is too small, the effects of inversion, editing, and style transfer are constrained. Image generation quality. The generation quality of StyleGAN determines the performance upper bound of our method. The images generated by the StyleGAN2 models have artifacts in the background on SkyTimelapse [47], and lack fine details and a sense of structure on TaiChi [34].", + "bbox": [ + 496, + 421, + 890, + 541 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Model training. Our approach is two-stage, requiring 7.5 and 9 GPU days each, which is more than the 8 GPU days of StyleGAN-V [35]. Despite this, StyleInV is as efficient when finetuning the hyperparameters of the video generator, since the image generator only needs to be trained once.", + "bbox": [ + 496, + 542, + 890, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2. Broader Impacts", + "text_level": 1, + "bbox": [ + 500, + 628, + 666, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We believe that the potential of StyleInV can be further exploited. Our method can provide a natural solution towards mega-pixel level video generation and StyleGAN-based editing, and it might in return promote the research of learning-based GAN inversion methods.", + "bbox": [ + 496, + 652, + 890, + 727 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As for the negative side, StyleInV may ease the synthesis of better-quality fake videos with threats. We believe that it can be alleviated by developing more advanced falsified media detection methods or contributing larger-scale and higher-quality forgery detection datasets.", + "bbox": [ + 496, + 728, + 890, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2022-01-030). It is also supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). We thank Shuai Yang for his help in this work.", + "bbox": [ + 496, + 803, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22859", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. StyleFlow: Attribute-conditioned exploration of StyleGAN-generated images using conditional continuous normalizing flows. ACM TOG, 40:1-21, 2021. 2", + "[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: Age transformation using a style-based regression model. ACM TOG, 40:1-12, 2021. 2", + "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. ReStyle: A residual-based StyleGAN encoder via iterative refinement. In ICCV, 2021. 2", + "[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit H Bermano. HyperStyle: StyleGAN inversion with hypernetworks for real image editing. In CVPR, 2022. 2", + "[5] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. ACM TOG, 38:1-11, 2019. 2", + "[6] David Bau, Jun-Yan Zhu, Jonas Wulff, William Peebles, Hendrik Strobelt, Bolei Zhou, and Antonio Torralba. Seeing what a GAN cannot generate. In ICCV, 2019. 2", + "[7] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurlPS, 2022. 2, 3, 5", + "[8] Lucy Chai, Jonas Wulff, and Phillip Isola. Using latent space regression to analyze and leverage compositionality in GANs. In ICLR, 2021. 2", + "[9] Lucy Chai, Jun-Yan Zhu, Eli Shechtman, Phillip Isola, and Richard Zhang. Ensembling with deep generative views. In CVPR, 2021. 2", + "[10] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 3", + "[11] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv preprint, arXiv:1907.06571, 2019. 1, 3", + "[12] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5", + "[13] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. StyleVideoGAN: A temporal generative model using a pretrained StyleGAN. In BMVC, 2021. 3", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In NeurIPS, 2017. 6", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3", + "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 3", + "[17] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. DeeperForensics-1.0: A large-scale dataset for real-world face forgery detection. In CVPR, 2020. 2, 5" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 5", + "[19] Emmanuel Kahembwe and Subramanian Ramamoorthy. Lower dimensional kernels for video discriminators. Neural Networks, 132:506-520, 2020. 3", + "[20] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 5, 8", + "[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 3", + "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 3, 5", + "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2", + "[24] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 3", + "[25] Jacek Naruniec, Leonhard Helminger, Christopher Schroers, and Romann M Weber. High-resolution neural face swapping for visual effects. In CGF, 2020. 5", + "[26] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: Text-driven manipulation of StyleGAN imagery. In ICCV, 2021. 5", + "[27] Justin NM Pinkney and Doron Adler. Resolution dependent GAN interpolation for controllable image synthesis between domains. arXiv preprint, arXiv:2010.05334, 2020. 5, 8", + "[28] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 3", + "[29] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In CVPR, 2021. 2, 5, 8", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3", + "[31] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint, arXiv:1803.09179, 2018. 2, 5", + "[32] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In ICCV, 2017. 1, 3", + "[33] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal GAN. IJCV, 128:2586-2606, 2020. 1, 3, 5", + "[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In NeurIPS, 2019. 2, 5, 9", + "[35] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. StyleGAN-V: A continuous video generator with the" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22860", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "price, image quality and perks of StyleGAN2. In CVPR, 2022. 2, 3, 4, 5, 6, 9", + "[36] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In ICLR, 2021. 1, 2, 3, 5, 6", + "[37] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. ACM TOG, 40:1-14, 2021. 2", + "[38] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In CVPR, 2018. 1, 3", + "[39] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint, arXiv:1812.01717, 2018. 6", + "[40] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. JMLR, 9(11), 2008. 4", + "[41] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. 2022. 3", + "[42] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. In NeurIPS, 2016. 1, 3", + "[43] Hui-Po Wang, Ning Yu, and Mario Fritz. Hijack-GAN: Unintended-use of pretrained, black-box GANs. In CVPR, 2021. 2", + "[44] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Lu Yuan, Gang Hua, and Nenghai Yu. E2Style: Improve the efficiency and effectiveness of StyleGAN inversion. TIP, 31:3267-3280, 2022. 2", + "[45] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: Disentangled controls for StyleGAN image generation. In CVPR, 2021. 2", + "[46] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. GAN inversion: A survey. TPAMI, 45:3121-3138, 2022. 2", + "[47] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In CVPR, 2018. 2, 5, 9", + "[48] Yangyang Xu, Yong Du, Wenpeng Xiao, Xuemiao Xu, and Shengfeng He. From continuity to editability: Inverting GANs with consecutive images. In ICCV, pages 13910-13918, 2021. 2", + "[49] Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, and Bolei Zhou. Generative hierarchical features from synthesizing images. In CVPR, 2021. 2", + "[50] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video generation using VQ-VAE and Transformers. arXiv preprint, arXiv:2104.10157, 2021. 1, 3", + "[51] Siyuan Yang, Lu Zhang, Yu Liu, Zhizhuo Jiang, and You He. Video diffusion models with local-global context guidance. In IJCAI, 2023. 3", + "[52] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In ICLR, 2022. 2, 3, 4, 5, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Qihang Zhang, Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Towards smooth video composition. In ICLR, 2023. 3", + "[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022.8", + "[55] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain GAN inversion for real image editing. In ECCV, 2020. 2", + "[56] Peihao Zhu, Rameen Abdal, Yipeng Qin, John Femiani, and Peter Wonka. Improved StyleGAN embedding: Where are the good latents? arXiv preprint, arXiv:2012.09036, 2020. 2", + "[57] Peiye Zhuang, Oluwasanmi Koyejo, and Alexander Schwing. Enjoy your editing: Controllable GANs for image editing via latent space navigation. In ICLR, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 318 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22861", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/ee18bf56-6e38-4711-9114-1ec06f896ae1_model.json b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/ee18bf56-6e38-4711-9114-1ec06f896ae1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c9a51e18a7505dddfb81605e0b498f451e01d0fb --- /dev/null +++ b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/ee18bf56-6e38-4711-9114-1ec06f896ae1_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.12, + 0.798, + 0.163 + ], + "angle": 0, + "content": "StyleInV: A Temporal Style Modulated Inversion Network for Unconditional Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.186, + 0.672, + 0.24 + ], + "angle": 0, + "content": "Yuhan Wang, Liming Jiang, Chen Change Loy \nS-Lab, Nanyang Technological University \n{yuhan004, liming002, ccloy}@ntu.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.474, + 0.699 + ], + "angle": 0, + "content": "Unconditional video generation is a challenging task that involves synthesizing high-quality videos that are both coherent and of extended duration. To address this challenge, researchers have used pretrained StyleGAN image generators for high-quality frame synthesis and focused on motion generator design. The motion generator is trained in an autoregressive manner using heavy 3D convolutional discriminators to ensure motion coherence during video generation. In this paper, we introduce a novel motion generator design that uses a learning-based inversion network for GAN. The encoder in our method captures rich and smooth priors from encoding images to latents, and given the latent of an initially generated frame as guidance, our method can generate smooth future latent by modulating the inversion encoder temporally. Our method enjoys the advantage of sparse training and naturally constrains the generation space of our motion generator with the inversion network guided by the initial frame, eliminating the need for heavy discriminators. Moreover, our method supports style transfer with simple fine-tuning when the encoder is paired with a pretrained StyleGAN generator. Extensive experiments conducted on various benchmarks demonstrate the superiority of our method in generating long and high-resolution videos with decent single-frame quality and temporal consistency. Code is available at https://github.com/johannwyh/StyleInV." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.21, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.885 + ], + "angle": 0, + "content": "Unconditional video generation aims at learning a generative model to create novel videos from latent vectors. Despite extensive studies [42, 32, 33, 38, 11, 50] in addressing this problem, it remains challenging to generate high-resolution videos with both favorable quality and motion coherence over a long-term duration. The core difficulties in this task lie in modeling consistent motion and managing the high memory consumption introduced by the addition of the temporal dimension." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To ensure high single-frame resolution and quality, many" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.275, + 0.865, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.416, + 0.846, + 0.43 + ], + "angle": 0, + "content": "(a) Autoregressive Framework (e.g., MoCoGAN-HD)" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.437, + 0.862, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.582, + 0.822, + 0.595 + ], + "angle": 0, + "content": "(b) Non-autoregressive Framework (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.607, + 0.895, + 0.728 + ], + "angle": 0, + "content": "Figure 1: A comparison between autoregressive and non-autoregressive pipeline: (a) Previous autoregressive motion generators require generating the whole clip for a 3D-convolution-based discriminator. (b) Our non-autoregressive motion generator, StyleInV, is an inversion network modulated by temporal style (as a random function of \\( t \\)), which enjoys sparse training using a 2D-convolution-based discriminator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "existing studies, such as MoCoGAN-HD [36], employ a powerful image generator such as StyleGAN [22] as a backbone to serve as a strong generative prior. This approach shifts the focus towards developing a robust motion generator that can capture temporally coherent motion. Most of these methods model motion in an auto-regressive manner, where the next latent is sampled conditioned on the previous one (see Fig. 1). However, this design has two main drawbacks. First, while good performance requires seeing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22851" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.445, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.351, + 0.473, + 0.457 + ], + "angle": 0, + "content": "Figure 2: Inverted latent space visualization and modulated inversion process: When the StyleGAN generator is trained with video frame data, \\(\\mathcal{W}\\) space is well clustered by human identities and provides promising inversion results. Thus, the modulated inversion process can easily find the target latent corresponding to the same identity (shifted to the next motion) as the source one." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.548 + ], + "angle": 0, + "content": "a long sequence of images, the use of heavy 3D discriminators limits its ability to be trained with longer videos. Second, the autoregressive motion generator can lead to motion collapse when extrapolating to generate longer videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.471, + 0.775 + ], + "angle": 0, + "content": "In this study, we present an effective framework for non-autoregressive motion generation that is capable of generating long and high-resolution videos. Our approach leverages learning-based Generative Adversarial Network (GAN) inversion, which learns the inverse mapping of GANs via an inversion network that consists of an encoder and a decoder1. To generate long and coherent videos, we exploit the unique characteristic of the inversion encoder, which captures a rich and smooth manifold between the mapping of images and latent. As illustrated in Fig. 1, to generate a sequence of smooth motion latents, we just need to provide the initial latent code and modulate the inversion encoder with temporal style codes, which are encodings of timestamps with randomness. The motion latents can then be mapped by a StyleGAN decoder to generate a video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.471, + 0.839 + ], + "angle": 0, + "content": "The proposed framework offers several advantages in a single unified framework. First, the use of an inversion network naturally constrains the generation space to stay consistent with the desired appearance, which is de" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.394 + ], + "angle": 0, + "content": "fined by the initial latent code. As demonstrated in Fig. 2, this leads to a significant benefit. Second, thanks to the flexibility of the inversion network in accepting temporal styles of arbitrary timestamps, the framework allows non-autoregressive generation and sparse training [52, 35]. These merits help alleviate the need for heavy discriminators to ensure temporal consistency, as is required in existing approaches. In our implementation, we only need to use a 2D convolutional discriminator instead of a 3D discriminator like MoCoGAN-HD. Third, Unlike existing state-of-the-art methods [52, 35, 7] that couple content and motion decoding in one synthesis network, our framework can naturally support content decoder fine-tuning on different image datasets. Specifically, after fine-tuning the decoder (e.g., StyleGAN2) on another image dataset with the mapping layers and low-resolution synthesis layers fixed, given the same sequence of synthesized motion latents, the generated video can possess the new style of the fine-tuning dataset while preserving the motion patterns of the video generated by the parent content decoder." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.394, + 0.893, + 0.651 + ], + "angle": 0, + "content": "The main contribution of this work is a novel motion generator that modulates a GAN inversion network. This is the first attempt to build such a generator, and it offers several advantages in a unified framework over existing approaches. These advantages include consistent generation, sparse training, and flexibility in supporting style transfer with simple fine-tuning. We additionally contribute a reformulation to the conventional sparse training, through first-frame-aware acyclic positional encoding (FFA-APE) and first-frame-aware sparse training (FFA-ST), to ensure that our motion generator can faithfully reconstruct the initial frame and that the generated video is smooth and continuous. Extensive experiments on DeeperForensics [17], FaceForensics [31], SkyTimelapse [47] and Tai-Chi-HD [34] datasets show that our model is comparable to or even better than state-of-the-art unconditional video generation methods [36, 52, 35] both qualitatively and quantitatively." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.664, + 0.642, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "GAN inversion. The goal of GAN inversion is to find the corresponding vector in the latent space of a pretrained GAN [22, 23] to reconstruct the input image. Existing methods can be classified into three categories [46]: (1) learning-based methods [8, 37, 49, 3, 2, 44, 29], which leverage an encoder network to directly map an image into a latent vector; (2) optimization-based methods [43, 1, 45, 48, 56, 57], which iteratively find the latent vector that best reconstructs the input image using gradient descent; and (3) hybrid models [6, 5, 9, 55], which initialize the iteration process with the result of an encoder network. The design of our motion generator follows the learning-based approach. Therefore, our method is trainable, efficient for single-image inference, and suitable for hierarchical mod-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.471, + 0.901 + ], + "angle": 0, + "content": "1In many contexts, the decoder is a StyleGAN, and the encoder learns to encode a given image to meaningful latent vectors in the StyleGAN space. There is a variety of image manipulation applications [46, 29, 4] developed based upon such an inversion framework." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22852" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "ulation. We devise the motion generator on the \\(\\mathcal{W}\\) space and use the StyleGAN generated latent as the initial content code to guide the modulated inversion process (see Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.14, + 0.471, + 0.624 + ], + "angle": 0, + "content": "Unconditional video generation. Unconditional video generation aims to model the distribution of real videos in a training dataset and generate videos from sampled noise vectors. Many recent studies on this topic are inspired by the success of GANs in image generation. VGAN [42] applies 3D convolutions in both the generator and discriminator, while TGAN [32] optimizes this design by decomposing the generator into an image generator, which is shared by the generation of each frame, and a motion generator2. This framework has been followed by most subsequent studies, such as MoCoGAN [38], which applies a content-motion decomposition. Some approaches [33, 11, 19] have focused on reducing the computational cost of the video discriminator, but the cost is still proportional to the video duration and resolution. Some recent methods have applied more advanced generative frameworks and techniques to unconditional video generation. For example, VideoGPT [50] uses VQ-VAE [28] and GPT [10] to formulate a non-GAN-based video generation approach. Recent studies have also explored unconditional video generation with higher resolution and longer duration. For example, Long-Video-GAN [7] develops a two-phase model that focuses on improving the long-term temporal dynamics of video generation. MoCoGAN-HD [36] and StyleVideoGAN [13] study the generation of latent trajectories in the latent space of a pretrained StyleGAN2 generator. Our approach is inspired by these studies, but differs in the design of the motion generator. Our motion generator is non-autoregressive, thus alleviating the use of heavy discriminators, and it is unique since it obtains the motion latent via modulating a GAN inversion network. This design allows us to attain better motion consistency and semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Recent works [52, 35] explored neural representation-based generators and trained them sparsely as an image GAN. StyleSV [53] improves this framework by introducing StyleGAN3 [21] architecture and several temporal designs. In our work, we extend the idea of sparse training to first-frame-aware sparse training, allowing it to be applied to a generation pipeline conditioned on the initial latent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.842 + ], + "angle": 0, + "content": "Diffusion-based video generation. The diffusion models [15, 30], a new paradigm for image generation tasks, have also achieved significant progress in the task of unconditional video generation [16, 24, 41, 51]. Despite their success, temporal consistency is still an open problem for diffusion models, and GAN-based models exhibit a clear advantage in terms of inference speed." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.888, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.436, + 0.895, + 0.514 + ], + "angle": 0, + "content": "Figure 3: From a typical inversion encoder to StyleInV: We add AdaIN layers at the end of each residual block to inject the temporal style, which is a combination of time positional encoding and the first frame latent code. Here \"A\" stands for a learned affine transform [22]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.534, + 0.637, + 0.552 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.807, + 0.575 + ], + "angle": 0, + "content": "3.1. Preliminaries of Inversion Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.688 + ], + "angle": 0, + "content": "An inversion encoder maps an input image to a vector in the \\(\\mathcal{W}\\) or \\(\\mathcal{W}+\\) latent space of a pretrained StyleGAN2 generator. The generated image that corresponds to this vector should faithfully reconstruct the details of the input image. Therefore, when based on \\(\\mathcal{W}\\) latent space, given an input image \\(\\mathbf{x}\\), the reconstruction process can be defined on top of the inversion network Inv as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.696, + 0.892, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} := G (\\operatorname {I n v} (\\mathbf {x})) := G (E (\\mathbf {x}) + \\overline {{\\mathbf {w}}}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.894, + 0.84 + ], + "angle": 0, + "content": "Here \\(E\\) and \\(G\\) denote the inversion encoder and StyleGAN generator, respectively. \\(\\overline{\\mathbf{w}}\\in \\mathbb{R}^{512}\\) denotes the average latent vector of the generator in the \\(\\mathcal{W}\\) latent space. In our implementation, the encoder \\(E\\) is a convolutional network backbone that outputs a 512-dimensional vector from the last layer embedding, as shown in Fig. 3(left). We build the encoder on the \\(\\mathcal{W}\\) latent space, which eases the design of temporal modulation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.887, + 0.864 + ], + "angle": 0, + "content": "3.2. Temporal Style Modulated Inversion Encoder" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "We observe that the latent space of a StyleGAN trained on a video dataset is typically well-clustered by its content" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "2In the original paper of TGAN [32], the authors called this module temporal generator, which is equivalent to the motion generator used in subsequent studies [38, 36, 52] and in our paper." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22853" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.227 + ], + "angle": 0, + "content": "subject. Figure 2 shows an example of human face videos, where we depict the results of inverting video clips of different identities into the \\(\\mathcal{W}\\) space and visualizing them with t-SNE [40]. It can be observed that the latent space is grouped by human identities. We also observe the same property in video datasets that follow other distributions. This phenomenon suggests that the inversion network inherits some important temporal priors that we could leverage to maintain motion consistency in generated videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.471, + 0.424 + ], + "angle": 0, + "content": "Motivated by this observation, we propose StyleInV, in which the motion latent is generated by modulating a GAN inversion network with temporal styles. Figure 3(right) illustrates the pipeline of our framework. The temporal style \\( s_t \\) of a timestamp \\( t \\) consists of two parts: the motion code \\( v_t \\) and the latent code of the initial frame \\( w_0 \\). Inspired by [35], we use an acyclic positional encoding module to compute a dynamic embedding of the timestamp \\( t \\). However, unlike [35], we make the embedding of the zero timestamp fixed, so this module becomes first-frame-aware. We provide more details in Section 3.3. The latent code \\( w_0 \\) of the initial frame is concatenated with the motion code for content-adaptive affine transform." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.424, + 0.47, + 0.5 + ], + "angle": 0, + "content": "The temporal style is injected into the inversion encoder through AdaIN layers at the end of each convolution block. With this design, the encoder \\( E \\) of StyleInV becomes a function of the initial latent code \\( w_{0} \\) and timestamp \\( t \\). The modulated inversion process can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.51, + 0.47, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {t} := G \\left(\\operatorname {S t y l e I n V} \\left(\\boldsymbol {w} _ {0}, t\\right)\\right) := G \\left(E \\left(G \\left(\\boldsymbol {w} _ {0}\\right), \\boldsymbol {s} _ {t}\\right) + \\boldsymbol {w} _ {0}\\right). \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.469, + 0.58 + ], + "angle": 0, + "content": "Notably, the output of \\( E \\) serves as the residual w.r.t. \\( w_0 \\), instead of \\( \\overline{w} \\). This modification provides more explicit content information guidance for the inversion encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.469, + 0.673 + ], + "angle": 0, + "content": "During training, we first train a raw inversion encoder following Eq. (1) on all video frames. Then, we use this network to initialize the weights of all convolution layers in the StyleInV encoder. Other parameters (e.g., FFA-APE and Affine Transforms) are randomly initialized. Finally, the entire StyleInV encoder is trained end-to-end." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.192, + 0.696 + ], + "angle": 0, + "content": "3.3. FFA-APE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.87 + ], + "angle": 0, + "content": "The original implementation of acyclic positional encoding (APE) [35] samples a series of noise vectors \\( \\boldsymbol{z}_{t_0}^{\\mathsf{m}},\\dots ,\\boldsymbol{z}_{t_n}^{\\mathsf{m}}\\sim \\mathcal{N}(\\mathbf{0},\\boldsymbol {I}) \\) where \\( t_i = i\\cdot \\delta^z \\). We call these temporal points anchor points. Here \\( \\delta^z \\) is a set constant distance between adjacent anchor points. Then, the noise vectors are mapped to tokens \\( \\boldsymbol{u}_{t_0},\\dots ,\\boldsymbol{u}_{t_n} \\) by a paddingless conv1d-based motion mapping network. The computation of the acyclic positional encoding \\( \\boldsymbol{v}_t \\) of arbitrary timestamp \\( t \\) is achieved by a scalable and learnable interpolation between the tokens of two adjacent anchor points that cover \\( t \\). The computation pipeline is shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In our non-autoregressive generation pipeline, the modulated inversion encoder needs to faithfully reconstruct the" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.22, + 0.894, + 0.312 + ], + "angle": 0, + "content": "Figure 4: FFA-APE: A simplified case to compute \\( v_{16} \\) for demonstrating the original acyclic positional encoding and our FFA-APE when \\( \\delta^z = 32 \\) and convld kernel size is 3. In FFA-APE, the encoding of zero timestamp \\( (u_0) \\) only depends on constant paddings and a constant noise vector, thus is fixed for any sampled noise vector sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.893, + 0.566 + ], + "angle": 0, + "content": "initial frame when the input timestamp is zero, making it necessary to fix the computation of the APE for the zero timestamp \\( v_{0} \\). The original APE computation for \\( v_{0} \\) is dynamic and depends on randomly sampled noise vectors, which can lead to dynamic output that is not desired. To address this, we devise a first-frame-aware acyclic positional encoding (FFA-APE) method that fixes \\( v_{0} \\) while maintaining the smoothness of APE (see Fig. 4). We achieve this by replacing the noise vector for the first anchor point with a learnable constant vector \\( c_{0}^{\\mathrm{m}} \\), and using left-sided conv1d layers with constant padding instead of the padding-less conv1d layers. This way, the value of \\( v_{0} \\) only depends on the constant vector \\( c_{0}^{\\mathrm{m}} \\) and the left-padded vectors, which are also constant. As a result, \\( v_{0} \\) is naturally fixed without affecting the continuity of positional encoding." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.6, + 0.59 + ], + "angle": 0, + "content": "3.4. FFA-ST" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.751 + ], + "angle": 0, + "content": "In this section, we introduce the first-frame-aware sparse training specially designed for our framework. Recent non-autoregressive video generation approaches [52, 35] use a discriminator design that only considers \\(k\\) frames \\(\\pmb{x}_{t_1},\\dots ,\\pmb{x}_{t_k}\\) for each video, distinguishing the realness of the input conditioned on the time difference of input frames \\(\\delta_{i} = t_{i + 1} - t_{i}\\). This training scheme is called sparse training. StyleGAN-V [35] has analyzed the choice of \\(k\\) and found that \\(k = 3\\) is ideal for most datasets. The discriminator is defined as \\(D(\\pmb{x}_{t_{1,2,3}},\\delta_{1,2})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We follow this training scheme to make full use of our non-autoregressive framework. Nonetheless, using only three randomly sampled timestamps to train the generator and discriminator can result in sharp transitions at the beginning of the generated video, where the generated \\( \\pmb{x}_0 \\) and \\( \\pmb{x}_1 \\) usually diverge too much, and sometimes even switch to another identity and never return. This happens because although we define the generation process of a video as a modulated inversion process of the start frame, the discriminator is unaware of it. The discriminator only focuses on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22854" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "the smoothness of generated latent trajectories, failing to ensure the motion generator produces frames that share the identity with the start frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.197 + ], + "angle": 0, + "content": "To solve this problem, we introduce the initial frame into the discriminator to enhance content consistency and motion smoothness. The adversarial loss for the first-frame-aware discriminator (FFA-D) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.207, + 0.469, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {y} _ {t _ {0, 1, 2, 3}} = G \\left(\\text {S t y l e I n V} \\left(w _ {0}, t _ {0, 1, 2, 3}\\right)\\right), \\\\ \\mathcal {L} _ {a d v} = \\mathbb {E} _ {\\boldsymbol {x} \\sim p _ {v}} \\left[ \\log D \\left(\\boldsymbol {x} _ {t _ {0, 1, 2, 3}}, \\delta_ {0, 1, 2}\\right) \\right] \\tag {3} \\\\ + \\mathbb {E} _ {w _ {0} \\sim p _ {\\mathcal {W}}} \\left[ \\log (1 - D (\\boldsymbol {y} _ {t _ {0, 1, 2, 3}}, \\delta_ {0, 1, 2}) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.469, + 0.336 + ], + "angle": 0, + "content": "where we specify \\( t_0 = 0 \\). Here, \\( p_v \\) and \\( p_{\\mathcal{W}} \\) denote the real data distribution and \\( \\mathcal{W} \\) latent space distribution, respectively. To explicitly enforce initial frame reconstruction, we use a \\( L_2 \\) loss for the generated \\( y_{t_0} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.346, + 0.469, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L _ {2}} = \\left\\| G \\left(w _ {0}\\right) - G \\left(\\text {S t y l e I n V} \\left(w _ {0}, 0\\right)\\right) \\right\\| _ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.373, + 0.469, + 0.403 + ], + "angle": 0, + "content": "Finally, we apply latent regularization [29, 26] to the encoder's output, so as to enhance content consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.413, + 0.469, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\sum_ {i = 0} ^ {3} | | E (G (w _ {0}), t _ {i}) | | _ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.449, + 0.469, + 0.478 + ], + "angle": 0, + "content": "The overall loss function for training our motion generator and the discriminator is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.49, + 0.469, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\min _ {E} \\max _ {D} \\mathcal {L} _ {a d v} + \\min _ {E} \\left(\\lambda_ {L _ {2}} \\mathcal {L} _ {L _ {2}} + \\lambda_ {r e g} \\mathcal {L} _ {r e g}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.469, + 0.582 + ], + "angle": 0, + "content": "Here \\(\\lambda_{L_2}\\) and \\(\\lambda_{reg}\\) are the loss hyperparameters. We also apply discriminator adaptive augmentation [20, 35] and \\(r1\\) regularization [22, 35] to further improve the training stability and generation quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.362, + 0.607 + ], + "angle": 0, + "content": "3.5. Finetuning-based Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.734 + ], + "angle": 0, + "content": "Our 'inversion encoder+decoder' framework can naturally take a pretrained StyleGAN model as the generator. And such a configuration allows the generator to be finetuned for different styles, and yet still able to use the motion generator for generating new video with styles. The capability is not possible with existing non-autoregressive video generation methods [52, 35] because they cannot be finetuned under an image GAN training scheme." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To achieve style transfer, as illustrated in Fig. 5, we fine-tune the pre-trained StyleGAN model using an image dataset, such as MegaCartoon [27], while keeping the mapping network and low-resolution \\((\\leqslant 32^2\\), coarse and middle layers in [22]) synthesis blocks fixed. This configuration maintains the distribution of the \\(\\mathcal{W}\\) space during finetuning. To improve identity preservation and reduce artifacts, we apply both a perceptual loss [18] and an identity loss [12] between the images generated by the original and fine-tuned StyleGAN. We show some visual results in Fig. 8. The style-transferred video maintains the" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.092, + 0.891, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.305, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Figure 5: Finetuning-based style transfer: our framework allows easy fine-tuning of decoder (a pretrained StyleGAN generator) to a new domain by freezing mapping and low-resolution \\((\\leqslant 32^{2})\\) layers. Standard identity loss and perceptual loss are applied to improve identity preservation and reduce artifacts. The style-transferred videos can then be generated by incorporating the StyleInV motion generator." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.892, + 0.558 + ], + "angle": 0, + "content": "same motion pattern as the video generated by the parent model, while adopting a new style from the fine-tuning image dataset. It is noteworthy that the finetuning process is independent of the video generation training. It means that the finetuning-based style transfer is \"plug-and-play\" as the fine-tuned image generator can be used on any StyleInV models. There is no additional computational cost in the inference either." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.633, + 0.59 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.599, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Datasets. We use four video datasets in our main experiments: DeeperForensics \\(256^{2}\\) [17], FaceForensics \\(256^{2}\\) [31], SkyTimelapse \\(256^{2}\\) [47] and TaiChi \\(256^{2}\\) [34]. The cropping strategy for DeeperForensics [17] and FaceForensics [31] is different. For DeeperForensics, we use a stabilized FFHQ [22] cropping strategy [25], while we follow the strategy firstly adopted by TGAN-V2 [33] for FaceForensics. Please refer to the supplementary material for a detailed discussion." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Baselines. We explore four state-of-the-art methods for comparison: MoCoGAN-HD [36], DIGAN [52], StyleGAN-V [35] and Long-Video-GAN [7]. Among these methods, MoCoGAN-HD and DIGAN require an explicit setting of the training clip length. We follow the default setting of their paper to set the clip length as 16 for both methods. This setting is identical to StyleGAN-V [35]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addition, on DeeperForensics, we explore an optimized setting on DIGAN and MoCoGAN-HD for a more fair comparison. For DIGAN, we directly increase the clip length to 128 frames. For MoCoGAN-HD, we apply the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22855" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "first-frame-aware sparse training to train its motion generator, so as to avoid using a heavy 3D discriminator, allowing it to be trained with 128-frame clips as other methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.289 + ], + "angle": 0, + "content": "Evaluation. We use Fréchet Inception Distance (FID) [14] and Fréchet Video Distance (FVD) [39] to evaluate all models quantitatively. In practice, we follow the metric calculation framework provided by StyleGAN-V [35] to first generate a fake video dataset with 2,048 synthesized clips, each of 128 frames. For FID, we sample 50k frames from real and fake video datasets to compute the result. For FVD, we compute \\(\\mathrm{FVD}_{16}\\) and \\(\\mathrm{FVD}_{128}\\) with the first 16 frames and all 128 frames of each clip, respectively. We use FID results to show the single-frame image quality of each method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.468, + 0.382 + ], + "angle": 0, + "content": "To ensure a fair comparison, we re-benchmark the quantitative results of every method on every dataset. We retrain all the baselines using the official paper setting, except for MoCoGAN-HD on SkyTimelapse, where an officially released checkpoint is available. For more implementation details, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.392, + 0.216, + 0.406 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.468, + 0.506 + ], + "angle": 0, + "content": "Quantitative results. Table 1 summarizes the quantitative results of our method compared to other baselines. Our method achieves competitive quantitative results on all the benchmarks. Notably, although MoCoGAN-HD and DIGAN are trained with clips of 16 frames, we still outperform them in terms of \\(\\mathrm{FVD}_{16}\\) metrics on all four datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Qualitative results. Figure 6 shows the qualitative comparison between our method and the baselines on all four datasets. MoCoGAN-HD and DIGAN both suffer from motion collapse, resulting in a degraded generation quality over time. StyleGAN-V shows an impressive visual performance on FaceForensics and SkyTimelapse, but it sometimes fails to maintain the identity and accessories on DeeperForensics and lacks diversity and magnitude of motion over a long time span on TaiChi (the subject gradually fixes at one state). Long-Video-GAN is exceptionally good at SkyTimelapse, but it cannot achieve similar performance on other datasets. It fails to maintain the identity on DeeperForensics, and its single-frame content on TaiChi lacks details and is inferior to other methods. The generated videos by Long-Video-GAN collapse on FaceForensics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.902 + ], + "angle": 0, + "content": "In contrast to existing methods, our method demonstrates stable results on all four datasets, particularly with superior identity preservation on human-face video and long-term generation quality on TaiChi. Although our method outperforms existing methods in terms of content quality, continuity, and quantitative results, the motion semantics of our generated videos on SkyTimelapse are inferior to those on other datasets. This could be one of the limitations of our work and an area for future improvement. Extended experiments. We present more in-depth comparisons in Table 2 and Fig. 7 by introducing training im" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.15 + ], + "angle": 0, + "content": "Table 1: FID, \\(\\mathrm{FVD}_{16}\\) and \\(\\mathrm{FVD}_{128}\\) results of video generation methods on (a) DeeperForensics \\(256^{2}\\), (b) FaceForensics \\(256^{2}\\), (c) TaiChi \\(256^{2}\\), and (d) SkyTimelapse \\(256^{2}\\). **Bolds** indicate best and **underline** indicate the second best." + }, + { + "type": "table_caption", + "bbox": [ + 0.618, + 0.157, + 0.774, + 0.172 + ], + "angle": 0, + "content": "(a) DeeperForensics \\(256^{2}\\)" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.178, + 0.878, + 0.279 + ], + "angle": 0, + "content": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD135.30101.07610.30
DIGAN191.9946.691060.27
StyleGAN-V59.5939.3368.81
Long-Video-GAN56.5474.77169.45
StyleInV (ours)54.0541.5853.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.627, + 0.282, + 0.766, + 0.295 + ], + "angle": 0, + "content": "(b) FaceForensics \\(256^{2}\\)" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.303, + 0.878, + 0.403 + ], + "angle": 0, + "content": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD24.45112.67486.69
DIGAN151.53146.621993.20
StyleGAN-V8.6452.92108.86
Long-Video-GAN40.40233.26567.78
StyleInV (ours)12.0647.88103.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.649, + 0.405, + 0.744, + 0.42 + ], + "angle": 0, + "content": "(c) TaiChi \\(256^{2}\\)" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.427, + 0.878, + 0.527 + ], + "angle": 0, + "content": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD73.61315.03622.95
DIGAN67.24196.77954.93
StyleGAN-V35.68254.74477.78
Long-Video-GAN43.90248.55502.65
StyleInV (ours)41.55185.72328.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.627, + 0.53, + 0.766, + 0.545 + ], + "angle": 0, + "content": "(d) SkyTimelapse \\(256^{2}\\)" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.551, + 0.878, + 0.651 + ], + "angle": 0, + "content": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD251.81696.584116.03
DIGAN32.83148.08269.43
StyleGAN-V16.9581.32197.83
Long-Video-GAN25.41116.50152.70
StyleInV (ours)14.3277.04194.25
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.665, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Table 2: FID, \\(\\mathrm{FVD}_{16}\\) and \\(\\mathrm{FVD}_{128}\\) results of extended experiments on DeeperForensics \\(256^{2}\\). We apply sparse training to MoCoGAN-HD [36] (#1) and change the preset clip length of DIGAN [52] to 128 (#2). **Bolds** indicate best. (-) indicates a smaller (better) quantitative result, while (+) indicates a larger (worse) one, compared with Table 1a." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.764, + 0.889, + 0.83 + ], + "angle": 0, + "content": "
#MethodFID (↓)FVD16 (↓)FVD128 (↓)
1[36] + Sparse Training55.84 (-)54.58 (-)129.13 (-)
2[52] + Clip 12874.80 (-)87.42 (+)95.80 (-)
3StyleInV (ours)54.0541.5853.93
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "provements to baselines. Increasing the clip length generally improves the results of MoCoGAN-HD and DIGAN, but they are still inferior to our method. Notably, training" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22856" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.476, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.295, + 0.358, + 0.308 + ], + "angle": 0, + "content": "(a) DeeperForensics \\(256^{2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.089, + 0.885, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.294, + 0.761, + 0.308 + ], + "angle": 0, + "content": "(b) FaceForensics \\(256^{2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.31, + 0.476, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.514, + 0.328, + 0.527 + ], + "angle": 0, + "content": "(c) TaiChi \\(256^{2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.31, + 0.885, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.514, + 0.759, + 0.528 + ], + "angle": 0, + "content": "(d) SkyTimelapse \\(256^{2}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.54, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Figure 6: Uncurated samples from the existing methods on DeeperForensics \\(256^{2}\\), FaceForensics \\(256^{2}\\), TaiChi \\(256^{2}\\) and SkyTimelapse \\(256^{2}\\), respectively. We sample a 128-frame video and display every 16 frames, starting from \\(t = 0\\)." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.587, + 0.48, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.587, + 0.89, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.719, + 0.89, + 0.75 + ], + "angle": 0, + "content": "Figure 7: Qualitative comparison of extended experiments. “M+ST” and “D-C128” correspond to Table 2 (#1) and (#2), respectively. Each row shows the first and last 128 frames of a 2056-frame (68.5s) video, displayed every 16 frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.776, + 0.47, + 0.897 + ], + "angle": 0, + "content": "with longer clips harms the short-term \\(\\mathrm{FVD}_{16}\\) result of DIGAN, which indicates its tradeoff between duration length and local temporal quality. Qualitatively, for both methods, the generated content is evidently improved within 128 frames, although the sparsely trained MoCoGAN-HD exhibits issues with identity switching. Motion collapse is still observed when MoCoGAN-HD and DIGAN generates long videos. In contrast, our method can stably generate ex" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.776, + 0.892, + 0.822 + ], + "angle": 0, + "content": "tremely long videos without motion collapse. Our method outperforms the sparsely trained MoCoGAN-HD, demonstrating the superiority of our motion generator design." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.616, + 0.848 + ], + "angle": 0, + "content": "4.2. Properties" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As discussed in Section 3.5, our method has the unique advantage over state-of-the-art methods, such as StyleGAN-V and Long-Video-GAN, on its high compati" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22857" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.493, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.092, + 0.888, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.253, + 0.892, + 0.285 + ], + "angle": 0, + "content": "Figure 8: Finetuning-based style transfer result. The 1st row is generated by the parent model trained on CelebV-HQ [54]. The 2nd, 3rd, and 4th row uses the StyleGAN generator fintuned on Cartoon, Arcane, and MetFace, respectively." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.307, + 0.468, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.609 + ], + "angle": 0, + "content": "Figure 9: Initial-frame conditioned generation and style transferred results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.425, + 0.656 + ], + "angle": 0, + "content": "bility with StyleGAN-based downstream techniques." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Finetuning-based style transfer. We train the parent model (motion generator and StyleGAN) on CelebV-HQ [54], as its rich identity makes it more suitable for transfer learning. To perform style transfer, we fine-tune the StyleGAN on the Cartoon [27], MetFace [20], and Arcane datasets following the procedure outlined in Section 3.5. In Fig. 8, we show examples where the same StyleInV-generated latent sequence is decoded by different but aligned StyleGAN generators. Our method achieves satisfactory results in terms of smooth video style transfer with well-aligned face structure, identity, and expression, demonstrating its desirable properties and potential for various applications. More results can be found in the supplementary video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Initial-frame conditioned generation. Our network supports generating a series of content given a real-world im" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.309, + 0.88, + 0.324 + ], + "angle": 0, + "content": "Table 3: Ablation result on the DeeperForensics dataset." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.336, + 0.89, + 0.446 + ], + "angle": 0, + "content": "
#MethodFID (↓)FVD16 (↓)FVD128 (↓)
1w/o inversion encoder54.3559.49152.82
2w/o FFA-APE55.2688.98144.52
3w/o Eq.(4) & Eq.(3)52.5567.4358.88
4w/o Eq.(3)53.9586.3259.76
5Ours54.0541.5853.93
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.623 + ], + "angle": 0, + "content": "age as the initial frame. We first inverse the image into the StyleGAN2 latent space with a pSp [29] encoder, which is trained to initialize the weights of StyleInV. We treat it as the 512-dimensional initial frame latent \\( w_{0} \\), then use it to generate a video with our StyleInV. The generated latent sequence can be also applied to a finetuned image generator to synthesize a style-transferred animation video. Through this pipeline, the real image is reconstructed twice, the first time is during the inversion process, while the second time is when synthesizing \\( G(\\text{StyleInV}(w_{0},0)) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.893, + 0.819 + ], + "angle": 0, + "content": "When the real images are sampled from the training dataset (see Fig. 9 first two rows), \\( G(\\text{StyleInV}(w_0, 0)) \\) can faithfully reconstruct the raw image and generate high-quality videos. We then test the generation quality for real images sampled out of the training set (see Fig. 9 last two rows, where we select Benedict Cumberbatch and Anne Hathaway). We use the StyleGAN2 generator and StyleInV model trained on CelebV-HQ [54] dataset as it is richer in its identities. The results show that our StyleInV network can still generate meaningful videos while reconstructing the initial frame decently, and the style transfer results are smooth and well-aligned. Please refer to the supplementary material for more results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.665, + 0.846 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Motion generator design. We explore two alternative motion generator designs. The first is the autoregressive MoCoGAN-HD design, which has been discussed in Sec-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22858" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.188, + 0.468, + 0.249 + ], + "angle": 0, + "content": "Figure 10: Ablating FFA-APE. Generate the first frame with a fixed \\( w_0 \\) but different temporal noise sequences. The model without FFA-APE fails to reconstruct the initial frame and generates the first frame with randomness." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.268, + 0.467, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.365, + 0.468, + 0.41 + ], + "angle": 0, + "content": "Figure 11: Ablating FFA-ST. The \\( L_{2} \\) loss ensures the initial frame reconstruction. But without the initial frame included in the discriminator, we still cannot preserve the identity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.468, + 0.534 + ], + "angle": 0, + "content": "tion 4.1. For the second design, we remove all Convs, AdaIN and affine transform layers in Fig. 3 and let the mapped temporal style be the output of inversion encoder, i.e., the residual w.r.t. initial frame latent \\(w_0\\). It largely harms identity preservation. Both \\(\\mathrm{FVD}_{16}\\) and \\(\\mathrm{FVD}_{128}\\) degrade significantly as is shown in Table 3(#1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.468, + 0.702 + ], + "angle": 0, + "content": "FFA-APE. We evaluate the importance of our first-frame-aware acyclic positional encoding (FFA-APE) module by replacing it with the original design proposed by [35]. As shown in Fig. 10, the dynamic embedding of the zero timestamp prevents the network from faithfully reconstructing the initial frame. In contrast, our full method can stably realize reconstruction. In addition, the L2 loss in Eq. (4) fails to converge for the ablation method, and its gradient further harms the learning of the positional encoding module, leading to a much worse quantitative result shown in Table 3 (#2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.9 + ], + "angle": 0, + "content": "FFA-ST. We conduct two ablation experiments for the FFA-ST modules. In the first experiment, we remove the initial frame from the discriminator and remove the reconstruction loss (Eq.(4)) (Table 3(#3)). In the second experiment, we only remove the initial frame from the discriminator while keeping the reconstruction loss (Table 3(#4)). As shown in Fig. 11, without the reconstruction loss, our model cannot reconstruct the initial frame accurately. With the \\( L_{2} \\) loss, the initial frame is reconstructed, but there is a sudden transition between the first two frames, and sometimes, the identity also changes, leading to a worse \\( \\mathrm{FVD}_{16} \\) result. These experiments demonstrate the importance of our first-frame-aware discriminator (FFA-D)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.619, + 0.105 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.297 + ], + "angle": 0, + "content": "We have presented a novel approach for unconditional video generation by employing a pretrained StyleGAN image generator. The proposed StyleInV motion generator generates latents in the StyleGAN2 latent space by modulating a learning-based inversion network, and thus capable of inheriting its informative priors of the initial latent. Our network features non-autoregressive training and uniquely supports fine-tuning based style transfer. Extensive experiments demonstrate the superiority of our method in generating long and high-resolution videos, outperforming state-of-the-art baselines. Here we also briefly discuss our limitations and broader impacts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.308, + 0.625, + 0.322 + ], + "angle": 0, + "content": "5.1. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.892, + 0.421 + ], + "angle": 0, + "content": "Inferior motion semantics on SkyTimelapse. Our motion semantics on SkyTimelapse [47] are inferior to those on other datasets. This could be due to different dataset characteristics, as videos in SkyTimelapse are not subject-centric and typically driven by global motions, which does not align perfectly with our model nature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.892, + 0.542 + ], + "angle": 0, + "content": "The impact of dataset identity richness. When the scale of facial identities in the video dataset is too small, the effects of inversion, editing, and style transfer are constrained. Image generation quality. The generation quality of StyleGAN determines the performance upper bound of our method. The images generated by the StyleGAN2 models have artifacts in the background on SkyTimelapse [47], and lack fine details and a sense of structure on TaiChi [34]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.618 + ], + "angle": 0, + "content": "Model training. Our approach is two-stage, requiring 7.5 and 9 GPU days each, which is more than the 8 GPU days of StyleGAN-V [35]. Despite this, StyleInV is as efficient when finetuning the hyperparameters of the video generator, since the image generator only needs to be trained once." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.629, + 0.667, + 0.645 + ], + "angle": 0, + "content": "5.2. Broader Impacts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.728 + ], + "angle": 0, + "content": "We believe that the potential of StyleInV can be further exploited. Our method can provide a natural solution towards mega-pixel level video generation and StyleGAN-based editing, and it might in return promote the research of learning-based GAN inversion methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.892, + 0.803 + ], + "angle": 0, + "content": "As for the negative side, StyleInV may ease the synthesis of better-quality fake videos with threats. We believe that it can be alleviated by developing more advanced falsified media detection methods or contributing larger-scale and higher-quality forgery detection datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.804, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2022-01-030). It is also supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). We thank Shuai Yang for his help in this work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22859" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. StyleFlow: Attribute-conditioned exploration of StyleGAN-generated images using conditional continuous normalizing flows. ACM TOG, 40:1-21, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: Age transformation using a style-based regression model. ACM TOG, 40:1-12, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.468, + 0.257 + ], + "angle": 0, + "content": "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. ReStyle: A residual-based StyleGAN encoder via iterative refinement. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit H Bermano. HyperStyle: StyleGAN inversion with hypernetworks for real image editing. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.468, + 0.356 + ], + "angle": 0, + "content": "[5] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. ACM TOG, 38:1-11, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.468, + 0.4 + ], + "angle": 0, + "content": "[6] David Bau, Jun-Yan Zhu, Jonas Wulff, William Peebles, Hendrik Strobelt, Bolei Zhou, and Antonio Torralba. Seeing what a GAN cannot generate. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.468, + 0.456 + ], + "angle": 0, + "content": "[7] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurlPS, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.468, + 0.499 + ], + "angle": 0, + "content": "[8] Lucy Chai, Jonas Wulff, and Phillip Isola. Using latent space regression to analyze and leverage compositionality in GANs. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.468, + 0.542 + ], + "angle": 0, + "content": "[9] Lucy Chai, Jun-Yan Zhu, Eli Shechtman, Phillip Isola, and Richard Zhang. Ensembling with deep generative views. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.468, + 0.586 + ], + "angle": 0, + "content": "[10] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.468, + 0.628 + ], + "angle": 0, + "content": "[11] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv preprint, arXiv:1907.06571, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.468, + 0.671 + ], + "angle": 0, + "content": "[12] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.674, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[13] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. StyleVideoGAN: A temporal generative model using a pretrained StyleGAN. In BMVC, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.468, + 0.801 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[17] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. DeeperForensics-1.0: A large-scale dataset for real-world face forgery detection. In CVPR, 2020. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[19] Emmanuel Kahembwe and Subramanian Ramamoorthy. Lower dimensional kernels for video discriminators. Neural Networks, 132:506-520, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[20] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[24] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[25] Jacek Naruniec, Leonhard Helminger, Christopher Schroers, and Romann M Weber. High-resolution neural face swapping for visual effects. In CGF, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[26] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: Text-driven manipulation of StyleGAN imagery. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[27] Justin NM Pinkney and Doron Adler. Resolution dependent GAN interpolation for controllable image synthesis between domains. arXiv preprint, arXiv:2010.05334, 2020. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[28] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[29] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In CVPR, 2021. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[31] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint, arXiv:1803.09179, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[32] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In ICCV, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[33] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal GAN. IJCV, 128:2586-2606, 2020. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In NeurIPS, 2019. 2, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[35] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. StyleGAN-V: A continuous video generator with the" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22860" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "price, image quality and perks of StyleGAN2. In CVPR, 2022. 2, 3, 4, 5, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[36] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In ICLR, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[37] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. ACM TOG, 40:1-14, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[38] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In CVPR, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[39] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint, arXiv:1812.01717, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[40] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. JMLR, 9(11), 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[41] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[42] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. In NeurIPS, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[43] Hui-Po Wang, Ning Yu, and Mario Fritz. Hijack-GAN: Unintended-use of pretrained, black-box GANs. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[44] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Lu Yuan, Gang Hua, and Nenghai Yu. E2Style: Improve the efficiency and effectiveness of StyleGAN inversion. TIP, 31:3267-3280, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[45] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: Disentangled controls for StyleGAN image generation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[46] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. GAN inversion: A survey. TPAMI, 45:3121-3138, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[47] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In CVPR, 2018. 2, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[48] Yangyang Xu, Yong Du, Wenpeng Xiao, Xuemiao Xu, and Shengfeng He. From continuity to editability: Inverting GANs with consecutive images. In ICCV, pages 13910-13918, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[49] Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, and Bolei Zhou. Generative hierarchical features from synthesizing images. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[50] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video generation using VQ-VAE and Transformers. arXiv preprint, arXiv:2104.10157, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[51] Siyuan Yang, Lu Zhang, Yu Liu, Zhizhuo Jiang, and You He. Video diffusion models with local-global context guidance. In IJCAI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[52] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In ICLR, 2022. 2, 3, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[53] Qihang Zhang, Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Towards smooth video composition. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[55] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain GAN inversion for real image editing. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[56] Peihao Zhu, Rameen Abdal, Yipeng Qin, John Femiani, and Peter Wonka. Improved StyleGAN embedding: Where are the good latents? arXiv preprint, arXiv:2012.09036, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[57] Peiye Zhuang, Oluwasanmi Koyejo, and Alexander Schwing. Enjoy your editing: Controllable GANs for image editing via latent space navigation. In ICLR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22861" + } + ] +] \ No newline at end of file diff --git a/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/ee18bf56-6e38-4711-9114-1ec06f896ae1_origin.pdf b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/ee18bf56-6e38-4711-9114-1ec06f896ae1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6516d20a3b4323990c71fb6df507e54941523861 --- /dev/null +++ b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/ee18bf56-6e38-4711-9114-1ec06f896ae1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec010d715337039b706c0aa853a875ea545eff0cb6dd16e11f0269086de15d53 +size 1871264 diff --git a/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/full.md b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f0f8c5a99df231b0d8911c12c58dcbc080eba0c2 --- /dev/null +++ b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/full.md @@ -0,0 +1,332 @@ +# StyleInV: A Temporal Style Modulated Inversion Network for Unconditional Video Generation + +Yuhan Wang, Liming Jiang, Chen Change Loy +S-Lab, Nanyang Technological University +{yuhan004, liming002, ccloy}@ntu.edu.sg + +# Abstract + +Unconditional video generation is a challenging task that involves synthesizing high-quality videos that are both coherent and of extended duration. To address this challenge, researchers have used pretrained StyleGAN image generators for high-quality frame synthesis and focused on motion generator design. The motion generator is trained in an autoregressive manner using heavy 3D convolutional discriminators to ensure motion coherence during video generation. In this paper, we introduce a novel motion generator design that uses a learning-based inversion network for GAN. The encoder in our method captures rich and smooth priors from encoding images to latents, and given the latent of an initially generated frame as guidance, our method can generate smooth future latent by modulating the inversion encoder temporally. Our method enjoys the advantage of sparse training and naturally constrains the generation space of our motion generator with the inversion network guided by the initial frame, eliminating the need for heavy discriminators. Moreover, our method supports style transfer with simple fine-tuning when the encoder is paired with a pretrained StyleGAN generator. Extensive experiments conducted on various benchmarks demonstrate the superiority of our method in generating long and high-resolution videos with decent single-frame quality and temporal consistency. Code is available at https://github.com/johannwyh/StyleInV. + +# 1. Introduction + +Unconditional video generation aims at learning a generative model to create novel videos from latent vectors. Despite extensive studies [42, 32, 33, 38, 11, 50] in addressing this problem, it remains challenging to generate high-resolution videos with both favorable quality and motion coherence over a long-term duration. The core difficulties in this task lie in modeling consistent motion and managing the high memory consumption introduced by the addition of the temporal dimension. + +To ensure high single-frame resolution and quality, many + +![](images/277a5ac1fb41b9074a00430b0e4b98da665a53bd83c976dfde04a4d851567ecc.jpg) +(a) Autoregressive Framework (e.g., MoCoGAN-HD) + +![](images/4f1dae7e75cc5c72e52c718ff9325e806423d2651a326c0f784991ffd34f9157.jpg) +(b) Non-autoregressive Framework (Ours) +Figure 1: A comparison between autoregressive and non-autoregressive pipeline: (a) Previous autoregressive motion generators require generating the whole clip for a 3D-convolution-based discriminator. (b) Our non-autoregressive motion generator, StyleInV, is an inversion network modulated by temporal style (as a random function of $t$ ), which enjoys sparse training using a 2D-convolution-based discriminator. + +existing studies, such as MoCoGAN-HD [36], employ a powerful image generator such as StyleGAN [22] as a backbone to serve as a strong generative prior. This approach shifts the focus towards developing a robust motion generator that can capture temporally coherent motion. Most of these methods model motion in an auto-regressive manner, where the next latent is sampled conditioned on the previous one (see Fig. 1). However, this design has two main drawbacks. First, while good performance requires seeing + +![](images/b14e68c39c107fbc4ea11e36709f5f1f825d7a28b201b9499b6f00d3e38f1705.jpg) +Figure 2: Inverted latent space visualization and modulated inversion process: When the StyleGAN generator is trained with video frame data, $\mathcal{W}$ space is well clustered by human identities and provides promising inversion results. Thus, the modulated inversion process can easily find the target latent corresponding to the same identity (shifted to the next motion) as the source one. + +a long sequence of images, the use of heavy 3D discriminators limits its ability to be trained with longer videos. Second, the autoregressive motion generator can lead to motion collapse when extrapolating to generate longer videos. + +In this study, we present an effective framework for non-autoregressive motion generation that is capable of generating long and high-resolution videos. Our approach leverages learning-based Generative Adversarial Network (GAN) inversion, which learns the inverse mapping of GANs via an inversion network that consists of an encoder and a decoder1. To generate long and coherent videos, we exploit the unique characteristic of the inversion encoder, which captures a rich and smooth manifold between the mapping of images and latent. As illustrated in Fig. 1, to generate a sequence of smooth motion latents, we just need to provide the initial latent code and modulate the inversion encoder with temporal style codes, which are encodings of timestamps with randomness. The motion latents can then be mapped by a StyleGAN decoder to generate a video. + +The proposed framework offers several advantages in a single unified framework. First, the use of an inversion network naturally constrains the generation space to stay consistent with the desired appearance, which is de + +fined by the initial latent code. As demonstrated in Fig. 2, this leads to a significant benefit. Second, thanks to the flexibility of the inversion network in accepting temporal styles of arbitrary timestamps, the framework allows non-autoregressive generation and sparse training [52, 35]. These merits help alleviate the need for heavy discriminators to ensure temporal consistency, as is required in existing approaches. In our implementation, we only need to use a 2D convolutional discriminator instead of a 3D discriminator like MoCoGAN-HD. Third, Unlike existing state-of-the-art methods [52, 35, 7] that couple content and motion decoding in one synthesis network, our framework can naturally support content decoder fine-tuning on different image datasets. Specifically, after fine-tuning the decoder (e.g., StyleGAN2) on another image dataset with the mapping layers and low-resolution synthesis layers fixed, given the same sequence of synthesized motion latents, the generated video can possess the new style of the fine-tuning dataset while preserving the motion patterns of the video generated by the parent content decoder. + +The main contribution of this work is a novel motion generator that modulates a GAN inversion network. This is the first attempt to build such a generator, and it offers several advantages in a unified framework over existing approaches. These advantages include consistent generation, sparse training, and flexibility in supporting style transfer with simple fine-tuning. We additionally contribute a reformulation to the conventional sparse training, through first-frame-aware acyclic positional encoding (FFA-APE) and first-frame-aware sparse training (FFA-ST), to ensure that our motion generator can faithfully reconstruct the initial frame and that the generated video is smooth and continuous. Extensive experiments on DeeperForensics [17], FaceForensics [31], SkyTimelapse [47] and Tai-Chi-HD [34] datasets show that our model is comparable to or even better than state-of-the-art unconditional video generation methods [36, 52, 35] both qualitatively and quantitatively. + +# 2. Related Work + +GAN inversion. The goal of GAN inversion is to find the corresponding vector in the latent space of a pretrained GAN [22, 23] to reconstruct the input image. Existing methods can be classified into three categories [46]: (1) learning-based methods [8, 37, 49, 3, 2, 44, 29], which leverage an encoder network to directly map an image into a latent vector; (2) optimization-based methods [43, 1, 45, 48, 56, 57], which iteratively find the latent vector that best reconstructs the input image using gradient descent; and (3) hybrid models [6, 5, 9, 55], which initialize the iteration process with the result of an encoder network. The design of our motion generator follows the learning-based approach. Therefore, our method is trainable, efficient for single-image inference, and suitable for hierarchical mod- + +ulation. We devise the motion generator on the $\mathcal{W}$ space and use the StyleGAN generated latent as the initial content code to guide the modulated inversion process (see Fig. 2). + +Unconditional video generation. Unconditional video generation aims to model the distribution of real videos in a training dataset and generate videos from sampled noise vectors. Many recent studies on this topic are inspired by the success of GANs in image generation. VGAN [42] applies 3D convolutions in both the generator and discriminator, while TGAN [32] optimizes this design by decomposing the generator into an image generator, which is shared by the generation of each frame, and a motion generator2. This framework has been followed by most subsequent studies, such as MoCoGAN [38], which applies a content-motion decomposition. Some approaches [33, 11, 19] have focused on reducing the computational cost of the video discriminator, but the cost is still proportional to the video duration and resolution. Some recent methods have applied more advanced generative frameworks and techniques to unconditional video generation. For example, VideoGPT [50] uses VQ-VAE [28] and GPT [10] to formulate a non-GAN-based video generation approach. Recent studies have also explored unconditional video generation with higher resolution and longer duration. For example, Long-Video-GAN [7] develops a two-phase model that focuses on improving the long-term temporal dynamics of video generation. MoCoGAN-HD [36] and StyleVideoGAN [13] study the generation of latent trajectories in the latent space of a pretrained StyleGAN2 generator. Our approach is inspired by these studies, but differs in the design of the motion generator. Our motion generator is non-autoregressive, thus alleviating the use of heavy discriminators, and it is unique since it obtains the motion latent via modulating a GAN inversion network. This design allows us to attain better motion consistency and semantics. + +Recent works [52, 35] explored neural representation-based generators and trained them sparsely as an image GAN. StyleSV [53] improves this framework by introducing StyleGAN3 [21] architecture and several temporal designs. In our work, we extend the idea of sparse training to first-frame-aware sparse training, allowing it to be applied to a generation pipeline conditioned on the initial latent. + +Diffusion-based video generation. The diffusion models [15, 30], a new paradigm for image generation tasks, have also achieved significant progress in the task of unconditional video generation [16, 24, 41, 51]. Despite their success, temporal consistency is still an open problem for diffusion models, and GAN-based models exhibit a clear advantage in terms of inference speed. + +![](images/2cf7ca05a5f0e6b5c34507ab42726cc8feeacf35c1acadb12f52db5aff4cea89.jpg) +Figure 3: From a typical inversion encoder to StyleInV: We add AdaIN layers at the end of each residual block to inject the temporal style, which is a combination of time positional encoding and the first frame latent code. Here "A" stands for a learned affine transform [22]. + +# 3. Methodology + +# 3.1. Preliminaries of Inversion Encoder + +An inversion encoder maps an input image to a vector in the $\mathcal{W}$ or $\mathcal{W}+$ latent space of a pretrained StyleGAN2 generator. The generated image that corresponds to this vector should faithfully reconstruct the details of the input image. Therefore, when based on $\mathcal{W}$ latent space, given an input image $\mathbf{x}$ , the reconstruction process can be defined on top of the inversion network Inv as: + +$$ +\hat {\mathbf {x}} := G (\operatorname {I n v} (\mathbf {x})) := G (E (\mathbf {x}) + \overline {{\mathbf {w}}}). \tag {1} +$$ + +Here $E$ and $G$ denote the inversion encoder and StyleGAN generator, respectively. $\overline{\mathbf{w}}\in \mathbb{R}^{512}$ denotes the average latent vector of the generator in the $\mathcal{W}$ latent space. In our implementation, the encoder $E$ is a convolutional network backbone that outputs a 512-dimensional vector from the last layer embedding, as shown in Fig. 3(left). We build the encoder on the $\mathcal{W}$ latent space, which eases the design of temporal modulation. + +# 3.2. Temporal Style Modulated Inversion Encoder + +We observe that the latent space of a StyleGAN trained on a video dataset is typically well-clustered by its content + +subject. Figure 2 shows an example of human face videos, where we depict the results of inverting video clips of different identities into the $\mathcal{W}$ space and visualizing them with t-SNE [40]. It can be observed that the latent space is grouped by human identities. We also observe the same property in video datasets that follow other distributions. This phenomenon suggests that the inversion network inherits some important temporal priors that we could leverage to maintain motion consistency in generated videos. + +Motivated by this observation, we propose StyleInV, in which the motion latent is generated by modulating a GAN inversion network with temporal styles. Figure 3(right) illustrates the pipeline of our framework. The temporal style $s_t$ of a timestamp $t$ consists of two parts: the motion code $v_t$ and the latent code of the initial frame $w_0$ . Inspired by [35], we use an acyclic positional encoding module to compute a dynamic embedding of the timestamp $t$ . However, unlike [35], we make the embedding of the zero timestamp fixed, so this module becomes first-frame-aware. We provide more details in Section 3.3. The latent code $w_0$ of the initial frame is concatenated with the motion code for content-adaptive affine transform. + +The temporal style is injected into the inversion encoder through AdaIN layers at the end of each convolution block. With this design, the encoder $E$ of StyleInV becomes a function of the initial latent code $w_{0}$ and timestamp $t$ . The modulated inversion process can be defined as: + +$$ +\hat {\mathbf {x}} _ {t} := G \left(\operatorname {S t y l e I n V} \left(\boldsymbol {w} _ {0}, t\right)\right) := G \left(E \left(G \left(\boldsymbol {w} _ {0}\right), \boldsymbol {s} _ {t}\right) + \boldsymbol {w} _ {0}\right). \quad (2) +$$ + +Notably, the output of $E$ serves as the residual w.r.t. $w_0$ , instead of $\overline{w}$ . This modification provides more explicit content information guidance for the inversion encoder. + +During training, we first train a raw inversion encoder following Eq. (1) on all video frames. Then, we use this network to initialize the weights of all convolution layers in the StyleInV encoder. Other parameters (e.g., FFA-APE and Affine Transforms) are randomly initialized. Finally, the entire StyleInV encoder is trained end-to-end. + +# 3.3. FFA-APE + +The original implementation of acyclic positional encoding (APE) [35] samples a series of noise vectors $\boldsymbol{z}_{t_0}^{\mathsf{m}},\dots ,\boldsymbol{z}_{t_n}^{\mathsf{m}}\sim \mathcal{N}(\mathbf{0},\boldsymbol {I})$ where $t_i = i\cdot \delta^z$ . We call these temporal points anchor points. Here $\delta^z$ is a set constant distance between adjacent anchor points. Then, the noise vectors are mapped to tokens $\boldsymbol{u}_{t_0},\dots ,\boldsymbol{u}_{t_n}$ by a paddingless conv1d-based motion mapping network. The computation of the acyclic positional encoding $\boldsymbol{v}_t$ of arbitrary timestamp $t$ is achieved by a scalable and learnable interpolation between the tokens of two adjacent anchor points that cover $t$ . The computation pipeline is shown in Fig. 4. + +In our non-autoregressive generation pipeline, the modulated inversion encoder needs to faithfully reconstruct the + +![](images/a8a92dc3843daf5f14c6b568af60c48db73ba8826ba843afb7b689d0728245fc.jpg) +Figure 4: FFA-APE: A simplified case to compute $v_{16}$ for demonstrating the original acyclic positional encoding and our FFA-APE when $\delta^z = 32$ and convld kernel size is 3. In FFA-APE, the encoding of zero timestamp $(u_0)$ only depends on constant paddings and a constant noise vector, thus is fixed for any sampled noise vector sequence. + +initial frame when the input timestamp is zero, making it necessary to fix the computation of the APE for the zero timestamp $v_{0}$ . The original APE computation for $v_{0}$ is dynamic and depends on randomly sampled noise vectors, which can lead to dynamic output that is not desired. To address this, we devise a first-frame-aware acyclic positional encoding (FFA-APE) method that fixes $v_{0}$ while maintaining the smoothness of APE (see Fig. 4). We achieve this by replacing the noise vector for the first anchor point with a learnable constant vector $c_{0}^{\mathrm{m}}$ , and using left-sided conv1d layers with constant padding instead of the padding-less conv1d layers. This way, the value of $v_{0}$ only depends on the constant vector $c_{0}^{\mathrm{m}}$ and the left-padded vectors, which are also constant. As a result, $v_{0}$ is naturally fixed without affecting the continuity of positional encoding. + +# 3.4. FFA-ST + +In this section, we introduce the first-frame-aware sparse training specially designed for our framework. Recent non-autoregressive video generation approaches [52, 35] use a discriminator design that only considers $k$ frames $\pmb{x}_{t_1},\dots ,\pmb{x}_{t_k}$ for each video, distinguishing the realness of the input conditioned on the time difference of input frames $\delta_{i} = t_{i + 1} - t_{i}$ . This training scheme is called sparse training. StyleGAN-V [35] has analyzed the choice of $k$ and found that $k = 3$ is ideal for most datasets. The discriminator is defined as $D(\pmb{x}_{t_{1,2,3}},\delta_{1,2})$ . + +We follow this training scheme to make full use of our non-autoregressive framework. Nonetheless, using only three randomly sampled timestamps to train the generator and discriminator can result in sharp transitions at the beginning of the generated video, where the generated $\pmb{x}_0$ and $\pmb{x}_1$ usually diverge too much, and sometimes even switch to another identity and never return. This happens because although we define the generation process of a video as a modulated inversion process of the start frame, the discriminator is unaware of it. The discriminator only focuses on + +the smoothness of generated latent trajectories, failing to ensure the motion generator produces frames that share the identity with the start frame. + +To solve this problem, we introduce the initial frame into the discriminator to enhance content consistency and motion smoothness. The adversarial loss for the first-frame-aware discriminator (FFA-D) can be written as: + +$$ +\begin{array}{l} \mathbf {y} _ {t _ {0, 1, 2, 3}} = G \left(\text {S t y l e I n V} \left(w _ {0}, t _ {0, 1, 2, 3}\right)\right), \\ \mathcal {L} _ {a d v} = \mathbb {E} _ {\boldsymbol {x} \sim p _ {v}} \left[ \log D \left(\boldsymbol {x} _ {t _ {0, 1, 2, 3}}, \delta_ {0, 1, 2}\right) \right] \tag {3} \\ + \mathbb {E} _ {w _ {0} \sim p _ {\mathcal {W}}} \left[ \log (1 - D (\boldsymbol {y} _ {t _ {0, 1, 2, 3}}, \delta_ {0, 1, 2}) \right], \\ \end{array} +$$ + +where we specify $t_0 = 0$ . Here, $p_v$ and $p_{\mathcal{W}}$ denote the real data distribution and $\mathcal{W}$ latent space distribution, respectively. To explicitly enforce initial frame reconstruction, we use a $L_2$ loss for the generated $y_{t_0}$ : + +$$ +\mathcal {L} _ {L _ {2}} = \left\| G \left(w _ {0}\right) - G \left(\text {S t y l e I n V} \left(w _ {0}, 0\right)\right) \right\| _ {2}. \tag {4} +$$ + +Finally, we apply latent regularization [29, 26] to the encoder's output, so as to enhance content consistency: + +$$ +\mathcal {L} _ {r e g} = \sum_ {i = 0} ^ {3} | | E (G (w _ {0}), t _ {i}) | | _ {2}. \tag {5} +$$ + +The overall loss function for training our motion generator and the discriminator is defined as: + +$$ +\min _ {E} \max _ {D} \mathcal {L} _ {a d v} + \min _ {E} \left(\lambda_ {L _ {2}} \mathcal {L} _ {L _ {2}} + \lambda_ {r e g} \mathcal {L} _ {r e g}\right). \tag {6} +$$ + +Here $\lambda_{L_2}$ and $\lambda_{reg}$ are the loss hyperparameters. We also apply discriminator adaptive augmentation [20, 35] and $r1$ regularization [22, 35] to further improve the training stability and generation quality. + +# 3.5. Finetuning-based Style Transfer + +Our 'inversion encoder+decoder' framework can naturally take a pretrained StyleGAN model as the generator. And such a configuration allows the generator to be finetuned for different styles, and yet still able to use the motion generator for generating new video with styles. The capability is not possible with existing non-autoregressive video generation methods [52, 35] because they cannot be finetuned under an image GAN training scheme. + +To achieve style transfer, as illustrated in Fig. 5, we fine-tune the pre-trained StyleGAN model using an image dataset, such as MegaCartoon [27], while keeping the mapping network and low-resolution $(\leqslant 32^2$ , coarse and middle layers in [22]) synthesis blocks fixed. This configuration maintains the distribution of the $\mathcal{W}$ space during finetuning. To improve identity preservation and reduce artifacts, we apply both a perceptual loss [18] and an identity loss [12] between the images generated by the original and fine-tuned StyleGAN. We show some visual results in Fig. 8. The style-transferred video maintains the + +![](images/555f9ea1ab613a2e8e335f4d9eb0bc8af194509ffc3946ec7326c60b3e51f314.jpg) +Figure 5: Finetuning-based style transfer: our framework allows easy fine-tuning of decoder (a pretrained StyleGAN generator) to a new domain by freezing mapping and low-resolution $(\leqslant 32^{2})$ layers. Standard identity loss and perceptual loss are applied to improve identity preservation and reduce artifacts. The style-transferred videos can then be generated by incorporating the StyleInV motion generator. + +same motion pattern as the video generated by the parent model, while adopting a new style from the fine-tuning image dataset. It is noteworthy that the finetuning process is independent of the video generation training. It means that the finetuning-based style transfer is "plug-and-play" as the fine-tuned image generator can be used on any StyleInV models. There is no additional computational cost in the inference either. + +# 4. Experiments + +Datasets. We use four video datasets in our main experiments: DeeperForensics $256^{2}$ [17], FaceForensics $256^{2}$ [31], SkyTimelapse $256^{2}$ [47] and TaiChi $256^{2}$ [34]. The cropping strategy for DeeperForensics [17] and FaceForensics [31] is different. For DeeperForensics, we use a stabilized FFHQ [22] cropping strategy [25], while we follow the strategy firstly adopted by TGAN-V2 [33] for FaceForensics. Please refer to the supplementary material for a detailed discussion. + +Baselines. We explore four state-of-the-art methods for comparison: MoCoGAN-HD [36], DIGAN [52], StyleGAN-V [35] and Long-Video-GAN [7]. Among these methods, MoCoGAN-HD and DIGAN require an explicit setting of the training clip length. We follow the default setting of their paper to set the clip length as 16 for both methods. This setting is identical to StyleGAN-V [35]. + +In addition, on DeeperForensics, we explore an optimized setting on DIGAN and MoCoGAN-HD for a more fair comparison. For DIGAN, we directly increase the clip length to 128 frames. For MoCoGAN-HD, we apply the + +first-frame-aware sparse training to train its motion generator, so as to avoid using a heavy 3D discriminator, allowing it to be trained with 128-frame clips as other methods. + +Evaluation. We use Fréchet Inception Distance (FID) [14] and Fréchet Video Distance (FVD) [39] to evaluate all models quantitatively. In practice, we follow the metric calculation framework provided by StyleGAN-V [35] to first generate a fake video dataset with 2,048 synthesized clips, each of 128 frames. For FID, we sample 50k frames from real and fake video datasets to compute the result. For FVD, we compute $\mathrm{FVD}_{16}$ and $\mathrm{FVD}_{128}$ with the first 16 frames and all 128 frames of each clip, respectively. We use FID results to show the single-frame image quality of each method. + +To ensure a fair comparison, we re-benchmark the quantitative results of every method on every dataset. We retrain all the baselines using the official paper setting, except for MoCoGAN-HD on SkyTimelapse, where an officially released checkpoint is available. For more implementation details, please refer to the supplementary material. + +# 4.1. Main Results + +Quantitative results. Table 1 summarizes the quantitative results of our method compared to other baselines. Our method achieves competitive quantitative results on all the benchmarks. Notably, although MoCoGAN-HD and DIGAN are trained with clips of 16 frames, we still outperform them in terms of $\mathrm{FVD}_{16}$ metrics on all four datasets. + +Qualitative results. Figure 6 shows the qualitative comparison between our method and the baselines on all four datasets. MoCoGAN-HD and DIGAN both suffer from motion collapse, resulting in a degraded generation quality over time. StyleGAN-V shows an impressive visual performance on FaceForensics and SkyTimelapse, but it sometimes fails to maintain the identity and accessories on DeeperForensics and lacks diversity and magnitude of motion over a long time span on TaiChi (the subject gradually fixes at one state). Long-Video-GAN is exceptionally good at SkyTimelapse, but it cannot achieve similar performance on other datasets. It fails to maintain the identity on DeeperForensics, and its single-frame content on TaiChi lacks details and is inferior to other methods. The generated videos by Long-Video-GAN collapse on FaceForensics. + +In contrast to existing methods, our method demonstrates stable results on all four datasets, particularly with superior identity preservation on human-face video and long-term generation quality on TaiChi. Although our method outperforms existing methods in terms of content quality, continuity, and quantitative results, the motion semantics of our generated videos on SkyTimelapse are inferior to those on other datasets. This could be one of the limitations of our work and an area for future improvement. Extended experiments. We present more in-depth comparisons in Table 2 and Fig. 7 by introducing training im + +Table 1: FID, $\mathrm{FVD}_{16}$ and $\mathrm{FVD}_{128}$ results of video generation methods on (a) DeeperForensics $256^{2}$ , (b) FaceForensics $256^{2}$ , (c) TaiChi $256^{2}$ , and (d) SkyTimelapse $256^{2}$ . **Bolds** indicate best and **underline** indicate the second best. + +(a) DeeperForensics $256^{2}$ + +
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD135.30101.07610.30
DIGAN191.9946.691060.27
StyleGAN-V59.5939.3368.81
Long-Video-GAN56.5474.77169.45
StyleInV (ours)54.0541.5853.93
+ +(b) FaceForensics $256^{2}$ + +
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD24.45112.67486.69
DIGAN151.53146.621993.20
StyleGAN-V8.6452.92108.86
Long-Video-GAN40.40233.26567.78
StyleInV (ours)12.0647.88103.63
+ +(c) TaiChi $256^{2}$ + +
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD73.61315.03622.95
DIGAN67.24196.77954.93
StyleGAN-V35.68254.74477.78
Long-Video-GAN43.90248.55502.65
StyleInV (ours)41.55185.72328.90
+ +(d) SkyTimelapse $256^{2}$ + +
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD251.81696.584116.03
DIGAN32.83148.08269.43
StyleGAN-V16.9581.32197.83
Long-Video-GAN25.41116.50152.70
StyleInV (ours)14.3277.04194.25
+ +Table 2: FID, $\mathrm{FVD}_{16}$ and $\mathrm{FVD}_{128}$ results of extended experiments on DeeperForensics $256^{2}$ . We apply sparse training to MoCoGAN-HD [36] (#1) and change the preset clip length of DIGAN [52] to 128 (#2). **Bolds** indicate best. (-) indicates a smaller (better) quantitative result, while (+) indicates a larger (worse) one, compared with Table 1a. + +
#MethodFID (↓)FVD16 (↓)FVD128 (↓)
1[36] + Sparse Training55.84 (-)54.58 (-)129.13 (-)
2[52] + Clip 12874.80 (-)87.42 (+)95.80 (-)
3StyleInV (ours)54.0541.5853.93
+ +provements to baselines. Increasing the clip length generally improves the results of MoCoGAN-HD and DIGAN, but they are still inferior to our method. Notably, training + +![](images/56532472b55d24987e00ab4a2a7ea17bcf8e592b0c366abea6ba417212b5770b.jpg) +(a) DeeperForensics $256^{2}$ + +![](images/aa0537a460bfc8d84f3504c1cd466626fb3b59f20d7125d0aa03596485022036.jpg) +(b) FaceForensics $256^{2}$ + +![](images/43ec2d0d21d64d2c460706b674201465506293d1307c2c8012a1adbfe48e9723.jpg) +(c) TaiChi $256^{2}$ + +![](images/534c12848b3af10c60bc89e47d58713f7b31efbb9efafdd5d49edf17dab4a01c.jpg) +(d) SkyTimelapse $256^{2}$ + +![](images/ea122f05fbb5a5bfe1ae4aea506e0b126b977781c803f94e5b79e2384b369130.jpg) +Figure 6: Uncurated samples from the existing methods on DeeperForensics $256^{2}$ , FaceForensics $256^{2}$ , TaiChi $256^{2}$ and SkyTimelapse $256^{2}$ , respectively. We sample a 128-frame video and display every 16 frames, starting from $t = 0$ . + +![](images/93fd0e7287387afd330d39731987c3ce921e7bcac9a1a197ccde13494cb39b12.jpg) +Figure 7: Qualitative comparison of extended experiments. “M+ST” and “D-C128” correspond to Table 2 (#1) and (#2), respectively. Each row shows the first and last 128 frames of a 2056-frame (68.5s) video, displayed every 16 frames. + +with longer clips harms the short-term $\mathrm{FVD}_{16}$ result of DIGAN, which indicates its tradeoff between duration length and local temporal quality. Qualitatively, for both methods, the generated content is evidently improved within 128 frames, although the sparsely trained MoCoGAN-HD exhibits issues with identity switching. Motion collapse is still observed when MoCoGAN-HD and DIGAN generates long videos. In contrast, our method can stably generate ex + +tremely long videos without motion collapse. Our method outperforms the sparsely trained MoCoGAN-HD, demonstrating the superiority of our motion generator design. + +# 4.2. Properties + +As discussed in Section 3.5, our method has the unique advantage over state-of-the-art methods, such as StyleGAN-V and Long-Video-GAN, on its high compati + +![](images/fffce52942f24b84dde75e8bad4ec7faad5f6fb379d3d890f26be0c5e61ee3e0.jpg) +Figure 8: Finetuning-based style transfer result. The 1st row is generated by the parent model trained on CelebV-HQ [54]. The 2nd, 3rd, and 4th row uses the StyleGAN generator fintuned on Cartoon, Arcane, and MetFace, respectively. + +![](images/ecdfb14ff0a42adaca4cec5a778eea248f2b918b169ecf1fa81d11a57d08d8b0.jpg) + +![](images/f09d88d1e8513870048be291257c1dd67a2f7f0d77367991c2b0ded16e7b516e.jpg) +Figure 9: Initial-frame conditioned generation and style transferred results. + +bility with StyleGAN-based downstream techniques. + +Finetuning-based style transfer. We train the parent model (motion generator and StyleGAN) on CelebV-HQ [54], as its rich identity makes it more suitable for transfer learning. To perform style transfer, we fine-tune the StyleGAN on the Cartoon [27], MetFace [20], and Arcane datasets following the procedure outlined in Section 3.5. In Fig. 8, we show examples where the same StyleInV-generated latent sequence is decoded by different but aligned StyleGAN generators. Our method achieves satisfactory results in terms of smooth video style transfer with well-aligned face structure, identity, and expression, demonstrating its desirable properties and potential for various applications. More results can be found in the supplementary video. + +Initial-frame conditioned generation. Our network supports generating a series of content given a real-world im + +Table 3: Ablation result on the DeeperForensics dataset. + +
#MethodFID (↓)FVD16 (↓)FVD128 (↓)
1w/o inversion encoder54.3559.49152.82
2w/o FFA-APE55.2688.98144.52
3w/o Eq.(4) & Eq.(3)52.5567.4358.88
4w/o Eq.(3)53.9586.3259.76
5Ours54.0541.5853.93
+ +age as the initial frame. We first inverse the image into the StyleGAN2 latent space with a pSp [29] encoder, which is trained to initialize the weights of StyleInV. We treat it as the 512-dimensional initial frame latent $w_{0}$ , then use it to generate a video with our StyleInV. The generated latent sequence can be also applied to a finetuned image generator to synthesize a style-transferred animation video. Through this pipeline, the real image is reconstructed twice, the first time is during the inversion process, while the second time is when synthesizing $G(\text{StyleInV}(w_{0},0))$ . + +When the real images are sampled from the training dataset (see Fig. 9 first two rows), $G(\text{StyleInV}(w_0, 0))$ can faithfully reconstruct the raw image and generate high-quality videos. We then test the generation quality for real images sampled out of the training set (see Fig. 9 last two rows, where we select Benedict Cumberbatch and Anne Hathaway). We use the StyleGAN2 generator and StyleInV model trained on CelebV-HQ [54] dataset as it is richer in its identities. The results show that our StyleInV network can still generate meaningful videos while reconstructing the initial frame decently, and the style transfer results are smooth and well-aligned. Please refer to the supplementary material for more results. + +# 4.3. Ablation Studies + +Motion generator design. We explore two alternative motion generator designs. The first is the autoregressive MoCoGAN-HD design, which has been discussed in Sec- + +![](images/5305fab939e19ec3b5aafaad54263943881abfb05bcfe9219646bba506746dc1.jpg) +Figure 10: Ablating FFA-APE. Generate the first frame with a fixed $w_0$ but different temporal noise sequences. The model without FFA-APE fails to reconstruct the initial frame and generates the first frame with randomness. + +![](images/b7dd93cd4ea4567fc4badbf1360ab4d02d8d8ebe048a67c9cf13adabe7d94316.jpg) +Figure 11: Ablating FFA-ST. The $L_{2}$ loss ensures the initial frame reconstruction. But without the initial frame included in the discriminator, we still cannot preserve the identity. + +tion 4.1. For the second design, we remove all Convs, AdaIN and affine transform layers in Fig. 3 and let the mapped temporal style be the output of inversion encoder, i.e., the residual w.r.t. initial frame latent $w_0$ . It largely harms identity preservation. Both $\mathrm{FVD}_{16}$ and $\mathrm{FVD}_{128}$ degrade significantly as is shown in Table 3(#1). + +FFA-APE. We evaluate the importance of our first-frame-aware acyclic positional encoding (FFA-APE) module by replacing it with the original design proposed by [35]. As shown in Fig. 10, the dynamic embedding of the zero timestamp prevents the network from faithfully reconstructing the initial frame. In contrast, our full method can stably realize reconstruction. In addition, the L2 loss in Eq. (4) fails to converge for the ablation method, and its gradient further harms the learning of the positional encoding module, leading to a much worse quantitative result shown in Table 3 (#2). + +FFA-ST. We conduct two ablation experiments for the FFA-ST modules. In the first experiment, we remove the initial frame from the discriminator and remove the reconstruction loss (Eq.(4)) (Table 3(#3)). In the second experiment, we only remove the initial frame from the discriminator while keeping the reconstruction loss (Table 3(#4)). As shown in Fig. 11, without the reconstruction loss, our model cannot reconstruct the initial frame accurately. With the $L_{2}$ loss, the initial frame is reconstructed, but there is a sudden transition between the first two frames, and sometimes, the identity also changes, leading to a worse $\mathrm{FVD}_{16}$ result. These experiments demonstrate the importance of our first-frame-aware discriminator (FFA-D). + +# 5. Conclusion + +We have presented a novel approach for unconditional video generation by employing a pretrained StyleGAN image generator. The proposed StyleInV motion generator generates latents in the StyleGAN2 latent space by modulating a learning-based inversion network, and thus capable of inheriting its informative priors of the initial latent. Our network features non-autoregressive training and uniquely supports fine-tuning based style transfer. Extensive experiments demonstrate the superiority of our method in generating long and high-resolution videos, outperforming state-of-the-art baselines. Here we also briefly discuss our limitations and broader impacts. + +# 5.1. Limitations + +Inferior motion semantics on SkyTimelapse. Our motion semantics on SkyTimelapse [47] are inferior to those on other datasets. This could be due to different dataset characteristics, as videos in SkyTimelapse are not subject-centric and typically driven by global motions, which does not align perfectly with our model nature. + +The impact of dataset identity richness. When the scale of facial identities in the video dataset is too small, the effects of inversion, editing, and style transfer are constrained. Image generation quality. The generation quality of StyleGAN determines the performance upper bound of our method. The images generated by the StyleGAN2 models have artifacts in the background on SkyTimelapse [47], and lack fine details and a sense of structure on TaiChi [34]. + +Model training. Our approach is two-stage, requiring 7.5 and 9 GPU days each, which is more than the 8 GPU days of StyleGAN-V [35]. Despite this, StyleInV is as efficient when finetuning the hyperparameters of the video generator, since the image generator only needs to be trained once. + +# 5.2. Broader Impacts + +We believe that the potential of StyleInV can be further exploited. Our method can provide a natural solution towards mega-pixel level video generation and StyleGAN-based editing, and it might in return promote the research of learning-based GAN inversion methods. + +As for the negative side, StyleInV may ease the synthesis of better-quality fake videos with threats. We believe that it can be alleviated by developing more advanced falsified media detection methods or contributing larger-scale and higher-quality forgery detection datasets. + +Acknowledgement. This work is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2022-01-030). It is also supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). We thank Shuai Yang for his help in this work. + +# References + +[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. StyleFlow: Attribute-conditioned exploration of StyleGAN-generated images using conditional continuous normalizing flows. ACM TOG, 40:1-21, 2021. 2 +[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: Age transformation using a style-based regression model. ACM TOG, 40:1-12, 2021. 2 +[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. ReStyle: A residual-based StyleGAN encoder via iterative refinement. In ICCV, 2021. 2 +[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit H Bermano. HyperStyle: StyleGAN inversion with hypernetworks for real image editing. In CVPR, 2022. 2 +[5] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. ACM TOG, 38:1-11, 2019. 2 +[6] David Bau, Jun-Yan Zhu, Jonas Wulff, William Peebles, Hendrik Strobelt, Bolei Zhou, and Antonio Torralba. Seeing what a GAN cannot generate. In ICCV, 2019. 2 +[7] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurlPS, 2022. 2, 3, 5 +[8] Lucy Chai, Jonas Wulff, and Phillip Isola. Using latent space regression to analyze and leverage compositionality in GANs. In ICLR, 2021. 2 +[9] Lucy Chai, Jun-Yan Zhu, Eli Shechtman, Phillip Isola, and Richard Zhang. Ensembling with deep generative views. In CVPR, 2021. 2 +[10] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 3 +[11] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv preprint, arXiv:1907.06571, 2019. 1, 3 +[12] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5 +[13] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. StyleVideoGAN: A temporal generative model using a pretrained StyleGAN. In BMVC, 2021. 3 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In NeurIPS, 2017. 6 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 +[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 3 +[17] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. DeeperForensics-1.0: A large-scale dataset for real-world face forgery detection. In CVPR, 2020. 2, 5 + +[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 5 +[19] Emmanuel Kahembwe and Subramanian Ramamoorthy. Lower dimensional kernels for video discriminators. Neural Networks, 132:506-520, 2020. 3 +[20] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 5, 8 +[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 3 +[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 3, 5 +[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2 +[24] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 3 +[25] Jacek Naruniec, Leonhard Helminger, Christopher Schroers, and Romann M Weber. High-resolution neural face swapping for visual effects. In CGF, 2020. 5 +[26] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: Text-driven manipulation of StyleGAN imagery. In ICCV, 2021. 5 +[27] Justin NM Pinkney and Doron Adler. Resolution dependent GAN interpolation for controllable image synthesis between domains. arXiv preprint, arXiv:2010.05334, 2020. 5, 8 +[28] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 3 +[29] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In CVPR, 2021. 2, 5, 8 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 +[31] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint, arXiv:1803.09179, 2018. 2, 5 +[32] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In ICCV, 2017. 1, 3 +[33] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal GAN. IJCV, 128:2586-2606, 2020. 1, 3, 5 +[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In NeurIPS, 2019. 2, 5, 9 +[35] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. StyleGAN-V: A continuous video generator with the + +price, image quality and perks of StyleGAN2. In CVPR, 2022. 2, 3, 4, 5, 6, 9 +[36] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In ICLR, 2021. 1, 2, 3, 5, 6 +[37] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. ACM TOG, 40:1-14, 2021. 2 +[38] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In CVPR, 2018. 1, 3 +[39] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint, arXiv:1812.01717, 2018. 6 +[40] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. JMLR, 9(11), 2008. 4 +[41] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. 2022. 3 +[42] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. In NeurIPS, 2016. 1, 3 +[43] Hui-Po Wang, Ning Yu, and Mario Fritz. Hijack-GAN: Unintended-use of pretrained, black-box GANs. In CVPR, 2021. 2 +[44] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Lu Yuan, Gang Hua, and Nenghai Yu. E2Style: Improve the efficiency and effectiveness of StyleGAN inversion. TIP, 31:3267-3280, 2022. 2 +[45] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: Disentangled controls for StyleGAN image generation. In CVPR, 2021. 2 +[46] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. GAN inversion: A survey. TPAMI, 45:3121-3138, 2022. 2 +[47] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In CVPR, 2018. 2, 5, 9 +[48] Yangyang Xu, Yong Du, Wenpeng Xiao, Xuemiao Xu, and Shengfeng He. From continuity to editability: Inverting GANs with consecutive images. In ICCV, pages 13910-13918, 2021. 2 +[49] Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, and Bolei Zhou. Generative hierarchical features from synthesizing images. In CVPR, 2021. 2 +[50] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video generation using VQ-VAE and Transformers. arXiv preprint, arXiv:2104.10157, 2021. 1, 3 +[51] Siyuan Yang, Lu Zhang, Yu Liu, Zhizhuo Jiang, and You He. Video diffusion models with local-global context guidance. In IJCAI, 2023. 3 +[52] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In ICLR, 2022. 2, 3, 4, 5, 6 + +[53] Qihang Zhang, Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Towards smooth video composition. In ICLR, 2023. 3 +[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022.8 +[55] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain GAN inversion for real image editing. In ECCV, 2020. 2 +[56] Peihao Zhu, Rameen Abdal, Yipeng Qin, John Femiani, and Peter Wonka. Improved StyleGAN embedding: Where are the good latents? arXiv preprint, arXiv:2012.09036, 2020. 2 +[57] Peiye Zhuang, Oluwasanmi Koyejo, and Alexander Schwing. Enjoy your editing: Controllable GANs for image editing via latent space navigation. In ICLR, 2021. 2 \ No newline at end of file diff --git a/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/images.zip b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e421e9c05f5183412f89506ccaddd7ea39aa3a11 --- /dev/null +++ b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:927692e81739fef0ed35724abfaea1ce7df9418aec5cb633fee2361884617531 +size 929167 diff --git a/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/layout.json b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d681226819cb407f7a0d4cae603df65e22e28b79 --- /dev/null +++ b/2023/StyleInV_ A Temporal Style Modulated Inversion Network for Unconditional Video Generation/layout.json @@ -0,0 +1,9381 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 106, + 95, + 488, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 95, + 488, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 95, + 488, + 129 + ], + "type": "text", + "content": "StyleInV: A Temporal Style Modulated Inversion Network for Unconditional Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 183, + 147, + 411, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 147, + 411, + 190 + ], + "spans": [ + { + "bbox": [ + 183, + 147, + 411, + 190 + ], + "type": "text", + "content": "Yuhan Wang, Liming Jiang, Chen Change Loy \nS-Lab, Nanyang Technological University \n{yuhan004, liming002, ccloy}@ntu.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 242, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 290, + 553 + ], + "type": "text", + "content": "Unconditional video generation is a challenging task that involves synthesizing high-quality videos that are both coherent and of extended duration. To address this challenge, researchers have used pretrained StyleGAN image generators for high-quality frame synthesis and focused on motion generator design. The motion generator is trained in an autoregressive manner using heavy 3D convolutional discriminators to ensure motion coherence during video generation. In this paper, we introduce a novel motion generator design that uses a learning-based inversion network for GAN. The encoder in our method captures rich and smooth priors from encoding images to latents, and given the latent of an initially generated frame as guidance, our method can generate smooth future latent by modulating the inversion encoder temporally. Our method enjoys the advantage of sparse training and naturally constrains the generation space of our motion generator with the inversion network guided by the initial frame, eliminating the need for heavy discriminators. Moreover, our method supports style transfer with simple fine-tuning when the encoder is paired with a pretrained StyleGAN generator. Extensive experiments conducted on various benchmarks demonstrate the superiority of our method in generating long and high-resolution videos with decent single-frame quality and temporal consistency. Code is available at https://github.com/johannwyh/StyleInV." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 700 + ], + "type": "text", + "content": "Unconditional video generation aims at learning a generative model to create novel videos from latent vectors. Despite extensive studies [42, 32, 33, 38, 11, 50] in addressing this problem, it remains challenging to generate high-resolution videos with both favorable quality and motion coherence over a long-term duration. The core difficulties in this task lie in modeling consistent motion and managing the high memory consumption introduced by the addition of the temporal dimension." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "To ensure high single-frame resolution and quality, many" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 217, + 529, + 323 + ], + "blocks": [ + { + "bbox": [ + 323, + 217, + 529, + 323 + ], + "lines": [ + { + "bbox": [ + 323, + 217, + 529, + 323 + ], + "spans": [ + { + "bbox": [ + 323, + 217, + 529, + 323 + ], + "type": "image", + "image_path": "277a5ac1fb41b9074a00430b0e4b98da665a53bd83c976dfde04a4d851567ecc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 329, + 517, + 340 + ], + "lines": [ + { + "bbox": [ + 333, + 329, + 517, + 340 + ], + "spans": [ + { + "bbox": [ + 333, + 329, + 517, + 340 + ], + "type": "text", + "content": "(a) Autoregressive Framework (e.g., MoCoGAN-HD)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 346, + 527, + 455 + ], + "blocks": [ + { + "bbox": [ + 328, + 346, + 527, + 455 + ], + "lines": [ + { + "bbox": [ + 328, + 346, + 527, + 455 + ], + "spans": [ + { + "bbox": [ + 328, + 346, + 527, + 455 + ], + "type": "image", + "image_path": "4f1dae7e75cc5c72e52c718ff9325e806423d2651a326c0f784991ffd34f9157.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 460, + 503, + 471 + ], + "lines": [ + { + "bbox": [ + 349, + 460, + 503, + 471 + ], + "spans": [ + { + "bbox": [ + 349, + 460, + 503, + 471 + ], + "type": "text", + "content": "(b) Non-autoregressive Framework (Ours)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "lines": [ + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "content": "Figure 1: A comparison between autoregressive and non-autoregressive pipeline: (a) Previous autoregressive motion generators require generating the whole clip for a 3D-convolution-based discriminator. (b) Our non-autoregressive motion generator, StyleInV, is an inversion network modulated by temporal style (as a random function of " + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "content": "), which enjoys sparse training using a 2D-convolution-based discriminator." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "existing studies, such as MoCoGAN-HD [36], employ a powerful image generator such as StyleGAN [22] as a backbone to serve as a strong generative prior. This approach shifts the focus towards developing a robust motion generator that can capture temporally coherent motion. Most of these methods model motion in an auto-regressive manner, where the next latent is sampled conditioned on the previous one (see Fig. 1). However, this design has two main drawbacks. First, while good performance requires seeing" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22851" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 272, + 266 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 272, + 266 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 272, + 266 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 272, + 266 + ], + "type": "image", + "image_path": "b14e68c39c107fbc4ea11e36709f5f1f825d7a28b201b9499b6f00d3e38f1705.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 289, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 361 + ], + "type": "text", + "content": "Figure 2: Inverted latent space visualization and modulated inversion process: When the StyleGAN generator is trained with video frame data, " + }, + { + "bbox": [ + 46, + 277, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 277, + 289, + 361 + ], + "type": "text", + "content": " space is well clustered by human identities and provides promising inversion results. Thus, the modulated inversion process can easily find the target latent corresponding to the same identity (shifted to the next motion) as the source one." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 434 + ], + "type": "text", + "content": "a long sequence of images, the use of heavy 3D discriminators limits its ability to be trained with longer videos. Second, the autoregressive motion generator can lead to motion collapse when extrapolating to generate longer videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 434, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 288, + 613 + ], + "type": "text", + "content": "In this study, we present an effective framework for non-autoregressive motion generation that is capable of generating long and high-resolution videos. Our approach leverages learning-based Generative Adversarial Network (GAN) inversion, which learns the inverse mapping of GANs via an inversion network that consists of an encoder and a decoder1. To generate long and coherent videos, we exploit the unique characteristic of the inversion encoder, which captures a rich and smooth manifold between the mapping of images and latent. As illustrated in Fig. 1, to generate a sequence of smooth motion latents, we just need to provide the initial latent code and modulate the inversion encoder with temporal style codes, which are encodings of timestamps with randomness. The motion latents can then be mapped by a StyleGAN decoder to generate a video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 615, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 288, + 664 + ], + "type": "text", + "content": "The proposed framework offers several advantages in a single unified framework. First, the use of an inversion network naturally constrains the generation space to stay consistent with the desired appearance, which is de" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 312 + ], + "type": "text", + "content": "fined by the initial latent code. As demonstrated in Fig. 2, this leads to a significant benefit. Second, thanks to the flexibility of the inversion network in accepting temporal styles of arbitrary timestamps, the framework allows non-autoregressive generation and sparse training [52, 35]. These merits help alleviate the need for heavy discriminators to ensure temporal consistency, as is required in existing approaches. In our implementation, we only need to use a 2D convolutional discriminator instead of a 3D discriminator like MoCoGAN-HD. Third, Unlike existing state-of-the-art methods [52, 35, 7] that couple content and motion decoding in one synthesis network, our framework can naturally support content decoder fine-tuning on different image datasets. Specifically, after fine-tuning the decoder (e.g., StyleGAN2) on another image dataset with the mapping layers and low-resolution synthesis layers fixed, given the same sequence of synthesized motion latents, the generated video can possess the new style of the fine-tuning dataset while preserving the motion patterns of the video generated by the parent content decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 312, + 546, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 546, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 546, + 515 + ], + "type": "text", + "content": "The main contribution of this work is a novel motion generator that modulates a GAN inversion network. This is the first attempt to build such a generator, and it offers several advantages in a unified framework over existing approaches. These advantages include consistent generation, sparse training, and flexibility in supporting style transfer with simple fine-tuning. We additionally contribute a reformulation to the conventional sparse training, through first-frame-aware acyclic positional encoding (FFA-APE) and first-frame-aware sparse training (FFA-ST), to ensure that our motion generator can faithfully reconstruct the initial frame and that the generated video is smooth and continuous. Extensive experiments on DeeperForensics [17], FaceForensics [31], SkyTimelapse [47] and Tai-Chi-HD [34] datasets show that our model is comparable to or even better than state-of-the-art unconditional video generation methods [36, 52, 35] both qualitatively and quantitatively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 525, + 392, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 392, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 392, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 545, + 546, + 713 + ], + "type": "text", + "content": "GAN inversion. The goal of GAN inversion is to find the corresponding vector in the latent space of a pretrained GAN [22, 23] to reconstruct the input image. Existing methods can be classified into three categories [46]: (1) learning-based methods [8, 37, 49, 3, 2, 44, 29], which leverage an encoder network to directly map an image into a latent vector; (2) optimization-based methods [43, 1, 45, 48, 56, 57], which iteratively find the latent vector that best reconstructs the input image using gradient descent; and (3) hybrid models [6, 5, 9, 55], which initialize the iteration process with the result of an encoder network. The design of our motion generator follows the learning-based approach. Therefore, our method is trainable, efficient for single-image inference, and suitable for hierarchical mod-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 288, + 713 + ], + "type": "text", + "content": "1In many contexts, the decoder is a StyleGAN, and the encoder learns to encode a given image to meaningful latent vectors in the StyleGAN space. There is a variety of image manipulation applications [46, 29, 4] developed based upon such an inversion framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22852" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "ulation. We devise the motion generator on the " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " space and use the StyleGAN generated latent as the initial content code to guide the modulated inversion process (see Fig. 2)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 288, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 288, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 288, + 494 + ], + "type": "text", + "content": "Unconditional video generation. Unconditional video generation aims to model the distribution of real videos in a training dataset and generate videos from sampled noise vectors. Many recent studies on this topic are inspired by the success of GANs in image generation. VGAN [42] applies 3D convolutions in both the generator and discriminator, while TGAN [32] optimizes this design by decomposing the generator into an image generator, which is shared by the generation of each frame, and a motion generator2. This framework has been followed by most subsequent studies, such as MoCoGAN [38], which applies a content-motion decomposition. Some approaches [33, 11, 19] have focused on reducing the computational cost of the video discriminator, but the cost is still proportional to the video duration and resolution. Some recent methods have applied more advanced generative frameworks and techniques to unconditional video generation. For example, VideoGPT [50] uses VQ-VAE [28] and GPT [10] to formulate a non-GAN-based video generation approach. Recent studies have also explored unconditional video generation with higher resolution and longer duration. For example, Long-Video-GAN [7] develops a two-phase model that focuses on improving the long-term temporal dynamics of video generation. MoCoGAN-HD [36] and StyleVideoGAN [13] study the generation of latent trajectories in the latent space of a pretrained StyleGAN2 generator. Our approach is inspired by these studies, but differs in the design of the motion generator. Our motion generator is non-autoregressive, thus alleviating the use of heavy discriminators, and it is unique since it obtains the motion latent via modulating a GAN inversion network. This design allows us to attain better motion consistency and semantics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 495, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 579 + ], + "type": "text", + "content": "Recent works [52, 35] explored neural representation-based generators and trained them sparsely as an image GAN. StyleSV [53] improves this framework by introducing StyleGAN3 [21] architecture and several temporal designs. In our work, we extend the idea of sparse training to first-frame-aware sparse training, allowing it to be applied to a generation pipeline conditioned on the initial latent." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 666 + ], + "type": "text", + "content": "Diffusion-based video generation. The diffusion models [15, 30], a new paradigm for image generation tasks, have also achieved significant progress in the task of unconditional video generation [16, 24, 41, 51]. Despite their success, temporal consistency is still an open problem for diffusion models, and GAN-based models exhibit a clear advantage in terms of inference speed." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 543, + 338 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 543, + 338 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 543, + 338 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 543, + 338 + ], + "type": "image", + "image_path": "2cf7ca05a5f0e6b5c34507ab42726cc8feeacf35c1acadb12f52db5aff4cea89.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 345, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 304, + 345, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 547, + 407 + ], + "type": "text", + "content": "Figure 3: From a typical inversion encoder to StyleInV: We add AdaIN layers at the end of each residual block to inject the temporal style, which is a combination of time positional encoding and the first frame latent code. Here \"A\" stands for a learned affine transform [22]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 422, + 389, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 389, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 389, + 437 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 442, + 493, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 493, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 493, + 455 + ], + "type": "text", + "content": "3.1. Preliminaries of Inversion Encoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": "An inversion encoder maps an input image to a vector in the " + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": " latent space of a pretrained StyleGAN2 generator. The generated image that corresponds to this vector should faithfully reconstruct the details of the input image. Therefore, when based on " + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": " latent space, given an input image " + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": ", the reconstruction process can be defined on top of the inversion network Inv as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 352, + 551, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 551, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 352, + 551, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} := G (\\operatorname {I n v} (\\mathbf {x})) := G (E (\\mathbf {x}) + \\overline {{\\mathbf {w}}}). \\tag {1}", + "image_path": "ab88e7ecb7a5955a0a56e757045da7273af8a8c3f667ffabcb86103d24f4e12a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "text", + "content": " denote the inversion encoder and StyleGAN generator, respectively. " + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{w}}\\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "text", + "content": " denotes the average latent vector of the generator in the " + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "text", + "content": " latent space. In our implementation, the encoder " + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "text", + "content": " is a convolutional network backbone that outputs a 512-dimensional vector from the last layer embedding, as shown in Fig. 3(left). We build the encoder on the " + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 569, + 547, + 665 + ], + "type": "text", + "content": " latent space, which eases the design of temporal modulation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 671, + 542, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 542, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 542, + 684 + ], + "type": "text", + "content": "3.2. Temporal Style Modulated Inversion Encoder" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "We observe that the latent space of a StyleGAN trained on a video dataset is typically well-clustered by its content" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "2In the original paper of TGAN [32], the authors called this module temporal generator, which is equivalent to the motion generator used in subsequent studies [38, 36, 52] and in our paper." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22853" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": "subject. Figure 2 shows an example of human face videos, where we depict the results of inverting video clips of different identities into the " + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": " space and visualizing them with t-SNE [40]. It can be observed that the latent space is grouped by human identities. We also observe the same property in video datasets that follow other distributions. This phenomenon suggests that the inversion network inherits some important temporal priors that we could leverage to maintain motion consistency in generated videos." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "text", + "content": "Motivated by this observation, we propose StyleInV, in which the motion latent is generated by modulating a GAN inversion network with temporal styles. Figure 3(right) illustrates the pipeline of our framework. The temporal style " + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "text", + "content": " of a timestamp " + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "text", + "content": " consists of two parts: the motion code " + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "text", + "content": " and the latent code of the initial frame " + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "text", + "content": ". Inspired by [35], we use an acyclic positional encoding module to compute a dynamic embedding of the timestamp " + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "text", + "content": ". However, unlike [35], we make the embedding of the zero timestamp fixed, so this module becomes first-frame-aware. We provide more details in Section 3.3. The latent code " + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 46, + 179, + 288, + 335 + ], + "type": "text", + "content": " of the initial frame is concatenated with the motion code for content-adaptive affine transform." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 335, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 396 + ], + "type": "text", + "content": "The temporal style is injected into the inversion encoder through AdaIN layers at the end of each convolution block. With this design, the encoder " + }, + { + "bbox": [ + 47, + 335, + 287, + 396 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 335, + 287, + 396 + ], + "type": "text", + "content": " of StyleInV becomes a function of the initial latent code " + }, + { + "bbox": [ + 47, + 335, + 287, + 396 + ], + "type": "inline_equation", + "content": "w_{0}" + }, + { + "bbox": [ + 47, + 335, + 287, + 396 + ], + "type": "text", + "content": " and timestamp " + }, + { + "bbox": [ + 47, + 335, + 287, + 396 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 335, + 287, + 396 + ], + "type": "text", + "content": ". The modulated inversion process can be defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 403, + 287, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 287, + 416 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {t} := G \\left(\\operatorname {S t y l e I n V} \\left(\\boldsymbol {w} _ {0}, t\\right)\\right) := G \\left(E \\left(G \\left(\\boldsymbol {w} _ {0}\\right), \\boldsymbol {s} _ {t}\\right) + \\boldsymbol {w} _ {0}\\right). \\quad (2)", + "image_path": "e699a18db48e1ae166d162dde1d2c9f575479615b39e2c656eda5bf6184f4afa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 424, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 459 + ], + "type": "text", + "content": "Notably, the output of " + }, + { + "bbox": [ + 46, + 424, + 287, + 459 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 424, + 287, + 459 + ], + "type": "text", + "content": " serves as the residual w.r.t. " + }, + { + "bbox": [ + 46, + 424, + 287, + 459 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 46, + 424, + 287, + 459 + ], + "type": "text", + "content": ", instead of " + }, + { + "bbox": [ + 46, + 424, + 287, + 459 + ], + "type": "inline_equation", + "content": "\\overline{w}" + }, + { + "bbox": [ + 46, + 424, + 287, + 459 + ], + "type": "text", + "content": ". This modification provides more explicit content information guidance for the inversion encoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "content": "During training, we first train a raw inversion encoder following Eq. (1) on all video frames. Then, we use this network to initialize the weights of all convolution layers in the StyleInV encoder. Other parameters (e.g., FFA-APE and Affine Transforms) are randomly initialized. Finally, the entire StyleInV encoder is trained end-to-end." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 117, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 117, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 117, + 551 + ], + "type": "text", + "content": "3.3. FFA-APE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": "The original implementation of acyclic positional encoding (APE) [35] samples a series of noise vectors " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_{t_0}^{\\mathsf{m}},\\dots ,\\boldsymbol{z}_{t_n}^{\\mathsf{m}}\\sim \\mathcal{N}(\\mathbf{0},\\boldsymbol {I})" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "t_i = i\\cdot \\delta^z" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": ". We call these temporal points anchor points. Here " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\delta^z" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": " is a set constant distance between adjacent anchor points. Then, the noise vectors are mapped to tokens " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_{t_0},\\dots ,\\boldsymbol{u}_{t_n}" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": " by a paddingless conv1d-based motion mapping network. The computation of the acyclic positional encoding " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_t" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": " of arbitrary timestamp " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": " is achieved by a scalable and learnable interpolation between the tokens of two adjacent anchor points that cover " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": ". The computation pipeline is shown in Fig. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In our non-autoregressive generation pipeline, the modulated inversion encoder needs to faithfully reconstruct the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 545, + 164 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 164 + ], + "type": "image", + "image_path": "a8a92dc3843daf5f14c6b568af60c48db73ba8826ba843afb7b689d0728245fc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 174, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 304, + 174, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 547, + 247 + ], + "type": "text", + "content": "Figure 4: FFA-APE: A simplified case to compute " + }, + { + "bbox": [ + 304, + 174, + 547, + 247 + ], + "type": "inline_equation", + "content": "v_{16}" + }, + { + "bbox": [ + 304, + 174, + 547, + 247 + ], + "type": "text", + "content": " for demonstrating the original acyclic positional encoding and our FFA-APE when " + }, + { + "bbox": [ + 304, + 174, + 547, + 247 + ], + "type": "inline_equation", + "content": "\\delta^z = 32" + }, + { + "bbox": [ + 304, + 174, + 547, + 247 + ], + "type": "text", + "content": " and convld kernel size is 3. In FFA-APE, the encoding of zero timestamp " + }, + { + "bbox": [ + 304, + 174, + 547, + 247 + ], + "type": "inline_equation", + "content": "(u_0)" + }, + { + "bbox": [ + 304, + 174, + 547, + 247 + ], + "type": "text", + "content": " only depends on constant paddings and a constant noise vector, thus is fixed for any sampled noise vector sequence." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "text", + "content": "initial frame when the input timestamp is zero, making it necessary to fix the computation of the APE for the zero timestamp " + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "text", + "content": ". The original APE computation for " + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "text", + "content": " is dynamic and depends on randomly sampled noise vectors, which can lead to dynamic output that is not desired. To address this, we devise a first-frame-aware acyclic positional encoding (FFA-APE) method that fixes " + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "text", + "content": " while maintaining the smoothness of APE (see Fig. 4). We achieve this by replacing the noise vector for the first anchor point with a learnable constant vector " + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "inline_equation", + "content": "c_{0}^{\\mathrm{m}}" + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "text", + "content": ", and using left-sided conv1d layers with constant padding instead of the padding-less conv1d layers. This way, the value of " + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "text", + "content": " only depends on the constant vector " + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "inline_equation", + "content": "c_{0}^{\\mathrm{m}}" + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "text", + "content": " and the left-padded vectors, which are also constant. As a result, " + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 304, + 267, + 546, + 448 + ], + "type": "text", + "content": " is naturally fixed without affecting the continuity of positional encoding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 455, + 367, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 367, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 367, + 467 + ], + "type": "text", + "content": "3.4. FFA-ST" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": "In this section, we introduce the first-frame-aware sparse training specially designed for our framework. Recent non-autoregressive video generation approaches [52, 35] use a discriminator design that only considers " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t_1},\\dots ,\\pmb{x}_{t_k}" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " for each video, distinguishing the realness of the input conditioned on the time difference of input frames " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\delta_{i} = t_{i + 1} - t_{i}" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": ". This training scheme is called sparse training. StyleGAN-V [35] has analyzed the choice of " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " and found that " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": " is ideal for most datasets. The discriminator is defined as " + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "inline_equation", + "content": "D(\\pmb{x}_{t_{1,2,3}},\\delta_{1,2})" + }, + { + "bbox": [ + 304, + 474, + 545, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "We follow this training scheme to make full use of our non-autoregressive framework. Nonetheless, using only three randomly sampled timestamps to train the generator and discriminator can result in sharp transitions at the beginning of the generated video, where the generated " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}_1" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": " usually diverge too much, and sometimes even switch to another identity and never return. This happens because although we define the generation process of a video as a modulated inversion process of the start frame, the discriminator is unaware of it. The discriminator only focuses on" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22854" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "the smoothness of generated latent trajectories, failing to ensure the motion generator produces frames that share the identity with the start frame." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 156 + ], + "type": "text", + "content": "To solve this problem, we introduce the initial frame into the discriminator to enhance content consistency and motion smoothness. The adversarial loss for the first-frame-aware discriminator (FFA-D) can be written as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 163, + 287, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 163, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 59, + 163, + 287, + 211 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {y} _ {t _ {0, 1, 2, 3}} = G \\left(\\text {S t y l e I n V} \\left(w _ {0}, t _ {0, 1, 2, 3}\\right)\\right), \\\\ \\mathcal {L} _ {a d v} = \\mathbb {E} _ {\\boldsymbol {x} \\sim p _ {v}} \\left[ \\log D \\left(\\boldsymbol {x} _ {t _ {0, 1, 2, 3}}, \\delta_ {0, 1, 2}\\right) \\right] \\tag {3} \\\\ + \\mathbb {E} _ {w _ {0} \\sim p _ {\\mathcal {W}}} \\left[ \\log (1 - D (\\boldsymbol {y} _ {t _ {0, 1, 2, 3}}, \\delta_ {0, 1, 2}) \\right], \\\\ \\end{array}", + "image_path": "6d72ad8e74c2f9caa63bab42289389f4dbf5eb9f296dfc2e7a7d96e2a4353d7c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "text", + "content": "where we specify " + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "inline_equation", + "content": "t_0 = 0" + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "inline_equation", + "content": "p_v" + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "inline_equation", + "content": "p_{\\mathcal{W}}" + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "text", + "content": " denote the real data distribution and " + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "text", + "content": " latent space distribution, respectively. To explicitly enforce initial frame reconstruction, we use a " + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "text", + "content": " loss for the generated " + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "inline_equation", + "content": "y_{t_0}" + }, + { + "bbox": [ + 46, + 217, + 287, + 266 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 274, + 287, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 274, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 80, + 274, + 287, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L _ {2}} = \\left\\| G \\left(w _ {0}\\right) - G \\left(\\text {S t y l e I n V} \\left(w _ {0}, 0\\right)\\right) \\right\\| _ {2}. \\tag {4}", + "image_path": "678b3da918bbd31dd80b68fc27251df1d4e9dcfc838bcc91426c3ba76dfd0e70.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "type": "text", + "content": "Finally, we apply latent regularization [29, 26] to the encoder's output, so as to enhance content consistency:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 327, + 287, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 327, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 97, + 327, + 287, + 348 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\sum_ {i = 0} ^ {3} | | E (G (w _ {0}), t _ {i}) | | _ {2}. \\tag {5}", + "image_path": "6d5a0af2ec4c78675252e8916e9536d253855db936abdbb274e2c3fc0bb4f0e1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 355, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 287, + 378 + ], + "type": "text", + "content": "The overall loss function for training our motion generator and the discriminator is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 388, + 287, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 388, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 75, + 388, + 287, + 406 + ], + "type": "interline_equation", + "content": "\\min _ {E} \\max _ {D} \\mathcal {L} _ {a d v} + \\min _ {E} \\left(\\lambda_ {L _ {2}} \\mathcal {L} _ {L _ {2}} + \\lambda_ {r e g} \\mathcal {L} _ {r e g}\\right). \\tag {6}", + "image_path": "b7be5252a1fc4df70a155fdadbaf0f570c83d98476f4a691be53a350b1ebd0a2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{L_2}" + }, + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "text", + "content": " are the loss hyperparameters. We also apply discriminator adaptive augmentation [20, 35] and " + }, + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "r1" + }, + { + "bbox": [ + 46, + 413, + 287, + 460 + ], + "type": "text", + "content": " regularization [22, 35] to further improve the training stability and generation quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 468, + 221, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 221, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 221, + 480 + ], + "type": "text", + "content": "3.5. Finetuning-based Style Transfer" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 581 + ], + "type": "text", + "content": "Our 'inversion encoder+decoder' framework can naturally take a pretrained StyleGAN model as the generator. And such a configuration allows the generator to be finetuned for different styles, and yet still able to use the motion generator for generating new video with styles. The capability is not possible with existing non-autoregressive video generation methods [52, 35] because they cannot be finetuned under an image GAN training scheme." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "To achieve style transfer, as illustrated in Fig. 5, we fine-tune the pre-trained StyleGAN model using an image dataset, such as MegaCartoon [27], while keeping the mapping network and low-resolution " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\leqslant 32^2" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": ", coarse and middle layers in [22]) synthesis blocks fixed. This configuration maintains the distribution of the " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " space during finetuning. To improve identity preservation and reduce artifacts, we apply both a perceptual loss [18] and an identity loss [12] between the images generated by the original and fine-tuned StyleGAN. We show some visual results in Fig. 8. The style-transferred video maintains the" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 545, + 230 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 545, + 230 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 545, + 230 + ], + "type": "image", + "image_path": "555f9ea1ab613a2e8e335f4d9eb0bc8af194509ffc3946ec7326c60b3e51f314.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 241, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 305, + 241, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 305, + 241, + 545, + 326 + ], + "type": "text", + "content": "Figure 5: Finetuning-based style transfer: our framework allows easy fine-tuning of decoder (a pretrained StyleGAN generator) to a new domain by freezing mapping and low-resolution " + }, + { + "bbox": [ + 305, + 241, + 545, + 326 + ], + "type": "inline_equation", + "content": "(\\leqslant 32^{2})" + }, + { + "bbox": [ + 305, + 241, + 545, + 326 + ], + "type": "text", + "content": " layers. Standard identity loss and perceptual loss are applied to improve identity preservation and reduce artifacts. The style-transferred videos can then be generated by incorporating the StyleInV motion generator." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 347, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 545, + 441 + ], + "type": "text", + "content": "same motion pattern as the video generated by the parent model, while adopting a new style from the fine-tuning image dataset. It is noteworthy that the finetuning process is independent of the video generation training. It means that the finetuning-based style transfer is \"plug-and-play\" as the fine-tuned image generator can be used on any StyleInV models. There is no additional computational cost in the inference either." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 453, + 387, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 387, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 387, + 467 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "type": "text", + "content": "Datasets. We use four video datasets in our main experiments: DeeperForensics " + }, + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "type": "text", + "content": " [17], FaceForensics " + }, + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "type": "text", + "content": " [31], SkyTimelapse " + }, + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "type": "text", + "content": " [47] and TaiChi " + }, + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 305, + 474, + 545, + 581 + ], + "type": "text", + "content": " [34]. The cropping strategy for DeeperForensics [17] and FaceForensics [31] is different. For DeeperForensics, we use a stabilized FFHQ [22] cropping strategy [25], while we follow the strategy firstly adopted by TGAN-V2 [33] for FaceForensics. Please refer to the supplementary material for a detailed discussion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 582, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 665 + ], + "type": "text", + "content": "Baselines. We explore four state-of-the-art methods for comparison: MoCoGAN-HD [36], DIGAN [52], StyleGAN-V [35] and Long-Video-GAN [7]. Among these methods, MoCoGAN-HD and DIGAN require an explicit setting of the training clip length. We follow the default setting of their paper to set the clip length as 16 for both methods. This setting is identical to StyleGAN-V [35]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "In addition, on DeeperForensics, we explore an optimized setting on DIGAN and MoCoGAN-HD for a more fair comparison. For DIGAN, we directly increase the clip length to 128 frames. For MoCoGAN-HD, we apply the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22855" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "first-frame-aware sparse training to train its motion generator, so as to avoid using a heavy 3D discriminator, allowing it to be trained with 128-frame clips as other methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 228 + ], + "type": "text", + "content": "Evaluation. We use Fréchet Inception Distance (FID) [14] and Fréchet Video Distance (FVD) [39] to evaluate all models quantitatively. In practice, we follow the metric calculation framework provided by StyleGAN-V [35] to first generate a fake video dataset with 2,048 synthesized clips, each of 128 frames. For FID, we sample 50k frames from real and fake video datasets to compute the result. For FVD, we compute " + }, + { + "bbox": [ + 46, + 109, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{16}" + }, + { + "bbox": [ + 46, + 109, + 286, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 109, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{128}" + }, + { + "bbox": [ + 46, + 109, + 286, + 228 + ], + "type": "text", + "content": " with the first 16 frames and all 128 frames of each clip, respectively. We use FID results to show the single-frame image quality of each method." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 229, + 286, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 286, + 302 + ], + "type": "text", + "content": "To ensure a fair comparison, we re-benchmark the quantitative results of every method on every dataset. We retrain all the baselines using the official paper setting, except for MoCoGAN-HD on SkyTimelapse, where an officially released checkpoint is available. For more implementation details, please refer to the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 310, + 132, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 132, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 132, + 321 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 329, + 286, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 286, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 286, + 400 + ], + "type": "text", + "content": "Quantitative results. Table 1 summarizes the quantitative results of our method compared to other baselines. Our method achieves competitive quantitative results on all the benchmarks. Notably, although MoCoGAN-HD and DIGAN are trained with clips of 16 frames, we still outperform them in terms of " + }, + { + "bbox": [ + 46, + 329, + 286, + 400 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{16}" + }, + { + "bbox": [ + 46, + 329, + 286, + 400 + ], + "type": "text", + "content": " metrics on all four datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 286, + 581 + ], + "type": "text", + "content": "Qualitative results. Figure 6 shows the qualitative comparison between our method and the baselines on all four datasets. MoCoGAN-HD and DIGAN both suffer from motion collapse, resulting in a degraded generation quality over time. StyleGAN-V shows an impressive visual performance on FaceForensics and SkyTimelapse, but it sometimes fails to maintain the identity and accessories on DeeperForensics and lacks diversity and magnitude of motion over a long time span on TaiChi (the subject gradually fixes at one state). Long-Video-GAN is exceptionally good at SkyTimelapse, but it cannot achieve similar performance on other datasets. It fails to maintain the identity on DeeperForensics, and its single-frame content on TaiChi lacks details and is inferior to other methods. The generated videos by Long-Video-GAN collapse on FaceForensics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 714 + ], + "type": "text", + "content": "In contrast to existing methods, our method demonstrates stable results on all four datasets, particularly with superior identity preservation on human-face video and long-term generation quality on TaiChi. Although our method outperforms existing methods in terms of content quality, continuity, and quantitative results, the motion semantics of our generated videos on SkyTimelapse are inferior to those on other datasets. This could be one of the limitations of our work and an area for future improvement. Extended experiments. We present more in-depth comparisons in Table 2 and Fig. 7 by introducing training im" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "text", + "content": "Table 1: FID, " + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{16}" + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{128}" + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "text", + "content": " results of video generation methods on (a) DeeperForensics " + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "text", + "content": ", (b) FaceForensics " + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "text", + "content": ", (c) TaiChi " + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "text", + "content": ", and (d) SkyTimelapse " + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 305, + 70, + 545, + 118 + ], + "type": "text", + "content": ". **Bolds** indicate best and **underline** indicate the second best." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 313, + 140, + 537, + 220 + ], + "blocks": [ + { + "bbox": [ + 378, + 124, + 473, + 136 + ], + "lines": [ + { + "bbox": [ + 378, + 124, + 473, + 136 + ], + "spans": [ + { + "bbox": [ + 378, + 124, + 473, + 136 + ], + "type": "text", + "content": "(a) DeeperForensics " + }, + { + "bbox": [ + 378, + 124, + 473, + 136 + ], + "type": "inline_equation", + "content": "256^{2}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 140, + 537, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 140, + 537, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 537, + 220 + ], + "type": "table", + "html": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD135.30101.07610.30
DIGAN191.9946.691060.27
StyleGAN-V59.5939.3368.81
Long-Video-GAN56.5474.77169.45
StyleInV (ours)54.0541.5853.93
", + "image_path": "a4bf78635fb69b58b0a3af5273817b40503d6d3b1c5cace12b96428dee08c88a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 313, + 239, + 537, + 319 + ], + "blocks": [ + { + "bbox": [ + 383, + 223, + 468, + 233 + ], + "lines": [ + { + "bbox": [ + 383, + 223, + 468, + 233 + ], + "spans": [ + { + "bbox": [ + 383, + 223, + 468, + 233 + ], + "type": "text", + "content": "(b) FaceForensics " + }, + { + "bbox": [ + 383, + 223, + 468, + 233 + ], + "type": "inline_equation", + "content": "256^{2}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 239, + 537, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 537, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 537, + 319 + ], + "type": "table", + "html": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD24.45112.67486.69
DIGAN151.53146.621993.20
StyleGAN-V8.6452.92108.86
Long-Video-GAN40.40233.26567.78
StyleInV (ours)12.0647.88103.63
", + "image_path": "8ffcfff41b90f82c21429be436830ed26686478724dfedc866b8283c909345bd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 313, + 338, + 537, + 417 + ], + "blocks": [ + { + "bbox": [ + 397, + 320, + 455, + 332 + ], + "lines": [ + { + "bbox": [ + 397, + 320, + 455, + 332 + ], + "spans": [ + { + "bbox": [ + 397, + 320, + 455, + 332 + ], + "type": "text", + "content": "(c) TaiChi " + }, + { + "bbox": [ + 397, + 320, + 455, + 332 + ], + "type": "inline_equation", + "content": "256^{2}" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 338, + 537, + 417 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 537, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 537, + 417 + ], + "type": "table", + "html": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD73.61315.03622.95
DIGAN67.24196.77954.93
StyleGAN-V35.68254.74477.78
Long-Video-GAN43.90248.55502.65
StyleInV (ours)41.55185.72328.90
", + "image_path": "485826393b94d8ba8e005d3e58b4e3253cde9cbb63fc897d58c19904d458faa4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 436, + 537, + 515 + ], + "blocks": [ + { + "bbox": [ + 383, + 419, + 468, + 431 + ], + "lines": [ + { + "bbox": [ + 383, + 419, + 468, + 431 + ], + "spans": [ + { + "bbox": [ + 383, + 419, + 468, + 431 + ], + "type": "text", + "content": "(d) SkyTimelapse " + }, + { + "bbox": [ + 383, + 419, + 468, + 431 + ], + "type": "inline_equation", + "content": "256^{2}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 436, + 537, + 515 + ], + "lines": [ + { + "bbox": [ + 315, + 436, + 537, + 515 + ], + "spans": [ + { + "bbox": [ + 315, + 436, + 537, + 515 + ], + "type": "table", + "html": "
MethodFID (↓)FVD16 (↓)FVD128 (↓)
MoCoGAN-HD251.81696.584116.03
DIGAN32.83148.08269.43
StyleGAN-V16.9581.32197.83
Long-Video-GAN25.41116.50152.70
StyleInV (ours)14.3277.04194.25
", + "image_path": "3f5a460fa021cb8189d1b4a370cc8f9436959b2f5ef3911485d47aacaafc2d07.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 526, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 545, + 598 + ], + "type": "text", + "content": "Table 2: FID, " + }, + { + "bbox": [ + 305, + 526, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{16}" + }, + { + "bbox": [ + 305, + 526, + 545, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 526, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{128}" + }, + { + "bbox": [ + 305, + 526, + 545, + 598 + ], + "type": "text", + "content": " results of extended experiments on DeeperForensics " + }, + { + "bbox": [ + 305, + 526, + 545, + 598 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 305, + 526, + 545, + 598 + ], + "type": "text", + "content": ". We apply sparse training to MoCoGAN-HD [36] (#1) and change the preset clip length of DIGAN [52] to 128 (#2). **Bolds** indicate best. (-) indicates a smaller (better) quantitative result, while (+) indicates a larger (worse) one, compared with Table 1a." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 309, + 605, + 544, + 657 + ], + "blocks": [ + { + "bbox": [ + 309, + 605, + 544, + 657 + ], + "lines": [ + { + "bbox": [ + 309, + 605, + 544, + 657 + ], + "spans": [ + { + "bbox": [ + 309, + 605, + 544, + 657 + ], + "type": "table", + "html": "
#MethodFID (↓)FVD16 (↓)FVD128 (↓)
1[36] + Sparse Training55.84 (-)54.58 (-)129.13 (-)
2[52] + Clip 12874.80 (-)87.42 (+)95.80 (-)
3StyleInV (ours)54.0541.5853.93
", + "image_path": "7f7e9f5f15910feea0330622064310a3409b11c157667c6abf7b3fdbc0f4708d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "provements to baselines. Increasing the clip length generally improves the results of MoCoGAN-HD and DIGAN, but they are still inferior to our method. Notably, training" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22856" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 291, + 228 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 291, + 228 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 291, + 228 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 291, + 228 + ], + "type": "image", + "image_path": "56532472b55d24987e00ab4a2a7ea17bcf8e592b0c366abea6ba417212b5770b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 233, + 219, + 243 + ], + "lines": [ + { + "bbox": [ + 124, + 233, + 219, + 243 + ], + "spans": [ + { + "bbox": [ + 124, + 233, + 219, + 243 + ], + "type": "text", + "content": "(a) DeeperForensics " + }, + { + "bbox": [ + 124, + 233, + 219, + 243 + ], + "type": "inline_equation", + "content": "256^{2}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 541, + 227 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 541, + 227 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 541, + 227 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 541, + 227 + ], + "type": "image", + "image_path": "aa0537a460bfc8d84f3504c1cd466626fb3b59f20d7125d0aa03596485022036.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 232, + 465, + 243 + ], + "lines": [ + { + "bbox": [ + 379, + 232, + 465, + 243 + ], + "spans": [ + { + "bbox": [ + 379, + 232, + 465, + 243 + ], + "type": "text", + "content": "(b) FaceForensics " + }, + { + "bbox": [ + 379, + 232, + 465, + 243 + ], + "type": "inline_equation", + "content": "256^{2}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 245, + 291, + 401 + ], + "blocks": [ + { + "bbox": [ + 52, + 245, + 291, + 401 + ], + "lines": [ + { + "bbox": [ + 52, + 245, + 291, + 401 + ], + "spans": [ + { + "bbox": [ + 52, + 245, + 291, + 401 + ], + "type": "image", + "image_path": "43ec2d0d21d64d2c460706b674201465506293d1307c2c8012a1adbfe48e9723.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 407, + 200, + 417 + ], + "lines": [ + { + "bbox": [ + 143, + 407, + 200, + 417 + ], + "spans": [ + { + "bbox": [ + 143, + 407, + 200, + 417 + ], + "type": "text", + "content": "(c) TaiChi " + }, + { + "bbox": [ + 143, + 407, + 200, + 417 + ], + "type": "inline_equation", + "content": "256^{2}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 245, + 541, + 401 + ], + "blocks": [ + { + "bbox": [ + 304, + 245, + 541, + 401 + ], + "lines": [ + { + "bbox": [ + 304, + 245, + 541, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 541, + 401 + ], + "type": "image", + "image_path": "534c12848b3af10c60bc89e47d58713f7b31efbb9efafdd5d49edf17dab4a01c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 407, + 464, + 418 + ], + "lines": [ + { + "bbox": [ + 380, + 407, + 464, + 418 + ], + "spans": [ + { + "bbox": [ + 380, + 407, + 464, + 418 + ], + "type": "text", + "content": "(d) SkyTimelapse " + }, + { + "bbox": [ + 380, + 407, + 464, + 418 + ], + "type": "inline_equation", + "content": "256^{2}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 464, + 293, + 559 + ], + "blocks": [ + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "text", + "content": "Figure 6: Uncurated samples from the existing methods on DeeperForensics " + }, + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "text", + "content": ", FaceForensics " + }, + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "text", + "content": ", TaiChi " + }, + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "text", + "content": " and SkyTimelapse " + }, + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "text", + "content": ", respectively. We sample a 128-frame video and display every 16 frames, starting from " + }, + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 427, + 545, + 452 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 464, + 293, + 559 + ], + "lines": [ + { + "bbox": [ + 50, + 464, + 293, + 559 + ], + "spans": [ + { + "bbox": [ + 50, + 464, + 293, + 559 + ], + "type": "image", + "image_path": "ea122f05fbb5a5bfe1ae4aea506e0b126b977781c803f94e5b79e2384b369130.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 294, + 464, + 544, + 559 + ], + "blocks": [ + { + "bbox": [ + 294, + 464, + 544, + 559 + ], + "lines": [ + { + "bbox": [ + 294, + 464, + 544, + 559 + ], + "spans": [ + { + "bbox": [ + 294, + 464, + 544, + 559 + ], + "type": "image", + "image_path": "93fd0e7287387afd330d39731987c3ce921e7bcac9a1a197ccde13494cb39b12.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 569, + 544, + 594 + ], + "lines": [ + { + "bbox": [ + 46, + 569, + 544, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 544, + 594 + ], + "type": "text", + "content": "Figure 7: Qualitative comparison of extended experiments. “M+ST” and “D-C128” correspond to Table 2 (#1) and (#2), respectively. Each row shows the first and last 128 frames of a 2056-frame (68.5s) video, displayed every 16 frames." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 614, + 287, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 614, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 614, + 287, + 710 + ], + "type": "text", + "content": "with longer clips harms the short-term " + }, + { + "bbox": [ + 46, + 614, + 287, + 710 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{16}" + }, + { + "bbox": [ + 46, + 614, + 287, + 710 + ], + "type": "text", + "content": " result of DIGAN, which indicates its tradeoff between duration length and local temporal quality. Qualitatively, for both methods, the generated content is evidently improved within 128 frames, although the sparsely trained MoCoGAN-HD exhibits issues with identity switching. Motion collapse is still observed when MoCoGAN-HD and DIGAN generates long videos. In contrast, our method can stably generate ex" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 614, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 614, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 305, + 614, + 545, + 651 + ], + "type": "text", + "content": "tremely long videos without motion collapse. Our method outperforms the sparsely trained MoCoGAN-HD, demonstrating the superiority of our motion generator design." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 658, + 376, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 376, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 376, + 671 + ], + "type": "text", + "content": "4.2. Properties" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "As discussed in Section 3.5, our method has the unique advantage over state-of-the-art methods, such as StyleGAN-V and Long-Video-GAN, on its high compati" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22857" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 301, + 191 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 301, + 191 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 301, + 191 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 301, + 191 + ], + "type": "image", + "image_path": "fffce52942f24b84dde75e8bad4ec7faad5f6fb379d3d890f26be0c5e61ee3e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 200, + 545, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 200, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 545, + 225 + ], + "type": "text", + "content": "Figure 8: Finetuning-based style transfer result. The 1st row is generated by the parent model trained on CelebV-HQ [54]. The 2nd, 3rd, and 4th row uses the StyleGAN generator fintuned on Cartoon, Arcane, and MetFace, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 303, + 72, + 543, + 191 + ], + "blocks": [ + { + "bbox": [ + 303, + 72, + 543, + 191 + ], + "lines": [ + { + "bbox": [ + 303, + 72, + 543, + 191 + ], + "spans": [ + { + "bbox": [ + 303, + 72, + 543, + 191 + ], + "type": "image", + "image_path": "ecdfb14ff0a42adaca4cec5a778eea248f2b918b169ecf1fa81d11a57d08d8b0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 243, + 286, + 449 + ], + "blocks": [ + { + "bbox": [ + 50, + 243, + 286, + 449 + ], + "lines": [ + { + "bbox": [ + 50, + 243, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 50, + 243, + 286, + 449 + ], + "type": "image", + "image_path": "f09d88d1e8513870048be291257c1dd67a2f7f0d77367991c2b0ded16e7b516e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 458, + 287, + 482 + ], + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 482 + ], + "type": "text", + "content": "Figure 9: Initial-frame conditioned generation and style transferred results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 507, + 260, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 260, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 260, + 519 + ], + "type": "text", + "content": "bility with StyleGAN-based downstream techniques." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 687 + ], + "type": "text", + "content": "Finetuning-based style transfer. We train the parent model (motion generator and StyleGAN) on CelebV-HQ [54], as its rich identity makes it more suitable for transfer learning. To perform style transfer, we fine-tune the StyleGAN on the Cartoon [27], MetFace [20], and Arcane datasets following the procedure outlined in Section 3.5. In Fig. 8, we show examples where the same StyleInV-generated latent sequence is decoded by different but aligned StyleGAN generators. Our method achieves satisfactory results in terms of smooth video style transfer with well-aligned face structure, identity, and expression, demonstrating its desirable properties and potential for various applications. More results can be found in the supplementary video." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Initial-frame conditioned generation. Our network supports generating a series of content given a real-world im" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 266, + 544, + 353 + ], + "blocks": [ + { + "bbox": [ + 312, + 244, + 538, + 256 + ], + "lines": [ + { + "bbox": [ + 312, + 244, + 538, + 256 + ], + "spans": [ + { + "bbox": [ + 312, + 244, + 538, + 256 + ], + "type": "text", + "content": "Table 3: Ablation result on the DeeperForensics dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 266, + 544, + 353 + ], + "lines": [ + { + "bbox": [ + 309, + 266, + 544, + 353 + ], + "spans": [ + { + "bbox": [ + 309, + 266, + 544, + 353 + ], + "type": "table", + "html": "
#MethodFID (↓)FVD16 (↓)FVD128 (↓)
1w/o inversion encoder54.3559.49152.82
2w/o FFA-APE55.2688.98144.52
3w/o Eq.(4) & Eq.(3)52.5567.4358.88
4w/o Eq.(3)53.9586.3259.76
5Ours54.0541.5853.93
", + "image_path": "57ce44e6928ac730e35f9cc8344300da4566e605a1aad58f6426a678d4db0dad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 373, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 493 + ], + "type": "text", + "content": "age as the initial frame. We first inverse the image into the StyleGAN2 latent space with a pSp [29] encoder, which is trained to initialize the weights of StyleInV. We treat it as the 512-dimensional initial frame latent " + }, + { + "bbox": [ + 304, + 373, + 545, + 493 + ], + "type": "inline_equation", + "content": "w_{0}" + }, + { + "bbox": [ + 304, + 373, + 545, + 493 + ], + "type": "text", + "content": ", then use it to generate a video with our StyleInV. The generated latent sequence can be also applied to a finetuned image generator to synthesize a style-transferred animation video. Through this pipeline, the real image is reconstructed twice, the first time is during the inversion process, while the second time is when synthesizing " + }, + { + "bbox": [ + 304, + 373, + 545, + 493 + ], + "type": "inline_equation", + "content": "G(\\text{StyleInV}(w_{0},0))" + }, + { + "bbox": [ + 304, + 373, + 545, + 493 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 494, + 546, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 546, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 546, + 648 + ], + "type": "text", + "content": "When the real images are sampled from the training dataset (see Fig. 9 first two rows), " + }, + { + "bbox": [ + 304, + 494, + 546, + 648 + ], + "type": "inline_equation", + "content": "G(\\text{StyleInV}(w_0, 0))" + }, + { + "bbox": [ + 304, + 494, + 546, + 648 + ], + "type": "text", + "content": " can faithfully reconstruct the raw image and generate high-quality videos. We then test the generation quality for real images sampled out of the training set (see Fig. 9 last two rows, where we select Benedict Cumberbatch and Anne Hathaway). We use the StyleGAN2 generator and StyleInV model trained on CelebV-HQ [54] dataset as it is richer in its identities. The results show that our StyleInV network can still generate meaningful videos while reconstructing the initial frame decently, and the style transfer results are smooth and well-aligned. Please refer to the supplementary material for more results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 658, + 406, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 406, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 406, + 670 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Motion generator design. We explore two alternative motion generator designs. The first is the autoregressive MoCoGAN-HD design, which has been discussed in Sec-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22858" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 286, + 140 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 140 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 140 + ], + "type": "image", + "image_path": "5305fab939e19ec3b5aafaad54263943881abfb05bcfe9219646bba506746dc1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 148, + 286, + 197 + ], + "lines": [ + { + "bbox": [ + 47, + 148, + 286, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 286, + 197 + ], + "type": "text", + "content": "Figure 10: Ablating FFA-APE. Generate the first frame with a fixed " + }, + { + "bbox": [ + 47, + 148, + 286, + 197 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 47, + 148, + 286, + 197 + ], + "type": "text", + "content": " but different temporal noise sequences. The model without FFA-APE fails to reconstruct the initial frame and generates the first frame with randomness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 212, + 285, + 280 + ], + "blocks": [ + { + "bbox": [ + 50, + 212, + 285, + 280 + ], + "lines": [ + { + "bbox": [ + 50, + 212, + 285, + 280 + ], + "spans": [ + { + "bbox": [ + 50, + 212, + 285, + 280 + ], + "type": "image", + "image_path": "b7dd93cd4ea4567fc4badbf1360ab4d02d8d8ebe048a67c9cf13adabe7d94316.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 289, + 286, + 324 + ], + "lines": [ + { + "bbox": [ + 47, + 289, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 286, + 324 + ], + "type": "text", + "content": "Figure 11: Ablating FFA-ST. The " + }, + { + "bbox": [ + 47, + 289, + 286, + 324 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 289, + 286, + 324 + ], + "type": "text", + "content": " loss ensures the initial frame reconstruction. But without the initial frame included in the discriminator, we still cannot preserve the identity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 286, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 286, + 422 + ], + "type": "text", + "content": "tion 4.1. For the second design, we remove all Convs, AdaIN and affine transform layers in Fig. 3 and let the mapped temporal style be the output of inversion encoder, i.e., the residual w.r.t. initial frame latent " + }, + { + "bbox": [ + 46, + 350, + 286, + 422 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 46, + 350, + 286, + 422 + ], + "type": "text", + "content": ". It largely harms identity preservation. Both " + }, + { + "bbox": [ + 46, + 350, + 286, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{16}" + }, + { + "bbox": [ + 46, + 350, + 286, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 350, + 286, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{128}" + }, + { + "bbox": [ + 46, + 350, + 286, + 422 + ], + "type": "text", + "content": " degrade significantly as is shown in Table 3(#1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 424, + 286, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 286, + 555 + ], + "type": "text", + "content": "FFA-APE. We evaluate the importance of our first-frame-aware acyclic positional encoding (FFA-APE) module by replacing it with the original design proposed by [35]. As shown in Fig. 10, the dynamic embedding of the zero timestamp prevents the network from faithfully reconstructing the initial frame. In contrast, our full method can stably realize reconstruction. In addition, the L2 loss in Eq. (4) fails to converge for the ablation method, and its gradient further harms the learning of the positional encoding module, leading to a much worse quantitative result shown in Table 3 (#2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": "FFA-ST. We conduct two ablation experiments for the FFA-ST modules. In the first experiment, we remove the initial frame from the discriminator and remove the reconstruction loss (Eq.(4)) (Table 3(#3)). In the second experiment, we only remove the initial frame from the discriminator while keeping the reconstruction loss (Table 3(#4)). As shown in Fig. 11, without the reconstruction loss, our model cannot reconstruct the initial frame accurately. With the " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": " loss, the initial frame is reconstructed, but there is a sudden transition between the first two frames, and sometimes, the identity also changes, leading to a worse " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{FVD}_{16}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": " result. These experiments demonstrate the importance of our first-frame-aware discriminator (FFA-D)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 378, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 378, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 378, + 83 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 545, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 235 + ], + "type": "text", + "content": "We have presented a novel approach for unconditional video generation by employing a pretrained StyleGAN image generator. The proposed StyleInV motion generator generates latents in the StyleGAN2 latent space by modulating a learning-based inversion network, and thus capable of inheriting its informative priors of the initial latent. Our network features non-autoregressive training and uniquely supports fine-tuning based style transfer. Extensive experiments demonstrate the superiority of our method in generating long and high-resolution videos, outperforming state-of-the-art baselines. Here we also briefly discuss our limitations and broader impacts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 243, + 382, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 382, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 382, + 255 + ], + "type": "text", + "content": "5.1. Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 262, + 545, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 333 + ], + "type": "text", + "content": "Inferior motion semantics on SkyTimelapse. Our motion semantics on SkyTimelapse [47] are inferior to those on other datasets. This could be due to different dataset characteristics, as videos in SkyTimelapse are not subject-centric and typically driven by global motions, which does not align perfectly with our model nature." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 334, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 545, + 429 + ], + "type": "text", + "content": "The impact of dataset identity richness. When the scale of facial identities in the video dataset is too small, the effects of inversion, editing, and style transfer are constrained. Image generation quality. The generation quality of StyleGAN determines the performance upper bound of our method. The images generated by the StyleGAN2 models have artifacts in the background on SkyTimelapse [47], and lack fine details and a sense of structure on TaiChi [34]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 489 + ], + "type": "text", + "content": "Model training. Our approach is two-stage, requiring 7.5 and 9 GPU days each, which is more than the 8 GPU days of StyleGAN-V [35]. Despite this, StyleInV is as efficient when finetuning the hyperparameters of the video generator, since the image generator only needs to be trained once." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 498, + 408, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 408, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 408, + 510 + ], + "type": "text", + "content": "5.2. Broader Impacts" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 517, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 576 + ], + "type": "text", + "content": "We believe that the potential of StyleInV can be further exploited. Our method can provide a natural solution towards mega-pixel level video generation and StyleGAN-based editing, and it might in return promote the research of learning-based GAN inversion methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 577, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 545, + 635 + ], + "type": "text", + "content": "As for the negative side, StyleInV may ease the synthesis of better-quality fake videos with threats. We believe that it can be alleviated by developing more advanced falsified media detection methods or contributing larger-scale and higher-quality forgery detection datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 636, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2022-01-030). It is also supported under the RIE2020 Industry Alignment Fund Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). We thank Shuai Yang for his help in this work." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22859" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. StyleFlow: Attribute-conditioned exploration of StyleGAN-generated images using conditional continuous normalizing flows. ACM TOG, 40:1-21, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "text", + "content": "[2] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: Age transformation using a style-based regression model. ACM TOG, 40:1-12, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "type": "text", + "content": "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. ReStyle: A residual-based StyleGAN encoder via iterative refinement. In ICCV, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 286, + 237 + ], + "type": "text", + "content": "[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit H Bermano. HyperStyle: StyleGAN inversion with hypernetworks for real image editing. In CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 286, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 286, + 281 + ], + "type": "text", + "content": "[5] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. ACM TOG, 38:1-11, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 286, + 316 + ], + "type": "text", + "content": "[6] David Bau, Jun-Yan Zhu, Jonas Wulff, William Peebles, Hendrik Strobelt, Bolei Zhou, and Antonio Torralba. Seeing what a GAN cannot generate. In ICCV, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 318, + 286, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 286, + 361 + ], + "type": "text", + "content": "[7] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei A Efros, and Tero Karras. Generating long videos of dynamic scenes. In NeurlPS, 2022. 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 363, + 286, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 286, + 395 + ], + "type": "text", + "content": "[8] Lucy Chai, Jonas Wulff, and Phillip Isola. Using latent space regression to analyze and leverage compositionality in GANs. In ICLR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 397, + 286, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 286, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 286, + 429 + ], + "type": "text", + "content": "[9] Lucy Chai, Jun-Yan Zhu, Eli Shechtman, Phillip Isola, and Richard Zhang. Ensembling with deep generative views. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 431, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 286, + 464 + ], + "type": "text", + "content": "[10] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 465, + 286, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 286, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 286, + 497 + ], + "type": "text", + "content": "[11] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv preprint, arXiv:1907.06571, 2019. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 499, + 286, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 286, + 531 + ], + "type": "text", + "content": "[12] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. ArcFace: Additive angular margin loss for deep face recognition. In CVPR, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 533, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 286, + 566 + ], + "type": "text", + "content": "[13] Gereon Fox, Ayush Tewari, Mohamed Elgharib, and Christian Theobalt. StyleVideoGAN: A temporal generative model using a pretrained StyleGAN. In BMVC, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 613, + 286, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 634 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 635, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 667 + ], + "type": "text", + "content": "[16] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[17] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. DeeperForensics-1.0: A large-scale dataset for real-world face forgery detection. In CVPR, 2020. 2, 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "text", + "content": "[19] Emmanuel Kahembwe and Subramanian Ramamoorthy. Lower dimensional kernels for video discriminators. Neural Networks, 132:506-520, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[20] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In NeurIPS, 2020. 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[21] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 275 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "text", + "content": "[24] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. VideoFusion: Decomposed diffusion models for high-quality video generation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 354 + ], + "type": "text", + "content": "[25] Jacek Naruniec, Leonhard Helminger, Christopher Schroers, and Romann M Weber. High-resolution neural face swapping for visual effects. In CGF, 2020. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 388 + ], + "type": "text", + "content": "[26] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. StyleCLIP: Text-driven manipulation of StyleGAN imagery. In ICCV, 2021. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 388, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 421 + ], + "type": "text", + "content": "[27] Justin NM Pinkney and Doron Adler. Resolution dependent GAN interpolation for controllable image synthesis between domains. arXiv preprint, arXiv:2010.05334, 2020. 5, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "text", + "content": "[28] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "text", + "content": "[29] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a StyleGAN encoder for image-to-image translation. In CVPR, 2021. 2, 5, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "text", + "content": "[31] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint, arXiv:1803.09179, 2018. 2, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 580, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 611 + ], + "type": "text", + "content": "[32] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In ICCV, 2017. 1, 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "text", + "content": "[33] Masaki Saito, Shunta Saito, Masanori Koyama, and Sosuke Kobayashi. Train sparsely, generate densely: Memory-efficient unsupervised training of high-resolution temporal GAN. IJCV, 128:2586-2606, 2020. 1, 3, 5" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[34] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In NeurIPS, 2019. 2, 5, 9" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[35] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. StyleGAN-V: A continuous video generator with the" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22860" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "text", + "content": "price, image quality and perks of StyleGAN2. In CVPR, 2022. 2, 3, 4, 5, 6, 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "text", + "content": "[36] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. In ICLR, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "text", + "content": "[37] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for StyleGAN image manipulation. ACM TOG, 40:1-14, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 205 + ], + "type": "text", + "content": "[38] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. In CVPR, 2018. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 206, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 249 + ], + "type": "text", + "content": "[39] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint, arXiv:1812.01717, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 250, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 271 + ], + "type": "text", + "content": "[40] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. JMLR, 9(11), 2008. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[41] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 305, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 336 + ], + "type": "text", + "content": "[42] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. In NeurIPS, 2016. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 338, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 369 + ], + "type": "text", + "content": "[43] Hui-Po Wang, Ning Yu, and Mario Fritz. Hijack-GAN: Unintended-use of pretrained, black-box GANs. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "text", + "content": "[44] Tianyi Wei, Dongdong Chen, Wenbo Zhou, Jing Liao, Weiming Zhang, Lu Yuan, Gang Hua, and Nenghai Yu. E2Style: Improve the efficiency and effectiveness of StyleGAN inversion. TIP, 31:3267-3280, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "text", + "content": "[45] Zongze Wu, Dani Lischinski, and Eli Shechtman. StyleSpace analysis: Disentangled controls for StyleGAN image generation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[46] Weihao Xia, Yulun Zhang, Yujiu Yang, Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang. GAN inversion: A survey. TPAMI, 45:3121-3138, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 482, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 524 + ], + "type": "text", + "content": "[47] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In CVPR, 2018. 2, 5, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "type": "text", + "content": "[48] Yangyang Xu, Yong Du, Wenpeng Xiao, Xuemiao Xu, and Shengfeng He. From continuity to editability: Inverting GANs with consecutive images. In ICCV, pages 13910-13918, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[49] Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, and Bolei Zhou. Generative hierarchical features from synthesizing images. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "text", + "content": "[50] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video generation using VQ-VAE and Transformers. arXiv preprint, arXiv:2104.10157, 2021. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 668 + ], + "type": "text", + "content": "[51] Siyuan Yang, Lu Zhang, Yu Liu, Zhizhuo Jiang, and You He. Video diffusion models with local-global context guidance. In IJCAI, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[52] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In ICLR, 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 252 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[53] Qihang Zhang, Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Towards smooth video composition. In ICLR, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[54] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022.8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 183 + ], + "type": "text", + "content": "[55] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain GAN inversion for real image editing. In ECCV, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 185, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 218 + ], + "type": "text", + "content": "[56] Peihao Zhu, Rameen Abdal, Yipeng Qin, John Femiani, and Peter Wonka. Improved StyleGAN embedding: Where are the good latents? arXiv preprint, arXiv:2012.09036, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "text", + "content": "[57] Peiye Zhuang, Oluwasanmi Koyejo, and Alexander Schwing. Enjoy your editing: Controllable GANs for image editing via latent space navigation. In ICLR, 2021. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22861" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/35be9e19-3939-4344-a259-82c8df74d69e_content_list.json b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/35be9e19-3939-4344-a259-82c8df74d69e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..621868e627c11ef297ebf42f14ce14cd6b6332ad --- /dev/null +++ b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/35be9e19-3939-4344-a259-82c8df74d69e_content_list.json @@ -0,0 +1,1829 @@ +[ + { + "type": "text", + "text": "StyleLipSync: Style-based Personalized Lip-sync Video Generation", + "text_level": 1, + "bbox": [ + 145, + 130, + 823, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taekyung Ki $^{1\\star}$ Dongchan Min $^{2\\star}$ $^{1}$ AITRICS, $^{2}$ Graduate School of AI, KAIST", + "bbox": [ + 305, + 180, + 663, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tkki@aitrics.com, alsehdcks95@kaist.ac.kr https://stylelipsync.github.io", + "bbox": [ + 303, + 219, + 666, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present StyleLipSync, a style-based personalized lip-sync video generative model that can generate identity-agnostic lip-synchronizing video from arbitrary audio. To generate a video of arbitrary identities, we leverage expressive lip prior from the semantically rich latent space of a pre-trained StyleGAN, where we can also design a video consistency with a linear transformation. In contrast to the previous lip-sync methods, we introduce pose-aware masking that dynamically locates the mask to improve the naturalness over frames by utilizing a 3D parametric mesh predictor frame by frame. Moreover, we propose a few-shot lip-sync adaptation method for an arbitrary person by introducing a sync regularizer that preserves lip-sync generalization while enhancing the person-specific visual information. Extensive experiments demonstrate that our model can generate accurate lip-sync videos even with the zero-shot setting and enhance characteristics of an unseen face using a few seconds of target video through the proposed adaptation method.", + "bbox": [ + 75, + 318, + 473, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 636, + 209, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past few years, advances in deep learning have altered the dynamics of video creation. Now, users can easily make and edit videos with the help of deep learning. In particular, the task of generating a talking head video has received great interest due to its various practical uses. It can be applied in many applications such as film dubbing into a different language, face-to-face live chats, and virtual avatars in games and videos. Thus, a lot of prior works [21, 34, 28, 46, 45, 27] have been studied to generate a talking head video that has accurate lip shapes according to arbitrary audio inputs.", + "bbox": [ + 75, + 662, + 468, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most of the prior works mainly focus on enhancing synchronization between lip shapes and audio input. Some of the previous methods [46, 9, 34] use intermediate structural", + "bbox": [ + 75, + 829, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "representations such as landmarks and 3D models. They predicted the representations from the audio input and synthesized a talking head video of a target person. However, they suffered from inaccurate lip-sync results since such representations are too sparse to produce fine-grained details in lip-syncing. Recently, another line of methods [28, 27] mapped input audio to latent space and leveraged it to construct the mouth region of the target identity. While it achieves satisfactory results in lip-syncing, it generated blurry lower faces which are visually implausible. Furthermore, most methods only consider synthesizing frame-by-frame, lacking temporal consistency at the video level.", + "bbox": [ + 496, + 287, + 893, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose StyleLipSync, a style-based lip-sync video generative model that can generate identity-agnostic lip-synchronizing video from the arbitrary audio input. Our model consists of the following components. First, different from a previous masking method [21, 28, 27, 6] which masks the entire lower half face, we propose Pose-aware Masking. We analyze that the previous masking method cause unpleasant artifacts and unnatural jaw moving in the generated videos. To circumvent this, we utilize a 3D face mesh predictor [11, 23] and generate lip masks with consideration of pose information and facial semantics such as jaw shape. Second, our image decoder is based on a style-based generator, namely StyleGAN [18, 19, 17]. StyleGANs have demonstrated their effectiveness in various facial generative tasks, including face editing[1, 2], face enhancement [42], and video generation [35, 24]. As a pre-trained StyleGAN already contains expressive and diverse face priors in style latent space [1], we leverage it to synthesize the high-fidelity lip region of the target person. Furthermore, thanks to the continuous and linear nature of the latent space [18, 13, 32], we linearly manipulate the style codes using the audio input to generate lip-synced video frames. Additionally, we propose Style-aware Masked Fusion to effectively adopt a skipconnection to our decoder, which helps to preserve the 2D structure of the image and improves lip fidelity. Finally, we propose a Moving-average based Latent Smoothing module that makes the latent trajectory smoother for enhancing the", + "bbox": [ + 496, + 477, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22841", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "temporal consistency in the synthesized talking head video.", + "bbox": [ + 76, + 90, + 467, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While our model can synthesize a talking head video of the target person, there is a slight identity gap between a generated video and the target person. The gap can be noticeable, for example, in racial faces, which are relatively scarce in the training data. One approach to addressing this issue is to fine-tune the generator on a few seconds of video of the target person to create a personalized model. Several fine-tuning methods [30, 25, 37] have already been demonstrated and widely adopted by the industry to achieve product-level quality. However, we analyze that simply fine-tuning the generator loses its ability to generalize to arbitrary audio inputs, which is critical for generating talking head videos. Therefore, to minimize the side effect, we propose a sync regularizer enforcing the audio generalization performance. The key idea is to leverage the audio from the training data, not from the target video. Specifically, we not only optimize the generator to reconstruct the target video but also synthesize the video corresponding to the randomly sampled audio from the training data and maintain a sync correlation between the synthesized video and the audio. As a result, we obtain a personalized lip-sync generative model that can synthesize a video of the target person for arbitrary audio. Our contributions are summarized as follows:", + "bbox": [ + 75, + 106, + 468, + 450 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present StyleLipSync, a lip-sync video generative model which generates lip-synchronizing video in-the-wild of $256 \\times 256$ resolution with accurate and natural lip movement from a given masked video frames, audio segment, and single reference image.", + "- We additionally propose a few-shot adaptation method for entirely unseen faces, which uses only a few seconds of video by introducing a sync regularizer to maintain audio generalization.", + "- Experimental results demonstrate that StyleLipSync achieves state-of-the-art performance in terms of lip-sync and visual quality, even with the zero-shot setting." + ], + "bbox": [ + 96, + 457, + 467, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 656, + 225, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Lip-sync Video Generation", + "text_level": 1, + "bbox": [ + 76, + 681, + 318, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lip-sync video generation aims to generate a talking face video with lip motion synchronized with the given input audio. Early works [21, 28] generate lip from the lower half masked face image corresponding to the input audio. Specifically, Wav2Lip [28] uses a pre-trained SyncNet [8] as a lip-sync expert which maximizes the correlation between the generated lip and the input audio. Similarly, we use lip-sync expert for audio-visual alignment, which is trained with a contrastive manner proposed in [24]. VideoRetalking [6] improves Wav2Lip [28] a two-stage manner, where it first generates low-resolution $(96\\times 96)$ video and then increases the resolution by using a single identity specific super-resolution network. In contrast, we propose", + "bbox": [ + 75, + 704, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a zero-shot model that directly generates lip-sync video of $256 \\times 256$ resolution and also propose a unseen face adaptation method then enhances the personal characteristics from a few-shot target video. SyncTalkFace [27] introduces a lip memory network, which encodes lip motion features into a discrete space in the training phase and retrieves a lip feature from query audio at the inference phase. In contrast to [27], we utilize continuous and linear latent space from a pre-trained style-based generator [19] to generate lip images with high fidelity and take video consistency into account in the latent space.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. GAN Prior", + "text_level": 1, + "bbox": [ + 500, + 267, + 620, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Style-based generators [18, 19, 17] demonstrate the power of their semantic latent spaces, namely $\\mathcal{W}$ , in image generation, image editing [13, 32], and video generation [35]. GAN-inversion [12, 1, 2, 29, 36] utilizes the pre-trained GANs to invert an image into corresponding latent code so one can manipulate the attributes of the image only within the latent space. Extended $\\mathcal{W}+$ have been shown its much expressive power. For instance, pSp [29] adopts the feature pyramid networks (FPNs) [22] to use $\\mathcal{W}+$ , which follow the nature of the progressive generation [15] of StyleGAN [18] and achieves state-of-the-art performance in image-to-image translation (e.g., face inpainting). Similarly, we utilize $\\mathcal{W}+$ for diverse and strong lip prior since we aim to build a lip-sync video generative model of arbitrary identity.", + "bbox": [ + 496, + 291, + 890, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works [42, 38, 41] not only adopt pre-trained GAN prior as their decoder but also introduce a skip-connection that concatenates the encoded and generated features, which helps the model preserve 2D spatial information. Specifically, GPEN [42] uses skip-connection that concatenates the encoded and generated features and achieves state-of-the-art performance in blind face restoration. StyleSwap [41] adopts it to face swap and introduces the ToMask branch that predicts the target facial attribute regions for swapping in a supervised manner. In contrast to those methods, we use an additive skip connection more efficient than the concatenation, along with the unsupervised predicted masked sum, which helps the decoder distinguish the target lip region from the whole face and therefore increases the lip image fidelity.", + "bbox": [ + 496, + 518, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Personalization", + "text_level": 1, + "bbox": [ + 500, + 756, + 653, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although GAN prior has been successful in various tasks [29, 36], it is still challenging to faithfully recover person-specific information that lies in the out-of-distribution [1, 2, 29, 36, 42]. Recently, a few-shot personalization has become an alternative to solving the problem [25]. Pivotal-tuning-inversion (PTI) [30] fine-tunes the image generator while freezing a single latent code, namely pivot, to compensate the person-specific information in the generative", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22842", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dc8017fdc410f69ed32e3e1e934127fa55afbd6387e54c6add6305322fd74a59.jpg", + "image_caption": [ + "Figure 1. A framework of StyleLipSync. We leverage a 3D parametric mesh predictor [11, 23] to obtain pose-aware masked frames $X_{1:T}$ which inherits the facial pose of input frames. Face encoder $\\mathbf{E}_{face}$ maps $X_{1:T}$ into 2D spatial features and then fed into the decoder $\\mathbf{G}$ through style-aware masked fusion (SaMF). Single reference image $X_{ref}$ and audio segments $A_{1:T}$ are mapped into latent space, followed by Moving-average based Latent Smoothing (MaLS). This module outputs smooth video latent codes $\\tilde{w}_{1:T} \\subseteq \\mathcal{W}+$ that represent temporally consistent lip movement. With the guidance of SaMFs and the smooth video latent codes $\\tilde{w}_{1:T}$ , StyleLipSync can generate temporally consistent lip-synced videos." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 869, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "process, not in the encoding process. MyStyle [25] adopts PTI [30] to image inpainting and semantic editing, by restricting the latent space to a subspace spanned by the multiple pivots from the few photos (roughly 100) of the target person. Stitch-it-in-time [37] adopts the pivotal-tuning to the video editing in a multi-stage manner, which leverages an off-the-shelf latent manipulation [31] to manipulate video in the latent space and then stitch it to the source video. Inspired by them, we propose a few-shot unseen face adaptation method that slightly fine-tunes the image decoder for a given latent code trajectory of a target identity and maintains the audio generalization by introducing a sync-regularizer.", + "bbox": [ + 75, + 428, + 472, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 638, + 169, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given lip masked video frames $X_{1:T} = (X_t)_{t=1}^T$ and audio segments $A_{1:T} = (A_t)_{t=1}^T$ , a lip-sync method generates video frames", + "bbox": [ + 75, + 662, + 468, + 708 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {1: T} = (\\hat {X} _ {t}) _ {t = 1} ^ {T}, \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 708, + 468, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{X}_{1:T}$ has a lip movement synchronized with the audio segments $A_{1:T}$ . In contrast to the previous lip-sync methods [21, 28, 27, 6], we leverage 3D parametric facial mesh predictor [11, 23] to compute lip mask so that the generator can be aware of semantically meaningful facial pose information (Section 3.1). We utilize a pre-trained StyleGAN [19] as our decoder $\\mathbf{G}$ . When audio encoder $\\mathbf{E}_{\\text{aud}}$ and reference encoder $\\mathbf{E}_{\\text{ref}}$ map their inputs into the latent space $\\mathcal{W}+$ (Section 3.3), the decoder $\\mathbf{G}$ generates lip-synced video frames $\\hat{X}_{1:T}$ from these latent codes (Section 3.2) guided by the proposed style-aware masked fusion. For", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "enhancing the temporal consistency, we propose a Moving-average based Latent Smoothing module, which learns local motion between the latent codes, and makes video latent trajectory smoother. Finally, sync loss [8, 24] is used for the audio-lip synchronization. The overall framework of our model is described in Figure 1.", + "bbox": [ + 496, + 428, + 893, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pose-aware Masking", + "text_level": 1, + "bbox": [ + 498, + 529, + 696, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic head motion is an important factor in the natural talking style. However, existing methods [21, 28, 27, 6] employ the rectangular lower-half mouth masking method without consideration of the pose information. It often fails to detect appropriate masking regions when the head moves dynamically, which leads to unpleasant artifacts and unnatural jaw movement in the generated videos (see the first row in Figure 5 for examples). To address this limitation, we use the face meshes by leveraging a 3D face mesh predictor [11], which captures 3D parameters and predicts dense face geometry. We predict the 3D parameters [4] and the mesh from given video frames. Then, the predicted expression parameter $\\delta \\in \\mathbb{R}^{64}$ is used to adjust the mesh to obtain naturally opened and closed mouth meshes. We normalize the mesh vertices using the predicted pose parameters $\\tau \\in \\mathbb{R}^3$ (translation), $\\gamma \\in \\mathrm{SO}(3)$ (rotation) and leave only the lower frontal vertices. These meshes are combined and projected onto the original 2D plane to finally get our pose-aware lip masks. Figure 2 illustrates the framework of the pose-aware masking. This masking not only captures the pose information but also inherits facial semantics such as jaw shape. Ablation studies in Section 5.4 show that the pose-aware masking helps the model to increase visual quality along", + "bbox": [ + 496, + 553, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22843", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ba1161b1b65e071e93fb3c0d187d891285d98f1066ebb6afec5bc6fededcbfc0.jpg", + "image_caption": [ + "Figure 2. Illustration of pose-aware masking. The expression parameter $\\delta \\in \\mathbb{R}^{64}$ and the pose parameter $\\tau \\in \\mathbb{R}^3$ , $\\gamma \\in \\mathrm{SO}(3)$ are used to compute the natural mask." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 472, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with dynamic pose.", + "bbox": [ + 76, + 263, + 209, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.Decoder", + "text_level": 1, + "bbox": [ + 76, + 287, + 179, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lip Prior from Style-based Generator. Generating the lip-sync videos from scratch for an arbitrary person is very hard since the mapping from audio to lip is basically one-to-many. In this paper, we leverage a style-based generator as our image decoder $\\mathbf{G}$ for the following two reasons. First, a pre-trained StyleGAN already contains expressive and diverse face priors [18, 19, 17] in the form of latent code, namely style code in the latent space $\\mathcal{W} + [2]$ . Those latent spaces enable us to better synthesize the lip region of the target person with the diverse lip prior. Second, the style codes form the continuous and linear [18, 13, 32] latent spaces, which enables us to design a high-level visual transformation, such as natural motion, only with a linear transformation of latent code [39]. Hence, we can generate a talking head video with smooth lip motion by simply manipulating the style codes using audio, which the previous lip-sync methods never take into account.", + "bbox": [ + 75, + 311, + 468, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style-aware Masked Fusion (SaMF). Recently, it has been explored that adopting skip-connections based on the concatenation to GAN-inversion helps to preserve the 2D spatial information of the input [41, 14, 42]. Similarly, we adopt an additive skip-connection to our model for effectively preserving the non-masked region of $X_{1:T}$ while faithfully utilizing the latent space.", + "bbox": [ + 75, + 568, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we propose style-aware masked fusion (SaMF) for efficiently preserving the 2D spatial feature and relieving the information gap between masked and non-masked regions. SaMFs are introduced at the beginning of the decoder blocks. The decoder $\\mathbf{G}$ consists of $L$ decoder blocks, each of which takes 2 style codes to modulate 3 convolution weights, as illustrated in Figure 3. The first style code in each decoder block modulates 2 different convolution weights, one for the convolution in the original block and the other for the SaMF. SaMF learns to predict a 1-channel mask $S_{t}^{l}$ of the current resolution from the encoded feature through the newly modulated convolution followed by the sigmoid, which is used for spatial weighted fusion of the encoded feature and the generated feature.", + "bbox": [ + 75, + 674, + 470, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given encoded feature $\\mathbf{E}_{face}^{l}(X_{t})$ and gen", + "bbox": [ + 94, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/30c7eef7cdb0eb181f35e80e5a3b167e23dd49eafbc9cfa07c176ef3e4493479.jpg", + "image_caption": [ + "Figure 3. Illustration of the decoder block. The encoded feature $\\mathbf{E}_{face}^{l}(X_{t})$ is injected into $l$ -th decoder block through Style-aware Masked Fusion (SaMF). Note that only the convolutions in SaMF are trainable, while the others are frozen during the training phase." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 890, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "erated feature $\\mathbf{G}^{l - 1}(X_t)$ of same dimension $\\mathbb{R}^{h\\times w\\times c}$ , SaMF firstly predicts a spatial mask $S_{t}^{l}\\in \\mathbb{R}^{h\\times w\\times 1}$ from $\\mathbf{E}_{face}^{l}(X_{t})$ and then output fused feature as follows:", + "bbox": [ + 498, + 362, + 890, + 411 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t} ^ {l} \\otimes \\mathbf {E} _ {\\text {f a c e}} ^ {l} \\left(X _ {t}\\right) + \\left(1 - S _ {t} ^ {l}\\right) \\otimes \\mathbf {G} ^ {l - 1} \\left(X _ {t}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 417, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ablation studies (Section 5.4) show that SaMFs improve the fidelity of the mouth since they separate masked and non-masked regions.", + "bbox": [ + 498, + 441, + 890, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.Encoders", + "text_level": 1, + "bbox": [ + 500, + 496, + 609, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model has three different encoders: face encoder $\\mathbf{E}_{face}$ , reference encoder $\\mathbf{E}_{ref}$ , and audio encoder $\\mathbf{E}_{aud}$ .", + "bbox": [ + 500, + 518, + 890, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Face encoder $\\mathbf{E}_{face}$ takes masked video frames $X_{1:T}$ as input and then outputs $l$ 2D spatial features $\\mathbf{E}_{face}^{l}(X_{t}) = \\{\\mathbf{E}_{face}^{l}(X_{t})\\mid l\\in [1,2,\\dots ,L]\\}$ for each $t$ . These features are injected into the decoder $\\mathbf{G}$ through the style-aware masked fusion to efficiently preserve 2D spatial structure, as described in Section 3.2.", + "bbox": [ + 498, + 550, + 890, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reference Encoder $\\mathbf{E}_{ref}$ maps a reference $X_{ref}$ into $2L$ reference style codes, each of which has 512 dimensions. We simply denote the reference style code as $w_{ref} = [w_{ref}^{1}|w_{ref}^{2}|\\dots |w_{ref}^{2L}] \\in \\mathbb{R}^{512 \\times 2L}$ .", + "bbox": [ + 498, + 642, + 890, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to $\\mathbf{E}_{ref}$ , audio encoder $\\mathbf{E}_{aud}$ maps a single audio segment into $2L$ audio style codes, each of which has 512 dimensions. As we use $T$ consecutive audio segments $A_{1:T}$ , $\\mathbf{E}_{aud}$ independently maps $A_{t}$ into $a_{t} = [a_{t}^{1}|a_{t}^{2}|\\dots |a_{t}^{2L}] \\in \\mathbb{R}^{512 \\times 2L}$ . We simply denote $a_{1:T} = (a_{t})_{t=1}^{T}$ for $T$ audio style codes. Please refer to our supplementary materials for the detailed encoder architectures.", + "bbox": [ + 498, + 703, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From these style codes $w_{ref}$ and $a_{1:T}$ , we compute target video's style codes over frames $w_{1:T}$ by", + "bbox": [ + 498, + 809, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {1: T} = \\left(w _ {1}, w _ {2}, \\dots , w _ {T}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 845, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{t} = w_{ref} + a_{t}\\in \\mathcal{W}+$ . We compute the style codes by simply adding these two different codes based on the", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22844", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "linearity [13, 32] of the latent space $\\mathcal{W}+$ . The style codes $w_{1:T}$ are then fed into the decoder $\\mathbf{G}$ through the learned affine transformations to generate synced lip motion.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Consistency. Thanks to the semantically rich latent space $\\mathcal{W}+$ , our model can generate an accurate lip-sync video frame by frame. However, this frame-wisely independent encoding of style codes turns out to lead inconsistent mouth movements in the final results. To remedy this, we assume that the generated style codes $w_{1:T}$ form a trajectory of the target video [40] in $\\mathcal{W}+$ and enforce the smooth local transition of the motion [24, 33, 39] into the trajectory. Toward this, we introduce Moving-average based Latent Smoothing (MaLS), each of which consists of a stack of the weighted moving-average [3] and 1D convolutions operating on the style codes along the time-axis. More precisely, we employ $2L$ MaLS for $L$ resolutions, each of which takes the $l$ -th component of $w_{t-1}, w_t$ , and $w_{t+1}$ as its inputs to learn the local difference between them, and then we inject the local motions into $w_{ref}$ to compute the smooth style code $\\tilde{w}_t$ :", + "bbox": [ + 75, + 137, + 470, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {w} _ {t} ^ {l} = w _ {r e f} ^ {l} + \\mathrm {M a L S} ^ {l} (w _ {t - 1: t + 1} ^ {l}), \\\\ = w _ {r e f} ^ {l} + \\operatorname {C o n v 1 D s} \\left(\\sum_ {\\tau = t - 1} ^ {t + 1} \\alpha_ {\\tau} \\cdot w _ {\\tau} ^ {l}\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 398, + 468, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{MaLS}^l$ denotes the $l$ -th MaLS, and $\\alpha_{\\tau}$ is the weight of the moving-average.", + "bbox": [ + 76, + 470, + 468, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With this smooth latent codes $\\tilde{w}_{1:T}$ , we compute the video frames:", + "bbox": [ + 76, + 502, + 468, + 531 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {1: T} = \\left(\\mathbf {G} \\left(\\tilde {w} _ {t}, \\mathbf {E} _ {\\text {f a c e}} \\left(X _ {t}\\right)\\right)\\right) _ {t = 1} ^ {T}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 537, + 468, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For better initialization [29], we add the average code $w_{avg}$ of the pre-trained generator to each $\\tilde{w}_t$ in (4).", + "bbox": [ + 76, + 563, + 467, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Objective", + "text_level": 1, + "bbox": [ + 76, + 601, + 258, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train StyleLipSync to reconstruct target video frames from corresponding audio. We randomly choose $T = 5$ consequent frames with corresponding audio segments and 1 single reference frame. Image perceptual loss [43] is used to minimize perceptual image distance between generated frames $\\hat{X}$ and ground-truth frames $Y$ :", + "bbox": [ + 76, + 625, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l p i p s} = \\sum_ {i = 1} ^ {N} \\left\\| \\phi^ {i} (\\hat {X}) - \\phi^ {i} (Y) \\right\\| _ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 722, + 468, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the number of feature extractor, $\\phi^i (\\cdot)$ is the $i$ th feature extractor, and $\\| \\cdot \\| _2$ is the $\\ell^2$ loss. Similar to [33], we use the multi-scale perceptual loss with 3 levels.", + "bbox": [ + 76, + 770, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For audio-visual alignment, we utilize SyncNet trained in a contrastive manner [24] that minimizes the cosine distance between generated frames $\\hat{X}$ and corresponding audio segment $A$ :", + "bbox": [ + 76, + 816, + 468, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s y n c}} = 1 - \\cos \\left(f _ {v} (\\hat {X}), f _ {a} (A)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1a5d8d6b21a61f6c4547432ad4bfd5c27165049cda41b35edd0040e6ed0528ad.jpg", + "image_caption": [ + "Figure 4. Adaptation for Unseen Face. We slightly tune the decoder $\\mathbf{G}_{\\theta}$ with the proposed sync regularizer $\\mathcal{R}_{sync}$ , while freezing all encoders' weight. Face encoder $\\mathbf{E}_{face}$ and SaMFs are omitted here for simplicity." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 890, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ denotes the cosine similarity. $f_{v}(\\cdot)$ and $f_{a}(\\cdot)$ are the frame and audio feature extractor, respectively. The final objective $\\mathcal{L}_{train}$ is computed as:", + "bbox": [ + 498, + 279, + 890, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a i n}} = \\lambda_ {1} \\mathcal {L} _ {\\text {l p i p s}} + \\lambda_ {2} \\mathcal {L} _ {\\text {s y n c}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 334, + 890, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are the balancing coefficients.", + "bbox": [ + 500, + 359, + 810, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Unseen Face Adaptation", + "text_level": 1, + "bbox": [ + 500, + 386, + 728, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although StyleLipSync successfully generates accurately lip-synced videos with high fidelity, the model would fail to exactly synthesize unseen faces lying in the out-of-distribution. This problem refers to the failure of id preservation. Therefore, to handle this, we fine-tune our decoder $\\mathbf{G}$ on a target person video to make a personal model that is better able to synthesize toward the target person.", + "bbox": [ + 496, + 411, + 890, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $X_{1:T}$ be masked video frames of unseen face, which corresponds to the audio segments $A_{1:T}$ , and $X_{ref}$ be a reference frame. The frozen encoders convert each input into the intermediate representations, $\\tilde{w}_{1:T}$ and $\\mathbf{E}_{face}(X_t))_{t=1}^T$ . Then we fine-tune the decoder $\\mathbf{G}_{\\theta}$ , now parameterized by $\\theta$ , by minimizing the distance between $(\\mathbf{G}_{\\theta}(\\tilde{w}_t,\\mathbf{E}_{face}(X_t)))_{t=1}^T$ and target frames, as same as (8). However, fine-tuning the decoder on a short video of a single identity leads to over-fitting and losing the lip-sync generality as the generator can memorize the target video [26]. To prevent the model from this scenario, we introduce a sync regularizer $\\mathcal{R}_{sync}$ to enforce audio generality to the decoder $\\mathbf{G}_{\\theta}$ by leveraging the audio from the training dataset, not from the target video. Formally, given audio segments $A_{1:T}^{\\prime}$ randomly chosen from the training data (Voxceleb2 [7]) and $w_{ref}$ , we compute smooth style codes $\\tilde{w}_{1:T}^{\\prime}$ , and then decode them to a synced video $X_{1:T}^{\\prime}$ . The sync regularizer $\\mathcal{R}_{sync}$ is defined as", + "bbox": [ + 496, + 517, + 892, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {\\text {s y n c}} = 1 - \\cos \\left(f _ {v} \\left(\\hat {X} ^ {\\prime}\\right), f _ {a} \\left(A ^ {\\prime}\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 797, + 890, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which enforces $\\mathbf{G}_{\\theta}$ to generate $\\hat{X}_{1:T}^{\\prime}$ aligning with $A_{1:T}^{\\prime}$ . The final object for a single person adaptation is given as follows:", + "bbox": [ + 496, + 824, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\underset {\\theta} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\text {t r a i n}} + \\lambda_ {R} \\mathcal {R} _ {\\text {s y n c}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 878, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22845", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/61ffa634159844f0c6b05d9e17ce424811b8b2a60edffcbb4166e3ea5f1ed34c.jpg", + "image_caption": [ + "(a) Reconstruction results on Voxceleb2 [7] test data.", + "Figure 5. Comparison with state-of-the-art methods. The different field of view comes from the pre-processing strategy of each model." + ], + "image_footnote": [], + "bbox": [ + 86, + 95, + 480, + 470 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/78368105a11b5e3fa251e766852549e3bd6d103febcc4822321b71cd11d95889.jpg", + "image_caption": [ + "(b) Cross-id results on HDTF [44]." + ], + "image_footnote": [], + "bbox": [ + 491, + 95, + 885, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_R$ is the regularizer coefficient. Ablation studies in Section 5.4 show that $\\mathcal{R}_{sync}$ regularizes the audio generality even if we use the audio from the training set.", + "bbox": [ + 76, + 523, + 468, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 580, + 209, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 604, + 171, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model on Voxceleb2 [7], which consists of in-the-wild talking face videos collected from YouTube. It contains more than 145,000 videos of about 6,100 identities in the train set and more than 4,900 videos of more than 110 identities in the test set. We convert all videos into frames with 25 fps and then crop and resize those frames into $256 \\times 256$ resolution, following the method of [33]. For more semantically rich face priors, we only use videos where the detected face bounding boxes' height (and width) is longer than 256. After pre-processing, the remaining 11051 videos and 340 videos are used as the training set and the test set, respectively. All audios are re-sampled to $16\\mathrm{kHz}$ and converted into a mel-spectrogram to be used as our audio representation similar to the method in [45, 24]. We also use HDTF [44] to further test our model for crossid experiments. It is widely used to evaluate high-resolution talking face generative models, where the head pose dynamics are not as significant as the Voxceleb2.", + "bbox": [ + 75, + 628, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 523, + 718, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We pre-train StyleGAN2 [19] on the Voxceleb2 [7] following the implementation of [16]. For training StyleLipSync and SyncNet, we set the frame length $T = 5$ , and employ Adam [20] optimizer with learning rate $10^{-4}$ throughout the training phase in both cases. All experiments are performed on 2 TITAN RTX GPUs. We set $\\lambda_{1} = 10$ , $\\lambda_{2} = 0.1$ , $\\lambda_{R} = 0.1$ for the objective (10), and $\\alpha_{t-1} = 0.25$ , $\\alpha_{t} = 0.5$ , $\\alpha_{t+1} = 0.25$ for MaLS (4). For all inference, we use the first frame as the reference image.", + "bbox": [ + 496, + 547, + 890, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For image quality metric, we use PSNR, and SSIM. We also calculate CSIM [10] in the cross-id experiment that measures the face similarity between the images in the pretrained face embedding space. For lip-sync quality metric, we use LMD, LSE-D, and LSE-C. LMD [5] is the absolute distance of facial landmarks between the target and generated frames. LSE-D and LSE-C are proposed in [28], where LSE-D measures the distance between the lip and audio representations, and LSE-C measures the lip-sync confidence, respectively.", + "bbox": [ + 496, + 684, + 892, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 845, + 620, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reconstruction Results. We compare the other state-of-the-art methods in the lip-sync (Wav2Lip [28]) and talking", + "bbox": [ + 500, + 869, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22846", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "face generation (ATVG [5], MakeItTalk [46], and PC-AVS [45]) on reconstruction results of Voxceleb2 test dataset. Table 1 shows that StyleLipSync outperforms all metrics except the LSE-C score. Wav2Lip [28] achieves the highest LSE-C score, however, it achieves low image quality metrics since it generates $96 \\times 96$ low-resolution videos. PC-AVS [45] achieves comparable lip-sync scores with StyleLipSync, however, achieves low image quality metrics than our model since it highly relies on its specific preprocessing and fails to generate in many cases. We also illustrate the qualitative results in Figure 4. Wav2Lip [28] produces a lip-synchronized video with considerable visual artifacts since it cannot adapt to the video of dynamic head pose. MakeItTalk [46] generates low lip-sync video since it uses sparse facial landmarks. PC-AVS [45] generates an accurate lip-synchronized video following the input head pose. However, it struggles to preserve the unseen identity involving visual artifacts. StyleLipSync generates the natural lip motion with high fidelity and preserves the input identity, which is comparable to the ground truth video.", + "bbox": [ + 75, + 90, + 472, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross Identity Results. To evaluate lip-sync generalization, we conduct cross-id experiments settings using unseen videos and audio. We randomly sample 10 videos of different identities and 150 audio without duplication from HDTF [44]. We use the first 10 seconds of videos and audio. For each video, we generate 15 lip-synced videos from 15 different audios, where those face-audio pairs are not from the same source. In Table 2, we report the LSE-D and LSE-C for lip-sync quality and the CSIM for face similarity. Wav2Lip [28] achieves a higher LSE-C score than StyleLipSync, however, it achieves a low CSIM score since it produces a low-resolution video with visual artifacts. MakeItTalk [46] achieves the best CSIM score, while the lip-sync quality is the worst since it uses sparse facial landmarks. PC-AVS [45] outperforms LSE-C while achieving the lowest CSIM since it can't preserve an unseen face's identity. StyleLipSync achieves the best LSE-D score and comparable CSIM score to MakeItTalk [46]. We show qualitative results of cross-id experiments with target lip references in Figure 4. PC-AVS [45] can generate accurate lip-synchronizing video compared to the target lip, while it fails to preserve facial details of the unseen face. MakeItTalk [46] produces a high-resolution and identity-preserving video, however, it is out-of-sync compared to the target. StyleLipSync generates a high-resolution lip synchronizing video compared to the target lip, without any visual artifacts.", + "bbox": [ + 75, + 395, + 472, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 815, + 243, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Studies on Zero-shot Model. Figure 6 and Table 3 summarize the ablation studies on our zero-shot method of reconstruction on Voxceleb2 test data. If we replace the pose-aware masking with standard rectangular", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6978e33c12a382fd8c505b0c36d3c2716c37d7b65d67e210c42d7102449856e.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of reconstruction on Voxceleb2 test data. The best score for each metric is in bold." + ], + "table_footnote": [], + "table_body": "
MethodVoxceleb2 (Reconstruction)
ImageLip-Sync
SSIM↑PSNR↑LMD↓LSE-D↓LSE-C↑
Wav2Lip96×96 [28]0.44813.5346.4226.9998.329
ATVG128×128 [5]0.46113.3497.1658.8215.421
MakeItTalk256×256 [46]0.41912.8683.64910.8953.624
PC-AVS224×224 [45]0.36913.2102.8127.2787.699
Ours256×2560.63119.6072.6966.6288.056
", + "bbox": [ + 503, + 123, + 893, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2c250f0d338fe21a19be6621f6bb277c0c6bd914cc7f47fcfd60b565c5f08dec.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of cross-identity results for unseen face. We report CSIM [10] as the image quality metric since there is no ground truth frames for the cross-id experiments. The best score for each metric is in bold." + ], + "table_footnote": [], + "table_body": "
MethodHDTF (Cross-id)
ImageLip-Sync
CSIM ↑LSE-D ↓LSE-C ↑
Wav2Lip96×96 [28]0.6567.0478.576
ATVG128×128 [5]0.2878.6686.040
MakeItTalk256×256 [46]0.77010.6414.725
PC-AVS224×224 [45]0.2386.9218.858
Ours256×2560.7376.8258.209
", + "bbox": [ + 509, + 292, + 883, + 414 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ebd4a00f648c629b285e20069f2cd95d64324f3ca4d80202ce20212112b912d2.jpg", + "table_caption": [ + "Table 3. Ablation study on zero-shot model. The best score for each metric is in **bold**." + ], + "table_footnote": [], + "table_body": "
Method (ours)Voceleb2 (Reconstruction)
ImageLip-Sync
SSIM ↑PSNR ↑LMD ↓LSE-D ↓LSE-C ↓
w/o Pose Mask0.60218.8673.0576.7717.748
w/o MaLS0.59318.1862.7406.9947.577
w/o SaMF0.59118.1812.7646.8387.780
Full0.63119.6072.6966.6288.056
", + "bbox": [ + 506, + 462, + 885, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "masking (w/o Pose Mask) in [28, 6, 27], considerable visual artifacts occur around the masked region since it is insufficient to capture the pose difference between the reference and the target. To validate SaMF, we replace the modulated convolutions in SaMFs with the standard convolutions. Figure 6(c) shows that SaMFs significantly improve lip region's fidelity since the modulated convolution helps to be aware of the lip style. As shown in Table 3, MaLS significantly improves lip-sync quality, which cannot be reflected in a single image. Please refers to our project page for ablation studies on MaLS.", + "bbox": [ + 496, + 580, + 890, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Studies on Unseen Face Adaptation. We conduct ablation studies on the proposed unseen face adaptation following the same setting in Table 2. Additionally, we use 60 seconds of video for each 10 personalized models and 15 audios of 10 seconds from different identities for inference. Figure 7(b) shows the lip-sync metrics according to the adaptation step. In the cases without the sync regularizer, the models lose the lip-sync generality, in other words, it memorizes the short target video as the adaptation phase proceeds. Introducing the sync regularizer with the sync", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22847", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/698d1f2001ae813ddf0bc7faa0819f15a4a6327ef2ad943f78924cd17ba9b61e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 90, + 173, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4708a0e1698c232cded03e42c3e435857ef6d8c8d9ac3aee3b0401285f920c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 90, + 271, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/619085aa427a0a96bb5669e0deb157594248c33aee01de3928ce3a7976d193de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 90, + 370, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b8d79bff406be7ef24efd9c84151e682448a47a3b79e53d3e6a35f4e0f455b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 90, + 467, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48b1c0f911668b356bf3202d0eb401bb8369a9db52e7ad656b2529e6453708af.jpg", + "image_caption": [ + "(a) Reference" + ], + "image_footnote": [], + "bbox": [ + 81, + 164, + 173, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51a18be1cd5024a406193b05721bcf4696e7ad7879c3d4ac875519e16b0f41ff.jpg", + "image_caption": [ + "(b) w/o Pose mask" + ], + "image_footnote": [], + "bbox": [ + 179, + 164, + 271, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/404f867ccb7b672d7bf5b7d2562e8d44493f7c58932ddcf5402b070e39821a03.jpg", + "image_caption": [ + "(c) w/o SaMF" + ], + "image_footnote": [], + "bbox": [ + 277, + 164, + 370, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/beace11418acd0a701ff56659bb5fbd5c76762bf6938d414471860ffe3c566be.jpg", + "image_caption": [ + "(d) Full", + "Figure 6. Qualitative comparison of zero-shot model." + ], + "image_footnote": [], + "bbox": [ + 375, + 164, + 467, + 236 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ad6a4c5ca6ff945801715a2bce8843df697cb1c5aeea983b42aa8d19e0c54276.jpg", + "table_caption": [ + "Table 4. Mean Opinion score (MOS) user study results with $95\\%$ confidence interval on cross-id setting. The score ranges in 1 to 5. The best score for each metric is in bold." + ], + "table_footnote": [], + "table_body": "
MethodUser Study (MOS)
Lip-sync AccuracyFace SimilarityVisual Quality
Wav2Lip [28]3.76 ± 0.202.98 ± 0.252.03 ± 0.21
ATVG [5]2.19 ± 0.212.45 ± 0.251.54 ± 0.17
MakeItTalk [46]2.32 ± 0.243.47 ± 0.232.95 ± 0.24
PC-AVS [45]3.28 ± 0.212.55 ± 0.222.51 ± 0.22
Ours4.01 ± 0.163.42 ± 0.223.55 ± 0.20
Ours (Personalized)3.52 ± 0.214.03 ± 0.173.64 ± 0.19
", + "bbox": [ + 84, + 330, + 460, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "loss stabilizes the lip-sync metrics and even improves them compared to the zero-shot results. If we use the sync regularizer without the sync loss, lip-sync quality is stabilized, however slightly lower than the zero-shot results due to the lack of ground-truth audio-visual correlation. Audio generalization for unseen face data is maintained even though we used audio from the learning data. Figure 7(a) supports the validity of the adaptation method. It shows that visual difference between the zero-shot results and the adaptation results. The zero-shot model can generate an accurate lip motion for the target audio, while it shows a slight difference in person-specific details compared to the reference images. Through the proposed adaptation method, personal-specific lip shape, teeth, and wrinkles are faithfully recovered while maintaining the lip motion of the zero-shot results.", + "bbox": [ + 75, + 472, + 467, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. User Study", + "text_level": 1, + "bbox": [ + 76, + 710, + 197, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further conduct a user study based on MOS (Mean opinion score) to compare the perceptual quality of each model, including our zero-shot and personalized model. 5 videos generated by each model in cross-id setting are used for this study. 20 participants scored lip-sync accuracy, face similarity, and visual quality of each video in the range of 1 to 5. As shown in Table 4, our models outperform all metrics. Specifically, our zero-shot model achieves the highest lip-sync accuracy, and our adaptation model achieves the highest score in face similarity and visual quality with competitive lip-sync accuracy.", + "bbox": [ + 75, + 734, + 467, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b822265a98179096967236193e26baf6594f688045c67653f7cab4a9f704b1b3.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 504, + 99, + 578, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f294511969cd368b685fbefd6096572670fe578201e4260dcfac6118b720cab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 99, + 612, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6cc51cc1828e797f22d508ade1792ee6550fb1af8676a064d05e789732ebf0b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 99, + 653, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/138cb89af7151cfd97a94847703ebb601befeb5c89d6388d9c95d5cb1af7a857.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 99, + 717, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b9d0caf51bc87ac06851bc69d154c9ca29b7e64ee1c3b91295994bcf60736b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 99, + 885, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1296d6d1f743fec74111cbfa48291a7ddcb0b50b8a8303963c96da3617c243af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 159, + 602, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b33774d7de86e7eac1bd941f7a06a64137f38960dbf4df109dc7660bcd91c2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 159, + 647, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5c51ddc39c4015f8c831c6eb8d33c5c8a82554020d9e845eb08c58545bab80fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 159, + 712, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed7fdd7a259014ca277e120771bd702cda400120f06a5b792889c4cc10cc2eea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 159, + 885, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7cac37e4dfb3a2d148ec5686177cb512a1ba646973558181088ac76c61b8ec5a.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 504, + 232, + 578, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/381fc29f755cc4b721118146ac6f62961058002e8c3e07e2fcc73ade466d27cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 232, + 612, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a28cda9da671a4dd4ace4823b7b930f9089cf94fef925aded126d0869d7c2fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 232, + 653, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1ac1c3637c4abd36a4ff76f6b663453ddba4e6b761cff16e23cb55f376572163.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 232, + 717, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/820a54989705e6f0b4ab47b889ec449dca315dd045bc27cd5a1349f193695c12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 232, + 885, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b77a2a94484b0da1c715e87fbbcd0590ba3ccb2304ecb088a63e2d5cce42cf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 292, + 602, + 324 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40994e0b5b587add63603696f30d796ba6ddb8a44992ff2d638f6080d2242180.jpg", + "image_caption": [ + "(a) Comparison of cross-id results of zero-shot and adaptation." + ], + "image_footnote": [], + "bbox": [ + 604, + 292, + 656, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec22210d6ed95457a1ac66d5d5f47f98f7a50c4d6b4d4c1fc38d5797d1842306.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 292, + 718, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/546dea6b7e3c73acf0e549d0d130dea73563c92340cbed7c016ce47afe784631.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 292, + 885, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4a6aa5c2fc86a5b5c34f4eba7acc9f621bdb2de268c3042f674f966967d5c890.jpg", + "image_caption": [ + "(b) Ablation study on the proposed sync regularizer.", + "Figure 7. Experimental results of the proposed adaptation method using the cross-id setting. It improves the person specific visuals while maintaining the lip-sync generality." + ], + "image_footnote": [], + "bbox": [ + 504, + 386, + 696, + 512 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/54c0008f606aed12cfba9029416b59432a1a18c28678d5e913e887b81ee7618e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 387, + 885, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 613, + 617, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed StyleLipSync, a lip-sync video generative model for arbitrary identity, which leverages expressive lip priors from a pre-trained style-based generator. In contrast to existing lip-sync generative models, we introduce pose-aware masking for lip region by utilizing a 3D parametric mesh predictor, which inherits the pose information in the mask itself. Designing a smooth lip motion by using the moving-average based latent smoothing in the continuous and linear latent space, StyleLipSync can generate temporally consistent lip motion. Furthermore, we propose a few-shot lip-sync adaptation method for a single person who lies in the out-of-distribution, which uses a few seconds of the target person's video. Experimental results show that our StyleLipSync can generate realistic lip-sync video from arbitrary audio even with the zero-shot setting, and the proposed adaptation method enhances the person-specific visual information without losing the lip-sync generality.", + "bbox": [ + 496, + 643, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22848", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Future Works. Since learning audiovisual representation requires a large number of different identities (e.g., Voxceleb2 [7]), extending our method to higher resolution $(512 \\times 512 \\uparrow)$ is very challenging. We consider it our future work to develop an effective encoding system for a limited number of identities using a style-based generator [18, 19, 17]. Designing an effective reference encoder to improve lip identity preservation in a zero-shot setting can be another future work.", + "bbox": [ + 75, + 90, + 468, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ethical Considerations. Since our method can generate a video of a specific person talking specific words only with a few seconds of video source, it has the potential for misuse. As discussed in [6], attaching visual watermarks on the generated videos can be a solution to it.", + "bbox": [ + 76, + 226, + 468, + 301 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 315, + 173, + 329 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019.", + "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020.", + "[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022.", + "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, pages 187-194, 1999.", + "[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019.", + "[6] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022.", + "[7] J. S. Chung, A. Nagrani, and A. Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018.", + "[8] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision, pages 251-263, 2016.", + "[9] Dipanjan Das, S. Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision, 2020.", + "[10] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019." + ], + "bbox": [ + 78, + 339, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019.", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 2014.", + "[13] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020.", + "[14] Jingwen He, Wu Shi, Kai Chen, Lean Fu, and Chao Dong. Gcfsr: a generative and controllable face super resolution method without facial and gan priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1898, 2022.", + "[15] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017.", + "[16] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020.", + "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021.", + "[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019.", + "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020.", + "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.", + "[21] Prajwal KR, Rudrabha Mukhopadhyay, Jerin Philip, Abhishek Jha, Vinay Namboodiri, and CV Jawahar. Towards automatic face-to-face translation. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1428-1436, 2019.", + "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017.", + "[23] Camillo Lugasre, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22849", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Dongchan Min, Minyoung Song, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. arXiv preprint arXiv:2208.10922, 2022.", + "[25] Yotam Nitzan, Kfir Aberman, Qiurui He, Orly Liba, Michal Yarom, Yossi Gandelsman, Inbar Mosseri, Yael Pritch, and Daniel Cohen-Or. Mystyle: A personalized generative prior. arXiv preprint arXiv:2203.17272, 2022.", + "[26] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021.", + "[27] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In 36th AAAI Conference on Artificial Intelligence (AAAI 22). Association for the Advancement of Artificial Intelligence, 2022.", + "[28] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020.", + "[29] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021.", + "[30] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022.", + "[31] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020.", + "[32] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021.", + "[33] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in Neural Information Processing Systems, 32, 2019.", + "[34] Justus Thies, Mohamed A. Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision, 2019.", + "[35] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021.", + "[36] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021.", + "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022.", + "[38] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9168-9178, 2021.", + "[39] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations, 2022.", + "[40] Weihao Xia, Yujiu Yang, and Jing-Hao Xue. Gan inversion for consistent video interpolation and manipulation. arXiv preprint arXiv:2208.11197, 2022.", + "[41] Zhiliang Xu, Hang Zhou, Zhibin Hong, Ziwei Liu, Jiaming Liu, Zhizhi Guo, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Styleswap: Style-based generator empowers robust face swapping. In European Conference on Computer Vision, pages 661-677, 2022.", + "[42] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021.", + "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018.", + "[44] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021.", + "[45] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021.", + "[46] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020." + ], + "bbox": [ + 503, + 92, + 890, + 698 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22850", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/35be9e19-3939-4344-a259-82c8df74d69e_model.json b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/35be9e19-3939-4344-a259-82c8df74d69e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3318f8c892962e21518add7a68dbcc77dc3ac28 --- /dev/null +++ b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/35be9e19-3939-4344-a259-82c8df74d69e_model.json @@ -0,0 +1,2420 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.825, + 0.154 + ], + "angle": 0, + "content": "StyleLipSync: Style-based Personalized Lip-sync Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.181, + 0.664, + 0.217 + ], + "angle": 0, + "content": "Taekyung Ki\\(^{1\\star}\\) Dongchan Min\\(^{2\\star}\\) \n\\(^{1}\\)AITRICS, \\(^{2}\\)Graduate School of AI, KAIST" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.22, + 0.668, + 0.252 + ], + "angle": 0, + "content": "tkki@aitrics.com, alsehdcks95@kaist.ac.kr https://stylelipsync.github.io" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.606 + ], + "angle": 0, + "content": "In this paper, we present StyleLipSync, a style-based personalized lip-sync video generative model that can generate identity-agnostic lip-synchronizing video from arbitrary audio. To generate a video of arbitrary identities, we leverage expressive lip prior from the semantically rich latent space of a pre-trained StyleGAN, where we can also design a video consistency with a linear transformation. In contrast to the previous lip-sync methods, we introduce pose-aware masking that dynamically locates the mask to improve the naturalness over frames by utilizing a 3D parametric mesh predictor frame by frame. Moreover, we propose a few-shot lip-sync adaptation method for an arbitrary person by introducing a sync regularizer that preserves lip-sync generalization while enhancing the person-specific visual information. Extensive experiments demonstrate that our model can generate accurate lip-sync videos even with the zero-shot setting and enhance characteristics of an unseen face using a few seconds of target video through the proposed adaptation method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.637, + 0.21, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.829 + ], + "angle": 0, + "content": "In the past few years, advances in deep learning have altered the dynamics of video creation. Now, users can easily make and edit videos with the help of deep learning. In particular, the task of generating a talking head video has received great interest due to its various practical uses. It can be applied in many applications such as film dubbing into a different language, face-to-face live chats, and virtual avatars in games and videos. Thus, a lot of prior works [21, 34, 28, 46, 45, 27] have been studied to generate a talking head video that has accurate lip shapes according to arbitrary audio inputs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.83, + 0.471, + 0.876 + ], + "angle": 0, + "content": "Most of the prior works mainly focus on enhancing synchronization between lip shapes and audio input. Some of the previous methods [46, 9, 34] use intermediate structural" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.471 + ], + "angle": 0, + "content": "representations such as landmarks and 3D models. They predicted the representations from the audio input and synthesized a talking head video of a target person. However, they suffered from inaccurate lip-sync results since such representations are too sparse to produce fine-grained details in lip-syncing. Recently, another line of methods [28, 27] mapped input audio to latent space and leveraged it to construct the mouth region of the target identity. While it achieves satisfactory results in lip-syncing, it generated blurry lower faces which are visually implausible. Furthermore, most methods only consider synthesizing frame-by-frame, lacking temporal consistency at the video level." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose StyleLipSync, a style-based lip-sync video generative model that can generate identity-agnostic lip-synchronizing video from the arbitrary audio input. Our model consists of the following components. First, different from a previous masking method [21, 28, 27, 6] which masks the entire lower half face, we propose Pose-aware Masking. We analyze that the previous masking method cause unpleasant artifacts and unnatural jaw moving in the generated videos. To circumvent this, we utilize a 3D face mesh predictor [11, 23] and generate lip masks with consideration of pose information and facial semantics such as jaw shape. Second, our image decoder is based on a style-based generator, namely StyleGAN [18, 19, 17]. StyleGANs have demonstrated their effectiveness in various facial generative tasks, including face editing[1, 2], face enhancement [42], and video generation [35, 24]. As a pre-trained StyleGAN already contains expressive and diverse face priors in style latent space [1], we leverage it to synthesize the high-fidelity lip region of the target person. Furthermore, thanks to the continuous and linear nature of the latent space [18, 13, 32], we linearly manipulate the style codes using the audio input to generate lip-synced video frames. Additionally, we propose Style-aware Masked Fusion to effectively adopt a skipconnection to our decoder, which helps to preserve the 2D structure of the image and improves lip fidelity. Finally, we propose a Moving-average based Latent Smoothing module that makes the latent trajectory smoother for enhancing the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22841" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.106 + ], + "angle": 0, + "content": "temporal consistency in the synthesized talking head video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.452 + ], + "angle": 0, + "content": "While our model can synthesize a talking head video of the target person, there is a slight identity gap between a generated video and the target person. The gap can be noticeable, for example, in racial faces, which are relatively scarce in the training data. One approach to addressing this issue is to fine-tune the generator on a few seconds of video of the target person to create a personalized model. Several fine-tuning methods [30, 25, 37] have already been demonstrated and widely adopted by the industry to achieve product-level quality. However, we analyze that simply fine-tuning the generator loses its ability to generalize to arbitrary audio inputs, which is critical for generating talking head videos. Therefore, to minimize the side effect, we propose a sync regularizer enforcing the audio generalization performance. The key idea is to leverage the audio from the training data, not from the target video. Specifically, we not only optimize the generator to reconstruct the target video but also synthesize the video corresponding to the randomly sampled audio from the training data and maintain a sync correlation between the synthesized video and the audio. As a result, we obtain a personalized lip-sync generative model that can synthesize a video of the target person for arbitrary audio. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.458, + 0.468, + 0.532 + ], + "angle": 0, + "content": "- We present StyleLipSync, a lip-sync video generative model which generates lip-synchronizing video in-the-wild of \\(256 \\times 256\\) resolution with accurate and natural lip movement from a given masked video frames, audio segment, and single reference image." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.536, + 0.468, + 0.595 + ], + "angle": 0, + "content": "- We additionally propose a few-shot adaptation method for entirely unseen faces, which uses only a few seconds of video by introducing a sync regularizer to maintain audio generalization." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.599, + 0.468, + 0.643 + ], + "angle": 0, + "content": "- Experimental results demonstrate that StyleLipSync achieves state-of-the-art performance in terms of lip-sync and visual quality, even with the zero-shot setting." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.458, + 0.468, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.657, + 0.227, + 0.672 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.32, + 0.697 + ], + "angle": 0, + "content": "2.1. Lip-sync Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Lip-sync video generation aims to generate a talking face video with lip motion synchronized with the given input audio. Early works [21, 28] generate lip from the lower half masked face image corresponding to the input audio. Specifically, Wav2Lip [28] uses a pre-trained SyncNet [8] as a lip-sync expert which maximizes the correlation between the generated lip and the input audio. Similarly, we use lip-sync expert for audio-visual alignment, which is trained with a contrastive manner proposed in [24]. VideoRetalking [6] improves Wav2Lip [28] a two-stage manner, where it first generates low-resolution \\((96\\times 96)\\) video and then increases the resolution by using a single identity specific super-resolution network. In contrast, we propose" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "a zero-shot model that directly generates lip-sync video of \\(256 \\times 256\\) resolution and also propose a unseen face adaptation method then enhances the personal characteristics from a few-shot target video. SyncTalkFace [27] introduces a lip memory network, which encodes lip motion features into a discrete space in the training phase and retrieves a lip feature from query audio at the inference phase. In contrast to [27], we utilize continuous and linear latent space from a pre-trained style-based generator [19] to generate lip images with high fidelity and take video consistency into account in the latent space." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.268, + 0.622, + 0.283 + ], + "angle": 0, + "content": "2.2. GAN Prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.519 + ], + "angle": 0, + "content": "Style-based generators [18, 19, 17] demonstrate the power of their semantic latent spaces, namely \\(\\mathcal{W}\\), in image generation, image editing [13, 32], and video generation [35]. GAN-inversion [12, 1, 2, 29, 36] utilizes the pre-trained GANs to invert an image into corresponding latent code so one can manipulate the attributes of the image only within the latent space. Extended \\(\\mathcal{W}+\\) have been shown its much expressive power. For instance, pSp [29] adopts the feature pyramid networks (FPNs) [22] to use \\(\\mathcal{W}+\\), which follow the nature of the progressive generation [15] of StyleGAN [18] and achieves state-of-the-art performance in image-to-image translation (e.g., face inpainting). Similarly, we utilize \\(\\mathcal{W}+\\) for diverse and strong lip prior since we aim to build a lip-sync video generative model of arbitrary identity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.746 + ], + "angle": 0, + "content": "Recent works [42, 38, 41] not only adopt pre-trained GAN prior as their decoder but also introduce a skip-connection that concatenates the encoded and generated features, which helps the model preserve 2D spatial information. Specifically, GPEN [42] uses skip-connection that concatenates the encoded and generated features and achieves state-of-the-art performance in blind face restoration. StyleSwap [41] adopts it to face swap and introduces the ToMask branch that predicts the target facial attribute regions for swapping in a supervised manner. In contrast to those methods, we use an additive skip connection more efficient than the concatenation, along with the unsupervised predicted masked sum, which helps the decoder distinguish the target lip region from the whole face and therefore increases the lip image fidelity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.655, + 0.771 + ], + "angle": 0, + "content": "2.3. Personalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although GAN prior has been successful in various tasks [29, 36], it is still challenging to faithfully recover person-specific information that lies in the out-of-distribution [1, 2, 29, 36, 42]. Recently, a few-shot personalization has become an alternative to solving the problem [25]. Pivotal-tuning-inversion (PTI) [30] fine-tunes the image generator while freezing a single latent code, namely pivot, to compensate the person-specific information in the generative" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22842" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.087, + 0.87, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.32, + 0.896, + 0.404 + ], + "angle": 0, + "content": "Figure 1. A framework of StyleLipSync. We leverage a 3D parametric mesh predictor [11, 23] to obtain pose-aware masked frames \\( X_{1:T} \\) which inherits the facial pose of input frames. Face encoder \\( \\mathbf{E}_{face} \\) maps \\( X_{1:T} \\) into 2D spatial features and then fed into the decoder \\( \\mathbf{G} \\) through style-aware masked fusion (SaMF). Single reference image \\( X_{ref} \\) and audio segments \\( A_{1:T} \\) are mapped into latent space, followed by Moving-average based Latent Smoothing (MaLS). This module outputs smooth video latent codes \\( \\tilde{w}_{1:T} \\subseteq \\mathcal{W}+ \\) that represent temporally consistent lip movement. With the guidance of SaMFs and the smooth video latent codes \\( \\tilde{w}_{1:T} \\), StyleLipSync can generate temporally consistent lip-synced videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.473, + 0.626 + ], + "angle": 0, + "content": "process, not in the encoding process. MyStyle [25] adopts PTI [30] to image inpainting and semantic editing, by restricting the latent space to a subspace spanned by the multiple pivots from the few photos (roughly 100) of the target person. Stitch-it-in-time [37] adopts the pivotal-tuning to the video editing in a multi-stage manner, which leverages an off-the-shelf latent manipulation [31] to manipulate video in the latent space and then stitch it to the source video. Inspired by them, we propose a few-shot unseen face adaptation method that slightly fine-tunes the image decoder for a given latent code trajectory of a target identity and maintains the audio generalization by introducing a sync-regularizer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.639, + 0.17, + 0.654 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.47, + 0.709 + ], + "angle": 0, + "content": "Given lip masked video frames \\( X_{1:T} = (X_t)_{t=1}^T \\) and audio segments \\( A_{1:T} = (A_t)_{t=1}^T \\), a lip-sync method generates video frames" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.709, + 0.47, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {1: T} = (\\hat {X} _ {t}) _ {t = 1} ^ {T}, \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.904 + ], + "angle": 0, + "content": "where \\(\\hat{X}_{1:T}\\) has a lip movement synchronized with the audio segments \\(A_{1:T}\\). In contrast to the previous lip-sync methods [21, 28, 27, 6], we leverage 3D parametric facial mesh predictor [11, 23] to compute lip mask so that the generator can be aware of semantically meaningful facial pose information (Section 3.1). We utilize a pre-trained StyleGAN [19] as our decoder \\(\\mathbf{G}\\). When audio encoder \\(\\mathbf{E}_{\\text{aud}}\\) and reference encoder \\(\\mathbf{E}_{\\text{ref}}\\) map their inputs into the latent space \\(\\mathcal{W}+\\) (Section 3.3), the decoder \\(\\mathbf{G}\\) generates lip-synced video frames \\(\\hat{X}_{1:T}\\) from these latent codes (Section 3.2) guided by the proposed style-aware masked fusion. For" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.895, + 0.52 + ], + "angle": 0, + "content": "enhancing the temporal consistency, we propose a Moving-average based Latent Smoothing module, which learns local motion between the latent codes, and makes video latent trajectory smoother. Finally, sync loss [8, 24] is used for the audio-lip synchronization. The overall framework of our model is described in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.53, + 0.697, + 0.547 + ], + "angle": 0, + "content": "3.1. Pose-aware Masking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Dynamic head motion is an important factor in the natural talking style. However, existing methods [21, 28, 27, 6] employ the rectangular lower-half mouth masking method without consideration of the pose information. It often fails to detect appropriate masking regions when the head moves dynamically, which leads to unpleasant artifacts and unnatural jaw movement in the generated videos (see the first row in Figure 5 for examples). To address this limitation, we use the face meshes by leveraging a 3D face mesh predictor [11], which captures 3D parameters and predicts dense face geometry. We predict the 3D parameters [4] and the mesh from given video frames. Then, the predicted expression parameter \\(\\delta \\in \\mathbb{R}^{64}\\) is used to adjust the mesh to obtain naturally opened and closed mouth meshes. We normalize the mesh vertices using the predicted pose parameters \\(\\tau \\in \\mathbb{R}^3\\) (translation), \\(\\gamma \\in \\mathrm{SO}(3)\\) (rotation) and leave only the lower frontal vertices. These meshes are combined and projected onto the original 2D plane to finally get our pose-aware lip masks. Figure 2 illustrates the framework of the pose-aware masking. This masking not only captures the pose information but also inherits facial semantics such as jaw shape. Ablation studies in Section 5.4 show that the pose-aware masking helps the model to increase visual quality along" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22843" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.473, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.216, + 0.47, + 0.258 + ], + "angle": 0, + "content": "Figure 2. Illustration of pose-aware masking. The expression parameter \\(\\delta \\in \\mathbb{R}^{64}\\) and the pose parameter \\(\\tau \\in \\mathbb{R}^3\\), \\(\\gamma \\in \\mathrm{SO}(3)\\) are used to compute the natural mask." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.265, + 0.21, + 0.28 + ], + "angle": 0, + "content": "with dynamic pose." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.289, + 0.18, + 0.303 + ], + "angle": 0, + "content": "3.2.Decoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.47, + 0.568 + ], + "angle": 0, + "content": "Lip Prior from Style-based Generator. Generating the lip-sync videos from scratch for an arbitrary person is very hard since the mapping from audio to lip is basically one-to-many. In this paper, we leverage a style-based generator as our image decoder \\(\\mathbf{G}\\) for the following two reasons. First, a pre-trained StyleGAN already contains expressive and diverse face priors [18, 19, 17] in the form of latent code, namely style code in the latent space \\(\\mathcal{W} + [2]\\). Those latent spaces enable us to better synthesize the lip region of the target person with the diverse lip prior. Second, the style codes form the continuous and linear [18, 13, 32] latent spaces, which enables us to design a high-level visual transformation, such as natural motion, only with a linear transformation of latent code [39]. Hence, we can generate a talking head video with smooth lip motion by simply manipulating the style codes using audio, which the previous lip-sync methods never take into account." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.674 + ], + "angle": 0, + "content": "Style-aware Masked Fusion (SaMF). Recently, it has been explored that adopting skip-connections based on the concatenation to GAN-inversion helps to preserve the 2D spatial information of the input [41, 14, 42]. Similarly, we adopt an additive skip-connection to our model for effectively preserving the non-masked region of \\( X_{1:T} \\) while faithfully utilizing the latent space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.885 + ], + "angle": 0, + "content": "Specifically, we propose style-aware masked fusion (SaMF) for efficiently preserving the 2D spatial feature and relieving the information gap between masked and non-masked regions. SaMFs are introduced at the beginning of the decoder blocks. The decoder \\(\\mathbf{G}\\) consists of \\(L\\) decoder blocks, each of which takes 2 style codes to modulate 3 convolution weights, as illustrated in Figure 3. The first style code in each decoder block modulates 2 different convolution weights, one for the convolution in the original block and the other for the SaMF. SaMF learns to predict a 1-channel mask \\(S_{t}^{l}\\) of the current resolution from the encoded feature through the newly modulated convolution followed by the sigmoid, which is used for spatial weighted fusion of the encoded feature and the generated feature." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.885, + 0.47, + 0.904 + ], + "angle": 0, + "content": "Formally, given encoded feature \\(\\mathbf{E}_{face}^{l}(X_{t})\\) and gen" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.891, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.301, + 0.893, + 0.357 + ], + "angle": 0, + "content": "Figure 3. Illustration of the decoder block. The encoded feature \\(\\mathbf{E}_{face}^{l}(X_{t})\\) is injected into \\(l\\)-th decoder block through Style-aware Masked Fusion (SaMF). Note that only the convolutions in SaMF are trainable, while the others are frozen during the training phase." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.363, + 0.892, + 0.412 + ], + "angle": 0, + "content": "erated feature \\(\\mathbf{G}^{l - 1}(X_t)\\) of same dimension \\(\\mathbb{R}^{h\\times w\\times c}\\), SaMF firstly predicts a spatial mask \\(S_{t}^{l}\\in \\mathbb{R}^{h\\times w\\times 1}\\) from \\(\\mathbf{E}_{face}^{l}(X_{t})\\) and then output fused feature as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.418, + 0.892, + 0.438 + ], + "angle": 0, + "content": "\\[\nS _ {t} ^ {l} \\otimes \\mathbf {E} _ {\\text {f a c e}} ^ {l} \\left(X _ {t}\\right) + \\left(1 - S _ {t} ^ {l}\\right) \\otimes \\mathbf {G} ^ {l - 1} \\left(X _ {t}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.443, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Ablation studies (Section 5.4) show that SaMFs improve the fidelity of the mouth since they separate masked and non-masked regions." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.497, + 0.61, + 0.511 + ], + "angle": 0, + "content": "3.3.Encoders" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.52, + 0.892, + 0.551 + ], + "angle": 0, + "content": "Our model has three different encoders: face encoder \\(\\mathbf{E}_{face}\\), reference encoder \\(\\mathbf{E}_{ref}\\), and audio encoder \\(\\mathbf{E}_{aud}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.551, + 0.892, + 0.642 + ], + "angle": 0, + "content": "Face encoder \\(\\mathbf{E}_{face}\\) takes masked video frames \\(X_{1:T}\\) as input and then outputs \\(l\\) 2D spatial features \\(\\mathbf{E}_{face}^{l}(X_{t}) = \\{\\mathbf{E}_{face}^{l}(X_{t})\\mid l\\in [1,2,\\dots ,L]\\}\\) for each \\(t\\). These features are injected into the decoder \\(\\mathbf{G}\\) through the style-aware masked fusion to efficiently preserve 2D spatial structure, as described in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.643, + 0.892, + 0.705 + ], + "angle": 0, + "content": "Reference Encoder \\(\\mathbf{E}_{ref}\\) maps a reference \\(X_{ref}\\) into \\(2L\\) reference style codes, each of which has 512 dimensions. We simply denote the reference style code as \\(w_{ref} = [w_{ref}^{1}|w_{ref}^{2}|\\dots |w_{ref}^{2L}] \\in \\mathbb{R}^{512 \\times 2L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Similar to \\(\\mathbf{E}_{ref}\\), audio encoder \\(\\mathbf{E}_{aud}\\) maps a single audio segment into \\(2L\\) audio style codes, each of which has 512 dimensions. As we use \\(T\\) consecutive audio segments \\(A_{1:T}\\), \\(\\mathbf{E}_{aud}\\) independently maps \\(A_{t}\\) into \\(a_{t} = [a_{t}^{1}|a_{t}^{2}|\\dots |a_{t}^{2L}] \\in \\mathbb{R}^{512 \\times 2L}\\). We simply denote \\(a_{1:T} = (a_{t})_{t=1}^{T}\\) for \\(T\\) audio style codes. Please refer to our supplementary materials for the detailed encoder architectures." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.84 + ], + "angle": 0, + "content": "From these style codes \\( w_{ref} \\) and \\( a_{1:T} \\), we compute target video's style codes over frames \\( w_{1:T} \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.847, + 0.892, + 0.865 + ], + "angle": 0, + "content": "\\[\nw _ {1: T} = \\left(w _ {1}, w _ {2}, \\dots , w _ {T}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(w_{t} = w_{ref} + a_{t}\\in \\mathcal{W}+\\). We compute the style codes by simply adding these two different codes based on the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22844" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "linearity [13, 32] of the latent space \\(\\mathcal{W}+\\). The style codes \\(w_{1:T}\\) are then fed into the decoder \\(\\mathbf{G}\\) through the learned affine transformations to generate synced lip motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.394 + ], + "angle": 0, + "content": "Temporal Consistency. Thanks to the semantically rich latent space \\(\\mathcal{W}+\\), our model can generate an accurate lip-sync video frame by frame. However, this frame-wisely independent encoding of style codes turns out to lead inconsistent mouth movements in the final results. To remedy this, we assume that the generated style codes \\(w_{1:T}\\) form a trajectory of the target video [40] in \\(\\mathcal{W}+\\) and enforce the smooth local transition of the motion [24, 33, 39] into the trajectory. Toward this, we introduce Moving-average based Latent Smoothing (MaLS), each of which consists of a stack of the weighted moving-average [3] and 1D convolutions operating on the style codes along the time-axis. More precisely, we employ \\(2L\\) MaLS for \\(L\\) resolutions, each of which takes the \\(l\\)-th component of \\(w_{t-1}, w_t\\), and \\(w_{t+1}\\) as its inputs to learn the local difference between them, and then we inject the local motions into \\(w_{ref}\\) to compute the smooth style code \\(\\tilde{w}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.399, + 0.47, + 0.464 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {w} _ {t} ^ {l} = w _ {r e f} ^ {l} + \\mathrm {M a L S} ^ {l} (w _ {t - 1: t + 1} ^ {l}), \\\\ = w _ {r e f} ^ {l} + \\operatorname {C o n v 1 D s} \\left(\\sum_ {\\tau = t - 1} ^ {t + 1} \\alpha_ {\\tau} \\cdot w _ {\\tau} ^ {l}\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.471, + 0.469, + 0.502 + ], + "angle": 0, + "content": "where \\(\\mathrm{MaLS}^l\\) denotes the \\(l\\)-th MaLS, and \\(\\alpha_{\\tau}\\) is the weight of the moving-average." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.503, + 0.469, + 0.532 + ], + "angle": 0, + "content": "With this smooth latent codes \\(\\tilde{w}_{1:T}\\), we compute the video frames:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.538, + 0.469, + 0.558 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {1: T} = \\left(\\mathbf {G} \\left(\\tilde {w} _ {t}, \\mathbf {E} _ {\\text {f a c e}} \\left(X _ {t}\\right)\\right)\\right) _ {t = 1} ^ {T}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.564, + 0.468, + 0.595 + ], + "angle": 0, + "content": "For better initialization [29], we add the average code \\( w_{avg} \\) of the pre-trained generator to each \\( \\tilde{w}_t \\) in (4)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.602, + 0.259, + 0.619 + ], + "angle": 0, + "content": "3.4. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.625, + 0.469, + 0.716 + ], + "angle": 0, + "content": "We train StyleLipSync to reconstruct target video frames from corresponding audio. We randomly choose \\( T = 5 \\) consequent frames with corresponding audio segments and 1 single reference frame. Image perceptual loss [43] is used to minimize perceptual image distance between generated frames \\( \\hat{X} \\) and ground-truth frames \\( Y \\):" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.723, + 0.469, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l p i p s} = \\sum_ {i = 1} ^ {N} \\left\\| \\phi^ {i} (\\hat {X}) - \\phi^ {i} (Y) \\right\\| _ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.771, + 0.469, + 0.817 + ], + "angle": 0, + "content": "where \\(N\\) is the number of feature extractor, \\(\\phi^i (\\cdot)\\) is the \\(i\\) th feature extractor, and \\(\\| \\cdot \\| _2\\) is the \\(\\ell^2\\) loss. Similar to [33], we use the multi-scale perceptual loss with 3 levels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.818, + 0.469, + 0.877 + ], + "angle": 0, + "content": "For audio-visual alignment, we utilize SyncNet trained in a contrastive manner [24] that minimizes the cosine distance between generated frames \\(\\hat{X}\\) and corresponding audio segment \\(A\\):" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.884, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s y n c}} = 1 - \\cos \\left(f _ {v} (\\hat {X}), f _ {a} (A)\\right), \\tag {7}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.892, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Adaptation for Unseen Face. We slightly tune the decoder \\(\\mathbf{G}_{\\theta}\\) with the proposed sync regularizer \\(\\mathcal{R}_{sync}\\), while freezing all encoders' weight. Face encoder \\(\\mathbf{E}_{face}\\) and SaMFs are omitted here for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.28, + 0.892, + 0.327 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) denotes the cosine similarity. \\(f_{v}(\\cdot)\\) and \\(f_{a}(\\cdot)\\) are the frame and audio feature extractor, respectively. The final objective \\(\\mathcal{L}_{train}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.335, + 0.892, + 0.352 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a i n}} = \\lambda_ {1} \\mathcal {L} _ {\\text {l p i p s}} + \\lambda_ {2} \\mathcal {L} _ {\\text {s y n c}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.36, + 0.812, + 0.375 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are the balancing coefficients." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.387, + 0.729, + 0.404 + ], + "angle": 0, + "content": "4. Unseen Face Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.518 + ], + "angle": 0, + "content": "Although StyleLipSync successfully generates accurately lip-synced videos with high fidelity, the model would fail to exactly synthesize unseen faces lying in the out-of-distribution. This problem refers to the failure of id preservation. Therefore, to handle this, we fine-tune our decoder \\(\\mathbf{G}\\) on a target person video to make a personal model that is better able to synthesize toward the target person." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.893, + 0.791 + ], + "angle": 0, + "content": "Let \\(X_{1:T}\\) be masked video frames of unseen face, which corresponds to the audio segments \\(A_{1:T}\\), and \\(X_{ref}\\) be a reference frame. The frozen encoders convert each input into the intermediate representations, \\(\\tilde{w}_{1:T}\\) and \\(\\mathbf{E}_{face}(X_t))_{t=1}^T\\). Then we fine-tune the decoder \\(\\mathbf{G}_{\\theta}\\), now parameterized by \\(\\theta\\), by minimizing the distance between \\((\\mathbf{G}_{\\theta}(\\tilde{w}_t,\\mathbf{E}_{face}(X_t)))_{t=1}^T\\) and target frames, as same as (8). However, fine-tuning the decoder on a short video of a single identity leads to over-fitting and losing the lip-sync generality as the generator can memorize the target video [26]. To prevent the model from this scenario, we introduce a sync regularizer \\(\\mathcal{R}_{sync}\\) to enforce audio generality to the decoder \\(\\mathbf{G}_{\\theta}\\) by leveraging the audio from the training dataset, not from the target video. Formally, given audio segments \\(A_{1:T}^{\\prime}\\) randomly chosen from the training data (Voxceleb2 [7]) and \\(w_{ref}\\), we compute smooth style codes \\(\\tilde{w}_{1:T}^{\\prime}\\), and then decode them to a synced video \\(X_{1:T}^{\\prime}\\). The sync regularizer \\(\\mathcal{R}_{sync}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.798, + 0.891, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {\\text {s y n c}} = 1 - \\cos \\left(f _ {v} \\left(\\hat {X} ^ {\\prime}\\right), f _ {a} \\left(A ^ {\\prime}\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.87 + ], + "angle": 0, + "content": "which enforces \\(\\mathbf{G}_{\\theta}\\) to generate \\(\\hat{X}_{1:T}^{\\prime}\\) aligning with \\(A_{1:T}^{\\prime}\\). The final object for a single person adaptation is given as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.88, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\underset {\\theta} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\text {t r a i n}} + \\lambda_ {R} \\mathcal {R} _ {\\text {s y n c}}, \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22845" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.097, + 0.482, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.478, + 0.424, + 0.491 + ], + "angle": 0, + "content": "(a) Reconstruction results on Voxceleb2 [7] test data." + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.096, + 0.887, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.478, + 0.78, + 0.491 + ], + "angle": 0, + "content": "(b) Cross-id results on HDTF [44]." + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.496, + 0.871, + 0.51 + ], + "angle": 0, + "content": "Figure 5. Comparison with state-of-the-art methods. The different field of view comes from the pre-processing strategy of each model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.525, + 0.47, + 0.571 + ], + "angle": 0, + "content": "where \\(\\lambda_R\\) is the regularizer coefficient. Ablation studies in Section 5.4 show that \\(\\mathcal{R}_{sync}\\) regularizes the audio generality even if we use the audio from the training set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.581, + 0.21, + 0.599 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.606, + 0.173, + 0.621 + ], + "angle": 0, + "content": "5.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.901 + ], + "angle": 0, + "content": "We train our model on Voxceleb2 [7], which consists of in-the-wild talking face videos collected from YouTube. It contains more than 145,000 videos of about 6,100 identities in the train set and more than 4,900 videos of more than 110 identities in the test set. We convert all videos into frames with 25 fps and then crop and resize those frames into \\(256 \\times 256\\) resolution, following the method of [33]. For more semantically rich face priors, we only use videos where the detected face bounding boxes' height (and width) is longer than 256. After pre-processing, the remaining 11051 videos and 340 videos are used as the training set and the test set, respectively. All audios are re-sampled to \\(16\\mathrm{kHz}\\) and converted into a mel-spectrogram to be used as our audio representation similar to the method in [45, 24]. We also use HDTF [44] to further test our model for crossid experiments. It is widely used to evaluate high-resolution talking face generative models, where the head pose dynamics are not as significant as the Voxceleb2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.719, + 0.54 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.683 + ], + "angle": 0, + "content": "We pre-train StyleGAN2 [19] on the Voxceleb2 [7] following the implementation of [16]. For training StyleLipSync and SyncNet, we set the frame length \\( T = 5 \\), and employ Adam [20] optimizer with learning rate \\( 10^{-4} \\) throughout the training phase in both cases. All experiments are performed on 2 TITAN RTX GPUs. We set \\( \\lambda_{1} = 10 \\), \\( \\lambda_{2} = 0.1 \\), \\( \\lambda_{R} = 0.1 \\) for the objective (10), and \\( \\alpha_{t-1} = 0.25 \\), \\( \\alpha_{t} = 0.5 \\), \\( \\alpha_{t+1} = 0.25 \\) for MaLS (4). For all inference, we use the first frame as the reference image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.893, + 0.837 + ], + "angle": 0, + "content": "For image quality metric, we use PSNR, and SSIM. We also calculate CSIM [10] in the cross-id experiment that measures the face similarity between the images in the pretrained face embedding space. For lip-sync quality metric, we use LMD, LSE-D, and LSE-C. LMD [5] is the absolute distance of facial landmarks between the target and generated frames. LSE-D and LSE-C are proposed in [28], where LSE-D measures the distance between the lip and audio representations, and LSE-C measures the lip-sync confidence, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.622, + 0.862 + ], + "angle": 0, + "content": "5.3. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Reconstruction Results. We compare the other state-of-the-art methods in the lip-sync (Wav2Lip [28]) and talking" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22846" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.394 + ], + "angle": 0, + "content": "face generation (ATVG [5], MakeItTalk [46], and PC-AVS [45]) on reconstruction results of Voxceleb2 test dataset. Table 1 shows that StyleLipSync outperforms all metrics except the LSE-C score. Wav2Lip [28] achieves the highest LSE-C score, however, it achieves low image quality metrics since it generates \\(96 \\times 96\\) low-resolution videos. PC-AVS [45] achieves comparable lip-sync scores with StyleLipSync, however, achieves low image quality metrics than our model since it highly relies on its specific preprocessing and fails to generate in many cases. We also illustrate the qualitative results in Figure 4. Wav2Lip [28] produces a lip-synchronized video with considerable visual artifacts since it cannot adapt to the video of dynamic head pose. MakeItTalk [46] generates low lip-sync video since it uses sparse facial landmarks. PC-AVS [45] generates an accurate lip-synchronized video following the input head pose. However, it struggles to preserve the unseen identity involving visual artifacts. StyleLipSync generates the natural lip motion with high fidelity and preserves the input identity, which is comparable to the ground truth video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.473, + 0.804 + ], + "angle": 0, + "content": "Cross Identity Results. To evaluate lip-sync generalization, we conduct cross-id experiments settings using unseen videos and audio. We randomly sample 10 videos of different identities and 150 audio without duplication from HDTF [44]. We use the first 10 seconds of videos and audio. For each video, we generate 15 lip-synced videos from 15 different audios, where those face-audio pairs are not from the same source. In Table 2, we report the LSE-D and LSE-C for lip-sync quality and the CSIM for face similarity. Wav2Lip [28] achieves a higher LSE-C score than StyleLipSync, however, it achieves a low CSIM score since it produces a low-resolution video with visual artifacts. MakeItTalk [46] achieves the best CSIM score, while the lip-sync quality is the worst since it uses sparse facial landmarks. PC-AVS [45] outperforms LSE-C while achieving the lowest CSIM since it can't preserve an unseen face's identity. StyleLipSync achieves the best LSE-D score and comparable CSIM score to MakeItTalk [46]. We show qualitative results of cross-id experiments with target lip references in Figure 4. PC-AVS [45] can generate accurate lip-synchronizing video compared to the target lip, while it fails to preserve facial details of the unseen face. MakeItTalk [46] produces a high-resolution and identity-preserving video, however, it is out-of-sync compared to the target. StyleLipSync generates a high-resolution lip synchronizing video compared to the target lip, without any visual artifacts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.245, + 0.832 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Ablation Studies on Zero-shot Model. Figure 6 and Table 3 summarize the ablation studies on our zero-shot method of reconstruction on Voxceleb2 test data. If we replace the pose-aware masking with standard rectangular" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.088, + 0.895, + 0.116 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of reconstruction on Voxceleb2 test data. The best score for each metric is in bold." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.124, + 0.895, + 0.217 + ], + "angle": 0, + "content": "
MethodVoxceleb2 (Reconstruction)
ImageLip-Sync
SSIM↑PSNR↑LMD↓LSE-D↓LSE-C↑
Wav2Lip96×96 [28]0.44813.5346.4226.9998.329
ATVG128×128 [5]0.46113.3497.1658.8215.421
MakeItTalk256×256 [46]0.41912.8683.64910.8953.624
PC-AVS224×224 [45]0.36913.2102.8127.2787.699
Ours256×2560.63119.6072.6966.6288.056
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.229, + 0.895, + 0.286 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of cross-identity results for unseen face. We report CSIM [10] as the image quality metric since there is no ground truth frames for the cross-id experiments. The best score for each metric is in bold." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.294, + 0.885, + 0.415 + ], + "angle": 0, + "content": "
MethodHDTF (Cross-id)
ImageLip-Sync
CSIM ↑LSE-D ↓LSE-C ↑
Wav2Lip96×96 [28]0.6567.0478.576
ATVG128×128 [5]0.2878.6686.040
MakeItTalk256×256 [46]0.77010.6414.725
PC-AVS224×224 [45]0.2386.9218.858
Ours256×2560.7376.8258.209
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.426, + 0.895, + 0.454 + ], + "angle": 0, + "content": "Table 3. Ablation study on zero-shot model. The best score for each metric is in **bold**." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.463, + 0.887, + 0.555 + ], + "angle": 0, + "content": "
Method (ours)Voceleb2 (Reconstruction)
ImageLip-Sync
SSIM ↑PSNR ↑LMD ↓LSE-D ↓LSE-C ↓
w/o Pose Mask0.60218.8673.0576.7717.748
w/o MaLS0.59318.1862.7406.9947.577
w/o SaMF0.59118.1812.7646.8387.780
Full0.63119.6072.6966.6288.056
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.748 + ], + "angle": 0, + "content": "masking (w/o Pose Mask) in [28, 6, 27], considerable visual artifacts occur around the masked region since it is insufficient to capture the pose difference between the reference and the target. To validate SaMF, we replace the modulated convolutions in SaMFs with the standard convolutions. Figure 6(c) shows that SaMFs significantly improve lip region's fidelity since the modulated convolution helps to be aware of the lip style. As shown in Table 3, MaLS significantly improves lip-sync quality, which cannot be reflected in a single image. Please refers to our project page for ablation studies on MaLS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Ablation Studies on Unseen Face Adaptation. We conduct ablation studies on the proposed unseen face adaptation following the same setting in Table 2. Additionally, we use 60 seconds of video for each 10 personalized models and 15 audios of 10 seconds from different identities for inference. Figure 7(b) shows the lip-sync metrics according to the adaptation step. In the cases without the sync regularizer, the models lose the lip-sync generality, in other words, it memorizes the short target video as the adaptation phase proceeds. Introducing the sync regularizer with the sync" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22847" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.174, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.091, + 0.272, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.091, + 0.371, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.091, + 0.468, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.165, + 0.174, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.165, + 0.272, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.165, + 0.371, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.165, + 0.468, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.244, + 0.162, + 0.254 + ], + "angle": 0, + "content": "(a) Reference" + }, + { + "type": "image_caption", + "bbox": [ + 0.181, + 0.244, + 0.271, + 0.254 + ], + "angle": 0, + "content": "(b) w/o Pose mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.244, + 0.358, + 0.254 + ], + "angle": 0, + "content": "(c) w/o SaMF" + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.244, + 0.442, + 0.254 + ], + "angle": 0, + "content": "(d) Full" + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.258, + 0.431, + 0.271 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of zero-shot model." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.284, + 0.468, + 0.325 + ], + "angle": 0, + "content": "Table 4. Mean Opinion score (MOS) user study results with \\(95\\%\\) confidence interval on cross-id setting. The score ranges in 1 to 5. The best score for each metric is in bold." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.332, + 0.462, + 0.451 + ], + "angle": 0, + "content": "
MethodUser Study (MOS)
Lip-sync AccuracyFace SimilarityVisual Quality
Wav2Lip [28]3.76 ± 0.202.98 ± 0.252.03 ± 0.21
ATVG [5]2.19 ± 0.212.45 ± 0.251.54 ± 0.17
MakeItTalk [46]2.32 ± 0.243.47 ± 0.232.95 ± 0.24
PC-AVS [45]3.28 ± 0.212.55 ± 0.222.51 ± 0.22
Ours4.01 ± 0.163.42 ± 0.223.55 ± 0.20
Ours (Personalized)3.52 ± 0.214.03 ± 0.173.64 ± 0.19
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.468, + 0.7 + ], + "angle": 0, + "content": "loss stabilizes the lip-sync metrics and even improves them compared to the zero-shot results. If we use the sync regularizer without the sync loss, lip-sync quality is stabilized, however slightly lower than the zero-shot results due to the lack of ground-truth audio-visual correlation. Audio generalization for unseen face data is maintained even though we used audio from the learning data. Figure 7(a) supports the validity of the adaptation method. It shows that visual difference between the zero-shot results and the adaptation results. The zero-shot model can generate an accurate lip motion for the target audio, while it shows a slight difference in person-specific details compared to the reference images. Through the proposed adaptation method, personal-specific lip shape, teeth, and wrinkles are faithfully recovered while maintaining the lip motion of the zero-shot results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.199, + 0.727 + ], + "angle": 0, + "content": "5.5. User Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.902 + ], + "angle": 0, + "content": "We further conduct a user study based on MOS (Mean opinion score) to compare the perceptual quality of each model, including our zero-shot and personalized model. 5 videos generated by each model in cross-id setting are used for this study. 20 participants scored lip-sync accuracy, face similarity, and visual quality of each video in the range of 1 to 5. As shown in Table 4, our models outperform all metrics. Specifically, our zero-shot model achieves the highest lip-sync accuracy, and our adaptation model achieves the highest score in face similarity and visual quality with competitive lip-sync accuracy." + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.092, + 0.564, + 0.099 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.1, + 0.58, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.1, + 0.613, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.101, + 0.655, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.101, + 0.718, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.101, + 0.886, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.16, + 0.604, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.16, + 0.648, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.16, + 0.713, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.16, + 0.887, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.226, + 0.564, + 0.233 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.233, + 0.58, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.233, + 0.613, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.233, + 0.655, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.233, + 0.718, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.233, + 0.886, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.293, + 0.604, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.293, + 0.657, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.293, + 0.719, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.293, + 0.887, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.359, + 0.861, + 0.372 + ], + "angle": 0, + "content": "(a) Comparison of cross-id results of zero-shot and adaptation." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.387, + 0.697, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.388, + 0.887, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.521, + 0.834, + 0.534 + ], + "angle": 0, + "content": "(b) Ablation study on the proposed sync regularizer." + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.539, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Figure 7. Experimental results of the proposed adaptation method using the cross-id setting. It improves the person specific visuals while maintaining the lip-sync generality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.614, + 0.619, + 0.63 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We proposed StyleLipSync, a lip-sync video generative model for arbitrary identity, which leverages expressive lip priors from a pre-trained style-based generator. In contrast to existing lip-sync generative models, we introduce pose-aware masking for lip region by utilizing a 3D parametric mesh predictor, which inherits the pose information in the mask itself. Designing a smooth lip motion by using the moving-average based latent smoothing in the continuous and linear latent space, StyleLipSync can generate temporally consistent lip motion. Furthermore, we propose a few-shot lip-sync adaptation method for a single person who lies in the out-of-distribution, which uses a few seconds of the target person's video. Experimental results show that our StyleLipSync can generate realistic lip-sync video from arbitrary audio even with the zero-shot setting, and the proposed adaptation method enhances the person-specific visual information without losing the lip-sync generality." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22848" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.226 + ], + "angle": 0, + "content": "Limitation and Future Works. Since learning audiovisual representation requires a large number of different identities (e.g., Voxceleb2 [7]), extending our method to higher resolution \\((512 \\times 512 \\uparrow)\\) is very challenging. We consider it our future work to develop an effective encoding system for a limited number of identities using a style-based generator [18, 19, 17]. Designing an effective reference encoder to improve lip identity preservation in a zero-shot setting can be another future work." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.227, + 0.47, + 0.303 + ], + "angle": 0, + "content": "Ethical Considerations. Since our method can generate a video of a specific person talking specific words only with a few seconds of video source, it has the potential for misuse. As discussed in [6], attaching visual watermarks on the generated videos can be a solution to it." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.316, + 0.174, + 0.33 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.47, + 0.397 + ], + "angle": 0, + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.472, + 0.452 + ], + "angle": 0, + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.509, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, pages 187-194, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.565, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.636, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[6] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.705, + 0.468, + 0.733 + ], + "angle": 0, + "content": "[7] J. S. Chung, A. Nagrani, and A. Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.734, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[8] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision, pages 251-263, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.776, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[9] Dipanjan Das, S. Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[10] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.34, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[11] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.223, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[13] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.281, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[14] Jingwen He, Wu Shi, Kai Chen, Lean Fu, and Chao Dong. Gcfsr: a generative and controllable face super resolution method without facial and gan priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1898, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.353, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[15] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.397, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[16] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.513, + 0.892, + 0.569 + ], + "angle": 0, + "content": "[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.571, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.643, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[21] Prajwal KR, Rudrabha Mukhopadhyay, Jerin Philip, Abhishek Jha, Vinay Namboodiri, and CV Jawahar. Towards automatic face-to-face translation. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1428-1436, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Camillo Lugasre, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22849" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[24] Dongchan Min, Minyoung Song, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. arXiv preprint arXiv:2208.10922, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[25] Yotam Nitzan, Kfir Aberman, Qiurui He, Orly Liba, Michal Yarom, Yossi Gandelsman, Inbar Mosseri, Yael Pritch, and Daniel Cohen-Or. Mystyle: A personalized generative prior. arXiv preprint arXiv:2203.17272, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[26] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[27] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In 36th AAAI Conference on Artificial Intelligence (AAAI 22). Association for the Advancement of Artificial Intelligence, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[28] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[29] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[30] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.532, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[31] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.589, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[32] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.646, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[33] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[34] Justus Thies, Mohamed A. Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[35] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[36] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[38] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9168-9178, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[39] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[40] Weihao Xia, Yujiu Yang, and Jing-Hao Xue. Gan inversion for consistent video interpolation and manipulation. arXiv preprint arXiv:2208.11197, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[41] Zhiliang Xu, Hang Zhou, Zhibin Hong, Ziwei Liu, Jiaming Liu, Zhizhi Guo, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Styleswap: Style-based generator empowers robust face swapping. In European Conference on Computer Vision, pages 661-677, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[42] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[44] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[45] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.644, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[46] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22850" + } + ] +] \ No newline at end of file diff --git a/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/35be9e19-3939-4344-a259-82c8df74d69e_origin.pdf b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/35be9e19-3939-4344-a259-82c8df74d69e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f9582683d6e7ef850a4d8592e915f9fe02a7fccc --- /dev/null +++ b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/35be9e19-3939-4344-a259-82c8df74d69e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71806cce51d0e34d4ca870f1337cfb9180d13aec4abb53d6db8ab78cbb3a5285 +size 2050650 diff --git a/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/full.md b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d8cdee035d6486292a6515d0ab5728344ee60533 --- /dev/null +++ b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/full.md @@ -0,0 +1,357 @@ +# StyleLipSync: Style-based Personalized Lip-sync Video Generation + +Taekyung Ki $^{1\star}$ Dongchan Min $^{2\star}$ $^{1}$ AITRICS, $^{2}$ Graduate School of AI, KAIST + +tkki@aitrics.com, alsehdcks95@kaist.ac.kr https://stylelipsync.github.io + +# Abstract + +In this paper, we present StyleLipSync, a style-based personalized lip-sync video generative model that can generate identity-agnostic lip-synchronizing video from arbitrary audio. To generate a video of arbitrary identities, we leverage expressive lip prior from the semantically rich latent space of a pre-trained StyleGAN, where we can also design a video consistency with a linear transformation. In contrast to the previous lip-sync methods, we introduce pose-aware masking that dynamically locates the mask to improve the naturalness over frames by utilizing a 3D parametric mesh predictor frame by frame. Moreover, we propose a few-shot lip-sync adaptation method for an arbitrary person by introducing a sync regularizer that preserves lip-sync generalization while enhancing the person-specific visual information. Extensive experiments demonstrate that our model can generate accurate lip-sync videos even with the zero-shot setting and enhance characteristics of an unseen face using a few seconds of target video through the proposed adaptation method. + +# 1. Introduction + +In the past few years, advances in deep learning have altered the dynamics of video creation. Now, users can easily make and edit videos with the help of deep learning. In particular, the task of generating a talking head video has received great interest due to its various practical uses. It can be applied in many applications such as film dubbing into a different language, face-to-face live chats, and virtual avatars in games and videos. Thus, a lot of prior works [21, 34, 28, 46, 45, 27] have been studied to generate a talking head video that has accurate lip shapes according to arbitrary audio inputs. + +Most of the prior works mainly focus on enhancing synchronization between lip shapes and audio input. Some of the previous methods [46, 9, 34] use intermediate structural + +representations such as landmarks and 3D models. They predicted the representations from the audio input and synthesized a talking head video of a target person. However, they suffered from inaccurate lip-sync results since such representations are too sparse to produce fine-grained details in lip-syncing. Recently, another line of methods [28, 27] mapped input audio to latent space and leveraged it to construct the mouth region of the target identity. While it achieves satisfactory results in lip-syncing, it generated blurry lower faces which are visually implausible. Furthermore, most methods only consider synthesizing frame-by-frame, lacking temporal consistency at the video level. + +In this paper, we propose StyleLipSync, a style-based lip-sync video generative model that can generate identity-agnostic lip-synchronizing video from the arbitrary audio input. Our model consists of the following components. First, different from a previous masking method [21, 28, 27, 6] which masks the entire lower half face, we propose Pose-aware Masking. We analyze that the previous masking method cause unpleasant artifacts and unnatural jaw moving in the generated videos. To circumvent this, we utilize a 3D face mesh predictor [11, 23] and generate lip masks with consideration of pose information and facial semantics such as jaw shape. Second, our image decoder is based on a style-based generator, namely StyleGAN [18, 19, 17]. StyleGANs have demonstrated their effectiveness in various facial generative tasks, including face editing[1, 2], face enhancement [42], and video generation [35, 24]. As a pre-trained StyleGAN already contains expressive and diverse face priors in style latent space [1], we leverage it to synthesize the high-fidelity lip region of the target person. Furthermore, thanks to the continuous and linear nature of the latent space [18, 13, 32], we linearly manipulate the style codes using the audio input to generate lip-synced video frames. Additionally, we propose Style-aware Masked Fusion to effectively adopt a skipconnection to our decoder, which helps to preserve the 2D structure of the image and improves lip fidelity. Finally, we propose a Moving-average based Latent Smoothing module that makes the latent trajectory smoother for enhancing the + +temporal consistency in the synthesized talking head video. + +While our model can synthesize a talking head video of the target person, there is a slight identity gap between a generated video and the target person. The gap can be noticeable, for example, in racial faces, which are relatively scarce in the training data. One approach to addressing this issue is to fine-tune the generator on a few seconds of video of the target person to create a personalized model. Several fine-tuning methods [30, 25, 37] have already been demonstrated and widely adopted by the industry to achieve product-level quality. However, we analyze that simply fine-tuning the generator loses its ability to generalize to arbitrary audio inputs, which is critical for generating talking head videos. Therefore, to minimize the side effect, we propose a sync regularizer enforcing the audio generalization performance. The key idea is to leverage the audio from the training data, not from the target video. Specifically, we not only optimize the generator to reconstruct the target video but also synthesize the video corresponding to the randomly sampled audio from the training data and maintain a sync correlation between the synthesized video and the audio. As a result, we obtain a personalized lip-sync generative model that can synthesize a video of the target person for arbitrary audio. Our contributions are summarized as follows: + +- We present StyleLipSync, a lip-sync video generative model which generates lip-synchronizing video in-the-wild of $256 \times 256$ resolution with accurate and natural lip movement from a given masked video frames, audio segment, and single reference image. +- We additionally propose a few-shot adaptation method for entirely unseen faces, which uses only a few seconds of video by introducing a sync regularizer to maintain audio generalization. +- Experimental results demonstrate that StyleLipSync achieves state-of-the-art performance in terms of lip-sync and visual quality, even with the zero-shot setting. + +# 2. Related Works + +# 2.1. Lip-sync Video Generation + +Lip-sync video generation aims to generate a talking face video with lip motion synchronized with the given input audio. Early works [21, 28] generate lip from the lower half masked face image corresponding to the input audio. Specifically, Wav2Lip [28] uses a pre-trained SyncNet [8] as a lip-sync expert which maximizes the correlation between the generated lip and the input audio. Similarly, we use lip-sync expert for audio-visual alignment, which is trained with a contrastive manner proposed in [24]. VideoRetalking [6] improves Wav2Lip [28] a two-stage manner, where it first generates low-resolution $(96\times 96)$ video and then increases the resolution by using a single identity specific super-resolution network. In contrast, we propose + +a zero-shot model that directly generates lip-sync video of $256 \times 256$ resolution and also propose a unseen face adaptation method then enhances the personal characteristics from a few-shot target video. SyncTalkFace [27] introduces a lip memory network, which encodes lip motion features into a discrete space in the training phase and retrieves a lip feature from query audio at the inference phase. In contrast to [27], we utilize continuous and linear latent space from a pre-trained style-based generator [19] to generate lip images with high fidelity and take video consistency into account in the latent space. + +# 2.2. GAN Prior + +Style-based generators [18, 19, 17] demonstrate the power of their semantic latent spaces, namely $\mathcal{W}$ , in image generation, image editing [13, 32], and video generation [35]. GAN-inversion [12, 1, 2, 29, 36] utilizes the pre-trained GANs to invert an image into corresponding latent code so one can manipulate the attributes of the image only within the latent space. Extended $\mathcal{W}+$ have been shown its much expressive power. For instance, pSp [29] adopts the feature pyramid networks (FPNs) [22] to use $\mathcal{W}+$ , which follow the nature of the progressive generation [15] of StyleGAN [18] and achieves state-of-the-art performance in image-to-image translation (e.g., face inpainting). Similarly, we utilize $\mathcal{W}+$ for diverse and strong lip prior since we aim to build a lip-sync video generative model of arbitrary identity. + +Recent works [42, 38, 41] not only adopt pre-trained GAN prior as their decoder but also introduce a skip-connection that concatenates the encoded and generated features, which helps the model preserve 2D spatial information. Specifically, GPEN [42] uses skip-connection that concatenates the encoded and generated features and achieves state-of-the-art performance in blind face restoration. StyleSwap [41] adopts it to face swap and introduces the ToMask branch that predicts the target facial attribute regions for swapping in a supervised manner. In contrast to those methods, we use an additive skip connection more efficient than the concatenation, along with the unsupervised predicted masked sum, which helps the decoder distinguish the target lip region from the whole face and therefore increases the lip image fidelity. + +# 2.3. Personalization + +Although GAN prior has been successful in various tasks [29, 36], it is still challenging to faithfully recover person-specific information that lies in the out-of-distribution [1, 2, 29, 36, 42]. Recently, a few-shot personalization has become an alternative to solving the problem [25]. Pivotal-tuning-inversion (PTI) [30] fine-tunes the image generator while freezing a single latent code, namely pivot, to compensate the person-specific information in the generative + +![](images/dc8017fdc410f69ed32e3e1e934127fa55afbd6387e54c6add6305322fd74a59.jpg) +Figure 1. A framework of StyleLipSync. We leverage a 3D parametric mesh predictor [11, 23] to obtain pose-aware masked frames $X_{1:T}$ which inherits the facial pose of input frames. Face encoder $\mathbf{E}_{face}$ maps $X_{1:T}$ into 2D spatial features and then fed into the decoder $\mathbf{G}$ through style-aware masked fusion (SaMF). Single reference image $X_{ref}$ and audio segments $A_{1:T}$ are mapped into latent space, followed by Moving-average based Latent Smoothing (MaLS). This module outputs smooth video latent codes $\tilde{w}_{1:T} \subseteq \mathcal{W}+$ that represent temporally consistent lip movement. With the guidance of SaMFs and the smooth video latent codes $\tilde{w}_{1:T}$ , StyleLipSync can generate temporally consistent lip-synced videos. + +process, not in the encoding process. MyStyle [25] adopts PTI [30] to image inpainting and semantic editing, by restricting the latent space to a subspace spanned by the multiple pivots from the few photos (roughly 100) of the target person. Stitch-it-in-time [37] adopts the pivotal-tuning to the video editing in a multi-stage manner, which leverages an off-the-shelf latent manipulation [31] to manipulate video in the latent space and then stitch it to the source video. Inspired by them, we propose a few-shot unseen face adaptation method that slightly fine-tunes the image decoder for a given latent code trajectory of a target identity and maintains the audio generalization by introducing a sync-regularizer. + +# 3. Method + +Given lip masked video frames $X_{1:T} = (X_t)_{t=1}^T$ and audio segments $A_{1:T} = (A_t)_{t=1}^T$ , a lip-sync method generates video frames + +$$ +\hat {X} _ {1: T} = (\hat {X} _ {t}) _ {t = 1} ^ {T}, \qquad (1) +$$ + +where $\hat{X}_{1:T}$ has a lip movement synchronized with the audio segments $A_{1:T}$ . In contrast to the previous lip-sync methods [21, 28, 27, 6], we leverage 3D parametric facial mesh predictor [11, 23] to compute lip mask so that the generator can be aware of semantically meaningful facial pose information (Section 3.1). We utilize a pre-trained StyleGAN [19] as our decoder $\mathbf{G}$ . When audio encoder $\mathbf{E}_{\text{aud}}$ and reference encoder $\mathbf{E}_{\text{ref}}$ map their inputs into the latent space $\mathcal{W}+$ (Section 3.3), the decoder $\mathbf{G}$ generates lip-synced video frames $\hat{X}_{1:T}$ from these latent codes (Section 3.2) guided by the proposed style-aware masked fusion. For + +enhancing the temporal consistency, we propose a Moving-average based Latent Smoothing module, which learns local motion between the latent codes, and makes video latent trajectory smoother. Finally, sync loss [8, 24] is used for the audio-lip synchronization. The overall framework of our model is described in Figure 1. + +# 3.1. Pose-aware Masking + +Dynamic head motion is an important factor in the natural talking style. However, existing methods [21, 28, 27, 6] employ the rectangular lower-half mouth masking method without consideration of the pose information. It often fails to detect appropriate masking regions when the head moves dynamically, which leads to unpleasant artifacts and unnatural jaw movement in the generated videos (see the first row in Figure 5 for examples). To address this limitation, we use the face meshes by leveraging a 3D face mesh predictor [11], which captures 3D parameters and predicts dense face geometry. We predict the 3D parameters [4] and the mesh from given video frames. Then, the predicted expression parameter $\delta \in \mathbb{R}^{64}$ is used to adjust the mesh to obtain naturally opened and closed mouth meshes. We normalize the mesh vertices using the predicted pose parameters $\tau \in \mathbb{R}^3$ (translation), $\gamma \in \mathrm{SO}(3)$ (rotation) and leave only the lower frontal vertices. These meshes are combined and projected onto the original 2D plane to finally get our pose-aware lip masks. Figure 2 illustrates the framework of the pose-aware masking. This masking not only captures the pose information but also inherits facial semantics such as jaw shape. Ablation studies in Section 5.4 show that the pose-aware masking helps the model to increase visual quality along + +![](images/ba1161b1b65e071e93fb3c0d187d891285d98f1066ebb6afec5bc6fededcbfc0.jpg) +Figure 2. Illustration of pose-aware masking. The expression parameter $\delta \in \mathbb{R}^{64}$ and the pose parameter $\tau \in \mathbb{R}^3$ , $\gamma \in \mathrm{SO}(3)$ are used to compute the natural mask. + +with dynamic pose. + +# 3.2.Decoder + +Lip Prior from Style-based Generator. Generating the lip-sync videos from scratch for an arbitrary person is very hard since the mapping from audio to lip is basically one-to-many. In this paper, we leverage a style-based generator as our image decoder $\mathbf{G}$ for the following two reasons. First, a pre-trained StyleGAN already contains expressive and diverse face priors [18, 19, 17] in the form of latent code, namely style code in the latent space $\mathcal{W} + [2]$ . Those latent spaces enable us to better synthesize the lip region of the target person with the diverse lip prior. Second, the style codes form the continuous and linear [18, 13, 32] latent spaces, which enables us to design a high-level visual transformation, such as natural motion, only with a linear transformation of latent code [39]. Hence, we can generate a talking head video with smooth lip motion by simply manipulating the style codes using audio, which the previous lip-sync methods never take into account. + +Style-aware Masked Fusion (SaMF). Recently, it has been explored that adopting skip-connections based on the concatenation to GAN-inversion helps to preserve the 2D spatial information of the input [41, 14, 42]. Similarly, we adopt an additive skip-connection to our model for effectively preserving the non-masked region of $X_{1:T}$ while faithfully utilizing the latent space. + +Specifically, we propose style-aware masked fusion (SaMF) for efficiently preserving the 2D spatial feature and relieving the information gap between masked and non-masked regions. SaMFs are introduced at the beginning of the decoder blocks. The decoder $\mathbf{G}$ consists of $L$ decoder blocks, each of which takes 2 style codes to modulate 3 convolution weights, as illustrated in Figure 3. The first style code in each decoder block modulates 2 different convolution weights, one for the convolution in the original block and the other for the SaMF. SaMF learns to predict a 1-channel mask $S_{t}^{l}$ of the current resolution from the encoded feature through the newly modulated convolution followed by the sigmoid, which is used for spatial weighted fusion of the encoded feature and the generated feature. + +Formally, given encoded feature $\mathbf{E}_{face}^{l}(X_{t})$ and gen + +![](images/30c7eef7cdb0eb181f35e80e5a3b167e23dd49eafbc9cfa07c176ef3e4493479.jpg) +Figure 3. Illustration of the decoder block. The encoded feature $\mathbf{E}_{face}^{l}(X_{t})$ is injected into $l$ -th decoder block through Style-aware Masked Fusion (SaMF). Note that only the convolutions in SaMF are trainable, while the others are frozen during the training phase. + +erated feature $\mathbf{G}^{l - 1}(X_t)$ of same dimension $\mathbb{R}^{h\times w\times c}$ , SaMF firstly predicts a spatial mask $S_{t}^{l}\in \mathbb{R}^{h\times w\times 1}$ from $\mathbf{E}_{face}^{l}(X_{t})$ and then output fused feature as follows: + +$$ +S _ {t} ^ {l} \otimes \mathbf {E} _ {\text {f a c e}} ^ {l} \left(X _ {t}\right) + \left(1 - S _ {t} ^ {l}\right) \otimes \mathbf {G} ^ {l - 1} \left(X _ {t}\right). \tag {2} +$$ + +Ablation studies (Section 5.4) show that SaMFs improve the fidelity of the mouth since they separate masked and non-masked regions. + +# 3.3.Encoders + +Our model has three different encoders: face encoder $\mathbf{E}_{face}$ , reference encoder $\mathbf{E}_{ref}$ , and audio encoder $\mathbf{E}_{aud}$ . + +Face encoder $\mathbf{E}_{face}$ takes masked video frames $X_{1:T}$ as input and then outputs $l$ 2D spatial features $\mathbf{E}_{face}^{l}(X_{t}) = \{\mathbf{E}_{face}^{l}(X_{t})\mid l\in [1,2,\dots ,L]\}$ for each $t$ . These features are injected into the decoder $\mathbf{G}$ through the style-aware masked fusion to efficiently preserve 2D spatial structure, as described in Section 3.2. + +Reference Encoder $\mathbf{E}_{ref}$ maps a reference $X_{ref}$ into $2L$ reference style codes, each of which has 512 dimensions. We simply denote the reference style code as $w_{ref} = [w_{ref}^{1}|w_{ref}^{2}|\dots |w_{ref}^{2L}] \in \mathbb{R}^{512 \times 2L}$ . + +Similar to $\mathbf{E}_{ref}$ , audio encoder $\mathbf{E}_{aud}$ maps a single audio segment into $2L$ audio style codes, each of which has 512 dimensions. As we use $T$ consecutive audio segments $A_{1:T}$ , $\mathbf{E}_{aud}$ independently maps $A_{t}$ into $a_{t} = [a_{t}^{1}|a_{t}^{2}|\dots |a_{t}^{2L}] \in \mathbb{R}^{512 \times 2L}$ . We simply denote $a_{1:T} = (a_{t})_{t=1}^{T}$ for $T$ audio style codes. Please refer to our supplementary materials for the detailed encoder architectures. + +From these style codes $w_{ref}$ and $a_{1:T}$ , we compute target video's style codes over frames $w_{1:T}$ by + +$$ +w _ {1: T} = \left(w _ {1}, w _ {2}, \dots , w _ {T}\right), \tag {3} +$$ + +where $w_{t} = w_{ref} + a_{t}\in \mathcal{W}+$ . We compute the style codes by simply adding these two different codes based on the + +linearity [13, 32] of the latent space $\mathcal{W}+$ . The style codes $w_{1:T}$ are then fed into the decoder $\mathbf{G}$ through the learned affine transformations to generate synced lip motion. + +Temporal Consistency. Thanks to the semantically rich latent space $\mathcal{W}+$ , our model can generate an accurate lip-sync video frame by frame. However, this frame-wisely independent encoding of style codes turns out to lead inconsistent mouth movements in the final results. To remedy this, we assume that the generated style codes $w_{1:T}$ form a trajectory of the target video [40] in $\mathcal{W}+$ and enforce the smooth local transition of the motion [24, 33, 39] into the trajectory. Toward this, we introduce Moving-average based Latent Smoothing (MaLS), each of which consists of a stack of the weighted moving-average [3] and 1D convolutions operating on the style codes along the time-axis. More precisely, we employ $2L$ MaLS for $L$ resolutions, each of which takes the $l$ -th component of $w_{t-1}, w_t$ , and $w_{t+1}$ as its inputs to learn the local difference between them, and then we inject the local motions into $w_{ref}$ to compute the smooth style code $\tilde{w}_t$ : + +$$ +\begin{array}{l} \tilde {w} _ {t} ^ {l} = w _ {r e f} ^ {l} + \mathrm {M a L S} ^ {l} (w _ {t - 1: t + 1} ^ {l}), \\ = w _ {r e f} ^ {l} + \operatorname {C o n v 1 D s} \left(\sum_ {\tau = t - 1} ^ {t + 1} \alpha_ {\tau} \cdot w _ {\tau} ^ {l}\right), \tag {4} \\ \end{array} +$$ + +where $\mathrm{MaLS}^l$ denotes the $l$ -th MaLS, and $\alpha_{\tau}$ is the weight of the moving-average. + +With this smooth latent codes $\tilde{w}_{1:T}$ , we compute the video frames: + +$$ +\hat {X} _ {1: T} = \left(\mathbf {G} \left(\tilde {w} _ {t}, \mathbf {E} _ {\text {f a c e}} \left(X _ {t}\right)\right)\right) _ {t = 1} ^ {T}. \tag {5} +$$ + +For better initialization [29], we add the average code $w_{avg}$ of the pre-trained generator to each $\tilde{w}_t$ in (4). + +# 3.4. Training Objective + +We train StyleLipSync to reconstruct target video frames from corresponding audio. We randomly choose $T = 5$ consequent frames with corresponding audio segments and 1 single reference frame. Image perceptual loss [43] is used to minimize perceptual image distance between generated frames $\hat{X}$ and ground-truth frames $Y$ : + +$$ +\mathcal {L} _ {l p i p s} = \sum_ {i = 1} ^ {N} \left\| \phi^ {i} (\hat {X}) - \phi^ {i} (Y) \right\| _ {2}, \tag {6} +$$ + +where $N$ is the number of feature extractor, $\phi^i (\cdot)$ is the $i$ th feature extractor, and $\| \cdot \| _2$ is the $\ell^2$ loss. Similar to [33], we use the multi-scale perceptual loss with 3 levels. + +For audio-visual alignment, we utilize SyncNet trained in a contrastive manner [24] that minimizes the cosine distance between generated frames $\hat{X}$ and corresponding audio segment $A$ : + +$$ +\mathcal {L} _ {\text {s y n c}} = 1 - \cos \left(f _ {v} (\hat {X}), f _ {a} (A)\right), \tag {7} +$$ + +![](images/1a5d8d6b21a61f6c4547432ad4bfd5c27165049cda41b35edd0040e6ed0528ad.jpg) +Figure 4. Adaptation for Unseen Face. We slightly tune the decoder $\mathbf{G}_{\theta}$ with the proposed sync regularizer $\mathcal{R}_{sync}$ , while freezing all encoders' weight. Face encoder $\mathbf{E}_{face}$ and SaMFs are omitted here for simplicity. + +where $\cos (\cdot ,\cdot)$ denotes the cosine similarity. $f_{v}(\cdot)$ and $f_{a}(\cdot)$ are the frame and audio feature extractor, respectively. The final objective $\mathcal{L}_{train}$ is computed as: + +$$ +\mathcal {L} _ {\text {t r a i n}} = \lambda_ {1} \mathcal {L} _ {\text {l p i p s}} + \lambda_ {2} \mathcal {L} _ {\text {s y n c}}, \tag {8} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are the balancing coefficients. + +# 4. Unseen Face Adaptation + +Although StyleLipSync successfully generates accurately lip-synced videos with high fidelity, the model would fail to exactly synthesize unseen faces lying in the out-of-distribution. This problem refers to the failure of id preservation. Therefore, to handle this, we fine-tune our decoder $\mathbf{G}$ on a target person video to make a personal model that is better able to synthesize toward the target person. + +Let $X_{1:T}$ be masked video frames of unseen face, which corresponds to the audio segments $A_{1:T}$ , and $X_{ref}$ be a reference frame. The frozen encoders convert each input into the intermediate representations, $\tilde{w}_{1:T}$ and $\mathbf{E}_{face}(X_t))_{t=1}^T$ . Then we fine-tune the decoder $\mathbf{G}_{\theta}$ , now parameterized by $\theta$ , by minimizing the distance between $(\mathbf{G}_{\theta}(\tilde{w}_t,\mathbf{E}_{face}(X_t)))_{t=1}^T$ and target frames, as same as (8). However, fine-tuning the decoder on a short video of a single identity leads to over-fitting and losing the lip-sync generality as the generator can memorize the target video [26]. To prevent the model from this scenario, we introduce a sync regularizer $\mathcal{R}_{sync}$ to enforce audio generality to the decoder $\mathbf{G}_{\theta}$ by leveraging the audio from the training dataset, not from the target video. Formally, given audio segments $A_{1:T}^{\prime}$ randomly chosen from the training data (Voxceleb2 [7]) and $w_{ref}$ , we compute smooth style codes $\tilde{w}_{1:T}^{\prime}$ , and then decode them to a synced video $X_{1:T}^{\prime}$ . The sync regularizer $\mathcal{R}_{sync}$ is defined as + +$$ +\mathcal {R} _ {\text {s y n c}} = 1 - \cos \left(f _ {v} \left(\hat {X} ^ {\prime}\right), f _ {a} \left(A ^ {\prime}\right)\right), \tag {9} +$$ + +which enforces $\mathbf{G}_{\theta}$ to generate $\hat{X}_{1:T}^{\prime}$ aligning with $A_{1:T}^{\prime}$ . The final object for a single person adaptation is given as follows: + +$$ +\theta^ {*} = \underset {\theta} {\operatorname {a r g m i n}} \mathcal {L} _ {\text {t r a i n}} + \lambda_ {R} \mathcal {R} _ {\text {s y n c}}, \tag {10} +$$ + +![](images/61ffa634159844f0c6b05d9e17ce424811b8b2a60edffcbb4166e3ea5f1ed34c.jpg) +(a) Reconstruction results on Voxceleb2 [7] test data. +Figure 5. Comparison with state-of-the-art methods. The different field of view comes from the pre-processing strategy of each model. + +![](images/78368105a11b5e3fa251e766852549e3bd6d103febcc4822321b71cd11d95889.jpg) +(b) Cross-id results on HDTF [44]. + +where $\lambda_R$ is the regularizer coefficient. Ablation studies in Section 5.4 show that $\mathcal{R}_{sync}$ regularizes the audio generality even if we use the audio from the training set. + +# 5. Experiments + +# 5.1. Dataset + +We train our model on Voxceleb2 [7], which consists of in-the-wild talking face videos collected from YouTube. It contains more than 145,000 videos of about 6,100 identities in the train set and more than 4,900 videos of more than 110 identities in the test set. We convert all videos into frames with 25 fps and then crop and resize those frames into $256 \times 256$ resolution, following the method of [33]. For more semantically rich face priors, we only use videos where the detected face bounding boxes' height (and width) is longer than 256. After pre-processing, the remaining 11051 videos and 340 videos are used as the training set and the test set, respectively. All audios are re-sampled to $16\mathrm{kHz}$ and converted into a mel-spectrogram to be used as our audio representation similar to the method in [45, 24]. We also use HDTF [44] to further test our model for crossid experiments. It is widely used to evaluate high-resolution talking face generative models, where the head pose dynamics are not as significant as the Voxceleb2. + +# 5.2. Implementation Details + +We pre-train StyleGAN2 [19] on the Voxceleb2 [7] following the implementation of [16]. For training StyleLipSync and SyncNet, we set the frame length $T = 5$ , and employ Adam [20] optimizer with learning rate $10^{-4}$ throughout the training phase in both cases. All experiments are performed on 2 TITAN RTX GPUs. We set $\lambda_{1} = 10$ , $\lambda_{2} = 0.1$ , $\lambda_{R} = 0.1$ for the objective (10), and $\alpha_{t-1} = 0.25$ , $\alpha_{t} = 0.5$ , $\alpha_{t+1} = 0.25$ for MaLS (4). For all inference, we use the first frame as the reference image. + +For image quality metric, we use PSNR, and SSIM. We also calculate CSIM [10] in the cross-id experiment that measures the face similarity between the images in the pretrained face embedding space. For lip-sync quality metric, we use LMD, LSE-D, and LSE-C. LMD [5] is the absolute distance of facial landmarks between the target and generated frames. LSE-D and LSE-C are proposed in [28], where LSE-D measures the distance between the lip and audio representations, and LSE-C measures the lip-sync confidence, respectively. + +# 5.3. Evaluation + +Reconstruction Results. We compare the other state-of-the-art methods in the lip-sync (Wav2Lip [28]) and talking + +face generation (ATVG [5], MakeItTalk [46], and PC-AVS [45]) on reconstruction results of Voxceleb2 test dataset. Table 1 shows that StyleLipSync outperforms all metrics except the LSE-C score. Wav2Lip [28] achieves the highest LSE-C score, however, it achieves low image quality metrics since it generates $96 \times 96$ low-resolution videos. PC-AVS [45] achieves comparable lip-sync scores with StyleLipSync, however, achieves low image quality metrics than our model since it highly relies on its specific preprocessing and fails to generate in many cases. We also illustrate the qualitative results in Figure 4. Wav2Lip [28] produces a lip-synchronized video with considerable visual artifacts since it cannot adapt to the video of dynamic head pose. MakeItTalk [46] generates low lip-sync video since it uses sparse facial landmarks. PC-AVS [45] generates an accurate lip-synchronized video following the input head pose. However, it struggles to preserve the unseen identity involving visual artifacts. StyleLipSync generates the natural lip motion with high fidelity and preserves the input identity, which is comparable to the ground truth video. + +Cross Identity Results. To evaluate lip-sync generalization, we conduct cross-id experiments settings using unseen videos and audio. We randomly sample 10 videos of different identities and 150 audio without duplication from HDTF [44]. We use the first 10 seconds of videos and audio. For each video, we generate 15 lip-synced videos from 15 different audios, where those face-audio pairs are not from the same source. In Table 2, we report the LSE-D and LSE-C for lip-sync quality and the CSIM for face similarity. Wav2Lip [28] achieves a higher LSE-C score than StyleLipSync, however, it achieves a low CSIM score since it produces a low-resolution video with visual artifacts. MakeItTalk [46] achieves the best CSIM score, while the lip-sync quality is the worst since it uses sparse facial landmarks. PC-AVS [45] outperforms LSE-C while achieving the lowest CSIM since it can't preserve an unseen face's identity. StyleLipSync achieves the best LSE-D score and comparable CSIM score to MakeItTalk [46]. We show qualitative results of cross-id experiments with target lip references in Figure 4. PC-AVS [45] can generate accurate lip-synchronizing video compared to the target lip, while it fails to preserve facial details of the unseen face. MakeItTalk [46] produces a high-resolution and identity-preserving video, however, it is out-of-sync compared to the target. StyleLipSync generates a high-resolution lip synchronizing video compared to the target lip, without any visual artifacts. + +# 5.4. Ablation Studies + +Ablation Studies on Zero-shot Model. Figure 6 and Table 3 summarize the ablation studies on our zero-shot method of reconstruction on Voxceleb2 test data. If we replace the pose-aware masking with standard rectangular + +Table 1. Quantitative comparison of reconstruction on Voxceleb2 test data. The best score for each metric is in bold. + +
MethodVoxceleb2 (Reconstruction)
ImageLip-Sync
SSIM↑PSNR↑LMD↓LSE-D↓LSE-C↑
Wav2Lip96×96 [28]0.44813.5346.4226.9998.329
ATVG128×128 [5]0.46113.3497.1658.8215.421
MakeItTalk256×256 [46]0.41912.8683.64910.8953.624
PC-AVS224×224 [45]0.36913.2102.8127.2787.699
Ours256×2560.63119.6072.6966.6288.056
+ +Table 2. Quantitative comparison of cross-identity results for unseen face. We report CSIM [10] as the image quality metric since there is no ground truth frames for the cross-id experiments. The best score for each metric is in bold. + +
MethodHDTF (Cross-id)
ImageLip-Sync
CSIM ↑LSE-D ↓LSE-C ↑
Wav2Lip96×96 [28]0.6567.0478.576
ATVG128×128 [5]0.2878.6686.040
MakeItTalk256×256 [46]0.77010.6414.725
PC-AVS224×224 [45]0.2386.9218.858
Ours256×2560.7376.8258.209
+ +Table 3. Ablation study on zero-shot model. The best score for each metric is in **bold**. + +
Method (ours)Voceleb2 (Reconstruction)
ImageLip-Sync
SSIM ↑PSNR ↑LMD ↓LSE-D ↓LSE-C ↓
w/o Pose Mask0.60218.8673.0576.7717.748
w/o MaLS0.59318.1862.7406.9947.577
w/o SaMF0.59118.1812.7646.8387.780
Full0.63119.6072.6966.6288.056
+ +masking (w/o Pose Mask) in [28, 6, 27], considerable visual artifacts occur around the masked region since it is insufficient to capture the pose difference between the reference and the target. To validate SaMF, we replace the modulated convolutions in SaMFs with the standard convolutions. Figure 6(c) shows that SaMFs significantly improve lip region's fidelity since the modulated convolution helps to be aware of the lip style. As shown in Table 3, MaLS significantly improves lip-sync quality, which cannot be reflected in a single image. Please refers to our project page for ablation studies on MaLS. + +Ablation Studies on Unseen Face Adaptation. We conduct ablation studies on the proposed unseen face adaptation following the same setting in Table 2. Additionally, we use 60 seconds of video for each 10 personalized models and 15 audios of 10 seconds from different identities for inference. Figure 7(b) shows the lip-sync metrics according to the adaptation step. In the cases without the sync regularizer, the models lose the lip-sync generality, in other words, it memorizes the short target video as the adaptation phase proceeds. Introducing the sync regularizer with the sync + +![](images/698d1f2001ae813ddf0bc7faa0819f15a4a6327ef2ad943f78924cd17ba9b61e.jpg) + +![](images/a4708a0e1698c232cded03e42c3e435857ef6d8c8d9ac3aee3b0401285f920c7.jpg) + +![](images/619085aa427a0a96bb5669e0deb157594248c33aee01de3928ce3a7976d193de.jpg) + +![](images/4b8d79bff406be7ef24efd9c84151e682448a47a3b79e53d3e6a35f4e0f455b6.jpg) + +![](images/48b1c0f911668b356bf3202d0eb401bb8369a9db52e7ad656b2529e6453708af.jpg) +(a) Reference + +![](images/51a18be1cd5024a406193b05721bcf4696e7ad7879c3d4ac875519e16b0f41ff.jpg) +(b) w/o Pose mask + +![](images/404f867ccb7b672d7bf5b7d2562e8d44493f7c58932ddcf5402b070e39821a03.jpg) +(c) w/o SaMF + +![](images/beace11418acd0a701ff56659bb5fbd5c76762bf6938d414471860ffe3c566be.jpg) +(d) Full +Figure 6. Qualitative comparison of zero-shot model. + +Table 4. Mean Opinion score (MOS) user study results with $95\%$ confidence interval on cross-id setting. The score ranges in 1 to 5. The best score for each metric is in bold. + +
MethodUser Study (MOS)
Lip-sync AccuracyFace SimilarityVisual Quality
Wav2Lip [28]3.76 ± 0.202.98 ± 0.252.03 ± 0.21
ATVG [5]2.19 ± 0.212.45 ± 0.251.54 ± 0.17
MakeItTalk [46]2.32 ± 0.243.47 ± 0.232.95 ± 0.24
PC-AVS [45]3.28 ± 0.212.55 ± 0.222.51 ± 0.22
Ours4.01 ± 0.163.42 ± 0.223.55 ± 0.20
Ours (Personalized)3.52 ± 0.214.03 ± 0.173.64 ± 0.19
+ +loss stabilizes the lip-sync metrics and even improves them compared to the zero-shot results. If we use the sync regularizer without the sync loss, lip-sync quality is stabilized, however slightly lower than the zero-shot results due to the lack of ground-truth audio-visual correlation. Audio generalization for unseen face data is maintained even though we used audio from the learning data. Figure 7(a) supports the validity of the adaptation method. It shows that visual difference between the zero-shot results and the adaptation results. The zero-shot model can generate an accurate lip motion for the target audio, while it shows a slight difference in person-specific details compared to the reference images. Through the proposed adaptation method, personal-specific lip shape, teeth, and wrinkles are faithfully recovered while maintaining the lip motion of the zero-shot results. + +# 5.5. User Study + +We further conduct a user study based on MOS (Mean opinion score) to compare the perceptual quality of each model, including our zero-shot and personalized model. 5 videos generated by each model in cross-id setting are used for this study. 20 participants scored lip-sync accuracy, face similarity, and visual quality of each video in the range of 1 to 5. As shown in Table 4, our models outperform all metrics. Specifically, our zero-shot model achieves the highest lip-sync accuracy, and our adaptation model achieves the highest score in face similarity and visual quality with competitive lip-sync accuracy. + +![](images/b822265a98179096967236193e26baf6594f688045c67653f7cab4a9f704b1b3.jpg) +Reference + +![](images/f294511969cd368b685fbefd6096572670fe578201e4260dcfac6118b720cab6.jpg) + +![](images/6cc51cc1828e797f22d508ade1792ee6550fb1af8676a064d05e789732ebf0b4.jpg) + +![](images/138cb89af7151cfd97a94847703ebb601befeb5c89d6388d9c95d5cb1af7a857.jpg) + +![](images/1b9d0caf51bc87ac06851bc69d154c9ca29b7e64ee1c3b91295994bcf60736b4.jpg) + +![](images/1296d6d1f743fec74111cbfa48291a7ddcb0b50b8a8303963c96da3617c243af.jpg) + +![](images/7b33774d7de86e7eac1bd941f7a06a64137f38960dbf4df109dc7660bcd91c2d.jpg) + +![](images/5c51ddc39c4015f8c831c6eb8d33c5c8a82554020d9e845eb08c58545bab80fc.jpg) + +![](images/ed7fdd7a259014ca277e120771bd702cda400120f06a5b792889c4cc10cc2eea.jpg) + +![](images/7cac37e4dfb3a2d148ec5686177cb512a1ba646973558181088ac76c61b8ec5a.jpg) +Reference + +![](images/381fc29f755cc4b721118146ac6f62961058002e8c3e07e2fcc73ade466d27cd.jpg) + +![](images/2a28cda9da671a4dd4ace4823b7b930f9089cf94fef925aded126d0869d7c2fe.jpg) + +![](images/1ac1c3637c4abd36a4ff76f6b663453ddba4e6b761cff16e23cb55f376572163.jpg) + +![](images/820a54989705e6f0b4ab47b889ec449dca315dd045bc27cd5a1349f193695c12.jpg) + +![](images/1b77a2a94484b0da1c715e87fbbcd0590ba3ccb2304ecb088a63e2d5cce42cf4.jpg) + +![](images/40994e0b5b587add63603696f30d796ba6ddb8a44992ff2d638f6080d2242180.jpg) +(a) Comparison of cross-id results of zero-shot and adaptation. + +![](images/ec22210d6ed95457a1ac66d5d5f47f98f7a50c4d6b4d4c1fc38d5797d1842306.jpg) + +![](images/546dea6b7e3c73acf0e549d0d130dea73563c92340cbed7c016ce47afe784631.jpg) + +![](images/4a6aa5c2fc86a5b5c34f4eba7acc9f621bdb2de268c3042f674f966967d5c890.jpg) +(b) Ablation study on the proposed sync regularizer. +Figure 7. Experimental results of the proposed adaptation method using the cross-id setting. It improves the person specific visuals while maintaining the lip-sync generality. + +![](images/54c0008f606aed12cfba9029416b59432a1a18c28678d5e913e887b81ee7618e.jpg) + +# 6. Conclusion + +We proposed StyleLipSync, a lip-sync video generative model for arbitrary identity, which leverages expressive lip priors from a pre-trained style-based generator. In contrast to existing lip-sync generative models, we introduce pose-aware masking for lip region by utilizing a 3D parametric mesh predictor, which inherits the pose information in the mask itself. Designing a smooth lip motion by using the moving-average based latent smoothing in the continuous and linear latent space, StyleLipSync can generate temporally consistent lip motion. Furthermore, we propose a few-shot lip-sync adaptation method for a single person who lies in the out-of-distribution, which uses a few seconds of the target person's video. Experimental results show that our StyleLipSync can generate realistic lip-sync video from arbitrary audio even with the zero-shot setting, and the proposed adaptation method enhances the person-specific visual information without losing the lip-sync generality. + +Limitation and Future Works. Since learning audiovisual representation requires a large number of different identities (e.g., Voxceleb2 [7]), extending our method to higher resolution $(512 \times 512 \uparrow)$ is very challenging. We consider it our future work to develop an effective encoding system for a limited number of identities using a style-based generator [18, 19, 17]. Designing an effective reference encoder to improve lip identity preservation in a zero-shot setting can be another future work. + +Ethical Considerations. Since our method can generate a video of a specific person talking specific words only with a few seconds of video source, it has the potential for misuse. As discussed in [6], attaching visual watermarks on the generated videos can be a solution to it. + +# References + +[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. +[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020. +[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022. +[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, pages 187-194, 1999. +[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. +[6] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. +[7] J. S. Chung, A. Nagrani, and A. Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. +[8] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision, pages 251-263, 2016. +[9] Dipanjan Das, S. Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision, 2020. +[10] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019. + +[11] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019. +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 2014. +[13] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020. +[14] Jingwen He, Wu Shi, Kai Chen, Lean Fu, and Chao Dong. Gcfsr: a generative and controllable face super resolution method without facial and gan priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1898, 2022. +[15] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. +[16] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. +[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. +[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. +[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. +[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[21] Prajwal KR, Rudrabha Mukhopadhyay, Jerin Philip, Abhishek Jha, Vinay Namboodiri, and CV Jawahar. Towards automatic face-to-face translation. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1428-1436, 2019. +[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. +[23] Camillo Lugasre, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. + +[24] Dongchan Min, Minyoung Song, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. arXiv preprint arXiv:2208.10922, 2022. +[25] Yotam Nitzan, Kfir Aberman, Qiurui He, Orly Liba, Michal Yarom, Yossi Gandelsman, Inbar Mosseri, Yael Pritch, and Daniel Cohen-Or. Mystyle: A personalized generative prior. arXiv preprint arXiv:2203.17272, 2022. +[26] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. +[27] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In 36th AAAI Conference on Artificial Intelligence (AAAI 22). Association for the Advancement of Artificial Intelligence, 2022. +[28] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. +[29] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. +[30] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. +[31] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020. +[32] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021. +[33] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in Neural Information Processing Systems, 32, 2019. +[34] Justus Thies, Mohamed A. Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision, 2019. +[35] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021. +[36] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. +[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial + +editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. +[38] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9168-9178, 2021. +[39] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations, 2022. +[40] Weihao Xia, Yujiu Yang, and Jing-Hao Xue. Gan inversion for consistent video interpolation and manipulation. arXiv preprint arXiv:2208.11197, 2022. +[41] Zhiliang Xu, Hang Zhou, Zhibin Hong, Ziwei Liu, Jiaming Liu, Zhizhi Guo, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Styleswap: Style-based generator empowers robust face swapping. In European Conference on Computer Vision, pages 661-677, 2022. +[42] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021. +[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. +[44] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. +[45] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021. +[46] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. \ No newline at end of file diff --git a/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/images.zip b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2b2b71fef9d02233727c943948095ec96427d90b --- /dev/null +++ b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cca1b2e423729d5770251b82072830883d405c4027f558262ce91ececfa7b84 +size 631241 diff --git a/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/layout.json b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..922d45c1e8f70c9efd801234d4a7b075c006ae61 --- /dev/null +++ b/2023/StyleLipSync_ Style-based Personalized Lip-sync Video Generation/layout.json @@ -0,0 +1,10615 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 504, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 504, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 504, + 121 + ], + "type": "text", + "content": "StyleLipSync: Style-based Personalized Lip-sync Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 143, + 406, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 143, + 406, + 171 + ], + "spans": [ + { + "bbox": [ + 187, + 143, + 406, + 171 + ], + "type": "text", + "content": "Taekyung Ki" + }, + { + "bbox": [ + 187, + 143, + 406, + 171 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 187, + 143, + 406, + 171 + ], + "type": "text", + "content": " Dongchan Min" + }, + { + "bbox": [ + 187, + 143, + 406, + 171 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 187, + 143, + 406, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 187, + 143, + 406, + 171 + ], + "type": "text", + "content": "AITRICS, " + }, + { + "bbox": [ + 187, + 143, + 406, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 187, + 143, + 406, + 171 + ], + "type": "text", + "content": "Graduate School of AI, KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 174, + 408, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 174, + 408, + 199 + ], + "spans": [ + { + "bbox": [ + 186, + 174, + 408, + 199 + ], + "type": "text", + "content": "tkki@aitrics.com, alsehdcks95@kaist.ac.kr https://stylelipsync.github.io" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 252, + 290, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 479 + ], + "type": "text", + "content": "In this paper, we present StyleLipSync, a style-based personalized lip-sync video generative model that can generate identity-agnostic lip-synchronizing video from arbitrary audio. To generate a video of arbitrary identities, we leverage expressive lip prior from the semantically rich latent space of a pre-trained StyleGAN, where we can also design a video consistency with a linear transformation. In contrast to the previous lip-sync methods, we introduce pose-aware masking that dynamically locates the mask to improve the naturalness over frames by utilizing a 3D parametric mesh predictor frame by frame. Moreover, we propose a few-shot lip-sync adaptation method for an arbitrary person by introducing a sync regularizer that preserves lip-sync generalization while enhancing the person-specific visual information. Extensive experiments demonstrate that our model can generate accurate lip-sync videos even with the zero-shot setting and enhance characteristics of an unseen face using a few seconds of target video through the proposed adaptation method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 504, + 128, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 128, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 128, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 525, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 656 + ], + "type": "text", + "content": "In the past few years, advances in deep learning have altered the dynamics of video creation. Now, users can easily make and edit videos with the help of deep learning. In particular, the task of generating a talking head video has received great interest due to its various practical uses. It can be applied in many applications such as film dubbing into a different language, face-to-face live chats, and virtual avatars in games and videos. Thus, a lot of prior works [21, 34, 28, 46, 45, 27] have been studied to generate a talking head video that has accurate lip shapes according to arbitrary audio inputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 657, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 657, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 657, + 288, + 693 + ], + "type": "text", + "content": "Most of the prior works mainly focus on enhancing synchronization between lip shapes and audio input. Some of the previous methods [46, 9, 34] use intermediate structural" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 228, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 373 + ], + "type": "text", + "content": "representations such as landmarks and 3D models. They predicted the representations from the audio input and synthesized a talking head video of a target person. However, they suffered from inaccurate lip-sync results since such representations are too sparse to produce fine-grained details in lip-syncing. Recently, another line of methods [28, 27] mapped input audio to latent space and leveraged it to construct the mouth region of the target identity. While it achieves satisfactory results in lip-syncing, it generated blurry lower faces which are visually implausible. Furthermore, most methods only consider synthesizing frame-by-frame, lacking temporal consistency at the video level." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we propose StyleLipSync, a style-based lip-sync video generative model that can generate identity-agnostic lip-synchronizing video from the arbitrary audio input. Our model consists of the following components. First, different from a previous masking method [21, 28, 27, 6] which masks the entire lower half face, we propose Pose-aware Masking. We analyze that the previous masking method cause unpleasant artifacts and unnatural jaw moving in the generated videos. To circumvent this, we utilize a 3D face mesh predictor [11, 23] and generate lip masks with consideration of pose information and facial semantics such as jaw shape. Second, our image decoder is based on a style-based generator, namely StyleGAN [18, 19, 17]. StyleGANs have demonstrated their effectiveness in various facial generative tasks, including face editing[1, 2], face enhancement [42], and video generation [35, 24]. As a pre-trained StyleGAN already contains expressive and diverse face priors in style latent space [1], we leverage it to synthesize the high-fidelity lip region of the target person. Furthermore, thanks to the continuous and linear nature of the latent space [18, 13, 32], we linearly manipulate the style codes using the audio input to generate lip-synced video frames. Additionally, we propose Style-aware Masked Fusion to effectively adopt a skipconnection to our decoder, which helps to preserve the 2D structure of the image and improves lip fidelity. Finally, we propose a Moving-average based Latent Smoothing module that makes the latent trajectory smoother for enhancing the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22841" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 83 + ], + "type": "text", + "content": "temporal consistency in the synthesized talking head video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 357 + ], + "type": "text", + "content": "While our model can synthesize a talking head video of the target person, there is a slight identity gap between a generated video and the target person. The gap can be noticeable, for example, in racial faces, which are relatively scarce in the training data. One approach to addressing this issue is to fine-tune the generator on a few seconds of video of the target person to create a personalized model. Several fine-tuning methods [30, 25, 37] have already been demonstrated and widely adopted by the industry to achieve product-level quality. However, we analyze that simply fine-tuning the generator loses its ability to generalize to arbitrary audio inputs, which is critical for generating talking head videos. Therefore, to minimize the side effect, we propose a sync regularizer enforcing the audio generalization performance. The key idea is to leverage the audio from the training data, not from the target video. Specifically, we not only optimize the generator to reconstruct the target video but also synthesize the video corresponding to the randomly sampled audio from the training data and maintain a sync correlation between the synthesized video and the audio. As a result, we obtain a personalized lip-sync generative model that can synthesize a video of the target person for arbitrary audio. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 362, + 286, + 509 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 59, + 362, + 286, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 362, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 59, + 362, + 286, + 421 + ], + "type": "text", + "content": "- We present StyleLipSync, a lip-sync video generative model which generates lip-synchronizing video in-the-wild of " + }, + { + "bbox": [ + 59, + 362, + 286, + 421 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 59, + 362, + 286, + 421 + ], + "type": "text", + "content": " resolution with accurate and natural lip movement from a given masked video frames, audio segment, and single reference image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 424, + 286, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 424, + 286, + 471 + ], + "spans": [ + { + "bbox": [ + 59, + 424, + 286, + 471 + ], + "type": "text", + "content": "- We additionally propose a few-shot adaptation method for entirely unseen faces, which uses only a few seconds of video by introducing a sync regularizer to maintain audio generalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 474, + 286, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 474, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 59, + 474, + 286, + 509 + ], + "type": "text", + "content": "- Experimental results demonstrate that StyleLipSync achieves state-of-the-art performance in terms of lip-sync and visual quality, even with the zero-shot setting." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 520, + 138, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 138, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 138, + 532 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 540, + 195, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 195, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 195, + 552 + ], + "type": "text", + "content": "2.1. Lip-sync Video Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "content": "Lip-sync video generation aims to generate a talking face video with lip motion synchronized with the given input audio. Early works [21, 28] generate lip from the lower half masked face image corresponding to the input audio. Specifically, Wav2Lip [28] uses a pre-trained SyncNet [8] as a lip-sync expert which maximizes the correlation between the generated lip and the input audio. Similarly, we use lip-sync expert for audio-visual alignment, which is trained with a contrastive manner proposed in [24]. VideoRetalking [6] improves Wav2Lip [28] a two-stage manner, where it first generates low-resolution " + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "inline_equation", + "content": "(96\\times 96)" + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "content": " video and then increases the resolution by using a single identity specific super-resolution network. In contrast, we propose" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "a zero-shot model that directly generates lip-sync video of " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " resolution and also propose a unseen face adaptation method then enhances the personal characteristics from a few-shot target video. SyncTalkFace [27] introduces a lip memory network, which encodes lip motion features into a discrete space in the training phase and retrieves a lip feature from query audio at the inference phase. In contrast to [27], we utilize continuous and linear latent space from a pre-trained style-based generator [19] to generate lip images with high fidelity and take video consistency into account in the latent space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 212, + 380, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 212, + 380, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 380, + 224 + ], + "type": "text", + "content": "2.2. GAN Prior" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "text", + "content": "Style-based generators [18, 19, 17] demonstrate the power of their semantic latent spaces, namely " + }, + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "text", + "content": ", in image generation, image editing [13, 32], and video generation [35]. GAN-inversion [12, 1, 2, 29, 36] utilizes the pre-trained GANs to invert an image into corresponding latent code so one can manipulate the attributes of the image only within the latent space. Extended " + }, + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "text", + "content": " have been shown its much expressive power. For instance, pSp [29] adopts the feature pyramid networks (FPNs) [22] to use " + }, + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "text", + "content": ", which follow the nature of the progressive generation [15] of StyleGAN [18] and achieves state-of-the-art performance in image-to-image translation (e.g., face inpainting). Similarly, we utilize " + }, + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 304, + 231, + 545, + 411 + ], + "type": "text", + "content": " for diverse and strong lip prior since we aim to build a lip-sync video generative model of arbitrary identity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 411, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 590 + ], + "type": "text", + "content": "Recent works [42, 38, 41] not only adopt pre-trained GAN prior as their decoder but also introduce a skip-connection that concatenates the encoded and generated features, which helps the model preserve 2D spatial information. Specifically, GPEN [42] uses skip-connection that concatenates the encoded and generated features and achieves state-of-the-art performance in blind face restoration. StyleSwap [41] adopts it to face swap and introduces the ToMask branch that predicts the target facial attribute regions for swapping in a supervised manner. In contrast to those methods, we use an additive skip connection more efficient than the concatenation, along with the unsupervised predicted masked sum, which helps the decoder distinguish the target lip region from the whole face and therefore increases the lip image fidelity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 599, + 400, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 400, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 400, + 610 + ], + "type": "text", + "content": "2.3. Personalization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Although GAN prior has been successful in various tasks [29, 36], it is still challenging to faithfully recover person-specific information that lies in the out-of-distribution [1, 2, 29, 36, 42]. Recently, a few-shot personalization has become an alternative to solving the problem [25]. Pivotal-tuning-inversion (PTI) [30] fine-tunes the image generator while freezing a single latent code, namely pivot, to compensate the person-specific information in the generative" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22842" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 532, + 251 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 532, + 251 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 532, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 532, + 251 + ], + "type": "image", + "image_path": "dc8017fdc410f69ed32e3e1e934127fa55afbd6387e54c6add6305322fd74a59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "text", + "content": "Figure 1. A framework of StyleLipSync. We leverage a 3D parametric mesh predictor [11, 23] to obtain pose-aware masked frames " + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "inline_equation", + "content": "X_{1:T}" + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "text", + "content": " which inherits the facial pose of input frames. Face encoder " + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{face}" + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "text", + "content": " maps " + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "inline_equation", + "content": "X_{1:T}" + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "text", + "content": " into 2D spatial features and then fed into the decoder " + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "text", + "content": " through style-aware masked fusion (SaMF). Single reference image " + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "inline_equation", + "content": "X_{ref}" + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "text", + "content": " and audio segments " + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "inline_equation", + "content": "A_{1:T}" + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "text", + "content": " are mapped into latent space, followed by Moving-average based Latent Smoothing (MaLS). This module outputs smooth video latent codes " + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "inline_equation", + "content": "\\tilde{w}_{1:T} \\subseteq \\mathcal{W}+" + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "text", + "content": " that represent temporally consistent lip movement. With the guidance of SaMFs and the smooth video latent codes " + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "inline_equation", + "content": "\\tilde{w}_{1:T}" + }, + { + "bbox": [ + 46, + 253, + 548, + 319 + ], + "type": "text", + "content": ", StyleLipSync can generate temporally consistent lip-synced videos." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 339, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 289, + 495 + ], + "type": "text", + "content": "process, not in the encoding process. MyStyle [25] adopts PTI [30] to image inpainting and semantic editing, by restricting the latent space to a subspace spanned by the multiple pivots from the few photos (roughly 100) of the target person. Stitch-it-in-time [37] adopts the pivotal-tuning to the video editing in a multi-stage manner, which leverages an off-the-shelf latent manipulation [31] to manipulate video in the latent space and then stitch it to the source video. Inspired by them, we propose a few-shot unseen face adaptation method that slightly fine-tunes the image decoder for a given latent code trajectory of a target identity and maintains the audio generalization by introducing a sync-regularizer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 506, + 104, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 104, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 104, + 517 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": "Given lip masked video frames " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "X_{1:T} = (X_t)_{t=1}^T" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": " and audio segments " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "A_{1:T} = (A_t)_{t=1}^T" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": ", a lip-sync method generates video frames" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 561, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 561, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 130, + 561, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {1: T} = (\\hat {X} _ {t}) _ {t = 1} ^ {T}, \\qquad (1)", + "image_path": "18110934368632a586a88c1ced953d3f3bca1a155cfdbbf9015721e7c59cb6d0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{X}_{1:T}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " has a lip movement synchronized with the audio segments " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "A_{1:T}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". In contrast to the previous lip-sync methods [21, 28, 27, 6], we leverage 3D parametric facial mesh predictor [11, 23] to compute lip mask so that the generator can be aware of semantically meaningful facial pose information (Section 3.1). We utilize a pre-trained StyleGAN [19] as our decoder " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". When audio encoder " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\text{aud}}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and reference encoder " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\text{ref}}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " map their inputs into the latent space " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " (Section 3.3), the decoder " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " generates lip-synced video frames " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{X}_{1:T}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " from these latent codes (Section 3.2) guided by the proposed style-aware masked fusion. For" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 339, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 547, + 411 + ], + "type": "text", + "content": "enhancing the temporal consistency, we propose a Moving-average based Latent Smoothing module, which learns local motion between the latent codes, and makes video latent trajectory smoother. Finally, sync loss [8, 24] is used for the audio-lip synchronization. The overall framework of our model is described in Figure 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 419, + 426, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 419, + 426, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 426, + 433 + ], + "type": "text", + "content": "3.1. Pose-aware Masking" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": "Dynamic head motion is an important factor in the natural talking style. However, existing methods [21, 28, 27, 6] employ the rectangular lower-half mouth masking method without consideration of the pose information. It often fails to detect appropriate masking regions when the head moves dynamically, which leads to unpleasant artifacts and unnatural jaw movement in the generated videos (see the first row in Figure 5 for examples). To address this limitation, we use the face meshes by leveraging a 3D face mesh predictor [11], which captures 3D parameters and predicts dense face geometry. We predict the 3D parameters [4] and the mesh from given video frames. Then, the predicted expression parameter " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^{64}" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": " is used to adjust the mesh to obtain naturally opened and closed mouth meshes. We normalize the mesh vertices using the predicted pose parameters " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\tau \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": " (translation), " + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathrm{SO}(3)" + }, + { + "bbox": [ + 304, + 438, + 547, + 715 + ], + "type": "text", + "content": " (rotation) and leave only the lower frontal vertices. These meshes are combined and projected onto the original 2D plane to finally get our pose-aware lip masks. Figure 2 illustrates the framework of the pose-aware masking. This masking not only captures the pose information but also inherits facial semantics such as jaw shape. Ablation studies in Section 5.4 show that the pose-aware masking helps the model to increase visual quality along" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22843" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 289, + 169 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 169 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 169 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 169 + ], + "type": "image", + "image_path": "ba1161b1b65e071e93fb3c0d187d891285d98f1066ebb6afec5bc6fededcbfc0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 287, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 204 + ], + "type": "text", + "content": "Figure 2. Illustration of pose-aware masking. The expression parameter " + }, + { + "bbox": [ + 46, + 171, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^{64}" + }, + { + "bbox": [ + 46, + 171, + 287, + 204 + ], + "type": "text", + "content": " and the pose parameter " + }, + { + "bbox": [ + 46, + 171, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\tau \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 171, + 287, + 204 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 171, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathrm{SO}(3)" + }, + { + "bbox": [ + 46, + 171, + 287, + 204 + ], + "type": "text", + "content": " are used to compute the natural mask." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 209, + 128, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 128, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 128, + 221 + ], + "type": "text", + "content": "with dynamic pose." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 228, + 110, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 110, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 110, + 239 + ], + "type": "text", + "content": "3.2.Decoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 247, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 449 + ], + "type": "text", + "content": "Lip Prior from Style-based Generator. Generating the lip-sync videos from scratch for an arbitrary person is very hard since the mapping from audio to lip is basically one-to-many. In this paper, we leverage a style-based generator as our image decoder " + }, + { + "bbox": [ + 46, + 247, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 46, + 247, + 287, + 449 + ], + "type": "text", + "content": " for the following two reasons. First, a pre-trained StyleGAN already contains expressive and diverse face priors [18, 19, 17] in the form of latent code, namely style code in the latent space " + }, + { + "bbox": [ + 46, + 247, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{W} + [2]" + }, + { + "bbox": [ + 46, + 247, + 287, + 449 + ], + "type": "text", + "content": ". Those latent spaces enable us to better synthesize the lip region of the target person with the diverse lip prior. Second, the style codes form the continuous and linear [18, 13, 32] latent spaces, which enables us to design a high-level visual transformation, such as natural motion, only with a linear transformation of latent code [39]. Hence, we can generate a talking head video with smooth lip motion by simply manipulating the style codes using audio, which the previous lip-sync methods never take into account." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 533 + ], + "type": "text", + "content": "Style-aware Masked Fusion (SaMF). Recently, it has been explored that adopting skip-connections based on the concatenation to GAN-inversion helps to preserve the 2D spatial information of the input [41, 14, 42]. Similarly, we adopt an additive skip-connection to our model for effectively preserving the non-masked region of " + }, + { + "bbox": [ + 46, + 450, + 287, + 533 + ], + "type": "inline_equation", + "content": "X_{1:T}" + }, + { + "bbox": [ + 46, + 450, + 287, + 533 + ], + "type": "text", + "content": " while faithfully utilizing the latent space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 700 + ], + "type": "text", + "content": "Specifically, we propose style-aware masked fusion (SaMF) for efficiently preserving the 2D spatial feature and relieving the information gap between masked and non-masked regions. SaMFs are introduced at the beginning of the decoder blocks. The decoder " + }, + { + "bbox": [ + 46, + 534, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 46, + 534, + 288, + 700 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 46, + 534, + 288, + 700 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 534, + 288, + 700 + ], + "type": "text", + "content": " decoder blocks, each of which takes 2 style codes to modulate 3 convolution weights, as illustrated in Figure 3. The first style code in each decoder block modulates 2 different convolution weights, one for the convolution in the original block and the other for the SaMF. SaMF learns to predict a 1-channel mask " + }, + { + "bbox": [ + 46, + 534, + 288, + 700 + ], + "type": "inline_equation", + "content": "S_{t}^{l}" + }, + { + "bbox": [ + 46, + 534, + 288, + 700 + ], + "type": "text", + "content": " of the current resolution from the encoded feature through the newly modulated convolution followed by the sigmoid, which is used for spatial weighted fusion of the encoded feature and the generated feature." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 700, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 58, + 700, + 287, + 715 + ], + "type": "text", + "content": "Formally, given encoded feature " + }, + { + "bbox": [ + 58, + 700, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{face}^{l}(X_{t})" + }, + { + "bbox": [ + 58, + 700, + 287, + 715 + ], + "type": "text", + "content": " and gen" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 545, + 235 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 235 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 235 + ], + "type": "image", + "image_path": "30c7eef7cdb0eb181f35e80e5a3b167e23dd49eafbc9cfa07c176ef3e4493479.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 238, + 546, + 282 + ], + "lines": [ + { + "bbox": [ + 305, + 238, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 546, + 282 + ], + "type": "text", + "content": "Figure 3. Illustration of the decoder block. The encoded feature " + }, + { + "bbox": [ + 305, + 238, + 546, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{face}^{l}(X_{t})" + }, + { + "bbox": [ + 305, + 238, + 546, + 282 + ], + "type": "text", + "content": " is injected into " + }, + { + "bbox": [ + 305, + 238, + 546, + 282 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 238, + 546, + 282 + ], + "type": "text", + "content": "-th decoder block through Style-aware Masked Fusion (SaMF). Note that only the convolutions in SaMF are trainable, while the others are frozen during the training phase." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "type": "text", + "content": "erated feature " + }, + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^{l - 1}(X_t)" + }, + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "type": "text", + "content": " of same dimension " + }, + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{h\\times w\\times c}" + }, + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "type": "text", + "content": ", SaMF firstly predicts a spatial mask " + }, + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "type": "inline_equation", + "content": "S_{t}^{l}\\in \\mathbb{R}^{h\\times w\\times 1}" + }, + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{face}^{l}(X_{t})" + }, + { + "bbox": [ + 305, + 287, + 545, + 326 + ], + "type": "text", + "content": " and then output fused feature as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 331, + 545, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 331, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 339, + 331, + 545, + 346 + ], + "type": "interline_equation", + "content": "S _ {t} ^ {l} \\otimes \\mathbf {E} _ {\\text {f a c e}} ^ {l} \\left(X _ {t}\\right) + \\left(1 - S _ {t} ^ {l}\\right) \\otimes \\mathbf {G} ^ {l - 1} \\left(X _ {t}\\right). \\tag {2}", + "image_path": "fb8fd3aa8c9bfafd176d8436ed406f9fcbf27236deb25e15e471a7304678af06.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 350, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 545, + 387 + ], + "type": "text", + "content": "Ablation studies (Section 5.4) show that SaMFs improve the fidelity of the mouth since they separate masked and non-masked regions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 393, + 373, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 373, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 373, + 404 + ], + "type": "text", + "content": "3.3.Encoders" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 411, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 545, + 436 + ], + "type": "text", + "content": "Our model has three different encoders: face encoder " + }, + { + "bbox": [ + 306, + 411, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{face}" + }, + { + "bbox": [ + 306, + 411, + 545, + 436 + ], + "type": "text", + "content": ", reference encoder " + }, + { + "bbox": [ + 306, + 411, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{ref}" + }, + { + "bbox": [ + 306, + 411, + 545, + 436 + ], + "type": "text", + "content": ", and audio encoder " + }, + { + "bbox": [ + 306, + 411, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{aud}" + }, + { + "bbox": [ + 306, + 411, + 545, + 436 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "text", + "content": "Face encoder " + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{face}" + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "text", + "content": " takes masked video frames " + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "inline_equation", + "content": "X_{1:T}" + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "text", + "content": " as input and then outputs " + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "text", + "content": " 2D spatial features " + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{face}^{l}(X_{t}) = \\{\\mathbf{E}_{face}^{l}(X_{t})\\mid l\\in [1,2,\\dots ,L]\\}" + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "text", + "content": ". These features are injected into the decoder " + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 305, + 436, + 545, + 508 + ], + "type": "text", + "content": " through the style-aware masked fusion to efficiently preserve 2D spatial structure, as described in Section 3.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "type": "text", + "content": "Reference Encoder " + }, + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{ref}" + }, + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "type": "text", + "content": " maps a reference " + }, + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "type": "inline_equation", + "content": "X_{ref}" + }, + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "type": "inline_equation", + "content": "2L" + }, + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "type": "text", + "content": " reference style codes, each of which has 512 dimensions. We simply denote the reference style code as " + }, + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "type": "inline_equation", + "content": "w_{ref} = [w_{ref}^{1}|w_{ref}^{2}|\\dots |w_{ref}^{2L}] \\in \\mathbb{R}^{512 \\times 2L}" + }, + { + "bbox": [ + 305, + 509, + 545, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": "Similar to " + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{ref}" + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": ", audio encoder " + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{aud}" + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": " maps a single audio segment into " + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "inline_equation", + "content": "2L" + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": " audio style codes, each of which has 512 dimensions. As we use " + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": " consecutive audio segments " + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "inline_equation", + "content": "A_{1:T}" + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{aud}" + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": " independently maps " + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "inline_equation", + "content": "a_{t} = [a_{t}^{1}|a_{t}^{2}|\\dots |a_{t}^{2L}] \\in \\mathbb{R}^{512 \\times 2L}" + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": ". We simply denote " + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "inline_equation", + "content": "a_{1:T} = (a_{t})_{t=1}^{T}" + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 557, + 545, + 639 + ], + "type": "text", + "content": " audio style codes. Please refer to our supplementary materials for the detailed encoder architectures." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "text", + "content": "From these style codes " + }, + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "inline_equation", + "content": "w_{ref}" + }, + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "inline_equation", + "content": "a_{1:T}" + }, + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "text", + "content": ", we compute target video's style codes over frames " + }, + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "inline_equation", + "content": "w_{1:T}" + }, + { + "bbox": [ + 305, + 641, + 545, + 665 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 369, + 670, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 670, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 369, + 670, + 545, + 685 + ], + "type": "interline_equation", + "content": "w _ {1: T} = \\left(w _ {1}, w _ {2}, \\dots , w _ {T}\\right), \\tag {3}", + "image_path": "373aacfe92e194309a6ead23838ebf263971e1aec013628be7ea93d90e54e196.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_{t} = w_{ref} + a_{t}\\in \\mathcal{W}+" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": ". We compute the style codes by simply adding these two different codes based on the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22844" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "linearity [13, 32] of the latent space " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": ". The style codes " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "w_{1:T}" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " are then fed into the decoder " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " through the learned affine transformations to generate synced lip motion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": "Temporal Consistency. Thanks to the semantically rich latent space " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": ", our model can generate an accurate lip-sync video frame by frame. However, this frame-wisely independent encoding of style codes turns out to lead inconsistent mouth movements in the final results. To remedy this, we assume that the generated style codes " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "w_{1:T}" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": " form a trajectory of the target video [40] in " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{W}+" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": " and enforce the smooth local transition of the motion [24, 33, 39] into the trajectory. Toward this, we introduce Moving-average based Latent Smoothing (MaLS), each of which consists of a stack of the weighted moving-average [3] and 1D convolutions operating on the style codes along the time-axis. More precisely, we employ " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "2L" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": " MaLS for " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": " resolutions, each of which takes the " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": "-th component of " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "w_{t-1}, w_t" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "w_{t+1}" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": " as its inputs to learn the local difference between them, and then we inject the local motions into " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "w_{ref}" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": " to compute the smooth style code " + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\tilde{w}_t" + }, + { + "bbox": [ + 46, + 109, + 288, + 312 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 71, + 316, + 287, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 316, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 71, + 316, + 287, + 367 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {w} _ {t} ^ {l} = w _ {r e f} ^ {l} + \\mathrm {M a L S} ^ {l} (w _ {t - 1: t + 1} ^ {l}), \\\\ = w _ {r e f} ^ {l} + \\operatorname {C o n v 1 D s} \\left(\\sum_ {\\tau = t - 1} ^ {t + 1} \\alpha_ {\\tau} \\cdot w _ {\\tau} ^ {l}\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "720f9b0c2d0e1199745908a98d4089d4719db53741fad8910482ea6b706494b8.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 373, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 287, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 373, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{MaLS}^l" + }, + { + "bbox": [ + 47, + 373, + 287, + 397 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 47, + 373, + 287, + 397 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 373, + 287, + 397 + ], + "type": "text", + "content": "-th MaLS, and " + }, + { + "bbox": [ + 47, + 373, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\alpha_{\\tau}" + }, + { + "bbox": [ + 47, + 373, + 287, + 397 + ], + "type": "text", + "content": " is the weight of the moving-average." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 398, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 398, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 287, + 421 + ], + "type": "text", + "content": "With this smooth latent codes " + }, + { + "bbox": [ + 47, + 398, + 287, + 421 + ], + "type": "inline_equation", + "content": "\\tilde{w}_{1:T}" + }, + { + "bbox": [ + 47, + 398, + 287, + 421 + ], + "type": "text", + "content": ", we compute the video frames:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 426, + 287, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 426, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 97, + 426, + 287, + 441 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {1: T} = \\left(\\mathbf {G} \\left(\\tilde {w} _ {t}, \\mathbf {E} _ {\\text {f a c e}} \\left(X _ {t}\\right)\\right)\\right) _ {t = 1} ^ {T}. \\tag {5}", + "image_path": "765a8ff494793e9bb788e75318fa17db988edcc8a2597867fdb695e8492b4aa7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 446, + 286, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 446, + 286, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 446, + 286, + 471 + ], + "type": "text", + "content": "For better initialization [29], we add the average code " + }, + { + "bbox": [ + 47, + 446, + 286, + 471 + ], + "type": "inline_equation", + "content": "w_{avg}" + }, + { + "bbox": [ + 47, + 446, + 286, + 471 + ], + "type": "text", + "content": " of the pre-trained generator to each " + }, + { + "bbox": [ + 47, + 446, + 286, + 471 + ], + "type": "inline_equation", + "content": "\\tilde{w}_t" + }, + { + "bbox": [ + 47, + 446, + 286, + 471 + ], + "type": "text", + "content": " in (4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 476, + 158, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 158, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 158, + 490 + ], + "type": "text", + "content": "3.4. Training Objective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 495, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 287, + 567 + ], + "type": "text", + "content": "We train StyleLipSync to reconstruct target video frames from corresponding audio. We randomly choose " + }, + { + "bbox": [ + 47, + 495, + 287, + 567 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 47, + 495, + 287, + 567 + ], + "type": "text", + "content": " consequent frames with corresponding audio segments and 1 single reference frame. Image perceptual loss [43] is used to minimize perceptual image distance between generated frames " + }, + { + "bbox": [ + 47, + 495, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 47, + 495, + 287, + 567 + ], + "type": "text", + "content": " and ground-truth frames " + }, + { + "bbox": [ + 47, + 495, + 287, + 567 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 495, + 287, + 567 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 96, + 572, + 287, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 572, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 96, + 572, + 287, + 604 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l p i p s} = \\sum_ {i = 1} ^ {N} \\left\\| \\phi^ {i} (\\hat {X}) - \\phi^ {i} (Y) \\right\\| _ {2}, \\tag {6}", + "image_path": "56ed3aa9848f443cf1c4f60765d2dc8860d72ebc2f4869d606495de63206bbdf.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "text", + "content": " is the number of feature extractor, " + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\phi^i (\\cdot)" + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "text", + "content": " th feature extractor, and " + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\ell^2" + }, + { + "bbox": [ + 47, + 610, + 287, + 647 + ], + "type": "text", + "content": " loss. Similar to [33], we use the multi-scale perceptual loss with 3 levels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 647, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 694 + ], + "type": "text", + "content": "For audio-visual alignment, we utilize SyncNet trained in a contrastive manner [24] that minimizes the cosine distance between generated frames " + }, + { + "bbox": [ + 47, + 647, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 47, + 647, + 287, + 694 + ], + "type": "text", + "content": " and corresponding audio segment " + }, + { + "bbox": [ + 47, + 647, + 287, + 694 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 647, + 287, + 694 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 97, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 97, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s y n c}} = 1 - \\cos \\left(f _ {v} (\\hat {X}), f _ {a} (A)\\right), \\tag {7}", + "image_path": "3c98339c0fcc382e39b1a0653c4a247abb85aa0944c1ced2911e192ba6494081.jpg" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 171 + ], + "type": "image", + "image_path": "1a5d8d6b21a61f6c4547432ad4bfd5c27165049cda41b35edd0040e6ed0528ad.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 171, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 216 + ], + "type": "text", + "content": "Figure 4. Adaptation for Unseen Face. We slightly tune the decoder " + }, + { + "bbox": [ + 304, + 171, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\theta}" + }, + { + "bbox": [ + 304, + 171, + 545, + 216 + ], + "type": "text", + "content": " with the proposed sync regularizer " + }, + { + "bbox": [ + 304, + 171, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{sync}" + }, + { + "bbox": [ + 304, + 171, + 545, + 216 + ], + "type": "text", + "content": ", while freezing all encoders' weight. Face encoder " + }, + { + "bbox": [ + 304, + 171, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{face}" + }, + { + "bbox": [ + 304, + 171, + 545, + 216 + ], + "type": "text", + "content": " and SaMFs are omitted here for simplicity." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "type": "text", + "content": " denotes the cosine similarity. " + }, + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "type": "inline_equation", + "content": "f_{v}(\\cdot)" + }, + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "type": "inline_equation", + "content": "f_{a}(\\cdot)" + }, + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "type": "text", + "content": " are the frame and audio feature extractor, respectively. The final objective " + }, + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{train}" + }, + { + "bbox": [ + 305, + 221, + 545, + 258 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 362, + 265, + 545, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 265, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 362, + 265, + 545, + 278 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a i n}} = \\lambda_ {1} \\mathcal {L} _ {\\text {l p i p s}} + \\lambda_ {2} \\mathcal {L} _ {\\text {s y n c}}, \\tag {8}", + "image_path": "7582d19c6b9bef7cb85801e68a67bec8c40839e07dfdb94e39929d6b7fa06034.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 285, + 496, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 496, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 496, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 285, + 496, + 297 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 306, + 285, + 496, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 285, + 496, + 297 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 306, + 285, + 496, + 297 + ], + "type": "text", + "content": " are the balancing coefficients." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 306, + 446, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 446, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 446, + 319 + ], + "type": "text", + "content": "4. Unseen Face Adaptation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 326, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 410 + ], + "type": "text", + "content": "Although StyleLipSync successfully generates accurately lip-synced videos with high fidelity, the model would fail to exactly synthesize unseen faces lying in the out-of-distribution. This problem refers to the failure of id preservation. Therefore, to handle this, we fine-tune our decoder " + }, + { + "bbox": [ + 304, + 326, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 304, + 326, + 545, + 410 + ], + "type": "text", + "content": " on a target person video to make a personal model that is better able to synthesize toward the target person." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "X_{1:T}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " be masked video frames of unseen face, which corresponds to the audio segments " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "A_{1:T}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "X_{ref}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " be a reference frame. The frozen encoders convert each input into the intermediate representations, " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\tilde{w}_{1:T}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{face}(X_t))_{t=1}^T" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": ". Then we fine-tune the decoder " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\theta}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": ", now parameterized by " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": ", by minimizing the distance between " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "(\\mathbf{G}_{\\theta}(\\tilde{w}_t,\\mathbf{E}_{face}(X_t)))_{t=1}^T" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " and target frames, as same as (8). However, fine-tuning the decoder on a short video of a single identity leads to over-fitting and losing the lip-sync generality as the generator can memorize the target video [26]. To prevent the model from this scenario, we introduce a sync regularizer " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{sync}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " to enforce audio generality to the decoder " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\theta}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " by leveraging the audio from the training dataset, not from the target video. Formally, given audio segments " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "A_{1:T}^{\\prime}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " randomly chosen from the training data (Voxceleb2 [7]) and " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "w_{ref}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": ", we compute smooth style codes " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\tilde{w}_{1:T}^{\\prime}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": ", and then decode them to a synced video " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "X_{1:T}^{\\prime}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": ". The sync regularizer " + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{sync}" + }, + { + "bbox": [ + 304, + 410, + 546, + 626 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 353, + 632, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 632, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 353, + 632, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {\\text {s y n c}} = 1 - \\cos \\left(f _ {v} \\left(\\hat {X} ^ {\\prime}\\right), f _ {a} \\left(A ^ {\\prime}\\right)\\right), \\tag {9}", + "image_path": "c30100006077a250c927c45b47577cfb16bf9a7a3d11822c72a4454d16a2f23c.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": "which enforces " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\theta}" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\hat{X}_{1:T}^{\\prime}" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " aligning with " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "A_{1:T}^{\\prime}" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": ". The final object for a single person adaptation is given as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 356, + 696, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 696, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 356, + 696, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\underset {\\theta} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\text {t r a i n}} + \\lambda_ {R} \\mathcal {R} _ {\\text {s y n c}}, \\tag {10}", + "image_path": "b4f2d88fc8b37bf3763b0e5cd7cc9aa1f5f06b781f77a8fe15d651ba629a04c3.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22845" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 294, + 373 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 294, + 373 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 294, + 373 + ], + "type": "image", + "image_path": "61ffa634159844f0c6b05d9e17ce424811b8b2a60edffcbb4166e3ea5f1ed34c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 378, + 259, + 388 + ], + "lines": [ + { + "bbox": [ + 87, + 378, + 259, + 388 + ], + "spans": [ + { + "bbox": [ + 87, + 378, + 259, + 388 + ], + "type": "text", + "content": "(a) Reconstruction results on Voxceleb2 [7] test data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 392, + 533, + 403 + ], + "lines": [ + { + "bbox": [ + 47, + 392, + 533, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 533, + 403 + ], + "type": "text", + "content": "Figure 5. Comparison with state-of-the-art methods. The different field of view comes from the pre-processing strategy of each model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 76, + 542, + 373 + ], + "blocks": [ + { + "bbox": [ + 301, + 76, + 542, + 373 + ], + "lines": [ + { + "bbox": [ + 301, + 76, + 542, + 373 + ], + "spans": [ + { + "bbox": [ + 301, + 76, + 542, + 373 + ], + "type": "image", + "image_path": "78368105a11b5e3fa251e766852549e3bd6d103febcc4822321b71cd11d95889.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 378, + 477, + 388 + ], + "lines": [ + { + "bbox": [ + 362, + 378, + 477, + 388 + ], + "spans": [ + { + "bbox": [ + 362, + 378, + 477, + 388 + ], + "type": "text", + "content": "(b) Cross-id results on HDTF [44]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 415, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 415, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\lambda_R" + }, + { + "bbox": [ + 47, + 415, + 287, + 452 + ], + "type": "text", + "content": " is the regularizer coefficient. Ablation studies in Section 5.4 show that " + }, + { + "bbox": [ + 47, + 415, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{sync}" + }, + { + "bbox": [ + 47, + 415, + 287, + 452 + ], + "type": "text", + "content": " regularizes the audio generality even if we use the audio from the training set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 128, + 474 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 479, + 105, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 105, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 105, + 491 + ], + "type": "text", + "content": "5.1. Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 713 + ], + "type": "text", + "content": "We train our model on Voxceleb2 [7], which consists of in-the-wild talking face videos collected from YouTube. It contains more than 145,000 videos of about 6,100 identities in the train set and more than 4,900 videos of more than 110 identities in the test set. We convert all videos into frames with 25 fps and then crop and resize those frames into " + }, + { + "bbox": [ + 46, + 498, + 289, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 498, + 289, + 713 + ], + "type": "text", + "content": " resolution, following the method of [33]. For more semantically rich face priors, we only use videos where the detected face bounding boxes' height (and width) is longer than 256. After pre-processing, the remaining 11051 videos and 340 videos are used as the training set and the test set, respectively. All audios are re-sampled to " + }, + { + "bbox": [ + 46, + 498, + 289, + 713 + ], + "type": "inline_equation", + "content": "16\\mathrm{kHz}" + }, + { + "bbox": [ + 46, + 498, + 289, + 713 + ], + "type": "text", + "content": " and converted into a mel-spectrogram to be used as our audio representation similar to the method in [45, 24]. We also use HDTF [44] to further test our model for crossid experiments. It is widely used to evaluate high-resolution talking face generative models, where the head pose dynamics are not as significant as the Voxceleb2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 415, + 440, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 440, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 440, + 427 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": "We pre-train StyleGAN2 [19] on the Voxceleb2 [7] following the implementation of [16]. For training StyleLipSync and SyncNet, we set the frame length " + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": ", and employ Adam [20] optimizer with learning rate " + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": " throughout the training phase in both cases. All experiments are performed on 2 TITAN RTX GPUs. We set " + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 10" + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.1" + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\lambda_{R} = 0.1" + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": " for the objective (10), and " + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\alpha_{t-1} = 0.25" + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 0.5" + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\alpha_{t+1} = 0.25" + }, + { + "bbox": [ + 304, + 434, + 545, + 540 + ], + "type": "text", + "content": " for MaLS (4). For all inference, we use the first frame as the reference image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 542, + 546, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 546, + 662 + ], + "type": "text", + "content": "For image quality metric, we use PSNR, and SSIM. We also calculate CSIM [10] in the cross-id experiment that measures the face similarity between the images in the pretrained face embedding space. For lip-sync quality metric, we use LMD, LSE-D, and LSE-C. LMD [5] is the absolute distance of facial landmarks between the target and generated frames. LSE-D and LSE-C are proposed in [28], where LSE-D measures the distance between the lip and audio representations, and LSE-C measures the lip-sync confidence, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 670, + 380, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 380, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 380, + 682 + ], + "type": "text", + "content": "5.3. Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": "Reconstruction Results. We compare the other state-of-the-art methods in the lip-sync (Wav2Lip [28]) and talking" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22846" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "face generation (ATVG [5], MakeItTalk [46], and PC-AVS [45]) on reconstruction results of Voxceleb2 test dataset. Table 1 shows that StyleLipSync outperforms all metrics except the LSE-C score. Wav2Lip [28] achieves the highest LSE-C score, however, it achieves low image quality metrics since it generates " + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": " low-resolution videos. PC-AVS [45] achieves comparable lip-sync scores with StyleLipSync, however, achieves low image quality metrics than our model since it highly relies on its specific preprocessing and fails to generate in many cases. We also illustrate the qualitative results in Figure 4. Wav2Lip [28] produces a lip-synchronized video with considerable visual artifacts since it cannot adapt to the video of dynamic head pose. MakeItTalk [46] generates low lip-sync video since it uses sparse facial landmarks. PC-AVS [45] generates an accurate lip-synchronized video following the input head pose. However, it struggles to preserve the unseen identity involving visual artifacts. StyleLipSync generates the natural lip motion with high fidelity and preserves the input identity, which is comparable to the ground truth video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 313, + 289, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 289, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 289, + 636 + ], + "type": "text", + "content": "Cross Identity Results. To evaluate lip-sync generalization, we conduct cross-id experiments settings using unseen videos and audio. We randomly sample 10 videos of different identities and 150 audio without duplication from HDTF [44]. We use the first 10 seconds of videos and audio. For each video, we generate 15 lip-synced videos from 15 different audios, where those face-audio pairs are not from the same source. In Table 2, we report the LSE-D and LSE-C for lip-sync quality and the CSIM for face similarity. Wav2Lip [28] achieves a higher LSE-C score than StyleLipSync, however, it achieves a low CSIM score since it produces a low-resolution video with visual artifacts. MakeItTalk [46] achieves the best CSIM score, while the lip-sync quality is the worst since it uses sparse facial landmarks. PC-AVS [45] outperforms LSE-C while achieving the lowest CSIM since it can't preserve an unseen face's identity. StyleLipSync achieves the best LSE-D score and comparable CSIM score to MakeItTalk [46]. We show qualitative results of cross-id experiments with target lip references in Figure 4. PC-AVS [45] can generate accurate lip-synchronizing video compared to the target lip, while it fails to preserve facial details of the unseen face. MakeItTalk [46] produces a high-resolution and identity-preserving video, however, it is out-of-sync compared to the target. StyleLipSync generates a high-resolution lip synchronizing video compared to the target lip, without any visual artifacts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 646, + 149, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 149, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 149, + 658 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "Ablation Studies on Zero-shot Model. Figure 6 and Table 3 summarize the ablation studies on our zero-shot method of reconstruction on Voxceleb2 test data. If we replace the pose-aware masking with standard rectangular" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 308, + 98, + 547, + 171 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 547, + 91 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 547, + 91 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 547, + 91 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of reconstruction on Voxceleb2 test data. The best score for each metric is in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 98, + 547, + 171 + ], + "lines": [ + { + "bbox": [ + 308, + 98, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 98, + 547, + 171 + ], + "type": "table", + "html": "
MethodVoxceleb2 (Reconstruction)
ImageLip-Sync
SSIM↑PSNR↑LMD↓LSE-D↓LSE-C↑
Wav2Lip96×96 [28]0.44813.5346.4226.9998.329
ATVG128×128 [5]0.46113.3497.1658.8215.421
MakeItTalk256×256 [46]0.41912.8683.64910.8953.624
PC-AVS224×224 [45]0.36913.2102.8127.2787.699
Ours256×2560.63119.6072.6966.6288.056
", + "image_path": "c6978e33c12a382fd8c505b0c36d3c2716c37d7b65d67e210c42d7102449856e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 312, + 232, + 541, + 328 + ], + "blocks": [ + { + "bbox": [ + 305, + 181, + 547, + 226 + ], + "lines": [ + { + "bbox": [ + 305, + 181, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 305, + 181, + 547, + 226 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of cross-identity results for unseen face. We report CSIM [10] as the image quality metric since there is no ground truth frames for the cross-id experiments. The best score for each metric is in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 232, + 541, + 328 + ], + "lines": [ + { + "bbox": [ + 312, + 232, + 541, + 328 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 541, + 328 + ], + "type": "table", + "html": "
MethodHDTF (Cross-id)
ImageLip-Sync
CSIM ↑LSE-D ↓LSE-C ↑
Wav2Lip96×96 [28]0.6567.0478.576
ATVG128×128 [5]0.2878.6686.040
MakeItTalk256×256 [46]0.77010.6414.725
PC-AVS224×224 [45]0.2386.9218.858
Ours256×2560.7376.8258.209
", + "image_path": "2c250f0d338fe21a19be6621f6bb277c0c6bd914cc7f47fcfd60b565c5f08dec.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 366, + 542, + 439 + ], + "blocks": [ + { + "bbox": [ + 305, + 337, + 547, + 359 + ], + "lines": [ + { + "bbox": [ + 305, + 337, + 547, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 337, + 547, + 359 + ], + "type": "text", + "content": "Table 3. Ablation study on zero-shot model. The best score for each metric is in **bold**." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 366, + 542, + 439 + ], + "lines": [ + { + "bbox": [ + 310, + 366, + 542, + 439 + ], + "spans": [ + { + "bbox": [ + 310, + 366, + 542, + 439 + ], + "type": "table", + "html": "
Method (ours)Voceleb2 (Reconstruction)
ImageLip-Sync
SSIM ↑PSNR ↑LMD ↓LSE-D ↓LSE-C ↓
w/o Pose Mask0.60218.8673.0576.7717.748
w/o MaLS0.59318.1862.7406.9947.577
w/o SaMF0.59118.1812.7646.8387.780
Full0.63119.6072.6966.6288.056
", + "image_path": "ebd4a00f648c629b285e20069f2cd95d64324f3ca4d80202ce20212112b912d2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 460, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 592 + ], + "type": "text", + "content": "masking (w/o Pose Mask) in [28, 6, 27], considerable visual artifacts occur around the masked region since it is insufficient to capture the pose difference between the reference and the target. To validate SaMF, we replace the modulated convolutions in SaMFs with the standard convolutions. Figure 6(c) shows that SaMFs significantly improve lip region's fidelity since the modulated convolution helps to be aware of the lip style. As shown in Table 3, MaLS significantly improves lip-sync quality, which cannot be reflected in a single image. Please refers to our project page for ablation studies on MaLS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Ablation Studies on Unseen Face Adaptation. We conduct ablation studies on the proposed unseen face adaptation following the same setting in Table 2. Additionally, we use 60 seconds of video for each 10 personalized models and 15 audios of 10 seconds from different identities for inference. Figure 7(b) shows the lip-sync metrics according to the adaptation step. In the cases without the sync regularizer, the models lose the lip-sync generality, in other words, it memorizes the short target video as the adaptation phase proceeds. Introducing the sync regularizer with the sync" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22847" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 106, + 129 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 106, + 129 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 106, + 129 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 106, + 129 + ], + "type": "image", + "image_path": "698d1f2001ae813ddf0bc7faa0819f15a4a6327ef2ad943f78924cd17ba9b61e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 72, + 166, + 129 + ], + "blocks": [ + { + "bbox": [ + 110, + 72, + 166, + 129 + ], + "lines": [ + { + "bbox": [ + 110, + 72, + 166, + 129 + ], + "spans": [ + { + "bbox": [ + 110, + 72, + 166, + 129 + ], + "type": "image", + "image_path": "a4708a0e1698c232cded03e42c3e435857ef6d8c8d9ac3aee3b0401285f920c7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 72, + 227, + 129 + ], + "blocks": [ + { + "bbox": [ + 170, + 72, + 227, + 129 + ], + "lines": [ + { + "bbox": [ + 170, + 72, + 227, + 129 + ], + "spans": [ + { + "bbox": [ + 170, + 72, + 227, + 129 + ], + "type": "image", + "image_path": "619085aa427a0a96bb5669e0deb157594248c33aee01de3928ce3a7976d193de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 286, + 129 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 286, + 129 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 286, + 129 + ], + "type": "image", + "image_path": "4b8d79bff406be7ef24efd9c84151e682448a47a3b79e53d3e6a35f4e0f455b6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 130, + 106, + 186 + ], + "blocks": [ + { + "bbox": [ + 50, + 130, + 106, + 186 + ], + "lines": [ + { + "bbox": [ + 50, + 130, + 106, + 186 + ], + "spans": [ + { + "bbox": [ + 50, + 130, + 106, + 186 + ], + "type": "image", + "image_path": "48b1c0f911668b356bf3202d0eb401bb8369a9db52e7ad656b2529e6453708af.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 193, + 99, + 201 + ], + "lines": [ + { + "bbox": [ + 57, + 193, + 99, + 201 + ], + "spans": [ + { + "bbox": [ + 57, + 193, + 99, + 201 + ], + "type": "text", + "content": "(a) Reference" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 110, + 130, + 166, + 187 + ], + "blocks": [ + { + "bbox": [ + 110, + 130, + 166, + 187 + ], + "lines": [ + { + "bbox": [ + 110, + 130, + 166, + 187 + ], + "spans": [ + { + "bbox": [ + 110, + 130, + 166, + 187 + ], + "type": "image", + "image_path": "51a18be1cd5024a406193b05721bcf4696e7ad7879c3d4ac875519e16b0f41ff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 193, + 165, + 201 + ], + "lines": [ + { + "bbox": [ + 110, + 193, + 165, + 201 + ], + "spans": [ + { + "bbox": [ + 110, + 193, + 165, + 201 + ], + "type": "text", + "content": "(b) w/o Pose mask" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 170, + 130, + 227, + 187 + ], + "blocks": [ + { + "bbox": [ + 170, + 130, + 227, + 187 + ], + "lines": [ + { + "bbox": [ + 170, + 130, + 227, + 187 + ], + "spans": [ + { + "bbox": [ + 170, + 130, + 227, + 187 + ], + "type": "image", + "image_path": "404f867ccb7b672d7bf5b7d2562e8d44493f7c58932ddcf5402b070e39821a03.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 193, + 219, + 201 + ], + "lines": [ + { + "bbox": [ + 176, + 193, + 219, + 201 + ], + "spans": [ + { + "bbox": [ + 176, + 193, + 219, + 201 + ], + "type": "text", + "content": "(c) w/o SaMF" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 230, + 130, + 286, + 187 + ], + "blocks": [ + { + "bbox": [ + 230, + 130, + 286, + 187 + ], + "lines": [ + { + "bbox": [ + 230, + 130, + 286, + 187 + ], + "spans": [ + { + "bbox": [ + 230, + 130, + 286, + 187 + ], + "type": "image", + "image_path": "beace11418acd0a701ff56659bb5fbd5c76762bf6938d414471860ffe3c566be.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 193, + 270, + 201 + ], + "lines": [ + { + "bbox": [ + 244, + 193, + 270, + 201 + ], + "spans": [ + { + "bbox": [ + 244, + 193, + 270, + 201 + ], + "type": "text", + "content": "(d) Full" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 204, + 263, + 214 + ], + "lines": [ + { + "bbox": [ + 70, + 204, + 263, + 214 + ], + "spans": [ + { + "bbox": [ + 70, + 204, + 263, + 214 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of zero-shot model." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 52, + 262, + 282, + 357 + ], + "blocks": [ + { + "bbox": [ + 47, + 224, + 286, + 257 + ], + "lines": [ + { + "bbox": [ + 47, + 224, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 286, + 257 + ], + "type": "text", + "content": "Table 4. Mean Opinion score (MOS) user study results with " + }, + { + "bbox": [ + 47, + 224, + 286, + 257 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 47, + 224, + 286, + 257 + ], + "type": "text", + "content": " confidence interval on cross-id setting. The score ranges in 1 to 5. The best score for each metric is in bold." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 262, + 282, + 357 + ], + "lines": [ + { + "bbox": [ + 52, + 262, + 282, + 357 + ], + "spans": [ + { + "bbox": [ + 52, + 262, + 282, + 357 + ], + "type": "table", + "html": "
MethodUser Study (MOS)
Lip-sync AccuracyFace SimilarityVisual Quality
Wav2Lip [28]3.76 ± 0.202.98 ± 0.252.03 ± 0.21
ATVG [5]2.19 ± 0.212.45 ± 0.251.54 ± 0.17
MakeItTalk [46]2.32 ± 0.243.47 ± 0.232.95 ± 0.24
PC-AVS [45]3.28 ± 0.212.55 ± 0.222.51 ± 0.22
Ours4.01 ± 0.163.42 ± 0.223.55 ± 0.20
Ours (Personalized)3.52 ± 0.214.03 ± 0.173.64 ± 0.19
", + "image_path": "ad6a4c5ca6ff945801715a2bce8843df697cb1c5aeea983b42aa8d19e0c54276.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 374, + 286, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 286, + 554 + ], + "type": "text", + "content": "loss stabilizes the lip-sync metrics and even improves them compared to the zero-shot results. If we use the sync regularizer without the sync loss, lip-sync quality is stabilized, however slightly lower than the zero-shot results due to the lack of ground-truth audio-visual correlation. Audio generalization for unseen face data is maintained even though we used audio from the learning data. Figure 7(a) supports the validity of the adaptation method. It shows that visual difference between the zero-shot results and the adaptation results. The zero-shot model can generate an accurate lip motion for the target audio, while it shows a slight difference in person-specific details compared to the reference images. Through the proposed adaptation method, personal-specific lip shape, teeth, and wrinkles are faithfully recovered while maintaining the lip motion of the zero-shot results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 563, + 121, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 121, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 121, + 575 + ], + "type": "text", + "content": "5.5. User Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 582, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 714 + ], + "type": "text", + "content": "We further conduct a user study based on MOS (Mean opinion score) to compare the perceptual quality of each model, including our zero-shot and personalized model. 5 videos generated by each model in cross-id setting are used for this study. 20 participants scored lip-sync accuracy, face similarity, and visual quality of each video in the range of 1 to 5. As shown in Table 4, our models outperform all metrics. Specifically, our zero-shot model achieves the highest lip-sync accuracy, and our adaptation model achieves the highest score in face similarity and visual quality with competitive lip-sync accuracy." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 309, + 79, + 354, + 125 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 345, + 78 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 345, + 78 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 345, + 78 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 79, + 354, + 125 + ], + "lines": [ + { + "bbox": [ + 309, + 79, + 354, + 125 + ], + "spans": [ + { + "bbox": [ + 309, + 79, + 354, + 125 + ], + "type": "image", + "image_path": "b822265a98179096967236193e26baf6594f688045c67653f7cab4a9f704b1b3.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 355, + 79, + 375, + 119 + ], + "blocks": [ + { + "bbox": [ + 355, + 79, + 375, + 119 + ], + "lines": [ + { + "bbox": [ + 355, + 79, + 375, + 119 + ], + "spans": [ + { + "bbox": [ + 355, + 79, + 375, + 119 + ], + "type": "image", + "image_path": "f294511969cd368b685fbefd6096572670fe578201e4260dcfac6118b720cab6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 375, + 79, + 400, + 119 + ], + "blocks": [ + { + "bbox": [ + 375, + 79, + 400, + 119 + ], + "lines": [ + { + "bbox": [ + 375, + 79, + 400, + 119 + ], + "spans": [ + { + "bbox": [ + 375, + 79, + 400, + 119 + ], + "type": "image", + "image_path": "6cc51cc1828e797f22d508ade1792ee6550fb1af8676a064d05e789732ebf0b4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 400, + 79, + 439, + 125 + ], + "blocks": [ + { + "bbox": [ + 400, + 79, + 439, + 125 + ], + "lines": [ + { + "bbox": [ + 400, + 79, + 439, + 125 + ], + "spans": [ + { + "bbox": [ + 400, + 79, + 439, + 125 + ], + "type": "image", + "image_path": "138cb89af7151cfd97a94847703ebb601befeb5c89d6388d9c95d5cb1af7a857.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 439, + 79, + 542, + 125 + ], + "blocks": [ + { + "bbox": [ + 439, + 79, + 542, + 125 + ], + "lines": [ + { + "bbox": [ + 439, + 79, + 542, + 125 + ], + "spans": [ + { + "bbox": [ + 439, + 79, + 542, + 125 + ], + "type": "image", + "image_path": "1b9d0caf51bc87ac06851bc69d154c9ca29b7e64ee1c3b91295994bcf60736b4.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 354, + 126, + 369, + 165 + ], + "blocks": [ + { + "bbox": [ + 354, + 126, + 369, + 165 + ], + "lines": [ + { + "bbox": [ + 354, + 126, + 369, + 165 + ], + "spans": [ + { + "bbox": [ + 354, + 126, + 369, + 165 + ], + "type": "image", + "image_path": "1296d6d1f743fec74111cbfa48291a7ddcb0b50b8a8303963c96da3617c243af.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 370, + 126, + 396, + 165 + ], + "blocks": [ + { + "bbox": [ + 370, + 126, + 396, + 165 + ], + "lines": [ + { + "bbox": [ + 370, + 126, + 396, + 165 + ], + "spans": [ + { + "bbox": [ + 370, + 126, + 396, + 165 + ], + "type": "image", + "image_path": "7b33774d7de86e7eac1bd941f7a06a64137f38960dbf4df109dc7660bcd91c2d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 397, + 126, + 436, + 165 + ], + "blocks": [ + { + "bbox": [ + 397, + 126, + 436, + 165 + ], + "lines": [ + { + "bbox": [ + 397, + 126, + 436, + 165 + ], + "spans": [ + { + "bbox": [ + 397, + 126, + 436, + 165 + ], + "type": "image", + "image_path": "5c51ddc39c4015f8c831c6eb8d33c5c8a82554020d9e845eb08c58545bab80fc.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 436, + 126, + 542, + 173 + ], + "blocks": [ + { + "bbox": [ + 436, + 126, + 542, + 173 + ], + "lines": [ + { + "bbox": [ + 436, + 126, + 542, + 173 + ], + "spans": [ + { + "bbox": [ + 436, + 126, + 542, + 173 + ], + "type": "image", + "image_path": "ed7fdd7a259014ca277e120771bd702cda400120f06a5b792889c4cc10cc2eea.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 309, + 184, + 354, + 230 + ], + "blocks": [ + { + "bbox": [ + 318, + 178, + 345, + 184 + ], + "lines": [ + { + "bbox": [ + 318, + 178, + 345, + 184 + ], + "spans": [ + { + "bbox": [ + 318, + 178, + 345, + 184 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 184, + 354, + 230 + ], + "lines": [ + { + "bbox": [ + 309, + 184, + 354, + 230 + ], + "spans": [ + { + "bbox": [ + 309, + 184, + 354, + 230 + ], + "type": "image", + "image_path": "7cac37e4dfb3a2d148ec5686177cb512a1ba646973558181088ac76c61b8ec5a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 355, + 184, + 375, + 230 + ], + "blocks": [ + { + "bbox": [ + 355, + 184, + 375, + 230 + ], + "lines": [ + { + "bbox": [ + 355, + 184, + 375, + 230 + ], + "spans": [ + { + "bbox": [ + 355, + 184, + 375, + 230 + ], + "type": "image", + "image_path": "381fc29f755cc4b721118146ac6f62961058002e8c3e07e2fcc73ade466d27cd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 375, + 184, + 400, + 230 + ], + "blocks": [ + { + "bbox": [ + 375, + 184, + 400, + 230 + ], + "lines": [ + { + "bbox": [ + 375, + 184, + 400, + 230 + ], + "spans": [ + { + "bbox": [ + 375, + 184, + 400, + 230 + ], + "type": "image", + "image_path": "2a28cda9da671a4dd4ace4823b7b930f9089cf94fef925aded126d0869d7c2fe.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 400, + 184, + 439, + 230 + ], + "blocks": [ + { + "bbox": [ + 400, + 184, + 439, + 230 + ], + "lines": [ + { + "bbox": [ + 400, + 184, + 439, + 230 + ], + "spans": [ + { + "bbox": [ + 400, + 184, + 439, + 230 + ], + "type": "image", + "image_path": "1ac1c3637c4abd36a4ff76f6b663453ddba4e6b761cff16e23cb55f376572163.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 439, + 184, + 542, + 230 + ], + "blocks": [ + { + "bbox": [ + 439, + 184, + 542, + 230 + ], + "lines": [ + { + "bbox": [ + 439, + 184, + 542, + 230 + ], + "spans": [ + { + "bbox": [ + 439, + 184, + 542, + 230 + ], + "type": "image", + "image_path": "820a54989705e6f0b4ab47b889ec449dca315dd045bc27cd5a1349f193695c12.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 354, + 232, + 369, + 257 + ], + "blocks": [ + { + "bbox": [ + 354, + 232, + 369, + 257 + ], + "lines": [ + { + "bbox": [ + 354, + 232, + 369, + 257 + ], + "spans": [ + { + "bbox": [ + 354, + 232, + 369, + 257 + ], + "type": "image", + "image_path": "1b77a2a94484b0da1c715e87fbbcd0590ba3ccb2304ecb088a63e2d5cce42cf4.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 370, + 232, + 402, + 278 + ], + "blocks": [ + { + "bbox": [ + 370, + 232, + 402, + 278 + ], + "lines": [ + { + "bbox": [ + 370, + 232, + 402, + 278 + ], + "spans": [ + { + "bbox": [ + 370, + 232, + 402, + 278 + ], + "type": "image", + "image_path": "40994e0b5b587add63603696f30d796ba6ddb8a44992ff2d638f6080d2242180.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 284, + 526, + 294 + ], + "lines": [ + { + "bbox": [ + 324, + 284, + 526, + 294 + ], + "spans": [ + { + "bbox": [ + 324, + 284, + 526, + 294 + ], + "type": "text", + "content": "(a) Comparison of cross-id results of zero-shot and adaptation." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 402, + 232, + 440, + 278 + ], + "blocks": [ + { + "bbox": [ + 402, + 232, + 440, + 278 + ], + "lines": [ + { + "bbox": [ + 402, + 232, + 440, + 278 + ], + "spans": [ + { + "bbox": [ + 402, + 232, + 440, + 278 + ], + "type": "image", + "image_path": "ec22210d6ed95457a1ac66d5d5f47f98f7a50c4d6b4d4c1fc38d5797d1842306.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 440, + 232, + 542, + 278 + ], + "blocks": [ + { + "bbox": [ + 440, + 232, + 542, + 278 + ], + "lines": [ + { + "bbox": [ + 440, + 232, + 542, + 278 + ], + "spans": [ + { + "bbox": [ + 440, + 232, + 542, + 278 + ], + "type": "image", + "image_path": "546dea6b7e3c73acf0e549d0d130dea73563c92340cbed7c016ce47afe784631.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 309, + 306, + 426, + 406 + ], + "blocks": [ + { + "bbox": [ + 309, + 306, + 426, + 406 + ], + "lines": [ + { + "bbox": [ + 309, + 306, + 426, + 406 + ], + "spans": [ + { + "bbox": [ + 309, + 306, + 426, + 406 + ], + "type": "image", + "image_path": "4a6aa5c2fc86a5b5c34f4eba7acc9f621bdb2de268c3042f674f966967d5c890.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 412, + 510, + 422 + ], + "lines": [ + { + "bbox": [ + 341, + 412, + 510, + 422 + ], + "spans": [ + { + "bbox": [ + 341, + 412, + 510, + 422 + ], + "type": "text", + "content": "(b) Ablation study on the proposed sync regularizer." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 426, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 306, + 426, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 545, + 460 + ], + "type": "text", + "content": "Figure 7. Experimental results of the proposed adaptation method using the cross-id setting. It improves the person specific visuals while maintaining the lip-sync generality." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 429, + 307, + 542, + 405 + ], + "blocks": [ + { + "bbox": [ + 429, + 307, + 542, + 405 + ], + "lines": [ + { + "bbox": [ + 429, + 307, + 542, + 405 + ], + "spans": [ + { + "bbox": [ + 429, + 307, + 542, + 405 + ], + "type": "image", + "image_path": "54c0008f606aed12cfba9029416b59432a1a18c28678d5e913e887b81ee7618e.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 486, + 378, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 486, + 378, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 486, + 378, + 498 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 714 + ], + "type": "text", + "content": "We proposed StyleLipSync, a lip-sync video generative model for arbitrary identity, which leverages expressive lip priors from a pre-trained style-based generator. In contrast to existing lip-sync generative models, we introduce pose-aware masking for lip region by utilizing a 3D parametric mesh predictor, which inherits the pose information in the mask itself. Designing a smooth lip motion by using the moving-average based latent smoothing in the continuous and linear latent space, StyleLipSync can generate temporally consistent lip motion. Furthermore, we propose a few-shot lip-sync adaptation method for a single person who lies in the out-of-distribution, which uses a few seconds of the target person's video. Experimental results show that our StyleLipSync can generate realistic lip-sync video from arbitrary audio even with the zero-shot setting, and the proposed adaptation method enhances the person-specific visual information without losing the lip-sync generality." + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22848" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "type": "text", + "content": "Limitation and Future Works. Since learning audiovisual representation requires a large number of different identities (e.g., Voxceleb2 [7]), extending our method to higher resolution " + }, + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "type": "inline_equation", + "content": "(512 \\times 512 \\uparrow)" + }, + { + "bbox": [ + 46, + 72, + 287, + 178 + ], + "type": "text", + "content": " is very challenging. We consider it our future work to develop an effective encoding system for a limited number of identities using a style-based generator [18, 19, 17]. Designing an effective reference encoder to improve lip identity preservation in a zero-shot setting can be another future work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 179, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 179, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 179, + 287, + 239 + ], + "type": "text", + "content": "Ethical Considerations. Since our method can generate a video of a specific person talking specific words only with a few seconds of video source, it has the potential for misuse. As discussed in [6], attaching visual watermarks on the generated videos can be a solution to it." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 250, + 106, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 106, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 106, + 261 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 269, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 269, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 287, + 314 + ], + "type": "text", + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 315, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 288, + 357 + ], + "type": "text", + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8296-8305, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "type": "text", + "content": "[3] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. arXiv preprint arXiv:2201.13433, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, pages 187-194, 1999." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 502 + ], + "type": "text", + "content": "[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 503, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 503, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 503, + 287, + 557 + ], + "type": "text", + "content": "[6] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 558, + 286, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 286, + 580 + ], + "type": "text", + "content": "[7] J. S. Chung, A. Nagrani, and A. Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 581, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 581, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 53, + 581, + 287, + 613 + ], + "type": "text", + "content": "[8] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision, pages 251-263, 2016." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 287, + 658 + ], + "type": "text", + "content": "[9] Dipanjan Das, S. Biswas, Sanjana Sinha, and Brojeshwar Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 712 + ], + "type": "text", + "content": "[10] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2019." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[11] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 2014." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 176, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 176, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 308, + 176, + 545, + 219 + ], + "type": "text", + "content": "[13] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 222, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 222, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 222, + 545, + 277 + ], + "type": "text", + "content": "[14] Jingwen He, Wu Shi, Kai Chen, Lean Fu, and Chao Dong. Gcfsr: a generative and controllable face super resolution method without facial and gan priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1898, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 279, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 279, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 279, + 545, + 312 + ], + "type": "text", + "content": "[15] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 314, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 314, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 314, + 545, + 358 + ], + "type": "text", + "content": "[16] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 360, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 404 + ], + "type": "text", + "content": "[17] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 406, + 545, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 406, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 308, + 406, + 545, + 450 + ], + "type": "text", + "content": "[18] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 452, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 452, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 452, + 545, + 507 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 509, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 509, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 308, + 509, + 545, + 541 + ], + "type": "text", + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 544, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 599 + ], + "type": "text", + "content": "[21] Prajwal KR, Rudrabha Mukhopadhyay, Jerin Philip, Abhishek Jha, Vinay Namboodiri, and CV Jawahar. Towards automatic face-to-face translation. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1428-1436, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 600, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 656 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[23] Camillo Lugasre, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22849" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[24] Dongchan Min, Minyoung Song, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. arXiv preprint arXiv:2208.10922, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 151 + ], + "type": "text", + "content": "[25] Yotam Nitzan, Kfir Aberman, Qiurui He, Orly Liba, Michal Yarom, Yossi Gandelsman, Inbar Mosseri, Yael Pritch, and Daniel Cohen-Or. Mystyle: A personalized generative prior. arXiv preprint arXiv:2203.17272, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "type": "text", + "content": "[26] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "type": "text", + "content": "[27] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In 36th AAAI Conference on Artificial Intelligence (AAAI 22). Association for the Advancement of Artificial Intelligence, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 319 + ], + "type": "text", + "content": "[28] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 320, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 287, + 374 + ], + "type": "text", + "content": "[29] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 376, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 287, + 419 + ], + "type": "text", + "content": "[30] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 421, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 421, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 421, + 287, + 464 + ], + "type": "text", + "content": "[31] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9243-9252, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 466, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 466, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 466, + 287, + 510 + ], + "type": "text", + "content": "[32] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1532-1540, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 511, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 511, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 511, + 287, + 555 + ], + "type": "text", + "content": "[33] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 556, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 287, + 600 + ], + "type": "text", + "content": "[34] Justus Thies, Mohamed A. Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. In European Conference on Computer Vision, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 287, + 644 + ], + "type": "text", + "content": "[35] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. arXiv preprint arXiv:2104.15069, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 689 + ], + "type": "text", + "content": "[36] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "type": "text", + "content": "[37] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 553 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "editing of real videos. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[38] Xintao Wang, Yu Li, Honglun Zhang, and Ying Shan. Towards real-world blind face restoration with generative facial prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9168-9178, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[39] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 229 + ], + "type": "text", + "content": "[40] Weihao Xia, Yujiu Yang, and Jing-Hao Xue. Gan inversion for consistent video interpolation and manipulation. arXiv preprint arXiv:2208.11197, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "type": "text", + "content": "[41] Zhiliang Xu, Hang Zhou, Zhibin Hong, Ziwei Liu, Jiaming Liu, Zhizhi Guo, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Styleswap: Style-based generator empowers robust face swapping. In European Conference on Computer Vision, pages 661-677, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "text", + "content": "[42] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 672-681, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 385 + ], + "type": "text", + "content": "[43] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "type": "text", + "content": "[44] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 443, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 545, + 508 + ], + "type": "text", + "content": "[45] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4176–4186, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 510, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 545, + 553 + ], + "type": "text", + "content": "[46] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22850" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/39699168-4600-4fd3-a205-b0cf85e92a6d_content_list.json b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/39699168-4600-4fd3-a205-b0cf85e92a6d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..335b35298898b48a997929c5f56929902b3c574b --- /dev/null +++ b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/39699168-4600-4fd3-a205-b0cf85e92a6d_content_list.json @@ -0,0 +1,1479 @@ +[ + { + "type": "text", + "text": "StylerDALLE: Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model", + "text_level": 1, + "bbox": [ + 161, + 168, + 806, + 206 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zipeng Xu1 Enver Sangineto2 Nicu Sebe1 \n1University of Trento, Italy 2University of Modena and Reggio Emilia, Italy {zipeng.xu, niculae.sebe}@unitn.it enver.sangineto@unimore.it", + "bbox": [ + 181, + 231, + 784, + 281 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/622024d084f6d673456c683a0a951f442c00dc1edfdb78694ae4887d20181334.jpg", + "image_caption": [ + "Figure 1: With a large-scale pretrained vector-quantized tokenizer (e.g., the dVAE of DALL-E) and CLIP, StylerDALLE can transfer various types of styles (indicated on the top), from abstract art styles to specific artist styles and more." + ], + "image_footnote": [], + "bbox": [ + 97, + 313, + 873, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 238, + 769, + 315, + 783 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the progress made in the style transfer task, most previous work focuses on transferring only relatively simple features like color or texture, while missing more abstract concepts such as overall art expression or painter", + "bbox": [ + 82, + 804, + 470, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "specific traits. However, these abstract semantics can be captured by models like DALL-E or CLIP, which have been trained using huge datasets of images and textual documents. In this paper, we propose StylerDALLE, a style transfer method that exploits both of these models and uses natural language to describe abstract art styles.", + "bbox": [ + 497, + 770, + 884, + 852 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 95, + 8, + 169, + 45 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 231, + 7, + 815, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7601", + "bbox": [ + 468, + 883, + 500, + 894 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Specifically, we formulate the language-guided style transfer task as a non-autoregressive token sequence translation, i.e., from input content image to output stylized image, in the discrete latent space of a large-scale pretrained vector-quantized tokenizer, e.g., the discrete variational auto-encoder (dVAE) of DALL-E. To incorporate style information, we propose a Reinforcement Learning strategy with CLIP-based language supervision that ensures stylization and content preservation simultaneously. Experimental results demonstrate the superiority of our method, which can effectively transfer art styles using language instructions at different granularities. Code is available at https://github.com/zipengxuc/StylerDALLE.", + "bbox": [ + 85, + 131, + 468, + 307 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 85, + 334, + 213, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the last few years, a lot of work has focused on the style transfer task using a reference image as a representative of the target style, where the goal is to transfer the style of the reference to a content image [14, 16, 28, 3, 22, 8, 9, 40]. Recent improvements in this field include: reducing the artifacts [2, 25, 5, 40], modeling the style-content relationship [28, 43, 24], increasing the generation diversity [21, 37, 46] and many others. However, art styles are usually abstract concepts, e.g., pop art, fauvism, and the style of Van Gogh. To transfer these abstract art styles to a content image, the low-level features (e.g., textures and colors), which are commonly extracted from a single reference image, are not enough. A possible solution is to collect a set of reference images that can be used, e.g., to train a Generative Adversarial Network (GAN) for artist-specific style transfer [36, 47, 19]. The disadvantage of this set-based representation is the effort required to collect sufficiently large style-specific data for training.", + "bbox": [ + 85, + 357, + 468, + 601 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, large-scale image generative models [32, 31, 35, 45] have shown their power to generate high-quality images of various types, e.g., realistic photo, cartoon, ukiyo-e print, or painting of a specific artist. Moreover, CLIP [30], which is trained with 400 million text-image pairs, learns good joint knowledge between language and vision. Can we leverage the power of these large-scale models for style transfer? In this paper, we propose StylerDALLE, a language-guided style transfer method that uses both the vector-quantized tokenizer of large-scale image generative methods and CLIP. There are three advantages of our method. Firstly, as compared to images, language is a more natural and efficient way to describe abstract art styles and enables more flexibility. Language can directly indicate art styles at different levels, e.g., \"Van Gogh style oil painting\" and \"Van Gogh starry night style oil painting\". Secondly, using CLIP and the language description of style saves the effort of collecting style images, as CLIP already learns style-related knowledge and thus can be used to provide", + "bbox": [ + 85, + 602, + 468, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "supervision. Thirdly, using a large-scale pretrained vector-quantized tokenizer potentially enables style transfer results to be closer to real artworks, as the model is trained from an enormous number of real-world images. In concrete, we propose a Non-Autoregressive Transformer (NAT) model [15], which translates tokens of a content image to tokens of a stylized image in the discrete latent space of a pretrained visual tokenizer, and a two-stage training paradigm.", + "bbox": [ + 502, + 131, + 882, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First of all, since both the input content image and output stylized image can be represented by a sequence of tokens via the large-scale pretrained vector-quantized tokenizer, we formulate the language-guided style transfer task as a token sequence translation. Specifically, we design a NAT model that learns to translate a token-based representation of a low-resolution image into a full-resolution representation, where the final token sequence contains appearance details specialized for the target style. The most important advantage of using NAT in image generation, with respect to the more common autoregressive transformer (AT) based generation, is that NAT is much faster than AT at inference time, as it allows a parallel token generation while AT generates in a token-by-token way.", + "bbox": [ + 502, + 244, + 882, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, we propose to use CLIP-based language supervision to ensure stylization and content preservation simultaneously, saving the effort to collect data and design dedicated losses. A similar solution that uses CLIP to do style transfer is CLIPStyler [20]. Since merely maximizing the CLIP similarity with respect to a textual style description is not enough for a style transfer task, which should also preserve the content of the source image, CLIPstyler introduces a hybrid of losses, including a content loss measured in an external pre-trained VGG network. Conversely, we propose an alternative direction, which avoids the need to tune the coefficients of different loss functions so as to keep the balance between style and content. Specifically, we introduce a two-stage training paradigm: 1) a self-supervised pretraining stage, where the model learns to add semantically coherent image details from a low-resolution image to a high-resolution image; and 2) a style-specific fine-tuning stage, where the model learns to incorporate style into the high-resolution image. Since the fine-tuning phase is built on top of the first stage, the translator is able to keep the semantic consistency with respect to the input image as learned during pretraining. Moreover, we create a textual prompt by concatenating both the style and the textual description of the image content (i.e., its caption). This way, the prompt simultaneously models both the target appearance (i.e., the style description) and the image content which should be preserved by the translation process. Finally, since our translator's output is discrete and there is no ground-truth tokens for a stylized image, we introduce a Reinforcement Learning (RL) approach to fine-tune the translator using a reward based on the CLIP similarity be", + "bbox": [ + 502, + 439, + 882, + 859 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7602", + "bbox": [ + 468, + 884, + 500, + 894 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tween the stylized image and textual prompt, enabling the model to explore the answers in the latent space of a pretrained vector-quantized tokenizer.", + "bbox": [ + 84, + 131, + 468, + 173 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We call our network StylerDALLE and we show that it can generate stylized images driven by different types of language guidance. Compared with previous language-guided and reference image-based transfer methods, our generated images are less inclined to produce artifacts or semantic errors. Moreover, they can capture abstract concepts related to the target style (e.g., the typical brushstrokes of the artist) besides low-level features like texture and colors. We illustrate the effectiveness of our method through qualitative results, quantitative results, and a user study.", + "bbox": [ + 84, + 174, + 472, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To conclude, our main contributions are:", + "bbox": [ + 104, + 311, + 369, + 324 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose StylerDALLE, a language-guided style transfer method that manipulates the discrete latent space of a pretrained vector-quantized tokenizer using a token sequence translation approach.", + "- We propose a non-autoregressive translation network that translates a low-resolution content image into a full-resolution image with style-specific details.", + "- We propose a two-stage training procedure, including an RL strategy to ensure stylization and content preservation using CLIP-based language supervision.", + "- Experimental results show that StylerDALLE can effectively transfer abstract style concepts, going beyond simple texture and color features while simultaneously preserving the semantic content of the translated scene." + ], + "bbox": [ + 102, + 335, + 470, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 85, + 577, + 223, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reference Image-Based Style Transfer. Gatys et al. [14] propose a neural style transfer method in which a pretrained CNN is used to extract content and style information from images, and to transfer the latter from an image to another. Following this pioneering work, a lot of interests have been attracted, with different methods focusing on different aspects of the topic, such as, e.g., diversified style transfer [37, 41], or attention mechanisms to fuse style and content [44, 28, 24]. A specific line of work focuses on artistic style transfer. For instance, Chen et al. [3] propose to use internal-external learning and contrastive learning with GANs to bridge the gap between human artworks and AI-created artworks. Wang et al. [40] introduce an aesthetic discriminator trained with a large corpus of human-created artworks. Other works train GANs using an artist-specific collection of images [36, 19, 4]. In contrast, we use the generic visual-language semantics embedded in the large-scale pretrained vector-quantized tokenizer and CLIP to avoid collecting style or artist-specific datasets.", + "bbox": [ + 84, + 601, + 472, + 860 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language-Guided Style Transfer. Very recently a few works have proposed transferring methods conditioned on a textual description of the style. For instance, Fu et al. [12] use contrastive learning to train a GAN for artistic style transfer, but they adopt descriptive language instructions rather than more abstract style concepts. Gal et al. [13] use the CLIP space for a domain adaptation of a pre-trained StyleGAN [17]. The method closest to our approach is CLIPStyler [20], where a patch-wise CLIP loss is used to train a U-Net [34]. However, to condition the style change while preserving the image content, CLIPStyler uses hybrid losses and a rejection threshold, introducing many hyperparameters. In our method, we only use CLIP-based language supervision to ensure the style and content, saving the efforts of designing losses and tuning hyperparameters.", + "bbox": [ + 497, + 131, + 885, + 337 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large-scale Text-to-Image Generation Models. Recently, text-to-image models trained with large or very large scale datasets [32, 31, 35, 45, 1] have attracted tremendous attention because of their excellent performance in generating high-quality images starting from a textual query. Inspired by their strong ability to synthesize various types of images, we study how to use large-scale text-to-image models for style transfer. Specifically, we focus on transformer-based methods, which are potentially less time-consuming as compared to diffusion model-based methods [26]. In addition, we propose to use a non-autoregressive transformer that can generate tokens in parallel.", + "bbox": [ + 497, + 355, + 885, + 519 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 499, + 532, + 626, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vector-Quantzied Image Tokenizer. Transformer-based text-to-image generative models [32, 10, 11, 45, 1] rely on a vector-quantized image tokenizer to produce a discrete representation of the images, e.g., DALL-E [32] has a dVAE [33] and PARTI [45] has a ViT-VQGAN. Despite the differences in implementing the tokenizers, the effects are the same. In more detail, an image $I$ is transformed into a $k \\times k$ grid of tokens $X = \\{x_{i,j}\\}_{i,j=1,\\dots,k}$ , using an encoder $E(\\cdot)$ . Each token $x_{i,j} \\in X = E(I)$ is an index of a codebook of embeddings ( $C = \\{e_1, \\ldots, e_M\\}$ , $1 \\leq x_{i,j} \\leq M$ ), built during the training, and corresponds to a patch in $I$ . A decoder $G(\\cdot)$ takes as input a grid of embeddings and reconstructs the original image: $\\tilde{I} = G(\\{e_{x_{i,j}}\\}_{i,j=1,\\dots,k})$ . Training in the transformer-based text-to-image generative models consists of two stages. The first stage is dedicated to training the image tokenizer, while in the second phase, a transformer is used to learn a prior distribution over the text and the image tokens. In StylerDALLE (Sec. 4) we only use the pretrained image tokenizer.", + "bbox": [ + 497, + 555, + 885, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Non-Autoregressive Transformer (NAT). Gu et al. [15] propose a NAT for natural language translation, which con", + "bbox": [ + 497, + 832, + 884, + 860 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7603", + "bbox": [ + 468, + 883, + 500, + 895 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c5af19da210c221852064406a7a8df89901eb20e8449b8810e17d021d9206e3a.jpg", + "image_caption": [ + "Figure 2: A schematic illustration of StylerDALLE. We propose a NAT to translate tokens of a half-resolution content image to tokens of a full-resolution stylized image. StylerDALLE consists of three parts: 1) image tokenization preprocessing, to obtain the discrete visual tokens from a pretrained image tokenizer; 2) self-supervised pretaining, to train the NAT to predict full-resolution content images from half-resolution content images; and 3) style-specific fine-tuning, to let the model adding style details in the up-scaling prediction via Reinforcement Learning with language-prompted CLIP guidance. Specifically, in our NAT, the decoder input is duplicated from the tokens of the half-resolution image, following the natural positional correspondence (as highlighted by orange rectangles in the left part) with the full-resolution image." + ], + "image_footnote": [], + "bbox": [ + 97, + 127, + 579, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/436d81afcc601f3abae24e8c55599a0127c0a64965102070dbdba812dbaf3a1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 128, + 875, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sits of an encoder and a decoder. The encoder takes as input a source sentence $X = \\{x_{1},\\dots,x_{N^{\\prime}}\\}$ of $N^{\\prime}$ tokens and outputs a distribution over possible output sentences $Y = \\{y_{1},\\dots,y_{N}\\}$ , where $Y$ is the translation of $X$ in the target language and, usually, $N\\neq N^{\\prime}$ . The main novelty of NAT with respect to AT is that, during training, NAT uses a conditional independent factorization for the target sentence and the following log-likelihood:", + "bbox": [ + 82, + 514, + 473, + 624 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = \\sum_ {n = 1} ^ {N} \\log p \\left(y _ {n} \\mid x _ {1: N ^ {\\prime}}; \\theta\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 627, + 468, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which differs from the common AT factorization in which the prediction of the $n$ -th token $(y_{n})$ depends on the previously predicted tokens $(p(y_{n}|y_{0:n-1}, x_{1:N'}; \\theta))$ . This conditional independence assumption makes it possible a parallel generation of $Y$ at inference time, largely accelerating the translation time with respect to AT models. Importantly, to make parallel generation possible, the encoder input $(X)$ is provided as input to the decoder as well, and individual tokens $(x_{n} \\in X)$ can be copied zero or more times, with the number of times each input is copied depending on a specific predicted \"fertility\" value. As we will see in Sec. 4.1, we do not need to predict fertilities because, in our case, the cardinality of the input copies is fixed and determined by the up-scaling task we use for the image translation.", + "bbox": [ + 82, + 669, + 472, + 860 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 499, + 512, + 589, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The language-guided style transfer task can be described as follows. Given an image $I$ , we want to generate a new image $I^s$ which preserves the semantic content of $I$ but changes its appearance according to a textual style description $t_s$ (e.g., \"cartoon\"). In StylerDALLE, we formulate this task as a visual-token based translation, from tokens of a content image to tokens of a stylized image. Specifically, given a content image $I$ we first downsample $I$ to get a half-resolution image $I'$ . Then, $I'$ is fed to the tokenization encoder (Sec. 3) which extracts a discrete grid of $k \\times k$ source tokens $X' = E(I')$ . $X'$ can now be \"translated\" into a target (discrete) representation $\\hat{Y}$ , where $\\hat{Y}$ is a grid at the original resolution $(2k \\times 2k)$ and $\\hat{Y} = f(X')$ , being $f(\\cdot)$ the translation function. Finally, $\\hat{Y}$ is fed into the tokenization decoder obtaining the stylized image $I^s = G(\\hat{Y})$ . In the following subsections, we describe the architecture of $f(\\cdot)$ and the way in which it is trained.", + "bbox": [ + 497, + 538, + 885, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Architecture and self-supervised pre-training", + "text_level": 1, + "bbox": [ + 497, + 782, + 877, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For our translation network $f(\\cdot)$ we use a NAT architecture [15] (Sec. 3) which we train from scratch using a self-supervised learning pretext task consisting in predicting the image at full resolution. Specifically, given the downsam-", + "bbox": [ + 497, + 804, + 885, + 860 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7604", + "bbox": [ + 468, + 883, + 502, + 895 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pled image $I'$ and its corresponding grid of tokens $X' = E(I')$ , we use the indexes in $X'$ to extract the corresponding embeddings from $C$ (Sec. 3). For each $x_{i,j} \\in X'$ , let $\\pmb{e}_{x_{i,j}}$ be the corresponding embedding in $C$ and let $N' = k^2$ . The so obtained set of embeddings is flattened into a sequence, and, for each element $\\pmb{e}_n$ of the sequence ( $1 \\leq n \\leq N'$ ), we add an absolute positional embedding [38] $\\pmb{p}_n$ , where $\\pmb{p}_n$ has the same dimension as $\\pmb{e}_n$ : $\\pmb{v}_n = \\pmb{e}_n + \\pmb{p}_n$ . The final sequence $V' = \\{\\pmb{v}_1, \\dots, \\pmb{v}_n, \\dots, \\pmb{v}_{N'}\\}$ is input to the encoder of $f(\\cdot)$ . Note that an alternative solution is to directly fed $f(\\cdot)$ with (a flattened version of) $X'$ and let $f(\\cdot)$ learn its own initial token embedding. However, using the embeddings in $C$ has the advantage of exploiting the image representation of the pre-trained image tokenizer. Moreover, from the original image $I$ we extract the ground truth $X = E(I)$ , which is flattened in a sequence of $N = 4k^2$ tokens.", + "bbox": [ + 84, + 131, + 472, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, following [15], we build a second sequence of input embeddings $V$ , with cardinality $N$ , which is fed to the decoder of $f(\\cdot)$ . As mentioned in Sec. 3, differently from NAT, we do not predict fertility values. Instead, we get the input of the decoder by simply replicating each element $x_{i,j} \\in X'$ according to the positional correspondences between the low-resolution image and high-resolution image (as in Fig. 2). The rationale behind this choice is that $f(\\cdot)$ is trained to predict the full-resolution image, and each encoder input $(\\pmb{e}_{x_{i,j}})$ corresponds to a patch in the subsampled image $I'$ and to 4 patches in the full-resolution image $I$ . Thus, initializing the decoder with 4 replicas of each source-image patch initial embedding provides a coarse-grained signal for the upsampling task. Similar to before, the embeddings extracted from $C$ are then flattened and added with a new positional encoding.", + "bbox": [ + 84, + 350, + 472, + 567 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both the encoder and the decoder have self-attention layers and no causal masking is used. However, following [15], in the decoder, we mask out each query position $(n)$ only from attending to itself. Using $V^{\\prime}$ and $V$ , $f(\\cdot)$ generates $N$ parallel posterior distributions over the visual vocabulary $(\\{1,\\dots,M\\})$ : $P = f_{\\theta}(V',V)$ , where $P$ is a $N\\times M$ matrix, $P_{n}\\in [0,1]^{M}$ and $P_{n}[y] = p_{\\theta}(Y_{n} = y|V^{\\prime},V)$ . Using $Y = \\{y_{1},\\ldots ,y_{n},\\ldots ,y_{N}\\}$ , $f(\\cdot)$ is trained to maximize:", + "bbox": [ + 84, + 567, + 470, + 677 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r e - t r a i n}} (\\theta) = \\sum_ {n = 1} ^ {N} \\log P _ {n} [ y _ {n} ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 688, + 468, + 725 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This pre-training stage is independent of the target style and it can be shared over different styles. After this stage, $f(\\cdot)$ is able to generate realistic low-level details (which are missing in $I'$ ). In the next, we describe how a specific style is incorporated in $f(\\cdot)$ using a fine-tuning phase.", + "bbox": [ + 84, + 732, + 468, + 802 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Style-specific fine-tuning", + "text_level": 1, + "bbox": [ + 84, + 810, + 307, + 826 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a style description provided with a textual sentence $t_s$ , the goal is to fine-tune the pre-trained translator", + "bbox": [ + 84, + 832, + 468, + 860 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\\(f(\\cdot)\\) (Sec. 4.1) to make it generate image details in the style of \\(t_s\\). We fine-tune only the decoder of \\(f(\\cdot)\\), keeping frozen the encoder. To ensure both stylization and content preservation, we design a prompt that consists of two parts, i.e., a style description \\(t_s\\) and an image caption \\(t_a\\), which describe the image content and can be obtained from a generic image captioning dataset. For instance, given \\(t_a = \"A man's hand is adjusting his black tie\" and \\(t_s = \"cartoon\"\\), we obtain \\(t_q = \"a cartoon of a man's hand is adjusting his black tie.\\). On the other hand, in order to represent the image generated by \\(f(\\cdot)\\), we first need to sample the distributions in \\(P\\) (Sec. 4.1), and we do so using multinomial sampling:", + "bbox": [ + 497, + 131, + 885, + 297 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} _ {n} = \\text {S a m p l i n g} \\left(P _ {n} [ y ]\\right), \\forall n \\in \\{1, \\dots , N \\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 546, + 321, + 882, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sampled sequence $\\hat{Y}$ is reshaped to a $2k\\times 2k$ grid and fed to the image detokenizer to get the final image $I^s = G(\\hat{Y})$ . Finally, using the CLIP visual and textual encoders we compute the cosine similarity on the CLIP space:", + "bbox": [ + 497, + 349, + 885, + 404 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr = \\operatorname {S i m} _ {C L I P} \\left(I ^ {s}, t _ {q}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 615, + 420, + 882, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, directly using Eq. 4 as the fine-tuning objective function is not possible because Eq. 3 is not differentiable. In addition, since there is no ground truth for the tokens of stylized images we use RL to encourage the model to explore the answers in the latent space of a pretrained vector-quantized model. We use the REINFORCE algorithm [42] that updates the parameters of $f_{\\theta}(\\cdot)$ using the CLIP-based reward $r$ , which could keep awarding the model for achieving better-stylized results until the model reaches a limit. This leads to the gradient estimate:", + "bbox": [ + 497, + 445, + 885, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta | _ {d}} \\mathcal {L} _ {\\text {f i n e - t u n e}} (\\theta | _ {d}) = \\sum_ {n = 1} ^ {N} r \\nabla_ {\\theta | _ {d}} \\log P _ {n} [ y _ {n} ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 592, + 882, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta|_d$ indicates the parameters of the decoder only (we found fine-tuning both could lead to content loss). By maximizing Eq. 5 we encourage $f(\\cdot)$ to generate images having both the content and the style of the prompt $t_q$ .", + "bbox": [ + 497, + 640, + 884, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With respect to the whole method, the reason we use down-sampled versions of the content image is that \"style\" is commonly assumed to be involved in the low-level visual details, such as colors, texture, painting strokes, etc. $X^{\\prime}$ , which in our formulation represents $I$ at a lower resolution, presumably keeps most of the content in $I$ discarding some details, this way facilitating the style translation process. Preliminary experiments (in Appendix A.2) in which we fed the encoder with tokenized full-resolution images led to poor results, demonstrating that the different cardinality between the source and the target sequence is an essential component in this translation process.", + "bbox": [ + 497, + 697, + 885, + 860 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7605", + "bbox": [ + 468, + 883, + 500, + 894 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b2252aab2985062c302386dad8df9e7dedf553e18ccb8c6db7e5fed30cdaecf5.jpg", + "image_caption": [ + "Figure 3: Qualitative results of StylerDALLE-1 on various styles: (a) Monet Sunrise, (b) Picasso cubism, (c) Van Gogh blue color, (d) Van Gogh purple color, (e) warm and relaxing, (f) ukiyo-e print, (g) fauvism, (h) pixel art illustration. The style references at the top are for illustration only (not used as input to the model)." + ], + "image_footnote": [], + "bbox": [ + 85, + 130, + 885, + 451 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 84, + 535, + 215, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Training Details", + "text_level": 1, + "bbox": [ + 84, + 560, + 243, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement our method based on two types of pretrained vector-quantized tokenizer: 1) the officially released dVAE of DALL-E 1 [27] and 2) the VQGAN of RuDALLE [39], and consequently obtain two groups of results, noting as StylerDALLE-1 and StylerDALLE-Ru, respectively. We use the two models in our experiment because they are open-sourced while our method is applicable to any large-scale pretrained vector-quantized tokenizer.", + "bbox": [ + 82, + 585, + 468, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train the model, we use the MS-COCO [23] trainset, which contains 83k images of common objects in daily scenes. In the self-supervised pretraining stage, we only use the images while in the style-specific fine-tuning stage we use both images and captions. The StylerDALLE NAT model consists of a 4-layer encoder and an 8-layer decoder while the attention head number is 8 and the hidden dimension is 512. We use Pytorch [29] to implement our method. We train the NAT model for 25 epochs with a learning rate of 1e-4. We use Adam [18] optimizer. In the style-specific fine-tuning stage, we use both the images and the captions. In particular, we utilize all the caption annotations to en", + "bbox": [ + 82, + 696, + 470, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "hance the model robustness, as usually human-beings annotate different captions of a single image. Notably, the caption annotations are only used at the fine-tuning stage. In other words, StylerDALLE does not need to use image caption as input at inference time. We only fine-tune the decoder of the NAT model, and keep the encoder frozen. We use Adam optimizer with a learning rate of 1e-6. For CLIP model, we use the CLIP ViT-B/32 model. For both training stages, the model is trained on a single RTX-A6000 GPU for 24 hours.", + "bbox": [ + 497, + 536, + 884, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 499, + 682, + 697, + 697 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the following, we present qualitative, quantitative, and user study results of our method and comparative methods, as well as comparisons with reference image-based methods. The additional implementation details, ablation study, inference time comparison, and additional experimental results are shown in Appendix A.1, A.2, A.3, and A.5, respectively.", + "bbox": [ + 497, + 703, + 884, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. In Fig. 1, Fig. 3 and Fig. 4, we show that our method can effectively transfer various types of styles, i.e., a) abstract art styles, e.g., \"fauvism\" and", + "bbox": [ + 497, + 818, + 884, + 860 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7606", + "bbox": [ + 468, + 883, + 502, + 894 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/884181b992be00983c15537b69b7135cf17d675bc107ea1c870f94aed16dcd14.jpg", + "image_caption": [ + "Figure 4: Qualitative results of StylerDALLE-Ru on various styles: (a) Monet, (b) Monet Paris, (c) Monet Venice, (d) Van Gogh, (e) Van Gogh Irises, (f) Van Gogh Almond, (g) Van Gogh Starry Night, (h) Van Gogh Sunflowers. The style references at the top are for illustration only (not used as input to the model)." + ], + "image_footnote": [], + "bbox": [ + 85, + 131, + 885, + 456 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c8db4ab767de157313dd3a3e3cfa48893f03929022444ee0412fab53749b0d58.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetStyle MethodFauvismMonetMonet SunriseMonetSunsetMonetParisVan Gogh IrisesVan GoghAlmond Blossoms
CoCoCLIPstyler-Fast26.1127.4624.7826.7027.6026.2327.05
Ours30.6127.7228.8130.6829.6328.1431.64
AFHQCLIPstyler-Fast26.5226.0823.2725.0525.4223.2824.31
Ours29.9626.5526.5829.1126.1028.0431.92
ImageNet-100CLIPstyler-Fast26.6327.1624.3326.1227.2826.3226.23
Ours30.2227.5328.2930.2827.6728.2631.60
", + "bbox": [ + 85, + 525, + 890, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Quantitative comparisons with CLIPstyler [20] on different types of styles and datasets.", + "bbox": [ + 171, + 669, + 794, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"pop art\"; b) artist-specific styles, e.g., \"Monet\" and \"Van Gogh\"; c) artist-specific styles with additional descriptions, e.g., \"Monet Paris\" and \"Van Gogh Sunflowers\"; d) artistic painting types, e.g., \"pixel art illustration\"; and e) emotional effects, e.g., \"warm and relaxing\". According to the qualitative results, we draw the following conclusions: 1) StylerDALLE can transfer abstract style concepts which go beyond the texture and color features and are similar to the typical trait of the artist/artistic target style; 2) each style corresponds to generated images that are different from those of other styles; 3) the image content is well preserved;", + "bbox": [ + 82, + 707, + 470, + 857 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and 4) StylerDALLE can be applied to open-domain content images (i.e., the image content can contain animals, human beings, daily objects, buildings, etc.).", + "bbox": [ + 497, + 707, + 884, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further, we compare StylerDALLE with the recent language-guided style transfer method CLIPStyler [20]. CLIPstyler proposes two methods: 1) CLIPStyler-Optimization, which optimizes a style transfer network for each content image, thus being time-consuming; and 2) CLIPStyler-Fast, which is the most comparable method to ours as it trains a network for each style, and then it can be used with any content images. As shown in Fig. 5,", + "bbox": [ + 497, + 750, + 885, + 860 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7607", + "bbox": [ + 468, + 883, + 500, + 894 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1e0263c65c1e8f6b7567708a4d3d752fef37a134ab05752768de1a0e77f6f6b.jpg", + "image_caption": [ + "Figure 5: Comparisons with language-guided methods. The styles are (a) fauxivism, (b) Monet, (c) Monet Impression Sunrise, (d) oil painting, (e) watercolor painting, (f) Van Gogh, (g) Van Gogh blue color. The style references at the top are for illustration only (not used as input to the model)." + ], + "image_footnote": [], + "bbox": [ + 90, + 133, + 885, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CLIPStyler-Optimization generates diverse stylized results but it suffers from inharmonious artifacts. For instance, in the column \"Monet\", there are many plants on the train. In addition, CLIPStyler-Optimization has the problems of writing the style text in the results, as in the results of \"fauvism\" and so on. On the other hand, the images generated by CLIPStyler-Fast do not contain artifacts but there is less variation among different styles. Importantly, it is hard to recognize the typical trait of each artistic style, and the main differences among the styles are the colors. In contrast, the results of StylerDALLE are much closer to the artworks of the specific artistic style, they show distinct differences among different styles, and they have no artifact issues.", + "bbox": [ + 82, + 502, + 472, + 682 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Results. To do quantitative analysis, we use the CLIP similarity score, formalized as score $= \\operatorname{Sim}_{\\text{CLIP}}(I, t_s)$ , which is computed between generated stylized images and textual description of the target style. In Tab. 1, we present the results of StylerDALLE-1 and CLIPstyler-Fast. Since both methods are applicable for arbitrary content images, we use the MS-COCO val-set, AFHQ val-set [6], and ImageNet-100 val-set [7] for evaluation. According to the results on multiple datasets, although the way we use CLIP is to compute rewards in Reinforcement Learning, instead of directly using CLIP-scores to optimize the network as CLIPstyler, we achieve comparable", + "bbox": [ + 82, + 696, + 472, + 860 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eb187c5d3c45803654fc7c732ffd43c68f64d13b571ed961c9260c64463903e6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(%)CLIPstyler-FastStyler DALLE-1Styler DALLE-Ru
Preference15.9015.3368.76
", + "bbox": [ + 509, + 502, + 877, + 557 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: Preference scores of user study.", + "bbox": [ + 559, + 565, + 823, + 581 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and even better quantitative results, indicating the effectiveness of our method.", + "bbox": [ + 497, + 607, + 884, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. We conduct a user study to see human opinions towards the stylized images coming from different methods: 1) CLIPstyler-Fast, 2) StylerDALLE-1, and 3) StylerDALLE-Ru. In specific, we collect opinions from 35 human subjects with a 30-question questionnaire. In each question, we ask them to select one stylized result that is closest to a target style. As in Tab. 2, among the three methods, StylerDALLE-Ru achieves the highest preference score of $68.76\\%$ , indicating the superiority of our method. We also find that humans prefer StylerDALLE-Ru much more than StylerDALLE-1. We infer this could be because the results from StylerDALLE-1 are blurry, as they are based on the dVAE of DALLE-1, and humans dislike blurry images. More user study details are given in A.4.", + "bbox": [ + 497, + 655, + 885, + 847 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7608", + "bbox": [ + 468, + 883, + 502, + 895 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b0b54a2c82d79c28f263e9f39ece067161b0b3b0925f56875a2d1ef1d306fa1.jpg", + "image_caption": [ + "ref1" + ], + "image_footnote": [], + "bbox": [ + 90, + 130, + 206, + 219 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e0feb0de3e0c2c5a9fc9bb933872ecdc0322323bc03da03a88d3dec12315f836.jpg", + "image_caption": [ + "ref2" + ], + "image_footnote": [], + "bbox": [ + 90, + 233, + 205, + 313 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/11eda235384a8422a2baa37953a7feeb2b9d6d084033a4f4a954c2e52e8fcff7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 208, + 130, + 302, + 197 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0cfb287d26fbaa9be473d4cd9ef793f0dd192c2fd6a7000d64dc693d8df35ba4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 208, + 199, + 302, + 265 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/994091be88005fc1d17d99a40e613dc5fd853ac1d06a9e6e68e7895d72cbbece.jpg", + "image_caption": [ + "Content" + ], + "image_footnote": [], + "bbox": [ + 208, + 268, + 302, + 334 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e6ca4c8b9c8e6178a471eae45b3e40457c4a648a387228f14f85b4942ecf3526.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 130, + 401, + 197 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/171073ceedee604c21cb9354618de42ab77aa38e7773e7725cd8ee1da469a53f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 199, + 401, + 265 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fe810f9f01a7f65ebf0e46ab7aba97d6fb463269591315c322d43b7268a45d71.jpg", + "image_caption": [ + "AesUST", + "w/ ref1" + ], + "image_footnote": [], + "bbox": [ + 307, + 268, + 401, + 334 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/083eaa866ced7f12ed3df653f4e1d9fb3b3db9b2364f7b42fbcf489249c32576.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 130, + 497, + 197 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f617f7921ea53fcf2c29d1bcb1101fbccb2d150010651d1d1a18494b70548c59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 199, + 497, + 265 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/40f346403ee108140b3b27570439c834d593212a0020279abf18303a7b604ab3.jpg", + "image_caption": [ + "StyTr2", + "w/ ref1", + "Figure 6: Comparisons with reference image-based methods. StylerDALLE is able to transfer more abstract concepts, e.g., specific painting strokes, and is less likely to produce semantic errors." + ], + "image_footnote": [], + "bbox": [ + 403, + 268, + 497, + 334 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/eceb251da3ec631aca88e5f5a832c52bbc04b8669e988b299757c7e6a5bc757a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 499, + 130, + 591, + 197 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/08588be497d109f3585502d97ae42489db351313a1abb3a91618fbe3b41416e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 497, + 199, + 591, + 265 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7cce431f6ace421c029a8ac4c53f3d9a7085851f4fee61bd17a0a5dd95fbdad1.jpg", + "image_caption": [ + "StylerDALLE-Ru", + "w/text" + ], + "image_footnote": [], + "bbox": [ + 497, + 268, + 591, + 334 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/03d4d90c86d7ba54403da3b313d5c42035601f253df708171efbc188ecec8a86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 130, + 690, + 197 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/de1d49cdb238f4e342b06bf82fe71262a93e38298c1c7796b953e3a576e478ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 199, + 690, + 265 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f1f0af7d616df3041b61a491d6e96740f43aa8c0bbe9eec698d07fdf9cd9a4ac.jpg", + "image_caption": [ + "AesUST", + "w/ ref2" + ], + "image_footnote": [], + "bbox": [ + 596, + 268, + 690, + 334 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b2193636062281c30a974fb3cedccc7edb41700c1925a2901ebedbd724c8dff6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 130, + 786, + 197 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ada740624fc6aa034bcccc2b8e1b0292e5e93adc3bd7952c551978d8548ce841.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 199, + 786, + 265 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/93bdab9d2e0f0328ddadd12b72d3d78a2bef774be9289d852e884f7039892177.jpg", + "image_caption": [ + "StyTr2", + "w/ ref2" + ], + "image_footnote": [], + "bbox": [ + 692, + 268, + 786, + 334 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/28b4a7e5e88ec6621d5e9b520d5b4dbcfc00fef1ba8ff75e5e6b82c9ed600638.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 786, + 130, + 880, + 197 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/75d20ebfc8f91a3c0c2e52ed5b538bd47b3b4a69fd2d391e7c18d0fdf24a7b29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 786, + 199, + 880, + 265 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9711048784a74d21cd2af45b0fcb449172b1d3db1c6cbefcb7216463a5c56195.jpg", + "image_caption": [ + "StylerDALLE-Ru", + "w/text" + ], + "image_footnote": [], + "bbox": [ + 786, + 268, + 880, + 334 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Comparisons with Reference Image-Based Methods. We compare StylerDALLE with state-of-the-art reference image-based methods: 1) AesUST [40], an arbitrary style transfer method which enhances the aesthetic reality using a GAN trained with a collection of artworks; and 2) StyTr2 [9], an arbitrary style transfer method which uses a transformer to eliminate the biased content representation issues of CNN-based methods. To make the comparison feasible, we show the results of AesUST and StyTr2 using two Van Gogh paintings as the reference images, and the results of StylerDALLE trained using the corresponding language description.", + "bbox": [ + 82, + 431, + 468, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As shown in Fig. 6, the results of StylerDALLE are distinctive from the results of reference image-based methods. In concrete, the results of both AesUST and StyleTr2 are mostly affected by the colors and the textures of the reference images, and to some extent in an unnatural way. For instance, on the bottom line, in the \"Van Gogh Irises\" stylized results of AesUST and StyleTr2, the textures of the irises get transferred to the orange cup. In transferring the style of \"Van Gogh Starry Night\", the objects in the results of AesUST and StyleTr2 are mostly in the same dark-sky blue as the reference image, making the scenes a bit in-realistic. By contrast, the colors and textures used in the reference style are well transferred while being appropriately applied to the contents, without changing the original semantics. Moreover, StylerDALLE achieves to transfer higher-level style features, e.g., the brushstroke, rather than merely colors and textures, leading to the results of a similar style to the target one. For example, the strokes in Starry Night are sharper as compared to the ones in Irises, and the differences are also", + "bbox": [ + 82, + 601, + 472, + 859 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "presented in the stylized results of StylerDALLE.", + "bbox": [ + 499, + 431, + 831, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 499, + 455, + 616, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present StylerDALLE, a language-guided style transfer method that leverages the power of the large-scale pretrained vector-quantized image tokenizer and CLIP. Specifically, inspired by natural language translation, we propose a non-autoregressive sequence translation approach to manipulate the discrete visual tokens, from the content image to the stylized image. We use Reinforcement Learning to include CLIP-based language supervision on the style and content. Differently from previous work, StylerDALLE can transfer abstract style concepts that are implicitly represented in the pretrained image tokenizer and CLIP and which cannot be easily obtained using reference images. Moreover, using the large-scale pretrained latent space as the basic image representation makes it possible to reduce the artifacts and the semantic incoherence better than the previous work that operates at the pixel level.", + "bbox": [ + 497, + 476, + 884, + 695 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by the MUR PNRR project FAIR (PE00000013) funded by the NextGenerationEU and by the PRIN project CREATIVE (Prot. 2020ZSL9F9).", + "bbox": [ + 497, + 709, + 882, + 764 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 774, + 594, + 788 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[1] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 3, 12", + "bbox": [ + 507, + 796, + 884, + 859 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7609", + "bbox": [ + 468, + 883, + 500, + 894 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[2] Haibo Chen, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, Dongming Lu, et al. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems, 34:26561-26573, 2021. 2", + "[3] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 872-881, June 2021. 2, 3", + "[4] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dualist: Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 872-881, 2021. 3", + "[5] Jiaxin Cheng, Ayush Jaiswal, Yue Wu, Pradeep Natarajan, and Prem Natarajan. Style-aware normalized loss for improving arbitrary style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-143, 2021. 2", + "[6] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 8", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 8", + "[8] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11326-11336, June 2022. 2", + "[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11326-11336, 2022. 2, 9", + "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 3", + "[11] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. Advances in Neural Information Processing Systems, 35, 2022. 3", + "[12] Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-driven artistic style transfer. In European Conference on Computer Vision (ECCV), 2022. 3", + "[13] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3", + "[14] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In" + ], + "bbox": [ + 87, + 133, + 470, + 859 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 3", + "[15] Jiatao Gu, James Bradbury, Caiming Xiong, Victor O. K. Li, and Richard Socher. Non-autoregressive neural machine translation. In ICLR, 2018. 2, 3, 4, 5", + "[16] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2", + "[17] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[19] Dmytro Kotovenko, Artsiom Sanakoyeu, Sabine Lang, and Bjorn Ommer. Content and style disentanglement for artistic style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4422-4431, 2019. 2, 3", + "[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 2, 3, 7", + "[21] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Diversified texture synthesis with feed-forward networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3920-3928, 2017. 2", + "[22] Tianwei Lin, Zhuoqi Ma, Fu Li, Dongliang He, Xin Li, Errui Ding, Nannan Wang, Jie Li, and Xinbo Gao. Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5141-5150, June 2021. 2", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6", + "[24] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 2, 3", + "[25] Yahui Liu, Enver Sangineto, Yajing Chen, Linchao Bao, Haoxian Zhang, Nicu Sebe, Bruno Lepri, Wei Wang, and Marco De Nadai. Smoothing the disentangled latent style space for unsupervised image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the" + ], + "bbox": [ + 502, + 134, + 884, + 859 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7610", + "bbox": [ + 468, + 883, + 500, + 894 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Conference on Computer Vision and Pattern Recognition, 2023. 3", + "[27] OpenAI. https://github.com/openai/dall-e. 6, 13", + "[28] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 2, 3", + "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3", + "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3", + "[33] Ali Razavi, Aïron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 3", + "[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI, 2015. 3", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3, 12", + "[36] Artsiom Sanakoyeu, Dmytro Kotovenko, Sabine Lang, and Björn Ommer. A style-aware content loss for real-time hd style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), pages 698-714, 10 2018. 2, 3", + "[37] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6924-6932, 2017. 2, 3", + "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5" + ], + "bbox": [ + 87, + 133, + 470, + 859 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Phil Wang. https://github.com/ai-forever/ru-dalle. 6, 13", + "[40] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: Towards aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia (ACM MM), 2022. 2, 3, 9", + "[41] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7789-7798, 2020. 3", + "[42] Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Mach. Learn., 8(3-4):229-256, may 1992. 5", + "[43] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2", + "[44] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1467-1475, 2019. 3", + "[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2, 3", + "[46] Yulun Zhang, Chen Fang, Yilin Wang, Zhaowen Wang, Zhe Lin, Yun Fu, and Jimei Yang. Multimodal style transfer via graph cuts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5943-5951, 2019. 2", + "[47] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + ], + "bbox": [ + 502, + 133, + 884, + 613 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7611", + "bbox": [ + 468, + 883, + 500, + 895 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/39699168-4600-4fd3-a205-b0cf85e92a6d_model.json b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/39699168-4600-4fd3-a205-b0cf85e92a6d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3e73ac6f84f7ad3d885085042f8c9468c7a20e07 --- /dev/null +++ b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/39699168-4600-4fd3-a205-b0cf85e92a6d_model.json @@ -0,0 +1,2125 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.096, + 0.009, + 0.171, + 0.046 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.232, + 0.008, + 0.816, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.162, + 0.169, + 0.808, + 0.208 + ], + "angle": 0, + "content": "StylerDALLE: Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.233, + 0.786, + 0.282 + ], + "angle": 0, + "content": "Zipeng Xu1 Enver Sangineto2 Nicu Sebe1 \n1University of Trento, Italy 2University of Modena and Reggio Emilia, Italy {zipeng.xu, niculae.sebe}@unitn.it enver.sangineto@unimore.it" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.315, + 0.874, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.73, + 0.885, + 0.759 + ], + "angle": 0, + "content": "Figure 1: With a large-scale pretrained vector-quantized tokenizer (e.g., the dVAE of DALL-E) and CLIP, StylerDALLE can transfer various types of styles (indicated on the top), from abstract art styles to specific artist styles and more." + }, + { + "type": "title", + "bbox": [ + 0.24, + 0.77, + 0.316, + 0.784 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.806, + 0.471, + 0.86 + ], + "angle": 0, + "content": "Despite the progress made in the style transfer task, most previous work focuses on transferring only relatively simple features like color or texture, while missing more abstract concepts such as overall art expression or painter" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.771, + 0.885, + 0.853 + ], + "angle": 0, + "content": "specific traits. However, these abstract semantics can be captured by models like DALL-E or CLIP, which have been trained using huge datasets of images and textual documents. In this paper, we propose StylerDALLE, a style transfer method that exploits both of these models and uses natural language to describe abstract art styles." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.884, + 0.501, + 0.895 + ], + "angle": 0, + "content": "7601" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.087, + 0.133, + 0.47, + 0.309 + ], + "angle": 0, + "content": "Specifically, we formulate the language-guided style transfer task as a non-autoregressive token sequence translation, i.e., from input content image to output stylized image, in the discrete latent space of a large-scale pretrained vector-quantized tokenizer, e.g., the discrete variational auto-encoder (dVAE) of DALL-E. To incorporate style information, we propose a Reinforcement Learning strategy with CLIP-based language supervision that ensures stylization and content preservation simultaneously. Experimental results demonstrate the superiority of our method, which can effectively transfer art styles using language instructions at different granularities. Code is available at https://github.com/zipengxuc/StylerDALLE." + }, + { + "type": "title", + "bbox": [ + 0.087, + 0.335, + 0.214, + 0.35 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.358, + 0.47, + 0.602 + ], + "angle": 0, + "content": "In the last few years, a lot of work has focused on the style transfer task using a reference image as a representative of the target style, where the goal is to transfer the style of the reference to a content image [14, 16, 28, 3, 22, 8, 9, 40]. Recent improvements in this field include: reducing the artifacts [2, 25, 5, 40], modeling the style-content relationship [28, 43, 24], increasing the generation diversity [21, 37, 46] and many others. However, art styles are usually abstract concepts, e.g., pop art, fauvism, and the style of Van Gogh. To transfer these abstract art styles to a content image, the low-level features (e.g., textures and colors), which are commonly extracted from a single reference image, are not enough. A possible solution is to collect a set of reference images that can be used, e.g., to train a Generative Adversarial Network (GAN) for artist-specific style transfer [36, 47, 19]. The disadvantage of this set-based representation is the effort required to collect sufficiently large style-specific data for training." + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.603, + 0.47, + 0.86 + ], + "angle": 0, + "content": "Recently, large-scale image generative models [32, 31, 35, 45] have shown their power to generate high-quality images of various types, e.g., realistic photo, cartoon, ukiyo-e print, or painting of a specific artist. Moreover, CLIP [30], which is trained with 400 million text-image pairs, learns good joint knowledge between language and vision. Can we leverage the power of these large-scale models for style transfer? In this paper, we propose StylerDALLE, a language-guided style transfer method that uses both the vector-quantized tokenizer of large-scale image generative methods and CLIP. There are three advantages of our method. Firstly, as compared to images, language is a more natural and efficient way to describe abstract art styles and enables more flexibility. Language can directly indicate art styles at different levels, e.g., \"Van Gogh style oil painting\" and \"Van Gogh starry night style oil painting\". Secondly, using CLIP and the language description of style saves the effort of collecting style images, as CLIP already learns style-related knowledge and thus can be used to provide" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.133, + 0.884, + 0.242 + ], + "angle": 0, + "content": "supervision. Thirdly, using a large-scale pretrained vector-quantized tokenizer potentially enables style transfer results to be closer to real artworks, as the model is trained from an enormous number of real-world images. In concrete, we propose a Non-Autoregressive Transformer (NAT) model [15], which translates tokens of a content image to tokens of a stylized image in the discrete latent space of a pretrained visual tokenizer, and a two-stage training paradigm." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.245, + 0.884, + 0.435 + ], + "angle": 0, + "content": "First of all, since both the input content image and output stylized image can be represented by a sequence of tokens via the large-scale pretrained vector-quantized tokenizer, we formulate the language-guided style transfer task as a token sequence translation. Specifically, we design a NAT model that learns to translate a token-based representation of a low-resolution image into a full-resolution representation, where the final token sequence contains appearance details specialized for the target style. The most important advantage of using NAT in image generation, with respect to the more common autoregressive transformer (AT) based generation, is that NAT is much faster than AT at inference time, as it allows a parallel token generation while AT generates in a token-by-token way." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.44, + 0.884, + 0.86 + ], + "angle": 0, + "content": "Secondly, we propose to use CLIP-based language supervision to ensure stylization and content preservation simultaneously, saving the effort to collect data and design dedicated losses. A similar solution that uses CLIP to do style transfer is CLIPStyler [20]. Since merely maximizing the CLIP similarity with respect to a textual style description is not enough for a style transfer task, which should also preserve the content of the source image, CLIPstyler introduces a hybrid of losses, including a content loss measured in an external pre-trained VGG network. Conversely, we propose an alternative direction, which avoids the need to tune the coefficients of different loss functions so as to keep the balance between style and content. Specifically, we introduce a two-stage training paradigm: 1) a self-supervised pretraining stage, where the model learns to add semantically coherent image details from a low-resolution image to a high-resolution image; and 2) a style-specific fine-tuning stage, where the model learns to incorporate style into the high-resolution image. Since the fine-tuning phase is built on top of the first stage, the translator is able to keep the semantic consistency with respect to the input image as learned during pretraining. Moreover, we create a textual prompt by concatenating both the style and the textual description of the image content (i.e., its caption). This way, the prompt simultaneously models both the target appearance (i.e., the style description) and the image content which should be preserved by the translation process. Finally, since our translator's output is discrete and there is no ground-truth tokens for a stylized image, we introduce a Reinforcement Learning (RL) approach to fine-tune the translator using a reward based on the CLIP similarity be" + }, + { + "type": "page_number", + "bbox": [ + 0.47, + 0.885, + 0.501, + 0.895 + ], + "angle": 0, + "content": "7602" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.085, + 0.133, + 0.47, + 0.174 + ], + "angle": 0, + "content": "tween the stylized image and textual prompt, enabling the model to explore the answers in the latent space of a pretrained vector-quantized tokenizer." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.175, + 0.473, + 0.31 + ], + "angle": 0, + "content": "We call our network StylerDALLE and we show that it can generate stylized images driven by different types of language guidance. Compared with previous language-guided and reference image-based transfer methods, our generated images are less inclined to produce artifacts or semantic errors. Moreover, they can capture abstract concepts related to the target style (e.g., the typical brushstrokes of the artist) besides low-level features like texture and colors. We illustrate the effectiveness of our method through qualitative results, quantitative results, and a user study." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.312, + 0.371, + 0.325 + ], + "angle": 0, + "content": "To conclude, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.336, + 0.47, + 0.392 + ], + "angle": 0, + "content": "- We propose StylerDALLE, a language-guided style transfer method that manipulates the discrete latent space of a pretrained vector-quantized tokenizer using a token sequence translation approach." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.403, + 0.472, + 0.445 + ], + "angle": 0, + "content": "- We propose a non-autoregressive translation network that translates a low-resolution content image into a full-resolution image with style-specific details." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.457, + 0.47, + 0.498 + ], + "angle": 0, + "content": "- We propose a two-stage training procedure, including an RL strategy to ensure stylization and content preservation using CLIP-based language supervision." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.51, + 0.47, + 0.565 + ], + "angle": 0, + "content": "- Experimental results show that StylerDALLE can effectively transfer abstract style concepts, going beyond simple texture and color features while simultaneously preserving the semantic content of the translated scene." + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.336, + 0.472, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.579, + 0.225, + 0.593 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.602, + 0.473, + 0.862 + ], + "angle": 0, + "content": "Reference Image-Based Style Transfer. Gatys et al. [14] propose a neural style transfer method in which a pretrained CNN is used to extract content and style information from images, and to transfer the latter from an image to another. Following this pioneering work, a lot of interests have been attracted, with different methods focusing on different aspects of the topic, such as, e.g., diversified style transfer [37, 41], or attention mechanisms to fuse style and content [44, 28, 24]. A specific line of work focuses on artistic style transfer. For instance, Chen et al. [3] propose to use internal-external learning and contrastive learning with GANs to bridge the gap between human artworks and AI-created artworks. Wang et al. [40] introduce an aesthetic discriminator trained with a large corpus of human-created artworks. Other works train GANs using an artist-specific collection of images [36, 19, 4]. In contrast, we use the generic visual-language semantics embedded in the large-scale pretrained vector-quantized tokenizer and CLIP to avoid collecting style or artist-specific datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.133, + 0.887, + 0.338 + ], + "angle": 0, + "content": "Language-Guided Style Transfer. Very recently a few works have proposed transferring methods conditioned on a textual description of the style. For instance, Fu et al. [12] use contrastive learning to train a GAN for artistic style transfer, but they adopt descriptive language instructions rather than more abstract style concepts. Gal et al. [13] use the CLIP space for a domain adaptation of a pre-trained StyleGAN [17]. The method closest to our approach is CLIPStyler [20], where a patch-wise CLIP loss is used to train a U-Net [34]. However, to condition the style change while preserving the image content, CLIPStyler uses hybrid losses and a rejection threshold, introducing many hyperparameters. In our method, we only use CLIP-based language supervision to ensure the style and content, saving the efforts of designing losses and tuning hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.887, + 0.52 + ], + "angle": 0, + "content": "Large-scale Text-to-Image Generation Models. Recently, text-to-image models trained with large or very large scale datasets [32, 31, 35, 45, 1] have attracted tremendous attention because of their excellent performance in generating high-quality images starting from a textual query. Inspired by their strong ability to synthesize various types of images, we study how to use large-scale text-to-image models for style transfer. Specifically, we focus on transformer-based methods, which are potentially less time-consuming as compared to diffusion model-based methods [26]. In addition, we propose to use a non-autoregressive transformer that can generate tokens in parallel." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.533, + 0.627, + 0.548 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.886, + 0.815 + ], + "angle": 0, + "content": "Vector-Quantzied Image Tokenizer. Transformer-based text-to-image generative models [32, 10, 11, 45, 1] rely on a vector-quantized image tokenizer to produce a discrete representation of the images, e.g., DALL-E [32] has a dVAE [33] and PARTI [45] has a ViT-VQGAN. Despite the differences in implementing the tokenizers, the effects are the same. In more detail, an image \\( I \\) is transformed into a \\( k \\times k \\) grid of tokens \\( X = \\{x_{i,j}\\}_{i,j=1,\\dots,k} \\), using an encoder \\( E(\\cdot) \\). Each token \\( x_{i,j} \\in X = E(I) \\) is an index of a codebook of embeddings (\\( C = \\{e_1, \\ldots, e_M\\} \\), \\( 1 \\leq x_{i,j} \\leq M \\)), built during the training, and corresponds to a patch in \\( I \\). A decoder \\( G(\\cdot) \\) takes as input a grid of embeddings and reconstructs the original image: \\( \\tilde{I} = G(\\{e_{x_{i,j}}\\}_{i,j=1,\\dots,k}) \\). Training in the transformer-based text-to-image generative models consists of two stages. The first stage is dedicated to training the image tokenizer, while in the second phase, a transformer is used to learn a prior distribution over the text and the image tokens. In StylerDALLE (Sec. 4) we only use the pretrained image tokenizer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.833, + 0.885, + 0.861 + ], + "angle": 0, + "content": "Non-Autoregressive Transformer (NAT). Gu et al. [15] propose a NAT for natural language translation, which con" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.884, + 0.502, + 0.896 + ], + "angle": 0, + "content": "7603" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.128, + 0.581, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.129, + 0.876, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.083, + 0.396, + 0.889, + 0.496 + ], + "angle": 0, + "content": "Figure 2: A schematic illustration of StylerDALLE. We propose a NAT to translate tokens of a half-resolution content image to tokens of a full-resolution stylized image. StylerDALLE consists of three parts: 1) image tokenization preprocessing, to obtain the discrete visual tokens from a pretrained image tokenizer; 2) self-supervised pretaining, to train the NAT to predict full-resolution content images from half-resolution content images; and 3) style-specific fine-tuning, to let the model adding style details in the up-scaling prediction via Reinforcement Learning with language-prompted CLIP guidance. Specifically, in our NAT, the decoder input is duplicated from the tokens of the half-resolution image, following the natural positional correspondence (as highlighted by orange rectangles in the left part) with the full-resolution image." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.515, + 0.475, + 0.625 + ], + "angle": 0, + "content": "sits of an encoder and a decoder. The encoder takes as input a source sentence \\( X = \\{x_{1},\\dots,x_{N^{\\prime}}\\} \\) of \\( N^{\\prime} \\) tokens and outputs a distribution over possible output sentences \\( Y = \\{y_{1},\\dots,y_{N}\\} \\), where \\( Y \\) is the translation of \\( X \\) in the target language and, usually, \\( N\\neq N^{\\prime} \\). The main novelty of NAT with respect to AT is that, during training, NAT uses a conditional independent factorization for the target sentence and the following log-likelihood:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.629, + 0.47, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = \\sum_ {n = 1} ^ {N} \\log p \\left(y _ {n} \\mid x _ {1: N ^ {\\prime}}; \\theta\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.67, + 0.473, + 0.862 + ], + "angle": 0, + "content": "which differs from the common AT factorization in which the prediction of the \\( n \\)-th token \\( (y_{n}) \\) depends on the previously predicted tokens \\( (p(y_{n}|y_{0:n-1}, x_{1:N'}; \\theta)) \\). This conditional independence assumption makes it possible a parallel generation of \\( Y \\) at inference time, largely accelerating the translation time with respect to AT models. Importantly, to make parallel generation possible, the encoder input \\( (X) \\) is provided as input to the decoder as well, and individual tokens \\( (x_{n} \\in X) \\) can be copied zero or more times, with the number of times each input is copied depending on a specific predicted \"fertility\" value. As we will see in Sec. 4.1, we do not need to predict fertilities because, in our case, the cardinality of the input copies is fixed and determined by the up-scaling task we use for the image translation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.59, + 0.528 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.886, + 0.771 + ], + "angle": 0, + "content": "The language-guided style transfer task can be described as follows. Given an image \\( I \\), we want to generate a new image \\( I^s \\) which preserves the semantic content of \\( I \\) but changes its appearance according to a textual style description \\( t_s \\) (e.g., \"cartoon\"). In StylerDALLE, we formulate this task as a visual-token based translation, from tokens of a content image to tokens of a stylized image. Specifically, given a content image \\( I \\) we first downsample \\( I \\) to get a half-resolution image \\( I' \\). Then, \\( I' \\) is fed to the tokenization encoder (Sec. 3) which extracts a discrete grid of \\( k \\times k \\) source tokens \\( X' = E(I') \\). \\( X' \\) can now be \"translated\" into a target (discrete) representation \\( \\hat{Y} \\), where \\( \\hat{Y} \\) is a grid at the original resolution \\( (2k \\times 2k) \\) and \\( \\hat{Y} = f(X') \\), being \\( f(\\cdot) \\) the translation function. Finally, \\( \\hat{Y} \\) is fed into the tokenization decoder obtaining the stylized image \\( I^s = G(\\hat{Y}) \\). In the following subsections, we describe the architecture of \\( f(\\cdot) \\) and the way in which it is trained." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.783, + 0.878, + 0.799 + ], + "angle": 0, + "content": "4.1. Architecture and self-supervised pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.805, + 0.886, + 0.862 + ], + "angle": 0, + "content": "For our translation network \\( f(\\cdot) \\) we use a NAT architecture [15] (Sec. 3) which we train from scratch using a self-supervised learning pretext task consisting in predicting the image at full resolution. Specifically, given the downsam-" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.884, + 0.503, + 0.896 + ], + "angle": 0, + "content": "7604" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.085, + 0.133, + 0.473, + 0.35 + ], + "angle": 0, + "content": "pled image \\(I'\\) and its corresponding grid of tokens \\(X' = E(I')\\), we use the indexes in \\(X'\\) to extract the corresponding embeddings from \\(C\\) (Sec. 3). For each \\(x_{i,j} \\in X'\\), let \\(\\pmb{e}_{x_{i,j}}\\) be the corresponding embedding in \\(C\\) and let \\(N' = k^2\\). The so obtained set of embeddings is flattened into a sequence, and, for each element \\(\\pmb{e}_n\\) of the sequence (\\(1 \\leq n \\leq N'\\)), we add an absolute positional embedding [38] \\(\\pmb{p}_n\\), where \\(\\pmb{p}_n\\) has the same dimension as \\(\\pmb{e}_n\\): \\(\\pmb{v}_n = \\pmb{e}_n + \\pmb{p}_n\\). The final sequence \\(V' = \\{\\pmb{v}_1, \\dots, \\pmb{v}_n, \\dots, \\pmb{v}_{N'}\\}\\) is input to the encoder of \\(f(\\cdot)\\). Note that an alternative solution is to directly fed \\(f(\\cdot)\\) with (a flattened version of) \\(X'\\) and let \\(f(\\cdot)\\) learn its own initial token embedding. However, using the embeddings in \\(C\\) has the advantage of exploiting the image representation of the pre-trained image tokenizer. Moreover, from the original image \\(I\\) we extract the ground truth \\(X = E(I)\\), which is flattened in a sequence of \\(N = 4k^2\\) tokens." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.351, + 0.473, + 0.568 + ], + "angle": 0, + "content": "Finally, following [15], we build a second sequence of input embeddings \\( V \\), with cardinality \\( N \\), which is fed to the decoder of \\( f(\\cdot) \\). As mentioned in Sec. 3, differently from NAT, we do not predict fertility values. Instead, we get the input of the decoder by simply replicating each element \\( x_{i,j} \\in X' \\) according to the positional correspondences between the low-resolution image and high-resolution image (as in Fig. 2). The rationale behind this choice is that \\( f(\\cdot) \\) is trained to predict the full-resolution image, and each encoder input \\( (\\pmb{e}_{x_{i,j}}) \\) corresponds to a patch in the subsampled image \\( I' \\) and to 4 patches in the full-resolution image \\( I \\). Thus, initializing the decoder with 4 replicas of each source-image patch initial embedding provides a coarse-grained signal for the upsampling task. Similar to before, the embeddings extracted from \\( C \\) are then flattened and added with a new positional encoding." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.568, + 0.472, + 0.678 + ], + "angle": 0, + "content": "Both the encoder and the decoder have self-attention layers and no causal masking is used. However, following [15], in the decoder, we mask out each query position \\((n)\\) only from attending to itself. Using \\(V^{\\prime}\\) and \\(V\\), \\(f(\\cdot)\\) generates \\(N\\) parallel posterior distributions over the visual vocabulary \\((\\{1,\\dots,M\\})\\): \\(P = f_{\\theta}(V',V)\\), where \\(P\\) is a \\(N\\times M\\) matrix, \\(P_{n}\\in [0,1]^{M}\\) and \\(P_{n}[y] = p_{\\theta}(Y_{n} = y|V^{\\prime},V)\\). Using \\(Y = \\{y_{1},\\ldots ,y_{n},\\ldots ,y_{N}\\}\\), \\(f(\\cdot)\\) is trained to maximize:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.689, + 0.47, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r e - t r a i n}} (\\theta) = \\sum_ {n = 1} ^ {N} \\log P _ {n} [ y _ {n} ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.733, + 0.47, + 0.803 + ], + "angle": 0, + "content": "This pre-training stage is independent of the target style and it can be shared over different styles. After this stage, \\( f(\\cdot) \\) is able to generate realistic low-level details (which are missing in \\( I' \\)). In the next, we describe how a specific style is incorporated in \\( f(\\cdot) \\) using a fine-tuning phase." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.811, + 0.308, + 0.827 + ], + "angle": 0, + "content": "4.2. Style-specific fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.833, + 0.47, + 0.861 + ], + "angle": 0, + "content": "Given a style description provided with a textual sentence \\( t_s \\), the goal is to fine-tune the pre-trained translator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.133, + 0.887, + 0.298 + ], + "angle": 0, + "content": "\\(f(\\cdot)\\) (Sec. 4.1) to make it generate image details in the style of \\(t_s\\). We fine-tune only the decoder of \\(f(\\cdot)\\), keeping frozen the encoder. To ensure both stylization and content preservation, we design a prompt that consists of two parts, i.e., a style description \\(t_s\\) and an image caption \\(t_a\\), which describe the image content and can be obtained from a generic image captioning dataset. For instance, given \\(t_a = \"A man's hand is adjusting his black tie\" and \\(t_s = \"cartoon\"\\), we obtain \\(t_q = \"a cartoon of a man's hand is adjusting his black tie.\\). On the other hand, in order to represent the image generated by \\(f(\\cdot)\\), we first need to sample the distributions in \\(P\\) (Sec. 4.1), and we do so using multinomial sampling:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.322, + 0.884, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} _ {n} = \\text {S a m p l i n g} \\left(P _ {n} [ y ]\\right), \\forall n \\in \\{1, \\dots , N \\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.886, + 0.405 + ], + "angle": 0, + "content": "The sampled sequence \\(\\hat{Y}\\) is reshaped to a \\(2k\\times 2k\\) grid and fed to the image detokenizer to get the final image \\(I^s = G(\\hat{Y})\\). Finally, using the CLIP visual and textual encoders we compute the cosine similarity on the CLIP space:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.422, + 0.884, + 0.437 + ], + "angle": 0, + "content": "\\[\nr = \\operatorname {S i m} _ {C L I P} \\left(I ^ {s}, t _ {q}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.886, + 0.581 + ], + "angle": 0, + "content": "However, directly using Eq. 4 as the fine-tuning objective function is not possible because Eq. 3 is not differentiable. In addition, since there is no ground truth for the tokens of stylized images we use RL to encourage the model to explore the answers in the latent space of a pretrained vector-quantized model. We use the REINFORCE algorithm [42] that updates the parameters of \\( f_{\\theta}(\\cdot) \\) using the CLIP-based reward \\( r \\), which could keep awarding the model for achieving better-stylized results until the model reaches a limit. This leads to the gradient estimate:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.593, + 0.884, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta | _ {d}} \\mathcal {L} _ {\\text {f i n e - t u n e}} (\\theta | _ {d}) = \\sum_ {n = 1} ^ {N} r \\nabla_ {\\theta | _ {d}} \\log P _ {n} [ y _ {n} ], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.885, + 0.697 + ], + "angle": 0, + "content": "where \\(\\theta|_d\\) indicates the parameters of the decoder only (we found fine-tuning both could lead to content loss). By maximizing Eq. 5 we encourage \\(f(\\cdot)\\) to generate images having both the content and the style of the prompt \\(t_q\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.886, + 0.861 + ], + "angle": 0, + "content": "With respect to the whole method, the reason we use down-sampled versions of the content image is that \"style\" is commonly assumed to be involved in the low-level visual details, such as colors, texture, painting strokes, etc. \\(X^{\\prime}\\), which in our formulation represents \\(I\\) at a lower resolution, presumably keeps most of the content in \\(I\\) discarding some details, this way facilitating the style translation process. Preliminary experiments (in Appendix A.2) in which we fed the encoder with tokenized full-resolution images led to poor results, demonstrating that the different cardinality between the source and the target sequence is an essential component in this translation process." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.884, + 0.502, + 0.895 + ], + "angle": 0, + "content": "7605" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.131, + 0.887, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.472, + 0.884, + 0.514 + ], + "angle": 0, + "content": "Figure 3: Qualitative results of StylerDALLE-1 on various styles: (a) Monet Sunrise, (b) Picasso cubism, (c) Van Gogh blue color, (d) Van Gogh purple color, (e) warm and relaxing, (f) ukiyo-e print, (g) fauvism, (h) pixel art illustration. The style references at the top are for illustration only (not used as input to the model)." + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.536, + 0.216, + 0.551 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.085, + 0.561, + 0.244, + 0.575 + ], + "angle": 0, + "content": "5.1. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.586, + 0.47, + 0.695 + ], + "angle": 0, + "content": "We implement our method based on two types of pretrained vector-quantized tokenizer: 1) the officially released dVAE of DALL-E 1 [27] and 2) the VQGAN of RuDALLE [39], and consequently obtain two groups of results, noting as StylerDALLE-1 and StylerDALLE-Ru, respectively. We use the two models in our experiment because they are open-sourced while our method is applicable to any large-scale pretrained vector-quantized tokenizer." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.697, + 0.471, + 0.86 + ], + "angle": 0, + "content": "To train the model, we use the MS-COCO [23] trainset, which contains 83k images of common objects in daily scenes. In the self-supervised pretraining stage, we only use the images while in the style-specific fine-tuning stage we use both images and captions. The StylerDALLE NAT model consists of a 4-layer encoder and an 8-layer decoder while the attention head number is 8 and the hidden dimension is 512. We use Pytorch [29] to implement our method. We train the NAT model for 25 epochs with a learning rate of 1e-4. We use Adam [18] optimizer. In the style-specific fine-tuning stage, we use both the images and the captions. In particular, we utilize all the caption annotations to en" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.885, + 0.673 + ], + "angle": 0, + "content": "hance the model robustness, as usually human-beings annotate different captions of a single image. Notably, the caption annotations are only used at the fine-tuning stage. In other words, StylerDALLE does not need to use image caption as input at inference time. We only fine-tune the decoder of the NAT model, and keep the encoder frozen. We use Adam optimizer with a learning rate of 1e-6. For CLIP model, we use the CLIP ViT-B/32 model. For both training stages, the model is trained on a single RTX-A6000 GPU for 24 hours." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.683, + 0.698, + 0.698 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.885, + 0.8 + ], + "angle": 0, + "content": "In the following, we present qualitative, quantitative, and user study results of our method and comparative methods, as well as comparisons with reference image-based methods. The additional implementation details, ablation study, inference time comparison, and additional experimental results are shown in Appendix A.1, A.2, A.3, and A.5, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.819, + 0.885, + 0.861 + ], + "angle": 0, + "content": "Qualitative Results. In Fig. 1, Fig. 3 and Fig. 4, we show that our method can effectively transfer various types of styles, i.e., a) abstract art styles, e.g., \"fauvism\" and" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.884, + 0.503, + 0.895 + ], + "angle": 0, + "content": "7606" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.133, + 0.887, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.472, + 0.889, + 0.515 + ], + "angle": 0, + "content": "Figure 4: Qualitative results of StylerDALLE-Ru on various styles: (a) Monet, (b) Monet Paris, (c) Monet Venice, (d) Van Gogh, (e) Van Gogh Irises, (f) Van Gogh Almond, (g) Van Gogh Starry Night, (h) Van Gogh Sunflowers. The style references at the top are for illustration only (not used as input to the model)." + }, + { + "type": "table", + "bbox": [ + 0.087, + 0.526, + 0.892, + 0.661 + ], + "angle": 0, + "content": "
DatasetStyle MethodFauvismMonetMonet SunriseMonetSunsetMonetParisVan Gogh IrisesVan GoghAlmond Blossoms
CoCoCLIPstyler-Fast26.1127.4624.7826.7027.6026.2327.05
Ours30.6127.7228.8130.6829.6328.1431.64
AFHQCLIPstyler-Fast26.5226.0823.2725.0525.4223.2824.31
Ours29.9626.5526.5829.1126.1028.0431.92
ImageNet-100CLIPstyler-Fast26.6327.1624.3326.1227.2826.3226.23
Ours30.2227.5328.2930.2827.6728.2631.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.173, + 0.67, + 0.796, + 0.684 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons with CLIPstyler [20] on different types of styles and datasets." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.708, + 0.472, + 0.858 + ], + "angle": 0, + "content": "\"pop art\"; b) artist-specific styles, e.g., \"Monet\" and \"Van Gogh\"; c) artist-specific styles with additional descriptions, e.g., \"Monet Paris\" and \"Van Gogh Sunflowers\"; d) artistic painting types, e.g., \"pixel art illustration\"; and e) emotional effects, e.g., \"warm and relaxing\". According to the qualitative results, we draw the following conclusions: 1) StylerDALLE can transfer abstract style concepts which go beyond the texture and color features and are similar to the typical trait of the artist/artistic target style; 2) each style corresponds to generated images that are different from those of other styles; 3) the image content is well preserved;" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.885, + 0.749 + ], + "angle": 0, + "content": "and 4) StylerDALLE can be applied to open-domain content images (i.e., the image content can contain animals, human beings, daily objects, buildings, etc.)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.886, + 0.861 + ], + "angle": 0, + "content": "Further, we compare StylerDALLE with the recent language-guided style transfer method CLIPStyler [20]. CLIPstyler proposes two methods: 1) CLIPStyler-Optimization, which optimizes a style transfer network for each content image, thus being time-consuming; and 2) CLIPStyler-Fast, which is the most comparable method to ours as it trains a network for each style, and then it can be used with any content images. As shown in Fig. 5," + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.884, + 0.502, + 0.895 + ], + "angle": 0, + "content": "7607" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.134, + 0.887, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.439, + 0.885, + 0.482 + ], + "angle": 0, + "content": "Figure 5: Comparisons with language-guided methods. The styles are (a) fauxivism, (b) Monet, (c) Monet Impression Sunrise, (d) oil painting, (e) watercolor painting, (f) Van Gogh, (g) Van Gogh blue color. The style references at the top are for illustration only (not used as input to the model)." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.504, + 0.473, + 0.683 + ], + "angle": 0, + "content": "CLIPStyler-Optimization generates diverse stylized results but it suffers from inharmonious artifacts. For instance, in the column \"Monet\", there are many plants on the train. In addition, CLIPStyler-Optimization has the problems of writing the style text in the results, as in the results of \"fauvism\" and so on. On the other hand, the images generated by CLIPStyler-Fast do not contain artifacts but there is less variation among different styles. Importantly, it is hard to recognize the typical trait of each artistic style, and the main differences among the styles are the colors. In contrast, the results of StylerDALLE are much closer to the artworks of the specific artistic style, they show distinct differences among different styles, and they have no artifact issues." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.697, + 0.473, + 0.862 + ], + "angle": 0, + "content": "Quantitative Results. To do quantitative analysis, we use the CLIP similarity score, formalized as score \\(= \\operatorname{Sim}_{\\text{CLIP}}(I, t_s)\\), which is computed between generated stylized images and textual description of the target style. In Tab. 1, we present the results of StylerDALLE-1 and CLIPstyler-Fast. Since both methods are applicable for arbitrary content images, we use the MS-COCO val-set, AFHQ val-set [6], and ImageNet-100 val-set [7] for evaluation. According to the results on multiple datasets, although the way we use CLIP is to compute rewards in Reinforcement Learning, instead of directly using CLIP-scores to optimize the network as CLIPstyler, we achieve comparable" + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.503, + 0.878, + 0.558 + ], + "angle": 0, + "content": "
(%)CLIPstyler-FastStyler DALLE-1Styler DALLE-Ru
Preference15.9015.3368.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.56, + 0.567, + 0.825, + 0.582 + ], + "angle": 0, + "content": "Table 2: Preference scores of user study." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.608, + 0.885, + 0.634 + ], + "angle": 0, + "content": "and even better quantitative results, indicating the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.887, + 0.848 + ], + "angle": 0, + "content": "User Study. We conduct a user study to see human opinions towards the stylized images coming from different methods: 1) CLIPstyler-Fast, 2) StylerDALLE-1, and 3) StylerDALLE-Ru. In specific, we collect opinions from 35 human subjects with a 30-question questionnaire. In each question, we ask them to select one stylized result that is closest to a target style. As in Tab. 2, among the three methods, StylerDALLE-Ru achieves the highest preference score of \\(68.76\\%\\), indicating the superiority of our method. We also find that humans prefer StylerDALLE-Ru much more than StylerDALLE-1. We infer this could be because the results from StylerDALLE-1 are blurry, as they are based on the dVAE of DALLE-1, and humans dislike blurry images. More user study details are given in A.4." + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.884, + 0.503, + 0.896 + ], + "angle": 0, + "content": "7608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.131, + 0.207, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.222, + 0.16, + 0.231 + ], + "angle": 0, + "content": "ref1" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.234, + 0.206, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.318, + 0.156, + 0.327 + ], + "angle": 0, + "content": "ref2" + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.131, + 0.304, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.2, + 0.304, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.269, + 0.304, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.337, + 0.278, + 0.346 + ], + "angle": 0, + "content": "Content" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.131, + 0.403, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.2, + 0.403, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.269, + 0.403, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.337, + 0.376, + 0.346 + ], + "angle": 0, + "content": "AesUST" + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.348, + 0.376, + 0.357 + ], + "angle": 0, + "content": "w/ ref1" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.131, + 0.498, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.2, + 0.498, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.269, + 0.498, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.337, + 0.469, + 0.346 + ], + "angle": 0, + "content": "StyTr2" + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.348, + 0.474, + 0.357 + ], + "angle": 0, + "content": "w/ ref1" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.131, + 0.593, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.2, + 0.593, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.269, + 0.593, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.337, + 0.586, + 0.346 + ], + "angle": 0, + "content": "StylerDALLE-Ru" + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.348, + 0.566, + 0.357 + ], + "angle": 0, + "content": "w/text" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.131, + 0.692, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.2, + 0.692, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.269, + 0.692, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.337, + 0.662, + 0.346 + ], + "angle": 0, + "content": "AesUST" + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.348, + 0.661, + 0.357 + ], + "angle": 0, + "content": "w/ ref2" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.131, + 0.787, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.2, + 0.787, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.269, + 0.787, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.337, + 0.765, + 0.346 + ], + "angle": 0, + "content": "StyTr2" + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.348, + 0.767, + 0.357 + ], + "angle": 0, + "content": "w/ ref2" + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.131, + 0.881, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.2, + 0.881, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.269, + 0.881, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.337, + 0.877, + 0.346 + ], + "angle": 0, + "content": "StylerDALLE-Ru" + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.348, + 0.858, + 0.357 + ], + "angle": 0, + "content": "w/text" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.38, + 0.884, + 0.409 + ], + "angle": 0, + "content": "Figure 6: Comparisons with reference image-based methods. StylerDALLE is able to transfer more abstract concepts, e.g., specific painting strokes, and is less likely to produce semantic errors." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.432, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Comparisons with Reference Image-Based Methods. We compare StylerDALLE with state-of-the-art reference image-based methods: 1) AesUST [40], an arbitrary style transfer method which enhances the aesthetic reality using a GAN trained with a collection of artworks; and 2) StyTr2 [9], an arbitrary style transfer method which uses a transformer to eliminate the biased content representation issues of CNN-based methods. To make the comparison feasible, we show the results of AesUST and StyTr2 using two Van Gogh paintings as the reference images, and the results of StylerDALLE trained using the corresponding language description." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.602, + 0.473, + 0.86 + ], + "angle": 0, + "content": "As shown in Fig. 6, the results of StylerDALLE are distinctive from the results of reference image-based methods. In concrete, the results of both AesUST and StyleTr2 are mostly affected by the colors and the textures of the reference images, and to some extent in an unnatural way. For instance, on the bottom line, in the \"Van Gogh Irises\" stylized results of AesUST and StyleTr2, the textures of the irises get transferred to the orange cup. In transferring the style of \"Van Gogh Starry Night\", the objects in the results of AesUST and StyleTr2 are mostly in the same dark-sky blue as the reference image, making the scenes a bit in-realistic. By contrast, the colors and textures used in the reference style are well transferred while being appropriately applied to the contents, without changing the original semantics. Moreover, StylerDALLE achieves to transfer higher-level style features, e.g., the brushstroke, rather than merely colors and textures, leading to the results of a similar style to the target one. For example, the strokes in Starry Night are sharper as compared to the ones in Irises, and the differences are also" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.432, + 0.833, + 0.446 + ], + "angle": 0, + "content": "presented in the stylized results of StylerDALLE." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.456, + 0.617, + 0.469 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.885, + 0.696 + ], + "angle": 0, + "content": "We present StylerDALLE, a language-guided style transfer method that leverages the power of the large-scale pretrained vector-quantized image tokenizer and CLIP. Specifically, inspired by natural language translation, we propose a non-autoregressive sequence translation approach to manipulate the discrete visual tokens, from the content image to the stylized image. We use Reinforcement Learning to include CLIP-based language supervision on the style and content. Differently from previous work, StylerDALLE can transfer abstract style concepts that are implicitly represented in the pretrained image tokenizer and CLIP and which cannot be easily obtained using reference images. Moreover, using the large-scale pretrained latent space as the basic image representation makes it possible to reduce the artifacts and the semantic incoherence better than the previous work that operates at the pixel level." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.71, + 0.884, + 0.765 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by the MUR PNRR project FAIR (PE00000013) funded by the NextGenerationEU and by the PRIN project CREATIVE (Prot. 2020ZSL9F9)." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.775, + 0.596, + 0.789 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.797, + 0.885, + 0.86 + ], + "angle": 0, + "content": "[1] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 3, 12" + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.884, + 0.502, + 0.895 + ], + "angle": 0, + "content": "7609" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.134, + 0.47, + 0.196 + ], + "angle": 0, + "content": "[2] Haibo Chen, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, Dongming Lu, et al. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems, 34:26561-26573, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.199, + 0.472, + 0.272 + ], + "angle": 0, + "content": "[3] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 872-881, June 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.274, + 0.471, + 0.336 + ], + "angle": 0, + "content": "[4] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dualist: Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 872-881, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.338, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[5] Jiaxin Cheng, Ayush Jaiswal, Yue Wu, Pradeep Natarajan, and Prem Natarajan. Style-aware normalized loss for improving arbitrary style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-143, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.402, + 0.47, + 0.45 + ], + "angle": 0, + "content": "[6] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.452, + 0.471, + 0.502 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.504, + 0.472, + 0.566 + ], + "angle": 0, + "content": "[8] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11326-11336, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.567, + 0.471, + 0.629 + ], + "angle": 0, + "content": "[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11326-11336, 2022. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.631, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.694, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[11] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. Advances in Neural Information Processing Systems, 35, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.746, + 0.47, + 0.783 + ], + "angle": 0, + "content": "[12] Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-driven artistic style transfer. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.785, + 0.47, + 0.834 + ], + "angle": 0, + "content": "[13] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.835, + 0.47, + 0.86 + ], + "angle": 0, + "content": "[14] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.134, + 0.472, + 0.86 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.135, + 0.885, + 0.16 + ], + "angle": 0, + "content": "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.162, + 0.884, + 0.198 + ], + "angle": 0, + "content": "[15] Jiatao Gu, James Bradbury, Caiming Xiong, Victor O. K. Li, and Richard Socher. Non-autoregressive neural machine translation. In ICLR, 2018. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.2, + 0.884, + 0.248 + ], + "angle": 0, + "content": "[16] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.885, + 0.313 + ], + "angle": 0, + "content": "[17] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.315, + 0.884, + 0.35 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.353, + 0.884, + 0.413 + ], + "angle": 0, + "content": "[19] Dmytro Kotovenko, Artsiom Sanakoyeu, Sabine Lang, and Bjorn Ommer. Content and style disentanglement for artistic style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4422-4431, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.416, + 0.885, + 0.466 + ], + "angle": 0, + "content": "[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.468, + 0.884, + 0.529 + ], + "angle": 0, + "content": "[21] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Diversified texture synthesis with feed-forward networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3920-3928, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.884, + 0.606 + ], + "angle": 0, + "content": "[22] Tianwei Lin, Zhuoqi Ma, Fu Li, Dongliang He, Xin Li, Errui Ding, Nannan Wang, Jie Li, and Xinbo Gao. Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5141-5150, June 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.884, + 0.669 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.671, + 0.884, + 0.744 + ], + "angle": 0, + "content": "[24] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.884, + 0.821 + ], + "angle": 0, + "content": "[25] Yahui Liu, Enver Sangineto, Yajing Chen, Linchao Bao, Haoxian Zhang, Nicu Sebe, Bruno Lepri, Wei Wang, and Marco De Nadai. Smoothing the disentangled latent style space for unsupervised image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.823, + 0.884, + 0.86 + ], + "angle": 0, + "content": "[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.135, + 0.885, + 0.86 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.884, + 0.502, + 0.895 + ], + "angle": 0, + "content": "7610" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.134, + 0.469, + 0.158 + ], + "angle": 0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.16, + 0.399, + 0.172 + ], + "angle": 0, + "content": "[27] OpenAI. https://github.com/openai/dall-e. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.173, + 0.471, + 0.222 + ], + "angle": 0, + "content": "[28] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.223, + 0.471, + 0.358 + ], + "angle": 0, + "content": "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.36, + 0.471, + 0.434 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.435, + 0.471, + 0.483 + ], + "angle": 0, + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.485, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.547, + 0.471, + 0.584 + ], + "angle": 0, + "content": "[33] Ali Razavi, Aïron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.585, + 0.471, + 0.634 + ], + "angle": 0, + "content": "[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.635, + 0.471, + 0.709 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.71, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[36] Artsiom Sanakoyeu, Dmytro Kotovenko, Sabine Lang, and Björn Ommer. A style-aware content loss for real-time hd style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), pages 698-714, 10 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.76, + 0.471, + 0.821 + ], + "angle": 0, + "content": "[37] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6924-6932, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.822, + 0.471, + 0.86 + ], + "angle": 0, + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.134, + 0.471, + 0.86 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.856, + 0.147 + ], + "angle": 0, + "content": "[39] Phil Wang. https://github.com/ai-forever/ru-dalle. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.885, + 0.209 + ], + "angle": 0, + "content": "[40] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: Towards aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia (ACM MM), 2022. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.212, + 0.885, + 0.273 + ], + "angle": 0, + "content": "[41] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7789-7798, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.275, + 0.885, + 0.312 + ], + "angle": 0, + "content": "[42] Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Mach. Learn., 8(3-4):229-256, may 1992. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.313, + 0.885, + 0.374 + ], + "angle": 0, + "content": "[43] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.885, + 0.437 + ], + "angle": 0, + "content": "[44] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1467-1475, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.885, + 0.501 + ], + "angle": 0, + "content": "[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.502, + 0.885, + 0.552 + ], + "angle": 0, + "content": "[46] Yulun Zhang, Chen Fang, Yilin Wang, Zhaowen Wang, Zhe Lin, Yun Fu, and Jimei Yang. Multimodal style transfer via graph cuts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5943-5951, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.885, + 0.614 + ], + "angle": 0, + "content": "[47] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.134, + 0.885, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.469, + 0.884, + 0.501, + 0.896 + ], + "angle": 0, + "content": "7611" + } + ] +] \ No newline at end of file diff --git a/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/39699168-4600-4fd3-a205-b0cf85e92a6d_origin.pdf b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/39699168-4600-4fd3-a205-b0cf85e92a6d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1b53abffb5bfbc2f2f49c9da45e2aebf012b9e34 --- /dev/null +++ b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/39699168-4600-4fd3-a205-b0cf85e92a6d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6579980c437d7102f5ae277b64b714b9867e30f3b7e8dae4b31a9a32e57c2463 +size 9980845 diff --git a/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/full.md b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b215e1b314f9247d43b9e3660e0997b3e876ae1b --- /dev/null +++ b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/full.md @@ -0,0 +1,288 @@ +# StylerDALLE: Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model + +Zipeng Xu1 Enver Sangineto2 Nicu Sebe1 +1University of Trento, Italy 2University of Modena and Reggio Emilia, Italy {zipeng.xu, niculae.sebe}@unitn.it enver.sangineto@unimore.it + +![](images/622024d084f6d673456c683a0a951f442c00dc1edfdb78694ae4887d20181334.jpg) +Figure 1: With a large-scale pretrained vector-quantized tokenizer (e.g., the dVAE of DALL-E) and CLIP, StylerDALLE can transfer various types of styles (indicated on the top), from abstract art styles to specific artist styles and more. + +# Abstract + +Despite the progress made in the style transfer task, most previous work focuses on transferring only relatively simple features like color or texture, while missing more abstract concepts such as overall art expression or painter + +specific traits. However, these abstract semantics can be captured by models like DALL-E or CLIP, which have been trained using huge datasets of images and textual documents. In this paper, we propose StylerDALLE, a style transfer method that exploits both of these models and uses natural language to describe abstract art styles. + +Specifically, we formulate the language-guided style transfer task as a non-autoregressive token sequence translation, i.e., from input content image to output stylized image, in the discrete latent space of a large-scale pretrained vector-quantized tokenizer, e.g., the discrete variational auto-encoder (dVAE) of DALL-E. To incorporate style information, we propose a Reinforcement Learning strategy with CLIP-based language supervision that ensures stylization and content preservation simultaneously. Experimental results demonstrate the superiority of our method, which can effectively transfer art styles using language instructions at different granularities. Code is available at https://github.com/zipengxuc/StylerDALLE. + +# 1. Introduction + +In the last few years, a lot of work has focused on the style transfer task using a reference image as a representative of the target style, where the goal is to transfer the style of the reference to a content image [14, 16, 28, 3, 22, 8, 9, 40]. Recent improvements in this field include: reducing the artifacts [2, 25, 5, 40], modeling the style-content relationship [28, 43, 24], increasing the generation diversity [21, 37, 46] and many others. However, art styles are usually abstract concepts, e.g., pop art, fauvism, and the style of Van Gogh. To transfer these abstract art styles to a content image, the low-level features (e.g., textures and colors), which are commonly extracted from a single reference image, are not enough. A possible solution is to collect a set of reference images that can be used, e.g., to train a Generative Adversarial Network (GAN) for artist-specific style transfer [36, 47, 19]. The disadvantage of this set-based representation is the effort required to collect sufficiently large style-specific data for training. + +Recently, large-scale image generative models [32, 31, 35, 45] have shown their power to generate high-quality images of various types, e.g., realistic photo, cartoon, ukiyo-e print, or painting of a specific artist. Moreover, CLIP [30], which is trained with 400 million text-image pairs, learns good joint knowledge between language and vision. Can we leverage the power of these large-scale models for style transfer? In this paper, we propose StylerDALLE, a language-guided style transfer method that uses both the vector-quantized tokenizer of large-scale image generative methods and CLIP. There are three advantages of our method. Firstly, as compared to images, language is a more natural and efficient way to describe abstract art styles and enables more flexibility. Language can directly indicate art styles at different levels, e.g., "Van Gogh style oil painting" and "Van Gogh starry night style oil painting". Secondly, using CLIP and the language description of style saves the effort of collecting style images, as CLIP already learns style-related knowledge and thus can be used to provide + +supervision. Thirdly, using a large-scale pretrained vector-quantized tokenizer potentially enables style transfer results to be closer to real artworks, as the model is trained from an enormous number of real-world images. In concrete, we propose a Non-Autoregressive Transformer (NAT) model [15], which translates tokens of a content image to tokens of a stylized image in the discrete latent space of a pretrained visual tokenizer, and a two-stage training paradigm. + +First of all, since both the input content image and output stylized image can be represented by a sequence of tokens via the large-scale pretrained vector-quantized tokenizer, we formulate the language-guided style transfer task as a token sequence translation. Specifically, we design a NAT model that learns to translate a token-based representation of a low-resolution image into a full-resolution representation, where the final token sequence contains appearance details specialized for the target style. The most important advantage of using NAT in image generation, with respect to the more common autoregressive transformer (AT) based generation, is that NAT is much faster than AT at inference time, as it allows a parallel token generation while AT generates in a token-by-token way. + +Secondly, we propose to use CLIP-based language supervision to ensure stylization and content preservation simultaneously, saving the effort to collect data and design dedicated losses. A similar solution that uses CLIP to do style transfer is CLIPStyler [20]. Since merely maximizing the CLIP similarity with respect to a textual style description is not enough for a style transfer task, which should also preserve the content of the source image, CLIPstyler introduces a hybrid of losses, including a content loss measured in an external pre-trained VGG network. Conversely, we propose an alternative direction, which avoids the need to tune the coefficients of different loss functions so as to keep the balance between style and content. Specifically, we introduce a two-stage training paradigm: 1) a self-supervised pretraining stage, where the model learns to add semantically coherent image details from a low-resolution image to a high-resolution image; and 2) a style-specific fine-tuning stage, where the model learns to incorporate style into the high-resolution image. Since the fine-tuning phase is built on top of the first stage, the translator is able to keep the semantic consistency with respect to the input image as learned during pretraining. Moreover, we create a textual prompt by concatenating both the style and the textual description of the image content (i.e., its caption). This way, the prompt simultaneously models both the target appearance (i.e., the style description) and the image content which should be preserved by the translation process. Finally, since our translator's output is discrete and there is no ground-truth tokens for a stylized image, we introduce a Reinforcement Learning (RL) approach to fine-tune the translator using a reward based on the CLIP similarity be + +tween the stylized image and textual prompt, enabling the model to explore the answers in the latent space of a pretrained vector-quantized tokenizer. + +We call our network StylerDALLE and we show that it can generate stylized images driven by different types of language guidance. Compared with previous language-guided and reference image-based transfer methods, our generated images are less inclined to produce artifacts or semantic errors. Moreover, they can capture abstract concepts related to the target style (e.g., the typical brushstrokes of the artist) besides low-level features like texture and colors. We illustrate the effectiveness of our method through qualitative results, quantitative results, and a user study. + +To conclude, our main contributions are: + +- We propose StylerDALLE, a language-guided style transfer method that manipulates the discrete latent space of a pretrained vector-quantized tokenizer using a token sequence translation approach. +- We propose a non-autoregressive translation network that translates a low-resolution content image into a full-resolution image with style-specific details. +- We propose a two-stage training procedure, including an RL strategy to ensure stylization and content preservation using CLIP-based language supervision. +- Experimental results show that StylerDALLE can effectively transfer abstract style concepts, going beyond simple texture and color features while simultaneously preserving the semantic content of the translated scene. + +# 2. Related Work + +Reference Image-Based Style Transfer. Gatys et al. [14] propose a neural style transfer method in which a pretrained CNN is used to extract content and style information from images, and to transfer the latter from an image to another. Following this pioneering work, a lot of interests have been attracted, with different methods focusing on different aspects of the topic, such as, e.g., diversified style transfer [37, 41], or attention mechanisms to fuse style and content [44, 28, 24]. A specific line of work focuses on artistic style transfer. For instance, Chen et al. [3] propose to use internal-external learning and contrastive learning with GANs to bridge the gap between human artworks and AI-created artworks. Wang et al. [40] introduce an aesthetic discriminator trained with a large corpus of human-created artworks. Other works train GANs using an artist-specific collection of images [36, 19, 4]. In contrast, we use the generic visual-language semantics embedded in the large-scale pretrained vector-quantized tokenizer and CLIP to avoid collecting style or artist-specific datasets. + +Language-Guided Style Transfer. Very recently a few works have proposed transferring methods conditioned on a textual description of the style. For instance, Fu et al. [12] use contrastive learning to train a GAN for artistic style transfer, but they adopt descriptive language instructions rather than more abstract style concepts. Gal et al. [13] use the CLIP space for a domain adaptation of a pre-trained StyleGAN [17]. The method closest to our approach is CLIPStyler [20], where a patch-wise CLIP loss is used to train a U-Net [34]. However, to condition the style change while preserving the image content, CLIPStyler uses hybrid losses and a rejection threshold, introducing many hyperparameters. In our method, we only use CLIP-based language supervision to ensure the style and content, saving the efforts of designing losses and tuning hyperparameters. + +Large-scale Text-to-Image Generation Models. Recently, text-to-image models trained with large or very large scale datasets [32, 31, 35, 45, 1] have attracted tremendous attention because of their excellent performance in generating high-quality images starting from a textual query. Inspired by their strong ability to synthesize various types of images, we study how to use large-scale text-to-image models for style transfer. Specifically, we focus on transformer-based methods, which are potentially less time-consuming as compared to diffusion model-based methods [26]. In addition, we propose to use a non-autoregressive transformer that can generate tokens in parallel. + +# 3. Background + +Vector-Quantzied Image Tokenizer. Transformer-based text-to-image generative models [32, 10, 11, 45, 1] rely on a vector-quantized image tokenizer to produce a discrete representation of the images, e.g., DALL-E [32] has a dVAE [33] and PARTI [45] has a ViT-VQGAN. Despite the differences in implementing the tokenizers, the effects are the same. In more detail, an image $I$ is transformed into a $k \times k$ grid of tokens $X = \{x_{i,j}\}_{i,j=1,\dots,k}$ , using an encoder $E(\cdot)$ . Each token $x_{i,j} \in X = E(I)$ is an index of a codebook of embeddings ( $C = \{e_1, \ldots, e_M\}$ , $1 \leq x_{i,j} \leq M$ ), built during the training, and corresponds to a patch in $I$ . A decoder $G(\cdot)$ takes as input a grid of embeddings and reconstructs the original image: $\tilde{I} = G(\{e_{x_{i,j}}\}_{i,j=1,\dots,k})$ . Training in the transformer-based text-to-image generative models consists of two stages. The first stage is dedicated to training the image tokenizer, while in the second phase, a transformer is used to learn a prior distribution over the text and the image tokens. In StylerDALLE (Sec. 4) we only use the pretrained image tokenizer. + +Non-Autoregressive Transformer (NAT). Gu et al. [15] propose a NAT for natural language translation, which con + +![](images/c5af19da210c221852064406a7a8df89901eb20e8449b8810e17d021d9206e3a.jpg) +Figure 2: A schematic illustration of StylerDALLE. We propose a NAT to translate tokens of a half-resolution content image to tokens of a full-resolution stylized image. StylerDALLE consists of three parts: 1) image tokenization preprocessing, to obtain the discrete visual tokens from a pretrained image tokenizer; 2) self-supervised pretaining, to train the NAT to predict full-resolution content images from half-resolution content images; and 3) style-specific fine-tuning, to let the model adding style details in the up-scaling prediction via Reinforcement Learning with language-prompted CLIP guidance. Specifically, in our NAT, the decoder input is duplicated from the tokens of the half-resolution image, following the natural positional correspondence (as highlighted by orange rectangles in the left part) with the full-resolution image. + +![](images/436d81afcc601f3abae24e8c55599a0127c0a64965102070dbdba812dbaf3a1e.jpg) + +sits of an encoder and a decoder. The encoder takes as input a source sentence $X = \{x_{1},\dots,x_{N^{\prime}}\}$ of $N^{\prime}$ tokens and outputs a distribution over possible output sentences $Y = \{y_{1},\dots,y_{N}\}$ , where $Y$ is the translation of $X$ in the target language and, usually, $N\neq N^{\prime}$ . The main novelty of NAT with respect to AT is that, during training, NAT uses a conditional independent factorization for the target sentence and the following log-likelihood: + +$$ +\mathcal {L} (\theta) = \sum_ {n = 1} ^ {N} \log p \left(y _ {n} \mid x _ {1: N ^ {\prime}}; \theta\right), \tag {1} +$$ + +which differs from the common AT factorization in which the prediction of the $n$ -th token $(y_{n})$ depends on the previously predicted tokens $(p(y_{n}|y_{0:n-1}, x_{1:N'}; \theta))$ . This conditional independence assumption makes it possible a parallel generation of $Y$ at inference time, largely accelerating the translation time with respect to AT models. Importantly, to make parallel generation possible, the encoder input $(X)$ is provided as input to the decoder as well, and individual tokens $(x_{n} \in X)$ can be copied zero or more times, with the number of times each input is copied depending on a specific predicted "fertility" value. As we will see in Sec. 4.1, we do not need to predict fertilities because, in our case, the cardinality of the input copies is fixed and determined by the up-scaling task we use for the image translation. + +# 4. Method + +The language-guided style transfer task can be described as follows. Given an image $I$ , we want to generate a new image $I^s$ which preserves the semantic content of $I$ but changes its appearance according to a textual style description $t_s$ (e.g., "cartoon"). In StylerDALLE, we formulate this task as a visual-token based translation, from tokens of a content image to tokens of a stylized image. Specifically, given a content image $I$ we first downsample $I$ to get a half-resolution image $I'$ . Then, $I'$ is fed to the tokenization encoder (Sec. 3) which extracts a discrete grid of $k \times k$ source tokens $X' = E(I')$ . $X'$ can now be "translated" into a target (discrete) representation $\hat{Y}$ , where $\hat{Y}$ is a grid at the original resolution $(2k \times 2k)$ and $\hat{Y} = f(X')$ , being $f(\cdot)$ the translation function. Finally, $\hat{Y}$ is fed into the tokenization decoder obtaining the stylized image $I^s = G(\hat{Y})$ . In the following subsections, we describe the architecture of $f(\cdot)$ and the way in which it is trained. + +# 4.1. Architecture and self-supervised pre-training + +For our translation network $f(\cdot)$ we use a NAT architecture [15] (Sec. 3) which we train from scratch using a self-supervised learning pretext task consisting in predicting the image at full resolution. Specifically, given the downsam- + +pled image $I'$ and its corresponding grid of tokens $X' = E(I')$ , we use the indexes in $X'$ to extract the corresponding embeddings from $C$ (Sec. 3). For each $x_{i,j} \in X'$ , let $\pmb{e}_{x_{i,j}}$ be the corresponding embedding in $C$ and let $N' = k^2$ . The so obtained set of embeddings is flattened into a sequence, and, for each element $\pmb{e}_n$ of the sequence ( $1 \leq n \leq N'$ ), we add an absolute positional embedding [38] $\pmb{p}_n$ , where $\pmb{p}_n$ has the same dimension as $\pmb{e}_n$ : $\pmb{v}_n = \pmb{e}_n + \pmb{p}_n$ . The final sequence $V' = \{\pmb{v}_1, \dots, \pmb{v}_n, \dots, \pmb{v}_{N'}\}$ is input to the encoder of $f(\cdot)$ . Note that an alternative solution is to directly fed $f(\cdot)$ with (a flattened version of) $X'$ and let $f(\cdot)$ learn its own initial token embedding. However, using the embeddings in $C$ has the advantage of exploiting the image representation of the pre-trained image tokenizer. Moreover, from the original image $I$ we extract the ground truth $X = E(I)$ , which is flattened in a sequence of $N = 4k^2$ tokens. + +Finally, following [15], we build a second sequence of input embeddings $V$ , with cardinality $N$ , which is fed to the decoder of $f(\cdot)$ . As mentioned in Sec. 3, differently from NAT, we do not predict fertility values. Instead, we get the input of the decoder by simply replicating each element $x_{i,j} \in X'$ according to the positional correspondences between the low-resolution image and high-resolution image (as in Fig. 2). The rationale behind this choice is that $f(\cdot)$ is trained to predict the full-resolution image, and each encoder input $(\pmb{e}_{x_{i,j}})$ corresponds to a patch in the subsampled image $I'$ and to 4 patches in the full-resolution image $I$ . Thus, initializing the decoder with 4 replicas of each source-image patch initial embedding provides a coarse-grained signal for the upsampling task. Similar to before, the embeddings extracted from $C$ are then flattened and added with a new positional encoding. + +Both the encoder and the decoder have self-attention layers and no causal masking is used. However, following [15], in the decoder, we mask out each query position $(n)$ only from attending to itself. Using $V^{\prime}$ and $V$ , $f(\cdot)$ generates $N$ parallel posterior distributions over the visual vocabulary $(\{1,\dots,M\})$ : $P = f_{\theta}(V',V)$ , where $P$ is a $N\times M$ matrix, $P_{n}\in [0,1]^{M}$ and $P_{n}[y] = p_{\theta}(Y_{n} = y|V^{\prime},V)$ . Using $Y = \{y_{1},\ldots ,y_{n},\ldots ,y_{N}\}$ , $f(\cdot)$ is trained to maximize: + +$$ +\mathcal {L} _ {\text {p r e - t r a i n}} (\theta) = \sum_ {n = 1} ^ {N} \log P _ {n} [ y _ {n} ]. \tag {2} +$$ + +This pre-training stage is independent of the target style and it can be shared over different styles. After this stage, $f(\cdot)$ is able to generate realistic low-level details (which are missing in $I'$ ). In the next, we describe how a specific style is incorporated in $f(\cdot)$ using a fine-tuning phase. + +# 4.2. Style-specific fine-tuning + +Given a style description provided with a textual sentence $t_s$ , the goal is to fine-tune the pre-trained translator + +\(f(\cdot)\) (Sec. 4.1) to make it generate image details in the style of \(t_s\). We fine-tune only the decoder of \(f(\cdot)\), keeping frozen the encoder. To ensure both stylization and content preservation, we design a prompt that consists of two parts, i.e., a style description \(t_s\) and an image caption \(t_a\), which describe the image content and can be obtained from a generic image captioning dataset. For instance, given \(t_a = "A man's hand is adjusting his black tie" and \(t_s = "cartoon"\), we obtain \(t_q = "a cartoon of a man's hand is adjusting his black tie.\). On the other hand, in order to represent the image generated by \(f(\cdot)\), we first need to sample the distributions in \(P\) (Sec. 4.1), and we do so using multinomial sampling: + +$$ +\hat {Y} _ {n} = \text {S a m p l i n g} \left(P _ {n} [ y ]\right), \forall n \in \{1, \dots , N \}. \tag {3} +$$ + +The sampled sequence $\hat{Y}$ is reshaped to a $2k\times 2k$ grid and fed to the image detokenizer to get the final image $I^s = G(\hat{Y})$ . Finally, using the CLIP visual and textual encoders we compute the cosine similarity on the CLIP space: + +$$ +r = \operatorname {S i m} _ {C L I P} \left(I ^ {s}, t _ {q}\right). \tag {4} +$$ + +However, directly using Eq. 4 as the fine-tuning objective function is not possible because Eq. 3 is not differentiable. In addition, since there is no ground truth for the tokens of stylized images we use RL to encourage the model to explore the answers in the latent space of a pretrained vector-quantized model. We use the REINFORCE algorithm [42] that updates the parameters of $f_{\theta}(\cdot)$ using the CLIP-based reward $r$ , which could keep awarding the model for achieving better-stylized results until the model reaches a limit. This leads to the gradient estimate: + +$$ +\nabla_ {\theta | _ {d}} \mathcal {L} _ {\text {f i n e - t u n e}} (\theta | _ {d}) = \sum_ {n = 1} ^ {N} r \nabla_ {\theta | _ {d}} \log P _ {n} [ y _ {n} ], \tag {5} +$$ + +where $\theta|_d$ indicates the parameters of the decoder only (we found fine-tuning both could lead to content loss). By maximizing Eq. 5 we encourage $f(\cdot)$ to generate images having both the content and the style of the prompt $t_q$ . + +With respect to the whole method, the reason we use down-sampled versions of the content image is that "style" is commonly assumed to be involved in the low-level visual details, such as colors, texture, painting strokes, etc. $X^{\prime}$ , which in our formulation represents $I$ at a lower resolution, presumably keeps most of the content in $I$ discarding some details, this way facilitating the style translation process. Preliminary experiments (in Appendix A.2) in which we fed the encoder with tokenized full-resolution images led to poor results, demonstrating that the different cardinality between the source and the target sequence is an essential component in this translation process. + +![](images/b2252aab2985062c302386dad8df9e7dedf553e18ccb8c6db7e5fed30cdaecf5.jpg) +Figure 3: Qualitative results of StylerDALLE-1 on various styles: (a) Monet Sunrise, (b) Picasso cubism, (c) Van Gogh blue color, (d) Van Gogh purple color, (e) warm and relaxing, (f) ukiyo-e print, (g) fauvism, (h) pixel art illustration. The style references at the top are for illustration only (not used as input to the model). + +# 5. Experiments + +# 5.1. Training Details + +We implement our method based on two types of pretrained vector-quantized tokenizer: 1) the officially released dVAE of DALL-E 1 [27] and 2) the VQGAN of RuDALLE [39], and consequently obtain two groups of results, noting as StylerDALLE-1 and StylerDALLE-Ru, respectively. We use the two models in our experiment because they are open-sourced while our method is applicable to any large-scale pretrained vector-quantized tokenizer. + +To train the model, we use the MS-COCO [23] trainset, which contains 83k images of common objects in daily scenes. In the self-supervised pretraining stage, we only use the images while in the style-specific fine-tuning stage we use both images and captions. The StylerDALLE NAT model consists of a 4-layer encoder and an 8-layer decoder while the attention head number is 8 and the hidden dimension is 512. We use Pytorch [29] to implement our method. We train the NAT model for 25 epochs with a learning rate of 1e-4. We use Adam [18] optimizer. In the style-specific fine-tuning stage, we use both the images and the captions. In particular, we utilize all the caption annotations to en + +hance the model robustness, as usually human-beings annotate different captions of a single image. Notably, the caption annotations are only used at the fine-tuning stage. In other words, StylerDALLE does not need to use image caption as input at inference time. We only fine-tune the decoder of the NAT model, and keep the encoder frozen. We use Adam optimizer with a learning rate of 1e-6. For CLIP model, we use the CLIP ViT-B/32 model. For both training stages, the model is trained on a single RTX-A6000 GPU for 24 hours. + +# 5.2. Experimental Results + +In the following, we present qualitative, quantitative, and user study results of our method and comparative methods, as well as comparisons with reference image-based methods. The additional implementation details, ablation study, inference time comparison, and additional experimental results are shown in Appendix A.1, A.2, A.3, and A.5, respectively. + +Qualitative Results. In Fig. 1, Fig. 3 and Fig. 4, we show that our method can effectively transfer various types of styles, i.e., a) abstract art styles, e.g., "fauvism" and + +![](images/884181b992be00983c15537b69b7135cf17d675bc107ea1c870f94aed16dcd14.jpg) +Figure 4: Qualitative results of StylerDALLE-Ru on various styles: (a) Monet, (b) Monet Paris, (c) Monet Venice, (d) Van Gogh, (e) Van Gogh Irises, (f) Van Gogh Almond, (g) Van Gogh Starry Night, (h) Van Gogh Sunflowers. The style references at the top are for illustration only (not used as input to the model). + +
DatasetStyle MethodFauvismMonetMonet SunriseMonetSunsetMonetParisVan Gogh IrisesVan GoghAlmond Blossoms
CoCoCLIPstyler-Fast26.1127.4624.7826.7027.6026.2327.05
Ours30.6127.7228.8130.6829.6328.1431.64
AFHQCLIPstyler-Fast26.5226.0823.2725.0525.4223.2824.31
Ours29.9626.5526.5829.1126.1028.0431.92
ImageNet-100CLIPstyler-Fast26.6327.1624.3326.1227.2826.3226.23
Ours30.2227.5328.2930.2827.6728.2631.60
+ +Table 1: Quantitative comparisons with CLIPstyler [20] on different types of styles and datasets. + +"pop art"; b) artist-specific styles, e.g., "Monet" and "Van Gogh"; c) artist-specific styles with additional descriptions, e.g., "Monet Paris" and "Van Gogh Sunflowers"; d) artistic painting types, e.g., "pixel art illustration"; and e) emotional effects, e.g., "warm and relaxing". According to the qualitative results, we draw the following conclusions: 1) StylerDALLE can transfer abstract style concepts which go beyond the texture and color features and are similar to the typical trait of the artist/artistic target style; 2) each style corresponds to generated images that are different from those of other styles; 3) the image content is well preserved; + +and 4) StylerDALLE can be applied to open-domain content images (i.e., the image content can contain animals, human beings, daily objects, buildings, etc.). + +Further, we compare StylerDALLE with the recent language-guided style transfer method CLIPStyler [20]. CLIPstyler proposes two methods: 1) CLIPStyler-Optimization, which optimizes a style transfer network for each content image, thus being time-consuming; and 2) CLIPStyler-Fast, which is the most comparable method to ours as it trains a network for each style, and then it can be used with any content images. As shown in Fig. 5, + +![](images/d1e0263c65c1e8f6b7567708a4d3d752fef37a134ab05752768de1a0e77f6f6b.jpg) +Figure 5: Comparisons with language-guided methods. The styles are (a) fauxivism, (b) Monet, (c) Monet Impression Sunrise, (d) oil painting, (e) watercolor painting, (f) Van Gogh, (g) Van Gogh blue color. The style references at the top are for illustration only (not used as input to the model). + +CLIPStyler-Optimization generates diverse stylized results but it suffers from inharmonious artifacts. For instance, in the column "Monet", there are many plants on the train. In addition, CLIPStyler-Optimization has the problems of writing the style text in the results, as in the results of "fauvism" and so on. On the other hand, the images generated by CLIPStyler-Fast do not contain artifacts but there is less variation among different styles. Importantly, it is hard to recognize the typical trait of each artistic style, and the main differences among the styles are the colors. In contrast, the results of StylerDALLE are much closer to the artworks of the specific artistic style, they show distinct differences among different styles, and they have no artifact issues. + +Quantitative Results. To do quantitative analysis, we use the CLIP similarity score, formalized as score $= \operatorname{Sim}_{\text{CLIP}}(I, t_s)$ , which is computed between generated stylized images and textual description of the target style. In Tab. 1, we present the results of StylerDALLE-1 and CLIPstyler-Fast. Since both methods are applicable for arbitrary content images, we use the MS-COCO val-set, AFHQ val-set [6], and ImageNet-100 val-set [7] for evaluation. According to the results on multiple datasets, although the way we use CLIP is to compute rewards in Reinforcement Learning, instead of directly using CLIP-scores to optimize the network as CLIPstyler, we achieve comparable + +
(%)CLIPstyler-FastStyler DALLE-1Styler DALLE-Ru
Preference15.9015.3368.76
+ +Table 2: Preference scores of user study. + +and even better quantitative results, indicating the effectiveness of our method. + +User Study. We conduct a user study to see human opinions towards the stylized images coming from different methods: 1) CLIPstyler-Fast, 2) StylerDALLE-1, and 3) StylerDALLE-Ru. In specific, we collect opinions from 35 human subjects with a 30-question questionnaire. In each question, we ask them to select one stylized result that is closest to a target style. As in Tab. 2, among the three methods, StylerDALLE-Ru achieves the highest preference score of $68.76\%$ , indicating the superiority of our method. We also find that humans prefer StylerDALLE-Ru much more than StylerDALLE-1. We infer this could be because the results from StylerDALLE-1 are blurry, as they are based on the dVAE of DALLE-1, and humans dislike blurry images. More user study details are given in A.4. + +![](images/1b0b54a2c82d79c28f263e9f39ece067161b0b3b0925f56875a2d1ef1d306fa1.jpg) +ref1 + +![](images/e0feb0de3e0c2c5a9fc9bb933872ecdc0322323bc03da03a88d3dec12315f836.jpg) +ref2 + +![](images/11eda235384a8422a2baa37953a7feeb2b9d6d084033a4f4a954c2e52e8fcff7.jpg) + +![](images/0cfb287d26fbaa9be473d4cd9ef793f0dd192c2fd6a7000d64dc693d8df35ba4.jpg) + +![](images/994091be88005fc1d17d99a40e613dc5fd853ac1d06a9e6e68e7895d72cbbece.jpg) +Content + +![](images/e6ca4c8b9c8e6178a471eae45b3e40457c4a648a387228f14f85b4942ecf3526.jpg) + +![](images/171073ceedee604c21cb9354618de42ab77aa38e7773e7725cd8ee1da469a53f.jpg) + +![](images/fe810f9f01a7f65ebf0e46ab7aba97d6fb463269591315c322d43b7268a45d71.jpg) +AesUST +w/ ref1 + +![](images/083eaa866ced7f12ed3df653f4e1d9fb3b3db9b2364f7b42fbcf489249c32576.jpg) + +![](images/f617f7921ea53fcf2c29d1bcb1101fbccb2d150010651d1d1a18494b70548c59.jpg) + +![](images/40f346403ee108140b3b27570439c834d593212a0020279abf18303a7b604ab3.jpg) +StyTr2 +w/ ref1 +Figure 6: Comparisons with reference image-based methods. StylerDALLE is able to transfer more abstract concepts, e.g., specific painting strokes, and is less likely to produce semantic errors. + +![](images/eceb251da3ec631aca88e5f5a832c52bbc04b8669e988b299757c7e6a5bc757a.jpg) + +![](images/08588be497d109f3585502d97ae42489db351313a1abb3a91618fbe3b41416e5.jpg) + +![](images/7cce431f6ace421c029a8ac4c53f3d9a7085851f4fee61bd17a0a5dd95fbdad1.jpg) +StylerDALLE-Ru +w/text + +![](images/03d4d90c86d7ba54403da3b313d5c42035601f253df708171efbc188ecec8a86.jpg) + +![](images/de1d49cdb238f4e342b06bf82fe71262a93e38298c1c7796b953e3a576e478ea.jpg) + +![](images/f1f0af7d616df3041b61a491d6e96740f43aa8c0bbe9eec698d07fdf9cd9a4ac.jpg) +AesUST +w/ ref2 + +![](images/b2193636062281c30a974fb3cedccc7edb41700c1925a2901ebedbd724c8dff6.jpg) + +![](images/ada740624fc6aa034bcccc2b8e1b0292e5e93adc3bd7952c551978d8548ce841.jpg) + +![](images/93bdab9d2e0f0328ddadd12b72d3d78a2bef774be9289d852e884f7039892177.jpg) +StyTr2 +w/ ref2 + +![](images/28b4a7e5e88ec6621d5e9b520d5b4dbcfc00fef1ba8ff75e5e6b82c9ed600638.jpg) + +![](images/75d20ebfc8f91a3c0c2e52ed5b538bd47b3b4a69fd2d391e7c18d0fdf24a7b29.jpg) + +![](images/9711048784a74d21cd2af45b0fcb449172b1d3db1c6cbefcb7216463a5c56195.jpg) +StylerDALLE-Ru +w/text + +Comparisons with Reference Image-Based Methods. We compare StylerDALLE with state-of-the-art reference image-based methods: 1) AesUST [40], an arbitrary style transfer method which enhances the aesthetic reality using a GAN trained with a collection of artworks; and 2) StyTr2 [9], an arbitrary style transfer method which uses a transformer to eliminate the biased content representation issues of CNN-based methods. To make the comparison feasible, we show the results of AesUST and StyTr2 using two Van Gogh paintings as the reference images, and the results of StylerDALLE trained using the corresponding language description. + +As shown in Fig. 6, the results of StylerDALLE are distinctive from the results of reference image-based methods. In concrete, the results of both AesUST and StyleTr2 are mostly affected by the colors and the textures of the reference images, and to some extent in an unnatural way. For instance, on the bottom line, in the "Van Gogh Irises" stylized results of AesUST and StyleTr2, the textures of the irises get transferred to the orange cup. In transferring the style of "Van Gogh Starry Night", the objects in the results of AesUST and StyleTr2 are mostly in the same dark-sky blue as the reference image, making the scenes a bit in-realistic. By contrast, the colors and textures used in the reference style are well transferred while being appropriately applied to the contents, without changing the original semantics. Moreover, StylerDALLE achieves to transfer higher-level style features, e.g., the brushstroke, rather than merely colors and textures, leading to the results of a similar style to the target one. For example, the strokes in Starry Night are sharper as compared to the ones in Irises, and the differences are also + +presented in the stylized results of StylerDALLE. + +# 6. Conclusion + +We present StylerDALLE, a language-guided style transfer method that leverages the power of the large-scale pretrained vector-quantized image tokenizer and CLIP. Specifically, inspired by natural language translation, we propose a non-autoregressive sequence translation approach to manipulate the discrete visual tokens, from the content image to the stylized image. We use Reinforcement Learning to include CLIP-based language supervision on the style and content. Differently from previous work, StylerDALLE can transfer abstract style concepts that are implicitly represented in the pretrained image tokenizer and CLIP and which cannot be easily obtained using reference images. Moreover, using the large-scale pretrained latent space as the basic image representation makes it possible to reduce the artifacts and the semantic incoherence better than the previous work that operates at the pixel level. + +Acknowledgment. This work was supported by the MUR PNRR project FAIR (PE00000013) funded by the NextGenerationEU and by the PRIN project CREATIVE (Prot. 2020ZSL9F9). + +# References + +[1] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 3, 12 + +[2] Haibo Chen, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, Dongming Lu, et al. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems, 34:26561-26573, 2021. 2 +[3] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 872-881, June 2021. 2, 3 +[4] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dualist: Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 872-881, 2021. 3 +[5] Jiaxin Cheng, Ayush Jaiswal, Yue Wu, Pradeep Natarajan, and Prem Natarajan. Style-aware normalized loss for improving arbitrary style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-143, 2021. 2 +[6] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 8 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 8 +[8] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11326-11336, June 2022. 2 +[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11326-11336, 2022. 2, 9 +[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 3 +[11] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. Advances in Neural Information Processing Systems, 35, 2022. 3 +[12] Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-driven artistic style transfer. In European Conference on Computer Vision (ECCV), 2022. 3 +[13] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3 +[14] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In + +Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 3 +[15] Jiatao Gu, James Bradbury, Caiming Xiong, Victor O. K. Li, and Richard Socher. Non-autoregressive neural machine translation. In ICLR, 2018. 2, 3, 4, 5 +[16] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2 +[17] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[19] Dmytro Kotovenko, Artsiom Sanakoyeu, Sabine Lang, and Bjorn Ommer. Content and style disentanglement for artistic style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4422-4431, 2019. 2, 3 +[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 2, 3, 7 +[21] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Diversified texture synthesis with feed-forward networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3920-3928, 2017. 2 +[22] Tianwei Lin, Zhuoqi Ma, Fu Li, Dongliang He, Xin Li, Errui Ding, Nannan Wang, Jie Li, and Xinbo Gao. Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5141-5150, June 2021. 2 +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6 +[24] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 2, 3 +[25] Yahui Liu, Enver Sangineto, Yajing Chen, Linchao Bao, Haoxian Zhang, Nicu Sebe, Bruno Lepri, Wei Wang, and Marco De Nadai. Smoothing the disentangled latent style space for unsupervised image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the + +IEEE Conference on Computer Vision and Pattern Recognition, 2023. 3 +[27] OpenAI. https://github.com/openai/dall-e. 6, 13 +[28] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 2, 3 +[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3 +[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3 +[33] Ali Razavi, Aïron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 3 +[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI, 2015. 3 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3, 12 +[36] Artsiom Sanakoyeu, Dmytro Kotovenko, Sabine Lang, and Björn Ommer. A style-aware content loss for real-time hd style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), pages 698-714, 10 2018. 2, 3 +[37] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6924-6932, 2017. 2, 3 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5 + +[39] Phil Wang. https://github.com/ai-forever/ru-dalle. 6, 13 +[40] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: Towards aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia (ACM MM), 2022. 2, 3, 9 +[41] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7789-7798, 2020. 3 +[42] Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Mach. Learn., 8(3-4):229-256, may 1992. 5 +[43] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2 +[44] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1467-1475, 2019. 3 +[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2, 3 +[46] Yulun Zhang, Chen Fang, Yilin Wang, Zhaowen Wang, Zhe Lin, Yun Fu, and Jimei Yang. Multimodal style transfer via graph cuts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5943-5951, 2019. 2 +[47] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2 \ No newline at end of file diff --git a/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/images.zip b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e8310d2ada501b49a1601fa212391f32b1fed913 --- /dev/null +++ b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:345cb47f98e711863a4bc8c84dd003626e9538c042fe649f3b8eba4fdb9eb601 +size 1371095 diff --git a/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/layout.json b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4632d2ed36f64e229017bc64ccae00a21b94620f --- /dev/null +++ b/2023/StylerDALLE_ Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model/layout.json @@ -0,0 +1,8794 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 142, + 480, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 142, + 480, + 174 + ], + "spans": [ + { + "bbox": [ + 96, + 142, + 480, + 174 + ], + "type": "text", + "content": "StylerDALLE: Language-Guided Style Transfer Using a Vector-Quantized Tokenizer of a Large-Scale Generative Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 195, + 467, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 195, + 467, + 237 + ], + "spans": [ + { + "bbox": [ + 108, + 195, + 467, + 237 + ], + "type": "text", + "content": "Zipeng Xu1 Enver Sangineto2 Nicu Sebe1 \n1University of Trento, Italy 2University of Modena and Reggio Emilia, Italy {zipeng.xu, niculae.sebe}@unitn.it enver.sangineto@unimore.it" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 264, + 520, + 605 + ], + "blocks": [ + { + "bbox": [ + 58, + 264, + 520, + 605 + ], + "lines": [ + { + "bbox": [ + 58, + 264, + 520, + 605 + ], + "spans": [ + { + "bbox": [ + 58, + 264, + 520, + 605 + ], + "type": "image", + "image_path": "622024d084f6d673456c683a0a951f442c00dc1edfdb78694ae4887d20181334.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 613, + 526, + 638 + ], + "lines": [ + { + "bbox": [ + 49, + 613, + 526, + 638 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 526, + 638 + ], + "type": "text", + "content": "Figure 1: With a large-scale pretrained vector-quantized tokenizer (e.g., the dVAE of DALL-E) and CLIP, StylerDALLE can transfer various types of styles (indicated on the top), from abstract art styles to specific artist styles and more." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 647, + 188, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 647, + 188, + 659 + ], + "spans": [ + { + "bbox": [ + 142, + 647, + 188, + 659 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 677, + 280, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 677, + 280, + 723 + ], + "spans": [ + { + "bbox": [ + 49, + 677, + 280, + 723 + ], + "type": "text", + "content": "Despite the progress made in the style transfer task, most previous work focuses on transferring only relatively simple features like color or texture, while missing more abstract concepts such as overall art expression or painter" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 648, + 526, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 648, + 526, + 717 + ], + "spans": [ + { + "bbox": [ + 296, + 648, + 526, + 717 + ], + "type": "text", + "content": "specific traits. However, these abstract semantics can be captured by models like DALL-E or CLIP, which have been trained using huge datasets of images and textual documents. In this paper, we propose StylerDALLE, a style transfer method that exploits both of these models and uses natural language to describe abstract art styles." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 7, + 101, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 7, + 101, + 38 + ], + "spans": [ + { + "bbox": [ + 57, + 7, + 101, + 38 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 6, + 485, + 40 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 6, + 485, + 40 + ], + "spans": [ + { + "bbox": [ + 138, + 6, + 485, + 40 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "spans": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "type": "text", + "content": "7601" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 51, + 111, + 279, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 111, + 279, + 259 + ], + "spans": [ + { + "bbox": [ + 51, + 111, + 279, + 259 + ], + "type": "text", + "content": "Specifically, we formulate the language-guided style transfer task as a non-autoregressive token sequence translation, i.e., from input content image to output stylized image, in the discrete latent space of a large-scale pretrained vector-quantized tokenizer, e.g., the discrete variational auto-encoder (dVAE) of DALL-E. To incorporate style information, we propose a Reinforcement Learning strategy with CLIP-based language supervision that ensures stylization and content preservation simultaneously. Experimental results demonstrate the superiority of our method, which can effectively transfer art styles using language instructions at different granularities. Code is available at https://github.com/zipengxuc/StylerDALLE." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 51, + 281, + 127, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 281, + 127, + 294 + ], + "spans": [ + { + "bbox": [ + 51, + 281, + 127, + 294 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 51, + 301, + 279, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 301, + 279, + 506 + ], + "spans": [ + { + "bbox": [ + 51, + 301, + 279, + 506 + ], + "type": "text", + "content": "In the last few years, a lot of work has focused on the style transfer task using a reference image as a representative of the target style, where the goal is to transfer the style of the reference to a content image [14, 16, 28, 3, 22, 8, 9, 40]. Recent improvements in this field include: reducing the artifacts [2, 25, 5, 40], modeling the style-content relationship [28, 43, 24], increasing the generation diversity [21, 37, 46] and many others. However, art styles are usually abstract concepts, e.g., pop art, fauvism, and the style of Van Gogh. To transfer these abstract art styles to a content image, the low-level features (e.g., textures and colors), which are commonly extracted from a single reference image, are not enough. A possible solution is to collect a set of reference images that can be used, e.g., to train a Generative Adversarial Network (GAN) for artist-specific style transfer [36, 47, 19]. The disadvantage of this set-based representation is the effort required to collect sufficiently large style-specific data for training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 507, + 279, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 507, + 279, + 723 + ], + "spans": [ + { + "bbox": [ + 51, + 507, + 279, + 723 + ], + "type": "text", + "content": "Recently, large-scale image generative models [32, 31, 35, 45] have shown their power to generate high-quality images of various types, e.g., realistic photo, cartoon, ukiyo-e print, or painting of a specific artist. Moreover, CLIP [30], which is trained with 400 million text-image pairs, learns good joint knowledge between language and vision. Can we leverage the power of these large-scale models for style transfer? In this paper, we propose StylerDALLE, a language-guided style transfer method that uses both the vector-quantized tokenizer of large-scale image generative methods and CLIP. There are three advantages of our method. Firstly, as compared to images, language is a more natural and efficient way to describe abstract art styles and enables more flexibility. Language can directly indicate art styles at different levels, e.g., \"Van Gogh style oil painting\" and \"Van Gogh starry night style oil painting\". Secondly, using CLIP and the language description of style saves the effort of collecting style images, as CLIP already learns style-related knowledge and thus can be used to provide" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 299, + 111, + 525, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 111, + 525, + 203 + ], + "spans": [ + { + "bbox": [ + 299, + 111, + 525, + 203 + ], + "type": "text", + "content": "supervision. Thirdly, using a large-scale pretrained vector-quantized tokenizer potentially enables style transfer results to be closer to real artworks, as the model is trained from an enormous number of real-world images. In concrete, we propose a Non-Autoregressive Transformer (NAT) model [15], which translates tokens of a content image to tokens of a stylized image in the discrete latent space of a pretrained visual tokenizer, and a two-stage training paradigm." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 299, + 206, + 525, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 206, + 525, + 365 + ], + "spans": [ + { + "bbox": [ + 299, + 206, + 525, + 365 + ], + "type": "text", + "content": "First of all, since both the input content image and output stylized image can be represented by a sequence of tokens via the large-scale pretrained vector-quantized tokenizer, we formulate the language-guided style transfer task as a token sequence translation. Specifically, we design a NAT model that learns to translate a token-based representation of a low-resolution image into a full-resolution representation, where the final token sequence contains appearance details specialized for the target style. The most important advantage of using NAT in image generation, with respect to the more common autoregressive transformer (AT) based generation, is that NAT is much faster than AT at inference time, as it allows a parallel token generation while AT generates in a token-by-token way." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 299, + 370, + 525, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 370, + 525, + 723 + ], + "spans": [ + { + "bbox": [ + 299, + 370, + 525, + 723 + ], + "type": "text", + "content": "Secondly, we propose to use CLIP-based language supervision to ensure stylization and content preservation simultaneously, saving the effort to collect data and design dedicated losses. A similar solution that uses CLIP to do style transfer is CLIPStyler [20]. Since merely maximizing the CLIP similarity with respect to a textual style description is not enough for a style transfer task, which should also preserve the content of the source image, CLIPstyler introduces a hybrid of losses, including a content loss measured in an external pre-trained VGG network. Conversely, we propose an alternative direction, which avoids the need to tune the coefficients of different loss functions so as to keep the balance between style and content. Specifically, we introduce a two-stage training paradigm: 1) a self-supervised pretraining stage, where the model learns to add semantically coherent image details from a low-resolution image to a high-resolution image; and 2) a style-specific fine-tuning stage, where the model learns to incorporate style into the high-resolution image. Since the fine-tuning phase is built on top of the first stage, the translator is able to keep the semantic consistency with respect to the input image as learned during pretraining. Moreover, we create a textual prompt by concatenating both the style and the textual description of the image content (i.e., its caption). This way, the prompt simultaneously models both the target appearance (i.e., the style description) and the image content which should be preserved by the translation process. Finally, since our translator's output is discrete and there is no ground-truth tokens for a stylized image, we introduce a Reinforcement Learning (RL) approach to fine-tune the translator using a reward based on the CLIP similarity be" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 744, + 298, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 744, + 298, + 752 + ], + "spans": [ + { + "bbox": [ + 279, + 744, + 298, + 752 + ], + "type": "text", + "content": "7602" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 50, + 111, + 279, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 111, + 279, + 146 + ], + "spans": [ + { + "bbox": [ + 50, + 111, + 279, + 146 + ], + "type": "text", + "content": "tween the stylized image and textual prompt, enabling the model to explore the answers in the latent space of a pretrained vector-quantized tokenizer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 147, + 281, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 147, + 281, + 260 + ], + "spans": [ + { + "bbox": [ + 50, + 147, + 281, + 260 + ], + "type": "text", + "content": "We call our network StylerDALLE and we show that it can generate stylized images driven by different types of language guidance. Compared with previous language-guided and reference image-based transfer methods, our generated images are less inclined to produce artifacts or semantic errors. Moreover, they can capture abstract concepts related to the target style (e.g., the typical brushstrokes of the artist) besides low-level features like texture and colors. We illustrate the effectiveness of our method through qualitative results, quantitative results, and a user study." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 262, + 220, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 220, + 273 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 220, + 273 + ], + "type": "text", + "content": "To conclude, our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 282, + 280, + 475 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 61, + 282, + 279, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 282, + 279, + 329 + ], + "spans": [ + { + "bbox": [ + 61, + 282, + 279, + 329 + ], + "type": "text", + "content": "- We propose StylerDALLE, a language-guided style transfer method that manipulates the discrete latent space of a pretrained vector-quantized tokenizer using a token sequence translation approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 338, + 280, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 338, + 280, + 374 + ], + "spans": [ + { + "bbox": [ + 61, + 338, + 280, + 374 + ], + "type": "text", + "content": "- We propose a non-autoregressive translation network that translates a low-resolution content image into a full-resolution image with style-specific details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 384, + 279, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 384, + 279, + 418 + ], + "spans": [ + { + "bbox": [ + 61, + 384, + 279, + 418 + ], + "type": "text", + "content": "- We propose a two-stage training procedure, including an RL strategy to ensure stylization and content preservation using CLIP-based language supervision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 428, + 279, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 428, + 279, + 475 + ], + "spans": [ + { + "bbox": [ + 61, + 428, + 279, + 475 + ], + "type": "text", + "content": "- Experimental results show that StylerDALLE can effectively transfer abstract style concepts, going beyond simple texture and color features while simultaneously preserving the semantic content of the translated scene." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 51, + 486, + 133, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 486, + 133, + 498 + ], + "spans": [ + { + "bbox": [ + 51, + 486, + 133, + 498 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 50, + 506, + 281, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 506, + 281, + 724 + ], + "spans": [ + { + "bbox": [ + 50, + 506, + 281, + 724 + ], + "type": "text", + "content": "Reference Image-Based Style Transfer. Gatys et al. [14] propose a neural style transfer method in which a pretrained CNN is used to extract content and style information from images, and to transfer the latter from an image to another. Following this pioneering work, a lot of interests have been attracted, with different methods focusing on different aspects of the topic, such as, e.g., diversified style transfer [37, 41], or attention mechanisms to fuse style and content [44, 28, 24]. A specific line of work focuses on artistic style transfer. For instance, Chen et al. [3] propose to use internal-external learning and contrastive learning with GANs to bridge the gap between human artworks and AI-created artworks. Wang et al. [40] introduce an aesthetic discriminator trained with a large corpus of human-created artworks. Other works train GANs using an artist-specific collection of images [36, 19, 4]. In contrast, we use the generic visual-language semantics embedded in the large-scale pretrained vector-quantized tokenizer and CLIP to avoid collecting style or artist-specific datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 111, + 527, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 111, + 527, + 284 + ], + "spans": [ + { + "bbox": [ + 296, + 111, + 527, + 284 + ], + "type": "text", + "content": "Language-Guided Style Transfer. Very recently a few works have proposed transferring methods conditioned on a textual description of the style. For instance, Fu et al. [12] use contrastive learning to train a GAN for artistic style transfer, but they adopt descriptive language instructions rather than more abstract style concepts. Gal et al. [13] use the CLIP space for a domain adaptation of a pre-trained StyleGAN [17]. The method closest to our approach is CLIPStyler [20], where a patch-wise CLIP loss is used to train a U-Net [34]. However, to condition the style change while preserving the image content, CLIPStyler uses hybrid losses and a rejection threshold, introducing many hyperparameters. In our method, we only use CLIP-based language supervision to ensure the style and content, saving the efforts of designing losses and tuning hyperparameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 299, + 527, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 299, + 527, + 437 + ], + "spans": [ + { + "bbox": [ + 296, + 299, + 527, + 437 + ], + "type": "text", + "content": "Large-scale Text-to-Image Generation Models. Recently, text-to-image models trained with large or very large scale datasets [32, 31, 35, 45, 1] have attracted tremendous attention because of their excellent performance in generating high-quality images starting from a textual query. Inspired by their strong ability to synthesize various types of images, we study how to use large-scale text-to-image models for style transfer. Specifically, we focus on transformer-based methods, which are potentially less time-consuming as compared to diffusion model-based methods [26]. In addition, we propose to use a non-autoregressive transformer that can generate tokens in parallel." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 448, + 373, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 448, + 373, + 460 + ], + "spans": [ + { + "bbox": [ + 297, + 448, + 373, + 460 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "spans": [ + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": "Vector-Quantzied Image Tokenizer. Transformer-based text-to-image generative models [32, 10, 11, 45, 1] rely on a vector-quantized image tokenizer to produce a discrete representation of the images, e.g., DALL-E [32] has a dVAE [33] and PARTI [45] has a ViT-VQGAN. Despite the differences in implementing the tokenizers, the effects are the same. In more detail, an image " + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": " is transformed into a " + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "inline_equation", + "content": "k \\times k" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": " grid of tokens " + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "inline_equation", + "content": "X = \\{x_{i,j}\\}_{i,j=1,\\dots,k}" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": ", using an encoder " + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": ". Each token " + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "inline_equation", + "content": "x_{i,j} \\in X = E(I)" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": " is an index of a codebook of embeddings (" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "inline_equation", + "content": "C = \\{e_1, \\ldots, e_M\\}" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "inline_equation", + "content": "1 \\leq x_{i,j} \\leq M" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": "), built during the training, and corresponds to a patch in " + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": ". A decoder " + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": " takes as input a grid of embeddings and reconstructs the original image: " + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "inline_equation", + "content": "\\tilde{I} = G(\\{e_{x_{i,j}}\\}_{i,j=1,\\dots,k})" + }, + { + "bbox": [ + 296, + 467, + 527, + 685 + ], + "type": "text", + "content": ". Training in the transformer-based text-to-image generative models consists of two stages. The first stage is dedicated to training the image tokenizer, while in the second phase, a transformer is used to learn a prior distribution over the text and the image tokens. In StylerDALLE (Sec. 4) we only use the pretrained image tokenizer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 700, + 526, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 700, + 526, + 724 + ], + "spans": [ + { + "bbox": [ + 296, + 700, + 526, + 724 + ], + "type": "text", + "content": "Non-Autoregressive Transformer (NAT). Gu et al. [15] propose a NAT for natural language translation, which con" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 743, + 298, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 743, + 298, + 753 + ], + "spans": [ + { + "bbox": [ + 279, + 743, + 298, + 753 + ], + "type": "text", + "content": "7603" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 107, + 345, + 318 + ], + "blocks": [ + { + "bbox": [ + 58, + 107, + 345, + 318 + ], + "lines": [ + { + "bbox": [ + 58, + 107, + 345, + 318 + ], + "spans": [ + { + "bbox": [ + 58, + 107, + 345, + 318 + ], + "type": "image", + "image_path": "c5af19da210c221852064406a7a8df89901eb20e8449b8810e17d021d9206e3a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 333, + 528, + 417 + ], + "lines": [ + { + "bbox": [ + 49, + 333, + 528, + 417 + ], + "spans": [ + { + "bbox": [ + 49, + 333, + 528, + 417 + ], + "type": "text", + "content": "Figure 2: A schematic illustration of StylerDALLE. We propose a NAT to translate tokens of a half-resolution content image to tokens of a full-resolution stylized image. StylerDALLE consists of three parts: 1) image tokenization preprocessing, to obtain the discrete visual tokens from a pretrained image tokenizer; 2) self-supervised pretaining, to train the NAT to predict full-resolution content images from half-resolution content images; and 3) style-specific fine-tuning, to let the model adding style details in the up-scaling prediction via Reinforcement Learning with language-prompted CLIP guidance. Specifically, in our NAT, the decoder input is duplicated from the tokens of the half-resolution image, following the natural positional correspondence (as highlighted by orange rectangles in the left part) with the full-resolution image." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 351, + 108, + 521, + 317 + ], + "blocks": [ + { + "bbox": [ + 351, + 108, + 521, + 317 + ], + "lines": [ + { + "bbox": [ + 351, + 108, + 521, + 317 + ], + "spans": [ + { + "bbox": [ + 351, + 108, + 521, + 317 + ], + "type": "image", + "image_path": "436d81afcc601f3abae24e8c55599a0127c0a64965102070dbdba812dbaf3a1e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "spans": [ + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "text", + "content": "sits of an encoder and a decoder. The encoder takes as input a source sentence " + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "inline_equation", + "content": "X = \\{x_{1},\\dots,x_{N^{\\prime}}\\}" + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "inline_equation", + "content": "N^{\\prime}" + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "text", + "content": " tokens and outputs a distribution over possible output sentences " + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "inline_equation", + "content": "Y = \\{y_{1},\\dots,y_{N}\\}" + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "text", + "content": " is the translation of " + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "text", + "content": " in the target language and, usually, " + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "inline_equation", + "content": "N\\neq N^{\\prime}" + }, + { + "bbox": [ + 49, + 433, + 282, + 525 + ], + "type": "text", + "content": ". The main novelty of NAT with respect to AT is that, during training, NAT uses a conditional independent factorization for the target sentence and the following log-likelihood:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 528, + 279, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 528, + 279, + 560 + ], + "spans": [ + { + "bbox": [ + 102, + 528, + 279, + 560 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = \\sum_ {n = 1} ^ {N} \\log p \\left(y _ {n} \\mid x _ {1: N ^ {\\prime}}; \\theta\\right), \\tag {1}", + "image_path": "17eb33523dd45fdf79a09d50014aad8ecd0e87a3aa7a42550e7a3645b9d92f35.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "spans": [ + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "text", + "content": "which differs from the common AT factorization in which the prediction of the " + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "text", + "content": "-th token " + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "inline_equation", + "content": "(y_{n})" + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "text", + "content": " depends on the previously predicted tokens " + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "inline_equation", + "content": "(p(y_{n}|y_{0:n-1}, x_{1:N'}; \\theta))" + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "text", + "content": ". This conditional independence assumption makes it possible a parallel generation of " + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "text", + "content": " at inference time, largely accelerating the translation time with respect to AT models. Importantly, to make parallel generation possible, the encoder input " + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "inline_equation", + "content": "(X)" + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "text", + "content": " is provided as input to the decoder as well, and individual tokens " + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "inline_equation", + "content": "(x_{n} \\in X)" + }, + { + "bbox": [ + 49, + 563, + 281, + 724 + ], + "type": "text", + "content": " can be copied zero or more times, with the number of times each input is copied depending on a specific predicted \"fertility\" value. As we will see in Sec. 4.1, we do not need to predict fertilities because, in our case, the cardinality of the input copies is fixed and determined by the up-scaling task we use for the image translation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 297, + 431, + 351, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 431, + 351, + 444 + ], + "spans": [ + { + "bbox": [ + 297, + 431, + 351, + 444 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "spans": [ + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": "The language-guided style transfer task can be described as follows. Given an image " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": ", we want to generate a new image " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "I^s" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " which preserves the semantic content of " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " but changes its appearance according to a textual style description " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " (e.g., \"cartoon\"). In StylerDALLE, we formulate this task as a visual-token based translation, from tokens of a content image to tokens of a stylized image. Specifically, given a content image " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " we first downsample " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " to get a half-resolution image " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " is fed to the tokenization encoder (Sec. 3) which extracts a discrete grid of " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "k \\times k" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " source tokens " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "X' = E(I')" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " can now be \"translated\" into a target (discrete) representation " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " is a grid at the original resolution " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "(2k \\times 2k)" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "\\hat{Y} = f(X')" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": ", being " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " the translation function. Finally, " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " is fed into the tokenization decoder obtaining the stylized image " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "I^s = G(\\hat{Y})" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": ". In the following subsections, we describe the architecture of " + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 296, + 453, + 527, + 648 + ], + "type": "text", + "content": " and the way in which it is trained." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 658, + 522, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 658, + 522, + 671 + ], + "spans": [ + { + "bbox": [ + 296, + 658, + 522, + 671 + ], + "type": "text", + "content": "4.1. Architecture and self-supervised pre-training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 677, + 527, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 677, + 527, + 724 + ], + "spans": [ + { + "bbox": [ + 296, + 677, + 527, + 724 + ], + "type": "text", + "content": "For our translation network " + }, + { + "bbox": [ + 296, + 677, + 527, + 724 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 296, + 677, + 527, + 724 + ], + "type": "text", + "content": " we use a NAT architecture [15] (Sec. 3) which we train from scratch using a self-supervised learning pretext task consisting in predicting the image at full resolution. Specifically, given the downsam-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 743, + 299, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 743, + 299, + 753 + ], + "spans": [ + { + "bbox": [ + 279, + 743, + 299, + 753 + ], + "type": "text", + "content": "7604" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "spans": [ + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": "pled image " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " and its corresponding grid of tokens " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "X' = E(I')" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": ", we use the indexes in " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " to extract the corresponding embeddings from " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " (Sec. 3). For each " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "x_{i,j} \\in X'" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "\\pmb{e}_{x_{i,j}}" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " be the corresponding embedding in " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "N' = k^2" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": ". The so obtained set of embeddings is flattened into a sequence, and, for each element " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "\\pmb{e}_n" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " of the sequence (" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "1 \\leq n \\leq N'" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": "), we add an absolute positional embedding [38] " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "\\pmb{p}_n" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "\\pmb{p}_n" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " has the same dimension as " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "\\pmb{e}_n" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "\\pmb{v}_n = \\pmb{e}_n + \\pmb{p}_n" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": ". The final sequence " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "V' = \\{\\pmb{v}_1, \\dots, \\pmb{v}_n, \\dots, \\pmb{v}_{N'}\\}" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " is input to the encoder of " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": ". Note that an alternative solution is to directly fed " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " with (a flattened version of) " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " learn its own initial token embedding. However, using the embeddings in " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " has the advantage of exploiting the image representation of the pre-trained image tokenizer. Moreover, from the original image " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " we extract the ground truth " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "X = E(I)" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": ", which is flattened in a sequence of " + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "inline_equation", + "content": "N = 4k^2" + }, + { + "bbox": [ + 50, + 111, + 281, + 294 + ], + "type": "text", + "content": " tokens." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "spans": [ + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "content": "Finally, following [15], we build a second sequence of input embeddings " + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "content": ", with cardinality " + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "content": ", which is fed to the decoder of " + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "content": ". As mentioned in Sec. 3, differently from NAT, we do not predict fertility values. Instead, we get the input of the decoder by simply replicating each element " + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "inline_equation", + "content": "x_{i,j} \\in X'" + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "content": " according to the positional correspondences between the low-resolution image and high-resolution image (as in Fig. 2). The rationale behind this choice is that " + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "content": " is trained to predict the full-resolution image, and each encoder input " + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "inline_equation", + "content": "(\\pmb{e}_{x_{i,j}})" + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "content": " corresponds to a patch in the subsampled image " + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "content": " and to 4 patches in the full-resolution image " + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "content": ". Thus, initializing the decoder with 4 replicas of each source-image patch initial embedding provides a coarse-grained signal for the upsampling task. Similar to before, the embeddings extracted from " + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 50, + 295, + 281, + 477 + ], + "type": "text", + "content": " are then flattened and added with a new positional encoding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "spans": [ + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": "Both the encoder and the decoder have self-attention layers and no causal masking is used. However, following [15], in the decoder, we mask out each query position " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "(n)" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": " only from attending to itself. Using " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "V^{\\prime}" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": " generates " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": " parallel posterior distributions over the visual vocabulary " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "(\\{1,\\dots,M\\})" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "P = f_{\\theta}(V',V)" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "N\\times M" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": " matrix, " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "P_{n}\\in [0,1]^{M}" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "P_{n}[y] = p_{\\theta}(Y_{n} = y|V^{\\prime},V)" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": ". Using " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "Y = \\{y_{1},\\ldots ,y_{n},\\ldots ,y_{N}\\}" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 50, + 477, + 280, + 570 + ], + "type": "text", + "content": " is trained to maximize:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 579, + 279, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 579, + 279, + 610 + ], + "spans": [ + { + "bbox": [ + 100, + 579, + 279, + 610 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r e - t r a i n}} (\\theta) = \\sum_ {n = 1} ^ {N} \\log P _ {n} [ y _ {n} ]. \\tag {2}", + "image_path": "c9ca9c7cebaddc721d5165361cff15a12056d1517659b121f29b808bddf407ec.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 50, + 616, + 279, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 616, + 279, + 675 + ], + "spans": [ + { + "bbox": [ + 50, + 616, + 279, + 675 + ], + "type": "text", + "content": "This pre-training stage is independent of the target style and it can be shared over different styles. After this stage, " + }, + { + "bbox": [ + 50, + 616, + 279, + 675 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 50, + 616, + 279, + 675 + ], + "type": "text", + "content": " is able to generate realistic low-level details (which are missing in " + }, + { + "bbox": [ + 50, + 616, + 279, + 675 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 50, + 616, + 279, + 675 + ], + "type": "text", + "content": "). In the next, we describe how a specific style is incorporated in " + }, + { + "bbox": [ + 50, + 616, + 279, + 675 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 50, + 616, + 279, + 675 + ], + "type": "text", + "content": " using a fine-tuning phase." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 682, + 183, + 695 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 682, + 183, + 695 + ], + "spans": [ + { + "bbox": [ + 50, + 682, + 183, + 695 + ], + "type": "text", + "content": "4.2. Style-specific fine-tuning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 50, + 700, + 279, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 700, + 279, + 724 + ], + "spans": [ + { + "bbox": [ + 50, + 700, + 279, + 724 + ], + "type": "text", + "content": "Given a style description provided with a textual sentence " + }, + { + "bbox": [ + 50, + 700, + 279, + 724 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 50, + 700, + 279, + 724 + ], + "type": "text", + "content": ", the goal is to fine-tune the pre-trained translator" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 111, + 527, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 111, + 527, + 250 + ], + "spans": [ + { + "bbox": [ + 296, + 111, + 527, + 250 + ], + "type": "text", + "content": "\\(f(\\cdot)\\) (Sec. 4.1) to make it generate image details in the style of \\(t_s\\). We fine-tune only the decoder of \\(f(\\cdot)\\), keeping frozen the encoder. To ensure both stylization and content preservation, we design a prompt that consists of two parts, i.e., a style description \\(t_s\\) and an image caption \\(t_a\\), which describe the image content and can be obtained from a generic image captioning dataset. For instance, given \\(t_a = \"A man's hand is adjusting his black tie\" and \\(t_s = \"cartoon\"\\), we obtain \\(t_q = \"a cartoon of a man's hand is adjusting his black tie.\\). On the other hand, in order to represent the image generated by \\(f(\\cdot)\\), we first need to sample the distributions in \\(P\\) (Sec. 4.1), and we do so using multinomial sampling:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 270, + 525, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 270, + 525, + 285 + ], + "spans": [ + { + "bbox": [ + 325, + 270, + 525, + 285 + ], + "type": "interline_equation", + "content": "\\hat {Y} _ {n} = \\text {S a m p l i n g} \\left(P _ {n} [ y ]\\right), \\forall n \\in \\{1, \\dots , N \\}. \\tag {3}", + "image_path": "78a6c193358ea1b9c85c6421fa51b532cd54cb61420d05022db22d62e36ebd6d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 294, + 527, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 294, + 527, + 340 + ], + "spans": [ + { + "bbox": [ + 296, + 294, + 527, + 340 + ], + "type": "text", + "content": "The sampled sequence " + }, + { + "bbox": [ + 296, + 294, + 527, + 340 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 296, + 294, + 527, + 340 + ], + "type": "text", + "content": " is reshaped to a " + }, + { + "bbox": [ + 296, + 294, + 527, + 340 + ], + "type": "inline_equation", + "content": "2k\\times 2k" + }, + { + "bbox": [ + 296, + 294, + 527, + 340 + ], + "type": "text", + "content": " grid and fed to the image detokenizer to get the final image " + }, + { + "bbox": [ + 296, + 294, + 527, + 340 + ], + "type": "inline_equation", + "content": "I^s = G(\\hat{Y})" + }, + { + "bbox": [ + 296, + 294, + 527, + 340 + ], + "type": "text", + "content": ". Finally, using the CLIP visual and textual encoders we compute the cosine similarity on the CLIP space:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 366, + 354, + 525, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 354, + 525, + 367 + ], + "spans": [ + { + "bbox": [ + 366, + 354, + 525, + 367 + ], + "type": "interline_equation", + "content": "r = \\operatorname {S i m} _ {C L I P} \\left(I ^ {s}, t _ {q}\\right). \\tag {4}", + "image_path": "00b0a52d051c0d205595650d11c17edc305a106fd632097839a771fee3be76b4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 375, + 527, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 375, + 527, + 488 + ], + "spans": [ + { + "bbox": [ + 296, + 375, + 527, + 488 + ], + "type": "text", + "content": "However, directly using Eq. 4 as the fine-tuning objective function is not possible because Eq. 3 is not differentiable. In addition, since there is no ground truth for the tokens of stylized images we use RL to encourage the model to explore the answers in the latent space of a pretrained vector-quantized model. We use the REINFORCE algorithm [42] that updates the parameters of " + }, + { + "bbox": [ + 296, + 375, + 527, + 488 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 296, + 375, + 527, + 488 + ], + "type": "text", + "content": " using the CLIP-based reward " + }, + { + "bbox": [ + 296, + 375, + 527, + 488 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 296, + 375, + 527, + 488 + ], + "type": "text", + "content": ", which could keep awarding the model for achieving better-stylized results until the model reaches a limit. This leads to the gradient estimate:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 498, + 525, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 498, + 525, + 530 + ], + "spans": [ + { + "bbox": [ + 321, + 498, + 525, + 530 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta | _ {d}} \\mathcal {L} _ {\\text {f i n e - t u n e}} (\\theta | _ {d}) = \\sum_ {n = 1} ^ {N} r \\nabla_ {\\theta | _ {d}} \\log P _ {n} [ y _ {n} ], \\tag {5}", + "image_path": "cc038758b5cc963bc0d1c530054712e60fc0c6a8a043edd400cd0370a143a71e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 539, + 526, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 539, + 526, + 586 + ], + "spans": [ + { + "bbox": [ + 296, + 539, + 526, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 539, + 526, + 586 + ], + "type": "inline_equation", + "content": "\\theta|_d" + }, + { + "bbox": [ + 296, + 539, + 526, + 586 + ], + "type": "text", + "content": " indicates the parameters of the decoder only (we found fine-tuning both could lead to content loss). By maximizing Eq. 5 we encourage " + }, + { + "bbox": [ + 296, + 539, + 526, + 586 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 296, + 539, + 526, + 586 + ], + "type": "text", + "content": " to generate images having both the content and the style of the prompt " + }, + { + "bbox": [ + 296, + 539, + 526, + 586 + ], + "type": "inline_equation", + "content": "t_q" + }, + { + "bbox": [ + 296, + 539, + 526, + 586 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 587, + 527, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 587, + 527, + 724 + ], + "spans": [ + { + "bbox": [ + 296, + 587, + 527, + 724 + ], + "type": "text", + "content": "With respect to the whole method, the reason we use down-sampled versions of the content image is that \"style\" is commonly assumed to be involved in the low-level visual details, such as colors, texture, painting strokes, etc. " + }, + { + "bbox": [ + 296, + 587, + 527, + 724 + ], + "type": "inline_equation", + "content": "X^{\\prime}" + }, + { + "bbox": [ + 296, + 587, + 527, + 724 + ], + "type": "text", + "content": ", which in our formulation represents " + }, + { + "bbox": [ + 296, + 587, + 527, + 724 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 296, + 587, + 527, + 724 + ], + "type": "text", + "content": " at a lower resolution, presumably keeps most of the content in " + }, + { + "bbox": [ + 296, + 587, + 527, + 724 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 296, + 587, + 527, + 724 + ], + "type": "text", + "content": " discarding some details, this way facilitating the style translation process. Preliminary experiments (in Appendix A.2) in which we fed the encoder with tokenized full-resolution images led to poor results, demonstrating that the different cardinality between the source and the target sequence is an essential component in this translation process." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "spans": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "type": "text", + "content": "7605" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 110, + 527, + 380 + ], + "blocks": [ + { + "bbox": [ + 51, + 110, + 527, + 380 + ], + "lines": [ + { + "bbox": [ + 51, + 110, + 527, + 380 + ], + "spans": [ + { + "bbox": [ + 51, + 110, + 527, + 380 + ], + "type": "image", + "image_path": "b2252aab2985062c302386dad8df9e7dedf553e18ccb8c6db7e5fed30cdaecf5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 396, + 525, + 432 + ], + "lines": [ + { + "bbox": [ + 49, + 396, + 525, + 432 + ], + "spans": [ + { + "bbox": [ + 49, + 396, + 525, + 432 + ], + "type": "text", + "content": "Figure 3: Qualitative results of StylerDALLE-1 on various styles: (a) Monet Sunrise, (b) Picasso cubism, (c) Van Gogh blue color, (d) Van Gogh purple color, (e) warm and relaxing, (f) ukiyo-e print, (g) fauvism, (h) pixel art illustration. The style references at the top are for illustration only (not used as input to the model)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 450, + 128, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 450, + 128, + 463 + ], + "spans": [ + { + "bbox": [ + 50, + 450, + 128, + 463 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 50, + 471, + 145, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 471, + 145, + 483 + ], + "spans": [ + { + "bbox": [ + 50, + 471, + 145, + 483 + ], + "type": "text", + "content": "5.1. Training Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 492, + 279, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 492, + 279, + 584 + ], + "spans": [ + { + "bbox": [ + 49, + 492, + 279, + 584 + ], + "type": "text", + "content": "We implement our method based on two types of pretrained vector-quantized tokenizer: 1) the officially released dVAE of DALL-E 1 [27] and 2) the VQGAN of RuDALLE [39], and consequently obtain two groups of results, noting as StylerDALLE-1 and StylerDALLE-Ru, respectively. We use the two models in our experiment because they are open-sourced while our method is applicable to any large-scale pretrained vector-quantized tokenizer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 586, + 280, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 586, + 280, + 723 + ], + "spans": [ + { + "bbox": [ + 49, + 586, + 280, + 723 + ], + "type": "text", + "content": "To train the model, we use the MS-COCO [23] trainset, which contains 83k images of common objects in daily scenes. In the self-supervised pretraining stage, we only use the images while in the style-specific fine-tuning stage we use both images and captions. The StylerDALLE NAT model consists of a 4-layer encoder and an 8-layer decoder while the attention head number is 8 and the hidden dimension is 512. We use Pytorch [29] to implement our method. We train the NAT model for 25 epochs with a learning rate of 1e-4. We use Adam [18] optimizer. In the style-specific fine-tuning stage, we use both the images and the captions. In particular, we utilize all the caption annotations to en" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 451, + 526, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 451, + 526, + 565 + ], + "spans": [ + { + "bbox": [ + 296, + 451, + 526, + 565 + ], + "type": "text", + "content": "hance the model robustness, as usually human-beings annotate different captions of a single image. Notably, the caption annotations are only used at the fine-tuning stage. In other words, StylerDALLE does not need to use image caption as input at inference time. We only fine-tune the decoder of the NAT model, and keep the encoder frozen. We use Adam optimizer with a learning rate of 1e-6. For CLIP model, we use the CLIP ViT-B/32 model. For both training stages, the model is trained on a single RTX-A6000 GPU for 24 hours." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 297, + 574, + 415, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 574, + 415, + 587 + ], + "spans": [ + { + "bbox": [ + 297, + 574, + 415, + 587 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 592, + 526, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 592, + 526, + 672 + ], + "spans": [ + { + "bbox": [ + 296, + 592, + 526, + 672 + ], + "type": "text", + "content": "In the following, we present qualitative, quantitative, and user study results of our method and comparative methods, as well as comparisons with reference image-based methods. The additional implementation details, ablation study, inference time comparison, and additional experimental results are shown in Appendix A.1, A.2, A.3, and A.5, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 688, + 526, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 688, + 526, + 724 + ], + "spans": [ + { + "bbox": [ + 296, + 688, + 526, + 724 + ], + "type": "text", + "content": "Qualitative Results. In Fig. 1, Fig. 3 and Fig. 4, we show that our method can effectively transfer various types of styles, i.e., a) abstract art styles, e.g., \"fauvism\" and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 743, + 299, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 743, + 299, + 752 + ], + "spans": [ + { + "bbox": [ + 279, + 743, + 299, + 752 + ], + "type": "text", + "content": "7606" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 111, + 527, + 384 + ], + "blocks": [ + { + "bbox": [ + 51, + 111, + 527, + 384 + ], + "lines": [ + { + "bbox": [ + 51, + 111, + 527, + 384 + ], + "spans": [ + { + "bbox": [ + 51, + 111, + 527, + 384 + ], + "type": "image", + "image_path": "884181b992be00983c15537b69b7135cf17d675bc107ea1c870f94aed16dcd14.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 396, + 528, + 433 + ], + "lines": [ + { + "bbox": [ + 49, + 396, + 528, + 433 + ], + "spans": [ + { + "bbox": [ + 49, + 396, + 528, + 433 + ], + "type": "text", + "content": "Figure 4: Qualitative results of StylerDALLE-Ru on various styles: (a) Monet, (b) Monet Paris, (c) Monet Venice, (d) Van Gogh, (e) Van Gogh Irises, (f) Van Gogh Almond, (g) Van Gogh Starry Night, (h) Van Gogh Sunflowers. The style references at the top are for illustration only (not used as input to the model)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 442, + 530, + 555 + ], + "blocks": [ + { + "bbox": [ + 51, + 442, + 530, + 555 + ], + "lines": [ + { + "bbox": [ + 51, + 442, + 530, + 555 + ], + "spans": [ + { + "bbox": [ + 51, + 442, + 530, + 555 + ], + "type": "table", + "html": "
DatasetStyle MethodFauvismMonetMonet SunriseMonetSunsetMonetParisVan Gogh IrisesVan GoghAlmond Blossoms
CoCoCLIPstyler-Fast26.1127.4624.7826.7027.6026.2327.05
Ours30.6127.7228.8130.6829.6328.1431.64
AFHQCLIPstyler-Fast26.5226.0823.2725.0525.4223.2824.31
Ours29.9626.5526.5829.1126.1028.0431.92
ImageNet-100CLIPstyler-Fast26.6327.1624.3326.1227.2826.3226.23
Ours30.2227.5328.2930.2827.6728.2631.60
", + "image_path": "c8db4ab767de157313dd3a3e3cfa48893f03929022444ee0412fab53749b0d58.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 563, + 473, + 575 + ], + "lines": [ + { + "bbox": [ + 102, + 563, + 473, + 575 + ], + "spans": [ + { + "bbox": [ + 102, + 563, + 473, + 575 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons with CLIPstyler [20] on different types of styles and datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 49, + 595, + 280, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 595, + 280, + 721 + ], + "spans": [ + { + "bbox": [ + 49, + 595, + 280, + 721 + ], + "type": "text", + "content": "\"pop art\"; b) artist-specific styles, e.g., \"Monet\" and \"Van Gogh\"; c) artist-specific styles with additional descriptions, e.g., \"Monet Paris\" and \"Van Gogh Sunflowers\"; d) artistic painting types, e.g., \"pixel art illustration\"; and e) emotional effects, e.g., \"warm and relaxing\". According to the qualitative results, we draw the following conclusions: 1) StylerDALLE can transfer abstract style concepts which go beyond the texture and color features and are similar to the typical trait of the artist/artistic target style; 2) each style corresponds to generated images that are different from those of other styles; 3) the image content is well preserved;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 296, + 595, + 526, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 595, + 526, + 629 + ], + "spans": [ + { + "bbox": [ + 296, + 595, + 526, + 629 + ], + "type": "text", + "content": "and 4) StylerDALLE can be applied to open-domain content images (i.e., the image content can contain animals, human beings, daily objects, buildings, etc.)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 631, + 527, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 631, + 527, + 724 + ], + "spans": [ + { + "bbox": [ + 296, + 631, + 527, + 724 + ], + "type": "text", + "content": "Further, we compare StylerDALLE with the recent language-guided style transfer method CLIPStyler [20]. CLIPstyler proposes two methods: 1) CLIPStyler-Optimization, which optimizes a style transfer network for each content image, thus being time-consuming; and 2) CLIPStyler-Fast, which is the most comparable method to ours as it trains a network for each style, and then it can be used with any content images. As shown in Fig. 5," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "spans": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "type": "text", + "content": "7607" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 112, + 527, + 355 + ], + "blocks": [ + { + "bbox": [ + 54, + 112, + 527, + 355 + ], + "lines": [ + { + "bbox": [ + 54, + 112, + 527, + 355 + ], + "spans": [ + { + "bbox": [ + 54, + 112, + 527, + 355 + ], + "type": "image", + "image_path": "d1e0263c65c1e8f6b7567708a4d3d752fef37a134ab05752768de1a0e77f6f6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 369, + 526, + 405 + ], + "lines": [ + { + "bbox": [ + 49, + 369, + 526, + 405 + ], + "spans": [ + { + "bbox": [ + 49, + 369, + 526, + 405 + ], + "type": "text", + "content": "Figure 5: Comparisons with language-guided methods. The styles are (a) fauxivism, (b) Monet, (c) Monet Impression Sunrise, (d) oil painting, (e) watercolor painting, (f) Van Gogh, (g) Van Gogh blue color. The style references at the top are for illustration only (not used as input to the model)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 423, + 281, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 423, + 281, + 574 + ], + "spans": [ + { + "bbox": [ + 49, + 423, + 281, + 574 + ], + "type": "text", + "content": "CLIPStyler-Optimization generates diverse stylized results but it suffers from inharmonious artifacts. For instance, in the column \"Monet\", there are many plants on the train. In addition, CLIPStyler-Optimization has the problems of writing the style text in the results, as in the results of \"fauvism\" and so on. On the other hand, the images generated by CLIPStyler-Fast do not contain artifacts but there is less variation among different styles. Importantly, it is hard to recognize the typical trait of each artistic style, and the main differences among the styles are the colors. In contrast, the results of StylerDALLE are much closer to the artworks of the specific artistic style, they show distinct differences among different styles, and they have no artifact issues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 586, + 281, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 586, + 281, + 724 + ], + "spans": [ + { + "bbox": [ + 49, + 586, + 281, + 724 + ], + "type": "text", + "content": "Quantitative Results. To do quantitative analysis, we use the CLIP similarity score, formalized as score " + }, + { + "bbox": [ + 49, + 586, + 281, + 724 + ], + "type": "inline_equation", + "content": "= \\operatorname{Sim}_{\\text{CLIP}}(I, t_s)" + }, + { + "bbox": [ + 49, + 586, + 281, + 724 + ], + "type": "text", + "content": ", which is computed between generated stylized images and textual description of the target style. In Tab. 1, we present the results of StylerDALLE-1 and CLIPstyler-Fast. Since both methods are applicable for arbitrary content images, we use the MS-COCO val-set, AFHQ val-set [6], and ImageNet-100 val-set [7] for evaluation. According to the results on multiple datasets, although the way we use CLIP is to compute rewards in Reinforcement Learning, instead of directly using CLIP-scores to optimize the network as CLIPstyler, we achieve comparable" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 303, + 423, + 522, + 469 + ], + "blocks": [ + { + "bbox": [ + 303, + 423, + 522, + 469 + ], + "lines": [ + { + "bbox": [ + 303, + 423, + 522, + 469 + ], + "spans": [ + { + "bbox": [ + 303, + 423, + 522, + 469 + ], + "type": "table", + "html": "
(%)CLIPstyler-FastStyler DALLE-1Styler DALLE-Ru
Preference15.9015.3368.76
", + "image_path": "eb187c5d3c45803654fc7c732ffd43c68f64d13b571ed961c9260c64463903e6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 333, + 476, + 490, + 489 + ], + "lines": [ + { + "bbox": [ + 333, + 476, + 490, + 489 + ], + "spans": [ + { + "bbox": [ + 333, + 476, + 490, + 489 + ], + "type": "text", + "content": "Table 2: Preference scores of user study." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 511, + 526, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 511, + 526, + 533 + ], + "spans": [ + { + "bbox": [ + 296, + 511, + 526, + 533 + ], + "type": "text", + "content": "and even better quantitative results, indicating the effectiveness of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 551, + 527, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 551, + 527, + 713 + ], + "spans": [ + { + "bbox": [ + 296, + 551, + 527, + 713 + ], + "type": "text", + "content": "User Study. We conduct a user study to see human opinions towards the stylized images coming from different methods: 1) CLIPstyler-Fast, 2) StylerDALLE-1, and 3) StylerDALLE-Ru. In specific, we collect opinions from 35 human subjects with a 30-question questionnaire. In each question, we ask them to select one stylized result that is closest to a target style. As in Tab. 2, among the three methods, StylerDALLE-Ru achieves the highest preference score of " + }, + { + "bbox": [ + 296, + 551, + 527, + 713 + ], + "type": "inline_equation", + "content": "68.76\\%" + }, + { + "bbox": [ + 296, + 551, + 527, + 713 + ], + "type": "text", + "content": ", indicating the superiority of our method. We also find that humans prefer StylerDALLE-Ru much more than StylerDALLE-1. We infer this could be because the results from StylerDALLE-1 are blurry, as they are based on the dVAE of DALLE-1, and humans dislike blurry images. More user study details are given in A.4." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 743, + 299, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 743, + 299, + 753 + ], + "spans": [ + { + "bbox": [ + 279, + 743, + 299, + 753 + ], + "type": "text", + "content": "7608" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 110, + 123, + 185 + ], + "blocks": [ + { + "bbox": [ + 54, + 110, + 123, + 185 + ], + "lines": [ + { + "bbox": [ + 54, + 110, + 123, + 185 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 123, + 185 + ], + "type": "image", + "image_path": "1b0b54a2c82d79c28f263e9f39ece067161b0b3b0925f56875a2d1ef1d306fa1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 186, + 95, + 194 + ], + "lines": [ + { + "bbox": [ + 82, + 186, + 95, + 194 + ], + "spans": [ + { + "bbox": [ + 82, + 186, + 95, + 194 + ], + "type": "text", + "content": "ref1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 54, + 196, + 122, + 264 + ], + "blocks": [ + { + "bbox": [ + 54, + 196, + 122, + 264 + ], + "lines": [ + { + "bbox": [ + 54, + 196, + 122, + 264 + ], + "spans": [ + { + "bbox": [ + 54, + 196, + 122, + 264 + ], + "type": "image", + "image_path": "e0feb0de3e0c2c5a9fc9bb933872ecdc0322323bc03da03a88d3dec12315f836.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 267, + 92, + 275 + ], + "lines": [ + { + "bbox": [ + 79, + 267, + 92, + 275 + ], + "spans": [ + { + "bbox": [ + 79, + 267, + 92, + 275 + ], + "type": "text", + "content": "ref2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 124, + 110, + 180, + 166 + ], + "blocks": [ + { + "bbox": [ + 124, + 110, + 180, + 166 + ], + "lines": [ + { + "bbox": [ + 124, + 110, + 180, + 166 + ], + "spans": [ + { + "bbox": [ + 124, + 110, + 180, + 166 + ], + "type": "image", + "image_path": "11eda235384a8422a2baa37953a7feeb2b9d6d084033a4f4a954c2e52e8fcff7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 124, + 168, + 180, + 223 + ], + "blocks": [ + { + "bbox": [ + 124, + 168, + 180, + 223 + ], + "lines": [ + { + "bbox": [ + 124, + 168, + 180, + 223 + ], + "spans": [ + { + "bbox": [ + 124, + 168, + 180, + 223 + ], + "type": "image", + "image_path": "0cfb287d26fbaa9be473d4cd9ef793f0dd192c2fd6a7000d64dc693d8df35ba4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 124, + 226, + 180, + 281 + ], + "blocks": [ + { + "bbox": [ + 124, + 226, + 180, + 281 + ], + "lines": [ + { + "bbox": [ + 124, + 226, + 180, + 281 + ], + "spans": [ + { + "bbox": [ + 124, + 226, + 180, + 281 + ], + "type": "image", + "image_path": "994091be88005fc1d17d99a40e613dc5fd853ac1d06a9e6e68e7895d72cbbece.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 283, + 165, + 290 + ], + "lines": [ + { + "bbox": [ + 139, + 283, + 165, + 290 + ], + "spans": [ + { + "bbox": [ + 139, + 283, + 165, + 290 + ], + "type": "text", + "content": "Content" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 183, + 110, + 239, + 166 + ], + "blocks": [ + { + "bbox": [ + 183, + 110, + 239, + 166 + ], + "lines": [ + { + "bbox": [ + 183, + 110, + 239, + 166 + ], + "spans": [ + { + "bbox": [ + 183, + 110, + 239, + 166 + ], + "type": "image", + "image_path": "e6ca4c8b9c8e6178a471eae45b3e40457c4a648a387228f14f85b4942ecf3526.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 183, + 168, + 239, + 223 + ], + "blocks": [ + { + "bbox": [ + 183, + 168, + 239, + 223 + ], + "lines": [ + { + "bbox": [ + 183, + 168, + 239, + 223 + ], + "spans": [ + { + "bbox": [ + 183, + 168, + 239, + 223 + ], + "type": "image", + "image_path": "171073ceedee604c21cb9354618de42ab77aa38e7773e7725cd8ee1da469a53f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 183, + 226, + 239, + 281 + ], + "blocks": [ + { + "bbox": [ + 183, + 226, + 239, + 281 + ], + "lines": [ + { + "bbox": [ + 183, + 226, + 239, + 281 + ], + "spans": [ + { + "bbox": [ + 183, + 226, + 239, + 281 + ], + "type": "image", + "image_path": "fe810f9f01a7f65ebf0e46ab7aba97d6fb463269591315c322d43b7268a45d71.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 283, + 223, + 290 + ], + "lines": [ + { + "bbox": [ + 199, + 283, + 223, + 290 + ], + "spans": [ + { + "bbox": [ + 199, + 283, + 223, + 290 + ], + "type": "text", + "content": "AesUST" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 199, + 292, + 223, + 300 + ], + "lines": [ + { + "bbox": [ + 199, + 292, + 223, + 300 + ], + "spans": [ + { + "bbox": [ + 199, + 292, + 223, + 300 + ], + "type": "text", + "content": "w/ ref1" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 240, + 110, + 296, + 166 + ], + "blocks": [ + { + "bbox": [ + 240, + 110, + 296, + 166 + ], + "lines": [ + { + "bbox": [ + 240, + 110, + 296, + 166 + ], + "spans": [ + { + "bbox": [ + 240, + 110, + 296, + 166 + ], + "type": "image", + "image_path": "083eaa866ced7f12ed3df653f4e1d9fb3b3db9b2364f7b42fbcf489249c32576.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 240, + 168, + 296, + 223 + ], + "blocks": [ + { + "bbox": [ + 240, + 168, + 296, + 223 + ], + "lines": [ + { + "bbox": [ + 240, + 168, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 240, + 168, + 296, + 223 + ], + "type": "image", + "image_path": "f617f7921ea53fcf2c29d1bcb1101fbccb2d150010651d1d1a18494b70548c59.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 240, + 226, + 296, + 281 + ], + "blocks": [ + { + "bbox": [ + 240, + 226, + 296, + 281 + ], + "lines": [ + { + "bbox": [ + 240, + 226, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 240, + 226, + 296, + 281 + ], + "type": "image", + "image_path": "40f346403ee108140b3b27570439c834d593212a0020279abf18303a7b604ab3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 283, + 279, + 290 + ], + "lines": [ + { + "bbox": [ + 258, + 283, + 279, + 290 + ], + "spans": [ + { + "bbox": [ + 258, + 283, + 279, + 290 + ], + "type": "text", + "content": "StyTr2" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 292, + 282, + 300 + ], + "lines": [ + { + "bbox": [ + 258, + 292, + 282, + 300 + ], + "spans": [ + { + "bbox": [ + 258, + 292, + 282, + 300 + ], + "type": "text", + "content": "w/ ref1" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 319, + 525, + 343 + ], + "lines": [ + { + "bbox": [ + 49, + 319, + 525, + 343 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 525, + 343 + ], + "type": "text", + "content": "Figure 6: Comparisons with reference image-based methods. StylerDALLE is able to transfer more abstract concepts, e.g., specific painting strokes, and is less likely to produce semantic errors." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 297, + 110, + 352, + 166 + ], + "blocks": [ + { + "bbox": [ + 297, + 110, + 352, + 166 + ], + "lines": [ + { + "bbox": [ + 297, + 110, + 352, + 166 + ], + "spans": [ + { + "bbox": [ + 297, + 110, + 352, + 166 + ], + "type": "image", + "image_path": "eceb251da3ec631aca88e5f5a832c52bbc04b8669e988b299757c7e6a5bc757a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 296, + 168, + 352, + 223 + ], + "blocks": [ + { + "bbox": [ + 296, + 168, + 352, + 223 + ], + "lines": [ + { + "bbox": [ + 296, + 168, + 352, + 223 + ], + "spans": [ + { + "bbox": [ + 296, + 168, + 352, + 223 + ], + "type": "image", + "image_path": "08588be497d109f3585502d97ae42489db351313a1abb3a91618fbe3b41416e5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 296, + 226, + 352, + 281 + ], + "blocks": [ + { + "bbox": [ + 296, + 226, + 352, + 281 + ], + "lines": [ + { + "bbox": [ + 296, + 226, + 352, + 281 + ], + "spans": [ + { + "bbox": [ + 296, + 226, + 352, + 281 + ], + "type": "image", + "image_path": "7cce431f6ace421c029a8ac4c53f3d9a7085851f4fee61bd17a0a5dd95fbdad1.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 283, + 348, + 290 + ], + "lines": [ + { + "bbox": [ + 299, + 283, + 348, + 290 + ], + "spans": [ + { + "bbox": [ + 299, + 283, + 348, + 290 + ], + "type": "text", + "content": "StylerDALLE-Ru" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 292, + 336, + 300 + ], + "lines": [ + { + "bbox": [ + 312, + 292, + 336, + 300 + ], + "spans": [ + { + "bbox": [ + 312, + 292, + 336, + 300 + ], + "type": "text", + "content": "w/text" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 355, + 110, + 411, + 166 + ], + "blocks": [ + { + "bbox": [ + 355, + 110, + 411, + 166 + ], + "lines": [ + { + "bbox": [ + 355, + 110, + 411, + 166 + ], + "spans": [ + { + "bbox": [ + 355, + 110, + 411, + 166 + ], + "type": "image", + "image_path": "03d4d90c86d7ba54403da3b313d5c42035601f253df708171efbc188ecec8a86.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 355, + 168, + 411, + 223 + ], + "blocks": [ + { + "bbox": [ + 355, + 168, + 411, + 223 + ], + "lines": [ + { + "bbox": [ + 355, + 168, + 411, + 223 + ], + "spans": [ + { + "bbox": [ + 355, + 168, + 411, + 223 + ], + "type": "image", + "image_path": "de1d49cdb238f4e342b06bf82fe71262a93e38298c1c7796b953e3a576e478ea.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 355, + 226, + 411, + 281 + ], + "blocks": [ + { + "bbox": [ + 355, + 226, + 411, + 281 + ], + "lines": [ + { + "bbox": [ + 355, + 226, + 411, + 281 + ], + "spans": [ + { + "bbox": [ + 355, + 226, + 411, + 281 + ], + "type": "image", + "image_path": "f1f0af7d616df3041b61a491d6e96740f43aa8c0bbe9eec698d07fdf9cd9a4ac.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 283, + 393, + 290 + ], + "lines": [ + { + "bbox": [ + 369, + 283, + 393, + 290 + ], + "spans": [ + { + "bbox": [ + 369, + 283, + 393, + 290 + ], + "type": "text", + "content": "AesUST" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 368, + 292, + 393, + 300 + ], + "lines": [ + { + "bbox": [ + 368, + 292, + 393, + 300 + ], + "spans": [ + { + "bbox": [ + 368, + 292, + 393, + 300 + ], + "type": "text", + "content": "w/ ref2" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 412, + 110, + 468, + 166 + ], + "blocks": [ + { + "bbox": [ + 412, + 110, + 468, + 166 + ], + "lines": [ + { + "bbox": [ + 412, + 110, + 468, + 166 + ], + "spans": [ + { + "bbox": [ + 412, + 110, + 468, + 166 + ], + "type": "image", + "image_path": "b2193636062281c30a974fb3cedccc7edb41700c1925a2901ebedbd724c8dff6.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 412, + 168, + 468, + 223 + ], + "blocks": [ + { + "bbox": [ + 412, + 168, + 468, + 223 + ], + "lines": [ + { + "bbox": [ + 412, + 168, + 468, + 223 + ], + "spans": [ + { + "bbox": [ + 412, + 168, + 468, + 223 + ], + "type": "image", + "image_path": "ada740624fc6aa034bcccc2b8e1b0292e5e93adc3bd7952c551978d8548ce841.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 412, + 226, + 468, + 281 + ], + "blocks": [ + { + "bbox": [ + 412, + 226, + 468, + 281 + ], + "lines": [ + { + "bbox": [ + 412, + 226, + 468, + 281 + ], + "spans": [ + { + "bbox": [ + 412, + 226, + 468, + 281 + ], + "type": "image", + "image_path": "93bdab9d2e0f0328ddadd12b72d3d78a2bef774be9289d852e884f7039892177.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 283, + 455, + 290 + ], + "lines": [ + { + "bbox": [ + 432, + 283, + 455, + 290 + ], + "spans": [ + { + "bbox": [ + 432, + 283, + 455, + 290 + ], + "type": "text", + "content": "StyTr2" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 292, + 456, + 300 + ], + "lines": [ + { + "bbox": [ + 432, + 292, + 456, + 300 + ], + "spans": [ + { + "bbox": [ + 432, + 292, + 456, + 300 + ], + "type": "text", + "content": "w/ ref2" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 468, + 110, + 524, + 166 + ], + "blocks": [ + { + "bbox": [ + 468, + 110, + 524, + 166 + ], + "lines": [ + { + "bbox": [ + 468, + 110, + 524, + 166 + ], + "spans": [ + { + "bbox": [ + 468, + 110, + 524, + 166 + ], + "type": "image", + "image_path": "28b4a7e5e88ec6621d5e9b520d5b4dbcfc00fef1ba8ff75e5e6b82c9ed600638.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 468, + 168, + 524, + 223 + ], + "blocks": [ + { + "bbox": [ + 468, + 168, + 524, + 223 + ], + "lines": [ + { + "bbox": [ + 468, + 168, + 524, + 223 + ], + "spans": [ + { + "bbox": [ + 468, + 168, + 524, + 223 + ], + "type": "image", + "image_path": "75d20ebfc8f91a3c0c2e52ed5b538bd47b3b4a69fd2d391e7c18d0fdf24a7b29.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 468, + 226, + 524, + 281 + ], + "blocks": [ + { + "bbox": [ + 468, + 226, + 524, + 281 + ], + "lines": [ + { + "bbox": [ + 468, + 226, + 524, + 281 + ], + "spans": [ + { + "bbox": [ + 468, + 226, + 524, + 281 + ], + "type": "image", + "image_path": "9711048784a74d21cd2af45b0fcb449172b1d3db1c6cbefcb7216463a5c56195.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 283, + 521, + 290 + ], + "lines": [ + { + "bbox": [ + 473, + 283, + 521, + 290 + ], + "spans": [ + { + "bbox": [ + 473, + 283, + 521, + 290 + ], + "type": "text", + "content": "StylerDALLE-Ru" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 487, + 292, + 510, + 300 + ], + "lines": [ + { + "bbox": [ + 487, + 292, + 510, + 300 + ], + "spans": [ + { + "bbox": [ + 487, + 292, + 510, + 300 + ], + "type": "text", + "content": "w/text" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 49, + 363, + 279, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 363, + 279, + 500 + ], + "spans": [ + { + "bbox": [ + 49, + 363, + 279, + 500 + ], + "type": "text", + "content": "Comparisons with Reference Image-Based Methods. We compare StylerDALLE with state-of-the-art reference image-based methods: 1) AesUST [40], an arbitrary style transfer method which enhances the aesthetic reality using a GAN trained with a collection of artworks; and 2) StyTr2 [9], an arbitrary style transfer method which uses a transformer to eliminate the biased content representation issues of CNN-based methods. To make the comparison feasible, we show the results of AesUST and StyTr2 using two Van Gogh paintings as the reference images, and the results of StylerDALLE trained using the corresponding language description." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 49, + 506, + 281, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 506, + 281, + 723 + ], + "spans": [ + { + "bbox": [ + 49, + 506, + 281, + 723 + ], + "type": "text", + "content": "As shown in Fig. 6, the results of StylerDALLE are distinctive from the results of reference image-based methods. In concrete, the results of both AesUST and StyleTr2 are mostly affected by the colors and the textures of the reference images, and to some extent in an unnatural way. For instance, on the bottom line, in the \"Van Gogh Irises\" stylized results of AesUST and StyleTr2, the textures of the irises get transferred to the orange cup. In transferring the style of \"Van Gogh Starry Night\", the objects in the results of AesUST and StyleTr2 are mostly in the same dark-sky blue as the reference image, making the scenes a bit in-realistic. By contrast, the colors and textures used in the reference style are well transferred while being appropriately applied to the contents, without changing the original semantics. Moreover, StylerDALLE achieves to transfer higher-level style features, e.g., the brushstroke, rather than merely colors and textures, leading to the results of a similar style to the target one. For example, the strokes in Starry Night are sharper as compared to the ones in Irises, and the differences are also" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 297, + 363, + 495, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 363, + 495, + 375 + ], + "spans": [ + { + "bbox": [ + 297, + 363, + 495, + 375 + ], + "type": "text", + "content": "presented in the stylized results of StylerDALLE." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 297, + 383, + 367, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 383, + 367, + 394 + ], + "spans": [ + { + "bbox": [ + 297, + 383, + 367, + 394 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 296, + 401, + 526, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 401, + 526, + 585 + ], + "spans": [ + { + "bbox": [ + 296, + 401, + 526, + 585 + ], + "type": "text", + "content": "We present StylerDALLE, a language-guided style transfer method that leverages the power of the large-scale pretrained vector-quantized image tokenizer and CLIP. Specifically, inspired by natural language translation, we propose a non-autoregressive sequence translation approach to manipulate the discrete visual tokens, from the content image to the stylized image. We use Reinforcement Learning to include CLIP-based language supervision on the style and content. Differently from previous work, StylerDALLE can transfer abstract style concepts that are implicitly represented in the pretrained image tokenizer and CLIP and which cannot be easily obtained using reference images. Moreover, using the large-scale pretrained latent space as the basic image representation makes it possible to reduce the artifacts and the semantic incoherence better than the previous work that operates at the pixel level." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 296, + 597, + 525, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 597, + 525, + 643 + ], + "spans": [ + { + "bbox": [ + 296, + 597, + 525, + 643 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by the MUR PNRR project FAIR (PE00000013) funded by the NextGenerationEU and by the PRIN project CREATIVE (Prot. 2020ZSL9F9)." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 298, + 651, + 354, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 651, + 354, + 663 + ], + "spans": [ + { + "bbox": [ + 298, + 651, + 354, + 663 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 302, + 670, + 526, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 670, + 526, + 723 + ], + "spans": [ + { + "bbox": [ + 302, + 670, + 526, + 723 + ], + "type": "text", + "content": "[1] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 3, 12" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "spans": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "type": "text", + "content": "7609" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 52, + 112, + 280, + 723 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 55, + 112, + 279, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 112, + 279, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 112, + 279, + 164 + ], + "type": "text", + "content": "[2] Haibo Chen, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, Dongming Lu, et al. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems, 34:26561-26573, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 167, + 280, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 280, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 280, + 228 + ], + "type": "text", + "content": "[3] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 872-881, June 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 230, + 280, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 280, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 280, + 282 + ], + "type": "text", + "content": "[4] Haibo Chen, Lei Zhao, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Dualist: Dual style-learning networks for artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 872-881, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 284, + 280, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 284, + 280, + 336 + ], + "spans": [ + { + "bbox": [ + 57, + 284, + 280, + 336 + ], + "type": "text", + "content": "[5] Jiaxin Cheng, Ayush Jaiswal, Yue Wu, Pradeep Natarajan, and Prem Natarajan. Style-aware normalized loss for improving arbitrary style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-143, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 338, + 279, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 338, + 279, + 378 + ], + "spans": [ + { + "bbox": [ + 57, + 338, + 279, + 378 + ], + "type": "text", + "content": "[6] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 380, + 280, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 380, + 280, + 422 + ], + "spans": [ + { + "bbox": [ + 57, + 380, + 280, + 422 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 423, + 280, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 423, + 280, + 476 + ], + "spans": [ + { + "bbox": [ + 57, + 423, + 280, + 476 + ], + "type": "text", + "content": "[8] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11326-11336, June 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 476, + 280, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 280, + 528 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 280, + 528 + ], + "type": "text", + "content": "[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11326-11336, 2022. 2, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 530, + 279, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 530, + 279, + 581 + ], + "spans": [ + { + "bbox": [ + 52, + 530, + 279, + 581 + ], + "type": "text", + "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 583, + 279, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 583, + 279, + 625 + ], + "spans": [ + { + "bbox": [ + 52, + 583, + 279, + 625 + ], + "type": "text", + "content": "[11] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. Advances in Neural Information Processing Systems, 35, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 627, + 279, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 627, + 279, + 658 + ], + "spans": [ + { + "bbox": [ + 52, + 627, + 279, + 658 + ], + "type": "text", + "content": "[12] Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-driven artistic style transfer. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 52, + 660, + 279, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 660, + 279, + 701 + ], + "spans": [ + { + "bbox": [ + 52, + 660, + 279, + 701 + ], + "type": "text", + "content": "[13] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 52, + 702, + 279, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 702, + 279, + 723 + ], + "spans": [ + { + "bbox": [ + 52, + 702, + 279, + 723 + ], + "type": "text", + "content": "[14] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 113, + 526, + 723 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 317, + 113, + 526, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 113, + 526, + 134 + ], + "spans": [ + { + "bbox": [ + 317, + 113, + 526, + 134 + ], + "type": "text", + "content": "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 136, + 525, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 136, + 525, + 166 + ], + "spans": [ + { + "bbox": [ + 299, + 136, + 525, + 166 + ], + "type": "text", + "content": "[15] Jiatao Gu, James Bradbury, Caiming Xiong, Victor O. K. Li, and Richard Socher. Non-autoregressive neural machine translation. In ICLR, 2018. 2, 3, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 168, + 525, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 168, + 525, + 208 + ], + "spans": [ + { + "bbox": [ + 299, + 168, + 525, + 208 + ], + "type": "text", + "content": "[16] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 211, + 526, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 211, + 526, + 263 + ], + "spans": [ + { + "bbox": [ + 299, + 211, + 526, + 263 + ], + "type": "text", + "content": "[17] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 264, + 525, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 264, + 525, + 294 + ], + "spans": [ + { + "bbox": [ + 299, + 264, + 525, + 294 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 296, + 525, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 296, + 525, + 347 + ], + "spans": [ + { + "bbox": [ + 299, + 296, + 525, + 347 + ], + "type": "text", + "content": "[19] Dmytro Kotovenko, Artsiom Sanakoyeu, Sabine Lang, and Bjorn Ommer. Content and style disentanglement for artistic style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4422-4431, 2019. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 349, + 526, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 349, + 526, + 391 + ], + "spans": [ + { + "bbox": [ + 299, + 349, + 526, + 391 + ], + "type": "text", + "content": "[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18062-18071, 2022. 2, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 393, + 525, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 393, + 525, + 444 + ], + "spans": [ + { + "bbox": [ + 299, + 393, + 525, + 444 + ], + "type": "text", + "content": "[21] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Diversified texture synthesis with feed-forward networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3920-3928, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 446, + 525, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 446, + 525, + 509 + ], + "spans": [ + { + "bbox": [ + 299, + 446, + 525, + 509 + ], + "type": "text", + "content": "[22] Tianwei Lin, Zhuoqi Ma, Fu Li, Dongliang He, Xin Li, Errui Ding, Nannan Wang, Jie Li, and Xinbo Gao. Drafting and revision: Laplacian pyramid network for fast high-quality artistic style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5141-5150, June 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 511, + 525, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 511, + 525, + 562 + ], + "spans": [ + { + "bbox": [ + 299, + 511, + 525, + 562 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 564, + 525, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 564, + 525, + 625 + ], + "spans": [ + { + "bbox": [ + 299, + 564, + 525, + 625 + ], + "type": "text", + "content": "[24] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 628, + 525, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 628, + 525, + 690 + ], + "spans": [ + { + "bbox": [ + 299, + 628, + 525, + 690 + ], + "type": "text", + "content": "[25] Yahui Liu, Enver Sangineto, Yajing Chen, Linchao Bao, Haoxian Zhang, Nicu Sebe, Bruno Lepri, Wei Wang, and Marco De Nadai. Smoothing the disentangled latent style space for unsupervised image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 692, + 525, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 692, + 525, + 723 + ], + "spans": [ + { + "bbox": [ + 299, + 692, + 525, + 723 + ], + "type": "text", + "content": "[26] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "spans": [ + { + "bbox": [ + 279, + 743, + 298, + 752 + ], + "type": "text", + "content": "7610" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 52, + 112, + 280, + 723 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 70, + 112, + 279, + 132 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 112, + 279, + 132 + ], + "spans": [ + { + "bbox": [ + 70, + 112, + 279, + 132 + ], + "type": "text", + "content": "IEEE Conference on Computer Vision and Pattern Recognition, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 134, + 237, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 134, + 237, + 144 + ], + "spans": [ + { + "bbox": [ + 52, + 134, + 237, + 144 + ], + "type": "text", + "content": "[27] OpenAI. https://github.com/openai/dall-e. 6, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 145, + 280, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 145, + 280, + 186 + ], + "spans": [ + { + "bbox": [ + 52, + 145, + 280, + 186 + ], + "type": "text", + "content": "[28] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 187, + 280, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 187, + 280, + 301 + ], + "spans": [ + { + "bbox": [ + 52, + 187, + 280, + 301 + ], + "type": "text", + "content": "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 302, + 280, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 302, + 280, + 364 + ], + "spans": [ + { + "bbox": [ + 52, + 302, + 280, + 364 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 365, + 280, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 365, + 280, + 406 + ], + "spans": [ + { + "bbox": [ + 52, + 365, + 280, + 406 + ], + "type": "text", + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 407, + 280, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 407, + 280, + 458 + ], + "spans": [ + { + "bbox": [ + 52, + 407, + 280, + 458 + ], + "type": "text", + "content": "[32] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 460, + 280, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 460, + 280, + 491 + ], + "spans": [ + { + "bbox": [ + 52, + 460, + 280, + 491 + ], + "type": "text", + "content": "[33] Ali Razavi, Aïron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 491, + 280, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 491, + 280, + 533 + ], + "spans": [ + { + "bbox": [ + 52, + 491, + 280, + 533 + ], + "type": "text", + "content": "[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 534, + 280, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 534, + 280, + 596 + ], + "spans": [ + { + "bbox": [ + 52, + 534, + 280, + 596 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 597, + 280, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 597, + 280, + 638 + ], + "spans": [ + { + "bbox": [ + 52, + 597, + 280, + 638 + ], + "type": "text", + "content": "[36] Artsiom Sanakoyeu, Dmytro Kotovenko, Sabine Lang, and Björn Ommer. A style-aware content loss for real-time hd style transfer. In Proceedings of the European Conference on Computer Vision (ECCV), pages 698-714, 10 2018. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 52, + 639, + 280, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 639, + 280, + 690 + ], + "spans": [ + { + "bbox": [ + 52, + 639, + 280, + 690 + ], + "type": "text", + "content": "[37] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6924-6932, 2017. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 52, + 691, + 280, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 691, + 280, + 723 + ], + "spans": [ + { + "bbox": [ + 52, + 691, + 280, + 723 + ], + "type": "text", + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 112, + 526, + 516 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 299, + 112, + 509, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 112, + 509, + 123 + ], + "spans": [ + { + "bbox": [ + 299, + 112, + 509, + 123 + ], + "type": "text", + "content": "[39] Phil Wang. https://github.com/ai-forever/ru-dalle. 6, 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 125, + 526, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 125, + 526, + 175 + ], + "spans": [ + { + "bbox": [ + 299, + 125, + 526, + 175 + ], + "type": "text", + "content": "[40] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: Towards aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia (ACM MM), 2022. 2, 3, 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 178, + 526, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 178, + 526, + 229 + ], + "spans": [ + { + "bbox": [ + 299, + 178, + 526, + 229 + ], + "type": "text", + "content": "[41] Zhizhong Wang, Lei Zhao, Haibo Chen, Lihong Qiu, Qihang Mo, Sihuan Lin, Wei Xing, and Dongming Lu. Diversified arbitrary style transfer via deep feature perturbation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7789-7798, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 231, + 526, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 231, + 526, + 262 + ], + "spans": [ + { + "bbox": [ + 299, + 231, + 526, + 262 + ], + "type": "text", + "content": "[42] Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Mach. Learn., 8(3-4):229-256, may 1992. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 263, + 526, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 263, + 526, + 314 + ], + "spans": [ + { + "bbox": [ + 299, + 263, + 526, + 314 + ], + "type": "text", + "content": "[43] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 316, + 526, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 316, + 526, + 367 + ], + "spans": [ + { + "bbox": [ + 299, + 316, + 526, + 367 + ], + "type": "text", + "content": "[44] Yuan Yao, Jianqiang Ren, Xuansong Xie, Weidong Liu, Yong-Jin Liu, and Jun Wang. Attention-aware multi-stroke style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1467-1475, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 369, + 526, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 369, + 526, + 421 + ], + "spans": [ + { + "bbox": [ + 299, + 369, + 526, + 421 + ], + "type": "text", + "content": "[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 422, + 526, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 422, + 526, + 464 + ], + "spans": [ + { + "bbox": [ + 299, + 422, + 526, + 464 + ], + "type": "text", + "content": "[46] Yulun Zhang, Chen Fang, Yilin Wang, Zhaowen Wang, Zhe Lin, Yun Fu, and Jimei Yang. Multimodal style transfer via graph cuts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5943-5951, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 465, + 526, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 465, + 526, + 516 + ], + "spans": [ + { + "bbox": [ + 299, + 465, + 526, + 516 + ], + "type": "text", + "content": "[47] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 279, + 743, + 298, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 279, + 743, + 298, + 753 + ], + "spans": [ + { + "bbox": [ + 279, + 743, + 298, + 753 + ], + "type": "text", + "content": "7611" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/000612a7-f65d-40f9-b28d-4f843782dfff_content_list.json b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/000612a7-f65d-40f9-b28d-4f843782dfff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..83f549c431abe934c558a202940e9f8b432ceb30 --- /dev/null +++ b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/000612a7-f65d-40f9-b28d-4f843782dfff_content_list.json @@ -0,0 +1,1843 @@ +[ + { + "type": "text", + "text": "SuS-X: Training-Free Name-Only Transfer of Vision-Language Models", + "text_level": 1, + "bbox": [ + 125, + 130, + 843, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vishaal Udandarao University of Cambridge vu214@cam.ac.uk", + "bbox": [ + 142, + 181, + 343, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ankush Gupta DeepMind, London ankushgupta@google.com", + "bbox": [ + 385, + 181, + 584, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samuel Albanie \nUniversity of Cambridge \nsma71@cam.ac.uk", + "bbox": [ + 627, + 181, + 826, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive Language-Image Pre-training (CLIP) has emerged as a simple yet effective way to train large-scale vision-language models. CLIP demonstrates impressive zero-shot classification and retrieval performance on diverse downstream tasks. However, to leverage its full potential, fine-tuning still appears to be necessary. Finetuning the entire CLIP model can be resource-intensive and unstable. Moreover, recent methods that aim to circumvent this need for fine-tuning still require access to images from the target task distribution. In this paper, we pursue a different approach and explore the regime of training-free \"name-only transfer\" in which the only knowledge we possess about the downstream task comprises the names of downstream target categories. We propose a novel method, $\\mathbf{SuS - X}$ , consisting of two key building blocks—\"SuS\" and \"TIP-X\", that requires neither intensive finetuning nor costly labelled data. $\\mathbf{SuS - X}$ achieves state-of-the-art (SoTA) zero-shot classification results on 19 benchmark datasets. We further show the utility of TIP-X in the training-free few-shot setting, where we again achieve SoTA results over strong training-free baselines. Code is available at https://github.com/vishaal27/SuS-X.", + "bbox": [ + 76, + 300, + 473, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 657, + 209, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language pre-training has taken the machine learning community by storm. A broad range of vision-language models (VLMs) [57, 42, 73, 1, 37] exhibiting exceptional transfer on tasks like classification [80, 84], cross-modal retrieval [67, 2] and segmentation [63, 27] have emerged. These models are now the de facto standard for downstream task transfer in the field of computer vision.", + "bbox": [ + 75, + 683, + 468, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One such prominent model, CLIP [57], is trained on 400M image-text pairs using a contrastive loss that maximises the similarities of paired image-text samples. CLIP pioneered the notion of zero-shot transfer in the vision-language setting: classification on unseen datasets. For a given classification task, CLIP converts the class labels", + "bbox": [ + 75, + 789, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/65ab95a0ab2821d0e722d4adc1e665a5f0696092fdea8374a9416f1748bdd02c.jpg", + "image_caption": [ + "Figure 1: Training-free name-only transfer. We propose $SuS-X$ , a framework for enhancing the zero-shot transfer abilities of VLMs like CLIP [57], BLIP [42] and TCL [72], without training. To achieve this, we propose a novel method $TIP-X$ , which adapts these VLMs using a curated support set ( $SuS$ ) that is not drawn from the target distribution. Our $SuS$ leverages one key piece of information about the task at hand: the names of the target categories." + ], + "image_footnote": [], + "bbox": [ + 511, + 275, + 880, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "into textual prompts (e.g. \"A photo of a .\", where represents the ground-truth text label for each class). It then computes similarities between all the class prompts and the query image, selecting the class with the highest image similarity as the predicted label (see Eq. (2)).", + "bbox": [ + 496, + 527, + 890, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CLIP's zero-shot performance is however limited by its pre-training distribution [24, 60, 21, 51]. If the downstream dataset diverges significantly from the pretraining image distribution, CLIP's zero-shot performance drastically drops [21]. To mitigate this, several lines of work propose to adapt CLIP on diverse downstream tasks—Tab. 1 briefly summarises these methods. Most of them employ fine-tuning on either labelled or unlabelled subsets of data from the target task. However, fine-tuning such an overparameterised model can be unstable and lead to overfitting [15, 25]. Furthermore, having access to the true distribution of the target task can be prohibitive in data-scarce environments [12, 4, 38] and online learning settings [14, 65].", + "bbox": [ + 496, + 604, + 892, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To alleviate these issues, in this paper, we aim to adapt CLIP and other VLMs for downstream classification in a name-only (requires only category names $^2$ , but no samples", + "bbox": [ + 496, + 801, + 893, + 849 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This idea of zero-shot transfer is distinct from the traditional zero-shot", + "bbox": [ + 94, + 886, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "classification setup introduced by Lampert et al. [41] in which the task is to generalise to classes not seen during training.", + "bbox": [ + 500, + 862, + 890, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2We use category and class interchangeably in this paper.", + "bbox": [ + 517, + 887, + 821, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2725", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/d10902f247af84db3c73432e490a8bc03c6906c082fff66c805e6f8f4a06f450.jpg", + "table_caption": [ + "Table 1: Taxonomy of CLIP adaptation methods for downstream classification. We underline the Zero-Shot CLIP model to signify that it is the base model that all others build on top of. *This method considers access to all test-set samples simultaneously, hence we still consider it zero-shot. †This method additionally uses class hierarchy maps." + ], + "table_footnote": [], + "table_body": "
MethodDoes not require trainingDoes not require labelled dataDoes not require target data distribution
Few-shot fine-tuning methodsLP-CLIP [57]XXX
CoOp [84]XXX
PLOT [11]XXX
LASP [9]XXX
SoftCPT [19]XXX
VT-CLIP [79]XXX
VPT [17]XXX
ProDA [45]XXX
CoCoOp [83]XXX
CLIP-Adapter [25]XXX
Intermediate methodsTIP-Adapter [80]✓XX
UPL [36]X✓X
SVL-Adapter [54]X✓X
TPT [48]X✓✓
CLIP+SYN [33]X✓✓
CaFo [78]X✓✓
Zero-shot methodsZero-Shot CLIP [57]✓✓✓
CALIP [31]✓✓✓
CLIP+DN [85]*✓✓✓
Training-free name-only transfer methodsCuPL [56]✓✓✓
VisDesc [49]✓✓✓
CHiLS [53]†✓✓✓
SuS-X (ours)✓✓✓
", + "bbox": [ + 163, + 145, + 805, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from the target task) and training-free fashion. We propose $SuS-X$ (see Fig. 1), consisting of two novel building blocks: (i) $SuS$ (Support Sets), our dynamic support set curation strategy that forgoes the need for samples from the target task, and (ii) $TIP-X$ , our main framework for performing zero-shot classification while being training-free. For a given downstream task, we first curate a support set by leveraging the task category labels, either in a parametric manner i.e., generating images from large-scale text-to-image models (e.g., Stable Diffusion [59]) or non-parametric manner i.e., retrieving real-world images from a large vision-language data bank (e.g., LAION-5B [61]). We then use the curated support set as a proxy few-shot dataset to inform our downstream predictions using $TIP-X$ , in a similar vein to recent few-shot adaptation methods [25, 80].", + "bbox": [ + 73, + 500, + 470, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our extensive experiments show that $SuS-X$ outperforms zero-shot methods on 19 benchmark datasets across three VLMs, namely, CLIP, BLIP and TCL by $4.60\\%$ , $5.97\\%$ and $11.37\\%$ absolute average accuracy respectively. We further extend the $TIP-X$ framework to the few-shot regime, outperforming previous SoTA methods in the training-free domain. Our main contributions are three-fold: (1) We propose $SuS-X$ , a SoTA method in the training-free name-only transfer setting for downstream adaptation of VLMs, (2) We present $SuS$ , an effective strategy for curating support sets using parametric or non-parametric methods to mitigate", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the lack of data samples available from the target task distribution, and (3) We propose $TIP-X$ , a novel training-free method for adapting VLMs to downstream classification in both the name-only transfer and few-shot regimes.", + "bbox": [ + 496, + 500, + 893, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 571, + 640, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language (VL) Foundation Models. Recent years have seen a Cambrian explosion in large-scale VL foundation models [6]. In a seminal work, Radford et al. [57] introduced CLIP, a large VLM trained on a web-scale corpus (400M image-text pairs), that exhibits strong downstream visual task performance. The introduction of CLIP inspired further development of VLMs [42, 1, 37, 18, 81, 75, 72, 10, 70, 26, 28, 43, 46, 74], each pre-trained on web-scale datasets to learn joint image-text representations. These representations can then be applied to tackle downstream tasks like semantic segmentation [63, 27], object detection [30, 20], image captioning [50, 3] and generative modelling [59, 58]. In this work, we adapt such VLMs in a training-free setting to diverse downstream tasks.", + "bbox": [ + 496, + 598, + 893, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adaptation of VL models. CLIP introduces a paradigm shift with its zero-shot transfer ability [57]. In this setup, none of the target dataset classes are known $a$ -priori and the task is to adapt implicitly at inference time to a given dataset. Since CLIP's training objective drives it to assign appropriate similarities to image-text pairs, it acquires the", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2726", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ability to perform zero-shot classification directly.", + "bbox": [ + 76, + 90, + 408, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by CLIP's zero-shot success, further work has sought to improve upon its performance. In Tab. 1, we characterise some of these methods along three major axes: (i) if the method requires training, (ii) if the method requires labelled samples from the target task, and (iii) if the method requires samples from the target task distribution3.", + "bbox": [ + 75, + 106, + 467, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we focus on the training-free name-only transfer regime—our goal is to adapt VLMs to target tasks without explicit training or access to samples from the target distribution. Instead, we assume access only to category names of target tasks. This formulation was recently considered for semantic segmentation, where it was called name-only transfer [62]—we likewise adopt this terminology. To the best of our knowledge, only two other concurrent approaches, CuPL [56] and VisDesc [49], operate in this regime. They use pre-trained language models to enhance textual prompts for zero-shot classification. By contrast, $SuS-X$ pursues a support set curation strategy to adapt VLMs using knowledge of category names. These approaches are complementary, and we find that they can be productively combined. Two other related works operating purely in the zero-shot setting are: (1) CALIP [31], which uses parameter-free attention on image-text features, and (2) CLIP+DN [85], which uses distribution normalisation. We compare with these four baselines in Sec. 4.", + "bbox": [ + 75, + 196, + 468, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SuS-X: Training-Free Name-Only Transfer", + "text_level": 1, + "bbox": [ + 76, + 494, + 460, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We describe the two main building blocks of $\\mathbf{SuS - X}$ (1) Support Set (SuS) construction, and (2) training-free inference using our novel TIP- $X$ method. Fig. 2 depicts our overall training-free name-only transfer framework.", + "bbox": [ + 76, + 520, + 467, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SuS Construction", + "text_level": 1, + "bbox": [ + 76, + 590, + 245, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow recent adaptation methods [80, 25] that use a small collection of labelled images to provide visual information to CLIP. However, differently from these methods, rather than accessing labelled images from the target distribution, we propose two methods (described next) to construct such a support set $(SuS)$ without such access.", + "bbox": [ + 75, + 614, + 467, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(I) Stable Diffusion Generation. Our first method leverages the powerful text-to-image generation model, Stable Diffusion [59]. We employ specific prompting strategies for generating salient and informative support images. Concretely, given a set of downstream textual class labels, $\\mathcal{T} = \\{t_1,t_2,\\dots ,t_C\\}$ , where $C$ denotes the number of categories, we prompt Stable Diffusion to generate $N$ images per class. In this way, we construct our support set of size $NC$ , with each image having its associated class label.", + "bbox": [ + 75, + 705, + 467, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By default, we prompt Stable Diffusion using the original CLIP prompts, i.e., \"A photo of a $<\\text{CLASS}>$ ,\" where $<\\text{CLASS}>$ is the class text label. To further diversify the generation process, we follow CuPL [56] to first generate customised textual prompts for each class by prompting GPT-3 [8] to output descriptions of the particular class. We then feed this customised set of prompts output by GPT-3 into Stable Diffusion for generating images. For example, to generate images from the \"dog\" class, we prompt GPT-3 to describe \"dogs\", and then prompt Stable Diffusion with the resulting descriptions. In section 4.4, we compare the performance of the default (called Photo) and this augmented prompting procedure (called CuPL). Unless otherwise specified, all our experiments with Stable Diffusion support sets use the CuPL strategy.", + "bbox": [ + 500, + 90, + 890, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(II) LAION-5B Retrieval. Our second method leverages the large-scale vision-language dataset, LAION-5B [61]. It contains 5.85 billion image-text pairs, pre-filtered by CLIP. Using LAION-5B, we retrieve task-specific images using class text prompts for constructing the support set. Concretely, given textual class labels, $\\mathcal{T} = \\{t_1,t_2,\\dots ,t_C\\}$ , we rank all images in LAION-5B by their CLIP image-text similarity to each text class label $t_i$ , where $i\\in [1,C]$ . We then use the top $N$ image matches as our support set for class $i$ , resulting in an NC-sized support set of images with their associated class labels. Note that curating supporting knowledge by search is a classical technique in computer vision [23] that was recently revisited in the task of semantic segmentation [63]. Here we adapt this idea to the name-only transfer classification setting. For efficient retrieval, we leverage the approximate nearest neighbour indices released by the authors4. Similar to the Stable Diffusion generation approach, we experiment with both Photo and CuPL prompting strategies for curating our LAION-5B support set (see Sec. 4.4). By default, we use Photo prompting for all our experiments with LAION-5B support sets.", + "bbox": [ + 496, + 318, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remark. SuS can be seen as a visual analogue to CuPL [56]—we augment VLMs with rich, class-specific images, instead of generating customised text descriptions.", + "bbox": [ + 496, + 633, + 890, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. TIP-X Inference", + "text_level": 1, + "bbox": [ + 500, + 688, + 656, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given our support set from the previous section, we now leverage it in a training-free method to inform CLIP's zero-shot predictions. We first briefly review CLIP zero-shot classification and TIP-Adapter [80] (a training-free adaptation method). We then highlight a critical shortcoming in TIP-Adapter due to uncalibrated intra-modal embedding distances, which we address in our method—TIP-X.", + "bbox": [ + 496, + 712, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Zero-shot CLIP. For classification into $C$ classes, CLIP converts class labels into text prompts and encodes them with its text encoder. Collectively, the encoded prompt vectors can be interpreted as a classifier weight matrix $W \\in$", + "bbox": [ + 496, + 816, + 890, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3Note that (iii) subsumes (ii). (ii) refers to access to labelled data samples from the target dataset whereas (iii) refers to a more general setting where the samples from the target dataset can be unlabelled. We distinguish between the two for clarity.", + "bbox": [ + 75, + 849, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4https://huggingface.co/datasets/laion/laion5B-index", + "bbox": [ + 516, + 886, + 797, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2727", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/77e4146946ebc30f1eacc2c17bfd9dd6cbdcc6e1c25465a38ed1eca0be48c405.jpg", + "image_caption": [ + "Figure 2: SuS-X for training-free name-only transfer. SuS-X consists of two core building blocks. (1) SuS (top right), a dynamic support set that we construct to infuse visual information into the VLM based only on knowledge of target category names. We construct support sets either in a parametric (generating images using Stable Diffusion) or non-parametric (retrieving images from LAION-5B) manner. (2) TIP-X (bottom right), our novel training-free method that leverages image-text distances to compute similarities between the support set and the test images. These similarities act as attention weights for the support set labels, and can directly be combined with the original logits from the VLM for classification." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 893, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{C\\times d}$ , where $d$ is embedding dimension. For a test set $T = \\{y_{1},y_{2},\\dots,y_{t}\\}$ comprising $t$ test images, CLIP's image encoder is applied to produce test image features:", + "bbox": [ + 76, + 446, + 470, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} = \\operatorname {C L I P I m a g e E n c o d e r} \\left(y _ {i}\\right), i \\in [ 1, t ], f _ {i} \\in \\mathbb {R} ^ {d} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 501, + 468, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf = \\operatorname {C o n c a t} ([ f _ {1}, f _ {2}, \\dots , f _ {t} ]), f \\in \\mathbb {R} ^ {t \\times d}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 523, + 401, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using $W$ and $f$ , CLIP performs classification by computing zero-shot logits (ZSL) via a dot product:", + "bbox": [ + 76, + 549, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {Z S L} = f W ^ {T} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 588, + 468, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TIP-Adapter. Given a $CK$ -sized $K$ -shot labelled dataset $D = \\{x_{1}, x_{2}, \\ldots, x_{CK}\\}^{5}$ from the target domain, TIP-Adapter [80] encodes $D$ using CLIP's image encoder:", + "bbox": [ + 76, + 614, + 468, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} = \\operatorname {C L I P I m a g e E n c o d e r} \\left(x _ {i}\\right), i \\in [ 1, C K ], F _ {i} \\in \\mathbb {R} ^ {d}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 669, + 460, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF = \\operatorname {C o n c a t} ([ F _ {1}, F _ {2}, \\dots , F _ {C K} ]), F \\in \\mathbb {R} ^ {C K \\times d} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 689, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It then converts each of the few-shot class labels to one-hot vectors $L \\in \\mathbb{R}^{CK \\times C}$ . Next, it computes an affinity matrix to capture the similarities between $F$ and $f$ :", + "bbox": [ + 76, + 722, + 468, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA = \\exp (- \\beta (1 - f F ^ {T})) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 776, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is a hyperparameter that modulates \"sharpness\". Finally, these affinities are used as attention weights over $L$ to produce logits that are blended with ZSL using a hyperparameter, $\\alpha$ :", + "bbox": [ + 76, + 803, + 468, + 862 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T L} = \\alpha A L + f W ^ {T} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 862, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivating TIP-X. TIP-Adapter benefits from the affinity computation between the test and few-shot image samples (Eq. (4)). This similarity is computed in CLIP's image space. However, prior research [76, 44, 66] has demonstrated the existence of a modality gap between CLIP's image and text spaces. This leads us to question if doing image-image similarity comparisons in CLIP's image space is optimal.", + "bbox": [ + 496, + 446, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3a shows the pairwise image-image, text-text and image-text cosine similarities of the ImageNet validation set CLIP embeddings. Clearly, the intra-modal and intermodal similarities are distributed differently—the intermodal similarities have small variance and mean, whereas the intra-modal similarities have larger means and variances. This mismatch happens because contrastive training of CLIP maximises the inter-modal cosine similarities of paired samples without regard to intra-modal similarities. This implies that the intra-image CLIP embedding similarities employed by TIP-Adapter may not reflect the true intra-image similarities. Fig. 3b illustrates this idea with a simple example. Consider two image embeddings that are required to be a distance $r$ away from a particular text embedding. The two image embeddings can satisfy this condition by being very close to each other or very far apart from each other. Fig. 3b shows that this constraint can be satisfied by any two arbitrary points on a hypersphere of radius $r$ . While we expect loose constraints to be imposed via transitivity, we nevertheless expect a lower quality of calibration in intra-modal (e.g., image-image) comparisons.", + "bbox": [ + 496, + 568, + 892, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TIP-X to the rescue. To get around the problem of un", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "5Note that a $K$ -shot labelled dataset for $C$ classes has a size $CK$", + "bbox": [ + 94, + 886, + 442, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2728", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dc5d814a059c4b7d2bca83c1126b28ed85053893ead31ef4d63b91debc11474a.jpg", + "image_caption": [ + "(a) Intra-modal and inter-modal CLIP cosine similarities. We observe quite distinct intra-modal and inter-modal cosine similarity distributions." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 459, + 205 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8cc77245b77bba9ffb8e7684f50a85b6f76fe56b73638526eee2d8efcfed72c1.jpg", + "image_caption": [ + "(b) Intra-modal degrees of freedom. Different intra-modal similarities can satisfy same inter-modal constraints, leaving room for poor calibration.", + "Figure 3: Our two-fold analysis motivating TIP-X" + ], + "image_footnote": [], + "bbox": [ + 84, + 260, + 460, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "calibrated intra-modal embedding distances in TIP-Adapter, we propose to use inter-modal distances as a bridge. More specifically, rather than computing similarities between the test features $(f\\in \\mathbb{R}^{t\\times d})$ and few-shot features $(F\\in \\mathbb{R}^{CK\\times d})$ in the image embedding space $(fF^T)$ , we use the image-text space. We first construct signatures by computing similarities of $f$ and $F$ with the text classifier weights $W$ :", + "bbox": [ + 75, + 489, + 468, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r} S = \\operatorname {s o f t m a x} \\left(F W ^ {T}\\right), S \\in \\mathbb {R} ^ {C K \\times C} \\\\ s = \\operatorname {s o f t m a x} \\left(f W ^ {T}\\right). s \\in \\mathbb {R} ^ {t \\times C} \\end{array} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 604, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These signatures comprise probability distributions encoding inter-modal affinities between the few-shot features and class text vectors, and likewise for the test features. We then construct our affinity matrix $M \\in \\mathbb{R}^{t \\times CK}$ by measuring the KL-divergence between the signatures as follows:", + "bbox": [ + 75, + 652, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i, j} = \\mathrm {K L} \\left(s _ {i} \\| S _ {j}\\right), i \\in [ 1, t ], j \\in [ 1, C K ] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 742, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_i$ represents the $i^{th}$ test signature for the $t$ test samples, and $S_j$ represents the $j^{th}$ few-shot signature. Since we are working with discrete probability distributions, we compute the KL-divergence as $\\mathbb{KL}(P||Q) = \\sum_{i} P_i \\log \\frac{P_i}{Q_i}$ .", + "bbox": [ + 75, + 763, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The construction of the affinity matrix $M$ can be seen as analogous to the affinity computation in TIP-Adapter (Eq. (4)). However, our affinity matrix construction removes direct reliance on the uncalibrated image-image similarities.", + "bbox": [ + 75, + 825, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, before using our affinity matrix $M$ as attention weights for $L$ (one-hot encoded class labels), we rescale (denoted by $\\psi$ ) the values of $M$ to have the same range (min, max values) as the TIP-Adapter affinities $(A)$ . Further, since our affinity matrix $M$ consists of KL-divergence values, the most similar samples will get small weights since their KL-divergence will be low (close to 0). To mitigate this, we simply negate the values in $M$ . We then blend our predicted logits with TL using a scalar $\\gamma$ :", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T X L} = f W ^ {T} + \\alpha A L + \\gamma \\psi (- M) L \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 232, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The entire $TIP-X$ method is shown in Fig. 2 (bottom right).", + "bbox": [ + 498, + 256, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. SuS-X: Combining SuS and TIP-X", + "text_level": 1, + "bbox": [ + 500, + 279, + 790, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our constructed support sets act as pseudo few-shot datasets, we directly replace the few-shot features $F$ in the $TIP - X$ framework with the features of our support set. We call our method $\\mathbf{SuS - X - LC}$ if we combine $TIP - X$ with the LAION-5B curated support set, and $\\mathbf{SuS - X - SD}$ when combined with the Stable Diffusion generated support set. These methods enable training-free name-only adaptation of zero-shot VLMs.", + "bbox": [ + 496, + 301, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 434, + 632, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we evaluate $\\mathbf{SuS - X}$ against strong baselines in the training-free zero-shot/name-only transfer regimes, across three VLMs. Next, we illustrate the adaptation of $TIP-X$ into the few-shot training-free regime. Finally, we ablate and analyse our method to provide additional insights.", + "bbox": [ + 496, + 459, + 890, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Training-free name-only transfer evaluation", + "text_level": 1, + "bbox": [ + 498, + 542, + 872, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. For a comprehensive evaluation, we test on 19 datasets spanning a wide range of object, scene and fine-grained categories: ImageNet [16], StanfordCars [39], UCF101 [64], Caltech101 [22], Caltech256 [29], Flowers102 [52], OxfordPets [55], Food101 [7], SUN397 [71], DTD [13], EuroSAT [34], FGVCAircraft [47], Country211 [57], CIFAR-10 [40], CIFAR-100 [40], Birdsnap [5], CUB [68], ImageNet-Sketch [69] and ImageNet-R [35]. Previous few-shot adaptation methods [77, 25, 82] benchmark on a subset of 11 of these 19 datasets. We report results on the 19-dataset suite in the main paper and compare results using only the 11-dataset subset in the supp. mat.", + "bbox": [ + 496, + 565, + 890, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Settings. We compare against six baselines. For zero-shot CLIP, we use prompt ensembling with 7 different prompt templates following $[57, 80]^6$ . We run CuPL7, VisDesc8 (name-only transfer) and CLIP+DN9", + "bbox": [ + 496, + 747, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "The 7 prompt templates are: \"itap of a .\", \"a origami .\", \"a bad photo of the .\", \"a photo of the large .\", \"a in a video game.\", \"art of the .\", and \"a photo of the small \".", + "bbox": [ + 500, + 814, + 890, + 862 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "7https://github.com/sarahpratt/CuPL", + "bbox": [ + 517, + 862, + 712, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "8https://github.com/sachit-menon/classify_by_description_release", + "bbox": [ + 517, + 875, + 861, + 887 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "9https://github.com/fengyuli2002/distribution-normalization", + "bbox": [ + 519, + 887, + 836, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2729", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(zero-shot transfer) using their official code. We also experiment with augmenting the CuPL prompts with the original prompt ensemble, and call it $\\mathrm{CuPL + e}$ . For CALIP (zero-shot transfer), in the absence of public code at the time of writing, we aim to reproduce their results using our own implementation. For our proposed methods, we report results using both SuS-X-LC and SuS-X-SD. For both methods, we use a fixed number of support samples per dataset (see supp. mat. for details). For CALIP and SuS-X, we conduct a hyperparameter search on the dataset validation sets. In Sec. 4.4 we perform a hyperparameter sensitivity test for a fair evaluation. By default, we use the ResNet-50 [32] backbone as CLIP's image encoder for all models.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main Results. In Tab. 2, we compare both variants of $SuS-X$ with the baselines. We report an average across 19 datasets. We also include results on ImageNet, EuroSAT, DTD, Birdsnap, ImageNet-R and ImageNet-Sketch (results on all 19 datasets in the supp. mat.). $SuS-X$ methods outperform zero-shot CLIP by $4.6\\%$ on average across all 19 datasets. We observe striking gains of $18\\%$ , $8\\%$ and $7\\%$ on EuroSAT, DTD and Birdsnap respectively. We also outperform the SoTA training-free adaptation methods—CuPL+ensemble and VisDesc by $1.1\\%$ and $3.1\\%$ on average respectively. To further probe where we attain the most gains, we plot the absolute improvement of our models over zero-shot CLIP in Fig. 4a. We observe large gains on fine-grained (Birdsnap, CUB, UCF101) and specialised (EuroSAT, DTD) datasets, demonstrating the utility of $SuS-X$ in injecting rich visual knowledge into zero-shot CLIP (additional fine-grained classification analysis in supp. mat.). We further compare $SuS-X$ to few-shot methods that use labelled samples from the true distribution in the supp. mat.—despite being at a disadvantage due to using no target distribution samples, $SuS-X$ is still competitive with these methods.", + "bbox": [ + 75, + 287, + 472, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Transfer to different VLMs", + "text_level": 1, + "bbox": [ + 76, + 630, + 326, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate transfer to VLMs other than CLIP, namely TCL [72] and BLIP [42]. We only retain image and text encoders of these models for computing features, while preserving all other experimental settings from Sec. 4.1. Tab. 3 shows our $SuS-X$ methods strongly outperform all baseline methods across both VLMs—we improve on zero-shot models by $11.37\\%$ and $5.97\\%$ on average across 19 datasets. This demonstrates that our method is not specific to CLIP, but can improve performance across different VLMs.", + "bbox": [ + 75, + 654, + 468, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Adapting to the few-shot regime", + "text_level": 1, + "bbox": [ + 76, + 801, + 362, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A key component of our $\\mathbf{SuS - X}$ method is $TIP - X$ . In the previous section, we showcased SoTA results in the training-free name-only transfer regime. Due to its formulation, $TIP - X$ can directly be extended to the few-shot regime, where our support sets are labelled samples from", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the target dataset rather than curated/generated samples. To evaluate $TIP-X$ on such real-world support sets, we conduct training-free few-shot classification using $TIP-X$ . We compare against the SoTA method in this regime—TIP-Adapter [80]. We report results on the 11-dataset subset used by TIP-Adapter on five different shot settings of the $K$ -shot classification task: 1, 2, 4, 8 and 16.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present average accuracy results on all shots in Fig. 4b—TIP-X outperforms both Zero-shot CLIP and TIP-Adapter (absolute gain of $0.91\\%$ across shots). Notably, on OxfordPets, we achieve $2.1\\%$ average gain. This further demonstrates the generalisability of the TIP-X method in transferring to the few-shot training-free setting.", + "bbox": [ + 496, + 198, + 893, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Analysis", + "text_level": 1, + "bbox": [ + 500, + 300, + 602, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct several ablations and provide additional visualisations to offer further insight into the $SuS - X$ method. Component Analysis. $SuS - X$ consists of two major building blocks— $SuS$ construction and $TIP - X$ . We compare the performance difference (with average accuracy across 19 datasets) of using $SuS$ with TIP-Adapter instead of $TIP - X$ in Tab. 4. We use both default ensemble prompts and CuPL prompts for CLIP's text classifier to break down the performance gains further. We note that both $SuS$ and $TIP - X$ are crucial for achieving the best results.", + "bbox": [ + 496, + 323, + 890, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transfer to different visual backbones. We evaluate the scalability of our model across different CLIP visual backbones—Fig. 4c shows that both $SuS-X$ variants consistently improve upon zero-shot CLIP across ResNet and VisionTransformer backbones of varying depths and sizes. $SuS$ size. We study the effect of varying support set size for $SuS-LC$ and $SuS-SD$ —we generate three different support sets with random seeds for support sizes of 1, 5, 10, 25, 50, 75 and 100 samples. From Fig. 6, we observe two broad trends—some tasks benefit (ImageNet-R, DTD) from having more support set samples while others do not (Country211, Flowers102). We suggest that this is connected to the domain gap between the true data distribution and support set samples—if the domain gap is large, it is inimical to provide a large support set, whereas if the domains are similar, providing more support samples always helps.", + "bbox": [ + 496, + 476, + 890, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SuS visualisation. We visualise samples from both support set construction methods on ImageNet in Fig. 5. It is hard to distinguish between the true ImageNet samples and the SuS samples—we can therefore construct support sets to mimic the true data distribution, with access to only the category names. A caveat is that the support set does not always capture the domain characteristics of the true distribution, leading to a domain gap (lighting conditions, diverse scene backgrounds, confounding objects etc). To fully close the gap to using true few-shot datasets as support sets [25, 80], further research into exact unsupervised domain matching of support sets and few-shot datasets is required.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2730", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e96f8a0401e24bcdcd977994b296a8ca67aad3ce5322c11042307c9dd13c4fa3.jpg", + "table_caption": [ + "Table 2: Training-free adaptation of CLIP on 19 datasets with RN50 visual backbone. The best and second best results for each dataset are bolded and underlined, respectively. Individual results for all 19 datasets are available in the supp. mat. *Average reported across 19 datasets. †Our re-implementation." + ], + "table_footnote": [], + "table_body": "
MethodAverage*ImageNet [16]ImageNet-R [35]ImageNet-Sketch [69]EuroSAT [34]DTD [13]Birdsnap [5]
Zero-shotZero-shot CLIP [57]52.2760.3159.3435.4226.8341.0130.56
CALIP [31]-60.57--38.9042.39-
CALIP [31]†52.3760.3159.3336.1026.9641.0230.68
CLIP+DN [85]53.0260.1660.3735.9528.3141.2131.23
Name-onlyCuPL [56]55.5061.4561.0235.1338.3848.6435.65
CuPL+e55.7661.6461.1735.8537.0647.4635.80
VisDesc [49]53.7659.6857.1633.7837.6041.9635.65
SuS-X-SD (ours)56.7361.8461.7636.3045.5750.5937.14
SuS-X-LC (ours)56.8761.8962.1037.8344.2349.2338.50
", + "bbox": [ + 80, + 136, + 883, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/32527ecad94a01223c8ddb456e081da891c38410cf32d9488d7416f6ea89bf0d.jpg", + "image_caption": [ + "Absolute improvement over Zero-shot CLIP", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 91, + 296, + 344, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14e7532a956a0d4f126bdd5b646c377c987386759076ca88628b58ab96a2171a.jpg", + "image_caption": [ + "Average over 11 datasets", + "(b)", + "Figure 4: (a) Comparison of $SuS-X$ with Zero-shot CLIP. (b) Results of training-free few-shot classification. (c) Performance comparison of $SuS-X$ across visual backbones." + ], + "image_footnote": [], + "bbox": [ + 367, + 296, + 599, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d9e8944f682abeb4400ffb894d99697357319df89a6c5bc970779c41d06a7c2.jpg", + "image_caption": [ + "Effect of visual backbone", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 625, + 295, + 864, + 428 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a7e4a781140fb074a71617f211bd1969958130f5df774f0daee5ff61d906bc1c.jpg", + "table_caption": [ + "Table 3: SuS- $X$ generalises to different VLMs. *Average reported across 19 datasets." + ], + "table_footnote": [], + "table_body": "
VLMMethodAverage*ImageNetEuroSATDTDBirdsnap
TCLZero-shot31.3835.5520.8028.554.51
CuPL34.7941.6026.3042.846.83
CuPL+e32.7941.3625.8841.966.60
VisDesc33.9440.4021.2734.285.69
SuS-X-SD41.4952.2928.7548.1713.60
SuS-X-LC42.7552.7736.9046.6317.93
BLIPZero-shot48.7350.5944.1044.6810.21
CuPL51.1152.9639.3752.9512.24
CuPL+e51.3653.0741.4853.3012.18
VisDesc49.9150.9442.2547.4511.69
SuS-X-SD53.2055.9345.3656.1516.95
SuS-X-LC54.6456.7551.6255.9123.78
", + "bbox": [ + 91, + 536, + 452, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompting strategies for SuS construction. Tab. 5 depicts the performance of Photo and CuPL prompting—best results are achieved with the LC-Photo and SD-CuPL strategies. We further compare the diversity of images produced by the two strategies on ImageNet11—from Tab. 5, it is evident that CuPL prompting leads to more diverse support sets as compared to Photo prompting.", + "bbox": [ + 75, + 731, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/98e386620cbbbbd2a7111af07d3532202d5a796911ed39f1c21f078010b6a7ad.jpg", + "table_caption": [ + "Table 4: Component Analysis of SuS-X." + ], + "table_footnote": [], + "table_body": "
Text PromptsMethodSuSTIP-XAverage Accuracy
DefaultZero-shot CLIP✗✗52.27
SuS-TIP-SD✓✗53.49 (+1.22%)
SuS-X-SD✓✓53.69 (+1.42%)
SuS-TIP-LC✓✗53.83 (+1.56%)
SuS-X-LC✓✓54.20 (+1.93%)
CuPL+eCuPL+e✗✗55.76 (+3.49%)
SuS-TIP-SD✓✗56.63 (+4.36%)
SuS-X-SD✓✓56.73 (+4.46%)
SuS-TIP-LC✓✗56.72 (+4.45%)
SuS-X-LC✓✓56.87 (+4.60%)
", + "bbox": [ + 519, + 521, + 867, + 675 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ffcd4bd4ee769c3d2bf71bc63ad8556a8a537366113ecb186685d832dfd3e86.jpg", + "table_caption": [ + "Table 5: Prompting strategies for SuS construction." + ], + "table_footnote": [], + "table_body": "
SuS methodAverage Acc.ImageNet Acc.Diversity
PhotoCuPLPhotoCuPLPhotoCuPL
LC56.8756.2061.8961.790.280.32
SD56.3256.7361.7961.840.170.20
", + "bbox": [ + 540, + 704, + 849, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameter Sensitivity. We perform a sensitivity test for our $\\gamma$ hyperparameter (refer Eq. 8) on ImageNet-R, OxfordPets, and DTD. We fix $\\alpha$ and $\\beta$ to be 1, and run a sweep over $\\gamma \\in [0,1]$ . From Tab. 6, we observe that moderate values of $\\gamma$ are typically preferred, and the variance of the accuracy values is small. However, note that for DTD, the optimal $\\gamma$ is slightly larger (0.75)—this is due to its spe", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "11We compute diversity as 1 minus the mean of the average pairwise image cosine-similarities within a class. A larger value implies low cosine similarities across images within a class, implying more diverse images. Alternatively, a smaller value implies less diverse images.", + "bbox": [ + 76, + 849, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2731", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d18fa62e710fa2cb7bdec58c7a0d214552fe2345c45490258849e8a037db35fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 87, + 230, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f30e3fa30e53e68b9ce23215771a116122563d84db7204f8010c888dac109fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 88, + 349, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b59335d45e4d8ae8477b266f42e33e18740a87c6baf7a51242a32e438a20859.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 88, + 473, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b98f1c92b5e2892eba622e4d81bfd2d82f68cf6de8eae6cba52859325bcab2e0.jpg", + "image_caption": [ + "(c) Australian Kelpie" + ], + "image_footnote": [], + "bbox": [ + 109, + 200, + 230, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7bbe1df296493689cf41d0f4a87a5f0d977c0a0ffa8c3e81f49eb2dfcf105971.jpg", + "image_caption": [ + "(a) Dishwasher" + ], + "image_footnote": [], + "bbox": [ + 232, + 200, + 349, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd8bfe97689358daf3387e7fd04e30072befb116efe1a0efa44ffbb418875fa0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 200, + 472, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a7e6385cc38b8904f59e6796353a576b1b1c1789eaa42b8669fea3d155f08ea0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 617, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b09202a6b171409825e3cc4e4e40d6ba37459ada3e2d3eb9cf53153ae669ff1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 88, + 736, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/01fa8d65683e1a8b6bb88b83bd2e7cf2f0826c501aef031233969bda89be6e9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 88, + 859, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/752d4480fca02e1b01d71e0a0b8ce85c475ac7fe5c00a791f31839622248878b.jpg", + "image_caption": [ + "(b) Split Rail Fence" + ], + "image_footnote": [], + "bbox": [ + 496, + 200, + 617, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5120b195c602f8fff73e92fe156e0335db8992f3cfdf8a0ea9232c75d5097a39.jpg", + "image_caption": [ + "(d) Bulbul" + ], + "image_footnote": [], + "bbox": [ + 619, + 200, + 736, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fc1e62077b2ec3595184dc4c257b07829677e278b3ff116aeddb945aefb471e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 200, + 859, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a991222a15d2d48b1787c4defc05f495bde3c4a6267215ffe0ae3ef44e323434.jpg", + "image_caption": [ + "Figure 5: Support samples from the generated SuS-SD, retrieved SuS-LC and true training distribution for ImageNet. By randomising the image order in each subfigure, we pose a challenge question—can you match the three images for each subfigure to their source i.e. SuS-SD, SuS-LC or ImageNet train set? The answers are provided at the bottom of the page10." + ], + "image_footnote": [], + "bbox": [ + 81, + 376, + 272, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f138b98170b35893f61ec645b6d15b42384751217cfed536317025dcbfed285.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 377, + 465, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f22fb2837c8403b787a8c0ed1353bd4c590a790c6ab63157cb2f0e2f5a03dfb.jpg", + "image_caption": [ + "(a) Tasks where larger support sets are beneficial", + "Figure 6: Effect of support size." + ], + "image_footnote": [], + "bbox": [ + 84, + 518, + 269, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56b686f9a5827ef7d5a2028efd8d1b52d5eece0792d62ff5a35ab57078b9d30e.jpg", + "image_caption": [ + "(b) Tasks where larger support sets are harmful" + ], + "image_footnote": [], + "bbox": [ + 269, + 518, + 460, + 635 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fe0802df17aab380caf9072ce51466bb46d59924a0fd1f3fabc45641b2178f1c.jpg", + "table_caption": [ + "Table 6: Hyperparameter sensitivity for $\\gamma$" + ], + "table_footnote": [], + "table_body": "
Datasetγ value
00.10.20.30.50.751
ImageNet-R60.8760.9861.0361.0561.0060.8960.65
OxfordPets76.7677.1777.5877.4477.1777.1776.90
DTD47.1647.1647.5147.6947.8747.9647.60
", + "bbox": [ + 93, + 705, + 450, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cialised nature which requires more guidance from the specialised support set to inform pre-trained CLIP. Previous few-shot adaptation works [25, 80] observed similar results. For more hyperparameter ablations, see the supp. mat.", + "bbox": [ + 75, + 806, + 467, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations and broader impact", + "text_level": 1, + "bbox": [ + 500, + 377, + 781, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While demonstrating promising results, we note some limitations of our approach: (1) To perform name-only transfer, we rely on CLIP having seen related concepts during pre-training. For rare concepts not seen during pretraining, transfer might not be feasible. (2) We employ LAION-5B [61] as a source of knowledge. While reasonable for a proof of concept, this data is relatively uncurated and may contain harmful content. As such, our approach is unsuitable for real-world deployment without careful mitigation strategies to address this. Similar arguments apply to Stable Diffusion [59].", + "bbox": [ + 496, + 405, + 890, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 595, + 617, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we studied the training-free name-only transfer paradigm for classification tasks with vision-language models. We systematically curated support sets with no access to samples from the target distribution and showed that they help improve CLIP's zero-shot predictions by providing rich, task-specific knowledge. We further motivated the $TIP-X$ framework through the observation that CLIP's intra-modal embedding spaces are not optimal for computing similarities. With these two building blocks, we demonstrated superior performance to prior state-of-the-art.", + "bbox": [ + 496, + 625, + 890, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the Isaac Newton Trust and an EPSRC access-to-HPC grant. SA would like to acknowledge the support of Z. Novak and N. Novak in enabling his contribution. VU would like to thank Gyungin Shin, Surabhi S. Nath, Jonathan Roberts, Vlad Bogolin, Kaiqu Liang and Anchit Jain for helpful discussions and feedback.", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "11 13T'uRnLdS(p)dsT'uRnL(o)TnRnLdS(q)uRnLdS'oT(e)", + "bbox": [ + 91, + 885, + 444, + 906 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2732", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022.", + "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022.", + "[3] Manuele Barraco, Marcella Cornia, Silvia Cascianelli, Lorenzo Baraldi, and Rita Cucchiara. The unreasonable effectiveness of clip features for image captioning: An experimental analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4662-4670, 2022.", + "[4] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018.", + "[5] Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N Belhumeur. Bird-snap: Large-scale fine-grained visual categorization of birds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2011-2018, 2014.", + "[6] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021.", + "[7] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European conference on computer vision, pages 446-461. Springer, 2014.", + "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020.", + "[9] Adrian Bulat and Georgios Tzimiropoulos. Language-aware soft prompting for vision & language foundation models. arXiv preprint arXiv:2210.01115, 2022.", + "[10] Delong Chen, Zhao Wu, Fan Liu, Zaiquan Yang, Yixiang Huang, Yiping Bao, and Erjin Zhou. Prototypical contrastive language image pretraining. arXiv preprint arXiv:2206.10996, 2022.", + "[11] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022.", + "[12] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6172-6180, 2018.", + "[13] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Andrea Cossu, Tinne Tuytelaars, Antonio Carta, Lucia Passaro, Vincenzo Lomonaco, and Davide Bacciu. Continual pre-training mitigates forgetting in language and vision. arXiv preprint arXiv:2205.09357, 2022.", + "[15] Guillaume Couairon, Matthijs Douze, Matthieu Cord, and Holger Schwenk. Embedding arithmetic of multimodal queries for image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4950-4958, 2022.", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[17] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Variational prompt tuning improves generalization of vision-language models. arXiv preprint arXiv:2210.02390, 2022.", + "[18] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021.", + "[19] Kun Ding, Ying Wang, Pengzhang Liu, Qiang Yu, Haojian Zhang, Shiming Xiang, and Chunhong Pan. Prompt tuning with soft context sharing for vision-language models. arXiv preprint arXiv:2208.13474, 2022.", + "[20] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022.", + "[21] Alex Fang, Gabriel Ilharco, Mitchell Wortsman, Yuhao Wan, Vaishaal Shankar, Achal Dave, and Ludwig Schmidt. Data determines distributional robustness in contrastive language image pre-training (clip). arXiv preprint arXiv:2205.01397, 2022.", + "[22] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004.", + "[23] Robert Fergus, Li Fei-Fei, Pietro Perona, and Andrew Zisserman. Learning object categories from google's image search. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1816-1823. IEEE, 2005.", + "[24] Benjamin Feuer, Ameya Joshi, and Chinmay Hegde. Caption supervision enables robust learners. arXiv preprint arXiv:2210.07396, 2022.", + "[25] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021.", + "[26] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022." + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2733", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021.", + "[28] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022.", + "[29] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007.", + "[30] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.", + "[31] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. arXiv preprint arXiv:2209.14169, 2022.", + "[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[33] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022.", + "[34] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019.", + "[35] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021.", + "[36] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022.", + "[37] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021.", + "[38] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell, 172(5):1122-1131, 2018.", + "[39] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013.", + "[40] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In 2009 IEEE conference on computer vision and pattern recognition, pages 951-958. IEEE, 2009.", + "[42] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022.", + "[43] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021.", + "[44] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022.", + "[45] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. arXiv preprint arXiv:2205.03340, 2022.", + "[46] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285, 2022.", + "[47] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013.", + "[48] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022.", + "[49] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022.", + "[50] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021.", + "[51] Thao Nguyen, Gabriel Ilharco, Mitchell Wortsman, Sewoong Oh, and Ludwig Schmidt. Quality not quantity: On the interaction between dataset design and robustness of clip. arXiv preprint arXiv:2208.05516, 2022.", + "[52] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008.", + "[53] Zachary Novack, Saurabh Garg, Julian McAuley, and Zachary C Lipton. Chils: Zero-shot image classification with hierarchical label sets. arXiv preprint arXiv:2302.02551, 2023.", + "[54] Omiros Pantazis, Gabriel Brostow, Kate Jones, and Oisin Mac Aodha. Svl-adapter: Self-supervised adapter for vision-language pretrained models. arXiv preprint arXiv:2210.03794, 2022.", + "[55] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2734", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. arXiv preprint arXiv:2209.03320, 2022.", + "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021.", + "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjarn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[60] Shibani Santurkar, Yann Dubois, Rohan Taori, Percy Liang, and Tatsunori Hashimoto. Is a caption worth a thousand images? a controlled study for representation learning. arXiv preprint arXiv:2207.07635, 2022.", + "[61] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022.", + "[62] Gyungin Shin, Weidi Xie, and Samuel Albanie. Named-mask: Distilling segmenters from complementary foundation models. arXiv preprint arXiv:2209.11228, 2022.", + "[63] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. arXiv preprint arXiv:2206.07045, 2022.", + "[64] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.", + "[65] Tejas Srinivasan, Ting-Yun Chang, Leticia Leonor Pinto Alva, Georgios Chochlakis, Mohammad Rostami, and Jesse Thomason. Climb: A continual learning benchmark for vision-and-language tasks. arXiv preprint arXiv:2206.09059, 2022.", + "[66] Vishaal Udandarao. Understanding and Fixing the Modality Gap in Vision-Language Models. Master's thesis, University of Cambridge, 2022.", + "[67] Vishaal Udandarao, Abhishek Maiti, Deepak Srivatsav, Suryatej Reddy Vyalla, Yifang Yin, and Rajiv Ratn Shah. Cobra: Contrastive bi-modal representation algorithm. arXiv preprint arXiv:2005.03687, 2020.", + "[68] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011.", + "[69] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021.", + "[71] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010.", + "[72] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022.", + "[73] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021.", + "[74] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. arXiv preprint arXiv:2207.12661, 2022.", + "[75] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022.", + "[76] Youngjae Yu, Jiwan Chung, Heeseung Yun, Jack Hessel, JaeSung Park, Ximing Lu, Prithviraj Ammanabrolu, Rowan Zellers, Ronan Le Bras, Gunhee Kim, et al. Multimodal knowledge alignment with reinforcement learning. arXiv preprint arXiv:2205.12630, 2022.", + "[77] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021.", + "[78] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Hongsheng Li, Yu Qiao, and Peng Gao. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. arXiv preprint arXiv:2303.02151, 2023.", + "[79] Renrui Zhang, Longtian Qiu, Wei Zhang, and Ziyao Zeng. Vt-clip: Enhancing vision-language models with visual-guided texts. arXiv preprint arXiv:2112.02399, 2021.", + "[80] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. arXiv preprint arXiv:2207.09519, 2022.", + "[81] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D Manning, and Curtis P Langlotz. Contrastive learning of medical visual representations from paired images and text. arXiv preprint arXiv:2010.00747, 2020.", + "[82] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. arXiv preprint arXiv:2109.01134, 2021." + ], + "bbox": [ + 501, + 90, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2735", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[83] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. arXiv preprint arXiv:2203.05557, 2022.", + "[84] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022.", + "[85] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser-Nam Lim. Distribution normalization: An \"effortless\" test-time augmentation for contrastively learned visual-language models. arXiv preprint arXiv:2302.11084, 2023." + ], + "bbox": [ + 78, + 90, + 467, + 246 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2736", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/000612a7-f65d-40f9-b28d-4f843782dfff_model.json b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/000612a7-f65d-40f9-b28d-4f843782dfff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..49d6fc922e5669be1c74c2dc183cb4fe849e6e66 --- /dev/null +++ b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/000612a7-f65d-40f9-b28d-4f843782dfff_model.json @@ -0,0 +1,2875 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.131, + 0.844, + 0.154 + ], + "angle": 0, + "content": "SuS-X: Training-Free Name-Only Transfer of Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.182, + 0.344, + 0.234 + ], + "angle": 0, + "content": "Vishaal Udandarao University of Cambridge vu214@cam.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.182, + 0.586, + 0.234 + ], + "angle": 0, + "content": "Ankush Gupta DeepMind, London ankushgupta@google.com" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.182, + 0.828, + 0.234 + ], + "angle": 0, + "content": "Samuel Albanie \nUniversity of Cambridge \nsma71@cam.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.474, + 0.633 + ], + "angle": 0, + "content": "Contrastive Language-Image Pre-training (CLIP) has emerged as a simple yet effective way to train large-scale vision-language models. CLIP demonstrates impressive zero-shot classification and retrieval performance on diverse downstream tasks. However, to leverage its full potential, fine-tuning still appears to be necessary. Finetuning the entire CLIP model can be resource-intensive and unstable. Moreover, recent methods that aim to circumvent this need for fine-tuning still require access to images from the target task distribution. In this paper, we pursue a different approach and explore the regime of training-free \"name-only transfer\" in which the only knowledge we possess about the downstream task comprises the names of downstream target categories. We propose a novel method, \\( \\mathbf{SuS - X} \\), consisting of two key building blocks—\"SuS\" and \"TIP-X\", that requires neither intensive finetuning nor costly labelled data. \\( \\mathbf{SuS - X} \\) achieves state-of-the-art (SoTA) zero-shot classification results on 19 benchmark datasets. We further show the utility of TIP-X in the training-free few-shot setting, where we again achieve SoTA results over strong training-free baselines. Code is available at https://github.com/vishaal27/SuS-X." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.659, + 0.21, + 0.675 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Vision-language pre-training has taken the machine learning community by storm. A broad range of vision-language models (VLMs) [57, 42, 73, 1, 37] exhibiting exceptional transfer on tasks like classification [80, 84], cross-modal retrieval [67, 2] and segmentation [63, 27] have emerged. These models are now the de facto standard for downstream task transfer in the field of computer vision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.47, + 0.88 + ], + "angle": 0, + "content": "One such prominent model, CLIP [57], is trained on 400M image-text pairs using a contrastive loss that maximises the similarities of paired image-text samples. CLIP pioneered the notion of zero-shot transfer in the vision-language setting: classification on unseen datasets. For a given classification task, CLIP converts the class labels" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.276, + 0.882, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.512 + ], + "angle": 0, + "content": "Figure 1: Training-free name-only transfer. We propose \\( SuS-X \\), a framework for enhancing the zero-shot transfer abilities of VLMs like CLIP [57], BLIP [42] and TCL [72], without training. To achieve this, we propose a novel method \\( TIP-X \\), which adapts these VLMs using a curated support set (\\( SuS \\)) that is not drawn from the target distribution. Our \\( SuS \\) leverages one key piece of information about the task at hand: the names of the target categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.604 + ], + "angle": 0, + "content": "into textual prompts (e.g. \"A photo of a .\", where represents the ground-truth text label for each class). It then computes similarities between all the class prompts and the query image, selecting the class with the highest image similarity as the predicted label (see Eq. (2))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.893, + 0.802 + ], + "angle": 0, + "content": "CLIP's zero-shot performance is however limited by its pre-training distribution [24, 60, 21, 51]. If the downstream dataset diverges significantly from the pretraining image distribution, CLIP's zero-shot performance drastically drops [21]. To mitigate this, several lines of work propose to adapt CLIP on diverse downstream tasks—Tab. 1 briefly summarises these methods. Most of them employ fine-tuning on either labelled or unlabelled subsets of data from the target task. However, fine-tuning such an overparameterised model can be unstable and lead to overfitting [15, 25]. Furthermore, having access to the true distribution of the target task can be prohibitive in data-scarce environments [12, 4, 38] and online learning settings [14, 65]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.894, + 0.85 + ], + "angle": 0, + "content": "To alleviate these issues, in this paper, we aim to adapt CLIP and other VLMs for downstream classification in a name-only (requires only category names\\(^2\\), but no samples" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.47, + 0.9 + ], + "angle": 0, + "content": "This idea of zero-shot transfer is distinct from the traditional zero-shot" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.892, + 0.888 + ], + "angle": 0, + "content": "classification setup introduced by Lampert et al. [41] in which the task is to generalise to classes not seen during training." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.822, + 0.901 + ], + "angle": 0, + "content": "2We use category and class interchangeably in this paper." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2725" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.138 + ], + "angle": 0, + "content": "Table 1: Taxonomy of CLIP adaptation methods for downstream classification. We underline the Zero-Shot CLIP model to signify that it is the base model that all others build on top of. *This method considers access to all test-set samples simultaneously, hence we still consider it zero-shot. †This method additionally uses class hierarchy maps." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.146, + 0.807, + 0.483 + ], + "angle": 0, + "content": "
MethodDoes not require trainingDoes not require labelled dataDoes not require target data distribution
Few-shot fine-tuning methodsLP-CLIP [57]XXX
CoOp [84]XXX
PLOT [11]XXX
LASP [9]XXX
SoftCPT [19]XXX
VT-CLIP [79]XXX
VPT [17]XXX
ProDA [45]XXX
CoCoOp [83]XXX
CLIP-Adapter [25]XXX
Intermediate methodsTIP-Adapter [80]✓XX
UPL [36]X✓X
SVL-Adapter [54]X✓X
TPT [48]X✓✓
CLIP+SYN [33]X✓✓
CaFo [78]X✓✓
Zero-shot methodsZero-Shot CLIP [57]✓✓✓
CALIP [31]✓✓✓
CLIP+DN [85]*✓✓✓
Training-free name-only transfer methodsCuPL [56]✓✓✓
VisDesc [49]✓✓✓
CHiLS [53]†✓✓✓
SuS-X (ours)✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.5, + 0.472, + 0.727 + ], + "angle": 0, + "content": "from the target task) and training-free fashion. We propose \\( SuS-X \\) (see Fig. 1), consisting of two novel building blocks: (i) \\( SuS \\) (Support Sets), our dynamic support set curation strategy that forgoes the need for samples from the target task, and (ii) \\( TIP-X \\), our main framework for performing zero-shot classification while being training-free. For a given downstream task, we first curate a support set by leveraging the task category labels, either in a parametric manner i.e., generating images from large-scale text-to-image models (e.g., Stable Diffusion [59]) or non-parametric manner i.e., retrieving real-world images from a large vision-language data bank (e.g., LAION-5B [61]). We then use the curated support set as a proxy few-shot dataset to inform our downstream predictions using \\( TIP-X \\), in a similar vein to recent few-shot adaptation methods [25, 80]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Our extensive experiments show that \\( SuS-X \\) outperforms zero-shot methods on 19 benchmark datasets across three VLMs, namely, CLIP, BLIP and TCL by \\( 4.60\\% \\), \\( 5.97\\% \\) and \\( 11.37\\% \\) absolute average accuracy respectively. We further extend the \\( TIP-X \\) framework to the few-shot regime, outperforming previous SoTA methods in the training-free domain. Our main contributions are three-fold: (1) We propose \\( SuS-X \\), a SoTA method in the training-free name-only transfer setting for downstream adaptation of VLMs, (2) We present \\( SuS \\), an effective strategy for curating support sets using parametric or non-parametric methods to mitigate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.895, + 0.561 + ], + "angle": 0, + "content": "the lack of data samples available from the target task distribution, and (3) We propose \\( TIP-X \\), a novel training-free method for adapting VLMs to downstream classification in both the name-only transfer and few-shot regimes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.642, + 0.59 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Vision-Language (VL) Foundation Models. Recent years have seen a Cambrian explosion in large-scale VL foundation models [6]. In a seminal work, Radford et al. [57] introduced CLIP, a large VLM trained on a web-scale corpus (400M image-text pairs), that exhibits strong downstream visual task performance. The introduction of CLIP inspired further development of VLMs [42, 1, 37, 18, 81, 75, 72, 10, 70, 26, 28, 43, 46, 74], each pre-trained on web-scale datasets to learn joint image-text representations. These representations can then be applied to tackle downstream tasks like semantic segmentation [63, 27], object detection [30, 20], image captioning [50, 3] and generative modelling [59, 58]. In this work, we adapt such VLMs in a training-free setting to diverse downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Adaptation of VL models. CLIP introduces a paradigm shift with its zero-shot transfer ability [57]. In this setup, none of the target dataset classes are known \\(a\\)-priori and the task is to adapt implicitly at inference time to a given dataset. Since CLIP's training objective drives it to assign appropriate similarities to image-text pairs, it acquires the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2726" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.409, + 0.106 + ], + "angle": 0, + "content": "ability to perform zero-shot classification directly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.197 + ], + "angle": 0, + "content": "Inspired by CLIP's zero-shot success, further work has sought to improve upon its performance. In Tab. 1, we characterise some of these methods along three major axes: (i) if the method requires training, (ii) if the method requires labelled samples from the target task, and (iii) if the method requires samples from the target task distribution3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.47, + 0.485 + ], + "angle": 0, + "content": "In this work, we focus on the training-free name-only transfer regime—our goal is to adapt VLMs to target tasks without explicit training or access to samples from the target distribution. Instead, we assume access only to category names of target tasks. This formulation was recently considered for semantic segmentation, where it was called name-only transfer [62]—we likewise adopt this terminology. To the best of our knowledge, only two other concurrent approaches, CuPL [56] and VisDesc [49], operate in this regime. They use pre-trained language models to enhance textual prompts for zero-shot classification. By contrast, \\( SuS-X \\) pursues a support set curation strategy to adapt VLMs using knowledge of category names. These approaches are complementary, and we find that they can be productively combined. Two other related works operating purely in the zero-shot setting are: (1) CALIP [31], which uses parameter-free attention on image-text features, and (2) CLIP+DN [85], which uses distribution normalisation. We compare with these four baselines in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.495, + 0.462, + 0.513 + ], + "angle": 0, + "content": "3. SuS-X: Training-Free Name-Only Transfer" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.521, + 0.468, + 0.582 + ], + "angle": 0, + "content": "We describe the two main building blocks of \\( \\mathbf{SuS - X} \\) (1) Support Set (SuS) construction, and (2) training-free inference using our novel TIP- \\( X \\) method. Fig. 2 depicts our overall training-free name-only transfer framework." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.246, + 0.607 + ], + "angle": 0, + "content": "3.1. SuS Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.468, + 0.705 + ], + "angle": 0, + "content": "We follow recent adaptation methods [80, 25] that use a small collection of labelled images to provide visual information to CLIP. However, differently from these methods, rather than accessing labelled images from the target distribution, we propose two methods (described next) to construct such a support set \\((SuS)\\) without such access." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.706, + 0.468, + 0.841 + ], + "angle": 0, + "content": "(I) Stable Diffusion Generation. Our first method leverages the powerful text-to-image generation model, Stable Diffusion [59]. We employ specific prompting strategies for generating salient and informative support images. Concretely, given a set of downstream textual class labels, \\(\\mathcal{T} = \\{t_1,t_2,\\dots ,t_C\\}\\), where \\(C\\) denotes the number of categories, we prompt Stable Diffusion to generate \\(N\\) images per class. In this way, we construct our support set of size \\(NC\\), with each image having its associated class label." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.318 + ], + "angle": 0, + "content": "By default, we prompt Stable Diffusion using the original CLIP prompts, i.e., \"A photo of a \\(<\\text{CLASS}>\\),\" where \\(<\\text{CLASS}>\\) is the class text label. To further diversify the generation process, we follow CuPL [56] to first generate customised textual prompts for each class by prompting GPT-3 [8] to output descriptions of the particular class. We then feed this customised set of prompts output by GPT-3 into Stable Diffusion for generating images. For example, to generate images from the \"dog\" class, we prompt GPT-3 to describe \"dogs\", and then prompt Stable Diffusion with the resulting descriptions. In section 4.4, we compare the performance of the default (called Photo) and this augmented prompting procedure (called CuPL). Unless otherwise specified, all our experiments with Stable Diffusion support sets use the CuPL strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.635 + ], + "angle": 0, + "content": "(II) LAION-5B Retrieval. Our second method leverages the large-scale vision-language dataset, LAION-5B [61]. It contains 5.85 billion image-text pairs, pre-filtered by CLIP. Using LAION-5B, we retrieve task-specific images using class text prompts for constructing the support set. Concretely, given textual class labels, \\(\\mathcal{T} = \\{t_1,t_2,\\dots ,t_C\\}\\), we rank all images in LAION-5B by their CLIP image-text similarity to each text class label \\(t_i\\), where \\(i\\in [1,C]\\). We then use the top \\(N\\) image matches as our support set for class \\(i\\), resulting in an NC-sized support set of images with their associated class labels. Note that curating supporting knowledge by search is a classical technique in computer vision [23] that was recently revisited in the task of semantic segmentation [63]. Here we adapt this idea to the name-only transfer classification setting. For efficient retrieval, we leverage the approximate nearest neighbour indices released by the authors4. Similar to the Stable Diffusion generation approach, we experiment with both Photo and CuPL prompting strategies for curating our LAION-5B support set (see Sec. 4.4). By default, we use Photo prompting for all our experiments with LAION-5B support sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.892, + 0.681 + ], + "angle": 0, + "content": "Remark. SuS can be seen as a visual analogue to CuPL [56]—we augment VLMs with rich, class-specific images, instead of generating customised text descriptions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.689, + 0.657, + 0.705 + ], + "angle": 0, + "content": "3.2. TIP-X Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Given our support set from the previous section, we now leverage it in a training-free method to inform CLIP's zero-shot predictions. We first briefly review CLIP zero-shot classification and TIP-Adapter [80] (a training-free adaptation method). We then highlight a critical shortcoming in TIP-Adapter due to uncalibrated intra-modal embedding distances, which we address in our method—TIP-X." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Zero-shot CLIP. For classification into \\( C \\) classes, CLIP converts class labels into text prompts and encodes them with its text encoder. Collectively, the encoded prompt vectors can be interpreted as a classifier weight matrix \\( W \\in \\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.468, + 0.901 + ], + "angle": 0, + "content": "3Note that (iii) subsumes (ii). (ii) refers to access to labelled data samples from the target dataset whereas (iii) refers to a more general setting where the samples from the target dataset can be unlabelled. We distinguish between the two for clarity." + }, + { + "type": "page_footnote", + "bbox": [ + 0.517, + 0.887, + 0.798, + 0.901 + ], + "angle": 0, + "content": "4https://huggingface.co/datasets/laion/laion5B-index" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2727" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.895, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.895, + 0.435 + ], + "angle": 0, + "content": "Figure 2: SuS-X for training-free name-only transfer. SuS-X consists of two core building blocks. (1) SuS (top right), a dynamic support set that we construct to infuse visual information into the VLM based only on knowledge of target category names. We construct support sets either in a parametric (generating images using Stable Diffusion) or non-parametric (retrieving images from LAION-5B) manner. (2) TIP-X (bottom right), our novel training-free method that leverages image-text distances to compute similarities between the support set and the test images. These similarities act as attention weights for the support set labels, and can directly be combined with the original logits from the VLM for classification." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.447, + 0.471, + 0.495 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{C\\times d}\\), where \\(d\\) is embedding dimension. For a test set \\(T = \\{y_{1},y_{2},\\dots,y_{t}\\}\\) comprising \\(t\\) test images, CLIP's image encoder is applied to produce test image features:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.502, + 0.469, + 0.528 + ], + "angle": 0, + "content": "\\[\nf _ {i} = \\operatorname {C L I P I m a g e E n c o d e r} \\left(y _ {i}\\right), i \\in [ 1, t ], f _ {i} \\in \\mathbb {R} ^ {d} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.524, + 0.402, + 0.542 + ], + "angle": 0, + "content": "\\[\nf = \\operatorname {C o n c a t} ([ f _ {1}, f _ {2}, \\dots , f _ {t} ]), f \\in \\mathbb {R} ^ {t \\times d}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.55, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Using \\( W \\) and \\( f \\), CLIP performs classification by computing zero-shot logits (ZSL) via a dot product:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.589, + 0.47, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathrm {Z S L} = f W ^ {T} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.616, + 0.47, + 0.663 + ], + "angle": 0, + "content": "TIP-Adapter. Given a \\( CK \\)-sized \\( K \\)-shot labelled dataset \\( D = \\{x_{1}, x_{2}, \\ldots, x_{CK}\\}^{5} \\) from the target domain, TIP-Adapter [80] encodes \\( D \\) using CLIP's image encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.67, + 0.462, + 0.688 + ], + "angle": 0, + "content": "\\[\nF _ {i} = \\operatorname {C L I P I m a g e E n c o d e r} \\left(x _ {i}\\right), i \\in [ 1, C K ], F _ {i} \\in \\mathbb {R} ^ {d}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.69, + 0.469, + 0.722 + ], + "angle": 0, + "content": "\\[\nF = \\operatorname {C o n c a t} ([ F _ {1}, F _ {2}, \\dots , F _ {C K} ]), F \\in \\mathbb {R} ^ {C K \\times d} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.723, + 0.47, + 0.768 + ], + "angle": 0, + "content": "It then converts each of the few-shot class labels to one-hot vectors \\(L \\in \\mathbb{R}^{CK \\times C}\\). Next, it computes an affinity matrix to capture the similarities between \\(F\\) and \\(f\\):" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.777, + 0.47, + 0.795 + ], + "angle": 0, + "content": "\\[\nA = \\exp (- \\beta (1 - f F ^ {T})) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.804, + 0.469, + 0.863 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a hyperparameter that modulates \"sharpness\". Finally, these affinities are used as attention weights over \\(L\\) to produce logits that are blended with ZSL using a hyperparameter, \\(\\alpha\\):" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.863, + 0.469, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T L} = \\alpha A L + f W ^ {T} \\tag {5}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.444, + 0.9 + ], + "angle": 0, + "content": "5Note that a \\(K\\) -shot labelled dataset for \\(C\\) classes has a size \\(CK\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Motivating TIP-X. TIP-Adapter benefits from the affinity computation between the test and few-shot image samples (Eq. (4)). This similarity is computed in CLIP's image space. However, prior research [76, 44, 66] has demonstrated the existence of a modality gap between CLIP's image and text spaces. This leads us to question if doing image-image similarity comparisons in CLIP's image space is optimal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.893, + 0.886 + ], + "angle": 0, + "content": "Fig. 3a shows the pairwise image-image, text-text and image-text cosine similarities of the ImageNet validation set CLIP embeddings. Clearly, the intra-modal and intermodal similarities are distributed differently—the intermodal similarities have small variance and mean, whereas the intra-modal similarities have larger means and variances. This mismatch happens because contrastive training of CLIP maximises the inter-modal cosine similarities of paired samples without regard to intra-modal similarities. This implies that the intra-image CLIP embedding similarities employed by TIP-Adapter may not reflect the true intra-image similarities. Fig. 3b illustrates this idea with a simple example. Consider two image embeddings that are required to be a distance \\( r \\) away from a particular text embedding. The two image embeddings can satisfy this condition by being very close to each other or very far apart from each other. Fig. 3b shows that this constraint can be satisfied by any two arbitrary points on a hypersphere of radius \\( r \\). While we expect loose constraints to be imposed via transitivity, we nevertheless expect a lower quality of calibration in intra-modal (e.g., image-image) comparisons." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "TIP-X to the rescue. To get around the problem of un" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2728" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.46, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.215, + 0.47, + 0.256 + ], + "angle": 0, + "content": "(a) Intra-modal and inter-modal CLIP cosine similarities. We observe quite distinct intra-modal and inter-modal cosine similarity distributions." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.261, + 0.462, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.396, + 0.47, + 0.438 + ], + "angle": 0, + "content": "(b) Intra-modal degrees of freedom. Different intra-modal similarities can satisfy same inter-modal constraints, leaving room for poor calibration." + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.45, + 0.446, + 0.466 + ], + "angle": 0, + "content": "Figure 3: Our two-fold analysis motivating TIP-X" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.47, + 0.597 + ], + "angle": 0, + "content": "calibrated intra-modal embedding distances in TIP-Adapter, we propose to use inter-modal distances as a bridge. More specifically, rather than computing similarities between the test features \\((f\\in \\mathbb{R}^{t\\times d})\\) and few-shot features \\((F\\in \\mathbb{R}^{CK\\times d})\\) in the image embedding space \\((fF^T)\\), we use the image-text space. We first construct signatures by computing similarities of \\(f\\) and \\(F\\) with the text classifier weights \\(W\\):" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.605, + 0.47, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r} S = \\operatorname {s o f t m a x} \\left(F W ^ {T}\\right), S \\in \\mathbb {R} ^ {C K \\times C} \\\\ s = \\operatorname {s o f t m a x} \\left(f W ^ {T}\\right). s \\in \\mathbb {R} ^ {t \\times C} \\end{array} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.73 + ], + "angle": 0, + "content": "These signatures comprise probability distributions encoding inter-modal affinities between the few-shot features and class text vectors, and likewise for the test features. We then construct our affinity matrix \\( M \\in \\mathbb{R}^{t \\times CK} \\) by measuring the KL-divergence between the signatures as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.743, + 0.469, + 0.76 + ], + "angle": 0, + "content": "\\[\nM _ {i, j} = \\mathrm {K L} \\left(s _ {i} \\| S _ {j}\\right), i \\in [ 1, t ], j \\in [ 1, C K ] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.826 + ], + "angle": 0, + "content": "where \\( s_i \\) represents the \\( i^{th} \\) test signature for the \\( t \\) test samples, and \\( S_j \\) represents the \\( j^{th} \\) few-shot signature. Since we are working with discrete probability distributions, we compute the KL-divergence as \\( \\mathbb{KL}(P||Q) = \\sum_{i} P_i \\log \\frac{P_i}{Q_i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.9 + ], + "angle": 0, + "content": "The construction of the affinity matrix \\(M\\) can be seen as analogous to the affinity computation in TIP-Adapter (Eq. (4)). However, our affinity matrix construction removes direct reliance on the uncalibrated image-image similarities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "Finally, before using our affinity matrix \\(M\\) as attention weights for \\(L\\) (one-hot encoded class labels), we rescale (denoted by \\(\\psi\\)) the values of \\(M\\) to have the same range (min, max values) as the TIP-Adapter affinities \\((A)\\). Further, since our affinity matrix \\(M\\) consists of KL-divergence values, the most similar samples will get small weights since their KL-divergence will be low (close to 0). To mitigate this, we simply negate the values in \\(M\\). We then blend our predicted logits with TL using a scalar \\(\\gamma\\):" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.233, + 0.892, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T X L} = f W ^ {T} + \\alpha A L + \\gamma \\psi (- M) L \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.257, + 0.891, + 0.273 + ], + "angle": 0, + "content": "The entire \\(TIP-X\\) method is shown in Fig. 2 (bottom right)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.28, + 0.791, + 0.295 + ], + "angle": 0, + "content": "3.3. SuS-X: Combining SuS and TIP-X" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.423 + ], + "angle": 0, + "content": "Since our constructed support sets act as pseudo few-shot datasets, we directly replace the few-shot features \\(F\\) in the \\(TIP - X\\) framework with the features of our support set. We call our method \\(\\mathbf{SuS - X - LC}\\) if we combine \\(TIP - X\\) with the LAION-5B curated support set, and \\(\\mathbf{SuS - X - SD}\\) when combined with the Stable Diffusion generated support set. These methods enable training-free name-only adaptation of zero-shot VLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.435, + 0.633, + 0.452 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.537 + ], + "angle": 0, + "content": "First, we evaluate \\( \\mathbf{SuS - X} \\) against strong baselines in the training-free zero-shot/name-only transfer regimes, across three VLMs. Next, we illustrate the adaptation of \\( TIP-X \\) into the few-shot training-free regime. Finally, we ablate and analyse our method to provide additional insights." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.543, + 0.874, + 0.559 + ], + "angle": 0, + "content": "4.1. Training-free name-only transfer evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Datasets. For a comprehensive evaluation, we test on 19 datasets spanning a wide range of object, scene and fine-grained categories: ImageNet [16], StanfordCars [39], UCF101 [64], Caltech101 [22], Caltech256 [29], Flowers102 [52], OxfordPets [55], Food101 [7], SUN397 [71], DTD [13], EuroSAT [34], FGVCAircraft [47], Country211 [57], CIFAR-10 [40], CIFAR-100 [40], Birdsnap [5], CUB [68], ImageNet-Sketch [69] and ImageNet-R [35]. Previous few-shot adaptation methods [77, 25, 82] benchmark on a subset of 11 of these 19 datasets. We report results on the 19-dataset suite in the main paper and compare results using only the 11-dataset subset in the supp. mat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.808 + ], + "angle": 0, + "content": "Experimental Settings. We compare against six baselines. For zero-shot CLIP, we use prompt ensembling with 7 different prompt templates following \\([57, 80]^6\\). We run CuPL7, VisDesc8 (name-only transfer) and CLIP+DN9" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.815, + 0.892, + 0.863 + ], + "angle": 0, + "content": "The 7 prompt templates are: \"itap of a .\", \"a origami .\", \"a bad photo of the .\", \"a photo of the large .\", \"a in a video game.\", \"art of the .\", and \"a photo of the small \"." + }, + { + "type": "page_footnote", + "bbox": [ + 0.519, + 0.863, + 0.714, + 0.876 + ], + "angle": 0, + "content": "7https://github.com/sarahpratt/CuPL" + }, + { + "type": "page_footnote", + "bbox": [ + 0.519, + 0.876, + 0.862, + 0.888 + ], + "angle": 0, + "content": "8https://github.com/sachit-menon/classify_by_description_release" + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.888, + 0.837, + 0.9 + ], + "angle": 0, + "content": "9https://github.com/fengyuli2002/distribution-normalization" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.815, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2729" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "(zero-shot transfer) using their official code. We also experiment with augmenting the CuPL prompts with the original prompt ensemble, and call it \\(\\mathrm{CuPL + e}\\). For CALIP (zero-shot transfer), in the absence of public code at the time of writing, we aim to reproduce their results using our own implementation. For our proposed methods, we report results using both SuS-X-LC and SuS-X-SD. For both methods, we use a fixed number of support samples per dataset (see supp. mat. for details). For CALIP and SuS-X, we conduct a hyperparameter search on the dataset validation sets. In Sec. 4.4 we perform a hyperparameter sensitivity test for a fair evaluation. By default, we use the ResNet-50 [32] backbone as CLIP's image encoder for all models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.621 + ], + "angle": 0, + "content": "Main Results. In Tab. 2, we compare both variants of \\( SuS-X \\) with the baselines. We report an average across 19 datasets. We also include results on ImageNet, EuroSAT, DTD, Birdsnap, ImageNet-R and ImageNet-Sketch (results on all 19 datasets in the supp. mat.). \\( SuS-X \\) methods outperform zero-shot CLIP by \\( 4.6\\% \\) on average across all 19 datasets. We observe striking gains of \\( 18\\% \\), \\( 8\\% \\) and \\( 7\\% \\) on EuroSAT, DTD and Birdsnap respectively. We also outperform the SoTA training-free adaptation methods—CuPL+ensemble and VisDesc by \\( 1.1\\% \\) and \\( 3.1\\% \\) on average respectively. To further probe where we attain the most gains, we plot the absolute improvement of our models over zero-shot CLIP in Fig. 4a. We observe large gains on fine-grained (Birdsnap, CUB, UCF101) and specialised (EuroSAT, DTD) datasets, demonstrating the utility of \\( SuS-X \\) in injecting rich visual knowledge into zero-shot CLIP (additional fine-grained classification analysis in supp. mat.). We further compare \\( SuS-X \\) to few-shot methods that use labelled samples from the true distribution in the supp. mat.—despite being at a disadvantage due to using no target distribution samples, \\( SuS-X \\) is still competitive with these methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.327, + 0.646 + ], + "angle": 0, + "content": "4.2. Transfer to different VLMs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.792 + ], + "angle": 0, + "content": "We evaluate transfer to VLMs other than CLIP, namely TCL [72] and BLIP [42]. We only retain image and text encoders of these models for computing features, while preserving all other experimental settings from Sec. 4.1. Tab. 3 shows our \\( SuS-X \\) methods strongly outperform all baseline methods across both VLMs—we improve on zero-shot models by \\( 11.37\\% \\) and \\( 5.97\\% \\) on average across 19 datasets. This demonstrates that our method is not specific to CLIP, but can improve performance across different VLMs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.363, + 0.818 + ], + "angle": 0, + "content": "4.3. Adapting to the few-shot regime" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "A key component of our \\( \\mathbf{SuS - X} \\) method is \\( TIP - X \\). In the previous section, we showcased SoTA results in the training-free name-only transfer regime. Due to its formulation, \\( TIP - X \\) can directly be extended to the few-shot regime, where our support sets are labelled samples from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "the target dataset rather than curated/generated samples. To evaluate \\( TIP-X \\) on such real-world support sets, we conduct training-free few-shot classification using \\( TIP-X \\). We compare against the SoTA method in this regime—TIP-Adapter [80]. We report results on the 11-dataset subset used by TIP-Adapter on five different shot settings of the \\( K \\)-shot classification task: 1, 2, 4, 8 and 16." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.895, + 0.29 + ], + "angle": 0, + "content": "We present average accuracy results on all shots in Fig. 4b—TIP-X outperforms both Zero-shot CLIP and TIP-Adapter (absolute gain of \\(0.91\\%\\) across shots). Notably, on OxfordPets, we achieve \\(2.1\\%\\) average gain. This further demonstrates the generalisability of the TIP-X method in transferring to the few-shot training-free setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.301, + 0.603, + 0.316 + ], + "angle": 0, + "content": "4.4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.476 + ], + "angle": 0, + "content": "We conduct several ablations and provide additional visualisations to offer further insight into the \\( SuS - X \\) method. Component Analysis. \\( SuS - X \\) consists of two major building blocks—\\( SuS \\) construction and \\( TIP - X \\). We compare the performance difference (with average accuracy across 19 datasets) of using \\( SuS \\) with TIP-Adapter instead of \\( TIP - X \\) in Tab. 4. We use both default ensemble prompts and CuPL prompts for CLIP's text classifier to break down the performance gains further. We note that both \\( SuS \\) and \\( TIP - X \\) are crucial for achieving the best results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.72 + ], + "angle": 0, + "content": "Transfer to different visual backbones. We evaluate the scalability of our model across different CLIP visual backbones—Fig. 4c shows that both \\( SuS-X \\) variants consistently improve upon zero-shot CLIP across ResNet and VisionTransformer backbones of varying depths and sizes. \\( SuS \\) size. We study the effect of varying support set size for \\( SuS-LC \\) and \\( SuS-SD \\)—we generate three different support sets with random seeds for support sizes of 1, 5, 10, 25, 50, 75 and 100 samples. From Fig. 6, we observe two broad trends—some tasks benefit (ImageNet-R, DTD) from having more support set samples while others do not (Country211, Flowers102). We suggest that this is connected to the domain gap between the true data distribution and support set samples—if the domain gap is large, it is inimical to provide a large support set, whereas if the domains are similar, providing more support samples always helps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SuS visualisation. We visualise samples from both support set construction methods on ImageNet in Fig. 5. It is hard to distinguish between the true ImageNet samples and the SuS samples—we can therefore construct support sets to mimic the true data distribution, with access to only the category names. A caveat is that the support set does not always capture the domain characteristics of the true distribution, leading to a domain gap (lighting conditions, diverse scene backgrounds, confounding objects etc). To fully close the gap to using true few-shot datasets as support sets [25, 80], further research into exact unsupervised domain matching of support sets and few-shot datasets is required." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2730" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Table 2: Training-free adaptation of CLIP on 19 datasets with RN50 visual backbone. The best and second best results for each dataset are bolded and underlined, respectively. Individual results for all 19 datasets are available in the supp. mat. *Average reported across 19 datasets. †Our re-implementation." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.137, + 0.885, + 0.279 + ], + "angle": 0, + "content": "
MethodAverage*ImageNet [16]ImageNet-R [35]ImageNet-Sketch [69]EuroSAT [34]DTD [13]Birdsnap [5]
Zero-shotZero-shot CLIP [57]52.2760.3159.3435.4226.8341.0130.56
CALIP [31]-60.57--38.9042.39-
CALIP [31]†52.3760.3159.3336.1026.9641.0230.68
CLIP+DN [85]53.0260.1660.3735.9528.3141.2131.23
Name-onlyCuPL [56]55.5061.4561.0235.1338.3848.6435.65
CuPL+e55.7661.6461.1735.8537.0647.4635.80
VisDesc [49]53.7659.6857.1633.7837.6041.9635.65
SuS-X-SD (ours)56.7361.8461.7636.3045.5750.5937.14
SuS-X-LC (ours)56.8761.8962.1037.8344.2349.2338.50
" + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.286, + 0.362, + 0.297 + ], + "angle": 0, + "content": "Absolute improvement over Zero-shot CLIP" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.297, + 0.345, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.432, + 0.23, + 0.445 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.287, + 0.565, + 0.297 + ], + "angle": 0, + "content": "Average over 11 datasets" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.297, + 0.6, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.432, + 0.493, + 0.444 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.286, + 0.825, + 0.296 + ], + "angle": 0, + "content": "Effect of visual backbone" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.296, + 0.865, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.432, + 0.756, + 0.444 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.45, + 0.893, + 0.481 + ], + "angle": 0, + "content": "Figure 4: (a) Comparison of \\( SuS-X \\) with Zero-shot CLIP. (b) Results of training-free few-shot classification. (c) Performance comparison of \\( SuS-X \\) across visual backbones." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.505, + 0.468, + 0.535 + ], + "angle": 0, + "content": "Table 3: SuS- \\( X \\) generalises to different VLMs. *Average reported across 19 datasets." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.537, + 0.453, + 0.715 + ], + "angle": 0, + "content": "
VLMMethodAverage*ImageNetEuroSATDTDBirdsnap
TCLZero-shot31.3835.5520.8028.554.51
CuPL34.7941.6026.3042.846.83
CuPL+e32.7941.3625.8841.966.60
VisDesc33.9440.4021.2734.285.69
SuS-X-SD41.4952.2928.7548.1713.60
SuS-X-LC42.7552.7736.9046.6317.93
BLIPZero-shot48.7350.5944.1044.6810.21
CuPL51.1152.9639.3752.9512.24
CuPL+e51.3653.0741.4853.3012.18
VisDesc49.9150.9442.2547.4511.69
SuS-X-SD53.2055.9345.3656.1516.95
SuS-X-LC54.6456.7551.6255.9123.78
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Prompting strategies for SuS construction. Tab. 5 depicts the performance of Photo and CuPL prompting—best results are achieved with the LC-Photo and SD-CuPL strategies. We further compare the diversity of images produced by the two strategies on ImageNet11—from Tab. 5, it is evident that CuPL prompting leads to more diverse support sets as compared to Photo prompting." + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.506, + 0.834, + 0.521 + ], + "angle": 0, + "content": "Table 4: Component Analysis of SuS-X." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.522, + 0.868, + 0.676 + ], + "angle": 0, + "content": "
Text PromptsMethodSuSTIP-XAverage Accuracy
DefaultZero-shot CLIP✗✗52.27
SuS-TIP-SD✓✗53.49 (+1.22%)
SuS-X-SD✓✓53.69 (+1.42%)
SuS-TIP-LC✓✗53.83 (+1.56%)
SuS-X-LC✓✓54.20 (+1.93%)
CuPL+eCuPL+e✗✗55.76 (+3.49%)
SuS-TIP-SD✓✗56.63 (+4.36%)
SuS-X-SD✓✓56.73 (+4.46%)
SuS-TIP-LC✓✗56.72 (+4.45%)
SuS-X-LC✓✓56.87 (+4.60%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.688, + 0.875, + 0.703 + ], + "angle": 0, + "content": "Table 5: Prompting strategies for SuS construction." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.705, + 0.851, + 0.775 + ], + "angle": 0, + "content": "
SuS methodAverage Acc.ImageNet Acc.Diversity
PhotoCuPLPhotoCuPLPhotoCuPL
LC56.8756.2061.8961.790.280.32
SD56.3256.7361.7961.840.170.20
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Hyperparameter Sensitivity. We perform a sensitivity test for our \\(\\gamma\\) hyperparameter (refer Eq. 8) on ImageNet-R, OxfordPets, and DTD. We fix \\(\\alpha\\) and \\(\\beta\\) to be 1, and run a sweep over \\(\\gamma \\in [0,1]\\). From Tab. 6, we observe that moderate values of \\(\\gamma\\) are typically preferred, and the variance of the accuracy values is small. However, note that for DTD, the optimal \\(\\gamma\\) is slightly larger (0.75)—this is due to its spe" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.851, + 0.47, + 0.902 + ], + "angle": 0, + "content": "11We compute diversity as 1 minus the mean of the average pairwise image cosine-similarities within a class. A larger value implies low cosine similarities across images within a class, implying more diverse images. Alternatively, a smaller value implies less diverse images." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2731" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.088, + 0.231, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.089, + 0.351, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.089, + 0.474, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.187, + 0.338, + 0.2 + ], + "angle": 0, + "content": "(a) Dishwasher" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.201, + 0.231, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.202, + 0.351, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.202, + 0.473, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.3, + 0.355, + 0.314 + ], + "angle": 0, + "content": "(c) Australian Kelpie" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.618, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.089, + 0.738, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.089, + 0.86, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.187, + 0.738, + 0.2 + ], + "angle": 0, + "content": "(b) Split Rail Fence" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.201, + 0.619, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.202, + 0.738, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.202, + 0.86, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.3, + 0.712, + 0.313 + ], + "angle": 0, + "content": "(d) Bulbul" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Figure 5: Support samples from the generated SuS-SD, retrieved SuS-LC and true training distribution for ImageNet. By randomising the image order in each subfigure, we pose a challenge question—can you match the three images for each subfigure to their source i.e. SuS-SD, SuS-LC or ImageNet train set? The answers are provided at the bottom of the page10." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.377, + 0.273, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.378, + 0.466, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.503, + 0.418, + 0.517 + ], + "angle": 0, + "content": "(a) Tasks where larger support sets are beneficial" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.519, + 0.27, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.519, + 0.462, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.644, + 0.414, + 0.657 + ], + "angle": 0, + "content": "(b) Tasks where larger support sets are harmful" + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.667, + 0.383, + 0.682 + ], + "angle": 0, + "content": "Figure 6: Effect of support size." + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.691, + 0.419, + 0.706 + ], + "angle": 0, + "content": "Table 6: Hyperparameter sensitivity for \\( \\gamma \\)" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.707, + 0.452, + 0.789 + ], + "angle": 0, + "content": "
Datasetγ value
00.10.20.30.50.751
ImageNet-R60.8760.9861.0361.0561.0060.8960.65
OxfordPets76.7677.1777.5877.4477.1777.1776.90
DTD47.1647.1647.5147.6947.8747.9647.60
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.468, + 0.87 + ], + "angle": 0, + "content": "cialised nature which requires more guidance from the specialised support set to inform pre-trained CLIP. Previous few-shot adaptation works [25, 80] observed similar results. For more hyperparameter ablations, see the supp. mat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.378, + 0.782, + 0.396 + ], + "angle": 0, + "content": "4.5. Limitations and broader impact" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.571 + ], + "angle": 0, + "content": "While demonstrating promising results, we note some limitations of our approach: (1) To perform name-only transfer, we rely on CLIP having seen related concepts during pre-training. For rare concepts not seen during pretraining, transfer might not be feasible. (2) We employ LAION-5B [61] as a source of knowledge. While reasonable for a proof of concept, this data is relatively uncurated and may contain harmful content. As such, our approach is unsuitable for real-world deployment without careful mitigation strategies to address this. Similar arguments apply to Stable Diffusion [59]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.596, + 0.619, + 0.612 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.79 + ], + "angle": 0, + "content": "In this paper, we studied the training-free name-only transfer paradigm for classification tasks with vision-language models. We systematically curated support sets with no access to samples from the target distribution and showed that they help improve CLIP's zero-shot predictions by providing rich, task-specific knowledge. We further motivated the \\( TIP-X \\) framework through the observation that CLIP's intra-modal embedding spaces are not optimal for computing similarities. With these two building blocks, we demonstrated superior performance to prior state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the Isaac Newton Trust and an EPSRC access-to-HPC grant. SA would like to acknowledge the support of Z. Novak and N. Novak in enabling his contribution. VU would like to thank Gyungin Shin, Surabhi S. Nath, Jonathan Roberts, Vlad Bogolin, Kaiqu Liang and Anchit Jain for helpful discussions and feedback." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.886, + 0.446, + 0.907 + ], + "angle": 0, + "content": "11 13T'uRnLdS(p)dsT'uRnL(o)TnRnLdS(q)uRnLdS'oT(e)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2732" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.47, + 0.313 + ], + "angle": 0, + "content": "[3] Manuele Barraco, Marcella Cornia, Silvia Cascianelli, Lorenzo Baraldi, and Rita Cucchiara. The unreasonable effectiveness of clip features for image captioning: An experimental analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4662-4670, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[5] Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N Belhumeur. Bird-snap: Large-scale fine-grained visual categorization of birds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2011-2018, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[7] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European conference on computer vision, pages 446-461. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[9] Adrian Bulat and Georgios Tzimiropoulos. Language-aware soft prompting for vision & language foundation models. arXiv preprint arXiv:2210.01115, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[10] Delong Chen, Zhao Wu, Fan Liu, Zaiquan Yang, Yixiang Huang, Yiping Bao, and Erjin Zhou. Prototypical contrastive language image pretraining. arXiv preprint arXiv:2206.10996, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[11] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[12] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6172-6180, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Andrea Cossu, Tinne Tuytelaars, Antonio Carta, Lucia Passaro, Vincenzo Lomonaco, and Davide Bacciu. Continual pre-training mitigates forgetting in language and vision. arXiv preprint arXiv:2205.09357, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[15] Guillaume Couairon, Matthijs Douze, Matthieu Cord, and Holger Schwenk. Embedding arithmetic of multimodal queries for image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4950-4958, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[17] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Variational prompt tuning improves generalization of vision-language models. arXiv preprint arXiv:2210.02390, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.894, + 0.403 + ], + "angle": 0, + "content": "[18] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[19] Kun Ding, Ying Wang, Pengzhang Liu, Qiang Yu, Haojian Zhang, Shiming Xiang, and Chunhong Pan. Prompt tuning with soft context sharing for vision-language models. arXiv preprint arXiv:2208.13474, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.894, + 0.531 + ], + "angle": 0, + "content": "[20] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[21] Alex Fang, Gabriel Ilharco, Mitchell Wortsman, Yuhao Wan, Vaishaal Shankar, Achal Dave, and Ludwig Schmidt. Data determines distributional robustness in contrastive language image pre-training (clip). arXiv preprint arXiv:2205.01397, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[22] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[23] Robert Fergus, Li Fei-Fei, Pietro Perona, and Andrew Zisserman. Learning object categories from google's image search. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1816-1823. IEEE, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[24] Benjamin Feuer, Ameya Joshi, and Chinmay Hegde. Caption supervision enables robust learners. arXiv preprint arXiv:2210.07396, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[25] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2733" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[27] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[28] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.223 + ], + "angle": 0, + "content": "[29] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.225, + 0.469, + 0.28 + ], + "angle": 0, + "content": "[30] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.469, + 0.338 + ], + "angle": 0, + "content": "[31] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. arXiv preprint arXiv:2209.14169, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.34, + 0.469, + 0.397 + ], + "angle": 0, + "content": "[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.399, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[33] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.469, + 0.526 + ], + "angle": 0, + "content": "[34] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.528, + 0.469, + 0.612 + ], + "angle": 0, + "content": "[35] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[36] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[37] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[38] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell, 172(5):1122-1131, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.815, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[39] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[40] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[41] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In 2009 IEEE conference on computer vision and pattern recognition, pages 951-958. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[42] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[43] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[44] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[45] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. arXiv preprint arXiv:2205.03340, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[46] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[47] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[48] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[49] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[50] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[51] Thao Nguyen, Gabriel Ilharco, Mitchell Wortsman, Sewoong Oh, and Ludwig Schmidt. Quality not quantity: On the interaction between dataset design and robustness of clip. arXiv preprint arXiv:2208.05516, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[52] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[53] Zachary Novack, Saurabh Garg, Julian McAuley, and Zachary C Lipton. Chils: Zero-shot image classification with hierarchical label sets. arXiv preprint arXiv:2302.02551, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[54] Omiros Pantazis, Gabriel Brostow, Kate Jones, and Oisin Mac Aodha. Svl-adapter: Self-supervised adapter for vision-language pretrained models. arXiv preprint arXiv:2210.03794, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[55] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2734" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[56] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. arXiv preprint arXiv:2209.03320, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.359 + ], + "angle": 0, + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjarn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[60] Shibani Santurkar, Yann Dubois, Rohan Taori, Percy Liang, and Tatsunori Hashimoto. Is a caption worth a thousand images? a controlled study for representation learning. arXiv preprint arXiv:2207.07635, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[61] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.502, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[62] Gyungin Shin, Weidi Xie, and Samuel Albanie. Named-mask: Distilling segmenters from complementary foundation models. arXiv preprint arXiv:2209.11228, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[63] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. arXiv preprint arXiv:2206.07045, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[64] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[65] Tejas Srinivasan, Ting-Yun Chang, Leticia Leonor Pinto Alva, Georgios Chochlakis, Mohammad Rostami, and Jesse Thomason. Climb: A continual learning benchmark for vision-and-language tasks. arXiv preprint arXiv:2206.09059, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[66] Vishaal Udandarao. Understanding and Fixing the Modality Gap in Vision-Language Models. Master's thesis, University of Cambridge, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[67] Vishaal Udandarao, Abhishek Maiti, Deepak Srivatsav, Suryatej Reddy Vyalla, Yifang Yin, and Rajiv Ratn Shah. Cobra: Contrastive bi-modal representation algorithm. arXiv preprint arXiv:2005.03687, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[68] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[69] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[70] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[71] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[72] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[73] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.892, + 0.435 + ], + "angle": 0, + "content": "[74] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. arXiv preprint arXiv:2207.12661, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.438, + 0.892, + 0.492 + ], + "angle": 0, + "content": "[75] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[76] Youngjae Yu, Jiwan Chung, Heeseung Yun, Jack Hessel, JaeSung Park, Ximing Lu, Prithviraj Ammanabrolu, Rowan Zellers, Ronan Le Bras, Gunhee Kim, et al. Multimodal knowledge alignment with reinforcement learning. arXiv preprint arXiv:2205.12630, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.569, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[77] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.627, + 0.892, + 0.694 + ], + "angle": 0, + "content": "[78] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Hongsheng Li, Yu Qiao, and Peng Gao. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. arXiv preprint arXiv:2303.02151, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.699, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[79] Renrui Zhang, Longtian Qiu, Wei Zhang, and Ziyao Zeng. Vt-clip: Enhancing vision-language models with visual-guided texts. arXiv preprint arXiv:2112.02399, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[80] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. arXiv preprint arXiv:2207.09519, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[81] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D Manning, and Curtis P Langlotz. Contrastive learning of medical visual representations from paired images and text. arXiv preprint arXiv:2010.00747, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[82] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. arXiv preprint arXiv:2109.01134, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "2735" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[83] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. arXiv preprint arXiv:2203.05557, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[84] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[85] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser-Nam Lim. Distribution normalization: An \"effortless\" test-time augmentation for contrastively learned visual-language models. arXiv preprint arXiv:2302.11084, 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2736" + } + ] +] \ No newline at end of file diff --git a/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/000612a7-f65d-40f9-b28d-4f843782dfff_origin.pdf b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/000612a7-f65d-40f9-b28d-4f843782dfff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..38188e4a81937e381e16dd99f5d9867f53d8879a --- /dev/null +++ b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/000612a7-f65d-40f9-b28d-4f843782dfff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:370f33a82b3545cc57d62f7ed4e855864982d10a8bbf614e6559d51c8ba1466c +size 5438856 diff --git a/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/full.md b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..079ba2cfa431a560f6601816c36fc22ce9dbf7b7 --- /dev/null +++ b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/full.md @@ -0,0 +1,371 @@ +# SuS-X: Training-Free Name-Only Transfer of Vision-Language Models + +Vishaal Udandarao University of Cambridge vu214@cam.ac.uk + +Ankush Gupta DeepMind, London ankushgupta@google.com + +Samuel Albanie +University of Cambridge +sma71@cam.ac.uk + +# Abstract + +Contrastive Language-Image Pre-training (CLIP) has emerged as a simple yet effective way to train large-scale vision-language models. CLIP demonstrates impressive zero-shot classification and retrieval performance on diverse downstream tasks. However, to leverage its full potential, fine-tuning still appears to be necessary. Finetuning the entire CLIP model can be resource-intensive and unstable. Moreover, recent methods that aim to circumvent this need for fine-tuning still require access to images from the target task distribution. In this paper, we pursue a different approach and explore the regime of training-free "name-only transfer" in which the only knowledge we possess about the downstream task comprises the names of downstream target categories. We propose a novel method, $\mathbf{SuS - X}$ , consisting of two key building blocks—"SuS" and "TIP-X", that requires neither intensive finetuning nor costly labelled data. $\mathbf{SuS - X}$ achieves state-of-the-art (SoTA) zero-shot classification results on 19 benchmark datasets. We further show the utility of TIP-X in the training-free few-shot setting, where we again achieve SoTA results over strong training-free baselines. Code is available at https://github.com/vishaal27/SuS-X. + +# 1. Introduction + +Vision-language pre-training has taken the machine learning community by storm. A broad range of vision-language models (VLMs) [57, 42, 73, 1, 37] exhibiting exceptional transfer on tasks like classification [80, 84], cross-modal retrieval [67, 2] and segmentation [63, 27] have emerged. These models are now the de facto standard for downstream task transfer in the field of computer vision. + +One such prominent model, CLIP [57], is trained on 400M image-text pairs using a contrastive loss that maximises the similarities of paired image-text samples. CLIP pioneered the notion of zero-shot transfer in the vision-language setting: classification on unseen datasets. For a given classification task, CLIP converts the class labels + +![](images/65ab95a0ab2821d0e722d4adc1e665a5f0696092fdea8374a9416f1748bdd02c.jpg) +Figure 1: Training-free name-only transfer. We propose $SuS-X$ , a framework for enhancing the zero-shot transfer abilities of VLMs like CLIP [57], BLIP [42] and TCL [72], without training. To achieve this, we propose a novel method $TIP-X$ , which adapts these VLMs using a curated support set ( $SuS$ ) that is not drawn from the target distribution. Our $SuS$ leverages one key piece of information about the task at hand: the names of the target categories. + +into textual prompts (e.g. "A photo of a .", where represents the ground-truth text label for each class). It then computes similarities between all the class prompts and the query image, selecting the class with the highest image similarity as the predicted label (see Eq. (2)). + +CLIP's zero-shot performance is however limited by its pre-training distribution [24, 60, 21, 51]. If the downstream dataset diverges significantly from the pretraining image distribution, CLIP's zero-shot performance drastically drops [21]. To mitigate this, several lines of work propose to adapt CLIP on diverse downstream tasks—Tab. 1 briefly summarises these methods. Most of them employ fine-tuning on either labelled or unlabelled subsets of data from the target task. However, fine-tuning such an overparameterised model can be unstable and lead to overfitting [15, 25]. Furthermore, having access to the true distribution of the target task can be prohibitive in data-scarce environments [12, 4, 38] and online learning settings [14, 65]. + +To alleviate these issues, in this paper, we aim to adapt CLIP and other VLMs for downstream classification in a name-only (requires only category names $^2$ , but no samples + +Table 1: Taxonomy of CLIP adaptation methods for downstream classification. We underline the Zero-Shot CLIP model to signify that it is the base model that all others build on top of. *This method considers access to all test-set samples simultaneously, hence we still consider it zero-shot. †This method additionally uses class hierarchy maps. + +
MethodDoes not require trainingDoes not require labelled dataDoes not require target data distribution
Few-shot fine-tuning methodsLP-CLIP [57]XXX
CoOp [84]XXX
PLOT [11]XXX
LASP [9]XXX
SoftCPT [19]XXX
VT-CLIP [79]XXX
VPT [17]XXX
ProDA [45]XXX
CoCoOp [83]XXX
CLIP-Adapter [25]XXX
Intermediate methodsTIP-Adapter [80]✓XX
UPL [36]X✓X
SVL-Adapter [54]X✓X
TPT [48]X✓✓
CLIP+SYN [33]X✓✓
CaFo [78]X✓✓
Zero-shot methodsZero-Shot CLIP [57]✓✓✓
CALIP [31]✓✓✓
CLIP+DN [85]*✓✓✓
Training-free name-only transfer methodsCuPL [56]✓✓✓
VisDesc [49]✓✓✓
CHiLS [53]†✓✓✓
SuS-X (ours)✓✓✓
+ +from the target task) and training-free fashion. We propose $SuS-X$ (see Fig. 1), consisting of two novel building blocks: (i) $SuS$ (Support Sets), our dynamic support set curation strategy that forgoes the need for samples from the target task, and (ii) $TIP-X$ , our main framework for performing zero-shot classification while being training-free. For a given downstream task, we first curate a support set by leveraging the task category labels, either in a parametric manner i.e., generating images from large-scale text-to-image models (e.g., Stable Diffusion [59]) or non-parametric manner i.e., retrieving real-world images from a large vision-language data bank (e.g., LAION-5B [61]). We then use the curated support set as a proxy few-shot dataset to inform our downstream predictions using $TIP-X$ , in a similar vein to recent few-shot adaptation methods [25, 80]. + +Our extensive experiments show that $SuS-X$ outperforms zero-shot methods on 19 benchmark datasets across three VLMs, namely, CLIP, BLIP and TCL by $4.60\%$ , $5.97\%$ and $11.37\%$ absolute average accuracy respectively. We further extend the $TIP-X$ framework to the few-shot regime, outperforming previous SoTA methods in the training-free domain. Our main contributions are three-fold: (1) We propose $SuS-X$ , a SoTA method in the training-free name-only transfer setting for downstream adaptation of VLMs, (2) We present $SuS$ , an effective strategy for curating support sets using parametric or non-parametric methods to mitigate + +the lack of data samples available from the target task distribution, and (3) We propose $TIP-X$ , a novel training-free method for adapting VLMs to downstream classification in both the name-only transfer and few-shot regimes. + +# 2. Related Work + +Vision-Language (VL) Foundation Models. Recent years have seen a Cambrian explosion in large-scale VL foundation models [6]. In a seminal work, Radford et al. [57] introduced CLIP, a large VLM trained on a web-scale corpus (400M image-text pairs), that exhibits strong downstream visual task performance. The introduction of CLIP inspired further development of VLMs [42, 1, 37, 18, 81, 75, 72, 10, 70, 26, 28, 43, 46, 74], each pre-trained on web-scale datasets to learn joint image-text representations. These representations can then be applied to tackle downstream tasks like semantic segmentation [63, 27], object detection [30, 20], image captioning [50, 3] and generative modelling [59, 58]. In this work, we adapt such VLMs in a training-free setting to diverse downstream tasks. + +Adaptation of VL models. CLIP introduces a paradigm shift with its zero-shot transfer ability [57]. In this setup, none of the target dataset classes are known $a$ -priori and the task is to adapt implicitly at inference time to a given dataset. Since CLIP's training objective drives it to assign appropriate similarities to image-text pairs, it acquires the + +ability to perform zero-shot classification directly. + +Inspired by CLIP's zero-shot success, further work has sought to improve upon its performance. In Tab. 1, we characterise some of these methods along three major axes: (i) if the method requires training, (ii) if the method requires labelled samples from the target task, and (iii) if the method requires samples from the target task distribution3. + +In this work, we focus on the training-free name-only transfer regime—our goal is to adapt VLMs to target tasks without explicit training or access to samples from the target distribution. Instead, we assume access only to category names of target tasks. This formulation was recently considered for semantic segmentation, where it was called name-only transfer [62]—we likewise adopt this terminology. To the best of our knowledge, only two other concurrent approaches, CuPL [56] and VisDesc [49], operate in this regime. They use pre-trained language models to enhance textual prompts for zero-shot classification. By contrast, $SuS-X$ pursues a support set curation strategy to adapt VLMs using knowledge of category names. These approaches are complementary, and we find that they can be productively combined. Two other related works operating purely in the zero-shot setting are: (1) CALIP [31], which uses parameter-free attention on image-text features, and (2) CLIP+DN [85], which uses distribution normalisation. We compare with these four baselines in Sec. 4. + +# 3. SuS-X: Training-Free Name-Only Transfer + +We describe the two main building blocks of $\mathbf{SuS - X}$ (1) Support Set (SuS) construction, and (2) training-free inference using our novel TIP- $X$ method. Fig. 2 depicts our overall training-free name-only transfer framework. + +# 3.1. SuS Construction + +We follow recent adaptation methods [80, 25] that use a small collection of labelled images to provide visual information to CLIP. However, differently from these methods, rather than accessing labelled images from the target distribution, we propose two methods (described next) to construct such a support set $(SuS)$ without such access. + +(I) Stable Diffusion Generation. Our first method leverages the powerful text-to-image generation model, Stable Diffusion [59]. We employ specific prompting strategies for generating salient and informative support images. Concretely, given a set of downstream textual class labels, $\mathcal{T} = \{t_1,t_2,\dots ,t_C\}$ , where $C$ denotes the number of categories, we prompt Stable Diffusion to generate $N$ images per class. In this way, we construct our support set of size $NC$ , with each image having its associated class label. + +By default, we prompt Stable Diffusion using the original CLIP prompts, i.e., "A photo of a $<\text{CLASS}>$ ," where $<\text{CLASS}>$ is the class text label. To further diversify the generation process, we follow CuPL [56] to first generate customised textual prompts for each class by prompting GPT-3 [8] to output descriptions of the particular class. We then feed this customised set of prompts output by GPT-3 into Stable Diffusion for generating images. For example, to generate images from the "dog" class, we prompt GPT-3 to describe "dogs", and then prompt Stable Diffusion with the resulting descriptions. In section 4.4, we compare the performance of the default (called Photo) and this augmented prompting procedure (called CuPL). Unless otherwise specified, all our experiments with Stable Diffusion support sets use the CuPL strategy. + +(II) LAION-5B Retrieval. Our second method leverages the large-scale vision-language dataset, LAION-5B [61]. It contains 5.85 billion image-text pairs, pre-filtered by CLIP. Using LAION-5B, we retrieve task-specific images using class text prompts for constructing the support set. Concretely, given textual class labels, $\mathcal{T} = \{t_1,t_2,\dots ,t_C\}$ , we rank all images in LAION-5B by their CLIP image-text similarity to each text class label $t_i$ , where $i\in [1,C]$ . We then use the top $N$ image matches as our support set for class $i$ , resulting in an NC-sized support set of images with their associated class labels. Note that curating supporting knowledge by search is a classical technique in computer vision [23] that was recently revisited in the task of semantic segmentation [63]. Here we adapt this idea to the name-only transfer classification setting. For efficient retrieval, we leverage the approximate nearest neighbour indices released by the authors4. Similar to the Stable Diffusion generation approach, we experiment with both Photo and CuPL prompting strategies for curating our LAION-5B support set (see Sec. 4.4). By default, we use Photo prompting for all our experiments with LAION-5B support sets. + +Remark. SuS can be seen as a visual analogue to CuPL [56]—we augment VLMs with rich, class-specific images, instead of generating customised text descriptions. + +# 3.2. TIP-X Inference + +Given our support set from the previous section, we now leverage it in a training-free method to inform CLIP's zero-shot predictions. We first briefly review CLIP zero-shot classification and TIP-Adapter [80] (a training-free adaptation method). We then highlight a critical shortcoming in TIP-Adapter due to uncalibrated intra-modal embedding distances, which we address in our method—TIP-X. + +Zero-shot CLIP. For classification into $C$ classes, CLIP converts class labels into text prompts and encodes them with its text encoder. Collectively, the encoded prompt vectors can be interpreted as a classifier weight matrix $W \in$ + +![](images/77e4146946ebc30f1eacc2c17bfd9dd6cbdcc6e1c25465a38ed1eca0be48c405.jpg) +Figure 2: SuS-X for training-free name-only transfer. SuS-X consists of two core building blocks. (1) SuS (top right), a dynamic support set that we construct to infuse visual information into the VLM based only on knowledge of target category names. We construct support sets either in a parametric (generating images using Stable Diffusion) or non-parametric (retrieving images from LAION-5B) manner. (2) TIP-X (bottom right), our novel training-free method that leverages image-text distances to compute similarities between the support set and the test images. These similarities act as attention weights for the support set labels, and can directly be combined with the original logits from the VLM for classification. + +$\mathbb{R}^{C\times d}$ , where $d$ is embedding dimension. For a test set $T = \{y_{1},y_{2},\dots,y_{t}\}$ comprising $t$ test images, CLIP's image encoder is applied to produce test image features: + +$$ +f _ {i} = \operatorname {C L I P I m a g e E n c o d e r} \left(y _ {i}\right), i \in [ 1, t ], f _ {i} \in \mathbb {R} ^ {d} \tag {1} +$$ + +$$ +f = \operatorname {C o n c a t} ([ f _ {1}, f _ {2}, \dots , f _ {t} ]), f \in \mathbb {R} ^ {t \times d} +$$ + +Using $W$ and $f$ , CLIP performs classification by computing zero-shot logits (ZSL) via a dot product: + +$$ +\mathrm {Z S L} = f W ^ {T} \tag {2} +$$ + +TIP-Adapter. Given a $CK$ -sized $K$ -shot labelled dataset $D = \{x_{1}, x_{2}, \ldots, x_{CK}\}^{5}$ from the target domain, TIP-Adapter [80] encodes $D$ using CLIP's image encoder: + +$$ +F _ {i} = \operatorname {C L I P I m a g e E n c o d e r} \left(x _ {i}\right), i \in [ 1, C K ], F _ {i} \in \mathbb {R} ^ {d} +$$ + +$$ +F = \operatorname {C o n c a t} ([ F _ {1}, F _ {2}, \dots , F _ {C K} ]), F \in \mathbb {R} ^ {C K \times d} \tag {3} +$$ + +It then converts each of the few-shot class labels to one-hot vectors $L \in \mathbb{R}^{CK \times C}$ . Next, it computes an affinity matrix to capture the similarities between $F$ and $f$ : + +$$ +A = \exp (- \beta (1 - f F ^ {T})) \tag {4} +$$ + +where $\beta$ is a hyperparameter that modulates "sharpness". Finally, these affinities are used as attention weights over $L$ to produce logits that are blended with ZSL using a hyperparameter, $\alpha$ : + +$$ +\mathrm {T L} = \alpha A L + f W ^ {T} \tag {5} +$$ + +Motivating TIP-X. TIP-Adapter benefits from the affinity computation between the test and few-shot image samples (Eq. (4)). This similarity is computed in CLIP's image space. However, prior research [76, 44, 66] has demonstrated the existence of a modality gap between CLIP's image and text spaces. This leads us to question if doing image-image similarity comparisons in CLIP's image space is optimal. + +Fig. 3a shows the pairwise image-image, text-text and image-text cosine similarities of the ImageNet validation set CLIP embeddings. Clearly, the intra-modal and intermodal similarities are distributed differently—the intermodal similarities have small variance and mean, whereas the intra-modal similarities have larger means and variances. This mismatch happens because contrastive training of CLIP maximises the inter-modal cosine similarities of paired samples without regard to intra-modal similarities. This implies that the intra-image CLIP embedding similarities employed by TIP-Adapter may not reflect the true intra-image similarities. Fig. 3b illustrates this idea with a simple example. Consider two image embeddings that are required to be a distance $r$ away from a particular text embedding. The two image embeddings can satisfy this condition by being very close to each other or very far apart from each other. Fig. 3b shows that this constraint can be satisfied by any two arbitrary points on a hypersphere of radius $r$ . While we expect loose constraints to be imposed via transitivity, we nevertheless expect a lower quality of calibration in intra-modal (e.g., image-image) comparisons. + +TIP-X to the rescue. To get around the problem of un + +![](images/dc5d814a059c4b7d2bca83c1126b28ed85053893ead31ef4d63b91debc11474a.jpg) +(a) Intra-modal and inter-modal CLIP cosine similarities. We observe quite distinct intra-modal and inter-modal cosine similarity distributions. + +![](images/8cc77245b77bba9ffb8e7684f50a85b6f76fe56b73638526eee2d8efcfed72c1.jpg) +(b) Intra-modal degrees of freedom. Different intra-modal similarities can satisfy same inter-modal constraints, leaving room for poor calibration. +Figure 3: Our two-fold analysis motivating TIP-X + +calibrated intra-modal embedding distances in TIP-Adapter, we propose to use inter-modal distances as a bridge. More specifically, rather than computing similarities between the test features $(f\in \mathbb{R}^{t\times d})$ and few-shot features $(F\in \mathbb{R}^{CK\times d})$ in the image embedding space $(fF^T)$ , we use the image-text space. We first construct signatures by computing similarities of $f$ and $F$ with the text classifier weights $W$ : + +$$ +\begin{array}{r} S = \operatorname {s o f t m a x} \left(F W ^ {T}\right), S \in \mathbb {R} ^ {C K \times C} \\ s = \operatorname {s o f t m a x} \left(f W ^ {T}\right). s \in \mathbb {R} ^ {t \times C} \end{array} \tag {6} +$$ + +These signatures comprise probability distributions encoding inter-modal affinities between the few-shot features and class text vectors, and likewise for the test features. We then construct our affinity matrix $M \in \mathbb{R}^{t \times CK}$ by measuring the KL-divergence between the signatures as follows: + +$$ +M _ {i, j} = \mathrm {K L} \left(s _ {i} \| S _ {j}\right), i \in [ 1, t ], j \in [ 1, C K ] \tag {7} +$$ + +where $s_i$ represents the $i^{th}$ test signature for the $t$ test samples, and $S_j$ represents the $j^{th}$ few-shot signature. Since we are working with discrete probability distributions, we compute the KL-divergence as $\mathbb{KL}(P||Q) = \sum_{i} P_i \log \frac{P_i}{Q_i}$ . + +The construction of the affinity matrix $M$ can be seen as analogous to the affinity computation in TIP-Adapter (Eq. (4)). However, our affinity matrix construction removes direct reliance on the uncalibrated image-image similarities. + +Finally, before using our affinity matrix $M$ as attention weights for $L$ (one-hot encoded class labels), we rescale (denoted by $\psi$ ) the values of $M$ to have the same range (min, max values) as the TIP-Adapter affinities $(A)$ . Further, since our affinity matrix $M$ consists of KL-divergence values, the most similar samples will get small weights since their KL-divergence will be low (close to 0). To mitigate this, we simply negate the values in $M$ . We then blend our predicted logits with TL using a scalar $\gamma$ : + +$$ +\mathrm {T X L} = f W ^ {T} + \alpha A L + \gamma \psi (- M) L \tag {8} +$$ + +The entire $TIP-X$ method is shown in Fig. 2 (bottom right). + +# 3.3. SuS-X: Combining SuS and TIP-X + +Since our constructed support sets act as pseudo few-shot datasets, we directly replace the few-shot features $F$ in the $TIP - X$ framework with the features of our support set. We call our method $\mathbf{SuS - X - LC}$ if we combine $TIP - X$ with the LAION-5B curated support set, and $\mathbf{SuS - X - SD}$ when combined with the Stable Diffusion generated support set. These methods enable training-free name-only adaptation of zero-shot VLMs. + +# 4. Experiments + +First, we evaluate $\mathbf{SuS - X}$ against strong baselines in the training-free zero-shot/name-only transfer regimes, across three VLMs. Next, we illustrate the adaptation of $TIP-X$ into the few-shot training-free regime. Finally, we ablate and analyse our method to provide additional insights. + +# 4.1. Training-free name-only transfer evaluation + +Datasets. For a comprehensive evaluation, we test on 19 datasets spanning a wide range of object, scene and fine-grained categories: ImageNet [16], StanfordCars [39], UCF101 [64], Caltech101 [22], Caltech256 [29], Flowers102 [52], OxfordPets [55], Food101 [7], SUN397 [71], DTD [13], EuroSAT [34], FGVCAircraft [47], Country211 [57], CIFAR-10 [40], CIFAR-100 [40], Birdsnap [5], CUB [68], ImageNet-Sketch [69] and ImageNet-R [35]. Previous few-shot adaptation methods [77, 25, 82] benchmark on a subset of 11 of these 19 datasets. We report results on the 19-dataset suite in the main paper and compare results using only the 11-dataset subset in the supp. mat. + +Experimental Settings. We compare against six baselines. For zero-shot CLIP, we use prompt ensembling with 7 different prompt templates following $[57, 80]^6$ . We run CuPL7, VisDesc8 (name-only transfer) and CLIP+DN9 + +(zero-shot transfer) using their official code. We also experiment with augmenting the CuPL prompts with the original prompt ensemble, and call it $\mathrm{CuPL + e}$ . For CALIP (zero-shot transfer), in the absence of public code at the time of writing, we aim to reproduce their results using our own implementation. For our proposed methods, we report results using both SuS-X-LC and SuS-X-SD. For both methods, we use a fixed number of support samples per dataset (see supp. mat. for details). For CALIP and SuS-X, we conduct a hyperparameter search on the dataset validation sets. In Sec. 4.4 we perform a hyperparameter sensitivity test for a fair evaluation. By default, we use the ResNet-50 [32] backbone as CLIP's image encoder for all models. + +Main Results. In Tab. 2, we compare both variants of $SuS-X$ with the baselines. We report an average across 19 datasets. We also include results on ImageNet, EuroSAT, DTD, Birdsnap, ImageNet-R and ImageNet-Sketch (results on all 19 datasets in the supp. mat.). $SuS-X$ methods outperform zero-shot CLIP by $4.6\%$ on average across all 19 datasets. We observe striking gains of $18\%$ , $8\%$ and $7\%$ on EuroSAT, DTD and Birdsnap respectively. We also outperform the SoTA training-free adaptation methods—CuPL+ensemble and VisDesc by $1.1\%$ and $3.1\%$ on average respectively. To further probe where we attain the most gains, we plot the absolute improvement of our models over zero-shot CLIP in Fig. 4a. We observe large gains on fine-grained (Birdsnap, CUB, UCF101) and specialised (EuroSAT, DTD) datasets, demonstrating the utility of $SuS-X$ in injecting rich visual knowledge into zero-shot CLIP (additional fine-grained classification analysis in supp. mat.). We further compare $SuS-X$ to few-shot methods that use labelled samples from the true distribution in the supp. mat.—despite being at a disadvantage due to using no target distribution samples, $SuS-X$ is still competitive with these methods. + +# 4.2. Transfer to different VLMs + +We evaluate transfer to VLMs other than CLIP, namely TCL [72] and BLIP [42]. We only retain image and text encoders of these models for computing features, while preserving all other experimental settings from Sec. 4.1. Tab. 3 shows our $SuS-X$ methods strongly outperform all baseline methods across both VLMs—we improve on zero-shot models by $11.37\%$ and $5.97\%$ on average across 19 datasets. This demonstrates that our method is not specific to CLIP, but can improve performance across different VLMs. + +# 4.3. Adapting to the few-shot regime + +A key component of our $\mathbf{SuS - X}$ method is $TIP - X$ . In the previous section, we showcased SoTA results in the training-free name-only transfer regime. Due to its formulation, $TIP - X$ can directly be extended to the few-shot regime, where our support sets are labelled samples from + +the target dataset rather than curated/generated samples. To evaluate $TIP-X$ on such real-world support sets, we conduct training-free few-shot classification using $TIP-X$ . We compare against the SoTA method in this regime—TIP-Adapter [80]. We report results on the 11-dataset subset used by TIP-Adapter on five different shot settings of the $K$ -shot classification task: 1, 2, 4, 8 and 16. + +We present average accuracy results on all shots in Fig. 4b—TIP-X outperforms both Zero-shot CLIP and TIP-Adapter (absolute gain of $0.91\%$ across shots). Notably, on OxfordPets, we achieve $2.1\%$ average gain. This further demonstrates the generalisability of the TIP-X method in transferring to the few-shot training-free setting. + +# 4.4. Analysis + +We conduct several ablations and provide additional visualisations to offer further insight into the $SuS - X$ method. Component Analysis. $SuS - X$ consists of two major building blocks— $SuS$ construction and $TIP - X$ . We compare the performance difference (with average accuracy across 19 datasets) of using $SuS$ with TIP-Adapter instead of $TIP - X$ in Tab. 4. We use both default ensemble prompts and CuPL prompts for CLIP's text classifier to break down the performance gains further. We note that both $SuS$ and $TIP - X$ are crucial for achieving the best results. + +Transfer to different visual backbones. We evaluate the scalability of our model across different CLIP visual backbones—Fig. 4c shows that both $SuS-X$ variants consistently improve upon zero-shot CLIP across ResNet and VisionTransformer backbones of varying depths and sizes. $SuS$ size. We study the effect of varying support set size for $SuS-LC$ and $SuS-SD$ —we generate three different support sets with random seeds for support sizes of 1, 5, 10, 25, 50, 75 and 100 samples. From Fig. 6, we observe two broad trends—some tasks benefit (ImageNet-R, DTD) from having more support set samples while others do not (Country211, Flowers102). We suggest that this is connected to the domain gap between the true data distribution and support set samples—if the domain gap is large, it is inimical to provide a large support set, whereas if the domains are similar, providing more support samples always helps. + +SuS visualisation. We visualise samples from both support set construction methods on ImageNet in Fig. 5. It is hard to distinguish between the true ImageNet samples and the SuS samples—we can therefore construct support sets to mimic the true data distribution, with access to only the category names. A caveat is that the support set does not always capture the domain characteristics of the true distribution, leading to a domain gap (lighting conditions, diverse scene backgrounds, confounding objects etc). To fully close the gap to using true few-shot datasets as support sets [25, 80], further research into exact unsupervised domain matching of support sets and few-shot datasets is required. + +Table 2: Training-free adaptation of CLIP on 19 datasets with RN50 visual backbone. The best and second best results for each dataset are bolded and underlined, respectively. Individual results for all 19 datasets are available in the supp. mat. *Average reported across 19 datasets. †Our re-implementation. + +
MethodAverage*ImageNet [16]ImageNet-R [35]ImageNet-Sketch [69]EuroSAT [34]DTD [13]Birdsnap [5]
Zero-shotZero-shot CLIP [57]52.2760.3159.3435.4226.8341.0130.56
CALIP [31]-60.57--38.9042.39-
CALIP [31]†52.3760.3159.3336.1026.9641.0230.68
CLIP+DN [85]53.0260.1660.3735.9528.3141.2131.23
Name-onlyCuPL [56]55.5061.4561.0235.1338.3848.6435.65
CuPL+e55.7661.6461.1735.8537.0647.4635.80
VisDesc [49]53.7659.6857.1633.7837.6041.9635.65
SuS-X-SD (ours)56.7361.8461.7636.3045.5750.5937.14
SuS-X-LC (ours)56.8761.8962.1037.8344.2349.2338.50
+ +![](images/32527ecad94a01223c8ddb456e081da891c38410cf32d9488d7416f6ea89bf0d.jpg) +Absolute improvement over Zero-shot CLIP +(a) + +![](images/14e7532a956a0d4f126bdd5b646c377c987386759076ca88628b58ab96a2171a.jpg) +Average over 11 datasets +(b) +Figure 4: (a) Comparison of $SuS-X$ with Zero-shot CLIP. (b) Results of training-free few-shot classification. (c) Performance comparison of $SuS-X$ across visual backbones. + +![](images/1d9e8944f682abeb4400ffb894d99697357319df89a6c5bc970779c41d06a7c2.jpg) +Effect of visual backbone +(c) + +Table 3: SuS- $X$ generalises to different VLMs. *Average reported across 19 datasets. + +
VLMMethodAverage*ImageNetEuroSATDTDBirdsnap
TCLZero-shot31.3835.5520.8028.554.51
CuPL34.7941.6026.3042.846.83
CuPL+e32.7941.3625.8841.966.60
VisDesc33.9440.4021.2734.285.69
SuS-X-SD41.4952.2928.7548.1713.60
SuS-X-LC42.7552.7736.9046.6317.93
BLIPZero-shot48.7350.5944.1044.6810.21
CuPL51.1152.9639.3752.9512.24
CuPL+e51.3653.0741.4853.3012.18
VisDesc49.9150.9442.2547.4511.69
SuS-X-SD53.2055.9345.3656.1516.95
SuS-X-LC54.6456.7551.6255.9123.78
+ +Prompting strategies for SuS construction. Tab. 5 depicts the performance of Photo and CuPL prompting—best results are achieved with the LC-Photo and SD-CuPL strategies. We further compare the diversity of images produced by the two strategies on ImageNet11—from Tab. 5, it is evident that CuPL prompting leads to more diverse support sets as compared to Photo prompting. + +Table 4: Component Analysis of SuS-X. + +
Text PromptsMethodSuSTIP-XAverage Accuracy
DefaultZero-shot CLIP✗✗52.27
SuS-TIP-SD✓✗53.49 (+1.22%)
SuS-X-SD✓✓53.69 (+1.42%)
SuS-TIP-LC✓✗53.83 (+1.56%)
SuS-X-LC✓✓54.20 (+1.93%)
CuPL+eCuPL+e✗✗55.76 (+3.49%)
SuS-TIP-SD✓✗56.63 (+4.36%)
SuS-X-SD✓✓56.73 (+4.46%)
SuS-TIP-LC✓✗56.72 (+4.45%)
SuS-X-LC✓✓56.87 (+4.60%)
+ +Table 5: Prompting strategies for SuS construction. + +
SuS methodAverage Acc.ImageNet Acc.Diversity
PhotoCuPLPhotoCuPLPhotoCuPL
LC56.8756.2061.8961.790.280.32
SD56.3256.7361.7961.840.170.20
+ +Hyperparameter Sensitivity. We perform a sensitivity test for our $\gamma$ hyperparameter (refer Eq. 8) on ImageNet-R, OxfordPets, and DTD. We fix $\alpha$ and $\beta$ to be 1, and run a sweep over $\gamma \in [0,1]$ . From Tab. 6, we observe that moderate values of $\gamma$ are typically preferred, and the variance of the accuracy values is small. However, note that for DTD, the optimal $\gamma$ is slightly larger (0.75)—this is due to its spe + +![](images/d18fa62e710fa2cb7bdec58c7a0d214552fe2345c45490258849e8a037db35fb.jpg) + +![](images/7f30e3fa30e53e68b9ce23215771a116122563d84db7204f8010c888dac109fe.jpg) + +![](images/7b59335d45e4d8ae8477b266f42e33e18740a87c6baf7a51242a32e438a20859.jpg) + +![](images/b98f1c92b5e2892eba622e4d81bfd2d82f68cf6de8eae6cba52859325bcab2e0.jpg) +(c) Australian Kelpie + +![](images/7bbe1df296493689cf41d0f4a87a5f0d977c0a0ffa8c3e81f49eb2dfcf105971.jpg) +(a) Dishwasher + +![](images/bd8bfe97689358daf3387e7fd04e30072befb116efe1a0efa44ffbb418875fa0.jpg) + +![](images/a7e6385cc38b8904f59e6796353a576b1b1c1789eaa42b8669fea3d155f08ea0.jpg) + +![](images/b09202a6b171409825e3cc4e4e40d6ba37459ada3e2d3eb9cf53153ae669ff1d.jpg) + +![](images/01fa8d65683e1a8b6bb88b83bd2e7cf2f0826c501aef031233969bda89be6e9c.jpg) + +![](images/752d4480fca02e1b01d71e0a0b8ce85c475ac7fe5c00a791f31839622248878b.jpg) +(b) Split Rail Fence + +![](images/5120b195c602f8fff73e92fe156e0335db8992f3cfdf8a0ea9232c75d5097a39.jpg) +(d) Bulbul + +![](images/1fc1e62077b2ec3595184dc4c257b07829677e278b3ff116aeddb945aefb471e.jpg) + +![](images/a991222a15d2d48b1787c4defc05f495bde3c4a6267215ffe0ae3ef44e323434.jpg) +Figure 5: Support samples from the generated SuS-SD, retrieved SuS-LC and true training distribution for ImageNet. By randomising the image order in each subfigure, we pose a challenge question—can you match the three images for each subfigure to their source i.e. SuS-SD, SuS-LC or ImageNet train set? The answers are provided at the bottom of the page10. + +![](images/8f138b98170b35893f61ec645b6d15b42384751217cfed536317025dcbfed285.jpg) + +![](images/2f22fb2837c8403b787a8c0ed1353bd4c590a790c6ab63157cb2f0e2f5a03dfb.jpg) +(a) Tasks where larger support sets are beneficial +Figure 6: Effect of support size. + +![](images/56b686f9a5827ef7d5a2028efd8d1b52d5eece0792d62ff5a35ab57078b9d30e.jpg) +(b) Tasks where larger support sets are harmful + +Table 6: Hyperparameter sensitivity for $\gamma$ + +
Datasetγ value
00.10.20.30.50.751
ImageNet-R60.8760.9861.0361.0561.0060.8960.65
OxfordPets76.7677.1777.5877.4477.1777.1776.90
DTD47.1647.1647.5147.6947.8747.9647.60
+ +cialised nature which requires more guidance from the specialised support set to inform pre-trained CLIP. Previous few-shot adaptation works [25, 80] observed similar results. For more hyperparameter ablations, see the supp. mat. + +# 4.5. Limitations and broader impact + +While demonstrating promising results, we note some limitations of our approach: (1) To perform name-only transfer, we rely on CLIP having seen related concepts during pre-training. For rare concepts not seen during pretraining, transfer might not be feasible. (2) We employ LAION-5B [61] as a source of knowledge. While reasonable for a proof of concept, this data is relatively uncurated and may contain harmful content. As such, our approach is unsuitable for real-world deployment without careful mitigation strategies to address this. Similar arguments apply to Stable Diffusion [59]. + +# 5. Conclusion + +In this paper, we studied the training-free name-only transfer paradigm for classification tasks with vision-language models. We systematically curated support sets with no access to samples from the target distribution and showed that they help improve CLIP's zero-shot predictions by providing rich, task-specific knowledge. We further motivated the $TIP-X$ framework through the observation that CLIP's intra-modal embedding spaces are not optimal for computing similarities. With these two building blocks, we demonstrated superior performance to prior state-of-the-art. + +Acknowledgements. This work was supported by the Isaac Newton Trust and an EPSRC access-to-HPC grant. SA would like to acknowledge the support of Z. Novak and N. Novak in enabling his contribution. VU would like to thank Gyungin Shin, Surabhi S. Nath, Jonathan Roberts, Vlad Bogolin, Kaiqu Liang and Anchit Jain for helpful discussions and feedback. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. +[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022. +[3] Manuele Barraco, Marcella Cornia, Silvia Cascianelli, Lorenzo Baraldi, and Rita Cucchiara. The unreasonable effectiveness of clip features for image captioning: An experimental analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4662-4670, 2022. +[4] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018. +[5] Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N Belhumeur. Bird-snap: Large-scale fine-grained visual categorization of birds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2011-2018, 2014. +[6] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. +[7] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European conference on computer vision, pages 446-461. Springer, 2014. +[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. +[9] Adrian Bulat and Georgios Tzimiropoulos. Language-aware soft prompting for vision & language foundation models. arXiv preprint arXiv:2210.01115, 2022. +[10] Delong Chen, Zhao Wu, Fan Liu, Zaiquan Yang, Yixiang Huang, Yiping Bao, and Erjin Zhou. Prototypical contrastive language image pretraining. arXiv preprint arXiv:2206.10996, 2022. +[11] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022. +[12] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6172-6180, 2018. +[13] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. + +[14] Andrea Cossu, Tinne Tuytelaars, Antonio Carta, Lucia Passaro, Vincenzo Lomonaco, and Davide Bacciu. Continual pre-training mitigates forgetting in language and vision. arXiv preprint arXiv:2205.09357, 2022. +[15] Guillaume Couairon, Matthijs Douze, Matthieu Cord, and Holger Schwenk. Embedding arithmetic of multimodal queries for image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4950-4958, 2022. +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[17] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Variational prompt tuning improves generalization of vision-language models. arXiv preprint arXiv:2210.02390, 2022. +[18] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. +[19] Kun Ding, Ying Wang, Pengzhang Liu, Qiang Yu, Haojian Zhang, Shiming Xiang, and Chunhong Pan. Prompt tuning with soft context sharing for vision-language models. arXiv preprint arXiv:2208.13474, 2022. +[20] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. +[21] Alex Fang, Gabriel Ilharco, Mitchell Wortsman, Yuhao Wan, Vaishaal Shankar, Achal Dave, and Ludwig Schmidt. Data determines distributional robustness in contrastive language image pre-training (clip). arXiv preprint arXiv:2205.01397, 2022. +[22] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. +[23] Robert Fergus, Li Fei-Fei, Pietro Perona, and Andrew Zisserman. Learning object categories from google's image search. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1816-1823. IEEE, 2005. +[24] Benjamin Feuer, Ameya Joshi, and Chinmay Hegde. Caption supervision enables robust learners. arXiv preprint arXiv:2210.07396, 2022. +[25] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. +[26] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022. + +[27] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021. +[28] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022. +[29] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. +[30] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. +[31] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. arXiv preprint arXiv:2209.14169, 2022. +[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[33] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. +[34] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. +[35] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. +[36] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. +[37] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. +[38] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell, 172(5):1122-1131, 2018. +[39] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. +[40] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. + +[41] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In 2009 IEEE conference on computer vision and pattern recognition, pages 951-958. IEEE, 2009. +[42] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. +[43] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. +[44] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022. +[45] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. arXiv preprint arXiv:2205.03340, 2022. +[46] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285, 2022. +[47] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. +[48] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. +[49] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. +[50] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. +[51] Thao Nguyen, Gabriel Ilharco, Mitchell Wortsman, Sewoong Oh, and Ludwig Schmidt. Quality not quantity: On the interaction between dataset design and robustness of clip. arXiv preprint arXiv:2208.05516, 2022. +[52] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. +[53] Zachary Novack, Saurabh Garg, Julian McAuley, and Zachary C Lipton. Chils: Zero-shot image classification with hierarchical label sets. arXiv preprint arXiv:2302.02551, 2023. +[54] Omiros Pantazis, Gabriel Brostow, Kate Jones, and Oisin Mac Aodha. Svl-adapter: Self-supervised adapter for vision-language pretrained models. arXiv preprint arXiv:2210.03794, 2022. +[55] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. + +[56] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. arXiv preprint arXiv:2209.03320, 2022. +[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. +[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. +[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjarn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[60] Shibani Santurkar, Yann Dubois, Rohan Taori, Percy Liang, and Tatsunori Hashimoto. Is a caption worth a thousand images? a controlled study for representation learning. arXiv preprint arXiv:2207.07635, 2022. +[61] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. +[62] Gyungin Shin, Weidi Xie, and Samuel Albanie. Named-mask: Distilling segmenters from complementary foundation models. arXiv preprint arXiv:2209.11228, 2022. +[63] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. arXiv preprint arXiv:2206.07045, 2022. +[64] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. +[65] Tejas Srinivasan, Ting-Yun Chang, Leticia Leonor Pinto Alva, Georgios Chochlakis, Mohammad Rostami, and Jesse Thomason. Climb: A continual learning benchmark for vision-and-language tasks. arXiv preprint arXiv:2206.09059, 2022. +[66] Vishaal Udandarao. Understanding and Fixing the Modality Gap in Vision-Language Models. Master's thesis, University of Cambridge, 2022. +[67] Vishaal Udandarao, Abhishek Maiti, Deepak Srivatsav, Suryatej Reddy Vyalla, Yifang Yin, and Rajiv Ratn Shah. Cobra: Contrastive bi-modal representation algorithm. arXiv preprint arXiv:2005.03687, 2020. +[68] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. +[69] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019. + +[70] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021. +[71] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. +[72] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. +[73] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. +[74] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. arXiv preprint arXiv:2207.12661, 2022. +[75] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. +[76] Youngjae Yu, Jiwan Chung, Heeseung Yun, Jack Hessel, JaeSung Park, Ximing Lu, Prithviraj Ammanabrolu, Rowan Zellers, Ronan Le Bras, Gunhee Kim, et al. Multimodal knowledge alignment with reinforcement learning. arXiv preprint arXiv:2205.12630, 2022. +[77] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021. +[78] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Hongsheng Li, Yu Qiao, and Peng Gao. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. arXiv preprint arXiv:2303.02151, 2023. +[79] Renrui Zhang, Longtian Qiu, Wei Zhang, and Ziyao Zeng. Vt-clip: Enhancing vision-language models with visual-guided texts. arXiv preprint arXiv:2112.02399, 2021. +[80] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. arXiv preprint arXiv:2207.09519, 2022. +[81] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D Manning, and Curtis P Langlotz. Contrastive learning of medical visual representations from paired images and text. arXiv preprint arXiv:2010.00747, 2020. +[82] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. arXiv preprint arXiv:2109.01134, 2021. + +[83] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. arXiv preprint arXiv:2203.05557, 2022. +[84] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. +[85] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser-Nam Lim. Distribution normalization: An "effortless" test-time augmentation for contrastively learned visual-language models. arXiv preprint arXiv:2302.11084, 2023. \ No newline at end of file diff --git a/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/images.zip b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..81e7943778515bc7374b868cb8963cec43f9c6bf --- /dev/null +++ b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:630df6989f0c97ae74eb1a9ef60726a45a18b84222ee4caac767aaa407f99c31 +size 651856 diff --git a/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/layout.json b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2180ef4309cffe450a28cb68aceaf6a95dfce185 --- /dev/null +++ b/2023/SuS-X_ Training-Free Name-Only Transfer of Vision-Language Models/layout.json @@ -0,0 +1,11510 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 516, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 516, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 516, + 121 + ], + "type": "text", + "content": "SuS-X: Training-Free Name-Only Transfer of Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 144, + 210, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 144, + 210, + 185 + ], + "spans": [ + { + "bbox": [ + 87, + 144, + 210, + 185 + ], + "type": "text", + "content": "Vishaal Udandarao University of Cambridge vu214@cam.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 236, + 144, + 358, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 144, + 358, + 185 + ], + "spans": [ + { + "bbox": [ + 236, + 144, + 358, + 185 + ], + "type": "text", + "content": "Ankush Gupta DeepMind, London ankushgupta@google.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 384, + 144, + 506, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 144, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 384, + 144, + 506, + 185 + ], + "type": "text", + "content": "Samuel Albanie \nUniversity of Cambridge \nsma71@cam.ac.uk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "type": "text", + "content": "Contrastive Language-Image Pre-training (CLIP) has emerged as a simple yet effective way to train large-scale vision-language models. CLIP demonstrates impressive zero-shot classification and retrieval performance on diverse downstream tasks. However, to leverage its full potential, fine-tuning still appears to be necessary. Finetuning the entire CLIP model can be resource-intensive and unstable. Moreover, recent methods that aim to circumvent this need for fine-tuning still require access to images from the target task distribution. In this paper, we pursue a different approach and explore the regime of training-free \"name-only transfer\" in which the only knowledge we possess about the downstream task comprises the names of downstream target categories. We propose a novel method, " + }, + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{SuS - X}" + }, + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "type": "text", + "content": ", consisting of two key building blocks—\"SuS\" and \"TIP-X\", that requires neither intensive finetuning nor costly labelled data. " + }, + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{SuS - X}" + }, + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "type": "text", + "content": " achieves state-of-the-art (SoTA) zero-shot classification results on 19 benchmark datasets. We further show the utility of TIP-X in the training-free few-shot setting, where we again achieve SoTA results over strong training-free baselines. Code is available at https://github.com/vishaal27/SuS-X." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 521, + 128, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 128, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 128, + 534 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 541, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 624 + ], + "type": "text", + "content": "Vision-language pre-training has taken the machine learning community by storm. A broad range of vision-language models (VLMs) [57, 42, 73, 1, 37] exhibiting exceptional transfer on tasks like classification [80, 84], cross-modal retrieval [67, 2] and segmentation [63, 27] have emerged. These models are now the de facto standard for downstream task transfer in the field of computer vision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 625, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 287, + 696 + ], + "type": "text", + "content": "One such prominent model, CLIP [57], is trained on 400M image-text pairs using a contrastive loss that maximises the similarities of paired image-text samples. CLIP pioneered the notion of zero-shot transfer in the vision-language setting: classification on unseen datasets. For a given classification task, CLIP converts the class labels" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 218, + 539, + 293 + ], + "blocks": [ + { + "bbox": [ + 313, + 218, + 539, + 293 + ], + "lines": [ + { + "bbox": [ + 313, + 218, + 539, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 539, + 293 + ], + "type": "image", + "image_path": "65ab95a0ab2821d0e722d4adc1e665a5f0696092fdea8374a9416f1748bdd02c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "type": "text", + "content": "Figure 1: Training-free name-only transfer. We propose " + }, + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "type": "text", + "content": ", a framework for enhancing the zero-shot transfer abilities of VLMs like CLIP [57], BLIP [42] and TCL [72], without training. To achieve this, we propose a novel method " + }, + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "type": "inline_equation", + "content": "TIP-X" + }, + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "type": "text", + "content": ", which adapts these VLMs using a curated support set (" + }, + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "type": "inline_equation", + "content": "SuS" + }, + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "type": "text", + "content": ") that is not drawn from the target distribution. Our " + }, + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "type": "inline_equation", + "content": "SuS" + }, + { + "bbox": [ + 304, + 307, + 547, + 405 + ], + "type": "text", + "content": " leverages one key piece of information about the task at hand: the names of the target categories." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 418, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 478 + ], + "type": "text", + "content": "into textual prompts (e.g. \"A photo of a .\", where represents the ground-truth text label for each class). It then computes similarities between all the class prompts and the query image, selecting the class with the highest image similarity as the predicted label (see Eq. (2))." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 479, + 546, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 546, + 635 + ], + "type": "text", + "content": "CLIP's zero-shot performance is however limited by its pre-training distribution [24, 60, 21, 51]. If the downstream dataset diverges significantly from the pretraining image distribution, CLIP's zero-shot performance drastically drops [21]. To mitigate this, several lines of work propose to adapt CLIP on diverse downstream tasks—Tab. 1 briefly summarises these methods. Most of them employ fine-tuning on either labelled or unlabelled subsets of data from the target task. However, fine-tuning such an overparameterised model can be unstable and lead to overfitting [15, 25]. Furthermore, having access to the true distribution of the target task can be prohibitive in data-scarce environments [12, 4, 38] and online learning settings [14, 65]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 635, + 547, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 547, + 673 + ], + "type": "text", + "content": "To alleviate these issues, in this paper, we aim to adapt CLIP and other VLMs for downstream classification in a name-only (requires only category names" + }, + { + "bbox": [ + 304, + 635, + 547, + 673 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 635, + 547, + 673 + ], + "type": "text", + "content": ", but no samples" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 287, + 712 + ], + "type": "text", + "content": "This idea of zero-shot transfer is distinct from the traditional zero-shot" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 683, + 545, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 545, + 703 + ], + "type": "text", + "content": "classification setup introduced by Lampert et al. [41] in which the task is to generalise to classes not seen during training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 703, + 503, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 503, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 503, + 713 + ], + "type": "text", + "content": "2We use category and class interchangeably in this paper." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2725" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 115, + 493, + 382 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": "Table 1: Taxonomy of CLIP adaptation methods for downstream classification. We underline the Zero-Shot CLIP model to signify that it is the base model that all others build on top of. *This method considers access to all test-set samples simultaneously, hence we still consider it zero-shot. †This method additionally uses class hierarchy maps." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 115, + 493, + 382 + ], + "lines": [ + { + "bbox": [ + 100, + 115, + 493, + 382 + ], + "spans": [ + { + "bbox": [ + 100, + 115, + 493, + 382 + ], + "type": "table", + "html": "
MethodDoes not require trainingDoes not require labelled dataDoes not require target data distribution
Few-shot fine-tuning methodsLP-CLIP [57]XXX
CoOp [84]XXX
PLOT [11]XXX
LASP [9]XXX
SoftCPT [19]XXX
VT-CLIP [79]XXX
VPT [17]XXX
ProDA [45]XXX
CoCoOp [83]XXX
CLIP-Adapter [25]XXX
Intermediate methodsTIP-Adapter [80]✓XX
UPL [36]X✓X
SVL-Adapter [54]X✓X
TPT [48]X✓✓
CLIP+SYN [33]X✓✓
CaFo [78]X✓✓
Zero-shot methodsZero-Shot CLIP [57]✓✓✓
CALIP [31]✓✓✓
CLIP+DN [85]*✓✓✓
Training-free name-only transfer methodsCuPL [56]✓✓✓
VisDesc [49]✓✓✓
CHiLS [53]†✓✓✓
SuS-X (ours)✓✓✓
", + "image_path": "d10902f247af84db3c73432e490a8bc03c6906c082fff66c805e6f8f4a06f450.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "type": "text", + "content": "from the target task) and training-free fashion. We propose " + }, + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "type": "text", + "content": " (see Fig. 1), consisting of two novel building blocks: (i) " + }, + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "type": "inline_equation", + "content": "SuS" + }, + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "type": "text", + "content": " (Support Sets), our dynamic support set curation strategy that forgoes the need for samples from the target task, and (ii) " + }, + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "type": "inline_equation", + "content": "TIP-X" + }, + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "type": "text", + "content": ", our main framework for performing zero-shot classification while being training-free. For a given downstream task, we first curate a support set by leveraging the task category labels, either in a parametric manner i.e., generating images from large-scale text-to-image models (e.g., Stable Diffusion [59]) or non-parametric manner i.e., retrieving real-world images from a large vision-language data bank (e.g., LAION-5B [61]). We then use the curated support set as a proxy few-shot dataset to inform our downstream predictions using " + }, + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "type": "inline_equation", + "content": "TIP-X" + }, + { + "bbox": [ + 45, + 396, + 288, + 575 + ], + "type": "text", + "content": ", in a similar vein to recent few-shot adaptation methods [25, 80]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Our extensive experiments show that " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " outperforms zero-shot methods on 19 benchmark datasets across three VLMs, namely, CLIP, BLIP and TCL by " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "4.60\\%" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "5.97\\%" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "11.37\\%" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " absolute average accuracy respectively. We further extend the " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "TIP-X" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " framework to the few-shot regime, outperforming previous SoTA methods in the training-free domain. Our main contributions are three-fold: (1) We propose " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", a SoTA method in the training-free name-only transfer setting for downstream adaptation of VLMs, (2) We present " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "SuS" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", an effective strategy for curating support sets using parametric or non-parametric methods to mitigate" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 396, + 547, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 444 + ], + "type": "text", + "content": "the lack of data samples available from the target task distribution, and (3) We propose " + }, + { + "bbox": [ + 304, + 396, + 547, + 444 + ], + "type": "inline_equation", + "content": "TIP-X" + }, + { + "bbox": [ + 304, + 396, + 547, + 444 + ], + "type": "text", + "content": ", a novel training-free method for adapting VLMs to downstream classification in both the name-only transfer and few-shot regimes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 453, + 392, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 392, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 392, + 467 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 474, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 641 + ], + "type": "text", + "content": "Vision-Language (VL) Foundation Models. Recent years have seen a Cambrian explosion in large-scale VL foundation models [6]. In a seminal work, Radford et al. [57] introduced CLIP, a large VLM trained on a web-scale corpus (400M image-text pairs), that exhibits strong downstream visual task performance. The introduction of CLIP inspired further development of VLMs [42, 1, 37, 18, 81, 75, 72, 10, 70, 26, 28, 43, 46, 74], each pre-trained on web-scale datasets to learn joint image-text representations. These representations can then be applied to tackle downstream tasks like semantic segmentation [63, 27], object detection [30, 20], image captioning [50, 3] and generative modelling [59, 58]. In this work, we adapt such VLMs in a training-free setting to diverse downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Adaptation of VL models. CLIP introduces a paradigm shift with its zero-shot transfer ability [57]. In this setup, none of the target dataset classes are known " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "-priori and the task is to adapt implicitly at inference time to a given dataset. Since CLIP's training objective drives it to assign appropriate similarities to image-text pairs, it acquires the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2726" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 250, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 250, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 250, + 83 + ], + "type": "text", + "content": "ability to perform zero-shot classification directly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 156 + ], + "type": "text", + "content": "Inspired by CLIP's zero-shot success, further work has sought to improve upon its performance. In Tab. 1, we characterise some of these methods along three major axes: (i) if the method requires training, (ii) if the method requires labelled samples from the target task, and (iii) if the method requires samples from the target task distribution3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 156, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 384 + ], + "type": "text", + "content": "In this work, we focus on the training-free name-only transfer regime—our goal is to adapt VLMs to target tasks without explicit training or access to samples from the target distribution. Instead, we assume access only to category names of target tasks. This formulation was recently considered for semantic segmentation, where it was called name-only transfer [62]—we likewise adopt this terminology. To the best of our knowledge, only two other concurrent approaches, CuPL [56] and VisDesc [49], operate in this regime. They use pre-trained language models to enhance textual prompts for zero-shot classification. By contrast, " + }, + { + "bbox": [ + 46, + 156, + 287, + 384 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 156, + 287, + 384 + ], + "type": "text", + "content": " pursues a support set curation strategy to adapt VLMs using knowledge of category names. These approaches are complementary, and we find that they can be productively combined. Two other related works operating purely in the zero-shot setting are: (1) CALIP [31], which uses parameter-free attention on image-text features, and (2) CLIP+DN [85], which uses distribution normalisation. We compare with these four baselines in Sec. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 392, + 282, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 282, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 282, + 406 + ], + "type": "text", + "content": "3. SuS-X: Training-Free Name-Only Transfer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 412, + 286, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 286, + 460 + ], + "type": "text", + "content": "We describe the two main building blocks of " + }, + { + "bbox": [ + 47, + 412, + 286, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{SuS - X}" + }, + { + "bbox": [ + 47, + 412, + 286, + 460 + ], + "type": "text", + "content": " (1) Support Set (SuS) construction, and (2) training-free inference using our novel TIP- " + }, + { + "bbox": [ + 47, + 412, + 286, + 460 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 412, + 286, + 460 + ], + "type": "text", + "content": " method. Fig. 2 depicts our overall training-free name-only transfer framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 468, + 150, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 150, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 150, + 480 + ], + "type": "text", + "content": "3.1. SuS Construction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 487, + 286, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 286, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 286, + 558 + ], + "type": "text", + "content": "We follow recent adaptation methods [80, 25] that use a small collection of labelled images to provide visual information to CLIP. However, differently from these methods, rather than accessing labelled images from the target distribution, we propose two methods (described next) to construct such a support set " + }, + { + "bbox": [ + 46, + 487, + 286, + 558 + ], + "type": "inline_equation", + "content": "(SuS)" + }, + { + "bbox": [ + 46, + 487, + 286, + 558 + ], + "type": "text", + "content": " without such access." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "type": "text", + "content": "(I) Stable Diffusion Generation. Our first method leverages the powerful text-to-image generation model, Stable Diffusion [59]. We employ specific prompting strategies for generating salient and informative support images. Concretely, given a set of downstream textual class labels, " + }, + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_1,t_2,\\dots ,t_C\\}" + }, + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "type": "text", + "content": " denotes the number of categories, we prompt Stable Diffusion to generate " + }, + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "type": "text", + "content": " images per class. In this way, we construct our support set of size " + }, + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "type": "inline_equation", + "content": "NC" + }, + { + "bbox": [ + 46, + 559, + 286, + 666 + ], + "type": "text", + "content": ", with each image having its associated class label." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 251 + ], + "type": "text", + "content": "By default, we prompt Stable Diffusion using the original CLIP prompts, i.e., \"A photo of a " + }, + { + "bbox": [ + 306, + 72, + 545, + 251 + ], + "type": "inline_equation", + "content": "<\\text{CLASS}>" + }, + { + "bbox": [ + 306, + 72, + 545, + 251 + ], + "type": "text", + "content": ",\" where " + }, + { + "bbox": [ + 306, + 72, + 545, + 251 + ], + "type": "inline_equation", + "content": "<\\text{CLASS}>" + }, + { + "bbox": [ + 306, + 72, + 545, + 251 + ], + "type": "text", + "content": " is the class text label. To further diversify the generation process, we follow CuPL [56] to first generate customised textual prompts for each class by prompting GPT-3 [8] to output descriptions of the particular class. We then feed this customised set of prompts output by GPT-3 into Stable Diffusion for generating images. For example, to generate images from the \"dog\" class, we prompt GPT-3 to describe \"dogs\", and then prompt Stable Diffusion with the resulting descriptions. In section 4.4, we compare the performance of the default (called Photo) and this augmented prompting procedure (called CuPL). Unless otherwise specified, all our experiments with Stable Diffusion support sets use the CuPL strategy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "text", + "content": "(II) LAION-5B Retrieval. Our second method leverages the large-scale vision-language dataset, LAION-5B [61]. It contains 5.85 billion image-text pairs, pre-filtered by CLIP. Using LAION-5B, we retrieve task-specific images using class text prompts for constructing the support set. Concretely, given textual class labels, " + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_1,t_2,\\dots ,t_C\\}" + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "text", + "content": ", we rank all images in LAION-5B by their CLIP image-text similarity to each text class label " + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "inline_equation", + "content": "i\\in [1,C]" + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "text", + "content": ". We then use the top " + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "text", + "content": " image matches as our support set for class " + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 252, + 545, + 502 + ], + "type": "text", + "content": ", resulting in an NC-sized support set of images with their associated class labels. Note that curating supporting knowledge by search is a classical technique in computer vision [23] that was recently revisited in the task of semantic segmentation [63]. Here we adapt this idea to the name-only transfer classification setting. For efficient retrieval, we leverage the approximate nearest neighbour indices released by the authors4. Similar to the Stable Diffusion generation approach, we experiment with both Photo and CuPL prompting strategies for curating our LAION-5B support set (see Sec. 4.4). By default, we use Photo prompting for all our experiments with LAION-5B support sets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "content": "Remark. SuS can be seen as a visual analogue to CuPL [56]—we augment VLMs with rich, class-specific images, instead of generating customised text descriptions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 545, + 402, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 402, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 402, + 558 + ], + "type": "text", + "content": "3.2. TIP-X Inference" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 647 + ], + "type": "text", + "content": "Given our support set from the previous section, we now leverage it in a training-free method to inform CLIP's zero-shot predictions. We first briefly review CLIP zero-shot classification and TIP-Adapter [80] (a training-free adaptation method). We then highlight a critical shortcoming in TIP-Adapter due to uncalibrated intra-modal embedding distances, which we address in our method—TIP-X." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": "Zero-shot CLIP. For classification into " + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": " classes, CLIP converts class labels into text prompts and encodes them with its text encoder. Collectively, the encoded prompt vectors can be interpreted as a classifier weight matrix " + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "inline_equation", + "content": "W \\in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 286, + 713 + ], + "type": "text", + "content": "3Note that (iii) subsumes (ii). (ii) refers to access to labelled data samples from the target dataset whereas (iii) refers to a more general setting where the samples from the target dataset can be unlabelled. We distinguish between the two for clarity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 702, + 488, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 488, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 488, + 713 + ], + "type": "text", + "content": "4https://huggingface.co/datasets/laion/laion5B-index" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2727" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 547, + 263 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 547, + 263 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 547, + 263 + ], + "type": "image", + "image_path": "77e4146946ebc30f1eacc2c17bfd9dd6cbdcc6e1c25465a38ed1eca0be48c405.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 547, + 344 + ], + "type": "text", + "content": "Figure 2: SuS-X for training-free name-only transfer. SuS-X consists of two core building blocks. (1) SuS (top right), a dynamic support set that we construct to infuse visual information into the VLM based only on knowledge of target category names. We construct support sets either in a parametric (generating images using Stable Diffusion) or non-parametric (retrieving images from LAION-5B) manner. (2) TIP-X (bottom right), our novel training-free method that leverages image-text distances to compute similarities between the support set and the test images. These similarities act as attention weights for the support set labels, and can directly be combined with the original logits from the VLM for classification." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 354, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 288, + 392 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times d}" + }, + { + "bbox": [ + 47, + 354, + 288, + 392 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 354, + 288, + 392 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 354, + 288, + 392 + ], + "type": "text", + "content": " is embedding dimension. For a test set " + }, + { + "bbox": [ + 47, + 354, + 288, + 392 + ], + "type": "inline_equation", + "content": "T = \\{y_{1},y_{2},\\dots,y_{t}\\}" + }, + { + "bbox": [ + 47, + 354, + 288, + 392 + ], + "type": "text", + "content": " comprising " + }, + { + "bbox": [ + 47, + 354, + 288, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 354, + 288, + 392 + ], + "type": "text", + "content": " test images, CLIP's image encoder is applied to produce test image features:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 397, + 287, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 287, + 418 + ], + "type": "interline_equation", + "content": "f _ {i} = \\operatorname {C L I P I m a g e E n c o d e r} \\left(y _ {i}\\right), i \\in [ 1, t ], f _ {i} \\in \\mathbb {R} ^ {d} \\tag {1}", + "image_path": "1dedb0faf46b1aff0dacaef9331a9019b2792f1ddef6407a6cacfb38cc3ac19a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 415, + 246, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 415, + 246, + 429 + ], + "spans": [ + { + "bbox": [ + 78, + 415, + 246, + 429 + ], + "type": "interline_equation", + "content": "f = \\operatorname {C o n c a t} ([ f _ {1}, f _ {2}, \\dots , f _ {t} ]), f \\in \\mathbb {R} ^ {t \\times d}", + "image_path": "5eb67ae2013a40f7b39fe7e2278fb6da9cafbd746dd3ae37feccf27162907fd3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 435, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 287, + 460 + ], + "type": "text", + "content": "Using " + }, + { + "bbox": [ + 47, + 435, + 287, + 460 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 435, + 287, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 435, + 287, + 460 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 435, + 287, + 460 + ], + "type": "text", + "content": ", CLIP performs classification by computing zero-shot logits (ZSL) via a dot product:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 466, + 287, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 466, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 138, + 466, + 287, + 480 + ], + "type": "interline_equation", + "content": "\\mathrm {Z S L} = f W ^ {T} \\tag {2}", + "image_path": "7106ff4c4d3fee8b62778eed21ee91a28f541809968b81707a25933457ca46e5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "type": "text", + "content": "TIP-Adapter. Given a " + }, + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "type": "inline_equation", + "content": "CK" + }, + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "type": "text", + "content": "-sized " + }, + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "type": "text", + "content": "-shot labelled dataset " + }, + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "type": "inline_equation", + "content": "D = \\{x_{1}, x_{2}, \\ldots, x_{CK}\\}^{5}" + }, + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "type": "text", + "content": " from the target domain, TIP-Adapter [80] encodes " + }, + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 487, + 287, + 525 + ], + "type": "text", + "content": " using CLIP's image encoder:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 530, + 282, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 530, + 282, + 544 + ], + "spans": [ + { + "bbox": [ + 52, + 530, + 282, + 544 + ], + "type": "interline_equation", + "content": "F _ {i} = \\operatorname {C L I P I m a g e E n c o d e r} \\left(x _ {i}\\right), i \\in [ 1, C K ], F _ {i} \\in \\mathbb {R} ^ {d}", + "image_path": "79bfaf68d78f784db37b6017da151b663cdd6ddc75b4fc876d6209e623e4023c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 546, + 287, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 546, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 69, + 546, + 287, + 571 + ], + "type": "interline_equation", + "content": "F = \\operatorname {C o n c a t} ([ F _ {1}, F _ {2}, \\dots , F _ {C K} ]), F \\in \\mathbb {R} ^ {C K \\times d} \\tag {3}", + "image_path": "085eefb1236b851009b47c25eeab6def58d452bb78fcbb3465c45cd7872a7630.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 572, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 287, + 608 + ], + "type": "text", + "content": "It then converts each of the few-shot class labels to one-hot vectors " + }, + { + "bbox": [ + 47, + 572, + 287, + 608 + ], + "type": "inline_equation", + "content": "L \\in \\mathbb{R}^{CK \\times C}" + }, + { + "bbox": [ + 47, + 572, + 287, + 608 + ], + "type": "text", + "content": ". Next, it computes an affinity matrix to capture the similarities between " + }, + { + "bbox": [ + 47, + 572, + 287, + 608 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 572, + 287, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 572, + 287, + 608 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 572, + 287, + 608 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 615, + 287, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 615, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 114, + 615, + 287, + 629 + ], + "type": "interline_equation", + "content": "A = \\exp (- \\beta (1 - f F ^ {T})) \\tag {4}", + "image_path": "32b8a13005246f83ffc8a510e96bfa55212fb2bbf3e3489de90ef56548d79c0c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 636, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 636, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 636, + 287, + 683 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 636, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 636, + 287, + 683 + ], + "type": "text", + "content": " is a hyperparameter that modulates \"sharpness\". Finally, these affinities are used as attention weights over " + }, + { + "bbox": [ + 47, + 636, + 287, + 683 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 636, + 287, + 683 + ], + "type": "text", + "content": " to produce logits that are blended with ZSL using a hyperparameter, " + }, + { + "bbox": [ + 47, + 636, + 287, + 683 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 636, + 287, + 683 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 125, + 683, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 683, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 125, + 683, + 287, + 696 + ], + "type": "interline_equation", + "content": "\\mathrm {T L} = \\alpha A L + f W ^ {T} \\tag {5}", + "image_path": "2b52f493e906a93934355071fdaff4392d92e2819f226165373ea041c7ac5e1f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": "Motivating TIP-X. TIP-Adapter benefits from the affinity computation between the test and few-shot image samples (Eq. (4)). This similarity is computed in CLIP's image space. However, prior research [76, 44, 66] has demonstrated the existence of a modality gap between CLIP's image and text spaces. This leads us to question if doing image-image similarity comparisons in CLIP's image space is optimal." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 450, + 546, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 546, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 546, + 701 + ], + "type": "text", + "content": "Fig. 3a shows the pairwise image-image, text-text and image-text cosine similarities of the ImageNet validation set CLIP embeddings. Clearly, the intra-modal and intermodal similarities are distributed differently—the intermodal similarities have small variance and mean, whereas the intra-modal similarities have larger means and variances. This mismatch happens because contrastive training of CLIP maximises the inter-modal cosine similarities of paired samples without regard to intra-modal similarities. This implies that the intra-image CLIP embedding similarities employed by TIP-Adapter may not reflect the true intra-image similarities. Fig. 3b illustrates this idea with a simple example. Consider two image embeddings that are required to be a distance " + }, + { + "bbox": [ + 304, + 450, + 546, + 701 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 450, + 546, + 701 + ], + "type": "text", + "content": " away from a particular text embedding. The two image embeddings can satisfy this condition by being very close to each other or very far apart from each other. Fig. 3b shows that this constraint can be satisfied by any two arbitrary points on a hypersphere of radius " + }, + { + "bbox": [ + 304, + 450, + 546, + 701 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 450, + 546, + 701 + ], + "type": "text", + "content": ". While we expect loose constraints to be imposed via transitivity, we nevertheless expect a lower quality of calibration in intra-modal (e.g., image-image) comparisons." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "TIP-X to the rescue. To get around the problem of un" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "text", + "content": "5Note that a " + }, + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "text", + "content": " -shot labelled dataset for " + }, + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "text", + "content": " classes has a size " + }, + { + "bbox": [ + 58, + 702, + 271, + 712 + ], + "type": "inline_equation", + "content": "CK" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2728" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 281, + 163 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 281, + 163 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 281, + 163 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 281, + 163 + ], + "type": "image", + "image_path": "dc5d814a059c4b7d2bca83c1126b28ed85053893ead31ef4d63b91debc11474a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 170, + 287, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 170, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 287, + 202 + ], + "type": "text", + "content": "(a) Intra-modal and inter-modal CLIP cosine similarities. We observe quite distinct intra-modal and inter-modal cosine similarity distributions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 206, + 282, + 306 + ], + "blocks": [ + { + "bbox": [ + 52, + 206, + 282, + 306 + ], + "lines": [ + { + "bbox": [ + 52, + 206, + 282, + 306 + ], + "spans": [ + { + "bbox": [ + 52, + 206, + 282, + 306 + ], + "type": "image", + "image_path": "8cc77245b77bba9ffb8e7684f50a85b6f76fe56b73638526eee2d8efcfed72c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 313, + 287, + 346 + ], + "lines": [ + { + "bbox": [ + 47, + 313, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 287, + 346 + ], + "type": "text", + "content": "(b) Intra-modal degrees of freedom. Different intra-modal similarities can satisfy same inter-modal constraints, leaving room for poor calibration." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 356, + 272, + 369 + ], + "lines": [ + { + "bbox": [ + 61, + 356, + 272, + 369 + ], + "spans": [ + { + "bbox": [ + 61, + 356, + 272, + 369 + ], + "type": "text", + "content": "Figure 3: Our two-fold analysis motivating TIP-X" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "text", + "content": "calibrated intra-modal embedding distances in TIP-Adapter, we propose to use inter-modal distances as a bridge. More specifically, rather than computing similarities between the test features " + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "inline_equation", + "content": "(f\\in \\mathbb{R}^{t\\times d})" + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "text", + "content": " and few-shot features " + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "inline_equation", + "content": "(F\\in \\mathbb{R}^{CK\\times d})" + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "text", + "content": " in the image embedding space " + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "inline_equation", + "content": "(fF^T)" + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "text", + "content": ", we use the image-text space. We first construct signatures by computing similarities of " + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "text", + "content": " with the text classifier weights " + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 388, + 287, + 472 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 479, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 479, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 89, + 479, + 287, + 506 + ], + "type": "interline_equation", + "content": "\\begin{array}{r} S = \\operatorname {s o f t m a x} \\left(F W ^ {T}\\right), S \\in \\mathbb {R} ^ {C K \\times C} \\\\ s = \\operatorname {s o f t m a x} \\left(f W ^ {T}\\right). s \\in \\mathbb {R} ^ {t \\times C} \\end{array} \\tag {6}", + "image_path": "e4f2a65891bc78444d8847c7da28092f5734d96c2fd40fc789774c369632775e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "type": "text", + "content": "These signatures comprise probability distributions encoding inter-modal affinities between the few-shot features and class text vectors, and likewise for the test features. We then construct our affinity matrix " + }, + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{t \\times CK}" + }, + { + "bbox": [ + 46, + 517, + 287, + 578 + ], + "type": "text", + "content": " by measuring the KL-divergence between the signatures as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 588, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 588, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 82, + 588, + 287, + 601 + ], + "type": "interline_equation", + "content": "M _ {i, j} = \\mathrm {K L} \\left(s _ {i} \\| S _ {j}\\right), i \\in [ 1, t ], j \\in [ 1, C K ] \\tag {7}", + "image_path": "e720624f13b0703c59fb44bbd1a40d63708d2339db87ec22d31fc092a7d2bdc2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "text", + "content": " test signature for the " + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "text", + "content": " test samples, and " + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "inline_equation", + "content": "S_j" + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "text", + "content": " few-shot signature. Since we are working with discrete probability distributions, we compute the KL-divergence as " + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\mathbb{KL}(P||Q) = \\sum_{i} P_i \\log \\frac{P_i}{Q_i}" + }, + { + "bbox": [ + 46, + 605, + 287, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": "The construction of the affinity matrix " + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": " can be seen as analogous to the affinity computation in TIP-Adapter (Eq. (4)). However, our affinity matrix construction removes direct reliance on the uncalibrated image-image similarities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "Finally, before using our affinity matrix " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " as attention weights for " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " (one-hot encoded class labels), we rescale (denoted by " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ") the values of " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " to have the same range (min, max values) as the TIP-Adapter affinities " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ". Further, since our affinity matrix " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " consists of KL-divergence values, the most similar samples will get small weights since their KL-divergence will be low (close to 0). To mitigate this, we simply negate the values in " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ". We then blend our predicted logits with TL using a scalar " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 184, + 545, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 184, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 351, + 184, + 545, + 198 + ], + "type": "interline_equation", + "content": "\\mathrm {T X L} = f W ^ {T} + \\alpha A L + \\gamma \\psi (- M) L \\tag {8}", + "image_path": "d047810cccc1a5e03ba8b052d78f57a2b795fd6f97bd87085fd4c640312d8b47.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 203, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 203, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 203, + 545, + 216 + ], + "type": "text", + "content": "The entire " + }, + { + "bbox": [ + 305, + 203, + 545, + 216 + ], + "type": "inline_equation", + "content": "TIP-X" + }, + { + "bbox": [ + 305, + 203, + 545, + 216 + ], + "type": "text", + "content": " method is shown in Fig. 2 (bottom right)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 221, + 484, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 221, + 484, + 233 + ], + "spans": [ + { + "bbox": [ + 306, + 221, + 484, + 233 + ], + "type": "text", + "content": "3.3. SuS-X: Combining SuS and TIP-X" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "text", + "content": "Since our constructed support sets act as pseudo few-shot datasets, we directly replace the few-shot features " + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "inline_equation", + "content": "TIP - X" + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "text", + "content": " framework with the features of our support set. We call our method " + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{SuS - X - LC}" + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "text", + "content": " if we combine " + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "inline_equation", + "content": "TIP - X" + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "text", + "content": " with the LAION-5B curated support set, and " + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{SuS - X - SD}" + }, + { + "bbox": [ + 304, + 239, + 545, + 335 + ], + "type": "text", + "content": " when combined with the Stable Diffusion generated support set. These methods enable training-free name-only adaptation of zero-shot VLMs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 344, + 387, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 387, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 387, + 357 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 364, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 425 + ], + "type": "text", + "content": "First, we evaluate " + }, + { + "bbox": [ + 304, + 364, + 545, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{SuS - X}" + }, + { + "bbox": [ + 304, + 364, + 545, + 425 + ], + "type": "text", + "content": " against strong baselines in the training-free zero-shot/name-only transfer regimes, across three VLMs. Next, we illustrate the adaptation of " + }, + { + "bbox": [ + 304, + 364, + 545, + 425 + ], + "type": "inline_equation", + "content": "TIP-X" + }, + { + "bbox": [ + 304, + 364, + 545, + 425 + ], + "type": "text", + "content": " into the few-shot training-free regime. Finally, we ablate and analyse our method to provide additional insights." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 430, + 534, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 534, + 442 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 534, + 442 + ], + "type": "text", + "content": "4.1. Training-free name-only transfer evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 448, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 592 + ], + "type": "text", + "content": "Datasets. For a comprehensive evaluation, we test on 19 datasets spanning a wide range of object, scene and fine-grained categories: ImageNet [16], StanfordCars [39], UCF101 [64], Caltech101 [22], Caltech256 [29], Flowers102 [52], OxfordPets [55], Food101 [7], SUN397 [71], DTD [13], EuroSAT [34], FGVCAircraft [47], Country211 [57], CIFAR-10 [40], CIFAR-100 [40], Birdsnap [5], CUB [68], ImageNet-Sketch [69] and ImageNet-R [35]. Previous few-shot adaptation methods [77, 25, 82] benchmark on a subset of 11 of these 19 datasets. We report results on the 19-dataset suite in the main paper and compare results using only the 11-dataset subset in the supp. mat." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 592, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 639 + ], + "type": "text", + "content": "Experimental Settings. We compare against six baselines. For zero-shot CLIP, we use prompt ensembling with 7 different prompt templates following " + }, + { + "bbox": [ + 304, + 592, + 545, + 639 + ], + "type": "inline_equation", + "content": "[57, 80]^6" + }, + { + "bbox": [ + 304, + 592, + 545, + 639 + ], + "type": "text", + "content": ". We run CuPL7, VisDesc8 (name-only transfer) and CLIP+DN9" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 645, + 545, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 545, + 683 + ], + "type": "text", + "content": "The 7 prompt templates are: \"itap of a .\", \"a origami .\", \"a bad photo of the .\", \"a photo of the large .\", \"a in a video game.\", \"art of the .\", and \"a photo of the small \"." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 683, + 436, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 683, + 436, + 693 + ], + "spans": [ + { + "bbox": [ + 317, + 683, + 436, + 693 + ], + "type": "text", + "content": "7https://github.com/sarahpratt/CuPL" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 693, + 527, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 693, + 527, + 703 + ], + "spans": [ + { + "bbox": [ + 317, + 693, + 527, + 703 + ], + "type": "text", + "content": "8https://github.com/sachit-menon/classify_by_description_release" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 703, + 512, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 703, + 512, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 703, + 512, + 712 + ], + "type": "text", + "content": "9https://github.com/fengyuli2002/distribution-normalization" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2729" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "(zero-shot transfer) using their official code. We also experiment with augmenting the CuPL prompts with the original prompt ensemble, and call it " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{CuPL + e}" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": ". For CALIP (zero-shot transfer), in the absence of public code at the time of writing, we aim to reproduce their results using our own implementation. For our proposed methods, we report results using both SuS-X-LC and SuS-X-SD. For both methods, we use a fixed number of support samples per dataset (see supp. mat. for details). For CALIP and SuS-X, we conduct a hyperparameter search on the dataset validation sets. In Sec. 4.4 we perform a hyperparameter sensitivity test for a fair evaluation. By default, we use the ResNet-50 [32] backbone as CLIP's image encoder for all models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": "Main Results. In Tab. 2, we compare both variants of " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": " with the baselines. We report an average across 19 datasets. We also include results on ImageNet, EuroSAT, DTD, Birdsnap, ImageNet-R and ImageNet-Sketch (results on all 19 datasets in the supp. mat.). " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": " methods outperform zero-shot CLIP by " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": " on average across all 19 datasets. We observe striking gains of " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": " on EuroSAT, DTD and Birdsnap respectively. We also outperform the SoTA training-free adaptation methods—CuPL+ensemble and VisDesc by " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": " on average respectively. To further probe where we attain the most gains, we plot the absolute improvement of our models over zero-shot CLIP in Fig. 4a. We observe large gains on fine-grained (Birdsnap, CUB, UCF101) and specialised (EuroSAT, DTD) datasets, demonstrating the utility of " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": " in injecting rich visual knowledge into zero-shot CLIP (additional fine-grained classification analysis in supp. mat.). We further compare " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": " to few-shot methods that use labelled samples from the true distribution in the supp. mat.—despite being at a disadvantage due to using no target distribution samples, " + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 228, + 289, + 491 + ], + "type": "text", + "content": " is still competitive with these methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 499, + 200, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 200, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 200, + 511 + ], + "type": "text", + "content": "4.2. Transfer to different VLMs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 518, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 627 + ], + "type": "text", + "content": "We evaluate transfer to VLMs other than CLIP, namely TCL [72] and BLIP [42]. We only retain image and text encoders of these models for computing features, while preserving all other experimental settings from Sec. 4.1. Tab. 3 shows our " + }, + { + "bbox": [ + 46, + 518, + 287, + 627 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 518, + 287, + 627 + ], + "type": "text", + "content": " methods strongly outperform all baseline methods across both VLMs—we improve on zero-shot models by " + }, + { + "bbox": [ + 46, + 518, + 287, + 627 + ], + "type": "inline_equation", + "content": "11.37\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 518, + 287, + 627 + ], + "type": "inline_equation", + "content": "5.97\\%" + }, + { + "bbox": [ + 46, + 518, + 287, + 627 + ], + "type": "text", + "content": " on average across 19 datasets. This demonstrates that our method is not specific to CLIP, but can improve performance across different VLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 222, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 222, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 222, + 647 + ], + "type": "text", + "content": "4.3. Adapting to the few-shot regime" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "A key component of our " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{SuS - X}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " method is " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "TIP - X" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ". In the previous section, we showcased SoTA results in the training-free name-only transfer regime. Due to its formulation, " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "TIP - X" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " can directly be extended to the few-shot regime, where our support sets are labelled samples from" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "the target dataset rather than curated/generated samples. To evaluate " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "TIP-X" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " on such real-world support sets, we conduct training-free few-shot classification using " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "TIP-X" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". We compare against the SoTA method in this regime—TIP-Adapter [80]. We report results on the 11-dataset subset used by TIP-Adapter on five different shot settings of the " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "-shot classification task: 1, 2, 4, 8 and 16." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 157, + 547, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 547, + 229 + ], + "type": "text", + "content": "We present average accuracy results on all shots in Fig. 4b—TIP-X outperforms both Zero-shot CLIP and TIP-Adapter (absolute gain of " + }, + { + "bbox": [ + 304, + 157, + 547, + 229 + ], + "type": "inline_equation", + "content": "0.91\\%" + }, + { + "bbox": [ + 304, + 157, + 547, + 229 + ], + "type": "text", + "content": " across shots). Notably, on OxfordPets, we achieve " + }, + { + "bbox": [ + 304, + 157, + 547, + 229 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 157, + 547, + 229 + ], + "type": "text", + "content": " average gain. This further demonstrates the generalisability of the TIP-X method in transferring to the few-shot training-free setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 238, + 369, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 369, + 250 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 369, + 250 + ], + "type": "text", + "content": "4.4. Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "text", + "content": "We conduct several ablations and provide additional visualisations to offer further insight into the " + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "inline_equation", + "content": "SuS - X" + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "text", + "content": " method. Component Analysis. " + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "inline_equation", + "content": "SuS - X" + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "text", + "content": " consists of two major building blocks—" + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "inline_equation", + "content": "SuS" + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "text", + "content": " construction and " + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "inline_equation", + "content": "TIP - X" + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "text", + "content": ". We compare the performance difference (with average accuracy across 19 datasets) of using " + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "inline_equation", + "content": "SuS" + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "text", + "content": " with TIP-Adapter instead of " + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "inline_equation", + "content": "TIP - X" + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "text", + "content": " in Tab. 4. We use both default ensemble prompts and CuPL prompts for CLIP's text classifier to break down the performance gains further. We note that both " + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "inline_equation", + "content": "SuS" + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "inline_equation", + "content": "TIP - X" + }, + { + "bbox": [ + 304, + 256, + 545, + 376 + ], + "type": "text", + "content": " are crucial for achieving the best results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "type": "text", + "content": "Transfer to different visual backbones. We evaluate the scalability of our model across different CLIP visual backbones—Fig. 4c shows that both " + }, + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "type": "text", + "content": " variants consistently improve upon zero-shot CLIP across ResNet and VisionTransformer backbones of varying depths and sizes. " + }, + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "type": "inline_equation", + "content": "SuS" + }, + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "type": "text", + "content": " size. We study the effect of varying support set size for " + }, + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "type": "inline_equation", + "content": "SuS-LC" + }, + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "type": "inline_equation", + "content": "SuS-SD" + }, + { + "bbox": [ + 304, + 377, + 545, + 570 + ], + "type": "text", + "content": "—we generate three different support sets with random seeds for support sizes of 1, 5, 10, 25, 50, 75 and 100 samples. From Fig. 6, we observe two broad trends—some tasks benefit (ImageNet-R, DTD) from having more support set samples while others do not (Country211, Flowers102). We suggest that this is connected to the domain gap between the true data distribution and support set samples—if the domain gap is large, it is inimical to provide a large support set, whereas if the domains are similar, providing more support samples always helps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "SuS visualisation. We visualise samples from both support set construction methods on ImageNet in Fig. 5. It is hard to distinguish between the true ImageNet samples and the SuS samples—we can therefore construct support sets to mimic the true data distribution, with access to only the category names. A caveat is that the support set does not always capture the domain characteristics of the true distribution, leading to a domain gap (lighting conditions, diverse scene backgrounds, confounding objects etc). To fully close the gap to using true few-shot datasets as support sets [25, 80], further research into exact unsupervised domain matching of support sets and few-shot datasets is required." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2730" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 108, + 541, + 220 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 106 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 106 + ], + "type": "text", + "content": "Table 2: Training-free adaptation of CLIP on 19 datasets with RN50 visual backbone. The best and second best results for each dataset are bolded and underlined, respectively. Individual results for all 19 datasets are available in the supp. mat. *Average reported across 19 datasets. †Our re-implementation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 108, + 541, + 220 + ], + "lines": [ + { + "bbox": [ + 49, + 108, + 541, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 108, + 541, + 220 + ], + "type": "table", + "html": "
MethodAverage*ImageNet [16]ImageNet-R [35]ImageNet-Sketch [69]EuroSAT [34]DTD [13]Birdsnap [5]
Zero-shotZero-shot CLIP [57]52.2760.3159.3435.4226.8341.0130.56
CALIP [31]-60.57--38.9042.39-
CALIP [31]†52.3760.3159.3336.1026.9641.0230.68
CLIP+DN [85]53.0260.1660.3735.9528.3141.2131.23
Name-onlyCuPL [56]55.5061.4561.0235.1338.3848.6435.65
CuPL+e55.7661.6461.1735.8537.0647.4635.80
VisDesc [49]53.7659.6857.1633.7837.6041.9635.65
SuS-X-SD (ours)56.7361.8461.7636.3045.5750.5937.14
SuS-X-LC (ours)56.8761.8962.1037.8344.2349.2338.50
", + "image_path": "e96f8a0401e24bcdcd977994b296a8ca67aad3ce5322c11042307c9dd13c4fa3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 235, + 211, + 339 + ], + "blocks": [ + { + "bbox": [ + 77, + 226, + 221, + 235 + ], + "lines": [ + { + "bbox": [ + 77, + 226, + 221, + 235 + ], + "spans": [ + { + "bbox": [ + 77, + 226, + 221, + 235 + ], + "type": "text", + "content": "Absolute improvement over Zero-shot CLIP" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 235, + 211, + 339 + ], + "lines": [ + { + "bbox": [ + 56, + 235, + 211, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 235, + 211, + 339 + ], + "type": "image", + "image_path": "32527ecad94a01223c8ddb456e081da891c38410cf32d9488d7416f6ea89bf0d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 342, + 140, + 352 + ], + "lines": [ + { + "bbox": [ + 129, + 342, + 140, + 352 + ], + "spans": [ + { + "bbox": [ + 129, + 342, + 140, + 352 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 225, + 235, + 367, + 340 + ], + "blocks": [ + { + "bbox": [ + 262, + 227, + 345, + 235 + ], + "lines": [ + { + "bbox": [ + 262, + 227, + 345, + 235 + ], + "spans": [ + { + "bbox": [ + 262, + 227, + 345, + 235 + ], + "type": "text", + "content": "Average over 11 datasets" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 225, + 235, + 367, + 340 + ], + "lines": [ + { + "bbox": [ + 225, + 235, + 367, + 340 + ], + "spans": [ + { + "bbox": [ + 225, + 235, + 367, + 340 + ], + "type": "image", + "image_path": "14e7532a956a0d4f126bdd5b646c377c987386759076ca88628b58ab96a2171a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 342, + 301, + 351 + ], + "lines": [ + { + "bbox": [ + 290, + 342, + 301, + 351 + ], + "spans": [ + { + "bbox": [ + 290, + 342, + 301, + 351 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 356, + 546, + 380 + ], + "lines": [ + { + "bbox": [ + 46, + 356, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 546, + 380 + ], + "type": "text", + "content": "Figure 4: (a) Comparison of " + }, + { + "bbox": [ + 46, + 356, + 546, + 380 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 356, + 546, + 380 + ], + "type": "text", + "content": " with Zero-shot CLIP. (b) Results of training-free few-shot classification. (c) Performance comparison of " + }, + { + "bbox": [ + 46, + 356, + 546, + 380 + ], + "type": "inline_equation", + "content": "SuS-X" + }, + { + "bbox": [ + 46, + 356, + 546, + 380 + ], + "type": "text", + "content": " across visual backbones." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 383, + 234, + 529, + 339 + ], + "blocks": [ + { + "bbox": [ + 421, + 226, + 504, + 234 + ], + "lines": [ + { + "bbox": [ + 421, + 226, + 504, + 234 + ], + "spans": [ + { + "bbox": [ + 421, + 226, + 504, + 234 + ], + "type": "text", + "content": "Effect of visual backbone" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 234, + 529, + 339 + ], + "lines": [ + { + "bbox": [ + 383, + 234, + 529, + 339 + ], + "spans": [ + { + "bbox": [ + 383, + 234, + 529, + 339 + ], + "type": "image", + "image_path": "1d9e8944f682abeb4400ffb894d99697357319df89a6c5bc970779c41d06a7c2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 342, + 462, + 351 + ], + "lines": [ + { + "bbox": [ + 451, + 342, + 462, + 351 + ], + "spans": [ + { + "bbox": [ + 451, + 342, + 462, + 351 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 56, + 425, + 277, + 566 + ], + "blocks": [ + { + "bbox": [ + 47, + 399, + 286, + 423 + ], + "lines": [ + { + "bbox": [ + 47, + 399, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 286, + 423 + ], + "type": "text", + "content": "Table 3: SuS- " + }, + { + "bbox": [ + 47, + 399, + 286, + 423 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 399, + 286, + 423 + ], + "type": "text", + "content": " generalises to different VLMs. *Average reported across 19 datasets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 425, + 277, + 566 + ], + "lines": [ + { + "bbox": [ + 56, + 425, + 277, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 425, + 277, + 566 + ], + "type": "table", + "html": "
VLMMethodAverage*ImageNetEuroSATDTDBirdsnap
TCLZero-shot31.3835.5520.8028.554.51
CuPL34.7941.6026.3042.846.83
CuPL+e32.7941.3625.8841.966.60
VisDesc33.9440.4021.2734.285.69
SuS-X-SD41.4952.2928.7548.1713.60
SuS-X-LC42.7552.7736.9046.6317.93
BLIPZero-shot48.7350.5944.1044.6810.21
CuPL51.1152.9639.3752.9512.24
CuPL+e51.3653.0741.4853.3012.18
VisDesc49.9150.9442.2547.4511.69
SuS-X-SD53.2055.9345.3656.1516.95
SuS-X-LC54.6456.7551.6255.9123.78
", + "image_path": "a7e4a781140fb074a71617f211bd1969958130f5df774f0daee5ff61d906bc1c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 664 + ], + "type": "text", + "content": "Prompting strategies for SuS construction. Tab. 5 depicts the performance of Photo and CuPL prompting—best results are achieved with the LC-Photo and SD-CuPL strategies. We further compare the diversity of images produced by the two strategies on ImageNet11—from Tab. 5, it is evident that CuPL prompting leads to more diverse support sets as compared to Photo prompting." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 318, + 413, + 531, + 535 + ], + "blocks": [ + { + "bbox": [ + 340, + 400, + 510, + 412 + ], + "lines": [ + { + "bbox": [ + 340, + 400, + 510, + 412 + ], + "spans": [ + { + "bbox": [ + 340, + 400, + 510, + 412 + ], + "type": "text", + "content": "Table 4: Component Analysis of SuS-X." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 413, + 531, + 535 + ], + "lines": [ + { + "bbox": [ + 318, + 413, + 531, + 535 + ], + "spans": [ + { + "bbox": [ + 318, + 413, + 531, + 535 + ], + "type": "table", + "html": "
Text PromptsMethodSuSTIP-XAverage Accuracy
DefaultZero-shot CLIP✗✗52.27
SuS-TIP-SD✓✗53.49 (+1.22%)
SuS-X-SD✓✓53.69 (+1.42%)
SuS-TIP-LC✓✗53.83 (+1.56%)
SuS-X-LC✓✓54.20 (+1.93%)
CuPL+eCuPL+e✗✗55.76 (+3.49%)
SuS-TIP-SD✓✗56.63 (+4.36%)
SuS-X-SD✓✓56.73 (+4.46%)
SuS-TIP-LC✓✗56.72 (+4.45%)
SuS-X-LC✓✓56.87 (+4.60%)
", + "image_path": "98e386620cbbbbd2a7111af07d3532202d5a796911ed39f1c21f078010b6a7ad.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 331, + 558, + 520, + 613 + ], + "blocks": [ + { + "bbox": [ + 317, + 544, + 535, + 556 + ], + "lines": [ + { + "bbox": [ + 317, + 544, + 535, + 556 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 535, + 556 + ], + "type": "text", + "content": "Table 5: Prompting strategies for SuS construction." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 558, + 520, + 613 + ], + "lines": [ + { + "bbox": [ + 331, + 558, + 520, + 613 + ], + "spans": [ + { + "bbox": [ + 331, + 558, + 520, + 613 + ], + "type": "table", + "html": "
SuS methodAverage Acc.ImageNet Acc.Diversity
PhotoCuPLPhotoCuPLPhotoCuPL
LC56.8756.2061.8961.790.280.32
SD56.3256.7361.7961.840.170.20
", + "image_path": "6ffcd4bd4ee769c3d2bf71bc63ad8556a8a537366113ecb186685d832dfd3e86.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "Hyperparameter Sensitivity. We perform a sensitivity test for our " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " hyperparameter (refer Eq. 8) on ImageNet-R, OxfordPets, and DTD. We fix " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " to be 1, and run a sweep over " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\gamma \\in [0,1]" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": ". From Tab. 6, we observe that moderate values of " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " are typically preferred, and the variance of the accuracy values is small. However, note that for DTD, the optimal " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " is slightly larger (0.75)—this is due to its spe" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 673, + 287, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 287, + 714 + ], + "type": "text", + "content": "11We compute diversity as 1 minus the mean of the average pairwise image cosine-similarities within a class. A larger value implies low cosine similarities across images within a class, implying more diverse images. Alternatively, a smaller value implies less diverse images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2731" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 69, + 141, + 144 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 141, + 144 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 141, + 144 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 141, + 144 + ], + "type": "image", + "image_path": "d18fa62e710fa2cb7bdec58c7a0d214552fe2345c45490258849e8a037db35fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 142, + 70, + 214, + 144 + ], + "blocks": [ + { + "bbox": [ + 142, + 70, + 214, + 144 + ], + "lines": [ + { + "bbox": [ + 142, + 70, + 214, + 144 + ], + "spans": [ + { + "bbox": [ + 142, + 70, + 214, + 144 + ], + "type": "image", + "image_path": "7f30e3fa30e53e68b9ce23215771a116122563d84db7204f8010c888dac109fe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 215, + 70, + 290, + 144 + ], + "blocks": [ + { + "bbox": [ + 215, + 70, + 290, + 144 + ], + "lines": [ + { + "bbox": [ + 215, + 70, + 290, + 144 + ], + "spans": [ + { + "bbox": [ + 215, + 70, + 290, + 144 + ], + "type": "image", + "image_path": "7b59335d45e4d8ae8477b266f42e33e18740a87c6baf7a51242a32e438a20859.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 159, + 141, + 233 + ], + "blocks": [ + { + "bbox": [ + 67, + 159, + 141, + 233 + ], + "lines": [ + { + "bbox": [ + 67, + 159, + 141, + 233 + ], + "spans": [ + { + "bbox": [ + 67, + 159, + 141, + 233 + ], + "type": "image", + "image_path": "b98f1c92b5e2892eba622e4d81bfd2d82f68cf6de8eae6cba52859325bcab2e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 139, + 237, + 217, + 248 + ], + "lines": [ + { + "bbox": [ + 139, + 237, + 217, + 248 + ], + "spans": [ + { + "bbox": [ + 139, + 237, + 217, + 248 + ], + "type": "text", + "content": "(c) Australian Kelpie" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 142, + 159, + 214, + 233 + ], + "blocks": [ + { + "bbox": [ + 149, + 148, + 206, + 158 + ], + "lines": [ + { + "bbox": [ + 149, + 148, + 206, + 158 + ], + "spans": [ + { + "bbox": [ + 149, + 148, + 206, + 158 + ], + "type": "text", + "content": "(a) Dishwasher" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 142, + 159, + 214, + 233 + ], + "lines": [ + { + "bbox": [ + 142, + 159, + 214, + 233 + ], + "spans": [ + { + "bbox": [ + 142, + 159, + 214, + 233 + ], + "type": "image", + "image_path": "7bbe1df296493689cf41d0f4a87a5f0d977c0a0ffa8c3e81f49eb2dfcf105971.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 215, + 159, + 289, + 233 + ], + "blocks": [ + { + "bbox": [ + 215, + 159, + 289, + 233 + ], + "lines": [ + { + "bbox": [ + 215, + 159, + 289, + 233 + ], + "spans": [ + { + "bbox": [ + 215, + 159, + 289, + 233 + ], + "type": "image", + "image_path": "bd8bfe97689358daf3387e7fd04e30072befb116efe1a0efa44ffbb418875fa0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 378, + 144 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 378, + 144 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 378, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 378, + 144 + ], + "type": "image", + "image_path": "a7e6385cc38b8904f59e6796353a576b1b1c1789eaa42b8669fea3d155f08ea0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 379, + 70, + 451, + 144 + ], + "blocks": [ + { + "bbox": [ + 379, + 70, + 451, + 144 + ], + "lines": [ + { + "bbox": [ + 379, + 70, + 451, + 144 + ], + "spans": [ + { + "bbox": [ + 379, + 70, + 451, + 144 + ], + "type": "image", + "image_path": "b09202a6b171409825e3cc4e4e40d6ba37459ada3e2d3eb9cf53153ae669ff1d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 452, + 70, + 526, + 144 + ], + "blocks": [ + { + "bbox": [ + 452, + 70, + 526, + 144 + ], + "lines": [ + { + "bbox": [ + 452, + 70, + 526, + 144 + ], + "spans": [ + { + "bbox": [ + 452, + 70, + 526, + 144 + ], + "type": "image", + "image_path": "01fa8d65683e1a8b6bb88b83bd2e7cf2f0826c501aef031233969bda89be6e9c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 304, + 159, + 378, + 233 + ], + "blocks": [ + { + "bbox": [ + 380, + 148, + 451, + 158 + ], + "lines": [ + { + "bbox": [ + 380, + 148, + 451, + 158 + ], + "spans": [ + { + "bbox": [ + 380, + 148, + 451, + 158 + ], + "type": "text", + "content": "(b) Split Rail Fence" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 159, + 378, + 233 + ], + "lines": [ + { + "bbox": [ + 304, + 159, + 378, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 378, + 233 + ], + "type": "image", + "image_path": "752d4480fca02e1b01d71e0a0b8ce85c475ac7fe5c00a791f31839622248878b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 379, + 159, + 451, + 233 + ], + "blocks": [ + { + "bbox": [ + 379, + 159, + 451, + 233 + ], + "lines": [ + { + "bbox": [ + 379, + 159, + 451, + 233 + ], + "spans": [ + { + "bbox": [ + 379, + 159, + 451, + 233 + ], + "type": "image", + "image_path": "5120b195c602f8fff73e92fe156e0335db8992f3cfdf8a0ea9232c75d5097a39.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 237, + 435, + 247 + ], + "lines": [ + { + "bbox": [ + 396, + 237, + 435, + 247 + ], + "spans": [ + { + "bbox": [ + 396, + 237, + 435, + 247 + ], + "type": "text", + "content": "(d) Bulbul" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 452, + 159, + 526, + 233 + ], + "blocks": [ + { + "bbox": [ + 452, + 159, + 526, + 233 + ], + "lines": [ + { + "bbox": [ + 452, + 159, + 526, + 233 + ], + "spans": [ + { + "bbox": [ + 452, + 159, + 526, + 233 + ], + "type": "image", + "image_path": "1fc1e62077b2ec3595184dc4c257b07829677e278b3ff116aeddb945aefb471e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 50, + 298, + 167, + 392 + ], + "blocks": [ + { + "bbox": [ + 46, + 252, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 545, + 289 + ], + "type": "text", + "content": "Figure 5: Support samples from the generated SuS-SD, retrieved SuS-LC and true training distribution for ImageNet. By randomising the image order in each subfigure, we pose a challenge question—can you match the three images for each subfigure to their source i.e. SuS-SD, SuS-LC or ImageNet train set? The answers are provided at the bottom of the page10." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 298, + 167, + 392 + ], + "lines": [ + { + "bbox": [ + 50, + 298, + 167, + 392 + ], + "spans": [ + { + "bbox": [ + 50, + 298, + 167, + 392 + ], + "type": "image", + "image_path": "a991222a15d2d48b1787c4defc05f495bde3c4a6267215ffe0ae3ef44e323434.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 167, + 299, + 285, + 392 + ], + "blocks": [ + { + "bbox": [ + 167, + 299, + 285, + 392 + ], + "lines": [ + { + "bbox": [ + 167, + 299, + 285, + 392 + ], + "spans": [ + { + "bbox": [ + 167, + 299, + 285, + 392 + ], + "type": "image", + "image_path": "8f138b98170b35893f61ec645b6d15b42384751217cfed536317025dcbfed285.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 52, + 411, + 165, + 503 + ], + "blocks": [ + { + "bbox": [ + 79, + 398, + 255, + 409 + ], + "lines": [ + { + "bbox": [ + 79, + 398, + 255, + 409 + ], + "spans": [ + { + "bbox": [ + 79, + 398, + 255, + 409 + ], + "type": "text", + "content": "(a) Tasks where larger support sets are beneficial" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 411, + 165, + 503 + ], + "lines": [ + { + "bbox": [ + 52, + 411, + 165, + 503 + ], + "spans": [ + { + "bbox": [ + 52, + 411, + 165, + 503 + ], + "type": "image", + "image_path": "2f22fb2837c8403b787a8c0ed1353bd4c590a790c6ab63157cb2f0e2f5a03dfb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 528, + 234, + 540 + ], + "lines": [ + { + "bbox": [ + 99, + 528, + 234, + 540 + ], + "spans": [ + { + "bbox": [ + 99, + 528, + 234, + 540 + ], + "type": "text", + "content": "Figure 6: Effect of support size." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 165, + 411, + 282, + 503 + ], + "blocks": [ + { + "bbox": [ + 165, + 411, + 282, + 503 + ], + "lines": [ + { + "bbox": [ + 165, + 411, + 282, + 503 + ], + "spans": [ + { + "bbox": [ + 165, + 411, + 282, + 503 + ], + "type": "image", + "image_path": "56b686f9a5827ef7d5a2028efd8d1b52d5eece0792d62ff5a35ab57078b9d30e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 510, + 253, + 520 + ], + "lines": [ + { + "bbox": [ + 82, + 510, + 253, + 520 + ], + "spans": [ + { + "bbox": [ + 82, + 510, + 253, + 520 + ], + "type": "text", + "content": "(b) Tasks where larger support sets are harmful" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 57, + 559, + 276, + 624 + ], + "blocks": [ + { + "bbox": [ + 77, + 547, + 256, + 559 + ], + "lines": [ + { + "bbox": [ + 77, + 547, + 256, + 559 + ], + "spans": [ + { + "bbox": [ + 77, + 547, + 256, + 559 + ], + "type": "text", + "content": "Table 6: Hyperparameter sensitivity for " + }, + { + "bbox": [ + 77, + 547, + 256, + 559 + ], + "type": "inline_equation", + "content": "\\gamma" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 559, + 276, + 624 + ], + "lines": [ + { + "bbox": [ + 57, + 559, + 276, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 276, + 624 + ], + "type": "table", + "html": "
Datasetγ value
00.10.20.30.50.751
ImageNet-R60.8760.9861.0361.0561.0060.8960.65
OxfordPets76.7677.1777.5877.4477.1777.1776.90
DTD47.1647.1647.5147.6947.8747.9647.60
", + "image_path": "fe0802df17aab380caf9072ce51466bb46d59924a0fd1f3fabc45641b2178f1c.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 639, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 286, + 689 + ], + "type": "text", + "content": "cialised nature which requires more guidance from the specialised support set to inform pre-trained CLIP. Previous few-shot adaptation works [25, 80] observed similar results. For more hyperparameter ablations, see the supp. mat." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 299, + 478, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 478, + 313 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 478, + 313 + ], + "type": "text", + "content": "4.5. Limitations and broader impact" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 321, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 452 + ], + "type": "text", + "content": "While demonstrating promising results, we note some limitations of our approach: (1) To perform name-only transfer, we rely on CLIP having seen related concepts during pre-training. For rare concepts not seen during pretraining, transfer might not be feasible. (2) We employ LAION-5B [61] as a source of knowledge. While reasonable for a proof of concept, this data is relatively uncurated and may contain harmful content. As such, our approach is unsuitable for real-world deployment without careful mitigation strategies to address this. Similar arguments apply to Stable Diffusion [59]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 472, + 378, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 378, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 378, + 484 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 495, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 625 + ], + "type": "text", + "content": "In this paper, we studied the training-free name-only transfer paradigm for classification tasks with vision-language models. We systematically curated support sets with no access to samples from the target distribution and showed that they help improve CLIP's zero-shot predictions by providing rich, task-specific knowledge. We further motivated the " + }, + { + "bbox": [ + 304, + 495, + 545, + 625 + ], + "type": "inline_equation", + "content": "TIP-X" + }, + { + "bbox": [ + 304, + 495, + 545, + 625 + ], + "type": "text", + "content": " framework through the observation that CLIP's intra-modal embedding spaces are not optimal for computing similarities. With these two building blocks, we demonstrated superior performance to prior state-of-the-art." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the Isaac Newton Trust and an EPSRC access-to-HPC grant. SA would like to acknowledge the support of Z. Novak and N. Novak in enabling his contribution. VU would like to thank Gyungin Shin, Surabhi S. Nath, Jonathan Roberts, Vlad Bogolin, Kaiqu Liang and Anchit Jain for helpful discussions and feedback." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 56, + 701, + 272, + 718 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 701, + 272, + 718 + ], + "spans": [ + { + "bbox": [ + 56, + 701, + 272, + 718 + ], + "type": "text", + "content": "11 13T'uRnLdS(p)dsT'uRnL(o)TnRnLdS(q)uRnLdS'oT(e)" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2732" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "type": "text", + "content": "[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 247 + ], + "type": "text", + "content": "[3] Manuele Barraco, Marcella Cornia, Silvia Cascianelli, Lorenzo Baraldi, and Rita Cucchiara. The unreasonable effectiveness of clip features for image captioning: An experimental analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4662-4670, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 282 + ], + "type": "text", + "content": "[4] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 339 + ], + "type": "text", + "content": "[5] Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N Belhumeur. Bird-snap: Large-scale fine-grained visual categorization of birds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2011-2018, 2014." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 396 + ], + "type": "text", + "content": "[6] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselat, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "text", + "content": "[7] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European conference on computer vision, pages 446-461. Springer, 2014." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 497 + ], + "type": "text", + "content": "[8] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 499, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 287, + 531 + ], + "type": "text", + "content": "[9] Adrian Bulat and Georgios Tzimiropoulos. Language-aware soft prompting for vision & language foundation models. arXiv preprint arXiv:2210.01115, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "text", + "content": "[10] Delong Chen, Zhao Wu, Fan Liu, Zaiquan Yang, Yixiang Huang, Yiping Bao, and Erjin Zhou. Prototypical contrastive language image pretraining. arXiv preprint arXiv:2206.10996, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "text", + "content": "[11] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. arXiv preprint arXiv:2210.01253, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[12] Gordon Christie, Neil Fendley, James Wilson, and Ryan Mukherjee. Functional map of the world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6172-6180, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Andrea Cossu, Tinne Tuytelaars, Antonio Carta, Lucia Passaro, Vincenzo Lomonaco, and Davide Bacciu. Continual pre-training mitigates forgetting in language and vision. arXiv preprint arXiv:2205.09357, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "text", + "content": "[15] Guillaume Couairon, Matthijs Douze, Matthieu Cord, and Holger Schwenk. Embedding arithmetic of multimodal queries for image retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4950-4958, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 218 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[17] Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor Guilherme Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. Variational prompt tuning improves generalization of vision-language models. arXiv preprint arXiv:2210.02390, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 547, + 319 + ], + "type": "text", + "content": "[18] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "text", + "content": "[19] Kun Ding, Ying Wang, Pengzhang Liu, Qiang Yu, Haojian Zhang, Shiming Xiang, and Chunhong Pan. Prompt tuning with soft context sharing for vision-language models. arXiv preprint arXiv:2208.13474, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 547, + 420 + ], + "type": "text", + "content": "[20] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 475 + ], + "type": "text", + "content": "[21] Alex Fang, Gabriel Ilharco, Mitchell Wortsman, Yuhao Wan, Vaishaal Shankar, Achal Dave, and Ludwig Schmidt. Data determines distributional robustness in contrastive language image pre-training (clip). arXiv preprint arXiv:2205.01397, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "text", + "content": "[22] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 534, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 587 + ], + "type": "text", + "content": "[23] Robert Fergus, Li Fei-Fei, Pietro Perona, and Andrew Zisserman. Learning object categories from google's image search. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 2, pages 1816-1823. IEEE, 2005." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 621 + ], + "type": "text", + "content": "[24] Benjamin Feuer, Ameya Joshi, and Chinmay Hegde. Caption supervision enables robust learners. arXiv preprint arXiv:2210.07396, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[25] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[26] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. arXiv preprint arXiv:2204.14095, 2022." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2733" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[27] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Open-vocabulary image segmentation. arXiv preprint arXiv:2112.12143, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 287, + 152 + ], + "type": "text", + "content": "[28] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. arXiv preprint arXiv:2205.14459, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 176 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 176 + ], + "type": "text", + "content": "[29] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 178, + 287, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 178, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 48, + 178, + 287, + 221 + ], + "type": "text", + "content": "[30] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 267 + ], + "type": "text", + "content": "[31] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. arXiv preprint arXiv:2209.14169, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 269, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 269, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 287, + 314 + ], + "type": "text", + "content": "[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "text", + "content": "[33] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 416 + ], + "type": "text", + "content": "[34] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 418, + 287, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 418, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 48, + 418, + 287, + 484 + ], + "type": "text", + "content": "[35] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 486, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 287, + 518 + ], + "type": "text", + "content": "[36] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 520, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 586 + ], + "type": "text", + "content": "[37] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 588, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 643 + ], + "type": "text", + "content": "[38] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell, 172(5):1122-1131, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 645, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 645, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 645, + 287, + 689 + ], + "type": "text", + "content": "[39] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[40] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[41] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In 2009 IEEE conference on computer vision and pattern recognition, pages 951-958. IEEE, 2009." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 172 + ], + "type": "text", + "content": "[42] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 227 + ], + "type": "text", + "content": "[43] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[44] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. arXiv preprint arXiv:2203.02053, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "text", + "content": "[45] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. arXiv preprint arXiv:2205.03340, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 348 + ], + "type": "text", + "content": "[46] Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. arXiv preprint arXiv:2207.07285, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "type": "text", + "content": "[47] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "type": "text", + "content": "[48] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "text", + "content": "[49] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "text", + "content": "[50] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "type": "text", + "content": "[51] Thao Nguyen, Gabriel Ilharco, Mitchell Wortsman, Sewoong Oh, and Ludwig Schmidt. Quality not quantity: On the interaction between dataset design and robustness of clip. arXiv preprint arXiv:2208.05516, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 581 + ], + "type": "text", + "content": "[52] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "text", + "content": "[53] Zachary Novack, Saurabh Garg, Julian McAuley, and Zachary C Lipton. Chils: Zero-shot image classification with hierarchical label sets. arXiv preprint arXiv:2302.02551, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "text", + "content": "[54] Omiros Pantazis, Gabriel Brostow, Kate Jones, and Oisin Mac Aodha. Svl-adapter: Self-supervised adapter for vision-language pretrained models. arXiv preprint arXiv:2210.03794, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[55] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2734" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[56] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. arXiv preprint arXiv:2209.03320, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 184 + ], + "type": "text", + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "type": "text", + "content": "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "text", + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjarn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "type": "text", + "content": "[60] Shibani Santurkar, Yann Dubois, Rohan Taori, Percy Liang, and Tatsunori Hashimoto. Is a caption worth a thousand images? a controlled study for representation learning. arXiv preprint arXiv:2207.07635, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "text", + "content": "[61] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 397, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 397, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 397, + 287, + 430 + ], + "type": "text", + "content": "[62] Gyungin Shin, Weidi Xie, and Samuel Albanie. Named-mask: Distilling segmenters from complementary foundation models. arXiv preprint arXiv:2209.11228, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 432, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 287, + 464 + ], + "type": "text", + "content": "[63] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. arXiv preprint arXiv:2206.07045, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "text", + "content": "[64] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 500, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 554 + ], + "type": "text", + "content": "[65] Tejas Srinivasan, Ting-Yun Chang, Leticia Leonor Pinto Alva, Georgios Chochlakis, Mohammad Rostami, and Jesse Thomason. Climb: A continual learning benchmark for vision-and-language tasks. arXiv preprint arXiv:2206.09059, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 588 + ], + "type": "text", + "content": "[66] Vishaal Udandarao. Understanding and Fixing the Modality Gap in Vision-Language Models. Master's thesis, University of Cambridge, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[67] Vishaal Udandarao, Abhishek Maiti, Deepak Srivatsav, Suryatej Reddy Vyalla, Yifang Yin, and Rajiv Ratn Shah. Cobra: Contrastive bi-modal representation algorithm. arXiv preprint arXiv:2005.03687, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "text", + "content": "[68] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[69] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, pages 10506-10518, 2019." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 115 + ], + "type": "text", + "content": "[70] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. arXiv preprint arXiv:2108.10904, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 174 + ], + "type": "text", + "content": "[71] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 176, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 545, + 241 + ], + "type": "text", + "content": "[72] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 244, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 545, + 288 + ], + "type": "text", + "content": "[73] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 289, + 545, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 344 + ], + "type": "text", + "content": "[74] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. arXiv preprint arXiv:2207.12661, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 346, + 545, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 346, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 307, + 346, + 545, + 389 + ], + "type": "text", + "content": "[75] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 393, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 448 + ], + "type": "text", + "content": "[76] Youngjae Yu, Jiwan Chung, Heeseung Yun, Jack Hessel, JaeSung Park, Ximing Lu, Prithviraj Ammanabrolu, Rowan Zellers, Ronan Le Bras, Gunhee Kim, et al. Multimodal knowledge alignment with reinforcement learning. arXiv preprint arXiv:2205.12630, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 450, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 450, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 450, + 545, + 494 + ], + "type": "text", + "content": "[77] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 496, + 545, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 496, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 307, + 496, + 545, + 549 + ], + "type": "text", + "content": "[78] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Hongsheng Li, Yu Qiao, and Peng Gao. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. arXiv preprint arXiv:2303.02151, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 553, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 553, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 553, + 545, + 586 + ], + "type": "text", + "content": "[79] Renrui Zhang, Longtian Qiu, Wei Zhang, and Ziyao Zeng. Vt-clip: Enhancing vision-language models with visual-guided texts. arXiv preprint arXiv:2112.02399, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "text", + "content": "[80] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. arXiv preprint arXiv:2207.09519, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 634, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 677 + ], + "type": "text", + "content": "[81] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D Manning, and Curtis P Langlotz. Contrastive learning of medical visual representations from paired images and text. arXiv preprint arXiv:2010.00747, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[82] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. arXiv preprint arXiv:2109.01134, 2021." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2735" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 195 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[83] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. arXiv preprint arXiv:2203.05557, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 286, + 149 + ], + "type": "text", + "content": "[84] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 195 + ], + "type": "text", + "content": "[85] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser-Nam Lim. Distribution normalization: An \"effortless\" test-time augmentation for contrastively learned visual-language models. arXiv preprint arXiv:2302.11084, 2023." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "2736" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/bc05a493-9765-4ef8-8b9c-fc9c90911ac5_content_list.json b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/bc05a493-9765-4ef8-8b9c-fc9c90911ac5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..57aa02b6089ad2d3abe0863bf64a6e7bcb5a6d24 --- /dev/null +++ b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/bc05a493-9765-4ef8-8b9c-fc9c90911ac5_content_list.json @@ -0,0 +1,1587 @@ +[ + { + "type": "text", + "text": "Subclass-balancing Contrastive Learning for Long-tailed Recognition", + "text_level": 1, + "bbox": [ + 133, + 130, + 834, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengkai Hou1, Jieyu Zhang2, Haonan Wang3, Tianyi Zhou4* \n1Jilin university, 2University of Washington, 3National University of Singapore \n4University of Maryland", + "bbox": [ + 173, + 179, + 795, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "houck20@mails.jlu.edu.cn, jieyuz2@cs.washington.edu, haonan.wang@u.nus.edu.tianyi@umd.edu", + "bbox": [ + 161, + 236, + 820, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long-tailed recognition with imbalanced class distribution naturally emerges in practical machine learning applications. Existing methods such as data reweighing, resampling, and supervised contrastive learning enforce the class balance with a price of introducing imbalance between instances of head class and tail class, which may ignore the underlying rich semantic substructures of the former and exaggerate the biases in the latter. We overcome these drawbacks by a novel \"subclass-balancing contrastive learning (SBCL)\" approach that clusters each head class into multiple subclasses of similar sizes as the tail classes and enforce representations to capture the two-layer class hierarchy between the original classes and their subclasses. Since the clustering is conducted in the representation space and updated during the course of training, the subclass labels preserve the semantic substructures of head classes. Meanwhile, it does not overemphasize tail class samples, so each individual instance contribute to the representation learning equally. Hence, our method achieves both the instance- and subclass-balance, while the original class labels are also learned through contrastive learning among subclasses from different classes. We evaluate SBCL over a list of long-tailed benchmark datasets and it achieves the state-of-the-art performance. In addition, we present extensive analyses and ablation studies of SBCL to verify its advantages. Our code is available at https://github.com/JackHck/subclass-balancing-contrastive-learning.", + "bbox": [ + 76, + 335, + 473, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In reality, the datasets often follow the Zipfian distribution over classes with a long tail [57, 75], i.e., a few classes (head classes) containing significantly more instances than the remaining tail classes. Such tail classes could be of great importance for high-stake applications, e.g., patient", + "bbox": [ + 75, + 801, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "class in medical diagnosis or accident class in autonomous driving [6, 55]. However, training on such class-imbalanced datasets can result in a severely biased model with noticeable performance drop in classification tasks [2, 4, 13, 47, 64, 68, 71].", + "bbox": [ + 500, + 305, + 893, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome the challenges posed by long-tailed data, data resampling [2, 4, 9, 54] and loss reweighing [5, 6, 15, 17] have been widely applied but they cannot fully leverage all the head-class samples. Very recent work discovered that supervised contrastive learning (SCL) [36] can achieve state-of-the-art (SOTA) performance on benchmark datasets of long-tailed recognition [33, 41]. Specifically, the $k$ -positive contrastive learning (KCL) [33] and its subsequent work targeted supervised contrastive learning (TSC) [41] revamp SCL by encouraging the learned feature space to be class-balanced and uniformly distributed. However, those methods enforcing class-balance often come with a price of instance-imbalance, i.e., each individual instance of tail classes would have much greater impact on model training than that of head classes.", + "bbox": [ + 496, + 383, + 895, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Such instance-imbalance can result in significant degradation of the performance on long-tailed recognition for several reasons. On the one hand, the limited samples in each tail class might not be sufficiently representative of the whole class. So even a small bias of them can be enormously exaggerated by class-balancing methods and result in sub-optimal learning of classifiers or representations. On the other hand, head classes usually have more complicated semantic substructures, e.g., multiple high-density regions of the data distribution, so simply downweighing samples of head classes and treating them equally can easily lose critical structural information. For example, images of a head class \"cat\" might be highly diverse in breeds and colors, which need to be captured by different features but downweighing or subsampling them may easily lose such information, while a tail class \"platypus\" might only contain a few similar images that are unlikely to cover all the representative features. Therefore, it is non-trivial to enforce both class-balance and instance-balance simultaneously in the same method.", + "bbox": [ + 496, + 613, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5395", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can we remove the negative impact of class-imbalance while still retain the advantages of instance-balance? In this paper, we achieve both through subclass-balancing contrastive learning (SBCL), a novel supervised contrastive learning defined on subclasses, which are the clusters within each head class, have comparable size as tail classes, and are adaptively updated during the training. Instead of sacrificing instance-balance for class-balance, our method achieves both instance- and subclass-balance by exploring the head-class structure in the learned representation space of the model-in-training. In particular, we propose a bi-granularity contrastive loss that enforces a sample (1) to be closer to samples from the same subclass than all the other samples; and (2) to be closer to samples from a different subclass but the same class than samples from any other subclasses. While the former learns representations with balanced and compact subclasses, the latter preserves the class structure on subclass level by encouraging the same class's subclasses to be closer to each other than to any different class's subclasses. Hence, it can learn an accurate classifier distinguishing original classes while enjoy both the instance- and subclass-balance.", + "bbox": [ + 75, + 90, + 472, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we apply SBCL for several visual recognition tasks to demonstrate SBCL superiority over other previous works (e.g., KCL [33], TSC [41]). To summarize, this paper makes the following contributions:", + "bbox": [ + 76, + 407, + 470, + 469 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a). We provide a new design principal of leveraging supervised contrastive learning for long-tailed recognition, i.e., aiming at achieving both instance- and subclass-balance instead of class-balance at the expense of instance-balance.", + "(b). We propose a novel instantiation of the aforementioned design principal, subclass-balancing contrastive learning (SBCL), which consists of two major components, namely, subclass-balancing adaptive clustering and bi-granularity contrastive loss.", + "(c). Empirically, we compare the SBCL against state-of-the-art methods on three visual tasks: image classification, object detection, and instance segmentation to demonstrate its effectiveness on handling class imbalance. We also conduct a series of experiments to analyze the efficacy of SBCL." + ], + "bbox": [ + 78, + 478, + 472, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and notations", + "text_level": 1, + "bbox": [ + 76, + 753, + 323, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long-tailed recognition. Long-tailed recognition aims to learn a classifier from a training dataset with long-tailed class distribution, i.e., a few classes contain many data (head classes) while most classes contain only a few data (tail classes), where the major challenge is to require model recognizing all classes equally well. Let $\\mathcal{D} = \\{x_i, y_i\\}_{i \\in [n]}$ be a long-tailed training dataset, where $x_i$ denotes a sample and $y_i \\in [C]$ denotes its label. Denote by $\\mathcal{D}_k \\subseteq \\mathcal{D}$ the set of", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "instances belonging to class $k$ and $n_k = |\\mathcal{D}_k|$ is the number of samples in class $k$ . The total number of training samples over $C$ classes is $n = \\sum_{k=1}^{C} n_k$ . Without loss of generality, we follow prior work [28, 34] to assume that the classes are sorted by cardinality in decreasing order (i.e., if $i < j$ , then $n_i \\geq n_j$ ), and $n_1 \\gg n_C$ . In addition, we define the imbalance ratio as $\\max_{k \\in [C]} (n_k) / \\min_{k \\in [C]} (n_k) = n_1 / n_C$ . Finally, let $f_\\theta(\\cdot)$ be a deep feature extractor, e.g., a neural network, parameterized by $\\theta$ and $\\mathbf{w}_c$ is the linear classifier of class $c$ , then the classifier we aim to learn is $h(x_i) = \\arg \\max_{c \\in [C]} \\mathbf{w}_c^\\top f_\\theta(x_i)$ .", + "bbox": [ + 496, + 90, + 895, + 261 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised contrastive learning. Recent studies have shown that supervised contrastive learning (SCL) [36] provides a strong performance gain for long-tailed recognition and its variants have achieved state-of-the-art (SOTA) performance [33, 41]. Specifically, SCL learns the feature extractor $f_{\\theta}(\\cdot)$ via maximizing the discriminativeness of positive instances, i.e., instances from the same class, and the learning objective for a single training data $(x_i, y_i)$ in a batch $\\mathcal{B} = \\{(x_i, y_i)\\}_{i \\in [N]}$ , is", + "bbox": [ + 496, + 275, + 893, + 414 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S C L} = \\sum_ {i = 1} ^ {N} - \\frac {1}{| \\tilde {P} _ {i} |} \\sum_ {z _ {p} \\in \\tilde {P} _ {i}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau\\right)}{\\sum_ {z _ {a} \\in \\tilde {V} _ {i}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 425, + 890, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature hyperparameter, $z_{i} = f_{\\theta}(x_{i})$ is the feature generated from $x_{i}$ , $V_{i} = \\{z_{i}\\}_{i\\in [N]}\\setminus \\{z_{i}\\}$ is the current batch of features except for $z_{i}$ , $P_{i} = \\{z_{j}\\in V_{i}:y_{j} = y_{i}\\}$ is a set of instances with the same label as $x_{i}$ . Finally, let $\\tilde{z}_i$ be the feature of $\\tilde{x}_i$ , the augmented version of $x_{i}$ , and for any set $S_{i}$ indexed by $i$ , we use $\\tilde{S}_i = S_i\\cup \\{\\tilde{z}_i\\}$ , e.g., $\\tilde{V}_i = V_i\\cup \\{\\tilde{z}_i\\}$ . However, for the long-tailed datasets, the feature spaces is dominated by head classes and thus have limited capability of semantic discrimination [33]. To address this, the $k$ -positive contrastive learning (KCL) [33] attempts to balance the feature space by keeping the number of positive instances in $\\tilde{P}_i$ equal for each class, leading to the following loss", + "bbox": [ + 496, + 477, + 893, + 675 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K C L} = \\sum_ {i = 1} ^ {N} - \\frac {1}{k + 1} \\sum_ {z _ {p} \\in \\tilde {P _ {i} ^ {k}}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau\\right)}{\\sum_ {z _ {a} \\in \\tilde {V _ {i}}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau\\right)} (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 684, + 890, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $P_{i}^{k}$ is a subset of $P_{i}$ with $k$ randomly drawn instances. Finally, the learned feature extractor $f_{\\theta}(\\cdot)$ is exploited in a sequel stage of training the classifier for long-tailed recognition [33, 41].", + "bbox": [ + 496, + 739, + 893, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 814, + 635, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As mentioned above, KCL and its sequels [33,41] balance the learning objective of SCL by picking the same number of positive instances for each class, i.e., $|P_i^k| = k$ in Eq. 2 no matter which class $x_i$ belongs to, however, we argue that", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5396", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f6d07b822ac00f11d4981192a1450f50fb0540bac0dc65db54ec7be3be90840d.jpg", + "image_caption": [ + "(a). Subclass-balancing adaptive clustering (Section 3.1)" + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 477, + 277 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2ec4e63648de9990919d8b2dbdd950dd170ee77df32b54e9bdf0b2be7a492cf6.jpg", + "image_caption": [ + "(b). Bi-granularity contrastive loss (Section 3.2)", + "Figure 1: Illustration of subclass-balancing contrastive learning (SBCL). It initially divides the head classes into multiple subclasses of comparable size. Then, during training, SBCL builds each sample to be closer to samples from the same subclass than samples from different subclasses but the same class, which are also made to be closer than samples from different classes." + ], + "image_footnote": [], + "bbox": [ + 480, + 90, + 870, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "such a class-balancing approach would inevitably introduce instance-imbalance: the instances of tail classes have much more chances to be engaged in the training than that of head class. Specifically, assume each class has no less than $k$ instances, then the probability of an instance of class $c$ being selected as positive instance is $p(c) = \\frac{k}{n_c}$ ; if the tail class $n_C$ has only $k$ instances and the imbalance ratio is $\\frac{n_1}{n_C} = 100$ , then we have $p(1) = \\frac{k}{100k} = 0.01$ while $p(C) = 1$ . We can see that when the instances of head class are selected once, that of tail class may already be trained 100 times. Thus, the training is immensely biased towards the few samples in each tail class. Besides, as tail classes only have very few instances that are not necessarily representative, the learned feature space might be unsatisfactory and sensitive to the training data of tail classes.", + "bbox": [ + 75, + 373, + 468, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we provide a new prospective of handling class-imbalance issue by contrastive learning: instead of aiming at class-balance at the expense of instance-imbalance, we propose to achieve both instance- and subclass-balance. We argue that head classes typically contain more diverse instances and thus have richer semantics in the training dataset. Therefore, it might be wise to break down the head classes into multiple semantically coherent subclasses, each of which consists of similar number of instance as tail classes. Built on this spirit, we develop subclass-balancing contrastive learning (SBCL), a new contrastive learning framework for long-tailed recognition (visualized in Figure 1) that achieves both instance- and subclass-balance. We present a concrete example that effectively demonstrates the superiority of SBCL compared to previous methods in Section 4.6.", + "bbox": [ + 75, + 604, + 470, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Subclass-balancing adaptive clustering", + "text_level": 1, + "bbox": [ + 76, + 845, + 410, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We break down the class into several \"subclasses\" to attack the imbalanced phenomenon. Particularly, given a class", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$c$ and the associated set of data $\\mathcal{D}_c$ , we employ a clustering algorithm of choice based on the features extracted by current feature extractor $f_{\\theta}(\\cdot)$ to divide $\\mathcal{D}_c$ into $m_c$ subclasses/clusters. We use $\\Gamma_c(x_i)$ to denote the cluster label of an instance $x_i$ of class $c$ . To ensure that the number of samples for each subclass is roughly the same, we propose a new cluster algorithm to divide the unit-length feature vectors, i.e., the features output by $f_{\\theta}(\\cdot)$ with additional unit-length normalization. The new proposed cluster algorithm is described in Algorithm 1.", + "bbox": [ + 496, + 373, + 893, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Algorithm 1, during the initialization of cluster centers, our approach gradually selects cluster centers by prioritizing the samples that are maximally distant from other existing cluster centers [3]. In the process of assigning vectors to their centers, we set a threshold $M$ as an upper limit of sample size in a cluster, which guarantees clusters of balanced sizes.", + "bbox": [ + 496, + 523, + 890, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the threshold $M$ is", + "bbox": [ + 517, + 614, + 730, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM = \\max \\left(n _ {C}, \\delta\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 637, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n_C$ is the size of tail class and the hyperparameter $\\delta$ controls the lower bound of sample size in clusters to prevent overly-small cluster. Note that we only apply clustering algorithm to classes which contain multiple instances while the tail classes remain unchanged. As a consequence, the size of each resultant cluster is similar to that of tail class $n_C$ . In addition, instead of only clustering once at the beginning, we update the cluster assignment adaptively based on the current feature extractor $f_{\\theta}(\\cdot)$ during training process and empirically show that such adaptive clustering outperforms only-once clustering in Section 4.5. We add an exhaustive description of Algorithm 1 and show the distribution of the sample size in clusters on the benchmark dataset in Appendix A.1.", + "bbox": [ + 496, + 657, + 890, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, by replacing the class labels of head classes used in SCL/KCL with the finer-grained cluster labels, we ensure the", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5397", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Subclass-balancing Adaptive Clustering" + ], + "code_body": "Require: Sample set $S = \\{x_{i}\\}_{i = 1}^{n}$ ; A threshold $M$ ; The number of iterations $K$ \nEnsure: Cluster assignments for samples in $s$ \nfor $k = 0$ to $K$ do if $k = 0$ then Choose the cluster centers $y_{j}$ which are farther away from previously selected centers. else Update the cluster centers $y_{j} = \\frac{1}{n_{j}}\\sum_{i = 1}^{n_{j}}x_{i}.$ $\\triangleright n_j$ is the number of samples in a cluster end if Construct the cluster center set $\\mathcal{C} = \\{y_j\\}_{j = 1}^m$ . $\\triangleright m$ is the number of cluster centers while $S\\neq \\phi$ do $\\triangleright$ Assign samples to centers $y_{i}$ Select the most similar pair $(x_{i},y_{j}) = \\arg \\max \\mathrm{cosine - similarity}(x,y)$ $x\\in S,y\\in C$ Assign the sample $x_{i}$ to the center $y_{j}$ . Delete the assigned sample $x_{i}$ from the sample set $S = S / \\{x_i\\}$ if $n_j\\geq M$ then Sample number in a cluster exceeds the threshold $M$ Delete the cluster center $y_{j}$ from the cluster center set $\\mathcal{C} = \\mathcal{C} / \\{y_j\\}$ end if end while end for", + "bbox": [ + 76, + 108, + 890, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "instance-balanced, i.e., each instance has similar chance of being selected regardless of its class. By breaking down the head classes, which typically contain more diverse instances, into multiple semantically coherent subclasses, we achieve subclass-balanced (instead of class-balanced) while maintain the rich semantics rendered by head classes in training dataset.", + "bbox": [ + 75, + 397, + 470, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Bi-granularity contrastive loss", + "text_level": 1, + "bbox": [ + 76, + 513, + 344, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now have two types of label for instances in head classes from different granularities: the coarse-grained class label and the fine-grained cluster label. A direct consequence of replacing class label in SCL/KCL with cluster label is that we no longer distinguish instances from different head classes, and therefore the boundaries between classes might be blurry, leading to sub-optimal feature space. As a remedy, we combine the contrastive loss of both class label and cluster label into the following one and reuse the notations of Eq. 1:", + "bbox": [ + 75, + 537, + 468, + 689 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {S B C L} = - \\sum_ {i = 1} ^ {N} \\left(\\frac {1}{| \\tilde {M} _ {i} |} \\sum_ {z _ {p} \\in \\tilde {M} _ {i}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau_ {1}\\right)}{\\sum z _ {a} \\in \\tilde {V} _ {i}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau_ {1}\\right) \\right. \\\\ + \\beta \\frac {1}{\\left| \\tilde {P} _ {i} \\right| - \\left| M _ {i} \\right|} \\sum_ {z _ {p} \\in \\tilde {P} _ {i} / M _ {i}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau_ {2}\\right)}{\\sum_ {z _ {a} \\in \\tilde {V} _ {i} / M _ {i}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau_ {2}\\right)}) \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 713, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{i} = \\{z_{j}\\in P_{i}:\\Gamma_{y_{i}}(x_{i}) = \\Gamma_{y_{i}}(x_{j})\\}$ is a set of instances with the same cluster label as $x_{i}$ . $\\beta$ is a hyperparameter that balances these two loss terms. The first term corresponds to the SCL loss with cluster labels, while the second term leverages the class label but does not consider the instances of the same cluster, i.e., the instances in $M_{i}$ are", + "bbox": [ + 75, + 809, + 470, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "removed in the second term. Such a design choice reflects the two types of positive instances for $z_{i}$ : (1) the instance in the same cluster and (2) the instance of the same class but in different clusters.", + "bbox": [ + 496, + 397, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to previous studies [27, 40, 61], the temperature $\\tau$ in contrastive loss is critical in controlling the local separation and global uniformity of the feature distribution. Specifically, for supervised contrastive learning, a low temperature makes relative high penalty on feature distribution, that actually encourages the features distribute more concentrically. As the temperature increases, the relative penalty tends to be more uniform which uniformizes the distribution of the features. Although the above objective explicitly considers the two types of label from different granularities, it still treats class and cluster label similarly. Intuitively, we expect instances of the same subclass to form a more concentrated cluster in feature space than those of the same class, since subclass naturally indicates finer-grained semantic coherence. To achieve this, we ensure the temperature $\\tau_{2} > \\tau_{1}$ and dynamically adjust $\\tau_{2}$ for each class according to its current level of concentration of the instances' feature. Following [40], for class $c$ we define $\\phi(c)$ as", + "bbox": [ + 496, + 460, + 893, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (c) = \\frac {\\sum_ {i = 1} ^ {n _ {c}} \\| z _ {i} - t _ {c} \\| _ {2}}{n _ {c} \\log \\left(n _ {c} + \\alpha\\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 748, + 890, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t_c$ is the centroid for the class $c$ , $\\alpha$ is a hyperparameter to ensure that $\\phi(c)$ is not overly-large, and $z_i$ corresponds to instances of class $c$ . From the formulation, we can see that if the current averaged distance to the class centroid is large or the class contains fewer data, thus the temperature will be set large to adopt the feature distribution of class $c$ during the training process. Then we define the temperature of class", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5398", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$c$ as", + "bbox": [ + 78, + 94, + 107, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {2} (c) = \\tau_ {1} \\cdot \\exp \\left(\\frac {\\phi (c)}{\\frac {1}{C} \\sum_ {i = 1} ^ {C} \\phi (i)}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 102, + 468, + 140 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "such that $\\tau_{2}(c)$ for class label is always larger than $\\tau_{1}$ for cluster label (since $\\phi(c) > 0$ ) and could reflect the current level of concentration of the instances in a class. In particular, the proposed $\\tau_{2}(c)$ encourages the features of instances in class $c$ to form a less tight cluster than that of a subclass (by $\\tau_{2}(c) > \\tau_{1}$ ) while adaptively adjust the temperature to prevent an overly-loose/dense cluster.", + "bbox": [ + 75, + 142, + 470, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training algorithm", + "text_level": 1, + "bbox": [ + 76, + 257, + 261, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we describe the overall training process of subclassbalancing contrastive learning and the algorithm can be found in Algorithm 2. First, the adaptive clustering (Section 3.1) could be noisy at the early stage of training [40,63]. Thus, we warm-up the feature extractor $f_{\\theta}(\\cdot)$ by a few epochs of training on ordinary SCL or KCL loss. In addition, our algorithm involves two adaptively-adjusting parts, namely, the cluster assignment and the temperature $\\tau_{2}(c)$ for each head class $c$ . Instead of updating these every epoch, we use a hyperparameter $K$ as the update interval, i.e., we update the cluster assignment and the temperature based on the current learned $f_{\\theta}(\\cdot)$ every $K$ epoch.", + "bbox": [ + 75, + 280, + 472, + 464 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 Training Algorithm" + ], + "code_body": "Require: Dataset $\\mathcal{D} = \\{x_i,y_i\\}_{i\\in [n]};$ ; The update interval of cluster assignment $K$ ; The number of warm-up epoch $T_0$ ; The total number of epoch $T$ ; The hyperparameters $\\beta$ and $\\delta$ \nEnsure: A trained feature extractor $f_{\\theta}(\\cdot)$ \n1: Initialize the model parameters $\\theta$ \n2: Train $f_{\\theta}(\\cdot)$ with SCL/KCL for $T_0$ epochs Warm-up stage \n3: for $t = T_0$ to $T$ do \n4: if $t\\% K = = 0$ then Update cluster and temperature \n5: Update the cluster assignment based on the current feature extractor $f_{\\theta}(x)$ \n6: Update the temperture $\\tau_{2}$ for each head class using Eq. 5 and Eq. 6 \n7: end if \n8: Train $f_{\\theta}(\\cdot)$ using Eq. 4 Subclass-balancing contrastive learning \n9: end for", + "bbox": [ + 76, + 494, + 470, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 746, + 200, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 771, + 264, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We consider three commonly used long-tailed recognition benchmark datasets: CIFAR-100-LT [6], ImageNet-LT [46], and iNaturalist 2018 [60]. The CIFAR-100-LT and ImageNet-LT datasets are artificially generated long-tailed datasets from the class-balanced datasets [38,52], and the iNaturalist 2018 dataset is a large-scale real-world dataset that exhibits long-tailed imbalance.", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We consider baseline methods of the following three categories: (1) class-balancing classifiers, including $\\tau$ -norm, LWS and cRT [34], which fixes the representation which trained by cross-entropy loss and trains the classifier with class-balanced sampling; (2) one-stage balancing loss, including CB loss [15], Focal loss [44], and LDAM loss [6]. These supervised distribution-aware loss makes the model to pay more attention on the minority class during training. (3) contrastive learning methods, including SCL [36], KCL [33], SwAV [8], PCL [40] and TSC [41] which train a feature extractor with the contrastive loss and then learn a classifier given the trained feature extractor.", + "bbox": [ + 496, + 90, + 893, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation protocol. Following [33, 34, 41], we implement SBCL, as well as other contrastive learning methods, in a two-stage framework. In the first stage, we train the feature extractor with a contrastive learning method, while in the second stage, we train a linear classifier on top of the learned representation. Specifically, for CIFAR-100-LT dataset, the linear classifier is trained with LDAM loss and class re-weighting [6]. For ImageNet-LT and iNaturalist 2018 datasets, the linear classifier is trained with CE loss and class-balanced sampling [34]. All results are averaged over 5 trials with different random seeds. We mainly report the overall top-1 accuracy. For the two large datasets, ImageNet-LT and iNaturalist 2018 datasets, following the previous work [46], we also report the accuracy of three disjoint subsets: Many-shot classes (classes with more than 100 samples), Medium-shot classes (classes with 20 to 100 samples), and Few-shot classes (classes under 20 samples). We show the implementation details and important hyperparameters in the Appendix A.2.", + "bbox": [ + 496, + 286, + 893, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main results", + "text_level": 1, + "bbox": [ + 500, + 589, + 633, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results on ImageNet-LT and iNaturalist 2018 are in Table 1. We can see that SBCL outperforms the baselines with a large margin over the two datasets. In addition, on iNaturalist 2018 dataset, SBCL outperforms the previous SOTA method by $0.7\\%$ on Many, $1.3\\%$ on Medium, $0.8\\%$ on Few and $1.1\\%$ on All, which shows the effectiveness of the proposed method in solving real-world long-tailed recognition problems such as natural species classification. Besides, SBCL is also better than existing contrastive learning method like KCL and TSC for all class splits, which demonstrates the effectiveness of the design principal of pursuing both instances- and subclass-balance in contrastive learning. Table 2 summarizes the results on CIFAR-100-LT dataset. For CIFAR-100-LT dataset, SBCL outperforms previous SOTA methods except for imbalance ratio 10. We hypothesize that it is because the tail class of CIFAR-100-LT with imbalance ratio 10 has multiple samples, which makes it hard to distinguish the performance of methods on the long-tailed recognition.", + "bbox": [ + 496, + 613, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5399", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c3cc4aabe92649c66e05e57e92305790920a471e1970a61156cda71ccc903b6d.jpg", + "table_caption": [ + "Table 1: Performance comparison on ImageNet-LT and iNaturalist 2018 datasets. Top-1 accuracy of ResNet-50 [26] is reported. The \"Many\", \"Medium\", \"Few\" and \"All\" denotes different groups. $\\dagger$ denotes our reproduced results of PCL, SwAV, and BYOL based on their official code. Other baselines' results on ImageNet-LT and iNaturalist 2018 are copied from Li et al. [41]." + ], + "table_footnote": [], + "table_body": "
BackboneImageNet-LTiNaturalist 2018
MethodsManyMediumFewAllManyMediumFewAll
CE64.033.85.841.672.263.057.261.7
Focal loss [44]51.040.820.843.7---61.3
CB-Focal [15]-------61.1
LDAM-DRW [6]60.446.930.749.8---64.6
OLTR [46]35.832.321.532.259.064.164.963.9
τ-norm [34]56.644.227.446.771.168.969.369.3
cRT [34]58.844.026.147.373.268.866.168.2
LWS [34]57.145.229.347.771.069.868.869.5
PCL† [40]34.726.112.327.548.545.941.744.5
SwAV† [8]37.528.315.630.151.948.443.747.0
BYOL† [21]37.728.916.330.652.348.644.147.2
SCL [36]61.447.028.249.8---66.4
KCL [33]62.449.029.551.5---68.6
TSC [41]63.549.730.452.472.670.667.869.7
SBCL63.851.331.253.473.371.968.670.8
", + "bbox": [ + 200, + 161, + 769, + 414 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ec996901eb0c11ab9b49fc5689dfb4d170692086748251ac9eae684ae1b1f68a.jpg", + "table_caption": [ + "Table 2: Performance comparison on CIFAR-100-LT. Top-1 accuracy of the ResNet-32 [26] under different imbalance ratios is reported. We also report the accuracy of our re-implemented important baselines $(\\dagger)$ in same setting on CIFAR-100-LT. The columns of \"Statistic (IR 100)\" are the results of different disjoint subsets on CIFAR-100-LT with imbalance ratio being 100." + ], + "table_footnote": [], + "table_body": "
MethodImbalance RatioStatistic (IR 100)
1005010ManyMediumFew
CE38.343.955.765.237.19.1
CB-CE [15]38.644.657.1---
Focal Loss [44]38.444.355.865.338.48.1
CB-Focal [15]38.745.258.065.037.610.3
CE-DRW [6]41.445.358.1---
CE-DRS [6]41.645.558.1---
LDAM [6]39.645.056.9---
LDAM-DRW [6]42.046.658.761.541.720.2
M2m-ERM [37]42.9-58.2---
M2m-LDAM [37]43.5-57.6---
cRT [34]43.346.858.164.044.818.1
LWS [34]43.146.458.1---
SCL [36] †42.145.254.862.842.018.4
KCL [33]42.846.357.6---
KCL†42.846.457.563.442.519.2
TSC [41]43.847.459.0---
TSC†43.547.658.763.743.220.4
SBCL44.948.757.964.445.322.2
", + "bbox": [ + 81, + 556, + 468, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Performance on Other Visual Tasks", + "text_level": 1, + "bbox": [ + 76, + 829, + 387, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There is a recent trend of using the contrastive learning to pretrain a feature extractor for downstream visual tasks other than image classification [24]. We are curious", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1f69d80d941548bec3e10ab4f538d0ae0ba133e24e3e455b39a286f98d9df5a5.jpg", + "table_caption": [ + "Table 3: Object detection results on PASCAL VOC." + ], + "table_footnote": [], + "table_body": "
MethodImageNetImageNet-LT
AP50APAP75AP50APAP75
random init.60.233.833.160.233.833.1
CE81.353.759.276.548.551.0
CL [24]81.356.162.778.251.556.5
KCL [33]82.355.562.179.752.657.9
SBCL81.956.262.880.653.458.8
", + "bbox": [ + 501, + 465, + 890, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "about two questions: (1) when the pretraining dataset is class-imbalanced, how the downstream performance is affected? (2) In such case, can our SBCL improve the learned feature extractor over existing contrastive learning baselines? To answer these questions, we use the object detection task of PASCAL VOC dataset as the evaluation suite and use ImageNet/ImageNet-LT datasets as class-balanced/imbalanced pretraining datasets. Following [24, 33], we first pretrain a feature extractor on ImageNet/ImageNet-LT then further finetune it for the downstream object detection tasks using Faster R-CNN [51] with R50-C4 backbone.", + "bbox": [ + 496, + 608, + 893, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experiment results are shown in Tables 3. From the results, we can see that pretraining on class-balanced data (ImageNet) leads to consistently better results than that on class-imbalanced dataset (ImageNet-LT) pretraining the model on the ImageNet and ImageNet-LT datasets by the SBCL can perform slightly better than other baselines. In addition, the proposed SBCL significantly outperforms baselines on class-imbalanced pretraining dataset, while achieve", + "bbox": [ + 496, + 779, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5400", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "comparable performance on class-balanced ones. For the representation which trained on the full ImageNet dataset, the performance advantage is not obvious. In Appendix A.1, we show additional experimental results of object detection and instance segmentation on COCO [45] dataset and SBCL also outperforms other baseline methods. Thus, we conclude that the proposed SBCL is not only helpful for image classification, but also other visual tasks.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Combining SBCL with SOTA methods", + "text_level": 1, + "bbox": [ + 76, + 223, + 410, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another line of research to address the long-tailed problem is the ensemble-based methods, such as MisLAS [72], WD [1], RIDE [62], PaCo [14], and BCL [74]. Here we show that SBCL can also be leveraged to boost the performance of these methods.", + "bbox": [ + 75, + 248, + 470, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MisLAS use label-aware smoothing and shifted batch normalization to improve the second stage (the classifier learning) on the long-tailed recognition, while our method is to improve the first stage of representation learning. Thus, MisLAS could be combined with SBCL and TSC. In Table 4, we report the results of combing MisLAS with both SBCL and TSC. The results show that both SBCL and TSC improve the performance of MisLAS and SBCL renders more performance boost than TSC.", + "bbox": [ + 75, + 324, + 468, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Weight decay (WD) [1] applies a penalty on the magnitudes of the weights in a model, which is particularly prominent for larger weights, resulting in the acquisition of more balanced parameters. WD and MaxNorm significantly enhances classification accuracy when applied to a classifier [1]. Table 4 presents the accuracy of the classifier trained by WD and MaxNorm, based on different representations learned by SBCL and TSC. The results demonstrate that SBCL improves the performance of the WD method [1] more effectively compared to TSC.", + "bbox": [ + 75, + 460, + 470, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To implement SBCL with RIDE which incorporates multiple models in a multi-expert framework, we follow [41] to simply replace the feature extractor on stage-1 training in RIDE with that trained with SBCL and keep the stage-2 routing training unchanged. As shown in Table 4, applying SBCL to RIDE improves its performance with a significant gap, outperforms the combination of TSC and RIDE on all different number of experts.", + "bbox": [ + 75, + 612, + 470, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PaCo and BCL propose new variants of supervised contrastive loss and jointly train both the proposed loss and classification loss to improve long-tail recognition, while we focus on the two stage pipeline, especially the first stage of representation learning. In this experiment, we show that using models pretrained with both TSC and SBCL as initialization could improve the performance of both PaCo and BCL. The results can be found in Table 4, and we can see that SBCL renders larger performance gain than TSC. The improvement over PaCo and BCL sheds lights on potential future work to evaluate the combination of multiple tech-", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "niques of long-tail recognition to achieve new SOTA results.", + "bbox": [ + 498, + 90, + 893, + 106 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4dc2de779ed4080e13cd29140b35b4b4f4f3dab1c6598e0f6d0803ef32efc108.jpg", + "table_caption": [ + "Table 4: Performance of the combination of SBCL and SOTA ensemble-based methods with ResNet-50 [26] on ImageNet-LT. $\\dagger$ denotes the results of our re-implemented TSC." + ], + "table_footnote": [], + "table_body": "
MethodManyMediumFewAll
MiSLAS [72]---52.7
WD & Max [1]62.550.441.553.9
RIDE [62] (3 experts)66.251.734.954.9
PaCo [14]64.455.733.756.0
BCL [74]67.954.236.657.1
TSC+ MiSLAS63.750.536.053.6
TSC+WD & Max63.851.941.655.1
TSC+ RIDE (3 experts)69.151.736.756.3
TSC+ RIDE (3 experts)†68.651.436.055.9
TSC+ PaCo66.455.835.757.1
TSC+ BCL69.056.337.858.7
SBCL + MiSLAS64.152.036.454.5
SBCL +WD & Max65.352.642.756.1
SBCL + RIDE (3 experts)69.252.436.956.8
SBCL + PaCo66.956.138.457.9
SBCL + BCL70.056.938.559.5
", + "bbox": [ + 501, + 179, + 893, + 426 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d087ae8fff96cf3d6f3f1d9ca4aadd937266713e6e398067dee12ac32769e651.jpg", + "table_caption": [ + "Table 5: Ablation study on different components of SBCL." + ], + "table_footnote": [], + "table_body": "
Warm-upAdaptive clusterDynamic temperatureCIFAR-100-LT
Imbalance Ratio1005010
✓✓44.047.957.2
✓✓43.847.256.5
✓✓43.847.857.0
✓✓✓44.948.757.9
", + "bbox": [ + 501, + 484, + 890, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation studies", + "text_level": 1, + "bbox": [ + 500, + 580, + 660, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Warm-up. As mentioned in Section 3.3, we train the feature extractor for several epochs using ordinary SCL or KCL as warm-up stage. As shown in Table 5, such a warm-up stage is beneficial since the performance drops when we remove the warm-up stage. This is likely because at the early stage of training, the extracted feature is not well-trained and the cluster assignment could be noisy and ineffective, hindering the efficacy of SBCL.", + "bbox": [ + 498, + 604, + 893, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive clustering. We are also curious about the efficacy of adaptive clustering and thus present the performance of SBCL with clustering only once and fixing cluster assignments during training. As shown in Table 5, without adaptive clustering, the performance decreases in all cases. The reason could be fixed cluster assignment is prone to noise when the model is not well-trained, while adaptive clustering would dynamically adjust the cluster assignments based on the current learned model, which is supposed to become better as the training proceeds.", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5401", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/76be709dc3c4b05881659dbcddda0a77b8fba9d70f2110aaa3dfcf52887147c3.jpg", + "image_caption": [ + "(a) Feature distance in subclasses." + ], + "image_footnote": [], + "bbox": [ + 158, + 90, + 444, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7a95efb41c4c36788cd109771490c8fdef84a519d462e13d307e2233431104ab.jpg", + "image_caption": [ + "(b) Feature distance in classes.", + "Figure 2: Feature distance of subclasses and classes on CIFAR-100-LT with imbalance ratio 100. We randomly sample instances from many-shot and medium-shot classes so that the size of each equals to that of few-shot classes." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 810, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic temperature. In Table 5, we also study the effectiveness of dynamic temperature (Section 3.2). We remove the dynamic temperature and simply set $\\tau_{2} > \\tau_{1}$ following [27]. With the fixed temperature $\\tau_{2}$ , the performance of SBCL is significantly worse than that with dynamic temperature. We speculate that this is because dynamic temperature could help prevent the instances of a class to form overly large or small cluster in the feature space and therefore lead to better learned representations. Additionally, to evaluate the impact of dynamic temperature on other baselines, we apply the dynamic temperature on TSC, as reported in Appendix A.1. In Appendix A.1, we also provide an additional analysis of other hyperparameters on SBCL.", + "bbox": [ + 75, + 330, + 472, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Illustrate the superiority of SBCL", + "text_level": 1, + "bbox": [ + 76, + 537, + 372, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we present an example that explicitly demonstrates the superiority of SBCL compared to other baselines. We fix the features learned from previous approaches and SBCL on the CIFAR-100-LT with imbalance ratio 100, and use these features for classification task. Table 6 shows that SBCL achieves the best performance.", + "bbox": [ + 75, + 561, + 470, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: Accuracy(%) of SBCL and other baselines on CIFAR-100-LT with imbalanced ratio 100.", + "bbox": [ + 75, + 654, + 468, + 683 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/088f1638f6e5761cf54e2085e1e2b42bfbb4892805335d141be3869db5931797.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCESCLKCLTSCSBCL
ACC(%)38.336.637.838.739.9
", + "bbox": [ + 109, + 695, + 431, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Feature distribution of SBCL", + "text_level": 1, + "bbox": [ + 76, + 753, + 336, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To analyze the representation learned by SBCL, we firstly define the euclidean distance between a given sample and other samples from the same/different classes as intra/inter-class distance. Concretely, the euclidean distance between a sample $z_{i}$ and a set $S$ is defined as $\\mathbf{D}(z_i,S) = \\frac{1}{|S|}\\sum_{z_j\\in S}\\| z_i - z_j\\| _2$ . Then, the intra- and interclass distance of sample $z_{i}$ can be defined as $\\mathbf{D}(z_i,P_i)$ and $\\mathbf{D}(z_i,\\mathcal{D} / P_i)$ separately; and the intra- and inter-subclass", + "bbox": [ + 75, + 777, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "distance of sample $z_{i}$ can be defined as $\\mathbf{D}(z_i,M_i)$ and $\\mathbf{D}(z_i,P_i / M_i)$ separately.", + "bbox": [ + 498, + 330, + 890, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 2 shows the average distance between features learned by SBCL in subclasses and classes on the CIFAR-100-LT. As shown in Figure 2a, the distance between samples from the same subclass is less than those from the same class but different subclasses. Meanwhile, in Figure 2b, the inter-class distance consistently exceeds the intra-class distance, indicating clear separation between features from different classes. Additionally, the inter-class distance remains stable, which suggests that each class is uniformly distributed on a hypersphere. The results in all indicate that the two-layer class hierarchy is successfully captured and feature distribution achieves the core idea of SBCL. The specific values of feature distance are displayed in Appendix A.1.", + "bbox": [ + 496, + 363, + 893, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 3 visualizes the feature distribution of subclasses in a head class and head classes in CIFAR-100-LT with imbalance ratio 100 trained by SBCL using t-SNE [59]. The representations learned by SBCL form some separated clusters, which be aligned with subclass labels and class labels. This suggests that SBCL can learn an embedding space where features from same subclasses/classes are pulled closely and features from different subclasses/class are separated.", + "bbox": [ + 496, + 560, + 893, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.8. Variation of per-class weight during training", + "text_level": 1, + "bbox": [ + 500, + 694, + 877, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Weight decay [1] achieves superior performance by effectively tuning weight decay during training. This study highlights the significance of regularizing network weights to become balanced for long-tailed recognition. Therefore, it is crucial to investigate whether SBCL can help the linear classifier to learn balanced weights through the training process.", + "bbox": [ + 496, + 717, + 893, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 shows the per-class weight norm of a linear classifier trained on top of features learned by SBCL in different training stages on CIFAR-100-LT. From the figure, we can see that as the training proceeds, the per-class weight norm becomes balanced even when training the linear classifier,", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5402", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c339c5d2ff0e13c098d1be84be93bb18c7192efa2eb41f976b78ebcf18a2f5c.jpg", + "image_caption": [ + "(a) Distribution of subclasses on a head class." + ], + "image_footnote": [], + "bbox": [ + 166, + 89, + 436, + 258 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fa7efb1b566fb25a5e16be51e8f0f10df6668a017219eef2636f4bec1f0a1ec2.jpg", + "image_caption": [ + "(b) Distribution of head classes on the dataset.", + "Figure 3: Feature distribution of subclasses and classes on CIFAR-100-LT with imbalance ratio 100. Color represents subclasses/classes." + ], + "image_footnote": [], + "bbox": [ + 534, + 90, + 803, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the original cross-entropy loss is used.", + "bbox": [ + 76, + 349, + 331, + 364 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3735514970833759f85269c51acb17992d9c8cf8be4b8d0fe523e44bc58578ff.jpg", + "image_caption": [ + "Figure 4: Change in weight norm of the linear classifier based on the representations trained by SBCL on CIFAR-100-LT with imbalance ratio 100 during training." + ], + "image_footnote": [], + "bbox": [ + 114, + 382, + 431, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Related work", + "text_level": 1, + "bbox": [ + 76, + 617, + 215, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Traditional methods for handling long-tailed recognition problem includes re-sampling and re-weighting. There are roughly two types of re-sampling techniques: over-sampling the minority classes [4, 5, 54, 70] and under-sampling the frequent classes [4, 23, 32]. The re-weighting techniques assign adaptive weights for different classes or even different samples. The vanilla scheme re-weights classes proportionally to the inverse of their frequency [29, 30, 64]. For class-level re-weighting methods, many loss functions including CB loss [15], LDAM loss [6] and Balanced softmax loss [50] were recently proposed, while instance-level re-weighting methods include Focal loss [44] and Influence-balanced loss [49]. Recently, two-stage algorithms have achieved remarkable performance for long-tailed recognition, such as classifier re-training (cRT) [34], learnable weight scaling (LWS) [34], and Mixup Shifted Label-Aware Smoothing model (MiSLAS) [72]. Meanwhile, bilateral branch net", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "work (BBN) [73] uses an additional network branch for rebalancing. RIDE [62] use multiple branches named experts, each learning to specialize in the entire classes. LADE [28] assumes the prior of test class distributions is available and accordingly post-adjust model predictions. PaCo [14] applies parametric class-wise learnable centers to rebalance in contrastive learning. BCL [74] proposes a multi-branch framework to achieve class-averaging and class-complement in the training process.", + "bbox": [ + 496, + 349, + 893, + 486 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To boost the performance of the two-stage algorithms, researchers have introduced supervised contrastive learning [36] to the first feature-learning stage and proposed $k$ -positive contrastive loss (KCL) [33] and targeted supervised contrastive learning (TSC) [41]. While achieving the state-of-the-art performance, these methods inject class-balance in the contrastive learning objective, inevitably leading to instance-imbalance during training. In this work, we instead propose to achieve both subclass- and instance-balance in the contrastive learning object. Our method is also related to recent studies of clustering-based deep unsupervised learning [7,8,18,43,65,66], especially those that leverage contrastive learning [22,40,42,63]. However, they target at general unsupervised representation learning scenario, while our method is tailored for long-tailed recognition where the training data is immensely class-imbalanced.", + "bbox": [ + 496, + 487, + 895, + 727 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 739, + 617, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we introduced Subclass-balancing Contrastive Learning (SBCL) for long-tailed recognition. It breaks down the head classes into multiple semantically-coherent subclasses via subclass-balancing adaptive clustering and incorporates a bi-granularity contrastive loss that encourages both subclass- and instance-balance. Extensive experiments on multiple datasets demonstrate that SBCL achieves state-of-the-art single-model performance on benchmark datasets for long-tailed recognition.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5403", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6897-6907, 2022. 7, 8", + "[2] Shin Ando and Chun Yuan Huang. Deep over-sampling framework for classifying imbalanced data. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 770-785. Springer, 2017. 1", + "[3] David Arthur and Sergei Vassilvitskii. K-means++ the advantages of careful seeding. In Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms, pages 1027-1035, 2007. 3, 14", + "[4] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106:249-259, 2018. 1, 9", + "[5] Jonathon Byrd and Zachary Lipton. What is the effect of importance weighting in deep learning? In International Conference on Machine Learning, pages 872-881. PMLR, 2019. 1, 9", + "[6] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in Neural Information Processing Systems, 32, 2019. 1, 5, 6, 9", + "[7] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In European Conference on Computer Vision, 2018. 9", + "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 2020. 5, 6, 9, 18", + "[9] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. Smote: synthetic minority over-sampling technique. Journal of artificial intelligence research, 16:321-357, 2002. 1", + "[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5008-5017, 2021.", + "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607. PMLR, 2020." + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020.", + "[13] Collobert, Ronan, Weston, and Jason. A unified architecture for natural language processing: deep neural networks with multitask learning. In International Conference on Machine Learning, 2008. 1", + "[14] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 7, 9", + "[15] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9268-9277, 2019. 1, 5, 6, 9", + "[16] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015.", + "[17] Qi Dong, Shaogang Gong, and Xiatian Zhu. Imbalanced deep learning by minority class incremental rectification. IEEE transactions on pattern analysis and machine intelligence, 41(6):1367-1381, 2018. 1", + "[18] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. Advances in Neural Information Processing Systems, 27, 2014. 9", + "[19] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018.", + "[20] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In Proceedings of the IEEE/cvf International Conference on computer vision, pages 6391-6400, 2019.", + "[21] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6, 18", + "[22] Yuanfan Guo, Minghao Xu, Jiawen Li, Bingbing Ni, Xuanyu Zhu, Zhenbang Sun, and Yi Xu. Hcsc: Hierarchical contrastive selective coding. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2022. 9" + ], + "bbox": [ + 501, + 90, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5404", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9):1263-1284, 2009. 9", + "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 6, 16, 17, 18", + "[25] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 16, 17", + "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6, 7", + "[27] David T Hoffmann, Nadine Behrmann, Juergen Gall, Thomas Brox, and Mehdi Noroozi. Ranking info noise contrastive estimation: Boosting contrastive learning via ranked positives. the Association for the Advance of Artificial Intelligence, 2022. 4, 8", + "[28] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, 2021. 2, 9", + "[29] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5375-5384, 2016. 9", + "[30] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Deep imbalanced learning for face recognition and attribute prediction. IEEE transactions on pattern analysis and machine intelligence, 42(11):2781-2794, 2019. 9", + "[31] Ahmet Iscen, André Araujo, Boqing Gong, and Cordelia Schmid. Class-balanced distillation for long-tailed visual recognition. Proceedings of the IEEE international conference on computer vision, 2021.", + "[32] Nathalie Japkowicz and Shaju Stephen. The class imbalance problem: A systematic study. Intelligent data analysis, 6(5):429-449, 2002. 9", + "[33] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 1, 2, 5, 6, 9, 16, 18", + "[34] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kalantidis. Decoupling representation and classifier for long-tailed recognition. International Conference on Learning Representations, 2019. 2, 5, 6, 9, 16", + "[35] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7482-7491, 2018.", + "[36] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 1, 2, 5, 6, 9", + "[37] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13896-13905, 2020. 6", + "[38] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[39] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955.", + "[40] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 4, 5, 6, 9, 18", + "[41] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6, 7, 9, 16, 17", + "[42] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. AAAI Conference on Artificial Intelligence, 2021. 9", + "[43] Renjie Liao, Alex Schwing, Richard Zemel, and Raquel Urtasun. Learning deep parsimonious representations. Advances in Neural Information Processing Systems, 29, 2016. 9", + "[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2980-2988, 2017. 5, 6, 9", + "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7" + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5405", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 5, 6", + "[47] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In European conference on computer vision, pages 181-196, 2018. 1", + "[48] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016.", + "[49] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 735-744, 2021. 9", + "[50] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. Advances in Neural Information Processing Systems, 33:4175–4186, 2020. 9", + "[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6, 17", + "[52] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 5", + "[53] Xie Saining, Girshick Ross, Dollar Piotr, Tu Zhuowen, and He Kaiming. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page 1492-1500, 2017.", + "[54] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In European conference on computer vision, pages 467-482. Springer, 2016. 1, 9", + "[55] Tony Shen, Ariel Lee, Carol Shen, and C Jimmy Lin. The long tail and rare disease research: the impact of next-generation sequencing for rare mendelian disorders. Genetics research, 97, 2015. 1", + "[56] Aditya Singh and Alessandro Bay. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in Neural Information Processing Systems, 2019." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Merrielle Spain and Pietro Perona. Measuring and predicting importance of objects in our visual world. 2007. 1", + "[58] AFM Uddin, Mst Monira, Wheemyung Shin, Tae-Choong Chung, Sung-Ho Bae, et al. Saliency: A saliency guided data augmentation strategy for better regularization. International Conference on Learning Representations, 2021.", + "[59] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8", + "[60] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 5", + "[61] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2021. 4", + "[62] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. International Conference on Learning Representations, 2021. 7, 9", + "[63] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021. 5, 9", + "[64] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. Advances in Neural Information Processing Systems, 30, 2017. 1, 9", + "[65] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016. 9", + "[66] Jianwei Yang, Devi Parikh, and Dhruv Batra. Joint unsupervised learning of deep representations and image clusters. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5147-5156, 2016. 9", + "[67] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. Advances in Neural Information Processing Systems, 33:19290-19301, 2020.", + "[68] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. International Conference on Machine Learning, 2019. 1" + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "5406", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Han-Jia Ye, Hong-You Chen, De-Chuan Zhan, and Wei-Lun Chao. Identifying and compensating for feature deviation in imbalanced deep learning. arXiv preprint arXiv:2001.01385, 2020.", + "[70] Qiaoyong Zhong, Chao Li, Yingying Zhang, H Sun, S Yang, D Xie, and S Pu. Towards good practices for recognition & detection. In CVPR workshops, volume 1, page 2, 2016. 9", + "[71] Yaoyao Zhong, Weihong Deng, Mei Wang, Jiani Hu, Jianteng Peng, Xunqiang Tao, and Yaohai Huang. Unequal-training for deep face recognition with long-tailed noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7812-7821, 2019. 1", + "[72] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16489-16498, 2021. 7, 9", + "[73] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, 2020. 9", + "[74] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 7, 9", + "[75] George Kingsley Zipf. The psycho-biology of language: An introduction to dynamic philology. Routledge, 2013. 1" + ], + "bbox": [ + 78, + 90, + 470, + 609 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "5407", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/bc05a493-9765-4ef8-8b9c-fc9c90911ac5_model.json b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/bc05a493-9765-4ef8-8b9c-fc9c90911ac5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f1db92bb0fc1576e027c2bfe37a2302b00788f64 --- /dev/null +++ b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/bc05a493-9765-4ef8-8b9c-fc9c90911ac5_model.json @@ -0,0 +1,2481 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.131, + 0.836, + 0.154 + ], + "angle": 0, + "content": "Subclass-balancing Contrastive Learning for Long-tailed Recognition" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.18, + 0.796, + 0.234 + ], + "angle": 0, + "content": "Chengkai Hou1, Jieyu Zhang2, Haonan Wang3, Tianyi Zhou4* \n1Jilin university, 2University of Washington, 3National University of Singapore \n4University of Maryland" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.237, + 0.821, + 0.269 + ], + "angle": 0, + "content": "houck20@mails.jlu.edu.cn, jieyuz2@cs.washington.edu, haonan.wang@u.nus.edu.tianyi@umd.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.336, + 0.474, + 0.761 + ], + "angle": 0, + "content": "Long-tailed recognition with imbalanced class distribution naturally emerges in practical machine learning applications. Existing methods such as data reweighing, resampling, and supervised contrastive learning enforce the class balance with a price of introducing imbalance between instances of head class and tail class, which may ignore the underlying rich semantic substructures of the former and exaggerate the biases in the latter. We overcome these drawbacks by a novel \"subclass-balancing contrastive learning (SBCL)\" approach that clusters each head class into multiple subclasses of similar sizes as the tail classes and enforce representations to capture the two-layer class hierarchy between the original classes and their subclasses. Since the clustering is conducted in the representation space and updated during the course of training, the subclass labels preserve the semantic substructures of head classes. Meanwhile, it does not overemphasize tail class samples, so each individual instance contribute to the representation learning equally. Hence, our method achieves both the instance- and subclass-balance, while the original class labels are also learned through contrastive learning among subclasses from different classes. We evaluate SBCL over a list of long-tailed benchmark datasets and it achieves the state-of-the-art performance. In addition, we present extensive analyses and ablation studies of SBCL to verify its advantages. Our code is available at https://github.com/JackHck/subclass-balancing-contrastive-learning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.778, + 0.21, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.471, + 0.88 + ], + "angle": 0, + "content": "In reality, the datasets often follow the Zipfian distribution over classes with a long tail [57, 75], i.e., a few classes (head classes) containing significantly more instances than the remaining tail classes. Such tail classes could be of great importance for high-stake applications, e.g., patient" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.306, + 0.895, + 0.381 + ], + "angle": 0, + "content": "class in medical diagnosis or accident class in autonomous driving [6, 55]. However, training on such class-imbalanced datasets can result in a severely biased model with noticeable performance drop in classification tasks [2, 4, 13, 47, 64, 68, 71]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.896, + 0.61 + ], + "angle": 0, + "content": "To overcome the challenges posed by long-tailed data, data resampling [2, 4, 9, 54] and loss reweighing [5, 6, 15, 17] have been widely applied but they cannot fully leverage all the head-class samples. Very recent work discovered that supervised contrastive learning (SCL) [36] can achieve state-of-the-art (SOTA) performance on benchmark datasets of long-tailed recognition [33, 41]. Specifically, the \\(k\\)-positive contrastive learning (KCL) [33] and its subsequent work targeted supervised contrastive learning (TSC) [41] revamp SCL by encouraging the learned feature space to be class-balanced and uniformly distributed. However, those methods enforcing class-balance often come with a price of instance-imbalance, i.e., each individual instance of tail classes would have much greater impact on model training than that of head classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Such instance-imbalance can result in significant degradation of the performance on long-tailed recognition for several reasons. On the one hand, the limited samples in each tail class might not be sufficiently representative of the whole class. So even a small bias of them can be enormously exaggerated by class-balancing methods and result in sub-optimal learning of classifiers or representations. On the other hand, head classes usually have more complicated semantic substructures, e.g., multiple high-density regions of the data distribution, so simply downweighing samples of head classes and treating them equally can easily lose critical structural information. For example, images of a head class \"cat\" might be highly diverse in breeds and colors, which need to be captured by different features but downweighing or subsampling them may easily lose such information, while a tail class \"platypus\" might only contain a few similar images that are unlikely to cover all the representative features. Therefore, it is non-trivial to enforce both class-balance and instance-balance simultaneously in the same method." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5395" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.408 + ], + "angle": 0, + "content": "Can we remove the negative impact of class-imbalance while still retain the advantages of instance-balance? In this paper, we achieve both through subclass-balancing contrastive learning (SBCL), a novel supervised contrastive learning defined on subclasses, which are the clusters within each head class, have comparable size as tail classes, and are adaptively updated during the training. Instead of sacrificing instance-balance for class-balance, our method achieves both instance- and subclass-balance by exploring the head-class structure in the learned representation space of the model-in-training. In particular, we propose a bi-granularity contrastive loss that enforces a sample (1) to be closer to samples from the same subclass than all the other samples; and (2) to be closer to samples from a different subclass but the same class than samples from any other subclasses. While the former learns representations with balanced and compact subclasses, the latter preserves the class structure on subclass level by encouraging the same class's subclasses to be closer to each other than to any different class's subclasses. Hence, it can learn an accurate classifier distinguishing original classes while enjoy both the instance- and subclass-balance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.409, + 0.472, + 0.47 + ], + "angle": 0, + "content": "In this paper, we apply SBCL for several visual recognition tasks to demonstrate SBCL superiority over other previous works (e.g., KCL [33], TSC [41]). To summarize, this paper makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.479, + 0.473, + 0.555 + ], + "angle": 0, + "content": "(a). We provide a new design principal of leveraging supervised contrastive learning for long-tailed recognition, i.e., aiming at achieving both instance- and subclass-balance instead of class-balance at the expense of instance-balance." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.565, + 0.472, + 0.642 + ], + "angle": 0, + "content": "(b). We propose a novel instantiation of the aforementioned design principal, subclass-balancing contrastive learning (SBCL), which consists of two major components, namely, subclass-balancing adaptive clustering and bi-granularity contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.65, + 0.473, + 0.742 + ], + "angle": 0, + "content": "(c). Empirically, we compare the SBCL against state-of-the-art methods on three visual tasks: image classification, object detection, and instance segmentation to demonstrate its effectiveness on handling class imbalance. We also conduct a series of experiments to analyze the efficacy of SBCL." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.479, + 0.473, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.325, + 0.772 + ], + "angle": 0, + "content": "2. Background and notations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Long-tailed recognition. Long-tailed recognition aims to learn a classifier from a training dataset with long-tailed class distribution, i.e., a few classes contain many data (head classes) while most classes contain only a few data (tail classes), where the major challenge is to require model recognizing all classes equally well. Let \\(\\mathcal{D} = \\{x_i, y_i\\}_{i \\in [n]}\\) be a long-tailed training dataset, where \\(x_i\\) denotes a sample and \\(y_i \\in [C]\\) denotes its label. Denote by \\(\\mathcal{D}_k \\subseteq \\mathcal{D}\\) the set of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.896, + 0.262 + ], + "angle": 0, + "content": "instances belonging to class \\( k \\) and \\( n_k = |\\mathcal{D}_k| \\) is the number of samples in class \\( k \\). The total number of training samples over \\( C \\) classes is \\( n = \\sum_{k=1}^{C} n_k \\). Without loss of generality, we follow prior work [28, 34] to assume that the classes are sorted by cardinality in decreasing order (i.e., if \\( i < j \\), then \\( n_i \\geq n_j \\)), and \\( n_1 \\gg n_C \\). In addition, we define the imbalance ratio as \\( \\max_{k \\in [C]} (n_k) / \\min_{k \\in [C]} (n_k) = n_1 / n_C \\). Finally, let \\( f_\\theta(\\cdot) \\) be a deep feature extractor, e.g., a neural network, parameterized by \\( \\theta \\) and \\( \\mathbf{w}_c \\) is the linear classifier of class \\( c \\), then the classifier we aim to learn is \\( h(x_i) = \\arg \\max_{c \\in [C]} \\mathbf{w}_c^\\top f_\\theta(x_i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.895, + 0.415 + ], + "angle": 0, + "content": "Supervised contrastive learning. Recent studies have shown that supervised contrastive learning (SCL) [36] provides a strong performance gain for long-tailed recognition and its variants have achieved state-of-the-art (SOTA) performance [33, 41]. Specifically, SCL learns the feature extractor \\( f_{\\theta}(\\cdot) \\) via maximizing the discriminativeness of positive instances, i.e., instances from the same class, and the learning objective for a single training data \\( (x_i, y_i) \\) in a batch \\( \\mathcal{B} = \\{(x_i, y_i)\\}_{i \\in [N]} \\), is" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.426, + 0.892, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S C L} = \\sum_ {i = 1} ^ {N} - \\frac {1}{| \\tilde {P} _ {i} |} \\sum_ {z _ {p} \\in \\tilde {P} _ {i}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau\\right)}{\\sum_ {z _ {a} \\in \\tilde {V} _ {i}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.676 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature hyperparameter, \\(z_{i} = f_{\\theta}(x_{i})\\) is the feature generated from \\(x_{i}\\), \\(V_{i} = \\{z_{i}\\}_{i\\in [N]}\\setminus \\{z_{i}\\}\\) is the current batch of features except for \\(z_{i}\\), \\(P_{i} = \\{z_{j}\\in V_{i}:y_{j} = y_{i}\\}\\) is a set of instances with the same label as \\(x_{i}\\). Finally, let \\(\\tilde{z}_i\\) be the feature of \\(\\tilde{x}_i\\), the augmented version of \\(x_{i}\\), and for any set \\(S_{i}\\) indexed by \\(i\\), we use \\(\\tilde{S}_i = S_i\\cup \\{\\tilde{z}_i\\}\\), e.g., \\(\\tilde{V}_i = V_i\\cup \\{\\tilde{z}_i\\}\\). However, for the long-tailed datasets, the feature spaces is dominated by head classes and thus have limited capability of semantic discrimination [33]. To address this, the \\(k\\)-positive contrastive learning (KCL) [33] attempts to balance the feature space by keeping the number of positive instances in \\(\\tilde{P}_i\\) equal for each class, leading to the following loss" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.685, + 0.892, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K C L} = \\sum_ {i = 1} ^ {N} - \\frac {1}{k + 1} \\sum_ {z _ {p} \\in \\tilde {P _ {i} ^ {k}}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau\\right)}{\\sum_ {z _ {a} \\in \\tilde {V _ {i}}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau\\right)} (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.895, + 0.803 + ], + "angle": 0, + "content": "where \\( P_{i}^{k} \\) is a subset of \\( P_{i} \\) with \\( k \\) randomly drawn instances. Finally, the learned feature extractor \\( f_{\\theta}(\\cdot) \\) is exploited in a sequel stage of training the classifier for long-tailed recognition [33, 41]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.636, + 0.833 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As mentioned above, KCL and its sequels [33,41] balance the learning objective of SCL by picking the same number of positive instances for each class, i.e., \\( |P_i^k| = k \\) in Eq. 2 no matter which class \\( x_i \\) belongs to, however, we argue that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5396" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.09, + 0.478, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.28, + 0.424, + 0.291 + ], + "angle": 0, + "content": "(a). Subclass-balancing adaptive clustering (Section 3.1)" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.091, + 0.871, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.28, + 0.799, + 0.291 + ], + "angle": 0, + "content": "(b). Bi-granularity contrastive loss (Section 3.2)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Figure 1: Illustration of subclass-balancing contrastive learning (SBCL). It initially divides the head classes into multiple subclasses of comparable size. Then, during training, SBCL builds each sample to be closer to samples from the same subclass than samples from different subclasses but the same class, which are also made to be closer than samples from different classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.47, + 0.603 + ], + "angle": 0, + "content": "such a class-balancing approach would inevitably introduce instance-imbalance: the instances of tail classes have much more chances to be engaged in the training than that of head class. Specifically, assume each class has no less than \\( k \\) instances, then the probability of an instance of class \\( c \\) being selected as positive instance is \\( p(c) = \\frac{k}{n_c} \\); if the tail class \\( n_C \\) has only \\( k \\) instances and the imbalance ratio is \\( \\frac{n_1}{n_C} = 100 \\), then we have \\( p(1) = \\frac{k}{100k} = 0.01 \\) while \\( p(C) = 1 \\). We can see that when the instances of head class are selected once, that of tail class may already be trained 100 times. Thus, the training is immensely biased towards the few samples in each tail class. Besides, as tail classes only have very few instances that are not necessarily representative, the learned feature space might be unsatisfactory and sensitive to the training data of tail classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.471, + 0.831 + ], + "angle": 0, + "content": "Here, we provide a new prospective of handling class-imbalance issue by contrastive learning: instead of aiming at class-balance at the expense of instance-imbalance, we propose to achieve both instance- and subclass-balance. We argue that head classes typically contain more diverse instances and thus have richer semantics in the training dataset. Therefore, it might be wise to break down the head classes into multiple semantically coherent subclasses, each of which consists of similar number of instance as tail classes. Built on this spirit, we develop subclass-balancing contrastive learning (SBCL), a new contrastive learning framework for long-tailed recognition (visualized in Figure 1) that achieves both instance- and subclass-balance. We present a concrete example that effectively demonstrates the superiority of SBCL compared to previous methods in Section 4.6." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.846, + 0.411, + 0.863 + ], + "angle": 0, + "content": "3.1. Subclass-balancing adaptive clustering" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We break down the class into several \"subclasses\" to attack the imbalanced phenomenon. Particularly, given a class" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.895, + 0.525 + ], + "angle": 0, + "content": "\\(c\\) and the associated set of data \\(\\mathcal{D}_c\\), we employ a clustering algorithm of choice based on the features extracted by current feature extractor \\(f_{\\theta}(\\cdot)\\) to divide \\(\\mathcal{D}_c\\) into \\(m_c\\) subclasses/clusters. We use \\(\\Gamma_c(x_i)\\) to denote the cluster label of an instance \\(x_i\\) of class \\(c\\). To ensure that the number of samples for each subclass is roughly the same, we propose a new cluster algorithm to divide the unit-length feature vectors, i.e., the features output by \\(f_{\\theta}(\\cdot)\\) with additional unit-length normalization. The new proposed cluster algorithm is described in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.892, + 0.615 + ], + "angle": 0, + "content": "In Algorithm 1, during the initialization of cluster centers, our approach gradually selects cluster centers by prioritizing the samples that are maximally distant from other existing cluster centers [3]. In the process of assigning vectors to their centers, we set a threshold \\( M \\) as an upper limit of sample size in a cluster, which guarantees clusters of balanced sizes." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.616, + 0.731, + 0.63 + ], + "angle": 0, + "content": "Specifically, the threshold \\(M\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.638, + 0.891, + 0.653 + ], + "angle": 0, + "content": "\\[\nM = \\max \\left(n _ {C}, \\delta\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.869 + ], + "angle": 0, + "content": "where \\( n_C \\) is the size of tail class and the hyperparameter \\( \\delta \\) controls the lower bound of sample size in clusters to prevent overly-small cluster. Note that we only apply clustering algorithm to classes which contain multiple instances while the tail classes remain unchanged. As a consequence, the size of each resultant cluster is similar to that of tail class \\( n_C \\). In addition, instead of only clustering once at the beginning, we update the cluster assignment adaptively based on the current feature extractor \\( f_{\\theta}(\\cdot) \\) during training process and empirically show that such adaptive clustering outperforms only-once clustering in Section 4.5. We add an exhaustive description of Algorithm 1 and show the distribution of the sample size in clusters on the benchmark dataset in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.9 + ], + "angle": 0, + "content": "Then, by replacing the class labels of head classes used in SCL/KCL with the finer-grained cluster labels, we ensure the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "5397" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.092, + 0.432, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Subclass-balancing Adaptive Clustering" + }, + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.109, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Require: Sample set \\(S = \\{x_{i}\\}_{i = 1}^{n}\\) ; A threshold \\(M\\) ; The number of iterations \\(K\\) \nEnsure: Cluster assignments for samples in \\(s\\) \nfor \\(k = 0\\) to \\(K\\) do if \\(k = 0\\) then Choose the cluster centers \\(y_{j}\\) which are farther away from previously selected centers. else Update the cluster centers \\(y_{j} = \\frac{1}{n_{j}}\\sum_{i = 1}^{n_{j}}x_{i}.\\) \\(\\triangleright n_j\\) is the number of samples in a cluster end if Construct the cluster center set \\(\\mathcal{C} = \\{y_j\\}_{j = 1}^m\\) . \\(\\triangleright m\\) is the number of cluster centers while \\(S\\neq \\phi\\) do \\(\\triangleright\\) Assign samples to centers \\(y_{i}\\) Select the most similar pair \\((x_{i},y_{j}) = \\arg \\max \\mathrm{cosine - similarity}(x,y)\\) \\(x\\in S,y\\in C\\) Assign the sample \\(x_{i}\\) to the center \\(y_{j}\\) . Delete the assigned sample \\(x_{i}\\) from the sample set \\(S = S / \\{x_i\\}\\) if \\(n_j\\geq M\\) then Sample number in a cluster exceeds the threshold \\(M\\) Delete the cluster center \\(y_{j}\\) from the cluster center set \\(\\mathcal{C} = \\mathcal{C} / \\{y_j\\}\\) end if end while end for" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.471, + 0.503 + ], + "angle": 0, + "content": "instance-balanced, i.e., each instance has similar chance of being selected regardless of its class. By breaking down the head classes, which typically contain more diverse instances, into multiple semantically coherent subclasses, we achieve subclass-balanced (instead of class-balanced) while maintain the rich semantics rendered by head classes in training dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.345, + 0.53 + ], + "angle": 0, + "content": "3.2. Bi-granularity contrastive loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.69 + ], + "angle": 0, + "content": "We now have two types of label for instances in head classes from different granularities: the coarse-grained class label and the fine-grained cluster label. A direct consequence of replacing class label in SCL/KCL with cluster label is that we no longer distinguish instances from different head classes, and therefore the boundaries between classes might be blurry, leading to sub-optimal feature space. As a remedy, we combine the contrastive loss of both class label and cluster label into the following one and reuse the notations of Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.714, + 0.469, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {S B C L} = - \\sum_ {i = 1} ^ {N} \\left(\\frac {1}{| \\tilde {M} _ {i} |} \\sum_ {z _ {p} \\in \\tilde {M} _ {i}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau_ {1}\\right)}{\\sum z _ {a} \\in \\tilde {V} _ {i}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau_ {1}\\right) \\right. \\\\ + \\beta \\frac {1}{\\left| \\tilde {P} _ {i} \\right| - \\left| M _ {i} \\right|} \\sum_ {z _ {p} \\in \\tilde {P} _ {i} / M _ {i}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau_ {2}\\right)}{\\sum_ {z _ {a} \\in \\tilde {V} _ {i} / M _ {i}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau_ {2}\\right)}) \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.9 + ], + "angle": 0, + "content": "where \\( M_{i} = \\{z_{j}\\in P_{i}:\\Gamma_{y_{i}}(x_{i}) = \\Gamma_{y_{i}}(x_{j})\\} \\) is a set of instances with the same cluster label as \\( x_{i} \\). \\( \\beta \\) is a hyperparameter that balances these two loss terms. The first term corresponds to the SCL loss with cluster labels, while the second term leverages the class label but does not consider the instances of the same cluster, i.e., the instances in \\( M_{i} \\) are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.457 + ], + "angle": 0, + "content": "removed in the second term. Such a design choice reflects the two types of positive instances for \\( z_{i} \\): (1) the instance in the same cluster and (2) the instance of the same class but in different clusters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.895, + 0.734 + ], + "angle": 0, + "content": "According to previous studies [27, 40, 61], the temperature \\(\\tau\\) in contrastive loss is critical in controlling the local separation and global uniformity of the feature distribution. Specifically, for supervised contrastive learning, a low temperature makes relative high penalty on feature distribution, that actually encourages the features distribute more concentrically. As the temperature increases, the relative penalty tends to be more uniform which uniformizes the distribution of the features. Although the above objective explicitly considers the two types of label from different granularities, it still treats class and cluster label similarly. Intuitively, we expect instances of the same subclass to form a more concentrated cluster in feature space than those of the same class, since subclass naturally indicates finer-grained semantic coherence. To achieve this, we ensure the temperature \\(\\tau_{2} > \\tau_{1}\\) and dynamically adjust \\(\\tau_{2}\\) for each class according to its current level of concentration of the instances' feature. Following [40], for class \\(c\\) we define \\(\\phi(c)\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.749, + 0.892, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\phi (c) = \\frac {\\sum_ {i = 1} ^ {n _ {c}} \\| z _ {i} - t _ {c} \\| _ {2}}{n _ {c} \\log \\left(n _ {c} + \\alpha\\right)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( t_c \\) is the centroid for the class \\( c \\), \\( \\alpha \\) is a hyperparameter to ensure that \\( \\phi(c) \\) is not overly-large, and \\( z_i \\) corresponds to instances of class \\( c \\). From the formulation, we can see that if the current averaged distance to the class centroid is large or the class contains fewer data, thus the temperature will be set large to adopt the feature distribution of class \\( c \\) during the training process. Then we define the temperature of class" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5398" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.079, + 0.095, + 0.108, + 0.105 + ], + "angle": 0, + "content": "\\(c\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.103, + 0.47, + 0.141 + ], + "angle": 0, + "content": "\\[\n\\tau_ {2} (c) = \\tau_ {1} \\cdot \\exp \\left(\\frac {\\phi (c)}{\\frac {1}{C} \\sum_ {i = 1} ^ {C} \\phi (i)}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.143, + 0.471, + 0.25 + ], + "angle": 0, + "content": "such that \\(\\tau_{2}(c)\\) for class label is always larger than \\(\\tau_{1}\\) for cluster label (since \\(\\phi(c) > 0\\)) and could reflect the current level of concentration of the instances in a class. In particular, the proposed \\(\\tau_{2}(c)\\) encourages the features of instances in class \\(c\\) to form a less tight cluster than that of a subclass (by \\(\\tau_{2}(c) > \\tau_{1}\\)) while adaptively adjust the temperature to prevent an overly-loose/dense cluster." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.258, + 0.262, + 0.274 + ], + "angle": 0, + "content": "3.3. Training algorithm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.473, + 0.465 + ], + "angle": 0, + "content": "Here, we describe the overall training process of subclassbalancing contrastive learning and the algorithm can be found in Algorithm 2. First, the adaptive clustering (Section 3.1) could be noisy at the early stage of training [40,63]. Thus, we warm-up the feature extractor \\( f_{\\theta}(\\cdot) \\) by a few epochs of training on ordinary SCL or KCL loss. In addition, our algorithm involves two adaptively-adjusting parts, namely, the cluster assignment and the temperature \\( \\tau_{2}(c) \\) for each head class \\( c \\). Instead of updating these every epoch, we use a hyperparameter \\( K \\) as the update interval, i.e., we update the cluster assignment and the temperature based on the current learned \\( f_{\\theta}(\\cdot) \\) every \\( K \\) epoch." + }, + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.477, + 0.298, + 0.492 + ], + "angle": 0, + "content": "Algorithm 2 Training Algorithm" + }, + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.495, + 0.472, + 0.718 + ], + "angle": 0, + "content": "Require: Dataset \\(\\mathcal{D} = \\{x_i,y_i\\}_{i\\in [n]};\\) ; The update interval of cluster assignment \\(K\\) ; The number of warm-up epoch \\(T_0\\) ; The total number of epoch \\(T\\) ; The hyperparameters \\(\\beta\\) and \\(\\delta\\) \nEnsure: A trained feature extractor \\(f_{\\theta}(\\cdot)\\) \n1: Initialize the model parameters \\(\\theta\\) \n2: Train \\(f_{\\theta}(\\cdot)\\) with SCL/KCL for \\(T_0\\) epochs Warm-up stage \n3: for \\(t = T_0\\) to \\(T\\) do \n4: if \\(t\\% K = = 0\\) then Update cluster and temperature \n5: Update the cluster assignment based on the current feature extractor \\(f_{\\theta}(x)\\) \n6: Update the temperture \\(\\tau_{2}\\) for each head class using Eq. 5 and Eq. 6 \n7: end if \n8: Train \\(f_{\\theta}(\\cdot)\\) using Eq. 4 Subclass-balancing contrastive learning \n9: end for" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.747, + 0.202, + 0.764 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.265, + 0.789 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Datasets. We consider three commonly used long-tailed recognition benchmark datasets: CIFAR-100-LT [6], ImageNet-LT [46], and iNaturalist 2018 [60]. The CIFAR-100-LT and ImageNet-LT datasets are artificially generated long-tailed datasets from the class-balanced datasets [38,52], and the iNaturalist 2018 dataset is a large-scale real-world dataset that exhibits long-tailed imbalance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.273 + ], + "angle": 0, + "content": "Baselines. We consider baseline methods of the following three categories: (1) class-balancing classifiers, including \\(\\tau\\)-norm, LWS and cRT [34], which fixes the representation which trained by cross-entropy loss and trains the classifier with class-balanced sampling; (2) one-stage balancing loss, including CB loss [15], Focal loss [44], and LDAM loss [6]. These supervised distribution-aware loss makes the model to pay more attention on the minority class during training. (3) contrastive learning methods, including SCL [36], KCL [33], SwAV [8], PCL [40] and TSC [41] which train a feature extractor with the contrastive loss and then learn a classifier given the trained feature extractor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.895, + 0.576 + ], + "angle": 0, + "content": "Evaluation protocol. Following [33, 34, 41], we implement SBCL, as well as other contrastive learning methods, in a two-stage framework. In the first stage, we train the feature extractor with a contrastive learning method, while in the second stage, we train a linear classifier on top of the learned representation. Specifically, for CIFAR-100-LT dataset, the linear classifier is trained with LDAM loss and class re-weighting [6]. For ImageNet-LT and iNaturalist 2018 datasets, the linear classifier is trained with CE loss and class-balanced sampling [34]. All results are averaged over 5 trials with different random seeds. We mainly report the overall top-1 accuracy. For the two large datasets, ImageNet-LT and iNaturalist 2018 datasets, following the previous work [46], we also report the accuracy of three disjoint subsets: Many-shot classes (classes with more than 100 samples), Medium-shot classes (classes with 20 to 100 samples), and Few-shot classes (classes under 20 samples). We show the implementation details and important hyperparameters in the Appendix A.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.634, + 0.604 + ], + "angle": 0, + "content": "4.2. Main results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The results on ImageNet-LT and iNaturalist 2018 are in Table 1. We can see that SBCL outperforms the baselines with a large margin over the two datasets. In addition, on iNaturalist 2018 dataset, SBCL outperforms the previous SOTA method by \\(0.7\\%\\) on Many, \\(1.3\\%\\) on Medium, \\(0.8\\%\\) on Few and \\(1.1\\%\\) on All, which shows the effectiveness of the proposed method in solving real-world long-tailed recognition problems such as natural species classification. Besides, SBCL is also better than existing contrastive learning method like KCL and TSC for all class splits, which demonstrates the effectiveness of the design principal of pursuing both instances- and subclass-balance in contrastive learning. Table 2 summarizes the results on CIFAR-100-LT dataset. For CIFAR-100-LT dataset, SBCL outperforms previous SOTA methods except for imbalance ratio 10. We hypothesize that it is because the tail class of CIFAR-100-LT with imbalance ratio 10 has multiple samples, which makes it hard to distinguish the performance of methods on the long-tailed recognition." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5399" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.089, + 0.892, + 0.151 + ], + "angle": 0, + "content": "Table 1: Performance comparison on ImageNet-LT and iNaturalist 2018 datasets. Top-1 accuracy of ResNet-50 [26] is reported. The \"Many\", \"Medium\", \"Few\" and \"All\" denotes different groups. \\(\\dagger\\) denotes our reproduced results of PCL, SwAV, and BYOL based on their official code. Other baselines' results on ImageNet-LT and iNaturalist 2018 are copied from Li et al. [41]." + }, + { + "type": "table", + "bbox": [ + 0.202, + 0.162, + 0.77, + 0.415 + ], + "angle": 0, + "content": "
BackboneImageNet-LTiNaturalist 2018
MethodsManyMediumFewAllManyMediumFewAll
CE64.033.85.841.672.263.057.261.7
Focal loss [44]51.040.820.843.7---61.3
CB-Focal [15]-------61.1
LDAM-DRW [6]60.446.930.749.8---64.6
OLTR [46]35.832.321.532.259.064.164.963.9
τ-norm [34]56.644.227.446.771.168.969.369.3
cRT [34]58.844.026.147.373.268.866.168.2
LWS [34]57.145.229.347.771.069.868.869.5
PCL† [40]34.726.112.327.548.545.941.744.5
SwAV† [8]37.528.315.630.151.948.443.747.0
BYOL† [21]37.728.916.330.652.348.644.147.2
SCL [36]61.447.028.249.8---66.4
KCL [33]62.449.029.551.5---68.6
TSC [41]63.549.730.452.472.670.667.869.7
SBCL63.851.331.253.473.371.968.670.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.438, + 0.472, + 0.545 + ], + "angle": 0, + "content": "Table 2: Performance comparison on CIFAR-100-LT. Top-1 accuracy of the ResNet-32 [26] under different imbalance ratios is reported. We also report the accuracy of our re-implemented important baselines \\((\\dagger)\\) in same setting on CIFAR-100-LT. The columns of \"Statistic (IR 100)\" are the results of different disjoint subsets on CIFAR-100-LT with imbalance ratio being 100." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.557, + 0.469, + 0.811 + ], + "angle": 0, + "content": "
MethodImbalance RatioStatistic (IR 100)
1005010ManyMediumFew
CE38.343.955.765.237.19.1
CB-CE [15]38.644.657.1---
Focal Loss [44]38.444.355.865.338.48.1
CB-Focal [15]38.745.258.065.037.610.3
CE-DRW [6]41.445.358.1---
CE-DRS [6]41.645.558.1---
LDAM [6]39.645.056.9---
LDAM-DRW [6]42.046.658.761.541.720.2
M2m-ERM [37]42.9-58.2---
M2m-LDAM [37]43.5-57.6---
cRT [34]43.346.858.164.044.818.1
LWS [34]43.146.458.1---
SCL [36] †42.145.254.862.842.018.4
KCL [33]42.846.357.6---
KCL†42.846.457.563.442.519.2
TSC [41]43.847.459.0---
TSC†43.547.658.763.743.220.4
SBCL44.948.757.964.445.322.2
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.388, + 0.845 + ], + "angle": 0, + "content": "4.3. Performance on Other Visual Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "There is a recent trend of using the contrastive learning to pretrain a feature extractor for downstream visual tasks other than image classification [24]. We are curious" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.439, + 0.867, + 0.454 + ], + "angle": 0, + "content": "Table 3: Object detection results on PASCAL VOC." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.466, + 0.892, + 0.573 + ], + "angle": 0, + "content": "
MethodImageNetImageNet-LT
AP50APAP75AP50APAP75
random init.60.233.833.160.233.833.1
CE81.353.759.276.548.551.0
CL [24]81.356.162.778.251.556.5
KCL [33]82.355.562.179.752.657.9
SBCL81.956.262.880.653.458.8
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.895, + 0.776 + ], + "angle": 0, + "content": "about two questions: (1) when the pretraining dataset is class-imbalanced, how the downstream performance is affected? (2) In such case, can our SBCL improve the learned feature extractor over existing contrastive learning baselines? To answer these questions, we use the object detection task of PASCAL VOC dataset as the evaluation suite and use ImageNet/ImageNet-LT datasets as class-balanced/imbalanced pretraining datasets. Following [24, 33], we first pretrain a feature extractor on ImageNet/ImageNet-LT then further finetune it for the downstream object detection tasks using Faster R-CNN [51] with R50-C4 backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The experiment results are shown in Tables 3. From the results, we can see that pretraining on class-balanced data (ImageNet) leads to consistently better results than that on class-imbalanced dataset (ImageNet-LT) pretraining the model on the ImageNet and ImageNet-LT datasets by the SBCL can perform slightly better than other baselines. In addition, the proposed SBCL significantly outperforms baselines on class-imbalanced pretraining dataset, while achieve" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5400" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "comparable performance on class-balanced ones. For the representation which trained on the full ImageNet dataset, the performance advantage is not obvious. In Appendix A.1, we show additional experimental results of object detection and instance segmentation on COCO [45] dataset and SBCL also outperforms other baseline methods. Thus, we conclude that the proposed SBCL is not only helpful for image classification, but also other visual tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.224, + 0.411, + 0.24 + ], + "angle": 0, + "content": "4.4. Combining SBCL with SOTA methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.249, + 0.471, + 0.323 + ], + "angle": 0, + "content": "Another line of research to address the long-tailed problem is the ensemble-based methods, such as MisLAS [72], WD [1], RIDE [62], PaCo [14], and BCL [74]. Here we show that SBCL can also be leveraged to boost the performance of these methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.46 + ], + "angle": 0, + "content": "MisLAS use label-aware smoothing and shifted batch normalization to improve the second stage (the classifier learning) on the long-tailed recognition, while our method is to improve the first stage of representation learning. Thus, MisLAS could be combined with SBCL and TSC. In Table 4, we report the results of combing MisLAS with both SBCL and TSC. The results show that both SBCL and TSC improve the performance of MisLAS and SBCL renders more performance boost than TSC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.471, + 0.612 + ], + "angle": 0, + "content": "Weight decay (WD) [1] applies a penalty on the magnitudes of the weights in a model, which is particularly prominent for larger weights, resulting in the acquisition of more balanced parameters. WD and MaxNorm significantly enhances classification accuracy when applied to a classifier [1]. Table 4 presents the accuracy of the classifier trained by WD and MaxNorm, based on different representations learned by SBCL and TSC. The results demonstrate that SBCL improves the performance of the WD method [1] more effectively compared to TSC." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.471, + 0.734 + ], + "angle": 0, + "content": "To implement SBCL with RIDE which incorporates multiple models in a multi-expert framework, we follow [41] to simply replace the feature extractor on stage-1 training in RIDE with that trained with SBCL and keep the stage-2 routing training unchanged. As shown in Table 4, applying SBCL to RIDE improves its performance with a significant gap, outperforms the combination of TSC and RIDE on all different number of experts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "PaCo and BCL propose new variants of supervised contrastive loss and jointly train both the proposed loss and classification loss to improve long-tail recognition, while we focus on the two stage pipeline, especially the first stage of representation learning. In this experiment, we show that using models pretrained with both TSC and SBCL as initialization could improve the performance of both PaCo and BCL. The results can be found in Table 4, and we can see that SBCL renders larger performance gain than TSC. The improvement over PaCo and BCL sheds lights on potential future work to evaluate the combination of multiple tech-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.107 + ], + "angle": 0, + "content": "niques of long-tail recognition to achieve new SOTA results." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.122, + 0.894, + 0.168 + ], + "angle": 0, + "content": "Table 4: Performance of the combination of SBCL and SOTA ensemble-based methods with ResNet-50 [26] on ImageNet-LT. \\(\\dagger\\) denotes the results of our re-implemented TSC." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.18, + 0.895, + 0.428 + ], + "angle": 0, + "content": "
MethodManyMediumFewAll
MiSLAS [72]---52.7
WD & Max [1]62.550.441.553.9
RIDE [62] (3 experts)66.251.734.954.9
PaCo [14]64.455.733.756.0
BCL [74]67.954.236.657.1
TSC+ MiSLAS63.750.536.053.6
TSC+WD & Max63.851.941.655.1
TSC+ RIDE (3 experts)69.151.736.756.3
TSC+ RIDE (3 experts)†68.651.436.055.9
TSC+ PaCo66.455.835.757.1
TSC+ BCL69.056.337.858.7
SBCL + MiSLAS64.152.036.454.5
SBCL +WD & Max65.352.642.756.1
SBCL + RIDE (3 experts)69.252.436.956.8
SBCL + PaCo66.956.138.457.9
SBCL + BCL70.056.938.559.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.459, + 0.889, + 0.474 + ], + "angle": 0, + "content": "Table 5: Ablation study on different components of SBCL." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.486, + 0.892, + 0.564 + ], + "angle": 0, + "content": "
Warm-upAdaptive clusterDynamic temperatureCIFAR-100-LT
Imbalance Ratio1005010
✓✓44.047.957.2
✓✓43.847.256.5
✓✓43.847.857.0
✓✓✓44.948.757.9
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.661, + 0.595 + ], + "angle": 0, + "content": "4.5. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.605, + 0.894, + 0.727 + ], + "angle": 0, + "content": "Warm-up. As mentioned in Section 3.3, we train the feature extractor for several epochs using ordinary SCL or KCL as warm-up stage. As shown in Table 5, such a warm-up stage is beneficial since the performance drops when we remove the warm-up stage. This is likely because at the early stage of training, the extracted feature is not well-trained and the cluster assignment could be noisy and ineffective, hindering the efficacy of SBCL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Adaptive clustering. We are also curious about the efficacy of adaptive clustering and thus present the performance of SBCL with clustering only once and fixing cluster assignments during training. As shown in Table 5, without adaptive clustering, the performance decreases in all cases. The reason could be fixed cluster assignment is prone to noise when the model is not well-trained, while adaptive clustering would dynamically adjust the cluster assignments based on the current learned model, which is supposed to become better as the training proceeds." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5401" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.091, + 0.445, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.25, + 0.402, + 0.262 + ], + "angle": 0, + "content": "(a) Feature distance in subclasses." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.092, + 0.812, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.25, + 0.76, + 0.262 + ], + "angle": 0, + "content": "(b) Feature distance in classes." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.893, + 0.307 + ], + "angle": 0, + "content": "Figure 2: Feature distance of subclasses and classes on CIFAR-100-LT with imbalance ratio 100. We randomly sample instances from many-shot and medium-shot classes so that the size of each equals to that of few-shot classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.473, + 0.53 + ], + "angle": 0, + "content": "Dynamic temperature. In Table 5, we also study the effectiveness of dynamic temperature (Section 3.2). We remove the dynamic temperature and simply set \\(\\tau_{2} > \\tau_{1}\\) following [27]. With the fixed temperature \\(\\tau_{2}\\), the performance of SBCL is significantly worse than that with dynamic temperature. We speculate that this is because dynamic temperature could help prevent the instances of a class to form overly large or small cluster in the feature space and therefore lead to better learned representations. Additionally, to evaluate the impact of dynamic temperature on other baselines, we apply the dynamic temperature on TSC, as reported in Appendix A.1. In Appendix A.1, we also provide an additional analysis of other hyperparameters on SBCL." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.539, + 0.373, + 0.556 + ], + "angle": 0, + "content": "4.6. Illustrate the superiority of SBCL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.471, + 0.653 + ], + "angle": 0, + "content": "In this section, we present an example that explicitly demonstrates the superiority of SBCL compared to other baselines. We fix the features learned from previous approaches and SBCL on the CIFAR-100-LT with imbalance ratio 100, and use these features for classification task. Table 6 shows that SBCL achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.684 + ], + "angle": 0, + "content": "Table 6: Accuracy(%) of SBCL and other baselines on CIFAR-100-LT with imbalanced ratio 100." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.696, + 0.433, + 0.73 + ], + "angle": 0, + "content": "
MethodCESCLKCLTSCSBCL
ACC(%)38.336.637.838.739.9
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.754, + 0.338, + 0.769 + ], + "angle": 0, + "content": "4.7. Feature distribution of SBCL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To analyze the representation learned by SBCL, we firstly define the euclidean distance between a given sample and other samples from the same/different classes as intra/inter-class distance. Concretely, the euclidean distance between a sample \\( z_{i} \\) and a set \\( S \\) is defined as \\( \\mathbf{D}(z_i,S) = \\frac{1}{|S|}\\sum_{z_j\\in S}\\| z_i - z_j\\| _2 \\). Then, the intra- and interclass distance of sample \\( z_{i} \\) can be defined as \\( \\mathbf{D}(z_i,P_i) \\) and \\( \\mathbf{D}(z_i,\\mathcal{D} / P_i) \\) separately; and the intra- and inter-subclass" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.332, + 0.892, + 0.363 + ], + "angle": 0, + "content": "distance of sample \\(z_{i}\\) can be defined as \\(\\mathbf{D}(z_i,M_i)\\) and \\(\\mathbf{D}(z_i,P_i / M_i)\\) separately." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.895, + 0.56 + ], + "angle": 0, + "content": "Figure 2 shows the average distance between features learned by SBCL in subclasses and classes on the CIFAR-100-LT. As shown in Figure 2a, the distance between samples from the same subclass is less than those from the same class but different subclasses. Meanwhile, in Figure 2b, the inter-class distance consistently exceeds the intra-class distance, indicating clear separation between features from different classes. Additionally, the inter-class distance remains stable, which suggests that each class is uniformly distributed on a hypersphere. The results in all indicate that the two-layer class hierarchy is successfully captured and feature distribution achieves the core idea of SBCL. The specific values of feature distance are displayed in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.895, + 0.682 + ], + "angle": 0, + "content": "Figure 3 visualizes the feature distribution of subclasses in a head class and head classes in CIFAR-100-LT with imbalance ratio 100 trained by SBCL using t-SNE [59]. The representations learned by SBCL form some separated clusters, which be aligned with subclass labels and class labels. This suggests that SBCL can learn an embedding space where features from same subclasses/classes are pulled closely and features from different subclasses/class are separated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.878, + 0.711 + ], + "angle": 0, + "content": "4.8. Variation of per-class weight during training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.824 + ], + "angle": 0, + "content": "Weight decay [1] achieves superior performance by effectively tuning weight decay during training. This study highlights the significance of regularizing network weights to become balanced for long-tailed recognition. Therefore, it is crucial to investigate whether SBCL can help the linear classifier to learn balanced weights through the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Figure 4 shows the per-class weight norm of a linear classifier trained on top of features learned by SBCL in different training stages on CIFAR-100-LT. From the figure, we can see that as the training proceeds, the per-class weight norm becomes balanced even when training the linear classifier," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5402" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.09, + 0.437, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.267, + 0.436, + 0.28 + ], + "angle": 0, + "content": "(a) Distribution of subclasses on a head class." + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.091, + 0.804, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.267, + 0.806, + 0.281 + ], + "angle": 0, + "content": "(b) Distribution of head classes on the dataset." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.893, + 0.325 + ], + "angle": 0, + "content": "Figure 3: Feature distribution of subclasses and classes on CIFAR-100-LT with imbalance ratio 100. Color represents subclasses/classes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.351, + 0.332, + 0.366 + ], + "angle": 0, + "content": "the original cross-entropy loss is used." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.383, + 0.432, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.564, + 0.472, + 0.61 + ], + "angle": 0, + "content": "Figure 4: Change in weight norm of the linear classifier based on the representations trained by SBCL on CIFAR-100-LT with imbalance ratio 100 during training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.216, + 0.634 + ], + "angle": 0, + "content": "5. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Traditional methods for handling long-tailed recognition problem includes re-sampling and re-weighting. There are roughly two types of re-sampling techniques: over-sampling the minority classes [4, 5, 54, 70] and under-sampling the frequent classes [4, 23, 32]. The re-weighting techniques assign adaptive weights for different classes or even different samples. The vanilla scheme re-weights classes proportionally to the inverse of their frequency [29, 30, 64]. For class-level re-weighting methods, many loss functions including CB loss [15], LDAM loss [6] and Balanced softmax loss [50] were recently proposed, while instance-level re-weighting methods include Focal loss [44] and Influence-balanced loss [49]. Recently, two-stage algorithms have achieved remarkable performance for long-tailed recognition, such as classifier re-training (cRT) [34], learnable weight scaling (LWS) [34], and Mixup Shifted Label-Aware Smoothing model (MiSLAS) [72]. Meanwhile, bilateral branch net" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.895, + 0.487 + ], + "angle": 0, + "content": "work (BBN) [73] uses an additional network branch for rebalancing. RIDE [62] use multiple branches named experts, each learning to specialize in the entire classes. LADE [28] assumes the prior of test class distributions is available and accordingly post-adjust model predictions. PaCo [14] applies parametric class-wise learnable centers to rebalance in contrastive learning. BCL [74] proposes a multi-branch framework to achieve class-averaging and class-complement in the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.896, + 0.728 + ], + "angle": 0, + "content": "To boost the performance of the two-stage algorithms, researchers have introduced supervised contrastive learning [36] to the first feature-learning stage and proposed \\( k \\)-positive contrastive loss (KCL) [33] and targeted supervised contrastive learning (TSC) [41]. While achieving the state-of-the-art performance, these methods inject class-balance in the contrastive learning objective, inevitably leading to instance-imbalance during training. In this work, we instead propose to achieve both subclass- and instance-balance in the contrastive learning object. Our method is also related to recent studies of clustering-based deep unsupervised learning [7,8,18,43,65,66], especially those that leverage contrastive learning [22,40,42,63]. However, they target at general unsupervised representation learning scenario, while our method is tailored for long-tailed recognition where the training data is immensely class-imbalanced." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.619, + 0.755 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduced Subclass-balancing Contrastive Learning (SBCL) for long-tailed recognition. It breaks down the head classes into multiple semantically-coherent subclasses via subclass-balancing adaptive clustering and incorporates a bi-granularity contrastive loss that encourages both subclass- and instance-balance. Extensive experiments on multiple datasets demonstrate that SBCL achieves state-of-the-art single-model performance on benchmark datasets for long-tailed recognition." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5403" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.471, + 0.19 + ], + "angle": 0, + "content": "[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6897-6907, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.194, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Shin Ando and Chun Yuan Huang. Deep over-sampling framework for classifying imbalanced data. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 770-785. Springer, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.273, + 0.471, + 0.333 + ], + "angle": 0, + "content": "[3] David Arthur and Sergei Vassilvitskii. K-means++ the advantages of careful seeding. In Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms, pages 1027-1035, 2007. 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.336, + 0.471, + 0.396 + ], + "angle": 0, + "content": "[4] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106:249-259, 2018. 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.4, + 0.471, + 0.46 + ], + "angle": 0, + "content": "[5] Jonathon Byrd and Zachary Lipton. What is the effect of importance weighting in deep learning? In International Conference on Machine Learning, pages 872-881. PMLR, 2019. 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.464, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[6] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in Neural Information Processing Systems, 32, 2019. 1, 5, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.541, + 0.471, + 0.602 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In European Conference on Computer Vision, 2018. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.606, + 0.471, + 0.68 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 2020. 5, 6, 9, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.684, + 0.471, + 0.744 + ], + "angle": 0, + "content": "[9] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. Smote: synthetic minority over-sampling technique. Journal of artificial intelligence research, 16:321-357, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.471, + 0.821 + ], + "angle": 0, + "content": "[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5008-5017, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.826, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.143, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[13] Collobert, Ronan, Weston, and Jason. A unified architecture for natural language processing: deep neural networks with multitask learning. In International Conference on Machine Learning, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.894, + 0.269 + ], + "angle": 0, + "content": "[14] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.894, + 0.349 + ], + "angle": 0, + "content": "[15] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9268-9277, 2019. 1, 5, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.355, + 0.894, + 0.43 + ], + "angle": 0, + "content": "[16] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.436, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[17] Qi Dong, Shaogang Gong, and Xiatian Zhu. Imbalanced deep learning by minority class incremental rectification. IEEE transactions on pattern analysis and machine intelligence, 41(6):1367-1381, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.502, + 0.894, + 0.577 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. Advances in Neural Information Processing Systems, 27, 2014. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.583, + 0.894, + 0.628 + ], + "angle": 0, + "content": "[19] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.894, + 0.709 + ], + "angle": 0, + "content": "[20] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In Proceedings of the IEEE/cvf International Conference on computer vision, pages 6391-6400, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.894, + 0.82 + ], + "angle": 0, + "content": "[21] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.826, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[22] Yuanfan Guo, Minghao Xu, Jiawen Li, Bingbing Ni, Xuanyu Zhu, Zhenbang Sun, and Yi Xu. Hcsc: Hierarchical contrastive selective coding. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2022. 9" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5404" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.137 + ], + "angle": 0, + "content": "[23] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9):1263-1284, 2009. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.141, + 0.472, + 0.216 + ], + "angle": 0, + "content": "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 6, 16, 17, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.472, + 0.28 + ], + "angle": 0, + "content": "[25] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.283, + 0.472, + 0.358 + ], + "angle": 0, + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.472, + 0.438 + ], + "angle": 0, + "content": "[27] David T Hoffmann, Nadine Behrmann, Juergen Gall, Thomas Brox, and Mehdi Noroozi. Ranking info noise contrastive estimation: Boosting contrastive learning via ranked positives. the Association for the Advance of Artificial Intelligence, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.472, + 0.532 + ], + "angle": 0, + "content": "[28] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, 2021. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.536, + 0.472, + 0.611 + ], + "angle": 0, + "content": "[29] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5375-5384, 2016. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.615, + 0.472, + 0.689 + ], + "angle": 0, + "content": "[30] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Deep imbalanced learning for face recognition and attribute prediction. IEEE transactions on pattern analysis and machine intelligence, 42(11):2781-2794, 2019. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.694, + 0.472, + 0.754 + ], + "angle": 0, + "content": "[31] Ahmet Iscen, André Araujo, Boqing Gong, and Cordelia Schmid. Class-balanced distillation for long-tailed visual recognition. Proceedings of the IEEE international conference on computer vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[32] Nathalie Japkowicz and Shaju Stephen. The class imbalance problem: A systematic study. Intelligent data analysis, 6(5):429-449, 2002. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.807, + 0.472, + 0.867 + ], + "angle": 0, + "content": "[33] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 1, 2, 5, 6, 9, 16, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.871, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[34] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Kalantidis. Decoupling representation and classifier for long-tailed recognition. International Conference on Learning Representations, 2019. 2, 5, 6, 9, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.141, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[35] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7482-7491, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.894, + 0.297 + ], + "angle": 0, + "content": "[36] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 1, 2, 5, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.303, + 0.894, + 0.378 + ], + "angle": 0, + "content": "[37] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13896-13905, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[38] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.894, + 0.463 + ], + "angle": 0, + "content": "[39] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.894, + 0.529 + ], + "angle": 0, + "content": "[40] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 4, 5, 6, 9, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.894, + 0.624 + ], + "angle": 0, + "content": "[41] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6, 7, 9, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.894, + 0.675 + ], + "angle": 0, + "content": "[42] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. AAAI Conference on Artificial Intelligence, 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.68, + 0.894, + 0.74 + ], + "angle": 0, + "content": "[43] Renjie Liao, Alex Schwing, Richard Zemel, and Raquel Urtasun. Learning deep parsimonious representations. Advances in Neural Information Processing Systems, 29, 2016. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.894, + 0.82 + ], + "angle": 0, + "content": "[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2980-2988, 2017. 5, 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5405" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.091, + 0.472, + 0.168 + ], + "angle": 0, + "content": "[46] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.171, + 0.472, + 0.261 + ], + "angle": 0, + "content": "[47] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In European conference on computer vision, pages 181-196, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[48] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.472, + 0.405 + ], + "angle": 0, + "content": "[49] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 735-744, 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.409, + 0.472, + 0.47 + ], + "angle": 0, + "content": "[50] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. Advances in Neural Information Processing Systems, 33:4175–4186, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.473, + 0.472, + 0.535 + ], + "angle": 0, + "content": "[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.472, + 0.628 + ], + "angle": 0, + "content": "[52] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.472, + 0.708 + ], + "angle": 0, + "content": "[53] Xie Saining, Girshick Ross, Dollar Piotr, Tu Zhuowen, and He Kaiming. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page 1492-1500, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.712, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[54] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In European conference on computer vision, pages 467-482. Springer, 2016. 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.472, + 0.836 + ], + "angle": 0, + "content": "[55] Tony Shen, Ariel Lee, Carol Shen, and C Jimmy Lin. The long tail and rare disease research: the impact of next-generation sequencing for rare mendelian disorders. Genetics research, 97, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.84, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[56] Aditya Singh and Alessandro Bay. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in Neural Information Processing Systems, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.091, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.091, + 0.893, + 0.136 + ], + "angle": 0, + "content": "[57] Merrielle Spain and Pietro Perona. Measuring and predicting importance of objects in our visual world. 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.142, + 0.895, + 0.218 + ], + "angle": 0, + "content": "[58] AFM Uddin, Mst Monira, Wheemyung Shin, Tae-Choong Chung, Sung-Ho Bae, et al. Saliency: A saliency guided data augmentation strategy for better regularization. International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.895, + 0.267 + ], + "angle": 0, + "content": "[59] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.273, + 0.895, + 0.363 + ], + "angle": 0, + "content": "[60] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.369, + 0.895, + 0.429 + ], + "angle": 0, + "content": "[61] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.895, + 0.493 + ], + "angle": 0, + "content": "[62] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. International Conference on Learning Representations, 2021. 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.499, + 0.895, + 0.574 + ], + "angle": 0, + "content": "[63] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021. 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.579, + 0.895, + 0.625 + ], + "angle": 0, + "content": "[64] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. Advances in Neural Information Processing Systems, 30, 2017. 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.895, + 0.69 + ], + "angle": 0, + "content": "[65] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.695, + 0.895, + 0.769 + ], + "angle": 0, + "content": "[66] Jianwei Yang, Devi Parikh, and Dhruv Batra. Joint unsupervised learning of deep representations and image clusters. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5147-5156, 2016. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.895, + 0.835 + ], + "angle": 0, + "content": "[67] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. Advances in Neural Information Processing Systems, 33:19290-19301, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[68] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. International Conference on Machine Learning, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.091, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "5406" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.091, + 0.472, + 0.152 + ], + "angle": 0, + "content": "[69] Han-Jia Ye, Hong-You Chen, De-Chuan Zhan, and Wei-Lun Chao. Identifying and compensating for feature deviation in imbalanced deep learning. arXiv preprint arXiv:2001.01385, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.156, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[70] Qiaoyong Zhong, Chao Li, Yingying Zhang, H Sun, S Yang, D Xie, and S Pu. Towards good practices for recognition & detection. In CVPR workshops, volume 1, page 2, 2016. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.472, + 0.311 + ], + "angle": 0, + "content": "[71] Yaoyao Zhong, Weihong Deng, Mei Wang, Jiani Hu, Jianteng Peng, Xunqiang Tao, and Yaohai Huang. Unequal-training for deep face recognition with long-tailed noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7812-7821, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.315, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[72] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16489-16498, 2021. 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.394, + 0.472, + 0.47 + ], + "angle": 0, + "content": "[73] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.473, + 0.472, + 0.563 + ], + "angle": 0, + "content": "[74] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.472, + 0.611 + ], + "angle": 0, + "content": "[75] George Kingsley Zipf. The psycho-biology of language: An introduction to dynamic philology. Routledge, 2013. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.091, + 0.472, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5407" + } + ] +] \ No newline at end of file diff --git a/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/bc05a493-9765-4ef8-8b9c-fc9c90911ac5_origin.pdf b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/bc05a493-9765-4ef8-8b9c-fc9c90911ac5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e196b126a2756e27d7e265ae34f61c32c79dd508 --- /dev/null +++ b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/bc05a493-9765-4ef8-8b9c-fc9c90911ac5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fec7abd74a3cacf462c58e5018bf2d0a399d671e0ae4a8bedddff679ab8e91c +size 1654297 diff --git a/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/full.md b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e4f00fd1b46ca8f837d62dfd5765999a81a7df5f --- /dev/null +++ b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/full.md @@ -0,0 +1,340 @@ +# Subclass-balancing Contrastive Learning for Long-tailed Recognition + +Chengkai Hou1, Jieyu Zhang2, Haonan Wang3, Tianyi Zhou4* +1Jilin university, 2University of Washington, 3National University of Singapore +4University of Maryland + +houck20@mails.jlu.edu.cn, jieyuz2@cs.washington.edu, haonan.wang@u.nus.edu.tianyi@umd.edu + +# Abstract + +Long-tailed recognition with imbalanced class distribution naturally emerges in practical machine learning applications. Existing methods such as data reweighing, resampling, and supervised contrastive learning enforce the class balance with a price of introducing imbalance between instances of head class and tail class, which may ignore the underlying rich semantic substructures of the former and exaggerate the biases in the latter. We overcome these drawbacks by a novel "subclass-balancing contrastive learning (SBCL)" approach that clusters each head class into multiple subclasses of similar sizes as the tail classes and enforce representations to capture the two-layer class hierarchy between the original classes and their subclasses. Since the clustering is conducted in the representation space and updated during the course of training, the subclass labels preserve the semantic substructures of head classes. Meanwhile, it does not overemphasize tail class samples, so each individual instance contribute to the representation learning equally. Hence, our method achieves both the instance- and subclass-balance, while the original class labels are also learned through contrastive learning among subclasses from different classes. We evaluate SBCL over a list of long-tailed benchmark datasets and it achieves the state-of-the-art performance. In addition, we present extensive analyses and ablation studies of SBCL to verify its advantages. Our code is available at https://github.com/JackHck/subclass-balancing-contrastive-learning. + +# 1. Introduction + +In reality, the datasets often follow the Zipfian distribution over classes with a long tail [57, 75], i.e., a few classes (head classes) containing significantly more instances than the remaining tail classes. Such tail classes could be of great importance for high-stake applications, e.g., patient + +class in medical diagnosis or accident class in autonomous driving [6, 55]. However, training on such class-imbalanced datasets can result in a severely biased model with noticeable performance drop in classification tasks [2, 4, 13, 47, 64, 68, 71]. + +To overcome the challenges posed by long-tailed data, data resampling [2, 4, 9, 54] and loss reweighing [5, 6, 15, 17] have been widely applied but they cannot fully leverage all the head-class samples. Very recent work discovered that supervised contrastive learning (SCL) [36] can achieve state-of-the-art (SOTA) performance on benchmark datasets of long-tailed recognition [33, 41]. Specifically, the $k$ -positive contrastive learning (KCL) [33] and its subsequent work targeted supervised contrastive learning (TSC) [41] revamp SCL by encouraging the learned feature space to be class-balanced and uniformly distributed. However, those methods enforcing class-balance often come with a price of instance-imbalance, i.e., each individual instance of tail classes would have much greater impact on model training than that of head classes. + +Such instance-imbalance can result in significant degradation of the performance on long-tailed recognition for several reasons. On the one hand, the limited samples in each tail class might not be sufficiently representative of the whole class. So even a small bias of them can be enormously exaggerated by class-balancing methods and result in sub-optimal learning of classifiers or representations. On the other hand, head classes usually have more complicated semantic substructures, e.g., multiple high-density regions of the data distribution, so simply downweighing samples of head classes and treating them equally can easily lose critical structural information. For example, images of a head class "cat" might be highly diverse in breeds and colors, which need to be captured by different features but downweighing or subsampling them may easily lose such information, while a tail class "platypus" might only contain a few similar images that are unlikely to cover all the representative features. Therefore, it is non-trivial to enforce both class-balance and instance-balance simultaneously in the same method. + +Can we remove the negative impact of class-imbalance while still retain the advantages of instance-balance? In this paper, we achieve both through subclass-balancing contrastive learning (SBCL), a novel supervised contrastive learning defined on subclasses, which are the clusters within each head class, have comparable size as tail classes, and are adaptively updated during the training. Instead of sacrificing instance-balance for class-balance, our method achieves both instance- and subclass-balance by exploring the head-class structure in the learned representation space of the model-in-training. In particular, we propose a bi-granularity contrastive loss that enforces a sample (1) to be closer to samples from the same subclass than all the other samples; and (2) to be closer to samples from a different subclass but the same class than samples from any other subclasses. While the former learns representations with balanced and compact subclasses, the latter preserves the class structure on subclass level by encouraging the same class's subclasses to be closer to each other than to any different class's subclasses. Hence, it can learn an accurate classifier distinguishing original classes while enjoy both the instance- and subclass-balance. + +In this paper, we apply SBCL for several visual recognition tasks to demonstrate SBCL superiority over other previous works (e.g., KCL [33], TSC [41]). To summarize, this paper makes the following contributions: + +(a). We provide a new design principal of leveraging supervised contrastive learning for long-tailed recognition, i.e., aiming at achieving both instance- and subclass-balance instead of class-balance at the expense of instance-balance. +(b). We propose a novel instantiation of the aforementioned design principal, subclass-balancing contrastive learning (SBCL), which consists of two major components, namely, subclass-balancing adaptive clustering and bi-granularity contrastive loss. +(c). Empirically, we compare the SBCL against state-of-the-art methods on three visual tasks: image classification, object detection, and instance segmentation to demonstrate its effectiveness on handling class imbalance. We also conduct a series of experiments to analyze the efficacy of SBCL. + +# 2. Background and notations + +Long-tailed recognition. Long-tailed recognition aims to learn a classifier from a training dataset with long-tailed class distribution, i.e., a few classes contain many data (head classes) while most classes contain only a few data (tail classes), where the major challenge is to require model recognizing all classes equally well. Let $\mathcal{D} = \{x_i, y_i\}_{i \in [n]}$ be a long-tailed training dataset, where $x_i$ denotes a sample and $y_i \in [C]$ denotes its label. Denote by $\mathcal{D}_k \subseteq \mathcal{D}$ the set of + +instances belonging to class $k$ and $n_k = |\mathcal{D}_k|$ is the number of samples in class $k$ . The total number of training samples over $C$ classes is $n = \sum_{k=1}^{C} n_k$ . Without loss of generality, we follow prior work [28, 34] to assume that the classes are sorted by cardinality in decreasing order (i.e., if $i < j$ , then $n_i \geq n_j$ ), and $n_1 \gg n_C$ . In addition, we define the imbalance ratio as $\max_{k \in [C]} (n_k) / \min_{k \in [C]} (n_k) = n_1 / n_C$ . Finally, let $f_\theta(\cdot)$ be a deep feature extractor, e.g., a neural network, parameterized by $\theta$ and $\mathbf{w}_c$ is the linear classifier of class $c$ , then the classifier we aim to learn is $h(x_i) = \arg \max_{c \in [C]} \mathbf{w}_c^\top f_\theta(x_i)$ . + +Supervised contrastive learning. Recent studies have shown that supervised contrastive learning (SCL) [36] provides a strong performance gain for long-tailed recognition and its variants have achieved state-of-the-art (SOTA) performance [33, 41]. Specifically, SCL learns the feature extractor $f_{\theta}(\cdot)$ via maximizing the discriminativeness of positive instances, i.e., instances from the same class, and the learning objective for a single training data $(x_i, y_i)$ in a batch $\mathcal{B} = \{(x_i, y_i)\}_{i \in [N]}$ , is + +$$ +\mathcal {L} _ {S C L} = \sum_ {i = 1} ^ {N} - \frac {1}{| \tilde {P} _ {i} |} \sum_ {z _ {p} \in \tilde {P} _ {i}} \log \frac {\exp \left(z _ {i} \cdot z _ {p} ^ {\top} / \tau\right)}{\sum_ {z _ {a} \in \tilde {V} _ {i}} \exp \left(z _ {i} \cdot z _ {a} ^ {\top} / \tau\right)}, \tag {1} +$$ + +where $\tau$ is the temperature hyperparameter, $z_{i} = f_{\theta}(x_{i})$ is the feature generated from $x_{i}$ , $V_{i} = \{z_{i}\}_{i\in [N]}\setminus \{z_{i}\}$ is the current batch of features except for $z_{i}$ , $P_{i} = \{z_{j}\in V_{i}:y_{j} = y_{i}\}$ is a set of instances with the same label as $x_{i}$ . Finally, let $\tilde{z}_i$ be the feature of $\tilde{x}_i$ , the augmented version of $x_{i}$ , and for any set $S_{i}$ indexed by $i$ , we use $\tilde{S}_i = S_i\cup \{\tilde{z}_i\}$ , e.g., $\tilde{V}_i = V_i\cup \{\tilde{z}_i\}$ . However, for the long-tailed datasets, the feature spaces is dominated by head classes and thus have limited capability of semantic discrimination [33]. To address this, the $k$ -positive contrastive learning (KCL) [33] attempts to balance the feature space by keeping the number of positive instances in $\tilde{P}_i$ equal for each class, leading to the following loss + +$$ +\mathcal {L} _ {K C L} = \sum_ {i = 1} ^ {N} - \frac {1}{k + 1} \sum_ {z _ {p} \in \tilde {P _ {i} ^ {k}}} \log \frac {\exp \left(z _ {i} \cdot z _ {p} ^ {\top} / \tau\right)}{\sum_ {z _ {a} \in \tilde {V _ {i}}} \exp \left(z _ {i} \cdot z _ {a} ^ {\top} / \tau\right)} (2) +$$ + +where $P_{i}^{k}$ is a subset of $P_{i}$ with $k$ randomly drawn instances. Finally, the learned feature extractor $f_{\theta}(\cdot)$ is exploited in a sequel stage of training the classifier for long-tailed recognition [33, 41]. + +# 3. Methodology + +As mentioned above, KCL and its sequels [33,41] balance the learning objective of SCL by picking the same number of positive instances for each class, i.e., $|P_i^k| = k$ in Eq. 2 no matter which class $x_i$ belongs to, however, we argue that + +![](images/f6d07b822ac00f11d4981192a1450f50fb0540bac0dc65db54ec7be3be90840d.jpg) +(a). Subclass-balancing adaptive clustering (Section 3.1) + +![](images/2ec4e63648de9990919d8b2dbdd950dd170ee77df32b54e9bdf0b2be7a492cf6.jpg) +(b). Bi-granularity contrastive loss (Section 3.2) +Figure 1: Illustration of subclass-balancing contrastive learning (SBCL). It initially divides the head classes into multiple subclasses of comparable size. Then, during training, SBCL builds each sample to be closer to samples from the same subclass than samples from different subclasses but the same class, which are also made to be closer than samples from different classes. + +such a class-balancing approach would inevitably introduce instance-imbalance: the instances of tail classes have much more chances to be engaged in the training than that of head class. Specifically, assume each class has no less than $k$ instances, then the probability of an instance of class $c$ being selected as positive instance is $p(c) = \frac{k}{n_c}$ ; if the tail class $n_C$ has only $k$ instances and the imbalance ratio is $\frac{n_1}{n_C} = 100$ , then we have $p(1) = \frac{k}{100k} = 0.01$ while $p(C) = 1$ . We can see that when the instances of head class are selected once, that of tail class may already be trained 100 times. Thus, the training is immensely biased towards the few samples in each tail class. Besides, as tail classes only have very few instances that are not necessarily representative, the learned feature space might be unsatisfactory and sensitive to the training data of tail classes. + +Here, we provide a new prospective of handling class-imbalance issue by contrastive learning: instead of aiming at class-balance at the expense of instance-imbalance, we propose to achieve both instance- and subclass-balance. We argue that head classes typically contain more diverse instances and thus have richer semantics in the training dataset. Therefore, it might be wise to break down the head classes into multiple semantically coherent subclasses, each of which consists of similar number of instance as tail classes. Built on this spirit, we develop subclass-balancing contrastive learning (SBCL), a new contrastive learning framework for long-tailed recognition (visualized in Figure 1) that achieves both instance- and subclass-balance. We present a concrete example that effectively demonstrates the superiority of SBCL compared to previous methods in Section 4.6. + +# 3.1. Subclass-balancing adaptive clustering + +We break down the class into several "subclasses" to attack the imbalanced phenomenon. Particularly, given a class + +$c$ and the associated set of data $\mathcal{D}_c$ , we employ a clustering algorithm of choice based on the features extracted by current feature extractor $f_{\theta}(\cdot)$ to divide $\mathcal{D}_c$ into $m_c$ subclasses/clusters. We use $\Gamma_c(x_i)$ to denote the cluster label of an instance $x_i$ of class $c$ . To ensure that the number of samples for each subclass is roughly the same, we propose a new cluster algorithm to divide the unit-length feature vectors, i.e., the features output by $f_{\theta}(\cdot)$ with additional unit-length normalization. The new proposed cluster algorithm is described in Algorithm 1. + +In Algorithm 1, during the initialization of cluster centers, our approach gradually selects cluster centers by prioritizing the samples that are maximally distant from other existing cluster centers [3]. In the process of assigning vectors to their centers, we set a threshold $M$ as an upper limit of sample size in a cluster, which guarantees clusters of balanced sizes. + +Specifically, the threshold $M$ is + +$$ +M = \max \left(n _ {C}, \delta\right) \tag {3} +$$ + +where $n_C$ is the size of tail class and the hyperparameter $\delta$ controls the lower bound of sample size in clusters to prevent overly-small cluster. Note that we only apply clustering algorithm to classes which contain multiple instances while the tail classes remain unchanged. As a consequence, the size of each resultant cluster is similar to that of tail class $n_C$ . In addition, instead of only clustering once at the beginning, we update the cluster assignment adaptively based on the current feature extractor $f_{\theta}(\cdot)$ during training process and empirically show that such adaptive clustering outperforms only-once clustering in Section 4.5. We add an exhaustive description of Algorithm 1 and show the distribution of the sample size in clusters on the benchmark dataset in Appendix A.1. + +Then, by replacing the class labels of head classes used in SCL/KCL with the finer-grained cluster labels, we ensure the + +Algorithm 1 Subclass-balancing Adaptive Clustering +Require: Sample set $S = \{x_{i}\}_{i = 1}^{n}$ ; A threshold $M$ ; The number of iterations $K$ +Ensure: Cluster assignments for samples in $s$ +for $k = 0$ to $K$ do if $k = 0$ then Choose the cluster centers $y_{j}$ which are farther away from previously selected centers. else Update the cluster centers $y_{j} = \frac{1}{n_{j}}\sum_{i = 1}^{n_{j}}x_{i}.$ $\triangleright n_j$ is the number of samples in a cluster end if Construct the cluster center set $\mathcal{C} = \{y_j\}_{j = 1}^m$ . $\triangleright m$ is the number of cluster centers while $S\neq \phi$ do $\triangleright$ Assign samples to centers $y_{i}$ Select the most similar pair $(x_{i},y_{j}) = \arg \max \mathrm{cosine - similarity}(x,y)$ $x\in S,y\in C$ Assign the sample $x_{i}$ to the center $y_{j}$ . Delete the assigned sample $x_{i}$ from the sample set $S = S / \{x_i\}$ if $n_j\geq M$ then Sample number in a cluster exceeds the threshold $M$ Delete the cluster center $y_{j}$ from the cluster center set $\mathcal{C} = \mathcal{C} / \{y_j\}$ end if end while end for + +instance-balanced, i.e., each instance has similar chance of being selected regardless of its class. By breaking down the head classes, which typically contain more diverse instances, into multiple semantically coherent subclasses, we achieve subclass-balanced (instead of class-balanced) while maintain the rich semantics rendered by head classes in training dataset. + +# 3.2. Bi-granularity contrastive loss + +We now have two types of label for instances in head classes from different granularities: the coarse-grained class label and the fine-grained cluster label. A direct consequence of replacing class label in SCL/KCL with cluster label is that we no longer distinguish instances from different head classes, and therefore the boundaries between classes might be blurry, leading to sub-optimal feature space. As a remedy, we combine the contrastive loss of both class label and cluster label into the following one and reuse the notations of Eq. 1: + +$$ +\begin{array}{l} \mathcal {L} _ {S B C L} = - \sum_ {i = 1} ^ {N} \left(\frac {1}{| \tilde {M} _ {i} |} \sum_ {z _ {p} \in \tilde {M} _ {i}} \log \frac {\exp \left(z _ {i} \cdot z _ {p} ^ {\top} / \tau_ {1}\right)}{\sum z _ {a} \in \tilde {V} _ {i}} \exp \left(z _ {i} \cdot z _ {a} ^ {\top} / \tau_ {1}\right) \right. \\ + \beta \frac {1}{\left| \tilde {P} _ {i} \right| - \left| M _ {i} \right|} \sum_ {z _ {p} \in \tilde {P} _ {i} / M _ {i}} \log \frac {\exp \left(z _ {i} \cdot z _ {p} ^ {\top} / \tau_ {2}\right)}{\sum_ {z _ {a} \in \tilde {V} _ {i} / M _ {i}} \exp \left(z _ {i} \cdot z _ {a} ^ {\top} / \tau_ {2}\right)}) \tag {4} \\ \end{array} +$$ + +where $M_{i} = \{z_{j}\in P_{i}:\Gamma_{y_{i}}(x_{i}) = \Gamma_{y_{i}}(x_{j})\}$ is a set of instances with the same cluster label as $x_{i}$ . $\beta$ is a hyperparameter that balances these two loss terms. The first term corresponds to the SCL loss with cluster labels, while the second term leverages the class label but does not consider the instances of the same cluster, i.e., the instances in $M_{i}$ are + +removed in the second term. Such a design choice reflects the two types of positive instances for $z_{i}$ : (1) the instance in the same cluster and (2) the instance of the same class but in different clusters. + +According to previous studies [27, 40, 61], the temperature $\tau$ in contrastive loss is critical in controlling the local separation and global uniformity of the feature distribution. Specifically, for supervised contrastive learning, a low temperature makes relative high penalty on feature distribution, that actually encourages the features distribute more concentrically. As the temperature increases, the relative penalty tends to be more uniform which uniformizes the distribution of the features. Although the above objective explicitly considers the two types of label from different granularities, it still treats class and cluster label similarly. Intuitively, we expect instances of the same subclass to form a more concentrated cluster in feature space than those of the same class, since subclass naturally indicates finer-grained semantic coherence. To achieve this, we ensure the temperature $\tau_{2} > \tau_{1}$ and dynamically adjust $\tau_{2}$ for each class according to its current level of concentration of the instances' feature. Following [40], for class $c$ we define $\phi(c)$ as + +$$ +\phi (c) = \frac {\sum_ {i = 1} ^ {n _ {c}} \| z _ {i} - t _ {c} \| _ {2}}{n _ {c} \log \left(n _ {c} + \alpha\right)} \tag {5} +$$ + +where $t_c$ is the centroid for the class $c$ , $\alpha$ is a hyperparameter to ensure that $\phi(c)$ is not overly-large, and $z_i$ corresponds to instances of class $c$ . From the formulation, we can see that if the current averaged distance to the class centroid is large or the class contains fewer data, thus the temperature will be set large to adopt the feature distribution of class $c$ during the training process. Then we define the temperature of class + +$c$ as + +$$ +\tau_ {2} (c) = \tau_ {1} \cdot \exp \left(\frac {\phi (c)}{\frac {1}{C} \sum_ {i = 1} ^ {C} \phi (i)}\right) \tag {6} +$$ + +such that $\tau_{2}(c)$ for class label is always larger than $\tau_{1}$ for cluster label (since $\phi(c) > 0$ ) and could reflect the current level of concentration of the instances in a class. In particular, the proposed $\tau_{2}(c)$ encourages the features of instances in class $c$ to form a less tight cluster than that of a subclass (by $\tau_{2}(c) > \tau_{1}$ ) while adaptively adjust the temperature to prevent an overly-loose/dense cluster. + +# 3.3. Training algorithm + +Here, we describe the overall training process of subclassbalancing contrastive learning and the algorithm can be found in Algorithm 2. First, the adaptive clustering (Section 3.1) could be noisy at the early stage of training [40,63]. Thus, we warm-up the feature extractor $f_{\theta}(\cdot)$ by a few epochs of training on ordinary SCL or KCL loss. In addition, our algorithm involves two adaptively-adjusting parts, namely, the cluster assignment and the temperature $\tau_{2}(c)$ for each head class $c$ . Instead of updating these every epoch, we use a hyperparameter $K$ as the update interval, i.e., we update the cluster assignment and the temperature based on the current learned $f_{\theta}(\cdot)$ every $K$ epoch. + +Algorithm 2 Training Algorithm +Require: Dataset $\mathcal{D} = \{x_i,y_i\}_{i\in [n]};$ ; The update interval of cluster assignment $K$ ; The number of warm-up epoch $T_0$ ; The total number of epoch $T$ ; The hyperparameters $\beta$ and $\delta$ +Ensure: A trained feature extractor $f_{\theta}(\cdot)$ +1: Initialize the model parameters $\theta$ +2: Train $f_{\theta}(\cdot)$ with SCL/KCL for $T_0$ epochs Warm-up stage +3: for $t = T_0$ to $T$ do +4: if $t\% K = = 0$ then Update cluster and temperature +5: Update the cluster assignment based on the current feature extractor $f_{\theta}(x)$ +6: Update the temperture $\tau_{2}$ for each head class using Eq. 5 and Eq. 6 +7: end if +8: Train $f_{\theta}(\cdot)$ using Eq. 4 Subclass-balancing contrastive learning +9: end for + +# 4. Experiment + +# 4.1. Experimental setup + +Datasets. We consider three commonly used long-tailed recognition benchmark datasets: CIFAR-100-LT [6], ImageNet-LT [46], and iNaturalist 2018 [60]. The CIFAR-100-LT and ImageNet-LT datasets are artificially generated long-tailed datasets from the class-balanced datasets [38,52], and the iNaturalist 2018 dataset is a large-scale real-world dataset that exhibits long-tailed imbalance. + +Baselines. We consider baseline methods of the following three categories: (1) class-balancing classifiers, including $\tau$ -norm, LWS and cRT [34], which fixes the representation which trained by cross-entropy loss and trains the classifier with class-balanced sampling; (2) one-stage balancing loss, including CB loss [15], Focal loss [44], and LDAM loss [6]. These supervised distribution-aware loss makes the model to pay more attention on the minority class during training. (3) contrastive learning methods, including SCL [36], KCL [33], SwAV [8], PCL [40] and TSC [41] which train a feature extractor with the contrastive loss and then learn a classifier given the trained feature extractor. + +Evaluation protocol. Following [33, 34, 41], we implement SBCL, as well as other contrastive learning methods, in a two-stage framework. In the first stage, we train the feature extractor with a contrastive learning method, while in the second stage, we train a linear classifier on top of the learned representation. Specifically, for CIFAR-100-LT dataset, the linear classifier is trained with LDAM loss and class re-weighting [6]. For ImageNet-LT and iNaturalist 2018 datasets, the linear classifier is trained with CE loss and class-balanced sampling [34]. All results are averaged over 5 trials with different random seeds. We mainly report the overall top-1 accuracy. For the two large datasets, ImageNet-LT and iNaturalist 2018 datasets, following the previous work [46], we also report the accuracy of three disjoint subsets: Many-shot classes (classes with more than 100 samples), Medium-shot classes (classes with 20 to 100 samples), and Few-shot classes (classes under 20 samples). We show the implementation details and important hyperparameters in the Appendix A.2. + +# 4.2. Main results + +The results on ImageNet-LT and iNaturalist 2018 are in Table 1. We can see that SBCL outperforms the baselines with a large margin over the two datasets. In addition, on iNaturalist 2018 dataset, SBCL outperforms the previous SOTA method by $0.7\%$ on Many, $1.3\%$ on Medium, $0.8\%$ on Few and $1.1\%$ on All, which shows the effectiveness of the proposed method in solving real-world long-tailed recognition problems such as natural species classification. Besides, SBCL is also better than existing contrastive learning method like KCL and TSC for all class splits, which demonstrates the effectiveness of the design principal of pursuing both instances- and subclass-balance in contrastive learning. Table 2 summarizes the results on CIFAR-100-LT dataset. For CIFAR-100-LT dataset, SBCL outperforms previous SOTA methods except for imbalance ratio 10. We hypothesize that it is because the tail class of CIFAR-100-LT with imbalance ratio 10 has multiple samples, which makes it hard to distinguish the performance of methods on the long-tailed recognition. + +Table 1: Performance comparison on ImageNet-LT and iNaturalist 2018 datasets. Top-1 accuracy of ResNet-50 [26] is reported. The "Many", "Medium", "Few" and "All" denotes different groups. $\dagger$ denotes our reproduced results of PCL, SwAV, and BYOL based on their official code. Other baselines' results on ImageNet-LT and iNaturalist 2018 are copied from Li et al. [41]. + +
BackboneImageNet-LTiNaturalist 2018
MethodsManyMediumFewAllManyMediumFewAll
CE64.033.85.841.672.263.057.261.7
Focal loss [44]51.040.820.843.7---61.3
CB-Focal [15]-------61.1
LDAM-DRW [6]60.446.930.749.8---64.6
OLTR [46]35.832.321.532.259.064.164.963.9
τ-norm [34]56.644.227.446.771.168.969.369.3
cRT [34]58.844.026.147.373.268.866.168.2
LWS [34]57.145.229.347.771.069.868.869.5
PCL† [40]34.726.112.327.548.545.941.744.5
SwAV† [8]37.528.315.630.151.948.443.747.0
BYOL† [21]37.728.916.330.652.348.644.147.2
SCL [36]61.447.028.249.8---66.4
KCL [33]62.449.029.551.5---68.6
TSC [41]63.549.730.452.472.670.667.869.7
SBCL63.851.331.253.473.371.968.670.8
+ +Table 2: Performance comparison on CIFAR-100-LT. Top-1 accuracy of the ResNet-32 [26] under different imbalance ratios is reported. We also report the accuracy of our re-implemented important baselines $(\dagger)$ in same setting on CIFAR-100-LT. The columns of "Statistic (IR 100)" are the results of different disjoint subsets on CIFAR-100-LT with imbalance ratio being 100. + +
MethodImbalance RatioStatistic (IR 100)
1005010ManyMediumFew
CE38.343.955.765.237.19.1
CB-CE [15]38.644.657.1---
Focal Loss [44]38.444.355.865.338.48.1
CB-Focal [15]38.745.258.065.037.610.3
CE-DRW [6]41.445.358.1---
CE-DRS [6]41.645.558.1---
LDAM [6]39.645.056.9---
LDAM-DRW [6]42.046.658.761.541.720.2
M2m-ERM [37]42.9-58.2---
M2m-LDAM [37]43.5-57.6---
cRT [34]43.346.858.164.044.818.1
LWS [34]43.146.458.1---
SCL [36] †42.145.254.862.842.018.4
KCL [33]42.846.357.6---
KCL†42.846.457.563.442.519.2
TSC [41]43.847.459.0---
TSC†43.547.658.763.743.220.4
SBCL44.948.757.964.445.322.2
+ +# 4.3. Performance on Other Visual Tasks + +There is a recent trend of using the contrastive learning to pretrain a feature extractor for downstream visual tasks other than image classification [24]. We are curious + +Table 3: Object detection results on PASCAL VOC. + +
MethodImageNetImageNet-LT
AP50APAP75AP50APAP75
random init.60.233.833.160.233.833.1
CE81.353.759.276.548.551.0
CL [24]81.356.162.778.251.556.5
KCL [33]82.355.562.179.752.657.9
SBCL81.956.262.880.653.458.8
+ +about two questions: (1) when the pretraining dataset is class-imbalanced, how the downstream performance is affected? (2) In such case, can our SBCL improve the learned feature extractor over existing contrastive learning baselines? To answer these questions, we use the object detection task of PASCAL VOC dataset as the evaluation suite and use ImageNet/ImageNet-LT datasets as class-balanced/imbalanced pretraining datasets. Following [24, 33], we first pretrain a feature extractor on ImageNet/ImageNet-LT then further finetune it for the downstream object detection tasks using Faster R-CNN [51] with R50-C4 backbone. + +The experiment results are shown in Tables 3. From the results, we can see that pretraining on class-balanced data (ImageNet) leads to consistently better results than that on class-imbalanced dataset (ImageNet-LT) pretraining the model on the ImageNet and ImageNet-LT datasets by the SBCL can perform slightly better than other baselines. In addition, the proposed SBCL significantly outperforms baselines on class-imbalanced pretraining dataset, while achieve + +comparable performance on class-balanced ones. For the representation which trained on the full ImageNet dataset, the performance advantage is not obvious. In Appendix A.1, we show additional experimental results of object detection and instance segmentation on COCO [45] dataset and SBCL also outperforms other baseline methods. Thus, we conclude that the proposed SBCL is not only helpful for image classification, but also other visual tasks. + +# 4.4. Combining SBCL with SOTA methods + +Another line of research to address the long-tailed problem is the ensemble-based methods, such as MisLAS [72], WD [1], RIDE [62], PaCo [14], and BCL [74]. Here we show that SBCL can also be leveraged to boost the performance of these methods. + +MisLAS use label-aware smoothing and shifted batch normalization to improve the second stage (the classifier learning) on the long-tailed recognition, while our method is to improve the first stage of representation learning. Thus, MisLAS could be combined with SBCL and TSC. In Table 4, we report the results of combing MisLAS with both SBCL and TSC. The results show that both SBCL and TSC improve the performance of MisLAS and SBCL renders more performance boost than TSC. + +Weight decay (WD) [1] applies a penalty on the magnitudes of the weights in a model, which is particularly prominent for larger weights, resulting in the acquisition of more balanced parameters. WD and MaxNorm significantly enhances classification accuracy when applied to a classifier [1]. Table 4 presents the accuracy of the classifier trained by WD and MaxNorm, based on different representations learned by SBCL and TSC. The results demonstrate that SBCL improves the performance of the WD method [1] more effectively compared to TSC. + +To implement SBCL with RIDE which incorporates multiple models in a multi-expert framework, we follow [41] to simply replace the feature extractor on stage-1 training in RIDE with that trained with SBCL and keep the stage-2 routing training unchanged. As shown in Table 4, applying SBCL to RIDE improves its performance with a significant gap, outperforms the combination of TSC and RIDE on all different number of experts. + +PaCo and BCL propose new variants of supervised contrastive loss and jointly train both the proposed loss and classification loss to improve long-tail recognition, while we focus on the two stage pipeline, especially the first stage of representation learning. In this experiment, we show that using models pretrained with both TSC and SBCL as initialization could improve the performance of both PaCo and BCL. The results can be found in Table 4, and we can see that SBCL renders larger performance gain than TSC. The improvement over PaCo and BCL sheds lights on potential future work to evaluate the combination of multiple tech- + +niques of long-tail recognition to achieve new SOTA results. + +Table 4: Performance of the combination of SBCL and SOTA ensemble-based methods with ResNet-50 [26] on ImageNet-LT. $\dagger$ denotes the results of our re-implemented TSC. + +
MethodManyMediumFewAll
MiSLAS [72]---52.7
WD & Max [1]62.550.441.553.9
RIDE [62] (3 experts)66.251.734.954.9
PaCo [14]64.455.733.756.0
BCL [74]67.954.236.657.1
TSC+ MiSLAS63.750.536.053.6
TSC+WD & Max63.851.941.655.1
TSC+ RIDE (3 experts)69.151.736.756.3
TSC+ RIDE (3 experts)†68.651.436.055.9
TSC+ PaCo66.455.835.757.1
TSC+ BCL69.056.337.858.7
SBCL + MiSLAS64.152.036.454.5
SBCL +WD & Max65.352.642.756.1
SBCL + RIDE (3 experts)69.252.436.956.8
SBCL + PaCo66.956.138.457.9
SBCL + BCL70.056.938.559.5
+ +Table 5: Ablation study on different components of SBCL. + +
Warm-upAdaptive clusterDynamic temperatureCIFAR-100-LT
Imbalance Ratio1005010
✓✓44.047.957.2
✓✓43.847.256.5
✓✓43.847.857.0
✓✓✓44.948.757.9
+ +# 4.5. Ablation studies + +Warm-up. As mentioned in Section 3.3, we train the feature extractor for several epochs using ordinary SCL or KCL as warm-up stage. As shown in Table 5, such a warm-up stage is beneficial since the performance drops when we remove the warm-up stage. This is likely because at the early stage of training, the extracted feature is not well-trained and the cluster assignment could be noisy and ineffective, hindering the efficacy of SBCL. + +Adaptive clustering. We are also curious about the efficacy of adaptive clustering and thus present the performance of SBCL with clustering only once and fixing cluster assignments during training. As shown in Table 5, without adaptive clustering, the performance decreases in all cases. The reason could be fixed cluster assignment is prone to noise when the model is not well-trained, while adaptive clustering would dynamically adjust the cluster assignments based on the current learned model, which is supposed to become better as the training proceeds. + +![](images/76be709dc3c4b05881659dbcddda0a77b8fba9d70f2110aaa3dfcf52887147c3.jpg) +(a) Feature distance in subclasses. + +![](images/7a95efb41c4c36788cd109771490c8fdef84a519d462e13d307e2233431104ab.jpg) +(b) Feature distance in classes. +Figure 2: Feature distance of subclasses and classes on CIFAR-100-LT with imbalance ratio 100. We randomly sample instances from many-shot and medium-shot classes so that the size of each equals to that of few-shot classes. + +Dynamic temperature. In Table 5, we also study the effectiveness of dynamic temperature (Section 3.2). We remove the dynamic temperature and simply set $\tau_{2} > \tau_{1}$ following [27]. With the fixed temperature $\tau_{2}$ , the performance of SBCL is significantly worse than that with dynamic temperature. We speculate that this is because dynamic temperature could help prevent the instances of a class to form overly large or small cluster in the feature space and therefore lead to better learned representations. Additionally, to evaluate the impact of dynamic temperature on other baselines, we apply the dynamic temperature on TSC, as reported in Appendix A.1. In Appendix A.1, we also provide an additional analysis of other hyperparameters on SBCL. + +# 4.6. Illustrate the superiority of SBCL + +In this section, we present an example that explicitly demonstrates the superiority of SBCL compared to other baselines. We fix the features learned from previous approaches and SBCL on the CIFAR-100-LT with imbalance ratio 100, and use these features for classification task. Table 6 shows that SBCL achieves the best performance. + +Table 6: Accuracy(%) of SBCL and other baselines on CIFAR-100-LT with imbalanced ratio 100. + +
MethodCESCLKCLTSCSBCL
ACC(%)38.336.637.838.739.9
+ +# 4.7. Feature distribution of SBCL + +To analyze the representation learned by SBCL, we firstly define the euclidean distance between a given sample and other samples from the same/different classes as intra/inter-class distance. Concretely, the euclidean distance between a sample $z_{i}$ and a set $S$ is defined as $\mathbf{D}(z_i,S) = \frac{1}{|S|}\sum_{z_j\in S}\| z_i - z_j\| _2$ . Then, the intra- and interclass distance of sample $z_{i}$ can be defined as $\mathbf{D}(z_i,P_i)$ and $\mathbf{D}(z_i,\mathcal{D} / P_i)$ separately; and the intra- and inter-subclass + +distance of sample $z_{i}$ can be defined as $\mathbf{D}(z_i,M_i)$ and $\mathbf{D}(z_i,P_i / M_i)$ separately. + +Figure 2 shows the average distance between features learned by SBCL in subclasses and classes on the CIFAR-100-LT. As shown in Figure 2a, the distance between samples from the same subclass is less than those from the same class but different subclasses. Meanwhile, in Figure 2b, the inter-class distance consistently exceeds the intra-class distance, indicating clear separation between features from different classes. Additionally, the inter-class distance remains stable, which suggests that each class is uniformly distributed on a hypersphere. The results in all indicate that the two-layer class hierarchy is successfully captured and feature distribution achieves the core idea of SBCL. The specific values of feature distance are displayed in Appendix A.1. + +Figure 3 visualizes the feature distribution of subclasses in a head class and head classes in CIFAR-100-LT with imbalance ratio 100 trained by SBCL using t-SNE [59]. The representations learned by SBCL form some separated clusters, which be aligned with subclass labels and class labels. This suggests that SBCL can learn an embedding space where features from same subclasses/classes are pulled closely and features from different subclasses/class are separated. + +# 4.8. Variation of per-class weight during training + +Weight decay [1] achieves superior performance by effectively tuning weight decay during training. This study highlights the significance of regularizing network weights to become balanced for long-tailed recognition. Therefore, it is crucial to investigate whether SBCL can help the linear classifier to learn balanced weights through the training process. + +Figure 4 shows the per-class weight norm of a linear classifier trained on top of features learned by SBCL in different training stages on CIFAR-100-LT. From the figure, we can see that as the training proceeds, the per-class weight norm becomes balanced even when training the linear classifier, + +![](images/4c339c5d2ff0e13c098d1be84be93bb18c7192efa2eb41f976b78ebcf18a2f5c.jpg) +(a) Distribution of subclasses on a head class. + +![](images/fa7efb1b566fb25a5e16be51e8f0f10df6668a017219eef2636f4bec1f0a1ec2.jpg) +(b) Distribution of head classes on the dataset. +Figure 3: Feature distribution of subclasses and classes on CIFAR-100-LT with imbalance ratio 100. Color represents subclasses/classes. + +the original cross-entropy loss is used. + +![](images/3735514970833759f85269c51acb17992d9c8cf8be4b8d0fe523e44bc58578ff.jpg) +Figure 4: Change in weight norm of the linear classifier based on the representations trained by SBCL on CIFAR-100-LT with imbalance ratio 100 during training. + +# 5. Related work + +Traditional methods for handling long-tailed recognition problem includes re-sampling and re-weighting. There are roughly two types of re-sampling techniques: over-sampling the minority classes [4, 5, 54, 70] and under-sampling the frequent classes [4, 23, 32]. The re-weighting techniques assign adaptive weights for different classes or even different samples. The vanilla scheme re-weights classes proportionally to the inverse of their frequency [29, 30, 64]. For class-level re-weighting methods, many loss functions including CB loss [15], LDAM loss [6] and Balanced softmax loss [50] were recently proposed, while instance-level re-weighting methods include Focal loss [44] and Influence-balanced loss [49]. Recently, two-stage algorithms have achieved remarkable performance for long-tailed recognition, such as classifier re-training (cRT) [34], learnable weight scaling (LWS) [34], and Mixup Shifted Label-Aware Smoothing model (MiSLAS) [72]. Meanwhile, bilateral branch net + +work (BBN) [73] uses an additional network branch for rebalancing. RIDE [62] use multiple branches named experts, each learning to specialize in the entire classes. LADE [28] assumes the prior of test class distributions is available and accordingly post-adjust model predictions. PaCo [14] applies parametric class-wise learnable centers to rebalance in contrastive learning. BCL [74] proposes a multi-branch framework to achieve class-averaging and class-complement in the training process. + +To boost the performance of the two-stage algorithms, researchers have introduced supervised contrastive learning [36] to the first feature-learning stage and proposed $k$ -positive contrastive loss (KCL) [33] and targeted supervised contrastive learning (TSC) [41]. While achieving the state-of-the-art performance, these methods inject class-balance in the contrastive learning objective, inevitably leading to instance-imbalance during training. In this work, we instead propose to achieve both subclass- and instance-balance in the contrastive learning object. Our method is also related to recent studies of clustering-based deep unsupervised learning [7,8,18,43,65,66], especially those that leverage contrastive learning [22,40,42,63]. However, they target at general unsupervised representation learning scenario, while our method is tailored for long-tailed recognition where the training data is immensely class-imbalanced. + +# 6. Conclusion + +In this paper, we introduced Subclass-balancing Contrastive Learning (SBCL) for long-tailed recognition. It breaks down the head classes into multiple semantically-coherent subclasses via subclass-balancing adaptive clustering and incorporates a bi-granularity contrastive loss that encourages both subclass- and instance-balance. Extensive experiments on multiple datasets demonstrate that SBCL achieves state-of-the-art single-model performance on benchmark datasets for long-tailed recognition. + +# References + +[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6897-6907, 2022. 7, 8 +[2] Shin Ando and Chun Yuan Huang. Deep over-sampling framework for classifying imbalanced data. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 770-785. Springer, 2017. 1 +[3] David Arthur and Sergei Vassilvitskii. K-means++ the advantages of careful seeding. In Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms, pages 1027-1035, 2007. 3, 14 +[4] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106:249-259, 2018. 1, 9 +[5] Jonathon Byrd and Zachary Lipton. What is the effect of importance weighting in deep learning? In International Conference on Machine Learning, pages 872-881. PMLR, 2019. 1, 9 +[6] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in Neural Information Processing Systems, 32, 2019. 1, 5, 6, 9 +[7] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In European Conference on Computer Vision, 2018. 9 +[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 2020. 5, 6, 9, 18 +[9] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. Smote: synthetic minority over-sampling technique. Journal of artificial intelligence research, 16:321-357, 2002. 1 +[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5008-5017, 2021. +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607. PMLR, 2020. + +[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. +[13] Collobert, Ronan, Weston, and Jason. A unified architecture for natural language processing: deep neural networks with multitask learning. In International Conference on Machine Learning, 2008. 1 +[14] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 7, 9 +[15] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9268-9277, 2019. 1, 5, 6, 9 +[16] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015. +[17] Qi Dong, Shaogang Gong, and Xiatian Zhu. Imbalanced deep learning by minority class incremental rectification. IEEE transactions on pattern analysis and machine intelligence, 41(6):1367-1381, 2018. 1 +[18] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. Advances in Neural Information Processing Systems, 27, 2014. 9 +[19] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. +[20] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In Proceedings of the IEEE/cvf International Conference on computer vision, pages 6391-6400, 2019. +[21] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6, 18 +[22] Yuanfan Guo, Minghao Xu, Jiawen Li, Bingbing Ni, Xuanyu Zhu, Zhenbang Sun, and Yi Xu. Hcsc: Hierarchical contrastive selective coding. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2022. 9 + +[23] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9):1263-1284, 2009. 9 +[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 6, 16, 17, 18 +[25] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 16, 17 +[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6, 7 +[27] David T Hoffmann, Nadine Behrmann, Juergen Gall, Thomas Brox, and Mehdi Noroozi. Ranking info noise contrastive estimation: Boosting contrastive learning via ranked positives. the Association for the Advance of Artificial Intelligence, 2022. 4, 8 +[28] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, 2021. 2, 9 +[29] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5375-5384, 2016. 9 +[30] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Deep imbalanced learning for face recognition and attribute prediction. IEEE transactions on pattern analysis and machine intelligence, 42(11):2781-2794, 2019. 9 +[31] Ahmet Iscen, André Araujo, Boqing Gong, and Cordelia Schmid. Class-balanced distillation for long-tailed visual recognition. Proceedings of the IEEE international conference on computer vision, 2021. +[32] Nathalie Japkowicz and Shaju Stephen. The class imbalance problem: A systematic study. Intelligent data analysis, 6(5):429-449, 2002. 9 +[33] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 1, 2, 5, 6, 9, 16, 18 +[34] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis + +Kalantidis. Decoupling representation and classifier for long-tailed recognition. International Conference on Learning Representations, 2019. 2, 5, 6, 9, 16 +[35] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7482-7491, 2018. +[36] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 1, 2, 5, 6, 9 +[37] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13896-13905, 2020. 6 +[38] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[39] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. +[40] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 4, 5, 6, 9, 18 +[41] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6, 7, 9, 16, 17 +[42] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. AAAI Conference on Artificial Intelligence, 2021. 9 +[43] Renjie Liao, Alex Schwing, Richard Zemel, and Raquel Urtasun. Learning deep parsimonious representations. Advances in Neural Information Processing Systems, 29, 2016. 9 +[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2980-2988, 2017. 5, 6, 9 +[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7 + +[46] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 5, 6 +[47] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In European conference on computer vision, pages 181-196, 2018. 1 +[48] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. +[49] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 735-744, 2021. 9 +[50] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. Advances in Neural Information Processing Systems, 33:4175–4186, 2020. 9 +[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6, 17 +[52] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 5 +[53] Xie Saining, Girshick Ross, Dollar Piotr, Tu Zhuowen, and He Kaiming. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page 1492-1500, 2017. +[54] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In European conference on computer vision, pages 467-482. Springer, 2016. 1, 9 +[55] Tony Shen, Ariel Lee, Carol Shen, and C Jimmy Lin. The long tail and rare disease research: the impact of next-generation sequencing for rare mendelian disorders. Genetics research, 97, 2015. 1 +[56] Aditya Singh and Alessandro Bay. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in Neural Information Processing Systems, 2019. + +[57] Merrielle Spain and Pietro Perona. Measuring and predicting importance of objects in our visual world. 2007. 1 +[58] AFM Uddin, Mst Monira, Wheemyung Shin, Tae-Choong Chung, Sung-Ho Bae, et al. Saliency: A saliency guided data augmentation strategy for better regularization. International Conference on Learning Representations, 2021. +[59] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8 +[60] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 5 +[61] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2021. 4 +[62] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. International Conference on Learning Representations, 2021. 7, 9 +[63] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021. 5, 9 +[64] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. Advances in Neural Information Processing Systems, 30, 2017. 1, 9 +[65] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016. 9 +[66] Jianwei Yang, Devi Parikh, and Dhruv Batra. Joint unsupervised learning of deep representations and image clusters. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5147-5156, 2016. 9 +[67] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. Advances in Neural Information Processing Systems, 33:19290-19301, 2020. +[68] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. International Conference on Machine Learning, 2019. 1 + +[69] Han-Jia Ye, Hong-You Chen, De-Chuan Zhan, and Wei-Lun Chao. Identifying and compensating for feature deviation in imbalanced deep learning. arXiv preprint arXiv:2001.01385, 2020. +[70] Qiaoyong Zhong, Chao Li, Yingying Zhang, H Sun, S Yang, D Xie, and S Pu. Towards good practices for recognition & detection. In CVPR workshops, volume 1, page 2, 2016. 9 +[71] Yaoyao Zhong, Weihong Deng, Mei Wang, Jiani Hu, Jianteng Peng, Xunqiang Tao, and Yaohai Huang. Unequal-training for deep face recognition with long-tailed noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7812-7821, 2019. 1 +[72] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16489-16498, 2021. 7, 9 +[73] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, 2020. 9 +[74] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 7, 9 +[75] George Kingsley Zipf. The psycho-biology of language: An introduction to dynamic philology. Routledge, 2013. 1 \ No newline at end of file diff --git a/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/images.zip b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb1c16d4276cc31a3768b1eed3a6168f67fae3a9 --- /dev/null +++ b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7a25058742eca65e4d65d65e94fc66c7ff9ddcc37b3184e44769ea8e6bec1dd +size 525161 diff --git a/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/layout.json b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e770f27428b9d8cbfa7e4d7edb8e8d1f4063f657 --- /dev/null +++ b/2023/Subclass-balancing Contrastive Learning for Long-tailed Recognition/layout.json @@ -0,0 +1,10670 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 511, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 511, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 511, + 121 + ], + "type": "text", + "content": "Subclass-balancing Contrastive Learning for Long-tailed Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 142, + 487, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 142, + 487, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 487, + 185 + ], + "type": "text", + "content": "Chengkai Hou1, Jieyu Zhang2, Haonan Wang3, Tianyi Zhou4* \n1Jilin university, 2University of Washington, 3National University of Singapore \n4University of Maryland" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 187, + 502, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 187, + 502, + 213 + ], + "spans": [ + { + "bbox": [ + 99, + 187, + 502, + 213 + ], + "type": "text", + "content": "houck20@mails.jlu.edu.cn, jieyuz2@cs.washington.edu, haonan.wang@u.nus.edu.tianyi@umd.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 266, + 290, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 290, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 290, + 602 + ], + "type": "text", + "content": "Long-tailed recognition with imbalanced class distribution naturally emerges in practical machine learning applications. Existing methods such as data reweighing, resampling, and supervised contrastive learning enforce the class balance with a price of introducing imbalance between instances of head class and tail class, which may ignore the underlying rich semantic substructures of the former and exaggerate the biases in the latter. We overcome these drawbacks by a novel \"subclass-balancing contrastive learning (SBCL)\" approach that clusters each head class into multiple subclasses of similar sizes as the tail classes and enforce representations to capture the two-layer class hierarchy between the original classes and their subclasses. Since the clustering is conducted in the representation space and updated during the course of training, the subclass labels preserve the semantic substructures of head classes. Meanwhile, it does not overemphasize tail class samples, so each individual instance contribute to the representation learning equally. Hence, our method achieves both the instance- and subclass-balance, while the original class labels are also learned through contrastive learning among subclasses from different classes. We evaluate SBCL over a list of long-tailed benchmark datasets and it achieves the state-of-the-art performance. In addition, we present extensive analyses and ablation studies of SBCL to verify its advantages. Our code is available at https://github.com/JackHck/subclass-balancing-contrastive-learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 635, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 288, + 696 + ], + "type": "text", + "content": "In reality, the datasets often follow the Zipfian distribution over classes with a long tail [57, 75], i.e., a few classes (head classes) containing significantly more instances than the remaining tail classes. Such tail classes could be of great importance for high-stake applications, e.g., patient" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 242, + 547, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 547, + 301 + ], + "type": "text", + "content": "class in medical diagnosis or accident class in autonomous driving [6, 55]. However, training on such class-imbalanced datasets can result in a severely biased model with noticeable performance drop in classification tasks [2, 4, 13, 47, 64, 68, 71]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 548, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 548, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 548, + 483 + ], + "type": "text", + "content": "To overcome the challenges posed by long-tailed data, data resampling [2, 4, 9, 54] and loss reweighing [5, 6, 15, 17] have been widely applied but they cannot fully leverage all the head-class samples. Very recent work discovered that supervised contrastive learning (SCL) [36] can achieve state-of-the-art (SOTA) performance on benchmark datasets of long-tailed recognition [33, 41]. Specifically, the " + }, + { + "bbox": [ + 304, + 304, + 548, + 483 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 304, + 548, + 483 + ], + "type": "text", + "content": "-positive contrastive learning (KCL) [33] and its subsequent work targeted supervised contrastive learning (TSC) [41] revamp SCL by encouraging the learned feature space to be class-balanced and uniformly distributed. However, those methods enforcing class-balance often come with a price of instance-imbalance, i.e., each individual instance of tail classes would have much greater impact on model training than that of head classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 548, + 713 + ], + "type": "text", + "content": "Such instance-imbalance can result in significant degradation of the performance on long-tailed recognition for several reasons. On the one hand, the limited samples in each tail class might not be sufficiently representative of the whole class. So even a small bias of them can be enormously exaggerated by class-balancing methods and result in sub-optimal learning of classifiers or representations. On the other hand, head classes usually have more complicated semantic substructures, e.g., multiple high-density regions of the data distribution, so simply downweighing samples of head classes and treating them equally can easily lose critical structural information. For example, images of a head class \"cat\" might be highly diverse in breeds and colors, which need to be captured by different features but downweighing or subsampling them may easily lose such information, while a tail class \"platypus\" might only contain a few similar images that are unlikely to cover all the representative features. Therefore, it is non-trivial to enforce both class-balance and instance-balance simultaneously in the same method." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5395" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 323 + ], + "type": "text", + "content": "Can we remove the negative impact of class-imbalance while still retain the advantages of instance-balance? In this paper, we achieve both through subclass-balancing contrastive learning (SBCL), a novel supervised contrastive learning defined on subclasses, which are the clusters within each head class, have comparable size as tail classes, and are adaptively updated during the training. Instead of sacrificing instance-balance for class-balance, our method achieves both instance- and subclass-balance by exploring the head-class structure in the learned representation space of the model-in-training. In particular, we propose a bi-granularity contrastive loss that enforces a sample (1) to be closer to samples from the same subclass than all the other samples; and (2) to be closer to samples from a different subclass but the same class than samples from any other subclasses. While the former learns representations with balanced and compact subclasses, the latter preserves the class structure on subclass level by encouraging the same class's subclasses to be closer to each other than to any different class's subclasses. Hence, it can learn an accurate classifier distinguishing original classes while enjoy both the instance- and subclass-balance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 323, + 288, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 288, + 372 + ], + "type": "text", + "content": "In this paper, we apply SBCL for several visual recognition tasks to demonstrate SBCL superiority over other previous works (e.g., KCL [33], TSC [41]). To summarize, this paper makes the following contributions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 379, + 289, + 587 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 379, + 289, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 379, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 48, + 379, + 289, + 439 + ], + "type": "text", + "content": "(a). We provide a new design principal of leveraging supervised contrastive learning for long-tailed recognition, i.e., aiming at achieving both instance- and subclass-balance instead of class-balance at the expense of instance-balance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 447, + 288, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 288, + 508 + ], + "type": "text", + "content": "(b). We propose a novel instantiation of the aforementioned design principal, subclass-balancing contrastive learning (SBCL), which consists of two major components, namely, subclass-balancing adaptive clustering and bi-granularity contrastive loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 514, + 289, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 289, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 289, + 587 + ], + "type": "text", + "content": "(c). Empirically, we compare the SBCL against state-of-the-art methods on three visual tasks: image classification, object detection, and instance segmentation to demonstrate its effectiveness on handling class imbalance. We also conduct a series of experiments to analyze the efficacy of SBCL." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 597, + 198, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 198, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 198, + 611 + ], + "type": "text", + "content": "2. Background and notations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Long-tailed recognition. Long-tailed recognition aims to learn a classifier from a training dataset with long-tailed class distribution, i.e., a few classes contain many data (head classes) while most classes contain only a few data (tail classes), where the major challenge is to require model recognizing all classes equally well. Let " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{x_i, y_i\\}_{i \\in [n]}" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": " be a long-tailed training dataset, where " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": " denotes a sample and " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "y_i \\in [C]" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": " denotes its label. Denote by " + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k \\subseteq \\mathcal{D}" + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": " the set of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": "instances belonging to class " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "n_k = |\\mathcal{D}_k|" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": " is the number of samples in class " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": ". The total number of training samples over " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": " classes is " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "n = \\sum_{k=1}^{C} n_k" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": ". Without loss of generality, we follow prior work [28, 34] to assume that the classes are sorted by cardinality in decreasing order (i.e., if " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "i < j" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "n_i \\geq n_j" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "n_1 \\gg n_C" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": ". In addition, we define the imbalance ratio as " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "\\max_{k \\in [C]} (n_k) / \\min_{k \\in [C]} (n_k) = n_1 / n_C" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": ". Finally, let " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "f_\\theta(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": " be a deep feature extractor, e.g., a neural network, parameterized by " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_c" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": " is the linear classifier of class " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": ", then the classifier we aim to learn is " + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "inline_equation", + "content": "h(x_i) = \\arg \\max_{c \\in [C]} \\mathbf{w}_c^\\top f_\\theta(x_i)" + }, + { + "bbox": [ + 304, + 72, + 548, + 207 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 218, + 547, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 547, + 328 + ], + "type": "text", + "content": "Supervised contrastive learning. Recent studies have shown that supervised contrastive learning (SCL) [36] provides a strong performance gain for long-tailed recognition and its variants have achieved state-of-the-art (SOTA) performance [33, 41]. Specifically, SCL learns the feature extractor " + }, + { + "bbox": [ + 304, + 218, + 547, + 328 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 218, + 547, + 328 + ], + "type": "text", + "content": " via maximizing the discriminativeness of positive instances, i.e., instances from the same class, and the learning objective for a single training data " + }, + { + "bbox": [ + 304, + 218, + 547, + 328 + ], + "type": "inline_equation", + "content": "(x_i, y_i)" + }, + { + "bbox": [ + 304, + 218, + 547, + 328 + ], + "type": "text", + "content": " in a batch " + }, + { + "bbox": [ + 304, + 218, + 547, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{(x_i, y_i)\\}_{i \\in [N]}" + }, + { + "bbox": [ + 304, + 218, + 547, + 328 + ], + "type": "text", + "content": ", is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 337, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 545, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S C L} = \\sum_ {i = 1} ^ {N} - \\frac {1}{| \\tilde {P} _ {i} |} \\sum_ {z _ {p} \\in \\tilde {P} _ {i}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau\\right)}{\\sum_ {z _ {a} \\in \\tilde {V} _ {i}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau\\right)}, \\tag {1}", + "image_path": "ede2c53ba656f572f87692ef9ad4c748b02015a964f7cb6472f7356f9c402f26.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": " is the temperature hyperparameter, " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "z_{i} = f_{\\theta}(x_{i})" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": " is the feature generated from " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "V_{i} = \\{z_{i}\\}_{i\\in [N]}\\setminus \\{z_{i}\\}" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": " is the current batch of features except for " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "P_{i} = \\{z_{j}\\in V_{i}:y_{j} = y_{i}\\}" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": " is a set of instances with the same label as " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": ". Finally, let " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "\\tilde{z}_i" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": " be the feature of " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": ", the augmented version of " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": ", and for any set " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": " indexed by " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "\\tilde{S}_i = S_i\\cup \\{\\tilde{z}_i\\}" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "\\tilde{V}_i = V_i\\cup \\{\\tilde{z}_i\\}" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": ". However, for the long-tailed datasets, the feature spaces is dominated by head classes and thus have limited capability of semantic discrimination [33]. To address this, the " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": "-positive contrastive learning (KCL) [33] attempts to balance the feature space by keeping the number of positive instances in " + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "inline_equation", + "content": "\\tilde{P}_i" + }, + { + "bbox": [ + 304, + 378, + 547, + 535 + ], + "type": "text", + "content": " equal for each class, leading to the following loss" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 542, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 542, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 542, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K C L} = \\sum_ {i = 1} ^ {N} - \\frac {1}{k + 1} \\sum_ {z _ {p} \\in \\tilde {P _ {i} ^ {k}}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau\\right)}{\\sum_ {z _ {a} \\in \\tilde {V _ {i}}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau\\right)} (2)", + "image_path": "9119308ef2d35123b4b29de6e875b06806e8760617c67b0371a060fb53a6fdd3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "type": "inline_equation", + "content": "P_{i}^{k}" + }, + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "type": "text", + "content": " is a subset of " + }, + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "type": "text", + "content": " randomly drawn instances. Finally, the learned feature extractor " + }, + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 586, + 547, + 635 + ], + "type": "text", + "content": " is exploited in a sequel stage of training the classifier for long-tailed recognition [33, 41]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 645, + 389, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 389, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 389, + 659 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "As mentioned above, KCL and its sequels [33,41] balance the learning objective of SCL by picking the same number of positive instances for each class, i.e., " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "|P_i^k| = k" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " in Eq. 2 no matter which class " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " belongs to, however, we argue that" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5396" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 292, + 220 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 292, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 292, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 292, + 220 + ], + "type": "image", + "image_path": "f6d07b822ac00f11d4981192a1450f50fb0540bac0dc65db54ec7be3be90840d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 221, + 259, + 230 + ], + "lines": [ + { + "bbox": [ + 99, + 221, + 259, + 230 + ], + "spans": [ + { + "bbox": [ + 99, + 221, + 259, + 230 + ], + "type": "text", + "content": "(a). Subclass-balancing adaptive clustering (Section 3.1)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 72, + 533, + 220 + ], + "blocks": [ + { + "bbox": [ + 294, + 72, + 533, + 220 + ], + "lines": [ + { + "bbox": [ + 294, + 72, + 533, + 220 + ], + "spans": [ + { + "bbox": [ + 294, + 72, + 533, + 220 + ], + "type": "image", + "image_path": "2ec4e63648de9990919d8b2dbdd950dd170ee77df32b54e9bdf0b2be7a492cf6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 221, + 488, + 230 + ], + "lines": [ + { + "bbox": [ + 355, + 221, + 488, + 230 + ], + "spans": [ + { + "bbox": [ + 355, + 221, + 488, + 230 + ], + "type": "text", + "content": "(b). Bi-granularity contrastive loss (Section 3.2)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 238, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 545, + 275 + ], + "type": "text", + "content": "Figure 1: Illustration of subclass-balancing contrastive learning (SBCL). It initially divides the head classes into multiple subclasses of comparable size. Then, during training, SBCL builds each sample to be closer to samples from the same subclass than samples from different subclasses but the same class, which are also made to be closer than samples from different classes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "text", + "content": "such a class-balancing approach would inevitably introduce instance-imbalance: the instances of tail classes have much more chances to be engaged in the training than that of head class. Specifically, assume each class has no less than " + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "text", + "content": " instances, then the probability of an instance of class " + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "text", + "content": " being selected as positive instance is " + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "inline_equation", + "content": "p(c) = \\frac{k}{n_c}" + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "text", + "content": "; if the tail class " + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "inline_equation", + "content": "n_C" + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "text", + "content": " has only " + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "text", + "content": " instances and the imbalance ratio is " + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\frac{n_1}{n_C} = 100" + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "text", + "content": ", then we have " + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "inline_equation", + "content": "p(1) = \\frac{k}{100k} = 0.01" + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "inline_equation", + "content": "p(C) = 1" + }, + { + "bbox": [ + 46, + 296, + 287, + 477 + ], + "type": "text", + "content": ". We can see that when the instances of head class are selected once, that of tail class may already be trained 100 times. Thus, the training is immensely biased towards the few samples in each tail class. Besides, as tail classes only have very few instances that are not necessarily representative, the learned feature space might be unsatisfactory and sensitive to the training data of tail classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 479, + 288, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 658 + ], + "type": "text", + "content": "Here, we provide a new prospective of handling class-imbalance issue by contrastive learning: instead of aiming at class-balance at the expense of instance-imbalance, we propose to achieve both instance- and subclass-balance. We argue that head classes typically contain more diverse instances and thus have richer semantics in the training dataset. Therefore, it might be wise to break down the head classes into multiple semantically coherent subclasses, each of which consists of similar number of instance as tail classes. Built on this spirit, we develop subclass-balancing contrastive learning (SBCL), a new contrastive learning framework for long-tailed recognition (visualized in Figure 1) that achieves both instance- and subclass-balance. We present a concrete example that effectively demonstrates the superiority of SBCL compared to previous methods in Section 4.6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 251, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 251, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 251, + 683 + ], + "type": "text", + "content": "3.1. Subclass-balancing adaptive clustering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "We break down the class into several \"subclasses\" to attack the imbalanced phenomenon. Particularly, given a class" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "text", + "content": " and the associated set of data " + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "text", + "content": ", we employ a clustering algorithm of choice based on the features extracted by current feature extractor " + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "text", + "content": " to divide " + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "inline_equation", + "content": "m_c" + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "text", + "content": " subclasses/clusters. We use " + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\Gamma_c(x_i)" + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "text", + "content": " to denote the cluster label of an instance " + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "text", + "content": ". To ensure that the number of samples for each subclass is roughly the same, we propose a new cluster algorithm to divide the unit-length feature vectors, i.e., the features output by " + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 296, + 547, + 415 + ], + "type": "text", + "content": " with additional unit-length normalization. The new proposed cluster algorithm is described in Algorithm 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 545, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 487 + ], + "type": "text", + "content": "In Algorithm 1, during the initialization of cluster centers, our approach gradually selects cluster centers by prioritizing the samples that are maximally distant from other existing cluster centers [3]. In the process of assigning vectors to their centers, we set a threshold " + }, + { + "bbox": [ + 304, + 415, + 545, + 487 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 415, + 545, + 487 + ], + "type": "text", + "content": " as an upper limit of sample size in a cluster, which guarantees clusters of balanced sizes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 487, + 447, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 487, + 447, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 487, + 447, + 498 + ], + "type": "text", + "content": "Specifically, the threshold " + }, + { + "bbox": [ + 317, + 487, + 447, + 498 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 317, + 487, + 447, + 498 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 393, + 505, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 505, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 393, + 505, + 545, + 517 + ], + "type": "interline_equation", + "content": "M = \\max \\left(n _ {C}, \\delta\\right) \\tag {3}", + "image_path": "e2215f09d123865c2115fcc5cf365b0ef66f15d5aeef0aecf31a5def1614c1da.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "type": "inline_equation", + "content": "n_C" + }, + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "type": "text", + "content": " is the size of tail class and the hyperparameter " + }, + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "type": "text", + "content": " controls the lower bound of sample size in clusters to prevent overly-small cluster. Note that we only apply clustering algorithm to classes which contain multiple instances while the tail classes remain unchanged. As a consequence, the size of each resultant cluster is similar to that of tail class " + }, + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "type": "inline_equation", + "content": "n_C" + }, + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "type": "text", + "content": ". In addition, instead of only clustering once at the beginning, we update the cluster assignment adaptively based on the current feature extractor " + }, + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 304, + 521, + 545, + 688 + ], + "type": "text", + "content": " during training process and empirically show that such adaptive clustering outperforms only-once clustering in Section 4.5. We add an exhaustive description of Algorithm 1 and show the distribution of the sample size in clusters on the benchmark dataset in Appendix A.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Then, by replacing the class labels of head classes used in SCL/KCL with the finer-grained cluster labels, we ensure the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5397" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 86, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 264, + 84 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 264, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 264, + 84 + ], + "type": "text", + "content": "Algorithm 1 Subclass-balancing Adaptive Clustering" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": "Require: Sample set " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "S = \\{x_{i}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " ; A threshold " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " ; The number of iterations " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " \nEnsure: Cluster assignments for samples in " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " do if " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " then Choose the cluster centers " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " which are farther away from previously selected centers. else Update the cluster centers " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "y_{j} = \\frac{1}{n_{j}}\\sum_{i = 1}^{n_{j}}x_{i}." + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\triangleright n_j" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " is the number of samples in a cluster end if Construct the cluster center set " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{y_j\\}_{j = 1}^m" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\triangleright m" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " is the number of cluster centers while " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "S\\neq \\phi" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " Assign samples to centers " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " Select the most similar pair " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "(x_{i},y_{j}) = \\arg \\max \\mathrm{cosine - similarity}(x,y)" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "x\\in S,y\\in C" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " Assign the sample " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " to the center " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " . Delete the assigned sample " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " from the sample set " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "S = S / \\{x_i\\}" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "n_j\\geq M" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " then Sample number in a cluster exceeds the threshold " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " Delete the cluster center " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " from the cluster center set " + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\mathcal{C} / \\{y_j\\}" + }, + { + "bbox": [ + 47, + 86, + 545, + 293 + ], + "type": "text", + "content": " end if end while end for" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 315, + 288, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 398 + ], + "type": "text", + "content": "instance-balanced, i.e., each instance has similar chance of being selected regardless of its class. By breaking down the head classes, which typically contain more diverse instances, into multiple semantically coherent subclasses, we achieve subclass-balanced (instead of class-balanced) while maintain the rich semantics rendered by head classes in training dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 211, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 211, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 211, + 419 + ], + "type": "text", + "content": "3.2. Bi-granularity contrastive loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 546 + ], + "type": "text", + "content": "We now have two types of label for instances in head classes from different granularities: the coarse-grained class label and the fine-grained cluster label. A direct consequence of replacing class label in SCL/KCL with cluster label is that we no longer distinguish instances from different head classes, and therefore the boundaries between classes might be blurry, leading to sub-optimal feature space. As a remedy, we combine the contrastive loss of both class label and cluster label into the following one and reuse the notations of Eq. 1:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 565, + 287, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 565, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 52, + 565, + 287, + 640 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {S B C L} = - \\sum_ {i = 1} ^ {N} \\left(\\frac {1}{| \\tilde {M} _ {i} |} \\sum_ {z _ {p} \\in \\tilde {M} _ {i}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau_ {1}\\right)}{\\sum z _ {a} \\in \\tilde {V} _ {i}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau_ {1}\\right) \\right. \\\\ + \\beta \\frac {1}{\\left| \\tilde {P} _ {i} \\right| - \\left| M _ {i} \\right|} \\sum_ {z _ {p} \\in \\tilde {P} _ {i} / M _ {i}} \\log \\frac {\\exp \\left(z _ {i} \\cdot z _ {p} ^ {\\top} / \\tau_ {2}\\right)}{\\sum_ {z _ {a} \\in \\tilde {V} _ {i} / M _ {i}} \\exp \\left(z _ {i} \\cdot z _ {a} ^ {\\top} / \\tau_ {2}\\right)}) \\tag {4} \\\\ \\end{array}", + "image_path": "c2c77650f981fb11b4dab32afcf7f3c310909213e5f694bd5037027f6bef572c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "inline_equation", + "content": "M_{i} = \\{z_{j}\\in P_{i}:\\Gamma_{y_{i}}(x_{i}) = \\Gamma_{y_{i}}(x_{j})\\}" + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": " is a set of instances with the same cluster label as " + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": " is a hyperparameter that balances these two loss terms. The first term corresponds to the SCL loss with cluster labels, while the second term leverages the class label but does not consider the instances of the same cluster, i.e., the instances in " + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": " are" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 315, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 361 + ], + "type": "text", + "content": "removed in the second term. Such a design choice reflects the two types of positive instances for " + }, + { + "bbox": [ + 304, + 315, + 545, + 361 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 315, + 545, + 361 + ], + "type": "text", + "content": ": (1) the instance in the same cluster and (2) the instance of the same class but in different clusters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "text", + "content": "According to previous studies [27, 40, 61], the temperature " + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "text", + "content": " in contrastive loss is critical in controlling the local separation and global uniformity of the feature distribution. Specifically, for supervised contrastive learning, a low temperature makes relative high penalty on feature distribution, that actually encourages the features distribute more concentrically. As the temperature increases, the relative penalty tends to be more uniform which uniformizes the distribution of the features. Although the above objective explicitly considers the two types of label from different granularities, it still treats class and cluster label similarly. Intuitively, we expect instances of the same subclass to form a more concentrated cluster in feature space than those of the same class, since subclass naturally indicates finer-grained semantic coherence. To achieve this, we ensure the temperature " + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\tau_{2} > \\tau_{1}" + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "text", + "content": " and dynamically adjust " + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "text", + "content": " for each class according to its current level of concentration of the instances' feature. Following [40], for class " + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "text", + "content": " we define " + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\phi(c)" + }, + { + "bbox": [ + 304, + 365, + 547, + 581 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 376, + 593, + 545, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 593, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 376, + 593, + 545, + 617 + ], + "type": "interline_equation", + "content": "\\phi (c) = \\frac {\\sum_ {i = 1} ^ {n _ {c}} \\| z _ {i} - t _ {c} \\| _ {2}}{n _ {c} \\log \\left(n _ {c} + \\alpha\\right)} \\tag {5}", + "image_path": "f0376e99c2bd38dcb52ee5f2881afd80799ea4e4679e721041de2c8f688224e3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " is the centroid for the class " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " is a hyperparameter to ensure that " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi(c)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " is not overly-large, and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " corresponds to instances of class " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". From the formulation, we can see that if the current averaged distance to the class centroid is large or the class contains fewer data, thus the temperature will be set large to adopt the feature distribution of class " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " during the training process. Then we define the temperature of class" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5398" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 75, + 66, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 75, + 66, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 75, + 66, + 83 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 48, + 75, + 66, + 83 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 101, + 81, + 287, + 111 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 81, + 287, + 111 + ], + "spans": [ + { + "bbox": [ + 101, + 81, + 287, + 111 + ], + "type": "interline_equation", + "content": "\\tau_ {2} (c) = \\tau_ {1} \\cdot \\exp \\left(\\frac {\\phi (c)}{\\frac {1}{C} \\sum_ {i = 1} ^ {C} \\phi (i)}\\right) \\tag {6}", + "image_path": "ee68817676bf47e62ec396a1d58c055b4f173a843d89c6d7c6dee126928d5ce1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "text", + "content": "such that " + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "inline_equation", + "content": "\\tau_{2}(c)" + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "text", + "content": " for class label is always larger than " + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "text", + "content": " for cluster label (since " + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "inline_equation", + "content": "\\phi(c) > 0" + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "text", + "content": ") and could reflect the current level of concentration of the instances in a class. In particular, the proposed " + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "inline_equation", + "content": "\\tau_{2}(c)" + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "text", + "content": " encourages the features of instances in class " + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "text", + "content": " to form a less tight cluster than that of a subclass (by " + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "inline_equation", + "content": "\\tau_{2}(c) > \\tau_{1}" + }, + { + "bbox": [ + 46, + 113, + 288, + 198 + ], + "type": "text", + "content": ") while adaptively adjust the temperature to prevent an overly-loose/dense cluster." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 204, + 160, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 160, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 160, + 217 + ], + "type": "text", + "content": "3.3. Training algorithm" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "text", + "content": "Here, we describe the overall training process of subclassbalancing contrastive learning and the algorithm can be found in Algorithm 2. First, the adaptive clustering (Section 3.1) could be noisy at the early stage of training [40,63]. Thus, we warm-up the feature extractor " + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "text", + "content": " by a few epochs of training on ordinary SCL or KCL loss. In addition, our algorithm involves two adaptively-adjusting parts, namely, the cluster assignment and the temperature " + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "inline_equation", + "content": "\\tau_{2}(c)" + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "text", + "content": " for each head class " + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "text", + "content": ". Instead of updating these every epoch, we use a hyperparameter " + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "text", + "content": " as the update interval, i.e., we update the cluster assignment and the temperature based on the current learned " + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "text", + "content": " every " + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 222, + 289, + 368 + ], + "type": "text", + "content": " epoch." + } + ] + } + ], + "index": 4 + }, + { + "type": "code", + "bbox": [ + 47, + 392, + 288, + 568 + ], + "blocks": [ + { + "bbox": [ + 48, + 377, + 182, + 389 + ], + "lines": [ + { + "bbox": [ + 48, + 377, + 182, + 389 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 182, + 389 + ], + "type": "text", + "content": "Algorithm 2 Training Algorithm" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "lines": [ + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": "Require: Dataset " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{x_i,y_i\\}_{i\\in [n]};" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " ; The update interval of cluster assignment " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " ; The number of warm-up epoch " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " ; The total number of epoch " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " ; The hyperparameters " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " \nEnsure: A trained feature extractor " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " \n1: Initialize the model parameters " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " \n2: Train " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " with SCL/KCL for " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " epochs Warm-up stage \n3: for " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "t = T_0" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " do \n4: if " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "t\\% K = = 0" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " then Update cluster and temperature \n5: Update the cluster assignment based on the current feature extractor " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x)" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " \n6: Update the temperture " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " for each head class using Eq. 5 and Eq. 6 \n7: end if \n8: Train " + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 47, + 392, + 288, + 568 + ], + "type": "text", + "content": " using Eq. 4 Subclass-balancing contrastive learning \n9: end for" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_body" + } + ], + "index": 6, + "sub_type": "algorithm" + }, + { + "bbox": [ + 47, + 591, + 123, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 123, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 123, + 605 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 611, + 162, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 162, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 162, + 624 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "Datasets. We consider three commonly used long-tailed recognition benchmark datasets: CIFAR-100-LT [6], ImageNet-LT [46], and iNaturalist 2018 [60]. The CIFAR-100-LT and ImageNet-LT datasets are artificially generated long-tailed datasets from the class-balanced datasets [38,52], and the iNaturalist 2018 dataset is a large-scale real-world dataset that exhibits long-tailed imbalance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": "Baselines. We consider baseline methods of the following three categories: (1) class-balancing classifiers, including " + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": "-norm, LWS and cRT [34], which fixes the representation which trained by cross-entropy loss and trains the classifier with class-balanced sampling; (2) one-stage balancing loss, including CB loss [15], Focal loss [44], and LDAM loss [6]. These supervised distribution-aware loss makes the model to pay more attention on the minority class during training. (3) contrastive learning methods, including SCL [36], KCL [33], SwAV [8], PCL [40] and TSC [41] which train a feature extractor with the contrastive loss and then learn a classifier given the trained feature extractor." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 227, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 547, + 456 + ], + "type": "text", + "content": "Evaluation protocol. Following [33, 34, 41], we implement SBCL, as well as other contrastive learning methods, in a two-stage framework. In the first stage, we train the feature extractor with a contrastive learning method, while in the second stage, we train a linear classifier on top of the learned representation. Specifically, for CIFAR-100-LT dataset, the linear classifier is trained with LDAM loss and class re-weighting [6]. For ImageNet-LT and iNaturalist 2018 datasets, the linear classifier is trained with CE loss and class-balanced sampling [34]. All results are averaged over 5 trials with different random seeds. We mainly report the overall top-1 accuracy. For the two large datasets, ImageNet-LT and iNaturalist 2018 datasets, following the previous work [46], we also report the accuracy of three disjoint subsets: Many-shot classes (classes with more than 100 samples), Medium-shot classes (classes with 20 to 100 samples), and Few-shot classes (classes under 20 samples). We show the implementation details and important hyperparameters in the Appendix A.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 467, + 388, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 388, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 388, + 478 + ], + "type": "text", + "content": "4.2. Main results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": "The results on ImageNet-LT and iNaturalist 2018 are in Table 1. We can see that SBCL outperforms the baselines with a large margin over the two datasets. In addition, on iNaturalist 2018 dataset, SBCL outperforms the previous SOTA method by " + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": " on Many, " + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": " on Medium, " + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": " on Few and " + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": " on All, which shows the effectiveness of the proposed method in solving real-world long-tailed recognition problems such as natural species classification. Besides, SBCL is also better than existing contrastive learning method like KCL and TSC for all class splits, which demonstrates the effectiveness of the design principal of pursuing both instances- and subclass-balance in contrastive learning. Table 2 summarizes the results on CIFAR-100-LT dataset. For CIFAR-100-LT dataset, SBCL outperforms previous SOTA methods except for imbalance ratio 10. We hypothesize that it is because the tail class of CIFAR-100-LT with imbalance ratio 10 has multiple samples, which makes it hard to distinguish the performance of methods on the long-tailed recognition." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5399" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 123, + 128, + 471, + 328 + ], + "blocks": [ + { + "bbox": [ + 45, + 70, + 545, + 119 + ], + "lines": [ + { + "bbox": [ + 45, + 70, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 45, + 70, + 545, + 119 + ], + "type": "text", + "content": "Table 1: Performance comparison on ImageNet-LT and iNaturalist 2018 datasets. Top-1 accuracy of ResNet-50 [26] is reported. The \"Many\", \"Medium\", \"Few\" and \"All\" denotes different groups. " + }, + { + "bbox": [ + 45, + 70, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 45, + 70, + 545, + 119 + ], + "type": "text", + "content": " denotes our reproduced results of PCL, SwAV, and BYOL based on their official code. Other baselines' results on ImageNet-LT and iNaturalist 2018 are copied from Li et al. [41]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 123, + 128, + 471, + 328 + ], + "lines": [ + { + "bbox": [ + 123, + 128, + 471, + 328 + ], + "spans": [ + { + "bbox": [ + 123, + 128, + 471, + 328 + ], + "type": "table", + "html": "
BackboneImageNet-LTiNaturalist 2018
MethodsManyMediumFewAllManyMediumFewAll
CE64.033.85.841.672.263.057.261.7
Focal loss [44]51.040.820.843.7---61.3
CB-Focal [15]-------61.1
LDAM-DRW [6]60.446.930.749.8---64.6
OLTR [46]35.832.321.532.259.064.164.963.9
τ-norm [34]56.644.227.446.771.168.969.369.3
cRT [34]58.844.026.147.373.268.866.168.2
LWS [34]57.145.229.347.771.069.868.869.5
PCL† [40]34.726.112.327.548.545.941.744.5
SwAV† [8]37.528.315.630.151.948.443.747.0
BYOL† [21]37.728.916.330.652.348.644.147.2
SCL [36]61.447.028.249.8---66.4
KCL [33]62.449.029.551.5---68.6
TSC [41]63.549.730.452.472.670.667.869.7
SBCL63.851.331.253.473.371.968.670.8
", + "image_path": "c3cc4aabe92649c66e05e57e92305790920a471e1970a61156cda71ccc903b6d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 441, + 287, + 642 + ], + "blocks": [ + { + "bbox": [ + 46, + 346, + 288, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 288, + 431 + ], + "type": "text", + "content": "Table 2: Performance comparison on CIFAR-100-LT. Top-1 accuracy of the ResNet-32 [26] under different imbalance ratios is reported. We also report the accuracy of our re-implemented important baselines " + }, + { + "bbox": [ + 46, + 346, + 288, + 431 + ], + "type": "inline_equation", + "content": "(\\dagger)" + }, + { + "bbox": [ + 46, + 346, + 288, + 431 + ], + "type": "text", + "content": " in same setting on CIFAR-100-LT. The columns of \"Statistic (IR 100)\" are the results of different disjoint subsets on CIFAR-100-LT with imbalance ratio being 100." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 441, + 287, + 642 + ], + "lines": [ + { + "bbox": [ + 50, + 441, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 50, + 441, + 287, + 642 + ], + "type": "table", + "html": "
MethodImbalance RatioStatistic (IR 100)
1005010ManyMediumFew
CE38.343.955.765.237.19.1
CB-CE [15]38.644.657.1---
Focal Loss [44]38.444.355.865.338.48.1
CB-Focal [15]38.745.258.065.037.610.3
CE-DRW [6]41.445.358.1---
CE-DRS [6]41.645.558.1---
LDAM [6]39.645.056.9---
LDAM-DRW [6]42.046.658.761.541.720.2
M2m-ERM [37]42.9-58.2---
M2m-LDAM [37]43.5-57.6---
cRT [34]43.346.858.164.044.818.1
LWS [34]43.146.458.1---
SCL [36] †42.145.254.862.842.018.4
KCL [33]42.846.357.6---
KCL†42.846.457.563.442.519.2
TSC [41]43.847.459.0---
TSC†43.547.658.763.743.220.4
SBCL44.948.757.964.445.322.2
", + "image_path": "ec996901eb0c11ab9b49fc5689dfb4d170692086748251ac9eae684ae1b1f68a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 657, + 237, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 237, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 237, + 669 + ], + "type": "text", + "content": "4.3. Performance on Other Visual Tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "There is a recent trend of using the contrastive learning to pretrain a feature extractor for downstream visual tasks other than image classification [24]. We are curious" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 369, + 545, + 453 + ], + "blocks": [ + { + "bbox": [ + 320, + 347, + 530, + 359 + ], + "lines": [ + { + "bbox": [ + 320, + 347, + 530, + 359 + ], + "spans": [ + { + "bbox": [ + 320, + 347, + 530, + 359 + ], + "type": "text", + "content": "Table 3: Object detection results on PASCAL VOC." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 369, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 307, + 369, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 369, + 545, + 453 + ], + "type": "table", + "html": "
MethodImageNetImageNet-LT
AP50APAP75AP50APAP75
random init.60.233.833.160.233.833.1
CE81.353.759.276.548.551.0
CL [24]81.356.162.778.251.556.5
KCL [33]82.355.562.179.752.657.9
SBCL81.956.262.880.653.458.8
", + "image_path": "1f69d80d941548bec3e10ab4f538d0ae0ba133e24e3e455b39a286f98d9df5a5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 482, + 547, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 547, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 547, + 614 + ], + "type": "text", + "content": "about two questions: (1) when the pretraining dataset is class-imbalanced, how the downstream performance is affected? (2) In such case, can our SBCL improve the learned feature extractor over existing contrastive learning baselines? To answer these questions, we use the object detection task of PASCAL VOC dataset as the evaluation suite and use ImageNet/ImageNet-LT datasets as class-balanced/imbalanced pretraining datasets. Following [24, 33], we first pretrain a feature extractor on ImageNet/ImageNet-LT then further finetune it for the downstream object detection tasks using Faster R-CNN [51] with R50-C4 backbone." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 714 + ], + "type": "text", + "content": "The experiment results are shown in Tables 3. From the results, we can see that pretraining on class-balanced data (ImageNet) leads to consistently better results than that on class-imbalanced dataset (ImageNet-LT) pretraining the model on the ImageNet and ImageNet-LT datasets by the SBCL can perform slightly better than other baselines. In addition, the proposed SBCL significantly outperforms baselines on class-imbalanced pretraining dataset, while achieve" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5400" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "comparable performance on class-balanced ones. For the representation which trained on the full ImageNet dataset, the performance advantage is not obvious. In Appendix A.1, we show additional experimental results of object detection and instance segmentation on COCO [45] dataset and SBCL also outperforms other baseline methods. Thus, we conclude that the proposed SBCL is not only helpful for image classification, but also other visual tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 177, + 251, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 251, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 251, + 190 + ], + "type": "text", + "content": "4.4. Combining SBCL with SOTA methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 197, + 288, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 197, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 288, + 255 + ], + "type": "text", + "content": "Another line of research to address the long-tailed problem is the ensemble-based methods, such as MisLAS [72], WD [1], RIDE [62], PaCo [14], and BCL [74]. Here we show that SBCL can also be leveraged to boost the performance of these methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 257, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 364 + ], + "type": "text", + "content": "MisLAS use label-aware smoothing and shifted batch normalization to improve the second stage (the classifier learning) on the long-tailed recognition, while our method is to improve the first stage of representation learning. Thus, MisLAS could be combined with SBCL and TSC. In Table 4, we report the results of combing MisLAS with both SBCL and TSC. The results show that both SBCL and TSC improve the performance of MisLAS and SBCL renders more performance boost than TSC." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 288, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 288, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 288, + 484 + ], + "type": "text", + "content": "Weight decay (WD) [1] applies a penalty on the magnitudes of the weights in a model, which is particularly prominent for larger weights, resulting in the acquisition of more balanced parameters. WD and MaxNorm significantly enhances classification accuracy when applied to a classifier [1]. Table 4 presents the accuracy of the classifier trained by WD and MaxNorm, based on different representations learned by SBCL and TSC. The results demonstrate that SBCL improves the performance of the WD method [1] more effectively compared to TSC." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 485, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 288, + 581 + ], + "type": "text", + "content": "To implement SBCL with RIDE which incorporates multiple models in a multi-expert framework, we follow [41] to simply replace the feature extractor on stage-1 training in RIDE with that trained with SBCL and keep the stage-2 routing training unchanged. As shown in Table 4, applying SBCL to RIDE improves its performance with a significant gap, outperforms the combination of TSC and RIDE on all different number of experts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "PaCo and BCL propose new variants of supervised contrastive loss and jointly train both the proposed loss and classification loss to improve long-tail recognition, while we focus on the two stage pipeline, especially the first stage of representation learning. In this experiment, we show that using models pretrained with both TSC and SBCL as initialization could improve the performance of both PaCo and BCL. The results can be found in Table 4, and we can see that SBCL renders larger performance gain than TSC. The improvement over PaCo and BCL sheds lights on potential future work to evaluate the combination of multiple tech-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 547, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 84 + ], + "type": "text", + "content": "niques of long-tail recognition to achieve new SOTA results." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 142, + 547, + 338 + ], + "blocks": [ + { + "bbox": [ + 305, + 96, + 547, + 133 + ], + "lines": [ + { + "bbox": [ + 305, + 96, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 547, + 133 + ], + "type": "text", + "content": "Table 4: Performance of the combination of SBCL and SOTA ensemble-based methods with ResNet-50 [26] on ImageNet-LT. " + }, + { + "bbox": [ + 305, + 96, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 96, + 547, + 133 + ], + "type": "text", + "content": " denotes the results of our re-implemented TSC." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 142, + 547, + 338 + ], + "lines": [ + { + "bbox": [ + 307, + 142, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 547, + 338 + ], + "type": "table", + "html": "
MethodManyMediumFewAll
MiSLAS [72]---52.7
WD & Max [1]62.550.441.553.9
RIDE [62] (3 experts)66.251.734.954.9
PaCo [14]64.455.733.756.0
BCL [74]67.954.236.657.1
TSC+ MiSLAS63.750.536.053.6
TSC+WD & Max63.851.941.655.1
TSC+ RIDE (3 experts)69.151.736.756.3
TSC+ RIDE (3 experts)†68.651.436.055.9
TSC+ PaCo66.455.835.757.1
TSC+ BCL69.056.337.858.7
SBCL + MiSLAS64.152.036.454.5
SBCL +WD & Max65.352.642.756.1
SBCL + RIDE (3 experts)69.252.436.956.8
SBCL + PaCo66.956.138.457.9
SBCL + BCL70.056.938.559.5
", + "image_path": "4dc2de779ed4080e13cd29140b35b4b4f4f3dab1c6598e0f6d0803ef32efc108.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 384, + 545, + 446 + ], + "blocks": [ + { + "bbox": [ + 306, + 363, + 544, + 375 + ], + "lines": [ + { + "bbox": [ + 306, + 363, + 544, + 375 + ], + "spans": [ + { + "bbox": [ + 306, + 363, + 544, + 375 + ], + "type": "text", + "content": "Table 5: Ablation study on different components of SBCL." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 384, + 545, + 446 + ], + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 446 + ], + "type": "table", + "html": "
Warm-upAdaptive clusterDynamic temperatureCIFAR-100-LT
Imbalance Ratio1005010
✓✓44.047.957.2
✓✓43.847.256.5
✓✓43.847.857.0
✓✓✓44.948.757.9
", + "image_path": "d087ae8fff96cf3d6f3f1d9ca4aadd937266713e6e398067dee12ac32769e651.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 460, + 404, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 404, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 404, + 471 + ], + "type": "text", + "content": "4.5. Ablation studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 479, + 547, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 547, + 575 + ], + "type": "text", + "content": "Warm-up. As mentioned in Section 3.3, we train the feature extractor for several epochs using ordinary SCL or KCL as warm-up stage. As shown in Table 5, such a warm-up stage is beneficial since the performance drops when we remove the warm-up stage. This is likely because at the early stage of training, the extracted feature is not well-trained and the cluster assignment could be noisy and ineffective, hindering the efficacy of SBCL." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Adaptive clustering. We are also curious about the efficacy of adaptive clustering and thus present the performance of SBCL with clustering only once and fixing cluster assignments during training. As shown in Table 5, without adaptive clustering, the performance decreases in all cases. The reason could be fixed cluster assignment is prone to noise when the model is not well-trained, while adaptive clustering would dynamically adjust the cluster assignments based on the current learned model, which is supposed to become better as the training proceeds." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5401" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 72, + 272, + 190 + ], + "blocks": [ + { + "bbox": [ + 97, + 72, + 272, + 190 + ], + "lines": [ + { + "bbox": [ + 97, + 72, + 272, + 190 + ], + "spans": [ + { + "bbox": [ + 97, + 72, + 272, + 190 + ], + "type": "image", + "image_path": "76be709dc3c4b05881659dbcddda0a77b8fba9d70f2110aaa3dfcf52887147c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 198, + 246, + 207 + ], + "lines": [ + { + "bbox": [ + 123, + 198, + 246, + 207 + ], + "spans": [ + { + "bbox": [ + 123, + 198, + 246, + 207 + ], + "type": "text", + "content": "(a) Feature distance in subclasses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 496, + 190 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 496, + 190 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 496, + 190 + ], + "type": "image", + "image_path": "7a95efb41c4c36788cd109771490c8fdef84a519d462e13d307e2233431104ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 198, + 465, + 207 + ], + "lines": [ + { + "bbox": [ + 353, + 198, + 465, + 207 + ], + "spans": [ + { + "bbox": [ + 353, + 198, + 465, + 207 + ], + "type": "text", + "content": "(b) Feature distance in classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 217, + 546, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 546, + 243 + ], + "type": "text", + "content": "Figure 2: Feature distance of subclasses and classes on CIFAR-100-LT with imbalance ratio 100. We randomly sample instances from many-shot and medium-shot classes so that the size of each equals to that of few-shot classes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 262, + 289, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 289, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 289, + 419 + ], + "type": "text", + "content": "Dynamic temperature. In Table 5, we also study the effectiveness of dynamic temperature (Section 3.2). We remove the dynamic temperature and simply set " + }, + { + "bbox": [ + 46, + 262, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\tau_{2} > \\tau_{1}" + }, + { + "bbox": [ + 46, + 262, + 289, + 419 + ], + "type": "text", + "content": " following [27]. With the fixed temperature " + }, + { + "bbox": [ + 46, + 262, + 289, + 419 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 46, + 262, + 289, + 419 + ], + "type": "text", + "content": ", the performance of SBCL is significantly worse than that with dynamic temperature. We speculate that this is because dynamic temperature could help prevent the instances of a class to form overly large or small cluster in the feature space and therefore lead to better learned representations. Additionally, to evaluate the impact of dynamic temperature on other baselines, we apply the dynamic temperature on TSC, as reported in Appendix A.1. In Appendix A.1, we also provide an additional analysis of other hyperparameters on SBCL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 426, + 228, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 228, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 228, + 440 + ], + "type": "text", + "content": "4.6. Illustrate the superiority of SBCL" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 288, + 517 + ], + "type": "text", + "content": "In this section, we present an example that explicitly demonstrates the superiority of SBCL compared to other baselines. We fix the features learned from previous approaches and SBCL on the CIFAR-100-LT with imbalance ratio 100, and use these features for classification task. Table 6 shows that SBCL achieves the best performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 518, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 541 + ], + "type": "text", + "content": "Table 6: Accuracy(%) of SBCL and other baselines on CIFAR-100-LT with imbalanced ratio 100." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 67, + 551, + 264, + 578 + ], + "blocks": [ + { + "bbox": [ + 67, + 551, + 264, + 578 + ], + "lines": [ + { + "bbox": [ + 67, + 551, + 264, + 578 + ], + "spans": [ + { + "bbox": [ + 67, + 551, + 264, + 578 + ], + "type": "table", + "html": "
MethodCESCLKCLTSCSBCL
ACC(%)38.336.637.838.739.9
", + "image_path": "088f1638f6e5761cf54e2085e1e2b42bfbb4892805335d141be3869db5931797.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 597, + 206, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 206, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 206, + 609 + ], + "type": "text", + "content": "4.7. Feature distribution of SBCL" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": "To analyze the representation learned by SBCL, we firstly define the euclidean distance between a given sample and other samples from the same/different classes as intra/inter-class distance. Concretely, the euclidean distance between a sample " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " and a set " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(z_i,S) = \\frac{1}{|S|}\\sum_{z_j\\in S}\\| z_i - z_j\\| _2" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": ". Then, the intra- and interclass distance of sample " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " can be defined as " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(z_i,P_i)" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(z_i,\\mathcal{D} / P_i)" + }, + { + "bbox": [ + 46, + 616, + 288, + 715 + ], + "type": "text", + "content": " separately; and the intra- and inter-subclass" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 262, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 287 + ], + "type": "text", + "content": "distance of sample " + }, + { + "bbox": [ + 305, + 262, + 545, + 287 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 305, + 262, + 545, + 287 + ], + "type": "text", + "content": " can be defined as " + }, + { + "bbox": [ + 305, + 262, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(z_i,M_i)" + }, + { + "bbox": [ + 305, + 262, + 545, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 262, + 545, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(z_i,P_i / M_i)" + }, + { + "bbox": [ + 305, + 262, + 545, + 287 + ], + "type": "text", + "content": " separately." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 288, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 547, + 443 + ], + "type": "text", + "content": "Figure 2 shows the average distance between features learned by SBCL in subclasses and classes on the CIFAR-100-LT. As shown in Figure 2a, the distance between samples from the same subclass is less than those from the same class but different subclasses. Meanwhile, in Figure 2b, the inter-class distance consistently exceeds the intra-class distance, indicating clear separation between features from different classes. Additionally, the inter-class distance remains stable, which suggests that each class is uniformly distributed on a hypersphere. The results in all indicate that the two-layer class hierarchy is successfully captured and feature distribution achieves the core idea of SBCL. The specific values of feature distance are displayed in Appendix A.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 444, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 547, + 540 + ], + "type": "text", + "content": "Figure 3 visualizes the feature distribution of subclasses in a head class and head classes in CIFAR-100-LT with imbalance ratio 100 trained by SBCL using t-SNE [59]. The representations learned by SBCL form some separated clusters, which be aligned with subclass labels and class labels. This suggests that SBCL can learn an embedding space where features from same subclasses/classes are pulled closely and features from different subclasses/class are separated." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 550, + 537, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 537, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 537, + 563 + ], + "type": "text", + "content": "4.8. Variation of per-class weight during training" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 568, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 652 + ], + "type": "text", + "content": "Weight decay [1] achieves superior performance by effectively tuning weight decay during training. This study highlights the significance of regularizing network weights to become balanced for long-tailed recognition. Therefore, it is crucial to investigate whether SBCL can help the linear classifier to learn balanced weights through the training process." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Figure 4 shows the per-class weight norm of a linear classifier trained on top of features learned by SBCL in different training stages on CIFAR-100-LT. From the figure, we can see that as the training proceeds, the per-class weight norm becomes balanced even when training the linear classifier," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5402" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 71, + 267, + 205 + ], + "blocks": [ + { + "bbox": [ + 102, + 71, + 267, + 205 + ], + "lines": [ + { + "bbox": [ + 102, + 71, + 267, + 205 + ], + "spans": [ + { + "bbox": [ + 102, + 71, + 267, + 205 + ], + "type": "image", + "image_path": "4c339c5d2ff0e13c098d1be84be93bb18c7192efa2eb41f976b78ebcf18a2f5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 211, + 266, + 221 + ], + "lines": [ + { + "bbox": [ + 101, + 211, + 266, + 221 + ], + "spans": [ + { + "bbox": [ + 101, + 211, + 266, + 221 + ], + "type": "text", + "content": "(a) Distribution of subclasses on a head class." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 327, + 72, + 492, + 205 + ], + "blocks": [ + { + "bbox": [ + 327, + 72, + 492, + 205 + ], + "lines": [ + { + "bbox": [ + 327, + 72, + 492, + 205 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 492, + 205 + ], + "type": "image", + "image_path": "fa7efb1b566fb25a5e16be51e8f0f10df6668a017219eef2636f4bec1f0a1ec2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 211, + 493, + 222 + ], + "lines": [ + { + "bbox": [ + 326, + 211, + 493, + 222 + ], + "spans": [ + { + "bbox": [ + 326, + 211, + 493, + 222 + ], + "type": "text", + "content": "(b) Distribution of head classes on the dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 232, + 546, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 546, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 546, + 257 + ], + "type": "text", + "content": "Figure 3: Feature distribution of subclasses and classes on CIFAR-100-LT with imbalance ratio 100. Color represents subclasses/classes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 277, + 203, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 203, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 203, + 289 + ], + "type": "text", + "content": "the original cross-entropy loss is used." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 70, + 303, + 264, + 434 + ], + "blocks": [ + { + "bbox": [ + 70, + 303, + 264, + 434 + ], + "lines": [ + { + "bbox": [ + 70, + 303, + 264, + 434 + ], + "spans": [ + { + "bbox": [ + 70, + 303, + 264, + 434 + ], + "type": "image", + "image_path": "3735514970833759f85269c51acb17992d9c8cf8be4b8d0fe523e44bc58578ff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 446, + 288, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 446, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 288, + 483 + ], + "type": "text", + "content": "Figure 4: Change in weight norm of the linear classifier based on the representations trained by SBCL on CIFAR-100-LT with imbalance ratio 100 during training." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 489, + 132, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 132, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 132, + 502 + ], + "type": "text", + "content": "5. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Traditional methods for handling long-tailed recognition problem includes re-sampling and re-weighting. There are roughly two types of re-sampling techniques: over-sampling the minority classes [4, 5, 54, 70] and under-sampling the frequent classes [4, 23, 32]. The re-weighting techniques assign adaptive weights for different classes or even different samples. The vanilla scheme re-weights classes proportionally to the inverse of their frequency [29, 30, 64]. For class-level re-weighting methods, many loss functions including CB loss [15], LDAM loss [6] and Balanced softmax loss [50] were recently proposed, while instance-level re-weighting methods include Focal loss [44] and Influence-balanced loss [49]. Recently, two-stage algorithms have achieved remarkable performance for long-tailed recognition, such as classifier re-training (cRT) [34], learnable weight scaling (LWS) [34], and Mixup Shifted Label-Aware Smoothing model (MiSLAS) [72]. Meanwhile, bilateral branch net" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 277, + 547, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 385 + ], + "type": "text", + "content": "work (BBN) [73] uses an additional network branch for rebalancing. RIDE [62] use multiple branches named experts, each learning to specialize in the entire classes. LADE [28] assumes the prior of test class distributions is available and accordingly post-adjust model predictions. PaCo [14] applies parametric class-wise learnable centers to rebalance in contrastive learning. BCL [74] proposes a multi-branch framework to achieve class-averaging and class-complement in the training process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 386, + 548, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 548, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 548, + 576 + ], + "type": "text", + "content": "To boost the performance of the two-stage algorithms, researchers have introduced supervised contrastive learning [36] to the first feature-learning stage and proposed " + }, + { + "bbox": [ + 304, + 386, + 548, + 576 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 386, + 548, + 576 + ], + "type": "text", + "content": "-positive contrastive loss (KCL) [33] and targeted supervised contrastive learning (TSC) [41]. While achieving the state-of-the-art performance, these methods inject class-balance in the contrastive learning objective, inevitably leading to instance-imbalance during training. In this work, we instead propose to achieve both subclass- and instance-balance in the contrastive learning object. Our method is also related to recent studies of clustering-based deep unsupervised learning [7,8,18,43,65,66], especially those that leverage contrastive learning [22,40,42,63]. However, they target at general unsupervised representation learning scenario, while our method is tailored for long-tailed recognition where the training data is immensely class-imbalanced." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 378, + 597 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we introduced Subclass-balancing Contrastive Learning (SBCL) for long-tailed recognition. It breaks down the head classes into multiple semantically-coherent subclasses via subclass-balancing adaptive clustering and incorporates a bi-granularity contrastive loss that encourages both subclass- and instance-balance. Extensive experiments on multiple datasets demonstrate that SBCL achieves state-of-the-art single-model performance on benchmark datasets for long-tailed recognition." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5403" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 150 + ], + "type": "text", + "content": "[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6897-6907, 2022. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 153, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 153, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 153, + 288, + 213 + ], + "type": "text", + "content": "[2] Shin Ando and Chun Yuan Huang. Deep over-sampling framework for classifying imbalanced data. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 770-785. Springer, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 216, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 216, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 216, + 288, + 263 + ], + "type": "text", + "content": "[3] David Arthur and Sergei Vassilvitskii. K-means++ the advantages of careful seeding. In Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms, pages 1027-1035, 2007. 3, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 266, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 288, + 313 + ], + "type": "text", + "content": "[4] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106:249-259, 2018. 1, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 316, + 288, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 288, + 364 + ], + "type": "text", + "content": "[5] Jonathon Byrd and Zachary Lipton. What is the effect of importance weighting in deep learning? In International Conference on Machine Learning, pages 872-881. PMLR, 2019. 1, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 367, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 367, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 54, + 367, + 288, + 425 + ], + "type": "text", + "content": "[6] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in Neural Information Processing Systems, 32, 2019. 1, 5, 6, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 428, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 428, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 54, + 428, + 288, + 476 + ], + "type": "text", + "content": "[7] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In European Conference on Computer Vision, 2018. 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 479, + 288, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 479, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 54, + 479, + 288, + 538 + ], + "type": "text", + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 2020. 5, 6, 9, 18" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 541, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 541, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 541, + 288, + 589 + ], + "type": "text", + "content": "[9] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. Smote: synthetic minority over-sampling technique. Journal of artificial intelligence research, 16:321-357, 2002. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 591, + 288, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 650 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 650 + ], + "type": "text", + "content": "[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5008-5017, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 654, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 654, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 654, + 288, + 712 + ], + "type": "text", + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 108 + ], + "type": "text", + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 113, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 113, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 547, + 160 + ], + "type": "text", + "content": "[13] Collobert, Ronan, Weston, and Jason. A unified architecture for natural language processing: deep neural networks with multitask learning. In International Conference on Machine Learning, 2008. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 165, + 547, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 547, + 213 + ], + "type": "text", + "content": "[14] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 7, 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 217, + 547, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 547, + 276 + ], + "type": "text", + "content": "[15] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9268-9277, 2019. 1, 5, 6, 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 281, + 547, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 281, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 281, + 547, + 340 + ], + "type": "text", + "content": "[16] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422-1430, 2015." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 345, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 345, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 345, + 547, + 392 + ], + "type": "text", + "content": "[17] Qi Dong, Shaogang Gong, and Xiatian Zhu. Imbalanced deep learning by minority class incremental rectification. IEEE transactions on pattern analysis and machine intelligence, 41(6):1367-1381, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 397, + 547, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 397, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 307, + 397, + 547, + 456 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. Advances in Neural Information Processing Systems, 27, 2014. 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 461, + 547, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 461, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 461, + 547, + 497 + ], + "type": "text", + "content": "[19] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 502, + 547, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 547, + 561 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 547, + 561 + ], + "type": "text", + "content": "[20] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In Proceedings of the IEEE/cvf International Conference on computer vision, pages 6391-6400, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 566, + 547, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 547, + 649 + ], + "type": "text", + "content": "[21] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6, 18" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 654, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 654, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 654, + 547, + 712 + ], + "type": "text", + "content": "[22] Yuanfan Guo, Minghao Xu, Jiawen Li, Bingbing Ni, Xuanyu Zhu, Zhenbang Sun, and Yi Xu. Hcsc: Hierarchical contrastive selective coding. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2022. 9" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5404" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 108 + ], + "type": "text", + "content": "[23] Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering, 21(9):1263-1284, 2009. 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 111, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 111, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 111, + 288, + 171 + ], + "type": "text", + "content": "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2020. 6, 16, 17, 18" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 288, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 221 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 221 + ], + "type": "text", + "content": "[25] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 16, 17" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 224, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 224, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 288, + 283 + ], + "type": "text", + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 346 + ], + "type": "text", + "content": "[27] David T Hoffmann, Nadine Behrmann, Juergen Gall, Thomas Brox, and Mehdi Noroozi. Ranking info noise contrastive estimation: Boosting contrastive learning via ranked positives. the Association for the Advance of Artificial Intelligence, 2022. 4, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 350, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 421 + ], + "type": "text", + "content": "[28] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, 2021. 2, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 424, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 424, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 424, + 288, + 483 + ], + "type": "text", + "content": "[29] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Learning deep representation for imbalanced classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5375-5384, 2016. 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 487, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 545 + ], + "type": "text", + "content": "[30] Chen Huang, Yining Li, Chen Change Loy, and Xiaou Tang. Deep imbalanced learning for face recognition and attribute prediction. IEEE transactions on pattern analysis and machine intelligence, 42(11):2781-2794, 2019. 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 549, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 549, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 549, + 288, + 597 + ], + "type": "text", + "content": "[31] Ahmet Iscen, André Araujo, Boqing Gong, and Cordelia Schmid. Class-balanced distillation for long-tailed visual recognition. Proceedings of the IEEE international conference on computer vision, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 600, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 635 + ], + "type": "text", + "content": "[32] Nathalie Japkowicz and Shaju Stephen. The class imbalance problem: A systematic study. Intelligent data analysis, 6(5):429-449, 2002. 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 639, + 288, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 639, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 48, + 639, + 288, + 686 + ], + "type": "text", + "content": "[33] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 1, 2, 5, 6, 9, 16, 18" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 689, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 288, + 713 + ], + "type": "text", + "content": "[34] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 108 + ], + "type": "text", + "content": "Kalantidis. Decoupling representation and classifier for long-tailed recognition. International Conference on Learning Representations, 2019. 2, 5, 6, 9, 16" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 111, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 111, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 111, + 547, + 172 + ], + "type": "text", + "content": "[35] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7482-7491, 2018." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 547, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 235 + ], + "type": "text", + "content": "[36] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 1, 2, 5, 6, 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 239, + 547, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 547, + 299 + ], + "type": "text", + "content": "[37] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13896-13905, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 303, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 327 + ], + "type": "text", + "content": "[38] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 331, + 547, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 547, + 366 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 547, + 366 + ], + "type": "text", + "content": "[39] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 371, + 547, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 547, + 418 + ], + "type": "text", + "content": "[40] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations, 2021. 4, 5, 6, 9, 18" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 422, + 547, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 547, + 494 + ], + "type": "text", + "content": "[41] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 5, 6, 7, 9, 16, 17" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 498, + 547, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 547, + 534 + ], + "type": "text", + "content": "[42] Yunfan Li, Peng Hu, Zitao Liu, Dezhong Peng, Joey Tianyi Zhou, and Xi Peng. Contrastive clustering. AAAI Conference on Artificial Intelligence, 2021. 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 538, + 547, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 538, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 538, + 547, + 586 + ], + "type": "text", + "content": "[43] Renjie Liao, Alex Schwing, Richard Zemel, and Raquel Urtasun. Learning deep parsimonious representations. Advances in Neural Information Processing Systems, 29, 2016. 9" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 590, + 547, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 649 + ], + "type": "text", + "content": "[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2980-2988, 2017. 5, 6, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 653, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 653, + 547, + 713 + ], + "type": "text", + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5405" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 133 + ], + "type": "text", + "content": "[46] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 135, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 135, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 135, + 288, + 206 + ], + "type": "text", + "content": "[47] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In European conference on computer vision, pages 181-196, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 209, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 258 + ], + "type": "text", + "content": "[48] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 261, + 288, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 320 + ], + "type": "text", + "content": "[49] Seulki Park, Jongin Lim, Younghan Jeon, and Jin Young Choi. Influence-balanced loss for imbalanced visual classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 735-744, 2021. 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 323, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 323, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 48, + 323, + 288, + 372 + ], + "type": "text", + "content": "[50] Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta-softmax for long-tailed visual recognition. Advances in Neural Information Processing Systems, 33:4175–4186, 2020. 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 374, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 374, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 48, + 374, + 288, + 423 + ], + "type": "text", + "content": "[51] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6, 17" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 426, + 288, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 497 + ], + "type": "text", + "content": "[52] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 501, + 288, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 560 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 560 + ], + "type": "text", + "content": "[53] Xie Saining, Girshick Ross, Dollar Piotr, Tu Zhuowen, and He Kaiming. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page 1492-1500, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 563, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 563, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 288, + 612 + ], + "type": "text", + "content": "[54] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In European conference on computer vision, pages 467-482. Springer, 2016. 1, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 614, + 288, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 662 + ], + "type": "text", + "content": "[55] Tony Shen, Ariel Lee, Carol Shen, and C Jimmy Lin. The long tail and rare disease research: the impact of next-generation sequencing for rare mendelian disorders. Genetics research, 97, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 665, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 665, + 288, + 713 + ], + "type": "text", + "content": "[56] Aditya Singh and Alessandro Bay. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in Neural Information Processing Systems, 2019." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 107 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 107 + ], + "type": "text", + "content": "[57] Merrielle Spain and Pietro Perona. Measuring and predicting importance of objects in our visual world. 2007. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 112, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 112, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 112, + 547, + 172 + ], + "type": "text", + "content": "[58] AFM Uddin, Mst Monira, Wheemyung Shin, Tae-Choong Chung, Sung-Ho Bae, et al. Saliency: A saliency guided data augmentation strategy for better regularization. International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 175, + 547, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 211 + ], + "type": "text", + "content": "[59] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 216, + 547, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 216, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 216, + 547, + 287 + ], + "type": "text", + "content": "[60] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 8769-8778, 2018. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 292, + 547, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 292, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 292, + 547, + 339 + ], + "type": "text", + "content": "[61] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 343, + 547, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 547, + 390 + ], + "type": "text", + "content": "[62] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. International Conference on Learning Representations, 2021. 7, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 395, + 547, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 395, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 547, + 454 + ], + "type": "text", + "content": "[63] Xudong Wang, Ziwei Liu, and Stella X Yu. Unsupervised feature learning by cross-level instance-group discrimination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12586-12595, 2021. 5, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 458, + 547, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 458, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 458, + 547, + 495 + ], + "type": "text", + "content": "[64] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. Advances in Neural Information Processing Systems, 30, 2017. 1, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 498, + 547, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 547, + 546 + ], + "type": "text", + "content": "[65] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pages 478-487. PMLR, 2016. 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 550, + 547, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 550, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 550, + 547, + 609 + ], + "type": "text", + "content": "[66] Jianwei Yang, Devi Parikh, and Dhruv Batra. Joint unsupervised learning of deep representations and image clusters. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5147-5156, 2016. 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 614, + 547, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 547, + 661 + ], + "type": "text", + "content": "[67] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. Advances in Neural Information Processing Systems, 33:19290-19301, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 665, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 665, + 547, + 713 + ], + "type": "text", + "content": "[68] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. International Conference on Machine Learning, 2019. 1" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5406" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 483 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 120 + ], + "type": "text", + "content": "[69] Han-Jia Ye, Hong-You Chen, De-Chuan Zhan, and Wei-Lun Chao. Identifying and compensating for feature deviation in imbalanced deep learning. arXiv preprint arXiv:2001.01385, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 123, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 123, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 123, + 288, + 171 + ], + "type": "text", + "content": "[70] Qiaoyong Zhong, Chao Li, Yingying Zhang, H Sun, S Yang, D Xie, and S Pu. Towards good practices for recognition & detection. In CVPR workshops, volume 1, page 2, 2016. 9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 246 + ], + "type": "text", + "content": "[71] Yaoyao Zhong, Weihong Deng, Mei Wang, Jiani Hu, Jianteng Peng, Xunqiang Tao, and Yaohai Huang. Unequal-training for deep face recognition with long-tailed noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7812-7821, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 249, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 249, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 288, + 308 + ], + "type": "text", + "content": "[72] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16489-16498, 2021. 7, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 312, + 288, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 312, + 288, + 372 + ], + "spans": [ + { + "bbox": [ + 48, + 312, + 288, + 372 + ], + "type": "text", + "content": "[73] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, 2020. 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 374, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 374, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 374, + 288, + 445 + ], + "type": "text", + "content": "[74] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 7, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 449, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 483 + ], + "type": "text", + "content": "[75] George Kingsley Zipf. The psycho-biology of language: An introduction to dynamic philology. Routledge, 2013. 1" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5407" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/92e5510f-e140-45f8-8597-b810b193a66f_content_list.json b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/92e5510f-e140-45f8-8597-b810b193a66f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6fee32f3af8d63bbf8834c6272494914cf1c2213 --- /dev/null +++ b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/92e5510f-e140-45f8-8597-b810b193a66f_content_list.json @@ -0,0 +1,1472 @@ +[ + { + "type": "text", + "text": "SupFusion: Supervised LiDAR-Camera Fusion for 3D Object Detection", + "text_level": 1, + "bbox": [ + 122, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiran Qin $^{1*}$ , Chaoqun Wang $^{1*}$ , Zijian Kang $^{2}$ , Ningning Ma $^{2}$ , Zhen Li $^{3}$ , Ruimao Zhang $^{1\\dagger}$", + "bbox": [ + 140, + 179, + 834, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Data Science, Shenzhen Research Institute of Big Data,", + "bbox": [ + 220, + 198, + 748, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen), China $^{2}$ NIO", + "bbox": [ + 156, + 215, + 810, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Science and Engineering, Future Intelligent Network Research Institute,", + "bbox": [ + 156, + 233, + 812, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen), China", + "bbox": [ + 181, + 252, + 787, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yiranqin@link, chaoqunwang@link, lizhen@, ruimaozhang@}cuhk.edu.hk, {zijiankang, ningningma}@nio.com", + "bbox": [ + 88, + 272, + 880, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR-Camera fusion-based 3D detection is a critical task for automatic driving. In recent years, many LiDAR-Camera fusion approaches sprung up and gained promising performances compared with single-modal detectors, but always lack carefully designed and effective supervision for the fusion process. In this paper, we propose a novel training strategy called SupFusion, which provides an auxiliary feature level supervision for effective LiDAR-Camera fusion and significantly boosts detection performance. Our strategy involves a data enhancement method named Polar Sampling, which densifies sparse objects and trains an assistant model to generate high-quality features as the supervision. These features are then used to train the LiDAR-Camera fusion model, where the fusion feature is optimized to simulate the generated high-quality features. Furthermore, we propose a simple yet effective deep fusion module, which contiguously gains superior performance compared with previous fusion methods with SupFusion strategy. In such a manner, our proposal shares the following advantages. Firstly, SupFusion introduces auxiliary feature-level supervision which could boost LiDAR-Camera detection performance without introducing extra inference costs. Secondly, the proposed deep fusion could continuously improve the detector's abilities. Our proposed SupFusion and deep fusion module is plug-and-play, we make extensive experiments to demonstrate its effectiveness. Specifically, we gain around $2\\%$ 3D mAP improvements on KITTI benchmark based on multiple LiDAR-Camera 3D detectors. Our code is available at https://github.com/IranQin/SupFusion.", + "bbox": [ + 73, + 364, + 473, + 806 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54814ff1800592069e17aa0443576449f7f4e859af0ae525027806812b972f98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 323, + 888, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ff36b3f8ea3bf9970e14f08d8bec410d40d31cd7ca77f7da2f65c9131bf1e89.jpg", + "image_caption": [ + "Figure 1: Top: The previous LiDAR-Camera 3D detector, the fusion module is optimized by detection loss. Bottom: Our proposed SupFusion, we introduce auxiliary supervision via high-quality features provided by an assistant model." + ], + "image_footnote": [], + "bbox": [ + 506, + 431, + 888, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 705, + 632, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR-Camera fusion-based 3D object detection is a critical and challenging task for autonomous driving and robotics [32, 35]. Previous approaches [20, 22, 34, 19] always project the camera inputs to LiDAR BEV or voxel space via intrinsic and extrinsic to align the LiDAR and camera features. Then, simple concatenation or summation is employed to obtain the fusion features for final detection. Also, some deep learning-based fusion approaches [1, 21] sprung up and gained promising performances. However, the previous fusion approaches always directly optimized the 3D/2D feature extraction as well as the fusion module by the", + "bbox": [ + 496, + 733, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Work done during an internship at NIO.", + "bbox": [ + 94, + 875, + 416, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22014", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "detection loss, which lacks carefully designed and effective supervision at the feature level and limits its performance.", + "bbox": [ + 76, + 90, + 467, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, distillation manners have shown great improvements in feature-level supervision for 3D detection. For instance, some approaches [5, 6] provide LiDAR features to guide the 2D backbone in estimating the depth information from camera inputs. Also, some approaches [19] provide LiDAR-Camera fusion features to supervise LiDAR backbone learns global and contextual presentation from LiDAR inputs. Introducing feature-level auxiliary supervision via simulating more robust and high-quality features, the detectors could boost marginal improvements. Motivated by this, a natural solution to handle LiDAR-Camera feature fusion is providing stronger and high-quality features and introducing auxiliary supervision for LiDAR-Camera 3D detection.", + "bbox": [ + 75, + 121, + 468, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, in this paper we propose a supervised LiDAR-Camera fusion method named SupFusion, to generate high-quality features and provide effective supervision for the fusion as well as feature extraction process, and further boost the LiDAR-Camera fusion-based 3D detection performances. Specifically, we first train an assistant model to provide high-quality features. To achieve it, different from previous approaches, which leverage larger models or extra data, we proposed a novel data enhancement method named Polar Sampling. The polar sampling could dynamically enhance the object's density from sparse LiDAR data, which is easier to detect and improve features quality, e.g., the feature could conclude accuracy detection results. Then, we simply train the LiDAR-Camera fusion-based detector with introduced auxiliary feature-level supervision. In this step, we feed the raw LiDAR and camera inputs into 3D/2D backbones and fusion modules to obtain fusion features. In one aspect, the fusion features are fed into the detection head for final prediction, which is decision-level supervision. In another aspect, the auxiliary supervision mimics the fusion feature to high-quality features, which are obtained via the pretrained assistant model and enhanced LiDAR data. In such a manner, the proposed feature-level supervision could lead the fusion module to generate more robust features, and further boost the detection performance. To better fuse the LiDAR and camera features, we propose a simple yet effective deep fusion module, which consists of stacked MLP blocks and dynamic fusion blocks. SupFusion could sufficiently excavate the deep fusion module capacity, and continuously improve detection accuracy.", + "bbox": [ + 75, + 316, + 470, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared with previous LiDAR-Camera 3D detectors, this is a brand new attempt to improve the effectiveness of the LiDAR-Camera fusion by introducing feature-level supervision, which consistently boosts 3D detection accuracy. The main contributions are summarized as follows.", + "bbox": [ + 76, + 770, + 470, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a novel supervised fusion training strategy named SupFusion, which mainly consists of a high-quality feature generation process and firstly propose", + "bbox": [ + 94, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "auxiliary feature-level supervision loss for robust fusion feature extraction and accuracy 3D detection to our knowledge.", + "bbox": [ + 529, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To obtain high-quality features in SupFusion, we propose a data enhancement method named Polar Sampling to densify the sparse objects. Furthermore, we propose an efficient deep fusion module to contiguously boost detection accuracy.", + "- We make extensive experiments based on multiple detectors with different fusion strategies, and gain around $2\\%$ mAP improvements on KITTI benchmark. The source code will be released after the blind review." + ], + "bbox": [ + 517, + 146, + 893, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 306, + 640, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Object Detection", + "text_level": 1, + "bbox": [ + 500, + 330, + 692, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera-based 3D Object Detection methods perform 3D detection on single- or multi-view images. Many methods are proposed to address monocular 3D detection[24, 23, 17, 51, 53, 27, 37, 36] since KITTI benchmark[9] only contains one from the camera. With the development of autonomous driving datasets which have more cameras covering a wider range of angles(nuScenes[2] and Waymo[33]), more and more methods[38, 40, 42] are developed that take multi-view images as input and get surprising performance compared to monocular methods.", + "bbox": [ + 496, + 354, + 893, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR-based 3D Object Detection methods perform 3D detection on a point cloud scene. Depending on the three kinds of point cloud data pretreatments, 3D detectors are also classified into three mainstream. Point-based method [31, 46, 50, 4] use row point cloud data, leveraging pointnet[25] and pointnet++[26] as 3D backbone networks to get point-wised features which are sent to point-wise 3D detection head. Pillar-based method [18, 41, 29] pretreat point cloud to pillars, gain speed up while still maintaining a good performance. Voxel-based methods[54, 45, 8] are widely used, point clouds are voxelized to voxel and sent to a 3D sparse convolutional[10] based 3D backbone to get bird's-eye-view features. Some methods[30] also try to gain more improvement by using both the voxel feature and the point feature.", + "bbox": [ + 496, + 505, + 893, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. LiDAR-Camera 3D Object Detection", + "text_level": 1, + "bbox": [ + 500, + 739, + 818, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR-camera fusion is a technique that combines data from LiDAR and camera sensors to obtain a more complete understanding of a scene in 3D. Due to the complementary nature of the features produced by these two modalities, researchers have developed methods that can be jointly optimized on both sensors. These methods can be classified into two categories based on their fusion mechanism. The first category is point-level fusion[14, 35, 47], where the image features are queried via the raw LiDAR points and then", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22015", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/57afa910eb861de1fe39472dea66f52a4aaa28851ec7bb490604103f32704b56.jpg", + "image_caption": [ + "Figure 2: The detector training procedure. For a given LiDAR-Camera sample, we extract the feature presentation via the corresponding 3D/2D encoders, and the proposed deep fusion module is employed to fuse the multi-modal features. Besides the detection loss $\\mathcal{L}_{Det}$ , we introduce the auxiliary feature-level supervision $\\mathcal{L}_{Sim}$ , to mimic the fusion feature $\\mathbf{f}_i^{lc}$ to the high-quality features $\\mathbf{f}_l^*$ , which is generated by the assistant model and enhanced data. Here the $\\odot$ , $\\oplus$ , and $\\odot$ indicate the element-wise multiplication, summation, and channel-wise concatenation. More details could be referred in Sec. 3.1" + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 885, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "concatenated back as additional point features. The second category is feature-level fusion, where the LiDAR points are projected into a specific feature space[48, 20, 19] or proposals level feature fusion[3, 16, 1, 43]. There are also methods that combine these two fusion manners, such as MVX-Net [32]. Recently, attention-based fusion methods have sprung up. TransFusion[1] uses the bounding box prediction of LiDAR features as a proposal to query the image feature, and then a Transformer-like architecture fuses the information back to LiDAR features. Similarly, DeepFusion[21] projects LiDAR features on each view image as queries and leverages cross-attention for the two modalities. Bird's-eye-view space feature fusion[22, 34] is proposed recently and gained great performance with different tasks.", + "bbox": [ + 75, + 416, + 472, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Knowledge Distillation for 3D Object Detection", + "text_level": 1, + "bbox": [ + 76, + 635, + 470, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object detection tasks differ from classification tasks in that they require consideration of both global and local features. While numerous distillation methods have been proposed to address this challenge, such as those developed by [39, 7, 11, 49], many of these methods require the creation of specific distillation regions or objects based on ground-truth data, bounding box predictions, attention, or gradients to mitigate the imbalance between foreground and background. In addition to 2D detection, there has been growing interest in cross-model knowledge distillation methods between RGB and LiDAR in 3D object detection. For example, [12, 28, 6, 5] have developed methods that transfer knowledge from LiDAR detectors to image detectors. [52, 15] has trained a teacher model similar to that proposed by [35] and developed a multilevel distillation framework to transfer semantic segmentation information to a student model. In this", + "bbox": [ + 75, + 659, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work, we want to further boost LiDAR-Camera detectors with an auxiliary feature level supervision.", + "bbox": [ + 498, + 416, + 890, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 462, + 635, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first overview the overall SupFusion pipeline in Sec. 3.1, and then detail introduces the main components, including the proposed Polar Sampling in Sec. 3.2, and deep fusion module in Sec. 3.3.", + "bbox": [ + 498, + 488, + 893, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall SupFusion Pipeline", + "text_level": 1, + "bbox": [ + 500, + 559, + 746, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a given LiDAR-Camera fusion-based 3D detector $\\mathbf{D}$ , which consists of a 3D Encoder $\\mathbf{E}_l$ , a 2D Encoder $\\mathbf{E}_c$ , a fusion module $\\mathbf{F}$ , and a detection head. Denotes $\\mathcal{D} = \\{\\left(\\mathbf{l}_1,\\mathbf{c}_1\\right),\\left(\\mathbf{l}_2,\\mathbf{c}_2\\right),\\dots,\\left(\\mathbf{l}_N,\\mathbf{c}_N\\right)\\}$ and $\\mathcal{V} = \\{\\mathbf{y}_1,\\mathbf{y}_2,\\dots,\\mathbf{y}_N\\}$ are the given dataset with $N$ samples, where $(\\mathbf{l}_i,\\mathbf{c}_i)$ are the paired LiDAR and camera inputs. The previous approaches feed the paired $(\\mathbf{l}_i,\\mathbf{c}_i)$ into 3D/2D encoders $\\mathbf{E}_l$ and $\\mathbf{E}_c$ to capture corresponding feature presentation $\\mathbf{f}_i^l$ and $\\mathbf{f}_i^c$ . Then, the paired features presentation is fed into the fusion module to obtain the fusion feature $\\mathbf{f}_i^{lc}$ , which is used for the final prediction. To better leverage the LiDAR and camera data, we propose a supervised LiDAR-Camera Fusion training strategy named SupFusion, which provides high-quality features for feature-level supervision. To end this, different from previous approaches, which adopt larger scale models or extra data input, we propose a data enhancement method named polar sampling to densify the sparse objects in LiDAR data. In such a manner, the data is easier and could easily generate high-quality features (e.g., the feature could conclude accuracy detection results) with lightweight models. The proposed polar sampling mainly", + "bbox": [ + 498, + 583, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22016", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8a3162c7f7ddf1873005e730012f3e3b23b8cb2001fef0969b259e2889da6775.jpg", + "image_caption": [ + "Figure 3: High-quality feature generation process. For any given LiDAR sample, we densify the sparse objects via polar pasting, which calculate the direction and rotation to query dense object from the database, and pasting to add extra points to sparse objects. We first train the assistant model via the enhanced data and feed the enhanced LiDAR data into the assistant model to generate high-quality features $\\mathbf{f}^*$ after it is convergent. More details could be referred in Sec. 3.1." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 885, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "densifies the sparse object points in LiDAR data. Specifically, for any sample $(\\mathbf{l}_i,\\mathbf{c}_i)$ , where $\\mathbf{l}_i = \\{l_i^1,l_i^2,\\dots,l_i^{n_i},l_i^s\\}$ contains $n_i$ objects points sets and one background points set. In the polar sampling, we add extra points to each object point set from raw LiDAR data, to obtain enhanced LiDAR data $\\mathbf{l}_i' = \\{(l_i^1,\\hat{l}_i^1),(l_i^2,\\hat{l}_i^2),\\dots,(l_i^{n_i},\\hat{l}_i^{n_i}),l_i^s\\}$ in polar pasting, where $\\hat{l}_i^k$ is the added points set for the $k$ -th objects in $i$ -th LiDAR sample and generated in polar grouping. We will detail introduce the polar sampling in Sec. 3.2.", + "bbox": [ + 75, + 464, + 472, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "High-quality Feature Generation. To provide feature-level supervision in our SupFusion, we adopt an assistant model to capture high-quality features from the enhanced data as shown in Fig. 3. Firstly, we train an assistant model to provide high-quality features. Specifically, for any sample in $\\mathcal{D}$ , we enhance the sparse LiDAR data $\\mathbf{l}_i$ to obtain enhanced data $\\mathbf{l}_i^{\\prime}$ via the polar pasting, which densifies the spare objects with added points set generated in polar grouping. Then, after the assistant model convergence (often with few epochs), we feed the enhanced sample $\\mathbf{l}_i^{\\prime}$ into the optimized assistant model, to capture high-quality features $\\mathbf{f}_i^*$ , which is used to train the LiDAR-Camera 3D detector. To better apply to the given LiDAR-Camera detector and easier implement, we simply adopt the LiDAR branch detector as the assistant model.", + "bbox": [ + 75, + 608, + 470, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detector Training. For any given LiDAR-Camera detector, we train the model with the proposed auxiliary supervision at the feature level. Specifically, given sample $(\\mathbf{l}_i,\\mathbf{c}_i)\\in \\mathcal{D}$ , we first feed the LiDAR and camera into the 3D", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and 2D Encoders $\\mathbf{E}_l$ and $\\mathbf{E}_s$ to capture the corresponding features $\\mathbf{f}_i^l$ and $\\mathbf{f}_i^c$ , which are fed into the fusion model to generate fusion features $\\mathbf{f}_i^{l,c}$ , and flow into the detection head for final prediction. Furthermore, the proposed auxiliary supervision is employed to mimic the fusion feature with high-quality features, which is generated by the pre-trained assistant model and enhanced LiDAR data. The above process can be formulated as:", + "bbox": [ + 496, + 464, + 893, + 584 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {((\\mathbf {l} _ {i}, \\mathbf {c} _ {i}), \\mathbf {y} _ {i}) \\sim \\mathcal {D}} \\mathcal {L} _ {\\text {d e t}} (\\mathbf {D} (\\mathbf {l} _ {i}, \\mathbf {c} _ {i}, \\theta), \\mathbf {y} _ {i}) (1) \\\\ + \\lambda \\mathcal {L} _ {\\text {S i m}} \\left(\\mathbf {f} _ {i} ^ {l c}, \\mathbf {f} _ {i} ^ {*}\\right), (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 597, + 892, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad \\mathbf {f} _ {i} ^ {l c} = \\mathbf {F} \\left(\\mathbf {E} _ {l} \\left(\\mathbf {l} _ {i}, \\theta_ {l}\\right), \\mathbf {E} _ {c} \\left(\\mathbf {c} _ {i}, \\theta_ {c}\\right), \\theta_ {f}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 647, + 892, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {i} ^ {*} = \\mathbf {E} _ {l} \\left(\\mathbf {I} _ {i} ^ {\\prime}, \\theta_ {l} ^ {*}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 666, + 890, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S i m} = \\left\\| \\psi \\left(\\mathbf {f} _ {i} ^ {l c}\\right) - \\mathbf {f} _ {i} ^ {*} \\right\\| _ {2} ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 686, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta^{*}$ is the optimized parameters of detector $\\mathbf{D}$ , $\\mathbf{f}_i^*$ is the high-quality features, which is generated from pre-trained assistant model LiDAR encoder $\\mathbf{E}_l$ with parameters $\\theta_l^*$ , $\\psi$ is a simple convolutional layer with kernel size 1. $\\lambda$ is the balanced weight and we set it as 1.0.", + "bbox": [ + 496, + 715, + 892, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Polar Sampling", + "text_level": 1, + "bbox": [ + 500, + 801, + 656, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To provide high-quality features via the assistant model in our proposed SupFusion, we propose a novel data enhancement method named Polar Sampling, to ease the sparse problem which frequently leads to the detection failers. To end this, we densify the sparse objects in LiDAR data with", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22017", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a similar dense object. Specifically, the polar sampling consists of two parts, including polar grouping and polar pasting. In polar grouping, we mainly build a database to store the dense objects, which is used in polar pasting to densify the sparse objects.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the characteristics of LiDAR sensors, the collected point cloud data naturally exists a particular density distribution, e.g., the objects have more points on the surface towards the LiDAR sensor, while few points on the opposite sides. The density distribution is mainly influenced by the direction and rotation, while the points' density mainly depends on the distances, e.g. the objects closer to the LiDAR sensor have denser points. Motivated by this, we aim to densify the sparse objects in longer distances with the dense objects in short distances, according to their direction and rotation to keep the density distributions. Specifically, we build polar coordinate systems for the whole scene as well as the objects based on the center of the scene and specific objects and define the positive direction of the LiDAR sensor as 0 degrees to measure the corresponding directions and rotations. Then, we collect the objects with similar density distribution (e.g. with similar direction and rotation) and generate one dense object for each group in the polar grouping, and used it to densify the sparse objects in polar pasting.", + "bbox": [ + 75, + 167, + 473, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Polar Grouping. As shown in Fig. 4, we build a database $\\mathcal{B}$ to store the generated dense object point set $\\hat{l}$ according to their directions and rotations in the polar grouping, which is noted as $\\alpha$ and $\\beta$ in Fig. 4.", + "bbox": [ + 76, + 454, + 470, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, we search the whole dataset and calculated their polar angles for all objects via the locations and provided rotations in the benchmarks. Secondly, we split the objects into groups according to their polar angles. Specifically, we manually divide the direction and rotation into $N$ groups. For any object point set $l$ , we could put it into the corresponding group according to the index:", + "bbox": [ + 75, + 513, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {l} ^ {\\prime} = \\left\\lfloor \\frac {\\alpha_ {l} \\cdot 2 \\pi}{N} \\right\\rfloor , \\quad \\beta_ {l} ^ {\\prime} = \\left\\lfloor \\frac {\\beta_ {l} \\cdot 2 \\pi}{N} \\right\\rfloor , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 627, + 468, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{l}^{\\prime},\\beta_{l}^{\\prime}\\in \\{1,2,\\dots,N\\}$ are the direction and rotation index of object $l$ , and we could put the points set into $\\mathcal{B}$ as: $\\mathcal{B}_C^{\\alpha_l',\\beta_l'}\\gets l$ , where $C$ is the label of object $l$ . Finally, for each object group $\\mathcal{B}_C^{\\alpha ',\\beta '} = \\{l_1,l_2,\\ldots ,l_n\\}$ contains $n$ similar objects, we select the densest $k$ objects and mix up to generate one dense and complete object $\\hat{l}$ , and $\\mathcal{B}_C^{\\alpha ',\\beta '} = \\{\\hat{l}\\}$ is used in polar pasting phase. To keep the density and avoid too much calculation in the model, we randomly select partial points in the dense object. Here we set $k = 10$ and select 5000 points for all objects.", + "bbox": [ + 75, + 665, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Polar Pasting. As shown in Fig. 2, the polar pasting is utilized to enhance the sparse LiDAR data to train the assistant model and generate high-quality features. Concretely, given a LiDAR sample $\\mathbf{l}_i = \\{l_i^1, l_i^2, \\dots, l_i^{n_i}, l_i^s\\}$ contains $n_i$ objects, for any object $l_i^k$ , we can calculate the direction", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/63df54b7c17ffde47cf418c685d34796aeec5bfef97b2c2010b4724f5e142881.jpg", + "image_caption": [ + "Figure 4: Polar grouping mainly builds a database to store the dense objects of each group in polar coordinate systems of direction and rotations for each class. More details could be referred to in Sec. 3.2." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 885, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and rotation the same as the grouping process and query the dense object $\\hat{l}_i^k$ from $\\mathcal{B}$ according to the label and indexes, which could be obtained from E.q. 6 In such a manner, we enhance all the objects in the sample and obtain the enhanced data: $\\mathbf{I}_i' = \\{(l_i^1,\\hat{l}_i^1),(l_i^2,\\hat{l}_i^2),\\dots,(l_i^{n_i},\\hat{l}_i^{n_i}),l_i^s\\}$", + "bbox": [ + 496, + 377, + 893, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Deep Fusion", + "text_level": 1, + "bbox": [ + 500, + 463, + 633, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mimic the high-quality features generated from enhanced LiDAR data, the fusion model aims to extract the missing information of the sparse objects from the rich color and contextual features in camera input. To end this, we propose the deep fusion module, to leverage image features and complete the LiDAR presentation. The proposed deep fusion mainly consists of a 3D learner and a 2D3D learner. The 3D learner is a simple convolutional layer, to transfer the 3D presentation into 2D space. Then, concatenating the 2D features and 3D presentation (e.g. in 2D space), the 2D3D learner is employed to fuse the LiDAR-Camera features. Finally, we weighed the fusion feature via an MLP and an activate function, adding back to the raw LiDAR features as the output of the deep fusion module. The 2D3D learner consists of stacked MLP blocks with depth $K$ , learning to leverage camera features to complete the LiDAR representation from the sparse objects, to simulate the high-quality features of dense LiDAR objects.", + "bbox": [ + 496, + 487, + 893, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 775, + 633, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiments Settings", + "text_level": 1, + "bbox": [ + 500, + 800, + 699, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. KITTI [9] is a large-scale public benchmark for autonomous driving, including 3712 training, 3769 validation, and 7518 testing samples with LiDAR and camera pair data. All objects in KITTI are divided into 'easy', 'moderate', and 'hard' difficulties ac", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22018", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/abd8761755900a4921d9c94a5dbbe1babd19a55af263c4ccb88b0e18968a65c2.jpg", + "table_caption": [ + "Table 1: The experiment results (mAP@R40 %) on KITTI validation set. We list the easy, moderate (mod.), and hard cases for three classes, and the overall performances. Here, L, LC, LC* indicate the results of the corresponding LiDAR detector, LiDAR-Camera fusion detectors, and our proposal. $\\Delta$ indicates the improvements. The best results are shown in bold expect $\\mathrm{L}^{\\dagger}$ , which is the assistant model, and test on the enhanced validation set. MVXNet is re-conduct based on mmdetection3d. PV-RCNN-LC and Voxel-RCNN-LC are re-conduct based on the open-source code of VFF [20]." + ], + "table_footnote": [], + "table_body": "
DetectorModalityOverallPedestrianCyclistCar
easymod.hardeasymod.hardeasymod.hardeasymod.hard
MVXNetL†88.4490.9388.5887.3387.6180.1580.0986.0186.1997.8999.1799.41
L73.4658.6455.1557.5450.1045.5674.6351.8148.9288.2173.9970.98
LC74.9963.0059.3264.4657.7653.0572.6955.2751.6587.8275.9873.26
LC*75.3764.7360.5263.1158.1353.7374.7957.6053.8288.2078.4674.01
Δ+0.38+1.73+1.20-1.35+0.37+0.68+2.10+2.33+2.17+0.38+2.48+0.75
PV-RCNN-LCL†94.9095.2095.1086.8686.7887.0798.9099.3398.6298.9599.5099.62
L81.3771.5568.0867.8059.5654.4085.0670.8367.9691.2684.2581.88
LC84.1674.5470.8670.4564.1659.5090.4774.4670.2591.5685.0182.83
LC*85.7776.0072.0074.2966.6761.3891.0076.1371.6692.0285.2082.94
Δ+1.61+1.46+1.14+3.84+2.51+1.88+0.53+1.67+1.41+0.46+0.19+0.11
Voxel-RCNN-LCL†99.2699.6699.7597.8498.0598.1794.3694.6094.8599.9099.9099.91
L82.0171.6968.4067.8460.9556.0386.1071.3967.0592.0982.7482.12
LC84.3674.4670.8071.3964.8960.2089.6472.9469.0392.0585.5583.17
LC*86.1375.9072.1773.9566.0661.8392.0276.1871.8392.4285.4782.83
Δ+1.77+1.44+1.37+2.56+1.17+1.63+2.38+3.24+2.80+0.37-0.08-0.34
", + "bbox": [ + 76, + 167, + 901, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cording to the object scales and occlusion ratio. We evaluate the detection performance by the mean Average Precision under 40 recall thresholds (mAP@R40) the same as the official benchmark and evaluation for 3 classes including Pedestrian, Cyclist, and Car. nuScenes [2] is a large-scale autonomous driving dataset for 3D detection and tracking tasks, including 700 training, 150 validation, and 150 testing sequences, with each 20s long and annotated with 3D bounding boxes. For 3D object detection, the official evaluation metrics include the mean Average Precision (mAP) and nuScenes detection score (NDS). mAP measures the localization precision using a threshold base on the birds-eye view. NDS is a weighted combination of mAP and regression accuracy of other object attributes, including box size, orientation, and translation.", + "bbox": [ + 75, + 542, + 470, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detectors. To demonstrate the effectiveness of our proposed method, we apply the SupFusion training strategy and deep fusion module on variance LiDAR-Camera detectors including MVXNet [32], PV-RCNN-LC [30], and Voxel-RCNN-LC [8] on KITTI, and SECOND-LC [45] on nuScenes benchmark. MVXNet is a voxel-based LiDAR-Camera 3D detector based on VoxelNet [54]. The camera features project to LiDAR space and concatenates with points features. Voxel-RCNN and PV-RCNN are LiDAR", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "based detectors, we re-conduct the LiDAR-Camera fusion approach following VFF [20] based on the open-source code, which is a previous state-of-the-art fusion approach based on the two detectors, mainly projects the camera features into voxel presentation and summarized fusion with the voxel features of LiDAR input. SECOND-LC introduces camera inputs based on the LiDAR-based SECOND detector, the same as BEVFusion [22], we concatenate the camera and LiDAR features in birds-eys-view spaces.", + "bbox": [ + 496, + 542, + 893, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implement Details. We conduct experiments on MVXNet, PV-RCNN-LC and Voxel-RCNN-LC, SECOND-LC following the open-sourced code and mmdetection3d on 8 Tesla A100 GPUs. All the architecture, configuration, and optimizer follow the settings reported in papers including voxel size, point cloud range, batch size, and learning rate. We modify the fusion method (e.g. concatenation, summarized) by our proposed deep fusion module and applied our SupFusion training strategy. For the deep fusion module, we set depth $K = 3$ . For the auxiliary supervision, we utilize simple L2 loss and set $\\lambda = 1.0$ . More details about how to apply SupFusion and deep fusion modules on different detectors are demonstrated in the Appendix.", + "bbox": [ + 496, + 704, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22019", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ef883c62103f4fca0d862bb9e9eb41c970fc371bf7c9b56850510705c2d18664.jpg", + "table_caption": [ + "Table 2: NDS and mAP(%) results on nuScenes benchmarks." + ], + "table_footnote": [], + "table_body": "
Modality3D2DNDSMAP
L†SECOND-90.3894.44
LSECOND-60.5252.28
LCSECONDSwin-T62.5755.26
LC*SECONDSwin-T64.58(+2.01)56.64(+1.38)
", + "bbox": [ + 78, + 106, + 470, + 189 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8371bc60cfbf5dab3e636adbd3a69be5dff4d9707cedfd245c2f7f3859356671.jpg", + "table_caption": [ + "Table 3: Results comparison with different fusion methods. Sup. indicates our proposed feature-level supervision. Ped., Cyc., and O.a. are short of Pedestrian, Cyclist, and Overall." + ], + "table_footnote": [], + "table_body": "
MethodSup.Ped.Cyc.CarO.a.Δ
SumX64.8972.9485.5574.46-
✓67.0374.6185.1175.59+1.13
ConcatX65.9072.8083.3774.02-
✓66.2374.9585.4275.52+1.50
OursX63.1872.8485.5873.86-
✓66.0676.1885.4775.90+2.04
", + "bbox": [ + 80, + 258, + 462, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 76, + 405, + 215, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall Performances. The 3D mAP@R40 comparison based on three detectors is shown in Tab. 1 for three classes and overall performances for each difficulty split. We can clearly observe that the LiDAR-Camera approaches (LC) are superior to the LiDAR-based detectors (L) by introducing extra camera inputs. By introducing polar sampling, the assistant model $(\\mathrm{L}^{\\dagger})$ could gain admirable performances (e.g. over $90\\%$ mAP) on the enhanced validation set. With the auxiliary supervision from high-quality features and our proposed deep fusion module, our proposal could contiguously boost detection accuracy. For instance, compared with the baseline (LC) model, our proposal could gain $+1.54\\%$ and $+1.24\\%$ 3D mAP improvements for the moderate and hard objects on average for the three detectors. Further, we also conduct experiments on nuScenes benchmark based on SECOND-LC as shown in Tab. 2, which gain $+2.01\\%$ and $+1.38\\%$ improvements for NDS and mAP. Also, we will prune the SupFusion training strategy and deep fusion module in Sec. 4.3 to detail discuss their effectiveness.", + "bbox": [ + 76, + 430, + 468, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Class-Aware Improvements Analysis. Compared with the baseline model, our SupFusion and deep fusion could boost the detection performances not only for the overall performances but also for each class including Pedestrian, Cyclist, and Car. Comparing the average improvements (e.g. moderate cases) for three classes, we can obtain the following observation: the cyclist gains the largest improvements $(+2.41\\%)$ while pedestrian and car gains $+1.35\\%$ and $+0.86\\%$ increasements respectively. The reason is obvious: (1) The car is easier to detect and gains the best results compared with pedestrians and cyclists, hence it is more difficult to improve. (2) Cyclist gains more improvements against", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1e9e49c9855715b01d2c53c18c5e074222e0c6898dc94875c76add60653d0a9e.jpg", + "table_caption": [ + "Table 4: Results comparison with different supervision loss. $\\mathcal{L}_{Sim}$ indicates the auxiliary supervision loss in Eq. 1." + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{Sim} \\)Ped.Cyc.CarO.a.\\( \\Delta \\)
None64.8972.9485.5574.46-
L165.7675.5785.4275.58+1.12
L266.0676.1885.4775.90+1.44
", + "bbox": [ + 522, + 109, + 864, + 179 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/267b20592852676054f3940296544a1c92aecf0a12da24ce9dba44ffe8ffeeaf.jpg", + "table_caption": [ + "Table 5: Results with different depth $K$ in deep fusion." + ], + "table_footnote": [], + "table_body": "
K1234
LC74.2374.0073.8673.22
LC*75.6475.7575.9075.88
", + "bbox": [ + 521, + 215, + 867, + 268 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cb2892380ab029889e33aa1f3765ff2b93ce702cadd0be5ea33825c666f25280.jpg", + "table_caption": [ + "Table 6: Results with different $N$ in polar sampling." + ], + "table_footnote": [], + "table_body": "
K14816
L†99.8099.6599.6699.43
LC*75.6875.7275.9075.85
", + "bbox": [ + 521, + 304, + 867, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pedestrian because pedestrians are non-grid, the generated dense object is not so well compared with Cyclist, hence gains lower performance improvements.", + "bbox": [ + 498, + 378, + 890, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 436, + 653, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To analyze the effectiveness of our proposed SupFusion and deep fusion module, we conduct ablation experiments based on Voxel-RCNN-LC on the KITTI benchmark.", + "bbox": [ + 498, + 460, + 890, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Supervision. In SupFusion training strategy, we mainly introduce a high-quality guided feature-level supervision and simply employ L2 loss as the auxiliary loss. Here we also adopt L1 loss and the results are shown in Tab. 4, from which we can observe that L1 loss gains comparable but slight improvements (e.g. $+1.12\\%$ versus $+1.44\\%$ ). To mimic the high-quality features by the fusion features, utilizing L1 and L2 to measure the feature distance is acceptable [13, 44].", + "bbox": [ + 496, + 506, + 890, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fusion Strategy. We conduct experiments with different fusion methods, including the concatenation and summation approaches employed in MVXNet and PV-RCNN-LC/Voxel-RCNN-LC, as well as our proposed deep fusion methods. As shown in Tab. 3, without auxiliary supervision in SupFusion, our proposed deep fusion and concatenation fusion gain comparable but lower performances, demonstrating introducing extra parameters in the fusion model would lead the model difficult to train and hard to optimize. By introducing auxiliary supervision, our proposal could gain the best performance and improvements compared with the baseline model. This demonstrates that the SupFusion could effectively promote the learning process, guiding the 3D/2D backbone and fusion model extracts meaningful robust fusion features from LiDAR and camera inputs. Furthermore, we also manually set variance depth in deep fusion, to rescale the fusion module size. The results are shown in Tab. 5, in which we can observe that by increasing the fusion depth,", + "bbox": [ + 496, + 628, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22020", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddabcb16c12cd858ea94c87677e1fd75f2ee47e590554831cf15cbaecb48c401.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 89, + 285, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0600ecff488cd288295c93125f848ca421168e636c92ced0f71a5107d651563b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 90, + 486, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16f895916f1513043ab5b173c7bc1b780a0994b19d9d84e417d8fc722585e1a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 90, + 883, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73b5d7c4475322f8b7f7388f041a91dffebf28bed7f1c8948de7bac5df24fa65.jpg", + "image_caption": [ + "Figure 5: Visualization results. The top row shows the raw sparse LiDAR data and the bottom row displays the enhanced data. Particularly, the points in magenta are the added points and row points are shown in black. We can densify the sparse objects with few points via the dense objects in database while keeping the density distributions." + ], + "image_footnote": [], + "bbox": [ + 84, + 244, + 285, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0330759fde550c5e00492c52f3f103d331a9cf1584f0b84f5104ac781ffc4fc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 244, + 486, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e008fd4c23467c92c78a7d036dfcf52d485497a58984ff8affc858d603433d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 244, + 883, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the baseline gets worse while our proposal gains better performances, demonstrating the effectiveness of our proposed SupFusion training strategy and deep fusion module again.", + "bbox": [ + 75, + 484, + 470, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Polar Sampling. In the proposed polar sampling data enhancement, we split the objects into $N \\times N$ groups according to their directions and rotations. To explore the influence of the split number $N$ , we manually set $N = \\{1, 4, 8, 16\\}$ to conduct experiments, and the results are shown as Tab. 6, from which we can observe that the split number $N$ in polar sampling affect the performances of assistant and our model in an extent. For instance, when setting $N = 1$ , we only generate one object for each class, the object is complete but destroys the density distribution, which boosts the assistant model performances but decreases the improvements in an extent (e.g. $+1.22\\%$ versus $+1.44\\%$ ), when setting $N$ larger, the objects in each object set may get fewer, because we split all the objects in the dataset into more groups, which may influence the quality of the dense object in the database in the polar grouping.", + "bbox": [ + 75, + 534, + 470, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assistant Model. We utilize the enhanced dense LiDAR data to generate high-quality features. By densifying the sparse objects, the assistant model with enhanced data could avoid detection failers easily and provide high-quality features. To better apply to the given LiDAR-Camera detector and easier implement, we simply adopt the LiDAR branch detector as the assistant model. Here, we also conduct experiments to keep the assistant model sharing the same archi-", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9f7ce5df54cd48b55dd60a29557100dd7bd8514b140ee6047667249cb1b5cb07.jpg", + "table_caption": [ + "Table 7: Results with different teacher models in SupFusion. LC* and LC** indicate utilize LiDAR and LiDAR-Camera based assistant model." + ], + "table_footnote": [], + "table_body": "
ModelPed.Cyc.CarO.a.Δ
LC64.8972.9485.5574.46-
LC*66.0676.1885.4775.90+1.44
LC**66.3975.7185.4775.86+1.40
", + "bbox": [ + 519, + 529, + 870, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tecture with the given detector, which is a LiDAR-Camera detector. The results are shown as Tab. 7, from which we can observe that utilizing LiDAR-base and LiDAR-Camera based detectors as the assistant model, the SupFusion gains comparable performances (e.g. LC* versus LC**).", + "bbox": [ + 496, + 627, + 893, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 710, + 683, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we visualize the row sparse LiDAR data and enhanced dense LiDAR in Fig. 5. In the left 2 columns, we can observe the added points in magenta could effectively densify the sparse points in black points for all classes including pedestrians, cyclists, and cars. In the right column, we can find that the polar sampling densifies the sparse objects with dense ones while keeping the objects' density distribution. As we could see, enhanced point cloud data have high integrity and obvious class representation which could be utilized to generate high-quality features as the fusion supervision. To better demonstrate the quality of feature-level", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22021", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "supervision, we visualize the feature of assistant, baseline, and our model in Appendix.", + "bbox": [ + 80, + 90, + 467, + 119 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5. Information leak", + "text_level": 1, + "bbox": [ + 80, + 130, + 240, + 145 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "There is no information leak if we are not misunderstanding. In practice, we build the database in Fig. 4 only using the training set in the polar grouping phase, the validation set is not used in polar sampling, as well as the model training process.", + "bbox": [ + 80, + 154, + 467, + 228 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 80, + 242, + 192, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a novel training strategy named SupFusion, to introduce auxiliary feature-level supervision for effective LiDAR-Camera fusion, and further boost 3D detection performances. Specifically, we introduce high-quality features to the LiDAR-Camera fusion-based 3D detector, which is used to provide auxiliary supervision for the 3D/2D feature extraction and multi-modal feature fusion. To achieve this, we propose a novel data enhancement method named polar sampling to densify the sparse objects, which is utilized to generate high-quality features via an assistant model. Furthermore, we proposed a simple yet effective deep fusion module, which leverages meaningful information from camera input and contiguously boosts the LiDAR-Camera fusion-based 3D detectors. This is the first approach to introducing feature-level supervision for LiDAR-Camera 3D detection and we conduct extensive experiments to demonstrate its effectiveness.", + "bbox": [ + 80, + 268, + 467, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledge", + "text_level": 1, + "bbox": [ + 80, + 537, + 210, + 555 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work is partially supported by the Young Scientists Fund of the National Natural Science Foundation of China under grant No. 62106154, by the Natural Science Foundation of Guangdong Province, China (General Program) under grant No.2022A1515011524, and by Shenzhen Science and Technology Program JCYJ20220818103001002 and by Shenzhen Science and Technology Program ZDSYS20211021111415025.", + "bbox": [ + 80, + 563, + 467, + 683 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22022", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In CVPR, pages 1090-1099, 2022.", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020.", + "[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, pages 1907-1915, 2017.", + "[4] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In ICCV, pages 9775-9784, 2019.", + "[5] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. arXiv preprint arXiv:2211.09386, 2022.", + "[6] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. arXiv preprint arXiv:2201.10830, 2022.", + "[7] Xing Dai, Zeren Jiang, Zhao Wu, Yiping Bao, Zhicheng Wang, Si Liu, and Erjin Zhou. General instance distillation for object detection. In CVPR, pages 7842-7851, 2021.", + "[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021.", + "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012.", + "[10] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018.", + "[11] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu, and Chang Xu. Distilling object detectors via decoupled features. In CVPR, pages 2154-2164, 2021.", + "[12] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In ICCV, pages 3153-3163, 2021.", + "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022.", + "[14] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Epnet: Enhancing point features with image semantics for 3d object detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 35-52. Springer, 2020.", + "[15] Bo Ju, Zhikang Zou, Xiaoqing Ye, Minyue Jiang, Xiao Tan, Errui Ding, and Jingdong Wang. Paint and distill: Boosting" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d object detection with semantic passing network. In ACMM, pages 5639-5648, 2022.", + "[16] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018.", + "[17] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In CVPR, pages 8973-8983, 2021.", + "[18] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019.", + "[19] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022.", + "[20] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In CVPR, pages 1120-1129, 2022.", + "[21] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V Le, et al. Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In CVPR, pages 17182-17191, 2022.", + "[22] Zhijian Liu, Haotian Tang, Alexander Amini, Xingyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In IEEE International Conference on Robotics and Automation (ICRA), 2023.", + "[23] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In ICCV, pages 15641-15650, 2021.", + "[24] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In ICCV, pages 3111-3121, 2021.", + "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017.", + "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017.", + "[27] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In CVPR, pages 8555-8564, 2021.", + "[28] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In CVPR, pages 9891-9901, 2022.", + "[29] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: Realtime and high-performance pillar-based 3d object detection. In ECCV, pages 35-52. Springer, 2022.", + "[30] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22023", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "voxel feature set abstraction for 3d object detection. In CVPR, pages 10529-10538, 2020.", + "[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointcnn: 3d object proposal generation and detection from point cloud. In CVPR, pages 770-779, 2019.", + "[32] Vishwanath A Sindagi, Yin Zhou, and Oncel Tuzel. Mvx-net: Multimodal voxelnet for 3d object detection. In 2019 International Conference on Robotics and Automation (ICRA), pages 7276-7282. IEEE, 2019.", + "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020.", + "[34] Kaicheng Yu Zhongyu Xia Zhiwei Lin Yongtao Wang Tao Tang Bing Wang Tingting Liang, Hongwei Xie and Zhi Tang. BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework. In Neural Information Processing Systems (NeurIPS), 2022.", + "[35] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In CVPR, pages 11794-11803, 2021.", + "[36] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3d object detection. In CVPR, pages 454-463, 2021.", + "[37] Li Wang, Li Zhang, Yi Zhu, Zhi Zhang, Tong He, Mu Li, and Xiangyang Xue. Progressive coordinate transforms for monocular 3d object detection. Advances in Neural Information Processing Systems, 34:13364-13377, 2021.", + "[38] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022.", + "[39] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, pages 4933-4942, 2019.", + "[40] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In ICCV, pages 913-922, 2021.", + "[41] Yue Wang, Alireza Fathi, Abhijit Kundu, David A Ross, Caroline Pantofaru, Tom Funkhouser, and Justin Solomon. Pillar-based object detection for autonomous driving. In ECCV, pages 18-34. Springer, 2020.", + "[42] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022.", + "[43] Xiaopei Wu, Liang Peng, Honghui Yang, Liang Xie, Chenxi Huang, Chengqi Deng, Haifeng Liu, and Deng Cai. Sparse fuse dense: Towards high quality 3d detection with depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5418-5427, 2022.", + "[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "framework for masked image modeling. In CVPR, pages 9653-9663, 2022.", + "[45] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018.", + "[46] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In ICCV, pages 1951-1960, 2019.", + "[47] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021.", + "[48] Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In ECCV, pages 720-736. Springer, 2020.", + "[49] Linfeng Zhang and Kaisheng Ma. Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In ICLR, 2021.", + "[50] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In CVPR, pages 18953-18962, 2022.", + "[51] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In CVPR, pages 3289-3298, 2021.", + "[52] Wu Zheng, Mingxuan Hong, Li Jiang, and Chi-Wing Fu. Boosting 3d object detection by simulating multimodality on point clouds. In CVPR, pages 13638-13647, 2022.", + "[53] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3d object detection: An extrinsic parameter free approach. In CVPR, pages 7556-7566, 2021.", + "[54] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, pages 4490-4499, 2018." + ], + "bbox": [ + 501, + 92, + 893, + 574 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22024", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/92e5510f-e140-45f8-8597-b810b193a66f_model.json b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/92e5510f-e140-45f8-8597-b810b193a66f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e8ec1392021682ef033ddaa0d2ce891d2ebe557b --- /dev/null +++ b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/92e5510f-e140-45f8-8597-b810b193a66f_model.json @@ -0,0 +1,2125 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "SupFusion: Supervised LiDAR-Camera Fusion for 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.18, + 0.836, + 0.2 + ], + "angle": 0, + "content": "Yiran Qin\\(^{1*}\\), Chaoqun Wang\\(^{1*}\\), Zijian Kang\\(^{2}\\), Ningning Ma\\(^{2}\\), Zhen Li\\(^{3}\\), Ruimao Zhang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.199, + 0.749, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Data Science, Shenzhen Research Institute of Big Data," + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.217, + 0.812, + 0.234 + ], + "angle": 0, + "content": "The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen), China \\(^{2}\\)NIO" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.234, + 0.813, + 0.252 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Science and Engineering, Future Intelligent Network Research Institute," + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.253, + 0.789, + 0.27 + ], + "angle": 0, + "content": "The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen), China" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.273, + 0.882, + 0.288 + ], + "angle": 0, + "content": "{yiranqin@link, chaoqunwang@link, lizhen@, ruimaozhang@}cuhk.edu.hk, {zijiankang, ningningma}@nio.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.366, + 0.474, + 0.807 + ], + "angle": 0, + "content": "LiDAR-Camera fusion-based 3D detection is a critical task for automatic driving. In recent years, many LiDAR-Camera fusion approaches sprung up and gained promising performances compared with single-modal detectors, but always lack carefully designed and effective supervision for the fusion process. In this paper, we propose a novel training strategy called SupFusion, which provides an auxiliary feature level supervision for effective LiDAR-Camera fusion and significantly boosts detection performance. Our strategy involves a data enhancement method named Polar Sampling, which densifies sparse objects and trains an assistant model to generate high-quality features as the supervision. These features are then used to train the LiDAR-Camera fusion model, where the fusion feature is optimized to simulate the generated high-quality features. Furthermore, we propose a simple yet effective deep fusion module, which contiguously gains superior performance compared with previous fusion methods with SupFusion strategy. In such a manner, our proposal shares the following advantages. Firstly, SupFusion introduces auxiliary feature-level supervision which could boost LiDAR-Camera detection performance without introducing extra inference costs. Secondly, the proposed deep fusion could continuously improve the detector's abilities. Our proposed SupFusion and deep fusion module is plug-and-play, we make extensive experiments to demonstrate its effectiveness. Specifically, we gain around \\(2\\%\\) 3D mAP improvements on KITTI benchmark based on multiple LiDAR-Camera 3D detectors. Our code is available at https://github.com/IranQin/SupFusion." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.324, + 0.889, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.432, + 0.889, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.612, + 0.896, + 0.675 + ], + "angle": 0, + "content": "Figure 1: Top: The previous LiDAR-Camera 3D detector, the fusion module is optimized by detection loss. Bottom: Our proposed SupFusion, we introduce auxiliary supervision via high-quality features provided by an assistant model." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.706, + 0.633, + 0.722 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.896, + 0.903 + ], + "angle": 0, + "content": "LiDAR-Camera fusion-based 3D object detection is a critical and challenging task for autonomous driving and robotics [32, 35]. Previous approaches [20, 22, 34, 19] always project the camera inputs to LiDAR BEV or voxel space via intrinsic and extrinsic to align the LiDAR and camera features. Then, simple concatenation or summation is employed to obtain the fusion features for final detection. Also, some deep learning-based fusion approaches [1, 21] sprung up and gained promising performances. However, the previous fusion approaches always directly optimized the 3D/2D feature extraction as well as the fusion module by the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.418, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution. Work done during an internship at NIO." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.418, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22014" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "detection loss, which lacks carefully designed and effective supervision at the feature level and limits its performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.317 + ], + "angle": 0, + "content": "In recent years, distillation manners have shown great improvements in feature-level supervision for 3D detection. For instance, some approaches [5, 6] provide LiDAR features to guide the 2D backbone in estimating the depth information from camera inputs. Also, some approaches [19] provide LiDAR-Camera fusion features to supervise LiDAR backbone learns global and contextual presentation from LiDAR inputs. Introducing feature-level auxiliary supervision via simulating more robust and high-quality features, the detectors could boost marginal improvements. Motivated by this, a natural solution to handle LiDAR-Camera feature fusion is providing stronger and high-quality features and introducing auxiliary supervision for LiDAR-Camera 3D detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.77 + ], + "angle": 0, + "content": "To this end, in this paper we propose a supervised LiDAR-Camera fusion method named SupFusion, to generate high-quality features and provide effective supervision for the fusion as well as feature extraction process, and further boost the LiDAR-Camera fusion-based 3D detection performances. Specifically, we first train an assistant model to provide high-quality features. To achieve it, different from previous approaches, which leverage larger models or extra data, we proposed a novel data enhancement method named Polar Sampling. The polar sampling could dynamically enhance the object's density from sparse LiDAR data, which is easier to detect and improve features quality, e.g., the feature could conclude accuracy detection results. Then, we simply train the LiDAR-Camera fusion-based detector with introduced auxiliary feature-level supervision. In this step, we feed the raw LiDAR and camera inputs into 3D/2D backbones and fusion modules to obtain fusion features. In one aspect, the fusion features are fed into the detection head for final prediction, which is decision-level supervision. In another aspect, the auxiliary supervision mimics the fusion feature to high-quality features, which are obtained via the pretrained assistant model and enhanced LiDAR data. In such a manner, the proposed feature-level supervision could lead the fusion module to generate more robust features, and further boost the detection performance. To better fuse the LiDAR and camera features, we propose a simple yet effective deep fusion module, which consists of stacked MLP blocks and dynamic fusion blocks. SupFusion could sufficiently excavate the deep fusion module capacity, and continuously improve detection accuracy." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.771, + 0.472, + 0.847 + ], + "angle": 0, + "content": "Compared with previous LiDAR-Camera 3D detectors, this is a brand new attempt to improve the effectiveness of the LiDAR-Camera fusion by introducing feature-level supervision, which consistently boosts 3D detection accuracy. The main contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "- We propose a novel supervised fusion training strategy named SupFusion, which mainly consists of a high-quality feature generation process and firstly propose" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "auxiliary feature-level supervision loss for robust fusion feature extraction and accuracy 3D detection to our knowledge." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.147, + 0.894, + 0.224 + ], + "angle": 0, + "content": "- To obtain high-quality features in SupFusion, we propose a data enhancement method named Polar Sampling to densify the sparse objects. Furthermore, we propose an efficient deep fusion module to contiguously boost detection accuracy." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.232, + 0.895, + 0.293 + ], + "angle": 0, + "content": "- We make extensive experiments based on multiple detectors with different fusion strategies, and gain around \\(2\\%\\) mAP improvements on KITTI benchmark. The source code will be released after the blind review." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.147, + 0.895, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.307, + 0.642, + 0.323 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.693, + 0.348 + ], + "angle": 0, + "content": "2.1. 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.895, + 0.505 + ], + "angle": 0, + "content": "Camera-based 3D Object Detection methods perform 3D detection on single- or multi-view images. Many methods are proposed to address monocular 3D detection[24, 23, 17, 51, 53, 27, 37, 36] since KITTI benchmark[9] only contains one from the camera. With the development of autonomous driving datasets which have more cameras covering a wider range of angles(nuScenes[2] and Waymo[33]), more and more methods[38, 40, 42] are developed that take multi-view images as input and get surprising performance compared to monocular methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.895, + 0.732 + ], + "angle": 0, + "content": "LiDAR-based 3D Object Detection methods perform 3D detection on a point cloud scene. Depending on the three kinds of point cloud data pretreatments, 3D detectors are also classified into three mainstream. Point-based method [31, 46, 50, 4] use row point cloud data, leveraging pointnet[25] and pointnet++[26] as 3D backbone networks to get point-wised features which are sent to point-wise 3D detection head. Pillar-based method [18, 41, 29] pretreat point cloud to pillars, gain speed up while still maintaining a good performance. Voxel-based methods[54, 45, 8] are widely used, point clouds are voxelized to voxel and sent to a 3D sparse convolutional[10] based 3D backbone to get bird's-eye-view features. Some methods[30] also try to gain more improvement by using both the voxel feature and the point feature." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.82, + 0.758 + ], + "angle": 0, + "content": "2.2. LiDAR-Camera 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "LiDAR-camera fusion is a technique that combines data from LiDAR and camera sensors to obtain a more complete understanding of a scene in 3D. Due to the complementary nature of the features produced by these two modalities, researchers have developed methods that can be jointly optimized on both sensors. These methods can be classified into two categories based on their fusion mechanism. The first category is point-level fusion[14, 35, 47], where the image features are queried via the raw LiDAR points and then" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22015" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.09, + 0.887, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.896, + 0.391 + ], + "angle": 0, + "content": "Figure 2: The detector training procedure. For a given LiDAR-Camera sample, we extract the feature presentation via the corresponding 3D/2D encoders, and the proposed deep fusion module is employed to fuse the multi-modal features. Besides the detection loss \\(\\mathcal{L}_{Det}\\), we introduce the auxiliary feature-level supervision \\(\\mathcal{L}_{Sim}\\), to mimic the fusion feature \\(\\mathbf{f}_i^{lc}\\) to the high-quality features \\(\\mathbf{f}_l^*\\), which is generated by the assistant model and enhanced data. Here the \\(\\odot\\), \\(\\oplus\\), and \\(\\odot\\) indicate the element-wise multiplication, summation, and channel-wise concatenation. More details could be referred in Sec. 3.1" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.473, + 0.629 + ], + "angle": 0, + "content": "concatenated back as additional point features. The second category is feature-level fusion, where the LiDAR points are projected into a specific feature space[48, 20, 19] or proposals level feature fusion[3, 16, 1, 43]. There are also methods that combine these two fusion manners, such as MVX-Net [32]. Recently, attention-based fusion methods have sprung up. TransFusion[1] uses the bounding box prediction of LiDAR features as a proposal to query the image feature, and then a Transformer-like architecture fuses the information back to LiDAR features. Similarly, DeepFusion[21] projects LiDAR features on each view image as queries and leverages cross-attention for the two modalities. Bird's-eye-view space feature fusion[22, 34] is proposed recently and gained great performance with different tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.471, + 0.654 + ], + "angle": 0, + "content": "2.3. Knowledge Distillation for 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Object detection tasks differ from classification tasks in that they require consideration of both global and local features. While numerous distillation methods have been proposed to address this challenge, such as those developed by [39, 7, 11, 49], many of these methods require the creation of specific distillation regions or objects based on ground-truth data, bounding box predictions, attention, or gradients to mitigate the imbalance between foreground and background. In addition to 2D detection, there has been growing interest in cross-model knowledge distillation methods between RGB and LiDAR in 3D object detection. For example, [12, 28, 6, 5] have developed methods that transfer knowledge from LiDAR detectors to image detectors. [52, 15] has trained a teacher model similar to that proposed by [35] and developed a multilevel distillation framework to transfer semantic segmentation information to a student model. In this" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.417, + 0.892, + 0.448 + ], + "angle": 0, + "content": "work, we want to further boost LiDAR-Camera detectors with an auxiliary feature level supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.463, + 0.637, + 0.481 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.895, + 0.549 + ], + "angle": 0, + "content": "In this section, we first overview the overall SupFusion pipeline in Sec. 3.1, and then detail introduces the main components, including the proposed Polar Sampling in Sec. 3.2, and deep fusion module in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.747, + 0.577 + ], + "angle": 0, + "content": "3.1. Overall SupFusion Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For a given LiDAR-Camera fusion-based 3D detector \\(\\mathbf{D}\\), which consists of a 3D Encoder \\(\\mathbf{E}_l\\), a 2D Encoder \\(\\mathbf{E}_c\\), a fusion module \\(\\mathbf{F}\\), and a detection head. Denotes \\(\\mathcal{D} = \\{\\left(\\mathbf{l}_1,\\mathbf{c}_1\\right),\\left(\\mathbf{l}_2,\\mathbf{c}_2\\right),\\dots,\\left(\\mathbf{l}_N,\\mathbf{c}_N\\right)\\}\\) and \\(\\mathcal{V} = \\{\\mathbf{y}_1,\\mathbf{y}_2,\\dots,\\mathbf{y}_N\\}\\) are the given dataset with \\(N\\) samples, where \\((\\mathbf{l}_i,\\mathbf{c}_i)\\) are the paired LiDAR and camera inputs. The previous approaches feed the paired \\((\\mathbf{l}_i,\\mathbf{c}_i)\\) into 3D/2D encoders \\(\\mathbf{E}_l\\) and \\(\\mathbf{E}_c\\) to capture corresponding feature presentation \\(\\mathbf{f}_i^l\\) and \\(\\mathbf{f}_i^c\\). Then, the paired features presentation is fed into the fusion module to obtain the fusion feature \\(\\mathbf{f}_i^{lc}\\), which is used for the final prediction. To better leverage the LiDAR and camera data, we propose a supervised LiDAR-Camera Fusion training strategy named SupFusion, which provides high-quality features for feature-level supervision. To end this, different from previous approaches, which adopt larger scale models or extra data input, we propose a data enhancement method named polar sampling to densify the sparse objects in LiDAR data. In such a manner, the data is easier and could easily generate high-quality features (e.g., the feature could conclude accuracy detection results) with lightweight models. The proposed polar sampling mainly" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22016" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.887, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.378, + 0.893, + 0.44 + ], + "angle": 0, + "content": "Figure 3: High-quality feature generation process. For any given LiDAR sample, we densify the sparse objects via polar pasting, which calculate the direction and rotation to query dense object from the database, and pasting to add extra points to sparse objects. We first train the assistant model via the enhanced data and feed the enhanced LiDAR data into the assistant model to generate high-quality features \\(\\mathbf{f}^*\\) after it is convergent. More details could be referred in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.473, + 0.604 + ], + "angle": 0, + "content": "densifies the sparse object points in LiDAR data. Specifically, for any sample \\((\\mathbf{l}_i,\\mathbf{c}_i)\\), where \\(\\mathbf{l}_i = \\{l_i^1,l_i^2,\\dots,l_i^{n_i},l_i^s\\}\\) contains \\(n_i\\) objects points sets and one background points set. In the polar sampling, we add extra points to each object point set from raw LiDAR data, to obtain enhanced LiDAR data \\(\\mathbf{l}_i' = \\{(l_i^1,\\hat{l}_i^1),(l_i^2,\\hat{l}_i^2),\\dots,(l_i^{n_i},\\hat{l}_i^{n_i}),l_i^s\\}\\) in polar pasting, where \\(\\hat{l}_i^k\\) is the added points set for the \\(k\\)-th objects in \\(i\\)-th LiDAR sample and generated in polar grouping. We will detail introduce the polar sampling in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.472, + 0.835 + ], + "angle": 0, + "content": "High-quality Feature Generation. To provide feature-level supervision in our SupFusion, we adopt an assistant model to capture high-quality features from the enhanced data as shown in Fig. 3. Firstly, we train an assistant model to provide high-quality features. Specifically, for any sample in \\(\\mathcal{D}\\), we enhance the sparse LiDAR data \\(\\mathbf{l}_i\\) to obtain enhanced data \\(\\mathbf{l}_i^{\\prime}\\) via the polar pasting, which densifies the spare objects with added points set generated in polar grouping. Then, after the assistant model convergence (often with few epochs), we feed the enhanced sample \\(\\mathbf{l}_i^{\\prime}\\) into the optimized assistant model, to capture high-quality features \\(\\mathbf{f}_i^*\\), which is used to train the LiDAR-Camera 3D detector. To better apply to the given LiDAR-Camera detector and easier implement, we simply adopt the LiDAR branch detector as the assistant model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Detector Training. For any given LiDAR-Camera detector, we train the model with the proposed auxiliary supervision at the feature level. Specifically, given sample \\((\\mathbf{l}_i,\\mathbf{c}_i)\\in \\mathcal{D}\\), we first feed the LiDAR and camera into the 3D" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.895, + 0.585 + ], + "angle": 0, + "content": "and 2D Encoders \\(\\mathbf{E}_l\\) and \\(\\mathbf{E}_s\\) to capture the corresponding features \\(\\mathbf{f}_i^l\\) and \\(\\mathbf{f}_i^c\\), which are fed into the fusion model to generate fusion features \\(\\mathbf{f}_i^{l,c}\\), and flow into the detection head for final prediction. Furthermore, the proposed auxiliary supervision is employed to mimic the fusion feature with high-quality features, which is generated by the pre-trained assistant model and enhanced LiDAR data. The above process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.598, + 0.893, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {((\\mathbf {l} _ {i}, \\mathbf {c} _ {i}), \\mathbf {y} _ {i}) \\sim \\mathcal {D}} \\mathcal {L} _ {\\text {d e t}} (\\mathbf {D} (\\mathbf {l} _ {i}, \\mathbf {c} _ {i}, \\theta), \\mathbf {y} _ {i}) (1) \\\\ + \\lambda \\mathcal {L} _ {\\text {S i m}} \\left(\\mathbf {f} _ {i} ^ {l c}, \\mathbf {f} _ {i} ^ {*}\\right), (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.648, + 0.893, + 0.665 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad \\mathbf {f} _ {i} ^ {l c} = \\mathbf {F} \\left(\\mathbf {E} _ {l} \\left(\\mathbf {l} _ {i}, \\theta_ {l}\\right), \\mathbf {E} _ {c} \\left(\\mathbf {c} _ {i}, \\theta_ {c}\\right), \\theta_ {f}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.667, + 0.892, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {i} ^ {*} = \\mathbf {E} _ {l} \\left(\\mathbf {I} _ {i} ^ {\\prime}, \\theta_ {l} ^ {*}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.687, + 0.892, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S i m} = \\left\\| \\psi \\left(\\mathbf {f} _ {i} ^ {l c}\\right) - \\mathbf {f} _ {i} ^ {*} \\right\\| _ {2} ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.893, + 0.792 + ], + "angle": 0, + "content": "where \\(\\theta^{*}\\) is the optimized parameters of detector \\(\\mathbf{D}\\), \\(\\mathbf{f}_i^*\\) is the high-quality features, which is generated from pre-trained assistant model LiDAR encoder \\(\\mathbf{E}_l\\) with parameters \\(\\theta_l^*\\), \\(\\psi\\) is a simple convolutional layer with kernel size 1. \\(\\lambda\\) is the balanced weight and we set it as 1.0." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.658, + 0.819 + ], + "angle": 0, + "content": "3.2. Polar Sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To provide high-quality features via the assistant model in our proposed SupFusion, we propose a novel data enhancement method named Polar Sampling, to ease the sparse problem which frequently leads to the detection failers. To end this, we densify the sparse objects in LiDAR data with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22017" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "a similar dense object. Specifically, the polar sampling consists of two parts, including polar grouping and polar pasting. In polar grouping, we mainly build a database to store the dense objects, which is used in polar pasting to densify the sparse objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.474, + 0.454 + ], + "angle": 0, + "content": "Considering the characteristics of LiDAR sensors, the collected point cloud data naturally exists a particular density distribution, e.g., the objects have more points on the surface towards the LiDAR sensor, while few points on the opposite sides. The density distribution is mainly influenced by the direction and rotation, while the points' density mainly depends on the distances, e.g. the objects closer to the LiDAR sensor have denser points. Motivated by this, we aim to densify the sparse objects in longer distances with the dense objects in short distances, according to their direction and rotation to keep the density distributions. Specifically, we build polar coordinate systems for the whole scene as well as the objects based on the center of the scene and specific objects and define the positive direction of the LiDAR sensor as 0 degrees to measure the corresponding directions and rotations. Then, we collect the objects with similar density distribution (e.g. with similar direction and rotation) and generate one dense object for each group in the polar grouping, and used it to densify the sparse objects in polar pasting." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.455, + 0.471, + 0.514 + ], + "angle": 0, + "content": "Polar Grouping. As shown in Fig. 4, we build a database \\(\\mathcal{B}\\) to store the generated dense object point set \\(\\hat{l}\\) according to their directions and rotations in the polar grouping, which is noted as \\(\\alpha\\) and \\(\\beta\\) in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.621 + ], + "angle": 0, + "content": "Firstly, we search the whole dataset and calculated their polar angles for all objects via the locations and provided rotations in the benchmarks. Secondly, we split the objects into groups according to their polar angles. Specifically, we manually divide the direction and rotation into \\( N \\) groups. For any object point set \\( l \\), we could put it into the corresponding group according to the index:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.628, + 0.47, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {l} ^ {\\prime} = \\left\\lfloor \\frac {\\alpha_ {l} \\cdot 2 \\pi}{N} \\right\\rfloor , \\quad \\beta_ {l} ^ {\\prime} = \\left\\lfloor \\frac {\\beta_ {l} \\cdot 2 \\pi}{N} \\right\\rfloor , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.47, + 0.825 + ], + "angle": 0, + "content": "where \\(\\alpha_{l}^{\\prime},\\beta_{l}^{\\prime}\\in \\{1,2,\\dots,N\\}\\) are the direction and rotation index of object \\(l\\) , and we could put the points set into \\(\\mathcal{B}\\) as: \\(\\mathcal{B}_C^{\\alpha_l',\\beta_l'}\\gets l\\) , where \\(C\\) is the label of object \\(l\\) . Finally, for each object group \\(\\mathcal{B}_C^{\\alpha ',\\beta '} = \\{l_1,l_2,\\ldots ,l_n\\}\\) contains \\(n\\) similar objects, we select the densest \\(k\\) objects and mix up to generate one dense and complete object \\(\\hat{l}\\) , and \\(\\mathcal{B}_C^{\\alpha ',\\beta '} = \\{\\hat{l}\\}\\) is used in polar pasting phase. To keep the density and avoid too much calculation in the model, we randomly select partial points in the dense object. Here we set \\(k = 10\\) and select 5000 points for all objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Polar Pasting. As shown in Fig. 2, the polar pasting is utilized to enhance the sparse LiDAR data to train the assistant model and generate high-quality features. Concretely, given a LiDAR sample \\(\\mathbf{l}_i = \\{l_i^1, l_i^2, \\dots, l_i^{n_i}, l_i^s\\}\\) contains \\(n_i\\) objects, for any object \\(l_i^k\\), we can calculate the direction" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.887, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 4: Polar grouping mainly builds a database to store the dense objects of each group in polar coordinate systems of direction and rotations for each class. More details could be referred to in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.895, + 0.456 + ], + "angle": 0, + "content": "and rotation the same as the grouping process and query the dense object \\(\\hat{l}_i^k\\) from \\(\\mathcal{B}\\) according to the label and indexes, which could be obtained from E.q. 6 In such a manner, we enhance all the objects in the sample and obtain the enhanced data: \\(\\mathbf{I}_i' = \\{(l_i^1,\\hat{l}_i^1),(l_i^2,\\hat{l}_i^2),\\dots,(l_i^{n_i},\\hat{l}_i^{n_i}),l_i^s\\}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.464, + 0.634, + 0.48 + ], + "angle": 0, + "content": "3.3. Deep Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.895, + 0.761 + ], + "angle": 0, + "content": "To mimic the high-quality features generated from enhanced LiDAR data, the fusion model aims to extract the missing information of the sparse objects from the rich color and contextual features in camera input. To end this, we propose the deep fusion module, to leverage image features and complete the LiDAR presentation. The proposed deep fusion mainly consists of a 3D learner and a 2D3D learner. The 3D learner is a simple convolutional layer, to transfer the 3D presentation into 2D space. Then, concatenating the 2D features and 3D presentation (e.g. in 2D space), the 2D3D learner is employed to fuse the LiDAR-Camera features. Finally, we weighed the fusion feature via an MLP and an activate function, adding back to the raw LiDAR features as the output of the deep fusion module. The 2D3D learner consists of stacked MLP blocks with depth \\( K \\), learning to leverage camera features to complete the LiDAR representation from the sparse objects, to simulate the high-quality features of dense LiDAR objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.776, + 0.634, + 0.793 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.7, + 0.818 + ], + "angle": 0, + "content": "4.1. Experiments Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. KITTI [9] is a large-scale public benchmark for autonomous driving, including 3712 training, 3769 validation, and 7518 testing samples with LiDAR and camera pair data. All objects in KITTI are divided into 'easy', 'moderate', and 'hard' difficulties ac" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22018" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.166 + ], + "angle": 0, + "content": "Table 1: The experiment results (mAP@R40 %) on KITTI validation set. We list the easy, moderate (mod.), and hard cases for three classes, and the overall performances. Here, L, LC, LC* indicate the results of the corresponding LiDAR detector, LiDAR-Camera fusion detectors, and our proposal. \\(\\Delta\\) indicates the improvements. The best results are shown in bold expect \\(\\mathrm{L}^{\\dagger}\\), which is the assistant model, and test on the enhanced validation set. MVXNet is re-conduct based on mmdetection3d. PV-RCNN-LC and Voxel-RCNN-LC are re-conduct based on the open-source code of VFF [20]." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.169, + 0.902, + 0.516 + ], + "angle": 0, + "content": "
DetectorModalityOverallPedestrianCyclistCar
easymod.hardeasymod.hardeasymod.hardeasymod.hard
MVXNetL†88.4490.9388.5887.3387.6180.1580.0986.0186.1997.8999.1799.41
L73.4658.6455.1557.5450.1045.5674.6351.8148.9288.2173.9970.98
LC74.9963.0059.3264.4657.7653.0572.6955.2751.6587.8275.9873.26
LC*75.3764.7360.5263.1158.1353.7374.7957.6053.8288.2078.4674.01
Δ+0.38+1.73+1.20-1.35+0.37+0.68+2.10+2.33+2.17+0.38+2.48+0.75
PV-RCNN-LCL†94.9095.2095.1086.8686.7887.0798.9099.3398.6298.9599.5099.62
L81.3771.5568.0867.8059.5654.4085.0670.8367.9691.2684.2581.88
LC84.1674.5470.8670.4564.1659.5090.4774.4670.2591.5685.0182.83
LC*85.7776.0072.0074.2966.6761.3891.0076.1371.6692.0285.2082.94
Δ+1.61+1.46+1.14+3.84+2.51+1.88+0.53+1.67+1.41+0.46+0.19+0.11
Voxel-RCNN-LCL†99.2699.6699.7597.8498.0598.1794.3694.6094.8599.9099.9099.91
L82.0171.6968.4067.8460.9556.0386.1071.3967.0592.0982.7482.12
LC84.3674.4670.8071.3964.8960.2089.6472.9469.0392.0585.5583.17
LC*86.1375.9072.1773.9566.0661.8392.0276.1871.8392.4285.4782.83
Δ+1.77+1.44+1.37+2.56+1.17+1.63+2.38+3.24+2.80+0.37-0.08-0.34
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.472, + 0.753 + ], + "angle": 0, + "content": "cording to the object scales and occlusion ratio. We evaluate the detection performance by the mean Average Precision under 40 recall thresholds (mAP@R40) the same as the official benchmark and evaluation for 3 classes including Pedestrian, Cyclist, and Car. nuScenes [2] is a large-scale autonomous driving dataset for 3D detection and tracking tasks, including 700 training, 150 validation, and 150 testing sequences, with each 20s long and annotated with 3D bounding boxes. For 3D object detection, the official evaluation metrics include the mean Average Precision (mAP) and nuScenes detection score (NDS). mAP measures the localization precision using a threshold base on the birds-eye view. NDS is a weighted combination of mAP and regression accuracy of other object attributes, including box size, orientation, and translation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Detectors. To demonstrate the effectiveness of our proposed method, we apply the SupFusion training strategy and deep fusion module on variance LiDAR-Camera detectors including MVXNet [32], PV-RCNN-LC [30], and Voxel-RCNN-LC [8] on KITTI, and SECOND-LC [45] on nuScenes benchmark. MVXNet is a voxel-based LiDAR-Camera 3D detector based on VoxelNet [54]. The camera features project to LiDAR space and concatenates with points features. Voxel-RCNN and PV-RCNN are LiDAR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.895, + 0.679 + ], + "angle": 0, + "content": "based detectors, we re-conduct the LiDAR-Camera fusion approach following VFF [20] based on the open-source code, which is a previous state-of-the-art fusion approach based on the two detectors, mainly projects the camera features into voxel presentation and summarized fusion with the voxel features of LiDAR input. SECOND-LC introduces camera inputs based on the LiDAR-based SECOND detector, the same as BEVFusion [22], we concatenate the camera and LiDAR features in birds-eys-view spaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Implement Details. We conduct experiments on MVXNet, PV-RCNN-LC and Voxel-RCNN-LC, SECOND-LC following the open-sourced code and mmdetection3d on 8 Tesla A100 GPUs. All the architecture, configuration, and optimizer follow the settings reported in papers including voxel size, point cloud range, batch size, and learning rate. We modify the fusion method (e.g. concatenation, summarized) by our proposed deep fusion module and applied our SupFusion training strategy. For the deep fusion module, we set depth \\( K = 3 \\). For the auxiliary supervision, we utilize simple L2 loss and set \\( \\lambda = 1.0 \\). More details about how to apply SupFusion and deep fusion modules on different detectors are demonstrated in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22019" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.09, + 0.468, + 0.104 + ], + "angle": 0, + "content": "Table 2: NDS and mAP(%) results on nuScenes benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.107, + 0.472, + 0.19 + ], + "angle": 0, + "content": "
Modality3D2DNDSMAP
L†SECOND-90.3894.44
LSECOND-60.5252.28
LCSECONDSwin-T62.5755.26
LC*SECONDSwin-T64.58(+2.01)56.64(+1.38)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.212, + 0.47, + 0.257 + ], + "angle": 0, + "content": "Table 3: Results comparison with different fusion methods. Sup. indicates our proposed feature-level supervision. Ped., Cyc., and O.a. are short of Pedestrian, Cyclist, and Overall." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.26, + 0.464, + 0.375 + ], + "angle": 0, + "content": "
MethodSup.Ped.Cyc.CarO.a.Δ
SumX64.8972.9485.5574.46-
✓67.0374.6185.1175.59+1.13
ConcatX65.9072.8083.3774.02-
✓66.2374.9585.4275.52+1.50
OursX63.1872.8485.5873.86-
✓66.0676.1885.4775.90+2.04
" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.406, + 0.216, + 0.421 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Overall Performances. The 3D mAP@R40 comparison based on three detectors is shown in Tab. 1 for three classes and overall performances for each difficulty split. We can clearly observe that the LiDAR-Camera approaches (LC) are superior to the LiDAR-based detectors (L) by introducing extra camera inputs. By introducing polar sampling, the assistant model \\((\\mathrm{L}^{\\dagger})\\) could gain admirable performances (e.g. over \\(90\\%\\) mAP) on the enhanced validation set. With the auxiliary supervision from high-quality features and our proposed deep fusion module, our proposal could contiguously boost detection accuracy. For instance, compared with the baseline (LC) model, our proposal could gain \\(+1.54\\%\\) and \\(+1.24\\%\\) 3D mAP improvements for the moderate and hard objects on average for the three detectors. Further, we also conduct experiments on nuScenes benchmark based on SECOND-LC as shown in Tab. 2, which gain \\(+2.01\\%\\) and \\(+1.38\\%\\) improvements for NDS and mAP. Also, we will prune the SupFusion training strategy and deep fusion module in Sec. 4.3 to detail discuss their effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Class-Aware Improvements Analysis. Compared with the baseline model, our SupFusion and deep fusion could boost the detection performances not only for the overall performances but also for each class including Pedestrian, Cyclist, and Car. Comparing the average improvements (e.g. moderate cases) for three classes, we can obtain the following observation: the cyclist gains the largest improvements \\((+2.41\\%)\\) while pedestrian and car gains \\(+1.35\\%\\) and \\(+0.86\\%\\) increasements respectively. The reason is obvious: (1) The car is easier to detect and gains the best results compared with pedestrians and cyclists, hence it is more difficult to improve. (2) Cyclist gains more improvements against" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.079, + 0.894, + 0.109 + ], + "angle": 0, + "content": "Table 4: Results comparison with different supervision loss. \\(\\mathcal{L}_{Sim}\\) indicates the auxiliary supervision loss in Eq. 1." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.111, + 0.866, + 0.18 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{Sim} \\)Ped.Cyc.CarO.a.\\( \\Delta \\)
None64.8972.9485.5574.46-
L165.7675.5785.4275.58+1.12
L266.0676.1885.4775.90+1.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.199, + 0.877, + 0.214 + ], + "angle": 0, + "content": "Table 5: Results with different depth \\( K \\) in deep fusion." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.216, + 0.868, + 0.269 + ], + "angle": 0, + "content": "
K1234
LC74.2374.0073.8673.22
LC*75.6475.7575.9075.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.288, + 0.866, + 0.304 + ], + "angle": 0, + "content": "Table 6: Results with different \\( N \\) in polar sampling." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.305, + 0.868, + 0.358 + ], + "angle": 0, + "content": "
K14816
L†99.8099.6599.6699.43
LC*75.6875.7275.9075.85
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.379, + 0.892, + 0.425 + ], + "angle": 0, + "content": "Pedestrian because pedestrians are non-grid, the generated dense object is not so well compared with Cyclist, hence gains lower performance improvements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.437, + 0.654, + 0.453 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.506 + ], + "angle": 0, + "content": "To analyze the effectiveness of our proposed SupFusion and deep fusion module, we conduct ablation experiments based on Voxel-RCNN-LC on the KITTI benchmark." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Supervision. In SupFusion training strategy, we mainly introduce a high-quality guided feature-level supervision and simply employ L2 loss as the auxiliary loss. Here we also adopt L1 loss and the results are shown in Tab. 4, from which we can observe that L1 loss gains comparable but slight improvements (e.g. \\(+1.12\\%\\) versus \\(+1.44\\%\\)). To mimic the high-quality features by the fusion features, utilizing L1 and L2 to measure the feature distance is acceptable [13, 44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Fusion Strategy. We conduct experiments with different fusion methods, including the concatenation and summation approaches employed in MVXNet and PV-RCNN-LC/Voxel-RCNN-LC, as well as our proposed deep fusion methods. As shown in Tab. 3, without auxiliary supervision in SupFusion, our proposed deep fusion and concatenation fusion gain comparable but lower performances, demonstrating introducing extra parameters in the fusion model would lead the model difficult to train and hard to optimize. By introducing auxiliary supervision, our proposal could gain the best performance and improvements compared with the baseline model. This demonstrates that the SupFusion could effectively promote the learning process, guiding the 3D/2D backbone and fusion model extracts meaningful robust fusion features from LiDAR and camera inputs. Furthermore, we also manually set variance depth in deep fusion, to rescale the fusion module size. The results are shown in Tab. 5, in which we can observe that by increasing the fusion depth," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22020" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.287, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.091, + 0.488, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.091, + 0.885, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.245, + 0.287, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.245, + 0.488, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.245, + 0.885, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.414, + 0.895, + 0.461 + ], + "angle": 0, + "content": "Figure 5: Visualization results. The top row shows the raw sparse LiDAR data and the bottom row displays the enhanced data. Particularly, the points in magenta are the added points and row points are shown in black. We can densify the sparse objects with few points via the dense objects in database while keeping the density distributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.471, + 0.533 + ], + "angle": 0, + "content": "the baseline gets worse while our proposal gains better performances, demonstrating the effectiveness of our proposed SupFusion training strategy and deep fusion module again." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Polar Sampling. In the proposed polar sampling data enhancement, we split the objects into \\( N \\times N \\) groups according to their directions and rotations. To explore the influence of the split number \\( N \\), we manually set \\( N = \\{1, 4, 8, 16\\} \\) to conduct experiments, and the results are shown as Tab. 6, from which we can observe that the split number \\( N \\) in polar sampling affect the performances of assistant and our model in an extent. For instance, when setting \\( N = 1 \\), we only generate one object for each class, the object is complete but destroys the density distribution, which boosts the assistant model performances but decreases the improvements in an extent (e.g. \\( +1.22\\% \\) versus \\( +1.44\\% \\)), when setting \\( N \\) larger, the objects in each object set may get fewer, because we split all the objects in the dataset into more groups, which may influence the quality of the dense object in the database in the polar grouping." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Assistant Model. We utilize the enhanced dense LiDAR data to generate high-quality features. By densifying the sparse objects, the assistant model with enhanced data could avoid detection failers easily and provide high-quality features. To better apply to the given LiDAR-Camera detector and easier implement, we simply adopt the LiDAR branch detector as the assistant model. Here, we also conduct experiments to keep the assistant model sharing the same archi-" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.519 + ], + "angle": 0, + "content": "Table 7: Results with different teacher models in SupFusion. LC* and LC** indicate utilize LiDAR and LiDAR-Camera based assistant model." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.53, + 0.871, + 0.601 + ], + "angle": 0, + "content": "
ModelPed.Cyc.CarO.a.Δ
LC64.8972.9485.5574.46-
LC*66.0676.1885.4775.90+1.44
LC**66.3975.7185.4775.86+1.40
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.894, + 0.705 + ], + "angle": 0, + "content": "tecture with the given detector, which is a LiDAR-Camera detector. The results are shown as Tab. 7, from which we can observe that utilizing LiDAR-base and LiDAR-Camera based detectors as the assistant model, the SupFusion gains comparable performances (e.g. LC* versus LC**)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.684, + 0.728 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Here we visualize the row sparse LiDAR data and enhanced dense LiDAR in Fig. 5. In the left 2 columns, we can observe the added points in magenta could effectively densify the sparse points in black points for all classes including pedestrians, cyclists, and cars. In the right column, we can find that the polar sampling densifies the sparse objects with dense ones while keeping the objects' density distribution. As we could see, enhanced point cloud data have high integrity and obvious class representation which could be utilized to generate high-quality features as the fusion supervision. To better demonstrate the quality of feature-level" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22021" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "supervision, we visualize the feature of assistant, baseline, and our model in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.131, + 0.241, + 0.146 + ], + "angle": 0, + "content": "4.5. Information leak" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.155, + 0.468, + 0.229 + ], + "angle": 0, + "content": "There is no information leak if we are not misunderstanding. In practice, we build the database in Fig. 4 only using the training set in the polar grouping phase, the validation set is not used in polar sampling, as well as the model training process." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.243, + 0.194, + 0.259 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.269, + 0.468, + 0.524 + ], + "angle": 0, + "content": "In this paper, we propose a novel training strategy named SupFusion, to introduce auxiliary feature-level supervision for effective LiDAR-Camera fusion, and further boost 3D detection performances. Specifically, we introduce high-quality features to the LiDAR-Camera fusion-based 3D detector, which is used to provide auxiliary supervision for the 3D/2D feature extraction and multi-modal feature fusion. To achieve this, we propose a novel data enhancement method named polar sampling to densify the sparse objects, which is utilized to generate high-quality features via an assistant model. Furthermore, we proposed a simple yet effective deep fusion module, which leverages meaningful information from camera input and contiguously boosts the LiDAR-Camera fusion-based 3D detectors. This is the first approach to introducing feature-level supervision for LiDAR-Camera 3D detection and we conduct extensive experiments to demonstrate its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.539, + 0.212, + 0.556 + ], + "angle": 0, + "content": "6. Acknowledge" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.564, + 0.468, + 0.684 + ], + "angle": 0, + "content": "The work is partially supported by the Young Scientists Fund of the National Natural Science Foundation of China under grant No. 62106154, by the Natural Science Foundation of Guangdong Province, China (General Program) under grant No.2022A1515011524, and by Shenzhen Science and Technology Program JCYJ20220818103001002 and by Shenzhen Science and Technology Program ZDSYS20211021111415025." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22022" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In CVPR, pages 1090-1099, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.472, + 0.299 + ], + "angle": 0, + "content": "[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, pages 1907-1915, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[4] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In ICCV, pages 9775-9784, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[5] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. arXiv preprint arXiv:2211.09386, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[6] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. arXiv preprint arXiv:2201.10830, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[7] Xing Dai, Zeren Jiang, Zhao Wu, Yiping Bao, Zhicheng Wang, Si Liu, and Erjin Zhou. General instance distillation for object detection. In CVPR, pages 7842-7851, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[10] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[11] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu, and Chang Xu. Distilling object detectors via decoupled features. In CVPR, pages 2154-2164, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[12] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In ICCV, pages 3153-3163, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[14] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Epnet: Enhancing point features with image semantics for 3d object detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 35-52. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[15] Bo Ju, Zhikang Zou, Xiaoqing Ye, Minyue Jiang, Xiao Tan, Errui Ding, and Jingdong Wang. Paint and distill: Boosting" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.12 + ], + "angle": 0, + "content": "3d object detection with semantic passing network. In ACMM, pages 5639-5648, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[16] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[17] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In CVPR, pages 8973-8983, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[18] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[19] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.893, + 0.389 + ], + "angle": 0, + "content": "[20] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In CVPR, pages 1120-1129, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.893, + 0.459 + ], + "angle": 0, + "content": "[21] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V Le, et al. Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In CVPR, pages 17182-17191, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.893, + 0.53 + ], + "angle": 0, + "content": "[22] Zhijian Liu, Haotian Tang, Alexander Amini, Xingyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In IEEE International Conference on Robotics and Automation (ICRA), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.893, + 0.573 + ], + "angle": 0, + "content": "[23] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In ICCV, pages 15641-15650, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[24] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In ICCV, pages 3111-3121, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.893, + 0.772 + ], + "angle": 0, + "content": "[27] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In CVPR, pages 8555-8564, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.893, + 0.829 + ], + "angle": 0, + "content": "[28] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In CVPR, pages 9891-9901, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.893, + 0.871 + ], + "angle": 0, + "content": "[29] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: Realtime and high-performance pillar-based 3d object detection. In ECCV, pages 35-52. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[30] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22023" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.12 + ], + "angle": 0, + "content": "voxel feature set abstraction for 3d object detection. In CVPR, pages 10529-10538, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.163 + ], + "angle": 0, + "content": "[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointcnn: 3d object proposal generation and detection from point cloud. In CVPR, pages 770-779, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[32] Vishwanath A Sindagi, Yin Zhou, and Oncel Tuzel. Mvx-net: Multimodal voxelnet for 3d object detection. In 2019 International Conference on Robotics and Automation (ICRA), pages 7276-7282. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.472, + 0.291 + ], + "angle": 0, + "content": "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.472, + 0.361 + ], + "angle": 0, + "content": "[34] Kaicheng Yu Zhongyu Xia Zhiwei Lin Yongtao Wang Tao Tang Bing Wang Tingting Liang, Hongwei Xie and Zhi Tang. BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework. In Neural Information Processing Systems (NeurIPS), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.472, + 0.404 + ], + "angle": 0, + "content": "[35] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In CVPR, pages 11794-11803, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.472, + 0.461 + ], + "angle": 0, + "content": "[36] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3d object detection. In CVPR, pages 454-463, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.472, + 0.517 + ], + "angle": 0, + "content": "[37] Li Wang, Li Zhang, Yi Zhu, Zhi Zhang, Tong He, Mu Li, and Xiangyang Xue. Progressive coordinate transforms for monocular 3d object detection. Advances in Neural Information Processing Systems, 34:13364-13377, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.472, + 0.573 + ], + "angle": 0, + "content": "[38] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.472, + 0.617 + ], + "angle": 0, + "content": "[39] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, pages 4933-4942, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.472, + 0.66 + ], + "angle": 0, + "content": "[40] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In ICCV, pages 913-922, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.472, + 0.717 + ], + "angle": 0, + "content": "[41] Yue Wang, Alireza Fathi, Abhijit Kundu, David A Ross, Caroline Pantofaru, Tom Funkhouser, and Justin Solomon. Pillar-based object detection for autonomous driving. In ECCV, pages 18-34. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.472, + 0.786 + ], + "angle": 0, + "content": "[42] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[43] Xiaopei Wu, Liang Peng, Honghui Yang, Liang Xie, Chenxi Huang, Chengqi Deng, Haifeng Liu, and Deng Cai. Sparse fuse dense: Towards high quality 3d detection with depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5418-5427, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.903 + ], + "angle": 0, + "content": "[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "framework for masked image modeling. In CVPR, pages 9653-9663, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.149 + ], + "angle": 0, + "content": "[45] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[46] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In ICCV, pages 1951-1960, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.894, + 0.235 + ], + "angle": 0, + "content": "[47] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.894, + 0.292 + ], + "angle": 0, + "content": "[48] Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In ECCV, pages 720-736. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.294, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[49] Linfeng Zhang and Kaisheng Ma. Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.894, + 0.39 + ], + "angle": 0, + "content": "[50] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In CVPR, pages 18953-18962, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.894, + 0.433 + ], + "angle": 0, + "content": "[51] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In CVPR, pages 3289-3298, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.894, + 0.477 + ], + "angle": 0, + "content": "[52] Wu Zheng, Mingxuan Hong, Li Jiang, and Chi-Wing Fu. Boosting 3d object detection by simulating multimodality on point clouds. In CVPR, pages 13638-13647, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.478, + 0.894, + 0.533 + ], + "angle": 0, + "content": "[53] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3d object detection: An extrinsic parameter free approach. In CVPR, pages 7556-7566, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[54] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, pages 4490-4499, 2018." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22024" + } + ] +] \ No newline at end of file diff --git a/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/92e5510f-e140-45f8-8597-b810b193a66f_origin.pdf b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/92e5510f-e140-45f8-8597-b810b193a66f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9e69bc028e2d714ba007c7a8e69d7a3defbda47c --- /dev/null +++ b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/92e5510f-e140-45f8-8597-b810b193a66f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:601387ca2d3f5bf73149f87e6b5faaf99a9ac7555d06eb181c810379b0f2a9b1 +size 4184104 diff --git a/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/full.md b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..11114f98b1ae4c07616b83c02dac722d21286b5b --- /dev/null +++ b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/full.md @@ -0,0 +1,290 @@ +# SupFusion: Supervised LiDAR-Camera Fusion for 3D Object Detection + +Yiran Qin $^{1*}$ , Chaoqun Wang $^{1*}$ , Zijian Kang $^{2}$ , Ningning Ma $^{2}$ , Zhen Li $^{3}$ , Ruimao Zhang $^{1\dagger}$ + +$^{1}$ School of Data Science, Shenzhen Research Institute of Big Data, + +The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen), China $^{2}$ NIO + +$^{3}$ School of Science and Engineering, Future Intelligent Network Research Institute, + +The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen), China + +{yiranqin@link, chaoqunwang@link, lizhen@, ruimaozhang@}cuhk.edu.hk, {zijiankang, ningningma}@nio.com + +# Abstract + +LiDAR-Camera fusion-based 3D detection is a critical task for automatic driving. In recent years, many LiDAR-Camera fusion approaches sprung up and gained promising performances compared with single-modal detectors, but always lack carefully designed and effective supervision for the fusion process. In this paper, we propose a novel training strategy called SupFusion, which provides an auxiliary feature level supervision for effective LiDAR-Camera fusion and significantly boosts detection performance. Our strategy involves a data enhancement method named Polar Sampling, which densifies sparse objects and trains an assistant model to generate high-quality features as the supervision. These features are then used to train the LiDAR-Camera fusion model, where the fusion feature is optimized to simulate the generated high-quality features. Furthermore, we propose a simple yet effective deep fusion module, which contiguously gains superior performance compared with previous fusion methods with SupFusion strategy. In such a manner, our proposal shares the following advantages. Firstly, SupFusion introduces auxiliary feature-level supervision which could boost LiDAR-Camera detection performance without introducing extra inference costs. Secondly, the proposed deep fusion could continuously improve the detector's abilities. Our proposed SupFusion and deep fusion module is plug-and-play, we make extensive experiments to demonstrate its effectiveness. Specifically, we gain around $2\%$ 3D mAP improvements on KITTI benchmark based on multiple LiDAR-Camera 3D detectors. Our code is available at https://github.com/IranQin/SupFusion. + +![](images/54814ff1800592069e17aa0443576449f7f4e859af0ae525027806812b972f98.jpg) + +![](images/9ff36b3f8ea3bf9970e14f08d8bec410d40d31cd7ca77f7da2f65c9131bf1e89.jpg) +Figure 1: Top: The previous LiDAR-Camera 3D detector, the fusion module is optimized by detection loss. Bottom: Our proposed SupFusion, we introduce auxiliary supervision via high-quality features provided by an assistant model. + +# 1. Introduction + +LiDAR-Camera fusion-based 3D object detection is a critical and challenging task for autonomous driving and robotics [32, 35]. Previous approaches [20, 22, 34, 19] always project the camera inputs to LiDAR BEV or voxel space via intrinsic and extrinsic to align the LiDAR and camera features. Then, simple concatenation or summation is employed to obtain the fusion features for final detection. Also, some deep learning-based fusion approaches [1, 21] sprung up and gained promising performances. However, the previous fusion approaches always directly optimized the 3D/2D feature extraction as well as the fusion module by the + +detection loss, which lacks carefully designed and effective supervision at the feature level and limits its performance. + +In recent years, distillation manners have shown great improvements in feature-level supervision for 3D detection. For instance, some approaches [5, 6] provide LiDAR features to guide the 2D backbone in estimating the depth information from camera inputs. Also, some approaches [19] provide LiDAR-Camera fusion features to supervise LiDAR backbone learns global and contextual presentation from LiDAR inputs. Introducing feature-level auxiliary supervision via simulating more robust and high-quality features, the detectors could boost marginal improvements. Motivated by this, a natural solution to handle LiDAR-Camera feature fusion is providing stronger and high-quality features and introducing auxiliary supervision for LiDAR-Camera 3D detection. + +To this end, in this paper we propose a supervised LiDAR-Camera fusion method named SupFusion, to generate high-quality features and provide effective supervision for the fusion as well as feature extraction process, and further boost the LiDAR-Camera fusion-based 3D detection performances. Specifically, we first train an assistant model to provide high-quality features. To achieve it, different from previous approaches, which leverage larger models or extra data, we proposed a novel data enhancement method named Polar Sampling. The polar sampling could dynamically enhance the object's density from sparse LiDAR data, which is easier to detect and improve features quality, e.g., the feature could conclude accuracy detection results. Then, we simply train the LiDAR-Camera fusion-based detector with introduced auxiliary feature-level supervision. In this step, we feed the raw LiDAR and camera inputs into 3D/2D backbones and fusion modules to obtain fusion features. In one aspect, the fusion features are fed into the detection head for final prediction, which is decision-level supervision. In another aspect, the auxiliary supervision mimics the fusion feature to high-quality features, which are obtained via the pretrained assistant model and enhanced LiDAR data. In such a manner, the proposed feature-level supervision could lead the fusion module to generate more robust features, and further boost the detection performance. To better fuse the LiDAR and camera features, we propose a simple yet effective deep fusion module, which consists of stacked MLP blocks and dynamic fusion blocks. SupFusion could sufficiently excavate the deep fusion module capacity, and continuously improve detection accuracy. + +Compared with previous LiDAR-Camera 3D detectors, this is a brand new attempt to improve the effectiveness of the LiDAR-Camera fusion by introducing feature-level supervision, which consistently boosts 3D detection accuracy. The main contributions are summarized as follows. + +- We propose a novel supervised fusion training strategy named SupFusion, which mainly consists of a high-quality feature generation process and firstly propose + +auxiliary feature-level supervision loss for robust fusion feature extraction and accuracy 3D detection to our knowledge. + +- To obtain high-quality features in SupFusion, we propose a data enhancement method named Polar Sampling to densify the sparse objects. Furthermore, we propose an efficient deep fusion module to contiguously boost detection accuracy. +- We make extensive experiments based on multiple detectors with different fusion strategies, and gain around $2\%$ mAP improvements on KITTI benchmark. The source code will be released after the blind review. + +# 2. Related Work + +# 2.1. 3D Object Detection + +Camera-based 3D Object Detection methods perform 3D detection on single- or multi-view images. Many methods are proposed to address monocular 3D detection[24, 23, 17, 51, 53, 27, 37, 36] since KITTI benchmark[9] only contains one from the camera. With the development of autonomous driving datasets which have more cameras covering a wider range of angles(nuScenes[2] and Waymo[33]), more and more methods[38, 40, 42] are developed that take multi-view images as input and get surprising performance compared to monocular methods. + +LiDAR-based 3D Object Detection methods perform 3D detection on a point cloud scene. Depending on the three kinds of point cloud data pretreatments, 3D detectors are also classified into three mainstream. Point-based method [31, 46, 50, 4] use row point cloud data, leveraging pointnet[25] and pointnet++[26] as 3D backbone networks to get point-wised features which are sent to point-wise 3D detection head. Pillar-based method [18, 41, 29] pretreat point cloud to pillars, gain speed up while still maintaining a good performance. Voxel-based methods[54, 45, 8] are widely used, point clouds are voxelized to voxel and sent to a 3D sparse convolutional[10] based 3D backbone to get bird's-eye-view features. Some methods[30] also try to gain more improvement by using both the voxel feature and the point feature. + +# 2.2. LiDAR-Camera 3D Object Detection + +LiDAR-camera fusion is a technique that combines data from LiDAR and camera sensors to obtain a more complete understanding of a scene in 3D. Due to the complementary nature of the features produced by these two modalities, researchers have developed methods that can be jointly optimized on both sensors. These methods can be classified into two categories based on their fusion mechanism. The first category is point-level fusion[14, 35, 47], where the image features are queried via the raw LiDAR points and then + +![](images/57afa910eb861de1fe39472dea66f52a4aaa28851ec7bb490604103f32704b56.jpg) +Figure 2: The detector training procedure. For a given LiDAR-Camera sample, we extract the feature presentation via the corresponding 3D/2D encoders, and the proposed deep fusion module is employed to fuse the multi-modal features. Besides the detection loss $\mathcal{L}_{Det}$ , we introduce the auxiliary feature-level supervision $\mathcal{L}_{Sim}$ , to mimic the fusion feature $\mathbf{f}_i^{lc}$ to the high-quality features $\mathbf{f}_l^*$ , which is generated by the assistant model and enhanced data. Here the $\odot$ , $\oplus$ , and $\odot$ indicate the element-wise multiplication, summation, and channel-wise concatenation. More details could be referred in Sec. 3.1 + +concatenated back as additional point features. The second category is feature-level fusion, where the LiDAR points are projected into a specific feature space[48, 20, 19] or proposals level feature fusion[3, 16, 1, 43]. There are also methods that combine these two fusion manners, such as MVX-Net [32]. Recently, attention-based fusion methods have sprung up. TransFusion[1] uses the bounding box prediction of LiDAR features as a proposal to query the image feature, and then a Transformer-like architecture fuses the information back to LiDAR features. Similarly, DeepFusion[21] projects LiDAR features on each view image as queries and leverages cross-attention for the two modalities. Bird's-eye-view space feature fusion[22, 34] is proposed recently and gained great performance with different tasks. + +# 2.3. Knowledge Distillation for 3D Object Detection + +Object detection tasks differ from classification tasks in that they require consideration of both global and local features. While numerous distillation methods have been proposed to address this challenge, such as those developed by [39, 7, 11, 49], many of these methods require the creation of specific distillation regions or objects based on ground-truth data, bounding box predictions, attention, or gradients to mitigate the imbalance between foreground and background. In addition to 2D detection, there has been growing interest in cross-model knowledge distillation methods between RGB and LiDAR in 3D object detection. For example, [12, 28, 6, 5] have developed methods that transfer knowledge from LiDAR detectors to image detectors. [52, 15] has trained a teacher model similar to that proposed by [35] and developed a multilevel distillation framework to transfer semantic segmentation information to a student model. In this + +work, we want to further boost LiDAR-Camera detectors with an auxiliary feature level supervision. + +# 3. Methodology + +In this section, we first overview the overall SupFusion pipeline in Sec. 3.1, and then detail introduces the main components, including the proposed Polar Sampling in Sec. 3.2, and deep fusion module in Sec. 3.3. + +# 3.1. Overall SupFusion Pipeline + +For a given LiDAR-Camera fusion-based 3D detector $\mathbf{D}$ , which consists of a 3D Encoder $\mathbf{E}_l$ , a 2D Encoder $\mathbf{E}_c$ , a fusion module $\mathbf{F}$ , and a detection head. Denotes $\mathcal{D} = \{\left(\mathbf{l}_1,\mathbf{c}_1\right),\left(\mathbf{l}_2,\mathbf{c}_2\right),\dots,\left(\mathbf{l}_N,\mathbf{c}_N\right)\}$ and $\mathcal{V} = \{\mathbf{y}_1,\mathbf{y}_2,\dots,\mathbf{y}_N\}$ are the given dataset with $N$ samples, where $(\mathbf{l}_i,\mathbf{c}_i)$ are the paired LiDAR and camera inputs. The previous approaches feed the paired $(\mathbf{l}_i,\mathbf{c}_i)$ into 3D/2D encoders $\mathbf{E}_l$ and $\mathbf{E}_c$ to capture corresponding feature presentation $\mathbf{f}_i^l$ and $\mathbf{f}_i^c$ . Then, the paired features presentation is fed into the fusion module to obtain the fusion feature $\mathbf{f}_i^{lc}$ , which is used for the final prediction. To better leverage the LiDAR and camera data, we propose a supervised LiDAR-Camera Fusion training strategy named SupFusion, which provides high-quality features for feature-level supervision. To end this, different from previous approaches, which adopt larger scale models or extra data input, we propose a data enhancement method named polar sampling to densify the sparse objects in LiDAR data. In such a manner, the data is easier and could easily generate high-quality features (e.g., the feature could conclude accuracy detection results) with lightweight models. The proposed polar sampling mainly + +![](images/8a3162c7f7ddf1873005e730012f3e3b23b8cb2001fef0969b259e2889da6775.jpg) +Figure 3: High-quality feature generation process. For any given LiDAR sample, we densify the sparse objects via polar pasting, which calculate the direction and rotation to query dense object from the database, and pasting to add extra points to sparse objects. We first train the assistant model via the enhanced data and feed the enhanced LiDAR data into the assistant model to generate high-quality features $\mathbf{f}^*$ after it is convergent. More details could be referred in Sec. 3.1. + +densifies the sparse object points in LiDAR data. Specifically, for any sample $(\mathbf{l}_i,\mathbf{c}_i)$ , where $\mathbf{l}_i = \{l_i^1,l_i^2,\dots,l_i^{n_i},l_i^s\}$ contains $n_i$ objects points sets and one background points set. In the polar sampling, we add extra points to each object point set from raw LiDAR data, to obtain enhanced LiDAR data $\mathbf{l}_i' = \{(l_i^1,\hat{l}_i^1),(l_i^2,\hat{l}_i^2),\dots,(l_i^{n_i},\hat{l}_i^{n_i}),l_i^s\}$ in polar pasting, where $\hat{l}_i^k$ is the added points set for the $k$ -th objects in $i$ -th LiDAR sample and generated in polar grouping. We will detail introduce the polar sampling in Sec. 3.2. + +High-quality Feature Generation. To provide feature-level supervision in our SupFusion, we adopt an assistant model to capture high-quality features from the enhanced data as shown in Fig. 3. Firstly, we train an assistant model to provide high-quality features. Specifically, for any sample in $\mathcal{D}$ , we enhance the sparse LiDAR data $\mathbf{l}_i$ to obtain enhanced data $\mathbf{l}_i^{\prime}$ via the polar pasting, which densifies the spare objects with added points set generated in polar grouping. Then, after the assistant model convergence (often with few epochs), we feed the enhanced sample $\mathbf{l}_i^{\prime}$ into the optimized assistant model, to capture high-quality features $\mathbf{f}_i^*$ , which is used to train the LiDAR-Camera 3D detector. To better apply to the given LiDAR-Camera detector and easier implement, we simply adopt the LiDAR branch detector as the assistant model. + +Detector Training. For any given LiDAR-Camera detector, we train the model with the proposed auxiliary supervision at the feature level. Specifically, given sample $(\mathbf{l}_i,\mathbf{c}_i)\in \mathcal{D}$ , we first feed the LiDAR and camera into the 3D + +and 2D Encoders $\mathbf{E}_l$ and $\mathbf{E}_s$ to capture the corresponding features $\mathbf{f}_i^l$ and $\mathbf{f}_i^c$ , which are fed into the fusion model to generate fusion features $\mathbf{f}_i^{l,c}$ , and flow into the detection head for final prediction. Furthermore, the proposed auxiliary supervision is employed to mimic the fusion feature with high-quality features, which is generated by the pre-trained assistant model and enhanced LiDAR data. The above process can be formulated as: + +$$ +\begin{array}{l} \theta^ {*} = \underset {\theta} {\arg \min } \mathbb {E} _ {((\mathbf {l} _ {i}, \mathbf {c} _ {i}), \mathbf {y} _ {i}) \sim \mathcal {D}} \mathcal {L} _ {\text {d e t}} (\mathbf {D} (\mathbf {l} _ {i}, \mathbf {c} _ {i}, \theta), \mathbf {y} _ {i}) (1) \\ + \lambda \mathcal {L} _ {\text {S i m}} \left(\mathbf {f} _ {i} ^ {l c}, \mathbf {f} _ {i} ^ {*}\right), (2) \\ \end{array} +$$ + +$$ +s. t. \quad \mathbf {f} _ {i} ^ {l c} = \mathbf {F} \left(\mathbf {E} _ {l} \left(\mathbf {l} _ {i}, \theta_ {l}\right), \mathbf {E} _ {c} \left(\mathbf {c} _ {i}, \theta_ {c}\right), \theta_ {f}\right), \tag {3} +$$ + +$$ +\mathbf {f} _ {i} ^ {*} = \mathbf {E} _ {l} \left(\mathbf {I} _ {i} ^ {\prime}, \theta_ {l} ^ {*}\right), \tag {4} +$$ + +$$ +\mathcal {L} _ {S i m} = \left\| \psi \left(\mathbf {f} _ {i} ^ {l c}\right) - \mathbf {f} _ {i} ^ {*} \right\| _ {2} ^ {2} \tag {5} +$$ + +where $\theta^{*}$ is the optimized parameters of detector $\mathbf{D}$ , $\mathbf{f}_i^*$ is the high-quality features, which is generated from pre-trained assistant model LiDAR encoder $\mathbf{E}_l$ with parameters $\theta_l^*$ , $\psi$ is a simple convolutional layer with kernel size 1. $\lambda$ is the balanced weight and we set it as 1.0. + +# 3.2. Polar Sampling + +To provide high-quality features via the assistant model in our proposed SupFusion, we propose a novel data enhancement method named Polar Sampling, to ease the sparse problem which frequently leads to the detection failers. To end this, we densify the sparse objects in LiDAR data with + +a similar dense object. Specifically, the polar sampling consists of two parts, including polar grouping and polar pasting. In polar grouping, we mainly build a database to store the dense objects, which is used in polar pasting to densify the sparse objects. + +Considering the characteristics of LiDAR sensors, the collected point cloud data naturally exists a particular density distribution, e.g., the objects have more points on the surface towards the LiDAR sensor, while few points on the opposite sides. The density distribution is mainly influenced by the direction and rotation, while the points' density mainly depends on the distances, e.g. the objects closer to the LiDAR sensor have denser points. Motivated by this, we aim to densify the sparse objects in longer distances with the dense objects in short distances, according to their direction and rotation to keep the density distributions. Specifically, we build polar coordinate systems for the whole scene as well as the objects based on the center of the scene and specific objects and define the positive direction of the LiDAR sensor as 0 degrees to measure the corresponding directions and rotations. Then, we collect the objects with similar density distribution (e.g. with similar direction and rotation) and generate one dense object for each group in the polar grouping, and used it to densify the sparse objects in polar pasting. + +Polar Grouping. As shown in Fig. 4, we build a database $\mathcal{B}$ to store the generated dense object point set $\hat{l}$ according to their directions and rotations in the polar grouping, which is noted as $\alpha$ and $\beta$ in Fig. 4. + +Firstly, we search the whole dataset and calculated their polar angles for all objects via the locations and provided rotations in the benchmarks. Secondly, we split the objects into groups according to their polar angles. Specifically, we manually divide the direction and rotation into $N$ groups. For any object point set $l$ , we could put it into the corresponding group according to the index: + +$$ +\alpha_ {l} ^ {\prime} = \left\lfloor \frac {\alpha_ {l} \cdot 2 \pi}{N} \right\rfloor , \quad \beta_ {l} ^ {\prime} = \left\lfloor \frac {\beta_ {l} \cdot 2 \pi}{N} \right\rfloor , \tag {6} +$$ + +where $\alpha_{l}^{\prime},\beta_{l}^{\prime}\in \{1,2,\dots,N\}$ are the direction and rotation index of object $l$ , and we could put the points set into $\mathcal{B}$ as: $\mathcal{B}_C^{\alpha_l',\beta_l'}\gets l$ , where $C$ is the label of object $l$ . Finally, for each object group $\mathcal{B}_C^{\alpha ',\beta '} = \{l_1,l_2,\ldots ,l_n\}$ contains $n$ similar objects, we select the densest $k$ objects and mix up to generate one dense and complete object $\hat{l}$ , and $\mathcal{B}_C^{\alpha ',\beta '} = \{\hat{l}\}$ is used in polar pasting phase. To keep the density and avoid too much calculation in the model, we randomly select partial points in the dense object. Here we set $k = 10$ and select 5000 points for all objects. + +Polar Pasting. As shown in Fig. 2, the polar pasting is utilized to enhance the sparse LiDAR data to train the assistant model and generate high-quality features. Concretely, given a LiDAR sample $\mathbf{l}_i = \{l_i^1, l_i^2, \dots, l_i^{n_i}, l_i^s\}$ contains $n_i$ objects, for any object $l_i^k$ , we can calculate the direction + +![](images/63df54b7c17ffde47cf418c685d34796aeec5bfef97b2c2010b4724f5e142881.jpg) +Figure 4: Polar grouping mainly builds a database to store the dense objects of each group in polar coordinate systems of direction and rotations for each class. More details could be referred to in Sec. 3.2. + +and rotation the same as the grouping process and query the dense object $\hat{l}_i^k$ from $\mathcal{B}$ according to the label and indexes, which could be obtained from E.q. 6 In such a manner, we enhance all the objects in the sample and obtain the enhanced data: $\mathbf{I}_i' = \{(l_i^1,\hat{l}_i^1),(l_i^2,\hat{l}_i^2),\dots,(l_i^{n_i},\hat{l}_i^{n_i}),l_i^s\}$ + +# 3.3. Deep Fusion + +To mimic the high-quality features generated from enhanced LiDAR data, the fusion model aims to extract the missing information of the sparse objects from the rich color and contextual features in camera input. To end this, we propose the deep fusion module, to leverage image features and complete the LiDAR presentation. The proposed deep fusion mainly consists of a 3D learner and a 2D3D learner. The 3D learner is a simple convolutional layer, to transfer the 3D presentation into 2D space. Then, concatenating the 2D features and 3D presentation (e.g. in 2D space), the 2D3D learner is employed to fuse the LiDAR-Camera features. Finally, we weighed the fusion feature via an MLP and an activate function, adding back to the raw LiDAR features as the output of the deep fusion module. The 2D3D learner consists of stacked MLP blocks with depth $K$ , learning to leverage camera features to complete the LiDAR representation from the sparse objects, to simulate the high-quality features of dense LiDAR objects. + +# 4. Experiments + +# 4.1. Experiments Settings + +Datasets and Evaluation Metrics. KITTI [9] is a large-scale public benchmark for autonomous driving, including 3712 training, 3769 validation, and 7518 testing samples with LiDAR and camera pair data. All objects in KITTI are divided into 'easy', 'moderate', and 'hard' difficulties ac + +Table 1: The experiment results (mAP@R40 %) on KITTI validation set. We list the easy, moderate (mod.), and hard cases for three classes, and the overall performances. Here, L, LC, LC* indicate the results of the corresponding LiDAR detector, LiDAR-Camera fusion detectors, and our proposal. $\Delta$ indicates the improvements. The best results are shown in bold expect $\mathrm{L}^{\dagger}$ , which is the assistant model, and test on the enhanced validation set. MVXNet is re-conduct based on mmdetection3d. PV-RCNN-LC and Voxel-RCNN-LC are re-conduct based on the open-source code of VFF [20]. + +
DetectorModalityOverallPedestrianCyclistCar
easymod.hardeasymod.hardeasymod.hardeasymod.hard
MVXNetL†88.4490.9388.5887.3387.6180.1580.0986.0186.1997.8999.1799.41
L73.4658.6455.1557.5450.1045.5674.6351.8148.9288.2173.9970.98
LC74.9963.0059.3264.4657.7653.0572.6955.2751.6587.8275.9873.26
LC*75.3764.7360.5263.1158.1353.7374.7957.6053.8288.2078.4674.01
Δ+0.38+1.73+1.20-1.35+0.37+0.68+2.10+2.33+2.17+0.38+2.48+0.75
PV-RCNN-LCL†94.9095.2095.1086.8686.7887.0798.9099.3398.6298.9599.5099.62
L81.3771.5568.0867.8059.5654.4085.0670.8367.9691.2684.2581.88
LC84.1674.5470.8670.4564.1659.5090.4774.4670.2591.5685.0182.83
LC*85.7776.0072.0074.2966.6761.3891.0076.1371.6692.0285.2082.94
Δ+1.61+1.46+1.14+3.84+2.51+1.88+0.53+1.67+1.41+0.46+0.19+0.11
Voxel-RCNN-LCL†99.2699.6699.7597.8498.0598.1794.3694.6094.8599.9099.9099.91
L82.0171.6968.4067.8460.9556.0386.1071.3967.0592.0982.7482.12
LC84.3674.4670.8071.3964.8960.2089.6472.9469.0392.0585.5583.17
LC*86.1375.9072.1773.9566.0661.8392.0276.1871.8392.4285.4782.83
Δ+1.77+1.44+1.37+2.56+1.17+1.63+2.38+3.24+2.80+0.37-0.08-0.34
+ +cording to the object scales and occlusion ratio. We evaluate the detection performance by the mean Average Precision under 40 recall thresholds (mAP@R40) the same as the official benchmark and evaluation for 3 classes including Pedestrian, Cyclist, and Car. nuScenes [2] is a large-scale autonomous driving dataset for 3D detection and tracking tasks, including 700 training, 150 validation, and 150 testing sequences, with each 20s long and annotated with 3D bounding boxes. For 3D object detection, the official evaluation metrics include the mean Average Precision (mAP) and nuScenes detection score (NDS). mAP measures the localization precision using a threshold base on the birds-eye view. NDS is a weighted combination of mAP and regression accuracy of other object attributes, including box size, orientation, and translation. + +Detectors. To demonstrate the effectiveness of our proposed method, we apply the SupFusion training strategy and deep fusion module on variance LiDAR-Camera detectors including MVXNet [32], PV-RCNN-LC [30], and Voxel-RCNN-LC [8] on KITTI, and SECOND-LC [45] on nuScenes benchmark. MVXNet is a voxel-based LiDAR-Camera 3D detector based on VoxelNet [54]. The camera features project to LiDAR space and concatenates with points features. Voxel-RCNN and PV-RCNN are LiDAR + +based detectors, we re-conduct the LiDAR-Camera fusion approach following VFF [20] based on the open-source code, which is a previous state-of-the-art fusion approach based on the two detectors, mainly projects the camera features into voxel presentation and summarized fusion with the voxel features of LiDAR input. SECOND-LC introduces camera inputs based on the LiDAR-based SECOND detector, the same as BEVFusion [22], we concatenate the camera and LiDAR features in birds-eys-view spaces. + +Implement Details. We conduct experiments on MVXNet, PV-RCNN-LC and Voxel-RCNN-LC, SECOND-LC following the open-sourced code and mmdetection3d on 8 Tesla A100 GPUs. All the architecture, configuration, and optimizer follow the settings reported in papers including voxel size, point cloud range, batch size, and learning rate. We modify the fusion method (e.g. concatenation, summarized) by our proposed deep fusion module and applied our SupFusion training strategy. For the deep fusion module, we set depth $K = 3$ . For the auxiliary supervision, we utilize simple L2 loss and set $\lambda = 1.0$ . More details about how to apply SupFusion and deep fusion modules on different detectors are demonstrated in the Appendix. + +Table 2: NDS and mAP(%) results on nuScenes benchmarks. + +
Modality3D2DNDSMAP
L†SECOND-90.3894.44
LSECOND-60.5252.28
LCSECONDSwin-T62.5755.26
LC*SECONDSwin-T64.58(+2.01)56.64(+1.38)
+ +Table 3: Results comparison with different fusion methods. Sup. indicates our proposed feature-level supervision. Ped., Cyc., and O.a. are short of Pedestrian, Cyclist, and Overall. + +
MethodSup.Ped.Cyc.CarO.a.Δ
SumX64.8972.9485.5574.46-
✓67.0374.6185.1175.59+1.13
ConcatX65.9072.8083.3774.02-
✓66.2374.9585.4275.52+1.50
OursX63.1872.8485.5873.86-
✓66.0676.1885.4775.90+2.04
+ +# 4.2. Main Results + +Overall Performances. The 3D mAP@R40 comparison based on three detectors is shown in Tab. 1 for three classes and overall performances for each difficulty split. We can clearly observe that the LiDAR-Camera approaches (LC) are superior to the LiDAR-based detectors (L) by introducing extra camera inputs. By introducing polar sampling, the assistant model $(\mathrm{L}^{\dagger})$ could gain admirable performances (e.g. over $90\%$ mAP) on the enhanced validation set. With the auxiliary supervision from high-quality features and our proposed deep fusion module, our proposal could contiguously boost detection accuracy. For instance, compared with the baseline (LC) model, our proposal could gain $+1.54\%$ and $+1.24\%$ 3D mAP improvements for the moderate and hard objects on average for the three detectors. Further, we also conduct experiments on nuScenes benchmark based on SECOND-LC as shown in Tab. 2, which gain $+2.01\%$ and $+1.38\%$ improvements for NDS and mAP. Also, we will prune the SupFusion training strategy and deep fusion module in Sec. 4.3 to detail discuss their effectiveness. + +Class-Aware Improvements Analysis. Compared with the baseline model, our SupFusion and deep fusion could boost the detection performances not only for the overall performances but also for each class including Pedestrian, Cyclist, and Car. Comparing the average improvements (e.g. moderate cases) for three classes, we can obtain the following observation: the cyclist gains the largest improvements $(+2.41\%)$ while pedestrian and car gains $+1.35\%$ and $+0.86\%$ increasements respectively. The reason is obvious: (1) The car is easier to detect and gains the best results compared with pedestrians and cyclists, hence it is more difficult to improve. (2) Cyclist gains more improvements against + +Table 4: Results comparison with different supervision loss. $\mathcal{L}_{Sim}$ indicates the auxiliary supervision loss in Eq. 1. + +
\( \mathcal{L}_{Sim} \)Ped.Cyc.CarO.a.\( \Delta \)
None64.8972.9485.5574.46-
L165.7675.5785.4275.58+1.12
L266.0676.1885.4775.90+1.44
+ +Table 5: Results with different depth $K$ in deep fusion. + +
K1234
LC74.2374.0073.8673.22
LC*75.6475.7575.9075.88
+ +Table 6: Results with different $N$ in polar sampling. + +
K14816
L†99.8099.6599.6699.43
LC*75.6875.7275.9075.85
+ +Pedestrian because pedestrians are non-grid, the generated dense object is not so well compared with Cyclist, hence gains lower performance improvements. + +# 4.3. Ablation Study + +To analyze the effectiveness of our proposed SupFusion and deep fusion module, we conduct ablation experiments based on Voxel-RCNN-LC on the KITTI benchmark. + +Supervision. In SupFusion training strategy, we mainly introduce a high-quality guided feature-level supervision and simply employ L2 loss as the auxiliary loss. Here we also adopt L1 loss and the results are shown in Tab. 4, from which we can observe that L1 loss gains comparable but slight improvements (e.g. $+1.12\%$ versus $+1.44\%$ ). To mimic the high-quality features by the fusion features, utilizing L1 and L2 to measure the feature distance is acceptable [13, 44]. + +Fusion Strategy. We conduct experiments with different fusion methods, including the concatenation and summation approaches employed in MVXNet and PV-RCNN-LC/Voxel-RCNN-LC, as well as our proposed deep fusion methods. As shown in Tab. 3, without auxiliary supervision in SupFusion, our proposed deep fusion and concatenation fusion gain comparable but lower performances, demonstrating introducing extra parameters in the fusion model would lead the model difficult to train and hard to optimize. By introducing auxiliary supervision, our proposal could gain the best performance and improvements compared with the baseline model. This demonstrates that the SupFusion could effectively promote the learning process, guiding the 3D/2D backbone and fusion model extracts meaningful robust fusion features from LiDAR and camera inputs. Furthermore, we also manually set variance depth in deep fusion, to rescale the fusion module size. The results are shown in Tab. 5, in which we can observe that by increasing the fusion depth, + +![](images/ddabcb16c12cd858ea94c87677e1fd75f2ee47e590554831cf15cbaecb48c401.jpg) + +![](images/0600ecff488cd288295c93125f848ca421168e636c92ced0f71a5107d651563b.jpg) + +![](images/16f895916f1513043ab5b173c7bc1b780a0994b19d9d84e417d8fc722585e1a4.jpg) + +![](images/73b5d7c4475322f8b7f7388f041a91dffebf28bed7f1c8948de7bac5df24fa65.jpg) +Figure 5: Visualization results. The top row shows the raw sparse LiDAR data and the bottom row displays the enhanced data. Particularly, the points in magenta are the added points and row points are shown in black. We can densify the sparse objects with few points via the dense objects in database while keeping the density distributions. + +![](images/0330759fde550c5e00492c52f3f103d331a9cf1584f0b84f5104ac781ffc4fc1.jpg) + +![](images/2e008fd4c23467c92c78a7d036dfcf52d485497a58984ff8affc858d603433d8.jpg) + +the baseline gets worse while our proposal gains better performances, demonstrating the effectiveness of our proposed SupFusion training strategy and deep fusion module again. + +Polar Sampling. In the proposed polar sampling data enhancement, we split the objects into $N \times N$ groups according to their directions and rotations. To explore the influence of the split number $N$ , we manually set $N = \{1, 4, 8, 16\}$ to conduct experiments, and the results are shown as Tab. 6, from which we can observe that the split number $N$ in polar sampling affect the performances of assistant and our model in an extent. For instance, when setting $N = 1$ , we only generate one object for each class, the object is complete but destroys the density distribution, which boosts the assistant model performances but decreases the improvements in an extent (e.g. $+1.22\%$ versus $+1.44\%$ ), when setting $N$ larger, the objects in each object set may get fewer, because we split all the objects in the dataset into more groups, which may influence the quality of the dense object in the database in the polar grouping. + +Assistant Model. We utilize the enhanced dense LiDAR data to generate high-quality features. By densifying the sparse objects, the assistant model with enhanced data could avoid detection failers easily and provide high-quality features. To better apply to the given LiDAR-Camera detector and easier implement, we simply adopt the LiDAR branch detector as the assistant model. Here, we also conduct experiments to keep the assistant model sharing the same archi- + +Table 7: Results with different teacher models in SupFusion. LC* and LC** indicate utilize LiDAR and LiDAR-Camera based assistant model. + +
ModelPed.Cyc.CarO.a.Δ
LC64.8972.9485.5574.46-
LC*66.0676.1885.4775.90+1.44
LC**66.3975.7185.4775.86+1.40
+ +tecture with the given detector, which is a LiDAR-Camera detector. The results are shown as Tab. 7, from which we can observe that utilizing LiDAR-base and LiDAR-Camera based detectors as the assistant model, the SupFusion gains comparable performances (e.g. LC* versus LC**). + +# 4.4. Qualitative Results + +Here we visualize the row sparse LiDAR data and enhanced dense LiDAR in Fig. 5. In the left 2 columns, we can observe the added points in magenta could effectively densify the sparse points in black points for all classes including pedestrians, cyclists, and cars. In the right column, we can find that the polar sampling densifies the sparse objects with dense ones while keeping the objects' density distribution. As we could see, enhanced point cloud data have high integrity and obvious class representation which could be utilized to generate high-quality features as the fusion supervision. To better demonstrate the quality of feature-level + +supervision, we visualize the feature of assistant, baseline, and our model in Appendix. + +# 4.5. Information leak + +There is no information leak if we are not misunderstanding. In practice, we build the database in Fig. 4 only using the training set in the polar grouping phase, the validation set is not used in polar sampling, as well as the model training process. + +# 5. Conclusion + +In this paper, we propose a novel training strategy named SupFusion, to introduce auxiliary feature-level supervision for effective LiDAR-Camera fusion, and further boost 3D detection performances. Specifically, we introduce high-quality features to the LiDAR-Camera fusion-based 3D detector, which is used to provide auxiliary supervision for the 3D/2D feature extraction and multi-modal feature fusion. To achieve this, we propose a novel data enhancement method named polar sampling to densify the sparse objects, which is utilized to generate high-quality features via an assistant model. Furthermore, we proposed a simple yet effective deep fusion module, which leverages meaningful information from camera input and contiguously boosts the LiDAR-Camera fusion-based 3D detectors. This is the first approach to introducing feature-level supervision for LiDAR-Camera 3D detection and we conduct extensive experiments to demonstrate its effectiveness. + +# 6. Acknowledge + +The work is partially supported by the Young Scientists Fund of the National Natural Science Foundation of China under grant No. 62106154, by the Natural Science Foundation of Guangdong Province, China (General Program) under grant No.2022A1515011524, and by Shenzhen Science and Technology Program JCYJ20220818103001002 and by Shenzhen Science and Technology Program ZDSYS20211021111415025. + +# References + +[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In CVPR, pages 1090-1099, 2022. +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. +[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, pages 1907-1915, 2017. +[4] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In ICCV, pages 9775-9784, 2019. +[5] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. arXiv preprint arXiv:2211.09386, 2022. +[6] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. arXiv preprint arXiv:2201.10830, 2022. +[7] Xing Dai, Zeren Jiang, Zhao Wu, Yiping Bao, Zhicheng Wang, Si Liu, and Erjin Zhou. General instance distillation for object detection. In CVPR, pages 7842-7851, 2021. +[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021. +[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. +[10] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018. +[11] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu, and Chang Xu. Distilling object detectors via decoupled features. In CVPR, pages 2154-2164, 2021. +[12] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In ICCV, pages 3153-3163, 2021. +[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. +[14] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Epnet: Enhancing point features with image semantics for 3d object detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 35-52. Springer, 2020. +[15] Bo Ju, Zhikang Zou, Xiaoqing Ye, Minyue Jiang, Xiao Tan, Errui Ding, and Jingdong Wang. Paint and distill: Boosting + +3d object detection with semantic passing network. In ACMM, pages 5639-5648, 2022. +[16] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018. +[17] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In CVPR, pages 8973-8983, 2021. +[18] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. +[19] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022. +[20] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In CVPR, pages 1120-1129, 2022. +[21] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V Le, et al. Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In CVPR, pages 17182-17191, 2022. +[22] Zhijian Liu, Haotian Tang, Alexander Amini, Xingyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In IEEE International Conference on Robotics and Automation (ICRA), 2023. +[23] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In ICCV, pages 15641-15650, 2021. +[24] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In ICCV, pages 3111-3121, 2021. +[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. +[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. +[27] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In CVPR, pages 8555-8564, 2021. +[28] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In CVPR, pages 9891-9901, 2022. +[29] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: Realtime and high-performance pillar-based 3d object detection. In ECCV, pages 35-52. Springer, 2022. +[30] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point + +voxel feature set abstraction for 3d object detection. In CVPR, pages 10529-10538, 2020. +[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointcnn: 3d object proposal generation and detection from point cloud. In CVPR, pages 770-779, 2019. +[32] Vishwanath A Sindagi, Yin Zhou, and Oncel Tuzel. Mvx-net: Multimodal voxelnet for 3d object detection. In 2019 International Conference on Robotics and Automation (ICRA), pages 7276-7282. IEEE, 2019. +[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020. +[34] Kaicheng Yu Zhongyu Xia Zhiwei Lin Yongtao Wang Tao Tang Bing Wang Tingting Liang, Hongwei Xie and Zhi Tang. BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework. In Neural Information Processing Systems (NeurIPS), 2022. +[35] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In CVPR, pages 11794-11803, 2021. +[36] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3d object detection. In CVPR, pages 454-463, 2021. +[37] Li Wang, Li Zhang, Yi Zhu, Zhi Zhang, Tong He, Mu Li, and Xiangyang Xue. Progressive coordinate transforms for monocular 3d object detection. Advances in Neural Information Processing Systems, 34:13364-13377, 2021. +[38] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022. +[39] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, pages 4933-4942, 2019. +[40] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In ICCV, pages 913-922, 2021. +[41] Yue Wang, Alireza Fathi, Abhijit Kundu, David A Ross, Caroline Pantofaru, Tom Funkhouser, and Justin Solomon. Pillar-based object detection for autonomous driving. In ECCV, pages 18-34. Springer, 2020. +[42] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. +[43] Xiaopei Wu, Liang Peng, Honghui Yang, Liang Xie, Chenxi Huang, Chengqi Deng, Haifeng Liu, and Deng Cai. Sparse fuse dense: Towards high quality 3d detection with depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5418-5427, 2022. +[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple + +framework for masked image modeling. In CVPR, pages 9653-9663, 2022. +[45] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. +[46] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In ICCV, pages 1951-1960, 2019. +[47] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021. +[48] Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In ECCV, pages 720-736. Springer, 2020. +[49] Linfeng Zhang and Kaisheng Ma. Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In ICLR, 2021. +[50] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In CVPR, pages 18953-18962, 2022. +[51] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In CVPR, pages 3289-3298, 2021. +[52] Wu Zheng, Mingxuan Hong, Li Jiang, and Chi-Wing Fu. Boosting 3d object detection by simulating multimodality on point clouds. In CVPR, pages 13638-13647, 2022. +[53] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3d object detection: An extrinsic parameter free approach. In CVPR, pages 7556-7566, 2021. +[54] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, pages 4490-4499, 2018. \ No newline at end of file diff --git a/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/images.zip b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0aa29a99ae1d299b3fcbee50dccddb8791d2d6c9 --- /dev/null +++ b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d70f003557d739219a1d1981353deb318047bbc6ab001ae29aef0789f3227e8 +size 597609 diff --git a/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/layout.json b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c6863e40c8db063bf9670d6f9e10b9cc58efa003 --- /dev/null +++ b/2023/SupFusion_ Supervised LiDAR-Camera Fusion for 3D Object Detection/layout.json @@ -0,0 +1,8670 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 518, + 121 + ], + "type": "text", + "content": "SupFusion: Supervised LiDAR-Camera Fusion for 3D Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "spans": [ + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "text", + "content": "Yiran Qin" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "text", + "content": ", Chaoqun Wang" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "text", + "content": ", Zijian Kang" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "text", + "content": ", Ningning Ma" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "text", + "content": ", Zhen Li" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "text", + "content": ", Ruimao Zhang" + }, + { + "bbox": [ + 86, + 142, + 511, + 158 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 157, + 458, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 157, + 458, + 171 + ], + "spans": [ + { + "bbox": [ + 135, + 157, + 458, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 157, + 458, + 171 + ], + "type": "text", + "content": "School of Data Science, Shenzhen Research Institute of Big Data," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 171, + 496, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 171, + 496, + 185 + ], + "spans": [ + { + "bbox": [ + 96, + 171, + 496, + 185 + ], + "type": "text", + "content": "The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen), China " + }, + { + "bbox": [ + 96, + 171, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 171, + 496, + 185 + ], + "type": "text", + "content": "NIO" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 185, + 497, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 185, + 497, + 199 + ], + "spans": [ + { + "bbox": [ + 96, + 185, + 497, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 185, + 497, + 199 + ], + "type": "text", + "content": "School of Science and Engineering, Future Intelligent Network Research Institute," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 200, + 482, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 200, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 111, + 200, + 482, + 213 + ], + "type": "text", + "content": "The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen), China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 216, + 539, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 216, + 539, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 216, + 539, + 228 + ], + "type": "text", + "content": "{yiranqin@link, chaoqunwang@link, lizhen@, ruimaozhang@}cuhk.edu.hk, {zijiankang, ningningma}@nio.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 289, + 290, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 289, + 290, + 639 + ], + "spans": [ + { + "bbox": [ + 45, + 289, + 290, + 639 + ], + "type": "text", + "content": "LiDAR-Camera fusion-based 3D detection is a critical task for automatic driving. In recent years, many LiDAR-Camera fusion approaches sprung up and gained promising performances compared with single-modal detectors, but always lack carefully designed and effective supervision for the fusion process. In this paper, we propose a novel training strategy called SupFusion, which provides an auxiliary feature level supervision for effective LiDAR-Camera fusion and significantly boosts detection performance. Our strategy involves a data enhancement method named Polar Sampling, which densifies sparse objects and trains an assistant model to generate high-quality features as the supervision. These features are then used to train the LiDAR-Camera fusion model, where the fusion feature is optimized to simulate the generated high-quality features. Furthermore, we propose a simple yet effective deep fusion module, which contiguously gains superior performance compared with previous fusion methods with SupFusion strategy. In such a manner, our proposal shares the following advantages. Firstly, SupFusion introduces auxiliary feature-level supervision which could boost LiDAR-Camera detection performance without introducing extra inference costs. Secondly, the proposed deep fusion could continuously improve the detector's abilities. Our proposed SupFusion and deep fusion module is plug-and-play, we make extensive experiments to demonstrate its effectiveness. Specifically, we gain around " + }, + { + "bbox": [ + 45, + 289, + 290, + 639 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 45, + 289, + 290, + 639 + ], + "type": "text", + "content": " 3D mAP improvements on KITTI benchmark based on multiple LiDAR-Camera 3D detectors. Our code is available at https://github.com/IranQin/SupFusion." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 256, + 544, + 339 + ], + "blocks": [ + { + "bbox": [ + 310, + 256, + 544, + 339 + ], + "lines": [ + { + "bbox": [ + 310, + 256, + 544, + 339 + ], + "spans": [ + { + "bbox": [ + 310, + 256, + 544, + 339 + ], + "type": "image", + "image_path": "54814ff1800592069e17aa0443576449f7f4e859af0ae525027806812b972f98.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 342, + 544, + 466 + ], + "blocks": [ + { + "bbox": [ + 310, + 342, + 544, + 466 + ], + "lines": [ + { + "bbox": [ + 310, + 342, + 544, + 466 + ], + "spans": [ + { + "bbox": [ + 310, + 342, + 544, + 466 + ], + "type": "image", + "image_path": "9ff36b3f8ea3bf9970e14f08d8bec410d40d31cd7ca77f7da2f65c9131bf1e89.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 484, + 548, + 534 + ], + "lines": [ + { + "bbox": [ + 305, + 484, + 548, + 534 + ], + "spans": [ + { + "bbox": [ + 305, + 484, + 548, + 534 + ], + "type": "text", + "content": "Figure 1: Top: The previous LiDAR-Camera 3D detector, the fusion module is optimized by detection loss. Bottom: Our proposed SupFusion, we introduce auxiliary supervision via high-quality features provided by an assistant model." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 559, + 387, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 387, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 387, + 571 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 581, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 548, + 715 + ], + "type": "text", + "content": "LiDAR-Camera fusion-based 3D object detection is a critical and challenging task for autonomous driving and robotics [32, 35]. Previous approaches [20, 22, 34, 19] always project the camera inputs to LiDAR BEV or voxel space via intrinsic and extrinsic to align the LiDAR and camera features. Then, simple concatenation or summation is employed to obtain the fusion features for final detection. Also, some deep learning-based fusion approaches [1, 21] sprung up and gained promising performances. However, the previous fusion approaches always directly optimized the 3D/2D feature extraction as well as the fusion module by the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 255, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 255, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 255, + 703 + ], + "type": "text", + "content": "*Equal contribution. Work done during an internship at NIO." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22014" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 95 + ], + "type": "text", + "content": "detection loss, which lacks carefully designed and effective supervision at the feature level and limits its performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 251 + ], + "type": "text", + "content": "In recent years, distillation manners have shown great improvements in feature-level supervision for 3D detection. For instance, some approaches [5, 6] provide LiDAR features to guide the 2D backbone in estimating the depth information from camera inputs. Also, some approaches [19] provide LiDAR-Camera fusion features to supervise LiDAR backbone learns global and contextual presentation from LiDAR inputs. Introducing feature-level auxiliary supervision via simulating more robust and high-quality features, the detectors could boost marginal improvements. Motivated by this, a natural solution to handle LiDAR-Camera feature fusion is providing stronger and high-quality features and introducing auxiliary supervision for LiDAR-Camera 3D detection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 288, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 609 + ], + "type": "text", + "content": "To this end, in this paper we propose a supervised LiDAR-Camera fusion method named SupFusion, to generate high-quality features and provide effective supervision for the fusion as well as feature extraction process, and further boost the LiDAR-Camera fusion-based 3D detection performances. Specifically, we first train an assistant model to provide high-quality features. To achieve it, different from previous approaches, which leverage larger models or extra data, we proposed a novel data enhancement method named Polar Sampling. The polar sampling could dynamically enhance the object's density from sparse LiDAR data, which is easier to detect and improve features quality, e.g., the feature could conclude accuracy detection results. Then, we simply train the LiDAR-Camera fusion-based detector with introduced auxiliary feature-level supervision. In this step, we feed the raw LiDAR and camera inputs into 3D/2D backbones and fusion modules to obtain fusion features. In one aspect, the fusion features are fed into the detection head for final prediction, which is decision-level supervision. In another aspect, the auxiliary supervision mimics the fusion feature to high-quality features, which are obtained via the pretrained assistant model and enhanced LiDAR data. In such a manner, the proposed feature-level supervision could lead the fusion module to generate more robust features, and further boost the detection performance. To better fuse the LiDAR and camera features, we propose a simple yet effective deep fusion module, which consists of stacked MLP blocks and dynamic fusion blocks. SupFusion could sufficiently excavate the deep fusion module capacity, and continuously improve detection accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 610, + 288, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 288, + 670 + ], + "type": "text", + "content": "Compared with previous LiDAR-Camera 3D detectors, this is a brand new attempt to improve the effectiveness of the LiDAR-Camera fusion by introducing feature-level supervision, which consistently boosts 3D detection accuracy. The main contributions are summarized as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 288, + 715 + ], + "type": "text", + "content": "- We propose a novel supervised fusion training strategy named SupFusion, which mainly consists of a high-quality feature generation process and firstly propose" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 324, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 109 + ], + "type": "text", + "content": "auxiliary feature-level supervision loss for robust fusion feature extraction and accuracy 3D detection to our knowledge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 116, + 547, + 232 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 116, + 547, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 116, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 317, + 116, + 547, + 177 + ], + "type": "text", + "content": "- To obtain high-quality features in SupFusion, we propose a data enhancement method named Polar Sampling to densify the sparse objects. Furthermore, we propose an efficient deep fusion module to contiguously boost detection accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 183, + 547, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 183, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 183, + 547, + 232 + ], + "type": "text", + "content": "- We make extensive experiments based on multiple detectors with different fusion strategies, and gain around " + }, + { + "bbox": [ + 317, + 183, + 547, + 232 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 317, + 183, + 547, + 232 + ], + "type": "text", + "content": " mAP improvements on KITTI benchmark. The source code will be released after the blind review." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 243, + 392, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 392, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 392, + 255 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 262, + 424, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 424, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 424, + 275 + ], + "type": "text", + "content": "2.1. 3D Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 281, + 547, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 399 + ], + "type": "text", + "content": "Camera-based 3D Object Detection methods perform 3D detection on single- or multi-view images. Many methods are proposed to address monocular 3D detection[24, 23, 17, 51, 53, 27, 37, 36] since KITTI benchmark[9] only contains one from the camera. With the development of autonomous driving datasets which have more cameras covering a wider range of angles(nuScenes[2] and Waymo[33]), more and more methods[38, 40, 42] are developed that take multi-view images as input and get surprising performance compared to monocular methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 400, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 579 + ], + "type": "text", + "content": "LiDAR-based 3D Object Detection methods perform 3D detection on a point cloud scene. Depending on the three kinds of point cloud data pretreatments, 3D detectors are also classified into three mainstream. Point-based method [31, 46, 50, 4] use row point cloud data, leveraging pointnet[25] and pointnet++[26] as 3D backbone networks to get point-wised features which are sent to point-wise 3D detection head. Pillar-based method [18, 41, 29] pretreat point cloud to pillars, gain speed up while still maintaining a good performance. Voxel-based methods[54, 45, 8] are widely used, point clouds are voxelized to voxel and sent to a 3D sparse convolutional[10] based 3D backbone to get bird's-eye-view features. Some methods[30] also try to gain more improvement by using both the voxel feature and the point feature." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 586, + 501, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 501, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 501, + 600 + ], + "type": "text", + "content": "2.2. LiDAR-Camera 3D Object Detection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "LiDAR-camera fusion is a technique that combines data from LiDAR and camera sensors to obtain a more complete understanding of a scene in 3D. Due to the complementary nature of the features produced by these two modalities, researchers have developed methods that can be jointly optimized on both sensors. These methods can be classified into two categories based on their fusion mechanism. The first category is point-level fusion[14, 35, 47], where the image features are queried via the raw LiDAR points and then" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22015" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 542, + 231 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 542, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 542, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 542, + 231 + ], + "type": "image", + "image_path": "57afa910eb861de1fe39472dea66f52a4aaa28851ec7bb490604103f32704b56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "text", + "content": "Figure 2: The detector training procedure. For a given LiDAR-Camera sample, we extract the feature presentation via the corresponding 3D/2D encoders, and the proposed deep fusion module is employed to fuse the multi-modal features. Besides the detection loss " + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Det}" + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "text", + "content": ", we introduce the auxiliary feature-level supervision " + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Sim}" + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "text", + "content": ", to mimic the fusion feature " + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{lc}" + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "text", + "content": " to the high-quality features " + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l^*" + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "text", + "content": ", which is generated by the assistant model and enhanced data. Here the " + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 248, + 548, + 309 + ], + "type": "text", + "content": " indicate the element-wise multiplication, summation, and channel-wise concatenation. More details could be referred in Sec. 3.1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 330, + 289, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 289, + 498 + ], + "type": "text", + "content": "concatenated back as additional point features. The second category is feature-level fusion, where the LiDAR points are projected into a specific feature space[48, 20, 19] or proposals level feature fusion[3, 16, 1, 43]. There are also methods that combine these two fusion manners, such as MVX-Net [32]. Recently, attention-based fusion methods have sprung up. TransFusion[1] uses the bounding box prediction of LiDAR features as a proposal to query the image feature, and then a Transformer-like architecture fuses the information back to LiDAR features. Similarly, DeepFusion[21] projects LiDAR features on each view image as queries and leverages cross-attention for the two modalities. Bird's-eye-view space feature fusion[22, 34] is proposed recently and gained great performance with different tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 503, + 288, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 288, + 517 + ], + "type": "text", + "content": "2.3. Knowledge Distillation for 3D Object Detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": "Object detection tasks differ from classification tasks in that they require consideration of both global and local features. While numerous distillation methods have been proposed to address this challenge, such as those developed by [39, 7, 11, 49], many of these methods require the creation of specific distillation regions or objects based on ground-truth data, bounding box predictions, attention, or gradients to mitigate the imbalance between foreground and background. In addition to 2D detection, there has been growing interest in cross-model knowledge distillation methods between RGB and LiDAR in 3D object detection. For example, [12, 28, 6, 5] have developed methods that transfer knowledge from LiDAR detectors to image detectors. [52, 15] has trained a teacher model similar to that proposed by [35] and developed a multilevel distillation framework to transfer semantic segmentation information to a student model. In this" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 545, + 354 + ], + "type": "text", + "content": "work, we want to further boost LiDAR-Camera detectors with an auxiliary feature level supervision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 366, + 389, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 389, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 389, + 380 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 387, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 547, + 434 + ], + "type": "text", + "content": "In this section, we first overview the overall SupFusion pipeline in Sec. 3.1, and then detail introduces the main components, including the proposed Polar Sampling in Sec. 3.2, and deep fusion module in Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 443, + 457, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 457, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 457, + 456 + ], + "type": "text", + "content": "3.1. Overall SupFusion Pipeline" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": "For a given LiDAR-Camera fusion-based 3D detector " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": ", which consists of a 3D Encoder " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_l" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": ", a 2D Encoder " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_c" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": ", a fusion module " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": ", and a detection head. Denotes " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\left(\\mathbf{l}_1,\\mathbf{c}_1\\right),\\left(\\mathbf{l}_2,\\mathbf{c}_2\\right),\\dots,\\left(\\mathbf{l}_N,\\mathbf{c}_N\\right)\\}" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\mathbf{y}_1,\\mathbf{y}_2,\\dots,\\mathbf{y}_N\\}" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": " are the given dataset with " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": " samples, where " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{l}_i,\\mathbf{c}_i)" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": " are the paired LiDAR and camera inputs. The previous approaches feed the paired " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{l}_i,\\mathbf{c}_i)" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": " into 3D/2D encoders " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_l" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_c" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": " to capture corresponding feature presentation " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^l" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^c" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": ". Then, the paired features presentation is fed into the fusion module to obtain the fusion feature " + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{lc}" + }, + { + "bbox": [ + 305, + 462, + 547, + 715 + ], + "type": "text", + "content": ", which is used for the final prediction. To better leverage the LiDAR and camera data, we propose a supervised LiDAR-Camera Fusion training strategy named SupFusion, which provides high-quality features for feature-level supervision. To end this, different from previous approaches, which adopt larger scale models or extra data input, we propose a data enhancement method named polar sampling to densify the sparse objects in LiDAR data. In such a manner, the data is easier and could easily generate high-quality features (e.g., the feature could conclude accuracy detection results) with lightweight models. The proposed polar sampling mainly" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22016" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 542, + 281 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 542, + 281 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 542, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 542, + 281 + ], + "type": "image", + "image_path": "8a3162c7f7ddf1873005e730012f3e3b23b8cb2001fef0969b259e2889da6775.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 299, + 546, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 546, + 348 + ], + "type": "text", + "content": "Figure 3: High-quality feature generation process. For any given LiDAR sample, we densify the sparse objects via polar pasting, which calculate the direction and rotation to query dense object from the database, and pasting to add extra points to sparse objects. We first train the assistant model via the enhanced data and feed the enhanced LiDAR data into the assistant model to generate high-quality features " + }, + { + "bbox": [ + 46, + 299, + 546, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^*" + }, + { + "bbox": [ + 46, + 299, + 546, + 348 + ], + "type": "text", + "content": " after it is convergent. More details could be referred in Sec. 3.1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "text", + "content": "densifies the sparse object points in LiDAR data. Specifically, for any sample " + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "inline_equation", + "content": "(\\mathbf{l}_i,\\mathbf{c}_i)" + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i = \\{l_i^1,l_i^2,\\dots,l_i^{n_i},l_i^s\\}" + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "text", + "content": " objects points sets and one background points set. In the polar sampling, we add extra points to each object point set from raw LiDAR data, to obtain enhanced LiDAR data " + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i' = \\{(l_i^1,\\hat{l}_i^1),(l_i^2,\\hat{l}_i^2),\\dots,(l_i^{n_i},\\hat{l}_i^{n_i}),l_i^s\\}" + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "text", + "content": " in polar pasting, where " + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "inline_equation", + "content": "\\hat{l}_i^k" + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "text", + "content": " is the added points set for the " + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "text", + "content": "-th objects in " + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 368, + 289, + 478 + ], + "type": "text", + "content": "-th LiDAR sample and generated in polar grouping. We will detail introduce the polar sampling in Sec. 3.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "text", + "content": "High-quality Feature Generation. To provide feature-level supervision in our SupFusion, we adopt an assistant model to capture high-quality features from the enhanced data as shown in Fig. 3. Firstly, we train an assistant model to provide high-quality features. Specifically, for any sample in " + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "text", + "content": ", we enhance the sparse LiDAR data " + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i" + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "text", + "content": " to obtain enhanced data " + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i^{\\prime}" + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "text", + "content": " via the polar pasting, which densifies the spare objects with added points set generated in polar grouping. Then, after the assistant model convergence (often with few epochs), we feed the enhanced sample " + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i^{\\prime}" + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "text", + "content": " into the optimized assistant model, to capture high-quality features " + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^*" + }, + { + "bbox": [ + 46, + 482, + 288, + 661 + ], + "type": "text", + "content": ", which is used to train the LiDAR-Camera 3D detector. To better apply to the given LiDAR-Camera detector and easier implement, we simply adopt the LiDAR branch detector as the assistant model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "Detector Training. For any given LiDAR-Camera detector, we train the model with the proposed auxiliary supervision at the feature level. Specifically, given sample " + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{l}_i,\\mathbf{c}_i)\\in \\mathcal{D}" + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": ", we first feed the LiDAR and camera into the 3D" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "text", + "content": "and 2D Encoders " + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_l" + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_s" + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "text", + "content": " to capture the corresponding features " + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^l" + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^c" + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "text", + "content": ", which are fed into the fusion model to generate fusion features " + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{l,c}" + }, + { + "bbox": [ + 304, + 368, + 547, + 463 + ], + "type": "text", + "content": ", and flow into the detection head for final prediction. Furthermore, the proposed auxiliary supervision is employed to mimic the fusion feature with high-quality features, which is generated by the pre-trained assistant model and enhanced LiDAR data. The above process can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 320, + 473, + 546, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 473, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 320, + 473, + 546, + 510 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {((\\mathbf {l} _ {i}, \\mathbf {c} _ {i}), \\mathbf {y} _ {i}) \\sim \\mathcal {D}} \\mathcal {L} _ {\\text {d e t}} (\\mathbf {D} (\\mathbf {l} _ {i}, \\mathbf {c} _ {i}, \\theta), \\mathbf {y} _ {i}) (1) \\\\ + \\lambda \\mathcal {L} _ {\\text {S i m}} \\left(\\mathbf {f} _ {i} ^ {l c}, \\mathbf {f} _ {i} ^ {*}\\right), (2) \\\\ \\end{array}", + "image_path": "bb9474fe6449e6d3534db49d631c72245845dd24c4781a7d2c495bed18cc1f28.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 333, + 513, + 546, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 513, + 546, + 526 + ], + "spans": [ + { + "bbox": [ + 333, + 513, + 546, + 526 + ], + "type": "interline_equation", + "content": "s. t. \\quad \\mathbf {f} _ {i} ^ {l c} = \\mathbf {F} \\left(\\mathbf {E} _ {l} \\left(\\mathbf {l} _ {i}, \\theta_ {l}\\right), \\mathbf {E} _ {c} \\left(\\mathbf {c} _ {i}, \\theta_ {c}\\right), \\theta_ {f}\\right), \\tag {3}", + "image_path": "0c8c23a94e436306a7dac0dc93e0a91e3cc1bd62a45c044d1f2f66e96b9616f5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 369, + 528, + 545, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 528, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 369, + 528, + 545, + 541 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {i} ^ {*} = \\mathbf {E} _ {l} \\left(\\mathbf {I} _ {i} ^ {\\prime}, \\theta_ {l} ^ {*}\\right), \\tag {4}", + "image_path": "ed65144b87fd7c3d5e4e680dffeddcc70ea74fe9dedff122c8742b30b6c3f3bf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 354, + 544, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 544, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 354, + 544, + 545, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S i m} = \\left\\| \\psi \\left(\\mathbf {f} _ {i} ^ {l c}\\right) - \\mathbf {f} _ {i} ^ {*} \\right\\| _ {2} ^ {2} \\tag {5}", + "image_path": "f6bbb2a2769c38bc83e0f12d78f0edda48adf4c46eb007329e42d26fd1eee100.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "text", + "content": " is the optimized parameters of detector " + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^*" + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "text", + "content": " is the high-quality features, which is generated from pre-trained assistant model LiDAR encoder " + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_l" + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "inline_equation", + "content": "\\theta_l^*" + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "text", + "content": " is a simple convolutional layer with kernel size 1. " + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 567, + 546, + 627 + ], + "type": "text", + "content": " is the balanced weight and we set it as 1.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 635, + 402, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 402, + 648 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 402, + 648 + ], + "type": "text", + "content": "3.2. Polar Sampling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "To provide high-quality features via the assistant model in our proposed SupFusion, we propose a novel data enhancement method named Polar Sampling, to ease the sparse problem which frequently leads to the detection failers. To end this, we densify the sparse objects in LiDAR data with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22017" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "a similar dense object. Specifically, the polar sampling consists of two parts, including polar grouping and polar pasting. In polar grouping, we mainly build a database to store the dense objects, which is used in polar pasting to densify the sparse objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 290, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 290, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 290, + 359 + ], + "type": "text", + "content": "Considering the characteristics of LiDAR sensors, the collected point cloud data naturally exists a particular density distribution, e.g., the objects have more points on the surface towards the LiDAR sensor, while few points on the opposite sides. The density distribution is mainly influenced by the direction and rotation, while the points' density mainly depends on the distances, e.g. the objects closer to the LiDAR sensor have denser points. Motivated by this, we aim to densify the sparse objects in longer distances with the dense objects in short distances, according to their direction and rotation to keep the density distributions. Specifically, we build polar coordinate systems for the whole scene as well as the objects based on the center of the scene and specific objects and define the positive direction of the LiDAR sensor as 0 degrees to measure the corresponding directions and rotations. Then, we collect the objects with similar density distribution (e.g. with similar direction and rotation) and generate one dense object for each group in the polar grouping, and used it to densify the sparse objects in polar pasting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "type": "text", + "content": "Polar Grouping. As shown in Fig. 4, we build a database " + }, + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "type": "text", + "content": " to store the generated dense object point set " + }, + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\hat{l}" + }, + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "type": "text", + "content": " according to their directions and rotations in the polar grouping, which is noted as " + }, + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 360, + 288, + 407 + ], + "type": "text", + "content": " in Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "content": "Firstly, we search the whole dataset and calculated their polar angles for all objects via the locations and provided rotations in the benchmarks. Secondly, we split the objects into groups according to their polar angles. Specifically, we manually divide the direction and rotation into " + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "content": " groups. For any object point set " + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 407, + 287, + 491 + ], + "type": "text", + "content": ", we could put it into the corresponding group according to the index:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 497, + 287, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 497, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 95, + 497, + 287, + 520 + ], + "type": "interline_equation", + "content": "\\alpha_ {l} ^ {\\prime} = \\left\\lfloor \\frac {\\alpha_ {l} \\cdot 2 \\pi}{N} \\right\\rfloor , \\quad \\beta_ {l} ^ {\\prime} = \\left\\lfloor \\frac {\\beta_ {l} \\cdot 2 \\pi}{N} \\right\\rfloor , \\tag {6}", + "image_path": "45d247c2dade87aaeaa744e81dba47e6374e8f73b305678838d228e2a8e33e1d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{\\prime},\\beta_{l}^{\\prime}\\in \\{1,2,\\dots,N\\}" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " are the direction and rotation index of object " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " , and we could put the points set into " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_C^{\\alpha_l',\\beta_l'}\\gets l" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " is the label of object " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " . Finally, for each object group " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_C^{\\alpha ',\\beta '} = \\{l_1,l_2,\\ldots ,l_n\\}" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " similar objects, we select the densest " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " objects and mix up to generate one dense and complete object " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{l}" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_C^{\\alpha ',\\beta '} = \\{\\hat{l}\\}" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " is used in polar pasting phase. To keep the density and avoid too much calculation in the model, we randomly select partial points in the dense object. Here we set " + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 46, + 527, + 287, + 653 + ], + "type": "text", + "content": " and select 5000 points for all objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "Polar Pasting. As shown in Fig. 2, the polar pasting is utilized to enhance the sparse LiDAR data to train the assistant model and generate high-quality features. Concretely, given a LiDAR sample " + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_i = \\{l_i^1, l_i^2, \\dots, l_i^{n_i}, l_i^s\\}" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": " objects, for any object " + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "l_i^k" + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": ", we can calculate the direction" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 542, + 217 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 542, + 217 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 542, + 217 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 542, + 217 + ], + "type": "image", + "image_path": "63df54b7c17ffde47cf418c685d34796aeec5bfef97b2c2010b4724f5e142881.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 228, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 275 + ], + "type": "text", + "content": "Figure 4: Polar grouping mainly builds a database to store the dense objects of each group in polar coordinate systems of direction and rotations for each class. More details could be referred to in Sec. 3.2." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 299, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 547, + 361 + ], + "type": "text", + "content": "and rotation the same as the grouping process and query the dense object " + }, + { + "bbox": [ + 304, + 299, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\hat{l}_i^k" + }, + { + "bbox": [ + 304, + 299, + 547, + 361 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 299, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 299, + 547, + 361 + ], + "type": "text", + "content": " according to the label and indexes, which could be obtained from E.q. 6 In such a manner, we enhance all the objects in the sample and obtain the enhanced data: " + }, + { + "bbox": [ + 304, + 299, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i' = \\{(l_i^1,\\hat{l}_i^1),(l_i^2,\\hat{l}_i^2),\\dots,(l_i^{n_i},\\hat{l}_i^{n_i}),l_i^s\\}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 367, + 388, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 367, + 388, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 367, + 388, + 380 + ], + "type": "text", + "content": "3.3. Deep Fusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 386, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 547, + 602 + ], + "type": "text", + "content": "To mimic the high-quality features generated from enhanced LiDAR data, the fusion model aims to extract the missing information of the sparse objects from the rich color and contextual features in camera input. To end this, we propose the deep fusion module, to leverage image features and complete the LiDAR presentation. The proposed deep fusion mainly consists of a 3D learner and a 2D3D learner. The 3D learner is a simple convolutional layer, to transfer the 3D presentation into 2D space. Then, concatenating the 2D features and 3D presentation (e.g. in 2D space), the 2D3D learner is employed to fuse the LiDAR-Camera features. Finally, we weighed the fusion feature via an MLP and an activate function, adding back to the raw LiDAR features as the output of the deep fusion module. The 2D3D learner consists of stacked MLP blocks with depth " + }, + { + "bbox": [ + 304, + 386, + 547, + 602 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 386, + 547, + 602 + ], + "type": "text", + "content": ", learning to leverage camera features to complete the LiDAR representation from the sparse objects, to simulate the high-quality features of dense LiDAR objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 614, + 388, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 614, + 388, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 614, + 388, + 628 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 634, + 428, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 428, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 428, + 647 + ], + "type": "text", + "content": "4.1. Experiments Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. KITTI [9] is a large-scale public benchmark for autonomous driving, including 3712 training, 3769 validation, and 7518 testing samples with LiDAR and camera pair data. All objects in KITTI are divided into 'easy', 'moderate', and 'hard' difficulties ac" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22018" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 133, + 552, + 408 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 131 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 131 + ], + "type": "text", + "content": "Table 1: The experiment results (mAP@R40 %) on KITTI validation set. We list the easy, moderate (mod.), and hard cases for three classes, and the overall performances. Here, L, LC, LC* indicate the results of the corresponding LiDAR detector, LiDAR-Camera fusion detectors, and our proposal. " + }, + { + "bbox": [ + 46, + 70, + 547, + 131 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 70, + 547, + 131 + ], + "type": "text", + "content": " indicates the improvements. The best results are shown in bold expect " + }, + { + "bbox": [ + 46, + 70, + 547, + 131 + ], + "type": "inline_equation", + "content": "\\mathrm{L}^{\\dagger}" + }, + { + "bbox": [ + 46, + 70, + 547, + 131 + ], + "type": "text", + "content": ", which is the assistant model, and test on the enhanced validation set. MVXNet is re-conduct based on mmdetection3d. PV-RCNN-LC and Voxel-RCNN-LC are re-conduct based on the open-source code of VFF [20]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 133, + 552, + 408 + ], + "lines": [ + { + "bbox": [ + 47, + 133, + 552, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 552, + 408 + ], + "type": "table", + "html": "
DetectorModalityOverallPedestrianCyclistCar
easymod.hardeasymod.hardeasymod.hardeasymod.hard
MVXNetL†88.4490.9388.5887.3387.6180.1580.0986.0186.1997.8999.1799.41
L73.4658.6455.1557.5450.1045.5674.6351.8148.9288.2173.9970.98
LC74.9963.0059.3264.4657.7653.0572.6955.2751.6587.8275.9873.26
LC*75.3764.7360.5263.1158.1353.7374.7957.6053.8288.2078.4674.01
Δ+0.38+1.73+1.20-1.35+0.37+0.68+2.10+2.33+2.17+0.38+2.48+0.75
PV-RCNN-LCL†94.9095.2095.1086.8686.7887.0798.9099.3398.6298.9599.5099.62
L81.3771.5568.0867.8059.5654.4085.0670.8367.9691.2684.2581.88
LC84.1674.5470.8670.4564.1659.5090.4774.4670.2591.5685.0182.83
LC*85.7776.0072.0074.2966.6761.3891.0076.1371.6692.0285.2082.94
Δ+1.61+1.46+1.14+3.84+2.51+1.88+0.53+1.67+1.41+0.46+0.19+0.11
Voxel-RCNN-LCL†99.2699.6699.7597.8498.0598.1794.3694.6094.8599.9099.9099.91
L82.0171.6968.4067.8460.9556.0386.1071.3967.0592.0982.7482.12
LC84.3674.4670.8071.3964.8960.2089.6472.9469.0392.0585.5583.17
LC*86.1375.9072.1773.9566.0661.8392.0276.1871.8392.4285.4782.83
Δ+1.77+1.44+1.37+2.56+1.17+1.63+2.38+3.24+2.80+0.37-0.08-0.34
", + "image_path": "abd8761755900a4921d9c94a5dbbe1babd19a55af263c4ccb88b0e18968a65c2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 430, + 288, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 288, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 288, + 596 + ], + "type": "text", + "content": "cording to the object scales and occlusion ratio. We evaluate the detection performance by the mean Average Precision under 40 recall thresholds (mAP@R40) the same as the official benchmark and evaluation for 3 classes including Pedestrian, Cyclist, and Car. nuScenes [2] is a large-scale autonomous driving dataset for 3D detection and tracking tasks, including 700 training, 150 validation, and 150 testing sequences, with each 20s long and annotated with 3D bounding boxes. For 3D object detection, the official evaluation metrics include the mean Average Precision (mAP) and nuScenes detection score (NDS). mAP measures the localization precision using a threshold base on the birds-eye view. NDS is a weighted combination of mAP and regression accuracy of other object attributes, including box size, orientation, and translation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Detectors. To demonstrate the effectiveness of our proposed method, we apply the SupFusion training strategy and deep fusion module on variance LiDAR-Camera detectors including MVXNet [32], PV-RCNN-LC [30], and Voxel-RCNN-LC [8] on KITTI, and SECOND-LC [45] on nuScenes benchmark. MVXNet is a voxel-based LiDAR-Camera 3D detector based on VoxelNet [54]. The camera features project to LiDAR space and concatenates with points features. Voxel-RCNN and PV-RCNN are LiDAR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 430, + 547, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 537 + ], + "type": "text", + "content": "based detectors, we re-conduct the LiDAR-Camera fusion approach following VFF [20] based on the open-source code, which is a previous state-of-the-art fusion approach based on the two detectors, mainly projects the camera features into voxel presentation and summarized fusion with the voxel features of LiDAR input. SECOND-LC introduces camera inputs based on the LiDAR-based SECOND detector, the same as BEVFusion [22], we concatenate the camera and LiDAR features in birds-eys-view spaces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": "Implement Details. We conduct experiments on MVXNet, PV-RCNN-LC and Voxel-RCNN-LC, SECOND-LC following the open-sourced code and mmdetection3d on 8 Tesla A100 GPUs. All the architecture, configuration, and optimizer follow the settings reported in papers including voxel size, point cloud range, batch size, and learning rate. We modify the fusion method (e.g. concatenation, summarized) by our proposed deep fusion module and applied our SupFusion training strategy. For the deep fusion module, we set depth " + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": ". For the auxiliary supervision, we utilize simple L2 loss and set " + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": ". More details about how to apply SupFusion and deep fusion modules on different detectors are demonstrated in the Appendix." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22019" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 84, + 288, + 150 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 286, + 82 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 286, + 82 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 286, + 82 + ], + "type": "text", + "content": "Table 2: NDS and mAP(%) results on nuScenes benchmarks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 84, + 288, + 150 + ], + "lines": [ + { + "bbox": [ + 48, + 84, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 84, + 288, + 150 + ], + "type": "table", + "html": "
Modality3D2DNDSMAP
L†SECOND-90.3894.44
LSECOND-60.5252.28
LCSECONDSwin-T62.5755.26
LC*SECONDSwin-T64.58(+2.01)56.64(+1.38)
", + "image_path": "ef883c62103f4fca0d862bb9e9eb41c970fc371bf7c9b56850510705c2d18664.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 205, + 283, + 297 + ], + "blocks": [ + { + "bbox": [ + 47, + 167, + 287, + 203 + ], + "lines": [ + { + "bbox": [ + 47, + 167, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 287, + 203 + ], + "type": "text", + "content": "Table 3: Results comparison with different fusion methods. Sup. indicates our proposed feature-level supervision. Ped., Cyc., and O.a. are short of Pedestrian, Cyclist, and Overall." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 205, + 283, + 297 + ], + "lines": [ + { + "bbox": [ + 49, + 205, + 283, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 205, + 283, + 297 + ], + "type": "table", + "html": "
MethodSup.Ped.Cyc.CarO.a.Δ
SumX64.8972.9485.5574.46-
✓67.0374.6185.1175.59+1.13
ConcatX65.9072.8083.3774.02-
✓66.2374.9585.4275.52+1.50
OursX63.1872.8485.5873.86-
✓66.0676.1885.4775.90+2.04
", + "image_path": "8371bc60cfbf5dab3e636adbd3a69be5dff4d9707cedfd245c2f7f3859356671.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 321, + 132, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 321, + 132, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 132, + 333 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "text", + "content": "Overall Performances. The 3D mAP@R40 comparison based on three detectors is shown in Tab. 1 for three classes and overall performances for each difficulty split. We can clearly observe that the LiDAR-Camera approaches (LC) are superior to the LiDAR-based detectors (L) by introducing extra camera inputs. By introducing polar sampling, the assistant model " + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "inline_equation", + "content": "(\\mathrm{L}^{\\dagger})" + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "text", + "content": " could gain admirable performances (e.g. over " + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "text", + "content": " mAP) on the enhanced validation set. With the auxiliary supervision from high-quality features and our proposed deep fusion module, our proposal could contiguously boost detection accuracy. For instance, compared with the baseline (LC) model, our proposal could gain " + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "inline_equation", + "content": "+1.54\\%" + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "inline_equation", + "content": "+1.24\\%" + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "text", + "content": " 3D mAP improvements for the moderate and hard objects on average for the three detectors. Further, we also conduct experiments on nuScenes benchmark based on SECOND-LC as shown in Tab. 2, which gain " + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "inline_equation", + "content": "+2.01\\%" + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "inline_equation", + "content": "+1.38\\%" + }, + { + "bbox": [ + 47, + 341, + 287, + 568 + ], + "type": "text", + "content": " improvements for NDS and mAP. Also, we will prune the SupFusion training strategy and deep fusion module in Sec. 4.3 to detail discuss their effectiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Class-Aware Improvements Analysis. Compared with the baseline model, our SupFusion and deep fusion could boost the detection performances not only for the overall performances but also for each class including Pedestrian, Cyclist, and Car. Comparing the average improvements (e.g. moderate cases) for three classes, we can obtain the following observation: the cyclist gains the largest improvements " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "(+2.41\\%)" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " while pedestrian and car gains " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "+1.35\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "+0.86\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " increasements respectively. The reason is obvious: (1) The car is easier to detect and gains the best results compared with pedestrians and cyclists, hence it is more difficult to improve. (2) Cyclist gains more improvements against" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 87, + 529, + 142 + ], + "blocks": [ + { + "bbox": [ + 306, + 62, + 547, + 86 + ], + "lines": [ + { + "bbox": [ + 306, + 62, + 547, + 86 + ], + "spans": [ + { + "bbox": [ + 306, + 62, + 547, + 86 + ], + "type": "text", + "content": "Table 4: Results comparison with different supervision loss. " + }, + { + "bbox": [ + 306, + 62, + 547, + 86 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Sim}" + }, + { + "bbox": [ + 306, + 62, + 547, + 86 + ], + "type": "text", + "content": " indicates the auxiliary supervision loss in Eq. 1." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 87, + 529, + 142 + ], + "lines": [ + { + "bbox": [ + 320, + 87, + 529, + 142 + ], + "spans": [ + { + "bbox": [ + 320, + 87, + 529, + 142 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{Sim} \\)Ped.Cyc.CarO.a.\\( \\Delta \\)
None64.8972.9485.5574.46-
L165.7675.5785.4275.58+1.12
L266.0676.1885.4775.90+1.44
", + "image_path": "1e9e49c9855715b01d2c53c18c5e074222e0c6898dc94875c76add60653d0a9e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 319, + 171, + 531, + 213 + ], + "blocks": [ + { + "bbox": [ + 315, + 157, + 536, + 169 + ], + "lines": [ + { + "bbox": [ + 315, + 157, + 536, + 169 + ], + "spans": [ + { + "bbox": [ + 315, + 157, + 536, + 169 + ], + "type": "text", + "content": "Table 5: Results with different depth " + }, + { + "bbox": [ + 315, + 157, + 536, + 169 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 315, + 157, + 536, + 169 + ], + "type": "text", + "content": " in deep fusion." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 171, + 531, + 213 + ], + "lines": [ + { + "bbox": [ + 319, + 171, + 531, + 213 + ], + "spans": [ + { + "bbox": [ + 319, + 171, + 531, + 213 + ], + "type": "table", + "html": "
K1234
LC74.2374.0073.8673.22
LC*75.6475.7575.9075.88
", + "image_path": "267b20592852676054f3940296544a1c92aecf0a12da24ce9dba44ffe8ffeeaf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 319, + 241, + 531, + 283 + ], + "blocks": [ + { + "bbox": [ + 320, + 228, + 529, + 240 + ], + "lines": [ + { + "bbox": [ + 320, + 228, + 529, + 240 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 529, + 240 + ], + "type": "text", + "content": "Table 6: Results with different " + }, + { + "bbox": [ + 320, + 228, + 529, + 240 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 320, + 228, + 529, + 240 + ], + "type": "text", + "content": " in polar sampling." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 241, + 531, + 283 + ], + "lines": [ + { + "bbox": [ + 319, + 241, + 531, + 283 + ], + "spans": [ + { + "bbox": [ + 319, + 241, + 531, + 283 + ], + "type": "table", + "html": "
K14816
L†99.8099.6599.6699.43
LC*75.6875.7275.9075.85
", + "image_path": "cb2892380ab029889e33aa1f3765ff2b93ce702cadd0be5ea33825c666f25280.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 300, + 545, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 336 + ], + "type": "text", + "content": "Pedestrian because pedestrians are non-grid, the generated dense object is not so well compared with Cyclist, hence gains lower performance improvements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 346, + 400, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 400, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 400, + 358 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 365, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 400 + ], + "type": "text", + "content": "To analyze the effectiveness of our proposed SupFusion and deep fusion module, we conduct ablation experiments based on Voxel-RCNN-LC on the KITTI benchmark." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": "Supervision. In SupFusion training strategy, we mainly introduce a high-quality guided feature-level supervision and simply employ L2 loss as the auxiliary loss. Here we also adopt L1 loss and the results are shown in Tab. 4, from which we can observe that L1 loss gains comparable but slight improvements (e.g. " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "+1.12\\%" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": " versus " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "+1.44\\%" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": "). To mimic the high-quality features by the fusion features, utilizing L1 and L2 to measure the feature distance is acceptable [13, 44]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": "Fusion Strategy. We conduct experiments with different fusion methods, including the concatenation and summation approaches employed in MVXNet and PV-RCNN-LC/Voxel-RCNN-LC, as well as our proposed deep fusion methods. As shown in Tab. 3, without auxiliary supervision in SupFusion, our proposed deep fusion and concatenation fusion gain comparable but lower performances, demonstrating introducing extra parameters in the fusion model would lead the model difficult to train and hard to optimize. By introducing auxiliary supervision, our proposal could gain the best performance and improvements compared with the baseline model. This demonstrates that the SupFusion could effectively promote the learning process, guiding the 3D/2D backbone and fusion model extracts meaningful robust fusion features from LiDAR and camera inputs. Furthermore, we also manually set variance depth in deep fusion, to rescale the fusion module size. The results are shown in Tab. 5, in which we can observe that by increasing the fusion depth," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22020" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 175, + 193 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 175, + 193 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 175, + 193 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 175, + 193 + ], + "type": "image", + "image_path": "ddabcb16c12cd858ea94c87677e1fd75f2ee47e590554831cf15cbaecb48c401.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 72, + 298, + 193 + ], + "blocks": [ + { + "bbox": [ + 176, + 72, + 298, + 193 + ], + "lines": [ + { + "bbox": [ + 176, + 72, + 298, + 193 + ], + "spans": [ + { + "bbox": [ + 176, + 72, + 298, + 193 + ], + "type": "image", + "image_path": "0600ecff488cd288295c93125f848ca421168e636c92ced0f71a5107d651563b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 541, + 193 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 541, + 193 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 541, + 193 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 541, + 193 + ], + "type": "image", + "image_path": "16f895916f1513043ab5b173c7bc1b780a0994b19d9d84e417d8fc722585e1a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 194, + 175, + 315 + ], + "blocks": [ + { + "bbox": [ + 52, + 194, + 175, + 315 + ], + "lines": [ + { + "bbox": [ + 52, + 194, + 175, + 315 + ], + "spans": [ + { + "bbox": [ + 52, + 194, + 175, + 315 + ], + "type": "image", + "image_path": "73b5d7c4475322f8b7f7388f041a91dffebf28bed7f1c8948de7bac5df24fa65.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 547, + 365 + ], + "type": "text", + "content": "Figure 5: Visualization results. The top row shows the raw sparse LiDAR data and the bottom row displays the enhanced data. Particularly, the points in magenta are the added points and row points are shown in black. We can densify the sparse objects with few points via the dense objects in database while keeping the density distributions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 176, + 194, + 298, + 315 + ], + "blocks": [ + { + "bbox": [ + 176, + 194, + 298, + 315 + ], + "lines": [ + { + "bbox": [ + 176, + 194, + 298, + 315 + ], + "spans": [ + { + "bbox": [ + 176, + 194, + 298, + 315 + ], + "type": "image", + "image_path": "0330759fde550c5e00492c52f3f103d331a9cf1584f0b84f5104ac781ffc4fc1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 300, + 194, + 541, + 315 + ], + "blocks": [ + { + "bbox": [ + 300, + 194, + 541, + 315 + ], + "lines": [ + { + "bbox": [ + 300, + 194, + 541, + 315 + ], + "spans": [ + { + "bbox": [ + 300, + 194, + 541, + 315 + ], + "type": "image", + "image_path": "2e008fd4c23467c92c78a7d036dfcf52d485497a58984ff8affc858d603433d8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 384, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 288, + 422 + ], + "type": "text", + "content": "the baseline gets worse while our proposal gains better performances, demonstrating the effectiveness of our proposed SupFusion training strategy and deep fusion module again." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "text", + "content": "Polar Sampling. In the proposed polar sampling data enhancement, we split the objects into " + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "text", + "content": " groups according to their directions and rotations. To explore the influence of the split number " + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "text", + "content": ", we manually set " + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "inline_equation", + "content": "N = \\{1, 4, 8, 16\\}" + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "text", + "content": " to conduct experiments, and the results are shown as Tab. 6, from which we can observe that the split number " + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "text", + "content": " in polar sampling affect the performances of assistant and our model in an extent. For instance, when setting " + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "text", + "content": ", we only generate one object for each class, the object is complete but destroys the density distribution, which boosts the assistant model performances but decreases the improvements in an extent (e.g. " + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "inline_equation", + "content": "+1.22\\%" + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "text", + "content": " versus " + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "inline_equation", + "content": "+1.44\\%" + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "text", + "content": "), when setting " + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 423, + 288, + 616 + ], + "type": "text", + "content": " larger, the objects in each object set may get fewer, because we split all the objects in the dataset into more groups, which may influence the quality of the dense object in the database in the polar grouping." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Assistant Model. We utilize the enhanced dense LiDAR data to generate high-quality features. By densifying the sparse objects, the assistant model with enhanced data could avoid detection failers easily and provide high-quality features. To better apply to the given LiDAR-Camera detector and easier implement, we simply adopt the LiDAR branch detector as the assistant model. Here, we also conduct experiments to keep the assistant model sharing the same archi-" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 318, + 419, + 533, + 475 + ], + "blocks": [ + { + "bbox": [ + 304, + 374, + 547, + 411 + ], + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 411 + ], + "type": "text", + "content": "Table 7: Results with different teacher models in SupFusion. LC* and LC** indicate utilize LiDAR and LiDAR-Camera based assistant model." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 419, + 533, + 475 + ], + "lines": [ + { + "bbox": [ + 318, + 419, + 533, + 475 + ], + "spans": [ + { + "bbox": [ + 318, + 419, + 533, + 475 + ], + "type": "table", + "html": "
ModelPed.Cyc.CarO.a.Δ
LC64.8972.9485.5574.46-
LC*66.0676.1885.4775.90+1.44
LC**66.3975.7185.4775.86+1.40
", + "image_path": "9f7ce5df54cd48b55dd60a29557100dd7bd8514b140ee6047667249cb1b5cb07.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 497, + 547, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 558 + ], + "type": "text", + "content": "tecture with the given detector, which is a LiDAR-Camera detector. The results are shown as Tab. 7, from which we can observe that utilizing LiDAR-base and LiDAR-Camera based detectors as the assistant model, the SupFusion gains comparable performances (e.g. LC* versus LC**)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 563, + 418, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 418, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 418, + 576 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Here we visualize the row sparse LiDAR data and enhanced dense LiDAR in Fig. 5. In the left 2 columns, we can observe the added points in magenta could effectively densify the sparse points in black points for all classes including pedestrians, cyclists, and cars. In the right column, we can find that the polar sampling densifies the sparse objects with dense ones while keeping the objects' density distribution. As we could see, enhanced point cloud data have high integrity and obvious class representation which could be utilized to generate high-quality features as the fusion supervision. To better demonstrate the quality of feature-level" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22021" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 95 + ], + "type": "text", + "content": "supervision, we visualize the feature of assistant, baseline, and our model in Appendix." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 103, + 147, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 103, + 147, + 115 + ], + "spans": [ + { + "bbox": [ + 49, + 103, + 147, + 115 + ], + "type": "text", + "content": "4.5. Information leak" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 122, + 286, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 122, + 286, + 181 + ], + "spans": [ + { + "bbox": [ + 49, + 122, + 286, + 181 + ], + "type": "text", + "content": "There is no information leak if we are not misunderstanding. In practice, we build the database in Fig. 4 only using the training set in the polar grouping phase, the validation set is not used in polar sampling, as well as the model training process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 192, + 118, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 192, + 118, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 192, + 118, + 205 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 213, + 286, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 213, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 49, + 213, + 286, + 415 + ], + "type": "text", + "content": "In this paper, we propose a novel training strategy named SupFusion, to introduce auxiliary feature-level supervision for effective LiDAR-Camera fusion, and further boost 3D detection performances. Specifically, we introduce high-quality features to the LiDAR-Camera fusion-based 3D detector, which is used to provide auxiliary supervision for the 3D/2D feature extraction and multi-modal feature fusion. To achieve this, we propose a novel data enhancement method named polar sampling to densify the sparse objects, which is utilized to generate high-quality features via an assistant model. Furthermore, we proposed a simple yet effective deep fusion module, which leverages meaningful information from camera input and contiguously boosts the LiDAR-Camera fusion-based 3D detectors. This is the first approach to introducing feature-level supervision for LiDAR-Camera 3D detection and we conduct extensive experiments to demonstrate its effectiveness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 426, + 129, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 129, + 440 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 129, + 440 + ], + "type": "text", + "content": "6. Acknowledge" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 446, + 286, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 446, + 286, + 541 + ], + "spans": [ + { + "bbox": [ + 49, + 446, + 286, + 541 + ], + "type": "text", + "content": "The work is partially supported by the Young Scientists Fund of the National Natural Science Foundation of China under grant No. 62106154, by the Natural Science Foundation of Guangdong Province, China (General Program) under grant No.2022A1515011524, and by Shenzhen Science and Technology Program JCYJ20220818103001002 and by Shenzhen Science and Technology Program ZDSYS20211021111415025." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "22022" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In CVPR, pages 1090-1099, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 236 + ], + "type": "text", + "content": "[3] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, pages 1907-1915, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 259 + ], + "type": "text", + "content": "[4] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In ICCV, pages 9775-9784, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 304 + ], + "type": "text", + "content": "[5] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. arXiv preprint arXiv:2211.09386, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "text", + "content": "[6] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. arXiv preprint arXiv:2201.10830, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "type": "text", + "content": "[7] Xing Dai, Zeren Jiang, Zhao Wu, Yiping Bao, Zhicheng Wang, Si Liu, and Erjin Zhou. General instance distillation for object detection. In CVPR, pages 7842-7851, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 441 + ], + "type": "text", + "content": "[8] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1201-1209, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 442, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 475 + ], + "type": "text", + "content": "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 476, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 518 + ], + "type": "text", + "content": "[10] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 554 + ], + "type": "text", + "content": "[11] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu, and Chang Xu. Distilling object detectors via decoupled features. In CVPR, pages 2154-2164, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[12] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In ICCV, pages 3153-3163, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 633 + ], + "type": "text", + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[14] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Epnet: Enhancing point features with image semantics for 3d object detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 35-52. Springer, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[15] Bo Ju, Zhikang Zou, Xiaoqing Ye, Minyue Jiang, Xiao Tan, Errui Ding, and Jingdong Wang. Paint and distill: Boosting" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 95 + ], + "type": "text", + "content": "3d object detection with semantic passing network. In ACMM, pages 5639-5648, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "text", + "content": "[16] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-8. IEEE, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 152, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 547, + 185 + ], + "type": "text", + "content": "[17] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. Groomed-nms: Grouped mathematically differentiable nms for monocular 3d object detection. In CVPR, pages 8973-8983, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 186, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 547, + 228 + ], + "type": "text", + "content": "[18] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 230, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 547, + 274 + ], + "type": "text", + "content": "[19] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 276, + 546, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 546, + 308 + ], + "type": "text", + "content": "[20] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In CVPR, pages 1120-1129, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 309, + 546, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 546, + 363 + ], + "type": "text", + "content": "[21] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V Le, et al. Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection. In CVPR, pages 17182-17191, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 365, + 546, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 546, + 419 + ], + "type": "text", + "content": "[22] Zhijian Liu, Haotian Tang, Alexander Amini, Xingyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In IEEE International Conference on Robotics and Automation (ICRA), 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 421, + 546, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 546, + 453 + ], + "type": "text", + "content": "[23] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In ICCV, pages 15641-15650, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "text", + "content": "[24] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In ICCV, pages 3111-3121, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 533 + ], + "type": "text", + "content": "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "text", + "content": "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 579, + 546, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 546, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 546, + 611 + ], + "type": "text", + "content": "[27] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In CVPR, pages 8555-8564, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 613, + 546, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 546, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 546, + 656 + ], + "type": "text", + "content": "[28] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In CVPR, pages 9891-9901, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 658, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 546, + 689 + ], + "type": "text", + "content": "[29] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: Realtime and high-performance pillar-based 3d object detection. In ECCV, pages 35-52. Springer, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 691, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 546, + 714 + ], + "type": "text", + "content": "[30] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22023" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "text", + "content": "voxel feature set abstraction for 3d object detection. In CVPR, pages 10529-10538, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 129 + ], + "type": "text", + "content": "[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointcnn: 3d object proposal generation and detection from point cloud. In CVPR, pages 770-779, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 173 + ], + "type": "text", + "content": "[32] Vishwanath A Sindagi, Yin Zhou, and Oncel Tuzel. Mvx-net: Multimodal voxelnet for 3d object detection. In 2019 International Conference on Robotics and Automation (ICRA), pages 7276-7282. IEEE, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 288, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 230 + ], + "type": "text", + "content": "[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 288, + 285 + ], + "type": "text", + "content": "[34] Kaicheng Yu Zhongyu Xia Zhiwei Lin Yongtao Wang Tao Tang Bing Wang Tingting Liang, Hongwei Xie and Zhi Tang. BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework. In Neural Information Processing Systems (NeurIPS), 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 288, + 319 + ], + "type": "text", + "content": "[35] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In CVPR, pages 11794-11803, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 288, + 365 + ], + "type": "text", + "content": "[36] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3d object detection. In CVPR, pages 454-463, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 366, + 288, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 288, + 409 + ], + "type": "text", + "content": "[37] Li Wang, Li Zhang, Yi Zhu, Zhi Zhang, Tong He, Mu Li, and Xiangyang Xue. Progressive coordinate transforms for monocular 3d object detection. Advances in Neural Information Processing Systems, 34:13364-13377, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 453 + ], + "type": "text", + "content": "[38] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 488 + ], + "type": "text", + "content": "[39] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, pages 4933-4942, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 490, + 288, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 288, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 288, + 522 + ], + "type": "text", + "content": "[40] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In ICCV, pages 913-922, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 524, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 288, + 567 + ], + "type": "text", + "content": "[41] Yue Wang, Alireza Fathi, Abhijit Kundu, David A Ross, Caroline Pantofaru, Tom Funkhouser, and Justin Solomon. Pillar-based object detection for autonomous driving. In ECCV, pages 18-34. Springer, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 569, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 622 + ], + "type": "text", + "content": "[42] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "text", + "content": "[43] Xiaopei Wu, Liang Peng, Honghui Yang, Liang Xie, Chenxi Huang, Chengqi Deng, Haifeng Liu, and Deng Cai. Sparse fuse dense: Towards high quality 3d detection with depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5418-5427, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 455 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "framework for masked image modeling. In CVPR, pages 9653-9663, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 547, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 118 + ], + "type": "text", + "content": "[45] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 119, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 547, + 152 + ], + "type": "text", + "content": "[46] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In ICCV, pages 1951-1960, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 153, + 547, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 547, + 186 + ], + "type": "text", + "content": "[47] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 34:16494-16507, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 547, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 231 + ], + "type": "text", + "content": "[48] Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In ECCV, pages 720-736. Springer, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 232, + 545, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 264 + ], + "type": "text", + "content": "[49] Linfeng Zhang and Kaisheng Ma. Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In ICLR, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 266, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 547, + 308 + ], + "type": "text", + "content": "[50] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In CVPR, pages 18953-18962, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 310, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 547, + 342 + ], + "type": "text", + "content": "[51] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In CVPR, pages 3289-3298, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 344, + 547, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 547, + 377 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 547, + 377 + ], + "type": "text", + "content": "[52] Wu Zheng, Mingxuan Hong, Li Jiang, and Chi-Wing Fu. Boosting 3d object detection by simulating multimodality on point clouds. In CVPR, pages 13638-13647, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 378, + 547, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 378, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 378, + 547, + 422 + ], + "type": "text", + "content": "[53] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3d object detection: An extrinsic parameter free approach. In CVPR, pages 7556-7566, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 423, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 545, + 455 + ], + "type": "text", + "content": "[54] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, pages 4490-4499, 2018." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22024" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Supervised Homography Learning with Realistic Dataset Generation/56098eb5-516e-4294-9a4a-ac98eb40dc0f_content_list.json b/2023/Supervised Homography Learning with Realistic Dataset Generation/56098eb5-516e-4294-9a4a-ac98eb40dc0f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..15492ba1c3048b9da2cc8e3bff48fcf654e626c2 --- /dev/null +++ b/2023/Supervised Homography Learning with Realistic Dataset Generation/56098eb5-516e-4294-9a4a-ac98eb40dc0f_content_list.json @@ -0,0 +1,1698 @@ +[ + { + "type": "text", + "text": "Supervised Homography Learning with Realistic Dataset Generation", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hai Jiang $^{1,2,*}$ , Haipeng Li $^{3,2,*}$ , Songchen Han $^{1,\\dagger}$ , Haoqiang Fan $^{2}$ , Bing Zeng $^{3}$ , Shuaicheng Liu $^{3,2,\\dagger}$ , Sichuan University ${}^{2}$ Megvii Technology", + "bbox": [ + 119, + 186, + 857, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3University of Electronic Science and Technology of China", + "bbox": [ + 254, + 223, + 720, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jianghai@stu.，hansongchen@}scu.edu.cn，{jianghai,lihaipeng,fhq,liushuaicheng}@megvii.com", + "bbox": [ + 86, + 243, + 887, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lihaipeng@std.,eezeng@,liushuaicheng@}uestc.edu.cn", + "bbox": [ + 263, + 262, + 712, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 311, + 313, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose an iterative framework, which consists of two phases: a generation phase and a training phase, to generate realistic training data and yield a supervised homography network. In the generation phase, given an unlabeled image pair, we utilize the pre-estimated dominant plane masks and homography of the pair, along with another sampled homography that serves as ground truth to generate a new labeled training pair with realistic motion. In the training phase, the generated data is used to train the supervised homography network, in which the training data is refined via a content consistency module and a quality assessment module. Once an iteration is finished, the trained network is used in the next data generation phase to update the pre-estimated homography. Through such an iterative strategy, the quality of the dataset and the performance of the network can be gradually and simultaneously improved. Experimental results show that our method achieves state-of-the-art performance and existing supervised methods can be also improved based on the generated dataset. Code and dataset are available at https://github.com/JianghaiSCU/RealSH.", + "bbox": [ + 76, + 343, + 473, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 689, + 209, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Homography estimation is a fundamental task in computer vision that has been widely used for high-dynamic range imaging [40, 24, 25], image stitching [32, 31], and video stabilization [23, 43]. Traditional methods typically use feature extraction and matching methods [26, 3] with outlier suppression [9, 2] to obtain feature matches of two images and subsequently solve direct linear transformation (DLT) [11] to obtain the homography matrix. However, these methods highly rely on the extracted matching keypoints and may crash in challenging scenes that lack", + "bbox": [ + 75, + 714, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cca1b9c37ccec0ef3c38635ef1ab380e38e4e77cf712d4b7404db8f32e7d0de2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 310, + 694, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc41691050c75e5230aa69d2a139f122481c0d909fa84bfd5bc243f252b5c401.jpg", + "image_caption": [ + "label criteria", + "(a) Previous strategy", + "Figure 1. The first row shows the target images generated by the previous strategy [6] and our proposed method, visualized by superimposing the source image warped by GT homography on the generated target image, where the misaligned pixels are represented as colored ghosts. As highlighted in yellow boxes, the dominant plane is fully aligned by the GT homography, proving that both our method and the previous strategy can satisfy the label criteria. In contrast, only our method can maintain realistic motion between the two images thus satisfying the realism criteria, as shown in the red boxes, the car is a moving object that cannot be aligned by a homography." + ], + "image_footnote": [], + "bbox": [ + 504, + 395, + 586, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1fd3a3f3f5d4305438a467b9263d9e80f1101b530b749babe3d672c8e5f297e.jpg", + "image_caption": [ + "realism criteria X" + ], + "image_footnote": [], + "bbox": [ + 589, + 395, + 692, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a8a20c97b43fb0e1072cc6fb8185c4d5ed0331d5af8f495affa85ef14d8e238.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 310, + 888, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7858f74645cd23c5a437eb66751531051e3ab051a43a8066827179cd52a3751.jpg", + "image_caption": [ + "label criteria", + "(b) Ours" + ], + "image_footnote": [], + "bbox": [ + 699, + 395, + 781, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c8429cefc18058be4258d158bab5f04d760ee9f47e7e3b28c9444f95f3b97b6.jpg", + "image_caption": [ + "realism criteria" + ], + "image_footnote": [], + "bbox": [ + 784, + 395, + 888, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sufficient high-quality feature matches. With the rise of deep learning, such problems have been partially solved, learning-based methods [22, 20, 13] take a pair of images as input and directly output the corresponding homography, thus are more robust than traditional methods due to their keypoint-free estimation strategy. The learning-based methods can be divided into two categories: supervised and unsupervised. At the moment, benefiting from the label-free characteristic, unsupervised methods can be trained on enormous amounts of real-world data, delivering better performance and generalization capability than supervised ones. We find that the lack of qualified training data is one of the main barriers to the development of supervised methods. Previously supervised dataset generation strategy [6] synthesizes an image pair by using pre-defined ground truth (GT) homography to warp a single image, such strategy considers the whole image as a single plane, neglecting par", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 875, + 202, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 96, + 887, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9806", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "allax and foreground motion in the real world. As shown in Fig. 1(a), the two images are fully aligned using the GT homography, but the car in the red box should be a moving object that cannot be aligned by a homography.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose an iterative framework, which is designed to generate image pairs that satisfy both label and realism criteria [10] for supervised homography learning and learns a state-of-the-art homography estimation network with the generated dataset. Specifically, given an unlabeled source and target image $(I_s$ and $I_t)$ captured in real-world scenes, we use the homography and dominant plane masks of the two images estimated by the homography estimation network and a pre-trained dominant plane detection network, respectively, along with a sampled GT homography to generate a new target image $I_t'$ . Then, we use the $I_s$ and $I_t'$ , together with the sampled GT homography, to compose a training sample of our supervised dataset. The dominant planes of $I_s$ and $I_t'$ can be fully aligned by the GT homography while the rest remain in realistic motion, making the two images satisfy both label criteria and realism criteria. As shown in Fig. 1(b), the dominant plane of the generated image pair can be fully aligned by the GT homography, while the scene parallax and dynamic objects are maintained in the foreground.", + "bbox": [ + 75, + 152, + 470, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, few artifacts could exist in the generated target image in early iterations, we, therefore, propose a content consistency module and a quality assessment module to eliminate the unexpected content and select high-quality image pairs for training, respectively. The selected image pairs are used to update the homography estimation network which is utilized to estimate homographies for image generation in the next iteration. During the iterative learning steps, the capability of the network is gradually improved, as well as the quality of the synthesized dataset. With our framework, any unlabeled image pairs can be used to generate training samples, thus addressing the lack of qualified datasets in supervised homography learning, which improves the performance of supervised methods and enables them to be better generalized to real-world scenes. In summary, our main contributions are threefold:", + "bbox": [ + 75, + 455, + 468, + 696 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an iterative deep framework to generate a realistic dataset from unlabeled real-world image pairs for supervised homography learning and simultaneously obtain a high-precision network based on the generated dataset.", + "- We propose a content consistency module and a quality assessment module, achieving the elimination of unexpected content and the selection of qualified data for training.", + "- Experimental results show that our method brings noticeable image realism improvement compared to the" + ], + "bbox": [ + 96, + 708, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prior dataset generation strategy and achieves state-of-the-art performance on public benchmarks.", + "bbox": [ + 531, + 90, + 888, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 135, + 640, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Traditional Method", + "text_level": 1, + "bbox": [ + 500, + 160, + 686, + 175 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional homography estimation methods can be divided into feature-based methods and optimization-based methods. Feature-based methods usually combine feature extraction and matching algorithms [26, 3, 37, 7, 34, 36] with outliner suppression approaches [9, 2], followed by solving DLT [11] to obtain homography. However, such methods would crash in challenging scenes where sufficient feature matches cannot be obtained. Optimization-based methods [5, 8] use the Lucas-Kanade algorithm [1] or calculate the sum of squared differences between two images to optimize an initialized homography iteratively. But optimization-based approaches are time-consuming and suffer from cumulative errors.", + "bbox": [ + 496, + 183, + 890, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Learning-based Method", + "text_level": 1, + "bbox": [ + 500, + 390, + 720, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the development of deep image alignment methods, such as optical flow [16, 21, 10, 27, 28] and dense correspondence [39, 38], the first deep homography estimation network was proposed by [6]. Deep learning-based methods can be divided into two categories: supervised and unsupervised. The former ones [15, 35, 4] utilize synthetic image pairs that are derived from single images for training, which suffer from lacking dynamic objects and realistic scene parallax, hindering their generalization ability to real-world scenes. Unsupervised methods [30, 14, 42, 41, 12] are more robust than supervised ones thanks to their label-free training strategies. However, existing unsupervised methods are mostly optimized by minimizing the photometric distance from the warped source to the target, being sensitive to dynamic objects and homogeneous regions [38]. Besides, unsupervised methods are unstable and difficult to converge during the training phase compared to supervised ones. Therefore, our work aims to generate a realistic dataset with homography labels to optimize the network in a supervised manner and enable the network to better generalize to unseen scenes.", + "bbox": [ + 496, + 412, + 890, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Deep Homography Estimation Dataset", + "text_level": 1, + "bbox": [ + 500, + 739, + 831, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The MS-COCO [18] and CA-unsup [42] datasets are commonly used for supervised and unsupervised homography learning. Recently, the GHOF dataset [17] is introduced for unsupervised gyroscope-guided optical flow and homography estimation. The MS-COCO dataset only contains source images, the target images are obtained by warping source images using pre-defined GT homographies. As such, the synthetic image pairs satisfy the label criteria but lack realism criteria. The CA-unsup and GHOF", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9807", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9bdb6ca1525b1d6ba4515bc4cd3f502cede4ef633bbfec9a9c277c127ce5a7e5.jpg", + "image_caption": [ + "Figure 2. The overall pipeline of our proposed framework. In the generation phase, we synthesize a new target image and form a training sample with the source image according to the dominant plane masks and homography of the original image pair estimated by D-Net and H-Net. In the training phase, we propose a content consistency module and a quality assessment module to prepare qualified image pairs for training H-Net that is adopted for image generation in the next iteration. Moreover, we perform the same pre-processing on the training pairs as in the previous method [6, 42, 41, 12], including central cropping, grayscale transformation, and normalization." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 339, + 303 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a9e0502a7291ea9d3caa37524b3a41a8582728f610a56bd505f1f5c11a532dfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 89, + 890, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "datasets collect image pairs from consecutive real-world video frames that satisfy the realism criteria but lack GT labels. In contrast, our method can synthesize realistic image pairs with labels from any unlabeled image pairs, instead of single images, to satisfy both label and realism criteria.", + "bbox": [ + 75, + 401, + 470, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 493, + 166, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 518, + 187, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline of our method is illustrated in Fig. 2. The core innovation of our method is that generating training data and estimating homography are mutually reinforcing. We, therefore, integrate dataset generation and network training into an iterative process, as shown in Fig. 2(a). Our method consists of two phases: the generation phase (G-phase) and the training phase (T-phase).", + "bbox": [ + 75, + 542, + 468, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- G-phase: Given an unlabeled source image $I_{s}$ and target image $I_{t}$ , we generate a new target image $I_{t}^{\\prime}$ according to a pre-defined homography $\\mathbf{H}_{gt}$ as well as the homography and dominant plane masks of the two images estimated by the homography estimation network (H-Net) denoted as $\\Theta_{H}$ and the dominant plane detection network (D-Net) as $\\Theta_{D}$ , forming a training sample as $x = (I_s,I_t',\\mathbf{H}_{gt})$ , i.e.,", + "bbox": [ + 75, + 651, + 468, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx = \\mathrm {G} \\left(I _ {s}, I _ {t}, \\Theta_ {D}, \\Theta_ {H}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 771, + 468, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- T-phase: Using the generated training samples $X = \\{x\\}$ to update the H-Net with the help of a content consistency module (CCM) denoted as $\\Theta_{c}$ and a quality assessment module (QAM) as $\\Theta_{q}$ , i.e,", + "bbox": [ + 76, + 801, + 468, + 863 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta_ {H} ^ {\\prime} = \\underset {\\theta} {\\arg \\min } \\mathcal {L} (X, \\Theta_ {H}, \\Theta_ {c}, \\Theta_ {q}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 877, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Theta_H^{\\prime}$ denotes the H-Net retrained on the training samples and $\\mathcal{L}$ is the learning objective.", + "bbox": [ + 498, + 400, + 888, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Performing G-phase and T-phase interactively can yield a qualified dataset and a high-precision network.", + "bbox": [ + 498, + 431, + 890, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Generation Phase", + "text_level": 1, + "bbox": [ + 500, + 470, + 671, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2(b), we aim to generate a new target image $I_{t}^{\\prime}$ according to the source image $I_{s}$ and target image $I_{t}$ along with a sampled homography $\\mathbf{H}_{gt}$ to serves as the label between $I_{s}$ and $I_{t}^{\\prime}$ , which is modeled by randomly selecting scaling, shearing, rotation, translation, and perspective factors from pre-defined small-baseline ranges. As a result, a training sample that satisfies both label and realism criteria [10] is generated as $(I_{s}, I_{t}^{\\prime}, \\mathbf{H}_{gt})$ .", + "bbox": [ + 496, + 494, + 890, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we first adopt a dominant plane detection network (D-Net), which is formed as a UNet-like structure and pre-trained in an unsupervised manner, to estimate the dominant plane masks $M_{s}$ and $M_{t}$ of the $I_{s}$ and $I_{t}$ , and segment the dominant plane region $P_{d}^{s} = I_{s} \\cdot M_{s}$ and nondominant plane region $P_{n}^{t} = I_{t} \\cdot (1 - M_{t})$ . By warping the $P_{d}^{s}$ with $\\mathbf{H}_{gt}$ to form the counterpart region of $I_{t}^{\\prime}$ , we ensure that their dominant planes are fully aligned. To model the realistic motion between two images, we construct the nondominant plane region of $I_{t}^{\\prime}$ from $I_{t}$ . One simple approach is to directly fuse the warped $P_{d}^{s}$ and $P_{n}^{t}$ as", + "bbox": [ + 496, + 617, + 892, + 784 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} ^ {\\prime} = \\mathcal {W} \\left(P _ {d} ^ {s} + P _ {n} ^ {t}, \\mathbf {H} _ {g t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 795, + 890, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{W}(\\cdot)$ represents the warp operation. However, there exists relative motion between the two images, and the masks may not be accurate and dense enough to segment continuous and complete regions, artlessly using Eq.(3) to form the new target image would produce discontinuities in", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9808", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8ef909d36ee711a2a2f3d3e746a6da9eca8b9fa18ff6ad80708d9e2373c575eb.jpg", + "image_caption": [ + "Source $I_{s}$" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 202, + 143 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cdbaa3701d8f669dc0497fa1040c2a14470ef3dfb763d86d0e84a60ed8baed3a.jpg", + "image_caption": [ + "Target $I_{t}$", + "Figure 3. Illustration of generated images by using different fusion functions. The image fused by using Eq.(3) leads to unsmooth edges and artifacts while using Eq.(4) renders a natural image." + ], + "image_footnote": [], + "bbox": [ + 83, + 154, + 202, + 207 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/efc4208481579acb72cb80d9870d6f2aaa353c9d9dd4b7674cae61a6a187652c.jpg", + "image_caption": [ + "$M_{s}$" + ], + "image_footnote": [], + "bbox": [ + 204, + 90, + 323, + 143 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/14747d079f88a84e02d6cec7618d274c211602127c081a6c717ad44c3dbbd86c.jpg", + "image_caption": [ + "1-Mt" + ], + "image_footnote": [], + "bbox": [ + 204, + 155, + 323, + 207 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a07ae69a55c9f555d46001a5dc03c3cc754a972b868e01c11d0b66f989270111.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 90, + 468, + 154 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d1493168d7b841bdcf5a10e97644dc69afbce5d6c0ab985860f001b0f117322.jpg", + "image_caption": [ + "$l_{t}^{\\prime}$ generated by using Eq.(4)" + ], + "image_footnote": [], + "bbox": [ + 348, + 154, + 467, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the fusion boundary and artifacts. To solve this problem, we construct the non-dominant plane region of $I_{t}^{\\prime}$ by warping the counterpart of $I_{t}$ with the cumulative multiplication result of the homography from target to source $\\mathbf{H}_{ts}$ and $\\mathbf{H}_{gt}$ , and the final $I_{t}^{\\prime}$ can be obtained as", + "bbox": [ + 75, + 284, + 468, + 361 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t} ^ {\\prime} = \\mathcal {W} \\left(P _ {d} ^ {s}, \\mathbf {H} _ {g t}\\right) + \\mathcal {W} \\left(P _ {n} ^ {t}, \\mathbf {H} _ {g t} \\mathbf {H} _ {t s}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 369, + 468, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, despite the inaccuracy of $M_{s}$ and $M_{t}$ , artifacts are avoided and image smoothness is guaranteed using Eq.(4).", + "bbox": [ + 75, + 393, + 468, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training Phase", + "text_level": 1, + "bbox": [ + 76, + 446, + 228, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through the G-phase, we obtain a large-scale realistic dataset for improving the homography estimation network (H-Net). As mentioned above, the quality of generated images depends on the accuracy of the estimated dominant plane masks and homography, few artifacts could exist in generated images in early iterations. To this end, we propose a content consistency module and a quality assessment module to refine the generated image and further select qualified image pairs for training, as shown in Fig. 2(c).", + "bbox": [ + 75, + 470, + 468, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Content Consistency Module. The content consistency module (CCM) is designed to eliminate artifacts in $I_{t}^{\\prime}$ to improve the content quality. As described in Sec. 3.2, the dominant plane of the $I_{s}$ and $I_{t}^{\\prime}$ can be fully aligned by the $\\mathbf{H}_{gt}$ as $P_{d}^{t} = \\mathcal{W}(P_{d}^{s},\\mathbf{H}_{gt})$ , and the $P_d^t$ and $P_d^s$ can be aligned by the $\\mathbf{H}_{ts}$ as $P_{d}^{s} = \\mathcal{W}(P_{d}^{t},\\mathbf{H}_{ts})$ once the $\\mathbf{H}_{ts}$ is accurate. Therefore, the dominant plane of $I_{t}^{\\prime}$ can be converted into the counterpart of $I_{t}$ as", + "bbox": [ + 75, + 607, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {d} ^ {t ^ {\\prime}} = \\mathcal {W} (\\mathcal {W} (P _ {d} ^ {t}, \\mathbf {H} _ {t s}), \\mathbf {H} _ {g t}) = \\mathcal {W} (P _ {d} ^ {t}, \\mathbf {H} _ {g t} \\mathbf {H} _ {t s}). \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 739, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, the non-dominant plane of $I_{t}^{\\prime}$ is obtained by warping the non-dominant plane of $I_{t}$ with the accumulated homography $\\mathbf{H}_{gt}\\mathbf{H}_{ts}$ . Therefore, where $\\mathbf{H}_{ts}$ is extremely accurate, the $I_{t}^{\\prime}$ can be obtained by warping $I_{t}$ with the accumulated homography as $I_{t}^{\\prime} = \\mathcal{W}(I_{t},\\mathbf{H}_{gt}\\mathbf{H}_{ts})$ , which is the assumption behind the CCM we designed. To achieve this, a content consistency loss $\\mathcal{L}_{ccl}$ is proposed as", + "bbox": [ + 75, + 768, + 468, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c c l} = \\left| \\mathcal {W} \\left(\\mathcal {F} \\left(\\hat {I} _ {t} ^ {\\prime}\\right), \\left(\\mathbf {H} _ {g t} \\mathbf {H} _ {t s}\\right) ^ {- 1}\\right) - \\mathcal {F} \\left(I _ {t}\\right) \\right| _ {1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 882, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/55e32fb461799ca7f8379e8c369446bb1dfdb19bc6777089496f8c2638be6141.jpg", + "image_caption": [ + "New Target $I_{t}^{1}$", + "(a) Illustration of the target image reconstructed by CCM." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 678, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ab95e9026e34242d5d3e77da500cc413c707956130091b09f727dce7bac324b4.jpg", + "image_caption": [ + "Reconstructed Target $\\hat{I}_t^{\\prime}$" + ], + "image_footnote": [], + "bbox": [ + 679, + 90, + 888, + 161 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88f7099e994891d0997bce70ecf42add3642d7f7f516c1bf72a279712e517c04.jpg", + "image_caption": [ + "Disturbed $I_{r}$" + ], + "image_footnote": [], + "bbox": [ + 501, + 203, + 630, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e299bc1bacbc87bf713afc6bc8fb5f6471e62c983b80528958efad1cd448046d.jpg", + "image_caption": [ + "Reconstructed $\\hat{I}_t^{\\prime}$", + "Original $I_{t}$" + ], + "image_footnote": [], + "bbox": [ + 632, + 203, + 759, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9eb171ef59a95e0ddd038865f0c3bfdda306bbb72fe049ac0c0decbeba781e05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 203, + 888, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f8ab21ad1a716a67a850a8dd5522923cc66dbc62a3331f0635850746b8412ca3.jpg", + "image_caption": [ + "(b) Illustration of the quality scores estimated by QAM.", + "Figure 4. Illustration of the capabilities of our proposed modules. (a) shows that the artifacts are successfully removed by applying our CCM, yielding high-quality images for training. (b) shows the quality scores of different images estimated by QAM, where high-quality images are used for training and inversely rejected." + ], + "image_footnote": [], + "bbox": [ + 542, + 257, + 857, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{I}_t^\\prime$ is the reconstructed result of $I_t^\\prime$ by CCM, i.e., $\\hat{I}_t^\\prime = \\Theta_c(I_t^\\prime)$ , and $\\mathcal{F}(\\cdot)$ is the feature extractor of H-Net. Instead of directly minimizing the content distance between $\\hat{I}_t^\\prime$ and $I_{t}$ we constrain the similarity of the features to prompt CCM to reconstruct $\\hat{I}_t^\\prime$ to be sharp as the real-world image, since the $\\mathbf{H}_{ts}$ is not accurate enough. As shown in Fig. 4(a), the artifacts in $I_t^\\prime$ have been successfully removed by CCM.", + "bbox": [ + 496, + 419, + 890, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quality Assessment Module. The quality assessment module (QAM) is designed to filter bad cases that cannot be reconstructed by CCM, further improving the quality of the training data. Specifically, the QAM is designed to be a classification network whose output $x \\in \\mathbb{R}$ ranges from 0 to 1, $x$ being closer to 1 means better quality and vice versa.", + "bbox": [ + 496, + 527, + 890, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this, we sample another homography to interfere with $\\mathbf{H}_{ts}$ since the quality of the generated image is partially relied on the accuracy of $\\mathbf{H}_{ts}$ , producing a disturbing image $I_{r}$ with artifacts through G-phase. The score $x_{r}$ obtained by feeding $I_{r}$ into QAM should be close to 0. In contrast, the original target image $I_{t}$ is captured from real-world scenes, so the score $x_{t}$ of $I_{t}$ should be close to 1. Thus, the quality assessment loss $\\mathcal{L}_{qal}$ is proposed to optimize the QAM as", + "bbox": [ + 496, + 619, + 890, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {q a l} = \\operatorname {B C E} \\left(x _ {t}, 1\\right) + \\operatorname {B C E} \\left(x _ {r}, 0\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 766, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{BCE}(\\cdot)$ denotes the binary cross-entropy loss. As a result, when the score $\\hat{x}_t^\\prime$ of $\\hat{I}_t^\\prime$ exceeds a certain threshold $\\tau$ , we consider $\\hat{I}_t^\\prime$ to be a high-quality image that can be used for training and inversely be rejected, as shown in Fig. 4(b).", + "bbox": [ + 496, + 794, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, with the help of CCM and QAM, qualified training pairs can be generated, satisfying both realism criteria and supervision labels to improve the H-Net. Through iter", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9809", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ative learning, the performance of the network and the quality of the generated dataset are mutually improved.", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Network Training", + "text_level": 1, + "bbox": [ + 76, + 130, + 251, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the training samples have GT labels, we can optimize the H-Net by minimizing the difference between $\\mathbf{H}_{gt}$ and the estimated $\\mathbf{H}_{st'}$ . However, as mentioned in [6], it is non-trivial to directly estimate a tomography, we follow previous works [15, 4, 35] to use the 4 corners offset vectors $D_{\\mathbf{H}_{st'}}$ and $D_{\\mathbf{H}_{gt}}$ as supervision objectives, the corresponding homography could be computed by solving DLT [11] using the offset vectors. In addition, we compute the bidirectional homography as the $\\mathbf{H}_{t's}$ should also be aligned with $\\mathbf{H}_{gt}^{-1}$ . The finally supervised loss $\\mathcal{L}_{sup}$ is expressed as", + "bbox": [ + 76, + 154, + 470, + 308 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s u p} = \\left| D _ {\\mathbf {H} _ {s t ^ {\\prime}}} - D _ {\\mathbf {H} _ {g t}} \\right| _ {1} + \\left| D _ {\\mathbf {H} _ {t ^ {\\prime} s}} - D _ {\\mathbf {H} _ {g t} ^ {- 1}} \\right| _ {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 330, + 468, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss $\\mathcal{L}_{\\text {total }}$ is expressed by combing the supervised loss, the content consistency loss, and the quality assessment loss as", + "bbox": [ + 76, + 361, + 468, + 406 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s u p}} + \\lambda_ {1} \\mathcal {L} _ {\\text {c c l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {q a l}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 419, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are empirically set as 0.5 and 0.1.", + "bbox": [ + 76, + 446, + 413, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 474, + 202, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 500, + 171, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MS-COCO. The MS-COCO dataset [18] is one of the widely used datasets in the fields of object detection [19], segmentation [29], etc. Previously supervised methods [6, 15, 35, 4] used it to prepare their training datasets by perturbing each corner of single images to produce GT homographies and warp the images using the homographies to generate training pairs.", + "bbox": [ + 76, + 522, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CA-Unsup. The CA-unsup dataset [42] contains 800k training pairs and 4.2k testing pairs that are captured from five types of real-world scenes, i.e., regular (RE), low texture (LT), low light (LL), small foreground (SF), and large foreground (LF), where the last four are challenging scenes for homography estimation. The training set contains only unlabeled image pairs, but for each pair of the test set, $6\\sim 10$ equally distributed matching points are manually marked for evaluation. Based on the CA-unsup dataset, we generated a new dataset that contains the same number of image pairs satisfying both label criteria and realism criteria as the CA-unusp dataset through our proposed method, named CA-sup. We show some examples of our CA-sup dataset in the supplementary materials.", + "bbox": [ + 75, + 628, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GHOF. The GHOF dataset [17] is designed to evaluate homography and optical flow estimation with the gyroscope readings, it contains 10k training data and 300 testing pairs under 5 different categories, including regular (RE), foggy", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(FOG), low light (LL), rainy (RAIN), and snowy (SNOW) scenes. Compared to the CA-unsup dataset, it contains more parallax motion variations and extreme foreground ratio which are challenging to the homography methods. The sparse $5\\sim 8$ correspondences are used to evaluate the results.", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 176, + 718, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our framework consists of D-Net, CCM, QAM, and H-Net. The D-Net is pre-trained on the CA-unsup dataset in an unsupervised manner, the detailed architecture and implementation details are described in the supplementary material. We choose BasesHomo [41], which balances performance and efficiency as the backbone of H-Net, and convert its output representation into 4 corners offset vectors. The CCM and QAM are training with the H-Net for 2 iterations, and the training parameters are consistent with the official implementation of BasesHomo for each iteration.", + "bbox": [ + 498, + 200, + 890, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 500, + 363, + 807, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Methods. We compare our method with three categories of existing homography estimation methods: 1) feature-based methods including SIFT [26], ORB [33], SuperPoint [7] with SuperGlue [34] (SPSG), and LoFTR [36], 2) supervised methods including DHN [6], LocalTrans [35], and IHN [4], 3) unsupervised methods including CAHomo [42], BasesHomo [41], and HomogAN [12]. For feature-based methods, we improve them with two different outlier rejection algorithms RANSAC [9] and MAGSAC [2], respectively. For supervised methods, we employ their models pre-trained on the MS-COCO dataset [18] and the models retrained on our CA-sup dataset denoted with $*$ for evaluation.", + "bbox": [ + 498, + 386, + 890, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Comparison. We report the quantitative results of all comparison methods on the CA-unsup test set in Table 1. The $\\mathcal{I}_{3\\times 3}$ in the first row refers to a $3\\times 3$ identity matrix as a \"no-warping\" homography for reference, of which the errors reflect the original distance between point pairs. As shown in Table 1, our method achieves state-of-the-art performance in all categories and outperforms the best existing method HomoGAN by $12.82\\%$ , with the points matching error (PME) reduced from 0.39 to 0.34. The feature-based methods can perform well in regular (RE) scenes, as sufficient matching points can be obtained, while our methods and HomoGAN reduce the error by 0.08 compared to SIFT+RANSAC. More specifically, our method produces a PME of 0.216 which is lower than 0.222 of HomoGAN. The small foreground (SF) and large foreground (LF) scenes contain dynamic objects that affect the accuracy of estimated homography, and our method has the lowest PME in these two categories, even when compared to the methods utilizing outlier removal masks for robust estimation, i.e., CAHomo and HomoGAN. In low texture (LT) and low light (LL) scenes, most learning-based", + "bbox": [ + 496, + 584, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9810", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8bd83622134a04ae22ebc7535ddd73df17a90b16ee17b7fabd9e0d1fdcced926.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
1)MethodsAVGRELTLLSFLF
2)\\( \\mathcal{I}_{3\\times 3} \\)6.70 (+1617.95%)7.75 (+3422.73%)7.65 (+1765.85%)7.21 (+1164.91%)7.53 (+1611.36%)3.39 (+993.55%)
3)SIFT [26]+RANSAC [9]1.41 (+261.54%)0.30 (+36.36%)1.34 (+226.83%)4.03 (+607.02%)0.81 (+84.09%)0.57 (+83.87%)
4)SIFT [26]+MAGSAC [2]1.34 (+243.59%)0.31 (+40.91%)1.72 (+319.51%)3.39 (+494.74%)0.80 (+81.82%)0.47 (51.61%)
5)ORB [33]+RANSAC [9]1.48 (+279.49%)0.85 (+286.36%)2.59 (+531.71%)1.67 (+192.98%)1.10 (+150.00%)1.24 (+300.00%)
6)ORB [33]+MAGSAC [2]1.69 (+333.33%)0.97 (+340.91%)3.34 (+714.63%)1.58 (+177.19%)1.15 (+161.36%)1.40 (+351.61%)
7)SPSG [7, 34]+RANSAC [9]0.71 (+82.05%)0.41 (+86.36%)0.87 (+112.20%)0.72 (+26.32%)0.80 (+81.81%)0.75 (+141.94%)
8)SPSG [7, 34]+MAGSAC [2]0.63 (+61.54%)0.36 (+63.64%)0.79 (+92.68%)0.70 (+22.81%)0.71 (+61.36%)0.70 (+125.81%)
9)LoFTR [36]+RANSAC [9]1.44 (+269.23%)0.56 (+154.55%)2.70 (+558.54%)1.36 (+138.60%)1.05 (+138.64%)1.52 (+390.32%)
10)LoFTR [36]+MAGSAC [2]1.39 (+256.41%)0.55 (+150.00%)2.57 (+526.83%)1.33 (+133.33%)1.05 (+138.64%)1.41 (+354.84%)
11)CAHomo [42]0.88 (+125.64%)0.73 (+231.82%)1.01 (+146.34%)1.03 (+80.70%)0.92 (+109.09%)0.70 (+125.81%)
12)BasesHomo [41]0.50 (+28.21%)0.29 (+31.82%)0.54 (+31.71%)0.65 (+14.04%)0.61 (+38.64%)0.41 (+32.26%)
13)HomoGAN [12]0.39 (+0.00%)0.22 (+0.00%)0.41 (+0.00%)0.57 (+0.00%)0.44 (+0.00%)0.31 (+0.00%)
14)DHN [6]2.87 (+635.90%)1.51 (+586.36%)4.48 (+992.68%)2.76 (+384.21%)2.62 (+495.45%)3.00 (+867.74%)
15)LocalTrans [35]4.21 (+978.26%)4.09 (+1757.59%)4.84 (+1081.14%)4.55 (+697.98%)5.30 (+1105.20%)2.25 (+624.29%)
16)IHN [4]4.67 (+1097.44%)4.85 (+2104.55%)5.54 (+1251.22%)5.10 (+794.74%)5.04 (+1045.45%)2.84 (+816.13%)
17)DHN* [6]1.89 (+384.95%)1.21 (+451.82%)2.13 (+418.54%)2.33 (+307.91%)1.72 (+291.23%)2.07 (+567.81%)
18)LocalTrans* [35]1.76 (+352.03%)1.25 (+467.41%)2.15 (+423.85%)1.90 (+232.61%)2.25 (+410.41%)1.28 (+311.77%)
19)IHN* [4]1.19 (+205.54%)0.72 (+227.09%)1.74 (+324.93%)1.18 (+107.42%)1.30 (+196.52%)1.01 (+225.58%)
20)Ours0.34 (-12.82%)0.22 (+0.00%)0.35 (-14.63%)0.44 (-22.81%)0.42 (-4.55%)0.29 (-6.45%)
", + "bbox": [ + 91, + 87, + 883, + 340 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f8e5f6c10b836410120d14082f51614126097547b28ab769abce8a4fccadd704.jpg", + "table_caption": [ + "Table 1. The point matching errors (PME) of our method and all comparison methods on the CA-sup [42] test set. The best and second-best results are highlighted in bold and underlined. The percentages in the parentheses indicate the relative change in comparison to the second-best results. SPSG indicates SuperPoint with SuperGlue, and * denotes the methods are retrained on our CA-sup dataset." + ], + "table_footnote": [], + "table_body": "
1)MethodsAVGREFOGLLRAINSNOW
2)I3×36.334.947.248.095.485.89
3)SIFT [26]+MAGSAC [2]3.900.644.019.770.704.40
4)ORB [33]+RANSAC [9]15.146.9231.2727.801.827.90
5)SPSG [7, 34]+MAGSAC [2]3.283.411.467.610.753.19
6)LoFTR [36]+MAGSAC [2]2.552.761.165.340.522.98
7)CAHomo [42]3.874.103.846.991.273.17
8)BasesHomo [41]2.282.021.434.900.782.29
9)HomoGAN [12]1.951.730.603.950.473.02
10)DHN [6]6.616.046.027.686.996.32
11)LocalTrans [35]5.724.066.495.955.786.34
12)IHN [4]8.177.108.719.346.579.13
13)DHN* [6]3.011.923.944.541.982.66
14)LocalTrans* [35]2.891.784.274.591.372.43
15)IHN* [4]2.592.213.054.700.982.03
16)Ours1.721.600.884.420.431.28
", + "bbox": [ + 84, + 414, + 465, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The point matching errors (PME) of our method and comparison methods on the GHOF [17] test set. The best and second-best results are highlighted in bold and underlined.", + "bbox": [ + 75, + 604, + 468, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods are more robust due to their keypoint-free estimation strategy, especially unsupervised ones. However, homogeneous region [38] dominates a large portion of the image in the LT and LL scenes, where the photometric loss between the two images is minor, so the performance of unsupervised methods in these two categories is not as well as in others. In contrast, our method outperforms the unsupervised methods in the LT and LL with errors reduced by at least $14.63\\%$ and $22.81\\%$ respectively, thanks to our photometric-free fully supervised learning strategy.", + "bbox": [ + 75, + 672, + 468, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, the generalization ability of supervised methods has been widely criticized. Therefore, we further evaluate feature-based methods that performed well on the CA-unsup dataset, learning-based methods, and our method on the GHOF test set to prove the effectiveness of our method.", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 2, our method achieves the best results in RAIN and SNOW scenes and the second-best results in RE, FOG and LL scenes, resulting in the lowest average PME. For regular scenes, even though feature-based methods can extract sufficient high-quality feature matches, we still outperform most of them. In addition, our CA-sup dataset does not contain images captured in the FOG and RAIN scenes, while our method is capable of generalizing these scenes, which proves the superiority of our method. On the other hand, previous supervised methods cannot generalize well to real-world scenes since their training data lack realistic motion. After retraining on our CA-sup dataset, their performance has been significantly improved, which further proves the effectiveness of our framework.", + "bbox": [ + 496, + 417, + 890, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. In Fig. 5 and Fig. 6, we provide the qualitative results of our method and competitive methods on the CA-unsup and GHOF test sets, respectively. Fig. 5(a) and (b) are from LF and SF scenes with dynamic objects, and our method produces more accurate results than feature-based and unsupervised methods which are sensitive to moving objects, as highlighted in the red and green boxes. Fig. 5(c) is from the LT scene, where feature-based methods all fail due to the lack of sufficient correspondence. On the contrary, learning-based methods perform well, but they are still not competitive as ours. In Fig. 6, we provide the comparison results in RAIN and FOG scenes of the GHOF test set, where our training set does not contain images captured in such scenes. As mentioned above, while unsupervised methods have better generalization ability in unseen scenes, we prove that supervised methods can also achieve satisfactory results after training on our generated dataset. For example, in Fig. 6(a), the pre", + "bbox": [ + 496, + 628, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9811", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f673f975c65a8c38f15579105572119d4d41f382c701d8674efc339dc9377038.jpg", + "image_caption": [ + "Figure 5. Qualitative results of our method and other competitive methods on the CA-unsup [42] test set. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, and the green boxes show the content difference between the two images in the error-prone regions. Best viewed by zooming in." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 877, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68fa1c046d9ac77b6e87c1d6cf62c962a876c9b4e420c2bb9588b998bd834e53.jpg", + "image_caption": [ + "Figure 6. Qualitative results of our method and other competitive methods on the GHOF [17] test set. Best viewed by zooming in." + ], + "image_footnote": [], + "bbox": [ + 81, + 507, + 464, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vious supervised methods can even obtain more accurate results than CAHomo, and our method performs the best.", + "bbox": [ + 76, + 806, + 468, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Dataset Generation Methods. We use images from the CA unsup dataset to generate a new dataset named CA-sup $^{\\dagger}$ using the previous dataset generation strategy [6], and retrain H-Net on it for comparison.", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/35d36683c4644e3e47e51e52b95e9572185fc21ed80f9c0f98eb8de40296e547.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetAVGRELTLLSFLF
DHN [6]CA-sup†1.370.791.701.281.231.83
OursCA-sup0.340.220.350.440.420.29
", + "bbox": [ + 511, + 508, + 883, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison with previous dataset generation method. We use the same source images to generate the training dataset and train the same network for comparison.", + "bbox": [ + 498, + 575, + 892, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results on the CA-unsup test set are illustrated in Table 3. The previous strategy generates image pairs from single images that meet label criteria only, causing unsatisfactory performance. In contrast, the image pairs contained in our CA-sup dataset can help the network to learn real-world motion representations, proving the importance of realistic scene motion for estimating homography.", + "bbox": [ + 496, + 643, + 890, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness Evaluation. Furthermore, we evaluate the robustness of all comparison methods on the CA-unsup test set by setting thresholds to calculate the proportion of inlier predictions. As shown in Fig. 7, we plot a series of curves where axis $X \\in [0.1, 3.0]$ is the threshold and axis $Y \\in [0.0, 1.0]$ is the percentage of inliers, higher position of curves represents better robustness. With a threshold of 0.1, our outlier percentage is $6.9\\%$ higher than the second-best method (15.8% vs. $8.9\\%$ ), and our method achieves a proportion of $99.9\\%$ with a threshold of 3.0.", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9812", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ed1caa4e02b272300721236eab804d93b5440ac7fcbed1b2dc33da15001c42c.jpg", + "image_caption": [ + "Figure 7. The proportion of inliers of our method and all comparison methods under various thresholds. The higher position of curves represents better robustness." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 472, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 299, + 230, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct a series of ablation studies to measure the impact of different component choices of our method. We use the CA-sup dataset for training, and quantitative results on the CA-unsup test set are illustrated in Table 4, detailed experiment settings are discussed below. For more ablation experiments, please refer to the supplementary material.", + "bbox": [ + 75, + 323, + 468, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset Generation Strategy. We conduct an experiment to evaluate the effectiveness of our dataset generation strategy by using G-phase only to generate training samples to train the network, denoted as Baseline in row 1 of Table 4. Even without using CCM and QAM, our method achieves state-of-the-art performance compared to the existing methods. Moreover, the homography estimation network, i.e., BasesHomo, trained on our generated dataset achieves a $24\\%$ performance gain compared to the network trained with the unsupervised learning strategy, further proving the superiority of our dataset generation strategy.", + "bbox": [ + 75, + 430, + 468, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Module Effectiveness. We conduct experiments to evaluate the effectiveness of our proposed CCM and QAM by removing them separately from the overall framework. As shown in rows 2-3 of Table 4, CCM and QAM can help to generate and select high-quality image pairs for training, achieving error reduction of 0.03 and 0.01, respectively, compared to the Baseline in row 1 using none of them. By using both CCM and QAM in the training phase, our method can achieve the best performance.", + "bbox": [ + 75, + 597, + 468, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different Homography Estimation Networks. As described in Sec. 4.2, we choose BasesHomo as the backbone of our H-Net. In rows 4-5 of Table 4, we further switch the backbone to CAHomo and HomoGAN, denoted as Real-CA and Real-GAN, respectively, to prove that our method also works not only on a specific architecture. Compared with the unsupervised results in rows 11 and 13 of Table 1, CAHomo obtains performance gain by $50\\%$ with our framework and HomoGAN reduces the error by 0.09, proving the generalization of our proposed framework.", + "bbox": [ + 75, + 733, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Iteration Times. The quality of generated dataset and", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9bdb8d27a5d21ff2045ffb3c0360890076d980abf5bd5127018bb53101fe6190.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AVGRELTLLSFLF
1)Baseline0.380.230.400.500.450.33
2)w/o CCM0.370.230.360.480.440.34
3)w/o QAM0.350.230.360.450.420.30
4)Real-CA [42]0.440.270.450.520.530.43
5)Real-GAN [12]0.300.180.300.390.370.28
6)1 iteration0.380.240.400.460.480.33
7)3 iterations0.330.210.340.440.400.28
8)Default0.340.220.350.440.420.29
", + "bbox": [ + 506, + 88, + 888, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Results of ablation studies, please refer to the text for more details. w/o denotes without.", + "bbox": [ + 498, + 248, + 890, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03ce53b99f9cacb75f67a218b60825cd6c73aa5278daa5ee43a042522901fe10.jpg", + "image_caption": [ + "Figure 8. Illustration of the generated image in different iterations." + ], + "image_footnote": [], + "bbox": [ + 501, + 289, + 890, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "network performance are mutually reinforcing through iterative learning. As shown in rows 6-8 of Table 4, the performance of H-Net improves with iteration increases, while higher-precision H-Net generates more natural images, as shown in Fig. 8, and high-quality images can improve performance. When the homography estimated by H-Net is accurate enough, there would be no artifacts and boundary discontinuities in the generated images. Notably, \"0 iteration\" indicates that the image is generated by using the homography estimated by unsupervised BasesHomo. However, a certain upper limit exists in our method and it will converge after several iterations. The performance gain of 3 iterations compared to 2 iterations, i.e., default, is not significant, and more iterations are not cost-effective compared to the resources spent per iteration.", + "bbox": [ + 496, + 385, + 892, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 630, + 619, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented an iterative deep framework that aims to generate a realistic dataset for supervised homography learning and obtain a high-precision network. We use the dominant plane masks estimated by D-Net and the homography estimated by H-Net to generate training pairs, and the CCM and QAM are proposed to prepare high-quality training data to update the H-Net. The trained network and the generated data can be improved iteratively, yielding a qualified dataset as well as a state-of-the-art network. Extensive experiments demonstrate the superiority of our method and the effectiveness of our newly proposed components, the performance of existing methods can be also improved using our dataset for training.", + "bbox": [ + 496, + 657, + 890, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by Sichuan Science and Technology Program of China under grant No.2023NSFSC0462.", + "bbox": [ + 498, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9813", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Simon Baker and Iain Matthews. Lucas-kanade 20 years on: A unifying framework. International Journal of Computer Vision, 56(3):221-255, 2004. 2", + "[2] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proc. CVPR, pages 10197-10205, 2019. 1, 2, 5, 6", + "[3] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In Proc. ECCV, pages 404-417, 2006. 1, 2", + "[4] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proc. CVPR, pages 1879-1888, 2022. 2, 5, 6", + "[5] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. Clkn: Cascaded lucas-kanade networks for image alignment. In Proc. CVPR, pages 2213–2221, 2017. 2", + "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 1, 2, 3, 5, 6, 7", + "[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPRW, pages 224-236, 2018. 2, 5, 6", + "[8] Tianjiao Ding, Yunchen Yang, Zhihui Zhu, Daniel P Robinson, René Vidal, Laurent Kneip, and Manolis C Tsakiris. Robust homography estimation via dual principal component pursuit. In Proc. CVPR, pages 6080-6089, 2020. 2", + "[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2, 5, 6", + "[10] Yunhui Han, Kunming Luo, Ao Luo, Jiangyu Liu, Haoqiang Fan, Guiming Luo, and Shuaicheng Liu. Realflow: Embased realistic optical flow dataset generation from videos. In Proc. ECCV, pages 288-305, 2022. 2, 3", + "[11] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 2, 5", + "[12] Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, Qijun Zhao, and Shuaicheng Liu. Unsupervised homography estimation with coplanarity-aware gan. In Proc. CVPR, pages 17663-17672, 2022. 2, 3, 5, 6, 8", + "[13] Hai Jiang, Haipeng Li, Yuhang Lu, Songchen Han, and Shuaicheng Liu. Semi-supervised deep large-baseline homography estimation with progressive equivalence constraint. In Proc. AAAI, pages 1024-1032, 2023. 1", + "[14] Dewi Endah Kharismawati, Hadi Ali Akbarpour, Rumana Aktar, Filiz Bunyak, Kannappan Palaniappan, and Toni Kazic. Cornet: Unsupervised deep homography estimation for agricultural aerial imagery. In Proc. ECCV, pages 400-417, 2020. 2", + "[15] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proc. CVPR, pages 7652-7661, 2020. 2, 5", + "[16] Haipeng Li, Kunming Luo, and Shuaicheng Liu. Gyroflow: Gyroscope-guided unsupervised optical flow learning. In Proc. ICCV, pages 12869-12878, 2021. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Haipeng Li, Kunming Luo, Bing Zeng, and Shuaicheng Liu. Gyroflow+: Gyroscope-guided unsupervised deep homography and optical flow learning. arXiv preprint arXiv:2301.10018, 2023. 2, 5, 6, 7", + "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755, 2014. 2, 5", + "[19] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International Journal of Computer Vision, 128(2):261-318, 2020. 5", + "[20] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(06):7885-7899, 2023. 1", + "[21] Shuaicheng Liu, Kunming Luo, Ao Luo, Chuan Wang, Fanman Meng, and Bing Zeng. Asflow: Unsupervised optical flow learning with adaptive pyramid sampling. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4282-4295, 2022. 2", + "[22] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(03):2849-2863, 2023. 1", + "[23] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM Transactions on Graphics, 32(4):1-10, 2013. 1", + "[24] Zhen Liu, Wenjie Lin, Xinpeng Li, Qing Rao, Ting Jiang, Mingyan Han, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. Adnet: Attention-guided deformable convolutional network for high dynamic range imaging. In Proc. CVPRW, pages 463-470, 2021. 1", + "[25] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proc. ECCV, pages 344-360, 2022. 1", + "[26] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 5, 6", + "[27] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In Proc. CVPR, pages 8906-8915, 2022. 2", + "[28] Ao Luo, Fan Yang, Kunming Luo, Xin Li, Haoqiang Fan, and Shuaicheng Liu. Learning optical flow with adaptive graph reasoning. In Proc. AAAI, pages 1890-1898, 2022. 2", + "[29] Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3523-3542, 2022. 5", + "[30] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters, 3(3):2346-2353, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9814", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Lang Nie, Chunyu Lin, Kang Liao, Meiqin Liu, and Yao Zhao. A view-free image stitching network based on global homography. Journal of Visual Communication and Image Representation, 73:102950, 2020. 1", + "[32] Lang Nie, Chunyu Lin, Kang Liao, and Yao Zhao. Learning edge-preserved image stitching from multi-scale deep homography. Neurocomputing, 491:533-543, 2022. 1", + "[33] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In Proc. ICCV, pages 2564-2571, 2011. 5, 6", + "[34] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. CVPR, pages 4938–4947, 2020. 2, 5, 6", + "[35] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. Localtrans: A multiscale local transformer network for cross-resolution homography estimation. In Proc. ICCV, pages 14890-14899, 2021. 2, 5, 6", + "[36] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proc. CVPR, pages 8922-8931, 2021. 2, 5, 6", + "[37] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proc. CVPR, pages 11016-11025, 2019. 2", + "[38] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proc. CVPR, pages 6258-6268, 2020. 2, 6", + "[39] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proc. CVPR, pages 5714-5724, 2021. 2", + "[40] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proc. ECCV, pages 117–132, 2018. 1", + "[41] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proc. ICCV, pages 13117-13125, 2021. 2, 3, 5, 6", + "[42] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proc. ECCV, pages 653-669, 2020. 2, 3, 5, 6, 7, 8", + "[43] Zhuofan Zhang, Zhen Liu, Bing Zeng, and Shuaicheng Liu. Minimum latency deep online video stabilization. arXiv preprint arXiv:2212.02073, 2022. 1" + ], + "bbox": [ + 78, + 90, + 468, + 757 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9815", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Supervised Homography Learning with Realistic Dataset Generation/56098eb5-516e-4294-9a4a-ac98eb40dc0f_model.json b/2023/Supervised Homography Learning with Realistic Dataset Generation/56098eb5-516e-4294-9a4a-ac98eb40dc0f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9272881da8dcfe09fbc9e7d5eae1416b9813c7e5 --- /dev/null +++ b/2023/Supervised Homography Learning with Realistic Dataset Generation/56098eb5-516e-4294-9a4a-ac98eb40dc0f_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Supervised Homography Learning with Realistic Dataset Generation" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.187, + 0.859, + 0.224 + ], + "angle": 0, + "content": "Hai Jiang\\(^{1,2,*}\\), Haipeng Li\\(^{3,2,*}\\), Songchen Han\\(^{1,\\dagger}\\), Haoqiang Fan\\(^{2}\\), Bing Zeng\\(^{3}\\), Shuaicheng Liu\\(^{3,2,\\dagger}\\), Sichuan University \\({}^{2}\\)Megvii Technology" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.224, + 0.722, + 0.242 + ], + "angle": 0, + "content": "3University of Electronic Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.244, + 0.888, + 0.26 + ], + "angle": 0, + "content": "{jianghai@stu.，hansongchen@}scu.edu.cn，{jianghai,lihaipeng,fhq,liushuaicheng}@megvii.com" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.263, + 0.714, + 0.277 + ], + "angle": 0, + "content": "{lihaipeng@std.,eezeng@,liushuaicheng@}uestc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.312, + 0.314, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.344, + 0.474, + 0.661 + ], + "angle": 0, + "content": "In this paper, we propose an iterative framework, which consists of two phases: a generation phase and a training phase, to generate realistic training data and yield a supervised homography network. In the generation phase, given an unlabeled image pair, we utilize the pre-estimated dominant plane masks and homography of the pair, along with another sampled homography that serves as ground truth to generate a new labeled training pair with realistic motion. In the training phase, the generated data is used to train the supervised homography network, in which the training data is refined via a content consistency module and a quality assessment module. Once an iteration is finished, the trained network is used in the next data generation phase to update the pre-estimated homography. Through such an iterative strategy, the quality of the dataset and the performance of the network can be gradually and simultaneously improved. Experimental results show that our method achieves state-of-the-art performance and existing supervised methods can be also improved based on the generated dataset. Code and dataset are available at https://github.com/JianghaiSCU/RealSH." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.69, + 0.21, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Homography estimation is a fundamental task in computer vision that has been widely used for high-dynamic range imaging [40, 24, 25], image stitching [32, 31], and video stabilization [23, 43]. Traditional methods typically use feature extraction and matching methods [26, 3] with outlier suppression [9, 2] to obtain feature matches of two images and subsequently solve direct linear transformation (DLT) [11] to obtain the homography matrix. However, these methods highly rely on the extracted matching keypoints and may crash in challenging scenes that lack" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.311, + 0.695, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.396, + 0.588, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.429, + 0.587, + 0.442 + ], + "angle": 0, + "content": "label criteria" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.396, + 0.693, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.429, + 0.69, + 0.441 + ], + "angle": 0, + "content": "realism criteria X" + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.444, + 0.662, + 0.457 + ], + "angle": 0, + "content": "(a) Previous strategy" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.311, + 0.89, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.396, + 0.782, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.429, + 0.772, + 0.442 + ], + "angle": 0, + "content": "label criteria" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.396, + 0.889, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.429, + 0.885, + 0.442 + ], + "angle": 0, + "content": "realism criteria" + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.444, + 0.822, + 0.456 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.462, + 0.895, + 0.616 + ], + "angle": 0, + "content": "Figure 1. The first row shows the target images generated by the previous strategy [6] and our proposed method, visualized by superimposing the source image warped by GT homography on the generated target image, where the misaligned pixels are represented as colored ghosts. As highlighted in yellow boxes, the dominant plane is fully aligned by the GT homography, proving that both our method and the previous strategy can satisfy the label criteria. In contrast, only our method can maintain realistic motion between the two images thus satisfying the realism criteria, as shown in the red boxes, the car is a moving object that cannot be aligned by a homography." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "sufficient high-quality feature matches. With the rise of deep learning, such problems have been partially solved, learning-based methods [22, 20, 13] take a pair of images as input and directly output the corresponding homography, thus are more robust than traditional methods due to their keypoint-free estimation strategy. The learning-based methods can be divided into two categories: supervised and unsupervised. At the moment, benefiting from the label-free characteristic, unsupervised methods can be trained on enormous amounts of real-world data, delivering better performance and generalization capability than supervised ones. We find that the lack of qualified training data is one of the main barriers to the development of supervised methods. Previously supervised dataset generation strategy [6] synthesizes an image pair by using pre-defined ground truth (GT) homography to warp a single image, such strategy considers the whole image as a single plane, neglecting par" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.204, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.225, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.225, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9806" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "allax and foreground motion in the real world. As shown in Fig. 1(a), the two images are fully aligned using the GT homography, but the car in the red box should be a moving object that cannot be aligned by a homography." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.455 + ], + "angle": 0, + "content": "In this work, we propose an iterative framework, which is designed to generate image pairs that satisfy both label and realism criteria [10] for supervised homography learning and learns a state-of-the-art homography estimation network with the generated dataset. Specifically, given an unlabeled source and target image \\((I_s\\) and \\(I_t)\\) captured in real-world scenes, we use the homography and dominant plane masks of the two images estimated by the homography estimation network and a pre-trained dominant plane detection network, respectively, along with a sampled GT homography to generate a new target image \\(I_t'\\). Then, we use the \\(I_s\\) and \\(I_t'\\), together with the sampled GT homography, to compose a training sample of our supervised dataset. The dominant planes of \\(I_s\\) and \\(I_t'\\) can be fully aligned by the GT homography while the rest remain in realistic motion, making the two images satisfy both label criteria and realism criteria. As shown in Fig. 1(b), the dominant plane of the generated image pair can be fully aligned by the GT homography, while the scene parallax and dynamic objects are maintained in the foreground." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.698 + ], + "angle": 0, + "content": "In addition, few artifacts could exist in the generated target image in early iterations, we, therefore, propose a content consistency module and a quality assessment module to eliminate the unexpected content and select high-quality image pairs for training, respectively. The selected image pairs are used to update the homography estimation network which is utilized to estimate homographies for image generation in the next iteration. During the iterative learning steps, the capability of the network is gradually improved, as well as the quality of the synthesized dataset. With our framework, any unlabeled image pairs can be used to generate training samples, thus addressing the lack of qualified datasets in supervised homography learning, which improves the performance of supervised methods and enables them to be better generalized to real-world scenes. In summary, our main contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.709, + 0.468, + 0.785 + ], + "angle": 0, + "content": "- We propose an iterative deep framework to generate a realistic dataset from unlabeled real-world image pairs for supervised homography learning and simultaneously obtain a high-precision network based on the generated dataset." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.798, + 0.468, + 0.858 + ], + "angle": 0, + "content": "- We propose a content consistency module and a quality assessment module, achieving the elimination of unexpected content and the selection of qualified data for training." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- Experimental results show that our method brings noticeable image realism improvement compared to the" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.709, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "prior dataset generation strategy and achieves state-of-the-art performance on public benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.136, + 0.642, + 0.152 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.161, + 0.687, + 0.176 + ], + "angle": 0, + "content": "2.1. Traditional Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Traditional homography estimation methods can be divided into feature-based methods and optimization-based methods. Feature-based methods usually combine feature extraction and matching algorithms [26, 3, 37, 7, 34, 36] with outliner suppression approaches [9, 2], followed by solving DLT [11] to obtain homography. However, such methods would crash in challenging scenes where sufficient feature matches cannot be obtained. Optimization-based methods [5, 8] use the Lucas-Kanade algorithm [1] or calculate the sum of squared differences between two images to optimize an initialized homography iteratively. But optimization-based approaches are time-consuming and suffer from cumulative errors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.722, + 0.407 + ], + "angle": 0, + "content": "2.2. Learning-based Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Following the development of deep image alignment methods, such as optical flow [16, 21, 10, 27, 28] and dense correspondence [39, 38], the first deep homography estimation network was proposed by [6]. Deep learning-based methods can be divided into two categories: supervised and unsupervised. The former ones [15, 35, 4] utilize synthetic image pairs that are derived from single images for training, which suffer from lacking dynamic objects and realistic scene parallax, hindering their generalization ability to real-world scenes. Unsupervised methods [30, 14, 42, 41, 12] are more robust than supervised ones thanks to their label-free training strategies. However, existing unsupervised methods are mostly optimized by minimizing the photometric distance from the warped source to the target, being sensitive to dynamic objects and homogeneous regions [38]. Besides, unsupervised methods are unstable and difficult to converge during the training phase compared to supervised ones. Therefore, our work aims to generate a realistic dataset with homography labels to optimize the network in a supervised manner and enable the network to better generalize to unseen scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.833, + 0.758 + ], + "angle": 0, + "content": "2.3. Deep Homography Estimation Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The MS-COCO [18] and CA-unsup [42] datasets are commonly used for supervised and unsupervised homography learning. Recently, the GHOF dataset [17] is introduced for unsupervised gyroscope-guided optical flow and homography estimation. The MS-COCO dataset only contains source images, the target images are obtained by warping source images using pre-defined GT homographies. As such, the synthetic image pairs satisfy the label criteria but lack realism criteria. The CA-unsup and GHOF" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.34, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.09, + 0.891, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.894, + 0.378 + ], + "angle": 0, + "content": "Figure 2. The overall pipeline of our proposed framework. In the generation phase, we synthesize a new target image and form a training sample with the source image according to the dominant plane masks and homography of the original image pair estimated by D-Net and H-Net. In the training phase, we propose a content consistency module and a quality assessment module to prepare qualified image pairs for training H-Net that is adopted for image generation in the next iteration. Moreover, we perform the same pre-processing on the training pairs as in the previous method [6, 42, 41, 12], including central cropping, grayscale transformation, and normalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.471, + 0.477 + ], + "angle": 0, + "content": "datasets collect image pairs from consecutive real-world video frames that satisfy the realism criteria but lack GT labels. In contrast, our method can synthesize realistic image pairs with labels from any unlabeled image pairs, instead of single images, to satisfy both label and realism criteria." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.494, + 0.168, + 0.509 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.52, + 0.188, + 0.535 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.469, + 0.65 + ], + "angle": 0, + "content": "The overall pipeline of our method is illustrated in Fig. 2. The core innovation of our method is that generating training data and estimating homography are mutually reinforcing. We, therefore, integrate dataset generation and network training into an iterative process, as shown in Fig. 2(a). Our method consists of two phases: the generation phase (G-phase) and the training phase (T-phase)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.469, + 0.759 + ], + "angle": 0, + "content": "- G-phase: Given an unlabeled source image \\(I_{s}\\) and target image \\(I_{t}\\), we generate a new target image \\(I_{t}^{\\prime}\\) according to a pre-defined homography \\(\\mathbf{H}_{gt}\\) as well as the homography and dominant plane masks of the two images estimated by the homography estimation network (H-Net) denoted as \\(\\Theta_{H}\\) and the dominant plane detection network (D-Net) as \\(\\Theta_{D}\\), forming a training sample as \\(x = (I_s,I_t',\\mathbf{H}_{gt})\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.772, + 0.469, + 0.787 + ], + "angle": 0, + "content": "\\[\nx = \\mathrm {G} \\left(I _ {s}, I _ {t}, \\Theta_ {D}, \\Theta_ {H}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.803, + 0.469, + 0.864 + ], + "angle": 0, + "content": "- T-phase: Using the generated training samples \\( X = \\{x\\} \\) to update the H-Net with the help of a content consistency module (CCM) denoted as \\( \\Theta_{c} \\) and a quality assessment module (QAM) as \\( \\Theta_{q} \\), i.e," + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.878, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\Theta_ {H} ^ {\\prime} = \\underset {\\theta} {\\arg \\min } \\mathcal {L} (X, \\Theta_ {H}, \\Theta_ {c}, \\Theta_ {q}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.401, + 0.89, + 0.432 + ], + "angle": 0, + "content": "where \\(\\Theta_H^{\\prime}\\) denotes the H-Net retrained on the training samples and \\(\\mathcal{L}\\) is the learning objective." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.432, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Performing G-phase and T-phase interactively can yield a qualified dataset and a high-precision network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.673, + 0.485 + ], + "angle": 0, + "content": "3.2. Generation Phase" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.892, + 0.619 + ], + "angle": 0, + "content": "As shown in Fig. 2(b), we aim to generate a new target image \\( I_{t}^{\\prime} \\) according to the source image \\( I_{s} \\) and target image \\( I_{t} \\) along with a sampled homography \\( \\mathbf{H}_{gt} \\) to serves as the label between \\( I_{s} \\) and \\( I_{t}^{\\prime} \\), which is modeled by randomly selecting scaling, shearing, rotation, translation, and perspective factors from pre-defined small-baseline ranges. As a result, a training sample that satisfies both label and realism criteria [10] is generated as \\( (I_{s}, I_{t}^{\\prime}, \\mathbf{H}_{gt}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.893, + 0.785 + ], + "angle": 0, + "content": "Specifically, we first adopt a dominant plane detection network (D-Net), which is formed as a UNet-like structure and pre-trained in an unsupervised manner, to estimate the dominant plane masks \\(M_{s}\\) and \\(M_{t}\\) of the \\(I_{s}\\) and \\(I_{t}\\), and segment the dominant plane region \\(P_{d}^{s} = I_{s} \\cdot M_{s}\\) and nondominant plane region \\(P_{n}^{t} = I_{t} \\cdot (1 - M_{t})\\). By warping the \\(P_{d}^{s}\\) with \\(\\mathbf{H}_{gt}\\) to form the counterpart region of \\(I_{t}^{\\prime}\\), we ensure that their dominant planes are fully aligned. To model the realistic motion between two images, we construct the nondominant plane region of \\(I_{t}^{\\prime}\\) from \\(I_{t}\\). One simple approach is to directly fuse the warped \\(P_{d}^{s}\\) and \\(P_{n}^{t}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.796, + 0.891, + 0.815 + ], + "angle": 0, + "content": "\\[\nI _ {t} ^ {\\prime} = \\mathcal {W} \\left(P _ {d} ^ {s} + P _ {n} ^ {t}, \\mathbf {H} _ {g t}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}(\\cdot)\\) represents the warp operation. However, there exists relative motion between the two images, and the masks may not be accurate and dense enough to segment continuous and complete regions, artlessly using Eq.(3) to form the new target image would produce discontinuities in" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.204, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.146, + 0.166, + 0.154 + ], + "angle": 0, + "content": "Source \\(I_{s}\\)" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.155, + 0.204, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.207, + 0.165, + 0.218 + ], + "angle": 0, + "content": "Target \\(I_{t}\\)" + }, + { + "type": "image", + "bbox": [ + 0.205, + 0.091, + 0.324, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.146, + 0.271, + 0.155 + ], + "angle": 0, + "content": "\\(M_{s}\\)" + }, + { + "type": "image", + "bbox": [ + 0.205, + 0.156, + 0.324, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.207, + 0.28, + 0.219 + ], + "angle": 0, + "content": "1-Mt" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.091, + 0.469, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.155, + 0.468, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.207, + 0.463, + 0.218 + ], + "angle": 0, + "content": "\\(l_{t}^{\\prime}\\) generated by using Eq.(4)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.47, + 0.265 + ], + "angle": 0, + "content": "Figure 3. Illustration of generated images by using different fusion functions. The image fused by using Eq.(3) leads to unsmooth edges and artifacts while using Eq.(4) renders a natural image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.47, + 0.362 + ], + "angle": 0, + "content": "the fusion boundary and artifacts. To solve this problem, we construct the non-dominant plane region of \\(I_{t}^{\\prime}\\) by warping the counterpart of \\(I_{t}\\) with the cumulative multiplication result of the homography from target to source \\(\\mathbf{H}_{ts}\\) and \\(\\mathbf{H}_{gt}\\), and the final \\(I_{t}^{\\prime}\\) can be obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.37, + 0.469, + 0.389 + ], + "angle": 0, + "content": "\\[\nI _ {t} ^ {\\prime} = \\mathcal {W} \\left(P _ {d} ^ {s}, \\mathbf {H} _ {g t}\\right) + \\mathcal {W} \\left(P _ {n} ^ {t}, \\mathbf {H} _ {g t} \\mathbf {H} _ {t s}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.469, + 0.441 + ], + "angle": 0, + "content": "As shown in Fig. 3, despite the inaccuracy of \\( M_{s} \\) and \\( M_{t} \\), artifacts are avoided and image smoothness is guaranteed using Eq.(4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.448, + 0.23, + 0.464 + ], + "angle": 0, + "content": "3.3. Training Phase" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.469, + 0.607 + ], + "angle": 0, + "content": "Through the G-phase, we obtain a large-scale realistic dataset for improving the homography estimation network (H-Net). As mentioned above, the quality of generated images depends on the accuracy of the estimated dominant plane masks and homography, few artifacts could exist in generated images in early iterations. To this end, we propose a content consistency module and a quality assessment module to refine the generated image and further select qualified image pairs for training, as shown in Fig. 2(c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.469, + 0.734 + ], + "angle": 0, + "content": "Content Consistency Module. The content consistency module (CCM) is designed to eliminate artifacts in \\(I_{t}^{\\prime}\\) to improve the content quality. As described in Sec. 3.2, the dominant plane of the \\(I_{s}\\) and \\(I_{t}^{\\prime}\\) can be fully aligned by the \\(\\mathbf{H}_{gt}\\) as \\(P_{d}^{t} = \\mathcal{W}(P_{d}^{s},\\mathbf{H}_{gt})\\) , and the \\(P_d^t\\) and \\(P_d^s\\) can be aligned by the \\(\\mathbf{H}_{ts}\\) as \\(P_{d}^{s} = \\mathcal{W}(P_{d}^{t},\\mathbf{H}_{ts})\\) once the \\(\\mathbf{H}_{ts}\\) is accurate. Therefore, the dominant plane of \\(I_{t}^{\\prime}\\) can be converted into the counterpart of \\(I_{t}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.741, + 0.469, + 0.763 + ], + "angle": 0, + "content": "\\[\nP _ {d} ^ {t ^ {\\prime}} = \\mathcal {W} (\\mathcal {W} (P _ {d} ^ {t}, \\mathbf {H} _ {t s}), \\mathbf {H} _ {g t}) = \\mathcal {W} (P _ {d} ^ {t}, \\mathbf {H} _ {g t} \\mathbf {H} _ {t s}). \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Besides, the non-dominant plane of \\(I_{t}^{\\prime}\\) is obtained by warping the non-dominant plane of \\(I_{t}\\) with the accumulated homography \\(\\mathbf{H}_{gt}\\mathbf{H}_{ts}\\). Therefore, where \\(\\mathbf{H}_{ts}\\) is extremely accurate, the \\(I_{t}^{\\prime}\\) can be obtained by warping \\(I_{t}\\) with the accumulated homography as \\(I_{t}^{\\prime} = \\mathcal{W}(I_{t},\\mathbf{H}_{gt}\\mathbf{H}_{ts})\\), which is the assumption behind the CCM we designed. To achieve this, a content consistency loss \\(\\mathcal{L}_{ccl}\\) is proposed as" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.883, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c c l} = \\left| \\mathcal {W} \\left(\\mathcal {F} \\left(\\hat {I} _ {t} ^ {\\prime}\\right), \\left(\\mathbf {H} _ {g t} \\mathbf {H} _ {t s}\\right) ^ {- 1}\\right) - \\mathcal {F} \\left(I _ {t}\\right) \\right| _ {1}, \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.679, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.163, + 0.617, + 0.175 + ], + "angle": 0, + "content": "New Target \\(I_{t}^{1}\\)" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.091, + 0.89, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.163, + 0.865, + 0.175 + ], + "angle": 0, + "content": "Reconstructed Target \\(\\hat{I}_t^{\\prime}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.177, + 0.83, + 0.188 + ], + "angle": 0, + "content": "(a) Illustration of the target image reconstructed by CCM." + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.192, + 0.597, + 0.203 + ], + "angle": 0, + "content": "Disturbed \\(I_{r}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.193, + 0.735, + 0.204 + ], + "angle": 0, + "content": "Reconstructed \\(\\hat{I}_t^{\\prime}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.192, + 0.89, + 0.204 + ], + "angle": 0, + "content": "Original \\(I_{t}\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.204, + 0.632, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.204, + 0.761, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.204, + 0.89, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.258, + 0.859, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.311, + 0.824, + 0.323 + ], + "angle": 0, + "content": "(b) Illustration of the quality scores estimated by QAM." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Figure 4. Illustration of the capabilities of our proposed modules. (a) shows that the artifacts are successfully removed by applying our CCM, yielding high-quality images for training. (b) shows the quality scores of different images estimated by QAM, where high-quality images are used for training and inversely rejected." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.527 + ], + "angle": 0, + "content": "where \\(\\hat{I}_t^\\prime\\) is the reconstructed result of \\(I_t^\\prime\\) by CCM, i.e., \\(\\hat{I}_t^\\prime = \\Theta_c(I_t^\\prime)\\), and \\(\\mathcal{F}(\\cdot)\\) is the feature extractor of H-Net. Instead of directly minimizing the content distance between \\(\\hat{I}_t^\\prime\\) and \\(I_{t}\\) we constrain the similarity of the features to prompt CCM to reconstruct \\(\\hat{I}_t^\\prime\\) to be sharp as the real-world image, since the \\(\\mathbf{H}_{ts}\\) is not accurate enough. As shown in Fig. 4(a), the artifacts in \\(I_t^\\prime\\) have been successfully removed by CCM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.619 + ], + "angle": 0, + "content": "Quality Assessment Module. The quality assessment module (QAM) is designed to filter bad cases that cannot be reconstructed by CCM, further improving the quality of the training data. Specifically, the QAM is designed to be a classification network whose output \\( x \\in \\mathbb{R} \\) ranges from 0 to 1, \\( x \\) being closer to 1 means better quality and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.755 + ], + "angle": 0, + "content": "To achieve this, we sample another homography to interfere with \\(\\mathbf{H}_{ts}\\) since the quality of the generated image is partially relied on the accuracy of \\(\\mathbf{H}_{ts}\\), producing a disturbing image \\(I_{r}\\) with artifacts through G-phase. The score \\(x_{r}\\) obtained by feeding \\(I_{r}\\) into QAM should be close to 0. In contrast, the original target image \\(I_{t}\\) is captured from real-world scenes, so the score \\(x_{t}\\) of \\(I_{t}\\) should be close to 1. Thus, the quality assessment loss \\(\\mathcal{L}_{qal}\\) is proposed to optimize the QAM as" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.767, + 0.891, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {q a l} = \\operatorname {B C E} \\left(x _ {t}, 1\\right) + \\operatorname {B C E} \\left(x _ {r}, 0\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.855 + ], + "angle": 0, + "content": "where \\(\\mathrm{BCE}(\\cdot)\\) denotes the binary cross-entropy loss. As a result, when the score \\(\\hat{x}_t^\\prime\\) of \\(\\hat{I}_t^\\prime\\) exceeds a certain threshold \\(\\tau\\), we consider \\(\\hat{I}_t^\\prime\\) to be a high-quality image that can be used for training and inversely be rejected, as shown in Fig. 4(b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Overall, with the help of CCM and QAM, qualified training pairs can be generated, satisfying both realism criteria and supervision labels to improve the H-Net. Through iter" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9809" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "ative learning, the performance of the network and the quality of the generated dataset are mutually improved." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.131, + 0.253, + 0.148 + ], + "angle": 0, + "content": "3.4. Network Training" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.155, + 0.471, + 0.309 + ], + "angle": 0, + "content": "Since the training samples have GT labels, we can optimize the H-Net by minimizing the difference between \\(\\mathbf{H}_{gt}\\) and the estimated \\(\\mathbf{H}_{st'}\\). However, as mentioned in [6], it is non-trivial to directly estimate a tomography, we follow previous works [15, 4, 35] to use the 4 corners offset vectors \\(D_{\\mathbf{H}_{st'}}\\) and \\(D_{\\mathbf{H}_{gt}}\\) as supervision objectives, the corresponding homography could be computed by solving DLT [11] using the offset vectors. In addition, we compute the bidirectional homography as the \\(\\mathbf{H}_{t's}\\) should also be aligned with \\(\\mathbf{H}_{gt}^{-1}\\). The finally supervised loss \\(\\mathcal{L}_{sup}\\) is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.332, + 0.47, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s u p} = \\left| D _ {\\mathbf {H} _ {s t ^ {\\prime}}} - D _ {\\mathbf {H} _ {g t}} \\right| _ {1} + \\left| D _ {\\mathbf {H} _ {t ^ {\\prime} s}} - D _ {\\mathbf {H} _ {g t} ^ {- 1}} \\right| _ {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.362, + 0.47, + 0.407 + ], + "angle": 0, + "content": "The total loss \\(\\mathcal{L}_{\\text {total }}\\) is expressed by combing the supervised loss, the content consistency loss, and the quality assessment loss as" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.42, + 0.469, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s u p}} + \\lambda_ {1} \\mathcal {L} _ {\\text {c c l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {q a l}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.447, + 0.415, + 0.464 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are empirically set as 0.5 and 0.1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.476, + 0.203, + 0.493 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.5, + 0.173, + 0.515 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.523, + 0.469, + 0.629 + ], + "angle": 0, + "content": "MS-COCO. The MS-COCO dataset [18] is one of the widely used datasets in the fields of object detection [19], segmentation [29], etc. Previously supervised methods [6, 15, 35, 4] used it to prepare their training datasets by perturbing each corner of single images to produce GT homographies and warp the images using the homographies to generate training pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.84 + ], + "angle": 0, + "content": "CA-Unsup. The CA-unsup dataset [42] contains 800k training pairs and 4.2k testing pairs that are captured from five types of real-world scenes, i.e., regular (RE), low texture (LT), low light (LL), small foreground (SF), and large foreground (LF), where the last four are challenging scenes for homography estimation. The training set contains only unlabeled image pairs, but for each pair of the test set, \\(6\\sim 10\\) equally distributed matching points are manually marked for evaluation. Based on the CA-unsup dataset, we generated a new dataset that contains the same number of image pairs satisfying both label criteria and realism criteria as the CA-unusp dataset through our proposed method, named CA-sup. We show some examples of our CA-sup dataset in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "GHOF. The GHOF dataset [17] is designed to evaluate homography and optical flow estimation with the gyroscope readings, it contains 10k training data and 300 testing pairs under 5 different categories, including regular (RE), foggy" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "(FOG), low light (LL), rainy (RAIN), and snowy (SNOW) scenes. Compared to the CA-unsup dataset, it contains more parallax motion variations and extreme foreground ratio which are challenging to the homography methods. The sparse \\(5\\sim 8\\) correspondences are used to evaluate the results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.178, + 0.719, + 0.194 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.201, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Our framework consists of D-Net, CCM, QAM, and H-Net. The D-Net is pre-trained on the CA-unsup dataset in an unsupervised manner, the detailed architecture and implementation details are described in the supplementary material. We choose BasesHomo [41], which balances performance and efficiency as the backbone of H-Net, and convert its output representation into 4 corners offset vectors. The CCM and QAM are training with the H-Net for 2 iterations, and the training parameters are consistent with the official implementation of BasesHomo for each iteration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.364, + 0.808, + 0.38 + ], + "angle": 0, + "content": "4.3. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.387, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Comparison Methods. We compare our method with three categories of existing homography estimation methods: 1) feature-based methods including SIFT [26], ORB [33], SuperPoint [7] with SuperGlue [34] (SPSG), and LoFTR [36], 2) supervised methods including DHN [6], LocalTrans [35], and IHN [4], 3) unsupervised methods including CAHomo [42], BasesHomo [41], and HomogAN [12]. For feature-based methods, we improve them with two different outlier rejection algorithms RANSAC [9] and MAGSAC [2], respectively. For supervised methods, we employ their models pre-trained on the MS-COCO dataset [18] and the models retrained on our CA-sup dataset denoted with \\(*\\) for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparison. We report the quantitative results of all comparison methods on the CA-unsup test set in Table 1. The \\(\\mathcal{I}_{3\\times 3}\\) in the first row refers to a \\(3\\times 3\\) identity matrix as a \"no-warping\" homography for reference, of which the errors reflect the original distance between point pairs. As shown in Table 1, our method achieves state-of-the-art performance in all categories and outperforms the best existing method HomoGAN by \\(12.82\\%\\), with the points matching error (PME) reduced from 0.39 to 0.34. The feature-based methods can perform well in regular (RE) scenes, as sufficient matching points can be obtained, while our methods and HomoGAN reduce the error by 0.08 compared to SIFT+RANSAC. More specifically, our method produces a PME of 0.216 which is lower than 0.222 of HomoGAN. The small foreground (SF) and large foreground (LF) scenes contain dynamic objects that affect the accuracy of estimated homography, and our method has the lowest PME in these two categories, even when compared to the methods utilizing outlier removal masks for robust estimation, i.e., CAHomo and HomoGAN. In low texture (LT) and low light (LL) scenes, most learning-based" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9810" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.088, + 0.885, + 0.342 + ], + "angle": 0, + "content": "
1)MethodsAVGRELTLLSFLF
2)\\( \\mathcal{I}_{3\\times 3} \\)6.70 (+1617.95%)7.75 (+3422.73%)7.65 (+1765.85%)7.21 (+1164.91%)7.53 (+1611.36%)3.39 (+993.55%)
3)SIFT [26]+RANSAC [9]1.41 (+261.54%)0.30 (+36.36%)1.34 (+226.83%)4.03 (+607.02%)0.81 (+84.09%)0.57 (+83.87%)
4)SIFT [26]+MAGSAC [2]1.34 (+243.59%)0.31 (+40.91%)1.72 (+319.51%)3.39 (+494.74%)0.80 (+81.82%)0.47 (51.61%)
5)ORB [33]+RANSAC [9]1.48 (+279.49%)0.85 (+286.36%)2.59 (+531.71%)1.67 (+192.98%)1.10 (+150.00%)1.24 (+300.00%)
6)ORB [33]+MAGSAC [2]1.69 (+333.33%)0.97 (+340.91%)3.34 (+714.63%)1.58 (+177.19%)1.15 (+161.36%)1.40 (+351.61%)
7)SPSG [7, 34]+RANSAC [9]0.71 (+82.05%)0.41 (+86.36%)0.87 (+112.20%)0.72 (+26.32%)0.80 (+81.81%)0.75 (+141.94%)
8)SPSG [7, 34]+MAGSAC [2]0.63 (+61.54%)0.36 (+63.64%)0.79 (+92.68%)0.70 (+22.81%)0.71 (+61.36%)0.70 (+125.81%)
9)LoFTR [36]+RANSAC [9]1.44 (+269.23%)0.56 (+154.55%)2.70 (+558.54%)1.36 (+138.60%)1.05 (+138.64%)1.52 (+390.32%)
10)LoFTR [36]+MAGSAC [2]1.39 (+256.41%)0.55 (+150.00%)2.57 (+526.83%)1.33 (+133.33%)1.05 (+138.64%)1.41 (+354.84%)
11)CAHomo [42]0.88 (+125.64%)0.73 (+231.82%)1.01 (+146.34%)1.03 (+80.70%)0.92 (+109.09%)0.70 (+125.81%)
12)BasesHomo [41]0.50 (+28.21%)0.29 (+31.82%)0.54 (+31.71%)0.65 (+14.04%)0.61 (+38.64%)0.41 (+32.26%)
13)HomoGAN [12]0.39 (+0.00%)0.22 (+0.00%)0.41 (+0.00%)0.57 (+0.00%)0.44 (+0.00%)0.31 (+0.00%)
14)DHN [6]2.87 (+635.90%)1.51 (+586.36%)4.48 (+992.68%)2.76 (+384.21%)2.62 (+495.45%)3.00 (+867.74%)
15)LocalTrans [35]4.21 (+978.26%)4.09 (+1757.59%)4.84 (+1081.14%)4.55 (+697.98%)5.30 (+1105.20%)2.25 (+624.29%)
16)IHN [4]4.67 (+1097.44%)4.85 (+2104.55%)5.54 (+1251.22%)5.10 (+794.74%)5.04 (+1045.45%)2.84 (+816.13%)
17)DHN* [6]1.89 (+384.95%)1.21 (+451.82%)2.13 (+418.54%)2.33 (+307.91%)1.72 (+291.23%)2.07 (+567.81%)
18)LocalTrans* [35]1.76 (+352.03%)1.25 (+467.41%)2.15 (+423.85%)1.90 (+232.61%)2.25 (+410.41%)1.28 (+311.77%)
19)IHN* [4]1.19 (+205.54%)0.72 (+227.09%)1.74 (+324.93%)1.18 (+107.42%)1.30 (+196.52%)1.01 (+225.58%)
20)Ours0.34 (-12.82%)0.22 (+0.00%)0.35 (-14.63%)0.44 (-22.81%)0.42 (-4.55%)0.29 (-6.45%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.351, + 0.893, + 0.395 + ], + "angle": 0, + "content": "Table 1. The point matching errors (PME) of our method and all comparison methods on the CA-sup [42] test set. The best and second-best results are highlighted in bold and underlined. The percentages in the parentheses indicate the relative change in comparison to the second-best results. SPSG indicates SuperPoint with SuperGlue, and * denotes the methods are retrained on our CA-sup dataset." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.415, + 0.466, + 0.596 + ], + "angle": 0, + "content": "
1)MethodsAVGREFOGLLRAINSNOW
2)I3×36.334.947.248.095.485.89
3)SIFT [26]+MAGSAC [2]3.900.644.019.770.704.40
4)ORB [33]+RANSAC [9]15.146.9231.2727.801.827.90
5)SPSG [7, 34]+MAGSAC [2]3.283.411.467.610.753.19
6)LoFTR [36]+MAGSAC [2]2.552.761.165.340.522.98
7)CAHomo [42]3.874.103.846.991.273.17
8)BasesHomo [41]2.282.021.434.900.782.29
9)HomoGAN [12]1.951.730.603.950.473.02
10)DHN [6]6.616.046.027.686.996.32
11)LocalTrans [35]5.724.066.495.955.786.34
12)IHN [4]8.177.108.719.346.579.13
13)DHN* [6]3.011.923.944.541.982.66
14)LocalTrans* [35]2.891.784.274.591.372.43
15)IHN* [4]2.592.213.054.700.982.03
16)Ours1.721.600.884.420.431.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.649 + ], + "angle": 0, + "content": "Table 2. The point matching errors (PME) of our method and comparison methods on the GHOF [17] test set. The best and second-best results are highlighted in bold and underlined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.825 + ], + "angle": 0, + "content": "methods are more robust due to their keypoint-free estimation strategy, especially unsupervised ones. However, homogeneous region [38] dominates a large portion of the image in the LT and LL scenes, where the photometric loss between the two images is minor, so the performance of unsupervised methods in these two categories is not as well as in others. In contrast, our method outperforms the unsupervised methods in the LT and LL with errors reduced by at least \\(14.63\\%\\) and \\(22.81\\%\\) respectively, thanks to our photometric-free fully supervised learning strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Moreover, the generalization ability of supervised methods has been widely criticized. Therefore, we further evaluate feature-based methods that performed well on the CA-unsup dataset, learning-based methods, and our method on the GHOF test set to prove the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.628 + ], + "angle": 0, + "content": "As shown in Table 2, our method achieves the best results in RAIN and SNOW scenes and the second-best results in RE, FOG and LL scenes, resulting in the lowest average PME. For regular scenes, even though feature-based methods can extract sufficient high-quality feature matches, we still outperform most of them. In addition, our CA-sup dataset does not contain images captured in the FOG and RAIN scenes, while our method is capable of generalizing these scenes, which proves the superiority of our method. On the other hand, previous supervised methods cannot generalize well to real-world scenes since their training data lack realistic motion. After retraining on our CA-sup dataset, their performance has been significantly improved, which further proves the effectiveness of our framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Qualitative Comparison. In Fig. 5 and Fig. 6, we provide the qualitative results of our method and competitive methods on the CA-unsup and GHOF test sets, respectively. Fig. 5(a) and (b) are from LF and SF scenes with dynamic objects, and our method produces more accurate results than feature-based and unsupervised methods which are sensitive to moving objects, as highlighted in the red and green boxes. Fig. 5(c) is from the LT scene, where feature-based methods all fail due to the lack of sufficient correspondence. On the contrary, learning-based methods perform well, but they are still not competitive as ours. In Fig. 6, we provide the comparison results in RAIN and FOG scenes of the GHOF test set, where our training set does not contain images captured in such scenes. As mentioned above, while unsupervised methods have better generalization ability in unseen scenes, we prove that supervised methods can also achieve satisfactory results after training on our generated dataset. For example, in Fig. 6(a), the pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9811" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.878, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.445, + 0.895, + 0.488 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of our method and other competitive methods on the CA-unsup [42] test set. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, and the green boxes show the content difference between the two images in the error-prone regions. Best viewed by zooming in." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.508, + 0.465, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.747, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of our method and other competitive methods on the GHOF [17] test set. Best viewed by zooming in." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.807, + 0.47, + 0.838 + ], + "angle": 0, + "content": "vious supervised methods can even obtain more accurate results than CAHomo, and our method performs the best." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Comparison with Dataset Generation Methods. We use images from the CA unsup dataset to generate a new dataset named CA-sup\\(^{\\dagger}\\) using the previous dataset generation strategy [6], and retrain H-Net on it for comparison." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.509, + 0.885, + 0.567 + ], + "angle": 0, + "content": "
DatasetAVGRELTLLSFLF
DHN [6]CA-sup†1.370.791.701.281.231.83
OursCA-sup0.340.220.350.440.420.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.576, + 0.893, + 0.619 + ], + "angle": 0, + "content": "Table 3. Comparison with previous dataset generation method. We use the same source images to generate the training dataset and train the same network for comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Quantitative results on the CA-unsup test set are illustrated in Table 3. The previous strategy generates image pairs from single images that meet label criteria only, causing unsatisfactory performance. In contrast, the image pairs contained in our CA-sup dataset can help the network to learn real-world motion representations, proving the importance of realistic scene motion for estimating homography." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Robustness Evaluation. Furthermore, we evaluate the robustness of all comparison methods on the CA-unsup test set by setting thresholds to calculate the proportion of inlier predictions. As shown in Fig. 7, we plot a series of curves where axis \\( X \\in [0.1, 3.0] \\) is the threshold and axis \\( Y \\in [0.0, 1.0] \\) is the percentage of inliers, higher position of curves represents better robustness. With a threshold of 0.1, our outlier percentage is \\( 6.9\\% \\) higher than the second-best method (15.8% vs. \\( 8.9\\% \\)), and our method achieves a proportion of \\( 99.9\\% \\) with a threshold of 3.0." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.089, + 0.473, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.233, + 0.472, + 0.276 + ], + "angle": 0, + "content": "Figure 7. The proportion of inliers of our method and all comparison methods under various thresholds. The higher position of curves represents better robustness." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.3, + 0.231, + 0.317 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.43 + ], + "angle": 0, + "content": "In this section, we conduct a series of ablation studies to measure the impact of different component choices of our method. We use the CA-sup dataset for training, and quantitative results on the CA-unsup test set are illustrated in Table 4, detailed experiment settings are discussed below. For more ablation experiments, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.597 + ], + "angle": 0, + "content": "Dataset Generation Strategy. We conduct an experiment to evaluate the effectiveness of our dataset generation strategy by using G-phase only to generate training samples to train the network, denoted as Baseline in row 1 of Table 4. Even without using CCM and QAM, our method achieves state-of-the-art performance compared to the existing methods. Moreover, the homography estimation network, i.e., BasesHomo, trained on our generated dataset achieves a \\(24\\%\\) performance gain compared to the network trained with the unsupervised learning strategy, further proving the superiority of our dataset generation strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Module Effectiveness. We conduct experiments to evaluate the effectiveness of our proposed CCM and QAM by removing them separately from the overall framework. As shown in rows 2-3 of Table 4, CCM and QAM can help to generate and select high-quality image pairs for training, achieving error reduction of 0.03 and 0.01, respectively, compared to the Baseline in row 1 using none of them. By using both CCM and QAM in the training phase, our method can achieve the best performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.47, + 0.885 + ], + "angle": 0, + "content": "Different Homography Estimation Networks. As described in Sec. 4.2, we choose BasesHomo as the backbone of our H-Net. In rows 4-5 of Table 4, we further switch the backbone to CAHomo and HomoGAN, denoted as Real-CA and Real-GAN, respectively, to prove that our method also works not only on a specific architecture. Compared with the unsupervised results in rows 11 and 13 of Table 1, CAHomo obtains performance gain by \\(50\\%\\) with our framework and HomoGAN reduces the error by 0.09, proving the generalization of our proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Iteration Times. The quality of generated dataset and" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.089, + 0.89, + 0.239 + ], + "angle": 0, + "content": "
AVGRELTLLSFLF
1)Baseline0.380.230.400.500.450.33
2)w/o CCM0.370.230.360.480.440.34
3)w/o QAM0.350.230.360.450.420.30
4)Real-CA [42]0.440.270.450.520.530.43
5)Real-GAN [12]0.300.180.300.390.370.28
6)1 iteration0.380.240.400.460.480.33
7)3 iterations0.330.210.340.440.400.28
8)Default0.340.220.350.440.420.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.249, + 0.892, + 0.275 + ], + "angle": 0, + "content": "Table 4. Results of ablation studies, please refer to the text for more details. w/o denotes without." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.29, + 0.892, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.343, + 0.892, + 0.358 + ], + "angle": 0, + "content": "Figure 8. Illustration of the generated image in different iterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.893, + 0.613 + ], + "angle": 0, + "content": "network performance are mutually reinforcing through iterative learning. As shown in rows 6-8 of Table 4, the performance of H-Net improves with iteration increases, while higher-precision H-Net generates more natural images, as shown in Fig. 8, and high-quality images can improve performance. When the homography estimated by H-Net is accurate enough, there would be no artifacts and boundary discontinuities in the generated images. Notably, \"0 iteration\" indicates that the image is generated by using the homography estimated by unsupervised BasesHomo. However, a certain upper limit exists in our method and it will converge after several iterations. The performance gain of 3 iterations compared to 2 iterations, i.e., default, is not significant, and more iterations are not cost-effective compared to the resources spent per iteration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.631, + 0.62, + 0.646 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.854 + ], + "angle": 0, + "content": "We have presented an iterative deep framework that aims to generate a realistic dataset for supervised homography learning and obtain a high-precision network. We use the dominant plane masks estimated by D-Net and the homography estimated by H-Net to generate training pairs, and the CCM and QAM are proposed to prepare high-quality training data to update the H-Net. The trained network and the generated data can be improved iteratively, yielding a qualified dataset as well as a state-of-the-art network. Extensive experiments demonstrate the superiority of our method and the effectiveness of our newly proposed components, the performance of existing methods can be also improved using our dataset for training." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by Sichuan Science and Technology Program of China under grant No.2023NSFSC0462." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9813" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Simon Baker and Iain Matthews. Lucas-kanade 20 years on: A unifying framework. International Journal of Computer Vision, 56(3):221-255, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[2] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proc. CVPR, pages 10197-10205, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.468, + 0.24 + ], + "angle": 0, + "content": "[3] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In Proc. ECCV, pages 404-417, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.468, + 0.282 + ], + "angle": 0, + "content": "[4] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proc. CVPR, pages 1879-1888, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.468, + 0.325 + ], + "angle": 0, + "content": "[5] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. Clkn: Cascaded lucas-kanade networks for image alignment. In Proc. CVPR, pages 2213–2221, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.468, + 0.422 + ], + "angle": 0, + "content": "[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPRW, pages 224-236, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[8] Tianjiao Ding, Yunchen Yang, Zhihui Zhu, Daniel P Robinson, René Vidal, Laurent Kneip, and Manolis C Tsakiris. Robust homography estimation via dual principal component pursuit. In Proc. CVPR, pages 6080-6089, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.468, + 0.591 + ], + "angle": 0, + "content": "[10] Yunhui Han, Kunming Luo, Ao Luo, Jiangyu Liu, Haoqiang Fan, Guiming Luo, and Shuaicheng Liu. Realflow: Embased realistic optical flow dataset generation from videos. In Proc. ECCV, pages 288-305, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.468, + 0.633 + ], + "angle": 0, + "content": "[11] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.468, + 0.689 + ], + "angle": 0, + "content": "[12] Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, Qijun Zhao, and Shuaicheng Liu. Unsupervised homography estimation with coplanarity-aware gan. In Proc. CVPR, pages 17663-17672, 2022. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.468, + 0.746 + ], + "angle": 0, + "content": "[13] Hai Jiang, Haipeng Li, Yuhang Lu, Songchen Han, and Shuaicheng Liu. Semi-supervised deep large-baseline homography estimation with progressive equivalence constraint. In Proc. AAAI, pages 1024-1032, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[14] Dewi Endah Kharismawati, Hadi Ali Akbarpour, Rumana Aktar, Filiz Bunyak, Kannappan Palaniappan, and Toni Kazic. Cornet: Unsupervised deep homography estimation for agricultural aerial imagery. In Proc. ECCV, pages 400-417, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[15] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proc. CVPR, pages 7652-7661, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[16] Haipeng Li, Kunming Luo, and Shuaicheng Liu. Gyroflow: Gyroscope-guided unsupervised optical flow learning. In Proc. ICCV, pages 12869-12878, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Haipeng Li, Kunming Luo, Bing Zeng, and Shuaicheng Liu. Gyroflow+: Gyroscope-guided unsupervised deep homography and optical flow learning. arXiv preprint arXiv:2301.10018, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[19] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International Journal of Computer Vision, 128(2):261-318, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[20] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(06):7885-7899, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[21] Shuaicheng Liu, Kunming Luo, Ao Luo, Chuan Wang, Fanman Meng, and Bing Zeng. Asflow: Unsupervised optical flow learning with adaptive pyramid sampling. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4282-4295, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[22] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(03):2849-2863, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[23] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM Transactions on Graphics, 32(4):1-10, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[24] Zhen Liu, Wenjie Lin, Xinpeng Li, Qing Rao, Ting Jiang, Mingyan Han, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. Adnet: Attention-guided deformable convolutional network for high dynamic range imaging. In Proc. CVPRW, pages 463-470, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[25] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proc. ECCV, pages 344-360, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[26] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[27] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In Proc. CVPR, pages 8906-8915, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[28] Ao Luo, Fan Yang, Kunming Luo, Xin Li, Haoqiang Fan, and Shuaicheng Liu. Learning optical flow with adaptive graph reasoning. In Proc. AAAI, pages 1890-1898, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[29] Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3523-3542, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[30] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters, 3(3):2346-2353, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9814" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Lang Nie, Chunyu Lin, Kang Liao, Meiqin Liu, and Yao Zhao. A view-free image stitching network based on global homography. Journal of Visual Communication and Image Representation, 73:102950, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[32] Lang Nie, Chunyu Lin, Kang Liao, and Yao Zhao. Learning edge-preserved image stitching from multi-scale deep homography. Neurocomputing, 491:533-543, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[33] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In Proc. ICCV, pages 2564-2571, 2011. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[34] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. CVPR, pages 4938–4947, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.291, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[35] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. Localtrans: A multiscale local transformer network for cross-resolution homography estimation. In Proc. ICCV, pages 14890-14899, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.348, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[36] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proc. CVPR, pages 8922-8931, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.404, + 0.468, + 0.459 + ], + "angle": 0, + "content": "[37] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proc. CVPR, pages 11016-11025, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.461, + 0.468, + 0.515 + ], + "angle": 0, + "content": "[38] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proc. CVPR, pages 6258-6268, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.518, + 0.468, + 0.559 + ], + "angle": 0, + "content": "[39] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proc. CVPR, pages 5714-5724, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.561, + 0.468, + 0.602 + ], + "angle": 0, + "content": "[40] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proc. ECCV, pages 117–132, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.604, + 0.468, + 0.658 + ], + "angle": 0, + "content": "[41] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proc. ICCV, pages 13117-13125, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.66, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[42] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proc. ECCV, pages 653-669, 2020. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.717, + 0.468, + 0.758 + ], + "angle": 0, + "content": "[43] Zhuofan Zhang, Zhen Liu, Bing Zeng, and Shuaicheng Liu. Minimum latency deep online video stabilization. arXiv preprint arXiv:2212.02073, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "9815" + } + ] +] \ No newline at end of file diff --git a/2023/Supervised Homography Learning with Realistic Dataset Generation/56098eb5-516e-4294-9a4a-ac98eb40dc0f_origin.pdf b/2023/Supervised Homography Learning with Realistic Dataset Generation/56098eb5-516e-4294-9a4a-ac98eb40dc0f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6ebbb389598174e14b7fc02c63ee375abcfff92a --- /dev/null +++ b/2023/Supervised Homography Learning with Realistic Dataset Generation/56098eb5-516e-4294-9a4a-ac98eb40dc0f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:316316aca85e23d2d972aa0faabed053bdc61cf3681e2bdb3f5d52f037069fca +size 8957077 diff --git a/2023/Supervised Homography Learning with Realistic Dataset Generation/full.md b/2023/Supervised Homography Learning with Realistic Dataset Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b9eb3e749fd0ba3874cfc448afb52607431b6aa9 --- /dev/null +++ b/2023/Supervised Homography Learning with Realistic Dataset Generation/full.md @@ -0,0 +1,333 @@ +# Supervised Homography Learning with Realistic Dataset Generation + +Hai Jiang $^{1,2,*}$ , Haipeng Li $^{3,2,*}$ , Songchen Han $^{1,\dagger}$ , Haoqiang Fan $^{2}$ , Bing Zeng $^{3}$ , Shuaicheng Liu $^{3,2,\dagger}$ , Sichuan University ${}^{2}$ Megvii Technology + +3University of Electronic Science and Technology of China + +{jianghai@stu.，hansongchen@}scu.edu.cn，{jianghai,lihaipeng,fhq,liushuaicheng}@megvii.com + +{lihaipeng@std.,eezeng@,liushuaicheng@}uestc.edu.cn + +# Abstract + +In this paper, we propose an iterative framework, which consists of two phases: a generation phase and a training phase, to generate realistic training data and yield a supervised homography network. In the generation phase, given an unlabeled image pair, we utilize the pre-estimated dominant plane masks and homography of the pair, along with another sampled homography that serves as ground truth to generate a new labeled training pair with realistic motion. In the training phase, the generated data is used to train the supervised homography network, in which the training data is refined via a content consistency module and a quality assessment module. Once an iteration is finished, the trained network is used in the next data generation phase to update the pre-estimated homography. Through such an iterative strategy, the quality of the dataset and the performance of the network can be gradually and simultaneously improved. Experimental results show that our method achieves state-of-the-art performance and existing supervised methods can be also improved based on the generated dataset. Code and dataset are available at https://github.com/JianghaiSCU/RealSH. + +# 1. Introduction + +Homography estimation is a fundamental task in computer vision that has been widely used for high-dynamic range imaging [40, 24, 25], image stitching [32, 31], and video stabilization [23, 43]. Traditional methods typically use feature extraction and matching methods [26, 3] with outlier suppression [9, 2] to obtain feature matches of two images and subsequently solve direct linear transformation (DLT) [11] to obtain the homography matrix. However, these methods highly rely on the extracted matching keypoints and may crash in challenging scenes that lack + +![](images/cca1b9c37ccec0ef3c38635ef1ab380e38e4e77cf712d4b7404db8f32e7d0de2.jpg) + +![](images/bc41691050c75e5230aa69d2a139f122481c0d909fa84bfd5bc243f252b5c401.jpg) +label criteria +(a) Previous strategy +Figure 1. The first row shows the target images generated by the previous strategy [6] and our proposed method, visualized by superimposing the source image warped by GT homography on the generated target image, where the misaligned pixels are represented as colored ghosts. As highlighted in yellow boxes, the dominant plane is fully aligned by the GT homography, proving that both our method and the previous strategy can satisfy the label criteria. In contrast, only our method can maintain realistic motion between the two images thus satisfying the realism criteria, as shown in the red boxes, the car is a moving object that cannot be aligned by a homography. + +![](images/b1fd3a3f3f5d4305438a467b9263d9e80f1101b530b749babe3d672c8e5f297e.jpg) +realism criteria X + +![](images/5a8a20c97b43fb0e1072cc6fb8185c4d5ed0331d5af8f495affa85ef14d8e238.jpg) + +![](images/a7858f74645cd23c5a437eb66751531051e3ab051a43a8066827179cd52a3751.jpg) +label criteria +(b) Ours + +![](images/4c8429cefc18058be4258d158bab5f04d760ee9f47e7e3b28c9444f95f3b97b6.jpg) +realism criteria + +sufficient high-quality feature matches. With the rise of deep learning, such problems have been partially solved, learning-based methods [22, 20, 13] take a pair of images as input and directly output the corresponding homography, thus are more robust than traditional methods due to their keypoint-free estimation strategy. The learning-based methods can be divided into two categories: supervised and unsupervised. At the moment, benefiting from the label-free characteristic, unsupervised methods can be trained on enormous amounts of real-world data, delivering better performance and generalization capability than supervised ones. We find that the lack of qualified training data is one of the main barriers to the development of supervised methods. Previously supervised dataset generation strategy [6] synthesizes an image pair by using pre-defined ground truth (GT) homography to warp a single image, such strategy considers the whole image as a single plane, neglecting par + +allax and foreground motion in the real world. As shown in Fig. 1(a), the two images are fully aligned using the GT homography, but the car in the red box should be a moving object that cannot be aligned by a homography. + +In this work, we propose an iterative framework, which is designed to generate image pairs that satisfy both label and realism criteria [10] for supervised homography learning and learns a state-of-the-art homography estimation network with the generated dataset. Specifically, given an unlabeled source and target image $(I_s$ and $I_t)$ captured in real-world scenes, we use the homography and dominant plane masks of the two images estimated by the homography estimation network and a pre-trained dominant plane detection network, respectively, along with a sampled GT homography to generate a new target image $I_t'$ . Then, we use the $I_s$ and $I_t'$ , together with the sampled GT homography, to compose a training sample of our supervised dataset. The dominant planes of $I_s$ and $I_t'$ can be fully aligned by the GT homography while the rest remain in realistic motion, making the two images satisfy both label criteria and realism criteria. As shown in Fig. 1(b), the dominant plane of the generated image pair can be fully aligned by the GT homography, while the scene parallax and dynamic objects are maintained in the foreground. + +In addition, few artifacts could exist in the generated target image in early iterations, we, therefore, propose a content consistency module and a quality assessment module to eliminate the unexpected content and select high-quality image pairs for training, respectively. The selected image pairs are used to update the homography estimation network which is utilized to estimate homographies for image generation in the next iteration. During the iterative learning steps, the capability of the network is gradually improved, as well as the quality of the synthesized dataset. With our framework, any unlabeled image pairs can be used to generate training samples, thus addressing the lack of qualified datasets in supervised homography learning, which improves the performance of supervised methods and enables them to be better generalized to real-world scenes. In summary, our main contributions are threefold: + +- We propose an iterative deep framework to generate a realistic dataset from unlabeled real-world image pairs for supervised homography learning and simultaneously obtain a high-precision network based on the generated dataset. +- We propose a content consistency module and a quality assessment module, achieving the elimination of unexpected content and the selection of qualified data for training. +- Experimental results show that our method brings noticeable image realism improvement compared to the + +prior dataset generation strategy and achieves state-of-the-art performance on public benchmarks. + +# 2. Related Work + +# 2.1. Traditional Method + +Traditional homography estimation methods can be divided into feature-based methods and optimization-based methods. Feature-based methods usually combine feature extraction and matching algorithms [26, 3, 37, 7, 34, 36] with outliner suppression approaches [9, 2], followed by solving DLT [11] to obtain homography. However, such methods would crash in challenging scenes where sufficient feature matches cannot be obtained. Optimization-based methods [5, 8] use the Lucas-Kanade algorithm [1] or calculate the sum of squared differences between two images to optimize an initialized homography iteratively. But optimization-based approaches are time-consuming and suffer from cumulative errors. + +# 2.2. Learning-based Method + +Following the development of deep image alignment methods, such as optical flow [16, 21, 10, 27, 28] and dense correspondence [39, 38], the first deep homography estimation network was proposed by [6]. Deep learning-based methods can be divided into two categories: supervised and unsupervised. The former ones [15, 35, 4] utilize synthetic image pairs that are derived from single images for training, which suffer from lacking dynamic objects and realistic scene parallax, hindering their generalization ability to real-world scenes. Unsupervised methods [30, 14, 42, 41, 12] are more robust than supervised ones thanks to their label-free training strategies. However, existing unsupervised methods are mostly optimized by minimizing the photometric distance from the warped source to the target, being sensitive to dynamic objects and homogeneous regions [38]. Besides, unsupervised methods are unstable and difficult to converge during the training phase compared to supervised ones. Therefore, our work aims to generate a realistic dataset with homography labels to optimize the network in a supervised manner and enable the network to better generalize to unseen scenes. + +# 2.3. Deep Homography Estimation Dataset + +The MS-COCO [18] and CA-unsup [42] datasets are commonly used for supervised and unsupervised homography learning. Recently, the GHOF dataset [17] is introduced for unsupervised gyroscope-guided optical flow and homography estimation. The MS-COCO dataset only contains source images, the target images are obtained by warping source images using pre-defined GT homographies. As such, the synthetic image pairs satisfy the label criteria but lack realism criteria. The CA-unsup and GHOF + +![](images/9bdb6ca1525b1d6ba4515bc4cd3f502cede4ef633bbfec9a9c277c127ce5a7e5.jpg) +Figure 2. The overall pipeline of our proposed framework. In the generation phase, we synthesize a new target image and form a training sample with the source image according to the dominant plane masks and homography of the original image pair estimated by D-Net and H-Net. In the training phase, we propose a content consistency module and a quality assessment module to prepare qualified image pairs for training H-Net that is adopted for image generation in the next iteration. Moreover, we perform the same pre-processing on the training pairs as in the previous method [6, 42, 41, 12], including central cropping, grayscale transformation, and normalization. + +![](images/a9e0502a7291ea9d3caa37524b3a41a8582728f610a56bd505f1f5c11a532dfc.jpg) + +datasets collect image pairs from consecutive real-world video frames that satisfy the realism criteria but lack GT labels. In contrast, our method can synthesize realistic image pairs with labels from any unlabeled image pairs, instead of single images, to satisfy both label and realism criteria. + +# 3. Method + +# 3.1. Overview + +The overall pipeline of our method is illustrated in Fig. 2. The core innovation of our method is that generating training data and estimating homography are mutually reinforcing. We, therefore, integrate dataset generation and network training into an iterative process, as shown in Fig. 2(a). Our method consists of two phases: the generation phase (G-phase) and the training phase (T-phase). + +- G-phase: Given an unlabeled source image $I_{s}$ and target image $I_{t}$ , we generate a new target image $I_{t}^{\prime}$ according to a pre-defined homography $\mathbf{H}_{gt}$ as well as the homography and dominant plane masks of the two images estimated by the homography estimation network (H-Net) denoted as $\Theta_{H}$ and the dominant plane detection network (D-Net) as $\Theta_{D}$ , forming a training sample as $x = (I_s,I_t',\mathbf{H}_{gt})$ , i.e., + +$$ +x = \mathrm {G} \left(I _ {s}, I _ {t}, \Theta_ {D}, \Theta_ {H}\right). \tag {1} +$$ + +- T-phase: Using the generated training samples $X = \{x\}$ to update the H-Net with the help of a content consistency module (CCM) denoted as $\Theta_{c}$ and a quality assessment module (QAM) as $\Theta_{q}$ , i.e, + +$$ +\Theta_ {H} ^ {\prime} = \underset {\theta} {\arg \min } \mathcal {L} (X, \Theta_ {H}, \Theta_ {c}, \Theta_ {q}), \tag {2} +$$ + +where $\Theta_H^{\prime}$ denotes the H-Net retrained on the training samples and $\mathcal{L}$ is the learning objective. + +Performing G-phase and T-phase interactively can yield a qualified dataset and a high-precision network. + +# 3.2. Generation Phase + +As shown in Fig. 2(b), we aim to generate a new target image $I_{t}^{\prime}$ according to the source image $I_{s}$ and target image $I_{t}$ along with a sampled homography $\mathbf{H}_{gt}$ to serves as the label between $I_{s}$ and $I_{t}^{\prime}$ , which is modeled by randomly selecting scaling, shearing, rotation, translation, and perspective factors from pre-defined small-baseline ranges. As a result, a training sample that satisfies both label and realism criteria [10] is generated as $(I_{s}, I_{t}^{\prime}, \mathbf{H}_{gt})$ . + +Specifically, we first adopt a dominant plane detection network (D-Net), which is formed as a UNet-like structure and pre-trained in an unsupervised manner, to estimate the dominant plane masks $M_{s}$ and $M_{t}$ of the $I_{s}$ and $I_{t}$ , and segment the dominant plane region $P_{d}^{s} = I_{s} \cdot M_{s}$ and nondominant plane region $P_{n}^{t} = I_{t} \cdot (1 - M_{t})$ . By warping the $P_{d}^{s}$ with $\mathbf{H}_{gt}$ to form the counterpart region of $I_{t}^{\prime}$ , we ensure that their dominant planes are fully aligned. To model the realistic motion between two images, we construct the nondominant plane region of $I_{t}^{\prime}$ from $I_{t}$ . One simple approach is to directly fuse the warped $P_{d}^{s}$ and $P_{n}^{t}$ as + +$$ +I _ {t} ^ {\prime} = \mathcal {W} \left(P _ {d} ^ {s} + P _ {n} ^ {t}, \mathbf {H} _ {g t}\right), \tag {3} +$$ + +where $\mathcal{W}(\cdot)$ represents the warp operation. However, there exists relative motion between the two images, and the masks may not be accurate and dense enough to segment continuous and complete regions, artlessly using Eq.(3) to form the new target image would produce discontinuities in + +![](images/8ef909d36ee711a2a2f3d3e746a6da9eca8b9fa18ff6ad80708d9e2373c575eb.jpg) +Source $I_{s}$ + +![](images/cdbaa3701d8f669dc0497fa1040c2a14470ef3dfb763d86d0e84a60ed8baed3a.jpg) +Target $I_{t}$ +Figure 3. Illustration of generated images by using different fusion functions. The image fused by using Eq.(3) leads to unsmooth edges and artifacts while using Eq.(4) renders a natural image. + +![](images/efc4208481579acb72cb80d9870d6f2aaa353c9d9dd4b7674cae61a6a187652c.jpg) +$M_{s}$ + +![](images/14747d079f88a84e02d6cec7618d274c211602127c081a6c717ad44c3dbbd86c.jpg) +1-Mt + +![](images/a07ae69a55c9f555d46001a5dc03c3cc754a972b868e01c11d0b66f989270111.jpg) + +![](images/5d1493168d7b841bdcf5a10e97644dc69afbce5d6c0ab985860f001b0f117322.jpg) +$l_{t}^{\prime}$ generated by using Eq.(4) + +the fusion boundary and artifacts. To solve this problem, we construct the non-dominant plane region of $I_{t}^{\prime}$ by warping the counterpart of $I_{t}$ with the cumulative multiplication result of the homography from target to source $\mathbf{H}_{ts}$ and $\mathbf{H}_{gt}$ , and the final $I_{t}^{\prime}$ can be obtained as + +$$ +I _ {t} ^ {\prime} = \mathcal {W} \left(P _ {d} ^ {s}, \mathbf {H} _ {g t}\right) + \mathcal {W} \left(P _ {n} ^ {t}, \mathbf {H} _ {g t} \mathbf {H} _ {t s}\right). \tag {4} +$$ + +As shown in Fig. 3, despite the inaccuracy of $M_{s}$ and $M_{t}$ , artifacts are avoided and image smoothness is guaranteed using Eq.(4). + +# 3.3. Training Phase + +Through the G-phase, we obtain a large-scale realistic dataset for improving the homography estimation network (H-Net). As mentioned above, the quality of generated images depends on the accuracy of the estimated dominant plane masks and homography, few artifacts could exist in generated images in early iterations. To this end, we propose a content consistency module and a quality assessment module to refine the generated image and further select qualified image pairs for training, as shown in Fig. 2(c). + +Content Consistency Module. The content consistency module (CCM) is designed to eliminate artifacts in $I_{t}^{\prime}$ to improve the content quality. As described in Sec. 3.2, the dominant plane of the $I_{s}$ and $I_{t}^{\prime}$ can be fully aligned by the $\mathbf{H}_{gt}$ as $P_{d}^{t} = \mathcal{W}(P_{d}^{s},\mathbf{H}_{gt})$ , and the $P_d^t$ and $P_d^s$ can be aligned by the $\mathbf{H}_{ts}$ as $P_{d}^{s} = \mathcal{W}(P_{d}^{t},\mathbf{H}_{ts})$ once the $\mathbf{H}_{ts}$ is accurate. Therefore, the dominant plane of $I_{t}^{\prime}$ can be converted into the counterpart of $I_{t}$ as + +$$ +P _ {d} ^ {t ^ {\prime}} = \mathcal {W} (\mathcal {W} (P _ {d} ^ {t}, \mathbf {H} _ {t s}), \mathbf {H} _ {g t}) = \mathcal {W} (P _ {d} ^ {t}, \mathbf {H} _ {g t} \mathbf {H} _ {t s}). \quad (5) +$$ + +Besides, the non-dominant plane of $I_{t}^{\prime}$ is obtained by warping the non-dominant plane of $I_{t}$ with the accumulated homography $\mathbf{H}_{gt}\mathbf{H}_{ts}$ . Therefore, where $\mathbf{H}_{ts}$ is extremely accurate, the $I_{t}^{\prime}$ can be obtained by warping $I_{t}$ with the accumulated homography as $I_{t}^{\prime} = \mathcal{W}(I_{t},\mathbf{H}_{gt}\mathbf{H}_{ts})$ , which is the assumption behind the CCM we designed. To achieve this, a content consistency loss $\mathcal{L}_{ccl}$ is proposed as + +$$ +\mathcal {L} _ {c c l} = \left| \mathcal {W} \left(\mathcal {F} \left(\hat {I} _ {t} ^ {\prime}\right), \left(\mathbf {H} _ {g t} \mathbf {H} _ {t s}\right) ^ {- 1}\right) - \mathcal {F} \left(I _ {t}\right) \right| _ {1}, \tag {6} +$$ + +![](images/55e32fb461799ca7f8379e8c369446bb1dfdb19bc6777089496f8c2638be6141.jpg) +New Target $I_{t}^{1}$ +(a) Illustration of the target image reconstructed by CCM. + +![](images/ab95e9026e34242d5d3e77da500cc413c707956130091b09f727dce7bac324b4.jpg) +Reconstructed Target $\hat{I}_t^{\prime}$ + +![](images/88f7099e994891d0997bce70ecf42add3642d7f7f516c1bf72a279712e517c04.jpg) +Disturbed $I_{r}$ + +![](images/e299bc1bacbc87bf713afc6bc8fb5f6471e62c983b80528958efad1cd448046d.jpg) +Reconstructed $\hat{I}_t^{\prime}$ +Original $I_{t}$ + +![](images/9eb171ef59a95e0ddd038865f0c3bfdda306bbb72fe049ac0c0decbeba781e05.jpg) + +![](images/f8ab21ad1a716a67a850a8dd5522923cc66dbc62a3331f0635850746b8412ca3.jpg) +(b) Illustration of the quality scores estimated by QAM. +Figure 4. Illustration of the capabilities of our proposed modules. (a) shows that the artifacts are successfully removed by applying our CCM, yielding high-quality images for training. (b) shows the quality scores of different images estimated by QAM, where high-quality images are used for training and inversely rejected. + +where $\hat{I}_t^\prime$ is the reconstructed result of $I_t^\prime$ by CCM, i.e., $\hat{I}_t^\prime = \Theta_c(I_t^\prime)$ , and $\mathcal{F}(\cdot)$ is the feature extractor of H-Net. Instead of directly minimizing the content distance between $\hat{I}_t^\prime$ and $I_{t}$ we constrain the similarity of the features to prompt CCM to reconstruct $\hat{I}_t^\prime$ to be sharp as the real-world image, since the $\mathbf{H}_{ts}$ is not accurate enough. As shown in Fig. 4(a), the artifacts in $I_t^\prime$ have been successfully removed by CCM. + +Quality Assessment Module. The quality assessment module (QAM) is designed to filter bad cases that cannot be reconstructed by CCM, further improving the quality of the training data. Specifically, the QAM is designed to be a classification network whose output $x \in \mathbb{R}$ ranges from 0 to 1, $x$ being closer to 1 means better quality and vice versa. + +To achieve this, we sample another homography to interfere with $\mathbf{H}_{ts}$ since the quality of the generated image is partially relied on the accuracy of $\mathbf{H}_{ts}$ , producing a disturbing image $I_{r}$ with artifacts through G-phase. The score $x_{r}$ obtained by feeding $I_{r}$ into QAM should be close to 0. In contrast, the original target image $I_{t}$ is captured from real-world scenes, so the score $x_{t}$ of $I_{t}$ should be close to 1. Thus, the quality assessment loss $\mathcal{L}_{qal}$ is proposed to optimize the QAM as + +$$ +\mathcal {L} _ {q a l} = \operatorname {B C E} \left(x _ {t}, 1\right) + \operatorname {B C E} \left(x _ {r}, 0\right), \tag {7} +$$ + +where $\mathrm{BCE}(\cdot)$ denotes the binary cross-entropy loss. As a result, when the score $\hat{x}_t^\prime$ of $\hat{I}_t^\prime$ exceeds a certain threshold $\tau$ , we consider $\hat{I}_t^\prime$ to be a high-quality image that can be used for training and inversely be rejected, as shown in Fig. 4(b). + +Overall, with the help of CCM and QAM, qualified training pairs can be generated, satisfying both realism criteria and supervision labels to improve the H-Net. Through iter + +ative learning, the performance of the network and the quality of the generated dataset are mutually improved. + +# 3.4. Network Training + +Since the training samples have GT labels, we can optimize the H-Net by minimizing the difference between $\mathbf{H}_{gt}$ and the estimated $\mathbf{H}_{st'}$ . However, as mentioned in [6], it is non-trivial to directly estimate a tomography, we follow previous works [15, 4, 35] to use the 4 corners offset vectors $D_{\mathbf{H}_{st'}}$ and $D_{\mathbf{H}_{gt}}$ as supervision objectives, the corresponding homography could be computed by solving DLT [11] using the offset vectors. In addition, we compute the bidirectional homography as the $\mathbf{H}_{t's}$ should also be aligned with $\mathbf{H}_{gt}^{-1}$ . The finally supervised loss $\mathcal{L}_{sup}$ is expressed as + +$$ +\mathcal {L} _ {s u p} = \left| D _ {\mathbf {H} _ {s t ^ {\prime}}} - D _ {\mathbf {H} _ {g t}} \right| _ {1} + \left| D _ {\mathbf {H} _ {t ^ {\prime} s}} - D _ {\mathbf {H} _ {g t} ^ {- 1}} \right| _ {1}. \tag {8} +$$ + +The total loss $\mathcal{L}_{\text {total }}$ is expressed by combing the supervised loss, the content consistency loss, and the quality assessment loss as + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {s u p}} + \lambda_ {1} \mathcal {L} _ {\text {c c l}} + \lambda_ {2} \mathcal {L} _ {\text {q a l}}, \tag {9} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are empirically set as 0.5 and 0.1. + +# 4. Experiment + +# 4.1. Dataset + +MS-COCO. The MS-COCO dataset [18] is one of the widely used datasets in the fields of object detection [19], segmentation [29], etc. Previously supervised methods [6, 15, 35, 4] used it to prepare their training datasets by perturbing each corner of single images to produce GT homographies and warp the images using the homographies to generate training pairs. + +CA-Unsup. The CA-unsup dataset [42] contains 800k training pairs and 4.2k testing pairs that are captured from five types of real-world scenes, i.e., regular (RE), low texture (LT), low light (LL), small foreground (SF), and large foreground (LF), where the last four are challenging scenes for homography estimation. The training set contains only unlabeled image pairs, but for each pair of the test set, $6\sim 10$ equally distributed matching points are manually marked for evaluation. Based on the CA-unsup dataset, we generated a new dataset that contains the same number of image pairs satisfying both label criteria and realism criteria as the CA-unusp dataset through our proposed method, named CA-sup. We show some examples of our CA-sup dataset in the supplementary materials. + +GHOF. The GHOF dataset [17] is designed to evaluate homography and optical flow estimation with the gyroscope readings, it contains 10k training data and 300 testing pairs under 5 different categories, including regular (RE), foggy + +(FOG), low light (LL), rainy (RAIN), and snowy (SNOW) scenes. Compared to the CA-unsup dataset, it contains more parallax motion variations and extreme foreground ratio which are challenging to the homography methods. The sparse $5\sim 8$ correspondences are used to evaluate the results. + +# 4.2. Implementation Details + +Our framework consists of D-Net, CCM, QAM, and H-Net. The D-Net is pre-trained on the CA-unsup dataset in an unsupervised manner, the detailed architecture and implementation details are described in the supplementary material. We choose BasesHomo [41], which balances performance and efficiency as the backbone of H-Net, and convert its output representation into 4 corners offset vectors. The CCM and QAM are training with the H-Net for 2 iterations, and the training parameters are consistent with the official implementation of BasesHomo for each iteration. + +# 4.3. Comparison with Existing Methods + +Comparison Methods. We compare our method with three categories of existing homography estimation methods: 1) feature-based methods including SIFT [26], ORB [33], SuperPoint [7] with SuperGlue [34] (SPSG), and LoFTR [36], 2) supervised methods including DHN [6], LocalTrans [35], and IHN [4], 3) unsupervised methods including CAHomo [42], BasesHomo [41], and HomogAN [12]. For feature-based methods, we improve them with two different outlier rejection algorithms RANSAC [9] and MAGSAC [2], respectively. For supervised methods, we employ their models pre-trained on the MS-COCO dataset [18] and the models retrained on our CA-sup dataset denoted with $*$ for evaluation. + +Quantitative Comparison. We report the quantitative results of all comparison methods on the CA-unsup test set in Table 1. The $\mathcal{I}_{3\times 3}$ in the first row refers to a $3\times 3$ identity matrix as a "no-warping" homography for reference, of which the errors reflect the original distance between point pairs. As shown in Table 1, our method achieves state-of-the-art performance in all categories and outperforms the best existing method HomoGAN by $12.82\%$ , with the points matching error (PME) reduced from 0.39 to 0.34. The feature-based methods can perform well in regular (RE) scenes, as sufficient matching points can be obtained, while our methods and HomoGAN reduce the error by 0.08 compared to SIFT+RANSAC. More specifically, our method produces a PME of 0.216 which is lower than 0.222 of HomoGAN. The small foreground (SF) and large foreground (LF) scenes contain dynamic objects that affect the accuracy of estimated homography, and our method has the lowest PME in these two categories, even when compared to the methods utilizing outlier removal masks for robust estimation, i.e., CAHomo and HomoGAN. In low texture (LT) and low light (LL) scenes, most learning-based + +
1)MethodsAVGRELTLLSFLF
2)\( \mathcal{I}_{3\times 3} \)6.70 (+1617.95%)7.75 (+3422.73%)7.65 (+1765.85%)7.21 (+1164.91%)7.53 (+1611.36%)3.39 (+993.55%)
3)SIFT [26]+RANSAC [9]1.41 (+261.54%)0.30 (+36.36%)1.34 (+226.83%)4.03 (+607.02%)0.81 (+84.09%)0.57 (+83.87%)
4)SIFT [26]+MAGSAC [2]1.34 (+243.59%)0.31 (+40.91%)1.72 (+319.51%)3.39 (+494.74%)0.80 (+81.82%)0.47 (51.61%)
5)ORB [33]+RANSAC [9]1.48 (+279.49%)0.85 (+286.36%)2.59 (+531.71%)1.67 (+192.98%)1.10 (+150.00%)1.24 (+300.00%)
6)ORB [33]+MAGSAC [2]1.69 (+333.33%)0.97 (+340.91%)3.34 (+714.63%)1.58 (+177.19%)1.15 (+161.36%)1.40 (+351.61%)
7)SPSG [7, 34]+RANSAC [9]0.71 (+82.05%)0.41 (+86.36%)0.87 (+112.20%)0.72 (+26.32%)0.80 (+81.81%)0.75 (+141.94%)
8)SPSG [7, 34]+MAGSAC [2]0.63 (+61.54%)0.36 (+63.64%)0.79 (+92.68%)0.70 (+22.81%)0.71 (+61.36%)0.70 (+125.81%)
9)LoFTR [36]+RANSAC [9]1.44 (+269.23%)0.56 (+154.55%)2.70 (+558.54%)1.36 (+138.60%)1.05 (+138.64%)1.52 (+390.32%)
10)LoFTR [36]+MAGSAC [2]1.39 (+256.41%)0.55 (+150.00%)2.57 (+526.83%)1.33 (+133.33%)1.05 (+138.64%)1.41 (+354.84%)
11)CAHomo [42]0.88 (+125.64%)0.73 (+231.82%)1.01 (+146.34%)1.03 (+80.70%)0.92 (+109.09%)0.70 (+125.81%)
12)BasesHomo [41]0.50 (+28.21%)0.29 (+31.82%)0.54 (+31.71%)0.65 (+14.04%)0.61 (+38.64%)0.41 (+32.26%)
13)HomoGAN [12]0.39 (+0.00%)0.22 (+0.00%)0.41 (+0.00%)0.57 (+0.00%)0.44 (+0.00%)0.31 (+0.00%)
14)DHN [6]2.87 (+635.90%)1.51 (+586.36%)4.48 (+992.68%)2.76 (+384.21%)2.62 (+495.45%)3.00 (+867.74%)
15)LocalTrans [35]4.21 (+978.26%)4.09 (+1757.59%)4.84 (+1081.14%)4.55 (+697.98%)5.30 (+1105.20%)2.25 (+624.29%)
16)IHN [4]4.67 (+1097.44%)4.85 (+2104.55%)5.54 (+1251.22%)5.10 (+794.74%)5.04 (+1045.45%)2.84 (+816.13%)
17)DHN* [6]1.89 (+384.95%)1.21 (+451.82%)2.13 (+418.54%)2.33 (+307.91%)1.72 (+291.23%)2.07 (+567.81%)
18)LocalTrans* [35]1.76 (+352.03%)1.25 (+467.41%)2.15 (+423.85%)1.90 (+232.61%)2.25 (+410.41%)1.28 (+311.77%)
19)IHN* [4]1.19 (+205.54%)0.72 (+227.09%)1.74 (+324.93%)1.18 (+107.42%)1.30 (+196.52%)1.01 (+225.58%)
20)Ours0.34 (-12.82%)0.22 (+0.00%)0.35 (-14.63%)0.44 (-22.81%)0.42 (-4.55%)0.29 (-6.45%)
+ +Table 1. The point matching errors (PME) of our method and all comparison methods on the CA-sup [42] test set. The best and second-best results are highlighted in bold and underlined. The percentages in the parentheses indicate the relative change in comparison to the second-best results. SPSG indicates SuperPoint with SuperGlue, and * denotes the methods are retrained on our CA-sup dataset. + +
1)MethodsAVGREFOGLLRAINSNOW
2)I3×36.334.947.248.095.485.89
3)SIFT [26]+MAGSAC [2]3.900.644.019.770.704.40
4)ORB [33]+RANSAC [9]15.146.9231.2727.801.827.90
5)SPSG [7, 34]+MAGSAC [2]3.283.411.467.610.753.19
6)LoFTR [36]+MAGSAC [2]2.552.761.165.340.522.98
7)CAHomo [42]3.874.103.846.991.273.17
8)BasesHomo [41]2.282.021.434.900.782.29
9)HomoGAN [12]1.951.730.603.950.473.02
10)DHN [6]6.616.046.027.686.996.32
11)LocalTrans [35]5.724.066.495.955.786.34
12)IHN [4]8.177.108.719.346.579.13
13)DHN* [6]3.011.923.944.541.982.66
14)LocalTrans* [35]2.891.784.274.591.372.43
15)IHN* [4]2.592.213.054.700.982.03
16)Ours1.721.600.884.420.431.28
+ +Table 2. The point matching errors (PME) of our method and comparison methods on the GHOF [17] test set. The best and second-best results are highlighted in bold and underlined. + +methods are more robust due to their keypoint-free estimation strategy, especially unsupervised ones. However, homogeneous region [38] dominates a large portion of the image in the LT and LL scenes, where the photometric loss between the two images is minor, so the performance of unsupervised methods in these two categories is not as well as in others. In contrast, our method outperforms the unsupervised methods in the LT and LL with errors reduced by at least $14.63\%$ and $22.81\%$ respectively, thanks to our photometric-free fully supervised learning strategy. + +Moreover, the generalization ability of supervised methods has been widely criticized. Therefore, we further evaluate feature-based methods that performed well on the CA-unsup dataset, learning-based methods, and our method on the GHOF test set to prove the effectiveness of our method. + +As shown in Table 2, our method achieves the best results in RAIN and SNOW scenes and the second-best results in RE, FOG and LL scenes, resulting in the lowest average PME. For regular scenes, even though feature-based methods can extract sufficient high-quality feature matches, we still outperform most of them. In addition, our CA-sup dataset does not contain images captured in the FOG and RAIN scenes, while our method is capable of generalizing these scenes, which proves the superiority of our method. On the other hand, previous supervised methods cannot generalize well to real-world scenes since their training data lack realistic motion. After retraining on our CA-sup dataset, their performance has been significantly improved, which further proves the effectiveness of our framework. + +Qualitative Comparison. In Fig. 5 and Fig. 6, we provide the qualitative results of our method and competitive methods on the CA-unsup and GHOF test sets, respectively. Fig. 5(a) and (b) are from LF and SF scenes with dynamic objects, and our method produces more accurate results than feature-based and unsupervised methods which are sensitive to moving objects, as highlighted in the red and green boxes. Fig. 5(c) is from the LT scene, where feature-based methods all fail due to the lack of sufficient correspondence. On the contrary, learning-based methods perform well, but they are still not competitive as ours. In Fig. 6, we provide the comparison results in RAIN and FOG scenes of the GHOF test set, where our training set does not contain images captured in such scenes. As mentioned above, while unsupervised methods have better generalization ability in unseen scenes, we prove that supervised methods can also achieve satisfactory results after training on our generated dataset. For example, in Fig. 6(a), the pre + +![](images/f673f975c65a8c38f15579105572119d4d41f382c701d8674efc339dc9377038.jpg) +Figure 5. Qualitative results of our method and other competitive methods on the CA-unsup [42] test set. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, and the green boxes show the content difference between the two images in the error-prone regions. Best viewed by zooming in. + +![](images/68fa1c046d9ac77b6e87c1d6cf62c962a876c9b4e420c2bb9588b998bd834e53.jpg) +Figure 6. Qualitative results of our method and other competitive methods on the GHOF [17] test set. Best viewed by zooming in. + +vious supervised methods can even obtain more accurate results than CAHomo, and our method performs the best. + +Comparison with Dataset Generation Methods. We use images from the CA unsup dataset to generate a new dataset named CA-sup $^{\dagger}$ using the previous dataset generation strategy [6], and retrain H-Net on it for comparison. + +
DatasetAVGRELTLLSFLF
DHN [6]CA-sup†1.370.791.701.281.231.83
OursCA-sup0.340.220.350.440.420.29
+ +Table 3. Comparison with previous dataset generation method. We use the same source images to generate the training dataset and train the same network for comparison. + +Quantitative results on the CA-unsup test set are illustrated in Table 3. The previous strategy generates image pairs from single images that meet label criteria only, causing unsatisfactory performance. In contrast, the image pairs contained in our CA-sup dataset can help the network to learn real-world motion representations, proving the importance of realistic scene motion for estimating homography. + +Robustness Evaluation. Furthermore, we evaluate the robustness of all comparison methods on the CA-unsup test set by setting thresholds to calculate the proportion of inlier predictions. As shown in Fig. 7, we plot a series of curves where axis $X \in [0.1, 3.0]$ is the threshold and axis $Y \in [0.0, 1.0]$ is the percentage of inliers, higher position of curves represents better robustness. With a threshold of 0.1, our outlier percentage is $6.9\%$ higher than the second-best method (15.8% vs. $8.9\%$ ), and our method achieves a proportion of $99.9\%$ with a threshold of 3.0. + +![](images/2ed1caa4e02b272300721236eab804d93b5440ac7fcbed1b2dc33da15001c42c.jpg) +Figure 7. The proportion of inliers of our method and all comparison methods under various thresholds. The higher position of curves represents better robustness. + +# 4.4. Ablation Study + +In this section, we conduct a series of ablation studies to measure the impact of different component choices of our method. We use the CA-sup dataset for training, and quantitative results on the CA-unsup test set are illustrated in Table 4, detailed experiment settings are discussed below. For more ablation experiments, please refer to the supplementary material. + +Dataset Generation Strategy. We conduct an experiment to evaluate the effectiveness of our dataset generation strategy by using G-phase only to generate training samples to train the network, denoted as Baseline in row 1 of Table 4. Even without using CCM and QAM, our method achieves state-of-the-art performance compared to the existing methods. Moreover, the homography estimation network, i.e., BasesHomo, trained on our generated dataset achieves a $24\%$ performance gain compared to the network trained with the unsupervised learning strategy, further proving the superiority of our dataset generation strategy. + +Module Effectiveness. We conduct experiments to evaluate the effectiveness of our proposed CCM and QAM by removing them separately from the overall framework. As shown in rows 2-3 of Table 4, CCM and QAM can help to generate and select high-quality image pairs for training, achieving error reduction of 0.03 and 0.01, respectively, compared to the Baseline in row 1 using none of them. By using both CCM and QAM in the training phase, our method can achieve the best performance. + +Different Homography Estimation Networks. As described in Sec. 4.2, we choose BasesHomo as the backbone of our H-Net. In rows 4-5 of Table 4, we further switch the backbone to CAHomo and HomoGAN, denoted as Real-CA and Real-GAN, respectively, to prove that our method also works not only on a specific architecture. Compared with the unsupervised results in rows 11 and 13 of Table 1, CAHomo obtains performance gain by $50\%$ with our framework and HomoGAN reduces the error by 0.09, proving the generalization of our proposed framework. + +Iteration Times. The quality of generated dataset and + +
AVGRELTLLSFLF
1)Baseline0.380.230.400.500.450.33
2)w/o CCM0.370.230.360.480.440.34
3)w/o QAM0.350.230.360.450.420.30
4)Real-CA [42]0.440.270.450.520.530.43
5)Real-GAN [12]0.300.180.300.390.370.28
6)1 iteration0.380.240.400.460.480.33
7)3 iterations0.330.210.340.440.400.28
8)Default0.340.220.350.440.420.29
+ +Table 4. Results of ablation studies, please refer to the text for more details. w/o denotes without. + +![](images/03ce53b99f9cacb75f67a218b60825cd6c73aa5278daa5ee43a042522901fe10.jpg) +Figure 8. Illustration of the generated image in different iterations. + +network performance are mutually reinforcing through iterative learning. As shown in rows 6-8 of Table 4, the performance of H-Net improves with iteration increases, while higher-precision H-Net generates more natural images, as shown in Fig. 8, and high-quality images can improve performance. When the homography estimated by H-Net is accurate enough, there would be no artifacts and boundary discontinuities in the generated images. Notably, "0 iteration" indicates that the image is generated by using the homography estimated by unsupervised BasesHomo. However, a certain upper limit exists in our method and it will converge after several iterations. The performance gain of 3 iterations compared to 2 iterations, i.e., default, is not significant, and more iterations are not cost-effective compared to the resources spent per iteration. + +# 5. Conclusion + +We have presented an iterative deep framework that aims to generate a realistic dataset for supervised homography learning and obtain a high-precision network. We use the dominant plane masks estimated by D-Net and the homography estimated by H-Net to generate training pairs, and the CCM and QAM are proposed to prepare high-quality training data to update the H-Net. The trained network and the generated data can be improved iteratively, yielding a qualified dataset as well as a state-of-the-art network. Extensive experiments demonstrate the superiority of our method and the effectiveness of our newly proposed components, the performance of existing methods can be also improved using our dataset for training. + +Acknowledgements. This work is supported by Sichuan Science and Technology Program of China under grant No.2023NSFSC0462. + +# References + +[1] Simon Baker and Iain Matthews. Lucas-kanade 20 years on: A unifying framework. International Journal of Computer Vision, 56(3):221-255, 2004. 2 +[2] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proc. CVPR, pages 10197-10205, 2019. 1, 2, 5, 6 +[3] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In Proc. ECCV, pages 404-417, 2006. 1, 2 +[4] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proc. CVPR, pages 1879-1888, 2022. 2, 5, 6 +[5] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. Clkn: Cascaded lucas-kanade networks for image alignment. In Proc. CVPR, pages 2213–2221, 2017. 2 +[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 1, 2, 3, 5, 6, 7 +[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPRW, pages 224-236, 2018. 2, 5, 6 +[8] Tianjiao Ding, Yunchen Yang, Zhihui Zhu, Daniel P Robinson, René Vidal, Laurent Kneip, and Manolis C Tsakiris. Robust homography estimation via dual principal component pursuit. In Proc. CVPR, pages 6080-6089, 2020. 2 +[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2, 5, 6 +[10] Yunhui Han, Kunming Luo, Ao Luo, Jiangyu Liu, Haoqiang Fan, Guiming Luo, and Shuaicheng Liu. Realflow: Embased realistic optical flow dataset generation from videos. In Proc. ECCV, pages 288-305, 2022. 2, 3 +[11] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 2, 5 +[12] Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, Qijun Zhao, and Shuaicheng Liu. Unsupervised homography estimation with coplanarity-aware gan. In Proc. CVPR, pages 17663-17672, 2022. 2, 3, 5, 6, 8 +[13] Hai Jiang, Haipeng Li, Yuhang Lu, Songchen Han, and Shuaicheng Liu. Semi-supervised deep large-baseline homography estimation with progressive equivalence constraint. In Proc. AAAI, pages 1024-1032, 2023. 1 +[14] Dewi Endah Kharismawati, Hadi Ali Akbarpour, Rumana Aktar, Filiz Bunyak, Kannappan Palaniappan, and Toni Kazic. Cornet: Unsupervised deep homography estimation for agricultural aerial imagery. In Proc. ECCV, pages 400-417, 2020. 2 +[15] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proc. CVPR, pages 7652-7661, 2020. 2, 5 +[16] Haipeng Li, Kunming Luo, and Shuaicheng Liu. Gyroflow: Gyroscope-guided unsupervised optical flow learning. In Proc. ICCV, pages 12869-12878, 2021. 2 + +[17] Haipeng Li, Kunming Luo, Bing Zeng, and Shuaicheng Liu. Gyroflow+: Gyroscope-guided unsupervised deep homography and optical flow learning. arXiv preprint arXiv:2301.10018, 2023. 2, 5, 6, 7 +[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755, 2014. 2, 5 +[19] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International Journal of Computer Vision, 128(2):261-318, 2020. 5 +[20] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(06):7885-7899, 2023. 1 +[21] Shuaicheng Liu, Kunming Luo, Ao Luo, Chuan Wang, Fanman Meng, and Bing Zeng. Asflow: Unsupervised optical flow learning with adaptive pyramid sampling. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4282-4295, 2022. 2 +[22] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(03):2849-2863, 2023. 1 +[23] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM Transactions on Graphics, 32(4):1-10, 2013. 1 +[24] Zhen Liu, Wenjie Lin, Xinpeng Li, Qing Rao, Ting Jiang, Mingyan Han, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. Adnet: Attention-guided deformable convolutional network for high dynamic range imaging. In Proc. CVPRW, pages 463-470, 2021. 1 +[25] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proc. ECCV, pages 344-360, 2022. 1 +[26] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 5, 6 +[27] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In Proc. CVPR, pages 8906-8915, 2022. 2 +[28] Ao Luo, Fan Yang, Kunming Luo, Xin Li, Haoqiang Fan, and Shuaicheng Liu. Learning optical flow with adaptive graph reasoning. In Proc. AAAI, pages 1890-1898, 2022. 2 +[29] Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3523-3542, 2022. 5 +[30] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters, 3(3):2346-2353, 2018. 2 + +[31] Lang Nie, Chunyu Lin, Kang Liao, Meiqin Liu, and Yao Zhao. A view-free image stitching network based on global homography. Journal of Visual Communication and Image Representation, 73:102950, 2020. 1 +[32] Lang Nie, Chunyu Lin, Kang Liao, and Yao Zhao. Learning edge-preserved image stitching from multi-scale deep homography. Neurocomputing, 491:533-543, 2022. 1 +[33] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In Proc. ICCV, pages 2564-2571, 2011. 5, 6 +[34] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. CVPR, pages 4938–4947, 2020. 2, 5, 6 +[35] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. Localtrans: A multiscale local transformer network for cross-resolution homography estimation. In Proc. ICCV, pages 14890-14899, 2021. 2, 5, 6 +[36] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proc. CVPR, pages 8922-8931, 2021. 2, 5, 6 +[37] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proc. CVPR, pages 11016-11025, 2019. 2 +[38] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proc. CVPR, pages 6258-6268, 2020. 2, 6 +[39] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proc. CVPR, pages 5714-5724, 2021. 2 +[40] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proc. ECCV, pages 117–132, 2018. 1 +[41] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proc. ICCV, pages 13117-13125, 2021. 2, 3, 5, 6 +[42] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proc. ECCV, pages 653-669, 2020. 2, 3, 5, 6, 7, 8 +[43] Zhuofan Zhang, Zhen Liu, Bing Zeng, and Shuaicheng Liu. Minimum latency deep online video stabilization. arXiv preprint arXiv:2212.02073, 2022. 1 \ No newline at end of file diff --git a/2023/Supervised Homography Learning with Realistic Dataset Generation/images.zip b/2023/Supervised Homography Learning with Realistic Dataset Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c4914d824b03359bf3d15756f3454cc2df929897 --- /dev/null +++ b/2023/Supervised Homography Learning with Realistic Dataset Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60fa1a8ea4e079457e7e5fefb0684dcd2283cba43bf32d4e123f7e5adf124c2e +size 845261 diff --git a/2023/Supervised Homography Learning with Realistic Dataset Generation/layout.json b/2023/Supervised Homography Learning with Realistic Dataset Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d7e071f2379e01846d76dfb99bc71d0ccd6ab75a --- /dev/null +++ b/2023/Supervised Homography Learning with Realistic Dataset Generation/layout.json @@ -0,0 +1,9899 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "text", + "content": "Supervised Homography Learning with Realistic Dataset Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "spans": [ + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "text", + "content": "Hai Jiang" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "text", + "content": ", Haipeng Li" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "inline_equation", + "content": "^{3,2,*}" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "text", + "content": ", Songchen Han" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "text", + "content": ", Haoqiang Fan" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "text", + "content": ", Bing Zeng" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "text", + "content": ", Shuaicheng Liu" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "inline_equation", + "content": "^{3,2,\\dagger}" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "text", + "content": ", Sichuan University " + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 73, + 148, + 525, + 177 + ], + "type": "text", + "content": "Megvii Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 177, + 441, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 177, + 441, + 191 + ], + "spans": [ + { + "bbox": [ + 156, + 177, + 441, + 191 + ], + "type": "text", + "content": "3University of Electronic Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 193, + 543, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 543, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 543, + 205 + ], + "type": "text", + "content": "{jianghai@stu.，hansongchen@}scu.edu.cn，{jianghai,lihaipeng,fhq,liushuaicheng}@megvii.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 161, + 208, + 436, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 208, + 436, + 219 + ], + "spans": [ + { + "bbox": [ + 161, + 208, + 436, + 219 + ], + "type": "text", + "content": "{lihaipeng@std.,eezeng@,liushuaicheng@}uestc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "spans": [ + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 272, + 290, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 290, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 290, + 523 + ], + "type": "text", + "content": "In this paper, we propose an iterative framework, which consists of two phases: a generation phase and a training phase, to generate realistic training data and yield a supervised homography network. In the generation phase, given an unlabeled image pair, we utilize the pre-estimated dominant plane masks and homography of the pair, along with another sampled homography that serves as ground truth to generate a new labeled training pair with realistic motion. In the training phase, the generated data is used to train the supervised homography network, in which the training data is refined via a content consistency module and a quality assessment module. Once an iteration is finished, the trained network is used in the next data generation phase to update the pre-estimated homography. Through such an iterative strategy, the quality of the dataset and the performance of the network can be gradually and simultaneously improved. Experimental results show that our method achieves state-of-the-art performance and existing supervised methods can be also improved based on the generated dataset. Code and dataset are available at https://github.com/JianghaiSCU/RealSH." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 546, + 128, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 128, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 128, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": "Homography estimation is a fundamental task in computer vision that has been widely used for high-dynamic range imaging [40, 24, 25], image stitching [32, 31], and video stabilization [23, 43]. Traditional methods typically use feature extraction and matching methods [26, 3] with outlier suppression [9, 2] to obtain feature matches of two images and subsequently solve direct linear transformation (DLT) [11] to obtain the homography matrix. However, these methods highly rely on the extracted matching keypoints and may crash in challenging scenes that lack" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 246, + 425, + 312 + ], + "blocks": [ + { + "bbox": [ + 309, + 246, + 425, + 312 + ], + "lines": [ + { + "bbox": [ + 309, + 246, + 425, + 312 + ], + "spans": [ + { + "bbox": [ + 309, + 246, + 425, + 312 + ], + "type": "image", + "image_path": "cca1b9c37ccec0ef3c38635ef1ab380e38e4e77cf712d4b7404db8f32e7d0de2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 313, + 359, + 339 + ], + "blocks": [ + { + "bbox": [ + 309, + 313, + 359, + 339 + ], + "lines": [ + { + "bbox": [ + 309, + 313, + 359, + 339 + ], + "spans": [ + { + "bbox": [ + 309, + 313, + 359, + 339 + ], + "type": "image", + "image_path": "bc41691050c75e5230aa69d2a139f122481c0d909fa84bfd5bc243f252b5c401.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 339, + 359, + 350 + ], + "lines": [ + { + "bbox": [ + 310, + 339, + 359, + 350 + ], + "spans": [ + { + "bbox": [ + 310, + 339, + 359, + 350 + ], + "type": "text", + "content": "label criteria" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 351, + 405, + 361 + ], + "lines": [ + { + "bbox": [ + 328, + 351, + 405, + 361 + ], + "spans": [ + { + "bbox": [ + 328, + 351, + 405, + 361 + ], + "type": "text", + "content": "(a) Previous strategy" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 365, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 487 + ], + "type": "text", + "content": "Figure 1. The first row shows the target images generated by the previous strategy [6] and our proposed method, visualized by superimposing the source image warped by GT homography on the generated target image, where the misaligned pixels are represented as colored ghosts. As highlighted in yellow boxes, the dominant plane is fully aligned by the GT homography, proving that both our method and the previous strategy can satisfy the label criteria. In contrast, only our method can maintain realistic motion between the two images thus satisfying the realism criteria, as shown in the red boxes, the car is a moving object that cannot be aligned by a homography." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 361, + 313, + 424, + 339 + ], + "blocks": [ + { + "bbox": [ + 361, + 313, + 424, + 339 + ], + "lines": [ + { + "bbox": [ + 361, + 313, + 424, + 339 + ], + "spans": [ + { + "bbox": [ + 361, + 313, + 424, + 339 + ], + "type": "image", + "image_path": "b1fd3a3f3f5d4305438a467b9263d9e80f1101b530b749babe3d672c8e5f297e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 339, + 422, + 349 + ], + "lines": [ + { + "bbox": [ + 365, + 339, + 422, + 349 + ], + "spans": [ + { + "bbox": [ + 365, + 339, + 422, + 349 + ], + "type": "text", + "content": "realism criteria X" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 427, + 246, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 427, + 246, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 427, + 246, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 427, + 246, + 544, + 312 + ], + "type": "image", + "image_path": "5a8a20c97b43fb0e1072cc6fb8185c4d5ed0331d5af8f495affa85ef14d8e238.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 428, + 313, + 478, + 339 + ], + "blocks": [ + { + "bbox": [ + 428, + 313, + 478, + 339 + ], + "lines": [ + { + "bbox": [ + 428, + 313, + 478, + 339 + ], + "spans": [ + { + "bbox": [ + 428, + 313, + 478, + 339 + ], + "type": "image", + "image_path": "a7858f74645cd23c5a437eb66751531051e3ab051a43a8066827179cd52a3751.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 339, + 472, + 350 + ], + "lines": [ + { + "bbox": [ + 430, + 339, + 472, + 350 + ], + "spans": [ + { + "bbox": [ + 430, + 339, + 472, + 350 + ], + "type": "text", + "content": "label criteria" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 351, + 503, + 361 + ], + "lines": [ + { + "bbox": [ + 470, + 351, + 503, + 361 + ], + "spans": [ + { + "bbox": [ + 470, + 351, + 503, + 361 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 480, + 313, + 544, + 339 + ], + "blocks": [ + { + "bbox": [ + 480, + 313, + 544, + 339 + ], + "lines": [ + { + "bbox": [ + 480, + 313, + 544, + 339 + ], + "spans": [ + { + "bbox": [ + 480, + 313, + 544, + 339 + ], + "type": "image", + "image_path": "4c8429cefc18058be4258d158bab5f04d760ee9f47e7e3b28c9444f95f3b97b6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 339, + 541, + 350 + ], + "lines": [ + { + "bbox": [ + 485, + 339, + 541, + 350 + ], + "spans": [ + { + "bbox": [ + 485, + 339, + 541, + 350 + ], + "type": "text", + "content": "realism criteria" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "sufficient high-quality feature matches. With the rise of deep learning, such problems have been partially solved, learning-based methods [22, 20, 13] take a pair of images as input and directly output the corresponding homography, thus are more robust than traditional methods due to their keypoint-free estimation strategy. The learning-based methods can be divided into two categories: supervised and unsupervised. At the moment, benefiting from the label-free characteristic, unsupervised methods can be trained on enormous amounts of real-world data, delivering better performance and generalization capability than supervised ones. We find that the lack of qualified training data is one of the main barriers to the development of supervised methods. Previously supervised dataset generation strategy [6] synthesizes an image pair by using pre-defined ground truth (GT) homography to warp a single image, such strategy considers the whole image as a single plane, neglecting par" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9806" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "allax and foreground motion in the real world. As shown in Fig. 1(a), the two images are fully aligned using the GT homography, but the car in the red box should be a moving object that cannot be aligned by a homography." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "text", + "content": "In this work, we propose an iterative framework, which is designed to generate image pairs that satisfy both label and realism criteria [10] for supervised homography learning and learns a state-of-the-art homography estimation network with the generated dataset. Specifically, given an unlabeled source and target image " + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "inline_equation", + "content": "(I_s" + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "inline_equation", + "content": "I_t)" + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "text", + "content": " captured in real-world scenes, we use the homography and dominant plane masks of the two images estimated by the homography estimation network and a pre-trained dominant plane detection network, respectively, along with a sampled GT homography to generate a new target image " + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "inline_equation", + "content": "I_t'" + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "text", + "content": ". Then, we use the " + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "inline_equation", + "content": "I_t'" + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "text", + "content": ", together with the sampled GT homography, to compose a training sample of our supervised dataset. The dominant planes of " + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "inline_equation", + "content": "I_t'" + }, + { + "bbox": [ + 46, + 121, + 288, + 360 + ], + "type": "text", + "content": " can be fully aligned by the GT homography while the rest remain in realistic motion, making the two images satisfy both label criteria and realism criteria. As shown in Fig. 1(b), the dominant plane of the generated image pair can be fully aligned by the GT homography, while the scene parallax and dynamic objects are maintained in the foreground." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 361, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 552 + ], + "type": "text", + "content": "In addition, few artifacts could exist in the generated target image in early iterations, we, therefore, propose a content consistency module and a quality assessment module to eliminate the unexpected content and select high-quality image pairs for training, respectively. The selected image pairs are used to update the homography estimation network which is utilized to estimate homographies for image generation in the next iteration. During the iterative learning steps, the capability of the network is gradually improved, as well as the quality of the synthesized dataset. With our framework, any unlabeled image pairs can be used to generate training samples, thus addressing the lack of qualified datasets in supervised homography learning, which improves the performance of supervised methods and enables them to be better generalized to real-world scenes. In summary, our main contributions are threefold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 561, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 561, + 286, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 561, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 59, + 561, + 286, + 621 + ], + "type": "text", + "content": "- We propose an iterative deep framework to generate a realistic dataset from unlabeled real-world image pairs for supervised homography learning and simultaneously obtain a high-precision network based on the generated dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 632, + 286, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 632, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 59, + 632, + 286, + 679 + ], + "type": "text", + "content": "- We propose a content consistency module and a quality assessment module, achieving the elimination of unexpected content and the selection of qualified data for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 689, + 286, + 713 + ], + "type": "text", + "content": "- Experimental results show that our method brings noticeable image realism improvement compared to the" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 544, + 96 + ], + "type": "text", + "content": "prior dataset generation strategy and achieves state-of-the-art performance on public benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 107, + 392, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 107, + 392, + 120 + ], + "spans": [ + { + "bbox": [ + 306, + 107, + 392, + 120 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 127, + 420, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 127, + 420, + 139 + ], + "spans": [ + { + "bbox": [ + 306, + 127, + 420, + 139 + ], + "type": "text", + "content": "2.1. Traditional Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 145, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 545, + 300 + ], + "type": "text", + "content": "Traditional homography estimation methods can be divided into feature-based methods and optimization-based methods. Feature-based methods usually combine feature extraction and matching algorithms [26, 3, 37, 7, 34, 36] with outliner suppression approaches [9, 2], followed by solving DLT [11] to obtain homography. However, such methods would crash in challenging scenes where sufficient feature matches cannot be obtained. Optimization-based methods [5, 8] use the Lucas-Kanade algorithm [1] or calculate the sum of squared differences between two images to optimize an initialized homography iteratively. But optimization-based approaches are time-consuming and suffer from cumulative errors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 309, + 441, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 441, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 441, + 322 + ], + "type": "text", + "content": "2.2. Learning-based Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 327, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 579 + ], + "type": "text", + "content": "Following the development of deep image alignment methods, such as optical flow [16, 21, 10, 27, 28] and dense correspondence [39, 38], the first deep homography estimation network was proposed by [6]. Deep learning-based methods can be divided into two categories: supervised and unsupervised. The former ones [15, 35, 4] utilize synthetic image pairs that are derived from single images for training, which suffer from lacking dynamic objects and realistic scene parallax, hindering their generalization ability to real-world scenes. Unsupervised methods [30, 14, 42, 41, 12] are more robust than supervised ones thanks to their label-free training strategies. However, existing unsupervised methods are mostly optimized by minimizing the photometric distance from the warped source to the target, being sensitive to dynamic objects and homogeneous regions [38]. Besides, unsupervised methods are unstable and difficult to converge during the training phase compared to supervised ones. Therefore, our work aims to generate a realistic dataset with homography labels to optimize the network in a supervised manner and enable the network to better generalize to unseen scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 586, + 509, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 509, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 509, + 600 + ], + "type": "text", + "content": "2.3. Deep Homography Estimation Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "The MS-COCO [18] and CA-unsup [42] datasets are commonly used for supervised and unsupervised homography learning. Recently, the GHOF dataset [17] is introduced for unsupervised gyroscope-guided optical flow and homography estimation. The MS-COCO dataset only contains source images, the target images are obtained by warping source images using pre-defined GT homographies. As such, the synthetic image pairs satisfy the label criteria but lack realism criteria. The CA-unsup and GHOF" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9807" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 208, + 240 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 208, + 240 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 208, + 240 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 208, + 240 + ], + "type": "image", + "image_path": "9bdb6ca1525b1d6ba4515bc4cd3f502cede4ef633bbfec9a9c277c127ce5a7e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 547, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 299 + ], + "type": "text", + "content": "Figure 2. The overall pipeline of our proposed framework. In the generation phase, we synthesize a new target image and form a training sample with the source image according to the dominant plane masks and homography of the original image pair estimated by D-Net and H-Net. In the training phase, we propose a content consistency module and a quality assessment module to prepare qualified image pairs for training H-Net that is adopted for image generation in the next iteration. Moreover, we perform the same pre-processing on the training pairs as in the previous method [6, 42, 41, 12], including central cropping, grayscale transformation, and normalization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 209, + 71, + 545, + 240 + ], + "blocks": [ + { + "bbox": [ + 209, + 71, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 209, + 71, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 209, + 71, + 545, + 240 + ], + "type": "image", + "image_path": "a9e0502a7291ea9d3caa37524b3a41a8582728f610a56bd505f1f5c11a532dfc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 318, + 288, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 377 + ], + "type": "text", + "content": "datasets collect image pairs from consecutive real-world video frames that satisfy the realism criteria but lack GT labels. In contrast, our method can synthesize realistic image pairs with labels from any unlabeled image pairs, instead of single images, to satisfy both label and realism criteria." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 391, + 102, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 102, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 102, + 403 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 411, + 115, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 115, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 115, + 423 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 430, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 514 + ], + "type": "text", + "content": "The overall pipeline of our method is illustrated in Fig. 2. The core innovation of our method is that generating training data and estimating homography are mutually reinforcing. We, therefore, integrate dataset generation and network training into an iterative process, as shown in Fig. 2(a). Our method consists of two phases: the generation phase (G-phase) and the training phase (T-phase)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "text", + "content": "- G-phase: Given an unlabeled source image " + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "text", + "content": " and target image " + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "text", + "content": ", we generate a new target image " + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "text", + "content": " according to a pre-defined homography " + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{gt}" + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "text", + "content": " as well as the homography and dominant plane masks of the two images estimated by the homography estimation network (H-Net) denoted as " + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\Theta_{H}" + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "text", + "content": " and the dominant plane detection network (D-Net) as " + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\Theta_{D}" + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "text", + "content": ", forming a training sample as " + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "inline_equation", + "content": "x = (I_s,I_t',\\mathbf{H}_{gt})" + }, + { + "bbox": [ + 46, + 516, + 287, + 601 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 611, + 287, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 611, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 116, + 611, + 287, + 623 + ], + "type": "interline_equation", + "content": "x = \\mathrm {G} \\left(I _ {s}, I _ {t}, \\Theta_ {D}, \\Theta_ {H}\\right). \\tag {1}", + "image_path": "c0e716e9b26d1912c4eea35092b436254bdf67009aff6c38a81487d1ddba3849.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "text", + "content": "- T-phase: Using the generated training samples " + }, + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "inline_equation", + "content": "X = \\{x\\}" + }, + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "text", + "content": " to update the H-Net with the help of a content consistency module (CCM) denoted as " + }, + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\Theta_{c}" + }, + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "text", + "content": " and a quality assessment module (QAM) as " + }, + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "inline_equation", + "content": "\\Theta_{q}" + }, + { + "bbox": [ + 47, + 635, + 287, + 684 + ], + "type": "text", + "content": ", i.e," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 93, + 695, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 695, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 93, + 695, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\Theta_ {H} ^ {\\prime} = \\underset {\\theta} {\\arg \\min } \\mathcal {L} (X, \\Theta_ {H}, \\Theta_ {c}, \\Theta_ {q}), \\tag {2}", + "image_path": "9c983cac7ee9774f69e0a28963d4e91ee3d038d67ea88df36f40a81bdb0d8fd6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 317, + 544, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 317, + 544, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 317, + 544, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 317, + 544, + 342 + ], + "type": "inline_equation", + "content": "\\Theta_H^{\\prime}" + }, + { + "bbox": [ + 305, + 317, + 544, + 342 + ], + "type": "text", + "content": " denotes the H-Net retrained on the training samples and " + }, + { + "bbox": [ + 305, + 317, + 544, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 317, + 544, + 342 + ], + "type": "text", + "content": " is the learning objective." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 342, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 365 + ], + "type": "text", + "content": "Performing G-phase and T-phase interactively can yield a qualified dataset and a high-precision network." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 373, + 411, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 411, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 411, + 384 + ], + "type": "text", + "content": "3.2. Generation Phase" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "text", + "content": "As shown in Fig. 2(b), we aim to generate a new target image " + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "text", + "content": " according to the source image " + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "text", + "content": " and target image " + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "text", + "content": " along with a sampled homography " + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{gt}" + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "text", + "content": " to serves as the label between " + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "text", + "content": ", which is modeled by randomly selecting scaling, shearing, rotation, translation, and perspective factors from pre-defined small-baseline ranges. As a result, a training sample that satisfies both label and realism criteria [10] is generated as " + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "inline_equation", + "content": "(I_{s}, I_{t}^{\\prime}, \\mathbf{H}_{gt})" + }, + { + "bbox": [ + 304, + 392, + 545, + 490 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": "Specifically, we first adopt a dominant plane detection network (D-Net), which is formed as a UNet-like structure and pre-trained in an unsupervised manner, to estimate the dominant plane masks " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": ", and segment the dominant plane region " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "P_{d}^{s} = I_{s} \\cdot M_{s}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": " and nondominant plane region " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "P_{n}^{t} = I_{t} \\cdot (1 - M_{t})" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": ". By warping the " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "P_{d}^{s}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{gt}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": " to form the counterpart region of " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": ", we ensure that their dominant planes are fully aligned. To model the realistic motion between two images, we construct the nondominant plane region of " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": ". One simple approach is to directly fuse the warped " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "P_{d}^{s}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "inline_equation", + "content": "P_{n}^{t}" + }, + { + "bbox": [ + 304, + 489, + 546, + 621 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 374, + 630, + 545, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 630, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 374, + 630, + 545, + 645 + ], + "type": "interline_equation", + "content": "I _ {t} ^ {\\prime} = \\mathcal {W} \\left(P _ {d} ^ {s} + P _ {n} ^ {t}, \\mathbf {H} _ {g t}\\right), \\tag {3}", + "image_path": "08e6af2dd369abe25361cbbfee66da1a407153c33696491cc2d04aa992c0296a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\cdot)" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " represents the warp operation. However, there exists relative motion between the two images, and the masks may not be accurate and dense enough to segment continuous and complete regions, artlessly using Eq.(3) to form the new target image would produce discontinuities in" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9808" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 124, + 114 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 124, + 114 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 124, + 114 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 124, + 114 + ], + "type": "image", + "image_path": "8ef909d36ee711a2a2f3d3e746a6da9eca8b9fa18ff6ad80708d9e2373c575eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 115, + 101, + 121 + ], + "lines": [ + { + "bbox": [ + 75, + 115, + 101, + 121 + ], + "spans": [ + { + "bbox": [ + 75, + 115, + 101, + 121 + ], + "type": "text", + "content": "Source " + }, + { + "bbox": [ + 75, + 115, + 101, + 121 + ], + "type": "inline_equation", + "content": "I_{s}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 122, + 124, + 164 + ], + "blocks": [ + { + "bbox": [ + 51, + 122, + 124, + 164 + ], + "lines": [ + { + "bbox": [ + 51, + 122, + 124, + 164 + ], + "spans": [ + { + "bbox": [ + 51, + 122, + 124, + 164 + ], + "type": "image", + "image_path": "cdbaa3701d8f669dc0497fa1040c2a14470ef3dfb763d86d0e84a60ed8baed3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 163, + 100, + 172 + ], + "lines": [ + { + "bbox": [ + 76, + 163, + 100, + 172 + ], + "spans": [ + { + "bbox": [ + 76, + 163, + 100, + 172 + ], + "type": "text", + "content": "Target " + }, + { + "bbox": [ + 76, + 163, + 100, + 172 + ], + "type": "inline_equation", + "content": "I_{t}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 176, + 287, + 209 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 287, + 209 + ], + "type": "text", + "content": "Figure 3. Illustration of generated images by using different fusion functions. The image fused by using Eq.(3) leads to unsmooth edges and artifacts while using Eq.(4) renders a natural image." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 125, + 72, + 198, + 114 + ], + "blocks": [ + { + "bbox": [ + 125, + 72, + 198, + 114 + ], + "lines": [ + { + "bbox": [ + 125, + 72, + 198, + 114 + ], + "spans": [ + { + "bbox": [ + 125, + 72, + 198, + 114 + ], + "type": "image", + "image_path": "efc4208481579acb72cb80d9870d6f2aaa353c9d9dd4b7674cae61a6a187652c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 115, + 165, + 122 + ], + "lines": [ + { + "bbox": [ + 155, + 115, + 165, + 122 + ], + "spans": [ + { + "bbox": [ + 155, + 115, + 165, + 122 + ], + "type": "inline_equation", + "content": "M_{s}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 125, + 123, + 198, + 164 + ], + "blocks": [ + { + "bbox": [ + 125, + 123, + 198, + 164 + ], + "lines": [ + { + "bbox": [ + 125, + 123, + 198, + 164 + ], + "spans": [ + { + "bbox": [ + 125, + 123, + 198, + 164 + ], + "type": "image", + "image_path": "14747d079f88a84e02d6cec7618d274c211602127c081a6c717ad44c3dbbd86c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 163, + 171, + 173 + ], + "lines": [ + { + "bbox": [ + 149, + 163, + 171, + 173 + ], + "spans": [ + { + "bbox": [ + 149, + 163, + 171, + 173 + ], + "type": "text", + "content": "1-Mt" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 198, + 72, + 287, + 122 + ], + "blocks": [ + { + "bbox": [ + 198, + 72, + 287, + 122 + ], + "lines": [ + { + "bbox": [ + 198, + 72, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 198, + 72, + 287, + 122 + ], + "type": "image", + "image_path": "a07ae69a55c9f555d46001a5dc03c3cc754a972b868e01c11d0b66f989270111.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 213, + 122, + 286, + 163 + ], + "blocks": [ + { + "bbox": [ + 213, + 122, + 286, + 163 + ], + "lines": [ + { + "bbox": [ + 213, + 122, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 213, + 122, + 286, + 163 + ], + "type": "image", + "image_path": "5d1493168d7b841bdcf5a10e97644dc69afbce5d6c0ab985860f001b0f117322.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 163, + 283, + 172 + ], + "lines": [ + { + "bbox": [ + 217, + 163, + 283, + 172 + ], + "spans": [ + { + "bbox": [ + 217, + 163, + 283, + 172 + ], + "type": "inline_equation", + "content": "l_{t}^{\\prime}" + }, + { + "bbox": [ + 217, + 163, + 283, + 172 + ], + "type": "text", + "content": " generated by using Eq.(4)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "text", + "content": "the fusion boundary and artifacts. To solve this problem, we construct the non-dominant plane region of " + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "text", + "content": " by warping the counterpart of " + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "text", + "content": " with the cumulative multiplication result of the homography from target to source " + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{ts}" + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{gt}" + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "text", + "content": ", and the final " + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 46, + 225, + 287, + 286 + ], + "type": "text", + "content": " can be obtained as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 87, + 293, + 287, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 293, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 87, + 293, + 287, + 308 + ], + "type": "interline_equation", + "content": "I _ {t} ^ {\\prime} = \\mathcal {W} \\left(P _ {d} ^ {s}, \\mathbf {H} _ {g t}\\right) + \\mathcal {W} \\left(P _ {n} ^ {t}, \\mathbf {H} _ {g t} \\mathbf {H} _ {t s}\\right). \\tag {4}", + "image_path": "fd750fd4e7026e52c13a5cd0901dced1b3ad6ac058bc573e007f5379d9eabd57.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 312, + 287, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 349 + ], + "type": "text", + "content": "As shown in Fig. 3, despite the inaccuracy of " + }, + { + "bbox": [ + 46, + 312, + 287, + 349 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 46, + 312, + 287, + 349 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 312, + 287, + 349 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 46, + 312, + 287, + 349 + ], + "type": "text", + "content": ", artifacts are avoided and image smoothness is guaranteed using Eq.(4)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 354, + 140, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 140, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 140, + 367 + ], + "type": "text", + "content": "3.3. Training Phase" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 373, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 480 + ], + "type": "text", + "content": "Through the G-phase, we obtain a large-scale realistic dataset for improving the homography estimation network (H-Net). As mentioned above, the quality of generated images depends on the accuracy of the estimated dominant plane masks and homography, few artifacts could exist in generated images in early iterations. To this end, we propose a content consistency module and a quality assessment module to refine the generated image and further select qualified image pairs for training, as shown in Fig. 2(c)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": "Content Consistency Module. The content consistency module (CCM) is designed to eliminate artifacts in " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " to improve the content quality. As described in Sec. 3.2, the dominant plane of the " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " can be fully aligned by the " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{gt}" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "P_{d}^{t} = \\mathcal{W}(P_{d}^{s},\\mathbf{H}_{gt})" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " , and the " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "P_d^t" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "P_d^s" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " can be aligned by the " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{ts}" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "P_{d}^{s} = \\mathcal{W}(P_{d}^{t},\\mathbf{H}_{ts})" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " once the " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{ts}" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " is accurate. Therefore, the dominant plane of " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " can be converted into the counterpart of " + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 481, + 287, + 581 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 586, + 287, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 586, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 59, + 586, + 287, + 604 + ], + "type": "interline_equation", + "content": "P _ {d} ^ {t ^ {\\prime}} = \\mathcal {W} (\\mathcal {W} (P _ {d} ^ {t}, \\mathbf {H} _ {t s}), \\mathbf {H} _ {g t}) = \\mathcal {W} (P _ {d} ^ {t}, \\mathbf {H} _ {g t} \\mathbf {H} _ {t s}). \\quad (5)", + "image_path": "8d22474c49f95151dc328af163d92d54911378b040d75f31763c6818033bfcc4.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": "Besides, the non-dominant plane of " + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": " is obtained by warping the non-dominant plane of " + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": " with the accumulated homography " + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{gt}\\mathbf{H}_{ts}" + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": ". Therefore, where " + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{ts}" + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": " is extremely accurate, the " + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": " can be obtained by warping " + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": " with the accumulated homography as " + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime} = \\mathcal{W}(I_{t},\\mathbf{H}_{gt}\\mathbf{H}_{ts})" + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": ", which is the assumption behind the CCM we designed. To achieve this, a content consistency loss " + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ccl}" + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": " is proposed as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 76, + 699, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 699, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 76, + 699, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c c l} = \\left| \\mathcal {W} \\left(\\mathcal {F} \\left(\\hat {I} _ {t} ^ {\\prime}\\right), \\left(\\mathbf {H} _ {g t} \\mathbf {H} _ {t s}\\right) ^ {- 1}\\right) - \\mathcal {F} \\left(I _ {t}\\right) \\right| _ {1}, \\tag {6}", + "image_path": "792569a0492866cd0bbcf5a08073a59d342f37684d3a2d19316837d94c2ec3f5.jpg" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 415, + 129 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 415, + 129 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 415, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 415, + 129 + ], + "type": "image", + "image_path": "55e32fb461799ca7f8379e8c369446bb1dfdb19bc6777089496f8c2638be6141.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 129, + 377, + 138 + ], + "lines": [ + { + "bbox": [ + 336, + 129, + 377, + 138 + ], + "spans": [ + { + "bbox": [ + 336, + 129, + 377, + 138 + ], + "type": "text", + "content": "New Target " + }, + { + "bbox": [ + 336, + 129, + 377, + 138 + ], + "type": "inline_equation", + "content": "I_{t}^{1}" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 140, + 507, + 148 + ], + "lines": [ + { + "bbox": [ + 345, + 140, + 507, + 148 + ], + "spans": [ + { + "bbox": [ + 345, + 140, + 507, + 148 + ], + "type": "text", + "content": "(a) Illustration of the target image reconstructed by CCM." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 416, + 72, + 544, + 128 + ], + "blocks": [ + { + "bbox": [ + 416, + 72, + 544, + 128 + ], + "lines": [ + { + "bbox": [ + 416, + 72, + 544, + 128 + ], + "spans": [ + { + "bbox": [ + 416, + 72, + 544, + 128 + ], + "type": "image", + "image_path": "ab95e9026e34242d5d3e77da500cc413c707956130091b09f727dce7bac324b4.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 129, + 529, + 138 + ], + "lines": [ + { + "bbox": [ + 460, + 129, + 529, + 138 + ], + "spans": [ + { + "bbox": [ + 460, + 129, + 529, + 138 + ], + "type": "text", + "content": "Reconstructed Target " + }, + { + "bbox": [ + 460, + 129, + 529, + 138 + ], + "type": "inline_equation", + "content": "\\hat{I}_t^{\\prime}" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 307, + 161, + 386, + 204 + ], + "blocks": [ + { + "bbox": [ + 329, + 152, + 365, + 160 + ], + "lines": [ + { + "bbox": [ + 329, + 152, + 365, + 160 + ], + "spans": [ + { + "bbox": [ + 329, + 152, + 365, + 160 + ], + "type": "text", + "content": "Disturbed " + }, + { + "bbox": [ + 329, + 152, + 365, + 160 + ], + "type": "inline_equation", + "content": "I_{r}" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 161, + 386, + 204 + ], + "lines": [ + { + "bbox": [ + 307, + 161, + 386, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 386, + 204 + ], + "type": "image", + "image_path": "88f7099e994891d0997bce70ecf42add3642d7f7f516c1bf72a279712e517c04.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 387, + 161, + 465, + 204 + ], + "blocks": [ + { + "bbox": [ + 402, + 152, + 449, + 161 + ], + "lines": [ + { + "bbox": [ + 402, + 152, + 449, + 161 + ], + "spans": [ + { + "bbox": [ + 402, + 152, + 449, + 161 + ], + "type": "text", + "content": "Reconstructed " + }, + { + "bbox": [ + 402, + 152, + 449, + 161 + ], + "type": "inline_equation", + "content": "\\hat{I}_t^{\\prime}" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 450, + 152, + 544, + 161 + ], + "lines": [ + { + "bbox": [ + 450, + 152, + 544, + 161 + ], + "spans": [ + { + "bbox": [ + 450, + 152, + 544, + 161 + ], + "type": "text", + "content": "Original " + }, + { + "bbox": [ + 450, + 152, + 544, + 161 + ], + "type": "inline_equation", + "content": "I_{t}" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 387, + 161, + 465, + 204 + ], + "lines": [ + { + "bbox": [ + 387, + 161, + 465, + 204 + ], + "spans": [ + { + "bbox": [ + 387, + 161, + 465, + 204 + ], + "type": "image", + "image_path": "e299bc1bacbc87bf713afc6bc8fb5f6471e62c983b80528958efad1cd448046d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 466, + 161, + 544, + 204 + ], + "blocks": [ + { + "bbox": [ + 466, + 161, + 544, + 204 + ], + "lines": [ + { + "bbox": [ + 466, + 161, + 544, + 204 + ], + "spans": [ + { + "bbox": [ + 466, + 161, + 544, + 204 + ], + "type": "image", + "image_path": "9eb171ef59a95e0ddd038865f0c3bfdda306bbb72fe049ac0c0decbeba781e05.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 332, + 204, + 525, + 244 + ], + "blocks": [ + { + "bbox": [ + 332, + 204, + 525, + 244 + ], + "lines": [ + { + "bbox": [ + 332, + 204, + 525, + 244 + ], + "spans": [ + { + "bbox": [ + 332, + 204, + 525, + 244 + ], + "type": "image", + "image_path": "f8ab21ad1a716a67a850a8dd5522923cc66dbc62a3331f0635850746b8412ca3.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 246, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 347, + 246, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 347, + 246, + 504, + 255 + ], + "type": "text", + "content": "(b) Illustration of the quality scores estimated by QAM." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 258, + 545, + 314 + ], + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 314 + ], + "type": "text", + "content": "Figure 4. Illustration of the capabilities of our proposed modules. (a) shows that the artifacts are successfully removed by applying our CCM, yielding high-quality images for training. (b) shows the quality scores of different images estimated by QAM, where high-quality images are used for training and inversely rejected." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\hat{I}_t^\\prime" + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "content": " is the reconstructed result of " + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "inline_equation", + "content": "I_t^\\prime" + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "content": " by CCM, i.e., " + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\hat{I}_t^\\prime = \\Theta_c(I_t^\\prime)" + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "content": " is the feature extractor of H-Net. Instead of directly minimizing the content distance between " + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\hat{I}_t^\\prime" + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "content": " we constrain the similarity of the features to prompt CCM to reconstruct " + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\hat{I}_t^\\prime" + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "content": " to be sharp as the real-world image, since the " + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{ts}" + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "content": " is not accurate enough. As shown in Fig. 4(a), the artifacts in " + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "inline_equation", + "content": "I_t^\\prime" + }, + { + "bbox": [ + 304, + 332, + 545, + 417 + ], + "type": "text", + "content": " have been successfully removed by CCM." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "content": "Quality Assessment Module. The quality assessment module (QAM) is designed to filter bad cases that cannot be reconstructed by CCM, further improving the quality of the training data. Specifically, the QAM is designed to be a classification network whose output " + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "content": " ranges from 0 to 1, " + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 418, + 545, + 490 + ], + "type": "text", + "content": " being closer to 1 means better quality and vice versa." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "content": "To achieve this, we sample another homography to interfere with " + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{ts}" + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "content": " since the quality of the generated image is partially relied on the accuracy of " + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{ts}" + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "content": ", producing a disturbing image " + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "content": " with artifacts through G-phase. The score " + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "content": " obtained by feeding " + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "content": " into QAM should be close to 0. In contrast, the original target image " + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "content": " is captured from real-world scenes, so the score " + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "content": " should be close to 1. Thus, the quality assessment loss " + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{qal}" + }, + { + "bbox": [ + 304, + 491, + 545, + 597 + ], + "type": "text", + "content": " is proposed to optimize the QAM as" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 353, + 607, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 607, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 353, + 607, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {q a l} = \\operatorname {B C E} \\left(x _ {t}, 1\\right) + \\operatorname {B C E} \\left(x _ {r}, 0\\right), \\tag {7}", + "image_path": "90d0968292914eeacedbc5c93b5331365c8856a2f15d086d28431b7a7eb29def.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{BCE}(\\cdot)" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": " denotes the binary cross-entropy loss. As a result, when the score " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\hat{x}_t^\\prime" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\hat{I}_t^\\prime" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": " exceeds a certain threshold " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": ", we consider " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\hat{I}_t^\\prime" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": " to be a high-quality image that can be used for training and inversely be rejected, as shown in Fig. 4(b)." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Overall, with the help of CCM and QAM, qualified training pairs can be generated, satisfying both realism criteria and supervision labels to improve the H-Net. Through iter" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9809" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "ative learning, the performance of the network and the quality of the generated dataset are mutually improved." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 103, + 154, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 154, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 154, + 117 + ], + "type": "text", + "content": "3.4. Network Training" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "text", + "content": "Since the training samples have GT labels, we can optimize the H-Net by minimizing the difference between " + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{gt}" + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "text", + "content": " and the estimated " + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{st'}" + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "text", + "content": ". However, as mentioned in [6], it is non-trivial to directly estimate a tomography, we follow previous works [15, 4, 35] to use the 4 corners offset vectors " + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "inline_equation", + "content": "D_{\\mathbf{H}_{st'}}" + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "inline_equation", + "content": "D_{\\mathbf{H}_{gt}}" + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "text", + "content": " as supervision objectives, the corresponding homography could be computed by solving DLT [11] using the offset vectors. In addition, we compute the bidirectional homography as the " + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{t's}" + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "text", + "content": " should also be aligned with " + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{gt}^{-1}" + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "text", + "content": ". The finally supervised loss " + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sup}" + }, + { + "bbox": [ + 47, + 122, + 288, + 244 + ], + "type": "text", + "content": " is expressed as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 262, + 287, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 262, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 64, + 262, + 287, + 279 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s u p} = \\left| D _ {\\mathbf {H} _ {s t ^ {\\prime}}} - D _ {\\mathbf {H} _ {g t}} \\right| _ {1} + \\left| D _ {\\mathbf {H} _ {t ^ {\\prime} s}} - D _ {\\mathbf {H} _ {g t} ^ {- 1}} \\right| _ {1}. \\tag {8}", + "image_path": "aab4a6fd4f7ee5d4106ef2065bd25681181c454fa61028fa849dbbac672fcdd8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 286, + 287, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 287, + 322 + ], + "type": "text", + "content": "The total loss " + }, + { + "bbox": [ + 47, + 286, + 287, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {total }}" + }, + { + "bbox": [ + 47, + 286, + 287, + 322 + ], + "type": "text", + "content": " is expressed by combing the supervised loss, the content consistency loss, and the quality assessment loss as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 332, + 287, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 332, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 94, + 332, + 287, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s u p}} + \\lambda_ {1} \\mathcal {L} _ {\\text {c c l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {q a l}}, \\tag {9}", + "image_path": "dba269dcefe9c2a631d4703ce41536661d952c3a4dbbb5a2bd980bff7ea52d72.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 354, + 253, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 253, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 253, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 354, + 253, + 367 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 47, + 354, + 253, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 354, + 253, + 367 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 47, + 354, + 253, + 367 + ], + "type": "text", + "content": " are empirically set as 0.5 and 0.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 376, + 124, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 124, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 124, + 390 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 396, + 105, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 105, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 105, + 407 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 414, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 414, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 287, + 498 + ], + "type": "text", + "content": "MS-COCO. The MS-COCO dataset [18] is one of the widely used datasets in the fields of object detection [19], segmentation [29], etc. Previously supervised methods [6, 15, 35, 4] used it to prepare their training datasets by perturbing each corner of single images to produce GT homographies and warp the images using the homographies to generate training pairs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "content": "CA-Unsup. The CA-unsup dataset [42] contains 800k training pairs and 4.2k testing pairs that are captured from five types of real-world scenes, i.e., regular (RE), low texture (LT), low light (LL), small foreground (SF), and large foreground (LF), where the last four are challenging scenes for homography estimation. The training set contains only unlabeled image pairs, but for each pair of the test set, " + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "inline_equation", + "content": "6\\sim 10" + }, + { + "bbox": [ + 46, + 498, + 287, + 665 + ], + "type": "text", + "content": " equally distributed matching points are manually marked for evaluation. Based on the CA-unsup dataset, we generated a new dataset that contains the same number of image pairs satisfying both label criteria and realism criteria as the CA-unusp dataset through our proposed method, named CA-sup. We show some examples of our CA-sup dataset in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "GHOF. The GHOF dataset [17] is designed to evaluate homography and optical flow estimation with the gyroscope readings, it contains 10k training data and 300 testing pairs under 5 different categories, including regular (RE), foggy" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "(FOG), low light (LL), rainy (RAIN), and snowy (SNOW) scenes. Compared to the CA-unsup dataset, it contains more parallax motion variations and extreme foreground ratio which are challenging to the homography methods. The sparse " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "5\\sim 8" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " correspondences are used to evaluate the results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 140, + 440, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 140, + 440, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 140, + 440, + 153 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 159, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 159, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 159, + 545, + 278 + ], + "type": "text", + "content": "Our framework consists of D-Net, CCM, QAM, and H-Net. The D-Net is pre-trained on the CA-unsup dataset in an unsupervised manner, the detailed architecture and implementation details are described in the supplementary material. We choose BasesHomo [41], which balances performance and efficiency as the backbone of H-Net, and convert its output representation into 4 corners offset vectors. The CCM and QAM are training with the H-Net for 2 iterations, and the training parameters are consistent with the official implementation of BasesHomo for each iteration." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 288, + 494, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 494, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 494, + 300 + ], + "type": "text", + "content": "4.3. Comparison with Existing Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 306, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 545, + 461 + ], + "type": "text", + "content": "Comparison Methods. We compare our method with three categories of existing homography estimation methods: 1) feature-based methods including SIFT [26], ORB [33], SuperPoint [7] with SuperGlue [34] (SPSG), and LoFTR [36], 2) supervised methods including DHN [6], LocalTrans [35], and IHN [4], 3) unsupervised methods including CAHomo [42], BasesHomo [41], and HomogAN [12]. For feature-based methods, we improve them with two different outlier rejection algorithms RANSAC [9] and MAGSAC [2], respectively. For supervised methods, we employ their models pre-trained on the MS-COCO dataset [18] and the models retrained on our CA-sup dataset denoted with " + }, + { + "bbox": [ + 305, + 306, + 545, + 461 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 305, + 306, + 545, + 461 + ], + "type": "text", + "content": " for evaluation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "text", + "content": "Quantitative Comparison. We report the quantitative results of all comparison methods on the CA-unsup test set in Table 1. The " + }, + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{3\\times 3}" + }, + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "text", + "content": " in the first row refers to a " + }, + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "text", + "content": " identity matrix as a \"no-warping\" homography for reference, of which the errors reflect the original distance between point pairs. As shown in Table 1, our method achieves state-of-the-art performance in all categories and outperforms the best existing method HomoGAN by " + }, + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "inline_equation", + "content": "12.82\\%" + }, + { + "bbox": [ + 304, + 463, + 546, + 713 + ], + "type": "text", + "content": ", with the points matching error (PME) reduced from 0.39 to 0.34. The feature-based methods can perform well in regular (RE) scenes, as sufficient matching points can be obtained, while our methods and HomoGAN reduce the error by 0.08 compared to SIFT+RANSAC. More specifically, our method produces a PME of 0.216 which is lower than 0.222 of HomoGAN. The small foreground (SF) and large foreground (LF) scenes contain dynamic objects that affect the accuracy of estimated homography, and our method has the lowest PME in these two categories, even when compared to the methods utilizing outlier removal masks for robust estimation, i.e., CAHomo and HomoGAN. In low texture (LT) and low light (LL) scenes, most learning-based" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9810" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 541, + 270 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 541, + 270 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 541, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 541, + 270 + ], + "type": "table", + "html": "
1)MethodsAVGRELTLLSFLF
2)\\( \\mathcal{I}_{3\\times 3} \\)6.70 (+1617.95%)7.75 (+3422.73%)7.65 (+1765.85%)7.21 (+1164.91%)7.53 (+1611.36%)3.39 (+993.55%)
3)SIFT [26]+RANSAC [9]1.41 (+261.54%)0.30 (+36.36%)1.34 (+226.83%)4.03 (+607.02%)0.81 (+84.09%)0.57 (+83.87%)
4)SIFT [26]+MAGSAC [2]1.34 (+243.59%)0.31 (+40.91%)1.72 (+319.51%)3.39 (+494.74%)0.80 (+81.82%)0.47 (51.61%)
5)ORB [33]+RANSAC [9]1.48 (+279.49%)0.85 (+286.36%)2.59 (+531.71%)1.67 (+192.98%)1.10 (+150.00%)1.24 (+300.00%)
6)ORB [33]+MAGSAC [2]1.69 (+333.33%)0.97 (+340.91%)3.34 (+714.63%)1.58 (+177.19%)1.15 (+161.36%)1.40 (+351.61%)
7)SPSG [7, 34]+RANSAC [9]0.71 (+82.05%)0.41 (+86.36%)0.87 (+112.20%)0.72 (+26.32%)0.80 (+81.81%)0.75 (+141.94%)
8)SPSG [7, 34]+MAGSAC [2]0.63 (+61.54%)0.36 (+63.64%)0.79 (+92.68%)0.70 (+22.81%)0.71 (+61.36%)0.70 (+125.81%)
9)LoFTR [36]+RANSAC [9]1.44 (+269.23%)0.56 (+154.55%)2.70 (+558.54%)1.36 (+138.60%)1.05 (+138.64%)1.52 (+390.32%)
10)LoFTR [36]+MAGSAC [2]1.39 (+256.41%)0.55 (+150.00%)2.57 (+526.83%)1.33 (+133.33%)1.05 (+138.64%)1.41 (+354.84%)
11)CAHomo [42]0.88 (+125.64%)0.73 (+231.82%)1.01 (+146.34%)1.03 (+80.70%)0.92 (+109.09%)0.70 (+125.81%)
12)BasesHomo [41]0.50 (+28.21%)0.29 (+31.82%)0.54 (+31.71%)0.65 (+14.04%)0.61 (+38.64%)0.41 (+32.26%)
13)HomoGAN [12]0.39 (+0.00%)0.22 (+0.00%)0.41 (+0.00%)0.57 (+0.00%)0.44 (+0.00%)0.31 (+0.00%)
14)DHN [6]2.87 (+635.90%)1.51 (+586.36%)4.48 (+992.68%)2.76 (+384.21%)2.62 (+495.45%)3.00 (+867.74%)
15)LocalTrans [35]4.21 (+978.26%)4.09 (+1757.59%)4.84 (+1081.14%)4.55 (+697.98%)5.30 (+1105.20%)2.25 (+624.29%)
16)IHN [4]4.67 (+1097.44%)4.85 (+2104.55%)5.54 (+1251.22%)5.10 (+794.74%)5.04 (+1045.45%)2.84 (+816.13%)
17)DHN* [6]1.89 (+384.95%)1.21 (+451.82%)2.13 (+418.54%)2.33 (+307.91%)1.72 (+291.23%)2.07 (+567.81%)
18)LocalTrans* [35]1.76 (+352.03%)1.25 (+467.41%)2.15 (+423.85%)1.90 (+232.61%)2.25 (+410.41%)1.28 (+311.77%)
19)IHN* [4]1.19 (+205.54%)0.72 (+227.09%)1.74 (+324.93%)1.18 (+107.42%)1.30 (+196.52%)1.01 (+225.58%)
20)Ours0.34 (-12.82%)0.22 (+0.00%)0.35 (-14.63%)0.44 (-22.81%)0.42 (-4.55%)0.29 (-6.45%)
", + "image_path": "8bd83622134a04ae22ebc7535ddd73df17a90b16ee17b7fabd9e0d1fdcced926.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 328, + 285, + 472 + ], + "blocks": [ + { + "bbox": [ + 46, + 277, + 546, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 546, + 312 + ], + "type": "text", + "content": "Table 1. The point matching errors (PME) of our method and all comparison methods on the CA-sup [42] test set. The best and second-best results are highlighted in bold and underlined. The percentages in the parentheses indicate the relative change in comparison to the second-best results. SPSG indicates SuperPoint with SuperGlue, and * denotes the methods are retrained on our CA-sup dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 328, + 285, + 472 + ], + "lines": [ + { + "bbox": [ + 52, + 328, + 285, + 472 + ], + "spans": [ + { + "bbox": [ + 52, + 328, + 285, + 472 + ], + "type": "table", + "html": "
1)MethodsAVGREFOGLLRAINSNOW
2)I3×36.334.947.248.095.485.89
3)SIFT [26]+MAGSAC [2]3.900.644.019.770.704.40
4)ORB [33]+RANSAC [9]15.146.9231.2727.801.827.90
5)SPSG [7, 34]+MAGSAC [2]3.283.411.467.610.753.19
6)LoFTR [36]+MAGSAC [2]2.552.761.165.340.522.98
7)CAHomo [42]3.874.103.846.991.273.17
8)BasesHomo [41]2.282.021.434.900.782.29
9)HomoGAN [12]1.951.730.603.950.473.02
10)DHN [6]6.616.046.027.686.996.32
11)LocalTrans [35]5.724.066.495.955.786.34
12)IHN [4]8.177.108.719.346.579.13
13)DHN* [6]3.011.923.944.541.982.66
14)LocalTrans* [35]2.891.784.274.591.372.43
15)IHN* [4]2.592.213.054.700.982.03
16)Ours1.721.600.884.420.431.28
", + "image_path": "f8e5f6c10b836410120d14082f51614126097547b28ab769abce8a4fccadd704.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 479, + 287, + 514 + ], + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 514 + ], + "type": "text", + "content": "Table 2. The point matching errors (PME) of our method and comparison methods on the GHOF [17] test set. The best and second-best results are highlighted in bold and underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": "methods are more robust due to their keypoint-free estimation strategy, especially unsupervised ones. However, homogeneous region [38] dominates a large portion of the image in the LT and LL scenes, where the photometric loss between the two images is minor, so the performance of unsupervised methods in these two categories is not as well as in others. In contrast, our method outperforms the unsupervised methods in the LT and LL with errors reduced by at least " + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "inline_equation", + "content": "14.63\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "inline_equation", + "content": "22.81\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 653 + ], + "type": "text", + "content": " respectively, thanks to our photometric-free fully supervised learning strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "Moreover, the generalization ability of supervised methods has been widely criticized. Therefore, we further evaluate feature-based methods that performed well on the CA-unsup dataset, learning-based methods, and our method on the GHOF test set to prove the effectiveness of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 331, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 497 + ], + "type": "text", + "content": "As shown in Table 2, our method achieves the best results in RAIN and SNOW scenes and the second-best results in RE, FOG and LL scenes, resulting in the lowest average PME. For regular scenes, even though feature-based methods can extract sufficient high-quality feature matches, we still outperform most of them. In addition, our CA-sup dataset does not contain images captured in the FOG and RAIN scenes, while our method is capable of generalizing these scenes, which proves the superiority of our method. On the other hand, previous supervised methods cannot generalize well to real-world scenes since their training data lack realistic motion. After retraining on our CA-sup dataset, their performance has been significantly improved, which further proves the effectiveness of our framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 714 + ], + "type": "text", + "content": "Qualitative Comparison. In Fig. 5 and Fig. 6, we provide the qualitative results of our method and competitive methods on the CA-unsup and GHOF test sets, respectively. Fig. 5(a) and (b) are from LF and SF scenes with dynamic objects, and our method produces more accurate results than feature-based and unsupervised methods which are sensitive to moving objects, as highlighted in the red and green boxes. Fig. 5(c) is from the LT scene, where feature-based methods all fail due to the lack of sufficient correspondence. On the contrary, learning-based methods perform well, but they are still not competitive as ours. In Fig. 6, we provide the comparison results in RAIN and FOG scenes of the GHOF test set, where our training set does not contain images captured in such scenes. As mentioned above, while unsupervised methods have better generalization ability in unseen scenes, we prove that supervised methods can also achieve satisfactory results after training on our generated dataset. For example, in Fig. 6(a), the pre" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9811" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 537, + 350 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 537, + 350 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 537, + 350 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 537, + 350 + ], + "type": "image", + "image_path": "f673f975c65a8c38f15579105572119d4d41f382c701d8674efc339dc9377038.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 352, + 547, + 386 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 547, + 386 + ], + "type": "text", + "content": "Figure 5. Qualitative results of our method and other competitive methods on the CA-unsup [42] test set. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, and the green boxes show the content difference between the two images in the error-prone regions. Best viewed by zooming in." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 402, + 284, + 589 + ], + "blocks": [ + { + "bbox": [ + 50, + 402, + 284, + 589 + ], + "lines": [ + { + "bbox": [ + 50, + 402, + 284, + 589 + ], + "spans": [ + { + "bbox": [ + 50, + 402, + 284, + 589 + ], + "type": "image", + "image_path": "68fa1c046d9ac77b6e87c1d6cf62c962a876c9b4e420c2bb9588b998bd834e53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 591, + 287, + 615 + ], + "lines": [ + { + "bbox": [ + 47, + 591, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 287, + 615 + ], + "type": "text", + "content": "Figure 6. Qualitative results of our method and other competitive methods on the GHOF [17] test set. Best viewed by zooming in." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "content": "vious supervised methods can even obtain more accurate results than CAHomo, and our method performs the best." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Comparison with Dataset Generation Methods. We use images from the CA unsup dataset to generate a new dataset named CA-sup" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " using the previous dataset generation strategy [6], and retrain H-Net on it for comparison." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 403, + 541, + 449 + ], + "blocks": [ + { + "bbox": [ + 313, + 403, + 541, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 541, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 541, + 449 + ], + "type": "table", + "html": "
DatasetAVGRELTLLSFLF
DHN [6]CA-sup†1.370.791.701.281.231.83
OursCA-sup0.340.220.350.440.420.29
", + "image_path": "35d36683c4644e3e47e51e52b95e9572185fc21ed80f9c0f98eb8de40296e547.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 456, + 546, + 490 + ], + "lines": [ + { + "bbox": [ + 305, + 456, + 546, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 546, + 490 + ], + "type": "text", + "content": "Table 3. Comparison with previous dataset generation method. We use the same source images to generate the training dataset and train the same network for comparison." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "content": "Quantitative results on the CA-unsup test set are illustrated in Table 3. The previous strategy generates image pairs from single images that meet label criteria only, causing unsatisfactory performance. In contrast, the image pairs contained in our CA-sup dataset can help the network to learn real-world motion representations, proving the importance of realistic scene motion for estimating homography." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Robustness Evaluation. Furthermore, we evaluate the robustness of all comparison methods on the CA-unsup test set by setting thresholds to calculate the proportion of inlier predictions. As shown in Fig. 7, we plot a series of curves where axis " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "X \\in [0.1, 3.0]" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is the threshold and axis " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "Y \\in [0.0, 1.0]" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is the percentage of inliers, higher position of curves represents better robustness. With a threshold of 0.1, our outlier percentage is " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " higher than the second-best method (15.8% vs. " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "), and our method achieves a proportion of " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "99.9\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " with a threshold of 3.0." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9812" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 289, + 180 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 180 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 180 + ], + "type": "image", + "image_path": "2ed1caa4e02b272300721236eab804d93b5440ac7fcbed1b2dc33da15001c42c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 184, + 288, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 288, + 218 + ], + "type": "text", + "content": "Figure 7. The proportion of inliers of our method and all comparison methods under various thresholds. The higher position of curves represents better robustness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 237, + 141, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 141, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 141, + 251 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 256, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 340 + ], + "type": "text", + "content": "In this section, we conduct a series of ablation studies to measure the impact of different component choices of our method. We use the CA-sup dataset for training, and quantitative results on the CA-unsup test set are illustrated in Table 4, detailed experiment settings are discussed below. For more ablation experiments, please refer to the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 341, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 472 + ], + "type": "text", + "content": "Dataset Generation Strategy. We conduct an experiment to evaluate the effectiveness of our dataset generation strategy by using G-phase only to generate training samples to train the network, denoted as Baseline in row 1 of Table 4. Even without using CCM and QAM, our method achieves state-of-the-art performance compared to the existing methods. Moreover, the homography estimation network, i.e., BasesHomo, trained on our generated dataset achieves a " + }, + { + "bbox": [ + 46, + 341, + 287, + 472 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 46, + 341, + 287, + 472 + ], + "type": "text", + "content": " performance gain compared to the network trained with the unsupervised learning strategy, further proving the superiority of our dataset generation strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": "Module Effectiveness. We conduct experiments to evaluate the effectiveness of our proposed CCM and QAM by removing them separately from the overall framework. As shown in rows 2-3 of Table 4, CCM and QAM can help to generate and select high-quality image pairs for training, achieving error reduction of 0.03 and 0.01, respectively, compared to the Baseline in row 1 using none of them. By using both CCM and QAM in the training phase, our method can achieve the best performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "type": "text", + "content": "Different Homography Estimation Networks. As described in Sec. 4.2, we choose BasesHomo as the backbone of our H-Net. In rows 4-5 of Table 4, we further switch the backbone to CAHomo and HomoGAN, denoted as Real-CA and Real-GAN, respectively, to prove that our method also works not only on a specific architecture. Compared with the unsupervised results in rows 11 and 13 of Table 1, CAHomo obtains performance gain by " + }, + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 581, + 287, + 700 + ], + "type": "text", + "content": " with our framework and HomoGAN reduces the error by 0.09, proving the generalization of our proposed framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Iteration Times. The quality of generated dataset and" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 544, + 189 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 544, + 189 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 544, + 189 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 544, + 189 + ], + "type": "table", + "html": "
AVGRELTLLSFLF
1)Baseline0.380.230.400.500.450.33
2)w/o CCM0.370.230.360.480.440.34
3)w/o QAM0.350.230.360.450.420.30
4)Real-CA [42]0.440.270.450.520.530.43
5)Real-GAN [12]0.300.180.300.390.370.28
6)1 iteration0.380.240.400.460.480.33
7)3 iterations0.330.210.340.440.400.28
8)Default0.340.220.350.440.420.29
", + "image_path": "9bdb8d27a5d21ff2045ffb3c0360890076d980abf5bd5127018bb53101fe6190.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 197, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 305, + 197, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 197, + 545, + 217 + ], + "type": "text", + "content": "Table 4. Results of ablation studies, please refer to the text for more details. w/o denotes without." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 229, + 545, + 270 + ], + "blocks": [ + { + "bbox": [ + 307, + 229, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 270 + ], + "type": "image", + "image_path": "03ce53b99f9cacb75f67a218b60825cd6c73aa5278daa5ee43a042522901fe10.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 271, + 545, + 283 + ], + "lines": [ + { + "bbox": [ + 306, + 271, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 306, + 271, + 545, + 283 + ], + "type": "text", + "content": "Figure 8. Illustration of the generated image in different iterations." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 305, + 546, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 546, + 485 + ], + "type": "text", + "content": "network performance are mutually reinforcing through iterative learning. As shown in rows 6-8 of Table 4, the performance of H-Net improves with iteration increases, while higher-precision H-Net generates more natural images, as shown in Fig. 8, and high-quality images can improve performance. When the homography estimated by H-Net is accurate enough, there would be no artifacts and boundary discontinuities in the generated images. Notably, \"0 iteration\" indicates that the image is generated by using the homography estimated by unsupervised BasesHomo. However, a certain upper limit exists in our method and it will converge after several iterations. The performance gain of 3 iterations compared to 2 iterations, i.e., default, is not significant, and more iterations are not cost-effective compared to the resources spent per iteration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 499, + 379, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 499, + 379, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 499, + 379, + 511 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 676 + ], + "type": "text", + "content": "We have presented an iterative deep framework that aims to generate a realistic dataset for supervised homography learning and obtain a high-precision network. We use the dominant plane masks estimated by D-Net and the homography estimated by H-Net to generate training pairs, and the CCM and QAM are proposed to prepare high-quality training data to update the H-Net. The trained network and the generated data can be improved iteratively, yielding a qualified dataset as well as a state-of-the-art network. Extensive experiments demonstrate the superiority of our method and the effectiveness of our newly proposed components, the performance of existing methods can be also improved using our dataset for training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by Sichuan Science and Technology Program of China under grant No.2023NSFSC0462." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9813" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Simon Baker and Iain Matthews. Lucas-kanade 20 years on: A unifying framework. International Journal of Computer Vision, 56(3):221-255, 2004. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "text", + "content": "[2] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proc. CVPR, pages 10197-10205, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 286, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 286, + 190 + ], + "type": "text", + "content": "[3] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In Proc. ECCV, pages 404-417, 2006. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 286, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 286, + 223 + ], + "type": "text", + "content": "[4] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proc. CVPR, pages 1879-1888, 2022. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 286, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 286, + 257 + ], + "type": "text", + "content": "[5] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. Clkn: Cascaded lucas-kanade networks for image alignment. In Proc. CVPR, pages 2213–2221, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 286, + 290 + ], + "type": "text", + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 292, + 286, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 286, + 334 + ], + "type": "text", + "content": "[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPRW, pages 224-236, 2018. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 335, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 379 + ], + "type": "text", + "content": "[8] Tianjiao Ding, Yunchen Yang, Zhihui Zhu, Daniel P Robinson, René Vidal, Laurent Kneip, and Manolis C Tsakiris. Robust homography estimation via dual principal component pursuit. In Proc. CVPR, pages 6080-6089, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 423 + ], + "type": "text", + "content": "[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 425, + 286, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 286, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 286, + 468 + ], + "type": "text", + "content": "[10] Yunhui Han, Kunming Luo, Ao Luo, Jiangyu Liu, Haoqiang Fan, Guiming Luo, and Shuaicheng Liu. Realflow: Embased realistic optical flow dataset generation from videos. In Proc. ECCV, pages 288-305, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 469, + 286, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 286, + 501 + ], + "type": "text", + "content": "[11] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 502, + 286, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 286, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 286, + 545 + ], + "type": "text", + "content": "[12] Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, Qijun Zhao, and Shuaicheng Liu. Unsupervised homography estimation with coplanarity-aware gan. In Proc. CVPR, pages 17663-17672, 2022. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 286, + 590 + ], + "type": "text", + "content": "[13] Hai Jiang, Haipeng Li, Yuhang Lu, Songchen Han, and Shuaicheng Liu. Semi-supervised deep large-baseline homography estimation with progressive equivalence constraint. In Proc. AAAI, pages 1024-1032, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 591, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 286, + 644 + ], + "type": "text", + "content": "[14] Dewi Endah Kharismawati, Hadi Ali Akbarpour, Rumana Aktar, Filiz Bunyak, Kannappan Palaniappan, and Toni Kazic. Cornet: Unsupervised deep homography estimation for agricultural aerial imagery. In Proc. ECCV, pages 400-417, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 679 + ], + "type": "text", + "content": "[15] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proc. CVPR, pages 7652-7661, 2020. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "text", + "content": "[16] Haipeng Li, Kunming Luo, and Shuaicheng Liu. Gyroflow: Gyroscope-guided unsupervised optical flow learning. In Proc. ICCV, pages 12869-12878, 2021. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[17] Haipeng Li, Kunming Luo, Bing Zeng, and Shuaicheng Liu. Gyroflow+: Gyroscope-guided unsupervised deep homography and optical flow learning. arXiv preprint arXiv:2301.10018, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755, 2014. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[19] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International Journal of Computer Vision, 128(2):261-318, 2020. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "text", + "content": "[20] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(06):7885-7899, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 264, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 318 + ], + "type": "text", + "content": "[21] Shuaicheng Liu, Kunming Luo, Ao Luo, Chuan Wang, Fanman Meng, and Bing Zeng. Asflow: Unsupervised optical flow learning with adaptive pyramid sampling. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4282-4295, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "type": "text", + "content": "[22] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(03):2849-2863, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 376, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 408 + ], + "type": "text", + "content": "[23] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM Transactions on Graphics, 32(4):1-10, 2013. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 410, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 464 + ], + "type": "text", + "content": "[24] Zhen Liu, Wenjie Lin, Xinpeng Li, Qing Rao, Ting Jiang, Mingyan Han, Haoqiang Fan, Jian Sun, and Shuaicheng Liu. Adnet: Attention-guided deformable convolutional network for high dynamic range imaging. In Proc. CVPRW, pages 463-470, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 466, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 498 + ], + "type": "text", + "content": "[25] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proc. ECCV, pages 344-360, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "text", + "content": "[26] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 1, 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "text", + "content": "[27] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In Proc. CVPR, pages 8906-8915, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "type": "text", + "content": "[28] Ao Luo, Fan Yang, Kunming Luo, Xin Li, Haoqiang Fan, and Shuaicheng Liu. Learning optical flow with adaptive graph reasoning. In Proc. AAAI, pages 1890-1898, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "type": "text", + "content": "[29] Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3523-3542, 2022. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[30] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters, 3(3):2346-2353, 2018. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9814" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 600 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Lang Nie, Chunyu Lin, Kang Liao, Meiqin Liu, and Yao Zhao. A view-free image stitching network based on global homography. Journal of Visual Communication and Image Representation, 73:102950, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "text", + "content": "[32] Lang Nie, Chunyu Lin, Kang Liao, and Yao Zhao. Learning edge-preserved image stitching from multi-scale deep homography. Neurocomputing, 491:533-543, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "text", + "content": "[33] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In Proc. ICCV, pages 2564-2571, 2011. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "type": "text", + "content": "[34] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. CVPR, pages 4938–4947, 2020. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 230, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 230, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 286, + 274 + ], + "type": "text", + "content": "[35] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. Localtrans: A multiscale local transformer network for cross-resolution homography estimation. In Proc. ICCV, pages 14890-14899, 2021. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 275, + 286, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 286, + 318 + ], + "type": "text", + "content": "[36] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proc. CVPR, pages 8922-8931, 2021. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 319, + 286, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 286, + 363 + ], + "type": "text", + "content": "[37] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proc. CVPR, pages 11016-11025, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 365, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 286, + 407 + ], + "type": "text", + "content": "[38] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proc. CVPR, pages 6258-6268, 2020. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 410, + 286, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 410, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 410, + 286, + 442 + ], + "type": "text", + "content": "[39] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proc. CVPR, pages 5714-5724, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 444, + 286, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 444, + 286, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 444, + 286, + 476 + ], + "type": "text", + "content": "[40] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proc. ECCV, pages 117–132, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 478, + 286, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 478, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 478, + 286, + 521 + ], + "type": "text", + "content": "[41] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proc. ICCV, pages 13117-13125, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 522, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 522, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 49, + 522, + 286, + 566 + ], + "type": "text", + "content": "[42] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proc. ECCV, pages 653-669, 2020. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 567, + 286, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 286, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 286, + 600 + ], + "type": "text", + "content": "[43] Zhuofan Zhang, Zhen Liu, Bing Zeng, and Shuaicheng Liu. Minimum latency deep online video stabilization. arXiv preprint arXiv:2212.02073, 2022. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9815" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Surface Extraction from Neural Unsigned Distance Fields/a0829e7a-732d-4761-bffa-ad35624b76ef_content_list.json b/2023/Surface Extraction from Neural Unsigned Distance Fields/a0829e7a-732d-4761-bffa-ad35624b76ef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e2003cfb11d34ee5602e0a3cbf3c94b5cb004d6d --- /dev/null +++ b/2023/Surface Extraction from Neural Unsigned Distance Fields/a0829e7a-732d-4761-bffa-ad35624b76ef_content_list.json @@ -0,0 +1,1489 @@ +[ + { + "type": "text", + "text": "Surface Extraction from Neural Unsigned Distance Fields", + "text_level": 1, + "bbox": [ + 192, + 130, + 777, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Congyi Zhang $^{1,2*}$ Guying Lin $^{1*}$ Lei Yang $^{1,2}$ Xin Li $^{3}$ Taku Komura $^{1}$ Scott Schaefer $^{3}$ John Keyser $^{3}$ Wenping Wang $^{3\\dagger}$", + "bbox": [ + 133, + 178, + 834, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong 2TransGP \n3Texas A&M University", + "bbox": [ + 299, + 239, + 666, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 311, + 313, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a method, named DualMesh-UDF, to extract a surface from unsigned distance functions (UDFs), encoded by neural networks, or neural UDFs. Neural UDFs are becoming increasingly popular for surface representation because of their versatility in presenting surfaces with arbitrary topologies, as opposed to the signed distance function that is limited to representing a closed surface. However, the applications of neural UDFs are hindered by the notorious difficulty in extracting the target surfaces they represent. Recent methods for surface extraction from a neural UDF suffer from significant geometric errors or topological artifacts due to two main difficulties: (1) A UDF does not exhibit sign changes; and (2) A neural UDF typically has substantial approximation errors.", + "bbox": [ + 73, + 344, + 472, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DualMesh-UDF addresses these two difficulties. Specifically, given a neural UDF encoding a target surface $\\bar{S}$ to be recovered, we first estimate the tangent planes of $\\bar{S}$ at a set of sample points close to $\\bar{S}$ . Next, we organize these sample points into local clusters, and for each local cluster, solve a linear least squares problem to determine a final surface point. These surface points are then connected to create the output mesh surface, which approximates the target surface. The robust estimation of the tangent planes of the target surface and the subsequent minimization problem constitute our core strategy, which contributes to the favorable performance of DualMesh-UDF over other competing methods. To efficiently implement this strategy, we employ an adaptive Octree. Within this framework, we estimate the location of a surface point in each of the octree cells identified as containing part of the target surface. Extensive experiments show that our method outperforms existing methods in terms of surface reconstruction quality while maintaining comparable computational efficiency.", + "bbox": [ + 75, + 556, + 472, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 311, + 630, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implicit surfaces are widely used for surface representation in computer vision and computer graphics. An implicit surface is usually defined as a level set of a function, such as the zero-level set of a signed distance function (SDF). Extracting a mesh representation of an implicit surface from its defining equation is therefore a critical task for surface visualization and processing. Recent advances in machine learning have given rise to a new kind of implicit surface, called a neural implicit surface. A neural implicit surface is a level-set of a function encoded by an MLP (multilayer perceptron) and has the advantage of compactness and inherent smoothness thanks to its MLP representation. SDFs or occupancy fields are widely used in these implicit representations [17, 15, 16, 4, 18, 20, 6, 1, 8, 14, 19].", + "bbox": [ + 496, + 338, + 893, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, neural implicit surfaces based on the SDF or occupancy fields require inside/outside labeling and thus can only represent orientable and closed surfaces. Hence, as an extension, unsigned distance functions (UDFs) have been used to represent surfaces of arbitrary topologies, including open surfaces with boundaries or non-orientable surfaces (e.g. the Möbius strip). Despite its versatility, applications of a UDF-based surface representation are severely hindered by the difficulty in extracting the target surface it represents, as shown in [5] and [7].", + "bbox": [ + 496, + 550, + 893, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Problem formulation: Suppose that a surface $\\bar{S}$ , called the target surface, is defined as the zero-level set of its unsigned distance function (UDF) $\\bar{F}(\\mathbf{p})$ . Then suppose that this UDF $\\bar{F}(\\mathbf{p})$ is approximated by a neural network with the resulting network-encoded UDF being referred to as the neural UDF, denoted by $F(\\mathbf{p})$ . Given a neural UDF $F(\\mathbf{p})$ , the surface extraction problem is to robustly extract a surface $\\mathcal{S}$ from $F(\\mathbf{p})$ such that $\\mathcal{S}$ well approximates the target surface $\\bar{S}$ .", + "bbox": [ + 496, + 702, + 893, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Challenges: The difficulty in surface extraction from a neural UDF arises from two aspects: (1) A UDF does not have zero-crossings (or sign changes) across the surface it represents. As a result, traditional mesh extraction methods", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 875, + 202, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22531", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that rely on zero-crossings (e.g. Marching Cubes [13, 11], Dual Contouring [10], and their variants) are not applicable to UDFs. (2) The MLP representation of a neural UDF tends to have significant approximation errors around the target surface (see detailed error characteristics of neural UDFs in Sec. 3). This makes it even more challenging to extract a high-quality approximation of the target surface.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several methods, MeshUDF [7], CAP-UDF [21] and Neural Dual Contouring (NDC) [3], have recently been proposed for extracting a mesh surface from a UDF. MeshUDF and CAP-UDF attempt to infer the gradients of a UDF on the grids and determine the sign changes of the estimated gradients, invoking the Marching Cubes method for mesh extraction. When applied to a neural UDF, the sign-change inference step of this method suffers from instability due to the non-negligible error introduced by the approximate MLP representation near the surface where the gradients of the ideal UDF are undefined. As a result, the extracted meshes are less accurate and often have topological errors (e.g. holes). The NDC method proposes a data-driven Dual Contouring approach to predict the position of mesh vertices and dual faces directly from the UDF data. When applied to a neural UDF, this method often produces meshes with considerable artifacts such as holes, zig-zags, etc.", + "bbox": [ + 75, + 198, + 470, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We develop a new strategy, consisting of novel sampling and efficient optimization techniques to address the difficulties in surface extraction from the neural UDF. Suppose the input is a neural UDF $F(\\mathbf{p})$ encoding the target surface $\\bar{S}$ to be recovered. Our strategy has two key steps: (1) computing approximate tangent planes of the target surface; and (2) local minimization for generating final surface points.", + "bbox": [ + 76, + 457, + 468, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Step (1), we first generate sample points $\\mathbf{p}_i$ around, but not too close to, the target surface $\\bar{S}$ , because the UDF values and gradients at locations too close to $\\bar{S}$ are relatively unreliable. Thus, $\\mathbf{p}_i$ are called off-surface sample points. For each $\\mathbf{p}_i$ , we use the UDF value $F(\\mathbf{p}_i)$ and its gradient $\\nabla F(\\mathbf{p}_i)$ to project $\\mathbf{p}_i$ towards the target surface to obtain point $\\mathbf{q}_i = \\mathbf{p}_i - F(\\mathbf{p}_i)\\mathbf{n}_i(\\mathbf{p}_i)$ where $\\mathbf{n}(\\mathbf{p}_i) = \\nabla F(\\mathbf{p}_i) / \\| \\nabla F(\\mathbf{p}_i)\\|$ [5]. These points $\\mathbf{q}_i$ are called projection points. Although the points $\\mathbf{q}_i$ are very close to the target surface $\\bar{S}$ , as we will show later, the noisy error in the neural UDF makes these points a poor approximation to the target surface.", + "bbox": [ + 75, + 564, + 468, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further improve surface accuracy, for each projection point $\\mathbf{q}_i$ we generate an estimated tangent plane $T_{i}$ of $\\bar{S}$ such that $T_{i}$ passes through $\\mathbf{q}_i$ and has the unit normal vector $\\mathbf{n}_i$ . Note that the normal vector $\\mathbf{n}_i$ of $T_{i}$ is set to be $\\mathbf{n}(\\mathbf{p}_i)$ rather than $\\mathbf{n}(\\mathbf{q}_i)$ since the former is a more reliable estimation. This is because the initial sample point $\\mathbf{p}_i$ is not too close to $\\bar{S}$ , so the gradient $\\nabla F(\\mathbf{p}_i)$ is less contaminated by the pronounced errors of the neural UDF close to $\\bar{S}$ .", + "bbox": [ + 75, + 747, + 468, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Step (2), the estimated tangent planes are organized into clusters, which may overlap. For each cluster of tan-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gent planes $T_{i}$ , we solve a linear least squares problem to produce a final surface point $\\mathbf{s}_i$ that minimizes the sum of its squared distances to the tangent planes $T_{i}$ . This minimization step based on tangent planes not only provides an accurate surface point but also allows us to faithfully reconstruct the sharp edges of the target surface. Finally, all the surface points $\\mathbf{s}_i$ from all the clusters are connected to form the output mesh surface to approximate the target surface $\\bar{S}$ .", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To efficiently implement the above strategy, our DualMesh-UDF method adopts an adaptive Octree structure to partition the space containing the target surface to regular cells. We developed efficient procedures to determine those cells that contain part of the target surface and perform the sampling and minimization procedures in each occupied cell. To connect the surface points to create the output mesh, we follow the Dual Contouring approach, connecting surface points residing in adjacent grid cells to create polygons dual to octree edges.", + "bbox": [ + 496, + 212, + 890, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that our DualMesh-UDF significantly outperforms existing methods in terms of surface reconstruction accuracy and sharp feature preservation.", + "bbox": [ + 496, + 363, + 890, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contribution of this work is a new algorithm to robustly and accurately extract a surface from a neural UDF. To overcome the inevitable approximation errors near the target surface and cut locus, we obtain robust estimation of surface tangent planes by leveraging off-surface sample points, use least square minimization to better predict the surface points, and achieve high-quality surface extraction results with sharp features better preserved compared to the state of the art. The code is available at https://github.com/cong-yi/DualMesh-UDF.", + "bbox": [ + 496, + 422, + 890, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 588, + 640, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Marching Cubes method [13] and its variants have been established as the de facto standard of converting distance fields to boundary mesh representations. Using the gradient information of the distance function, Extended Marching Cubes [11] and Dual Contouring [10] are both capable of producing meshes with faithfully preserved sharp features (e.g., corners and edges).", + "bbox": [ + 496, + 613, + 890, + 719 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, all of these methods require inside/outside labeling on the sampling grid, which is either a regular grid of cubes or an adaptive octree grid of cells, to determine whether any zero-level set surface of the signed distance field crosses a particular grid cell. This requirement for sign changes limits the application of these methods to only SDFs or its variants that have sign changes across the underlying surface. Hence, they cannot be directly used to extract a mesh surface from a UDF, which is now often used to represent arbitrary surfaces such as open surfaces or non-orientable surfaces [5].", + "bbox": [ + 496, + 719, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method is similar to [10] only in the sense that we", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22532", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "follow [10] to solve a quadratic error function (QEF) to estimate a surface point per grid cell. But unlike [10], where sign changes are available and the gradients at the zero-crossings are reliable, unsigned distance fields do not have sign changes and are non-differentiable at the zero level set. Additionally, we will show that the information around the zero-level set of a neural UDF is unreliable. Due to these two reasons, we alternatively make use of the spatial sample points that are off the target surface in the neural UDF to find reliably estimated tangent planes and formulate a QEF for estimating the final surface points.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Three relevant methods, MeshUDF [7], CAP-UDF [21] and Neural Dual Contouring (NDC) [3], have recently been proposed to extract meshes from UDFs. MeshUDF [7] and CAP-UDF [21] use the gradient information of the UDF to assign different signs (+ or -) to grid points on different sides of the underlying surface, thus invoking Marching Cubes to extract the surface according to the sign labels. However, the quality of the estimated sign labels can be significantly affected by the accuracy of the neural UDF. As we will show later, the neural UDF becomes less accurate around the target surface it represents, contaminating the inferred sign labels and explaining the poor performance of these two methods. Their use of Marching Cubes is incapable of preserving sharp corners or edges, while the use of a regular grid incurs high memory overhead when a high grid resolution is needed to resolve shape details. In contrast, our DualMesh-UDF method preserves sharp features and uses an adaptive octree grid to reduce computational expense even with a high grid resolution. Furthermore, estimating the surface point location is done by solving a least square problem (QEF) which is more robust than the gradient-based sign-labeling strategy used in MeshUDF in terms of the topology of the resulting meshes.", + "bbox": [ + 75, + 257, + 472, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The NDC method [3] uses a data-driven approach to train a neural network that predicts vertex position per regular grid cell and the overall dual faces directly from a UDF. However, as a data-driven method, NDC's performance is largely influenced by how accurately the network is trained. We show in the extensive experiments that our explicit geometry-based method consistently outperforms NDC in terms of reconstruction accuracy, preserving sharp features, and preserving smooth boundaries in the original shapes.", + "bbox": [ + 75, + 604, + 470, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Error Characteristics of Neural UDF", + "text_level": 1, + "bbox": [ + 76, + 753, + 413, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first analyze the characteristics of the errors introduced by the MLP representation used to approximate a UDF, which explains why extracting a mesh surface from a neural UDF is difficult. This analysis also provides a foundation for justifying design choices in our method to overcome these challenging characteristics and achieve robust surface extraction.", + "bbox": [ + 75, + 779, + 468, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin with an ideal UDF $\\bar{F} (\\mathbf{x})$ that represents a target", + "bbox": [ + 96, + 883, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/00d45700a31ea47de216c005386887ea9c56d3523ee59cc76476625a517b3573.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 622, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fd8ef8b328b682963334d419c86c6fb01dcb07f7c1c320abc9a38c0fd1430253.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 635, + 88, + 756, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f5c821f1752547cf0e458ffc709e97aebb268826a5c39066652d74795953ef5e.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 767, + 88, + 888, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/571f375db8d69f60fbb67690305687ec8cec126e822a96b130e973197dbcb7e5.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 500, + 215, + 624, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cd8b8f23ab744ac4c0ef400d0bf30158946b909e777704829624c150d4084f32.jpg", + "image_caption": [ + "(e)", + "Figure 1: A 2D illustrative example. We observe that the neural UDF tends to have larger errors near the zero-level set $\\bar{S}$ and its cut locus. (a) The target shape (the red solid curve), its cut locus (green dashed), and the induced GT UDF; (b) Approximation errors of the neural UDF to the GT UDF; (c) Gradient direction errors between the neural UDF and the GT UDF; For the local region outlined with the white box around the corner, we show the close-up views of the distance error and the gradient direction error of the region in (e) and (f), respectively; (d) The GT UDF and its approximation by a neural UDF showing that the neural UDF is positive and smooth at the zero-level set of the ideal UDF, but the ideal UDF is non-differentiable." + ], + "image_footnote": [], + "bbox": [ + 635, + 196, + 756, + 289 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/70976584046c56c46c7c569dedf3de2be1c326b92592cd56e330b3128e955b6e.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 767, + 196, + 888, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "surface $\\bar{S}$ defined by the zero-level set of $\\bar{F}(\\mathbf{x})$ , that is, $\\bar{S} = \\{\\mathbf{x} | \\bar{F}(\\mathbf{x}) = 0\\} \\subset \\mathbb{R}^3$ . Note that the UDF $\\bar{F}(\\mathbf{x})$ is nonnegative, so it does not have sign changes across the surface $\\bar{S}$ . Furthermore, $\\bar{F}(\\mathbf{x})$ is not differentiable at the surface $\\bar{S}$ or the surface's cut locus [9].", + "bbox": [ + 496, + 537, + 893, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Now suppose that the ideal UDF $\\bar{F}(\\mathbf{x})$ is approximated by an MLP, denoted by $F(\\mathbf{x})$ . The approximation errors of $F(\\mathbf{x})$ to $\\bar{F}(\\mathbf{x})$ are significant in the narrow region around the surface and a narrow region around the cut locus of $\\bar{S}$ , because the neural UDF $F(\\mathbf{x})$ is inherently smooth and thus poorly approximates the ideal $\\bar{F}(\\mathbf{x})$ , which is non-smooth at the surface $\\bar{S}$ and its cut locus. Due to these errors, in general, we have $F(\\mathbf{x}) > 0$ for any $\\mathbf{x} \\in \\bar{S}$ when the MLP is differentiable.", + "bbox": [ + 496, + 614, + 893, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For ease of visualization, we will use a 2D example (see Fig. 1) to illustrate the characteristic behaviors of the approximation errors of $F(\\mathbf{x})$ , whose zero-level set defines an open curve (red in Fig. 1(a)). The characteristics of the approximation errors of $F(\\mathbf{x})$ in 3D space are similar. Detailed visualization and analysis of the error behaviors in 3D space are provided in the supplementary materials. Note", + "bbox": [ + 496, + 753, + 893, + 859 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 The cut locus of a surface is a set of points such that each point of the set has two or more distinct closest points on the surface.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22533", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that the cut locus (Fig. 1(a)) of $\\bar{S}$ touches the sharp features of $\\bar{S}$ . Therefore the neural UDF $F(\\mathbf{x})$ has even more significant errors in terms of both the distance value and the gradient direction around the sharp features (see the error maps of the distance values and the gradient directions in Fig. 1(e,f), respectively). Consequently, in 3D space, the resulting inaccurate distance values and unreliable gradient vectors of the neural UDF $F(\\mathbf{x})$ make it hard to faithfully reconstruct the target surface, especially to preserve the sharp edges in the extracted surface.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To recap, given a neural UDF $F(\\mathbf{x})$ , there are mainly two reasons for the difficulty in extracting a surface $\\bar{S}_M$ to approximate the target surface $\\bar{S}$ defined by the ideal UDF $\\bar{F}(\\mathbf{x})$ : (1) the given neural UDF $F(\\mathbf{x})$ is usually a poor approximation of the ideal UDF $\\bar{F}(\\mathbf{x})$ around the target surface $\\bar{S}$ and its cut locus, where $\\bar{F}(\\mathbf{x})$ is non-differentiable; and (2) the neural UDF $F(\\mathbf{x})$ does not, in general, have a zero-crossing around the target surface $\\bar{S}$ , thus no well-defined surface is associated with it. These issues, faced by current deep neural networks, make it hard to estimate the location of the surface in a numerically stable manner and thus motivate us to develop the two filtering criteria described in Sec. 4.3.", + "bbox": [ + 75, + 242, + 473, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 450, + 171, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method consists of three major designs: (1) a quadratic error function (QEF) that solves for a surface point within a cell for mesh extraction (Sec. 4.1); (2) an adaptive octree data structure that produces high-resolution grid cells while reducing the number of QEF solves required (Sec. 4.2); and (3) a point filtering strategy (Sec. 4.3) for a neural UDF, whose distance values and gradient directions are considerably more noisy compared to the GT as discussed in Sec. 3.", + "bbox": [ + 75, + 476, + 468, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed DualMesh-UDF pipeline has three main steps. Firstly, it employs an adaptive tree to partition the domain of the UDF containing the target surface $\\bar{S}$ . Then, it detects non-empty cells (defined in Sec. 4.2) based on a two-step cell-shape intersection detection method, and solves one surface point per non-empty leaf cell in a least squares manner. For a neural network, two distance bounds are introduced to filter out sample points that may introduce noisy information and contaminate the localization of the target surface. Lastly, since there are no sign changes in the UDFs, we construct the mesh faces by checking each edge shared by four incident non-empty cells in the octree.", + "bbox": [ + 75, + 612, + 470, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. QEF to locate surface points", + "text_level": 1, + "bbox": [ + 76, + 801, + 333, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a cell that contains part of the target surface, we present a procedure to estimate a surface point in this nonempty cell, as illustrated in Fig. 2 for 2D demonstrations. For a neural UDF, we aim to compute a reconstructed point per non-empty cell. For a set of sample points $\\{\\mathbf{p}_i\\}_{i=1}^m$ in", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a6322235f5ca0a395f8b313a48591d9a740331f89fa465344a519d0367fb6b12.jpg", + "image_caption": [ + "(a) A neural UDF case", + "Figure 2: The QEF formulation for the unsigned distance field. The orange lines represent the target surface to be reconstructed. The gray lines in (a) stand for the estimated tangent planes contributed by the corresponding sample points. The red point is the solution to the QEF problem." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 694, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0a30c13cafd837e28f4f7ac3b0c596e4abc65b6588d03750e0661db5f48d35c3.jpg", + "image_caption": [ + "(b) An ideal UDF case" + ], + "image_footnote": [], + "bbox": [ + 697, + 89, + 887, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the given cell, we query the neural UDF $F(\\mathbf{x})$ to get the approximate distance value and the normalized gradient vector at $\\mathbf{p}_i$ , i.e., $d_i = F(\\mathbf{p}_i)$ and $\\mathbf{n}_i = \\nabla F(\\mathbf{p}_i) / \\| \\nabla F(\\mathbf{p}_i)\\|$ , respectively (see Fig. 2a).", + "bbox": [ + 496, + 376, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each of the points contributes to an estimated tangent plane of the target surface $\\bar{S}$ : $\\mathbf{n}_i\\cdot (\\mathbf{p}_i - \\mathbf{x}) - d_i = 0$ . Considering the inaccurate nature of the neural UDF, we define the estimated surface point to be the point that is the closest to all these tangent planes and compute it by minimizing a quadratic error function (a least squares problem):", + "bbox": [ + 496, + 438, + 893, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} = \\arg \\min _ {\\mathbf {x}} \\sum_ {i = 1} ^ {m} \\left(\\mathbf {n} _ {i} \\cdot \\left(\\mathbf {p} _ {i} - \\mathbf {x}\\right) - d _ {i}\\right) ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 541, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m$ is the number of points $\\{\\mathbf{p}_i\\}$ and $\\mathbf{x}$ is a surface point to be solved. In our implementation, we solve the linear least squares problem in Eqn. 1 of the form $\\min \\| Ax - b\\|^2$ . For a non-empty cell, this procedure can yield a reconstructed point $\\mathbf{v}$ , which the target surface $\\bar{S}$ approximately crosses.", + "bbox": [ + 496, + 592, + 890, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, when $A$ is nearly singular, the solution to Eqn. 1 can be located near the boundary of the cell or even outside of it, which would compromise the quality of extracted mesh surfaces. Note that while [10] proposed an approach with an additional regularization point to stabilize such cases, their approach relies on the intersection points at the zero-level set of an SDF with the cell edges. However, in our setting, the zero-level set of the UDF is elusive or unavailable, therefore we do not have such intersections. We propose an alternative approach based on singular value analysis of the matrix $A$ . We denote the three singular val", + "bbox": [ + 496, + 684, + 890, + 851 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1While $\\| \\nabla F\\|$ should be 1 for a distance function, note that since $\\nabla F$ is drawn from a neural representation, it may differ slightly, and thus need normalization.", + "bbox": [ + 500, + 862, + 893, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22534", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ues of the matrix $A$ as $\\sigma_0\\geq \\sigma_1\\geq \\sigma_2\\geq 0$ and consider three cases:", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) In the non-singular case, where all three singular values are much larger than 0, $\\mathbf{v}$ corresponds to a sharp feature point that can be solved directly from Eqn. 1.", + "2) If only $\\sigma_{2} \\approx 0$ , the solution space of $\\mathbf{v}$ corresponds to a linear edge within the cell. This line has the same direction with the singular vector corresponding to $\\sigma_{2}$ and passes the solution of Eqn. 1. We then compute the intersections between this line and all 6 faces of the cell, yielding two intersecting points. We set $\\mathbf{v}$ as the midpoint of these two points.", + "3) In the case where both $\\sigma_{1,2} \\approx 0$ , the part of the target surface enclosed in this cell is approximately planar. Similar to the edge case, we can formulate a plane function from the corresponding singular vectors. By computing the intersecting points between all 12 edges of the cell and this plane, we obtain multiple intersecting points. We use the centroid of these points as $\\mathbf{v}$ ." + ], + "bbox": [ + 75, + 122, + 468, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This way, in the degenerate cases (2 and 3), the point $\\mathbf{v}$ is still a good approximate solution to Eqn. 1 and well positioned inside the cell.", + "bbox": [ + 75, + 383, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a neural UDF (e.g., Fig. 2a), the UDF values $\\{d_i\\}$ and the gradient directions $\\{\\mathbf{n}_i\\}$ are approximated and often unreliable. Consequently, the tangent plane computed from each sample point $\\mathbf{p}_i$ may not exactly align with the target surface. Solving Eqn. 1 yields a least squares solution that is numerically robust to the approximation errors in neural UDF and can lead to an accurate estimation of the surface point in this non-empty cell. For an ideal UDF case, our method is also applicable to yield an accurate result as shown in Fig. 2b.", + "bbox": [ + 75, + 429, + 468, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, to enhance the robustness of our method, we also design a point filtering strategy to remove unreliable sample points, especially those near the target surface and its cut locus, where larger approximation errors exist in the MLP-encoded UDF. The details of this filtering strategy are explained in Sec. 4.3.", + "bbox": [ + 75, + 582, + 468, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differences to DC [10]. Note that the dual contouring (DC) method [10] designs a QEF using intersection points between the surface and cell edges, along with the gradient directions at these intersection points. This approach is not applicable to our objective of extracting surface meshes from a neural UDF. This is because (1) it is difficult to find a reliable intersection point, as an exact zero-level set that crosses an edge may not exist, and (2) the gradient directions in the region with lower UDF values (i.e., closer to the target surface) are highly unstable and thus unreliable for estimating the tangent plane. Hence, our method leverages off-surface points that are sufficiently far from the target surface, as indicated by their UDF values (see Sec. 4.3). Our approach thus has less stringent requirements than the DC method (as presented in [10]), yet produces satisfactory", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6219206ecfbcdd1522f94ca054eae62d26b2ac41bdcd5b58425703e4f7cdc5bb.jpg", + "image_caption": [ + "Figure 3: The pipeline of our mesh extraction method." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 890, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "results even on noisy neural UDFs as demonstrated by our results.", + "bbox": [ + 498, + 263, + 890, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Octree Design and Subdivision", + "text_level": 1, + "bbox": [ + 498, + 308, + 769, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To efficiently process high-resolution data and speed up the mesh extraction process, we employ an octree data structure. As shown in Fig. 3, we recursively subdivide the cells unless they are categorized as empty according to the following criterion or they reach the maximal depth. After the subdivision, all leaf nodes that are non-empty will invoke the QEF procedure to solve for the dual points.", + "bbox": [ + 496, + 333, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Checking obviously empty cells. Firstly, we design a checking condition to help quickly prune obviously empty cells. Given an octree cell $C$ , let $\\mathbf{c}_0$ denote its center and $d_0$ denote the UDF value at $\\mathbf{c}_0$ , that is, $d_0 = F(\\mathbf{c}_0)$ . Then we have the following sufficient condition for cell $C$ to be empty:", + "bbox": [ + 496, + 441, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd _ {0} > \\operatorname {D i a g} (C) / 2, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 537, + 890, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{Diag}(C)$ is the diagonal length of cell $C$ . Since the UDF distance indicates the distance between a spatial point and the target surface, this criterion determines if the target surface lies outside the sphere centered at $\\mathbf{c}_0$ with radius $\\operatorname{Diag}(C)/2$ . Given that this sphere contains the entire octree cell, the cell will not contain any portion of the surface if Eqn. 2 is satisfied. Otherwise, the cell will first be categorized as an unsolved cell. Since neural UDF is a reasonable approximation of the ideal UDF, we tailor Eqn. 2 to $d_0 > \\operatorname{Diag}(C)/2 + \\epsilon$ , where $\\epsilon > 0$ is a tolerance for the approximation error and set to $\\epsilon = 2 \\times 10^{-3}$ throughout all experiments. Empirically, we observed this pruning strategy did not affect the quality of the results.", + "bbox": [ + 496, + 565, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Maximum octree depth. Having a predefined maximal depth of the octree is critical, not only in terms of limiting the amount of computation but also because of the inaccuracy of the neural UDF near the target surface. Specifically, if the octree keeps subdividing the space that contains the target surface, after a certain depth the cells will become too small and the sample points $\\mathbf{p}$ within these cells will lie in the unreliable region of the neural UDF, which will eventually lead to unstable estimation.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22535", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Point filtering for neural UDFs", + "text_level": 1, + "bbox": [ + 76, + 90, + 349, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since our method makes use of the gradient directions and the distance values to estimate a surface point in each non-empty grid cell (see Eqn. 1), the quality of the extracted mesh heavily depends on the approximation accuracy of the neural UDF $F(\\mathbf{x})$ . Consider a neighborhood of a sharp feature of the target surface, as shown in the zoom-in views of Fig. 1(e,f). The distance errors and gradient direction errors are more significant around the target surface and its cut locus.", + "bbox": [ + 75, + 113, + 470, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hence, motivated by our observation and analysis in Sec. 3, to avoid building the QEF using points from these regions, we use the following criteria to filter out unreliable points to enhance the robustness of the QEF solution to the noisy characteristics of the MLP-encoded UDFs.", + "bbox": [ + 75, + 250, + 468, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Criterion 1: Removing sample points potentially near the surface. A candidate point $\\mathbf{p}_i$ is considered too close to the target surface if $F(\\mathbf{p}_i) < \\delta_1$ , where $\\delta_1$ is a preset filtering threshold; such points are discarded. This criterion ensures the sample points are from a region where the distance errors are expected to be relatively small.", + "bbox": [ + 75, + 325, + 468, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Criterion 2: Removing sample points whose projections have large UDF values. While the previous criterion rejects the sample points that are too close to the target surface and thus avoids the erroneous approximation at those regions, some sample points may still have larger UDF errors even far from the target surface (e.g. near the cut locus). We observe that these sample points will not be projected to regions near the target surface. Hence, if the UDF of the projected point $\\{\\mathbf{q}_i\\}$ is large, we consider it to be an incorrect estimate and thus reject the corresponding sample point $\\{\\mathbf{p}_i\\}$ as unreliable. To this end, we introduce the second filtering criterion as follows, $F(\\mathbf{q}_i) > \\delta_2$ , where $\\delta_2$ is another preset filtering bound; again, the corresponding points $\\mathbf{p}_i$ are discarded.", + "bbox": [ + 75, + 416, + 470, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Creating Output Mesh Surfaces", + "text_level": 1, + "bbox": [ + 76, + 635, + 357, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider the regular grid of cubic cells at the maximum depth of the octree. Then each edge of the grid is shared by four cubic cells surrounding the edge. To build the initial mesh connectivity, for each edge of the grid, we examine each of its four incident cells. Similar to [10], the mesh connection rule is designed as follows: if all four incident cells are non-empty, a quad-face candidate that connects the four dual points in these cells will be constructed. To ensure the correct connectivity, we validate and triangulate the quad faces. We further examine if the normals of the face candidates can consistently reflect their geometric property, being a sharp corner, part of a sharp edge, or part of a plane as classified by the SVD shape analysis (Sec. 4.1). For example, all correct triangulated faces incident to a surface point classified as part of a plane would have all their normals parallel to the singular vector corresponding to the", + "bbox": [ + 75, + 659, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "largest singular value from SVD shape analysis, while the normals of triangulated faces incident to a surface point being part of a sharp edge should be orthogonal to its singular vector direction. This way, we reject all inconsistent face candidates.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also provide a practical approach to make sure the output mesh is manifold when the desired target surface is manifold: With the help of the octree cells and the definition of reconstructed dual points, it is feasible to generate an auxiliary blocky model by moving all mesh vertices to the centroids of the corresponding cells. By extracting the outer envelope of this auxiliary model, we obtain a manifold structure. Finally, we tessellate the surface points using the connectivity from that manifold structure to produce a manifold mesh. This approach is crucial for applications that require manifold meshes (e.g. parameterization, remeshing, or shape analysis).", + "bbox": [ + 496, + 167, + 890, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 498, + 364, + 633, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental details and metrics", + "text_level": 1, + "bbox": [ + 500, + 390, + 790, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental setting. We rescaled all the shapes to a bounding cube with a side length of 2, centered at the origin. To improve efficiency, we share sample points between cells, i.e. $m = 27$ points in each non-empty leaf cell, including 8 corner points, 12 edge midpoints, 6 face midpoints, and 1 centroid point. We set the filtering parameters $\\delta_{1} = 2 \\times 10^{-3}$ and $\\delta_{2} = 2 \\times 10^{-3}$ to filter the sampled points as described in Sec. 4.3. We notice that more sample points may bring marginal performance improvement but incur computational overhead in our ablation study.", + "bbox": [ + 496, + 415, + 890, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given that the grid resolution of $128^{3}$ and $256^{3}$ are commonly used in related works, we compare our method with these prior techniques using a maximal octree resolution of $128^{3}$ and $256^{3}$ . We also discuss how the resolution will affect our results by varying the maximum depth of the octree to have a resolution from $64^{3}$ to $512^{3}$ in our supplementary material.", + "bbox": [ + 496, + 566, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MLP architecture. To overfit single shapes using individual MLPs, we employed the MLP implementation provided by [18]. The activation functions are the Sine activation, except for the last one which is a SoftPlus $(\\beta = 100)$ activation to ensure the output value is non-negative. All neural UDFs in the experiments were trained with the described MLP implementation. All MLP networks were trained for $3k$ iterations with the ADAM optimizer to minimize the difference between the predicted and the GT unsigned distance fields. For more general tasks, we also test an MLP with latent codes that represent a shape space, following the network and training settings in [5]. We report the timing performance on a Linux desktop with an Intel Core™ i7-10870H CPU and an NVIDIA GeForce RTX 3090 graphics card. We elaborate on the loss function as", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22536", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5b0ccebcaeb2fa41e52fe066cb3f788591bbd72644ac4a0147bfd3051ae304f5.jpg", + "image_caption": [ + "Figure 4: Meshes extracted from neural UDFs. We show 6 examples and compare our results to the mesh extraction results in MeshUDF [7], CAP-UDF [21] and NDC [3]. Our method preserves sharp geometric features better, yields results without undesirable holes, and reconstructs the original smooth boundaries of open surfaces." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 872, + 847 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22537", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a31f2ae1fb6e62cb60ddb3689ce4f74be1770bb54ecb677b28800d93afd8caa2.jpg", + "table_caption": [ + "Table 1: Quantitative comparison between the results obtained by our method and those by the competing methods, MeshUDF, CAP-UDF, and NDC. The average performance on the dataset containing 354 shapes is reported. The Chamfer distance (CD) and the Hausdorff Distance (HD) are scaled by $10^{-4}$ and $10^{-3}$ , respectively. T1 and T2 stand for the time spent (seconds) on the mesh extraction and that on the UDF query, respectively." + ], + "table_footnote": [], + "table_body": "
MGNThingi10KABCRunning time
CD↓F-score↑HD↓CD↓F-score↑HD↓CD↓F-score↑HD↓T1T2
1283Ours2.3898.0911.911.9797.519.213.6993.4111.470.2971.184
MeshUDF [7]4.7690.0622.394.5790.7815.338.4987.7423.110.3131.316
CAP-UDF [21]13.5387.1060.0419.5584.8557.4927.7577.4353.466.4664.764
NDC [3]3.3295.6112.042.7496.449.693.9592.6514.042.3581.324
2563Ours2.0398.967.351.6098.296.782.2296.578.201.4925.508
MeshUDF [7]3.4595.1012.83.3394.138.913.2593.8511.491.60010.27
CAP-UDF [21]8.4592.8849.4012.1790.5347.2821.6783.8843.3349.41637.336
NDC [3]2.5498.047.842.4297.167.832.3696.1011.389.36310.256
", + "bbox": [ + 119, + 160, + 849, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/86e5515a7df0ce7d4b9c44bbdd36e2985cdbf3e0560d590bd37fa66618978c81.jpg", + "table_caption": [ + "Table 2: Quantitative comparison between our method and three competing methods, MeshUDF, CAP-UDF, and NDC on a pre-trained UDF network with latent codes, on the ShapeNet dataset. In this table, CD and HD are scaled by $10^{-3}$ and $10^{-2}$ respectively. The F-score is calculated with a threshold of 0.006." + ], + "table_footnote": [], + "table_body": "
CD↓F-score↑HD ↓
Ours3.4182.303.92
MeshUDF [7]4.1079.354.67
CAP-UDF [21]7.7070.307.21
NDC [3]7.1678.3812.38
", + "bbox": [ + 117, + 411, + 426, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "well as other details regarding training the neural UDFs in the supplementary materials.", + "bbox": [ + 75, + 517, + 468, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metrics. To evaluate the performance of our method and the other methods, we adopt the following metrics, i.e. the double-sided Chamfer distance (CD), the F-score based on CD, and the Hausdorff distance (HD). The Chamfer distance reflects the overall quality of the extracted surface mesh as compared to the GT shape. The F-score indicates the percentage of points that are reconstructed correctly under a threshold (set to 0.001 for shape-overfitting neural UDFs, and 0.01 for shared UDF network with latent codes). The Hausdorff distance can reveal if there is anything missing or redundant in the reconstructed geometry (e.g. a hole or a floating piece).", + "bbox": [ + 75, + 547, + 470, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Comparison with SOTA methods", + "text_level": 1, + "bbox": [ + 76, + 739, + 367, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compared our results with three existing methods: 1) MeshUDF [7], 2) the standalone mesh extraction module presented in CAP-UDF [21], and 3) the UNDC presented in NDC [3]. The first two methods use gradients of the UDF to estimate sign changes in the field to invoke the Marching Cubes method to extract a surface mesh from the UDF. NDC proposes a data-driven approach to extract a surface mesh from the grid-based representation of an implicit field. Shape-overfitting neural UDFs", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first compare our results to those obtained by these three methods on a shape collection consisting of shapes from four public sources: 1) 100 from the Thingi10K Dataset [22] containing 3D printing models; 2) 134 from the MGN Dataset [2] containing garments with open boundaries; 3) 100 from the ABC Dataset [12] containing CAD models; and 4) 20 commonly used shapes in geometric processing research. In addition, we also compare different methods on a Möbius strip, which is a non-orientable surface. We fit each shape with an independent neural UDF and apply mesh extraction to each of these neural UDFs.", + "bbox": [ + 496, + 311, + 893, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 reports the performance of different methods on this shape collection. Our method outperforms the other three competing methods in terms of all three quantitative metrics. While all of these methods use uniform grids, our octree structure results in increased computing efficiency. MeshUDF and CAP-UDF would likely be accelerated by adopting an octree structure, however, it would be nontrivial to adopt an octree structure for NDC. For a fair comparison, we also test our method without using the octree acceleration approach for the resolution of $128^{3}$ ; by doing so, the T1 time increases to 2.70s, and the T2 time increases to 5.54s.", + "bbox": [ + 496, + 478, + 893, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To qualitatively compare results obtained by different methods, we show the mesh surfaces extracted from neural UDFs for several shapes in Fig. 4. We can see that our results are higher quality than those produced by MeshUDF and CAP-UDF. These two methods cannot preserve sharp geometric features due to the use of the Marching Cubes method. Compared to NDC - a data-driven method - our method also produces consistently better results both quantitatively and qualitatively. Specifically, some mesh surfaces extracted by NDC are less smooth than ours. One example (the Pants) is shown in the fourth row in Fig. 4. The staircase artifact observed may be attributed to NDC having been trained on the ABC dataset [12] containing only mechanical components.", + "bbox": [ + 496, + 659, + 893, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method is the only method that recovers open boundaries faithfully, while the other compared methods", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22538", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bed410780710cbba2ea8c8d245e2d57bc6f2e685860ef795653e52403a38e93b.jpg", + "table_caption": [ + "Table 3: Quantitative comparison between two different sampling number settings (27 and 125 per cell respectively). The Chamfer distance (CD) and the Hausdorff Distance (HD) are scaled by $10^{-4}$ and $10^{-3}$ , respectively. T1 and T2 stand for the time spent (seconds) on the mesh extraction and that on the UDF query, respectively." + ], + "table_footnote": [], + "table_body": "
MGNThingi10KABCRunning time
CD↓F-score↑HD↓CD↓F-score↑HD↓CD↓F-score↑HD↓T1T2
272.3898.0911.911.9797.519.213.6993.4111.470.2971.184
1252.3498.1111.771.9197.529.023.7092.8911.120.9657.608
", + "bbox": [ + 119, + 145, + 849, + 207 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "show staircase artifacts or even redundant pieces near the open boundaries; see the first two rows of Fig. 4. Also, unexpected holes can be observed on the mesh surface extracted by the competing methods from the neural UDFs. Empirically, our method produces quality results without redundant pieces or unexpected holes, generating clear-cut boundaries of the open surfaces as shown in the top four rows of Fig. 4, and reproducing the sharp features as shown in the bottom two rows of the same figure.", + "bbox": [ + 75, + 231, + 472, + 366 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "More results and comparisons (also on the GT UDF) are presented in our Supplementary Materials, which further validate the superior performance of our method.", + "bbox": [ + 75, + 366, + 470, + 412 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Shared, pre-trained UDF network with latent codes. We consider a single neural network with latent codes trained to represent the entire shape space, where each shape is associated with a unique latent code. Specifically, we test our method and the competing methods on the pre-trained UDF network provided by [5] representing 300 ShapeNet car models. Although this UDF network is less accurate than the other overfitting-based network settings considered earlier, our method still outperforms the other methods by a significant margin as reported in Tab. 2, showing the versatility and robustness of our method.", + "bbox": [ + 75, + 412, + 470, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 588, + 227, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conducted an ablation study on our method using sampling numbers (the number $m$ in Eqn. 1) of $27(3^{3})$ and $125(5^{3})$ to justify our design choice to use 27 as our sampling number. In Tab. 3, we demonstrate that using 125 sample points per cell results in a marginal improvement but incurs a significantly higher computing cost.", + "bbox": [ + 75, + 611, + 468, + 703 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.4. Limitations", + "text_level": 1, + "bbox": [ + 76, + 710, + 204, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While DualMesh-UDF demonstrates the ability to accurately extract surfaces from neural UDFs, some limitations remain. First, the extracted surface cannot be adaptively subdivided with respect to sharp features or fine geometry details. To faithfully reconstruct the fine geometry details, we need to pre-define a sufficient depth of the octree and solve the QEF problem in each of its non-empty leaf nodes. Second, to share the UDF and gradient values at sampled points between adjacent octree cells, we adopt a regular grid sampling strategy. A more flexible and adaptive sampling strategy may bring further improvement to our method.", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 229, + 619, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have presented a method for extracting high-quality surface meshes from unsigned distance fields. In order to attain robust performance on MLP-encoded neural UDFs, we discuss the characteristics of the approximation errors of the neural UDF and develop an adaptive octree-based method to effectively localize the target surface embedded in the given UDF. Extensive experiments show that our method outperforms the SOTA methods and produces high-quality results with sharp geometry features, with clear open boundaries, and free of undesirable holes.", + "bbox": [ + 496, + 255, + 890, + 405 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 498, + 420, + 679, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is partly supported by the Innovation and Technology Commission of the HKSAR Government through the InnoHK initiative. The authors would like to thank Yanhong Lin and Ruixing Jia for their generous help on the project.", + "bbox": [ + 496, + 446, + 893, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 536, + 596, + 551 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In International Conference on Learning Representations, 2021.", + "[2] Bharat Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5419-5429, 2019.", + "[3] Zhiqin Chen, Andrea Tagliasacchi, Thomas Funkhouser, and Hao Zhang. Neural dual contouring. ACM Trans. Graph., 41(4), jul 2022.", + "[4] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6968-6979, 2020.", + "[5] Julian Chibane, Mohamad Aymen mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21638-21652. Curran Associates, Inc., 2020.", + "[6] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learn" + ], + "bbox": [ + 506, + 560, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22539", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing shapes. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 3789-3799. PMLR, 13-18 Jul 2020.", + "[7] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 576-592, Cham, 2022. Springer Nature Switzerland.", + "[8] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, and Thomas Funkhouser. Local implicit grid representations for 3d scenes. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6000-6009, 2020.", + "[9] M.W. Jones, J.A. Baerentzen, and M. Sramek. 3d distance fields: a survey of techniques and applications. IEEE Transactions on Visualization and Computer Graphics, 12(4):581-599, 2006.", + "[10] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. ACM Trans. Graph., 21(3):339-346, jul 2002.", + "[11] Leif P. Kobbelt, Mario Botsch, Ulrich Schwanecke, and Hans-Peter Seidel. Feature sensitive surface extraction from volume data. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '01, page 57-66, New York, NY, USA, 2001. Association for Computing Machinery.", + "[12] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.", + "[13] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '87, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery.", + "[14] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. ACM Trans. Graph., 40(4), jul 2021.", + "[15] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4455-4465, 2019.", + "[16] Mateusz Michalkiewicz, Jhony Kaesemodel Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4742-4751, 2019.", + "[17] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019.", + "[18] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7462-7473. Curran Associates, Inc., 2020.", + "[19] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11353-11362, 2021.", + "[20] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7537-7547. Curran Associates, Inc., 2020.", + "[21] Junsheng Zhou, Baorui Ma, Yu-Shen Liu, Yi Fang, and Zhizhong Han. Learning consistency-aware unsigned distance functions progressively from raw point clouds. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022.", + "[22] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016." + ], + "bbox": [ + 503, + 92, + 890, + 526 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22540", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Surface Extraction from Neural Unsigned Distance Fields/a0829e7a-732d-4761-bffa-ad35624b76ef_model.json b/2023/Surface Extraction from Neural Unsigned Distance Fields/a0829e7a-732d-4761-bffa-ad35624b76ef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4b6cc9d5577fdad127cd3e3c3d95c65c0f0e8a06 --- /dev/null +++ b/2023/Surface Extraction from Neural Unsigned Distance Fields/a0829e7a-732d-4761-bffa-ad35624b76ef_model.json @@ -0,0 +1,1870 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.194, + 0.131, + 0.778, + 0.154 + ], + "angle": 0, + "content": "Surface Extraction from Neural Unsigned Distance Fields" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.179, + 0.836, + 0.227 + ], + "angle": 0, + "content": "Congyi Zhang\\(^{1,2*}\\) Guying Lin\\(^{1*}\\) Lei Yang\\(^{1,2}\\) Xin Li\\(^{3}\\) Taku Komura\\(^{1}\\) Scott Schaefer\\(^{3}\\) John Keyser\\(^{3}\\) Wenping Wang\\(^{3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.241, + 0.668, + 0.278 + ], + "angle": 0, + "content": "1The University of Hong Kong 2TransGP \n3Texas A&M University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.312, + 0.314, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.345, + 0.473, + 0.557 + ], + "angle": 0, + "content": "We propose a method, named DualMesh-UDF, to extract a surface from unsigned distance functions (UDFs), encoded by neural networks, or neural UDFs. Neural UDFs are becoming increasingly popular for surface representation because of their versatility in presenting surfaces with arbitrary topologies, as opposed to the signed distance function that is limited to representing a closed surface. However, the applications of neural UDFs are hindered by the notorious difficulty in extracting the target surfaces they represent. Recent methods for surface extraction from a neural UDF suffer from significant geometric errors or topological artifacts due to two main difficulties: (1) A UDF does not exhibit sign changes; and (2) A neural UDF typically has substantial approximation errors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.473, + 0.846 + ], + "angle": 0, + "content": "DualMesh-UDF addresses these two difficulties. Specifically, given a neural UDF encoding a target surface \\(\\bar{S}\\) to be recovered, we first estimate the tangent planes of \\(\\bar{S}\\) at a set of sample points close to \\(\\bar{S}\\). Next, we organize these sample points into local clusters, and for each local cluster, solve a linear least squares problem to determine a final surface point. These surface points are then connected to create the output mesh surface, which approximates the target surface. The robust estimation of the tangent planes of the target surface and the subsequent minimization problem constitute our core strategy, which contributes to the favorable performance of DualMesh-UDF over other competing methods. To efficiently implement this strategy, we employ an adaptive Octree. Within this framework, we estimate the location of a surface point in each of the octree cells identified as containing part of the target surface. Extensive experiments show that our method outperforms existing methods in terms of surface reconstruction quality while maintaining comparable computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.312, + 0.632, + 0.329 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.895, + 0.55 + ], + "angle": 0, + "content": "Implicit surfaces are widely used for surface representation in computer vision and computer graphics. An implicit surface is usually defined as a level set of a function, such as the zero-level set of a signed distance function (SDF). Extracting a mesh representation of an implicit surface from its defining equation is therefore a critical task for surface visualization and processing. Recent advances in machine learning have given rise to a new kind of implicit surface, called a neural implicit surface. A neural implicit surface is a level-set of a function encoded by an MLP (multilayer perceptron) and has the advantage of compactness and inherent smoothness thanks to its MLP representation. SDFs or occupancy fields are widely used in these implicit representations [17, 15, 16, 4, 18, 20, 6, 1, 8, 14, 19]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.895, + 0.702 + ], + "angle": 0, + "content": "However, neural implicit surfaces based on the SDF or occupancy fields require inside/outside labeling and thus can only represent orientable and closed surfaces. Hence, as an extension, unsigned distance functions (UDFs) have been used to represent surfaces of arbitrary topologies, including open surfaces with boundaries or non-orientable surfaces (e.g. the Möbius strip). Despite its versatility, applications of a UDF-based surface representation are severely hindered by the difficulty in extracting the target surface it represents, as shown in [5] and [7]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.838 + ], + "angle": 0, + "content": "Problem formulation: Suppose that a surface \\(\\bar{S}\\), called the target surface, is defined as the zero-level set of its unsigned distance function (UDF) \\(\\bar{F}(\\mathbf{p})\\). Then suppose that this UDF \\(\\bar{F}(\\mathbf{p})\\) is approximated by a neural network with the resulting network-encoded UDF being referred to as the neural UDF, denoted by \\(F(\\mathbf{p})\\). Given a neural UDF \\(F(\\mathbf{p})\\), the surface extraction problem is to robustly extract a surface \\(\\mathcal{S}\\) from \\(F(\\mathbf{p})\\) such that \\(\\mathcal{S}\\) well approximates the target surface \\(\\bar{S}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Challenges: The difficulty in surface extraction from a neural UDF arises from two aspects: (1) A UDF does not have zero-crossings (or sign changes) across the surface it represents. As a result, traditional mesh extraction methods" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.203, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.221, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22531" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "that rely on zero-crossings (e.g. Marching Cubes [13, 11], Dual Contouring [10], and their variants) are not applicable to UDFs. (2) The MLP representation of a neural UDF tends to have significant approximation errors around the target surface (see detailed error characteristics of neural UDFs in Sec. 3). This makes it even more challenging to extract a high-quality approximation of the target surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.456 + ], + "angle": 0, + "content": "Several methods, MeshUDF [7], CAP-UDF [21] and Neural Dual Contouring (NDC) [3], have recently been proposed for extracting a mesh surface from a UDF. MeshUDF and CAP-UDF attempt to infer the gradients of a UDF on the grids and determine the sign changes of the estimated gradients, invoking the Marching Cubes method for mesh extraction. When applied to a neural UDF, the sign-change inference step of this method suffers from instability due to the non-negligible error introduced by the approximate MLP representation near the surface where the gradients of the ideal UDF are undefined. As a result, the extracted meshes are less accurate and often have topological errors (e.g. holes). The NDC method proposes a data-driven Dual Contouring approach to predict the position of mesh vertices and dual faces directly from the UDF data. When applied to a neural UDF, this method often produces meshes with considerable artifacts such as holes, zig-zags, etc." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.458, + 0.47, + 0.563 + ], + "angle": 0, + "content": "We develop a new strategy, consisting of novel sampling and efficient optimization techniques to address the difficulties in surface extraction from the neural UDF. Suppose the input is a neural UDF \\( F(\\mathbf{p}) \\) encoding the target surface \\( \\bar{S} \\) to be recovered. Our strategy has two key steps: (1) computing approximate tangent planes of the target surface; and (2) local minimization for generating final surface points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.746 + ], + "angle": 0, + "content": "In Step (1), we first generate sample points \\(\\mathbf{p}_i\\) around, but not too close to, the target surface \\(\\bar{S}\\), because the UDF values and gradients at locations too close to \\(\\bar{S}\\) are relatively unreliable. Thus, \\(\\mathbf{p}_i\\) are called off-surface sample points. For each \\(\\mathbf{p}_i\\), we use the UDF value \\(F(\\mathbf{p}_i)\\) and its gradient \\(\\nabla F(\\mathbf{p}_i)\\) to project \\(\\mathbf{p}_i\\) towards the target surface to obtain point \\(\\mathbf{q}_i = \\mathbf{p}_i - F(\\mathbf{p}_i)\\mathbf{n}_i(\\mathbf{p}_i)\\) where \\(\\mathbf{n}(\\mathbf{p}_i) = \\nabla F(\\mathbf{p}_i) / \\| \\nabla F(\\mathbf{p}_i)\\|\\) [5]. These points \\(\\mathbf{q}_i\\) are called projection points. Although the points \\(\\mathbf{q}_i\\) are very close to the target surface \\(\\bar{S}\\), as we will show later, the noisy error in the neural UDF makes these points a poor approximation to the target surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.47, + 0.869 + ], + "angle": 0, + "content": "To further improve surface accuracy, for each projection point \\(\\mathbf{q}_i\\) we generate an estimated tangent plane \\(T_{i}\\) of \\(\\bar{S}\\) such that \\(T_{i}\\) passes through \\(\\mathbf{q}_i\\) and has the unit normal vector \\(\\mathbf{n}_i\\). Note that the normal vector \\(\\mathbf{n}_i\\) of \\(T_{i}\\) is set to be \\(\\mathbf{n}(\\mathbf{p}_i)\\) rather than \\(\\mathbf{n}(\\mathbf{q}_i)\\) since the former is a more reliable estimation. This is because the initial sample point \\(\\mathbf{p}_i\\) is not too close to \\(\\bar{S}\\), so the gradient \\(\\nabla F(\\mathbf{p}_i)\\) is less contaminated by the pronounced errors of the neural UDF close to \\(\\bar{S}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In Step (2), the estimated tangent planes are organized into clusters, which may overlap. For each cluster of tan-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "gent planes \\( T_{i} \\), we solve a linear least squares problem to produce a final surface point \\( \\mathbf{s}_i \\) that minimizes the sum of its squared distances to the tangent planes \\( T_{i} \\). This minimization step based on tangent planes not only provides an accurate surface point but also allows us to faithfully reconstruct the sharp edges of the target surface. Finally, all the surface points \\( \\mathbf{s}_i \\) from all the clusters are connected to form the output mesh surface to approximate the target surface \\( \\bar{S} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.363 + ], + "angle": 0, + "content": "To efficiently implement the above strategy, our DualMesh-UDF method adopts an adaptive Octree structure to partition the space containing the target surface to regular cells. We developed efficient procedures to determine those cells that contain part of the target surface and perform the sampling and minimization procedures in each occupied cell. To connect the surface points to create the output mesh, we follow the Dual Contouring approach, connecting surface points residing in adjacent grid cells to create polygons dual to octree edges." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that our DualMesh-UDF significantly outperforms existing methods in terms of surface reconstruction accuracy and sharp feature preservation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.575 + ], + "angle": 0, + "content": "The main contribution of this work is a new algorithm to robustly and accurately extract a surface from a neural UDF. To overcome the inevitable approximation errors near the target surface and cut locus, we obtain robust estimation of surface tangent planes by leveraging off-surface sample points, use least square minimization to better predict the surface points, and achieve high-quality surface extraction results with sharp features better preserved compared to the state of the art. The code is available at https://github.com/cong-yi/DualMesh-UDF." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.642, + 0.605 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.72 + ], + "angle": 0, + "content": "The Marching Cubes method [13] and its variants have been established as the de facto standard of converting distance fields to boundary mesh representations. Using the gradient information of the distance function, Extended Marching Cubes [11] and Dual Contouring [10] are both capable of producing meshes with faithfully preserved sharp features (e.g., corners and edges)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.885 + ], + "angle": 0, + "content": "However, all of these methods require inside/outside labeling on the sampling grid, which is either a regular grid of cubes or an adaptive octree grid of cells, to determine whether any zero-level set surface of the signed distance field crosses a particular grid cell. This requirement for sign changes limits the application of these methods to only SDFs or its variants that have sign changes across the underlying surface. Hence, they cannot be directly used to extract a mesh surface from a UDF, which is now often used to represent arbitrary surfaces such as open surfaces or non-orientable surfaces [5]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our method is similar to [10] only in the sense that we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22532" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.257 + ], + "angle": 0, + "content": "follow [10] to solve a quadratic error function (QEF) to estimate a surface point per grid cell. But unlike [10], where sign changes are available and the gradients at the zero-crossings are reliable, unsigned distance fields do not have sign changes and are non-differentiable at the zero level set. Additionally, we will show that the information around the zero-level set of a neural UDF is unreliable. Due to these two reasons, we alternatively make use of the spatial sample points that are off the target surface in the neural UDF to find reliably estimated tangent planes and formulate a QEF for estimating the final surface points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.605 + ], + "angle": 0, + "content": "Three relevant methods, MeshUDF [7], CAP-UDF [21] and Neural Dual Contouring (NDC) [3], have recently been proposed to extract meshes from UDFs. MeshUDF [7] and CAP-UDF [21] use the gradient information of the UDF to assign different signs (+ or -) to grid points on different sides of the underlying surface, thus invoking Marching Cubes to extract the surface according to the sign labels. However, the quality of the estimated sign labels can be significantly affected by the accuracy of the neural UDF. As we will show later, the neural UDF becomes less accurate around the target surface it represents, contaminating the inferred sign labels and explaining the poor performance of these two methods. Their use of Marching Cubes is incapable of preserving sharp corners or edges, while the use of a regular grid incurs high memory overhead when a high grid resolution is needed to resolve shape details. In contrast, our DualMesh-UDF method preserves sharp features and uses an adaptive octree grid to reduce computational expense even with a high grid resolution. Furthermore, estimating the surface point location is done by solving a least square problem (QEF) which is more robust than the gradient-based sign-labeling strategy used in MeshUDF in terms of the topology of the resulting meshes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.472, + 0.742 + ], + "angle": 0, + "content": "The NDC method [3] uses a data-driven approach to train a neural network that predicts vertex position per regular grid cell and the overall dual faces directly from a UDF. However, as a data-driven method, NDC's performance is largely influenced by how accurately the network is trained. We show in the extensive experiments that our explicit geometry-based method consistently outperforms NDC in terms of reconstruction accuracy, preserving sharp features, and preserving smooth boundaries in the original shapes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.414, + 0.772 + ], + "angle": 0, + "content": "3. Error Characteristics of Neural UDF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.884 + ], + "angle": 0, + "content": "We first analyze the characteristics of the errors introduced by the MLP representation used to approximate a UDF, which explains why extracting a mesh surface from a neural UDF is difficult. This analysis also provides a foundation for justifying design choices in our method to overcome these challenging characteristics and achieve robust surface extraction." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.885, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We begin with an ideal UDF \\(\\bar{F} (\\mathbf{x})\\) that represents a target" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.623, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.183, + 0.574, + 0.193 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.757, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.183, + 0.704, + 0.193 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.089, + 0.89, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.183, + 0.837, + 0.193 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.217, + 0.625, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.29, + 0.571, + 0.301 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.197, + 0.757, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.291, + 0.701, + 0.301 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.197, + 0.89, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.82, + 0.291, + 0.834, + 0.301 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.509 + ], + "angle": 0, + "content": "Figure 1: A 2D illustrative example. We observe that the neural UDF tends to have larger errors near the zero-level set \\(\\bar{S}\\) and its cut locus. (a) The target shape (the red solid curve), its cut locus (green dashed), and the induced GT UDF; (b) Approximation errors of the neural UDF to the GT UDF; (c) Gradient direction errors between the neural UDF and the GT UDF; For the local region outlined with the white box around the corner, we show the close-up views of the distance error and the gradient direction error of the region in (e) and (f), respectively; (d) The GT UDF and its approximation by a neural UDF showing that the neural UDF is positive and smooth at the zero-level set of the ideal UDF, but the ideal UDF is non-differentiable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.894, + 0.615 + ], + "angle": 0, + "content": "surface \\(\\bar{S}\\) defined by the zero-level set of \\(\\bar{F}(\\mathbf{x})\\), that is, \\(\\bar{S} = \\{\\mathbf{x} | \\bar{F}(\\mathbf{x}) = 0\\} \\subset \\mathbb{R}^3\\). Note that the UDF \\(\\bar{F}(\\mathbf{x})\\) is nonnegative, so it does not have sign changes across the surface \\(\\bar{S}\\). Furthermore, \\(\\bar{F}(\\mathbf{x})\\) is not differentiable at the surface \\(\\bar{S}\\) or the surface's cut locus [9]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.895, + 0.752 + ], + "angle": 0, + "content": "Now suppose that the ideal UDF \\(\\bar{F}(\\mathbf{x})\\) is approximated by an MLP, denoted by \\(F(\\mathbf{x})\\). The approximation errors of \\(F(\\mathbf{x})\\) to \\(\\bar{F}(\\mathbf{x})\\) are significant in the narrow region around the surface and a narrow region around the cut locus of \\(\\bar{S}\\), because the neural UDF \\(F(\\mathbf{x})\\) is inherently smooth and thus poorly approximates the ideal \\(\\bar{F}(\\mathbf{x})\\), which is non-smooth at the surface \\(\\bar{S}\\) and its cut locus. Due to these errors, in general, we have \\(F(\\mathbf{x}) > 0\\) for any \\(\\mathbf{x} \\in \\bar{S}\\) when the MLP is differentiable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.754, + 0.894, + 0.86 + ], + "angle": 0, + "content": "For ease of visualization, we will use a 2D example (see Fig. 1) to illustrate the characteristic behaviors of the approximation errors of \\( F(\\mathbf{x}) \\), whose zero-level set defines an open curve (red in Fig. 1(a)). The characteristics of the approximation errors of \\( F(\\mathbf{x}) \\) in 3D space are similar. Detailed visualization and analysis of the error behaviors in 3D space are provided in the supplementary materials. Note" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1 The cut locus of a surface is a set of points such that each point of the set has two or more distinct closest points on the surface." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22533" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.242 + ], + "angle": 0, + "content": "that the cut locus (Fig. 1(a)) of \\(\\bar{S}\\) touches the sharp features of \\(\\bar{S}\\). Therefore the neural UDF \\(F(\\mathbf{x})\\) has even more significant errors in terms of both the distance value and the gradient direction around the sharp features (see the error maps of the distance values and the gradient directions in Fig. 1(e,f), respectively). Consequently, in 3D space, the resulting inaccurate distance values and unreliable gradient vectors of the neural UDF \\(F(\\mathbf{x})\\) make it hard to faithfully reconstruct the target surface, especially to preserve the sharp edges in the extracted surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.474, + 0.439 + ], + "angle": 0, + "content": "To recap, given a neural UDF \\( F(\\mathbf{x}) \\), there are mainly two reasons for the difficulty in extracting a surface \\( \\bar{S}_M \\) to approximate the target surface \\( \\bar{S} \\) defined by the ideal UDF \\( \\bar{F}(\\mathbf{x}) \\): (1) the given neural UDF \\( F(\\mathbf{x}) \\) is usually a poor approximation of the ideal UDF \\( \\bar{F}(\\mathbf{x}) \\) around the target surface \\( \\bar{S} \\) and its cut locus, where \\( \\bar{F}(\\mathbf{x}) \\) is non-differentiable; and (2) the neural UDF \\( F(\\mathbf{x}) \\) does not, in general, have a zero-crossing around the target surface \\( \\bar{S} \\), thus no well-defined surface is associated with it. These issues, faced by current deep neural networks, make it hard to estimate the location of the surface in a numerically stable manner and thus motivate us to develop the two filtering criteria described in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.451, + 0.172, + 0.467 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.612 + ], + "angle": 0, + "content": "Our method consists of three major designs: (1) a quadratic error function (QEF) that solves for a surface point within a cell for mesh extraction (Sec. 4.1); (2) an adaptive octree data structure that produces high-resolution grid cells while reducing the number of QEF solves required (Sec. 4.2); and (3) a point filtering strategy (Sec. 4.3) for a neural UDF, whose distance values and gradient directions are considerably more noisy compared to the GT as discussed in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.471, + 0.794 + ], + "angle": 0, + "content": "The proposed DualMesh-UDF pipeline has three main steps. Firstly, it employs an adaptive tree to partition the domain of the UDF containing the target surface \\(\\bar{S}\\). Then, it detects non-empty cells (defined in Sec. 4.2) based on a two-step cell-shape intersection detection method, and solves one surface point per non-empty leaf cell in a least squares manner. For a neural network, two distance bounds are introduced to filter out sample points that may introduce noisy information and contaminate the localization of the target surface. Lastly, since there are no sign changes in the UDFs, we construct the mesh faces by checking each edge shared by four incident non-empty cells in the octree." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.334, + 0.819 + ], + "angle": 0, + "content": "4.1. QEF to locate surface points" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.904 + ], + "angle": 0, + "content": "Given a cell that contains part of the target surface, we present a procedure to estimate a surface point in this nonempty cell, as illustrated in Fig. 2 for 2D demonstrations. For a neural UDF, we aim to compute a reconstructed point per non-empty cell. For a set of sample points \\(\\{\\mathbf{p}_i\\}_{i=1}^m\\) in" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.696, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.247, + 0.669, + 0.261 + ], + "angle": 0, + "content": "(a) A neural UDF case" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.09, + 0.888, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.247, + 0.862, + 0.261 + ], + "angle": 0, + "content": "(b) An ideal UDF case" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.273, + 0.895, + 0.351 + ], + "angle": 0, + "content": "Figure 2: The QEF formulation for the unsigned distance field. The orange lines represent the target surface to be reconstructed. The gray lines in (a) stand for the estimated tangent planes contributed by the corresponding sample points. The red point is the solution to the QEF problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.438 + ], + "angle": 0, + "content": "the given cell, we query the neural UDF \\( F(\\mathbf{x}) \\) to get the approximate distance value and the normalized gradient vector at \\( \\mathbf{p}_i \\), i.e., \\( d_i = F(\\mathbf{p}_i) \\) and \\( \\mathbf{n}_i = \\nabla F(\\mathbf{p}_i) / \\| \\nabla F(\\mathbf{p}_i)\\| \\), respectively (see Fig. 2a)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.894, + 0.531 + ], + "angle": 0, + "content": "Each of the points contributes to an estimated tangent plane of the target surface \\(\\bar{S}\\): \\(\\mathbf{n}_i\\cdot (\\mathbf{p}_i - \\mathbf{x}) - d_i = 0\\). Considering the inaccurate nature of the neural UDF, we define the estimated surface point to be the point that is the closest to all these tangent planes and compute it by minimizing a quadratic error function (a least squares problem):" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.542, + 0.892, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} = \\arg \\min _ {\\mathbf {x}} \\sum_ {i = 1} ^ {m} \\left(\\mathbf {n} _ {i} \\cdot \\left(\\mathbf {p} _ {i} - \\mathbf {x}\\right) - d _ {i}\\right) ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.684 + ], + "angle": 0, + "content": "where \\(m\\) is the number of points \\(\\{\\mathbf{p}_i\\}\\) and \\(\\mathbf{x}\\) is a surface point to be solved. In our implementation, we solve the linear least squares problem in Eqn. 1 of the form \\(\\min \\| Ax - b\\|^2\\). For a non-empty cell, this procedure can yield a reconstructed point \\(\\mathbf{v}\\), which the target surface \\(\\bar{S}\\) approximately crosses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.852 + ], + "angle": 0, + "content": "However, when \\( A \\) is nearly singular, the solution to Eqn. 1 can be located near the boundary of the cell or even outside of it, which would compromise the quality of extracted mesh surfaces. Note that while [10] proposed an approach with an additional regularization point to stabilize such cases, their approach relies on the intersection points at the zero-level set of an SDF with the cell edges. However, in our setting, the zero-level set of the UDF is elusive or unavailable, therefore we do not have such intersections. We propose an alternative approach based on singular value analysis of the matrix \\( A \\). We denote the three singular val" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.863, + 0.894, + 0.9 + ], + "angle": 0, + "content": "1While \\( \\| \\nabla F\\| \\) should be 1 for a distance function, note that since \\( \\nabla F \\) is drawn from a neural representation, it may differ slightly, and thus need normalization." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22534" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "ues of the matrix \\(A\\) as \\(\\sigma_0\\geq \\sigma_1\\geq \\sigma_2\\geq 0\\) and consider three cases:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.469, + 0.168 + ], + "angle": 0, + "content": "1) In the non-singular case, where all three singular values are much larger than 0, \\(\\mathbf{v}\\) corresponds to a sharp feature point that can be solved directly from Eqn. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.469, + 0.274 + ], + "angle": 0, + "content": "2) If only \\(\\sigma_{2} \\approx 0\\), the solution space of \\(\\mathbf{v}\\) corresponds to a linear edge within the cell. This line has the same direction with the singular vector corresponding to \\(\\sigma_{2}\\) and passes the solution of Eqn. 1. We then compute the intersections between this line and all 6 faces of the cell, yielding two intersecting points. We set \\(\\mathbf{v}\\) as the midpoint of these two points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.469, + 0.382 + ], + "angle": 0, + "content": "3) In the case where both \\(\\sigma_{1,2} \\approx 0\\), the part of the target surface enclosed in this cell is approximately planar. Similar to the edge case, we can formulate a plane function from the corresponding singular vectors. By computing the intersecting points between all 12 edges of the cell and this plane, we obtain multiple intersecting points. We use the centroid of these points as \\(\\mathbf{v}\\)." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.123, + 0.469, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.469, + 0.428 + ], + "angle": 0, + "content": "This way, in the degenerate cases (2 and 3), the point \\(\\mathbf{v}\\) is still a good approximate solution to Eqn. 1 and well positioned inside the cell." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.469, + 0.581 + ], + "angle": 0, + "content": "In a neural UDF (e.g., Fig. 2a), the UDF values \\(\\{d_i\\}\\) and the gradient directions \\(\\{\\mathbf{n}_i\\}\\) are approximated and often unreliable. Consequently, the tangent plane computed from each sample point \\(\\mathbf{p}_i\\) may not exactly align with the target surface. Solving Eqn. 1 yields a least squares solution that is numerically robust to the approximation errors in neural UDF and can lead to an accurate estimation of the surface point in this non-empty cell. For an ideal UDF case, our method is also applicable to yield an accurate result as shown in Fig. 2b." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.469, + 0.672 + ], + "angle": 0, + "content": "Furthermore, to enhance the robustness of our method, we also design a point filtering strategy to remove unreliable sample points, especially those near the target surface and its cut locus, where larger approximation errors exist in the MLP-encoded UDF. The details of this filtering strategy are explained in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Differences to DC [10]. Note that the dual contouring (DC) method [10] designs a QEF using intersection points between the surface and cell edges, along with the gradient directions at these intersection points. This approach is not applicable to our objective of extracting surface meshes from a neural UDF. This is because (1) it is difficult to find a reliable intersection point, as an exact zero-level set that crosses an edge may not exist, and (2) the gradient directions in the region with lower UDF values (i.e., closer to the target surface) are highly unstable and thus unreliable for estimating the tangent plane. Hence, our method leverages off-surface points that are sufficiently far from the target surface, as indicated by their UDF values (see Sec. 4.3). Our approach thus has less stringent requirements than the DC method (as presented in [10]), yet produces satisfactory" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.218, + 0.875, + 0.233 + ], + "angle": 0, + "content": "Figure 3: The pipeline of our mesh extraction method." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.264, + 0.892, + 0.293 + ], + "angle": 0, + "content": "results even on noisy neural UDFs as demonstrated by our results." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.309, + 0.771, + 0.325 + ], + "angle": 0, + "content": "4.2. Octree Design and Subdivision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.44 + ], + "angle": 0, + "content": "To efficiently process high-resolution data and speed up the mesh extraction process, we employ an octree data structure. As shown in Fig. 3, we recursively subdivide the cells unless they are categorized as empty according to the following criterion or they reach the maximal depth. After the subdivision, all leaf nodes that are non-empty will invoke the QEF procedure to solve for the dual points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Checking obviously empty cells. Firstly, we design a checking condition to help quickly prune obviously empty cells. Given an octree cell \\( C \\), let \\( \\mathbf{c}_0 \\) denote its center and \\( d_0 \\) denote the UDF value at \\( \\mathbf{c}_0 \\), that is, \\( d_0 = F(\\mathbf{c}_0) \\). Then we have the following sufficient condition for cell \\( C \\) to be empty:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.538, + 0.891, + 0.555 + ], + "angle": 0, + "content": "\\[\nd _ {0} > \\operatorname {D i a g} (C) / 2, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.763 + ], + "angle": 0, + "content": "where \\(\\operatorname{Diag}(C)\\) is the diagonal length of cell \\(C\\). Since the UDF distance indicates the distance between a spatial point and the target surface, this criterion determines if the target surface lies outside the sphere centered at \\(\\mathbf{c}_0\\) with radius \\(\\operatorname{Diag}(C)/2\\). Given that this sphere contains the entire octree cell, the cell will not contain any portion of the surface if Eqn. 2 is satisfied. Otherwise, the cell will first be categorized as an unsolved cell. Since neural UDF is a reasonable approximation of the ideal UDF, we tailor Eqn. 2 to \\(d_0 > \\operatorname{Diag}(C)/2 + \\epsilon\\), where \\(\\epsilon > 0\\) is a tolerance for the approximation error and set to \\(\\epsilon = 2 \\times 10^{-3}\\) throughout all experiments. Empirically, we observed this pruning strategy did not affect the quality of the results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Maximum octree depth. Having a predefined maximal depth of the octree is critical, not only in terms of limiting the amount of computation but also because of the inaccuracy of the neural UDF near the target surface. Specifically, if the octree keeps subdividing the space that contains the target surface, after a certain depth the cells will become too small and the sample points \\(\\mathbf{p}\\) within these cells will lie in the unreliable region of the neural UDF, which will eventually lead to unstable estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22535" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.35, + 0.108 + ], + "angle": 0, + "content": "4.3. Point filtering for neural UDFs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.471, + 0.249 + ], + "angle": 0, + "content": "Since our method makes use of the gradient directions and the distance values to estimate a surface point in each non-empty grid cell (see Eqn. 1), the quality of the extracted mesh heavily depends on the approximation accuracy of the neural UDF \\( F(\\mathbf{x}) \\). Consider a neighborhood of a sharp feature of the target surface, as shown in the zoom-in views of Fig. 1(e,f). The distance errors and gradient direction errors are more significant around the target surface and its cut locus." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.326 + ], + "angle": 0, + "content": "Hence, motivated by our observation and analysis in Sec. 3, to avoid building the QEF using points from these regions, we use the following criteria to filter out unreliable points to enhance the robustness of the QEF solution to the noisy characteristics of the MLP-encoded UDFs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.47, + 0.416 + ], + "angle": 0, + "content": "Criterion 1: Removing sample points potentially near the surface. A candidate point \\(\\mathbf{p}_i\\) is considered too close to the target surface if \\(F(\\mathbf{p}_i) < \\delta_1\\), where \\(\\delta_1\\) is a preset filtering threshold; such points are discarded. This criterion ensures the sample points are from a region where the distance errors are expected to be relatively small." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.471, + 0.627 + ], + "angle": 0, + "content": "Criterion 2: Removing sample points whose projections have large UDF values. While the previous criterion rejects the sample points that are too close to the target surface and thus avoids the erroneous approximation at those regions, some sample points may still have larger UDF errors even far from the target surface (e.g. near the cut locus). We observe that these sample points will not be projected to regions near the target surface. Hence, if the UDF of the projected point \\(\\{\\mathbf{q}_i\\}\\) is large, we consider it to be an incorrect estimate and thus reject the corresponding sample point \\(\\{\\mathbf{p}_i\\}\\) as unreliable. To this end, we introduce the second filtering criterion as follows, \\(F(\\mathbf{q}_i) > \\delta_2\\), where \\(\\delta_2\\) is another preset filtering bound; again, the corresponding points \\(\\mathbf{p}_i\\) are discarded." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.359, + 0.653 + ], + "angle": 0, + "content": "4.4. Creating Output Mesh Surfaces" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We consider the regular grid of cubic cells at the maximum depth of the octree. Then each edge of the grid is shared by four cubic cells surrounding the edge. To build the initial mesh connectivity, for each edge of the grid, we examine each of its four incident cells. Similar to [10], the mesh connection rule is designed as follows: if all four incident cells are non-empty, a quad-face candidate that connects the four dual points in these cells will be constructed. To ensure the correct connectivity, we validate and triangulate the quad faces. We further examine if the normals of the face candidates can consistently reflect their geometric property, being a sharp corner, part of a sharp edge, or part of a plane as classified by the SVD shape analysis (Sec. 4.1). For example, all correct triangulated faces incident to a surface point classified as part of a plane would have all their normals parallel to the singular vector corresponding to the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "largest singular value from SVD shape analysis, while the normals of triangulated faces incident to a surface point being part of a sharp edge should be orthogonal to its singular vector direction. This way, we reject all inconsistent face candidates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.892, + 0.35 + ], + "angle": 0, + "content": "We also provide a practical approach to make sure the output mesh is manifold when the desired target surface is manifold: With the help of the octree cells and the definition of reconstructed dual points, it is feasible to generate an auxiliary blocky model by moving all mesh vertices to the centroids of the corresponding cells. By extracting the outer envelope of this auxiliary model, we obtain a manifold structure. Finally, we tessellate the surface points using the connectivity from that manifold structure to produce a manifold mesh. This approach is crucial for applications that require manifold meshes (e.g. parameterization, remeshing, or shape analysis)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.366, + 0.634, + 0.383 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.791, + 0.408 + ], + "angle": 0, + "content": "5.1. Experimental details and metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Experimental setting. We rescaled all the shapes to a bounding cube with a side length of 2, centered at the origin. To improve efficiency, we share sample points between cells, i.e. \\( m = 27 \\) points in each non-empty leaf cell, including 8 corner points, 12 edge midpoints, 6 face midpoints, and 1 centroid point. We set the filtering parameters \\( \\delta_{1} = 2 \\times 10^{-3} \\) and \\( \\delta_{2} = 2 \\times 10^{-3} \\) to filter the sampled points as described in Sec. 4.3. We notice that more sample points may bring marginal performance improvement but incur computational overhead in our ablation study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Given that the grid resolution of \\(128^{3}\\) and \\(256^{3}\\) are commonly used in related works, we compare our method with these prior techniques using a maximal octree resolution of \\(128^{3}\\) and \\(256^{3}\\). We also discuss how the resolution will affect our results by varying the maximum depth of the octree to have a resolution from \\(64^{3}\\) to \\(512^{3}\\) in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "MLP architecture. To overfit single shapes using individual MLPs, we employed the MLP implementation provided by [18]. The activation functions are the Sine activation, except for the last one which is a SoftPlus \\((\\beta = 100)\\) activation to ensure the output value is non-negative. All neural UDFs in the experiments were trained with the described MLP implementation. All MLP networks were trained for \\(3k\\) iterations with the ADAM optimizer to minimize the difference between the predicted and the GT unsigned distance fields. For more general tasks, we also test an MLP with latent codes that represent a shape space, following the network and training settings in [5]. We report the timing performance on a Linux desktop with an Intel Core™ i7-10870H CPU and an NVIDIA GeForce RTX 3090 graphics card. We elaborate on the loss function as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22536" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.874, + 0.848 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.857, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Figure 4: Meshes extracted from neural UDFs. We show 6 examples and compare our results to the mesh extraction results in MeshUDF [7], CAP-UDF [21] and NDC [3]. Our method preserves sharp geometric features better, yields results without undesirable holes, and reconstructs the original smooth boundaries of open surfaces." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22537" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.152 + ], + "angle": 0, + "content": "Table 1: Quantitative comparison between the results obtained by our method and those by the competing methods, MeshUDF, CAP-UDF, and NDC. The average performance on the dataset containing 354 shapes is reported. The Chamfer distance (CD) and the Hausdorff Distance (HD) are scaled by \\(10^{-4}\\) and \\(10^{-3}\\), respectively. T1 and T2 stand for the time spent (seconds) on the mesh extraction and that on the UDF query, respectively." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.161, + 0.851, + 0.288 + ], + "angle": 0, + "content": "
MGNThingi10KABCRunning time
CD↓F-score↑HD↓CD↓F-score↑HD↓CD↓F-score↑HD↓T1T2
1283Ours2.3898.0911.911.9797.519.213.6993.4111.470.2971.184
MeshUDF [7]4.7690.0622.394.5790.7815.338.4987.7423.110.3131.316
CAP-UDF [21]13.5387.1060.0419.5584.8557.4927.7577.4353.466.4664.764
NDC [3]3.3295.6112.042.7496.449.693.9592.6514.042.3581.324
2563Ours2.0398.967.351.6098.296.782.2296.578.201.4925.508
MeshUDF [7]3.4595.1012.83.3394.138.913.2593.8511.491.60010.27
CAP-UDF [21]8.4592.8849.4012.1790.5347.2821.6783.8843.3349.41637.336
NDC [3]2.5498.047.842.4297.167.832.3696.1011.389.36310.256
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.31, + 0.472, + 0.4 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison between our method and three competing methods, MeshUDF, CAP-UDF, and NDC on a pre-trained UDF network with latent codes, on the ShapeNet dataset. In this table, CD and HD are scaled by \\(10^{-3}\\) and \\(10^{-2}\\) respectively. The F-score is calculated with a threshold of 0.006." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.412, + 0.427, + 0.492 + ], + "angle": 0, + "content": "
CD↓F-score↑HD ↓
Ours3.4182.303.92
MeshUDF [7]4.1079.354.67
CAP-UDF [21]7.7070.307.21
NDC [3]7.1678.3812.38
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.548 + ], + "angle": 0, + "content": "well as other details regarding training the neural UDFs in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.471, + 0.732 + ], + "angle": 0, + "content": "Metrics. To evaluate the performance of our method and the other methods, we adopt the following metrics, i.e. the double-sided Chamfer distance (CD), the F-score based on CD, and the Hausdorff distance (HD). The Chamfer distance reflects the overall quality of the extracted surface mesh as compared to the GT shape. The F-score indicates the percentage of points that are reconstructed correctly under a threshold (set to 0.001 for shape-overfitting neural UDFs, and 0.01 for shared UDF network with latent codes). The Hausdorff distance can reveal if there is anything missing or redundant in the reconstructed geometry (e.g. a hole or a floating piece)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.368, + 0.756 + ], + "angle": 0, + "content": "5.2. Comparison with SOTA methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We compared our results with three existing methods: 1) MeshUDF [7], 2) the standalone mesh extraction module presented in CAP-UDF [21], and 3) the UNDC presented in NDC [3]. The first two methods use gradients of the UDF to estimate sign changes in the field to invoke the Marching Cubes method to extract a surface mesh from the UDF. NDC proposes a data-driven approach to extract a surface mesh from the grid-based representation of an implicit field. Shape-overfitting neural UDFs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.477 + ], + "angle": 0, + "content": "We first compare our results to those obtained by these three methods on a shape collection consisting of shapes from four public sources: 1) 100 from the Thingi10K Dataset [22] containing 3D printing models; 2) 134 from the MGN Dataset [2] containing garments with open boundaries; 3) 100 from the ABC Dataset [12] containing CAD models; and 4) 20 commonly used shapes in geometric processing research. In addition, we also compare different methods on a Möbius strip, which is a non-orientable surface. We fit each shape with an independent neural UDF and apply mesh extraction to each of these neural UDFs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.895, + 0.658 + ], + "angle": 0, + "content": "Table 1 reports the performance of different methods on this shape collection. Our method outperforms the other three competing methods in terms of all three quantitative metrics. While all of these methods use uniform grids, our octree structure results in increased computing efficiency. MeshUDF and CAP-UDF would likely be accelerated by adopting an octree structure, however, it would be nontrivial to adopt an octree structure for NDC. For a fair comparison, we also test our method without using the octree acceleration approach for the resolution of \\(128^{3}\\); by doing so, the T1 time increases to 2.70s, and the T2 time increases to 5.54s." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.87 + ], + "angle": 0, + "content": "To qualitatively compare results obtained by different methods, we show the mesh surfaces extracted from neural UDFs for several shapes in Fig. 4. We can see that our results are higher quality than those produced by MeshUDF and CAP-UDF. These two methods cannot preserve sharp geometric features due to the use of the Marching Cubes method. Compared to NDC - a data-driven method - our method also produces consistently better results both quantitatively and qualitatively. Specifically, some mesh surfaces extracted by NDC are less smooth than ours. One example (the Pants) is shown in the fourth row in Fig. 4. The staircase artifact observed may be attributed to NDC having been trained on the ABC dataset [12] containing only mechanical components." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Our method is the only method that recovers open boundaries faithfully, while the other compared methods" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22538" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.138 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison between two different sampling number settings (27 and 125 per cell respectively). The Chamfer distance (CD) and the Hausdorff Distance (HD) are scaled by \\(10^{-4}\\) and \\(10^{-3}\\), respectively. T1 and T2 stand for the time spent (seconds) on the mesh extraction and that on the UDF query, respectively." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.146, + 0.851, + 0.208 + ], + "angle": 0, + "content": "
MGNThingi10KABCRunning time
CD↓F-score↑HD↓CD↓F-score↑HD↓CD↓F-score↑HD↓T1T2
272.3898.0911.911.9797.519.213.6993.4111.470.2971.184
1252.3498.1111.771.9197.529.023.7092.8911.120.9657.608
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.473, + 0.367 + ], + "angle": 0, + "content": "show staircase artifacts or even redundant pieces near the open boundaries; see the first two rows of Fig. 4. Also, unexpected holes can be observed on the mesh surface extracted by the competing methods from the neural UDFs. Empirically, our method produces quality results without redundant pieces or unexpected holes, generating clear-cut boundaries of the open surfaces as shown in the top four rows of Fig. 4, and reproducing the sharp features as shown in the bottom two rows of the same figure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.472, + 0.413 + ], + "angle": 0, + "content": "More results and comparisons (also on the GT UDF) are presented in our Supplementary Materials, which further validate the superior performance of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.472, + 0.581 + ], + "angle": 0, + "content": "Shared, pre-trained UDF network with latent codes. We consider a single neural network with latent codes trained to represent the entire shape space, where each shape is associated with a unique latent code. Specifically, we test our method and the competing methods on the pre-trained UDF network provided by [5] representing 300 ShapeNet car models. Although this UDF network is less accurate than the other overfitting-based network settings considered earlier, our method still outperforms the other methods by a significant margin as reported in Tab. 2, showing the versatility and robustness of our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.228, + 0.605 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.704 + ], + "angle": 0, + "content": "We conducted an ablation study on our method using sampling numbers (the number \\( m \\) in Eqn. 1) of \\( 27(3^{3}) \\) and \\( 125(5^{3}) \\) to justify our design choice to use 27 as our sampling number. In Tab. 3, we demonstrate that using 125 sample points per cell results in a marginal improvement but incurs a significantly higher computing cost." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.205, + 0.726 + ], + "angle": 0, + "content": "5.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "While DualMesh-UDF demonstrates the ability to accurately extract surfaces from neural UDFs, some limitations remain. First, the extracted surface cannot be adaptively subdivided with respect to sharp features or fine geometry details. To faithfully reconstruct the fine geometry details, we need to pre-define a sufficient depth of the octree and solve the QEF problem in each of its non-empty leaf nodes. Second, to share the UDF and gradient values at sampled points between adjacent octree cells, we adopt a regular grid sampling strategy. A more flexible and adaptive sampling strategy may bring further improvement to our method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.23, + 0.62, + 0.246 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.406 + ], + "angle": 0, + "content": "We have presented a method for extracting high-quality surface meshes from unsigned distance fields. In order to attain robust performance on MLP-encoded neural UDFs, we discuss the characteristics of the approximation errors of the neural UDF and develop an adaptive octree-based method to effectively localize the target surface embedded in the given UDF. Extensive experiments show that our method outperforms the SOTA methods and produces high-quality results with sharp geometry features, with clear open boundaries, and free of undesirable holes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.421, + 0.681, + 0.438 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.894, + 0.524 + ], + "angle": 0, + "content": "This research is partly supported by the Innovation and Technology Commission of the HKSAR Government through the InnoHK initiative. The authors would like to thank Yanhong Lin and Ruixing Jia for their generous help on the project." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.537, + 0.597, + 0.552 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.561, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[1] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.605, + 0.895, + 0.673 + ], + "angle": 0, + "content": "[2] Bharat Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5419-5429, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.675, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[3] Zhiqin Chen, Andrea Tagliasacchi, Thomas Funkhouser, and Hao Zhang. Neural dual contouring. ACM Trans. Graph., 41(4), jul 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.718, + 0.894, + 0.786 + ], + "angle": 0, + "content": "[4] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6968-6979, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.789, + 0.895, + 0.872 + ], + "angle": 0, + "content": "[5] Julian Chibane, Mohamad Aymen mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21638-21652. Curran Associates, Inc., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.873, + 0.895, + 0.904 + ], + "angle": 0, + "content": "[6] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learn" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.561, + 0.895, + 0.904 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22539" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "ing shapes. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 3789-3799. PMLR, 13-18 Jul 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[7] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 576-592, Cham, 2022. Springer Nature Switzerland." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.235, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[8] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, and Thomas Funkhouser. Local implicit grid representations for 3d scenes. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6000-6009, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[9] M.W. Jones, J.A. Baerentzen, and M. Sramek. 3d distance fields: a survey of techniques and applications. IEEE Transactions on Visualization and Computer Graphics, 12(4):581-599, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[10] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. ACM Trans. Graph., 21(3):339-346, jul 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.406, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[11] Leif P. Kobbelt, Mario Botsch, Ulrich Schwanecke, and Hans-Peter Seidel. Feature sensitive surface extraction from volume data. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '01, page 57-66, New York, NY, USA, 2001. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[12] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[13] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '87, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.468, + 0.715 + ], + "angle": 0, + "content": "[14] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. ACM Trans. Graph., 40(4), jul 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[15] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4455-4465, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[16] Mateusz Michalkiewicz, Jhony Kaesemodel Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4742-4751, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[17] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[18] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7462-7473. Curran Associates, Inc., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[19] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11353-11362, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.4 + ], + "angle": 0, + "content": "[20] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7537-7547. Curran Associates, Inc., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.402, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[21] Junsheng Zhou, Baorui Ma, Yu-Shen Liu, Yi Fang, and Zhizhong Han. Learning consistency-aware unsigned distance functions progressively from raw point clouds. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.487, + 0.892, + 0.527 + ], + "angle": 0, + "content": "[22] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22540" + } + ] +] \ No newline at end of file diff --git a/2023/Surface Extraction from Neural Unsigned Distance Fields/a0829e7a-732d-4761-bffa-ad35624b76ef_origin.pdf b/2023/Surface Extraction from Neural Unsigned Distance Fields/a0829e7a-732d-4761-bffa-ad35624b76ef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..53160e1325b431091495c13a2fedc663015569f1 --- /dev/null +++ b/2023/Surface Extraction from Neural Unsigned Distance Fields/a0829e7a-732d-4761-bffa-ad35624b76ef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:280b177e822adcc9ff750de20d33f7ec999bb5ede99af75c2e0d56e3d23958d9 +size 3804790 diff --git a/2023/Surface Extraction from Neural Unsigned Distance Fields/full.md b/2023/Surface Extraction from Neural Unsigned Distance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5f53c9f6b9422c4d134745de13fd3803dbed1a4c --- /dev/null +++ b/2023/Surface Extraction from Neural Unsigned Distance Fields/full.md @@ -0,0 +1,263 @@ +# Surface Extraction from Neural Unsigned Distance Fields + +Congyi Zhang $^{1,2*}$ Guying Lin $^{1*}$ Lei Yang $^{1,2}$ Xin Li $^{3}$ Taku Komura $^{1}$ Scott Schaefer $^{3}$ John Keyser $^{3}$ Wenping Wang $^{3\dagger}$ + +1The University of Hong Kong 2TransGP +3Texas A&M University + +# Abstract + +We propose a method, named DualMesh-UDF, to extract a surface from unsigned distance functions (UDFs), encoded by neural networks, or neural UDFs. Neural UDFs are becoming increasingly popular for surface representation because of their versatility in presenting surfaces with arbitrary topologies, as opposed to the signed distance function that is limited to representing a closed surface. However, the applications of neural UDFs are hindered by the notorious difficulty in extracting the target surfaces they represent. Recent methods for surface extraction from a neural UDF suffer from significant geometric errors or topological artifacts due to two main difficulties: (1) A UDF does not exhibit sign changes; and (2) A neural UDF typically has substantial approximation errors. + +DualMesh-UDF addresses these two difficulties. Specifically, given a neural UDF encoding a target surface $\bar{S}$ to be recovered, we first estimate the tangent planes of $\bar{S}$ at a set of sample points close to $\bar{S}$ . Next, we organize these sample points into local clusters, and for each local cluster, solve a linear least squares problem to determine a final surface point. These surface points are then connected to create the output mesh surface, which approximates the target surface. The robust estimation of the tangent planes of the target surface and the subsequent minimization problem constitute our core strategy, which contributes to the favorable performance of DualMesh-UDF over other competing methods. To efficiently implement this strategy, we employ an adaptive Octree. Within this framework, we estimate the location of a surface point in each of the octree cells identified as containing part of the target surface. Extensive experiments show that our method outperforms existing methods in terms of surface reconstruction quality while maintaining comparable computational efficiency. + +# 1. Introduction + +Implicit surfaces are widely used for surface representation in computer vision and computer graphics. An implicit surface is usually defined as a level set of a function, such as the zero-level set of a signed distance function (SDF). Extracting a mesh representation of an implicit surface from its defining equation is therefore a critical task for surface visualization and processing. Recent advances in machine learning have given rise to a new kind of implicit surface, called a neural implicit surface. A neural implicit surface is a level-set of a function encoded by an MLP (multilayer perceptron) and has the advantage of compactness and inherent smoothness thanks to its MLP representation. SDFs or occupancy fields are widely used in these implicit representations [17, 15, 16, 4, 18, 20, 6, 1, 8, 14, 19]. + +However, neural implicit surfaces based on the SDF or occupancy fields require inside/outside labeling and thus can only represent orientable and closed surfaces. Hence, as an extension, unsigned distance functions (UDFs) have been used to represent surfaces of arbitrary topologies, including open surfaces with boundaries or non-orientable surfaces (e.g. the Möbius strip). Despite its versatility, applications of a UDF-based surface representation are severely hindered by the difficulty in extracting the target surface it represents, as shown in [5] and [7]. + +Problem formulation: Suppose that a surface $\bar{S}$ , called the target surface, is defined as the zero-level set of its unsigned distance function (UDF) $\bar{F}(\mathbf{p})$ . Then suppose that this UDF $\bar{F}(\mathbf{p})$ is approximated by a neural network with the resulting network-encoded UDF being referred to as the neural UDF, denoted by $F(\mathbf{p})$ . Given a neural UDF $F(\mathbf{p})$ , the surface extraction problem is to robustly extract a surface $\mathcal{S}$ from $F(\mathbf{p})$ such that $\mathcal{S}$ well approximates the target surface $\bar{S}$ . + +Challenges: The difficulty in surface extraction from a neural UDF arises from two aspects: (1) A UDF does not have zero-crossings (or sign changes) across the surface it represents. As a result, traditional mesh extraction methods + +that rely on zero-crossings (e.g. Marching Cubes [13, 11], Dual Contouring [10], and their variants) are not applicable to UDFs. (2) The MLP representation of a neural UDF tends to have significant approximation errors around the target surface (see detailed error characteristics of neural UDFs in Sec. 3). This makes it even more challenging to extract a high-quality approximation of the target surface. + +Several methods, MeshUDF [7], CAP-UDF [21] and Neural Dual Contouring (NDC) [3], have recently been proposed for extracting a mesh surface from a UDF. MeshUDF and CAP-UDF attempt to infer the gradients of a UDF on the grids and determine the sign changes of the estimated gradients, invoking the Marching Cubes method for mesh extraction. When applied to a neural UDF, the sign-change inference step of this method suffers from instability due to the non-negligible error introduced by the approximate MLP representation near the surface where the gradients of the ideal UDF are undefined. As a result, the extracted meshes are less accurate and often have topological errors (e.g. holes). The NDC method proposes a data-driven Dual Contouring approach to predict the position of mesh vertices and dual faces directly from the UDF data. When applied to a neural UDF, this method often produces meshes with considerable artifacts such as holes, zig-zags, etc. + +We develop a new strategy, consisting of novel sampling and efficient optimization techniques to address the difficulties in surface extraction from the neural UDF. Suppose the input is a neural UDF $F(\mathbf{p})$ encoding the target surface $\bar{S}$ to be recovered. Our strategy has two key steps: (1) computing approximate tangent planes of the target surface; and (2) local minimization for generating final surface points. + +In Step (1), we first generate sample points $\mathbf{p}_i$ around, but not too close to, the target surface $\bar{S}$ , because the UDF values and gradients at locations too close to $\bar{S}$ are relatively unreliable. Thus, $\mathbf{p}_i$ are called off-surface sample points. For each $\mathbf{p}_i$ , we use the UDF value $F(\mathbf{p}_i)$ and its gradient $\nabla F(\mathbf{p}_i)$ to project $\mathbf{p}_i$ towards the target surface to obtain point $\mathbf{q}_i = \mathbf{p}_i - F(\mathbf{p}_i)\mathbf{n}_i(\mathbf{p}_i)$ where $\mathbf{n}(\mathbf{p}_i) = \nabla F(\mathbf{p}_i) / \| \nabla F(\mathbf{p}_i)\|$ [5]. These points $\mathbf{q}_i$ are called projection points. Although the points $\mathbf{q}_i$ are very close to the target surface $\bar{S}$ , as we will show later, the noisy error in the neural UDF makes these points a poor approximation to the target surface. + +To further improve surface accuracy, for each projection point $\mathbf{q}_i$ we generate an estimated tangent plane $T_{i}$ of $\bar{S}$ such that $T_{i}$ passes through $\mathbf{q}_i$ and has the unit normal vector $\mathbf{n}_i$ . Note that the normal vector $\mathbf{n}_i$ of $T_{i}$ is set to be $\mathbf{n}(\mathbf{p}_i)$ rather than $\mathbf{n}(\mathbf{q}_i)$ since the former is a more reliable estimation. This is because the initial sample point $\mathbf{p}_i$ is not too close to $\bar{S}$ , so the gradient $\nabla F(\mathbf{p}_i)$ is less contaminated by the pronounced errors of the neural UDF close to $\bar{S}$ . + +In Step (2), the estimated tangent planes are organized into clusters, which may overlap. For each cluster of tan- + +gent planes $T_{i}$ , we solve a linear least squares problem to produce a final surface point $\mathbf{s}_i$ that minimizes the sum of its squared distances to the tangent planes $T_{i}$ . This minimization step based on tangent planes not only provides an accurate surface point but also allows us to faithfully reconstruct the sharp edges of the target surface. Finally, all the surface points $\mathbf{s}_i$ from all the clusters are connected to form the output mesh surface to approximate the target surface $\bar{S}$ . + +To efficiently implement the above strategy, our DualMesh-UDF method adopts an adaptive Octree structure to partition the space containing the target surface to regular cells. We developed efficient procedures to determine those cells that contain part of the target surface and perform the sampling and minimization procedures in each occupied cell. To connect the surface points to create the output mesh, we follow the Dual Contouring approach, connecting surface points residing in adjacent grid cells to create polygons dual to octree edges. + +Extensive experiments demonstrate that our DualMesh-UDF significantly outperforms existing methods in terms of surface reconstruction accuracy and sharp feature preservation. + +The main contribution of this work is a new algorithm to robustly and accurately extract a surface from a neural UDF. To overcome the inevitable approximation errors near the target surface and cut locus, we obtain robust estimation of surface tangent planes by leveraging off-surface sample points, use least square minimization to better predict the surface points, and achieve high-quality surface extraction results with sharp features better preserved compared to the state of the art. The code is available at https://github.com/cong-yi/DualMesh-UDF. + +# 2. Related Work + +The Marching Cubes method [13] and its variants have been established as the de facto standard of converting distance fields to boundary mesh representations. Using the gradient information of the distance function, Extended Marching Cubes [11] and Dual Contouring [10] are both capable of producing meshes with faithfully preserved sharp features (e.g., corners and edges). + +However, all of these methods require inside/outside labeling on the sampling grid, which is either a regular grid of cubes or an adaptive octree grid of cells, to determine whether any zero-level set surface of the signed distance field crosses a particular grid cell. This requirement for sign changes limits the application of these methods to only SDFs or its variants that have sign changes across the underlying surface. Hence, they cannot be directly used to extract a mesh surface from a UDF, which is now often used to represent arbitrary surfaces such as open surfaces or non-orientable surfaces [5]. + +Our method is similar to [10] only in the sense that we + +follow [10] to solve a quadratic error function (QEF) to estimate a surface point per grid cell. But unlike [10], where sign changes are available and the gradients at the zero-crossings are reliable, unsigned distance fields do not have sign changes and are non-differentiable at the zero level set. Additionally, we will show that the information around the zero-level set of a neural UDF is unreliable. Due to these two reasons, we alternatively make use of the spatial sample points that are off the target surface in the neural UDF to find reliably estimated tangent planes and formulate a QEF for estimating the final surface points. + +Three relevant methods, MeshUDF [7], CAP-UDF [21] and Neural Dual Contouring (NDC) [3], have recently been proposed to extract meshes from UDFs. MeshUDF [7] and CAP-UDF [21] use the gradient information of the UDF to assign different signs (+ or -) to grid points on different sides of the underlying surface, thus invoking Marching Cubes to extract the surface according to the sign labels. However, the quality of the estimated sign labels can be significantly affected by the accuracy of the neural UDF. As we will show later, the neural UDF becomes less accurate around the target surface it represents, contaminating the inferred sign labels and explaining the poor performance of these two methods. Their use of Marching Cubes is incapable of preserving sharp corners or edges, while the use of a regular grid incurs high memory overhead when a high grid resolution is needed to resolve shape details. In contrast, our DualMesh-UDF method preserves sharp features and uses an adaptive octree grid to reduce computational expense even with a high grid resolution. Furthermore, estimating the surface point location is done by solving a least square problem (QEF) which is more robust than the gradient-based sign-labeling strategy used in MeshUDF in terms of the topology of the resulting meshes. + +The NDC method [3] uses a data-driven approach to train a neural network that predicts vertex position per regular grid cell and the overall dual faces directly from a UDF. However, as a data-driven method, NDC's performance is largely influenced by how accurately the network is trained. We show in the extensive experiments that our explicit geometry-based method consistently outperforms NDC in terms of reconstruction accuracy, preserving sharp features, and preserving smooth boundaries in the original shapes. + +# 3. Error Characteristics of Neural UDF + +We first analyze the characteristics of the errors introduced by the MLP representation used to approximate a UDF, which explains why extracting a mesh surface from a neural UDF is difficult. This analysis also provides a foundation for justifying design choices in our method to overcome these challenging characteristics and achieve robust surface extraction. + +We begin with an ideal UDF $\bar{F} (\mathbf{x})$ that represents a target + +![](images/00d45700a31ea47de216c005386887ea9c56d3523ee59cc76476625a517b3573.jpg) +(a) + +![](images/fd8ef8b328b682963334d419c86c6fb01dcb07f7c1c320abc9a38c0fd1430253.jpg) +(b) + +![](images/f5c821f1752547cf0e458ffc709e97aebb268826a5c39066652d74795953ef5e.jpg) +(c) + +![](images/571f375db8d69f60fbb67690305687ec8cec126e822a96b130e973197dbcb7e5.jpg) +(d) + +![](images/cd8b8f23ab744ac4c0ef400d0bf30158946b909e777704829624c150d4084f32.jpg) +(e) +Figure 1: A 2D illustrative example. We observe that the neural UDF tends to have larger errors near the zero-level set $\bar{S}$ and its cut locus. (a) The target shape (the red solid curve), its cut locus (green dashed), and the induced GT UDF; (b) Approximation errors of the neural UDF to the GT UDF; (c) Gradient direction errors between the neural UDF and the GT UDF; For the local region outlined with the white box around the corner, we show the close-up views of the distance error and the gradient direction error of the region in (e) and (f), respectively; (d) The GT UDF and its approximation by a neural UDF showing that the neural UDF is positive and smooth at the zero-level set of the ideal UDF, but the ideal UDF is non-differentiable. + +![](images/70976584046c56c46c7c569dedf3de2be1c326b92592cd56e330b3128e955b6e.jpg) +(f) + +surface $\bar{S}$ defined by the zero-level set of $\bar{F}(\mathbf{x})$ , that is, $\bar{S} = \{\mathbf{x} | \bar{F}(\mathbf{x}) = 0\} \subset \mathbb{R}^3$ . Note that the UDF $\bar{F}(\mathbf{x})$ is nonnegative, so it does not have sign changes across the surface $\bar{S}$ . Furthermore, $\bar{F}(\mathbf{x})$ is not differentiable at the surface $\bar{S}$ or the surface's cut locus [9]. + +Now suppose that the ideal UDF $\bar{F}(\mathbf{x})$ is approximated by an MLP, denoted by $F(\mathbf{x})$ . The approximation errors of $F(\mathbf{x})$ to $\bar{F}(\mathbf{x})$ are significant in the narrow region around the surface and a narrow region around the cut locus of $\bar{S}$ , because the neural UDF $F(\mathbf{x})$ is inherently smooth and thus poorly approximates the ideal $\bar{F}(\mathbf{x})$ , which is non-smooth at the surface $\bar{S}$ and its cut locus. Due to these errors, in general, we have $F(\mathbf{x}) > 0$ for any $\mathbf{x} \in \bar{S}$ when the MLP is differentiable. + +For ease of visualization, we will use a 2D example (see Fig. 1) to illustrate the characteristic behaviors of the approximation errors of $F(\mathbf{x})$ , whose zero-level set defines an open curve (red in Fig. 1(a)). The characteristics of the approximation errors of $F(\mathbf{x})$ in 3D space are similar. Detailed visualization and analysis of the error behaviors in 3D space are provided in the supplementary materials. Note + +that the cut locus (Fig. 1(a)) of $\bar{S}$ touches the sharp features of $\bar{S}$ . Therefore the neural UDF $F(\mathbf{x})$ has even more significant errors in terms of both the distance value and the gradient direction around the sharp features (see the error maps of the distance values and the gradient directions in Fig. 1(e,f), respectively). Consequently, in 3D space, the resulting inaccurate distance values and unreliable gradient vectors of the neural UDF $F(\mathbf{x})$ make it hard to faithfully reconstruct the target surface, especially to preserve the sharp edges in the extracted surface. + +To recap, given a neural UDF $F(\mathbf{x})$ , there are mainly two reasons for the difficulty in extracting a surface $\bar{S}_M$ to approximate the target surface $\bar{S}$ defined by the ideal UDF $\bar{F}(\mathbf{x})$ : (1) the given neural UDF $F(\mathbf{x})$ is usually a poor approximation of the ideal UDF $\bar{F}(\mathbf{x})$ around the target surface $\bar{S}$ and its cut locus, where $\bar{F}(\mathbf{x})$ is non-differentiable; and (2) the neural UDF $F(\mathbf{x})$ does not, in general, have a zero-crossing around the target surface $\bar{S}$ , thus no well-defined surface is associated with it. These issues, faced by current deep neural networks, make it hard to estimate the location of the surface in a numerically stable manner and thus motivate us to develop the two filtering criteria described in Sec. 4.3. + +# 4. Method + +Our method consists of three major designs: (1) a quadratic error function (QEF) that solves for a surface point within a cell for mesh extraction (Sec. 4.1); (2) an adaptive octree data structure that produces high-resolution grid cells while reducing the number of QEF solves required (Sec. 4.2); and (3) a point filtering strategy (Sec. 4.3) for a neural UDF, whose distance values and gradient directions are considerably more noisy compared to the GT as discussed in Sec. 3. + +The proposed DualMesh-UDF pipeline has three main steps. Firstly, it employs an adaptive tree to partition the domain of the UDF containing the target surface $\bar{S}$ . Then, it detects non-empty cells (defined in Sec. 4.2) based on a two-step cell-shape intersection detection method, and solves one surface point per non-empty leaf cell in a least squares manner. For a neural network, two distance bounds are introduced to filter out sample points that may introduce noisy information and contaminate the localization of the target surface. Lastly, since there are no sign changes in the UDFs, we construct the mesh faces by checking each edge shared by four incident non-empty cells in the octree. + +# 4.1. QEF to locate surface points + +Given a cell that contains part of the target surface, we present a procedure to estimate a surface point in this nonempty cell, as illustrated in Fig. 2 for 2D demonstrations. For a neural UDF, we aim to compute a reconstructed point per non-empty cell. For a set of sample points $\{\mathbf{p}_i\}_{i=1}^m$ in + +![](images/a6322235f5ca0a395f8b313a48591d9a740331f89fa465344a519d0367fb6b12.jpg) +(a) A neural UDF case +Figure 2: The QEF formulation for the unsigned distance field. The orange lines represent the target surface to be reconstructed. The gray lines in (a) stand for the estimated tangent planes contributed by the corresponding sample points. The red point is the solution to the QEF problem. + +![](images/0a30c13cafd837e28f4f7ac3b0c596e4abc65b6588d03750e0661db5f48d35c3.jpg) +(b) An ideal UDF case + +the given cell, we query the neural UDF $F(\mathbf{x})$ to get the approximate distance value and the normalized gradient vector at $\mathbf{p}_i$ , i.e., $d_i = F(\mathbf{p}_i)$ and $\mathbf{n}_i = \nabla F(\mathbf{p}_i) / \| \nabla F(\mathbf{p}_i)\|$ , respectively (see Fig. 2a). + +Each of the points contributes to an estimated tangent plane of the target surface $\bar{S}$ : $\mathbf{n}_i\cdot (\mathbf{p}_i - \mathbf{x}) - d_i = 0$ . Considering the inaccurate nature of the neural UDF, we define the estimated surface point to be the point that is the closest to all these tangent planes and compute it by minimizing a quadratic error function (a least squares problem): + +$$ +\mathbf {v} = \arg \min _ {\mathbf {x}} \sum_ {i = 1} ^ {m} \left(\mathbf {n} _ {i} \cdot \left(\mathbf {p} _ {i} - \mathbf {x}\right) - d _ {i}\right) ^ {2}, \tag {1} +$$ + +where $m$ is the number of points $\{\mathbf{p}_i\}$ and $\mathbf{x}$ is a surface point to be solved. In our implementation, we solve the linear least squares problem in Eqn. 1 of the form $\min \| Ax - b\|^2$ . For a non-empty cell, this procedure can yield a reconstructed point $\mathbf{v}$ , which the target surface $\bar{S}$ approximately crosses. + +However, when $A$ is nearly singular, the solution to Eqn. 1 can be located near the boundary of the cell or even outside of it, which would compromise the quality of extracted mesh surfaces. Note that while [10] proposed an approach with an additional regularization point to stabilize such cases, their approach relies on the intersection points at the zero-level set of an SDF with the cell edges. However, in our setting, the zero-level set of the UDF is elusive or unavailable, therefore we do not have such intersections. We propose an alternative approach based on singular value analysis of the matrix $A$ . We denote the three singular val + +ues of the matrix $A$ as $\sigma_0\geq \sigma_1\geq \sigma_2\geq 0$ and consider three cases: + +1) In the non-singular case, where all three singular values are much larger than 0, $\mathbf{v}$ corresponds to a sharp feature point that can be solved directly from Eqn. 1. +2) If only $\sigma_{2} \approx 0$ , the solution space of $\mathbf{v}$ corresponds to a linear edge within the cell. This line has the same direction with the singular vector corresponding to $\sigma_{2}$ and passes the solution of Eqn. 1. We then compute the intersections between this line and all 6 faces of the cell, yielding two intersecting points. We set $\mathbf{v}$ as the midpoint of these two points. +3) In the case where both $\sigma_{1,2} \approx 0$ , the part of the target surface enclosed in this cell is approximately planar. Similar to the edge case, we can formulate a plane function from the corresponding singular vectors. By computing the intersecting points between all 12 edges of the cell and this plane, we obtain multiple intersecting points. We use the centroid of these points as $\mathbf{v}$ . + +This way, in the degenerate cases (2 and 3), the point $\mathbf{v}$ is still a good approximate solution to Eqn. 1 and well positioned inside the cell. + +In a neural UDF (e.g., Fig. 2a), the UDF values $\{d_i\}$ and the gradient directions $\{\mathbf{n}_i\}$ are approximated and often unreliable. Consequently, the tangent plane computed from each sample point $\mathbf{p}_i$ may not exactly align with the target surface. Solving Eqn. 1 yields a least squares solution that is numerically robust to the approximation errors in neural UDF and can lead to an accurate estimation of the surface point in this non-empty cell. For an ideal UDF case, our method is also applicable to yield an accurate result as shown in Fig. 2b. + +Furthermore, to enhance the robustness of our method, we also design a point filtering strategy to remove unreliable sample points, especially those near the target surface and its cut locus, where larger approximation errors exist in the MLP-encoded UDF. The details of this filtering strategy are explained in Sec. 4.3. + +Differences to DC [10]. Note that the dual contouring (DC) method [10] designs a QEF using intersection points between the surface and cell edges, along with the gradient directions at these intersection points. This approach is not applicable to our objective of extracting surface meshes from a neural UDF. This is because (1) it is difficult to find a reliable intersection point, as an exact zero-level set that crosses an edge may not exist, and (2) the gradient directions in the region with lower UDF values (i.e., closer to the target surface) are highly unstable and thus unreliable for estimating the tangent plane. Hence, our method leverages off-surface points that are sufficiently far from the target surface, as indicated by their UDF values (see Sec. 4.3). Our approach thus has less stringent requirements than the DC method (as presented in [10]), yet produces satisfactory + +![](images/6219206ecfbcdd1522f94ca054eae62d26b2ac41bdcd5b58425703e4f7cdc5bb.jpg) +Figure 3: The pipeline of our mesh extraction method. + +results even on noisy neural UDFs as demonstrated by our results. + +# 4.2. Octree Design and Subdivision + +To efficiently process high-resolution data and speed up the mesh extraction process, we employ an octree data structure. As shown in Fig. 3, we recursively subdivide the cells unless they are categorized as empty according to the following criterion or they reach the maximal depth. After the subdivision, all leaf nodes that are non-empty will invoke the QEF procedure to solve for the dual points. + +Checking obviously empty cells. Firstly, we design a checking condition to help quickly prune obviously empty cells. Given an octree cell $C$ , let $\mathbf{c}_0$ denote its center and $d_0$ denote the UDF value at $\mathbf{c}_0$ , that is, $d_0 = F(\mathbf{c}_0)$ . Then we have the following sufficient condition for cell $C$ to be empty: + +$$ +d _ {0} > \operatorname {D i a g} (C) / 2, \tag {2} +$$ + +where $\operatorname{Diag}(C)$ is the diagonal length of cell $C$ . Since the UDF distance indicates the distance between a spatial point and the target surface, this criterion determines if the target surface lies outside the sphere centered at $\mathbf{c}_0$ with radius $\operatorname{Diag}(C)/2$ . Given that this sphere contains the entire octree cell, the cell will not contain any portion of the surface if Eqn. 2 is satisfied. Otherwise, the cell will first be categorized as an unsolved cell. Since neural UDF is a reasonable approximation of the ideal UDF, we tailor Eqn. 2 to $d_0 > \operatorname{Diag}(C)/2 + \epsilon$ , where $\epsilon > 0$ is a tolerance for the approximation error and set to $\epsilon = 2 \times 10^{-3}$ throughout all experiments. Empirically, we observed this pruning strategy did not affect the quality of the results. + +Maximum octree depth. Having a predefined maximal depth of the octree is critical, not only in terms of limiting the amount of computation but also because of the inaccuracy of the neural UDF near the target surface. Specifically, if the octree keeps subdividing the space that contains the target surface, after a certain depth the cells will become too small and the sample points $\mathbf{p}$ within these cells will lie in the unreliable region of the neural UDF, which will eventually lead to unstable estimation. + +# 4.3. Point filtering for neural UDFs + +Since our method makes use of the gradient directions and the distance values to estimate a surface point in each non-empty grid cell (see Eqn. 1), the quality of the extracted mesh heavily depends on the approximation accuracy of the neural UDF $F(\mathbf{x})$ . Consider a neighborhood of a sharp feature of the target surface, as shown in the zoom-in views of Fig. 1(e,f). The distance errors and gradient direction errors are more significant around the target surface and its cut locus. + +Hence, motivated by our observation and analysis in Sec. 3, to avoid building the QEF using points from these regions, we use the following criteria to filter out unreliable points to enhance the robustness of the QEF solution to the noisy characteristics of the MLP-encoded UDFs. + +Criterion 1: Removing sample points potentially near the surface. A candidate point $\mathbf{p}_i$ is considered too close to the target surface if $F(\mathbf{p}_i) < \delta_1$ , where $\delta_1$ is a preset filtering threshold; such points are discarded. This criterion ensures the sample points are from a region where the distance errors are expected to be relatively small. + +Criterion 2: Removing sample points whose projections have large UDF values. While the previous criterion rejects the sample points that are too close to the target surface and thus avoids the erroneous approximation at those regions, some sample points may still have larger UDF errors even far from the target surface (e.g. near the cut locus). We observe that these sample points will not be projected to regions near the target surface. Hence, if the UDF of the projected point $\{\mathbf{q}_i\}$ is large, we consider it to be an incorrect estimate and thus reject the corresponding sample point $\{\mathbf{p}_i\}$ as unreliable. To this end, we introduce the second filtering criterion as follows, $F(\mathbf{q}_i) > \delta_2$ , where $\delta_2$ is another preset filtering bound; again, the corresponding points $\mathbf{p}_i$ are discarded. + +# 4.4. Creating Output Mesh Surfaces + +We consider the regular grid of cubic cells at the maximum depth of the octree. Then each edge of the grid is shared by four cubic cells surrounding the edge. To build the initial mesh connectivity, for each edge of the grid, we examine each of its four incident cells. Similar to [10], the mesh connection rule is designed as follows: if all four incident cells are non-empty, a quad-face candidate that connects the four dual points in these cells will be constructed. To ensure the correct connectivity, we validate and triangulate the quad faces. We further examine if the normals of the face candidates can consistently reflect their geometric property, being a sharp corner, part of a sharp edge, or part of a plane as classified by the SVD shape analysis (Sec. 4.1). For example, all correct triangulated faces incident to a surface point classified as part of a plane would have all their normals parallel to the singular vector corresponding to the + +largest singular value from SVD shape analysis, while the normals of triangulated faces incident to a surface point being part of a sharp edge should be orthogonal to its singular vector direction. This way, we reject all inconsistent face candidates. + +We also provide a practical approach to make sure the output mesh is manifold when the desired target surface is manifold: With the help of the octree cells and the definition of reconstructed dual points, it is feasible to generate an auxiliary blocky model by moving all mesh vertices to the centroids of the corresponding cells. By extracting the outer envelope of this auxiliary model, we obtain a manifold structure. Finally, we tessellate the surface points using the connectivity from that manifold structure to produce a manifold mesh. This approach is crucial for applications that require manifold meshes (e.g. parameterization, remeshing, or shape analysis). + +# 5. Experiments + +# 5.1. Experimental details and metrics + +Experimental setting. We rescaled all the shapes to a bounding cube with a side length of 2, centered at the origin. To improve efficiency, we share sample points between cells, i.e. $m = 27$ points in each non-empty leaf cell, including 8 corner points, 12 edge midpoints, 6 face midpoints, and 1 centroid point. We set the filtering parameters $\delta_{1} = 2 \times 10^{-3}$ and $\delta_{2} = 2 \times 10^{-3}$ to filter the sampled points as described in Sec. 4.3. We notice that more sample points may bring marginal performance improvement but incur computational overhead in our ablation study. + +Given that the grid resolution of $128^{3}$ and $256^{3}$ are commonly used in related works, we compare our method with these prior techniques using a maximal octree resolution of $128^{3}$ and $256^{3}$ . We also discuss how the resolution will affect our results by varying the maximum depth of the octree to have a resolution from $64^{3}$ to $512^{3}$ in our supplementary material. + +MLP architecture. To overfit single shapes using individual MLPs, we employed the MLP implementation provided by [18]. The activation functions are the Sine activation, except for the last one which is a SoftPlus $(\beta = 100)$ activation to ensure the output value is non-negative. All neural UDFs in the experiments were trained with the described MLP implementation. All MLP networks were trained for $3k$ iterations with the ADAM optimizer to minimize the difference between the predicted and the GT unsigned distance fields. For more general tasks, we also test an MLP with latent codes that represent a shape space, following the network and training settings in [5]. We report the timing performance on a Linux desktop with an Intel Core™ i7-10870H CPU and an NVIDIA GeForce RTX 3090 graphics card. We elaborate on the loss function as + +![](images/5b0ccebcaeb2fa41e52fe066cb3f788591bbd72644ac4a0147bfd3051ae304f5.jpg) +Figure 4: Meshes extracted from neural UDFs. We show 6 examples and compare our results to the mesh extraction results in MeshUDF [7], CAP-UDF [21] and NDC [3]. Our method preserves sharp geometric features better, yields results without undesirable holes, and reconstructs the original smooth boundaries of open surfaces. + +Table 1: Quantitative comparison between the results obtained by our method and those by the competing methods, MeshUDF, CAP-UDF, and NDC. The average performance on the dataset containing 354 shapes is reported. The Chamfer distance (CD) and the Hausdorff Distance (HD) are scaled by $10^{-4}$ and $10^{-3}$ , respectively. T1 and T2 stand for the time spent (seconds) on the mesh extraction and that on the UDF query, respectively. + +
MGNThingi10KABCRunning time
CD↓F-score↑HD↓CD↓F-score↑HD↓CD↓F-score↑HD↓T1T2
1283Ours2.3898.0911.911.9797.519.213.6993.4111.470.2971.184
MeshUDF [7]4.7690.0622.394.5790.7815.338.4987.7423.110.3131.316
CAP-UDF [21]13.5387.1060.0419.5584.8557.4927.7577.4353.466.4664.764
NDC [3]3.3295.6112.042.7496.449.693.9592.6514.042.3581.324
2563Ours2.0398.967.351.6098.296.782.2296.578.201.4925.508
MeshUDF [7]3.4595.1012.83.3394.138.913.2593.8511.491.60010.27
CAP-UDF [21]8.4592.8849.4012.1790.5347.2821.6783.8843.3349.41637.336
NDC [3]2.5498.047.842.4297.167.832.3696.1011.389.36310.256
+ +Table 2: Quantitative comparison between our method and three competing methods, MeshUDF, CAP-UDF, and NDC on a pre-trained UDF network with latent codes, on the ShapeNet dataset. In this table, CD and HD are scaled by $10^{-3}$ and $10^{-2}$ respectively. The F-score is calculated with a threshold of 0.006. + +
CD↓F-score↑HD ↓
Ours3.4182.303.92
MeshUDF [7]4.1079.354.67
CAP-UDF [21]7.7070.307.21
NDC [3]7.1678.3812.38
+ +well as other details regarding training the neural UDFs in the supplementary materials. + +Metrics. To evaluate the performance of our method and the other methods, we adopt the following metrics, i.e. the double-sided Chamfer distance (CD), the F-score based on CD, and the Hausdorff distance (HD). The Chamfer distance reflects the overall quality of the extracted surface mesh as compared to the GT shape. The F-score indicates the percentage of points that are reconstructed correctly under a threshold (set to 0.001 for shape-overfitting neural UDFs, and 0.01 for shared UDF network with latent codes). The Hausdorff distance can reveal if there is anything missing or redundant in the reconstructed geometry (e.g. a hole or a floating piece). + +# 5.2. Comparison with SOTA methods + +We compared our results with three existing methods: 1) MeshUDF [7], 2) the standalone mesh extraction module presented in CAP-UDF [21], and 3) the UNDC presented in NDC [3]. The first two methods use gradients of the UDF to estimate sign changes in the field to invoke the Marching Cubes method to extract a surface mesh from the UDF. NDC proposes a data-driven approach to extract a surface mesh from the grid-based representation of an implicit field. Shape-overfitting neural UDFs + +We first compare our results to those obtained by these three methods on a shape collection consisting of shapes from four public sources: 1) 100 from the Thingi10K Dataset [22] containing 3D printing models; 2) 134 from the MGN Dataset [2] containing garments with open boundaries; 3) 100 from the ABC Dataset [12] containing CAD models; and 4) 20 commonly used shapes in geometric processing research. In addition, we also compare different methods on a Möbius strip, which is a non-orientable surface. We fit each shape with an independent neural UDF and apply mesh extraction to each of these neural UDFs. + +Table 1 reports the performance of different methods on this shape collection. Our method outperforms the other three competing methods in terms of all three quantitative metrics. While all of these methods use uniform grids, our octree structure results in increased computing efficiency. MeshUDF and CAP-UDF would likely be accelerated by adopting an octree structure, however, it would be nontrivial to adopt an octree structure for NDC. For a fair comparison, we also test our method without using the octree acceleration approach for the resolution of $128^{3}$ ; by doing so, the T1 time increases to 2.70s, and the T2 time increases to 5.54s. + +To qualitatively compare results obtained by different methods, we show the mesh surfaces extracted from neural UDFs for several shapes in Fig. 4. We can see that our results are higher quality than those produced by MeshUDF and CAP-UDF. These two methods cannot preserve sharp geometric features due to the use of the Marching Cubes method. Compared to NDC - a data-driven method - our method also produces consistently better results both quantitatively and qualitatively. Specifically, some mesh surfaces extracted by NDC are less smooth than ours. One example (the Pants) is shown in the fourth row in Fig. 4. The staircase artifact observed may be attributed to NDC having been trained on the ABC dataset [12] containing only mechanical components. + +Our method is the only method that recovers open boundaries faithfully, while the other compared methods + +Table 3: Quantitative comparison between two different sampling number settings (27 and 125 per cell respectively). The Chamfer distance (CD) and the Hausdorff Distance (HD) are scaled by $10^{-4}$ and $10^{-3}$ , respectively. T1 and T2 stand for the time spent (seconds) on the mesh extraction and that on the UDF query, respectively. + +
MGNThingi10KABCRunning time
CD↓F-score↑HD↓CD↓F-score↑HD↓CD↓F-score↑HD↓T1T2
272.3898.0911.911.9797.519.213.6993.4111.470.2971.184
1252.3498.1111.771.9197.529.023.7092.8911.120.9657.608
+ +show staircase artifacts or even redundant pieces near the open boundaries; see the first two rows of Fig. 4. Also, unexpected holes can be observed on the mesh surface extracted by the competing methods from the neural UDFs. Empirically, our method produces quality results without redundant pieces or unexpected holes, generating clear-cut boundaries of the open surfaces as shown in the top four rows of Fig. 4, and reproducing the sharp features as shown in the bottom two rows of the same figure. + +More results and comparisons (also on the GT UDF) are presented in our Supplementary Materials, which further validate the superior performance of our method. + +Shared, pre-trained UDF network with latent codes. We consider a single neural network with latent codes trained to represent the entire shape space, where each shape is associated with a unique latent code. Specifically, we test our method and the competing methods on the pre-trained UDF network provided by [5] representing 300 ShapeNet car models. Although this UDF network is less accurate than the other overfitting-based network settings considered earlier, our method still outperforms the other methods by a significant margin as reported in Tab. 2, showing the versatility and robustness of our method. + +# 5.3. Ablation study + +We conducted an ablation study on our method using sampling numbers (the number $m$ in Eqn. 1) of $27(3^{3})$ and $125(5^{3})$ to justify our design choice to use 27 as our sampling number. In Tab. 3, we demonstrate that using 125 sample points per cell results in a marginal improvement but incurs a significantly higher computing cost. + +# 5.4. Limitations + +While DualMesh-UDF demonstrates the ability to accurately extract surfaces from neural UDFs, some limitations remain. First, the extracted surface cannot be adaptively subdivided with respect to sharp features or fine geometry details. To faithfully reconstruct the fine geometry details, we need to pre-define a sufficient depth of the octree and solve the QEF problem in each of its non-empty leaf nodes. Second, to share the UDF and gradient values at sampled points between adjacent octree cells, we adopt a regular grid sampling strategy. A more flexible and adaptive sampling strategy may bring further improvement to our method. + +# 6. Conclusion + +We have presented a method for extracting high-quality surface meshes from unsigned distance fields. In order to attain robust performance on MLP-encoded neural UDFs, we discuss the characteristics of the approximation errors of the neural UDF and develop an adaptive octree-based method to effectively localize the target surface embedded in the given UDF. Extensive experiments show that our method outperforms the SOTA methods and produces high-quality results with sharp geometry features, with clear open boundaries, and free of undesirable holes. + +# 7. Acknowledgement + +This research is partly supported by the Innovation and Technology Commission of the HKSAR Government through the InnoHK initiative. The authors would like to thank Yanhong Lin and Ruixing Jia for their generous help on the project. + +# References + +[1] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In International Conference on Learning Representations, 2021. +[2] Bharat Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5419-5429, 2019. +[3] Zhiqin Chen, Andrea Tagliasacchi, Thomas Funkhouser, and Hao Zhang. Neural dual contouring. ACM Trans. Graph., 41(4), jul 2022. +[4] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6968-6979, 2020. +[5] Julian Chibane, Mohamad Aymen mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21638-21652. Curran Associates, Inc., 2020. +[6] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learn + +ing shapes. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 3789-3799. PMLR, 13-18 Jul 2020. +[7] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 576-592, Cham, 2022. Springer Nature Switzerland. +[8] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, and Thomas Funkhouser. Local implicit grid representations for 3d scenes. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6000-6009, 2020. +[9] M.W. Jones, J.A. Baerentzen, and M. Sramek. 3d distance fields: a survey of techniques and applications. IEEE Transactions on Visualization and Computer Graphics, 12(4):581-599, 2006. +[10] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. ACM Trans. Graph., 21(3):339-346, jul 2002. +[11] Leif P. Kobbelt, Mario Botsch, Ulrich Schwanecke, and Hans-Peter Seidel. Feature sensitive surface extraction from volume data. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '01, page 57-66, New York, NY, USA, 2001. Association for Computing Machinery. +[12] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. +[13] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '87, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery. +[14] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. ACM Trans. Graph., 40(4), jul 2021. +[15] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4455-4465, 2019. +[16] Mateusz Michalkiewicz, Jhony Kaesemodel Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4742-4751, 2019. +[17] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. + +In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. +[18] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7462-7473. Curran Associates, Inc., 2020. +[19] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11353-11362, 2021. +[20] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7537-7547. Curran Associates, Inc., 2020. +[21] Junsheng Zhou, Baorui Ma, Yu-Shen Liu, Yi Fang, and Zhizhong Han. Learning consistency-aware unsigned distance functions progressively from raw point clouds. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. +[22] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016. \ No newline at end of file diff --git a/2023/Surface Extraction from Neural Unsigned Distance Fields/images.zip b/2023/Surface Extraction from Neural Unsigned Distance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b74c909dede083c5286348d4384c48171633ae47 --- /dev/null +++ b/2023/Surface Extraction from Neural Unsigned Distance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4dc563e9f322550fa6ec0455b7ebb3473589389c7dfa262090d5609e711d9519 +size 446225 diff --git a/2023/Surface Extraction from Neural Unsigned Distance Fields/layout.json b/2023/Surface Extraction from Neural Unsigned Distance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f31bef1420aacb0afb1cc76286f82d7258331e6 --- /dev/null +++ b/2023/Surface Extraction from Neural Unsigned Distance Fields/layout.json @@ -0,0 +1,8982 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 476, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 476, + 121 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 476, + 121 + ], + "type": "text", + "content": "Surface Extraction from Neural Unsigned Distance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "spans": [ + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "text", + "content": "Congyi Zhang" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "text", + "content": " Guying Lin" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "text", + "content": " Lei Yang" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "text", + "content": " Xin Li" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "text", + "content": " Taku Komura" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "text", + "content": " Scott Schaefer" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "text", + "content": " John Keyser" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "text", + "content": " Wenping Wang" + }, + { + "bbox": [ + 82, + 141, + 511, + 179 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 183, + 190, + 408, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 190, + 408, + 220 + ], + "spans": [ + { + "bbox": [ + 183, + 190, + 408, + 220 + ], + "type": "text", + "content": "1The University of Hong Kong 2TransGP \n3Texas A&M University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "spans": [ + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 273, + 289, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 273, + 289, + 441 + ], + "spans": [ + { + "bbox": [ + 45, + 273, + 289, + 441 + ], + "type": "text", + "content": "We propose a method, named DualMesh-UDF, to extract a surface from unsigned distance functions (UDFs), encoded by neural networks, or neural UDFs. Neural UDFs are becoming increasingly popular for surface representation because of their versatility in presenting surfaces with arbitrary topologies, as opposed to the signed distance function that is limited to representing a closed surface. However, the applications of neural UDFs are hindered by the notorious difficulty in extracting the target surfaces they represent. Recent methods for surface extraction from a neural UDF suffer from significant geometric errors or topological artifacts due to two main difficulties: (1) A UDF does not exhibit sign changes; and (2) A neural UDF typically has substantial approximation errors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 441, + 289, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 289, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 289, + 670 + ], + "type": "text", + "content": "DualMesh-UDF addresses these two difficulties. Specifically, given a neural UDF encoding a target surface " + }, + { + "bbox": [ + 46, + 441, + 289, + 670 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 441, + 289, + 670 + ], + "type": "text", + "content": " to be recovered, we first estimate the tangent planes of " + }, + { + "bbox": [ + 46, + 441, + 289, + 670 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 441, + 289, + 670 + ], + "type": "text", + "content": " at a set of sample points close to " + }, + { + "bbox": [ + 46, + 441, + 289, + 670 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 441, + 289, + 670 + ], + "type": "text", + "content": ". Next, we organize these sample points into local clusters, and for each local cluster, solve a linear least squares problem to determine a final surface point. These surface points are then connected to create the output mesh surface, which approximates the target surface. The robust estimation of the tangent planes of the target surface and the subsequent minimization problem constitute our core strategy, which contributes to the favorable performance of DualMesh-UDF over other competing methods. To efficiently implement this strategy, we employ an adaptive Octree. Within this framework, we estimate the location of a surface point in each of the octree cells identified as containing part of the target surface. Extensive experiments show that our method outperforms existing methods in terms of surface reconstruction quality while maintaining comparable computational efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 247, + 386, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 386, + 260 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 386, + 260 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 547, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 435 + ], + "type": "text", + "content": "Implicit surfaces are widely used for surface representation in computer vision and computer graphics. An implicit surface is usually defined as a level set of a function, such as the zero-level set of a signed distance function (SDF). Extracting a mesh representation of an implicit surface from its defining equation is therefore a critical task for surface visualization and processing. Recent advances in machine learning have given rise to a new kind of implicit surface, called a neural implicit surface. A neural implicit surface is a level-set of a function encoded by an MLP (multilayer perceptron) and has the advantage of compactness and inherent smoothness thanks to its MLP representation. SDFs or occupancy fields are widely used in these implicit representations [17, 15, 16, 4, 18, 20, 6, 1, 8, 14, 19]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 436, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 547, + 555 + ], + "type": "text", + "content": "However, neural implicit surfaces based on the SDF or occupancy fields require inside/outside labeling and thus can only represent orientable and closed surfaces. Hence, as an extension, unsigned distance functions (UDFs) have been used to represent surfaces of arbitrary topologies, including open surfaces with boundaries or non-orientable surfaces (e.g. the Möbius strip). Despite its versatility, applications of a UDF-based surface representation are severely hindered by the difficulty in extracting the target surface it represents, as shown in [5] and [7]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "content": "Problem formulation: Suppose that a surface " + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "content": ", called the target surface, is defined as the zero-level set of its unsigned distance function (UDF) " + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "content": ". Then suppose that this UDF " + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "content": " is approximated by a neural network with the resulting network-encoded UDF being referred to as the neural UDF, denoted by " + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "inline_equation", + "content": "F(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "content": ". Given a neural UDF " + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "inline_equation", + "content": "F(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "content": ", the surface extraction problem is to robustly extract a surface " + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "inline_equation", + "content": "F(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "content": " well approximates the target surface " + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 556, + 547, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Challenges: The difficulty in surface extraction from a neural UDF arises from two aspects: (1) A UDF does not have zero-crossings (or sign changes) across the surface it represents. As a result, traditional mesh extraction methods" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 124, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22531" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "that rely on zero-crossings (e.g. Marching Cubes [13, 11], Dual Contouring [10], and their variants) are not applicable to UDFs. (2) The MLP representation of a neural UDF tends to have significant approximation errors around the target surface (see detailed error characteristics of neural UDFs in Sec. 3). This makes it even more challenging to extract a high-quality approximation of the target surface." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 361 + ], + "type": "text", + "content": "Several methods, MeshUDF [7], CAP-UDF [21] and Neural Dual Contouring (NDC) [3], have recently been proposed for extracting a mesh surface from a UDF. MeshUDF and CAP-UDF attempt to infer the gradients of a UDF on the grids and determine the sign changes of the estimated gradients, invoking the Marching Cubes method for mesh extraction. When applied to a neural UDF, the sign-change inference step of this method suffers from instability due to the non-negligible error introduced by the approximate MLP representation near the surface where the gradients of the ideal UDF are undefined. As a result, the extracted meshes are less accurate and often have topological errors (e.g. holes). The NDC method proposes a data-driven Dual Contouring approach to predict the position of mesh vertices and dual faces directly from the UDF data. When applied to a neural UDF, this method often produces meshes with considerable artifacts such as holes, zig-zags, etc." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 362, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 287, + 445 + ], + "type": "text", + "content": "We develop a new strategy, consisting of novel sampling and efficient optimization techniques to address the difficulties in surface extraction from the neural UDF. Suppose the input is a neural UDF " + }, + { + "bbox": [ + 47, + 362, + 287, + 445 + ], + "type": "inline_equation", + "content": "F(\\mathbf{p})" + }, + { + "bbox": [ + 47, + 362, + 287, + 445 + ], + "type": "text", + "content": " encoding the target surface " + }, + { + "bbox": [ + 47, + 362, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 47, + 362, + 287, + 445 + ], + "type": "text", + "content": " to be recovered. Our strategy has two key steps: (1) computing approximate tangent planes of the target surface; and (2) local minimization for generating final surface points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": "In Step (1), we first generate sample points " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": " around, but not too close to, the target surface " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": ", because the UDF values and gradients at locations too close to " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": " are relatively unreliable. Thus, " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": " are called off-surface sample points. For each " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": ", we use the UDF value " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "F(\\mathbf{p}_i)" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": " and its gradient " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\nabla F(\\mathbf{p}_i)" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": " to project " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": " towards the target surface to obtain point " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i = \\mathbf{p}_i - F(\\mathbf{p}_i)\\mathbf{n}_i(\\mathbf{p}_i)" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(\\mathbf{p}_i) = \\nabla F(\\mathbf{p}_i) / \\| \\nabla F(\\mathbf{p}_i)\\|" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": " [5]. These points " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": " are called projection points. Although the points " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": " are very close to the target surface " + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 447, + 287, + 590 + ], + "type": "text", + "content": ", as we will show later, the noisy error in the neural UDF makes these points a poor approximation to the target surface." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": "To further improve surface accuracy, for each projection point " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " we generate an estimated tangent plane " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " passes through " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " and has the unit normal vector " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": ". Note that the normal vector " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " is set to be " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(\\mathbf{p}_i)" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " rather than " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(\\mathbf{q}_i)" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " since the former is a more reliable estimation. This is because the initial sample point " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " is not too close to " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": ", so the gradient " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\nabla F(\\mathbf{p}_i)" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": " is less contaminated by the pronounced errors of the neural UDF close to " + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 592, + 287, + 688 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In Step (2), the estimated tangent planes are organized into clusters, which may overlap. For each cluster of tan-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "gent planes " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": ", we solve a linear least squares problem to produce a final surface point " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " that minimizes the sum of its squared distances to the tangent planes " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": ". This minimization step based on tangent planes not only provides an accurate surface point but also allows us to faithfully reconstruct the sharp edges of the target surface. Finally, all the surface points " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " from all the clusters are connected to form the output mesh surface to approximate the target surface " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 168, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 287 + ], + "type": "text", + "content": "To efficiently implement the above strategy, our DualMesh-UDF method adopts an adaptive Octree structure to partition the space containing the target surface to regular cells. We developed efficient procedures to determine those cells that contain part of the target surface and perform the sampling and minimization procedures in each occupied cell. To connect the surface points to create the output mesh, we follow the Dual Contouring approach, connecting surface points residing in adjacent grid cells to create polygons dual to octree edges." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 288, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 334 + ], + "type": "text", + "content": "Extensive experiments demonstrate that our DualMesh-UDF significantly outperforms existing methods in terms of surface reconstruction accuracy and sharp feature preservation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 455 + ], + "type": "text", + "content": "The main contribution of this work is a new algorithm to robustly and accurately extract a surface from a neural UDF. To overcome the inevitable approximation errors near the target surface and cut locus, we obtain robust estimation of surface tangent planes by leveraging off-surface sample points, use least square minimization to better predict the surface points, and achieve high-quality surface extraction results with sharp features better preserved compared to the state of the art. The code is available at https://github.com/cong-yi/DualMesh-UDF." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 466, + 392, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 392, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 392, + 479 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 570 + ], + "type": "text", + "content": "The Marching Cubes method [13] and its variants have been established as the de facto standard of converting distance fields to boundary mesh representations. Using the gradient information of the distance function, Extended Marching Cubes [11] and Dual Contouring [10] are both capable of producing meshes with faithfully preserved sharp features (e.g., corners and edges)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": "However, all of these methods require inside/outside labeling on the sampling grid, which is either a regular grid of cubes or an adaptive octree grid of cells, to determine whether any zero-level set surface of the signed distance field crosses a particular grid cell. This requirement for sign changes limits the application of these methods to only SDFs or its variants that have sign changes across the underlying surface. Hence, they cannot be directly used to extract a mesh surface from a UDF, which is now often used to represent arbitrary surfaces such as open surfaces or non-orientable surfaces [5]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Our method is similar to [10] only in the sense that we" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22532" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "follow [10] to solve a quadratic error function (QEF) to estimate a surface point per grid cell. But unlike [10], where sign changes are available and the gradients at the zero-crossings are reliable, unsigned distance fields do not have sign changes and are non-differentiable at the zero level set. Additionally, we will show that the information around the zero-level set of a neural UDF is unreliable. Due to these two reasons, we alternatively make use of the spatial sample points that are off the target surface in the neural UDF to find reliably estimated tangent planes and formulate a QEF for estimating the final surface points." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 479 + ], + "type": "text", + "content": "Three relevant methods, MeshUDF [7], CAP-UDF [21] and Neural Dual Contouring (NDC) [3], have recently been proposed to extract meshes from UDFs. MeshUDF [7] and CAP-UDF [21] use the gradient information of the UDF to assign different signs (+ or -) to grid points on different sides of the underlying surface, thus invoking Marching Cubes to extract the surface according to the sign labels. However, the quality of the estimated sign labels can be significantly affected by the accuracy of the neural UDF. As we will show later, the neural UDF becomes less accurate around the target surface it represents, contaminating the inferred sign labels and explaining the poor performance of these two methods. Their use of Marching Cubes is incapable of preserving sharp corners or edges, while the use of a regular grid incurs high memory overhead when a high grid resolution is needed to resolve shape details. In contrast, our DualMesh-UDF method preserves sharp features and uses an adaptive octree grid to reduce computational expense even with a high grid resolution. Furthermore, estimating the surface point location is done by solving a least square problem (QEF) which is more robust than the gradient-based sign-labeling strategy used in MeshUDF in terms of the topology of the resulting meshes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 479, + 288, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 587 + ], + "type": "text", + "content": "The NDC method [3] uses a data-driven approach to train a neural network that predicts vertex position per regular grid cell and the overall dual faces directly from a UDF. However, as a data-driven method, NDC's performance is largely influenced by how accurately the network is trained. We show in the extensive experiments that our explicit geometry-based method consistently outperforms NDC in terms of reconstruction accuracy, preserving sharp features, and preserving smooth boundaries in the original shapes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 597, + 253, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 253, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 253, + 611 + ], + "type": "text", + "content": "3. Error Characteristics of Neural UDF" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 700 + ], + "type": "text", + "content": "We first analyze the characteristics of the errors introduced by the MLP representation used to approximate a UDF, which explains why extracting a mesh surface from a neural UDF is difficult. This analysis also provides a foundation for justifying design choices in our method to overcome these challenging characteristics and achieve robust surface extraction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 700, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 700, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 700, + 288, + 714 + ], + "type": "text", + "content": "We begin with an ideal UDF " + }, + { + "bbox": [ + 59, + 700, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\bar{F} (\\mathbf{x})" + }, + { + "bbox": [ + 59, + 700, + 288, + 714 + ], + "type": "text", + "content": " that represents a target" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 381, + 144 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 381, + 144 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 381, + 144 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 381, + 144 + ], + "type": "image", + "image_path": "00d45700a31ea47de216c005386887ea9c56d3523ee59cc76476625a517b3573.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 144, + 351, + 152 + ], + "lines": [ + { + "bbox": [ + 342, + 144, + 351, + 152 + ], + "spans": [ + { + "bbox": [ + 342, + 144, + 351, + 152 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 463, + 144 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 463, + 144 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 463, + 144 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 463, + 144 + ], + "type": "image", + "image_path": "fd8ef8b328b682963334d419c86c6fb01dcb07f7c1c320abc9a38c0fd1430253.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 144, + 430, + 152 + ], + "lines": [ + { + "bbox": [ + 421, + 144, + 430, + 152 + ], + "spans": [ + { + "bbox": [ + 421, + 144, + 430, + 152 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 470, + 70, + 544, + 144 + ], + "blocks": [ + { + "bbox": [ + 470, + 70, + 544, + 144 + ], + "lines": [ + { + "bbox": [ + 470, + 70, + 544, + 144 + ], + "spans": [ + { + "bbox": [ + 470, + 70, + 544, + 144 + ], + "type": "image", + "image_path": "f5c821f1752547cf0e458ffc709e97aebb268826a5c39066652d74795953ef5e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 144, + 512, + 152 + ], + "lines": [ + { + "bbox": [ + 503, + 144, + 512, + 152 + ], + "spans": [ + { + "bbox": [ + 503, + 144, + 512, + 152 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 171, + 382, + 225 + ], + "blocks": [ + { + "bbox": [ + 306, + 171, + 382, + 225 + ], + "lines": [ + { + "bbox": [ + 306, + 171, + 382, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 382, + 225 + ], + "type": "image", + "image_path": "571f375db8d69f60fbb67690305687ec8cec126e822a96b130e973197dbcb7e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 229, + 349, + 238 + ], + "lines": [ + { + "bbox": [ + 340, + 229, + 349, + 238 + ], + "spans": [ + { + "bbox": [ + 340, + 229, + 349, + 238 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 389, + 156, + 463, + 229 + ], + "blocks": [ + { + "bbox": [ + 389, + 156, + 463, + 229 + ], + "lines": [ + { + "bbox": [ + 389, + 156, + 463, + 229 + ], + "spans": [ + { + "bbox": [ + 389, + 156, + 463, + 229 + ], + "type": "image", + "image_path": "cd8b8f23ab744ac4c0ef400d0bf30158946b909e777704829624c150d4084f32.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 230, + 429, + 238 + ], + "lines": [ + { + "bbox": [ + 419, + 230, + 429, + 238 + ], + "spans": [ + { + "bbox": [ + 419, + 230, + 429, + 238 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 247, + 547, + 403 + ], + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 403 + ], + "type": "text", + "content": "Figure 1: A 2D illustrative example. We observe that the neural UDF tends to have larger errors near the zero-level set " + }, + { + "bbox": [ + 304, + 247, + 547, + 403 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 247, + 547, + 403 + ], + "type": "text", + "content": " and its cut locus. (a) The target shape (the red solid curve), its cut locus (green dashed), and the induced GT UDF; (b) Approximation errors of the neural UDF to the GT UDF; (c) Gradient direction errors between the neural UDF and the GT UDF; For the local region outlined with the white box around the corner, we show the close-up views of the distance error and the gradient direction error of the region in (e) and (f), respectively; (d) The GT UDF and its approximation by a neural UDF showing that the neural UDF is positive and smooth at the zero-level set of the ideal UDF, but the ideal UDF is non-differentiable." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 470, + 156, + 544, + 229 + ], + "blocks": [ + { + "bbox": [ + 470, + 156, + 544, + 229 + ], + "lines": [ + { + "bbox": [ + 470, + 156, + 544, + 229 + ], + "spans": [ + { + "bbox": [ + 470, + 156, + 544, + 229 + ], + "type": "image", + "image_path": "70976584046c56c46c7c569dedf3de2be1c326b92592cd56e330b3128e955b6e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 230, + 510, + 238 + ], + "lines": [ + { + "bbox": [ + 501, + 230, + 510, + 238 + ], + "spans": [ + { + "bbox": [ + 501, + 230, + 510, + 238 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "text", + "content": "surface " + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "text", + "content": " defined by the zero-level set of " + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "text", + "content": ", that is, " + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\bar{S} = \\{\\mathbf{x} | \\bar{F}(\\mathbf{x}) = 0\\} \\subset \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "text", + "content": ". Note that the UDF " + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "text", + "content": " is nonnegative, so it does not have sign changes across the surface " + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "text", + "content": ". Furthermore, " + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "text", + "content": " is not differentiable at the surface " + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 426, + 547, + 487 + ], + "type": "text", + "content": " or the surface's cut locus [9]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": "Now suppose that the ideal UDF " + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": " is approximated by an MLP, denoted by " + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": ". The approximation errors of " + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": " are significant in the narrow region around the surface and a narrow region around the cut locus of " + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": ", because the neural UDF " + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": " is inherently smooth and thus poorly approximates the ideal " + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": ", which is non-smooth at the surface " + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": " and its cut locus. Due to these errors, in general, we have " + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x}) > 0" + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\bar{S}" + }, + { + "bbox": [ + 304, + 487, + 547, + 595 + ], + "type": "text", + "content": " when the MLP is differentiable." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 597, + 547, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 547, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 547, + 681 + ], + "type": "text", + "content": "For ease of visualization, we will use a 2D example (see Fig. 1) to illustrate the characteristic behaviors of the approximation errors of " + }, + { + "bbox": [ + 304, + 597, + 547, + 681 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 597, + 547, + 681 + ], + "type": "text", + "content": ", whose zero-level set defines an open curve (red in Fig. 1(a)). The characteristics of the approximation errors of " + }, + { + "bbox": [ + 304, + 597, + 547, + 681 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 597, + 547, + 681 + ], + "type": "text", + "content": " in 3D space are similar. Detailed visualization and analysis of the error behaviors in 3D space are provided in the supplementary materials. Note" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1 The cut locus of a surface is a set of points such that each point of the set has two or more distinct closest points on the surface." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22533" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "that the cut locus (Fig. 1(a)) of " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " touches the sharp features of " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": ". Therefore the neural UDF " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " has even more significant errors in terms of both the distance value and the gradient direction around the sharp features (see the error maps of the distance values and the gradient directions in Fig. 1(e,f), respectively). Consequently, in 3D space, the resulting inaccurate distance values and unreliable gradient vectors of the neural UDF " + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": " make it hard to faithfully reconstruct the target surface, especially to preserve the sharp edges in the extracted surface." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": "To recap, given a neural UDF " + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": ", there are mainly two reasons for the difficulty in extracting a surface " + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "inline_equation", + "content": "\\bar{S}_M" + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": " to approximate the target surface " + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": " defined by the ideal UDF " + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": ": (1) the given neural UDF " + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": " is usually a poor approximation of the ideal UDF " + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": " around the target surface " + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": " and its cut locus, where " + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "inline_equation", + "content": "\\bar{F}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": " is non-differentiable; and (2) the neural UDF " + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": " does not, in general, have a zero-crossing around the target surface " + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 192, + 290, + 347 + ], + "type": "text", + "content": ", thus no well-defined surface is associated with it. These issues, faced by current deep neural networks, make it hard to estimate the location of the surface in a numerically stable manner and thus motivate us to develop the two filtering criteria described in Sec. 4.3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 357, + 105, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 105, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 105, + 369 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 377, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 484 + ], + "type": "text", + "content": "Our method consists of three major designs: (1) a quadratic error function (QEF) that solves for a surface point within a cell for mesh extraction (Sec. 4.1); (2) an adaptive octree data structure that produces high-resolution grid cells while reducing the number of QEF solves required (Sec. 4.2); and (3) a point filtering strategy (Sec. 4.3) for a neural UDF, whose distance values and gradient directions are considerably more noisy compared to the GT as discussed in Sec. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "type": "text", + "content": "The proposed DualMesh-UDF pipeline has three main steps. Firstly, it employs an adaptive tree to partition the domain of the UDF containing the target surface " + }, + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 46, + 485, + 288, + 628 + ], + "type": "text", + "content": ". Then, it detects non-empty cells (defined in Sec. 4.2) based on a two-step cell-shape intersection detection method, and solves one surface point per non-empty leaf cell in a least squares manner. For a neural network, two distance bounds are introduced to filter out sample points that may introduce noisy information and contaminate the localization of the target surface. Lastly, since there are no sign changes in the UDFs, we construct the mesh faces by checking each edge shared by four incident non-empty cells in the octree." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 204, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 204, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 204, + 648 + ], + "type": "text", + "content": "4.1. QEF to locate surface points" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Given a cell that contains part of the target surface, we present a procedure to estimate a surface point in this nonempty cell, as illustrated in Fig. 2 for 2D demonstrations. For a neural UDF, we aim to compute a reconstructed point per non-empty cell. For a set of sample points " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_i\\}_{i=1}^m" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " in" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 425, + 192 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 425, + 192 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 425, + 192 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 425, + 192 + ], + "type": "image", + "image_path": "a6322235f5ca0a395f8b313a48591d9a740331f89fa465344a519d0367fb6b12.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 195, + 409, + 206 + ], + "lines": [ + { + "bbox": [ + 325, + 195, + 409, + 206 + ], + "spans": [ + { + "bbox": [ + 325, + 195, + 409, + 206 + ], + "type": "text", + "content": "(a) A neural UDF case" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 216, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 304, + 216, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 547, + 277 + ], + "type": "text", + "content": "Figure 2: The QEF formulation for the unsigned distance field. The orange lines represent the target surface to be reconstructed. The gray lines in (a) stand for the estimated tangent planes contributed by the corresponding sample points. The red point is the solution to the QEF problem." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 543, + 192 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 543, + 192 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 543, + 192 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 543, + 192 + ], + "type": "image", + "image_path": "0a30c13cafd837e28f4f7ac3b0c596e4abc65b6588d03750e0661db5f48d35c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 195, + 527, + 206 + ], + "lines": [ + { + "bbox": [ + 444, + 195, + 527, + 206 + ], + "spans": [ + { + "bbox": [ + 444, + 195, + 527, + 206 + ], + "type": "text", + "content": "(b) An ideal UDF case" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "text", + "content": "the given cell, we query the neural UDF " + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "text", + "content": " to get the approximate distance value and the normalized gradient vector at " + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "inline_equation", + "content": "d_i = F(\\mathbf{p}_i)" + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i = \\nabla F(\\mathbf{p}_i) / \\| \\nabla F(\\mathbf{p}_i)\\|" + }, + { + "bbox": [ + 304, + 298, + 545, + 346 + ], + "type": "text", + "content": ", respectively (see Fig. 2a)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 347, + 547, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 420 + ], + "type": "text", + "content": "Each of the points contributes to an estimated tangent plane of the target surface " + }, + { + "bbox": [ + 304, + 347, + 547, + 420 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 347, + 547, + 420 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 347, + 547, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_i\\cdot (\\mathbf{p}_i - \\mathbf{x}) - d_i = 0" + }, + { + "bbox": [ + 304, + 347, + 547, + 420 + ], + "type": "text", + "content": ". Considering the inaccurate nature of the neural UDF, we define the estimated surface point to be the point that is the closest to all these tangent planes and compute it by minimizing a quadratic error function (a least squares problem):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 429, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 429, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 343, + 429, + 545, + 460 + ], + "type": "interline_equation", + "content": "\\mathbf {v} = \\arg \\min _ {\\mathbf {x}} \\sum_ {i = 1} ^ {m} \\left(\\mathbf {n} _ {i} \\cdot \\left(\\mathbf {p} _ {i} - \\mathbf {x}\\right) - d _ {i}\\right) ^ {2}, \\tag {1}", + "image_path": "d3707900b4e78ae978df5bcef61dfad0c7c59bf7e36966bd5f9d121617157757.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "text", + "content": " is the number of points " + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_i\\}" + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "text", + "content": " is a surface point to be solved. In our implementation, we solve the linear least squares problem in Eqn. 1 of the form " + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\min \\| Ax - b\\|^2" + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "text", + "content": ". For a non-empty cell, this procedure can yield a reconstructed point " + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "text", + "content": ", which the target surface " + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 304, + 469, + 545, + 541 + ], + "type": "text", + "content": " approximately crosses." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 542, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 674 + ], + "type": "text", + "content": "However, when " + }, + { + "bbox": [ + 304, + 542, + 545, + 674 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 542, + 545, + 674 + ], + "type": "text", + "content": " is nearly singular, the solution to Eqn. 1 can be located near the boundary of the cell or even outside of it, which would compromise the quality of extracted mesh surfaces. Note that while [10] proposed an approach with an additional regularization point to stabilize such cases, their approach relies on the intersection points at the zero-level set of an SDF with the cell edges. However, in our setting, the zero-level set of the UDF is elusive or unavailable, therefore we do not have such intersections. We propose an alternative approach based on singular value analysis of the matrix " + }, + { + "bbox": [ + 304, + 542, + 545, + 674 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 542, + 545, + 674 + ], + "type": "text", + "content": ". We denote the three singular val" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 683, + 547, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 547, + 712 + ], + "type": "text", + "content": "1While " + }, + { + "bbox": [ + 306, + 683, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\| \\nabla F\\|" + }, + { + "bbox": [ + 306, + 683, + 547, + 712 + ], + "type": "text", + "content": " should be 1 for a distance function, note that since " + }, + { + "bbox": [ + 306, + 683, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\nabla F" + }, + { + "bbox": [ + 306, + 683, + 547, + 712 + ], + "type": "text", + "content": " is drawn from a neural representation, it may differ slightly, and thus need normalization." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22534" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "ues of the matrix " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\sigma_0\\geq \\sigma_1\\geq \\sigma_2\\geq 0" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": " and consider three cases:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 287, + 302 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 46, + 97, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 287, + 133 + ], + "type": "text", + "content": "1) In the non-singular case, where all three singular values are much larger than 0, " + }, + { + "bbox": [ + 46, + 97, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 97, + 287, + 133 + ], + "type": "text", + "content": " corresponds to a sharp feature point that can be solved directly from Eqn. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "type": "text", + "content": "2) If only " + }, + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "type": "inline_equation", + "content": "\\sigma_{2} \\approx 0" + }, + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "type": "text", + "content": ", the solution space of " + }, + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "type": "text", + "content": " corresponds to a linear edge within the cell. This line has the same direction with the singular vector corresponding to " + }, + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "type": "inline_equation", + "content": "\\sigma_{2}" + }, + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "type": "text", + "content": " and passes the solution of Eqn. 1. We then compute the intersections between this line and all 6 faces of the cell, yielding two intersecting points. We set " + }, + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 134, + 287, + 217 + ], + "type": "text", + "content": " as the midpoint of these two points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "type": "text", + "content": "3) In the case where both " + }, + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "type": "inline_equation", + "content": "\\sigma_{1,2} \\approx 0" + }, + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "type": "text", + "content": ", the part of the target surface enclosed in this cell is approximately planar. Similar to the edge case, we can formulate a plane function from the corresponding singular vectors. By computing the intersecting points between all 12 edges of the cell and this plane, we obtain multiple intersecting points. We use the centroid of these points as " + }, + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 304, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 338 + ], + "type": "text", + "content": "This way, in the degenerate cases (2 and 3), the point " + }, + { + "bbox": [ + 46, + 304, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 304, + 287, + 338 + ], + "type": "text", + "content": " is still a good approximate solution to Eqn. 1 and well positioned inside the cell." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "type": "text", + "content": "In a neural UDF (e.g., Fig. 2a), the UDF values " + }, + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\{d_i\\}" + }, + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "type": "text", + "content": " and the gradient directions " + }, + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{n}_i\\}" + }, + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "type": "text", + "content": " are approximated and often unreliable. Consequently, the tangent plane computed from each sample point " + }, + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 340, + 287, + 460 + ], + "type": "text", + "content": " may not exactly align with the target surface. Solving Eqn. 1 yields a least squares solution that is numerically robust to the approximation errors in neural UDF and can lead to an accurate estimation of the surface point in this non-empty cell. For an ideal UDF case, our method is also applicable to yield an accurate result as shown in Fig. 2b." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 461, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 532 + ], + "type": "text", + "content": "Furthermore, to enhance the robustness of our method, we also design a point filtering strategy to remove unreliable sample points, especially those near the target surface and its cut locus, where larger approximation errors exist in the MLP-encoded UDF. The details of this filtering strategy are explained in Sec. 4.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "Differences to DC [10]. Note that the dual contouring (DC) method [10] designs a QEF using intersection points between the surface and cell edges, along with the gradient directions at these intersection points. This approach is not applicable to our objective of extracting surface meshes from a neural UDF. This is because (1) it is difficult to find a reliable intersection point, as an exact zero-level set that crosses an edge may not exist, and (2) the gradient directions in the region with lower UDF values (i.e., closer to the target surface) are highly unstable and thus unreliable for estimating the tangent plane. Hence, our method leverages off-surface points that are sufficiently far from the target surface, as indicated by their UDF values (see Sec. 4.3). Our approach thus has less stringent requirements than the DC method (as presented in [10]), yet produces satisfactory" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 545, + 164 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 164 + ], + "type": "image", + "image_path": "6219206ecfbcdd1522f94ca054eae62d26b2ac41bdcd5b58425703e4f7cdc5bb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 172, + 535, + 184 + ], + "lines": [ + { + "bbox": [ + 315, + 172, + 535, + 184 + ], + "spans": [ + { + "bbox": [ + 315, + 172, + 535, + 184 + ], + "type": "text", + "content": "Figure 3: The pipeline of our mesh extraction method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 209, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 209, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 545, + 232 + ], + "type": "text", + "content": "results even on noisy neural UDFs as demonstrated by our results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 244, + 471, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 244, + 471, + 257 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 471, + 257 + ], + "type": "text", + "content": "4.2. Octree Design and Subdivision" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 264, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 348 + ], + "type": "text", + "content": "To efficiently process high-resolution data and speed up the mesh extraction process, we employ an octree data structure. As shown in Fig. 3, we recursively subdivide the cells unless they are categorized as empty according to the following criterion or they reach the maximal depth. After the subdivision, all leaf nodes that are non-empty will invoke the QEF procedure to solve for the dual points." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "content": "Checking obviously empty cells. Firstly, we design a checking condition to help quickly prune obviously empty cells. Given an octree cell " + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_0" + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "content": " denote its center and " + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "d_0" + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "content": " denote the UDF value at " + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_0" + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "content": ", that is, " + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "d_0 = F(\\mathbf{c}_0)" + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "content": ". Then we have the following sufficient condition for cell " + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "content": " to be empty:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 388, + 426, + 545, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 426, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 388, + 426, + 545, + 439 + ], + "type": "interline_equation", + "content": "d _ {0} > \\operatorname {D i a g} (C) / 2, \\tag {2}", + "image_path": "6783e9a1150dfb24cde36e039f3e9c882ce09d589301d5cf225f169f527694da.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\operatorname{Diag}(C)" + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "content": " is the diagonal length of cell " + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "content": ". Since the UDF distance indicates the distance between a spatial point and the target surface, this criterion determines if the target surface lies outside the sphere centered at " + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_0" + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "content": " with radius " + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\operatorname{Diag}(C)/2" + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "content": ". Given that this sphere contains the entire octree cell, the cell will not contain any portion of the surface if Eqn. 2 is satisfied. Otherwise, the cell will first be categorized as an unsolved cell. Since neural UDF is a reasonable approximation of the ideal UDF, we tailor Eqn. 2 to " + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "inline_equation", + "content": "d_0 > \\operatorname{Diag}(C)/2 + \\epsilon" + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon > 0" + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "content": " is a tolerance for the approximation error and set to " + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\epsilon = 2 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "content": " throughout all experiments. Empirically, we observed this pruning strategy did not affect the quality of the results." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Maximum octree depth. Having a predefined maximal depth of the octree is critical, not only in terms of limiting the amount of computation but also because of the inaccuracy of the neural UDF near the target surface. Specifically, if the octree keeps subdividing the space that contains the target surface, after a certain depth the cells will become too small and the sample points " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " within these cells will lie in the unreliable region of the neural UDF, which will eventually lead to unstable estimation." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22535" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 214, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 214, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 214, + 85 + ], + "type": "text", + "content": "4.3. Point filtering for neural UDFs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 288, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 288, + 197 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 288, + 197 + ], + "type": "text", + "content": "Since our method makes use of the gradient directions and the distance values to estimate a surface point in each non-empty grid cell (see Eqn. 1), the quality of the extracted mesh heavily depends on the approximation accuracy of the neural UDF " + }, + { + "bbox": [ + 46, + 90, + 288, + 197 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 90, + 288, + 197 + ], + "type": "text", + "content": ". Consider a neighborhood of a sharp feature of the target surface, as shown in the zoom-in views of Fig. 1(e,f). The distance errors and gradient direction errors are more significant around the target surface and its cut locus." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 258 + ], + "type": "text", + "content": "Hence, motivated by our observation and analysis in Sec. 3, to avoid building the QEF using points from these regions, we use the following criteria to filter out unreliable points to enhance the robustness of the QEF solution to the noisy characteristics of the MLP-encoded UDFs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 329 + ], + "type": "text", + "content": "Criterion 1: Removing sample points potentially near the surface. A candidate point " + }, + { + "bbox": [ + 46, + 258, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 258, + 287, + 329 + ], + "type": "text", + "content": " is considered too close to the target surface if " + }, + { + "bbox": [ + 46, + 258, + 287, + 329 + ], + "type": "inline_equation", + "content": "F(\\mathbf{p}_i) < \\delta_1" + }, + { + "bbox": [ + 46, + 258, + 287, + 329 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 258, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\delta_1" + }, + { + "bbox": [ + 46, + 258, + 287, + 329 + ], + "type": "text", + "content": " is a preset filtering threshold; such points are discarded. This criterion ensures the sample points are from a region where the distance errors are expected to be relatively small." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "text", + "content": "Criterion 2: Removing sample points whose projections have large UDF values. While the previous criterion rejects the sample points that are too close to the target surface and thus avoids the erroneous approximation at those regions, some sample points may still have larger UDF errors even far from the target surface (e.g. near the cut locus). We observe that these sample points will not be projected to regions near the target surface. Hence, if the UDF of the projected point " + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q}_i\\}" + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "text", + "content": " is large, we consider it to be an incorrect estimate and thus reject the corresponding sample point " + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_i\\}" + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "text", + "content": " as unreliable. To this end, we introduce the second filtering criterion as follows, " + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "inline_equation", + "content": "F(\\mathbf{q}_i) > \\delta_2" + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "inline_equation", + "content": "\\delta_2" + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "text", + "content": " is another preset filtering bound; again, the corresponding points " + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 330, + 288, + 496 + ], + "type": "text", + "content": " are discarded." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 503, + 219, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 219, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 219, + 517 + ], + "type": "text", + "content": "4.4. Creating Output Mesh Surfaces" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 714 + ], + "type": "text", + "content": "We consider the regular grid of cubic cells at the maximum depth of the octree. Then each edge of the grid is shared by four cubic cells surrounding the edge. To build the initial mesh connectivity, for each edge of the grid, we examine each of its four incident cells. Similar to [10], the mesh connection rule is designed as follows: if all four incident cells are non-empty, a quad-face candidate that connects the four dual points in these cells will be constructed. To ensure the correct connectivity, we validate and triangulate the quad faces. We further examine if the normals of the face candidates can consistently reflect their geometric property, being a sharp corner, part of a sharp edge, or part of a plane as classified by the SVD shape analysis (Sec. 4.1). For example, all correct triangulated faces incident to a surface point classified as part of a plane would have all their normals parallel to the singular vector corresponding to the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "largest singular value from SVD shape analysis, while the normals of triangulated faces incident to a surface point being part of a sharp edge should be orthogonal to its singular vector direction. This way, we reject all inconsistent face candidates." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 133, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 277 + ], + "type": "text", + "content": "We also provide a practical approach to make sure the output mesh is manifold when the desired target surface is manifold: With the help of the octree cells and the definition of reconstructed dual points, it is feasible to generate an auxiliary blocky model by moving all mesh vertices to the centroids of the corresponding cells. By extracting the outer envelope of this auxiliary model, we obtain a manifold structure. Finally, we tessellate the surface points using the connectivity from that manifold structure to produce a manifold mesh. This approach is crucial for applications that require manifold meshes (e.g. parameterization, remeshing, or shape analysis)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 289, + 388, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 388, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 388, + 303 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 309, + 484, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 484, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 484, + 323 + ], + "type": "text", + "content": "5.1. Experimental details and metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "type": "text", + "content": "Experimental setting. We rescaled all the shapes to a bounding cube with a side length of 2, centered at the origin. To improve efficiency, we share sample points between cells, i.e. " + }, + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "type": "inline_equation", + "content": "m = 27" + }, + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "type": "text", + "content": " points in each non-empty leaf cell, including 8 corner points, 12 edge midpoints, 6 face midpoints, and 1 centroid point. We set the filtering parameters " + }, + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\delta_{1} = 2 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\delta_{2} = 2 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 329, + 545, + 449 + ], + "type": "text", + "content": " to filter the sampled points as described in Sec. 4.3. We notice that more sample points may bring marginal performance improvement but incur computational overhead in our ablation study." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": "Given that the grid resolution of " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "128^{3}" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "256^{3}" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": " are commonly used in related works, we compare our method with these prior techniques using a maximal octree resolution of " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "128^{3}" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "256^{3}" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": ". We also discuss how the resolution will affect our results by varying the maximum depth of the octree to have a resolution from " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "inline_equation", + "content": "512^{3}" + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": " in our supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "MLP architecture. To overfit single shapes using individual MLPs, we employed the MLP implementation provided by [18]. The activation functions are the Sine activation, except for the last one which is a SoftPlus " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\beta = 100)" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " activation to ensure the output value is non-negative. All neural UDFs in the experiments were trained with the described MLP implementation. All MLP networks were trained for " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "3k" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " iterations with the ADAM optimizer to minimize the difference between the predicted and the GT unsigned distance fields. For more general tasks, we also test an MLP with latent codes that represent a shape space, following the network and training settings in [5]. We report the timing performance on a Linux desktop with an Intel Core™ i7-10870H CPU and an NVIDIA GeForce RTX 3090 graphics card. We elaborate on the loss function as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22536" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 534, + 671 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 534, + 671 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 534, + 671 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 534, + 671 + ], + "type": "image", + "image_path": "5b0ccebcaeb2fa41e52fe066cb3f788591bbd72644ac4a0147bfd3051ae304f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 678, + 548, + 715 + ], + "lines": [ + { + "bbox": [ + 46, + 678, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 678, + 548, + 715 + ], + "type": "text", + "content": "Figure 4: Meshes extracted from neural UDFs. We show 6 examples and compare our results to the mesh extraction results in MeshUDF [7], CAP-UDF [21] and NDC [3]. Our method preserves sharp geometric features better, yields results without undesirable holes, and reconstructs the original smooth boundaries of open surfaces." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22537" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 127, + 520, + 228 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "type": "text", + "content": "Table 1: Quantitative comparison between the results obtained by our method and those by the competing methods, MeshUDF, CAP-UDF, and NDC. The average performance on the dataset containing 354 shapes is reported. The Chamfer distance (CD) and the Hausdorff Distance (HD) are scaled by " + }, + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 70, + 547, + 120 + ], + "type": "text", + "content": ", respectively. T1 and T2 stand for the time spent (seconds) on the mesh extraction and that on the UDF query, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 127, + 520, + 228 + ], + "lines": [ + { + "bbox": [ + 73, + 127, + 520, + 228 + ], + "spans": [ + { + "bbox": [ + 73, + 127, + 520, + 228 + ], + "type": "table", + "html": "
MGNThingi10KABCRunning time
CD↓F-score↑HD↓CD↓F-score↑HD↓CD↓F-score↑HD↓T1T2
1283Ours2.3898.0911.911.9797.519.213.6993.4111.470.2971.184
MeshUDF [7]4.7690.0622.394.5790.7815.338.4987.7423.110.3131.316
CAP-UDF [21]13.5387.1060.0419.5584.8557.4927.7577.4353.466.4664.764
NDC [3]3.3295.6112.042.7496.449.693.9592.6514.042.3581.324
2563Ours2.0398.967.351.6098.296.782.2296.578.201.4925.508
MeshUDF [7]3.4595.1012.83.3394.138.913.2593.8511.491.60010.27
CAP-UDF [21]8.4592.8849.4012.1790.5347.2821.6783.8843.3349.41637.336
NDC [3]2.5498.047.842.4297.167.832.3696.1011.389.36310.256
", + "image_path": "a31f2ae1fb6e62cb60ddb3689ce4f74be1770bb54ecb677b28800d93afd8caa2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 72, + 326, + 261, + 389 + ], + "blocks": [ + { + "bbox": [ + 46, + 245, + 288, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 245, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 288, + 316 + ], + "type": "text", + "content": "Table 2: Quantitative comparison between our method and three competing methods, MeshUDF, CAP-UDF, and NDC on a pre-trained UDF network with latent codes, on the ShapeNet dataset. In this table, CD and HD are scaled by " + }, + { + "bbox": [ + 46, + 245, + 288, + 316 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 245, + 288, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 245, + 288, + 316 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 46, + 245, + 288, + 316 + ], + "type": "text", + "content": " respectively. The F-score is calculated with a threshold of 0.006." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 326, + 261, + 389 + ], + "lines": [ + { + "bbox": [ + 72, + 326, + 261, + 389 + ], + "spans": [ + { + "bbox": [ + 72, + 326, + 261, + 389 + ], + "type": "table", + "html": "
CD↓F-score↑HD ↓
Ours3.4182.303.92
MeshUDF [7]4.1079.354.67
CAP-UDF [21]7.7070.307.21
NDC [3]7.1678.3812.38
", + "image_path": "86e5515a7df0ce7d4b9c44bbdd36e2985cdbf3e0560d590bd37fa66618978c81.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 410, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 434 + ], + "type": "text", + "content": "well as other details regarding training the neural UDFs in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 434, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 288, + 579 + ], + "type": "text", + "content": "Metrics. To evaluate the performance of our method and the other methods, we adopt the following metrics, i.e. the double-sided Chamfer distance (CD), the F-score based on CD, and the Hausdorff distance (HD). The Chamfer distance reflects the overall quality of the extracted surface mesh as compared to the GT shape. The F-score indicates the percentage of points that are reconstructed correctly under a threshold (set to 0.001 for shape-overfitting neural UDFs, and 0.01 for shared UDF network with latent codes). The Hausdorff distance can reveal if there is anything missing or redundant in the reconstructed geometry (e.g. a hole or a floating piece)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 225, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 225, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 225, + 598 + ], + "type": "text", + "content": "5.2. Comparison with SOTA methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "We compared our results with three existing methods: 1) MeshUDF [7], 2) the standalone mesh extraction module presented in CAP-UDF [21], and 3) the UNDC presented in NDC [3]. The first two methods use gradients of the UDF to estimate sign changes in the field to invoke the Marching Cubes method to extract a surface mesh from the UDF. NDC proposes a data-driven approach to extract a surface mesh from the grid-based representation of an implicit field. Shape-overfitting neural UDFs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 247, + 547, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 377 + ], + "type": "text", + "content": "We first compare our results to those obtained by these three methods on a shape collection consisting of shapes from four public sources: 1) 100 from the Thingi10K Dataset [22] containing 3D printing models; 2) 134 from the MGN Dataset [2] containing garments with open boundaries; 3) 100 from the ABC Dataset [12] containing CAD models; and 4) 20 commonly used shapes in geometric processing research. In addition, we also compare different methods on a Möbius strip, which is a non-orientable surface. We fit each shape with an independent neural UDF and apply mesh extraction to each of these neural UDFs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 379, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 547, + 521 + ], + "type": "text", + "content": "Table 1 reports the performance of different methods on this shape collection. Our method outperforms the other three competing methods in terms of all three quantitative metrics. While all of these methods use uniform grids, our octree structure results in increased computing efficiency. MeshUDF and CAP-UDF would likely be accelerated by adopting an octree structure, however, it would be nontrivial to adopt an octree structure for NDC. For a fair comparison, we also test our method without using the octree acceleration approach for the resolution of " + }, + { + "bbox": [ + 304, + 379, + 547, + 521 + ], + "type": "inline_equation", + "content": "128^{3}" + }, + { + "bbox": [ + 304, + 379, + 547, + 521 + ], + "type": "text", + "content": "; by doing so, the T1 time increases to 2.70s, and the T2 time increases to 5.54s." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 522, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 689 + ], + "type": "text", + "content": "To qualitatively compare results obtained by different methods, we show the mesh surfaces extracted from neural UDFs for several shapes in Fig. 4. We can see that our results are higher quality than those produced by MeshUDF and CAP-UDF. These two methods cannot preserve sharp geometric features due to the use of the Marching Cubes method. Compared to NDC - a data-driven method - our method also produces consistently better results both quantitatively and qualitatively. Specifically, some mesh surfaces extracted by NDC are less smooth than ours. One example (the Pants) is shown in the fourth row in Fig. 4. The staircase artifact observed may be attributed to NDC having been trained on the ABC dataset [12] containing only mechanical components." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Our method is the only method that recovers open boundaries faithfully, while the other compared methods" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22538" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 115, + 520, + 164 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": "Table 3: Quantitative comparison between two different sampling number settings (27 and 125 per cell respectively). The Chamfer distance (CD) and the Hausdorff Distance (HD) are scaled by " + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": ", respectively. T1 and T2 stand for the time spent (seconds) on the mesh extraction and that on the UDF query, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 115, + 520, + 164 + ], + "lines": [ + { + "bbox": [ + 73, + 115, + 520, + 164 + ], + "spans": [ + { + "bbox": [ + 73, + 115, + 520, + 164 + ], + "type": "table", + "html": "
MGNThingi10KABCRunning time
CD↓F-score↑HD↓CD↓F-score↑HD↓CD↓F-score↑HD↓T1T2
272.3898.0911.911.9797.519.213.6993.4111.470.2971.184
1252.3498.1111.771.9197.529.023.7092.8911.120.9657.608
", + "image_path": "bed410780710cbba2ea8c8d245e2d57bc6f2e685860ef795653e52403a38e93b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 183, + 289, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 289, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 289, + 290 + ], + "type": "text", + "content": "show staircase artifacts or even redundant pieces near the open boundaries; see the first two rows of Fig. 4. Also, unexpected holes can be observed on the mesh surface extracted by the competing methods from the neural UDFs. Empirically, our method produces quality results without redundant pieces or unexpected holes, generating clear-cut boundaries of the open surfaces as shown in the top four rows of Fig. 4, and reproducing the sharp features as shown in the bottom two rows of the same figure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 288, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 288, + 327 + ], + "type": "text", + "content": "More results and comparisons (also on the GT UDF) are presented in our Supplementary Materials, which further validate the superior performance of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 288, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 288, + 460 + ], + "type": "text", + "content": "Shared, pre-trained UDF network with latent codes. We consider a single neural network with latent codes trained to represent the entire shape space, where each shape is associated with a unique latent code. Specifically, we test our method and the competing methods on the pre-trained UDF network provided by [5] representing 300 ShapeNet car models. Although this UDF network is less accurate than the other overfitting-based network settings considered earlier, our method still outperforms the other methods by a significant margin as reported in Tab. 2, showing the versatility and robustness of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 466, + 139, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 139, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 139, + 479 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 484, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 557 + ], + "type": "text", + "content": "We conducted an ablation study on our method using sampling numbers (the number " + }, + { + "bbox": [ + 46, + 484, + 287, + 557 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 484, + 287, + 557 + ], + "type": "text", + "content": " in Eqn. 1) of " + }, + { + "bbox": [ + 46, + 484, + 287, + 557 + ], + "type": "inline_equation", + "content": "27(3^{3})" + }, + { + "bbox": [ + 46, + 484, + 287, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 484, + 287, + 557 + ], + "type": "inline_equation", + "content": "125(5^{3})" + }, + { + "bbox": [ + 46, + 484, + 287, + 557 + ], + "type": "text", + "content": " to justify our design choice to use 27 as our sampling number. In Tab. 3, we demonstrate that using 125 sample points per cell results in a marginal improvement but incurs a significantly higher computing cost." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 125, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 125, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 125, + 574 + ], + "type": "text", + "content": "5.4. Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "While DualMesh-UDF demonstrates the ability to accurately extract surfaces from neural UDFs, some limitations remain. First, the extracted surface cannot be adaptively subdivided with respect to sharp features or fine geometry details. To faithfully reconstruct the fine geometry details, we need to pre-define a sufficient depth of the octree and solve the QEF problem in each of its non-empty leaf nodes. Second, to share the UDF and gradient values at sampled points between adjacent octree cells, we adopt a regular grid sampling strategy. A more flexible and adaptive sampling strategy may bring further improvement to our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 182, + 379, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 379, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 379, + 194 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 202, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 321 + ], + "type": "text", + "content": "We have presented a method for extracting high-quality surface meshes from unsigned distance fields. In order to attain robust performance on MLP-encoded neural UDFs, we discuss the characteristics of the approximation errors of the neural UDF and develop an adaptive octree-based method to effectively localize the target surface embedded in the given UDF. Extensive experiments show that our method outperforms the SOTA methods and produces high-quality results with sharp geometry features, with clear open boundaries, and free of undesirable holes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 333, + 416, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 333, + 416, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 333, + 416, + 346 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 415 + ], + "type": "text", + "content": "This research is partly supported by the Innovation and Technology Commission of the HKSAR Government through the InnoHK initiative. The authors would like to thank Yanhong Lin and Ruixing Jia for their generous help on the project." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 425, + 365, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 365, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 365, + 437 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 444, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 310, + 444, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 444, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 310, + 444, + 545, + 478 + ], + "type": "text", + "content": "[1] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 479, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 479, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 311, + 479, + 547, + 533 + ], + "type": "text", + "content": "[2] Bharat Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5419-5429, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 311, + 534, + 545, + 567 + ], + "type": "text", + "content": "[3] Zhiqin Chen, Andrea Tagliasacchi, Thomas Funkhouser, and Hao Zhang. Neural dual contouring. ACM Trans. Graph., 41(4), jul 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 568, + 547, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 568, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 311, + 568, + 547, + 622 + ], + "type": "text", + "content": "[4] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6968-6979, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 624, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 624, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 311, + 624, + 547, + 690 + ], + "type": "text", + "content": "[5] Julian Chibane, Mohamad Aymen mir, and Gerard Pons-Moll. Neural unsigned distance fields for implicit function learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21638-21652. Curran Associates, Inc., 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 691, + 547, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 691, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 311, + 691, + 547, + 715 + ], + "type": "text", + "content": "[6] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learn" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22539" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "ing shapes. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 3789-3799. PMLR, 13-18 Jul 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 185 + ], + "type": "text", + "content": "[7] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 576-592, Cham, 2022. Springer Nature Switzerland." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 186, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 288, + 239 + ], + "type": "text", + "content": "[8] Chiyu Jiang, Avneesh Sud, Ameesh Makadia, Jingwei Huang, Matthias Nießner, and Thomas Funkhouser. Local implicit grid representations for 3d scenes. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6000-6009, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 241, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 288, + 285 + ], + "type": "text", + "content": "[9] M.W. Jones, J.A. Baerentzen, and M. Sramek. 3d distance fields: a survey of techniques and applications. IEEE Transactions on Visualization and Computer Graphics, 12(4):581-599, 2006." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "text", + "content": "[10] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. ACM Trans. Graph., 21(3):339-346, jul 2002." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 387 + ], + "type": "text", + "content": "[11] Leif P. Kobbelt, Mario Botsch, Ulrich Schwanecke, and Hans-Peter Seidel. Feature sensitive surface extraction from volume data. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '01, page 57-66, New York, NY, USA, 2001. Association for Computing Machinery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 453 + ], + "type": "text", + "content": "[12] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 455, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 521 + ], + "type": "text", + "content": "[13] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '87, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 523, + 286, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 286, + 566 + ], + "type": "text", + "content": "[14] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. ACM Trans. Graph., 40(4), jul 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 622 + ], + "type": "text", + "content": "[15] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4455-4465, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[16] Mateusz Michalkiewicz, Jhony Kaesemodel Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4742-4751, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[17] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 417 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "type": "text", + "content": "[18] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7462-7473. Curran Associates, Inc., 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "text", + "content": "[19] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11353-11362, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 230, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 316 + ], + "type": "text", + "content": "[20] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 7537-7547. Curran Associates, Inc., 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 318, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 318, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 318, + 545, + 384 + ], + "type": "text", + "content": "[21] Junsheng Zhou, Baorui Ma, Yu-Shen Liu, Yi Fang, and Zhizhong Han. Learning consistency-aware unsigned distance functions progressively from raw point clouds. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 385, + 545, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 385, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 308, + 385, + 545, + 417 + ], + "type": "text", + "content": "[22] Qingnan Zhou and Alec Jacobson. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv:1605.04797, 2016." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22540" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/f357e4d1-f9a1-4f72-b3d1-9e59cc154adf_content_list.json b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/f357e4d1-f9a1-4f72-b3d1-9e59cc154adf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d4d79ed67a7bb6a1cf422d513eca4244dfbd86b5 --- /dev/null +++ b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/f357e4d1-f9a1-4f72-b3d1-9e59cc154adf_content_list.json @@ -0,0 +1,1463 @@ +[ + { + "type": "text", + "text": "Surface Normal Clustering for Implicit Representation of Manhattan Scenes", + "text_level": 1, + "bbox": [ + 99, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nikola Popovic1, Danda Pani Paudel1,2, Luc Van Gool1,2", + "bbox": [ + 259, + 179, + 707, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Computer Vision Laboratory, ETH Zurich, Switzerland \n2INSAIT, Sofia University, Bulgaria", + "bbox": [ + 259, + 198, + 707, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{nipopovic, paudel, vangool}@vision.ee.ethz.ch", + "bbox": [ + 281, + 236, + 687, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis and 3D modeling using implicit neural field representation are shown to be very effective for calibrated multi-view cameras. Such representations are known to benefit from additional geometric and semantic supervision. Most existing methods that exploit additional supervision require dense pixel-wise labels or localized scene priors. These methods cannot benefit from high-level vague scene priors provided in terms of scenes' descriptions. In this work, we aim to leverage the geometric prior of Manhattan scenes to improve the implicit neural radiance field representations. More precisely, we assume that only the knowledge of the indoor scene (under investigation) being Manhattan is known - with no additional information whatsoever - with an unknown Manhattan coordinate frame. Such high-level prior is used to self-supervise the surface normals derived explicitly in the implicit neural fields. Our modeling allows us to cluster the derived normals and exploit their orthogonality constraints for self-supervision. Our exhaustive experiments on datasets of diverse indoor scenes demonstrate the significant benefit of the proposed method over the established baselines. The source code will be available at https://github.com/nikola3794/normal-clustering-nerf.", + "bbox": [ + 75, + 316, + 472, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 693, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Above $80\\%$ images ever taken are estimated to involve human-made architectural structures, with a substantial share of indoor scenes [11]. These scenes often exhibit strong structural regularities, including flat and texture-poor surfaces in some axis-aligned Cartesian coordinate system - also known as Manhattan world [7, 12]. Paradoxically, these regularities may hinder the 3D modeling process if it is unaware of the human-made scene priors. In fact, several computer vision works have even benefited from the knowledge of the Manhattan world for the task of 3D scene reconstruction [43, 12], camera localization [19], self-calibration [50], and more [37, 35, 29, 57].", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For calibrated multi-view cameras, 3D inversion using implicit neural representations [24, 47, 54] is becoming increasingly popular due to their remarkable performance and recent efficiency developments [55, 30, 25, 51]. Meanwhile, such representations are known to benefit from additional supervision in the form of depth [1, 9, 32], normals [18], semantics [45, 58, 17, 15], local-regularization [46, 44, 26], local planar patches [21], or their combinations [56, 14, 21]. In this context, a notable recent work ManhattanSDF [14] demonstrates the benefit of exploiting the high-level geometric prior for structured scenes. More precisely, ManhattanSDF [14] uses the known semantic regions to impose the planar geometry prior of floors and walls under the Manhattan scene assumption. During this process, the exact normal of the floor and partial normals of the walls are assumed to be known, with respect to the camera coordinates.", + "bbox": [ + 496, + 287, + 893, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We aim to improve the 3D neural radiance field representations for calibrated multi-view cameras in indoor Manhattan scenes, with no further assumptions. In other words, we consider that the structural and semantic information is not available, unlike ManhattanSDF [14]. In addition to the floor and walls used in ManhattanSDF [14], we wish to exploit the Manhattan prior of many other indoor scene parts (e.g. tables and wardrobes). More importantly, we consider that the Manhattan coordinate frame is also unknown. Our assumptions (of unknown semantics and Manhattan frame) on one hand make our setting very practical. On the other hand, those assumptions make the problem of exploiting the Manhattan scene prior for 3D inversion very challenging.", + "bbox": [ + 496, + 532, + 893, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The virtue of the Manhattan world assumption comes from its simplicity, allowing us to intuitively reason about the geometry of a wide range of complex scenes/objects such as cities, buildings, and furniture. However, such reasoning often requires the axis-aligned Cartesian coordinate frame, also known as the Manhattan frame (MF), to be known [37]. Unfortunately, recovering the Manhattan frame directly from images is not an easy task [7, 37]. Therefore, several methods have been developed until recently [10, 4, 36, 16, 13] to recover the Manhattan frame, relying upon the known 3D reconstruction or image primi", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17860", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5775e0ce3426ff3c30121e5471a1a73efcd3a6787f004adaf4ff305745a799c9.jpg", + "image_caption": [ + "Figure 1: Illustration of the proposed method. We compute one surface normal for each ray triplet, using 3D surface points derived from rendered depths (left). The computed normals (from many ray triplets) are clustered to obtain the MF (middle). The non-Manhattan and noisy surfaces are handled by a robust orthogonal normals search, which is later used for self-supervision through the Manhattan prior (right)." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 326, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a7eab9213b88cf5b4434efa310c8a6b9ddeb780d2f63ac251c9e6b7af5deea13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 88, + 676, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bb6479a6e5384015fae3d31865f05580f1766214e15d38251e4f102df4b94293.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 84, + 893, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tives (eg. lines, planes). We wish to exploit the Manhattan prior for improving the 3D representation, without needing to know MF beforehand. Instead, our experiments reveal that knowing the MF beforehand offers no additional benefit in Manhattan-prior aware radiance field representation.", + "bbox": [ + 75, + 340, + 468, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a method that jointly learns the Manhattan frame and neural radiance field, from calibrated multi-view in indoor Manhattan scenes, in an end-to-end manner using the recent efficient backbone of Instant-NGP [25]. The proposed method requires no additional information to exploit the Manhattan prior and relies on the explicitly derived normals in the implicit neural fields. We use batches of three neighboring rays, whose effective surface's local piece-wise planarity is assumed to derive the explicit normals by algebraic means. In pure, complete, and enclosed Manhattan scenes, these normals form six clusters corresponding to three orthogonal and other three parallel counterpart planes. However real scenes consist of non-Manhattan scene parts and missing planes. Therefore, we use a robust method that uses minimal three orthogonal clusters to recover the sought Manhattan frame. As in the literature, we seek a rotation matrix whose entries are directly derived from the orthogonal clusters of normals, to align the Manhattan frame. The recovered MF is then used to encourage the derived normals to be axis-aligned for self-supervision. A graphical illustration of our method is presented in Figure 1. Our extensive experiments demonstrate the robustness and utility of our method in improving the implicit 3D in neural radiance field representations.", + "bbox": [ + 75, + 417, + 468, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The major contributions of our paper are listed below:", + "bbox": [ + 96, + 780, + 452, + 794 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We address the problem of exploiting the structured-scene knowledge without requiring any dense or localized scene priors, for the first time in this paper.", + "- We present a method that successfully exploits the Manhattan scene prior with an unknown Manhattan frame. The proposed method also recovers the unknown frame.", + "- We demonstrate the robustness and utility of our method" + ], + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on three indoor datasets, where our method improves the established baselines significantly. These datasets consist of $200+$ scenes, making our method tested in significantly more scenes than the state-of-the-art methods.", + "bbox": [ + 498, + 340, + 890, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 419, + 648, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit neural representation of 3D: Since the foundational work of Mildenhall et al. [24], the implicit neural representation of 3D scenes has advanced on various fronts. These fronts include representation [54, 40], generalization [49, 28, 20, 22, 34], generation [27, 6], and efficient methods [30, 55, 25]. We rely on a recent method Instant-NGP [25] developed by Mueller et al., as our backbone. Our choice is made primarily based on the computational efficiency during both training and inference. Thanks to the offered computational efficiency, we are able to conduct large-scale experiments on several scenes.", + "bbox": [ + 496, + 446, + 890, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Auxiliary supervision methods: In addition to the images, other inputs such as depth [33, 39, 9], semantics [58, 15, 17, 41], normal [52, 53], and their combinations [56, 14, 21] are shown to be beneficial on improving the neural radiance field representation. In this regard, these auxiliary supervisions often use ground-truth labels. It is needless to say that the need for ground-truth supervision is not desired, whenever possible. Therefore, recent methods use labels predicted by some pre-trained networks [56] or recovered from the structure-from-motion (SfM) pipeline [9, 21]. One notable work ManhattanSDF [14] exploits the Manhattan prior without requiring any SfM reconstruction. However, ManhattanSDF requires (a) semantics of the scene parts and (b) the Manhattan frame, to be known. We argue that such labels required for auxiliary supervision cannot always be relied upon, due to domain gaps, poor reconstruction of texture-less regions, and additional computational needs, to list a few. Therefore, we do not use any additional labels for auxiliary supervision.", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17861", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Manhattan frame estimation: Since the early works of Bernard [2], Manhattan structure reasoning is done directly on/from images by detecting the so-called the vanishing-points (VPs) [23, 42]. In fact, the problem of detecting three orthogonal VPs is equivalent to finding the MF in 3D for the calibrated multi-view setting [3]. Note that the knowledge of Manhattan structure has been used in several computer vision works [43, 12, 19, 50, 37, 35, 29, 57]. When unknown, most methods implicitly or explicitly estimate the MF to leverage the Manhattan scene prior. In [37], Straub et al. have demonstrated that the MF can be efficiently represented in and recovered from the space of surface normals. We use a similar formulation as [37], using the surface normals derived explicitly from the implicit neural fields.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background and Notations", + "text_level": 1, + "bbox": [ + 76, + 316, + 328, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Manhattan frame (MF) is a coordinate system that is defined by the structure building orthogonal planes of Manhattan scenes. We consider unknown MF since the scene planes and their geometric relationships are unknown. Let the unknown MF differs from the world-frame (WF), used for the 3D representation, by rotation $\\mathsf{R} \\in SO(3)$ . We denote three orthogonal axes in MF by $\\mathcal{E} = \\{\\mathsf{e}_x, \\mathsf{e}_y, \\mathsf{e}_z\\}$ . Without loss of generality, let the axes' coordinates be $\\mathsf{e}_x = [1,0,0]^{\\mathsf{T}}$ , $\\mathsf{e}_y = [0,1,0]^{\\mathsf{T}}$ , and $\\mathsf{e}_z = [0,0,1]^{\\mathsf{T}}$ . Note that these axes align with the normals of the respective scene building planes, in the WF. Let $\\mathcal{N} = \\{\\mathsf{n}_i\\}_{i=1}^m$ be a set of 3D normals of all the scene planes. Then, the rotation $\\mathsf{R}$ , from world to Manhattan frame, aligns the normals $\\mathsf{n} \\in \\mathcal{N}$ to the Manhattan axes $\\mathsf{e} \\in \\mathcal{E}$ , i.e., $\\mathsf{Rn}_i \\in \\mathcal{E}$ , $\\forall \\mathsf{n}_i \\in \\mathcal{N}$ .", + "bbox": [ + 75, + 340, + 472, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are interested to recover $\\mathsf{R}$ from a set of normals $\\mathcal{N}$ . For this, the above set-to-set assignment alone is not sufficient. This requires the element-wise assignment between sets $\\mathcal{N}$ and $\\mathcal{E}$ . To do so, we divide the set $\\mathcal{N}$ into three orthogonal subsets $\\mathcal{N}_x, \\mathcal{N}_y,$ and $\\mathcal{N}_z$ . Now, for any triplet of $\\{\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z\\}$ from the corresponding orthogonal subsets, we aim to establish the condition $[\\mathsf{e}_x, \\mathsf{e}_y, \\mathsf{e}_z] = \\mathsf{R}[\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z]$ . Note that the assignment condition results into $\\mathsf{R} = [\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z]^{\\top}$ . Therefore, the problem of recovering MF from a given set of normals boils down to finding three normals from orthogonal subsets. At this point, one issue regarding robustness remains pending. More precisely, we wish to recover $\\mathsf{R}$ for a noisy set of normals $\\mathcal{N}$ , with potentially overwhelmingly many outliers.", + "bbox": [ + 75, + 553, + 472, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use a robust method to recover $\\mathsf{R}$ from given normals $\\mathcal{N}$ . In principle, $\\mathsf{R}$ can be recovered from minimal two orthogonal normals, with one additional normal for disambiguation by sign correction. The recovered rotation can then be validated by consensus for robust recovery [37]. Alternatively, one can also perform the robust estimation of the orthogonal subsets of $\\mathcal{N}$ , followed by solving $[\\mathsf{e}_x, \\mathsf{e}_y, \\mathsf{e}_z] = \\mathsf{R}[\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z], \\forall \\mathsf{n}_x \\in \\mathcal{N}_x, \\mathsf{n}_y \\in \\mathcal{N}_y, \\mathsf{n}_z \\in \\mathcal{N}_z$ for $\\mathsf{R} \\in SO(3)$ . For computational reasons, we estimate", + "bbox": [ + 75, + 765, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the robust subsets by clustering. The orthogonal subsets are then obtained by choosing the three most orthogonal clusters. Later, the obtained cluster centers are used to estimate the MF (or to enforce its existence) represented by R. Please refer to Section 4.2 for more details.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 498, + 181, + 591, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From a set of calibrated images, we model the 3D scene using the neural radiance field representation. In the process, we wish to exploit the Manhattan scene prior, without the knowledge of the Manhattan frame. This problem is addressed by jointly optimizing for the neural radiance field and the Manhattan frame estimation. The complete pipeline of our method is illustrated in Figure 2. As shown, our method consists of three units: (a) Explicit normal modeling, (b) Robust estimation of orthogonal normals, and (3) Self-supervision by Manhattan prior. In the first stage, we shoot a batch of three rays which allows us to estimate the surface normal using an algebraic method. As the estimated normals are bound to be noisy (with outliers), we perform their clustering to obtain the most orthogonal clusters (representing the Manhattan frame) in a robust manner. The obtained clusters are then used to estimate the sought MF, in the form of a rotation matrix, using the method discussed in Section 3. In the final step, we use the estimated Manhattan frame to encourage the close-by normals to be Manhattan-like and to enforce a stricter orthogonality constraint for self-supervision. In the following, we will present the details of three units of our method in different Subsections.", + "bbox": [ + 496, + 207, + 893, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Explicit Normal Modelling", + "text_level": 1, + "bbox": [ + 498, + 549, + 743, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At any given view, we consider the color $c = C_{\\theta}(\\mathsf{r})$ and depth $d = D_{\\theta}(\\mathsf{r})$ for a ray $\\mathsf{r}$ emanating from the corresponding camera center $\\mathsf{o}$ is obtained using the volumetric rendering of the implicit neural radiance field, represented by a neural network parameterized by $\\theta$ as in [24]. Then, the location of the 3D surface point is given by, $\\mathsf{x} = \\mathsf{o} + d.\\mathsf{r}$ . We process triplets of three neighboring rays. Let $\\mathcal{T} = \\{\\mathsf{r}_1,\\mathsf{r}_2,\\mathsf{r}_3\\}$ be such a triplet, whose corresponding surface points are given by $\\mathcal{X} = \\{\\times_1,\\times_2,\\times_3\\}$ . Now, for the ray triplet $\\mathcal{T}$ , we obtain the explicit surface normal $\\mathsf{n}$ with the help of the point triplet $\\mathcal{X}$ , using the following mapping and algebraic operation,", + "bbox": [ + 496, + 573, + 893, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} \\rightarrow \\mathcal {X} \\rightarrow v = \\left(x _ {1} - x _ {2}\\right) \\times \\left(x _ {1} - x _ {3}\\right)\\rightarrow n = \\frac {\\operatorname {s i g n} (\\mathrm {o} ^ {\\mathrm {T}} v) v}{\\| v \\|}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 766, + 893, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that $\\nu$ is a vector orthogonal to the plane passing through 3D points in triplet $\\mathcal{T}$ . We obtain the oriented normal $\\mathsf{n}$ by normalizing $\\nu$ and correcting its sign by ensuring that the camera center $\\mathsf{o}$ is in front of the estimated plane. A graphical illustration of the surface normal estimation is provided in Figure 1 on the left. We select a random set", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17862", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d3faeefa59b922e950680a21324596f55cda204969108ec9d3f44a70a0fcc6e2.jpg", + "image_caption": [ + "Figure 2: The complete pipeline of our method. We use grid features (GF) and an MLP to represent the radiance field. The explicit normals are derived using the depths obtained from volume rendering. The Manhattan scene prior is exploited by clustering the estimated normals to enforce their orthogonality." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 869, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of ray triplets $\\{\\mathcal{T}_i\\}_{i = 1}^m$ from multiple cameras. These ray triplets provide us the surface point triplets $\\{\\mathcal{X}_i\\}_{i = 1}^m$ . We use point triplets to explicitly derive the surface normals $\\mathcal{N} = \\{\\mathfrak{n}_i\\}_{i = 1}^m$ , using (1). These normals are later used to recover the unknown Manhattan frame. The simplicity of the explicit normals computed in this paper makes them easy to compute and handle. If needed, normals of different sizes of surface regions could also be computed similarly.", + "bbox": [ + 75, + 277, + 472, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Robust Estimation of Orthogonal Normals", + "text_level": 1, + "bbox": [ + 76, + 406, + 439, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We are interested to recover the MF from a set of noisy surface normals $\\mathcal{N} = \\{\\mathfrak{n}_i\\}_{i=1}^m$ . Recall Section 3, the MF can be obtained by robustly recovering a set of three orthogonal normals from $\\mathcal{N}$ . To do so, we first cluster $\\mathcal{N}$ into $k$ clusters $\\{\\mathcal{C}_i\\}_{i=1}^k$ , by using the well-known $k$ -means clustering algorithm [5]. During the clustering, every centroid is $L_2$ normalized after each iteration, to ensure that they are unit vectors to represent surface normals. In a perfect Manhattan world, there exist only six clusters corresponding to the three orthogonal MF axes $\\mathcal{E} = \\{\\mathbf{e}_x, \\mathbf{e}_y, \\mathbf{e}_z\\}$ and their parallel counterparts. However, real scenes consist of surfaces of different orientations. An additional source of non-Manhattan normals comes from the inaccuracy in the normal estimation. Therefore, we use a clustering technique to robustly recover the orthogonal normals. In the following, we present how the orthogonal clusters are selected from the set of clusters $\\{\\mathcal{C}_i\\}_{i=1}^k$ . Note from Section 3, the selected orthogonal clusters are considered to represent MF defining sets $\\mathcal{N}_x, \\mathcal{N}_y$ , and $\\mathcal{N}_z$ . We proceed by first selecting three orthogonal clusters, say $\\mathcal{N}_1, \\mathcal{N}_2$ , and $\\mathcal{N}_3$ . Later, we assign them to $\\mathcal{N}_x, \\mathcal{N}_y$ , and $\\mathcal{N}_z$ to recover MF.", + "bbox": [ + 75, + 429, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For notational ease, we pair clusters and centroids as $\\mathcal{U} = \\left\\{\\left(\\mathcal{C}_i,\\mathsf{c}_i\\right)\\right\\}_{i = 1}^k$ , where the centroids are computed by taking the average across the corresponding cluster such that $\\mathsf{c}_i = \\frac{1}{|\\mathcal{C}_i|}\\sum_{\\mathsf{n}\\in \\mathcal{C}_i}\\mathsf{n},$ followed by normalizing to a unit vector $\\mathsf{c}_i = \\frac{\\mathsf{c}_i}{\\|\\mathsf{c}_i\\|}$ . Leveraging the Manhattan prior through the assumption that Manhattan surfaces dominate the scene, we pick $\\mathsf{n}_1 = \\mathsf{c}_f$ , where $f = \\arg \\max_{i}|\\mathcal{C}_i|$ In other words, we pick the centroid of the largest cluster, as one of the MF axes. Then, we obtain $\\mathsf{n}_2 = \\mathsf{c}_s$ and $\\mathsf{n}_3 = \\mathsf{c}_t$ as a solution of the optimization problem", + "bbox": [ + 75, + 746, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 $(\\mathcal{N}_1,\\mathcal{N}_2,\\mathcal{N}_3,\\mathsf{R}) =$ findManhattanFrame(N)", + "text_level": 1, + "bbox": [ + 500, + 277, + 854, + 292 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Cluster normals $\\mathcal{N}$ into, $\\mathcal{U} = \\{(\\mathcal{C}_i,\\mathsf{c}_i)\\}_{i = 1}^k$ using k-means.", + "2. For the largest cluster $\\mathcal{C} \\in \\mathcal{U}$ , assign $(\\mathcal{C}, \\mathfrak{c}) \\to (\\mathcal{N}_1, \\mathfrak{n}_1)$ .", + "3. Find $\\mathcal{C}_s\\in \\mathcal{U},\\mathcal{C}_t\\in \\mathcal{U}$ minimizing $|\\mathsf{c}_s^\\top \\mathsf{n}_1| + |\\mathsf{n}_1^\\top \\mathsf{c}_t| + |\\mathsf{c}_s^\\top \\mathsf{c}_t|$", + "4. Assign $(\\mathcal{C}_s,\\mathsf{c}_s)\\to (\\mathcal{N}_2,\\mathsf{n}_2),(\\mathcal{C}_t,\\mathsf{c}_t)\\to (\\mathcal{N}_3,.,\\mathsf{n}_3).$", + "5. $|\\mathbf{e}_z^{\\mathsf{T}}\\mathfrak{n}_1|\\leq \\frac{1}{\\sqrt{2}}?\\mathfrak{n}_1\\to \\mathfrak{n}_z:(|\\mathbf{e}_y^{\\mathsf{T}}\\mathfrak{n}_1|\\leq \\frac{1}{\\sqrt{2}}?\\mathfrak{n}_1\\to \\mathfrak{n}_y:\\mathfrak{n}_1\\to \\mathfrak{n}_x).$", + "6. Perform the remaining $\\{\\mathsf{n}_1,\\mathsf{n}_2,\\mathsf{n}_3\\} \\to \\{\\mathsf{n}_x,\\mathsf{n}_y,\\mathsf{n}_z\\}$" + ], + "bbox": [ + 514, + 297, + 883, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as in step 5 to the closest corresponding canonical axes.", + "bbox": [ + 532, + 385, + 861, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "7. Return $\\mathcal{N}_1, \\mathcal{N}_2, \\mathcal{N}_3$ , and $\\mathsf{R} = [\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z]^\\top$ .", + "bbox": [ + 514, + 398, + 787, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$s, t = \\arg \\min_{i,j} |\\mathsf{c}_i^\\top \\mathsf{n}_1| + |\\mathsf{n}_1^\\top \\mathsf{c}_j| + |\\mathsf{c}_i^\\top \\mathsf{c}_j|$ . In other words, we find two additional cluster centroids providing the most orthogonal triplet. We further merge all selected clusters with their opposites. This is achieved by comparing all cluster pairs. Whenever two centroids are nearby, but opposite in sign, the corresponding clusters are merged with the appropriate sign correction. The procedure of finding $\\{\\mathcal{N}_1, \\mathcal{N}_2, \\mathcal{N}_3\\}$ is illustrated in Figure 1 (middle and right), and is summarized in Algorithm 1 from step 1-4. Although the orthogonal clusters with their centroids are sufficient to exploit the desired Manhattan prior, we may wish to recover the MF in the form of a rotation matrix. As such, any arrangement of $\\{\\mathsf{n}_1, \\mathsf{n}_2, \\mathsf{n}_3\\}$ as a valid rotation matrix offers us a valid MF. We may however often be interested to recover the one which is closest to the world frame. For this reason, we suggest keeping the largest cluster's centroid paired to the closes canonical axis in $\\mathcal{E}$ . Similarly, we also align one more axis, whereas the last remaining axis gets paired by default. We summarize how to recover MF closest to the world frame in Algorithm 1 in steps 5-7.", + "bbox": [ + 496, + 425, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Self-supervision by Manhattan Prior", + "text_level": 1, + "bbox": [ + 498, + 738, + 818, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The supervision of the implicit neural radiance field, parameterized by $\\theta$ , is achieved through jointly optimizing three loss terms. The first one is the photometric loss computed as follows, $\\mathcal{L}_{img} = \\frac{1}{|\\mathcal{R}|}\\sum_{\\mathbf{r}\\in \\mathcal{R}}\\| C_{\\theta}(\\mathbf{r}) - C(\\mathbf{r})\\| _2^2$ where $C(\\mathbf{r})$ is the ground-truth color, and $\\mathcal{R}$ is the set of rays going through sampled pixel triplets. This loss term is responsible for facilitating the learning of the implicit 3D representation of the investigated scene [24].", + "bbox": [ + 496, + 762, + 890, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Losses from Manhattan prior: We exploit the Manhattan", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17863", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scene prior for self-supervision, in order to improve the implicit 3D representation without additional ground truth labels. We do so by using the clusters $\\{\\mathcal{N}_1,\\mathcal{N}_2,\\mathcal{N}_3\\}$ obtained from Algorithm 1. More precisely, for cluster-centroid pairs $\\{(\\mathcal{N}_i,\\mathsf{n}_i)\\}_{i = 1}^3$ we use the following two losses,", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c t r} = \\frac {1}{3} \\sum_ {i = 1} ^ {3} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {\\mathfrak {n} \\in \\mathcal {N} _ {i}} \\| 1 - \\mathfrak {n} _ {i} ^ {\\top} \\mathfrak {n} \\| _ {1} + \\| \\mathfrak {n} _ {i} - \\mathfrak {n} \\| _ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 179, + 468, + 219 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o r t} = \\frac {1}{3} \\left(\\left| n _ {1} ^ {\\top} n _ {2} \\right| + \\left| n _ {1} ^ {\\top} n _ {3} \\right| + \\left| n _ {2} ^ {\\top} n _ {3} \\right|\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 223, + 468, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, the loss $\\mathcal{L}_{ctr}$ encourages tighter clusters, while the loss $\\mathcal{L}_{ort}$ enforces the orthogonality among the three clusters. The final loss used to optimize $\\theta$ is then given by $\\mathcal{L} = \\mathcal{L}_{img} + \\lambda_{ctr}\\mathcal{L}_{ctr} + \\lambda_{ort}\\mathcal{L}_{ort}$ , where $\\lambda_{ctr}$ and $\\lambda_{ort}$ are the hyperparameters. Please, refer to Figure 2 for a schematic summary of all three losses.", + "bbox": [ + 76, + 262, + 468, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 367, + 209, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Baselines, Metrics, and Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 392, + 468, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "InstantNGP [25] (baseline): This method represents the scene as a multi-resolution voxel grid and leverages a hash table of trainable feature vectors, which are used to represent grid elements. The representation is further processed with a small MLP. This allows InstantNGP to be very computationally efficient, requiring in our case around 30 minutes per scene to train and evaluate. Therefore, we use it as the backbone of our method and as our baseline.", + "bbox": [ + 75, + 415, + 468, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ManhattanDF [14]: This method exploits the Manhattan prior by supervising the explicit normals of the floors to align with the known $\\mathsf{n}_z$ , as well as by supervising normals of the walls to align with two learned orthogonal axes which are also orthogonal to $\\mathsf{n}_z$ . To achieve this, apart from the RGB images, this method relies on knowing the wall and floor semantics, as well as knowing the exact floor axis $\\mathsf{n}_z$ (MF partially known). We implemented it on top of the In-stantNGP backbone with density field estimation.", + "bbox": [ + 75, + 537, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ours: We implement our proposed method from Section 4 on top of InstantNGP with density field estimation.", + "bbox": [ + 75, + 672, + 468, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ours (MF known): We modify our method by assuming that the MF is fully known. We do so by adding additional loss terms $\\mathcal{L}_{man_i} = \\|1 - \\mathsf{n}_i^{\\mathrm{T}}\\mathsf{m}\\|_1 + \\| \\mathsf{n}_i - \\mathsf{m}\\|_1$ for each orthogonal cluster centroid $\\mathsf{n}_i$ and its closest MF axis $\\mathfrak{m}$ (or a closer opposite counterpart). Thus, we explicitly guide the orthogonal triplet to align with the known MF.", + "bbox": [ + 75, + 704, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics: To quantitatively evaluate novel view rendering, we use peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) [48]. To evaluate the extracted surface normals of novel rendered views, we use the median angular error. To partially evaluate the quality of the learned implicit 3D structure, we utilize the mean absolute error (MAE) and the root mean square error (RMSE) on the", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "depth obtained by volume rendering. Finally, to evaluate the recovered MF, we calculate the absolute error between the yaw, pitch, and roll angles of the recovered frame and the MF. All metrics are averaged across scenes, except for the yaw, pitch, and roll errors for which the median is reported.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details: We turn on $\\mathcal{L}_{\\text {ort }}$ and $\\mathcal{L}_{\\text {centr }}$ after 500 steps and linearly increase their weights to the specified values over the next 2500 steps. Also, we randomly sample rays for one-third of every batch size and select their left and upper neighbor to form a triplet to facilitate obtaining explicit surface normals. For other implementation details, please refer to the supplementary material.", + "bbox": [ + 496, + 167, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that in addition to RGB images, ManhattanDF [14] requires knowing the wall and floor semantics, and the exact floor axis $\\mathsf{n}_z$ . In contrast, our method only requires RGB images and the assumed Manhattan prior to holding true. Therefore, this comparison aims to get a better insight into what can be achieved without leveraging additional labels.", + "bbox": [ + 496, + 273, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Datasets", + "text_level": 1, + "bbox": [ + 500, + 377, + 601, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hypersim [31] is a photorealistic synthetic dataset consisting of indoor scenes. It was created by leveraging a large repository of scenes created by professional artists, with 461 indoor scenes in total. This dataset is geometry-rich, containing a lot of details and lighting sources. Each scene has one or more camera trajectories available, where each trajectory has up to 100 views rendered in $1024 \\times 768$ . For each scene, camera calibration information is provided, as well as detailed per-pixel labels such as depth and surface normals. After cleaning up the scenes by discarding a few with insufficient camera views, as well as other problems, we were left with 435 scenes. We also kept only one camera trajectory per scene. We then evaluated the InstantNGP [25] baseline on all 435 scenes and made three divisions based on its PSNR on unseen views. Hypersim-A contains 20 scenes where the baseline was performing well above average, Hypersim-B contains 20 scenes where the baseline had near-average performance, and Hypersim-C contains 10 scenes where the baseline had a below-average performance. For each scene, we randomly assigned half views to the training split and the rest to the test split. We note that the test split often contains views that were partially or completely unobserved during training. We also note that Hypersim is visually very realistic, geometry-rich, and challenging, considering it is synthetic. This can be subjectively observed by inspecting rendered views.", + "bbox": [ + 496, + 401, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ScanNet [8] is a real-world dataset consisting of indoor scenes. It was collected using a scalable RGB-D capture system that includes automated surface reconstruction and crowd-sourced semantic annotation. The dataset contains 1613 indoor scenes, which are annotated with ground-truth camera poses, surface reconstructions, and instance-level semantic segmentations. We use three scenes that were used", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17864", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fcb4dae230da0a5edc477ae9bc02dd022a4f47d8cac5743f2b4d9ad5d9826736.jpg", + "table_caption": [ + "Table 1: Experiments on Hypersim. We observe that our method consistently outperforms the baseline, as well as the ManhattanDF. This is very interesting since, unlike ManhattanDF, we do not use any additional labels during training. Finally, we see that it does not matter for our method whether the MF is known beforehand or not. Therefore, the additional knowledge of MF is neither necessary nor helpful." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑Normals°↓Pitch°↓Roll°↓Yaw°↓D-MAE↓D-RMSE↓
Scenes AInstantNGP [25] (baseline)25.860.87157.126.186.4619.250.0640.102
ManhattanDF [14]26.510.86840.691.231.015.080.0530.087
Ours27.200.86437.300.400.500.520.0530.093
Ours (MF known)27.210.85635.590.250.260.450.0520.091
Scenes BInstantNGP [25] (baseline)20.750.81160.126.067.9215.870.1050.151
ManhattanDF [14]21.870.82650.502.552.0611.690.0790.121
Ours22.450.81654.081.191.351.810.0800.127
Ours (MF known)22.510.81350.590.510.650.550.0780.126
Scenes CInstantNGP [25] (baseline)17.790.74064.297.454.5515.140.1300.174
ManhattanDF [14]18.330.77056.083.203.4110.250.1030.147
Ours19.430.76854.795.372.244.240.0940.133
Ours (MF known)19.290.76455.123.644.039.480.0940.135
194 scenesInstantNGP [25] (baseline)20.470.78361.346.566.9921.750.1040.146
ManhattanDF [14]20.940.79452.811.722.3213.480.0970.140
Ours21.630.78652.011.871.943.770.0850.126
", + "bbox": [ + 106, + 135, + 859, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in [14], where one-tenth of all views were uniformly sampled, leaving 303-477 views per scene. For each scene, the training and test split both contain half of the total views.", + "bbox": [ + 75, + 359, + 468, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Replica [38] is a synthetic dataset featuring a diverse set of 18 indoor scenes. Each scene is equipped with photorealistic textures, allowing one to render realistic images from arbitrary camera poses. We use five scenes that were rendered and prepared in [58], where the rendered views also contain semantic segmentation labels. Each scene contains 900 views generated from random 6-DOF trajectories similar to hand-held camera motions in $640 \\times 480$ resolution. We select 75 views for training and 75 for testing in each scene.", + "bbox": [ + 75, + 406, + 470, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Results", + "text_level": 1, + "bbox": [ + 76, + 558, + 171, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments on Hypersim: We summarize the experiments performed on Hypersim in Table 1. Our proposed method achieves clear improvements in comparison to the InstantNGP baseline, in terms of novel-view rendering, normals estimation, depth estimation, as well as MF recovery. This is consistent across different scene difficulties, namely splits A, B, and C. Furthermore, our method also outperforms the ManhattanDF SotA. This is very interesting since, unlike ManhattanDF, we do not use any additional labels during training, other than the ground truth RGB. The ManhattanDF has partial access to ground truth MF axes, as well as wall and floor semantics, which directly implies having sparse surface normals ground truth. However, there are lots of Manhattan objects and areas in realistic scenes such as in Hypersim, other than the walls and floors. Our self-supervised method leverages the presence of many such objects and areas, to facilitate imposing geometrical constraints on the implicit representation during learning. This can be visually observed in Figure 3. Moreover, we observe that it does not matter for our method whether the MF is known beforehand, since our method recovers the", + "bbox": [ + 75, + 583, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/931cc711d1421ab11c9a7fda1fcc78f8f7fed9e3591eae4d34bbccc38f485229.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 359, + 694, + 446 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/712b45e1e66298b516e9dde262be8b9d4f5749c2a89ceb19f419e817ffea2ed6.jpg", + "image_caption": [ + "(a) InstantNGP [25] (baseline)", + "(c) Ours", + "Figure 3: Leveraging the Manhattan prior (Hypersim-A). Our method leverages the presence of many Manhattan objects and areas in the scene in a self-supervised fashion, which facilitates imposing geometrical constraints during learning. Our method offers plausible normals – sometimes with missing details – that help to better model the radiance fields. In contrast, ManhattanDF leverages the Manhattan prior through labels of only walls and floors." + ], + "image_footnote": [], + "bbox": [ + 501, + 465, + 694, + 551 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16bba075f6b5187fecd171f0d629f1cdfc5791a03dab16f32f4405548dc9dd55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 359, + 890, + 446 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3eb669f9bee957446c057f629488a204a95f5f16a38826bd2ec3710f28e4e2b.jpg", + "image_caption": [ + "(b) ManhattanDF [14]", + "(d) Grond truth" + ], + "image_footnote": [], + "bbox": [ + 741, + 467, + 885, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MF in the clustering step automatically, and therefore similar performance is achieved in both cases. This is particularly exciting because our self-supervised method performs similarly to the supervised one (using the groundtruth MF). Finally, we show large-scale experiments on a larger set of 194 scenes, which lead to the same conclusions. We observed that the semantic loss of the ManhattanDF sometimes causes convergence issues on difficult scenes, which we partially alleviated with class weighting and label smoothing. Therefore, we report results on only 194 scenes where ManhattanDF converged. Additional results on all 435 scenes are provided in the supplementary.", + "bbox": [ + 496, + 689, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments on ScanNet: In Table 2, we examine the behavior of our proposed method on real-world indoor scenes.", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17865", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f03d4b048099205cde62a4252f74d5eed9eb1b2da95cea442227198e5b90389.jpg", + "image_caption": [ + "Figure 4: Qualitative results. Our method leverages many Manhattan objects and surfaces in the scene, which improves the implicit geometrical representation compared to the baseline. Unlike ManhattanDF, our method relies on many cues other than the walls & floors, which leads to a better representation of some objects (e.g. the white table in the bottom right corner of the example from columns 1-3). For more qualitative results, please refer to the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 883, + 381 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cf0341691a7b9bfa5729bb90d193db3e989a73b3a48a1b4b00acc03fba60a65c.jpg", + "table_caption": [ + "Table 2: Experiments on real-world ScanNet data. Our method outperforms both the baseline and ManhattanDF. This is the case both with and without supervising with sparse depth from SfM." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM ↑Depth↓-MAE
InstantNGP [25] (baseline)17.780.5870.119
RegNeRF [26]18.730.6180.102
ManhattanDF [14]18.680.6140.112
Ours20.790.6430.072
+ additional sparse depth from SfM
InstantNGP [25] (baseline)20.700.6310.048
ManhattanDF [14]21.530.6400.052
Ours22.250.6670.033
", + "bbox": [ + 83, + 496, + 460, + 617 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8671ba17b52bee8000d7a534316c372aef89032d485420de16837ba776e79574.jpg", + "table_caption": [ + "Table 3: Experiments on Replica. Our method outperforms the baseline, and it shows similar performance as ManhattanDF, which leverages additional labels during training." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM ↑Depth↓-MAE
InstantNGP [25] (baseline)34.300.9440.022
Semantic-NeRF [58]34.080.9380.014
ManhattanDF [14]35.240.9440.008
Ours35.130.9440.011
", + "bbox": [ + 83, + 667, + 460, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method achieves clear improvements in comparison to the InstantNGP baseline, as well as to ManhattanDF, in terms of all measured metrics. Additionally, following the experimental setting of ManhattanDF [14], we train all methods with additional sparse depth supervision – where the sparse depth is obtained from the SfM pipeline. Our proposed method is also superior in this setting.", + "bbox": [ + 75, + 750, + 468, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiments on Replica: We summarize experiments performed on the Replica dataset in Table 3. It can be observed that the InstantNGP baseline already performs very", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20b87505bcc4efa99a6f596cbe137a2db8ddb6c04dcb80aee8ec3af502bf34e7.jpg", + "image_caption": [ + "Improvements of our method w.r.t. scene difficulty", + "Figure 5: Improvements vs. scene difficulty. Our method achieves the biggest improvements on scenes of hard and moderate difficulty, by leveraging many Manhattan objects and surfaces." + ], + "image_footnote": [], + "bbox": [ + 516, + 460, + 870, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "well on this dataset, leaving not much room for further improvements. Therefore, we consider Replica as an easier dataset. Nevertheless, our method still performs better than the baseline. As expected, ManhattanDF also improves the baseline similarly. Recall that the ManhattanDF uses additional labels for supervision. We also note that Replica has noticeably more walls and floors, and less of other objects, compared to other more complex datasets.", + "bbox": [ + 496, + 628, + 890, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results: We depict the visual results of the discussed experiments in Figure 4. Our method leverages many Manhattan objects and surfaces in the scene, which improves the geometrical structure of 3D compared to the InstantNGP baseline. This is visible in surface normals and depth, obtained using volume rendering. Furthermore, we observe that, unlike ManhattanDF, our method relies on many Manhattan cues other than the walls & floors. This leads to a better representation of some such objects, e.g. the white table in the bottom right corner of the example", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17866", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0ce2ab559bfd8f9e784ac719d77cb8e1839ba8b7bbf8dc1a873ce7cec4c347ba.jpg", + "table_caption": [ + "Table 4: Sparse training views. Our proposed method clearly exhibits the best performance, for all cases of input view sparsity." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM ↑Depth↓-MAE
12 viewsInstantNGP [25]18.020.7060.138
ManhattanDF [14]19.450.7500.116
Ours20.500.7600.104
9 viewsInstantNGP [25]16.790.6610.154
ManhattanDF [14]18.040.7140.130
Ours19.140.7280.120
6 viewsInstantNGP [25]15.750.5820.178
ManhattanDF [14]16.000.6390.159
Ours16.670.6670.158
", + "bbox": [ + 80, + 107, + 462, + 239 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d0b92525a1588d91f51e30861c5a3f75a5f297b137937512c854f573d46ddd65.jpg", + "table_caption": [ + "Table 5: Ablation study on Hypersim-A. Both of our proposed loss terms contribute to the overall performance." + ], + "table_footnote": [], + "table_body": "
PSNR↑Norm.°↓Depth↓Rot.°↓
Only Limg25.8657.120.06410.63
+ Lort27.2150.070.0585.39
+ Lctr27.0636.090.0530.57
+ Lort + Lctr (Ours)27.2037.300.0530.47
Ours + MF known27.2135.590.0520.32
Ours (no delay)27.0635.780.0540.39
Ours (no w lin.)27.0137.860.0550.60
", + "bbox": [ + 78, + 276, + 467, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from columns 1-3. Moreover, we observe that our method is able to cope with difficult scenes and views, where other methods struggle. For more qualitative results, please refer to the supplementary material.", + "bbox": [ + 75, + 401, + 468, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Improvements with respect to scene difficulty: We analyze the improvements by our method for different scene difficulties. We decided on scene difficulty based on the novel-view rendering performance of the InstantNGP baseline. In Figure 5, we see that our method brings the most benefits for scenes of hard and moderate difficulties, thanks to the Manhattan scene prior.", + "bbox": [ + 75, + 462, + 468, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse training views: In order to gain more insight, we examine the behavior of training neural radiance fields with sparse training input views. The results on the Hypersim-A dataset are presented in Table 4. Our proposed method clearly outperforms the InstantNGP baseline, as well as ManhattanDF, when trained with 6, 9, and 12 input views.", + "bbox": [ + 75, + 566, + 468, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finding the MF: We test the robustness of our proposed method for finding the Manhattan frame, by introducing a simultaneous offset $\\alpha$ in the yaw, pitch, and roll on the canonical MF. The experiments are performed on Hypersim A and can be found in Figure 6. In Figure 6a we observe that the novel-view rendering quality remains largely robust to the rotation offset $\\alpha$ . We note that we increase the scene bounding box by the same factor for all experiments in Figure 6, to make sure that objects remain within the voxel grid for maximal $\\alpha$ . This slightly decreases the resolution of grid element representations, so PSNR is slightly lower than in Table 1. Furthermore, in Figure 6b we see that the MF estimation remains robust to the rotation offset.", + "bbox": [ + 75, + 657, + 468, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study: We report our ablation study in Table 5. Both of our proposed losses contribute to the overall performance. Furthermore, turning on $\\mathcal{L}_{\\text{ort}}$ and $\\mathcal{L}_{\\text{centr}}$ after", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c81a6c5747db5b77d5cf4c8f7a5a0a2d9c326fdac240db57c8ed609ff0fc073c.jpg", + "image_caption": [ + "Rotation offset effects on novel-view rendering", + "(a) Novel-view rendering quality remains robust to the rotation offset $\\alpha$ ." + ], + "image_footnote": [], + "bbox": [ + 542, + 101, + 846, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa753d85a2f79945af025e76515d98191ad5919d49b0f2de87582429492c77a2.jpg", + "image_caption": [ + "Rotation offset effects on Manhattan frame estimation", + "(b) MF estimation remains robust to the rotation offset $\\alpha$", + "Figure 6: Finding the MF. We test the robustness of our method by introducing the rotation offsets on the canonical MF." + ], + "image_footnote": [], + "bbox": [ + 542, + 202, + 844, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "500 steps, and linearly increasing their weight also helps slightly. Please refer to the supplementary for more details.", + "bbox": [ + 500, + 340, + 890, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 383, + 617, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrated the possibility of exploiting the Manhattan scene prior without needing any additional supervision. This is achieved by performing robust clustering of explicit normals, followed by the search of the Manhattan frame (MF) whose existence is based on the assumed prior. The sought MF is obtained from the orthogonal clusters, which are later used to self-supervise the neural representation learning. Our self-supervision encourages the normals of Manhattan surfaces to group into three orthogonal directions. Our experiments on three indoor datasets demonstrate that the proposed method not only benefits from building parts (such as walls and floors) but also exploits many other Manhattan scene parts (such as furniture). Both quantitative and qualitative evaluations reveal the benefit of the proposed method in terms of, improved performance over the established baselines and competitive results to state-of-the-art methods that use additional labels for supervision. Our method has the potential to be extended in other higher expressiveness scene priors, such as Atlanta world and the mixture of Manhattan frames.", + "bbox": [ + 496, + 409, + 890, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: One limitation of our method is that it sometimes produces surface normals with missing details or \"blocky\" artifacts. Nevertheless, this usually offers better novel-view RGB rendering, compared to not imposing any structure priors. Another limitation is that our method is not beneficial for very easy scenes. For a detailed discussion of limitations, please refer to the supplementary material.", + "bbox": [ + 496, + 710, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 830, + 658, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was co-financed by Innosuisse under the project Know Where To Look, Grant 59189.1 IP-ICT. Also, Carlos Oliveira helped with splendid visual renderings.", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17867", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022.", + "[2] Stephen T Barnard. Interpreting perspective images. Artificial intelligence, 21(4):435-462, 1983.", + "[3] Jean-Charles Bazin, Hongdong Li, In So Kweon, Cedric Demonceaux, Pascal Vasseur, and Katsushi Ikeuchi. A branch-and-bound approach to correspondence and grouping problems. IEEE transactions on pattern analysis and machine intelligence, 35(7):1565-1576, 2012.", + "[4] Jean-Charles Bazin, Yongduek Seo, Cédric Demonceaux, Pascal Vasseur, Katsushi Ikeuchi, Inso Kweon, and Marc Pollefeys. Globally optimal line clustering and vanishing point estimation in manhattan world. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 638–645. IEEE, 2012.", + "[5] Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag, Berlin, Heidelberg, 2006.", + "[6] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022.", + "[7] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 941-947. IEEE, 1999.", + "[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017.", + "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022.", + "[10] Patrick Denis, James H Elder, and Francisco J Estrada. Efficient edge-based methods for estimating Manhattan frames in urban imagery. In European conference on computer vision, pages 197-210. Springer, 2008.", + "[11] Clara Fernandez-Labrador. Indoor scene understanding using non-conventional cameras. PhD thesis, 2020.", + "[12] Yasutaka Furukawa, Brian Curless, Steven M Seitz, and Richard Szeliski. Manhattan-world stereo. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1422-1429. IEEE, 2009.", + "[13] Wuwei Ge, Yu Song, Baichao Zhang, and Zehua Dong. Globally optimal and efficient Manhattan frame estimation by delimiting rotation search space. In Proceedings of the" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 15213-15221, 2021.", + "[14] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022.", + "[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021.", + "[16] Kyungdon Joo, Tae-Hyun Oh, Junsik Kim, and In So Kweon. Globally optimal Manhattan frame estimation in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1763-1771, 2016.", + "[17] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022.", + "[18] Fu Li, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerf-posed: A first-reconstruct-thenregress approach for weakly-supervised 6d object pose estimation. arXiv preprint arXiv:2203.04802, 2022.", + "[19] Haoang Li, Jian Yao, Jean-Charles Bazin, Xiaohu Lu, Yazhou Xing, and Kang Liu. A monocular slam system leveraging structural regularity in manhattan world. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 2518-2525. IEEE, 2018.", + "[20] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021.", + "[21] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15702-15712, 2022.", + "[22] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021.", + "[23] Gerard F McLean and D Kotturi. Vanishing point detection by line clustering. IEEE Transactions on pattern analysis and machine intelligence, 17(11):1090-1095, 1995.", + "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. cite arxiv:2003.08934Comment: ECCV 2020 (oral). Project page with videos and code: http://tancik.com/nerf.", + "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022." + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17868", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022.", + "[27] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021.", + "[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021.", + "[29] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in manhattan world. In Proceedings of the IEEE International Conference on Computer Vision, pages 882-890, 2017.", + "[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021.", + "[31] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021.", + "[32] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022.", + "[33] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022.", + "[34] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lučić, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6229-6238, 2022.", + "[35] Gautam Singh and Jana Kosecka. Visual loop closing using gist descriptors in manhattan world. In ICRA omnidirectional vision workshop, pages 4042-4047, 2010.", + "[36] Julian Straub, Nishchal Bhandari, John J Leonard, and John W Fisher. Real-time manhattan world rotation estimation in 3d. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1913-1920. IEEE, 2015.", + "[37] Julian Straub, Oren Freifeld, Guy Rosman, John J Leonard, and John W Fisher. The Manhattan frame model—manhattan world inference in the space of surface normals. IEEE transactions on pattern analysis and machine intelligence, 40(1):235-249, 2017." + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.", + "[39] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021.", + "[40] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light field neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8269-8279, 2022.", + "[41] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2022.", + "[42] Frank A Van den Heuvel. 3d reconstruction from a single image using geometric constraints. ISPRS Journal of Photogrammetry and Remote Sensing, 53(6):354-368, 1998.", + "[43] Carlos A Vanegas, Daniel G Aliaga, and Bedrich Benes. Building reconstruction using manhattan-world grammars. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 358-365. IEEE, 2010.", + "[44] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022.", + "[45] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021.", + "[46] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. arXiv preprint arXiv:2206.13597, 2022.", + "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021.", + "[48] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004.", + "[49] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17869", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Horst Wildenauer and Allan Hanbury. Robust camera self-calibration from monocular images of manhattan worlds. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2831-2838. IEEE, 2012.", + "[51] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022.", + "[52] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K Wong. Ps-nerf: Neural inverse rendering for multi-view photometric stereo. In European Conference on Computer Vision, pages 266–284. Springer, 2022.", + "[53] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K Wong. S3-nerf: Neural reflectance field from shading and shadow under a single viewpoint. arXiv preprint arXiv:2210.08936, 2022.", + "[54] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021.", + "[55] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021.", + "[56] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022.", + "[57] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5657-5665, 2016.", + "[58] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021." + ], + "bbox": [ + 78, + 90, + 467, + 641 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17870", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/f357e4d1-f9a1-4f72-b3d1-9e59cc154adf_model.json b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/f357e4d1-f9a1-4f72-b3d1-9e59cc154adf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ad67adc617ea45158c826a28925b334cff72966 --- /dev/null +++ b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/f357e4d1-f9a1-4f72-b3d1-9e59cc154adf_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "Surface Normal Clustering for Implicit Representation of Manhattan Scenes" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.18, + 0.708, + 0.199 + ], + "angle": 0, + "content": "Nikola Popovic1, Danda Pani Paudel1,2, Luc Van Gool1,2" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.199, + 0.708, + 0.234 + ], + "angle": 0, + "content": "1Computer Vision Laboratory, ETH Zurich, Switzerland \n2INSAIT, Sofia University, Bulgaria" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.237, + 0.688, + 0.252 + ], + "angle": 0, + "content": "{nipopovic, paudel, vangool}@vision.ee.ethz.ch" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.665 + ], + "angle": 0, + "content": "Novel view synthesis and 3D modeling using implicit neural field representation are shown to be very effective for calibrated multi-view cameras. Such representations are known to benefit from additional geometric and semantic supervision. Most existing methods that exploit additional supervision require dense pixel-wise labels or localized scene priors. These methods cannot benefit from high-level vague scene priors provided in terms of scenes' descriptions. In this work, we aim to leverage the geometric prior of Manhattan scenes to improve the implicit neural radiance field representations. More precisely, we assume that only the knowledge of the indoor scene (under investigation) being Manhattan is known - with no additional information whatsoever - with an unknown Manhattan coordinate frame. Such high-level prior is used to self-supervise the surface normals derived explicitly in the implicit neural fields. Our modeling allows us to cluster the derived normals and exploit their orthogonality constraints for self-supervision. Our exhaustive experiments on datasets of diverse indoor scenes demonstrate the significant benefit of the proposed method over the established baselines. The source code will be available at https://github.com/nikola3794/normal-clustering-nerf." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.694, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Above \\(80\\%\\) images ever taken are estimated to involve human-made architectural structures, with a substantial share of indoor scenes [11]. These scenes often exhibit strong structural regularities, including flat and texture-poor surfaces in some axis-aligned Cartesian coordinate system - also known as Manhattan world [7, 12]. Paradoxically, these regularities may hinder the 3D modeling process if it is unaware of the human-made scene priors. In fact, several computer vision works have even benefited from the knowledge of the Manhattan world for the task of 3D scene reconstruction [43, 12], camera localization [19], self-calibration [50], and more [37, 35, 29, 57]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.53 + ], + "angle": 0, + "content": "For calibrated multi-view cameras, 3D inversion using implicit neural representations [24, 47, 54] is becoming increasingly popular due to their remarkable performance and recent efficiency developments [55, 30, 25, 51]. Meanwhile, such representations are known to benefit from additional supervision in the form of depth [1, 9, 32], normals [18], semantics [45, 58, 17, 15], local-regularization [46, 44, 26], local planar patches [21], or their combinations [56, 14, 21]. In this context, a notable recent work ManhattanSDF [14] demonstrates the benefit of exploiting the high-level geometric prior for structured scenes. More precisely, ManhattanSDF [14] uses the known semantic regions to impose the planar geometry prior of floors and walls under the Manhattan scene assumption. During this process, the exact normal of the floor and partial normals of the walls are assumed to be known, with respect to the camera coordinates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.895, + 0.732 + ], + "angle": 0, + "content": "We aim to improve the 3D neural radiance field representations for calibrated multi-view cameras in indoor Manhattan scenes, with no further assumptions. In other words, we consider that the structural and semantic information is not available, unlike ManhattanSDF [14]. In addition to the floor and walls used in ManhattanSDF [14], we wish to exploit the Manhattan prior of many other indoor scene parts (e.g. tables and wardrobes). More importantly, we consider that the Manhattan coordinate frame is also unknown. Our assumptions (of unknown semantics and Manhattan frame) on one hand make our setting very practical. On the other hand, those assumptions make the problem of exploiting the Manhattan scene prior for 3D inversion very challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.896, + 0.903 + ], + "angle": 0, + "content": "The virtue of the Manhattan world assumption comes from its simplicity, allowing us to intuitively reason about the geometry of a wide range of complex scenes/objects such as cities, buildings, and furniture. However, such reasoning often requires the axis-aligned Cartesian coordinate frame, also known as the Manhattan frame (MF), to be known [37]. Unfortunately, recovering the Manhattan frame directly from images is not an easy task [7, 37]. Therefore, several methods have been developed until recently [10, 4, 36, 16, 13] to recover the Manhattan frame, relying upon the known 3D reconstruction or image primi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "17860" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.327, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.089, + 0.677, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.085, + 0.895, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.264, + 0.895, + 0.327 + ], + "angle": 0, + "content": "Figure 1: Illustration of the proposed method. We compute one surface normal for each ray triplet, using 3D surface points derived from rendered depths (left). The computed normals (from many ray triplets) are clustered to obtain the MF (middle). The non-Manhattan and noisy surfaces are handled by a robust orthogonal normals search, which is later used for self-supervision through the Manhattan prior (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.469, + 0.417 + ], + "angle": 0, + "content": "tives (eg. lines, planes). We wish to exploit the Manhattan prior for improving the 3D representation, without needing to know MF beforehand. Instead, our experiments reveal that knowing the MF beforehand offers no additional benefit in Manhattan-prior aware radiance field representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.78 + ], + "angle": 0, + "content": "In this work, we propose a method that jointly learns the Manhattan frame and neural radiance field, from calibrated multi-view in indoor Manhattan scenes, in an end-to-end manner using the recent efficient backbone of Instant-NGP [25]. The proposed method requires no additional information to exploit the Manhattan prior and relies on the explicitly derived normals in the implicit neural fields. We use batches of three neighboring rays, whose effective surface's local piece-wise planarity is assumed to derive the explicit normals by algebraic means. In pure, complete, and enclosed Manhattan scenes, these normals form six clusters corresponding to three orthogonal and other three parallel counterpart planes. However real scenes consist of non-Manhattan scene parts and missing planes. Therefore, we use a robust method that uses minimal three orthogonal clusters to recover the sought Manhattan frame. As in the literature, we seek a rotation matrix whose entries are directly derived from the orthogonal clusters of normals, to align the Manhattan frame. The recovered MF is then used to encourage the derived normals to be axis-aligned for self-supervision. A graphical illustration of our method is presented in Figure 1. Our extensive experiments demonstrate the robustness and utility of our method in improving the implicit 3D in neural radiance field representations." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.781, + 0.454, + 0.795 + ], + "angle": 0, + "content": "The major contributions of our paper are listed below:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.841 + ], + "angle": 0, + "content": "- We address the problem of exploiting the structured-scene knowledge without requiring any dense or localized scene priors, for the first time in this paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.886 + ], + "angle": 0, + "content": "- We present a method that successfully exploits the Manhattan scene prior with an unknown Manhattan frame. The proposed method also recovers the unknown frame." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- We demonstrate the robustness and utility of our method" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.341, + 0.892, + 0.402 + ], + "angle": 0, + "content": "on three indoor datasets, where our method improves the established baselines significantly. These datasets consist of \\(200+\\) scenes, making our method tested in significantly more scenes than the state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.42, + 0.65, + 0.435 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Implicit neural representation of 3D: Since the foundational work of Mildenhall et al. [24], the implicit neural representation of 3D scenes has advanced on various fronts. These fronts include representation [54, 40], generalization [49, 28, 20, 22, 34], generation [27, 6], and efficient methods [30, 55, 25]. We rely on a recent method Instant-NGP [25] developed by Mueller et al., as our backbone. Our choice is made primarily based on the computational efficiency during both training and inference. Thanks to the offered computational efficiency, we are able to conduct large-scale experiments on several scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Auxiliary supervision methods: In addition to the images, other inputs such as depth [33, 39, 9], semantics [58, 15, 17, 41], normal [52, 53], and their combinations [56, 14, 21] are shown to be beneficial on improving the neural radiance field representation. In this regard, these auxiliary supervisions often use ground-truth labels. It is needless to say that the need for ground-truth supervision is not desired, whenever possible. Therefore, recent methods use labels predicted by some pre-trained networks [56] or recovered from the structure-from-motion (SfM) pipeline [9, 21]. One notable work ManhattanSDF [14] exploits the Manhattan prior without requiring any SfM reconstruction. However, ManhattanSDF requires (a) semantics of the scene parts and (b) the Manhattan frame, to be known. We argue that such labels required for auxiliary supervision cannot always be relied upon, due to domain gaps, poor reconstruction of texture-less regions, and additional computational needs, to list a few. Therefore, we do not use any additional labels for auxiliary supervision." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17861" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "Manhattan frame estimation: Since the early works of Bernard [2], Manhattan structure reasoning is done directly on/from images by detecting the so-called the vanishing-points (VPs) [23, 42]. In fact, the problem of detecting three orthogonal VPs is equivalent to finding the MF in 3D for the calibrated multi-view setting [3]. Note that the knowledge of Manhattan structure has been used in several computer vision works [43, 12, 19, 50, 37, 35, 29, 57]. When unknown, most methods implicitly or explicitly estimate the MF to leverage the Manhattan scene prior. In [37], Straub et al. have demonstrated that the MF can be efficiently represented in and recovered from the space of surface normals. We use a similar formulation as [37], using the surface normals derived explicitly from the implicit neural fields." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.317, + 0.33, + 0.334 + ], + "angle": 0, + "content": "3. Background and Notations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.473, + 0.553 + ], + "angle": 0, + "content": "Manhattan frame (MF) is a coordinate system that is defined by the structure building orthogonal planes of Manhattan scenes. We consider unknown MF since the scene planes and their geometric relationships are unknown. Let the unknown MF differs from the world-frame (WF), used for the 3D representation, by rotation \\(\\mathsf{R} \\in SO(3)\\). We denote three orthogonal axes in MF by \\(\\mathcal{E} = \\{\\mathsf{e}_x, \\mathsf{e}_y, \\mathsf{e}_z\\}\\). Without loss of generality, let the axes' coordinates be \\(\\mathsf{e}_x = [1,0,0]^{\\mathsf{T}}\\), \\(\\mathsf{e}_y = [0,1,0]^{\\mathsf{T}}\\), and \\(\\mathsf{e}_z = [0,0,1]^{\\mathsf{T}}\\). Note that these axes align with the normals of the respective scene building planes, in the WF. Let \\(\\mathcal{N} = \\{\\mathsf{n}_i\\}_{i=1}^m\\) be a set of 3D normals of all the scene planes. Then, the rotation \\(\\mathsf{R}\\), from world to Manhattan frame, aligns the normals \\(\\mathsf{n} \\in \\mathcal{N}\\) to the Manhattan axes \\(\\mathsf{e} \\in \\mathcal{E}\\), i.e., \\(\\mathsf{Rn}_i \\in \\mathcal{E}\\), \\(\\forall \\mathsf{n}_i \\in \\mathcal{N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.473, + 0.765 + ], + "angle": 0, + "content": "We are interested to recover \\( \\mathsf{R} \\) from a set of normals \\( \\mathcal{N} \\). For this, the above set-to-set assignment alone is not sufficient. This requires the element-wise assignment between sets \\( \\mathcal{N} \\) and \\( \\mathcal{E} \\). To do so, we divide the set \\( \\mathcal{N} \\) into three orthogonal subsets \\( \\mathcal{N}_x, \\mathcal{N}_y, \\) and \\( \\mathcal{N}_z \\). Now, for any triplet of \\( \\{\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z\\} \\) from the corresponding orthogonal subsets, we aim to establish the condition \\( [\\mathsf{e}_x, \\mathsf{e}_y, \\mathsf{e}_z] = \\mathsf{R}[\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z] \\). Note that the assignment condition results into \\( \\mathsf{R} = [\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z]^{\\top} \\). Therefore, the problem of recovering MF from a given set of normals boils down to finding three normals from orthogonal subsets. At this point, one issue regarding robustness remains pending. More precisely, we wish to recover \\( \\mathsf{R} \\) for a noisy set of normals \\( \\mathcal{N} \\), with potentially overwhelmingly many outliers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We use a robust method to recover \\( \\mathsf{R} \\) from given normals \\( \\mathcal{N} \\). In principle, \\( \\mathsf{R} \\) can be recovered from minimal two orthogonal normals, with one additional normal for disambiguation by sign correction. The recovered rotation can then be validated by consensus for robust recovery [37]. Alternatively, one can also perform the robust estimation of the orthogonal subsets of \\( \\mathcal{N} \\), followed by solving \\( [\\mathsf{e}_x, \\mathsf{e}_y, \\mathsf{e}_z] = \\mathsf{R}[\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z], \\forall \\mathsf{n}_x \\in \\mathcal{N}_x, \\mathsf{n}_y \\in \\mathcal{N}_y, \\mathsf{n}_z \\in \\mathcal{N}_z \\) for \\( \\mathsf{R} \\in SO(3) \\). For computational reasons, we estimate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "the robust subsets by clustering. The orthogonal subsets are then obtained by choosing the three most orthogonal clusters. Later, the obtained cluster centers are used to estimate the MF (or to enforce its existence) represented by R. Please refer to Section 4.2 for more details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.182, + 0.593, + 0.198 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.208, + 0.895, + 0.54 + ], + "angle": 0, + "content": "From a set of calibrated images, we model the 3D scene using the neural radiance field representation. In the process, we wish to exploit the Manhattan scene prior, without the knowledge of the Manhattan frame. This problem is addressed by jointly optimizing for the neural radiance field and the Manhattan frame estimation. The complete pipeline of our method is illustrated in Figure 2. As shown, our method consists of three units: (a) Explicit normal modeling, (b) Robust estimation of orthogonal normals, and (3) Self-supervision by Manhattan prior. In the first stage, we shoot a batch of three rays which allows us to estimate the surface normal using an algebraic method. As the estimated normals are bound to be noisy (with outliers), we perform their clustering to obtain the most orthogonal clusters (representing the Manhattan frame) in a robust manner. The obtained clusters are then used to estimate the sought MF, in the form of a rotation matrix, using the method discussed in Section 3. In the final step, we use the estimated Manhattan frame to encourage the close-by normals to be Manhattan-like and to enforce a stricter orthogonality constraint for self-supervision. In the following, we will present the details of three units of our method in different Subsections." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.55, + 0.744, + 0.567 + ], + "angle": 0, + "content": "4.1. Explicit Normal Modelling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.895, + 0.756 + ], + "angle": 0, + "content": "At any given view, we consider the color \\(c = C_{\\theta}(\\mathsf{r})\\) and depth \\(d = D_{\\theta}(\\mathsf{r})\\) for a ray \\(\\mathsf{r}\\) emanating from the corresponding camera center \\(\\mathsf{o}\\) is obtained using the volumetric rendering of the implicit neural radiance field, represented by a neural network parameterized by \\(\\theta\\) as in [24]. Then, the location of the 3D surface point is given by, \\(\\mathsf{x} = \\mathsf{o} + d.\\mathsf{r}\\). We process triplets of three neighboring rays. Let \\(\\mathcal{T} = \\{\\mathsf{r}_1,\\mathsf{r}_2,\\mathsf{r}_3\\}\\) be such a triplet, whose corresponding surface points are given by \\(\\mathcal{X} = \\{\\times_1,\\times_2,\\times_3\\}\\). Now, for the ray triplet \\(\\mathcal{T}\\), we obtain the explicit surface normal \\(\\mathsf{n}\\) with the help of the point triplet \\(\\mathcal{X}\\), using the following mapping and algebraic operation," + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.767, + 0.895, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} \\rightarrow \\mathcal {X} \\rightarrow v = \\left(x _ {1} - x _ {2}\\right) \\times \\left(x _ {1} - x _ {3}\\right)\\rightarrow n = \\frac {\\operatorname {s i g n} (\\mathrm {o} ^ {\\mathrm {T}} v) v}{\\| v \\|}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Note that \\(\\nu\\) is a vector orthogonal to the plane passing through 3D points in triplet \\(\\mathcal{T}\\). We obtain the oriented normal \\(\\mathsf{n}\\) by normalizing \\(\\nu\\) and correcting its sign by ensuring that the camera center \\(\\mathsf{o}\\) is in front of the estimated plane. A graphical illustration of the surface normal estimation is provided in Figure 1 on the left. We select a random set" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17862" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.086, + 0.87, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.217, + 0.895, + 0.263 + ], + "angle": 0, + "content": "Figure 2: The complete pipeline of our method. We use grid features (GF) and an MLP to represent the radiance field. The explicit normals are derived using the depths obtained from volume rendering. The Manhattan scene prior is exploited by clustering the estimated normals to enforce their orthogonality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.473, + 0.4 + ], + "angle": 0, + "content": "of ray triplets \\(\\{\\mathcal{T}_i\\}_{i = 1}^m\\) from multiple cameras. These ray triplets provide us the surface point triplets \\(\\{\\mathcal{X}_i\\}_{i = 1}^m\\). We use point triplets to explicitly derive the surface normals \\(\\mathcal{N} = \\{\\mathfrak{n}_i\\}_{i = 1}^m\\), using (1). These normals are later used to recover the unknown Manhattan frame. The simplicity of the explicit normals computed in this paper makes them easy to compute and handle. If needed, normals of different sizes of surface regions could also be computed similarly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.407, + 0.441, + 0.423 + ], + "angle": 0, + "content": "4.2. Robust Estimation of Orthogonal Normals" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.746 + ], + "angle": 0, + "content": "We are interested to recover the MF from a set of noisy surface normals \\(\\mathcal{N} = \\{\\mathfrak{n}_i\\}_{i=1}^m\\). Recall Section 3, the MF can be obtained by robustly recovering a set of three orthogonal normals from \\(\\mathcal{N}\\). To do so, we first cluster \\(\\mathcal{N}\\) into \\(k\\) clusters \\(\\{\\mathcal{C}_i\\}_{i=1}^k\\), by using the well-known \\(k\\)-means clustering algorithm [5]. During the clustering, every centroid is \\(L_2\\) normalized after each iteration, to ensure that they are unit vectors to represent surface normals. In a perfect Manhattan world, there exist only six clusters corresponding to the three orthogonal MF axes \\(\\mathcal{E} = \\{\\mathbf{e}_x, \\mathbf{e}_y, \\mathbf{e}_z\\}\\) and their parallel counterparts. However, real scenes consist of surfaces of different orientations. An additional source of non-Manhattan normals comes from the inaccuracy in the normal estimation. Therefore, we use a clustering technique to robustly recover the orthogonal normals. In the following, we present how the orthogonal clusters are selected from the set of clusters \\(\\{\\mathcal{C}_i\\}_{i=1}^k\\). Note from Section 3, the selected orthogonal clusters are considered to represent MF defining sets \\(\\mathcal{N}_x, \\mathcal{N}_y\\), and \\(\\mathcal{N}_z\\). We proceed by first selecting three orthogonal clusters, say \\(\\mathcal{N}_1, \\mathcal{N}_2\\), and \\(\\mathcal{N}_3\\). Later, we assign them to \\(\\mathcal{N}_x, \\mathcal{N}_y\\), and \\(\\mathcal{N}_z\\) to recover MF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.471, + 0.902 + ], + "angle": 0, + "content": "For notational ease, we pair clusters and centroids as \\(\\mathcal{U} = \\left\\{\\left(\\mathcal{C}_i,\\mathsf{c}_i\\right)\\right\\}_{i = 1}^k\\) , where the centroids are computed by taking the average across the corresponding cluster such that \\(\\mathsf{c}_i = \\frac{1}{|\\mathcal{C}_i|}\\sum_{\\mathsf{n}\\in \\mathcal{C}_i}\\mathsf{n},\\) followed by normalizing to a unit vector \\(\\mathsf{c}_i = \\frac{\\mathsf{c}_i}{\\|\\mathsf{c}_i\\|}\\) . Leveraging the Manhattan prior through the assumption that Manhattan surfaces dominate the scene, we pick \\(\\mathsf{n}_1 = \\mathsf{c}_f\\) , where \\(f = \\arg \\max_{i}|\\mathcal{C}_i|\\) In other words, we pick the centroid of the largest cluster, as one of the MF axes. Then, we obtain \\(\\mathsf{n}_2 = \\mathsf{c}_s\\) and \\(\\mathsf{n}_3 = \\mathsf{c}_t\\) as a solution of the optimization problem" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.279, + 0.856, + 0.294 + ], + "angle": 0, + "content": "Algorithm 1 \\((\\mathcal{N}_1,\\mathcal{N}_2,\\mathcal{N}_3,\\mathsf{R}) =\\) findManhattanFrame(N)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.299, + 0.882, + 0.315 + ], + "angle": 0, + "content": "1. Cluster normals \\(\\mathcal{N}\\) into, \\(\\mathcal{U} = \\{(\\mathcal{C}_i,\\mathsf{c}_i)\\}_{i = 1}^k\\) using k-means." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.315, + 0.865, + 0.328 + ], + "angle": 0, + "content": "2. For the largest cluster \\(\\mathcal{C} \\in \\mathcal{U}\\), assign \\((\\mathcal{C}, \\mathfrak{c}) \\to (\\mathcal{N}_1, \\mathfrak{n}_1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.328, + 0.88, + 0.342 + ], + "angle": 0, + "content": "3. Find \\(\\mathcal{C}_s\\in \\mathcal{U},\\mathcal{C}_t\\in \\mathcal{U}\\) minimizing \\(|\\mathsf{c}_s^\\top \\mathsf{n}_1| + |\\mathsf{n}_1^\\top \\mathsf{c}_t| + |\\mathsf{c}_s^\\top \\mathsf{c}_t|\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.342, + 0.837, + 0.356 + ], + "angle": 0, + "content": "4. Assign \\((\\mathcal{C}_s,\\mathsf{c}_s)\\to (\\mathcal{N}_2,\\mathsf{n}_2),(\\mathcal{C}_t,\\mathsf{c}_t)\\to (\\mathcal{N}_3,.,\\mathsf{n}_3).\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.356, + 0.884, + 0.372 + ], + "angle": 0, + "content": "5. \\(|\\mathbf{e}_z^{\\mathsf{T}}\\mathfrak{n}_1|\\leq \\frac{1}{\\sqrt{2}}?\\mathfrak{n}_1\\to \\mathfrak{n}_z:(|\\mathbf{e}_y^{\\mathsf{T}}\\mathfrak{n}_1|\\leq \\frac{1}{\\sqrt{2}}?\\mathfrak{n}_1\\to \\mathfrak{n}_y:\\mathfrak{n}_1\\to \\mathfrak{n}_x).\\)" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.372, + 0.839, + 0.385 + ], + "angle": 0, + "content": "6. Perform the remaining \\(\\{\\mathsf{n}_1,\\mathsf{n}_2,\\mathsf{n}_3\\} \\to \\{\\mathsf{n}_x,\\mathsf{n}_y,\\mathsf{n}_z\\}\\)" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.299, + 0.884, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.386, + 0.862, + 0.399 + ], + "angle": 0, + "content": "as in step 5 to the closest corresponding canonical axes." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.399, + 0.789, + 0.414 + ], + "angle": 0, + "content": "7. Return \\(\\mathcal{N}_1, \\mathcal{N}_2, \\mathcal{N}_3\\), and \\(\\mathsf{R} = [\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z]^\\top\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.73 + ], + "angle": 0, + "content": "\\(s, t = \\arg \\min_{i,j} |\\mathsf{c}_i^\\top \\mathsf{n}_1| + |\\mathsf{n}_1^\\top \\mathsf{c}_j| + |\\mathsf{c}_i^\\top \\mathsf{c}_j|\\). In other words, we find two additional cluster centroids providing the most orthogonal triplet. We further merge all selected clusters with their opposites. This is achieved by comparing all cluster pairs. Whenever two centroids are nearby, but opposite in sign, the corresponding clusters are merged with the appropriate sign correction. The procedure of finding \\(\\{\\mathcal{N}_1, \\mathcal{N}_2, \\mathcal{N}_3\\}\\) is illustrated in Figure 1 (middle and right), and is summarized in Algorithm 1 from step 1-4. Although the orthogonal clusters with their centroids are sufficient to exploit the desired Manhattan prior, we may wish to recover the MF in the form of a rotation matrix. As such, any arrangement of \\(\\{\\mathsf{n}_1, \\mathsf{n}_2, \\mathsf{n}_3\\}\\) as a valid rotation matrix offers us a valid MF. We may however often be interested to recover the one which is closest to the world frame. For this reason, we suggest keeping the largest cluster's centroid paired to the closes canonical axis in \\(\\mathcal{E}\\). Similarly, we also align one more axis, whereas the last remaining axis gets paired by default. We summarize how to recover MF closest to the world frame in Algorithm 1 in steps 5-7." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.739, + 0.819, + 0.755 + ], + "angle": 0, + "content": "4.3. Self-supervision by Manhattan Prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.885 + ], + "angle": 0, + "content": "The supervision of the implicit neural radiance field, parameterized by \\(\\theta\\), is achieved through jointly optimizing three loss terms. The first one is the photometric loss computed as follows, \\(\\mathcal{L}_{img} = \\frac{1}{|\\mathcal{R}|}\\sum_{\\mathbf{r}\\in \\mathcal{R}}\\| C_{\\theta}(\\mathbf{r}) - C(\\mathbf{r})\\| _2^2\\) where \\(C(\\mathbf{r})\\) is the ground-truth color, and \\(\\mathcal{R}\\) is the set of rays going through sampled pixel triplets. This loss term is responsible for facilitating the learning of the implicit 3D representation of the investigated scene [24]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Losses from Manhattan prior: We exploit the Manhattan" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17863" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.169 + ], + "angle": 0, + "content": "scene prior for self-supervision, in order to improve the implicit 3D representation without additional ground truth labels. We do so by using the clusters \\(\\{\\mathcal{N}_1,\\mathcal{N}_2,\\mathcal{N}_3\\}\\) obtained from Algorithm 1. More precisely, for cluster-centroid pairs \\(\\{(\\mathcal{N}_i,\\mathsf{n}_i)\\}_{i = 1}^3\\) we use the following two losses," + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.18, + 0.47, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c t r} = \\frac {1}{3} \\sum_ {i = 1} ^ {3} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {\\mathfrak {n} \\in \\mathcal {N} _ {i}} \\| 1 - \\mathfrak {n} _ {i} ^ {\\top} \\mathfrak {n} \\| _ {1} + \\| \\mathfrak {n} _ {i} - \\mathfrak {n} \\| _ {1}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.224, + 0.469, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o r t} = \\frac {1}{3} \\left(\\left| n _ {1} ^ {\\top} n _ {2} \\right| + \\left| n _ {1} ^ {\\top} n _ {3} \\right| + \\left| n _ {2} ^ {\\top} n _ {3} \\right|\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.263, + 0.469, + 0.354 + ], + "angle": 0, + "content": "Here, the loss \\(\\mathcal{L}_{ctr}\\) encourages tighter clusters, while the loss \\(\\mathcal{L}_{ort}\\) enforces the orthogonality among the three clusters. The final loss used to optimize \\(\\theta\\) is then given by \\(\\mathcal{L} = \\mathcal{L}_{img} + \\lambda_{ctr}\\mathcal{L}_{ctr} + \\lambda_{ort}\\mathcal{L}_{ort}\\), where \\(\\lambda_{ctr}\\) and \\(\\lambda_{ort}\\) are the hyperparameters. Please, refer to Figure 2 for a schematic summary of all three losses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.368, + 0.21, + 0.385 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.469, + 0.409 + ], + "angle": 0, + "content": "5.1. Baselines, Metrics, and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.469, + 0.536 + ], + "angle": 0, + "content": "InstantNGP [25] (baseline): This method represents the scene as a multi-resolution voxel grid and leverages a hash table of trainable feature vectors, which are used to represent grid elements. The representation is further processed with a small MLP. This allows InstantNGP to be very computationally efficient, requiring in our case around 30 minutes per scene to train and evaluate. Therefore, we use it as the backbone of our method and as our baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.469, + 0.673 + ], + "angle": 0, + "content": "ManhattanDF [14]: This method exploits the Manhattan prior by supervising the explicit normals of the floors to align with the known \\(\\mathsf{n}_z\\), as well as by supervising normals of the walls to align with two learned orthogonal axes which are also orthogonal to \\(\\mathsf{n}_z\\). To achieve this, apart from the RGB images, this method relies on knowing the wall and floor semantics, as well as knowing the exact floor axis \\(\\mathsf{n}_z\\) (MF partially known). We implemented it on top of the In-stantNGP backbone with density field estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.703 + ], + "angle": 0, + "content": "Ours: We implement our proposed method from Section 4 on top of InstantNGP with density field estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Ours (MF known): We modify our method by assuming that the MF is fully known. We do so by adding additional loss terms \\(\\mathcal{L}_{man_i} = \\|1 - \\mathsf{n}_i^{\\mathrm{T}}\\mathsf{m}\\|_1 + \\| \\mathsf{n}_i - \\mathsf{m}\\|_1\\) for each orthogonal cluster centroid \\(\\mathsf{n}_i\\) and its closest MF axis \\(\\mathfrak{m}\\) (or a closer opposite counterpart). Thus, we explicitly guide the orthogonal triplet to align with the known MF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Metrics: To quantitatively evaluate novel view rendering, we use peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) [48]. To evaluate the extracted surface normals of novel rendered views, we use the median angular error. To partially evaluate the quality of the learned implicit 3D structure, we utilize the mean absolute error (MAE) and the root mean square error (RMSE) on the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "depth obtained by volume rendering. Finally, to evaluate the recovered MF, we calculate the absolute error between the yaw, pitch, and roll angles of the recovered frame and the MF. All metrics are averaged across scenes, except for the yaw, pitch, and roll errors for which the median is reported." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Implementation Details: We turn on \\(\\mathcal{L}_{\\text {ort }}\\) and \\(\\mathcal{L}_{\\text {centr }}\\) after 500 steps and linearly increase their weights to the specified values over the next 2500 steps. Also, we randomly sample rays for one-third of every batch size and select their left and upper neighbor to form a triplet to facilitate obtaining explicit surface normals. For other implementation details, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Note that in addition to RGB images, ManhattanDF [14] requires knowing the wall and floor semantics, and the exact floor axis \\(\\mathsf{n}_z\\). In contrast, our method only requires RGB images and the assumed Manhattan prior to holding true. Therefore, this comparison aims to get a better insight into what can be achieved without leveraging additional labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.378, + 0.602, + 0.392 + ], + "angle": 0, + "content": "5.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Hypersim [31] is a photorealistic synthetic dataset consisting of indoor scenes. It was created by leveraging a large repository of scenes created by professional artists, with 461 indoor scenes in total. This dataset is geometry-rich, containing a lot of details and lighting sources. Each scene has one or more camera trajectories available, where each trajectory has up to 100 views rendered in \\(1024 \\times 768\\). For each scene, camera calibration information is provided, as well as detailed per-pixel labels such as depth and surface normals. After cleaning up the scenes by discarding a few with insufficient camera views, as well as other problems, we were left with 435 scenes. We also kept only one camera trajectory per scene. We then evaluated the InstantNGP [25] baseline on all 435 scenes and made three divisions based on its PSNR on unseen views. Hypersim-A contains 20 scenes where the baseline was performing well above average, Hypersim-B contains 20 scenes where the baseline had near-average performance, and Hypersim-C contains 10 scenes where the baseline had a below-average performance. For each scene, we randomly assigned half views to the training split and the rest to the test split. We note that the test split often contains views that were partially or completely unobserved during training. We also note that Hypersim is visually very realistic, geometry-rich, and challenging, considering it is synthetic. This can be subjectively observed by inspecting rendered views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ScanNet [8] is a real-world dataset consisting of indoor scenes. It was collected using a scalable RGB-D capture system that includes automated surface reconstruction and crowd-sourced semantic annotation. The dataset contains 1613 indoor scenes, which are annotated with ground-truth camera poses, surface reconstructions, and instance-level semantic segmentations. We use three scenes that were used" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17864" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.132 + ], + "angle": 0, + "content": "Table 1: Experiments on Hypersim. We observe that our method consistently outperforms the baseline, as well as the ManhattanDF. This is very interesting since, unlike ManhattanDF, we do not use any additional labels during training. Finally, we see that it does not matter for our method whether the MF is known beforehand or not. Therefore, the additional knowledge of MF is neither necessary nor helpful." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.136, + 0.861, + 0.345 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑Normals°↓Pitch°↓Roll°↓Yaw°↓D-MAE↓D-RMSE↓
Scenes AInstantNGP [25] (baseline)25.860.87157.126.186.4619.250.0640.102
ManhattanDF [14]26.510.86840.691.231.015.080.0530.087
Ours27.200.86437.300.400.500.520.0530.093
Ours (MF known)27.210.85635.590.250.260.450.0520.091
Scenes BInstantNGP [25] (baseline)20.750.81160.126.067.9215.870.1050.151
ManhattanDF [14]21.870.82650.502.552.0611.690.0790.121
Ours22.450.81654.081.191.351.810.0800.127
Ours (MF known)22.510.81350.590.510.650.550.0780.126
Scenes CInstantNGP [25] (baseline)17.790.74064.297.454.5515.140.1300.174
ManhattanDF [14]18.330.77056.083.203.4110.250.1030.147
Ours19.430.76854.795.372.244.240.0940.133
Ours (MF known)19.290.76455.123.644.039.480.0940.135
194 scenesInstantNGP [25] (baseline)20.470.78361.346.566.9921.750.1040.146
ManhattanDF [14]20.940.79452.811.722.3213.480.0970.140
Ours21.630.78652.011.871.943.770.0850.126
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.47, + 0.404 + ], + "angle": 0, + "content": "in [14], where one-tenth of all views were uniformly sampled, leaving 303-477 views per scene. For each scene, the training and test split both contain half of the total views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.471, + 0.543 + ], + "angle": 0, + "content": "Replica [38] is a synthetic dataset featuring a diverse set of 18 indoor scenes. Each scene is equipped with photorealistic textures, allowing one to render realistic images from arbitrary camera poses. We use five scenes that were rendered and prepared in [58], where the rendered views also contain semantic segmentation labels. Each scene contains 900 views generated from random 6-DOF trajectories similar to hand-held camera motions in \\(640 \\times 480\\) resolution. We select 75 views for training and 75 for testing in each scene." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.172, + 0.573 + ], + "angle": 0, + "content": "5.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Experiments on Hypersim: We summarize the experiments performed on Hypersim in Table 1. Our proposed method achieves clear improvements in comparison to the InstantNGP baseline, in terms of novel-view rendering, normals estimation, depth estimation, as well as MF recovery. This is consistent across different scene difficulties, namely splits A, B, and C. Furthermore, our method also outperforms the ManhattanDF SotA. This is very interesting since, unlike ManhattanDF, we do not use any additional labels during training, other than the ground truth RGB. The ManhattanDF has partial access to ground truth MF axes, as well as wall and floor semantics, which directly implies having sparse surface normals ground truth. However, there are lots of Manhattan objects and areas in realistic scenes such as in Hypersim, other than the walls and floors. Our self-supervised method leverages the presence of many such objects and areas, to facilitate imposing geometrical constraints on the implicit representation during learning. This can be visually observed in Figure 3. Moreover, we observe that it does not matter for our method whether the MF is known beforehand, since our method recovers the" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.361, + 0.695, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.453, + 0.68, + 0.465 + ], + "angle": 0, + "content": "(a) InstantNGP [25] (baseline)" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.467, + 0.695, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.559, + 0.621, + 0.571 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.361, + 0.892, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.453, + 0.855, + 0.466 + ], + "angle": 0, + "content": "(b) ManhattanDF [14]" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.468, + 0.887, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.558, + 0.833, + 0.57 + ], + "angle": 0, + "content": "(d) Grond truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Figure 3: Leveraging the Manhattan prior (Hypersim-A). Our method leverages the presence of many Manhattan objects and areas in the scene in a self-supervised fashion, which facilitates imposing geometrical constraints during learning. Our method offers plausible normals – sometimes with missing details – that help to better model the radiance fields. In contrast, ManhattanDF leverages the Manhattan prior through labels of only walls and floors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.87 + ], + "angle": 0, + "content": "MF in the clustering step automatically, and therefore similar performance is achieved in both cases. This is particularly exciting because our self-supervised method performs similarly to the supervised one (using the groundtruth MF). Finally, we show large-scale experiments on a larger set of 194 scenes, which lead to the same conclusions. We observed that the semantic loss of the ManhattanDF sometimes causes convergence issues on difficult scenes, which we partially alleviated with class weighting and label smoothing. Therefore, we report results on only 194 scenes where ManhattanDF converged. Additional results on all 435 scenes are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Experiments on ScanNet: In Table 2, we examine the behavior of our proposed method on real-world indoor scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17865" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.885, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.383, + 0.895, + 0.444 + ], + "angle": 0, + "content": "Figure 4: Qualitative results. Our method leverages many Manhattan objects and surfaces in the scene, which improves the implicit geometrical representation compared to the baseline. Unlike ManhattanDF, our method relies on many cues other than the walls & floors, which leads to a better representation of some objects (e.g. the white table in the bottom right corner of the example from columns 1-3). For more qualitative results, please refer to the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.451, + 0.471, + 0.494 + ], + "angle": 0, + "content": "Table 2: Experiments on real-world ScanNet data. Our method outperforms both the baseline and ManhattanDF. This is the case both with and without supervising with sparse depth from SfM." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.497, + 0.462, + 0.618 + ], + "angle": 0, + "content": "
PSNR↑SSIM ↑Depth↓-MAE
InstantNGP [25] (baseline)17.780.5870.119
RegNeRF [26]18.730.6180.102
ManhattanDF [14]18.680.6140.112
Ours20.790.6430.072
+ additional sparse depth from SfM
InstantNGP [25] (baseline)20.700.6310.048
ManhattanDF [14]21.530.6400.052
Ours22.250.6670.033
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.622, + 0.47, + 0.665 + ], + "angle": 0, + "content": "Table 3: Experiments on Replica. Our method outperforms the baseline, and it shows similar performance as ManhattanDF, which leverages additional labels during training." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.669, + 0.462, + 0.737 + ], + "angle": 0, + "content": "
PSNR↑SSIM ↑Depth↓-MAE
InstantNGP [25] (baseline)34.300.9440.022
Semantic-NeRF [58]34.080.9380.014
ManhattanDF [14]35.240.9440.008
Ours35.130.9440.011
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Our method achieves clear improvements in comparison to the InstantNGP baseline, as well as to ManhattanDF, in terms of all measured metrics. Additionally, following the experimental setting of ManhattanDF [14], we train all methods with additional sparse depth supervision – where the sparse depth is obtained from the SfM pipeline. Our proposed method is also superior in this setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Experiments on Replica: We summarize experiments performed on the Replica dataset in Table 3. It can be observed that the InstantNGP baseline already performs very" + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.45, + 0.854, + 0.462 + ], + "angle": 0, + "content": "Improvements of our method w.r.t. scene difficulty" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.462, + 0.872, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.569, + 0.892, + 0.611 + ], + "angle": 0, + "content": "Figure 5: Improvements vs. scene difficulty. Our method achieves the biggest improvements on scenes of hard and moderate difficulty, by leveraging many Manhattan objects and surfaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.749 + ], + "angle": 0, + "content": "well on this dataset, leaving not much room for further improvements. Therefore, we consider Replica as an easier dataset. Nevertheless, our method still performs better than the baseline. As expected, ManhattanDF also improves the baseline similarly. Recall that the ManhattanDF uses additional labels for supervision. We also note that Replica has noticeably more walls and floors, and less of other objects, compared to other more complex datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Qualitative results: We depict the visual results of the discussed experiments in Figure 4. Our method leverages many Manhattan objects and surfaces in the scene, which improves the geometrical structure of 3D compared to the InstantNGP baseline. This is visible in surface normals and depth, obtained using volume rendering. Furthermore, we observe that, unlike ManhattanDF, our method relies on many Manhattan cues other than the walls & floors. This leads to a better representation of some such objects, e.g. the white table in the bottom right corner of the example" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17866" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.079, + 0.468, + 0.107 + ], + "angle": 0, + "content": "Table 4: Sparse training views. Our proposed method clearly exhibits the best performance, for all cases of input view sparsity." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.108, + 0.464, + 0.24 + ], + "angle": 0, + "content": "
PSNR↑SSIM ↑Depth↓-MAE
12 viewsInstantNGP [25]18.020.7060.138
ManhattanDF [14]19.450.7500.116
Ours20.500.7600.104
9 viewsInstantNGP [25]16.790.6610.154
ManhattanDF [14]18.040.7140.130
Ours19.140.7280.120
6 viewsInstantNGP [25]15.750.5820.178
ManhattanDF [14]16.000.6390.159
Ours16.670.6670.158
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.249, + 0.468, + 0.275 + ], + "angle": 0, + "content": "Table 5: Ablation study on Hypersim-A. Both of our proposed loss terms contribute to the overall performance." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.277, + 0.468, + 0.388 + ], + "angle": 0, + "content": "
PSNR↑Norm.°↓Depth↓Rot.°↓
Only Limg25.8657.120.06410.63
+ Lort27.2150.070.0585.39
+ Lctr27.0636.090.0530.57
+ Lort + Lctr (Ours)27.2037.300.0530.47
Ours + MF known27.2135.590.0520.32
Ours (no delay)27.0635.780.0540.39
Ours (no w lin.)27.0137.860.0550.60
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.469, + 0.462 + ], + "angle": 0, + "content": "from columns 1-3. Moreover, we observe that our method is able to cope with difficult scenes and views, where other methods struggle. For more qualitative results, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.568 + ], + "angle": 0, + "content": "Improvements with respect to scene difficulty: We analyze the improvements by our method for different scene difficulties. We decided on scene difficulty based on the novel-view rendering performance of the InstantNGP baseline. In Figure 5, we see that our method brings the most benefits for scenes of hard and moderate difficulties, thanks to the Manhattan scene prior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Sparse training views: In order to gain more insight, we examine the behavior of training neural radiance fields with sparse training input views. The results on the Hypersim-A dataset are presented in Table 4. Our proposed method clearly outperforms the InstantNGP baseline, as well as ManhattanDF, when trained with 6, 9, and 12 input views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Finding the MF: We test the robustness of our proposed method for finding the Manhattan frame, by introducing a simultaneous offset \\(\\alpha\\) in the yaw, pitch, and roll on the canonical MF. The experiments are performed on Hypersim A and can be found in Figure 6. In Figure 6a we observe that the novel-view rendering quality remains largely robust to the rotation offset \\(\\alpha\\). We note that we increase the scene bounding box by the same factor for all experiments in Figure 6, to make sure that objects remain within the voxel grid for maximal \\(\\alpha\\). This slightly decreases the resolution of grid element representations, so PSNR is slightly lower than in Table 1. Furthermore, in Figure 6b we see that the MF estimation remains robust to the rotation offset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Ablation study: We report our ablation study in Table 5. Both of our proposed losses contribute to the overall performance. Furthermore, turning on \\(\\mathcal{L}_{\\text{ort}}\\) and \\(\\mathcal{L}_{\\text{centr}}\\) after" + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.092, + 0.82, + 0.102 + ], + "angle": 0, + "content": "Rotation offset effects on novel-view rendering" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.102, + 0.847, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.179, + 0.885, + 0.19 + ], + "angle": 0, + "content": "(a) Novel-view rendering quality remains robust to the rotation offset \\(\\alpha\\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.193, + 0.839, + 0.203 + ], + "angle": 0, + "content": "Rotation offset effects on Manhattan frame estimation" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.203, + 0.846, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.281, + 0.846, + 0.291 + ], + "angle": 0, + "content": "(b) MF estimation remains robust to the rotation offset \\(\\alpha\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.298, + 0.891, + 0.326 + ], + "angle": 0, + "content": "Figure 6: Finding the MF. We test the robustness of our method by introducing the rotation offsets on the canonical MF." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.341, + 0.891, + 0.372 + ], + "angle": 0, + "content": "500 steps, and linearly increasing their weight also helps slightly. Please refer to the supplementary for more details." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.385, + 0.619, + 0.4 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.71 + ], + "angle": 0, + "content": "We demonstrated the possibility of exploiting the Manhattan scene prior without needing any additional supervision. This is achieved by performing robust clustering of explicit normals, followed by the search of the Manhattan frame (MF) whose existence is based on the assumed prior. The sought MF is obtained from the orthogonal clusters, which are later used to self-supervise the neural representation learning. Our self-supervision encourages the normals of Manhattan surfaces to group into three orthogonal directions. Our experiments on three indoor datasets demonstrate that the proposed method not only benefits from building parts (such as walls and floors) but also exploits many other Manhattan scene parts (such as furniture). Both quantitative and qualitative evaluations reveal the benefit of the proposed method in terms of, improved performance over the established baselines and competitive results to state-of-the-art methods that use additional labels for supervision. Our method has the potential to be extended in other higher expressiveness scene priors, such as Atlanta world and the mixture of Manhattan frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Limitations: One limitation of our method is that it sometimes produces surface normals with missing details or \"blocky\" artifacts. Nevertheless, this usually offers better novel-view RGB rendering, compared to not imposing any structure priors. Another limitation is that our method is not beneficial for very easy scenes. For a detailed discussion of limitations, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.831, + 0.659, + 0.847 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.891, + 0.901 + ], + "angle": 0, + "content": "This research was co-financed by Innosuisse under the project Know Where To Look, Grant 59189.1 IP-ICT. Also, Carlos Oliveira helped with splendid visual renderings." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17867" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] Stephen T Barnard. Interpreting perspective images. Artificial intelligence, 21(4):435-462, 1983." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] Jean-Charles Bazin, Hongdong Li, In So Kweon, Cedric Demonceaux, Pascal Vasseur, and Katsushi Ikeuchi. A branch-and-bound approach to correspondence and grouping problems. IEEE transactions on pattern analysis and machine intelligence, 35(7):1565-1576, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[4] Jean-Charles Bazin, Yongduek Seo, Cédric Demonceaux, Pascal Vasseur, Katsushi Ikeuchi, Inso Kweon, and Marc Pollefeys. Globally optimal line clustering and vanishing point estimation in manhattan world. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 638–645. IEEE, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.471, + 0.415 + ], + "angle": 0, + "content": "[5] Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag, Berlin, Heidelberg, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.471, + 0.5 + ], + "angle": 0, + "content": "[6] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.471, + 0.57 + ], + "angle": 0, + "content": "[7] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 941-947. IEEE, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.573, + 0.471, + 0.641 + ], + "angle": 0, + "content": "[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.644, + 0.471, + 0.713 + ], + "angle": 0, + "content": "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[10] Patrick Denis, James H Elder, and Francisco J Estrada. Efficient edge-based methods for estimating Manhattan frames in urban imagery. In European conference on computer vision, pages 197-210. Springer, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[11] Clara Fernandez-Labrador. Indoor scene understanding using non-conventional cameras. PhD thesis, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[12] Yasutaka Furukawa, Brian Curless, Steven M Seitz, and Richard Szeliski. Manhattan-world stereo. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1422-1429. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Wuwei Ge, Yu Song, Baichao Zhang, and Zehua Dong. Globally optimal and efficient Manhattan frame estimation by delimiting rotation search space. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 15213-15221, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[14] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[16] Kyungdon Joo, Tae-Hyun Oh, Junsik Kim, and In So Kweon. Globally optimal Manhattan frame estimation in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1763-1771, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[17] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[18] Fu Li, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerf-posed: A first-reconstruct-thenregress approach for weakly-supervised 6d object pose estimation. arXiv preprint arXiv:2203.04802, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[19] Haoang Li, Jian Yao, Jean-Charles Bazin, Xiaohu Lu, Yazhou Xing, and Kang Liu. A monocular slam system leveraging structural regularity in manhattan world. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 2518-2525. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[20] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[21] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15702-15712, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[22] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[23] Gerard F McLean and D Kotturi. Vanishing point detection by line clustering. IEEE Transactions on pattern analysis and machine intelligence, 17(11):1090-1095, 1995." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. cite arxiv:2003.08934Comment: ECCV 2020 (oral). Project page with videos and code: http://tancik.com/nerf." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17868" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[26] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.176, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[27] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.233, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[29] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in manhattan world. In Proceedings of the IEEE International Conference on Computer Vision, pages 882-890, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.359, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.428, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[31] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.511, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[32] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.581, + 0.471, + 0.622 + ], + "angle": 0, + "content": "[33] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.623, + 0.471, + 0.719 + ], + "angle": 0, + "content": "[34] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lučić, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6229-6238, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.72, + 0.471, + 0.762 + ], + "angle": 0, + "content": "[35] Gautam Singh and Jana Kosecka. Visual loop closing using gist descriptors in manhattan world. In ICRA omnidirectional vision workshop, pages 4042-4047, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.763, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[36] Julian Straub, Nishchal Bhandari, John J Leonard, and John W Fisher. Real-time manhattan world rotation estimation in 3d. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1913-1920. IEEE, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[37] Julian Straub, Oren Freifeld, Guy Rosman, John J Leonard, and John W Fisher. The Manhattan frame model—manhattan world inference in the space of surface normals. IEEE transactions on pattern analysis and machine intelligence, 40(1):235-249, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[38] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[39] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[40] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light field neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8269-8279, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[41] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[42] Frank A Van den Heuvel. 3d reconstruction from a single image using geometric constraints. ISPRS Journal of Photogrammetry and Remote Sensing, 53(6):354-368, 1998." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[43] Carlos A Vanegas, Daniel G Aliaga, and Bedrich Benes. Building reconstruction using manhattan-world grammars. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 358-365. IEEE, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[44] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[45] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[46] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. arXiv preprint arXiv:2206.13597, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[48] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[49] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17869" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.147 + ], + "angle": 0, + "content": "[50] Horst Wildenauer and Allan Hanbury. Robust camera self-calibration from monocular images of manhattan worlds. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2831-2838. IEEE, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.219 + ], + "angle": 0, + "content": "[51] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.22, + 0.468, + 0.276 + ], + "angle": 0, + "content": "[52] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K Wong. Ps-nerf: Neural inverse rendering for multi-view photometric stereo. In European Conference on Computer Vision, pages 266–284. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.468, + 0.331 + ], + "angle": 0, + "content": "[53] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K Wong. S3-nerf: Neural reflectance field from shading and shadow under a single viewpoint. arXiv preprint arXiv:2210.08936, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[54] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.375, + 0.468, + 0.444 + ], + "angle": 0, + "content": "[55] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.446, + 0.468, + 0.5 + ], + "angle": 0, + "content": "[56] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.502, + 0.468, + 0.571 + ], + "angle": 0, + "content": "[57] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5657-5665, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.573, + 0.468, + 0.642 + ], + "angle": 0, + "content": "[58] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17870" + } + ] +] \ No newline at end of file diff --git a/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/f357e4d1-f9a1-4f72-b3d1-9e59cc154adf_origin.pdf b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/f357e4d1-f9a1-4f72-b3d1-9e59cc154adf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..64ac2a07bc66883384b087b7e25bd30e196e2477 --- /dev/null +++ b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/f357e4d1-f9a1-4f72-b3d1-9e59cc154adf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbc1856bb0e335c86fbfaac53bc5ef16db87e61d86f798abc02ced765ba8c4ff +size 1698084 diff --git a/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/full.md b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad7b52910e9e7b5a53fd9082efddeecca6b2769c --- /dev/null +++ b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/full.md @@ -0,0 +1,300 @@ +# Surface Normal Clustering for Implicit Representation of Manhattan Scenes + +Nikola Popovic1, Danda Pani Paudel1,2, Luc Van Gool1,2 + +1Computer Vision Laboratory, ETH Zurich, Switzerland +2INSAIT, Sofia University, Bulgaria + +{nipopovic, paudel, vangool}@vision.ee.ethz.ch + +# Abstract + +Novel view synthesis and 3D modeling using implicit neural field representation are shown to be very effective for calibrated multi-view cameras. Such representations are known to benefit from additional geometric and semantic supervision. Most existing methods that exploit additional supervision require dense pixel-wise labels or localized scene priors. These methods cannot benefit from high-level vague scene priors provided in terms of scenes' descriptions. In this work, we aim to leverage the geometric prior of Manhattan scenes to improve the implicit neural radiance field representations. More precisely, we assume that only the knowledge of the indoor scene (under investigation) being Manhattan is known - with no additional information whatsoever - with an unknown Manhattan coordinate frame. Such high-level prior is used to self-supervise the surface normals derived explicitly in the implicit neural fields. Our modeling allows us to cluster the derived normals and exploit their orthogonality constraints for self-supervision. Our exhaustive experiments on datasets of diverse indoor scenes demonstrate the significant benefit of the proposed method over the established baselines. The source code will be available at https://github.com/nikola3794/normal-clustering-nerf. + +# 1. Introduction + +Above $80\%$ images ever taken are estimated to involve human-made architectural structures, with a substantial share of indoor scenes [11]. These scenes often exhibit strong structural regularities, including flat and texture-poor surfaces in some axis-aligned Cartesian coordinate system - also known as Manhattan world [7, 12]. Paradoxically, these regularities may hinder the 3D modeling process if it is unaware of the human-made scene priors. In fact, several computer vision works have even benefited from the knowledge of the Manhattan world for the task of 3D scene reconstruction [43, 12], camera localization [19], self-calibration [50], and more [37, 35, 29, 57]. + +For calibrated multi-view cameras, 3D inversion using implicit neural representations [24, 47, 54] is becoming increasingly popular due to their remarkable performance and recent efficiency developments [55, 30, 25, 51]. Meanwhile, such representations are known to benefit from additional supervision in the form of depth [1, 9, 32], normals [18], semantics [45, 58, 17, 15], local-regularization [46, 44, 26], local planar patches [21], or their combinations [56, 14, 21]. In this context, a notable recent work ManhattanSDF [14] demonstrates the benefit of exploiting the high-level geometric prior for structured scenes. More precisely, ManhattanSDF [14] uses the known semantic regions to impose the planar geometry prior of floors and walls under the Manhattan scene assumption. During this process, the exact normal of the floor and partial normals of the walls are assumed to be known, with respect to the camera coordinates. + +We aim to improve the 3D neural radiance field representations for calibrated multi-view cameras in indoor Manhattan scenes, with no further assumptions. In other words, we consider that the structural and semantic information is not available, unlike ManhattanSDF [14]. In addition to the floor and walls used in ManhattanSDF [14], we wish to exploit the Manhattan prior of many other indoor scene parts (e.g. tables and wardrobes). More importantly, we consider that the Manhattan coordinate frame is also unknown. Our assumptions (of unknown semantics and Manhattan frame) on one hand make our setting very practical. On the other hand, those assumptions make the problem of exploiting the Manhattan scene prior for 3D inversion very challenging. + +The virtue of the Manhattan world assumption comes from its simplicity, allowing us to intuitively reason about the geometry of a wide range of complex scenes/objects such as cities, buildings, and furniture. However, such reasoning often requires the axis-aligned Cartesian coordinate frame, also known as the Manhattan frame (MF), to be known [37]. Unfortunately, recovering the Manhattan frame directly from images is not an easy task [7, 37]. Therefore, several methods have been developed until recently [10, 4, 36, 16, 13] to recover the Manhattan frame, relying upon the known 3D reconstruction or image primi + +![](images/5775e0ce3426ff3c30121e5471a1a73efcd3a6787f004adaf4ff305745a799c9.jpg) +Figure 1: Illustration of the proposed method. We compute one surface normal for each ray triplet, using 3D surface points derived from rendered depths (left). The computed normals (from many ray triplets) are clustered to obtain the MF (middle). The non-Manhattan and noisy surfaces are handled by a robust orthogonal normals search, which is later used for self-supervision through the Manhattan prior (right). + +![](images/a7eab9213b88cf5b4434efa310c8a6b9ddeb780d2f63ac251c9e6b7af5deea13.jpg) + +![](images/bb6479a6e5384015fae3d31865f05580f1766214e15d38251e4f102df4b94293.jpg) + +tives (eg. lines, planes). We wish to exploit the Manhattan prior for improving the 3D representation, without needing to know MF beforehand. Instead, our experiments reveal that knowing the MF beforehand offers no additional benefit in Manhattan-prior aware radiance field representation. + +In this work, we propose a method that jointly learns the Manhattan frame and neural radiance field, from calibrated multi-view in indoor Manhattan scenes, in an end-to-end manner using the recent efficient backbone of Instant-NGP [25]. The proposed method requires no additional information to exploit the Manhattan prior and relies on the explicitly derived normals in the implicit neural fields. We use batches of three neighboring rays, whose effective surface's local piece-wise planarity is assumed to derive the explicit normals by algebraic means. In pure, complete, and enclosed Manhattan scenes, these normals form six clusters corresponding to three orthogonal and other three parallel counterpart planes. However real scenes consist of non-Manhattan scene parts and missing planes. Therefore, we use a robust method that uses minimal three orthogonal clusters to recover the sought Manhattan frame. As in the literature, we seek a rotation matrix whose entries are directly derived from the orthogonal clusters of normals, to align the Manhattan frame. The recovered MF is then used to encourage the derived normals to be axis-aligned for self-supervision. A graphical illustration of our method is presented in Figure 1. Our extensive experiments demonstrate the robustness and utility of our method in improving the implicit 3D in neural radiance field representations. + +The major contributions of our paper are listed below: + +- We address the problem of exploiting the structured-scene knowledge without requiring any dense or localized scene priors, for the first time in this paper. +- We present a method that successfully exploits the Manhattan scene prior with an unknown Manhattan frame. The proposed method also recovers the unknown frame. +- We demonstrate the robustness and utility of our method + +on three indoor datasets, where our method improves the established baselines significantly. These datasets consist of $200+$ scenes, making our method tested in significantly more scenes than the state-of-the-art methods. + +# 2. Related Works + +Implicit neural representation of 3D: Since the foundational work of Mildenhall et al. [24], the implicit neural representation of 3D scenes has advanced on various fronts. These fronts include representation [54, 40], generalization [49, 28, 20, 22, 34], generation [27, 6], and efficient methods [30, 55, 25]. We rely on a recent method Instant-NGP [25] developed by Mueller et al., as our backbone. Our choice is made primarily based on the computational efficiency during both training and inference. Thanks to the offered computational efficiency, we are able to conduct large-scale experiments on several scenes. + +Auxiliary supervision methods: In addition to the images, other inputs such as depth [33, 39, 9], semantics [58, 15, 17, 41], normal [52, 53], and their combinations [56, 14, 21] are shown to be beneficial on improving the neural radiance field representation. In this regard, these auxiliary supervisions often use ground-truth labels. It is needless to say that the need for ground-truth supervision is not desired, whenever possible. Therefore, recent methods use labels predicted by some pre-trained networks [56] or recovered from the structure-from-motion (SfM) pipeline [9, 21]. One notable work ManhattanSDF [14] exploits the Manhattan prior without requiring any SfM reconstruction. However, ManhattanSDF requires (a) semantics of the scene parts and (b) the Manhattan frame, to be known. We argue that such labels required for auxiliary supervision cannot always be relied upon, due to domain gaps, poor reconstruction of texture-less regions, and additional computational needs, to list a few. Therefore, we do not use any additional labels for auxiliary supervision. + +Manhattan frame estimation: Since the early works of Bernard [2], Manhattan structure reasoning is done directly on/from images by detecting the so-called the vanishing-points (VPs) [23, 42]. In fact, the problem of detecting three orthogonal VPs is equivalent to finding the MF in 3D for the calibrated multi-view setting [3]. Note that the knowledge of Manhattan structure has been used in several computer vision works [43, 12, 19, 50, 37, 35, 29, 57]. When unknown, most methods implicitly or explicitly estimate the MF to leverage the Manhattan scene prior. In [37], Straub et al. have demonstrated that the MF can be efficiently represented in and recovered from the space of surface normals. We use a similar formulation as [37], using the surface normals derived explicitly from the implicit neural fields. + +# 3. Background and Notations + +Manhattan frame (MF) is a coordinate system that is defined by the structure building orthogonal planes of Manhattan scenes. We consider unknown MF since the scene planes and their geometric relationships are unknown. Let the unknown MF differs from the world-frame (WF), used for the 3D representation, by rotation $\mathsf{R} \in SO(3)$ . We denote three orthogonal axes in MF by $\mathcal{E} = \{\mathsf{e}_x, \mathsf{e}_y, \mathsf{e}_z\}$ . Without loss of generality, let the axes' coordinates be $\mathsf{e}_x = [1,0,0]^{\mathsf{T}}$ , $\mathsf{e}_y = [0,1,0]^{\mathsf{T}}$ , and $\mathsf{e}_z = [0,0,1]^{\mathsf{T}}$ . Note that these axes align with the normals of the respective scene building planes, in the WF. Let $\mathcal{N} = \{\mathsf{n}_i\}_{i=1}^m$ be a set of 3D normals of all the scene planes. Then, the rotation $\mathsf{R}$ , from world to Manhattan frame, aligns the normals $\mathsf{n} \in \mathcal{N}$ to the Manhattan axes $\mathsf{e} \in \mathcal{E}$ , i.e., $\mathsf{Rn}_i \in \mathcal{E}$ , $\forall \mathsf{n}_i \in \mathcal{N}$ . + +We are interested to recover $\mathsf{R}$ from a set of normals $\mathcal{N}$ . For this, the above set-to-set assignment alone is not sufficient. This requires the element-wise assignment between sets $\mathcal{N}$ and $\mathcal{E}$ . To do so, we divide the set $\mathcal{N}$ into three orthogonal subsets $\mathcal{N}_x, \mathcal{N}_y,$ and $\mathcal{N}_z$ . Now, for any triplet of $\{\mathsf{n}_x, \mathsf{n}_y, \mathsf{n}_z\}$ from the corresponding orthogonal subsets, we aim to establish the condition $[\mathsf{e}_x, \mathsf{e}_y, \mathsf{e}_z] = \mathsf{R}[\mathsf{n}_x, \mathsf{n}_y, \mathsf{n}_z]$ . Note that the assignment condition results into $\mathsf{R} = [\mathsf{n}_x, \mathsf{n}_y, \mathsf{n}_z]^{\top}$ . Therefore, the problem of recovering MF from a given set of normals boils down to finding three normals from orthogonal subsets. At this point, one issue regarding robustness remains pending. More precisely, we wish to recover $\mathsf{R}$ for a noisy set of normals $\mathcal{N}$ , with potentially overwhelmingly many outliers. + +We use a robust method to recover $\mathsf{R}$ from given normals $\mathcal{N}$ . In principle, $\mathsf{R}$ can be recovered from minimal two orthogonal normals, with one additional normal for disambiguation by sign correction. The recovered rotation can then be validated by consensus for robust recovery [37]. Alternatively, one can also perform the robust estimation of the orthogonal subsets of $\mathcal{N}$ , followed by solving $[\mathsf{e}_x, \mathsf{e}_y, \mathsf{e}_z] = \mathsf{R}[\mathsf{n}_x, \mathsf{n}_y, \mathsf{n}_z], \forall \mathsf{n}_x \in \mathcal{N}_x, \mathsf{n}_y \in \mathcal{N}_y, \mathsf{n}_z \in \mathcal{N}_z$ for $\mathsf{R} \in SO(3)$ . For computational reasons, we estimate + +the robust subsets by clustering. The orthogonal subsets are then obtained by choosing the three most orthogonal clusters. Later, the obtained cluster centers are used to estimate the MF (or to enforce its existence) represented by R. Please refer to Section 4.2 for more details. + +# 4. Method + +From a set of calibrated images, we model the 3D scene using the neural radiance field representation. In the process, we wish to exploit the Manhattan scene prior, without the knowledge of the Manhattan frame. This problem is addressed by jointly optimizing for the neural radiance field and the Manhattan frame estimation. The complete pipeline of our method is illustrated in Figure 2. As shown, our method consists of three units: (a) Explicit normal modeling, (b) Robust estimation of orthogonal normals, and (3) Self-supervision by Manhattan prior. In the first stage, we shoot a batch of three rays which allows us to estimate the surface normal using an algebraic method. As the estimated normals are bound to be noisy (with outliers), we perform their clustering to obtain the most orthogonal clusters (representing the Manhattan frame) in a robust manner. The obtained clusters are then used to estimate the sought MF, in the form of a rotation matrix, using the method discussed in Section 3. In the final step, we use the estimated Manhattan frame to encourage the close-by normals to be Manhattan-like and to enforce a stricter orthogonality constraint for self-supervision. In the following, we will present the details of three units of our method in different Subsections. + +# 4.1. Explicit Normal Modelling + +At any given view, we consider the color $c = C_{\theta}(\mathsf{r})$ and depth $d = D_{\theta}(\mathsf{r})$ for a ray $\mathsf{r}$ emanating from the corresponding camera center $\mathsf{o}$ is obtained using the volumetric rendering of the implicit neural radiance field, represented by a neural network parameterized by $\theta$ as in [24]. Then, the location of the 3D surface point is given by, $\mathsf{x} = \mathsf{o} + d.\mathsf{r}$ . We process triplets of three neighboring rays. Let $\mathcal{T} = \{\mathsf{r}_1,\mathsf{r}_2,\mathsf{r}_3\}$ be such a triplet, whose corresponding surface points are given by $\mathcal{X} = \{\times_1,\times_2,\times_3\}$ . Now, for the ray triplet $\mathcal{T}$ , we obtain the explicit surface normal $\mathsf{n}$ with the help of the point triplet $\mathcal{X}$ , using the following mapping and algebraic operation, + +$$ +\mathcal {T} \rightarrow \mathcal {X} \rightarrow v = \left(x _ {1} - x _ {2}\right) \times \left(x _ {1} - x _ {3}\right)\rightarrow n = \frac {\operatorname {s i g n} (\mathrm {o} ^ {\mathrm {T}} v) v}{\| v \|}. \tag {1} +$$ + +Note that $\nu$ is a vector orthogonal to the plane passing through 3D points in triplet $\mathcal{T}$ . We obtain the oriented normal $\mathsf{n}$ by normalizing $\nu$ and correcting its sign by ensuring that the camera center $\mathsf{o}$ is in front of the estimated plane. A graphical illustration of the surface normal estimation is provided in Figure 1 on the left. We select a random set + +![](images/d3faeefa59b922e950680a21324596f55cda204969108ec9d3f44a70a0fcc6e2.jpg) +Figure 2: The complete pipeline of our method. We use grid features (GF) and an MLP to represent the radiance field. The explicit normals are derived using the depths obtained from volume rendering. The Manhattan scene prior is exploited by clustering the estimated normals to enforce their orthogonality. + +of ray triplets $\{\mathcal{T}_i\}_{i = 1}^m$ from multiple cameras. These ray triplets provide us the surface point triplets $\{\mathcal{X}_i\}_{i = 1}^m$ . We use point triplets to explicitly derive the surface normals $\mathcal{N} = \{\mathfrak{n}_i\}_{i = 1}^m$ , using (1). These normals are later used to recover the unknown Manhattan frame. The simplicity of the explicit normals computed in this paper makes them easy to compute and handle. If needed, normals of different sizes of surface regions could also be computed similarly. + +# 4.2. Robust Estimation of Orthogonal Normals + +We are interested to recover the MF from a set of noisy surface normals $\mathcal{N} = \{\mathfrak{n}_i\}_{i=1}^m$ . Recall Section 3, the MF can be obtained by robustly recovering a set of three orthogonal normals from $\mathcal{N}$ . To do so, we first cluster $\mathcal{N}$ into $k$ clusters $\{\mathcal{C}_i\}_{i=1}^k$ , by using the well-known $k$ -means clustering algorithm [5]. During the clustering, every centroid is $L_2$ normalized after each iteration, to ensure that they are unit vectors to represent surface normals. In a perfect Manhattan world, there exist only six clusters corresponding to the three orthogonal MF axes $\mathcal{E} = \{\mathbf{e}_x, \mathbf{e}_y, \mathbf{e}_z\}$ and their parallel counterparts. However, real scenes consist of surfaces of different orientations. An additional source of non-Manhattan normals comes from the inaccuracy in the normal estimation. Therefore, we use a clustering technique to robustly recover the orthogonal normals. In the following, we present how the orthogonal clusters are selected from the set of clusters $\{\mathcal{C}_i\}_{i=1}^k$ . Note from Section 3, the selected orthogonal clusters are considered to represent MF defining sets $\mathcal{N}_x, \mathcal{N}_y$ , and $\mathcal{N}_z$ . We proceed by first selecting three orthogonal clusters, say $\mathcal{N}_1, \mathcal{N}_2$ , and $\mathcal{N}_3$ . Later, we assign them to $\mathcal{N}_x, \mathcal{N}_y$ , and $\mathcal{N}_z$ to recover MF. + +For notational ease, we pair clusters and centroids as $\mathcal{U} = \left\{\left(\mathcal{C}_i,\mathsf{c}_i\right)\right\}_{i = 1}^k$ , where the centroids are computed by taking the average across the corresponding cluster such that $\mathsf{c}_i = \frac{1}{|\mathcal{C}_i|}\sum_{\mathsf{n}\in \mathcal{C}_i}\mathsf{n},$ followed by normalizing to a unit vector $\mathsf{c}_i = \frac{\mathsf{c}_i}{\|\mathsf{c}_i\|}$ . Leveraging the Manhattan prior through the assumption that Manhattan surfaces dominate the scene, we pick $\mathsf{n}_1 = \mathsf{c}_f$ , where $f = \arg \max_{i}|\mathcal{C}_i|$ In other words, we pick the centroid of the largest cluster, as one of the MF axes. Then, we obtain $\mathsf{n}_2 = \mathsf{c}_s$ and $\mathsf{n}_3 = \mathsf{c}_t$ as a solution of the optimization problem + +# Algorithm 1 $(\mathcal{N}_1,\mathcal{N}_2,\mathcal{N}_3,\mathsf{R}) =$ findManhattanFrame(N) + +1. Cluster normals $\mathcal{N}$ into, $\mathcal{U} = \{(\mathcal{C}_i,\mathsf{c}_i)\}_{i = 1}^k$ using k-means. +2. For the largest cluster $\mathcal{C} \in \mathcal{U}$ , assign $(\mathcal{C}, \mathfrak{c}) \to (\mathcal{N}_1, \mathfrak{n}_1)$ . +3. Find $\mathcal{C}_s\in \mathcal{U},\mathcal{C}_t\in \mathcal{U}$ minimizing $|\mathsf{c}_s^\top \mathsf{n}_1| + |\mathsf{n}_1^\top \mathsf{c}_t| + |\mathsf{c}_s^\top \mathsf{c}_t|$ +4. Assign $(\mathcal{C}_s,\mathsf{c}_s)\to (\mathcal{N}_2,\mathsf{n}_2),(\mathcal{C}_t,\mathsf{c}_t)\to (\mathcal{N}_3,.,\mathsf{n}_3).$ +5. $|\mathbf{e}_z^{\mathsf{T}}\mathfrak{n}_1|\leq \frac{1}{\sqrt{2}}?\mathfrak{n}_1\to \mathfrak{n}_z:(|\mathbf{e}_y^{\mathsf{T}}\mathfrak{n}_1|\leq \frac{1}{\sqrt{2}}?\mathfrak{n}_1\to \mathfrak{n}_y:\mathfrak{n}_1\to \mathfrak{n}_x).$ +6. Perform the remaining $\{\mathsf{n}_1,\mathsf{n}_2,\mathsf{n}_3\} \to \{\mathsf{n}_x,\mathsf{n}_y,\mathsf{n}_z\}$ + +as in step 5 to the closest corresponding canonical axes. + +7. Return $\mathcal{N}_1, \mathcal{N}_2, \mathcal{N}_3$ , and $\mathsf{R} = [\mathsf{n}_x, \mathsf{n}_y, \mathsf{n}_z]^\top$ . + +$s, t = \arg \min_{i,j} |\mathsf{c}_i^\top \mathsf{n}_1| + |\mathsf{n}_1^\top \mathsf{c}_j| + |\mathsf{c}_i^\top \mathsf{c}_j|$ . In other words, we find two additional cluster centroids providing the most orthogonal triplet. We further merge all selected clusters with their opposites. This is achieved by comparing all cluster pairs. Whenever two centroids are nearby, but opposite in sign, the corresponding clusters are merged with the appropriate sign correction. The procedure of finding $\{\mathcal{N}_1, \mathcal{N}_2, \mathcal{N}_3\}$ is illustrated in Figure 1 (middle and right), and is summarized in Algorithm 1 from step 1-4. Although the orthogonal clusters with their centroids are sufficient to exploit the desired Manhattan prior, we may wish to recover the MF in the form of a rotation matrix. As such, any arrangement of $\{\mathsf{n}_1, \mathsf{n}_2, \mathsf{n}_3\}$ as a valid rotation matrix offers us a valid MF. We may however often be interested to recover the one which is closest to the world frame. For this reason, we suggest keeping the largest cluster's centroid paired to the closes canonical axis in $\mathcal{E}$ . Similarly, we also align one more axis, whereas the last remaining axis gets paired by default. We summarize how to recover MF closest to the world frame in Algorithm 1 in steps 5-7. + +# 4.3. Self-supervision by Manhattan Prior + +The supervision of the implicit neural radiance field, parameterized by $\theta$ , is achieved through jointly optimizing three loss terms. The first one is the photometric loss computed as follows, $\mathcal{L}_{img} = \frac{1}{|\mathcal{R}|}\sum_{\mathbf{r}\in \mathcal{R}}\| C_{\theta}(\mathbf{r}) - C(\mathbf{r})\| _2^2$ where $C(\mathbf{r})$ is the ground-truth color, and $\mathcal{R}$ is the set of rays going through sampled pixel triplets. This loss term is responsible for facilitating the learning of the implicit 3D representation of the investigated scene [24]. + +Losses from Manhattan prior: We exploit the Manhattan + +scene prior for self-supervision, in order to improve the implicit 3D representation without additional ground truth labels. We do so by using the clusters $\{\mathcal{N}_1,\mathcal{N}_2,\mathcal{N}_3\}$ obtained from Algorithm 1. More precisely, for cluster-centroid pairs $\{(\mathcal{N}_i,\mathsf{n}_i)\}_{i = 1}^3$ we use the following two losses, + +$$ +\mathcal {L} _ {c t r} = \frac {1}{3} \sum_ {i = 1} ^ {3} \frac {1}{| \mathcal {N} _ {i} |} \sum_ {\mathfrak {n} \in \mathcal {N} _ {i}} \| 1 - \mathfrak {n} _ {i} ^ {\top} \mathfrak {n} \| _ {1} + \| \mathfrak {n} _ {i} - \mathfrak {n} \| _ {1}, \tag {2} +$$ + +$$ +\mathcal {L} _ {o r t} = \frac {1}{3} \left(\left| n _ {1} ^ {\top} n _ {2} \right| + \left| n _ {1} ^ {\top} n _ {3} \right| + \left| n _ {2} ^ {\top} n _ {3} \right|\right). \tag {3} +$$ + +Here, the loss $\mathcal{L}_{ctr}$ encourages tighter clusters, while the loss $\mathcal{L}_{ort}$ enforces the orthogonality among the three clusters. The final loss used to optimize $\theta$ is then given by $\mathcal{L} = \mathcal{L}_{img} + \lambda_{ctr}\mathcal{L}_{ctr} + \lambda_{ort}\mathcal{L}_{ort}$ , where $\lambda_{ctr}$ and $\lambda_{ort}$ are the hyperparameters. Please, refer to Figure 2 for a schematic summary of all three losses. + +# 5. Experiments + +# 5.1. Baselines, Metrics, and Implementation Details + +InstantNGP [25] (baseline): This method represents the scene as a multi-resolution voxel grid and leverages a hash table of trainable feature vectors, which are used to represent grid elements. The representation is further processed with a small MLP. This allows InstantNGP to be very computationally efficient, requiring in our case around 30 minutes per scene to train and evaluate. Therefore, we use it as the backbone of our method and as our baseline. + +ManhattanDF [14]: This method exploits the Manhattan prior by supervising the explicit normals of the floors to align with the known $\mathsf{n}_z$ , as well as by supervising normals of the walls to align with two learned orthogonal axes which are also orthogonal to $\mathsf{n}_z$ . To achieve this, apart from the RGB images, this method relies on knowing the wall and floor semantics, as well as knowing the exact floor axis $\mathsf{n}_z$ (MF partially known). We implemented it on top of the In-stantNGP backbone with density field estimation. + +Ours: We implement our proposed method from Section 4 on top of InstantNGP with density field estimation. + +Ours (MF known): We modify our method by assuming that the MF is fully known. We do so by adding additional loss terms $\mathcal{L}_{man_i} = \|1 - \mathsf{n}_i^{\mathrm{T}}\mathsf{m}\|_1 + \| \mathsf{n}_i - \mathsf{m}\|_1$ for each orthogonal cluster centroid $\mathsf{n}_i$ and its closest MF axis $\mathfrak{m}$ (or a closer opposite counterpart). Thus, we explicitly guide the orthogonal triplet to align with the known MF. + +Metrics: To quantitatively evaluate novel view rendering, we use peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) [48]. To evaluate the extracted surface normals of novel rendered views, we use the median angular error. To partially evaluate the quality of the learned implicit 3D structure, we utilize the mean absolute error (MAE) and the root mean square error (RMSE) on the + +depth obtained by volume rendering. Finally, to evaluate the recovered MF, we calculate the absolute error between the yaw, pitch, and roll angles of the recovered frame and the MF. All metrics are averaged across scenes, except for the yaw, pitch, and roll errors for which the median is reported. + +Implementation Details: We turn on $\mathcal{L}_{\text {ort }}$ and $\mathcal{L}_{\text {centr }}$ after 500 steps and linearly increase their weights to the specified values over the next 2500 steps. Also, we randomly sample rays for one-third of every batch size and select their left and upper neighbor to form a triplet to facilitate obtaining explicit surface normals. For other implementation details, please refer to the supplementary material. + +Note that in addition to RGB images, ManhattanDF [14] requires knowing the wall and floor semantics, and the exact floor axis $\mathsf{n}_z$ . In contrast, our method only requires RGB images and the assumed Manhattan prior to holding true. Therefore, this comparison aims to get a better insight into what can be achieved without leveraging additional labels. + +# 5.2. Datasets + +Hypersim [31] is a photorealistic synthetic dataset consisting of indoor scenes. It was created by leveraging a large repository of scenes created by professional artists, with 461 indoor scenes in total. This dataset is geometry-rich, containing a lot of details and lighting sources. Each scene has one or more camera trajectories available, where each trajectory has up to 100 views rendered in $1024 \times 768$ . For each scene, camera calibration information is provided, as well as detailed per-pixel labels such as depth and surface normals. After cleaning up the scenes by discarding a few with insufficient camera views, as well as other problems, we were left with 435 scenes. We also kept only one camera trajectory per scene. We then evaluated the InstantNGP [25] baseline on all 435 scenes and made three divisions based on its PSNR on unseen views. Hypersim-A contains 20 scenes where the baseline was performing well above average, Hypersim-B contains 20 scenes where the baseline had near-average performance, and Hypersim-C contains 10 scenes where the baseline had a below-average performance. For each scene, we randomly assigned half views to the training split and the rest to the test split. We note that the test split often contains views that were partially or completely unobserved during training. We also note that Hypersim is visually very realistic, geometry-rich, and challenging, considering it is synthetic. This can be subjectively observed by inspecting rendered views. + +ScanNet [8] is a real-world dataset consisting of indoor scenes. It was collected using a scalable RGB-D capture system that includes automated surface reconstruction and crowd-sourced semantic annotation. The dataset contains 1613 indoor scenes, which are annotated with ground-truth camera poses, surface reconstructions, and instance-level semantic segmentations. We use three scenes that were used + +Table 1: Experiments on Hypersim. We observe that our method consistently outperforms the baseline, as well as the ManhattanDF. This is very interesting since, unlike ManhattanDF, we do not use any additional labels during training. Finally, we see that it does not matter for our method whether the MF is known beforehand or not. Therefore, the additional knowledge of MF is neither necessary nor helpful. + +
PSNR↑SSIM↑Normals°↓Pitch°↓Roll°↓Yaw°↓D-MAE↓D-RMSE↓
Scenes AInstantNGP [25] (baseline)25.860.87157.126.186.4619.250.0640.102
ManhattanDF [14]26.510.86840.691.231.015.080.0530.087
Ours27.200.86437.300.400.500.520.0530.093
Ours (MF known)27.210.85635.590.250.260.450.0520.091
Scenes BInstantNGP [25] (baseline)20.750.81160.126.067.9215.870.1050.151
ManhattanDF [14]21.870.82650.502.552.0611.690.0790.121
Ours22.450.81654.081.191.351.810.0800.127
Ours (MF known)22.510.81350.590.510.650.550.0780.126
Scenes CInstantNGP [25] (baseline)17.790.74064.297.454.5515.140.1300.174
ManhattanDF [14]18.330.77056.083.203.4110.250.1030.147
Ours19.430.76854.795.372.244.240.0940.133
Ours (MF known)19.290.76455.123.644.039.480.0940.135
194 scenesInstantNGP [25] (baseline)20.470.78361.346.566.9921.750.1040.146
ManhattanDF [14]20.940.79452.811.722.3213.480.0970.140
Ours21.630.78652.011.871.943.770.0850.126
+ +in [14], where one-tenth of all views were uniformly sampled, leaving 303-477 views per scene. For each scene, the training and test split both contain half of the total views. + +Replica [38] is a synthetic dataset featuring a diverse set of 18 indoor scenes. Each scene is equipped with photorealistic textures, allowing one to render realistic images from arbitrary camera poses. We use five scenes that were rendered and prepared in [58], where the rendered views also contain semantic segmentation labels. Each scene contains 900 views generated from random 6-DOF trajectories similar to hand-held camera motions in $640 \times 480$ resolution. We select 75 views for training and 75 for testing in each scene. + +# 5.3. Results + +Experiments on Hypersim: We summarize the experiments performed on Hypersim in Table 1. Our proposed method achieves clear improvements in comparison to the InstantNGP baseline, in terms of novel-view rendering, normals estimation, depth estimation, as well as MF recovery. This is consistent across different scene difficulties, namely splits A, B, and C. Furthermore, our method also outperforms the ManhattanDF SotA. This is very interesting since, unlike ManhattanDF, we do not use any additional labels during training, other than the ground truth RGB. The ManhattanDF has partial access to ground truth MF axes, as well as wall and floor semantics, which directly implies having sparse surface normals ground truth. However, there are lots of Manhattan objects and areas in realistic scenes such as in Hypersim, other than the walls and floors. Our self-supervised method leverages the presence of many such objects and areas, to facilitate imposing geometrical constraints on the implicit representation during learning. This can be visually observed in Figure 3. Moreover, we observe that it does not matter for our method whether the MF is known beforehand, since our method recovers the + +![](images/931cc711d1421ab11c9a7fda1fcc78f8f7fed9e3591eae4d34bbccc38f485229.jpg) + +![](images/712b45e1e66298b516e9dde262be8b9d4f5749c2a89ceb19f419e817ffea2ed6.jpg) +(a) InstantNGP [25] (baseline) +(c) Ours +Figure 3: Leveraging the Manhattan prior (Hypersim-A). Our method leverages the presence of many Manhattan objects and areas in the scene in a self-supervised fashion, which facilitates imposing geometrical constraints during learning. Our method offers plausible normals – sometimes with missing details – that help to better model the radiance fields. In contrast, ManhattanDF leverages the Manhattan prior through labels of only walls and floors. + +![](images/16bba075f6b5187fecd171f0d629f1cdfc5791a03dab16f32f4405548dc9dd55.jpg) + +![](images/f3eb669f9bee957446c057f629488a204a95f5f16a38826bd2ec3710f28e4e2b.jpg) +(b) ManhattanDF [14] +(d) Grond truth + +MF in the clustering step automatically, and therefore similar performance is achieved in both cases. This is particularly exciting because our self-supervised method performs similarly to the supervised one (using the groundtruth MF). Finally, we show large-scale experiments on a larger set of 194 scenes, which lead to the same conclusions. We observed that the semantic loss of the ManhattanDF sometimes causes convergence issues on difficult scenes, which we partially alleviated with class weighting and label smoothing. Therefore, we report results on only 194 scenes where ManhattanDF converged. Additional results on all 435 scenes are provided in the supplementary. + +Experiments on ScanNet: In Table 2, we examine the behavior of our proposed method on real-world indoor scenes. + +![](images/7f03d4b048099205cde62a4252f74d5eed9eb1b2da95cea442227198e5b90389.jpg) +Figure 4: Qualitative results. Our method leverages many Manhattan objects and surfaces in the scene, which improves the implicit geometrical representation compared to the baseline. Unlike ManhattanDF, our method relies on many cues other than the walls & floors, which leads to a better representation of some objects (e.g. the white table in the bottom right corner of the example from columns 1-3). For more qualitative results, please refer to the supplementary material. + +Table 2: Experiments on real-world ScanNet data. Our method outperforms both the baseline and ManhattanDF. This is the case both with and without supervising with sparse depth from SfM. + +
PSNR↑SSIM ↑Depth↓-MAE
InstantNGP [25] (baseline)17.780.5870.119
RegNeRF [26]18.730.6180.102
ManhattanDF [14]18.680.6140.112
Ours20.790.6430.072
+ additional sparse depth from SfM
InstantNGP [25] (baseline)20.700.6310.048
ManhattanDF [14]21.530.6400.052
Ours22.250.6670.033
+ +Table 3: Experiments on Replica. Our method outperforms the baseline, and it shows similar performance as ManhattanDF, which leverages additional labels during training. + +
PSNR↑SSIM ↑Depth↓-MAE
InstantNGP [25] (baseline)34.300.9440.022
Semantic-NeRF [58]34.080.9380.014
ManhattanDF [14]35.240.9440.008
Ours35.130.9440.011
+ +Our method achieves clear improvements in comparison to the InstantNGP baseline, as well as to ManhattanDF, in terms of all measured metrics. Additionally, following the experimental setting of ManhattanDF [14], we train all methods with additional sparse depth supervision – where the sparse depth is obtained from the SfM pipeline. Our proposed method is also superior in this setting. + +Experiments on Replica: We summarize experiments performed on the Replica dataset in Table 3. It can be observed that the InstantNGP baseline already performs very + +![](images/20b87505bcc4efa99a6f596cbe137a2db8ddb6c04dcb80aee8ec3af502bf34e7.jpg) +Improvements of our method w.r.t. scene difficulty +Figure 5: Improvements vs. scene difficulty. Our method achieves the biggest improvements on scenes of hard and moderate difficulty, by leveraging many Manhattan objects and surfaces. + +well on this dataset, leaving not much room for further improvements. Therefore, we consider Replica as an easier dataset. Nevertheless, our method still performs better than the baseline. As expected, ManhattanDF also improves the baseline similarly. Recall that the ManhattanDF uses additional labels for supervision. We also note that Replica has noticeably more walls and floors, and less of other objects, compared to other more complex datasets. + +Qualitative results: We depict the visual results of the discussed experiments in Figure 4. Our method leverages many Manhattan objects and surfaces in the scene, which improves the geometrical structure of 3D compared to the InstantNGP baseline. This is visible in surface normals and depth, obtained using volume rendering. Furthermore, we observe that, unlike ManhattanDF, our method relies on many Manhattan cues other than the walls & floors. This leads to a better representation of some such objects, e.g. the white table in the bottom right corner of the example + +Table 4: Sparse training views. Our proposed method clearly exhibits the best performance, for all cases of input view sparsity. + +
PSNR↑SSIM ↑Depth↓-MAE
12 viewsInstantNGP [25]18.020.7060.138
ManhattanDF [14]19.450.7500.116
Ours20.500.7600.104
9 viewsInstantNGP [25]16.790.6610.154
ManhattanDF [14]18.040.7140.130
Ours19.140.7280.120
6 viewsInstantNGP [25]15.750.5820.178
ManhattanDF [14]16.000.6390.159
Ours16.670.6670.158
+ +Table 5: Ablation study on Hypersim-A. Both of our proposed loss terms contribute to the overall performance. + +
PSNR↑Norm.°↓Depth↓Rot.°↓
Only Limg25.8657.120.06410.63
+ Lort27.2150.070.0585.39
+ Lctr27.0636.090.0530.57
+ Lort + Lctr (Ours)27.2037.300.0530.47
Ours + MF known27.2135.590.0520.32
Ours (no delay)27.0635.780.0540.39
Ours (no w lin.)27.0137.860.0550.60
+ +from columns 1-3. Moreover, we observe that our method is able to cope with difficult scenes and views, where other methods struggle. For more qualitative results, please refer to the supplementary material. + +Improvements with respect to scene difficulty: We analyze the improvements by our method for different scene difficulties. We decided on scene difficulty based on the novel-view rendering performance of the InstantNGP baseline. In Figure 5, we see that our method brings the most benefits for scenes of hard and moderate difficulties, thanks to the Manhattan scene prior. + +Sparse training views: In order to gain more insight, we examine the behavior of training neural radiance fields with sparse training input views. The results on the Hypersim-A dataset are presented in Table 4. Our proposed method clearly outperforms the InstantNGP baseline, as well as ManhattanDF, when trained with 6, 9, and 12 input views. + +Finding the MF: We test the robustness of our proposed method for finding the Manhattan frame, by introducing a simultaneous offset $\alpha$ in the yaw, pitch, and roll on the canonical MF. The experiments are performed on Hypersim A and can be found in Figure 6. In Figure 6a we observe that the novel-view rendering quality remains largely robust to the rotation offset $\alpha$ . We note that we increase the scene bounding box by the same factor for all experiments in Figure 6, to make sure that objects remain within the voxel grid for maximal $\alpha$ . This slightly decreases the resolution of grid element representations, so PSNR is slightly lower than in Table 1. Furthermore, in Figure 6b we see that the MF estimation remains robust to the rotation offset. + +Ablation study: We report our ablation study in Table 5. Both of our proposed losses contribute to the overall performance. Furthermore, turning on $\mathcal{L}_{\text{ort}}$ and $\mathcal{L}_{\text{centr}}$ after + +![](images/c81a6c5747db5b77d5cf4c8f7a5a0a2d9c326fdac240db57c8ed609ff0fc073c.jpg) +Rotation offset effects on novel-view rendering +(a) Novel-view rendering quality remains robust to the rotation offset $\alpha$ . + +![](images/aa753d85a2f79945af025e76515d98191ad5919d49b0f2de87582429492c77a2.jpg) +Rotation offset effects on Manhattan frame estimation +(b) MF estimation remains robust to the rotation offset $\alpha$ +Figure 6: Finding the MF. We test the robustness of our method by introducing the rotation offsets on the canonical MF. + +500 steps, and linearly increasing their weight also helps slightly. Please refer to the supplementary for more details. + +# 6. Conclusion + +We demonstrated the possibility of exploiting the Manhattan scene prior without needing any additional supervision. This is achieved by performing robust clustering of explicit normals, followed by the search of the Manhattan frame (MF) whose existence is based on the assumed prior. The sought MF is obtained from the orthogonal clusters, which are later used to self-supervise the neural representation learning. Our self-supervision encourages the normals of Manhattan surfaces to group into three orthogonal directions. Our experiments on three indoor datasets demonstrate that the proposed method not only benefits from building parts (such as walls and floors) but also exploits many other Manhattan scene parts (such as furniture). Both quantitative and qualitative evaluations reveal the benefit of the proposed method in terms of, improved performance over the established baselines and competitive results to state-of-the-art methods that use additional labels for supervision. Our method has the potential to be extended in other higher expressiveness scene priors, such as Atlanta world and the mixture of Manhattan frames. + +Limitations: One limitation of our method is that it sometimes produces surface normals with missing details or "blocky" artifacts. Nevertheless, this usually offers better novel-view RGB rendering, compared to not imposing any structure priors. Another limitation is that our method is not beneficial for very easy scenes. For a detailed discussion of limitations, please refer to the supplementary material. + +# Acknowledgments + +This research was co-financed by Innosuisse under the project Know Where To Look, Grant 59189.1 IP-ICT. Also, Carlos Oliveira helped with splendid visual renderings. + +# References + +[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022. +[2] Stephen T Barnard. Interpreting perspective images. Artificial intelligence, 21(4):435-462, 1983. +[3] Jean-Charles Bazin, Hongdong Li, In So Kweon, Cedric Demonceaux, Pascal Vasseur, and Katsushi Ikeuchi. A branch-and-bound approach to correspondence and grouping problems. IEEE transactions on pattern analysis and machine intelligence, 35(7):1565-1576, 2012. +[4] Jean-Charles Bazin, Yongduek Seo, Cédric Demonceaux, Pascal Vasseur, Katsushi Ikeuchi, Inso Kweon, and Marc Pollefeys. Globally optimal line clustering and vanishing point estimation in manhattan world. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 638–645. IEEE, 2012. +[5] Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag, Berlin, Heidelberg, 2006. +[6] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. +[7] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 941-947. IEEE, 1999. +[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. +[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. +[10] Patrick Denis, James H Elder, and Francisco J Estrada. Efficient edge-based methods for estimating Manhattan frames in urban imagery. In European conference on computer vision, pages 197-210. Springer, 2008. +[11] Clara Fernandez-Labrador. Indoor scene understanding using non-conventional cameras. PhD thesis, 2020. +[12] Yasutaka Furukawa, Brian Curless, Steven M Seitz, and Richard Szeliski. Manhattan-world stereo. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1422-1429. IEEE, 2009. +[13] Wuwei Ge, Yu Song, Baichao Zhang, and Zehua Dong. Globally optimal and efficient Manhattan frame estimation by delimiting rotation search space. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 15213-15221, 2021. +[14] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022. +[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. +[16] Kyungdon Joo, Tae-Hyun Oh, Junsik Kim, and In So Kweon. Globally optimal Manhattan frame estimation in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1763-1771, 2016. +[17] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. +[18] Fu Li, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerf-posed: A first-reconstruct-thenregress approach for weakly-supervised 6d object pose estimation. arXiv preprint arXiv:2203.04802, 2022. +[19] Haoang Li, Jian Yao, Jean-Charles Bazin, Xiaohu Lu, Yazhou Xing, and Kang Liu. A monocular slam system leveraging structural regularity in manhattan world. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 2518-2525. IEEE, 2018. +[20] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. +[21] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15702-15712, 2022. +[22] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. +[23] Gerard F McLean and D Kotturi. Vanishing point detection by line clustering. IEEE Transactions on pattern analysis and machine intelligence, 17(11):1090-1095, 1995. +[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. cite arxiv:2003.08934Comment: ECCV 2020 (oral). Project page with videos and code: http://tancik.com/nerf. +[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. + +[26] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. +[27] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. +[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. +[29] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in manhattan world. In Proceedings of the IEEE International Conference on Computer Vision, pages 882-890, 2017. +[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. +[31] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. +[32] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022. +[33] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. +[34] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lučić, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6229-6238, 2022. +[35] Gautam Singh and Jana Kosecka. Visual loop closing using gist descriptors in manhattan world. In ICRA omnidirectional vision workshop, pages 4042-4047, 2010. +[36] Julian Straub, Nishchal Bhandari, John J Leonard, and John W Fisher. Real-time manhattan world rotation estimation in 3d. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1913-1920. IEEE, 2015. +[37] Julian Straub, Oren Freifeld, Guy Rosman, John J Leonard, and John W Fisher. The Manhattan frame model—manhattan world inference in the space of surface normals. IEEE transactions on pattern analysis and machine intelligence, 40(1):235-249, 2017. + +[38] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. +[39] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. +[40] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light field neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8269-8279, 2022. +[41] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2022. +[42] Frank A Van den Heuvel. 3d reconstruction from a single image using geometric constraints. ISPRS Journal of Photogrammetry and Remote Sensing, 53(6):354-368, 1998. +[43] Carlos A Vanegas, Daniel G Aliaga, and Bedrich Benes. Building reconstruction using manhattan-world grammars. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 358-365. IEEE, 2010. +[44] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022. +[45] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021. +[46] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. arXiv preprint arXiv:2206.13597, 2022. +[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021. +[48] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. +[49] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. + +[50] Horst Wildenauer and Allan Hanbury. Robust camera self-calibration from monocular images of manhattan worlds. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2831-2838. IEEE, 2012. +[51] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. +[52] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K Wong. Ps-nerf: Neural inverse rendering for multi-view photometric stereo. In European Conference on Computer Vision, pages 266–284. Springer, 2022. +[53] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K Wong. S3-nerf: Neural reflectance field from shading and shadow under a single viewpoint. arXiv preprint arXiv:2210.08936, 2022. +[54] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. +[55] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. +[56] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. +[57] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5657-5665, 2016. +[58] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. \ No newline at end of file diff --git a/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/images.zip b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..499dceb409fb43004f525f8c74d276bf8cf83de6 --- /dev/null +++ b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:879bfaf0f2be2369744761966b876dc7872f22777ede1d8299382498328bb1ff +size 633431 diff --git a/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/layout.json b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..698615ba394cafc9d8a1705637a4ae9cb5d34ec6 --- /dev/null +++ b/2023/Surface Normal Clustering for Implicit Representation of Manhattan Scenes/layout.json @@ -0,0 +1,9271 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "text", + "content": "Surface Normal Clustering for Implicit Representation of Manhattan Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 142, + 433, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 142, + 433, + 157 + ], + "spans": [ + { + "bbox": [ + 159, + 142, + 433, + 157 + ], + "type": "text", + "content": "Nikola Popovic1, Danda Pani Paudel1,2, Luc Van Gool1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 157, + 433, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 157, + 433, + 185 + ], + "spans": [ + { + "bbox": [ + 159, + 157, + 433, + 185 + ], + "type": "text", + "content": "1Computer Vision Laboratory, ETH Zurich, Switzerland \n2INSAIT, Sofia University, Bulgaria" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 187, + 421, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 187, + 421, + 199 + ], + "spans": [ + { + "bbox": [ + 172, + 187, + 421, + 199 + ], + "type": "text", + "content": "{nipopovic, paudel, vangool}@vision.ee.ethz.ch" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 251, + 289, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 526 + ], + "type": "text", + "content": "Novel view synthesis and 3D modeling using implicit neural field representation are shown to be very effective for calibrated multi-view cameras. Such representations are known to benefit from additional geometric and semantic supervision. Most existing methods that exploit additional supervision require dense pixel-wise labels or localized scene priors. These methods cannot benefit from high-level vague scene priors provided in terms of scenes' descriptions. In this work, we aim to leverage the geometric prior of Manhattan scenes to improve the implicit neural radiance field representations. More precisely, we assume that only the knowledge of the indoor scene (under investigation) being Manhattan is known - with no additional information whatsoever - with an unknown Manhattan coordinate frame. Such high-level prior is used to self-supervise the surface normals derived explicitly in the implicit neural fields. Our modeling allows us to cluster the derived normals and exploit their orthogonality constraints for self-supervision. Our exhaustive experiments on datasets of diverse indoor scenes demonstrate the significant benefit of the proposed method over the established baselines. The source code will be available at https://github.com/nikola3794/normal-clustering-nerf." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Above " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " images ever taken are estimated to involve human-made architectural structures, with a substantial share of indoor scenes [11]. These scenes often exhibit strong structural regularities, including flat and texture-poor surfaces in some axis-aligned Cartesian coordinate system - also known as Manhattan world [7, 12]. Paradoxically, these regularities may hinder the 3D modeling process if it is unaware of the human-made scene priors. In fact, several computer vision works have even benefited from the knowledge of the Manhattan world for the task of 3D scene reconstruction [43, 12], camera localization [19], self-calibration [50], and more [37, 35, 29, 57]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 228, + 547, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 419 + ], + "type": "text", + "content": "For calibrated multi-view cameras, 3D inversion using implicit neural representations [24, 47, 54] is becoming increasingly popular due to their remarkable performance and recent efficiency developments [55, 30, 25, 51]. Meanwhile, such representations are known to benefit from additional supervision in the form of depth [1, 9, 32], normals [18], semantics [45, 58, 17, 15], local-regularization [46, 44, 26], local planar patches [21], or their combinations [56, 14, 21]. In this context, a notable recent work ManhattanSDF [14] demonstrates the benefit of exploiting the high-level geometric prior for structured scenes. More precisely, ManhattanSDF [14] uses the known semantic regions to impose the planar geometry prior of floors and walls under the Manhattan scene assumption. During this process, the exact normal of the floor and partial normals of the walls are assumed to be known, with respect to the camera coordinates." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 422, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 579 + ], + "type": "text", + "content": "We aim to improve the 3D neural radiance field representations for calibrated multi-view cameras in indoor Manhattan scenes, with no further assumptions. In other words, we consider that the structural and semantic information is not available, unlike ManhattanSDF [14]. In addition to the floor and walls used in ManhattanSDF [14], we wish to exploit the Manhattan prior of many other indoor scene parts (e.g. tables and wardrobes). More importantly, we consider that the Manhattan coordinate frame is also unknown. Our assumptions (of unknown semantics and Manhattan frame) on one hand make our setting very practical. On the other hand, those assumptions make the problem of exploiting the Manhattan scene prior for 3D inversion very challenging." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "The virtue of the Manhattan world assumption comes from its simplicity, allowing us to intuitively reason about the geometry of a wide range of complex scenes/objects such as cities, buildings, and furniture. However, such reasoning often requires the axis-aligned Cartesian coordinate frame, also known as the Manhattan frame (MF), to be known [37]. Unfortunately, recovering the Manhattan frame directly from images is not an easy task [7, 37]. Therefore, several methods have been developed until recently [10, 4, 36, 16, 13] to recover the Manhattan frame, relying upon the known 3D reconstruction or image primi" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17860" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 200, + 205 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 200, + 205 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 200, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 200, + 205 + ], + "type": "image", + "image_path": "5775e0ce3426ff3c30121e5471a1a73efcd3a6787f004adaf4ff305745a799c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 258 + ], + "type": "text", + "content": "Figure 1: Illustration of the proposed method. We compute one surface normal for each ray triplet, using 3D surface points derived from rendered depths (left). The computed normals (from many ray triplets) are clustered to obtain the MF (middle). The non-Manhattan and noisy surfaces are handled by a robust orthogonal normals search, which is later used for self-supervision through the Manhattan prior (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 201, + 70, + 414, + 205 + ], + "blocks": [ + { + "bbox": [ + 201, + 70, + 414, + 205 + ], + "lines": [ + { + "bbox": [ + 201, + 70, + 414, + 205 + ], + "spans": [ + { + "bbox": [ + 201, + 70, + 414, + 205 + ], + "type": "image", + "image_path": "a7eab9213b88cf5b4434efa310c8a6b9ddeb780d2f63ac251c9e6b7af5deea13.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 414, + 67, + 547, + 205 + ], + "blocks": [ + { + "bbox": [ + 414, + 67, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 414, + 67, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 414, + 67, + 547, + 205 + ], + "type": "image", + "image_path": "bb6479a6e5384015fae3d31865f05580f1766214e15d38251e4f102df4b94293.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 287, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 330 + ], + "type": "text", + "content": "tives (eg. lines, planes). We wish to exploit the Manhattan prior for improving the 3D representation, without needing to know MF beforehand. Instead, our experiments reveal that knowing the MF beforehand offers no additional benefit in Manhattan-prior aware radiance field representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 331, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 617 + ], + "type": "text", + "content": "In this work, we propose a method that jointly learns the Manhattan frame and neural radiance field, from calibrated multi-view in indoor Manhattan scenes, in an end-to-end manner using the recent efficient backbone of Instant-NGP [25]. The proposed method requires no additional information to exploit the Manhattan prior and relies on the explicitly derived normals in the implicit neural fields. We use batches of three neighboring rays, whose effective surface's local piece-wise planarity is assumed to derive the explicit normals by algebraic means. In pure, complete, and enclosed Manhattan scenes, these normals form six clusters corresponding to three orthogonal and other three parallel counterpart planes. However real scenes consist of non-Manhattan scene parts and missing planes. Therefore, we use a robust method that uses minimal three orthogonal clusters to recover the sought Manhattan frame. As in the literature, we seek a rotation matrix whose entries are directly derived from the orthogonal clusters of normals, to align the Manhattan frame. The recovered MF is then used to encourage the derived normals to be axis-aligned for self-supervision. A graphical illustration of our method is presented in Figure 1. Our extensive experiments demonstrate the robustness and utility of our method in improving the implicit 3D in neural radiance field representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 618, + 277, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 618, + 277, + 629 + ], + "spans": [ + { + "bbox": [ + 59, + 618, + 277, + 629 + ], + "type": "text", + "content": "The major contributions of our paper are listed below:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 46, + 630, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 666 + ], + "type": "text", + "content": "- We address the problem of exploiting the structured-scene knowledge without requiring any dense or localized scene priors, for the first time in this paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 666, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 701 + ], + "type": "text", + "content": "- We present a method that successfully exploits the Manhattan scene prior with an unknown Manhattan frame. The proposed method also recovers the unknown frame." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "- We demonstrate the robustness and utility of our method" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 270, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 318 + ], + "type": "text", + "content": "on three indoor datasets, where our method improves the established baselines significantly. These datasets consist of " + }, + { + "bbox": [ + 305, + 270, + 545, + 318 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 305, + 270, + 545, + 318 + ], + "type": "text", + "content": " scenes, making our method tested in significantly more scenes than the state-of-the-art methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 332, + 397, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 332, + 397, + 344 + ], + "spans": [ + { + "bbox": [ + 306, + 332, + 397, + 344 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 484 + ], + "type": "text", + "content": "Implicit neural representation of 3D: Since the foundational work of Mildenhall et al. [24], the implicit neural representation of 3D scenes has advanced on various fronts. These fronts include representation [54, 40], generalization [49, 28, 20, 22, 34], generation [27, 6], and efficient methods [30, 55, 25]. We rely on a recent method Instant-NGP [25] developed by Mueller et al., as our backbone. Our choice is made primarily based on the computational efficiency during both training and inference. Thanks to the offered computational efficiency, we are able to conduct large-scale experiments on several scenes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "Auxiliary supervision methods: In addition to the images, other inputs such as depth [33, 39, 9], semantics [58, 15, 17, 41], normal [52, 53], and their combinations [56, 14, 21] are shown to be beneficial on improving the neural radiance field representation. In this regard, these auxiliary supervisions often use ground-truth labels. It is needless to say that the need for ground-truth supervision is not desired, whenever possible. Therefore, recent methods use labels predicted by some pre-trained networks [56] or recovered from the structure-from-motion (SfM) pipeline [9, 21]. One notable work ManhattanSDF [14] exploits the Manhattan prior without requiring any SfM reconstruction. However, ManhattanSDF requires (a) semantics of the scene parts and (b) the Manhattan frame, to be known. We argue that such labels required for auxiliary supervision cannot always be relied upon, due to domain gaps, poor reconstruction of texture-less regions, and additional computational needs, to list a few. Therefore, we do not use any additional labels for auxiliary supervision." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17861" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "Manhattan frame estimation: Since the early works of Bernard [2], Manhattan structure reasoning is done directly on/from images by detecting the so-called the vanishing-points (VPs) [23, 42]. In fact, the problem of detecting three orthogonal VPs is equivalent to finding the MF in 3D for the calibrated multi-view setting [3]. Note that the knowledge of Manhattan structure has been used in several computer vision works [43, 12, 19, 50, 37, 35, 29, 57]. When unknown, most methods implicitly or explicitly estimate the MF to leverage the Manhattan scene prior. In [37], Straub et al. have demonstrated that the MF can be efficiently represented in and recovered from the space of surface normals. We use a similar formulation as [37], using the surface normals derived explicitly from the implicit neural fields." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 251, + 201, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 201, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 201, + 264 + ], + "type": "text", + "content": "3. Background and Notations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": "Manhattan frame (MF) is a coordinate system that is defined by the structure building orthogonal planes of Manhattan scenes. We consider unknown MF since the scene planes and their geometric relationships are unknown. Let the unknown MF differs from the world-frame (WF), used for the 3D representation, by rotation " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathsf{R} \\in SO(3)" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": ". We denote three orthogonal axes in MF by " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{\\mathsf{e}_x, \\mathsf{e}_y, \\mathsf{e}_z\\}" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": ". Without loss of generality, let the axes' coordinates be " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathsf{e}_x = [1,0,0]^{\\mathsf{T}}" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathsf{e}_y = [0,1,0]^{\\mathsf{T}}" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathsf{e}_z = [0,0,1]^{\\mathsf{T}}" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": ". Note that these axes align with the normals of the respective scene building planes, in the WF. Let " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\{\\mathsf{n}_i\\}_{i=1}^m" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": " be a set of 3D normals of all the scene planes. Then, the rotation " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathsf{R}" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": ", from world to Manhattan frame, aligns the normals " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathsf{n} \\in \\mathcal{N}" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": " to the Manhattan axes " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathsf{e} \\in \\mathcal{E}" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\mathsf{Rn}_i \\in \\mathcal{E}" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "inline_equation", + "content": "\\forall \\mathsf{n}_i \\in \\mathcal{N}" + }, + { + "bbox": [ + 46, + 270, + 289, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": "We are interested to recover " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\mathsf{R}" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": " from a set of normals " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": ". For this, the above set-to-set assignment alone is not sufficient. This requires the element-wise assignment between sets " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": ". To do so, we divide the set " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": " into three orthogonal subsets " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_x, \\mathcal{N}_y," + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_z" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": ". Now, for any triplet of " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z\\}" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": " from the corresponding orthogonal subsets, we aim to establish the condition " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "[\\mathsf{e}_x, \\mathsf{e}_y, \\mathsf{e}_z] = \\mathsf{R}[\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z]" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": ". Note that the assignment condition results into " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\mathsf{R} = [\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z]^{\\top}" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": ". Therefore, the problem of recovering MF from a given set of normals boils down to finding three normals from orthogonal subsets. At this point, one issue regarding robustness remains pending. More precisely, we wish to recover " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\mathsf{R}" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": " for a noisy set of normals " + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 438, + 289, + 605 + ], + "type": "text", + "content": ", with potentially overwhelmingly many outliers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": "We use a robust method to recover " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{R}" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": " from given normals " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": ". In principle, " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{R}" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": " can be recovered from minimal two orthogonal normals, with one additional normal for disambiguation by sign correction. The recovered rotation can then be validated by consensus for robust recovery [37]. Alternatively, one can also perform the robust estimation of the orthogonal subsets of " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": ", followed by solving " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "[\\mathsf{e}_x, \\mathsf{e}_y, \\mathsf{e}_z] = \\mathsf{R}[\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z], \\forall \\mathsf{n}_x \\in \\mathcal{N}_x, \\mathsf{n}_y \\in \\mathcal{N}_y, \\mathsf{n}_z \\in \\mathcal{N}_z" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{R} \\in SO(3)" + }, + { + "bbox": [ + 46, + 606, + 289, + 714 + ], + "type": "text", + "content": ". For computational reasons, we estimate" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "the robust subsets by clustering. The orthogonal subsets are then obtained by choosing the three most orthogonal clusters. Later, the obtained cluster centers are used to estimate the MF (or to enforce its existence) represented by R. Please refer to Section 4.2 for more details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 144, + 362, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 362, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 362, + 156 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 164, + 547, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 164, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 164, + 547, + 427 + ], + "type": "text", + "content": "From a set of calibrated images, we model the 3D scene using the neural radiance field representation. In the process, we wish to exploit the Manhattan scene prior, without the knowledge of the Manhattan frame. This problem is addressed by jointly optimizing for the neural radiance field and the Manhattan frame estimation. The complete pipeline of our method is illustrated in Figure 2. As shown, our method consists of three units: (a) Explicit normal modeling, (b) Robust estimation of orthogonal normals, and (3) Self-supervision by Manhattan prior. In the first stage, we shoot a batch of three rays which allows us to estimate the surface normal using an algebraic method. As the estimated normals are bound to be noisy (with outliers), we perform their clustering to obtain the most orthogonal clusters (representing the Manhattan frame) in a robust manner. The obtained clusters are then used to estimate the sought MF, in the form of a rotation matrix, using the method discussed in Section 3. In the final step, we use the estimated Manhattan frame to encourage the close-by normals to be Manhattan-like and to enforce a stricter orthogonality constraint for self-supervision. In the following, we will present the details of three units of our method in different Subsections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 435, + 455, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 435, + 455, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 455, + 449 + ], + "type": "text", + "content": "4.1. Explicit Normal Modelling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": "At any given view, we consider the color " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "c = C_{\\theta}(\\mathsf{r})" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "d = D_{\\theta}(\\mathsf{r})" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": " for a ray " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathsf{r}" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": " emanating from the corresponding camera center " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathsf{o}" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": " is obtained using the volumetric rendering of the implicit neural radiance field, represented by a neural network parameterized by " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": " as in [24]. Then, the location of the 3D surface point is given by, " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathsf{x} = \\mathsf{o} + d.\\mathsf{r}" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": ". We process triplets of three neighboring rays. Let " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\mathsf{r}_1,\\mathsf{r}_2,\\mathsf{r}_3\\}" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": " be such a triplet, whose corresponding surface points are given by " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\times_1,\\times_2,\\times_3\\}" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": ". Now, for the ray triplet " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": ", we obtain the explicit surface normal " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathsf{n}" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": " with the help of the point triplet " + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 454, + 547, + 598 + ], + "type": "text", + "content": ", using the following mapping and algebraic operation," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 607, + 547, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 607, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 310, + 607, + 547, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {T} \\rightarrow \\mathcal {X} \\rightarrow v = \\left(x _ {1} - x _ {2}\\right) \\times \\left(x _ {1} - x _ {3}\\right)\\rightarrow n = \\frac {\\operatorname {s i g n} (\\mathrm {o} ^ {\\mathrm {T}} v) v}{\\| v \\|}. \\tag {1}", + "image_path": "e395906aa92a5894678440fb7df801a1067094437a93b97dbebdde8eb4130376.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " is a vector orthogonal to the plane passing through 3D points in triplet " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": ". We obtain the oriented normal " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathsf{n}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " by normalizing " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " and correcting its sign by ensuring that the camera center " + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathsf{o}" + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": " is in front of the estimated plane. A graphical illustration of the surface normal estimation is provided in Figure 1 on the left. We select a random set" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17862" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 532, + 168 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 532, + 168 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 532, + 168 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 532, + 168 + ], + "type": "image", + "image_path": "d3faeefa59b922e950680a21324596f55cda204969108ec9d3f44a70a0fcc6e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 171, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 45, + 171, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 45, + 171, + 547, + 208 + ], + "type": "text", + "content": "Figure 2: The complete pipeline of our method. We use grid features (GF) and an MLP to represent the radiance field. The explicit normals are derived using the depths obtained from volume rendering. The Manhattan scene prior is exploited by clustering the estimated normals to enforce their orthogonality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 289, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 289, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 289, + 316 + ], + "type": "text", + "content": "of ray triplets " + }, + { + "bbox": [ + 46, + 220, + 289, + 316 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_i\\}_{i = 1}^m" + }, + { + "bbox": [ + 46, + 220, + 289, + 316 + ], + "type": "text", + "content": " from multiple cameras. These ray triplets provide us the surface point triplets " + }, + { + "bbox": [ + 46, + 220, + 289, + 316 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{X}_i\\}_{i = 1}^m" + }, + { + "bbox": [ + 46, + 220, + 289, + 316 + ], + "type": "text", + "content": ". We use point triplets to explicitly derive the surface normals " + }, + { + "bbox": [ + 46, + 220, + 289, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\{\\mathfrak{n}_i\\}_{i = 1}^m" + }, + { + "bbox": [ + 46, + 220, + 289, + 316 + ], + "type": "text", + "content": ", using (1). These normals are later used to recover the unknown Manhattan frame. The simplicity of the explicit normals computed in this paper makes them easy to compute and handle. If needed, normals of different sizes of surface regions could also be computed similarly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 322, + 269, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 269, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 269, + 335 + ], + "type": "text", + "content": "4.2. Robust Estimation of Orthogonal Normals" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": "We are interested to recover the MF from a set of noisy surface normals " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\{\\mathfrak{n}_i\\}_{i=1}^m" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": ". Recall Section 3, the MF can be obtained by robustly recovering a set of three orthogonal normals from " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": ". To do so, we first cluster " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": " clusters " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{C}_i\\}_{i=1}^k" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": ", by using the well-known " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": "-means clustering algorithm [5]. During the clustering, every centroid is " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": " normalized after each iteration, to ensure that they are unit vectors to represent surface normals. In a perfect Manhattan world, there exist only six clusters corresponding to the three orthogonal MF axes " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{\\mathbf{e}_x, \\mathbf{e}_y, \\mathbf{e}_z\\}" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": " and their parallel counterparts. However, real scenes consist of surfaces of different orientations. An additional source of non-Manhattan normals comes from the inaccuracy in the normal estimation. Therefore, we use a clustering technique to robustly recover the orthogonal normals. In the following, we present how the orthogonal clusters are selected from the set of clusters " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{C}_i\\}_{i=1}^k" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": ". Note from Section 3, the selected orthogonal clusters are considered to represent MF defining sets " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_x, \\mathcal{N}_y" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_z" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": ". We proceed by first selecting three orthogonal clusters, say " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_1, \\mathcal{N}_2" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_3" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": ". Later, we assign them to " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_x, \\mathcal{N}_y" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_z" + }, + { + "bbox": [ + 46, + 340, + 287, + 590 + ], + "type": "text", + "content": " to recover MF." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "text", + "content": "For notational ease, we pair clusters and centroids as " + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{U} = \\left\\{\\left(\\mathcal{C}_i,\\mathsf{c}_i\\right)\\right\\}_{i = 1}^k" + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "text", + "content": " , where the centroids are computed by taking the average across the corresponding cluster such that " + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{c}_i = \\frac{1}{|\\mathcal{C}_i|}\\sum_{\\mathsf{n}\\in \\mathcal{C}_i}\\mathsf{n}," + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "text", + "content": " followed by normalizing to a unit vector " + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{c}_i = \\frac{\\mathsf{c}_i}{\\|\\mathsf{c}_i\\|}" + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "text", + "content": " . Leveraging the Manhattan prior through the assumption that Manhattan surfaces dominate the scene, we pick " + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{n}_1 = \\mathsf{c}_f" + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "inline_equation", + "content": "f = \\arg \\max_{i}|\\mathcal{C}_i|" + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "text", + "content": " In other words, we pick the centroid of the largest cluster, as one of the MF axes. Then, we obtain " + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{n}_2 = \\mathsf{c}_s" + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{n}_3 = \\mathsf{c}_t" + }, + { + "bbox": [ + 46, + 591, + 288, + 714 + ], + "type": "text", + "content": " as a solution of the optimization problem" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 220, + 523, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 523, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 523, + 232 + ], + "type": "text", + "content": "Algorithm 1 " + }, + { + "bbox": [ + 306, + 220, + 523, + 232 + ], + "type": "inline_equation", + "content": "(\\mathcal{N}_1,\\mathcal{N}_2,\\mathcal{N}_3,\\mathsf{R}) =" + }, + { + "bbox": [ + 306, + 220, + 523, + 232 + ], + "type": "text", + "content": " findManhattanFrame(N)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 236, + 541, + 304 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 315, + 236, + 539, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 236, + 539, + 249 + ], + "spans": [ + { + "bbox": [ + 315, + 236, + 539, + 249 + ], + "type": "text", + "content": "1. Cluster normals " + }, + { + "bbox": [ + 315, + 236, + 539, + 249 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 315, + 236, + 539, + 249 + ], + "type": "text", + "content": " into, " + }, + { + "bbox": [ + 315, + 236, + 539, + 249 + ], + "type": "inline_equation", + "content": "\\mathcal{U} = \\{(\\mathcal{C}_i,\\mathsf{c}_i)\\}_{i = 1}^k" + }, + { + "bbox": [ + 315, + 236, + 539, + 249 + ], + "type": "text", + "content": " using k-means." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 249, + 529, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 249, + 529, + 259 + ], + "spans": [ + { + "bbox": [ + 315, + 249, + 529, + 259 + ], + "type": "text", + "content": "2. For the largest cluster " + }, + { + "bbox": [ + 315, + 249, + 529, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\in \\mathcal{U}" + }, + { + "bbox": [ + 315, + 249, + 529, + 259 + ], + "type": "text", + "content": ", assign " + }, + { + "bbox": [ + 315, + 249, + 529, + 259 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}, \\mathfrak{c}) \\to (\\mathcal{N}_1, \\mathfrak{n}_1)" + }, + { + "bbox": [ + 315, + 249, + 529, + 259 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 259, + 538, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 259, + 538, + 270 + ], + "spans": [ + { + "bbox": [ + 315, + 259, + 538, + 270 + ], + "type": "text", + "content": "3. Find " + }, + { + "bbox": [ + 315, + 259, + 538, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_s\\in \\mathcal{U},\\mathcal{C}_t\\in \\mathcal{U}" + }, + { + "bbox": [ + 315, + 259, + 538, + 270 + ], + "type": "text", + "content": " minimizing " + }, + { + "bbox": [ + 315, + 259, + 538, + 270 + ], + "type": "inline_equation", + "content": "|\\mathsf{c}_s^\\top \\mathsf{n}_1| + |\\mathsf{n}_1^\\top \\mathsf{c}_t| + |\\mathsf{c}_s^\\top \\mathsf{c}_t|" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 270, + 512, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 270, + 512, + 281 + ], + "spans": [ + { + "bbox": [ + 315, + 270, + 512, + 281 + ], + "type": "text", + "content": "4. Assign " + }, + { + "bbox": [ + 315, + 270, + 512, + 281 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}_s,\\mathsf{c}_s)\\to (\\mathcal{N}_2,\\mathsf{n}_2),(\\mathcal{C}_t,\\mathsf{c}_t)\\to (\\mathcal{N}_3,.,\\mathsf{n}_3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 281, + 541, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 281, + 541, + 294 + ], + "spans": [ + { + "bbox": [ + 315, + 281, + 541, + 294 + ], + "type": "text", + "content": "5. " + }, + { + "bbox": [ + 315, + 281, + 541, + 294 + ], + "type": "inline_equation", + "content": "|\\mathbf{e}_z^{\\mathsf{T}}\\mathfrak{n}_1|\\leq \\frac{1}{\\sqrt{2}}?\\mathfrak{n}_1\\to \\mathfrak{n}_z:(|\\mathbf{e}_y^{\\mathsf{T}}\\mathfrak{n}_1|\\leq \\frac{1}{\\sqrt{2}}?\\mathfrak{n}_1\\to \\mathfrak{n}_y:\\mathfrak{n}_1\\to \\mathfrak{n}_x)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 294, + 513, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 513, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 513, + 304 + ], + "type": "text", + "content": "6. Perform the remaining " + }, + { + "bbox": [ + 316, + 294, + 513, + 304 + ], + "type": "inline_equation", + "content": "\\{\\mathsf{n}_1,\\mathsf{n}_2,\\mathsf{n}_3\\} \\to \\{\\mathsf{n}_x,\\mathsf{n}_y,\\mathsf{n}_z\\}" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 305, + 527, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 305, + 527, + 316 + ], + "spans": [ + { + "bbox": [ + 326, + 305, + 527, + 316 + ], + "type": "text", + "content": "as in step 5 to the closest corresponding canonical axes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 316, + 482, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 316, + 482, + 327 + ], + "spans": [ + { + "bbox": [ + 315, + 316, + 482, + 327 + ], + "type": "text", + "content": "7. Return " + }, + { + "bbox": [ + 315, + 316, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_1, \\mathcal{N}_2, \\mathcal{N}_3" + }, + { + "bbox": [ + 315, + 316, + 482, + 327 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 315, + 316, + 482, + 327 + ], + "type": "inline_equation", + "content": "\\mathsf{R} = [\\mathsf{n}_x, \\mathsf{n}_y, \\mathsf{n}_z]^\\top" + }, + { + "bbox": [ + 315, + 316, + 482, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 337, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 578 + ], + "type": "inline_equation", + "content": "s, t = \\arg \\min_{i,j} |\\mathsf{c}_i^\\top \\mathsf{n}_1| + |\\mathsf{n}_1^\\top \\mathsf{c}_j| + |\\mathsf{c}_i^\\top \\mathsf{c}_j|" + }, + { + "bbox": [ + 304, + 337, + 545, + 578 + ], + "type": "text", + "content": ". In other words, we find two additional cluster centroids providing the most orthogonal triplet. We further merge all selected clusters with their opposites. This is achieved by comparing all cluster pairs. Whenever two centroids are nearby, but opposite in sign, the corresponding clusters are merged with the appropriate sign correction. The procedure of finding " + }, + { + "bbox": [ + 304, + 337, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{N}_1, \\mathcal{N}_2, \\mathcal{N}_3\\}" + }, + { + "bbox": [ + 304, + 337, + 545, + 578 + ], + "type": "text", + "content": " is illustrated in Figure 1 (middle and right), and is summarized in Algorithm 1 from step 1-4. Although the orthogonal clusters with their centroids are sufficient to exploit the desired Manhattan prior, we may wish to recover the MF in the form of a rotation matrix. As such, any arrangement of " + }, + { + "bbox": [ + 304, + 337, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\{\\mathsf{n}_1, \\mathsf{n}_2, \\mathsf{n}_3\\}" + }, + { + "bbox": [ + 304, + 337, + 545, + 578 + ], + "type": "text", + "content": " as a valid rotation matrix offers us a valid MF. We may however often be interested to recover the one which is closest to the world frame. For this reason, we suggest keeping the largest cluster's centroid paired to the closes canonical axis in " + }, + { + "bbox": [ + 304, + 337, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 337, + 545, + 578 + ], + "type": "text", + "content": ". Similarly, we also align one more axis, whereas the last remaining axis gets paired by default. We summarize how to recover MF closest to the world frame in Algorithm 1 in steps 5-7." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 585, + 501, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 585, + 501, + 597 + ], + "spans": [ + { + "bbox": [ + 305, + 585, + 501, + 597 + ], + "type": "text", + "content": "4.3. Self-supervision by Manhattan Prior" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "type": "text", + "content": "The supervision of the implicit neural radiance field, parameterized by " + }, + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "type": "text", + "content": ", is achieved through jointly optimizing three loss terms. The first one is the photometric loss computed as follows, " + }, + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{img} = \\frac{1}{|\\mathcal{R}|}\\sum_{\\mathbf{r}\\in \\mathcal{R}}\\| C_{\\theta}(\\mathbf{r}) - C(\\mathbf{r})\\| _2^2" + }, + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r})" + }, + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "type": "text", + "content": " is the ground-truth color, and " + }, + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 604, + 545, + 700 + ], + "type": "text", + "content": " is the set of rays going through sampled pixel triplets. This loss term is responsible for facilitating the learning of the implicit 3D representation of the investigated scene [24]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Losses from Manhattan prior: We exploit the Manhattan" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17863" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "scene prior for self-supervision, in order to improve the implicit 3D representation without additional ground truth labels. We do so by using the clusters " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{N}_1,\\mathcal{N}_2,\\mathcal{N}_3\\}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " obtained from Algorithm 1. More precisely, for cluster-centroid pairs " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\{(\\mathcal{N}_i,\\mathsf{n}_i)\\}_{i = 1}^3" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " we use the following two losses," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 142, + 287, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 287, + 174 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c t r} = \\frac {1}{3} \\sum_ {i = 1} ^ {3} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {\\mathfrak {n} \\in \\mathcal {N} _ {i}} \\| 1 - \\mathfrak {n} _ {i} ^ {\\top} \\mathfrak {n} \\| _ {1} + \\| \\mathfrak {n} _ {i} - \\mathfrak {n} \\| _ {1}, \\tag {2}", + "image_path": "037f9553057eed97ded2ec135a3a94ea59b4c30c23e7ffe454aca261ab158289.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 177, + 287, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 177, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 287, + 200 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o r t} = \\frac {1}{3} \\left(\\left| n _ {1} ^ {\\top} n _ {2} \\right| + \\left| n _ {1} ^ {\\top} n _ {3} \\right| + \\left| n _ {2} ^ {\\top} n _ {3} \\right|\\right). \\tag {3}", + "image_path": "e0f6493df23a641c702684b3c6b7916c39336582bc41c8ed775e404bb4b2ea02.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "text", + "content": "Here, the loss " + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ctr}" + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "text", + "content": " encourages tighter clusters, while the loss " + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ort}" + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "text", + "content": " enforces the orthogonality among the three clusters. The final loss used to optimize " + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "text", + "content": " is then given by " + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{img} + \\lambda_{ctr}\\mathcal{L}_{ctr} + \\lambda_{ort}\\mathcal{L}_{ort}" + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\lambda_{ctr}" + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\lambda_{ort}" + }, + { + "bbox": [ + 47, + 208, + 287, + 280 + ], + "type": "text", + "content": " are the hyperparameters. Please, refer to Figure 2 for a schematic summary of all three losses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 291, + 128, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 128, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 128, + 304 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 311, + 287, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 287, + 323 + ], + "type": "text", + "content": "5.1. Baselines, Metrics, and Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 329, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 424 + ], + "type": "text", + "content": "InstantNGP [25] (baseline): This method represents the scene as a multi-resolution voxel grid and leverages a hash table of trainable feature vectors, which are used to represent grid elements. The representation is further processed with a small MLP. This allows InstantNGP to be very computationally efficient, requiring in our case around 30 minutes per scene to train and evaluate. Therefore, we use it as the backbone of our method and as our baseline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "content": "ManhattanDF [14]: This method exploits the Manhattan prior by supervising the explicit normals of the floors to align with the known " + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathsf{n}_z" + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "content": ", as well as by supervising normals of the walls to align with two learned orthogonal axes which are also orthogonal to " + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathsf{n}_z" + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "content": ". To achieve this, apart from the RGB images, this method relies on knowing the wall and floor semantics, as well as knowing the exact floor axis " + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathsf{n}_z" + }, + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "content": " (MF partially known). We implemented it on top of the In-stantNGP backbone with density field estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 556 + ], + "type": "text", + "content": "Ours: We implement our proposed method from Section 4 on top of InstantNGP with density field estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "content": "Ours (MF known): We modify our method by assuming that the MF is fully known. We do so by adding additional loss terms " + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{man_i} = \\|1 - \\mathsf{n}_i^{\\mathrm{T}}\\mathsf{m}\\|_1 + \\| \\mathsf{n}_i - \\mathsf{m}\\|_1" + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "content": " for each orthogonal cluster centroid " + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathsf{n}_i" + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "content": " and its closest MF axis " + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathfrak{m}" + }, + { + "bbox": [ + 46, + 558, + 287, + 629 + ], + "type": "text", + "content": " (or a closer opposite counterpart). Thus, we explicitly guide the orthogonal triplet to align with the known MF." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Metrics: To quantitatively evaluate novel view rendering, we use peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) [48]. To evaluate the extracted surface normals of novel rendered views, we use the median angular error. To partially evaluate the quality of the learned implicit 3D structure, we utilize the mean absolute error (MAE) and the root mean square error (RMSE) on the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "depth obtained by volume rendering. Finally, to evaluate the recovered MF, we calculate the absolute error between the yaw, pitch, and roll angles of the recovered frame and the MF. All metrics are averaged across scenes, except for the yaw, pitch, and roll errors for which the median is reported." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "content": "Implementation Details: We turn on " + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {ort }}" + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {centr }}" + }, + { + "bbox": [ + 304, + 133, + 545, + 216 + ], + "type": "text", + "content": " after 500 steps and linearly increase their weights to the specified values over the next 2500 steps. Also, we randomly sample rays for one-third of every batch size and select their left and upper neighbor to form a triplet to facilitate obtaining explicit surface normals. For other implementation details, please refer to the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 217, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 289 + ], + "type": "text", + "content": "Note that in addition to RGB images, ManhattanDF [14] requires knowing the wall and floor semantics, and the exact floor axis " + }, + { + "bbox": [ + 304, + 217, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathsf{n}_z" + }, + { + "bbox": [ + 304, + 217, + 545, + 289 + ], + "type": "text", + "content": ". In contrast, our method only requires RGB images and the assumed Manhattan prior to holding true. Therefore, this comparison aims to get a better insight into what can be achieved without leveraging additional labels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 299, + 368, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 368, + 310 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 368, + 310 + ], + "type": "text", + "content": "5.2. Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 318, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 628 + ], + "type": "text", + "content": "Hypersim [31] is a photorealistic synthetic dataset consisting of indoor scenes. It was created by leveraging a large repository of scenes created by professional artists, with 461 indoor scenes in total. This dataset is geometry-rich, containing a lot of details and lighting sources. Each scene has one or more camera trajectories available, where each trajectory has up to 100 views rendered in " + }, + { + "bbox": [ + 304, + 318, + 545, + 628 + ], + "type": "inline_equation", + "content": "1024 \\times 768" + }, + { + "bbox": [ + 304, + 318, + 545, + 628 + ], + "type": "text", + "content": ". For each scene, camera calibration information is provided, as well as detailed per-pixel labels such as depth and surface normals. After cleaning up the scenes by discarding a few with insufficient camera views, as well as other problems, we were left with 435 scenes. We also kept only one camera trajectory per scene. We then evaluated the InstantNGP [25] baseline on all 435 scenes and made three divisions based on its PSNR on unseen views. Hypersim-A contains 20 scenes where the baseline was performing well above average, Hypersim-B contains 20 scenes where the baseline had near-average performance, and Hypersim-C contains 10 scenes where the baseline had a below-average performance. For each scene, we randomly assigned half views to the training split and the rest to the test split. We note that the test split often contains views that were partially or completely unobserved during training. We also note that Hypersim is visually very realistic, geometry-rich, and challenging, considering it is synthetic. This can be subjectively observed by inspecting rendered views." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "ScanNet [8] is a real-world dataset consisting of indoor scenes. It was collected using a scalable RGB-D capture system that includes automated surface reconstruction and crowd-sourced semantic annotation. The dataset contains 1613 indoor scenes, which are annotated with ground-truth camera poses, surface reconstructions, and instance-level semantic segmentations. We use three scenes that were used" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17864" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 107, + 526, + 273 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 104 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 104 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 104 + ], + "type": "text", + "content": "Table 1: Experiments on Hypersim. We observe that our method consistently outperforms the baseline, as well as the ManhattanDF. This is very interesting since, unlike ManhattanDF, we do not use any additional labels during training. Finally, we see that it does not matter for our method whether the MF is known beforehand or not. Therefore, the additional knowledge of MF is neither necessary nor helpful." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 107, + 526, + 273 + ], + "lines": [ + { + "bbox": [ + 65, + 107, + 526, + 273 + ], + "spans": [ + { + "bbox": [ + 65, + 107, + 526, + 273 + ], + "type": "table", + "html": "
PSNR↑SSIM↑Normals°↓Pitch°↓Roll°↓Yaw°↓D-MAE↓D-RMSE↓
Scenes AInstantNGP [25] (baseline)25.860.87157.126.186.4619.250.0640.102
ManhattanDF [14]26.510.86840.691.231.015.080.0530.087
Ours27.200.86437.300.400.500.520.0530.093
Ours (MF known)27.210.85635.590.250.260.450.0520.091
Scenes BInstantNGP [25] (baseline)20.750.81160.126.067.9215.870.1050.151
ManhattanDF [14]21.870.82650.502.552.0611.690.0790.121
Ours22.450.81654.081.191.351.810.0800.127
Ours (MF known)22.510.81350.590.510.650.550.0780.126
Scenes CInstantNGP [25] (baseline)17.790.74064.297.454.5515.140.1300.174
ManhattanDF [14]18.330.77056.083.203.4110.250.1030.147
Ours19.430.76854.795.372.244.240.0940.133
Ours (MF known)19.290.76455.123.644.039.480.0940.135
194 scenesInstantNGP [25] (baseline)20.470.78361.346.566.9921.750.1040.146
ManhattanDF [14]20.940.79452.811.722.3213.480.0970.140
Ours21.630.78652.011.871.943.770.0850.126
", + "image_path": "fcb4dae230da0a5edc477ae9bc02dd022a4f47d8cac5743f2b4d9ad5d9826736.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 285, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 319 + ], + "type": "text", + "content": "in [14], where one-tenth of all views were uniformly sampled, leaving 303-477 views per scene. For each scene, the training and test split both contain half of the total views." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": "Replica [38] is a synthetic dataset featuring a diverse set of 18 indoor scenes. Each scene is equipped with photorealistic textures, allowing one to render realistic images from arbitrary camera poses. We use five scenes that were rendered and prepared in [58], where the rendered views also contain semantic segmentation labels. Each scene contains 900 views generated from random 6-DOF trajectories similar to hand-held camera motions in " + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 46, + 322, + 288, + 430 + ], + "type": "text", + "content": " resolution. We select 75 views for training and 75 for testing in each scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 442, + 105, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 105, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 105, + 453 + ], + "type": "text", + "content": "5.3. Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": "Experiments on Hypersim: We summarize the experiments performed on Hypersim in Table 1. Our proposed method achieves clear improvements in comparison to the InstantNGP baseline, in terms of novel-view rendering, normals estimation, depth estimation, as well as MF recovery. This is consistent across different scene difficulties, namely splits A, B, and C. Furthermore, our method also outperforms the ManhattanDF SotA. This is very interesting since, unlike ManhattanDF, we do not use any additional labels during training, other than the ground truth RGB. The ManhattanDF has partial access to ground truth MF axes, as well as wall and floor semantics, which directly implies having sparse surface normals ground truth. However, there are lots of Manhattan objects and areas in realistic scenes such as in Hypersim, other than the walls and floors. Our self-supervised method leverages the presence of many such objects and areas, to facilitate imposing geometrical constraints on the implicit representation during learning. This can be visually observed in Figure 3. Moreover, we observe that it does not matter for our method whether the MF is known beforehand, since our method recovers the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 285, + 425, + 354 + ], + "blocks": [ + { + "bbox": [ + 306, + 285, + 425, + 354 + ], + "lines": [ + { + "bbox": [ + 306, + 285, + 425, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 425, + 354 + ], + "type": "image", + "image_path": "931cc711d1421ab11c9a7fda1fcc78f8f7fed9e3591eae4d34bbccc38f485229.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 369, + 425, + 437 + ], + "blocks": [ + { + "bbox": [ + 317, + 358, + 416, + 368 + ], + "lines": [ + { + "bbox": [ + 317, + 358, + 416, + 368 + ], + "spans": [ + { + "bbox": [ + 317, + 358, + 416, + 368 + ], + "type": "text", + "content": "(a) InstantNGP [25] (baseline)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 369, + 425, + 437 + ], + "lines": [ + { + "bbox": [ + 307, + 369, + 425, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 369, + 425, + 437 + ], + "type": "image", + "image_path": "712b45e1e66298b516e9dde262be8b9d4f5749c2a89ceb19f419e817ffea2ed6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 442, + 380, + 452 + ], + "lines": [ + { + "bbox": [ + 352, + 442, + 380, + 452 + ], + "spans": [ + { + "bbox": [ + 352, + 442, + 380, + 452 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 456, + 545, + 533 + ], + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 533 + ], + "type": "text", + "content": "Figure 3: Leveraging the Manhattan prior (Hypersim-A). Our method leverages the presence of many Manhattan objects and areas in the scene in a self-supervised fashion, which facilitates imposing geometrical constraints during learning. Our method offers plausible normals – sometimes with missing details – that help to better model the radiance fields. In contrast, ManhattanDF leverages the Manhattan prior through labels of only walls and floors." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 427, + 285, + 545, + 354 + ], + "blocks": [ + { + "bbox": [ + 427, + 285, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 427, + 285, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 427, + 285, + 545, + 354 + ], + "type": "image", + "image_path": "16bba075f6b5187fecd171f0d629f1cdfc5791a03dab16f32f4405548dc9dd55.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 454, + 370, + 542, + 437 + ], + "blocks": [ + { + "bbox": [ + 451, + 358, + 523, + 369 + ], + "lines": [ + { + "bbox": [ + 451, + 358, + 523, + 369 + ], + "spans": [ + { + "bbox": [ + 451, + 358, + 523, + 369 + ], + "type": "text", + "content": "(b) ManhattanDF [14]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 370, + 542, + 437 + ], + "lines": [ + { + "bbox": [ + 454, + 370, + 542, + 437 + ], + "spans": [ + { + "bbox": [ + 454, + 370, + 542, + 437 + ], + "type": "image", + "image_path": "f3eb669f9bee957446c057f629488a204a95f5f16a38826bd2ec3710f28e4e2b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 441, + 509, + 451 + ], + "lines": [ + { + "bbox": [ + 459, + 441, + 509, + 451 + ], + "spans": [ + { + "bbox": [ + 459, + 441, + 509, + 451 + ], + "type": "text", + "content": "(d) Grond truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 689 + ], + "type": "text", + "content": "MF in the clustering step automatically, and therefore similar performance is achieved in both cases. This is particularly exciting because our self-supervised method performs similarly to the supervised one (using the groundtruth MF). Finally, we show large-scale experiments on a larger set of 194 scenes, which lead to the same conclusions. We observed that the semantic loss of the ManhattanDF sometimes causes convergence issues on difficult scenes, which we partially alleviated with class weighting and label smoothing. Therefore, we report results on only 194 scenes where ManhattanDF converged. Additional results on all 435 scenes are provided in the supplementary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Experiments on ScanNet: In Table 2, we examine the behavior of our proposed method on real-world indoor scenes." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17865" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 541, + 302 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 541, + 302 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 541, + 302 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 541, + 302 + ], + "type": "image", + "image_path": "7f03d4b048099205cde62a4252f74d5eed9eb1b2da95cea442227198e5b90389.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 303, + 547, + 351 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 547, + 351 + ], + "type": "text", + "content": "Figure 4: Qualitative results. Our method leverages many Manhattan objects and surfaces in the scene, which improves the implicit geometrical representation compared to the baseline. Unlike ManhattanDF, our method relies on many cues other than the walls & floors, which leads to a better representation of some objects (e.g. the white table in the bottom right corner of the example from columns 1-3). For more qualitative results, please refer to the supplementary material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 393, + 282, + 489 + ], + "blocks": [ + { + "bbox": [ + 47, + 357, + 288, + 391 + ], + "lines": [ + { + "bbox": [ + 47, + 357, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 288, + 391 + ], + "type": "text", + "content": "Table 2: Experiments on real-world ScanNet data. Our method outperforms both the baseline and ManhattanDF. This is the case both with and without supervising with sparse depth from SfM." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 393, + 282, + 489 + ], + "lines": [ + { + "bbox": [ + 51, + 393, + 282, + 489 + ], + "spans": [ + { + "bbox": [ + 51, + 393, + 282, + 489 + ], + "type": "table", + "html": "
PSNR↑SSIM ↑Depth↓-MAE
InstantNGP [25] (baseline)17.780.5870.119
RegNeRF [26]18.730.6180.102
ManhattanDF [14]18.680.6140.112
Ours20.790.6430.072
+ additional sparse depth from SfM
InstantNGP [25] (baseline)20.700.6310.048
ManhattanDF [14]21.530.6400.052
Ours22.250.6670.033
", + "image_path": "cf0341691a7b9bfa5729bb90d193db3e989a73b3a48a1b4b00acc03fba60a65c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 51, + 529, + 282, + 583 + ], + "blocks": [ + { + "bbox": [ + 47, + 492, + 287, + 526 + ], + "lines": [ + { + "bbox": [ + 47, + 492, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 287, + 526 + ], + "type": "text", + "content": "Table 3: Experiments on Replica. Our method outperforms the baseline, and it shows similar performance as ManhattanDF, which leverages additional labels during training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 529, + 282, + 583 + ], + "lines": [ + { + "bbox": [ + 51, + 529, + 282, + 583 + ], + "spans": [ + { + "bbox": [ + 51, + 529, + 282, + 583 + ], + "type": "table", + "html": "
PSNR↑SSIM ↑Depth↓-MAE
InstantNGP [25] (baseline)34.300.9440.022
Semantic-NeRF [58]34.080.9380.014
ManhattanDF [14]35.240.9440.008
Ours35.130.9440.011
", + "image_path": "8671ba17b52bee8000d7a534316c372aef89032d485420de16837ba776e79574.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "content": "Our method achieves clear improvements in comparison to the InstantNGP baseline, as well as to ManhattanDF, in terms of all measured metrics. Additionally, following the experimental setting of ManhattanDF [14], we train all methods with additional sparse depth supervision – where the sparse depth is obtained from the SfM pipeline. Our proposed method is also superior in this setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Experiments on Replica: We summarize experiments performed on the Replica dataset in Table 3. It can be observed that the InstantNGP baseline already performs very" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 365, + 533, + 449 + ], + "blocks": [ + { + "bbox": [ + 347, + 356, + 522, + 365 + ], + "lines": [ + { + "bbox": [ + 347, + 356, + 522, + 365 + ], + "spans": [ + { + "bbox": [ + 347, + 356, + 522, + 365 + ], + "type": "text", + "content": "Improvements of our method w.r.t. scene difficulty" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 365, + 533, + 449 + ], + "lines": [ + { + "bbox": [ + 316, + 365, + 533, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 533, + 449 + ], + "type": "image", + "image_path": "20b87505bcc4efa99a6f596cbe137a2db8ddb6c04dcb80aee8ec3af502bf34e7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 450, + 545, + 483 + ], + "lines": [ + { + "bbox": [ + 306, + 450, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 545, + 483 + ], + "type": "text", + "content": "Figure 5: Improvements vs. scene difficulty. Our method achieves the biggest improvements on scenes of hard and moderate difficulty, by leveraging many Manhattan objects and surfaces." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 593 + ], + "type": "text", + "content": "well on this dataset, leaving not much room for further improvements. Therefore, we consider Replica as an easier dataset. Nevertheless, our method still performs better than the baseline. As expected, ManhattanDF also improves the baseline similarly. Recall that the ManhattanDF uses additional labels for supervision. We also note that Replica has noticeably more walls and floors, and less of other objects, compared to other more complex datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Qualitative results: We depict the visual results of the discussed experiments in Figure 4. Our method leverages many Manhattan objects and surfaces in the scene, which improves the geometrical structure of 3D compared to the InstantNGP baseline. This is visible in surface normals and depth, obtained using volume rendering. Furthermore, we observe that, unlike ManhattanDF, our method relies on many Manhattan cues other than the walls & floors. This leads to a better representation of some such objects, e.g. the white table in the bottom right corner of the example" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17866" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 85, + 283, + 190 + ], + "blocks": [ + { + "bbox": [ + 47, + 62, + 286, + 84 + ], + "lines": [ + { + "bbox": [ + 47, + 62, + 286, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 62, + 286, + 84 + ], + "type": "text", + "content": "Table 4: Sparse training views. Our proposed method clearly exhibits the best performance, for all cases of input view sparsity." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 85, + 283, + 190 + ], + "lines": [ + { + "bbox": [ + 49, + 85, + 283, + 190 + ], + "spans": [ + { + "bbox": [ + 49, + 85, + 283, + 190 + ], + "type": "table", + "html": "
PSNR↑SSIM ↑Depth↓-MAE
12 viewsInstantNGP [25]18.020.7060.138
ManhattanDF [14]19.450.7500.116
Ours20.500.7600.104
9 viewsInstantNGP [25]16.790.6610.154
ManhattanDF [14]18.040.7140.130
Ours19.140.7280.120
6 viewsInstantNGP [25]15.750.5820.178
ManhattanDF [14]16.000.6390.159
Ours16.670.6670.158
", + "image_path": "0ce2ab559bfd8f9e784ac719d77cb8e1839ba8b7bbf8dc1a873ce7cec4c347ba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 219, + 286, + 307 + ], + "blocks": [ + { + "bbox": [ + 47, + 197, + 286, + 217 + ], + "lines": [ + { + "bbox": [ + 47, + 197, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 197, + 286, + 217 + ], + "type": "text", + "content": "Table 5: Ablation study on Hypersim-A. Both of our proposed loss terms contribute to the overall performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 219, + 286, + 307 + ], + "lines": [ + { + "bbox": [ + 48, + 219, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 286, + 307 + ], + "type": "table", + "html": "
PSNR↑Norm.°↓Depth↓Rot.°↓
Only Limg25.8657.120.06410.63
+ Lort27.2150.070.0585.39
+ Lctr27.0636.090.0530.57
+ Lort + Lctr (Ours)27.2037.300.0530.47
Ours + MF known27.2135.590.0520.32
Ours (no delay)27.0635.780.0540.39
Ours (no w lin.)27.0137.860.0550.60
", + "image_path": "d0b92525a1588d91f51e30861c5a3f75a5f297b137937512c854f573d46ddd65.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 318, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 365 + ], + "type": "text", + "content": "from columns 1-3. Moreover, we observe that our method is able to cope with difficult scenes and views, where other methods struggle. For more qualitative results, please refer to the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 366, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 449 + ], + "type": "text", + "content": "Improvements with respect to scene difficulty: We analyze the improvements by our method for different scene difficulties. We decided on scene difficulty based on the novel-view rendering performance of the InstantNGP baseline. In Figure 5, we see that our method brings the most benefits for scenes of hard and moderate difficulties, thanks to the Manhattan scene prior." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 449, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 521 + ], + "type": "text", + "content": "Sparse training views: In order to gain more insight, we examine the behavior of training neural radiance fields with sparse training input views. The results on the Hypersim-A dataset are presented in Table 4. Our proposed method clearly outperforms the InstantNGP baseline, as well as ManhattanDF, when trained with 6, 9, and 12 input views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "text", + "content": "Finding the MF: We test the robustness of our proposed method for finding the Manhattan frame, by introducing a simultaneous offset " + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "text", + "content": " in the yaw, pitch, and roll on the canonical MF. The experiments are performed on Hypersim A and can be found in Figure 6. In Figure 6a we observe that the novel-view rendering quality remains largely robust to the rotation offset " + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "text", + "content": ". We note that we increase the scene bounding box by the same factor for all experiments in Figure 6, to make sure that objects remain within the voxel grid for maximal " + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 521, + 287, + 676 + ], + "type": "text", + "content": ". This slightly decreases the resolution of grid element representations, so PSNR is slightly lower than in Table 1. Furthermore, in Figure 6b we see that the MF estimation remains robust to the rotation offset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Ablation study: We report our ablation study in Table 5. Both of our proposed losses contribute to the overall performance. Furthermore, turning on " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{ort}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{centr}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " after" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 332, + 80, + 518, + 141 + ], + "blocks": [ + { + "bbox": [ + 365, + 72, + 501, + 80 + ], + "lines": [ + { + "bbox": [ + 365, + 72, + 501, + 80 + ], + "spans": [ + { + "bbox": [ + 365, + 72, + 501, + 80 + ], + "type": "text", + "content": "Rotation offset effects on novel-view rendering" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 80, + 518, + 141 + ], + "lines": [ + { + "bbox": [ + 332, + 80, + 518, + 141 + ], + "spans": [ + { + "bbox": [ + 332, + 80, + 518, + 141 + ], + "type": "image", + "image_path": "c81a6c5747db5b77d5cf4c8f7a5a0a2d9c326fdac240db57c8ed609ff0fc073c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 141, + 541, + 150 + ], + "lines": [ + { + "bbox": [ + 310, + 141, + 541, + 150 + ], + "spans": [ + { + "bbox": [ + 310, + 141, + 541, + 150 + ], + "type": "text", + "content": "(a) Novel-view rendering quality remains robust to the rotation offset " + }, + { + "bbox": [ + 310, + 141, + 541, + 150 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 310, + 141, + 541, + 150 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 332, + 160, + 517, + 221 + ], + "blocks": [ + { + "bbox": [ + 350, + 152, + 513, + 160 + ], + "lines": [ + { + "bbox": [ + 350, + 152, + 513, + 160 + ], + "spans": [ + { + "bbox": [ + 350, + 152, + 513, + 160 + ], + "type": "text", + "content": "Rotation offset effects on Manhattan frame estimation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 160, + 517, + 221 + ], + "lines": [ + { + "bbox": [ + 332, + 160, + 517, + 221 + ], + "spans": [ + { + "bbox": [ + 332, + 160, + 517, + 221 + ], + "type": "image", + "image_path": "aa753d85a2f79945af025e76515d98191ad5919d49b0f2de87582429492c77a2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 222, + 517, + 230 + ], + "lines": [ + { + "bbox": [ + 333, + 222, + 517, + 230 + ], + "spans": [ + { + "bbox": [ + 333, + 222, + 517, + 230 + ], + "type": "text", + "content": "(b) MF estimation remains robust to the rotation offset " + }, + { + "bbox": [ + 333, + 222, + 517, + 230 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 236, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 306, + 236, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 545, + 258 + ], + "type": "text", + "content": "Figure 6: Finding the MF. We test the robustness of our method by introducing the rotation offsets on the canonical MF." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 270, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 545, + 294 + ], + "type": "text", + "content": "500 steps, and linearly increasing their weight also helps slightly. Please refer to the supplementary for more details." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 304, + 378, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 378, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 378, + 316 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 324, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 562 + ], + "type": "text", + "content": "We demonstrated the possibility of exploiting the Manhattan scene prior without needing any additional supervision. This is achieved by performing robust clustering of explicit normals, followed by the search of the Manhattan frame (MF) whose existence is based on the assumed prior. The sought MF is obtained from the orthogonal clusters, which are later used to self-supervise the neural representation learning. Our self-supervision encourages the normals of Manhattan surfaces to group into three orthogonal directions. Our experiments on three indoor datasets demonstrate that the proposed method not only benefits from building parts (such as walls and floors) but also exploits many other Manhattan scene parts (such as furniture). Both quantitative and qualitative evaluations reveal the benefit of the proposed method in terms of, improved performance over the established baselines and competitive results to state-of-the-art methods that use additional labels for supervision. Our method has the potential to be extended in other higher expressiveness scene priors, such as Atlanta world and the mixture of Manhattan frames." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 647 + ], + "type": "text", + "content": "Limitations: One limitation of our method is that it sometimes produces surface normals with missing details or \"blocky\" artifacts. Nevertheless, this usually offers better novel-view RGB rendering, compared to not imposing any structure priors. Another limitation is that our method is not beneficial for very easy scenes. For a detailed discussion of limitations, please refer to the supplementary material." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 658, + 403, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 403, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 403, + 670 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "This research was co-financed by Innosuisse under the project Know Where To Look, Grant 59189.1 IP-ICT. Also, Carlos Oliveira helped with splendid visual renderings." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17867" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290-6301, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 169 + ], + "type": "text", + "content": "[2] Stephen T Barnard. Interpreting perspective images. Artificial intelligence, 21(4):435-462, 1983." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 226 + ], + "type": "text", + "content": "[3] Jean-Charles Bazin, Hongdong Li, In So Kweon, Cedric Demonceaux, Pascal Vasseur, and Katsushi Ikeuchi. A branch-and-bound approach to correspondence and grouping problems. IEEE transactions on pattern analysis and machine intelligence, 35(7):1565-1576, 2012." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 293 + ], + "type": "text", + "content": "[4] Jean-Charles Bazin, Yongduek Seo, Cédric Demonceaux, Pascal Vasseur, Katsushi Ikeuchi, Inso Kweon, and Marc Pollefeys. Globally optimal line clustering and vanishing point estimation in manhattan world. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 638–645. IEEE, 2012." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 288, + 328 + ], + "type": "text", + "content": "[5] Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag, Berlin, Heidelberg, 2006." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 330, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 288, + 396 + ], + "type": "text", + "content": "[6] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 288, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 451 + ], + "type": "text", + "content": "[7] James M Coughlan and Alan L Yuille. Manhattan world: Compass direction from a single image by bayesian inference. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 941-947. IEEE, 1999." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 453, + 288, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 288, + 507 + ], + "type": "text", + "content": "[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 288, + 564 + ], + "type": "text", + "content": "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[10] Patrick Denis, James H Elder, and Francisco J Estrada. Efficient edge-based methods for estimating Manhattan frames in urban imagery. In European conference on computer vision, pages 197-210. Springer, 2008." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 633 + ], + "type": "text", + "content": "[11] Clara Fernandez-Labrador. Indoor scene understanding using non-conventional cameras. PhD thesis, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[12] Yasutaka Furukawa, Brian Curless, Steven M Seitz, and Richard Szeliski. Manhattan-world stereo. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1422-1429. IEEE, 2009." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[13] Wuwei Ge, Yu Song, Baichao Zhang, and Zehua Dong. Globally optimal and efficient Manhattan frame estimation by delimiting rotation search space. In Proceedings of the" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 15213-15221, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[14] Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5511-5520, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 239 + ], + "type": "text", + "content": "[16] Kyungdon Joo, Tae-Hyun Oh, Junsik Kim, and In So Kweon. Globally optimal Manhattan frame estimation in real-time. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1763-1771, 2016." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 304 + ], + "type": "text", + "content": "[17] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "text", + "content": "[18] Fu Li, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerf-posed: A first-reconstruct-thenregress approach for weakly-supervised 6d object pose estimation. arXiv preprint arXiv:2203.04802, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[19] Haoang Li, Jian Yao, Jean-Charles Bazin, Xiaohu Lu, Yazhou Xing, and Kang Liu. A monocular slam system leveraging structural regularity in manhattan world. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 2518-2525. IEEE, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "text", + "content": "[20] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "text", + "content": "[21] Zhi-Hao Lin, Wei-Chiu Ma, Hao-Yu Hsu, Yu-Chiang Frank Wang, and Shenlong Wang. Neurmips: Neural mixture of planar experts for view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15702-15712, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 569 + ], + "type": "text", + "content": "[22] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 570, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 603 + ], + "type": "text", + "content": "[23] Gerard F McLean and D Kotturi. Vanishing point detection by line clustering. IEEE Transactions on pattern analysis and machine intelligence, 17(11):1090-1095, 1995." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 604, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 669 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. cite arxiv:2003.08934Comment: ECCV 2020 (oral). Project page with videos and code: http://tancik.com/nerf." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17868" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[26] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 139, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 139, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 139, + 288, + 183 + ], + "type": "text", + "content": "[27] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 184, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 184, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 184, + 288, + 239 + ], + "type": "text", + "content": "[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 288, + 282 + ], + "type": "text", + "content": "[29] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in manhattan world. In Proceedings of the IEEE International Conference on Computer Vision, pages 882-890, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 284, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 284, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 288, + 337 + ], + "type": "text", + "content": "[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 338, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 288, + 403 + ], + "type": "text", + "content": "[31] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 404, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 49, + 404, + 288, + 459 + ], + "type": "text", + "content": "[32] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12892-12901, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 460, + 288, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 288, + 492 + ], + "type": "text", + "content": "[33] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 493, + 288, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 493, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 49, + 493, + 288, + 569 + ], + "type": "text", + "content": "[34] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lučić, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6229-6238, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 570, + 288, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 570, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 49, + 570, + 288, + 603 + ], + "type": "text", + "content": "[35] Gautam Singh and Jana Kosecka. Visual loop closing using gist descriptors in manhattan world. In ICRA omnidirectional vision workshop, pages 4042-4047, 2010." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 604, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 604, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 49, + 604, + 288, + 658 + ], + "type": "text", + "content": "[36] Julian Straub, Nishchal Bhandari, John J Leonard, and John W Fisher. Real-time manhattan world rotation estimation in 3d. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1913-1920. IEEE, 2015." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 288, + 712 + ], + "type": "text", + "content": "[37] Julian Straub, Oren Freifeld, Guy Rosman, John J Leonard, and John W Fisher. The Manhattan frame model—manhattan world inference in the space of surface normals. IEEE transactions on pattern analysis and machine intelligence, 40(1):235-249, 2017." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 183 + ], + "type": "text", + "content": "[38] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 228 + ], + "type": "text", + "content": "[39] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 229, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 545, + 274 + ], + "type": "text", + "content": "[40] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light field neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8269-8279, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 274, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 329 + ], + "type": "text", + "content": "[41] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "type": "text", + "content": "[42] Frank A Van den Heuvel. 3d reconstruction from a single image using geometric constraints. ISPRS Journal of Photogrammetry and Remote Sensing, 53(6):354-368, 1998." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[43] Carlos A Vanegas, Daniel G Aliaga, and Bedrich Benes. Building reconstruction using manhattan-world grammars. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 358-365. IEEE, 2010." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 410, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 465 + ], + "type": "text", + "content": "[44] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 466, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 521 + ], + "type": "text", + "content": "[45] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. Nesf: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 522, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 545, + 567 + ], + "type": "text", + "content": "[46] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. arXiv preprint arXiv:2206.13597, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 622 + ], + "type": "text", + "content": "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "text", + "content": "[48] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[49] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17869" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 508 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 116 + ], + "type": "text", + "content": "[50] Horst Wildenauer and Allan Hanbury. Robust camera self-calibration from monocular images of manhattan worlds. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2831-2838. IEEE, 2012." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 173 + ], + "type": "text", + "content": "[51] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 174, + 286, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 174, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 174, + 286, + 218 + ], + "type": "text", + "content": "[52] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K Wong. Ps-nerf: Neural inverse rendering for multi-view photometric stereo. In European Conference on Computer Vision, pages 266–284. Springer, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 286, + 262 + ], + "type": "text", + "content": "[53] Wenqi Yang, Guanying Chen, Chaofeng Chen, Zhenfang Chen, and Kwan-Yee K Wong. S3-nerf: Neural reflectance field from shading and shadow under a single viewpoint. arXiv preprint arXiv:2210.08936, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 286, + 296 + ], + "type": "text", + "content": "[54] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 297, + 286, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 297, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 49, + 297, + 286, + 351 + ], + "type": "text", + "content": "[55] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 353, + 286, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 353, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 49, + 353, + 286, + 396 + ], + "type": "text", + "content": "[56] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 397, + 286, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 397, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 49, + 397, + 286, + 452 + ], + "type": "text", + "content": "[57] Menghua Zhai, Scott Workman, and Nathan Jacobs. Detecting vanishing points using global image context in a non-manhattan world. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5657-5665, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 453, + 286, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 453, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 49, + 453, + 286, + 508 + ], + "type": "text", + "content": "[58] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17870" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/e47c0301-5f63-4201-a2ee-b44e21fee063_content_list.json b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/e47c0301-5f63-4201-a2ee-b44e21fee063_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c6ada0d6e4ee43c410ab6eab89f8794cd7da3390 --- /dev/null +++ b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/e47c0301-5f63-4201-a2ee-b44e21fee063_content_list.json @@ -0,0 +1,1509 @@ +[ + { + "type": "text", + "text": "SURFSUP: Learning Fluid Simulation for Novel Surfaces", + "text_level": 1, + "bbox": [ + 196, + 130, + 774, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arjun Mani\\*, Ishaan Preetam Chandratreya\\*, Elliot Creager\\*3, Carl Vondrick1, Richard Zemel1 \n1 Columbia University 2 University of Toronto 3 Vector Institute \nsurfsup.cs.columbia.edu", + "bbox": [ + 102, + 186, + 866, + 243 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39ac34228a60edba9ac06071898a4fe9d2f9da8e50af07105250e807be33b435.jpg", + "image_caption": [ + "Figure 1: We introduce SURFSUP, which learns to predict how fluid particles interact with novel, complex 3D surfaces." + ], + "image_footnote": [], + "bbox": [ + 122, + 268, + 849, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 614, + 313, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modeling the mechanics of fluid in complex scenes is vital to applications in design, graphics, and robotics. Learning-based methods provide fast and differentiable fluid simulators, however most prior work is unable to accurately model how fluids interact with genuinely novel surfaces not seen during training. We introduce SURFSUP, a framework that represents objects implicitly using signed distance functions (SDFs), rather than an explicit representation of meshes or particles. This continuous representation of geometry enables more accurate simulation of fluid-object interactions over long time periods while simultaneously making computation more efficient. Moreover, SURFSUP trained on simple shape primitives generalizes considerably out-of-distribution, even to complex real-world scenes and objects. Finally, we show we can invert our model to design simple objects to manipulate fluid flow.", + "bbox": [ + 73, + 637, + 473, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 614, + 632, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Across engineering and science, the simulation of fluid dynamics has become an invaluable tool, and it will be a crucial component for building visual systems that are capable of understanding and interacting with real-world environments. Recently, a new class of simulators has emerged that learn the dynamics of physical systems from data [13, 32, 25]. These simulators offer several advantages over classical simulators by increasing speed, reducing reliance on hand-crafted dynamics models, and providing differentiable rollouts for solving inverse problems. Recent approaches based on graph neural networks (GNNs) have in particular shown impressive accuracy and generalization on a wide range of fluids represented as particle or mesh-based systems [25, 23].", + "bbox": [ + 496, + 641, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In a wide variety of applications areas, fluid simulation must properly handle interaction with solid structures to attain practical relevance. Many problems involving", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally.", + "bbox": [ + 93, + 886, + 254, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14225", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fluid simulation are fundamentally about object design; for example, designing aerodynamic shapes for cars and airplanes. In robotics too, an intelligent robot would have to understand fluid-surface interactions; for example, to carry a mug of hot coffee to a person without spilling the liquid from the container. In both cases, the need to accurately model how fluids interact with novel solid surfaces1—with new shapes, configurations, orientations, and so on—motivates an challenging out-of-distribution (OOD) generalization subproblem for fluid simulation.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we tackle this problem by introducing Surface implicit sUbstitution for particles (SURFSUP), a framework for simulating the interactions of fluids and surfaces by representing surfaces using implicit representations [19]. Current GNN-based approaches represent surfaces explicitly, discretizing them into particles and augmenting the graph accordingly [25]. Implicit representations, in particular signed distance functions which represent surfaces as zero-level sets of functions, offer several compelling advantages. They can smoothly represent continuous surfaces, provide rich geometric information (e.g. whether a point is in the interior/exterior of the object), and can scale easily to large objects and scenes. Due to these properties, many state-of-the-art generative models for 3D shapes learn shapes with signed distance functions. We argue that these properties are useful not just for statics but also dynamics: in order to accurately model fluid-surface interactions for potentially complex surfaces, smooth and informative representations of local surface geometry are needed. In turn, we posit that these locally informative representations will enable fluid simulators to appropriately model interaction with never-before-seen surfaces.", + "bbox": [ + 75, + 242, + 472, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concretely, we propose an approach that integrates signed distance representations of surfaces [22, 19] with graph neural network models [26, 25]. Our architecture follows the encode-process-decode paradigm [25], where the particles in the physical system at time $T$ are encoded into a graph, followed by several message-passing steps, and finally a decoder predicts the dynamics (accelerations). We represent the surface using its signed distance function, and exploit geometric properties of SDFs to model the distance and orientation of fluid particles with respect to the surface.", + "bbox": [ + 75, + 575, + 472, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments show that our approach can model fluid-surface interactions over long rollouts with a high degree of accuracy and visual realism. SURFSUP can model surfaces far beyond its training domain; when trained on simple primitive shapes (e.g. spheres, cones, toruses), we generalize zero-shot to complex real-world objects and scenes represented using neural implicit generative models (Fig. 1). We compare to a method which represents the surface as particles inside the graph network simulator, and show that our model achieves better generalization and efficiency, and", + "bbox": [ + 75, + 726, + 472, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can better handle complex surface geometry. Finally, we demonstrate the potential of SURFSUP to solve inverse problems across two object design tasks. Overall, our paper bridges learned physical models for dynamics and 3D implicit representations, and takes a step towards AI models which can understand and help design the physical world.", + "bbox": [ + 496, + 90, + 893, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 195, + 640, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning Differentiable Simulators. Using machine learning for predicting and modeling complex physical systems is a rapidly growing area of research. Learned physics simulators have been shown to accurately simulate colliding rigid objects, deformable objects, fluids, and other complex systems [13, 3, 12, 32, 25, 21].", + "bbox": [ + 496, + 220, + 893, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent progress in learned fluid simulation has been driven by the Lagrangian viewpoint, which tracks the motion of fluid particles over time. Graph neural networks (GNNs) have in particular emerged as effective forward models for fluid simulation, and have demonstrated better stability and generalization compared to grid-based, convolutional approaches. [13, 14, 27, 25]. Sanchez-Gonzalez et al. [25] introduced a GNN architecture which can accurately and generalizably model fluid behavior over long rollouts; most subsequent work builds off their approach [23, 16]. However, most of these works generally consider simple surfaces; e.g. only cube-shaped containers for fluids. There has been little work focused on fluid-surface interactions; Mayr et al. [16] introduce an approach specifically for triangularized meshes. All these approaches use particles or meshes to represent surfaces; we will argue that implicit representations enable learning more accurate, generalizable physical dynamics.", + "bbox": [ + 496, + 311, + 893, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some works have explored the possibility of using learned fluid simulators for solving inverse problems [27, 12, 2]. Allen et al. [2] showed that graph network simulators can be used to solve design tasks by rolling out fluid trajectories and using gradient-based optimization. In 3D, they show that a 2D plane with a learned height field can be used to split a stream of fluid to fall into several locations on the ground. SDFs have the advantage of being able to learn geometry in a more smooth and unconstrained way, which could help solve inverse problems.", + "bbox": [ + 496, + 584, + 893, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Implicit Representations. Implicit representations have become state-of-the-art for generative models of 3D objects [17, 22, 6, 29, 30]. These models represent the surface as a function $F(x)$ . There are several choices for $F$ . Park et al. [22] show that signed distance functions $F(x)$ are particularly effective and informative representations. Compared to point clouds [1, 28] or voxels [7, 34], models based on SDF representations can smoothly represent complex geometry, are compact and efficient, and can be learned effectively [29, 6]. Follow-up work has studied learning \"dual\" generative models that jointly learn an", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "For brevity we refer to surfaces of solid objects simply as \"surfaces\".", + "bbox": [ + 94, + 886, + 465, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14226", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "uncontrained representation and one composed of primitive shapes [10, 33].", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit Representations for Physical Dynamics. Some classical particle-based fluid simulators based on the Smoothed Particle Hydrodynamics (SPH) formalism [20] have transitioned from particle-based to implicit surface representations. [11, 9, 4]. In these simulators implicit surfaces allow for smoother fluid motion with less artifacts [11]. For general-purpose learned simulators (GNNs), there is little prior work using implicit rather than particle-based representations of surfaces.", + "bbox": [ + 75, + 121, + 467, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another line of work combines implicit generative models with physical simulators [18, 24]. Mezghanni et al. [18] generate more physically stable SDF shapes by backpropagating gradients through a physical simulation layer. Their simulation mainly considers Newtonian dynamics. Our contribution is an effective, generalizable forward model for fluids that represents surfaces accurately and efficiently, and can also be used for solving inverse problems.", + "bbox": [ + 75, + 257, + 467, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 393, + 166, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Setup: Particle-Based Fluid Simulation", + "text_level": 1, + "bbox": [ + 76, + 417, + 419, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We address the problem of learning the physical dynamics of fluids interacting with surfaces. Given the state of a physical system $X^0$ and some initial conditions (e.g. gravity, other external forces), rolling out a fluid simulation over $T$ steps produces a sequence of states $X^{(1)},\\dots,X^{(T)}$ . In this paper, we focus on particle-based simulation; the state of the system at time $t$ is represented by a set of particles $X^{(t)} = (p_1^{(t)},\\dots,p_n^{(t)})$ , where $p_i$ is the position of particle $i$ in 3D space. A rollout over $T$ steps is produced by repeatedly applying a forward model $S_{\\theta}:X^{(t)}\\to X^{(t + 1)}$ , which takes in the particle positions at time $t$ (and optionally a short history $t - 1,\\ldots$ ) and produces the next-step particle positions. While the forward model can be used to directly predict $X^{(t + 1)}$ , most methods instead predict intermediate values $Y$ , typically next-state accelerations for each particle, which can then be numerically integrated (e.g. Euler integration) to produce $X^{(t + 1)}$ ; concretely, our task is to learn the parameters $\\theta$ of this forward model $f_{\\theta}:X^{(t)}\\to Y$ .", + "bbox": [ + 75, + 441, + 467, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Graph Networks for Particle-Based Simulation", + "text_level": 1, + "bbox": [ + 76, + 726, + 467, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Graph neural networks lend themselves naturally to particle-based simulation, since they compute pairwise interactions between nodes and aggregate these interactions. Graph Network-based Simulators (GNS), a model architecture introduced by Sanchez-Gonzalez et al. [25], have in particular shown state-of-the-art performance on simulating fluids over long rollouts. These methods follow an \"encode-process-decode\" paradigm. First, the state $X^{(t)}$ is encoded into a graph, where nodes $v_{i}$ are instantiated using a set of features (e.g. previous velocities, parti", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cle type, etc.); nodes $v_{i}, v_{j}$ are then connected with an edge $e_{ij}$ if their distance is less than a connectivity radius $\\epsilon$ . Each node and edge is associated with an embedding. The \"processor\" stage then passes the encoded graph through $M$ message-passing steps. Each message-passing step takes the current graph $\\{(v_{i}, v_{j}, e_{ij})\\}^{(m)}$ . An edge module first computes the updated edge embeddings $e_{ij}^{(m+1)} = f_{e}(v_{i}^{(m)}, v_{j}^{(m)}, e_{ij}^{(m)})$ ; then the edge embeddings are aggregated to compute updated node embeddings $v_{i}^{(m+1)} = f_{n}(v_{i}^{(m)}, \\sum_{j \\in N(i)} e_{ij}^{(m)})$ , where $N(i)$ are the set of neighbors of node $i$ . $f_{e}$ and $f_{n}$ are generally MLPs with 2-3 hidden layers, and the weights are shared across edges/nodes (although differing across message passing steps). Finally, the \"decoder\" step applies an MLP $f_{D}$ to the final node embeddings to predict the accelerations $\\{y_{i}\\}$ .", + "bbox": [ + 496, + 90, + 890, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our problem, we further assume that there exists some rigid surface $S$ that interacts with the fluid throughout the simulation. The GNS approach handles surfaces by discretizing the surface into particles (i.e. a point cloud) and adding these particles $\\{s_i\\}$ to the graph (denoting them by a different particle type, and masking during position updates) [25]. Follow-up work for mesh-based simulation handles surfaces similarly by representing the surface as an irregular mesh and adding nodes for mesh vertices [23].", + "bbox": [ + 496, + 321, + 890, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Modeling Fluid-Surface Interactions with SDFs", + "text_level": 1, + "bbox": [ + 500, + 465, + 890, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose to model the interactions of fluids and surfaces with signed distance function (SDF) representations of surfaces. An SDF represents a watertight surface as a function $F(p) = s$ , which takes a spatial point $p \\in \\mathbb{R}^3$ and outputs its distance $s$ to the closest corresponding point on", + "bbox": [ + 496, + 489, + 890, + 566 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a76562248cb915f5d8fc7915cb5b0f084083b8fe295f0466c397ec3f7aa34009.jpg", + "image_caption": [ + "Figure 2: Rigid surfaces can be represented compactly and implicitly as the level set of a SDF: $\\{x|F(x) = 0\\}$ . Evaluating $F$ and its derivatives at particle position $p$ provides useful information for simulating the interaction of that particle with the surface." + ], + "image_footnote": [], + "bbox": [ + 573, + 585, + 820, + 799 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14227", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the surface. The function is \"signed\", such that it takes on positive values for points outside the surface and negative values in the surface interior. The surface of the object is implicitly represented as the zero-level set $F(p) = 0$ . Explicit representations of the surface can be constructed by applying meshing algorithms such as Marching Cubes [15].", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Almost every surface can be described by an SDF. For several primitive shapes SDFs can be written down analytically (e.g. spheres, cones, cylinders); boolean operations on these shapes such as unions and intersections give rise to more complex shapes [31]. SDFs can also be represented discretely as voxel grids; or more recently for finer geometry, as neural SDFs $F_{\\theta}(p)$ which learn to accurately approximate $F$ from mesh-based supervision.", + "bbox": [ + 75, + 183, + 467, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Locality and Geometric Information for Fluid-Surface Interactions", + "text_level": 1, + "bbox": [ + 76, + 325, + 467, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A key insight of our method is that fluid-surface interactions are inherently about local rather than global surface geometry. The interaction of a fluid with a surface depends on a small local region of the surface, over which the geometry can have limited variation. This suggests that an effective representation of local surface geometry, trained on a variety of object shapes, can generalize OOD to shapes that may look different globally. By contrast, particles are not an effective representation of surface geometry; for any moderately complex surface, discretization leads to artifacts such as non-smooth fluid motion and even penetration of the surface boundary (we confirm this empirically).", + "bbox": [ + 75, + 366, + 467, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By contrast, we argue that SDFs are a smooth, richly informative representation of local surface geometry for learning generalizable fluid-surface interactions. Our key insight is shown in Figure 2. Given a solid surface represented as an SDF $F(p)$ , we can use properties of the SDF to inform the behavior of a fluid particle as it reaches the surface. The value of the SDF $F(p)$ gives the distance from the particle at point $p$ to the closest point on the surface $p'$ . Moreover, the SDF gradient $\\nabla F(p)$ provides the unit-norm displacement vector from the particle to its closest surface point $\\frac{(p - p')}{||p - p'||}$ . This can be shown by observing that the SDF decreases most quickly on the line from $p$ to $p'$ . Thus, the SDF and its gradient provide the distance and orientation of the fluid particle w.r.t. to its closest surface point; as the particle approaches the surface, its acceleration will more closely approach $\\nabla F(p)$ and it will tend to rebound. Moreover, since the gradient varies smoothly because of the continuity of $F$ , these features do not suffer from artifacts.", + "bbox": [ + 75, + 547, + 467, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Higher order derivatives of the SDF can also be used to gain geometric information. For example, we could march to the surface by computing $p - F(x)\\nabla F(x)$ . At the surface, the Hessian $\\nabla^2 F(x)$ is the shape operator; its eigenvalues provide the principal directions of surface curvature", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6187da8e669f4f3de12fe83663c39f6f7f68d55627cc19ade6c9b04a1092a8c9.jpg", + "image_caption": [ + "Figure 3: Given fluid particles and the SDF, our method augments the constructed graph with features from the SDF. We then apply message-passing and predict dynamics, weighting near-surface errors more highly in our loss." + ], + "image_footnote": [], + "bbox": [ + 506, + 95, + 893, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\kappa_{1}$ and $\\kappa_{2}$ , and the mean curvature is $H = \\frac{1}{2} \\mathrm{Tr}(\\nabla^2 F)$ . We can also estimate the surface curvature via finite differences by slightly perturbing the gradient vector on each axis and computing the SDF (Fig. 2). In this paper, we only use the SDF and its gradient as features without curvature information; this can be seen as an effective theory where the relevant \"local region\" of the surface is a single closest point on the surface. As we show, modeling surfaces using SDF representations helps our fluid simulator generalize beyond the available training data by modeling fluid-surface interactions for unseen surfaces.", + "bbox": [ + 498, + 334, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Model", + "text_level": 1, + "bbox": [ + 500, + 532, + 584, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Having developed our ideas about SDFs and modeling fluid-surface interactions, we now describe our approach, which we call Surface implicit sUbstitution for particles (SURFSUP). We learn a forward predictor $f_{\\theta}:X^{(t)},F\\to Y$ , which takes as input a set of particle positions and a surface implicitly parametrized by an SDF $F$ . We adopt the encode-process-decode paradigm for our model, but make two key changes. In the encoder step, for each fluid particle $p_i$ we add the value of the SDF $F(p_i)$ and its gradient $\\nabla F(p_i)$ to the initial node representations, such that the new representation is $v_{i}^{\\prime (0)} = [v_{i}^{(0)},F(p_{i}),\\nabla F(p_{i})]$ . This is the way that the surface is encoded into the model; subsequently no particles representing the surface are added to the graph. The updated node embeddings are then passed through the \"processor\" with several message-passing steps and then decoded to predict the accelerations as standard. The other modification we make is to weight particles close to the surface higher during training when predicting dynamics. This has the effect of teaching the model to prioritize modeling fluid-surface interactions accurately. Since the model is predicting next-step accelerations, our loss becomes:", + "bbox": [ + 498, + 563, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14228", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = \\sum_ {F (p _ {i}) > \\alpha} \\left\\| a _ {i} ^ {(t)} - \\hat {a} _ {i} ^ {(t)} \\right\\| ^ {2} + \\sum_ {F (p _ {i}) \\leq \\alpha} \\lambda \\left\\| a _ {i} ^ {(t)} - \\hat {a} _ {i} ^ {(t)} \\right\\| ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 114, + 468, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we set $\\alpha$ based on the \"neighborhood radius\" of the classical simulator we train against, and set $\\lambda = 5$ . We found that this weighting helped the model learn surface interactions more quickly without sacrificing overall accuracy on fluid prediction. Note that the SDF allows loss weighting to be implemented in a very natural way, by simple thresholding. Our approach is shown in Fig. 3.", + "bbox": [ + 76, + 161, + 468, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Other Advantages of SDF Representations", + "text_level": 1, + "bbox": [ + 76, + 276, + 439, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have argued for SDFs from the perspective of achieving smooth, information-rich representations of local surface geometry. Here we mention other advantages of SDFs, each of which we demonstrate in our experimental results.", + "bbox": [ + 76, + 300, + 468, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficient message passing in the GNN. By modeling surfaces implicitly, we reduce the number of unique particles needed to compose the GNN's state. This allows us to perform more lightweight and computationally efficient simulation, as well as run the model for more message-passing steps. This advantage becomes especially clear for large objects/scenes or objects with complex surface geometry.", + "bbox": [ + 76, + 364, + 468, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shape Design with SDFs. SDFs are a useful paradigm for shape design and assembly. This is because compositions of shapes can be generated by union, intersection, and difference operations; e.g. the union of two SDFs $F_{1}$ and $F_{2}$ is $\\min(F_{1}, F_{2})$ . The construction of shapes using primitives and these operations is known as Constructive Solid Geometry (CSG), and is implemented in many CAD packages for object design [8]. Moreover the recent emergence of neural SDF models as state-of-the-art 3D shape representations, offers another compelling avenue for shape design with end-to-end learning. These suggest that good forward models with SDFs could be leveraged for design problems.", + "bbox": [ + 76, + 474, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 670, + 215, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Fluid Domains", + "text_level": 1, + "bbox": [ + 76, + 695, + 225, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our approach on complex prediction tasks involving fluids interacting with a variety of surfaces. Our main contribution here is a 3D fluid dataset consisting of water interacting with a range of primitive shapes, across shape parameters, positions, and orientations. For the ground-truth simulator we use SPlisHSPlasH [5], a 3D-particle-based fluid simulator based on the SPH technique.", + "bbox": [ + 76, + 719, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PrimShapes Dataset. We train SURFSUP on a 3D water dataset, where the fluid interacts with various shapes. We initialize a block of fluid directly above the shape, then let it fall due to gravity. The entire scene takes place in a $1 \\times 1 \\times 2$ m container. A random rotation is applied to", + "bbox": [ + 76, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the base shape which sits on the bottom of the container, and the initial position of the shape is also randomized. We extract 800 time-steps from the ground-truth fluid rollout (with $\\Delta t = 0.005$ ); this corresponds to 4000 steps for the classical simulator, which requires very small time-steps for stability [25].", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use five \"primitive\" shapes in our dataset: spheres, boxes, cylinders, cones, and toruses. These shapes have SDFs with analytical forms. (E.g. the simplest form is the origin-centered, radius- $r$ sphere with SDF $F(p) = ||p||_2 - r$ ; the others tend to be more complex). We emphasize that SURFSUP is trained only on simulations containing these primitive shapes and generalizes zero-shot to complex objects and scenes. Our key insight is that accurate fluid-surface simulation depends on understanding local surface geometry, of which a significant diversity is expressed by these shapes (like curvature, holes, differently-sloped faces, corners). Leveraging this and the properties of SDFs, SURFSUP can generalize to interactions with surfaces of entirely novel global geometry. Our training and test set consists of 1000 and 100 simulations respectively. See Supp. Fig 6 for example initializations.", + "bbox": [ + 496, + 181, + 892, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Generalization Test-Sets", + "text_level": 1, + "bbox": [ + 500, + 441, + 723, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We create a range of qualitative and quantitative evaluation scenarios to test our model's generalization ability. The first three below assess generalization to novel analytical SDFs (e.g. unions $\\min(f_1, f_2)$ or differences $\\max(f_1, -f_2)$ ), and the last to real-world shapes represented as neural SDFs $f_\\theta$ .", + "bbox": [ + 496, + 465, + 890, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prim-OOD (60 simulations). This contains the same shapes with OOD shape parameters from the training set (e.g. cylinder with higher radius $r$ and lower height $h$ ).", + "bbox": [ + 496, + 551, + 890, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prim-Unions (40). This test set includes all possible pairwise unions which 'merge' two of the primitive shapes. See Supp. for examples.", + "bbox": [ + 496, + 598, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Funnels (12). We examine the model's ability to handle analytical \"differences\" of shapes by creating funnels, formed by carving an inverted cone out of a cylinder.", + "bbox": [ + 496, + 643, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "'Complex' Scenes (7). As the key, challenging test of our model's generalization, we simulate fluid interacting with several complex real-world objects. To do so, we leverage neural SDFs of shapes, which can represent complex shapes with high accuracy and fidelity (Sec. 3.3). Instead of an analytical SDF $f(p)$ , we use a trained neural SDF $f_{\\theta}(p)$ to generate our SDF features during our forward pass. To test our model's ability to generalize zero-shot to complex objects/scenes, we train a single SIREN model [30] for 7 fine-grained objects (including the coral, lion, and room scenes shown in Fig. 1, Fig. 4). For each object we create a fluid scene using the ground-truth simulator, and evaluate SURFSUP on this small Complex-Scenes dataset. To represent the geometry at an appropriate scale, these objects are", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14229", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e2fadaabcd31e738f56913c883d697e7eb5585d05c2026638e906365a8af7677.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Testing SetChamferChamfer SurfaceNumber InsideMean SDF Inside
SURFSUPGNS [25]SURFSUPGNSSURFSUPGNSSURFSUPGNS
Primitives0.02970.02854.7045.45102000-1.04e-11-2.48e-05
Primitives-OOD0.03210.034310.95118.44532433006-1.65e-06-0.0006
Primitives-Unions0.03800.034824.55326.3071000063000-4.46e-05-0.001
Funnels0.03620.05219.99013.66327204-2.51e-09-0.001
Complex-Scenes0.06380.265850.521276.17640513257722-0.0011-0.026
", + "bbox": [ + 84, + 88, + 733, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Results on in-distribution (Primitives) and OOD test sets, comparing SURFSUP to GNS [25] which represents the surface as particles.", + "bbox": [ + 741, + 89, + 890, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scaled 2x in the fluid scene compared to previous datasets. To test our model on families of shapes, which gives us the capacity to solve inverse problems, we train a variational DeepSDF model [22] for the ShapeNet chair and bowl categories. See Supp. for details on these scenes and models.", + "bbox": [ + 75, + 237, + 468, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Training Details", + "text_level": 1, + "bbox": [ + 76, + 324, + 236, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For all experiments, we train our model on the PrimShapes dataset. The training data is pairs $X^{(t)}, X^{(t+1)}$ from the ground-truth simulator. We use a graph network (GN) architecture with 10 message-passing steps and train with Adam for 2M steps (see Supp. for details).", + "bbox": [ + 75, + 348, + 468, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Metrics and Baseline", + "text_level": 1, + "bbox": [ + 76, + 436, + 274, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although we train on next-step prediction, we measure performance on full rollouts. We use mean Chamfer distance between the predicted particles $P$ and ground-truth $G$ to measure overall rollout agreement (averaging across timesteps). Compared to MSE, Chamfer distance is permutation-invariant and therefore a better measure of similarity between fluid flows [25]. This metric, however, can be noisy since it averages over long rollouts on chaotic systems with many particles. Since we focus on the quality of fluid-surface interactions, we also compute a \"Chamfer surface\" metric that focuses on prediction errors near the surface. For each time-step, we consider the predicted point cloud only near the surface (where $F(p) < \\alpha$ for threshold $\\alpha$ ) and measure the Chamfer distance of $P_{\\alpha}$ to $G$ , summing over particles; we do this symmetrically for $G$ . Intuitively this focuses on discrepancies between $P$ and $G$ for fluid motion near the surface. See Supp. for further details.", + "bbox": [ + 73, + 460, + 468, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since surface penetration is a serious failure mode, we also measure the number of particles inside the surface through the rollout (summing over the number of timesteps). Additionally, we compute the average SDF value for particles inside the surface, indicating the extent of surface penetration (since SDFs become increasingly negative deeper into the surface interior).", + "bbox": [ + 75, + 718, + 468, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline. Our main comparison is to the graph network simulator (GNS) proposed by Sanchez Gonzalez et al. [25]. Their approach represents surfaces explicitly through discretization into particles, which are added to the graph and participate in the message-passing. To implement this base", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "line, we convert all the SDFs in our datasets to meshes using Marching Cubes [15]. From this mesh we uniformly sample 2000 surface particles (comparable to the number of fluid particles in our simulations) and add them to the graph. In all our experiments we compare our implicit SURFSUP approach to this baseline with explicit surface particles.", + "bbox": [ + 496, + 237, + 890, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While the baseline method could in principle be extended to include local surface information (e.g. using surface normal estimates for each particle as additional features, or mesh connectivity), we omit such an extension here. This is because any explicit representation of large surfaces quickly becomes computationally infeasible, especially for large objects/scenes or complex geometry. Whereas the simulation quality of SURFSUP scales with the capacity of the SDF network $f_{\\theta}$ , the baseline quality scales with the number of particles. We show in Sec 5 that SURFSUP still outperforms (with lower computational cost and inference time) a baseline with more particles.", + "bbox": [ + 496, + 329, + 892, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 500, + 527, + 591, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SURFSUP can accurately predict fluid behavior over long rollouts, interacting with a range of complex surfaces (Fig. 4). Across the different analytical primitive shapes in our test sets and different orientations, we can generate accurate and visually realistic predictions of fluid behavior. See the Supp. Material for videos of these, and all, rollouts in the paper. Our key result is the following: we find that SURFSUP can perform effective simulation with surfaces that are significantly out-of-distribution, and even generalize to complex real-world shapes and scenes. This demonstrates that SURFSUP can leverage the rich geometric information provided by the SDF for generalizing to novel surfaces expressed as neural SDFs.", + "bbox": [ + 496, + 551, + 890, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results and Comparison to Baseline. We evaluate SURFSUP on rollouts from ID and OOD test sets of shapes with analytical SDFs, and on complex scenes, with comparison to the GNS baseline. Results are shown in Table 1. Across all test sets, our model's fluid predictions display significantly lower surface penetration than the baseline (e.g. 324 vs. 33006 particles on average for the primitives+OOD test set). Moreover, for those few particles which have penetrated the surface, the average SDF value for our method is nearly zero; this implies that any", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14230", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/740a1f905fe8766f130c663479ba46e507a6d88bed38d43b2288966d2e3941d4.jpg", + "image_caption": [ + "Figure 4: Although SURFSUP is only trained on primitive shapes, our model generalizes to complex and real-world shapes without any additional training. We visualize our predictions (middle) for two scenes. In comparison, the baseline (top) often generates unrealistic predictions, such as having water penetrate solid objects. In many cases, the predictions from the baseline cause the water to disappear because it falls through the floor (top right)." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 500, + 410 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94c75b747b8057e4de536fbe705e2fc8cf51ad5f836569d506b66cf7f7f5489f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 90, + 890, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "particle that penetrates the boundary $(F(p) < 0)$ is quickly projected outward to the region $F(p) > 0$ . On Chamfer distance near the surface, which measures overall accuracy of fluid-surface interactions beyond surface penetration, our method consistently improves over the baseline. Overall Chamfer distance between the two models is generally comparable. This is unsurprising given that small differences in particle positions along long time-horizons, potentially far from the surface, can accumulate error. (Sec. 4.3). See Supp. Fig. 5 for rollouts on rows 1-4 of Table 1, indicating the stability and accuracy of our method's dynamics.", + "bbox": [ + 75, + 508, + 472, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalizing to Complex Scenes We evaluate SURFSUP on chairs and bowls from Shapenet, and fine-grained objects and scenes in the real world (See Fig. 1, Fig. 4, Fig. 5, and Fig. 8, also videos in Supp.). Remarkably SURFSUP trained only with primitive shapes generalizes zero-shot to complex shapes/scenes represented as neural SDFs. This is notable because our chosen real-world objects differ widely (see the variation in Fig. 4 and Fig. 5), and second, neural SDFs are imperfect compared to analytical SDFs (e.g. not fulfilling the property $||\\nabla f|| = 1$ ). Despite these challenges, SURFSUP produces rollouts with accurate and realistic fluid dynamics, capturing fine-grained behavior near complex surfaces. Our method also avoids unintuitive artifacts, such as surface penetration, seen in the GNS baseline.", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results (Table 1, row 'Complex-Scenes') indicate that our method outperforms the baseline most significantly on complex real-world scenes. These results elucidate key generalization advantages of our method in handling surfaces with complex geometry and large scenes at scale. For the coral scene (Fig. 1), our method's fluid rolls accurately capture and percolate through the fine tentacle structure of the coral. The GNS baseline does not display this ability (Supp. Fig 1). Our method also scales accurately to large scenes such as the room scene in Fig. 1. By contrast, the baseline's predictions consistently exhibit penetration of the furniture surfaces in the scene due to insufficient resolution (see Fig. 6). Recall (Sec. 4.3) that our scaling law is the capacity of the neural SDF network $f_{\\theta}$ instead of the number of surface particles.", + "bbox": [ + 496, + 508, + 893, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency of Simulation. We study comparing SURFSUP as we increase the number of surface particles for the GNS baseline (which in our experiments is afforded 2000 surface particles). Note that there is a tradeoff between performance and computational cost (memory, time) for the baseline; more particles improves surface resolution while also increasing graph sizes. We examine a scene in the Primitives-Unions test set (union of a sphere and box), for which GNS's predictions display significant surface penetration, and increase surface particles from 2000 to 5000.", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14231", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6d6c3ac613d53f68d3380410ae407fae369cd4d5714ed082d23af32a23e6110.jpg", + "image_caption": [ + "Figure 5: SURFSUP predictions on ShapeNet objects." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 434, + 296 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e12890e6b577710cd29589628ed05516253bc683865d57411114968c4c9d8302.jpg", + "image_caption": [ + "Figure 7: Comparing our method vs. the baseline with more surface particles. (a) Left shows how surface penetration is affected; even as the baseline improves with more surface particles, our method maintains an order magnitude improvement (red line, 490 particles). Right shows that the graph size increases significantly with the number of surface particles; we measure the max number of edges during the rollout, which indicates memory cost. (b) Final rollout timesteps for our method and the GNS at 2000, 3000, and 4000 particles, which reflect the quantitative results." + ], + "image_footnote": [], + "bbox": [ + 81, + 339, + 478, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Correspondingly GNS's fluid predictions improve, with less surface penetration (Fig. 7, top left and bottom). However, the baseline with 5000 particles still shows an order magnitude more penetration than our method (12285 vs. 490 particles). Moreover its computational cost, measured as the maximum particle graph size (# edges) during the rollout, almost doubles with 5000 surface particles (Fig. 7, top right). In contrast SURFSUP achieves near-zero surface-penetration with constant, smaller fluid graphs, overcoming the baseline's cost-performance tradeoff. We similarly analyze the room scene (Fig. 1), increasing the number of surface particles to 5K and 10K. Remarkably, SURFSUP here outperforms even a 10000-particle baseline (Ours 125.39 vs. baseline 194.98 Chamfer Surface; Supp.for details). This suggests fundamental advantages in scaling for simulation with implicit surfaces. We perform an inference time study (Supp.) showing that our method achieves speedups over GNS and the ground-truth simulator.", + "bbox": [ + 75, + 536, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scaling up SURFSUP. As a final benchmark of unusual size and complexity, we assessed SURFSUP on a natural scene with complex topography (Puncak Jaya mountains, Indonesia). This scene was $40 \\times 27 \\times 6 \\, \\text{m}^3$ in size (6840 m³ compared to roughly 10 m³ for previous scenes), and contained 68,600 fluid particles (order magnitude more than", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c371af9480e01e61b8acaa4a77163a3092fda8e9e707452ba9ecd97762c3c2a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 88, + 710, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78f72b7d0a30b416c34797665abf16019a263d25ada706c90a379143130ecf6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 179, + 890, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/552256be592444af0fad124718e80260b43e23781b885f9b38b4491cf671bc07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 90, + 905, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ba7745f9abf574b5124c008e99dac88e100ebeb2a72516a9db68801b4434cd5.jpg", + "image_caption": [ + "Figure 6: Fine-grained comparison of SURFSUP and GNS on four parts of the room scene. Our dynamics accurately model the furniture while the baseline's leak into surfaces.", + "Figure 8: Predicted fluid rollouts from SURFSUP on different parts of a $40 \\times 27 \\times 6$ m³ mountain scene." + ], + "image_footnote": [], + "bbox": [ + 503, + 444, + 911, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "previous scenes, see Supp. for details). Fig. 8 shows the predictions of SURFSUP, which correctly handle complex features such as slopes, confluences, and valleys between the mountains. Note that while GNS would require $O(1M)$ particles to represent this scene, we leave the network size for SIREN unchanged from smaller scenes. These results demonstrate the potential of our method for large-scale, realistic dynamics. Please see Supp. for rollout videos.", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14232", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inverse Design. Finally, we evaluate SURFSUP on its potential to solve object design tasks. We show the ability of our model to design both parametrized shapes, by optimizing the shape parameters with gradients, as well as shapes represented by neural SDF models.", + "bbox": [ + 75, + 90, + 472, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Parametrized Design Tasks. We consider the funnel, which is an inverted capped cone carved out of a cylinder, for which we have derived an analytical SDF. This shape is parametrized by the cylinder radius $R$ , height $H$ , cone height $h = H$ ; and the radii $r_1$ and $r_2$ of the bottom and top circles that \"cap\" the cone, with $r_1 > r_2$ .", + "bbox": [ + 75, + 169, + 468, + 261 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our setup, we drop a block of fluid directly onto the object. We consider two tasks: designing a bowl where the task is to contain the fluid inside the object, and a funnel where the task is to concentrate the water onto a location onto the ground. We optimize $r_1$ (which we call the \"capture radius\" that captures fluid into the object) and $r_2$ (the \"filter radius\" which decides how much water to filter onto the ground). For the funnel, the desired solution is large $r_1$ and small $r_2$ , and for the bowl we want $r_2 \\rightarrow 0$ . If our learned model is performing well, the model should be able to discover these solutions.", + "bbox": [ + 75, + 262, + 470, + 428 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We roll out our model for 50 (bowl) and 75 (funnel) timesteps. Our reward measures the log probability of a Gaussian centered at the desired location of the fluid particles. For the bowl task, we use the bottom of the object, i.e. $\\mathcal{N}((0.,0.,h),\\Sigma)$ , where $h$ is the object height. For the funnel we use a 2D Gaussian $\\mathcal{N}((0.,0.),\\Sigma)$ and measure the reward on particles that reach the ground during training. For a given design, we roll out our model and obtain gradients, which we use to iteratively refine the design.", + "bbox": [ + 75, + 431, + 468, + 568 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Qualitative results are shown in Fig. 9. Starting from a sub-optimal design with low $r_1$ and $r_2 > 0$ , our model successfully converges for the bowl; similarly the funnel converges from a suboptimal to an optimal solution. This shows that over long rollouts and hundreds of message-passing steps with dense fluid-surface interactions, our model's gradients are useful for solving inverse design problems. We also find that design is robust across different initializations of $r_1$ and $r_2$ (details in Supp.).", + "bbox": [ + 75, + 569, + 470, + 707 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Latent Space Design. We further examined whether SURF-SUP could directly optimize in the latent space of a DeepSDF model, trained on ShapeNet chairs. For this investigation, as a proof-of-concept we aimed to discover a novel chair that traps and contains falling water (i.e. maximizes the \"bowl reward\" above). Given a DeepSDF model $f_{\\theta}(p,z)$ with latent code $z_0$ , we iteratively refine the design by rolling out our simulator for 100 time-steps, computing a reward, and updating $z$ via gradient-based optimization. Across multiple trials, our model converges from an suboptimal initial chair to a final design that can contain water and retains the semantics of a chair (Fig. 10).", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5ca803e74f2c3bea259c6a6ba3ff5a72dc98f854e31885982d55bcb6120f0e23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 90, + 624, + 200 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/68f77f40fb030f28717d6bfa71d3a2f3cd06d1ecaf763ad82430e86390afcc14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 90, + 743, + 205 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/91dea91e98a968b4c31f3e8ac9df33ddf919237c6d436cd1b22e5cbfaaa228b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 90, + 875, + 205 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/626cc9c5ff3c2f95e55a716f5b67e0e8e1d1d6cb671e11b0eb72342d03f5892c.jpg", + "image_caption": [ + "Initialization" + ], + "image_footnote": [], + "bbox": [ + 500, + 209, + 625, + 321 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/71fbad5635e23a684bd2bdaec298f0a8dc587f8e11151d40647b52c4ff6a5705.jpg", + "image_caption": [ + "Intermediate" + ], + "image_footnote": [], + "bbox": [ + 638, + 209, + 743, + 321 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d3a17ae70925fdab343280ed35aa90b68e6e2d3c49041a48cd252eb4b996ad15.jpg", + "image_caption": [ + "Final", + "Figure 9: Design optimization; top is the funnel task, bottom the bowl task. For each frame, the design is shown near the top of the container and a timestep from the fluid rollout is shown below. Initial, intermediate, and final designs are shown; note that the reward increases through optimization and final designs approximate a funnel and bowl well." + ], + "image_footnote": [], + "bbox": [ + 774, + 209, + 875, + 321 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a03ff94c8af6f1fbdff8aee63b1e62846b521f40cb104c172223cf09c04b8dae.jpg", + "image_caption": [ + "Initial Design" + ], + "image_footnote": [], + "bbox": [ + 506, + 472, + 591, + 554 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/94c16d356dcdaf51926be6fa5baa27f61d753250d04f43d47aeb7246d5b8a3bd.jpg", + "image_caption": [ + "Final Design" + ], + "image_footnote": [], + "bbox": [ + 633, + 472, + 722, + 553 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c0bdf3c72a59d9e696ca3cfe3af6a0eecced8146a4ddcae17b0e7041f42703b2.jpg", + "image_caption": [ + "Reward Convergence", + "Figure 10: We differentiate through SURFSUP to directly optimize (in the latent space of a DeepSDF) a chair that holds water dropped onto it." + ], + "image_footnote": [], + "bbox": [ + 743, + 474, + 888, + 556 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 619, + 619, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we have taken a first step towards integrating implicit 3D representations into the dynamical simulation of physical models. SURFSUP realizes this approach by incorporating the implicit representation of solid surfaces (via SDFs) into a graph neural network fluid simulator. As such, SURFSUP is able to accurately model how fluids interact with surfaces, including highly complex surfaces unseen in training. We find that promoting this type of OOD generalization is helpful in both forward and inverse contexts.", + "bbox": [ + 496, + 643, + 890, + 779 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements: We thank Huy Ha for help with visualizations of fluid rollouts. We thank Jan Bender for helpful feedback on SPlisHSPlasH, and Kelsey Allen, Tobias Pfaff, and Alvaro Sanchez-Gonzalez for advice on graph network simulators. We thank Max Helman for early setup work on SPlisHSPlasH. This research is partially supported by the NSF STC for Learning the Earth with Artificial Intelligence and Physics, and the NSF NRI Award #1925157. AM is supported by the NSF fellowship.", + "bbox": [ + 496, + 780, + 892, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14233", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas J Guibas. Learning representations and generative models for 3d point clouds. arXiv preprint arXiv:1707.02392, 2017.", + "[2] Kelsey R Allen, Tatiana Lopez-Guevara, Kimberly Stachenfeld, Alvaro Sanchez-Gonzalez, Peter Battaglia, Jessica Hamrick, and Tobias Pfaff. Physical design using differentiable learned simulators. arXiv preprint arXiv:2202.00728, 2022.", + "[3] Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interaction networks for learning about objects, relations and physics. Advances in neural information processing systems, 29, 2016.", + "[4] Jan Bender and Dan Koschier. Divergence-free smoothed particle hydrodynamics. In Proceedings of the 14th ACM SIGGRAPH / Eurographics Symposium on Computer Animation, SCA '15, page 147-155, New York, NY, USA, 2015. Association for Computing Machinery.", + "[5] Jan Bender and Dan Koschier. Divergence-free smoothed particle hydrodynamics. In Proceedings of the 14th ACM SIGGRAPH / Eurographics Symposium on Computer Animation, SCA '15, page 147-155, New York, NY, USA, 2015. Association for Computing Machinery.", + "[6] Siddhartha Chaudhuri, Daniel Ritchie, Jiajun Wu, Kai Xu, and Hao Zhang. Learning generative models of 3d structures. In Computer Graphics Forum, volume 39, pages 643-666. Wiley Online Library, 2020.", + "[7] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), 2016.", + "[8] James D Foley, Foley Dan Van, Andries Van Dam, Steven K Feiner, and John F Hughes. Computer graphics: principles and practice, volume 12110. Addison-Wesley Professional, 1996.", + "[9] Christoph Gissler, Andreas Peer, Stefan Band, Jan Bender, and Matthias Teschner. Interlinked sph pressure solvers for strong fluid-rigid coupling. ACM Transactions on Graphics (TOG), 38(1):1-13, 2019.", + "[10] Zekun Hao, Hadar Averbuch-Elor, Noah Snavely, and Serge Belongie. Dualsdf: Semantic shape manipulation using a two-level representation. pages 7631–7641, 2020.", + "[11] Dan Koschier and Jan Bender. Density maps for improved sph boundary handling. In Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 1-10, 2017.", + "[12] Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. In ICLR, 2019.", + "[13] Yunzhu Li, Jiajun Wu, Jun-Yan Zhu, Joshua B Tenenbaum, Antonio Torralba, and Russ Tedrake. Propagation networks for model-based control under partial observation. In 2019 International Conference on Robotics and Automation (ICRA), pages 1205-1211. IEEE, 2019." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Zhijian Liu, William T Freeman, Joshua B Tenenbaum, and Jiajun Wu. Physical primitive decomposition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-19, 2018.", + "[15] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987.", + "[16] Andreas Mayr, Sebastian Lehner, Arno Mayrhofer, Christoph Kloss, Sepp Hochreiter, and Johannes Brandstetter. Boundary graph neural networks for 3d simulations. arXiv preprint arXiv:2106.11299, 2021.", + "[17] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[18] Mariem Mezghanni, Théo Bodrito, Malika Boulkenafed, and Maks Ovsjanikov. Physical simulation layer for accurate 3d modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13514-13523, 2022.", + "[19] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019.", + "[20] J. J. Monaghan. Smoothed particle hydrodynamics. Annual Review of Astronomy and Astrophysics, 30:543-574, Jan. 1992.", + "[21] Damian Mrowca, Chengxu Zhuang, Elias Wang, Nick Haber, Li F Fei-Fei, Josh Tenenbaum, and Daniel L Yamins. Flexible neural representation for physics prediction. Advances in neural information processing systems, 31, 2018.", + "[22] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019.", + "[23] Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh-based simulation with graph networks. In International Conference on Learning Representations, 2020.", + "[24] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 22468-22478. Curran Associates, Inc., 2020.", + "[25] Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In International Conference on Machine Learning, pages 8459-8468. PMLR, 2020.", + "[26] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20(1):61-80, 2008." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14234", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] C. Schenck and D. Fox. Spnets: Differentiable fluid dynamics for deep neural networks. In Proceedings of the Second Conference on Robot Learning (CoRL), Zurich, Switzerland, 2018.", + "[28] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. 2019.", + "[29] Vincent Sitzmann, Eric R. Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In Proc. Neurips, 2020.", + "[30] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Proc. NeurIPS, 2020.", + "[31] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In Computer Vision and Pattern Recognition (CVPR), 2017.", + "[32] Benjamin Ummenhofer, Lukas Prantl, Nils Thuerey, and Vladlen Koltun. Lagrangian fluid simulation with continuous convolutions. In International Conference on Learning Representations, 2019.", + "[33] Subeesh Vasu, Nicolas Talabot, Artem Lukoianov, Pierre Baqué, Jonathan Donier, and Pascal Fua. Hybridsdf: Combining deep implicit shapes and geometric primitives for 3d shape representation and manipulation. In International Conference on 3D Vision, 2022.", + "[34] Zhirong Wu, S. Song, A. Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and J. Xiao. 3d shapenets: A deep representation for volumetric shapes. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, Los Alamitos, CA, USA, jun 2015. IEEE Computer Society." + ], + "bbox": [ + 78, + 90, + 468, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14235", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/e47c0301-5f63-4201-a2ee-b44e21fee063_model.json b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/e47c0301-5f63-4201-a2ee-b44e21fee063_model.json new file mode 100644 index 0000000000000000000000000000000000000000..60571fd34f8fe6d871b3149dbfeb28e5224b8bf4 --- /dev/null +++ b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/e47c0301-5f63-4201-a2ee-b44e21fee063_model.json @@ -0,0 +1,1949 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.197, + 0.131, + 0.775, + 0.154 + ], + "angle": 0, + "content": "SURFSUP: Learning Fluid Simulation for Novel Surfaces" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.188, + 0.867, + 0.244 + ], + "angle": 0, + "content": "Arjun Mani\\*, Ishaan Preetam Chandratreya\\*, Elliot Creager\\*3, Carl Vondrick1, Richard Zemel1 \n1 Columbia University 2 University of Toronto 3 Vector Institute \nsurfsup.cs.columbia.edu" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.27, + 0.85, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.587, + 0.877, + 0.605 + ], + "angle": 0, + "content": "Figure 1: We introduce SURFSUP, which learns to predict how fluid particles interact with novel, complex 3D surfaces." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.616, + 0.314, + 0.631 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.638, + 0.474, + 0.88 + ], + "angle": 0, + "content": "Modeling the mechanics of fluid in complex scenes is vital to applications in design, graphics, and robotics. Learning-based methods provide fast and differentiable fluid simulators, however most prior work is unable to accurately model how fluids interact with genuinely novel surfaces not seen during training. We introduce SURFSUP, a framework that represents objects implicitly using signed distance functions (SDFs), rather than an explicit representation of meshes or particles. This continuous representation of geometry enables more accurate simulation of fluid-object interactions over long time periods while simultaneously making computation more efficient. Moreover, SURFSUP trained on simple shape primitives generalizes considerably out-of-distribution, even to complex real-world scenes and objects. Finally, we show we can invert our model to design simple objects to manipulate fluid flow." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.616, + 0.633, + 0.631 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Across engineering and science, the simulation of fluid dynamics has become an invaluable tool, and it will be a crucial component for building visual systems that are capable of understanding and interacting with real-world environments. Recently, a new class of simulators has emerged that learn the dynamics of physical systems from data [13, 32, 25]. These simulators offer several advantages over classical simulators by increasing speed, reducing reliance on hand-crafted dynamics models, and providing differentiable rollouts for solving inverse problems. Recent approaches based on graph neural networks (GNNs) have in particular shown impressive accuracy and generalization on a wide range of fluids represented as particle or mesh-based systems [25, 23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In a wide variety of applications areas, fluid simulation must properly handle interaction with solid structures to attain practical relevance. Many problems involving" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.887, + 0.255, + 0.902 + ], + "angle": 0, + "content": "*Authors contributed equally." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14225" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "fluid simulation are fundamentally about object design; for example, designing aerodynamic shapes for cars and airplanes. In robotics too, an intelligent robot would have to understand fluid-surface interactions; for example, to carry a mug of hot coffee to a person without spilling the liquid from the container. In both cases, the need to accurately model how fluids interact with novel solid surfaces1—with new shapes, configurations, orientations, and so on—motivates an challenging out-of-distribution (OOD) generalization subproblem for fluid simulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.575 + ], + "angle": 0, + "content": "In this paper, we tackle this problem by introducing Surface implicit sUbstitution for particles (SURFSUP), a framework for simulating the interactions of fluids and surfaces by representing surfaces using implicit representations [19]. Current GNN-based approaches represent surfaces explicitly, discretizing them into particles and augmenting the graph accordingly [25]. Implicit representations, in particular signed distance functions which represent surfaces as zero-level sets of functions, offer several compelling advantages. They can smoothly represent continuous surfaces, provide rich geometric information (e.g. whether a point is in the interior/exterior of the object), and can scale easily to large objects and scenes. Due to these properties, many state-of-the-art generative models for 3D shapes learn shapes with signed distance functions. We argue that these properties are useful not just for statics but also dynamics: in order to accurately model fluid-surface interactions for potentially complex surfaces, smooth and informative representations of local surface geometry are needed. In turn, we posit that these locally informative representations will enable fluid simulators to appropriately model interaction with never-before-seen surfaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.473, + 0.727 + ], + "angle": 0, + "content": "Concretely, we propose an approach that integrates signed distance representations of surfaces [22, 19] with graph neural network models [26, 25]. Our architecture follows the encode-process-decode paradigm [25], where the particles in the physical system at time \\( T \\) are encoded into a graph, followed by several message-passing steps, and finally a decoder predicts the dynamics (accelerations). We represent the surface using its signed distance function, and exploit geometric properties of SDFs to model the distance and orientation of fluid particles with respect to the surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.473, + 0.879 + ], + "angle": 0, + "content": "Experiments show that our approach can model fluid-surface interactions over long rollouts with a high degree of accuracy and visual realism. SURFSUP can model surfaces far beyond its training domain; when trained on simple primitive shapes (e.g. spheres, cones, toruses), we generalize zero-shot to complex real-world objects and scenes represented using neural implicit generative models (Fig. 1). We compare to a method which represents the surface as particles inside the graph network simulator, and show that our model achieves better generalization and efficiency, and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "can better handle complex surface geometry. Finally, we demonstrate the potential of SURFSUP to solve inverse problems across two object design tasks. Overall, our paper bridges learned physical models for dynamics and 3D implicit representations, and takes a step towards AI models which can understand and help design the physical world." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.196, + 0.642, + 0.212 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.221, + 0.895, + 0.312 + ], + "angle": 0, + "content": "Learning Differentiable Simulators. Using machine learning for predicting and modeling complex physical systems is a rapidly growing area of research. Learned physics simulators have been shown to accurately simulate colliding rigid objects, deformable objects, fluids, and other complex systems [13, 3, 12, 32, 25, 21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Recent progress in learned fluid simulation has been driven by the Lagrangian viewpoint, which tracks the motion of fluid particles over time. Graph neural networks (GNNs) have in particular emerged as effective forward models for fluid simulation, and have demonstrated better stability and generalization compared to grid-based, convolutional approaches. [13, 14, 27, 25]. Sanchez-Gonzalez et al. [25] introduced a GNN architecture which can accurately and generalizably model fluid behavior over long rollouts; most subsequent work builds off their approach [23, 16]. However, most of these works generally consider simple surfaces; e.g. only cube-shaped containers for fluids. There has been little work focused on fluid-surface interactions; Mayr et al. [16] introduce an approach specifically for triangularized meshes. All these approaches use particles or meshes to represent surfaces; we will argue that implicit representations enable learning more accurate, generalizable physical dynamics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.895, + 0.735 + ], + "angle": 0, + "content": "Some works have explored the possibility of using learned fluid simulators for solving inverse problems [27, 12, 2]. Allen et al. [2] showed that graph network simulators can be used to solve design tasks by rolling out fluid trajectories and using gradient-based optimization. In 3D, they show that a 2D plane with a learned height field can be used to split a stream of fluid to fall into several locations on the ground. SDFs have the advantage of being able to learn geometry in a more smooth and unconstrained way, which could help solve inverse problems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "3D Implicit Representations. Implicit representations have become state-of-the-art for generative models of 3D objects [17, 22, 6, 29, 30]. These models represent the surface as a function \\( F(x) \\). There are several choices for \\( F \\). Park et al. [22] show that signed distance functions \\( F(x) \\) are particularly effective and informative representations. Compared to point clouds [1, 28] or voxels [7, 34], models based on SDF representations can smoothly represent complex geometry, are compact and efficient, and can be learned effectively [29, 6]. Follow-up work has studied learning \"dual\" generative models that jointly learn an" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.466, + 0.902 + ], + "angle": 0, + "content": "For brevity we refer to surfaces of solid objects simply as \"surfaces\"." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "14226" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "uncontrained representation and one composed of primitive shapes [10, 33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.257 + ], + "angle": 0, + "content": "Implicit Representations for Physical Dynamics. Some classical particle-based fluid simulators based on the Smoothed Particle Hydrodynamics (SPH) formalism [20] have transitioned from particle-based to implicit surface representations. [11, 9, 4]. In these simulators implicit surfaces allow for smoother fluid motion with less artifacts [11]. For general-purpose learned simulators (GNNs), there is little prior work using implicit rather than particle-based representations of surfaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.468, + 0.379 + ], + "angle": 0, + "content": "Another line of work combines implicit generative models with physical simulators [18, 24]. Mezghanni et al. [18] generate more physically stable SDF shapes by backpropagating gradients through a physical simulation layer. Their simulation mainly considers Newtonian dynamics. Our contribution is an effective, generalizable forward model for fluids that represents surfaces accurately and efficiently, and can also be used for solving inverse problems." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.394, + 0.168, + 0.409 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.419, + 0.421, + 0.435 + ], + "angle": 0, + "content": "3.1. Setup: Particle-Based Fluid Simulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.468, + 0.716 + ], + "angle": 0, + "content": "We address the problem of learning the physical dynamics of fluids interacting with surfaces. Given the state of a physical system \\( X^0 \\) and some initial conditions (e.g. gravity, other external forces), rolling out a fluid simulation over \\( T \\) steps produces a sequence of states \\( X^{(1)},\\dots,X^{(T)} \\). In this paper, we focus on particle-based simulation; the state of the system at time \\( t \\) is represented by a set of particles \\( X^{(t)} = (p_1^{(t)},\\dots,p_n^{(t)}) \\), where \\( p_i \\) is the position of particle \\( i \\) in 3D space. A rollout over \\( T \\) steps is produced by repeatedly applying a forward model \\( S_{\\theta}:X^{(t)}\\to X^{(t + 1)} \\), which takes in the particle positions at time \\( t \\) (and optionally a short history \\( t - 1,\\ldots \\)) and produces the next-step particle positions. While the forward model can be used to directly predict \\( X^{(t + 1)} \\), most methods instead predict intermediate values \\( Y \\), typically next-state accelerations for each particle, which can then be numerically integrated (e.g. Euler integration) to produce \\( X^{(t + 1)} \\); concretely, our task is to learn the parameters \\( \\theta \\) of this forward model \\( f_{\\theta}:X^{(t)}\\to Y \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.468, + 0.743 + ], + "angle": 0, + "content": "3.2. Graph Networks for Particle-Based Simulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Graph neural networks lend themselves naturally to particle-based simulation, since they compute pairwise interactions between nodes and aggregate these interactions. Graph Network-based Simulators (GNS), a model architecture introduced by Sanchez-Gonzalez et al. [25], have in particular shown state-of-the-art performance on simulating fluids over long rollouts. These methods follow an \"encode-process-decode\" paradigm. First, the state \\( X^{(t)} \\) is encoded into a graph, where nodes \\( v_{i} \\) are instantiated using a set of features (e.g. previous velocities, parti" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.322 + ], + "angle": 0, + "content": "cle type, etc.); nodes \\( v_{i}, v_{j} \\) are then connected with an edge \\( e_{ij} \\) if their distance is less than a connectivity radius \\( \\epsilon \\). Each node and edge is associated with an embedding. The \"processor\" stage then passes the encoded graph through \\( M \\) message-passing steps. Each message-passing step takes the current graph \\( \\{(v_{i}, v_{j}, e_{ij})\\}^{(m)} \\). An edge module first computes the updated edge embeddings \\( e_{ij}^{(m+1)} = f_{e}(v_{i}^{(m)}, v_{j}^{(m)}, e_{ij}^{(m)}) \\); then the edge embeddings are aggregated to compute updated node embeddings \\( v_{i}^{(m+1)} = f_{n}(v_{i}^{(m)}, \\sum_{j \\in N(i)} e_{ij}^{(m)}) \\), where \\( N(i) \\) are the set of neighbors of node \\( i \\). \\( f_{e} \\) and \\( f_{n} \\) are generally MLPs with 2-3 hidden layers, and the weights are shared across edges/nodes (although differing across message passing steps). Finally, the \"decoder\" step applies an MLP \\( f_{D} \\) to the final node embeddings to predict the accelerations \\( \\{y_{i}\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.458 + ], + "angle": 0, + "content": "In our problem, we further assume that there exists some rigid surface \\(S\\) that interacts with the fluid throughout the simulation. The GNS approach handles surfaces by discretizing the surface into particles (i.e. a point cloud) and adding these particles \\(\\{s_i\\}\\) to the graph (denoting them by a different particle type, and masking during position updates) [25]. Follow-up work for mesh-based simulation handles surfaces similarly by representing the surface as an irregular mesh and adding nodes for mesh vertices [23]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.467, + 0.892, + 0.483 + ], + "angle": 0, + "content": "3.3. Modeling Fluid-Surface Interactions with SDFs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.567 + ], + "angle": 0, + "content": "We propose to model the interactions of fluids and surfaces with signed distance function (SDF) representations of surfaces. An SDF represents a watertight surface as a function \\( F(p) = s \\), which takes a spatial point \\( p \\in \\mathbb{R}^3 \\) and outputs its distance \\( s \\) to the closest corresponding point on" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.587, + 0.821, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Figure 2: Rigid surfaces can be represented compactly and implicitly as the level set of a SDF: \\(\\{x|F(x) = 0\\}\\). Evaluating \\(F\\) and its derivatives at particle position \\(p\\) provides useful information for simulating the interaction of that particle with the surface." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14227" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.183 + ], + "angle": 0, + "content": "the surface. The function is \"signed\", such that it takes on positive values for points outside the surface and negative values in the surface interior. The surface of the object is implicitly represented as the zero-level set \\( F(p) = 0 \\). Explicit representations of the surface can be constructed by applying meshing algorithms such as Marching Cubes [15]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.468, + 0.304 + ], + "angle": 0, + "content": "Almost every surface can be described by an SDF. For several primitive shapes SDFs can be written down analytically (e.g. spheres, cones, cylinders); boolean operations on these shapes such as unions and intersections give rise to more complex shapes [31]. SDFs can also be represented discretely as voxel grids; or more recently for finer geometry, as neural SDFs \\( F_{\\theta}(p) \\) which learn to accurately approximate \\( F \\) from mesh-based supervision." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.326, + 0.468, + 0.356 + ], + "angle": 0, + "content": "3.3.1 Locality and Geometric Information for Fluid-Surface Interactions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.468, + 0.547 + ], + "angle": 0, + "content": "A key insight of our method is that fluid-surface interactions are inherently about local rather than global surface geometry. The interaction of a fluid with a surface depends on a small local region of the surface, over which the geometry can have limited variation. This suggests that an effective representation of local surface geometry, trained on a variety of object shapes, can generalize OOD to shapes that may look different globally. By contrast, particles are not an effective representation of surface geometry; for any moderately complex surface, discretization leads to artifacts such as non-smooth fluid motion and even penetration of the surface boundary (we confirm this empirically)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.468, + 0.824 + ], + "angle": 0, + "content": "By contrast, we argue that SDFs are a smooth, richly informative representation of local surface geometry for learning generalizable fluid-surface interactions. Our key insight is shown in Figure 2. Given a solid surface represented as an SDF \\( F(p) \\), we can use properties of the SDF to inform the behavior of a fluid particle as it reaches the surface. The value of the SDF \\( F(p) \\) gives the distance from the particle at point \\( p \\) to the closest point on the surface \\( p' \\). Moreover, the SDF gradient \\( \\nabla F(p) \\) provides the unit-norm displacement vector from the particle to its closest surface point \\( \\frac{(p - p')}{||p - p'||} \\). This can be shown by observing that the SDF decreases most quickly on the line from \\( p \\) to \\( p' \\). Thus, the SDF and its gradient provide the distance and orientation of the fluid particle w.r.t. to its closest surface point; as the particle approaches the surface, its acceleration will more closely approach \\( \\nabla F(p) \\) and it will tend to rebound. Moreover, since the gradient varies smoothly because of the continuity of \\( F \\), these features do not suffer from artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Higher order derivatives of the SDF can also be used to gain geometric information. For example, we could march to the surface by computing \\( p - F(x)\\nabla F(x) \\). At the surface, the Hessian \\( \\nabla^2 F(x) \\) is the shape operator; its eigenvalues provide the principal directions of surface curvature" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.097, + 0.895, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.24, + 0.892, + 0.301 + ], + "angle": 0, + "content": "Figure 3: Given fluid particles and the SDF, our method augments the constructed graph with features from the SDF. We then apply message-passing and predict dynamics, weighting near-surface errors more highly in our loss." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.335, + 0.892, + 0.501 + ], + "angle": 0, + "content": "\\(\\kappa_{1}\\) and \\(\\kappa_{2}\\), and the mean curvature is \\(H = \\frac{1}{2} \\mathrm{Tr}(\\nabla^2 F)\\). We can also estimate the surface curvature via finite differences by slightly perturbing the gradient vector on each axis and computing the SDF (Fig. 2). In this paper, we only use the SDF and its gradient as features without curvature information; this can be seen as an effective theory where the relevant \"local region\" of the surface is a single closest point on the surface. As we show, modeling surfaces using SDF representations helps our fluid simulator generalize beyond the available training data by modeling fluid-surface interactions for unseen surfaces." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.533, + 0.586, + 0.547 + ], + "angle": 0, + "content": "3.4. Model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.564, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Having developed our ideas about SDFs and modeling fluid-surface interactions, we now describe our approach, which we call Surface implicit sUbstitution for particles (SURFSUP). We learn a forward predictor \\( f_{\\theta}:X^{(t)},F\\to Y \\), which takes as input a set of particle positions and a surface implicitly parametrized by an SDF \\( F \\). We adopt the encode-process-decode paradigm for our model, but make two key changes. In the encoder step, for each fluid particle \\( p_i \\) we add the value of the SDF \\( F(p_i) \\) and its gradient \\( \\nabla F(p_i) \\) to the initial node representations, such that the new representation is \\( v_{i}^{\\prime (0)} = [v_{i}^{(0)},F(p_{i}),\\nabla F(p_{i})] \\). This is the way that the surface is encoded into the model; subsequently no particles representing the surface are added to the graph. The updated node embeddings are then passed through the \"processor\" with several message-passing steps and then decoded to predict the accelerations as standard. The other modification we make is to weight particles close to the surface higher during training when predicting dynamics. This has the effect of teaching the model to prioritize modeling fluid-surface interactions accurately. Since the model is predicting next-step accelerations, our loss becomes:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14228" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.085, + 0.115, + 0.47, + 0.151 + ], + "angle": 0, + "content": "\\[\nL = \\sum_ {F (p _ {i}) > \\alpha} \\left\\| a _ {i} ^ {(t)} - \\hat {a} _ {i} ^ {(t)} \\right\\| ^ {2} + \\sum_ {F (p _ {i}) \\leq \\alpha} \\lambda \\left\\| a _ {i} ^ {(t)} - \\hat {a} _ {i} ^ {(t)} \\right\\| ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.162, + 0.47, + 0.268 + ], + "angle": 0, + "content": "In our experiments, we set \\(\\alpha\\) based on the \"neighborhood radius\" of the classical simulator we train against, and set \\(\\lambda = 5\\). We found that this weighting helped the model learn surface interactions more quickly without sacrificing overall accuracy on fluid prediction. Note that the SDF allows loss weighting to be implemented in a very natural way, by simple thresholding. Our approach is shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.277, + 0.44, + 0.294 + ], + "angle": 0, + "content": "3.5. Other Advantages of SDF Representations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.469, + 0.362 + ], + "angle": 0, + "content": "We have argued for SDFs from the perspective of achieving smooth, information-rich representations of local surface geometry. Here we mention other advantages of SDFs, each of which we demonstrate in our experimental results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.469, + 0.472 + ], + "angle": 0, + "content": "Efficient message passing in the GNN. By modeling surfaces implicitly, we reduce the number of unique particles needed to compose the GNN's state. This allows us to perform more lightweight and computationally efficient simulation, as well as run the model for more message-passing steps. This advantage becomes especially clear for large objects/scenes or objects with complex surface geometry." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.476, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Shape Design with SDFs. SDFs are a useful paradigm for shape design and assembly. This is because compositions of shapes can be generated by union, intersection, and difference operations; e.g. the union of two SDFs \\( F_{1} \\) and \\( F_{2} \\) is \\( \\min(F_{1}, F_{2}) \\). The construction of shapes using primitives and these operations is known as Constructive Solid Geometry (CSG), and is implemented in many CAD packages for object design [8]. Moreover the recent emergence of neural SDF models as state-of-the-art 3D shape representations, offers another compelling avenue for shape design with end-to-end learning. These suggest that good forward models with SDFs could be leveraged for design problems." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.671, + 0.216, + 0.688 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.696, + 0.227, + 0.71 + ], + "angle": 0, + "content": "4.1. Fluid Domains" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.469, + 0.825 + ], + "angle": 0, + "content": "We evaluate our approach on complex prediction tasks involving fluids interacting with a variety of surfaces. Our main contribution here is a 3D fluid dataset consisting of water interacting with a range of primitive shapes, across shape parameters, positions, and orientations. For the ground-truth simulator we use SPlisHSPlasH [5], a 3D-particle-based fluid simulator based on the SPH technique." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "PrimShapes Dataset. We train SURFSUP on a 3D water dataset, where the fluid interacts with various shapes. We initialize a block of fluid directly above the shape, then let it fall due to gravity. The entire scene takes place in a \\(1 \\times 1 \\times 2\\) m container. A random rotation is applied to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "the base shape which sits on the bottom of the container, and the initial position of the shape is also randomized. We extract 800 time-steps from the ground-truth fluid rollout (with \\(\\Delta t = 0.005\\)); this corresponds to 4000 steps for the classical simulator, which requires very small time-steps for stability [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.893, + 0.424 + ], + "angle": 0, + "content": "We use five \"primitive\" shapes in our dataset: spheres, boxes, cylinders, cones, and toruses. These shapes have SDFs with analytical forms. (E.g. the simplest form is the origin-centered, radius- \\( r \\) sphere with SDF \\( F(p) = ||p||_2 - r \\); the others tend to be more complex). We emphasize that SURFSUP is trained only on simulations containing these primitive shapes and generalizes zero-shot to complex objects and scenes. Our key insight is that accurate fluid-surface simulation depends on understanding local surface geometry, of which a significant diversity is expressed by these shapes (like curvature, holes, differently-sloped faces, corners). Leveraging this and the properties of SDFs, SURFSUP can generalize to interactions with surfaces of entirely novel global geometry. Our training and test set consists of 1000 and 100 simulations respectively. See Supp. Fig 6 for example initializations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.443, + 0.724, + 0.457 + ], + "angle": 0, + "content": "4.1.1 Generalization Test-Sets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.543 + ], + "angle": 0, + "content": "We create a range of qualitative and quantitative evaluation scenarios to test our model's generalization ability. The first three below assess generalization to novel analytical SDFs (e.g. unions \\(\\min(f_1, f_2)\\) or differences \\(\\max(f_1, -f_2)\\)), and the last to real-world shapes represented as neural SDFs \\(f_\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.599 + ], + "angle": 0, + "content": "Prim-OOD (60 simulations). This contains the same shapes with OOD shape parameters from the training set (e.g. cylinder with higher radius \\( r \\) and lower height \\( h \\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Prim-Unions (40). This test set includes all possible pairwise unions which 'merge' two of the primitive shapes. See Supp. for examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Funnels (12). We examine the model's ability to handle analytical \"differences\" of shapes by creating funnels, formed by carving an inverted cone out of a cylinder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "'Complex' Scenes (7). As the key, challenging test of our model's generalization, we simulate fluid interacting with several complex real-world objects. To do so, we leverage neural SDFs of shapes, which can represent complex shapes with high accuracy and fidelity (Sec. 3.3). Instead of an analytical SDF \\( f(p) \\), we use a trained neural SDF \\( f_{\\theta}(p) \\) to generate our SDF features during our forward pass. To test our model's ability to generalize zero-shot to complex objects/scenes, we train a single SIREN model [30] for 7 fine-grained objects (including the coral, lion, and room scenes shown in Fig. 1, Fig. 4). For each object we create a fluid scene using the ground-truth simulator, and evaluate SURFSUP on this small Complex-Scenes dataset. To represent the geometry at an appropriate scale, these objects are" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14229" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.734, + 0.203 + ], + "angle": 0, + "content": "
Testing SetChamferChamfer SurfaceNumber InsideMean SDF Inside
SURFSUPGNS [25]SURFSUPGNSSURFSUPGNSSURFSUPGNS
Primitives0.02970.02854.7045.45102000-1.04e-11-2.48e-05
Primitives-OOD0.03210.034310.95118.44532433006-1.65e-06-0.0006
Primitives-Unions0.03800.034824.55326.3071000063000-4.46e-05-0.001
Funnels0.03620.05219.99013.66327204-2.51e-09-0.001
Complex-Scenes0.06380.265850.521276.17640513257722-0.0011-0.026
" + }, + { + "type": "table_caption", + "bbox": [ + 0.743, + 0.09, + 0.892, + 0.211 + ], + "angle": 0, + "content": "Table 1: Results on in-distribution (Primitives) and OOD test sets, comparing SURFSUP to GNS [25] which represents the surface as particles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.238, + 0.47, + 0.314 + ], + "angle": 0, + "content": "scaled 2x in the fluid scene compared to previous datasets. To test our model on families of shapes, which gives us the capacity to solve inverse problems, we train a variational DeepSDF model [22] for the ShapeNet chair and bowl categories. See Supp. for details on these scenes and models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.325, + 0.238, + 0.342 + ], + "angle": 0, + "content": "4.2. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.469, + 0.426 + ], + "angle": 0, + "content": "For all experiments, we train our model on the PrimShapes dataset. The training data is pairs \\( X^{(t)}, X^{(t+1)} \\) from the ground-truth simulator. We use a graph network (GN) architecture with 10 message-passing steps and train with Adam for 2M steps (see Supp. for details)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.275, + 0.453 + ], + "angle": 0, + "content": "4.3. Metrics and Baseline" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.461, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Although we train on next-step prediction, we measure performance on full rollouts. We use mean Chamfer distance between the predicted particles \\( P \\) and ground-truth \\( G \\) to measure overall rollout agreement (averaging across timesteps). Compared to MSE, Chamfer distance is permutation-invariant and therefore a better measure of similarity between fluid flows [25]. This metric, however, can be noisy since it averages over long rollouts on chaotic systems with many particles. Since we focus on the quality of fluid-surface interactions, we also compute a \"Chamfer surface\" metric that focuses on prediction errors near the surface. For each time-step, we consider the predicted point cloud only near the surface (where \\( F(p) < \\alpha \\) for threshold \\( \\alpha \\)) and measure the Chamfer distance of \\( P_{\\alpha} \\) to \\( G \\), summing over particles; we do this symmetrically for \\( G \\). Intuitively this focuses on discrepancies between \\( P \\) and \\( G \\) for fluid motion near the surface. See Supp. for further details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Since surface penetration is a serious failure mode, we also measure the number of particles inside the surface through the rollout (summing over the number of timesteps). Additionally, we compute the average SDF value for particles inside the surface, indicating the extent of surface penetration (since SDFs become increasingly negative deeper into the surface interior)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Baseline. Our main comparison is to the graph network simulator (GNS) proposed by Sanchez Gonzalez et al. [25]. Their approach represents surfaces explicitly through discretization into particles, which are added to the graph and participate in the message-passing. To implement this base" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.238, + 0.892, + 0.329 + ], + "angle": 0, + "content": "line, we convert all the SDFs in our datasets to meshes using Marching Cubes [15]. From this mesh we uniformly sample 2000 surface particles (comparable to the number of fluid particles in our simulations) and add them to the graph. In all our experiments we compare our implicit SURFSUP approach to this baseline with explicit surface particles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.893, + 0.512 + ], + "angle": 0, + "content": "While the baseline method could in principle be extended to include local surface information (e.g. using surface normal estimates for each particle as additional features, or mesh connectivity), we omit such an extension here. This is because any explicit representation of large surfaces quickly becomes computationally infeasible, especially for large objects/scenes or complex geometry. Whereas the simulation quality of SURFSUP scales with the capacity of the SDF network \\( f_{\\theta} \\), the baseline quality scales with the number of particles. We show in Sec 5 that SURFSUP still outperforms (with lower computational cost and inference time) a baseline with more particles." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.593, + 0.545 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.748 + ], + "angle": 0, + "content": "SURFSUP can accurately predict fluid behavior over long rollouts, interacting with a range of complex surfaces (Fig. 4). Across the different analytical primitive shapes in our test sets and different orientations, we can generate accurate and visually realistic predictions of fluid behavior. See the Supp. Material for videos of these, and all, rollouts in the paper. Our key result is the following: we find that SURFSUP can perform effective simulation with surfaces that are significantly out-of-distribution, and even generalize to complex real-world shapes and scenes. This demonstrates that SURFSUP can leverage the rich geometric information provided by the SDF for generalizing to novel surfaces expressed as neural SDFs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Quantitative Results and Comparison to Baseline. We evaluate SURFSUP on rollouts from ID and OOD test sets of shapes with analytical SDFs, and on complex scenes, with comparison to the GNS baseline. Results are shown in Table 1. Across all test sets, our model's fluid predictions display significantly lower surface penetration than the baseline (e.g. 324 vs. 33006 particles on average for the primitives+OOD test set). Moreover, for those few particles which have penetrated the surface, the average SDF value for our method is nearly zero; this implies that any" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14230" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.5, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.091, + 0.891, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.422, + 0.895, + 0.486 + ], + "angle": 0, + "content": "Figure 4: Although SURFSUP is only trained on primitive shapes, our model generalizes to complex and real-world shapes without any additional training. We visualize our predictions (middle) for two scenes. In comparison, the baseline (top) often generates unrealistic predictions, such as having water penetrate solid objects. In many cases, the predictions from the baseline cause the water to disappear because it falls through the floor (top right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.473, + 0.678 + ], + "angle": 0, + "content": "particle that penetrates the boundary \\((F(p) < 0)\\) is quickly projected outward to the region \\(F(p) > 0\\). On Chamfer distance near the surface, which measures overall accuracy of fluid-surface interactions beyond surface penetration, our method consistently improves over the baseline. Overall Chamfer distance between the two models is generally comparable. This is unsurprising given that small differences in particle positions along long time-horizons, potentially far from the surface, can accumulate error. (Sec. 4.3). See Supp. Fig. 5 for rollouts on rows 1-4 of Table 1, indicating the stability and accuracy of our method's dynamics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Generalizing to Complex Scenes We evaluate SURFSUP on chairs and bowls from Shapenet, and fine-grained objects and scenes in the real world (See Fig. 1, Fig. 4, Fig. 5, and Fig. 8, also videos in Supp.). Remarkably SURFSUP trained only with primitive shapes generalizes zero-shot to complex shapes/scenes represented as neural SDFs. This is notable because our chosen real-world objects differ widely (see the variation in Fig. 4 and Fig. 5), and second, neural SDFs are imperfect compared to analytical SDFs (e.g. not fulfilling the property \\(||\\nabla f|| = 1\\)). Despite these challenges, SURFSUP produces rollouts with accurate and realistic fluid dynamics, capturing fine-grained behavior near complex surfaces. Our method also avoids unintuitive artifacts, such as surface penetration, seen in the GNS baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.895, + 0.738 + ], + "angle": 0, + "content": "Quantitative results (Table 1, row 'Complex-Scenes') indicate that our method outperforms the baseline most significantly on complex real-world scenes. These results elucidate key generalization advantages of our method in handling surfaces with complex geometry and large scenes at scale. For the coral scene (Fig. 1), our method's fluid rolls accurately capture and percolate through the fine tentacle structure of the coral. The GNS baseline does not display this ability (Supp. Fig 1). Our method also scales accurately to large scenes such as the room scene in Fig. 1. By contrast, the baseline's predictions consistently exhibit penetration of the furniture surfaces in the scene due to insufficient resolution (see Fig. 6). Recall (Sec. 4.3) that our scaling law is the capacity of the neural SDF network \\( f_{\\theta} \\) instead of the number of surface particles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Efficiency of Simulation. We study comparing SURFSUP as we increase the number of surface particles for the GNS baseline (which in our experiments is afforded 2000 surface particles). Note that there is a tradeoff between performance and computational cost (memory, time) for the baseline; more particles improves surface resolution while also increasing graph sizes. We examine a scene in the Primitives-Unions test set (union of a sphere and box), for which GNS's predictions display significant surface penetration, and increase surface particles from 2000 to 5000." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "14231" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.436, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.31, + 0.451, + 0.326 + ], + "angle": 0, + "content": "Figure 5: SURFSUP predictions on ShapeNet objects." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.34, + 0.48, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.476, + 0.471, + 0.522 + ], + "angle": 0, + "content": "Figure 6: Fine-grained comparison of SURFSUP and GNS on four parts of the room scene. Our dynamics accurately model the furniture while the baseline's leak into surfaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Correspondingly GNS's fluid predictions improve, with less surface penetration (Fig. 7, top left and bottom). However, the baseline with 5000 particles still shows an order magnitude more penetration than our method (12285 vs. 490 particles). Moreover its computational cost, measured as the maximum particle graph size (# edges) during the rollout, almost doubles with 5000 surface particles (Fig. 7, top right). In contrast SURFSUP achieves near-zero surface-penetration with constant, smaller fluid graphs, overcoming the baseline's cost-performance tradeoff. We similarly analyze the room scene (Fig. 1), increasing the number of surface particles to 5K and 10K. Remarkably, SURFSUP here outperforms even a 10000-particle baseline (Ours 125.39 vs. baseline 194.98 Chamfer Surface; Supp.for details). This suggests fundamental advantages in scaling for simulation with implicit surfaces. We perform an inference time study (Supp.) showing that our method achieves speedups over GNS and the ground-truth simulator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Scaling up SURFSUP. As a final benchmark of unusual size and complexity, we assessed SURFSUP on a natural scene with complex topography (Puncak Jaya mountains, Indonesia). This scene was \\(40 \\times 27 \\times 6 \\, \\text{m}^3\\) in size (6840 m³ compared to roughly 10 m³ for previous scenes), and contained 68,600 fluid particles (order magnitude more than" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.712, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.18, + 0.892, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.091, + 0.906, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.276, + 0.895, + 0.428 + ], + "angle": 0, + "content": "Figure 7: Comparing our method vs. the baseline with more surface particles. (a) Left shows how surface penetration is affected; even as the baseline improves with more surface particles, our method maintains an order magnitude improvement (red line, 490 particles). Right shows that the graph size increases significantly with the number of surface particles; we measure the max number of edges during the rollout, which indicates memory cost. (b) Final rollout timesteps for our method and the GNS at 2000, 3000, and 4000 particles, which reflect the quantitative results." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.445, + 0.912, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.736, + 0.892, + 0.767 + ], + "angle": 0, + "content": "Figure 8: Predicted fluid rollouts from SURFSUP on different parts of a \\(40 \\times 27 \\times 6\\) m³ mountain scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "previous scenes, see Supp. for details). Fig. 8 shows the predictions of SURFSUP, which correctly handle complex features such as slopes, confluences, and valleys between the mountains. Note that while GNS would require \\( O(1M) \\) particles to represent this scene, we leave the network size for SIREN unchanged from smaller scenes. These results demonstrate the potential of our method for large-scale, realistic dynamics. Please see Supp. for rollout videos." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14232" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "Inverse Design. Finally, we evaluate SURFSUP on its potential to solve object design tasks. We show the ability of our model to design both parametrized shapes, by optimizing the shape parameters with gradients, as well as shapes represented by neural SDF models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.47, + 0.262 + ], + "angle": 0, + "content": "Parametrized Design Tasks. We consider the funnel, which is an inverted capped cone carved out of a cylinder, for which we have derived an analytical SDF. This shape is parametrized by the cylinder radius \\( R \\), height \\( H \\), cone height \\( h = H \\); and the radii \\( r_1 \\) and \\( r_2 \\) of the bottom and top circles that \"cap\" the cone, with \\( r_1 > r_2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.471, + 0.429 + ], + "angle": 0, + "content": "In our setup, we drop a block of fluid directly onto the object. We consider two tasks: designing a bowl where the task is to contain the fluid inside the object, and a funnel where the task is to concentrate the water onto a location onto the ground. We optimize \\( r_1 \\) (which we call the \"capture radius\" that captures fluid into the object) and \\( r_2 \\) (the \"filter radius\" which decides how much water to filter onto the ground). For the funnel, the desired solution is large \\( r_1 \\) and small \\( r_2 \\), and for the bowl we want \\( r_2 \\rightarrow 0 \\). If our learned model is performing well, the model should be able to discover these solutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.569 + ], + "angle": 0, + "content": "We roll out our model for 50 (bowl) and 75 (funnel) timesteps. Our reward measures the log probability of a Gaussian centered at the desired location of the fluid particles. For the bowl task, we use the bottom of the object, i.e. \\(\\mathcal{N}((0.,0.,h),\\Sigma)\\), where \\(h\\) is the object height. For the funnel we use a 2D Gaussian \\(\\mathcal{N}((0.,0.),\\Sigma)\\) and measure the reward on particles that reach the ground during training. For a given design, we roll out our model and obtain gradients, which we use to iteratively refine the design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.471, + 0.708 + ], + "angle": 0, + "content": "Qualitative results are shown in Fig. 9. Starting from a sub-optimal design with low \\( r_1 \\) and \\( r_2 > 0 \\), our model successfully converges for the bowl; similarly the funnel converges from a suboptimal to an optimal solution. This shows that over long rollouts and hundreds of message-passing steps with dense fluid-surface interactions, our model's gradients are useful for solving inverse design problems. We also find that design is robust across different initializations of \\( r_1 \\) and \\( r_2 \\) (details in Supp.)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Latent Space Design. We further examined whether SURF-SUP could directly optimize in the latent space of a DeepSDF model, trained on ShapeNet chairs. For this investigation, as a proof-of-concept we aimed to discover a novel chair that traps and contains falling water (i.e. maximizes the \"bowl reward\" above). Given a DeepSDF model \\( f_{\\theta}(p,z) \\) with latent code \\( z_0 \\), we iteratively refine the design by rolling out our simulator for 100 time-steps, computing a reward, and updating \\( z \\) via gradient-based optimization. Across multiple trials, our model converges from an suboptimal initial chair to a final design that can contain water and retains the semantics of a chair (Fig. 10)." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.092, + 0.625, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.092, + 0.745, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.092, + 0.877, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.21, + 0.626, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.324, + 0.604, + 0.334 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.21, + 0.744, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.324, + 0.728, + 0.334 + ], + "angle": 0, + "content": "Intermediate" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.21, + 0.877, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.324, + 0.837, + 0.334 + ], + "angle": 0, + "content": "Final" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.346, + 0.895, + 0.438 + ], + "angle": 0, + "content": "Figure 9: Design optimization; top is the funnel task, bottom the bowl task. For each frame, the design is shown near the top of the container and a timestep from the fluid rollout is shown below. Initial, intermediate, and final designs are shown; note that the reward increases through optimization and final designs approximate a funnel and bowl well." + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.45, + 0.585, + 0.465 + ], + "angle": 0, + "content": "Initial Design" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.473, + 0.593, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.452, + 0.719, + 0.465 + ], + "angle": 0, + "content": "Final Design" + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.473, + 0.723, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.452, + 0.88, + 0.464 + ], + "angle": 0, + "content": "Reward Convergence" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.476, + 0.89, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.616 + ], + "angle": 0, + "content": "Figure 10: We differentiate through SURFSUP to directly optimize (in the latent space of a DeepSDF) a chair that holds water dropped onto it." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.62, + 0.635 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In this paper, we have taken a first step towards integrating implicit 3D representations into the dynamical simulation of physical models. SURFSUP realizes this approach by incorporating the implicit representation of solid surfaces (via SDFs) into a graph neural network fluid simulator. As such, SURFSUP is able to accurately model how fluids interact with surfaces, including highly complex surfaces unseen in training. We find that promoting this type of OOD generalization is helpful in both forward and inverse contexts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements: We thank Huy Ha for help with visualizations of fluid rollouts. We thank Jan Bender for helpful feedback on SPlisHSPlasH, and Kelsey Allen, Tobias Pfaff, and Alvaro Sanchez-Gonzalez for advice on graph network simulators. We thank Max Helman for early setup work on SPlisHSPlasH. This research is partially supported by the NSF STC for Learning the Earth with Artificial Intelligence and Physics, and the NSF NRI Award #1925157. AM is supported by the NSF fellowship." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14233" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas J Guibas. Learning representations and generative models for 3d point clouds. arXiv preprint arXiv:1707.02392, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Kelsey R Allen, Tatiana Lopez-Guevara, Kimberly Stachenfeld, Alvaro Sanchez-Gonzalez, Peter Battaglia, Jessica Hamrick, and Tobias Pfaff. Physical design using differentiable learned simulators. arXiv preprint arXiv:2202.00728, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interaction networks for learning about objects, relations and physics. Advances in neural information processing systems, 29, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Jan Bender and Dan Koschier. Divergence-free smoothed particle hydrodynamics. In Proceedings of the 14th ACM SIGGRAPH / Eurographics Symposium on Computer Animation, SCA '15, page 147-155, New York, NY, USA, 2015. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[5] Jan Bender and Dan Koschier. Divergence-free smoothed particle hydrodynamics. In Proceedings of the 14th ACM SIGGRAPH / Eurographics Symposium on Computer Animation, SCA '15, page 147-155, New York, NY, USA, 2015. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[6] Siddhartha Chaudhuri, Daniel Ritchie, Jiajun Wu, Kai Xu, and Hao Zhang. Learning generative models of 3d structures. In Computer Graphics Forum, volume 39, pages 643-666. Wiley Online Library, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[8] James D Foley, Foley Dan Van, Andries Van Dam, Steven K Feiner, and John F Hughes. Computer graphics: principles and practice, volume 12110. Addison-Wesley Professional, 1996." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[9] Christoph Gissler, Andreas Peer, Stefan Band, Jan Bender, and Matthias Teschner. Interlinked sph pressure solvers for strong fluid-rigid coupling. ACM Transactions on Graphics (TOG), 38(1):1-13, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Zekun Hao, Hadar Averbuch-Elor, Noah Snavely, and Serge Belongie. Dualsdf: Semantic shape manipulation using a two-level representation. pages 7631–7641, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[11] Dan Koschier and Jan Bender. Density maps for improved sph boundary handling. In Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 1-10, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[12] Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. In ICLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Yunzhu Li, Jiajun Wu, Jun-Yan Zhu, Joshua B Tenenbaum, Antonio Torralba, and Russ Tedrake. Propagation networks for model-based control under partial observation. In 2019 International Conference on Robotics and Automation (ICRA), pages 1205-1211. IEEE, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Zhijian Liu, William T Freeman, Joshua B Tenenbaum, and Jiajun Wu. Physical primitive decomposition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-19, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[16] Andreas Mayr, Sebastian Lehner, Arno Mayrhofer, Christoph Kloss, Sepp Hochreiter, and Johannes Brandstetter. Boundary graph neural networks for 3d simulations. arXiv preprint arXiv:2106.11299, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[17] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[18] Mariem Mezghanni, Théo Bodrito, Malika Boulkenafed, and Maks Ovsjanikov. Physical simulation layer for accurate 3d modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13514-13523, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[19] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[20] J. J. Monaghan. Smoothed particle hydrodynamics. Annual Review of Astronomy and Astrophysics, 30:543-574, Jan. 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Damian Mrowca, Chengxu Zhuang, Elias Wang, Nick Haber, Li F Fei-Fei, Josh Tenenbaum, and Daniel L Yamins. Flexible neural representation for physics prediction. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[22] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[23] Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh-based simulation with graph networks. In International Conference on Learning Representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[24] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 22468-22478. Curran Associates, Inc., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[25] Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In International Conference on Machine Learning, pages 8459-8468. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20(1):61-80, 2008." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14234" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[27] C. Schenck and D. Fox. Spnets: Differentiable fluid dynamics for deep neural networks. In Proceedings of the Second Conference on Robot Learning (CoRL), Zurich, Switzerland, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[28] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[29] Vincent Sitzmann, Eric R. Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In Proc. Neurips, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[30] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Proc. NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.291, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[31] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In Computer Vision and Pattern Recognition (CVPR), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.348, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[32] Benjamin Ummenhofer, Lukas Prantl, Nils Thuerey, and Vladlen Koltun. Lagrangian fluid simulation with continuous convolutions. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.468, + 0.473 + ], + "angle": 0, + "content": "[33] Subeesh Vasu, Nicolas Talabot, Artem Lukoianov, Pierre Baqué, Jonathan Donier, and Pascal Fua. Hybridsdf: Combining deep implicit shapes and geometric primitives for 3d shape representation and manipulation. In International Conference on 3D Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.475, + 0.468, + 0.558 + ], + "angle": 0, + "content": "[34] Zhirong Wu, S. Song, A. Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and J. Xiao. 3d shapenets: A deep representation for volumetric shapes. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, Los Alamitos, CA, USA, jun 2015. IEEE Computer Society." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14235" + } + ] +] \ No newline at end of file diff --git a/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/e47c0301-5f63-4201-a2ee-b44e21fee063_origin.pdf b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/e47c0301-5f63-4201-a2ee-b44e21fee063_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e2b9972cee4ad0a660223751fab216345d94c195 --- /dev/null +++ b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/e47c0301-5f63-4201-a2ee-b44e21fee063_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1368ae7212d8450f17db834a3307f9e1c7eb7d5f17c939551efeaedd4a17c3c9 +size 10069938 diff --git a/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/full.md b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6fc6db7c44b49129c18e32babe1a132947612eb7 --- /dev/null +++ b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/full.md @@ -0,0 +1,270 @@ +# SURFSUP: Learning Fluid Simulation for Novel Surfaces + +Arjun Mani\*, Ishaan Preetam Chandratreya\*, Elliot Creager\*3, Carl Vondrick1, Richard Zemel1 +1 Columbia University 2 University of Toronto 3 Vector Institute +surfsup.cs.columbia.edu + +![](images/39ac34228a60edba9ac06071898a4fe9d2f9da8e50af07105250e807be33b435.jpg) +Figure 1: We introduce SURFSUP, which learns to predict how fluid particles interact with novel, complex 3D surfaces. + +# Abstract + +Modeling the mechanics of fluid in complex scenes is vital to applications in design, graphics, and robotics. Learning-based methods provide fast and differentiable fluid simulators, however most prior work is unable to accurately model how fluids interact with genuinely novel surfaces not seen during training. We introduce SURFSUP, a framework that represents objects implicitly using signed distance functions (SDFs), rather than an explicit representation of meshes or particles. This continuous representation of geometry enables more accurate simulation of fluid-object interactions over long time periods while simultaneously making computation more efficient. Moreover, SURFSUP trained on simple shape primitives generalizes considerably out-of-distribution, even to complex real-world scenes and objects. Finally, we show we can invert our model to design simple objects to manipulate fluid flow. + +# 1. Introduction + +Across engineering and science, the simulation of fluid dynamics has become an invaluable tool, and it will be a crucial component for building visual systems that are capable of understanding and interacting with real-world environments. Recently, a new class of simulators has emerged that learn the dynamics of physical systems from data [13, 32, 25]. These simulators offer several advantages over classical simulators by increasing speed, reducing reliance on hand-crafted dynamics models, and providing differentiable rollouts for solving inverse problems. Recent approaches based on graph neural networks (GNNs) have in particular shown impressive accuracy and generalization on a wide range of fluids represented as particle or mesh-based systems [25, 23]. + +In a wide variety of applications areas, fluid simulation must properly handle interaction with solid structures to attain practical relevance. Many problems involving + +fluid simulation are fundamentally about object design; for example, designing aerodynamic shapes for cars and airplanes. In robotics too, an intelligent robot would have to understand fluid-surface interactions; for example, to carry a mug of hot coffee to a person without spilling the liquid from the container. In both cases, the need to accurately model how fluids interact with novel solid surfaces1—with new shapes, configurations, orientations, and so on—motivates an challenging out-of-distribution (OOD) generalization subproblem for fluid simulation. + +In this paper, we tackle this problem by introducing Surface implicit sUbstitution for particles (SURFSUP), a framework for simulating the interactions of fluids and surfaces by representing surfaces using implicit representations [19]. Current GNN-based approaches represent surfaces explicitly, discretizing them into particles and augmenting the graph accordingly [25]. Implicit representations, in particular signed distance functions which represent surfaces as zero-level sets of functions, offer several compelling advantages. They can smoothly represent continuous surfaces, provide rich geometric information (e.g. whether a point is in the interior/exterior of the object), and can scale easily to large objects and scenes. Due to these properties, many state-of-the-art generative models for 3D shapes learn shapes with signed distance functions. We argue that these properties are useful not just for statics but also dynamics: in order to accurately model fluid-surface interactions for potentially complex surfaces, smooth and informative representations of local surface geometry are needed. In turn, we posit that these locally informative representations will enable fluid simulators to appropriately model interaction with never-before-seen surfaces. + +Concretely, we propose an approach that integrates signed distance representations of surfaces [22, 19] with graph neural network models [26, 25]. Our architecture follows the encode-process-decode paradigm [25], where the particles in the physical system at time $T$ are encoded into a graph, followed by several message-passing steps, and finally a decoder predicts the dynamics (accelerations). We represent the surface using its signed distance function, and exploit geometric properties of SDFs to model the distance and orientation of fluid particles with respect to the surface. + +Experiments show that our approach can model fluid-surface interactions over long rollouts with a high degree of accuracy and visual realism. SURFSUP can model surfaces far beyond its training domain; when trained on simple primitive shapes (e.g. spheres, cones, toruses), we generalize zero-shot to complex real-world objects and scenes represented using neural implicit generative models (Fig. 1). We compare to a method which represents the surface as particles inside the graph network simulator, and show that our model achieves better generalization and efficiency, and + +can better handle complex surface geometry. Finally, we demonstrate the potential of SURFSUP to solve inverse problems across two object design tasks. Overall, our paper bridges learned physical models for dynamics and 3D implicit representations, and takes a step towards AI models which can understand and help design the physical world. + +# 2. Related Work + +Learning Differentiable Simulators. Using machine learning for predicting and modeling complex physical systems is a rapidly growing area of research. Learned physics simulators have been shown to accurately simulate colliding rigid objects, deformable objects, fluids, and other complex systems [13, 3, 12, 32, 25, 21]. + +Recent progress in learned fluid simulation has been driven by the Lagrangian viewpoint, which tracks the motion of fluid particles over time. Graph neural networks (GNNs) have in particular emerged as effective forward models for fluid simulation, and have demonstrated better stability and generalization compared to grid-based, convolutional approaches. [13, 14, 27, 25]. Sanchez-Gonzalez et al. [25] introduced a GNN architecture which can accurately and generalizably model fluid behavior over long rollouts; most subsequent work builds off their approach [23, 16]. However, most of these works generally consider simple surfaces; e.g. only cube-shaped containers for fluids. There has been little work focused on fluid-surface interactions; Mayr et al. [16] introduce an approach specifically for triangularized meshes. All these approaches use particles or meshes to represent surfaces; we will argue that implicit representations enable learning more accurate, generalizable physical dynamics. + +Some works have explored the possibility of using learned fluid simulators for solving inverse problems [27, 12, 2]. Allen et al. [2] showed that graph network simulators can be used to solve design tasks by rolling out fluid trajectories and using gradient-based optimization. In 3D, they show that a 2D plane with a learned height field can be used to split a stream of fluid to fall into several locations on the ground. SDFs have the advantage of being able to learn geometry in a more smooth and unconstrained way, which could help solve inverse problems. + +3D Implicit Representations. Implicit representations have become state-of-the-art for generative models of 3D objects [17, 22, 6, 29, 30]. These models represent the surface as a function $F(x)$ . There are several choices for $F$ . Park et al. [22] show that signed distance functions $F(x)$ are particularly effective and informative representations. Compared to point clouds [1, 28] or voxels [7, 34], models based on SDF representations can smoothly represent complex geometry, are compact and efficient, and can be learned effectively [29, 6]. Follow-up work has studied learning "dual" generative models that jointly learn an + +uncontrained representation and one composed of primitive shapes [10, 33]. + +Implicit Representations for Physical Dynamics. Some classical particle-based fluid simulators based on the Smoothed Particle Hydrodynamics (SPH) formalism [20] have transitioned from particle-based to implicit surface representations. [11, 9, 4]. In these simulators implicit surfaces allow for smoother fluid motion with less artifacts [11]. For general-purpose learned simulators (GNNs), there is little prior work using implicit rather than particle-based representations of surfaces. + +Another line of work combines implicit generative models with physical simulators [18, 24]. Mezghanni et al. [18] generate more physically stable SDF shapes by backpropagating gradients through a physical simulation layer. Their simulation mainly considers Newtonian dynamics. Our contribution is an effective, generalizable forward model for fluids that represents surfaces accurately and efficiently, and can also be used for solving inverse problems. + +# 3. Method + +# 3.1. Setup: Particle-Based Fluid Simulation + +We address the problem of learning the physical dynamics of fluids interacting with surfaces. Given the state of a physical system $X^0$ and some initial conditions (e.g. gravity, other external forces), rolling out a fluid simulation over $T$ steps produces a sequence of states $X^{(1)},\dots,X^{(T)}$ . In this paper, we focus on particle-based simulation; the state of the system at time $t$ is represented by a set of particles $X^{(t)} = (p_1^{(t)},\dots,p_n^{(t)})$ , where $p_i$ is the position of particle $i$ in 3D space. A rollout over $T$ steps is produced by repeatedly applying a forward model $S_{\theta}:X^{(t)}\to X^{(t + 1)}$ , which takes in the particle positions at time $t$ (and optionally a short history $t - 1,\ldots$ ) and produces the next-step particle positions. While the forward model can be used to directly predict $X^{(t + 1)}$ , most methods instead predict intermediate values $Y$ , typically next-state accelerations for each particle, which can then be numerically integrated (e.g. Euler integration) to produce $X^{(t + 1)}$ ; concretely, our task is to learn the parameters $\theta$ of this forward model $f_{\theta}:X^{(t)}\to Y$ . + +# 3.2. Graph Networks for Particle-Based Simulation + +Graph neural networks lend themselves naturally to particle-based simulation, since they compute pairwise interactions between nodes and aggregate these interactions. Graph Network-based Simulators (GNS), a model architecture introduced by Sanchez-Gonzalez et al. [25], have in particular shown state-of-the-art performance on simulating fluids over long rollouts. These methods follow an "encode-process-decode" paradigm. First, the state $X^{(t)}$ is encoded into a graph, where nodes $v_{i}$ are instantiated using a set of features (e.g. previous velocities, parti + +cle type, etc.); nodes $v_{i}, v_{j}$ are then connected with an edge $e_{ij}$ if their distance is less than a connectivity radius $\epsilon$ . Each node and edge is associated with an embedding. The "processor" stage then passes the encoded graph through $M$ message-passing steps. Each message-passing step takes the current graph $\{(v_{i}, v_{j}, e_{ij})\}^{(m)}$ . An edge module first computes the updated edge embeddings $e_{ij}^{(m+1)} = f_{e}(v_{i}^{(m)}, v_{j}^{(m)}, e_{ij}^{(m)})$ ; then the edge embeddings are aggregated to compute updated node embeddings $v_{i}^{(m+1)} = f_{n}(v_{i}^{(m)}, \sum_{j \in N(i)} e_{ij}^{(m)})$ , where $N(i)$ are the set of neighbors of node $i$ . $f_{e}$ and $f_{n}$ are generally MLPs with 2-3 hidden layers, and the weights are shared across edges/nodes (although differing across message passing steps). Finally, the "decoder" step applies an MLP $f_{D}$ to the final node embeddings to predict the accelerations $\{y_{i}\}$ . + +In our problem, we further assume that there exists some rigid surface $S$ that interacts with the fluid throughout the simulation. The GNS approach handles surfaces by discretizing the surface into particles (i.e. a point cloud) and adding these particles $\{s_i\}$ to the graph (denoting them by a different particle type, and masking during position updates) [25]. Follow-up work for mesh-based simulation handles surfaces similarly by representing the surface as an irregular mesh and adding nodes for mesh vertices [23]. + +# 3.3. Modeling Fluid-Surface Interactions with SDFs + +We propose to model the interactions of fluids and surfaces with signed distance function (SDF) representations of surfaces. An SDF represents a watertight surface as a function $F(p) = s$ , which takes a spatial point $p \in \mathbb{R}^3$ and outputs its distance $s$ to the closest corresponding point on + +![](images/a76562248cb915f5d8fc7915cb5b0f084083b8fe295f0466c397ec3f7aa34009.jpg) +Figure 2: Rigid surfaces can be represented compactly and implicitly as the level set of a SDF: $\{x|F(x) = 0\}$ . Evaluating $F$ and its derivatives at particle position $p$ provides useful information for simulating the interaction of that particle with the surface. + +the surface. The function is "signed", such that it takes on positive values for points outside the surface and negative values in the surface interior. The surface of the object is implicitly represented as the zero-level set $F(p) = 0$ . Explicit representations of the surface can be constructed by applying meshing algorithms such as Marching Cubes [15]. + +Almost every surface can be described by an SDF. For several primitive shapes SDFs can be written down analytically (e.g. spheres, cones, cylinders); boolean operations on these shapes such as unions and intersections give rise to more complex shapes [31]. SDFs can also be represented discretely as voxel grids; or more recently for finer geometry, as neural SDFs $F_{\theta}(p)$ which learn to accurately approximate $F$ from mesh-based supervision. + +# 3.3.1 Locality and Geometric Information for Fluid-Surface Interactions + +A key insight of our method is that fluid-surface interactions are inherently about local rather than global surface geometry. The interaction of a fluid with a surface depends on a small local region of the surface, over which the geometry can have limited variation. This suggests that an effective representation of local surface geometry, trained on a variety of object shapes, can generalize OOD to shapes that may look different globally. By contrast, particles are not an effective representation of surface geometry; for any moderately complex surface, discretization leads to artifacts such as non-smooth fluid motion and even penetration of the surface boundary (we confirm this empirically). + +By contrast, we argue that SDFs are a smooth, richly informative representation of local surface geometry for learning generalizable fluid-surface interactions. Our key insight is shown in Figure 2. Given a solid surface represented as an SDF $F(p)$ , we can use properties of the SDF to inform the behavior of a fluid particle as it reaches the surface. The value of the SDF $F(p)$ gives the distance from the particle at point $p$ to the closest point on the surface $p'$ . Moreover, the SDF gradient $\nabla F(p)$ provides the unit-norm displacement vector from the particle to its closest surface point $\frac{(p - p')}{||p - p'||}$ . This can be shown by observing that the SDF decreases most quickly on the line from $p$ to $p'$ . Thus, the SDF and its gradient provide the distance and orientation of the fluid particle w.r.t. to its closest surface point; as the particle approaches the surface, its acceleration will more closely approach $\nabla F(p)$ and it will tend to rebound. Moreover, since the gradient varies smoothly because of the continuity of $F$ , these features do not suffer from artifacts. + +Higher order derivatives of the SDF can also be used to gain geometric information. For example, we could march to the surface by computing $p - F(x)\nabla F(x)$ . At the surface, the Hessian $\nabla^2 F(x)$ is the shape operator; its eigenvalues provide the principal directions of surface curvature + +![](images/6187da8e669f4f3de12fe83663c39f6f7f68d55627cc19ade6c9b04a1092a8c9.jpg) +Figure 3: Given fluid particles and the SDF, our method augments the constructed graph with features from the SDF. We then apply message-passing and predict dynamics, weighting near-surface errors more highly in our loss. + +$\kappa_{1}$ and $\kappa_{2}$ , and the mean curvature is $H = \frac{1}{2} \mathrm{Tr}(\nabla^2 F)$ . We can also estimate the surface curvature via finite differences by slightly perturbing the gradient vector on each axis and computing the SDF (Fig. 2). In this paper, we only use the SDF and its gradient as features without curvature information; this can be seen as an effective theory where the relevant "local region" of the surface is a single closest point on the surface. As we show, modeling surfaces using SDF representations helps our fluid simulator generalize beyond the available training data by modeling fluid-surface interactions for unseen surfaces. + +# 3.4. Model + +Having developed our ideas about SDFs and modeling fluid-surface interactions, we now describe our approach, which we call Surface implicit sUbstitution for particles (SURFSUP). We learn a forward predictor $f_{\theta}:X^{(t)},F\to Y$ , which takes as input a set of particle positions and a surface implicitly parametrized by an SDF $F$ . We adopt the encode-process-decode paradigm for our model, but make two key changes. In the encoder step, for each fluid particle $p_i$ we add the value of the SDF $F(p_i)$ and its gradient $\nabla F(p_i)$ to the initial node representations, such that the new representation is $v_{i}^{\prime (0)} = [v_{i}^{(0)},F(p_{i}),\nabla F(p_{i})]$ . This is the way that the surface is encoded into the model; subsequently no particles representing the surface are added to the graph. The updated node embeddings are then passed through the "processor" with several message-passing steps and then decoded to predict the accelerations as standard. The other modification we make is to weight particles close to the surface higher during training when predicting dynamics. This has the effect of teaching the model to prioritize modeling fluid-surface interactions accurately. Since the model is predicting next-step accelerations, our loss becomes: + +$$ +L = \sum_ {F (p _ {i}) > \alpha} \left\| a _ {i} ^ {(t)} - \hat {a} _ {i} ^ {(t)} \right\| ^ {2} + \sum_ {F (p _ {i}) \leq \alpha} \lambda \left\| a _ {i} ^ {(t)} - \hat {a} _ {i} ^ {(t)} \right\| ^ {2} \tag {1} +$$ + +In our experiments, we set $\alpha$ based on the "neighborhood radius" of the classical simulator we train against, and set $\lambda = 5$ . We found that this weighting helped the model learn surface interactions more quickly without sacrificing overall accuracy on fluid prediction. Note that the SDF allows loss weighting to be implemented in a very natural way, by simple thresholding. Our approach is shown in Fig. 3. + +# 3.5. Other Advantages of SDF Representations + +We have argued for SDFs from the perspective of achieving smooth, information-rich representations of local surface geometry. Here we mention other advantages of SDFs, each of which we demonstrate in our experimental results. + +Efficient message passing in the GNN. By modeling surfaces implicitly, we reduce the number of unique particles needed to compose the GNN's state. This allows us to perform more lightweight and computationally efficient simulation, as well as run the model for more message-passing steps. This advantage becomes especially clear for large objects/scenes or objects with complex surface geometry. + +Shape Design with SDFs. SDFs are a useful paradigm for shape design and assembly. This is because compositions of shapes can be generated by union, intersection, and difference operations; e.g. the union of two SDFs $F_{1}$ and $F_{2}$ is $\min(F_{1}, F_{2})$ . The construction of shapes using primitives and these operations is known as Constructive Solid Geometry (CSG), and is implemented in many CAD packages for object design [8]. Moreover the recent emergence of neural SDF models as state-of-the-art 3D shape representations, offers another compelling avenue for shape design with end-to-end learning. These suggest that good forward models with SDFs could be leveraged for design problems. + +# 4. Experiments + +# 4.1. Fluid Domains + +We evaluate our approach on complex prediction tasks involving fluids interacting with a variety of surfaces. Our main contribution here is a 3D fluid dataset consisting of water interacting with a range of primitive shapes, across shape parameters, positions, and orientations. For the ground-truth simulator we use SPlisHSPlasH [5], a 3D-particle-based fluid simulator based on the SPH technique. + +PrimShapes Dataset. We train SURFSUP on a 3D water dataset, where the fluid interacts with various shapes. We initialize a block of fluid directly above the shape, then let it fall due to gravity. The entire scene takes place in a $1 \times 1 \times 2$ m container. A random rotation is applied to + +the base shape which sits on the bottom of the container, and the initial position of the shape is also randomized. We extract 800 time-steps from the ground-truth fluid rollout (with $\Delta t = 0.005$ ); this corresponds to 4000 steps for the classical simulator, which requires very small time-steps for stability [25]. + +We use five "primitive" shapes in our dataset: spheres, boxes, cylinders, cones, and toruses. These shapes have SDFs with analytical forms. (E.g. the simplest form is the origin-centered, radius- $r$ sphere with SDF $F(p) = ||p||_2 - r$ ; the others tend to be more complex). We emphasize that SURFSUP is trained only on simulations containing these primitive shapes and generalizes zero-shot to complex objects and scenes. Our key insight is that accurate fluid-surface simulation depends on understanding local surface geometry, of which a significant diversity is expressed by these shapes (like curvature, holes, differently-sloped faces, corners). Leveraging this and the properties of SDFs, SURFSUP can generalize to interactions with surfaces of entirely novel global geometry. Our training and test set consists of 1000 and 100 simulations respectively. See Supp. Fig 6 for example initializations. + +# 4.1.1 Generalization Test-Sets + +We create a range of qualitative and quantitative evaluation scenarios to test our model's generalization ability. The first three below assess generalization to novel analytical SDFs (e.g. unions $\min(f_1, f_2)$ or differences $\max(f_1, -f_2)$ ), and the last to real-world shapes represented as neural SDFs $f_\theta$ . + +Prim-OOD (60 simulations). This contains the same shapes with OOD shape parameters from the training set (e.g. cylinder with higher radius $r$ and lower height $h$ ). + +Prim-Unions (40). This test set includes all possible pairwise unions which 'merge' two of the primitive shapes. See Supp. for examples. + +Funnels (12). We examine the model's ability to handle analytical "differences" of shapes by creating funnels, formed by carving an inverted cone out of a cylinder. + +'Complex' Scenes (7). As the key, challenging test of our model's generalization, we simulate fluid interacting with several complex real-world objects. To do so, we leverage neural SDFs of shapes, which can represent complex shapes with high accuracy and fidelity (Sec. 3.3). Instead of an analytical SDF $f(p)$ , we use a trained neural SDF $f_{\theta}(p)$ to generate our SDF features during our forward pass. To test our model's ability to generalize zero-shot to complex objects/scenes, we train a single SIREN model [30] for 7 fine-grained objects (including the coral, lion, and room scenes shown in Fig. 1, Fig. 4). For each object we create a fluid scene using the ground-truth simulator, and evaluate SURFSUP on this small Complex-Scenes dataset. To represent the geometry at an appropriate scale, these objects are + +
Testing SetChamferChamfer SurfaceNumber InsideMean SDF Inside
SURFSUPGNS [25]SURFSUPGNSSURFSUPGNSSURFSUPGNS
Primitives0.02970.02854.7045.45102000-1.04e-11-2.48e-05
Primitives-OOD0.03210.034310.95118.44532433006-1.65e-06-0.0006
Primitives-Unions0.03800.034824.55326.3071000063000-4.46e-05-0.001
Funnels0.03620.05219.99013.66327204-2.51e-09-0.001
Complex-Scenes0.06380.265850.521276.17640513257722-0.0011-0.026
+ +Table 1: Results on in-distribution (Primitives) and OOD test sets, comparing SURFSUP to GNS [25] which represents the surface as particles. + +scaled 2x in the fluid scene compared to previous datasets. To test our model on families of shapes, which gives us the capacity to solve inverse problems, we train a variational DeepSDF model [22] for the ShapeNet chair and bowl categories. See Supp. for details on these scenes and models. + +# 4.2. Training Details + +For all experiments, we train our model on the PrimShapes dataset. The training data is pairs $X^{(t)}, X^{(t+1)}$ from the ground-truth simulator. We use a graph network (GN) architecture with 10 message-passing steps and train with Adam for 2M steps (see Supp. for details). + +# 4.3. Metrics and Baseline + +Although we train on next-step prediction, we measure performance on full rollouts. We use mean Chamfer distance between the predicted particles $P$ and ground-truth $G$ to measure overall rollout agreement (averaging across timesteps). Compared to MSE, Chamfer distance is permutation-invariant and therefore a better measure of similarity between fluid flows [25]. This metric, however, can be noisy since it averages over long rollouts on chaotic systems with many particles. Since we focus on the quality of fluid-surface interactions, we also compute a "Chamfer surface" metric that focuses on prediction errors near the surface. For each time-step, we consider the predicted point cloud only near the surface (where $F(p) < \alpha$ for threshold $\alpha$ ) and measure the Chamfer distance of $P_{\alpha}$ to $G$ , summing over particles; we do this symmetrically for $G$ . Intuitively this focuses on discrepancies between $P$ and $G$ for fluid motion near the surface. See Supp. for further details. + +Since surface penetration is a serious failure mode, we also measure the number of particles inside the surface through the rollout (summing over the number of timesteps). Additionally, we compute the average SDF value for particles inside the surface, indicating the extent of surface penetration (since SDFs become increasingly negative deeper into the surface interior). + +Baseline. Our main comparison is to the graph network simulator (GNS) proposed by Sanchez Gonzalez et al. [25]. Their approach represents surfaces explicitly through discretization into particles, which are added to the graph and participate in the message-passing. To implement this base + +line, we convert all the SDFs in our datasets to meshes using Marching Cubes [15]. From this mesh we uniformly sample 2000 surface particles (comparable to the number of fluid particles in our simulations) and add them to the graph. In all our experiments we compare our implicit SURFSUP approach to this baseline with explicit surface particles. + +While the baseline method could in principle be extended to include local surface information (e.g. using surface normal estimates for each particle as additional features, or mesh connectivity), we omit such an extension here. This is because any explicit representation of large surfaces quickly becomes computationally infeasible, especially for large objects/scenes or complex geometry. Whereas the simulation quality of SURFSUP scales with the capacity of the SDF network $f_{\theta}$ , the baseline quality scales with the number of particles. We show in Sec 5 that SURFSUP still outperforms (with lower computational cost and inference time) a baseline with more particles. + +# 5. Results + +SURFSUP can accurately predict fluid behavior over long rollouts, interacting with a range of complex surfaces (Fig. 4). Across the different analytical primitive shapes in our test sets and different orientations, we can generate accurate and visually realistic predictions of fluid behavior. See the Supp. Material for videos of these, and all, rollouts in the paper. Our key result is the following: we find that SURFSUP can perform effective simulation with surfaces that are significantly out-of-distribution, and even generalize to complex real-world shapes and scenes. This demonstrates that SURFSUP can leverage the rich geometric information provided by the SDF for generalizing to novel surfaces expressed as neural SDFs. + +Quantitative Results and Comparison to Baseline. We evaluate SURFSUP on rollouts from ID and OOD test sets of shapes with analytical SDFs, and on complex scenes, with comparison to the GNS baseline. Results are shown in Table 1. Across all test sets, our model's fluid predictions display significantly lower surface penetration than the baseline (e.g. 324 vs. 33006 particles on average for the primitives+OOD test set). Moreover, for those few particles which have penetrated the surface, the average SDF value for our method is nearly zero; this implies that any + +![](images/740a1f905fe8766f130c663479ba46e507a6d88bed38d43b2288966d2e3941d4.jpg) +Figure 4: Although SURFSUP is only trained on primitive shapes, our model generalizes to complex and real-world shapes without any additional training. We visualize our predictions (middle) for two scenes. In comparison, the baseline (top) often generates unrealistic predictions, such as having water penetrate solid objects. In many cases, the predictions from the baseline cause the water to disappear because it falls through the floor (top right). + +![](images/94c75b747b8057e4de536fbe705e2fc8cf51ad5f836569d506b66cf7f7f5489f.jpg) + +particle that penetrates the boundary $(F(p) < 0)$ is quickly projected outward to the region $F(p) > 0$ . On Chamfer distance near the surface, which measures overall accuracy of fluid-surface interactions beyond surface penetration, our method consistently improves over the baseline. Overall Chamfer distance between the two models is generally comparable. This is unsurprising given that small differences in particle positions along long time-horizons, potentially far from the surface, can accumulate error. (Sec. 4.3). See Supp. Fig. 5 for rollouts on rows 1-4 of Table 1, indicating the stability and accuracy of our method's dynamics. + +Generalizing to Complex Scenes We evaluate SURFSUP on chairs and bowls from Shapenet, and fine-grained objects and scenes in the real world (See Fig. 1, Fig. 4, Fig. 5, and Fig. 8, also videos in Supp.). Remarkably SURFSUP trained only with primitive shapes generalizes zero-shot to complex shapes/scenes represented as neural SDFs. This is notable because our chosen real-world objects differ widely (see the variation in Fig. 4 and Fig. 5), and second, neural SDFs are imperfect compared to analytical SDFs (e.g. not fulfilling the property $||\nabla f|| = 1$ ). Despite these challenges, SURFSUP produces rollouts with accurate and realistic fluid dynamics, capturing fine-grained behavior near complex surfaces. Our method also avoids unintuitive artifacts, such as surface penetration, seen in the GNS baseline. + +Quantitative results (Table 1, row 'Complex-Scenes') indicate that our method outperforms the baseline most significantly on complex real-world scenes. These results elucidate key generalization advantages of our method in handling surfaces with complex geometry and large scenes at scale. For the coral scene (Fig. 1), our method's fluid rolls accurately capture and percolate through the fine tentacle structure of the coral. The GNS baseline does not display this ability (Supp. Fig 1). Our method also scales accurately to large scenes such as the room scene in Fig. 1. By contrast, the baseline's predictions consistently exhibit penetration of the furniture surfaces in the scene due to insufficient resolution (see Fig. 6). Recall (Sec. 4.3) that our scaling law is the capacity of the neural SDF network $f_{\theta}$ instead of the number of surface particles. + +Efficiency of Simulation. We study comparing SURFSUP as we increase the number of surface particles for the GNS baseline (which in our experiments is afforded 2000 surface particles). Note that there is a tradeoff between performance and computational cost (memory, time) for the baseline; more particles improves surface resolution while also increasing graph sizes. We examine a scene in the Primitives-Unions test set (union of a sphere and box), for which GNS's predictions display significant surface penetration, and increase surface particles from 2000 to 5000. + +![](images/d6d6c3ac613d53f68d3380410ae407fae369cd4d5714ed082d23af32a23e6110.jpg) +Figure 5: SURFSUP predictions on ShapeNet objects. + +![](images/e12890e6b577710cd29589628ed05516253bc683865d57411114968c4c9d8302.jpg) +Figure 7: Comparing our method vs. the baseline with more surface particles. (a) Left shows how surface penetration is affected; even as the baseline improves with more surface particles, our method maintains an order magnitude improvement (red line, 490 particles). Right shows that the graph size increases significantly with the number of surface particles; we measure the max number of edges during the rollout, which indicates memory cost. (b) Final rollout timesteps for our method and the GNS at 2000, 3000, and 4000 particles, which reflect the quantitative results. + +Correspondingly GNS's fluid predictions improve, with less surface penetration (Fig. 7, top left and bottom). However, the baseline with 5000 particles still shows an order magnitude more penetration than our method (12285 vs. 490 particles). Moreover its computational cost, measured as the maximum particle graph size (# edges) during the rollout, almost doubles with 5000 surface particles (Fig. 7, top right). In contrast SURFSUP achieves near-zero surface-penetration with constant, smaller fluid graphs, overcoming the baseline's cost-performance tradeoff. We similarly analyze the room scene (Fig. 1), increasing the number of surface particles to 5K and 10K. Remarkably, SURFSUP here outperforms even a 10000-particle baseline (Ours 125.39 vs. baseline 194.98 Chamfer Surface; Supp.for details). This suggests fundamental advantages in scaling for simulation with implicit surfaces. We perform an inference time study (Supp.) showing that our method achieves speedups over GNS and the ground-truth simulator. + +Scaling up SURFSUP. As a final benchmark of unusual size and complexity, we assessed SURFSUP on a natural scene with complex topography (Puncak Jaya mountains, Indonesia). This scene was $40 \times 27 \times 6 \, \text{m}^3$ in size (6840 m³ compared to roughly 10 m³ for previous scenes), and contained 68,600 fluid particles (order magnitude more than + +![](images/c371af9480e01e61b8acaa4a77163a3092fda8e9e707452ba9ecd97762c3c2a1.jpg) + +![](images/78f72b7d0a30b416c34797665abf16019a263d25ada706c90a379143130ecf6e.jpg) + +![](images/552256be592444af0fad124718e80260b43e23781b885f9b38b4491cf671bc07.jpg) + +![](images/6ba7745f9abf574b5124c008e99dac88e100ebeb2a72516a9db68801b4434cd5.jpg) +Figure 6: Fine-grained comparison of SURFSUP and GNS on four parts of the room scene. Our dynamics accurately model the furniture while the baseline's leak into surfaces. +Figure 8: Predicted fluid rollouts from SURFSUP on different parts of a $40 \times 27 \times 6$ m³ mountain scene. + +previous scenes, see Supp. for details). Fig. 8 shows the predictions of SURFSUP, which correctly handle complex features such as slopes, confluences, and valleys between the mountains. Note that while GNS would require $O(1M)$ particles to represent this scene, we leave the network size for SIREN unchanged from smaller scenes. These results demonstrate the potential of our method for large-scale, realistic dynamics. Please see Supp. for rollout videos. + +Inverse Design. Finally, we evaluate SURFSUP on its potential to solve object design tasks. We show the ability of our model to design both parametrized shapes, by optimizing the shape parameters with gradients, as well as shapes represented by neural SDF models. + +Parametrized Design Tasks. We consider the funnel, which is an inverted capped cone carved out of a cylinder, for which we have derived an analytical SDF. This shape is parametrized by the cylinder radius $R$ , height $H$ , cone height $h = H$ ; and the radii $r_1$ and $r_2$ of the bottom and top circles that "cap" the cone, with $r_1 > r_2$ . + +In our setup, we drop a block of fluid directly onto the object. We consider two tasks: designing a bowl where the task is to contain the fluid inside the object, and a funnel where the task is to concentrate the water onto a location onto the ground. We optimize $r_1$ (which we call the "capture radius" that captures fluid into the object) and $r_2$ (the "filter radius" which decides how much water to filter onto the ground). For the funnel, the desired solution is large $r_1$ and small $r_2$ , and for the bowl we want $r_2 \rightarrow 0$ . If our learned model is performing well, the model should be able to discover these solutions. + +We roll out our model for 50 (bowl) and 75 (funnel) timesteps. Our reward measures the log probability of a Gaussian centered at the desired location of the fluid particles. For the bowl task, we use the bottom of the object, i.e. $\mathcal{N}((0.,0.,h),\Sigma)$ , where $h$ is the object height. For the funnel we use a 2D Gaussian $\mathcal{N}((0.,0.),\Sigma)$ and measure the reward on particles that reach the ground during training. For a given design, we roll out our model and obtain gradients, which we use to iteratively refine the design. + +Qualitative results are shown in Fig. 9. Starting from a sub-optimal design with low $r_1$ and $r_2 > 0$ , our model successfully converges for the bowl; similarly the funnel converges from a suboptimal to an optimal solution. This shows that over long rollouts and hundreds of message-passing steps with dense fluid-surface interactions, our model's gradients are useful for solving inverse design problems. We also find that design is robust across different initializations of $r_1$ and $r_2$ (details in Supp.). + +Latent Space Design. We further examined whether SURF-SUP could directly optimize in the latent space of a DeepSDF model, trained on ShapeNet chairs. For this investigation, as a proof-of-concept we aimed to discover a novel chair that traps and contains falling water (i.e. maximizes the "bowl reward" above). Given a DeepSDF model $f_{\theta}(p,z)$ with latent code $z_0$ , we iteratively refine the design by rolling out our simulator for 100 time-steps, computing a reward, and updating $z$ via gradient-based optimization. Across multiple trials, our model converges from an suboptimal initial chair to a final design that can contain water and retains the semantics of a chair (Fig. 10). + +![](images/5ca803e74f2c3bea259c6a6ba3ff5a72dc98f854e31885982d55bcb6120f0e23.jpg) + +![](images/68f77f40fb030f28717d6bfa71d3a2f3cd06d1ecaf763ad82430e86390afcc14.jpg) + +![](images/91dea91e98a968b4c31f3e8ac9df33ddf919237c6d436cd1b22e5cbfaaa228b5.jpg) + +![](images/626cc9c5ff3c2f95e55a716f5b67e0e8e1d1d6cb671e11b0eb72342d03f5892c.jpg) +Initialization + +![](images/71fbad5635e23a684bd2bdaec298f0a8dc587f8e11151d40647b52c4ff6a5705.jpg) +Intermediate + +![](images/d3a17ae70925fdab343280ed35aa90b68e6e2d3c49041a48cd252eb4b996ad15.jpg) +Final +Figure 9: Design optimization; top is the funnel task, bottom the bowl task. For each frame, the design is shown near the top of the container and a timestep from the fluid rollout is shown below. Initial, intermediate, and final designs are shown; note that the reward increases through optimization and final designs approximate a funnel and bowl well. + +![](images/a03ff94c8af6f1fbdff8aee63b1e62846b521f40cb104c172223cf09c04b8dae.jpg) +Initial Design + +![](images/94c16d356dcdaf51926be6fa5baa27f61d753250d04f43d47aeb7246d5b8a3bd.jpg) +Final Design + +![](images/c0bdf3c72a59d9e696ca3cfe3af6a0eecced8146a4ddcae17b0e7041f42703b2.jpg) +Reward Convergence +Figure 10: We differentiate through SURFSUP to directly optimize (in the latent space of a DeepSDF) a chair that holds water dropped onto it. + +# 6. Conclusion + +In this paper, we have taken a first step towards integrating implicit 3D representations into the dynamical simulation of physical models. SURFSUP realizes this approach by incorporating the implicit representation of solid surfaces (via SDFs) into a graph neural network fluid simulator. As such, SURFSUP is able to accurately model how fluids interact with surfaces, including highly complex surfaces unseen in training. We find that promoting this type of OOD generalization is helpful in both forward and inverse contexts. + +Acknowledgements: We thank Huy Ha for help with visualizations of fluid rollouts. We thank Jan Bender for helpful feedback on SPlisHSPlasH, and Kelsey Allen, Tobias Pfaff, and Alvaro Sanchez-Gonzalez for advice on graph network simulators. We thank Max Helman for early setup work on SPlisHSPlasH. This research is partially supported by the NSF STC for Learning the Earth with Artificial Intelligence and Physics, and the NSF NRI Award #1925157. AM is supported by the NSF fellowship. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas J Guibas. Learning representations and generative models for 3d point clouds. arXiv preprint arXiv:1707.02392, 2017. +[2] Kelsey R Allen, Tatiana Lopez-Guevara, Kimberly Stachenfeld, Alvaro Sanchez-Gonzalez, Peter Battaglia, Jessica Hamrick, and Tobias Pfaff. Physical design using differentiable learned simulators. arXiv preprint arXiv:2202.00728, 2022. +[3] Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interaction networks for learning about objects, relations and physics. Advances in neural information processing systems, 29, 2016. +[4] Jan Bender and Dan Koschier. Divergence-free smoothed particle hydrodynamics. In Proceedings of the 14th ACM SIGGRAPH / Eurographics Symposium on Computer Animation, SCA '15, page 147-155, New York, NY, USA, 2015. Association for Computing Machinery. +[5] Jan Bender and Dan Koschier. Divergence-free smoothed particle hydrodynamics. In Proceedings of the 14th ACM SIGGRAPH / Eurographics Symposium on Computer Animation, SCA '15, page 147-155, New York, NY, USA, 2015. Association for Computing Machinery. +[6] Siddhartha Chaudhuri, Daniel Ritchie, Jiajun Wu, Kai Xu, and Hao Zhang. Learning generative models of 3d structures. In Computer Graphics Forum, volume 39, pages 643-666. Wiley Online Library, 2020. +[7] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), 2016. +[8] James D Foley, Foley Dan Van, Andries Van Dam, Steven K Feiner, and John F Hughes. Computer graphics: principles and practice, volume 12110. Addison-Wesley Professional, 1996. +[9] Christoph Gissler, Andreas Peer, Stefan Band, Jan Bender, and Matthias Teschner. Interlinked sph pressure solvers for strong fluid-rigid coupling. ACM Transactions on Graphics (TOG), 38(1):1-13, 2019. +[10] Zekun Hao, Hadar Averbuch-Elor, Noah Snavely, and Serge Belongie. Dualsdf: Semantic shape manipulation using a two-level representation. pages 7631–7641, 2020. +[11] Dan Koschier and Jan Bender. Density maps for improved sph boundary handling. In Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 1-10, 2017. +[12] Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. In ICLR, 2019. +[13] Yunzhu Li, Jiajun Wu, Jun-Yan Zhu, Joshua B Tenenbaum, Antonio Torralba, and Russ Tedrake. Propagation networks for model-based control under partial observation. In 2019 International Conference on Robotics and Automation (ICRA), pages 1205-1211. IEEE, 2019. + +[14] Zhijian Liu, William T Freeman, Joshua B Tenenbaum, and Jiajun Wu. Physical primitive decomposition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-19, 2018. +[15] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. +[16] Andreas Mayr, Sebastian Lehner, Arno Mayrhofer, Christoph Kloss, Sepp Hochreiter, and Johannes Brandstetter. Boundary graph neural networks for 3d simulations. arXiv preprint arXiv:2106.11299, 2021. +[17] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. +[18] Mariem Mezghanni, Théo Bodrito, Malika Boulkenafed, and Maks Ovsjanikov. Physical simulation layer for accurate 3d modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13514-13523, 2022. +[19] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019. +[20] J. J. Monaghan. Smoothed particle hydrodynamics. Annual Review of Astronomy and Astrophysics, 30:543-574, Jan. 1992. +[21] Damian Mrowca, Chengxu Zhuang, Elias Wang, Nick Haber, Li F Fei-Fei, Josh Tenenbaum, and Daniel L Yamins. Flexible neural representation for physics prediction. Advances in neural information processing systems, 31, 2018. +[22] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. +[23] Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh-based simulation with graph networks. In International Conference on Learning Representations, 2020. +[24] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 22468-22478. Curran Associates, Inc., 2020. +[25] Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In International Conference on Machine Learning, pages 8459-8468. PMLR, 2020. +[26] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20(1):61-80, 2008. + +[27] C. Schenck and D. Fox. Spnets: Differentiable fluid dynamics for deep neural networks. In Proceedings of the Second Conference on Robot Learning (CoRL), Zurich, Switzerland, 2018. +[28] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. 2019. +[29] Vincent Sitzmann, Eric R. Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In Proc. Neurips, 2020. +[30] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Proc. NeurIPS, 2020. +[31] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In Computer Vision and Pattern Recognition (CVPR), 2017. +[32] Benjamin Ummenhofer, Lukas Prantl, Nils Thuerey, and Vladlen Koltun. Lagrangian fluid simulation with continuous convolutions. In International Conference on Learning Representations, 2019. +[33] Subeesh Vasu, Nicolas Talabot, Artem Lukoianov, Pierre Baqué, Jonathan Donier, and Pascal Fua. Hybridsdf: Combining deep implicit shapes and geometric primitives for 3d shape representation and manipulation. In International Conference on 3D Vision, 2022. +[34] Zhirong Wu, S. Song, A. Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and J. Xiao. 3d shapenets: A deep representation for volumetric shapes. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, Los Alamitos, CA, USA, jun 2015. IEEE Computer Society. \ No newline at end of file diff --git a/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/images.zip b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eccde39d3880400802f18534a7be137beec156f2 --- /dev/null +++ b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68ad4c652486e20c0880dae5e8541f1df102368cb063672ce897047b15d7a5c2 +size 576443 diff --git a/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/layout.json b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d3565e310db8a7fcecd2d545272fa290e98956f4 --- /dev/null +++ b/2023/SurfsUP_ Learning Fluid Simulation for Novel Surfaces/layout.json @@ -0,0 +1,8529 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 120, + 103, + 474, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 103, + 474, + 121 + ], + "spans": [ + { + "bbox": [ + 120, + 103, + 474, + 121 + ], + "type": "text", + "content": "SURFSUP: Learning Fluid Simulation for Novel Surfaces" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 148, + 530, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 148, + 530, + 193 + ], + "spans": [ + { + "bbox": [ + 63, + 148, + 530, + 193 + ], + "type": "text", + "content": "Arjun Mani\\*, Ishaan Preetam Chandratreya\\*, Elliot Creager\\*3, Carl Vondrick1, Richard Zemel1 \n1 Columbia University 2 University of Toronto 3 Vector Institute \nsurfsup.cs.columbia.edu" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 75, + 213, + 520, + 452 + ], + "blocks": [ + { + "bbox": [ + 75, + 213, + 520, + 452 + ], + "lines": [ + { + "bbox": [ + 75, + 213, + 520, + 452 + ], + "spans": [ + { + "bbox": [ + 75, + 213, + 520, + 452 + ], + "type": "image", + "image_path": "39ac34228a60edba9ac06071898a4fe9d2f9da8e50af07105250e807be33b435.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 464, + 536, + 479 + ], + "lines": [ + { + "bbox": [ + 56, + 464, + 536, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 536, + 479 + ], + "type": "text", + "content": "Figure 1: We introduce SURFSUP, which learns to predict how fluid particles interact with novel, complex 3D surfaces." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "spans": [ + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 505, + 290, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 505, + 290, + 696 + ], + "spans": [ + { + "bbox": [ + 45, + 505, + 290, + 696 + ], + "type": "text", + "content": "Modeling the mechanics of fluid in complex scenes is vital to applications in design, graphics, and robotics. Learning-based methods provide fast and differentiable fluid simulators, however most prior work is unable to accurately model how fluids interact with genuinely novel surfaces not seen during training. We introduce SURFSUP, a framework that represents objects implicitly using signed distance functions (SDFs), rather than an explicit representation of meshes or particles. This continuous representation of geometry enables more accurate simulation of fluid-object interactions over long time periods while simultaneously making computation more efficient. Moreover, SURFSUP trained on simple shape primitives generalizes considerably out-of-distribution, even to complex real-world scenes and objects. Finally, we show we can invert our model to design simple objects to manipulate fluid flow." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 487, + 387, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 387, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 387, + 499 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 508, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 676 + ], + "type": "text", + "content": "Across engineering and science, the simulation of fluid dynamics has become an invaluable tool, and it will be a crucial component for building visual systems that are capable of understanding and interacting with real-world environments. Recently, a new class of simulators has emerged that learn the dynamics of physical systems from data [13, 32, 25]. These simulators offer several advantages over classical simulators by increasing speed, reducing reliance on hand-crafted dynamics models, and providing differentiable rollouts for solving inverse problems. Recent approaches based on graph neural networks (GNNs) have in particular shown impressive accuracy and generalization on a wide range of fluids represented as particle or mesh-based systems [25, 23]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "In a wide variety of applications areas, fluid simulation must properly handle interaction with solid structures to attain practical relevance. Many problems involving" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 702, + 156, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 156, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 156, + 714 + ], + "type": "text", + "content": "*Authors contributed equally." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14225" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "fluid simulation are fundamentally about object design; for example, designing aerodynamic shapes for cars and airplanes. In robotics too, an intelligent robot would have to understand fluid-surface interactions; for example, to carry a mug of hot coffee to a person without spilling the liquid from the container. In both cases, the need to accurately model how fluids interact with novel solid surfaces1—with new shapes, configurations, orientations, and so on—motivates an challenging out-of-distribution (OOD) generalization subproblem for fluid simulation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 455 + ], + "type": "text", + "content": "In this paper, we tackle this problem by introducing Surface implicit sUbstitution for particles (SURFSUP), a framework for simulating the interactions of fluids and surfaces by representing surfaces using implicit representations [19]. Current GNN-based approaches represent surfaces explicitly, discretizing them into particles and augmenting the graph accordingly [25]. Implicit representations, in particular signed distance functions which represent surfaces as zero-level sets of functions, offer several compelling advantages. They can smoothly represent continuous surfaces, provide rich geometric information (e.g. whether a point is in the interior/exterior of the object), and can scale easily to large objects and scenes. Due to these properties, many state-of-the-art generative models for 3D shapes learn shapes with signed distance functions. We argue that these properties are useful not just for statics but also dynamics: in order to accurately model fluid-surface interactions for potentially complex surfaces, smooth and informative representations of local surface geometry are needed. In turn, we posit that these locally informative representations will enable fluid simulators to appropriately model interaction with never-before-seen surfaces." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 456, + 289, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 289, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 289, + 575 + ], + "type": "text", + "content": "Concretely, we propose an approach that integrates signed distance representations of surfaces [22, 19] with graph neural network models [26, 25]. Our architecture follows the encode-process-decode paradigm [25], where the particles in the physical system at time " + }, + { + "bbox": [ + 46, + 456, + 289, + 575 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 456, + 289, + 575 + ], + "type": "text", + "content": " are encoded into a graph, followed by several message-passing steps, and finally a decoder predicts the dynamics (accelerations). We represent the surface using its signed distance function, and exploit geometric properties of SDFs to model the distance and orientation of fluid particles with respect to the surface." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 575, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 289, + 696 + ], + "type": "text", + "content": "Experiments show that our approach can model fluid-surface interactions over long rollouts with a high degree of accuracy and visual realism. SURFSUP can model surfaces far beyond its training domain; when trained on simple primitive shapes (e.g. spheres, cones, toruses), we generalize zero-shot to complex real-world objects and scenes represented using neural implicit generative models (Fig. 1). We compare to a method which represents the surface as particles inside the graph network simulator, and show that our model achieves better generalization and efficiency, and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": "can better handle complex surface geometry. Finally, we demonstrate the potential of SURFSUP to solve inverse problems across two object design tasks. Overall, our paper bridges learned physical models for dynamics and 3D implicit representations, and takes a step towards AI models which can understand and help design the physical world." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 155, + 392, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 155, + 392, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 392, + 167 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 175, + 547, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 547, + 247 + ], + "type": "text", + "content": "Learning Differentiable Simulators. Using machine learning for predicting and modeling complex physical systems is a rapidly growing area of research. Learned physics simulators have been shown to accurately simulate colliding rigid objects, deformable objects, fluids, and other complex systems [13, 3, 12, 32, 25, 21]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 247, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 462 + ], + "type": "text", + "content": "Recent progress in learned fluid simulation has been driven by the Lagrangian viewpoint, which tracks the motion of fluid particles over time. Graph neural networks (GNNs) have in particular emerged as effective forward models for fluid simulation, and have demonstrated better stability and generalization compared to grid-based, convolutional approaches. [13, 14, 27, 25]. Sanchez-Gonzalez et al. [25] introduced a GNN architecture which can accurately and generalizably model fluid behavior over long rollouts; most subsequent work builds off their approach [23, 16]. However, most of these works generally consider simple surfaces; e.g. only cube-shaped containers for fluids. There has been little work focused on fluid-surface interactions; Mayr et al. [16] introduce an approach specifically for triangularized meshes. All these approaches use particles or meshes to represent surfaces; we will argue that implicit representations enable learning more accurate, generalizable physical dynamics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 463, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 547, + 582 + ], + "type": "text", + "content": "Some works have explored the possibility of using learned fluid simulators for solving inverse problems [27, 12, 2]. Allen et al. [2] showed that graph network simulators can be used to solve design tasks by rolling out fluid trajectories and using gradient-based optimization. In 3D, they show that a 2D plane with a learned height field can be used to split a stream of fluid to fall into several locations on the ground. SDFs have the advantage of being able to learn geometry in a more smooth and unconstrained way, which could help solve inverse problems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "3D Implicit Representations. Implicit representations have become state-of-the-art for generative models of 3D objects [17, 22, 6, 29, 30]. These models represent the surface as a function " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "F(x)" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": ". There are several choices for " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": ". Park et al. [22] show that signed distance functions " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "F(x)" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " are particularly effective and informative representations. Compared to point clouds [1, 28] or voxels [7, 34], models based on SDF representations can smoothly represent complex geometry, are compact and efficient, and can be learned effectively [29, 6]. Follow-up work has studied learning \"dual\" generative models that jointly learn an" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 285, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 285, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 285, + 714 + ], + "type": "text", + "content": "For brevity we refer to surfaces of solid objects simply as \"surfaces\"." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14226" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "uncontrained representation and one composed of primitive shapes [10, 33]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "type": "text", + "content": "Implicit Representations for Physical Dynamics. Some classical particle-based fluid simulators based on the Smoothed Particle Hydrodynamics (SPH) formalism [20] have transitioned from particle-based to implicit surface representations. [11, 9, 4]. In these simulators implicit surfaces allow for smoother fluid motion with less artifacts [11]. For general-purpose learned simulators (GNNs), there is little prior work using implicit rather than particle-based representations of surfaces." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 204, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 286, + 300 + ], + "type": "text", + "content": "Another line of work combines implicit generative models with physical simulators [18, 24]. Mezghanni et al. [18] generate more physically stable SDF shapes by backpropagating gradients through a physical simulation layer. Their simulation mainly considers Newtonian dynamics. Our contribution is an effective, generalizable forward model for fluids that represents surfaces accurately and efficiently, and can also be used for solving inverse problems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 102, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 102, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 102, + 323 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 331, + 257, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 257, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 257, + 344 + ], + "type": "text", + "content": "3.1. Setup: Particle-Based Fluid Simulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": "We address the problem of learning the physical dynamics of fluids interacting with surfaces. Given the state of a physical system " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "X^0" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": " and some initial conditions (e.g. gravity, other external forces), rolling out a fluid simulation over " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": " steps produces a sequence of states " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "X^{(1)},\\dots,X^{(T)}" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": ". In this paper, we focus on particle-based simulation; the state of the system at time " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": " is represented by a set of particles " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "X^{(t)} = (p_1^{(t)},\\dots,p_n^{(t)})" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": " is the position of particle " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": " in 3D space. A rollout over " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": " steps is produced by repeatedly applying a forward model " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "S_{\\theta}:X^{(t)}\\to X^{(t + 1)}" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": ", which takes in the particle positions at time " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": " (and optionally a short history " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "t - 1,\\ldots" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": ") and produces the next-step particle positions. While the forward model can be used to directly predict " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "X^{(t + 1)}" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": ", most methods instead predict intermediate values " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": ", typically next-state accelerations for each particle, which can then be numerically integrated (e.g. Euler integration) to produce " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "X^{(t + 1)}" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": "; concretely, our task is to learn the parameters " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": " of this forward model " + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "inline_equation", + "content": "f_{\\theta}:X^{(t)}\\to Y" + }, + { + "bbox": [ + 46, + 350, + 286, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 575, + 286, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 286, + 588 + ], + "type": "text", + "content": "3.2. Graph Networks for Particle-Based Simulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Graph neural networks lend themselves naturally to particle-based simulation, since they compute pairwise interactions between nodes and aggregate these interactions. Graph Network-based Simulators (GNS), a model architecture introduced by Sanchez-Gonzalez et al. [25], have in particular shown state-of-the-art performance on simulating fluids over long rollouts. These methods follow an \"encode-process-decode\" paradigm. First, the state " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "X^{(t)}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " is encoded into a graph, where nodes " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " are instantiated using a set of features (e.g. previous velocities, parti" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": "cle type, etc.); nodes " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "v_{i}, v_{j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": " are then connected with an edge " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": " if their distance is less than a connectivity radius " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": ". Each node and edge is associated with an embedding. The \"processor\" stage then passes the encoded graph through " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": " message-passing steps. Each message-passing step takes the current graph " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\{(v_{i}, v_{j}, e_{ij})\\}^{(m)}" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": ". An edge module first computes the updated edge embeddings " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "e_{ij}^{(m+1)} = f_{e}(v_{i}^{(m)}, v_{j}^{(m)}, e_{ij}^{(m)})" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": "; then the edge embeddings are aggregated to compute updated node embeddings " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "v_{i}^{(m+1)} = f_{n}(v_{i}^{(m)}, \\sum_{j \\in N(i)} e_{ij}^{(m)})" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "N(i)" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": " are the set of neighbors of node " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": " are generally MLPs with 2-3 hidden layers, and the weights are shared across edges/nodes (although differing across message passing steps). Finally, the \"decoder\" step applies an MLP " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "f_{D}" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": " to the final node embeddings to predict the accelerations " + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\{y_{i}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 255, + 545, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 362 + ], + "type": "text", + "content": "In our problem, we further assume that there exists some rigid surface " + }, + { + "bbox": [ + 304, + 255, + 545, + 362 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 255, + 545, + 362 + ], + "type": "text", + "content": " that interacts with the fluid throughout the simulation. The GNS approach handles surfaces by discretizing the surface into particles (i.e. a point cloud) and adding these particles " + }, + { + "bbox": [ + 304, + 255, + 545, + 362 + ], + "type": "inline_equation", + "content": "\\{s_i\\}" + }, + { + "bbox": [ + 304, + 255, + 545, + 362 + ], + "type": "text", + "content": " to the graph (denoting them by a different particle type, and masking during position updates) [25]. Follow-up work for mesh-based simulation handles surfaces similarly by representing the surface as an irregular mesh and adding nodes for mesh vertices [23]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 369, + 545, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 545, + 382 + ], + "type": "text", + "content": "3.3. Modeling Fluid-Surface Interactions with SDFs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 388, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 449 + ], + "type": "text", + "content": "We propose to model the interactions of fluids and surfaces with signed distance function (SDF) representations of surfaces. An SDF represents a watertight surface as a function " + }, + { + "bbox": [ + 304, + 388, + 545, + 449 + ], + "type": "inline_equation", + "content": "F(p) = s" + }, + { + "bbox": [ + 304, + 388, + 545, + 449 + ], + "type": "text", + "content": ", which takes a spatial point " + }, + { + "bbox": [ + 304, + 388, + 545, + 449 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 388, + 545, + 449 + ], + "type": "text", + "content": " and outputs its distance " + }, + { + "bbox": [ + 304, + 388, + 545, + 449 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 388, + 545, + 449 + ], + "type": "text", + "content": " to the closest corresponding point on" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 351, + 464, + 502, + 633 + ], + "blocks": [ + { + "bbox": [ + 351, + 464, + 502, + 633 + ], + "lines": [ + { + "bbox": [ + 351, + 464, + 502, + 633 + ], + "spans": [ + { + "bbox": [ + 351, + 464, + 502, + 633 + ], + "type": "image", + "image_path": "a76562248cb915f5d8fc7915cb5b0f084083b8fe295f0466c397ec3f7aa34009.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 642, + 545, + 700 + ], + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 700 + ], + "type": "text", + "content": "Figure 2: Rigid surfaces can be represented compactly and implicitly as the level set of a SDF: " + }, + { + "bbox": [ + 304, + 642, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\{x|F(x) = 0\\}" + }, + { + "bbox": [ + 304, + 642, + 545, + 700 + ], + "type": "text", + "content": ". Evaluating " + }, + { + "bbox": [ + 304, + 642, + 545, + 700 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 642, + 545, + 700 + ], + "type": "text", + "content": " and its derivatives at particle position " + }, + { + "bbox": [ + 304, + 642, + 545, + 700 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 642, + 545, + 700 + ], + "type": "text", + "content": " provides useful information for simulating the interaction of that particle with the surface." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14227" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "the surface. The function is \"signed\", such that it takes on positive values for points outside the surface and negative values in the surface interior. The surface of the object is implicitly represented as the zero-level set " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "F(p) = 0" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": ". Explicit representations of the surface can be constructed by applying meshing algorithms such as Marching Cubes [15]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 286, + 240 + ], + "type": "text", + "content": "Almost every surface can be described by an SDF. For several primitive shapes SDFs can be written down analytically (e.g. spheres, cones, cylinders); boolean operations on these shapes such as unions and intersections give rise to more complex shapes [31]. SDFs can also be represented discretely as voxel grids; or more recently for finer geometry, as neural SDFs " + }, + { + "bbox": [ + 46, + 145, + 286, + 240 + ], + "type": "inline_equation", + "content": "F_{\\theta}(p)" + }, + { + "bbox": [ + 46, + 145, + 286, + 240 + ], + "type": "text", + "content": " which learn to accurately approximate " + }, + { + "bbox": [ + 46, + 145, + 286, + 240 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 145, + 286, + 240 + ], + "type": "text", + "content": " from mesh-based supervision." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 258, + 286, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 286, + 281 + ], + "type": "text", + "content": "3.3.1 Locality and Geometric Information for Fluid-Surface Interactions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 286, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 286, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 286, + 433 + ], + "type": "text", + "content": "A key insight of our method is that fluid-surface interactions are inherently about local rather than global surface geometry. The interaction of a fluid with a surface depends on a small local region of the surface, over which the geometry can have limited variation. This suggests that an effective representation of local surface geometry, trained on a variety of object shapes, can generalize OOD to shapes that may look different globally. By contrast, particles are not an effective representation of surface geometry; for any moderately complex surface, discretization leads to artifacts such as non-smooth fluid motion and even penetration of the surface boundary (we confirm this empirically)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": "By contrast, we argue that SDFs are a smooth, richly informative representation of local surface geometry for learning generalizable fluid-surface interactions. Our key insight is shown in Figure 2. Given a solid surface represented as an SDF " + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "inline_equation", + "content": "F(p)" + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": ", we can use properties of the SDF to inform the behavior of a fluid particle as it reaches the surface. The value of the SDF " + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "inline_equation", + "content": "F(p)" + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": " gives the distance from the particle at point " + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": " to the closest point on the surface " + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": ". Moreover, the SDF gradient " + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "inline_equation", + "content": "\\nabla F(p)" + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": " provides the unit-norm displacement vector from the particle to its closest surface point " + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "inline_equation", + "content": "\\frac{(p - p')}{||p - p'||}" + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": ". This can be shown by observing that the SDF decreases most quickly on the line from " + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": ". Thus, the SDF and its gradient provide the distance and orientation of the fluid particle w.r.t. to its closest surface point; as the particle approaches the surface, its acceleration will more closely approach " + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "inline_equation", + "content": "\\nabla F(p)" + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": " and it will tend to rebound. Moreover, since the gradient varies smoothly because of the continuity of " + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 434, + 286, + 652 + ], + "type": "text", + "content": ", these features do not suffer from artifacts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Higher order derivatives of the SDF can also be used to gain geometric information. For example, we could march to the surface by computing " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "p - F(x)\\nabla F(x)" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": ". At the surface, the Hessian " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\nabla^2 F(x)" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " is the shape operator; its eigenvalues provide the principal directions of surface curvature" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 76, + 547, + 182 + ], + "blocks": [ + { + "bbox": [ + 310, + 76, + 547, + 182 + ], + "lines": [ + { + "bbox": [ + 310, + 76, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 310, + 76, + 547, + 182 + ], + "type": "image", + "image_path": "6187da8e669f4f3de12fe83663c39f6f7f68d55627cc19ade6c9b04a1092a8c9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 190, + 545, + 238 + ], + "lines": [ + { + "bbox": [ + 305, + 190, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 545, + 238 + ], + "type": "text", + "content": "Figure 3: Given fluid particles and the SDF, our method augments the constructed graph with features from the SDF. We then apply message-passing and predict dynamics, weighting near-surface errors more highly in our loss." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 265, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\kappa_{1}" + }, + { + "bbox": [ + 305, + 265, + 545, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 265, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\kappa_{2}" + }, + { + "bbox": [ + 305, + 265, + 545, + 396 + ], + "type": "text", + "content": ", and the mean curvature is " + }, + { + "bbox": [ + 305, + 265, + 545, + 396 + ], + "type": "inline_equation", + "content": "H = \\frac{1}{2} \\mathrm{Tr}(\\nabla^2 F)" + }, + { + "bbox": [ + 305, + 265, + 545, + 396 + ], + "type": "text", + "content": ". We can also estimate the surface curvature via finite differences by slightly perturbing the gradient vector on each axis and computing the SDF (Fig. 2). In this paper, we only use the SDF and its gradient as features without curvature information; this can be seen as an effective theory where the relevant \"local region\" of the surface is a single closest point on the surface. As we show, modeling surfaces using SDF representations helps our fluid simulator generalize beyond the available training data by modeling fluid-surface interactions for unseen surfaces." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 422, + 358, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 358, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 358, + 433 + ], + "type": "text", + "content": "3.4. Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "text", + "content": "Having developed our ideas about SDFs and modeling fluid-surface interactions, we now describe our approach, which we call Surface implicit sUbstitution for particles (SURFSUP). We learn a forward predictor " + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}:X^{(t)},F\\to Y" + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "text", + "content": ", which takes as input a set of particle positions and a surface implicitly parametrized by an SDF " + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "text", + "content": ". We adopt the encode-process-decode paradigm for our model, but make two key changes. In the encoder step, for each fluid particle " + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "text", + "content": " we add the value of the SDF " + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "inline_equation", + "content": "F(p_i)" + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "text", + "content": " and its gradient " + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\nabla F(p_i)" + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "text", + "content": " to the initial node representations, such that the new representation is " + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime (0)} = [v_{i}^{(0)},F(p_{i}),\\nabla F(p_{i})]" + }, + { + "bbox": [ + 305, + 446, + 545, + 713 + ], + "type": "text", + "content": ". This is the way that the surface is encoded into the model; subsequently no particles representing the surface are added to the graph. The updated node embeddings are then passed through the \"processor\" with several message-passing steps and then decoded to predict the accelerations as standard. The other modification we make is to weight particles close to the surface higher during training when predicting dynamics. This has the effect of teaching the model to prioritize modeling fluid-surface interactions accurately. Since the model is predicting next-step accelerations, our loss becomes:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14228" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 52, + 91, + 287, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 91, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 52, + 91, + 287, + 119 + ], + "type": "interline_equation", + "content": "L = \\sum_ {F (p _ {i}) > \\alpha} \\left\\| a _ {i} ^ {(t)} - \\hat {a} _ {i} ^ {(t)} \\right\\| ^ {2} + \\sum_ {F (p _ {i}) \\leq \\alpha} \\lambda \\left\\| a _ {i} ^ {(t)} - \\hat {a} _ {i} ^ {(t)} \\right\\| ^ {2} \\tag {1}", + "image_path": "814ea7fe8ef350d0546a1611ec01c97937575f960641648cadcb4f4aef3082d1.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 128, + 287, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 128, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 287, + 212 + ], + "type": "text", + "content": "In our experiments, we set " + }, + { + "bbox": [ + 47, + 128, + 287, + 212 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 128, + 287, + 212 + ], + "type": "text", + "content": " based on the \"neighborhood radius\" of the classical simulator we train against, and set " + }, + { + "bbox": [ + 47, + 128, + 287, + 212 + ], + "type": "inline_equation", + "content": "\\lambda = 5" + }, + { + "bbox": [ + 47, + 128, + 287, + 212 + ], + "type": "text", + "content": ". We found that this weighting helped the model learn surface interactions more quickly without sacrificing overall accuracy on fluid prediction. Note that the SDF allows loss weighting to be implemented in a very natural way, by simple thresholding. Our approach is shown in Fig. 3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 219, + 269, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 219, + 269, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 219, + 269, + 232 + ], + "type": "text", + "content": "3.5. Other Advantages of SDF Representations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 238, + 287, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 287, + 286 + ], + "type": "text", + "content": "We have argued for SDFs from the perspective of achieving smooth, information-rich representations of local surface geometry. Here we mention other advantages of SDFs, each of which we demonstrate in our experimental results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 289, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 287, + 373 + ], + "type": "text", + "content": "Efficient message passing in the GNN. By modeling surfaces implicitly, we reduce the number of unique particles needed to compose the GNN's state. This allows us to perform more lightweight and computationally efficient simulation, as well as run the model for more message-passing steps. This advantage becomes especially clear for large objects/scenes or objects with complex surface geometry." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 376, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 521 + ], + "type": "text", + "content": "Shape Design with SDFs. SDFs are a useful paradigm for shape design and assembly. This is because compositions of shapes can be generated by union, intersection, and difference operations; e.g. the union of two SDFs " + }, + { + "bbox": [ + 47, + 376, + 287, + 521 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 47, + 376, + 287, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 376, + 287, + 521 + ], + "type": "inline_equation", + "content": "F_{2}" + }, + { + "bbox": [ + 47, + 376, + 287, + 521 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 47, + 376, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\min(F_{1}, F_{2})" + }, + { + "bbox": [ + 47, + 376, + 287, + 521 + ], + "type": "text", + "content": ". The construction of shapes using primitives and these operations is known as Constructive Solid Geometry (CSG), and is implemented in many CAD packages for object design [8]. Moreover the recent emergence of neural SDF models as state-of-the-art 3D shape representations, offers another compelling avenue for shape design with end-to-end learning. These suggest that good forward models with SDFs could be leveraged for design problems." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 531, + 132, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 132, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 132, + 544 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 551, + 138, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 138, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 138, + 562 + ], + "type": "text", + "content": "4.1. Fluid Domains" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 570, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 653 + ], + "type": "text", + "content": "We evaluate our approach on complex prediction tasks involving fluids interacting with a variety of surfaces. Our main contribution here is a 3D fluid dataset consisting of water interacting with a range of primitive shapes, across shape parameters, positions, and orientations. For the ground-truth simulator we use SPlisHSPlasH [5], a 3D-particle-based fluid simulator based on the SPH technique." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": "PrimShapes Dataset. We train SURFSUP on a 3D water dataset, where the fluid interacts with various shapes. We initialize a block of fluid directly above the shape, then let it fall due to gravity. The entire scene takes place in a " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1 \\times 2" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " m container. A random rotation is applied to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "the base shape which sits on the bottom of the container, and the initial position of the shape is also randomized. We extract 800 time-steps from the ground-truth fluid rollout (with " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\Delta t = 0.005" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "); this corresponds to 4000 steps for the classical simulator, which requires very small time-steps for stability [25]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 144, + 546, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 335 + ], + "type": "text", + "content": "We use five \"primitive\" shapes in our dataset: spheres, boxes, cylinders, cones, and toruses. These shapes have SDFs with analytical forms. (E.g. the simplest form is the origin-centered, radius- " + }, + { + "bbox": [ + 304, + 144, + 546, + 335 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 144, + 546, + 335 + ], + "type": "text", + "content": " sphere with SDF " + }, + { + "bbox": [ + 304, + 144, + 546, + 335 + ], + "type": "inline_equation", + "content": "F(p) = ||p||_2 - r" + }, + { + "bbox": [ + 304, + 144, + 546, + 335 + ], + "type": "text", + "content": "; the others tend to be more complex). We emphasize that SURFSUP is trained only on simulations containing these primitive shapes and generalizes zero-shot to complex objects and scenes. Our key insight is that accurate fluid-surface simulation depends on understanding local surface geometry, of which a significant diversity is expressed by these shapes (like curvature, holes, differently-sloped faces, corners). Leveraging this and the properties of SDFs, SURFSUP can generalize to interactions with surfaces of entirely novel global geometry. Our training and test set consists of 1000 and 100 simulations respectively. See Supp. Fig 6 for example initializations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 350, + 443, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 443, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 443, + 361 + ], + "type": "text", + "content": "4.1.1 Generalization Test-Sets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 369, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 430 + ], + "type": "text", + "content": "We create a range of qualitative and quantitative evaluation scenarios to test our model's generalization ability. The first three below assess generalization to novel analytical SDFs (e.g. unions " + }, + { + "bbox": [ + 304, + 369, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\min(f_1, f_2)" + }, + { + "bbox": [ + 304, + 369, + 545, + 430 + ], + "type": "text", + "content": " or differences " + }, + { + "bbox": [ + 304, + 369, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\max(f_1, -f_2)" + }, + { + "bbox": [ + 304, + 369, + 545, + 430 + ], + "type": "text", + "content": "), and the last to real-world shapes represented as neural SDFs " + }, + { + "bbox": [ + 304, + 369, + 545, + 430 + ], + "type": "inline_equation", + "content": "f_\\theta" + }, + { + "bbox": [ + 304, + 369, + 545, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 474 + ], + "type": "text", + "content": "Prim-OOD (60 simulations). This contains the same shapes with OOD shape parameters from the training set (e.g. cylinder with higher radius " + }, + { + "bbox": [ + 304, + 437, + 545, + 474 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 437, + 545, + 474 + ], + "type": "text", + "content": " and lower height " + }, + { + "bbox": [ + 304, + 437, + 545, + 474 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 437, + 545, + 474 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 510 + ], + "type": "text", + "content": "Prim-Unions (40). This test set includes all possible pairwise unions which 'merge' two of the primitive shapes. See Supp. for examples." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 545 + ], + "type": "text", + "content": "Funnels (12). We examine the model's ability to handle analytical \"differences\" of shapes by creating funnels, formed by carving an inverted cone out of a cylinder." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "'Complex' Scenes (7). As the key, challenging test of our model's generalization, we simulate fluid interacting with several complex real-world objects. To do so, we leverage neural SDFs of shapes, which can represent complex shapes with high accuracy and fidelity (Sec. 3.3). Instead of an analytical SDF " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "f(p)" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", we use a trained neural SDF " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}(p)" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " to generate our SDF features during our forward pass. To test our model's ability to generalize zero-shot to complex objects/scenes, we train a single SIREN model [30] for 7 fine-grained objects (including the coral, lion, and room scenes shown in Fig. 1, Fig. 4). For each object we create a fluid scene using the ground-truth simulator, and evaluate SURFSUP on this small Complex-Scenes dataset. To represent the geometry at an appropriate scale, these objects are" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14229" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 449, + 160 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 449, + 160 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 449, + 160 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 449, + 160 + ], + "type": "table", + "html": "
Testing SetChamferChamfer SurfaceNumber InsideMean SDF Inside
SURFSUPGNS [25]SURFSUPGNSSURFSUPGNSSURFSUPGNS
Primitives0.02970.02854.7045.45102000-1.04e-11-2.48e-05
Primitives-OOD0.03210.034310.95118.44532433006-1.65e-06-0.0006
Primitives-Unions0.03800.034824.55326.3071000063000-4.46e-05-0.001
Funnels0.03620.05219.99013.66327204-2.51e-09-0.001
Complex-Scenes0.06380.265850.521276.17640513257722-0.0011-0.026
", + "image_path": "e2fadaabcd31e738f56913c883d697e7eb5585d05c2026638e906365a8af7677.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 454, + 71, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 454, + 71, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 454, + 71, + 545, + 167 + ], + "type": "text", + "content": "Table 1: Results on in-distribution (Primitives) and OOD test sets, comparing SURFSUP to GNS [25] which represents the surface as particles." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 287, + 248 + ], + "type": "text", + "content": "scaled 2x in the fluid scene compared to previous datasets. To test our model on families of shapes, which gives us the capacity to solve inverse problems, we train a variational DeepSDF model [22] for the ShapeNet chair and bowl categories. See Supp. for details on these scenes and models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 257, + 145, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 145, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 145, + 270 + ], + "type": "text", + "content": "4.2. Training Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 276, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 337 + ], + "type": "text", + "content": "For all experiments, we train our model on the PrimShapes dataset. The training data is pairs " + }, + { + "bbox": [ + 46, + 276, + 287, + 337 + ], + "type": "inline_equation", + "content": "X^{(t)}, X^{(t+1)}" + }, + { + "bbox": [ + 46, + 276, + 287, + 337 + ], + "type": "text", + "content": " from the ground-truth simulator. We use a graph network (GN) architecture with 10 message-passing steps and train with Adam for 2M steps (see Supp. for details)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 346, + 168, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 168, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 168, + 358 + ], + "type": "text", + "content": "4.3. Metrics and Baseline" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "content": "Although we train on next-step prediction, we measure performance on full rollouts. We use mean Chamfer distance between the predicted particles " + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "content": " and ground-truth " + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "content": " to measure overall rollout agreement (averaging across timesteps). Compared to MSE, Chamfer distance is permutation-invariant and therefore a better measure of similarity between fluid flows [25]. This metric, however, can be noisy since it averages over long rollouts on chaotic systems with many particles. Since we focus on the quality of fluid-surface interactions, we also compute a \"Chamfer surface\" metric that focuses on prediction errors near the surface. For each time-step, we consider the predicted point cloud only near the surface (where " + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "inline_equation", + "content": "F(p) < \\alpha" + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "content": " for threshold " + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "content": ") and measure the Chamfer distance of " + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "inline_equation", + "content": "P_{\\alpha}" + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "content": ", summing over particles; we do this symmetrically for " + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "content": ". Intuitively this focuses on discrepancies between " + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 365, + 287, + 568 + ], + "type": "text", + "content": " for fluid motion near the surface. See Supp. for further details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 569, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 653 + ], + "type": "text", + "content": "Since surface penetration is a serious failure mode, we also measure the number of particles inside the surface through the rollout (summing over the number of timesteps). Additionally, we compute the average SDF value for particles inside the surface, indicating the extent of surface penetration (since SDFs become increasingly negative deeper into the surface interior)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Baseline. Our main comparison is to the graph network simulator (GNS) proposed by Sanchez Gonzalez et al. [25]. Their approach represents surfaces explicitly through discretization into particles, which are added to the graph and participate in the message-passing. To implement this base" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 188, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 545, + 260 + ], + "type": "text", + "content": "line, we convert all the SDFs in our datasets to meshes using Marching Cubes [15]. From this mesh we uniformly sample 2000 surface particles (comparable to the number of fluid particles in our simulations) and add them to the graph. In all our experiments we compare our implicit SURFSUP approach to this baseline with explicit surface particles." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 261, + 546, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 546, + 405 + ], + "type": "text", + "content": "While the baseline method could in principle be extended to include local surface information (e.g. using surface normal estimates for each particle as additional features, or mesh connectivity), we omit such an extension here. This is because any explicit representation of large surfaces quickly becomes computationally infeasible, especially for large objects/scenes or complex geometry. Whereas the simulation quality of SURFSUP scales with the capacity of the SDF network " + }, + { + "bbox": [ + 304, + 261, + 546, + 405 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 261, + 546, + 405 + ], + "type": "text", + "content": ", the baseline quality scales with the number of particles. We show in Sec 5 that SURFSUP still outperforms (with lower computational cost and inference time) a baseline with more particles." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 418, + 362, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 362, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 362, + 431 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 592 + ], + "type": "text", + "content": "SURFSUP can accurately predict fluid behavior over long rollouts, interacting with a range of complex surfaces (Fig. 4). Across the different analytical primitive shapes in our test sets and different orientations, we can generate accurate and visually realistic predictions of fluid behavior. See the Supp. Material for videos of these, and all, rollouts in the paper. Our key result is the following: we find that SURFSUP can perform effective simulation with surfaces that are significantly out-of-distribution, and even generalize to complex real-world shapes and scenes. This demonstrates that SURFSUP can leverage the rich geometric information provided by the SDF for generalizing to novel surfaces expressed as neural SDFs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Quantitative Results and Comparison to Baseline. We evaluate SURFSUP on rollouts from ID and OOD test sets of shapes with analytical SDFs, and on complex scenes, with comparison to the GNS baseline. Results are shown in Table 1. Across all test sets, our model's fluid predictions display significantly lower surface penetration than the baseline (e.g. 324 vs. 33006 particles on average for the primitives+OOD test set). Moreover, for those few particles which have penetrated the surface, the average SDF value for our method is nearly zero; this implies that any" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14230" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 306, + 325 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 306, + 325 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 306, + 325 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 306, + 325 + ], + "type": "image", + "image_path": "740a1f905fe8766f130c663479ba46e507a6d88bed38d43b2288966d2e3941d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 334, + 547, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 547, + 384 + ], + "type": "text", + "content": "Figure 4: Although SURFSUP is only trained on primitive shapes, our model generalizes to complex and real-world shapes without any additional training. We visualize our predictions (middle) for two scenes. In comparison, the baseline (top) often generates unrealistic predictions, such as having water penetrate solid objects. In many cases, the predictions from the baseline cause the water to disappear because it falls through the floor (top right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 545, + 325 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 545, + 325 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 545, + 325 + ], + "type": "image", + "image_path": "94c75b747b8057e4de536fbe705e2fc8cf51ad5f836569d506b66cf7f7f5489f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 403, + 289, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 536 + ], + "type": "text", + "content": "particle that penetrates the boundary " + }, + { + "bbox": [ + 46, + 403, + 289, + 536 + ], + "type": "inline_equation", + "content": "(F(p) < 0)" + }, + { + "bbox": [ + 46, + 403, + 289, + 536 + ], + "type": "text", + "content": " is quickly projected outward to the region " + }, + { + "bbox": [ + 46, + 403, + 289, + 536 + ], + "type": "inline_equation", + "content": "F(p) > 0" + }, + { + "bbox": [ + 46, + 403, + 289, + 536 + ], + "type": "text", + "content": ". On Chamfer distance near the surface, which measures overall accuracy of fluid-surface interactions beyond surface penetration, our method consistently improves over the baseline. Overall Chamfer distance between the two models is generally comparable. This is unsurprising given that small differences in particle positions along long time-horizons, potentially far from the surface, can accumulate error. (Sec. 4.3). See Supp. Fig. 5 for rollouts on rows 1-4 of Table 1, indicating the stability and accuracy of our method's dynamics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "Generalizing to Complex Scenes We evaluate SURFSUP on chairs and bowls from Shapenet, and fine-grained objects and scenes in the real world (See Fig. 1, Fig. 4, Fig. 5, and Fig. 8, also videos in Supp.). Remarkably SURFSUP trained only with primitive shapes generalizes zero-shot to complex shapes/scenes represented as neural SDFs. This is notable because our chosen real-world objects differ widely (see the variation in Fig. 4 and Fig. 5), and second, neural SDFs are imperfect compared to analytical SDFs (e.g. not fulfilling the property " + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "inline_equation", + "content": "||\\nabla f|| = 1" + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "). Despite these challenges, SURFSUP produces rollouts with accurate and realistic fluid dynamics, capturing fine-grained behavior near complex surfaces. Our method also avoids unintuitive artifacts, such as surface penetration, seen in the GNS baseline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 403, + 547, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 584 + ], + "type": "text", + "content": "Quantitative results (Table 1, row 'Complex-Scenes') indicate that our method outperforms the baseline most significantly on complex real-world scenes. These results elucidate key generalization advantages of our method in handling surfaces with complex geometry and large scenes at scale. For the coral scene (Fig. 1), our method's fluid rolls accurately capture and percolate through the fine tentacle structure of the coral. The GNS baseline does not display this ability (Supp. Fig 1). Our method also scales accurately to large scenes such as the room scene in Fig. 1. By contrast, the baseline's predictions consistently exhibit penetration of the furniture surfaces in the scene due to insufficient resolution (see Fig. 6). Recall (Sec. 4.3) that our scaling law is the capacity of the neural SDF network " + }, + { + "bbox": [ + 304, + 403, + 547, + 584 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 403, + 547, + 584 + ], + "type": "text", + "content": " instead of the number of surface particles." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Efficiency of Simulation. We study comparing SURFSUP as we increase the number of surface particles for the GNS baseline (which in our experiments is afforded 2000 surface particles). Note that there is a tradeoff between performance and computational cost (memory, time) for the baseline; more particles improves surface resolution while also increasing graph sizes. We examine a scene in the Primitives-Unions test set (union of a sphere and box), for which GNS's predictions display significant surface penetration, and increase surface particles from 2000 to 5000." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "14231" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 266, + 235 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 266, + 235 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 266, + 235 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 266, + 235 + ], + "type": "image", + "image_path": "d6d6c3ac613d53f68d3380410ae407fae369cd4d5714ed082d23af32a23e6110.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 245, + 276, + 258 + ], + "lines": [ + { + "bbox": [ + 58, + 245, + 276, + 258 + ], + "spans": [ + { + "bbox": [ + 58, + 245, + 276, + 258 + ], + "type": "text", + "content": "Figure 5: SURFSUP predictions on ShapeNet objects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 269, + 293, + 365 + ], + "blocks": [ + { + "bbox": [ + 50, + 269, + 293, + 365 + ], + "lines": [ + { + "bbox": [ + 50, + 269, + 293, + 365 + ], + "spans": [ + { + "bbox": [ + 50, + 269, + 293, + 365 + ], + "type": "image", + "image_path": "e12890e6b577710cd29589628ed05516253bc683865d57411114968c4c9d8302.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 218, + 547, + 338 + ], + "lines": [ + { + "bbox": [ + 304, + 218, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 547, + 338 + ], + "type": "text", + "content": "Figure 7: Comparing our method vs. the baseline with more surface particles. (a) Left shows how surface penetration is affected; even as the baseline improves with more surface particles, our method maintains an order magnitude improvement (red line, 490 particles). Right shows that the graph size increases significantly with the number of surface particles; we measure the max number of edges during the rollout, which indicates memory cost. (b) Final rollout timesteps for our method and the GNS at 2000, 3000, and 4000 particles, which reflect the quantitative results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 425, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 640 + ], + "type": "text", + "content": "Correspondingly GNS's fluid predictions improve, with less surface penetration (Fig. 7, top left and bottom). However, the baseline with 5000 particles still shows an order magnitude more penetration than our method (12285 vs. 490 particles). Moreover its computational cost, measured as the maximum particle graph size (# edges) during the rollout, almost doubles with 5000 surface particles (Fig. 7, top right). In contrast SURFSUP achieves near-zero surface-penetration with constant, smaller fluid graphs, overcoming the baseline's cost-performance tradeoff. We similarly analyze the room scene (Fig. 1), increasing the number of surface particles to 5K and 10K. Remarkably, SURFSUP here outperforms even a 10000-particle baseline (Ours 125.39 vs. baseline 194.98 Chamfer Surface; Supp.for details). This suggests fundamental advantages in scaling for simulation with implicit surfaces. We perform an inference time study (Supp.) showing that our method achieves speedups over GNS and the ground-truth simulator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Scaling up SURFSUP. As a final benchmark of unusual size and complexity, we assessed SURFSUP on a natural scene with complex topography (Puncak Jaya mountains, Indonesia). This scene was " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "40 \\times 27 \\times 6 \\, \\text{m}^3" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " in size (6840 m³ compared to roughly 10 m³ for previous scenes), and contained 68,600 fluid particles (order magnitude more than" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 435, + 140 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 435, + 140 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 435, + 140 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 435, + 140 + ], + "type": "image", + "image_path": "c371af9480e01e61b8acaa4a77163a3092fda8e9e707452ba9ecd97762c3c2a1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 311, + 142, + 545, + 210 + ], + "blocks": [ + { + "bbox": [ + 311, + 142, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 311, + 142, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 311, + 142, + 545, + 210 + ], + "type": "image", + "image_path": "78f72b7d0a30b416c34797665abf16019a263d25ada706c90a379143130ecf6e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 439, + 72, + 554, + 140 + ], + "blocks": [ + { + "bbox": [ + 439, + 72, + 554, + 140 + ], + "lines": [ + { + "bbox": [ + 439, + 72, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 439, + 72, + 554, + 140 + ], + "type": "image", + "image_path": "552256be592444af0fad124718e80260b43e23781b885f9b38b4491cf671bc07.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 352, + 558, + 575 + ], + "blocks": [ + { + "bbox": [ + 46, + 376, + 288, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 413 + ], + "type": "text", + "content": "Figure 6: Fine-grained comparison of SURFSUP and GNS on four parts of the room scene. Our dynamics accurately model the furniture while the baseline's leak into surfaces." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 352, + 558, + 575 + ], + "lines": [ + { + "bbox": [ + 308, + 352, + 558, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 352, + 558, + 575 + ], + "type": "image", + "image_path": "6ba7745f9abf574b5124c008e99dac88e100ebeb2a72516a9db68801b4434cd5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "text", + "content": "Figure 8: Predicted fluid rollouts from SURFSUP on different parts of a " + }, + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "inline_equation", + "content": "40 \\times 27 \\times 6" + }, + { + "bbox": [ + 305, + 582, + 545, + 607 + ], + "type": "text", + "content": " m³ mountain scene." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "previous scenes, see Supp. for details). Fig. 8 shows the predictions of SURFSUP, which correctly handle complex features such as slopes, confluences, and valleys between the mountains. Note that while GNS would require " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "O(1M)" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " particles to represent this scene, we leave the network size for SIREN unchanged from smaller scenes. These results demonstrate the potential of our method for large-scale, realistic dynamics. Please see Supp. for rollout videos." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14232" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "Inverse Design. Finally, we evaluate SURFSUP on its potential to solve object design tasks. We show the ability of our model to design both parametrized shapes, by optimizing the shape parameters with gradients, as well as shapes represented by neural SDF models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "text", + "content": "Parametrized Design Tasks. We consider the funnel, which is an inverted capped cone carved out of a cylinder, for which we have derived an analytical SDF. This shape is parametrized by the cylinder radius " + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "text", + "content": ", height " + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "text", + "content": ", cone height " + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "inline_equation", + "content": "h = H" + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "text", + "content": "; and the radii " + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "text", + "content": " of the bottom and top circles that \"cap\" the cone, with " + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "inline_equation", + "content": "r_1 > r_2" + }, + { + "bbox": [ + 46, + 134, + 287, + 207 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "text", + "content": "In our setup, we drop a block of fluid directly onto the object. We consider two tasks: designing a bowl where the task is to contain the fluid inside the object, and a funnel where the task is to concentrate the water onto a location onto the ground. We optimize " + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "text", + "content": " (which we call the \"capture radius\" that captures fluid into the object) and " + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "text", + "content": " (the \"filter radius\" which decides how much water to filter onto the ground). For the funnel, the desired solution is large " + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "text", + "content": " and small " + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "text", + "content": ", and for the bowl we want " + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "inline_equation", + "content": "r_2 \\rightarrow 0" + }, + { + "bbox": [ + 46, + 208, + 288, + 339 + ], + "type": "text", + "content": ". If our learned model is performing well, the model should be able to discover these solutions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "type": "text", + "content": "We roll out our model for 50 (bowl) and 75 (funnel) timesteps. Our reward measures the log probability of a Gaussian centered at the desired location of the fluid particles. For the bowl task, we use the bottom of the object, i.e. " + }, + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{N}((0.,0.,h),\\Sigma)" + }, + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "type": "text", + "content": " is the object height. For the funnel we use a 2D Gaussian " + }, + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{N}((0.,0.),\\Sigma)" + }, + { + "bbox": [ + 46, + 342, + 287, + 450 + ], + "type": "text", + "content": " and measure the reward on particles that reach the ground during training. For a given design, we roll out our model and obtain gradients, which we use to iteratively refine the design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "type": "text", + "content": "Qualitative results are shown in Fig. 9. Starting from a sub-optimal design with low " + }, + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "type": "inline_equation", + "content": "r_2 > 0" + }, + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "type": "text", + "content": ", our model successfully converges for the bowl; similarly the funnel converges from a suboptimal to an optimal solution. This shows that over long rollouts and hundreds of message-passing steps with dense fluid-surface interactions, our model's gradients are useful for solving inverse design problems. We also find that design is robust across different initializations of " + }, + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 46, + 451, + 288, + 560 + ], + "type": "text", + "content": " (details in Supp.)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Latent Space Design. We further examined whether SURF-SUP could directly optimize in the latent space of a DeepSDF model, trained on ShapeNet chairs. For this investigation, as a proof-of-concept we aimed to discover a novel chair that traps and contains falling water (i.e. maximizes the \"bowl reward\" above). Given a DeepSDF model " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}(p,z)" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " with latent code " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", we iteratively refine the design by rolling out our simulator for 100 time-steps, computing a reward, and updating " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " via gradient-based optimization. Across multiple trials, our model converges from an suboptimal initial chair to a final design that can contain water and retains the semantics of a chair (Fig. 10)." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 382, + 159 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 382, + 159 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 382, + 159 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 382, + 159 + ], + "type": "image", + "image_path": "5ca803e74f2c3bea259c6a6ba3ff5a72dc98f854e31885982d55bcb6120f0e23.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 455, + 163 + ], + "blocks": [ + { + "bbox": [ + 391, + 72, + 455, + 163 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 455, + 163 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 455, + 163 + ], + "type": "image", + "image_path": "68f77f40fb030f28717d6bfa71d3a2f3cd06d1ecaf763ad82430e86390afcc14.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 474, + 72, + 536, + 163 + ], + "blocks": [ + { + "bbox": [ + 474, + 72, + 536, + 163 + ], + "lines": [ + { + "bbox": [ + 474, + 72, + 536, + 163 + ], + "spans": [ + { + "bbox": [ + 474, + 72, + 536, + 163 + ], + "type": "image", + "image_path": "91dea91e98a968b4c31f3e8ac9df33ddf919237c6d436cd1b22e5cbfaaa228b5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 166, + 383, + 255 + ], + "blocks": [ + { + "bbox": [ + 306, + 166, + 383, + 255 + ], + "lines": [ + { + "bbox": [ + 306, + 166, + 383, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 166, + 383, + 255 + ], + "type": "image", + "image_path": "626cc9c5ff3c2f95e55a716f5b67e0e8e1d1d6cb671e11b0eb72342d03f5892c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 256, + 369, + 264 + ], + "lines": [ + { + "bbox": [ + 331, + 256, + 369, + 264 + ], + "spans": [ + { + "bbox": [ + 331, + 256, + 369, + 264 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 391, + 166, + 455, + 255 + ], + "blocks": [ + { + "bbox": [ + 391, + 166, + 455, + 255 + ], + "lines": [ + { + "bbox": [ + 391, + 166, + 455, + 255 + ], + "spans": [ + { + "bbox": [ + 391, + 166, + 455, + 255 + ], + "type": "image", + "image_path": "71fbad5635e23a684bd2bdaec298f0a8dc587f8e11151d40647b52c4ff6a5705.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 256, + 445, + 264 + ], + "lines": [ + { + "bbox": [ + 405, + 256, + 445, + 264 + ], + "spans": [ + { + "bbox": [ + 405, + 256, + 445, + 264 + ], + "type": "text", + "content": "Intermediate" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 474, + 166, + 536, + 255 + ], + "blocks": [ + { + "bbox": [ + 474, + 166, + 536, + 255 + ], + "lines": [ + { + "bbox": [ + 474, + 166, + 536, + 255 + ], + "spans": [ + { + "bbox": [ + 474, + 166, + 536, + 255 + ], + "type": "image", + "image_path": "d3a17ae70925fdab343280ed35aa90b68e6e2d3c49041a48cd252eb4b996ad15.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 256, + 512, + 264 + ], + "lines": [ + { + "bbox": [ + 495, + 256, + 512, + 264 + ], + "spans": [ + { + "bbox": [ + 495, + 256, + 512, + 264 + ], + "type": "text", + "content": "Final" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 547, + 346 + ], + "type": "text", + "content": "Figure 9: Design optimization; top is the funnel task, bottom the bowl task. For each frame, the design is shown near the top of the container and a timestep from the fluid rollout is shown below. Initial, intermediate, and final designs are shown; note that the reward increases through optimization and final designs approximate a funnel and bowl well." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 374, + 362, + 439 + ], + "blocks": [ + { + "bbox": [ + 310, + 356, + 358, + 368 + ], + "lines": [ + { + "bbox": [ + 310, + 356, + 358, + 368 + ], + "spans": [ + { + "bbox": [ + 310, + 356, + 358, + 368 + ], + "type": "text", + "content": "Initial Design" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 374, + 362, + 439 + ], + "lines": [ + { + "bbox": [ + 310, + 374, + 362, + 439 + ], + "spans": [ + { + "bbox": [ + 310, + 374, + 362, + 439 + ], + "type": "image", + "image_path": "a03ff94c8af6f1fbdff8aee63b1e62846b521f40cb104c172223cf09c04b8dae.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 388, + 374, + 442, + 438 + ], + "blocks": [ + { + "bbox": [ + 397, + 357, + 440, + 368 + ], + "lines": [ + { + "bbox": [ + 397, + 357, + 440, + 368 + ], + "spans": [ + { + "bbox": [ + 397, + 357, + 440, + 368 + ], + "type": "text", + "content": "Final Design" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 374, + 442, + 438 + ], + "lines": [ + { + "bbox": [ + 388, + 374, + 442, + 438 + ], + "spans": [ + { + "bbox": [ + 388, + 374, + 442, + 438 + ], + "type": "image", + "image_path": "94c16d356dcdaf51926be6fa5baa27f61d753250d04f43d47aeb7246d5b8a3bd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 455, + 376, + 544, + 441 + ], + "blocks": [ + { + "bbox": [ + 468, + 357, + 538, + 367 + ], + "lines": [ + { + "bbox": [ + 468, + 357, + 538, + 367 + ], + "spans": [ + { + "bbox": [ + 468, + 357, + 538, + 367 + ], + "type": "text", + "content": "Reward Convergence" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 455, + 376, + 544, + 441 + ], + "lines": [ + { + "bbox": [ + 455, + 376, + 544, + 441 + ], + "spans": [ + { + "bbox": [ + 455, + 376, + 544, + 441 + ], + "type": "image", + "image_path": "c0bdf3c72a59d9e696ca3cfe3af6a0eecced8146a4ddcae17b0e7041f42703b2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 450, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 487 + ], + "type": "text", + "content": "Figure 10: We differentiate through SURFSUP to directly optimize (in the latent space of a DeepSDF) a chair that holds water dropped onto it." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 491, + 379, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 379, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 379, + 502 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "content": "In this paper, we have taken a first step towards integrating implicit 3D representations into the dynamical simulation of physical models. SURFSUP realizes this approach by incorporating the implicit representation of solid surfaces (via SDFs) into a graph neural network fluid simulator. As such, SURFSUP is able to accurately model how fluids interact with surfaces, including highly complex surfaces unseen in training. We find that promoting this type of OOD generalization is helpful in both forward and inverse contexts." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": "Acknowledgements: We thank Huy Ha for help with visualizations of fluid rollouts. We thank Jan Bender for helpful feedback on SPlisHSPlasH, and Kelsey Allen, Tobias Pfaff, and Alvaro Sanchez-Gonzalez for advice on graph network simulators. We thank Max Helman for early setup work on SPlisHSPlasH. This research is partially supported by the NSF STC for Learning the Earth with Artificial Intelligence and Physics, and the NSF NRI Award #1925157. AM is supported by the NSF fellowship." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14233" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas J Guibas. Learning representations and generative models for 3d point clouds. arXiv preprint arXiv:1707.02392, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] Kelsey R Allen, Tatiana Lopez-Guevara, Kimberly Stachenfeld, Alvaro Sanchez-Gonzalez, Peter Battaglia, Jessica Hamrick, and Tobias Pfaff. Physical design using differentiable learned simulators. arXiv preprint arXiv:2202.00728, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "text", + "content": "[3] Peter Battaglia, Razvan Pascanu, Matthew Lai, Danilo Jimenez Rezende, et al. Interaction networks for learning about objects, relations and physics. Advances in neural information processing systems, 29, 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "text", + "content": "[4] Jan Bender and Dan Koschier. Divergence-free smoothed particle hydrodynamics. In Proceedings of the 14th ACM SIGGRAPH / Eurographics Symposium on Computer Animation, SCA '15, page 147-155, New York, NY, USA, 2015. Association for Computing Machinery." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "text", + "content": "[5] Jan Bender and Dan Koschier. Divergence-free smoothed particle hydrodynamics. In Proceedings of the 14th ACM SIGGRAPH / Eurographics Symposium on Computer Animation, SCA '15, page 147-155, New York, NY, USA, 2015. Association for Computing Machinery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "type": "text", + "content": "[6] Siddhartha Chaudhuri, Daniel Ritchie, Jiajun Wu, Kai Xu, and Hao Zhang. Learning generative models of 3d structures. In Computer Graphics Forum, volume 39, pages 643-666. Wiley Online Library, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 391, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 446 + ], + "type": "text", + "content": "[7] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 490 + ], + "type": "text", + "content": "[8] James D Foley, Foley Dan Van, Andries Van Dam, Steven K Feiner, and John F Hughes. Computer graphics: principles and practice, volume 12110. Addison-Wesley Professional, 1996." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 534 + ], + "type": "text", + "content": "[9] Christoph Gissler, Andreas Peer, Stefan Band, Jan Bender, and Matthias Teschner. Interlinked sph pressure solvers for strong fluid-rigid coupling. ACM Transactions on Graphics (TOG), 38(1):1-13, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[10] Zekun Hao, Hadar Averbuch-Elor, Noah Snavely, and Serge Belongie. Dualsdf: Semantic shape manipulation using a two-level representation. pages 7631–7641, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "text", + "content": "[11] Dan Koschier and Jan Bender. Density maps for improved sph boundary handling. In Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pages 1-10, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[12] Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. In ICLR, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Yunzhu Li, Jiajun Wu, Jun-Yan Zhu, Joshua B Tenenbaum, Antonio Torralba, and Russ Tedrake. Propagation networks for model-based control under partial observation. In 2019 International Conference on Robotics and Automation (ICRA), pages 1205-1211. IEEE, 2019." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Zhijian Liu, William T Freeman, Joshua B Tenenbaum, and Jiajun Wu. Physical primitive decomposition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-19, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[15] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[16] Andreas Mayr, Sebastian Lehner, Arno Mayrhofer, Christoph Kloss, Sepp Hochreiter, and Johannes Brandstetter. Boundary graph neural networks for 3d simulations. arXiv preprint arXiv:2106.11299, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "type": "text", + "content": "[17] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 304 + ], + "type": "text", + "content": "[18] Mariem Mezghanni, Théo Bodrito, Malika Boulkenafed, and Maks Ovsjanikov. Physical simulation layer for accurate 3d modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13514-13523, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 360 + ], + "type": "text", + "content": "[19] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 392 + ], + "type": "text", + "content": "[20] J. J. Monaghan. Smoothed particle hydrodynamics. Annual Review of Astronomy and Astrophysics, 30:543-574, Jan. 1992." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "text", + "content": "[21] Damian Mrowca, Chengxu Zhuang, Elias Wang, Nick Haber, Li F Fei-Fei, Josh Tenenbaum, and Daniel L Yamins. Flexible neural representation for physics prediction. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 437, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 492 + ], + "type": "text", + "content": "[22] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 536 + ], + "type": "text", + "content": "[23] Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh-based simulation with graph networks. In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 537, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 613 + ], + "type": "text", + "content": "[24] Edoardo Remelli, Artem Lukoianov, Stephan Richter, Benoit Guillard, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Meshsdf: Differentiable iso-surface extraction. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 22468-22478. Curran Associates, Inc., 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 668 + ], + "type": "text", + "content": "[25] Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In International Conference on Machine Learning, pages 8459-8468. PMLR, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[26] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE transactions on neural networks, 20(1):61-80, 2008." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14234" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 441 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[27] C. Schenck and D. Fox. Spnets: Differentiable fluid dynamics for deep neural networks. In Proceedings of the Second Conference on Robot Learning (CoRL), Zurich, Switzerland, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "text", + "content": "[28] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 185 + ], + "type": "text", + "content": "[29] Vincent Sitzmann, Eric R. Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In Proc. Neurips, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 228 + ], + "type": "text", + "content": "[30] Vincent Sitzmann, Julien N.P. Martel, Alexander W. Bergman, David B. Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Proc. NeurIPS, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 230, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 230, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 286, + 274 + ], + "type": "text", + "content": "[31] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In Computer Vision and Pattern Recognition (CVPR), 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 275, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 286, + 319 + ], + "type": "text", + "content": "[32] Benjamin Ummenhofer, Lukas Prantl, Nils Thuerey, and Vladlen Koltun. Lagrangian fluid simulation with continuous convolutions. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 320, + 286, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 286, + 374 + ], + "type": "text", + "content": "[33] Subeesh Vasu, Nicolas Talabot, Artem Lukoianov, Pierre Baqué, Jonathan Donier, and Pascal Fua. Hybridsdf: Combining deep implicit shapes and geometric primitives for 3d shape representation and manipulation. In International Conference on 3D Vision, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 376, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 286, + 441 + ], + "type": "text", + "content": "[34] Zhirong Wu, S. Song, A. Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and J. Xiao. 3d shapenets: A deep representation for volumetric shapes. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, Los Alamitos, CA, USA, jun 2015. IEEE Computer Society." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14235" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/0c733c34-bebb-47b3-8e25-0904ab5a04e1_content_list.json b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/0c733c34-bebb-47b3-8e25-0904ab5a04e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dba58419e316457c5f3c37e385afa9d37cd908ce --- /dev/null +++ b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/0c733c34-bebb-47b3-8e25-0904ab5a04e1_content_list.json @@ -0,0 +1,1590 @@ +[ + { + "type": "text", + "text": "SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving", + "text_level": 1, + "bbox": [ + 76, + 130, + 893, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Wei $^{1,2}$ , Linqing Zhao $^{3}$ , Wenzhao Zheng $^{1,2}$ , Zheng Zhu $^{4}$ , Jie Zhou $^{1,2}$ , Jiwen Lu $^{1,2\\dagger}$", + "bbox": [ + 156, + 178, + 818, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Beijing National Research Center for Information Science and Technology, China $^{2}$ Department of Automation, Tsinghua University, China", + "bbox": [ + 158, + 214, + 810, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Electrical and Information Engineering, Tianjin University, China $^{4}$ PhiGent Robotics", + "bbox": [ + 181, + 251, + 787, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{y-wei19, zhengwz18}@mails.tsinghua.edu.cn; linqingzhao@tju.edu.cn; zhengzhu@ieee.org; {jzhou, lujiwen}@tsinghua.edu.cn", + "bbox": [ + 187, + 289, + 769, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 356, + 313, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene understanding plays a vital role in vision-based autonomous driving. While most existing methods focus on 3D object detection, they have difficulty describing real-world objects of arbitrary shapes and infinite classes. Towards a more comprehensive perception of a 3D scene, in this paper, we propose a SurroundOcc method to predict the 3D occupancy with multi-camera images. We first extract multi-scale features for each image and adopt spatial 2D-3D attention to lift them to the 3D volume space. Then we apply 3D convolutions to progressively upsample the volume features and impose supervision on multiple levels. To obtain dense occupancy prediction, we design a pipeline to generate dense occupancy ground truth without expansive occupancy annotations. Specifically, we fuse multi-frame LiDAR scans of dynamic objects and static scenes separately. Then we adopt Poisson Reconstruction to fill the holes and voxelize the mesh to get dense occupancy labels. Extensive experiments on nuScenes and SemanticKITTI datasets demonstrate the superiority of our method. Code and dataset are available at https://github.com/weiyithu/SurroundOcc.", + "bbox": [ + 75, + 388, + 473, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 734, + 209, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding the 3D geometry of the surrounding scene serves as the basic procedure in an autonomous driving system. While LiDAR is a direct solution to capture this geometric information, it suffers from high-cost sensors and sparse scanned points, limiting its further application. Recently, vision-centric autonomous driving has attracted extensive attention as a promising direction. Taking multi-", + "bbox": [ + 75, + 758, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e033e84fd6801e416a851495c265db993a2a24650da8180cae346402ffac67be.jpg", + "image_caption": [ + "Figure 1. The overview of SurroundOcc. Given multi-camera images, our method can predict volumetric occupancy of surrounding 3D scenes. To train the network, we design a pipeline to generate dense occupancy labels with sparse LiDAR points. Better viewed when zoomed in." + ], + "image_footnote": [], + "bbox": [ + 509, + 356, + 879, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "camera images as inputs, it has demonstrated competitive performance in various 3D perception tasks including depth estimation [17, 58], 3D object detection [30, 27, 36, 34, 20], and semantic map construction [66, 19, 1].", + "bbox": [ + 496, + 685, + 892, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although multi-camera 3D object detection plays an important role [30, 27, 34, 20] in vision-based 3D perception, it is easy to suffer from the long-tail problem and difficult to recognize all classes of objects in the real world. Complementary to 3D detection, reconstructing surrounding 3D scenes can better help the downstream perception tasks. Recent works [17, 58] incorporate information from multiple views and predict surrounding depth maps. However, depth maps only predict the nearest occupied point in each optical ray and are unable to recover the occluded parts of the 3D", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21729", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scene. Different from depth-based methods, another trend [7, 21] is to directly predict the 3D occupancy of the scene. It describes a 3D scene by assigning an occupied probability to each voxel in the 3D space. We advocate 3D occupancy to be a good 3D representation for multi-camera scene reconstruction, which naturally guarantees the multicamera geometry consistency and is able to recover occluded parts. Also, it is flexible to extend to other 3D downstream tasks such as 3D semantic segmentation [69, 73, 15]. As one of the pioneering works, MonoScene [7] infers the dense 3D voxelized semantic scene with monocular images. However, simply fusing multi-camera results with cross-camera post-processing will lead to low performance [30]. TPVFormer [21] uses sparse LiDAR points as supervision, which results in sparse occupancy prediction.", + "bbox": [ + 80, + 90, + 467, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, we propose a SurroundOcc method, which aims to predict dense and accurate 3D occupancy with multi-camera images input. We first use a 2D backbone network to extract multi-scale feature maps from each image. Then we perform 2D-3D spatial attention to lift multi-camera image information to 3D volume features instead of BEV features. A 3D convolution network is then employed to progressively upsample the low-resolution volume features and fuse them with high-resolution ones to obtain fine-grained 3D representations. At each level, we use a decayed weighted loss to supervise the network. To get dense predictions, we need dense occupancy labels. However, the mainstream multi-camera dataset nuScenes [6] only provides sparse LiDAR points. To avoid expensive occupancy annotations, we devise a pipeline to generate dense occupancy ground truth only with the existing 3D detection and 3D semantic segmentation labels. Specifically, we first combine multi-frame points of dynamic objects and static scenes respectively. Then we leverage Poisson Reconstruction [23] algorithm to further fill the holes. Finally, NN and voxelization are used to obtain dense 3D occupancy labels.", + "bbox": [ + 80, + 318, + 467, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the dense occupancy ground truth, we train the model and compare it with other state-of-the-art methods on nuScenes [6] dataset. Both the quantitative results and visualizations demonstrate the effectiveness of our method. Moreover, we further conduct experiments on SemanticKITTI dataset [2]. Although our method is not designed for the monocular setting, it achieves state-of-the-art performance on the monocular 3D semantic scene completion benchmark.", + "bbox": [ + 80, + 635, + 467, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 784, + 215, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Voxel-based Scene Representation: How to effectively represent a 3D scene lies at the core of autonomous driving perception. Voxel-based scene representation voxelizes the 3D space into discretized voxels and describes each voxel by a vector feature. It has empowered the success of numerous methods on the lidar segmentation [33, 52, 11, 62, 61]", + "bbox": [ + 80, + 809, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and 3D scene completion [7, 46, 9, 25, 59, 29, 28] tasks. For the 3D occupancy prediction task, we also advocate the voxel representation as it is more suitable to model the occupancy field of a 3D scene. MonoScene [7] is the first work to reconstruct outdoor scenes using only RGB inputs. TPVFormer [21] further generalizes it to multi-camera 3D semantic occupancy prediction. However, its lack of dense supervision results in sparse occupancy prediction. Differently, we devise a pipeline to generate dense occupancy ground truth for training and our occupancy prediction is much denser.", + "bbox": [ + 503, + 90, + 888, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Scene Reconstruction: 3D reconstruction [42, 49, 48, 41, 71, 16, 4, 38, 50, 5] is a traditional but important topic in computer vision. One way is through depth estimation which predicts a depth value for each pixel in the image. While early methods require full depth annotations to supervise the depth estimation model [14, 24, 67], later research focuses on self-supervised depth estimation as it does not require intensive human annotations [71, 16, 4, 68, 63, 57, 70, 44, 8]. Recently, SurroundDepth [58] further incorporates the interactions between surrounding views to capture more spatial correlations. Different from depth estimation, 3D scene reconstruction methods [22, 38, 50, 5, 7] directly reconstruct a comprehensive and accurate 3D geometry of a scene. SurfaceNet [22] employs a 3D convolutional network to transform RGB colors to 3D surface occupancy from two images. Atlas [38] further extends it to the multi-view setting and utilized learned features to predict occupancy. NeuralRecon [50] and TransformerFusion [5] fuse the learned image features from different views in an online manner for more accurate 3D reconstructions. However, most of these 3D scene reconstruction methods are designed for indoor scenes, which are different from the multi-camera setting in the outdoor environment.", + "bbox": [ + 503, + 277, + 888, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-based 3D Perception: The lack of direct geometric input demands vision-based 3D surround perception methods to infer the 3D scene geometry based on semantic cues. Depth-based methods explicitly predict depth maps for image inputs to extract 3D geometric information of the scene before perception [55, 37, 45, 40, 27, 36, 31, 43, 20, 66]. The widely adopted pipeline is to predict categorical depth distributions and leverage them to project semantic features into 3D space [43]. Other methods implicitly learn 3D features without producing explicit depth maps [54, 30, 56, 35, 53, 64]. For example, BEV-Former [30] adopts cross attention to progressively refine BEV grid features from 2D image features. While most existing works employ BEV representation to describe a scene, we propose to reconstruct a 3D scene using volumetric occupancy representation, which provides the more fine-grained and comprehensive modeling of the scene.", + "bbox": [ + 503, + 643, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21730", + "bbox": [ + 478, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d38e3f70e6d739f038bff0ca54d376242c9fa9a65063ff3c44c6a2c7560b09f1.jpg", + "image_caption": [ + "Figure 2. The pipeline of the proposed method. First, we use a backbone to extract multi-scale features of multi-camera images. Then we adopt 2D-3D spatial attention to fuse multi-camera information and construct 3D volume features in a multi-scale fashion. Finally, the 3D deconvolution layer is used to upsample 3D volumes and occupancy prediction is supervised in each level." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 872, + 385 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/29ec71971466d5dc9bb444121d02f1aca7bdeb259bd16d67458e4c8d6471c8db.jpg", + "image_caption": [ + "Figure 3. The comparison of 3D-based and BEV-based cross-view attention. The 3D-based attention can better preserve 3D information. For each 3D volume query, we project it to the corresponding 2D views to sample features." + ], + "image_footnote": [], + "bbox": [ + 96, + 449, + 450, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 632, + 187, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 657, + 279, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we aim to predict the 3D occupancy of surrounding scenes with multi-camera images $I = \\{I^1, I^2, \\dots, I^N\\}$ . Formally, the 3D occupancy prediction is represented as:", + "bbox": [ + 76, + 680, + 468, + 742 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV = G \\left(I ^ {1}, I ^ {2}, \\dots I ^ {N}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 750, + 468, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $G$ is an neural network and $V \\in \\mathbb{R}^{H \\times W \\times Z}$ is the 3D occupancy. The value of $V$ is between 0 and 1, representing the occupied probability of the grids. Lifting $V$ to an $(L, H, W, Z)$ tensor, we can obtain the 3D semantic occupancy, where $L$ is the class number and class 0 means non-occupied grids.", + "bbox": [ + 75, + 777, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D occupancy is a good representation for multi-camera 3D scene reconstruction. First, since 3D occupancy is", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "predicted in 3D space, it theoretically satisfies the multicamera consistency. Second, it is possible for networks to predict occluded areas according to the surrounding semantic information, which is unavailable in depth estimation. Third, 3D occupancy is easy to extend to other downstream tasks, such as 3D semantic segmentation and scene flow estimation.", + "bbox": [ + 496, + 452, + 892, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 569, + 609, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows the pipeline of our method. Given a set of surrounding multi-camera images, we first use a backbone network (e.g. ResNet-101 [18]) to extract $N$ cameras' and $M$ levels' multi-scale features $X = \\{\\{X_{i}^{j}\\}_{i = 1}^{N}\\}_{j = 1}^{M}$ . For each level, we use a transformer to fuse multi-camera features with spatial cross attention. The output of the 2D-3D spatial attention layer is a 3D volume feature instead of the BEV feature. Then the 3D convolution network is utilized to upsample and combine multi-scale volume features. The occupancy prediction in each level is supervised by the generated dense occupancy ground truth with a decayed loss weight.", + "bbox": [ + 496, + 594, + 892, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. 2D-3D Spatial Attention", + "text_level": 1, + "bbox": [ + 500, + 786, + 723, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many 3D scene reconstruction methods [7, 38] integrate multi-view 2D features into 3D space by reprojecting 2D features back to the 3D volumes with known poses. The grid feature is calculated by simply averaging all 2D features in this grid. However, this kind of method assumes that different views contribute equally to the 3D volumes,", + "bbox": [ + 496, + 809, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21731", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2af4c56df14331738a606ad726512032134c97593e1d4b6bfafc82e536b4789e.jpg", + "image_caption": [ + "Figure 4. Dense occupancy ground truth generation. We first traverse all frames to stitch the multi-frame LiDAR points of dynamic objects and static scenes separately, and then merge them into a complete scene. Subsequently, we employ Poisson Reconstruction to densify the points and voxelize the resulting mesh to obtain a dense 3D occupancy. Finally, we use the Nearest Neighbor (NN) algorithm to assign semantic labels to dense voxels." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 870, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which does not always hold true, especially when some views are occluded or blurred.", + "bbox": [ + 76, + 444, + 470, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To tackle the issue, we leverage cross-view attention to fuse multi-camera features. We project 3D reference points to 2D views and use deformable attention [72, 30] to query points and aggregate information. As shown in Figure 3, instead of 2D BEV queries, we build 3D volume queries to further reserve 3D space information. Specifically, 3D volume queries are defined as $Q \\in \\mathbb{R}^{C \\times H \\times W \\times Z}$ . For each query, we project its corresponding 3D point to 2D views according to the given intrinsic and extrinsic. We only use the views that the 3D reference point hits. Then we sample 2D features around these projected 2D positions. The output $F \\in \\mathbb{R}^{C \\times H \\times W \\times Z}$ of the cross-view attention layer is a weighted sum of sampled features according to the deformable attention mechanism:", + "bbox": [ + 75, + 474, + 472, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {D e f o r m A t t n} (q, p, x) = \\sum_ {i = 1} ^ {N _ {\\text {h e a d}}} \\mathcal {W} _ {i} \\sum_ {j = 1} ^ {N _ {\\text {k e y}}} \\mathcal {A} _ {i j} \\cdot \\mathcal {W} _ {i} ^ {\\prime} x (p + \\Delta p _ {i j}) \\\\ F ^ {p} = \\frac {1}{\\left| \\mathcal {V} _ {\\mathrm {h i t}} \\right|} \\sum_ {i \\in \\mathcal {V} _ {\\mathrm {h i t}}} \\operatorname {D e f o r m A t t n} \\left(Q ^ {p}, \\mathcal {P} \\left(q ^ {p}, i\\right), X _ {i}\\right) \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 698, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F^p$ and $Q^p$ indicate the $p$ th element of the output features and 3D volume queries. $q^p$ is the corresponding 3D positions of queries and $P$ is the 3D to 2D project function. $\\mathcal{V}_{\\mathrm{hit}}$ represents the hit views of 3D query points. $W_i$ and $W_i'$ are the learnable weights and $A_{ij} \\in [0,1]$ is the attention weight calculated by the dot product of query and key. $x(p + \\Delta p_{ij})$ is the 2D feature at location $p + \\Delta p_{ij}$ . In-", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "stead of performing expensive 3D self-attention, we use the 3D convolution to interact features between neighboring 3D voxels.", + "bbox": [ + 498, + 444, + 892, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Multi-scale Occupancy Prediction", + "text_level": 1, + "bbox": [ + 500, + 503, + 797, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further extend the 2D-3D spatial attention to a multiscale fashion. Different from 3D detection task, 3D scene reconstruction needs more low-level features to help the network learn fine-grained details. To tackle the issue, we design a 2D-3D U-Net architecture. Specifically, given multi-scale 2D features $\\{\\{X_i^j\\}_{i=1}^N\\}_{j=1}^M$ , we adopt different number of 2D-3D spatial attention layers to extract multiscale 3D volume features $\\{F_j \\in \\mathbb{R}^{C_j \\times H_j \\times W_j \\times Z_j}\\}_{j=1}^M$ . Then we upsample $j - 1$ th level 3D volume features $Y_{j - 1}$ with 3D deconvolution layer and fuse it with $F_j$ :", + "bbox": [ + 496, + 529, + 893, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {j} = F _ {j} + \\operatorname {D e c o n v} \\left(Y _ {j - 1}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 703, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each level, the network outputs an occupancy prediction result with different resolution $V_{j} \\in \\mathbb{R}^{C_{j} \\times H_{j} \\times W_{j} \\times Z_{j}}$ . To get powerful both high-level and low-level 3D features, the network is supervised at each scale. Specifically, we use the cross-entropy loss and scene-class affinity loss introduced in [7] as the supervision signals. For 3D semantic occupancy prediction, we adopt a multi-class cross-entropy loss and for 3D scene reconstruction we change it to a two-class formulation. Since the high-resolution prediction is more important, we use a decayed loss weight $\\alpha_{j} = \\frac{1}{2^{j}}$ for $j$ th level supervision.", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21732", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/befe25c5c2be50f59761225271f81ae62a880c8ff4c3c50f89d2de21a7359483.jpg", + "image_caption": [ + "(a) RGB image" + ], + "image_footnote": [], + "bbox": [ + 91, + 94, + 269, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fe3c2fac10799e7f284f22f272e6938950a55124465693c3d19df49b0b2f7f8d.jpg", + "image_caption": [ + "(b) single-frame LiDAR points" + ], + "image_footnote": [], + "bbox": [ + 274, + 94, + 457, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/61a6689ca4744de07150ce9187367390cec506b8640b31025c40642f350bac87.jpg", + "image_caption": [ + "(c) Sparse occupancy labels" + ], + "image_footnote": [], + "bbox": [ + 91, + 214, + 269, + 306 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee806322d9ecd4e4c68cd87a944a03ea148711d27db15d27d89aca2cdb5044ec.jpg", + "image_caption": [ + "(d) Dense occupancy labels", + "Figure 5. Comparison on different occupancy labels. Compared with single-frame LiDAR points and the sparse occupancy converted from multi-frame points, our dense voxels are able to provide more realistic occupancy labels." + ], + "image_footnote": [], + "bbox": [ + 277, + 214, + 455, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Dense Occupancy Ground Truth", + "text_level": 1, + "bbox": [ + 76, + 398, + 375, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we find that the network supervised by sparse LiDAR points is unable to predict dense enough occupancy. Thus, it is necessary to generate dense occupancy labels. However, as mentioned in SemanticKITTI [2], it is complex and needs huge human efforts to annotate the dense occupancy of a 3D scene which has millions of voxels. To this end, we design a pipeline to generate dense occupancy ground truth leveraging existing 3D detection and 3D semantic segmentation labels without extra human annotations, which is shown in Figure 4. An intuitive way is to directly transform a multi-frame LiDAR point sequence into a unified coordinate system, and then voxelize the concatenated dense points into voxel grids. However, such a straightforward solution is only applicable to completely static scenes and ignores moving objects. Moreover, multiframe point clouds are not dense enough and there still exist many holes, which will result in wrong occupancy labels. To address these issues, we propose to stitch multi-frame LiDAR points of dynamic objects and static scenes separately. In addition, we adopt Poisson Reconstruction [23] to fill up the holes and voxelize the obtained mesh to get dense volumetric occupancy. Since the LiDAR scans surface points, our method also generates surface occupancy.", + "bbox": [ + 75, + 425, + 472, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Multi-frame Point Cloud Stitching", + "text_level": 1, + "bbox": [ + 76, + 785, + 377, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a two-stream pipeline to stitch static scenes and objects separately and merge them into a complete scene before voxelization. Specifically, for each frame, we first cut out movable objects from the LiDAR points according to 3D bounding box labels, so that we can obtain the 3D points of a static scene and movable objects. After travers-", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing all frames in the scene, we integrate the collected static scene segments and object segments into a set respectively. To combine the multi-frame segments, we then transform their coordinates into the world coordinate system via the known calibrated matrices and ego-poses. We denote the transformed static scene segments and object segments as $P_{ss} = \\{P_{ss}^{1}, P_{ss}^{2}, \\dots, P_{ss}^{n}\\}$ and $P_{os} = \\{P_{os}^{1}, P_{os}^{2}, \\dots, P_{os}^{m}\\}$ , where $n$ and $m$ are the numbers of frames and objects in the sequence, respectively. Note that the same objects in different frames can be recognized according to the bounding box index. Therefore, we can represent the whole static scene as $P_{s} = [P_{ss}^{1}, P_{ss}^{2}, \\dots, P_{ss}^{n}]$ while the objects as $P_{o} = [P_{os}^{1}, P_{os}^{2}, \\dots, P_{os}^{m}]$ , respectively, where $[\\cdot]$ is the concatenation operator. Finally, according to the objects' locations and ego-pose of the current frame, the 3D points of this frame can be obtained by merging static scene and objects: $P = [T_{s}(Ps), T_{o}(Po)]$ , where $T_{s}$ and $T_{o}$ are the transformations of static scenes and objects from world coordinate system to the current frame coordinate system. In this way, the occupancy labels of the current frame leverage the LiDAR points of all frames in the sequence.", + "bbox": [ + 496, + 90, + 893, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Densifying with Poisson Reconstruction", + "text_level": 1, + "bbox": [ + 498, + 417, + 839, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the density of $P$ is much larger than a single-frame LiDAR, there still exists many interspaces and the points are not evenly distributed, which is caused by the limited LiDAR beams. To address this, we first compute the normal vectors according to the spatial distribution in local neighborhoods. Then we reconstruct $P$ to a triangular mesh $\\mathcal{M}$ via Poisson Surface Reconstruction [23], whose input is point cloud with normal vectors, and the output is a triangular mesh (see Figure 4). The obtained mesh $\\mathcal{M} = \\{\\mathcal{V},\\mathcal{E}\\}$ fills up the holes of point clouds with evenly distributed vertices $\\mathcal{V}$ , so that we can further convert the mesh into dense voxel $V_{d}$ .", + "bbox": [ + 496, + 441, + 890, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Semantic Labeling with NN Algorithm", + "text_level": 1, + "bbox": [ + 498, + 633, + 831, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Having obtained the occupancy of dense voxels $V_{d}$ , we aim to assign semantic labels to each voxel, since position reconstruction can only be applied to 3D space, not semantic space. To this end, we propose to leverage Nearest Neighbors (NN) algorithm to search the nearest semantic label for each voxel. Specifically, we first voxelize $P$ with semantics into voxels $V_{s}$ , which are sparser than $V_{d}$ due to limited LiDAR beams. Then for each occupied voxel in $V_{d}$ , we use NN to search the nearest voxel in $V_{s}$ and assign the semantic label to it. Note that this process can be accelerated by parallel computing on the GPU. Thus, all occupied voxels in $V_{d}$ can obtain their semantic labels from $V_{s}$ .", + "bbox": [ + 496, + 657, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5 shows a detailed visual comparison between single-frame LiDAR points, sparse occupancy labels and dense occupancy labels. We observe that our dense voxels can provide much more realistic occupancy labels with", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21733", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/11d8037f397f4bf34fe9b6c5f5135deafc53f715c9b6731ebd7859dcdc4e23f6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSC IoUSSC mIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmanmadevegetation
MonoScene [7]23.967.314.030.358.008.042.900.281.160.674.014.3527.725.2015.1311.299.0314.86
Atlas [38]28.6615.0010.645.6819.6624.948.908.846.473.2810.4216.2134.8615.4621.8920.9511.2120.54
BEVFormer [30]30.5016.7514.226.5823.4628.288.6610.776.644.0511.2017.7837.2818.0022.8822.1713.8022.21
TPVFormer [21]11.5111.6616.147.1722.6317.138.8311.3910.468.239.4317.028.0713.6413.8510.344.907.37
TPVFormer*30.8617.1015.965.3123.8627.329.798.747.095.2010.9719.2238.8721.2524.2623.1511.7320.81
SurroundOcc31.4920.3020.5911.6828.0630.8610.7015.1414.0912.0614.3822.2637.2923.7024.4922.7714.8921.86
", + "bbox": [ + 80, + 88, + 893, + 247 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3f6cfb565dbbcbc1dd140bc2dc300618ac7edf40854253d8f9035dd2a71981fb.jpg", + "table_caption": [ + "Table 1. 3D semantic occupancy prediction results on nuScenes validation set. Except TPVFormer [21], all methods are trained with dense occupancy labels. To fairly compare, we further use dense ground truth to train the TPVFormer, which is denoted as TPVFormer*." + ], + "table_footnote": [], + "table_body": "
Accmeanp∈P(minp*∈P* ||p-p*||)
Compmeanp*∈P*(minp∈P ||p-p*)||
Precmeanp∈P(minp*∈P* ||p-p*||<0.5)
Recalmeanp*∈P*(minp∈P ||p-p*||<0.5)
CDAcc + Comp
F-score(2 × Prec × Recal)/(Prec + Recal)
", + "bbox": [ + 112, + 295, + 434, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Evaluation metrics for 3D scene reconstruction. $p$ and ${p}^{ * }$ are the predicted and ground truth point clouds.", + "bbox": [ + 75, + 387, + 467, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "clear semantic boundaries.", + "bbox": [ + 76, + 430, + 253, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We think it is not trivial to propagate the sparse semantic label to a dense one since it is a ill-posed problem. The proposed multi-frame point cloud stitching can aggregate multi-frame semantic information and provide dense enough reference points for NN. However, we find that NN is sensitive to the annotation noise in original LiDAR semantic labels, and we will try to solve it as the future work.", + "bbox": [ + 75, + 445, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 564, + 209, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 589, + 266, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset: We conduct multi-camera experiments on nuScenes dataset [6], which is a large-scale autonomous driving dataset. Since the 3D semantic and 3D detection labels are unavailable in test set and we cannot generate dense occupancy labels, we use the training set to train the model and validation set for evaluation. The occupancy prediction range is set as $[-50m, 50m]$ for $X, Y$ axis and $[-5m, 3m]$ for $Z$ axis. The final output occupancy has the shape $200x200x16$ with $0.5m$ voxel size. The input image resolution is $1600x900$ .", + "bbox": [ + 75, + 613, + 468, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further demonstrate the effectiveness of our method, we conduct monocular semantic scene completion experiment on SemanticKITTI dataset [2]. SemanticKITTI has annotated outdoor LiDAR scans with 21 semantic labels. The ground truth is voxelized as $256 \\times 256 \\times 32$ grid with $0.2 \\mathrm{~m}$ voxel size. We evaluate our model on the test set.", + "bbox": [ + 75, + 763, + 467, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: The whole network architecture obtains $M = 4$ levels and we do not add the skip connection in level 0. For nuScenes dataset, we adopt ResNet101-DCN", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[18, 13] with the initial weight from FCOS3D [54] as the backbone to extract image features. The features of stage 1,2,3 are fed to FPN [32] and used as multi-scale image features. The number of 2D-3D spatial attention layers are set as 1, 3, 6 for three levels. For SemanticKITTI dataset, following MonoScene [7], we use a pretrained EfficientNetB7 [51] as the backbone to generate multi-scale image features. We also adopt FPN to further fuse the features of different levels. We set the number of 2D-3D spatial attention layers as 1, 3, 8. All experiments are conducted on 8 RTX 3090s. For Possion Reconstruction, we accumulate both key-frame and non key-frame data. In details, we use LiDAR frames at $20\\mathrm{Hz}$ , resulting in 400 frames and around 13 million points for a 20 second sequence.", + "bbox": [ + 496, + 297, + 890, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: For 3D semantic occupancy prediction, we use the intersection over union (IoU) of occupied voxels, ignoring their semantic class as the evaluation metric of the scene completion (SC) task and the mIoU of all semantic classes for the SSC task.", + "bbox": [ + 496, + 517, + 890, + 590 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {I o U} = \\frac {T P}{T P + F P + F N} \\\\ \\mathrm {m I o U} = \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\frac {T P _ {i}}{T P _ {i} + F P _ {i} + F N _ {i}} \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 614, + 890, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $TP$ , $FP$ , $FN$ indicate the number of true positive, false positive, and false negative predictions. $C$ is the class number.", + "bbox": [ + 496, + 712, + 890, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For 3D scene reconstruction, we first convert occupancy prediction to point clouds. Following [38, 5], we use 3D metrics for evaluation, which is shown in Table 2. Chamfer distance (CD) and F-score are the main metrics since they consider both precision and recall. Please refer to the supplementary for more evaluation metric details. For all evaluation, we adopt dense occupancy as ground truth since sparse LiDAR points cannot fully evaluate the quality of occupancy reconstruction.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21734", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8c19caf5a1437ad4c76d12b820e5cf6cd8428b9c2d6deccea63b3f17977ab220.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSC IoUSSC mIoUroad (15.36%)sidewalk (11.13%)parking (1.12%)other-gmd (0.56%)building (0.41%)car (3.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation (39.5%)trunk (0.51%)terrain (9.17%)person (0.07%)bicyclist (0.07%)motorcyclist. (0.05%)fence (3.90%)pole (0.29%)traf.-sign (0.08%)
LMSCNet [47]31.387.0746.7019.5013.503.1010.3014.300.300.000.0010.800.0010.400.000.000.005.400.000.00
3DSketch [10]26.856.2337.7019.800.000.0012.1017.100.000.000.0012.100.0016.100.000.000.003.400.000.00
AICNet [26]23.937.0939.3018.3019.801.609.6015.300.700.000.009.601.9013.500.000.000.005.000.100.00
JS3C-Net [60]34.008.9747.3021.7019.902.8012.7020.100.800.000.004.1014.203.1012.400.000.208.701.900.30
MonoScene [7]34.1611.0854.7027.1024.805.7014.4018.803.300.500.704.4014.902.4019.501.001.400.4011.103.302.10
TPVFormer [21]34.2511.2655.1027.2027.406.5014.8019.203.701.000.502.3013.902.6020.401.102.400.3011.002.901.50
SurroundOcc34.7211.8656.9028.3030.206.8015.2020.601.401.601.204.4014.903.4019.301.402.000.1011.303.902.40
", + "bbox": [ + 94, + 89, + 877, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56757af439a200be096222c7ef88a3bad8357f44f1df00d37ba264a46fcffe98.jpg", + "image_caption": [ + "Figure 6. An example of challenging scenes. Although the quality of RGB images degrades in rainy days and nights, our method can still predict detailed occupancy. Better viewed when zoomed in." + ], + "image_footnote": [ + "driveable surface", + "car bus", + "truck terrain", + "Vegetation", + "sidewalk", + "other flat", + "pedestrian", + "□ bicycle", + "manmade", + "motorcycle", + "barrier", + "construction vehicle", + "■trailer", + "traffic cone" + ], + "bbox": [ + 84, + 315, + 480, + 486 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7778f66dab2dfc31cf808705a282c890eadaf684494049addac93a1051620c2e.jpg", + "table_caption": [ + "Table 3. Monocular Semantic scene completion results on SemanticKITTI test set. For fair comparison, we use the performances of RGB-inferred versions of the first four methods, which are reported in MonoScene [7]. Although our method is not designed for monocular perception, we still outperform other methods for a large margin." + ], + "table_footnote": [], + "table_body": "
MethodAcc ↓Comp ↓Prec ↑Recall ↑CD ↓F-score ↑
SurroundDepth [58]1.7471.3840.2610.3533.1300.293
AdaBins [3]1.9891.2870.2330.3473.2750.271
NeWCRFs [65]2.1631.2330.2140.3483.3960.257
Atlas [38]0.6791.6850.4070.5462.3650.458
TransformerFusion [5]0.7711.4340.3750.5912.2050.453
SurroundOcc0.7241.2260.4140.6021.9500.483
", + "bbox": [ + 81, + 571, + 475, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. 3D scene reconstruction results on nuScenes validation set. F-score and CD are the main metrics.", + "bbox": [ + 75, + 683, + 468, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. 3D Semantic Occupancy Prediction", + "text_level": 1, + "bbox": [ + 76, + 724, + 383, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first conduct multi-camera 3D semantic occupancy prediction on nuScenes [6] dataset and compare with several state-of-the-art methods [7, 30, 21, 38]. For MonoScene [7], we concatenate the multi-camera occupancy predictions and voxelize them with the $0.5\\mathrm{m}$ voxel size, which is same as our setting. For BEVFormer [30], we add a 3D segmentation head to predict semantic occupancy. As shown in Table 1, our method achieves state-of-the-art performance. We also show some qualitative results in Figure 6 and Figure 7. See supplementary material for more", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "video demos and qualitative comparisons.", + "bbox": [ + 500, + 573, + 777, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Especially in Figure 6, we show rainy day and night visualization. Although the quality of RGB images degrades in these two challenging scenes, our method can still predict fine details. The LiDAR sensor suffers from rainy days and easily misses points. With multi-frame aggregation and Possion Reconstruction, we dramatically densify the labels and provide strong supervision, which is crucial for the challenging scenarios. Moreover, we conduct color jitter augmentation, which increases the robustness of brightness change. We also note that some parts cannot be observed by LiDAR sensor, such as the back side of the car. Surprisingly, our model can predict the complete shape according to the surrounding information.", + "bbox": [ + 496, + 593, + 890, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further demonstrate the superiority of our method, we also conduct monocular 3D semantic scene completion on SemanticKITTI dataset [2]. Table 3 shows the results. Although our method is not designed for monocular perception and cross-view attention will be ineffective for the monocular setting, our method still achieves state-of-the-art performance on this benchmark.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21735", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/49dbfd2783c00604cd068d40974c91bf58f0640594968081a3bfef3af483f971.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSC IoUSSC mIoU
w/o spatial attention29.7817.34
BEV-based attention30.4518.94
Ours31.4920.30
", + "bbox": [ + 124, + 88, + 419, + 159 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5a8db4dfd2321d8ea64fea70143884fde2ab374abfebb61041a1137da82ec1bf.jpg", + "table_caption": [ + "Table 5. The ablation study of 2D-3D spatial attention. \"w/o spatial attention\" indicates that we average all multi-camera features in a grid." + ], + "table_footnote": [], + "table_body": "
MethodSC IoUSSC mIoU
w/o multi-scale structure30.4118.22
w/o multi-scale supervision31.1619.73
Ours31.4920.30
", + "bbox": [ + 99, + 218, + 444, + 291 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/91c1707f7207cd49eb2a50bfe4fd0e543b8a0d34a33ebce95b2000d22dd405c7.jpg", + "table_caption": [ + "Table 6. The ablation study of multi-scale occupancy prediction. \"w/o multi-scale structure\" means that we do not add multi-scale skip connection." + ], + "table_footnote": [], + "table_body": "
SupervisionSC IoUSSC mIoU
sparse LiDAR points11.9612.17
sparse occupancy labels30.5818.83
dense occupancy labels31.4920.30
", + "bbox": [ + 112, + 349, + 431, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. 3D Scene Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 482, + 303, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another important application of 3D occupancy prediction is 3D scene reconstruction. Due to this reason, we further evaluate 3D reconstruction performance without using multi-class semantic labels. We do the comparison with state-of-the-art multi-camera depth estimation methods (SurroundDepth [58]), monocular depth estimation methods (AdaBins [3] and NeWCRFs [65]) and 3D reconstruction method (Atlas [38] and TransformerFusion [5]). For the self-supervised methods SurroundDepth [58], we use depth ground truth to supervise them. To evaluate depth estimation methods in 3D space, following [38, 5], we run TSDF fusion [12, 39] to fuse multi-camera depth as point clouds. Note that we fuse the multi-camera depth maps of the same timestamp but not the multi-frame depths. Thus, there is no need to specially deal with movable objects. As shown in Table 4, SurroundOcc achieves state-of-the-art performance on most metrics, and outperforms other methods by a large margin, which verifies the effectiveness of the proposed method.", + "bbox": [ + 75, + 505, + 467, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 801, + 230, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2D-3D Spatial Attention: Table 5 shows the ablation results for 2D-3D spatial attention. Without spatial attention, we directly average all multi-camera features in a grid. However, we find this straightforward fusion method performs worse than spatial attention. The potential reason is", + "bbox": [ + 76, + 824, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that the contribution of each view is different for a 3D grid. Moreover, the ablation study shows that 3D-based cross-view attention is more effective than BEV-based cross-view attention since it can preserve 3D space information.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-scale Occupancy Prediction: We conduct an ablation study on multi-scale structure and multi-scale supervision in Table 6. The experimental results show that these two multi-scale designs can boost the performance. The multi-scale skip connection can help the network learn low-level fine-grained features, which is important for detailed high-resolution occupancy prediction. Moreover, the 3D volume features in all levels will be enhanced by multi-scale supervision.", + "bbox": [ + 496, + 151, + 890, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Dense Occupancy Supervision: The results in Table 7 demonstrate the importance of using dense occupancy as ground truth. Compared with sparse LiDAR points, the sparse occupancy labels fuse multi-frame points and can provide more powerful supervision. The possion reconstruction and NN algorithm can fill the holes and further densify the occupancy labels, which will boost model's performance. Figure 7 can better illustrate the effectiveness of dense supervision. We can see that the model trained with our occupancy labels can predict much denser occupancy than that trained with LiDAR points.", + "bbox": [ + 496, + 287, + 890, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Model Efficiency", + "text_level": 1, + "bbox": [ + 500, + 463, + 665, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the inference time and inference memory of different methods in Table 8. The experiments are conducted on one RTX 3090 with six multi-camera images, whose resolutions are $1600 \\times 900$ . We find that our method can achieve both high performance and efficiency. Compared with BEVFormer, our method slightly increases inference time and memory and we think the increased burden is acceptable", + "bbox": [ + 496, + 486, + 890, + 607 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f1f8c22a4fee7eeb9819d4c9a94d42750155aeb95bda4dd58ad720713edde0e0.jpg", + "table_caption": [ + "Table 7. The ablation study of dense occupancy supervision. The model trained with our dense occupancy ground truth is much better than that trained with sparse LiDAR points." + ], + "table_footnote": [], + "table_body": "
MethodLatency (s)Memory (G)
SurroundDepth [58]0.7312.4
NeWCRFs [65]1.0714.5
Adabins [3]0.7515.5
BEVFormer [30]0.314.5
TPVFormer [21]0.325.1
MonoScene [7]0.8720.3
Ours0.345.9
", + "bbox": [ + 521, + 619, + 867, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. The model efficiency of different methods. The experiments are conducted on one RTX 3090 with six multi-camera images, whose resolutions are $1600 \\times 900$ .", + "bbox": [ + 500, + 743, + 890, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 500, + 814, + 772, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Currently, we only explore single-frame occupancy prediction. However, for the downstream modules, e.g. motion prediction and planning, occupancy flow is more important. As the future work, we will design a framework to build", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21736", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1de631bc0bc46376cd8b433450b678318c793058c142c4a6a6d8315a796d275.jpg", + "image_caption": [ + "Figure 7. Visualizations on nuScenes validation set. Our generated dense occupancy labels are much denser than sparse LiDAR points. Trained with dense groundtruth, the network can predict better and denser occupancy. Better viewed when zoomed in." + ], + "image_footnote": [], + "bbox": [ + 78, + 79, + 906, + 681 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "occupancy flow dataset and utilize multi-frame surrounding images as the inputs. Moreover, LiDAR data is not always available. Self-supervised occupancy prediction with only RGB data is a valuable but challenging direction.", + "bbox": [ + 75, + 724, + 470, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 799, + 194, + 815 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose SurroundOcc for multi-camera 3D occupancy prediction. We utilize 2D-3D spatial attention to integrate 2D features to the 3D volume in a multiscale fashion, which is further upsampled and fused by the 3D deconvolution layer. Moreover, we devise a pipeline to", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "generate dense occupancy ground truth. We stitch multi-frame LiDAR points of dynamic objects and static scenes separately and utilize Poisson Reconstruction to fill the holes. The comparison on nuScenes and SemanticKITTI datasets demonstrates the superiority of our method.", + "bbox": [ + 496, + 724, + 890, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 815, + 660, + 832 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603.", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21737", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. arXiv preprint arXiv:2203.13641, 2022.", + "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019.", + "[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021.", + "[4] Jia-Wang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised Scale-consistent Depth and Ego-motion Learning from Monocular Video. In NeurIPS, pages 35-45, 2019.", + "[5] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021.", + "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020.", + "[7] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022.", + "[8] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018.", + "[9] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020.", + "[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020.", + "[11] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021.", + "[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In SIGGRAPH, pages 303-312, 1996.", + "[13] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, pages 764-773, 2017.", + "[14] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In ICCV, pages 2650-2658, 2015.", + "[15] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In ICCV, pages 2918-2927, 2021.", + "[16] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In ICCV, pages 3828-3838, 2019." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Vitor Guizilini, Igor Vasiljevic, Rares Ambrus, Greg Shakhnarovich, and Adrien Gaidon. Full surround monodepth from multiple cameras. IEEE Robotics and Automation Letters, 7(2):5397-5404, 2022.", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016.", + "[19] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, pages 15273-15282, 2021.", + "[20] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021.", + "[21] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817, 2023.", + "[22] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In ICCV, pages 2307-2315, 2017.", + "[23] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006.", + "[24] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019.", + "[25] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020.", + "[26] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020.", + "[27] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022.", + "[28] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, et al. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023.", + "[29] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023.", + "[30] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chong-hao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022.", + "[31] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21738", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022.", + "[32] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017.", + "[33] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. Amvnet: Assertion-based multi-view fusion network for lidar semantic segmentation. arXiv preprint arXiv:2012.04934, 2020.", + "[34] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In ECCV, pages 531-548. Springer, 2022.", + "[35] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022.", + "[36] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022.", + "[37] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In ICCV, 2019.", + "[38] Zak Murez, Tarrence van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In ECCV, pages 414-431, 2020.", + "[39] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In ISMAR, pages 127-136, 2011.", + "[40] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In ICCV, 2021.", + "[41] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019.", + "[42] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, pages 523-540. Springer, 2020.", + "[43] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210, 2020.", + "[44] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive Collaboration: Joint Unsupervised Learning of Depth, Camera Motion, Optical Flow and Motion Segmentation. In CVPR, pages 12240-12249, 2019.", + "[45] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In CVPR, 2021.", + "[46] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020." + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119. IEEE, 2020.", + "[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019.", + "[49] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017.", + "[50] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In CVPR, pages 15598-15607, 2021.", + "[51] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114. PMLR, 2019.", + "[52] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In ECCV, pages 685–702, 2020.", + "[53] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In CoRL, 2022.", + "[54] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In ICCV, 2021.", + "[55] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, 2019.", + "[56] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin M. Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In CoRL, 2021.", + "[57] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In ICCV, pages 379–387, 2017.", + "[58] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In CoRL, 2022.", + "[59] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, volume 35, pages 3101-3109, 2021.", + "[60] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, volume 35, pages 3101-3109, 2021.", + "[61] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet:" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21739", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022.", + "[62] Maosheng Ye, Rui Wan, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Drinet++: Efficient voxel-as-point point cloud segmentation. arXiv preprint arXiv: 2111.08318, 2021.", + "[63] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In CVPR, pages 1983-1992, 2018.", + "[64] Chen Yongjian, Tai Lei, Sun Kai, and Li Mingyang. Monopair: Monocular 3d object detection using pairwise spatial relationships. In CVPR, 2020.", + "[65] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. New crfs: Neural window fully-connected crfs for monocular depth estimation. In CVPR, 2022.", + "[66] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022.", + "[67] Zhenyu Zhang, Chunyan Xu, Jian Yang, Junbin Gao, and Zhen Cui. Progressive hard-mining network for monocular depth estimation. TIP, 27(8):3691-3702, 2018.", + "[68] Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. Towards Better Generalization: Joint Depth-Pose Learning without PoseNet. In CVPR, 2020.", + "[69] Hui Zhou, Xinge Zhu, Xiao Song, Yuexin Ma, Zhe Wang, Hongsheng Li, and Dahua Lin. Cylinder3d: An effective 3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550, 2020.", + "[70] Junsheng Zhou, Yuwang Wang, Kaihuai Qin, and Wenjun Zeng. Moving Indoor: Unsupervised Video Depth Learning in Challenging Environments. In ICCV, pages 8618-8627, 2019.", + "[71] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851-1858, 2017.", + "[72] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020.", + "[73] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021." + ], + "bbox": [ + 78, + 90, + 468, + 702 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21740", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/0c733c34-bebb-47b3-8e25-0904ab5a04e1_model.json b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/0c733c34-bebb-47b3-8e25-0904ab5a04e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5024a6fc0bb662a7b1d34395a6fa415d7fd4f9ba --- /dev/null +++ b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/0c733c34-bebb-47b3-8e25-0904ab5a04e1_model.json @@ -0,0 +1,2578 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.131, + 0.895, + 0.156 + ], + "angle": 0, + "content": "SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.179, + 0.819, + 0.202 + ], + "angle": 0, + "content": "Yi Wei\\(^{1,2}\\), Linqing Zhao\\(^{3}\\), Wenzhao Zheng\\(^{1,2}\\), Zheng Zhu\\(^{4}\\), Jie Zhou\\(^{1,2}\\), Jiwen Lu\\(^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.215, + 0.812, + 0.251 + ], + "angle": 0, + "content": "\\(^{1}\\)Beijing National Research Center for Information Science and Technology, China \\(^{2}\\)Department of Automation, Tsinghua University, China" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.252, + 0.788, + 0.286 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Electrical and Information Engineering, Tianjin University, China \\(^{4}\\)PhiGent Robotics" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.29, + 0.77, + 0.324 + ], + "angle": 0, + "content": "{y-wei19, zhengwz18}@mails.tsinghua.edu.cn; linqingzhao@tju.edu.cn; zhengzhu@ieee.org; {jzhou, lujiwen}@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.357, + 0.314, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.474, + 0.707 + ], + "angle": 0, + "content": "3D scene understanding plays a vital role in vision-based autonomous driving. While most existing methods focus on 3D object detection, they have difficulty describing real-world objects of arbitrary shapes and infinite classes. Towards a more comprehensive perception of a 3D scene, in this paper, we propose a SurroundOcc method to predict the 3D occupancy with multi-camera images. We first extract multi-scale features for each image and adopt spatial 2D-3D attention to lift them to the 3D volume space. Then we apply 3D convolutions to progressively upsample the volume features and impose supervision on multiple levels. To obtain dense occupancy prediction, we design a pipeline to generate dense occupancy ground truth without expansive occupancy annotations. Specifically, we fuse multi-frame LiDAR scans of dynamic objects and static scenes separately. Then we adopt Poisson Reconstruction to fill the holes and voxelize the mesh to get dense occupancy labels. Extensive experiments on nuScenes and SemanticKITTI datasets demonstrate the superiority of our method. Code and dataset are available at https://github.com/weiyithu/SurroundOcc." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.735, + 0.21, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Understanding the 3D geometry of the surrounding scene serves as the basic procedure in an autonomous driving system. While LiDAR is a direct solution to capture this geometric information, it suffers from high-cost sensors and sparse scanned points, limiting its further application. Recently, vision-centric autonomous driving has attracted extensive attention as a promising direction. Taking multi-" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.357, + 0.88, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.603, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Figure 1. The overview of SurroundOcc. Given multi-camera images, our method can predict volumetric occupancy of surrounding 3D scenes. To train the network, we design a pipeline to generate dense occupancy labels with sparse LiDAR points. Better viewed when zoomed in." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.893, + 0.746 + ], + "angle": 0, + "content": "camera images as inputs, it has demonstrated competitive performance in various 3D perception tasks including depth estimation [17, 58], 3D object detection [30, 27, 36, 34, 20], and semantic map construction [66, 19, 1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Although multi-camera 3D object detection plays an important role [30, 27, 34, 20] in vision-based 3D perception, it is easy to suffer from the long-tail problem and difficult to recognize all classes of objects in the real world. Complementary to 3D detection, reconstructing surrounding 3D scenes can better help the downstream perception tasks. Recent works [17, 58] incorporate information from multiple views and predict surrounding depth maps. However, depth maps only predict the nearest occupied point in each optical ray and are unable to recover the occluded parts of the 3D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21729" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.317 + ], + "angle": 0, + "content": "scene. Different from depth-based methods, another trend [7, 21] is to directly predict the 3D occupancy of the scene. It describes a 3D scene by assigning an occupied probability to each voxel in the 3D space. We advocate 3D occupancy to be a good 3D representation for multi-camera scene reconstruction, which naturally guarantees the multicamera geometry consistency and is able to recover occluded parts. Also, it is flexible to extend to other 3D downstream tasks such as 3D semantic segmentation [69, 73, 15]. As one of the pioneering works, MonoScene [7] infers the dense 3D voxelized semantic scene with monocular images. However, simply fusing multi-camera results with cross-camera post-processing will lead to low performance [30]. TPVFormer [21] uses sparse LiDAR points as supervision, which results in sparse occupancy prediction." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.319, + 0.468, + 0.634 + ], + "angle": 0, + "content": "To address this, we propose a SurroundOcc method, which aims to predict dense and accurate 3D occupancy with multi-camera images input. We first use a 2D backbone network to extract multi-scale feature maps from each image. Then we perform 2D-3D spatial attention to lift multi-camera image information to 3D volume features instead of BEV features. A 3D convolution network is then employed to progressively upsample the low-resolution volume features and fuse them with high-resolution ones to obtain fine-grained 3D representations. At each level, we use a decayed weighted loss to supervise the network. To get dense predictions, we need dense occupancy labels. However, the mainstream multi-camera dataset nuScenes [6] only provides sparse LiDAR points. To avoid expensive occupancy annotations, we devise a pipeline to generate dense occupancy ground truth only with the existing 3D detection and 3D semantic segmentation labels. Specifically, we first combine multi-frame points of dynamic objects and static scenes respectively. Then we leverage Poisson Reconstruction [23] algorithm to further fill the holes. Finally, NN and voxelization are used to obtain dense 3D occupancy labels." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.636, + 0.468, + 0.77 + ], + "angle": 0, + "content": "With the dense occupancy ground truth, we train the model and compare it with other state-of-the-art methods on nuScenes [6] dataset. Both the quantitative results and visualizations demonstrate the effectiveness of our method. Moreover, we further conduct experiments on SemanticKITTI dataset [2]. Although our method is not designed for the monocular setting, it achieves state-of-the-art performance on the monocular 3D semantic scene completion benchmark." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.785, + 0.216, + 0.8 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.81, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Voxel-based Scene Representation: How to effectively represent a 3D scene lies at the core of autonomous driving perception. Voxel-based scene representation voxelizes the 3D space into discretized voxels and describes each voxel by a vector feature. It has empowered the success of numerous methods on the lidar segmentation [33, 52, 11, 62, 61]" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.256 + ], + "angle": 0, + "content": "and 3D scene completion [7, 46, 9, 25, 59, 29, 28] tasks. For the 3D occupancy prediction task, we also advocate the voxel representation as it is more suitable to model the occupancy field of a 3D scene. MonoScene [7] is the first work to reconstruct outdoor scenes using only RGB inputs. TPVFormer [21] further generalizes it to multi-camera 3D semantic occupancy prediction. However, its lack of dense supervision results in sparse occupancy prediction. Differently, we devise a pipeline to generate dense occupancy ground truth for training and our occupancy prediction is much denser." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.278, + 0.89, + 0.623 + ], + "angle": 0, + "content": "3D Scene Reconstruction: 3D reconstruction [42, 49, 48, 41, 71, 16, 4, 38, 50, 5] is a traditional but important topic in computer vision. One way is through depth estimation which predicts a depth value for each pixel in the image. While early methods require full depth annotations to supervise the depth estimation model [14, 24, 67], later research focuses on self-supervised depth estimation as it does not require intensive human annotations [71, 16, 4, 68, 63, 57, 70, 44, 8]. Recently, SurroundDepth [58] further incorporates the interactions between surrounding views to capture more spatial correlations. Different from depth estimation, 3D scene reconstruction methods [22, 38, 50, 5, 7] directly reconstruct a comprehensive and accurate 3D geometry of a scene. SurfaceNet [22] employs a 3D convolutional network to transform RGB colors to 3D surface occupancy from two images. Atlas [38] further extends it to the multi-view setting and utilized learned features to predict occupancy. NeuralRecon [50] and TransformerFusion [5] fuse the learned image features from different views in an online manner for more accurate 3D reconstructions. However, most of these 3D scene reconstruction methods are designed for indoor scenes, which are different from the multi-camera setting in the outdoor environment." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.645, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Vision-based 3D Perception: The lack of direct geometric input demands vision-based 3D surround perception methods to infer the 3D scene geometry based on semantic cues. Depth-based methods explicitly predict depth maps for image inputs to extract 3D geometric information of the scene before perception [55, 37, 45, 40, 27, 36, 31, 43, 20, 66]. The widely adopted pipeline is to predict categorical depth distributions and leverage them to project semantic features into 3D space [43]. Other methods implicitly learn 3D features without producing explicit depth maps [54, 30, 56, 35, 53, 64]. For example, BEV-Former [30] adopts cross attention to progressively refine BEV grid features from 2D image features. While most existing works employ BEV representation to describe a scene, we propose to reconstruct a 3D scene using volumetric occupancy representation, which provides the more fine-grained and comprehensive modeling of the scene." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21730" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.874, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.39, + 0.893, + 0.434 + ], + "angle": 0, + "content": "Figure 2. The pipeline of the proposed method. First, we use a backbone to extract multi-scale features of multi-camera images. Then we adopt 2D-3D spatial attention to fuse multi-camera information and construct 3D volume features in a multi-scale fashion. Finally, the 3D deconvolution layer is used to upsample 3D volumes and occupancy prediction is supervised in each level." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.45, + 0.452, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.565, + 0.471, + 0.622 + ], + "angle": 0, + "content": "Figure 3. The comparison of 3D-based and BEV-based cross-view attention. The 3D-based attention can better preserve 3D information. For each 3D volume query, we project it to the corresponding 2D views to sample features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.633, + 0.188, + 0.65 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.658, + 0.281, + 0.673 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.681, + 0.47, + 0.743 + ], + "angle": 0, + "content": "In this work, we aim to predict the 3D occupancy of surrounding scenes with multi-camera images \\( I = \\{I^1, I^2, \\dots, I^N\\} \\). Formally, the 3D occupancy prediction is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.751, + 0.469, + 0.77 + ], + "angle": 0, + "content": "\\[\nV = G \\left(I ^ {1}, I ^ {2}, \\dots I ^ {N}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.469, + 0.87 + ], + "angle": 0, + "content": "where \\( G \\) is an neural network and \\( V \\in \\mathbb{R}^{H \\times W \\times Z} \\) is the 3D occupancy. The value of \\( V \\) is between 0 and 1, representing the occupied probability of the grids. Lifting \\( V \\) to an \\( (L, H, W, Z) \\) tensor, we can obtain the 3D semantic occupancy, where \\( L \\) is the class number and class 0 means non-occupied grids." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "3D occupancy is a good representation for multi-camera 3D scene reconstruction. First, since 3D occupancy is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.893, + 0.558 + ], + "angle": 0, + "content": "predicted in 3D space, it theoretically satisfies the multicamera consistency. Second, it is possible for networks to predict occluded areas according to the surrounding semantic information, which is unavailable in depth estimation. Third, 3D occupancy is easy to extend to other downstream tasks, such as 3D semantic segmentation and scene flow estimation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.57, + 0.611, + 0.585 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.893, + 0.776 + ], + "angle": 0, + "content": "Figure 2 shows the pipeline of our method. Given a set of surrounding multi-camera images, we first use a backbone network (e.g. ResNet-101 [18]) to extract \\(N\\) cameras' and \\(M\\) levels' multi-scale features \\(X = \\{\\{X_{i}^{j}\\}_{i = 1}^{N}\\}_{j = 1}^{M}\\). For each level, we use a transformer to fuse multi-camera features with spatial cross attention. The output of the 2D-3D spatial attention layer is a 3D volume feature instead of the BEV feature. Then the 3D convolution network is utilized to upsample and combine multi-scale volume features. The occupancy prediction in each level is supervised by the generated dense occupancy ground truth with a decayed loss weight." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.724, + 0.803 + ], + "angle": 0, + "content": "3.3. 2D-3D Spatial Attention" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Many 3D scene reconstruction methods [7, 38] integrate multi-view 2D features into 3D space by reprojecting 2D features back to the 3D volumes with known poses. The grid feature is calculated by simply averaging all 2D features in this grid. However, this kind of method assumes that different views contribute equally to the 3D volumes," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21731" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.872, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.377, + 0.895, + 0.435 + ], + "angle": 0, + "content": "Figure 4. Dense occupancy ground truth generation. We first traverse all frames to stitch the multi-frame LiDAR points of dynamic objects and static scenes separately, and then merge them into a complete scene. Subsequently, we employ Poisson Reconstruction to densify the points and voxelize the resulting mesh to obtain a dense 3D occupancy. Finally, we use the Nearest Neighbor (NN) algorithm to assign semantic labels to dense voxels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.445, + 0.471, + 0.474 + ], + "angle": 0, + "content": "which does not always hold true, especially when some views are occluded or blurred." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.688 + ], + "angle": 0, + "content": "To tackle the issue, we leverage cross-view attention to fuse multi-camera features. We project 3D reference points to 2D views and use deformable attention [72, 30] to query points and aggregate information. As shown in Figure 3, instead of 2D BEV queries, we build 3D volume queries to further reserve 3D space information. Specifically, 3D volume queries are defined as \\( Q \\in \\mathbb{R}^{C \\times H \\times W \\times Z} \\). For each query, we project its corresponding 3D point to 2D views according to the given intrinsic and extrinsic. We only use the views that the 3D reference point hits. Then we sample 2D features around these projected 2D positions. The output \\( F \\in \\mathbb{R}^{C \\times H \\times W \\times Z} \\) of the cross-view attention layer is a weighted sum of sampled features according to the deformable attention mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.699, + 0.47, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {D e f o r m A t t n} (q, p, x) = \\sum_ {i = 1} ^ {N _ {\\text {h e a d}}} \\mathcal {W} _ {i} \\sum_ {j = 1} ^ {N _ {\\text {k e y}}} \\mathcal {A} _ {i j} \\cdot \\mathcal {W} _ {i} ^ {\\prime} x (p + \\Delta p _ {i j}) \\\\ F ^ {p} = \\frac {1}{\\left| \\mathcal {V} _ {\\mathrm {h i t}} \\right|} \\sum_ {i \\in \\mathcal {V} _ {\\mathrm {h i t}}} \\operatorname {D e f o r m A t t n} \\left(Q ^ {p}, \\mathcal {P} \\left(q ^ {p}, i\\right), X _ {i}\\right) \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.904 + ], + "angle": 0, + "content": "where \\( F^p \\) and \\( Q^p \\) indicate the \\( p \\)th element of the output features and 3D volume queries. \\( q^p \\) is the corresponding 3D positions of queries and \\( P \\) is the 3D to 2D project function. \\( \\mathcal{V}_{\\mathrm{hit}} \\) represents the hit views of 3D query points. \\( W_i \\) and \\( W_i' \\) are the learnable weights and \\( A_{ij} \\in [0,1] \\) is the attention weight calculated by the dot product of query and key. \\( x(p + \\Delta p_{ij}) \\) is the 2D feature at location \\( p + \\Delta p_{ij} \\). In-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.445, + 0.893, + 0.49 + ], + "angle": 0, + "content": "stead of performing expensive 3D self-attention, we use the 3D convolution to interact features between neighboring 3D voxels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.799, + 0.523 + ], + "angle": 0, + "content": "3.4. Multi-scale Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.895, + 0.683 + ], + "angle": 0, + "content": "We further extend the 2D-3D spatial attention to a multiscale fashion. Different from 3D detection task, 3D scene reconstruction needs more low-level features to help the network learn fine-grained details. To tackle the issue, we design a 2D-3D U-Net architecture. Specifically, given multi-scale 2D features \\(\\{\\{X_i^j\\}_{i=1}^N\\}_{j=1}^M\\), we adopt different number of 2D-3D spatial attention layers to extract multiscale 3D volume features \\(\\{F_j \\in \\mathbb{R}^{C_j \\times H_j \\times W_j \\times Z_j}\\}_{j=1}^M\\). Then we upsample \\(j - 1\\)th level 3D volume features \\(Y_{j - 1}\\) with 3D deconvolution layer and fuse it with \\(F_j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.704, + 0.892, + 0.722 + ], + "angle": 0, + "content": "\\[\nY _ {j} = F _ {j} + \\operatorname {D e c o n v} \\left(Y _ {j - 1}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For each level, the network outputs an occupancy prediction result with different resolution \\( V_{j} \\in \\mathbb{R}^{C_{j} \\times H_{j} \\times W_{j} \\times Z_{j}} \\). To get powerful both high-level and low-level 3D features, the network is supervised at each scale. Specifically, we use the cross-entropy loss and scene-class affinity loss introduced in [7] as the supervision signals. For 3D semantic occupancy prediction, we adopt a multi-class cross-entropy loss and for 3D scene reconstruction we change it to a two-class formulation. Since the high-resolution prediction is more important, we use a decayed loss weight \\( \\alpha_{j} = \\frac{1}{2^{j}} \\) for \\( j \\)th level supervision." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21732" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.095, + 0.271, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.192, + 0.224, + 0.205 + ], + "angle": 0, + "content": "(a) RGB image" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.095, + 0.458, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.192, + 0.451, + 0.205 + ], + "angle": 0, + "content": "(b) single-frame LiDAR points" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.215, + 0.27, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.313, + 0.256, + 0.325 + ], + "angle": 0, + "content": "(c) Sparse occupancy labels" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.215, + 0.456, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.313, + 0.442, + 0.325 + ], + "angle": 0, + "content": "(d) Dense occupancy labels" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.328, + 0.47, + 0.384 + ], + "angle": 0, + "content": "Figure 5. Comparison on different occupancy labels. Compared with single-frame LiDAR points and the sparse occupancy converted from multi-frame points, our dense voxels are able to provide more realistic occupancy labels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.399, + 0.377, + 0.417 + ], + "angle": 0, + "content": "4. Dense Occupancy Ground Truth" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.473, + 0.775 + ], + "angle": 0, + "content": "In our experiments, we find that the network supervised by sparse LiDAR points is unable to predict dense enough occupancy. Thus, it is necessary to generate dense occupancy labels. However, as mentioned in SemanticKITTI [2], it is complex and needs huge human efforts to annotate the dense occupancy of a 3D scene which has millions of voxels. To this end, we design a pipeline to generate dense occupancy ground truth leveraging existing 3D detection and 3D semantic segmentation labels without extra human annotations, which is shown in Figure 4. An intuitive way is to directly transform a multi-frame LiDAR point sequence into a unified coordinate system, and then voxelize the concatenated dense points into voxel grids. However, such a straightforward solution is only applicable to completely static scenes and ignores moving objects. Moreover, multiframe point clouds are not dense enough and there still exist many holes, which will result in wrong occupancy labels. To address these issues, we propose to stitch multi-frame LiDAR points of dynamic objects and static scenes separately. In addition, we adopt Poisson Reconstruction [23] to fill up the holes and voxelize the obtained mesh to get dense volumetric occupancy. Since the LiDAR scans surface points, our method also generates surface occupancy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.379, + 0.803 + ], + "angle": 0, + "content": "4.1. Multi-frame Point Cloud Stitching" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We propose a two-stream pipeline to stitch static scenes and objects separately and merge them into a complete scene before voxelization. Specifically, for each frame, we first cut out movable objects from the LiDAR points according to 3D bounding box labels, so that we can obtain the 3D points of a static scene and movable objects. After travers-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.41 + ], + "angle": 0, + "content": "ing all frames in the scene, we integrate the collected static scene segments and object segments into a set respectively. To combine the multi-frame segments, we then transform their coordinates into the world coordinate system via the known calibrated matrices and ego-poses. We denote the transformed static scene segments and object segments as \\( P_{ss} = \\{P_{ss}^{1}, P_{ss}^{2}, \\dots, P_{ss}^{n}\\} \\) and \\( P_{os} = \\{P_{os}^{1}, P_{os}^{2}, \\dots, P_{os}^{m}\\} \\), where \\( n \\) and \\( m \\) are the numbers of frames and objects in the sequence, respectively. Note that the same objects in different frames can be recognized according to the bounding box index. Therefore, we can represent the whole static scene as \\( P_{s} = [P_{ss}^{1}, P_{ss}^{2}, \\dots, P_{ss}^{n}] \\) while the objects as \\( P_{o} = [P_{os}^{1}, P_{os}^{2}, \\dots, P_{os}^{m}] \\), respectively, where \\( [\\cdot] \\) is the concatenation operator. Finally, according to the objects' locations and ego-pose of the current frame, the 3D points of this frame can be obtained by merging static scene and objects: \\( P = [T_{s}(Ps), T_{o}(Po)] \\), where \\( T_{s} \\) and \\( T_{o} \\) are the transformations of static scenes and objects from world coordinate system to the current frame coordinate system. In this way, the occupancy labels of the current frame leverage the LiDAR points of all frames in the sequence." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.419, + 0.84, + 0.435 + ], + "angle": 0, + "content": "4.2. Densifying with Poisson Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.625 + ], + "angle": 0, + "content": "While the density of \\( P \\) is much larger than a single-frame LiDAR, there still exists many interspaces and the points are not evenly distributed, which is caused by the limited LiDAR beams. To address this, we first compute the normal vectors according to the spatial distribution in local neighborhoods. Then we reconstruct \\( P \\) to a triangular mesh \\( \\mathcal{M} \\) via Poisson Surface Reconstruction [23], whose input is point cloud with normal vectors, and the output is a triangular mesh (see Figure 4). The obtained mesh \\( \\mathcal{M} = \\{\\mathcal{V},\\mathcal{E}\\} \\) fills up the holes of point clouds with evenly distributed vertices \\( \\mathcal{V} \\), so that we can further convert the mesh into dense voxel \\( V_{d} \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.635, + 0.833, + 0.652 + ], + "angle": 0, + "content": "4.3. Semantic Labeling with NN Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Having obtained the occupancy of dense voxels \\( V_{d} \\), we aim to assign semantic labels to each voxel, since position reconstruction can only be applied to 3D space, not semantic space. To this end, we propose to leverage Nearest Neighbors (NN) algorithm to search the nearest semantic label for each voxel. Specifically, we first voxelize \\( P \\) with semantics into voxels \\( V_{s} \\), which are sparser than \\( V_{d} \\) due to limited LiDAR beams. Then for each occupied voxel in \\( V_{d} \\), we use NN to search the nearest voxel in \\( V_{s} \\) and assign the semantic label to it. Note that this process can be accelerated by parallel computing on the GPU. Thus, all occupied voxels in \\( V_{d} \\) can obtain their semantic labels from \\( V_{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Figure 5 shows a detailed visual comparison between single-frame LiDAR points, sparse occupancy labels and dense occupancy labels. We observe that our dense voxels can provide much more realistic occupancy labels with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21733" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.895, + 0.248 + ], + "angle": 0, + "content": "
MethodSC IoUSSC mIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmanmadevegetation
MonoScene [7]23.967.314.030.358.008.042.900.281.160.674.014.3527.725.2015.1311.299.0314.86
Atlas [38]28.6615.0010.645.6819.6624.948.908.846.473.2810.4216.2134.8615.4621.8920.9511.2120.54
BEVFormer [30]30.5016.7514.226.5823.4628.288.6610.776.644.0511.2017.7837.2818.0022.8822.1713.8022.21
TPVFormer [21]11.5111.6616.147.1722.6317.138.8311.3910.468.239.4317.028.0713.6413.8510.344.907.37
TPVFormer*30.8617.1015.965.3123.8627.329.798.747.095.2010.9719.2238.8721.2524.2623.1511.7320.81
SurroundOcc31.4920.3020.5911.6828.0630.8610.7015.1414.0912.0614.3822.2637.2923.7024.4922.7714.8921.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.253, + 0.895, + 0.283 + ], + "angle": 0, + "content": "Table 1. 3D semantic occupancy prediction results on nuScenes validation set. Except TPVFormer [21], all methods are trained with dense occupancy labels. To fairly compare, we further use dense ground truth to train the TPVFormer, which is denoted as TPVFormer*." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.296, + 0.436, + 0.386 + ], + "angle": 0, + "content": "
Accmeanp∈P(minp*∈P* ||p-p*||)
Compmeanp*∈P*(minp∈P ||p-p*)||
Precmeanp∈P(minp*∈P* ||p-p*||<0.5)
Recalmeanp*∈P*(minp∈P ||p-p*||<0.5)
CDAcc + Comp
F-score(2 × Prec × Recal)/(Prec + Recal)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.388, + 0.468, + 0.416 + ], + "angle": 0, + "content": "Table 2. Evaluation metrics for 3D scene reconstruction. \\( p \\) and \\( {p}^{ * } \\) are the predicted and ground truth point clouds." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.254, + 0.444 + ], + "angle": 0, + "content": "clear semantic boundaries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.552 + ], + "angle": 0, + "content": "We think it is not trivial to propagate the sparse semantic label to a dense one since it is a ill-posed problem. The proposed multi-frame point cloud stitching can aggregate multi-frame semantic information and provide dense enough reference points for NN. However, we find that NN is sensitive to the annotation noise in original LiDAR semantic labels, and we will try to solve it as the future work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.565, + 0.21, + 0.582 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.267, + 0.607 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Dataset: We conduct multi-camera experiments on nuScenes dataset [6], which is a large-scale autonomous driving dataset. Since the 3D semantic and 3D detection labels are unavailable in test set and we cannot generate dense occupancy labels, we use the training set to train the model and validation set for evaluation. The occupancy prediction range is set as \\([-50m, 50m]\\) for \\(X, Y\\) axis and \\([-5m, 3m]\\) for \\(Z\\) axis. The final output occupancy has the shape \\(200x200x16\\) with \\(0.5m\\) voxel size. The input image resolution is \\(1600x900\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.854 + ], + "angle": 0, + "content": "To further demonstrate the effectiveness of our method, we conduct monocular semantic scene completion experiment on SemanticKITTI dataset [2]. SemanticKITTI has annotated outdoor LiDAR scans with 21 semantic labels. The ground truth is voxelized as \\(256 \\times 256 \\times 32\\) grid with \\(0.2 \\mathrm{~m}\\) voxel size. We evaluate our model on the test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Implementation Details: The whole network architecture obtains \\( M = 4 \\) levels and we do not add the skip connection in level 0. For nuScenes dataset, we adopt ResNet101-DCN" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[18, 13] with the initial weight from FCOS3D [54] as the backbone to extract image features. The features of stage 1,2,3 are fed to FPN [32] and used as multi-scale image features. The number of 2D-3D spatial attention layers are set as 1, 3, 6 for three levels. For SemanticKITTI dataset, following MonoScene [7], we use a pretrained EfficientNetB7 [51] as the backbone to generate multi-scale image features. We also adopt FPN to further fuse the features of different levels. We set the number of 2D-3D spatial attention layers as 1, 3, 8. All experiments are conducted on 8 RTX 3090s. For Possion Reconstruction, we accumulate both key-frame and non key-frame data. In details, we use LiDAR frames at \\(20\\mathrm{Hz}\\), resulting in 400 frames and around 13 million points for a 20 second sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.592 + ], + "angle": 0, + "content": "Evaluation Metrics: For 3D semantic occupancy prediction, we use the intersection over union (IoU) of occupied voxels, ignoring their semantic class as the evaluation metric of the scene completion (SC) task and the mIoU of all semantic classes for the SSC task." + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.615, + 0.892, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {I o U} = \\frac {T P}{T P + F P + F N} \\\\ \\mathrm {m I o U} = \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\frac {T P _ {i}}{T P _ {i} + F P _ {i} + F N _ {i}} \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.757 + ], + "angle": 0, + "content": "where \\( TP \\), \\( FP \\), \\( FN \\) indicate the number of true positive, false positive, and false negative predictions. \\( C \\) is the class number." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For 3D scene reconstruction, we first convert occupancy prediction to point clouds. Following [38, 5], we use 3D metrics for evaluation, which is shown in Table 2. Chamfer distance (CD) and F-score are the main metrics since they consider both precision and recall. Please refer to the supplementary for more evaluation metric details. For all evaluation, we adopt dense occupancy as ground truth since sparse LiDAR points cannot fully evaluate the quality of occupancy reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21734" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.878, + 0.259 + ], + "angle": 0, + "content": "
MethodSC IoUSSC mIoUroad (15.36%)sidewalk (11.13%)parking (1.12%)other-gmd (0.56%)building (0.41%)car (3.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation (39.5%)trunk (0.51%)terrain (9.17%)person (0.07%)bicyclist (0.07%)motorcyclist. (0.05%)fence (3.90%)pole (0.29%)traf.-sign (0.08%)
LMSCNet [47]31.387.0746.7019.5013.503.1010.3014.300.300.000.0010.800.0010.400.000.000.005.400.000.00
3DSketch [10]26.856.2337.7019.800.000.0012.1017.100.000.000.0012.100.0016.100.000.000.003.400.000.00
AICNet [26]23.937.0939.3018.3019.801.609.6015.300.700.000.009.601.9013.500.000.000.005.000.100.00
JS3C-Net [60]34.008.9747.3021.7019.902.8012.7020.100.800.000.004.1014.203.1012.400.000.208.701.900.30
MonoScene [7]34.1611.0854.7027.1024.805.7014.4018.803.300.500.704.4014.902.4019.501.001.400.4011.103.302.10
TPVFormer [21]34.2511.2655.1027.2027.406.5014.8019.203.701.000.502.3013.902.6020.401.102.400.3011.002.901.50
SurroundOcc34.7211.8656.9028.3030.206.8015.2020.601.401.601.204.4014.903.4019.301.402.000.1011.303.902.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.262, + 0.894, + 0.304 + ], + "angle": 0, + "content": "Table 3. Monocular Semantic scene completion results on SemanticKITTI test set. For fair comparison, we use the performances of RGB-inferred versions of the first four methods, which are reported in MonoScene [7]. Although our method is not designed for monocular perception, we still outperform other methods for a large margin." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.316, + 0.481, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.109, + 0.491, + 0.214, + 0.504 + ], + "angle": 0, + "content": "driveable surface" + }, + { + "type": "image_footnote", + "bbox": [ + 0.232, + 0.491, + 0.321, + 0.504 + ], + "angle": 0, + "content": "car bus" + }, + { + "type": "image_footnote", + "bbox": [ + 0.336, + 0.491, + 0.45, + 0.504 + ], + "angle": 0, + "content": "truck terrain" + }, + { + "type": "image_footnote", + "bbox": [ + 0.462, + 0.491, + 0.533, + 0.504 + ], + "angle": 0, + "content": "Vegetation" + }, + { + "type": "image_footnote", + "bbox": [ + 0.551, + 0.491, + 0.614, + 0.504 + ], + "angle": 0, + "content": "sidewalk" + }, + { + "type": "image_footnote", + "bbox": [ + 0.631, + 0.491, + 0.696, + 0.504 + ], + "angle": 0, + "content": "other flat" + }, + { + "type": "image_footnote", + "bbox": [ + 0.715, + 0.491, + 0.783, + 0.504 + ], + "angle": 0, + "content": "pedestrian" + }, + { + "type": "image_footnote", + "bbox": [ + 0.8, + 0.491, + 0.856, + 0.504 + ], + "angle": 0, + "content": "□ bicycle" + }, + { + "type": "image_footnote", + "bbox": [ + 0.109, + 0.511, + 0.221, + 0.523 + ], + "angle": 0, + "content": "manmade" + }, + { + "type": "image_footnote", + "bbox": [ + 0.251, + 0.511, + 0.326, + 0.523 + ], + "angle": 0, + "content": "motorcycle" + }, + { + "type": "image_footnote", + "bbox": [ + 0.349, + 0.511, + 0.402, + 0.523 + ], + "angle": 0, + "content": "barrier" + }, + { + "type": "image_footnote", + "bbox": [ + 0.426, + 0.511, + 0.545, + 0.523 + ], + "angle": 0, + "content": "construction vehicle" + }, + { + "type": "image_footnote", + "bbox": [ + 0.569, + 0.511, + 0.619, + 0.523 + ], + "angle": 0, + "content": "■trailer" + }, + { + "type": "image_footnote", + "bbox": [ + 0.642, + 0.511, + 0.718, + 0.523 + ], + "angle": 0, + "content": "traffic cone" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.531, + 0.893, + 0.56 + ], + "angle": 0, + "content": "Figure 6. An example of challenging scenes. Although the quality of RGB images degrades in rainy days and nights, our method can still predict detailed occupancy. Better viewed when zoomed in." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.572, + 0.477, + 0.677 + ], + "angle": 0, + "content": "
MethodAcc ↓Comp ↓Prec ↑Recall ↑CD ↓F-score ↑
SurroundDepth [58]1.7471.3840.2610.3533.1300.293
AdaBins [3]1.9891.2870.2330.3473.2750.271
NeWCRFs [65]2.1631.2330.2140.3483.3960.257
Atlas [38]0.6791.6850.4070.5462.3650.458
TransformerFusion [5]0.7711.4340.3750.5912.2050.453
SurroundOcc0.7241.2260.4140.6021.9500.483
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.71 + ], + "angle": 0, + "content": "Table 4. 3D scene reconstruction results on nuScenes validation set. F-score and CD are the main metrics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.385, + 0.742 + ], + "angle": 0, + "content": "5.2. 3D Semantic Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We first conduct multi-camera 3D semantic occupancy prediction on nuScenes [6] dataset and compare with several state-of-the-art methods [7, 30, 21, 38]. For MonoScene [7], we concatenate the multi-camera occupancy predictions and voxelize them with the \\(0.5\\mathrm{m}\\) voxel size, which is same as our setting. For BEVFormer [30], we add a 3D segmentation head to predict semantic occupancy. As shown in Table 1, our method achieves state-of-the-art performance. We also show some qualitative results in Figure 6 and Figure 7. See supplementary material for more" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.574, + 0.779, + 0.589 + ], + "angle": 0, + "content": "video demos and qualitative comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Especially in Figure 6, we show rainy day and night visualization. Although the quality of RGB images degrades in these two challenging scenes, our method can still predict fine details. The LiDAR sensor suffers from rainy days and easily misses points. With multi-frame aggregation and Possion Reconstruction, we dramatically densify the labels and provide strong supervision, which is crucial for the challenging scenarios. Moreover, we conduct color jitter augmentation, which increases the robustness of brightness change. We also note that some parts cannot be observed by LiDAR sensor, such as the back side of the car. Surprisingly, our model can predict the complete shape according to the surrounding information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To further demonstrate the superiority of our method, we also conduct monocular 3D semantic scene completion on SemanticKITTI dataset [2]. Table 3 shows the results. Although our method is not designed for monocular perception and cross-view attention will be ineffective for the monocular setting, our method still achieves state-of-the-art performance on this benchmark." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21735" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.125, + 0.089, + 0.421, + 0.16 + ], + "angle": 0, + "content": "
MethodSC IoUSSC mIoU
w/o spatial attention29.7817.34
BEV-based attention30.4518.94
Ours31.4920.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.165, + 0.47, + 0.207 + ], + "angle": 0, + "content": "Table 5. The ablation study of 2D-3D spatial attention. \"w/o spatial attention\" indicates that we average all multi-camera features in a grid." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.219, + 0.445, + 0.292 + ], + "angle": 0, + "content": "
MethodSC IoUSSC mIoU
w/o multi-scale structure30.4118.22
w/o multi-scale supervision31.1619.73
Ours31.4920.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.296, + 0.468, + 0.339 + ], + "angle": 0, + "content": "Table 6. The ablation study of multi-scale occupancy prediction. \"w/o multi-scale structure\" means that we do not add multi-scale skip connection." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.35, + 0.432, + 0.422 + ], + "angle": 0, + "content": "
SupervisionSC IoUSSC mIoU
sparse LiDAR points11.9612.17
sparse occupancy labels30.5818.83
dense occupancy labels31.4920.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.427, + 0.468, + 0.469 + ], + "angle": 0, + "content": "Table 7. The ablation study of dense occupancy supervision. The model trained with our dense occupancy ground truth is much better than that trained with sparse LiDAR points." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.483, + 0.305, + 0.497 + ], + "angle": 0, + "content": "5.3. 3D Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.468, + 0.793 + ], + "angle": 0, + "content": "Another important application of 3D occupancy prediction is 3D scene reconstruction. Due to this reason, we further evaluate 3D reconstruction performance without using multi-class semantic labels. We do the comparison with state-of-the-art multi-camera depth estimation methods (SurroundDepth [58]), monocular depth estimation methods (AdaBins [3] and NeWCRFs [65]) and 3D reconstruction method (Atlas [38] and TransformerFusion [5]). For the self-supervised methods SurroundDepth [58], we use depth ground truth to supervise them. To evaluate depth estimation methods in 3D space, following [38, 5], we run TSDF fusion [12, 39] to fuse multi-camera depth as point clouds. Note that we fuse the multi-camera depth maps of the same timestamp but not the multi-frame depths. Thus, there is no need to specially deal with movable objects. As shown in Table 4, SurroundOcc achieves state-of-the-art performance on most metrics, and outperforms other methods by a large margin, which verifies the effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.231, + 0.818 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "2D-3D Spatial Attention: Table 5 shows the ablation results for 2D-3D spatial attention. Without spatial attention, we directly average all multi-camera features in a grid. However, we find this straightforward fusion method performs worse than spatial attention. The potential reason is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "that the contribution of each view is different for a 3D grid. Moreover, the ablation study shows that 3D-based cross-view attention is more effective than BEV-based cross-view attention since it can preserve 3D space information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Multi-scale Occupancy Prediction: We conduct an ablation study on multi-scale structure and multi-scale supervision in Table 6. The experimental results show that these two multi-scale designs can boost the performance. The multi-scale skip connection can help the network learn low-level fine-grained features, which is important for detailed high-resolution occupancy prediction. Moreover, the 3D volume features in all levels will be enhanced by multi-scale supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.455 + ], + "angle": 0, + "content": "- Dense Occupancy Supervision: The results in Table 7 demonstrate the importance of using dense occupancy as ground truth. Compared with sparse LiDAR points, the sparse occupancy labels fuse multi-frame points and can provide more powerful supervision. The possion reconstruction and NN algorithm can fill the holes and further densify the occupancy labels, which will boost model's performance. Figure 7 can better illustrate the effectiveness of dense supervision. We can see that the model trained with our occupancy labels can predict much denser occupancy than that trained with LiDAR points." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.464, + 0.666, + 0.48 + ], + "angle": 0, + "content": "5.5. Model Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.608 + ], + "angle": 0, + "content": "We compare the inference time and inference memory of different methods in Table 8. The experiments are conducted on one RTX 3090 with six multi-camera images, whose resolutions are \\(1600 \\times 900\\). We find that our method can achieve both high performance and efficiency. Compared with BEVFormer, our method slightly increases inference time and memory and we think the increased burden is acceptable" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.62, + 0.868, + 0.742 + ], + "angle": 0, + "content": "
MethodLatency (s)Memory (G)
SurroundDepth [58]0.7312.4
NeWCRFs [65]1.0714.5
Adabins [3]0.7515.5
BEVFormer [30]0.314.5
TPVFormer [21]0.325.1
MonoScene [7]0.8720.3
Ours0.345.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.744, + 0.892, + 0.786 + ], + "angle": 0, + "content": "Table 8. The model efficiency of different methods. The experiments are conducted on one RTX 3090 with six multi-camera images, whose resolutions are \\(1600 \\times 900\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.773, + 0.831 + ], + "angle": 0, + "content": "6. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Currently, we only explore single-frame occupancy prediction. However, for the downstream modules, e.g. motion prediction and planning, occupancy flow is more important. As the future work, we will design a framework to build" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21736" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.08, + 0.907, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.684, + 0.893, + 0.713 + ], + "angle": 0, + "content": "Figure 7. Visualizations on nuScenes validation set. Our generated dense occupancy labels are much denser than sparse LiDAR points. Trained with dense groundtruth, the network can predict better and denser occupancy. Better viewed when zoomed in." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.471, + 0.788 + ], + "angle": 0, + "content": "occupancy flow dataset and utilize multi-frame surrounding images as the inputs. Moreover, LiDAR data is not always available. Self-supervised occupancy prediction with only RGB data is a valuable but challenging direction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.196, + 0.816 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose SurroundOcc for multi-camera 3D occupancy prediction. We utilize 2D-3D spatial attention to integrate 2D features to the 3D volume in a multiscale fashion, which is further upsampled and fused by the 3D deconvolution layer. Moreover, we devise a pipeline to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.726, + 0.892, + 0.803 + ], + "angle": 0, + "content": "generate dense occupancy ground truth. We stitch multi-frame LiDAR points of dynamic objects and static scenes separately and utilize Poisson Reconstruction to fill the holes. The comparison on nuScenes and SemanticKITTI datasets demonstrates the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.661, + 0.833 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21737" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. arXiv preprint arXiv:2203.13641, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.47, + 0.257 + ], + "angle": 0, + "content": "[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.47, + 0.312 + ], + "angle": 0, + "content": "[4] Jia-Wang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised Scale-consistent Depth and Ego-motion Learning from Monocular Video. In NeurIPS, pages 35-45, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.47, + 0.368 + ], + "angle": 0, + "content": "[5] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.437 + ], + "angle": 0, + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.479 + ], + "angle": 0, + "content": "[7] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[8] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[9] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[11] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In SIGGRAPH, pages 303-312, 1996." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[13] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, pages 764-773, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[14] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In ICCV, pages 2650-2658, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[15] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In ICCV, pages 2918-2927, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[16] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In ICCV, pages 3828-3838, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[17] Vitor Guizilini, Igor Vasiljevic, Rares Ambrus, Greg Shakhnarovich, and Adrien Gaidon. Full surround monodepth from multiple cameras. IEEE Robotics and Automation Letters, 7(2):5397-5404, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[19] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, pages 15273-15282, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[20] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[21] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[22] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In ICCV, pages 2307-2315, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[23] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[24] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[25] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[26] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[27] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[28] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, et al. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[29] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[30] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chong-hao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[31] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21738" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.175 + ], + "angle": 0, + "content": "[32] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[33] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. Amvnet: Assertion-based multi-view fusion network for lidar semantic segmentation. arXiv preprint arXiv:2012.04934, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[34] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In ECCV, pages 531-548. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[35] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[36] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[37] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[38] Zak Murez, Tarrence van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In ECCV, pages 414-431, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[39] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In ISMAR, pages 127-136, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[40] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[41] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[42] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, pages 523-540. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[43] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[44] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive Collaboration: Joint Unsupervised Learning of Depth, Camera Motion, Optical Flow and Motion Segmentation. In CVPR, pages 12240-12249, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[45] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[46] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[47] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[49] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[50] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In CVPR, pages 15598-15607, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[51] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[52] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In ECCV, pages 685–702, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[53] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In CoRL, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[54] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[55] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[56] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin M. Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In CoRL, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[57] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In ICCV, pages 379–387, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[58] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In CoRL, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[59] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, volume 35, pages 3101-3109, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[60] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, volume 35, pages 3101-3109, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[61] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21739" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.163 + ], + "angle": 0, + "content": "[62] Maosheng Ye, Rui Wan, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Drinet++: Efficient voxel-as-point point cloud segmentation. arXiv preprint arXiv: 2111.08318, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[63] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In CVPR, pages 1983-1992, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[64] Chen Yongjian, Tai Lei, Sun Kai, and Li Mingyang. Monopair: Monocular 3d object detection using pairwise spatial relationships. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[65] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. New crfs: Neural window fully-connected crfs for monocular depth estimation. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[66] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.405 + ], + "angle": 0, + "content": "[67] Zhenyu Zhang, Chunyan Xu, Jian Yang, Junbin Gao, and Zhen Cui. Progressive hard-mining network for monocular depth estimation. TIP, 27(8):3691-3702, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[68] Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. Towards Better Generalization: Joint Depth-Pose Learning without PoseNet. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[69] Hui Zhou, Xinge Zhu, Xiao Song, Yuexin Ma, Zhe Wang, Hongsheng Li, and Dahua Lin. Cylinder3d: An effective 3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[70] Junsheng Zhou, Yuwang Wang, Kaihuai Qin, and Wenjun Zeng. Moving Indoor: Unsupervised Video Depth Learning in Challenging Environments. In ICCV, pages 8618-8627, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.604 + ], + "angle": 0, + "content": "[71] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851-1858, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[72] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[73] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21740" + } + ] +] \ No newline at end of file diff --git a/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/0c733c34-bebb-47b3-8e25-0904ab5a04e1_origin.pdf b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/0c733c34-bebb-47b3-8e25-0904ab5a04e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..accc421941d4168f8e8dc33efb5212af8b117c64 --- /dev/null +++ b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/0c733c34-bebb-47b3-8e25-0904ab5a04e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:934c0ff2139f57dc588dc50c8ea55c84dd8dbc2168c44aca60893e4b29bb7228 +size 8875216 diff --git a/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/full.md b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..60144a137de5d188a0b7612c9efc77d1ad383f61 --- /dev/null +++ b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/full.md @@ -0,0 +1,332 @@ +# SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving + +Yi Wei $^{1,2}$ , Linqing Zhao $^{3}$ , Wenzhao Zheng $^{1,2}$ , Zheng Zhu $^{4}$ , Jie Zhou $^{1,2}$ , Jiwen Lu $^{1,2\dagger}$ + +$^{1}$ Beijing National Research Center for Information Science and Technology, China $^{2}$ Department of Automation, Tsinghua University, China + +$^{3}$ School of Electrical and Information Engineering, Tianjin University, China $^{4}$ PhiGent Robotics + +{y-wei19, zhengwz18}@mails.tsinghua.edu.cn; linqingzhao@tju.edu.cn; zhengzhu@ieee.org; {jzhou, lujiwen}@tsinghua.edu.cn + +# Abstract + +3D scene understanding plays a vital role in vision-based autonomous driving. While most existing methods focus on 3D object detection, they have difficulty describing real-world objects of arbitrary shapes and infinite classes. Towards a more comprehensive perception of a 3D scene, in this paper, we propose a SurroundOcc method to predict the 3D occupancy with multi-camera images. We first extract multi-scale features for each image and adopt spatial 2D-3D attention to lift them to the 3D volume space. Then we apply 3D convolutions to progressively upsample the volume features and impose supervision on multiple levels. To obtain dense occupancy prediction, we design a pipeline to generate dense occupancy ground truth without expansive occupancy annotations. Specifically, we fuse multi-frame LiDAR scans of dynamic objects and static scenes separately. Then we adopt Poisson Reconstruction to fill the holes and voxelize the mesh to get dense occupancy labels. Extensive experiments on nuScenes and SemanticKITTI datasets demonstrate the superiority of our method. Code and dataset are available at https://github.com/weiyithu/SurroundOcc. + +# 1. Introduction + +Understanding the 3D geometry of the surrounding scene serves as the basic procedure in an autonomous driving system. While LiDAR is a direct solution to capture this geometric information, it suffers from high-cost sensors and sparse scanned points, limiting its further application. Recently, vision-centric autonomous driving has attracted extensive attention as a promising direction. Taking multi- + +![](images/e033e84fd6801e416a851495c265db993a2a24650da8180cae346402ffac67be.jpg) +Figure 1. The overview of SurroundOcc. Given multi-camera images, our method can predict volumetric occupancy of surrounding 3D scenes. To train the network, we design a pipeline to generate dense occupancy labels with sparse LiDAR points. Better viewed when zoomed in. + +camera images as inputs, it has demonstrated competitive performance in various 3D perception tasks including depth estimation [17, 58], 3D object detection [30, 27, 36, 34, 20], and semantic map construction [66, 19, 1]. + +Although multi-camera 3D object detection plays an important role [30, 27, 34, 20] in vision-based 3D perception, it is easy to suffer from the long-tail problem and difficult to recognize all classes of objects in the real world. Complementary to 3D detection, reconstructing surrounding 3D scenes can better help the downstream perception tasks. Recent works [17, 58] incorporate information from multiple views and predict surrounding depth maps. However, depth maps only predict the nearest occupied point in each optical ray and are unable to recover the occluded parts of the 3D + +scene. Different from depth-based methods, another trend [7, 21] is to directly predict the 3D occupancy of the scene. It describes a 3D scene by assigning an occupied probability to each voxel in the 3D space. We advocate 3D occupancy to be a good 3D representation for multi-camera scene reconstruction, which naturally guarantees the multicamera geometry consistency and is able to recover occluded parts. Also, it is flexible to extend to other 3D downstream tasks such as 3D semantic segmentation [69, 73, 15]. As one of the pioneering works, MonoScene [7] infers the dense 3D voxelized semantic scene with monocular images. However, simply fusing multi-camera results with cross-camera post-processing will lead to low performance [30]. TPVFormer [21] uses sparse LiDAR points as supervision, which results in sparse occupancy prediction. + +To address this, we propose a SurroundOcc method, which aims to predict dense and accurate 3D occupancy with multi-camera images input. We first use a 2D backbone network to extract multi-scale feature maps from each image. Then we perform 2D-3D spatial attention to lift multi-camera image information to 3D volume features instead of BEV features. A 3D convolution network is then employed to progressively upsample the low-resolution volume features and fuse them with high-resolution ones to obtain fine-grained 3D representations. At each level, we use a decayed weighted loss to supervise the network. To get dense predictions, we need dense occupancy labels. However, the mainstream multi-camera dataset nuScenes [6] only provides sparse LiDAR points. To avoid expensive occupancy annotations, we devise a pipeline to generate dense occupancy ground truth only with the existing 3D detection and 3D semantic segmentation labels. Specifically, we first combine multi-frame points of dynamic objects and static scenes respectively. Then we leverage Poisson Reconstruction [23] algorithm to further fill the holes. Finally, NN and voxelization are used to obtain dense 3D occupancy labels. + +With the dense occupancy ground truth, we train the model and compare it with other state-of-the-art methods on nuScenes [6] dataset. Both the quantitative results and visualizations demonstrate the effectiveness of our method. Moreover, we further conduct experiments on SemanticKITTI dataset [2]. Although our method is not designed for the monocular setting, it achieves state-of-the-art performance on the monocular 3D semantic scene completion benchmark. + +# 2. Related Work + +Voxel-based Scene Representation: How to effectively represent a 3D scene lies at the core of autonomous driving perception. Voxel-based scene representation voxelizes the 3D space into discretized voxels and describes each voxel by a vector feature. It has empowered the success of numerous methods on the lidar segmentation [33, 52, 11, 62, 61] + +and 3D scene completion [7, 46, 9, 25, 59, 29, 28] tasks. For the 3D occupancy prediction task, we also advocate the voxel representation as it is more suitable to model the occupancy field of a 3D scene. MonoScene [7] is the first work to reconstruct outdoor scenes using only RGB inputs. TPVFormer [21] further generalizes it to multi-camera 3D semantic occupancy prediction. However, its lack of dense supervision results in sparse occupancy prediction. Differently, we devise a pipeline to generate dense occupancy ground truth for training and our occupancy prediction is much denser. + +3D Scene Reconstruction: 3D reconstruction [42, 49, 48, 41, 71, 16, 4, 38, 50, 5] is a traditional but important topic in computer vision. One way is through depth estimation which predicts a depth value for each pixel in the image. While early methods require full depth annotations to supervise the depth estimation model [14, 24, 67], later research focuses on self-supervised depth estimation as it does not require intensive human annotations [71, 16, 4, 68, 63, 57, 70, 44, 8]. Recently, SurroundDepth [58] further incorporates the interactions between surrounding views to capture more spatial correlations. Different from depth estimation, 3D scene reconstruction methods [22, 38, 50, 5, 7] directly reconstruct a comprehensive and accurate 3D geometry of a scene. SurfaceNet [22] employs a 3D convolutional network to transform RGB colors to 3D surface occupancy from two images. Atlas [38] further extends it to the multi-view setting and utilized learned features to predict occupancy. NeuralRecon [50] and TransformerFusion [5] fuse the learned image features from different views in an online manner for more accurate 3D reconstructions. However, most of these 3D scene reconstruction methods are designed for indoor scenes, which are different from the multi-camera setting in the outdoor environment. + +Vision-based 3D Perception: The lack of direct geometric input demands vision-based 3D surround perception methods to infer the 3D scene geometry based on semantic cues. Depth-based methods explicitly predict depth maps for image inputs to extract 3D geometric information of the scene before perception [55, 37, 45, 40, 27, 36, 31, 43, 20, 66]. The widely adopted pipeline is to predict categorical depth distributions and leverage them to project semantic features into 3D space [43]. Other methods implicitly learn 3D features without producing explicit depth maps [54, 30, 56, 35, 53, 64]. For example, BEV-Former [30] adopts cross attention to progressively refine BEV grid features from 2D image features. While most existing works employ BEV representation to describe a scene, we propose to reconstruct a 3D scene using volumetric occupancy representation, which provides the more fine-grained and comprehensive modeling of the scene. + +![](images/d38e3f70e6d739f038bff0ca54d376242c9fa9a65063ff3c44c6a2c7560b09f1.jpg) +Figure 2. The pipeline of the proposed method. First, we use a backbone to extract multi-scale features of multi-camera images. Then we adopt 2D-3D spatial attention to fuse multi-camera information and construct 3D volume features in a multi-scale fashion. Finally, the 3D deconvolution layer is used to upsample 3D volumes and occupancy prediction is supervised in each level. + +![](images/29ec71971466d5dc9bb444121d02f1aca7bdeb259bd16d67458e4c8d6471c8db.jpg) +Figure 3. The comparison of 3D-based and BEV-based cross-view attention. The 3D-based attention can better preserve 3D information. For each 3D volume query, we project it to the corresponding 2D views to sample features. + +# 3. Approach + +# 3.1. Problem Formulation + +In this work, we aim to predict the 3D occupancy of surrounding scenes with multi-camera images $I = \{I^1, I^2, \dots, I^N\}$ . Formally, the 3D occupancy prediction is represented as: + +$$ +V = G \left(I ^ {1}, I ^ {2}, \dots I ^ {N}\right) \tag {1} +$$ + +where $G$ is an neural network and $V \in \mathbb{R}^{H \times W \times Z}$ is the 3D occupancy. The value of $V$ is between 0 and 1, representing the occupied probability of the grids. Lifting $V$ to an $(L, H, W, Z)$ tensor, we can obtain the 3D semantic occupancy, where $L$ is the class number and class 0 means non-occupied grids. + +3D occupancy is a good representation for multi-camera 3D scene reconstruction. First, since 3D occupancy is + +predicted in 3D space, it theoretically satisfies the multicamera consistency. Second, it is possible for networks to predict occluded areas according to the surrounding semantic information, which is unavailable in depth estimation. Third, 3D occupancy is easy to extend to other downstream tasks, such as 3D semantic segmentation and scene flow estimation. + +# 3.2. Overview + +Figure 2 shows the pipeline of our method. Given a set of surrounding multi-camera images, we first use a backbone network (e.g. ResNet-101 [18]) to extract $N$ cameras' and $M$ levels' multi-scale features $X = \{\{X_{i}^{j}\}_{i = 1}^{N}\}_{j = 1}^{M}$ . For each level, we use a transformer to fuse multi-camera features with spatial cross attention. The output of the 2D-3D spatial attention layer is a 3D volume feature instead of the BEV feature. Then the 3D convolution network is utilized to upsample and combine multi-scale volume features. The occupancy prediction in each level is supervised by the generated dense occupancy ground truth with a decayed loss weight. + +# 3.3. 2D-3D Spatial Attention + +Many 3D scene reconstruction methods [7, 38] integrate multi-view 2D features into 3D space by reprojecting 2D features back to the 3D volumes with known poses. The grid feature is calculated by simply averaging all 2D features in this grid. However, this kind of method assumes that different views contribute equally to the 3D volumes, + +![](images/2af4c56df14331738a606ad726512032134c97593e1d4b6bfafc82e536b4789e.jpg) +Figure 4. Dense occupancy ground truth generation. We first traverse all frames to stitch the multi-frame LiDAR points of dynamic objects and static scenes separately, and then merge them into a complete scene. Subsequently, we employ Poisson Reconstruction to densify the points and voxelize the resulting mesh to obtain a dense 3D occupancy. Finally, we use the Nearest Neighbor (NN) algorithm to assign semantic labels to dense voxels. + +which does not always hold true, especially when some views are occluded or blurred. + +To tackle the issue, we leverage cross-view attention to fuse multi-camera features. We project 3D reference points to 2D views and use deformable attention [72, 30] to query points and aggregate information. As shown in Figure 3, instead of 2D BEV queries, we build 3D volume queries to further reserve 3D space information. Specifically, 3D volume queries are defined as $Q \in \mathbb{R}^{C \times H \times W \times Z}$ . For each query, we project its corresponding 3D point to 2D views according to the given intrinsic and extrinsic. We only use the views that the 3D reference point hits. Then we sample 2D features around these projected 2D positions. The output $F \in \mathbb{R}^{C \times H \times W \times Z}$ of the cross-view attention layer is a weighted sum of sampled features according to the deformable attention mechanism: + +$$ +\begin{array}{l} \operatorname {D e f o r m A t t n} (q, p, x) = \sum_ {i = 1} ^ {N _ {\text {h e a d}}} \mathcal {W} _ {i} \sum_ {j = 1} ^ {N _ {\text {k e y}}} \mathcal {A} _ {i j} \cdot \mathcal {W} _ {i} ^ {\prime} x (p + \Delta p _ {i j}) \\ F ^ {p} = \frac {1}{\left| \mathcal {V} _ {\mathrm {h i t}} \right|} \sum_ {i \in \mathcal {V} _ {\mathrm {h i t}}} \operatorname {D e f o r m A t t n} \left(Q ^ {p}, \mathcal {P} \left(q ^ {p}, i\right), X _ {i}\right) \tag {2} \\ \end{array} +$$ + +where $F^p$ and $Q^p$ indicate the $p$ th element of the output features and 3D volume queries. $q^p$ is the corresponding 3D positions of queries and $P$ is the 3D to 2D project function. $\mathcal{V}_{\mathrm{hit}}$ represents the hit views of 3D query points. $W_i$ and $W_i'$ are the learnable weights and $A_{ij} \in [0,1]$ is the attention weight calculated by the dot product of query and key. $x(p + \Delta p_{ij})$ is the 2D feature at location $p + \Delta p_{ij}$ . In- + +stead of performing expensive 3D self-attention, we use the 3D convolution to interact features between neighboring 3D voxels. + +# 3.4. Multi-scale Occupancy Prediction + +We further extend the 2D-3D spatial attention to a multiscale fashion. Different from 3D detection task, 3D scene reconstruction needs more low-level features to help the network learn fine-grained details. To tackle the issue, we design a 2D-3D U-Net architecture. Specifically, given multi-scale 2D features $\{\{X_i^j\}_{i=1}^N\}_{j=1}^M$ , we adopt different number of 2D-3D spatial attention layers to extract multiscale 3D volume features $\{F_j \in \mathbb{R}^{C_j \times H_j \times W_j \times Z_j}\}_{j=1}^M$ . Then we upsample $j - 1$ th level 3D volume features $Y_{j - 1}$ with 3D deconvolution layer and fuse it with $F_j$ : + +$$ +Y _ {j} = F _ {j} + \operatorname {D e c o n v} \left(Y _ {j - 1}\right) \tag {3} +$$ + +For each level, the network outputs an occupancy prediction result with different resolution $V_{j} \in \mathbb{R}^{C_{j} \times H_{j} \times W_{j} \times Z_{j}}$ . To get powerful both high-level and low-level 3D features, the network is supervised at each scale. Specifically, we use the cross-entropy loss and scene-class affinity loss introduced in [7] as the supervision signals. For 3D semantic occupancy prediction, we adopt a multi-class cross-entropy loss and for 3D scene reconstruction we change it to a two-class formulation. Since the high-resolution prediction is more important, we use a decayed loss weight $\alpha_{j} = \frac{1}{2^{j}}$ for $j$ th level supervision. + +![](images/befe25c5c2be50f59761225271f81ae62a880c8ff4c3c50f89d2de21a7359483.jpg) +(a) RGB image + +![](images/fe3c2fac10799e7f284f22f272e6938950a55124465693c3d19df49b0b2f7f8d.jpg) +(b) single-frame LiDAR points + +![](images/61a6689ca4744de07150ce9187367390cec506b8640b31025c40642f350bac87.jpg) +(c) Sparse occupancy labels + +![](images/ee806322d9ecd4e4c68cd87a944a03ea148711d27db15d27d89aca2cdb5044ec.jpg) +(d) Dense occupancy labels +Figure 5. Comparison on different occupancy labels. Compared with single-frame LiDAR points and the sparse occupancy converted from multi-frame points, our dense voxels are able to provide more realistic occupancy labels. + +# 4. Dense Occupancy Ground Truth + +In our experiments, we find that the network supervised by sparse LiDAR points is unable to predict dense enough occupancy. Thus, it is necessary to generate dense occupancy labels. However, as mentioned in SemanticKITTI [2], it is complex and needs huge human efforts to annotate the dense occupancy of a 3D scene which has millions of voxels. To this end, we design a pipeline to generate dense occupancy ground truth leveraging existing 3D detection and 3D semantic segmentation labels without extra human annotations, which is shown in Figure 4. An intuitive way is to directly transform a multi-frame LiDAR point sequence into a unified coordinate system, and then voxelize the concatenated dense points into voxel grids. However, such a straightforward solution is only applicable to completely static scenes and ignores moving objects. Moreover, multiframe point clouds are not dense enough and there still exist many holes, which will result in wrong occupancy labels. To address these issues, we propose to stitch multi-frame LiDAR points of dynamic objects and static scenes separately. In addition, we adopt Poisson Reconstruction [23] to fill up the holes and voxelize the obtained mesh to get dense volumetric occupancy. Since the LiDAR scans surface points, our method also generates surface occupancy. + +# 4.1. Multi-frame Point Cloud Stitching + +We propose a two-stream pipeline to stitch static scenes and objects separately and merge them into a complete scene before voxelization. Specifically, for each frame, we first cut out movable objects from the LiDAR points according to 3D bounding box labels, so that we can obtain the 3D points of a static scene and movable objects. After travers- + +ing all frames in the scene, we integrate the collected static scene segments and object segments into a set respectively. To combine the multi-frame segments, we then transform their coordinates into the world coordinate system via the known calibrated matrices and ego-poses. We denote the transformed static scene segments and object segments as $P_{ss} = \{P_{ss}^{1}, P_{ss}^{2}, \dots, P_{ss}^{n}\}$ and $P_{os} = \{P_{os}^{1}, P_{os}^{2}, \dots, P_{os}^{m}\}$ , where $n$ and $m$ are the numbers of frames and objects in the sequence, respectively. Note that the same objects in different frames can be recognized according to the bounding box index. Therefore, we can represent the whole static scene as $P_{s} = [P_{ss}^{1}, P_{ss}^{2}, \dots, P_{ss}^{n}]$ while the objects as $P_{o} = [P_{os}^{1}, P_{os}^{2}, \dots, P_{os}^{m}]$ , respectively, where $[\cdot]$ is the concatenation operator. Finally, according to the objects' locations and ego-pose of the current frame, the 3D points of this frame can be obtained by merging static scene and objects: $P = [T_{s}(Ps), T_{o}(Po)]$ , where $T_{s}$ and $T_{o}$ are the transformations of static scenes and objects from world coordinate system to the current frame coordinate system. In this way, the occupancy labels of the current frame leverage the LiDAR points of all frames in the sequence. + +# 4.2. Densifying with Poisson Reconstruction + +While the density of $P$ is much larger than a single-frame LiDAR, there still exists many interspaces and the points are not evenly distributed, which is caused by the limited LiDAR beams. To address this, we first compute the normal vectors according to the spatial distribution in local neighborhoods. Then we reconstruct $P$ to a triangular mesh $\mathcal{M}$ via Poisson Surface Reconstruction [23], whose input is point cloud with normal vectors, and the output is a triangular mesh (see Figure 4). The obtained mesh $\mathcal{M} = \{\mathcal{V},\mathcal{E}\}$ fills up the holes of point clouds with evenly distributed vertices $\mathcal{V}$ , so that we can further convert the mesh into dense voxel $V_{d}$ . + +# 4.3. Semantic Labeling with NN Algorithm + +Having obtained the occupancy of dense voxels $V_{d}$ , we aim to assign semantic labels to each voxel, since position reconstruction can only be applied to 3D space, not semantic space. To this end, we propose to leverage Nearest Neighbors (NN) algorithm to search the nearest semantic label for each voxel. Specifically, we first voxelize $P$ with semantics into voxels $V_{s}$ , which are sparser than $V_{d}$ due to limited LiDAR beams. Then for each occupied voxel in $V_{d}$ , we use NN to search the nearest voxel in $V_{s}$ and assign the semantic label to it. Note that this process can be accelerated by parallel computing on the GPU. Thus, all occupied voxels in $V_{d}$ can obtain their semantic labels from $V_{s}$ . + +Figure 5 shows a detailed visual comparison between single-frame LiDAR points, sparse occupancy labels and dense occupancy labels. We observe that our dense voxels can provide much more realistic occupancy labels with + +
MethodSC IoUSSC mIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmanmadevegetation
MonoScene [7]23.967.314.030.358.008.042.900.281.160.674.014.3527.725.2015.1311.299.0314.86
Atlas [38]28.6615.0010.645.6819.6624.948.908.846.473.2810.4216.2134.8615.4621.8920.9511.2120.54
BEVFormer [30]30.5016.7514.226.5823.4628.288.6610.776.644.0511.2017.7837.2818.0022.8822.1713.8022.21
TPVFormer [21]11.5111.6616.147.1722.6317.138.8311.3910.468.239.4317.028.0713.6413.8510.344.907.37
TPVFormer*30.8617.1015.965.3123.8627.329.798.747.095.2010.9719.2238.8721.2524.2623.1511.7320.81
SurroundOcc31.4920.3020.5911.6828.0630.8610.7015.1414.0912.0614.3822.2637.2923.7024.4922.7714.8921.86
+ +Table 1. 3D semantic occupancy prediction results on nuScenes validation set. Except TPVFormer [21], all methods are trained with dense occupancy labels. To fairly compare, we further use dense ground truth to train the TPVFormer, which is denoted as TPVFormer*. + +
Accmeanp∈P(minp*∈P* ||p-p*||)
Compmeanp*∈P*(minp∈P ||p-p*)||
Precmeanp∈P(minp*∈P* ||p-p*||<0.5)
Recalmeanp*∈P*(minp∈P ||p-p*||<0.5)
CDAcc + Comp
F-score(2 × Prec × Recal)/(Prec + Recal)
+ +Table 2. Evaluation metrics for 3D scene reconstruction. $p$ and ${p}^{ * }$ are the predicted and ground truth point clouds. + +clear semantic boundaries. + +We think it is not trivial to propagate the sparse semantic label to a dense one since it is a ill-posed problem. The proposed multi-frame point cloud stitching can aggregate multi-frame semantic information and provide dense enough reference points for NN. However, we find that NN is sensitive to the annotation noise in original LiDAR semantic labels, and we will try to solve it as the future work. + +# 5. Experiments + +# 5.1. Experimental Setup + +Dataset: We conduct multi-camera experiments on nuScenes dataset [6], which is a large-scale autonomous driving dataset. Since the 3D semantic and 3D detection labels are unavailable in test set and we cannot generate dense occupancy labels, we use the training set to train the model and validation set for evaluation. The occupancy prediction range is set as $[-50m, 50m]$ for $X, Y$ axis and $[-5m, 3m]$ for $Z$ axis. The final output occupancy has the shape $200x200x16$ with $0.5m$ voxel size. The input image resolution is $1600x900$ . + +To further demonstrate the effectiveness of our method, we conduct monocular semantic scene completion experiment on SemanticKITTI dataset [2]. SemanticKITTI has annotated outdoor LiDAR scans with 21 semantic labels. The ground truth is voxelized as $256 \times 256 \times 32$ grid with $0.2 \mathrm{~m}$ voxel size. We evaluate our model on the test set. + +Implementation Details: The whole network architecture obtains $M = 4$ levels and we do not add the skip connection in level 0. For nuScenes dataset, we adopt ResNet101-DCN + +[18, 13] with the initial weight from FCOS3D [54] as the backbone to extract image features. The features of stage 1,2,3 are fed to FPN [32] and used as multi-scale image features. The number of 2D-3D spatial attention layers are set as 1, 3, 6 for three levels. For SemanticKITTI dataset, following MonoScene [7], we use a pretrained EfficientNetB7 [51] as the backbone to generate multi-scale image features. We also adopt FPN to further fuse the features of different levels. We set the number of 2D-3D spatial attention layers as 1, 3, 8. All experiments are conducted on 8 RTX 3090s. For Possion Reconstruction, we accumulate both key-frame and non key-frame data. In details, we use LiDAR frames at $20\mathrm{Hz}$ , resulting in 400 frames and around 13 million points for a 20 second sequence. + +Evaluation Metrics: For 3D semantic occupancy prediction, we use the intersection over union (IoU) of occupied voxels, ignoring their semantic class as the evaluation metric of the scene completion (SC) task and the mIoU of all semantic classes for the SSC task. + +$$ +\begin{array}{l} \mathrm {I o U} = \frac {T P}{T P + F P + F N} \\ \mathrm {m I o U} = \frac {1}{C} \sum_ {i = 1} ^ {C} \frac {T P _ {i}}{T P _ {i} + F P _ {i} + F N _ {i}} \tag {4} \\ \end{array} +$$ + +where $TP$ , $FP$ , $FN$ indicate the number of true positive, false positive, and false negative predictions. $C$ is the class number. + +For 3D scene reconstruction, we first convert occupancy prediction to point clouds. Following [38, 5], we use 3D metrics for evaluation, which is shown in Table 2. Chamfer distance (CD) and F-score are the main metrics since they consider both precision and recall. Please refer to the supplementary for more evaluation metric details. For all evaluation, we adopt dense occupancy as ground truth since sparse LiDAR points cannot fully evaluate the quality of occupancy reconstruction. + +
MethodSC IoUSSC mIoUroad (15.36%)sidewalk (11.13%)parking (1.12%)other-gmd (0.56%)building (0.41%)car (3.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation (39.5%)trunk (0.51%)terrain (9.17%)person (0.07%)bicyclist (0.07%)motorcyclist. (0.05%)fence (3.90%)pole (0.29%)traf.-sign (0.08%)
LMSCNet [47]31.387.0746.7019.5013.503.1010.3014.300.300.000.0010.800.0010.400.000.000.005.400.000.00
3DSketch [10]26.856.2337.7019.800.000.0012.1017.100.000.000.0012.100.0016.100.000.000.003.400.000.00
AICNet [26]23.937.0939.3018.3019.801.609.6015.300.700.000.009.601.9013.500.000.000.005.000.100.00
JS3C-Net [60]34.008.9747.3021.7019.902.8012.7020.100.800.000.004.1014.203.1012.400.000.208.701.900.30
MonoScene [7]34.1611.0854.7027.1024.805.7014.4018.803.300.500.704.4014.902.4019.501.001.400.4011.103.302.10
TPVFormer [21]34.2511.2655.1027.2027.406.5014.8019.203.701.000.502.3013.902.6020.401.102.400.3011.002.901.50
SurroundOcc34.7211.8656.9028.3030.206.8015.2020.601.401.601.204.4014.903.4019.301.402.000.1011.303.902.40
+ +Figure 6. An example of challenging scenes. Although the quality of RGB images degrades in rainy days and nights, our method can still predict detailed occupancy. Better viewed when zoomed in. +![](images/56757af439a200be096222c7ef88a3bad8357f44f1df00d37ba264a46fcffe98.jpg) +driveable surface +car bus +truck terrain +Vegetation +sidewalk +other flat +pedestrian +□ bicycle +manmade +motorcycle +barrier +construction vehicle +■trailer +traffic cone + +Table 3. Monocular Semantic scene completion results on SemanticKITTI test set. For fair comparison, we use the performances of RGB-inferred versions of the first four methods, which are reported in MonoScene [7]. Although our method is not designed for monocular perception, we still outperform other methods for a large margin. + +
MethodAcc ↓Comp ↓Prec ↑Recall ↑CD ↓F-score ↑
SurroundDepth [58]1.7471.3840.2610.3533.1300.293
AdaBins [3]1.9891.2870.2330.3473.2750.271
NeWCRFs [65]2.1631.2330.2140.3483.3960.257
Atlas [38]0.6791.6850.4070.5462.3650.458
TransformerFusion [5]0.7711.4340.3750.5912.2050.453
SurroundOcc0.7241.2260.4140.6021.9500.483
+ +Table 4. 3D scene reconstruction results on nuScenes validation set. F-score and CD are the main metrics. + +# 5.2. 3D Semantic Occupancy Prediction + +We first conduct multi-camera 3D semantic occupancy prediction on nuScenes [6] dataset and compare with several state-of-the-art methods [7, 30, 21, 38]. For MonoScene [7], we concatenate the multi-camera occupancy predictions and voxelize them with the $0.5\mathrm{m}$ voxel size, which is same as our setting. For BEVFormer [30], we add a 3D segmentation head to predict semantic occupancy. As shown in Table 1, our method achieves state-of-the-art performance. We also show some qualitative results in Figure 6 and Figure 7. See supplementary material for more + +video demos and qualitative comparisons. + +Especially in Figure 6, we show rainy day and night visualization. Although the quality of RGB images degrades in these two challenging scenes, our method can still predict fine details. The LiDAR sensor suffers from rainy days and easily misses points. With multi-frame aggregation and Possion Reconstruction, we dramatically densify the labels and provide strong supervision, which is crucial for the challenging scenarios. Moreover, we conduct color jitter augmentation, which increases the robustness of brightness change. We also note that some parts cannot be observed by LiDAR sensor, such as the back side of the car. Surprisingly, our model can predict the complete shape according to the surrounding information. + +To further demonstrate the superiority of our method, we also conduct monocular 3D semantic scene completion on SemanticKITTI dataset [2]. Table 3 shows the results. Although our method is not designed for monocular perception and cross-view attention will be ineffective for the monocular setting, our method still achieves state-of-the-art performance on this benchmark. + +
MethodSC IoUSSC mIoU
w/o spatial attention29.7817.34
BEV-based attention30.4518.94
Ours31.4920.30
+ +Table 5. The ablation study of 2D-3D spatial attention. "w/o spatial attention" indicates that we average all multi-camera features in a grid. + +
MethodSC IoUSSC mIoU
w/o multi-scale structure30.4118.22
w/o multi-scale supervision31.1619.73
Ours31.4920.30
+ +Table 6. The ablation study of multi-scale occupancy prediction. "w/o multi-scale structure" means that we do not add multi-scale skip connection. + +
SupervisionSC IoUSSC mIoU
sparse LiDAR points11.9612.17
sparse occupancy labels30.5818.83
dense occupancy labels31.4920.30
+ +# 5.3. 3D Scene Reconstruction + +Another important application of 3D occupancy prediction is 3D scene reconstruction. Due to this reason, we further evaluate 3D reconstruction performance without using multi-class semantic labels. We do the comparison with state-of-the-art multi-camera depth estimation methods (SurroundDepth [58]), monocular depth estimation methods (AdaBins [3] and NeWCRFs [65]) and 3D reconstruction method (Atlas [38] and TransformerFusion [5]). For the self-supervised methods SurroundDepth [58], we use depth ground truth to supervise them. To evaluate depth estimation methods in 3D space, following [38, 5], we run TSDF fusion [12, 39] to fuse multi-camera depth as point clouds. Note that we fuse the multi-camera depth maps of the same timestamp but not the multi-frame depths. Thus, there is no need to specially deal with movable objects. As shown in Table 4, SurroundOcc achieves state-of-the-art performance on most metrics, and outperforms other methods by a large margin, which verifies the effectiveness of the proposed method. + +# 5.4. Ablation Study + +2D-3D Spatial Attention: Table 5 shows the ablation results for 2D-3D spatial attention. Without spatial attention, we directly average all multi-camera features in a grid. However, we find this straightforward fusion method performs worse than spatial attention. The potential reason is + +that the contribution of each view is different for a 3D grid. Moreover, the ablation study shows that 3D-based cross-view attention is more effective than BEV-based cross-view attention since it can preserve 3D space information. + +Multi-scale Occupancy Prediction: We conduct an ablation study on multi-scale structure and multi-scale supervision in Table 6. The experimental results show that these two multi-scale designs can boost the performance. The multi-scale skip connection can help the network learn low-level fine-grained features, which is important for detailed high-resolution occupancy prediction. Moreover, the 3D volume features in all levels will be enhanced by multi-scale supervision. + +- Dense Occupancy Supervision: The results in Table 7 demonstrate the importance of using dense occupancy as ground truth. Compared with sparse LiDAR points, the sparse occupancy labels fuse multi-frame points and can provide more powerful supervision. The possion reconstruction and NN algorithm can fill the holes and further densify the occupancy labels, which will boost model's performance. Figure 7 can better illustrate the effectiveness of dense supervision. We can see that the model trained with our occupancy labels can predict much denser occupancy than that trained with LiDAR points. + +# 5.5. Model Efficiency + +We compare the inference time and inference memory of different methods in Table 8. The experiments are conducted on one RTX 3090 with six multi-camera images, whose resolutions are $1600 \times 900$ . We find that our method can achieve both high performance and efficiency. Compared with BEVFormer, our method slightly increases inference time and memory and we think the increased burden is acceptable + +Table 7. The ablation study of dense occupancy supervision. The model trained with our dense occupancy ground truth is much better than that trained with sparse LiDAR points. + +
MethodLatency (s)Memory (G)
SurroundDepth [58]0.7312.4
NeWCRFs [65]1.0714.5
Adabins [3]0.7515.5
BEVFormer [30]0.314.5
TPVFormer [21]0.325.1
MonoScene [7]0.8720.3
Ours0.345.9
+ +Table 8. The model efficiency of different methods. The experiments are conducted on one RTX 3090 with six multi-camera images, whose resolutions are $1600 \times 900$ . + +# 6. Limitations and Future Work + +Currently, we only explore single-frame occupancy prediction. However, for the downstream modules, e.g. motion prediction and planning, occupancy flow is more important. As the future work, we will design a framework to build + +![](images/a1de631bc0bc46376cd8b433450b678318c793058c142c4a6a6d8315a796d275.jpg) +Figure 7. Visualizations on nuScenes validation set. Our generated dense occupancy labels are much denser than sparse LiDAR points. Trained with dense groundtruth, the network can predict better and denser occupancy. Better viewed when zoomed in. + +occupancy flow dataset and utilize multi-frame surrounding images as the inputs. Moreover, LiDAR data is not always available. Self-supervised occupancy prediction with only RGB data is a valuable but challenging direction. + +# 7. Conclusion + +In this paper, we propose SurroundOcc for multi-camera 3D occupancy prediction. We utilize 2D-3D spatial attention to integrate 2D features to the 3D volume in a multiscale fashion, which is further upsampled and fused by the 3D deconvolution layer. Moreover, we devise a pipeline to + +generate dense occupancy ground truth. We stitch multi-frame LiDAR points of dynamic objects and static scenes separately and utilize Poisson Reconstruction to fill the holes. The comparison on nuScenes and SemanticKITTI datasets demonstrates the superiority of our method. + +# Acknowledgement + +This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603. + +# References + +[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. arXiv preprint arXiv:2203.13641, 2022. +[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. +[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021. +[4] Jia-Wang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised Scale-consistent Depth and Ego-motion Learning from Monocular Video. In NeurIPS, pages 35-45, 2019. +[5] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021. +[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. +[7] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022. +[8] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018. +[9] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020. +[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020. +[11] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021. +[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In SIGGRAPH, pages 303-312, 1996. +[13] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, pages 764-773, 2017. +[14] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In ICCV, pages 2650-2658, 2015. +[15] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In ICCV, pages 2918-2927, 2021. +[16] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In ICCV, pages 3828-3838, 2019. + +[17] Vitor Guizilini, Igor Vasiljevic, Rares Ambrus, Greg Shakhnarovich, and Adrien Gaidon. Full surround monodepth from multiple cameras. IEEE Robotics and Automation Letters, 7(2):5397-5404, 2022. +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. +[19] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, pages 15273-15282, 2021. +[20] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. +[21] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817, 2023. +[22] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In ICCV, pages 2307-2315, 2017. +[23] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. +[24] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. +[25] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020. +[26] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020. +[27] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. +[28] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, et al. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. +[29] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. +[30] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chong-hao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022. +[31] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi + +Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022. +[32] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. +[33] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. Amvnet: Assertion-based multi-view fusion network for lidar semantic segmentation. arXiv preprint arXiv:2012.04934, 2020. +[34] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In ECCV, pages 531-548. Springer, 2022. +[35] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022. +[36] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022. +[37] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In ICCV, 2019. +[38] Zak Murez, Tarrence van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In ECCV, pages 414-431, 2020. +[39] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In ISMAR, pages 127-136, 2011. +[40] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In ICCV, 2021. +[41] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019. +[42] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, pages 523-540. Springer, 2020. +[43] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210, 2020. +[44] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive Collaboration: Joint Unsupervised Learning of Depth, Camera Motion, Optical Flow and Motion Segmentation. In CVPR, pages 12240-12249, 2019. +[45] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In CVPR, 2021. +[46] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. + +[47] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119. IEEE, 2020. +[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019. +[49] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. +[50] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In CVPR, pages 15598-15607, 2021. +[51] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114. PMLR, 2019. +[52] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In ECCV, pages 685–702, 2020. +[53] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In CoRL, 2022. +[54] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In ICCV, 2021. +[55] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, 2019. +[56] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin M. Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In CoRL, 2021. +[57] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In ICCV, pages 379–387, 2017. +[58] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In CoRL, 2022. +[59] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, volume 35, pages 3101-3109, 2021. +[60] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, volume 35, pages 3101-3109, 2021. +[61] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet: + +Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022. +[62] Maosheng Ye, Rui Wan, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Drinet++: Efficient voxel-as-point point cloud segmentation. arXiv preprint arXiv: 2111.08318, 2021. +[63] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In CVPR, pages 1983-1992, 2018. +[64] Chen Yongjian, Tai Lei, Sun Kai, and Li Mingyang. Monopair: Monocular 3d object detection using pairwise spatial relationships. In CVPR, 2020. +[65] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. New crfs: Neural window fully-connected crfs for monocular depth estimation. In CVPR, 2022. +[66] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. +[67] Zhenyu Zhang, Chunyan Xu, Jian Yang, Junbin Gao, and Zhen Cui. Progressive hard-mining network for monocular depth estimation. TIP, 27(8):3691-3702, 2018. +[68] Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. Towards Better Generalization: Joint Depth-Pose Learning without PoseNet. In CVPR, 2020. +[69] Hui Zhou, Xinge Zhu, Xiao Song, Yuexin Ma, Zhe Wang, Hongsheng Li, and Dahua Lin. Cylinder3d: An effective 3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550, 2020. +[70] Junsheng Zhou, Yuwang Wang, Kaihuai Qin, and Wenjun Zeng. Moving Indoor: Unsupervised Video Depth Learning in Challenging Environments. In ICCV, pages 8618-8627, 2019. +[71] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851-1858, 2017. +[72] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. +[73] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021. \ No newline at end of file diff --git a/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/images.zip b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c603afa8f5849d7c379f27b7ebec125305d163cb --- /dev/null +++ b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e148cadcad62b237a8a0deb3fed15e508df6c3e85ec18fb33b32843c6450a0a +size 1144092 diff --git a/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/layout.json b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4bf4c40fe925b517a3c91b9d8a3d8f4cc106ed7e --- /dev/null +++ b/2023/SurroundOcc_ Multi-camera 3D Occupancy Prediction for Autonomous Driving/layout.json @@ -0,0 +1,9564 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 47, + 103, + 547, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 547, + 123 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 547, + 123 + ], + "type": "text", + "content": "SurroundOcc: Multi-Camera 3D Occupancy Prediction for Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "spans": [ + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "text", + "content": "Yi Wei" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "text", + "content": ", Linqing Zhao" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "text", + "content": ", Wenzhao Zheng" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "text", + "content": ", Zheng Zhu" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "text", + "content": ", Jie Zhou" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "text", + "content": ", Jiwen Lu" + }, + { + "bbox": [ + 96, + 141, + 501, + 159 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 170, + 496, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 170, + 496, + 198 + ], + "spans": [ + { + "bbox": [ + 97, + 170, + 496, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 170, + 496, + 198 + ], + "type": "text", + "content": "Beijing National Research Center for Information Science and Technology, China " + }, + { + "bbox": [ + 97, + 170, + 496, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 170, + 496, + 198 + ], + "type": "text", + "content": "Department of Automation, Tsinghua University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 199, + 482, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 199, + 482, + 226 + ], + "spans": [ + { + "bbox": [ + 111, + 199, + 482, + 226 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 199, + 482, + 226 + ], + "type": "text", + "content": "School of Electrical and Information Engineering, Tianjin University, China " + }, + { + "bbox": [ + 111, + 199, + 482, + 226 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 111, + 199, + 482, + 226 + ], + "type": "text", + "content": "PhiGent Robotics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 229, + 471, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 229, + 471, + 256 + ], + "spans": [ + { + "bbox": [ + 115, + 229, + 471, + 256 + ], + "type": "text", + "content": "{y-wei19, zhengwz18}@mails.tsinghua.edu.cn; linqingzhao@tju.edu.cn; zhengzhu@ieee.org; {jzhou, lujiwen}@tsinghua.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "spans": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 308, + 290, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 290, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 290, + 559 + ], + "type": "text", + "content": "3D scene understanding plays a vital role in vision-based autonomous driving. While most existing methods focus on 3D object detection, they have difficulty describing real-world objects of arbitrary shapes and infinite classes. Towards a more comprehensive perception of a 3D scene, in this paper, we propose a SurroundOcc method to predict the 3D occupancy with multi-camera images. We first extract multi-scale features for each image and adopt spatial 2D-3D attention to lift them to the 3D volume space. Then we apply 3D convolutions to progressively upsample the volume features and impose supervision on multiple levels. To obtain dense occupancy prediction, we design a pipeline to generate dense occupancy ground truth without expansive occupancy annotations. Specifically, we fuse multi-frame LiDAR scans of dynamic objects and static scenes separately. Then we adopt Poisson Reconstruction to fill the holes and voxelize the mesh to get dense occupancy labels. Extensive experiments on nuScenes and SemanticKITTI datasets demonstrate the superiority of our method. Code and dataset are available at https://github.com/weiyithu/SurroundOcc." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 601, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 687 + ], + "type": "text", + "content": "Understanding the 3D geometry of the surrounding scene serves as the basic procedure in an autonomous driving system. While LiDAR is a direct solution to capture this geometric information, it suffers from high-cost sensors and sparse scanned points, limiting its further application. Recently, vision-centric autonomous driving has attracted extensive attention as a promising direction. Taking multi-" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 282, + 538, + 473 + ], + "blocks": [ + { + "bbox": [ + 312, + 282, + 538, + 473 + ], + "lines": [ + { + "bbox": [ + 312, + 282, + 538, + 473 + ], + "spans": [ + { + "bbox": [ + 312, + 282, + 538, + 473 + ], + "type": "image", + "image_path": "e033e84fd6801e416a851495c265db993a2a24650da8180cae346402ffac67be.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 477, + 547, + 533 + ], + "lines": [ + { + "bbox": [ + 305, + 477, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 547, + 533 + ], + "type": "text", + "content": "Figure 1. The overview of SurroundOcc. Given multi-camera images, our method can predict volumetric occupancy of surrounding 3D scenes. To train the network, we design a pipeline to generate dense occupancy labels with sparse LiDAR points. Better viewed when zoomed in." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 543, + 546, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 546, + 590 + ], + "type": "text", + "content": "camera images as inputs, it has demonstrated competitive performance in various 3D perception tasks including depth estimation [17, 58], 3D object detection [30, 27, 36, 34, 20], and semantic map construction [66, 19, 1]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Although multi-camera 3D object detection plays an important role [30, 27, 34, 20] in vision-based 3D perception, it is easy to suffer from the long-tail problem and difficult to recognize all classes of objects in the real world. Complementary to 3D detection, reconstructing surrounding 3D scenes can better help the downstream perception tasks. Recent works [17, 58] incorporate information from multiple views and predict surrounding depth maps. However, depth maps only predict the nearest occupied point in each optical ray and are unable to recover the occluded parts of the 3D" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21729" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 251 + ], + "type": "text", + "content": "scene. Different from depth-based methods, another trend [7, 21] is to directly predict the 3D occupancy of the scene. It describes a 3D scene by assigning an occupied probability to each voxel in the 3D space. We advocate 3D occupancy to be a good 3D representation for multi-camera scene reconstruction, which naturally guarantees the multicamera geometry consistency and is able to recover occluded parts. Also, it is flexible to extend to other 3D downstream tasks such as 3D semantic segmentation [69, 73, 15]. As one of the pioneering works, MonoScene [7] infers the dense 3D voxelized semantic scene with monocular images. However, simply fusing multi-camera results with cross-camera post-processing will lead to low performance [30]. TPVFormer [21] uses sparse LiDAR points as supervision, which results in sparse occupancy prediction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 252, + 286, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 252, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 49, + 252, + 286, + 502 + ], + "type": "text", + "content": "To address this, we propose a SurroundOcc method, which aims to predict dense and accurate 3D occupancy with multi-camera images input. We first use a 2D backbone network to extract multi-scale feature maps from each image. Then we perform 2D-3D spatial attention to lift multi-camera image information to 3D volume features instead of BEV features. A 3D convolution network is then employed to progressively upsample the low-resolution volume features and fuse them with high-resolution ones to obtain fine-grained 3D representations. At each level, we use a decayed weighted loss to supervise the network. To get dense predictions, we need dense occupancy labels. However, the mainstream multi-camera dataset nuScenes [6] only provides sparse LiDAR points. To avoid expensive occupancy annotations, we devise a pipeline to generate dense occupancy ground truth only with the existing 3D detection and 3D semantic segmentation labels. Specifically, we first combine multi-frame points of dynamic objects and static scenes respectively. Then we leverage Poisson Reconstruction [23] algorithm to further fill the holes. Finally, NN and voxelization are used to obtain dense 3D occupancy labels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 503, + 286, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 503, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 49, + 503, + 286, + 609 + ], + "type": "text", + "content": "With the dense occupancy ground truth, we train the model and compare it with other state-of-the-art methods on nuScenes [6] dataset. Both the quantitative results and visualizations demonstrate the effectiveness of our method. Moreover, we further conduct experiments on SemanticKITTI dataset [2]. Although our method is not designed for the monocular setting, it achieves state-of-the-art performance on the monocular 3D semantic scene completion benchmark." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 621, + 132, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 621, + 132, + 633 + ], + "spans": [ + { + "bbox": [ + 49, + 621, + 132, + 633 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 641, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 641, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 641, + 286, + 712 + ], + "type": "text", + "content": "Voxel-based Scene Representation: How to effectively represent a 3D scene lies at the core of autonomous driving perception. Voxel-based scene representation voxelizes the 3D space into discretized voxels and describes each voxel by a vector feature. It has empowered the success of numerous methods on the lidar segmentation [33, 52, 11, 62, 61]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 72, + 544, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 202 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 202 + ], + "type": "text", + "content": "and 3D scene completion [7, 46, 9, 25, 59, 29, 28] tasks. For the 3D occupancy prediction task, we also advocate the voxel representation as it is more suitable to model the occupancy field of a 3D scene. MonoScene [7] is the first work to reconstruct outdoor scenes using only RGB inputs. TPVFormer [21] further generalizes it to multi-camera 3D semantic occupancy prediction. However, its lack of dense supervision results in sparse occupancy prediction. Differently, we devise a pipeline to generate dense occupancy ground truth for training and our occupancy prediction is much denser." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 220, + 544, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 544, + 493 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 544, + 493 + ], + "type": "text", + "content": "3D Scene Reconstruction: 3D reconstruction [42, 49, 48, 41, 71, 16, 4, 38, 50, 5] is a traditional but important topic in computer vision. One way is through depth estimation which predicts a depth value for each pixel in the image. While early methods require full depth annotations to supervise the depth estimation model [14, 24, 67], later research focuses on self-supervised depth estimation as it does not require intensive human annotations [71, 16, 4, 68, 63, 57, 70, 44, 8]. Recently, SurroundDepth [58] further incorporates the interactions between surrounding views to capture more spatial correlations. Different from depth estimation, 3D scene reconstruction methods [22, 38, 50, 5, 7] directly reconstruct a comprehensive and accurate 3D geometry of a scene. SurfaceNet [22] employs a 3D convolutional network to transform RGB colors to 3D surface occupancy from two images. Atlas [38] further extends it to the multi-view setting and utilized learned features to predict occupancy. NeuralRecon [50] and TransformerFusion [5] fuse the learned image features from different views in an online manner for more accurate 3D reconstructions. However, most of these 3D scene reconstruction methods are designed for indoor scenes, which are different from the multi-camera setting in the outdoor environment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 510, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 544, + 712 + ], + "type": "text", + "content": "Vision-based 3D Perception: The lack of direct geometric input demands vision-based 3D surround perception methods to infer the 3D scene geometry based on semantic cues. Depth-based methods explicitly predict depth maps for image inputs to extract 3D geometric information of the scene before perception [55, 37, 45, 40, 27, 36, 31, 43, 20, 66]. The widely adopted pipeline is to predict categorical depth distributions and leverage them to project semantic features into 3D space [43]. Other methods implicitly learn 3D features without producing explicit depth maps [54, 30, 56, 35, 53, 64]. For example, BEV-Former [30] adopts cross attention to progressively refine BEV grid features from 2D image features. While most existing works employ BEV representation to describe a scene, we propose to reconstruct a 3D scene using volumetric occupancy representation, which provides the more fine-grained and comprehensive modeling of the scene." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "type": "text", + "content": "21730" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 534, + 305 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 534, + 305 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 534, + 305 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 534, + 305 + ], + "type": "image", + "image_path": "d38e3f70e6d739f038bff0ca54d376242c9fa9a65063ff3c44c6a2c7560b09f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 546, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 546, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 546, + 343 + ], + "type": "text", + "content": "Figure 2. The pipeline of the proposed method. First, we use a backbone to extract multi-scale features of multi-camera images. Then we adopt 2D-3D spatial attention to fuse multi-camera information and construct 3D volume features in a multi-scale fashion. Finally, the 3D deconvolution layer is used to upsample 3D volumes and occupancy prediction is supervised in each level." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 356, + 276, + 443 + ], + "blocks": [ + { + "bbox": [ + 59, + 356, + 276, + 443 + ], + "lines": [ + { + "bbox": [ + 59, + 356, + 276, + 443 + ], + "spans": [ + { + "bbox": [ + 59, + 356, + 276, + 443 + ], + "type": "image", + "image_path": "29ec71971466d5dc9bb444121d02f1aca7bdeb259bd16d67458e4c8d6471c8db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 447, + 288, + 492 + ], + "lines": [ + { + "bbox": [ + 46, + 447, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 288, + 492 + ], + "type": "text", + "content": "Figure 3. The comparison of 3D-based and BEV-based cross-view attention. The 3D-based attention can better preserve 3D information. For each 3D volume query, we project it to the corresponding 2D views to sample features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 501, + 115, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 115, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 115, + 514 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 521, + 171, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 171, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 171, + 533 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "type": "text", + "content": "In this work, we aim to predict the 3D occupancy of surrounding scenes with multi-camera images " + }, + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "type": "inline_equation", + "content": "I = \\{I^1, I^2, \\dots, I^N\\}" + }, + { + "bbox": [ + 47, + 539, + 287, + 588 + ], + "type": "text", + "content": ". Formally, the 3D occupancy prediction is represented as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 594, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 594, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 119, + 594, + 287, + 609 + ], + "type": "interline_equation", + "content": "V = G \\left(I ^ {1}, I ^ {2}, \\dots I ^ {N}\\right) \\tag {1}", + "image_path": "57c6dd022205ad3e2e04264e249073974c7719850b3dffd7dca522954ea06432.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " is an neural network and " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{H \\times W \\times Z}" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " is the 3D occupancy. The value of " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " is between 0 and 1, representing the occupied probability of the grids. Lifting " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " to an " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "(L, H, W, Z)" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " tensor, we can obtain the 3D semantic occupancy, where " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " is the class number and class 0 means non-occupied grids." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "3D occupancy is a good representation for multi-camera 3D scene reconstruction. First, since 3D occupancy is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 358, + 546, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 546, + 441 + ], + "type": "text", + "content": "predicted in 3D space, it theoretically satisfies the multicamera consistency. Second, it is possible for networks to predict occluded areas according to the surrounding semantic information, which is unavailable in depth estimation. Third, 3D occupancy is easy to extend to other downstream tasks, such as 3D semantic segmentation and scene flow estimation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 451, + 373, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 451, + 373, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 451, + 373, + 463 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 471, + 546, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 546, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 546, + 614 + ], + "type": "text", + "content": "Figure 2 shows the pipeline of our method. Given a set of surrounding multi-camera images, we first use a backbone network (e.g. ResNet-101 [18]) to extract " + }, + { + "bbox": [ + 304, + 471, + 546, + 614 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 471, + 546, + 614 + ], + "type": "text", + "content": " cameras' and " + }, + { + "bbox": [ + 304, + 471, + 546, + 614 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 471, + 546, + 614 + ], + "type": "text", + "content": " levels' multi-scale features " + }, + { + "bbox": [ + 304, + 471, + 546, + 614 + ], + "type": "inline_equation", + "content": "X = \\{\\{X_{i}^{j}\\}_{i = 1}^{N}\\}_{j = 1}^{M}" + }, + { + "bbox": [ + 304, + 471, + 546, + 614 + ], + "type": "text", + "content": ". For each level, we use a transformer to fuse multi-camera features with spatial cross attention. The output of the 2D-3D spatial attention layer is a 3D volume feature instead of the BEV feature. Then the 3D convolution network is utilized to upsample and combine multi-scale volume features. The occupancy prediction in each level is supervised by the generated dense occupancy ground truth with a decayed loss weight." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 623, + 443, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 443, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 443, + 635 + ], + "type": "text", + "content": "3.3. 2D-3D Spatial Attention" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "content": "Many 3D scene reconstruction methods [7, 38] integrate multi-view 2D features into 3D space by reprojecting 2D features back to the 3D volumes with known poses. The grid feature is calculated by simply averaging all 2D features in this grid. However, this kind of method assumes that different views contribute equally to the 3D volumes," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21731" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 533, + 296 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 533, + 296 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 533, + 296 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 533, + 296 + ], + "type": "image", + "image_path": "2af4c56df14331738a606ad726512032134c97593e1d4b6bfafc82e536b4789e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 298, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 298, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 547, + 344 + ], + "type": "text", + "content": "Figure 4. Dense occupancy ground truth generation. We first traverse all frames to stitch the multi-frame LiDAR points of dynamic objects and static scenes separately, and then merge them into a complete scene. Subsequently, we employ Poisson Reconstruction to densify the points and voxelize the resulting mesh to obtain a dense 3D occupancy. Finally, we use the Nearest Neighbor (NN) algorithm to assign semantic labels to dense voxels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 352, + 288, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 288, + 375 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 288, + 375 + ], + "type": "text", + "content": "which does not always hold true, especially when some views are occluded or blurred." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 544 + ], + "type": "text", + "content": "To tackle the issue, we leverage cross-view attention to fuse multi-camera features. We project 3D reference points to 2D views and use deformable attention [72, 30] to query points and aggregate information. As shown in Figure 3, instead of 2D BEV queries, we build 3D volume queries to further reserve 3D space information. Specifically, 3D volume queries are defined as " + }, + { + "bbox": [ + 46, + 376, + 289, + 544 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{C \\times H \\times W \\times Z}" + }, + { + "bbox": [ + 46, + 376, + 289, + 544 + ], + "type": "text", + "content": ". For each query, we project its corresponding 3D point to 2D views according to the given intrinsic and extrinsic. We only use the views that the 3D reference point hits. Then we sample 2D features around these projected 2D positions. The output " + }, + { + "bbox": [ + 46, + 376, + 289, + 544 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{C \\times H \\times W \\times Z}" + }, + { + "bbox": [ + 46, + 376, + 289, + 544 + ], + "type": "text", + "content": " of the cross-view attention layer is a weighted sum of sampled features according to the deformable attention mechanism:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 50, + 553, + 287, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 553, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 50, + 553, + 287, + 628 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {D e f o r m A t t n} (q, p, x) = \\sum_ {i = 1} ^ {N _ {\\text {h e a d}}} \\mathcal {W} _ {i} \\sum_ {j = 1} ^ {N _ {\\text {k e y}}} \\mathcal {A} _ {i j} \\cdot \\mathcal {W} _ {i} ^ {\\prime} x (p + \\Delta p _ {i j}) \\\\ F ^ {p} = \\frac {1}{\\left| \\mathcal {V} _ {\\mathrm {h i t}} \\right|} \\sum_ {i \\in \\mathcal {V} _ {\\mathrm {h i t}}} \\operatorname {D e f o r m A t t n} \\left(Q ^ {p}, \\mathcal {P} \\left(q ^ {p}, i\\right), X _ {i}\\right) \\tag {2} \\\\ \\end{array}", + "image_path": "94035bfd125a06326c6baeaffe22a40032e9d3a92104e6a1441453799a987d2c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "F^p" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "Q^p" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " indicate the " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "th element of the output features and 3D volume queries. " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "q^p" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " is the corresponding 3D positions of queries and " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " is the 3D to 2D project function. " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{hit}}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " represents the hit views of 3D query points. " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "W_i" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "W_i'" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " are the learnable weights and " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "A_{ij} \\in [0,1]" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " is the attention weight calculated by the dot product of query and key. " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "x(p + \\Delta p_{ij})" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " is the 2D feature at location " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "p + \\Delta p_{ij}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ". In-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 352, + 546, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 352, + 546, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 546, + 388 + ], + "type": "text", + "content": "stead of performing expensive 3D self-attention, we use the 3D convolution to interact features between neighboring 3D voxels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 399, + 488, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 488, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 488, + 414 + ], + "type": "text", + "content": "3.4. Multi-scale Occupancy Prediction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "text", + "content": "We further extend the 2D-3D spatial attention to a multiscale fashion. Different from 3D detection task, 3D scene reconstruction needs more low-level features to help the network learn fine-grained details. To tackle the issue, we design a 2D-3D U-Net architecture. Specifically, given multi-scale 2D features " + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\{\\{X_i^j\\}_{i=1}^N\\}_{j=1}^M" + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "text", + "content": ", we adopt different number of 2D-3D spatial attention layers to extract multiscale 3D volume features " + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\{F_j \\in \\mathbb{R}^{C_j \\times H_j \\times W_j \\times Z_j}\\}_{j=1}^M" + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "text", + "content": ". Then we upsample " + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "inline_equation", + "content": "j - 1" + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "text", + "content": "th level 3D volume features " + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "inline_equation", + "content": "Y_{j - 1}" + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "text", + "content": " with 3D deconvolution layer and fuse it with " + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "inline_equation", + "content": "F_j" + }, + { + "bbox": [ + 304, + 419, + 547, + 540 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 371, + 557, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 557, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 371, + 557, + 545, + 571 + ], + "type": "interline_equation", + "content": "Y _ {j} = F _ {j} + \\operatorname {D e c o n v} \\left(Y _ {j - 1}\\right) \\tag {3}", + "image_path": "710b2441959a843b0afd64fe580ad0523e8de0b00061eaa5dd28b02af87ed681.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "For each level, the network outputs an occupancy prediction result with different resolution " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "V_{j} \\in \\mathbb{R}^{C_{j} \\times H_{j} \\times W_{j} \\times Z_{j}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ". To get powerful both high-level and low-level 3D features, the network is supervised at each scale. Specifically, we use the cross-entropy loss and scene-class affinity loss introduced in [7] as the supervision signals. For 3D semantic occupancy prediction, we adopt a multi-class cross-entropy loss and for 3D scene reconstruction we change it to a two-class formulation. Since the high-resolution prediction is more important, we use a decayed loss weight " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{j} = \\frac{1}{2^{j}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "th level supervision." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21732" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 75, + 165, + 148 + ], + "blocks": [ + { + "bbox": [ + 56, + 75, + 165, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 75, + 165, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 75, + 165, + 148 + ], + "type": "image", + "image_path": "befe25c5c2be50f59761225271f81ae62a880c8ff4c3c50f89d2de21a7359483.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 152, + 137, + 162 + ], + "lines": [ + { + "bbox": [ + 84, + 152, + 137, + 162 + ], + "spans": [ + { + "bbox": [ + 84, + 152, + 137, + 162 + ], + "type": "text", + "content": "(a) RGB image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 75, + 280, + 148 + ], + "blocks": [ + { + "bbox": [ + 168, + 75, + 280, + 148 + ], + "lines": [ + { + "bbox": [ + 168, + 75, + 280, + 148 + ], + "spans": [ + { + "bbox": [ + 168, + 75, + 280, + 148 + ], + "type": "image", + "image_path": "fe3c2fac10799e7f284f22f272e6938950a55124465693c3d19df49b0b2f7f8d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 152, + 276, + 162 + ], + "lines": [ + { + "bbox": [ + 173, + 152, + 276, + 162 + ], + "spans": [ + { + "bbox": [ + 173, + 152, + 276, + 162 + ], + "type": "text", + "content": "(b) single-frame LiDAR points" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 170, + 165, + 243 + ], + "blocks": [ + { + "bbox": [ + 56, + 170, + 165, + 243 + ], + "lines": [ + { + "bbox": [ + 56, + 170, + 165, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 165, + 243 + ], + "type": "image", + "image_path": "61a6689ca4744de07150ce9187367390cec506b8640b31025c40642f350bac87.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 247, + 156, + 257 + ], + "lines": [ + { + "bbox": [ + 64, + 247, + 156, + 257 + ], + "spans": [ + { + "bbox": [ + 64, + 247, + 156, + 257 + ], + "type": "text", + "content": "(c) Sparse occupancy labels" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 170, + 170, + 279, + 243 + ], + "blocks": [ + { + "bbox": [ + 170, + 170, + 279, + 243 + ], + "lines": [ + { + "bbox": [ + 170, + 170, + 279, + 243 + ], + "spans": [ + { + "bbox": [ + 170, + 170, + 279, + 243 + ], + "type": "image", + "image_path": "ee806322d9ecd4e4c68cd87a944a03ea148711d27db15d27d89aca2cdb5044ec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 247, + 270, + 257 + ], + "lines": [ + { + "bbox": [ + 179, + 247, + 270, + 257 + ], + "spans": [ + { + "bbox": [ + 179, + 247, + 270, + 257 + ], + "type": "text", + "content": "(d) Dense occupancy labels" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 259, + 287, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 287, + 304 + ], + "type": "text", + "content": "Figure 5. Comparison on different occupancy labels. Compared with single-frame LiDAR points and the sparse occupancy converted from multi-frame points, our dense voxels are able to provide more realistic occupancy labels." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 316, + 230, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 230, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 230, + 330 + ], + "type": "text", + "content": "4. Dense Occupancy Ground Truth" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 337, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 289, + 613 + ], + "type": "text", + "content": "In our experiments, we find that the network supervised by sparse LiDAR points is unable to predict dense enough occupancy. Thus, it is necessary to generate dense occupancy labels. However, as mentioned in SemanticKITTI [2], it is complex and needs huge human efforts to annotate the dense occupancy of a 3D scene which has millions of voxels. To this end, we design a pipeline to generate dense occupancy ground truth leveraging existing 3D detection and 3D semantic segmentation labels without extra human annotations, which is shown in Figure 4. An intuitive way is to directly transform a multi-frame LiDAR point sequence into a unified coordinate system, and then voxelize the concatenated dense points into voxel grids. However, such a straightforward solution is only applicable to completely static scenes and ignores moving objects. Moreover, multiframe point clouds are not dense enough and there still exist many holes, which will result in wrong occupancy labels. To address these issues, we propose to stitch multi-frame LiDAR points of dynamic objects and static scenes separately. In addition, we adopt Poisson Reconstruction [23] to fill up the holes and voxelize the obtained mesh to get dense volumetric occupancy. Since the LiDAR scans surface points, our method also generates surface occupancy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 622, + 231, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 231, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 231, + 635 + ], + "type": "text", + "content": "4.1. Multi-frame Point Cloud Stitching" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "We propose a two-stream pipeline to stitch static scenes and objects separately and merge them into a complete scene before voxelization. Specifically, for each frame, we first cut out movable objects from the LiDAR points according to 3D bounding box labels, so that we can obtain the 3D points of a static scene and movable objects. After travers-" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": "ing all frames in the scene, we integrate the collected static scene segments and object segments into a set respectively. To combine the multi-frame segments, we then transform their coordinates into the world coordinate system via the known calibrated matrices and ego-poses. We denote the transformed static scene segments and object segments as " + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "inline_equation", + "content": "P_{ss} = \\{P_{ss}^{1}, P_{ss}^{2}, \\dots, P_{ss}^{n}\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "inline_equation", + "content": "P_{os} = \\{P_{os}^{1}, P_{os}^{2}, \\dots, P_{os}^{m}\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": " are the numbers of frames and objects in the sequence, respectively. Note that the same objects in different frames can be recognized according to the bounding box index. Therefore, we can represent the whole static scene as " + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "inline_equation", + "content": "P_{s} = [P_{ss}^{1}, P_{ss}^{2}, \\dots, P_{ss}^{n}]" + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": " while the objects as " + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "inline_equation", + "content": "P_{o} = [P_{os}^{1}, P_{os}^{2}, \\dots, P_{os}^{m}]" + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": " is the concatenation operator. Finally, according to the objects' locations and ego-pose of the current frame, the 3D points of this frame can be obtained by merging static scene and objects: " + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "inline_equation", + "content": "P = [T_{s}(Ps), T_{o}(Po)]" + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "inline_equation", + "content": "T_{s}" + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "inline_equation", + "content": "T_{o}" + }, + { + "bbox": [ + 304, + 72, + 547, + 324 + ], + "type": "text", + "content": " are the transformations of static scenes and objects from world coordinate system to the current frame coordinate system. In this way, the occupancy labels of the current frame leverage the LiDAR points of all frames in the sequence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 331, + 514, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 514, + 344 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 514, + 344 + ], + "type": "text", + "content": "4.2. Densifying with Poisson Reconstruction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "text", + "content": "While the density of " + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "text", + "content": " is much larger than a single-frame LiDAR, there still exists many interspaces and the points are not evenly distributed, which is caused by the limited LiDAR beams. To address this, we first compute the normal vectors according to the spatial distribution in local neighborhoods. Then we reconstruct " + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "text", + "content": " to a triangular mesh " + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "text", + "content": " via Poisson Surface Reconstruction [23], whose input is point cloud with normal vectors, and the output is a triangular mesh (see Figure 4). The obtained mesh " + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mathcal{V},\\mathcal{E}\\}" + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "text", + "content": " fills up the holes of point clouds with evenly distributed vertices " + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "text", + "content": ", so that we can further convert the mesh into dense voxel " + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 304, + 350, + 545, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 502, + 509, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 509, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 509, + 516 + ], + "type": "text", + "content": "4.3. Semantic Labeling with NN Algorithm" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": "Having obtained the occupancy of dense voxels " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": ", we aim to assign semantic labels to each voxel, since position reconstruction can only be applied to 3D space, not semantic space. To this end, we propose to leverage Nearest Neighbors (NN) algorithm to search the nearest semantic label for each voxel. Specifically, we first voxelize " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " with semantics into voxels " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "V_{s}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": ", which are sparser than " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " due to limited LiDAR beams. Then for each occupied voxel in " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": ", we use NN to search the nearest voxel in " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "V_{s}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " and assign the semantic label to it. Note that this process can be accelerated by parallel computing on the GPU. Thus, all occupied voxels in " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "V_{d}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " can obtain their semantic labels from " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "V_{s}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Figure 5 shows a detailed visual comparison between single-frame LiDAR points, sparse occupancy labels and dense occupancy labels. We observe that our dense voxels can provide much more realistic occupancy labels with" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21733" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 547, + 196 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 196 + ], + "type": "table", + "html": "
MethodSC IoUSSC mIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmanmadevegetation
MonoScene [7]23.967.314.030.358.008.042.900.281.160.674.014.3527.725.2015.1311.299.0314.86
Atlas [38]28.6615.0010.645.6819.6624.948.908.846.473.2810.4216.2134.8615.4621.8920.9511.2120.54
BEVFormer [30]30.5016.7514.226.5823.4628.288.6610.776.644.0511.2017.7837.2818.0022.8822.1713.8022.21
TPVFormer [21]11.5111.6616.147.1722.6317.138.8311.3910.468.239.4317.028.0713.6413.8510.344.907.37
TPVFormer*30.8617.1015.965.3123.8627.329.798.747.095.2010.9719.2238.8721.2524.2623.1511.7320.81
SurroundOcc31.4920.3020.5911.6828.0630.8610.7015.1414.0912.0614.3822.2637.2923.7024.4922.7714.8921.86
", + "image_path": "11d8037f397f4bf34fe9b6c5f5135deafc53f715c9b6731ebd7859dcdc4e23f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 234, + 266, + 305 + ], + "blocks": [ + { + "bbox": [ + 46, + 200, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 200, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 547, + 224 + ], + "type": "text", + "content": "Table 1. 3D semantic occupancy prediction results on nuScenes validation set. Except TPVFormer [21], all methods are trained with dense occupancy labels. To fairly compare, we further use dense ground truth to train the TPVFormer, which is denoted as TPVFormer*." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 234, + 266, + 305 + ], + "lines": [ + { + "bbox": [ + 69, + 234, + 266, + 305 + ], + "spans": [ + { + "bbox": [ + 69, + 234, + 266, + 305 + ], + "type": "table", + "html": "
Accmeanp∈P(minp*∈P* ||p-p*||)
Compmeanp*∈P*(minp∈P ||p-p*)||
Precmeanp∈P(minp*∈P* ||p-p*||<0.5)
Recalmeanp*∈P*(minp∈P ||p-p*||<0.5)
CDAcc + Comp
F-score(2 × Prec × Recal)/(Prec + Recal)
", + "image_path": "3f6cfb565dbbcbc1dd140bc2dc300618ac7edf40854253d8f9035dd2a71981fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 307, + 286, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 286, + 329 + ], + "type": "text", + "content": "Table 2. Evaluation metrics for 3D scene reconstruction. " + }, + { + "bbox": [ + 46, + 307, + 286, + 329 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 307, + 286, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 307, + 286, + 329 + ], + "type": "inline_equation", + "content": "{p}^{ * }" + }, + { + "bbox": [ + 46, + 307, + 286, + 329 + ], + "type": "text", + "content": " are the predicted and ground truth point clouds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 341, + 155, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 155, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 155, + 351 + ], + "type": "text", + "content": "clear semantic boundaries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 437 + ], + "type": "text", + "content": "We think it is not trivial to propagate the sparse semantic label to a dense one since it is a ill-posed problem. The proposed multi-frame point cloud stitching can aggregate multi-frame semantic information and provide dense enough reference points for NN. However, we find that NN is sensitive to the annotation noise in original LiDAR semantic labels, and we will try to solve it as the future work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 447, + 128, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 128, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 128, + 460 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 467, + 163, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 163, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 163, + 480 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "text", + "content": "Dataset: We conduct multi-camera experiments on nuScenes dataset [6], which is a large-scale autonomous driving dataset. Since the 3D semantic and 3D detection labels are unavailable in test set and we cannot generate dense occupancy labels, we use the training set to train the model and validation set for evaluation. The occupancy prediction range is set as " + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "inline_equation", + "content": "[-50m, 50m]" + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "inline_equation", + "content": "X, Y" + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "text", + "content": " axis and " + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "inline_equation", + "content": "[-5m, 3m]" + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "text", + "content": " axis. The final output occupancy has the shape " + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "inline_equation", + "content": "200x200x16" + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "inline_equation", + "content": "0.5m" + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "text", + "content": " voxel size. The input image resolution is " + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "inline_equation", + "content": "1600x900" + }, + { + "bbox": [ + 46, + 486, + 287, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "text", + "content": "To further demonstrate the effectiveness of our method, we conduct monocular semantic scene completion experiment on SemanticKITTI dataset [2]. SemanticKITTI has annotated outdoor LiDAR scans with 21 semantic labels. The ground truth is voxelized as " + }, + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 32" + }, + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "text", + "content": " grid with " + }, + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "inline_equation", + "content": "0.2 \\mathrm{~m}" + }, + { + "bbox": [ + 46, + 605, + 286, + 676 + ], + "type": "text", + "content": " voxel size. We evaluate our model on the test set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Implementation Details: The whole network architecture obtains " + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": " levels and we do not add the skip connection in level 0. For nuScenes dataset, we adopt ResNet101-DCN" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 236, + 545, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 404 + ], + "type": "text", + "content": "[18, 13] with the initial weight from FCOS3D [54] as the backbone to extract image features. The features of stage 1,2,3 are fed to FPN [32] and used as multi-scale image features. The number of 2D-3D spatial attention layers are set as 1, 3, 6 for three levels. For SemanticKITTI dataset, following MonoScene [7], we use a pretrained EfficientNetB7 [51] as the backbone to generate multi-scale image features. We also adopt FPN to further fuse the features of different levels. We set the number of 2D-3D spatial attention layers as 1, 3, 8. All experiments are conducted on 8 RTX 3090s. For Possion Reconstruction, we accumulate both key-frame and non key-frame data. In details, we use LiDAR frames at " + }, + { + "bbox": [ + 304, + 236, + 545, + 404 + ], + "type": "inline_equation", + "content": "20\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 236, + 545, + 404 + ], + "type": "text", + "content": ", resulting in 400 frames and around 13 million points for a 20 second sequence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 410, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 468 + ], + "type": "text", + "content": "Evaluation Metrics: For 3D semantic occupancy prediction, we use the intersection over union (IoU) of occupied voxels, ignoring their semantic class as the evaluation metric of the scene completion (SC) task and the mIoU of all semantic classes for the SSC task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 487, + 545, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 487, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 353, + 487, + 545, + 545 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {I o U} = \\frac {T P}{T P + F P + F N} \\\\ \\mathrm {m I o U} = \\frac {1}{C} \\sum_ {i = 1} ^ {C} \\frac {T P _ {i}}{T P _ {i} + F P _ {i} + F N _ {i}} \\tag {4} \\\\ \\end{array}", + "image_path": "a93290dd23013a93bbbc92f373b547d12e8a6361891356f2f76f1dc3ee9c6243.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "type": "inline_equation", + "content": "TP" + }, + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "type": "text", + "content": " indicate the number of true positive, false positive, and false negative predictions. " + }, + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 564, + 545, + 599 + ], + "type": "text", + "content": " is the class number." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "For 3D scene reconstruction, we first convert occupancy prediction to point clouds. Following [38, 5], we use 3D metrics for evaluation, which is shown in Table 2. Chamfer distance (CD) and F-score are the main metrics since they consider both precision and recall. Please refer to the supplementary for more evaluation metric details. For all evaluation, we adopt dense occupancy as ground truth since sparse LiDAR points cannot fully evaluate the quality of occupancy reconstruction." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21734" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 537, + 205 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 537, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 537, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 537, + 205 + ], + "type": "table", + "html": "
MethodSC IoUSSC mIoUroad (15.36%)sidewalk (11.13%)parking (1.12%)other-gmd (0.56%)building (0.41%)car (3.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation (39.5%)trunk (0.51%)terrain (9.17%)person (0.07%)bicyclist (0.07%)motorcyclist. (0.05%)fence (3.90%)pole (0.29%)traf.-sign (0.08%)
LMSCNet [47]31.387.0746.7019.5013.503.1010.3014.300.300.000.0010.800.0010.400.000.000.005.400.000.00
3DSketch [10]26.856.2337.7019.800.000.0012.1017.100.000.000.0012.100.0016.100.000.000.003.400.000.00
AICNet [26]23.937.0939.3018.3019.801.609.6015.300.700.000.009.601.9013.500.000.000.005.000.100.00
JS3C-Net [60]34.008.9747.3021.7019.902.8012.7020.100.800.000.004.1014.203.1012.400.000.208.701.900.30
MonoScene [7]34.1611.0854.7027.1024.805.7014.4018.803.300.500.704.4014.902.4019.501.001.400.4011.103.302.10
TPVFormer [21]34.2511.2655.1027.2027.406.5014.8019.203.701.000.502.3013.902.6020.401.102.400.3011.002.901.50
SurroundOcc34.7211.8656.9028.3030.206.8015.2020.601.401.601.204.4014.903.4019.301.402.000.1011.303.902.40
", + "image_path": "8c19caf5a1437ad4c76d12b820e5cf6cd8428b9c2d6deccea63b3f17977ab220.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 250, + 294, + 385 + ], + "blocks": [ + { + "bbox": [ + 52, + 250, + 294, + 385 + ], + "lines": [ + { + "bbox": [ + 52, + 250, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 52, + 250, + 294, + 385 + ], + "type": "image", + "image_path": "56757af439a200be096222c7ef88a3bad8357f44f1df00d37ba264a46fcffe98.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 388, + 130, + 399 + ], + "lines": [ + { + "bbox": [ + 66, + 388, + 130, + 399 + ], + "spans": [ + { + "bbox": [ + 66, + 388, + 130, + 399 + ], + "type": "text", + "content": "driveable surface" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 141, + 388, + 196, + 399 + ], + "lines": [ + { + "bbox": [ + 141, + 388, + 196, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 196, + 399 + ], + "type": "text", + "content": "car bus" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 205, + 388, + 275, + 399 + ], + "lines": [ + { + "bbox": [ + 205, + 388, + 275, + 399 + ], + "spans": [ + { + "bbox": [ + 205, + 388, + 275, + 399 + ], + "type": "text", + "content": "truck terrain" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 282, + 388, + 326, + 399 + ], + "lines": [ + { + "bbox": [ + 282, + 388, + 326, + 399 + ], + "spans": [ + { + "bbox": [ + 282, + 388, + 326, + 399 + ], + "type": "text", + "content": "Vegetation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 337, + 388, + 375, + 399 + ], + "lines": [ + { + "bbox": [ + 337, + 388, + 375, + 399 + ], + "spans": [ + { + "bbox": [ + 337, + 388, + 375, + 399 + ], + "type": "text", + "content": "sidewalk" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 386, + 388, + 425, + 399 + ], + "lines": [ + { + "bbox": [ + 386, + 388, + 425, + 399 + ], + "spans": [ + { + "bbox": [ + 386, + 388, + 425, + 399 + ], + "type": "text", + "content": "other flat" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 437, + 388, + 479, + 399 + ], + "lines": [ + { + "bbox": [ + 437, + 388, + 479, + 399 + ], + "spans": [ + { + "bbox": [ + 437, + 388, + 479, + 399 + ], + "type": "text", + "content": "pedestrian" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 489, + 388, + 523, + 399 + ], + "lines": [ + { + "bbox": [ + 489, + 388, + 523, + 399 + ], + "spans": [ + { + "bbox": [ + 489, + 388, + 523, + 399 + ], + "type": "text", + "content": "□ bicycle" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 66, + 404, + 135, + 414 + ], + "lines": [ + { + "bbox": [ + 66, + 404, + 135, + 414 + ], + "spans": [ + { + "bbox": [ + 66, + 404, + 135, + 414 + ], + "type": "text", + "content": "manmade" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 153, + 404, + 199, + 414 + ], + "lines": [ + { + "bbox": [ + 153, + 404, + 199, + 414 + ], + "spans": [ + { + "bbox": [ + 153, + 404, + 199, + 414 + ], + "type": "text", + "content": "motorcycle" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 213, + 404, + 246, + 414 + ], + "lines": [ + { + "bbox": [ + 213, + 404, + 246, + 414 + ], + "spans": [ + { + "bbox": [ + 213, + 404, + 246, + 414 + ], + "type": "text", + "content": "barrier" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 260, + 404, + 333, + 414 + ], + "lines": [ + { + "bbox": [ + 260, + 404, + 333, + 414 + ], + "spans": [ + { + "bbox": [ + 260, + 404, + 333, + 414 + ], + "type": "text", + "content": "construction vehicle" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 348, + 404, + 378, + 414 + ], + "lines": [ + { + "bbox": [ + 348, + 404, + 378, + 414 + ], + "spans": [ + { + "bbox": [ + 348, + 404, + 378, + 414 + ], + "type": "text", + "content": "■trailer" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 392, + 404, + 439, + 414 + ], + "lines": [ + { + "bbox": [ + 392, + 404, + 439, + 414 + ], + "spans": [ + { + "bbox": [ + 392, + 404, + 439, + 414 + ], + "type": "text", + "content": "traffic cone" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 420, + 546, + 443 + ], + "lines": [ + { + "bbox": [ + 46, + 420, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 546, + 443 + ], + "type": "text", + "content": "Figure 6. An example of challenging scenes. Although the quality of RGB images degrades in rainy days and nights, our method can still predict detailed occupancy. Better viewed when zoomed in." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 453, + 291, + 536 + ], + "blocks": [ + { + "bbox": [ + 46, + 207, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 547, + 240 + ], + "type": "text", + "content": "Table 3. Monocular Semantic scene completion results on SemanticKITTI test set. For fair comparison, we use the performances of RGB-inferred versions of the first four methods, which are reported in MonoScene [7]. Although our method is not designed for monocular perception, we still outperform other methods for a large margin." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 453, + 291, + 536 + ], + "lines": [ + { + "bbox": [ + 50, + 453, + 291, + 536 + ], + "spans": [ + { + "bbox": [ + 50, + 453, + 291, + 536 + ], + "type": "table", + "html": "
MethodAcc ↓Comp ↓Prec ↑Recall ↑CD ↓F-score ↑
SurroundDepth [58]1.7471.3840.2610.3533.1300.293
AdaBins [3]1.9891.2870.2330.3473.2750.271
NeWCRFs [65]2.1631.2330.2140.3483.3960.257
Atlas [38]0.6791.6850.4070.5462.3650.458
TransformerFusion [5]0.7711.4340.3750.5912.2050.453
SurroundOcc0.7241.2260.4140.6021.9500.483
", + "image_path": "7778f66dab2dfc31cf808705a282c890eadaf684494049addac93a1051620c2e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 541, + 287, + 562 + ], + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 562 + ], + "type": "text", + "content": "Table 4. 3D scene reconstruction results on nuScenes validation set. F-score and CD are the main metrics." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 574, + 235, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 235, + 587 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 235, + 587 + ], + "type": "text", + "content": "5.2. 3D Semantic Occupancy Prediction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "We first conduct multi-camera 3D semantic occupancy prediction on nuScenes [6] dataset and compare with several state-of-the-art methods [7, 30, 21, 38]. For MonoScene [7], we concatenate the multi-camera occupancy predictions and voxelize them with the " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " voxel size, which is same as our setting. For BEVFormer [30], we add a 3D segmentation head to predict semantic occupancy. As shown in Table 1, our method achieves state-of-the-art performance. We also show some qualitative results in Figure 6 and Figure 7. See supplementary material for more" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 454, + 476, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 476, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 476, + 466 + ], + "type": "text", + "content": "video demos and qualitative comparisons." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 470, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 625 + ], + "type": "text", + "content": "Especially in Figure 6, we show rainy day and night visualization. Although the quality of RGB images degrades in these two challenging scenes, our method can still predict fine details. The LiDAR sensor suffers from rainy days and easily misses points. With multi-frame aggregation and Possion Reconstruction, we dramatically densify the labels and provide strong supervision, which is crucial for the challenging scenarios. Moreover, we conduct color jitter augmentation, which increases the robustness of brightness change. We also note that some parts cannot be observed by LiDAR sensor, such as the back side of the car. Surprisingly, our model can predict the complete shape according to the surrounding information." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "To further demonstrate the superiority of our method, we also conduct monocular 3D semantic scene completion on SemanticKITTI dataset [2]. Table 3 shows the results. Although our method is not designed for monocular perception and cross-view attention will be ineffective for the monocular setting, our method still achieves state-of-the-art performance on this benchmark." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21735" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 70, + 257, + 126 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 257, + 126 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 257, + 126 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 257, + 126 + ], + "type": "table", + "html": "
MethodSC IoUSSC mIoU
w/o spatial attention29.7817.34
BEV-based attention30.4518.94
Ours31.4920.30
", + "image_path": "49dbfd2783c00604cd068d40974c91bf58f0640594968081a3bfef3af483f971.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 173, + 272, + 231 + ], + "blocks": [ + { + "bbox": [ + 47, + 130, + 287, + 163 + ], + "lines": [ + { + "bbox": [ + 47, + 130, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 130, + 287, + 163 + ], + "type": "text", + "content": "Table 5. The ablation study of 2D-3D spatial attention. \"w/o spatial attention\" indicates that we average all multi-camera features in a grid." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 173, + 272, + 231 + ], + "lines": [ + { + "bbox": [ + 61, + 173, + 272, + 231 + ], + "spans": [ + { + "bbox": [ + 61, + 173, + 272, + 231 + ], + "type": "table", + "html": "
MethodSC IoUSSC mIoU
w/o multi-scale structure30.4118.22
w/o multi-scale supervision31.1619.73
Ours31.4920.30
", + "image_path": "5a8db4dfd2321d8ea64fea70143884fde2ab374abfebb61041a1137da82ec1bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 277, + 264, + 334 + ], + "blocks": [ + { + "bbox": [ + 47, + 234, + 286, + 268 + ], + "lines": [ + { + "bbox": [ + 47, + 234, + 286, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 286, + 268 + ], + "type": "text", + "content": "Table 6. The ablation study of multi-scale occupancy prediction. \"w/o multi-scale structure\" means that we do not add multi-scale skip connection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 277, + 264, + 334 + ], + "lines": [ + { + "bbox": [ + 69, + 277, + 264, + 334 + ], + "spans": [ + { + "bbox": [ + 69, + 277, + 264, + 334 + ], + "type": "table", + "html": "
SupervisionSC IoUSSC mIoU
sparse LiDAR points11.9612.17
sparse occupancy labels30.5818.83
dense occupancy labels31.4920.30
", + "image_path": "91c1707f7207cd49eb2a50bfe4fd0e543b8a0d34a33ebce95b2000d22dd405c7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 382, + 186, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 186, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 186, + 393 + ], + "type": "text", + "content": "5.3. 3D Scene Reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 400, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 286, + 628 + ], + "type": "text", + "content": "Another important application of 3D occupancy prediction is 3D scene reconstruction. Due to this reason, we further evaluate 3D reconstruction performance without using multi-class semantic labels. We do the comparison with state-of-the-art multi-camera depth estimation methods (SurroundDepth [58]), monocular depth estimation methods (AdaBins [3] and NeWCRFs [65]) and 3D reconstruction method (Atlas [38] and TransformerFusion [5]). For the self-supervised methods SurroundDepth [58], we use depth ground truth to supervise them. To evaluate depth estimation methods in 3D space, following [38, 5], we run TSDF fusion [12, 39] to fuse multi-camera depth as point clouds. Note that we fuse the multi-camera depth maps of the same timestamp but not the multi-frame depths. Thus, there is no need to specially deal with movable objects. As shown in Table 4, SurroundOcc achieves state-of-the-art performance on most metrics, and outperforms other methods by a large margin, which verifies the effectiveness of the proposed method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 141, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 141, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 141, + 647 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 286, + 713 + ], + "type": "text", + "content": "2D-3D Spatial Attention: Table 5 shows the ablation results for 2D-3D spatial attention. Without spatial attention, we directly average all multi-camera features in a grid. However, we find this straightforward fusion method performs worse than spatial attention. The potential reason is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "that the contribution of each view is different for a 3D grid. Moreover, the ablation study shows that 3D-based cross-view attention is more effective than BEV-based cross-view attention since it can preserve 3D space information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 120, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 227 + ], + "type": "text", + "content": "Multi-scale Occupancy Prediction: We conduct an ablation study on multi-scale structure and multi-scale supervision in Table 6. The experimental results show that these two multi-scale designs can boost the performance. The multi-scale skip connection can help the network learn low-level fine-grained features, which is important for detailed high-resolution occupancy prediction. Moreover, the 3D volume features in all levels will be enhanced by multi-scale supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 228, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 360 + ], + "type": "text", + "content": "- Dense Occupancy Supervision: The results in Table 7 demonstrate the importance of using dense occupancy as ground truth. Compared with sparse LiDAR points, the sparse occupancy labels fuse multi-frame points and can provide more powerful supervision. The possion reconstruction and NN algorithm can fill the holes and further densify the occupancy labels, which will boost model's performance. Figure 7 can better illustrate the effectiveness of dense supervision. We can see that the model trained with our occupancy labels can predict much denser occupancy than that trained with LiDAR points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 367, + 407, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 367, + 407, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 367, + 407, + 380 + ], + "type": "text", + "content": "5.5. Model Efficiency" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "type": "text", + "content": "We compare the inference time and inference memory of different methods in Table 8. The experiments are conducted on one RTX 3090 with six multi-camera images, whose resolutions are " + }, + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "type": "inline_equation", + "content": "1600 \\times 900" + }, + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "type": "text", + "content": ". We find that our method can achieve both high performance and efficiency. Compared with BEVFormer, our method slightly increases inference time and memory and we think the increased burden is acceptable" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 319, + 491, + 531, + 587 + ], + "blocks": [ + { + "bbox": [ + 47, + 338, + 286, + 371 + ], + "lines": [ + { + "bbox": [ + 47, + 338, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 286, + 371 + ], + "type": "text", + "content": "Table 7. The ablation study of dense occupancy supervision. The model trained with our dense occupancy ground truth is much better than that trained with sparse LiDAR points." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 491, + 531, + 587 + ], + "lines": [ + { + "bbox": [ + 319, + 491, + 531, + 587 + ], + "spans": [ + { + "bbox": [ + 319, + 491, + 531, + 587 + ], + "type": "table", + "html": "
MethodLatency (s)Memory (G)
SurroundDepth [58]0.7312.4
NeWCRFs [65]1.0714.5
Adabins [3]0.7515.5
BEVFormer [30]0.314.5
TPVFormer [21]0.325.1
MonoScene [7]0.8720.3
Ours0.345.9
", + "image_path": "f1f8c22a4fee7eeb9819d4c9a94d42750155aeb95bda4dd58ad720713edde0e0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 589, + 545, + 622 + ], + "lines": [ + { + "bbox": [ + 306, + 589, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 306, + 589, + 545, + 622 + ], + "type": "text", + "content": "Table 8. The model efficiency of different methods. The experiments are conducted on one RTX 3090 with six multi-camera images, whose resolutions are " + }, + { + "bbox": [ + 306, + 589, + 545, + 622 + ], + "type": "inline_equation", + "content": "1600 \\times 900" + }, + { + "bbox": [ + 306, + 589, + 545, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 645, + 473, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 473, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 473, + 658 + ], + "type": "text", + "content": "6. Limitations and Future Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Currently, we only explore single-frame occupancy prediction. However, for the downstream modules, e.g. motion prediction and planning, occupancy flow is more important. As the future work, we will design a framework to build" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21736" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 63, + 555, + 540 + ], + "blocks": [ + { + "bbox": [ + 48, + 63, + 555, + 540 + ], + "lines": [ + { + "bbox": [ + 48, + 63, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 63, + 555, + 540 + ], + "type": "image", + "image_path": "a1de631bc0bc46376cd8b433450b678318c793058c142c4a6a6d8315a796d275.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 541, + 546, + 564 + ], + "lines": [ + { + "bbox": [ + 46, + 541, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 546, + 564 + ], + "type": "text", + "content": "Figure 7. Visualizations on nuScenes validation set. Our generated dense occupancy labels are much denser than sparse LiDAR points. Trained with dense groundtruth, the network can predict better and denser occupancy. Better viewed when zoomed in." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 574, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 288, + 624 + ], + "type": "text", + "content": "occupancy flow dataset and utilize multi-frame surrounding images as the inputs. Moreover, LiDAR data is not always available. Self-supervised occupancy prediction with only RGB data is a valuable but challenging direction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 633, + 119, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 119, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 119, + 646 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "In this paper, we propose SurroundOcc for multi-camera 3D occupancy prediction. We utilize 2D-3D spatial attention to integrate 2D features to the 3D volume in a multiscale fashion, which is further upsampled and fused by the 3D deconvolution layer. Moreover, we devise a pipeline to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 635 + ], + "type": "text", + "content": "generate dense occupancy ground truth. We stitch multi-frame LiDAR points of dynamic objects and static scenes separately and utilize Poisson Reconstruction to fill the holes. The comparison on nuScenes and SemanticKITTI datasets demonstrates the superiority of our method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 646, + 404, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 404, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 404, + 659 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21737" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. arXiv preprint arXiv:2203.13641, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "text", + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 203 + ], + "type": "text", + "content": "[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 287, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 287, + 247 + ], + "type": "text", + "content": "[4] Jia-Wang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised Scale-consistent Depth and Ego-motion Learning from Monocular Video. In NeurIPS, pages 35-45, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 287, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 291 + ], + "type": "text", + "content": "[5] Aljaz Bozic, Pablo Palafox, Justus Thies, Angela Dai, and Matthias Nießner. Transformerfusion: Monocular rgb scene reconstruction using transformers. Advances in Neural Information Processing Systems, 34:1403-1414, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "type": "text", + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 379 + ], + "type": "text", + "content": "[7] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 402 + ], + "type": "text", + "content": "[8] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In CVPR, pages 5410-5418, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[9] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 447, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 491 + ], + "type": "text", + "content": "[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 535 + ], + "type": "text", + "content": "[11] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, pages 12547-12556, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In SIGGRAPH, pages 303-312, 1996." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "text", + "content": "[13] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, pages 764-773, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[14] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In ICCV, pages 2650-2658, 2015." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[15] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In ICCV, pages 2918-2927, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[16] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In ICCV, pages 3828-3838, 2019." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[17] Vitor Guizilini, Igor Vasiljevic, Rares Ambrus, Greg Shakhnarovich, and Adrien Gaidon. Full surround monodepth from multiple cameras. IEEE Robotics and Automation Letters, 7(2):5397-5404, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[19] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In ICCV, pages 15273-15282, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 207, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 241 + ], + "type": "text", + "content": "[20] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 242, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 285 + ], + "type": "text", + "content": "[21] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 319 + ], + "type": "text", + "content": "[22] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In ICCV, pages 2307-2315, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "text", + "content": "[23] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[24] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 410, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 444 + ], + "type": "text", + "content": "[25] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 445, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 478 + ], + "type": "text", + "content": "[26] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 479, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 522 + ], + "type": "text", + "content": "[27] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "text", + "content": "[28] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, et al. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 578, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 645 + ], + "type": "text", + "content": "[29] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 646, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 690 + ], + "type": "text", + "content": "[30] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chong-hao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[31] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21738" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "Tang. Bevfusion: A simple and robust lidar-camera fusion framework. arXiv preprint arXiv:2205.13790, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 138 + ], + "type": "text", + "content": "[32] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[33] Venice Erin Liong, Thi Ngoc Tho Nguyen, Sergi Widjaja, Dhananjai Sharma, and Zhuang Jie Chong. Amvnet: Assertion-based multi-view fusion network for lidar semantic segmentation. arXiv preprint arXiv:2012.04934, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "text", + "content": "[34] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In ECCV, pages 531-548. Springer, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 250 + ], + "type": "text", + "content": "[35] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv preprint arXiv:2203.05625, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 251, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 293 + ], + "type": "text", + "content": "[36] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela Rus, and Song Han. Bevfusion: Multitask multi-sensor fusion with unified bird's-eye view representation. arXiv preprint arXiv:2205.13542, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "type": "text", + "content": "[37] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving. In ICCV, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "text", + "content": "[38] Zak Murez, Tarrence van As, James Bartolozzi, Ayan Sinha, Vijay Badrinarayanan, and Andrew Rabinovich. Atlas: End-to-end 3d scene reconstruction from posed images. In ECCV, pages 414-431, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "type": "text", + "content": "[39] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In ISMAR, pages 127-136, 2011." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "text", + "content": "[40] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In ICCV, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[41] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, pages 165-174, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 515, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 548 + ], + "type": "text", + "content": "[42] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, pages 523-540. Springer, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 581 + ], + "type": "text", + "content": "[43] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 635 + ], + "type": "text", + "content": "[44] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive Collaboration: Joint Unsupervised Learning of Depth, Camera Motion, Optical Flow and Motion Segmentation. In CVPR, pages 12240-12249, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "text", + "content": "[45] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In CVPR, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[46] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[47] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119. IEEE, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "text", + "content": "[49] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[50] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In CVPR, pages 15598-15607, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[51] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114. PMLR, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[52] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution. In ECCV, pages 685–702, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "type": "text", + "content": "[53] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In CoRL, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 355, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 387 + ], + "type": "text", + "content": "[54] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In ICCV, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 432 + ], + "type": "text", + "content": "[55] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In CVPR, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "type": "text", + "content": "[56] Yue Wang, Vitor Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin M. Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In CoRL, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 533 + ], + "type": "text", + "content": "[57] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In ICCV, pages 379–387, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 578 + ], + "type": "text", + "content": "[58] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Yongming Rao, Guan Huang, Jiwen Lu, and Jie Zhou. Surrounddepth: Entangling surrounding views for self-supervised multi-camera depth estimation. In CoRL, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "text", + "content": "[59] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, volume 35, pages 3101-3109, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[60] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, volume 35, pages 3101-3109, 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 714 + ], + "type": "text", + "content": "[61] Dongqiangzi Ye, Zixiang Zhou, Weijia Chen, Yufei Xie, Yu Wang, Panqu Wang, and Hassan Foroosh. Lidarmultinet:" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21739" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 556 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "Towards a unified multi-task network for lidar perception. arXiv preprint arXiv:2209.09385, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[62] Maosheng Ye, Rui Wan, Shuangjie Xu, Tongyi Cao, and Qifeng Chen. Drinet++: Efficient voxel-as-point point cloud segmentation. arXiv preprint arXiv: 2111.08318, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "text", + "content": "[63] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In CVPR, pages 1983-1992, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "text", + "content": "[64] Chen Yongjian, Tai Lei, Sun Kai, and Li Mingyang. Monopair: Monocular 3d object detection using pairwise spatial relationships. In CVPR, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "text", + "content": "[65] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. New crfs: Neural window fully-connected crfs for monocular depth estimation. In CVPR, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "type": "text", + "content": "[66] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 320 + ], + "type": "text", + "content": "[67] Zhenyu Zhang, Chunyan Xu, Jian Yang, Junbin Gao, and Zhen Cui. Progressive hard-mining network for monocular depth estimation. TIP, 27(8):3691-3702, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 322, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 287, + 354 + ], + "type": "text", + "content": "[68] Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. Towards Better Generalization: Joint Depth-Pose Learning without PoseNet. In CVPR, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "text", + "content": "[69] Hui Zhou, Xinge Zhu, Xiao Song, Yuexin Ma, Zhe Wang, Hongsheng Li, and Dahua Lin. Cylinder3d: An effective 3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 400, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 443 + ], + "type": "text", + "content": "[70] Junsheng Zhou, Yuwang Wang, Kaihuai Qin, and Wenjun Zeng. Moving Indoor: Unsupervised Video Depth Learning in Challenging Environments. In ICCV, pages 8618-8627, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 287, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 478 + ], + "type": "text", + "content": "[71] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851-1858, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 511 + ], + "type": "text", + "content": "[72] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 513, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 556 + ], + "type": "text", + "content": "[73] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21740" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/3fb2491f-940c-4884-86b5-1f7be4430698_content_list.json b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/3fb2491f-940c-4884-86b5-1f7be4430698_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0203b8e0dd825582b45e5fa60ebdb40db1b7374d --- /dev/null +++ b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/3fb2491f-940c-4884-86b5-1f7be4430698_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "text", + "text": "SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications", + "text_level": 1, + "bbox": [ + 153, + 128, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abdelrahman Shaker $^{1*}$ Muhammad Maaz $^{1}$ Hanoona Rasheed $^{1}$ Salman Khan $^{1}$ Ming-Hsuan Yang $^{2,3,4}$ Fahad Shahbaz Khan $^{1,5}$ $^{1}$ Mohamed bin Zayed University of AI $^{2}$ University of California, Merced \n $^{3}$ Yonsei University $^{4}$ Google Research $^{5}$ Linköping University", + "bbox": [ + 148, + 202, + 820, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-attention has become a defacto choice for capturing global context in various vision applications. However, its quadratic computational complexity with respect to image resolution limits its use in real-time applications, especially for deployment on resource-constrained mobile devices. Although hybrid approaches have been proposed to combine the advantages of convolutions and self-attention for a better speed-accuracy trade-off, the expensive matrix multiplication operations in self-attention remain a bottleneck. In this work, we introduce a novel efficient additive attention mechanism that effectively replaces the quadratic matrix multiplication operations with linear element-wise multiplications. Our design shows that the key-value interaction can be replaced with a linear layer without sacrificing any accuracy. Unlike previous state-of-the-art methods, our efficient formulation of self-attention enables its usage at all stages of the network. Using our proposed efficient additive attention, we build a series of models called \"Swift-Former\" which achieves state-of-the-art performance in terms of both accuracy and mobile inference speed. Our small variant achieves $78.5\\%$ top-1 ImageNet-1K accuracy with only 0.8 ms latency on iPhone 14, which is more accurate and $2\\times$ faster compared to MobileViT-v2. Our code and models: https://tinyurl.com/5ft8v46w", + "bbox": [ + 75, + 340, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 731, + 209, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, transformer models have shown remarkable success in various vision applications such as classification [8, 44, 24, 23, 9], detection [2, 59, 33, 56, 28], and segmentation [4, 40]. However, deploying these models on resource-constrained mobile devices for real-time applications remains challenging due to their inherently complex nature [20, 29]. Specifically, vision transformers (ViTs) rely on global self-attention, which has a quadratic com", + "bbox": [ + 75, + 756, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e465b2c4ede2ca863f0d496b5efb8565cf4f6b70e777f2e89c77cd8852cf50ca.jpg", + "image_caption": [ + "Figure 1: Latency vs Accuracy Comparison. Compared to the recent EfficientFormer-L1 [20], our SwiftFormer-L1 achieves an absolute gain of $1.7\\%$ in terms of top-1 accuracy with the same latency and without requiring any neural architecture search." + ], + "image_footnote": [], + "bbox": [ + 500, + 306, + 893, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plexity with respect to the input image resolution, making it impractical for deployment on low-powered mobile devices [31]. As a result, convolutional neural networks (CNNs) are still the preferred choice for real-time deployment on mobile devices, primarily because the convolution operation is computationally efficient [39, 15]. However, a major limitation of CNNs is their reliance on local connections and stationary weights, which can limit their ability to adapt to variable input resolutions and capture long-range dependencies in the data. Therefore, developing more efficient and flexible models that combine the strengths of both CNNs and transformers is critical, particularly for mobile devices with limited computational resources.", + "bbox": [ + 496, + 684, + 893, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve this goal, several hybrid approaches have", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author: abdelrahman.youssief@mbzai.ac.ae", + "bbox": [ + 99, + 886, + 426, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17425", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "been proposed that use lightweight CNN modules in the high-resolution early stages and self-attention in the low-resolution later stages [55, 29, 20]. This approach effectively increases the receptive field of the network and strives to achieve a trade-off between speed and accuracy. Furthermore, different efficient variants of computing self-attention have been proposed to reduce the model complexity. These include computing attention across feature dimensions to implicitly model the global context [29], computing attention within local windows [24], pooling spatial features before applying self-attention [9], and sparsely attending to a fixed number of tokens [34], to name a few.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these approaches effectively reduce network complexity, they still involve inefficient matrix multiplication operations that significantly impact latency on mobile devices. To address this issue, Mehta et al. [31] propose a separable self-attention mechanism that replaces matrix multiplication operations to element-wise multiplications. This is achieved by projecting queries to context scores, followed by element-wise multiplication with keys to calculate context vectors for encoding global context.", + "bbox": [ + 75, + 277, + 472, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose efficient additive attention, which eliminates the need for expensive matrix multiplication operations in computing self-attention. Additionally, we propose to compute the global context using only the query-key interactions followed by a linear transformation, without requiring explicit key-value interactions. This significantly reduces the computational complexity and enables us to use the proposed attention block in all stages of the network. Our contributions are as follows:", + "bbox": [ + 75, + 419, + 472, + 555 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce efficient additive attention, a new approach for computing self-attention in vision backbones that eliminates the need for expensive matrix multiplication operations, significantly reducing the computational complexity of the model.", + "- Unlike previous methods, our proposed efficient attention design can be used at all stages of the network, enabling more effective contextual information capture and achieving superior speed-accuracy trade-off.", + "- We build a series of efficient generic classification models called \"SwiftFormer\", which utilize our proposed efficient additive attention. Our small model achieves $78.5\\%$ top-1 ImageNet-1K [7] accuracy while running at only $0.8~\\mathrm{ms}$ latency on iPhone 14. Moreover, our large model achieves $83.0\\%$ accuracy with a latency of only $1.9~\\mathrm{ms}$ . Our model achieves state-of-the-art performance, outperforming recent MobileViTv2 [31] and EfficientFormer [20] by obtaining a better trade-off between accuracy and latency (see Fig. 1)." + ], + "bbox": [ + 96, + 580, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient CNNs: Designing efficient CNNs for mobile vision applications has received much attention in recent years. MobileNet architectures [16, 39, 15] propose depthwise separable convolutions as well as efficient inverted residual blocks for improved performance on various vision tasks. Other methods aim to improve the efficiency by leveraging depth-wise dilated convolutions [32], channel shuffling and pointwise group convolutions [57, 27], network pruning [11, 50], low bit-width [1, 17], and neural architecture search [41, 15]. CNN-based methods are well-performing, efficient, and fast to train and run on edge devices, resulting in widespread usage in the industry. However, they are spatially local and lack global interaction between the features, which deeply affects their performance. Efficient Transformers: ViTs [8] have been widely used in numerous vision tasks, and significant advances have been made in terms of data efficiency [22, 44], transformer architecture [30, 20, 3, 54], and token mechanisms [43, 52]. Reducing the number of visual tokens is a major modification in the transformer architecture for efficient deployment. Instead of using a fixed feature representation through the whole architecture, some methods employ a hierarchical design where the resolution is gradually decreased through the stages, including down-sampling techniques [9, 40, 38, 14] and pyramidal structures [47, 49]. Recently, a few methods [37, 10, 34] propose token sparsification techniques to encode only a subset of the most informative tokens.", + "bbox": [ + 496, + 114, + 890, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Numerous approaches have recently been proposed to reduce the quadratic complexity of self-attention, the computational bottleneck in transformer-based architectures, by computing its approximated variants [34, 29, 31, 46, 19, 5, 45]. EdgeViT [34] uses a global sparse attention module attending only to a few tokens to improve the efficiency, while [48] down-samples the key and value vectors that lead to a better efficiency-accuracy trade-off. GCViT [12] uses a local-global attention mechanism by employing the interactions between a global query and local key and value to capture the contextual information. This approach enables the model to encode both local and global representations efficiently. EdgeNeXt [29] adopts transposed self-attention operation to compute the attention maps across the channel dimension instead of the spatial dimension, followed by token mixing, to have a linear complexity with respect to the number of tokens. Reformer [19] replaces the dot-product attention with a locality-sensitive hashing to group the tokens and reduced the complexity from $O(n^{2})$ to $O(n\\log n)$ . However, this design is only efficient on longer sequences, which is typically not the case for ViTs. LinFormer [46] is a low-rank matrix factorization method that approximates the self-attention matrix with a low-rank matrix, reducing the complexity from $O(n^{2})$ to $O(n)$ . Although matrix factorization methods theoretically reduce the complexity of", + "bbox": [ + 496, + 523, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17426", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "self-attention, they use expensive projections for computing attention, which may not reflect the reduction in FLOPs and parameters into actual speed on mobile platforms.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although these methods show promise and have reduced the complexity of self-attention theoretically, they are inadequate for reducing the inference speed for mobile deployment. Since the complexity of the multi-headed self-attention (MHSA) is higher in the earlier stage compared to the last stages, EfficientFormer [20] incorporates MHSA in the last stage only to learn contextual information from the high-level features without increasing the inference speed significantly. Recently, MobileViT-v2 [31] proposes separable self-attention that uses element-wise operations instead of the dot-product to compute the attention maps with linear complexity. Different from the existing approaches, we propose a consistent hybrid design with an efficient additive attention mechanism to model the contextual information with linear complexity. Instead of capturing the pairwise interactions between keys, queries, and values using the dot-product, we use element-wise operations with learnable attention weights to model the interactions between query and keys only, leading to better inference speed.", + "bbox": [ + 75, + 136, + 470, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 436, + 168, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation: To motivate our method, we first distinguish three desirable characteristics to be considered when designing an efficient yet accurate approach for resource constraint mobile devices.", + "bbox": [ + 75, + 462, + 468, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient Global Context Modeling: As discussed earlier, most existing approaches either employ the standard MHSA or an approximated variant to learn the global context. However, they struggle to operate as fast as MobileNets on resource-constrained devices. This is likely due to the computation-intensive multiplicative operations during attention computation or reliance on advanced reshaping and indexing operations in these approaches. For instance, the recent MobileViT-v2 [31] is $2 \\times$ slower than MobileNet-v2 [39]. Instead of using matrix multiplications, we argue that encoding the global context using an efficient additive attention design can reduce the operations with respect to the number of tokens. This is expected to help operate at comparable speed and model size, while achieving superior accuracy compared to MobileNets.", + "bbox": [ + 75, + 523, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rethinking key-value interactions: Other than multiplicative operations during attention computation, additive attention has been recently explored in the NLP domain [51]. However, in the standard form, it performs three-step processing to model query, key, and value interactions. Each step feeds into the subsequent one, thereby requiring sequential processing. Here, we rethink the additive attention for vision tasks by alleviating the need to compute explicit interactions between key-value. We argue that eliminating key-value interactions and replacing them with a sim", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ple linear transformation empirically encodes better contextual representation. Our design encodes only global queries and key interactions to learn the global contextual information, followed by a linear transformation to compute global context-aware attention weights.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consistent Hybrid Design: Most existing works employ MHSA or the approximated variant in the last stages, while avoiding its usage in the earlier stages. This is because the computational complexity of MHSA grows quadratically with the length of the tokens, making it impractical to incorporate during the initial stages. This constraint adds to the design complexity and requires a careful selection of stages, where MHSA can be applied. In contrast, our proposed SwiftFormer module has linear complexity with respect to the token length and can be incorporated in all stages to learn consistent global context at each scale. This consistency improves the model performance and makes it more generalizable and scalable for high-resolution images.", + "bbox": [ + 496, + 167, + 890, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of Attention Modules", + "text_level": 1, + "bbox": [ + 498, + 369, + 777, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision transformer models are built upon the self-attention (see Fig. 2 (a)), which can effectively model the interactions between the input tokens. Specifically, the self-attention has $\\mathbf{x}$ as an input, where $\\mathbf{x} \\in \\mathbb{R}^{n \\times d}$ , comprising $n$ tokens with $d$ -dimensional embedding vector. The input $\\mathbf{x}$ is projected to query $(\\mathbf{Q})$ , key $(\\mathbf{K})$ , and value $(\\mathbf{V})$ using three matrices, $\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}$ , and $\\mathbf{W}_{\\mathbf{V}}$ . Each self-attention layer comprises $h$ heads, which allows the model to attend to different views of the input. The self-attention can be described as:", + "bbox": [ + 496, + 393, + 890, + 529 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\cdot \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right) \\cdot \\mathbf {V}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 536, + 890, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The attention scores between each pair of tokens in $\\mathbf{Q}$ and $\\mathbf{K}$ are computed using the dot-product operation. Next, these scores are normalized followed by Softmax to weigh the interactions between the tokens. Finally, the weighted interactions are multiplied by $\\mathbf{V}$ using the dot-product operation to produce the final weighted output. Overall, the complexity of the self-attention is $O(n^{2} \\cdot d)$ , where $n$ is the number of tokens and $d$ is the hidden dimension. The computational and memory demands of $\\mathbf{Q} \\cdot \\mathbf{K}^{\\top}$ increase quadratically as the number of tokens grows, leading to slow inference speed and high memory usage, making it impractical to run in real-time for long sequences.", + "bbox": [ + 496, + 575, + 890, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate this issue, [29] proposes the transpose self-attention (see Fig. 2 (b)) to reduce the complexity from quadratic to linear with respect to the number of tokens. Here, the dot-product operation is applied across the channel dimension instead of the spatial dimension. This allows the model to learn feature maps with implicit contextual representation. The attention can be described as:", + "bbox": [ + 496, + 757, + 890, + 863 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = \\mathbf {V} \\cdot \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} ^ {\\top} \\cdot \\mathbf {K}}{\\sqrt {d}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 869, + 890, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17427", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f4dadbfd779cf46e7485bcf975029c11607f9bcbb982d8d877b7d4929d7444b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 89, + 251, + 277 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ce019951c2435c475affca2889a026a83ec51408895987e0368a1cf3bf5aaaf0.jpg", + "image_caption": [ + "Figure 2: Comparison with different self-attention modules. (a) is a typical self-attention used in ViTs [8]. (b) is the transpose self-attention used in EdgeNeXt [29], where the self-attention operation is applied across channel feature dimensions $(\\mathrm{d} \\times \\mathrm{d})$ instead of the spatial dimension $(\\mathrm{n} \\times \\mathrm{n})$ . (c) is the separable self-attention of MobileViT-v2 [31], it uses element-wise operations to compute the context vector from the interactions of $\\mathbf{Q}$ and $\\mathbf{K}$ matrices. Then, the context vector is multiplied by $\\mathbf{V}$ matrix to produce the final output. (d) Our proposed efficient additive self-attention. Here, the query matrix is multiplied by learnable weights and pooled to produce global queries. Then, the matrix $\\mathbf{K}$ is element-wise multiplied by the broadcasted global queries, resulting the global context representation." + ], + "image_footnote": [], + "bbox": [ + 135, + 290, + 202, + 309 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f73e2b92d00512d0387eee4f53d0739fae7e8873798e735034a3b05707448349.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 92, + 442, + 270 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cbe2ad1a39b283ffae4a407beae8ce0dd6c57e10a8120f259dfdb3aa10d0adfe.jpg", + "image_caption": [ + "Dot product" + ], + "image_footnote": [], + "bbox": [ + 359, + 290, + 380, + 308 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9f7fad8cfecfe1437fd5f3906b308ad92ee323de6f1981a7292d4e13a27fa0d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 92, + 647, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/64098f4a824891d3903e14023254a3e540ca1176cb161bb76e4af57ab5eb9d3e.jpg", + "image_caption": [ + "Broadcasted element-wise multiplication" + ], + "image_footnote": [], + "bbox": [ + 446, + 292, + 465, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1f2df7d2f6545780db8f993cde62f0d32290584c1462b44568b79e49b0f5508b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 92, + 867, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The transpose self-attention has a computational complexity of $O(n \\cdot d^2)$ . While this complexity scales linearly with the number of tokens $n$ , it remains quadratic with respect to the feature dimension $d$ . Further, the dot-product operation is still utilized between the query and key matrices.", + "bbox": [ + 75, + 467, + 468, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The separable self-attention mechanism (see Fig. 2 (c)) aims to address the bottleneck of the standard self-attention. Here, the interactions between the queries $(\\mathbf{Q})$ , keys $(\\mathbf{K})$ , and values $(\\mathbf{V})$ are encoded using element-wise operations. First, the query matrix $\\mathbf{Q}$ is projected to produce a vector $\\mathbf{q}$ of dimensions $n \\times 1$ , and then fed into Softmax to generate the context scores, which captures the importance of each query element. Then, the context scores are multiplied by the key matrix $\\mathbf{K}$ and pooled to compute a context vector, which encodes the contextual information. Finally, the context vector is multiplied element-wise with the value matrix $\\mathbf{V}$ to propagate the contextual information and produce the final output $\\hat{\\mathbf{x}}$ . It can be summarized as:", + "bbox": [ + 75, + 542, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = \\mathbf {V} * \\sum \\mathbf {K} * \\operatorname {S o f t m a x} (\\mathbf {q}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 748, + 468, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $*$ denotes the element-wise multiplication operation.", + "bbox": [ + 76, + 777, + 464, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Efficient Additive Attention", + "text_level": 1, + "bbox": [ + 76, + 801, + 326, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The typical additive attention mechanism in NLP captures the global context by utilizing pairwise interactions between the tokens via element-wise multiplications instead of using dot-product operation. It encodes the relevance scores for the contextual information of the input sequence based on", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the interactions of the three attention components (Q, K, V). In contrast, we show that key-value interactions can be removed without sacrificing the performance and only focusing on effectively encoding query-key interactions by incorporating a linear projection layer is sufficient to learn the relationship between the tokens (see Fig. 2 (d)). This approach, named efficient additive attention, has a faster inference speed and produces more robust contextual representations as demonstrated by our performance on image classification, object detection, and segmentation tasks (Sec. 4). Specifically, the input embedding matrix $\\mathbf{x}$ is transformed into query (Q) and key (K) using two matrices $\\mathbf{W_q}$ , $\\mathbf{W_k}$ , where Q, $\\mathbf{K} \\in \\mathbb{R}^{n \\times d}$ , $\\mathbf{W_q}$ , $\\mathbf{W_k} \\in \\mathbb{R}^{d \\times d}$ , $n$ is the token length and $d$ is the dimensions of the embedding vector. Next, the query matrix Q is multiplied by learnable parameter vector $\\mathbf{w}_a \\in \\mathbb{R}^d$ to learn the attention weights of the query, producing global attention query vector $\\alpha \\in \\mathbb{R}^n$ as:", + "bbox": [ + 496, + 467, + 890, + 726 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = \\mathbf {Q} \\cdot \\mathbf {w} _ {a} / \\sqrt {d} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 738, + 890, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the query matrix is pooled based on the learned attention weights, resulting in a single global query vector $\\mathbf{q} \\in \\mathbb{R}^d$ as follows:", + "bbox": [ + 496, + 763, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} = \\sum_ {i = 1} ^ {n} \\alpha_ {i} * \\mathbf {Q} _ {\\mathbf {i}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 820, + 890, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, the interactions between the global query vector $\\mathbf{q} \\in \\mathbb{R}^d$ and key matrix $\\mathbf{K} \\in \\mathbb{R}^{n \\times d}$ are encoded using", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17428", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6ef999bfc9198042cd8cfcd4d703bf261ddae56ca0f7c80b9cc8830ae68e2988.jpg", + "image_caption": [ + "Figure 3: Top Row: Overview of our proposed architecture. The input image is fed into the patch embedding layer, followed by hierarchical stages at four different scales $\\{\\frac{1}{4},\\frac{1}{8},\\frac{1}{16},\\frac{1}{32}\\}$ . Each stage is consistent and compose of Conv. Encoder blocks followed by SwiftFormer Encoder. Between two consecutive stages, we incorporate downsampling layer to reduce the spatial size by a factor of two and increase the feature dimensions. Bottom Row: We show the design of the Conv. Eencoder (left) and the SwiftFormer Encoder (right). The Conv. Encoder is designed to learn effective local representations and consists of $3\\times 3$ depth-wise convolutions followed by two point-wise convolutions for channel mixing. The SwiftFormer Encoder aims to learn enriched local-global representations. It begins with local convolutional layers to extract local features, followed by the efficient additive attention module (see Fig. 2 (d)) and linear layers." + ], + "image_footnote": [], + "bbox": [ + 80, + 92, + 467, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the element-wise product to form global context $(\\mathbb{R}^{n\\times d})$ . This matrix shares a similarity with the attention matrix in MHSA and captures information from every token and has the flexibility to learn the correlation in the input sequence. However, it is comparatively inexpensive to compute compared to MHSA and has linear complexity with the token length. Inspired by the transformer architecture, we employ a linear transformation layer to query-key interactions to learn the hidden representation of the tokens. The output of the efficient additive attention $\\hat{\\mathbf{x}}$ can be described as:", + "bbox": [ + 75, + 550, + 468, + 702 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} = \\hat {\\mathbf {Q}} + \\mathbf {T} (\\mathbf {K} * \\mathbf {q}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 709, + 468, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{Q}}$ denotes to the normalized query matrix, $\\mathbf{T}$ denotes to the linear transformation.", + "bbox": [ + 76, + 746, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. SwiftFormer Architecture", + "text_level": 1, + "bbox": [ + 76, + 786, + 313, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our SwiftFormer is based on the recently introduced EfficientFormer [20]. The main idea of EfficientFormer is to introduce 4D MetaBlocks based on PoolFormer [53] to learn local representations efficiently, while using 3D MetaBlocks based on self-attention to encode global context. However, the performance of EfficientFormer is lim", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ited by two design choices. Firstly, it uses ineffective token mixing, and secondly, it only employs 3D MetaBlocks in the last stage due to quadratic complexity of MHSA. This likely leads to inconsistent and insufficient contextual representation. To address these limitations, our SwiftFormer improves the token mixing by using a simple yet effective Conv. Encoder. Further, we introduce efficient additive attention module that can be incorporated in all stages (Sec. 3.2). This leads to more consistent learning of local-global representations. It is worth mentioning that EfficientFormer employs a latency-driven slimming method to obtain optimal configurations for its model variants, which leads to maximizing the speed. In contrast, our SwiftFormer models are built without using any neural architecture search.", + "bbox": [ + 496, + 90, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3 shows an overview of our proposed architecture. The main components are: (i) Effective Conv. Encoder, and (ii) SwiftFormer Encoder. In contrast to other hybrid designs, the proposed architecture is consistent and has Conv. Encoders followed by SwiftFormer Encoder in all stages. Our architecture extracts hierarchical features at four different scales across four stages. At the beginning of the network, the input image of size $H \\times W \\times 3$ is fed through Patch Embedding layer, implemented with two $3 \\times 3$ convolutions with a stride of 2, resulting $\\frac{H}{4} \\times \\frac{W}{4} \\times C_1$ feature maps. Then, the output feature maps are fed into the first stage, which begins with Conv. Encoder to extract spatial features, followed by SwiftFormer to learn the local-global information. Between two consecutive stages, there is a downsampling layer to increase the channel dimension and reduce the token length. Next, the resulting feature maps are subsequently fed into the second, third, and fourth stages of the architecture, producing $\\frac{H}{8} \\times \\frac{W}{8} \\times C_2$ , $\\frac{H}{16} \\times \\frac{W}{16} \\times C_3$ , and $\\frac{H}{32} \\times \\frac{W}{32} \\times C_4$ dimensional feature maps, respectively. Hence, each stage learns local-global features at different scales of the input image, which allows the network to have enriched representation.", + "bbox": [ + 496, + 303, + 892, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effective Conv. Encoder: The baseline EfficientFormer [20] employs $3 \\times 3$ average pooling layers as a local token mixer, similar to PoolFormer [53]. Although PoolFormer layers are known for their fast inference speed, replacing them with depth-wise convolutions does not increase the latency. Further, it improves the performance without increasing the parameters and latency. Specifically, the features maps $\\mathcal{X}_i$ are fed into $3 \\times 3$ depth-wise convolution (DWConv) followed by Batch Normalization (BN). Then, the resulting features are fed into two point-wise convolutions $(\\mathrm{Conv}_1)$ alongside GeLU activation. Finally, we incorporate a skip connection to enable information to flow across the network. The Conv. Encoder is defined as:", + "bbox": [ + 496, + 635, + 893, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {X}} _ {i} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {C o n v} _ {1, G} \\left(\\operatorname {D W C o n v} _ {B N} \\left(\\mathcal {X} _ {i}\\right)\\right)\\right) + \\mathcal {X} _ {i}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 840, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_i$ refers to the input features, $\\mathrm{Conv}_{1,G}$ refers to point-wise convolution followed by GeLU, DWConv $_{BN}$", + "bbox": [ + 496, + 869, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17429", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "refers to depth-wise convolution followed by Batch Normalization, and $\\hat{\\mathcal{X}}_i$ refers to the output feature maps.", + "bbox": [ + 75, + 90, + 467, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SwiftFormer Encoder: This module is carefully designed to efficiently encode enriched local-global representation in each stage. As shown in Fig. 3, the initial block of the SwiftFormer Encoder is composed of $3 \\times 3$ depth-wise convolution followed by point-wise convolution, which enables the module to learn spatial information and encode local representation. Then, the resulting feature maps are fed into the efficient additive attention block, which aims to learn contextual information at each scale of the input size. Finally, the output feature maps are fed into a Linear block, which composes of two $1 \\times 1$ point-wise convolution layers, Batch Normalization, and GeLU activation to generate non-linear features. The SwiftFormer Encoder is described as:", + "bbox": [ + 75, + 121, + 467, + 316 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {X}} _ {i} = \\operatorname {C o n v} _ {1} (\\operatorname {D W C o n v} _ {B N} (\\hat {\\mathcal {X}} _ {i})),\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 327, + 380, + 345 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {X}} _ {i} = \\mathrm {Q K} (\\hat {\\mathcal {X}} _ {i}) + \\hat {\\mathcal {X}} _ {i}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 349, + 468, + 367 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {X}} _ {i + 1} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {C o n v} _ {B N, 1, G} \\left(\\hat {\\mathcal {X}} _ {i}\\right)\\right) + \\hat {\\mathcal {X}} _ {i}.\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 371, + 410, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv}_{BN,1,G}$ denotes to Batch Normalization, followed by, $1\\times 1$ Conv layer, followed by GeLU, and QK denotes the efficient additive attention (explained in Sec. 6).", + "bbox": [ + 76, + 401, + 467, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 460, + 207, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our SwiftFormer models across four downstream tasks: classification on ImageNet-1K [7], object detection and instance segmentation on MS-COCO 2017 [21], and semantic segmentation on ADE20K [58].", + "bbox": [ + 75, + 487, + 467, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 559, + 294, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ImageNet-1K [7]: All of our models are trained from scratch on ImageNet-1K dataset for 300 epochs with AdamW optimizer [26] and cosine learning rate scheduler with an initial learning rate of $1\\mathrm{e}^{-3}$ . We use a linear warm-up for 5 epochs. We use an image resolution of $224\\times 224$ for both training and testing. Following the training recipe of [20], we use the same teacher model for distillation [36]. The experiments are conducted with PyTorch 1.12 [35] using 8 NVIDIA A100 GPUs. The latency is measured using iPhone 14 (iOS 16), and the throughput is measured using A100 40 GB GPU. For latency measurements, we compile the models using CoreML library [6] and perform inference with a batch size of 1. For the throughput on A100, the inference is performed using a batch size of 128.", + "bbox": [ + 75, + 582, + 467, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MS-COCO 2017 [21]: We use our ImageNet pre-trained models as the backbones in Mask-RCNN framework for object detection and instance segmentation on MS-COCO 2017 dataset. The dataset contains 118K training and 5K validation images. Following [20], we finetune our models for 12 epochs with an image size of $1333 \\times 800$ and batch size of 32 using AdamW optimizer. We use learning rate of", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$2e^{-4}$ and report the performance for detection and instance segmentation in terms of mean average precision (mAP).", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ADE20K [58]: The dataset comprises 20K training and 2K validation images and contains 150 class categories for scene parsing. Similar to [20] we use our ImageNet pre-trained models to extract image features and semantic FPN [18] as a decoder for segmentation. The model is trained with an image size of $512 \\times 512$ for 40K iterations with a batch size of 32 using AdamW optimizer. We use poly learning rate scheduling with an initial learning rate of $2\\mathrm{e}^{-4}$ . We report the semantic segmentation performance in terms of mean intersection over union (mIoU).", + "bbox": [ + 496, + 121, + 890, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Baseline Comparison", + "text_level": 1, + "bbox": [ + 498, + 282, + 699, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 illustrates the impact of integrating our proposed contributions into the baseline EfficientFormer-L1 [20] model in terms of ImageNet-1K top-1 accuracy and inference speed. The first row shows the results of the baseline model, which only includes the self-attention based transformer block in the final stage of the network and achieves a top-1 accuracy of $79.2\\%$ with a latency of $1.1~\\mathrm{ms}$ on an iPhone 14 mobile device. The second row replaces the pool mixers in the baseline model with our proposed Conv. Encoder, resulting in an improvement in performance to $79.9\\%$ while maintaining the same latency. In the third row, we replace the transformer block in the baseline with our proposed SwiftFormer Encoder built on the efficient additive attention. Although the performance drops by $0.2\\%$ , the inference speed improves by $0.1~\\mathrm{ms}$ , and the model has linear complexity with the number of tokens. This enables us to integrate the SwiftFormer Encoder into all stages and achieve better performance while maintaining the same inference speed as of baseline (first versus last row).", + "bbox": [ + 496, + 306, + 890, + 594 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6b40b45ea97a109b9ede1b45154284d8d7fa0244a013478b245d43f78e847285.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLatency (ms)Top-1 (%)
EfficientFormer-L1 (Baseline)1.179.2
+ Replace Pool Mixer by effective Conv. Encoder1.179.9
+ Replace Transformer block by SwiftFormer Encoder1.079.7
+ Incorporate SwiftFormer block across all stages1.180.9
", + "bbox": [ + 503, + 606, + 888, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Baseline comparison between our SwiftFormer-L1 and EfficientFormer-L1 [20] on the ImageNet-1K dataset. The latency is measured on iPhone14 Neural Engine.", + "bbox": [ + 498, + 681, + 890, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Image Classification", + "text_level": 1, + "bbox": [ + 498, + 756, + 691, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 presents a comparison of our proposed SwiftFormer models (XS, S, L1, and L3) with previous state-of-the-art ConvNets, transformer-based, and hybrid models. We show that our models set new state-of-the-art results, and outperform the recently introduced EfficientFormer [20] and MobileViT-v2 [31] in all model variants. This comprehensive evaluation shows the advantage of our proposed models in terms of both accuracy and latency on mobile devices.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17430", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fa959215b49c283ceec6f8f0be11eeaceda165136ba5a279f04be1d7d3339bc8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTypeLatency (ms) ↓Throughput (A100) ↑Params(M) ↓GMACs ↓Neural SearchTop-1(%) ↑
MobileNet-v2×1.0 [39]ConvNet0.898893.50.3✘71.8
MobileNet-v3-Large×0.75 [15]ConvNet0.8109344.00.2✓73.3
EdgeViT-XXS [34]Hybrid1.759654.10.6✘74.4
MobileViT-XS [30]Hybrid1.537072.30.7✘74.8
SwiftFormer-XSHybrid0.760343.50.6✘75.7
MobileNet-v2×1.4 [39]ConvNet0.974476.10.6✘74.7
MobileNet-v3-Large [15]ConvNet0.9103515.40.3✓75.1
EfficientNet-b0 [42]ConvNet1.385375.30.4✓77.1
DeiT-T [44]Transformer3.858605.73.8✘72.2
EdgeViT-XS [34]Hybrid2.748126.71.1✘77.5
MobileViT-v2×1.0 [31]Hybrid1.732014.91.8✘78.1
SwiftFormer-SHybrid0.850516.11.0✘78.5
MobileFormer-508M [3]Hybrid3.0444314.00.5✘79.3
PoolFormer-S12 [53]Pool1.2322712.01.8✘77.2
EdgeViT-S [34]Hybrid3.5425613.11.9✘81.0
EfficientFormer-L1 [20]Hybrid1.1504612.31.3✓79.2
MobileViT-v2×1.5 [31]Hybrid3.4235610.64.0✘80.4
SwiftFormer-L1Hybrid1.1446912.11.6✘80.9
ResNet-50 [13]ConvNet1.9483525.54.1✘78.5
PoolFormer-S36 [53]Pool2.8111431.05.0✘81.4
ConvNeXt-T [25]ConvNet2.5323528.64.5✘82.1
DeiT-S [44]Transformer9.9299022.54.5✘81.8
Swin-T [25]TransformerNA263528.34.5✘81.3
MobileViT-v2×2.0 [31]Hybrid5.0190618.57.5✘81.2
EfficientFormer-L3 [20]Hybrid2.0269131.33.9✓82.4
SwiftFormer-L3Hybrid1.9289028.54.0✘83.0
", + "bbox": [ + 89, + 88, + 879, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Comparison of our proposed SwiftFormer with the state-of-the-art counterpart models on ImageNet-1K. The latency is measured on iPhone 14 Neural Engine (iOS 16) and the throughput is measured on Nvidia A100 GPU. Our models run faster than MobileNets, Hybrid, and Transformer models, with a better trade-off between accuracy and model complexity. The error for the latency measurement is less than $\\pm 0.1$ ms. The input size remains consistent at $224\\times 224$ for all models, except for the MobileViT [30] and MobileViT-v2 [31] variants, which employ a larger dimension of $256\\times 256$ . Our results are shown in bold for all model variants.", + "bbox": [ + 75, + 429, + 890, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with ConvNets: Our SwiftFormer models surpass the widely used lightweight CNNs counterparts significantly in terms of top-1 accuracy, while running faster than the highly optimized MobileNet-v2 and MobileNet-v3 on an iPhone 14 mobile device. Specifically, our SwiftFormer-XS runs $0.1\\mathrm{ms}$ faster than MobileNet-v2 $\\times 1.0$ and MobileNet-v3-Large $\\times 0.75$ and achieve better top-1 accuracy with a margin of $3.9\\%$ and $2.4\\%$ respectively. Our SwiftFormer-S runs faster than EfficientNet-b0 [42] by $1.6\\times$ and achieves $1.4\\%$ higher top-1 accuracy. Further, our SwiftFormer-L3 achieves $4.5\\%$ and $0.9\\%$ gain in top-1 accuracy over ResNet-50 and ConvNeXt-T, respectively, while running at the same latency as ResNet-50 and $1.3\\times$ faster than ConvNeXt-T. This demonstrates that our SwiftFormer models, powered by our proposed efficient additive attention, run faster than the lightweight CNN models on mobile devices and achieve superior performance. Recent device-level optimizations for CNN-based models, such as dedicated hardware implementations for convolutions with batch normalization and non-linearity, likely contribute to the high throughput of fully CNN-based models on A100.", + "bbox": [ + 75, + 547, + 472, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with transformer models: Although transformer models usually outperform CNN-based models in", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "terms of accuracy, they tend to suffer from high latency when running on resource-constrained mobile devices. For instance, DeiT-S, which has a similar model size to ResNet-50 and achieves higher top-1 accuracy by $3.3\\%$ , but ResNet-50 runs approximately $5.2\\times$ faster on an iPhone 14 mobile device. In contrast, our SwiftFormer-L3 model achieves $1.2\\%$ higher accuracy than DeiT-S, while running at the same speed as ResNet-50. Further our SwiftFormer-S model runs approximately $4.7\\times$ faster than DeiT-T on an iPhone 14 mobile device and has $6.3\\%$ better accuracy.", + "bbox": [ + 496, + 547, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with hybrid models: Although most existing hybrid approaches achieve higher accuracy compared to their lightweight CNN counterparts, they still underperform the fully CNN-based models in terms of latency due to the quadratic complexity of multi-head self-attention. For example, EdgeViT-XXS runs at approximately $2 \\times$ slower compared to MobileNet-v3-Large $\\times 0.75$ . On the other hand, our SwiftFormer-XS has better latency as compared to lightweight CNNs and approximately is $2 \\times$ faster than EdgeViT-XXS and MobileViT-XS, with an overall $1.3\\%$ and $0.9\\%$ higher top-1 accuracy respectively. Further, our SwiftFormer-L1 model is $3 \\times$ faster than the state-of-the-art MobileViT-v2 $\\times 1.5$ with $0.5\\%$ better top-1 ac", + "bbox": [ + 496, + 704, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17431", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e49862393aaa0bd44e73b967b0784e6bdb1d4b23b216c41afad3870eb217a7ee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneDetection & Instance SegmentationSemantic mIoU(%)
\\(AP^{box}\\)\\(AP^{box}_{50}\\)\\(AP^{box}_{75}\\)\\(AP^{mask}\\)\\(AP^{mask}_{50}\\)\\(AP^{mask}_{75}\\)
ResNet18 [13]34.054.036.731.251.032.732.9
PoolFormer-S12 [53]37.359.040.134.655.836.937.2
EfficientFormer-L1 [20]37.960.341.035.457.337.338.9
SwiftFormer-L141.263.244.838.160.240.741.4
ResNet50 [13]38.058.641.434.455.136.736.7
PVT-S [48]40.462.943.837.860.140.339.9
PoolFormer-S24 [53]40.162.243.437.059.139.640.3
Swin-T [24]42.264.446.239.164.642.041.5
EfficientFormer-L3 [20]41.463.944.738.161.040.443.5
SwiftFormer-L342.764.446.739.161.741.843.9
", + "bbox": [ + 169, + 88, + 802, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Results using SwiftFormer as a backbone on dense prediction tasks: Object detection and instance segmentation on COCO, whereas semantic segmentation on ADE20K. Our approach outperforms EfficientFormer on all three tasks.", + "bbox": [ + 75, + 273, + 893, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/295cd82e9390bbd1cea73896c3504b1298ca26843493b0b6f9729fbacfc24afa.jpg", + "image_caption": [ + "Figure 4: Results on COCO. Examples for object detection and instance segmentation on the COCO 2017 validation set. Our SwiftFormer-L1 model can accurately detect and segment instances in images." + ], + "image_footnote": [], + "bbox": [ + 78, + 318, + 259, + 406 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5dc630255cce0a020d28a90ae146c07d7d839652d0a2b22b81111e95ce85774d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 319, + 352, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f5511ed9b71067fa4539f7760bfa92c0596035ed6a24f3bba2d43339a03243a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 319, + 529, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5cc1fe4e0aa983b347a43fdec61313bce1af823ff345d17e6bd8e999e35ad09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 319, + 710, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/914771bd7f12b94018c4bc67e706dbaeed42cc6cb34010d9f6db9333962d3b02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 319, + 887, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a639e07910c3252abbe5de29bccc18a469cc7f820884a8dbc265144dceed7cce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 472, + 212, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5615d815e280269b13c6c24ea83df91e691da0872cb42b00e5d61d7ecbd747df.jpg", + "image_caption": [ + "Figure 5: Qualitative results on ADE20K. The qualitative examples for semantic segmentation on the ADE20K validation set. Top: Ground truth masks. Bottom: The semantic segmentation results. Our model can accurately segment various indoor and outdoor scenes." + ], + "image_footnote": [], + "bbox": [ + 80, + 551, + 212, + 630 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/272a5b097220bf0f4808b79c284096d7c3f9f447a39ff2753eab15774ee72050.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 472, + 349, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9279f7c30948903d961f0289bff25484ed1442fda3204ff4234cb952b451091f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 551, + 349, + 630 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/712290b37e3e0bc4db7044f856c47bbc5772196fea7774ffb7c40bbece5ce73b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 472, + 482, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9cae50157efffee9b9b32a47d3c380451df466982222e9f62b8e040372385f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 551, + 482, + 630 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29f03f414fc104cde489597b2b32bded815b852573c168daa1a0ce3376db00b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 472, + 607, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac21309e543873f5de17f3015c2d60f2d2c64f7662e2ab09547b47005b402c35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 551, + 607, + 630 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3bdb2e9c3057e1537369146c340d0a5b2a01ea82b951f34f1752174a0b4e2ff1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 472, + 753, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/272b347c72b6d1de59e27791cc153832d17b40d4221de4471c00d320b54d8e16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 551, + 753, + 630 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9bcfdba7e715d8f7f5ea90f643413750e7b96561e79287d0c47f7ff062fc2cb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 472, + 888, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78589171dd9e4d09c4e83b14ec5acad1bdb7eda1aa02e32500087a9ca91b3c21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 551, + 888, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "curacy. Our SwiftFormer-L3 model achieves $83.0\\%$ top-1 accuracy and runs at $1.9~\\mathrm{ms}$ , which is $2.6\\times$ faster than MobileViT-v2 $\\times 2.0$ with an absolute $1.8\\%$ accuracy gain.", + "bbox": [ + 75, + 715, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Knowledge distillation (KD): It is worth mentioning that we use the same teacher model of DeiT [44] and EfficientFormer [20] in Table 2. We perform a baseline comparison to show the efficiency of our method without using KD. EfficientFormer-L1 achieves $77.3\\%$ accuracy w/o KD, compared to $79.2\\%$ with KD. Our SwiftFormer-L1 achieves $79.8\\%$ w/o KD, compared to $80.9\\%$ with KD. This shows that our approach achieves a higher absolute gain of $2.5\\%$ over EfficientFormer-L1, when not using KD.", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our SwiftFormer-XS and SwiftFormer-S w/o KD achieve $74.1\\%$ and $77.0\\%$ , respectively.", + "bbox": [ + 498, + 717, + 890, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Object Detection and Instance Segmentation", + "text_level": 1, + "bbox": [ + 498, + 756, + 877, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 compares the object detection and instance segmentation results of Mask-RCNN with different lightweight backbones. Our SwiftFormer-L1 backbone achieves 41.2 AP box, surpassing the lightweight ResNet18 and PoolFormer-S12 backbones by 7.2 and 3.9 points respectively. Further, it performs better than the previous state-of-the-art EfficientFormer-L1 backbone by 3.3 AP box. For instance segmentation, our method achieves 38.1 AP mask", + "bbox": [ + 496, + 780, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17432", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "score which is 2.7 points better than the previous state-of-the-art. Similar trend is observed for SwinfFormer-L3 backbone, which surpasses the previous state-of-the-art EfficientFormer-L3 backbone by 1.3 points and 1.0 points in AP box and mask respectively. The improvement in the downstream detection and instance segmentation tasks illustrates the effectiveness of our SwiftFormer backbone models for the dense prediction tasks.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5. Semantic Segmentation", + "text_level": 1, + "bbox": [ + 76, + 222, + 294, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 3 shows the semantic segmentation results of SwiftFormer backbone-based models as compared to previously proposed backbones. We achieve $41.4\\%$ mean intersection over union score using SwiftFormer-L1, surpassing ResNet18 by $8.5\\%$ , PoolFormer-S12 by $4.2\\%$ , and the state-of-the-art EfficientFormer-L1 by $2.5\\%$ . Similarly, our SwiftFormer-L3 backbone-based segmentation models achieve $43.9\\mathrm{mIoU}$ , surpassing all previous methods.", + "bbox": [ + 75, + 246, + 468, + 367 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 378, + 261, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 4 shows the qualitative results of our detection and instance segmentation model on COCO dataset. We also show in Fig. 5 qualitative examples for semantic segmentation results on the ADE20K dataset. Additionally, we demonstrate in Fig. 6 a baseline comparison with the baseline EfficientFormer-L1. The baseline misclassifies a bird as a sheep in the first example and as a bear in the second example, while misclassifies kites as birds in the third example. Our SwiftFormer-L1 accurately detects and segments objects in these examples, which demonstrates the ability of SwiftFormer to accurately identify and segment objects within a diverse range of indoor and outdoor settings.", + "bbox": [ + 75, + 402, + 470, + 583 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d53b4d09eb86ba9b0d12a966bb820e78bb017b3f58dfd2d3d4801b702e9d4481.jpg", + "image_caption": [ + "Figure 6: Qualitative comparison between EfficientFormer-L1 (top) and our SwiftFormer-L1 (bottom) on images from COCO val set for detection and instance segmentation." + ], + "image_footnote": [], + "bbox": [ + 81, + 597, + 465, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 829, + 225, + 847 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To show the effectiveness of the proposed efficient additive attention, we conduct a comparison with other attention variants over SwiftFormer-L1 in Table. 4. It is notable that", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/42ae0c9507ac5999a06d217faffda781199c0427f9ae7572024d5bc76ca0a68d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Attention MethodLatency (ms)↓Throughput↑Top-1 (%)↑
Self-attention (Heads=2)8.5331680.6
Self-attention (Heads=4)18.0303581.1
Transpose self-attention (Heads=2)2.0367180.2
Separable self-attention1.4394080.6
Efficient Additive Attention (Ours)1.1446980.9
", + "bbox": [ + 504, + 88, + 890, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4: Quantitative analysis of Fig. 2. We replace our attention module with these methods in the same setting.", + "bbox": [ + 500, + 170, + 890, + 200 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/346f957dec657f7ad933f2eb7e9d5a7a677f80e5eb4224f0e646efe85a1d35dc.jpg", + "image_caption": [ + "Figure 7: Visualization of the learned global query vector of the efficient additive attention from SwiftFormer-L1 model at different stages. Both the first and last stages show that the global context is captured effectively." + ], + "image_footnote": [], + "bbox": [ + 501, + 215, + 890, + 367 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the proposed efficient additive attention achieves the best trade-off between latency and top-1 accuracy. To demonstrate further insights on the interpretability of the proposed efficient additive attention, we visualize the learned global attention query vector $\\alpha$ of Eq. 4 of the first and last stages of the network. The global query vector of the efficient additive attention aligns with the semantics of the image and hence acts as an informative source in the additive attention mechanism. Also, it localizes the features of the objects accurately with fine details in the first stage.", + "bbox": [ + 496, + 465, + 890, + 618 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 632, + 619, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Transformers have gained popularity in vision applications due to their effective use of self-attention computation. However, their use in mobile vision applications is challenging due to the quadratic nature of self-attention, which is computationally expensive on mobile devices. In this work, we propose a novel efficient additive attention that replaces the expensive matrix multiplication operations with element-wise operations and eliminates explicit keys-values interaction. Our proposed attention is linear with respect to the input tokens and used in all stages of the network.", + "bbox": [ + 496, + 657, + 890, + 809 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although we achieve a superior balance between accuracy and efficiency, we observe that in the case of dense small object detection, our approach is less accurate likely due to not using positional encoding or attention biases. Our future work is to improve the performance of efficient additive attention in dense prediction tasks.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17433", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adrian Bulat and Georgios Tzimiropoulos. Bit-mixer: Mixed-precision networks with runtime bit-width selection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021.", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020.", + "[3] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[4] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[5] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In Advances in Neural Information Processing Systems, 2021.", + "[6] CoreMLTools. Use coremltools to convert models from third-party libraries to core ml., 2021.", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2009.", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[9] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021.", + "[10] Mohsen Fayyaz, Soroush Abbasi Kouhpayegani, Farnoush Rezaei Jafari, Eric Sommerlade, Hamid Reza Vaezi Joze, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers. 2022.", + "[11] TSong Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016.", + "[12] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023.", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021.", + "[15] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenet-v3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019.", + "[16] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017.", + "[17] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, 2018.", + "[18] A. Kirillov, R. Girshick, K. He, and P. Dollar. Panoptic feature pyramid networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[19] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In International Conference on Learning Representations, 2020.", + "[20] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. 2022.", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014.", + "[22] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco De Nadai. Efficient training of visual transformers with small datasets. In Advances in Neural Information Processing Systems, 2021.", + "[23] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021.", + "[25] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019.", + "[27] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In European Conference on Computer Vision, 2018." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17434", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In European Conference on Computer Vision, 2022.", + "[29] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: Efficiently amalgamated cnn-transformer architecture for mobile vision applications. In CADL2022, 2022.", + "[30] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations, 2022.", + "[31] Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. Transactions on Machine Learning Research, 2023.", + "[32] Sachin Mehta, Mohammad Rastegari, Linda Shapiro, and Hannaneh Hajishirzi. Espnetv2: A light-weight, power efficient, and general purpose convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019.", + "[33] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021.", + "[34] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martinez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In European Conference on Computer Vision, 2022.", + "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019.", + "[36] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In CVPR, 2020.", + "[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems, 2021.", + "[38] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In Advances in Neural Information Processing Systems, 2021.", + "[39] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenet-v2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 2018.", + "[40] Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan." + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Unetr++: Delving into efficient and accurate 3d medical image segmentation. arXiv:2212.04497, 2022.", + "[41] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019.", + "[42] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, 2019.", + "[43] Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, et al. Mlp-mixer: An all-mlp architecture for vision. In Advances in Neural Information Processing Systems, 2021.", + "[44] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2021.", + "[45] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022.", + "[46] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020.", + "[47] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In CVPR, 2021.", + "[48] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021.", + "[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media, 2022.", + "[50] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. In Advances in Neural Information Processing Systems, 2016.", + "[51] Chuhan Wu, Fangzhao Wu, Tao Qi, and Yongfeng Huang. Fastformer: Additive attention can be all you need. arXiv preprint arXiv:2108.09084, 2021.", + "[52] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022.", + "[53] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022.", + "[54] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Edgeformer: Improving light-weight convnets by learning from vision transformers. arXiv preprint arXiv:2203.03952, 2022." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17435", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Parc-net: Position aware circular convolution with merits from convnets and transformer. In European Conference on Computer Vision, 2022.", + "[56] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, 2022.", + "[57] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018.", + "[58] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.", + "[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021." + ], + "bbox": [ + 78, + 90, + 468, + 402 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17436", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/3fb2491f-940c-4884-86b5-1f7be4430698_model.json b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/3fb2491f-940c-4884-86b5-1f7be4430698_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3f96b8ead607ca902b85ebf6aa35733d995c9e5e --- /dev/null +++ b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/3fb2491f-940c-4884-86b5-1f7be4430698_model.json @@ -0,0 +1,2402 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.13, + 0.819, + 0.178 + ], + "angle": 0, + "content": "SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.203, + 0.821, + 0.278 + ], + "angle": 0, + "content": "Abdelrahman Shaker\\(^{1*}\\) Muhammad Maaz\\(^{1}\\) Hanoona Rasheed\\(^{1}\\) Salman Khan\\(^{1}\\) Ming-Hsuan Yang\\(^{2,3,4}\\) Fahad Shahbaz Khan\\(^{1,5}\\) \n\\(^{1}\\)Mohamed bin Zayed University of AI \\(^{2}\\)University of California, Merced \n\\(^{3}\\)Yonsei University \\(^{4}\\)Google Research \\(^{5}\\)Linköping University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.474, + 0.705 + ], + "angle": 0, + "content": "Self-attention has become a defacto choice for capturing global context in various vision applications. However, its quadratic computational complexity with respect to image resolution limits its use in real-time applications, especially for deployment on resource-constrained mobile devices. Although hybrid approaches have been proposed to combine the advantages of convolutions and self-attention for a better speed-accuracy trade-off, the expensive matrix multiplication operations in self-attention remain a bottleneck. In this work, we introduce a novel efficient additive attention mechanism that effectively replaces the quadratic matrix multiplication operations with linear element-wise multiplications. Our design shows that the key-value interaction can be replaced with a linear layer without sacrificing any accuracy. Unlike previous state-of-the-art methods, our efficient formulation of self-attention enables its usage at all stages of the network. Using our proposed efficient additive attention, we build a series of models called \"Swift-Former\" which achieves state-of-the-art performance in terms of both accuracy and mobile inference speed. Our small variant achieves \\(78.5\\%\\) top-1 ImageNet-1K accuracy with only 0.8 ms latency on iPhone 14, which is more accurate and \\(2\\times\\) faster compared to MobileViT-v2. Our code and models: https://tinyurl.com/5ft8v46w" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.732, + 0.21, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.47, + 0.879 + ], + "angle": 0, + "content": "In recent years, transformer models have shown remarkable success in various vision applications such as classification [8, 44, 24, 23, 9], detection [2, 59, 33, 56, 28], and segmentation [4, 40]. However, deploying these models on resource-constrained mobile devices for real-time applications remains challenging due to their inherently complex nature [20, 29]. Specifically, vision transformers (ViTs) rely on global self-attention, which has a quadratic com" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.308, + 0.895, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.573, + 0.895, + 0.648 + ], + "angle": 0, + "content": "Figure 1: Latency vs Accuracy Comparison. Compared to the recent EfficientFormer-L1 [20], our SwiftFormer-L1 achieves an absolute gain of \\(1.7\\%\\) in terms of top-1 accuracy with the same latency and without requiring any neural architecture search." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.895, + 0.881 + ], + "angle": 0, + "content": "plexity with respect to the input image resolution, making it impractical for deployment on low-powered mobile devices [31]. As a result, convolutional neural networks (CNNs) are still the preferred choice for real-time deployment on mobile devices, primarily because the convolution operation is computationally efficient [39, 15]. However, a major limitation of CNNs is their reliance on local connections and stationary weights, which can limit their ability to adapt to variable input resolutions and capture long-range dependencies in the data. Therefore, developing more efficient and flexible models that combine the strengths of both CNNs and transformers is critical, particularly for mobile devices with limited computational resources." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To achieve this goal, several hybrid approaches have" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.428, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author: abdelrahman.youssief@mbzai.ac.ae" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17425" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "been proposed that use lightweight CNN modules in the high-resolution early stages and self-attention in the low-resolution later stages [55, 29, 20]. This approach effectively increases the receptive field of the network and strives to achieve a trade-off between speed and accuracy. Furthermore, different efficient variants of computing self-attention have been proposed to reduce the model complexity. These include computing attention across feature dimensions to implicitly model the global context [29], computing attention within local windows [24], pooling spatial features before applying self-attention [9], and sparsely attending to a fixed number of tokens [34], to name a few." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.473, + 0.416 + ], + "angle": 0, + "content": "Although these approaches effectively reduce network complexity, they still involve inefficient matrix multiplication operations that significantly impact latency on mobile devices. To address this issue, Mehta et al. [31] propose a separable self-attention mechanism that replaces matrix multiplication operations to element-wise multiplications. This is achieved by projecting queries to context scores, followed by element-wise multiplication with keys to calculate context vectors for encoding global context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.473, + 0.556 + ], + "angle": 0, + "content": "In this work, we propose efficient additive attention, which eliminates the need for expensive matrix multiplication operations in computing self-attention. Additionally, we propose to compute the global context using only the query-key interactions followed by a linear transformation, without requiring explicit key-value interactions. This significantly reduces the computational complexity and enables us to use the proposed attention block in all stages of the network. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.582, + 0.47, + 0.658 + ], + "angle": 0, + "content": "- We introduce efficient additive attention, a new approach for computing self-attention in vision backbones that eliminates the need for expensive matrix multiplication operations, significantly reducing the computational complexity of the model." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.673, + 0.47, + 0.735 + ], + "angle": 0, + "content": "- Unlike previous methods, our proposed efficient attention design can be used at all stages of the network, enabling more effective contextual information capture and achieving superior speed-accuracy trade-off." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- We build a series of efficient generic classification models called \"SwiftFormer\", which utilize our proposed efficient additive attention. Our small model achieves \\(78.5\\%\\) top-1 ImageNet-1K [7] accuracy while running at only \\(0.8~\\mathrm{ms}\\) latency on iPhone 14. Moreover, our large model achieves \\(83.0\\%\\) accuracy with a latency of only \\(1.9~\\mathrm{ms}\\). Our model achieves state-of-the-art performance, outperforming recent MobileViTv2 [31] and EfficientFormer [20] by obtaining a better trade-off between accuracy and latency (see Fig. 1)." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.582, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Efficient CNNs: Designing efficient CNNs for mobile vision applications has received much attention in recent years. MobileNet architectures [16, 39, 15] propose depthwise separable convolutions as well as efficient inverted residual blocks for improved performance on various vision tasks. Other methods aim to improve the efficiency by leveraging depth-wise dilated convolutions [32], channel shuffling and pointwise group convolutions [57, 27], network pruning [11, 50], low bit-width [1, 17], and neural architecture search [41, 15]. CNN-based methods are well-performing, efficient, and fast to train and run on edge devices, resulting in widespread usage in the industry. However, they are spatially local and lack global interaction between the features, which deeply affects their performance. Efficient Transformers: ViTs [8] have been widely used in numerous vision tasks, and significant advances have been made in terms of data efficiency [22, 44], transformer architecture [30, 20, 3, 54], and token mechanisms [43, 52]. Reducing the number of visual tokens is a major modification in the transformer architecture for efficient deployment. Instead of using a fixed feature representation through the whole architecture, some methods employ a hierarchical design where the resolution is gradually decreased through the stages, including down-sampling techniques [9, 40, 38, 14] and pyramidal structures [47, 49]. Recently, a few methods [37, 10, 34] propose token sparsification techniques to encode only a subset of the most informative tokens." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Numerous approaches have recently been proposed to reduce the quadratic complexity of self-attention, the computational bottleneck in transformer-based architectures, by computing its approximated variants [34, 29, 31, 46, 19, 5, 45]. EdgeViT [34] uses a global sparse attention module attending only to a few tokens to improve the efficiency, while [48] down-samples the key and value vectors that lead to a better efficiency-accuracy trade-off. GCViT [12] uses a local-global attention mechanism by employing the interactions between a global query and local key and value to capture the contextual information. This approach enables the model to encode both local and global representations efficiently. EdgeNeXt [29] adopts transposed self-attention operation to compute the attention maps across the channel dimension instead of the spatial dimension, followed by token mixing, to have a linear complexity with respect to the number of tokens. Reformer [19] replaces the dot-product attention with a locality-sensitive hashing to group the tokens and reduced the complexity from \\( O(n^{2}) \\) to \\( O(n\\log n) \\). However, this design is only efficient on longer sequences, which is typically not the case for ViTs. LinFormer [46] is a low-rank matrix factorization method that approximates the self-attention matrix with a low-rank matrix, reducing the complexity from \\( O(n^{2}) \\) to \\( O(n) \\). Although matrix factorization methods theoretically reduce the complexity of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "17426" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "self-attention, they use expensive projections for computing attention, which may not reflect the reduction in FLOPs and parameters into actual speed on mobile platforms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.471, + 0.424 + ], + "angle": 0, + "content": "Although these methods show promise and have reduced the complexity of self-attention theoretically, they are inadequate for reducing the inference speed for mobile deployment. Since the complexity of the multi-headed self-attention (MHSA) is higher in the earlier stage compared to the last stages, EfficientFormer [20] incorporates MHSA in the last stage only to learn contextual information from the high-level features without increasing the inference speed significantly. Recently, MobileViT-v2 [31] proposes separable self-attention that uses element-wise operations instead of the dot-product to compute the attention maps with linear complexity. Different from the existing approaches, we propose a consistent hybrid design with an efficient additive attention mechanism to model the contextual information with linear complexity. Instead of capturing the pairwise interactions between keys, queries, and values using the dot-product, we use element-wise operations with learnable attention weights to model the interactions between query and keys only, leading to better inference speed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.438, + 0.169, + 0.453 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.522 + ], + "angle": 0, + "content": "Motivation: To motivate our method, we first distinguish three desirable characteristics to be considered when designing an efficient yet accurate approach for resource constraint mobile devices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Efficient Global Context Modeling: As discussed earlier, most existing approaches either employ the standard MHSA or an approximated variant to learn the global context. However, they struggle to operate as fast as MobileNets on resource-constrained devices. This is likely due to the computation-intensive multiplicative operations during attention computation or reliance on advanced reshaping and indexing operations in these approaches. For instance, the recent MobileViT-v2 [31] is \\(2 \\times\\) slower than MobileNet-v2 [39]. Instead of using matrix multiplications, we argue that encoding the global context using an efficient additive attention design can reduce the operations with respect to the number of tokens. This is expected to help operate at comparable speed and model size, while achieving superior accuracy compared to MobileNets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Rethinking key-value interactions: Other than multiplicative operations during attention computation, additive attention has been recently explored in the NLP domain [51]. However, in the standard form, it performs three-step processing to model query, key, and value interactions. Each step feeds into the subsequent one, thereby requiring sequential processing. Here, we rethink the additive attention for vision tasks by alleviating the need to compute explicit interactions between key-value. We argue that eliminating key-value interactions and replacing them with a sim" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "ple linear transformation empirically encodes better contextual representation. Our design encodes only global queries and key interactions to learn the global contextual information, followed by a linear transformation to compute global context-aware attention weights." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.364 + ], + "angle": 0, + "content": "Consistent Hybrid Design: Most existing works employ MHSA or the approximated variant in the last stages, while avoiding its usage in the earlier stages. This is because the computational complexity of MHSA grows quadratically with the length of the tokens, making it impractical to incorporate during the initial stages. This constraint adds to the design complexity and requires a careful selection of stages, where MHSA can be applied. In contrast, our proposed SwiftFormer module has linear complexity with respect to the token length and can be incorporated in all stages to learn consistent global context at each scale. This consistency improves the model performance and makes it more generalizable and scalable for high-resolution images." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.371, + 0.779, + 0.387 + ], + "angle": 0, + "content": "3.1. Overview of Attention Modules" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.53 + ], + "angle": 0, + "content": "Vision transformer models are built upon the self-attention (see Fig. 2 (a)), which can effectively model the interactions between the input tokens. Specifically, the self-attention has \\(\\mathbf{x}\\) as an input, where \\(\\mathbf{x} \\in \\mathbb{R}^{n \\times d}\\), comprising \\(n\\) tokens with \\(d\\)-dimensional embedding vector. The input \\(\\mathbf{x}\\) is projected to query \\((\\mathbf{Q})\\), key \\((\\mathbf{K})\\), and value \\((\\mathbf{V})\\) using three matrices, \\(\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}\\), and \\(\\mathbf{W}_{\\mathbf{V}}\\). Each self-attention layer comprises \\(h\\) heads, which allows the model to attend to different views of the input. The self-attention can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.537, + 0.892, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\cdot \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right) \\cdot \\mathbf {V}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.758 + ], + "angle": 0, + "content": "The attention scores between each pair of tokens in \\(\\mathbf{Q}\\) and \\(\\mathbf{K}\\) are computed using the dot-product operation. Next, these scores are normalized followed by Softmax to weigh the interactions between the tokens. Finally, the weighted interactions are multiplied by \\(\\mathbf{V}\\) using the dot-product operation to produce the final weighted output. Overall, the complexity of the self-attention is \\(O(n^{2} \\cdot d)\\), where \\(n\\) is the number of tokens and \\(d\\) is the hidden dimension. The computational and memory demands of \\(\\mathbf{Q} \\cdot \\mathbf{K}^{\\top}\\) increase quadratically as the number of tokens grows, leading to slow inference speed and high memory usage, making it impractical to run in real-time for long sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.864 + ], + "angle": 0, + "content": "To alleviate this issue, [29] proposes the transpose self-attention (see Fig. 2 (b)) to reduce the complexity from quadratic to linear with respect to the number of tokens. Here, the dot-product operation is applied across the channel dimension instead of the spatial dimension. This allows the model to learn feature maps with implicit contextual representation. The attention can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.87, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = \\mathbf {V} \\cdot \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} ^ {\\top} \\cdot \\mathbf {K}}{\\sqrt {d}}\\right). \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17427" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.252, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.291, + 0.204, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.093, + 0.444, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.291, + 0.382, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.293, + 0.442, + 0.306 + ], + "angle": 0, + "content": "Dot product" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.093, + 0.648, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.293, + 0.467, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.294, + 0.657, + 0.307 + ], + "angle": 0, + "content": "Broadcasted element-wise multiplication" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.093, + 0.868, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Figure 2: Comparison with different self-attention modules. (a) is a typical self-attention used in ViTs [8]. (b) is the transpose self-attention used in EdgeNeXt [29], where the self-attention operation is applied across channel feature dimensions \\((\\mathrm{d} \\times \\mathrm{d})\\) instead of the spatial dimension \\((\\mathrm{n} \\times \\mathrm{n})\\). (c) is the separable self-attention of MobileViT-v2 [31], it uses element-wise operations to compute the context vector from the interactions of \\(\\mathbf{Q}\\) and \\(\\mathbf{K}\\) matrices. Then, the context vector is multiplied by \\(\\mathbf{V}\\) matrix to produce the final output. (d) Our proposed efficient additive self-attention. Here, the query matrix is multiplied by learnable weights and pooled to produce global queries. Then, the matrix \\(\\mathbf{K}\\) is element-wise multiplied by the broadcasted global queries, resulting the global context representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.469, + 0.544 + ], + "angle": 0, + "content": "The transpose self-attention has a computational complexity of \\( O(n \\cdot d^2) \\). While this complexity scales linearly with the number of tokens \\( n \\), it remains quadratic with respect to the feature dimension \\( d \\). Further, the dot-product operation is still utilized between the query and key matrices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.74 + ], + "angle": 0, + "content": "The separable self-attention mechanism (see Fig. 2 (c)) aims to address the bottleneck of the standard self-attention. Here, the interactions between the queries \\((\\mathbf{Q})\\), keys \\((\\mathbf{K})\\), and values \\((\\mathbf{V})\\) are encoded using element-wise operations. First, the query matrix \\(\\mathbf{Q}\\) is projected to produce a vector \\(\\mathbf{q}\\) of dimensions \\(n \\times 1\\), and then fed into Softmax to generate the context scores, which captures the importance of each query element. Then, the context scores are multiplied by the key matrix \\(\\mathbf{K}\\) and pooled to compute a context vector, which encodes the contextual information. Finally, the context vector is multiplied element-wise with the value matrix \\(\\mathbf{V}\\) to propagate the contextual information and produce the final output \\(\\hat{\\mathbf{x}}\\). It can be summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.749, + 0.469, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = \\mathbf {V} * \\sum \\mathbf {K} * \\operatorname {S o f t m a x} (\\mathbf {q}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.779, + 0.465, + 0.794 + ], + "angle": 0, + "content": "Here, \\(*\\) denotes the element-wise multiplication operation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.327, + 0.817 + ], + "angle": 0, + "content": "3.2. Efficient Additive Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The typical additive attention mechanism in NLP captures the global context by utilizing pairwise interactions between the tokens via element-wise multiplications instead of using dot-product operation. It encodes the relevance scores for the contextual information of the input sequence based on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.727 + ], + "angle": 0, + "content": "the interactions of the three attention components (Q, K, V). In contrast, we show that key-value interactions can be removed without sacrificing the performance and only focusing on effectively encoding query-key interactions by incorporating a linear projection layer is sufficient to learn the relationship between the tokens (see Fig. 2 (d)). This approach, named efficient additive attention, has a faster inference speed and produces more robust contextual representations as demonstrated by our performance on image classification, object detection, and segmentation tasks (Sec. 4). Specifically, the input embedding matrix \\(\\mathbf{x}\\) is transformed into query (Q) and key (K) using two matrices \\(\\mathbf{W_q}\\), \\(\\mathbf{W_k}\\), where Q, \\(\\mathbf{K} \\in \\mathbb{R}^{n \\times d}\\), \\(\\mathbf{W_q}\\), \\(\\mathbf{W_k} \\in \\mathbb{R}^{d \\times d}\\), \\(n\\) is the token length and \\(d\\) is the dimensions of the embedding vector. Next, the query matrix Q is multiplied by learnable parameter vector \\(\\mathbf{w}_a \\in \\mathbb{R}^d\\) to learn the attention weights of the query, producing global attention query vector \\(\\alpha \\in \\mathbb{R}^n\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.739, + 0.892, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\alpha = \\mathbf {Q} \\cdot \\mathbf {w} _ {a} / \\sqrt {d} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Then, the query matrix is pooled based on the learned attention weights, resulting in a single global query vector \\(\\mathbf{q} \\in \\mathbb{R}^d\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.821, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} = \\sum_ {i = 1} ^ {n} \\alpha_ {i} * \\mathbf {Q} _ {\\mathbf {i}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Next, the interactions between the global query vector \\(\\mathbf{q} \\in \\mathbb{R}^d\\) and key matrix \\(\\mathbf{K} \\in \\mathbb{R}^{n \\times d}\\) are encoded using" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17428" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.093, + 0.468, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.473, + 0.526 + ], + "angle": 0, + "content": "Figure 3: Top Row: Overview of our proposed architecture. The input image is fed into the patch embedding layer, followed by hierarchical stages at four different scales \\(\\{\\frac{1}{4},\\frac{1}{8},\\frac{1}{16},\\frac{1}{32}\\}\\). Each stage is consistent and compose of Conv. Encoder blocks followed by SwiftFormer Encoder. Between two consecutive stages, we incorporate downsampling layer to reduce the spatial size by a factor of two and increase the feature dimensions. Bottom Row: We show the design of the Conv. Eencoder (left) and the SwiftFormer Encoder (right). The Conv. Encoder is designed to learn effective local representations and consists of \\(3\\times 3\\) depth-wise convolutions followed by two point-wise convolutions for channel mixing. The SwiftFormer Encoder aims to learn enriched local-global representations. It begins with local convolutional layers to extract local features, followed by the efficient additive attention module (see Fig. 2 (d)) and linear layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.703 + ], + "angle": 0, + "content": "the element-wise product to form global context \\((\\mathbb{R}^{n\\times d})\\). This matrix shares a similarity with the attention matrix in MHSA and captures information from every token and has the flexibility to learn the correlation in the input sequence. However, it is comparatively inexpensive to compute compared to MHSA and has linear complexity with the token length. Inspired by the transformer architecture, we employ a linear transformation layer to query-key interactions to learn the hidden representation of the tokens. The output of the efficient additive attention \\(\\hat{\\mathbf{x}}\\) can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.71, + 0.47, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} = \\hat {\\mathbf {Q}} + \\mathbf {T} (\\mathbf {K} * \\mathbf {q}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.747, + 0.47, + 0.779 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{Q}}\\) denotes to the normalized query matrix, \\(\\mathbf{T}\\) denotes to the linear transformation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.315, + 0.803 + ], + "angle": 0, + "content": "3.3. SwiftFormer Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our SwiftFormer is based on the recently introduced EfficientFormer [20]. The main idea of EfficientFormer is to introduce 4D MetaBlocks based on PoolFormer [53] to learn local representations efficiently, while using 3D MetaBlocks based on self-attention to encode global context. However, the performance of EfficientFormer is lim" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.303 + ], + "angle": 0, + "content": "ited by two design choices. Firstly, it uses ineffective token mixing, and secondly, it only employs 3D MetaBlocks in the last stage due to quadratic complexity of MHSA. This likely leads to inconsistent and insufficient contextual representation. To address these limitations, our SwiftFormer improves the token mixing by using a simple yet effective Conv. Encoder. Further, we introduce efficient additive attention module that can be incorporated in all stages (Sec. 3.2). This leads to more consistent learning of local-global representations. It is worth mentioning that EfficientFormer employs a latency-driven slimming method to obtain optimal configurations for its model variants, which leads to maximizing the speed. In contrast, our SwiftFormer models are built without using any neural architecture search." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.893, + 0.635 + ], + "angle": 0, + "content": "Fig. 3 shows an overview of our proposed architecture. The main components are: (i) Effective Conv. Encoder, and (ii) SwiftFormer Encoder. In contrast to other hybrid designs, the proposed architecture is consistent and has Conv. Encoders followed by SwiftFormer Encoder in all stages. Our architecture extracts hierarchical features at four different scales across four stages. At the beginning of the network, the input image of size \\( H \\times W \\times 3 \\) is fed through Patch Embedding layer, implemented with two \\( 3 \\times 3 \\) convolutions with a stride of 2, resulting \\( \\frac{H}{4} \\times \\frac{W}{4} \\times C_1 \\) feature maps. Then, the output feature maps are fed into the first stage, which begins with Conv. Encoder to extract spatial features, followed by SwiftFormer to learn the local-global information. Between two consecutive stages, there is a downsampling layer to increase the channel dimension and reduce the token length. Next, the resulting feature maps are subsequently fed into the second, third, and fourth stages of the architecture, producing \\( \\frac{H}{8} \\times \\frac{W}{8} \\times C_2 \\), \\( \\frac{H}{16} \\times \\frac{W}{16} \\times C_3 \\), and \\( \\frac{H}{32} \\times \\frac{W}{32} \\times C_4 \\) dimensional feature maps, respectively. Hence, each stage learns local-global features at different scales of the input image, which allows the network to have enriched representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.894, + 0.832 + ], + "angle": 0, + "content": "Effective Conv. Encoder: The baseline EfficientFormer [20] employs \\(3 \\times 3\\) average pooling layers as a local token mixer, similar to PoolFormer [53]. Although PoolFormer layers are known for their fast inference speed, replacing them with depth-wise convolutions does not increase the latency. Further, it improves the performance without increasing the parameters and latency. Specifically, the features maps \\(\\mathcal{X}_i\\) are fed into \\(3 \\times 3\\) depth-wise convolution (DWConv) followed by Batch Normalization (BN). Then, the resulting features are fed into two point-wise convolutions \\((\\mathrm{Conv}_1)\\) alongside GeLU activation. Finally, we incorporate a skip connection to enable information to flow across the network. The Conv. Encoder is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.842, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {X}} _ {i} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {C o n v} _ {1, G} \\left(\\operatorname {D W C o n v} _ {B N} \\left(\\mathcal {X} _ {i}\\right)\\right)\\right) + \\mathcal {X} _ {i}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_i\\) refers to the input features, \\(\\mathrm{Conv}_{1,G}\\) refers to point-wise convolution followed by GeLU, DWConv\\(_{BN}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17429" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "refers to depth-wise convolution followed by Batch Normalization, and \\(\\hat{\\mathcal{X}}_i\\) refers to the output feature maps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.318 + ], + "angle": 0, + "content": "SwiftFormer Encoder: This module is carefully designed to efficiently encode enriched local-global representation in each stage. As shown in Fig. 3, the initial block of the SwiftFormer Encoder is composed of \\(3 \\times 3\\) depth-wise convolution followed by point-wise convolution, which enables the module to learn spatial information and encode local representation. Then, the resulting feature maps are fed into the efficient additive attention block, which aims to learn contextual information at each scale of the input size. Finally, the output feature maps are fed into a Linear block, which composes of two \\(1 \\times 1\\) point-wise convolution layers, Batch Normalization, and GeLU activation to generate non-linear features. The SwiftFormer Encoder is described as:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.328, + 0.382, + 0.347 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {X}} _ {i} = \\operatorname {C o n v} _ {1} (\\operatorname {D W C o n v} _ {B N} (\\hat {\\mathcal {X}} _ {i})),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.351, + 0.469, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {X}} _ {i} = \\mathrm {Q K} (\\hat {\\mathcal {X}} _ {i}) + \\hat {\\mathcal {X}} _ {i}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.372, + 0.411, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {X}} _ {i + 1} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {C o n v} _ {B N, 1, G} \\left(\\hat {\\mathcal {X}} _ {i}\\right)\\right) + \\hat {\\mathcal {X}} _ {i}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.402, + 0.468, + 0.448 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv}_{BN,1,G}\\) denotes to Batch Normalization, followed by, \\(1\\times 1\\) Conv layer, followed by GeLU, and QK denotes the efficient additive attention (explained in Sec. 6)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.462, + 0.209, + 0.479 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.468, + 0.549 + ], + "angle": 0, + "content": "We evaluate our SwiftFormer models across four downstream tasks: classification on ImageNet-1K [7], object detection and instance segmentation on MS-COCO 2017 [21], and semantic segmentation on ADE20K [58]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.295, + 0.575 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.468, + 0.794 + ], + "angle": 0, + "content": "ImageNet-1K [7]: All of our models are trained from scratch on ImageNet-1K dataset for 300 epochs with AdamW optimizer [26] and cosine learning rate scheduler with an initial learning rate of \\(1\\mathrm{e}^{-3}\\). We use a linear warm-up for 5 epochs. We use an image resolution of \\(224\\times 224\\) for both training and testing. Following the training recipe of [20], we use the same teacher model for distillation [36]. The experiments are conducted with PyTorch 1.12 [35] using 8 NVIDIA A100 GPUs. The latency is measured using iPhone 14 (iOS 16), and the throughput is measured using A100 40 GB GPU. For latency measurements, we compile the models using CoreML library [6] and perform inference with a batch size of 1. For the throughput on A100, the inference is performed using a batch size of 128." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "MS-COCO 2017 [21]: We use our ImageNet pre-trained models as the backbones in Mask-RCNN framework for object detection and instance segmentation on MS-COCO 2017 dataset. The dataset contains 118K training and 5K validation images. Following [20], we finetune our models for 12 epochs with an image size of \\(1333 \\times 800\\) and batch size of 32 using AdamW optimizer. We use learning rate of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "\\(2e^{-4}\\) and report the performance for detection and instance segmentation in terms of mean average precision (mAP)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.273 + ], + "angle": 0, + "content": "ADE20K [58]: The dataset comprises 20K training and 2K validation images and contains 150 class categories for scene parsing. Similar to [20] we use our ImageNet pre-trained models to extract image features and semantic FPN [18] as a decoder for segmentation. The model is trained with an image size of \\(512 \\times 512\\) for 40K iterations with a batch size of 32 using AdamW optimizer. We use poly learning rate scheduling with an initial learning rate of \\(2\\mathrm{e}^{-4}\\). We report the semantic segmentation performance in terms of mean intersection over union (mIoU)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.284, + 0.7, + 0.3 + ], + "angle": 0, + "content": "4.2. Baseline Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Table 1 illustrates the impact of integrating our proposed contributions into the baseline EfficientFormer-L1 [20] model in terms of ImageNet-1K top-1 accuracy and inference speed. The first row shows the results of the baseline model, which only includes the self-attention based transformer block in the final stage of the network and achieves a top-1 accuracy of \\(79.2\\%\\) with a latency of \\(1.1~\\mathrm{ms}\\) on an iPhone 14 mobile device. The second row replaces the pool mixers in the baseline model with our proposed Conv. Encoder, resulting in an improvement in performance to \\(79.9\\%\\) while maintaining the same latency. In the third row, we replace the transformer block in the baseline with our proposed SwiftFormer Encoder built on the efficient additive attention. Although the performance drops by \\(0.2\\%\\), the inference speed improves by \\(0.1~\\mathrm{ms}\\), and the model has linear complexity with the number of tokens. This enables us to integrate the SwiftFormer Encoder into all stages and achieve better performance while maintaining the same inference speed as of baseline (first versus last row)." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.607, + 0.89, + 0.673 + ], + "angle": 0, + "content": "
MethodLatency (ms)Top-1 (%)
EfficientFormer-L1 (Baseline)1.179.2
+ Replace Pool Mixer by effective Conv. Encoder1.179.9
+ Replace Transformer block by SwiftFormer Encoder1.079.7
+ Incorporate SwiftFormer block across all stages1.180.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.683, + 0.891, + 0.73 + ], + "angle": 0, + "content": "Table 1: Baseline comparison between our SwiftFormer-L1 and EfficientFormer-L1 [20] on the ImageNet-1K dataset. The latency is measured on iPhone14 Neural Engine." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.692, + 0.772 + ], + "angle": 0, + "content": "4.3. Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 2 presents a comparison of our proposed SwiftFormer models (XS, S, L1, and L3) with previous state-of-the-art ConvNets, transformer-based, and hybrid models. We show that our models set new state-of-the-art results, and outperform the recently introduced EfficientFormer [20] and MobileViT-v2 [31] in all model variants. This comprehensive evaluation shows the advantage of our proposed models in terms of both accuracy and latency on mobile devices." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17430" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.88, + 0.42 + ], + "angle": 0, + "content": "
ModelTypeLatency (ms) ↓Throughput (A100) ↑Params(M) ↓GMACs ↓Neural SearchTop-1(%) ↑
MobileNet-v2×1.0 [39]ConvNet0.898893.50.3✘71.8
MobileNet-v3-Large×0.75 [15]ConvNet0.8109344.00.2✓73.3
EdgeViT-XXS [34]Hybrid1.759654.10.6✘74.4
MobileViT-XS [30]Hybrid1.537072.30.7✘74.8
SwiftFormer-XSHybrid0.760343.50.6✘75.7
MobileNet-v2×1.4 [39]ConvNet0.974476.10.6✘74.7
MobileNet-v3-Large [15]ConvNet0.9103515.40.3✓75.1
EfficientNet-b0 [42]ConvNet1.385375.30.4✓77.1
DeiT-T [44]Transformer3.858605.73.8✘72.2
EdgeViT-XS [34]Hybrid2.748126.71.1✘77.5
MobileViT-v2×1.0 [31]Hybrid1.732014.91.8✘78.1
SwiftFormer-SHybrid0.850516.11.0✘78.5
MobileFormer-508M [3]Hybrid3.0444314.00.5✘79.3
PoolFormer-S12 [53]Pool1.2322712.01.8✘77.2
EdgeViT-S [34]Hybrid3.5425613.11.9✘81.0
EfficientFormer-L1 [20]Hybrid1.1504612.31.3✓79.2
MobileViT-v2×1.5 [31]Hybrid3.4235610.64.0✘80.4
SwiftFormer-L1Hybrid1.1446912.11.6✘80.9
ResNet-50 [13]ConvNet1.9483525.54.1✘78.5
PoolFormer-S36 [53]Pool2.8111431.05.0✘81.4
ConvNeXt-T [25]ConvNet2.5323528.64.5✘82.1
DeiT-S [44]Transformer9.9299022.54.5✘81.8
Swin-T [25]TransformerNA263528.34.5✘81.3
MobileViT-v2×2.0 [31]Hybrid5.0190618.57.5✘81.2
EfficientFormer-L3 [20]Hybrid2.0269131.33.9✓82.4
SwiftFormer-L3Hybrid1.9289028.54.0✘83.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.43, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Table 2: Comparison of our proposed SwiftFormer with the state-of-the-art counterpart models on ImageNet-1K. The latency is measured on iPhone 14 Neural Engine (iOS 16) and the throughput is measured on Nvidia A100 GPU. Our models run faster than MobileNets, Hybrid, and Transformer models, with a better trade-off between accuracy and model complexity. The error for the latency measurement is less than \\(\\pm 0.1\\) ms. The input size remains consistent at \\(224\\times 224\\) for all models, except for the MobileViT [30] and MobileViT-v2 [31] variants, which employ a larger dimension of \\(256\\times 256\\). Our results are shown in bold for all model variants." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.473, + 0.866 + ], + "angle": 0, + "content": "Comparison with ConvNets: Our SwiftFormer models surpass the widely used lightweight CNNs counterparts significantly in terms of top-1 accuracy, while running faster than the highly optimized MobileNet-v2 and MobileNet-v3 on an iPhone 14 mobile device. Specifically, our SwiftFormer-XS runs \\(0.1\\mathrm{ms}\\) faster than MobileNet-v2 \\(\\times 1.0\\) and MobileNet-v3-Large \\(\\times 0.75\\) and achieve better top-1 accuracy with a margin of \\(3.9\\%\\) and \\(2.4\\%\\) respectively. Our SwiftFormer-S runs faster than EfficientNet-b0 [42] by \\(1.6\\times\\) and achieves \\(1.4\\%\\) higher top-1 accuracy. Further, our SwiftFormer-L3 achieves \\(4.5\\%\\) and \\(0.9\\%\\) gain in top-1 accuracy over ResNet-50 and ConvNeXt-T, respectively, while running at the same latency as ResNet-50 and \\(1.3\\times\\) faster than ConvNeXt-T. This demonstrates that our SwiftFormer models, powered by our proposed efficient additive attention, run faster than the lightweight CNN models on mobile devices and achieve superior performance. Recent device-level optimizations for CNN-based models, such as dedicated hardware implementations for convolutions with batch normalization and non-linearity, likely contribute to the high throughput of fully CNN-based models on A100." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Comparison with transformer models: Although transformer models usually outperform CNN-based models in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.701 + ], + "angle": 0, + "content": "terms of accuracy, they tend to suffer from high latency when running on resource-constrained mobile devices. For instance, DeiT-S, which has a similar model size to ResNet-50 and achieves higher top-1 accuracy by \\(3.3\\%\\), but ResNet-50 runs approximately \\(5.2\\times\\) faster on an iPhone 14 mobile device. In contrast, our SwiftFormer-L3 model achieves \\(1.2\\%\\) higher accuracy than DeiT-S, while running at the same speed as ResNet-50. Further our SwiftFormer-S model runs approximately \\(4.7\\times\\) faster than DeiT-T on an iPhone 14 mobile device and has \\(6.3\\%\\) better accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Comparison with hybrid models: Although most existing hybrid approaches achieve higher accuracy compared to their lightweight CNN counterparts, they still underperform the fully CNN-based models in terms of latency due to the quadratic complexity of multi-head self-attention. For example, EdgeViT-XXS runs at approximately \\(2 \\times\\) slower compared to MobileNet-v3-Large \\(\\times 0.75\\). On the other hand, our SwiftFormer-XS has better latency as compared to lightweight CNNs and approximately is \\(2 \\times\\) faster than EdgeViT-XXS and MobileViT-XS, with an overall \\(1.3\\%\\) and \\(0.9\\%\\) higher top-1 accuracy respectively. Further, our SwiftFormer-L1 model is \\(3 \\times\\) faster than the state-of-the-art MobileViT-v2 \\(\\times 1.5\\) with \\(0.5\\%\\) better top-1 ac" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17431" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.171, + 0.089, + 0.803, + 0.264 + ], + "angle": 0, + "content": "
BackboneDetection & Instance SegmentationSemantic mIoU(%)
\\(AP^{box}\\)\\(AP^{box}_{50}\\)\\(AP^{box}_{75}\\)\\(AP^{mask}\\)\\(AP^{mask}_{50}\\)\\(AP^{mask}_{75}\\)
ResNet18 [13]34.054.036.731.251.032.732.9
PoolFormer-S12 [53]37.359.040.134.655.836.937.2
EfficientFormer-L1 [20]37.960.341.035.457.337.338.9
SwiftFormer-L141.263.244.838.160.240.741.4
ResNet50 [13]38.058.641.434.455.136.736.7
PVT-S [48]40.462.943.837.860.140.339.9
PoolFormer-S24 [53]40.162.243.437.059.139.640.3
Swin-T [24]42.264.446.239.164.642.041.5
EfficientFormer-L3 [20]41.463.944.738.161.040.443.5
SwiftFormer-L342.764.446.739.161.741.843.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.274, + 0.895, + 0.307 + ], + "angle": 0, + "content": "Table 3: Results using SwiftFormer as a backbone on dense prediction tasks: Object detection and instance segmentation on COCO, whereas semantic segmentation on ADE20K. Our approach outperforms EfficientFormer on all three tasks." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.319, + 0.26, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.32, + 0.353, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.32, + 0.531, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.32, + 0.712, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.32, + 0.888, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.429, + 0.893, + 0.461 + ], + "angle": 0, + "content": "Figure 4: Results on COCO. Examples for object detection and instance segmentation on the COCO 2017 validation set. Our SwiftFormer-L1 model can accurately detect and segment instances in images." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.473, + 0.214, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.553, + 0.214, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.473, + 0.35, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.553, + 0.35, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.473, + 0.483, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.553, + 0.483, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.473, + 0.608, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.553, + 0.608, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.473, + 0.754, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.553, + 0.754, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.473, + 0.889, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.553, + 0.889, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.645, + 0.895, + 0.69 + ], + "angle": 0, + "content": "Figure 5: Qualitative results on ADE20K. The qualitative examples for semantic segmentation on the ADE20K validation set. Top: Ground truth masks. Bottom: The semantic segmentation results. Our model can accurately segment various indoor and outdoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.469, + 0.763 + ], + "angle": 0, + "content": "curacy. Our SwiftFormer-L3 model achieves \\(83.0\\%\\) top-1 accuracy and runs at \\(1.9~\\mathrm{ms}\\), which is \\(2.6\\times\\) faster than MobileViT-v2 \\(\\times 2.0\\) with an absolute \\(1.8\\%\\) accuracy gain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Knowledge distillation (KD): It is worth mentioning that we use the same teacher model of DeiT [44] and EfficientFormer [20] in Table 2. We perform a baseline comparison to show the efficiency of our method without using KD. EfficientFormer-L1 achieves \\(77.3\\%\\) accuracy w/o KD, compared to \\(79.2\\%\\) with KD. Our SwiftFormer-L1 achieves \\(79.8\\%\\) w/o KD, compared to \\(80.9\\%\\) with KD. This shows that our approach achieves a higher absolute gain of \\(2.5\\%\\) over EfficientFormer-L1, when not using KD." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.718, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Our SwiftFormer-XS and SwiftFormer-S w/o KD achieve \\(74.1\\%\\) and \\(77.0\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.878, + 0.774 + ], + "angle": 0, + "content": "4.4. Object Detection and Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Table 3 compares the object detection and instance segmentation results of Mask-RCNN with different lightweight backbones. Our SwiftFormer-L1 backbone achieves 41.2 AP box, surpassing the lightweight ResNet18 and PoolFormer-S12 backbones by 7.2 and 3.9 points respectively. Further, it performs better than the previous state-of-the-art EfficientFormer-L1 backbone by 3.3 AP box. For instance segmentation, our method achieves 38.1 AP mask" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17432" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "score which is 2.7 points better than the previous state-of-the-art. Similar trend is observed for SwinfFormer-L3 backbone, which surpasses the previous state-of-the-art EfficientFormer-L3 backbone by 1.3 points and 1.0 points in AP box and mask respectively. The improvement in the downstream detection and instance segmentation tasks illustrates the effectiveness of our SwiftFormer backbone models for the dense prediction tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.223, + 0.295, + 0.24 + ], + "angle": 0, + "content": "4.5. Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.47, + 0.368 + ], + "angle": 0, + "content": "Table 3 shows the semantic segmentation results of SwiftFormer backbone-based models as compared to previously proposed backbones. We achieve \\(41.4\\%\\) mean intersection over union score using SwiftFormer-L1, surpassing ResNet18 by \\(8.5\\%\\), PoolFormer-S12 by \\(4.2\\%\\), and the state-of-the-art EfficientFormer-L1 by \\(2.5\\%\\). Similarly, our SwiftFormer-L3 backbone-based segmentation models achieve \\(43.9\\mathrm{mIoU}\\), surpassing all previous methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.379, + 0.262, + 0.394 + ], + "angle": 0, + "content": "4.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.471, + 0.584 + ], + "angle": 0, + "content": "Fig. 4 shows the qualitative results of our detection and instance segmentation model on COCO dataset. We also show in Fig. 5 qualitative examples for semantic segmentation results on the ADE20K dataset. Additionally, we demonstrate in Fig. 6 a baseline comparison with the baseline EfficientFormer-L1. The baseline misclassifies a bird as a sheep in the first example and as a bear in the second example, while misclassifies kites as birds in the third example. Our SwiftFormer-L1 accurately detects and segments objects in these examples, which demonstrates the ability of SwiftFormer to accurately identify and segment objects within a diverse range of indoor and outdoor settings." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.598, + 0.467, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.751, + 0.47, + 0.798 + ], + "angle": 0, + "content": "Figure 6: Qualitative comparison between EfficientFormer-L1 (top) and our SwiftFormer-L1 (bottom) on images from COCO val set for detection and instance segmentation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.226, + 0.848 + ], + "angle": 0, + "content": "5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To show the effectiveness of the proposed efficient additive attention, we conduct a comparison with other attention variants over SwiftFormer-L1 in Table. 4. It is notable that" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.161 + ], + "angle": 0, + "content": "
Attention MethodLatency (ms)↓Throughput↑Top-1 (%)↑
Self-attention (Heads=2)8.5331680.6
Self-attention (Heads=4)18.0303581.1
Transpose self-attention (Heads=2)2.0367180.2
Separable self-attention1.4394080.6
Efficient Additive Attention (Ours)1.1446980.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.171, + 0.892, + 0.202 + ], + "angle": 0, + "content": "Table 4: Quantitative analysis of Fig. 2. We replace our attention module with these methods in the same setting." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.217, + 0.892, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.379, + 0.893, + 0.44 + ], + "angle": 0, + "content": "Figure 7: Visualization of the learned global query vector of the efficient additive attention from SwiftFormer-L1 model at different stages. Both the first and last stages show that the global context is captured effectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.619 + ], + "angle": 0, + "content": "the proposed efficient additive attention achieves the best trade-off between latency and top-1 accuracy. To demonstrate further insights on the interpretability of the proposed efficient additive attention, we visualize the learned global attention query vector \\(\\alpha\\) of Eq. 4 of the first and last stages of the network. The global query vector of the efficient additive attention aligns with the semantics of the image and hence acts as an informative source in the additive attention mechanism. Also, it localizes the features of the objects accurately with fine details in the first stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.62, + 0.648 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Transformers have gained popularity in vision applications due to their effective use of self-attention computation. However, their use in mobile vision applications is challenging due to the quadratic nature of self-attention, which is computationally expensive on mobile devices. In this work, we propose a novel efficient additive attention that replaces the expensive matrix multiplication operations with element-wise operations and eliminates explicit keys-values interaction. Our proposed attention is linear with respect to the input tokens and used in all stages of the network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although we achieve a superior balance between accuracy and efficiency, we observe that in the case of dense small object detection, our approach is less accurate likely due to not using positional encoding or attention biases. Our future work is to improve the performance of efficient additive attention in dense prediction tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17433" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Adrian Bulat and Georgios Tzimiropoulos. Bit-mixer: Mixed-precision networks with runtime bit-width selection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[5] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[6] CoreMLTools. Use coremltools to convert models from third-party libraries to core ml., 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.605 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[9] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[10] Mohsen Fayyaz, Soroush Abbasi Kouhpayegani, Farnoush Rezaei Jafari, Eric Sommerlade, Hamid Reza Vaezi Joze, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers. 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[11] TSong Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[12] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[14] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[15] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenet-v3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[16] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.28, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[17] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.351, + 0.892, + 0.394 + ], + "angle": 0, + "content": "[18] A. Kirillov, R. Girshick, K. He, and P. Dollar. Panoptic feature pyramid networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.396, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[19] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In International Conference on Learning Representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.44, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[20] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.54 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.542, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[22] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco De Nadai. Efficient training of visual transformers with small datasets. In Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.6, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[23] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.672, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.743, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[25] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In European Conference on Computer Vision, 2018." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17434" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[29] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: Efficiently amalgamated cnn-transformer architecture for mobile vision applications. In CADL2022, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[30] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.329 + ], + "angle": 0, + "content": "[31] Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. Transactions on Machine Learning Research, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[32] Sachin Mehta, Mohammad Rastegari, Linda Shapiro, and Hannaneh Hajishirzi. Espnetv2: A light-weight, power efficient, and general purpose convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.469 + ], + "angle": 0, + "content": "[33] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[34] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martinez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.649 + ], + "angle": 0, + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[36] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[38] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[39] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenet-v2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[40] Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Unetr++: Delving into efficient and accurate 3d medical image segmentation. arXiv:2212.04497, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[41] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[42] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[43] Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, et al. Mlp-mixer: An all-mlp architecture for vision. In Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[44] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[45] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[46] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[47] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[48] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[50] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. In Advances in Neural Information Processing Systems, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[51] Chuhan Wu, Fangzhao Wu, Tao Qi, and Yongfeng Huang. Fastformer: Additive attention can be all you need. arXiv preprint arXiv:2108.09084, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[52] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[53] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Edgeformer: Improving light-weight convnets by learning from vision transformers. arXiv preprint arXiv:2203.03952, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17435" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[55] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Parc-net: Position aware circular convolution with merits from convnets and transformer. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[56] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[57] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[58] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17436" + } + ] +] \ No newline at end of file diff --git a/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/3fb2491f-940c-4884-86b5-1f7be4430698_origin.pdf b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/3fb2491f-940c-4884-86b5-1f7be4430698_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ccb95834ccc029e02c6a45f93b35e3ac0239ae3f --- /dev/null +++ b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/3fb2491f-940c-4884-86b5-1f7be4430698_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8e737a267f8770a0afaef416f8a62d7a1d8ffde85e0d0aca7953d25374fffb0 +size 4378382 diff --git a/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/full.md b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6e00a8bcf10f718e36603c7cbab389fb5b47f3d7 --- /dev/null +++ b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/full.md @@ -0,0 +1,344 @@ +# SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications + +Abdelrahman Shaker $^{1*}$ Muhammad Maaz $^{1}$ Hanoona Rasheed $^{1}$ Salman Khan $^{1}$ Ming-Hsuan Yang $^{2,3,4}$ Fahad Shahbaz Khan $^{1,5}$ $^{1}$ Mohamed bin Zayed University of AI $^{2}$ University of California, Merced + $^{3}$ Yonsei University $^{4}$ Google Research $^{5}$ Linköping University + +# Abstract + +Self-attention has become a defacto choice for capturing global context in various vision applications. However, its quadratic computational complexity with respect to image resolution limits its use in real-time applications, especially for deployment on resource-constrained mobile devices. Although hybrid approaches have been proposed to combine the advantages of convolutions and self-attention for a better speed-accuracy trade-off, the expensive matrix multiplication operations in self-attention remain a bottleneck. In this work, we introduce a novel efficient additive attention mechanism that effectively replaces the quadratic matrix multiplication operations with linear element-wise multiplications. Our design shows that the key-value interaction can be replaced with a linear layer without sacrificing any accuracy. Unlike previous state-of-the-art methods, our efficient formulation of self-attention enables its usage at all stages of the network. Using our proposed efficient additive attention, we build a series of models called "Swift-Former" which achieves state-of-the-art performance in terms of both accuracy and mobile inference speed. Our small variant achieves $78.5\%$ top-1 ImageNet-1K accuracy with only 0.8 ms latency on iPhone 14, which is more accurate and $2\times$ faster compared to MobileViT-v2. Our code and models: https://tinyurl.com/5ft8v46w + +# 1. Introduction + +In recent years, transformer models have shown remarkable success in various vision applications such as classification [8, 44, 24, 23, 9], detection [2, 59, 33, 56, 28], and segmentation [4, 40]. However, deploying these models on resource-constrained mobile devices for real-time applications remains challenging due to their inherently complex nature [20, 29]. Specifically, vision transformers (ViTs) rely on global self-attention, which has a quadratic com + +![](images/e465b2c4ede2ca863f0d496b5efb8565cf4f6b70e777f2e89c77cd8852cf50ca.jpg) +Figure 1: Latency vs Accuracy Comparison. Compared to the recent EfficientFormer-L1 [20], our SwiftFormer-L1 achieves an absolute gain of $1.7\%$ in terms of top-1 accuracy with the same latency and without requiring any neural architecture search. + +plexity with respect to the input image resolution, making it impractical for deployment on low-powered mobile devices [31]. As a result, convolutional neural networks (CNNs) are still the preferred choice for real-time deployment on mobile devices, primarily because the convolution operation is computationally efficient [39, 15]. However, a major limitation of CNNs is their reliance on local connections and stationary weights, which can limit their ability to adapt to variable input resolutions and capture long-range dependencies in the data. Therefore, developing more efficient and flexible models that combine the strengths of both CNNs and transformers is critical, particularly for mobile devices with limited computational resources. + +To achieve this goal, several hybrid approaches have + +been proposed that use lightweight CNN modules in the high-resolution early stages and self-attention in the low-resolution later stages [55, 29, 20]. This approach effectively increases the receptive field of the network and strives to achieve a trade-off between speed and accuracy. Furthermore, different efficient variants of computing self-attention have been proposed to reduce the model complexity. These include computing attention across feature dimensions to implicitly model the global context [29], computing attention within local windows [24], pooling spatial features before applying self-attention [9], and sparsely attending to a fixed number of tokens [34], to name a few. + +Although these approaches effectively reduce network complexity, they still involve inefficient matrix multiplication operations that significantly impact latency on mobile devices. To address this issue, Mehta et al. [31] propose a separable self-attention mechanism that replaces matrix multiplication operations to element-wise multiplications. This is achieved by projecting queries to context scores, followed by element-wise multiplication with keys to calculate context vectors for encoding global context. + +In this work, we propose efficient additive attention, which eliminates the need for expensive matrix multiplication operations in computing self-attention. Additionally, we propose to compute the global context using only the query-key interactions followed by a linear transformation, without requiring explicit key-value interactions. This significantly reduces the computational complexity and enables us to use the proposed attention block in all stages of the network. Our contributions are as follows: + +- We introduce efficient additive attention, a new approach for computing self-attention in vision backbones that eliminates the need for expensive matrix multiplication operations, significantly reducing the computational complexity of the model. +- Unlike previous methods, our proposed efficient attention design can be used at all stages of the network, enabling more effective contextual information capture and achieving superior speed-accuracy trade-off. +- We build a series of efficient generic classification models called "SwiftFormer", which utilize our proposed efficient additive attention. Our small model achieves $78.5\%$ top-1 ImageNet-1K [7] accuracy while running at only $0.8~\mathrm{ms}$ latency on iPhone 14. Moreover, our large model achieves $83.0\%$ accuracy with a latency of only $1.9~\mathrm{ms}$ . Our model achieves state-of-the-art performance, outperforming recent MobileViTv2 [31] and EfficientFormer [20] by obtaining a better trade-off between accuracy and latency (see Fig. 1). + +# 2. Related Work + +Efficient CNNs: Designing efficient CNNs for mobile vision applications has received much attention in recent years. MobileNet architectures [16, 39, 15] propose depthwise separable convolutions as well as efficient inverted residual blocks for improved performance on various vision tasks. Other methods aim to improve the efficiency by leveraging depth-wise dilated convolutions [32], channel shuffling and pointwise group convolutions [57, 27], network pruning [11, 50], low bit-width [1, 17], and neural architecture search [41, 15]. CNN-based methods are well-performing, efficient, and fast to train and run on edge devices, resulting in widespread usage in the industry. However, they are spatially local and lack global interaction between the features, which deeply affects their performance. Efficient Transformers: ViTs [8] have been widely used in numerous vision tasks, and significant advances have been made in terms of data efficiency [22, 44], transformer architecture [30, 20, 3, 54], and token mechanisms [43, 52]. Reducing the number of visual tokens is a major modification in the transformer architecture for efficient deployment. Instead of using a fixed feature representation through the whole architecture, some methods employ a hierarchical design where the resolution is gradually decreased through the stages, including down-sampling techniques [9, 40, 38, 14] and pyramidal structures [47, 49]. Recently, a few methods [37, 10, 34] propose token sparsification techniques to encode only a subset of the most informative tokens. + +Numerous approaches have recently been proposed to reduce the quadratic complexity of self-attention, the computational bottleneck in transformer-based architectures, by computing its approximated variants [34, 29, 31, 46, 19, 5, 45]. EdgeViT [34] uses a global sparse attention module attending only to a few tokens to improve the efficiency, while [48] down-samples the key and value vectors that lead to a better efficiency-accuracy trade-off. GCViT [12] uses a local-global attention mechanism by employing the interactions between a global query and local key and value to capture the contextual information. This approach enables the model to encode both local and global representations efficiently. EdgeNeXt [29] adopts transposed self-attention operation to compute the attention maps across the channel dimension instead of the spatial dimension, followed by token mixing, to have a linear complexity with respect to the number of tokens. Reformer [19] replaces the dot-product attention with a locality-sensitive hashing to group the tokens and reduced the complexity from $O(n^{2})$ to $O(n\log n)$ . However, this design is only efficient on longer sequences, which is typically not the case for ViTs. LinFormer [46] is a low-rank matrix factorization method that approximates the self-attention matrix with a low-rank matrix, reducing the complexity from $O(n^{2})$ to $O(n)$ . Although matrix factorization methods theoretically reduce the complexity of + +self-attention, they use expensive projections for computing attention, which may not reflect the reduction in FLOPs and parameters into actual speed on mobile platforms. + +Although these methods show promise and have reduced the complexity of self-attention theoretically, they are inadequate for reducing the inference speed for mobile deployment. Since the complexity of the multi-headed self-attention (MHSA) is higher in the earlier stage compared to the last stages, EfficientFormer [20] incorporates MHSA in the last stage only to learn contextual information from the high-level features without increasing the inference speed significantly. Recently, MobileViT-v2 [31] proposes separable self-attention that uses element-wise operations instead of the dot-product to compute the attention maps with linear complexity. Different from the existing approaches, we propose a consistent hybrid design with an efficient additive attention mechanism to model the contextual information with linear complexity. Instead of capturing the pairwise interactions between keys, queries, and values using the dot-product, we use element-wise operations with learnable attention weights to model the interactions between query and keys only, leading to better inference speed. + +# 3. Method + +Motivation: To motivate our method, we first distinguish three desirable characteristics to be considered when designing an efficient yet accurate approach for resource constraint mobile devices. + +Efficient Global Context Modeling: As discussed earlier, most existing approaches either employ the standard MHSA or an approximated variant to learn the global context. However, they struggle to operate as fast as MobileNets on resource-constrained devices. This is likely due to the computation-intensive multiplicative operations during attention computation or reliance on advanced reshaping and indexing operations in these approaches. For instance, the recent MobileViT-v2 [31] is $2 \times$ slower than MobileNet-v2 [39]. Instead of using matrix multiplications, we argue that encoding the global context using an efficient additive attention design can reduce the operations with respect to the number of tokens. This is expected to help operate at comparable speed and model size, while achieving superior accuracy compared to MobileNets. + +Rethinking key-value interactions: Other than multiplicative operations during attention computation, additive attention has been recently explored in the NLP domain [51]. However, in the standard form, it performs three-step processing to model query, key, and value interactions. Each step feeds into the subsequent one, thereby requiring sequential processing. Here, we rethink the additive attention for vision tasks by alleviating the need to compute explicit interactions between key-value. We argue that eliminating key-value interactions and replacing them with a sim + +ple linear transformation empirically encodes better contextual representation. Our design encodes only global queries and key interactions to learn the global contextual information, followed by a linear transformation to compute global context-aware attention weights. + +Consistent Hybrid Design: Most existing works employ MHSA or the approximated variant in the last stages, while avoiding its usage in the earlier stages. This is because the computational complexity of MHSA grows quadratically with the length of the tokens, making it impractical to incorporate during the initial stages. This constraint adds to the design complexity and requires a careful selection of stages, where MHSA can be applied. In contrast, our proposed SwiftFormer module has linear complexity with respect to the token length and can be incorporated in all stages to learn consistent global context at each scale. This consistency improves the model performance and makes it more generalizable and scalable for high-resolution images. + +# 3.1. Overview of Attention Modules + +Vision transformer models are built upon the self-attention (see Fig. 2 (a)), which can effectively model the interactions between the input tokens. Specifically, the self-attention has $\mathbf{x}$ as an input, where $\mathbf{x} \in \mathbb{R}^{n \times d}$ , comprising $n$ tokens with $d$ -dimensional embedding vector. The input $\mathbf{x}$ is projected to query $(\mathbf{Q})$ , key $(\mathbf{K})$ , and value $(\mathbf{V})$ using three matrices, $\mathbf{W}_{\mathbf{Q}}, \mathbf{W}_{\mathbf{K}}$ , and $\mathbf{W}_{\mathbf{V}}$ . Each self-attention layer comprises $h$ heads, which allows the model to attend to different views of the input. The self-attention can be described as: + +$$ +\hat {\mathbf {x}} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \cdot \mathbf {K} ^ {\top}}{\sqrt {d}}\right) \cdot \mathbf {V}. \tag {1} +$$ + +The attention scores between each pair of tokens in $\mathbf{Q}$ and $\mathbf{K}$ are computed using the dot-product operation. Next, these scores are normalized followed by Softmax to weigh the interactions between the tokens. Finally, the weighted interactions are multiplied by $\mathbf{V}$ using the dot-product operation to produce the final weighted output. Overall, the complexity of the self-attention is $O(n^{2} \cdot d)$ , where $n$ is the number of tokens and $d$ is the hidden dimension. The computational and memory demands of $\mathbf{Q} \cdot \mathbf{K}^{\top}$ increase quadratically as the number of tokens grows, leading to slow inference speed and high memory usage, making it impractical to run in real-time for long sequences. + +To alleviate this issue, [29] proposes the transpose self-attention (see Fig. 2 (b)) to reduce the complexity from quadratic to linear with respect to the number of tokens. Here, the dot-product operation is applied across the channel dimension instead of the spatial dimension. This allows the model to learn feature maps with implicit contextual representation. The attention can be described as: + +$$ +\hat {\mathbf {x}} = \mathbf {V} \cdot \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} ^ {\top} \cdot \mathbf {K}}{\sqrt {d}}\right). \tag {2} +$$ + +![](images/f4dadbfd779cf46e7485bcf975029c11607f9bcbb982d8d877b7d4929d7444b7.jpg) + +![](images/ce019951c2435c475affca2889a026a83ec51408895987e0368a1cf3bf5aaaf0.jpg) +Figure 2: Comparison with different self-attention modules. (a) is a typical self-attention used in ViTs [8]. (b) is the transpose self-attention used in EdgeNeXt [29], where the self-attention operation is applied across channel feature dimensions $(\mathrm{d} \times \mathrm{d})$ instead of the spatial dimension $(\mathrm{n} \times \mathrm{n})$ . (c) is the separable self-attention of MobileViT-v2 [31], it uses element-wise operations to compute the context vector from the interactions of $\mathbf{Q}$ and $\mathbf{K}$ matrices. Then, the context vector is multiplied by $\mathbf{V}$ matrix to produce the final output. (d) Our proposed efficient additive self-attention. Here, the query matrix is multiplied by learnable weights and pooled to produce global queries. Then, the matrix $\mathbf{K}$ is element-wise multiplied by the broadcasted global queries, resulting the global context representation. + +![](images/f73e2b92d00512d0387eee4f53d0739fae7e8873798e735034a3b05707448349.jpg) + +![](images/cbe2ad1a39b283ffae4a407beae8ce0dd6c57e10a8120f259dfdb3aa10d0adfe.jpg) +Dot product + +![](images/9f7fad8cfecfe1437fd5f3906b308ad92ee323de6f1981a7292d4e13a27fa0d3.jpg) + +![](images/64098f4a824891d3903e14023254a3e540ca1176cb161bb76e4af57ab5eb9d3e.jpg) +Broadcasted element-wise multiplication + +![](images/1f2df7d2f6545780db8f993cde62f0d32290584c1462b44568b79e49b0f5508b.jpg) + +The transpose self-attention has a computational complexity of $O(n \cdot d^2)$ . While this complexity scales linearly with the number of tokens $n$ , it remains quadratic with respect to the feature dimension $d$ . Further, the dot-product operation is still utilized between the query and key matrices. + +The separable self-attention mechanism (see Fig. 2 (c)) aims to address the bottleneck of the standard self-attention. Here, the interactions between the queries $(\mathbf{Q})$ , keys $(\mathbf{K})$ , and values $(\mathbf{V})$ are encoded using element-wise operations. First, the query matrix $\mathbf{Q}$ is projected to produce a vector $\mathbf{q}$ of dimensions $n \times 1$ , and then fed into Softmax to generate the context scores, which captures the importance of each query element. Then, the context scores are multiplied by the key matrix $\mathbf{K}$ and pooled to compute a context vector, which encodes the contextual information. Finally, the context vector is multiplied element-wise with the value matrix $\mathbf{V}$ to propagate the contextual information and produce the final output $\hat{\mathbf{x}}$ . It can be summarized as: + +$$ +\hat {\mathbf {x}} = \mathbf {V} * \sum \mathbf {K} * \operatorname {S o f t m a x} (\mathbf {q}). \tag {3} +$$ + +Here, $*$ denotes the element-wise multiplication operation. + +# 3.2. Efficient Additive Attention + +The typical additive attention mechanism in NLP captures the global context by utilizing pairwise interactions between the tokens via element-wise multiplications instead of using dot-product operation. It encodes the relevance scores for the contextual information of the input sequence based on + +the interactions of the three attention components (Q, K, V). In contrast, we show that key-value interactions can be removed without sacrificing the performance and only focusing on effectively encoding query-key interactions by incorporating a linear projection layer is sufficient to learn the relationship between the tokens (see Fig. 2 (d)). This approach, named efficient additive attention, has a faster inference speed and produces more robust contextual representations as demonstrated by our performance on image classification, object detection, and segmentation tasks (Sec. 4). Specifically, the input embedding matrix $\mathbf{x}$ is transformed into query (Q) and key (K) using two matrices $\mathbf{W_q}$ , $\mathbf{W_k}$ , where Q, $\mathbf{K} \in \mathbb{R}^{n \times d}$ , $\mathbf{W_q}$ , $\mathbf{W_k} \in \mathbb{R}^{d \times d}$ , $n$ is the token length and $d$ is the dimensions of the embedding vector. Next, the query matrix Q is multiplied by learnable parameter vector $\mathbf{w}_a \in \mathbb{R}^d$ to learn the attention weights of the query, producing global attention query vector $\alpha \in \mathbb{R}^n$ as: + +$$ +\alpha = \mathbf {Q} \cdot \mathbf {w} _ {a} / \sqrt {d} \tag {4} +$$ + +Then, the query matrix is pooled based on the learned attention weights, resulting in a single global query vector $\mathbf{q} \in \mathbb{R}^d$ as follows: + +$$ +\mathbf {q} = \sum_ {i = 1} ^ {n} \alpha_ {i} * \mathbf {Q} _ {\mathbf {i}}. \tag {5} +$$ + +Next, the interactions between the global query vector $\mathbf{q} \in \mathbb{R}^d$ and key matrix $\mathbf{K} \in \mathbb{R}^{n \times d}$ are encoded using + +![](images/6ef999bfc9198042cd8cfcd4d703bf261ddae56ca0f7c80b9cc8830ae68e2988.jpg) +Figure 3: Top Row: Overview of our proposed architecture. The input image is fed into the patch embedding layer, followed by hierarchical stages at four different scales $\{\frac{1}{4},\frac{1}{8},\frac{1}{16},\frac{1}{32}\}$ . Each stage is consistent and compose of Conv. Encoder blocks followed by SwiftFormer Encoder. Between two consecutive stages, we incorporate downsampling layer to reduce the spatial size by a factor of two and increase the feature dimensions. Bottom Row: We show the design of the Conv. Eencoder (left) and the SwiftFormer Encoder (right). The Conv. Encoder is designed to learn effective local representations and consists of $3\times 3$ depth-wise convolutions followed by two point-wise convolutions for channel mixing. The SwiftFormer Encoder aims to learn enriched local-global representations. It begins with local convolutional layers to extract local features, followed by the efficient additive attention module (see Fig. 2 (d)) and linear layers. + +the element-wise product to form global context $(\mathbb{R}^{n\times d})$ . This matrix shares a similarity with the attention matrix in MHSA and captures information from every token and has the flexibility to learn the correlation in the input sequence. However, it is comparatively inexpensive to compute compared to MHSA and has linear complexity with the token length. Inspired by the transformer architecture, we employ a linear transformation layer to query-key interactions to learn the hidden representation of the tokens. The output of the efficient additive attention $\hat{\mathbf{x}}$ can be described as: + +$$ +\hat {\mathbf {x}} = \hat {\mathbf {Q}} + \mathbf {T} (\mathbf {K} * \mathbf {q}). \tag {6} +$$ + +where $\hat{\mathbf{Q}}$ denotes to the normalized query matrix, $\mathbf{T}$ denotes to the linear transformation. + +# 3.3. SwiftFormer Architecture + +Our SwiftFormer is based on the recently introduced EfficientFormer [20]. The main idea of EfficientFormer is to introduce 4D MetaBlocks based on PoolFormer [53] to learn local representations efficiently, while using 3D MetaBlocks based on self-attention to encode global context. However, the performance of EfficientFormer is lim + +ited by two design choices. Firstly, it uses ineffective token mixing, and secondly, it only employs 3D MetaBlocks in the last stage due to quadratic complexity of MHSA. This likely leads to inconsistent and insufficient contextual representation. To address these limitations, our SwiftFormer improves the token mixing by using a simple yet effective Conv. Encoder. Further, we introduce efficient additive attention module that can be incorporated in all stages (Sec. 3.2). This leads to more consistent learning of local-global representations. It is worth mentioning that EfficientFormer employs a latency-driven slimming method to obtain optimal configurations for its model variants, which leads to maximizing the speed. In contrast, our SwiftFormer models are built without using any neural architecture search. + +Fig. 3 shows an overview of our proposed architecture. The main components are: (i) Effective Conv. Encoder, and (ii) SwiftFormer Encoder. In contrast to other hybrid designs, the proposed architecture is consistent and has Conv. Encoders followed by SwiftFormer Encoder in all stages. Our architecture extracts hierarchical features at four different scales across four stages. At the beginning of the network, the input image of size $H \times W \times 3$ is fed through Patch Embedding layer, implemented with two $3 \times 3$ convolutions with a stride of 2, resulting $\frac{H}{4} \times \frac{W}{4} \times C_1$ feature maps. Then, the output feature maps are fed into the first stage, which begins with Conv. Encoder to extract spatial features, followed by SwiftFormer to learn the local-global information. Between two consecutive stages, there is a downsampling layer to increase the channel dimension and reduce the token length. Next, the resulting feature maps are subsequently fed into the second, third, and fourth stages of the architecture, producing $\frac{H}{8} \times \frac{W}{8} \times C_2$ , $\frac{H}{16} \times \frac{W}{16} \times C_3$ , and $\frac{H}{32} \times \frac{W}{32} \times C_4$ dimensional feature maps, respectively. Hence, each stage learns local-global features at different scales of the input image, which allows the network to have enriched representation. + +Effective Conv. Encoder: The baseline EfficientFormer [20] employs $3 \times 3$ average pooling layers as a local token mixer, similar to PoolFormer [53]. Although PoolFormer layers are known for their fast inference speed, replacing them with depth-wise convolutions does not increase the latency. Further, it improves the performance without increasing the parameters and latency. Specifically, the features maps $\mathcal{X}_i$ are fed into $3 \times 3$ depth-wise convolution (DWConv) followed by Batch Normalization (BN). Then, the resulting features are fed into two point-wise convolutions $(\mathrm{Conv}_1)$ alongside GeLU activation. Finally, we incorporate a skip connection to enable information to flow across the network. The Conv. Encoder is defined as: + +$$ +\hat {\mathcal {X}} _ {i} = \operatorname {C o n v} _ {1} \left(\operatorname {C o n v} _ {1, G} \left(\operatorname {D W C o n v} _ {B N} \left(\mathcal {X} _ {i}\right)\right)\right) + \mathcal {X} _ {i}. \tag {7} +$$ + +where $\mathcal{X}_i$ refers to the input features, $\mathrm{Conv}_{1,G}$ refers to point-wise convolution followed by GeLU, DWConv $_{BN}$ + +refers to depth-wise convolution followed by Batch Normalization, and $\hat{\mathcal{X}}_i$ refers to the output feature maps. + +SwiftFormer Encoder: This module is carefully designed to efficiently encode enriched local-global representation in each stage. As shown in Fig. 3, the initial block of the SwiftFormer Encoder is composed of $3 \times 3$ depth-wise convolution followed by point-wise convolution, which enables the module to learn spatial information and encode local representation. Then, the resulting feature maps are fed into the efficient additive attention block, which aims to learn contextual information at each scale of the input size. Finally, the output feature maps are fed into a Linear block, which composes of two $1 \times 1$ point-wise convolution layers, Batch Normalization, and GeLU activation to generate non-linear features. The SwiftFormer Encoder is described as: + +$$ +\hat {\mathcal {X}} _ {i} = \operatorname {C o n v} _ {1} (\operatorname {D W C o n v} _ {B N} (\hat {\mathcal {X}} _ {i})), +$$ + +$$ +\hat {\mathcal {X}} _ {i} = \mathrm {Q K} (\hat {\mathcal {X}} _ {i}) + \hat {\mathcal {X}} _ {i}, \tag {8} +$$ + +$$ +\hat {\mathcal {X}} _ {i + 1} = \operatorname {C o n v} _ {1} \left(\operatorname {C o n v} _ {B N, 1, G} \left(\hat {\mathcal {X}} _ {i}\right)\right) + \hat {\mathcal {X}} _ {i}. +$$ + +where $\mathrm{Conv}_{BN,1,G}$ denotes to Batch Normalization, followed by, $1\times 1$ Conv layer, followed by GeLU, and QK denotes the efficient additive attention (explained in Sec. 6). + +# 4. Experiments + +We evaluate our SwiftFormer models across four downstream tasks: classification on ImageNet-1K [7], object detection and instance segmentation on MS-COCO 2017 [21], and semantic segmentation on ADE20K [58]. + +# 4.1. Implementation Details + +ImageNet-1K [7]: All of our models are trained from scratch on ImageNet-1K dataset for 300 epochs with AdamW optimizer [26] and cosine learning rate scheduler with an initial learning rate of $1\mathrm{e}^{-3}$ . We use a linear warm-up for 5 epochs. We use an image resolution of $224\times 224$ for both training and testing. Following the training recipe of [20], we use the same teacher model for distillation [36]. The experiments are conducted with PyTorch 1.12 [35] using 8 NVIDIA A100 GPUs. The latency is measured using iPhone 14 (iOS 16), and the throughput is measured using A100 40 GB GPU. For latency measurements, we compile the models using CoreML library [6] and perform inference with a batch size of 1. For the throughput on A100, the inference is performed using a batch size of 128. + +MS-COCO 2017 [21]: We use our ImageNet pre-trained models as the backbones in Mask-RCNN framework for object detection and instance segmentation on MS-COCO 2017 dataset. The dataset contains 118K training and 5K validation images. Following [20], we finetune our models for 12 epochs with an image size of $1333 \times 800$ and batch size of 32 using AdamW optimizer. We use learning rate of + +$2e^{-4}$ and report the performance for detection and instance segmentation in terms of mean average precision (mAP). + +ADE20K [58]: The dataset comprises 20K training and 2K validation images and contains 150 class categories for scene parsing. Similar to [20] we use our ImageNet pre-trained models to extract image features and semantic FPN [18] as a decoder for segmentation. The model is trained with an image size of $512 \times 512$ for 40K iterations with a batch size of 32 using AdamW optimizer. We use poly learning rate scheduling with an initial learning rate of $2\mathrm{e}^{-4}$ . We report the semantic segmentation performance in terms of mean intersection over union (mIoU). + +# 4.2. Baseline Comparison + +Table 1 illustrates the impact of integrating our proposed contributions into the baseline EfficientFormer-L1 [20] model in terms of ImageNet-1K top-1 accuracy and inference speed. The first row shows the results of the baseline model, which only includes the self-attention based transformer block in the final stage of the network and achieves a top-1 accuracy of $79.2\%$ with a latency of $1.1~\mathrm{ms}$ on an iPhone 14 mobile device. The second row replaces the pool mixers in the baseline model with our proposed Conv. Encoder, resulting in an improvement in performance to $79.9\%$ while maintaining the same latency. In the third row, we replace the transformer block in the baseline with our proposed SwiftFormer Encoder built on the efficient additive attention. Although the performance drops by $0.2\%$ , the inference speed improves by $0.1~\mathrm{ms}$ , and the model has linear complexity with the number of tokens. This enables us to integrate the SwiftFormer Encoder into all stages and achieve better performance while maintaining the same inference speed as of baseline (first versus last row). + +
MethodLatency (ms)Top-1 (%)
EfficientFormer-L1 (Baseline)1.179.2
+ Replace Pool Mixer by effective Conv. Encoder1.179.9
+ Replace Transformer block by SwiftFormer Encoder1.079.7
+ Incorporate SwiftFormer block across all stages1.180.9
+ +Table 1: Baseline comparison between our SwiftFormer-L1 and EfficientFormer-L1 [20] on the ImageNet-1K dataset. The latency is measured on iPhone14 Neural Engine. + +# 4.3. Image Classification + +Table 2 presents a comparison of our proposed SwiftFormer models (XS, S, L1, and L3) with previous state-of-the-art ConvNets, transformer-based, and hybrid models. We show that our models set new state-of-the-art results, and outperform the recently introduced EfficientFormer [20] and MobileViT-v2 [31] in all model variants. This comprehensive evaluation shows the advantage of our proposed models in terms of both accuracy and latency on mobile devices. + +
ModelTypeLatency (ms) ↓Throughput (A100) ↑Params(M) ↓GMACs ↓Neural SearchTop-1(%) ↑
MobileNet-v2×1.0 [39]ConvNet0.898893.50.3✘71.8
MobileNet-v3-Large×0.75 [15]ConvNet0.8109344.00.2✓73.3
EdgeViT-XXS [34]Hybrid1.759654.10.6✘74.4
MobileViT-XS [30]Hybrid1.537072.30.7✘74.8
SwiftFormer-XSHybrid0.760343.50.6✘75.7
MobileNet-v2×1.4 [39]ConvNet0.974476.10.6✘74.7
MobileNet-v3-Large [15]ConvNet0.9103515.40.3✓75.1
EfficientNet-b0 [42]ConvNet1.385375.30.4✓77.1
DeiT-T [44]Transformer3.858605.73.8✘72.2
EdgeViT-XS [34]Hybrid2.748126.71.1✘77.5
MobileViT-v2×1.0 [31]Hybrid1.732014.91.8✘78.1
SwiftFormer-SHybrid0.850516.11.0✘78.5
MobileFormer-508M [3]Hybrid3.0444314.00.5✘79.3
PoolFormer-S12 [53]Pool1.2322712.01.8✘77.2
EdgeViT-S [34]Hybrid3.5425613.11.9✘81.0
EfficientFormer-L1 [20]Hybrid1.1504612.31.3✓79.2
MobileViT-v2×1.5 [31]Hybrid3.4235610.64.0✘80.4
SwiftFormer-L1Hybrid1.1446912.11.6✘80.9
ResNet-50 [13]ConvNet1.9483525.54.1✘78.5
PoolFormer-S36 [53]Pool2.8111431.05.0✘81.4
ConvNeXt-T [25]ConvNet2.5323528.64.5✘82.1
DeiT-S [44]Transformer9.9299022.54.5✘81.8
Swin-T [25]TransformerNA263528.34.5✘81.3
MobileViT-v2×2.0 [31]Hybrid5.0190618.57.5✘81.2
EfficientFormer-L3 [20]Hybrid2.0269131.33.9✓82.4
SwiftFormer-L3Hybrid1.9289028.54.0✘83.0
+ +Table 2: Comparison of our proposed SwiftFormer with the state-of-the-art counterpart models on ImageNet-1K. The latency is measured on iPhone 14 Neural Engine (iOS 16) and the throughput is measured on Nvidia A100 GPU. Our models run faster than MobileNets, Hybrid, and Transformer models, with a better trade-off between accuracy and model complexity. The error for the latency measurement is less than $\pm 0.1$ ms. The input size remains consistent at $224\times 224$ for all models, except for the MobileViT [30] and MobileViT-v2 [31] variants, which employ a larger dimension of $256\times 256$ . Our results are shown in bold for all model variants. + +Comparison with ConvNets: Our SwiftFormer models surpass the widely used lightweight CNNs counterparts significantly in terms of top-1 accuracy, while running faster than the highly optimized MobileNet-v2 and MobileNet-v3 on an iPhone 14 mobile device. Specifically, our SwiftFormer-XS runs $0.1\mathrm{ms}$ faster than MobileNet-v2 $\times 1.0$ and MobileNet-v3-Large $\times 0.75$ and achieve better top-1 accuracy with a margin of $3.9\%$ and $2.4\%$ respectively. Our SwiftFormer-S runs faster than EfficientNet-b0 [42] by $1.6\times$ and achieves $1.4\%$ higher top-1 accuracy. Further, our SwiftFormer-L3 achieves $4.5\%$ and $0.9\%$ gain in top-1 accuracy over ResNet-50 and ConvNeXt-T, respectively, while running at the same latency as ResNet-50 and $1.3\times$ faster than ConvNeXt-T. This demonstrates that our SwiftFormer models, powered by our proposed efficient additive attention, run faster than the lightweight CNN models on mobile devices and achieve superior performance. Recent device-level optimizations for CNN-based models, such as dedicated hardware implementations for convolutions with batch normalization and non-linearity, likely contribute to the high throughput of fully CNN-based models on A100. + +Comparison with transformer models: Although transformer models usually outperform CNN-based models in + +terms of accuracy, they tend to suffer from high latency when running on resource-constrained mobile devices. For instance, DeiT-S, which has a similar model size to ResNet-50 and achieves higher top-1 accuracy by $3.3\%$ , but ResNet-50 runs approximately $5.2\times$ faster on an iPhone 14 mobile device. In contrast, our SwiftFormer-L3 model achieves $1.2\%$ higher accuracy than DeiT-S, while running at the same speed as ResNet-50. Further our SwiftFormer-S model runs approximately $4.7\times$ faster than DeiT-T on an iPhone 14 mobile device and has $6.3\%$ better accuracy. + +Comparison with hybrid models: Although most existing hybrid approaches achieve higher accuracy compared to their lightweight CNN counterparts, they still underperform the fully CNN-based models in terms of latency due to the quadratic complexity of multi-head self-attention. For example, EdgeViT-XXS runs at approximately $2 \times$ slower compared to MobileNet-v3-Large $\times 0.75$ . On the other hand, our SwiftFormer-XS has better latency as compared to lightweight CNNs and approximately is $2 \times$ faster than EdgeViT-XXS and MobileViT-XS, with an overall $1.3\%$ and $0.9\%$ higher top-1 accuracy respectively. Further, our SwiftFormer-L1 model is $3 \times$ faster than the state-of-the-art MobileViT-v2 $\times 1.5$ with $0.5\%$ better top-1 ac + +
BackboneDetection & Instance SegmentationSemantic mIoU(%)
\(AP^{box}\)\(AP^{box}_{50}\)\(AP^{box}_{75}\)\(AP^{mask}\)\(AP^{mask}_{50}\)\(AP^{mask}_{75}\)
ResNet18 [13]34.054.036.731.251.032.732.9
PoolFormer-S12 [53]37.359.040.134.655.836.937.2
EfficientFormer-L1 [20]37.960.341.035.457.337.338.9
SwiftFormer-L141.263.244.838.160.240.741.4
ResNet50 [13]38.058.641.434.455.136.736.7
PVT-S [48]40.462.943.837.860.140.339.9
PoolFormer-S24 [53]40.162.243.437.059.139.640.3
Swin-T [24]42.264.446.239.164.642.041.5
EfficientFormer-L3 [20]41.463.944.738.161.040.443.5
SwiftFormer-L342.764.446.739.161.741.843.9
+ +Table 3: Results using SwiftFormer as a backbone on dense prediction tasks: Object detection and instance segmentation on COCO, whereas semantic segmentation on ADE20K. Our approach outperforms EfficientFormer on all three tasks. + +![](images/295cd82e9390bbd1cea73896c3504b1298ca26843493b0b6f9729fbacfc24afa.jpg) +Figure 4: Results on COCO. Examples for object detection and instance segmentation on the COCO 2017 validation set. Our SwiftFormer-L1 model can accurately detect and segment instances in images. + +![](images/5dc630255cce0a020d28a90ae146c07d7d839652d0a2b22b81111e95ce85774d.jpg) + +![](images/f5511ed9b71067fa4539f7760bfa92c0596035ed6a24f3bba2d43339a03243a4.jpg) + +![](images/d5cc1fe4e0aa983b347a43fdec61313bce1af823ff345d17e6bd8e999e35ad09.jpg) + +![](images/914771bd7f12b94018c4bc67e706dbaeed42cc6cb34010d9f6db9333962d3b02.jpg) + +![](images/a639e07910c3252abbe5de29bccc18a469cc7f820884a8dbc265144dceed7cce.jpg) + +![](images/5615d815e280269b13c6c24ea83df91e691da0872cb42b00e5d61d7ecbd747df.jpg) +Figure 5: Qualitative results on ADE20K. The qualitative examples for semantic segmentation on the ADE20K validation set. Top: Ground truth masks. Bottom: The semantic segmentation results. Our model can accurately segment various indoor and outdoor scenes. + +![](images/272a5b097220bf0f4808b79c284096d7c3f9f447a39ff2753eab15774ee72050.jpg) + +![](images/9279f7c30948903d961f0289bff25484ed1442fda3204ff4234cb952b451091f.jpg) + +![](images/712290b37e3e0bc4db7044f856c47bbc5772196fea7774ffb7c40bbece5ce73b.jpg) + +![](images/a9cae50157efffee9b9b32a47d3c380451df466982222e9f62b8e040372385f1.jpg) + +![](images/29f03f414fc104cde489597b2b32bded815b852573c168daa1a0ce3376db00b3.jpg) + +![](images/ac21309e543873f5de17f3015c2d60f2d2c64f7662e2ab09547b47005b402c35.jpg) + +![](images/3bdb2e9c3057e1537369146c340d0a5b2a01ea82b951f34f1752174a0b4e2ff1.jpg) + +![](images/272b347c72b6d1de59e27791cc153832d17b40d4221de4471c00d320b54d8e16.jpg) + +![](images/9bcfdba7e715d8f7f5ea90f643413750e7b96561e79287d0c47f7ff062fc2cb8.jpg) + +![](images/78589171dd9e4d09c4e83b14ec5acad1bdb7eda1aa02e32500087a9ca91b3c21.jpg) + +curacy. Our SwiftFormer-L3 model achieves $83.0\%$ top-1 accuracy and runs at $1.9~\mathrm{ms}$ , which is $2.6\times$ faster than MobileViT-v2 $\times 2.0$ with an absolute $1.8\%$ accuracy gain. + +Knowledge distillation (KD): It is worth mentioning that we use the same teacher model of DeiT [44] and EfficientFormer [20] in Table 2. We perform a baseline comparison to show the efficiency of our method without using KD. EfficientFormer-L1 achieves $77.3\%$ accuracy w/o KD, compared to $79.2\%$ with KD. Our SwiftFormer-L1 achieves $79.8\%$ w/o KD, compared to $80.9\%$ with KD. This shows that our approach achieves a higher absolute gain of $2.5\%$ over EfficientFormer-L1, when not using KD. + +Our SwiftFormer-XS and SwiftFormer-S w/o KD achieve $74.1\%$ and $77.0\%$ , respectively. + +# 4.4. Object Detection and Instance Segmentation + +Table 3 compares the object detection and instance segmentation results of Mask-RCNN with different lightweight backbones. Our SwiftFormer-L1 backbone achieves 41.2 AP box, surpassing the lightweight ResNet18 and PoolFormer-S12 backbones by 7.2 and 3.9 points respectively. Further, it performs better than the previous state-of-the-art EfficientFormer-L1 backbone by 3.3 AP box. For instance segmentation, our method achieves 38.1 AP mask + +score which is 2.7 points better than the previous state-of-the-art. Similar trend is observed for SwinfFormer-L3 backbone, which surpasses the previous state-of-the-art EfficientFormer-L3 backbone by 1.3 points and 1.0 points in AP box and mask respectively. The improvement in the downstream detection and instance segmentation tasks illustrates the effectiveness of our SwiftFormer backbone models for the dense prediction tasks. + +# 4.5. Semantic Segmentation + +Table 3 shows the semantic segmentation results of SwiftFormer backbone-based models as compared to previously proposed backbones. We achieve $41.4\%$ mean intersection over union score using SwiftFormer-L1, surpassing ResNet18 by $8.5\%$ , PoolFormer-S12 by $4.2\%$ , and the state-of-the-art EfficientFormer-L1 by $2.5\%$ . Similarly, our SwiftFormer-L3 backbone-based segmentation models achieve $43.9\mathrm{mIoU}$ , surpassing all previous methods. + +# 4.6. Qualitative Results + +Fig. 4 shows the qualitative results of our detection and instance segmentation model on COCO dataset. We also show in Fig. 5 qualitative examples for semantic segmentation results on the ADE20K dataset. Additionally, we demonstrate in Fig. 6 a baseline comparison with the baseline EfficientFormer-L1. The baseline misclassifies a bird as a sheep in the first example and as a bear in the second example, while misclassifies kites as birds in the third example. Our SwiftFormer-L1 accurately detects and segments objects in these examples, which demonstrates the ability of SwiftFormer to accurately identify and segment objects within a diverse range of indoor and outdoor settings. + +![](images/d53b4d09eb86ba9b0d12a966bb820e78bb017b3f58dfd2d3d4801b702e9d4481.jpg) +Figure 6: Qualitative comparison between EfficientFormer-L1 (top) and our SwiftFormer-L1 (bottom) on images from COCO val set for detection and instance segmentation. + +# 5. Ablation study + +To show the effectiveness of the proposed efficient additive attention, we conduct a comparison with other attention variants over SwiftFormer-L1 in Table. 4. It is notable that + +
Attention MethodLatency (ms)↓Throughput↑Top-1 (%)↑
Self-attention (Heads=2)8.5331680.6
Self-attention (Heads=4)18.0303581.1
Transpose self-attention (Heads=2)2.0367180.2
Separable self-attention1.4394080.6
Efficient Additive Attention (Ours)1.1446980.9
+ +Table 4: Quantitative analysis of Fig. 2. We replace our attention module with these methods in the same setting. + +![](images/346f957dec657f7ad933f2eb7e9d5a7a677f80e5eb4224f0e646efe85a1d35dc.jpg) +Figure 7: Visualization of the learned global query vector of the efficient additive attention from SwiftFormer-L1 model at different stages. Both the first and last stages show that the global context is captured effectively. + +the proposed efficient additive attention achieves the best trade-off between latency and top-1 accuracy. To demonstrate further insights on the interpretability of the proposed efficient additive attention, we visualize the learned global attention query vector $\alpha$ of Eq. 4 of the first and last stages of the network. The global query vector of the efficient additive attention aligns with the semantics of the image and hence acts as an informative source in the additive attention mechanism. Also, it localizes the features of the objects accurately with fine details in the first stage. + +# 6. Conclusion + +Transformers have gained popularity in vision applications due to their effective use of self-attention computation. However, their use in mobile vision applications is challenging due to the quadratic nature of self-attention, which is computationally expensive on mobile devices. In this work, we propose a novel efficient additive attention that replaces the expensive matrix multiplication operations with element-wise operations and eliminates explicit keys-values interaction. Our proposed attention is linear with respect to the input tokens and used in all stages of the network. + +Although we achieve a superior balance between accuracy and efficiency, we observe that in the case of dense small object detection, our approach is less accurate likely due to not using positional encoding or attention biases. Our future work is to improve the performance of efficient additive attention in dense prediction tasks. + +# References + +[1] Adrian Bulat and Georgios Tzimiropoulos. Bit-mixer: Mixed-precision networks with runtime bit-width selection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020. +[3] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[4] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[5] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In Advances in Neural Information Processing Systems, 2021. +[6] CoreMLTools. Use coremltools to convert models from third-party libraries to core ml., 2021. +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2009. +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[9] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. +[10] Mohsen Fayyaz, Soroush Abbasi Kouhpayegani, Farnoush Rezaei Jafari, Eric Sommerlade, Hamid Reza Vaezi Joze, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers. 2022. +[11] TSong Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016. +[12] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023. +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. + +[14] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021. +[15] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenet-v3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019. +[16] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017. +[17] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, 2018. +[18] A. Kirillov, R. Girshick, K. He, and P. Dollar. Panoptic feature pyramid networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[19] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In International Conference on Learning Representations, 2020. +[20] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. 2022. +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. +[22] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco De Nadai. Efficient training of visual transformers with small datasets. In Advances in Neural Information Processing Systems, 2021. +[23] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. +[25] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. +[27] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In European Conference on Computer Vision, 2018. + +[28] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In European Conference on Computer Vision, 2022. +[29] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: Efficiently amalgamated cnn-transformer architecture for mobile vision applications. In CADL2022, 2022. +[30] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations, 2022. +[31] Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. Transactions on Machine Learning Research, 2023. +[32] Sachin Mehta, Mohammad Rastegari, Linda Shapiro, and Hannaneh Hajishirzi. Espnetv2: A light-weight, power efficient, and general purpose convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. +[33] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. +[34] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martinez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In European Conference on Computer Vision, 2022. +[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019. +[36] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In CVPR, 2020. +[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems, 2021. +[38] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In Advances in Neural Information Processing Systems, 2021. +[39] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenet-v2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 2018. +[40] Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. + +Unetr++: Delving into efficient and accurate 3d medical image segmentation. arXiv:2212.04497, 2022. +[41] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. +[42] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, 2019. +[43] Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, et al. Mlp-mixer: An all-mlp architecture for vision. In Advances in Neural Information Processing Systems, 2021. +[44] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2021. +[45] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. +[46] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020. +[47] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In CVPR, 2021. +[48] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021. +[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media, 2022. +[50] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. In Advances in Neural Information Processing Systems, 2016. +[51] Chuhan Wu, Fangzhao Wu, Tao Qi, and Yongfeng Huang. Fastformer: Additive attention can be all you need. arXiv preprint arXiv:2108.09084, 2021. +[52] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. +[53] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. +[54] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Edgeformer: Improving light-weight convnets by learning from vision transformers. arXiv preprint arXiv:2203.03952, 2022. + +[55] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Parc-net: Position aware circular convolution with merits from convnets and transformer. In European Conference on Computer Vision, 2022. +[56] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, 2022. +[57] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. +[58] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. +[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. \ No newline at end of file diff --git a/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/images.zip b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bf631f2ecaf0abd46bb8b0470b0b51317465dc46 --- /dev/null +++ b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7078ad3990a0677156cd2ea8be7e18da8ef3f1845eb4de8cd44c4dc62004bf71 +size 675587 diff --git a/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/layout.json b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..31344cf5036b05a9777553c85087de59e7528500 --- /dev/null +++ b/2023/SwiftFormer_ Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications/layout.json @@ -0,0 +1,10560 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 102, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 102, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 94, + 102, + 501, + 140 + ], + "type": "text", + "content": "SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "spans": [ + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": "Abdelrahman Shaker" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": " Muhammad Maaz" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": " Hanoona Rasheed" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": " Salman Khan" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": " Ming-Hsuan Yang" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{2,3,4}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": " Fahad Shahbaz Khan" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": "Mohamed bin Zayed University of AI " + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": "University of California, Merced \n" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": "Yonsei University " + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": "Google Research " + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 91, + 160, + 502, + 220 + ], + "type": "text", + "content": "Linköping University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 270, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 558 + ], + "type": "text", + "content": "Self-attention has become a defacto choice for capturing global context in various vision applications. However, its quadratic computational complexity with respect to image resolution limits its use in real-time applications, especially for deployment on resource-constrained mobile devices. Although hybrid approaches have been proposed to combine the advantages of convolutions and self-attention for a better speed-accuracy trade-off, the expensive matrix multiplication operations in self-attention remain a bottleneck. In this work, we introduce a novel efficient additive attention mechanism that effectively replaces the quadratic matrix multiplication operations with linear element-wise multiplications. Our design shows that the key-value interaction can be replaced with a linear layer without sacrificing any accuracy. Unlike previous state-of-the-art methods, our efficient formulation of self-attention enables its usage at all stages of the network. Using our proposed efficient additive attention, we build a series of models called \"Swift-Former\" which achieves state-of-the-art performance in terms of both accuracy and mobile inference speed. Our small variant achieves " + }, + { + "bbox": [ + 46, + 270, + 290, + 558 + ], + "type": "inline_equation", + "content": "78.5\\%" + }, + { + "bbox": [ + 46, + 270, + 290, + 558 + ], + "type": "text", + "content": " top-1 ImageNet-1K accuracy with only 0.8 ms latency on iPhone 14, which is more accurate and " + }, + { + "bbox": [ + 46, + 270, + 290, + 558 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 46, + 270, + 290, + 558 + ], + "type": "text", + "content": " faster compared to MobileViT-v2. Our code and models: https://tinyurl.com/5ft8v46w" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "type": "text", + "content": "In recent years, transformer models have shown remarkable success in various vision applications such as classification [8, 44, 24, 23, 9], detection [2, 59, 33, 56, 28], and segmentation [4, 40]. However, deploying these models on resource-constrained mobile devices for real-time applications remains challenging due to their inherently complex nature [20, 29]. Specifically, vision transformers (ViTs) rely on global self-attention, which has a quadratic com" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 243, + 547, + 446 + ], + "blocks": [ + { + "bbox": [ + 306, + 243, + 547, + 446 + ], + "lines": [ + { + "bbox": [ + 306, + 243, + 547, + 446 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 547, + 446 + ], + "type": "image", + "image_path": "e465b2c4ede2ca863f0d496b5efb8565cf4f6b70e777f2e89c77cd8852cf50ca.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 453, + 547, + 513 + ], + "lines": [ + { + "bbox": [ + 305, + 453, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 547, + 513 + ], + "type": "text", + "content": "Figure 1: Latency vs Accuracy Comparison. Compared to the recent EfficientFormer-L1 [20], our SwiftFormer-L1 achieves an absolute gain of " + }, + { + "bbox": [ + 305, + 453, + 547, + 513 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 305, + 453, + 547, + 513 + ], + "type": "text", + "content": " in terms of top-1 accuracy with the same latency and without requiring any neural architecture search." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 542, + 547, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 547, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 547, + 697 + ], + "type": "text", + "content": "plexity with respect to the input image resolution, making it impractical for deployment on low-powered mobile devices [31]. As a result, convolutional neural networks (CNNs) are still the preferred choice for real-time deployment on mobile devices, primarily because the convolution operation is computationally efficient [39, 15]. However, a major limitation of CNNs is their reliance on local connections and stationary weights, which can limit their ability to adapt to variable input resolutions and capture long-range dependencies in the data. Therefore, developing more efficient and flexible models that combine the strengths of both CNNs and transformers is critical, particularly for mobile devices with limited computational resources." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "To achieve this goal, several hybrid approaches have" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 261, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 261, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 261, + 713 + ], + "type": "text", + "content": "* Corresponding author: abdelrahman.youssief@mbzai.ac.ae" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17425" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "been proposed that use lightweight CNN modules in the high-resolution early stages and self-attention in the low-resolution later stages [55, 29, 20]. This approach effectively increases the receptive field of the network and strives to achieve a trade-off between speed and accuracy. Furthermore, different efficient variants of computing self-attention have been proposed to reduce the model complexity. These include computing attention across feature dimensions to implicitly model the global context [29], computing attention within local windows [24], pooling spatial features before applying self-attention [9], and sparsely attending to a fixed number of tokens [34], to name a few." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 289, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 289, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 289, + 329 + ], + "type": "text", + "content": "Although these approaches effectively reduce network complexity, they still involve inefficient matrix multiplication operations that significantly impact latency on mobile devices. To address this issue, Mehta et al. [31] propose a separable self-attention mechanism that replaces matrix multiplication operations to element-wise multiplications. This is achieved by projecting queries to context scores, followed by element-wise multiplication with keys to calculate context vectors for encoding global context." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 332, + 289, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 289, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 289, + 440 + ], + "type": "text", + "content": "In this work, we propose efficient additive attention, which eliminates the need for expensive matrix multiplication operations in computing self-attention. Additionally, we propose to compute the global context using only the query-key interactions followed by a linear transformation, without requiring explicit key-value interactions. This significantly reduces the computational complexity and enables us to use the proposed attention block in all stages of the network. Our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 460, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 460, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 460, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 59, + 460, + 287, + 521 + ], + "type": "text", + "content": "- We introduce efficient additive attention, a new approach for computing self-attention in vision backbones that eliminates the need for expensive matrix multiplication operations, significantly reducing the computational complexity of the model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 533, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 533, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 59, + 533, + 287, + 582 + ], + "type": "text", + "content": "- Unlike previous methods, our proposed efficient attention design can be used at all stages of the network, enabling more effective contextual information capture and achieving superior speed-accuracy trade-off." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "type": "text", + "content": "- We build a series of efficient generic classification models called \"SwiftFormer\", which utilize our proposed efficient additive attention. Our small model achieves " + }, + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "78.5\\%" + }, + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "type": "text", + "content": " top-1 ImageNet-1K [7] accuracy while running at only " + }, + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.8~\\mathrm{ms}" + }, + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "type": "text", + "content": " latency on iPhone 14. Moreover, our large model achieves " + }, + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "83.0\\%" + }, + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "type": "text", + "content": " accuracy with a latency of only " + }, + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "1.9~\\mathrm{ms}" + }, + { + "bbox": [ + 59, + 594, + 288, + 714 + ], + "type": "text", + "content": ". Our model achieves state-of-the-art performance, outperforming recent MobileViTv2 [31] and EfficientFormer [20] by obtaining a better trade-off between accuracy and latency (see Fig. 1)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 414 + ], + "type": "text", + "content": "Efficient CNNs: Designing efficient CNNs for mobile vision applications has received much attention in recent years. MobileNet architectures [16, 39, 15] propose depthwise separable convolutions as well as efficient inverted residual blocks for improved performance on various vision tasks. Other methods aim to improve the efficiency by leveraging depth-wise dilated convolutions [32], channel shuffling and pointwise group convolutions [57, 27], network pruning [11, 50], low bit-width [1, 17], and neural architecture search [41, 15]. CNN-based methods are well-performing, efficient, and fast to train and run on edge devices, resulting in widespread usage in the industry. However, they are spatially local and lack global interaction between the features, which deeply affects their performance. Efficient Transformers: ViTs [8] have been widely used in numerous vision tasks, and significant advances have been made in terms of data efficiency [22, 44], transformer architecture [30, 20, 3, 54], and token mechanisms [43, 52]. Reducing the number of visual tokens is a major modification in the transformer architecture for efficient deployment. Instead of using a fixed feature representation through the whole architecture, some methods employ a hierarchical design where the resolution is gradually decreased through the stages, including down-sampling techniques [9, 40, 38, 14] and pyramidal structures [47, 49]. Recently, a few methods [37, 10, 34] propose token sparsification techniques to encode only a subset of the most informative tokens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": "Numerous approaches have recently been proposed to reduce the quadratic complexity of self-attention, the computational bottleneck in transformer-based architectures, by computing its approximated variants [34, 29, 31, 46, 19, 5, 45]. EdgeViT [34] uses a global sparse attention module attending only to a few tokens to improve the efficiency, while [48] down-samples the key and value vectors that lead to a better efficiency-accuracy trade-off. GCViT [12] uses a local-global attention mechanism by employing the interactions between a global query and local key and value to capture the contextual information. This approach enables the model to encode both local and global representations efficiently. EdgeNeXt [29] adopts transposed self-attention operation to compute the attention maps across the channel dimension instead of the spatial dimension, followed by token mixing, to have a linear complexity with respect to the number of tokens. Reformer [19] replaces the dot-product attention with a locality-sensitive hashing to group the tokens and reduced the complexity from " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "O(n^{2})" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "O(n\\log n)" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": ". However, this design is only efficient on longer sequences, which is typically not the case for ViTs. LinFormer [46] is a low-rank matrix factorization method that approximates the self-attention matrix with a low-rank matrix, reducing the complexity from " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "O(n^{2})" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "O(n)" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": ". Although matrix factorization methods theoretically reduce the complexity of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17426" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "self-attention, they use expensive projections for computing attention, which may not reflect the reduction in FLOPs and parameters into actual speed on mobile platforms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 288, + 335 + ], + "type": "text", + "content": "Although these methods show promise and have reduced the complexity of self-attention theoretically, they are inadequate for reducing the inference speed for mobile deployment. Since the complexity of the multi-headed self-attention (MHSA) is higher in the earlier stage compared to the last stages, EfficientFormer [20] incorporates MHSA in the last stage only to learn contextual information from the high-level features without increasing the inference speed significantly. Recently, MobileViT-v2 [31] proposes separable self-attention that uses element-wise operations instead of the dot-product to compute the attention maps with linear complexity. Different from the existing approaches, we propose a consistent hybrid design with an efficient additive attention mechanism to model the contextual information with linear complexity. Instead of capturing the pairwise interactions between keys, queries, and values using the dot-product, we use element-wise operations with learnable attention weights to model the interactions between query and keys only, leading to better inference speed." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 346, + 103, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 103, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 103, + 358 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 366, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 413 + ], + "type": "text", + "content": "Motivation: To motivate our method, we first distinguish three desirable characteristics to be considered when designing an efficient yet accurate approach for resource constraint mobile devices." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 594 + ], + "type": "text", + "content": "Efficient Global Context Modeling: As discussed earlier, most existing approaches either employ the standard MHSA or an approximated variant to learn the global context. However, they struggle to operate as fast as MobileNets on resource-constrained devices. This is likely due to the computation-intensive multiplicative operations during attention computation or reliance on advanced reshaping and indexing operations in these approaches. For instance, the recent MobileViT-v2 [31] is " + }, + { + "bbox": [ + 46, + 415, + 287, + 594 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 415, + 287, + 594 + ], + "type": "text", + "content": " slower than MobileNet-v2 [39]. Instead of using matrix multiplications, we argue that encoding the global context using an efficient additive attention design can reduce the operations with respect to the number of tokens. This is expected to help operate at comparable speed and model size, while achieving superior accuracy compared to MobileNets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Rethinking key-value interactions: Other than multiplicative operations during attention computation, additive attention has been recently explored in the NLP domain [51]. However, in the standard form, it performs three-step processing to model query, key, and value interactions. Each step feeds into the subsequent one, thereby requiring sequential processing. Here, we rethink the additive attention for vision tasks by alleviating the need to compute explicit interactions between key-value. We argue that eliminating key-value interactions and replacing them with a sim" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "ple linear transformation empirically encodes better contextual representation. Our design encodes only global queries and key interactions to learn the global contextual information, followed by a linear transformation to compute global context-aware attention weights." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 133, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 288 + ], + "type": "text", + "content": "Consistent Hybrid Design: Most existing works employ MHSA or the approximated variant in the last stages, while avoiding its usage in the earlier stages. This is because the computational complexity of MHSA grows quadratically with the length of the tokens, making it impractical to incorporate during the initial stages. This constraint adds to the design complexity and requires a careful selection of stages, where MHSA can be applied. In contrast, our proposed SwiftFormer module has linear complexity with respect to the token length and can be incorporated in all stages to learn consistent global context at each scale. This consistency improves the model performance and makes it more generalizable and scalable for high-resolution images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 293, + 476, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 476, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 476, + 306 + ], + "type": "text", + "content": "3.1. Overview of Attention Modules" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": "Vision transformer models are built upon the self-attention (see Fig. 2 (a)), which can effectively model the interactions between the input tokens. Specifically, the self-attention has " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": " as an input, where " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": ", comprising " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": " tokens with " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": "-dimensional embedding vector. The input " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": " is projected to query " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q})" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "(\\mathbf{K})" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "(\\mathbf{V})" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": " using three matrices, " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{V}}" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": ". Each self-attention layer comprises " + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": " heads, which allows the model to attend to different views of the input. The self-attention can be described as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 359, + 425, + 545, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 425, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 359, + 425, + 545, + 452 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\cdot \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right) \\cdot \\mathbf {V}. \\tag {1}", + "image_path": "de91193a7c5215ddb59920152970eb37150ba8848c3c6e63e452dffd943ad577.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "text", + "content": "The attention scores between each pair of tokens in " + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "text", + "content": " are computed using the dot-product operation. Next, these scores are normalized followed by Softmax to weigh the interactions between the tokens. Finally, the weighted interactions are multiplied by " + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "text", + "content": " using the dot-product operation to produce the final weighted output. Overall, the complexity of the self-attention is " + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "inline_equation", + "content": "O(n^{2} \\cdot d)" + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "text", + "content": " is the number of tokens and " + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "text", + "content": " is the hidden dimension. The computational and memory demands of " + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\cdot \\mathbf{K}^{\\top}" + }, + { + "bbox": [ + 304, + 456, + 545, + 600 + ], + "type": "text", + "content": " increase quadratically as the number of tokens grows, leading to slow inference speed and high memory usage, making it impractical to run in real-time for long sequences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 600, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 684 + ], + "type": "text", + "content": "To alleviate this issue, [29] proposes the transpose self-attention (see Fig. 2 (b)) to reduce the complexity from quadratic to linear with respect to the number of tokens. Here, the dot-product operation is applied across the channel dimension instead of the spatial dimension. This allows the model to learn feature maps with implicit contextual representation. The attention can be described as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 689, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 689, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 359, + 689, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = \\mathbf {V} \\cdot \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} ^ {\\top} \\cdot \\mathbf {K}}{\\sqrt {d}}\\right). \\tag {2}", + "image_path": "c8fa075f6fdeb7491705cb025a2257cd368f0e94b00363408756d15b27a86bc8.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17427" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 154, + 220 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 154, + 220 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 154, + 220 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 154, + 220 + ], + "type": "image", + "image_path": "f4dadbfd779cf46e7485bcf975029c11607f9bcbb982d8d877b7d4929d7444b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 230, + 124, + 245 + ], + "blocks": [ + { + "bbox": [ + 83, + 230, + 124, + 245 + ], + "lines": [ + { + "bbox": [ + 83, + 230, + 124, + 245 + ], + "spans": [ + { + "bbox": [ + 83, + 230, + 124, + 245 + ], + "type": "image", + "image_path": "ce019951c2435c475affca2889a026a83ec51408895987e0368a1cf3bf5aaaf0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "text", + "content": "Figure 2: Comparison with different self-attention modules. (a) is a typical self-attention used in ViTs [8]. (b) is the transpose self-attention used in EdgeNeXt [29], where the self-attention operation is applied across channel feature dimensions " + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "inline_equation", + "content": "(\\mathrm{d} \\times \\mathrm{d})" + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "text", + "content": " instead of the spatial dimension " + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "inline_equation", + "content": "(\\mathrm{n} \\times \\mathrm{n})" + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "text", + "content": ". (c) is the separable self-attention of MobileViT-v2 [31], it uses element-wise operations to compute the context vector from the interactions of " + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "text", + "content": " matrices. Then, the context vector is multiplied by " + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "text", + "content": " matrix to produce the final output. (d) Our proposed efficient additive self-attention. Here, the query matrix is multiplied by learnable weights and pooled to produce global queries. Then, the matrix " + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 265, + 545, + 350 + ], + "type": "text", + "content": " is element-wise multiplied by the broadcasted global queries, resulting the global context representation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 157, + 73, + 271, + 214 + ], + "blocks": [ + { + "bbox": [ + 157, + 73, + 271, + 214 + ], + "lines": [ + { + "bbox": [ + 157, + 73, + 271, + 214 + ], + "spans": [ + { + "bbox": [ + 157, + 73, + 271, + 214 + ], + "type": "image", + "image_path": "f73e2b92d00512d0387eee4f53d0739fae7e8873798e735034a3b05707448349.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 230, + 233, + 244 + ], + "blocks": [ + { + "bbox": [ + 220, + 230, + 233, + 244 + ], + "lines": [ + { + "bbox": [ + 220, + 230, + 233, + 244 + ], + "spans": [ + { + "bbox": [ + 220, + 230, + 233, + 244 + ], + "type": "image", + "image_path": "cbe2ad1a39b283ffae4a407beae8ce0dd6c57e10a8120f259dfdb3aa10d0adfe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 232, + 270, + 242 + ], + "lines": [ + { + "bbox": [ + 235, + 232, + 270, + 242 + ], + "spans": [ + { + "bbox": [ + 235, + 232, + 270, + 242 + ], + "type": "text", + "content": "Dot product" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 286, + 73, + 396, + 217 + ], + "blocks": [ + { + "bbox": [ + 286, + 73, + 396, + 217 + ], + "lines": [ + { + "bbox": [ + 286, + 73, + 396, + 217 + ], + "spans": [ + { + "bbox": [ + 286, + 73, + 396, + 217 + ], + "type": "image", + "image_path": "9f7fad8cfecfe1437fd5f3906b308ad92ee323de6f1981a7292d4e13a27fa0d3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 273, + 232, + 285, + 243 + ], + "blocks": [ + { + "bbox": [ + 273, + 232, + 285, + 243 + ], + "lines": [ + { + "bbox": [ + 273, + 232, + 285, + 243 + ], + "spans": [ + { + "bbox": [ + 273, + 232, + 285, + 243 + ], + "type": "image", + "image_path": "64098f4a824891d3903e14023254a3e540ca1176cb161bb76e4af57ab5eb9d3e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 232, + 402, + 243 + ], + "lines": [ + { + "bbox": [ + 287, + 232, + 402, + 243 + ], + "spans": [ + { + "bbox": [ + 287, + 232, + 402, + 243 + ], + "type": "text", + "content": "Broadcasted element-wise multiplication" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 408, + 73, + 531, + 254 + ], + "blocks": [ + { + "bbox": [ + 408, + 73, + 531, + 254 + ], + "lines": [ + { + "bbox": [ + 408, + 73, + 531, + 254 + ], + "spans": [ + { + "bbox": [ + 408, + 73, + 531, + 254 + ], + "type": "image", + "image_path": "1f2df7d2f6545780db8f993cde62f0d32290584c1462b44568b79e49b0f5508b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 370, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 287, + 430 + ], + "type": "text", + "content": "The transpose self-attention has a computational complexity of " + }, + { + "bbox": [ + 46, + 370, + 287, + 430 + ], + "type": "inline_equation", + "content": "O(n \\cdot d^2)" + }, + { + "bbox": [ + 46, + 370, + 287, + 430 + ], + "type": "text", + "content": ". While this complexity scales linearly with the number of tokens " + }, + { + "bbox": [ + 46, + 370, + 287, + 430 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 370, + 287, + 430 + ], + "type": "text", + "content": ", it remains quadratic with respect to the feature dimension " + }, + { + "bbox": [ + 46, + 370, + 287, + 430 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 370, + 287, + 430 + ], + "type": "text", + "content": ". Further, the dot-product operation is still utilized between the query and key matrices." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": "The separable self-attention mechanism (see Fig. 2 (c)) aims to address the bottleneck of the standard self-attention. Here, the interactions between the queries " + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q})" + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": ", keys " + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "inline_equation", + "content": "(\\mathbf{K})" + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": ", and values " + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "inline_equation", + "content": "(\\mathbf{V})" + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": " are encoded using element-wise operations. First, the query matrix " + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": " is projected to produce a vector " + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": " of dimensions " + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "inline_equation", + "content": "n \\times 1" + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": ", and then fed into Softmax to generate the context scores, which captures the importance of each query element. Then, the context scores are multiplied by the key matrix " + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": " and pooled to compute a context vector, which encodes the contextual information. Finally, the context vector is multiplied element-wise with the value matrix " + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": " to propagate the contextual information and produce the final output " + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 430, + 287, + 586 + ], + "type": "text", + "content": ". It can be summarized as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 99, + 593, + 287, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 593, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 99, + 593, + 287, + 610 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = \\mathbf {V} * \\sum \\mathbf {K} * \\operatorname {S o f t m a x} (\\mathbf {q}). \\tag {3}", + "image_path": "3f77bc807ecacb62eefa423a2d6006197b9fcfa14befb35accfb0e2e10b61811.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 616, + 284, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 284, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 284, + 628 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 616, + 284, + 628 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 47, + 616, + 284, + 628 + ], + "type": "text", + "content": " denotes the element-wise multiplication operation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 635, + 200, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 200, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 200, + 647 + ], + "type": "text", + "content": "3.2. Efficient Additive Attention" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "The typical additive attention mechanism in NLP captures the global context by utilizing pairwise interactions between the tokens via element-wise multiplications instead of using dot-product operation. It encodes the relevance scores for the contextual information of the input sequence based on" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": "the interactions of the three attention components (Q, K, V). In contrast, we show that key-value interactions can be removed without sacrificing the performance and only focusing on effectively encoding query-key interactions by incorporating a linear projection layer is sufficient to learn the relationship between the tokens (see Fig. 2 (d)). This approach, named efficient additive attention, has a faster inference speed and produces more robust contextual representations as demonstrated by our performance on image classification, object detection, and segmentation tasks (Sec. 4). Specifically, the input embedding matrix " + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": " is transformed into query (Q) and key (K) using two matrices " + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{W_q}" + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{W_k}" + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": ", where Q, " + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{W_q}" + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{W_k} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": " is the token length and " + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": " is the dimensions of the embedding vector. Next, the query matrix Q is multiplied by learnable parameter vector " + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_a \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": " to learn the attention weights of the query, producing global attention query vector " + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 304, + 370, + 545, + 575 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 390, + 585, + 545, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 585, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 390, + 585, + 545, + 600 + ], + "type": "interline_equation", + "content": "\\alpha = \\mathbf {Q} \\cdot \\mathbf {w} _ {a} / \\sqrt {d} \\tag {4}", + "image_path": "f3955fca0e818fbab1f418b027c1a6742b7a0d84d65ef36a63dacc84a252cc22.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": "Then, the query matrix is pooled based on the learned attention weights, resulting in a single global query vector " + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 389, + 650, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 650, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 389, + 650, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {q} = \\sum_ {i = 1} ^ {n} \\alpha_ {i} * \\mathbf {Q} _ {\\mathbf {i}}. \\tag {5}", + "image_path": "2d8b687cc7fcc7104fdcf9a57e61370a9a3ff634e369e4020ebb75d4af98b1fa.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Next, the interactions between the global query vector " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " and key matrix " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " are encoded using" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17428" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 73, + 286, + 201 + ], + "blocks": [ + { + "bbox": [ + 49, + 73, + 286, + 201 + ], + "lines": [ + { + "bbox": [ + 49, + 73, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 49, + 73, + 286, + 201 + ], + "type": "image", + "image_path": "6ef999bfc9198042cd8cfcd4d703bf261ddae56ca0f7c80b9cc8830ae68e2988.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 289, + 416 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 289, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 289, + 416 + ], + "type": "text", + "content": "Figure 3: Top Row: Overview of our proposed architecture. The input image is fed into the patch embedding layer, followed by hierarchical stages at four different scales " + }, + { + "bbox": [ + 46, + 213, + 289, + 416 + ], + "type": "inline_equation", + "content": "\\{\\frac{1}{4},\\frac{1}{8},\\frac{1}{16},\\frac{1}{32}\\}" + }, + { + "bbox": [ + 46, + 213, + 289, + 416 + ], + "type": "text", + "content": ". Each stage is consistent and compose of Conv. Encoder blocks followed by SwiftFormer Encoder. Between two consecutive stages, we incorporate downsampling layer to reduce the spatial size by a factor of two and increase the feature dimensions. Bottom Row: We show the design of the Conv. Eencoder (left) and the SwiftFormer Encoder (right). The Conv. Encoder is designed to learn effective local representations and consists of " + }, + { + "bbox": [ + 46, + 213, + 289, + 416 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 46, + 213, + 289, + 416 + ], + "type": "text", + "content": " depth-wise convolutions followed by two point-wise convolutions for channel mixing. The SwiftFormer Encoder aims to learn enriched local-global representations. It begins with local convolutional layers to extract local features, followed by the efficient additive attention module (see Fig. 2 (d)) and linear layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 436, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 556 + ], + "type": "text", + "content": "the element-wise product to form global context " + }, + { + "bbox": [ + 46, + 436, + 287, + 556 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^{n\\times d})" + }, + { + "bbox": [ + 46, + 436, + 287, + 556 + ], + "type": "text", + "content": ". This matrix shares a similarity with the attention matrix in MHSA and captures information from every token and has the flexibility to learn the correlation in the input sequence. However, it is comparatively inexpensive to compute compared to MHSA and has linear complexity with the token length. Inspired by the transformer architecture, we employ a linear transformation layer to query-key interactions to learn the hidden representation of the tokens. The output of the efficient additive attention " + }, + { + "bbox": [ + 46, + 436, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 436, + 287, + 556 + ], + "type": "text", + "content": " can be described as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 562, + 287, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 562, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 121, + 562, + 287, + 582 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} = \\hat {\\mathbf {Q}} + \\mathbf {T} (\\mathbf {K} * \\mathbf {q}). \\tag {6}", + "image_path": "74e53e1dc31a0c3edd9eea07a7362bf04b127ed9a3499f15766fe44861ff40b6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 591, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 287, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 591, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Q}}" + }, + { + "bbox": [ + 47, + 591, + 287, + 616 + ], + "type": "text", + "content": " denotes to the normalized query matrix, " + }, + { + "bbox": [ + 47, + 591, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 47, + 591, + 287, + 616 + ], + "type": "text", + "content": " denotes to the linear transformation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 623, + 192, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 192, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 192, + 635 + ], + "type": "text", + "content": "3.3. SwiftFormer Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Our SwiftFormer is based on the recently introduced EfficientFormer [20]. The main idea of EfficientFormer is to introduce 4D MetaBlocks based on PoolFormer [53] to learn local representations efficiently, while using 3D MetaBlocks based on self-attention to encode global context. However, the performance of EfficientFormer is lim" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 239 + ], + "type": "text", + "content": "ited by two design choices. Firstly, it uses ineffective token mixing, and secondly, it only employs 3D MetaBlocks in the last stage due to quadratic complexity of MHSA. This likely leads to inconsistent and insufficient contextual representation. To address these limitations, our SwiftFormer improves the token mixing by using a simple yet effective Conv. Encoder. Further, we introduce efficient additive attention module that can be incorporated in all stages (Sec. 3.2). This leads to more consistent learning of local-global representations. It is worth mentioning that EfficientFormer employs a latency-driven slimming method to obtain optimal configurations for its model variants, which leads to maximizing the speed. In contrast, our SwiftFormer models are built without using any neural architecture search." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "text", + "content": "Fig. 3 shows an overview of our proposed architecture. The main components are: (i) Effective Conv. Encoder, and (ii) SwiftFormer Encoder. In contrast to other hybrid designs, the proposed architecture is consistent and has Conv. Encoders followed by SwiftFormer Encoder in all stages. Our architecture extracts hierarchical features at four different scales across four stages. At the beginning of the network, the input image of size " + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "inline_equation", + "content": "H \\times W \\times 3" + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "text", + "content": " is fed through Patch Embedding layer, implemented with two " + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "text", + "content": " convolutions with a stride of 2, resulting " + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "inline_equation", + "content": "\\frac{H}{4} \\times \\frac{W}{4} \\times C_1" + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "text", + "content": " feature maps. Then, the output feature maps are fed into the first stage, which begins with Conv. Encoder to extract spatial features, followed by SwiftFormer to learn the local-global information. Between two consecutive stages, there is a downsampling layer to increase the channel dimension and reduce the token length. Next, the resulting feature maps are subsequently fed into the second, third, and fourth stages of the architecture, producing " + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "inline_equation", + "content": "\\frac{H}{8} \\times \\frac{W}{8} \\times C_2" + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "inline_equation", + "content": "\\frac{H}{16} \\times \\frac{W}{16} \\times C_3" + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "inline_equation", + "content": "\\frac{H}{32} \\times \\frac{W}{32} \\times C_4" + }, + { + "bbox": [ + 304, + 240, + 546, + 502 + ], + "type": "text", + "content": " dimensional feature maps, respectively. Hence, each stage learns local-global features at different scales of the input image, which allows the network to have enriched representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "type": "text", + "content": "Effective Conv. Encoder: The baseline EfficientFormer [20] employs " + }, + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "type": "text", + "content": " average pooling layers as a local token mixer, similar to PoolFormer [53]. Although PoolFormer layers are known for their fast inference speed, replacing them with depth-wise convolutions does not increase the latency. Further, it improves the performance without increasing the parameters and latency. Specifically, the features maps " + }, + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "type": "text", + "content": " are fed into " + }, + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "type": "text", + "content": " depth-wise convolution (DWConv) followed by Batch Normalization (BN). Then, the resulting features are fed into two point-wise convolutions " + }, + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "type": "inline_equation", + "content": "(\\mathrm{Conv}_1)" + }, + { + "bbox": [ + 304, + 503, + 547, + 658 + ], + "type": "text", + "content": " alongside GeLU activation. Finally, we incorporate a skip connection to enable information to flow across the network. The Conv. Encoder is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 666, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 666, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 319, + 666, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {X}} _ {i} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {C o n v} _ {1, G} \\left(\\operatorname {D W C o n v} _ {B N} \\left(\\mathcal {X} _ {i}\\right)\\right)\\right) + \\mathcal {X} _ {i}. \\tag {7}", + "image_path": "cb9dac13a2d1afc71bc5176aec2619ac72acbca170bf7354ee1d0f4bdffeaf6c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_i" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " refers to the input features, " + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{1,G}" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "text", + "content": " refers to point-wise convolution followed by GeLU, DWConv" + }, + { + "bbox": [ + 304, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "_{BN}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17429" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "refers to depth-wise convolution followed by Batch Normalization, and " + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{X}}_i" + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": " refers to the output feature maps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "type": "text", + "content": "SwiftFormer Encoder: This module is carefully designed to efficiently encode enriched local-global representation in each stage. As shown in Fig. 3, the initial block of the SwiftFormer Encoder is composed of " + }, + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "type": "text", + "content": " depth-wise convolution followed by point-wise convolution, which enables the module to learn spatial information and encode local representation. Then, the resulting feature maps are fed into the efficient additive attention block, which aims to learn contextual information at each scale of the input size. Finally, the output feature maps are fed into a Linear block, which composes of two " + }, + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "type": "text", + "content": " point-wise convolution layers, Batch Normalization, and GeLU activation to generate non-linear features. The SwiftFormer Encoder is described as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 99, + 259, + 233, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 259, + 233, + 274 + ], + "spans": [ + { + "bbox": [ + 99, + 259, + 233, + 274 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {X}} _ {i} = \\operatorname {C o n v} _ {1} (\\operatorname {D W C o n v} _ {B N} (\\hat {\\mathcal {X}} _ {i})),", + "image_path": "fe852f0dfc0c9a230d11d97693e558131c01e27115325b80aab2bb87e293f243.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 277, + 287, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 277, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 126, + 277, + 287, + 291 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {X}} _ {i} = \\mathrm {Q K} (\\hat {\\mathcal {X}} _ {i}) + \\hat {\\mathcal {X}} _ {i}, \\tag {8}", + "image_path": "f39a31701d9db0c11a9cc0e3a7c277c8657436f44fbf2c66fb44c64e7fc4d1e3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 294, + 251, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 294, + 251, + 308 + ], + "spans": [ + { + "bbox": [ + 83, + 294, + 251, + 308 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {X}} _ {i + 1} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {C o n v} _ {B N, 1, G} \\left(\\hat {\\mathcal {X}} _ {i}\\right)\\right) + \\hat {\\mathcal {X}} _ {i}.", + "image_path": "2902782eab09ff82e8d295f2ba11d7f7c83e487aa572fd8bef8f89b146dfa700.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 318, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 286, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 318, + 286, + 354 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{BN,1,G}" + }, + { + "bbox": [ + 47, + 318, + 286, + 354 + ], + "type": "text", + "content": " denotes to Batch Normalization, followed by, " + }, + { + "bbox": [ + 47, + 318, + 286, + 354 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 47, + 318, + 286, + 354 + ], + "type": "text", + "content": " Conv layer, followed by GeLU, and QK denotes the efficient additive attention (explained in Sec. 6)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 365, + 127, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 127, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 127, + 379 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 386, + 286, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 286, + 434 + ], + "type": "text", + "content": "We evaluate our SwiftFormer models across four downstream tasks: classification on ImageNet-1K [7], object detection and instance segmentation on MS-COCO 2017 [21], and semantic segmentation on ADE20K [58]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 443, + 180, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 180, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 180, + 455 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 461, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 286, + 628 + ], + "type": "text", + "content": "ImageNet-1K [7]: All of our models are trained from scratch on ImageNet-1K dataset for 300 epochs with AdamW optimizer [26] and cosine learning rate scheduler with an initial learning rate of " + }, + { + "bbox": [ + 46, + 461, + 286, + 628 + ], + "type": "inline_equation", + "content": "1\\mathrm{e}^{-3}" + }, + { + "bbox": [ + 46, + 461, + 286, + 628 + ], + "type": "text", + "content": ". We use a linear warm-up for 5 epochs. We use an image resolution of " + }, + { + "bbox": [ + 46, + 461, + 286, + 628 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 46, + 461, + 286, + 628 + ], + "type": "text", + "content": " for both training and testing. Following the training recipe of [20], we use the same teacher model for distillation [36]. The experiments are conducted with PyTorch 1.12 [35] using 8 NVIDIA A100 GPUs. The latency is measured using iPhone 14 (iOS 16), and the throughput is measured using A100 40 GB GPU. For latency measurements, we compile the models using CoreML library [6] and perform inference with a batch size of 1. For the throughput on A100, the inference is performed using a batch size of 128." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "MS-COCO 2017 [21]: We use our ImageNet pre-trained models as the backbones in Mask-RCNN framework for object detection and instance segmentation on MS-COCO 2017 dataset. The dataset contains 118K training and 5K validation images. Following [20], we finetune our models for 12 epochs with an image size of " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "1333 \\times 800" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " and batch size of 32 using AdamW optimizer. We use learning rate of" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "2e^{-4}" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " and report the performance for detection and instance segmentation in terms of mean average precision (mAP)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 96, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 216 + ], + "type": "text", + "content": "ADE20K [58]: The dataset comprises 20K training and 2K validation images and contains 150 class categories for scene parsing. Similar to [20] we use our ImageNet pre-trained models to extract image features and semantic FPN [18] as a decoder for segmentation. The model is trained with an image size of " + }, + { + "bbox": [ + 304, + 96, + 545, + 216 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 96, + 545, + 216 + ], + "type": "text", + "content": " for 40K iterations with a batch size of 32 using AdamW optimizer. We use poly learning rate scheduling with an initial learning rate of " + }, + { + "bbox": [ + 304, + 96, + 545, + 216 + ], + "type": "inline_equation", + "content": "2\\mathrm{e}^{-4}" + }, + { + "bbox": [ + 304, + 96, + 545, + 216 + ], + "type": "text", + "content": ". We report the semantic segmentation performance in terms of mean intersection over union (mIoU)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 224, + 428, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 428, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 428, + 237 + ], + "type": "text", + "content": "4.2. Baseline Comparison" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "text", + "content": "Table 1 illustrates the impact of integrating our proposed contributions into the baseline EfficientFormer-L1 [20] model in terms of ImageNet-1K top-1 accuracy and inference speed. The first row shows the results of the baseline model, which only includes the self-attention based transformer block in the final stage of the network and achieves a top-1 accuracy of " + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "inline_equation", + "content": "79.2\\%" + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "text", + "content": " with a latency of " + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "inline_equation", + "content": "1.1~\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "text", + "content": " on an iPhone 14 mobile device. The second row replaces the pool mixers in the baseline model with our proposed Conv. Encoder, resulting in an improvement in performance to " + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "inline_equation", + "content": "79.9\\%" + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "text", + "content": " while maintaining the same latency. In the third row, we replace the transformer block in the baseline with our proposed SwiftFormer Encoder built on the efficient additive attention. Although the performance drops by " + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "text", + "content": ", the inference speed improves by " + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "inline_equation", + "content": "0.1~\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 243, + 545, + 471 + ], + "type": "text", + "content": ", and the model has linear complexity with the number of tokens. This enables us to integrate the SwiftFormer Encoder into all stages and achieve better performance while maintaining the same inference speed as of baseline (first versus last row)." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 308, + 480, + 544, + 533 + ], + "blocks": [ + { + "bbox": [ + 308, + 480, + 544, + 533 + ], + "lines": [ + { + "bbox": [ + 308, + 480, + 544, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 480, + 544, + 533 + ], + "type": "table", + "html": "
MethodLatency (ms)Top-1 (%)
EfficientFormer-L1 (Baseline)1.179.2
+ Replace Pool Mixer by effective Conv. Encoder1.179.9
+ Replace Transformer block by SwiftFormer Encoder1.079.7
+ Incorporate SwiftFormer block across all stages1.180.9
", + "image_path": "6b40b45ea97a109b9ede1b45154284d8d7fa0244a013478b245d43f78e847285.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 540, + 545, + 578 + ], + "lines": [ + { + "bbox": [ + 305, + 540, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 545, + 578 + ], + "type": "text", + "content": "Table 1: Baseline comparison between our SwiftFormer-L1 and EfficientFormer-L1 [20] on the ImageNet-1K dataset. The latency is measured on iPhone14 Neural Engine." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 599, + 423, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 423, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 423, + 611 + ], + "type": "text", + "content": "4.3. Image Classification" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Table 2 presents a comparison of our proposed SwiftFormer models (XS, S, L1, and L3) with previous state-of-the-art ConvNets, transformer-based, and hybrid models. We show that our models set new state-of-the-art results, and outperform the recently introduced EfficientFormer [20] and MobileViT-v2 [31] in all model variants. This comprehensive evaluation shows the advantage of our proposed models in terms of both accuracy and latency on mobile devices." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17430" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 538, + 332 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 538, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 538, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 538, + 332 + ], + "type": "table", + "html": "
ModelTypeLatency (ms) ↓Throughput (A100) ↑Params(M) ↓GMACs ↓Neural SearchTop-1(%) ↑
MobileNet-v2×1.0 [39]ConvNet0.898893.50.3✘71.8
MobileNet-v3-Large×0.75 [15]ConvNet0.8109344.00.2✓73.3
EdgeViT-XXS [34]Hybrid1.759654.10.6✘74.4
MobileViT-XS [30]Hybrid1.537072.30.7✘74.8
SwiftFormer-XSHybrid0.760343.50.6✘75.7
MobileNet-v2×1.4 [39]ConvNet0.974476.10.6✘74.7
MobileNet-v3-Large [15]ConvNet0.9103515.40.3✓75.1
EfficientNet-b0 [42]ConvNet1.385375.30.4✓77.1
DeiT-T [44]Transformer3.858605.73.8✘72.2
EdgeViT-XS [34]Hybrid2.748126.71.1✘77.5
MobileViT-v2×1.0 [31]Hybrid1.732014.91.8✘78.1
SwiftFormer-SHybrid0.850516.11.0✘78.5
MobileFormer-508M [3]Hybrid3.0444314.00.5✘79.3
PoolFormer-S12 [53]Pool1.2322712.01.8✘77.2
EdgeViT-S [34]Hybrid3.5425613.11.9✘81.0
EfficientFormer-L1 [20]Hybrid1.1504612.31.3✓79.2
MobileViT-v2×1.5 [31]Hybrid3.4235610.64.0✘80.4
SwiftFormer-L1Hybrid1.1446912.11.6✘80.9
ResNet-50 [13]ConvNet1.9483525.54.1✘78.5
PoolFormer-S36 [53]Pool2.8111431.05.0✘81.4
ConvNeXt-T [25]ConvNet2.5323528.64.5✘82.1
DeiT-S [44]Transformer9.9299022.54.5✘81.8
Swin-T [25]TransformerNA263528.34.5✘81.3
MobileViT-v2×2.0 [31]Hybrid5.0190618.57.5✘81.2
EfficientFormer-L3 [20]Hybrid2.0269131.33.9✓82.4
SwiftFormer-L3Hybrid1.9289028.54.0✘83.0
", + "image_path": "fa959215b49c283ceec6f8f0be11eeaceda165136ba5a279f04be1d7d3339bc8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 340, + 545, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 340, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 545, + 413 + ], + "type": "text", + "content": "Table 2: Comparison of our proposed SwiftFormer with the state-of-the-art counterpart models on ImageNet-1K. The latency is measured on iPhone 14 Neural Engine (iOS 16) and the throughput is measured on Nvidia A100 GPU. Our models run faster than MobileNets, Hybrid, and Transformer models, with a better trade-off between accuracy and model complexity. The error for the latency measurement is less than " + }, + { + "bbox": [ + 46, + 340, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\pm 0.1" + }, + { + "bbox": [ + 46, + 340, + 545, + 413 + ], + "type": "text", + "content": " ms. The input size remains consistent at " + }, + { + "bbox": [ + 46, + 340, + 545, + 413 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 46, + 340, + 545, + 413 + ], + "type": "text", + "content": " for all models, except for the MobileViT [30] and MobileViT-v2 [31] variants, which employ a larger dimension of " + }, + { + "bbox": [ + 46, + 340, + 545, + 413 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 46, + 340, + 545, + 413 + ], + "type": "text", + "content": ". Our results are shown in bold for all model variants." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": "Comparison with ConvNets: Our SwiftFormer models surpass the widely used lightweight CNNs counterparts significantly in terms of top-1 accuracy, while running faster than the highly optimized MobileNet-v2 and MobileNet-v3 on an iPhone 14 mobile device. Specifically, our SwiftFormer-XS runs " + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": " faster than MobileNet-v2 " + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "inline_equation", + "content": "\\times 1.0" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": " and MobileNet-v3-Large " + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "inline_equation", + "content": "\\times 0.75" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": " and achieve better top-1 accuracy with a margin of " + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": " respectively. Our SwiftFormer-S runs faster than EfficientNet-b0 [42] by " + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "inline_equation", + "content": "1.6\\times" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": " and achieves " + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": " higher top-1 accuracy. Further, our SwiftFormer-L3 achieves " + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": " gain in top-1 accuracy over ResNet-50 and ConvNeXt-T, respectively, while running at the same latency as ResNet-50 and " + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "inline_equation", + "content": "1.3\\times" + }, + { + "bbox": [ + 46, + 434, + 289, + 685 + ], + "type": "text", + "content": " faster than ConvNeXt-T. This demonstrates that our SwiftFormer models, powered by our proposed efficient additive attention, run faster than the lightweight CNN models on mobile devices and achieve superior performance. Recent device-level optimizations for CNN-based models, such as dedicated hardware implementations for convolutions with batch normalization and non-linearity, likely contribute to the high throughput of fully CNN-based models on A100." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Comparison with transformer models: Although transformer models usually outperform CNN-based models in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "text", + "content": "terms of accuracy, they tend to suffer from high latency when running on resource-constrained mobile devices. For instance, DeiT-S, which has a similar model size to ResNet-50 and achieves higher top-1 accuracy by " + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "text", + "content": ", but ResNet-50 runs approximately " + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "inline_equation", + "content": "5.2\\times" + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "text", + "content": " faster on an iPhone 14 mobile device. In contrast, our SwiftFormer-L3 model achieves " + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "text", + "content": " higher accuracy than DeiT-S, while running at the same speed as ResNet-50. Further our SwiftFormer-S model runs approximately " + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "inline_equation", + "content": "4.7\\times" + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "text", + "content": " faster than DeiT-T on an iPhone 14 mobile device and has " + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 304, + 434, + 545, + 555 + ], + "type": "text", + "content": " better accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "content": "Comparison with hybrid models: Although most existing hybrid approaches achieve higher accuracy compared to their lightweight CNN counterparts, they still underperform the fully CNN-based models in terms of latency due to the quadratic complexity of multi-head self-attention. For example, EdgeViT-XXS runs at approximately " + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "content": " slower compared to MobileNet-v3-Large " + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\times 0.75" + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "content": ". On the other hand, our SwiftFormer-XS has better latency as compared to lightweight CNNs and approximately is " + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "content": " faster than EdgeViT-XXS and MobileViT-XS, with an overall " + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "content": " higher top-1 accuracy respectively. Further, our SwiftFormer-L1 model is " + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "content": " faster than the state-of-the-art MobileViT-v2 " + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\times 1.5" + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "content": " better top-1 ac" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17431" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 104, + 70, + 491, + 209 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 491, + 209 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 491, + 209 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 491, + 209 + ], + "type": "table", + "html": "
BackboneDetection & Instance SegmentationSemantic mIoU(%)
\\(AP^{box}\\)\\(AP^{box}_{50}\\)\\(AP^{box}_{75}\\)\\(AP^{mask}\\)\\(AP^{mask}_{50}\\)\\(AP^{mask}_{75}\\)
ResNet18 [13]34.054.036.731.251.032.732.9
PoolFormer-S12 [53]37.359.040.134.655.836.937.2
EfficientFormer-L1 [20]37.960.341.035.457.337.338.9
SwiftFormer-L141.263.244.838.160.240.741.4
ResNet50 [13]38.058.641.434.455.136.736.7
PVT-S [48]40.462.943.837.860.140.339.9
PoolFormer-S24 [53]40.162.243.437.059.139.640.3
Swin-T [24]42.264.446.239.164.642.041.5
EfficientFormer-L3 [20]41.463.944.738.161.040.443.5
SwiftFormer-L342.764.446.739.161.741.843.9
", + "image_path": "e49862393aaa0bd44e73b967b0784e6bdb1d4b23b216c41afad3870eb217a7ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 547, + 243 + ], + "type": "text", + "content": "Table 3: Results using SwiftFormer as a backbone on dense prediction tasks: Object detection and instance segmentation on COCO, whereas semantic segmentation on ADE20K. Our approach outperforms EfficientFormer on all three tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 252, + 159, + 322 + ], + "blocks": [ + { + "bbox": [ + 48, + 252, + 159, + 322 + ], + "lines": [ + { + "bbox": [ + 48, + 252, + 159, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 159, + 322 + ], + "type": "image", + "image_path": "295cd82e9390bbd1cea73896c3504b1298ca26843493b0b6f9729fbacfc24afa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 339, + 546, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 339, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 546, + 365 + ], + "type": "text", + "content": "Figure 4: Results on COCO. Examples for object detection and instance segmentation on the COCO 2017 validation set. Our SwiftFormer-L1 model can accurately detect and segment instances in images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 253, + 216, + 326 + ], + "blocks": [ + { + "bbox": [ + 159, + 253, + 216, + 326 + ], + "lines": [ + { + "bbox": [ + 159, + 253, + 216, + 326 + ], + "spans": [ + { + "bbox": [ + 159, + 253, + 216, + 326 + ], + "type": "image", + "image_path": "5dc630255cce0a020d28a90ae146c07d7d839652d0a2b22b81111e95ce85774d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 216, + 253, + 324, + 327 + ], + "blocks": [ + { + "bbox": [ + 216, + 253, + 324, + 327 + ], + "lines": [ + { + "bbox": [ + 216, + 253, + 324, + 327 + ], + "spans": [ + { + "bbox": [ + 216, + 253, + 324, + 327 + ], + "type": "image", + "image_path": "f5511ed9b71067fa4539f7760bfa92c0596035ed6a24f3bba2d43339a03243a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 324, + 253, + 435, + 327 + ], + "blocks": [ + { + "bbox": [ + 324, + 253, + 435, + 327 + ], + "lines": [ + { + "bbox": [ + 324, + 253, + 435, + 327 + ], + "spans": [ + { + "bbox": [ + 324, + 253, + 435, + 327 + ], + "type": "image", + "image_path": "d5cc1fe4e0aa983b347a43fdec61313bce1af823ff345d17e6bd8e999e35ad09.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 435, + 253, + 543, + 327 + ], + "blocks": [ + { + "bbox": [ + 435, + 253, + 543, + 327 + ], + "lines": [ + { + "bbox": [ + 435, + 253, + 543, + 327 + ], + "spans": [ + { + "bbox": [ + 435, + 253, + 543, + 327 + ], + "type": "image", + "image_path": "914771bd7f12b94018c4bc67e706dbaeed42cc6cb34010d9f6db9333962d3b02.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 374, + 130, + 437 + ], + "blocks": [ + { + "bbox": [ + 48, + 374, + 130, + 437 + ], + "lines": [ + { + "bbox": [ + 48, + 374, + 130, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 374, + 130, + 437 + ], + "type": "image", + "image_path": "a639e07910c3252abbe5de29bccc18a469cc7f820884a8dbc265144dceed7cce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 437, + 130, + 499 + ], + "blocks": [ + { + "bbox": [ + 49, + 437, + 130, + 499 + ], + "lines": [ + { + "bbox": [ + 49, + 437, + 130, + 499 + ], + "spans": [ + { + "bbox": [ + 49, + 437, + 130, + 499 + ], + "type": "image", + "image_path": "5615d815e280269b13c6c24ea83df91e691da0872cb42b00e5d61d7ecbd747df.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 510, + 547, + 546 + ], + "lines": [ + { + "bbox": [ + 46, + 510, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 547, + 546 + ], + "type": "text", + "content": "Figure 5: Qualitative results on ADE20K. The qualitative examples for semantic segmentation on the ADE20K validation set. Top: Ground truth masks. Bottom: The semantic segmentation results. Our model can accurately segment various indoor and outdoor scenes." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 132, + 374, + 214, + 437 + ], + "blocks": [ + { + "bbox": [ + 132, + 374, + 214, + 437 + ], + "lines": [ + { + "bbox": [ + 132, + 374, + 214, + 437 + ], + "spans": [ + { + "bbox": [ + 132, + 374, + 214, + 437 + ], + "type": "image", + "image_path": "272a5b097220bf0f4808b79c284096d7c3f9f447a39ff2753eab15774ee72050.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 132, + 437, + 214, + 499 + ], + "blocks": [ + { + "bbox": [ + 132, + 437, + 214, + 499 + ], + "lines": [ + { + "bbox": [ + 132, + 437, + 214, + 499 + ], + "spans": [ + { + "bbox": [ + 132, + 437, + 214, + 499 + ], + "type": "image", + "image_path": "9279f7c30948903d961f0289bff25484ed1442fda3204ff4234cb952b451091f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 215, + 374, + 295, + 437 + ], + "blocks": [ + { + "bbox": [ + 215, + 374, + 295, + 437 + ], + "lines": [ + { + "bbox": [ + 215, + 374, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 215, + 374, + 295, + 437 + ], + "type": "image", + "image_path": "712290b37e3e0bc4db7044f856c47bbc5772196fea7774ffb7c40bbece5ce73b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 215, + 437, + 295, + 499 + ], + "blocks": [ + { + "bbox": [ + 215, + 437, + 295, + 499 + ], + "lines": [ + { + "bbox": [ + 215, + 437, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 215, + 437, + 295, + 499 + ], + "type": "image", + "image_path": "a9cae50157efffee9b9b32a47d3c380451df466982222e9f62b8e040372385f1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 295, + 374, + 372, + 437 + ], + "blocks": [ + { + "bbox": [ + 295, + 374, + 372, + 437 + ], + "lines": [ + { + "bbox": [ + 295, + 374, + 372, + 437 + ], + "spans": [ + { + "bbox": [ + 295, + 374, + 372, + 437 + ], + "type": "image", + "image_path": "29f03f414fc104cde489597b2b32bded815b852573c168daa1a0ce3376db00b3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 295, + 437, + 372, + 499 + ], + "blocks": [ + { + "bbox": [ + 295, + 437, + 372, + 499 + ], + "lines": [ + { + "bbox": [ + 295, + 437, + 372, + 499 + ], + "spans": [ + { + "bbox": [ + 295, + 437, + 372, + 499 + ], + "type": "image", + "image_path": "ac21309e543873f5de17f3015c2d60f2d2c64f7662e2ab09547b47005b402c35.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 372, + 374, + 461, + 437 + ], + "blocks": [ + { + "bbox": [ + 372, + 374, + 461, + 437 + ], + "lines": [ + { + "bbox": [ + 372, + 374, + 461, + 437 + ], + "spans": [ + { + "bbox": [ + 372, + 374, + 461, + 437 + ], + "type": "image", + "image_path": "3bdb2e9c3057e1537369146c340d0a5b2a01ea82b951f34f1752174a0b4e2ff1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 372, + 437, + 461, + 499 + ], + "blocks": [ + { + "bbox": [ + 372, + 437, + 461, + 499 + ], + "lines": [ + { + "bbox": [ + 372, + 437, + 461, + 499 + ], + "spans": [ + { + "bbox": [ + 372, + 437, + 461, + 499 + ], + "type": "image", + "image_path": "272b347c72b6d1de59e27791cc153832d17b40d4221de4471c00d320b54d8e16.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 461, + 374, + 544, + 437 + ], + "blocks": [ + { + "bbox": [ + 461, + 374, + 544, + 437 + ], + "lines": [ + { + "bbox": [ + 461, + 374, + 544, + 437 + ], + "spans": [ + { + "bbox": [ + 461, + 374, + 544, + 437 + ], + "type": "image", + "image_path": "9bcfdba7e715d8f7f5ea90f643413750e7b96561e79287d0c47f7ff062fc2cb8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 461, + 437, + 544, + 499 + ], + "blocks": [ + { + "bbox": [ + 461, + 437, + 544, + 499 + ], + "lines": [ + { + "bbox": [ + 461, + 437, + 544, + 499 + ], + "spans": [ + { + "bbox": [ + 461, + 437, + 544, + 499 + ], + "type": "image", + "image_path": "78589171dd9e4d09c4e83b14ec5acad1bdb7eda1aa02e32500087a9ca91b3c21.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": "curacy. Our SwiftFormer-L3 model achieves " + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "inline_equation", + "content": "83.0\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": " top-1 accuracy and runs at " + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "inline_equation", + "content": "1.9~\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "inline_equation", + "content": "2.6\\times" + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": " faster than MobileViT-v2 " + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\times 2.0" + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": " with an absolute " + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 46, + 567, + 287, + 604 + ], + "type": "text", + "content": " accuracy gain." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Knowledge distillation (KD): It is worth mentioning that we use the same teacher model of DeiT [44] and EfficientFormer [20] in Table 2. We perform a baseline comparison to show the efficiency of our method without using KD. EfficientFormer-L1 achieves " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "77.3\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " accuracy w/o KD, compared to " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "79.2\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " with KD. Our SwiftFormer-L1 achieves " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "79.8\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " w/o KD, compared to " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "80.9\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " with KD. This shows that our approach achieves a higher absolute gain of " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " over EfficientFormer-L1, when not using KD." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 568, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 568, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 305, + 568, + 545, + 593 + ], + "type": "text", + "content": "Our SwiftFormer-XS and SwiftFormer-S w/o KD achieve " + }, + { + "bbox": [ + 305, + 568, + 545, + 593 + ], + "type": "inline_equation", + "content": "74.1\\%" + }, + { + "bbox": [ + 305, + 568, + 545, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 568, + 545, + 593 + ], + "type": "inline_equation", + "content": "77.0\\%" + }, + { + "bbox": [ + 305, + 568, + 545, + 593 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 599, + 537, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 537, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 537, + 613 + ], + "type": "text", + "content": "4.4. Object Detection and Instance Segmentation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "content": "Table 3 compares the object detection and instance segmentation results of Mask-RCNN with different lightweight backbones. Our SwiftFormer-L1 backbone achieves 41.2 AP box, surpassing the lightweight ResNet18 and PoolFormer-S12 backbones by 7.2 and 3.9 points respectively. Further, it performs better than the previous state-of-the-art EfficientFormer-L1 backbone by 3.3 AP box. For instance segmentation, our method achieves 38.1 AP mask" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17432" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "score which is 2.7 points better than the previous state-of-the-art. Similar trend is observed for SwinfFormer-L3 backbone, which surpasses the previous state-of-the-art EfficientFormer-L3 backbone by 1.3 points and 1.0 points in AP box and mask respectively. The improvement in the downstream detection and instance segmentation tasks illustrates the effectiveness of our SwiftFormer backbone models for the dense prediction tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 176, + 180, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 180, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 180, + 190 + ], + "type": "text", + "content": "4.5. Semantic Segmentation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "text", + "content": "Table 3 shows the semantic segmentation results of SwiftFormer backbone-based models as compared to previously proposed backbones. We achieve " + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "inline_equation", + "content": "41.4\\%" + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "text", + "content": " mean intersection over union score using SwiftFormer-L1, surpassing ResNet18 by " + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "text", + "content": ", PoolFormer-S12 by " + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "text", + "content": ", and the state-of-the-art EfficientFormer-L1 by " + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "text", + "content": ". Similarly, our SwiftFormer-L3 backbone-based segmentation models achieve " + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "inline_equation", + "content": "43.9\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 195, + 287, + 291 + ], + "type": "text", + "content": ", surpassing all previous methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 300, + 160, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 160, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 160, + 312 + ], + "type": "text", + "content": "4.6. Qualitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 462 + ], + "type": "text", + "content": "Fig. 4 shows the qualitative results of our detection and instance segmentation model on COCO dataset. We also show in Fig. 5 qualitative examples for semantic segmentation results on the ADE20K dataset. Additionally, we demonstrate in Fig. 6 a baseline comparison with the baseline EfficientFormer-L1. The baseline misclassifies a bird as a sheep in the first example and as a bear in the second example, while misclassifies kites as birds in the third example. Our SwiftFormer-L1 accurately detects and segments objects in these examples, which demonstrates the ability of SwiftFormer to accurately identify and segment objects within a diverse range of indoor and outdoor settings." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 473, + 285, + 583 + ], + "blocks": [ + { + "bbox": [ + 50, + 473, + 285, + 583 + ], + "lines": [ + { + "bbox": [ + 50, + 473, + 285, + 583 + ], + "spans": [ + { + "bbox": [ + 50, + 473, + 285, + 583 + ], + "type": "image", + "image_path": "d53b4d09eb86ba9b0d12a966bb820e78bb017b3f58dfd2d3d4801b702e9d4481.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 594, + 287, + 632 + ], + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 632 + ], + "type": "text", + "content": "Figure 6: Qualitative comparison between EfficientFormer-L1 (top) and our SwiftFormer-L1 (bottom) on images from COCO val set for detection and instance segmentation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 657, + 138, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 138, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 138, + 671 + ], + "type": "text", + "content": "5. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "To show the effectiveness of the proposed efficient additive attention, we conduct a comparison with other attention variants over SwiftFormer-L1 in Table. 4. It is notable that" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 127 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 545, + 127 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 127 + ], + "type": "table", + "html": "
Attention MethodLatency (ms)↓Throughput↑Top-1 (%)↑
Self-attention (Heads=2)8.5331680.6
Self-attention (Heads=4)18.0303581.1
Transpose self-attention (Heads=2)2.0367180.2
Separable self-attention1.4394080.6
Efficient Additive Attention (Ours)1.1446980.9
", + "image_path": "42ae0c9507ac5999a06d217faffda781199c0427f9ae7572024d5bc76ca0a68d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 135, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 306, + 135, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 306, + 135, + 545, + 159 + ], + "type": "text", + "content": "Table 4: Quantitative analysis of Fig. 2. We replace our attention module with these methods in the same setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 171, + 545, + 291 + ], + "blocks": [ + { + "bbox": [ + 307, + 171, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 307, + 171, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 545, + 291 + ], + "type": "image", + "image_path": "346f957dec657f7ad933f2eb7e9d5a7a677f80e5eb4224f0e646efe85a1d35dc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 300, + 546, + 348 + ], + "lines": [ + { + "bbox": [ + 305, + 300, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 546, + 348 + ], + "type": "text", + "content": "Figure 7: Visualization of the learned global query vector of the efficient additive attention from SwiftFormer-L1 model at different stages. Both the first and last stages show that the global context is captured effectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 369, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 490 + ], + "type": "text", + "content": "the proposed efficient additive attention achieves the best trade-off between latency and top-1 accuracy. To demonstrate further insights on the interpretability of the proposed efficient additive attention, we visualize the learned global attention query vector " + }, + { + "bbox": [ + 304, + 369, + 545, + 490 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 369, + 545, + 490 + ], + "type": "text", + "content": " of Eq. 4 of the first and last stages of the network. The global query vector of the efficient additive attention aligns with the semantics of the image and hence acts as an informative source in the additive attention mechanism. Also, it localizes the features of the objects accurately with fine details in the first stage." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 501, + 379, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 379, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 379, + 513 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": "Transformers have gained popularity in vision applications due to their effective use of self-attention computation. However, their use in mobile vision applications is challenging due to the quadratic nature of self-attention, which is computationally expensive on mobile devices. In this work, we propose a novel efficient additive attention that replaces the expensive matrix multiplication operations with element-wise operations and eliminates explicit keys-values interaction. Our proposed attention is linear with respect to the input tokens and used in all stages of the network." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Although we achieve a superior balance between accuracy and efficiency, we observe that in the case of dense small object detection, our approach is less accurate likely due to not using positional encoding or attention biases. Our future work is to improve the performance of efficient additive attention in dense prediction tasks." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17433" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Adrian Bulat and Georgios Tzimiropoulos. Bit-mixer: Mixed-precision networks with runtime bit-width selection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 235 + ], + "type": "text", + "content": "[3] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 290 + ], + "type": "text", + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "text", + "content": "[5] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 368 + ], + "type": "text", + "content": "[6] CoreMLTools. Use coremltools to convert models from third-party libraries to core ml., 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2009." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "type": "text", + "content": "[9] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "text", + "content": "[10] Mohsen Fayyaz, Soroush Abbasi Kouhpayegani, Farnoush Rezaei Jafari, Eric Sommerlade, Hamid Reza Vaezi Joze, Hamed Pirsiavash, and Juergen Gall. Adaptive token sampling for efficient vision transformers. 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[11] TSong Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "text", + "content": "[12] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In International Conference on Machine Learning, pages 12633-12646. PMLR, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 163 + ], + "type": "text", + "content": "[15] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenet-v3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 164, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 219 + ], + "type": "text", + "content": "[16] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 221, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 221, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 221, + 545, + 275 + ], + "type": "text", + "content": "[17] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In CVPR, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 277, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 312 + ], + "type": "text", + "content": "[18] A. Kirillov, R. Girshick, K. He, and P. Dollar. Panoptic feature pyramid networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 313, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 313, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 313, + 545, + 346 + ], + "type": "text", + "content": "[19] Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 348, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 348, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 348, + 545, + 381 + ], + "type": "text", + "content": "[20] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 427 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 429, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 429, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 308, + 429, + 545, + 473 + ], + "type": "text", + "content": "[22] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco De Nadai. Efficient training of visual transformers with small datasets. In Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 529 + ], + "type": "text", + "content": "[23] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 532, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 532, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 532, + 545, + 586 + ], + "type": "text", + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 588, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 588, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 588, + 545, + 633 + ], + "type": "text", + "content": "[25] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[27] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In European Conference on Computer Vision, 2018." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17434" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multimodal transformer. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "text", + "content": "[29] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: Efficiently amalgamated cnn-transformer architecture for mobile vision applications. In CADL2022, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 227 + ], + "type": "text", + "content": "[30] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 260 + ], + "type": "text", + "content": "[31] Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. Transactions on Machine Learning Research, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "type": "text", + "content": "[32] Sachin Mehta, Mohammad Rastegari, Linda Shapiro, and Hannaneh Hajishirzi. Espnetv2: A light-weight, power efficient, and general purpose convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 371 + ], + "type": "text", + "content": "[33] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 372, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 426 + ], + "type": "text", + "content": "[34] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martinez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 514 + ], + "type": "text", + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 514, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 547 + ], + "type": "text", + "content": "[36] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In CVPR, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "text", + "content": "[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[38] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. In Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 691 + ], + "type": "text", + "content": "[39] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenet-v2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition*, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[40] Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Unetr++: Delving into efficient and accurate 3d medical image segmentation. arXiv:2212.04497, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[41] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[42] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[43] Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, et al. Mlp-mixer: An all-mlp architecture for vision. In Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "text", + "content": "[44] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "text", + "content": "[45] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "text", + "content": "[46] Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[47] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In CVPR, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 487 + ], + "type": "text", + "content": "[48] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 568-578, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 488, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 533 + ], + "type": "text", + "content": "[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 533, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 577 + ], + "type": "text", + "content": "[50] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. In Advances in Neural Information Processing Systems, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "text", + "content": "[51] Chuhan Wu, Fangzhao Wu, Tao Qi, and Yongfeng Huang. Fastformer: Additive attention can be all you need. arXiv preprint arXiv:2108.09084, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "text", + "content": "[52] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[53] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[54] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Edgeformer: Improving light-weight convnets by learning from vision transformers. arXiv preprint arXiv:2203.03952, 2022." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17435" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 319 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[55] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Parc-net: Position aware circular convolution with merits from convnets and transformer. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[56] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "text", + "content": "[57] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 229, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 287, + 274 + ], + "type": "text", + "content": "[58] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "text", + "content": "[59] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17436" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/0bea32db-db74-4a9f-85a7-6fe6526e5e78_content_list.json b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/0bea32db-db74-4a9f-85a7-6fe6526e5e78_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef90ed97f7c742df4bac1c1ceba2014d282a9c15 --- /dev/null +++ b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/0bea32db-db74-4a9f-85a7-6fe6526e5e78_content_list.json @@ -0,0 +1,1499 @@ +[ + { + "type": "text", + "text": "SwinLSTM: Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM", + "text_level": 1, + "bbox": [ + 120, + 128, + 849, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Song Tang* Chuang Li* Pu Zhang RongNian Tang† Hainan University {songtang, lc, zhangpu, rn.tang}@hainanu.edu.cn", + "bbox": [ + 225, + 202, + 743, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Integrating CNNs and RNNs to capture spatiotemporal dependencies is a prevalent strategy for spatiotemporal prediction tasks. However, the property of CNNs to learn local spatial information decreases their efficiency in capturing spatiotemporal dependencies, thereby limiting their prediction accuracy. In this paper, we propose a new recurrent cell, SwinLSTM, which integrates Swin Transformer blocks and the simplified LSTM, an extension that replaces the convolutional structure in ConvLSTM with the self-attention mechanism. Furthermore, we construct a network with SwinLSTM cell as the core for spatiotemporal prediction. Without using unique tricks, SwinLSTM outperforms state-of-the-art methods on Moving MNIST, Human3.6m, TaxiBJ, and KTH datasets. In particular, it exhibits a significant improvement in prediction accuracy compared to ConvLSTM. Our competitive experimental results demonstrate that learning global spatial dependencies is more advantageous for models to capture spatiotemporal dependencies. We hope that SwinLSTM can serve as a solid baseline to promote the advancement of spatiotemporal prediction accuracy. The codes are publicly available at https://github.com/SongTang-x/SwinLSTM.", + "bbox": [ + 75, + 323, + 473, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 684, + 209, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spatiotemporal prediction has received increasing attention in recent years due to it can benefit many practical applications, e.g., precipitation forecasting [24, 25, 33], autonomous driving [1, 16], and traffic flow prediction [38, 40]. However, the complex physical dynamics and chaotic properties of spatiotemporal predictive learning make it challenging for purely data-driven deep learning methods to make accurate predictions. Existing methods [3, 8, 19, 24, 31-34, 39] integrate CNNs and RNNs to learn spatiotemporal dependencies in spatiotemporal data to improve prediction accuracy. To capture the spatial and temporal depen", + "bbox": [ + 75, + 710, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d80079b0730ca42fd2873ec5adbcd4a1bca9bbcef15ab45bfc32a4982dae5c03.jpg", + "image_caption": [ + "Figure 1. Comparison of input-to-state and state-to-state transitions in three different recurrent cells. (a) FC-LSTM utilizes linear operations to process 1D vectors. (b) ConvLSTM employs 2D convolutions to process 3D tensors. (c) The proposed SwinLSTM leverages self-attention to process 2D matrices." + ], + "image_footnote": [], + "bbox": [ + 524, + 303, + 874, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dencies simultaneously, ConvLSTM [24] extends the fully connected LSTM (FC-LSTM) [26] by replacing linear operations with convolutional operations. Subsequently, several variants of ConvLSTM are proposed. PredRNN [33] and MIM [34] modify the internal structure of the LSTM unit. E3D-LSTM [32] integrates 3D-Convs into LSTMs. PhyD-Net [8] leverages a CNN-based module to disentangle physical dynamics. Admittedly, these methods achieve impressive results on spatiotemporal prediction tasks. However, convolution operators focus on capturing local features and relations and are inefficient for modeling global spatial information [4]. Although the receptive field can be enlarged by stacking convolution layers, the effective receptive field only reaches a fraction of the theoretical receptive field in practice [21]. Therefore, it can be inferred that the CNN-based models may prove to be ineffective in capturing spa", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 94, + 886, + 336, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13470", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tiotemporal dependencies, owing to the inherent locality of the CNN architecture, leading to restricted accuracy in predictions.", + "bbox": [ + 75, + 90, + 468, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the Vision Transformer (ViT) [6] introduced the Transformer [29] that directly models long-range dependencies into the vision domain. ViT applies a standard Transformer for image classification and attains excellent results with sufficient data. Its outstanding achievement has attracted more researchers to apply Transformer to computer vision. Subsequently, some variants [20, 27, 36, 43] of ViT emerged, with tremendous success on different vision tasks. Notably, the Swin Transformer [20] has exhibited exceptional performance across diverse visual tasks, owing to its unique utilization of local attention and shift window mechanism.", + "bbox": [ + 75, + 136, + 468, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by this, we propose SwinLSTM, a new recurrent cell. Specifically, we integrate Swin Transformer blocks and a simplified LSTM module to extract spatiotemporal representations. In addition, we construct a predictive network with SwinLSTM as the core to capture spatial and temporal dependencies for spatiotemporal prediction tasks. As illustrated in Figure 2 (c), we first split an input image at the current time step into a sequence of image patches. Subsequently, the flattened image patches are fed to the patch embedding layer. Then, the SwinLSTM layer receives the transformed patches or the hidden states transformed by the previous layer (Patch Merging or Patch Expanding), and the cell and hidden states of the previous time step to extract the spatiotemporal representations. Finally, the reconstruction layer decodes the spatiotemporal representations to generate the next frame.", + "bbox": [ + 75, + 316, + 468, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper can be summarized as follows:", + "bbox": [ + 76, + 559, + 468, + 588 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new recurrent cell, named SwinLSTM (section 3.2), which is able to efficiently extract spatiotemporal representations.", + "- We introduce a new architecture (section 3.1) for spatiotemporal prediction tasks, which can efficiently model spatial and temporal dependencies.", + "- We evaluate the effectiveness of the proposed model on Moving MNIST, TaxiBJ, Human3.6m, and KTH. Experimental results show that SwinLSTM achieves excellent performance on four datasets." + ], + "bbox": [ + 96, + 599, + 467, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 784, + 217, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CNN-Based Models Prior models integrating CNNs and RNNs employ various strategies to better capture spatiotemporal dependencies to improve prediction accuracy. ConvLSTM [24] extends FC-LSTM [26] by replacing fully connected operations with convolutional operations to learn spatiotemporal dependencies. PredRNN [33] proposes the", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spatiotemporal LSTM (ST-LSTM) module, which simultaneously models spatiotemporal information by transferring hidden states in horizontal and vertical directions. PredRNN++ [31] designs a Gradient Highway unit to solve the vanishing gradient problem in PredRNN. E3D-LSTM [32] replaces the 2D convolution in the ST-LSTM module with 3D convolution, allowing the ST-LSTM module to remember more previous information and achieve better prediction performance. MIM [34] replaces the forget gate in the ST-LSTM module with two recurrent units to solve the nonstationary information learning in prediction. CrevNet [39] proposes a CNN-based reversible network to learn complex spatiotemporal dependencies. PhyDNet [8] introduces physical knowledge into a CNN-based model to improve prediction quality. The above models [8, 24, 31-34, 39] improve their ability to capture spatiotemporal dependencies from different perspectives and achieve excellent results. However, despite their widespread usage, convolutional methods have inherent limitations in capturing spatiotemporal dependencies due to their local nature. In order to overcome this challenge, we propose the adoption of a global modeling approach (specifically, Swin Transformer blocks) for learning spatial dependencies, thereby enhancing the model's capacity to capture spatiotemporal dependencies and improve prediction performance.", + "bbox": [ + 496, + 90, + 890, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers The widespread use of Transformer [29] in the field of natural language processing (NLP) led researchers to introduce it to the vision domain. ViT [6] pioneered the direct application of the Transformer architecture to image classification and achieved great success. However, its excellent performance is based on training on large datasets, which leads to its unsatisfactory performance on smaller datasets. To solve this problem, DeiT [27] proposes several training strategies to allow ViT to perform well on the smaller ImageNet-1K [5] dataset. Subsequently, some variants [28,37,41,42] of ViT achieved impressive results on various vision tasks. In particular, Swin Transformer [20] achieves outstanding performance on image classification, semantic segmentation, and object detection tasks due to its shifted window mechanism and hierarchical design. In this paper, we try to integrate the Swin Transformer blocks and the simplified LSTM to form a SwinLSTM recurrent cell, and use it as the core to build a model to capture temporal and spatial dependencies to perform spatiotemporal prediction tasks.", + "bbox": [ + 496, + 489, + 890, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 806, + 589, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 500, + 830, + 694, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The overall predictive architecture is depicted in Figure 2 (b and c). We build our base model and a deeper model, called SwinLSTM-B and SwinLSTM-D, respectively. First,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13471", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/306cca612b7971dc9199bfe31f85b3daaaf2610bb80edf878f0b3318a2575048.jpg", + "image_caption": [ + "(a) SwinLSTM Cell" + ], + "image_footnote": [], + "bbox": [ + 81, + 95, + 465, + 324 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/927ea7a06f3f0a1bc05b227abace829515cdb813ec6194c281b47de3d06d4836.jpg", + "image_caption": [ + "(b) SwinLSTM-B" + ], + "image_footnote": [], + "bbox": [ + 106, + 349, + 472, + 493 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7d2337a89868c5816927b213093ba7d1d06694cb70a16513613652390f90c1f9.jpg", + "image_caption": [ + "(c) SwinLSTM-D" + ], + "image_footnote": [], + "bbox": [ + 540, + 99, + 883, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an image at the time step $t$ is split into non-overlapping patches. The patch size is $P^2$ , and $P$ equals 2 or 4 in our implementation. Therefore, the feature dimension of each patch is $C \\cdot P^2$ ( $C$ denotes the number of channels). Subsequently, the image patches are flattened and fed to the patch embedding layer, which linearly maps the original features of the patches to an arbitrary dimension. Then, for SwinLSTM-B, the SwinLSTM layer receives the transformed image patches, the hidden state $\\mathcal{H}_{t-1}$ , and cell state $\\mathcal{C}_{t-1}$ to generate the hidden state $\\mathcal{H}_t$ and cell state $\\mathcal{C}_t$ , where the $\\mathcal{H}_t$ is duplicated into two copies, one for the reconstruction layer, the other together with $\\mathcal{C}_t$ for the SwinLSTM layer at the next time step. For SwinLSTM-D, we increase the number of SwinLSTM Cells, as well as add Patch Merging and Patch Expanding [2] layers, where the former is utilized for downsampling and the latter for upsampling. $\\mathcal{H}_t^{l=m}$ and $\\mathcal{C}_t^{l=m}$ refer to the hidden state and cell state at time step $t$ from layer $m$ . Finally, the reconstruction layer maps the hidden state $\\mathcal{H}_t$ from the SwinLSTM layer to the input size to get the predicted frame for the next time step.", + "bbox": [ + 76, + 594, + 472, + 898 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3990bddc9e028974e4316785eeb1f685101e9e6bfbdd8c23abf2233d618a3b6d.jpg", + "image_caption": [ + "Figure 2. (a): The detailed structure of the proposed recurrent cell: SwinLSTM. STB and LP denote Swin Transformer blocks and Linear Projection. (b): The architecture of the base model with a single SwinLSTM cell. (c): The architecture of the deep model with multiple SwinLSTM cells.", + "Figure 3. Swin Transformer Block." + ], + "image_footnote": [], + "bbox": [ + 566, + 595, + 820, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both SwinLSTM-B and SwinLSTM-D were trained using the same training approach. The specific training process for SwinLSTM-B is presented in Algorithm 1. During", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13472", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Training pipeline" + ], + "code_body": "1: Require: $\\alpha$ , the learning rate. $m$ , the batch size. $n$ , the number of training epochs. $S$ , the number of frames in the input (target) sequence. \n2: Initialize parameters of network $(\\theta)$ . \n3: for $i = 0, 1, \\ldots, n-1$ do \n4: for number of batches do \n5: Sample $\\{[X_{0:S-1};Y_{0:S-1}]^{(j)}\\}_{j=1}^m$ a batch from the training set. \n6: Phase 1: Warm-up Phase \n7: Initialize $\\mathcal{H}_0, \\mathcal{C}_0$ \n8: for $t = 0, 1, \\ldots, S-2$ do \n9: Get $\\hat{X}_{t+1}, \\mathcal{H}_{t+1}, \\mathcal{C}_{t+1} = P(X_t, \\mathcal{H}_t, \\mathcal{C}_t)$ \n10: end for \n11: Phase 2: Prediction Phase \n12: $\\hat{I}_0 \\gets X_{S-1}, \\mathcal{H}_0 \\gets \\mathcal{H}_{S-1}, \\mathcal{C}_0 \\gets \\mathcal{C}_{S-1}$ \n13: for $t = 0, 1, \\ldots, S-1$ do \n14: Get $\\hat{I}_{t+1}, \\mathcal{H}_{t+1}, \\mathcal{C}_{t+1} = P(\\hat{I}_t, \\mathcal{H}_t, \\mathcal{C}_t)$ \n15: end for \n16: $\\mathcal{L} \\gets \\mathcal{L}^{\\mathrm{p}}([\\hat{X}_{1:S-1};\\hat{I}_{1:S}], [X_{1:S-1};Y_{0:S-1}])$ \n17: Update $\\theta \\gets \\theta - \\alpha \\nabla_\\theta \\mathcal{L}$ \n18: end for \n19: end for", + "bbox": [ + 84, + 109, + 468, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the Warm-up phase, we take the frames of the input sequence as input to the model. However, in the Prediction phase, we use the output of the model at the previous time step as the input to the model at the current time step. In particular, we concatenate the outputs of the Warm-up and Prediction phases to compute the loss. When the prediction loss function $\\mathcal{L}^{\\mathrm{P}}$ is $L_{2}$ , its formula is shown in eq. 1.", + "bbox": [ + 76, + 472, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {p}} = \\left\\| \\left[ \\hat {X} _ {1: S - 1}; \\hat {I} _ {1: S} \\right] - \\left[ X _ {1: S - 1}; Y _ {0: S - 1} \\right] \\right\\| _ {2} ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 609, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the prediction loss function $\\mathcal{L}^{\\mathrm{p}}$ is $L_{1} + L_{2}$ , its formula is shown in eq. 2.", + "bbox": [ + 76, + 642, + 468, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} ^ {\\mathrm {p}} = \\| \\left[ \\hat {X} _ {1: S - 1}; \\hat {I} _ {1: S} \\right] - \\left[ X _ {1: S - 1}; Y _ {0: S - 1} \\right] \\| _ {1} \\tag {2} \\\\ + \\| \\left[ \\hat {X} _ {1: S - 1}; \\hat {I} _ {1: S} \\right] - \\left[ X _ {1: S - 1}; Y _ {0: S - 1} \\right] \\| _ {2} ^ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 686, + 468, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In eq. 1 and eq. 2, $\\hat{X}_{1:S-1}$ represents S-1 predicted frames in the Warm-up phase, $\\hat{I}_{1:S}$ denotes S predicted frames in the Prediction phase, and $X_{1:S-1}$ and $Y_{0:S-1}$ represent S-1 input frames and S ground truth future frames, respectively.", + "bbox": [ + 76, + 742, + 468, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SwinLSTM Module", + "text_level": 1, + "bbox": [ + 76, + 815, + 266, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ConvLSTM [24] introduces the convolution operators into input-to-state and state-to-state transitions to overcome the shortcomings of FC-LSTM [26] in processing spatiotemporal data. The key equations of ConvLSTM are as", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "follows:", + "bbox": [ + 500, + 90, + 557, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} i _ {t} = \\sigma \\left(W _ {x i} * \\mathcal {X} _ {t} + W _ {h i} * \\mathcal {H} _ {t - 1} + b _ {i}\\right) \\\\ f _ {t} = \\sigma \\left(W _ {x f} * \\mathcal {X} _ {t} + W _ {h f} * \\mathcal {H} _ {t - 1} + b _ {f}\\right) \\\\ o _ {t} = \\sigma \\left(W _ {x o} * \\mathcal {X} _ {t} + W _ {h o} * \\mathcal {H} _ {t - 1} + b _ {o}\\right) \\\\ \\mathcal {C} _ {t} = f _ {t} \\circ \\mathcal {C} _ {t - 1} + i _ {t} \\circ \\tanh \\left(W _ {x c} * \\mathcal {X} _ {t} + W _ {h c} * \\mathcal {H} _ {t - 1} + b _ {c}\\right) \\\\ \\mathcal {H} _ {t} = o _ {t} \\circ \\tanh \\left(\\mathcal {C} _ {t}\\right) \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 114, + 890, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $*$ denotes the convolution operator and $\\circ$ denotes the Hadamard product. $i_t, f_t, o_t$ are input, forget, output gate. Different from convolution operators to extract local correlations, the self-attention mechanism captures global spatial dependencies by computing similarity scores across all positions. Thus, we remove all weights $W$ and biases $b$ in eq. 3 to obtain eq. 4:", + "bbox": [ + 496, + 232, + 893, + 339 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ni _ {t} = f _ {t} = o _ {t} = \\sigma \\left(\\mathcal {X} _ {t} + \\mathcal {H} _ {t - 1}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 349, + 772, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {t} = f _ {t} \\circ \\mathcal {C} _ {t - 1} + i _ {t} \\circ \\tanh \\left(\\mathcal {X} _ {t} + \\mathcal {H} _ {t - 1}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 367, + 890, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {t} = o _ {t} \\circ \\operatorname {t a n h} \\left(\\mathcal {C} _ {t}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 387, + 694, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Obviously, $i_t = f_t = o_t$ in eq. 4. Therefore, we fuse the three gates: $i_t, f_t, o_t$ into one gate, named filter gate $\\mathcal{F}_t$ .", + "bbox": [ + 500, + 412, + 888, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The detailed structure of the SwinLSTM Module is presented in Figure 2 (a). In SwinLSTM, the information of cell states $\\mathcal{C}_t$ and hidden states $\\mathcal{H}_t$ is updated horizontally to capture long-term and short-term temporal dependencies. Meanwhile, the Swin Transformer blocks vertically learn global spatial dependencies. The key equations of SwinLSTM are shown in eq. 5, where STB denotes the Swin Transformer blocks and LP denotes the Linear Projection:", + "bbox": [ + 496, + 443, + 890, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {t} = \\sigma \\left(\\text {S T B} \\left(\\text {L P} \\left(\\mathcal {X} _ {t}; \\mathcal {H} _ {t - 1}\\right)\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 574, + 733, + 590 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {t} = \\mathcal {F} _ {t} \\circ (\\operatorname {t a n h} (\\operatorname {S T B} (\\operatorname {L P} (\\mathcal {X} _ {t}; \\mathcal {H} _ {t - 1}))) + \\mathcal {C} _ {t - 1}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 593, + 890, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {t} = \\mathcal {F} _ {t} \\circ \\operatorname {t a n h} \\left(\\mathcal {C} _ {t}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 612, + 658, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Swin Transformer Block", + "text_level": 1, + "bbox": [ + 500, + 638, + 723, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The global multi-head self-attention (MSA) mechanism in the Vision Transformer [6] computes the relationship between a token and all other tokens. Such a computation leads to quadratic computational complexity related to the number of tokens, which is not friendly to dense prediction tasks. To alleviate the computational burden of global self-attention and improve the modeling power, Swin Transformer [20] proposes window-based multi-head self-attention (W-MSA) and shifted-window-based multi-head self-attention (SW-MSA).", + "bbox": [ + 496, + 662, + 890, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} ^ {l} = \\operatorname {W - M S A} \\left(\\ln \\left(\\mathbf {z} ^ {l - 1}\\right)\\right) + \\mathbf {z} ^ {l - 1}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 821, + 812, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {l} = \\operatorname {M L P} \\left(\\ln \\left(\\hat {\\mathbf {z}} ^ {l}\\right)\\right) + \\hat {\\mathbf {z}} ^ {l} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 843, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} ^ {l + 1} = \\operatorname {S W - M S A} \\left(\\operatorname {L N} \\left(\\mathbf {z} ^ {l}\\right)\\right) + \\mathbf {z} ^ {l}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 863, + 805, + 883 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {l + 1} = \\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\hat {\\mathbf {z}} ^ {l + 1}\\right)\\right) + \\hat {\\mathbf {z}} ^ {l + 1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 883, + 805, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13473", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0e125cb317286c322cbb1eadf056c58f36d4f864f9d408c9f25e6020a00559b3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelSwinLSTMsSTBPatch sizeResolutionTrainTest
Moving MNISTSwinLSTM-D4(2, 6, 6, 2)2(64, 64, 1)10 → 1010 → 10
Human3.6mSwinLSTM-B1122(128, 128, 3)4 → 44 → 4
KTHSwinLSTM-B164(128, 128, 1)10 → 1010 → 20/40
TaxiBJSwinLSTM-B1124(32, 32, 2)4 → 44 → 4
", + "bbox": [ + 88, + 88, + 883, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 3, the first Swin Transformer block is based on W-MSA, consisting of two LayerNorm layers and a 2-layer MLP with GELU non-linearity. A LayerNorm layer and a residual connection are applied before and after each W-MSA module and MLP module, respectively. The second Swin Transformer block is the same as the first except that the W-MSA block in it is replaced by an SW-MSA block. The key equations of Swin Transformer blocks are shown in eq. 6.", + "bbox": [ + 75, + 258, + 472, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 410, + 210, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our proposed model on four commonly used datasets: Moving MNIST [26], TaxiBJ [40], Human3.6m [10], and KTH [23], and present quantitative comparison results and visual examples for each dataset, illustrating the effectiveness and generalizability of our proposed neural network. In addition, we perform ablation studies (section 4.5) and feature map visualization (section 4.6).", + "bbox": [ + 75, + 436, + 468, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementations", + "text_level": 1, + "bbox": [ + 76, + 566, + 243, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $L_{2}$ loss for the MovingMNIST dataset, $L_{1} + L_{2}$ loss for the Human3.6m, the TaxiBJ, and the KTH datasets. The input and target frames of each dataset are normalized to the intensity of [0, 1]. Each model is optimized with the Adam optimizer [15]. We train our model on an Nvidia RTX A5000 GPU. More detailed parameters for each dataset are listed in Table 1.", + "bbox": [ + 75, + 590, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 705, + 261, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt the Mean Squared Error (MSE), the Mean Absolute Error (MAE), the Peak Signal to Noise Ratio (PSNR), and the Structural Similarity Index Measure (SSIM) [35] as metrics to evaluate the quality of predictions. All metrics are averaged over predicted frames. Lower MAE and MSE or higher SSIM and PSNR indicate better prediction accuracy.", + "bbox": [ + 75, + 729, + 468, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Datasets", + "text_level": 1, + "bbox": [ + 76, + 845, + 181, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moving MNIST. The moving MNIST dataset [26] is a widely used synthetic dataset. In our experiments, the", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/65da243f3c753efe876d49deb839d6ff451d2c12b6c963bacb20167aa20e6cd1.jpg", + "table_caption": [ + "Table 1. Experimental setup. SwinLSTMs denotes the number of the SwinLSTM Cells in predictive network. STB denotes the number of the Swin Transformer blocks in SwinLSTM cell. Patch size indicates the patch token size. Train and Test represent the number of input and output frames during training and testing." + ], + "table_footnote": [], + "table_body": "
MethodMSE ↓SSIM ↑
ConvLSTM [24]103.30.707
DFN [11]89.00.726
FRNN [22]69.70.813
VPN [14]64.10.870
PredRNN [33]56.80.867
CausalLSTM [31]46.50.898
MIM [34]44.20.910
E3D-LSTM [32]41.30.910
LMC [19]41.50.924
MAU [3]27.60.937
PhyDNet [8]24.40.947
CrevNet [39]22.30.949
SwinLSTM17.70.962
", + "bbox": [ + 550, + 256, + 844, + 491 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f17749da804411e1d7a2c379fc8fce8e7dddb77d0e5853eef218b84594227ea0.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of SwinLSTM and other methods on Moving MNIST. Each model observes 10 frames and predicts the subsequent 10 frames. Lower MSE and higher SSIM indicate better predictions." + ], + "table_footnote": [], + "table_body": "
MethodMAE /100 ↓MSE /10 ↓SSIM ↑
FRNN [22]19.049.80.771
ConvLSTM [24]18.950.40.776
PredRNN [33]18.948.40.781
CausalLSTM [31]17.245.80.851
MIM [34]17.842.90.790
E3D-LSTM [32]16.646.40.869
PhyDNet [8]16.236.90.901
SwinLSTM11.933.20.913
", + "bbox": [ + 504, + 575, + 888, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Quantitative evaluation of different models on the Human3.6m dataset.", + "bbox": [ + 498, + 734, + 893, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "method for generating moving MNIST sequences follows [26], each sequence contains 20 frames, and the first 10 frames and the last 10 frames are used for input and target, respectively. The two handwritten digits in each frame are randomly sampled in the MNIST dataset [17], moving and bouncing around a $64 \\times 64$ pixel black canva with a fixed speed and angle. We can get an infinite number of se", + "bbox": [ + 498, + 794, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13474", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2c6d5ab6851191cf91ee23c8a403ede3eb6cf8a63f442171d6e1ee4899893af0.jpg", + "image_caption": [ + "(a) Moving MNIST" + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 553, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf9d06cb8607db915681d9aa75d73ae17745b33c0c0c2407efd6c5e85b2c1459.jpg", + "image_caption": [ + "(b) TaxiBJ" + ], + "image_footnote": [], + "bbox": [ + 578, + 88, + 885, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/92ee2b4a05418d2827fe2067bae898eb17cd26a8b87d48011f53c7d277b6ba9c.jpg", + "image_caption": [ + "(c) KTH (Jogging $10\\rightarrow 20$" + ], + "image_footnote": [], + "bbox": [ + 84, + 311, + 547, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95fe46f60339601a28abb7b13977f35de90e2a06f63ccc676bfb5aecb8e009da.jpg", + "image_caption": [ + "(d) Human3.6m" + ], + "image_footnote": [], + "bbox": [ + 576, + 311, + 874, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cff07aa7e7fd1d7d9d6de58bbd6539daabd873d3eea2f0ef5ce485dc2e04f234.jpg", + "image_caption": [ + "(e) KTH (Walking $10\\rightarrow 40$", + "Figure 4. Qualitative results of SwinLSTM on four datasets. The first line is the input, the second line is the ground truth, and the third line is the prediction of SwinLSTM. For Moving MNIST, we add the prediction results of ConvLSTM as the fourth line. For TaxiBJ, we add the absolute difference between the predictions and ground truths as the fourth line. For KTH action, (c) is the jogging action when the model predicts 20 frames based on the previous 10 frames, and (e) is the walking action when the model observes 10 frames and predicts the next 40 frames." + ], + "image_footnote": [], + "bbox": [ + 86, + 494, + 875, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quences by applying different speeds and angles to different handwritten digits. We adopt 10000 sequences for training and a fixed set containing 10000 sequences for testing.", + "bbox": [ + 75, + 768, + 470, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human3.6m. The human3.6m dataset [10] contains 3.6 million different human poses and corresponding images. Following [34], we only use the 'walking' scenario. Subjects S1, S5, S6, S7, S8 are used for training, and S9, S11 are used for testing. The images in the dataset are resized", + "bbox": [ + 75, + 824, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "from $1000 \\times 1000 \\times 3$ to $128 \\times 128 \\times 3$ . We train the models to predict the next 4 RGB frames from 4 observations.", + "bbox": [ + 496, + 768, + 890, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "KTH. The KTH dataset [23] contains 25 individuals performing 6 categories of human actions (walking, jogging, running, boxing, hand waving, and hand clapping) in 4 different scenarios. We follow the experimental setup in [30], resizing each image to $128 \\times 128$ and using persons 1-16 for training and 17-25 for testing. The models predict 10", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13475", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff62ff564e7132fff56175e92d510bb9f842f2e6bcbe0b42911dd9114fe7c16f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodKTH (10 → 20)KTH (10 → 40)
SSIM ↑PSNR ↑SSIM ↑PSNR ↑
ConvLSTM [24]0.71223.580.63922.85
SAVP [7]0.74625.380.70123.97
FRNN [22]0.77126.120.67823.77
DFN [11]0.79427.260.65223.01
PredRNN [33]0.83927.550.70324.16
VarNet [13]0.84328.480.73925.37
SVAP-VAE [18]0.85227.770.81126.18
PredRNN++ [31]0.86528.470.74125.21
E3d-LSTM [32]0.87929.310.81027.24
STMFANet [12]0.89329.850.85127.56
SwinLSTM0.90334.340.87933.15
", + "bbox": [ + 81, + 88, + 467, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames from 10 observations at training time and 20 or 40 frames at inference time.", + "bbox": [ + 75, + 369, + 468, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TaxiBJ. TaxiBJ [40] contains complex real-world taxi trajectory data collected from taxicab GPS monitors in Beijing. Each frame in TaxiBJ is a $32 \\times 32 \\times 2$ heat map, where the last dimension represents the flow of traffic entering and leaving the same area. We follow the experimental setup in [40] and use the last four weeks of data for testing and the rest for training. We use 4 observations to predict the next 4 consecutive frames. We adopted per-frame MSE as metrics.", + "bbox": [ + 73, + 398, + 468, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Main results", + "text_level": 1, + "bbox": [ + 76, + 545, + 210, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2, 3, 4 and 5 present quantitative comparisons with previous state-of-the-art models on four datasets. On Moving MNIST, SwinLSTM outperforms previous CNN-based models and achieves the new state-of-the-art results. Specifically, compared to ConvLSTM [24], SwinLSTM reduces the MSE from 103.3 to 17.7 and increases the SSIM from 0.707 to 0.962. On both Human3.6m and KTH, SwinLSTM achieves the state-of-the-art, especially with huge gains of 4.49 (10→20) and 5.59 (10→40) in PSNR on KTH. In the case of TaxiBJ, our proposed SwinLSTM achieved a remarkable reduction in MSE for each predicted frame compared to previous models. Additionally, the interframe MSE difference was smaller than that of the compared models. The results indicate SwinLSTM's great potential for an efficient and generalizable technique for spatiotemporal prediction.", + "bbox": [ + 73, + 568, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 4, we present qualitative results of SwinLSTM on all datasets. (a) presents a qualitative analysis of SwinLSTM and ConvLSTM on an extremely hard case of Moving MNIST. In the presented images, two digits are continuously intertwined, posing a challenge to the model's ability to accurately predict their motion trajectories and ap", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bbc049a94613f712675c108cad2926a1e2d0dc9093025c7f65c89ab8bc9954b7.jpg", + "table_caption": [ + "Table 4. Quantitative evaluation on the KTH test set. We present the model observing 10 frames to predict 20 or 40 frames, and all metrics are averaged over the predicted frames. Higher SSIM and PSNR indicate better prediction quality." + ], + "table_footnote": [], + "table_body": "
MethodFrame1 ↓Frame2 ↓Frame3 ↓Frame4 ↓
VPN [14]0.7441.0311.2511.444
ST-ResNet [40]0.6880.9391.1301.288
FRNN [22]0.6820.8230.9891.183
PredRNN [33]0.6340.9341.0471.263
PredRNN++ [31]0.6410.8550.9791.158
E3d-LSTM [32]0.6200.7730.8881.158
MIM [34]0.5540.7370.8870.999
SwinLSTM0.3240.4010.4730.525
", + "bbox": [ + 504, + 88, + 890, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Quantitative results on the TaxiBJ dataset. We report the per-frame MSE of the 4 predictions.", + "bbox": [ + 498, + 232, + 890, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pearances. As demonstrated, both models effectively capture the trajectory of the digits. However, in terms of appearance, ConvLSTM's prediction results become increasingly blurred as the number of time steps increases, while SwinLSTM's predictions maintain high similarity throughout. SwinLSTM, as an extension of ConvLSTM, utilizes LSTM [9] to capture temporal dependencies, and employs Swin Transformer blocks [20] to model global spatial information. In contrast, ConvLSTM uses convolution to model local spatial information. The significant performance gap between the two models demonstrates that learning global spatial information can better assist the model in capturing spatiotemporal dependencies and thus improve prediction accuracy.", + "bbox": [ + 496, + 289, + 890, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To facilitate the observation of the similarity between the predicted results and the ground truth in (b), we visualize the absolute difference between them. For (c), (d), and (e), we observe that SwinLSTM makes accurate predictions of human actions and scenes, where (c) and (e) show visual examples of 2 different categories of actions predicted by SwinLSTM, Jogging and walking respectively. The model receives 10 frames during training and predicts 10 frames, whereas it predicts 20 or 40 frames during testing. Expanding the predicted number of frames during testing poses a significant challenge to the model, requiring it to effectively learn spatiotemporal dependencies from the spatiotemporal data. Our experiments on the KTH dataset demonstrate that SwinLSTM outperforms alternative models in capturing spatiotemporal dependencies.", + "bbox": [ + 496, + 501, + 892, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 739, + 653, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform ablation studies on TaxiBJ and Human3.6m to analyze the impact of different elements on model performance. Specifically, we discuss three major elements: the reconstruction layer, patch size, and the number of Swin Transformer blocks.", + "bbox": [ + 496, + 763, + 890, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The reconstruction layer. The role of the reconstruction layer is to decode the spatiotemporal representations extracted by the SwinLSTM cell. We conduct experiments on transposed convolution, bilinear interpolation, and lin", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13476", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ear projection to analyze the impact of different decoding methods. Table 6 shows that the transposed convolution performs much better than the other two methods.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/99e5949812878446f56a2371a0fce273024e5b07d76c2a8592c9af73f2b2642d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
The reconstruction layerTaxiBJHuman3.6m
MSESSIMMSE /10SSIM
Transposed convolution0.3900.98033.20.913
Bilinear interpolation0.7940.96139.30.887
Linear Projection0.4150.97972.50.773
", + "bbox": [ + 81, + 152, + 465, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Patch size. The size of the image patch directly determines the length of the input token sequence. The smaller the size, the longer the length. To explore the effect of different patch sizes on prediction accuracy, we conduct experiments with patch sizes set to 2, 4, and 8 on TaxiBJ and Human3.6m. Table 7 shows that the optimal patch size varies across datasets, which indicates that adjusting the patch size to an appropriate size is beneficial to improve the generalization of SwinLSTM.", + "bbox": [ + 75, + 290, + 468, + 425 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/39193643c34aa768961dc60e961402c8e3036d86e996648346dd53b0cbd47f72.jpg", + "table_caption": [ + "Table 6. Ablation study on the three different methods of reconstruction layer with SwinLSTM on TaxiBJ and Human3.6m." + ], + "table_footnote": [], + "table_body": "
Patch sizeTaxiBJHuman3.6m
MSESSIMMSE /10SSIM
20.4170.97933.20.913
40.3900.98040.20.885
80.3930.98051.90.839
", + "bbox": [ + 98, + 439, + 449, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study on different patch sizes on TaxiBJ and Human3.6m.", + "bbox": [ + 76, + 542, + 468, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The number of Swin Transformer Blocks. Based on the mechanism of Swin Transformer blocks (STB) modeling global spatial information, it is evident that the learning capacity of the model varies with the number of STB utilized. We explored the effect of the number of STB from 2 to 18 on model performance. Figure 5 present the results of different numbers of STB on TaxiBJ and Human3.6m.", + "bbox": [ + 75, + 594, + 468, + 700 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2bbfb7588ca692fef4ac54889a5d9903fef69d9834b18da4c5c0170f96cfce41.jpg", + "image_caption": [ + "(a) TaxiBJ", + "Figure 5. Ablation study on the different numbers of STB with SwinLSTM on TaxiBJ and Human3.6m." + ], + "image_footnote": [], + "bbox": [ + 58, + 715, + 263, + 834 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c077b8cec9c4fe7a472412c80103501ec376f5b10eaabda9fcc501e709bfb8a.jpg", + "image_caption": [ + "(b) Human3.6m" + ], + "image_footnote": [], + "bbox": [ + 267, + 715, + 472, + 834 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/425f266a1907d3cec76545d82c24edd699af5f0a0814d732a51990a8b5545d90.jpg", + "image_caption": [ + "Figure 6. Feature map visualization of a sample in the Moving MNIST test set. Hidden States and Cell States are the feature maps of $\\mathcal{H}_t$ and $\\mathcal{C}_t$ of the recurrent cell SwinLSTM (section 3.2) at different time steps. STB-2,STB-4, and STB-6 are the feature maps of the 2nd, 4th, and 6th layers of the Swin Transformer blocks (section 3.3), respectively." + ], + "image_footnote": [], + "bbox": [ + 519, + 93, + 874, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Feature map Visualization", + "text_level": 1, + "bbox": [ + 500, + 609, + 736, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To analyze the mechanism of the SwinLSTM cell (section 3.2), we randomly selected a sample from the Moving MNIST test set and visualized the feature maps during inference. Figure 6 shows the results, where white arrows have been added to the images in the third row to facilitate the observation of digit trajectories. The Hidden States capture the position of digits at the current time step, whereas the Cell States memorize the digit trajectory. Meanwhile, we conducted a visualization of the feature maps generated by the second, fourth, and sixth Swin Transformer blocks of the SwinLSTM-B. By observing the feature maps of STB-2 to STB-6, it is evident that STB gradually learns global spatial correlations as the number of interactions of window self-attention increases (section 3.3). By analyzing the visualizations generated by our model, we aimed to gain valuable insights into its performance and behavior. These insights will inform future optimizations and improvements.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13477", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 89, + 194, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose SwinLSTM, a new recurrent cell that incorporates Swin Transformer blocks and simplified LSTM, and constructs a predictive network for spatiotemporal prediction tasks based on it. Through extensive experimentation, it has been demonstrated that the proposed method, SwinLSTM, achieves excellent performance on various datasets, including Moving MNIST, TaxiBJ, Human3.6m, and KTH. These impressive results showcase the effectiveness and generalization ability of SwinLSTM. Notably, SwinLSTM outperforms its predecessor, ConvLSTM, in prediction accuracy, which suggests that learning global spatial dependencies allows the model to better capture spatiotemporal dependencies. Overall, this study highlights the potential of SwinLSTM as a promising approach for spatiotemporal modeling.", + "bbox": [ + 75, + 114, + 472, + 344 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 357, + 236, + 376 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Natural Science Foundation Innovation Research Team Project of Hainan Province (320CXTD431), the National Key Research and Development Program of China (2021YFB1507104), the National Natural Science Foundation of China (32060413), and the Major Science and Technology Project of Hainan Province (ZDKJ2020013), which are gratefully acknowledged.", + "bbox": [ + 75, + 383, + 468, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 505, + 174, + 520 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Apratim Bhattacharyya, Mario Fritz, and Bernt Schiele. Long-term on-board prediction of people in traffic scenes under uncertainty. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4194-4202, 2018. 1", + "[2] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 205-218. Springer, 2023. 3", + "[3] Zheng Chang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, Yan Ye, Xiang Xinguang, and Wen Gao. Mau: A motion-aware unit for video prediction and beyond. Advances in Neural Information Processing Systems, 34:26950-26962, 2021. 1, 5", + "[4] Yunpeng Chen, Yannis Kalantidis, Jianshu Li, Shuicheng Yan, and Jiashi Feng. A^2-nets: Double attention networks. Advances in neural information processing systems, 31, 2018. 1", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 86, + 530, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.2,4", + "[7] Hang Gao, Huazhe Xu, Qi-Zhi Cai, Ruth Wang, Fisher Yu, and Trevor Darrell. Disentangling propagation and generation for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9006-9015, 2019. 7", + "[8] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11474-11484, 2020. 1, 2, 5", + "[9] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 7", + "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 5, 6", + "[11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 5, 7", + "[12] Beibei Jin, Yu Hu, Qiankun Tang, Jingyu Niu, Zhiping Shi, Yinhe Han, and Xiaowei Li. Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4554-4563, 2020. 7", + "[13] Beibei Jin, Yu Hu, Yiming Zeng, Qiankun Tang, Shice Liu, and Jing Ye. Varnet: Exploring variations for unsupervised video prediction. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5801-5806. IEEE, 2018. 7", + "[14] Nal Kalchbrenner, Aäron Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyls, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. In International Conference on Machine Learning, pages 1771-1779. PMLR, 2017. 5, 7", + "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 5", + "[16] Yong-Hoon Kwon and Min-Gyu Park. Predicting future frames using retrospective cycle gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1811-1820, 2019. 1", + "[17] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 5", + "[18] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 7", + "[19] Sangmin Lee, Hak Gu Kim, Dae Hwi Choi, Hyung-II Kim, and Yong Man Ro. Video prediction recalling long-term motion context via memory alignment learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3054-3063, 2021. 1, 5", + "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13478", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 4, 7", + "[21] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. Advances in neural information processing systems, 29, 2016. 1", + "[22] Marc Oliu, Javier Selva, and Sergio Escalera. Folded recurrent neural networks for future video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 716-731, 2018. 5, 7", + "[23] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a localsvm approach. In Proceedings of the 17th International Conference on Pattern Recognition, 2004.ICPR 2004., volume 3,pages 32-36.IEEE,2004.5,6", + "[24] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 1, 2, 4, 5, 7", + "[25] Xingjian Shi, Zhihan Gao, Leonard Lausen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Deep learning for precipitation nowcasting: A benchmark and a new model. Advances in neural information processing systems, 30, 2017. 1", + "[26] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 1, 2, 4, 5", + "[27] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2", + "[28] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 2", + "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[30] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017. 6", + "[31] Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and S Yu Philip. Predrnn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning. In International Conference on Machine Learning, pages 5123-5132. PMLR, 2018. 1, 2, 5, 7", + "[32] Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, and Li Fei-Fei. Eidetic 3d LSTM: A model for video prediction and beyond. In International conference on learning representations, 2018. 1, 2, 5, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and Philip S Yu. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in neural information processing systems, 30, 2017. 1, 2, 5, 7", + "[34] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154-9162, 2019. 1, 2, 5, 6, 7", + "[35] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[36] Yu-Huan Wu, Yun Liu, Xin Zhan, and Ming-Ming Cheng. P2t: Pyramid pooling transformer for scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[37] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2", + "[38] Ziru Xu, Yunbo Wang, Mingsheng Long, Jianmin Wang, and M KLiss. Predcnn: Predictive learning with cascade convolutions. In IJCAI, pages 2940-2947, 2018. 1", + "[39] Wei Yu, Yichao Lu, Steve Easterbrook, and Sanja Fidler. Efficient and information-preserving future frame prediction and beyond. 2020. 1, 2, 5", + "[40] Junbo Zhang, Yu Zheng, and Dekang Qi. Deep spatiotemporal residual networks for citywide crowd flows prediction. In Thirty-first AAAI conference on artificial intelligence, 2017. 1, 5, 7", + "[41] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3008, 2021. 2", + "[42] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 2", + "[43] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 795 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13479", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/0bea32db-db74-4a9f-85a7-6fe6526e5e78_model.json b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/0bea32db-db74-4a9f-85a7-6fe6526e5e78_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a05ebc798c2cddb47fbaed77d1503e8a77e2bf24 --- /dev/null +++ b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/0bea32db-db74-4a9f-85a7-6fe6526e5e78_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.13, + 0.85, + 0.175 + ], + "angle": 0, + "content": "SwinLSTM: Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.203, + 0.744, + 0.259 + ], + "angle": 0, + "content": "Song Tang* Chuang Li* Pu Zhang RongNian Tang† Hainan University {songtang, lc, zhangpu, rn.tang}@hainanu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.657 + ], + "angle": 0, + "content": "Integrating CNNs and RNNs to capture spatiotemporal dependencies is a prevalent strategy for spatiotemporal prediction tasks. However, the property of CNNs to learn local spatial information decreases their efficiency in capturing spatiotemporal dependencies, thereby limiting their prediction accuracy. In this paper, we propose a new recurrent cell, SwinLSTM, which integrates Swin Transformer blocks and the simplified LSTM, an extension that replaces the convolutional structure in ConvLSTM with the self-attention mechanism. Furthermore, we construct a network with SwinLSTM cell as the core for spatiotemporal prediction. Without using unique tricks, SwinLSTM outperforms state-of-the-art methods on Moving MNIST, Human3.6m, TaxiBJ, and KTH datasets. In particular, it exhibits a significant improvement in prediction accuracy compared to ConvLSTM. Our competitive experimental results demonstrate that learning global spatial dependencies is more advantageous for models to capture spatiotemporal dependencies. We hope that SwinLSTM can serve as a solid baseline to promote the advancement of spatiotemporal prediction accuracy. The codes are publicly available at https://github.com/SongTang-x/SwinLSTM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.685, + 0.21, + 0.701 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.47, + 0.878 + ], + "angle": 0, + "content": "Spatiotemporal prediction has received increasing attention in recent years due to it can benefit many practical applications, e.g., precipitation forecasting [24, 25, 33], autonomous driving [1, 16], and traffic flow prediction [38, 40]. However, the complex physical dynamics and chaotic properties of spatiotemporal predictive learning make it challenging for purely data-driven deep learning methods to make accurate predictions. Existing methods [3, 8, 19, 24, 31-34, 39] integrate CNNs and RNNs to learn spatiotemporal dependencies in spatiotemporal data to improve prediction accuracy. To capture the spatial and temporal depen" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.304, + 0.875, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.562, + 0.895, + 0.633 + ], + "angle": 0, + "content": "Figure 1. Comparison of input-to-state and state-to-state transitions in three different recurrent cells. (a) FC-LSTM utilizes linear operations to process 1D vectors. (b) ConvLSTM employs 2D convolutions to process 3D tensors. (c) The proposed SwinLSTM leverages self-attention to process 2D matrices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "dencies simultaneously, ConvLSTM [24] extends the fully connected LSTM (FC-LSTM) [26] by replacing linear operations with convolutional operations. Subsequently, several variants of ConvLSTM are proposed. PredRNN [33] and MIM [34] modify the internal structure of the LSTM unit. E3D-LSTM [32] integrates 3D-Convs into LSTMs. PhyD-Net [8] leverages a CNN-based module to disentangle physical dynamics. Admittedly, these methods achieve impressive results on spatiotemporal prediction tasks. However, convolution operators focus on capturing local features and relations and are inefficient for modeling global spatial information [4]. Although the receptive field can be enlarged by stacking convolution layers, the effective receptive field only reaches a fraction of the theoretical receptive field in practice [21]. Therefore, it can be inferred that the CNN-based models may prove to be ineffective in capturing spa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.338, + 0.902 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13470" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "tiotemporal dependencies, owing to the inherent locality of the CNN architecture, leading to restricted accuracy in predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.317 + ], + "angle": 0, + "content": "Recently, the Vision Transformer (ViT) [6] introduced the Transformer [29] that directly models long-range dependencies into the vision domain. ViT applies a standard Transformer for image classification and attains excellent results with sufficient data. Its outstanding achievement has attracted more researchers to apply Transformer to computer vision. Subsequently, some variants [20, 27, 36, 43] of ViT emerged, with tremendous success on different vision tasks. Notably, the Swin Transformer [20] has exhibited exceptional performance across diverse visual tasks, owing to its unique utilization of local attention and shift window mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.47, + 0.558 + ], + "angle": 0, + "content": "Inspired by this, we propose SwinLSTM, a new recurrent cell. Specifically, we integrate Swin Transformer blocks and a simplified LSTM module to extract spatiotemporal representations. In addition, we construct a predictive network with SwinLSTM as the core to capture spatial and temporal dependencies for spatiotemporal prediction tasks. As illustrated in Figure 2 (c), we first split an input image at the current time step into a sequence of image patches. Subsequently, the flattened image patches are fed to the patch embedding layer. Then, the SwinLSTM layer receives the transformed patches or the hidden states transformed by the previous layer (Patch Merging or Patch Expanding), and the cell and hidden states of the previous time step to extract the spatiotemporal representations. Finally, the reconstruction layer decodes the spatiotemporal representations to generate the next frame." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.56, + 0.47, + 0.589 + ], + "angle": 0, + "content": "The contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.6, + 0.468, + 0.644 + ], + "angle": 0, + "content": "- We propose a new recurrent cell, named SwinLSTM (section 3.2), which is able to efficiently extract spatiotemporal representations." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.655, + 0.468, + 0.7 + ], + "angle": 0, + "content": "- We introduce a new architecture (section 3.1) for spatiotemporal prediction tasks, which can efficiently model spatial and temporal dependencies." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.711, + 0.468, + 0.77 + ], + "angle": 0, + "content": "- We evaluate the effectiveness of the proposed model on Moving MNIST, TaxiBJ, Human3.6m, and KTH. Experimental results show that SwinLSTM achieves excellent performance on four datasets." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.6, + 0.468, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.218, + 0.8 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "CNN-Based Models Prior models integrating CNNs and RNNs employ various strategies to better capture spatiotemporal dependencies to improve prediction accuracy. ConvLSTM [24] extends FC-LSTM [26] by replacing fully connected operations with convolutional operations to learn spatiotemporal dependencies. PredRNN [33] proposes the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Spatiotemporal LSTM (ST-LSTM) module, which simultaneously models spatiotemporal information by transferring hidden states in horizontal and vertical directions. PredRNN++ [31] designs a Gradient Highway unit to solve the vanishing gradient problem in PredRNN. E3D-LSTM [32] replaces the 2D convolution in the ST-LSTM module with 3D convolution, allowing the ST-LSTM module to remember more previous information and achieve better prediction performance. MIM [34] replaces the forget gate in the ST-LSTM module with two recurrent units to solve the nonstationary information learning in prediction. CrevNet [39] proposes a CNN-based reversible network to learn complex spatiotemporal dependencies. PhyDNet [8] introduces physical knowledge into a CNN-based model to improve prediction quality. The above models [8, 24, 31-34, 39] improve their ability to capture spatiotemporal dependencies from different perspectives and achieve excellent results. However, despite their widespread usage, convolutional methods have inherent limitations in capturing spatiotemporal dependencies due to their local nature. In order to overcome this challenge, we propose the adoption of a global modeling approach (specifically, Swin Transformer blocks) for learning spatial dependencies, thereby enhancing the model's capacity to capture spatiotemporal dependencies and improve prediction performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Vision Transformers The widespread use of Transformer [29] in the field of natural language processing (NLP) led researchers to introduce it to the vision domain. ViT [6] pioneered the direct application of the Transformer architecture to image classification and achieved great success. However, its excellent performance is based on training on large datasets, which leads to its unsatisfactory performance on smaller datasets. To solve this problem, DeiT [27] proposes several training strategies to allow ViT to perform well on the smaller ImageNet-1K [5] dataset. Subsequently, some variants [28,37,41,42] of ViT achieved impressive results on various vision tasks. In particular, Swin Transformer [20] achieves outstanding performance on image classification, semantic segmentation, and object detection tasks due to its shifted window mechanism and hierarchical design. In this paper, we try to integrate the Swin Transformer blocks and the simplified LSTM to form a SwinLSTM recurrent cell, and use it as the core to build a model to capture temporal and spatial dependencies to perform spatiotemporal prediction tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.591, + 0.822 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.695, + 0.847 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The overall predictive architecture is depicted in Figure 2 (b and c). We build our base model and a deeper model, called SwinLSTM-B and SwinLSTM-D, respectively. First," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13471" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.096, + 0.466, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.328, + 0.337, + 0.342 + ], + "angle": 0, + "content": "(a) SwinLSTM Cell" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.351, + 0.473, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.497, + 0.315, + 0.51 + ], + "angle": 0, + "content": "(b) SwinLSTM-B" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.1, + 0.885, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.494, + 0.766, + 0.508 + ], + "angle": 0, + "content": "(c) SwinLSTM-D" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.526, + 0.894, + 0.568 + ], + "angle": 0, + "content": "Figure 2. (a): The detailed structure of the proposed recurrent cell: SwinLSTM. STB and LP denote Swin Transformer blocks and Linear Projection. (b): The architecture of the base model with a single SwinLSTM cell. (c): The architecture of the deep model with multiple SwinLSTM cells." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.595, + 0.473, + 0.899 + ], + "angle": 0, + "content": "an image at the time step \\( t \\) is split into non-overlapping patches. The patch size is \\( P^2 \\), and \\( P \\) equals 2 or 4 in our implementation. Therefore, the feature dimension of each patch is \\( C \\cdot P^2 \\) (\\( C \\) denotes the number of channels). Subsequently, the image patches are flattened and fed to the patch embedding layer, which linearly maps the original features of the patches to an arbitrary dimension. Then, for SwinLSTM-B, the SwinLSTM layer receives the transformed image patches, the hidden state \\( \\mathcal{H}_{t-1} \\), and cell state \\( \\mathcal{C}_{t-1} \\) to generate the hidden state \\( \\mathcal{H}_t \\) and cell state \\( \\mathcal{C}_t \\), where the \\( \\mathcal{H}_t \\) is duplicated into two copies, one for the reconstruction layer, the other together with \\( \\mathcal{C}_t \\) for the SwinLSTM layer at the next time step. For SwinLSTM-D, we increase the number of SwinLSTM Cells, as well as add Patch Merging and Patch Expanding [2] layers, where the former is utilized for downsampling and the latter for upsampling. \\( \\mathcal{H}_t^{l=m} \\) and \\( \\mathcal{C}_t^{l=m} \\) refer to the hidden state and cell state at time step \\( t \\) from layer \\( m \\). Finally, the reconstruction layer maps the hidden state \\( \\mathcal{H}_t \\) from the SwinLSTM layer to the input size to get the predicted frame for the next time step." + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.597, + 0.821, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.809, + 0.802, + 0.824 + ], + "angle": 0, + "content": "Figure 3. Swin Transformer Block." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Both SwinLSTM-B and SwinLSTM-D were trained using the same training approach. The specific training process for SwinLSTM-B is presented in Algorithm 1. During" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13472" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.08, + 0.092, + 0.284, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Training pipeline" + }, + { + "type": "algorithm", + "bbox": [ + 0.085, + 0.111, + 0.47, + 0.445 + ], + "angle": 0, + "content": "1: Require: \\(\\alpha\\), the learning rate. \\(m\\), the batch size. \\(n\\), the number of training epochs. \\(S\\), the number of frames in the input (target) sequence. \n2: Initialize parameters of network \\((\\theta)\\). \n3: for \\(i = 0, 1, \\ldots, n-1\\) do \n4: for number of batches do \n5: Sample \\(\\{[X_{0:S-1};Y_{0:S-1}]^{(j)}\\}_{j=1}^m\\) a batch from the training set. \n6: Phase 1: Warm-up Phase \n7: Initialize \\(\\mathcal{H}_0, \\mathcal{C}_0\\) \n8: for \\(t = 0, 1, \\ldots, S-2\\) do \n9: Get \\(\\hat{X}_{t+1}, \\mathcal{H}_{t+1}, \\mathcal{C}_{t+1} = P(X_t, \\mathcal{H}_t, \\mathcal{C}_t)\\) \n10: end for \n11: Phase 2: Prediction Phase \n12: \\(\\hat{I}_0 \\gets X_{S-1}, \\mathcal{H}_0 \\gets \\mathcal{H}_{S-1}, \\mathcal{C}_0 \\gets \\mathcal{C}_{S-1}\\) \n13: for \\(t = 0, 1, \\ldots, S-1\\) do \n14: Get \\(\\hat{I}_{t+1}, \\mathcal{H}_{t+1}, \\mathcal{C}_{t+1} = P(\\hat{I}_t, \\mathcal{H}_t, \\mathcal{C}_t)\\) \n15: end for \n16: \\(\\mathcal{L} \\gets \\mathcal{L}^{\\mathrm{p}}([\\hat{X}_{1:S-1};\\hat{I}_{1:S}], [X_{1:S-1};Y_{0:S-1}])\\) \n17: Update \\(\\theta \\gets \\theta - \\alpha \\nabla_\\theta \\mathcal{L}\\) \n18: end for \n19: end for" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.473, + 0.47, + 0.581 + ], + "angle": 0, + "content": "the Warm-up phase, we take the frames of the input sequence as input to the model. However, in the Prediction phase, we use the output of the model at the previous time step as the input to the model at the current time step. In particular, we concatenate the outputs of the Warm-up and Prediction phases to compute the loss. When the prediction loss function \\(\\mathcal{L}^{\\mathrm{P}}\\) is \\(L_{2}\\), its formula is shown in eq. 1." + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.61, + 0.469, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {p}} = \\left\\| \\left[ \\hat {X} _ {1: S - 1}; \\hat {I} _ {1: S} \\right] - \\left[ X _ {1: S - 1}; Y _ {0: S - 1} \\right] \\right\\| _ {2} ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.643, + 0.47, + 0.675 + ], + "angle": 0, + "content": "When the prediction loss function \\(\\mathcal{L}^{\\mathrm{p}}\\) is \\(L_{1} + L_{2}\\), its formula is shown in eq. 2." + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.687, + 0.469, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} ^ {\\mathrm {p}} = \\| \\left[ \\hat {X} _ {1: S - 1}; \\hat {I} _ {1: S} \\right] - \\left[ X _ {1: S - 1}; Y _ {0: S - 1} \\right] \\| _ {1} \\tag {2} \\\\ + \\| \\left[ \\hat {X} _ {1: S - 1}; \\hat {I} _ {1: S} \\right] - \\left[ X _ {1: S - 1}; Y _ {0: S - 1} \\right] \\| _ {2} ^ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.743, + 0.469, + 0.805 + ], + "angle": 0, + "content": "In eq. 1 and eq. 2, \\(\\hat{X}_{1:S-1}\\) represents S-1 predicted frames in the Warm-up phase, \\(\\hat{I}_{1:S}\\) denotes S predicted frames in the Prediction phase, and \\(X_{1:S-1}\\) and \\(Y_{0:S-1}\\) represent S-1 input frames and S ground truth future frames, respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.816, + 0.267, + 0.831 + ], + "angle": 0, + "content": "3.2. SwinLSTM Module" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "ConvLSTM [24] introduces the convolution operators into input-to-state and state-to-state transitions to overcome the shortcomings of FC-LSTM [26] in processing spatiotemporal data. The key equations of ConvLSTM are as" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.558, + 0.105 + ], + "angle": 0, + "content": "follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.115, + 0.892, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} i _ {t} = \\sigma \\left(W _ {x i} * \\mathcal {X} _ {t} + W _ {h i} * \\mathcal {H} _ {t - 1} + b _ {i}\\right) \\\\ f _ {t} = \\sigma \\left(W _ {x f} * \\mathcal {X} _ {t} + W _ {h f} * \\mathcal {H} _ {t - 1} + b _ {f}\\right) \\\\ o _ {t} = \\sigma \\left(W _ {x o} * \\mathcal {X} _ {t} + W _ {h o} * \\mathcal {H} _ {t - 1} + b _ {o}\\right) \\\\ \\mathcal {C} _ {t} = f _ {t} \\circ \\mathcal {C} _ {t - 1} + i _ {t} \\circ \\tanh \\left(W _ {x c} * \\mathcal {X} _ {t} + W _ {h c} * \\mathcal {H} _ {t - 1} + b _ {c}\\right) \\\\ \\mathcal {H} _ {t} = o _ {t} \\circ \\tanh \\left(\\mathcal {C} _ {t}\\right) \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.894, + 0.34 + ], + "angle": 0, + "content": "where \\(*\\) denotes the convolution operator and \\(\\circ\\) denotes the Hadamard product. \\(i_t, f_t, o_t\\) are input, forget, output gate. Different from convolution operators to extract local correlations, the self-attention mechanism captures global spatial dependencies by computing similarity scores across all positions. Thus, we remove all weights \\(W\\) and biases \\(b\\) in eq. 3 to obtain eq. 4:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.35, + 0.773, + 0.366 + ], + "angle": 0, + "content": "\\[\ni _ {t} = f _ {t} = o _ {t} = \\sigma \\left(\\mathcal {X} _ {t} + \\mathcal {H} _ {t - 1}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.368, + 0.891, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {t} = f _ {t} \\circ \\mathcal {C} _ {t - 1} + i _ {t} \\circ \\tanh \\left(\\mathcal {X} _ {t} + \\mathcal {H} _ {t - 1}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.388, + 0.695, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {t} = o _ {t} \\circ \\operatorname {t a n h} \\left(\\mathcal {C} _ {t}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.414, + 0.89, + 0.444 + ], + "angle": 0, + "content": "Obviously, \\( i_t = f_t = o_t \\) in eq. 4. Therefore, we fuse the three gates: \\( i_t, f_t, o_t \\) into one gate, named filter gate \\( \\mathcal{F}_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.565 + ], + "angle": 0, + "content": "The detailed structure of the SwinLSTM Module is presented in Figure 2 (a). In SwinLSTM, the information of cell states \\( \\mathcal{C}_t \\) and hidden states \\( \\mathcal{H}_t \\) is updated horizontally to capture long-term and short-term temporal dependencies. Meanwhile, the Swin Transformer blocks vertically learn global spatial dependencies. The key equations of SwinLSTM are shown in eq. 5, where STB denotes the Swin Transformer blocks and LP denotes the Linear Projection:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.575, + 0.735, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {t} = \\sigma \\left(\\text {S T B} \\left(\\text {L P} \\left(\\mathcal {X} _ {t}; \\mathcal {H} _ {t - 1}\\right)\\right)\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.594, + 0.891, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {t} = \\mathcal {F} _ {t} \\circ (\\operatorname {t a n h} (\\operatorname {S T B} (\\operatorname {L P} (\\mathcal {X} _ {t}; \\mathcal {H} _ {t - 1}))) + \\mathcal {C} _ {t - 1}) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.613, + 0.66, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {t} = \\mathcal {F} _ {t} \\circ \\operatorname {t a n h} \\left(\\mathcal {C} _ {t}\\right)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.725, + 0.654 + ], + "angle": 0, + "content": "3.3. Swin Transformer Block" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.892, + 0.814 + ], + "angle": 0, + "content": "The global multi-head self-attention (MSA) mechanism in the Vision Transformer [6] computes the relationship between a token and all other tokens. Such a computation leads to quadratic computational complexity related to the number of tokens, which is not friendly to dense prediction tasks. To alleviate the computational burden of global self-attention and improve the modeling power, Swin Transformer [20] proposes window-based multi-head self-attention (W-MSA) and shifted-window-based multi-head self-attention (SW-MSA)." + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.822, + 0.813, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} ^ {l} = \\operatorname {W - M S A} \\left(\\ln \\left(\\mathbf {z} ^ {l - 1}\\right)\\right) + \\mathbf {z} ^ {l - 1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.844, + 0.891, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {l} = \\operatorname {M L P} \\left(\\ln \\left(\\hat {\\mathbf {z}} ^ {l}\\right)\\right) + \\hat {\\mathbf {z}} ^ {l} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.864, + 0.807, + 0.884 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} ^ {l + 1} = \\operatorname {S W - M S A} \\left(\\operatorname {L N} \\left(\\mathbf {z} ^ {l}\\right)\\right) + \\mathbf {z} ^ {l}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.885, + 0.807, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {l + 1} = \\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\hat {\\mathbf {z}} ^ {l + 1}\\right)\\right) + \\hat {\\mathbf {z}} ^ {l + 1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13473" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.089, + 0.089, + 0.885, + 0.182 + ], + "angle": 0, + "content": "
DatasetModelSwinLSTMsSTBPatch sizeResolutionTrainTest
Moving MNISTSwinLSTM-D4(2, 6, 6, 2)2(64, 64, 1)10 → 1010 → 10
Human3.6mSwinLSTM-B1122(128, 128, 3)4 → 44 → 4
KTHSwinLSTM-B164(128, 128, 1)10 → 1010 → 20/40
TaxiBJSwinLSTM-B1124(32, 32, 2)4 → 44 → 4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.191, + 0.895, + 0.237 + ], + "angle": 0, + "content": "Table 1. Experimental setup. SwinLSTMs denotes the number of the SwinLSTM Cells in predictive network. STB denotes the number of the Swin Transformer blocks in SwinLSTM cell. Patch size indicates the patch token size. Train and Test represent the number of input and output frames during training and testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.473, + 0.398 + ], + "angle": 0, + "content": "As shown in Figure 3, the first Swin Transformer block is based on W-MSA, consisting of two LayerNorm layers and a 2-layer MLP with GELU non-linearity. A LayerNorm layer and a residual connection are applied before and after each W-MSA module and MLP module, respectively. The second Swin Transformer block is the same as the first except that the W-MSA block in it is replaced by an SW-MSA block. The key equations of Swin Transformer blocks are shown in eq. 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.411, + 0.212, + 0.429 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.47, + 0.557 + ], + "angle": 0, + "content": "We evaluate our proposed model on four commonly used datasets: Moving MNIST [26], TaxiBJ [40], Human3.6m [10], and KTH [23], and present quantitative comparison results and visual examples for each dataset, illustrating the effectiveness and generalizability of our proposed neural network. In addition, we perform ablation studies (section 4.5) and feature map visualization (section 4.6)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.568, + 0.245, + 0.584 + ], + "angle": 0, + "content": "4.1. Implementations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.698 + ], + "angle": 0, + "content": "We use \\( L_{2} \\) loss for the MovingMNIST dataset, \\( L_{1} + L_{2} \\) loss for the Human3.6m, the TaxiBJ, and the KTH datasets. The input and target frames of each dataset are normalized to the intensity of [0, 1]. Each model is optimized with the Adam optimizer [15]. We train our model on an Nvidia RTX A5000 GPU. More detailed parameters for each dataset are listed in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.707, + 0.262, + 0.723 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.839 + ], + "angle": 0, + "content": "We adopt the Mean Squared Error (MSE), the Mean Absolute Error (MAE), the Peak Signal to Noise Ratio (PSNR), and the Structural Similarity Index Measure (SSIM) [35] as metrics to evaluate the quality of predictions. All metrics are averaged over predicted frames. Lower MAE and MSE or higher SSIM and PSNR indicate better prediction accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.182, + 0.862 + ], + "angle": 0, + "content": "4.3. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Moving MNIST. The moving MNIST dataset [26] is a widely used synthetic dataset. In our experiments, the" + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.257, + 0.846, + 0.492 + ], + "angle": 0, + "content": "
MethodMSE ↓SSIM ↑
ConvLSTM [24]103.30.707
DFN [11]89.00.726
FRNN [22]69.70.813
VPN [14]64.10.870
PredRNN [33]56.80.867
CausalLSTM [31]46.50.898
MIM [34]44.20.910
E3D-LSTM [32]41.30.910
LMC [19]41.50.924
MAU [3]27.60.937
PhyDNet [8]24.40.947
CrevNet [39]22.30.949
SwinLSTM17.70.962
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.502, + 0.895, + 0.56 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of SwinLSTM and other methods on Moving MNIST. Each model observes 10 frames and predicts the subsequent 10 frames. Lower MSE and higher SSIM indicate better predictions." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.577, + 0.89, + 0.726 + ], + "angle": 0, + "content": "
MethodMAE /100 ↓MSE /10 ↓SSIM ↑
FRNN [22]19.049.80.771
ConvLSTM [24]18.950.40.776
PredRNN [33]18.948.40.781
CausalLSTM [31]17.245.80.851
MIM [34]17.842.90.790
E3D-LSTM [32]16.646.40.869
PhyDNet [8]16.236.90.901
SwinLSTM11.933.20.913
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.735, + 0.895, + 0.763 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation of different models on the Human3.6m dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "method for generating moving MNIST sequences follows [26], each sequence contains 20 frames, and the first 10 frames and the last 10 frames are used for input and target, respectively. The two handwritten digits in each frame are randomly sampled in the MNIST dataset [17], moving and bouncing around a \\(64 \\times 64\\) pixel black canva with a fixed speed and angle. We can get an infinite number of se" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13474" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.554, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.285, + 0.406, + 0.3 + ], + "angle": 0, + "content": "(a) Moving MNIST" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.089, + 0.887, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.286, + 0.79, + 0.299 + ], + "angle": 0, + "content": "(b) TaxiBJ" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.313, + 0.549, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.466, + 0.428, + 0.481 + ], + "angle": 0, + "content": "(c) KTH (Jogging \\(10\\rightarrow 20\\)" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.313, + 0.875, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.466, + 0.807, + 0.479 + ], + "angle": 0, + "content": "(d) Human3.6m" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.495, + 0.877, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.648, + 0.599, + 0.662 + ], + "angle": 0, + "content": "(e) KTH (Walking \\(10\\rightarrow 40\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.673, + 0.895, + 0.743 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of SwinLSTM on four datasets. The first line is the input, the second line is the ground truth, and the third line is the prediction of SwinLSTM. For Moving MNIST, we add the prediction results of ConvLSTM as the fourth line. For TaxiBJ, we add the absolute difference between the predictions and ground truths as the fourth line. For KTH action, (c) is the jogging action when the model predicts 20 frames based on the previous 10 frames, and (e) is the walking action when the model observes 10 frames and predicts the next 40 frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.471, + 0.816 + ], + "angle": 0, + "content": "quences by applying different speeds and angles to different handwritten digits. We adopt 10000 sequences for training and a fixed set containing 10000 sequences for testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Human3.6m. The human3.6m dataset [10] contains 3.6 million different human poses and corresponding images. Following [34], we only use the 'walking' scenario. Subjects S1, S5, S6, S7, S8 are used for training, and S9, S11 are used for testing. The images in the dataset are resized" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.769, + 0.892, + 0.801 + ], + "angle": 0, + "content": "from \\(1000 \\times 1000 \\times 3\\) to \\(128 \\times 128 \\times 3\\). We train the models to predict the next 4 RGB frames from 4 observations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "KTH. The KTH dataset [23] contains 25 individuals performing 6 categories of human actions (walking, jogging, running, boxing, hand waving, and hand clapping) in 4 different scenarios. We follow the experimental setup in [30], resizing each image to \\(128 \\times 128\\) and using persons 1-16 for training and 17-25 for testing. The models predict 10" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13475" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.277 + ], + "angle": 0, + "content": "
MethodKTH (10 → 20)KTH (10 → 40)
SSIM ↑PSNR ↑SSIM ↑PSNR ↑
ConvLSTM [24]0.71223.580.63922.85
SAVP [7]0.74625.380.70123.97
FRNN [22]0.77126.120.67823.77
DFN [11]0.79427.260.65223.01
PredRNN [33]0.83927.550.70324.16
VarNet [13]0.84328.480.73925.37
SVAP-VAE [18]0.85227.770.81126.18
PredRNN++ [31]0.86528.470.74125.21
E3d-LSTM [32]0.87929.310.81027.24
STMFANet [12]0.89329.850.85127.56
SwinLSTM0.90334.340.87933.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.287, + 0.471, + 0.344 + ], + "angle": 0, + "content": "Table 4. Quantitative evaluation on the KTH test set. We present the model observing 10 frames to predict 20 or 40 frames, and all metrics are averaged over the predicted frames. Higher SSIM and PSNR indicate better prediction quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.469, + 0.399 + ], + "angle": 0, + "content": "frames from 10 observations at training time and 20 or 40 frames at inference time." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.4, + 0.47, + 0.535 + ], + "angle": 0, + "content": "TaxiBJ. TaxiBJ [40] contains complex real-world taxi trajectory data collected from taxicab GPS monitors in Beijing. Each frame in TaxiBJ is a \\(32 \\times 32 \\times 2\\) heat map, where the last dimension represents the flow of traffic entering and leaving the same area. We follow the experimental setup in [40] and use the last four weeks of data for testing and the rest for training. We use 4 observations to predict the next 4 consecutive frames. We adopted per-frame MSE as metrics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.212, + 0.56 + ], + "angle": 0, + "content": "4.4. Main results" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.569, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Table 2, 3, 4 and 5 present quantitative comparisons with previous state-of-the-art models on four datasets. On Moving MNIST, SwinLSTM outperforms previous CNN-based models and achieves the new state-of-the-art results. Specifically, compared to ConvLSTM [24], SwinLSTM reduces the MSE from 103.3 to 17.7 and increases the SSIM from 0.707 to 0.962. On both Human3.6m and KTH, SwinLSTM achieves the state-of-the-art, especially with huge gains of 4.49 (10→20) and 5.59 (10→40) in PSNR on KTH. In the case of TaxiBJ, our proposed SwinLSTM achieved a remarkable reduction in MSE for each predicted frame compared to previous models. Additionally, the interframe MSE difference was smaller than that of the compared models. The results indicate SwinLSTM's great potential for an efficient and generalizable technique for spatiotemporal prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In Figure 4, we present qualitative results of SwinLSTM on all datasets. (a) presents a qualitative analysis of SwinLSTM and ConvLSTM on an extremely hard case of Moving MNIST. In the presented images, two digits are continuously intertwined, posing a challenge to the model's ability to accurately predict their motion trajectories and ap" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.223 + ], + "angle": 0, + "content": "
MethodFrame1 ↓Frame2 ↓Frame3 ↓Frame4 ↓
VPN [14]0.7441.0311.2511.444
ST-ResNet [40]0.6880.9391.1301.288
FRNN [22]0.6820.8230.9891.183
PredRNN [33]0.6340.9341.0471.263
PredRNN++ [31]0.6410.8550.9791.158
E3d-LSTM [32]0.6200.7730.8881.158
MIM [34]0.5540.7370.8870.999
SwinLSTM0.3240.4010.4730.525
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.233, + 0.892, + 0.261 + ], + "angle": 0, + "content": "Table 5. Quantitative results on the TaxiBJ dataset. We report the per-frame MSE of the 4 predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.501 + ], + "angle": 0, + "content": "pearances. As demonstrated, both models effectively capture the trajectory of the digits. However, in terms of appearance, ConvLSTM's prediction results become increasingly blurred as the number of time steps increases, while SwinLSTM's predictions maintain high similarity throughout. SwinLSTM, as an extension of ConvLSTM, utilizes LSTM [9] to capture temporal dependencies, and employs Swin Transformer blocks [20] to model global spatial information. In contrast, ConvLSTM uses convolution to model local spatial information. The significant performance gap between the two models demonstrates that learning global spatial information can better assist the model in capturing spatiotemporal dependencies and thus improve prediction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.893, + 0.729 + ], + "angle": 0, + "content": "To facilitate the observation of the similarity between the predicted results and the ground truth in (b), we visualize the absolute difference between them. For (c), (d), and (e), we observe that SwinLSTM makes accurate predictions of human actions and scenes, where (c) and (e) show visual examples of 2 different categories of actions predicted by SwinLSTM, Jogging and walking respectively. The model receives 10 frames during training and predicts 10 frames, whereas it predicts 20 or 40 frames during testing. Expanding the predicted number of frames during testing poses a significant challenge to the model, requiring it to effectively learn spatiotemporal dependencies from the spatiotemporal data. Our experiments on the KTH dataset demonstrate that SwinLSTM outperforms alternative models in capturing spatiotemporal dependencies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.655, + 0.757 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.839 + ], + "angle": 0, + "content": "In this section, we perform ablation studies on TaxiBJ and Human3.6m to analyze the impact of different elements on model performance. Specifically, we discuss three major elements: the reconstruction layer, patch size, and the number of Swin Transformer blocks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The reconstruction layer. The role of the reconstruction layer is to decode the spatiotemporal representations extracted by the SwinLSTM cell. We conduct experiments on transposed convolution, bilinear interpolation, and lin" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13476" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "ear projection to analyze the impact of different decoding methods. Table 6 shows that the transposed convolution performs much better than the other two methods." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.153, + 0.466, + 0.232 + ], + "angle": 0, + "content": "
The reconstruction layerTaxiBJHuman3.6m
MSESSIMMSE /10SSIM
Transposed convolution0.3900.98033.20.913
Bilinear interpolation0.7940.96139.30.887
Linear Projection0.4150.97972.50.773
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.242, + 0.468, + 0.27 + ], + "angle": 0, + "content": "Table 6. Ablation study on the three different methods of reconstruction layer with SwinLSTM on TaxiBJ and Human3.6m." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.469, + 0.426 + ], + "angle": 0, + "content": "Patch size. The size of the image patch directly determines the length of the input token sequence. The smaller the size, the longer the length. To explore the effect of different patch sizes on prediction accuracy, we conduct experiments with patch sizes set to 2, 4, and 8 on TaxiBJ and Human3.6m. Table 7 shows that the optimal patch size varies across datasets, which indicates that adjusting the patch size to an appropriate size is beneficial to improve the generalization of SwinLSTM." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.44, + 0.45, + 0.533 + ], + "angle": 0, + "content": "
Patch sizeTaxiBJHuman3.6m
MSESSIMMSE /10SSIM
20.4170.97933.20.913
40.3900.98040.20.885
80.3930.98051.90.839
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.543, + 0.47, + 0.571 + ], + "angle": 0, + "content": "Table 7. Ablation study on different patch sizes on TaxiBJ and Human3.6m." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.469, + 0.701 + ], + "angle": 0, + "content": "The number of Swin Transformer Blocks. Based on the mechanism of Swin Transformer blocks (STB) modeling global spatial information, it is evident that the learning capacity of the model varies with the number of STB utilized. We explored the effect of the number of STB from 2 to 18 on model performance. Figure 5 present the results of different numbers of STB on TaxiBJ and Human3.6m." + }, + { + "type": "image", + "bbox": [ + 0.06, + 0.717, + 0.264, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.841, + 0.213, + 0.855 + ], + "angle": 0, + "content": "(a) TaxiBJ" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.717, + 0.473, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.841, + 0.434, + 0.856 + ], + "angle": 0, + "content": "(b) Human3.6m" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.868, + 0.469, + 0.896 + ], + "angle": 0, + "content": "Figure 5. Ablation study on the different numbers of STB with SwinLSTM on TaxiBJ and Human3.6m." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.094, + 0.875, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.48, + 0.892, + 0.565 + ], + "angle": 0, + "content": "Figure 6. Feature map visualization of a sample in the Moving MNIST test set. Hidden States and Cell States are the feature maps of \\(\\mathcal{H}_t\\) and \\(\\mathcal{C}_t\\) of the recurrent cell SwinLSTM (section 3.2) at different time steps. STB-2,STB-4, and STB-6 are the feature maps of the 2nd, 4th, and 6th layers of the Swin Transformer blocks (section 3.3), respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.611, + 0.738, + 0.627 + ], + "angle": 0, + "content": "4.6. Feature map Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To analyze the mechanism of the SwinLSTM cell (section 3.2), we randomly selected a sample from the Moving MNIST test set and visualized the feature maps during inference. Figure 6 shows the results, where white arrows have been added to the images in the third row to facilitate the observation of digit trajectories. The Hidden States capture the position of digits at the current time step, whereas the Cell States memorize the digit trajectory. Meanwhile, we conducted a visualization of the feature maps generated by the second, fourth, and sixth Swin Transformer blocks of the SwinLSTM-B. By observing the feature maps of STB-2 to STB-6, it is evident that STB gradually learns global spatial correlations as the number of interactions of window self-attention increases (section 3.3). By analyzing the visualizations generated by our model, we aimed to gain valuable insights into its performance and behavior. These insights will inform future optimizations and improvements." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13477" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.196, + 0.106 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.345 + ], + "angle": 0, + "content": "In this paper, we propose SwinLSTM, a new recurrent cell that incorporates Swin Transformer blocks and simplified LSTM, and constructs a predictive network for spatiotemporal prediction tasks based on it. Through extensive experimentation, it has been demonstrated that the proposed method, SwinLSTM, achieves excellent performance on various datasets, including Moving MNIST, TaxiBJ, Human3.6m, and KTH. These impressive results showcase the effectiveness and generalization ability of SwinLSTM. Notably, SwinLSTM outperforms its predecessor, ConvLSTM, in prediction accuracy, which suggests that learning global spatial dependencies allows the model to better capture spatiotemporal dependencies. Overall, this study highlights the potential of SwinLSTM as a promising approach for spatiotemporal modeling." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.358, + 0.238, + 0.377 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.47, + 0.492 + ], + "angle": 0, + "content": "This work was supported by the Natural Science Foundation Innovation Research Team Project of Hainan Province (320CXTD431), the National Key Research and Development Program of China (2021YFB1507104), the National Natural Science Foundation of China (32060413), and the Major Science and Technology Project of Hainan Province (ZDKJ2020013), which are gratefully acknowledged." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.506, + 0.175, + 0.521 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.6 + ], + "angle": 0, + "content": "[1] Apratim Bhattacharyya, Mario Fritz, and Bernt Schiele. Long-term on-board prediction of people in traffic scenes under uncertainty. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4194-4202, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.603, + 0.472, + 0.686 + ], + "angle": 0, + "content": "[2] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 205-218. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.688, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[3] Zheng Chang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, Yan Ye, Xiang Xinguang, and Wen Gao. Mau: A motion-aware unit for video prediction and beyond. Advances in Neural Information Processing Systems, 34:26950-26962, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.759, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[4] Yunpeng Chen, Yannis Kalantidis, Jianshu Li, Shuicheng Yan, and Jiashi Feng. A^2-nets: Double attention networks. Advances in neural information processing systems, 31, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.816, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.531, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.2,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[7] Hang Gao, Huazhe Xu, Qi-Zhi Cai, Ruth Wang, Fisher Yu, and Trevor Darrell. Disentangling propagation and generation for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9006-9015, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.218, + 0.894, + 0.286 + ], + "angle": 0, + "content": "[8] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11474-11484, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.288, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[9] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.894, + 0.509 + ], + "angle": 0, + "content": "[12] Beibei Jin, Yu Hu, Qiankun Tang, Jingyu Niu, Zhiping Shi, Yinhe Han, and Xiaowei Li. Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4554-4563, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[13] Beibei Jin, Yu Hu, Yiming Zeng, Qiankun Tang, Shice Liu, and Jing Ye. Varnet: Exploring variations for unsupervised video prediction. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5801-5806. IEEE, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[14] Nal Kalchbrenner, Aäron Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyls, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. In International Conference on Machine Learning, pages 1771-1779. PMLR, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.894, + 0.719 + ], + "angle": 0, + "content": "[16] Yong-Hoon Kwon and Min-Gyu Park. Predicting future frames using retrospective cycle gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1811-1820, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[17] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[18] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[19] Sangmin Lee, Hak Gu Kim, Dae Hwi Choi, Hyung-II Kim, and Yong Man Ro. Video prediction recalling long-term motion context via memory alignment learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3054-3063, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13478" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[21] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. Advances in neural information processing systems, 29, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[22] Marc Oliu, Javier Selva, and Sergio Escalera. Folded recurrent neural networks for future video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 716-731, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[23] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a localsvm approach. In Proceedings of the 17th International Conference on Pattern Recognition, 2004.ICPR 2004., volume 3,pages 32-36.IEEE,2004.5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[24] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[25] Xingjian Shi, Zhihan Gao, Leonard Lausen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Deep learning for precipitation nowcasting: A benchmark and a new model. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[26] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[27] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[28] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[30] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[31] Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and S Yu Philip. Predrnn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning. In International Conference on Machine Learning, pages 5123-5132. PMLR, 2018. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[32] Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, and Li Fei-Fei. Eidetic 3d LSTM: A model for video prediction and beyond. In International conference on learning representations, 2018. 1, 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[33] Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and Philip S Yu. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in neural information processing systems, 30, 2017. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[34] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154-9162, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[35] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[36] Yu-Huan Wu, Yun Liu, Xin Zhan, and Ming-Ming Cheng. P2t: Pyramid pooling transformer for scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[37] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.432, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[38] Ziru Xu, Yunbo Wang, Mingsheng Long, Jianmin Wang, and M KLiss. Predcnn: Predictive learning with cascade convolutions. In IJCAI, pages 2940-2947, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[39] Wei Yu, Yichao Lu, Steve Easterbrook, and Sanja Fidler. Efficient and information-preserving future frame prediction and beyond. 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[40] Junbo Zhang, Yu Zheng, and Dekang Qi. Deep spatiotemporal residual networks for citywide crowd flows prediction. In Thirty-first AAAI conference on artificial intelligence, 2017. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[41] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3008, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[42] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[43] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13479" + } + ] +] \ No newline at end of file diff --git a/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/0bea32db-db74-4a9f-85a7-6fe6526e5e78_origin.pdf b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/0bea32db-db74-4a9f-85a7-6fe6526e5e78_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..57d0e7a2dddf93f2274b9441d6535ee727e5c12e --- /dev/null +++ b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/0bea32db-db74-4a9f-85a7-6fe6526e5e78_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cea98b83e53ff04520fed865bd51d13acc5c9296dc9ad1c3b1869b80b67d058 +size 1165251 diff --git a/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/full.md b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4e9ba3ea07acc5c455ac2caca86116cbe1499dcc --- /dev/null +++ b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/full.md @@ -0,0 +1,325 @@ +# SwinLSTM: Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM + +Song Tang* Chuang Li* Pu Zhang RongNian Tang† Hainan University {songtang, lc, zhangpu, rn.tang}@hainanu.edu.cn + +# Abstract + +Integrating CNNs and RNNs to capture spatiotemporal dependencies is a prevalent strategy for spatiotemporal prediction tasks. However, the property of CNNs to learn local spatial information decreases their efficiency in capturing spatiotemporal dependencies, thereby limiting their prediction accuracy. In this paper, we propose a new recurrent cell, SwinLSTM, which integrates Swin Transformer blocks and the simplified LSTM, an extension that replaces the convolutional structure in ConvLSTM with the self-attention mechanism. Furthermore, we construct a network with SwinLSTM cell as the core for spatiotemporal prediction. Without using unique tricks, SwinLSTM outperforms state-of-the-art methods on Moving MNIST, Human3.6m, TaxiBJ, and KTH datasets. In particular, it exhibits a significant improvement in prediction accuracy compared to ConvLSTM. Our competitive experimental results demonstrate that learning global spatial dependencies is more advantageous for models to capture spatiotemporal dependencies. We hope that SwinLSTM can serve as a solid baseline to promote the advancement of spatiotemporal prediction accuracy. The codes are publicly available at https://github.com/SongTang-x/SwinLSTM. + +# 1. Introduction + +Spatiotemporal prediction has received increasing attention in recent years due to it can benefit many practical applications, e.g., precipitation forecasting [24, 25, 33], autonomous driving [1, 16], and traffic flow prediction [38, 40]. However, the complex physical dynamics and chaotic properties of spatiotemporal predictive learning make it challenging for purely data-driven deep learning methods to make accurate predictions. Existing methods [3, 8, 19, 24, 31-34, 39] integrate CNNs and RNNs to learn spatiotemporal dependencies in spatiotemporal data to improve prediction accuracy. To capture the spatial and temporal depen + +![](images/d80079b0730ca42fd2873ec5adbcd4a1bca9bbcef15ab45bfc32a4982dae5c03.jpg) +Figure 1. Comparison of input-to-state and state-to-state transitions in three different recurrent cells. (a) FC-LSTM utilizes linear operations to process 1D vectors. (b) ConvLSTM employs 2D convolutions to process 3D tensors. (c) The proposed SwinLSTM leverages self-attention to process 2D matrices. + +dencies simultaneously, ConvLSTM [24] extends the fully connected LSTM (FC-LSTM) [26] by replacing linear operations with convolutional operations. Subsequently, several variants of ConvLSTM are proposed. PredRNN [33] and MIM [34] modify the internal structure of the LSTM unit. E3D-LSTM [32] integrates 3D-Convs into LSTMs. PhyD-Net [8] leverages a CNN-based module to disentangle physical dynamics. Admittedly, these methods achieve impressive results on spatiotemporal prediction tasks. However, convolution operators focus on capturing local features and relations and are inefficient for modeling global spatial information [4]. Although the receptive field can be enlarged by stacking convolution layers, the effective receptive field only reaches a fraction of the theoretical receptive field in practice [21]. Therefore, it can be inferred that the CNN-based models may prove to be ineffective in capturing spa + +tiotemporal dependencies, owing to the inherent locality of the CNN architecture, leading to restricted accuracy in predictions. + +Recently, the Vision Transformer (ViT) [6] introduced the Transformer [29] that directly models long-range dependencies into the vision domain. ViT applies a standard Transformer for image classification and attains excellent results with sufficient data. Its outstanding achievement has attracted more researchers to apply Transformer to computer vision. Subsequently, some variants [20, 27, 36, 43] of ViT emerged, with tremendous success on different vision tasks. Notably, the Swin Transformer [20] has exhibited exceptional performance across diverse visual tasks, owing to its unique utilization of local attention and shift window mechanism. + +Inspired by this, we propose SwinLSTM, a new recurrent cell. Specifically, we integrate Swin Transformer blocks and a simplified LSTM module to extract spatiotemporal representations. In addition, we construct a predictive network with SwinLSTM as the core to capture spatial and temporal dependencies for spatiotemporal prediction tasks. As illustrated in Figure 2 (c), we first split an input image at the current time step into a sequence of image patches. Subsequently, the flattened image patches are fed to the patch embedding layer. Then, the SwinLSTM layer receives the transformed patches or the hidden states transformed by the previous layer (Patch Merging or Patch Expanding), and the cell and hidden states of the previous time step to extract the spatiotemporal representations. Finally, the reconstruction layer decodes the spatiotemporal representations to generate the next frame. + +The contributions of this paper can be summarized as follows: + +- We propose a new recurrent cell, named SwinLSTM (section 3.2), which is able to efficiently extract spatiotemporal representations. +- We introduce a new architecture (section 3.1) for spatiotemporal prediction tasks, which can efficiently model spatial and temporal dependencies. +- We evaluate the effectiveness of the proposed model on Moving MNIST, TaxiBJ, Human3.6m, and KTH. Experimental results show that SwinLSTM achieves excellent performance on four datasets. + +# 2. Related Work + +CNN-Based Models Prior models integrating CNNs and RNNs employ various strategies to better capture spatiotemporal dependencies to improve prediction accuracy. ConvLSTM [24] extends FC-LSTM [26] by replacing fully connected operations with convolutional operations to learn spatiotemporal dependencies. PredRNN [33] proposes the + +Spatiotemporal LSTM (ST-LSTM) module, which simultaneously models spatiotemporal information by transferring hidden states in horizontal and vertical directions. PredRNN++ [31] designs a Gradient Highway unit to solve the vanishing gradient problem in PredRNN. E3D-LSTM [32] replaces the 2D convolution in the ST-LSTM module with 3D convolution, allowing the ST-LSTM module to remember more previous information and achieve better prediction performance. MIM [34] replaces the forget gate in the ST-LSTM module with two recurrent units to solve the nonstationary information learning in prediction. CrevNet [39] proposes a CNN-based reversible network to learn complex spatiotemporal dependencies. PhyDNet [8] introduces physical knowledge into a CNN-based model to improve prediction quality. The above models [8, 24, 31-34, 39] improve their ability to capture spatiotemporal dependencies from different perspectives and achieve excellent results. However, despite their widespread usage, convolutional methods have inherent limitations in capturing spatiotemporal dependencies due to their local nature. In order to overcome this challenge, we propose the adoption of a global modeling approach (specifically, Swin Transformer blocks) for learning spatial dependencies, thereby enhancing the model's capacity to capture spatiotemporal dependencies and improve prediction performance. + +Vision Transformers The widespread use of Transformer [29] in the field of natural language processing (NLP) led researchers to introduce it to the vision domain. ViT [6] pioneered the direct application of the Transformer architecture to image classification and achieved great success. However, its excellent performance is based on training on large datasets, which leads to its unsatisfactory performance on smaller datasets. To solve this problem, DeiT [27] proposes several training strategies to allow ViT to perform well on the smaller ImageNet-1K [5] dataset. Subsequently, some variants [28,37,41,42] of ViT achieved impressive results on various vision tasks. In particular, Swin Transformer [20] achieves outstanding performance on image classification, semantic segmentation, and object detection tasks due to its shifted window mechanism and hierarchical design. In this paper, we try to integrate the Swin Transformer blocks and the simplified LSTM to form a SwinLSTM recurrent cell, and use it as the core to build a model to capture temporal and spatial dependencies to perform spatiotemporal prediction tasks. + +# 3. Method + +# 3.1. Overall Architecture + +The overall predictive architecture is depicted in Figure 2 (b and c). We build our base model and a deeper model, called SwinLSTM-B and SwinLSTM-D, respectively. First, + +![](images/306cca612b7971dc9199bfe31f85b3daaaf2610bb80edf878f0b3318a2575048.jpg) +(a) SwinLSTM Cell + +![](images/927ea7a06f3f0a1bc05b227abace829515cdb813ec6194c281b47de3d06d4836.jpg) +(b) SwinLSTM-B + +![](images/7d2337a89868c5816927b213093ba7d1d06694cb70a16513613652390f90c1f9.jpg) +(c) SwinLSTM-D + +an image at the time step $t$ is split into non-overlapping patches. The patch size is $P^2$ , and $P$ equals 2 or 4 in our implementation. Therefore, the feature dimension of each patch is $C \cdot P^2$ ( $C$ denotes the number of channels). Subsequently, the image patches are flattened and fed to the patch embedding layer, which linearly maps the original features of the patches to an arbitrary dimension. Then, for SwinLSTM-B, the SwinLSTM layer receives the transformed image patches, the hidden state $\mathcal{H}_{t-1}$ , and cell state $\mathcal{C}_{t-1}$ to generate the hidden state $\mathcal{H}_t$ and cell state $\mathcal{C}_t$ , where the $\mathcal{H}_t$ is duplicated into two copies, one for the reconstruction layer, the other together with $\mathcal{C}_t$ for the SwinLSTM layer at the next time step. For SwinLSTM-D, we increase the number of SwinLSTM Cells, as well as add Patch Merging and Patch Expanding [2] layers, where the former is utilized for downsampling and the latter for upsampling. $\mathcal{H}_t^{l=m}$ and $\mathcal{C}_t^{l=m}$ refer to the hidden state and cell state at time step $t$ from layer $m$ . Finally, the reconstruction layer maps the hidden state $\mathcal{H}_t$ from the SwinLSTM layer to the input size to get the predicted frame for the next time step. + +![](images/3990bddc9e028974e4316785eeb1f685101e9e6bfbdd8c23abf2233d618a3b6d.jpg) +Figure 2. (a): The detailed structure of the proposed recurrent cell: SwinLSTM. STB and LP denote Swin Transformer blocks and Linear Projection. (b): The architecture of the base model with a single SwinLSTM cell. (c): The architecture of the deep model with multiple SwinLSTM cells. +Figure 3. Swin Transformer Block. + +Both SwinLSTM-B and SwinLSTM-D were trained using the same training approach. The specific training process for SwinLSTM-B is presented in Algorithm 1. During + +Algorithm 1 Training pipeline +1: Require: $\alpha$ , the learning rate. $m$ , the batch size. $n$ , the number of training epochs. $S$ , the number of frames in the input (target) sequence. +2: Initialize parameters of network $(\theta)$ . +3: for $i = 0, 1, \ldots, n-1$ do +4: for number of batches do +5: Sample $\{[X_{0:S-1};Y_{0:S-1}]^{(j)}\}_{j=1}^m$ a batch from the training set. +6: Phase 1: Warm-up Phase +7: Initialize $\mathcal{H}_0, \mathcal{C}_0$ +8: for $t = 0, 1, \ldots, S-2$ do +9: Get $\hat{X}_{t+1}, \mathcal{H}_{t+1}, \mathcal{C}_{t+1} = P(X_t, \mathcal{H}_t, \mathcal{C}_t)$ +10: end for +11: Phase 2: Prediction Phase +12: $\hat{I}_0 \gets X_{S-1}, \mathcal{H}_0 \gets \mathcal{H}_{S-1}, \mathcal{C}_0 \gets \mathcal{C}_{S-1}$ +13: for $t = 0, 1, \ldots, S-1$ do +14: Get $\hat{I}_{t+1}, \mathcal{H}_{t+1}, \mathcal{C}_{t+1} = P(\hat{I}_t, \mathcal{H}_t, \mathcal{C}_t)$ +15: end for +16: $\mathcal{L} \gets \mathcal{L}^{\mathrm{p}}([\hat{X}_{1:S-1};\hat{I}_{1:S}], [X_{1:S-1};Y_{0:S-1}])$ +17: Update $\theta \gets \theta - \alpha \nabla_\theta \mathcal{L}$ +18: end for +19: end for + +the Warm-up phase, we take the frames of the input sequence as input to the model. However, in the Prediction phase, we use the output of the model at the previous time step as the input to the model at the current time step. In particular, we concatenate the outputs of the Warm-up and Prediction phases to compute the loss. When the prediction loss function $\mathcal{L}^{\mathrm{P}}$ is $L_{2}$ , its formula is shown in eq. 1. + +$$ +\mathcal {L} ^ {\mathrm {p}} = \left\| \left[ \hat {X} _ {1: S - 1}; \hat {I} _ {1: S} \right] - \left[ X _ {1: S - 1}; Y _ {0: S - 1} \right] \right\| _ {2} ^ {2} \tag {1} +$$ + +When the prediction loss function $\mathcal{L}^{\mathrm{p}}$ is $L_{1} + L_{2}$ , its formula is shown in eq. 2. + +$$ +\begin{array}{l} \mathcal {L} ^ {\mathrm {p}} = \| \left[ \hat {X} _ {1: S - 1}; \hat {I} _ {1: S} \right] - \left[ X _ {1: S - 1}; Y _ {0: S - 1} \right] \| _ {1} \tag {2} \\ + \| \left[ \hat {X} _ {1: S - 1}; \hat {I} _ {1: S} \right] - \left[ X _ {1: S - 1}; Y _ {0: S - 1} \right] \| _ {2} ^ {2} \\ \end{array} +$$ + +In eq. 1 and eq. 2, $\hat{X}_{1:S-1}$ represents S-1 predicted frames in the Warm-up phase, $\hat{I}_{1:S}$ denotes S predicted frames in the Prediction phase, and $X_{1:S-1}$ and $Y_{0:S-1}$ represent S-1 input frames and S ground truth future frames, respectively. + +# 3.2. SwinLSTM Module + +ConvLSTM [24] introduces the convolution operators into input-to-state and state-to-state transitions to overcome the shortcomings of FC-LSTM [26] in processing spatiotemporal data. The key equations of ConvLSTM are as + +follows: + +$$ +\begin{array}{l} i _ {t} = \sigma \left(W _ {x i} * \mathcal {X} _ {t} + W _ {h i} * \mathcal {H} _ {t - 1} + b _ {i}\right) \\ f _ {t} = \sigma \left(W _ {x f} * \mathcal {X} _ {t} + W _ {h f} * \mathcal {H} _ {t - 1} + b _ {f}\right) \\ o _ {t} = \sigma \left(W _ {x o} * \mathcal {X} _ {t} + W _ {h o} * \mathcal {H} _ {t - 1} + b _ {o}\right) \\ \mathcal {C} _ {t} = f _ {t} \circ \mathcal {C} _ {t - 1} + i _ {t} \circ \tanh \left(W _ {x c} * \mathcal {X} _ {t} + W _ {h c} * \mathcal {H} _ {t - 1} + b _ {c}\right) \\ \mathcal {H} _ {t} = o _ {t} \circ \tanh \left(\mathcal {C} _ {t}\right) \tag {3} \\ \end{array} +$$ + +where $*$ denotes the convolution operator and $\circ$ denotes the Hadamard product. $i_t, f_t, o_t$ are input, forget, output gate. Different from convolution operators to extract local correlations, the self-attention mechanism captures global spatial dependencies by computing similarity scores across all positions. Thus, we remove all weights $W$ and biases $b$ in eq. 3 to obtain eq. 4: + +$$ +i _ {t} = f _ {t} = o _ {t} = \sigma \left(\mathcal {X} _ {t} + \mathcal {H} _ {t - 1}\right) +$$ + +$$ +\mathcal {C} _ {t} = f _ {t} \circ \mathcal {C} _ {t - 1} + i _ {t} \circ \tanh \left(\mathcal {X} _ {t} + \mathcal {H} _ {t - 1}\right) \tag {4} +$$ + +$$ +\mathcal {H} _ {t} = o _ {t} \circ \operatorname {t a n h} \left(\mathcal {C} _ {t}\right) +$$ + +Obviously, $i_t = f_t = o_t$ in eq. 4. Therefore, we fuse the three gates: $i_t, f_t, o_t$ into one gate, named filter gate $\mathcal{F}_t$ . + +The detailed structure of the SwinLSTM Module is presented in Figure 2 (a). In SwinLSTM, the information of cell states $\mathcal{C}_t$ and hidden states $\mathcal{H}_t$ is updated horizontally to capture long-term and short-term temporal dependencies. Meanwhile, the Swin Transformer blocks vertically learn global spatial dependencies. The key equations of SwinLSTM are shown in eq. 5, where STB denotes the Swin Transformer blocks and LP denotes the Linear Projection: + +$$ +\mathcal {F} _ {t} = \sigma \left(\text {S T B} \left(\text {L P} \left(\mathcal {X} _ {t}; \mathcal {H} _ {t - 1}\right)\right)\right) +$$ + +$$ +\mathcal {C} _ {t} = \mathcal {F} _ {t} \circ (\operatorname {t a n h} (\operatorname {S T B} (\operatorname {L P} (\mathcal {X} _ {t}; \mathcal {H} _ {t - 1}))) + \mathcal {C} _ {t - 1}) \tag {5} +$$ + +$$ +\mathcal {H} _ {t} = \mathcal {F} _ {t} \circ \operatorname {t a n h} \left(\mathcal {C} _ {t}\right) +$$ + +# 3.3. Swin Transformer Block + +The global multi-head self-attention (MSA) mechanism in the Vision Transformer [6] computes the relationship between a token and all other tokens. Such a computation leads to quadratic computational complexity related to the number of tokens, which is not friendly to dense prediction tasks. To alleviate the computational burden of global self-attention and improve the modeling power, Swin Transformer [20] proposes window-based multi-head self-attention (W-MSA) and shifted-window-based multi-head self-attention (SW-MSA). + +$$ +\hat {\mathbf {z}} ^ {l} = \operatorname {W - M S A} \left(\ln \left(\mathbf {z} ^ {l - 1}\right)\right) + \mathbf {z} ^ {l - 1} +$$ + +$$ +\mathbf {z} ^ {l} = \operatorname {M L P} \left(\ln \left(\hat {\mathbf {z}} ^ {l}\right)\right) + \hat {\mathbf {z}} ^ {l} \tag {6} +$$ + +$$ +\hat {\mathbf {z}} ^ {l + 1} = \operatorname {S W - M S A} \left(\operatorname {L N} \left(\mathbf {z} ^ {l}\right)\right) + \mathbf {z} ^ {l} +$$ + +$$ +\mathbf {z} ^ {l + 1} = \operatorname {M L P} \left(\operatorname {L N} \left(\hat {\mathbf {z}} ^ {l + 1}\right)\right) + \hat {\mathbf {z}} ^ {l + 1} +$$ + +
DatasetModelSwinLSTMsSTBPatch sizeResolutionTrainTest
Moving MNISTSwinLSTM-D4(2, 6, 6, 2)2(64, 64, 1)10 → 1010 → 10
Human3.6mSwinLSTM-B1122(128, 128, 3)4 → 44 → 4
KTHSwinLSTM-B164(128, 128, 1)10 → 1010 → 20/40
TaxiBJSwinLSTM-B1124(32, 32, 2)4 → 44 → 4
+ +As shown in Figure 3, the first Swin Transformer block is based on W-MSA, consisting of two LayerNorm layers and a 2-layer MLP with GELU non-linearity. A LayerNorm layer and a residual connection are applied before and after each W-MSA module and MLP module, respectively. The second Swin Transformer block is the same as the first except that the W-MSA block in it is replaced by an SW-MSA block. The key equations of Swin Transformer blocks are shown in eq. 6. + +# 4. Experiments + +We evaluate our proposed model on four commonly used datasets: Moving MNIST [26], TaxiBJ [40], Human3.6m [10], and KTH [23], and present quantitative comparison results and visual examples for each dataset, illustrating the effectiveness and generalizability of our proposed neural network. In addition, we perform ablation studies (section 4.5) and feature map visualization (section 4.6). + +# 4.1. Implementations + +We use $L_{2}$ loss for the MovingMNIST dataset, $L_{1} + L_{2}$ loss for the Human3.6m, the TaxiBJ, and the KTH datasets. The input and target frames of each dataset are normalized to the intensity of [0, 1]. Each model is optimized with the Adam optimizer [15]. We train our model on an Nvidia RTX A5000 GPU. More detailed parameters for each dataset are listed in Table 1. + +# 4.2. Evaluation Metrics + +We adopt the Mean Squared Error (MSE), the Mean Absolute Error (MAE), the Peak Signal to Noise Ratio (PSNR), and the Structural Similarity Index Measure (SSIM) [35] as metrics to evaluate the quality of predictions. All metrics are averaged over predicted frames. Lower MAE and MSE or higher SSIM and PSNR indicate better prediction accuracy. + +# 4.3. Datasets + +Moving MNIST. The moving MNIST dataset [26] is a widely used synthetic dataset. In our experiments, the + +Table 1. Experimental setup. SwinLSTMs denotes the number of the SwinLSTM Cells in predictive network. STB denotes the number of the Swin Transformer blocks in SwinLSTM cell. Patch size indicates the patch token size. Train and Test represent the number of input and output frames during training and testing. + +
MethodMSE ↓SSIM ↑
ConvLSTM [24]103.30.707
DFN [11]89.00.726
FRNN [22]69.70.813
VPN [14]64.10.870
PredRNN [33]56.80.867
CausalLSTM [31]46.50.898
MIM [34]44.20.910
E3D-LSTM [32]41.30.910
LMC [19]41.50.924
MAU [3]27.60.937
PhyDNet [8]24.40.947
CrevNet [39]22.30.949
SwinLSTM17.70.962
+ +Table 2. Quantitative comparison of SwinLSTM and other methods on Moving MNIST. Each model observes 10 frames and predicts the subsequent 10 frames. Lower MSE and higher SSIM indicate better predictions. + +
MethodMAE /100 ↓MSE /10 ↓SSIM ↑
FRNN [22]19.049.80.771
ConvLSTM [24]18.950.40.776
PredRNN [33]18.948.40.781
CausalLSTM [31]17.245.80.851
MIM [34]17.842.90.790
E3D-LSTM [32]16.646.40.869
PhyDNet [8]16.236.90.901
SwinLSTM11.933.20.913
+ +Table 3. Quantitative evaluation of different models on the Human3.6m dataset. + +method for generating moving MNIST sequences follows [26], each sequence contains 20 frames, and the first 10 frames and the last 10 frames are used for input and target, respectively. The two handwritten digits in each frame are randomly sampled in the MNIST dataset [17], moving and bouncing around a $64 \times 64$ pixel black canva with a fixed speed and angle. We can get an infinite number of se + +![](images/2c6d5ab6851191cf91ee23c8a403ede3eb6cf8a63f442171d6e1ee4899893af0.jpg) +(a) Moving MNIST + +![](images/cf9d06cb8607db915681d9aa75d73ae17745b33c0c0c2407efd6c5e85b2c1459.jpg) +(b) TaxiBJ + +![](images/92ee2b4a05418d2827fe2067bae898eb17cd26a8b87d48011f53c7d277b6ba9c.jpg) +(c) KTH (Jogging $10\rightarrow 20$ + +![](images/95fe46f60339601a28abb7b13977f35de90e2a06f63ccc676bfb5aecb8e009da.jpg) +(d) Human3.6m + +![](images/cff07aa7e7fd1d7d9d6de58bbd6539daabd873d3eea2f0ef5ce485dc2e04f234.jpg) +(e) KTH (Walking $10\rightarrow 40$ +Figure 4. Qualitative results of SwinLSTM on four datasets. The first line is the input, the second line is the ground truth, and the third line is the prediction of SwinLSTM. For Moving MNIST, we add the prediction results of ConvLSTM as the fourth line. For TaxiBJ, we add the absolute difference between the predictions and ground truths as the fourth line. For KTH action, (c) is the jogging action when the model predicts 20 frames based on the previous 10 frames, and (e) is the walking action when the model observes 10 frames and predicts the next 40 frames. + +quences by applying different speeds and angles to different handwritten digits. We adopt 10000 sequences for training and a fixed set containing 10000 sequences for testing. + +Human3.6m. The human3.6m dataset [10] contains 3.6 million different human poses and corresponding images. Following [34], we only use the 'walking' scenario. Subjects S1, S5, S6, S7, S8 are used for training, and S9, S11 are used for testing. The images in the dataset are resized + +from $1000 \times 1000 \times 3$ to $128 \times 128 \times 3$ . We train the models to predict the next 4 RGB frames from 4 observations. + +KTH. The KTH dataset [23] contains 25 individuals performing 6 categories of human actions (walking, jogging, running, boxing, hand waving, and hand clapping) in 4 different scenarios. We follow the experimental setup in [30], resizing each image to $128 \times 128$ and using persons 1-16 for training and 17-25 for testing. The models predict 10 + +
MethodKTH (10 → 20)KTH (10 → 40)
SSIM ↑PSNR ↑SSIM ↑PSNR ↑
ConvLSTM [24]0.71223.580.63922.85
SAVP [7]0.74625.380.70123.97
FRNN [22]0.77126.120.67823.77
DFN [11]0.79427.260.65223.01
PredRNN [33]0.83927.550.70324.16
VarNet [13]0.84328.480.73925.37
SVAP-VAE [18]0.85227.770.81126.18
PredRNN++ [31]0.86528.470.74125.21
E3d-LSTM [32]0.87929.310.81027.24
STMFANet [12]0.89329.850.85127.56
SwinLSTM0.90334.340.87933.15
+ +frames from 10 observations at training time and 20 or 40 frames at inference time. + +TaxiBJ. TaxiBJ [40] contains complex real-world taxi trajectory data collected from taxicab GPS monitors in Beijing. Each frame in TaxiBJ is a $32 \times 32 \times 2$ heat map, where the last dimension represents the flow of traffic entering and leaving the same area. We follow the experimental setup in [40] and use the last four weeks of data for testing and the rest for training. We use 4 observations to predict the next 4 consecutive frames. We adopted per-frame MSE as metrics. + +# 4.4. Main results + +Table 2, 3, 4 and 5 present quantitative comparisons with previous state-of-the-art models on four datasets. On Moving MNIST, SwinLSTM outperforms previous CNN-based models and achieves the new state-of-the-art results. Specifically, compared to ConvLSTM [24], SwinLSTM reduces the MSE from 103.3 to 17.7 and increases the SSIM from 0.707 to 0.962. On both Human3.6m and KTH, SwinLSTM achieves the state-of-the-art, especially with huge gains of 4.49 (10→20) and 5.59 (10→40) in PSNR on KTH. In the case of TaxiBJ, our proposed SwinLSTM achieved a remarkable reduction in MSE for each predicted frame compared to previous models. Additionally, the interframe MSE difference was smaller than that of the compared models. The results indicate SwinLSTM's great potential for an efficient and generalizable technique for spatiotemporal prediction. + +In Figure 4, we present qualitative results of SwinLSTM on all datasets. (a) presents a qualitative analysis of SwinLSTM and ConvLSTM on an extremely hard case of Moving MNIST. In the presented images, two digits are continuously intertwined, posing a challenge to the model's ability to accurately predict their motion trajectories and ap + +Table 4. Quantitative evaluation on the KTH test set. We present the model observing 10 frames to predict 20 or 40 frames, and all metrics are averaged over the predicted frames. Higher SSIM and PSNR indicate better prediction quality. + +
MethodFrame1 ↓Frame2 ↓Frame3 ↓Frame4 ↓
VPN [14]0.7441.0311.2511.444
ST-ResNet [40]0.6880.9391.1301.288
FRNN [22]0.6820.8230.9891.183
PredRNN [33]0.6340.9341.0471.263
PredRNN++ [31]0.6410.8550.9791.158
E3d-LSTM [32]0.6200.7730.8881.158
MIM [34]0.5540.7370.8870.999
SwinLSTM0.3240.4010.4730.525
+ +Table 5. Quantitative results on the TaxiBJ dataset. We report the per-frame MSE of the 4 predictions. + +pearances. As demonstrated, both models effectively capture the trajectory of the digits. However, in terms of appearance, ConvLSTM's prediction results become increasingly blurred as the number of time steps increases, while SwinLSTM's predictions maintain high similarity throughout. SwinLSTM, as an extension of ConvLSTM, utilizes LSTM [9] to capture temporal dependencies, and employs Swin Transformer blocks [20] to model global spatial information. In contrast, ConvLSTM uses convolution to model local spatial information. The significant performance gap between the two models demonstrates that learning global spatial information can better assist the model in capturing spatiotemporal dependencies and thus improve prediction accuracy. + +To facilitate the observation of the similarity between the predicted results and the ground truth in (b), we visualize the absolute difference between them. For (c), (d), and (e), we observe that SwinLSTM makes accurate predictions of human actions and scenes, where (c) and (e) show visual examples of 2 different categories of actions predicted by SwinLSTM, Jogging and walking respectively. The model receives 10 frames during training and predicts 10 frames, whereas it predicts 20 or 40 frames during testing. Expanding the predicted number of frames during testing poses a significant challenge to the model, requiring it to effectively learn spatiotemporal dependencies from the spatiotemporal data. Our experiments on the KTH dataset demonstrate that SwinLSTM outperforms alternative models in capturing spatiotemporal dependencies. + +# 4.5. Ablation Study + +In this section, we perform ablation studies on TaxiBJ and Human3.6m to analyze the impact of different elements on model performance. Specifically, we discuss three major elements: the reconstruction layer, patch size, and the number of Swin Transformer blocks. + +The reconstruction layer. The role of the reconstruction layer is to decode the spatiotemporal representations extracted by the SwinLSTM cell. We conduct experiments on transposed convolution, bilinear interpolation, and lin + +ear projection to analyze the impact of different decoding methods. Table 6 shows that the transposed convolution performs much better than the other two methods. + +
The reconstruction layerTaxiBJHuman3.6m
MSESSIMMSE /10SSIM
Transposed convolution0.3900.98033.20.913
Bilinear interpolation0.7940.96139.30.887
Linear Projection0.4150.97972.50.773
+ +Patch size. The size of the image patch directly determines the length of the input token sequence. The smaller the size, the longer the length. To explore the effect of different patch sizes on prediction accuracy, we conduct experiments with patch sizes set to 2, 4, and 8 on TaxiBJ and Human3.6m. Table 7 shows that the optimal patch size varies across datasets, which indicates that adjusting the patch size to an appropriate size is beneficial to improve the generalization of SwinLSTM. + +Table 6. Ablation study on the three different methods of reconstruction layer with SwinLSTM on TaxiBJ and Human3.6m. + +
Patch sizeTaxiBJHuman3.6m
MSESSIMMSE /10SSIM
20.4170.97933.20.913
40.3900.98040.20.885
80.3930.98051.90.839
+ +Table 7. Ablation study on different patch sizes on TaxiBJ and Human3.6m. + +The number of Swin Transformer Blocks. Based on the mechanism of Swin Transformer blocks (STB) modeling global spatial information, it is evident that the learning capacity of the model varies with the number of STB utilized. We explored the effect of the number of STB from 2 to 18 on model performance. Figure 5 present the results of different numbers of STB on TaxiBJ and Human3.6m. + +![](images/2bbfb7588ca692fef4ac54889a5d9903fef69d9834b18da4c5c0170f96cfce41.jpg) +(a) TaxiBJ +Figure 5. Ablation study on the different numbers of STB with SwinLSTM on TaxiBJ and Human3.6m. + +![](images/0c077b8cec9c4fe7a472412c80103501ec376f5b10eaabda9fcc501e709bfb8a.jpg) +(b) Human3.6m + +![](images/425f266a1907d3cec76545d82c24edd699af5f0a0814d732a51990a8b5545d90.jpg) +Figure 6. Feature map visualization of a sample in the Moving MNIST test set. Hidden States and Cell States are the feature maps of $\mathcal{H}_t$ and $\mathcal{C}_t$ of the recurrent cell SwinLSTM (section 3.2) at different time steps. STB-2,STB-4, and STB-6 are the feature maps of the 2nd, 4th, and 6th layers of the Swin Transformer blocks (section 3.3), respectively. + +# 4.6. Feature map Visualization + +To analyze the mechanism of the SwinLSTM cell (section 3.2), we randomly selected a sample from the Moving MNIST test set and visualized the feature maps during inference. Figure 6 shows the results, where white arrows have been added to the images in the third row to facilitate the observation of digit trajectories. The Hidden States capture the position of digits at the current time step, whereas the Cell States memorize the digit trajectory. Meanwhile, we conducted a visualization of the feature maps generated by the second, fourth, and sixth Swin Transformer blocks of the SwinLSTM-B. By observing the feature maps of STB-2 to STB-6, it is evident that STB gradually learns global spatial correlations as the number of interactions of window self-attention increases (section 3.3). By analyzing the visualizations generated by our model, we aimed to gain valuable insights into its performance and behavior. These insights will inform future optimizations and improvements. + +# 5. Conclusion + +In this paper, we propose SwinLSTM, a new recurrent cell that incorporates Swin Transformer blocks and simplified LSTM, and constructs a predictive network for spatiotemporal prediction tasks based on it. Through extensive experimentation, it has been demonstrated that the proposed method, SwinLSTM, achieves excellent performance on various datasets, including Moving MNIST, TaxiBJ, Human3.6m, and KTH. These impressive results showcase the effectiveness and generalization ability of SwinLSTM. Notably, SwinLSTM outperforms its predecessor, ConvLSTM, in prediction accuracy, which suggests that learning global spatial dependencies allows the model to better capture spatiotemporal dependencies. Overall, this study highlights the potential of SwinLSTM as a promising approach for spatiotemporal modeling. + +# Acknowledgement + +This work was supported by the Natural Science Foundation Innovation Research Team Project of Hainan Province (320CXTD431), the National Key Research and Development Program of China (2021YFB1507104), the National Natural Science Foundation of China (32060413), and the Major Science and Technology Project of Hainan Province (ZDKJ2020013), which are gratefully acknowledged. + +# References + +[1] Apratim Bhattacharyya, Mario Fritz, and Bernt Schiele. Long-term on-board prediction of people in traffic scenes under uncertainty. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4194-4202, 2018. 1 +[2] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 205-218. Springer, 2023. 3 +[3] Zheng Chang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, Yan Ye, Xiang Xinguang, and Wen Gao. Mau: A motion-aware unit for video prediction and beyond. Advances in Neural Information Processing Systems, 34:26950-26962, 2021. 1, 5 +[4] Yunpeng Chen, Yannis Kalantidis, Jianshu Li, Shuicheng Yan, and Jiashi Feng. A^2-nets: Double attention networks. Advances in neural information processing systems, 31, 2018. 1 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.2,4 +[7] Hang Gao, Huazhe Xu, Qi-Zhi Cai, Ruth Wang, Fisher Yu, and Trevor Darrell. Disentangling propagation and generation for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9006-9015, 2019. 7 +[8] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11474-11484, 2020. 1, 2, 5 +[9] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 7 +[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 5, 6 +[11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 5, 7 +[12] Beibei Jin, Yu Hu, Qiankun Tang, Jingyu Niu, Zhiping Shi, Yinhe Han, and Xiaowei Li. Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4554-4563, 2020. 7 +[13] Beibei Jin, Yu Hu, Yiming Zeng, Qiankun Tang, Shice Liu, and Jing Ye. Varnet: Exploring variations for unsupervised video prediction. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5801-5806. IEEE, 2018. 7 +[14] Nal Kalchbrenner, Aäron Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyls, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. In International Conference on Machine Learning, pages 1771-1779. PMLR, 2017. 5, 7 +[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 5 +[16] Yong-Hoon Kwon and Min-Gyu Park. Predicting future frames using retrospective cycle gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1811-1820, 2019. 1 +[17] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 5 +[18] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 7 +[19] Sangmin Lee, Hak Gu Kim, Dae Hwi Choi, Hyung-II Kim, and Yong Man Ro. Video prediction recalling long-term motion context via memory alignment learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3054-3063, 2021. 1, 5 +[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: + +Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 4, 7 +[21] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. Advances in neural information processing systems, 29, 2016. 1 +[22] Marc Oliu, Javier Selva, and Sergio Escalera. Folded recurrent neural networks for future video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 716-731, 2018. 5, 7 +[23] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a localsvm approach. In Proceedings of the 17th International Conference on Pattern Recognition, 2004.ICPR 2004., volume 3,pages 32-36.IEEE,2004.5,6 +[24] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 1, 2, 4, 5, 7 +[25] Xingjian Shi, Zhihan Gao, Leonard Lausen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Deep learning for precipitation nowcasting: A benchmark and a new model. Advances in neural information processing systems, 30, 2017. 1 +[26] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 1, 2, 4, 5 +[27] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2 +[28] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 2 +[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[30] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017. 6 +[31] Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and S Yu Philip. Predrnn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning. In International Conference on Machine Learning, pages 5123-5132. PMLR, 2018. 1, 2, 5, 7 +[32] Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, and Li Fei-Fei. Eidetic 3d LSTM: A model for video prediction and beyond. In International conference on learning representations, 2018. 1, 2, 5, 7 + +[33] Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and Philip S Yu. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in neural information processing systems, 30, 2017. 1, 2, 5, 7 +[34] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154-9162, 2019. 1, 2, 5, 6, 7 +[35] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[36] Yu-Huan Wu, Yun Liu, Xin Zhan, and Ming-Ming Cheng. P2t: Pyramid pooling transformer for scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[37] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2 +[38] Ziru Xu, Yunbo Wang, Mingsheng Long, Jianmin Wang, and M KLiss. Predcnn: Predictive learning with cascade convolutions. In IJCAI, pages 2940-2947, 2018. 1 +[39] Wei Yu, Yichao Lu, Steve Easterbrook, and Sanja Fidler. Efficient and information-preserving future frame prediction and beyond. 2020. 1, 2, 5 +[40] Junbo Zhang, Yu Zheng, and Dekang Qi. Deep spatiotemporal residual networks for citywide crowd flows prediction. In Thirty-first AAAI conference on artificial intelligence, 2017. 1, 5, 7 +[41] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3008, 2021. 2 +[42] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 2 +[43] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2 \ No newline at end of file diff --git a/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/images.zip b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6798eb1b23033dbabbcf999e75c53cff57de910b --- /dev/null +++ b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e76d44dcd7353f04a931ea9a3ca99e2a573c1e1c3868cc8b5e8fca7699e285d +size 710349 diff --git a/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/layout.json b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7cb5c038e5be0f2c884803cf7c7a6ba1a7e9620f --- /dev/null +++ b/2023/SwinLSTM_ Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM/layout.json @@ -0,0 +1,7711 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 520, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 520, + 138 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 520, + 138 + ], + "type": "text", + "content": "SwinLSTM: Improving Spatiotemporal Prediction Accuracy using Swin Transformer and LSTM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 455, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 455, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 455, + 205 + ], + "type": "text", + "content": "Song Tang* Chuang Li* Pu Zhang RongNian Tang† Hainan University {songtang, lc, zhangpu, rn.tang}@hainanu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 256, + 290, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 520 + ], + "type": "text", + "content": "Integrating CNNs and RNNs to capture spatiotemporal dependencies is a prevalent strategy for spatiotemporal prediction tasks. However, the property of CNNs to learn local spatial information decreases their efficiency in capturing spatiotemporal dependencies, thereby limiting their prediction accuracy. In this paper, we propose a new recurrent cell, SwinLSTM, which integrates Swin Transformer blocks and the simplified LSTM, an extension that replaces the convolutional structure in ConvLSTM with the self-attention mechanism. Furthermore, we construct a network with SwinLSTM cell as the core for spatiotemporal prediction. Without using unique tricks, SwinLSTM outperforms state-of-the-art methods on Moving MNIST, Human3.6m, TaxiBJ, and KTH datasets. In particular, it exhibits a significant improvement in prediction accuracy compared to ConvLSTM. Our competitive experimental results demonstrate that learning global spatial dependencies is more advantageous for models to capture spatiotemporal dependencies. We hope that SwinLSTM can serve as a solid baseline to promote the advancement of spatiotemporal prediction accuracy. The codes are publicly available at https://github.com/SongTang-x/SwinLSTM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 128, + 555 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 563, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 695 + ], + "type": "text", + "content": "Spatiotemporal prediction has received increasing attention in recent years due to it can benefit many practical applications, e.g., precipitation forecasting [24, 25, 33], autonomous driving [1, 16], and traffic flow prediction [38, 40]. However, the complex physical dynamics and chaotic properties of spatiotemporal predictive learning make it challenging for purely data-driven deep learning methods to make accurate predictions. Existing methods [3, 8, 19, 24, 31-34, 39] integrate CNNs and RNNs to learn spatiotemporal dependencies in spatiotemporal data to improve prediction accuracy. To capture the spatial and temporal depen" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 240, + 535, + 435 + ], + "blocks": [ + { + "bbox": [ + 321, + 240, + 535, + 435 + ], + "lines": [ + { + "bbox": [ + 321, + 240, + 535, + 435 + ], + "spans": [ + { + "bbox": [ + 321, + 240, + 535, + 435 + ], + "type": "image", + "image_path": "d80079b0730ca42fd2873ec5adbcd4a1bca9bbcef15ab45bfc32a4982dae5c03.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 445, + 547, + 501 + ], + "lines": [ + { + "bbox": [ + 305, + 445, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 547, + 501 + ], + "type": "text", + "content": "Figure 1. Comparison of input-to-state and state-to-state transitions in three different recurrent cells. (a) FC-LSTM utilizes linear operations to process 1D vectors. (b) ConvLSTM employs 2D convolutions to process 3D tensors. (c) The proposed SwinLSTM leverages self-attention to process 2D matrices." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "dencies simultaneously, ConvLSTM [24] extends the fully connected LSTM (FC-LSTM) [26] by replacing linear operations with convolutional operations. Subsequently, several variants of ConvLSTM are proposed. PredRNN [33] and MIM [34] modify the internal structure of the LSTM unit. E3D-LSTM [32] integrates 3D-Convs into LSTMs. PhyD-Net [8] leverages a CNN-based module to disentangle physical dynamics. Admittedly, these methods achieve impressive results on spatiotemporal prediction tasks. However, convolution operators focus on capturing local features and relations and are inefficient for modeling global spatial information [4]. Although the receptive field can be enlarged by stacking convolution layers, the effective receptive field only reaches a fraction of the theoretical receptive field in practice [21]. Therefore, it can be inferred that the CNN-based models may prove to be ineffective in capturing spa" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 206, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 206, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 206, + 714 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13470" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "content": "tiotemporal dependencies, owing to the inherent locality of the CNN architecture, leading to restricted accuracy in predictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 251 + ], + "type": "text", + "content": "Recently, the Vision Transformer (ViT) [6] introduced the Transformer [29] that directly models long-range dependencies into the vision domain. ViT applies a standard Transformer for image classification and attains excellent results with sufficient data. Its outstanding achievement has attracted more researchers to apply Transformer to computer vision. Subsequently, some variants [20, 27, 36, 43] of ViT emerged, with tremendous success on different vision tasks. Notably, the Swin Transformer [20] has exhibited exceptional performance across diverse visual tasks, owing to its unique utilization of local attention and shift window mechanism." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 441 + ], + "type": "text", + "content": "Inspired by this, we propose SwinLSTM, a new recurrent cell. Specifically, we integrate Swin Transformer blocks and a simplified LSTM module to extract spatiotemporal representations. In addition, we construct a predictive network with SwinLSTM as the core to capture spatial and temporal dependencies for spatiotemporal prediction tasks. As illustrated in Figure 2 (c), we first split an input image at the current time step into a sequence of image patches. Subsequently, the flattened image patches are fed to the patch embedding layer. Then, the SwinLSTM layer receives the transformed patches or the hidden states transformed by the previous layer (Patch Merging or Patch Expanding), and the cell and hidden states of the previous time step to extract the spatiotemporal representations. Finally, the reconstruction layer decodes the spatiotemporal representations to generate the next frame." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 443, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 287, + 466 + ], + "type": "text", + "content": "The contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 475, + 286, + 609 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 475, + 286, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 475, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 59, + 475, + 286, + 510 + ], + "type": "text", + "content": "- We propose a new recurrent cell, named SwinLSTM (section 3.2), which is able to efficiently extract spatiotemporal representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 518, + 286, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 518, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 59, + 518, + 286, + 554 + ], + "type": "text", + "content": "- We introduce a new architecture (section 3.1) for spatiotemporal prediction tasks, which can efficiently model spatial and temporal dependencies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 563, + 286, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 563, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 59, + 563, + 286, + 609 + ], + "type": "text", + "content": "- We evaluate the effectiveness of the proposed model on Moving MNIST, TaxiBJ, Human3.6m, and KTH. Experimental results show that SwinLSTM achieves excellent performance on four datasets." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 621, + 133, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 133, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 133, + 633 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "CNN-Based Models Prior models integrating CNNs and RNNs employ various strategies to better capture spatiotemporal dependencies to improve prediction accuracy. ConvLSTM [24] extends FC-LSTM [26] by replacing fully connected operations with convolutional operations to learn spatiotemporal dependencies. PredRNN [33] proposes the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 371 + ], + "type": "text", + "content": "Spatiotemporal LSTM (ST-LSTM) module, which simultaneously models spatiotemporal information by transferring hidden states in horizontal and vertical directions. PredRNN++ [31] designs a Gradient Highway unit to solve the vanishing gradient problem in PredRNN. E3D-LSTM [32] replaces the 2D convolution in the ST-LSTM module with 3D convolution, allowing the ST-LSTM module to remember more previous information and achieve better prediction performance. MIM [34] replaces the forget gate in the ST-LSTM module with two recurrent units to solve the nonstationary information learning in prediction. CrevNet [39] proposes a CNN-based reversible network to learn complex spatiotemporal dependencies. PhyDNet [8] introduces physical knowledge into a CNN-based model to improve prediction quality. The above models [8, 24, 31-34, 39] improve their ability to capture spatiotemporal dependencies from different perspectives and achieve excellent results. However, despite their widespread usage, convolutional methods have inherent limitations in capturing spatiotemporal dependencies due to their local nature. In order to overcome this challenge, we propose the adoption of a global modeling approach (specifically, Swin Transformer blocks) for learning spatial dependencies, thereby enhancing the model's capacity to capture spatiotemporal dependencies and improve prediction performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 388, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 626 + ], + "type": "text", + "content": "Vision Transformers The widespread use of Transformer [29] in the field of natural language processing (NLP) led researchers to introduce it to the vision domain. ViT [6] pioneered the direct application of the Transformer architecture to image classification and achieved great success. However, its excellent performance is based on training on large datasets, which leads to its unsatisfactory performance on smaller datasets. To solve this problem, DeiT [27] proposes several training strategies to allow ViT to perform well on the smaller ImageNet-1K [5] dataset. Subsequently, some variants [28,37,41,42] of ViT achieved impressive results on various vision tasks. In particular, Swin Transformer [20] achieves outstanding performance on image classification, semantic segmentation, and object detection tasks due to its shifted window mechanism and hierarchical design. In this paper, we try to integrate the Swin Transformer blocks and the simplified LSTM to form a SwinLSTM recurrent cell, and use it as the core to build a model to capture temporal and spatial dependencies to perform spatiotemporal prediction tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 639, + 361, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 361, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 361, + 651 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 658, + 425, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 425, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 425, + 670 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "The overall predictive architecture is depicted in Figure 2 (b and c). We build our base model and a deeper model, called SwinLSTM-B and SwinLSTM-D, respectively. First," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13471" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 76, + 285, + 257 + ], + "blocks": [ + { + "bbox": [ + 50, + 76, + 285, + 257 + ], + "lines": [ + { + "bbox": [ + 50, + 76, + 285, + 257 + ], + "spans": [ + { + "bbox": [ + 50, + 76, + 285, + 257 + ], + "type": "image", + "image_path": "306cca612b7971dc9199bfe31f85b3daaaf2610bb80edf878f0b3318a2575048.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 259, + 206, + 270 + ], + "lines": [ + { + "bbox": [ + 137, + 259, + 206, + 270 + ], + "spans": [ + { + "bbox": [ + 137, + 259, + 206, + 270 + ], + "type": "text", + "content": "(a) SwinLSTM Cell" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 277, + 289, + 391 + ], + "blocks": [ + { + "bbox": [ + 65, + 277, + 289, + 391 + ], + "lines": [ + { + "bbox": [ + 65, + 277, + 289, + 391 + ], + "spans": [ + { + "bbox": [ + 65, + 277, + 289, + 391 + ], + "type": "image", + "image_path": "927ea7a06f3f0a1bc05b227abace829515cdb813ec6194c281b47de3d06d4836.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 393, + 192, + 403 + ], + "lines": [ + { + "bbox": [ + 130, + 393, + 192, + 403 + ], + "spans": [ + { + "bbox": [ + 130, + 393, + 192, + 403 + ], + "type": "text", + "content": "(b) SwinLSTM-B" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 331, + 79, + 541, + 388 + ], + "blocks": [ + { + "bbox": [ + 331, + 79, + 541, + 388 + ], + "lines": [ + { + "bbox": [ + 331, + 79, + 541, + 388 + ], + "spans": [ + { + "bbox": [ + 331, + 79, + 541, + 388 + ], + "type": "image", + "image_path": "7d2337a89868c5816927b213093ba7d1d06694cb70a16513613652390f90c1f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 391, + 468, + 402 + ], + "lines": [ + { + "bbox": [ + 406, + 391, + 468, + 402 + ], + "spans": [ + { + "bbox": [ + 406, + 391, + 468, + 402 + ], + "type": "text", + "content": "(c) SwinLSTM-D" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": "an image at the time step " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " is split into non-overlapping patches. The patch size is " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "P^2" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " equals 2 or 4 in our implementation. Therefore, the feature dimension of each patch is " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "C \\cdot P^2" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " denotes the number of channels). Subsequently, the image patches are flattened and fed to the patch embedding layer, which linearly maps the original features of the patches to an arbitrary dimension. Then, for SwinLSTM-B, the SwinLSTM layer receives the transformed image patches, the hidden state " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{t-1}" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": ", and cell state " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{t-1}" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " to generate the hidden state " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " and cell state " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " is duplicated into two copies, one for the reconstruction layer, the other together with " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " for the SwinLSTM layer at the next time step. For SwinLSTM-D, we increase the number of SwinLSTM Cells, as well as add Patch Merging and Patch Expanding [2] layers, where the former is utilized for downsampling and the latter for upsampling. " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t^{l=m}" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t^{l=m}" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " refer to the hidden state and cell state at time step " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " from layer " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": ". Finally, the reconstruction layer maps the hidden state " + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 47, + 471, + 289, + 712 + ], + "type": "text", + "content": " from the SwinLSTM layer to the input size to get the predicted frame for the next time step." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 347, + 472, + 502, + 628 + ], + "blocks": [ + { + "bbox": [ + 46, + 416, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 46, + 416, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 547, + 449 + ], + "type": "text", + "content": "Figure 2. (a): The detailed structure of the proposed recurrent cell: SwinLSTM. STB and LP denote Swin Transformer blocks and Linear Projection. (b): The architecture of the base model with a single SwinLSTM cell. (c): The architecture of the deep model with multiple SwinLSTM cells." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 472, + 502, + 628 + ], + "lines": [ + { + "bbox": [ + 347, + 472, + 502, + 628 + ], + "spans": [ + { + "bbox": [ + 347, + 472, + 502, + 628 + ], + "type": "image", + "image_path": "3990bddc9e028974e4316785eeb1f685101e9e6bfbdd8c23abf2233d618a3b6d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 640, + 490, + 652 + ], + "lines": [ + { + "bbox": [ + 361, + 640, + 490, + 652 + ], + "spans": [ + { + "bbox": [ + 361, + 640, + 490, + 652 + ], + "type": "text", + "content": "Figure 3. Swin Transformer Block." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Both SwinLSTM-B and SwinLSTM-D were trained using the same training approach. The specific training process for SwinLSTM-B is presented in Algorithm 1. During" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "13472" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 52, + 87, + 287, + 352 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 173, + 84 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 173, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 173, + 84 + ], + "type": "text", + "content": "Algorithm 1 Training pipeline" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "lines": [ + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": "1: Require: " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": ", the learning rate. " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": ", the batch size. " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": ", the number of training epochs. " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": ", the number of frames in the input (target) sequence. \n2: Initialize parameters of network " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "(\\theta)" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": ". \n3: for " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "i = 0, 1, \\ldots, n-1" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": " do \n4: for number of batches do \n5: Sample " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\{[X_{0:S-1};Y_{0:S-1}]^{(j)}\\}_{j=1}^m" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": " a batch from the training set. \n6: Phase 1: Warm-up Phase \n7: Initialize " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_0, \\mathcal{C}_0" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": " \n8: for " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "t = 0, 1, \\ldots, S-2" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": " do \n9: Get " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\hat{X}_{t+1}, \\mathcal{H}_{t+1}, \\mathcal{C}_{t+1} = P(X_t, \\mathcal{H}_t, \\mathcal{C}_t)" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": " \n10: end for \n11: Phase 2: Prediction Phase \n12: " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\hat{I}_0 \\gets X_{S-1}, \\mathcal{H}_0 \\gets \\mathcal{H}_{S-1}, \\mathcal{C}_0 \\gets \\mathcal{C}_{S-1}" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": " \n13: for " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "t = 0, 1, \\ldots, S-1" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": " do \n14: Get " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t+1}, \\mathcal{H}_{t+1}, \\mathcal{C}_{t+1} = P(\\hat{I}_t, \\mathcal{H}_t, \\mathcal{C}_t)" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": " \n15: end for \n16: " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{L} \\gets \\mathcal{L}^{\\mathrm{p}}([\\hat{X}_{1:S-1};\\hat{I}_{1:S}], [X_{1:S-1};Y_{0:S-1}])" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": " \n17: Update " + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\theta \\gets \\theta - \\alpha \\nabla_\\theta \\mathcal{L}" + }, + { + "bbox": [ + 52, + 87, + 287, + 352 + ], + "type": "text", + "content": " \n18: end for \n19: end for" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 47, + 374, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 287, + 460 + ], + "type": "text", + "content": "the Warm-up phase, we take the frames of the input sequence as input to the model. However, in the Prediction phase, we use the output of the model at the previous time step as the input to the model at the current time step. In particular, we concatenate the outputs of the Warm-up and Prediction phases to compute the loss. When the prediction loss function " + }, + { + "bbox": [ + 47, + 374, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{P}}" + }, + { + "bbox": [ + 47, + 374, + 287, + 460 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 47, + 374, + 287, + 460 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 374, + 287, + 460 + ], + "type": "text", + "content": ", its formula is shown in eq. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 483, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 483, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 85, + 483, + 287, + 498 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {p}} = \\left\\| \\left[ \\hat {X} _ {1: S - 1}; \\hat {I} _ {1: S} \\right] - \\left[ X _ {1: S - 1}; Y _ {0: S - 1} \\right] \\right\\| _ {2} ^ {2} \\tag {1}", + "image_path": "7a6a91dda5c18a0dd90f4a6a9f01ba49366f6dcb7b930b34916be327b4de44b5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 509, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 287, + 534 + ], + "type": "text", + "content": "When the prediction loss function " + }, + { + "bbox": [ + 47, + 509, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{p}}" + }, + { + "bbox": [ + 47, + 509, + 287, + 534 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 47, + 509, + 287, + 534 + ], + "type": "inline_equation", + "content": "L_{1} + L_{2}" + }, + { + "bbox": [ + 47, + 509, + 287, + 534 + ], + "type": "text", + "content": ", its formula is shown in eq. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 544, + 287, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 544, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 78, + 544, + 287, + 576 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} ^ {\\mathrm {p}} = \\| \\left[ \\hat {X} _ {1: S - 1}; \\hat {I} _ {1: S} \\right] - \\left[ X _ {1: S - 1}; Y _ {0: S - 1} \\right] \\| _ {1} \\tag {2} \\\\ + \\| \\left[ \\hat {X} _ {1: S - 1}; \\hat {I} _ {1: S} \\right] - \\left[ X _ {1: S - 1}; Y _ {0: S - 1} \\right] \\| _ {2} ^ {2} \\\\ \\end{array}", + "image_path": "391360485966678d2f5c3cdd761039de6b38fb6b1cbcfadf75345ae279d85bdd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "type": "text", + "content": "In eq. 1 and eq. 2, " + }, + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\hat{X}_{1:S-1}" + }, + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "type": "text", + "content": " represents S-1 predicted frames in the Warm-up phase, " + }, + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\hat{I}_{1:S}" + }, + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "type": "text", + "content": " denotes S predicted frames in the Prediction phase, and " + }, + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "type": "inline_equation", + "content": "X_{1:S-1}" + }, + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "type": "inline_equation", + "content": "Y_{0:S-1}" + }, + { + "bbox": [ + 47, + 588, + 287, + 637 + ], + "type": "text", + "content": " represent S-1 input frames and S ground truth future frames, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 646, + 163, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 163, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 163, + 658 + ], + "type": "text", + "content": "3.2. SwinLSTM Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "ConvLSTM [24] introduces the convolution operators into input-to-state and state-to-state transitions to overcome the shortcomings of FC-LSTM [26] in processing spatiotemporal data. The key equations of ConvLSTM are as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 341, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 341, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 341, + 83 + ], + "type": "text", + "content": "follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 91, + 545, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 545, + 176 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} i _ {t} = \\sigma \\left(W _ {x i} * \\mathcal {X} _ {t} + W _ {h i} * \\mathcal {H} _ {t - 1} + b _ {i}\\right) \\\\ f _ {t} = \\sigma \\left(W _ {x f} * \\mathcal {X} _ {t} + W _ {h f} * \\mathcal {H} _ {t - 1} + b _ {f}\\right) \\\\ o _ {t} = \\sigma \\left(W _ {x o} * \\mathcal {X} _ {t} + W _ {h o} * \\mathcal {H} _ {t - 1} + b _ {o}\\right) \\\\ \\mathcal {C} _ {t} = f _ {t} \\circ \\mathcal {C} _ {t - 1} + i _ {t} \\circ \\tanh \\left(W _ {x c} * \\mathcal {X} _ {t} + W _ {h c} * \\mathcal {H} _ {t - 1} + b _ {c}\\right) \\\\ \\mathcal {H} _ {t} = o _ {t} \\circ \\tanh \\left(\\mathcal {C} _ {t}\\right) \\tag {3} \\\\ \\end{array}", + "image_path": "03a22273d39215cf82149c733e4aa10d5cde7c655ee3fc78b4f40743b1120a55.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "text", + "content": " denotes the convolution operator and " + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "text", + "content": " denotes the Hadamard product. " + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "inline_equation", + "content": "i_t, f_t, o_t" + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "text", + "content": " are input, forget, output gate. Different from convolution operators to extract local correlations, the self-attention mechanism captures global spatial dependencies by computing similarity scores across all positions. Thus, we remove all weights " + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "text", + "content": " and biases " + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 184, + 547, + 269 + ], + "type": "text", + "content": " in eq. 3 to obtain eq. 4:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 277, + 473, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 277, + 473, + 289 + ], + "spans": [ + { + "bbox": [ + 347, + 277, + 473, + 289 + ], + "type": "interline_equation", + "content": "i _ {t} = f _ {t} = o _ {t} = \\sigma \\left(\\mathcal {X} _ {t} + \\mathcal {H} _ {t - 1}\\right)", + "image_path": "be73b4035ced79ac31b35d949971ffcf5cb5a1278e38fa576fc6e2483f8aad86.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 345, + 291, + 545, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 291, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 345, + 291, + 545, + 304 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {t} = f _ {t} \\circ \\mathcal {C} _ {t - 1} + i _ {t} \\circ \\tanh \\left(\\mathcal {X} _ {t} + \\mathcal {H} _ {t - 1}\\right) \\tag {4}", + "image_path": "1796053c6b7861478f6489bbc534afb9fa66b7a05585a3cf6f327f82c682732b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 307, + 425, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 307, + 425, + 319 + ], + "spans": [ + { + "bbox": [ + 345, + 307, + 425, + 319 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {t} = o _ {t} \\circ \\operatorname {t a n h} \\left(\\mathcal {C} _ {t}\\right)", + "image_path": "a432de421947417ed85534736e8592206adf8113423f30178128a244de1bad66.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 327, + 544, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 544, + 351 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 544, + 351 + ], + "type": "text", + "content": "Obviously, " + }, + { + "bbox": [ + 306, + 327, + 544, + 351 + ], + "type": "inline_equation", + "content": "i_t = f_t = o_t" + }, + { + "bbox": [ + 306, + 327, + 544, + 351 + ], + "type": "text", + "content": " in eq. 4. Therefore, we fuse the three gates: " + }, + { + "bbox": [ + 306, + 327, + 544, + 351 + ], + "type": "inline_equation", + "content": "i_t, f_t, o_t" + }, + { + "bbox": [ + 306, + 327, + 544, + 351 + ], + "type": "text", + "content": " into one gate, named filter gate " + }, + { + "bbox": [ + 306, + 327, + 544, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 306, + 327, + 544, + 351 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 351, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 447 + ], + "type": "text", + "content": "The detailed structure of the SwinLSTM Module is presented in Figure 2 (a). In SwinLSTM, the information of cell states " + }, + { + "bbox": [ + 304, + 351, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 304, + 351, + 545, + 447 + ], + "type": "text", + "content": " and hidden states " + }, + { + "bbox": [ + 304, + 351, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 304, + 351, + 545, + 447 + ], + "type": "text", + "content": " is updated horizontally to capture long-term and short-term temporal dependencies. Meanwhile, the Swin Transformer blocks vertically learn global spatial dependencies. The key equations of SwinLSTM are shown in eq. 5, where STB denotes the Swin Transformer blocks and LP denotes the Linear Projection:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 455, + 449, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 455, + 449, + 468 + ], + "spans": [ + { + "bbox": [ + 318, + 455, + 449, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {t} = \\sigma \\left(\\text {S T B} \\left(\\text {L P} \\left(\\mathcal {X} _ {t}; \\mathcal {H} _ {t - 1}\\right)\\right)\\right)", + "image_path": "acebfca430613a5f2b84b5d53e1d5bbac151afbc0bf4b6bf5eff81f35f429aa8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 470, + 545, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 470, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 320, + 470, + 545, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {t} = \\mathcal {F} _ {t} \\circ (\\operatorname {t a n h} (\\operatorname {S T B} (\\operatorname {L P} (\\mathcal {X} _ {t}; \\mathcal {H} _ {t - 1}))) + \\mathcal {C} _ {t - 1}) \\tag {5}", + "image_path": "d8b422d54fca5aaf46fee9a6c52284baf172ff827e251a7d7373c94391b1156e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 485, + 403, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 485, + 403, + 498 + ], + "spans": [ + { + "bbox": [ + 318, + 485, + 403, + 498 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {t} = \\mathcal {F} _ {t} \\circ \\operatorname {t a n h} \\left(\\mathcal {C} _ {t}\\right)", + "image_path": "4b9d8d04dbba686a91b88dbe4f68e748302c590857672b590be291551eeb96c1.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 506, + 443, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 443, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 443, + 517 + ], + "type": "text", + "content": "3.3. Swin Transformer Block" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 525, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 644 + ], + "type": "text", + "content": "The global multi-head self-attention (MSA) mechanism in the Vision Transformer [6] computes the relationship between a token and all other tokens. Such a computation leads to quadratic computational complexity related to the number of tokens, which is not friendly to dense prediction tasks. To alleviate the computational burden of global self-attention and improve the modeling power, Swin Transformer [20] proposes window-based multi-head self-attention (W-MSA) and shifted-window-based multi-head self-attention (SW-MSA)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 352, + 651, + 497, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 651, + 497, + 666 + ], + "spans": [ + { + "bbox": [ + 352, + 651, + 497, + 666 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} ^ {l} = \\operatorname {W - M S A} \\left(\\ln \\left(\\mathbf {z} ^ {l - 1}\\right)\\right) + \\mathbf {z} ^ {l - 1}", + "image_path": "1b2d608ae49a1a65f8a6991f09ac7e34289273ddb32880e190d802f0475bfd97.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 354, + 668, + 545, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 668, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 354, + 668, + 545, + 689 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {l} = \\operatorname {M L P} \\left(\\ln \\left(\\hat {\\mathbf {z}} ^ {l}\\right)\\right) + \\hat {\\mathbf {z}} ^ {l} \\tag {6}", + "image_path": "021768394f37cae0cc226c319bf01304f07e79c5642c3746d2abab244ee024d8.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 354, + 684, + 493, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 684, + 493, + 700 + ], + "spans": [ + { + "bbox": [ + 354, + 684, + 493, + 700 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} ^ {l + 1} = \\operatorname {S W - M S A} \\left(\\operatorname {L N} \\left(\\mathbf {z} ^ {l}\\right)\\right) + \\mathbf {z} ^ {l}", + "image_path": "d8d53d6b9a226baa0fe7d8343eb3f2038dc160ffb105fb4af05338b31aebd11e.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 354, + 700, + 493, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 700, + 493, + 715 + ], + "spans": [ + { + "bbox": [ + 354, + 700, + 493, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {l + 1} = \\operatorname {M L P} \\left(\\operatorname {L N} \\left(\\hat {\\mathbf {z}} ^ {l + 1}\\right)\\right) + \\hat {\\mathbf {z}} ^ {l + 1}", + "image_path": "35902e30df76bb6676f3b15dad62cbd27eea5b0a59210ff5219e6ec1e9ce9cd1.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13473" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 70, + 541, + 144 + ], + "blocks": [ + { + "bbox": [ + 54, + 70, + 541, + 144 + ], + "lines": [ + { + "bbox": [ + 54, + 70, + 541, + 144 + ], + "spans": [ + { + "bbox": [ + 54, + 70, + 541, + 144 + ], + "type": "table", + "html": "
DatasetModelSwinLSTMsSTBPatch sizeResolutionTrainTest
Moving MNISTSwinLSTM-D4(2, 6, 6, 2)2(64, 64, 1)10 → 1010 → 10
Human3.6mSwinLSTM-B1122(128, 128, 3)4 → 44 → 4
KTHSwinLSTM-B164(128, 128, 1)10 → 1010 → 20/40
TaxiBJSwinLSTM-B1124(32, 32, 2)4 → 44 → 4
", + "image_path": "0e125cb317286c322cbb1eadf056c58f36d4f864f9d408c9f25e6020a00559b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 315 + ], + "type": "text", + "content": "As shown in Figure 3, the first Swin Transformer block is based on W-MSA, consisting of two LayerNorm layers and a 2-layer MLP with GELU non-linearity. A LayerNorm layer and a residual connection are applied before and after each W-MSA module and MLP module, respectively. The second Swin Transformer block is the same as the first except that the W-MSA block in it is replaced by an SW-MSA block. The key equations of Swin Transformer blocks are shown in eq. 6." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 325, + 129, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 129, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 129, + 339 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 441 + ], + "type": "text", + "content": "We evaluate our proposed model on four commonly used datasets: Moving MNIST [26], TaxiBJ [40], Human3.6m [10], and KTH [23], and present quantitative comparison results and visual examples for each dataset, illustrating the effectiveness and generalizability of our proposed neural network. In addition, we perform ablation studies (section 4.5) and feature map visualization (section 4.6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 449, + 149, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 149, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 149, + 462 + ], + "type": "text", + "content": "4.1. Implementations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": " loss for the MovingMNIST dataset, " + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "inline_equation", + "content": "L_{1} + L_{2}" + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": " loss for the Human3.6m, the TaxiBJ, and the KTH datasets. The input and target frames of each dataset are normalized to the intensity of [0, 1]. Each model is optimized with the Adam optimizer [15]. We train our model on an Nvidia RTX A5000 GPU. More detailed parameters for each dataset are listed in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 559, + 160, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 160, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 160, + 572 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 578, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 664 + ], + "type": "text", + "content": "We adopt the Mean Squared Error (MSE), the Mean Absolute Error (MAE), the Peak Signal to Noise Ratio (PSNR), and the Structural Similarity Index Measure (SSIM) [35] as metrics to evaluate the quality of predictions. All metrics are averaged over predicted frames. Lower MAE and MSE or higher SSIM and PSNR indicate better prediction accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 111, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 111, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 111, + 682 + ], + "type": "text", + "content": "4.3. Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Moving MNIST. The moving MNIST dataset [26] is a widely used synthetic dataset. In our experiments, the" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 337, + 203, + 517, + 389 + ], + "blocks": [ + { + "bbox": [ + 46, + 151, + 547, + 187 + ], + "lines": [ + { + "bbox": [ + 46, + 151, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 151, + 547, + 187 + ], + "type": "text", + "content": "Table 1. Experimental setup. SwinLSTMs denotes the number of the SwinLSTM Cells in predictive network. STB denotes the number of the Swin Transformer blocks in SwinLSTM cell. Patch size indicates the patch token size. Train and Test represent the number of input and output frames during training and testing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 203, + 517, + 389 + ], + "lines": [ + { + "bbox": [ + 337, + 203, + 517, + 389 + ], + "spans": [ + { + "bbox": [ + 337, + 203, + 517, + 389 + ], + "type": "table", + "html": "
MethodMSE ↓SSIM ↑
ConvLSTM [24]103.30.707
DFN [11]89.00.726
FRNN [22]69.70.813
VPN [14]64.10.870
PredRNN [33]56.80.867
CausalLSTM [31]46.50.898
MIM [34]44.20.910
E3D-LSTM [32]41.30.910
LMC [19]41.50.924
MAU [3]27.60.937
PhyDNet [8]24.40.947
CrevNet [39]22.30.949
SwinLSTM17.70.962
", + "image_path": "65da243f3c753efe876d49deb839d6ff451d2c12b6c963bacb20167aa20e6cd1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 456, + 544, + 574 + ], + "blocks": [ + { + "bbox": [ + 305, + 397, + 547, + 443 + ], + "lines": [ + { + "bbox": [ + 305, + 397, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 397, + 547, + 443 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of SwinLSTM and other methods on Moving MNIST. Each model observes 10 frames and predicts the subsequent 10 frames. Lower MSE and higher SSIM indicate better predictions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 456, + 544, + 574 + ], + "lines": [ + { + "bbox": [ + 309, + 456, + 544, + 574 + ], + "spans": [ + { + "bbox": [ + 309, + 456, + 544, + 574 + ], + "type": "table", + "html": "
MethodMAE /100 ↓MSE /10 ↓SSIM ↑
FRNN [22]19.049.80.771
ConvLSTM [24]18.950.40.776
PredRNN [33]18.948.40.781
CausalLSTM [31]17.245.80.851
MIM [34]17.842.90.790
E3D-LSTM [32]16.646.40.869
PhyDNet [8]16.236.90.901
SwinLSTM11.933.20.913
", + "image_path": "f17749da804411e1d7a2c379fc8fce8e7dddb77d0e5853eef218b84594227ea0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 582, + 547, + 604 + ], + "lines": [ + { + "bbox": [ + 305, + 582, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 547, + 604 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation of different models on the Human3.6m dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 547, + 715 + ], + "type": "text", + "content": "method for generating moving MNIST sequences follows [26], each sequence contains 20 frames, and the first 10 frames and the last 10 frames are used for input and target, respectively. The two handwritten digits in each frame are randomly sampled in the MNIST dataset [17], moving and bouncing around a " + }, + { + "bbox": [ + 305, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 305, + 629, + 547, + 715 + ], + "type": "text", + "content": " pixel black canva with a fixed speed and angle. We can get an infinite number of se" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13474" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 339, + 223 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 339, + 223 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 339, + 223 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 339, + 223 + ], + "type": "image", + "image_path": "2c6d5ab6851191cf91ee23c8a403ede3eb6cf8a63f442171d6e1ee4899893af0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 225, + 248, + 237 + ], + "lines": [ + { + "bbox": [ + 180, + 225, + 248, + 237 + ], + "spans": [ + { + "bbox": [ + 180, + 225, + 248, + 237 + ], + "type": "text", + "content": "(a) Moving MNIST" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 354, + 70, + 542, + 222 + ], + "blocks": [ + { + "bbox": [ + 354, + 70, + 542, + 222 + ], + "lines": [ + { + "bbox": [ + 354, + 70, + 542, + 222 + ], + "spans": [ + { + "bbox": [ + 354, + 70, + 542, + 222 + ], + "type": "image", + "image_path": "cf9d06cb8607db915681d9aa75d73ae17745b33c0c0c2407efd6c5e85b2c1459.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 226, + 483, + 236 + ], + "lines": [ + { + "bbox": [ + 444, + 226, + 483, + 236 + ], + "spans": [ + { + "bbox": [ + 444, + 226, + 483, + 236 + ], + "type": "text", + "content": "(b) TaxiBJ" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 247, + 335, + 365 + ], + "blocks": [ + { + "bbox": [ + 52, + 247, + 335, + 365 + ], + "lines": [ + { + "bbox": [ + 52, + 247, + 335, + 365 + ], + "spans": [ + { + "bbox": [ + 52, + 247, + 335, + 365 + ], + "type": "image", + "image_path": "92ee2b4a05418d2827fe2067bae898eb17cd26a8b87d48011f53c7d277b6ba9c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 369, + 261, + 380 + ], + "lines": [ + { + "bbox": [ + 165, + 369, + 261, + 380 + ], + "spans": [ + { + "bbox": [ + 165, + 369, + 261, + 380 + ], + "type": "text", + "content": "(c) KTH (Jogging " + }, + { + "bbox": [ + 165, + 369, + 261, + 380 + ], + "type": "inline_equation", + "content": "10\\rightarrow 20" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 353, + 247, + 535, + 365 + ], + "blocks": [ + { + "bbox": [ + 353, + 247, + 535, + 365 + ], + "lines": [ + { + "bbox": [ + 353, + 247, + 535, + 365 + ], + "spans": [ + { + "bbox": [ + 353, + 247, + 535, + 365 + ], + "type": "image", + "image_path": "95fe46f60339601a28abb7b13977f35de90e2a06f63ccc676bfb5aecb8e009da.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 369, + 493, + 379 + ], + "lines": [ + { + "bbox": [ + 436, + 369, + 493, + 379 + ], + "spans": [ + { + "bbox": [ + 436, + 369, + 493, + 379 + ], + "type": "text", + "content": "(d) Human3.6m" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 53, + 392, + 536, + 509 + ], + "blocks": [ + { + "bbox": [ + 53, + 392, + 536, + 509 + ], + "lines": [ + { + "bbox": [ + 53, + 392, + 536, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 536, + 509 + ], + "type": "image", + "image_path": "cff07aa7e7fd1d7d9d6de58bbd6539daabd873d3eea2f0ef5ce485dc2e04f234.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 513, + 366, + 524 + ], + "lines": [ + { + "bbox": [ + 268, + 513, + 366, + 524 + ], + "spans": [ + { + "bbox": [ + 268, + 513, + 366, + 524 + ], + "type": "text", + "content": "(e) KTH (Walking " + }, + { + "bbox": [ + 268, + 513, + 366, + 524 + ], + "type": "inline_equation", + "content": "10\\rightarrow 40" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 533, + 547, + 588 + ], + "lines": [ + { + "bbox": [ + 45, + 533, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 45, + 533, + 547, + 588 + ], + "type": "text", + "content": "Figure 4. Qualitative results of SwinLSTM on four datasets. The first line is the input, the second line is the ground truth, and the third line is the prediction of SwinLSTM. For Moving MNIST, we add the prediction results of ConvLSTM as the fourth line. For TaxiBJ, we add the absolute difference between the predictions and ground truths as the fourth line. For KTH action, (c) is the jogging action when the model predicts 20 frames based on the previous 10 frames, and (e) is the walking action when the model observes 10 frames and predicts the next 40 frames." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 609, + 288, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 288, + 646 + ], + "type": "text", + "content": "quences by applying different speeds and angles to different handwritten digits. We adopt 10000 sequences for training and a fixed set containing 10000 sequences for testing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 289, + 715 + ], + "type": "text", + "content": "Human3.6m. The human3.6m dataset [10] contains 3.6 million different human poses and corresponding images. Following [34], we only use the 'walking' scenario. Subjects S1, S5, S6, S7, S8 are used for training, and S9, S11 are used for testing. The images in the dataset are resized" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 609, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 545, + 634 + ], + "type": "text", + "content": "from " + }, + { + "bbox": [ + 304, + 609, + 545, + 634 + ], + "type": "inline_equation", + "content": "1000 \\times 1000 \\times 3" + }, + { + "bbox": [ + 304, + 609, + 545, + 634 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 609, + 545, + 634 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\times 3" + }, + { + "bbox": [ + 304, + 609, + 545, + 634 + ], + "type": "text", + "content": ". We train the models to predict the next 4 RGB frames from 4 observations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "KTH. The KTH dataset [23] contains 25 individuals performing 6 categories of human actions (walking, jogging, running, boxing, hand waving, and hand clapping) in 4 different scenarios. We follow the experimental setup in [30], resizing each image to " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " and using persons 1-16 for training and 17-25 for testing. The models predict 10" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13475" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 219 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 219 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 219 + ], + "type": "table", + "html": "
MethodKTH (10 → 20)KTH (10 → 40)
SSIM ↑PSNR ↑SSIM ↑PSNR ↑
ConvLSTM [24]0.71223.580.63922.85
SAVP [7]0.74625.380.70123.97
FRNN [22]0.77126.120.67823.77
DFN [11]0.79427.260.65223.01
PredRNN [33]0.83927.550.70324.16
VarNet [13]0.84328.480.73925.37
SVAP-VAE [18]0.85227.770.81126.18
PredRNN++ [31]0.86528.470.74125.21
E3d-LSTM [32]0.87929.310.81027.24
STMFANet [12]0.89329.850.85127.56
SwinLSTM0.90334.340.87933.15
", + "image_path": "ff62ff564e7132fff56175e92d510bb9f842f2e6bcbe0b42911dd9114fe7c16f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 316 + ], + "type": "text", + "content": "frames from 10 observations at training time and 20 or 40 frames at inference time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 316, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 316, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 45, + 316, + 287, + 423 + ], + "type": "text", + "content": "TaxiBJ. TaxiBJ [40] contains complex real-world taxi trajectory data collected from taxicab GPS monitors in Beijing. Each frame in TaxiBJ is a " + }, + { + "bbox": [ + 45, + 316, + 287, + 423 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 2" + }, + { + "bbox": [ + 45, + 316, + 287, + 423 + ], + "type": "text", + "content": " heat map, where the last dimension represents the flow of traffic entering and leaving the same area. We follow the experimental setup in [40] and use the last four weeks of data for testing and the rest for training. We use 4 observations to predict the next 4 consecutive frames. We adopted per-frame MSE as metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 432, + 129, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 129, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 129, + 443 + ], + "type": "text", + "content": "4.4. Main results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 450, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 450, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 45, + 450, + 287, + 641 + ], + "type": "text", + "content": "Table 2, 3, 4 and 5 present quantitative comparisons with previous state-of-the-art models on four datasets. On Moving MNIST, SwinLSTM outperforms previous CNN-based models and achieves the new state-of-the-art results. Specifically, compared to ConvLSTM [24], SwinLSTM reduces the MSE from 103.3 to 17.7 and increases the SSIM from 0.707 to 0.962. On both Human3.6m and KTH, SwinLSTM achieves the state-of-the-art, especially with huge gains of 4.49 (10→20) and 5.59 (10→40) in PSNR on KTH. In the case of TaxiBJ, our proposed SwinLSTM achieved a remarkable reduction in MSE for each predicted frame compared to previous models. Additionally, the interframe MSE difference was smaller than that of the compared models. The results indicate SwinLSTM's great potential for an efficient and generalizable technique for spatiotemporal prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "In Figure 4, we present qualitative results of SwinLSTM on all datasets. (a) presents a qualitative analysis of SwinLSTM and ConvLSTM on an extremely hard case of Moving MNIST. In the presented images, two digits are continuously intertwined, posing a challenge to the model's ability to accurately predict their motion trajectories and ap" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 176 + ], + "blocks": [ + { + "bbox": [ + 46, + 227, + 288, + 272 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 288, + 272 + ], + "type": "text", + "content": "Table 4. Quantitative evaluation on the KTH test set. We present the model observing 10 frames to predict 20 or 40 frames, and all metrics are averaged over the predicted frames. Higher SSIM and PSNR indicate better prediction quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 176 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 176 + ], + "type": "table", + "html": "
MethodFrame1 ↓Frame2 ↓Frame3 ↓Frame4 ↓
VPN [14]0.7441.0311.2511.444
ST-ResNet [40]0.6880.9391.1301.288
FRNN [22]0.6820.8230.9891.183
PredRNN [33]0.6340.9341.0471.263
PredRNN++ [31]0.6410.8550.9791.158
E3d-LSTM [32]0.6200.7730.8881.158
MIM [34]0.5540.7370.8870.999
SwinLSTM0.3240.4010.4730.525
", + "image_path": "bbc049a94613f712675c108cad2926a1e2d0dc9093025c7f65c89ab8bc9954b7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 184, + 545, + 206 + ], + "lines": [ + { + "bbox": [ + 305, + 184, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 305, + 184, + 545, + 206 + ], + "type": "text", + "content": "Table 5. Quantitative results on the TaxiBJ dataset. We report the per-frame MSE of the 4 predictions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 229, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 396 + ], + "type": "text", + "content": "pearances. As demonstrated, both models effectively capture the trajectory of the digits. However, in terms of appearance, ConvLSTM's prediction results become increasingly blurred as the number of time steps increases, while SwinLSTM's predictions maintain high similarity throughout. SwinLSTM, as an extension of ConvLSTM, utilizes LSTM [9] to capture temporal dependencies, and employs Swin Transformer blocks [20] to model global spatial information. In contrast, ConvLSTM uses convolution to model local spatial information. The significant performance gap between the two models demonstrates that learning global spatial information can better assist the model in capturing spatiotemporal dependencies and thus improve prediction accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 397, + 546, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 546, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 546, + 577 + ], + "type": "text", + "content": "To facilitate the observation of the similarity between the predicted results and the ground truth in (b), we visualize the absolute difference between them. For (c), (d), and (e), we observe that SwinLSTM makes accurate predictions of human actions and scenes, where (c) and (e) show visual examples of 2 different categories of actions predicted by SwinLSTM, Jogging and walking respectively. The model receives 10 frames during training and predicts 10 frames, whereas it predicts 20 or 40 frames during testing. Expanding the predicted number of frames during testing poses a significant challenge to the model, requiring it to effectively learn spatiotemporal dependencies from the spatiotemporal data. Our experiments on the KTH dataset demonstrate that SwinLSTM outperforms alternative models in capturing spatiotemporal dependencies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 586, + 400, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 400, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 400, + 599 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "content": "In this section, we perform ablation studies on TaxiBJ and Human3.6m to analyze the impact of different elements on model performance. Specifically, we discuss three major elements: the reconstruction layer, patch size, and the number of Swin Transformer blocks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "The reconstruction layer. The role of the reconstruction layer is to decode the spatiotemporal representations extracted by the SwinLSTM cell. We conduct experiments on transposed convolution, bilinear interpolation, and lin" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13476" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "ear projection to analyze the impact of different decoding methods. Table 6 shows that the transposed convolution performs much better than the other two methods." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 121, + 285, + 183 + ], + "blocks": [ + { + "bbox": [ + 50, + 121, + 285, + 183 + ], + "lines": [ + { + "bbox": [ + 50, + 121, + 285, + 183 + ], + "spans": [ + { + "bbox": [ + 50, + 121, + 285, + 183 + ], + "type": "table", + "html": "
The reconstruction layerTaxiBJHuman3.6m
MSESSIMMSE /10SSIM
Transposed convolution0.3900.98033.20.913
Bilinear interpolation0.7940.96139.30.887
Linear Projection0.4150.97972.50.773
", + "image_path": "99e5949812878446f56a2371a0fce273024e5b07d76c2a8592c9af73f2b2642d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 230, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 287, + 337 + ], + "type": "text", + "content": "Patch size. The size of the image patch directly determines the length of the input token sequence. The smaller the size, the longer the length. To explore the effect of different patch sizes on prediction accuracy, we conduct experiments with patch sizes set to 2, 4, and 8 on TaxiBJ and Human3.6m. Table 7 shows that the optimal patch size varies across datasets, which indicates that adjusting the patch size to an appropriate size is beneficial to improve the generalization of SwinLSTM." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 60, + 348, + 275, + 422 + ], + "blocks": [ + { + "bbox": [ + 47, + 191, + 286, + 213 + ], + "lines": [ + { + "bbox": [ + 47, + 191, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 286, + 213 + ], + "type": "text", + "content": "Table 6. Ablation study on the three different methods of reconstruction layer with SwinLSTM on TaxiBJ and Human3.6m." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 348, + 275, + 422 + ], + "lines": [ + { + "bbox": [ + 60, + 348, + 275, + 422 + ], + "spans": [ + { + "bbox": [ + 60, + 348, + 275, + 422 + ], + "type": "table", + "html": "
Patch sizeTaxiBJHuman3.6m
MSESSIMMSE /10SSIM
20.4170.97933.20.913
40.3900.98040.20.885
80.3930.98051.90.839
", + "image_path": "39193643c34aa768961dc60e961402c8e3036d86e996648346dd53b0cbd47f72.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 430, + 287, + 452 + ], + "lines": [ + { + "bbox": [ + 47, + 430, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 287, + 452 + ], + "type": "text", + "content": "Table 7. Ablation study on different patch sizes on TaxiBJ and Human3.6m." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 555 + ], + "type": "text", + "content": "The number of Swin Transformer Blocks. Based on the mechanism of Swin Transformer blocks (STB) modeling global spatial information, it is evident that the learning capacity of the model varies with the number of STB utilized. We explored the effect of the number of STB from 2 to 18 on model performance. Figure 5 present the results of different numbers of STB on TaxiBJ and Human3.6m." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 36, + 567, + 161, + 661 + ], + "blocks": [ + { + "bbox": [ + 36, + 567, + 161, + 661 + ], + "lines": [ + { + "bbox": [ + 36, + 567, + 161, + 661 + ], + "spans": [ + { + "bbox": [ + 36, + 567, + 161, + 661 + ], + "type": "image", + "image_path": "2bbfb7588ca692fef4ac54889a5d9903fef69d9834b18da4c5c0170f96cfce41.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 666, + 130, + 677 + ], + "lines": [ + { + "bbox": [ + 87, + 666, + 130, + 677 + ], + "spans": [ + { + "bbox": [ + 87, + 666, + 130, + 677 + ], + "type": "text", + "content": "(a) TaxiBJ" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 687, + 287, + 709 + ], + "lines": [ + { + "bbox": [ + 46, + 687, + 287, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 687, + 287, + 709 + ], + "type": "text", + "content": "Figure 5. Ablation study on the different numbers of STB with SwinLSTM on TaxiBJ and Human3.6m." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 164, + 567, + 289, + 661 + ], + "blocks": [ + { + "bbox": [ + 164, + 567, + 289, + 661 + ], + "lines": [ + { + "bbox": [ + 164, + 567, + 289, + 661 + ], + "spans": [ + { + "bbox": [ + 164, + 567, + 289, + 661 + ], + "type": "image", + "image_path": "0c077b8cec9c4fe7a472412c80103501ec376f5b10eaabda9fcc501e709bfb8a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 666, + 265, + 677 + ], + "lines": [ + { + "bbox": [ + 200, + 666, + 265, + 677 + ], + "spans": [ + { + "bbox": [ + 200, + 666, + 265, + 677 + ], + "type": "text", + "content": "(b) Human3.6m" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 74, + 535, + 369 + ], + "blocks": [ + { + "bbox": [ + 318, + 74, + 535, + 369 + ], + "lines": [ + { + "bbox": [ + 318, + 74, + 535, + 369 + ], + "spans": [ + { + "bbox": [ + 318, + 74, + 535, + 369 + ], + "type": "image", + "image_path": "425f266a1907d3cec76545d82c24edd699af5f0a0814d732a51990a8b5545d90.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 380, + 545, + 447 + ], + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 447 + ], + "type": "text", + "content": "Figure 6. Feature map visualization of a sample in the Moving MNIST test set. Hidden States and Cell States are the feature maps of " + }, + { + "bbox": [ + 305, + 380, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_t" + }, + { + "bbox": [ + 305, + 380, + 545, + 447 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 380, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 305, + 380, + 545, + 447 + ], + "type": "text", + "content": " of the recurrent cell SwinLSTM (section 3.2) at different time steps. STB-2,STB-4, and STB-6 are the feature maps of the 2nd, 4th, and 6th layers of the Swin Transformer blocks (section 3.3), respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 483, + 451, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 451, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 451, + 496 + ], + "type": "text", + "content": "4.6. Feature map Visualization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "To analyze the mechanism of the SwinLSTM cell (section 3.2), we randomly selected a sample from the Moving MNIST test set and visualized the feature maps during inference. Figure 6 shows the results, where white arrows have been added to the images in the third row to facilitate the observation of digit trajectories. The Hidden States capture the position of digits at the current time step, whereas the Cell States memorize the digit trajectory. Meanwhile, we conducted a visualization of the feature maps generated by the second, fourth, and sixth Swin Transformer blocks of the SwinLSTM-B. By observing the feature maps of STB-2 to STB-6, it is evident that STB gradually learns global spatial correlations as the number of interactions of window self-attention increases (section 3.3). By analyzing the visualizations generated by our model, we aimed to gain valuable insights into its performance and behavior. These insights will inform future optimizations and improvements." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13477" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 273 + ], + "type": "text", + "content": "In this paper, we propose SwinLSTM, a new recurrent cell that incorporates Swin Transformer blocks and simplified LSTM, and constructs a predictive network for spatiotemporal prediction tasks based on it. Through extensive experimentation, it has been demonstrated that the proposed method, SwinLSTM, achieves excellent performance on various datasets, including Moving MNIST, TaxiBJ, Human3.6m, and KTH. These impressive results showcase the effectiveness and generalization ability of SwinLSTM. Notably, SwinLSTM outperforms its predecessor, ConvLSTM, in prediction accuracy, which suggests that learning global spatial dependencies allows the model to better capture spatiotemporal dependencies. Overall, this study highlights the potential of SwinLSTM as a promising approach for spatiotemporal modeling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 283, + 145, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 145, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 145, + 298 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 389 + ], + "type": "text", + "content": "This work was supported by the Natural Science Foundation Innovation Research Team Project of Hainan Province (320CXTD431), the National Key Research and Development Program of China (2021YFB1507104), the National Natural Science Foundation of China (32060413), and the Major Science and Technology Project of Hainan Province (ZDKJ2020013), which are gratefully acknowledged." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 400, + 107, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 107, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 107, + 412 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 420, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "type": "text", + "content": "[1] Apratim Bhattacharyya, Mario Fritz, and Bernt Schiele. Long-term on-board prediction of people in traffic scenes under uncertainty. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4194-4202, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 477, + 288, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 288, + 543 + ], + "type": "text", + "content": "[2] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Computer Vision-ECCV 2022 Workshops: Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, pages 205-218. Springer, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 599 + ], + "type": "text", + "content": "[3] Zheng Chang, Xinfeng Zhang, Shanshe Wang, Siwei Ma, Yan Ye, Xiang Xinguang, and Wen Gao. Mau: A motion-aware unit for video prediction and beyond. Advances in Neural Information Processing Systems, 34:26950-26962, 2021. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 601, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 287, + 643 + ], + "type": "text", + "content": "[4] Yunpeng Chen, Yannis Kalantidis, Jianshu Li, Shuicheng Yan, and Jiashi Feng. A^2-nets: Double attention networks. Advances in neural information processing systems, 31, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 646, + 287, + 689 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.2,4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 118, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 547, + 171 + ], + "type": "text", + "content": "[7] Hang Gao, Huazhe Xu, Qi-Zhi Cai, Ruth Wang, Fisher Yu, and Trevor Darrell. Disentangling propagation and generation for video prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9006-9015, 2019. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 172, + 547, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 172, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 312, + 172, + 547, + 226 + ], + "type": "text", + "content": "[8] Vincent Le Guen and Nicolas Thome. Disentangling physical dynamics from unknown factors for unsupervised video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11474-11484, 2020. 1, 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 228, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 545, + 249 + ], + "type": "text", + "content": "[9] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 250, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 304 + ], + "type": "text", + "content": "[10] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 304, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 337 + ], + "type": "text", + "content": "[11] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 338, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 547, + 403 + ], + "type": "text", + "content": "[12] Beibei Jin, Yu Hu, Qiankun Tang, Jingyu Niu, Zhiping Shi, Yinhe Han, and Xiaowei Li. Exploring spatial-temporal multi-frequency analysis for high-fidelity and temporal-consistency video prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4554-4563, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "type": "text", + "content": "[13] Beibei Jin, Yu Hu, Yiming Zeng, Qiankun Tang, Shice Liu, and Jing Ye. Varnet: Exploring variations for unsupervised video prediction. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5801-5806. IEEE, 2018. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[14] Nal Kalchbrenner, Aäron Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyls, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. In International Conference on Machine Learning, pages 1771-1779. PMLR, 2017. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "type": "text", + "content": "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 526, + 547, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 547, + 569 + ], + "type": "text", + "content": "[16] Yong-Hoon Kwon and Min-Gyu Park. Predicting future frames using retrospective cycle gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1811-1820, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 570, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 592 + ], + "type": "text", + "content": "[17] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 593, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 634 + ], + "type": "text", + "content": "[18] Alex X Lee, Richard Zhang, Frederik Ebert, Pieter Abbeel, Chelsea Finn, and Sergey Levine. Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "text", + "content": "[19] Sangmin Lee, Hak Gu Kim, Dae Hwi Choi, Hyung-II Kim, and Yong Man Ro. Video prediction recalling long-term motion context via memory alignment learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3054-3063, 2021. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13478" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2, 4, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "text", + "content": "[21] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. Advances in neural information processing systems, 29, 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "text", + "content": "[22] Marc Oliu, Javier Selva, and Sergio Escalera. Folded recurrent neural networks for future video prediction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 716-731, 2018. 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[23] Christian Schuldt, Ivan Laptev, and Barbara Caputo. Recognizing human actions: a localsvm approach. In Proceedings of the 17th International Conference on Pattern Recognition, 2004.ICPR 2004., volume 3,pages 32-36.IEEE,2004.5,6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "text", + "content": "[24] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems, 28, 2015. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 353 + ], + "type": "text", + "content": "[25] Xingjian Shi, Zhihan Gao, Leonard Lausen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Deep learning for precipitation nowcasting: A benchmark and a new model. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "text", + "content": "[26] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning, pages 843-852. PMLR, 2015. 1, 2, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 400, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 454 + ], + "type": "text", + "content": "[27] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 520 + ], + "type": "text", + "content": "[28] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12894–12904, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 611 + ], + "type": "text", + "content": "[30] Ruben Villegas, Jimei Yang, Seunghoon Hong, Xunyu Lin, and Honglak Lee. Decomposing motion and content for natural video sequence prediction. arXiv preprint arXiv:1706.08033, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[31] Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and S Yu Philip. Predrnn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning. In International Conference on Machine Learning, pages 5123-5132. PMLR, 2018. 1, 2, 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[32] Yunbo Wang, Lu Jiang, Ming-Hsuan Yang, Li-Jia Li, Mingsheng Long, and Li Fei-Fei. Eidetic 3d LSTM: A model for video prediction and beyond. In International conference on learning representations, 2018. 1, 2, 5, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 630 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[33] Yunbo Wang, Mingsheng Long, Jianmin Wang, Zhifeng Gao, and Philip S Yu. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in neural information processing systems, 30, 2017. 1, 2, 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "type": "text", + "content": "[34] Yunbo Wang, Jianjin Zhang, Hongyu Zhu, Mingsheng Long, Jianmin Wang, and Philip S Yu. Memory in memory: A predictive neural network for learning higher-order nonstationarity from spatiotemporal dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9154-9162, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 196, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 238 + ], + "type": "text", + "content": "[35] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 283 + ], + "type": "text", + "content": "[36] Yu-Huan Wu, Yun Liu, Xin Zhan, and Ming-Ming Cheng. P2t: Pyramid pooling transformer for scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 339 + ], + "type": "text", + "content": "[37] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 373 + ], + "type": "text", + "content": "[38] Ziru Xu, Yunbo Wang, Mingsheng Long, Jianmin Wang, and M KLiss. Predcnn: Predictive learning with cascade convolutions. In IJCAI, pages 2940-2947, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 375, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 407 + ], + "type": "text", + "content": "[39] Wei Yu, Yichao Lu, Steve Easterbrook, and Sanja Fidler. Efficient and information-preserving future frame prediction and beyond. 2020. 1, 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "type": "text", + "content": "[40] Junbo Zhang, Yu Zheng, and Dekang Qi. Deep spatiotemporal residual networks for citywide crowd flows prediction. In Thirty-first AAAI conference on artificial intelligence, 2017. 1, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 454, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 508 + ], + "type": "text", + "content": "[41] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3008, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 553 + ], + "type": "text", + "content": "[42] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 555, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 630 + ], + "type": "text", + "content": "[43] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6881-6890, 2021. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13479" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/f251ae55-2ca7-4578-ad92-a51119c290a6_content_list.json b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/f251ae55-2ca7-4578-ad92-a51119c290a6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..559ed921c237baa128e88d59c102b04805037f67 --- /dev/null +++ b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/f251ae55-2ca7-4578-ad92-a51119c290a6_content_list.json @@ -0,0 +1,1747 @@ +[ + { + "type": "text", + "text": "SynBody: Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling", + "text_level": 1, + "bbox": [ + 174, + 130, + 795, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhitao Yang $^{1,*}$ Zhongang Cai $^{1,2,3,*}$ Haiyi Mei $^{1,*}$ Shuai Liu $^{2,*}$ Zhaoxi Chen $^{3,*}$ Weiye Xiao $^{1}$ Yukun Wei $^{1}$ Zhongfei Qing $^{1}$ Chen Wei $^{1}$ Bo Dai $^{2}$ Wayne Wu $^{2}$ Chen Qian $^{1}$ Dahua Lin $^{4}$ Ziwei Liu $^{3,\\dagger}$ Lei Yang $^{1,2,\\dagger}$ SenseTime Research Shanghai AI Laboratory \n $^{3}$ S-Lab, Nanyang Technological University The Chinese University of Hong Kong *Equal Contribution Corresponding Author https://synbody.github.io/", + "bbox": [ + 138, + 202, + 825, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d714c9f937ad5905a6fd480c195d1321a856f80ab64e94095837f3ce9f8b99c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 345, + 594, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fdc330cf86a5e26591a6b140ceaa7afc71afcd6ea8156a3482eaf86df79c5644.jpg", + "image_caption": [ + "Human Pose and Shape Estimation" + ], + "image_footnote": [], + "bbox": [ + 99, + 561, + 331, + 662 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5bcf5e98b5647f462842e1562e5caf76c30415ebdf7ee24014c0147c4b26733b.jpg", + "image_caption": [ + "Novel View Synthesis" + ], + "image_footnote": [], + "bbox": [ + 344, + 564, + 588, + 659 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9a25979a5c4acb49c79e014b8a3b17f6e74815aa26d26cad2be2d9a73cc66e5.jpg", + "image_caption": [ + "Various Environments" + ], + "image_footnote": [], + "bbox": [ + 594, + 345, + 864, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79763334979363e9def77ecf007ce4b0df9ecd49bde7e3d9c5e62365f05f83a7.jpg", + "image_caption": [ + "Layered Human Models" + ], + "image_footnote": [], + "bbox": [ + 637, + 411, + 820, + 541 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66844f706f30d6da5ef784237c2b597e83eb7aefc0807140d7e7b615db517eab.jpg", + "image_caption": [ + "Rich Annotations", + "Figure 1: SynBody is a large-scale synthetic dataset with a massive number of subjects and high-quality annotations. It supports various research topics, including human pose and shape estimation (HPS) and novel view synthesis for human (Human NeRF)." + ], + "image_footnote": [], + "bbox": [ + 622, + 561, + 841, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 720, + 313, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synthetic data has emerged as a promising source for 3D human research as it offers low-cost access to large-scale human datasets. To advance the diversity and annotation quality of human models, we introduce a new synthetic dataset, SynBody, with three appealing features: 1) a clothed parametric human model that can generate a diverse range of subjects; 2) the layered human representation that naturally offers high-quality 3D annotations to support multiple tasks; 3) a scalable system for produc", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing realistic data to facilitate real-world tasks. The dataset comprises 1.2M images with corresponding accurate 3D annotations, covering 10,000 human body models, 1,187 actions, and various viewpoints. The dataset includes two subsets for human pose and shape estimation as well as human neural rendering. Extensive experiments on SynBody indicate that it substantially enhances both SMPL and SMPL-X estimation. Furthermore, the incorporation of layered annotations offers a valuable training resource for investigating the Human Neural Radiance Fields(NeRF).", + "bbox": [ + 496, + 722, + 893, + 875 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20282", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 80, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The fields of 3D human perception [21, 23-25, 37, 53] and human reconstruction [13, 14, 26, 40, 41] have become increasingly important, but the lack of available data has limited their development. Collecting real human data on a large scale is challenging due to privacy concerns and time constraints. Therefore, exploring the use of synthetic human datasets has become a critical avenue of research.", + "bbox": [ + 80, + 114, + 467, + 219 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the great potential, existing synthetic datasets [5, 8,38,47] suffer from limitations such as the number of available human models and the quality of annotations. The main reason lies in that synthetic human datasets rely on real scans for rendering, which poses three key obstacles. Firstly, it is challenging to expand the types of body shapes, poses, and clothing available in the dataset. Secondly, as the human models are scanned with clothing, the 3D annotations obtained through fitting are prone to errors. Thirdly, it is difficult to obtain annotations of body and clothing separately. To address these issues, we develop a new synthetic dataset termed SynBody. The dataset includes 1.2 million frames with corresponding ground-truth 3D human body annotations. It covers 10,000 human body models, 1,187 motions, and 26,960 video clips with 2.7M SMPL/SMPL-X annotations.", + "bbox": [ + 80, + 220, + 467, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At the heart of SynBody is the layered parametric human model, which constructs the clothed human model in a bottom-up manner. SMPL-X [39] is a widely used parametric human model, capable of sampling human models with various body shapes. However, it lacks the ability to model clothing, limiting its applicability when synthesizing realistic human models. To overcome this limitation, we introduce SMPL-XL, a parametric human model based on SMPL-X in a layered representation. SMPL-XL enriches the SMPL-X model in three aspects: (1) Hair system: adding hair and beards to the FLAME [27] model, with 32 types of hair and 13 types of beards; (2) Garment and accessories: adding procedural clothes to the SMPL-X body, including coats, shirts, pants, skirts, shoes, and glasses; (3) Texture: in addition to adding rich geometry, SMPL-XL also adds rich textures for sampling various skin colors and clothing textures.", + "bbox": [ + 80, + 463, + 467, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The designed SMPL-XL is capable of automatically generating a large number of human models with high-quality annotations. We therefore generate 10,000 clothed human models by sampling various body shapes, clothing styles, hairstyles, accessories, and textures. Notably, the use of the SMPL-X model as the base body model guarantees that the parametric human annotations are always accurate, obviating the need for the necessity for annotations through fitting. Furthermore, as the clothes are explicitly attached to the surface of the human body, layered annotations for body and clothes are available.", + "bbox": [ + 80, + 719, + 467, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To generate a large-scale dataset with high diversity and", + "bbox": [ + 99, + 885, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "high-quality annotations, we design a scalable and automatic system to render images and annotations. We first animate the 10,000 dressed human models by retargeting motions from a large motion library [31]. Subsequently, we design an algorithm to place human models in the scene without piercing. Multiple cameras are then placed by evaluating self-occlusion, inter-occlusion, and view diversity, and the rendering module renders the assets into images with corresponding annotations.", + "bbox": [ + 503, + 90, + 888, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With SynBody, we launch two tracks that support human pose and shape estimation and human neural rendering, respectively. Experiments show that SynBody is more effective than AGORA under the same amount of training data for human pose and shape estimation. With diverse and large-scale training data, SynBody achieves significant performance gains on both SMPL and SMPL-X estimation. In terms of human neural rendering using neural radiance fields (i.e. NeRF [34]), benchmarking existing approaches on SynBody shows that it has comparable performance as real human data. Furthermore, with the layered annotations which offer accurate SMPL parameters, we observe that current human NeRF approaches are sensitive to the accuracy of estimated SMPL.", + "bbox": [ + 503, + 227, + 888, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, SynBody is a large-scale synthetic dataset for human perception and modeling, with three main contributions: (1) It constructs clothed subjects and samples 10,000 animatable subjects, which is an order of magnitude higher than existing datasets. (2) The clothed subjects are constructed with an explicit cloth model, thus it provides layered 3D annotations of the human body and clothing, which are not available in previous datasets. (3) Experiments on SynBody achieve promising results on both human perception and modeling, emphasizing the importance of diversity and annotation quality for downstream tasks.", + "bbox": [ + 503, + 439, + 888, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 503, + 618, + 645, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Parametric Models. Several 3D human parametric models, such as SMPL [30], SMPL-X [39], and GHUM [51], have been developed to generate 3D human meshes from parameters that represent the human pose and shape using linear blend skinning. SMPL-X [39] extends SMPL [30] by combining FLAME [27] and MANO [43] for the head and hands, respectively, and is trained on a large number of real scans to provide a strong basis for shape variations. However, SMPL-X only produces naked body meshes, and we aim to enhance its realism by building a layered parametric model that includes hair, clothes, and accessories. The proposed model leverages the shape basis of SMPL-X while providing realistic dressed human meshes.", + "bbox": [ + 503, + 643, + 888, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Pose and Shape Estimation. Several methods [21, 23-25, 37, 49, 52-54] have been proposed to estimate 3D human pose and shape parameters. HMR [21] directly regresses these parameters in an end-to-end manner, while", + "bbox": [ + 503, + 840, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20283", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/758d1e68346d5545e2d8df43abcd27bb85831b975a04687e8f610493b5ae2f6e.jpg", + "table_caption": [ + "Table 1: Comparisons of 3d human dataset. We compare SynBody with existing datasets. We divide datasets into three types: real (R), synthetic (S), and mixed (M). SynBody constructs 10,000 animatable subjects, which is an order of magnitude higher than any existing datasets and brings competitive scale and diversity. \"ITW\" stands for \"In-the-Wild\" in the table." + ], + "table_footnote": [], + "table_body": "
DatasetTypeITWVideo#Views#SMPL#Seq#Subj.#MotionsGT format
HumanEva [45]R-✓4/7NA7463DJ
Human3.6M [15]R-✓4312K83911153DJ, SMPL
MPI-INF-3DHP [32]M✓✓1496K16883DJ
3DPW [48]R✓✓132K6018*SMPL
Panoptic Studio [20]R-✓480736K480~100*3DJ
EFT [19]R✓-1129KNAManyNASMPL
ZJU-MoCap [41]R-✓21180K999SMPL,mask
SURREAL [47]S✓✓16.5MNA1452KSMPL
AGORA [38]S✓-1173KNA>350NASMPL, SMPL-X, mask
HSPACE [5]S✓✓5-NA100×16100GHUM/L, mask
GTA-Human [8]S✓✓11.4M20K>60020KSMPL
BEDLAM [6]S✓✓-1M10K2712,311SMPL-X, mask
SynBodyS✓✓42.7M27K10,0001,187SMPL, SMPL-X, mask
", + "bbox": [ + 96, + 145, + 875, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SPIN uses an optimization step [25] to guide the learning process towards pseudo-3D labels. PARE [24] employs part attention to tackle occlusion, and VIBE [23] leverages temporal information in videos for SMPL estimation. Apart from building low-cost data collection solutions [7], synthetic datasets have become a promising alternative to efficiently scale up data. SURREAL [47] renders textured SMPL body models in real-image backgrounds but does not account for cloth geometry, resulting in unrealistic subjects. AGORA [38] renders real human scans in a virtual world and provides high-quality synthetic data for image-based approaches. HSPACE [5] places animated human models in various scenes to provide training data for video-based methods, and increases the variation of human shape via refitting. GTA-Human [8] captures videos and optimizes corresponding SMPL annotations from video games. However, the diversity of subjects in current datasets is limited by either body shapes or cloth types.", + "bbox": [ + 75, + 376, + 472, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Expressive Human Pose and Shape Estimation. As face and hand are also crucial for human perception, some efforts [35,44,55] have been made to whole-body human pose and shape estimation. ExPose [9] introduces three experts to predict parameters for body, hand, and face, and merge them in a copy-paste strategy. Hand4Whole [35] further improves the prediction of wrist and finger poses by leveraging selected hand joint features. Predicting both hands and faces makes the dataset much more difficult to obtain than just predicting the body. In order to increase the diversity of real data, AGORA [38] provides image-based SMPL-X annotations by fitting SMPL-X on the scanned human models. While a very recent work, BEDLAM [6], also enhances the SMPL-X model with clothing, hair, and changes in skin tones. However, it employs a manual approach for creating garments and utilizes pre-made skin textures, making", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "it challenging to generate a large number of human body models. In contrast, our body model construction process is fully automated, encompassing clothing, hair, textures, and more. This enables easier scalability to generate human body models on a large scale.", + "bbox": [ + 496, + 376, + 893, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human NeRF. NeRF [34] has demonstrated impressive photo-realistic view synthesis by learning implicit fields of density and color. Yet, human motions are more challenging to learn due to dynamic deformation fields. NeuralBody [41] incorporates prior from a statistical body template to learn dynamic sequence, while Animatable NeRF [40] proposes to reconstruct an animatable human model that generalizes to new poses. Furthermore, NHP [26] and KeypointNeRF [33] achieve generalizability to unseen identities and poses. Several datasets have been adapted to study human NeRF. ZJU-MoCap [41] captures 9 human subjects with 21 synchronized cameras, providing fitted human body model parameters as well as the foreground mask. Human3.6M [16] collects 11 human subjects with 4 cameras, using a marker-based motion capture system. A-NeRF [46] generates a synthetic dataset using SURREAL [47] to study factors that affect visual quality.", + "bbox": [ + 496, + 455, + 895, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the aforementioned tasks are interrelated and often draw upon similar datasets, as shown in Table 1, datasets are limited in two key aspects. Firstly, obtaining real human models is challenging, which restricts the scale of these datasets. Secondly, 3D annotations are typically acquired through optimization, which introduces errors and cannot provide layered annotations. In contrast, with the designed SMPL-XL model, SynBody provides 10,000 subjects with diverse body shapes and clothing, along with layered annotations that include accurate SMPL and SMPL-X. Built on top of this layered human model, SynBody comprises two subsets for human mesh recovery and human NeRF, respec", + "bbox": [ + 496, + 719, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20284", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tively, and features the same 10,000 diverse subjects.", + "bbox": [ + 76, + 90, + 426, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Synthetic Data Generation System", + "text_level": 1, + "bbox": [ + 76, + 119, + 390, + 137 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fb00fd8f6e55eb1b726a486e2ce356fe79dcb390b82dd46f504310ab71d16968.jpg", + "image_caption": [ + "Figure 2: Synthetic data generation system. It consists of 4 components: 1) human model creation to generate layered human models, 2) motion retargeting to drive human models, 3) scene composition to place actors and cameras, and 4) rendering and outputs to generate multi-modal dataset." + ], + "image_footnote": [], + "bbox": [ + 76, + 157, + 467, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to movie production [4], our system comprises 4 components (depicted in Figure 2): (1) a layered parametric human model creation service, a scalable process to generate layered human models, (2) a motion retargeting module to apply motions from various sources to layered human models, (3) a scene composition module to place 3D actors and objects into a 3D scene, setting up cameras, (4) a 3D rendering engine and a multi-modal data annotation generator. Our infrastructure enables the generation of high-quality synthetic data for various computer vision tasks.", + "bbox": [ + 75, + 353, + 468, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Layered Parametric Human Model", + "text_level": 1, + "bbox": [ + 76, + 513, + 383, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parametric human models like SMPL-X [39] provide the ability of create rigged body models with various body shapes. However, the lack of available textures limits its application in data generation. We designed a module to perform an automatic process that combines SMPL-X with procedural garments and accessories, hair system and textures, producing realistic and diverse body models.", + "bbox": [ + 75, + 537, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Body shape. SMPL-X [39] body model has a 3D mesh whose vertex locations are controlled by parameters for pose $\\theta$ , shape $\\beta$ , and facial expression $\\psi$ . By modifying shape $\\beta$ , we obtain 3D meshes of various human heights and weights. And by alternating pose $\\theta$ , meshes can be driven to perform various poses.", + "bbox": [ + 75, + 643, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Garment Model. Our garment is generated as a separate layer on top of the body. Following the industrial garment-making workflow, we designed garment patterns of various styles. Notice that different parts of garment pieces are connected with sewing lines, e.g., the red line in Figure 4 (a). Then, we stitch patterns onto the body at T-Pose utilizing a physical simulator [1]. Specifically, we first manually move the garment pieces to roughly align them with the body. During simulation, vertices between two ends of the seam gradually shrink until they completely pooled together. Figure 4 (b) demonstrates the final draped gar", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ment. For garment animation, we bind every vertex in the garment to the closest point in the body. Then, the skinnig weight and blend shape of the body mesh are assigned to garment vertices. This makes our garment model easily integrated with existing skeletal pipelines with little computational overhead.", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Particle hair system. Our method uses a prefabricated particle system to generate realistic hair on a head-shaped mesh. To achieve this, a template of hairstyle or facial hair $T_{hair}$ would attach to the vertices on the mesh which are marked with different areas $V_{hair}$ (including fringe, top, temporal bone, occipital bone, and bottom area). Designers draw multiple sets of guidelines $L_{guide}$ with varying shapes on different areas. Each set of guidelines comprises a collection of Bezier curves that are utilized to accurately constrain the flow of hair strands. Furthermore, the shape of the hair strands can be adjusted by length $P_{length}$ and curliness $P_{curliness}$ . The entire hair system is composed of multiple sub-particle systems $\\{V_{hair}, L_{guide}, P_{length}, P_{curliness}\\}$ .", + "bbox": [ + 496, + 181, + 892, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accessories. We also add template accessories $T_{\\text{accessories}}$ to our model, such as glasses, shoes, hats, and headphones. All accessories are pre-assembled on an SMPL-X template with a uniform body shape. These accessories can be transferred to models with the same topology, ensuring consistent deformation across all models in accordance with changes in the shape $\\beta$ of SMPL-X. The corresponding bone weights of the human body model's vertices are transferred to the nearest accessories' vertices, ensuring that these accessories are correctly driven by the armature.", + "bbox": [ + 496, + 378, + 890, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Procedural texture. The procedural garment textures $T_{\\text{procedural}} = \\{T_{\\text{pattern}}, T_{\\text{decals}}, T_{\\text{bump}}, P_{\\text{mapping}}\\}$ used in clothing are composed of multiple pre-set textures by alpha blending as demonstrated in Figure 4 (c). The pattern texture $T_{\\text{pattern}}$ and decals texture $T_{\\text{decals}}$ serve as layered masks to elevate the style and color, while the bump texture $T_{\\text{bump}}$ functions as a height map, adding detailed normal information to the clothing's surface. The mapping parameters $P_{\\text{mapping}}$ control the coordination of all the textures in the UV space. Besides, to procedurally create body mesh textures, we build a similar texture template SMPL-X, in which we pre-draw layered masks to separate different body features, such as skin, lips, and freckles, allowing for color adjustment and blending in specific regions as in Figure 3.", + "bbox": [ + 496, + 531, + 892, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining all the elements above, as shown in Figure 5, we obtain human body models with the same high quality as RenderPeople [2]. Besides, it is challenging to expand the types of body shape and clothing available for real scans, but our model can be easily scaled by randomly sampling each component.", + "bbox": [ + 496, + 743, + 890, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Motion Retargeting", + "text_level": 1, + "bbox": [ + 500, + 845, + 687, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Retargeting of skeletal animations. Our motion retargeting module allows for the transfer of motion data from vari", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20285", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0b734aa6d69cd069dfd6dde50d42f857708d02b1bd3d32670ae5dff13f7e59d1.jpg", + "image_caption": [ + "(a) Naked human model generation" + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 439, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e8bd40e463f3736985786ed243195e9f19de9a3c722317ffea4cbd0a7965b967.jpg", + "image_caption": [ + "(b) Layered human model generation" + ], + "image_footnote": [], + "bbox": [ + 547, + 87, + 880, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ccf98bd231a782d1e61ecc3c19426e73222c0eeaf820d538e1b0e3bea8563a85.jpg", + "image_caption": [ + "Figure 3: Layered parametric human model creation: (a) Generation of a naked human model using parametric body model [39], procedural body texture colors, and particle hair System; (b) Integration of garments and accessories onto the naked body model.", + "Figure 4: Demonstrations of Garment and Procedural Texture. (a) garment pattern with sewing lines, (b) simulated garment under the body in canonical space, (c) procedural textured garment whose style and color are elevated by layered masks provided by patterns and decals." + ], + "image_footnote": [], + "bbox": [ + 81, + 303, + 460, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ous sources, such as academic motion datasets, motion captures, and artist-crafted sources, to SMPL-XL model skeletons. Despite variations in bone names, bone lengths, and rest pose bone rotations, the source skeletons are typically structurally similar to the target skeletons.", + "bbox": [ + 75, + 590, + 468, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose frames in motion clips contain root translation $T(t) \\in \\mathbb{R}^3$ and rotations of each bone in the corresponding parent bone's space $\\{R_i(t) \\in \\mathbb{R}^3 | i = 0,1,\\dots,n\\}$ . Following the forward kinematics (FK) manner, each bone's rotation relative to the model space can be calculated:", + "bbox": [ + 75, + 666, + 470, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {R} _ {i} (t) = \\widehat {R} _ {p (i)} (t) \\cdot R _ {i} (t) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 747, + 468, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widehat{R}_i(t)$ is the rotation of $R_{i}(t)$ in model space at $t$ frame, and $p(i)$ indicates the parent of $i$ . Specially, root bones have no parents so $\\widehat{R}_0(t) = R_0(t)$ .", + "bbox": [ + 75, + 777, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To retarget motion from one skeleton to another, we assume the source and the target motion can drive the corresponding bones of both skeletons to the same rotation in model space. We have T-posed frames of skeletons by manually posing them as T-posing, and treat T-posed as all mo", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/06e4ca65c4543ff09c958eb3f037498fd5e1b563189f7d448867a30383c07b32.jpg", + "image_caption": [ + "Figure 5: A Demonstration of comparison between commonly used human models in existing synthetic datasets [5, 8,38,47] and our SMPL-XL. We obtain high-quality models equal to RenderPeople [2], both are much better than SUR-REAL [47], and our model has the capability of scaling up easily by sampling various assets and body shapes." + ], + "image_footnote": [], + "bbox": [ + 511, + 303, + 883, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tions' first frame pose. So motions relative to their T-posedr frame can be easily obtained:", + "bbox": [ + 496, + 592, + 890, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {R} _ {i} (t) = \\widehat {R} _ {i} (0) \\cdot \\widehat {R} _ {i _ {- s r c}} (t) \\cdot \\widehat {R} _ {i _ {- s r c}} ^ {- 1} (0). \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 628, + 890, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering skeletons have different bone lengths, which could result in \"sliding feet\" artifacts on target skeletons. We simply scale $T$ according to the ratio of pelvis bones' height to mitigate them: $T(t) = \\frac{H_{pelvis}}{H_{pelvis,src}} \\cdot T_{src}(t)$ .", + "bbox": [ + 496, + 655, + 890, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SMPL-X Annotations. Considering SMPL-XL body shapes are sampled from SMPL-X, the shape $\\beta$ can be derived directly from the corresponding SMPL-X. Pose $\\theta$ in the model space is calculated in the motion retargeting module. And in the scene composition module, models are placed in a 3D scene with camera space locations $T_{c}$ and rotations $R_{c}$ . Pose $\\theta_{c}$ in camera space is calculated by applying those world transformations to $\\theta$ . So SMPL-X annotations are constructed with $\\{\\beta, \\theta_{c}\\}$ .", + "bbox": [ + 496, + 720, + 892, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SMPL Annotations. Although SMPL-XL naturally provides accurate SMPL-X annotations, SMPL cannot be derived directly. Thus, we need to refit the SMPL parameters.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20286", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The optimization process consists of two steps. First, we fit the shape $\\beta$ of all human models under the T-pose to its corresponding SMPL-X. Secondly, for each sequence, it is initialized with fitted $\\beta$ and its original pose. we fix $\\beta$ while fitting the body pose $\\theta$ for SMPL.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Scene Placement", + "text_level": 1, + "bbox": [ + 76, + 175, + 243, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To place $N_{p}$ subjects in a large scene with $N_{o}$ objects, we primarily follow three principles: standing on the ground, avoiding human-object and human-human penetration. To prevent subjects from floating in the air, the root position of a subject should align with the ground height. A sequential decision-making approach is used to find a suitable position for each subject, i.e., placing one subject at a time. An object is represented by $o_{i} = \\{x_{i},y_{i},l_{i},w_{i}\\} \\in \\mathbb{R}^{4}$ , where $\\{x_i,y_i\\}$ is the center of the object's axis-aligned bounding box projected onto the ground with length $l_{i}$ and width $w_{i}$ . Different from static objects, to avoid collisions between moving subjects at any one time, a subject with specified body shape and motion is simplified to $q_{i} = \\{x_{i},y_{i},l_{i},w_{i}\\} \\in \\mathbb{R}^{4}$ , which is the smallest axis-aligned box that envelops all bounding boxes across frames.", + "bbox": [ + 75, + 199, + 468, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To avoid human-object penetration and potential human-human collision, the solution $p_i^* = \\{x_i^*, y_i^*\\}$ of a character with the shape of $\\{l_i, w_i\\}$ should satisfy:", + "bbox": [ + 75, + 426, + 468, + 470 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(\\left\\{x _ {i} ^ {*}, y _ {i} ^ {*}, l _ {i}, w _ {i} \\right\\}, o _ {j}\\right) = 0, j = 1, \\dots , N _ {o} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 481, + 468, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(\\left\\{x _ {i} ^ {*}, y _ {i} ^ {*}, l _ {i}, w _ {i} \\right\\}, q _ {k}\\right) = 0, k = 1, \\dots , N _ {p}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 523, + 468, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I(\\text{box}_1, \\text{box}_2)$ denotes the overlapping area between two boxes and $N_p$ is the number of subjects already placed in the scene. Besides, distance constraint is used to prevent subjects from excessive dispersal. The problem is solved by grid search. Typically multiple solutions are available and we randomly sample one each time.", + "bbox": [ + 75, + 547, + 468, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Camera Placement", + "text_level": 1, + "bbox": [ + 76, + 646, + 261, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once the positions of all subjects have been organized, the next task involves placing $N_{c}$ cameras in suitable locations. To ensure that cameras would not be placed inside any objects or subjects, the 3D version of Eq. (3) and Eq. (4) are applied. In addition, we evaluate the suitability of a candidate camera by the following metrics: the distance from the camera to the subjects, the camera's pitch angle, and the degree of occlusion of each subject in the camera's view.", + "bbox": [ + 75, + 670, + 468, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The distance denoted as $L$ from the mean position of all subjects $\\bar{p}$ to the camera is restricted. $L_{max}$ is set to prevent an unreasonably small proportion of subjects in the image. To control the visibility of subjects, $L_{min}$ is defined as $\\frac{\\lambda}{\\sin(\\alpha / 2)}\\max_i||pv_i - \\bar{p} ||_2$ , where $i = 1,\\dots ,N_c$ , $\\alpha$ is the field of view of the camera, and $\\lambda$ is a hyperparameter that determines the probability of all subjects being within", + "bbox": [ + 75, + 790, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the camera's view. To estimate the degree of occlusion for each subject, rays are randomly and uniformly cast from the camera to each subject's body, and the percentage of blocked rays by other objects is determined. More details can be found in the Sup. Mat.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Rendering and Annotations", + "text_level": 1, + "bbox": [ + 500, + 178, + 750, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using a high-quality rendering pipeline in Unreal Engine 5, SynBody is rendered in multi-view with large 3D environments from the Unreal Marketplace for rich background information and dynamic lighting. Leveraging the G-buffer [12], our system simultaneously generates photorealistic RGB images and annotations, along with accurate ground truth for segmentation masks, optical flow, depth maps, normal maps, and other ground-truth labels.", + "bbox": [ + 496, + 200, + 893, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. SynBody Dataset", + "text_level": 1, + "bbox": [ + 500, + 332, + 668, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset Statistics", + "text_level": 1, + "bbox": [ + 500, + 358, + 668, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SynBody comprises 10,000 unique SMPL-XL models randomly created with different body shapes and genders. Each model is then combined with the following assets: (1) hairstyles $T_{hair}$ sampled from 45 particle hairs; (2) garments $G_{tmp}$ sampled from 68 clothing models containing multiple outfits; (3) procedural texture $T_{procedural}$ generated by sampling $T_{pattern}$ , $T_{decals}$ , and $T_{bump}$ from 1,038 template textures, and color values are randomly sampled; (4) accessories $T_{accessories}$ sampled from 46 template assets. To ensure the validity of the motions, we select a subset from AMASS [31]. Based on the BABEL annotations [42], we excluded interactive motions and non-ground motions(e.g., swim), as well as filtered out motions with a duration shorter than 2 seconds, leading to a subset with 1,187 motions. For each sequence, we randomly selected 4 human models, and each model is assigned a randomly chosen motion lasting between 2 and 10 seconds. For motions exceeding 2 seconds, we randomly extract a 2-second segment to ensure that each video clip has a length of 60 frames. To enhance the diversity of viewpoints, 4 view positions are generated for each sequence, where each one is randomly sampled from the surface of semi-spheres with varying radii. The system generates 26,960 sequences and 1.2M images with annotations, utilizing 6 vast and realistic scenes created by professional artists. 2.7M SMPL/SMPL-X annotations are provided, excluding highly occluded subjects. More details can be found in the Sup. Mat.", + "bbox": [ + 496, + 381, + 893, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Human Pose and Shape Estimation", + "text_level": 1, + "bbox": [ + 500, + 801, + 807, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With 2.7 million SMPLs, we leverage a pretrained regressor to extract 3D keypoints and then project them onto 2D space. Following standard practice in top-down human mesh recovery methods, we generate bounding boxes from the resulting 2D keypoints.", + "bbox": [ + 496, + 824, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "footer", + "text": "20287", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6701fe21f33d247dbaffe8211ed43442e0f78f54bb9c5b48c76b786b4369e12e.jpg", + "table_caption": [ + "Table 2: Training popular baseline methods (image-based and video-based) with SynBody on 3DPW test set [48]. R: standard real datasets. S: SynBody." + ], + "table_footnote": [], + "table_body": "
MethodDatasetsMPJPEPA-MPJPEPVE
HMR [21]R112.3067.50141.92
HMRR + S95.0157.62116.10
SPIN [25]R96.9059.20119.70
SPINR + S84.1453.67103.79
PARE [24]R81.7949.36105.27
PARER + S78.9848.46103.86
VIBE [23]R94.8857.08108.59
VIBER + S93.0457.00107.23
", + "bbox": [ + 81, + 140, + 467, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Human Neural Rendering", + "text_level": 1, + "bbox": [ + 76, + 304, + 316, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the flexibility of our pipeline, we render a total of 100 multi-view sequences with diverse motions and appearances for benchmarking human NeRFs. All sequences have a length of 300 frames with a resolution of $1024 \\times 1024$ , where the motion sequence is randomly sampled from AMASS [31]. Each sequence contains 8 views whose camera positions have uniformly distributed azimuth angles around the human body. We use RGB renderings, binary foreground masks, camera parameters, and SMPL parameters for the benchmark. Thanks to our layered design, we can offer accurate SMPL parameters for human NeRFs, which acts as an important prior for a majority of methods, while keeping the diversity in clothes and motions.", + "bbox": [ + 75, + 328, + 468, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 539, + 202, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we study the usefulness of SynBody for two popular research directions: human pose and shape estimation and human NeRF.", + "bbox": [ + 75, + 564, + 468, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Human Pose and Shape Estimation", + "text_level": 1, + "bbox": [ + 76, + 619, + 383, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Estimating 3D humans represented by SMPL (body-only) and SMPL-X (body, hands, and face) parameters from monocular 2D input has gained substantial attention. SynBody emerges as a scalable and effective solution to tackle the scarcity of paired data in these fields.", + "bbox": [ + 75, + 643, + 468, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parametric Human Models. SMPL [30] represents high-dimension human mesh $M(\\theta, \\beta) \\in \\mathbb{R}^{6890 \\times 3}$ as low-dimension pose parameters $\\theta \\in \\mathbb{R}^{72}$ and shape parameters $\\beta \\in \\mathbb{R}^{10}$ . In addition to SMPL, SMPL-X [39] further consists of additional left and right hand pose parameters $(\\theta_{lh}, \\theta_{rh} \\in \\mathbb{R}^{15 \\times 3})$ , jaw joint rotation $(\\theta_{lh} \\in \\mathbb{R}^3)$ and face expression parameters $(\\phi_f \\in \\mathbb{R}^{10})$ for an expressive human representation.", + "bbox": [ + 75, + 719, + 468, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Benchmarks.** We conduct experiments on two mainstream benchmarks: 3DPW [48] and AGORA [38]. As a widely used benchmark for in-the-wild evaluation, 3DPW encompasses diverse data collected through a mobile phone cam", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3bb97336d600857919730d2b23a6e28a71b025bdc44652cd6a2b629481ce4872.jpg", + "table_caption": [ + "Table 3: Training popular baseline methods (image-based only as AGORA does not provide videos but static images only) with SynBody on AGORA validation set [38]. R: standard real datasets. S: SynBody." + ], + "table_footnote": [], + "table_body": "
MethodDatasetsMPJPEPA-MPJPEPVE
HMR [21]R226.7187.72248.35
HMRR + S199.5177.97210.37
SPIN [25]R212.9179.76217.88
SPINR + S196.8176.06205.83
PARE [24]R178.1567.13189.73
PARER + S169.9364.37179.81
", + "bbox": [ + 516, + 154, + 877, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "era, paired with SMPL annotations. As for AGORA, it is a recent synthetic dataset that features challenging scenes with person-person occlusion. AGORA provides both SMPL and SMPL-X annotations.", + "bbox": [ + 496, + 305, + 890, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the quality of predicted parametric human models, we employ the standard metrics: MPJPE (Mean Per Joint Position Error), which is calculated as the average $L_{2}$ distance over 3D keypoints regressed from parametric models; PA-MPJPE (Procrustes Aligned Mean Per Joint Position Error), which is MPJPE with applying Procrustes Alignment [11] on the predicted keypoints to match the ground truth before error computation; PVE (Per Vertex Error), which is the average $L_{2}$ distance between predicted and ground truth mesh vertices.", + "bbox": [ + 496, + 366, + 892, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.1 SMPL Estimation", + "text_level": 1, + "bbox": [ + 500, + 541, + 679, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods. To gauge the usefulness of SynBody, we conduct experiments with four milestone works: HMR [21], SPIN [25], and PARE [24] are image-based methods and VIBE [23] is a video-based method. We follow the hyperparameter configurations provided in MMHuman3D [10].", + "bbox": [ + 496, + 566, + 890, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Data. In the following experiments, we refer to the standard baskets of real datasets for SMPL estimation as \"R\". For HMR and SPIN, R consists of H36M [16], MPI-INF-3DHP [32], LSP [17], LSPET [18], MPII [3] and MSCOCO [29], whereas in VIBE, R consists of MPI-INF-3DHP and InstaVariety [22]. \"S\" denotes SynBody data with 2.7M training instances with SMPL annotations.", + "bbox": [ + 496, + 643, + 890, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Main Results. We collate experiment results on 3DPW in Table 2. Each baseline is fine-tuned with a mixture of real datasets (R) and SynBody (S), leading to significant performance gains across different baseline models, even on strong baseline PARE [24] (2.81 mm and 1.41 mm improvements on MPJPE and PVE). Notably, SynBody can be used to train video-based methods such as VIBE as it contains video sequences instead of static images. Table 3 illustrates that training with SynBody exhibits remarkable improvements across methods, when compared with", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "footer", + "text": "20288", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9088f1e0e8c6a5eba3e196a484217c549eb830f322c7c1678c1d24b59ade6080.jpg", + "image_caption": [ + "Figure 6: Impact of the amount of SynBody data. The horizontal axis represents the amount of SynBody data that is varied by multiples of real data. The baseline model is HMR and tested on the 3DPW test set." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 467, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d4bfe06201984c4bff4bf562dac3fd0b4361414141ffcd1ff6cf2ebde64c7706.jpg", + "table_caption": [ + "Table 4: Comparison between SynBody-100K and AGORA on 3DPW test set. “A” means AGORA, and “S100K” means SynBody-100K which is a subset of SynBody datasets with over 100K SMPL annotations. For a fair comparison, the total number of SMPL annotations is close to that of AGORA." + ], + "table_footnote": [], + "table_body": "
MethodDatasetsMPJPEPA-MPJPEPVE
HMRR + A101.6157.85123.82
HMRR + S100K95.2857.68119.18
SPINR + A88.4454.97110.35
SPINR + S100K85.5254.12105.45
PARER + A85.3448.39109.77
PARER + S100K79.4247.80102.45
", + "bbox": [ + 84, + 380, + 460, + 498 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1096d790206f3be848a2034fe808f0ddec82521a305b37864ac84a2af93d7f00.jpg", + "table_caption": [ + "Table 5: Effectiveness of SMPL-XL. R: real datasets (MSCOCO, MPI-INF-3DHP, Human3.6M, LSP, LSPET, MPII). S100K: downsampled SynBody with 100K instances.*: SURREAL-style." + ], + "table_footnote": [], + "table_body": "
MethodDatasetsMPJPEPA-MPJPEPVE
PARER81.849.4105.3
PARER+S100K*82.249.7105.6
PARER+S100K79.447.8102.5
", + "bbox": [ + 93, + 592, + 450, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the baseline models in AGORA validation set.", + "bbox": [ + 75, + 700, + 382, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Data Scale. As our generation system can easily scale up the data size, we train HMR from scratch to study the influence of synthetic data scale. Figure 6 demonstrates that adding more SynBody data generally leads to better performance. These experiments confirm that synthetic data is a valuable complement to real data and serves as a readily scalable training source to supplement the typically limited real data.", + "bbox": [ + 75, + 717, + 467, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with AGORA. To evaluate the quality of SynBody, we randomly sample a subset with 100K training instances, \"SynBody-100K\", which is comparable in size with the popular synthetic dataset AGORA. In Table 4, we", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: SMPL-X estimation with OSX [28] as the baseline on AGORA validation set (AGORA-val) [38] and 3DPW test set [28]. R: real datasets (MSCOCO, MPII, and Human3.6M). S: SynBody. AGORA-val uses PVE (mm) whereas 3DPW uses MPJPE (mm) and PA-MPJPE (mm). Note the original OSX uses the SMPL head for 3DPW, which we modify to the SMPL-X head.", + "bbox": [ + 496, + 88, + 890, + 195 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4eb766633801c4c87a05929194f14b9c0f2212697229e0e9c60b67de5a5fc51b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDatasetsAGORA3DPW (Body)
AllHandsFaceMPJPEPA-MPJPE
OSXR168.670.677.2110.263.5
OSXR+S600K155.863.572.092.259.3
", + "bbox": [ + 503, + 200, + 888, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "show that finetuning with SynBody outperforms that with AGORA across different baseline methods. The results highlight that SynBody demonstrates competitive traits as a training source.", + "bbox": [ + 496, + 305, + 890, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of SMPL-XL. SMPL-XL enables layered human modeling which cannot be achieved with SURREAL/RenderPeople as Figure 5. We investigate the importance to synthesize data with these realistic considerations of SMPL-XL in Table 5. We compare SynBody with SURREAL-style data that renders body and cloth texture on human mesh surfaces, without actual cloth geometry or hair. It is observed that SURREAL-style data leads to significant performance degradation.", + "bbox": [ + 496, + 366, + 890, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1.2 SMPL-X Estimation", + "text_level": 1, + "bbox": [ + 500, + 526, + 694, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Also known as expressive human pose and shape estimation, SMPL-X estimation requires recovery of body, hands, and face parameters.", + "bbox": [ + 496, + 551, + 890, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Method and Training Data. In Table 6, we conduct experiments with the recent SoTA, OSX [28], as the base model. We compare the original baseline with training with a mixture of real datasets (\"R\" denotes COCO [29], MPII [3] and Human3.6M [16] with pseudo ground truth generated by NeuralAnnot [36]) and SynBody (\"S600K\" here denotes a downsampled set of 600K instances as the real datasets are smaller for SMPL-X than that for SMPL estimation). Note that in the OSX paper, the values reported on 3DPW are obtained using a SMPL head. Here we standardize the output using the same SMPL-X head for both AGORA and 3DPW.", + "bbox": [ + 496, + 597, + 890, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Main Results. We observe that adding SynBody in the training of OSX leads to a significant improvement in both overall estimation (more than $10\\mathrm{mm}$ ) and part-level (more than $5\\mathrm{mm}$ for hands and face) estimation on AGORA. Moreover, in a fair comparison with the same model architecture, training with SynBody leads to $18\\mathrm{mm}$ improvements in MPJPE on the 3DPW test set (here we follow the standard protocol to evaluate 14 key joints as 3DPW does not provide SMPL-X annotations). We speculate that it", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20289", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0b3a4abd395918a1cfbf5dfcf5c291e8d18950acd96dc8774206a232f1838344.jpg", + "table_caption": [ + "Table 7: Benchmark of NeRF-based methods for 3D human neural rendering on SynBody." + ], + "table_footnote": [], + "table_body": "
MethodNovel ViewNovel PoseNovel Identity
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [34]19.390.8620.16219.610.8240.201---
NeuralBody [41]28.940.9660.05725.020.9440.080---
HumanNeRF [50]28.320.9630.06621.970.8790.108---
AnimNeRF [40]27.490.9640.05626.210.9500.068---
NHP [26]25.660.9530.07624.180.9450.08022.460.9270.103
", + "bbox": [ + 116, + 108, + 856, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "is difficult to obtain accurate SMPL-X annotations, which is mitigated by SynBody as it provides large-scale high-quality SMPL-X labels paired with images that are rendered with diverse backgrounds and lighting conditions, which are beneficial to training a high-performing expressive human parametric model recovery.", + "bbox": [ + 75, + 250, + 468, + 339 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.2. Human NeRF", + "text_level": 1, + "bbox": [ + 76, + 347, + 220, + 361 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this section, we benchmark popular NeRF-based methods for 3D humans on SynBody, validating the effectiveness and great potential of our dataset in human neural rendering. Our benchmark is built upon three perspectives according to the purpose of synthesis tasks, which can be categorized into novel view, novel pose, and novel identity. Methods for Benchmark. We benchmark five methods in total, including the vanilla NeRF [34], NeuralBody [41] and HumanNeRF [50] for novel view synthesis, AnimNeRF [40] for novel pose synthesis, NHP [26] for generalizable human NeRF (novel identity synthesis). Except for NHP, all methods are trained in a person-specific manner, taking 4 views of the first 250 frames for training and the rest views and frames for evaluations.", + "bbox": [ + 75, + 371, + 468, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Evaluation Protocols. We follow [34, 50] to evaluate all methods using three standard metrics: Peak Signal-to-Noise Ratio (PSNR), Structural SIMilarity index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS [56]). To reach a consensus among all methods, all metrics are computed over the whole image with a black background.", + "bbox": [ + 75, + 582, + 468, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Main Results. Benchmark results are reported in Table 7. We observe that all methods achieve comparable performances as real human data on SynBody. Models (NeuralBody and AnimNeRF) which rely on accurate SMPL estimation and blending weights perform better on novel poses. We attribute it to our layer-wise design which offers ground truth SMPL parameters for human NeRF training. Besides, we present visualization results in Figure 7, observing that the diverse motions and appearances, as well as loose garments in SynBody, pose further challenges in the field of neural rendering of 3D humans.", + "bbox": [ + 75, + 672, + 470, + 838 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 845, + 194, + 859 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present SynBody, a large-scale synthetic dataset that features a substantial number of subjects and high-quality", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/3abc8211dd7460ae755962e39207c11572e83457af81e3b75ed119812d13b301.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 501, + 248, + 622, + 342 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4731a51771d0956d16f12af922897f06670a5aae89df6c7d02f4259edbaef5af.jpg", + "image_caption": [ + "NeRF" + ], + "image_footnote": [], + "bbox": [ + 637, + 250, + 756, + 342 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/626a7baf5cef8df6ed5b464cd3c69d56205fce8bfd44ad0e69773fe1e224c323.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 250, + 890, + 342 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7818d9e5f61954e8601842e765120f1474f2fa929407d320c7dc613ffab1151a.jpg", + "image_caption": [ + "HumanNeRF" + ], + "image_footnote": [], + "bbox": [ + 501, + 354, + 622, + 449 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/60c2fe7cd237e9bed2005b7578ed82fe43c385707e65a540a3deb4cfd748f000.jpg", + "image_caption": [ + "AnimNeRF" + ], + "image_footnote": [], + "bbox": [ + 637, + 354, + 756, + 449 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/295aa1aa17272125d0346c224615a1ce63b47877b7f4fe1b22201fa5e45453fa.jpg", + "image_caption": [ + "NeuralBody", + "NHP", + "Figure 7: Novel view synthesis of different human NeRF methods on SynBody." + ], + "image_footnote": [], + "bbox": [ + 769, + 354, + 890, + 449 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "3D annotations. At the core is a clothed human model with multiple layers of representation. Our experiments demonstrate the effectiveness of SynBody on both human mesh recovery and human NeRF. Future research can leverage SynBody for developing and evaluating methods to predict body and cloth simultaneously. Furthermore, the high controllability of the synthetic dataset offers ample opportunities for further improvements, such as the incorporation of contact labels for human-scene interaction.", + "bbox": [ + 496, + 527, + 890, + 664 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Societal Impacts. Even though SynBody is a synthetic dataset, the assets utilized in its creation might not be well-balanced. Although hairstyles and skin colors are chosen at random, they are designed to avoid racial biases. Yet, other elements, such as body shapes and clothing, might not be as balanced, posing a potential for bias in the resultant human models.", + "bbox": [ + 496, + 669, + 890, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This study is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2021-08-019), the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20290", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marvelous designer, 2023. https://www.marvelousdesigner.com.4", + "[2] Render people, 2023. https://renderpeople.com. 4,5", + "[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 7, 8", + "[4] Rao Anyi, Jiang Xuekun, Guo Yuwei, Xu Linning, Yang Lei, Jin Libiao, Lin Dahua, and Dai Bo. Dynamic storyboard generation in an engine-based virtual environment for video production. arXiv preprint arXiv:2301.12688, 2023. 4", + "[5] Eduard Gabriel Bazavan, Andrei Zanfir, Mihai Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Hspace: Synthetic parametric humans animated in complex environments. arXiv preprint arXiv:2112.12867, 2021. 2, 3, 5", + "[6] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 3", + "[7] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, Fangzhou Hong, Mingyuan Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. HuMMan: Multi-modal 4d human dataset for versatile sensing and modeling. In 17th European Conference on Computer Vision, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 557-577. Springer, 2022. 3", + "[8] Zhongang Cai, Mingyuan Zhang, Jiawei Ren, Chen Wei, Daxuan Ren, Zhengyu Lin, Haiyu Zhao, Lei Yang, and Zwei Liu. Playing for 3d human recovery. arXiv preprint arXiv:2110.07588, 2021. 2, 3, 5", + "[9] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 20-40. Springer, 2020. 3", + "[10] MMHuman3D Contributors. Openmmlab 3d human parametric model toolbox and benchmark. https://github.com/open-mmlab/mmhuman3d, 2021.7", + "[11] J. C. Gower. Generalized procrustes analysis. Psychometrika, 1975. 7", + "[12] Shawn Hargreaves and Mark Harris. Deferred shading. In Game Developers Conference, volume 2, page 31, 2004. 6", + "[13] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022.2", + "[14] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3", + "[16] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014. 3, 7, 8", + "[17] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC, pages 1-11. British Machine Vision Association, 2010. 7", + "[18] Sam Johnson and M. Everingham. Learning effective human pose estimation from inaccurate annotation. CVPR 2011, pages 1465-1472, 2011. 7", + "[19] H. Joo, N. Neverova, and A. Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. ArXiv, abs/2004.03686, 2020. 3", + "[20] H. Joo, Tomas Simon, Xulong Li, H. Liu, L. Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart C. Nabbe, I. Matthews, T. Kanade, S. Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:190-204, 2019. 3", + "[21] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, pages 7122-7131, 2018. 2, 7", + "[22] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In Computer Vision and Pattern Recognition (CVPR), 2019. 7", + "[23] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 3, 7", + "[24] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. arXiv preprint arXiv:2104.08527, 2021. 2, 3, 7", + "[25] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 3, 7", + "[26] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 9", + "[27] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2", + "[28] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. 2023. 8", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20291", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 7", + "[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 2, 6, 7", + "[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 3, 7", + "[33] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In ECCV, 2022. 3", + "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 9", + "[35] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In CVPR, pages 2308-2317, 2022. 3", + "[36] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannot: Neural annotator for 3d human mesh training sets. In Proceedings of the Conference on Computer Vision and Pattern Recognition, pages 2299-2307, 2022. 8", + "[37] Hui En Pang, Zhongang Cai, Lei Yang, Tianwei Zhang, and Ziwei Liu. Benchmarking and analyzing 3d human pose and shape estimation beyond algorithms. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2", + "[38] Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regression analysis. In Proceedings Conf. on Computer Vision and Pattern Recognition (CVPR), June 2021. 2, 3, 5, 7, 8", + "[39] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 2, 4, 5, 7", + "[40] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2, 3, 9", + "[41] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 9", + "[42] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In CVPR, pages 722-731, 2021. 6" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2", + "[44] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the International Conference on Computer Vision, pages 1749-1759, 2021. 3", + "[45] L. Sigal, A. O. Balan, and Michael J. Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision, 87:4-27, 2009. 3", + "[46] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In Advances in Neural Information Processing Systems, 2021. 3", + "[47] Gül Varol, J. Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, I. Laptev, and C. Schmid. Learning from synthetic humans. CVPR, pages 4627-4635, 2017. 2, 3, 5", + "[48] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 3, 7", + "[49] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. arXiv preprint arXiv:2303.13796, 2023. 2", + "[50] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In CVPR, pages 16210–16220, June 2022. 9", + "[51] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3d human shape and articulated pose models. In CVPR, pages 6184-6193, 2020. 2", + "[52] Ailing Zeng, Xuan Ju, Lei Yang, Ruiyuan Gao, Xizhou Zhu, Bo Dai, and Qiang Xu. Deciwatch: A simple baseline for $10 \\times$ efficient 2d and 3d pose estimation. In ECCV, pages 607-624. Springer, 2022. 2", + "[53] Ailing Zeng, Xiao Sun, Lei Yang, Nanxuan Zhao, Minhao Liu, and Qiang Xu. Learning skeletal graph neural networks for hard 3d pose estimation. In ICCV, pages 11436-11445, 2021. 2", + "[54] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: a plug-and-play network for refining human poses in videos. In ECCV, pages 625-642. Springer, 2022. 2", + "[55] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. arXiv preprint arXiv:2207.06400, 2022.3", + "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 9" + ], + "bbox": [ + 501, + 92, + 890, + 896 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20292", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/f251ae55-2ca7-4578-ad92-a51119c290a6_model.json b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/f251ae55-2ca7-4578-ad92-a51119c290a6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e0ee889cf61a4c999d4366109c1f977971585bb5 --- /dev/null +++ b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/f251ae55-2ca7-4578-ad92-a51119c290a6_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.131, + 0.797, + 0.177 + ], + "angle": 0, + "content": "SynBody: Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.203, + 0.826, + 0.328 + ], + "angle": 0, + "content": "Zhitao Yang\\(^{1,*}\\) Zhongang Cai\\(^{1,2,3,*}\\) Haiyi Mei\\(^{1,*}\\) Shuai Liu\\(^{2,*}\\) Zhaoxi Chen\\(^{3,*}\\) Weiye Xiao\\(^{1}\\) Yukun Wei\\(^{1}\\) Zhongfei Qing\\(^{1}\\) Chen Wei\\(^{1}\\) Bo Dai\\(^{2}\\) Wayne Wu\\(^{2}\\) Chen Qian\\(^{1}\\) Dahua Lin\\(^{4}\\) Ziwei Liu\\(^{3,\\dagger}\\) Lei Yang\\(^{1,2,\\dagger}\\) SenseTime Research Shanghai AI Laboratory \n\\(^{3}\\)S-Lab, Nanyang Technological University The Chinese University of Hong Kong *Equal Contribution Corresponding Author https://synbody.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.346, + 0.596, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.562, + 0.333, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.665, + 0.337, + 0.678 + ], + "angle": 0, + "content": "Human Pose and Shape Estimation" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.565, + 0.589, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.664, + 0.536, + 0.678 + ], + "angle": 0, + "content": "Novel View Synthesis" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.346, + 0.865, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.392, + 0.8, + 0.405 + ], + "angle": 0, + "content": "Various Environments" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.412, + 0.821, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.544, + 0.808, + 0.557 + ], + "angle": 0, + "content": "Layered Human Models" + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.562, + 0.842, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.665, + 0.792, + 0.677 + ], + "angle": 0, + "content": "Rich Annotations" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.68, + 0.893, + 0.709 + ], + "angle": 0, + "content": "Figure 1: SynBody is a large-scale synthetic dataset with a massive number of subjects and high-quality annotations. It supports various research topics, including human pose and shape estimation (HPS) and novel view synthesis for human (Human NeRF)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.721, + 0.314, + 0.737 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Synthetic data has emerged as a promising source for 3D human research as it offers low-cost access to large-scale human datasets. To advance the diversity and annotation quality of human models, we introduce a new synthetic dataset, SynBody, with three appealing features: 1) a clothed parametric human model that can generate a diverse range of subjects; 2) the layered human representation that naturally offers high-quality 3D annotations to support multiple tasks; 3) a scalable system for produc" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.895, + 0.875 + ], + "angle": 0, + "content": "ing realistic data to facilitate real-world tasks. The dataset comprises 1.2M images with corresponding accurate 3D annotations, covering 10,000 human body models, 1,187 actions, and various viewpoints. The dataset includes two subsets for human pose and shape estimation as well as human neural rendering. Extensive experiments on SynBody indicate that it substantially enhances both SMPL and SMPL-X estimation. Furthermore, the incorporation of layered annotations offers a valuable training resource for investigating the Human Neural Radiance Fields(NeRF)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20282" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.09, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.115, + 0.468, + 0.22 + ], + "angle": 0, + "content": "The fields of 3D human perception [21, 23-25, 37, 53] and human reconstruction [13, 14, 26, 40, 41] have become increasingly important, but the lack of available data has limited their development. Collecting real human data on a large scale is challenging due to privacy concerns and time constraints. Therefore, exploring the use of synthetic human datasets has become a critical avenue of research." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.222, + 0.468, + 0.461 + ], + "angle": 0, + "content": "Despite the great potential, existing synthetic datasets [5, 8,38,47] suffer from limitations such as the number of available human models and the quality of annotations. The main reason lies in that synthetic human datasets rely on real scans for rendering, which poses three key obstacles. Firstly, it is challenging to expand the types of body shapes, poses, and clothing available in the dataset. Secondly, as the human models are scanned with clothing, the 3D annotations obtained through fitting are prone to errors. Thirdly, it is difficult to obtain annotations of body and clothing separately. To address these issues, we develop a new synthetic dataset termed SynBody. The dataset includes 1.2 million frames with corresponding ground-truth 3D human body annotations. It covers 10,000 human body models, 1,187 motions, and 26,960 video clips with 2.7M SMPL/SMPL-X annotations." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.464, + 0.468, + 0.718 + ], + "angle": 0, + "content": "At the heart of SynBody is the layered parametric human model, which constructs the clothed human model in a bottom-up manner. SMPL-X [39] is a widely used parametric human model, capable of sampling human models with various body shapes. However, it lacks the ability to model clothing, limiting its applicability when synthesizing realistic human models. To overcome this limitation, we introduce SMPL-XL, a parametric human model based on SMPL-X in a layered representation. SMPL-XL enriches the SMPL-X model in three aspects: (1) Hair system: adding hair and beards to the FLAME [27] model, with 32 types of hair and 13 types of beards; (2) Garment and accessories: adding procedural clothes to the SMPL-X body, including coats, shirts, pants, skirts, shoes, and glasses; (3) Texture: in addition to adding rich geometry, SMPL-XL also adds rich textures for sampling various skin colors and clothing textures." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.72, + 0.468, + 0.884 + ], + "angle": 0, + "content": "The designed SMPL-XL is capable of automatically generating a large number of human models with high-quality annotations. We therefore generate 10,000 clothed human models by sampling various body shapes, clothing styles, hairstyles, accessories, and textures. Notably, the use of the SMPL-X model as the base body model guarantees that the parametric human annotations are always accurate, obviating the need for the necessity for annotations through fitting. Furthermore, as the clothes are explicitly attached to the surface of the human body, layered annotations for body and clothes are available." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To generate a large-scale dataset with high diversity and" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.227 + ], + "angle": 0, + "content": "high-quality annotations, we design a scalable and automatic system to render images and annotations. We first animate the 10,000 dressed human models by retargeting motions from a large motion library [31]. Subsequently, we design an algorithm to place human models in the scene without piercing. Multiple cameras are then placed by evaluating self-occlusion, inter-occlusion, and view diversity, and the rendering module renders the assets into images with corresponding annotations." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.228, + 0.89, + 0.437 + ], + "angle": 0, + "content": "With SynBody, we launch two tracks that support human pose and shape estimation and human neural rendering, respectively. Experiments show that SynBody is more effective than AGORA under the same amount of training data for human pose and shape estimation. With diverse and large-scale training data, SynBody achieves significant performance gains on both SMPL and SMPL-X estimation. In terms of human neural rendering using neural radiance fields (i.e. NeRF [34]), benchmarking existing approaches on SynBody shows that it has comparable performance as real human data. Furthermore, with the layered annotations which offer accurate SMPL parameters, we observe that current human NeRF approaches are sensitive to the accuracy of estimated SMPL." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.44, + 0.89, + 0.604 + ], + "angle": 0, + "content": "In summary, SynBody is a large-scale synthetic dataset for human perception and modeling, with three main contributions: (1) It constructs clothed subjects and samples 10,000 animatable subjects, which is an order of magnitude higher than existing datasets. (2) The clothed subjects are constructed with an explicit cloth model, thus it provides layered 3D annotations of the human body and clothing, which are not available in previous datasets. (3) Experiments on SynBody achieve promising results on both human perception and modeling, emphasizing the importance of diversity and annotation quality for downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.619, + 0.647, + 0.634 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.645, + 0.89, + 0.84 + ], + "angle": 0, + "content": "Human Parametric Models. Several 3D human parametric models, such as SMPL [30], SMPL-X [39], and GHUM [51], have been developed to generate 3D human meshes from parameters that represent the human pose and shape using linear blend skinning. SMPL-X [39] extends SMPL [30] by combining FLAME [27] and MANO [43] for the head and hands, respectively, and is trained on a large number of real scans to provide a strong basis for shape variations. However, SMPL-X only produces naked body meshes, and we aim to enhance its realism by building a layered parametric model that includes hair, clothes, and accessories. The proposed model leverages the shape basis of SMPL-X while providing realistic dressed human meshes." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.841, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Human Pose and Shape Estimation. Several methods [21, 23-25, 37, 49, 52-54] have been proposed to estimate 3D human pose and shape parameters. HMR [21] directly regresses these parameters in an end-to-end manner, while" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20283" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.138 + ], + "angle": 0, + "content": "Table 1: Comparisons of 3d human dataset. We compare SynBody with existing datasets. We divide datasets into three types: real (R), synthetic (S), and mixed (M). SynBody constructs 10,000 animatable subjects, which is an order of magnitude higher than any existing datasets and brings competitive scale and diversity. \"ITW\" stands for \"In-the-Wild\" in the table." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.146, + 0.877, + 0.361 + ], + "angle": 0, + "content": "
DatasetTypeITWVideo#Views#SMPL#Seq#Subj.#MotionsGT format
HumanEva [45]R-✓4/7NA7463DJ
Human3.6M [15]R-✓4312K83911153DJ, SMPL
MPI-INF-3DHP [32]M✓✓1496K16883DJ
3DPW [48]R✓✓132K6018*SMPL
Panoptic Studio [20]R-✓480736K480~100*3DJ
EFT [19]R✓-1129KNAManyNASMPL
ZJU-MoCap [41]R-✓21180K999SMPL,mask
SURREAL [47]S✓✓16.5MNA1452KSMPL
AGORA [38]S✓-1173KNA>350NASMPL, SMPL-X, mask
HSPACE [5]S✓✓5-NA100×16100GHUM/L, mask
GTA-Human [8]S✓✓11.4M20K>60020KSMPL
BEDLAM [6]S✓✓-1M10K2712,311SMPL-X, mask
SynBodyS✓✓42.7M27K10,0001,187SMPL, SMPL-X, mask
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.473, + 0.651 + ], + "angle": 0, + "content": "SPIN uses an optimization step [25] to guide the learning process towards pseudo-3D labels. PARE [24] employs part attention to tackle occlusion, and VIBE [23] leverages temporal information in videos for SMPL estimation. Apart from building low-cost data collection solutions [7], synthetic datasets have become a promising alternative to efficiently scale up data. SURREAL [47] renders textured SMPL body models in real-image backgrounds but does not account for cloth geometry, resulting in unrealistic subjects. AGORA [38] renders real human scans in a virtual world and provides high-quality synthetic data for image-based approaches. HSPACE [5] places animated human models in various scenes to provide training data for video-based methods, and increases the variation of human shape via refitting. GTA-Human [8] captures videos and optimizes corresponding SMPL annotations from video games. However, the diversity of subjects in current datasets is limited by either body shapes or cloth types." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Expressive Human Pose and Shape Estimation. As face and hand are also crucial for human perception, some efforts [35,44,55] have been made to whole-body human pose and shape estimation. ExPose [9] introduces three experts to predict parameters for body, hand, and face, and merge them in a copy-paste strategy. Hand4Whole [35] further improves the prediction of wrist and finger poses by leveraging selected hand joint features. Predicting both hands and faces makes the dataset much more difficult to obtain than just predicting the body. In order to increase the diversity of real data, AGORA [38] provides image-based SMPL-X annotations by fitting SMPL-X on the scanned human models. While a very recent work, BEDLAM [6], also enhances the SMPL-X model with clothing, hair, and changes in skin tones. However, it employs a manual approach for creating garments and utilizes pre-made skin textures, making" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.895, + 0.454 + ], + "angle": 0, + "content": "it challenging to generate a large number of human body models. In contrast, our body model construction process is fully automated, encompassing clothing, hair, textures, and more. This enables easier scalability to generate human body models on a large scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.896, + 0.717 + ], + "angle": 0, + "content": "Human NeRF. NeRF [34] has demonstrated impressive photo-realistic view synthesis by learning implicit fields of density and color. Yet, human motions are more challenging to learn due to dynamic deformation fields. NeuralBody [41] incorporates prior from a statistical body template to learn dynamic sequence, while Animatable NeRF [40] proposes to reconstruct an animatable human model that generalizes to new poses. Furthermore, NHP [26] and KeypointNeRF [33] achieve generalizability to unseen identities and poses. Several datasets have been adapted to study human NeRF. ZJU-MoCap [41] captures 9 human subjects with 21 synchronized cameras, providing fitted human body model parameters as well as the foreground mask. Human3.6M [16] collects 11 human subjects with 4 cameras, using a marker-based motion capture system. A-NeRF [46] generates a synthetic dataset using SURREAL [47] to study factors that affect visual quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.903 + ], + "angle": 0, + "content": "While the aforementioned tasks are interrelated and often draw upon similar datasets, as shown in Table 1, datasets are limited in two key aspects. Firstly, obtaining real human models is challenging, which restricts the scale of these datasets. Secondly, 3D annotations are typically acquired through optimization, which introduces errors and cannot provide layered annotations. In contrast, with the designed SMPL-XL model, SynBody provides 10,000 subjects with diverse body shapes and clothing, along with layered annotations that include accurate SMPL and SMPL-X. Built on top of this layered human model, SynBody comprises two subsets for human mesh recovery and human NeRF, respec" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20284" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.427, + 0.108 + ], + "angle": 0, + "content": "tively, and features the same 10,000 diverse subjects." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.121, + 0.391, + 0.138 + ], + "angle": 0, + "content": "3. Synthetic Data Generation System" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.159, + 0.468, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.471, + 0.345 + ], + "angle": 0, + "content": "Figure 2: Synthetic data generation system. It consists of 4 components: 1) human model creation to generate layered human models, 2) motion retargeting to drive human models, 3) scene composition to place actors and cameras, and 4) rendering and outputs to generate multi-modal dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.506 + ], + "angle": 0, + "content": "Similar to movie production [4], our system comprises 4 components (depicted in Figure 2): (1) a layered parametric human model creation service, a scalable process to generate layered human models, (2) a motion retargeting module to apply motions from various sources to layered human models, (3) a scene composition module to place 3D actors and objects into a 3D scene, setting up cameras, (4) a 3D rendering engine and a multi-modal data annotation generator. Our infrastructure enables the generation of high-quality synthetic data for various computer vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.384, + 0.53 + ], + "angle": 0, + "content": "3.1. Layered Parametric Human Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.469, + 0.644 + ], + "angle": 0, + "content": "Parametric human models like SMPL-X [39] provide the ability of create rigged body models with various body shapes. However, the lack of available textures limits its application in data generation. We designed a module to perform an automatic process that combines SMPL-X with procedural garments and accessories, hair system and textures, producing realistic and diverse body models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.734 + ], + "angle": 0, + "content": "Body shape. SMPL-X [39] body model has a 3D mesh whose vertex locations are controlled by parameters for pose \\(\\theta\\), shape \\(\\beta\\), and facial expression \\(\\psi\\). By modifying shape \\(\\beta\\), we obtain 3D meshes of various human heights and weights. And by alternating pose \\(\\theta\\), meshes can be driven to perform various poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Garment Model. Our garment is generated as a separate layer on top of the body. Following the industrial garment-making workflow, we designed garment patterns of various styles. Notice that different parts of garment pieces are connected with sewing lines, e.g., the red line in Figure 4 (a). Then, we stitch patterns onto the body at T-Pose utilizing a physical simulator [1]. Specifically, we first manually move the garment pieces to roughly align them with the body. During simulation, vertices between two ends of the seam gradually shrink until they completely pooled together. Figure 4 (b) demonstrates the final draped gar" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.181 + ], + "angle": 0, + "content": "ment. For garment animation, we bind every vertex in the garment to the closest point in the body. Then, the skinnig weight and blend shape of the body mesh are assigned to garment vertices. This makes our garment model easily integrated with existing skeletal pipelines with little computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.893, + 0.38 + ], + "angle": 0, + "content": "Particle hair system. Our method uses a prefabricated particle system to generate realistic hair on a head-shaped mesh. To achieve this, a template of hairstyle or facial hair \\( T_{hair} \\) would attach to the vertices on the mesh which are marked with different areas \\( V_{hair} \\) (including fringe, top, temporal bone, occipital bone, and bottom area). Designers draw multiple sets of guidelines \\( L_{guide} \\) with varying shapes on different areas. Each set of guidelines comprises a collection of Bezier curves that are utilized to accurately constrain the flow of hair strands. Furthermore, the shape of the hair strands can be adjusted by length \\( P_{length} \\) and curliness \\( P_{curliness} \\). The entire hair system is composed of multiple sub-particle systems \\( \\{V_{hair}, L_{guide}, P_{length}, P_{curliness}\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Accessories. We also add template accessories \\(T_{\\text{accessories}}\\) to our model, such as glasses, shoes, hats, and headphones. All accessories are pre-assembled on an SMPL-X template with a uniform body shape. These accessories can be transferred to models with the same topology, ensuring consistent deformation across all models in accordance with changes in the shape \\(\\beta\\) of SMPL-X. The corresponding bone weights of the human body model's vertices are transferred to the nearest accessories' vertices, ensuring that these accessories are correctly driven by the armature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.893, + 0.743 + ], + "angle": 0, + "content": "Procedural texture. The procedural garment textures \\( T_{\\text{procedural}} = \\{T_{\\text{pattern}}, T_{\\text{decals}}, T_{\\text{bump}}, P_{\\text{mapping}}\\} \\) used in clothing are composed of multiple pre-set textures by alpha blending as demonstrated in Figure 4 (c). The pattern texture \\( T_{\\text{pattern}} \\) and decals texture \\( T_{\\text{decals}} \\) serve as layered masks to elevate the style and color, while the bump texture \\( T_{\\text{bump}} \\) functions as a height map, adding detailed normal information to the clothing's surface. The mapping parameters \\( P_{\\text{mapping}} \\) control the coordination of all the textures in the UV space. Besides, to procedurally create body mesh textures, we build a similar texture template SMPL-X, in which we pre-draw layered masks to separate different body features, such as skin, lips, and freckles, allowing for color adjustment and blending in specific regions as in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Combining all the elements above, as shown in Figure 5, we obtain human body models with the same high quality as RenderPeople [2]. Besides, it is challenging to expand the types of body shape and clothing available for real scans, but our model can be easily scaled by randomly sampling each component." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.688, + 0.863 + ], + "angle": 0, + "content": "3.2. Motion Retargeting" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Retargeting of skeletal animations. Our motion retargeting module allows for the transfer of motion data from vari" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20285" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.441, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.217, + 0.383, + 0.232 + ], + "angle": 0, + "content": "(a) Naked human model generation" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.088, + 0.882, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.216, + 0.838, + 0.232 + ], + "angle": 0, + "content": "(b) Layered human model generation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.235, + 0.894, + 0.281 + ], + "angle": 0, + "content": "Figure 3: Layered parametric human model creation: (a) Generation of a naked human model using parametric body model [39], procedural body texture colors, and particle hair System; (b) Integration of garments and accessories onto the naked body model." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.304, + 0.462, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.498, + 0.471, + 0.575 + ], + "angle": 0, + "content": "Figure 4: Demonstrations of Garment and Procedural Texture. (a) garment pattern with sewing lines, (b) simulated garment under the body in canonical space, (c) procedural textured garment whose style and color are elevated by layered masks provided by patterns and decals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.667 + ], + "angle": 0, + "content": "ous sources, such as academic motion datasets, motion captures, and artist-crafted sources, to SMPL-XL model skeletons. Despite variations in bone names, bone lengths, and rest pose bone rotations, the source skeletons are typically structurally similar to the target skeletons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.471, + 0.743 + ], + "angle": 0, + "content": "Pose frames in motion clips contain root translation \\( T(t) \\in \\mathbb{R}^3 \\) and rotations of each bone in the corresponding parent bone's space \\( \\{R_i(t) \\in \\mathbb{R}^3 | i = 0,1,\\dots,n\\} \\). Following the forward kinematics (FK) manner, each bone's rotation relative to the model space can be calculated:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.748, + 0.47, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\widehat {R} _ {i} (t) = \\widehat {R} _ {p (i)} (t) \\cdot R _ {i} (t) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.47, + 0.826 + ], + "angle": 0, + "content": "where \\(\\widehat{R}_i(t)\\) is the rotation of \\(R_{i}(t)\\) in model space at \\(t\\) frame, and \\(p(i)\\) indicates the parent of \\(i\\). Specially, root bones have no parents so \\(\\widehat{R}_0(t) = R_0(t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To retarget motion from one skeleton to another, we assume the source and the target motion can drive the corresponding bones of both skeletons to the same rotation in model space. We have T-posed frames of skeletons by manually posing them as T-posing, and treat T-posed as all mo" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.304, + 0.885, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.485, + 0.893, + 0.577 + ], + "angle": 0, + "content": "Figure 5: A Demonstration of comparison between commonly used human models in existing synthetic datasets [5, 8,38,47] and our SMPL-XL. We obtain high-quality models equal to RenderPeople [2], both are much better than SUR-REAL [47], and our model has the capability of scaling up easily by sampling various assets and body shapes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.623 + ], + "angle": 0, + "content": "tions' first frame pose. So motions relative to their T-posedr frame can be easily obtained:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.629, + 0.892, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\widehat {R} _ {i} (t) = \\widehat {R} _ {i} (0) \\cdot \\widehat {R} _ {i _ {- s r c}} (t) \\cdot \\widehat {R} _ {i _ {- s r c}} ^ {- 1} (0). \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.721 + ], + "angle": 0, + "content": "Considering skeletons have different bone lengths, which could result in \"sliding feet\" artifacts on target skeletons. We simply scale \\( T \\) according to the ratio of pelvis bones' height to mitigate them: \\( T(t) = \\frac{H_{pelvis}}{H_{pelvis,src}} \\cdot T_{src}(t) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.893, + 0.856 + ], + "angle": 0, + "content": "SMPL-X Annotations. Considering SMPL-XL body shapes are sampled from SMPL-X, the shape \\(\\beta\\) can be derived directly from the corresponding SMPL-X. Pose \\(\\theta\\) in the model space is calculated in the motion retargeting module. And in the scene composition module, models are placed in a 3D scene with camera space locations \\(T_{c}\\) and rotations \\(R_{c}\\). Pose \\(\\theta_{c}\\) in camera space is calculated by applying those world transformations to \\(\\theta\\). So SMPL-X annotations are constructed with \\(\\{\\beta, \\theta_{c}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SMPL Annotations. Although SMPL-XL naturally provides accurate SMPL-X annotations, SMPL cannot be derived directly. Thus, we need to refit the SMPL parameters." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20286" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "The optimization process consists of two steps. First, we fit the shape \\(\\beta\\) of all human models under the T-pose to its corresponding SMPL-X. Secondly, for each sequence, it is initialized with fitted \\(\\beta\\) and its original pose. we fix \\(\\beta\\) while fitting the body pose \\(\\theta\\) for SMPL." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.176, + 0.245, + 0.192 + ], + "angle": 0, + "content": "3.3. Scene Placement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.47, + 0.426 + ], + "angle": 0, + "content": "To place \\(N_{p}\\) subjects in a large scene with \\(N_{o}\\) objects, we primarily follow three principles: standing on the ground, avoiding human-object and human-human penetration. To prevent subjects from floating in the air, the root position of a subject should align with the ground height. A sequential decision-making approach is used to find a suitable position for each subject, i.e., placing one subject at a time. An object is represented by \\(o_{i} = \\{x_{i},y_{i},l_{i},w_{i}\\} \\in \\mathbb{R}^{4}\\), where \\(\\{x_i,y_i\\}\\) is the center of the object's axis-aligned bounding box projected onto the ground with length \\(l_{i}\\) and width \\(w_{i}\\). Different from static objects, to avoid collisions between moving subjects at any one time, a subject with specified body shape and motion is simplified to \\(q_{i} = \\{x_{i},y_{i},l_{i},w_{i}\\} \\in \\mathbb{R}^{4}\\), which is the smallest axis-aligned box that envelops all bounding boxes across frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.472 + ], + "angle": 0, + "content": "To avoid human-object penetration and potential human-human collision, the solution \\( p_i^* = \\{x_i^*, y_i^*\\} \\) of a character with the shape of \\( \\{l_i, w_i\\} \\) should satisfy:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.482, + 0.47, + 0.5 + ], + "angle": 0, + "content": "\\[\nI \\left(\\left\\{x _ {i} ^ {*}, y _ {i} ^ {*}, l _ {i}, w _ {i} \\right\\}, o _ {j}\\right) = 0, j = 1, \\dots , N _ {o} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.525, + 0.47, + 0.543 + ], + "angle": 0, + "content": "\\[\nI \\left(\\left\\{x _ {i} ^ {*}, y _ {i} ^ {*}, l _ {i}, w _ {i} \\right\\}, q _ {k}\\right) = 0, k = 1, \\dots , N _ {p}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.639 + ], + "angle": 0, + "content": "where \\( I(\\text{box}_1, \\text{box}_2) \\) denotes the overlapping area between two boxes and \\( N_p \\) is the number of subjects already placed in the scene. Besides, distance constraint is used to prevent subjects from excessive dispersal. The problem is solved by grid search. Typically multiple solutions are available and we randomly sample one each time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.647, + 0.262, + 0.663 + ], + "angle": 0, + "content": "3.4. Camera Placement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.791 + ], + "angle": 0, + "content": "Once the positions of all subjects have been organized, the next task involves placing \\( N_{c} \\) cameras in suitable locations. To ensure that cameras would not be placed inside any objects or subjects, the 3D version of Eq. (3) and Eq. (4) are applied. In addition, we evaluate the suitability of a candidate camera by the following metrics: the distance from the camera to the subjects, the camera's pitch angle, and the degree of occlusion of each subject in the camera's view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.791, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The distance denoted as \\( L \\) from the mean position of all subjects \\( \\bar{p} \\) to the camera is restricted. \\( L_{max} \\) is set to prevent an unreasonably small proportion of subjects in the image. To control the visibility of subjects, \\( L_{min} \\) is defined as \\( \\frac{\\lambda}{\\sin(\\alpha / 2)}\\max_i||pv_i - \\bar{p} ||_2 \\), where \\( i = 1,\\dots ,N_c \\), \\( \\alpha \\) is the field of view of the camera, and \\( \\lambda \\) is a hyperparameter that determines the probability of all subjects being within" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.168 + ], + "angle": 0, + "content": "the camera's view. To estimate the degree of occlusion for each subject, rays are randomly and uniformly cast from the camera to each subject's body, and the percentage of blocked rays by other objects is determined. More details can be found in the Sup. Mat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.179, + 0.75, + 0.195 + ], + "angle": 0, + "content": "3.5. Rendering and Annotations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.202, + 0.894, + 0.324 + ], + "angle": 0, + "content": "Using a high-quality rendering pipeline in Unreal Engine 5, SynBody is rendered in multi-view with large 3D environments from the Unreal Marketplace for rich background information and dynamic lighting. Leveraging the G-buffer [12], our system simultaneously generates photorealistic RGB images and annotations, along with accurate ground truth for segmentation masks, optical flow, depth maps, normal maps, and other ground-truth labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.333, + 0.669, + 0.351 + ], + "angle": 0, + "content": "4. SynBody Dataset" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.359, + 0.669, + 0.373 + ], + "angle": 0, + "content": "4.1. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.895, + 0.791 + ], + "angle": 0, + "content": "SynBody comprises 10,000 unique SMPL-XL models randomly created with different body shapes and genders. Each model is then combined with the following assets: (1) hairstyles \\( T_{hair} \\) sampled from 45 particle hairs; (2) garments \\( G_{tmp} \\) sampled from 68 clothing models containing multiple outfits; (3) procedural texture \\( T_{procedural} \\) generated by sampling \\( T_{pattern} \\), \\( T_{decals} \\), and \\( T_{bump} \\) from 1,038 template textures, and color values are randomly sampled; (4) accessories \\( T_{accessories} \\) sampled from 46 template assets. To ensure the validity of the motions, we select a subset from AMASS [31]. Based on the BABEL annotations [42], we excluded interactive motions and non-ground motions(e.g., swim), as well as filtered out motions with a duration shorter than 2 seconds, leading to a subset with 1,187 motions. For each sequence, we randomly selected 4 human models, and each model is assigned a randomly chosen motion lasting between 2 and 10 seconds. For motions exceeding 2 seconds, we randomly extract a 2-second segment to ensure that each video clip has a length of 60 frames. To enhance the diversity of viewpoints, 4 view positions are generated for each sequence, where each one is randomly sampled from the surface of semi-spheres with varying radii. The system generates 26,960 sequences and 1.2M images with annotations, utilizing 6 vast and realistic scenes created by professional artists. 2.7M SMPL/SMPL-X annotations are provided, excluding highly occluded subjects. More details can be found in the Sup. Mat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.808, + 0.818 + ], + "angle": 0, + "content": "4.2. Human Pose and Shape Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.902 + ], + "angle": 0, + "content": "With 2.7 million SMPLs, we leverage a pretrained regressor to extract 3D keypoints and then project them onto 2D space. Following standard practice in top-down human mesh recovery methods, we generate bounding boxes from the resulting 2D keypoints." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20287" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.136 + ], + "angle": 0, + "content": "Table 2: Training popular baseline methods (image-based and video-based) with SynBody on 3DPW test set [48]. R: standard real datasets. S: SynBody." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.141, + 0.468, + 0.29 + ], + "angle": 0, + "content": "
MethodDatasetsMPJPEPA-MPJPEPVE
HMR [21]R112.3067.50141.92
HMRR + S95.0157.62116.10
SPIN [25]R96.9059.20119.70
SPINR + S84.1453.67103.79
PARE [24]R81.7949.36105.27
PARER + S78.9848.46103.86
VIBE [23]R94.8857.08108.59
VIBER + S93.0457.00107.23
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.305, + 0.317, + 0.323 + ], + "angle": 0, + "content": "4.3. Human Neural Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.47, + 0.526 + ], + "angle": 0, + "content": "Given the flexibility of our pipeline, we render a total of 100 multi-view sequences with diverse motions and appearances for benchmarking human NeRFs. All sequences have a length of 300 frames with a resolution of \\(1024 \\times 1024\\), where the motion sequence is randomly sampled from AMASS [31]. Each sequence contains 8 views whose camera positions have uniformly distributed azimuth angles around the human body. We use RGB renderings, binary foreground masks, camera parameters, and SMPL parameters for the benchmark. Thanks to our layered design, we can offer accurate SMPL parameters for human NeRFs, which acts as an important prior for a majority of methods, while keeping the diversity in clothes and motions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.54, + 0.203, + 0.557 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.61 + ], + "angle": 0, + "content": "In this section, we study the usefulness of SynBody for two popular research directions: human pose and shape estimation and human NeRF." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.384, + 0.637 + ], + "angle": 0, + "content": "5.1. Human Pose and Shape Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.719 + ], + "angle": 0, + "content": "Estimating 3D humans represented by SMPL (body-only) and SMPL-X (body, hands, and face) parameters from monocular 2D input has gained substantial attention. SynBody emerges as a scalable and effective solution to tackle the scarcity of paired data in these fields." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.84 + ], + "angle": 0, + "content": "Parametric Human Models. SMPL [30] represents high-dimension human mesh \\( M(\\theta, \\beta) \\in \\mathbb{R}^{6890 \\times 3} \\) as low-dimension pose parameters \\( \\theta \\in \\mathbb{R}^{72} \\) and shape parameters \\( \\beta \\in \\mathbb{R}^{10} \\). In addition to SMPL, SMPL-X [39] further consists of additional left and right hand pose parameters \\( (\\theta_{lh}, \\theta_{rh} \\in \\mathbb{R}^{15 \\times 3}) \\), jaw joint rotation \\( (\\theta_{lh} \\in \\mathbb{R}^3) \\) and face expression parameters \\( (\\phi_f \\in \\mathbb{R}^{10}) \\) for an expressive human representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "**Benchmarks.** We conduct experiments on two mainstream benchmarks: 3DPW [48] and AGORA [38]. As a widely used benchmark for in-the-wild evaluation, 3DPW encompasses diverse data collected through a mobile phone cam" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.151 + ], + "angle": 0, + "content": "Table 3: Training popular baseline methods (image-based only as AGORA does not provide videos but static images only) with SynBody on AGORA validation set [38]. R: standard real datasets. S: SynBody." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.155, + 0.878, + 0.268 + ], + "angle": 0, + "content": "
MethodDatasetsMPJPEPA-MPJPEPVE
HMR [21]R226.7187.72248.35
HMRR + S199.5177.97210.37
SPIN [25]R212.9179.76217.88
SPINR + S196.8176.06205.83
PARE [24]R178.1567.13189.73
PARER + S169.9364.37179.81
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.892, + 0.366 + ], + "angle": 0, + "content": "era, paired with SMPL annotations. As for AGORA, it is a recent synthetic dataset that features challenging scenes with person-person occlusion. AGORA provides both SMPL and SMPL-X annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.893, + 0.519 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the quality of predicted parametric human models, we employ the standard metrics: MPJPE (Mean Per Joint Position Error), which is calculated as the average \\( L_{2} \\) distance over 3D keypoints regressed from parametric models; PA-MPJPE (Procrustes Aligned Mean Per Joint Position Error), which is MPJPE with applying Procrustes Alignment [11] on the predicted keypoints to match the ground truth before error computation; PVE (Per Vertex Error), which is the average \\( L_{2} \\) distance between predicted and ground truth mesh vertices." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.542, + 0.681, + 0.557 + ], + "angle": 0, + "content": "5.1.1 SMPL Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Methods. To gauge the usefulness of SynBody, we conduct experiments with four milestone works: HMR [21], SPIN [25], and PARE [24] are image-based methods and VIBE [23] is a video-based method. We follow the hyperparameter configurations provided in MMHuman3D [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Training Data. In the following experiments, we refer to the standard baskets of real datasets for SMPL estimation as \"R\". For HMR and SPIN, R consists of H36M [16], MPI-INF-3DHP [32], LSP [17], LSPET [18], MPII [3] and MSCOCO [29], whereas in VIBE, R consists of MPI-INF-3DHP and InstaVariety [22]. \"S\" denotes SynBody data with 2.7M training instances with SMPL annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Main Results. We collate experiment results on 3DPW in Table 2. Each baseline is fine-tuned with a mixture of real datasets (R) and SynBody (S), leading to significant performance gains across different baseline models, even on strong baseline PARE [24] (2.81 mm and 1.41 mm improvements on MPJPE and PVE). Notably, SynBody can be used to train video-based methods such as VIBE as it contains video sequences instead of static images. Table 3 illustrates that training with SynBody exhibits remarkable improvements across methods, when compared with" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20288" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.468, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.222, + 0.47, + 0.281 + ], + "angle": 0, + "content": "Figure 6: Impact of the amount of SynBody data. The horizontal axis represents the amount of SynBody data that is varied by multiples of real data. The baseline model is HMR and tested on the 3DPW test set." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.301, + 0.468, + 0.376 + ], + "angle": 0, + "content": "Table 4: Comparison between SynBody-100K and AGORA on 3DPW test set. “A” means AGORA, and “S100K” means SynBody-100K which is a subset of SynBody datasets with over 100K SMPL annotations. For a fair comparison, the total number of SMPL annotations is close to that of AGORA." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.381, + 0.462, + 0.499 + ], + "angle": 0, + "content": "
MethodDatasetsMPJPEPA-MPJPEPVE
HMRR + A101.6157.85123.82
HMRR + S100K95.2857.68119.18
SPINR + A88.4454.97110.35
SPINR + S100K85.5254.12105.45
PARER + A85.3448.39109.77
PARER + S100K79.4247.80102.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.527, + 0.468, + 0.587 + ], + "angle": 0, + "content": "Table 5: Effectiveness of SMPL-XL. R: real datasets (MSCOCO, MPI-INF-3DHP, Human3.6M, LSP, LSPET, MPII). S100K: downsampled SynBody with 100K instances.*: SURREAL-style." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.593, + 0.451, + 0.665 + ], + "angle": 0, + "content": "
MethodDatasetsMPJPEPA-MPJPEPVE
PARER81.849.4105.3
PARER+S100K*82.249.7105.6
PARER+S100K79.447.8102.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.383, + 0.716 + ], + "angle": 0, + "content": "the baseline models in AGORA validation set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.468, + 0.838 + ], + "angle": 0, + "content": "Impact of Data Scale. As our generation system can easily scale up the data size, we train HMR from scratch to study the influence of synthetic data scale. Figure 6 demonstrates that adding more SynBody data generally leads to better performance. These experiments confirm that synthetic data is a valuable complement to real data and serves as a readily scalable training source to supplement the typically limited real data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Comparison with AGORA. To evaluate the quality of SynBody, we randomly sample a subset with 100K training instances, \"SynBody-100K\", which is comparable in size with the popular synthetic dataset AGORA. In Table 4, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.196 + ], + "angle": 0, + "content": "Table 6: SMPL-X estimation with OSX [28] as the baseline on AGORA validation set (AGORA-val) [38] and 3DPW test set [28]. R: real datasets (MSCOCO, MPII, and Human3.6M). S: SynBody. AGORA-val uses PVE (mm) whereas 3DPW uses MPJPE (mm) and PA-MPJPE (mm). Note the original OSX uses the SMPL head for 3DPW, which we modify to the SMPL-X head." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.201, + 0.89, + 0.268 + ], + "angle": 0, + "content": "
MethodDatasetsAGORA3DPW (Body)
AllHandsFaceMPJPEPA-MPJPE
OSXR168.670.677.2110.263.5
OSXR+S600K155.863.572.092.259.3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.892, + 0.366 + ], + "angle": 0, + "content": "show that finetuning with SynBody outperforms that with AGORA across different baseline methods. The results highlight that SynBody demonstrates competitive traits as a training source." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.503 + ], + "angle": 0, + "content": "Effectiveness of SMPL-XL. SMPL-XL enables layered human modeling which cannot be achieved with SURREAL/RenderPeople as Figure 5. We investigate the importance to synthesize data with these realistic considerations of SMPL-XL in Table 5. We compare SynBody with SURREAL-style data that renders body and cloth texture on human mesh surfaces, without actual cloth geometry or hair. It is observed that SURREAL-style data leads to significant performance degradation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.527, + 0.696, + 0.541 + ], + "angle": 0, + "content": "5.1.2 SMPL-X Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.891, + 0.597 + ], + "angle": 0, + "content": "Also known as expressive human pose and shape estimation, SMPL-X estimation requires recovery of body, hands, and face parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Method and Training Data. In Table 6, we conduct experiments with the recent SoTA, OSX [28], as the base model. We compare the original baseline with training with a mixture of real datasets (\"R\" denotes COCO [29], MPII [3] and Human3.6M [16] with pseudo ground truth generated by NeuralAnnot [36]) and SynBody (\"S600K\" here denotes a downsampled set of 600K instances as the real datasets are smaller for SMPL-X than that for SMPL estimation). Note that in the OSX paper, the values reported on 3DPW are obtained using a SMPL head. Here we standardize the output using the same SMPL-X head for both AGORA and 3DPW." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Main Results. We observe that adding SynBody in the training of OSX leads to a significant improvement in both overall estimation (more than \\(10\\mathrm{mm}\\)) and part-level (more than \\(5\\mathrm{mm}\\) for hands and face) estimation on AGORA. Moreover, in a fair comparison with the same model architecture, training with SynBody leads to \\(18\\mathrm{mm}\\) improvements in MPJPE on the 3DPW test set (here we follow the standard protocol to evaluate 14 key joints as 3DPW does not provide SMPL-X annotations). We speculate that it" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20289" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.186, + 0.09, + 0.782, + 0.105 + ], + "angle": 0, + "content": "Table 7: Benchmark of NeRF-based methods for 3D human neural rendering on SynBody." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.109, + 0.857, + 0.226 + ], + "angle": 0, + "content": "
MethodNovel ViewNovel PoseNovel Identity
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [34]19.390.8620.16219.610.8240.201---
NeuralBody [41]28.940.9660.05725.020.9440.080---
HumanNeRF [50]28.320.9630.06621.970.8790.108---
AnimNeRF [40]27.490.9640.05626.210.9500.068---
NHP [26]25.660.9530.07624.180.9450.08022.460.9270.103
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.34 + ], + "angle": 0, + "content": "is difficult to obtain accurate SMPL-X annotations, which is mitigated by SynBody as it provides large-scale high-quality SMPL-X labels paired with images that are rendered with diverse backgrounds and lighting conditions, which are beneficial to training a high-performing expressive human parametric model recovery." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.348, + 0.221, + 0.362 + ], + "angle": 0, + "content": "5.2. Human NeRF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.581 + ], + "angle": 0, + "content": "In this section, we benchmark popular NeRF-based methods for 3D humans on SynBody, validating the effectiveness and great potential of our dataset in human neural rendering. Our benchmark is built upon three perspectives according to the purpose of synthesis tasks, which can be categorized into novel view, novel pose, and novel identity. Methods for Benchmark. We benchmark five methods in total, including the vanilla NeRF [34], NeuralBody [41] and HumanNeRF [50] for novel view synthesis, AnimNeRF [40] for novel pose synthesis, NHP [26] for generalizable human NeRF (novel identity synthesis). Except for NHP, all methods are trained in a person-specific manner, taking 4 views of the first 250 frames for training and the rest views and frames for evaluations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Evaluation Protocols. We follow [34, 50] to evaluate all methods using three standard metrics: Peak Signal-to-Noise Ratio (PSNR), Structural SIMilarity index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS [56]). To reach a consensus among all methods, all metrics are computed over the whole image with a black background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.471, + 0.839 + ], + "angle": 0, + "content": "Main Results. Benchmark results are reported in Table 7. We observe that all methods achieve comparable performances as real human data on SynBody. Models (NeuralBody and AnimNeRF) which rely on accurate SMPL estimation and blending weights perform better on novel poses. We attribute it to our layer-wise design which offers ground truth SMPL parameters for human NeRF training. Besides, we present visualization results in Figure 7, observing that the diverse motions and appearances, as well as loose garments in SynBody, pose further challenges in the field of neural rendering of 3D humans." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.195, + 0.86 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We present SynBody, a large-scale synthetic dataset that features a substantial number of subjects and high-quality" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.249, + 0.623, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.344, + 0.599, + 0.354 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.25, + 0.757, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.344, + 0.712, + 0.354 + ], + "angle": 0, + "content": "NeRF" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.25, + 0.892, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.344, + 0.863, + 0.355 + ], + "angle": 0, + "content": "NeuralBody" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.356, + 0.623, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.452, + 0.598, + 0.461 + ], + "angle": 0, + "content": "HumanNeRF" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.356, + 0.757, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.452, + 0.726, + 0.461 + ], + "angle": 0, + "content": "AnimNeRF" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.356, + 0.892, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.452, + 0.845, + 0.461 + ], + "angle": 0, + "content": "NHP" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.47, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Figure 7: Novel view synthesis of different human NeRF methods on SynBody." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.665 + ], + "angle": 0, + "content": "3D annotations. At the core is a clothed human model with multiple layers of representation. Our experiments demonstrate the effectiveness of SynBody on both human mesh recovery and human NeRF. Future research can leverage SynBody for developing and evaluating methods to predict body and cloth simultaneously. Furthermore, the high controllability of the synthetic dataset offers ample opportunities for further improvements, such as the incorporation of contact labels for human-scene interaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.774 + ], + "angle": 0, + "content": "Societal Impacts. Even though SynBody is a synthetic dataset, the assets utilized in its creation might not be well-balanced. Although hairstyles and skin colors are chosen at random, they are designed to avoid racial biases. Yet, other elements, such as body shapes and clothing, might not be as balanced, posing a potential for bias in the resultant human models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This study is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2021-08-019), the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20290" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Marvelous designer, 2023. https://www.marvelousdesigner.com.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.17 + ], + "angle": 0, + "content": "[2] Render people, 2023. https://renderpeople.com. 4,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.468, + 0.24 + ], + "angle": 0, + "content": "[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.468, + 0.297 + ], + "angle": 0, + "content": "[4] Rao Anyi, Jiang Xuekun, Guo Yuwei, Xu Linning, Yang Lei, Jin Libiao, Lin Dahua, and Dai Bo. Dynamic storyboard generation in an engine-based virtual environment for video production. arXiv preprint arXiv:2301.12688, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[5] Eduard Gabriel Bazavan, Andrei Zanfir, Mihai Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Hspace: Synthetic parametric humans animated in complex environments. arXiv preprint arXiv:2112.12867, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.468, + 0.437 + ], + "angle": 0, + "content": "[6] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.468, + 0.548 + ], + "angle": 0, + "content": "[7] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, Fangzhou Hong, Mingyuan Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. HuMMan: Multi-modal 4d human dataset for versatile sensing and modeling. In 17th European Conference on Computer Vision, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 557-577. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.468, + 0.605 + ], + "angle": 0, + "content": "[8] Zhongang Cai, Mingyuan Zhang, Jiawei Ren, Chen Wei, Daxuan Ren, Zhengyu Lin, Haiyu Zhao, Lei Yang, and Zwei Liu. Playing for 3d human recovery. arXiv preprint arXiv:2110.07588, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.468, + 0.689 + ], + "angle": 0, + "content": "[9] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 20-40. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.468, + 0.731 + ], + "angle": 0, + "content": "[10] MMHuman3D Contributors. Openmmlab 3d human parametric model toolbox and benchmark. https://github.com/open-mmlab/mmhuman3d, 2021.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[11] J. C. Gower. Generalized procrustes analysis. Psychometrika, 1975. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[12] Shawn Hargreaves and Mark Harris. Deferred shading. In Game Developers Conference, volume 2, page 31, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.842 + ], + "angle": 0, + "content": "[13] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[14] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[15] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[16] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[17] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC, pages 1-11. British Machine Vision Association, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[18] Sam Johnson and M. Everingham. Learning effective human pose estimation from inaccurate annotation. CVPR 2011, pages 1465-1472, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[19] H. Joo, N. Neverova, and A. Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. ArXiv, abs/2004.03686, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[20] H. Joo, Tomas Simon, Xulong Li, H. Liu, L. Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart C. Nabbe, I. Matthews, T. Kanade, S. Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:190-204, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[21] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, pages 7122-7131, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[22] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In Computer Vision and Pattern Recognition (CVPR), 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[23] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[24] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. arXiv preprint arXiv:2104.08527, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[25] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[26] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[27] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[28] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20291" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[30] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.471, + 0.278 + ], + "angle": 0, + "content": "[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.469, + 0.336 + ], + "angle": 0, + "content": "[33] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.339, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.469, + 0.438 + ], + "angle": 0, + "content": "[35] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In CVPR, pages 2308-2317, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[36] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannot: Neural annotator for 3d human mesh training sets. In Proceedings of the Conference on Computer Vision and Pattern Recognition, pages 2299-2307, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.5, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[37] Hui En Pang, Zhongang Cai, Lei Yang, Tianwei Zhang, and Ziwei Liu. Benchmarking and analyzing 3d human pose and shape estimation beyond algorithms. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[38] Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regression analysis. In Proceedings Conf. on Computer Vision and Pattern Recognition (CVPR), June 2021. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[39] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[40] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[41] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In CVPR, pages 722-731, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[43] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[44] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the International Conference on Computer Vision, pages 1749-1759, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[45] L. Sigal, A. O. Balan, and Michael J. Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision, 87:4-27, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[46] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In Advances in Neural Information Processing Systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[47] Gül Varol, J. Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, I. Laptev, and C. Schmid. Learning from synthetic humans. CVPR, pages 4627-4635, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.343, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[48] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.398, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[49] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. arXiv preprint arXiv:2303.13796, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.468, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[50] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In CVPR, pages 16210–16220, June 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.537, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[51] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3d human shape and articulated pose models. In CVPR, pages 6184-6193, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.606, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[52] Ailing Zeng, Xuan Ju, Lei Yang, Ruiyuan Gao, Xizhou Zhu, Bo Dai, and Qiang Xu. Deciwatch: A simple baseline for \\(10 \\times\\) efficient 2d and 3d pose estimation. In ECCV, pages 607-624. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[53] Ailing Zeng, Xiao Sun, Lei Yang, Nanxuan Zhao, Minhao Liu, and Qiang Xu. Learning skeletal graph neural networks for hard 3d pose estimation. In ICCV, pages 11436-11445, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[54] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: a plug-and-play network for refining human poses in videos. In ECCV, pages 625-642. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[55] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. arXiv preprint arXiv:2207.06400, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.829, + 0.892, + 0.897 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 9" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20292" + } + ] +] \ No newline at end of file diff --git a/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/f251ae55-2ca7-4578-ad92-a51119c290a6_origin.pdf b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/f251ae55-2ca7-4578-ad92-a51119c290a6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7d7883daf5c38dd81484486742994918c662a45f --- /dev/null +++ b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/f251ae55-2ca7-4578-ad92-a51119c290a6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b36fb1db3a876f237607cadc55ec9ba7470cccf771f4df3109af98aaccf2d77 +size 8831274 diff --git a/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/full.md b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5547f3d19bc942c6690aba208ea7ecf89de4bcb9 --- /dev/null +++ b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/full.md @@ -0,0 +1,348 @@ +# SynBody: Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling + +Zhitao Yang $^{1,*}$ Zhongang Cai $^{1,2,3,*}$ Haiyi Mei $^{1,*}$ Shuai Liu $^{2,*}$ Zhaoxi Chen $^{3,*}$ Weiye Xiao $^{1}$ Yukun Wei $^{1}$ Zhongfei Qing $^{1}$ Chen Wei $^{1}$ Bo Dai $^{2}$ Wayne Wu $^{2}$ Chen Qian $^{1}$ Dahua Lin $^{4}$ Ziwei Liu $^{3,\dagger}$ Lei Yang $^{1,2,\dagger}$ SenseTime Research Shanghai AI Laboratory + $^{3}$ S-Lab, Nanyang Technological University The Chinese University of Hong Kong *Equal Contribution Corresponding Author https://synbody.github.io/ + +![](images/9d714c9f937ad5905a6fd480c195d1321a856f80ab64e94095837f3ce9f8b99c.jpg) + +![](images/fdc330cf86a5e26591a6b140ceaa7afc71afcd6ea8156a3482eaf86df79c5644.jpg) +Human Pose and Shape Estimation + +![](images/5bcf5e98b5647f462842e1562e5caf76c30415ebdf7ee24014c0147c4b26733b.jpg) +Novel View Synthesis + +![](images/e9a25979a5c4acb49c79e014b8a3b17f6e74815aa26d26cad2be2d9a73cc66e5.jpg) +Various Environments + +![](images/79763334979363e9def77ecf007ce4b0df9ecd49bde7e3d9c5e62365f05f83a7.jpg) +Layered Human Models + +![](images/66844f706f30d6da5ef784237c2b597e83eb7aefc0807140d7e7b615db517eab.jpg) +Rich Annotations +Figure 1: SynBody is a large-scale synthetic dataset with a massive number of subjects and high-quality annotations. It supports various research topics, including human pose and shape estimation (HPS) and novel view synthesis for human (Human NeRF). + +# Abstract + +Synthetic data has emerged as a promising source for 3D human research as it offers low-cost access to large-scale human datasets. To advance the diversity and annotation quality of human models, we introduce a new synthetic dataset, SynBody, with three appealing features: 1) a clothed parametric human model that can generate a diverse range of subjects; 2) the layered human representation that naturally offers high-quality 3D annotations to support multiple tasks; 3) a scalable system for produc + +ing realistic data to facilitate real-world tasks. The dataset comprises 1.2M images with corresponding accurate 3D annotations, covering 10,000 human body models, 1,187 actions, and various viewpoints. The dataset includes two subsets for human pose and shape estimation as well as human neural rendering. Extensive experiments on SynBody indicate that it substantially enhances both SMPL and SMPL-X estimation. Furthermore, the incorporation of layered annotations offers a valuable training resource for investigating the Human Neural Radiance Fields(NeRF). + +# 1. Introduction + +The fields of 3D human perception [21, 23-25, 37, 53] and human reconstruction [13, 14, 26, 40, 41] have become increasingly important, but the lack of available data has limited their development. Collecting real human data on a large scale is challenging due to privacy concerns and time constraints. Therefore, exploring the use of synthetic human datasets has become a critical avenue of research. + +Despite the great potential, existing synthetic datasets [5, 8,38,47] suffer from limitations such as the number of available human models and the quality of annotations. The main reason lies in that synthetic human datasets rely on real scans for rendering, which poses three key obstacles. Firstly, it is challenging to expand the types of body shapes, poses, and clothing available in the dataset. Secondly, as the human models are scanned with clothing, the 3D annotations obtained through fitting are prone to errors. Thirdly, it is difficult to obtain annotations of body and clothing separately. To address these issues, we develop a new synthetic dataset termed SynBody. The dataset includes 1.2 million frames with corresponding ground-truth 3D human body annotations. It covers 10,000 human body models, 1,187 motions, and 26,960 video clips with 2.7M SMPL/SMPL-X annotations. + +At the heart of SynBody is the layered parametric human model, which constructs the clothed human model in a bottom-up manner. SMPL-X [39] is a widely used parametric human model, capable of sampling human models with various body shapes. However, it lacks the ability to model clothing, limiting its applicability when synthesizing realistic human models. To overcome this limitation, we introduce SMPL-XL, a parametric human model based on SMPL-X in a layered representation. SMPL-XL enriches the SMPL-X model in three aspects: (1) Hair system: adding hair and beards to the FLAME [27] model, with 32 types of hair and 13 types of beards; (2) Garment and accessories: adding procedural clothes to the SMPL-X body, including coats, shirts, pants, skirts, shoes, and glasses; (3) Texture: in addition to adding rich geometry, SMPL-XL also adds rich textures for sampling various skin colors and clothing textures. + +The designed SMPL-XL is capable of automatically generating a large number of human models with high-quality annotations. We therefore generate 10,000 clothed human models by sampling various body shapes, clothing styles, hairstyles, accessories, and textures. Notably, the use of the SMPL-X model as the base body model guarantees that the parametric human annotations are always accurate, obviating the need for the necessity for annotations through fitting. Furthermore, as the clothes are explicitly attached to the surface of the human body, layered annotations for body and clothes are available. + +To generate a large-scale dataset with high diversity and + +high-quality annotations, we design a scalable and automatic system to render images and annotations. We first animate the 10,000 dressed human models by retargeting motions from a large motion library [31]. Subsequently, we design an algorithm to place human models in the scene without piercing. Multiple cameras are then placed by evaluating self-occlusion, inter-occlusion, and view diversity, and the rendering module renders the assets into images with corresponding annotations. + +With SynBody, we launch two tracks that support human pose and shape estimation and human neural rendering, respectively. Experiments show that SynBody is more effective than AGORA under the same amount of training data for human pose and shape estimation. With diverse and large-scale training data, SynBody achieves significant performance gains on both SMPL and SMPL-X estimation. In terms of human neural rendering using neural radiance fields (i.e. NeRF [34]), benchmarking existing approaches on SynBody shows that it has comparable performance as real human data. Furthermore, with the layered annotations which offer accurate SMPL parameters, we observe that current human NeRF approaches are sensitive to the accuracy of estimated SMPL. + +In summary, SynBody is a large-scale synthetic dataset for human perception and modeling, with three main contributions: (1) It constructs clothed subjects and samples 10,000 animatable subjects, which is an order of magnitude higher than existing datasets. (2) The clothed subjects are constructed with an explicit cloth model, thus it provides layered 3D annotations of the human body and clothing, which are not available in previous datasets. (3) Experiments on SynBody achieve promising results on both human perception and modeling, emphasizing the importance of diversity and annotation quality for downstream tasks. + +# 2. Related Works + +Human Parametric Models. Several 3D human parametric models, such as SMPL [30], SMPL-X [39], and GHUM [51], have been developed to generate 3D human meshes from parameters that represent the human pose and shape using linear blend skinning. SMPL-X [39] extends SMPL [30] by combining FLAME [27] and MANO [43] for the head and hands, respectively, and is trained on a large number of real scans to provide a strong basis for shape variations. However, SMPL-X only produces naked body meshes, and we aim to enhance its realism by building a layered parametric model that includes hair, clothes, and accessories. The proposed model leverages the shape basis of SMPL-X while providing realistic dressed human meshes. + +Human Pose and Shape Estimation. Several methods [21, 23-25, 37, 49, 52-54] have been proposed to estimate 3D human pose and shape parameters. HMR [21] directly regresses these parameters in an end-to-end manner, while + +Table 1: Comparisons of 3d human dataset. We compare SynBody with existing datasets. We divide datasets into three types: real (R), synthetic (S), and mixed (M). SynBody constructs 10,000 animatable subjects, which is an order of magnitude higher than any existing datasets and brings competitive scale and diversity. "ITW" stands for "In-the-Wild" in the table. + +
DatasetTypeITWVideo#Views#SMPL#Seq#Subj.#MotionsGT format
HumanEva [45]R-✓4/7NA7463DJ
Human3.6M [15]R-✓4312K83911153DJ, SMPL
MPI-INF-3DHP [32]M✓✓1496K16883DJ
3DPW [48]R✓✓132K6018*SMPL
Panoptic Studio [20]R-✓480736K480~100*3DJ
EFT [19]R✓-1129KNAManyNASMPL
ZJU-MoCap [41]R-✓21180K999SMPL,mask
SURREAL [47]S✓✓16.5MNA1452KSMPL
AGORA [38]S✓-1173KNA>350NASMPL, SMPL-X, mask
HSPACE [5]S✓✓5-NA100×16100GHUM/L, mask
GTA-Human [8]S✓✓11.4M20K>60020KSMPL
BEDLAM [6]S✓✓-1M10K2712,311SMPL-X, mask
SynBodyS✓✓42.7M27K10,0001,187SMPL, SMPL-X, mask
+ +SPIN uses an optimization step [25] to guide the learning process towards pseudo-3D labels. PARE [24] employs part attention to tackle occlusion, and VIBE [23] leverages temporal information in videos for SMPL estimation. Apart from building low-cost data collection solutions [7], synthetic datasets have become a promising alternative to efficiently scale up data. SURREAL [47] renders textured SMPL body models in real-image backgrounds but does not account for cloth geometry, resulting in unrealistic subjects. AGORA [38] renders real human scans in a virtual world and provides high-quality synthetic data for image-based approaches. HSPACE [5] places animated human models in various scenes to provide training data for video-based methods, and increases the variation of human shape via refitting. GTA-Human [8] captures videos and optimizes corresponding SMPL annotations from video games. However, the diversity of subjects in current datasets is limited by either body shapes or cloth types. + +Expressive Human Pose and Shape Estimation. As face and hand are also crucial for human perception, some efforts [35,44,55] have been made to whole-body human pose and shape estimation. ExPose [9] introduces three experts to predict parameters for body, hand, and face, and merge them in a copy-paste strategy. Hand4Whole [35] further improves the prediction of wrist and finger poses by leveraging selected hand joint features. Predicting both hands and faces makes the dataset much more difficult to obtain than just predicting the body. In order to increase the diversity of real data, AGORA [38] provides image-based SMPL-X annotations by fitting SMPL-X on the scanned human models. While a very recent work, BEDLAM [6], also enhances the SMPL-X model with clothing, hair, and changes in skin tones. However, it employs a manual approach for creating garments and utilizes pre-made skin textures, making + +it challenging to generate a large number of human body models. In contrast, our body model construction process is fully automated, encompassing clothing, hair, textures, and more. This enables easier scalability to generate human body models on a large scale. + +Human NeRF. NeRF [34] has demonstrated impressive photo-realistic view synthesis by learning implicit fields of density and color. Yet, human motions are more challenging to learn due to dynamic deformation fields. NeuralBody [41] incorporates prior from a statistical body template to learn dynamic sequence, while Animatable NeRF [40] proposes to reconstruct an animatable human model that generalizes to new poses. Furthermore, NHP [26] and KeypointNeRF [33] achieve generalizability to unseen identities and poses. Several datasets have been adapted to study human NeRF. ZJU-MoCap [41] captures 9 human subjects with 21 synchronized cameras, providing fitted human body model parameters as well as the foreground mask. Human3.6M [16] collects 11 human subjects with 4 cameras, using a marker-based motion capture system. A-NeRF [46] generates a synthetic dataset using SURREAL [47] to study factors that affect visual quality. + +While the aforementioned tasks are interrelated and often draw upon similar datasets, as shown in Table 1, datasets are limited in two key aspects. Firstly, obtaining real human models is challenging, which restricts the scale of these datasets. Secondly, 3D annotations are typically acquired through optimization, which introduces errors and cannot provide layered annotations. In contrast, with the designed SMPL-XL model, SynBody provides 10,000 subjects with diverse body shapes and clothing, along with layered annotations that include accurate SMPL and SMPL-X. Built on top of this layered human model, SynBody comprises two subsets for human mesh recovery and human NeRF, respec + +tively, and features the same 10,000 diverse subjects. + +# 3. Synthetic Data Generation System + +![](images/fb00fd8f6e55eb1b726a486e2ce356fe79dcb390b82dd46f504310ab71d16968.jpg) +Figure 2: Synthetic data generation system. It consists of 4 components: 1) human model creation to generate layered human models, 2) motion retargeting to drive human models, 3) scene composition to place actors and cameras, and 4) rendering and outputs to generate multi-modal dataset. + +Similar to movie production [4], our system comprises 4 components (depicted in Figure 2): (1) a layered parametric human model creation service, a scalable process to generate layered human models, (2) a motion retargeting module to apply motions from various sources to layered human models, (3) a scene composition module to place 3D actors and objects into a 3D scene, setting up cameras, (4) a 3D rendering engine and a multi-modal data annotation generator. Our infrastructure enables the generation of high-quality synthetic data for various computer vision tasks. + +# 3.1. Layered Parametric Human Model + +Parametric human models like SMPL-X [39] provide the ability of create rigged body models with various body shapes. However, the lack of available textures limits its application in data generation. We designed a module to perform an automatic process that combines SMPL-X with procedural garments and accessories, hair system and textures, producing realistic and diverse body models. + +Body shape. SMPL-X [39] body model has a 3D mesh whose vertex locations are controlled by parameters for pose $\theta$ , shape $\beta$ , and facial expression $\psi$ . By modifying shape $\beta$ , we obtain 3D meshes of various human heights and weights. And by alternating pose $\theta$ , meshes can be driven to perform various poses. + +Garment Model. Our garment is generated as a separate layer on top of the body. Following the industrial garment-making workflow, we designed garment patterns of various styles. Notice that different parts of garment pieces are connected with sewing lines, e.g., the red line in Figure 4 (a). Then, we stitch patterns onto the body at T-Pose utilizing a physical simulator [1]. Specifically, we first manually move the garment pieces to roughly align them with the body. During simulation, vertices between two ends of the seam gradually shrink until they completely pooled together. Figure 4 (b) demonstrates the final draped gar + +ment. For garment animation, we bind every vertex in the garment to the closest point in the body. Then, the skinnig weight and blend shape of the body mesh are assigned to garment vertices. This makes our garment model easily integrated with existing skeletal pipelines with little computational overhead. + +Particle hair system. Our method uses a prefabricated particle system to generate realistic hair on a head-shaped mesh. To achieve this, a template of hairstyle or facial hair $T_{hair}$ would attach to the vertices on the mesh which are marked with different areas $V_{hair}$ (including fringe, top, temporal bone, occipital bone, and bottom area). Designers draw multiple sets of guidelines $L_{guide}$ with varying shapes on different areas. Each set of guidelines comprises a collection of Bezier curves that are utilized to accurately constrain the flow of hair strands. Furthermore, the shape of the hair strands can be adjusted by length $P_{length}$ and curliness $P_{curliness}$ . The entire hair system is composed of multiple sub-particle systems $\{V_{hair}, L_{guide}, P_{length}, P_{curliness}\}$ . + +Accessories. We also add template accessories $T_{\text{accessories}}$ to our model, such as glasses, shoes, hats, and headphones. All accessories are pre-assembled on an SMPL-X template with a uniform body shape. These accessories can be transferred to models with the same topology, ensuring consistent deformation across all models in accordance with changes in the shape $\beta$ of SMPL-X. The corresponding bone weights of the human body model's vertices are transferred to the nearest accessories' vertices, ensuring that these accessories are correctly driven by the armature. + +Procedural texture. The procedural garment textures $T_{\text{procedural}} = \{T_{\text{pattern}}, T_{\text{decals}}, T_{\text{bump}}, P_{\text{mapping}}\}$ used in clothing are composed of multiple pre-set textures by alpha blending as demonstrated in Figure 4 (c). The pattern texture $T_{\text{pattern}}$ and decals texture $T_{\text{decals}}$ serve as layered masks to elevate the style and color, while the bump texture $T_{\text{bump}}$ functions as a height map, adding detailed normal information to the clothing's surface. The mapping parameters $P_{\text{mapping}}$ control the coordination of all the textures in the UV space. Besides, to procedurally create body mesh textures, we build a similar texture template SMPL-X, in which we pre-draw layered masks to separate different body features, such as skin, lips, and freckles, allowing for color adjustment and blending in specific regions as in Figure 3. + +Combining all the elements above, as shown in Figure 5, we obtain human body models with the same high quality as RenderPeople [2]. Besides, it is challenging to expand the types of body shape and clothing available for real scans, but our model can be easily scaled by randomly sampling each component. + +# 3.2. Motion Retargeting + +Retargeting of skeletal animations. Our motion retargeting module allows for the transfer of motion data from vari + +![](images/0b734aa6d69cd069dfd6dde50d42f857708d02b1bd3d32670ae5dff13f7e59d1.jpg) +(a) Naked human model generation + +![](images/e8bd40e463f3736985786ed243195e9f19de9a3c722317ffea4cbd0a7965b967.jpg) +(b) Layered human model generation + +![](images/ccf98bd231a782d1e61ecc3c19426e73222c0eeaf820d538e1b0e3bea8563a85.jpg) +Figure 3: Layered parametric human model creation: (a) Generation of a naked human model using parametric body model [39], procedural body texture colors, and particle hair System; (b) Integration of garments and accessories onto the naked body model. +Figure 4: Demonstrations of Garment and Procedural Texture. (a) garment pattern with sewing lines, (b) simulated garment under the body in canonical space, (c) procedural textured garment whose style and color are elevated by layered masks provided by patterns and decals. + +ous sources, such as academic motion datasets, motion captures, and artist-crafted sources, to SMPL-XL model skeletons. Despite variations in bone names, bone lengths, and rest pose bone rotations, the source skeletons are typically structurally similar to the target skeletons. + +Pose frames in motion clips contain root translation $T(t) \in \mathbb{R}^3$ and rotations of each bone in the corresponding parent bone's space $\{R_i(t) \in \mathbb{R}^3 | i = 0,1,\dots,n\}$ . Following the forward kinematics (FK) manner, each bone's rotation relative to the model space can be calculated: + +$$ +\widehat {R} _ {i} (t) = \widehat {R} _ {p (i)} (t) \cdot R _ {i} (t) \tag {1} +$$ + +where $\widehat{R}_i(t)$ is the rotation of $R_{i}(t)$ in model space at $t$ frame, and $p(i)$ indicates the parent of $i$ . Specially, root bones have no parents so $\widehat{R}_0(t) = R_0(t)$ . + +To retarget motion from one skeleton to another, we assume the source and the target motion can drive the corresponding bones of both skeletons to the same rotation in model space. We have T-posed frames of skeletons by manually posing them as T-posing, and treat T-posed as all mo + +![](images/06e4ca65c4543ff09c958eb3f037498fd5e1b563189f7d448867a30383c07b32.jpg) +Figure 5: A Demonstration of comparison between commonly used human models in existing synthetic datasets [5, 8,38,47] and our SMPL-XL. We obtain high-quality models equal to RenderPeople [2], both are much better than SUR-REAL [47], and our model has the capability of scaling up easily by sampling various assets and body shapes. + +tions' first frame pose. So motions relative to their T-posedr frame can be easily obtained: + +$$ +\widehat {R} _ {i} (t) = \widehat {R} _ {i} (0) \cdot \widehat {R} _ {i _ {- s r c}} (t) \cdot \widehat {R} _ {i _ {- s r c}} ^ {- 1} (0). \qquad (2) +$$ + +Considering skeletons have different bone lengths, which could result in "sliding feet" artifacts on target skeletons. We simply scale $T$ according to the ratio of pelvis bones' height to mitigate them: $T(t) = \frac{H_{pelvis}}{H_{pelvis,src}} \cdot T_{src}(t)$ . + +SMPL-X Annotations. Considering SMPL-XL body shapes are sampled from SMPL-X, the shape $\beta$ can be derived directly from the corresponding SMPL-X. Pose $\theta$ in the model space is calculated in the motion retargeting module. And in the scene composition module, models are placed in a 3D scene with camera space locations $T_{c}$ and rotations $R_{c}$ . Pose $\theta_{c}$ in camera space is calculated by applying those world transformations to $\theta$ . So SMPL-X annotations are constructed with $\{\beta, \theta_{c}\}$ . + +SMPL Annotations. Although SMPL-XL naturally provides accurate SMPL-X annotations, SMPL cannot be derived directly. Thus, we need to refit the SMPL parameters. + +The optimization process consists of two steps. First, we fit the shape $\beta$ of all human models under the T-pose to its corresponding SMPL-X. Secondly, for each sequence, it is initialized with fitted $\beta$ and its original pose. we fix $\beta$ while fitting the body pose $\theta$ for SMPL. + +# 3.3. Scene Placement + +To place $N_{p}$ subjects in a large scene with $N_{o}$ objects, we primarily follow three principles: standing on the ground, avoiding human-object and human-human penetration. To prevent subjects from floating in the air, the root position of a subject should align with the ground height. A sequential decision-making approach is used to find a suitable position for each subject, i.e., placing one subject at a time. An object is represented by $o_{i} = \{x_{i},y_{i},l_{i},w_{i}\} \in \mathbb{R}^{4}$ , where $\{x_i,y_i\}$ is the center of the object's axis-aligned bounding box projected onto the ground with length $l_{i}$ and width $w_{i}$ . Different from static objects, to avoid collisions between moving subjects at any one time, a subject with specified body shape and motion is simplified to $q_{i} = \{x_{i},y_{i},l_{i},w_{i}\} \in \mathbb{R}^{4}$ , which is the smallest axis-aligned box that envelops all bounding boxes across frames. + +To avoid human-object penetration and potential human-human collision, the solution $p_i^* = \{x_i^*, y_i^*\}$ of a character with the shape of $\{l_i, w_i\}$ should satisfy: + +$$ +I \left(\left\{x _ {i} ^ {*}, y _ {i} ^ {*}, l _ {i}, w _ {i} \right\}, o _ {j}\right) = 0, j = 1, \dots , N _ {o} \tag {3} +$$ + +$$ +I \left(\left\{x _ {i} ^ {*}, y _ {i} ^ {*}, l _ {i}, w _ {i} \right\}, q _ {k}\right) = 0, k = 1, \dots , N _ {p}, \tag {4} +$$ + +where $I(\text{box}_1, \text{box}_2)$ denotes the overlapping area between two boxes and $N_p$ is the number of subjects already placed in the scene. Besides, distance constraint is used to prevent subjects from excessive dispersal. The problem is solved by grid search. Typically multiple solutions are available and we randomly sample one each time. + +# 3.4. Camera Placement + +Once the positions of all subjects have been organized, the next task involves placing $N_{c}$ cameras in suitable locations. To ensure that cameras would not be placed inside any objects or subjects, the 3D version of Eq. (3) and Eq. (4) are applied. In addition, we evaluate the suitability of a candidate camera by the following metrics: the distance from the camera to the subjects, the camera's pitch angle, and the degree of occlusion of each subject in the camera's view. + +The distance denoted as $L$ from the mean position of all subjects $\bar{p}$ to the camera is restricted. $L_{max}$ is set to prevent an unreasonably small proportion of subjects in the image. To control the visibility of subjects, $L_{min}$ is defined as $\frac{\lambda}{\sin(\alpha / 2)}\max_i||pv_i - \bar{p} ||_2$ , where $i = 1,\dots ,N_c$ , $\alpha$ is the field of view of the camera, and $\lambda$ is a hyperparameter that determines the probability of all subjects being within + +the camera's view. To estimate the degree of occlusion for each subject, rays are randomly and uniformly cast from the camera to each subject's body, and the percentage of blocked rays by other objects is determined. More details can be found in the Sup. Mat. + +# 3.5. Rendering and Annotations + +Using a high-quality rendering pipeline in Unreal Engine 5, SynBody is rendered in multi-view with large 3D environments from the Unreal Marketplace for rich background information and dynamic lighting. Leveraging the G-buffer [12], our system simultaneously generates photorealistic RGB images and annotations, along with accurate ground truth for segmentation masks, optical flow, depth maps, normal maps, and other ground-truth labels. + +# 4. SynBody Dataset + +# 4.1. Dataset Statistics + +SynBody comprises 10,000 unique SMPL-XL models randomly created with different body shapes and genders. Each model is then combined with the following assets: (1) hairstyles $T_{hair}$ sampled from 45 particle hairs; (2) garments $G_{tmp}$ sampled from 68 clothing models containing multiple outfits; (3) procedural texture $T_{procedural}$ generated by sampling $T_{pattern}$ , $T_{decals}$ , and $T_{bump}$ from 1,038 template textures, and color values are randomly sampled; (4) accessories $T_{accessories}$ sampled from 46 template assets. To ensure the validity of the motions, we select a subset from AMASS [31]. Based on the BABEL annotations [42], we excluded interactive motions and non-ground motions(e.g., swim), as well as filtered out motions with a duration shorter than 2 seconds, leading to a subset with 1,187 motions. For each sequence, we randomly selected 4 human models, and each model is assigned a randomly chosen motion lasting between 2 and 10 seconds. For motions exceeding 2 seconds, we randomly extract a 2-second segment to ensure that each video clip has a length of 60 frames. To enhance the diversity of viewpoints, 4 view positions are generated for each sequence, where each one is randomly sampled from the surface of semi-spheres with varying radii. The system generates 26,960 sequences and 1.2M images with annotations, utilizing 6 vast and realistic scenes created by professional artists. 2.7M SMPL/SMPL-X annotations are provided, excluding highly occluded subjects. More details can be found in the Sup. Mat. + +# 4.2. Human Pose and Shape Estimation + +With 2.7 million SMPLs, we leverage a pretrained regressor to extract 3D keypoints and then project them onto 2D space. Following standard practice in top-down human mesh recovery methods, we generate bounding boxes from the resulting 2D keypoints. + +Table 2: Training popular baseline methods (image-based and video-based) with SynBody on 3DPW test set [48]. R: standard real datasets. S: SynBody. + +
MethodDatasetsMPJPEPA-MPJPEPVE
HMR [21]R112.3067.50141.92
HMRR + S95.0157.62116.10
SPIN [25]R96.9059.20119.70
SPINR + S84.1453.67103.79
PARE [24]R81.7949.36105.27
PARER + S78.9848.46103.86
VIBE [23]R94.8857.08108.59
VIBER + S93.0457.00107.23
+ +# 4.3. Human Neural Rendering + +Given the flexibility of our pipeline, we render a total of 100 multi-view sequences with diverse motions and appearances for benchmarking human NeRFs. All sequences have a length of 300 frames with a resolution of $1024 \times 1024$ , where the motion sequence is randomly sampled from AMASS [31]. Each sequence contains 8 views whose camera positions have uniformly distributed azimuth angles around the human body. We use RGB renderings, binary foreground masks, camera parameters, and SMPL parameters for the benchmark. Thanks to our layered design, we can offer accurate SMPL parameters for human NeRFs, which acts as an important prior for a majority of methods, while keeping the diversity in clothes and motions. + +# 5. Experiment + +In this section, we study the usefulness of SynBody for two popular research directions: human pose and shape estimation and human NeRF. + +# 5.1. Human Pose and Shape Estimation + +Estimating 3D humans represented by SMPL (body-only) and SMPL-X (body, hands, and face) parameters from monocular 2D input has gained substantial attention. SynBody emerges as a scalable and effective solution to tackle the scarcity of paired data in these fields. + +Parametric Human Models. SMPL [30] represents high-dimension human mesh $M(\theta, \beta) \in \mathbb{R}^{6890 \times 3}$ as low-dimension pose parameters $\theta \in \mathbb{R}^{72}$ and shape parameters $\beta \in \mathbb{R}^{10}$ . In addition to SMPL, SMPL-X [39] further consists of additional left and right hand pose parameters $(\theta_{lh}, \theta_{rh} \in \mathbb{R}^{15 \times 3})$ , jaw joint rotation $(\theta_{lh} \in \mathbb{R}^3)$ and face expression parameters $(\phi_f \in \mathbb{R}^{10})$ for an expressive human representation. + +**Benchmarks.** We conduct experiments on two mainstream benchmarks: 3DPW [48] and AGORA [38]. As a widely used benchmark for in-the-wild evaluation, 3DPW encompasses diverse data collected through a mobile phone cam + +Table 3: Training popular baseline methods (image-based only as AGORA does not provide videos but static images only) with SynBody on AGORA validation set [38]. R: standard real datasets. S: SynBody. + +
MethodDatasetsMPJPEPA-MPJPEPVE
HMR [21]R226.7187.72248.35
HMRR + S199.5177.97210.37
SPIN [25]R212.9179.76217.88
SPINR + S196.8176.06205.83
PARE [24]R178.1567.13189.73
PARER + S169.9364.37179.81
+ +era, paired with SMPL annotations. As for AGORA, it is a recent synthetic dataset that features challenging scenes with person-person occlusion. AGORA provides both SMPL and SMPL-X annotations. + +Evaluation Metrics. To evaluate the quality of predicted parametric human models, we employ the standard metrics: MPJPE (Mean Per Joint Position Error), which is calculated as the average $L_{2}$ distance over 3D keypoints regressed from parametric models; PA-MPJPE (Procrustes Aligned Mean Per Joint Position Error), which is MPJPE with applying Procrustes Alignment [11] on the predicted keypoints to match the ground truth before error computation; PVE (Per Vertex Error), which is the average $L_{2}$ distance between predicted and ground truth mesh vertices. + +# 5.1.1 SMPL Estimation + +Methods. To gauge the usefulness of SynBody, we conduct experiments with four milestone works: HMR [21], SPIN [25], and PARE [24] are image-based methods and VIBE [23] is a video-based method. We follow the hyperparameter configurations provided in MMHuman3D [10]. + +Training Data. In the following experiments, we refer to the standard baskets of real datasets for SMPL estimation as "R". For HMR and SPIN, R consists of H36M [16], MPI-INF-3DHP [32], LSP [17], LSPET [18], MPII [3] and MSCOCO [29], whereas in VIBE, R consists of MPI-INF-3DHP and InstaVariety [22]. "S" denotes SynBody data with 2.7M training instances with SMPL annotations. + +Main Results. We collate experiment results on 3DPW in Table 2. Each baseline is fine-tuned with a mixture of real datasets (R) and SynBody (S), leading to significant performance gains across different baseline models, even on strong baseline PARE [24] (2.81 mm and 1.41 mm improvements on MPJPE and PVE). Notably, SynBody can be used to train video-based methods such as VIBE as it contains video sequences instead of static images. Table 3 illustrates that training with SynBody exhibits remarkable improvements across methods, when compared with + +![](images/9088f1e0e8c6a5eba3e196a484217c549eb830f322c7c1678c1d24b59ade6080.jpg) +Figure 6: Impact of the amount of SynBody data. The horizontal axis represents the amount of SynBody data that is varied by multiples of real data. The baseline model is HMR and tested on the 3DPW test set. + +Table 4: Comparison between SynBody-100K and AGORA on 3DPW test set. “A” means AGORA, and “S100K” means SynBody-100K which is a subset of SynBody datasets with over 100K SMPL annotations. For a fair comparison, the total number of SMPL annotations is close to that of AGORA. + +
MethodDatasetsMPJPEPA-MPJPEPVE
HMRR + A101.6157.85123.82
HMRR + S100K95.2857.68119.18
SPINR + A88.4454.97110.35
SPINR + S100K85.5254.12105.45
PARER + A85.3448.39109.77
PARER + S100K79.4247.80102.45
+ +Table 5: Effectiveness of SMPL-XL. R: real datasets (MSCOCO, MPI-INF-3DHP, Human3.6M, LSP, LSPET, MPII). S100K: downsampled SynBody with 100K instances.*: SURREAL-style. + +
MethodDatasetsMPJPEPA-MPJPEPVE
PARER81.849.4105.3
PARER+S100K*82.249.7105.6
PARER+S100K79.447.8102.5
+ +the baseline models in AGORA validation set. + +Impact of Data Scale. As our generation system can easily scale up the data size, we train HMR from scratch to study the influence of synthetic data scale. Figure 6 demonstrates that adding more SynBody data generally leads to better performance. These experiments confirm that synthetic data is a valuable complement to real data and serves as a readily scalable training source to supplement the typically limited real data. + +Comparison with AGORA. To evaluate the quality of SynBody, we randomly sample a subset with 100K training instances, "SynBody-100K", which is comparable in size with the popular synthetic dataset AGORA. In Table 4, we + +Table 6: SMPL-X estimation with OSX [28] as the baseline on AGORA validation set (AGORA-val) [38] and 3DPW test set [28]. R: real datasets (MSCOCO, MPII, and Human3.6M). S: SynBody. AGORA-val uses PVE (mm) whereas 3DPW uses MPJPE (mm) and PA-MPJPE (mm). Note the original OSX uses the SMPL head for 3DPW, which we modify to the SMPL-X head. + +
MethodDatasetsAGORA3DPW (Body)
AllHandsFaceMPJPEPA-MPJPE
OSXR168.670.677.2110.263.5
OSXR+S600K155.863.572.092.259.3
+ +show that finetuning with SynBody outperforms that with AGORA across different baseline methods. The results highlight that SynBody demonstrates competitive traits as a training source. + +Effectiveness of SMPL-XL. SMPL-XL enables layered human modeling which cannot be achieved with SURREAL/RenderPeople as Figure 5. We investigate the importance to synthesize data with these realistic considerations of SMPL-XL in Table 5. We compare SynBody with SURREAL-style data that renders body and cloth texture on human mesh surfaces, without actual cloth geometry or hair. It is observed that SURREAL-style data leads to significant performance degradation. + +# 5.1.2 SMPL-X Estimation + +Also known as expressive human pose and shape estimation, SMPL-X estimation requires recovery of body, hands, and face parameters. + +Method and Training Data. In Table 6, we conduct experiments with the recent SoTA, OSX [28], as the base model. We compare the original baseline with training with a mixture of real datasets ("R" denotes COCO [29], MPII [3] and Human3.6M [16] with pseudo ground truth generated by NeuralAnnot [36]) and SynBody ("S600K" here denotes a downsampled set of 600K instances as the real datasets are smaller for SMPL-X than that for SMPL estimation). Note that in the OSX paper, the values reported on 3DPW are obtained using a SMPL head. Here we standardize the output using the same SMPL-X head for both AGORA and 3DPW. + +Main Results. We observe that adding SynBody in the training of OSX leads to a significant improvement in both overall estimation (more than $10\mathrm{mm}$ ) and part-level (more than $5\mathrm{mm}$ for hands and face) estimation on AGORA. Moreover, in a fair comparison with the same model architecture, training with SynBody leads to $18\mathrm{mm}$ improvements in MPJPE on the 3DPW test set (here we follow the standard protocol to evaluate 14 key joints as 3DPW does not provide SMPL-X annotations). We speculate that it + +Table 7: Benchmark of NeRF-based methods for 3D human neural rendering on SynBody. + +
MethodNovel ViewNovel PoseNovel Identity
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [34]19.390.8620.16219.610.8240.201---
NeuralBody [41]28.940.9660.05725.020.9440.080---
HumanNeRF [50]28.320.9630.06621.970.8790.108---
AnimNeRF [40]27.490.9640.05626.210.9500.068---
NHP [26]25.660.9530.07624.180.9450.08022.460.9270.103
+ +is difficult to obtain accurate SMPL-X annotations, which is mitigated by SynBody as it provides large-scale high-quality SMPL-X labels paired with images that are rendered with diverse backgrounds and lighting conditions, which are beneficial to training a high-performing expressive human parametric model recovery. + +# 5.2. Human NeRF + +In this section, we benchmark popular NeRF-based methods for 3D humans on SynBody, validating the effectiveness and great potential of our dataset in human neural rendering. Our benchmark is built upon three perspectives according to the purpose of synthesis tasks, which can be categorized into novel view, novel pose, and novel identity. Methods for Benchmark. We benchmark five methods in total, including the vanilla NeRF [34], NeuralBody [41] and HumanNeRF [50] for novel view synthesis, AnimNeRF [40] for novel pose synthesis, NHP [26] for generalizable human NeRF (novel identity synthesis). Except for NHP, all methods are trained in a person-specific manner, taking 4 views of the first 250 frames for training and the rest views and frames for evaluations. + +Evaluation Protocols. We follow [34, 50] to evaluate all methods using three standard metrics: Peak Signal-to-Noise Ratio (PSNR), Structural SIMilarity index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS [56]). To reach a consensus among all methods, all metrics are computed over the whole image with a black background. + +Main Results. Benchmark results are reported in Table 7. We observe that all methods achieve comparable performances as real human data on SynBody. Models (NeuralBody and AnimNeRF) which rely on accurate SMPL estimation and blending weights perform better on novel poses. We attribute it to our layer-wise design which offers ground truth SMPL parameters for human NeRF training. Besides, we present visualization results in Figure 7, observing that the diverse motions and appearances, as well as loose garments in SynBody, pose further challenges in the field of neural rendering of 3D humans. + +# 6. Conclusion + +We present SynBody, a large-scale synthetic dataset that features a substantial number of subjects and high-quality + +![](images/3abc8211dd7460ae755962e39207c11572e83457af81e3b75ed119812d13b301.jpg) +Ground Truth + +![](images/4731a51771d0956d16f12af922897f06670a5aae89df6c7d02f4259edbaef5af.jpg) +NeRF + +![](images/626a7baf5cef8df6ed5b464cd3c69d56205fce8bfd44ad0e69773fe1e224c323.jpg) + +![](images/7818d9e5f61954e8601842e765120f1474f2fa929407d320c7dc613ffab1151a.jpg) +HumanNeRF + +![](images/60c2fe7cd237e9bed2005b7578ed82fe43c385707e65a540a3deb4cfd748f000.jpg) +AnimNeRF + +![](images/295aa1aa17272125d0346c224615a1ce63b47877b7f4fe1b22201fa5e45453fa.jpg) +NeuralBody +NHP +Figure 7: Novel view synthesis of different human NeRF methods on SynBody. + +3D annotations. At the core is a clothed human model with multiple layers of representation. Our experiments demonstrate the effectiveness of SynBody on both human mesh recovery and human NeRF. Future research can leverage SynBody for developing and evaluating methods to predict body and cloth simultaneously. Furthermore, the high controllability of the synthetic dataset offers ample opportunities for further improvements, such as the incorporation of contact labels for human-scene interaction. + +Societal Impacts. Even though SynBody is a synthetic dataset, the assets utilized in its creation might not be well-balanced. Although hairstyles and skin colors are chosen at random, they are designed to avoid racial biases. Yet, other elements, such as body shapes and clothing, might not be as balanced, posing a potential for bias in the resultant human models. + +Acknowledgement. This study is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2021-08-019), the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Marvelous designer, 2023. https://www.marvelousdesigner.com.4 +[2] Render people, 2023. https://renderpeople.com. 4,5 +[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 7, 8 +[4] Rao Anyi, Jiang Xuekun, Guo Yuwei, Xu Linning, Yang Lei, Jin Libiao, Lin Dahua, and Dai Bo. Dynamic storyboard generation in an engine-based virtual environment for video production. arXiv preprint arXiv:2301.12688, 2023. 4 +[5] Eduard Gabriel Bazavan, Andrei Zanfir, Mihai Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Hspace: Synthetic parametric humans animated in complex environments. arXiv preprint arXiv:2112.12867, 2021. 2, 3, 5 +[6] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 3 +[7] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, Fangzhou Hong, Mingyuan Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. HuMMan: Multi-modal 4d human dataset for versatile sensing and modeling. In 17th European Conference on Computer Vision, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 557-577. Springer, 2022. 3 +[8] Zhongang Cai, Mingyuan Zhang, Jiawei Ren, Chen Wei, Daxuan Ren, Zhengyu Lin, Haiyu Zhao, Lei Yang, and Zwei Liu. Playing for 3d human recovery. arXiv preprint arXiv:2110.07588, 2021. 2, 3, 5 +[9] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 20-40. Springer, 2020. 3 +[10] MMHuman3D Contributors. Openmmlab 3d human parametric model toolbox and benchmark. https://github.com/open-mmlab/mmhuman3d, 2021.7 +[11] J. C. Gower. Generalized procrustes analysis. Psychometrika, 1975. 7 +[12] Shawn Hargreaves and Mark Harris. Deferred shading. In Game Developers Conference, volume 2, page 31, 2004. 6 +[13] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022.2 +[14] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2 + +[15] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3 +[16] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014. 3, 7, 8 +[17] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC, pages 1-11. British Machine Vision Association, 2010. 7 +[18] Sam Johnson and M. Everingham. Learning effective human pose estimation from inaccurate annotation. CVPR 2011, pages 1465-1472, 2011. 7 +[19] H. Joo, N. Neverova, and A. Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. ArXiv, abs/2004.03686, 2020. 3 +[20] H. Joo, Tomas Simon, Xulong Li, H. Liu, L. Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart C. Nabbe, I. Matthews, T. Kanade, S. Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:190-204, 2019. 3 +[21] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, pages 7122-7131, 2018. 2, 7 +[22] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In Computer Vision and Pattern Recognition (CVPR), 2019. 7 +[23] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 3, 7 +[24] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. arXiv preprint arXiv:2104.08527, 2021. 2, 3, 7 +[25] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 3, 7 +[26] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 9 +[27] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2 +[28] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. 2023. 8 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7, 8 + +[30] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 7 +[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 2, 6, 7 +[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 3, 7 +[33] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In ECCV, 2022. 3 +[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 9 +[35] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In CVPR, pages 2308-2317, 2022. 3 +[36] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannot: Neural annotator for 3d human mesh training sets. In Proceedings of the Conference on Computer Vision and Pattern Recognition, pages 2299-2307, 2022. 8 +[37] Hui En Pang, Zhongang Cai, Lei Yang, Tianwei Zhang, and Ziwei Liu. Benchmarking and analyzing 3d human pose and shape estimation beyond algorithms. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2 +[38] Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regression analysis. In Proceedings Conf. on Computer Vision and Pattern Recognition (CVPR), June 2021. 2, 3, 5, 7, 8 +[39] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 2, 4, 5, 7 +[40] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2, 3, 9 +[41] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 9 +[42] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In CVPR, pages 722-731, 2021. 6 + +[43] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2 +[44] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the International Conference on Computer Vision, pages 1749-1759, 2021. 3 +[45] L. Sigal, A. O. Balan, and Michael J. Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision, 87:4-27, 2009. 3 +[46] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In Advances in Neural Information Processing Systems, 2021. 3 +[47] Gül Varol, J. Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, I. Laptev, and C. Schmid. Learning from synthetic humans. CVPR, pages 4627-4635, 2017. 2, 3, 5 +[48] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 3, 7 +[49] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. arXiv preprint arXiv:2303.13796, 2023. 2 +[50] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In CVPR, pages 16210–16220, June 2022. 9 +[51] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3d human shape and articulated pose models. In CVPR, pages 6184-6193, 2020. 2 +[52] Ailing Zeng, Xuan Ju, Lei Yang, Ruiyuan Gao, Xizhou Zhu, Bo Dai, and Qiang Xu. Deciwatch: A simple baseline for $10 \times$ efficient 2d and 3d pose estimation. In ECCV, pages 607-624. Springer, 2022. 2 +[53] Ailing Zeng, Xiao Sun, Lei Yang, Nanxuan Zhao, Minhao Liu, and Qiang Xu. Learning skeletal graph neural networks for hard 3d pose estimation. In ICCV, pages 11436-11445, 2021. 2 +[54] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: a plug-and-play network for refining human poses in videos. In ECCV, pages 625-642. Springer, 2022. 2 +[55] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. arXiv preprint arXiv:2207.06400, 2022.3 +[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 9 \ No newline at end of file diff --git a/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/images.zip b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..764dfffce3c72b9fa88b70ccd229e6f5d1cb36d9 --- /dev/null +++ b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84d2257d78019e83f8246e90e5f0de67cd7cc7de9dce89fd89d538f7e044cd9b +size 544833 diff --git a/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/layout.json b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..880b4297c54f1aefff4b85a3290e907289a71899 --- /dev/null +++ b/2023/SynBody_ Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling/layout.json @@ -0,0 +1,9621 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 487, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 487, + 140 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 487, + 140 + ], + "type": "text", + "content": "SynBody: Synthetic Dataset with Layered Human Models for 3D Human Perception and Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": "Zhitao Yang" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Zhongang Cai" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{1,2,3,*}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Haiyi Mei" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Shuai Liu" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Zhaoxi Chen" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Weiye Xiao" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Yukun Wei" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Zhongfei Qing" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Chen Wei" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Bo Dai" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Wayne Wu" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Chen Qian" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Dahua Lin" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Ziwei Liu" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{3,\\dagger}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " Lei Yang" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": " SenseTime Research Shanghai AI Laboratory \n" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 160, + 505, + 259 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University The Chinese University of Hong Kong *Equal Contribution Corresponding Author https://synbody.github.io/" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 274, + 364, + 445 + ], + "blocks": [ + { + "bbox": [ + 60, + 274, + 364, + 445 + ], + "lines": [ + { + "bbox": [ + 60, + 274, + 364, + 445 + ], + "spans": [ + { + "bbox": [ + 60, + 274, + 364, + 445 + ], + "type": "image", + "image_path": "9d714c9f937ad5905a6fd480c195d1321a856f80ab64e94095837f3ce9f8b99c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 445, + 203, + 525 + ], + "blocks": [ + { + "bbox": [ + 61, + 445, + 203, + 525 + ], + "lines": [ + { + "bbox": [ + 61, + 445, + 203, + 525 + ], + "spans": [ + { + "bbox": [ + 61, + 445, + 203, + 525 + ], + "type": "image", + "image_path": "fdc330cf86a5e26591a6b140ceaa7afc71afcd6ea8156a3482eaf86df79c5644.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 526, + 206, + 536 + ], + "lines": [ + { + "bbox": [ + 64, + 526, + 206, + 536 + ], + "spans": [ + { + "bbox": [ + 64, + 526, + 206, + 536 + ], + "type": "text", + "content": "Human Pose and Shape Estimation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 211, + 447, + 360, + 522 + ], + "blocks": [ + { + "bbox": [ + 211, + 447, + 360, + 522 + ], + "lines": [ + { + "bbox": [ + 211, + 447, + 360, + 522 + ], + "spans": [ + { + "bbox": [ + 211, + 447, + 360, + 522 + ], + "type": "image", + "image_path": "5bcf5e98b5647f462842e1562e5caf76c30415ebdf7ee24014c0147c4b26733b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 525, + 328, + 536 + ], + "lines": [ + { + "bbox": [ + 241, + 525, + 328, + 536 + ], + "spans": [ + { + "bbox": [ + 241, + 525, + 328, + 536 + ], + "type": "text", + "content": "Novel View Synthesis" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 364, + 274, + 529, + 308 + ], + "blocks": [ + { + "bbox": [ + 364, + 274, + 529, + 308 + ], + "lines": [ + { + "bbox": [ + 364, + 274, + 529, + 308 + ], + "spans": [ + { + "bbox": [ + 364, + 274, + 529, + 308 + ], + "type": "image", + "image_path": "e9a25979a5c4acb49c79e014b8a3b17f6e74815aa26d26cad2be2d9a73cc66e5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 310, + 489, + 320 + ], + "lines": [ + { + "bbox": [ + 400, + 310, + 489, + 320 + ], + "spans": [ + { + "bbox": [ + 400, + 310, + 489, + 320 + ], + "type": "text", + "content": "Various Environments" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 390, + 326, + 502, + 429 + ], + "blocks": [ + { + "bbox": [ + 390, + 326, + 502, + 429 + ], + "lines": [ + { + "bbox": [ + 390, + 326, + 502, + 429 + ], + "spans": [ + { + "bbox": [ + 390, + 326, + 502, + 429 + ], + "type": "image", + "image_path": "79763334979363e9def77ecf007ce4b0df9ecd49bde7e3d9c5e62365f05f83a7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 430, + 494, + 441 + ], + "lines": [ + { + "bbox": [ + 399, + 430, + 494, + 441 + ], + "spans": [ + { + "bbox": [ + 399, + 430, + 494, + 441 + ], + "type": "text", + "content": "Layered Human Models" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 381, + 445, + 515, + 523 + ], + "blocks": [ + { + "bbox": [ + 381, + 445, + 515, + 523 + ], + "lines": [ + { + "bbox": [ + 381, + 445, + 515, + 523 + ], + "spans": [ + { + "bbox": [ + 381, + 445, + 515, + 523 + ], + "type": "image", + "image_path": "66844f706f30d6da5ef784237c2b597e83eb7aefc0807140d7e7b615db517eab.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 526, + 484, + 536 + ], + "lines": [ + { + "bbox": [ + 416, + 526, + 484, + 536 + ], + "spans": [ + { + "bbox": [ + 416, + 526, + 484, + 536 + ], + "type": "text", + "content": "Rich Annotations" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 538, + 546, + 561 + ], + "lines": [ + { + "bbox": [ + 46, + 538, + 546, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 546, + 561 + ], + "type": "text", + "content": "Figure 1: SynBody is a large-scale synthetic dataset with a massive number of subjects and high-quality annotations. It supports various research topics, including human pose and shape estimation (HPS) and novel view synthesis for human (Human NeRF)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 571, + 192, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 571, + 192, + 583 + ], + "spans": [ + { + "bbox": [ + 143, + 571, + 192, + 583 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Synthetic data has emerged as a promising source for 3D human research as it offers low-cost access to large-scale human datasets. To advance the diversity and annotation quality of human models, we introduce a new synthetic dataset, SynBody, with three appealing features: 1) a clothed parametric human model that can generate a diverse range of subjects; 2) the layered human representation that naturally offers high-quality 3D annotations to support multiple tasks; 3) a scalable system for produc" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 547, + 693 + ], + "type": "text", + "content": "ing realistic data to facilitate real-world tasks. The dataset comprises 1.2M images with corresponding accurate 3D annotations, covering 10,000 human body models, 1,187 actions, and various viewpoints. The dataset includes two subsets for human pose and shape estimation as well as human neural rendering. Extensive experiments on SynBody indicate that it substantially enhances both SMPL and SMPL-X estimation. Furthermore, the incorporation of layered annotations offers a valuable training resource for investigating the Human Neural Radiance Fields(NeRF)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20282" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 91, + 286, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 91, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 286, + 174 + ], + "type": "text", + "content": "The fields of 3D human perception [21, 23-25, 37, 53] and human reconstruction [13, 14, 26, 40, 41] have become increasingly important, but the lack of available data has limited their development. Collecting real human data on a large scale is challenging due to privacy concerns and time constraints. Therefore, exploring the use of synthetic human datasets has become a critical avenue of research." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 175, + 286, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 286, + 365 + ], + "type": "text", + "content": "Despite the great potential, existing synthetic datasets [5, 8,38,47] suffer from limitations such as the number of available human models and the quality of annotations. The main reason lies in that synthetic human datasets rely on real scans for rendering, which poses three key obstacles. Firstly, it is challenging to expand the types of body shapes, poses, and clothing available in the dataset. Secondly, as the human models are scanned with clothing, the 3D annotations obtained through fitting are prone to errors. Thirdly, it is difficult to obtain annotations of body and clothing separately. To address these issues, we develop a new synthetic dataset termed SynBody. The dataset includes 1.2 million frames with corresponding ground-truth 3D human body annotations. It covers 10,000 human body models, 1,187 motions, and 26,960 video clips with 2.7M SMPL/SMPL-X annotations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 367, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 367, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 49, + 367, + 286, + 568 + ], + "type": "text", + "content": "At the heart of SynBody is the layered parametric human model, which constructs the clothed human model in a bottom-up manner. SMPL-X [39] is a widely used parametric human model, capable of sampling human models with various body shapes. However, it lacks the ability to model clothing, limiting its applicability when synthesizing realistic human models. To overcome this limitation, we introduce SMPL-XL, a parametric human model based on SMPL-X in a layered representation. SMPL-XL enriches the SMPL-X model in three aspects: (1) Hair system: adding hair and beards to the FLAME [27] model, with 32 types of hair and 13 types of beards; (2) Garment and accessories: adding procedural clothes to the SMPL-X body, including coats, shirts, pants, skirts, shoes, and glasses; (3) Texture: in addition to adding rich geometry, SMPL-XL also adds rich textures for sampling various skin colors and clothing textures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 570, + 286, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 570, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 49, + 570, + 286, + 700 + ], + "type": "text", + "content": "The designed SMPL-XL is capable of automatically generating a large number of human models with high-quality annotations. We therefore generate 10,000 clothed human models by sampling various body shapes, clothing styles, hairstyles, accessories, and textures. Notably, the use of the SMPL-X model as the base body model guarantees that the parametric human annotations are always accurate, obviating the need for the necessity for annotations through fitting. Furthermore, as the clothes are explicitly attached to the surface of the human body, layered annotations for body and clothes are available." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 701, + 286, + 713 + ], + "type": "text", + "content": "To generate a large-scale dataset with high diversity and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 72, + 544, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 179 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 179 + ], + "type": "text", + "content": "high-quality annotations, we design a scalable and automatic system to render images and annotations. We first animate the 10,000 dressed human models by retargeting motions from a large motion library [31]. Subsequently, we design an algorithm to place human models in the scene without piercing. Multiple cameras are then placed by evaluating self-occlusion, inter-occlusion, and view diversity, and the rendering module renders the assets into images with corresponding annotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 180, + 544, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 180, + 544, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 180, + 544, + 346 + ], + "type": "text", + "content": "With SynBody, we launch two tracks that support human pose and shape estimation and human neural rendering, respectively. Experiments show that SynBody is more effective than AGORA under the same amount of training data for human pose and shape estimation. With diverse and large-scale training data, SynBody achieves significant performance gains on both SMPL and SMPL-X estimation. In terms of human neural rendering using neural radiance fields (i.e. NeRF [34]), benchmarking existing approaches on SynBody shows that it has comparable performance as real human data. Furthermore, with the layered annotations which offer accurate SMPL parameters, we observe that current human NeRF approaches are sensitive to the accuracy of estimated SMPL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 348, + 544, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 348, + 544, + 478 + ], + "spans": [ + { + "bbox": [ + 308, + 348, + 544, + 478 + ], + "type": "text", + "content": "In summary, SynBody is a large-scale synthetic dataset for human perception and modeling, with three main contributions: (1) It constructs clothed subjects and samples 10,000 animatable subjects, which is an order of magnitude higher than existing datasets. (2) The clothed subjects are constructed with an explicit cloth model, thus it provides layered 3D annotations of the human body and clothing, which are not available in previous datasets. (3) Experiments on SynBody achieve promising results on both human perception and modeling, emphasizing the importance of diversity and annotation quality for downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 490, + 395, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 395, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 395, + 502 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 510, + 544, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 544, + 665 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 544, + 665 + ], + "type": "text", + "content": "Human Parametric Models. Several 3D human parametric models, such as SMPL [30], SMPL-X [39], and GHUM [51], have been developed to generate 3D human meshes from parameters that represent the human pose and shape using linear blend skinning. SMPL-X [39] extends SMPL [30] by combining FLAME [27] and MANO [43] for the head and hands, respectively, and is trained on a large number of real scans to provide a strong basis for shape variations. However, SMPL-X only produces naked body meshes, and we aim to enhance its realism by building a layered parametric model that includes hair, clothes, and accessories. The proposed model leverages the shape basis of SMPL-X while providing realistic dressed human meshes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 666, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 666, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 666, + 544, + 712 + ], + "type": "text", + "content": "Human Pose and Shape Estimation. Several methods [21, 23-25, 37, 49, 52-54] have been proposed to estimate 3D human pose and shape parameters. HMR [21] directly regresses these parameters in an end-to-end manner, while" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "20283" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 115, + 536, + 285 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": "Table 1: Comparisons of 3d human dataset. We compare SynBody with existing datasets. We divide datasets into three types: real (R), synthetic (S), and mixed (M). SynBody constructs 10,000 animatable subjects, which is an order of magnitude higher than any existing datasets and brings competitive scale and diversity. \"ITW\" stands for \"In-the-Wild\" in the table." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 115, + 536, + 285 + ], + "lines": [ + { + "bbox": [ + 59, + 115, + 536, + 285 + ], + "spans": [ + { + "bbox": [ + 59, + 115, + 536, + 285 + ], + "type": "table", + "html": "
DatasetTypeITWVideo#Views#SMPL#Seq#Subj.#MotionsGT format
HumanEva [45]R-✓4/7NA7463DJ
Human3.6M [15]R-✓4312K83911153DJ, SMPL
MPI-INF-3DHP [32]M✓✓1496K16883DJ
3DPW [48]R✓✓132K6018*SMPL
Panoptic Studio [20]R-✓480736K480~100*3DJ
EFT [19]R✓-1129KNAManyNASMPL
ZJU-MoCap [41]R-✓21180K999SMPL,mask
SURREAL [47]S✓✓16.5MNA1452KSMPL
AGORA [38]S✓-1173KNA>350NASMPL, SMPL-X, mask
HSPACE [5]S✓✓5-NA100×16100GHUM/L, mask
GTA-Human [8]S✓✓11.4M20K>60020KSMPL
BEDLAM [6]S✓✓-1M10K2712,311SMPL-X, mask
SynBodyS✓✓42.7M27K10,0001,187SMPL, SMPL-X, mask
", + "image_path": "758d1e68346d5545e2d8df43abcd27bb85831b975a04687e8f610493b5ae2f6e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 298, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 289, + 515 + ], + "type": "text", + "content": "SPIN uses an optimization step [25] to guide the learning process towards pseudo-3D labels. PARE [24] employs part attention to tackle occlusion, and VIBE [23] leverages temporal information in videos for SMPL estimation. Apart from building low-cost data collection solutions [7], synthetic datasets have become a promising alternative to efficiently scale up data. SURREAL [47] renders textured SMPL body models in real-image backgrounds but does not account for cloth geometry, resulting in unrealistic subjects. AGORA [38] renders real human scans in a virtual world and provides high-quality synthetic data for image-based approaches. HSPACE [5] places animated human models in various scenes to provide training data for video-based methods, and increases the variation of human shape via refitting. GTA-Human [8] captures videos and optimizes corresponding SMPL annotations from video games. However, the diversity of subjects in current datasets is limited by either body shapes or cloth types." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "Expressive Human Pose and Shape Estimation. As face and hand are also crucial for human perception, some efforts [35,44,55] have been made to whole-body human pose and shape estimation. ExPose [9] introduces three experts to predict parameters for body, hand, and face, and merge them in a copy-paste strategy. Hand4Whole [35] further improves the prediction of wrist and finger poses by leveraging selected hand joint features. Predicting both hands and faces makes the dataset much more difficult to obtain than just predicting the body. In order to increase the diversity of real data, AGORA [38] provides image-based SMPL-X annotations by fitting SMPL-X on the scanned human models. While a very recent work, BEDLAM [6], also enhances the SMPL-X model with clothing, hair, and changes in skin tones. However, it employs a manual approach for creating garments and utilizes pre-made skin textures, making" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 298, + 547, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 547, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 547, + 359 + ], + "type": "text", + "content": "it challenging to generate a large number of human body models. In contrast, our body model construction process is fully automated, encompassing clothing, hair, textures, and more. This enables easier scalability to generate human body models on a large scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 361, + 548, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 548, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 548, + 567 + ], + "type": "text", + "content": "Human NeRF. NeRF [34] has demonstrated impressive photo-realistic view synthesis by learning implicit fields of density and color. Yet, human motions are more challenging to learn due to dynamic deformation fields. NeuralBody [41] incorporates prior from a statistical body template to learn dynamic sequence, while Animatable NeRF [40] proposes to reconstruct an animatable human model that generalizes to new poses. Furthermore, NHP [26] and KeypointNeRF [33] achieve generalizability to unseen identities and poses. Several datasets have been adapted to study human NeRF. ZJU-MoCap [41] captures 9 human subjects with 21 synchronized cameras, providing fitted human body model parameters as well as the foreground mask. Human3.6M [16] collects 11 human subjects with 4 cameras, using a marker-based motion capture system. A-NeRF [46] generates a synthetic dataset using SURREAL [47] to study factors that affect visual quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": "While the aforementioned tasks are interrelated and often draw upon similar datasets, as shown in Table 1, datasets are limited in two key aspects. Firstly, obtaining real human models is challenging, which restricts the scale of these datasets. Secondly, 3D annotations are typically acquired through optimization, which introduces errors and cannot provide layered annotations. In contrast, with the designed SMPL-XL model, SynBody provides 10,000 subjects with diverse body shapes and clothing, along with layered annotations that include accurate SMPL and SMPL-X. Built on top of this layered human model, SynBody comprises two subsets for human mesh recovery and human NeRF, respec" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20284" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 261, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 261, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 261, + 85 + ], + "type": "text", + "content": "tively, and features the same 10,000 diverse subjects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 95, + 239, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 95, + 239, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 95, + 239, + 109 + ], + "type": "text", + "content": "3. Synthetic Data Generation System" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 47, + 125, + 286, + 205 + ], + "blocks": [ + { + "bbox": [ + 47, + 125, + 286, + 205 + ], + "lines": [ + { + "bbox": [ + 47, + 125, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 286, + 205 + ], + "type": "image", + "image_path": "fb00fd8f6e55eb1b726a486e2ce356fe79dcb390b82dd46f504310ab71d16968.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 288, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 288, + 273 + ], + "type": "text", + "content": "Figure 2: Synthetic data generation system. It consists of 4 components: 1) human model creation to generate layered human models, 2) motion retargeting to drive human models, 3) scene composition to place actors and cameras, and 4) rendering and outputs to generate multi-modal dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 400 + ], + "type": "text", + "content": "Similar to movie production [4], our system comprises 4 components (depicted in Figure 2): (1) a layered parametric human model creation service, a scalable process to generate layered human models, (2) a motion retargeting module to apply motions from various sources to layered human models, (3) a scene composition module to place 3D actors and objects into a 3D scene, setting up cameras, (4) a 3D rendering engine and a multi-modal data annotation generator. Our infrastructure enables the generation of high-quality synthetic data for various computer vision tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 407, + 235, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 235, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 235, + 419 + ], + "type": "text", + "content": "3.1. Layered Parametric Human Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 510 + ], + "type": "text", + "content": "Parametric human models like SMPL-X [39] provide the ability of create rigged body models with various body shapes. However, the lack of available textures limits its application in data generation. We designed a module to perform an automatic process that combines SMPL-X with procedural garments and accessories, hair system and textures, producing realistic and diverse body models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "text", + "content": "Body shape. SMPL-X [39] body model has a 3D mesh whose vertex locations are controlled by parameters for pose " + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "text", + "content": ", shape " + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "text", + "content": ", and facial expression " + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "text", + "content": ". By modifying shape " + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "text", + "content": ", we obtain 3D meshes of various human heights and weights. And by alternating pose " + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 510, + 287, + 581 + ], + "type": "text", + "content": ", meshes can be driven to perform various poses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "Garment Model. Our garment is generated as a separate layer on top of the body. Following the industrial garment-making workflow, we designed garment patterns of various styles. Notice that different parts of garment pieces are connected with sewing lines, e.g., the red line in Figure 4 (a). Then, we stitch patterns onto the body at T-Pose utilizing a physical simulator [1]. Specifically, we first manually move the garment pieces to roughly align them with the body. During simulation, vertices between two ends of the seam gradually shrink until they completely pooled together. Figure 4 (b) demonstrates the final draped gar" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "ment. For garment animation, we bind every vertex in the garment to the closest point in the body. Then, the skinnig weight and blend shape of the body mesh are assigned to garment vertices. This makes our garment model easily integrated with existing skeletal pipelines with little computational overhead." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "text", + "content": "Particle hair system. Our method uses a prefabricated particle system to generate realistic hair on a head-shaped mesh. To achieve this, a template of hairstyle or facial hair " + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "inline_equation", + "content": "T_{hair}" + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "text", + "content": " would attach to the vertices on the mesh which are marked with different areas " + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "inline_equation", + "content": "V_{hair}" + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "text", + "content": " (including fringe, top, temporal bone, occipital bone, and bottom area). Designers draw multiple sets of guidelines " + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "inline_equation", + "content": "L_{guide}" + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "text", + "content": " with varying shapes on different areas. Each set of guidelines comprises a collection of Bezier curves that are utilized to accurately constrain the flow of hair strands. Furthermore, the shape of the hair strands can be adjusted by length " + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "inline_equation", + "content": "P_{length}" + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "text", + "content": " and curliness " + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "inline_equation", + "content": "P_{curliness}" + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "text", + "content": ". The entire hair system is composed of multiple sub-particle systems " + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "inline_equation", + "content": "\\{V_{hair}, L_{guide}, P_{length}, P_{curliness}\\}" + }, + { + "bbox": [ + 304, + 144, + 546, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 300, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 420 + ], + "type": "text", + "content": "Accessories. We also add template accessories " + }, + { + "bbox": [ + 304, + 300, + 545, + 420 + ], + "type": "inline_equation", + "content": "T_{\\text{accessories}}" + }, + { + "bbox": [ + 304, + 300, + 545, + 420 + ], + "type": "text", + "content": " to our model, such as glasses, shoes, hats, and headphones. All accessories are pre-assembled on an SMPL-X template with a uniform body shape. These accessories can be transferred to models with the same topology, ensuring consistent deformation across all models in accordance with changes in the shape " + }, + { + "bbox": [ + 304, + 300, + 545, + 420 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 300, + 545, + 420 + ], + "type": "text", + "content": " of SMPL-X. The corresponding bone weights of the human body model's vertices are transferred to the nearest accessories' vertices, ensuring that these accessories are correctly driven by the armature." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "text", + "content": "Procedural texture. The procedural garment textures " + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "inline_equation", + "content": "T_{\\text{procedural}} = \\{T_{\\text{pattern}}, T_{\\text{decals}}, T_{\\text{bump}}, P_{\\text{mapping}}\\}" + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "text", + "content": " used in clothing are composed of multiple pre-set textures by alpha blending as demonstrated in Figure 4 (c). The pattern texture " + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "inline_equation", + "content": "T_{\\text{pattern}}" + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "text", + "content": " and decals texture " + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "inline_equation", + "content": "T_{\\text{decals}}" + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "text", + "content": " serve as layered masks to elevate the style and color, while the bump texture " + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "inline_equation", + "content": "T_{\\text{bump}}" + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "text", + "content": " functions as a height map, adding detailed normal information to the clothing's surface. The mapping parameters " + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "inline_equation", + "content": "P_{\\text{mapping}}" + }, + { + "bbox": [ + 304, + 421, + 546, + 588 + ], + "type": "text", + "content": " control the coordination of all the textures in the UV space. Besides, to procedurally create body mesh textures, we build a similar texture template SMPL-X, in which we pre-draw layered masks to separate different body features, such as skin, lips, and freckles, allowing for color adjustment and blending in specific regions as in Figure 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 545, + 661 + ], + "type": "text", + "content": "Combining all the elements above, as shown in Figure 5, we obtain human body models with the same high quality as RenderPeople [2]. Besides, it is challenging to expand the types of body shape and clothing available for real scans, but our model can be easily scaled by randomly sampling each component." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 670, + 421, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 421, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 421, + 683 + ], + "type": "text", + "content": "3.2. Motion Retargeting" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Retargeting of skeletal animations. Our motion retargeting module allows for the transfer of motion data from vari" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20285" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 269, + 168 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 269, + 168 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 269, + 168 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 269, + 168 + ], + "type": "image", + "image_path": "0b734aa6d69cd069dfd6dde50d42f857708d02b1bd3d32670ae5dff13f7e59d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 171, + 234, + 183 + ], + "lines": [ + { + "bbox": [ + 90, + 171, + 234, + 183 + ], + "spans": [ + { + "bbox": [ + 90, + 171, + 234, + 183 + ], + "type": "text", + "content": "(a) Naked human model generation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 335, + 69, + 539, + 165 + ], + "blocks": [ + { + "bbox": [ + 335, + 69, + 539, + 165 + ], + "lines": [ + { + "bbox": [ + 335, + 69, + 539, + 165 + ], + "spans": [ + { + "bbox": [ + 335, + 69, + 539, + 165 + ], + "type": "image", + "image_path": "e8bd40e463f3736985786ed243195e9f19de9a3c722317ffea4cbd0a7965b967.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 171, + 512, + 183 + ], + "lines": [ + { + "bbox": [ + 361, + 171, + 512, + 183 + ], + "spans": [ + { + "bbox": [ + 361, + 171, + 512, + 183 + ], + "type": "text", + "content": "(b) Layered human model generation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 240, + 282, + 387 + ], + "blocks": [ + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 547, + 222 + ], + "type": "text", + "content": "Figure 3: Layered parametric human model creation: (a) Generation of a naked human model using parametric body model [39], procedural body texture colors, and particle hair System; (b) Integration of garments and accessories onto the naked body model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 240, + 282, + 387 + ], + "lines": [ + { + "bbox": [ + 50, + 240, + 282, + 387 + ], + "spans": [ + { + "bbox": [ + 50, + 240, + 282, + 387 + ], + "type": "image", + "image_path": "ccf98bd231a782d1e61ecc3c19426e73222c0eeaf820d538e1b0e3bea8563a85.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 394, + 288, + 455 + ], + "lines": [ + { + "bbox": [ + 46, + 394, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 288, + 455 + ], + "type": "text", + "content": "Figure 4: Demonstrations of Garment and Procedural Texture. (a) garment pattern with sewing lines, (b) simulated garment under the body in canonical space, (c) procedural textured garment whose style and color are elevated by layered masks provided by patterns and decals." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 528 + ], + "type": "text", + "content": "ous sources, such as academic motion datasets, motion captures, and artist-crafted sources, to SMPL-XL model skeletons. Despite variations in bone names, bone lengths, and rest pose bone rotations, the source skeletons are typically structurally similar to the target skeletons." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 528, + 288, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 288, + 588 + ], + "type": "text", + "content": "Pose frames in motion clips contain root translation " + }, + { + "bbox": [ + 46, + 528, + 288, + 588 + ], + "type": "inline_equation", + "content": "T(t) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 528, + 288, + 588 + ], + "type": "text", + "content": " and rotations of each bone in the corresponding parent bone's space " + }, + { + "bbox": [ + 46, + 528, + 288, + 588 + ], + "type": "inline_equation", + "content": "\\{R_i(t) \\in \\mathbb{R}^3 | i = 0,1,\\dots,n\\}" + }, + { + "bbox": [ + 46, + 528, + 288, + 588 + ], + "type": "text", + "content": ". Following the forward kinematics (FK) manner, each bone's rotation relative to the model space can be calculated:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 592, + 287, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 592, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 116, + 592, + 287, + 608 + ], + "type": "interline_equation", + "content": "\\widehat {R} _ {i} (t) = \\widehat {R} _ {p (i)} (t) \\cdot R _ {i} (t) \\tag {1}", + "image_path": "481b69a48e0c8ac4759b4f2d19ace0544b4bab6cc8223752fed0452b10a266f0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\widehat{R}_i(t)" + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "text", + "content": " is the rotation of " + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "inline_equation", + "content": "R_{i}(t)" + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "text", + "content": " in model space at " + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "text", + "content": " frame, and " + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "inline_equation", + "content": "p(i)" + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "text", + "content": " indicates the parent of " + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "text", + "content": ". Specially, root bones have no parents so " + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\widehat{R}_0(t) = R_0(t)" + }, + { + "bbox": [ + 46, + 616, + 287, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "To retarget motion from one skeleton to another, we assume the source and the target motion can drive the corresponding bones of both skeletons to the same rotation in model space. We have T-posed frames of skeletons by manually posing them as T-posing, and treat T-posed as all mo" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 313, + 240, + 541, + 377 + ], + "blocks": [ + { + "bbox": [ + 313, + 240, + 541, + 377 + ], + "lines": [ + { + "bbox": [ + 313, + 240, + 541, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 541, + 377 + ], + "type": "image", + "image_path": "06e4ca65c4543ff09c958eb3f037498fd5e1b563189f7d448867a30383c07b32.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 384, + 546, + 456 + ], + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 456 + ], + "type": "text", + "content": "Figure 5: A Demonstration of comparison between commonly used human models in existing synthetic datasets [5, 8,38,47] and our SMPL-XL. We obtain high-quality models equal to RenderPeople [2], both are much better than SUR-REAL [47], and our model has the capability of scaling up easily by sampling various assets and body shapes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 469, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 493 + ], + "type": "text", + "content": "tions' first frame pose. So motions relative to their T-posedr frame can be easily obtained:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 349, + 498, + 545, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 498, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 349, + 498, + 545, + 514 + ], + "type": "interline_equation", + "content": "\\widehat {R} _ {i} (t) = \\widehat {R} _ {i} (0) \\cdot \\widehat {R} _ {i _ {- s r c}} (t) \\cdot \\widehat {R} _ {i _ {- s r c}} ^ {- 1} (0). \\qquad (2)", + "image_path": "ec1594e490fe7db7bd884161e4fc16b8f565e9aed7ccbfe31618b77f8d3e979e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 519, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 571 + ], + "type": "text", + "content": "Considering skeletons have different bone lengths, which could result in \"sliding feet\" artifacts on target skeletons. We simply scale " + }, + { + "bbox": [ + 304, + 519, + 545, + 571 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 519, + 545, + 571 + ], + "type": "text", + "content": " according to the ratio of pelvis bones' height to mitigate them: " + }, + { + "bbox": [ + 304, + 519, + 545, + 571 + ], + "type": "inline_equation", + "content": "T(t) = \\frac{H_{pelvis}}{H_{pelvis,src}} \\cdot T_{src}(t)" + }, + { + "bbox": [ + 304, + 519, + 545, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "text", + "content": "SMPL-X Annotations. Considering SMPL-XL body shapes are sampled from SMPL-X, the shape " + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "text", + "content": " can be derived directly from the corresponding SMPL-X. Pose " + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "text", + "content": " in the model space is calculated in the motion retargeting module. And in the scene composition module, models are placed in a 3D scene with camera space locations " + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "text", + "content": " and rotations " + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "inline_equation", + "content": "R_{c}" + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "text", + "content": ". Pose " + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\theta_{c}" + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "text", + "content": " in camera space is calculated by applying those world transformations to " + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "text", + "content": ". So SMPL-X annotations are constructed with " + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\{\\beta, \\theta_{c}\\}" + }, + { + "bbox": [ + 304, + 571, + 546, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "SMPL Annotations. Although SMPL-XL naturally provides accurate SMPL-X annotations, SMPL cannot be derived directly. Thus, we need to refit the SMPL parameters." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20286" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "The optimization process consists of two steps. First, we fit the shape " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " of all human models under the T-pose to its corresponding SMPL-X. Secondly, for each sequence, it is initialized with fitted " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " and its original pose. we fix " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " while fitting the body pose " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": " for SMPL." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 139, + 149, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 139, + 149, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 149, + 152 + ], + "type": "text", + "content": "3.3. Scene Placement" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "text", + "content": "To place " + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "text", + "content": " subjects in a large scene with " + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "inline_equation", + "content": "N_{o}" + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "text", + "content": " objects, we primarily follow three principles: standing on the ground, avoiding human-object and human-human penetration. To prevent subjects from floating in the air, the root position of a subject should align with the ground height. A sequential decision-making approach is used to find a suitable position for each subject, i.e., placing one subject at a time. An object is represented by " + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "inline_equation", + "content": "o_{i} = \\{x_{i},y_{i},l_{i},w_{i}\\} \\in \\mathbb{R}^{4}" + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\{x_i,y_i\\}" + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "text", + "content": " is the center of the object's axis-aligned bounding box projected onto the ground with length " + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "text", + "content": ". Different from static objects, to avoid collisions between moving subjects at any one time, a subject with specified body shape and motion is simplified to " + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "inline_equation", + "content": "q_{i} = \\{x_{i},y_{i},l_{i},w_{i}\\} \\in \\mathbb{R}^{4}" + }, + { + "bbox": [ + 46, + 158, + 287, + 337 + ], + "type": "text", + "content": ", which is the smallest axis-aligned box that envelops all bounding boxes across frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 373 + ], + "type": "text", + "content": "To avoid human-object penetration and potential human-human collision, the solution " + }, + { + "bbox": [ + 46, + 338, + 287, + 373 + ], + "type": "inline_equation", + "content": "p_i^* = \\{x_i^*, y_i^*\\}" + }, + { + "bbox": [ + 46, + 338, + 287, + 373 + ], + "type": "text", + "content": " of a character with the shape of " + }, + { + "bbox": [ + 46, + 338, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\{l_i, w_i\\}" + }, + { + "bbox": [ + 46, + 338, + 287, + 373 + ], + "type": "text", + "content": " should satisfy:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 381, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 381, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 82, + 381, + 287, + 396 + ], + "type": "interline_equation", + "content": "I \\left(\\left\\{x _ {i} ^ {*}, y _ {i} ^ {*}, l _ {i}, w _ {i} \\right\\}, o _ {j}\\right) = 0, j = 1, \\dots , N _ {o} \\tag {3}", + "image_path": "106b721678b5050a6b6cbb690be2b8ddf90f17a94401fdbea76652297e97e43d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 415, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 415, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 80, + 415, + 287, + 430 + ], + "type": "interline_equation", + "content": "I \\left(\\left\\{x _ {i} ^ {*}, y _ {i} ^ {*}, l _ {i}, w _ {i} \\right\\}, q _ {k}\\right) = 0, k = 1, \\dots , N _ {p}, \\tag {4}", + "image_path": "5b821208348ac1adfbe2fe1949c1f6609fb3c93fff276696551ad7b608d43f9b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 434, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 434, + 287, + 506 + ], + "type": "inline_equation", + "content": "I(\\text{box}_1, \\text{box}_2)" + }, + { + "bbox": [ + 46, + 434, + 287, + 506 + ], + "type": "text", + "content": " denotes the overlapping area between two boxes and " + }, + { + "bbox": [ + 46, + 434, + 287, + 506 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 46, + 434, + 287, + 506 + ], + "type": "text", + "content": " is the number of subjects already placed in the scene. Besides, distance constraint is used to prevent subjects from excessive dispersal. The problem is solved by grid search. Typically multiple solutions are available and we randomly sample one each time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 512, + 160, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 160, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 160, + 525 + ], + "type": "text", + "content": "3.4. Camera Placement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 531, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 626 + ], + "type": "text", + "content": "Once the positions of all subjects have been organized, the next task involves placing " + }, + { + "bbox": [ + 46, + 531, + 287, + 626 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 46, + 531, + 287, + 626 + ], + "type": "text", + "content": " cameras in suitable locations. To ensure that cameras would not be placed inside any objects or subjects, the 3D version of Eq. (3) and Eq. (4) are applied. In addition, we evaluate the suitability of a candidate camera by the following metrics: the distance from the camera to the subjects, the camera's pitch angle, and the degree of occlusion of each subject in the camera's view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": "The distance denoted as " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " from the mean position of all subjects " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\bar{p}" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " to the camera is restricted. " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "L_{max}" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " is set to prevent an unreasonably small proportion of subjects in the image. To control the visibility of subjects, " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "L_{min}" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\lambda}{\\sin(\\alpha / 2)}\\max_i||pv_i - \\bar{p} ||_2" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,N_c" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " is the field of view of the camera, and " + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 626, + 288, + 714 + ], + "type": "text", + "content": " is a hyperparameter that determines the probability of all subjects being within" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "the camera's view. To estimate the degree of occlusion for each subject, rays are randomly and uniformly cast from the camera to each subject's body, and the percentage of blocked rays by other objects is determined. More details can be found in the Sup. Mat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 141, + 459, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 141, + 459, + 154 + ], + "spans": [ + { + "bbox": [ + 306, + 141, + 459, + 154 + ], + "type": "text", + "content": "3.5. Rendering and Annotations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 159, + 547, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 547, + 256 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 547, + 256 + ], + "type": "text", + "content": "Using a high-quality rendering pipeline in Unreal Engine 5, SynBody is rendered in multi-view with large 3D environments from the Unreal Marketplace for rich background information and dynamic lighting. Leveraging the G-buffer [12], our system simultaneously generates photorealistic RGB images and annotations, along with accurate ground truth for segmentation masks, optical flow, depth maps, normal maps, and other ground-truth labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 263, + 409, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 263, + 409, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 263, + 409, + 277 + ], + "type": "text", + "content": "4. SynBody Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 284, + 409, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 409, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 409, + 295 + ], + "type": "text", + "content": "4.1. Dataset Statistics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "text", + "content": "SynBody comprises 10,000 unique SMPL-XL models randomly created with different body shapes and genders. Each model is then combined with the following assets: (1) hairstyles " + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "inline_equation", + "content": "T_{hair}" + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "text", + "content": " sampled from 45 particle hairs; (2) garments " + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "inline_equation", + "content": "G_{tmp}" + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "text", + "content": " sampled from 68 clothing models containing multiple outfits; (3) procedural texture " + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "inline_equation", + "content": "T_{procedural}" + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "text", + "content": " generated by sampling " + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "inline_equation", + "content": "T_{pattern}" + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "inline_equation", + "content": "T_{decals}" + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "inline_equation", + "content": "T_{bump}" + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "text", + "content": " from 1,038 template textures, and color values are randomly sampled; (4) accessories " + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "inline_equation", + "content": "T_{accessories}" + }, + { + "bbox": [ + 304, + 302, + 547, + 626 + ], + "type": "text", + "content": " sampled from 46 template assets. To ensure the validity of the motions, we select a subset from AMASS [31]. Based on the BABEL annotations [42], we excluded interactive motions and non-ground motions(e.g., swim), as well as filtered out motions with a duration shorter than 2 seconds, leading to a subset with 1,187 motions. For each sequence, we randomly selected 4 human models, and each model is assigned a randomly chosen motion lasting between 2 and 10 seconds. For motions exceeding 2 seconds, we randomly extract a 2-second segment to ensure that each video clip has a length of 60 frames. To enhance the diversity of viewpoints, 4 view positions are generated for each sequence, where each one is randomly sampled from the surface of semi-spheres with varying radii. The system generates 26,960 sequences and 1.2M images with annotations, utilizing 6 vast and realistic scenes created by professional artists. 2.7M SMPL/SMPL-X annotations are provided, excluding highly occluded subjects. More details can be found in the Sup. Mat." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 494, + 647 + ], + "type": "text", + "content": "4.2. Human Pose and Shape Estimation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": "With 2.7 million SMPLs, we leverage a pretrained regressor to extract 3D keypoints and then project them onto 2D space. Following standard practice in top-down human mesh recovery methods, we generate bounding boxes from the resulting 2D keypoints." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20287" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 111, + 286, + 229 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 107 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 107 + ], + "type": "text", + "content": "Table 2: Training popular baseline methods (image-based and video-based) with SynBody on 3DPW test set [48]. R: standard real datasets. S: SynBody." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 111, + 286, + 229 + ], + "lines": [ + { + "bbox": [ + 50, + 111, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 50, + 111, + 286, + 229 + ], + "type": "table", + "html": "
MethodDatasetsMPJPEPA-MPJPEPVE
HMR [21]R112.3067.50141.92
HMRR + S95.0157.62116.10
SPIN [25]R96.9059.20119.70
SPINR + S84.1453.67103.79
PARE [24]R81.7949.36105.27
PARER + S78.9848.46103.86
VIBE [23]R94.8857.08108.59
VIBER + S93.0457.00107.23
", + "image_path": "6701fe21f33d247dbaffe8211ed43442e0f78f54bb9c5b48c76b786b4369e12e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 241, + 194, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 194, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 194, + 255 + ], + "type": "text", + "content": "4.3. Human Neural Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 260, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 287, + 416 + ], + "type": "text", + "content": "Given the flexibility of our pipeline, we render a total of 100 multi-view sequences with diverse motions and appearances for benchmarking human NeRFs. All sequences have a length of 300 frames with a resolution of " + }, + { + "bbox": [ + 46, + 260, + 287, + 416 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 46, + 260, + 287, + 416 + ], + "type": "text", + "content": ", where the motion sequence is randomly sampled from AMASS [31]. Each sequence contains 8 views whose camera positions have uniformly distributed azimuth angles around the human body. We use RGB renderings, binary foreground masks, camera parameters, and SMPL parameters for the benchmark. Thanks to our layered design, we can offer accurate SMPL parameters for human NeRFs, which acts as an important prior for a majority of methods, while keeping the diversity in clothes and motions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 427, + 124, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 124, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 124, + 441 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 483 + ], + "type": "text", + "content": "In this section, we study the usefulness of SynBody for two popular research directions: human pose and shape estimation and human NeRF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 491, + 235, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 235, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 235, + 504 + ], + "type": "text", + "content": "5.1. Human Pose and Shape Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 569 + ], + "type": "text", + "content": "Estimating 3D humans represented by SMPL (body-only) and SMPL-X (body, hands, and face) parameters from monocular 2D input has gained substantial attention. SynBody emerges as a scalable and effective solution to tackle the scarcity of paired data in these fields." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": "Parametric Human Models. SMPL [30] represents high-dimension human mesh " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "M(\\theta, \\beta) \\in \\mathbb{R}^{6890 \\times 3}" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": " as low-dimension pose parameters " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{72}" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": " and shape parameters " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": ". In addition to SMPL, SMPL-X [39] further consists of additional left and right hand pose parameters " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "(\\theta_{lh}, \\theta_{rh} \\in \\mathbb{R}^{15 \\times 3})" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": ", jaw joint rotation " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "(\\theta_{lh} \\in \\mathbb{R}^3)" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": " and face expression parameters " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "(\\phi_f \\in \\mathbb{R}^{10})" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": " for an expressive human representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "**Benchmarks.** We conduct experiments on two mainstream benchmarks: 3DPW [48] and AGORA [38]. As a widely used benchmark for in-the-wild evaluation, 3DPW encompasses diverse data collected through a mobile phone cam" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 122, + 537, + 212 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 119 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 119 + ], + "type": "text", + "content": "Table 3: Training popular baseline methods (image-based only as AGORA does not provide videos but static images only) with SynBody on AGORA validation set [38]. R: standard real datasets. S: SynBody." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 122, + 537, + 212 + ], + "lines": [ + { + "bbox": [ + 316, + 122, + 537, + 212 + ], + "spans": [ + { + "bbox": [ + 316, + 122, + 537, + 212 + ], + "type": "table", + "html": "
MethodDatasetsMPJPEPA-MPJPEPVE
HMR [21]R226.7187.72248.35
HMRR + S199.5177.97210.37
SPIN [25]R212.9179.76217.88
SPINR + S196.8176.06205.83
PARE [24]R178.1567.13189.73
PARER + S169.9364.37179.81
", + "image_path": "3bb97336d600857919730d2b23a6e28a71b025bdc44652cd6a2b629481ce4872.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 242, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 545, + 289 + ], + "type": "text", + "content": "era, paired with SMPL annotations. As for AGORA, it is a recent synthetic dataset that features challenging scenes with person-person occlusion. AGORA provides both SMPL and SMPL-X annotations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 290, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 546, + 411 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the quality of predicted parametric human models, we employ the standard metrics: MPJPE (Mean Per Joint Position Error), which is calculated as the average " + }, + { + "bbox": [ + 304, + 290, + 546, + 411 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 290, + 546, + 411 + ], + "type": "text", + "content": " distance over 3D keypoints regressed from parametric models; PA-MPJPE (Procrustes Aligned Mean Per Joint Position Error), which is MPJPE with applying Procrustes Alignment [11] on the predicted keypoints to match the ground truth before error computation; PVE (Per Vertex Error), which is the average " + }, + { + "bbox": [ + 304, + 290, + 546, + 411 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 290, + 546, + 411 + ], + "type": "text", + "content": " distance between predicted and ground truth mesh vertices." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 429, + 416, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 429, + 416, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 416, + 441 + ], + "type": "text", + "content": "5.1.1 SMPL Estimation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 449, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 509 + ], + "type": "text", + "content": "Methods. To gauge the usefulness of SynBody, we conduct experiments with four milestone works: HMR [21], SPIN [25], and PARE [24] are image-based methods and VIBE [23] is a video-based method. We follow the hyperparameter configurations provided in MMHuman3D [10]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 593 + ], + "type": "text", + "content": "Training Data. In the following experiments, we refer to the standard baskets of real datasets for SMPL estimation as \"R\". For HMR and SPIN, R consists of H36M [16], MPI-INF-3DHP [32], LSP [17], LSPET [18], MPII [3] and MSCOCO [29], whereas in VIBE, R consists of MPI-INF-3DHP and InstaVariety [22]. \"S\" denotes SynBody data with 2.7M training instances with SMPL annotations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Main Results. We collate experiment results on 3DPW in Table 2. Each baseline is fine-tuned with a mixture of real datasets (R) and SynBody (S), leading to significant performance gains across different baseline models, even on strong baseline PARE [24] (2.81 mm and 1.41 mm improvements on MPJPE and PVE). Notably, SynBody can be used to train video-based methods such as VIBE as it contains video sequences instead of static images. Table 3 illustrates that training with SynBody exhibits remarkable improvements across methods, when compared with" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20288" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 286, + 171 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 286, + 171 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 286, + 171 + ], + "type": "image", + "image_path": "9088f1e0e8c6a5eba3e196a484217c549eb830f322c7c1678c1d24b59ade6080.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 287, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 222 + ], + "type": "text", + "content": "Figure 6: Impact of the amount of SynBody data. The horizontal axis represents the amount of SynBody data that is varied by multiples of real data. The baseline model is HMR and tested on the 3DPW test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 301, + 282, + 395 + ], + "blocks": [ + { + "bbox": [ + 46, + 238, + 286, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 286, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 286, + 297 + ], + "type": "text", + "content": "Table 4: Comparison between SynBody-100K and AGORA on 3DPW test set. “A” means AGORA, and “S100K” means SynBody-100K which is a subset of SynBody datasets with over 100K SMPL annotations. For a fair comparison, the total number of SMPL annotations is close to that of AGORA." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 301, + 282, + 395 + ], + "lines": [ + { + "bbox": [ + 52, + 301, + 282, + 395 + ], + "spans": [ + { + "bbox": [ + 52, + 301, + 282, + 395 + ], + "type": "table", + "html": "
MethodDatasetsMPJPEPA-MPJPEPVE
HMRR + A101.6157.85123.82
HMRR + S100K95.2857.68119.18
SPINR + A88.4454.97110.35
SPINR + S100K85.5254.12105.45
PARER + A85.3448.39109.77
PARER + S100K79.4247.80102.45
", + "image_path": "d4bfe06201984c4bff4bf562dac3fd0b4361414141ffcd1ff6cf2ebde64c7706.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 469, + 276, + 526 + ], + "blocks": [ + { + "bbox": [ + 46, + 417, + 286, + 464 + ], + "lines": [ + { + "bbox": [ + 46, + 417, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 286, + 464 + ], + "type": "text", + "content": "Table 5: Effectiveness of SMPL-XL. R: real datasets (MSCOCO, MPI-INF-3DHP, Human3.6M, LSP, LSPET, MPII). S100K: downsampled SynBody with 100K instances.*: SURREAL-style." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 469, + 276, + 526 + ], + "lines": [ + { + "bbox": [ + 57, + 469, + 276, + 526 + ], + "spans": [ + { + "bbox": [ + 57, + 469, + 276, + 526 + ], + "type": "table", + "html": "
MethodDatasetsMPJPEPA-MPJPEPVE
PARER81.849.4105.3
PARER+S100K*82.249.7105.6
PARER+S100K79.447.8102.5
", + "image_path": "1096d790206f3be848a2034fe808f0ddec82521a305b37864ac84a2af93d7f00.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 555, + 234, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 234, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 234, + 567 + ], + "type": "text", + "content": "the baseline models in AGORA validation set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 568, + 286, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 286, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 286, + 663 + ], + "type": "text", + "content": "Impact of Data Scale. As our generation system can easily scale up the data size, we train HMR from scratch to study the influence of synthetic data scale. Figure 6 demonstrates that adding more SynBody data generally leads to better performance. These experiments confirm that synthetic data is a valuable complement to real data and serves as a readily scalable training source to supplement the typically limited real data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Comparison with AGORA. To evaluate the quality of SynBody, we randomly sample a subset with 100K training instances, \"SynBody-100K\", which is comparable in size with the popular synthetic dataset AGORA. In Table 4, we" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 70, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 155 + ], + "type": "text", + "content": "Table 6: SMPL-X estimation with OSX [28] as the baseline on AGORA validation set (AGORA-val) [38] and 3DPW test set [28]. R: real datasets (MSCOCO, MPII, and Human3.6M). S: SynBody. AGORA-val uses PVE (mm) whereas 3DPW uses MPJPE (mm) and PA-MPJPE (mm). Note the original OSX uses the SMPL head for 3DPW, which we modify to the SMPL-X head." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 159, + 544, + 212 + ], + "blocks": [ + { + "bbox": [ + 308, + 159, + 544, + 212 + ], + "lines": [ + { + "bbox": [ + 308, + 159, + 544, + 212 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 544, + 212 + ], + "type": "table", + "html": "
MethodDatasetsAGORA3DPW (Body)
AllHandsFaceMPJPEPA-MPJPE
OSXR168.670.677.2110.263.5
OSXR+S600K155.863.572.092.259.3
", + "image_path": "4eb766633801c4c87a05929194f14b9c0f2212697229e0e9c60b67de5a5fc51b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 242, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 545, + 289 + ], + "type": "text", + "content": "show that finetuning with SynBody outperforms that with AGORA across different baseline methods. The results highlight that SynBody demonstrates competitive traits as a training source." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 290, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 398 + ], + "type": "text", + "content": "Effectiveness of SMPL-XL. SMPL-XL enables layered human modeling which cannot be achieved with SURREAL/RenderPeople as Figure 5. We investigate the importance to synthesize data with these realistic considerations of SMPL-XL in Table 5. We compare SynBody with SURREAL-style data that renders body and cloth texture on human mesh surfaces, without actual cloth geometry or hair. It is observed that SURREAL-style data leads to significant performance degradation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 417, + 425, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 425, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 425, + 428 + ], + "type": "text", + "content": "5.1.2 SMPL-X Estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 472 + ], + "type": "text", + "content": "Also known as expressive human pose and shape estimation, SMPL-X estimation requires recovery of body, hands, and face parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": "Method and Training Data. In Table 6, we conduct experiments with the recent SoTA, OSX [28], as the base model. We compare the original baseline with training with a mixture of real datasets (\"R\" denotes COCO [29], MPII [3] and Human3.6M [16] with pseudo ground truth generated by NeuralAnnot [36]) and SynBody (\"S600K\" here denotes a downsampled set of 600K instances as the real datasets are smaller for SMPL-X than that for SMPL estimation). Note that in the OSX paper, the values reported on 3DPW are obtained using a SMPL head. Here we standardize the output using the same SMPL-X head for both AGORA and 3DPW." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Main Results. We observe that adding SynBody in the training of OSX leads to a significant improvement in both overall estimation (more than " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ") and part-level (more than " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "5\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " for hands and face) estimation on AGORA. Moreover, in a fair comparison with the same model architecture, training with SynBody leads to " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "18\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " improvements in MPJPE on the 3DPW test set (here we follow the standard protocol to evaluate 14 key joints as 3DPW does not provide SMPL-X annotations). We speculate that it" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20289" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 86, + 524, + 178 + ], + "blocks": [ + { + "bbox": [ + 113, + 71, + 478, + 83 + ], + "lines": [ + { + "bbox": [ + 113, + 71, + 478, + 83 + ], + "spans": [ + { + "bbox": [ + 113, + 71, + 478, + 83 + ], + "type": "text", + "content": "Table 7: Benchmark of NeRF-based methods for 3D human neural rendering on SynBody." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 86, + 524, + 178 + ], + "lines": [ + { + "bbox": [ + 71, + 86, + 524, + 178 + ], + "spans": [ + { + "bbox": [ + 71, + 86, + 524, + 178 + ], + "type": "table", + "html": "
MethodNovel ViewNovel PoseNovel Identity
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [34]19.390.8620.16219.610.8240.201---
NeuralBody [41]28.940.9660.05725.020.9440.080---
HumanNeRF [50]28.320.9630.06621.970.8790.108---
AnimNeRF [40]27.490.9640.05626.210.9500.068---
NHP [26]25.660.9530.07624.180.9450.08022.460.9270.103
", + "image_path": "0b3a4abd395918a1cfbf5dfcf5c291e8d18950acd96dc8774206a232f1838344.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 269 + ], + "type": "text", + "content": "is difficult to obtain accurate SMPL-X annotations, which is mitigated by SynBody as it provides large-scale high-quality SMPL-X labels paired with images that are rendered with diverse backgrounds and lighting conditions, which are beneficial to training a high-performing expressive human parametric model recovery." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 275, + 135, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 135, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 135, + 286 + ], + "type": "text", + "content": "5.2. Human NeRF" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 294, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 460 + ], + "type": "text", + "content": "In this section, we benchmark popular NeRF-based methods for 3D humans on SynBody, validating the effectiveness and great potential of our dataset in human neural rendering. Our benchmark is built upon three perspectives according to the purpose of synthesis tasks, which can be categorized into novel view, novel pose, and novel identity. Methods for Benchmark. We benchmark five methods in total, including the vanilla NeRF [34], NeuralBody [41] and HumanNeRF [50] for novel view synthesis, AnimNeRF [40] for novel pose synthesis, NHP [26] for generalizable human NeRF (novel identity synthesis). Except for NHP, all methods are trained in a person-specific manner, taking 4 views of the first 250 frames for training and the rest views and frames for evaluations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 461, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 533 + ], + "type": "text", + "content": "Evaluation Protocols. We follow [34, 50] to evaluate all methods using three standard metrics: Peak Signal-to-Noise Ratio (PSNR), Structural SIMilarity index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS [56]). To reach a consensus among all methods, all metrics are computed over the whole image with a black background." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 664 + ], + "type": "text", + "content": "Main Results. Benchmark results are reported in Table 7. We observe that all methods achieve comparable performances as real human data on SynBody. Models (NeuralBody and AnimNeRF) which rely on accurate SMPL estimation and blending weights perform better on novel poses. We attribute it to our layer-wise design which offers ground truth SMPL parameters for human NeRF training. Besides, we present visualization results in Figure 7, observing that the diverse motions and appearances, as well as loose garments in SynBody, pose further challenges in the field of neural rendering of 3D humans." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We present SynBody, a large-scale synthetic dataset that features a substantial number of subjects and high-quality" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 197, + 381, + 271 + ], + "blocks": [ + { + "bbox": [ + 307, + 197, + 381, + 271 + ], + "lines": [ + { + "bbox": [ + 307, + 197, + 381, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 381, + 271 + ], + "type": "image", + "image_path": "3abc8211dd7460ae755962e39207c11572e83457af81e3b75ed119812d13b301.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 272, + 366, + 280 + ], + "lines": [ + { + "bbox": [ + 321, + 272, + 366, + 280 + ], + "spans": [ + { + "bbox": [ + 321, + 272, + 366, + 280 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 390, + 198, + 463, + 271 + ], + "blocks": [ + { + "bbox": [ + 390, + 198, + 463, + 271 + ], + "lines": [ + { + "bbox": [ + 390, + 198, + 463, + 271 + ], + "spans": [ + { + "bbox": [ + 390, + 198, + 463, + 271 + ], + "type": "image", + "image_path": "4731a51771d0956d16f12af922897f06670a5aae89df6c7d02f4259edbaef5af.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 272, + 435, + 280 + ], + "lines": [ + { + "bbox": [ + 416, + 272, + 435, + 280 + ], + "spans": [ + { + "bbox": [ + 416, + 272, + 435, + 280 + ], + "type": "text", + "content": "NeRF" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 471, + 198, + 545, + 271 + ], + "blocks": [ + { + "bbox": [ + 471, + 198, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 471, + 198, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 471, + 198, + 545, + 271 + ], + "type": "image", + "image_path": "626a7baf5cef8df6ed5b464cd3c69d56205fce8bfd44ad0e69773fe1e224c323.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 281, + 381, + 356 + ], + "blocks": [ + { + "bbox": [ + 307, + 281, + 381, + 356 + ], + "lines": [ + { + "bbox": [ + 307, + 281, + 381, + 356 + ], + "spans": [ + { + "bbox": [ + 307, + 281, + 381, + 356 + ], + "type": "image", + "image_path": "7818d9e5f61954e8601842e765120f1474f2fa929407d320c7dc613ffab1151a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 357, + 365, + 365 + ], + "lines": [ + { + "bbox": [ + 322, + 357, + 365, + 365 + ], + "spans": [ + { + "bbox": [ + 322, + 357, + 365, + 365 + ], + "type": "text", + "content": "HumanNeRF" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 390, + 281, + 463, + 356 + ], + "blocks": [ + { + "bbox": [ + 390, + 281, + 463, + 356 + ], + "lines": [ + { + "bbox": [ + 390, + 281, + 463, + 356 + ], + "spans": [ + { + "bbox": [ + 390, + 281, + 463, + 356 + ], + "type": "image", + "image_path": "60c2fe7cd237e9bed2005b7578ed82fe43c385707e65a540a3deb4cfd748f000.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 357, + 444, + 365 + ], + "lines": [ + { + "bbox": [ + 407, + 357, + 444, + 365 + ], + "spans": [ + { + "bbox": [ + 407, + 357, + 444, + 365 + ], + "type": "text", + "content": "AnimNeRF" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 471, + 281, + 545, + 356 + ], + "blocks": [ + { + "bbox": [ + 488, + 272, + 528, + 281 + ], + "lines": [ + { + "bbox": [ + 488, + 272, + 528, + 281 + ], + "spans": [ + { + "bbox": [ + 488, + 272, + 528, + 281 + ], + "type": "text", + "content": "NeuralBody" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 281, + 545, + 356 + ], + "lines": [ + { + "bbox": [ + 471, + 281, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 471, + 281, + 545, + 356 + ], + "type": "image", + "image_path": "295aa1aa17272125d0346c224615a1ce63b47877b7f4fe1b22201fa5e45453fa.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 357, + 517, + 365 + ], + "lines": [ + { + "bbox": [ + 500, + 357, + 517, + 365 + ], + "spans": [ + { + "bbox": [ + 500, + 357, + 517, + 365 + ], + "type": "text", + "content": "NHP" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 372, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 306, + 372, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 545, + 396 + ], + "type": "text", + "content": "Figure 7: Novel view synthesis of different human NeRF methods on SynBody." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 418, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 526 + ], + "type": "text", + "content": "3D annotations. At the core is a clothed human model with multiple layers of representation. Our experiments demonstrate the effectiveness of SynBody on both human mesh recovery and human NeRF. Future research can leverage SynBody for developing and evaluating methods to predict body and cloth simultaneously. Furthermore, the high controllability of the synthetic dataset offers ample opportunities for further improvements, such as the incorporation of contact labels for human-scene interaction." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 530, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 613 + ], + "type": "text", + "content": "Societal Impacts. Even though SynBody is a synthetic dataset, the assets utilized in its creation might not be well-balanced. Although hairstyles and skin colors are chosen at random, they are designed to avoid racial biases. Yet, other elements, such as body shapes and clothing, might not be as balanced, posing a potential for bias in the resultant human models." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This study is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2021-08-019), the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20290" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Marvelous designer, 2023. https://www.marvelousdesigner.com.4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 134 + ], + "type": "text", + "content": "[2] Render people, 2023. https://renderpeople.com. 4,5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 136, + 286, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 286, + 190 + ], + "type": "text", + "content": "[3] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 286, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 286, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 286, + 235 + ], + "type": "text", + "content": "[4] Rao Anyi, Jiang Xuekun, Guo Yuwei, Xu Linning, Yang Lei, Jin Libiao, Lin Dahua, and Dai Bo. Dynamic storyboard generation in an engine-based virtual environment for video production. arXiv preprint arXiv:2301.12688, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 290 + ], + "type": "text", + "content": "[5] Eduard Gabriel Bazavan, Andrei Zanfir, Mihai Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Hspace: Synthetic parametric humans animated in complex environments. arXiv preprint arXiv:2112.12867, 2021. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 286, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 286, + 346 + ], + "type": "text", + "content": "[6] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 286, + 434 + ], + "type": "text", + "content": "[7] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, Fangzhou Hong, Mingyuan Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. HuMMan: Multi-modal 4d human dataset for versatile sensing and modeling. In 17th European Conference on Computer Vision, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 557-577. Springer, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 286, + 479 + ], + "type": "text", + "content": "[8] Zhongang Cai, Mingyuan Zhang, Jiawei Ren, Chen Wei, Daxuan Ren, Zhengyu Lin, Haiyu Zhao, Lei Yang, and Zwei Liu. Playing for 3d human recovery. arXiv preprint arXiv:2110.07588, 2021. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 286, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 286, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 286, + 545 + ], + "type": "text", + "content": "[9] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 20-40. Springer, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 286, + 578 + ], + "type": "text", + "content": "[10] MMHuman3D Contributors. Openmmlab 3d human parametric model toolbox and benchmark. https://github.com/open-mmlab/mmhuman3d, 2021.7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 286, + 601 + ], + "type": "text", + "content": "[11] J. C. Gower. Generalized procrustes analysis. Psychometrika, 1975. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 624 + ], + "type": "text", + "content": "[12] Shawn Hargreaves and Mark Harris. Deferred shading. In Game Developers Conference, volume 2, page 31, 2004. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 286, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 666 + ], + "type": "text", + "content": "[13] Fangzhou Hong, Zhaoxi Chen, Yushi Lan, Liang Pan, and Ziwei Liu. Eva3d: Compositional 3d human generation from 2d image collections. arXiv preprint arXiv:2210.04888, 2022.2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[14] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[15] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[16] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014. 3, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 226 + ], + "type": "text", + "content": "[17] Sam Johnson and Mark Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC, pages 1-11. British Machine Vision Association, 2010. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "text", + "content": "[18] Sam Johnson and M. Everingham. Learning effective human pose estimation from inaccurate annotation. CVPR 2011, pages 1465-1472, 2011. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "text", + "content": "[19] H. Joo, N. Neverova, and A. Vedaldi. Exemplar fine-tuning for 3d human pose fitting towards in-the-wild 3d human pose estimation. ArXiv, abs/2004.03686, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 358 + ], + "type": "text", + "content": "[20] H. Joo, Tomas Simon, Xulong Li, H. Liu, L. Tan, Lin Gui, Sean Banerjee, Timothy Godisart, Bart C. Nabbe, I. Matthews, T. Kanade, S. Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:190-204, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 392 + ], + "type": "text", + "content": "[21] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, pages 7122-7131, 2018. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "text", + "content": "[22] Angjoo Kanazawa, Jason Y. Zhang, Panna Felsen, and Jitendra Malik. Learning 3d human dynamics from video. In Computer Vision and Pattern Recognition (CVPR), 2019. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 459 + ], + "type": "text", + "content": "[23] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, pages 5253-5263, 2020. 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[24] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. arXiv preprint arXiv:2104.08527, 2021. 2, 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 547 + ], + "type": "text", + "content": "[25] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[26] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34, 2021. 2, 3, 9" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "text", + "content": "[27] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "type": "text", + "content": "[28] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. 2023. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 7, 8" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20291" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[30] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "type": "text", + "content": "[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 288, + 220 + ], + "type": "text", + "content": "[32] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In international conference on 3D vision (3DV), pages 506-516. IEEE, 2017. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 222, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 287, + 266 + ], + "type": "text", + "content": "[33] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In ECCV, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 314, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 346 + ], + "type": "text", + "content": "[35] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In CVPR, pages 2308-2317, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[36] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralannot: Neural annotator for 3d human mesh training sets. In Proceedings of the Conference on Computer Vision and Pattern Recognition, pages 2299-2307, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 396, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 287, + 449 + ], + "type": "text", + "content": "[37] Hui En Pang, Zhongang Cai, Lei Yang, Tianwei Zhang, and Ziwei Liu. Benchmarking and analyzing 3d human pose and shape estimation beyond algorithms. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 453, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 507 + ], + "type": "text", + "content": "[38] Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regression analysis. In Proceedings Conf. on Computer Vision and Pattern Recognition (CVPR), June 2021. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 563 + ], + "type": "text", + "content": "[39] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 2, 4, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 609 + ], + "type": "text", + "content": "[40] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 2, 3, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "text", + "content": "[41] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 3, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[42] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In CVPR, pages 722-731, 2021. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 710 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[43] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[44] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the International Conference on Computer Vision, pages 1749-1759, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "text", + "content": "[45] L. Sigal, A. O. Balan, and Michael J. Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision, 87:4-27, 2009. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 237 + ], + "type": "text", + "content": "[46] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In Advances in Neural Information Processing Systems, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 238, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 545, + 270 + ], + "type": "text", + "content": "[47] Gül Varol, J. Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, I. Laptev, and C. Schmid. Learning from synthetic humans. CVPR, pages 4627-4635, 2017. 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 271, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 271, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 545, + 314 + ], + "type": "text", + "content": "[48] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 315, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 315, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 545, + 369 + ], + "type": "text", + "content": "[49] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. arXiv preprint arXiv:2303.13796, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 370, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 370, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 370, + 545, + 422 + ], + "type": "text", + "content": "[50] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In CVPR, pages 16210–16220, June 2022. 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 425, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 425, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 307, + 425, + 545, + 478 + ], + "type": "text", + "content": "[51] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3d human shape and articulated pose models. In CVPR, pages 6184-6193, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 479, + 545, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 523 + ], + "type": "text", + "content": "[52] Ailing Zeng, Xuan Ju, Lei Yang, Ruiyuan Gao, Xizhou Zhu, Bo Dai, and Qiang Xu. Deciwatch: A simple baseline for " + }, + { + "bbox": [ + 307, + 479, + 545, + 523 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 307, + 479, + 545, + 523 + ], + "type": "text", + "content": " efficient 2d and 3d pose estimation. In ECCV, pages 607-624. Springer, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 524, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 524, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 545, + 566 + ], + "type": "text", + "content": "[53] Ailing Zeng, Xiao Sun, Lei Yang, Nanxuan Zhao, Minhao Liu, and Qiang Xu. Learning skeletal graph neural networks for hard 3d pose estimation. In ICCV, pages 11436-11445, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "type": "text", + "content": "[54] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: a plug-and-play network for refining human poses in videos. In ECCV, pages 625-642. Springer, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 612, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 655 + ], + "type": "text", + "content": "[55] Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: Towards well-aligned full-body model regression from monocular images. arXiv preprint arXiv:2207.06400, 2022.3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 656, + 545, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 656, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 307, + 656, + 545, + 710 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 9" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20292" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/d13aa22d-32ff-439b-9d2f-7deb1ed92309_content_list.json b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/d13aa22d-32ff-439b-9d2f-7deb1ed92309_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4826b50ff3922e88cee501dbe6ac0d852e696fdd --- /dev/null +++ b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/d13aa22d-32ff-439b-9d2f-7deb1ed92309_content_list.json @@ -0,0 +1,1870 @@ +[ + { + "type": "text", + "text": "Synchronize Feature Extracting and Matching: A Single Branch Framework for 3D Object Tracking", + "text_level": 1, + "bbox": [ + 80, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Teli Ma $^{1,2*}$ , Mengmeng Wang $^{2*}$ , Jimin Xiao $^{3}$ , Huifeng Wu $^{4}$ , Yong Liu $^{2\\dagger}$ $^{1}$ The Hong Kong University of Science and Technology, Guangzhou $^{2}$ Zhejiang University \n $^{3}$ Xi'an Jiaotong-Liverpool University $^{4}$ Hangzhou Dianzi University", + "bbox": [ + 117, + 202, + 848, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tmal84@connect.hkust-gz.edu.cn mengmengwang@zju.edu.cn", + "bbox": [ + 235, + 260, + 730, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jimin.xiao@xjtlu.edu.cn whf@hdu.edu.cn yongliu@iipc.zju.edu.cn", + "bbox": [ + 194, + 277, + 774, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siamese network has been a de facto benchmark framework for 3D LiDAR object tracking with a shared-parametric encoder extracting features from template and search region, respectively. This paradigm relies heavily on an additional matching network to model the cross-correlation/similarity of the template and search region. In this paper, we forsake the conventional Siamese paradigm and propose a novel single-branch framework, SyncTrack, synchronizing the feature extracting and matching to avoid forwarding encoder twice for template and search region as well as introducing extra parameters of matching network. The synchronization mechanism is based on the dynamic affinity of the Transformer, and an in-depth analysis of the relevance is provided theoretically. Moreover, based on the synchronization, we introduce a novel Attentive Points-Sampling strategy into the Transformer layers (APST), replacing the random/Farthest Points Sampling (FPS) method with sampling under the supervision of attentive relations between the template and search region. It implies connecting point-wise sampling with the feature learning, beneficial to aggregating more distinctive and geometric features for tracking with sparse points. Extensive experiments on two benchmark datasets (KITTI and NuScenes) show that SyncTrack achieves state-of-the-art performance in real-time tracking.", + "bbox": [ + 73, + 358, + 473, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 763, + 209, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With recent advances in autonomous driving, 3D vision tasks based on LiDAR are becoming increasingly popular in the visual community. Among these tasks, 3D LiDAR single object tracking (SOT) aims to track a specific target object in a 3D video with the knowledge of 3D bounding", + "bbox": [ + 75, + 789, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34e95910b51c9d23fd6adbd9b0e7028940a773dffc409aba5ca465a243bbc55a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 335, + 877, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f82eb7be74c14ff2f728acb6c8423c38537bd8b731d1c78035d371a60030d7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 450, + 877, + 544 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1dd7571ea949c6276854de24591f2dc474c069cd633a58e6ecd1948342536455.jpg", + "image_caption": [ + "Figure 1. The comparison of (a) Siamese network based trackers, (b) previous single-branch, two-stage framework M2Track [51] with (c) our SyncTrack, which is a single-branch and single-stage framework." + ], + "image_footnote": [], + "bbox": [ + 511, + 545, + 877, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "box in the initial frame. This task meets numerous challenges, such as LiDAR point cloud sparsity, occlusions, and fast motions.", + "bbox": [ + 498, + 733, + 890, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing methods [32, 50, 19, 12, 33, 52, 20, 9, 39] of 3D SOT mainly adopt a Siamese-like backbone and incorporate an additional matching network to cope with the tracking challenges as shown in Fig. 1(a). Trackers based on the Siamese-like backbone separate feature extraction of template and search region, forwarding the two kinds of features with shared model parameters, respectively. Subsequently, an extra matching network is introduced to fuse", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding Author", + "bbox": [ + 96, + 887, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9953", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the extracted template and search region features to model the correlation or similarity between them. However, such a paradigm restricts the feature interaction to a post-matching network, correlating the template and search region insufficiently with merely the high-level extracted features. In other words, the matching process posterior to the encoder is incapable of modeling the relations of multi-scale features intra-backbone. Moreover, a standalone matching network results in extra model parameters and computational overheads, let alone the double forwarding process of the Siamese-backbone to extract the template and search region features. M2Track [51] proposed a motion-centric paradigm to replace the Siamese-like structure, constructing a spatial-temporal point cloud to predict the motion. However, they still rely heavily on an additional motion transformation network, which requires extra training input, to integrate the extracted template features into the search region, and another two-stage refinement network is leveraged to ensure the performance as illustrated in Fig. 1(b). Based on aforementioned problems, we ask the question: Can feature extracting and matching be conducted simultaneously in a simple way?", + "bbox": [ + 76, + 90, + 472, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The answer is Yes and the solution is implied in the dynamic global reasoning property of Transformer [35, 7, 10]. Specifically, the affinity matrix of all tokens can be constructed dynamically via continuous computation of the key and query vectors in the attention mechanism. The spatial context is aggregated using affinity to attend features. Intuitively, this affinity matrix can intrinsically serve as the matching matrix for intermediate feature interactions between the template and search region if we merge them into one input of the Transformer layers. Therefore, we propose a single-branch and single-stage framework equipped with a Transformer-based backbone instead of the conventional Siamese-like PointNet++ [31] backbone, as shown in Fig 1(c). The framework is dubbed as SyncTrack, as the Transformer backbone synchronizes the feature extracting and matching process. The SyncTrack is composed of a simple backbone and prediction head, omitting the complex matching network design and motion state estimation, depending merely on point-wise features.", + "bbox": [ + 76, + 428, + 472, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, 3D point clouds have unique properties such as sparsity [54, 24], density variance, and implicit geometric features in data locality [29]. For example, $51\\%$ samples of KITTI [11] Car category have less than 100 points [19]. These problems are further aggravated when point clouds are grouped and down-sampled to formulate multi-scale point-wise feature maps. The phenomena stresses the vitality of point clouds sampling, aiming to improve the pointwise perception efficiency with limited points. PTTR [52] proposed to sample the input point clouds before the backbone, utilizing the $L_{2}$ distance as the similarity metric for sampling. However, as down-sampling layer by layer is", + "bbox": [ + 76, + 719, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "essential in tracking backbone for multi-scale feature fusion to strengthen representation, it is reasonable to consider the sampling strategy in the backbone. Therefore, we propose the attentive sampling strategy based on the attention map between the template and search region, and equip each Transformer layer with our sampling module as shown in Fig. 2(b). We name the Transformer containing Attentive Points-Sampling as APST. Specifically, the attentive response from template tokens to search region tokens is considered, as the positively respond tokens are more likely to be in the foreground and should be preserved for feature extracting. By contrast, as Fig. 3 shows, random sampling easily falls into the perceptive confusion as selected points hardly contain geometric features due to randomness.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our paper can be summarized as follows:", + "bbox": [ + 500, + 303, + 893, + 333 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a single-branch and single-stage framework for real-time 3D LiDAR SOT dubbed SyncTrack, without Siamese-like forward propagation and a standalone matching network. We ingeniously leverage the dynamic affinity characteristic of the self-attention mechanism to synchronize the feature extracting and matching. A detailed analysis is provided to explain the synchronizing mechanism.", + "- We propose a novel APST to build the backbone, replacing the random/FPS $^1$ down-sampling of point clouds with attentive sampling to preserve more target-relevant points, and thus improving the perceptive capability of feature extracting.", + "- Extensive results show that our method has achieved new state-of-the-art performance on the KITTI and NuScenes datasets in real-time tracking, up to $2.8\\%$ and $1.4\\%$ on mean results with a high-speed of around 45 fps. Besides, SyncTrack exhibits good scalability in both width and depth." + ], + "bbox": [ + 517, + 344, + 893, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 679, + 651, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 2D Visual Tracking", + "text_level": 1, + "bbox": [ + 500, + 704, + 687, + 720 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advances in 2D video [53, 27, 26] and object tracking approaches [30, 16, 36, 23, 34, 48] stimulate the development of the 3D tracking, and the methods are evolving all the time. Early 2D SOT methods mainly focus on the classifier design like structured output SVM [36, 14], correlation filter [17, 25]. With the prevalence of deep learning, end-to-end trainable trackers emerged in the visual community. The Siamese-like structure [6, 18, 15, 23, 46, 42, 28] based methods have been popular in the tracking field and", + "bbox": [ + 496, + 728, + 893, + 864 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1In this paper, we use FPS to refer to Farthest Points Sampling and fps to denote frames per second.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9954", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f420362eb4b670a23493bb4274b368f8c835a6d9695f7078062b20156234e0fc.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 104, + 643, + 388 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/da9622d72d628dbe5d780a8a1bb78c8b4ac0631f2ca2728f6ba497c130ff0d6e.jpg", + "image_caption": [ + "(b)", + "Figure 2. (a). The overall framework of SyncTrack and specific illustration of the single-branch backbone, the template and search region input are concatenated for feature extraction and matching synchronously. (b). The structure of the Attentive Points-Sampling Transformer (APST), which samples search region points after multi-head attention." + ], + "image_footnote": [], + "bbox": [ + 666, + 101, + 883, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "have numerous variants. SiamFC [1] is a pioneering work integrating feature correlation into a fully convolutional Siamese network for visual tracking. Subsequently, improvements like introducing detection components such as region proposal network [23, 22, 8], discriminating fore and back-ground [49], anchor-free detecting [13] etc.. Recently, vision transformers have been introduced into 2D SOT [4, 38, 43, 41] to exploit the long-range modeling of the attention mechanism to fuse the features effectively. Moreover, the one-stream network trackers based on transformers are proposed in 2D tracking. The OSTrack [44] proposes joint feature learning and relation modeling, masking a proportion of image patches to save computational overheads. SimTrack [3] concatenates the template and search input, improving the patch embedding method with a Foveal window strategy. Our work draws inspiration from the aforementioned trend in 2D vision. However, it is specifically tailored to address the challenges of the 3D Single Object Tracking (SOT) problem, taking into account the unique characteristics of point clouds. Furthermore, we analyze the success of the transformer-based single-branch backbone, attributing to the dynamic affinity of the attention mechanism, which enables the synchronization of feature extraction and matching processes.", + "bbox": [ + 75, + 479, + 472, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.3D Visual Tracking", + "text_level": 1, + "bbox": [ + 500, + 479, + 686, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we only discuss LiDAR-based 3D object tracking. Till to now, almost all the 3D tracking methods [12, 32, 9, 45, 5, 33, 52, 50, 19, 20, 37] are based on the Siamese structure. The pioneering work in this field is the SC3D [12], which initially defines the task. SC3D employs cosine similarity to measure the resemblance between template and search region features, and incorporates shape completion during training to enhance appearance refinement. Trackers following SC3D make advancements from two perspectives. Firstly, they enhance the matching network [32, 50, 19, 52, 5, 39, 51]. For instance, MLVSNet [39] uses the CBAM module [40] to enhance the vote cluster features with both channel and spatial attention. STNet [20] employs cross- and self-attention modules to enhance the interaction between the extracted template and search region features, boosting their feature-level integration. M2Track introduces a motion-centric paradigm and motion estimation module to correlate the template and search features instead of appearance matching. All these methods depend on a standalone module to match the features. Secondly, the trackers [32, 9, 19, 39, 33] have attempted to improve the prediction head part. P2B [32] employs Hough Voting to predict the target location, and [50, 33, 39] all follow the voting strategy to make predictions. 3D-SiamRPN [9] uses an RPN head to predict the final results. LTTR [9] and V2B [19] use center-based regres", + "bbox": [ + 496, + 507, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9955", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sions to predict several object properties. However, few efforts are made in exploring the encoder/backbone of trackers as PointNet++ [31] is the feature extractor by default. In this study, we focus on the backbone design and incorporate the matching process directly into the backbone, significantly streamlining the tracker network.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 196, + 168, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first define the 3D SOT task in Sec. 3.1. Then, a detailed introduction of the single-branch framework is included in the Sec. 3.2. Based on the single-branch framework, how the feature extracting and matching are synchronized is elaborated in (Sec. 3.3). Moreover, we propose the Attentive Points-Sampling Transformer to build the single-branch backbone and sampling search region tokens with a strategy of attentive sampling, as shown in Sec. 3.4. The decoder head and losses are described in Sec. 3.5.", + "bbox": [ + 75, + 222, + 470, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 76, + 385, + 261, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the configuration of 3D LiDAR single object detection (SOT) task, the 3D bounding box (BBox) is defined as $(x,y,z,w,l,h,\\theta)\\in \\mathbb{R}^7$ , where the $(x,y,z)$ represents the coordinate center of the BBox and $(w,l,h),\\theta$ stand for the BBox size and heading angle (the rotation around the up-axis) respectively. Generally, the BBox size is assumed to be fixed by default even when the target object is non-rigid, thus minimizing the dimensions of BBox from $\\mathbb{R}^7$ to $\\mathbb{R}^4$ . Given a sequence of temporally-connected point clouds $\\{\\mathcal{P}_i\\}_{i = 1}^T$ ( $T$ is the number of points in each frame) and a initial BBox $\\mathcal{B}_1$ of the target, the goal of SOT is to localize the target BBoxes $\\{\\mathcal{B}_i\\}_{i = 2}^T$ in all frames online. Following the previous manner, a template point cloud $\\mathcal{P}^t = \\{p_i^t\\}_{i = 1}^{N_t}$ and a search region $\\mathcal{P}^s = \\{p_i^s\\}_{i = 1}^{N_s}$ are generated, where $N_{t}$ and $N_{s}$ are number of template and search region points. The template $P^t$ is generated by cropping and centering the target in the initial frame based on the initial BBox.", + "bbox": [ + 75, + 409, + 468, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Single-Branch Structure", + "text_level": 1, + "bbox": [ + 76, + 680, + 302, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose to replace the conventional Siamese-like backbone paradigm with a sole backbone, eliminating the double forward process of Siamese structure. Therefore, template and search region seeds are concatenated to do forward propagation. The Transformers' property of long-range relation-modeling is the intrinsic merit to tackle the concatenated template and search region seeds. Based on that, we leverage the self-attention modules to build the single-branch backbone as shown in Fig. 2(a). In our approach, we utilize a Query & Group module to sample the template seeds, denoted as $\\mathcal{P}^t\\in \\mathbb{R}^{N_t\\times C}$ , using the FPS method before jointly forwarding. This module also groups the k-nearest points to aggregate features. However, when", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "it comes to the search region point cloud, represented as $\\mathcal{P}^s\\in \\mathbb{R}^{N_s\\times C}$ , we solely employ the Group module to aggregate neighbor information without reducing the number of search points. Subsequently, the template and search seeds are concatenated, incorporating a joint parametric positional embedding for the localization of tokens. This process is as follows:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} ^ {t} = \\mathrm {k N N} (\\operatorname {F P S} (\\mathcal {P} ^ {t})), \\mathcal {P} ^ {t} \\in \\mathbb {R} ^ {N _ {t} ^ {\\prime} \\times C ^ {\\prime}}, N _ {t} ^ {\\prime} < N _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 203, + 846, + 222 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} ^ {s} = \\mathrm {k N N} \\left(\\mathcal {P} ^ {s}\\right), \\mathcal {P} ^ {s} \\in \\mathbb {R} ^ {N _ {s} \\times C ^ {\\prime}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 224, + 890, + 244 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} ^ {t s} = \\left[ \\mathcal {T} ^ {t}; \\mathcal {T} ^ {s} \\right] + \\mathbf {p _ {e}}, \\mathcal {T} ^ {t s} \\in \\mathbb {R} ^ {\\left(N _ {t} + N _ {s}\\right) \\times C ^ {\\prime}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 247, + 828, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterward, linear layers are leveraged to project the input tokens into query, key, and value latent and the head-wise joint attention map is calculated to model the intra- & interrelations of template and search tokens.", + "bbox": [ + 496, + 276, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {t s}, K ^ {t s}, V ^ {t s} = W _ {q} (\\mathcal {T} ^ {t s}), W _ {k} (\\mathcal {T} ^ {t s}), W _ {v} (\\mathcal {T} ^ {t s}),\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 357, + 846, + 376 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {m} ^ {t s} = \\operatorname {S o f t m a x} \\frac {Q _ {m} ^ {t s} \\left(K _ {m} ^ {t s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 378, + 890, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the joint head-wise attention map $\\mathcal{A}_m^{ts}$ of template and search tokens, the features extracted by multi-head attention are:", + "bbox": [ + 496, + 424, + 890, + 468 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} ^ {t s} = \\left[ \\mathcal {A} _ {1} ^ {t s} V _ {1} ^ {t s}, \\mathcal {A} _ {2} ^ {t s} V _ {2} ^ {t s}, \\dots , \\mathcal {A} _ {M} ^ {t s} V _ {M} ^ {t s} \\right] \\mathcal {W} + \\mathcal {T} ^ {t s}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 478, + 890, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M$ is the number of the attention heads and $\\mathcal{W}$ is the weight of a MLP.", + "bbox": [ + 496, + 506, + 890, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Synchronize Feature Extracting and Matching", + "text_level": 1, + "bbox": [ + 498, + 546, + 890, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We illustrate how the single-branch structure can synchronize the process of feature extracting and matching with a simple backbone. The formula of joint attention $\\mathcal{A}_m^{ts}$ in Eq. 2 can be expanded as ( $\\sigma$ represents Softmax operation):", + "bbox": [ + 496, + 569, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left[ \\mathcal {A} _ {m} ^ {t}; \\mathcal {A} _ {m} ^ {s} \\right] = \\sigma \\frac {\\left[ Q _ {m} ^ {t} ; Q _ {m} ^ {s} \\right] \\left[ K _ {m} ^ {t} ; K _ {m} ^ {s} \\right] ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} \\\\ = \\left[ \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {t} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}, \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {t} \\left(K _ {m} ^ {s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}; \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}, \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} \\right]. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 652, + 916, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The new extracted search region features can be obtained based on Eq 4 as:", + "bbox": [ + 496, + 741, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {T} _ {m} ^ {\\prime s} = \\sigma \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} V _ {m} ^ {t} + \\sigma \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} V _ {m} ^ {s}, \\tag {5} \\\\ [ V _ {m} ^ {t}; V _ {m} ^ {s} ] = V _ {m} ^ {t s}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 779, + 890, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "depending on the projected features $[V_m^t,V_m^s ]$ of both the template and search region from last layer. The attention queried from search region to template, $\\sigma \\frac{Q_m^s(K_m^t)^\\top}{\\sqrt{C' / M}}$ , is the", + "bbox": [ + 496, + 845, + 890, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9956", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26d8e974add66a293459be24a512e840c37ae2966bead8c0ac7ef90ab62195cd.jpg", + "image_caption": [ + "Template" + ], + "image_footnote": [], + "bbox": [ + 89, + 95, + 222, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8efe66a72fe2bef8d12dac791820c82d1e9149f4c63fe5fcdcd88221b0dc90ef.jpg", + "image_caption": [ + "Random Sampling" + ], + "image_footnote": [], + "bbox": [ + 233, + 123, + 328, + 248 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/30d056bab8c6fb337463f32f1cfb4f668bfda65fae82b3f2904df11ead0cee6e.jpg", + "image_caption": [ + "Attentive Sampling", + "Figure 3. Comparisons of random sampling with our attentive sampling in point-wise down-sampling. We find that our attentive sampling approach excels in selecting more representative points, including those that contribute to the formation of the head, arms, and legs, which contain distinctive geometric features. In contrast, random sampling include more points that are far away from the target (we have removed such points for better visualization)." + ], + "image_footnote": [], + "bbox": [ + 344, + 98, + 460, + 246 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "matching matrix that guides to aggregate highly-relevant template features. Moreover, the matching matrix is dynamic, which is determined by the changing query and key latent of search and template features as:", + "bbox": [ + 75, + 422, + 470, + 482 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} = \\boldsymbol {\\sigma} \\frac {W _ {q , m} ^ {s} \\mathcal {T} _ {m} ^ {s} \\left(\\mathcal {T} _ {m} ^ {t}\\right) ^ {\\top} \\left(W _ {k , m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 493, + 468, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ni. e. \\qquad \\sigma \\frac {Q _ {m} ^ {s} (K _ {m} ^ {t}) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} \\propto [ \\mathcal {T} _ {m} ^ {s}; \\mathcal {T} _ {m} ^ {t} ].\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 535, + 352, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To conclude, the synchronization attributes to the dynamic mechanism of matching matrix, continuously adapting the matching relations according to the extracted features of template and search region.", + "bbox": [ + 75, + 582, + 468, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparisons with Siamese Network. Previous Siamese-like trackers can be summarized as the paradigm of Extracting then Matching. If we name the feature extracting and matching as $\\phi$ and $\\delta$ , then the objective of this paradigm can be concluded as $max(\\delta(\\phi(\\mathcal{P}^t), \\phi(\\mathcal{P}^s))$ for simplicity. It means the model training is to shorten the distances between correlated parts of the template and search region based on the extracted features from the backbone. However, compared to our approach, this matching mechanism is relatively static since it occurs only after feature extraction, resulting in inadequate modeling of inter-backbone relations. On the other hand, our single-branch framework facilitates dynamic interaction between the search region seeds and the template across all layers of the backbone. This allows for comprehensive learning of relations, encompassing both the local representation from early layers and the global representation from later ones.", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Attentive Points-Sampling Transformer", + "text_level": 1, + "bbox": [ + 500, + 90, + 841, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the synchronizing mechanism of feature extracting and matching, we propose to replace the Point-wise Transformer [47] with a novel Attentive Points-Sampling Transformer (APST). APST is proposed based on the observation that backbones utilized in previous 3D LiDAR trackers always adopt a non-parametric strategy of down-sampling the search region features/tokens with farthest point sampling (FPS) or randomly sampling to reduce the points as introduced in PointNet++ [31]. Nevertheless, this non-parametric sampling method lacks learnability and controllability since no parameters associated with the sampling process are updated during model training. Consequently, the final performance is compromised due to the significant influence of the points-center in LiDAR, which determines the effectiveness of extracting features around the foreground, as shown in Fig. 3.", + "bbox": [ + 496, + 119, + 890, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, we introduce the APST, which involves the selection of points based on attentive relations between the tokens of the template and search region as illustrated in sub-figure (b) of Fig. 2. The attentive response of template tokens reacting to search region tokens is considered as positively responding search tokens are more likely to be in the foreground. In that case, we segment the attention map and separate the $\\mathcal{A}_m^{t\\rightarrow s} = \\sigma \\frac{Q_m^t(K_m^s)^{\\top}}{\\sqrt{C' / M}}$ from the Eq. 4, averaging along the dimension of template-tokens attention among all transformer heads to acquire the response scores. To ensure the maximum response of search tokens to template ones, a group of search region indexes, named $\\Omega_s$ , is dynamically updated as:", + "bbox": [ + 496, + 368, + 892, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {s} ^ {*} = \\arg \\max _ {\\Omega_ {s}} \\frac {1}{N _ {t}} \\frac {1}{M} \\sum_ {t = 0} ^ {N _ {t}} \\sum_ {m = 0} ^ {M} \\frac {Q _ {m} ^ {t}}{\\sqrt {C ^ {\\prime} / M}} \\left(K _ {m} ^ {s}\\right) ^ {\\top} \\Big | _ {s \\in \\Omega_ {s}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 597, + 890, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the optimal solution $\\Omega_s^*$ , the search region tokens are sampled to decrease the number of tokens after multi-head self-attention and concatenate with template ones. Note that the template tokens are sampled with the FPS method as introduced in Sec. 3.2, and only the search region tokens are sampled with the attentive sampling method.", + "bbox": [ + 496, + 652, + 890, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The dynamic-affinity property of the Transformer suggests that the attention map is influenced by the latent tokens projected through learnable linear layers. Hence, it is reasonable to assert that sampling tokens guided by the attention map is linked to model learning, as the attention map is generated based on updated parameters. As a result, this type of token/point sampling proves advantageous in providing prior knowledge for centroid selection and enhancing the efficient aggregation of representational information, as shown in Fig. 3.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9957", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Decoder and Losses", + "text_level": 1, + "bbox": [ + 76, + 90, + 266, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the encoded point-wise features of various scales, a multi-scale feature fusion module is adopted to fuse these features and output the features with the same origin input size, feeding into the decoder part for final predictions. Following the V2B [19], the features are voxelized as a volumetric representation and 3D convolutions are utilized on the encoded features. Afterward, the BEV feature maps are acquired by pooling on the $z$ -axis for regression. Focal loss and L1 loss are leveraged for classification and BBox center offset and rotation regression, respectively. A detailed introduction is presented in the supplementary.", + "bbox": [ + 75, + 113, + 472, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 292, + 209, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setups", + "text_level": 1, + "bbox": [ + 76, + 316, + 261, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We set the number of input points as $N_{t} = 512$ and $N_{s} = 1024$ for template and search regions by randomly duplicating and discarding, respectively. The encoder backbone is merely consisted of three layers with Attentive Points-Sampling Transformers, and the number of both template and search region points output by each layer are 256, 128 and 64. Moreover, each layer's feature dimensions in the encoder are 32, 64 and 128, respectively, whereas the final features for the prediction head are with 32 channels. Note that the heads for all APST are 2 by default. In the voxelization process, the region $[(x_{min}, x_{max}), (y_{min}, y_{max}), (z_{min}, z_{max})]$ is defined as $[(-5.6, 5.6), (-3.6, 3.6), (-2.4, 2.4)]$ to contain most target points. The voxel size $(v_{x}, v_{y}, v_{z})$ is set to $(0.3, 0.3, 0.3)$ . For the detection head, four decomposed 3D (stride of 2,1,2,1 along the $z$ -axis) and 2D convolution blocks (stride of 2,1,1,2) are leveraged to strengthen the feature aggregation.", + "bbox": [ + 75, + 340, + 472, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and Testing. We train the model for 40 epochs with a batch size of 64. The Adam optimizer [21] is adopted with the initial learning rate of 0.001 and reduced by 5 every 10 epochs (every 2 epochs for nuScenes). The classification loss has a weight $\\lambda_{cls}$ of 1 and the regression loss has a weight $\\lambda_{reg}$ of 1.", + "bbox": [ + 76, + 612, + 468, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following the previous methods [32, 50], we measure the Success and Precision of the tracker. To be specific, Success is defined as the IoU between predicted boxes and the ground truth, and Precision measures the AUC (Area Under Curve) of the distance between predicted and ground truth boxes within the range of [0,2] meters.", + "bbox": [ + 75, + 702, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Trackers", + "text_level": 1, + "bbox": [ + 76, + 816, + 447, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on KITTI. KITTI [11] is one of the most popular datasets used in mobile robotics and autonomous driving. The tracking benchmark of KITTI consists of 21 training sequences and 29 test sequences. Following the previous", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods [12, 51, 50], we split the training sequences into train/val/test splits due to the inaccessibility of the testing labels, scenes 0-16 for training, scenes 7-18 for validation and scenes 19-20 for testing, respectively.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the SyncTrack with other state-of-the-art methods from the pioneering SC3D [12] to the most recent siamese network STNet [20] and single branch network M2Track [51], as shown in Table 1. We separate the trackers into Siamese Network and Single Branch Network categories to compare with our proposed SyncTrack. Compared with siamese structured trackers, the SyncTrack achieves the best results on rigid and non-rigid object tracking, outperforming current tracking methods based on siamese networks on most specific categories and the overall mean results. STNet [20] is the state-of-the-art siamese-based tracking method with self-attention and cross-attention modules to match the template and search-region features. Our SyncTrack outperforms the STNet by a relatively large margin on Car, Van and Pedestrian categories under the evaluation of Success metric. Also, SyncTrack surpasses the previous best Mean results by $2.8\\%$ and $1.8\\%$ on Success and Precision metrics, respectively.", + "bbox": [ + 496, + 152, + 892, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also compare the SyncTrack with the only current single-branch tracker M2Track [51]. Our SyncTrack outperforms M2Track by a large margin, up to $7.8\\%$ in Success in the category of Car whereas the M2Track is better in the category of Pedestrian. However, for the comprehensive evaluation, the Mean performance of all frames, our SyncTrack outperforms the M2Track by $1.2\\%$ on the Success metric.", + "bbox": [ + 496, + 428, + 893, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on nuScenes. The nuScenes dataset [2] contains 1000 driving scenes collected from Boston and Singapore with a diverse and exciting set of driving maneuvers, traffic situations, and unexpected behaviors. In the configurations of LiDAR-based tracking methods, the train/val/test sets make up 700/150/150 of the whole 1000 scenes, respectively. Officially, the train set is evenly split into 'train track' and 'train detect' to remedy overfitting. Following [19], we train our model with \"train track\" split and test it on the val set.", + "bbox": [ + 496, + 550, + 893, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that nuScenes dataset only annotates keyframes and provides official interpolated results for the remaining frames, so there are two configurations for this dataset. The first is from [50, 51] which trains and tests both only on the keyframes. The second one is from [19, 20], which trains and tests on all the frames. The results based on these two configurations are different. We believe that the motion in key frames is substantial, which does not conform to the practical applications. Therefore, we train and test all the frames in this paper. We train the previous methods on the nuScenes by ourselves using their official codes to compare with our SyncTrack as many results are missing or only reported by testing nuScenes test-split with a pre", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9958", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b0670bf50791558c65499039528c182139f1abbaebd28e136c5824a9ec4dead8.jpg", + "table_caption": [ + "Table 1. Success/Precision comparisons among our SyncTrack and the state-of-the-art methods on the KITTI datasets. Mean shows frame-level averaging results. Bold and underline denote the best performance and the second-best performance, respectively. Improvements over previous state-of-the-arts are shown in Italic and color." + ], + "table_footnote": [], + "table_body": "
MethodsCar (6424)Cyclist (308)Van (1248)Pedestrian (6088)Mean (14068)
SuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecision
Siamese Network
SC3D [12]41.357.941.570.440.447.018.237.831.248.5
SC3D-RPN[45]36.351.043.081.4--17.947.8--
P2B [32]56.272.832.144.740.848.428.749.642.460.0
MLVSNet [39]56.074.034.344.552.061.434.161.145.766.6
3DSiamRPN [9]58.276.236.149.045.652.835.256.246.664.9
LTTR [5]65.077.166.289.935.845.633.256.848.765.8
PTT [33]67.881.837.247.343.652.544.972.055.174.2
BAT [50]60.577.733.745.452.467.042.170.151.272.8
V2B [19]70.581.340.849.750.158.048.373.558.475.2
PTTR [52]65.277.465.190.552.561.850.981.658.477.8
STNet [20]72.184.073.593.758.070.649.977.261.380.1
SyncTrack73.385.073.193.860.370.054.780.564.181.9
improvement+1.2+1.0-0.4+0.1+2.3-0.6+3.8-1.1+2.8+1.8
Single Branch Network
M2Track [51]65.580.873.293.553.870.761.588.262.983.4
SyncTrack73.385.073.193.860.370.054.780.564.181.9
improvement+7.8+4.2-0.1+0.3+2.3-0.7-6.8-7.7+1.2-1.5
", + "bbox": [ + 80, + 128, + 888, + 409 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1f9c169ca5dd59fe7aefd859c47d18f77eecd9a9ed0fee6200a31ee738b67ffb.jpg", + "table_caption": [ + "Table 2. Comparison among SyncTrack and the state-of-the-art methods on the nuScenes datasets. Mean shows the average result weighed by frame numbers. Bold and underline denote the best performance and the second-best performance, respectively. Improvements over previous state-of-the-arts are shown in Italic and color." + ], + "table_footnote": [], + "table_body": "
MethodsCar (15578)Bicycle (501)Truck (3710)Pedestrian (8019)Mean (27808)
SuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecision
SC3D [12]24.525.916.618.832.530.613.814.722.323.2
P2B [32]32.835.219.726.616.211.119.226.626.429.3
BAT [50]26.528.817.822.816.510.619.428.223.027.9
V2B [19]32.934.520.327.528.723.820.127.428.430.9
STNet [20]35.737.222.329.333.532.420.127.830.733.7
M2Track [51]31.433.922.629.830.128.820.728.028.031.4
SyncTrack36.738.123.830.439.438.619.127.831.835.1
improvement+1.0+0.9+1.2+0.6+5.9+6.2-1.6-0.4+1.1+1.4
", + "bbox": [ + 80, + 459, + 888, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained KITTI model.", + "bbox": [ + 76, + 631, + 222, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 2, SyncTrack performs significantly better than other trackers on the mean results of four categories. Specifically, the SyncTrack yields the best results on both metrics and on most categories except the Pedestrian, which is lower than M2Track [51] and BAT [50] by a minor margin (1.6% and 0.4%). However, in the Truck category, our SyncTrack outperforms state-of-the-art by a large margin, up to 5.9% and 6.2% on Success and Precision respectively.", + "bbox": [ + 75, + 652, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational Cost Comparison. We analyze the computational overheads and inference speed of SyncTrack and compare it with other trackers in Table 3. The reported results are tested by ourselves with official codebases with a single Titan RTX GPU on Car category of KITTI. It can be seen that SyncTrack achieves the best Success performance with the lowest computational complexity (2.51", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "G). Compared with the most current Siamese-based tracker STNet [20] and single-branch tracker M2Track, our SyncTrack has the fewest number of parameters and fastest inference speed, satisfying the demand of real-time tracking.", + "bbox": [ + 498, + 630, + 890, + 691 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a94613f4ef14400464e05f202b1805009238ef185d7c7767c9e06b85ad2dfc52.jpg", + "table_caption": [ + "Table 3. The computational cost of different trackers." + ], + "table_footnote": [], + "table_body": "
MethodsParametersFLOPsFPSSuccess
SC3D [12]6.45 M20.07 G641.3
P2B [32]1.34 M4.28 G4856.2
BAT [50]1.47 M5.53 G5460.5
V2B [19]1.36 M5.57 G3970.5
STNet [20]1.66 M3.14 G3672.1
M2Track [51]2.24 M2.54 G3765.5
SyncTrack1.47 M2.51 G4573.3
", + "bbox": [ + 513, + 748, + 879, + 862 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9959", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5801c758bbe5b150f08e940d7f5d216476ca9b0b6cd0b70a58ab08361ec1058b.jpg", + "image_caption": [ + "C" + ], + "image_footnote": [], + "bbox": [ + 145, + 88, + 279, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d9921175ab92df7386da8408bddcaad1b13c1000f55f515b1239cd7b5234897.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 88, + 421, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23375ee6f4e24ce85bf2143aa1392ae979c4d64a5168b22558f1bdc488fb5399.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 88, + 563, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e83a98dca80efba7d737f083db9c9ed48c3f96aa3b3a7b1be8a0cad3dc7c0524.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 88, + 707, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/75df7b179aaeb778fd509c2e9ca0ae0e01809b048263226b2a992d0facaa6dcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 88, + 851, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c400235239773bea98973b17f2d604ef767ebd031a0c6b83cfed6d6b667d6c5d.jpg", + "image_caption": [ + "Cyclist" + ], + "image_footnote": [], + "bbox": [ + 145, + 196, + 279, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b93847361ea53ab19398a7698ce710e1c05df01192ac06fd04b94dc2b1ab8386.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 196, + 421, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/853a7293570efed7d2ad5eec3764433c6ef52ed4b68d995aa6c0bade13ab4c46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 196, + 563, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d00c77b875d7d2dd42fd7966042455ad3d2aab82ab997233f52b9e3873df9728.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 196, + 707, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fe662d6f998f80de7f73932e75985ad017ac0c4d2d2ff661ce555dc3b63dd57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 196, + 851, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21c25b9b289b0a8498af9ec8349c22908c4bc19bf9664b7d4a693eba4646f901.jpg", + "image_caption": [ + "A" + ], + "image_footnote": [], + "bbox": [ + 145, + 306, + 279, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/24f8cfb64039ec8d215784e527b78f06dfd693e808d78015208293e78eb05032.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 306, + 421, + 410 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/baccd366d1e5ad84995dc518dc5a517f0c83468dbed80800406db066377b7f26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 308, + 563, + 410 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1560e93f0f00465798e7fc131758ad4791d1987a144fa901e4e45b435028a0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 308, + 707, + 410 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04e23faea3658d040771c416ea59fba8cc94bc2724b347cea12e93618a6eb85b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 308, + 851, + 410 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a2add9e7555554f8a6db3acc16e4031d4ee2c0a12054349b12f20476508c780.jpg", + "image_caption": [ + "Peperrnnnne" + ], + "image_footnote": [], + "bbox": [ + 145, + 412, + 279, + 516 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c10da3371a8b869ac70f1398ecfc42b8d55948c54d76723cd51565fe2552479.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 289, + 412, + 421, + 517 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97ea566193899caa919442eb0c5781062fcdbf3b85e57e5c8896738de9f8c725.jpg", + "image_caption": [ + "STNet" + ], + "image_footnote": [], + "bbox": [ + 429, + 412, + 563, + 517 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c364da645eae84d1be3f235bf6fb37a1e1539f2b2d5691d668b569449d9ec4b.jpg", + "image_caption": [ + "SyncTrack" + ], + "image_footnote": [], + "bbox": [ + 573, + 412, + 707, + 517 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1f32a3afcf2b1d4751d7b10dbd7ac31297f3531b042d2ac04bec885d9fa26a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 412, + 851, + 517 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a529280928d04447fab05c3a53a3f3f8d050cb88e980d9a83cc93ddffc477169.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 521, + 334, + 536 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4aa7209e962675ca066240ef85d094fc591ac4969422355ad49c59a0bde47b58.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with STNet. We compare SyncTrack with STNet over four tracking sequences of different categories in KITTI. The predictions of SyncTrack fits the ground-truth boxes better." + ], + "image_footnote": [], + "bbox": [ + 460, + 521, + 480, + 537 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3bd6151f15a312ed904c1966f7a420e28b60d9f89c4f0c64d8ba44ad7b6bf4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 521, + 578, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Generalization Ability", + "text_level": 1, + "bbox": [ + 76, + 592, + 284, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the generalization ability of SyncTrack, we pre-train the model on the KITTI dataset and test it directly on the nuScenes dataset without fine-tuning. The results are shown in Table 4. It can be observed that SyncTrack outperforms other methods on the mean results of four categories by a large margin. SyncTrack not only achieves a good balance between inference speed and tracking accuracy, but also generalizes very well to new domains.", + "bbox": [ + 75, + 618, + 468, + 739 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/da86ef11e76b0dabba9d8b756c29dfe2f9cc29478f978162701c7d9f560100c8.jpg", + "table_caption": [ + "Table 4. Testing results on the nuScenes for generalization ability." + ], + "table_footnote": [], + "table_body": "
MethodsCar (15578)Bicycle (501)Truck (3710)Pedestrian (8019)Mean (27808)
SC3D [12]25.0/27.117.0/18.225.7/21.914.2/16.221.8/23.1
P2B [32]27.0/29.220.0/26.421.5/16.215.9/22.022.9/25.3
BAT [50]22.5/24.117.0/18.819.3/15.817.3/24.520.5/23.0
V2B [19]31.3/35.122.2/19.121.7/19.117.3/23.425.8/29.0
STNet [20]32.2/36.121.2/29.222.3/16.819.1/27.226.9/30.8
SyncTrack32.8/36.321.7/28.323.9/19.219.3/27.127.5/31.2
", + "bbox": [ + 81, + 767, + 468, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Scalability of Backbone", + "text_level": 1, + "bbox": [ + 500, + 592, + 718, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We prove that our SyncTrack has good scalability in large-scale datasets like nuScenes. The basic model of SyncTrack has merely one Transformer layer in each stage to ensure real-time performance for tracking. However, the backbone of SyncTrack is scalable in both depth and width. We name the basic model as SyncTrack-Small. Furthermore, the SyncTrack-Mid has three Transformer layers at every stage, with nine layers total. As for the SyncTrack-Large, it doubles the number of feature channels of the SyncTrack-Mid to [256, 128, 64] for every stage. Table 5 reveals that performance improves when scalability increases not only in depth (Small v.s. Mid), but also in width (Mid v.s. Large).", + "bbox": [ + 496, + 617, + 890, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 828, + 663, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct comprehensive ablations to evaluate the components of SyncTrack.", + "bbox": [ + 498, + 853, + 890, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compare with Siamese Structure. The synchronized fea", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9960", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/877bd4f8441dd09b57216728424dfb90ba3c96a5aadd8b66e26548005f9729f5.jpg", + "table_caption": [ + "Table 5. The scalability of SyncTrack on nuScenes, (S, M, L represent small, middle, large model size of SyncTrack)." + ], + "table_footnote": [], + "table_body": "
Scale#ParamFLOPsSuccessPrecision
BicycleS1.47 M2.51 G23.830.4
M1.82 M2.63 G25.033.6
L3.98 M5.37 G25.634.1
TruckS1.47 M2.51 G39.438.6
M1.82 M2.63 G40.138.8
L3.98 M5.37 G40.538.9
", + "bbox": [ + 80, + 114, + 467, + 218 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ture extracting and matching mechanism effectively aggregates features and models the relation. To make comparisons, we split the single branch into a Siamese structure based on SyncTrack. A matching network is added to the Siamese backbone for correlating the features. The shape-aware feature learning network in V2B [19] and iterative coarse-to-fine correlation network in STNet [20] are chosen as matching networks to compare with our single-branch framework. From Table 6, the single-branch structure of SyncTrack is quite significant, and when we split the branch and add a matcher to correlate, the performance drops heavily.", + "bbox": [ + 75, + 233, + 467, + 414 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/bad53b6a91bec4b6c5d5a166a2d663b3ffcc0be99f586923c4d3f3c697be330c.jpg", + "table_caption": [ + "Table 6. Ablations of proposed Single-Branch and Siamese Structure on Car category of KITTI and nuScenes, (matches1 is from the V2B [19] and matches2 is from the STNet [20])." + ], + "table_footnote": [], + "table_body": "
StructureKITTInuScenes
SuccessPrecisionSuccessPrecision
Siamese+catcher170.482.435.336.8
Siamese+catcher271.883.735.437.1
Single-Branch73.385.036.738.1
", + "bbox": [ + 81, + 468, + 465, + 544 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Attentive Sampling. In this paper, we integrate attentive sampling into the multi-head Transformers for selecting search region points to aggregate neighborhood features. We ablate such configuration by performing attentive sampling on template tokens and both template and search region tokens, as well as comparing with standard Transformer without attentive sampling (using random sampling and random/FPS sampling) as shown in Table 7. It can be concluded that the pattern of performing attentive sampling on search region tokens in Transformers is the best. We hypothesize that sampling template tokens attentively is meaningless as template points are target-centric, and search regions' responses include much noise from the background. Therefore, it is inefficient to down-sample template tokens based on attentive responses.", + "bbox": [ + 75, + 561, + 467, + 787 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6. Visualization", + "text_level": 1, + "bbox": [ + 76, + 801, + 212, + 816 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Figure 4, we present visualization results obtained from LiDAR video sequences taken from the KITTI dataset. These visualizations show the motion pattern of objects belonging to four categories within the KITTI dataset. Obviously, our SyncTrack excels in accurately tracking the in", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e6811661c633d061d23e3e466ed143908f8eb08eee412646b7a9c54a69478a68.jpg", + "table_caption": [ + "Table 7. Ablations of attentive sampling in APST on Car category of KITTI." + ], + "table_footnote": [], + "table_body": "
TemplateSearch RegionSuccessPrecision
√APST ×randomXX70.982.8
✓X67.678.8
X✓73.385.0
✓✓69.681.1
√APST ×FPSXX71.182.8
✓X68.078.4
X✓73.285.0
✓✓70.282.7
", + "bbox": [ + 504, + 112, + 887, + 246 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tended target and predicting bounding boxes when compared to STNet [20]. This achievement is primarily attributed to the dynamic and abundant feature interactions that occur between the template and search region seeds in SyncTrack. These interactions enable our algorithm to effectively distinguish the foreground from the background, leading to superior performance.", + "bbox": [ + 496, + 258, + 890, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 377, + 617, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose SyncTrack, a novel single-branch and single-stage framework for 3D LiDAR single object tracking. SyncTrack replaces the conventional Siamese-like backbones with a single-branch one, synchronizing the feature extracting and matching without an additional matching network. Moreover, the Attentive Points-Sampling Transformer is proposed for building the backbone, and sampling search region points attentively rather than randomly. Our SyncTrack achieves good tracking performance in accuracy, efficiency, and scalability. We hope it can help motivate further research on more simple yet efficient 3D trackers.", + "bbox": [ + 496, + 402, + 890, + 582 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations discussion. Compared with motion-centric tracking framework like M2Track [51], we find our SyncTrack achieves limited performance on tiny-sized and slow-moving objects like pedestrian. We attribute it to the global reasoning mechanism of self-attention. Specifically, the semantic density of small-sized objects' token is much lower, which hinders effective informative interactions between tokens when performing self-attention. The fact that transformer-based method STNet [20] outperforms CNN-based M2track on all classes except the pedestrian (Table 1) also supports this hypothesis.", + "bbox": [ + 496, + 584, + 890, + 750 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 763, + 596, + 777 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In European conference on computer vision, pages 850–865. Springer, 2016. 3", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of" + ], + "bbox": [ + 506, + 787, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9961", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6", + "[3] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: A simplified architecture for visual object tracking. arXiv preprint arXiv:2203.05328, 2022. 3", + "[4] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8126-8135, 2021. 3", + "[5] Yubo Cui, Zheng Fang, Jiayao Shan, Zuoxu Gu, and Sifan Zhou. 3d object tracking with transformer. arXiv preprint arXiv:2110.14921, 2021. 3, 7", + "[6] Xingping Dong, Jianbing Shen, Dongming Wu, Kan Guo, Xiaogang Jin, and Fatih Porikli. Quadruplet network with one-shot learning for fast visual object tracking. IEEE Transactions on Image Processing, 28(7):3516-3527, 2019. 2", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[8] Heng Fan and Haibin Ling. Siamese cascaded region proposal networks for real-time visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7952-7961, 2019. 3", + "[9] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3d-siamrpn: an end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 1, 3, 7", + "[10] Peng Gao, Teli Ma, Hongsheng Li, Ziyi Lin, Jifeng Dai, and Yu Qiao. Mcmae: Masked convolution meets masked autoencoders. Advances in Neural Information Processing Systems, 35:35632-35644, 2022. 2", + "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 6", + "[12] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3d siamese tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1359-1368, 2019. 1, 3, 6, 7, 8", + "[13] Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. Siamcar: Siamese fully convolutional classification and regression for visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6269-6277, 2020. 3", + "[14] Sam Hare, Stuart Golodetz, Amir Saffari, Vibhav Vineet, Ming-Ming Cheng, Stephen L Hicks, and Philip HS Torr. Struck: Structured output tracking with kernels. IEEE transactions on pattern analysis and machine intelligence, 38(10):2096-2109, 2015. 2", + "[15] Anfeng He, Chong Luo, Xinmei Tian, and Wenjun Zeng. A twofold siamese network for real-time object tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4834-4843, 2018. 2" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 fps with deep regression networks. In European Conference on Computer Vision, pages 749-765. Springer, 2016. 2", + "[17] João F Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista. High-speed tracking with kernelized correlation filters. IEEE transactions on pattern analysis and machine intelligence, 37(3):583-596, 2014. 2", + "[18] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 2", + "[19] Le Hui, Lingpeng Wang, Mingmei Cheng, Jin Xie, and Jian Yang. 3d siamese voxel-to-bev tracker for sparse point clouds. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 3, 6, 7, 8, 9", + "[20] Le Hui, Lingpeng Wang, Linghua Tang, Kaihao Lan, Jin Xie, and Jian Yang. 3d siamese transformer network for single object tracking on point clouds. arXiv preprint arXiv:2207.11995, 2022. 1, 3, 6, 7, 8, 9", + "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[22] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4282-4291, 2019. 3", + "[23] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8971-8980, 2018. 2, 3", + "[24] Rong Li, Anh-Quan Cao, and Raoul de Charette. Coarse3d: Class-prototypes for contrastive learning in weakly-supervised 3d point cloud segmentation. arXiv preprint arXiv:2210.01784, 2022. 2", + "[25] Yang Li and Jianke Zhu. A scale adaptive kernel correlation filter tracker with feature integration. In European conference on computer vision, pages 254-265. Springer, 2014. 2", + "[26] Junwei Liang, Lu Jiang, Kevin Murphy, Ting Yu, and Alexander Hauptmann. The garden of forking paths: Towards multi-future trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10508-10518, 2020. 2", + "[27] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5725-5734, 2019. 2", + "[28] Xiankai Lu, Chao Ma, Jianbing Shen, Xiaokang Yang, Ian Reid, and Ming-Hsuan Yang. Deep object tracking with shrinkage loss. IEEE transactions on pattern analysis and machine intelligence, 2020. 2", + "[29] Jiageng Mao, Xiaogang Wang, and Hongsheng Li. Interpolated convolutional networks for 3d point cloud understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1578-1587, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9962", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Hyeonseob Nam and Bohyung Han. Learning multi-domain convolutional neural networks for visual tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4293-4302, 2016. 2", + "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4, 5", + "[32] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2b: Point-to-box network for 3d object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6329-6338, 2020. 1, 3, 6, 7, 8", + "[33] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. Ptt: Point-track-transformer module for 3d single object tracking in point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 1, 3, 7", + "[34] Jack Valmadre, Luca Bertinetto, Joao Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2805-2813, 2017. 2", + "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[36] Mengmeng Wang, Yong Liu, and Zeyi Huang. Large margin object tracking with circulant feature maps. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4021-4029, 2017. 2", + "[37] Mengmeng Wang, Teli Ma, Xingxing Zuo, Jiajun Lv, and Yong Liu. Correlation pyramid network for 3d single object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3215-3224, 2023. 3", + "[38] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1571-1580, 2021. 3", + "[39] Zhoutao Wang, Qian Xie, Yu-Kun Lai, Jing Wu, Kun Long, and Jun Wang. Mlvsnet: Multi-level voting siamese network for 3d visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3101-3110, 2021. 1, 3, 7", + "[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3", + "[41] Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8751-8760, 2022. 3", + "[42] Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, pages 12549-12556, 2020. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10448-10457, 2021. 3", + "[44] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In European Conference on Computer Vision, pages 341-357. Springer, 2022. 3", + "[45] Jesus Zarzar, Silvio Giancola, and Bernard Ghanem. Efficient bird eye view proposals for 3d siamese tracking. arXiv preprint arXiv:1903.10168, 2019. 3, 7", + "[46] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4591-4600, 2019. 2", + "[47] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 5", + "[48] Haojie Zhao, Gang Yang, Dong Wang, and Huchuan Lu. Deep mutual learning for visual object tracking. Pattern Recognition, 112:107796, 2021. 2", + "[49] Shaochuan Zhao, Tianyang Xu, Xiao-Jun Wu, and Xue-Feng Zhu. Adaptive feature fusion for visual object tracking. Pattern Recognition, 111:107679, 2021. 3", + "[50] Chaoda Zheng, Xu Yan, Jiantao Gao, Weibing Zhao, Wei Zhang, Zhen Li, and Shuguang Cui. Box-aware feature enhancement for single object tracking on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13199-13208, 2021. 1, 3, 6, 7, 8", + "[51] Chaoda Zheng, Xu Yan, Haiming Zhang, Baoyuan Wang, Shenghui Cheng, Shuguang Cui, and Zhen Li. Beyond 3d siamese tracking: A motion-centric paradigm for 3d single object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8111-8120, 2022. 1, 2, 3, 6, 7, 9", + "[52] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu. Pptr: Relational 3d point cloud object tracking with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8531-8540, 2022. 1, 2, 3, 7", + "[53] Jiaming Zhou, Kun-Yu Lin, Haoxin Li, and Wei-Shi Zheng. Graph-based high-order relation modeling for long-term action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8984-8993, 2021. 2", + "[54] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 839 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9963", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/d13aa22d-32ff-439b-9d2f-7deb1ed92309_model.json b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/d13aa22d-32ff-439b-9d2f-7deb1ed92309_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5ba88a281047cf94221e55207ac28e35261c5afc --- /dev/null +++ b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/d13aa22d-32ff-439b-9d2f-7deb1ed92309_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.13, + 0.89, + 0.178 + ], + "angle": 0, + "content": "Synchronize Feature Extracting and Matching: A Single Branch Framework for 3D Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.203, + 0.849, + 0.258 + ], + "angle": 0, + "content": "Teli Ma\\(^{1,2*}\\), Mengmeng Wang\\(^{2*}\\), Jimin Xiao\\(^{3}\\), Huifeng Wu\\(^{4}\\), Yong Liu\\(^{2\\dagger}\\) \n\\(^{1}\\)The Hong Kong University of Science and Technology, Guangzhou \\(^{2}\\)Zhejiang University \n\\(^{3}\\)Xi'an Jiaotong-Liverpool University \\(^{4}\\)Hangzhou Dianzi University" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.261, + 0.732, + 0.275 + ], + "angle": 0, + "content": "tmal84@connect.hkust-gz.edu.cn mengmengwang@zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.278, + 0.776, + 0.293 + ], + "angle": 0, + "content": "jimin.xiao@xjtlu.edu.cn whf@hdu.edu.cn yongliu@iipc.zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.359, + 0.474, + 0.737 + ], + "angle": 0, + "content": "Siamese network has been a de facto benchmark framework for 3D LiDAR object tracking with a shared-parametric encoder extracting features from template and search region, respectively. This paradigm relies heavily on an additional matching network to model the cross-correlation/similarity of the template and search region. In this paper, we forsake the conventional Siamese paradigm and propose a novel single-branch framework, SyncTrack, synchronizing the feature extracting and matching to avoid forwarding encoder twice for template and search region as well as introducing extra parameters of matching network. The synchronization mechanism is based on the dynamic affinity of the Transformer, and an in-depth analysis of the relevance is provided theoretically. Moreover, based on the synchronization, we introduce a novel Attentive Points-Sampling strategy into the Transformer layers (APST), replacing the random/Farthest Points Sampling (FPS) method with sampling under the supervision of attentive relations between the template and search region. It implies connecting point-wise sampling with the feature learning, beneficial to aggregating more distinctive and geometric features for tracking with sparse points. Extensive experiments on two benchmark datasets (KITTI and NuScenes) show that SyncTrack achieves state-of-the-art performance in real-time tracking." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.765, + 0.21, + 0.781 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.47, + 0.867 + ], + "angle": 0, + "content": "With recent advances in autonomous driving, 3D vision tasks based on LiDAR are becoming increasingly popular in the visual community. Among these tasks, 3D LiDAR single object tracking (SOT) aims to track a specific target object in a 3D video with the knowledge of 3D bounding" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.336, + 0.879, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.452, + 0.878, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.546, + 0.878, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.656, + 0.892, + 0.711 + ], + "angle": 0, + "content": "Figure 1. The comparison of (a) Siamese network based trackers, (b) previous single-branch, two-stage framework M2Track [51] with (c) our SyncTrack, which is a single-branch and single-stage framework." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.734, + 0.892, + 0.779 + ], + "angle": 0, + "content": "box in the initial frame. This task meets numerous challenges, such as LiDAR point cloud sparsity, occlusions, and fast motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Most existing methods [32, 50, 19, 12, 33, 52, 20, 9, 39] of 3D SOT mainly adopt a Siamese-like backbone and incorporate an additional matching network to cope with the tracking challenges as shown in Fig. 1(a). Trackers based on the Siamese-like backbone separate feature extraction of template and search region, forwarding the two kinds of features with shared model parameters, respectively. Subsequently, an extra matching network is introduced to fuse" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.208, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.225, + 0.9 + ], + "angle": 0, + "content": "†Corresponding Author" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.225, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9953" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.425 + ], + "angle": 0, + "content": "the extracted template and search region features to model the correlation or similarity between them. However, such a paradigm restricts the feature interaction to a post-matching network, correlating the template and search region insufficiently with merely the high-level extracted features. In other words, the matching process posterior to the encoder is incapable of modeling the relations of multi-scale features intra-backbone. Moreover, a standalone matching network results in extra model parameters and computational overheads, let alone the double forwarding process of the Siamese-backbone to extract the template and search region features. M2Track [51] proposed a motion-centric paradigm to replace the Siamese-like structure, constructing a spatial-temporal point cloud to predict the motion. However, they still rely heavily on an additional motion transformation network, which requires extra training input, to integrate the extracted template features into the search region, and another two-stage refinement network is leveraged to ensure the performance as illustrated in Fig. 1(b). Based on aforementioned problems, we ask the question: Can feature extracting and matching be conducted simultaneously in a simple way?" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.429, + 0.473, + 0.716 + ], + "angle": 0, + "content": "The answer is Yes and the solution is implied in the dynamic global reasoning property of Transformer [35, 7, 10]. Specifically, the affinity matrix of all tokens can be constructed dynamically via continuous computation of the key and query vectors in the attention mechanism. The spatial context is aggregated using affinity to attend features. Intuitively, this affinity matrix can intrinsically serve as the matching matrix for intermediate feature interactions between the template and search region if we merge them into one input of the Transformer layers. Therefore, we propose a single-branch and single-stage framework equipped with a Transformer-based backbone instead of the conventional Siamese-like PointNet++ [31] backbone, as shown in Fig 1(c). The framework is dubbed as SyncTrack, as the Transformer backbone synchronizes the feature extracting and matching process. The SyncTrack is composed of a simple backbone and prediction head, omitting the complex matching network design and motion state estimation, depending merely on point-wise features." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.473, + 0.904 + ], + "angle": 0, + "content": "However, 3D point clouds have unique properties such as sparsity [54, 24], density variance, and implicit geometric features in data locality [29]. For example, \\(51\\%\\) samples of KITTI [11] Car category have less than 100 points [19]. These problems are further aggravated when point clouds are grouped and down-sampled to formulate multi-scale point-wise feature maps. The phenomena stresses the vitality of point clouds sampling, aiming to improve the pointwise perception efficiency with limited points. PTTR [52] proposed to sample the input point clouds before the backbone, utilizing the \\(L_{2}\\) distance as the similarity metric for sampling. However, as down-sampling layer by layer is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "essential in tracking backbone for multi-scale feature fusion to strengthen representation, it is reasonable to consider the sampling strategy in the backbone. Therefore, we propose the attentive sampling strategy based on the attention map between the template and search region, and equip each Transformer layer with our sampling module as shown in Fig. 2(b). We name the Transformer containing Attentive Points-Sampling as APST. Specifically, the attentive response from template tokens to search region tokens is considered, as the positively respond tokens are more likely to be in the foreground and should be preserved for feature extracting. By contrast, as Fig. 3 shows, random sampling easily falls into the perceptive confusion as selected points hardly contain geometric features due to randomness." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.304, + 0.895, + 0.334 + ], + "angle": 0, + "content": "The main contributions of our paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.345, + 0.892, + 0.466 + ], + "angle": 0, + "content": "- We introduce a single-branch and single-stage framework for real-time 3D LiDAR SOT dubbed SyncTrack, without Siamese-like forward propagation and a standalone matching network. We ingeniously leverage the dynamic affinity characteristic of the self-attention mechanism to synchronize the feature extracting and matching. A detailed analysis is provided to explain the synchronizing mechanism." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.477, + 0.892, + 0.554 + ], + "angle": 0, + "content": "- We propose a novel APST to build the backbone, replacing the random/FPS\\(^1\\) down-sampling of point clouds with attentive sampling to preserve more target-relevant points, and thus improving the perceptive capability of feature extracting." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.564, + 0.894, + 0.655 + ], + "angle": 0, + "content": "- Extensive results show that our method has achieved new state-of-the-art performance on the KITTI and NuScenes datasets in real-time tracking, up to \\(2.8\\%\\) and \\(1.4\\%\\) on mean results with a high-speed of around 45 fps. Besides, SyncTrack exhibits good scalability in both width and depth." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.345, + 0.894, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.652, + 0.696 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.705, + 0.688, + 0.722 + ], + "angle": 0, + "content": "2.1. 2D Visual Tracking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.894, + 0.866 + ], + "angle": 0, + "content": "Advances in 2D video [53, 27, 26] and object tracking approaches [30, 16, 36, 23, 34, 48] stimulate the development of the 3D tracking, and the methods are evolving all the time. Early 2D SOT methods mainly focus on the classifier design like structured output SVM [36, 14], correlation filter [17, 25]. With the prevalence of deep learning, end-to-end trainable trackers emerged in the visual community. The Siamese-like structure [6, 18, 15, 23, 46, 42, 28] based methods have been popular in the tracking field and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1In this paper, we use FPS to refer to Farthest Points Sampling and fps to denote frames per second." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9954" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.106, + 0.645, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.394, + 0.409, + 0.407 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.102, + 0.885, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.394, + 0.784, + 0.407 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.413, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Figure 2. (a). The overall framework of SyncTrack and specific illustration of the single-branch backbone, the template and search region input are concatenated for feature extraction and matching synchronously. (b). The structure of the Attentive Points-Sampling Transformer (APST), which samples search region points after multi-head attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.473, + 0.844 + ], + "angle": 0, + "content": "have numerous variants. SiamFC [1] is a pioneering work integrating feature correlation into a fully convolutional Siamese network for visual tracking. Subsequently, improvements like introducing detection components such as region proposal network [23, 22, 8], discriminating fore and back-ground [49], anchor-free detecting [13] etc.. Recently, vision transformers have been introduced into 2D SOT [4, 38, 43, 41] to exploit the long-range modeling of the attention mechanism to fuse the features effectively. Moreover, the one-stream network trackers based on transformers are proposed in 2D tracking. The OSTrack [44] proposes joint feature learning and relation modeling, masking a proportion of image patches to save computational overheads. SimTrack [3] concatenates the template and search input, improving the patch embedding method with a Foveal window strategy. Our work draws inspiration from the aforementioned trend in 2D vision. However, it is specifically tailored to address the challenges of the 3D Single Object Tracking (SOT) problem, taking into account the unique characteristics of point clouds. Furthermore, we analyze the success of the transformer-based single-branch backbone, attributing to the dynamic affinity of the attention mechanism, which enables the synchronization of feature extraction and matching processes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.48, + 0.687, + 0.496 + ], + "angle": 0, + "content": "2.2.3D Visual Tracking" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this paper, we only discuss LiDAR-based 3D object tracking. Till to now, almost all the 3D tracking methods [12, 32, 9, 45, 5, 33, 52, 50, 19, 20, 37] are based on the Siamese structure. The pioneering work in this field is the SC3D [12], which initially defines the task. SC3D employs cosine similarity to measure the resemblance between template and search region features, and incorporates shape completion during training to enhance appearance refinement. Trackers following SC3D make advancements from two perspectives. Firstly, they enhance the matching network [32, 50, 19, 52, 5, 39, 51]. For instance, MLVSNet [39] uses the CBAM module [40] to enhance the vote cluster features with both channel and spatial attention. STNet [20] employs cross- and self-attention modules to enhance the interaction between the extracted template and search region features, boosting their feature-level integration. M2Track introduces a motion-centric paradigm and motion estimation module to correlate the template and search features instead of appearance matching. All these methods depend on a standalone module to match the features. Secondly, the trackers [32, 9, 19, 39, 33] have attempted to improve the prediction head part. P2B [32] employs Hough Voting to predict the target location, and [50, 33, 39] all follow the voting strategy to make predictions. 3D-SiamRPN [9] uses an RPN head to predict the final results. LTTR [9] and V2B [19] use center-based regres" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9955" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "sions to predict several object properties. However, few efforts are made in exploring the encoder/backbone of trackers as PointNet++ [31] is the feature extractor by default. In this study, we focus on the backbone design and incorporate the matching process directly into the backbone, significantly streamlining the tracker network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.197, + 0.169, + 0.213 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.223, + 0.471, + 0.375 + ], + "angle": 0, + "content": "In this section, we first define the 3D SOT task in Sec. 3.1. Then, a detailed introduction of the single-branch framework is included in the Sec. 3.2. Based on the single-branch framework, how the feature extracting and matching are synchronized is elaborated in (Sec. 3.3). Moreover, we propose the Attentive Points-Sampling Transformer to build the single-branch backbone and sampling search region tokens with a strategy of attentive sampling, as shown in Sec. 3.4. The decoder head and losses are described in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.386, + 0.262, + 0.401 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.671 + ], + "angle": 0, + "content": "In the configuration of 3D LiDAR single object detection (SOT) task, the 3D bounding box (BBox) is defined as \\((x,y,z,w,l,h,\\theta)\\in \\mathbb{R}^7\\), where the \\((x,y,z)\\) represents the coordinate center of the BBox and \\((w,l,h),\\theta\\) stand for the BBox size and heading angle (the rotation around the up-axis) respectively. Generally, the BBox size is assumed to be fixed by default even when the target object is non-rigid, thus minimizing the dimensions of BBox from \\(\\mathbb{R}^7\\) to \\(\\mathbb{R}^4\\). Given a sequence of temporally-connected point clouds \\(\\{\\mathcal{P}_i\\}_{i = 1}^T\\) (\\(T\\) is the number of points in each frame) and a initial BBox \\(\\mathcal{B}_1\\) of the target, the goal of SOT is to localize the target BBoxes \\(\\{\\mathcal{B}_i\\}_{i = 2}^T\\) in all frames online. Following the previous manner, a template point cloud \\(\\mathcal{P}^t = \\{p_i^t\\}_{i = 1}^{N_t}\\) and a search region \\(\\mathcal{P}^s = \\{p_i^s\\}_{i = 1}^{N_s}\\) are generated, where \\(N_{t}\\) and \\(N_{s}\\) are number of template and search region points. The template \\(P^t\\) is generated by cropping and centering the target in the initial frame based on the initial BBox." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.303, + 0.696 + ], + "angle": 0, + "content": "3.2. Single-Branch Structure" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We propose to replace the conventional Siamese-like backbone paradigm with a sole backbone, eliminating the double forward process of Siamese structure. Therefore, template and search region seeds are concatenated to do forward propagation. The Transformers' property of long-range relation-modeling is the intrinsic merit to tackle the concatenated template and search region seeds. Based on that, we leverage the self-attention modules to build the single-branch backbone as shown in Fig. 2(a). In our approach, we utilize a Query & Group module to sample the template seeds, denoted as \\(\\mathcal{P}^t\\in \\mathbb{R}^{N_t\\times C}\\), using the FPS method before jointly forwarding. This module also groups the k-nearest points to aggregate features. However, when" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "it comes to the search region point cloud, represented as \\(\\mathcal{P}^s\\in \\mathbb{R}^{N_s\\times C}\\), we solely employ the Group module to aggregate neighbor information without reducing the number of search points. Subsequently, the template and search seeds are concatenated, incorporating a joint parametric positional embedding for the localization of tokens. This process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.204, + 0.847, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} ^ {t} = \\mathrm {k N N} (\\operatorname {F P S} (\\mathcal {P} ^ {t})), \\mathcal {P} ^ {t} \\in \\mathbb {R} ^ {N _ {t} ^ {\\prime} \\times C ^ {\\prime}}, N _ {t} ^ {\\prime} < N _ {t},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.226, + 0.892, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} ^ {s} = \\mathrm {k N N} \\left(\\mathcal {P} ^ {s}\\right), \\mathcal {P} ^ {s} \\in \\mathbb {R} ^ {N _ {s} \\times C ^ {\\prime}}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.248, + 0.829, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} ^ {t s} = \\left[ \\mathcal {T} ^ {t}; \\mathcal {T} ^ {s} \\right] + \\mathbf {p _ {e}}, \\mathcal {T} ^ {t s} \\in \\mathbb {R} ^ {\\left(N _ {t} + N _ {s}\\right) \\times C ^ {\\prime}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Afterward, linear layers are leveraged to project the input tokens into query, key, and value latent and the head-wise joint attention map is calculated to model the intra- & interrelations of template and search tokens." + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.358, + 0.847, + 0.377 + ], + "angle": 0, + "content": "\\[\nQ ^ {t s}, K ^ {t s}, V ^ {t s} = W _ {q} (\\mathcal {T} ^ {t s}), W _ {k} (\\mathcal {T} ^ {t s}), W _ {v} (\\mathcal {T} ^ {t s}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.379, + 0.891, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {m} ^ {t s} = \\operatorname {S o f t m a x} \\frac {Q _ {m} ^ {t s} \\left(K _ {m} ^ {t s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Based on the joint head-wise attention map \\(\\mathcal{A}_m^{ts}\\) of template and search tokens, the features extracted by multi-head attention are:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.479, + 0.892, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} ^ {t s} = \\left[ \\mathcal {A} _ {1} ^ {t s} V _ {1} ^ {t s}, \\mathcal {A} _ {2} ^ {t s} V _ {2} ^ {t s}, \\dots , \\mathcal {A} _ {M} ^ {t s} V _ {M} ^ {t s} \\right] \\mathcal {W} + \\mathcal {T} ^ {t s}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.891, + 0.538 + ], + "angle": 0, + "content": "where \\(M\\) is the number of the attention heads and \\(\\mathcal{W}\\) is the weight of a MLP." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.547, + 0.892, + 0.563 + ], + "angle": 0, + "content": "3.3. Synchronize Feature Extracting and Matching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.645 + ], + "angle": 0, + "content": "We illustrate how the single-branch structure can synchronize the process of feature extracting and matching with a simple backbone. The formula of joint attention \\( \\mathcal{A}_m^{ts} \\) in Eq. 2 can be expanded as ( \\( \\sigma \\) represents Softmax operation):" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.654, + 0.918, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left[ \\mathcal {A} _ {m} ^ {t}; \\mathcal {A} _ {m} ^ {s} \\right] = \\sigma \\frac {\\left[ Q _ {m} ^ {t} ; Q _ {m} ^ {s} \\right] \\left[ K _ {m} ^ {t} ; K _ {m} ^ {s} \\right] ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} \\\\ = \\left[ \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {t} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}, \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {t} \\left(K _ {m} ^ {s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}; \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}, \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} \\right]. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.892, + 0.772 + ], + "angle": 0, + "content": "The new extracted search region features can be obtained based on Eq 4 as:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.78, + 0.891, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {T} _ {m} ^ {\\prime s} = \\sigma \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} V _ {m} ^ {t} + \\sigma \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} V _ {m} ^ {s}, \\tag {5} \\\\ [ V _ {m} ^ {t}; V _ {m} ^ {s} ] = V _ {m} ^ {t s}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.847, + 0.892, + 0.905 + ], + "angle": 0, + "content": "depending on the projected features \\([V_m^t,V_m^s ]\\) of both the template and search region from last layer. The attention queried from search region to template, \\(\\sigma \\frac{Q_m^s(K_m^t)^\\top}{\\sqrt{C' / M}}\\) , is the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9956" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.096, + 0.223, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.256, + 0.184, + 0.269 + ], + "angle": 0, + "content": "Template" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.124, + 0.33, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.25, + 0.307, + 0.276 + ], + "angle": 0, + "content": "Random Sampling" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.099, + 0.462, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.25, + 0.429, + 0.276 + ], + "angle": 0, + "content": "Attentive Sampling" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.473, + 0.397 + ], + "angle": 0, + "content": "Figure 3. Comparisons of random sampling with our attentive sampling in point-wise down-sampling. We find that our attentive sampling approach excels in selecting more representative points, including those that contribute to the formation of the head, arms, and legs, which contain distinctive geometric features. In contrast, random sampling include more points that are far away from the target (we have removed such points for better visualization)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.471, + 0.483 + ], + "angle": 0, + "content": "matching matrix that guides to aggregate highly-relevant template features. Moreover, the matching matrix is dynamic, which is determined by the changing query and key latent of search and template features as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.494, + 0.47, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} = \\boldsymbol {\\sigma} \\frac {W _ {q , m} ^ {s} \\mathcal {T} _ {m} ^ {s} \\left(\\mathcal {T} _ {m} ^ {t}\\right) ^ {\\top} \\left(W _ {k , m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.536, + 0.353, + 0.573 + ], + "angle": 0, + "content": "\\[\ni. e. \\qquad \\sigma \\frac {Q _ {m} ^ {s} (K _ {m} ^ {t}) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} \\propto [ \\mathcal {T} _ {m} ^ {s}; \\mathcal {T} _ {m} ^ {t} ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.643 + ], + "angle": 0, + "content": "To conclude, the synchronization attributes to the dynamic mechanism of matching matrix, continuously adapting the matching relations according to the extracted features of template and search region." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Comparisons with Siamese Network. Previous Siamese-like trackers can be summarized as the paradigm of Extracting then Matching. If we name the feature extracting and matching as \\(\\phi\\) and \\(\\delta\\), then the objective of this paradigm can be concluded as \\(max(\\delta(\\phi(\\mathcal{P}^t), \\phi(\\mathcal{P}^s))\\) for simplicity. It means the model training is to shorten the distances between correlated parts of the template and search region based on the extracted features from the backbone. However, compared to our approach, this matching mechanism is relatively static since it occurs only after feature extraction, resulting in inadequate modeling of inter-backbone relations. On the other hand, our single-branch framework facilitates dynamic interaction between the search region seeds and the template across all layers of the backbone. This allows for comprehensive learning of relations, encompassing both the local representation from early layers and the global representation from later ones." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.842, + 0.108 + ], + "angle": 0, + "content": "3.4. Attentive Points-Sampling Transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Based on the synchronizing mechanism of feature extracting and matching, we propose to replace the Point-wise Transformer [47] with a novel Attentive Points-Sampling Transformer (APST). APST is proposed based on the observation that backbones utilized in previous 3D LiDAR trackers always adopt a non-parametric strategy of down-sampling the search region features/tokens with farthest point sampling (FPS) or randomly sampling to reduce the points as introduced in PointNet++ [31]. Nevertheless, this non-parametric sampling method lacks learnability and controllability since no parameters associated with the sampling process are updated during model training. Consequently, the final performance is compromised due to the significant influence of the points-center in LiDAR, which determines the effectiveness of extracting features around the foreground, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.893, + 0.575 + ], + "angle": 0, + "content": "Therefore, we introduce the APST, which involves the selection of points based on attentive relations between the tokens of the template and search region as illustrated in sub-figure (b) of Fig. 2. The attentive response of template tokens reacting to search region tokens is considered as positively responding search tokens are more likely to be in the foreground. In that case, we segment the attention map and separate the \\( \\mathcal{A}_m^{t\\rightarrow s} = \\sigma \\frac{Q_m^t(K_m^s)^{\\top}}{\\sqrt{C' / M}} \\) from the Eq. 4, averaging along the dimension of template-tokens attention among all transformer heads to acquire the response scores. To ensure the maximum response of search tokens to template ones, a group of search region indexes, named \\( \\Omega_s \\), is dynamically updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.598, + 0.891, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {s} ^ {*} = \\arg \\max _ {\\Omega_ {s}} \\frac {1}{N _ {t}} \\frac {1}{M} \\sum_ {t = 0} ^ {N _ {t}} \\sum_ {m = 0} ^ {M} \\frac {Q _ {m} ^ {t}}{\\sqrt {C ^ {\\prime} / M}} \\left(K _ {m} ^ {s}\\right) ^ {\\top} \\Big | _ {s \\in \\Omega_ {s}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Based on the optimal solution \\(\\Omega_s^*\\), the search region tokens are sampled to decrease the number of tokens after multi-head self-attention and concatenate with template ones. Note that the template tokens are sampled with the FPS method as introduced in Sec. 3.2, and only the search region tokens are sampled with the attentive sampling method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The dynamic-affinity property of the Transformer suggests that the attention map is influenced by the latent tokens projected through learnable linear layers. Hence, it is reasonable to assert that sampling tokens guided by the attention map is linked to model learning, as the attention map is generated based on updated parameters. As a result, this type of token/point sampling proves advantageous in providing prior knowledge for centroid selection and enhancing the efficient aggregation of representational information, as shown in Fig. 3." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9957" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.267, + 0.106 + ], + "angle": 0, + "content": "3.5. Decoder and Losses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.473, + 0.283 + ], + "angle": 0, + "content": "With the encoded point-wise features of various scales, a multi-scale feature fusion module is adopted to fuse these features and output the features with the same origin input size, feeding into the decoder part for final predictions. Following the V2B [19], the features are voxelized as a volumetric representation and 3D convolutions are utilized on the encoded features. Afterward, the BEV feature maps are acquired by pooling on the \\(z\\)-axis for regression. Focal loss and L1 loss are leveraged for classification and BBox center offset and rotation regression, respectively. A detailed introduction is presented in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.293, + 0.21, + 0.31 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.318, + 0.262, + 0.334 + ], + "angle": 0, + "content": "4.1. Experiment Setups" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.473, + 0.612 + ], + "angle": 0, + "content": "Implementation Details. We set the number of input points as \\( N_{t} = 512 \\) and \\( N_{s} = 1024 \\) for template and search regions by randomly duplicating and discarding, respectively. The encoder backbone is merely consisted of three layers with Attentive Points-Sampling Transformers, and the number of both template and search region points output by each layer are 256, 128 and 64. Moreover, each layer's feature dimensions in the encoder are 32, 64 and 128, respectively, whereas the final features for the prediction head are with 32 channels. Note that the heads for all APST are 2 by default. In the voxelization process, the region \\( [(x_{min}, x_{max}), (y_{min}, y_{max}), (z_{min}, z_{max})] \\) is defined as \\( [(-5.6, 5.6), (-3.6, 3.6), (-2.4, 2.4)] \\) to contain most target points. The voxel size \\( (v_{x}, v_{y}, v_{z}) \\) is set to \\( (0.3, 0.3, 0.3) \\). For the detection head, four decomposed 3D (stride of 2,1,2,1 along the \\( z \\)-axis) and 2D convolution blocks (stride of 2,1,1,2) are leveraged to strengthen the feature aggregation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.613, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Training and Testing. We train the model for 40 epochs with a batch size of 64. The Adam optimizer [21] is adopted with the initial learning rate of 0.001 and reduced by 5 every 10 epochs (every 2 epochs for nuScenes). The classification loss has a weight \\(\\lambda_{cls}\\) of 1 and the regression loss has a weight \\(\\lambda_{reg}\\) of 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Evaluation Metrics. Following the previous methods [32, 50], we measure the Success and Precision of the tracker. To be specific, Success is defined as the IoU between predicted boxes and the ground truth, and Precision measures the AUC (Area Under Curve) of the distance between predicted and ground truth boxes within the range of [0,2] meters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.449, + 0.834 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Trackers" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Results on KITTI. KITTI [11] is one of the most popular datasets used in mobile robotics and autonomous driving. The tracking benchmark of KITTI consists of 21 training sequences and 29 test sequences. Following the previous" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "methods [12, 51, 50], we split the training sequences into train/val/test splits due to the inaccessibility of the testing labels, scenes 0-16 for training, scenes 7-18 for validation and scenes 19-20 for testing, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.893, + 0.427 + ], + "angle": 0, + "content": "We compare the SyncTrack with other state-of-the-art methods from the pioneering SC3D [12] to the most recent siamese network STNet [20] and single branch network M2Track [51], as shown in Table 1. We separate the trackers into Siamese Network and Single Branch Network categories to compare with our proposed SyncTrack. Compared with siamese structured trackers, the SyncTrack achieves the best results on rigid and non-rigid object tracking, outperforming current tracking methods based on siamese networks on most specific categories and the overall mean results. STNet [20] is the state-of-the-art siamese-based tracking method with self-attention and cross-attention modules to match the template and search-region features. Our SyncTrack outperforms the STNet by a relatively large margin on Car, Van and Pedestrian categories under the evaluation of Success metric. Also, SyncTrack surpasses the previous best Mean results by \\(2.8\\%\\) and \\(1.8\\%\\) on Success and Precision metrics, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.894, + 0.548 + ], + "angle": 0, + "content": "We also compare the SyncTrack with the only current single-branch tracker M2Track [51]. Our SyncTrack outperforms M2Track by a large margin, up to \\(7.8\\%\\) in Success in the category of Car whereas the M2Track is better in the category of Pedestrian. However, for the comprehensive evaluation, the Mean performance of all frames, our SyncTrack outperforms the M2Track by \\(1.2\\%\\) on the Success metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.894, + 0.702 + ], + "angle": 0, + "content": "Results on nuScenes. The nuScenes dataset [2] contains 1000 driving scenes collected from Boston and Singapore with a diverse and exciting set of driving maneuvers, traffic situations, and unexpected behaviors. In the configurations of LiDAR-based tracking methods, the train/val/test sets make up 700/150/150 of the whole 1000 scenes, respectively. Officially, the train set is evenly split into 'train track' and 'train detect' to remedy overfitting. Following [19], we train our model with \"train track\" split and test it on the val set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Note that nuScenes dataset only annotates keyframes and provides official interpolated results for the remaining frames, so there are two configurations for this dataset. The first is from [50, 51] which trains and tests both only on the keyframes. The second one is from [19, 20], which trains and tests on all the frames. The results based on these two configurations are different. We believe that the motion in key frames is substantial, which does not conform to the practical applications. Therefore, we train and test all the frames in this paper. We train the previous methods on the nuScenes by ourselves using their official codes to compare with our SyncTrack as many results are missing or only reported by testing nuScenes test-split with a pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9958" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.892, + 0.128 + ], + "angle": 0, + "content": "Table 1. Success/Precision comparisons among our SyncTrack and the state-of-the-art methods on the KITTI datasets. Mean shows frame-level averaging results. Bold and underline denote the best performance and the second-best performance, respectively. Improvements over previous state-of-the-arts are shown in Italic and color." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.129, + 0.89, + 0.41 + ], + "angle": 0, + "content": "
MethodsCar (6424)Cyclist (308)Van (1248)Pedestrian (6088)Mean (14068)
SuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecision
Siamese Network
SC3D [12]41.357.941.570.440.447.018.237.831.248.5
SC3D-RPN[45]36.351.043.081.4--17.947.8--
P2B [32]56.272.832.144.740.848.428.749.642.460.0
MLVSNet [39]56.074.034.344.552.061.434.161.145.766.6
3DSiamRPN [9]58.276.236.149.045.652.835.256.246.664.9
LTTR [5]65.077.166.289.935.845.633.256.848.765.8
PTT [33]67.881.837.247.343.652.544.972.055.174.2
BAT [50]60.577.733.745.452.467.042.170.151.272.8
V2B [19]70.581.340.849.750.158.048.373.558.475.2
PTTR [52]65.277.465.190.552.561.850.981.658.477.8
STNet [20]72.184.073.593.758.070.649.977.261.380.1
SyncTrack73.385.073.193.860.370.054.780.564.181.9
improvement+1.2+1.0-0.4+0.1+2.3-0.6+3.8-1.1+2.8+1.8
Single Branch Network
M2Track [51]65.580.873.293.553.870.761.588.262.983.4
SyncTrack73.385.073.193.860.370.054.780.564.181.9
improvement+7.8+4.2-0.1+0.3+2.3-0.7-6.8-7.7+1.2-1.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.42, + 0.893, + 0.46 + ], + "angle": 0, + "content": "Table 2. Comparison among SyncTrack and the state-of-the-art methods on the nuScenes datasets. Mean shows the average result weighed by frame numbers. Bold and underline denote the best performance and the second-best performance, respectively. Improvements over previous state-of-the-arts are shown in Italic and color." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.46, + 0.89, + 0.607 + ], + "angle": 0, + "content": "
MethodsCar (15578)Bicycle (501)Truck (3710)Pedestrian (8019)Mean (27808)
SuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecision
SC3D [12]24.525.916.618.832.530.613.814.722.323.2
P2B [32]32.835.219.726.616.211.119.226.626.429.3
BAT [50]26.528.817.822.816.510.619.428.223.027.9
V2B [19]32.934.520.327.528.723.820.127.428.430.9
STNet [20]35.737.222.329.333.532.420.127.830.733.7
M2Track [51]31.433.922.629.830.128.820.728.028.031.4
SyncTrack36.738.123.830.439.438.619.127.831.835.1
improvement+1.0+0.9+1.2+0.6+5.9+6.2-1.6-0.4+1.1+1.4
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.632, + 0.223, + 0.646 + ], + "angle": 0, + "content": "trained KITTI model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.789 + ], + "angle": 0, + "content": "As shown in Table 2, SyncTrack performs significantly better than other trackers on the mean results of four categories. Specifically, the SyncTrack yields the best results on both metrics and on most categories except the Pedestrian, which is lower than M2Track [51] and BAT [50] by a minor margin (1.6% and 0.4%). However, in the Truck category, our SyncTrack outperforms state-of-the-art by a large margin, up to 5.9% and 6.2% on Success and Precision respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Computational Cost Comparison. We analyze the computational overheads and inference speed of SyncTrack and compare it with other trackers in Table 3. The reported results are tested by ourselves with official codebases with a single Titan RTX GPU on Car category of KITTI. It can be seen that SyncTrack achieves the best Success performance with the lowest computational complexity (2.51" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.631, + 0.892, + 0.693 + ], + "angle": 0, + "content": "G). Compared with the most current Siamese-based tracker STNet [20] and single-branch tracker M2Track, our SyncTrack has the fewest number of parameters and fastest inference speed, satisfying the demand of real-time tracking." + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.736, + 0.854, + 0.749 + ], + "angle": 0, + "content": "Table 3. The computational cost of different trackers." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.749, + 0.88, + 0.863 + ], + "angle": 0, + "content": "
MethodsParametersFLOPsFPSSuccess
SC3D [12]6.45 M20.07 G641.3
P2B [32]1.34 M4.28 G4856.2
BAT [50]1.47 M5.53 G5460.5
V2B [19]1.36 M5.57 G3970.5
STNet [20]1.66 M3.14 G3672.1
M2Track [51]2.24 M2.54 G3765.5
SyncTrack1.47 M2.51 G4573.3
" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9959" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.128, + 0.139, + 0.152 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.089, + 0.28, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.089, + 0.423, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.089, + 0.565, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.089, + 0.708, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.089, + 0.852, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.228, + 0.14, + 0.266 + ], + "angle": 0, + "content": "Cyclist" + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.198, + 0.28, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.198, + 0.423, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.198, + 0.565, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.198, + 0.708, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.198, + 0.852, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.359, + 0.138, + 0.38 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.308, + 0.28, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.308, + 0.423, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.309, + 0.565, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.309, + 0.708, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.309, + 0.852, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.433, + 0.138, + 0.489 + ], + "angle": 0, + "content": "Peperrnnnne" + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.414, + 0.28, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.414, + 0.423, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.414, + 0.565, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.414, + 0.708, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.414, + 0.852, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.522, + 0.336, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.524, + 0.437, + 0.537 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.522, + 0.481, + 0.538 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.524, + 0.531, + 0.537 + ], + "angle": 0, + "content": "STNet" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.522, + 0.58, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.524, + 0.657, + 0.537 + ], + "angle": 0, + "content": "SyncTrack" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.541, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with STNet. We compare SyncTrack with STNet over four tracking sequences of different categories in KITTI. The predictions of SyncTrack fits the ground-truth boxes better." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.593, + 0.285, + 0.61 + ], + "angle": 0, + "content": "4.3. Generalization Ability" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.47, + 0.74 + ], + "angle": 0, + "content": "To evaluate the generalization ability of SyncTrack, we pre-train the model on the KITTI dataset and test it directly on the nuScenes dataset without fine-tuning. The results are shown in Table 4. It can be observed that SyncTrack outperforms other methods on the mean results of four categories by a large margin. SyncTrack not only achieves a good balance between inference speed and tracking accuracy, but also generalizes very well to new domains." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.754, + 0.468, + 0.768 + ], + "angle": 0, + "content": "Table 4. Testing results on the nuScenes for generalization ability." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.768, + 0.47, + 0.88 + ], + "angle": 0, + "content": "
MethodsCar (15578)Bicycle (501)Truck (3710)Pedestrian (8019)Mean (27808)
SC3D [12]25.0/27.117.0/18.225.7/21.914.2/16.221.8/23.1
P2B [32]27.0/29.220.0/26.421.5/16.215.9/22.022.9/25.3
BAT [50]22.5/24.117.0/18.819.3/15.817.3/24.520.5/23.0
V2B [19]31.3/35.122.2/19.121.7/19.117.3/23.425.8/29.0
STNet [20]32.2/36.121.2/29.222.3/16.819.1/27.226.9/30.8
SyncTrack32.8/36.321.7/28.323.9/19.219.3/27.127.5/31.2
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.593, + 0.719, + 0.61 + ], + "angle": 0, + "content": "4.4. Scalability of Backbone" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.816 + ], + "angle": 0, + "content": "We prove that our SyncTrack has good scalability in large-scale datasets like nuScenes. The basic model of SyncTrack has merely one Transformer layer in each stage to ensure real-time performance for tracking. However, the backbone of SyncTrack is scalable in both depth and width. We name the basic model as SyncTrack-Small. Furthermore, the SyncTrack-Mid has three Transformer layers at every stage, with nine layers total. As for the SyncTrack-Large, it doubles the number of feature channels of the SyncTrack-Mid to [256, 128, 64] for every stage. Table 5 reveals that performance improves when scalability increases not only in depth (Small v.s. Mid), but also in width (Mid v.s. Large)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.665, + 0.844 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.854, + 0.891, + 0.884 + ], + "angle": 0, + "content": "We conduct comprehensive ablations to evaluate the components of SyncTrack." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Compare with Siamese Structure. The synchronized fea" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9960" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.088, + 0.468, + 0.114 + ], + "angle": 0, + "content": "Table 5. The scalability of SyncTrack on nuScenes, (S, M, L represent small, middle, large model size of SyncTrack)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.115, + 0.468, + 0.219 + ], + "angle": 0, + "content": "
Scale#ParamFLOPsSuccessPrecision
BicycleS1.47 M2.51 G23.830.4
M1.82 M2.63 G25.033.6
L3.98 M5.37 G25.634.1
TruckS1.47 M2.51 G39.438.6
M1.82 M2.63 G40.138.8
L3.98 M5.37 G40.538.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.468, + 0.415 + ], + "angle": 0, + "content": "ture extracting and matching mechanism effectively aggregates features and models the relation. To make comparisons, we split the single branch into a Siamese structure based on SyncTrack. A matching network is added to the Siamese backbone for correlating the features. The shape-aware feature learning network in V2B [19] and iterative coarse-to-fine correlation network in STNet [20] are chosen as matching networks to compare with our single-branch framework. From Table 6, the single-branch structure of SyncTrack is quite significant, and when we split the branch and add a matcher to correlate, the performance drops heavily." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.429, + 0.468, + 0.469 + ], + "angle": 0, + "content": "Table 6. Ablations of proposed Single-Branch and Siamese Structure on Car category of KITTI and nuScenes, (matches1 is from the V2B [19] and matches2 is from the STNet [20])." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.469, + 0.466, + 0.545 + ], + "angle": 0, + "content": "
StructureKITTInuScenes
SuccessPrecisionSuccessPrecision
Siamese+catcher170.482.435.336.8
Siamese+catcher271.883.735.437.1
Single-Branch73.385.036.738.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.468, + 0.789 + ], + "angle": 0, + "content": "Attentive Sampling. In this paper, we integrate attentive sampling into the multi-head Transformers for selecting search region points to aggregate neighborhood features. We ablate such configuration by performing attentive sampling on template tokens and both template and search region tokens, as well as comparing with standard Transformer without attentive sampling (using random sampling and random/FPS sampling) as shown in Table 7. It can be concluded that the pattern of performing attentive sampling on search region tokens in Transformers is the best. We hypothesize that sampling template tokens attentively is meaningless as template points are target-centric, and search regions' responses include much noise from the background. Therefore, it is inefficient to down-sample template tokens based on attentive responses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.214, + 0.817 + ], + "angle": 0, + "content": "4.6. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In Figure 4, we present visualization results obtained from LiDAR video sequences taken from the KITTI dataset. These visualizations show the motion pattern of objects belonging to four categories within the KITTI dataset. Obviously, our SyncTrack excels in accurately tracking the in" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.088, + 0.892, + 0.112 + ], + "angle": 0, + "content": "Table 7. Ablations of attentive sampling in APST on Car category of KITTI." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.113, + 0.888, + 0.247 + ], + "angle": 0, + "content": "
TemplateSearch RegionSuccessPrecision
√APST ×randomXX70.982.8
✓X67.678.8
X✓73.385.0
✓✓69.681.1
√APST ×FPSXX71.182.8
✓X68.078.4
X✓73.285.0
✓✓70.282.7
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.366 + ], + "angle": 0, + "content": "tended target and predicting bounding boxes when compared to STNet [20]. This achievement is primarily attributed to the dynamic and abundant feature interactions that occur between the template and search region seeds in SyncTrack. These interactions enable our algorithm to effectively distinguish the foreground from the background, leading to superior performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.378, + 0.618, + 0.394 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.892, + 0.583 + ], + "angle": 0, + "content": "In this paper, we propose SyncTrack, a novel single-branch and single-stage framework for 3D LiDAR single object tracking. SyncTrack replaces the conventional Siamese-like backbones with a single-branch one, synchronizing the feature extracting and matching without an additional matching network. Moreover, the Attentive Points-Sampling Transformer is proposed for building the backbone, and sampling search region points attentively rather than randomly. Our SyncTrack achieves good tracking performance in accuracy, efficiency, and scalability. We hope it can help motivate further research on more simple yet efficient 3D trackers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.751 + ], + "angle": 0, + "content": "Limitations discussion. Compared with motion-centric tracking framework like M2Track [51], we find our SyncTrack achieves limited performance on tiny-sized and slow-moving objects like pedestrian. We attribute it to the global reasoning mechanism of self-attention. Specifically, the semantic density of small-sized objects' token is much lower, which hinders effective informative interactions between tokens when performing self-attention. The fact that transformer-based method STNet [20] outperforms CNN-based M2track on all classes except the pedestrian (Table 1) also supports this hypothesis." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.764, + 0.597, + 0.779 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[1] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In European conference on computer vision, pages 850–865. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.845, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.789, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9961" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[3] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: A simplified architecture for visual object tracking. arXiv preprint arXiv:2203.05328, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.179, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[4] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8126-8135, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.236, + 0.471, + 0.277 + ], + "angle": 0, + "content": "[5] Yubo Cui, Zheng Fang, Jiayao Shan, Zuoxu Gu, and Sifan Zhou. 3d object tracking with transformer. arXiv preprint arXiv:2110.14921, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.279, + 0.471, + 0.333 + ], + "angle": 0, + "content": "[6] Xingping Dong, Jianbing Shen, Dongming Wu, Kan Guo, Xiaogang Jin, and Fatih Porikli. Quadruplet network with one-shot learning for fast visual object tracking. IEEE Transactions on Image Processing, 28(7):3516-3527, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.335, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.42, + 0.471, + 0.475 + ], + "angle": 0, + "content": "[8] Heng Fan and Haibin Ling. Siamese cascaded region proposal networks for real-time visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7952-7961, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.478, + 0.471, + 0.532 + ], + "angle": 0, + "content": "[9] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3d-siamrpn: an end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[10] Peng Gao, Teli Ma, Hongsheng Li, Ziyi Lin, Jifeng Dai, and Yu Qiao. Mcmae: Masked convolution meets masked autoencoders. Advances in Neural Information Processing Systems, 35:35632-35644, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.47, + 0.646 + ], + "angle": 0, + "content": "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[12] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3d siamese tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1359-1368, 2019. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[13] Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. Siamcar: Siamese fully convolutional classification and regression for visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6269-6277, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[14] Sam Hare, Stuart Golodetz, Amir Saffari, Vibhav Vineet, Ming-Ming Cheng, Stephen L Hicks, and Philip HS Torr. Struck: Structured output tracking with kernels. IEEE transactions on pattern analysis and machine intelligence, 38(10):2096-2109, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Anfeng He, Chong Luo, Xinmei Tian, and Wenjun Zeng. A twofold siamese network for real-time object tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4834-4843, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 fps with deep regression networks. In European Conference on Computer Vision, pages 749-765. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[17] João F Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista. High-speed tracking with kernelized correlation filters. IEEE transactions on pattern analysis and machine intelligence, 37(3):583-596, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[18] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[19] Le Hui, Lingpeng Wang, Mingmei Cheng, Jin Xie, and Jian Yang. 3d siamese voxel-to-bev tracker for sparse point clouds. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 3, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[20] Le Hui, Lingpeng Wang, Linghua Tang, Kaihao Lan, Jin Xie, and Jian Yang. 3d siamese transformer network for single object tracking on point clouds. arXiv preprint arXiv:2207.11995, 2022. 1, 3, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.411 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[22] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4282-4291, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[23] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8971-8980, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[24] Rong Li, Anh-Quan Cao, and Raoul de Charette. Coarse3d: Class-prototypes for contrastive learning in weakly-supervised 3d point cloud segmentation. arXiv preprint arXiv:2210.01784, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[25] Yang Li and Jianke Zhu. A scale adaptive kernel correlation filter tracker with feature integration. In European conference on computer vision, pages 254-265. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.652, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[26] Junwei Liang, Lu Jiang, Kevin Murphy, Ting Yu, and Alexander Hauptmann. The garden of forking paths: Towards multi-future trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10508-10518, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[27] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5725-5734, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[28] Xiankai Lu, Chao Ma, Jianbing Shen, Xiaokang Yang, Ian Reid, and Ming-Hsuan Yang. Deep object tracking with shrinkage loss. IEEE transactions on pattern analysis and machine intelligence, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Jiageng Mao, Xiaogang Wang, and Hongsheng Li. Interpolated convolutional networks for 3d point cloud understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1578-1587, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9962" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[30] Hyeonseob Nam and Bohyung Han. Learning multi-domain convolutional neural networks for visual tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4293-4302, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.204 + ], + "angle": 0, + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[32] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2b: Point-to-box network for 3d object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6329-6338, 2020. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[33] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. Ptt: Point-track-transformer module for 3d single object tracking in point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[34] Jack Valmadre, Luca Bertinetto, Joao Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2805-2813, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[36] Mengmeng Wang, Yong Liu, and Zeyi Huang. Large margin object tracking with circulant feature maps. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4021-4029, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.593 + ], + "angle": 0, + "content": "[37] Mengmeng Wang, Teli Ma, Xingxing Zuo, Jiajun Lv, and Yong Liu. Correlation pyramid network for 3d single object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3215-3224, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[38] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1571-1580, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[39] Zhoutao Wang, Qian Xie, Yu-Kun Lai, Jing Wu, Kun Long, and Jun Wang. Mlvsnet: Multi-level voting siamese network for 3d visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3101-3110, 2021. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[41] Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8751-8760, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[42] Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, pages 12549-12556, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.16 + ], + "angle": 0, + "content": "[43] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10448-10457, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[44] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In European Conference on Computer Vision, pages 341-357. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.893, + 0.261 + ], + "angle": 0, + "content": "[45] Jesus Zarzar, Silvio Giancola, and Bernard Ghanem. Efficient bird eye view proposals for 3d siamese tracking. arXiv preprint arXiv:1903.10168, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.893, + 0.318 + ], + "angle": 0, + "content": "[46] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4591-4600, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.893, + 0.374 + ], + "angle": 0, + "content": "[47] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.893, + 0.416 + ], + "angle": 0, + "content": "[48] Haojie Zhao, Gang Yang, Dong Wang, and Huchuan Lu. Deep mutual learning for visual object tracking. Pattern Recognition, 112:107796, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[49] Shaochuan Zhao, Tianyang Xu, Xiao-Jun Wu, and Xue-Feng Zhu. Adaptive feature fusion for visual object tracking. Pattern Recognition, 111:107679, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.893, + 0.531 + ], + "angle": 0, + "content": "[50] Chaoda Zheng, Xu Yan, Jiantao Gao, Weibing Zhao, Wei Zhang, Zhen Li, and Shuguang Cui. Box-aware feature enhancement for single object tracking on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13199-13208, 2021. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.894, + 0.615 + ], + "angle": 0, + "content": "[51] Chaoda Zheng, Xu Yan, Haiming Zhang, Baoyuan Wang, Shenghui Cheng, Shuguang Cui, and Zhen Li. Beyond 3d siamese tracking: A motion-centric paradigm for 3d single object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8111-8120, 2022. 1, 2, 3, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.616, + 0.893, + 0.698 + ], + "angle": 0, + "content": "[52] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu. Pptr: Relational 3d point cloud object tracking with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8531-8540, 2022. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.893, + 0.769 + ], + "angle": 0, + "content": "[53] Jiaming Zhou, Kun-Yu Lin, Haoxin Li, and Wei-Shi Zheng. Graph-based high-order relation modeling for long-term action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8984-8993, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.77, + 0.893, + 0.84 + ], + "angle": 0, + "content": "[54] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9963" + } + ] +] \ No newline at end of file diff --git a/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/d13aa22d-32ff-439b-9d2f-7deb1ed92309_origin.pdf b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/d13aa22d-32ff-439b-9d2f-7deb1ed92309_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..deef4352aee8a88b3b773d7f9cca7c94ef12b3e6 --- /dev/null +++ b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/d13aa22d-32ff-439b-9d2f-7deb1ed92309_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0827f644fcba921f0b622ebe8a16a9db15331ea3c419f22c2ba1b37e67dd2ced +size 7566969 diff --git a/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/full.md b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..60b6ee35111c9ac9da9884993af5a0944128306c --- /dev/null +++ b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/full.md @@ -0,0 +1,370 @@ +# Synchronize Feature Extracting and Matching: A Single Branch Framework for 3D Object Tracking + +Teli Ma $^{1,2*}$ , Mengmeng Wang $^{2*}$ , Jimin Xiao $^{3}$ , Huifeng Wu $^{4}$ , Yong Liu $^{2\dagger}$ $^{1}$ The Hong Kong University of Science and Technology, Guangzhou $^{2}$ Zhejiang University + $^{3}$ Xi'an Jiaotong-Liverpool University $^{4}$ Hangzhou Dianzi University + +tmal84@connect.hkust-gz.edu.cn mengmengwang@zju.edu.cn + +jimin.xiao@xjtlu.edu.cn whf@hdu.edu.cn yongliu@iipc.zju.edu.cn + +# Abstract + +Siamese network has been a de facto benchmark framework for 3D LiDAR object tracking with a shared-parametric encoder extracting features from template and search region, respectively. This paradigm relies heavily on an additional matching network to model the cross-correlation/similarity of the template and search region. In this paper, we forsake the conventional Siamese paradigm and propose a novel single-branch framework, SyncTrack, synchronizing the feature extracting and matching to avoid forwarding encoder twice for template and search region as well as introducing extra parameters of matching network. The synchronization mechanism is based on the dynamic affinity of the Transformer, and an in-depth analysis of the relevance is provided theoretically. Moreover, based on the synchronization, we introduce a novel Attentive Points-Sampling strategy into the Transformer layers (APST), replacing the random/Farthest Points Sampling (FPS) method with sampling under the supervision of attentive relations between the template and search region. It implies connecting point-wise sampling with the feature learning, beneficial to aggregating more distinctive and geometric features for tracking with sparse points. Extensive experiments on two benchmark datasets (KITTI and NuScenes) show that SyncTrack achieves state-of-the-art performance in real-time tracking. + +# 1. Introduction + +With recent advances in autonomous driving, 3D vision tasks based on LiDAR are becoming increasingly popular in the visual community. Among these tasks, 3D LiDAR single object tracking (SOT) aims to track a specific target object in a 3D video with the knowledge of 3D bounding + +![](images/34e95910b51c9d23fd6adbd9b0e7028940a773dffc409aba5ca465a243bbc55a.jpg) + +![](images/8f82eb7be74c14ff2f728acb6c8423c38537bd8b731d1c78035d371a60030d7e.jpg) + +![](images/1dd7571ea949c6276854de24591f2dc474c069cd633a58e6ecd1948342536455.jpg) +Figure 1. The comparison of (a) Siamese network based trackers, (b) previous single-branch, two-stage framework M2Track [51] with (c) our SyncTrack, which is a single-branch and single-stage framework. + +box in the initial frame. This task meets numerous challenges, such as LiDAR point cloud sparsity, occlusions, and fast motions. + +Most existing methods [32, 50, 19, 12, 33, 52, 20, 9, 39] of 3D SOT mainly adopt a Siamese-like backbone and incorporate an additional matching network to cope with the tracking challenges as shown in Fig. 1(a). Trackers based on the Siamese-like backbone separate feature extraction of template and search region, forwarding the two kinds of features with shared model parameters, respectively. Subsequently, an extra matching network is introduced to fuse + +the extracted template and search region features to model the correlation or similarity between them. However, such a paradigm restricts the feature interaction to a post-matching network, correlating the template and search region insufficiently with merely the high-level extracted features. In other words, the matching process posterior to the encoder is incapable of modeling the relations of multi-scale features intra-backbone. Moreover, a standalone matching network results in extra model parameters and computational overheads, let alone the double forwarding process of the Siamese-backbone to extract the template and search region features. M2Track [51] proposed a motion-centric paradigm to replace the Siamese-like structure, constructing a spatial-temporal point cloud to predict the motion. However, they still rely heavily on an additional motion transformation network, which requires extra training input, to integrate the extracted template features into the search region, and another two-stage refinement network is leveraged to ensure the performance as illustrated in Fig. 1(b). Based on aforementioned problems, we ask the question: Can feature extracting and matching be conducted simultaneously in a simple way? + +The answer is Yes and the solution is implied in the dynamic global reasoning property of Transformer [35, 7, 10]. Specifically, the affinity matrix of all tokens can be constructed dynamically via continuous computation of the key and query vectors in the attention mechanism. The spatial context is aggregated using affinity to attend features. Intuitively, this affinity matrix can intrinsically serve as the matching matrix for intermediate feature interactions between the template and search region if we merge them into one input of the Transformer layers. Therefore, we propose a single-branch and single-stage framework equipped with a Transformer-based backbone instead of the conventional Siamese-like PointNet++ [31] backbone, as shown in Fig 1(c). The framework is dubbed as SyncTrack, as the Transformer backbone synchronizes the feature extracting and matching process. The SyncTrack is composed of a simple backbone and prediction head, omitting the complex matching network design and motion state estimation, depending merely on point-wise features. + +However, 3D point clouds have unique properties such as sparsity [54, 24], density variance, and implicit geometric features in data locality [29]. For example, $51\%$ samples of KITTI [11] Car category have less than 100 points [19]. These problems are further aggravated when point clouds are grouped and down-sampled to formulate multi-scale point-wise feature maps. The phenomena stresses the vitality of point clouds sampling, aiming to improve the pointwise perception efficiency with limited points. PTTR [52] proposed to sample the input point clouds before the backbone, utilizing the $L_{2}$ distance as the similarity metric for sampling. However, as down-sampling layer by layer is + +essential in tracking backbone for multi-scale feature fusion to strengthen representation, it is reasonable to consider the sampling strategy in the backbone. Therefore, we propose the attentive sampling strategy based on the attention map between the template and search region, and equip each Transformer layer with our sampling module as shown in Fig. 2(b). We name the Transformer containing Attentive Points-Sampling as APST. Specifically, the attentive response from template tokens to search region tokens is considered, as the positively respond tokens are more likely to be in the foreground and should be preserved for feature extracting. By contrast, as Fig. 3 shows, random sampling easily falls into the perceptive confusion as selected points hardly contain geometric features due to randomness. + +The main contributions of our paper can be summarized as follows: + +- We introduce a single-branch and single-stage framework for real-time 3D LiDAR SOT dubbed SyncTrack, without Siamese-like forward propagation and a standalone matching network. We ingeniously leverage the dynamic affinity characteristic of the self-attention mechanism to synchronize the feature extracting and matching. A detailed analysis is provided to explain the synchronizing mechanism. +- We propose a novel APST to build the backbone, replacing the random/FPS $^1$ down-sampling of point clouds with attentive sampling to preserve more target-relevant points, and thus improving the perceptive capability of feature extracting. +- Extensive results show that our method has achieved new state-of-the-art performance on the KITTI and NuScenes datasets in real-time tracking, up to $2.8\%$ and $1.4\%$ on mean results with a high-speed of around 45 fps. Besides, SyncTrack exhibits good scalability in both width and depth. + +# 2. Related Works + +# 2.1. 2D Visual Tracking + +Advances in 2D video [53, 27, 26] and object tracking approaches [30, 16, 36, 23, 34, 48] stimulate the development of the 3D tracking, and the methods are evolving all the time. Early 2D SOT methods mainly focus on the classifier design like structured output SVM [36, 14], correlation filter [17, 25]. With the prevalence of deep learning, end-to-end trainable trackers emerged in the visual community. The Siamese-like structure [6, 18, 15, 23, 46, 42, 28] based methods have been popular in the tracking field and + +![](images/f420362eb4b670a23493bb4274b368f8c835a6d9695f7078062b20156234e0fc.jpg) +(a) + +![](images/da9622d72d628dbe5d780a8a1bb78c8b4ac0631f2ca2728f6ba497c130ff0d6e.jpg) +(b) +Figure 2. (a). The overall framework of SyncTrack and specific illustration of the single-branch backbone, the template and search region input are concatenated for feature extraction and matching synchronously. (b). The structure of the Attentive Points-Sampling Transformer (APST), which samples search region points after multi-head attention. + +have numerous variants. SiamFC [1] is a pioneering work integrating feature correlation into a fully convolutional Siamese network for visual tracking. Subsequently, improvements like introducing detection components such as region proposal network [23, 22, 8], discriminating fore and back-ground [49], anchor-free detecting [13] etc.. Recently, vision transformers have been introduced into 2D SOT [4, 38, 43, 41] to exploit the long-range modeling of the attention mechanism to fuse the features effectively. Moreover, the one-stream network trackers based on transformers are proposed in 2D tracking. The OSTrack [44] proposes joint feature learning and relation modeling, masking a proportion of image patches to save computational overheads. SimTrack [3] concatenates the template and search input, improving the patch embedding method with a Foveal window strategy. Our work draws inspiration from the aforementioned trend in 2D vision. However, it is specifically tailored to address the challenges of the 3D Single Object Tracking (SOT) problem, taking into account the unique characteristics of point clouds. Furthermore, we analyze the success of the transformer-based single-branch backbone, attributing to the dynamic affinity of the attention mechanism, which enables the synchronization of feature extraction and matching processes. + +# 2.2.3D Visual Tracking + +In this paper, we only discuss LiDAR-based 3D object tracking. Till to now, almost all the 3D tracking methods [12, 32, 9, 45, 5, 33, 52, 50, 19, 20, 37] are based on the Siamese structure. The pioneering work in this field is the SC3D [12], which initially defines the task. SC3D employs cosine similarity to measure the resemblance between template and search region features, and incorporates shape completion during training to enhance appearance refinement. Trackers following SC3D make advancements from two perspectives. Firstly, they enhance the matching network [32, 50, 19, 52, 5, 39, 51]. For instance, MLVSNet [39] uses the CBAM module [40] to enhance the vote cluster features with both channel and spatial attention. STNet [20] employs cross- and self-attention modules to enhance the interaction between the extracted template and search region features, boosting their feature-level integration. M2Track introduces a motion-centric paradigm and motion estimation module to correlate the template and search features instead of appearance matching. All these methods depend on a standalone module to match the features. Secondly, the trackers [32, 9, 19, 39, 33] have attempted to improve the prediction head part. P2B [32] employs Hough Voting to predict the target location, and [50, 33, 39] all follow the voting strategy to make predictions. 3D-SiamRPN [9] uses an RPN head to predict the final results. LTTR [9] and V2B [19] use center-based regres + +sions to predict several object properties. However, few efforts are made in exploring the encoder/backbone of trackers as PointNet++ [31] is the feature extractor by default. In this study, we focus on the backbone design and incorporate the matching process directly into the backbone, significantly streamlining the tracker network. + +# 3. Method + +In this section, we first define the 3D SOT task in Sec. 3.1. Then, a detailed introduction of the single-branch framework is included in the Sec. 3.2. Based on the single-branch framework, how the feature extracting and matching are synchronized is elaborated in (Sec. 3.3). Moreover, we propose the Attentive Points-Sampling Transformer to build the single-branch backbone and sampling search region tokens with a strategy of attentive sampling, as shown in Sec. 3.4. The decoder head and losses are described in Sec. 3.5. + +# 3.1. Problem Definition + +In the configuration of 3D LiDAR single object detection (SOT) task, the 3D bounding box (BBox) is defined as $(x,y,z,w,l,h,\theta)\in \mathbb{R}^7$ , where the $(x,y,z)$ represents the coordinate center of the BBox and $(w,l,h),\theta$ stand for the BBox size and heading angle (the rotation around the up-axis) respectively. Generally, the BBox size is assumed to be fixed by default even when the target object is non-rigid, thus minimizing the dimensions of BBox from $\mathbb{R}^7$ to $\mathbb{R}^4$ . Given a sequence of temporally-connected point clouds $\{\mathcal{P}_i\}_{i = 1}^T$ ( $T$ is the number of points in each frame) and a initial BBox $\mathcal{B}_1$ of the target, the goal of SOT is to localize the target BBoxes $\{\mathcal{B}_i\}_{i = 2}^T$ in all frames online. Following the previous manner, a template point cloud $\mathcal{P}^t = \{p_i^t\}_{i = 1}^{N_t}$ and a search region $\mathcal{P}^s = \{p_i^s\}_{i = 1}^{N_s}$ are generated, where $N_{t}$ and $N_{s}$ are number of template and search region points. The template $P^t$ is generated by cropping and centering the target in the initial frame based on the initial BBox. + +# 3.2. Single-Branch Structure + +We propose to replace the conventional Siamese-like backbone paradigm with a sole backbone, eliminating the double forward process of Siamese structure. Therefore, template and search region seeds are concatenated to do forward propagation. The Transformers' property of long-range relation-modeling is the intrinsic merit to tackle the concatenated template and search region seeds. Based on that, we leverage the self-attention modules to build the single-branch backbone as shown in Fig. 2(a). In our approach, we utilize a Query & Group module to sample the template seeds, denoted as $\mathcal{P}^t\in \mathbb{R}^{N_t\times C}$ , using the FPS method before jointly forwarding. This module also groups the k-nearest points to aggregate features. However, when + +it comes to the search region point cloud, represented as $\mathcal{P}^s\in \mathbb{R}^{N_s\times C}$ , we solely employ the Group module to aggregate neighbor information without reducing the number of search points. Subsequently, the template and search seeds are concatenated, incorporating a joint parametric positional embedding for the localization of tokens. This process is as follows: + +$$ +\mathcal {T} ^ {t} = \mathrm {k N N} (\operatorname {F P S} (\mathcal {P} ^ {t})), \mathcal {P} ^ {t} \in \mathbb {R} ^ {N _ {t} ^ {\prime} \times C ^ {\prime}}, N _ {t} ^ {\prime} < N _ {t}, +$$ + +$$ +\mathcal {T} ^ {s} = \mathrm {k N N} \left(\mathcal {P} ^ {s}\right), \mathcal {P} ^ {s} \in \mathbb {R} ^ {N _ {s} \times C ^ {\prime}}, \tag {1} +$$ + +$$ +\mathcal {T} ^ {t s} = \left[ \mathcal {T} ^ {t}; \mathcal {T} ^ {s} \right] + \mathbf {p _ {e}}, \mathcal {T} ^ {t s} \in \mathbb {R} ^ {\left(N _ {t} + N _ {s}\right) \times C ^ {\prime}}. +$$ + +Afterward, linear layers are leveraged to project the input tokens into query, key, and value latent and the head-wise joint attention map is calculated to model the intra- & interrelations of template and search tokens. + +$$ +Q ^ {t s}, K ^ {t s}, V ^ {t s} = W _ {q} (\mathcal {T} ^ {t s}), W _ {k} (\mathcal {T} ^ {t s}), W _ {v} (\mathcal {T} ^ {t s}), +$$ + +$$ +\mathcal {A} _ {m} ^ {t s} = \operatorname {S o f t m a x} \frac {Q _ {m} ^ {t s} \left(K _ {m} ^ {t s}\right) ^ {\top}}{\sqrt {C ^ {\prime} / M}}. \tag {2} +$$ + +Based on the joint head-wise attention map $\mathcal{A}_m^{ts}$ of template and search tokens, the features extracted by multi-head attention are: + +$$ +\mathcal {T} ^ {t s} = \left[ \mathcal {A} _ {1} ^ {t s} V _ {1} ^ {t s}, \mathcal {A} _ {2} ^ {t s} V _ {2} ^ {t s}, \dots , \mathcal {A} _ {M} ^ {t s} V _ {M} ^ {t s} \right] \mathcal {W} + \mathcal {T} ^ {t s}, \tag {3} +$$ + +where $M$ is the number of the attention heads and $\mathcal{W}$ is the weight of a MLP. + +# 3.3. Synchronize Feature Extracting and Matching + +We illustrate how the single-branch structure can synchronize the process of feature extracting and matching with a simple backbone. The formula of joint attention $\mathcal{A}_m^{ts}$ in Eq. 2 can be expanded as ( $\sigma$ represents Softmax operation): + +$$ +\begin{array}{l} \left[ \mathcal {A} _ {m} ^ {t}; \mathcal {A} _ {m} ^ {s} \right] = \sigma \frac {\left[ Q _ {m} ^ {t} ; Q _ {m} ^ {s} \right] \left[ K _ {m} ^ {t} ; K _ {m} ^ {s} \right] ^ {\top}}{\sqrt {C ^ {\prime} / M}} \\ = \left[ \boldsymbol {\sigma} \frac {Q _ {m} ^ {t} \left(K _ {m} ^ {t}\right) ^ {\top}}{\sqrt {C ^ {\prime} / M}}, \boldsymbol {\sigma} \frac {Q _ {m} ^ {t} \left(K _ {m} ^ {s}\right) ^ {\top}}{\sqrt {C ^ {\prime} / M}}; \boldsymbol {\sigma} \frac {Q _ {m} ^ {s} \left(K _ {m} ^ {t}\right) ^ {\top}}{\sqrt {C ^ {\prime} / M}}, \boldsymbol {\sigma} \frac {Q _ {m} ^ {s} \left(K _ {m} ^ {s}\right) ^ {\top}}{\sqrt {C ^ {\prime} / M}} \right]. \tag {4} \\ \end{array} +$$ + +The new extracted search region features can be obtained based on Eq 4 as: + +$$ +\begin{array}{l} \mathcal {T} _ {m} ^ {\prime s} = \sigma \frac {Q _ {m} ^ {s} \left(K _ {m} ^ {t}\right) ^ {\top}}{\sqrt {C ^ {\prime} / M}} V _ {m} ^ {t} + \sigma \frac {Q _ {m} ^ {s} \left(K _ {m} ^ {s}\right) ^ {\top}}{\sqrt {C ^ {\prime} / M}} V _ {m} ^ {s}, \tag {5} \\ [ V _ {m} ^ {t}; V _ {m} ^ {s} ] = V _ {m} ^ {t s}, \\ \end{array} +$$ + +depending on the projected features $[V_m^t,V_m^s ]$ of both the template and search region from last layer. The attention queried from search region to template, $\sigma \frac{Q_m^s(K_m^t)^\top}{\sqrt{C' / M}}$ , is the + +![](images/26d8e974add66a293459be24a512e840c37ae2966bead8c0ac7ef90ab62195cd.jpg) +Template + +![](images/8efe66a72fe2bef8d12dac791820c82d1e9149f4c63fe5fcdcd88221b0dc90ef.jpg) +Random Sampling + +![](images/30d056bab8c6fb337463f32f1cfb4f668bfda65fae82b3f2904df11ead0cee6e.jpg) +Attentive Sampling +Figure 3. Comparisons of random sampling with our attentive sampling in point-wise down-sampling. We find that our attentive sampling approach excels in selecting more representative points, including those that contribute to the formation of the head, arms, and legs, which contain distinctive geometric features. In contrast, random sampling include more points that are far away from the target (we have removed such points for better visualization). + +matching matrix that guides to aggregate highly-relevant template features. Moreover, the matching matrix is dynamic, which is determined by the changing query and key latent of search and template features as: + +$$ +\boldsymbol {\sigma} \frac {Q _ {m} ^ {s} \left(K _ {m} ^ {t}\right) ^ {\top}}{\sqrt {C ^ {\prime} / M}} = \boldsymbol {\sigma} \frac {W _ {q , m} ^ {s} \mathcal {T} _ {m} ^ {s} \left(\mathcal {T} _ {m} ^ {t}\right) ^ {\top} \left(W _ {k , m} ^ {t}\right) ^ {\top}}{\sqrt {C ^ {\prime} / M}}, \tag {6} +$$ + +$$ +i. e. \qquad \sigma \frac {Q _ {m} ^ {s} (K _ {m} ^ {t}) ^ {\top}}{\sqrt {C ^ {\prime} / M}} \propto [ \mathcal {T} _ {m} ^ {s}; \mathcal {T} _ {m} ^ {t} ]. +$$ + +To conclude, the synchronization attributes to the dynamic mechanism of matching matrix, continuously adapting the matching relations according to the extracted features of template and search region. + +Comparisons with Siamese Network. Previous Siamese-like trackers can be summarized as the paradigm of Extracting then Matching. If we name the feature extracting and matching as $\phi$ and $\delta$ , then the objective of this paradigm can be concluded as $max(\delta(\phi(\mathcal{P}^t), \phi(\mathcal{P}^s))$ for simplicity. It means the model training is to shorten the distances between correlated parts of the template and search region based on the extracted features from the backbone. However, compared to our approach, this matching mechanism is relatively static since it occurs only after feature extraction, resulting in inadequate modeling of inter-backbone relations. On the other hand, our single-branch framework facilitates dynamic interaction between the search region seeds and the template across all layers of the backbone. This allows for comprehensive learning of relations, encompassing both the local representation from early layers and the global representation from later ones. + +# 3.4. Attentive Points-Sampling Transformer + +Based on the synchronizing mechanism of feature extracting and matching, we propose to replace the Point-wise Transformer [47] with a novel Attentive Points-Sampling Transformer (APST). APST is proposed based on the observation that backbones utilized in previous 3D LiDAR trackers always adopt a non-parametric strategy of down-sampling the search region features/tokens with farthest point sampling (FPS) or randomly sampling to reduce the points as introduced in PointNet++ [31]. Nevertheless, this non-parametric sampling method lacks learnability and controllability since no parameters associated with the sampling process are updated during model training. Consequently, the final performance is compromised due to the significant influence of the points-center in LiDAR, which determines the effectiveness of extracting features around the foreground, as shown in Fig. 3. + +Therefore, we introduce the APST, which involves the selection of points based on attentive relations between the tokens of the template and search region as illustrated in sub-figure (b) of Fig. 2. The attentive response of template tokens reacting to search region tokens is considered as positively responding search tokens are more likely to be in the foreground. In that case, we segment the attention map and separate the $\mathcal{A}_m^{t\rightarrow s} = \sigma \frac{Q_m^t(K_m^s)^{\top}}{\sqrt{C' / M}}$ from the Eq. 4, averaging along the dimension of template-tokens attention among all transformer heads to acquire the response scores. To ensure the maximum response of search tokens to template ones, a group of search region indexes, named $\Omega_s$ , is dynamically updated as: + +$$ +\Omega_ {s} ^ {*} = \arg \max _ {\Omega_ {s}} \frac {1}{N _ {t}} \frac {1}{M} \sum_ {t = 0} ^ {N _ {t}} \sum_ {m = 0} ^ {M} \frac {Q _ {m} ^ {t}}{\sqrt {C ^ {\prime} / M}} \left(K _ {m} ^ {s}\right) ^ {\top} \Big | _ {s \in \Omega_ {s}}, \tag {7} +$$ + +Based on the optimal solution $\Omega_s^*$ , the search region tokens are sampled to decrease the number of tokens after multi-head self-attention and concatenate with template ones. Note that the template tokens are sampled with the FPS method as introduced in Sec. 3.2, and only the search region tokens are sampled with the attentive sampling method. + +The dynamic-affinity property of the Transformer suggests that the attention map is influenced by the latent tokens projected through learnable linear layers. Hence, it is reasonable to assert that sampling tokens guided by the attention map is linked to model learning, as the attention map is generated based on updated parameters. As a result, this type of token/point sampling proves advantageous in providing prior knowledge for centroid selection and enhancing the efficient aggregation of representational information, as shown in Fig. 3. + +# 3.5. Decoder and Losses + +With the encoded point-wise features of various scales, a multi-scale feature fusion module is adopted to fuse these features and output the features with the same origin input size, feeding into the decoder part for final predictions. Following the V2B [19], the features are voxelized as a volumetric representation and 3D convolutions are utilized on the encoded features. Afterward, the BEV feature maps are acquired by pooling on the $z$ -axis for regression. Focal loss and L1 loss are leveraged for classification and BBox center offset and rotation regression, respectively. A detailed introduction is presented in the supplementary. + +# 4. Experiments + +# 4.1. Experiment Setups + +Implementation Details. We set the number of input points as $N_{t} = 512$ and $N_{s} = 1024$ for template and search regions by randomly duplicating and discarding, respectively. The encoder backbone is merely consisted of three layers with Attentive Points-Sampling Transformers, and the number of both template and search region points output by each layer are 256, 128 and 64. Moreover, each layer's feature dimensions in the encoder are 32, 64 and 128, respectively, whereas the final features for the prediction head are with 32 channels. Note that the heads for all APST are 2 by default. In the voxelization process, the region $[(x_{min}, x_{max}), (y_{min}, y_{max}), (z_{min}, z_{max})]$ is defined as $[(-5.6, 5.6), (-3.6, 3.6), (-2.4, 2.4)]$ to contain most target points. The voxel size $(v_{x}, v_{y}, v_{z})$ is set to $(0.3, 0.3, 0.3)$ . For the detection head, four decomposed 3D (stride of 2,1,2,1 along the $z$ -axis) and 2D convolution blocks (stride of 2,1,1,2) are leveraged to strengthen the feature aggregation. + +Training and Testing. We train the model for 40 epochs with a batch size of 64. The Adam optimizer [21] is adopted with the initial learning rate of 0.001 and reduced by 5 every 10 epochs (every 2 epochs for nuScenes). The classification loss has a weight $\lambda_{cls}$ of 1 and the regression loss has a weight $\lambda_{reg}$ of 1. + +Evaluation Metrics. Following the previous methods [32, 50], we measure the Success and Precision of the tracker. To be specific, Success is defined as the IoU between predicted boxes and the ground truth, and Precision measures the AUC (Area Under Curve) of the distance between predicted and ground truth boxes within the range of [0,2] meters. + +# 4.2. Comparison with State-of-the-Art Trackers + +Results on KITTI. KITTI [11] is one of the most popular datasets used in mobile robotics and autonomous driving. The tracking benchmark of KITTI consists of 21 training sequences and 29 test sequences. Following the previous + +methods [12, 51, 50], we split the training sequences into train/val/test splits due to the inaccessibility of the testing labels, scenes 0-16 for training, scenes 7-18 for validation and scenes 19-20 for testing, respectively. + +We compare the SyncTrack with other state-of-the-art methods from the pioneering SC3D [12] to the most recent siamese network STNet [20] and single branch network M2Track [51], as shown in Table 1. We separate the trackers into Siamese Network and Single Branch Network categories to compare with our proposed SyncTrack. Compared with siamese structured trackers, the SyncTrack achieves the best results on rigid and non-rigid object tracking, outperforming current tracking methods based on siamese networks on most specific categories and the overall mean results. STNet [20] is the state-of-the-art siamese-based tracking method with self-attention and cross-attention modules to match the template and search-region features. Our SyncTrack outperforms the STNet by a relatively large margin on Car, Van and Pedestrian categories under the evaluation of Success metric. Also, SyncTrack surpasses the previous best Mean results by $2.8\%$ and $1.8\%$ on Success and Precision metrics, respectively. + +We also compare the SyncTrack with the only current single-branch tracker M2Track [51]. Our SyncTrack outperforms M2Track by a large margin, up to $7.8\%$ in Success in the category of Car whereas the M2Track is better in the category of Pedestrian. However, for the comprehensive evaluation, the Mean performance of all frames, our SyncTrack outperforms the M2Track by $1.2\%$ on the Success metric. + +Results on nuScenes. The nuScenes dataset [2] contains 1000 driving scenes collected from Boston and Singapore with a diverse and exciting set of driving maneuvers, traffic situations, and unexpected behaviors. In the configurations of LiDAR-based tracking methods, the train/val/test sets make up 700/150/150 of the whole 1000 scenes, respectively. Officially, the train set is evenly split into 'train track' and 'train detect' to remedy overfitting. Following [19], we train our model with "train track" split and test it on the val set. + +Note that nuScenes dataset only annotates keyframes and provides official interpolated results for the remaining frames, so there are two configurations for this dataset. The first is from [50, 51] which trains and tests both only on the keyframes. The second one is from [19, 20], which trains and tests on all the frames. The results based on these two configurations are different. We believe that the motion in key frames is substantial, which does not conform to the practical applications. Therefore, we train and test all the frames in this paper. We train the previous methods on the nuScenes by ourselves using their official codes to compare with our SyncTrack as many results are missing or only reported by testing nuScenes test-split with a pre + +Table 1. Success/Precision comparisons among our SyncTrack and the state-of-the-art methods on the KITTI datasets. Mean shows frame-level averaging results. Bold and underline denote the best performance and the second-best performance, respectively. Improvements over previous state-of-the-arts are shown in Italic and color. + +
MethodsCar (6424)Cyclist (308)Van (1248)Pedestrian (6088)Mean (14068)
SuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecision
Siamese Network
SC3D [12]41.357.941.570.440.447.018.237.831.248.5
SC3D-RPN[45]36.351.043.081.4--17.947.8--
P2B [32]56.272.832.144.740.848.428.749.642.460.0
MLVSNet [39]56.074.034.344.552.061.434.161.145.766.6
3DSiamRPN [9]58.276.236.149.045.652.835.256.246.664.9
LTTR [5]65.077.166.289.935.845.633.256.848.765.8
PTT [33]67.881.837.247.343.652.544.972.055.174.2
BAT [50]60.577.733.745.452.467.042.170.151.272.8
V2B [19]70.581.340.849.750.158.048.373.558.475.2
PTTR [52]65.277.465.190.552.561.850.981.658.477.8
STNet [20]72.184.073.593.758.070.649.977.261.380.1
SyncTrack73.385.073.193.860.370.054.780.564.181.9
improvement+1.2+1.0-0.4+0.1+2.3-0.6+3.8-1.1+2.8+1.8
Single Branch Network
M2Track [51]65.580.873.293.553.870.761.588.262.983.4
SyncTrack73.385.073.193.860.370.054.780.564.181.9
improvement+7.8+4.2-0.1+0.3+2.3-0.7-6.8-7.7+1.2-1.5
+ +Table 2. Comparison among SyncTrack and the state-of-the-art methods on the nuScenes datasets. Mean shows the average result weighed by frame numbers. Bold and underline denote the best performance and the second-best performance, respectively. Improvements over previous state-of-the-arts are shown in Italic and color. + +
MethodsCar (15578)Bicycle (501)Truck (3710)Pedestrian (8019)Mean (27808)
SuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecision
SC3D [12]24.525.916.618.832.530.613.814.722.323.2
P2B [32]32.835.219.726.616.211.119.226.626.429.3
BAT [50]26.528.817.822.816.510.619.428.223.027.9
V2B [19]32.934.520.327.528.723.820.127.428.430.9
STNet [20]35.737.222.329.333.532.420.127.830.733.7
M2Track [51]31.433.922.629.830.128.820.728.028.031.4
SyncTrack36.738.123.830.439.438.619.127.831.835.1
improvement+1.0+0.9+1.2+0.6+5.9+6.2-1.6-0.4+1.1+1.4
+ +trained KITTI model. + +As shown in Table 2, SyncTrack performs significantly better than other trackers on the mean results of four categories. Specifically, the SyncTrack yields the best results on both metrics and on most categories except the Pedestrian, which is lower than M2Track [51] and BAT [50] by a minor margin (1.6% and 0.4%). However, in the Truck category, our SyncTrack outperforms state-of-the-art by a large margin, up to 5.9% and 6.2% on Success and Precision respectively. + +Computational Cost Comparison. We analyze the computational overheads and inference speed of SyncTrack and compare it with other trackers in Table 3. The reported results are tested by ourselves with official codebases with a single Titan RTX GPU on Car category of KITTI. It can be seen that SyncTrack achieves the best Success performance with the lowest computational complexity (2.51 + +G). Compared with the most current Siamese-based tracker STNet [20] and single-branch tracker M2Track, our SyncTrack has the fewest number of parameters and fastest inference speed, satisfying the demand of real-time tracking. + +Table 3. The computational cost of different trackers. + +
MethodsParametersFLOPsFPSSuccess
SC3D [12]6.45 M20.07 G641.3
P2B [32]1.34 M4.28 G4856.2
BAT [50]1.47 M5.53 G5460.5
V2B [19]1.36 M5.57 G3970.5
STNet [20]1.66 M3.14 G3672.1
M2Track [51]2.24 M2.54 G3765.5
SyncTrack1.47 M2.51 G4573.3
+ +![](images/5801c758bbe5b150f08e940d7f5d216476ca9b0b6cd0b70a58ab08361ec1058b.jpg) +C + +![](images/0d9921175ab92df7386da8408bddcaad1b13c1000f55f515b1239cd7b5234897.jpg) + +![](images/23375ee6f4e24ce85bf2143aa1392ae979c4d64a5168b22558f1bdc488fb5399.jpg) + +![](images/e83a98dca80efba7d737f083db9c9ed48c3f96aa3b3a7b1be8a0cad3dc7c0524.jpg) + +![](images/75df7b179aaeb778fd509c2e9ca0ae0e01809b048263226b2a992d0facaa6dcc.jpg) + +![](images/c400235239773bea98973b17f2d604ef767ebd031a0c6b83cfed6d6b667d6c5d.jpg) +Cyclist + +![](images/b93847361ea53ab19398a7698ce710e1c05df01192ac06fd04b94dc2b1ab8386.jpg) + +![](images/853a7293570efed7d2ad5eec3764433c6ef52ed4b68d995aa6c0bade13ab4c46.jpg) + +![](images/d00c77b875d7d2dd42fd7966042455ad3d2aab82ab997233f52b9e3873df9728.jpg) + +![](images/1fe662d6f998f80de7f73932e75985ad017ac0c4d2d2ff661ce555dc3b63dd57.jpg) + +![](images/21c25b9b289b0a8498af9ec8349c22908c4bc19bf9664b7d4a693eba4646f901.jpg) +A + +![](images/24f8cfb64039ec8d215784e527b78f06dfd693e808d78015208293e78eb05032.jpg) + +![](images/baccd366d1e5ad84995dc518dc5a517f0c83468dbed80800406db066377b7f26.jpg) + +![](images/c1560e93f0f00465798e7fc131758ad4791d1987a144fa901e4e45b435028a0a.jpg) + +![](images/04e23faea3658d040771c416ea59fba8cc94bc2724b347cea12e93618a6eb85b.jpg) + +![](images/8a2add9e7555554f8a6db3acc16e4031d4ee2c0a12054349b12f20476508c780.jpg) +Peperrnnnne + +![](images/8c10da3371a8b869ac70f1398ecfc42b8d55948c54d76723cd51565fe2552479.jpg) +Ground Truth + +![](images/97ea566193899caa919442eb0c5781062fcdbf3b85e57e5c8896738de9f8c725.jpg) +STNet + +![](images/1c364da645eae84d1be3f235bf6fb37a1e1539f2b2d5691d668b569449d9ec4b.jpg) +SyncTrack + +![](images/a1f32a3afcf2b1d4751d7b10dbd7ac31297f3531b042d2ac04bec885d9fa26a6.jpg) + +![](images/a529280928d04447fab05c3a53a3f3f8d050cb88e980d9a83cc93ddffc477169.jpg) + +![](images/4aa7209e962675ca066240ef85d094fc591ac4969422355ad49c59a0bde47b58.jpg) +Figure 4. Qualitative comparison with STNet. We compare SyncTrack with STNet over four tracking sequences of different categories in KITTI. The predictions of SyncTrack fits the ground-truth boxes better. + +![](images/a3bd6151f15a312ed904c1966f7a420e28b60d9f89c4f0c64d8ba44ad7b6bf4b.jpg) + +# 4.3. Generalization Ability + +To evaluate the generalization ability of SyncTrack, we pre-train the model on the KITTI dataset and test it directly on the nuScenes dataset without fine-tuning. The results are shown in Table 4. It can be observed that SyncTrack outperforms other methods on the mean results of four categories by a large margin. SyncTrack not only achieves a good balance between inference speed and tracking accuracy, but also generalizes very well to new domains. + +Table 4. Testing results on the nuScenes for generalization ability. + +
MethodsCar (15578)Bicycle (501)Truck (3710)Pedestrian (8019)Mean (27808)
SC3D [12]25.0/27.117.0/18.225.7/21.914.2/16.221.8/23.1
P2B [32]27.0/29.220.0/26.421.5/16.215.9/22.022.9/25.3
BAT [50]22.5/24.117.0/18.819.3/15.817.3/24.520.5/23.0
V2B [19]31.3/35.122.2/19.121.7/19.117.3/23.425.8/29.0
STNet [20]32.2/36.121.2/29.222.3/16.819.1/27.226.9/30.8
SyncTrack32.8/36.321.7/28.323.9/19.219.3/27.127.5/31.2
+ +# 4.4. Scalability of Backbone + +We prove that our SyncTrack has good scalability in large-scale datasets like nuScenes. The basic model of SyncTrack has merely one Transformer layer in each stage to ensure real-time performance for tracking. However, the backbone of SyncTrack is scalable in both depth and width. We name the basic model as SyncTrack-Small. Furthermore, the SyncTrack-Mid has three Transformer layers at every stage, with nine layers total. As for the SyncTrack-Large, it doubles the number of feature channels of the SyncTrack-Mid to [256, 128, 64] for every stage. Table 5 reveals that performance improves when scalability increases not only in depth (Small v.s. Mid), but also in width (Mid v.s. Large). + +# 4.5. Ablation Studies + +We conduct comprehensive ablations to evaluate the components of SyncTrack. + +Compare with Siamese Structure. The synchronized fea + +Table 5. The scalability of SyncTrack on nuScenes, (S, M, L represent small, middle, large model size of SyncTrack). + +
Scale#ParamFLOPsSuccessPrecision
BicycleS1.47 M2.51 G23.830.4
M1.82 M2.63 G25.033.6
L3.98 M5.37 G25.634.1
TruckS1.47 M2.51 G39.438.6
M1.82 M2.63 G40.138.8
L3.98 M5.37 G40.538.9
+ +ture extracting and matching mechanism effectively aggregates features and models the relation. To make comparisons, we split the single branch into a Siamese structure based on SyncTrack. A matching network is added to the Siamese backbone for correlating the features. The shape-aware feature learning network in V2B [19] and iterative coarse-to-fine correlation network in STNet [20] are chosen as matching networks to compare with our single-branch framework. From Table 6, the single-branch structure of SyncTrack is quite significant, and when we split the branch and add a matcher to correlate, the performance drops heavily. + +Table 6. Ablations of proposed Single-Branch and Siamese Structure on Car category of KITTI and nuScenes, (matches1 is from the V2B [19] and matches2 is from the STNet [20]). + +
StructureKITTInuScenes
SuccessPrecisionSuccessPrecision
Siamese+catcher170.482.435.336.8
Siamese+catcher271.883.735.437.1
Single-Branch73.385.036.738.1
+ +Attentive Sampling. In this paper, we integrate attentive sampling into the multi-head Transformers for selecting search region points to aggregate neighborhood features. We ablate such configuration by performing attentive sampling on template tokens and both template and search region tokens, as well as comparing with standard Transformer without attentive sampling (using random sampling and random/FPS sampling) as shown in Table 7. It can be concluded that the pattern of performing attentive sampling on search region tokens in Transformers is the best. We hypothesize that sampling template tokens attentively is meaningless as template points are target-centric, and search regions' responses include much noise from the background. Therefore, it is inefficient to down-sample template tokens based on attentive responses. + +# 4.6. Visualization + +In Figure 4, we present visualization results obtained from LiDAR video sequences taken from the KITTI dataset. These visualizations show the motion pattern of objects belonging to four categories within the KITTI dataset. Obviously, our SyncTrack excels in accurately tracking the in + +Table 7. Ablations of attentive sampling in APST on Car category of KITTI. + +
TemplateSearch RegionSuccessPrecision
√APST ×randomXX70.982.8
✓X67.678.8
X✓73.385.0
✓✓69.681.1
√APST ×FPSXX71.182.8
✓X68.078.4
X✓73.285.0
✓✓70.282.7
+ +tended target and predicting bounding boxes when compared to STNet [20]. This achievement is primarily attributed to the dynamic and abundant feature interactions that occur between the template and search region seeds in SyncTrack. These interactions enable our algorithm to effectively distinguish the foreground from the background, leading to superior performance. + +# 5. Conclusion + +In this paper, we propose SyncTrack, a novel single-branch and single-stage framework for 3D LiDAR single object tracking. SyncTrack replaces the conventional Siamese-like backbones with a single-branch one, synchronizing the feature extracting and matching without an additional matching network. Moreover, the Attentive Points-Sampling Transformer is proposed for building the backbone, and sampling search region points attentively rather than randomly. Our SyncTrack achieves good tracking performance in accuracy, efficiency, and scalability. We hope it can help motivate further research on more simple yet efficient 3D trackers. + +Limitations discussion. Compared with motion-centric tracking framework like M2Track [51], we find our SyncTrack achieves limited performance on tiny-sized and slow-moving objects like pedestrian. We attribute it to the global reasoning mechanism of self-attention. Specifically, the semantic density of small-sized objects' token is much lower, which hinders effective informative interactions between tokens when performing self-attention. The fact that transformer-based method STNet [20] outperforms CNN-based M2track on all classes except the pedestrian (Table 1) also supports this hypothesis. + +# References + +[1] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In European conference on computer vision, pages 850–865. Springer, 2016. 3 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6 +[3] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: A simplified architecture for visual object tracking. arXiv preprint arXiv:2203.05328, 2022. 3 +[4] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8126-8135, 2021. 3 +[5] Yubo Cui, Zheng Fang, Jiayao Shan, Zuoxu Gu, and Sifan Zhou. 3d object tracking with transformer. arXiv preprint arXiv:2110.14921, 2021. 3, 7 +[6] Xingping Dong, Jianbing Shen, Dongming Wu, Kan Guo, Xiaogang Jin, and Fatih Porikli. Quadruplet network with one-shot learning for fast visual object tracking. IEEE Transactions on Image Processing, 28(7):3516-3527, 2019. 2 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[8] Heng Fan and Haibin Ling. Siamese cascaded region proposal networks for real-time visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7952-7961, 2019. 3 +[9] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3d-siamrpn: an end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 1, 3, 7 +[10] Peng Gao, Teli Ma, Hongsheng Li, Ziyi Lin, Jifeng Dai, and Yu Qiao. Mcmae: Masked convolution meets masked autoencoders. Advances in Neural Information Processing Systems, 35:35632-35644, 2022. 2 +[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 6 +[12] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3d siamese tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1359-1368, 2019. 1, 3, 6, 7, 8 +[13] Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. Siamcar: Siamese fully convolutional classification and regression for visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6269-6277, 2020. 3 +[14] Sam Hare, Stuart Golodetz, Amir Saffari, Vibhav Vineet, Ming-Ming Cheng, Stephen L Hicks, and Philip HS Torr. Struck: Structured output tracking with kernels. IEEE transactions on pattern analysis and machine intelligence, 38(10):2096-2109, 2015. 2 +[15] Anfeng He, Chong Luo, Xinmei Tian, and Wenjun Zeng. A twofold siamese network for real-time object tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4834-4843, 2018. 2 + +[16] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 fps with deep regression networks. In European Conference on Computer Vision, pages 749-765. Springer, 2016. 2 +[17] João F Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista. High-speed tracking with kernelized correlation filters. IEEE transactions on pattern analysis and machine intelligence, 37(3):583-596, 2014. 2 +[18] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 2 +[19] Le Hui, Lingpeng Wang, Mingmei Cheng, Jin Xie, and Jian Yang. 3d siamese voxel-to-bev tracker for sparse point clouds. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 3, 6, 7, 8, 9 +[20] Le Hui, Lingpeng Wang, Linghua Tang, Kaihao Lan, Jin Xie, and Jian Yang. 3d siamese transformer network for single object tracking on point clouds. arXiv preprint arXiv:2207.11995, 2022. 1, 3, 6, 7, 8, 9 +[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[22] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4282-4291, 2019. 3 +[23] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8971-8980, 2018. 2, 3 +[24] Rong Li, Anh-Quan Cao, and Raoul de Charette. Coarse3d: Class-prototypes for contrastive learning in weakly-supervised 3d point cloud segmentation. arXiv preprint arXiv:2210.01784, 2022. 2 +[25] Yang Li and Jianke Zhu. A scale adaptive kernel correlation filter tracker with feature integration. In European conference on computer vision, pages 254-265. Springer, 2014. 2 +[26] Junwei Liang, Lu Jiang, Kevin Murphy, Ting Yu, and Alexander Hauptmann. The garden of forking paths: Towards multi-future trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10508-10518, 2020. 2 +[27] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5725-5734, 2019. 2 +[28] Xiankai Lu, Chao Ma, Jianbing Shen, Xiaokang Yang, Ian Reid, and Ming-Hsuan Yang. Deep object tracking with shrinkage loss. IEEE transactions on pattern analysis and machine intelligence, 2020. 2 +[29] Jiageng Mao, Xiaogang Wang, and Hongsheng Li. Interpolated convolutional networks for 3d point cloud understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1578-1587, 2019. 2 + +[30] Hyeonseob Nam and Bohyung Han. Learning multi-domain convolutional neural networks for visual tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4293-4302, 2016. 2 +[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4, 5 +[32] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2b: Point-to-box network for 3d object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6329-6338, 2020. 1, 3, 6, 7, 8 +[33] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. Ptt: Point-track-transformer module for 3d single object tracking in point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 1, 3, 7 +[34] Jack Valmadre, Luca Bertinetto, Joao Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2805-2813, 2017. 2 +[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[36] Mengmeng Wang, Yong Liu, and Zeyi Huang. Large margin object tracking with circulant feature maps. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4021-4029, 2017. 2 +[37] Mengmeng Wang, Teli Ma, Xingxing Zuo, Jiajun Lv, and Yong Liu. Correlation pyramid network for 3d single object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3215-3224, 2023. 3 +[38] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1571-1580, 2021. 3 +[39] Zhoutao Wang, Qian Xie, Yu-Kun Lai, Jing Wu, Kun Long, and Jun Wang. Mlvsnet: Multi-level voting siamese network for 3d visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3101-3110, 2021. 1, 3, 7 +[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3 +[41] Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8751-8760, 2022. 3 +[42] Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, pages 12549-12556, 2020. 2 + +[43] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10448-10457, 2021. 3 +[44] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In European Conference on Computer Vision, pages 341-357. Springer, 2022. 3 +[45] Jesus Zarzar, Silvio Giancola, and Bernard Ghanem. Efficient bird eye view proposals for 3d siamese tracking. arXiv preprint arXiv:1903.10168, 2019. 3, 7 +[46] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4591-4600, 2019. 2 +[47] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 5 +[48] Haojie Zhao, Gang Yang, Dong Wang, and Huchuan Lu. Deep mutual learning for visual object tracking. Pattern Recognition, 112:107796, 2021. 2 +[49] Shaochuan Zhao, Tianyang Xu, Xiao-Jun Wu, and Xue-Feng Zhu. Adaptive feature fusion for visual object tracking. Pattern Recognition, 111:107679, 2021. 3 +[50] Chaoda Zheng, Xu Yan, Jiantao Gao, Weibing Zhao, Wei Zhang, Zhen Li, and Shuguang Cui. Box-aware feature enhancement for single object tracking on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13199-13208, 2021. 1, 3, 6, 7, 8 +[51] Chaoda Zheng, Xu Yan, Haiming Zhang, Baoyuan Wang, Shenghui Cheng, Shuguang Cui, and Zhen Li. Beyond 3d siamese tracking: A motion-centric paradigm for 3d single object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8111-8120, 2022. 1, 2, 3, 6, 7, 9 +[52] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu. Pptr: Relational 3d point cloud object tracking with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8531-8540, 2022. 1, 2, 3, 7 +[53] Jiaming Zhou, Kun-Yu Lin, Haoxin Li, and Wei-Shi Zheng. Graph-based high-order relation modeling for long-term action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8984-8993, 2021. 2 +[54] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021. 2 \ No newline at end of file diff --git a/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/images.zip b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9bac55a7cb5f8066f34c26551d2b6e8376dc8285 --- /dev/null +++ b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4052957b081b7486ee08b5303cdf4fe51e65ab0c0f93f31ab74e38898bd68375 +size 828132 diff --git a/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/layout.json b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d515d9f0df01c8712290a4912632676f571db137 --- /dev/null +++ b/2023/Synchronize Feature Extracting and Matching_ A Single Branch Framework for 3D Object Tracking/layout.json @@ -0,0 +1,9207 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 49, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 544, + 140 + ], + "type": "text", + "content": "Synchronize Feature Extracting and Matching: A Single Branch Framework for 3D Object Tracking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "spans": [ + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "text", + "content": "Teli Ma" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "text", + "content": ", Mengmeng Wang" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "text", + "content": ", Jimin Xiao" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "text", + "content": ", Huifeng Wu" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "text", + "content": ", Yong Liu" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology, Guangzhou " + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "text", + "content": "Zhejiang University \n" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "text", + "content": "Xi'an Jiaotong-Liverpool University " + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 72, + 160, + 519, + 204 + ], + "type": "text", + "content": "Hangzhou Dianzi University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 206, + 447, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 206, + 447, + 217 + ], + "spans": [ + { + "bbox": [ + 144, + 206, + 447, + 217 + ], + "type": "text", + "content": "tmal84@connect.hkust-gz.edu.cn mengmengwang@zju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 220, + 474, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 220, + 474, + 232 + ], + "spans": [ + { + "bbox": [ + 119, + 220, + 474, + 232 + ], + "type": "text", + "content": "jimin.xiao@xjtlu.edu.cn whf@hdu.edu.cn yongliu@iipc.zju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 284, + 290, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 284, + 290, + 583 + ], + "spans": [ + { + "bbox": [ + 45, + 284, + 290, + 583 + ], + "type": "text", + "content": "Siamese network has been a de facto benchmark framework for 3D LiDAR object tracking with a shared-parametric encoder extracting features from template and search region, respectively. This paradigm relies heavily on an additional matching network to model the cross-correlation/similarity of the template and search region. In this paper, we forsake the conventional Siamese paradigm and propose a novel single-branch framework, SyncTrack, synchronizing the feature extracting and matching to avoid forwarding encoder twice for template and search region as well as introducing extra parameters of matching network. The synchronization mechanism is based on the dynamic affinity of the Transformer, and an in-depth analysis of the relevance is provided theoretically. Moreover, based on the synchronization, we introduce a novel Attentive Points-Sampling strategy into the Transformer layers (APST), replacing the random/Farthest Points Sampling (FPS) method with sampling under the supervision of attentive relations between the template and search region. It implies connecting point-wise sampling with the feature learning, beneficial to aggregating more distinctive and geometric features for tracking with sparse points. Extensive experiments on two benchmark datasets (KITTI and NuScenes) show that SyncTrack achieves state-of-the-art performance in real-time tracking." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 625, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 287, + 686 + ], + "type": "text", + "content": "With recent advances in autonomous driving, 3D vision tasks based on LiDAR are becoming increasingly popular in the visual community. Among these tasks, 3D LiDAR single object tracking (SOT) aims to track a specific target object in a 3D video with the knowledge of 3D bounding" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 266, + 537, + 357 + ], + "blocks": [ + { + "bbox": [ + 313, + 266, + 537, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 266, + 537, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 537, + 357 + ], + "type": "image", + "image_path": "34e95910b51c9d23fd6adbd9b0e7028940a773dffc409aba5ca465a243bbc55a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 357, + 537, + 431 + ], + "blocks": [ + { + "bbox": [ + 313, + 357, + 537, + 431 + ], + "lines": [ + { + "bbox": [ + 313, + 357, + 537, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 537, + 431 + ], + "type": "image", + "image_path": "8f82eb7be74c14ff2f728acb6c8423c38537bd8b731d1c78035d371a60030d7e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 313, + 432, + 537, + 509 + ], + "blocks": [ + { + "bbox": [ + 313, + 432, + 537, + 509 + ], + "lines": [ + { + "bbox": [ + 313, + 432, + 537, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 537, + 509 + ], + "type": "image", + "image_path": "1dd7571ea949c6276854de24591f2dc474c069cd633a58e6ecd1948342536455.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 519, + 545, + 563 + ], + "lines": [ + { + "bbox": [ + 306, + 519, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 519, + 545, + 563 + ], + "type": "text", + "content": "Figure 1. The comparison of (a) Siamese network based trackers, (b) previous single-branch, two-stage framework M2Track [51] with (c) our SyncTrack, which is a single-branch and single-stage framework." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 581, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 581, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 305, + 581, + 545, + 616 + ], + "type": "text", + "content": "box in the initial frame. This task meets numerous challenges, such as LiDAR point cloud sparsity, occlusions, and fast motions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "Most existing methods [32, 50, 19, 12, 33, 52, 20, 9, 39] of 3D SOT mainly adopt a Siamese-like backbone and incorporate an additional matching network to cope with the tracking challenges as shown in Fig. 1(a). Trackers based on the Siamese-like backbone separate feature extraction of template and search region, forwarding the two kinds of features with shared model parameters, respectively. Subsequently, an extra matching network is introduced to fuse" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 137, + 712 + ], + "type": "text", + "content": "†Corresponding Author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9953" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 336 + ], + "type": "text", + "content": "the extracted template and search region features to model the correlation or similarity between them. However, such a paradigm restricts the feature interaction to a post-matching network, correlating the template and search region insufficiently with merely the high-level extracted features. In other words, the matching process posterior to the encoder is incapable of modeling the relations of multi-scale features intra-backbone. Moreover, a standalone matching network results in extra model parameters and computational overheads, let alone the double forwarding process of the Siamese-backbone to extract the template and search region features. M2Track [51] proposed a motion-centric paradigm to replace the Siamese-like structure, constructing a spatial-temporal point cloud to predict the motion. However, they still rely heavily on an additional motion transformation network, which requires extra training input, to integrate the extracted template features into the search region, and another two-stage refinement network is leveraged to ensure the performance as illustrated in Fig. 1(b). Based on aforementioned problems, we ask the question: Can feature extracting and matching be conducted simultaneously in a simple way?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 339, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 289, + 567 + ], + "type": "text", + "content": "The answer is Yes and the solution is implied in the dynamic global reasoning property of Transformer [35, 7, 10]. Specifically, the affinity matrix of all tokens can be constructed dynamically via continuous computation of the key and query vectors in the attention mechanism. The spatial context is aggregated using affinity to attend features. Intuitively, this affinity matrix can intrinsically serve as the matching matrix for intermediate feature interactions between the template and search region if we merge them into one input of the Transformer layers. Therefore, we propose a single-branch and single-stage framework equipped with a Transformer-based backbone instead of the conventional Siamese-like PointNet++ [31] backbone, as shown in Fig 1(c). The framework is dubbed as SyncTrack, as the Transformer backbone synchronizes the feature extracting and matching process. The SyncTrack is composed of a simple backbone and prediction head, omitting the complex matching network design and motion state estimation, depending merely on point-wise features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "text", + "content": "However, 3D point clouds have unique properties such as sparsity [54, 24], density variance, and implicit geometric features in data locality [29]. For example, " + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "text", + "content": " samples of KITTI [11] Car category have less than 100 points [19]. These problems are further aggravated when point clouds are grouped and down-sampled to formulate multi-scale point-wise feature maps. The phenomena stresses the vitality of point clouds sampling, aiming to improve the pointwise perception efficiency with limited points. PTTR [52] proposed to sample the input point clouds before the backbone, utilizing the " + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 570, + 289, + 715 + ], + "type": "text", + "content": " distance as the similarity metric for sampling. However, as down-sampling layer by layer is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "essential in tracking backbone for multi-scale feature fusion to strengthen representation, it is reasonable to consider the sampling strategy in the backbone. Therefore, we propose the attentive sampling strategy based on the attention map between the template and search region, and equip each Transformer layer with our sampling module as shown in Fig. 2(b). We name the Transformer containing Attentive Points-Sampling as APST. Specifically, the attentive response from template tokens to search region tokens is considered, as the positively respond tokens are more likely to be in the foreground and should be preserved for feature extracting. By contrast, as Fig. 3 shows, random sampling easily falls into the perceptive confusion as selected points hardly contain geometric features due to randomness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 240, + 547, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 547, + 264 + ], + "type": "text", + "content": "The main contributions of our paper can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 273, + 547, + 518 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 273, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 545, + 369 + ], + "type": "text", + "content": "- We introduce a single-branch and single-stage framework for real-time 3D LiDAR SOT dubbed SyncTrack, without Siamese-like forward propagation and a standalone matching network. We ingeniously leverage the dynamic affinity characteristic of the self-attention mechanism to synchronize the feature extracting and matching. A detailed analysis is provided to explain the synchronizing mechanism." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 377, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 545, + 438 + ], + "type": "text", + "content": "- We propose a novel APST to build the backbone, replacing the random/FPS" + }, + { + "bbox": [ + 317, + 377, + 545, + 438 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 317, + 377, + 545, + 438 + ], + "type": "text", + "content": " down-sampling of point clouds with attentive sampling to preserve more target-relevant points, and thus improving the perceptive capability of feature extracting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 446, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 446, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 317, + 446, + 547, + 518 + ], + "type": "text", + "content": "- Extensive results show that our method has achieved new state-of-the-art performance on the KITTI and NuScenes datasets in real-time tracking, up to " + }, + { + "bbox": [ + 317, + 446, + 547, + 518 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 317, + 446, + 547, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 446, + 547, + 518 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 317, + 446, + 547, + 518 + ], + "type": "text", + "content": " on mean results with a high-speed of around 45 fps. Besides, SyncTrack exhibits good scalability in both width and depth." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 538, + 399, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 399, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 399, + 551 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 558, + 421, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 421, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 421, + 571 + ], + "type": "text", + "content": "2.1. 2D Visual Tracking" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 577, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 547, + 685 + ], + "type": "text", + "content": "Advances in 2D video [53, 27, 26] and object tracking approaches [30, 16, 36, 23, 34, 48] stimulate the development of the 3D tracking, and the methods are evolving all the time. Early 2D SOT methods mainly focus on the classifier design like structured output SVM [36, 14], correlation filter [17, 25]. With the prevalence of deep learning, end-to-end trainable trackers emerged in the visual community. The Siamese-like structure [6, 18, 15, 23, 46, 42, 28] based methods have been popular in the tracking field and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1In this paper, we use FPS to refer to Farthest Points Sampling and fps to denote frames per second." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9954" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 83, + 394, + 308 + ], + "blocks": [ + { + "bbox": [ + 50, + 83, + 394, + 308 + ], + "lines": [ + { + "bbox": [ + 50, + 83, + 394, + 308 + ], + "spans": [ + { + "bbox": [ + 50, + 83, + 394, + 308 + ], + "type": "image", + "image_path": "f420362eb4b670a23493bb4274b368f8c835a6d9695f7078062b20156234e0fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 312, + 250, + 322 + ], + "lines": [ + { + "bbox": [ + 238, + 312, + 250, + 322 + ], + "spans": [ + { + "bbox": [ + 238, + 312, + 250, + 322 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 408, + 80, + 541, + 306 + ], + "blocks": [ + { + "bbox": [ + 408, + 80, + 541, + 306 + ], + "lines": [ + { + "bbox": [ + 408, + 80, + 541, + 306 + ], + "spans": [ + { + "bbox": [ + 408, + 80, + 541, + 306 + ], + "type": "image", + "image_path": "da9622d72d628dbe5d780a8a1bb78c8b4ac0631f2ca2728f6ba497c130ff0d6e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 312, + 479, + 322 + ], + "lines": [ + { + "bbox": [ + 466, + 312, + 479, + 322 + ], + "spans": [ + { + "bbox": [ + 466, + 312, + 479, + 322 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 327, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 546, + 361 + ], + "type": "text", + "content": "Figure 2. (a). The overall framework of SyncTrack and specific illustration of the single-branch backbone, the template and search region input are concatenated for feature extraction and matching synchronously. (b). The structure of the Attentive Points-Sampling Transformer (APST), which samples search region points after multi-head attention." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 380, + 289, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 289, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 289, + 668 + ], + "type": "text", + "content": "have numerous variants. SiamFC [1] is a pioneering work integrating feature correlation into a fully convolutional Siamese network for visual tracking. Subsequently, improvements like introducing detection components such as region proposal network [23, 22, 8], discriminating fore and back-ground [49], anchor-free detecting [13] etc.. Recently, vision transformers have been introduced into 2D SOT [4, 38, 43, 41] to exploit the long-range modeling of the attention mechanism to fuse the features effectively. Moreover, the one-stream network trackers based on transformers are proposed in 2D tracking. The OSTrack [44] proposes joint feature learning and relation modeling, masking a proportion of image patches to save computational overheads. SimTrack [3] concatenates the template and search input, improving the patch embedding method with a Foveal window strategy. Our work draws inspiration from the aforementioned trend in 2D vision. However, it is specifically tailored to address the challenges of the 3D Single Object Tracking (SOT) problem, taking into account the unique characteristics of point clouds. Furthermore, we analyze the success of the transformer-based single-branch backbone, attributing to the dynamic affinity of the attention mechanism, which enables the synchronization of feature extraction and matching processes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 380, + 420, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 420, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 420, + 392 + ], + "type": "text", + "content": "2.2.3D Visual Tracking" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 402, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 715 + ], + "type": "text", + "content": "In this paper, we only discuss LiDAR-based 3D object tracking. Till to now, almost all the 3D tracking methods [12, 32, 9, 45, 5, 33, 52, 50, 19, 20, 37] are based on the Siamese structure. The pioneering work in this field is the SC3D [12], which initially defines the task. SC3D employs cosine similarity to measure the resemblance between template and search region features, and incorporates shape completion during training to enhance appearance refinement. Trackers following SC3D make advancements from two perspectives. Firstly, they enhance the matching network [32, 50, 19, 52, 5, 39, 51]. For instance, MLVSNet [39] uses the CBAM module [40] to enhance the vote cluster features with both channel and spatial attention. STNet [20] employs cross- and self-attention modules to enhance the interaction between the extracted template and search region features, boosting their feature-level integration. M2Track introduces a motion-centric paradigm and motion estimation module to correlate the template and search features instead of appearance matching. All these methods depend on a standalone module to match the features. Secondly, the trackers [32, 9, 19, 39, 33] have attempted to improve the prediction head part. P2B [32] employs Hough Voting to predict the target location, and [50, 33, 39] all follow the voting strategy to make predictions. 3D-SiamRPN [9] uses an RPN head to predict the final results. LTTR [9] and V2B [19] use center-based regres" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9955" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "sions to predict several object properties. However, few efforts are made in exploring the encoder/backbone of trackers as PointNet++ [31] is the feature extractor by default. In this study, we focus on the backbone design and incorporate the matching process directly into the backbone, significantly streamlining the tracker network." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 103, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 103, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 103, + 168 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 176, + 288, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 176, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 288, + 297 + ], + "type": "text", + "content": "In this section, we first define the 3D SOT task in Sec. 3.1. Then, a detailed introduction of the single-branch framework is included in the Sec. 3.2. Based on the single-branch framework, how the feature extracting and matching are synchronized is elaborated in (Sec. 3.3). Moreover, we propose the Attentive Points-Sampling Transformer to build the single-branch backbone and sampling search region tokens with a strategy of attentive sampling, as shown in Sec. 3.4. The decoder head and losses are described in Sec. 3.5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 305, + 160, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 160, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 160, + 317 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": "In the configuration of 3D LiDAR single object detection (SOT) task, the 3D bounding box (BBox) is defined as " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "(x,y,z,w,l,h,\\theta)\\in \\mathbb{R}^7" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " represents the coordinate center of the BBox and " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "(w,l,h),\\theta" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " stand for the BBox size and heading angle (the rotation around the up-axis) respectively. Generally, the BBox size is assumed to be fixed by default even when the target object is non-rigid, thus minimizing the dimensions of BBox from " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^7" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^4" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": ". Given a sequence of temporally-connected point clouds " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}_i\\}_{i = 1}^T" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " is the number of points in each frame) and a initial BBox " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_1" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " of the target, the goal of SOT is to localize the target BBoxes " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i\\}_{i = 2}^T" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " in all frames online. Following the previous manner, a template point cloud " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t = \\{p_i^t\\}_{i = 1}^{N_t}" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " and a search region " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s = \\{p_i^s\\}_{i = 1}^{N_s}" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " are generated, where " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " are number of template and search region points. The template " + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "inline_equation", + "content": "P^t" + }, + { + "bbox": [ + 46, + 324, + 287, + 531 + ], + "type": "text", + "content": " is generated by cropping and centering the target in the initial frame based on the initial BBox." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 539, + 185, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 185, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 185, + 551 + ], + "type": "text", + "content": "3.2. Single-Branch Structure" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "We propose to replace the conventional Siamese-like backbone paradigm with a sole backbone, eliminating the double forward process of Siamese structure. Therefore, template and search region seeds are concatenated to do forward propagation. The Transformers' property of long-range relation-modeling is the intrinsic merit to tackle the concatenated template and search region seeds. Based on that, we leverage the self-attention modules to build the single-branch backbone as shown in Fig. 2(a). In our approach, we utilize a Query & Group module to sample the template seeds, denoted as " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t\\in \\mathbb{R}^{N_t\\times C}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", using the FPS method before jointly forwarding. This module also groups the k-nearest points to aggregate features. However, when" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "it comes to the search region point cloud, represented as " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^s\\in \\mathbb{R}^{N_s\\times C}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", we solely employ the Group module to aggregate neighbor information without reducing the number of search points. Subsequently, the template and search seeds are concatenated, incorporating a joint parametric positional embedding for the localization of tokens. This process is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 161, + 518, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 161, + 518, + 176 + ], + "spans": [ + { + "bbox": [ + 321, + 161, + 518, + 176 + ], + "type": "interline_equation", + "content": "\\mathcal {T} ^ {t} = \\mathrm {k N N} (\\operatorname {F P S} (\\mathcal {P} ^ {t})), \\mathcal {P} ^ {t} \\in \\mathbb {R} ^ {N _ {t} ^ {\\prime} \\times C ^ {\\prime}}, N _ {t} ^ {\\prime} < N _ {t},", + "image_path": "37997ad3c78eb1b5f2dea2e4bd038486ae3a6bb65cbd1c0fbc33b96bccc65534.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 332, + 178, + 545, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 178, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 332, + 178, + 545, + 194 + ], + "type": "interline_equation", + "content": "\\mathcal {T} ^ {s} = \\mathrm {k N N} \\left(\\mathcal {P} ^ {s}\\right), \\mathcal {P} ^ {s} \\in \\mathbb {R} ^ {N _ {s} \\times C ^ {\\prime}}, \\tag {1}", + "image_path": "761ace4168aeb0778e6f50ac2196952416eb6c957f182148e1f78171f576a0e7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 196, + 507, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 196, + 507, + 212 + ], + "spans": [ + { + "bbox": [ + 328, + 196, + 507, + 212 + ], + "type": "interline_equation", + "content": "\\mathcal {T} ^ {t s} = \\left[ \\mathcal {T} ^ {t}; \\mathcal {T} ^ {s} \\right] + \\mathbf {p _ {e}}, \\mathcal {T} ^ {t s} \\in \\mathbb {R} ^ {\\left(N _ {t} + N _ {s}\\right) \\times C ^ {\\prime}}.", + "image_path": "d7fa65470118f80a3fb47f462c4b136938efd39abaaeabb7da31afef31f6dc11.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 219, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 266 + ], + "type": "text", + "content": "Afterward, linear layers are leveraged to project the input tokens into query, key, and value latent and the head-wise joint attention map is calculated to model the intra- & interrelations of template and search tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 283, + 518, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 283, + 518, + 298 + ], + "spans": [ + { + "bbox": [ + 321, + 283, + 518, + 298 + ], + "type": "interline_equation", + "content": "Q ^ {t s}, K ^ {t s}, V ^ {t s} = W _ {q} (\\mathcal {T} ^ {t s}), W _ {k} (\\mathcal {T} ^ {t s}), W _ {v} (\\mathcal {T} ^ {t s}),", + "image_path": "055222547c900e2c0f13d52c5dfe318cb9a8c2684fe437b8b6896d1159d0527c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 352, + 300, + 545, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 300, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 352, + 300, + 545, + 329 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {m} ^ {t s} = \\operatorname {S o f t m a x} \\frac {Q _ {m} ^ {t s} \\left(K _ {m} ^ {t s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}. \\tag {2}", + "image_path": "3aad38ee4a941697a00509bc3697f1af69ff0d250c1d59209195f56161172d14.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 336, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 371 + ], + "type": "text", + "content": "Based on the joint head-wise attention map " + }, + { + "bbox": [ + 304, + 336, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_m^{ts}" + }, + { + "bbox": [ + 304, + 336, + 545, + 371 + ], + "type": "text", + "content": " of template and search tokens, the features extracted by multi-head attention are:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 379, + 545, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 379, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 319, + 379, + 545, + 395 + ], + "type": "interline_equation", + "content": "\\mathcal {T} ^ {t s} = \\left[ \\mathcal {A} _ {1} ^ {t s} V _ {1} ^ {t s}, \\mathcal {A} _ {2} ^ {t s} V _ {2} ^ {t s}, \\dots , \\mathcal {A} _ {M} ^ {t s} V _ {M} ^ {t s} \\right] \\mathcal {W} + \\mathcal {T} ^ {t s}, \\tag {3}", + "image_path": "13f4c8cd257926e6c3baed02fd93b4411e936e2e9d9ad46a21a97da0f822f495.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 401, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 401, + 545, + 426 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 401, + 545, + 426 + ], + "type": "text", + "content": " is the number of the attention heads and " + }, + { + "bbox": [ + 304, + 401, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 401, + 545, + 426 + ], + "type": "text", + "content": " is the weight of a MLP." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 433, + 545, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 433, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 433, + 545, + 445 + ], + "type": "text", + "content": "3.3. Synchronize Feature Extracting and Matching" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 451, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 510 + ], + "type": "text", + "content": "We illustrate how the single-branch structure can synchronize the process of feature extracting and matching with a simple backbone. The formula of joint attention " + }, + { + "bbox": [ + 304, + 451, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_m^{ts}" + }, + { + "bbox": [ + 304, + 451, + 545, + 510 + ], + "type": "text", + "content": " in Eq. 2 can be expanded as ( " + }, + { + "bbox": [ + 304, + 451, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 451, + 545, + 510 + ], + "type": "text", + "content": " represents Softmax operation):" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 517, + 561, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 561, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 561, + 586 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left[ \\mathcal {A} _ {m} ^ {t}; \\mathcal {A} _ {m} ^ {s} \\right] = \\sigma \\frac {\\left[ Q _ {m} ^ {t} ; Q _ {m} ^ {s} \\right] \\left[ K _ {m} ^ {t} ; K _ {m} ^ {s} \\right] ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} \\\\ = \\left[ \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {t} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}, \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {t} \\left(K _ {m} ^ {s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}; \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}, \\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} \\right]. \\tag {4} \\\\ \\end{array}", + "image_path": "1b4c24fa12721a124390533b3c6f8823e7f310b99418bd7f93694d52d29d3e23.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 587, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 587, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 587, + 545, + 611 + ], + "type": "text", + "content": "The new extracted search region features can be obtained based on Eq 4 as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 331, + 617, + 545, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 617, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 331, + 617, + 545, + 663 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {T} _ {m} ^ {\\prime s} = \\sigma \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} V _ {m} ^ {t} + \\sigma \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {s}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} V _ {m} ^ {s}, \\tag {5} \\\\ [ V _ {m} ^ {t}; V _ {m} ^ {s} ] = V _ {m} ^ {t s}, \\\\ \\end{array}", + "image_path": "dff80c4702b977e100089cb23cf25e6c8a3d757aace57fa43caacff4e1b5614b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 670, + 545, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 670, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 670, + 545, + 716 + ], + "type": "text", + "content": "depending on the projected features " + }, + { + "bbox": [ + 304, + 670, + 545, + 716 + ], + "type": "inline_equation", + "content": "[V_m^t,V_m^s ]" + }, + { + "bbox": [ + 304, + 670, + 545, + 716 + ], + "type": "text", + "content": " of both the template and search region from last layer. The attention queried from search region to template, " + }, + { + "bbox": [ + 304, + 670, + 545, + 716 + ], + "type": "inline_equation", + "content": "\\sigma \\frac{Q_m^s(K_m^t)^\\top}{\\sqrt{C' / M}}" + }, + { + "bbox": [ + 304, + 670, + 545, + 716 + ], + "type": "text", + "content": " , is the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9956" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 76, + 136, + 200 + ], + "blocks": [ + { + "bbox": [ + 55, + 76, + 136, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 76, + 136, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 76, + 136, + 200 + ], + "type": "image", + "image_path": "26d8e974add66a293459be24a512e840c37ae2966bead8c0ac7ef90ab62195cd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 202, + 112, + 213 + ], + "lines": [ + { + "bbox": [ + 77, + 202, + 112, + 213 + ], + "spans": [ + { + "bbox": [ + 77, + 202, + 112, + 213 + ], + "type": "text", + "content": "Template" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 143, + 98, + 201, + 197 + ], + "blocks": [ + { + "bbox": [ + 143, + 98, + 201, + 197 + ], + "lines": [ + { + "bbox": [ + 143, + 98, + 201, + 197 + ], + "spans": [ + { + "bbox": [ + 143, + 98, + 201, + 197 + ], + "type": "image", + "image_path": "8efe66a72fe2bef8d12dac791820c82d1e9149f4c63fe5fcdcd88221b0dc90ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 198, + 187, + 218 + ], + "lines": [ + { + "bbox": [ + 153, + 198, + 187, + 218 + ], + "spans": [ + { + "bbox": [ + 153, + 198, + 187, + 218 + ], + "type": "text", + "content": "Random Sampling" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 211, + 78, + 282, + 195 + ], + "blocks": [ + { + "bbox": [ + 211, + 78, + 282, + 195 + ], + "lines": [ + { + "bbox": [ + 211, + 78, + 282, + 195 + ], + "spans": [ + { + "bbox": [ + 211, + 78, + 282, + 195 + ], + "type": "image", + "image_path": "30d056bab8c6fb337463f32f1cfb4f668bfda65fae82b3f2904df11ead0cee6e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 198, + 262, + 218 + ], + "lines": [ + { + "bbox": [ + 227, + 198, + 262, + 218 + ], + "spans": [ + { + "bbox": [ + 227, + 198, + 262, + 218 + ], + "type": "text", + "content": "Attentive Sampling" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 225, + 289, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 289, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 289, + 314 + ], + "type": "text", + "content": "Figure 3. Comparisons of random sampling with our attentive sampling in point-wise down-sampling. We find that our attentive sampling approach excels in selecting more representative points, including those that contribute to the formation of the head, arms, and legs, which contain distinctive geometric features. In contrast, random sampling include more points that are far away from the target (we have removed such points for better visualization)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 335, + 288, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 382 + ], + "type": "text", + "content": "matching matrix that guides to aggregate highly-relevant template features. Moreover, the matching matrix is dynamic, which is determined by the changing query and key latent of search and template features as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 391, + 287, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 391, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 73, + 391, + 287, + 427 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\sigma} \\frac {Q _ {m} ^ {s} \\left(K _ {m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} = \\boldsymbol {\\sigma} \\frac {W _ {q , m} ^ {s} \\mathcal {T} _ {m} ^ {s} \\left(\\mathcal {T} _ {m} ^ {t}\\right) ^ {\\top} \\left(W _ {k , m} ^ {t}\\right) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}}, \\tag {6}", + "image_path": "059c8a140467887d5ddd8b09767f4349b4f5010168a8642e4b11463f7e5ae496.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 424, + 216, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 424, + 216, + 453 + ], + "spans": [ + { + "bbox": [ + 75, + 424, + 216, + 453 + ], + "type": "interline_equation", + "content": "i. e. \\qquad \\sigma \\frac {Q _ {m} ^ {s} (K _ {m} ^ {t}) ^ {\\top}}{\\sqrt {C ^ {\\prime} / M}} \\propto [ \\mathcal {T} _ {m} ^ {s}; \\mathcal {T} _ {m} ^ {t} ].", + "image_path": "41898b45d3cc6780e6866480f527f09f2cc2d0931447be50bd501b9afd470310.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 461, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 509 + ], + "type": "text", + "content": "To conclude, the synchronization attributes to the dynamic mechanism of matching matrix, continuously adapting the matching relations according to the extracted features of template and search region." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "Comparisons with Siamese Network. Previous Siamese-like trackers can be summarized as the paradigm of Extracting then Matching. If we name the feature extracting and matching as " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": ", then the objective of this paradigm can be concluded as " + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "inline_equation", + "content": "max(\\delta(\\phi(\\mathcal{P}^t), \\phi(\\mathcal{P}^s))" + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": " for simplicity. It means the model training is to shorten the distances between correlated parts of the template and search region based on the extracted features from the backbone. However, compared to our approach, this matching mechanism is relatively static since it occurs only after feature extraction, resulting in inadequate modeling of inter-backbone relations. On the other hand, our single-branch framework facilitates dynamic interaction between the search region seeds and the template across all layers of the backbone. This allows for comprehensive learning of relations, encompassing both the local representation from early layers and the global representation from later ones." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 515, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 515, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 515, + 85 + ], + "type": "text", + "content": "3.4. Attentive Points-Sampling Transformer" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 95, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 287 + ], + "type": "text", + "content": "Based on the synchronizing mechanism of feature extracting and matching, we propose to replace the Point-wise Transformer [47] with a novel Attentive Points-Sampling Transformer (APST). APST is proposed based on the observation that backbones utilized in previous 3D LiDAR trackers always adopt a non-parametric strategy of down-sampling the search region features/tokens with farthest point sampling (FPS) or randomly sampling to reduce the points as introduced in PointNet++ [31]. Nevertheless, this non-parametric sampling method lacks learnability and controllability since no parameters associated with the sampling process are updated during model training. Consequently, the final performance is compromised due to the significant influence of the points-center in LiDAR, which determines the effectiveness of extracting features around the foreground, as shown in Fig. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 292, + 546, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 546, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 546, + 455 + ], + "type": "text", + "content": "Therefore, we introduce the APST, which involves the selection of points based on attentive relations between the tokens of the template and search region as illustrated in sub-figure (b) of Fig. 2. The attentive response of template tokens reacting to search region tokens is considered as positively responding search tokens are more likely to be in the foreground. In that case, we segment the attention map and separate the " + }, + { + "bbox": [ + 304, + 292, + 546, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_m^{t\\rightarrow s} = \\sigma \\frac{Q_m^t(K_m^s)^{\\top}}{\\sqrt{C' / M}}" + }, + { + "bbox": [ + 304, + 292, + 546, + 455 + ], + "type": "text", + "content": " from the Eq. 4, averaging along the dimension of template-tokens attention among all transformer heads to acquire the response scores. To ensure the maximum response of search tokens to template ones, a group of search region indexes, named " + }, + { + "bbox": [ + 304, + 292, + 546, + 455 + ], + "type": "inline_equation", + "content": "\\Omega_s" + }, + { + "bbox": [ + 304, + 292, + 546, + 455 + ], + "type": "text", + "content": ", is dynamically updated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 473, + 545, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 473, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 314, + 473, + 545, + 516 + ], + "type": "interline_equation", + "content": "\\Omega_ {s} ^ {*} = \\arg \\max _ {\\Omega_ {s}} \\frac {1}{N _ {t}} \\frac {1}{M} \\sum_ {t = 0} ^ {N _ {t}} \\sum_ {m = 0} ^ {M} \\frac {Q _ {m} ^ {t}}{\\sqrt {C ^ {\\prime} / M}} \\left(K _ {m} ^ {s}\\right) ^ {\\top} \\Big | _ {s \\in \\Omega_ {s}}, \\tag {7}", + "image_path": "4b436d9b9753ceb17552ea97e6eb6e9946caff30b9a25a9e955824fb365beb0c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "type": "text", + "content": "Based on the optimal solution " + }, + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\Omega_s^*" + }, + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "type": "text", + "content": ", the search region tokens are sampled to decrease the number of tokens after multi-head self-attention and concatenate with template ones. Note that the template tokens are sampled with the FPS method as introduced in Sec. 3.2, and only the search region tokens are sampled with the attentive sampling method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "The dynamic-affinity property of the Transformer suggests that the attention map is influenced by the latent tokens projected through learnable linear layers. Hence, it is reasonable to assert that sampling tokens guided by the attention map is linked to model learning, as the attention map is generated based on updated parameters. As a result, this type of token/point sampling proves advantageous in providing prior knowledge for centroid selection and enhancing the efficient aggregation of representational information, as shown in Fig. 3." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9957" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 163, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 163, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 163, + 83 + ], + "type": "text", + "content": "3.5. Decoder and Losses" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 289, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 289, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 289, + 224 + ], + "type": "text", + "content": "With the encoded point-wise features of various scales, a multi-scale feature fusion module is adopted to fuse these features and output the features with the same origin input size, feeding into the decoder part for final predictions. Following the V2B [19], the features are voxelized as a volumetric representation and 3D convolutions are utilized on the encoded features. Afterward, the BEV feature maps are acquired by pooling on the " + }, + { + "bbox": [ + 46, + 90, + 289, + 224 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 90, + 289, + 224 + ], + "type": "text", + "content": "-axis for regression. Focal loss and L1 loss are leveraged for classification and BBox center offset and rotation regression, respectively. A detailed introduction is presented in the supplementary." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 232, + 128, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 232, + 128, + 245 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 128, + 245 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 251, + 160, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 160, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 160, + 264 + ], + "type": "text", + "content": "4.1. Experiment Setups" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "text", + "content": "Implementation Details. We set the number of input points as " + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "inline_equation", + "content": "N_{t} = 512" + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "inline_equation", + "content": "N_{s} = 1024" + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "text", + "content": " for template and search regions by randomly duplicating and discarding, respectively. The encoder backbone is merely consisted of three layers with Attentive Points-Sampling Transformers, and the number of both template and search region points output by each layer are 256, 128 and 64. Moreover, each layer's feature dimensions in the encoder are 32, 64 and 128, respectively, whereas the final features for the prediction head are with 32 channels. Note that the heads for all APST are 2 by default. In the voxelization process, the region " + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "inline_equation", + "content": "[(x_{min}, x_{max}), (y_{min}, y_{max}), (z_{min}, z_{max})]" + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "inline_equation", + "content": "[(-5.6, 5.6), (-3.6, 3.6), (-2.4, 2.4)]" + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "text", + "content": " to contain most target points. The voxel size " + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "inline_equation", + "content": "(v_{x}, v_{y}, v_{z})" + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "inline_equation", + "content": "(0.3, 0.3, 0.3)" + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "text", + "content": ". For the detection head, four decomposed 3D (stride of 2,1,2,1 along the " + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 270, + 289, + 484 + ], + "type": "text", + "content": "-axis) and 2D convolution blocks (stride of 2,1,1,2) are leveraged to strengthen the feature aggregation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 485, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 287, + 557 + ], + "type": "text", + "content": "Training and Testing. We train the model for 40 epochs with a batch size of 64. The Adam optimizer [21] is adopted with the initial learning rate of 0.001 and reduced by 5 every 10 epochs (every 2 epochs for nuScenes). The classification loss has a weight " + }, + { + "bbox": [ + 47, + 485, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 47, + 485, + 287, + 557 + ], + "type": "text", + "content": " of 1 and the regression loss has a weight " + }, + { + "bbox": [ + 47, + 485, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 47, + 485, + 287, + 557 + ], + "type": "text", + "content": " of 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 640 + ], + "type": "text", + "content": "Evaluation Metrics. Following the previous methods [32, 50], we measure the Success and Precision of the tracker. To be specific, Success is defined as the IoU between predicted boxes and the ground truth, and Precision measures the AUC (Area Under Curve) of the distance between predicted and ground truth boxes within the range of [0,2] meters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 274, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 274, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 274, + 660 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Trackers" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Results on KITTI. KITTI [11] is one of the most popular datasets used in mobile robotics and autonomous driving. The tracking benchmark of KITTI consists of 21 training sequences and 29 test sequences. Following the previous" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "methods [12, 51, 50], we split the training sequences into train/val/test splits due to the inaccessibility of the testing labels, scenes 0-16 for training, scenes 7-18 for validation and scenes 19-20 for testing, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 121, + 546, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 338 + ], + "type": "text", + "content": "We compare the SyncTrack with other state-of-the-art methods from the pioneering SC3D [12] to the most recent siamese network STNet [20] and single branch network M2Track [51], as shown in Table 1. We separate the trackers into Siamese Network and Single Branch Network categories to compare with our proposed SyncTrack. Compared with siamese structured trackers, the SyncTrack achieves the best results on rigid and non-rigid object tracking, outperforming current tracking methods based on siamese networks on most specific categories and the overall mean results. STNet [20] is the state-of-the-art siamese-based tracking method with self-attention and cross-attention modules to match the template and search-region features. Our SyncTrack outperforms the STNet by a relatively large margin on Car, Van and Pedestrian categories under the evaluation of Success metric. Also, SyncTrack surpasses the previous best Mean results by " + }, + { + "bbox": [ + 304, + 121, + 546, + 338 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 304, + 121, + 546, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 121, + 546, + 338 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 304, + 121, + 546, + 338 + ], + "type": "text", + "content": " on Success and Precision metrics, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 339, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 547, + 434 + ], + "type": "text", + "content": "We also compare the SyncTrack with the only current single-branch tracker M2Track [51]. Our SyncTrack outperforms M2Track by a large margin, up to " + }, + { + "bbox": [ + 304, + 339, + 547, + 434 + ], + "type": "inline_equation", + "content": "7.8\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 434 + ], + "type": "text", + "content": " in Success in the category of Car whereas the M2Track is better in the category of Pedestrian. However, for the comprehensive evaluation, the Mean performance of all frames, our SyncTrack outperforms the M2Track by " + }, + { + "bbox": [ + 304, + 339, + 547, + 434 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 434 + ], + "type": "text", + "content": " on the Success metric." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 436, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 547, + 555 + ], + "type": "text", + "content": "Results on nuScenes. The nuScenes dataset [2] contains 1000 driving scenes collected from Boston and Singapore with a diverse and exciting set of driving maneuvers, traffic situations, and unexpected behaviors. In the configurations of LiDAR-based tracking methods, the train/val/test sets make up 700/150/150 of the whole 1000 scenes, respectively. Officially, the train set is evenly split into 'train track' and 'train detect' to remedy overfitting. Following [19], we train our model with \"train track\" split and test it on the val set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Note that nuScenes dataset only annotates keyframes and provides official interpolated results for the remaining frames, so there are two configurations for this dataset. The first is from [50, 51] which trains and tests both only on the keyframes. The second one is from [19, 20], which trains and tests on all the frames. The results based on these two configurations are different. We believe that the motion in key frames is substantial, which does not conform to the practical applications. Therefore, we train and test all the frames in this paper. We train the previous methods on the nuScenes by ourselves using their official codes to compare with our SyncTrack as many results are missing or only reported by testing nuScenes test-split with a pre" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9958" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 102, + 544, + 324 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 101 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 101 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 101 + ], + "type": "text", + "content": "Table 1. Success/Precision comparisons among our SyncTrack and the state-of-the-art methods on the KITTI datasets. Mean shows frame-level averaging results. Bold and underline denote the best performance and the second-best performance, respectively. Improvements over previous state-of-the-arts are shown in Italic and color." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 102, + 544, + 324 + ], + "lines": [ + { + "bbox": [ + 49, + 102, + 544, + 324 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 544, + 324 + ], + "type": "table", + "html": "
MethodsCar (6424)Cyclist (308)Van (1248)Pedestrian (6088)Mean (14068)
SuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecision
Siamese Network
SC3D [12]41.357.941.570.440.447.018.237.831.248.5
SC3D-RPN[45]36.351.043.081.4--17.947.8--
P2B [32]56.272.832.144.740.848.428.749.642.460.0
MLVSNet [39]56.074.034.344.552.061.434.161.145.766.6
3DSiamRPN [9]58.276.236.149.045.652.835.256.246.664.9
LTTR [5]65.077.166.289.935.845.633.256.848.765.8
PTT [33]67.881.837.247.343.652.544.972.055.174.2
BAT [50]60.577.733.745.452.467.042.170.151.272.8
V2B [19]70.581.340.849.750.158.048.373.558.475.2
PTTR [52]65.277.465.190.552.561.850.981.658.477.8
STNet [20]72.184.073.593.758.070.649.977.261.380.1
SyncTrack73.385.073.193.860.370.054.780.564.181.9
improvement+1.2+1.0-0.4+0.1+2.3-0.6+3.8-1.1+2.8+1.8
Single Branch Network
M2Track [51]65.580.873.293.553.870.761.588.262.983.4
SyncTrack73.385.073.193.860.370.054.780.564.181.9
improvement+7.8+4.2-0.1+0.3+2.3-0.7-6.8-7.7+1.2-1.5
", + "image_path": "b0670bf50791558c65499039528c182139f1abbaebd28e136c5824a9ec4dead8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 364, + 544, + 480 + ], + "blocks": [ + { + "bbox": [ + 47, + 332, + 546, + 364 + ], + "lines": [ + { + "bbox": [ + 47, + 332, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 546, + 364 + ], + "type": "text", + "content": "Table 2. Comparison among SyncTrack and the state-of-the-art methods on the nuScenes datasets. Mean shows the average result weighed by frame numbers. Bold and underline denote the best performance and the second-best performance, respectively. Improvements over previous state-of-the-arts are shown in Italic and color." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 364, + 544, + 480 + ], + "lines": [ + { + "bbox": [ + 49, + 364, + 544, + 480 + ], + "spans": [ + { + "bbox": [ + 49, + 364, + 544, + 480 + ], + "type": "table", + "html": "
MethodsCar (15578)Bicycle (501)Truck (3710)Pedestrian (8019)Mean (27808)
SuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecisionSuccessPrecision
SC3D [12]24.525.916.618.832.530.613.814.722.323.2
P2B [32]32.835.219.726.616.211.119.226.626.429.3
BAT [50]26.528.817.822.816.510.619.428.223.027.9
V2B [19]32.934.520.327.528.723.820.127.428.430.9
STNet [20]35.737.222.329.333.532.420.127.830.733.7
M2Track [51]31.433.922.629.830.128.820.728.028.031.4
SyncTrack36.738.123.830.439.438.619.127.831.835.1
improvement+1.0+0.9+1.2+0.6+5.9+6.2-1.6-0.4+1.1+1.4
", + "image_path": "1f9c169ca5dd59fe7aefd859c47d18f77eecd9a9ed0fee6200a31ee738b67ffb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 500, + 136, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 136, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 136, + 511 + ], + "type": "text", + "content": "trained KITTI model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": "As shown in Table 2, SyncTrack performs significantly better than other trackers on the mean results of four categories. Specifically, the SyncTrack yields the best results on both metrics and on most categories except the Pedestrian, which is lower than M2Track [51] and BAT [50] by a minor margin (1.6% and 0.4%). However, in the Truck category, our SyncTrack outperforms state-of-the-art by a large margin, up to 5.9% and 6.2% on Success and Precision respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Computational Cost Comparison. We analyze the computational overheads and inference speed of SyncTrack and compare it with other trackers in Table 3. The reported results are tested by ourselves with official codebases with a single Titan RTX GPU on Car category of KITTI. It can be seen that SyncTrack achieves the best Success performance with the lowest computational complexity (2.51" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 499, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 499, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 305, + 499, + 545, + 548 + ], + "type": "text", + "content": "G). Compared with the most current Siamese-based tracker STNet [20] and single-branch tracker M2Track, our SyncTrack has the fewest number of parameters and fastest inference speed, satisfying the demand of real-time tracking." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 314, + 593, + 538, + 683 + ], + "blocks": [ + { + "bbox": [ + 329, + 582, + 522, + 593 + ], + "lines": [ + { + "bbox": [ + 329, + 582, + 522, + 593 + ], + "spans": [ + { + "bbox": [ + 329, + 582, + 522, + 593 + ], + "type": "text", + "content": "Table 3. The computational cost of different trackers." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 593, + 538, + 683 + ], + "lines": [ + { + "bbox": [ + 314, + 593, + 538, + 683 + ], + "spans": [ + { + "bbox": [ + 314, + 593, + 538, + 683 + ], + "type": "table", + "html": "
MethodsParametersFLOPsFPSSuccess
SC3D [12]6.45 M20.07 G641.3
P2B [32]1.34 M4.28 G4856.2
BAT [50]1.47 M5.53 G5460.5
V2B [19]1.36 M5.57 G3970.5
STNet [20]1.66 M3.14 G3672.1
M2Track [51]2.24 M2.54 G3765.5
SyncTrack1.47 M2.51 G4573.3
", + "image_path": "a94613f4ef14400464e05f202b1805009238ef185d7c7767c9e06b85ad2dfc52.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9959" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 70, + 171, + 152 + ], + "blocks": [ + { + "bbox": [ + 74, + 101, + 85, + 120 + ], + "lines": [ + { + "bbox": [ + 74, + 101, + 85, + 120 + ], + "spans": [ + { + "bbox": [ + 74, + 101, + 85, + 120 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 89, + 70, + 171, + 152 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 171, + 152 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 171, + 152 + ], + "type": "image", + "image_path": "5801c758bbe5b150f08e940d7f5d216476ca9b0b6cd0b70a58ab08361ec1058b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 258, + 152 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 258, + 152 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 258, + 152 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 258, + 152 + ], + "type": "image", + "image_path": "0d9921175ab92df7386da8408bddcaad1b13c1000f55f515b1239cd7b5234897.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 263, + 70, + 345, + 152 + ], + "blocks": [ + { + "bbox": [ + 263, + 70, + 345, + 152 + ], + "lines": [ + { + "bbox": [ + 263, + 70, + 345, + 152 + ], + "spans": [ + { + "bbox": [ + 263, + 70, + 345, + 152 + ], + "type": "image", + "image_path": "23375ee6f4e24ce85bf2143aa1392ae979c4d64a5168b22558f1bdc488fb5399.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 351, + 70, + 433, + 152 + ], + "blocks": [ + { + "bbox": [ + 351, + 70, + 433, + 152 + ], + "lines": [ + { + "bbox": [ + 351, + 70, + 433, + 152 + ], + "spans": [ + { + "bbox": [ + 351, + 70, + 433, + 152 + ], + "type": "image", + "image_path": "e83a98dca80efba7d737f083db9c9ed48c3f96aa3b3a7b1be8a0cad3dc7c0524.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 440, + 70, + 521, + 152 + ], + "blocks": [ + { + "bbox": [ + 440, + 70, + 521, + 152 + ], + "lines": [ + { + "bbox": [ + 440, + 70, + 521, + 152 + ], + "spans": [ + { + "bbox": [ + 440, + 70, + 521, + 152 + ], + "type": "image", + "image_path": "75df7b179aaeb778fd509c2e9ca0ae0e01809b048263226b2a992d0facaa6dcc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 89, + 156, + 171, + 239 + ], + "blocks": [ + { + "bbox": [ + 74, + 180, + 85, + 210 + ], + "lines": [ + { + "bbox": [ + 74, + 180, + 85, + 210 + ], + "spans": [ + { + "bbox": [ + 74, + 180, + 85, + 210 + ], + "type": "text", + "content": "Cyclist" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 89, + 156, + 171, + 239 + ], + "lines": [ + { + "bbox": [ + 89, + 156, + 171, + 239 + ], + "spans": [ + { + "bbox": [ + 89, + 156, + 171, + 239 + ], + "type": "image", + "image_path": "c400235239773bea98973b17f2d604ef767ebd031a0c6b83cfed6d6b667d6c5d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 176, + 156, + 258, + 239 + ], + "blocks": [ + { + "bbox": [ + 176, + 156, + 258, + 239 + ], + "lines": [ + { + "bbox": [ + 176, + 156, + 258, + 239 + ], + "spans": [ + { + "bbox": [ + 176, + 156, + 258, + 239 + ], + "type": "image", + "image_path": "b93847361ea53ab19398a7698ce710e1c05df01192ac06fd04b94dc2b1ab8386.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 263, + 156, + 345, + 239 + ], + "blocks": [ + { + "bbox": [ + 263, + 156, + 345, + 239 + ], + "lines": [ + { + "bbox": [ + 263, + 156, + 345, + 239 + ], + "spans": [ + { + "bbox": [ + 263, + 156, + 345, + 239 + ], + "type": "image", + "image_path": "853a7293570efed7d2ad5eec3764433c6ef52ed4b68d995aa6c0bade13ab4c46.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 351, + 156, + 433, + 239 + ], + "blocks": [ + { + "bbox": [ + 351, + 156, + 433, + 239 + ], + "lines": [ + { + "bbox": [ + 351, + 156, + 433, + 239 + ], + "spans": [ + { + "bbox": [ + 351, + 156, + 433, + 239 + ], + "type": "image", + "image_path": "d00c77b875d7d2dd42fd7966042455ad3d2aab82ab997233f52b9e3873df9728.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 440, + 156, + 521, + 239 + ], + "blocks": [ + { + "bbox": [ + 440, + 156, + 521, + 239 + ], + "lines": [ + { + "bbox": [ + 440, + 156, + 521, + 239 + ], + "spans": [ + { + "bbox": [ + 440, + 156, + 521, + 239 + ], + "type": "image", + "image_path": "1fe662d6f998f80de7f73932e75985ad017ac0c4d2d2ff661ce555dc3b63dd57.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 89, + 243, + 171, + 326 + ], + "blocks": [ + { + "bbox": [ + 74, + 284, + 84, + 300 + ], + "lines": [ + { + "bbox": [ + 74, + 284, + 84, + 300 + ], + "spans": [ + { + "bbox": [ + 74, + 284, + 84, + 300 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 89, + 243, + 171, + 326 + ], + "lines": [ + { + "bbox": [ + 89, + 243, + 171, + 326 + ], + "spans": [ + { + "bbox": [ + 89, + 243, + 171, + 326 + ], + "type": "image", + "image_path": "21c25b9b289b0a8498af9ec8349c22908c4bc19bf9664b7d4a693eba4646f901.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 176, + 243, + 258, + 325 + ], + "blocks": [ + { + "bbox": [ + 176, + 243, + 258, + 325 + ], + "lines": [ + { + "bbox": [ + 176, + 243, + 258, + 325 + ], + "spans": [ + { + "bbox": [ + 176, + 243, + 258, + 325 + ], + "type": "image", + "image_path": "24f8cfb64039ec8d215784e527b78f06dfd693e808d78015208293e78eb05032.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 263, + 244, + 345, + 325 + ], + "blocks": [ + { + "bbox": [ + 263, + 244, + 345, + 325 + ], + "lines": [ + { + "bbox": [ + 263, + 244, + 345, + 325 + ], + "spans": [ + { + "bbox": [ + 263, + 244, + 345, + 325 + ], + "type": "image", + "image_path": "baccd366d1e5ad84995dc518dc5a517f0c83468dbed80800406db066377b7f26.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 351, + 244, + 433, + 325 + ], + "blocks": [ + { + "bbox": [ + 351, + 244, + 433, + 325 + ], + "lines": [ + { + "bbox": [ + 351, + 244, + 433, + 325 + ], + "spans": [ + { + "bbox": [ + 351, + 244, + 433, + 325 + ], + "type": "image", + "image_path": "c1560e93f0f00465798e7fc131758ad4791d1987a144fa901e4e45b435028a0a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 440, + 244, + 521, + 325 + ], + "blocks": [ + { + "bbox": [ + 440, + 244, + 521, + 325 + ], + "lines": [ + { + "bbox": [ + 440, + 244, + 521, + 325 + ], + "spans": [ + { + "bbox": [ + 440, + 244, + 521, + 325 + ], + "type": "image", + "image_path": "04e23faea3658d040771c416ea59fba8cc94bc2724b347cea12e93618a6eb85b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 89, + 327, + 171, + 409 + ], + "blocks": [ + { + "bbox": [ + 74, + 342, + 84, + 387 + ], + "lines": [ + { + "bbox": [ + 74, + 342, + 84, + 387 + ], + "spans": [ + { + "bbox": [ + 74, + 342, + 84, + 387 + ], + "type": "text", + "content": "Peperrnnnne" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 89, + 327, + 171, + 409 + ], + "lines": [ + { + "bbox": [ + 89, + 327, + 171, + 409 + ], + "spans": [ + { + "bbox": [ + 89, + 327, + 171, + 409 + ], + "type": "image", + "image_path": "8a2add9e7555554f8a6db3acc16e4031d4ee2c0a12054349b12f20476508c780.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 177, + 327, + 258, + 410 + ], + "blocks": [ + { + "bbox": [ + 177, + 327, + 258, + 410 + ], + "lines": [ + { + "bbox": [ + 177, + 327, + 258, + 410 + ], + "spans": [ + { + "bbox": [ + 177, + 327, + 258, + 410 + ], + "type": "image", + "image_path": "8c10da3371a8b869ac70f1398ecfc42b8d55948c54d76723cd51565fe2552479.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 415, + 267, + 425 + ], + "lines": [ + { + "bbox": [ + 207, + 415, + 267, + 425 + ], + "spans": [ + { + "bbox": [ + 207, + 415, + 267, + 425 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 263, + 327, + 345, + 410 + ], + "blocks": [ + { + "bbox": [ + 263, + 327, + 345, + 410 + ], + "lines": [ + { + "bbox": [ + 263, + 327, + 345, + 410 + ], + "spans": [ + { + "bbox": [ + 263, + 327, + 345, + 410 + ], + "type": "image", + "image_path": "97ea566193899caa919442eb0c5781062fcdbf3b85e57e5c8896738de9f8c725.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 415, + 324, + 425 + ], + "lines": [ + { + "bbox": [ + 296, + 415, + 324, + 425 + ], + "spans": [ + { + "bbox": [ + 296, + 415, + 324, + 425 + ], + "type": "text", + "content": "STNet" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 351, + 327, + 433, + 410 + ], + "blocks": [ + { + "bbox": [ + 351, + 327, + 433, + 410 + ], + "lines": [ + { + "bbox": [ + 351, + 327, + 433, + 410 + ], + "spans": [ + { + "bbox": [ + 351, + 327, + 433, + 410 + ], + "type": "image", + "image_path": "1c364da645eae84d1be3f235bf6fb37a1e1539f2b2d5691d668b569449d9ec4b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 415, + 402, + 425 + ], + "lines": [ + { + "bbox": [ + 356, + 415, + 402, + 425 + ], + "spans": [ + { + "bbox": [ + 356, + 415, + 402, + 425 + ], + "type": "text", + "content": "SyncTrack" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 440, + 327, + 521, + 410 + ], + "blocks": [ + { + "bbox": [ + 440, + 327, + 521, + 410 + ], + "lines": [ + { + "bbox": [ + 440, + 327, + 521, + 410 + ], + "spans": [ + { + "bbox": [ + 440, + 327, + 521, + 410 + ], + "type": "image", + "image_path": "a1f32a3afcf2b1d4751d7b10dbd7ac31297f3531b042d2ac04bec885d9fa26a6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 194, + 413, + 205, + 425 + ], + "blocks": [ + { + "bbox": [ + 194, + 413, + 205, + 425 + ], + "lines": [ + { + "bbox": [ + 194, + 413, + 205, + 425 + ], + "spans": [ + { + "bbox": [ + 194, + 413, + 205, + 425 + ], + "type": "image", + "image_path": "a529280928d04447fab05c3a53a3f3f8d050cb88e980d9a83cc93ddffc477169.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 282, + 413, + 294, + 426 + ], + "blocks": [ + { + "bbox": [ + 282, + 413, + 294, + 426 + ], + "lines": [ + { + "bbox": [ + 282, + 413, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 282, + 413, + 294, + 426 + ], + "type": "image", + "image_path": "4aa7209e962675ca066240ef85d094fc591ac4969422355ad49c59a0bde47b58.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 428, + 545, + 451 + ], + "lines": [ + { + "bbox": [ + 46, + 428, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 428, + 545, + 451 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with STNet. We compare SyncTrack with STNet over four tracking sequences of different categories in KITTI. The predictions of SyncTrack fits the ground-truth boxes better." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 343, + 413, + 354, + 425 + ], + "blocks": [ + { + "bbox": [ + 343, + 413, + 354, + 425 + ], + "lines": [ + { + "bbox": [ + 343, + 413, + 354, + 425 + ], + "spans": [ + { + "bbox": [ + 343, + 413, + 354, + 425 + ], + "type": "image", + "image_path": "a3bd6151f15a312ed904c1966f7a420e28b60d9f89c4f0c64d8ba44ad7b6bf4b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 47, + 469, + 174, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 174, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 174, + 483 + ], + "type": "text", + "content": "4.3. Generalization Ability" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 490, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 287, + 586 + ], + "type": "text", + "content": "To evaluate the generalization ability of SyncTrack, we pre-train the model on the KITTI dataset and test it directly on the nuScenes dataset without fine-tuning. The results are shown in Table 4. It can be observed that SyncTrack outperforms other methods on the mean results of four categories by a large margin. SyncTrack not only achieves a good balance between inference speed and tracking accuracy, but also generalizes very well to new domains." + } + ] + } + ], + "index": 32 + }, + { + "type": "table", + "bbox": [ + 50, + 608, + 287, + 696 + ], + "blocks": [ + { + "bbox": [ + 47, + 597, + 286, + 608 + ], + "lines": [ + { + "bbox": [ + 47, + 597, + 286, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 286, + 608 + ], + "type": "text", + "content": "Table 4. Testing results on the nuScenes for generalization ability." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 608, + 287, + 696 + ], + "lines": [ + { + "bbox": [ + 50, + 608, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 50, + 608, + 287, + 696 + ], + "type": "table", + "html": "
MethodsCar (15578)Bicycle (501)Truck (3710)Pedestrian (8019)Mean (27808)
SC3D [12]25.0/27.117.0/18.225.7/21.914.2/16.221.8/23.1
P2B [32]27.0/29.220.0/26.421.5/16.215.9/22.022.9/25.3
BAT [50]22.5/24.117.0/18.819.3/15.817.3/24.520.5/23.0
V2B [19]31.3/35.122.2/19.121.7/19.117.3/23.425.8/29.0
STNet [20]32.2/36.121.2/29.222.3/16.819.1/27.226.9/30.8
SyncTrack32.8/36.321.7/28.323.9/19.219.3/27.127.5/31.2
", + "image_path": "da86ef11e76b0dabba9d8b756c29dfe2f9cc29478f978162701c7d9f560100c8.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 469, + 440, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 469, + 440, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 469, + 440, + 483 + ], + "type": "text", + "content": "4.4. Scalability of Backbone" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 489, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 646 + ], + "type": "text", + "content": "We prove that our SyncTrack has good scalability in large-scale datasets like nuScenes. The basic model of SyncTrack has merely one Transformer layer in each stage to ensure real-time performance for tracking. However, the backbone of SyncTrack is scalable in both depth and width. We name the basic model as SyncTrack-Small. Furthermore, the SyncTrack-Mid has three Transformer layers at every stage, with nine layers total. As for the SyncTrack-Large, it doubles the number of feature channels of the SyncTrack-Mid to [256, 128, 64] for every stage. Table 5 reveals that performance improves when scalability increases not only in depth (Small v.s. Mid), but also in width (Mid v.s. Large)." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 656, + 406, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 406, + 668 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 406, + 668 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 305, + 676, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 676, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 305, + 676, + 545, + 700 + ], + "type": "text", + "content": "We conduct comprehensive ablations to evaluate the components of SyncTrack." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Compare with Siamese Structure. The synchronized fea" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9960" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 91, + 286, + 173 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 286, + 90 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 286, + 90 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 286, + 90 + ], + "type": "text", + "content": "Table 5. The scalability of SyncTrack on nuScenes, (S, M, L represent small, middle, large model size of SyncTrack)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 91, + 286, + 173 + ], + "lines": [ + { + "bbox": [ + 49, + 91, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 286, + 173 + ], + "type": "table", + "html": "
Scale#ParamFLOPsSuccessPrecision
BicycleS1.47 M2.51 G23.830.4
M1.82 M2.63 G25.033.6
L3.98 M5.37 G25.634.1
TruckS1.47 M2.51 G39.438.6
M1.82 M2.63 G40.138.8
L3.98 M5.37 G40.538.9
", + "image_path": "877bd4f8441dd09b57216728424dfb90ba3c96a5aadd8b66e26548005f9729f5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 185, + 286, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 286, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 286, + 328 + ], + "type": "text", + "content": "ture extracting and matching mechanism effectively aggregates features and models the relation. To make comparisons, we split the single branch into a Siamese structure based on SyncTrack. A matching network is added to the Siamese backbone for correlating the features. The shape-aware feature learning network in V2B [19] and iterative coarse-to-fine correlation network in STNet [20] are chosen as matching networks to compare with our single-branch framework. From Table 6, the single-branch structure of SyncTrack is quite significant, and when we split the branch and add a matcher to correlate, the performance drops heavily." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 371, + 285, + 431 + ], + "blocks": [ + { + "bbox": [ + 47, + 339, + 286, + 371 + ], + "lines": [ + { + "bbox": [ + 47, + 339, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 286, + 371 + ], + "type": "text", + "content": "Table 6. Ablations of proposed Single-Branch and Siamese Structure on Car category of KITTI and nuScenes, (matches1 is from the V2B [19] and matches2 is from the STNet [20])." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 371, + 285, + 431 + ], + "lines": [ + { + "bbox": [ + 50, + 371, + 285, + 431 + ], + "spans": [ + { + "bbox": [ + 50, + 371, + 285, + 431 + ], + "type": "table", + "html": "
StructureKITTInuScenes
SuccessPrecisionSuccessPrecision
Siamese+catcher170.482.435.336.8
Siamese+catcher271.883.735.437.1
Single-Branch73.385.036.738.1
", + "image_path": "bad53b6a91bec4b6c5d5a166a2d663b3ffcc0be99f586923c4d3f3c697be330c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 286, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 286, + 624 + ], + "type": "text", + "content": "Attentive Sampling. In this paper, we integrate attentive sampling into the multi-head Transformers for selecting search region points to aggregate neighborhood features. We ablate such configuration by performing attentive sampling on template tokens and both template and search region tokens, as well as comparing with standard Transformer without attentive sampling (using random sampling and random/FPS sampling) as shown in Table 7. It can be concluded that the pattern of performing attentive sampling on search region tokens in Transformers is the best. We hypothesize that sampling template tokens attentively is meaningless as template points are target-centric, and search regions' responses include much noise from the background. Therefore, it is inefficient to down-sample template tokens based on attentive responses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 130, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 130, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 130, + 647 + ], + "type": "text", + "content": "4.6. Visualization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "In Figure 4, we present visualization results obtained from LiDAR video sequences taken from the KITTI dataset. These visualizations show the motion pattern of objects belonging to four categories within the KITTI dataset. Obviously, our SyncTrack excels in accurately tracking the in" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 89, + 543, + 195 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 545, + 88 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 545, + 88 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 545, + 88 + ], + "type": "text", + "content": "Table 7. Ablations of attentive sampling in APST on Car category of KITTI." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 89, + 543, + 195 + ], + "lines": [ + { + "bbox": [ + 309, + 89, + 543, + 195 + ], + "spans": [ + { + "bbox": [ + 309, + 89, + 543, + 195 + ], + "type": "table", + "html": "
TemplateSearch RegionSuccessPrecision
√APST ×randomXX70.982.8
✓X67.678.8
X✓73.385.0
✓✓69.681.1
√APST ×FPSXX71.182.8
✓X68.078.4
X✓73.285.0
✓✓70.282.7
", + "image_path": "e6811661c633d061d23e3e466ed143908f8eb08eee412646b7a9c54a69478a68.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 205, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 289 + ], + "type": "text", + "content": "tended target and predicting bounding boxes when compared to STNet [20]. This achievement is primarily attributed to the dynamic and abundant feature interactions that occur between the template and search region seeds in SyncTrack. These interactions enable our algorithm to effectively distinguish the foreground from the background, leading to superior performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 299, + 378, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 378, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 378, + 312 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 319, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 461 + ], + "type": "text", + "content": "In this paper, we propose SyncTrack, a novel single-branch and single-stage framework for 3D LiDAR single object tracking. SyncTrack replaces the conventional Siamese-like backbones with a single-branch one, synchronizing the feature extracting and matching without an additional matching network. Moreover, the Attentive Points-Sampling Transformer is proposed for building the backbone, and sampling search region points attentively rather than randomly. Our SyncTrack achieves good tracking performance in accuracy, efficiency, and scalability. We hope it can help motivate further research on more simple yet efficient 3D trackers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 463, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 594 + ], + "type": "text", + "content": "Limitations discussion. Compared with motion-centric tracking framework like M2Track [51], we find our SyncTrack achieves limited performance on tiny-sized and slow-moving objects like pedestrian. We attribute it to the global reasoning mechanism of self-attention. Specifically, the semantic density of small-sized objects' token is much lower, which hinders effective informative interactions between tokens when performing self-attention. The fact that transformer-based method STNet [20] outperforms CNN-based M2track on all classes except the pedestrian (Table 1) also supports this hypothesis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 605, + 365, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 365, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 365, + 616 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 624, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 310, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 310, + 624, + 545, + 668 + ], + "type": "text", + "content": "[1] Luca Bertinetto, Jack Valmadre, Joao F Henriques, Andrea Vedaldi, and Philip HS Torr. Fully-convolutional siamese networks for object tracking. In European conference on computer vision, pages 850–865. Springer, 2016. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 669, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 669, + 547, + 713 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9961" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 140 + ], + "type": "text", + "content": "[3] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: A simplified architecture for visual object tracking. arXiv preprint arXiv:2203.05328, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 288, + 185 + ], + "type": "text", + "content": "[4] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8126-8135, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 288, + 219 + ], + "type": "text", + "content": "[5] Yubo Cui, Zheng Fang, Jiayao Shan, Zuoxu Gu, and Sifan Zhou. 3d object tracking with transformer. arXiv preprint arXiv:2110.14921, 2021. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 263 + ], + "type": "text", + "content": "[6] Xingping Dong, Jianbing Shen, Dongming Wu, Kan Guo, Xiaogang Jin, and Fatih Porikli. Quadruplet network with one-shot learning for fast visual object tracking. IEEE Transactions on Image Processing, 28(7):3516-3527, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 265, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 265, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 265, + 288, + 330 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 332, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 332, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 332, + 288, + 376 + ], + "type": "text", + "content": "[8] Heng Fan and Haibin Ling. Siamese cascaded region proposal networks for real-time visual tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7952-7961, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 378, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 378, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 53, + 378, + 288, + 421 + ], + "type": "text", + "content": "[9] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3d-siamrpn: an end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 1, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[10] Peng Gao, Teli Ma, Hongsheng Li, Ziyi Lin, Jifeng Dai, and Yu Qiao. Mcmae: Masked convolution meets masked autoencoders. Advances in Neural Information Processing Systems, 35:35632-35644, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "type": "text", + "content": "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "type": "text", + "content": "[12] Silvio Giancola, Jesus Zarzar, and Bernard Ghanem. Leveraging shape completion for 3d siamese tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1359-1368, 2019. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "text", + "content": "[13] Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. Siamcar: Siamese fully convolutional classification and regression for visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6269-6277, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[14] Sam Hare, Stuart Golodetz, Amir Saffari, Vibhav Vineet, Ming-Ming Cheng, Stephen L Hicks, and Philip HS Torr. Struck: Structured output tracking with kernels. IEEE transactions on pattern analysis and machine intelligence, 38(10):2096-2109, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[15] Anfeng He, Chong Luo, Xinmei Tian, and Wenjun Zeng. A twofold siamese network for real-time object tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4834-4843, 2018. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 fps with deep regression networks. In European Conference on Computer Vision, pages 749-765. Springer, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[17] João F Henriques, Rui Caseiro, Pedro Martins, and Jorge Batista. High-speed tracking with kernelized correlation filters. IEEE transactions on pattern analysis and machine intelligence, 37(3):583-596, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[18] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "type": "text", + "content": "[19] Le Hui, Lingpeng Wang, Mingmei Cheng, Jin Xie, and Jian Yang. 3d siamese voxel-to-bev tracker for sparse point clouds. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 3, 6, 7, 8, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "text", + "content": "[20] Le Hui, Lingpeng Wang, Linghua Tang, Kaihao Lan, Jin Xie, and Jian Yang. 3d siamese transformer network for single object tracking on point clouds. arXiv preprint arXiv:2207.11995, 2022. 1, 3, 6, 7, 8, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 545, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 325 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[22] Bo Li, Wei Wu, Qiang Wang, Fangyi Zhang, Junliang Xing, and Junjie Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4282-4291, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "text", + "content": "[23] Bo Li, Junjie Yan, Wei Wu, Zheng Zhu, and Xiaolin Hu. High performance visual tracking with siamese region proposal network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8971-8980, 2018. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "text", + "content": "[24] Rong Li, Anh-Quan Cao, and Raoul de Charette. Coarse3d: Class-prototypes for contrastive learning in weakly-supervised 3d point cloud segmentation. arXiv preprint arXiv:2210.01784, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[25] Yang Li and Jianke Zhu. A scale adaptive kernel correlation filter tracker with feature integration. In European conference on computer vision, pages 254-265. Springer, 2014. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 516, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 516, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 516, + 545, + 569 + ], + "type": "text", + "content": "[26] Junwei Liang, Lu Jiang, Kevin Murphy, Ting Yu, and Alexander Hauptmann. The garden of forking paths: Towards multi-future trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10508-10518, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 624 + ], + "type": "text", + "content": "[27] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5725-5734, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[28] Xiankai Lu, Chao Ma, Jianbing Shen, Xiaokang Yang, Ian Reid, and Ming-Hsuan Yang. Deep object tracking with shrinkage loss. IEEE transactions on pattern analysis and machine intelligence, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[29] Jiageng Mao, Xiaogang Wang, and Hongsheng Li. Interpolated convolutional networks for 3d point cloud understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1578-1587, 2019. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9962" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[30] Hyeonseob Nam and Bohyung Han. Learning multi-domain convolutional neural networks for visual tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4293-4302, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 161 + ], + "type": "text", + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "type": "text", + "content": "[32] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2b: Point-to-box network for 3d object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6329-6338, 2020. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "text", + "content": "[33] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. Ptt: Point-track-transformer module for 3d single object tracking in point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 1, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 326 + ], + "type": "text", + "content": "[34] Jack Valmadre, Luca Bertinetto, Joao Henriques, Andrea Vedaldi, and Philip HS Torr. End-to-end representation learning for correlation filter based tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2805-2813, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 370 + ], + "type": "text", + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 415 + ], + "type": "text", + "content": "[36] Mengmeng Wang, Yong Liu, and Zeyi Huang. Large margin object tracking with circulant feature maps. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4021-4029, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 469 + ], + "type": "text", + "content": "[37] Mengmeng Wang, Teli Ma, Xingxing Zuo, Jiajun Lv, and Yong Liu. Correlation pyramid network for 3d single object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3215-3224, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "text", + "content": "[38] Ning Wang, Wengang Zhou, Jie Wang, and Houqiang Li. Transformer meets tracker: Exploiting temporal context for robust visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1571-1580, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 580 + ], + "type": "text", + "content": "[39] Zhoutao Wang, Qian Xie, Yu-Kun Lai, Jing Wu, Kun Long, and Jun Wang. Mlvsnet: Multi-level voting siamese network for 3d visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3101-3110, 2021. 1, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "text", + "content": "[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": "[41] Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8751-8760, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[42] Yinda Xu, Zeyu Wang, Zuoxin Li, Ye Yuan, and Gang Yu. Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines. In AAAI, pages 12549-12556, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 665 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 126 + ], + "type": "text", + "content": "[43] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10448-10457, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "text", + "content": "[44] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In European Conference on Computer Vision, pages 341-357. Springer, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 546, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 546, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 546, + 206 + ], + "type": "text", + "content": "[45] Jesus Zarzar, Silvio Giancola, and Bernard Ghanem. Efficient bird eye view proposals for 3d siamese tracking. arXiv preprint arXiv:1903.10168, 2019. 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 207, + 546, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 546, + 251 + ], + "type": "text", + "content": "[46] Zhipeng Zhang and Houwen Peng. Deeper and wider siamese networks for real-time visual tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4591-4600, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 252, + 546, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 546, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 546, + 296 + ], + "type": "text", + "content": "[47] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259-16268, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 546, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 329 + ], + "type": "text", + "content": "[48] Haojie Zhao, Gang Yang, Dong Wang, and Huchuan Lu. Deep mutual learning for visual object tracking. Pattern Recognition, 112:107796, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "type": "text", + "content": "[49] Shaochuan Zhao, Tianyang Xu, Xiao-Jun Wu, and Xue-Feng Zhu. Adaptive feature fusion for visual object tracking. Pattern Recognition, 111:107679, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 546, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 420 + ], + "type": "text", + "content": "[50] Chaoda Zheng, Xu Yan, Jiantao Gao, Weibing Zhao, Wei Zhang, Zhen Li, and Shuguang Cui. Box-aware feature enhancement for single object tracking on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13199-13208, 2021. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 420, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 547, + 487 + ], + "type": "text", + "content": "[51] Chaoda Zheng, Xu Yan, Haiming Zhang, Baoyuan Wang, Shenghui Cheng, Shuguang Cui, and Zhen Li. Beyond 3d siamese tracking: A motion-centric paradigm for 3d single object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8111-8120, 2022. 1, 2, 3, 6, 7, 9" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 487, + 546, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 546, + 552 + ], + "type": "text", + "content": "[52] Changqing Zhou, Zhipeng Luo, Yueru Luo, Tianrui Liu, Liang Pan, Zhongang Cai, Haiyu Zhao, and Shijian Lu. Pptr: Relational 3d point cloud object tracking with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8531-8540, 2022. 1, 2, 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 554, + 546, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 546, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 546, + 609 + ], + "type": "text", + "content": "[53] Jiaming Zhou, Kun-Yu Lin, Haoxin Li, and Wei-Shi Zheng. Graph-based high-order relation modeling for long-term action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8984-8993, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 609, + 546, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 546, + 665 + ], + "type": "text", + "content": "[54] Zhuangwei Zhuang, Rong Li, Kui Jia, Qicheng Wang, Yuanqing Li, and Mingkui Tan. Perception-aware multi-sensor fusion for 3d lidar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16280-16290, 2021. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9963" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/39947c2e-9c93-4c2e-8f0c-e7f92b1d1960_content_list.json b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/39947c2e-9c93-4c2e-8f0c-e7f92b1d1960_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..699448f5377018977d54e1b144311d07c6658000 --- /dev/null +++ b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/39947c2e-9c93-4c2e-8f0c-e7f92b1d1960_content_list.json @@ -0,0 +1,1676 @@ +[ + { + "type": "text", + "text": "Synthesizing Diverse Human Motions in 3D Indoor Scenes", + "text_level": 1, + "bbox": [ + 189, + 130, + 779, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kaifeng Zhao1 Yan Zhang1 Shaofei Wang1 Thabo Beeler2 Siyu Tang1 ETH Zürich 2Google", + "bbox": [ + 171, + 178, + 795, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kaifeng.zhao, yan.zhang, shaofei.wang, siyu.tang}@inf.ethz.ch thabo.beeler@gmail.com", + "bbox": [ + 210, + 219, + 759, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92394e5bf069ee29114dba10b1473a40d93f8ac25a090519db68cc9e4ffe965b.jpg", + "image_caption": [ + "Figure 1: In this work, we propose a method to generate a sequence of natural human-scene interaction events in real-world complex scenes as illustrated here. The human first walks to sit on a stool (yellow to red), then walk to another chair to sit down (red to magenta), and finally walk to and lie on the sofa (magenta to blue)." + ], + "image_footnote": [], + "bbox": [ + 78, + 286, + 890, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 597, + 313, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a novel method for populating 3D indoor scenes with virtual humans that can navigate in the environment and interact with objects in a realistic manner. Existing approaches rely on high-quality training sequences that contain captured human motions and the 3D scenes they interact with. However, such interaction data are costly, difficult to capture, and can hardly cover the full range of plausible human-scene interactions in complex indoor environments. To address these challenges, we propose a reinforcement learning-based approach that enables virtual humans to navigate in 3D scenes and interact with objects realistically and autonomously, driven by learned motion control policies. The motion control policies employ latent motion action spaces, which correspond to realistic motion primitives and are learned from large-scale motion capture data using a powerful generative motion model. For navigation in a 3D environment, we propose a scene-aware policy", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with novel state and reward designs for collision avoidance. Combined with navigation mesh-based path-finding algorithms to generate intermediate waypoints, our approach enables the synthesis of diverse human motions navigating in 3D indoor scenes and avoiding obstacles. To generate fine-grained human-object interactions, we carefully curate interaction goal guidance using a marker-based body representation and leverage features based on the signed distance field (SDF) to encode human-scene proximity relations. Our method can synthesize realistic and diverse human-object interactions (e.g., sitting on a chair and then getting up) even for out-of-distribution test scenarios with different object shapes, orientations, starting body positions, and poses. Experimental results demonstrate that our approach outperforms state-of-the-art human-scene interaction synthesis methods in terms of both motion naturalness and diversity. Code, models, and demonstrative video results are available at: https://zkf1997.github.io/DIMOS.", + "bbox": [ + 496, + 599, + 893, + 871 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14738", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Simulating how humans interact with environments plays an essential role in many applications, such as generating training data for machine learning algorithms, and simulating autonomous agents in AR/VR and computer games. Although this task is highly related to character animation in computer graphics, most existing character animation methods (e.g. [4, 5, 16]) focus on improving the realism and controllability of character movements. With the traditional character animation workflows, one can produce high-quality animations but can hardly generate autonomous and spontaneous natural human motions interacting with the surroundings in diverse plausible ways as real humans. Previous learning-based interaction synthesis methods [12, 42, 57] require simultaneously capturing human motion and scenes for supervision. However, capturing such training data is costly and challenging, resulting in a notably limited spectrum of human-scene interaction motions and difficulties in handling unseen interaction scenarios. This restriction also results in inferior motion quality of synthesized virtual humans.", + "bbox": [ + 76, + 114, + 468, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this problem, we leverage reinforcement learning (RL) [45] to solve our task. By formulating goals as rewards, perception as states, and latent variables of deep generative models as actions, we can synthesize continuous, stochastic, plausible, and spontaneous motions of virtual humans to inhabit the digital world. Although existing RL-based motion synthesis approaches (e.g. [26, 33, 60]) can effectively generate natural motions to achieve goals, their generated virtual humans can only interact with simple scenes, rather than complex environments with functional furniture and diverse objects. For example, GAMMA [60] employs generative motion primitives and a policy network that are generalizable across diverse human body shapes, but it can only synthesize waypoint-reaching locomotions. The trained digital humans are not aware of how to perform actions like sitting on a chair or lying on a sofa, and frequently inter-penetrate with the scene geometry.", + "bbox": [ + 76, + 417, + 468, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, we propose a novel framework to learn both scene and interaction-aware motion control policies for synthesizing realistic and diverse human-scene interaction motions. First, in order to improve the physical plausibility of the synthesized human motions, we design a new scene-aware policy to help virtual humans avoid collisions with scene objects. Specifically, we use a 2D occupancy-based local walkability map to incorporate scene information into the locomotion policy. In addition, we add features derived from the signed distance from body markers to the object surface and the gradient direction of the signed distance to encode the proximity between humans and objects for object interaction policies. Second, in order to achieve controllable object interactions, we provide fine-grained guidances based on surface mark", + "bbox": [ + 76, + 674, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ers [58] of a human body performing the target interactions. Specifically, we use COINS [61] to generate human bodies interacting with scene objects given the interaction semantics, and then use the body markers as the interaction guidance for motion synthesis. Combined with navigation mesh-based path-finding algorithms to generate intermediate waypoints in 3D scenes, virtual humans can autonomously reach target locations in complex environments and mimic target poses in a variety of plausible ways.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We train the policy networks in synthetic scenes consisting of randomized objects to learn generalizable scene-aware locomotion and fine-grained object interactions. With this framework, we investigate how to synthesize diverse in-scene motions consisting of locomotion, sitting, and lying. We empirically evaluate the motion realism and expressiveness of our proposed method, and compare it with state-of-the-art methods. The results show that our approach consistently outperforms the baselines in terms of diversity, physical plausibility, and perceptual scores.", + "bbox": [ + 496, + 227, + 892, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we aim to let virtual humans inhabit virtual environments, and present these contributions:", + "bbox": [ + 500, + 378, + 890, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a reinforcement learning-based framework to generate realistic and diverse motions of virtual humans in complex indoor scenes.", + "2. We propose to use body surface markers as detailed interaction goals for fine-grained human-object interaction synthesis and leverage COINS [61] to generate articulated 3D human bodies based on interaction semantics to make virtual humans controllable via interaction semantics and fine-grained body poses.", + "3. We design scene and interaction-aware policies to enable virtual humans to navigate in 3D scenes while avoiding collisions, to interact with scene objects, and to continuously perform sequences of activities in complex scenes." + ], + "bbox": [ + 511, + 417, + 890, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 664, + 640, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human motion synthesis. Generating high-quality human motion has been widely explored in computer vision and graphics. Motion graph [22] and motion matching [4, 5, 63] generate motion by searching suitable clips in datasets and blending them automatically. Starke et al. [41-43] use phase-conditioned neural networks to synthesize character animations and interaction with objects. Zhang et al. [58, 60] model motion as sequences of surface markers on the parametric SMPL-X [31] body model, and train autoregressive networks on the large-scale mocap dataset AMASS [29] in order to produce diverse motions of bodies with various shapes. Peng et al. [33] use imitation learning with a skill discovery objective to learn a general motion skill space for physically simulated characters. Tang et al.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14739", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[46] trains a motion manifold model of consecutive frames for real-time motion transition. Recent works [23, 24] propose generative methods to synthesize motions from single or a few example motion sequences. Transformer-based models have been designed to predict or generate stochastic motions conditioned on action categories [35], texts [36, 48], gaze [62], and others. More recently, motion diffusion models [1, 2, 49, 50, 54] achieve appealing performance on motion synthesis conditioned on various control signals and demonstrate flexible motion editing.", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion control and RL-based motion synthesis. Various motion control methods have been proposed to constrain body movements or guide the body to reach goals. Sampling-based motion control methods [27, 28] generate multiple samples at each step and select the samples that best match the targets. Goal-conditional generation networks [10, 17, 21] are applied for motion control. However, such methods may produce invalid results when the train-test domain gap is large. Optimization-based motion control methods [18, 53] leverage the learned generative motion model as regularization and optimize the motion latent variables to fit the decoded motion to the goals. Motion diffusion models [49, 54] implement control via classifier-free text guidance or gradually reprojecting the generated motion onto the physically plausible space at individual denoising steps. Human motions can be formulated as a Markov decision process, and synthesized and controlled via RL. Imitation learning methods [3, 30, 32, 34] trains policy networks to control humanoids to imitate reference motion and complete given tasks. Peng et al. [33] propose a skill discovery objective apart from the imitation objective to learn a latent space for general motion skills and train substream task policies leveraging the general motion space. Follow-up works [20, 47] combine the character controllers with language-based selection or finite state machine to compose more complex movements. Ling et al. [26] and Zhang et al. [60] leverage the latent variable of the generative motion model as the action, and train policies to generate high-fidelity perpetual motions.", + "bbox": [ + 75, + 261, + 472, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human-scene interaction synthesis. Synthesizing natural human-scene interactions is an appealing direction in recent years. Object affordance [8, 9, 19, 38] provides cues about how humans can interact with scene objects. Static bodies interacting with scene objects can be generated based on the scene semantics, geometries, and interaction semantics [14, 25, 56, 59, 61]. Wang et al. [52] propose to train generative models for human motion in scenes conditioned on scene pointclouds. Wang et al. [51] propose a three-stage method to synthesize human motions in scenes, which first puts static bodies in scenes, then generates global paths, and afterwards synthesizes in-between", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "motions. Starke [42] propose an autoregressive neural state machine to synthesize character animations interacting with objects. Hassan et al. [12] and Zhang et al. [57] propose similar networks as in [42], and incorporate stochastic motion generation and finer-grained motion control modules. Peng et al. [33] trains policy to control physically-simulated characters to hit a box. Hassan et al. [15] extend the method to generate interactions like carrying a box or sitting on a chair. Zhang et al. [55] learns physically simulated tennis players from broadcast videos.", + "bbox": [ + 496, + 90, + 893, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ours versus others. Our method is most similar to GAMMA [60] and SAMP [12]. GAMMA learns generalizable motion models and policies across human bodies of diverse identities and shapes, without goal-motion paired training data. Despite producing high-fidelity motions, its results are limited to locomotion in the scene, and frequently collide with scene objects. SAMP learns conditional VAEs to produce sitting and lying actions in living environments, with object-motion paired data. Its generated motion has visible artifacts such as foot-ground skating. Our method combines their merits and eliminates their individual disadvantages. We extend the RL-based framework proposed in GAMMA by incorporating fine-grained motion controls (guided by interaction semantics) and scene interaction modules, so as to generate human-scene interactions in complex daily living environments. Compared to SAMP, our produced motion is more diverse, more physically plausible, and can be guided by fine-grained body surface markers. For systematic comparisons, please refer to Section 4.", + "bbox": [ + 496, + 277, + 893, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 604, + 591, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 632, + 640, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SMPL-X [31] and body representation. We use SMPL-X to represent 3D human bodies in our work. Given shape parameter $\\beta \\in \\mathbb{R}^{10}$ and body pose parameters $\\theta \\in \\mathbb{R}^{63}$ , SMPL-X produces a posed body mesh with a fixed topology of 10475 vertices. To place the body in a scene, the root location $r \\in \\mathbb{R}^3$ and the orientation $\\phi \\in \\mathfrak{so}(3)$ w.r.t. the scene coordinates are additionally needed. Since facial expressions and hand gestures are not our focus, we leave their parameters as the default values. In addition, we follow [58, 60] to represent the body in motion by the SSM_67 body surface marker placement. A motion sequence is then formulated as $X = \\{\\pmb{x}_1, \\dots, \\pmb{x}_N\\}$ , where $N$ is the length of motion and $\\pmb{x}_i \\in \\mathbb{R}^{67 \\times 3}$ denotes the body marker 3D locations at frame $i$ . The marker locations are relative to a canonical coordinate frame centered at the body pelvis in the first frame.", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14740", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "GAMMA [60]. GAMMA can synthesize stochastic, perpetual, and realistic goal-reaching actions in 3D scenes. It comprises generative motion primitive models, RL-based control, and tree-based search to implement gradient-free test-time optimization. The motion primitive is formulated by a CVAE model to generate uncertain marker motions for 0.25 seconds into the future given a motion seed, followed by a MLP-based body regressor to yield SMPL-X parameters. Long-term random motion can be generated by running the motion primitive model recursively. The RL-based control is implemented by learning a policy within a simulation area. The actor-critic framework [45] and the PPO algorithm [39] are applied to update the policy network. An additional motion prior term is used to ensure the motion appears natural. During testing, the generated motion primitives are stored in a tree where only the best K primitives at each layer are preserved in order to discard low-quality sampling results.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "COINS[61]. COINS generates physically plausible static human-scene interactions with instance-level semantic control. Given the point cloud of an object and action labels, COINS can generate static bodies interacting with the given object based on the specified action, e.g., sitting, lying, or touching. COINS leverages transformer-based generative models trained on a human-scene interaction dataset [13] to first generate a plausible body pelvis for interaction and then the posed body. The generated bodies are further optimized to improve the physical plausibility and to match the predicted action-dependent contact areas with objects. Such generated static bodies capture the characteristics of the human-scene interaction process and can be used as fine-grained interaction guidance.", + "bbox": [ + 75, + 382, + 470, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. RL-based Framework to Inhabit the Virtual", + "text_level": 1, + "bbox": [ + 76, + 604, + 452, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, we propose a motion synthesis framework that enables virtual humans to navigate in complex indoor scenes and interact with various scene objects, e.g., sitting on a chair. Compared to the GAMMA framework [60], our method incorporates scene information into the states to better handle complex human-scene interactions. Also, we use body markers as goals to provide fine-grained guidance on how to drive the body for the target interactions. With modularized path-finding methods and static person-scene interaction generation methods, our framework can synthesize realistic human motions in complex 3D environments. In our work, we use COINS [61] to generate static person-scene interactions from interaction semantics given as 'action-object' pairs. The walking path can be either generated by hand, or by automatic path-finding algorithms like A* [11, 40].", + "bbox": [ + 75, + 628, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We formulate our motion synthesis tasks with reinforcement learning. At each time step, a virtual human perceives", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6ec21d51e278f04132cbdb51576586bf4f954801548852095d8834d3b7535a56.jpg", + "image_caption": [ + "Figure 2: Illustration of our proposed human-scene interaction synthesis framework, which consists of learned motion primitive (actions), alongside locomotion and interaction policies generating latent actions conditioned on scenes and interaction goals. By integrating navigation mesh-based path-finding and static human-scene interaction generation methods, we can synthesize realistic motion sequences for virtual humans with fine-grained controls." + ], + "image_footnote": [], + "bbox": [ + 526, + 93, + 883, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "its state $s_t$ in the environment and samples an action $\\pmb{a}_t$ from its policy model $\\pi(\\pmb{a}_t | s_t)$ . Based on its motion model, it advances its motion state, and obtains a new perception state $s_{t+1}$ . A reward $r_t = r(s_t, \\pmb{a}_t, s_{t+1})$ is calculated, tailored to different tasks.", + "bbox": [ + 496, + 492, + 890, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The motion model and the action. We leverage the CVAE-based generative motion primitive [60] as our motion model, and use its latent variables as actions. We train the model conditioned on 1 or 2 past frames using the combination of the SAMP [12] and AMASS [29] motion capture datasets, to learn a latent motion primitive space covering motion skills for human-scene interactions. Each latent variable $z$ in the motion primitive space is regarded as an action and can be decoded to a short clip of motion.", + "bbox": [ + 496, + 585, + 892, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The state. The state is formulated by", + "bbox": [ + 500, + 739, + 759, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {t} = \\left(\\boldsymbol {X} _ {s}, I, G\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 767, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X_{s} \\in \\mathbb{R}^{M \\times 67 \\times 3}$ is the body markers motion seed that represents a motion history of $M$ frames. $I$ and $G$ denote the person-scene interaction feature and the goal-reaching feature, respectively. The interaction feature and goal-reaching feature vary among the locomotion and object interaction tasks. We introduce the detailed formulation in Sec. 3.3 and 3.4.", + "bbox": [ + 496, + 792, + 893, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14741", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The rewards. The rewards evaluate how well the virtual human performs locomotion and fine-grained object interaction tasks. We formulate rewards as", + "bbox": [ + 76, + 90, + 468, + 135 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr = r _ {\\text {g o a l}} + r _ {\\text {c o n t a c t}} + r _ {\\text {p e n e}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 148, + 468, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r_{\\text{goal}}$ , $r_{\\text{contact}}$ , and $r_{\\text{pene}}$ represent the rewards for goal-reaching, foot-ground contact, and penetration avoidance, respectively. Specifically, the contact reward $r_{\\text{contact}}$ encourages foot-floor contact and discourages foot skating, and is defined as:", + "bbox": [ + 76, + 172, + 468, + 247 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\text {c o n t a c t}} = e ^ {- \\left(\\left| \\min x _ {z} \\right| - 0. 0 5\\right) _ {+}} \\cdot e ^ {- \\left(\\min \\| x _ {v e l} \\| _ {2} - 0. 0 7 5\\right) _ {+}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 253, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F$ is the set of foot markers, $\\pmb{x}_z$ is the height of the markers, $\\pmb{x}_{vel}$ is the velocity of the markers, and $(\\cdot)_+$ denotes clipping negative values. There are tolerance thresholds of $0.05\\mathrm{m}$ for foot-floor distance and $0.075\\mathrm{m/s}$ for skating, following GAMMA [60]. The other two reward terms are action-specific and introduced in Sec. 3.3 and 3.4.", + "bbox": [ + 76, + 282, + 468, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy network and training. We use the actor-critic algorithm [45] to learn the policy, where a policy network and a value network are trained jointly. The policy network generates a diagonal Gaussian distribution representing the stochastic action distribution given a state, while the value network outputs the value estimation for each state.", + "bbox": [ + 76, + 391, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Like in GAMMA [60], these two networks are jointly trained by minimizing:", + "bbox": [ + 76, + 481, + 468, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {P P O} + \\mathbb {E} \\left[ \\left(r _ {t} - V (s _ {t})\\right) ^ {2} \\right] \\\\ + \\alpha \\operatorname {K L} - \\operatorname {d i v} (\\pi (z | s) | | \\mathcal {N} (0, I)), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 520, + 380, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the first term is the PPO [39] loss, the second updates the value estimation of the critic networks, and the third Kullback-Leibler divergence term regularizes motion in the latent space [60].", + "bbox": [ + 76, + 565, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tree sampling for test-time optimization. Given the stochastic nature of our Gaussian policies, sampling motions from the generated action distributions can yield motion primitive results of various qualities. Therefore, we follow [60] to use tree-based sampling during inference to discard motion primitives with inferior goal-reaching and scene interaction scores. Specifically, we sample multiple latent actions at each time step and selectively keep the best K samples, utilizing the same rewards used to train the policies as the selection criteria. This tree-sampling technique yields improved synthesis results of higher quality.", + "bbox": [ + 76, + 643, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following, we elaborate on the design of states and rewards tailored to different actions, namely locomotion and fine-grained object interaction. By combining the learned policies, long-term and coherent motions can be composed by rolling out the initial body and switching between the locomotion and object interaction stages.", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a373935eda6a2537d67285d94c779f0187ce05f72ee8bdced70a8d92d0c566eb.jpg", + "image_caption": [ + "Figure 3: Illustration of the scene-aware locomotion policy network. The locomotion policy state consists of the body markers, the goal-reaching feature of normalized direction vectors from markers to the goal pelvis, and the interaction feature of a 2D binary map indicating the walkability (red: non-walkable area, blue: walkable area) of the local $1.6m \\times 1.6m$ square area. The locomotion policy network employs the actor-critic architecture and shares the state encoder." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 885, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Scene-Aware Locomotion Synthesis", + "text_level": 1, + "bbox": [ + 500, + 396, + 807, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Navigating in cluttered scenes here means the human body moving to a target while avoiding collisions with scene objects. Our key idea is to incorporate the walkability information of the surrounding environment into the states and use collision rewards to train the locomotion policy in order to avoid scene collisions. Specifically, we represent the walkability of the environment surrounding the human agent using a 2D binary map $\\mathcal{M} \\in \\{0,1\\}^{16 \\times 16}$ as illustrated in Fig. 3.", + "bbox": [ + 498, + 419, + 890, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The walkability map is defined in the human's local coordinates and covers a $1.6m \\times 1.6m$ area centered at the body pelvis and aligned with the body facing orientation. It consists of a $16 \\times 16$ cell grid where each cell stores a binary value indicating whether this cell is walkable or not. This local walkability map enables the policy to sense surrounding obstacles.", + "bbox": [ + 498, + 555, + 890, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Referring to Eq. 1, the person-scene interaction feature is specified by", + "bbox": [ + 498, + 661, + 890, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI = \\operatorname {v e c} (\\mathcal {M}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 691, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in which $\\mathrm{vec}(\\cdot)$ denotes vectorization. The goal-reaching feature is specified by", + "bbox": [ + 498, + 710, + 890, + 741 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG = \\left(\\tilde {\\boldsymbol {g}} _ {p} - \\boldsymbol {X} _ {s}\\right) _ {n}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 750, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_s$ and $\\tilde{\\pmb{g}}_p \\in \\mathbb{R}^{M \\times 67 \\times 3}$ are the body marker seed representing $M$ frames history of motion, and the broadcasted target pelvis location relative to the body-centered canonical coordinate, respectively. $(\\tilde{\\pmb{g}}_p - \\mathbf{X}_s)_n$ denotes the normalized vectors pointing from each marker to the goal pelvis.", + "bbox": [ + 498, + 771, + 890, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The rewards contributing to Eq. 2 are defined as", + "bbox": [ + 517, + 863, + 836, + 877 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {p e n e} = e ^ {- | \\mathcal {M} _ {0} \\cap \\mathcal {B} _ {x y} (X) |}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 883, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14742", + "bbox": [ + 478, + 924, + 517, + 935 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{M}_0$ denotes the non-walkable cells in the walkability map, $\\mathcal{B}_{xy}(\\cdot)$ denotes the 2D bounding box of the body markers $X$ , $\\cap$ denotes their intersection, and $|\\cdot|$ denotes the number of non-walkable cells overlapping with the human bounding box.", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\text {g o a l}} = r _ {\\text {d i s t}} + r _ {\\text {o r i}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 184, + 468, + 200 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nr _ {d i s t} = 1 - \\left(\\left\\| \\boldsymbol {p} - \\boldsymbol {g} _ {p} \\right\\| _ {2} - 0. 0 5\\right) _ {+}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 223, + 468, + 242 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nr _ {o r i} = \\frac {\\left\\langle \\boldsymbol {o} , \\boldsymbol {g} _ {p} - \\boldsymbol {p} \\right\\rangle}{2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 260, + 468, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $r_{dist}$ encourages the body pelvis $p$ to be close to the pelvis goal $g_p$ and $r_{ori}$ encourages the body facing direction $o$ to be aligned with the direction from the current body pelvis $p$ to the pelvis goal $g_p$ .", + "bbox": [ + 76, + 297, + 468, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Fine-grained Object Interaction Synthesis", + "text_level": 1, + "bbox": [ + 76, + 368, + 434, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To synthesize fine-grained human-object interactions, e.g. sitting on a chair or lying on a sofa, we use body marker goals as guidance, and model the proximity between the body surface and the scene object in a compact way. The goal marker sets can be generated by static person-scene interaction methods such as [14, 56, 61]. We use COINS [61] to generate the static goal interaction body for its performance and controllability of interaction semantics.", + "bbox": [ + 75, + 391, + 468, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the marker-based goal guidance, we incorporate the proximity relations between humans and objects into the states. Specifically, we use the signed distance from each marker to the surface of the scene object, as well as the gradient direction of the signed distance to represent the proximity relationship, as illustrated in Fig. 4. Both the signed distance and its gradient direction are calculated using the object's signed distance field (SDF) $\\Psi_{O}$ .", + "bbox": [ + 75, + 513, + 468, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Referring to Eq. 1, the person-scene interaction feature is formulated as", + "bbox": [ + 76, + 635, + 468, + 662 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI = \\left[ \\Psi_ {O} \\left(\\boldsymbol {X} _ {s}\\right), \\nabla \\Psi_ {O} \\left(\\boldsymbol {X} _ {s}\\right) \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 678, + 468, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in which $\\Psi_O\\in \\mathbb{R}^{M\\times 67}$ and $\\nabla \\Psi_{O}\\in \\mathbb{R}^{M\\times 201}$ denote the SDF values and the gradient at each marker location in the $M$ frames, respectively, and $[\\cdot ,\\cdot ]$ denotes feature concatenation. The goal-reaching feature is formulated as", + "bbox": [ + 76, + 705, + 468, + 767 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nG = \\left[ \\left(\\tilde {\\boldsymbol {g}} _ {m} - \\boldsymbol {X} _ {s}\\right) _ {n}, \\| \\tilde {\\boldsymbol {g}} _ {m} - \\boldsymbol {X} _ {s} \\| _ {2} \\right], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 780, + 468, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in which $\\mathbf{X}_s \\in \\mathbb{R}^{M \\times 67 \\times 3}$ denotes the body markers seed, $\\tilde{\\pmb{g}}_m \\in \\mathbb{R}^{M \\times 67 \\times 3}$ denotes the broadcasted goal body markers, $(\\tilde{\\pmb{g}}_m - \\pmb{X}_s)_n$ denotes the normalized vector representing the direction from each marker to the corresponding target marker, $\\| \\tilde{\\pmb{g}}_m - X_s \\|_2$ denotes the distance from each marker to target marker.", + "bbox": [ + 76, + 808, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c4e160cfb625594ef9c430e44de2487f539b9e80123502c95f3b5e8007a25ed4.jpg", + "image_caption": [ + "Figure 4: Illustration of the object interaction policy network. The interaction policy state consists of the body markers, the goal-reaching features of both distance and direction from current markers to the goal markers, and the interaction features of the signed distances from each marker to the object surfaces and the signed distance gradient at each marker location. Such interaction features encode the human-object proximity relationship." + ], + "image_footnote": [], + "bbox": [ + 500, + 87, + 885, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The interaction policy is trained using the reward defined in Eq. 2 with the following interaction-specific goal reward and penetration reward:", + "bbox": [ + 498, + 391, + 890, + 436 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\text {g o a l}} = 1 - \\left(\\| \\boldsymbol {x} - \\boldsymbol {g} _ {m} \\| _ {2} - 0. 0 5\\right) _ {+}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 450, + 890, + 468 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nr _ {p e n e} = e ^ {- \\frac {1}{| V |} \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {| V |} \\left| \\left(\\Psi_ {O} (\\boldsymbol {v} _ {t i})\\right) _ {-} \\right|}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 484, + 890, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $|V|$ being the SMPL-X mesh vertices, $T$ denotes the number of frames in each motion primitive (equals to 10 in our study). The distance reward encourages the final frame body markers $x$ to be close to the goal body markers $g_{m}$ . The penetration reward penalizes all the body vertices within a motion primitive that have negative SDF values. We use body vertices instead of joints because human-object contact happens on the body surface and can be better detected using vertices.", + "bbox": [ + 496, + 513, + 890, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, we train the interaction policy with a mixture of 'sit/lie down' and 'stand up' tasks. This training scheme enables the human agent to also learn how to stand up and transit from object interaction back to locomotion, which enables the synthesis of a sequence of interaction activities as in Fig. 1.", + "bbox": [ + 496, + 650, + 890, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 753, + 624, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motion Datasets. We combine the large-scale motion capture dataset AMASS [29] with SAMP [12] motion data to train the motion primitive models. Each sequence is first subsampled to 40 FPS and then split into 10 frames and 100 frames clips. Each motion clip is canonicalized using the first frame body. Specifically, we select AMASS sequences annotated with 'sit' or 'lie' in BABEL [37] and all motion data of SAMP to train the motion primitive model. We train", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14743", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/66d2e9bfaef14894623e45c6aa22439f0de79e7fed4b183f916e198f4c0ed074.jpg", + "table_caption": [ + "Table 1: Evaluation of the locomotion task. The up/down arrows denote the score is the higher/lower the better and the best results are in boldface." + ], + "table_footnote": [], + "table_body": "
time ↓avg. dist ↓contact ↑loco pene ↑
SAMP [12]5.970.140.840.94
GAMMA [60]3.870.030.940.94
Ours6.430.040.990.95
", + "bbox": [ + 81, + 145, + 459, + 199 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "separate motion primitive models for locomotion and interactions using task-related data. We observe that extending SAMP motion data with AMASS dataset is the key to making interaction policies work.", + "bbox": [ + 75, + 227, + 468, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Policy Training Environments. To train the scene-aware locomotion policy, we randomly generate synthetic cluttered scenes consisting of random objects from ShapeNet [6]. Random initial and target location pairs are sampled in the walkable areas using navigation meshes to train the locomotion policy. To train the interaction policy, we use the static person-object interaction data from PROX [13] and retargeted to ShapeNet objects. We first randomly sample a furniture and interaction body goal from the retargeted PROX data. Then we sample the initial body with random poses and locations in front of the object to train the interaction policy. We also randomly swap the initial and goal body to learn both 'sit/lie down' and 'stand up' motions. Please refer to Supp. Mat. for more details.", + "bbox": [ + 75, + 316, + 472, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Locomotion in 3D Scenes", + "text_level": 1, + "bbox": [ + 76, + 542, + 307, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We randomly generate test scenes for locomotion in the same way as the training scenes. The virtual human is instructed to move from the random start point to the random target point while avoiding penetration with scene objects.", + "bbox": [ + 75, + 569, + 468, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines and metrics. We compare our method with SAMP [12] and GAMMA [60] for locomotion. The SAMP results are recorded by running the released Unity demo. The start and termination are manually determined so the reported completion time may be slightly higher than the actual time due to human response time. The evaluation metrics for locomotion include: 1) time from start point to target point or reaching the time limit, measured in seconds. 2) the average distance from the final body to the targets, measured in meters. 3) foot contact score encouraging the lowest feet joints on the floor and discouraging foot skating as defined in Eq. 3. We use body joints instead of markers to calculate the contact score because the marker set annotation for the SAMP body is missing. 4) locomotion penetration score indicating the percentage of body vertices that are inside the walkable areas according to the navigation mesh.", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23cac8bda4be6cdde7c954ad3f3a9963844c75e07d4db5031a16b8ee855b88bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 88, + 694, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba5f7e324de279aab27e3347ff8ac17cb880930c85e80c848216f4baf9d18385.jpg", + "image_caption": [ + "Figure 5: Demonstration of locomotion tasks where GAMMA [60] (left) collides with the obstacles (red bodies) while our scene-aware locomotion policy (right) avoids collision. The yellow circles denote the specified waypoints." + ], + "image_footnote": [], + "bbox": [ + 519, + 202, + 694, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6c56c14fc6383ab105da0b9de95c6944b39d916d8e346c64581cff149f52559.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 88, + 872, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f64c7f963b62d65641324cc8ba336fdf83781c957d24b4af6b52440f73972df4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 202, + 872, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Table 1 shows the empirical evaluation results. Our method achieves a higher contact score (0.99) than both GAMMA (0.94) and SAMP (0.84) which indicates better foot-floor contact and less foot skating. Moreover, our method achieves the highest penetration score which indicates our scene-aware policy can better avoid scene collisions. Fig. 5 shows examples of locomotion tasks where GAMMA collides into the scenes while our scene-aware policy successfully avoids penetration. Compared to GAMMA trained in the same environments, we observe our scene-aware policy learns more conservative behavior with lower moving speed and spent more time (6.43s) walking to the target locations, just like a human afraid of stepping on surrounding traps. All the methods can reach the target location within a reasonable distance.", + "bbox": [ + 496, + 409, + 893, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Fine-Grained Human-Object Interaction", + "text_level": 1, + "bbox": [ + 500, + 645, + 851, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the object interaction task on 10 unseen objects (3 armchairs, 3 straight chairs, 3 sofas, 1 L-sofa) from ShapeNet [6]. We use the object size annotation of ShapeNet and manually filter unrealistic-sized objects. The virtual human is randomly placed in front of the target object and then instructed to perform the interaction, stay for around 2 seconds, and then stand up. We evaluate two interactions of sitting and lying separately.", + "bbox": [ + 496, + 669, + 893, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines and metrics. We compare our method with SAMP [12] for the object interaction task. The evaluation metrics for the interaction tasks are: 1) time of completing the object interaction task. 2) foot contact score as defined in Eq. 3. Note that the foot contact score does not always reflect the motion quality for lying tasks because the foot can", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14744", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4b0570ed2641973cc4c8c8273edae0d8b2bbd80d3e7a81a80fc28ef324564287.jpg", + "table_caption": [ + "Table 2: Evaluation of the interaction tasks. The up/down arrows denote the score is the higher/ lower the better and the best results are in boldface." + ], + "table_footnote": [], + "table_body": "
time ↓contact ↑pene. mean ↓pene. max ↓
SAMP [12] sit8.630.8911.9145.22
Ours sit4.090.971.9110.61
SAMP [12] lie12.550.7344.77238.81
Ours lie4.200.789.9044.61
", + "bbox": [ + 81, + 145, + 467, + 209 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/374b540cd5b261e574e784cec96c564c94f1b956630047236833d26581418edb.jpg", + "image_caption": [ + "Figure 6: Demonstration of various generated object interactions. Each row shows generated interactions with the same object starting from random initial body locations and orientations. Colors from yellow to red denote time." + ], + "image_footnote": [], + "bbox": [ + 84, + 224, + 459, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "often be off the floor during lying. 3) interaction penetration score for each frame is defined as:", + "bbox": [ + 75, + 542, + 468, + 571 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i n t e r - p e n e} = \\sum_ {v _ {i} \\in V} | (\\Psi_ {O} (v _ {i})) _ {-} |, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 584, + 468, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $\\Psi_{O}$ is the object signed distance field, $(\\cdot)_{-}$ clips all positive distance values to zero, and $V$ is the body vertivces. We show both the average penetration over time and the maximum penetration in one sequence.", + "bbox": [ + 75, + 625, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. Tab. 2 shows the evaluation results. Our method achieves a significantly higher foot contact score, indicating more natural motion. Our method also achieves lower mean and maximum penetration, which means our method generates more physically plausible results. Moreover, our method can complete the interaction tasks much faster than SAMP. This is because SAMP does not generalize well to unseen random body initialization and needs a longer time to start performing interactions. Qualitative results demonstrating the various object interactions generated by our method are shown in Fig. 6. Our object interaction policy generalizes to random initial body locations and orientations, as well as novel objects of various shapes.", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2214b1eaecfa4674311e718587f630b48b5b01d0e7129fdcfa01d93f12ba717e.jpg", + "image_caption": [ + "Figure 7: Our method can synthesize natural human movements in scene reconstructions that are significantly different from the training scenes. We show the synthesis results in one Paris street [7] (left figure) and one indoor room [13] (right figure). Transparent to solid colors denote time." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 697, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a28c54f5987d85f635beb86f2f0649b3d61b9e6f3c7579ab603cdaacdf649914.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 88, + 887, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aaf976264e7d4a1cc707adccbd805c716ddb753692ae2fac988aad8ad6cb1661.jpg", + "table_caption": [ + "Table 3: Evaluation of interaction sequences synthesis. Here 'contact' denotes the foot contact score, 'loco. pene.' is the percentage of body vertices inside the walkable areas, 'inter. pene. mean/max' denotes the mean and maximum penetration with interaction objects, and 'perceptual' denotes the ratio of being chosen as perceptually more natural. The best results are in boldface." + ], + "table_footnote": [], + "table_body": "
SAMP [12]Ours
contact ↑0.870.96
loco. pene. ↑0.620.72
inter. pene. mean↓15.613.40
inter. pene. max ↓101.2539.68
perceptual. ↑0.150.85
", + "bbox": [ + 557, + 440, + 831, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Interaction Sequences in 3D Scenes", + "text_level": 1, + "bbox": [ + 500, + 570, + 805, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Humans often continuously perform sequences of interactions with various objects in complex real-world scenes as shown in Fig. 1. Such interaction sequences are combinations of alternating locomotions and fine-grained interactions which we evaluate in Sec. 4.1 and Sec. 4.2 respectively. We conduct the empirical evaluation in real scene scans from Replica [44] and compare our method with SAMP. We select a list of interactable objects in the scene and instruct the virtual human to walk to the objects to perform interactions one by one. The evaluation metrics include the foot contact score, locomotion penetration score, and the mean and maximum interaction penetration score. We also conducted a perceptual study where participants are shown a side-by-side comparison of results from two methods and asked to choose the one perceptually more natural. We report the rate of being chosen as the better.", + "bbox": [ + 496, + 595, + 890, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The evaluation results are shown in Tab. 3. Our method generates high-quality results of human-scene interactions in cluttered environments. Our results achieve higher contact scores and less penetration with the scenes compared", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14745", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ad54229f9a833f2876b4e05fe0eee95fa873d2ebbf1a944755d71543e0434989.jpg", + "table_caption": [ + "Table 4: Rewards ablation studies results, where '-Contact' and '-Penetration' denote policies trained without the floor contact and penetration avoidance reward, respectively." + ], + "table_footnote": [], + "table_body": "
time ↓contact ↑pene. mean ↓pene. max ↓
Ours4.090.971.9110.61
- Contact3.750.8819.4960.10
- Penetration4.030.9516.8045.50
", + "bbox": [ + 78, + 145, + 465, + 199 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to SAMP. In addition, our method generates perceptually more natural results according to perceptual study.", + "bbox": [ + 76, + 223, + 467, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 263, + 241, + 279 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We perform ablation studies on the foot-ground contact reward and penetration avoidance reward of the interaction policies to substantiate the significance of these rewards. We train two ablation interaction policies using identical networks and environments, differing only in the exclusion of one of these rewards in each case. The quantitative metrics are reported in Tab. 4. The removal of the penetration avoidance reward yields a marked escalation in detected human-scene penetration. The removal of the footfloor contact reward yields notably both inferior contact and penetration scores, accompanied by observed erratic synthesized motions that either remain suspended in the air or penetrate the floor. These empirical findings underscore the crucial role played by the foot-floor contact and penetration avoidance rewards in the learning of interaction policies.", + "bbox": [ + 75, + 287, + 468, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 76, + 527, + 323, + 542 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations and Future Work. Our current method has various limitations that could be improved in future works. First, our method does not fully resolve penetrations with scene objects and floors since our method only uses rewards to encourage avoiding penetration with scenes, which does not impose hard constraints for penetrations. The combination of our method with physics simulation holds the potential to effectively address and resolve penetration issues. Moreover, the lying motions generated by our method are not as natural as the sitting motions because the motion primitive model fails to learn a comprehensive action space for lying given limited training data. Specifically, the available AMASS motion capture data for lying (167 seconds) is significantly less than the data for sitting (5K seconds). To overcome this issue, we aim to explore more data-efficient learning methods and scalable methods to collect human-scene interaction data. Furthermore, our locomotion policy is now limited to flat-floor scenes due to its reliance on the 2D occupancy-based walkability map. However, to broaden the applicability of our approach to more complex scenes such as uneven outdoor terrains and multi-floor buildings requiring walking upstairs, it will be necessary to replace the walkability map with a more suitable environment sens", + "bbox": [ + 75, + 551, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1b3ba8a7622d1b69fa19703f59b2f29b84b6eb9eea06cc7768ef58a2f8e989b0.jpg", + "image_caption": [ + "Inter-penetrations", + "Figure 8: Limitations. Penetrations remain observable in our results (left) since our method only encourages avoiding collision with rewards. Furthermore, the absence of sufficient training data for lying motions impedes the motion primitive model's ability to acquire a comprehensive lying motion space, leading to degraded motion (right)." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 686, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4e843826538511eb4c3b5a33b88fcca3d7d00b20dccde7f5f1bf36a8d6e79063.jpg", + "image_caption": [ + "Unnatural poses" + ], + "image_footnote": [], + "bbox": [ + 687, + 90, + 888, + 203 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ing mechanism. In addition, our method is restricted to interactions with static scenes. However, in the real world, humans are exposed to dynamic interactions with movable objects and scenes involving autonomous agents including other humans, animals, and vehicles. Extension to dynamic interactions will enable tackling a broader range of intricate and dynamic applications that mirror the complexities inherent in the real world.", + "bbox": [ + 496, + 354, + 890, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we leverage reinforcement learning to establish a framework to synthesize diverse human motions in indoor scenes, which is stochastic, realistic, and perpetual. The proposed method has large potential to improve many applications such as daily-living activity simulation, synthetic data creation, architecture design, and so on. Compared to existing methods, our method realizes fine-grained control by using body surface keypoints as targets, and achieves autonomous body-scene collision avoidance by incorporating scene information into the states and the rewards. Experiments show that our method effectively enables virtual humans to inhabit the virtual, and outperforms baselines consistently.", + "bbox": [ + 496, + 493, + 892, + 690 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. We sincerely acknowledge the anonymous reviewers for their insightful suggestions. This work was supported by the SNSF project grant 200021 204840 and SDSC PhD fellowship. Lingchen Yang provided valuable suggestions on visualization.", + "bbox": [ + 496, + 708, + 890, + 785 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14746", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. ISBN: 0730-0301 Publisher: ACM New York, NY, USA. 3", + "[2] Tenglong Ao, Zeyi Zhang, and Libin Liu. GestureDiffuCLIP: Gesture diffusion model with CLIP latents. arXiv preprint arXiv:2303.14613, 2023. 3", + "[3] Kevin Bergamin, Simon Clavet, Daniel Holden, and James Richard Forbes. DReCon: data-driven responsive control of physics-based characters. ACM Transactions On Graphics (TOG), 38(6):1-11, 2019. Publisher: ACM New York, NY, USA. 3", + "[4] Michael Buttner. Machine learning for motion synthesis and character control in games. Proc. of I3D 2019, 2, 2019. 2", + "[5] Michael Böttner and Simon Clavet. Motion matching-the road to next gen animation. Proc. of Nucl. ai, 2015(1):2, 2015. 2", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, and Hao Su. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 7", + "[7] Jean-Emmanuel Deschaud, David Duque, Jean Pierre Richa, Santiago Velasco-Forero, Beatrix Marcotegui, and François Goulette. Paris-carla-3d: A real and synthetic outdoor point cloud dataset for challenging tasks in 3d mapping. Remote Sensing, 13(22), 2021. 8", + "[8] Helmut Grabner, Juergen Gall, and Luc Van Gool. What makes a chair a chair? pages 1529-1536. IEEE, 2011. 3", + "[9] Abhinav Gupta, Scott Satkin, Alexei A Efros, and Martial Hebert. From 3d scene geometry to human workspace. pages 1961-1968. IEEE, 2011. 3", + "[10] Ikhsanul Habibie, Daniel Holden, Jonathan Schwarz, Joe Yearsley, and Taku Komura. A recurrent variational autoencoder for human motion synthesis. 2017. 3", + "[11] Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100-107, 1968. Publisher: IEEE. 4", + "[12] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. pages 11374-11384, 2021. 2, 3, 4, 6, 7, 8", + "[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3D human pose ambiguities with 3D scene constraints. pages 2282-2292, 2019. 4, 7, 8", + "[14] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3D scenes by learning human-scene interaction. pages 14708-14718, 2021. 3, 6", + "[15] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing Physical Character-Scene Interactions, Feb. 2023. arXiv:2302.00883 [cs]. 3", + "[16] Daniel Holden, Ouussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Graphics (TOG), 39(4):53-1, 2020. Publisher: ACM New York, NY, USA. 2", + "[17] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. Publisher: ACM New York, NY, USA. 3", + "[18] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. Publisher: ACM New York, NY, USA. 3", + "[19] Ruizhen Hu, Zihao Yan, Jingwen Zhang, Oliver Van Kaick, Ariel Shamir, Hao Zhang, and Hui Huang. Predictive and generative neural networks for object functionality. arXiv preprint arXiv:2006.15520, 2020. 3", + "[20] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. PADI: Language-Directed Physics-Based Character Control. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 3", + "[21] Kacper Kania, Marek Kowalski, and Tomasz Trzciński. TrajeVAE: Controllable Human Motion Generation from Trajectories. arXiv preprint arXiv:2104.00351, 2021. 3", + "[22] Lucas Kovar, Michael Gleicher, and Frédéric Pighin. Motion graphs. In ACM SIGGRAPH 2008 classes, pages 1-10. 2008. 2", + "[23] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. Ganimator: Neural motion synthesis from a single sequence. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. Publisher: ACM New York, NY, USA. 3", + "[24] Weiyu Li, Xuelin Chen, Peizhuo Li, Olga Sorkine-Hornung, and Baoquan Chen. Example-based Motion Synthesis via Generative Motion Matching. arXiv preprint arXiv:2306.00378, 2023. 3", + "[25] Xueting Li, Sifei Liu, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Putting humans in a scene: Learning affordance in 3d indoor environments. pages 12368-12376, 2019. 3", + "[26] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. Publisher: ACM New York, NY, USA. 2, 3", + "[27] Libin Liu, KangKang Yin, and Baining Guo. Improving sampling-based motion control. volume 34, pages 415-423. Wiley Online Library, 2015. Issue: 2. 3", + "[28] Libin Liu, KangKang Yin, Michiel Van de Panne, Tianjia Shao, and Weiwei Xu. Sampling-based contact-rich motion control. In ACM SIGGRAPH 2010 papers, pages 1-10. 2010. 3", + "[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. pages 5442-5451, 2019. 2, 4, 6", + "[30] Josh Merel, Leonard Hasenclever, Alexandre Galashov, Arun Ahuja, Vu Pham, Greg Wayne, Yee Whye Teh, and Nicolas Heess. Neural probabilistic motor primitives for humanoid control. arXiv preprint arXiv:1811.11711, 2018. 3", + "[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face," + ], + "bbox": [ + 501, + 92, + 890, + 890 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14747", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and body from a single image. pages 10975-10985, 2019. 2, 3", + "[32] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. Publisher: ACM New York, NY, USA. 3", + "[33] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. Publisher: ACM New York, NY, USA. 2, 3", + "[34] Xue Bin Peng, Angjoo Kanazawa, Jitendra Malik, Pieter Abbeel, and Sergey Levine. Sfv: Reinforcement learning of physical skills from videos. ACM Transactions On Graphics (TOG), 37(6):1-14, 2018. Publisher: ACM New York, NY, USA. 3", + "[35] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. pages 10985-10995, 2021. 3", + "[36] Mathis Petrovich, Michael J Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. pages 480-497. Springer, 2022. 3", + "[37] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. BABEL: bodies, action and behavior with English labels. pages 722-731, 2021. 6", + "[38] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. SceneGrok: Inferring action maps in 3D environments. ACM transactions on graphics (TOG), 33(6):1-10, 2014. Publisher: ACM New York, NY, USA. 3", + "[39] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 4, 5", + "[40] Greg Snook. Simplified 3D movement and pathfinding using navigation meshes. Game programming gems, 1(1):288-304, 2000. Publisher: Charles River Media Newton Centre, MA, USA. 4", + "[41] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. Publisher: ACM New York, NY, USA. 2", + "[42] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 2, 3", + "[43] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. Publisher: ACM New York, NY, USA. 2", + "[44] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, and Shobhit Verma. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.8", + "[45] Richard S Sutton and Andrew G Barto. Introduction to reinforcement learning, volume 135. MIT press Cambridge, 1998. 2, 4, 5", + "[46] Xiangjun Tang, He Wang, Bo Hu, Xu Gong, Ruifan Yi, Qi" + ], + "bbox": [ + 78, + 90, + 468, + 888 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "long Kou, and Xiaogang Jin. Real-time controllable motion transition for characters. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. Publisher: ACM New York, NY, USA. 3", + "[47] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. CALM: Conditional Adversarial Latent Models for Directable Virtual Characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3", + "[48] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. pages 358–374. Springer, 2022. 3", + "[49] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3", + "[50] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. EDGE: Editable Dance Generation From Music. arXiv preprint arXiv:2211.10658, 2022. 3", + "[51] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. pages 20460-20469, 2022. 3", + "[52] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. pages 9401-9411, 2021. 3", + "[53] Zhiyong Wang, Jinxiang Chai, and Shihong Xia. Combining recurrent neural networks and adversarial training for human motion synthesis and control. IEEE transactions on visualization and computer graphics, 27(1):14-28, 2019. Publisher: IEEE. 3", + "[54] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. PhysDiff: Physics-Guided Human Motion Diffusion Model. arXiv preprint arXiv:2212.02500, 2022.3", + "[55] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning Physically Simulated Tennis Skills from Broadcast Videos. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. ISBN: 0730-0301 Publisher: ACM New York, NY, USA. 3", + "[56] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. PLACE: Proximity learning of articulation and contact in 3D environments. pages 642-651. IEEE, 2020. 3, 6", + "[57] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: towards controllable human-chair interactions. pages 518-535. Springer, 2022. 2, 3", + "[58] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. pages 3372-3382, 2021. 2, 3", + "[59] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes without people. pages 6194-6204, 2020. 3", + "[60] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3D scenes. pages 20481-20491, 2022. 2, 3, 4, 5, 7", + "[61] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. pages 311-327. Springer, 2022. 2, 3, 4, 6", + "[62] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao" + ], + "bbox": [ + 501, + 92, + 890, + 888 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14748", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. pages 676-694. Springer, 2022. 3", + "[63] Fabio Zinno. MI tutorial day: From motion matching to motion synthesis, and all the hurdles in between. Proc. of GDC 2019, 2, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 172 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14749", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/39947c2e-9c93-4c2e-8f0c-e7f92b1d1960_model.json b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/39947c2e-9c93-4c2e-8f0c-e7f92b1d1960_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a5a551489af63d61cc161ce8e41237a1abef93f1 --- /dev/null +++ b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/39947c2e-9c93-4c2e-8f0c-e7f92b1d1960_model.json @@ -0,0 +1,2446 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.131, + 0.78, + 0.154 + ], + "angle": 0, + "content": "Synthesizing Diverse Human Motions in 3D Indoor Scenes" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.179, + 0.797, + 0.218 + ], + "angle": 0, + "content": "Kaifeng Zhao1 Yan Zhang1 Shaofei Wang1 Thabo Beeler2 Siyu Tang1 ETH Zürich 2Google" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.22, + 0.76, + 0.252 + ], + "angle": 0, + "content": "{kaifeng.zhao, yan.zhang, shaofei.wang, siyu.tang}@inf.ethz.ch thabo.beeler@gmail.com" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.287, + 0.891, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.529, + 0.894, + 0.575 + ], + "angle": 0, + "content": "Figure 1: In this work, we propose a method to generate a sequence of natural human-scene interaction events in real-world complex scenes as illustrated here. The human first walks to sit on a stool (yellow to red), then walk to another chair to sit down (red to magenta), and finally walk to and lie on the sofa (magenta to blue)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.598, + 0.314, + 0.615 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We present a novel method for populating 3D indoor scenes with virtual humans that can navigate in the environment and interact with objects in a realistic manner. Existing approaches rely on high-quality training sequences that contain captured human motions and the 3D scenes they interact with. However, such interaction data are costly, difficult to capture, and can hardly cover the full range of plausible human-scene interactions in complex indoor environments. To address these challenges, we propose a reinforcement learning-based approach that enables virtual humans to navigate in 3D scenes and interact with objects realistically and autonomously, driven by learned motion control policies. The motion control policies employ latent motion action spaces, which correspond to realistic motion primitives and are learned from large-scale motion capture data using a powerful generative motion model. For navigation in a 3D environment, we propose a scene-aware policy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.895, + 0.872 + ], + "angle": 0, + "content": "with novel state and reward designs for collision avoidance. Combined with navigation mesh-based path-finding algorithms to generate intermediate waypoints, our approach enables the synthesis of diverse human motions navigating in 3D indoor scenes and avoiding obstacles. To generate fine-grained human-object interactions, we carefully curate interaction goal guidance using a marker-based body representation and leverage features based on the signed distance field (SDF) to encode human-scene proximity relations. Our method can synthesize realistic and diverse human-object interactions (e.g., sitting on a chair and then getting up) even for out-of-distribution test scenarios with different object shapes, orientations, starting body positions, and poses. Experimental results demonstrate that our approach outperforms state-of-the-art human-scene interaction synthesis methods in terms of both motion naturalness and diversity. Code, models, and demonstrative video results are available at: https://zkf1997.github.io/DIMOS." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "14738" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.417 + ], + "angle": 0, + "content": "Simulating how humans interact with environments plays an essential role in many applications, such as generating training data for machine learning algorithms, and simulating autonomous agents in AR/VR and computer games. Although this task is highly related to character animation in computer graphics, most existing character animation methods (e.g. [4, 5, 16]) focus on improving the realism and controllability of character movements. With the traditional character animation workflows, one can produce high-quality animations but can hardly generate autonomous and spontaneous natural human motions interacting with the surroundings in diverse plausible ways as real humans. Previous learning-based interaction synthesis methods [12, 42, 57] require simultaneously capturing human motion and scenes for supervision. However, capturing such training data is costly and challenging, resulting in a notably limited spectrum of human-scene interaction motions and difficulties in handling unseen interaction scenarios. This restriction also results in inferior motion quality of synthesized virtual humans." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.418, + 0.47, + 0.675 + ], + "angle": 0, + "content": "To address this problem, we leverage reinforcement learning (RL) [45] to solve our task. By formulating goals as rewards, perception as states, and latent variables of deep generative models as actions, we can synthesize continuous, stochastic, plausible, and spontaneous motions of virtual humans to inhabit the digital world. Although existing RL-based motion synthesis approaches (e.g. [26, 33, 60]) can effectively generate natural motions to achieve goals, their generated virtual humans can only interact with simple scenes, rather than complex environments with functional furniture and diverse objects. For example, GAMMA [60] employs generative motion primitives and a policy network that are generalizable across diverse human body shapes, but it can only synthesize waypoint-reaching locomotions. The trained digital humans are not aware of how to perform actions like sitting on a chair or lying on a sofa, and frequently inter-penetrate with the scene geometry." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To overcome these limitations, we propose a novel framework to learn both scene and interaction-aware motion control policies for synthesizing realistic and diverse human-scene interaction motions. First, in order to improve the physical plausibility of the synthesized human motions, we design a new scene-aware policy to help virtual humans avoid collisions with scene objects. Specifically, we use a 2D occupancy-based local walkability map to incorporate scene information into the locomotion policy. In addition, we add features derived from the signed distance from body markers to the object surface and the gradient direction of the signed distance to encode the proximity between humans and objects for object interaction policies. Second, in order to achieve controllable object interactions, we provide fine-grained guidances based on surface mark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "ers [58] of a human body performing the target interactions. Specifically, we use COINS [61] to generate human bodies interacting with scene objects given the interaction semantics, and then use the body markers as the interaction guidance for motion synthesis. Combined with navigation mesh-based path-finding algorithms to generate intermediate waypoints in 3D scenes, virtual humans can autonomously reach target locations in complex environments and mimic target poses in a variety of plausible ways." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.893, + 0.378 + ], + "angle": 0, + "content": "We train the policy networks in synthetic scenes consisting of randomized objects to learn generalizable scene-aware locomotion and fine-grained object interactions. With this framework, we investigate how to synthesize diverse in-scene motions consisting of locomotion, sitting, and lying. We empirically evaluate the motion realism and expressiveness of our proposed method, and compare it with state-of-the-art methods. The results show that our approach consistently outperforms the baselines in terms of diversity, physical plausibility, and perceptual scores." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.379, + 0.892, + 0.409 + ], + "angle": 0, + "content": "In summary, we aim to let virtual humans inhabit virtual environments, and present these contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.419, + 0.891, + 0.464 + ], + "angle": 0, + "content": "1. We propose a reinforcement learning-based framework to generate realistic and diverse motions of virtual humans in complex indoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.474, + 0.892, + 0.567 + ], + "angle": 0, + "content": "2. We propose to use body surface markers as detailed interaction goals for fine-grained human-object interaction synthesis and leverage COINS [61] to generate articulated 3D human bodies based on interaction semantics to make virtual humans controllable via interaction semantics and fine-grained body poses." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.575, + 0.892, + 0.651 + ], + "angle": 0, + "content": "3. We design scene and interaction-aware policies to enable virtual humans to navigate in 3D scenes while avoiding collisions, to interact with scene objects, and to continuously perform sequences of activities in complex scenes." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.419, + 0.892, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.642, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Human motion synthesis. Generating high-quality human motion has been widely explored in computer vision and graphics. Motion graph [22] and motion matching [4, 5, 63] generate motion by searching suitable clips in datasets and blending them automatically. Starke et al. [41-43] use phase-conditioned neural networks to synthesize character animations and interaction with objects. Zhang et al. [58, 60] model motion as sequences of surface markers on the parametric SMPL-X [31] body model, and train autoregressive networks on the large-scale mocap dataset AMASS [29] in order to produce diverse motions of bodies with various shapes. Peng et al. [33] use imitation learning with a skill discovery objective to learn a general motion skill space for physically simulated characters. Tang et al." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "14739" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.244 + ], + "angle": 0, + "content": "[46] trains a motion manifold model of consecutive frames for real-time motion transition. Recent works [23, 24] propose generative methods to synthesize motions from single or a few example motion sequences. Transformer-based models have been designed to predict or generate stochastic motions conditioned on action categories [35], texts [36, 48], gaze [62], and others. More recently, motion diffusion models [1, 2, 49, 50, 54] achieve appealing performance on motion synthesis conditioned on various control signals and demonstrate flexible motion editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.701 + ], + "angle": 0, + "content": "Motion control and RL-based motion synthesis. Various motion control methods have been proposed to constrain body movements or guide the body to reach goals. Sampling-based motion control methods [27, 28] generate multiple samples at each step and select the samples that best match the targets. Goal-conditional generation networks [10, 17, 21] are applied for motion control. However, such methods may produce invalid results when the train-test domain gap is large. Optimization-based motion control methods [18, 53] leverage the learned generative motion model as regularization and optimize the motion latent variables to fit the decoded motion to the goals. Motion diffusion models [49, 54] implement control via classifier-free text guidance or gradually reprojecting the generated motion onto the physically plausible space at individual denoising steps. Human motions can be formulated as a Markov decision process, and synthesized and controlled via RL. Imitation learning methods [3, 30, 32, 34] trains policy networks to control humanoids to imitate reference motion and complete given tasks. Peng et al. [33] propose a skill discovery objective apart from the imitation objective to learn a latent space for general motion skills and train substream task policies leveraging the general motion space. Follow-up works [20, 47] combine the character controllers with language-based selection or finite state machine to compose more complex movements. Ling et al. [26] and Zhang et al. [60] leverage the latent variable of the generative motion model as the action, and train policies to generate high-fidelity perpetual motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Human-scene interaction synthesis. Synthesizing natural human-scene interactions is an appealing direction in recent years. Object affordance [8, 9, 19, 38] provides cues about how humans can interact with scene objects. Static bodies interacting with scene objects can be generated based on the scene semantics, geometries, and interaction semantics [14, 25, 56, 59, 61]. Wang et al. [52] propose to train generative models for human motion in scenes conditioned on scene pointclouds. Wang et al. [51] propose a three-stage method to synthesize human motions in scenes, which first puts static bodies in scenes, then generates global paths, and afterwards synthesizes in-between" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.245 + ], + "angle": 0, + "content": "motions. Starke [42] propose an autoregressive neural state machine to synthesize character animations interacting with objects. Hassan et al. [12] and Zhang et al. [57] propose similar networks as in [42], and incorporate stochastic motion generation and finer-grained motion control modules. Peng et al. [33] trains policy to control physically-simulated characters to hit a box. Hassan et al. [15] extend the method to generate interactions like carrying a box or sitting on a chair. Zhang et al. [55] learns physically simulated tennis players from broadcast videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.895, + 0.583 + ], + "angle": 0, + "content": "Ours versus others. Our method is most similar to GAMMA [60] and SAMP [12]. GAMMA learns generalizable motion models and policies across human bodies of diverse identities and shapes, without goal-motion paired training data. Despite producing high-fidelity motions, its results are limited to locomotion in the scene, and frequently collide with scene objects. SAMP learns conditional VAEs to produce sitting and lying actions in living environments, with object-motion paired data. Its generated motion has visible artifacts such as foot-ground skating. Our method combines their merits and eliminates their individual disadvantages. We extend the RL-based framework proposed in GAMMA by incorporating fine-grained motion controls (guided by interaction semantics) and scene interaction modules, so as to generate human-scene interactions in complex daily living environments. Compared to SAMP, our produced motion is more diverse, more physically plausible, and can be guided by fine-grained body surface markers. For systematic comparisons, please refer to Section 4." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.605, + 0.593, + 0.62 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.642, + 0.648 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "SMPL-X [31] and body representation. We use SMPL-X to represent 3D human bodies in our work. Given shape parameter \\(\\beta \\in \\mathbb{R}^{10}\\) and body pose parameters \\(\\theta \\in \\mathbb{R}^{63}\\), SMPL-X produces a posed body mesh with a fixed topology of 10475 vertices. To place the body in a scene, the root location \\(r \\in \\mathbb{R}^3\\) and the orientation \\(\\phi \\in \\mathfrak{so}(3)\\) w.r.t. the scene coordinates are additionally needed. Since facial expressions and hand gestures are not our focus, we leave their parameters as the default values. In addition, we follow [58, 60] to represent the body in motion by the SSM_67 body surface marker placement. A motion sequence is then formulated as \\(X = \\{\\pmb{x}_1, \\dots, \\pmb{x}_N\\}\\), where \\(N\\) is the length of motion and \\(\\pmb{x}_i \\in \\mathbb{R}^{67 \\times 3}\\) denotes the body marker 3D locations at frame \\(i\\). The marker locations are relative to a canonical coordinate frame centered at the body pelvis in the first frame." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.521, + 0.938 + ], + "angle": 0, + "content": "14740" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.364 + ], + "angle": 0, + "content": "GAMMA [60]. GAMMA can synthesize stochastic, perpetual, and realistic goal-reaching actions in 3D scenes. It comprises generative motion primitive models, RL-based control, and tree-based search to implement gradient-free test-time optimization. The motion primitive is formulated by a CVAE model to generate uncertain marker motions for 0.25 seconds into the future given a motion seed, followed by a MLP-based body regressor to yield SMPL-X parameters. Long-term random motion can be generated by running the motion primitive model recursively. The RL-based control is implemented by learning a policy within a simulation area. The actor-critic framework [45] and the PPO algorithm [39] are applied to update the policy network. An additional motion prior term is used to ensure the motion appears natural. During testing, the generated motion primitives are stored in a tree where only the best K primitives at each layer are preserved in order to discard low-quality sampling results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.472, + 0.595 + ], + "angle": 0, + "content": "COINS[61]. COINS generates physically plausible static human-scene interactions with instance-level semantic control. Given the point cloud of an object and action labels, COINS can generate static bodies interacting with the given object based on the specified action, e.g., sitting, lying, or touching. COINS leverages transformer-based generative models trained on a human-scene interaction dataset [13] to first generate a plausible body pelvis for interaction and then the posed body. The generated bodies are further optimized to improve the physical plausibility and to match the predicted action-dependent contact areas with objects. Such generated static bodies capture the characteristics of the human-scene interaction process and can be used as fine-grained interaction guidance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.453, + 0.621 + ], + "angle": 0, + "content": "3.2. RL-based Framework to Inhabit the Virtual" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.87 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, we propose a motion synthesis framework that enables virtual humans to navigate in complex indoor scenes and interact with various scene objects, e.g., sitting on a chair. Compared to the GAMMA framework [60], our method incorporates scene information into the states to better handle complex human-scene interactions. Also, we use body markers as goals to provide fine-grained guidance on how to drive the body for the target interactions. With modularized path-finding methods and static person-scene interaction generation methods, our framework can synthesize realistic human motions in complex 3D environments. In our work, we use COINS [61] to generate static person-scene interactions from interaction semantics given as 'action-object' pairs. The walking path can be either generated by hand, or by automatic path-finding algorithms like A* [11, 40]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We formulate our motion synthesis tasks with reinforcement learning. At each time step, a virtual human perceives" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.094, + 0.885, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.345, + 0.895, + 0.468 + ], + "angle": 0, + "content": "Figure 2: Illustration of our proposed human-scene interaction synthesis framework, which consists of learned motion primitive (actions), alongside locomotion and interaction policies generating latent actions conditioned on scenes and interaction goals. By integrating navigation mesh-based path-finding and static human-scene interaction generation methods, we can synthesize realistic motion sequences for virtual humans with fine-grained controls." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.569 + ], + "angle": 0, + "content": "its state \\( s_t \\) in the environment and samples an action \\( \\pmb{a}_t \\) from its policy model \\( \\pi(\\pmb{a}_t | s_t) \\). Based on its motion model, it advances its motion state, and obtains a new perception state \\( s_{t+1} \\). A reward \\( r_t = r(s_t, \\pmb{a}_t, s_{t+1}) \\) is calculated, tailored to different tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.893, + 0.724 + ], + "angle": 0, + "content": "The motion model and the action. We leverage the CVAE-based generative motion primitive [60] as our motion model, and use its latent variables as actions. We train the model conditioned on 1 or 2 past frames using the combination of the SAMP [12] and AMASS [29] motion capture datasets, to learn a latent motion primitive space covering motion skills for human-scene interactions. Each latent variable \\( z \\) in the motion primitive space is regarded as an action and can be decoded to a short clip of motion." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.741, + 0.76, + 0.756 + ], + "angle": 0, + "content": "The state. The state is formulated by" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.768, + 0.892, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {t} = \\left(\\boldsymbol {X} _ {s}, I, G\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.895, + 0.9 + ], + "angle": 0, + "content": "where \\( X_{s} \\in \\mathbb{R}^{M \\times 67 \\times 3} \\) is the body markers motion seed that represents a motion history of \\( M \\) frames. \\( I \\) and \\( G \\) denote the person-scene interaction feature and the goal-reaching feature, respectively. The interaction feature and goal-reaching feature vary among the locomotion and object interaction tasks. We introduce the detailed formulation in Sec. 3.3 and 3.4." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "14741" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "The rewards. The rewards evaluate how well the virtual human performs locomotion and fine-grained object interaction tasks. We formulate rewards as" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.149, + 0.469, + 0.163 + ], + "angle": 0, + "content": "\\[\nr = r _ {\\text {g o a l}} + r _ {\\text {c o n t a c t}} + r _ {\\text {p e n e}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.173, + 0.469, + 0.248 + ], + "angle": 0, + "content": "where \\( r_{\\text{goal}} \\), \\( r_{\\text{contact}} \\), and \\( r_{\\text{pene}} \\) represent the rewards for goal-reaching, foot-ground contact, and penetration avoidance, respectively. Specifically, the contact reward \\( r_{\\text{contact}} \\) encourages foot-floor contact and discourages foot skating, and is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.255, + 0.469, + 0.274 + ], + "angle": 0, + "content": "\\[\nr _ {\\text {c o n t a c t}} = e ^ {- \\left(\\left| \\min x _ {z} \\right| - 0. 0 5\\right) _ {+}} \\cdot e ^ {- \\left(\\min \\| x _ {v e l} \\| _ {2} - 0. 0 7 5\\right) _ {+}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.283, + 0.469, + 0.373 + ], + "angle": 0, + "content": "where \\(F\\) is the set of foot markers, \\(\\pmb{x}_z\\) is the height of the markers, \\(\\pmb{x}_{vel}\\) is the velocity of the markers, and \\((\\cdot)_+\\) denotes clipping negative values. There are tolerance thresholds of \\(0.05\\mathrm{m}\\) for foot-floor distance and \\(0.075\\mathrm{m/s}\\) for skating, following GAMMA [60]. The other two reward terms are action-specific and introduced in Sec. 3.3 and 3.4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.392, + 0.469, + 0.481 + ], + "angle": 0, + "content": "Policy network and training. We use the actor-critic algorithm [45] to learn the policy, where a policy network and a value network are trained jointly. The policy network generates a diagonal Gaussian distribution representing the stochastic action distribution given a state, while the value network outputs the value estimation for each state." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.482, + 0.469, + 0.514 + ], + "angle": 0, + "content": "Like in GAMMA [60], these two networks are jointly trained by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.521, + 0.382, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {P P O} + \\mathbb {E} \\left[ \\left(r _ {t} - V (s _ {t})\\right) ^ {2} \\right] \\\\ + \\alpha \\operatorname {K L} - \\operatorname {d i v} (\\pi (z | s) | | \\mathcal {N} (0, I)), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.566, + 0.469, + 0.627 + ], + "angle": 0, + "content": "where the first term is the PPO [39] loss, the second updates the value estimation of the critic networks, and the third Kullback-Leibler divergence term regularizes motion in the latent space [60]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.645, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Tree sampling for test-time optimization. Given the stochastic nature of our Gaussian policies, sampling motions from the generated action distributions can yield motion primitive results of various qualities. Therefore, we follow [60] to use tree-based sampling during inference to discard motion primitives with inferior goal-reaching and scene interaction scores. Specifically, we sample multiple latent actions at each time step and selectively keep the best K samples, utilizing the same rewards used to train the policies as the selection criteria. This tree-sampling technique yields improved synthesis results of higher quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In the following, we elaborate on the design of states and rewards tailored to different actions, namely locomotion and fine-grained object interaction. By combining the learned policies, long-term and coherent motions can be composed by rolling out the initial body and switching between the locomotion and object interaction stages." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.887, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.254, + 0.892, + 0.374 + ], + "angle": 0, + "content": "Figure 3: Illustration of the scene-aware locomotion policy network. The locomotion policy state consists of the body markers, the goal-reaching feature of normalized direction vectors from markers to the goal pelvis, and the interaction feature of a 2D binary map indicating the walkability (red: non-walkable area, blue: walkable area) of the local \\(1.6m \\times 1.6m\\) square area. The locomotion policy network employs the actor-critic architecture and shares the state encoder." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.397, + 0.808, + 0.413 + ], + "angle": 0, + "content": "3.3. Scene-Aware Locomotion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.42, + 0.892, + 0.555 + ], + "angle": 0, + "content": "Navigating in cluttered scenes here means the human body moving to a target while avoiding collisions with scene objects. Our key idea is to incorporate the walkability information of the surrounding environment into the states and use collision rewards to train the locomotion policy in order to avoid scene collisions. Specifically, we represent the walkability of the environment surrounding the human agent using a 2D binary map \\(\\mathcal{M} \\in \\{0,1\\}^{16 \\times 16}\\) as illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.556, + 0.892, + 0.661 + ], + "angle": 0, + "content": "The walkability map is defined in the human's local coordinates and covers a \\(1.6m \\times 1.6m\\) area centered at the body pelvis and aligned with the body facing orientation. It consists of a \\(16 \\times 16\\) cell grid where each cell stores a binary value indicating whether this cell is walkable or not. This local walkability map enables the policy to sense surrounding obstacles." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.662, + 0.891, + 0.692 + ], + "angle": 0, + "content": "Referring to Eq. 1, the person-scene interaction feature is specified by" + }, + { + "type": "equation", + "bbox": [ + 0.65, + 0.692, + 0.891, + 0.707 + ], + "angle": 0, + "content": "\\[\nI = \\operatorname {v e c} (\\mathcal {M}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.712, + 0.891, + 0.742 + ], + "angle": 0, + "content": "in which \\(\\mathrm{vec}(\\cdot)\\) denotes vectorization. The goal-reaching feature is specified by" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.75, + 0.891, + 0.765 + ], + "angle": 0, + "content": "\\[\nG = \\left(\\tilde {\\boldsymbol {g}} _ {p} - \\boldsymbol {X} _ {s}\\right) _ {n}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.772, + 0.892, + 0.862 + ], + "angle": 0, + "content": "where \\( \\mathbf{X}_s \\) and \\( \\tilde{\\pmb{g}}_p \\in \\mathbb{R}^{M \\times 67 \\times 3} \\) are the body marker seed representing \\( M \\) frames history of motion, and the broadcasted target pelvis location relative to the body-centered canonical coordinate, respectively. \\( (\\tilde{\\pmb{g}}_p - \\mathbf{X}_s)_n \\) denotes the normalized vectors pointing from each marker to the goal pelvis." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.864, + 0.838, + 0.878 + ], + "angle": 0, + "content": "The rewards contributing to Eq. 2 are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.884, + 0.891, + 0.902 + ], + "angle": 0, + "content": "\\[\nr _ {p e n e} = e ^ {- | \\mathcal {M} _ {0} \\cap \\mathcal {B} _ {x y} (X) |}, \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.936 + ], + "angle": 0, + "content": "14742" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}_0\\) denotes the non-walkable cells in the walkability map, \\(\\mathcal{B}_{xy}(\\cdot)\\) denotes the 2D bounding box of the body markers \\(X\\), \\(\\cap\\) denotes their intersection, and \\(|\\cdot|\\) denotes the number of non-walkable cells overlapping with the human bounding box." + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.185, + 0.469, + 0.201 + ], + "angle": 0, + "content": "\\[\nr _ {\\text {g o a l}} = r _ {\\text {d i s t}} + r _ {\\text {o r i}}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.224, + 0.469, + 0.243 + ], + "angle": 0, + "content": "\\[\nr _ {d i s t} = 1 - \\left(\\left\\| \\boldsymbol {p} - \\boldsymbol {g} _ {p} \\right\\| _ {2} - 0. 0 5\\right) _ {+}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.261, + 0.469, + 0.293 + ], + "angle": 0, + "content": "\\[\nr _ {o r i} = \\frac {\\left\\langle \\boldsymbol {o} , \\boldsymbol {g} _ {p} - \\boldsymbol {p} \\right\\rangle}{2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.298, + 0.469, + 0.359 + ], + "angle": 0, + "content": "where \\( r_{dist} \\) encourages the body pelvis \\( p \\) to be close to the pelvis goal \\( g_p \\) and \\( r_{ori} \\) encourages the body facing direction \\( o \\) to be aligned with the direction from the current body pelvis \\( p \\) to the pelvis goal \\( g_p \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.369, + 0.436, + 0.385 + ], + "angle": 0, + "content": "3.4. Fine-grained Object Interaction Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.469, + 0.514 + ], + "angle": 0, + "content": "To synthesize fine-grained human-object interactions, e.g. sitting on a chair or lying on a sofa, we use body marker goals as guidance, and model the proximity between the body surface and the scene object in a compact way. The goal marker sets can be generated by static person-scene interaction methods such as [14, 56, 61]. We use COINS [61] to generate the static goal interaction body for its performance and controllability of interaction semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.469, + 0.635 + ], + "angle": 0, + "content": "In addition to the marker-based goal guidance, we incorporate the proximity relations between humans and objects into the states. Specifically, we use the signed distance from each marker to the surface of the scene object, as well as the gradient direction of the signed distance to represent the proximity relationship, as illustrated in Fig. 4. Both the signed distance and its gradient direction are calculated using the object's signed distance field (SDF) \\(\\Psi_{O}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.636, + 0.469, + 0.664 + ], + "angle": 0, + "content": "Referring to Eq. 1, the person-scene interaction feature is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.679, + 0.469, + 0.696 + ], + "angle": 0, + "content": "\\[\nI = \\left[ \\Psi_ {O} \\left(\\boldsymbol {X} _ {s}\\right), \\nabla \\Psi_ {O} \\left(\\boldsymbol {X} _ {s}\\right) \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.706, + 0.469, + 0.768 + ], + "angle": 0, + "content": "in which \\(\\Psi_O\\in \\mathbb{R}^{M\\times 67}\\) and \\(\\nabla \\Psi_{O}\\in \\mathbb{R}^{M\\times 201}\\) denote the SDF values and the gradient at each marker location in the \\(M\\) frames, respectively, and \\([\\cdot ,\\cdot ]\\) denotes feature concatenation. The goal-reaching feature is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.781, + 0.469, + 0.799 + ], + "angle": 0, + "content": "\\[\nG = \\left[ \\left(\\tilde {\\boldsymbol {g}} _ {m} - \\boldsymbol {X} _ {s}\\right) _ {n}, \\| \\tilde {\\boldsymbol {g}} _ {m} - \\boldsymbol {X} _ {s} \\| _ {2} \\right], \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.809, + 0.469, + 0.901 + ], + "angle": 0, + "content": "in which \\( \\mathbf{X}_s \\in \\mathbb{R}^{M \\times 67 \\times 3} \\) denotes the body markers seed, \\( \\tilde{\\pmb{g}}_m \\in \\mathbb{R}^{M \\times 67 \\times 3} \\) denotes the broadcasted goal body markers, \\( (\\tilde{\\pmb{g}}_m - \\pmb{X}_s)_n \\) denotes the normalized vector representing the direction from each marker to the corresponding target marker, \\( \\| \\tilde{\\pmb{g}}_m - X_s \\|_2 \\) denotes the distance from each marker to target marker." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.088, + 0.887, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.244, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Figure 4: Illustration of the object interaction policy network. The interaction policy state consists of the body markers, the goal-reaching features of both distance and direction from current markers to the goal markers, and the interaction features of the signed distances from each marker to the object surfaces and the signed distance gradient at each marker location. Such interaction features encode the human-object proximity relationship." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.392, + 0.892, + 0.437 + ], + "angle": 0, + "content": "The interaction policy is trained using the reward defined in Eq. 2 with the following interaction-specific goal reward and penetration reward:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.452, + 0.891, + 0.469 + ], + "angle": 0, + "content": "\\[\nr _ {\\text {g o a l}} = 1 - \\left(\\| \\boldsymbol {x} - \\boldsymbol {g} _ {m} \\| _ {2} - 0. 0 5\\right) _ {+}, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.486, + 0.891, + 0.508 + ], + "angle": 0, + "content": "\\[\nr _ {p e n e} = e ^ {- \\frac {1}{| V |} \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {| V |} \\left| \\left(\\Psi_ {O} (\\boldsymbol {v} _ {t i})\\right) _ {-} \\right|}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.892, + 0.65 + ], + "angle": 0, + "content": "with \\( |V| \\) being the SMPL-X mesh vertices, \\( T \\) denotes the number of frames in each motion primitive (equals to 10 in our study). The distance reward encourages the final frame body markers \\( x \\) to be close to the goal body markers \\( g_{m} \\). The penetration reward penalizes all the body vertices within a motion primitive that have negative SDF values. We use body vertices instead of joints because human-object contact happens on the body surface and can be better detected using vertices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.741 + ], + "angle": 0, + "content": "Moreover, we train the interaction policy with a mixture of 'sit/lie down' and 'stand up' tasks. This training scheme enables the human agent to also learn how to stand up and transit from object interaction back to locomotion, which enables the synthesis of a sequence of interaction activities as in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.625, + 0.772 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Motion Datasets. We combine the large-scale motion capture dataset AMASS [29] with SAMP [12] motion data to train the motion primitive models. Each sequence is first subsampled to 40 FPS and then split into 10 frames and 100 frames clips. Each motion clip is canonicalized using the first frame body. Specifically, we select AMASS sequences annotated with 'sit' or 'lie' in BABEL [37] and all motion data of SAMP to train the motion primitive model. We train" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "14743" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.135 + ], + "angle": 0, + "content": "Table 1: Evaluation of the locomotion task. The up/down arrows denote the score is the higher/lower the better and the best results are in boldface." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.146, + 0.46, + 0.2 + ], + "angle": 0, + "content": "
time ↓avg. dist ↓contact ↑loco pene ↑
SAMP [12]5.970.140.840.94
GAMMA [60]3.870.030.940.94
Ours6.430.040.990.95
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.47, + 0.289 + ], + "angle": 0, + "content": "separate motion primitive models for locomotion and interactions using task-related data. We observe that extending SAMP motion data with AMASS dataset is the key to making interaction policies work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.473, + 0.53 + ], + "angle": 0, + "content": "Policy Training Environments. To train the scene-aware locomotion policy, we randomly generate synthetic cluttered scenes consisting of random objects from ShapeNet [6]. Random initial and target location pairs are sampled in the walkable areas using navigation meshes to train the locomotion policy. To train the interaction policy, we use the static person-object interaction data from PROX [13] and retargeted to ShapeNet objects. We first randomly sample a furniture and interaction body goal from the retargeted PROX data. Then we sample the initial body with random poses and locations in front of the object to train the interaction policy. We also randomly swap the initial and goal body to learn both 'sit/lie down' and 'stand up' motions. Please refer to Supp. Mat. for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.308, + 0.56 + ], + "angle": 0, + "content": "4.1. Locomotion in 3D Scenes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.632 + ], + "angle": 0, + "content": "We randomly generate test scenes for locomotion in the same way as the training scenes. The virtual human is instructed to move from the random start point to the random target point while avoiding penetration with scene objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Baselines and metrics. We compare our method with SAMP [12] and GAMMA [60] for locomotion. The SAMP results are recorded by running the released Unity demo. The start and termination are manually determined so the reported completion time may be slightly higher than the actual time due to human response time. The evaluation metrics for locomotion include: 1) time from start point to target point or reaching the time limit, measured in seconds. 2) the average distance from the final body to the targets, measured in meters. 3) foot contact score encouraging the lowest feet joints on the floor and discouraging foot skating as defined in Eq. 3. We use body joints instead of markers to calculate the contact score because the marker set annotation for the SAMP body is missing. 4) locomotion penetration score indicating the percentage of body vertices that are inside the walkable areas according to the navigation mesh." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.696, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.203, + 0.696, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.089, + 0.874, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.203, + 0.874, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.322, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 5: Demonstration of locomotion tasks where GAMMA [60] (left) collides with the obstacles (red bodies) while our scene-aware locomotion policy (right) avoids collision. The yellow circles denote the specified waypoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.895, + 0.637 + ], + "angle": 0, + "content": "Results. Table 1 shows the empirical evaluation results. Our method achieves a higher contact score (0.99) than both GAMMA (0.94) and SAMP (0.84) which indicates better foot-floor contact and less foot skating. Moreover, our method achieves the highest penetration score which indicates our scene-aware policy can better avoid scene collisions. Fig. 5 shows examples of locomotion tasks where GAMMA collides into the scenes while our scene-aware policy successfully avoids penetration. Compared to GAMMA trained in the same environments, we observe our scene-aware policy learns more conservative behavior with lower moving speed and spent more time (6.43s) walking to the target locations, just like a human afraid of stepping on surrounding traps. All the methods can reach the target location within a reasonable distance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.852, + 0.663 + ], + "angle": 0, + "content": "4.2. Fine-Grained Human-Object Interaction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.895, + 0.792 + ], + "angle": 0, + "content": "We evaluate the object interaction task on 10 unseen objects (3 armchairs, 3 straight chairs, 3 sofas, 1 L-sofa) from ShapeNet [6]. We use the object size annotation of ShapeNet and manually filter unrealistic-sized objects. The virtual human is randomly placed in front of the target object and then instructed to perform the interaction, stay for around 2 seconds, and then stand up. We evaluate two interactions of sitting and lying separately." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Baselines and metrics. We compare our method with SAMP [12] for the object interaction task. The evaluation metrics for the interaction tasks are: 1) time of completing the object interaction task. 2) foot contact score as defined in Eq. 3. Note that the foot contact score does not always reflect the motion quality for lying tasks because the foot can" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.521, + 0.938 + ], + "angle": 0, + "content": "14744" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.471, + 0.134 + ], + "angle": 0, + "content": "Table 2: Evaluation of the interaction tasks. The up/down arrows denote the score is the higher/ lower the better and the best results are in boldface." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.146, + 0.468, + 0.21 + ], + "angle": 0, + "content": "
time ↓contact ↑pene. mean ↓pene. max ↓
SAMP [12] sit8.630.8911.9145.22
Ours sit4.090.971.9110.61
SAMP [12] lie12.550.7344.77238.81
Ours lie4.200.789.9044.61
" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.226, + 0.46, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.515 + ], + "angle": 0, + "content": "Figure 6: Demonstration of various generated object interactions. Each row shows generated interactions with the same object starting from random initial body locations and orientations. Colors from yellow to red denote time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.47, + 0.573 + ], + "angle": 0, + "content": "often be off the floor during lying. 3) interaction penetration score for each frame is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.585, + 0.47, + 0.618 + ], + "angle": 0, + "content": "\\[\ns _ {i n t e r - p e n e} = \\sum_ {v _ {i} \\in V} | (\\Psi_ {O} (v _ {i})) _ {-} |, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.687 + ], + "angle": 0, + "content": "where \\(\\Psi_{O}\\) is the object signed distance field, \\((\\cdot)_{-}\\) clips all positive distance values to zero, and \\(V\\) is the body vertivces. We show both the average penetration over time and the maximum penetration in one sequence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Results. Tab. 2 shows the evaluation results. Our method achieves a significantly higher foot contact score, indicating more natural motion. Our method also achieves lower mean and maximum penetration, which means our method generates more physically plausible results. Moreover, our method can complete the interaction tasks much faster than SAMP. This is because SAMP does not generalize well to unseen random body initialization and needs a longer time to start performing interactions. Qualitative results demonstrating the various object interactions generated by our method are shown in Fig. 6. Our object interaction policy generalizes to random initial body locations and orientations, as well as novel objects of various shapes." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.698, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.089, + 0.888, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.305 + ], + "angle": 0, + "content": "Figure 7: Our method can synthesize natural human movements in scene reconstructions that are significantly different from the training scenes. We show the synthesis results in one Paris street [7] (left figure) and one indoor room [13] (right figure). Transparent to solid colors denote time." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.43 + ], + "angle": 0, + "content": "Table 3: Evaluation of interaction sequences synthesis. Here 'contact' denotes the foot contact score, 'loco. pene.' is the percentage of body vertices inside the walkable areas, 'inter. pene. mean/max' denotes the mean and maximum penetration with interaction objects, and 'perceptual' denotes the ratio of being chosen as perceptually more natural. The best results are in boldface." + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.441, + 0.832, + 0.542 + ], + "angle": 0, + "content": "
SAMP [12]Ours
contact ↑0.870.96
loco. pene. ↑0.620.72
inter. pene. mean↓15.613.40
inter. pene. max ↓101.2539.68
perceptual. ↑0.150.85
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.571, + 0.807, + 0.587 + ], + "angle": 0, + "content": "4.3. Interaction Sequences in 3D Scenes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Humans often continuously perform sequences of interactions with various objects in complex real-world scenes as shown in Fig. 1. Such interaction sequences are combinations of alternating locomotions and fine-grained interactions which we evaluate in Sec. 4.1 and Sec. 4.2 respectively. We conduct the empirical evaluation in real scene scans from Replica [44] and compare our method with SAMP. We select a list of interactable objects in the scene and instruct the virtual human to walk to the objects to perform interactions one by one. The evaluation metrics include the foot contact score, locomotion penetration score, and the mean and maximum interaction penetration score. We also conducted a perceptual study where participants are shown a side-by-side comparison of results from two methods and asked to choose the one perceptually more natural. We report the rate of being chosen as the better." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The evaluation results are shown in Tab. 3. Our method generates high-quality results of human-scene interactions in cluttered environments. Our results achieve higher contact scores and less penetration with the scenes compared" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "14745" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.136 + ], + "angle": 0, + "content": "Table 4: Rewards ablation studies results, where '-Contact' and '-Penetration' denote policies trained without the floor contact and penetration avoidance reward, respectively." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.146, + 0.466, + 0.2 + ], + "angle": 0, + "content": "
time ↓contact ↑pene. mean ↓pene. max ↓
Ours4.090.971.9110.61
- Contact3.750.8819.4960.10
- Penetration4.030.9516.8045.50
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.224, + 0.468, + 0.256 + ], + "angle": 0, + "content": "to SAMP. In addition, our method generates perceptually more natural results according to perceptual study." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.265, + 0.242, + 0.28 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.47, + 0.515 + ], + "angle": 0, + "content": "We perform ablation studies on the foot-ground contact reward and penetration avoidance reward of the interaction policies to substantiate the significance of these rewards. We train two ablation interaction policies using identical networks and environments, differing only in the exclusion of one of these rewards in each case. The quantitative metrics are reported in Tab. 4. The removal of the penetration avoidance reward yields a marked escalation in detected human-scene penetration. The removal of the footfloor contact reward yields notably both inferior contact and penetration scores, accompanied by observed erratic synthesized motions that either remain suspended in the air or penetrate the floor. These empirical findings underscore the crucial role played by the foot-floor contact and penetration avoidance rewards in the learning of interaction policies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.325, + 0.543 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Limitations and Future Work. Our current method has various limitations that could be improved in future works. First, our method does not fully resolve penetrations with scene objects and floors since our method only uses rewards to encourage avoiding penetration with scenes, which does not impose hard constraints for penetrations. The combination of our method with physics simulation holds the potential to effectively address and resolve penetration issues. Moreover, the lying motions generated by our method are not as natural as the sitting motions because the motion primitive model fails to learn a comprehensive action space for lying given limited training data. Specifically, the available AMASS motion capture data for lying (167 seconds) is significantly less than the data for sitting (5K seconds). To overcome this issue, we aim to explore more data-efficient learning methods and scalable methods to collect human-scene interaction data. Furthermore, our locomotion policy is now limited to flat-floor scenes due to its reliance on the 2D occupancy-based walkability map. However, to broaden the applicability of our approach to more complex scenes such as uneven outdoor terrains and multi-floor buildings requiring walking upstairs, it will be necessary to replace the walkability map with a more suitable environment sens" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.687, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.21, + 0.653, + 0.224 + ], + "angle": 0, + "content": "Inter-penetrations" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.091, + 0.89, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.21, + 0.849, + 0.224 + ], + "angle": 0, + "content": "Unnatural poses" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.238, + 0.892, + 0.329 + ], + "angle": 0, + "content": "Figure 8: Limitations. Penetrations remain observable in our results (left) since our method only encourages avoiding collision with rewards. Furthermore, the absence of sufficient training data for lying motions impedes the motion primitive model's ability to acquire a comprehensive lying motion space, leading to degraded motion (right)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.476 + ], + "angle": 0, + "content": "ing mechanism. In addition, our method is restricted to interactions with static scenes. However, in the real world, humans are exposed to dynamic interactions with movable objects and scenes involving autonomous agents including other humans, animals, and vehicles. Extension to dynamic interactions will enable tackling a broader range of intricate and dynamic applications that mirror the complexities inherent in the real world." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.893, + 0.691 + ], + "angle": 0, + "content": "Conclusion. In this paper, we leverage reinforcement learning to establish a framework to synthesize diverse human motions in indoor scenes, which is stochastic, realistic, and perpetual. The proposed method has large potential to improve many applications such as daily-living activity simulation, synthetic data creation, architecture design, and so on. Compared to existing methods, our method realizes fine-grained control by using body surface keypoints as targets, and achieves autonomous body-scene collision avoidance by incorporating scene information into the states and the rewards. Experiments show that our method effectively enables virtual humans to inhabit the virtual, and outperforms baselines consistently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.892, + 0.786 + ], + "angle": 0, + "content": "Acknowledgements. We sincerely acknowledge the anonymous reviewers for their insightful suggestions. This work was supported by the SNSF project grant 200021 204840 and SDSC PhD fellowship. Lingchen Yang provided valuable suggestions on visualization." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "14746" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. ISBN: 0730-0301 Publisher: ACM New York, NY, USA. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.184, + 0.471, + 0.224 + ], + "angle": 0, + "content": "[2] Tenglong Ao, Zeyi Zhang, and Libin Liu. GestureDiffuCLIP: Gesture diffusion model with CLIP latents. arXiv preprint arXiv:2303.14613, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.224, + 0.471, + 0.292 + ], + "angle": 0, + "content": "[3] Kevin Bergamin, Simon Clavet, Daniel Holden, and James Richard Forbes. DReCon: data-driven responsive control of physics-based characters. ACM Transactions On Graphics (TOG), 38(6):1-11, 2019. Publisher: ACM New York, NY, USA. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.292, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[4] Michael Buttner. Machine learning for motion synthesis and character control in games. Proc. of I3D 2019, 2, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.319, + 0.471, + 0.358 + ], + "angle": 0, + "content": "[5] Michael Böttner and Simon Clavet. Motion matching-the road to next gen animation. Proc. of Nucl. ai, 2015(1):2, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, and Hao Su. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.471, + 0.495 + ], + "angle": 0, + "content": "[7] Jean-Emmanuel Deschaud, David Duque, Jean Pierre Richa, Santiago Velasco-Forero, Beatrix Marcotegui, and François Goulette. Paris-carla-3d: A real and synthetic outdoor point cloud dataset for challenging tasks in 3d mapping. Remote Sensing, 13(22), 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[8] Helmut Grabner, Juergen Gall, and Luc Van Gool. What makes a chair a chair? pages 1529-1536. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[9] Abhinav Gupta, Scott Satkin, Alexei A Efros, and Martial Hebert. From 3d scene geometry to human workspace. pages 1961-1968. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.471, + 0.602 + ], + "angle": 0, + "content": "[10] Ikhsanul Habibie, Daniel Holden, Jonathan Schwarz, Joe Yearsley, and Taku Komura. A recurrent variational autoencoder for human motion synthesis. 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.471, + 0.655 + ], + "angle": 0, + "content": "[11] Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100-107, 1968. Publisher: IEEE. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.471, + 0.71 + ], + "angle": 0, + "content": "[12] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. pages 11374-11384, 2021. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.71, + 0.471, + 0.763 + ], + "angle": 0, + "content": "[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3D human pose ambiguities with 3D scene constraints. pages 2282-2292, 2019. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[14] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3D scenes by learning human-scene interaction. pages 14708-14718, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[15] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing Physical Character-Scene Interactions, Feb. 2023. arXiv:2302.00883 [cs]. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[16] Daniel Holden, Ouussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Graphics (TOG), 39(4):53-1, 2020. Publisher: ACM New York, NY, USA. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[17] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. Publisher: ACM New York, NY, USA. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.174, + 0.892, + 0.228 + ], + "angle": 0, + "content": "[18] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. Publisher: ACM New York, NY, USA. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.228, + 0.892, + 0.282 + ], + "angle": 0, + "content": "[19] Ruizhen Hu, Zihao Yan, Jingwen Zhang, Oliver Van Kaick, Ariel Shamir, Hao Zhang, and Hui Huang. Predictive and generative neural networks for object functionality. arXiv preprint arXiv:2006.15520, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.282, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[20] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. PADI: Language-Directed Physics-Based Character Control. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.892, + 0.378 + ], + "angle": 0, + "content": "[21] Kacper Kania, Marek Kowalski, and Tomasz Trzciński. TrajeVAE: Controllable Human Motion Generation from Trajectories. arXiv preprint arXiv:2104.00351, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[22] Lucas Kovar, Michael Gleicher, and Frédéric Pighin. Motion graphs. In ACM SIGGRAPH 2008 classes, pages 1-10. 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.416, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[23] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. Ganimator: Neural motion synthesis from a single sequence. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. Publisher: ACM New York, NY, USA. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.485, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[24] Weiyu Li, Xuelin Chen, Peizhuo Li, Olga Sorkine-Hornung, and Baoquan Chen. Example-based Motion Synthesis via Generative Motion Matching. arXiv preprint arXiv:2306.00378, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[25] Xueting Li, Sifei Liu, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Putting humans in a scene: Learning affordance in 3d indoor environments. pages 12368-12376, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.593, + 0.892, + 0.647 + ], + "angle": 0, + "content": "[26] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. Publisher: ACM New York, NY, USA. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[27] Libin Liu, KangKang Yin, and Baining Guo. Improving sampling-based motion control. volume 34, pages 415-423. Wiley Online Library, 2015. Issue: 2. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[28] Libin Liu, KangKang Yin, Michiel Van de Panne, Tianjia Shao, and Weiwei Xu. Sampling-based contact-rich motion control. In ACM SIGGRAPH 2010 papers, pages 1-10. 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.741, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. pages 5442-5451, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.796, + 0.892, + 0.85 + ], + "angle": 0, + "content": "[30] Josh Merel, Leonard Hasenclever, Alexandre Galashov, Arun Ahuja, Vu Pham, Greg Wayne, Yee Whye Teh, and Nicolas Heess. Neural probabilistic motor primitives for humanoid control. arXiv preprint arXiv:1811.11711, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.85, + 0.892, + 0.891 + ], + "angle": 0, + "content": "[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "14747" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.118 + ], + "angle": 0, + "content": "and body from a single image. pages 10975-10985, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.12, + 0.469, + 0.187 + ], + "angle": 0, + "content": "[32] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. Publisher: ACM New York, NY, USA. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.187, + 0.469, + 0.255 + ], + "angle": 0, + "content": "[33] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. Publisher: ACM New York, NY, USA. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.256, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[34] Xue Bin Peng, Angjoo Kanazawa, Jitendra Malik, Pieter Abbeel, and Sergey Levine. Sfv: Reinforcement learning of physical skills from videos. ACM Transactions On Graphics (TOG), 37(6):1-14, 2018. Publisher: ACM New York, NY, USA. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.323, + 0.469, + 0.363 + ], + "angle": 0, + "content": "[35] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. pages 10985-10995, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.468, + 0.404 + ], + "angle": 0, + "content": "[36] Mathis Petrovich, Michael J Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. pages 480-497. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[37] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. BABEL: bodies, action and behavior with English labels. pages 722-731, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[38] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. SceneGrok: Inferring action maps in 3D environments. ACM transactions on graphics (TOG), 33(6):1-10, 2014. Publisher: ACM New York, NY, USA. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[39] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[40] Greg Snook. Simplified 3D movement and pathfinding using navigation meshes. Game programming gems, 1(1):288-304, 2000. Publisher: Charles River Media Newton Centre, MA, USA. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[41] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. Publisher: ACM New York, NY, USA. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[42] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.469, + 0.768 + ], + "angle": 0, + "content": "[43] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. Publisher: ACM New York, NY, USA. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.769, + 0.469, + 0.835 + ], + "angle": 0, + "content": "[44] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, and Shobhit Verma. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.837, + 0.469, + 0.876 + ], + "angle": 0, + "content": "[45] Richard S Sutton and Andrew G Barto. Introduction to reinforcement learning, volume 135. MIT press Cambridge, 1998. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.877, + 0.469, + 0.89 + ], + "angle": 0, + "content": "[46] Xiangjun Tang, He Wang, Bo Hu, Xu Gong, Ruifan Yi, Qi" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "long Kou, and Xiaogang Jin. Real-time controllable motion transition for characters. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. Publisher: ACM New York, NY, USA. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.147, + 0.892, + 0.213 + ], + "angle": 0, + "content": "[47] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. CALM: Conditional Adversarial Latent Models for Directable Virtual Characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.215, + 0.892, + 0.256 + ], + "angle": 0, + "content": "[48] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. pages 358–374. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.256, + 0.892, + 0.296 + ], + "angle": 0, + "content": "[49] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[50] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. EDGE: Editable Dance Generation From Music. arXiv preprint arXiv:2211.10658, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[51] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. pages 20460-20469, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[52] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. pages 9401-9411, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.416, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[53] Zhiyong Wang, Jinxiang Chai, and Shihong Xia. Combining recurrent neural networks and adversarial training for human motion synthesis and control. IEEE transactions on visualization and computer graphics, 27(1):14-28, 2019. Publisher: IEEE. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[54] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. PhysDiff: Physics-Guided Human Motion Diffusion Model. arXiv preprint arXiv:2212.02500, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.605 + ], + "angle": 0, + "content": "[55] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning Physically Simulated Tennis Skills from Broadcast Videos. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. ISBN: 0730-0301 Publisher: ACM New York, NY, USA. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.606, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[56] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. PLACE: Proximity learning of articulation and contact in 3D environments. pages 642-651. IEEE, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[57] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: towards controllable human-chair interactions. pages 518-535. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[58] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. pages 3372-3382, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.755, + 0.892, + 0.795 + ], + "angle": 0, + "content": "[59] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes without people. pages 6194-6204, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.795, + 0.892, + 0.822 + ], + "angle": 0, + "content": "[60] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3D scenes. pages 20481-20491, 2022. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.822, + 0.892, + 0.874 + ], + "angle": 0, + "content": "[61] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. pages 311-327. Springer, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.89 + ], + "angle": 0, + "content": "[62] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "14748" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. pages 676-694. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.173 + ], + "angle": 0, + "content": "[63] Fabio Zinno. MI tutorial day: From motion matching to motion synthesis, and all the hurdles in between. Proc. of GDC 2019, 2, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "14749" + } + ] +] \ No newline at end of file diff --git a/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/39947c2e-9c93-4c2e-8f0c-e7f92b1d1960_origin.pdf b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/39947c2e-9c93-4c2e-8f0c-e7f92b1d1960_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bc3a5807b57dadeac8ea90e86b9a98173c601161 --- /dev/null +++ b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/39947c2e-9c93-4c2e-8f0c-e7f92b1d1960_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:368f4c73290b2e730724f7af69f9f11c95e1f90be385b5b9f1fa8f162b3bfa29 +size 15280587 diff --git a/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/full.md b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5dd2386a6a480ac8b760b817e019aea344f7b2fc --- /dev/null +++ b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/full.md @@ -0,0 +1,355 @@ +# Synthesizing Diverse Human Motions in 3D Indoor Scenes + +Kaifeng Zhao1 Yan Zhang1 Shaofei Wang1 Thabo Beeler2 Siyu Tang1 ETH Zürich 2Google + +{kaifeng.zhao, yan.zhang, shaofei.wang, siyu.tang}@inf.ethz.ch thabo.beeler@gmail.com + +![](images/92394e5bf069ee29114dba10b1473a40d93f8ac25a090519db68cc9e4ffe965b.jpg) +Figure 1: In this work, we propose a method to generate a sequence of natural human-scene interaction events in real-world complex scenes as illustrated here. The human first walks to sit on a stool (yellow to red), then walk to another chair to sit down (red to magenta), and finally walk to and lie on the sofa (magenta to blue). + +# Abstract + +We present a novel method for populating 3D indoor scenes with virtual humans that can navigate in the environment and interact with objects in a realistic manner. Existing approaches rely on high-quality training sequences that contain captured human motions and the 3D scenes they interact with. However, such interaction data are costly, difficult to capture, and can hardly cover the full range of plausible human-scene interactions in complex indoor environments. To address these challenges, we propose a reinforcement learning-based approach that enables virtual humans to navigate in 3D scenes and interact with objects realistically and autonomously, driven by learned motion control policies. The motion control policies employ latent motion action spaces, which correspond to realistic motion primitives and are learned from large-scale motion capture data using a powerful generative motion model. For navigation in a 3D environment, we propose a scene-aware policy + +with novel state and reward designs for collision avoidance. Combined with navigation mesh-based path-finding algorithms to generate intermediate waypoints, our approach enables the synthesis of diverse human motions navigating in 3D indoor scenes and avoiding obstacles. To generate fine-grained human-object interactions, we carefully curate interaction goal guidance using a marker-based body representation and leverage features based on the signed distance field (SDF) to encode human-scene proximity relations. Our method can synthesize realistic and diverse human-object interactions (e.g., sitting on a chair and then getting up) even for out-of-distribution test scenarios with different object shapes, orientations, starting body positions, and poses. Experimental results demonstrate that our approach outperforms state-of-the-art human-scene interaction synthesis methods in terms of both motion naturalness and diversity. Code, models, and demonstrative video results are available at: https://zkf1997.github.io/DIMOS. + +# 1. Introduction + +Simulating how humans interact with environments plays an essential role in many applications, such as generating training data for machine learning algorithms, and simulating autonomous agents in AR/VR and computer games. Although this task is highly related to character animation in computer graphics, most existing character animation methods (e.g. [4, 5, 16]) focus on improving the realism and controllability of character movements. With the traditional character animation workflows, one can produce high-quality animations but can hardly generate autonomous and spontaneous natural human motions interacting with the surroundings in diverse plausible ways as real humans. Previous learning-based interaction synthesis methods [12, 42, 57] require simultaneously capturing human motion and scenes for supervision. However, capturing such training data is costly and challenging, resulting in a notably limited spectrum of human-scene interaction motions and difficulties in handling unseen interaction scenarios. This restriction also results in inferior motion quality of synthesized virtual humans. + +To address this problem, we leverage reinforcement learning (RL) [45] to solve our task. By formulating goals as rewards, perception as states, and latent variables of deep generative models as actions, we can synthesize continuous, stochastic, plausible, and spontaneous motions of virtual humans to inhabit the digital world. Although existing RL-based motion synthesis approaches (e.g. [26, 33, 60]) can effectively generate natural motions to achieve goals, their generated virtual humans can only interact with simple scenes, rather than complex environments with functional furniture and diverse objects. For example, GAMMA [60] employs generative motion primitives and a policy network that are generalizable across diverse human body shapes, but it can only synthesize waypoint-reaching locomotions. The trained digital humans are not aware of how to perform actions like sitting on a chair or lying on a sofa, and frequently inter-penetrate with the scene geometry. + +To overcome these limitations, we propose a novel framework to learn both scene and interaction-aware motion control policies for synthesizing realistic and diverse human-scene interaction motions. First, in order to improve the physical plausibility of the synthesized human motions, we design a new scene-aware policy to help virtual humans avoid collisions with scene objects. Specifically, we use a 2D occupancy-based local walkability map to incorporate scene information into the locomotion policy. In addition, we add features derived from the signed distance from body markers to the object surface and the gradient direction of the signed distance to encode the proximity between humans and objects for object interaction policies. Second, in order to achieve controllable object interactions, we provide fine-grained guidances based on surface mark + +ers [58] of a human body performing the target interactions. Specifically, we use COINS [61] to generate human bodies interacting with scene objects given the interaction semantics, and then use the body markers as the interaction guidance for motion synthesis. Combined with navigation mesh-based path-finding algorithms to generate intermediate waypoints in 3D scenes, virtual humans can autonomously reach target locations in complex environments and mimic target poses in a variety of plausible ways. + +We train the policy networks in synthetic scenes consisting of randomized objects to learn generalizable scene-aware locomotion and fine-grained object interactions. With this framework, we investigate how to synthesize diverse in-scene motions consisting of locomotion, sitting, and lying. We empirically evaluate the motion realism and expressiveness of our proposed method, and compare it with state-of-the-art methods. The results show that our approach consistently outperforms the baselines in terms of diversity, physical plausibility, and perceptual scores. + +In summary, we aim to let virtual humans inhabit virtual environments, and present these contributions: + +1. We propose a reinforcement learning-based framework to generate realistic and diverse motions of virtual humans in complex indoor scenes. +2. We propose to use body surface markers as detailed interaction goals for fine-grained human-object interaction synthesis and leverage COINS [61] to generate articulated 3D human bodies based on interaction semantics to make virtual humans controllable via interaction semantics and fine-grained body poses. +3. We design scene and interaction-aware policies to enable virtual humans to navigate in 3D scenes while avoiding collisions, to interact with scene objects, and to continuously perform sequences of activities in complex scenes. + +# 2. Related Work + +Human motion synthesis. Generating high-quality human motion has been widely explored in computer vision and graphics. Motion graph [22] and motion matching [4, 5, 63] generate motion by searching suitable clips in datasets and blending them automatically. Starke et al. [41-43] use phase-conditioned neural networks to synthesize character animations and interaction with objects. Zhang et al. [58, 60] model motion as sequences of surface markers on the parametric SMPL-X [31] body model, and train autoregressive networks on the large-scale mocap dataset AMASS [29] in order to produce diverse motions of bodies with various shapes. Peng et al. [33] use imitation learning with a skill discovery objective to learn a general motion skill space for physically simulated characters. Tang et al. + +[46] trains a motion manifold model of consecutive frames for real-time motion transition. Recent works [23, 24] propose generative methods to synthesize motions from single or a few example motion sequences. Transformer-based models have been designed to predict or generate stochastic motions conditioned on action categories [35], texts [36, 48], gaze [62], and others. More recently, motion diffusion models [1, 2, 49, 50, 54] achieve appealing performance on motion synthesis conditioned on various control signals and demonstrate flexible motion editing. + +Motion control and RL-based motion synthesis. Various motion control methods have been proposed to constrain body movements or guide the body to reach goals. Sampling-based motion control methods [27, 28] generate multiple samples at each step and select the samples that best match the targets. Goal-conditional generation networks [10, 17, 21] are applied for motion control. However, such methods may produce invalid results when the train-test domain gap is large. Optimization-based motion control methods [18, 53] leverage the learned generative motion model as regularization and optimize the motion latent variables to fit the decoded motion to the goals. Motion diffusion models [49, 54] implement control via classifier-free text guidance or gradually reprojecting the generated motion onto the physically plausible space at individual denoising steps. Human motions can be formulated as a Markov decision process, and synthesized and controlled via RL. Imitation learning methods [3, 30, 32, 34] trains policy networks to control humanoids to imitate reference motion and complete given tasks. Peng et al. [33] propose a skill discovery objective apart from the imitation objective to learn a latent space for general motion skills and train substream task policies leveraging the general motion space. Follow-up works [20, 47] combine the character controllers with language-based selection or finite state machine to compose more complex movements. Ling et al. [26] and Zhang et al. [60] leverage the latent variable of the generative motion model as the action, and train policies to generate high-fidelity perpetual motions. + +Human-scene interaction synthesis. Synthesizing natural human-scene interactions is an appealing direction in recent years. Object affordance [8, 9, 19, 38] provides cues about how humans can interact with scene objects. Static bodies interacting with scene objects can be generated based on the scene semantics, geometries, and interaction semantics [14, 25, 56, 59, 61]. Wang et al. [52] propose to train generative models for human motion in scenes conditioned on scene pointclouds. Wang et al. [51] propose a three-stage method to synthesize human motions in scenes, which first puts static bodies in scenes, then generates global paths, and afterwards synthesizes in-between + +motions. Starke [42] propose an autoregressive neural state machine to synthesize character animations interacting with objects. Hassan et al. [12] and Zhang et al. [57] propose similar networks as in [42], and incorporate stochastic motion generation and finer-grained motion control modules. Peng et al. [33] trains policy to control physically-simulated characters to hit a box. Hassan et al. [15] extend the method to generate interactions like carrying a box or sitting on a chair. Zhang et al. [55] learns physically simulated tennis players from broadcast videos. + +Ours versus others. Our method is most similar to GAMMA [60] and SAMP [12]. GAMMA learns generalizable motion models and policies across human bodies of diverse identities and shapes, without goal-motion paired training data. Despite producing high-fidelity motions, its results are limited to locomotion in the scene, and frequently collide with scene objects. SAMP learns conditional VAEs to produce sitting and lying actions in living environments, with object-motion paired data. Its generated motion has visible artifacts such as foot-ground skating. Our method combines their merits and eliminates their individual disadvantages. We extend the RL-based framework proposed in GAMMA by incorporating fine-grained motion controls (guided by interaction semantics) and scene interaction modules, so as to generate human-scene interactions in complex daily living environments. Compared to SAMP, our produced motion is more diverse, more physically plausible, and can be guided by fine-grained body surface markers. For systematic comparisons, please refer to Section 4. + +# 3. Method + +# 3.1. Preliminaries + +SMPL-X [31] and body representation. We use SMPL-X to represent 3D human bodies in our work. Given shape parameter $\beta \in \mathbb{R}^{10}$ and body pose parameters $\theta \in \mathbb{R}^{63}$ , SMPL-X produces a posed body mesh with a fixed topology of 10475 vertices. To place the body in a scene, the root location $r \in \mathbb{R}^3$ and the orientation $\phi \in \mathfrak{so}(3)$ w.r.t. the scene coordinates are additionally needed. Since facial expressions and hand gestures are not our focus, we leave their parameters as the default values. In addition, we follow [58, 60] to represent the body in motion by the SSM_67 body surface marker placement. A motion sequence is then formulated as $X = \{\pmb{x}_1, \dots, \pmb{x}_N\}$ , where $N$ is the length of motion and $\pmb{x}_i \in \mathbb{R}^{67 \times 3}$ denotes the body marker 3D locations at frame $i$ . The marker locations are relative to a canonical coordinate frame centered at the body pelvis in the first frame. + +GAMMA [60]. GAMMA can synthesize stochastic, perpetual, and realistic goal-reaching actions in 3D scenes. It comprises generative motion primitive models, RL-based control, and tree-based search to implement gradient-free test-time optimization. The motion primitive is formulated by a CVAE model to generate uncertain marker motions for 0.25 seconds into the future given a motion seed, followed by a MLP-based body regressor to yield SMPL-X parameters. Long-term random motion can be generated by running the motion primitive model recursively. The RL-based control is implemented by learning a policy within a simulation area. The actor-critic framework [45] and the PPO algorithm [39] are applied to update the policy network. An additional motion prior term is used to ensure the motion appears natural. During testing, the generated motion primitives are stored in a tree where only the best K primitives at each layer are preserved in order to discard low-quality sampling results. + +COINS[61]. COINS generates physically plausible static human-scene interactions with instance-level semantic control. Given the point cloud of an object and action labels, COINS can generate static bodies interacting with the given object based on the specified action, e.g., sitting, lying, or touching. COINS leverages transformer-based generative models trained on a human-scene interaction dataset [13] to first generate a plausible body pelvis for interaction and then the posed body. The generated bodies are further optimized to improve the physical plausibility and to match the predicted action-dependent contact areas with objects. Such generated static bodies capture the characteristics of the human-scene interaction process and can be used as fine-grained interaction guidance. + +# 3.2. RL-based Framework to Inhabit the Virtual + +As illustrated in Fig. 2, we propose a motion synthesis framework that enables virtual humans to navigate in complex indoor scenes and interact with various scene objects, e.g., sitting on a chair. Compared to the GAMMA framework [60], our method incorporates scene information into the states to better handle complex human-scene interactions. Also, we use body markers as goals to provide fine-grained guidance on how to drive the body for the target interactions. With modularized path-finding methods and static person-scene interaction generation methods, our framework can synthesize realistic human motions in complex 3D environments. In our work, we use COINS [61] to generate static person-scene interactions from interaction semantics given as 'action-object' pairs. The walking path can be either generated by hand, or by automatic path-finding algorithms like A* [11, 40]. + +We formulate our motion synthesis tasks with reinforcement learning. At each time step, a virtual human perceives + +![](images/6ec21d51e278f04132cbdb51576586bf4f954801548852095d8834d3b7535a56.jpg) +Figure 2: Illustration of our proposed human-scene interaction synthesis framework, which consists of learned motion primitive (actions), alongside locomotion and interaction policies generating latent actions conditioned on scenes and interaction goals. By integrating navigation mesh-based path-finding and static human-scene interaction generation methods, we can synthesize realistic motion sequences for virtual humans with fine-grained controls. + +its state $s_t$ in the environment and samples an action $\pmb{a}_t$ from its policy model $\pi(\pmb{a}_t | s_t)$ . Based on its motion model, it advances its motion state, and obtains a new perception state $s_{t+1}$ . A reward $r_t = r(s_t, \pmb{a}_t, s_{t+1})$ is calculated, tailored to different tasks. + +The motion model and the action. We leverage the CVAE-based generative motion primitive [60] as our motion model, and use its latent variables as actions. We train the model conditioned on 1 or 2 past frames using the combination of the SAMP [12] and AMASS [29] motion capture datasets, to learn a latent motion primitive space covering motion skills for human-scene interactions. Each latent variable $z$ in the motion primitive space is regarded as an action and can be decoded to a short clip of motion. + +The state. The state is formulated by + +$$ +\boldsymbol {s} _ {t} = \left(\boldsymbol {X} _ {s}, I, G\right), \tag {1} +$$ + +where $X_{s} \in \mathbb{R}^{M \times 67 \times 3}$ is the body markers motion seed that represents a motion history of $M$ frames. $I$ and $G$ denote the person-scene interaction feature and the goal-reaching feature, respectively. The interaction feature and goal-reaching feature vary among the locomotion and object interaction tasks. We introduce the detailed formulation in Sec. 3.3 and 3.4. + +The rewards. The rewards evaluate how well the virtual human performs locomotion and fine-grained object interaction tasks. We formulate rewards as + +$$ +r = r _ {\text {g o a l}} + r _ {\text {c o n t a c t}} + r _ {\text {p e n e}}, \tag {2} +$$ + +where $r_{\text{goal}}$ , $r_{\text{contact}}$ , and $r_{\text{pene}}$ represent the rewards for goal-reaching, foot-ground contact, and penetration avoidance, respectively. Specifically, the contact reward $r_{\text{contact}}$ encourages foot-floor contact and discourages foot skating, and is defined as: + +$$ +r _ {\text {c o n t a c t}} = e ^ {- \left(\left| \min x _ {z} \right| - 0. 0 5\right) _ {+}} \cdot e ^ {- \left(\min \| x _ {v e l} \| _ {2} - 0. 0 7 5\right) _ {+}}, \tag {3} +$$ + +where $F$ is the set of foot markers, $\pmb{x}_z$ is the height of the markers, $\pmb{x}_{vel}$ is the velocity of the markers, and $(\cdot)_+$ denotes clipping negative values. There are tolerance thresholds of $0.05\mathrm{m}$ for foot-floor distance and $0.075\mathrm{m/s}$ for skating, following GAMMA [60]. The other two reward terms are action-specific and introduced in Sec. 3.3 and 3.4. + +Policy network and training. We use the actor-critic algorithm [45] to learn the policy, where a policy network and a value network are trained jointly. The policy network generates a diagonal Gaussian distribution representing the stochastic action distribution given a state, while the value network outputs the value estimation for each state. + +Like in GAMMA [60], these two networks are jointly trained by minimizing: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {P P O} + \mathbb {E} \left[ \left(r _ {t} - V (s _ {t})\right) ^ {2} \right] \\ + \alpha \operatorname {K L} - \operatorname {d i v} (\pi (z | s) | | \mathcal {N} (0, I)), \\ \end{array} +$$ + +where the first term is the PPO [39] loss, the second updates the value estimation of the critic networks, and the third Kullback-Leibler divergence term regularizes motion in the latent space [60]. + +Tree sampling for test-time optimization. Given the stochastic nature of our Gaussian policies, sampling motions from the generated action distributions can yield motion primitive results of various qualities. Therefore, we follow [60] to use tree-based sampling during inference to discard motion primitives with inferior goal-reaching and scene interaction scores. Specifically, we sample multiple latent actions at each time step and selectively keep the best K samples, utilizing the same rewards used to train the policies as the selection criteria. This tree-sampling technique yields improved synthesis results of higher quality. + +In the following, we elaborate on the design of states and rewards tailored to different actions, namely locomotion and fine-grained object interaction. By combining the learned policies, long-term and coherent motions can be composed by rolling out the initial body and switching between the locomotion and object interaction stages. + +![](images/a373935eda6a2537d67285d94c779f0187ce05f72ee8bdced70a8d92d0c566eb.jpg) +Figure 3: Illustration of the scene-aware locomotion policy network. The locomotion policy state consists of the body markers, the goal-reaching feature of normalized direction vectors from markers to the goal pelvis, and the interaction feature of a 2D binary map indicating the walkability (red: non-walkable area, blue: walkable area) of the local $1.6m \times 1.6m$ square area. The locomotion policy network employs the actor-critic architecture and shares the state encoder. + +# 3.3. Scene-Aware Locomotion Synthesis + +Navigating in cluttered scenes here means the human body moving to a target while avoiding collisions with scene objects. Our key idea is to incorporate the walkability information of the surrounding environment into the states and use collision rewards to train the locomotion policy in order to avoid scene collisions. Specifically, we represent the walkability of the environment surrounding the human agent using a 2D binary map $\mathcal{M} \in \{0,1\}^{16 \times 16}$ as illustrated in Fig. 3. + +The walkability map is defined in the human's local coordinates and covers a $1.6m \times 1.6m$ area centered at the body pelvis and aligned with the body facing orientation. It consists of a $16 \times 16$ cell grid where each cell stores a binary value indicating whether this cell is walkable or not. This local walkability map enables the policy to sense surrounding obstacles. + +Referring to Eq. 1, the person-scene interaction feature is specified by + +$$ +I = \operatorname {v e c} (\mathcal {M}), \tag {5} +$$ + +in which $\mathrm{vec}(\cdot)$ denotes vectorization. The goal-reaching feature is specified by + +$$ +G = \left(\tilde {\boldsymbol {g}} _ {p} - \boldsymbol {X} _ {s}\right) _ {n}, \tag {6} +$$ + +where $\mathbf{X}_s$ and $\tilde{\pmb{g}}_p \in \mathbb{R}^{M \times 67 \times 3}$ are the body marker seed representing $M$ frames history of motion, and the broadcasted target pelvis location relative to the body-centered canonical coordinate, respectively. $(\tilde{\pmb{g}}_p - \mathbf{X}_s)_n$ denotes the normalized vectors pointing from each marker to the goal pelvis. + +The rewards contributing to Eq. 2 are defined as + +$$ +r _ {p e n e} = e ^ {- | \mathcal {M} _ {0} \cap \mathcal {B} _ {x y} (X) |}, \tag {7} +$$ + +where $\mathcal{M}_0$ denotes the non-walkable cells in the walkability map, $\mathcal{B}_{xy}(\cdot)$ denotes the 2D bounding box of the body markers $X$ , $\cap$ denotes their intersection, and $|\cdot|$ denotes the number of non-walkable cells overlapping with the human bounding box. + +$$ +r _ {\text {g o a l}} = r _ {\text {d i s t}} + r _ {\text {o r i}}, \tag {8} +$$ + +$$ +r _ {d i s t} = 1 - \left(\left\| \boldsymbol {p} - \boldsymbol {g} _ {p} \right\| _ {2} - 0. 0 5\right) _ {+}, \tag {9} +$$ + +$$ +r _ {o r i} = \frac {\left\langle \boldsymbol {o} , \boldsymbol {g} _ {p} - \boldsymbol {p} \right\rangle}{2}, \tag {10} +$$ + +where $r_{dist}$ encourages the body pelvis $p$ to be close to the pelvis goal $g_p$ and $r_{ori}$ encourages the body facing direction $o$ to be aligned with the direction from the current body pelvis $p$ to the pelvis goal $g_p$ . + +# 3.4. Fine-grained Object Interaction Synthesis + +To synthesize fine-grained human-object interactions, e.g. sitting on a chair or lying on a sofa, we use body marker goals as guidance, and model the proximity between the body surface and the scene object in a compact way. The goal marker sets can be generated by static person-scene interaction methods such as [14, 56, 61]. We use COINS [61] to generate the static goal interaction body for its performance and controllability of interaction semantics. + +In addition to the marker-based goal guidance, we incorporate the proximity relations between humans and objects into the states. Specifically, we use the signed distance from each marker to the surface of the scene object, as well as the gradient direction of the signed distance to represent the proximity relationship, as illustrated in Fig. 4. Both the signed distance and its gradient direction are calculated using the object's signed distance field (SDF) $\Psi_{O}$ . + +Referring to Eq. 1, the person-scene interaction feature is formulated as + +$$ +I = \left[ \Psi_ {O} \left(\boldsymbol {X} _ {s}\right), \nabla \Psi_ {O} \left(\boldsymbol {X} _ {s}\right) \right], \tag {11} +$$ + +in which $\Psi_O\in \mathbb{R}^{M\times 67}$ and $\nabla \Psi_{O}\in \mathbb{R}^{M\times 201}$ denote the SDF values and the gradient at each marker location in the $M$ frames, respectively, and $[\cdot ,\cdot ]$ denotes feature concatenation. The goal-reaching feature is formulated as + +$$ +G = \left[ \left(\tilde {\boldsymbol {g}} _ {m} - \boldsymbol {X} _ {s}\right) _ {n}, \| \tilde {\boldsymbol {g}} _ {m} - \boldsymbol {X} _ {s} \| _ {2} \right], \tag {12} +$$ + +in which $\mathbf{X}_s \in \mathbb{R}^{M \times 67 \times 3}$ denotes the body markers seed, $\tilde{\pmb{g}}_m \in \mathbb{R}^{M \times 67 \times 3}$ denotes the broadcasted goal body markers, $(\tilde{\pmb{g}}_m - \pmb{X}_s)_n$ denotes the normalized vector representing the direction from each marker to the corresponding target marker, $\| \tilde{\pmb{g}}_m - X_s \|_2$ denotes the distance from each marker to target marker. + +![](images/c4e160cfb625594ef9c430e44de2487f539b9e80123502c95f3b5e8007a25ed4.jpg) +Figure 4: Illustration of the object interaction policy network. The interaction policy state consists of the body markers, the goal-reaching features of both distance and direction from current markers to the goal markers, and the interaction features of the signed distances from each marker to the object surfaces and the signed distance gradient at each marker location. Such interaction features encode the human-object proximity relationship. + +The interaction policy is trained using the reward defined in Eq. 2 with the following interaction-specific goal reward and penetration reward: + +$$ +r _ {\text {g o a l}} = 1 - \left(\| \boldsymbol {x} - \boldsymbol {g} _ {m} \| _ {2} - 0. 0 5\right) _ {+}, \tag {13} +$$ + +$$ +r _ {p e n e} = e ^ {- \frac {1}{| V |} \frac {1}{T} \sum_ {t = 1} ^ {T} \sum_ {i = 1} ^ {| V |} \left| \left(\Psi_ {O} (\boldsymbol {v} _ {t i})\right) _ {-} \right|}, \tag {14} +$$ + +with $|V|$ being the SMPL-X mesh vertices, $T$ denotes the number of frames in each motion primitive (equals to 10 in our study). The distance reward encourages the final frame body markers $x$ to be close to the goal body markers $g_{m}$ . The penetration reward penalizes all the body vertices within a motion primitive that have negative SDF values. We use body vertices instead of joints because human-object contact happens on the body surface and can be better detected using vertices. + +Moreover, we train the interaction policy with a mixture of 'sit/lie down' and 'stand up' tasks. This training scheme enables the human agent to also learn how to stand up and transit from object interaction back to locomotion, which enables the synthesis of a sequence of interaction activities as in Fig. 1. + +# 4. Experiment + +Motion Datasets. We combine the large-scale motion capture dataset AMASS [29] with SAMP [12] motion data to train the motion primitive models. Each sequence is first subsampled to 40 FPS and then split into 10 frames and 100 frames clips. Each motion clip is canonicalized using the first frame body. Specifically, we select AMASS sequences annotated with 'sit' or 'lie' in BABEL [37] and all motion data of SAMP to train the motion primitive model. We train + +Table 1: Evaluation of the locomotion task. The up/down arrows denote the score is the higher/lower the better and the best results are in boldface. + +
time ↓avg. dist ↓contact ↑loco pene ↑
SAMP [12]5.970.140.840.94
GAMMA [60]3.870.030.940.94
Ours6.430.040.990.95
+ +separate motion primitive models for locomotion and interactions using task-related data. We observe that extending SAMP motion data with AMASS dataset is the key to making interaction policies work. + +Policy Training Environments. To train the scene-aware locomotion policy, we randomly generate synthetic cluttered scenes consisting of random objects from ShapeNet [6]. Random initial and target location pairs are sampled in the walkable areas using navigation meshes to train the locomotion policy. To train the interaction policy, we use the static person-object interaction data from PROX [13] and retargeted to ShapeNet objects. We first randomly sample a furniture and interaction body goal from the retargeted PROX data. Then we sample the initial body with random poses and locations in front of the object to train the interaction policy. We also randomly swap the initial and goal body to learn both 'sit/lie down' and 'stand up' motions. Please refer to Supp. Mat. for more details. + +# 4.1. Locomotion in 3D Scenes + +We randomly generate test scenes for locomotion in the same way as the training scenes. The virtual human is instructed to move from the random start point to the random target point while avoiding penetration with scene objects. + +Baselines and metrics. We compare our method with SAMP [12] and GAMMA [60] for locomotion. The SAMP results are recorded by running the released Unity demo. The start and termination are manually determined so the reported completion time may be slightly higher than the actual time due to human response time. The evaluation metrics for locomotion include: 1) time from start point to target point or reaching the time limit, measured in seconds. 2) the average distance from the final body to the targets, measured in meters. 3) foot contact score encouraging the lowest feet joints on the floor and discouraging foot skating as defined in Eq. 3. We use body joints instead of markers to calculate the contact score because the marker set annotation for the SAMP body is missing. 4) locomotion penetration score indicating the percentage of body vertices that are inside the walkable areas according to the navigation mesh. + +![](images/23cac8bda4be6cdde7c954ad3f3a9963844c75e07d4db5031a16b8ee855b88bd.jpg) + +![](images/ba5f7e324de279aab27e3347ff8ac17cb880930c85e80c848216f4baf9d18385.jpg) +Figure 5: Demonstration of locomotion tasks where GAMMA [60] (left) collides with the obstacles (red bodies) while our scene-aware locomotion policy (right) avoids collision. The yellow circles denote the specified waypoints. + +![](images/c6c56c14fc6383ab105da0b9de95c6944b39d916d8e346c64581cff149f52559.jpg) + +![](images/f64c7f963b62d65641324cc8ba336fdf83781c957d24b4af6b52440f73972df4.jpg) + +Results. Table 1 shows the empirical evaluation results. Our method achieves a higher contact score (0.99) than both GAMMA (0.94) and SAMP (0.84) which indicates better foot-floor contact and less foot skating. Moreover, our method achieves the highest penetration score which indicates our scene-aware policy can better avoid scene collisions. Fig. 5 shows examples of locomotion tasks where GAMMA collides into the scenes while our scene-aware policy successfully avoids penetration. Compared to GAMMA trained in the same environments, we observe our scene-aware policy learns more conservative behavior with lower moving speed and spent more time (6.43s) walking to the target locations, just like a human afraid of stepping on surrounding traps. All the methods can reach the target location within a reasonable distance. + +# 4.2. Fine-Grained Human-Object Interaction + +We evaluate the object interaction task on 10 unseen objects (3 armchairs, 3 straight chairs, 3 sofas, 1 L-sofa) from ShapeNet [6]. We use the object size annotation of ShapeNet and manually filter unrealistic-sized objects. The virtual human is randomly placed in front of the target object and then instructed to perform the interaction, stay for around 2 seconds, and then stand up. We evaluate two interactions of sitting and lying separately. + +Baselines and metrics. We compare our method with SAMP [12] for the object interaction task. The evaluation metrics for the interaction tasks are: 1) time of completing the object interaction task. 2) foot contact score as defined in Eq. 3. Note that the foot contact score does not always reflect the motion quality for lying tasks because the foot can + +Table 2: Evaluation of the interaction tasks. The up/down arrows denote the score is the higher/ lower the better and the best results are in boldface. + +
time ↓contact ↑pene. mean ↓pene. max ↓
SAMP [12] sit8.630.8911.9145.22
Ours sit4.090.971.9110.61
SAMP [12] lie12.550.7344.77238.81
Ours lie4.200.789.9044.61
+ +![](images/374b540cd5b261e574e784cec96c564c94f1b956630047236833d26581418edb.jpg) +Figure 6: Demonstration of various generated object interactions. Each row shows generated interactions with the same object starting from random initial body locations and orientations. Colors from yellow to red denote time. + +often be off the floor during lying. 3) interaction penetration score for each frame is defined as: + +$$ +s _ {i n t e r - p e n e} = \sum_ {v _ {i} \in V} | (\Psi_ {O} (v _ {i})) _ {-} |, \tag {15} +$$ + +where $\Psi_{O}$ is the object signed distance field, $(\cdot)_{-}$ clips all positive distance values to zero, and $V$ is the body vertivces. We show both the average penetration over time and the maximum penetration in one sequence. + +Results. Tab. 2 shows the evaluation results. Our method achieves a significantly higher foot contact score, indicating more natural motion. Our method also achieves lower mean and maximum penetration, which means our method generates more physically plausible results. Moreover, our method can complete the interaction tasks much faster than SAMP. This is because SAMP does not generalize well to unseen random body initialization and needs a longer time to start performing interactions. Qualitative results demonstrating the various object interactions generated by our method are shown in Fig. 6. Our object interaction policy generalizes to random initial body locations and orientations, as well as novel objects of various shapes. + +![](images/2214b1eaecfa4674311e718587f630b48b5b01d0e7129fdcfa01d93f12ba717e.jpg) +Figure 7: Our method can synthesize natural human movements in scene reconstructions that are significantly different from the training scenes. We show the synthesis results in one Paris street [7] (left figure) and one indoor room [13] (right figure). Transparent to solid colors denote time. + +![](images/a28c54f5987d85f635beb86f2f0649b3d61b9e6f3c7579ab603cdaacdf649914.jpg) + +Table 3: Evaluation of interaction sequences synthesis. Here 'contact' denotes the foot contact score, 'loco. pene.' is the percentage of body vertices inside the walkable areas, 'inter. pene. mean/max' denotes the mean and maximum penetration with interaction objects, and 'perceptual' denotes the ratio of being chosen as perceptually more natural. The best results are in boldface. + +
SAMP [12]Ours
contact ↑0.870.96
loco. pene. ↑0.620.72
inter. pene. mean↓15.613.40
inter. pene. max ↓101.2539.68
perceptual. ↑0.150.85
+ +# 4.3. Interaction Sequences in 3D Scenes + +Humans often continuously perform sequences of interactions with various objects in complex real-world scenes as shown in Fig. 1. Such interaction sequences are combinations of alternating locomotions and fine-grained interactions which we evaluate in Sec. 4.1 and Sec. 4.2 respectively. We conduct the empirical evaluation in real scene scans from Replica [44] and compare our method with SAMP. We select a list of interactable objects in the scene and instruct the virtual human to walk to the objects to perform interactions one by one. The evaluation metrics include the foot contact score, locomotion penetration score, and the mean and maximum interaction penetration score. We also conducted a perceptual study where participants are shown a side-by-side comparison of results from two methods and asked to choose the one perceptually more natural. We report the rate of being chosen as the better. + +The evaluation results are shown in Tab. 3. Our method generates high-quality results of human-scene interactions in cluttered environments. Our results achieve higher contact scores and less penetration with the scenes compared + +Table 4: Rewards ablation studies results, where '-Contact' and '-Penetration' denote policies trained without the floor contact and penetration avoidance reward, respectively. + +
time ↓contact ↑pene. mean ↓pene. max ↓
Ours4.090.971.9110.61
- Contact3.750.8819.4960.10
- Penetration4.030.9516.8045.50
+ +to SAMP. In addition, our method generates perceptually more natural results according to perceptual study. + +# 4.4. Ablation Studies + +We perform ablation studies on the foot-ground contact reward and penetration avoidance reward of the interaction policies to substantiate the significance of these rewards. We train two ablation interaction policies using identical networks and environments, differing only in the exclusion of one of these rewards in each case. The quantitative metrics are reported in Tab. 4. The removal of the penetration avoidance reward yields a marked escalation in detected human-scene penetration. The removal of the footfloor contact reward yields notably both inferior contact and penetration scores, accompanied by observed erratic synthesized motions that either remain suspended in the air or penetrate the floor. These empirical findings underscore the crucial role played by the foot-floor contact and penetration avoidance rewards in the learning of interaction policies. + +# 5. Discussion and Conclusion + +Limitations and Future Work. Our current method has various limitations that could be improved in future works. First, our method does not fully resolve penetrations with scene objects and floors since our method only uses rewards to encourage avoiding penetration with scenes, which does not impose hard constraints for penetrations. The combination of our method with physics simulation holds the potential to effectively address and resolve penetration issues. Moreover, the lying motions generated by our method are not as natural as the sitting motions because the motion primitive model fails to learn a comprehensive action space for lying given limited training data. Specifically, the available AMASS motion capture data for lying (167 seconds) is significantly less than the data for sitting (5K seconds). To overcome this issue, we aim to explore more data-efficient learning methods and scalable methods to collect human-scene interaction data. Furthermore, our locomotion policy is now limited to flat-floor scenes due to its reliance on the 2D occupancy-based walkability map. However, to broaden the applicability of our approach to more complex scenes such as uneven outdoor terrains and multi-floor buildings requiring walking upstairs, it will be necessary to replace the walkability map with a more suitable environment sens + +![](images/1b3ba8a7622d1b69fa19703f59b2f29b84b6eb9eea06cc7768ef58a2f8e989b0.jpg) +Inter-penetrations +Figure 8: Limitations. Penetrations remain observable in our results (left) since our method only encourages avoiding collision with rewards. Furthermore, the absence of sufficient training data for lying motions impedes the motion primitive model's ability to acquire a comprehensive lying motion space, leading to degraded motion (right). + +![](images/4e843826538511eb4c3b5a33b88fcca3d7d00b20dccde7f5f1bf36a8d6e79063.jpg) +Unnatural poses + +ing mechanism. In addition, our method is restricted to interactions with static scenes. However, in the real world, humans are exposed to dynamic interactions with movable objects and scenes involving autonomous agents including other humans, animals, and vehicles. Extension to dynamic interactions will enable tackling a broader range of intricate and dynamic applications that mirror the complexities inherent in the real world. + +Conclusion. In this paper, we leverage reinforcement learning to establish a framework to synthesize diverse human motions in indoor scenes, which is stochastic, realistic, and perpetual. The proposed method has large potential to improve many applications such as daily-living activity simulation, synthetic data creation, architecture design, and so on. Compared to existing methods, our method realizes fine-grained control by using body surface keypoints as targets, and achieves autonomous body-scene collision avoidance by incorporating scene information into the states and the rewards. Experiments show that our method effectively enables virtual humans to inhabit the virtual, and outperforms baselines consistently. + +Acknowledgements. We sincerely acknowledge the anonymous reviewers for their insightful suggestions. This work was supported by the SNSF project grant 200021 204840 and SDSC PhD fellowship. Lingchen Yang provided valuable suggestions on visualization. + +# References + +[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. ISBN: 0730-0301 Publisher: ACM New York, NY, USA. 3 +[2] Tenglong Ao, Zeyi Zhang, and Libin Liu. GestureDiffuCLIP: Gesture diffusion model with CLIP latents. arXiv preprint arXiv:2303.14613, 2023. 3 +[3] Kevin Bergamin, Simon Clavet, Daniel Holden, and James Richard Forbes. DReCon: data-driven responsive control of physics-based characters. ACM Transactions On Graphics (TOG), 38(6):1-11, 2019. Publisher: ACM New York, NY, USA. 3 +[4] Michael Buttner. Machine learning for motion synthesis and character control in games. Proc. of I3D 2019, 2, 2019. 2 +[5] Michael Böttner and Simon Clavet. Motion matching-the road to next gen animation. Proc. of Nucl. ai, 2015(1):2, 2015. 2 +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, and Hao Su. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 7 +[7] Jean-Emmanuel Deschaud, David Duque, Jean Pierre Richa, Santiago Velasco-Forero, Beatrix Marcotegui, and François Goulette. Paris-carla-3d: A real and synthetic outdoor point cloud dataset for challenging tasks in 3d mapping. Remote Sensing, 13(22), 2021. 8 +[8] Helmut Grabner, Juergen Gall, and Luc Van Gool. What makes a chair a chair? pages 1529-1536. IEEE, 2011. 3 +[9] Abhinav Gupta, Scott Satkin, Alexei A Efros, and Martial Hebert. From 3d scene geometry to human workspace. pages 1961-1968. IEEE, 2011. 3 +[10] Ikhsanul Habibie, Daniel Holden, Jonathan Schwarz, Joe Yearsley, and Taku Komura. A recurrent variational autoencoder for human motion synthesis. 2017. 3 +[11] Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100-107, 1968. Publisher: IEEE. 4 +[12] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. pages 11374-11384, 2021. 2, 3, 4, 6, 7, 8 +[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3D human pose ambiguities with 3D scene constraints. pages 2282-2292, 2019. 4, 7, 8 +[14] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3D scenes by learning human-scene interaction. pages 14708-14718, 2021. 3, 6 +[15] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing Physical Character-Scene Interactions, Feb. 2023. arXiv:2302.00883 [cs]. 3 +[16] Daniel Holden, Ouussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions + +on Graphics (TOG), 39(4):53-1, 2020. Publisher: ACM New York, NY, USA. 2 +[17] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. Publisher: ACM New York, NY, USA. 3 +[18] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. Publisher: ACM New York, NY, USA. 3 +[19] Ruizhen Hu, Zihao Yan, Jingwen Zhang, Oliver Van Kaick, Ariel Shamir, Hao Zhang, and Hui Huang. Predictive and generative neural networks for object functionality. arXiv preprint arXiv:2006.15520, 2020. 3 +[20] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. PADI: Language-Directed Physics-Based Character Control. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 3 +[21] Kacper Kania, Marek Kowalski, and Tomasz Trzciński. TrajeVAE: Controllable Human Motion Generation from Trajectories. arXiv preprint arXiv:2104.00351, 2021. 3 +[22] Lucas Kovar, Michael Gleicher, and Frédéric Pighin. Motion graphs. In ACM SIGGRAPH 2008 classes, pages 1-10. 2008. 2 +[23] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. Ganimator: Neural motion synthesis from a single sequence. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. Publisher: ACM New York, NY, USA. 3 +[24] Weiyu Li, Xuelin Chen, Peizhuo Li, Olga Sorkine-Hornung, and Baoquan Chen. Example-based Motion Synthesis via Generative Motion Matching. arXiv preprint arXiv:2306.00378, 2023. 3 +[25] Xueting Li, Sifei Liu, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Putting humans in a scene: Learning affordance in 3d indoor environments. pages 12368-12376, 2019. 3 +[26] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. Publisher: ACM New York, NY, USA. 2, 3 +[27] Libin Liu, KangKang Yin, and Baining Guo. Improving sampling-based motion control. volume 34, pages 415-423. Wiley Online Library, 2015. Issue: 2. 3 +[28] Libin Liu, KangKang Yin, Michiel Van de Panne, Tianjia Shao, and Weiwei Xu. Sampling-based contact-rich motion control. In ACM SIGGRAPH 2010 papers, pages 1-10. 2010. 3 +[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. pages 5442-5451, 2019. 2, 4, 6 +[30] Josh Merel, Leonard Hasenclever, Alexandre Galashov, Arun Ahuja, Vu Pham, Greg Wayne, Yee Whye Teh, and Nicolas Heess. Neural probabilistic motor primitives for humanoid control. arXiv preprint arXiv:1811.11711, 2018. 3 +[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, + +and body from a single image. pages 10975-10985, 2019. 2, 3 +[32] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. Publisher: ACM New York, NY, USA. 3 +[33] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. Publisher: ACM New York, NY, USA. 2, 3 +[34] Xue Bin Peng, Angjoo Kanazawa, Jitendra Malik, Pieter Abbeel, and Sergey Levine. Sfv: Reinforcement learning of physical skills from videos. ACM Transactions On Graphics (TOG), 37(6):1-14, 2018. Publisher: ACM New York, NY, USA. 3 +[35] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. pages 10985-10995, 2021. 3 +[36] Mathis Petrovich, Michael J Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. pages 480-497. Springer, 2022. 3 +[37] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. BABEL: bodies, action and behavior with English labels. pages 722-731, 2021. 6 +[38] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. SceneGrok: Inferring action maps in 3D environments. ACM transactions on graphics (TOG), 33(6):1-10, 2014. Publisher: ACM New York, NY, USA. 3 +[39] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 4, 5 +[40] Greg Snook. Simplified 3D movement and pathfinding using navigation meshes. Game programming gems, 1(1):288-304, 2000. Publisher: Charles River Media Newton Centre, MA, USA. 4 +[41] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. Publisher: ACM New York, NY, USA. 2 +[42] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 2, 3 +[43] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. Publisher: ACM New York, NY, USA. 2 +[44] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, and Shobhit Verma. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.8 +[45] Richard S Sutton and Andrew G Barto. Introduction to reinforcement learning, volume 135. MIT press Cambridge, 1998. 2, 4, 5 +[46] Xiangjun Tang, He Wang, Bo Hu, Xu Gong, Ruifan Yi, Qi + +long Kou, and Xiaogang Jin. Real-time controllable motion transition for characters. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. Publisher: ACM New York, NY, USA. 3 +[47] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. CALM: Conditional Adversarial Latent Models for Directable Virtual Characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3 +[48] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. pages 358–374. Springer, 2022. 3 +[49] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3 +[50] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. EDGE: Editable Dance Generation From Music. arXiv preprint arXiv:2211.10658, 2022. 3 +[51] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. pages 20460-20469, 2022. 3 +[52] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. pages 9401-9411, 2021. 3 +[53] Zhiyong Wang, Jinxiang Chai, and Shihong Xia. Combining recurrent neural networks and adversarial training for human motion synthesis and control. IEEE transactions on visualization and computer graphics, 27(1):14-28, 2019. Publisher: IEEE. 3 +[54] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. PhysDiff: Physics-Guided Human Motion Diffusion Model. arXiv preprint arXiv:2212.02500, 2022.3 +[55] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning Physically Simulated Tennis Skills from Broadcast Videos. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. ISBN: 0730-0301 Publisher: ACM New York, NY, USA. 3 +[56] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. PLACE: Proximity learning of articulation and contact in 3D environments. pages 642-651. IEEE, 2020. 3, 6 +[57] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: towards controllable human-chair interactions. pages 518-535. Springer, 2022. 2, 3 +[58] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. pages 3372-3382, 2021. 2, 3 +[59] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes without people. pages 6194-6204, 2020. 3 +[60] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3D scenes. pages 20481-20491, 2022. 2, 3, 4, 5, 7 +[61] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. pages 311-327. Springer, 2022. 2, 3, 4, 6 +[62] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao + +Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. pages 676-694. Springer, 2022. 3 +[63] Fabio Zinno. MI tutorial day: From motion matching to motion synthesis, and all the hurdles in between. Proc. of GDC 2019, 2, 2019. 2 \ No newline at end of file diff --git a/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/images.zip b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9b5441193fe4f43ad9339a7d40d69e8af592430b --- /dev/null +++ b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19b126e6dcc3f85a15dffe1123ac09a952baf3711b361d88f8f06d7b9174f628 +size 371910 diff --git a/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/layout.json b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5cd132c135a432061b8ef3a16af198abbfa1d17a --- /dev/null +++ b/2023/Synthesizing Diverse Human Motions in 3D Indoor Scenes/layout.json @@ -0,0 +1,8807 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 477, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 477, + 121 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 477, + 121 + ], + "type": "text", + "content": "Synthesizing Diverse Human Motions in 3D Indoor Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 141, + 487, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 141, + 487, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 487, + 172 + ], + "type": "text", + "content": "Kaifeng Zhao1 Yan Zhang1 Shaofei Wang1 Thabo Beeler2 Siyu Tang1 ETH Zürich 2Google" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 174, + 465, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 174, + 465, + 199 + ], + "spans": [ + { + "bbox": [ + 129, + 174, + 465, + 199 + ], + "type": "text", + "content": "{kaifeng.zhao, yan.zhang, shaofei.wang, siyu.tang}@inf.ethz.ch thabo.beeler@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 227, + 545, + 417 + ], + "blocks": [ + { + "bbox": [ + 48, + 227, + 545, + 417 + ], + "lines": [ + { + "bbox": [ + 48, + 227, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 48, + 227, + 545, + 417 + ], + "type": "image", + "image_path": "92394e5bf069ee29114dba10b1473a40d93f8ac25a090519db68cc9e4ffe965b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 418, + 547, + 455 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 547, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 547, + 455 + ], + "type": "text", + "content": "Figure 1: In this work, we propose a method to generate a sequence of natural human-scene interaction events in real-world complex scenes as illustrated here. The human first walks to sit on a stool (yellow to red), then walk to another chair to sit down (red to magenta), and finally walk to and lie on the sofa (magenta to blue)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 473, + 192, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 473, + 192, + 487 + ], + "spans": [ + { + "bbox": [ + 143, + 473, + 192, + 487 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "We present a novel method for populating 3D indoor scenes with virtual humans that can navigate in the environment and interact with objects in a realistic manner. Existing approaches rely on high-quality training sequences that contain captured human motions and the 3D scenes they interact with. However, such interaction data are costly, difficult to capture, and can hardly cover the full range of plausible human-scene interactions in complex indoor environments. To address these challenges, we propose a reinforcement learning-based approach that enables virtual humans to navigate in 3D scenes and interact with objects realistically and autonomously, driven by learned motion control policies. The motion control policies employ latent motion action spaces, which correspond to realistic motion primitives and are learned from large-scale motion capture data using a powerful generative motion model. For navigation in a 3D environment, we propose a scene-aware policy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 475, + 547, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 690 + ], + "type": "text", + "content": "with novel state and reward designs for collision avoidance. Combined with navigation mesh-based path-finding algorithms to generate intermediate waypoints, our approach enables the synthesis of diverse human motions navigating in 3D indoor scenes and avoiding obstacles. To generate fine-grained human-object interactions, we carefully curate interaction goal guidance using a marker-based body representation and leverage features based on the signed distance field (SDF) to encode human-scene proximity relations. Our method can synthesize realistic and diverse human-object interactions (e.g., sitting on a chair and then getting up) even for out-of-distribution test scenarios with different object shapes, orientations, starting body positions, and poses. Experimental results demonstrate that our approach outperforms state-of-the-art human-scene interaction synthesis methods in terms of both motion naturalness and diversity. Code, models, and demonstrative video results are available at: https://zkf1997.github.io/DIMOS." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "14738" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 330 + ], + "type": "text", + "content": "Simulating how humans interact with environments plays an essential role in many applications, such as generating training data for machine learning algorithms, and simulating autonomous agents in AR/VR and computer games. Although this task is highly related to character animation in computer graphics, most existing character animation methods (e.g. [4, 5, 16]) focus on improving the realism and controllability of character movements. With the traditional character animation workflows, one can produce high-quality animations but can hardly generate autonomous and spontaneous natural human motions interacting with the surroundings in diverse plausible ways as real humans. Previous learning-based interaction synthesis methods [12, 42, 57] require simultaneously capturing human motion and scenes for supervision. However, capturing such training data is costly and challenging, resulting in a notably limited spectrum of human-scene interaction motions and difficulties in handling unseen interaction scenarios. This restriction also results in inferior motion quality of synthesized virtual humans." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 331, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 287, + 534 + ], + "type": "text", + "content": "To address this problem, we leverage reinforcement learning (RL) [45] to solve our task. By formulating goals as rewards, perception as states, and latent variables of deep generative models as actions, we can synthesize continuous, stochastic, plausible, and spontaneous motions of virtual humans to inhabit the digital world. Although existing RL-based motion synthesis approaches (e.g. [26, 33, 60]) can effectively generate natural motions to achieve goals, their generated virtual humans can only interact with simple scenes, rather than complex environments with functional furniture and diverse objects. For example, GAMMA [60] employs generative motion primitives and a policy network that are generalizable across diverse human body shapes, but it can only synthesize waypoint-reaching locomotions. The trained digital humans are not aware of how to perform actions like sitting on a chair or lying on a sofa, and frequently inter-penetrate with the scene geometry." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 534, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 288, + 714 + ], + "type": "text", + "content": "To overcome these limitations, we propose a novel framework to learn both scene and interaction-aware motion control policies for synthesizing realistic and diverse human-scene interaction motions. First, in order to improve the physical plausibility of the synthesized human motions, we design a new scene-aware policy to help virtual humans avoid collisions with scene objects. Specifically, we use a 2D occupancy-based local walkability map to incorporate scene information into the locomotion policy. In addition, we add features derived from the signed distance from body markers to the object surface and the gradient direction of the signed distance to encode the proximity between humans and objects for object interaction policies. Second, in order to achieve controllable object interactions, we provide fine-grained guidances based on surface mark" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "ers [58] of a human body performing the target interactions. Specifically, we use COINS [61] to generate human bodies interacting with scene objects given the interaction semantics, and then use the body markers as the interaction guidance for motion synthesis. Combined with navigation mesh-based path-finding algorithms to generate intermediate waypoints in 3D scenes, virtual humans can autonomously reach target locations in complex environments and mimic target poses in a variety of plausible ways." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 180, + 546, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 546, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 546, + 299 + ], + "type": "text", + "content": "We train the policy networks in synthetic scenes consisting of randomized objects to learn generalizable scene-aware locomotion and fine-grained object interactions. With this framework, we investigate how to synthesize diverse in-scene motions consisting of locomotion, sitting, and lying. We empirically evaluate the motion realism and expressiveness of our proposed method, and compare it with state-of-the-art methods. The results show that our approach consistently outperforms the baselines in terms of diversity, physical plausibility, and perceptual scores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 300, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 323 + ], + "type": "text", + "content": "In summary, we aim to let virtual humans inhabit virtual environments, and present these contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 545, + 515 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 331, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 331, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 314, + 331, + 545, + 367 + ], + "type": "text", + "content": "1. We propose a reinforcement learning-based framework to generate realistic and diverse motions of virtual humans in complex indoor scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 375, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 545, + 449 + ], + "type": "text", + "content": "2. We propose to use body surface markers as detailed interaction goals for fine-grained human-object interaction synthesis and leverage COINS [61] to generate articulated 3D human bodies based on interaction semantics to make virtual humans controllable via interaction semantics and fine-grained body poses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 455, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 545, + 515 + ], + "type": "text", + "content": "3. We design scene and interaction-aware policies to enable virtual humans to navigate in 3D scenes while avoiding collisions, to interact with scene objects, and to continuously perform sequences of activities in complex scenes." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 526, + 392, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 392, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 392, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Human motion synthesis. Generating high-quality human motion has been widely explored in computer vision and graphics. Motion graph [22] and motion matching [4, 5, 63] generate motion by searching suitable clips in datasets and blending them automatically. Starke et al. [41-43] use phase-conditioned neural networks to synthesize character animations and interaction with objects. Zhang et al. [58, 60] model motion as sequences of surface markers on the parametric SMPL-X [31] body model, and train autoregressive networks on the large-scale mocap dataset AMASS [29] in order to produce diverse motions of bodies with various shapes. Peng et al. [33] use imitation learning with a skill discovery objective to learn a general motion skill space for physically simulated characters. Tang et al." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "14739" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "[46] trains a motion manifold model of consecutive frames for real-time motion transition. Recent works [23, 24] propose generative methods to synthesize motions from single or a few example motion sequences. Transformer-based models have been designed to predict or generate stochastic motions conditioned on action categories [35], texts [36, 48], gaze [62], and others. More recently, motion diffusion models [1, 2, 49, 50, 54] achieve appealing performance on motion synthesis conditioned on various control signals and demonstrate flexible motion editing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 555 + ], + "type": "text", + "content": "Motion control and RL-based motion synthesis. Various motion control methods have been proposed to constrain body movements or guide the body to reach goals. Sampling-based motion control methods [27, 28] generate multiple samples at each step and select the samples that best match the targets. Goal-conditional generation networks [10, 17, 21] are applied for motion control. However, such methods may produce invalid results when the train-test domain gap is large. Optimization-based motion control methods [18, 53] leverage the learned generative motion model as regularization and optimize the motion latent variables to fit the decoded motion to the goals. Motion diffusion models [49, 54] implement control via classifier-free text guidance or gradually reprojecting the generated motion onto the physically plausible space at individual denoising steps. Human motions can be formulated as a Markov decision process, and synthesized and controlled via RL. Imitation learning methods [3, 30, 32, 34] trains policy networks to control humanoids to imitate reference motion and complete given tasks. Peng et al. [33] propose a skill discovery objective apart from the imitation objective to learn a latent space for general motion skills and train substream task policies leveraging the general motion space. Follow-up works [20, 47] combine the character controllers with language-based selection or finite state machine to compose more complex movements. Ling et al. [26] and Zhang et al. [60] leverage the latent variable of the generative motion model as the action, and train policies to generate high-fidelity perpetual motions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Human-scene interaction synthesis. Synthesizing natural human-scene interactions is an appealing direction in recent years. Object affordance [8, 9, 19, 38] provides cues about how humans can interact with scene objects. Static bodies interacting with scene objects can be generated based on the scene semantics, geometries, and interaction semantics [14, 25, 56, 59, 61]. Wang et al. [52] propose to train generative models for human motion in scenes conditioned on scene pointclouds. Wang et al. [51] propose a three-stage method to synthesize human motions in scenes, which first puts static bodies in scenes, then generates global paths, and afterwards synthesizes in-between" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "content": "motions. Starke [42] propose an autoregressive neural state machine to synthesize character animations interacting with objects. Hassan et al. [12] and Zhang et al. [57] propose similar networks as in [42], and incorporate stochastic motion generation and finer-grained motion control modules. Peng et al. [33] trains policy to control physically-simulated characters to hit a box. Hassan et al. [15] extend the method to generate interactions like carrying a box or sitting on a chair. Zhang et al. [55] learns physically simulated tennis players from broadcast videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 220, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 547, + 461 + ], + "type": "text", + "content": "Ours versus others. Our method is most similar to GAMMA [60] and SAMP [12]. GAMMA learns generalizable motion models and policies across human bodies of diverse identities and shapes, without goal-motion paired training data. Despite producing high-fidelity motions, its results are limited to locomotion in the scene, and frequently collide with scene objects. SAMP learns conditional VAEs to produce sitting and lying actions in living environments, with object-motion paired data. Its generated motion has visible artifacts such as foot-ground skating. Our method combines their merits and eliminates their individual disadvantages. We extend the RL-based framework proposed in GAMMA by incorporating fine-grained motion controls (guided by interaction semantics) and scene interaction modules, so as to generate human-scene interactions in complex daily living environments. Compared to SAMP, our produced motion is more diverse, more physically plausible, and can be guided by fine-grained body surface markers. For systematic comparisons, please refer to Section 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 479, + 362, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 362, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 362, + 491 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 501, + 392, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 392, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 392, + 513 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "SMPL-X [31] and body representation. We use SMPL-X to represent 3D human bodies in our work. Given shape parameter " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": " and body pose parameters " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{63}" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": ", SMPL-X produces a posed body mesh with a fixed topology of 10475 vertices. To place the body in a scene, the root location " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "r \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": " and the orientation " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\phi \\in \\mathfrak{so}(3)" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": " w.r.t. the scene coordinates are additionally needed. Since facial expressions and hand gestures are not our focus, we leave their parameters as the default values. In addition, we follow [58, 60] to represent the body in motion by the SSM_67 body surface marker placement. A motion sequence is then formulated as " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "X = \\{\\pmb{x}_1, \\dots, \\pmb{x}_N\\}" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": " is the length of motion and " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i \\in \\mathbb{R}^{67 \\times 3}" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": " denotes the body marker 3D locations at frame " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": ". The marker locations are relative to a canonical coordinate frame centered at the body pelvis in the first frame." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "14740" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "GAMMA [60]. GAMMA can synthesize stochastic, perpetual, and realistic goal-reaching actions in 3D scenes. It comprises generative motion primitive models, RL-based control, and tree-based search to implement gradient-free test-time optimization. The motion primitive is formulated by a CVAE model to generate uncertain marker motions for 0.25 seconds into the future given a motion seed, followed by a MLP-based body regressor to yield SMPL-X parameters. Long-term random motion can be generated by running the motion primitive model recursively. The RL-based control is implemented by learning a policy within a simulation area. The actor-critic framework [45] and the PPO algorithm [39] are applied to update the policy network. An additional motion prior term is used to ensure the motion appears natural. During testing, the generated motion primitives are stored in a tree where only the best K primitives at each layer are preserved in order to discard low-quality sampling results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 288, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 288, + 471 + ], + "type": "text", + "content": "COINS[61]. COINS generates physically plausible static human-scene interactions with instance-level semantic control. Given the point cloud of an object and action labels, COINS can generate static bodies interacting with the given object based on the specified action, e.g., sitting, lying, or touching. COINS leverages transformer-based generative models trained on a human-scene interaction dataset [13] to first generate a plausible body pelvis for interaction and then the posed body. The generated bodies are further optimized to improve the physical plausibility and to match the predicted action-dependent contact areas with objects. Such generated static bodies capture the characteristics of the human-scene interaction process and can be used as fine-grained interaction guidance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 479, + 277, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 277, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 277, + 491 + ], + "type": "text", + "content": "3.2. RL-based Framework to Inhabit the Virtual" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 689 + ], + "type": "text", + "content": "As illustrated in Fig. 2, we propose a motion synthesis framework that enables virtual humans to navigate in complex indoor scenes and interact with various scene objects, e.g., sitting on a chair. Compared to the GAMMA framework [60], our method incorporates scene information into the states to better handle complex human-scene interactions. Also, we use body markers as goals to provide fine-grained guidance on how to drive the body for the target interactions. With modularized path-finding methods and static person-scene interaction generation methods, our framework can synthesize realistic human motions in complex 3D environments. In our work, we use COINS [61] to generate static person-scene interactions from interaction semantics given as 'action-object' pairs. The walking path can be either generated by hand, or by automatic path-finding algorithms like A* [11, 40]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We formulate our motion synthesis tasks with reinforcement learning. At each time step, a virtual human perceives" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 74, + 541, + 266 + ], + "blocks": [ + { + "bbox": [ + 322, + 74, + 541, + 266 + ], + "lines": [ + { + "bbox": [ + 322, + 74, + 541, + 266 + ], + "spans": [ + { + "bbox": [ + 322, + 74, + 541, + 266 + ], + "type": "image", + "image_path": "6ec21d51e278f04132cbdb51576586bf4f954801548852095d8834d3b7535a56.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 273, + 547, + 370 + ], + "lines": [ + { + "bbox": [ + 304, + 273, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 547, + 370 + ], + "type": "text", + "content": "Figure 2: Illustration of our proposed human-scene interaction synthesis framework, which consists of learned motion primitive (actions), alongside locomotion and interaction policies generating latent actions conditioned on scenes and interaction goals. By integrating navigation mesh-based path-finding and static human-scene interaction generation methods, we can synthesize realistic motion sequences for virtual humans with fine-grained controls." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "text", + "content": "its state " + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "text", + "content": " in the environment and samples an action " + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\pmb{a}_t" + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "text", + "content": " from its policy model " + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "inline_equation", + "content": "\\pi(\\pmb{a}_t | s_t)" + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "text", + "content": ". Based on its motion model, it advances its motion state, and obtains a new perception state " + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "inline_equation", + "content": "s_{t+1}" + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "text", + "content": ". A reward " + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "inline_equation", + "content": "r_t = r(s_t, \\pmb{a}_t, s_{t+1})" + }, + { + "bbox": [ + 304, + 390, + 545, + 450 + ], + "type": "text", + "content": " is calculated, tailored to different tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 464, + 546, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 546, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 546, + 573 + ], + "type": "text", + "content": "The motion model and the action. We leverage the CVAE-based generative motion primitive [60] as our motion model, and use its latent variables as actions. We train the model conditioned on 1 or 2 past frames using the combination of the SAMP [12] and AMASS [29] motion capture datasets, to learn a latent motion primitive space covering motion skills for human-scene interactions. Each latent variable " + }, + { + "bbox": [ + 304, + 464, + 546, + 573 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 464, + 546, + 573 + ], + "type": "text", + "content": " in the motion primitive space is regarded as an action and can be decoded to a short clip of motion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 586, + 465, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 465, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 465, + 598 + ], + "type": "text", + "content": "The state. The state is formulated by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 390, + 608, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 608, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 390, + 608, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {t} = \\left(\\boldsymbol {X} _ {s}, I, G\\right), \\tag {1}", + "image_path": "7f9218c8a1a36a3d0b07b8b9947482961f301ca7e465bb3ba007945a3b91a427.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "inline_equation", + "content": "X_{s} \\in \\mathbb{R}^{M \\times 67 \\times 3}" + }, + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "text", + "content": " is the body markers motion seed that represents a motion history of " + }, + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "text", + "content": " frames. " + }, + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 628, + 547, + 712 + ], + "type": "text", + "content": " denote the person-scene interaction feature and the goal-reaching feature, respectively. The interaction feature and goal-reaching feature vary among the locomotion and object interaction tasks. We introduce the detailed formulation in Sec. 3.3 and 3.4." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "14741" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 107 + ], + "type": "text", + "content": "The rewards. The rewards evaluate how well the virtual human performs locomotion and fine-grained object interaction tasks. We formulate rewards as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 118, + 287, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 118, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 287, + 129 + ], + "type": "interline_equation", + "content": "r = r _ {\\text {g o a l}} + r _ {\\text {c o n t a c t}} + r _ {\\text {p e n e}}, \\tag {2}", + "image_path": "19a4101c8b95c70498edd678fded904f8a4792eda4389c3a4e013be761f2a402.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "type": "inline_equation", + "content": "r_{\\text{goal}}" + }, + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "type": "inline_equation", + "content": "r_{\\text{contact}}" + }, + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "type": "inline_equation", + "content": "r_{\\text{pene}}" + }, + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "type": "text", + "content": " represent the rewards for goal-reaching, foot-ground contact, and penetration avoidance, respectively. Specifically, the contact reward " + }, + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "type": "inline_equation", + "content": "r_{\\text{contact}}" + }, + { + "bbox": [ + 47, + 137, + 287, + 196 + ], + "type": "text", + "content": " encourages foot-floor contact and discourages foot skating, and is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 201, + 287, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 201, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 52, + 201, + 287, + 217 + ], + "type": "interline_equation", + "content": "r _ {\\text {c o n t a c t}} = e ^ {- \\left(\\left| \\min x _ {z} \\right| - 0. 0 5\\right) _ {+}} \\cdot e ^ {- \\left(\\min \\| x _ {v e l} \\| _ {2} - 0. 0 7 5\\right) _ {+}}, \\tag {3}", + "image_path": "00c3d1031f164a7a10ee48b96d9779d3ea06ba3a9a9cfecd66af1d3d932835fe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "text", + "content": " is the set of foot markers, " + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\pmb{x}_z" + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "text", + "content": " is the height of the markers, " + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{vel}" + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "text", + "content": " is the velocity of the markers, and " + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "inline_equation", + "content": "(\\cdot)_+" + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "text", + "content": " denotes clipping negative values. There are tolerance thresholds of " + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "inline_equation", + "content": "0.05\\mathrm{m}" + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "text", + "content": " for foot-floor distance and " + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "inline_equation", + "content": "0.075\\mathrm{m/s}" + }, + { + "bbox": [ + 47, + 224, + 287, + 295 + ], + "type": "text", + "content": " for skating, following GAMMA [60]. The other two reward terms are action-specific and introduced in Sec. 3.3 and 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 310, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 287, + 380 + ], + "type": "text", + "content": "Policy network and training. We use the actor-critic algorithm [45] to learn the policy, where a policy network and a value network are trained jointly. The policy network generates a diagonal Gaussian distribution representing the stochastic action distribution given a state, while the value network outputs the value estimation for each state." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 381, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 287, + 407 + ], + "type": "text", + "content": "Like in GAMMA [60], these two networks are jointly trained by minimizing:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 412, + 233, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 412, + 233, + 441 + ], + "spans": [ + { + "bbox": [ + 99, + 412, + 233, + 441 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {P P O} + \\mathbb {E} \\left[ \\left(r _ {t} - V (s _ {t})\\right) ^ {2} \\right] \\\\ + \\alpha \\operatorname {K L} - \\operatorname {d i v} (\\pi (z | s) | | \\mathcal {N} (0, I)), \\\\ \\end{array}", + "image_path": "91a2c9e2e358be42d868a61fc86d877fad8eaeb1114a5ef85e2f52cdaf2dbe42.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 448, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 448, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 448, + 287, + 496 + ], + "type": "text", + "content": "where the first term is the PPO [39] loss, the second updates the value estimation of the critic networks, and the third Kullback-Leibler divergence term regularizes motion in the latent space [60]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 510, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 287, + 641 + ], + "type": "text", + "content": "Tree sampling for test-time optimization. Given the stochastic nature of our Gaussian policies, sampling motions from the generated action distributions can yield motion primitive results of various qualities. Therefore, we follow [60] to use tree-based sampling during inference to discard motion primitives with inferior goal-reaching and scene interaction scores. Specifically, we sample multiple latent actions at each time step and selectively keep the best K samples, utilizing the same rewards used to train the policies as the selection criteria. This tree-sampling technique yields improved synthesis results of higher quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "In the following, we elaborate on the design of states and rewards tailored to different actions, namely locomotion and fine-grained object interaction. By combining the learned policies, long-term and coherent motions can be composed by rolling out the initial body and switching between the locomotion and object interaction stages." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 542, + 194 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 542, + 194 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 542, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 542, + 194 + ], + "type": "image", + "image_path": "a373935eda6a2537d67285d94c779f0187ce05f72ee8bdced70a8d92d0c566eb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 201, + 545, + 296 + ], + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 296 + ], + "type": "text", + "content": "Figure 3: Illustration of the scene-aware locomotion policy network. The locomotion policy state consists of the body markers, the goal-reaching feature of normalized direction vectors from markers to the goal pelvis, and the interaction feature of a 2D binary map indicating the walkability (red: non-walkable area, blue: walkable area) of the local " + }, + { + "bbox": [ + 305, + 201, + 545, + 296 + ], + "type": "inline_equation", + "content": "1.6m \\times 1.6m" + }, + { + "bbox": [ + 305, + 201, + 545, + 296 + ], + "type": "text", + "content": " square area. The locomotion policy network employs the actor-critic architecture and shares the state encoder." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 314, + 494, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 314, + 494, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 314, + 494, + 327 + ], + "type": "text", + "content": "3.3. Scene-Aware Locomotion Synthesis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 332, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 332, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 545, + 439 + ], + "type": "text", + "content": "Navigating in cluttered scenes here means the human body moving to a target while avoiding collisions with scene objects. Our key idea is to incorporate the walkability information of the surrounding environment into the states and use collision rewards to train the locomotion policy in order to avoid scene collisions. Specifically, we represent the walkability of the environment surrounding the human agent using a 2D binary map " + }, + { + "bbox": [ + 305, + 332, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\{0,1\\}^{16 \\times 16}" + }, + { + "bbox": [ + 305, + 332, + 545, + 439 + ], + "type": "text", + "content": " as illustrated in Fig. 3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 440, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 440, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 305, + 440, + 545, + 523 + ], + "type": "text", + "content": "The walkability map is defined in the human's local coordinates and covers a " + }, + { + "bbox": [ + 305, + 440, + 545, + 523 + ], + "type": "inline_equation", + "content": "1.6m \\times 1.6m" + }, + { + "bbox": [ + 305, + 440, + 545, + 523 + ], + "type": "text", + "content": " area centered at the body pelvis and aligned with the body facing orientation. It consists of a " + }, + { + "bbox": [ + 305, + 440, + 545, + 523 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 305, + 440, + 545, + 523 + ], + "type": "text", + "content": " cell grid where each cell stores a binary value indicating whether this cell is walkable or not. This local walkability map enables the policy to sense surrounding obstacles." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 524, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 524, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 305, + 524, + 545, + 548 + ], + "type": "text", + "content": "Referring to Eq. 1, the person-scene interaction feature is specified by" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 397, + 548, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 548, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 397, + 548, + 545, + 559 + ], + "type": "interline_equation", + "content": "I = \\operatorname {v e c} (\\mathcal {M}), \\tag {5}", + "image_path": "e5c27472d82476d7cc350f676fd44c37f9a7885be8ca393bcd18ccabac8f6e1d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 563, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 545, + 587 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 305, + 563, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\mathrm{vec}(\\cdot)" + }, + { + "bbox": [ + 305, + 563, + 545, + 587 + ], + "type": "text", + "content": " denotes vectorization. The goal-reaching feature is specified by" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 388, + 594, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 594, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 388, + 594, + 545, + 605 + ], + "type": "interline_equation", + "content": "G = \\left(\\tilde {\\boldsymbol {g}} _ {p} - \\boldsymbol {X} _ {s}\\right) _ {n}, \\tag {6}", + "image_path": "87ea070b6f925d6b3f2d0e31ebcd9cf5ea09e9e7d17589810419f7047302a491.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_s" + }, + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{g}}_p \\in \\mathbb{R}^{M \\times 67 \\times 3}" + }, + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "type": "text", + "content": " are the body marker seed representing " + }, + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "type": "text", + "content": " frames history of motion, and the broadcasted target pelvis location relative to the body-centered canonical coordinate, respectively. " + }, + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "type": "inline_equation", + "content": "(\\tilde{\\pmb{g}}_p - \\mathbf{X}_s)_n" + }, + { + "bbox": [ + 305, + 611, + 545, + 682 + ], + "type": "text", + "content": " denotes the normalized vectors pointing from each marker to the goal pelvis." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 684, + 512, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 684, + 512, + 695 + ], + "spans": [ + { + "bbox": [ + 317, + 684, + 512, + 695 + ], + "type": "text", + "content": "The rewards contributing to Eq. 2 are defined as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 374, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 374, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "r _ {p e n e} = e ^ {- | \\mathcal {M} _ {0} \\cap \\mathcal {B} _ {x y} (X) |}, \\tag {7}", + "image_path": "a30d2e6a6a9839a8c99f13264ea958a57fce1cbe811db21700ecf1396f45cc15.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 741 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 741 + ], + "type": "text", + "content": "14742" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " denotes the non-walkable cells in the walkability map, " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{xy}(\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " denotes the 2D bounding box of the body markers " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\cap" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " denotes their intersection, and " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " denotes the number of non-walkable cells overlapping with the human bounding box." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 146, + 287, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 146, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 123, + 146, + 287, + 159 + ], + "type": "interline_equation", + "content": "r _ {\\text {g o a l}} = r _ {\\text {d i s t}} + r _ {\\text {o r i}}, \\tag {8}", + "image_path": "fe13d03d6cb98e98ace01d3d150a8d7b3ef4c345f08bd873e02ba2528815b25a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 96, + 177, + 287, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 177, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 96, + 177, + 287, + 192 + ], + "type": "interline_equation", + "content": "r _ {d i s t} = 1 - \\left(\\left\\| \\boldsymbol {p} - \\boldsymbol {g} _ {p} \\right\\| _ {2} - 0. 0 5\\right) _ {+}, \\tag {9}", + "image_path": "32625b65cebc137516692d06f76e77fe93c45316452d556ef9d063314b015b68.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 206, + 287, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 206, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 124, + 206, + 287, + 232 + ], + "type": "interline_equation", + "content": "r _ {o r i} = \\frac {\\left\\langle \\boldsymbol {o} , \\boldsymbol {g} _ {p} - \\boldsymbol {p} \\right\\rangle}{2}, \\tag {10}", + "image_path": "50c783e0992f01a462fb1ea68e5cd8ce1f9398ba8c31376a29ab7b2ab4603470.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "inline_equation", + "content": "r_{dist}" + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "content": " encourages the body pelvis " + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "content": " to be close to the pelvis goal " + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "inline_equation", + "content": "g_p" + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "inline_equation", + "content": "r_{ori}" + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "content": " encourages the body facing direction " + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "content": " to be aligned with the direction from the current body pelvis " + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "content": " to the pelvis goal " + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "inline_equation", + "content": "g_p" + }, + { + "bbox": [ + 47, + 236, + 287, + 284 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 292, + 266, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 266, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 266, + 304 + ], + "type": "text", + "content": "3.4. Fine-grained Object Interaction Synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 310, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 287, + 407 + ], + "type": "text", + "content": "To synthesize fine-grained human-object interactions, e.g. sitting on a chair or lying on a sofa, we use body marker goals as guidance, and model the proximity between the body surface and the scene object in a compact way. The goal marker sets can be generated by static person-scene interaction methods such as [14, 56, 61]. We use COINS [61] to generate the static goal interaction body for its performance and controllability of interaction semantics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 407, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 502 + ], + "type": "text", + "content": "In addition to the marker-based goal guidance, we incorporate the proximity relations between humans and objects into the states. Specifically, we use the signed distance from each marker to the surface of the scene object, as well as the gradient direction of the signed distance to represent the proximity relationship, as illustrated in Fig. 4. Both the signed distance and its gradient direction are calculated using the object's signed distance field (SDF) " + }, + { + "bbox": [ + 46, + 407, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\Psi_{O}" + }, + { + "bbox": [ + 46, + 407, + 287, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 503, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 287, + 525 + ], + "type": "text", + "content": "Referring to Eq. 1, the person-scene interaction feature is formulated as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 537, + 287, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 537, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 110, + 537, + 287, + 551 + ], + "type": "interline_equation", + "content": "I = \\left[ \\Psi_ {O} \\left(\\boldsymbol {X} _ {s}\\right), \\nabla \\Psi_ {O} \\left(\\boldsymbol {X} _ {s}\\right) \\right], \\tag {11}", + "image_path": "c4a800790ee67193964b6a8a814c3e0c178acb6cbacfaa7bdd35e94a396c6cc8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\Psi_O\\in \\mathbb{R}^{M\\times 67}" + }, + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\nabla \\Psi_{O}\\in \\mathbb{R}^{M\\times 201}" + }, + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "type": "text", + "content": " denote the SDF values and the gradient at each marker location in the " + }, + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "type": "text", + "content": " frames, respectively, and " + }, + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 47, + 559, + 287, + 608 + ], + "type": "text", + "content": " denotes feature concatenation. The goal-reaching feature is formulated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 95, + 618, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 618, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 95, + 618, + 287, + 632 + ], + "type": "interline_equation", + "content": "G = \\left[ \\left(\\tilde {\\boldsymbol {g}} _ {m} - \\boldsymbol {X} _ {s}\\right) _ {n}, \\| \\tilde {\\boldsymbol {g}} _ {m} - \\boldsymbol {X} _ {s} \\| _ {2} \\right], \\tag {12}", + "image_path": "1272d797622acb7c43529c2f910b572e5ea7033dc37a41e6b37fc6b933ee46ee.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_s \\in \\mathbb{R}^{M \\times 67 \\times 3}" + }, + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "type": "text", + "content": " denotes the body markers seed, " + }, + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{g}}_m \\in \\mathbb{R}^{M \\times 67 \\times 3}" + }, + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "type": "text", + "content": " denotes the broadcasted goal body markers, " + }, + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\tilde{\\pmb{g}}_m - \\pmb{X}_s)_n" + }, + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "type": "text", + "content": " denotes the normalized vector representing the direction from each marker to the corresponding target marker, " + }, + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\| \\tilde{\\pmb{g}}_m - X_s \\|_2" + }, + { + "bbox": [ + 47, + 640, + 287, + 713 + ], + "type": "text", + "content": " denotes the distance from each marker to target marker." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 69, + 542, + 185 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 542, + 185 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 542, + 185 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 542, + 185 + ], + "type": "image", + "image_path": "c4e160cfb625594ef9c430e44de2487f539b9e80123502c95f3b5e8007a25ed4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 193, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 305, + 193, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 193, + 545, + 289 + ], + "type": "text", + "content": "Figure 4: Illustration of the object interaction policy network. The interaction policy state consists of the body markers, the goal-reaching features of both distance and direction from current markers to the goal markers, and the interaction features of the signed distances from each marker to the object surfaces and the signed distance gradient at each marker location. Such interaction features encode the human-object proximity relationship." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 310, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 545, + 346 + ], + "type": "text", + "content": "The interaction policy is trained using the reward defined in Eq. 2 with the following interaction-specific goal reward and penetration reward:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 357, + 545, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 357, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 351, + 357, + 545, + 371 + ], + "type": "interline_equation", + "content": "r _ {\\text {g o a l}} = 1 - \\left(\\| \\boldsymbol {x} - \\boldsymbol {g} _ {m} \\| _ {2} - 0. 0 5\\right) _ {+}, \\tag {13}", + "image_path": "1cdb838855ef548064b2e2dd6b6a3f4a6c4855bb68ce68a4879fc944fa6f84c7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 344, + 384, + 545, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 384, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 344, + 384, + 545, + 402 + ], + "type": "interline_equation", + "content": "r _ {p e n e} = e ^ {- \\frac {1}{| V |} \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {| V |} \\left| \\left(\\Psi_ {O} (\\boldsymbol {v} _ {t i})\\right) _ {-} \\right|}, \\tag {14}", + "image_path": "4dc0c62b44b6b54d6a7c7e0feb8bd147f591b7c2f1f1a4bc63ea85496bcfbc7b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "text", + "content": " being the SMPL-X mesh vertices, " + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "text", + "content": " denotes the number of frames in each motion primitive (equals to 10 in our study). The distance reward encourages the final frame body markers " + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "text", + "content": " to be close to the goal body markers " + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "inline_equation", + "content": "g_{m}" + }, + { + "bbox": [ + 304, + 407, + 545, + 514 + ], + "type": "text", + "content": ". The penetration reward penalizes all the body vertices within a motion primitive that have negative SDF values. We use body vertices instead of joints because human-object contact happens on the body surface and can be better detected using vertices." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 515, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 586 + ], + "type": "text", + "content": "Moreover, we train the interaction policy with a mixture of 'sit/lie down' and 'stand up' tasks. This training scheme enables the human agent to also learn how to stand up and transit from object interaction back to locomotion, which enables the synthesis of a sequence of interaction activities as in Fig. 1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 597, + 382, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 382, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 382, + 611 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Motion Datasets. We combine the large-scale motion capture dataset AMASS [29] with SAMP [12] motion data to train the motion primitive models. Each sequence is first subsampled to 40 FPS and then split into 10 frames and 100 frames clips. Each motion clip is canonicalized using the first frame body. Specifically, we select AMASS sequences annotated with 'sit' or 'lie' in BABEL [37] and all motion data of SAMP to train the motion primitive model. We train" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "14743" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 115, + 281, + 158 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 106 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 106 + ], + "type": "text", + "content": "Table 1: Evaluation of the locomotion task. The up/down arrows denote the score is the higher/lower the better and the best results are in boldface." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 115, + 281, + 158 + ], + "lines": [ + { + "bbox": [ + 50, + 115, + 281, + 158 + ], + "spans": [ + { + "bbox": [ + 50, + 115, + 281, + 158 + ], + "type": "table", + "html": "
time ↓avg. dist ↓contact ↑loco pene ↑
SAMP [12]5.970.140.840.94
GAMMA [60]3.870.030.940.94
Ours6.430.040.990.95
", + "image_path": "66d2e9bfaef14894623e45c6aa22439f0de79e7fed4b183f916e198f4c0ed074.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 180, + 287, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 228 + ], + "type": "text", + "content": "separate motion primitive models for locomotion and interactions using task-related data. We observe that extending SAMP motion data with AMASS dataset is the key to making interaction policies work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 419 + ], + "type": "text", + "content": "Policy Training Environments. To train the scene-aware locomotion policy, we randomly generate synthetic cluttered scenes consisting of random objects from ShapeNet [6]. Random initial and target location pairs are sampled in the walkable areas using navigation meshes to train the locomotion policy. To train the interaction policy, we use the static person-object interaction data from PROX [13] and retargeted to ShapeNet objects. We first randomly sample a furniture and interaction body goal from the retargeted PROX data. Then we sample the initial body with random poses and locations in front of the object to train the interaction policy. We also randomly swap the initial and goal body to learn both 'sit/lie down' and 'stand up' motions. Please refer to Supp. Mat. for more details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 430, + 188, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 188, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 188, + 443 + ], + "type": "text", + "content": "4.1. Locomotion in 3D Scenes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 451, + 287, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 500 + ], + "type": "text", + "content": "We randomly generate test scenes for locomotion in the same way as the training scenes. The virtual human is instructed to move from the random start point to the random target point while avoiding penetration with scene objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "Baselines and metrics. We compare our method with SAMP [12] and GAMMA [60] for locomotion. The SAMP results are recorded by running the released Unity demo. The start and termination are manually determined so the reported completion time may be slightly higher than the actual time due to human response time. The evaluation metrics for locomotion include: 1) time from start point to target point or reaching the time limit, measured in seconds. 2) the average distance from the final body to the targets, measured in meters. 3) foot contact score encouraging the lowest feet joints on the floor and discouraging foot skating as defined in Eq. 3. We use body joints instead of markers to calculate the contact score because the marker set annotation for the SAMP body is missing. 4) locomotion penetration score indicating the percentage of body vertices that are inside the walkable areas according to the navigation mesh." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 425, + 160 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 425, + 160 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 425, + 160 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 425, + 160 + ], + "type": "image", + "image_path": "23cac8bda4be6cdde7c954ad3f3a9963844c75e07d4db5031a16b8ee855b88bd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 160, + 425, + 250 + ], + "blocks": [ + { + "bbox": [ + 318, + 160, + 425, + 250 + ], + "lines": [ + { + "bbox": [ + 318, + 160, + 425, + 250 + ], + "spans": [ + { + "bbox": [ + 318, + 160, + 425, + 250 + ], + "type": "image", + "image_path": "ba5f7e324de279aab27e3347ff8ac17cb880930c85e80c848216f4baf9d18385.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 255, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 305, + 255, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 547, + 304 + ], + "type": "text", + "content": "Figure 5: Demonstration of locomotion tasks where GAMMA [60] (left) collides with the obstacles (red bodies) while our scene-aware locomotion policy (right) avoids collision. The yellow circles denote the specified waypoints." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 426, + 70, + 534, + 160 + ], + "blocks": [ + { + "bbox": [ + 426, + 70, + 534, + 160 + ], + "lines": [ + { + "bbox": [ + 426, + 70, + 534, + 160 + ], + "spans": [ + { + "bbox": [ + 426, + 70, + 534, + 160 + ], + "type": "image", + "image_path": "c6c56c14fc6383ab105da0b9de95c6944b39d916d8e346c64581cff149f52559.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 160, + 534, + 250 + ], + "blocks": [ + { + "bbox": [ + 426, + 160, + 534, + 250 + ], + "lines": [ + { + "bbox": [ + 426, + 160, + 534, + 250 + ], + "spans": [ + { + "bbox": [ + 426, + 160, + 534, + 250 + ], + "type": "image", + "image_path": "f64c7f963b62d65641324cc8ba336fdf83781c957d24b4af6b52440f73972df4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 324, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 547, + 504 + ], + "type": "text", + "content": "Results. Table 1 shows the empirical evaluation results. Our method achieves a higher contact score (0.99) than both GAMMA (0.94) and SAMP (0.84) which indicates better foot-floor contact and less foot skating. Moreover, our method achieves the highest penetration score which indicates our scene-aware policy can better avoid scene collisions. Fig. 5 shows examples of locomotion tasks where GAMMA collides into the scenes while our scene-aware policy successfully avoids penetration. Compared to GAMMA trained in the same environments, we observe our scene-aware policy learns more conservative behavior with lower moving speed and spent more time (6.43s) walking to the target locations, just like a human afraid of stepping on surrounding traps. All the methods can reach the target location within a reasonable distance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 511, + 521, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 521, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 521, + 525 + ], + "type": "text", + "content": "4.2. Fine-Grained Human-Object Interaction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 530, + 547, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 547, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 547, + 627 + ], + "type": "text", + "content": "We evaluate the object interaction task on 10 unseen objects (3 armchairs, 3 straight chairs, 3 sofas, 1 L-sofa) from ShapeNet [6]. We use the object size annotation of ShapeNet and manually filter unrealistic-sized objects. The virtual human is randomly placed in front of the target object and then instructed to perform the interaction, stay for around 2 seconds, and then stand up. We evaluate two interactions of sitting and lying separately." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Baselines and metrics. We compare our method with SAMP [12] for the object interaction task. The evaluation metrics for the interaction tasks are: 1) time of completing the object interaction task. 2) foot contact score as defined in Eq. 3. Note that the foot contact score does not always reflect the motion quality for lying tasks because the foot can" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "14744" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 115, + 286, + 166 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 106 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 106 + ], + "type": "text", + "content": "Table 2: Evaluation of the interaction tasks. The up/down arrows denote the score is the higher/ lower the better and the best results are in boldface." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 115, + 286, + 166 + ], + "lines": [ + { + "bbox": [ + 50, + 115, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 50, + 115, + 286, + 166 + ], + "type": "table", + "html": "
time ↓contact ↑pene. mean ↓pene. max ↓
SAMP [12] sit8.630.8911.9145.22
Ours sit4.090.971.9110.61
SAMP [12] lie12.550.7344.77238.81
Ours lie4.200.789.9044.61
", + "image_path": "4b0570ed2641973cc4c8c8273edae0d8b2bbd80d3e7a81a80fc28ef324564287.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 178, + 281, + 351 + ], + "blocks": [ + { + "bbox": [ + 52, + 178, + 281, + 351 + ], + "lines": [ + { + "bbox": [ + 52, + 178, + 281, + 351 + ], + "spans": [ + { + "bbox": [ + 52, + 178, + 281, + 351 + ], + "type": "image", + "image_path": "374b540cd5b261e574e784cec96c564c94f1b956630047236833d26581418edb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 360, + 287, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 407 + ], + "type": "text", + "content": "Figure 6: Demonstration of various generated object interactions. Each row shows generated interactions with the same object starting from random initial body locations and orientations. Colors from yellow to red denote time." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 453 + ], + "type": "text", + "content": "often be off the floor during lying. 3) interaction penetration score for each frame is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 463, + 287, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 463, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 100, + 463, + 287, + 489 + ], + "type": "interline_equation", + "content": "s _ {i n t e r - p e n e} = \\sum_ {v _ {i} \\in V} | (\\Psi_ {O} (v _ {i})) _ {-} |, \\tag {15}", + "image_path": "6b1c7dcd326f29085f709f92bf907fbe6a251237707450b4cbaa02b2ef7112af.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 495, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 495, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\Psi_{O}" + }, + { + "bbox": [ + 46, + 495, + 287, + 544 + ], + "type": "text", + "content": " is the object signed distance field, " + }, + { + "bbox": [ + 46, + 495, + 287, + 544 + ], + "type": "inline_equation", + "content": "(\\cdot)_{-}" + }, + { + "bbox": [ + 46, + 495, + 287, + 544 + ], + "type": "text", + "content": " clips all positive distance values to zero, and " + }, + { + "bbox": [ + 46, + 495, + 287, + 544 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 495, + 287, + 544 + ], + "type": "text", + "content": " is the body vertivces. We show both the average penetration over time and the maximum penetration in one sequence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "Results. Tab. 2 shows the evaluation results. Our method achieves a significantly higher foot contact score, indicating more natural motion. Our method also achieves lower mean and maximum penetration, which means our method generates more physically plausible results. Moreover, our method can complete the interaction tasks much faster than SAMP. This is because SAMP does not generalize well to unseen random body initialization and needs a longer time to start performing interactions. Qualitative results demonstrating the various object interactions generated by our method are shown in Fig. 6. Our object interaction policy generalizes to random initial body locations and orientations, as well as novel objects of various shapes." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 427, + 171 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 427, + 171 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 427, + 171 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 427, + 171 + ], + "type": "image", + "image_path": "2214b1eaecfa4674311e718587f630b48b5b01d0e7129fdcfa01d93f12ba717e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 180, + 545, + 241 + ], + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 241 + ], + "type": "text", + "content": "Figure 7: Our method can synthesize natural human movements in scene reconstructions that are significantly different from the training scenes. We show the synthesis results in one Paris street [7] (left figure) and one indoor room [13] (right figure). Transparent to solid colors denote time." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 543, + 171 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 543, + 171 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 543, + 171 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 543, + 171 + ], + "type": "image", + "image_path": "a28c54f5987d85f635beb86f2f0649b3d61b9e6f3c7579ab603cdaacdf649914.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 341, + 349, + 509, + 429 + ], + "blocks": [ + { + "bbox": [ + 304, + 256, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 340 + ], + "type": "text", + "content": "Table 3: Evaluation of interaction sequences synthesis. Here 'contact' denotes the foot contact score, 'loco. pene.' is the percentage of body vertices inside the walkable areas, 'inter. pene. mean/max' denotes the mean and maximum penetration with interaction objects, and 'perceptual' denotes the ratio of being chosen as perceptually more natural. The best results are in boldface." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 341, + 349, + 509, + 429 + ], + "lines": [ + { + "bbox": [ + 341, + 349, + 509, + 429 + ], + "spans": [ + { + "bbox": [ + 341, + 349, + 509, + 429 + ], + "type": "table", + "html": "
SAMP [12]Ours
contact ↑0.870.96
loco. pene. ↑0.620.72
inter. pene. mean↓15.613.40
inter. pene. max ↓101.2539.68
perceptual. ↑0.150.85
", + "image_path": "aaf976264e7d4a1cc707adccbd805c716ddb753692ae2fac988aad8ad6cb1661.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 452, + 493, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 493, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 493, + 464 + ], + "type": "text", + "content": "4.3. Interaction Sequences in 3D Scenes" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 472, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 663 + ], + "type": "text", + "content": "Humans often continuously perform sequences of interactions with various objects in complex real-world scenes as shown in Fig. 1. Such interaction sequences are combinations of alternating locomotions and fine-grained interactions which we evaluate in Sec. 4.1 and Sec. 4.2 respectively. We conduct the empirical evaluation in real scene scans from Replica [44] and compare our method with SAMP. We select a list of interactable objects in the scene and instruct the virtual human to walk to the objects to perform interactions one by one. The evaluation metrics include the foot contact score, locomotion penetration score, and the mean and maximum interaction penetration score. We also conducted a perceptual study where participants are shown a side-by-side comparison of results from two methods and asked to choose the one perceptually more natural. We report the rate of being chosen as the better." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "The evaluation results are shown in Tab. 3. Our method generates high-quality results of human-scene interactions in cluttered environments. Our results achieve higher contact scores and less penetration with the scenes compared" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "14745" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 115, + 285, + 158 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 107 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 107 + ], + "type": "text", + "content": "Table 4: Rewards ablation studies results, where '-Contact' and '-Penetration' denote policies trained without the floor contact and penetration avoidance reward, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 115, + 285, + 158 + ], + "lines": [ + { + "bbox": [ + 48, + 115, + 285, + 158 + ], + "spans": [ + { + "bbox": [ + 48, + 115, + 285, + 158 + ], + "type": "table", + "html": "
time ↓contact ↑pene. mean ↓pene. max ↓
Ours4.090.971.9110.61
- Contact3.750.8819.4960.10
- Penetration4.030.9516.8045.50
", + "image_path": "ad54229f9a833f2876b4e05fe0eee95fa873d2ebbf1a944755d71543e0434989.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 177, + 286, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 286, + 202 + ], + "type": "text", + "content": "to SAMP. In addition, our method generates perceptually more natural results according to perceptual study." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 209, + 148, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 148, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 148, + 221 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 228, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 407 + ], + "type": "text", + "content": "We perform ablation studies on the foot-ground contact reward and penetration avoidance reward of the interaction policies to substantiate the significance of these rewards. We train two ablation interaction policies using identical networks and environments, differing only in the exclusion of one of these rewards in each case. The quantitative metrics are reported in Tab. 4. The removal of the penetration avoidance reward yields a marked escalation in detected human-scene penetration. The removal of the footfloor contact reward yields notably both inferior contact and penetration scores, accompanied by observed erratic synthesized motions that either remain suspended in the air or penetrate the floor. These empirical findings underscore the crucial role played by the foot-floor contact and penetration avoidance rewards in the learning of interaction policies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 418, + 198, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 198, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 198, + 430 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 714 + ], + "type": "text", + "content": "Limitations and Future Work. Our current method has various limitations that could be improved in future works. First, our method does not fully resolve penetrations with scene objects and floors since our method only uses rewards to encourage avoiding penetration with scenes, which does not impose hard constraints for penetrations. The combination of our method with physics simulation holds the potential to effectively address and resolve penetration issues. Moreover, the lying motions generated by our method are not as natural as the sitting motions because the motion primitive model fails to learn a comprehensive action space for lying given limited training data. Specifically, the available AMASS motion capture data for lying (167 seconds) is significantly less than the data for sitting (5K seconds). To overcome this issue, we aim to explore more data-efficient learning methods and scalable methods to collect human-scene interaction data. Furthermore, our locomotion policy is now limited to flat-floor scenes due to its reliance on the 2D occupancy-based walkability map. However, to broaden the applicability of our approach to more complex scenes such as uneven outdoor terrains and multi-floor buildings requiring walking upstairs, it will be necessary to replace the walkability map with a more suitable environment sens" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 420, + 162 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 420, + 162 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 420, + 162 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 420, + 162 + ], + "type": "image", + "image_path": "1b3ba8a7622d1b69fa19703f59b2f29b84b6eb9eea06cc7768ef58a2f8e989b0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 166, + 399, + 177 + ], + "lines": [ + { + "bbox": [ + 326, + 166, + 399, + 177 + ], + "spans": [ + { + "bbox": [ + 326, + 166, + 399, + 177 + ], + "type": "text", + "content": "Inter-penetrations" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 188, + 545, + 260 + ], + "lines": [ + { + "bbox": [ + 305, + 188, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 545, + 260 + ], + "type": "text", + "content": "Figure 8: Limitations. Penetrations remain observable in our results (left) since our method only encourages avoiding collision with rewards. Furthermore, the absence of sufficient training data for lying motions impedes the motion primitive model's ability to acquire a comprehensive lying motion space, leading to degraded motion (right)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 421, + 72, + 544, + 161 + ], + "blocks": [ + { + "bbox": [ + 421, + 72, + 544, + 161 + ], + "lines": [ + { + "bbox": [ + 421, + 72, + 544, + 161 + ], + "spans": [ + { + "bbox": [ + 421, + 72, + 544, + 161 + ], + "type": "image", + "image_path": "4e843826538511eb4c3b5a33b88fcca3d7d00b20dccde7f5f1bf36a8d6e79063.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 166, + 519, + 177 + ], + "lines": [ + { + "bbox": [ + 453, + 166, + 519, + 177 + ], + "spans": [ + { + "bbox": [ + 453, + 166, + 519, + 177 + ], + "type": "text", + "content": "Unnatural poses" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "type": "text", + "content": "ing mechanism. In addition, our method is restricted to interactions with static scenes. However, in the real world, humans are exposed to dynamic interactions with movable objects and scenes involving autonomous agents including other humans, animals, and vehicles. Extension to dynamic interactions will enable tackling a broader range of intricate and dynamic applications that mirror the complexities inherent in the real world." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 391, + 546, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 546, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 546, + 547 + ], + "type": "text", + "content": "Conclusion. In this paper, we leverage reinforcement learning to establish a framework to synthesize diverse human motions in indoor scenes, which is stochastic, realistic, and perpetual. The proposed method has large potential to improve many applications such as daily-living activity simulation, synthetic data creation, architecture design, and so on. Compared to existing methods, our method realizes fine-grained control by using body surface keypoints as targets, and achieves autonomous body-scene collision avoidance by incorporating scene information into the states and the rewards. Experiments show that our method effectively enables virtual humans to inhabit the virtual, and outperforms baselines consistently." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 561, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 545, + 622 + ], + "type": "text", + "content": "Acknowledgements. We sincerely acknowledge the anonymous reviewers for their insightful suggestions. This work was supported by the SNSF project grant 200021 204840 and SDSC PhD fellowship. Lingchen Yang provided valuable suggestions on visualization." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "14746" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. ISBN: 0730-0301 Publisher: ACM New York, NY, USA. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 145, + 288, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 145, + 288, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 145, + 288, + 177 + ], + "type": "text", + "content": "[2] Tenglong Ao, Zeyi Zhang, and Libin Liu. GestureDiffuCLIP: Gesture diffusion model with CLIP latents. arXiv preprint arXiv:2303.14613, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 177, + 288, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 177, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 53, + 177, + 288, + 231 + ], + "type": "text", + "content": "[3] Kevin Bergamin, Simon Clavet, Daniel Holden, and James Richard Forbes. DReCon: data-driven responsive control of physics-based characters. ACM Transactions On Graphics (TOG), 38(6):1-11, 2019. Publisher: ACM New York, NY, USA. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 231, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 288, + 252 + ], + "type": "text", + "content": "[4] Michael Buttner. Machine learning for motion synthesis and character control in games. Proc. of I3D 2019, 2, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 252, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 252, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 252, + 288, + 283 + ], + "type": "text", + "content": "[5] Michael Böttner and Simon Clavet. Motion matching-the road to next gen animation. Proc. of Nucl. ai, 2015(1):2, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 337 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, and Hao Su. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 337, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 288, + 392 + ], + "type": "text", + "content": "[7] Jean-Emmanuel Deschaud, David Duque, Jean Pierre Richa, Santiago Velasco-Forero, Beatrix Marcotegui, and François Goulette. Paris-carla-3d: A real and synthetic outdoor point cloud dataset for challenging tasks in 3d mapping. Remote Sensing, 13(22), 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 413 + ], + "type": "text", + "content": "[8] Helmut Grabner, Juergen Gall, and Luc Van Gool. What makes a chair a chair? pages 1529-1536. IEEE, 2011. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 443 + ], + "type": "text", + "content": "[9] Abhinav Gupta, Scott Satkin, Alexei A Efros, and Martial Hebert. From 3d scene geometry to human workspace. pages 1961-1968. IEEE, 2011. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 444, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 476 + ], + "type": "text", + "content": "[10] Ikhsanul Habibie, Daniel Holden, Jonathan Schwarz, Joe Yearsley, and Taku Komura. A recurrent variational autoencoder for human motion synthesis. 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 476, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 288, + 518 + ], + "type": "text", + "content": "[11] Peter E Hart, Nils J Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics, 4(2):100-107, 1968. Publisher: IEEE. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 519, + 288, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 288, + 562 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 288, + 562 + ], + "type": "text", + "content": "[12] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. pages 11374-11384, 2021. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 562, + 288, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 562, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 48, + 562, + 288, + 604 + ], + "type": "text", + "content": "[13] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3D human pose ambiguities with 3D scene constraints. pages 2282-2292, 2019. 4, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 604, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 288, + 647 + ], + "type": "text", + "content": "[14] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3D scenes by learning human-scene interaction. pages 14708-14718, 2021. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "text", + "content": "[15] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing Physical Character-Scene Interactions, Feb. 2023. arXiv:2302.00883 [cs]. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[16] Daniel Holden, Ouussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 705 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Graphics (TOG), 39(4):53-1, 2020. Publisher: ACM New York, NY, USA. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 95, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 137 + ], + "type": "text", + "content": "[17] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. Publisher: ACM New York, NY, USA. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 137, + 545, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 137, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 545, + 180 + ], + "type": "text", + "content": "[18] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. Publisher: ACM New York, NY, USA. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 180, + 545, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 180, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 307, + 180, + 545, + 223 + ], + "type": "text", + "content": "[19] Ruizhen Hu, Zihao Yan, Jingwen Zhang, Oliver Van Kaick, Ariel Shamir, Hao Zhang, and Hui Huang. Predictive and generative neural networks for object functionality. arXiv preprint arXiv:2006.15520, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "type": "text", + "content": "[20] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. PADI: Language-Directed Physics-Based Character Control. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 266, + 545, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 299 + ], + "type": "text", + "content": "[21] Kacper Kania, Marek Kowalski, and Tomasz Trzciński. TrajeVAE: Controllable Human Motion Generation from Trajectories. arXiv preprint arXiv:2104.00351, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 299, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 329 + ], + "type": "text", + "content": "[22] Lucas Kovar, Michael Gleicher, and Frédéric Pighin. Motion graphs. In ACM SIGGRAPH 2008 classes, pages 1-10. 2008. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 329, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 329, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 329, + 545, + 384 + ], + "type": "text", + "content": "[23] Peizhuo Li, Kfir Aberman, Zihan Zhang, Rana Hanocka, and Olga Sorkine-Hornung. Ganimator: Neural motion synthesis from a single sequence. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. Publisher: ACM New York, NY, USA. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 384, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 426 + ], + "type": "text", + "content": "[24] Weiyu Li, Xuelin Chen, Peizhuo Li, Olga Sorkine-Hornung, and Baoquan Chen. Example-based Motion Synthesis via Generative Motion Matching. arXiv preprint arXiv:2306.00378, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "text", + "content": "[25] Xueting Li, Sifei Liu, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. Putting humans in a scene: Learning affordance in 3d indoor environments. pages 12368-12376, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 469, + 545, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 469, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 307, + 469, + 545, + 512 + ], + "type": "text", + "content": "[26] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion vaes. ACM Transactions on Graphics (TOG), 39(4):40-1, 2020. Publisher: ACM New York, NY, USA. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "text", + "content": "[27] Libin Liu, KangKang Yin, and Baining Guo. Improving sampling-based motion control. volume 34, pages 415-423. Wiley Online Library, 2015. Issue: 2. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "text", + "content": "[28] Libin Liu, KangKang Yin, Michiel Van de Panne, Tianjia Shao, and Weiwei Xu. Sampling-based contact-rich motion control. In ACM SIGGRAPH 2010 papers, pages 1-10. 2010. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 586, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 586, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 307, + 586, + 545, + 630 + ], + "type": "text", + "content": "[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. AMASS: Archive of motion capture as surface shapes. pages 5442-5451, 2019. 2, 4, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 630, + 545, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 630, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 307, + 630, + 545, + 673 + ], + "type": "text", + "content": "[30] Josh Merel, Leonard Hasenclever, Alexandre Galashov, Arun Ahuja, Vu Pham, Greg Wayne, Yee Whye Teh, and Nicolas Heess. Neural probabilistic motor primitives for humanoid control. arXiv preprint arXiv:1811.11711, 2018. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 673, + 545, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 673, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 307, + 673, + 545, + 705 + ], + "type": "text", + "content": "[31] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face," + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "14747" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 704 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 93 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 93 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 93 + ], + "type": "text", + "content": "and body from a single image. pages 10975-10985, 2019. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 148 + ], + "type": "text", + "content": "[32] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. Publisher: ACM New York, NY, USA. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 148, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 148, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 148, + 287, + 201 + ], + "type": "text", + "content": "[33] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1-17, 2022. Publisher: ACM New York, NY, USA. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 202, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 202, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 202, + 287, + 255 + ], + "type": "text", + "content": "[34] Xue Bin Peng, Angjoo Kanazawa, Jitendra Malik, Pieter Abbeel, and Sergey Levine. Sfv: Reinforcement learning of physical skills from videos. ACM Transactions On Graphics (TOG), 37(6):1-14, 2018. Publisher: ACM New York, NY, USA. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 255, + 287, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 287 + ], + "type": "text", + "content": "[35] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. pages 10985-10995, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 286, + 319 + ], + "type": "text", + "content": "[36] Mathis Petrovich, Michael J Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. pages 480-497. Springer, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 363 + ], + "type": "text", + "content": "[37] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. BABEL: bodies, action and behavior with English labels. pages 722-731, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 363, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 287, + 415 + ], + "type": "text", + "content": "[38] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. SceneGrok: Inferring action maps in 3D environments. ACM transactions on graphics (TOG), 33(6):1-10, 2014. Publisher: ACM New York, NY, USA. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 448 + ], + "type": "text", + "content": "[39] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 490 + ], + "type": "text", + "content": "[40] Greg Snook. Simplified 3D movement and pathfinding using navigation meshes. Game programming gems, 1(1):288-304, 2000. Publisher: Charles River Media Newton Centre, MA, USA. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 533 + ], + "type": "text", + "content": "[41] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. Publisher: ACM New York, NY, USA. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 565 + ], + "type": "text", + "content": "[42] Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character-scene interactions. ACM Trans. Graph., 38(6):209-1, 2019. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 566, + 287, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 608 + ], + "type": "text", + "content": "[43] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39(4):54-1, 2020. Publisher: ACM New York, NY, USA. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 609, + 287, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 287, + 661 + ], + "type": "text", + "content": "[44] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, and Shobhit Verma. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 662, + 287, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 662, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 48, + 662, + 287, + 693 + ], + "type": "text", + "content": "[45] Richard S Sutton and Andrew G Barto. Introduction to reinforcement learning, volume 135. MIT press Cambridge, 1998. 2, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 694, + 287, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 694, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 694, + 287, + 704 + ], + "type": "text", + "content": "[46] Xiangjun Tang, He Wang, Bo Hu, Xu Gong, Ruifan Yi, Qi" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 704 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 115 + ], + "type": "text", + "content": "long Kou, and Xiaogang Jin. Real-time controllable motion transition for characters. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. Publisher: ACM New York, NY, USA. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 116, + 545, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 116, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 545, + 168 + ], + "type": "text", + "content": "[47] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. CALM: Conditional Adversarial Latent Models for Directable Virtual Characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 170, + 545, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 170, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 307, + 170, + 545, + 202 + ], + "type": "text", + "content": "[48] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. pages 358–374. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 202, + 545, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 202, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 307, + 202, + 545, + 234 + ], + "type": "text", + "content": "[49] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 234, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 266 + ], + "type": "text", + "content": "[50] Jonathan Tseng, Rodrigo Castellon, and C Karen Liu. EDGE: Editable Dance Generation From Music. arXiv preprint arXiv:2211.10658, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 266, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 297 + ], + "type": "text", + "content": "[51] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. pages 20460-20469, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 297, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 329 + ], + "type": "text", + "content": "[52] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. pages 9401-9411, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 329, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 329, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 329, + 545, + 382 + ], + "type": "text", + "content": "[53] Zhiyong Wang, Jinxiang Chai, and Shihong Xia. Combining recurrent neural networks and adversarial training for human motion synthesis and control. IEEE transactions on visualization and computer graphics, 27(1):14-28, 2019. Publisher: IEEE. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "text", + "content": "[54] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. PhysDiff: Physics-Guided Human Motion Diffusion Model. arXiv preprint arXiv:2212.02500, 2022.3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 415, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 479 + ], + "type": "text", + "content": "[55] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning Physically Simulated Tennis Skills from Broadcast Videos. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. ISBN: 0730-0301 Publisher: ACM New York, NY, USA. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "text", + "content": "[56] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. PLACE: Proximity learning of articulation and contact in 3D environments. pages 642-651. IEEE, 2020. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "text", + "content": "[57] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: towards controllable human-chair interactions. pages 518-535. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 566, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 545, + 597 + ], + "type": "text", + "content": "[58] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. pages 3372-3382, 2021. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 597, + 545, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 597, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 307, + 597, + 545, + 629 + ], + "type": "text", + "content": "[59] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes without people. pages 6194-6204, 2020. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 629, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 629, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 307, + 629, + 545, + 651 + ], + "type": "text", + "content": "[60] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3D scenes. pages 20481-20491, 2022. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 651, + 545, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 651, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 307, + 651, + 545, + 692 + ], + "type": "text", + "content": "[61] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. pages 311-327. Springer, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 692, + 545, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 704 + ], + "type": "text", + "content": "[62] Yang Zheng, Yanchao Yang, Kaichun Mo, Jiaman Li, Tao" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "14748" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "Yu, Yebin Liu, C Karen Liu, and Leonidas J Guibas. Gimo: Gaze-informed human motion prediction in context. pages 676-694. Springer, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 137 + ], + "type": "text", + "content": "[63] Fabio Zinno. MI tutorial day: From motion matching to motion synthesis, and all the hurdles in between. Proc. of GDC 2019, 2, 2019. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "14749" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/b759f55b-503f-4164-9793-b1ffe02056fb_content_list.json b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/b759f55b-503f-4164-9793-b1ffe02056fb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b9fb693b5a640af5cacb36f378092ec97509a28 --- /dev/null +++ b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/b759f55b-503f-4164-9793-b1ffe02056fb_content_list.json @@ -0,0 +1,1623 @@ +[ + { + "type": "text", + "text": "TALL:,Thumbsail Layout for Deepfake Video Detection", + "text_level": 1, + "bbox": [ + 200, + 130, + 772, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuting Xu $^{1,3*}$ , Jian Liang $^{2,4}$ , Gengyun Jia $^{5}$ , Ziming Yang $^{1,3}$ , Yanhao Zhang $^{6}$ , Ran He $^{2,4\\dagger}$", + "bbox": [ + 122, + 178, + 854, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences", + "bbox": [ + 210, + 198, + 759, + 215 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{2}$ CRIPAC & MAIS, Institute of Automation, Chinese Academy of Sciences", + "$^{3}$ School of Cyber Security, UCAS $^{4}$ School of Artificial Intelligence, UCAS" + ], + "bbox": [ + 184, + 215, + 785, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ School of Communications and Information Engineering, NJUPT $^{6}$ OPPO Research Institute", + "bbox": [ + 112, + 251, + 856, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yuting.xu@cripac.ia.ac.cn, liangjian92@gmail.com, rhe@nlpr.ia.ac.cn", + "bbox": [ + 186, + 272, + 781, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The growing threats of deepfakes to society and cybersecurity have raised enormous public concerns, and increasing efforts have been devoted to this critical topic of deepfake video detection. Existing video methods achieve good performance but are computationally intensive. This paper introduces a simple yet effective strategy named Thumbnail Layout (TALL), which transforms a video clip into a pre-defined layout to realize the preservation of spatial and temporal dependencies. Specifically, consecutive frames are masked in a fixed position in each frame to improve generalization, then resized to sub-images and rearranged into a pre-defined layout as the thumbnail. TALL is model-agnostic and extremely simple by only modifying a few lines of code. Inspired by the success of vision transformers, we incorporate TALL into Swin Transformer, forming an efficient and effective method TALL-Swin. Extensive experiments on intra-dataset and cross-dataset validate the validity and superiority of TALL and SOTA TALL-Swin. TALL-Swin achieves $90.79\\%$ AUC on the challenging cross-dataset task, FaceForensics++ $\\rightarrow$ CelebDF. The code is available at https://github.com/rainy-xu/TALL4Deepfake.", + "bbox": [ + 75, + 354, + 473, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deepfakes generate and manipulate facial appearances to deceive viewers through generation techniques [61, 48]. With the remarkable success of generative adversarial networks [14, 27], deepfake products have become photorealistic that humans can not distinguish. These deepfake products [21, 49] may be misused for malicious purposes, leading to severe trust issues and security problems, such as financial fraud, identity theft, and celebrity impersonation.", + "bbox": [ + 75, + 742, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2648b852e6f75e080588e0ff8d14afab55c72aee5ab1b02647d8b2fd55fcd11.jpg", + "image_caption": [ + "Figure 1. The AUC and FLOPs trade-off of different backbones. Image-level backbones with TALL enjoy comparable accuracy-cost trade-offs with the 3DCNN and video transformer family on the unseen Celeb-DF dataset. All models with the same setting are trained on the FF++ (HQ) dataset." + ], + "image_footnote": [], + "bbox": [ + 511, + 324, + 882, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ation [53, 40]. The rapid development of social media accelerates the abuse of deepfakes. Therefore, it is crucial to develop advanced detection methods to protect the data privacy of individual users.", + "bbox": [ + 496, + 643, + 893, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most previous image-based methods [22, 69] perform well on intra-dataset, but their generalizability needs to be improved. Recent research has focused on video-based methods to detect deepfake by modeling spatio-temporal dependencies. There are subtle spatio-temporal inconsistencies between frames since the deepfake algorithms are executed frame by frame. The core of video-level approaches for deepfake detection is capturing inconsistencies through temporal modeling. Existing deepfake video detection methods generally follow two directions. Some methods [15, 16] use two-branch networks or modules to learn spatial and temporal information separately and then fuse them. However, these two-branch approaches may frag", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when she was a student in CRIPAC.", + "bbox": [ + 94, + 875, + 401, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22658", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ment spatiotemporal cooperation and lead to subtle artifacts being neglected. Others directly use classic temporal models such as LSTM and 3D-CNNs. These methods are computationally intensive. The current rise of transformers for vision task backbones has prompted the emergence of corresponding deepfake detection methods. They are accompanied by significant computational complexity that makes them challenging to deploy and use, despite breakthroughs in performance. To enjoy benefits from both image and video methods, we are curious to see whether it is possible to append information about the temporal dimension to the image dimension.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work develops a simple yet effective Thornail Layout (TALL) for deepfake detection by spatio-temporal modeling. TALL is computationally cheap and retains both temporal and spatial information. In detail, we use dense sampling to extract multiple clips in the video and then randomly select four consecutive frames in the video segment. Subsequently, a block is masked at a fixed position in each frame. Finally, the frames are resized as sub-image and sequentially rearranged into a pre-defined layout as a thumbnail, which has the same size as the clip frames. As shown in Figure 1, TALL brings two advantages compared to the previous spatio-temporal modeling methods for deepfake detection: (1) TALL contains local and global contextual deepfake patterns. (2) TALL is a model-agnostic method for spatio-temporal modeling deepfake patterns at zero computation and zero parameters.", + "bbox": [ + 75, + 273, + 472, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we discover that the better temporal modeling capabilities backbone has, the better performance TALL achieves. Based on the proposed TALL, we complement a baseline for video deepfake detection based on Swin Transformer [36], called TALL-Swin. We validate TALL-Swin on four popular benchmark datasets, including FaceForensics++, Celeb-DF, DFDC, and DeeperForensics. Our method gains a remarkable improvement over the state-of-the-art approaches. The main contributions of our paper are summarized as follows:", + "bbox": [ + 75, + 517, + 472, + 670 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide a new perspective for an efficient strategy for video deepfake detection called Thumbnail Layout (TALL), which incorporates both spatial-temporal dependencies, and allows the model to capture spatial-temporal inconsistencies.", + "- We propose a spatio-temporal modeling method called TALL-Swin, which efficiently captures the inconsistencies between deepfake video frames.", + "- Extensive experiments demonstrate the validity of our proposed TALL and TALL-Swin. TALL-Swin outperforms previous methods in both intra-dataset and cross-dataset scenarios." + ], + "bbox": [ + 96, + 684, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 642, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image-Level Deepfake Detection", + "text_level": 1, + "bbox": [ + 500, + 114, + 785, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Typically, existing deepfake detection methods fall into two categories: image-level and video-level methods. The image-level methods [25, 13] always exploit the artifacts of deepfake images in the spatial domain, such as discrepancies between local regions [42, 59], grid-like structure in frequency space [10], and differences in global texture statistics [37] that provide specific clues to distinguish deepfakes from the real images. F3Net [43] and FDFL [31] utilize the same pipeline that utilizes frequency-aware features and RGB information to capture the traces in different input spaces separately. RFM [54] and Multi-att [65] propose an attention-guided data augmentation mechanism to guide detectors to discover undetectable deepfake clues. Face X-ray [32] and PCL [67] provide effective ways to outline the boundary of the forged face for detecting deepfakes. ICT [11] exploits an identity extraction module to detect identity inconsistency in the suspect image. Similarly, M2tr [55] detects local inconsistencies within frames at different spatial levels. Generally, image-level methods suffer over-fitting issues when a specific technique manipulates the images, and they ignore temporal information.", + "bbox": [ + 496, + 137, + 893, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Video-Level Deepfake Detection", + "text_level": 1, + "bbox": [ + 500, + 465, + 782, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve the generalization of deepfake detectors, many studies generate diversity and generic deepfake data, while other studies capture the temporal incoherence of fake videos as generic clues. Some recent works propose detecting temporal inconsistency using well-designed spatiotemporal neural networks, and others [15, 16] attempt to add modules to image models that capture temporal information. STIL [15] formulates deepfake video detection as a spatial and temporal inconsistency learning process and integrates both spatial and temporal features in a unified 2D CNN framework. FTCN [68] detects temporal-related artifacts instead of spatial artifacts to promote generalization. LipForensics [18] is proposed to learn high-level semantic irregularity in mouth movement in the generated video. RealForensics [17] uses auxiliary data sets during training in exchange for generalization at the cost of higher computational demands. The video-based methods achieve strong generalization but suffer from large computational overhead. To reduce computational costs, we propose TALL which gathers consecutive video frames into thumbnails for learning spatio-temporal consistency.", + "bbox": [ + 496, + 488, + 893, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Deepfake Detection with Vision Transformer", + "text_level": 1, + "bbox": [ + 500, + 816, + 879, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, ViT [12] has achieved impressive performance in computer vision tasks [24, 23, 44]. Many studies extend the ViT for deepfake detection [66, 58]. These methods achieve better performance compared to CNN-based", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22659", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models, but also sacrifice computational efficiency. A few works [55, 66] attempt to extend the transformer for deepfake detection due to the advent of the visual transformer (ViT) and the impressive ability to model long-range data, different from two-branch architectures that capture short-range and long-range temporal inconsistencies with a single-branch model. ICT [11] aims to detect identity consistency in deepfake video but may fail in detecting face reenactment and entire face synthesis results. DFLL [28] extract the UV texture map to help the transformer to detect deepfakes, which may disrupt the continuity between video frames. DFTD [29] leverages ViT to consider both global and local information but ignores the problem of excessive model arithmetic requirements. Although the transformer-based approaches achieve promising performance, they are accompanied by significant computational complexity that makes them challenging to deploy and use, and the long-range dependencies may be insufficiently exploited in detection models. Swin Transformer [36] produces a hierarchical feature representation and has linear computational complexity concerning input image size, which is suitable as a general-purpose backbone for various vision tasks. In this paper, we cooperate with Swin Transformer to form our robust and efficient method TALL-Swin.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 465, + 171, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TALL is a deepfake video detection strategy that transforms a video clip into an all-in-one thumbnail without the extra computational overhead. In the following sections, we begin with the motivation of TALL for deepfake detection in Section 3.1. Then we present the technical details of the TALL in Section 3.2. Finally, a generalizable SwINTALL baseline is introduced to explore subtle artifacts in Section 3.3. More details of the layout design are presented in the Appendix.", + "bbox": [ + 75, + 491, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 76, + 635, + 199, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While recent studies have attempted to address noticeable flaws through techniques like slight motion blurring and temporal consistency loss, subtle spatio-temporal artifacts still remain. These artifacts are important for detecting deepfakes, but they introduce two problems: 1) video-based models are less efficient, and 2) analyzing information over long distances may overlook local artifacts, which are critical for deepfake detection. To address these challenges, we propose the TALL strategy, which naturally incorporates temporal information into image-level tasks without disrupting spatial information. This approach enables the image-level model to detect deepfakes in videos. Furthermore, we discovered that TALL provides even greater performance gains when combined with a powerful spatial model, resulting in the TALL-Swin.", + "bbox": [ + 75, + 659, + 468, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In detail, TALL arranges consecutive frames in the", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9380b024a1cd3eed710e5f8204bc0c4fabb7b49b9666b2f3aca592a19a7c0fda.jpg", + "image_caption": [ + "(a) Thumbnail Layout (TALL)", + "Figure 2. Illustration of the TALL and shifted window process for computing self-attention in the TALL." + ], + "image_footnote": [], + "bbox": [ + 509, + 87, + 709, + 243 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e79a74cc79cb3363864ed670b85229736b93aab6050a999cedcc01393ba14187.jpg", + "image_caption": [ + "(b) Feature Map" + ], + "image_footnote": [], + "bbox": [ + 709, + 90, + 885, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "temporal order in a compact $2 \\times 2$ layout, in line with the calculation theory of convolution and shifted window. TALL contains both spatial and temporal information so that model can learn both intra-frame artifacts and interframe inconsistency and obtains comparable performance to video-based methods. Here we use the shifted window to explain TALL's mechanism. As illustrated in Figure 2 (a), the model computes self-attention while accounting for spatial dependencies across sub-images (represented by the solid red box). When the window spans multiple sub-images (represented by the red dash box), the model is able to capture temporal inconsistencies between frames. Moreover, TALL leverages both local and global contexts of deepfake patterns to ensure robust modeling capabilities for short and long-range spatial dependencies. Compared to previous methods, we anticipate that TALL strikes a balance between speed and accuracy, sacrificing a little spatial information while preserving performance. Based on the fact that attention-based models are better at handling contextual features and that the Swin-Transformer uses shifted windows to reduce computation and memory, we further complement TALL-Swin baseline for video deepfake detection.", + "bbox": [ + 496, + 320, + 892, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Thumbnail Layout (TALL)", + "text_level": 1, + "bbox": [ + 500, + 678, + 745, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video $V \\in \\mathbb{R}^{T \\times C \\times H \\times W}$ , where $T$ is the frame length of the video, $C$ is the number of channels, and $H \\times W$ is the resolution of the frames. Assuming each video contains $N$ clips, we divide a video into $N$ equal segments of length $T / N$ and then sample consecutive $t$ (set to 4 by default) frames from the segments at random locations to form one clip. Then, the thumbnail $I$ is rearranged of sub-images $(C \\times \\frac{H}{\\sqrt{t}} \\times \\frac{W}{\\sqrt{t}})$ that are resized from the above $t$ frames. To maximize the utility of TALL, we mask the organized $N$ square masks of the thumbnail. It is based on two core designs: 1) The position of the masks is random between different sub-images, which retains the advantages of the Cutout [8] that encourages the network to focus more", + "bbox": [ + 496, + 700, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22660", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Pseudocode of TALL in a PyTorch-like style." + ], + "code_body": "# x: one clip of video (T*C*H*W) \n# T: frame number of clip \n# C: channels; s: mask size \n# d: length of frame included in the thumbnail \n# r: rows of thumbnail \n# x_tall: thumbnail image (224*224) \n#TALL's augmentation strategy \nh = np.random.randint(H) \nw = np.random.randint(W) \n#the mask position is fixed for each frame \nm = np.ones((H, W)) \nh1 = np.clip(h - s // 2, 0, H) \nh2 = np.clip(+s // 2, 0, H) \nw1 = np.clip(w - s // 2, 0, W) \nw2 = np.clip(w + s // 2, 0, W) \nm[h1: h2, w1: w2] = 0 \nm = torch.from_numpy(m) \nm = maskexpand_as(x) \nx = x * m \n#TALL: generation of the thumbnail \nx = x.view(-1, H, W).unsqueeze(0) \nx = x.view((-1, C*d) + x_tall.size() [2:]) \nx = rearrange(x, 'b (th tw c) h w -> b c (th h) (tw w)', th=r, c=C) \nx_tall = interpolate(x, size=H)", + "bbox": [ + 76, + 116, + 450, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "on complementary and less prominent features. 2) We fix the position of the mask within a clip to take advantage of the fact that most deepfake videos are frame-by-frame tampered with, thus forcing the model to detect inconsistencies between adjacent frames of the deepfake videos. We do not allow the mask to appear on the seams of the thumbnail but allow for partial mask inclusion in the thumbnail. The detailed procedure of TALL is summarized in Algorithm 1.", + "bbox": [ + 75, + 460, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. TALL-Swin", + "text_level": 1, + "bbox": [ + 76, + 590, + 205, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To balance efficiency and model performance for spatiotemporal feature learning and to leverage the benefits of attention-based models, we enhanced a baseline deepfake detection model called TALL-Swin by incorporating the Swin Transformer [36]. Given the characteristics of TALL, we slightly modified the window size of Swin-B in TALL-Swin. We first enlarge the window size of the first three stages of the model so that the interaction between frames in the thumbnail becomes more frequent, forcing the model to learn more detailed spatio-temporal dependencies. Next, we set the window size of the last stage to be the same as the feature map size, enabling the window to perform global attention computations while TALL-Swin captures global spatial-temporal dependencies. As a result, the size of the last layer of the feature map became smaller, reducing the window size without introducing any additional computational overhead. Consequently, the window sizes for the four stages of TALL-Swin are [14, 14, 14, 7]. Note that the patch merging process makes TALL-Swin captures a more", + "bbox": [ + 75, + 613, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "comprehensive range of dependencies through hierarchical representations, as shown in Figure 2 (b).", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a video of length $T$ , each frame contains $N$ patches, and the window contains $P$ patches. To demonstrate the superiority of TALL-Swin in terms of computational consumption, we show below the computational complexity of the image-level transformer and video-level transformer, including ViT [12], Swin [36], ViViT [2], and TALL-Swin respectively:", + "bbox": [ + 496, + 119, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {\\mathrm {V i T}} = 4 T N C ^ {2} + 2 T N ^ {2} C,\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 237, + 769, + 253 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {\\mathrm {S w i n}} = 4 T N C ^ {2} + 2 T P N C, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 255, + 890, + 279 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {\\mathrm {V i V i T}} = 4 T N C + 2 T ^ {2} N ^ {2} C,\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 273, + 782, + 290 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {\\mathrm {T A L L - S w i n}} = T N C ^ {2} + \\frac {1}{2} T P N C.\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 292, + 816, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TALL-Swin has the lowest computational complexity compared to image and video-level transformer methods. Subsequent experiments will demonstrate that TALL-Swin maintains performance, albeit at the sacrifice of some spatial information.", + "bbox": [ + 496, + 320, + 890, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cross-entropy loss is employed to optimize the TALL-Swin, which is defined as:", + "bbox": [ + 498, + 396, + 890, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C E} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} y _ {i} \\log \\mathcal {F} \\left(x _ {i}\\right) + (1 - y _ {i}) (\\log (1 - \\mathcal {F} \\left(x _ {i}\\right)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 433, + 890, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{i}$ indicates input clip, $y_{i}$ denotes the label of clip, $n$ is the number of clip, $\\mathcal{F}$ is TALL-Swin.", + "bbox": [ + 498, + 476, + 890, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 520, + 632, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 545, + 581, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. Following previous works [18, 17, 68], we evaluate the TALL and TALL-Swin on four widely used datasets. FaceForensics++ [45] is a most-used benchmark on intra-dataset deepfake detection, consisting of 1,000 real videos and 4,000 fake videos in four different manipulations: DeepFake [7], FaceSwap [38], Face2Face [52], and NeuralTextures [51]. Besides, FaceForensics++ contains multiple video qualities, e.g. high quality (HQ), low quality (LQ) and RAW. Celeb-DF (CDF) [34] is a popular benchmark on cross-dataset, which contains 5,693 deepfake videos generated from celebrities. The improved compositing process was used to improve the various visual artifacts presented in the video. Celeb-DF is also suitable for deepfake detection tasks with a reference set. DFDC [9] is a large-scale benchmark developed for Deepfake Detection Challenge. This dataset includes 124k videos from 3,426 paid actors. The existing deepfake detection methods do perform not very well on DFDC due to their sophisticated deepfake techniques. DeeperForensics (DFO) [26] includes 60,000 videos with 17.6 million frames for deepfake detection, whose videos vary in identity, pose, expression, emotion, lighting conditions, and blend shape with high quality.", + "bbox": [ + 496, + 568, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22661", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d47850939f4b6a3d246b79ba545e9a3edbbf95adcb1b8dad80bb055de495114d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Performance of different backbones. TALL consistently improves the accuracy over different image-level models. We show the AUC, FLOPs, and number of parameters for each model on the cross-dataset scenario. All models are trained on $\\mathrm{FF} + +$ (HQ). $\\checkmark$ indicates the model enables temporal modeling. \\* indicates our implementation. PT indicates pre-train. 1K and 21K indicate the model pre-trained on ImageNet-1K and 21K respectively. The best results are bold." + ], + "table_body": "
ModelsTemp.CDFDFDCFLOPsParamsPT
I3D-RGB* [4]✓78.2465.58222.7G25M1K
R3D-50* [19]✓79.6367.73296.6G46M1K
ResNet50* [20]×76.3864.0125.5G21M1K
+TALL✓80.9065.5425.5G21M1K
EffNetB4* [50]×78.1966.818.3G19M1K
+TALL✓83.3767.158.3G19M1K
VTN [41]✓83.2073.50296.6G46M21K
VidTR [63]✓83.3073.30117G93M21K
ViViT* [2]✓86.9674.61628G310M21K
ISTVT [64]✓84.1074.20455.8G--
ViT-B* [12]×82.3372.6455.4G84M21K
+TALL✓86.5874.1055.4G84M21K
Swin-B* [36]×83.1373.0147.5G86M21K
TALL-Swin✓90.7976.7847.5G86M21K
", + "bbox": [ + 81, + 88, + 460, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We use MTCNN to detect face for each frame in the deepfake videos, only extract the maximum area bounding box and add $30\\%$ face crop size from each side as in LipForensics [18]. The ImageNet-21K pretrained Swin-B model is used as our backbone. Excluding ablation experiments, we sample 8 clips using dense sampling, each clip contains 4 frames. The size of the thumbnail is $224 \\times 224$ . Following Swin Transformer [36], Adam [30] optimization is used with a learning rate of 1.5e-5 and batch size of 4, using a cosine decay learning rate scheduler and 10 epochs of linear warm-up. We adopt Acc. (accuracy) and AUC (Area Under Receiver Operating Characteristic Curve) as the evaluation metrics for extensive experiments. To ensure a fair comparison, we calculate video-level predictions for the image-based method and average the predictions across the entire video (following previous works [18, 16, 35, 68]). Note that results are directly cited from published papers if we follow the same setting.", + "bbox": [ + 75, + 505, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Scaling over Backbones", + "text_level": 1, + "bbox": [ + 76, + 786, + 294, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To verify our assumption, we adopt several image-level backbones commonly used for deepfake detection for comparison with the video-level backbones. As shown in Table 1 above the double horizontal line, we first compare the accuracy and complexity of the CNN-based video and image backbones. Although I3D [4] and R3D [19] achieve", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "better performance than vanilla ResNet50 [20] and EfficientNet [50], the computation costs are huge, such as R3D-50 with 296G FLOPs. For ResNet and EfficientNet who added TALL, ResNet achieves better AUC both on CDF (76.38 VS 80.93) and DFDC (64.01 VS 65.54) datasets, and EfficientNet achieves $5.18\\%$ better AUC on CDF.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The second section contains the video and image transformers. Compared to video transformers, the image-based ViT and Swin fail to achieve better performance due to the lack of temporal modeling. For example, ViViT achieves $86.96\\%$ AUC on CDF, which is $3.6\\%$ higher than Swin although ViViT with $13\\times$ more computation. By way of contrast, $\\mathrm{ViT + TALL}$ achieves $86.58\\%$ AUC on CDF with 55.4G FLOPs, which is comparable to AUC with ViViT but with low computation. Accordingly, Swin's performance was significantly improved with the addition of TALL without computation increment. On the other hand, TALL boosts higher performance on models with learned long-range dependencies. e.g., ResNet+TALL $(+4.5\\%)$ on CDF and $+1.5\\%$ on DFDC) vs. Swin+TALL $(+7.6\\%)$ on CDF and $+3.6\\%$ on DFDC). These two section results demonstrate that TALL provides both spatial and temporal information and enables the model to learn spatial and temporal inconsistencies for video deepfake detection.", + "bbox": [ + 496, + 189, + 892, + 460 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/63ab115643577cd65dd2d27bef29b6d0783a3ffd068bc9fab1d7656687ad2f2a.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Intra-dataset evaluations. We report the video-level Acc. $(\\%)$ and AUC $(\\%)$ on the FF++ dataset. HQ indicates high quality, and LQ indicates low quality." + ], + "table_body": "
MethodsFF++ (HQ)FF++ (LQ)
Acc.AUCAcc.AUC
MesoNet [1]83.10-70.47-
Xception [6]95.7396.3086.8689.30
Face X-ray [32]-87.35-61.60
Two-branch [39]96.4398.7086.3486.59
Add-Net [70]96.7897.7487.5091.01
F3-Net [43]97.5298.1090.4390.43
FDFL [31]96.6999.3089.0092.40
Multi-Att [65]97.6099.2988.6990.40
RECCE [3]97.0699.3291.0395.02
LipForensics [18]98.8099.7094.2098.10
DFDT [29]98.1899.2692.6794.48
ADT [56]92.0596.3081.4882.52
ST-M2TR [55]-99.42-95.31
VTN [41]98.47-94.02-
VidTR [63]97.42-92.12-
ViViT* [2]92.60-88.02-
ISTVT [64]99.00-96.15-
TALL-Swin98.6599.8792.8294.57
", + "bbox": [ + 501, + 489, + 888, + 829 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22662", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 90, + 441, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intra-dataset evaluations. Following ISTVT [64], we show the results of the $\\mathrm{FF}++$ dataset under both Low Quality (LQ) and High Quality (HQ) videos, and report comparisons against several advanced methods in Table 2. We can observe that advanced video-based transformers have better results than CNN-based methods. Compared to video-based transformer methods, TALL-Swin has comparable performance and lower consumption to the previous video transformer method with HQ settings. However, TALL-Swin gets unsatisfactory results with the LQ setting. The LQ setting is obtained by severely compressing the videos. So the reason for the result may be that TALL scales the frame to a smaller size, causing more spatial information to be lost in the frame. We will investigate the possibility of other designs to further improve performance in the LQ setting.", + "bbox": [ + 75, + 113, + 472, + 340 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/02e03602c7d07f0597c648418a18420c6be6b44a7ba86c65d1742993416db232.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCDFDFDCFShDFoAvg.
Xception [6]73.7070.9072.0084.5075.28
CNN-aug [57]75.6072.1065.7074.4071.95
CNN-GRU [46]69.8068.9080.8074.1073.40
Patch-based [5]69.6065.6057.8081.8068.70
Face X-Ray [32]79.5065.5092.8086.8081.15
Multi-Att [65]75.7068.1066.0077.7071.88
DSP-FWA [33]69.5067.3065.5050.2063.13
LipForensics [18]82.4073.5097.1097.6087.65
FTCN [68]86.9074.0098.8098.8089.63
RealForensics [17]86.9075.9099.7099.3090.45
DFDT [29]88.3076.1097.8096.9089.70
VTN [41]83.2073.5098.7097.7088.30
VidTR [63]83.5073.3098.0097.9088.10
ViViT* [2]86.9674.6199.4199.1990.05
ISTVT [64]84.1074.2099.3098.6089.10
TALL-Swin90.7976.7899.6799.6291.71
", + "bbox": [ + 83, + 354, + 460, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Generalization to unseen datasets. We report the video-level AUC (\\%) on four unseen datasets: Celeb-DF (CDF), DFDC, FaceShifter (FSh), and DeeperForensics (DFo).", + "bbox": [ + 75, + 646, + 468, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization to unseen datasets. In addition to the intra-dataset comparisons, we also investigate the generalization ability of our method. Adhering to the deepfake video detection cross-dataset protocol [18], we train a model on $\\mathrm{FF}++$ (HQ) then test on Celeb-DF (CDF), DFDC, FaceShifter (FSh), and DeeperForensics (DFo) datasets. As shown in Table 3: (1) Video-based methods generally have better results than image-based methods, which shows that temporal information is helpful for the deepfake video detection task. For example, Lip outperforms Face X-ray's AUC by a wide margin. In addition, most transformer-based models have higher performance than CNN-based models. For the transformer-based models, both achieved", + "bbox": [ + 75, + 704, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "an average AUC of $88\\%$ , while the best CNN-based video-level models only achieved $87\\%$ . (2) TALL-Swin achieves state-of-the-art results on Celeb-DF, DFDC, and Deeper-Forensics datasets, and also beats its competitors on Celeb-DF dataset by a large margin $(3.8\\%)$ . The results demonstrate that TALL-Swin performs well when encountering unseen datasets with better generalization ability than previous video transformer methods.", + "bbox": [ + 496, + 90, + 893, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67bba2563ef26452db382cf417d3f1da6d2be31f271f9fa32b71b43dc6e1e2e1.jpg", + "image_caption": [ + "Figure 3. Saliency map visualization of TALL-Swin on different datasets. The first four rows of samples are from the FF++ dataset, and the last four rows are from the unseen datasets." + ], + "image_footnote": [], + "bbox": [ + 511, + 222, + 888, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of saliency map visualization. We adopt Grad-CAM [47] to visualize where the TALL-Swin is paying its attention to the deepfake faces. In Figure 3, we give the results on intra-dataset and cross-dataset scenarios. All models are trained on $\\mathrm{FF} + +$ (HQ). It can be observed in the first four rows of Figure 3 that TALL-Swin captures method-specific artifacts. Note that the DF transfers the face region from a source video to a target, and the NT only modifies the facial expressions corresponding to the mouth region. TALL-Swin corresponds to focus on the face region and the mouth region. Furthermore, our model traces the more generalized artifacts that are independent of manipulation methods, e.g., blending boundaries (CDF), and abnormal motions in the clip (DFDC, Fsh, Dfo).", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22663", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5b466759f3094b0caecd7262ea43e40088af880ef34f19ca52484688d391b08.jpg", + "image_caption": [], + "image_footnote": [ + "Xception Face X-ray ICT Patch-based CNN-GRU LipForensics TALL-Swin" + ], + "bbox": [ + 81, + 113, + 295, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9c8c8662d00ed5404ab6de3103706e6d41d22a449160c4a34fd368fdebe3a9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 113, + 493, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7b257598955b38b8906a145e1afced6ed0b20b2ce4e1c02ecef460c9f66632f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 113, + 691, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f894b31f0b4f348eb464cab653a19b18908a33ca7358a86a0486885792461b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 113, + 888, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1288e83695a1a47eafc8e248fd431a864848075e0e0f2b481c6b972c2ef75ab7.jpg", + "image_caption": [ + "Figure 4. Robustness to various unseen corruptions. We report the video-level AUC $(\\%)$ of our methods under five different levels of seven particular types of corruption. \"Average\" denotes the mean across all corruptions at each severity level. Our TALL-Swin is more robust than previous methods for all corruptions." + ], + "image_footnote": [], + "bbox": [ + 80, + 270, + 295, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9471f592a87e976ec6610e2702a13b8da8f7c970ae161d7246166f975e8baca1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 270, + 493, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b21159c95b3a03adbf4842f84a81ce5140cd6d08b6d0aa5ed8025853777083c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 270, + 691, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e65bbfe61523b63ee11bbbf46dccf1e03b5dee93a39b68ce6d023f8ae9db92d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 270, + 888, + 436 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/31ce922a2577f7fb9a1a77063a9d3e92be408c7ccc26bbb7bfcf14d115aff431.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCleanSaturationContrastBlockNoiseBlurPixelCompressAvg.
Xception (ICCV'19) [6]99.899.398.699.753.860.274.262.178.3
CNN-GRU (CVPRW'19) [46]99.999.098.897.947.971.586.574.582.3
CNN-aug (CVPR'20) [57]99.899.399.195.254.776.591.272.584.1
Patch-based (ECCV'20) [5]99.984.374.299.250.054.456.753.467.5
Face X-ray (CVPR'20) [32]99.897.688.599.149.863.888.655.277.5
LipForensics (ICCV'21) [18]99.999.999.687.473.896.195.695.692.5
FTCN (ICCV'21) [68]99.499.496.797.153.195.898.286.489.5
RealForensics (CVPR'22) [17]99.899.899.698.979.795.398.497.695.6
TALL-Swin w/o mask100.0100.0100.099.883.597.398.497.996.7
TALL-Swin100.0100.0100.0100.085.397.698.598.197.1
", + "bbox": [ + 84, + 496, + 883, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Average robustness to unseen corruptions. Average Video-level AUC (%) across five intensity levels for each corruption type proposed in DFo [26]. \"Avg\" indicates the mean across all corruptions and all levels.", + "bbox": [ + 75, + 680, + 892, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to unseen perturbations. Deepfake detectors must be robust to common perturbations, given that video propagation on social media causes video compression, noise addition, etc. We also study the performance of robustness to unseen perturbations. Following RealForensics [17], the experiment applies seven unseen perturbations to fake videos at five intensity levels. In Figure 4, we show results of increasing the severity of each corruption. We can observe that other methods degrade dramatically as the perturbations become more severe. TALL-Swin still has a high performance. However, TALL-Swin degrades when", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the Gaussian noise reaches level five. Table 4 presents the average AUC across all intensity levels for corruption types. We observe that our method is significantly more robust to most perturbations than other methods. The good robustness may be from both the design of TALL and the proposed mask augmentation. The main reason may be the consecutive multi-frame input. We empirically consider that the key to deepfake detection is local inconsistency, the continuous frame design has less redundant information, ensuring that the model finds locally important clues.", + "bbox": [ + 496, + 734, + 893, + 886 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22664", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c8a499cba746e8873144849dd2f06d2a7bfb91df4a44f8e8d85499bad080da71.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 120, + 88, + 151, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3fb8489399ebb8d07696b41d794cdcb57ea4bf936005c3bc7d07cc0e1403dfc.jpg", + "image_caption": [ + "Figure 5. Illustration of different layout designs." + ], + "image_footnote": [], + "bbox": [ + 171, + 127, + 277, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9dd06e16684826bf5dff944678b7f2b73766ac45bb6fc7836e624b0dd6e9004d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 107, + 352, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6a081a7e563d69a532187f0d2fe7af46d89f58a5e4d20d9cae25f9fd9692e0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 107, + 426, + 166 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 234, + 228, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform the ablation study to analyze the effects of each component and hyper-parameter in TALL-Swin. All experiments are trained on $\\mathrm{FF}++$ (HQ) and tested on the CDF and DFDC datasets.", + "bbox": [ + 75, + 257, + 470, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of different layouts. We train a TALL-Swin model on $\\mathrm{FF}++(\\mathrm{HQ})$ for each layout illustrated in Figure 5, to analyze in which layout of the thumbnails the model learns the strongest generalization of the spatial-temporal dependence of the deepfake patterns. As shown in Table 5, the model with a compact layout like Figure 5 (d) has good generalization ability on the unseen datasets. A compact layout like Figure 5 (d) may help the model to learn the temporal dependence across frames because such a form provides the shortest distance between any two images.", + "bbox": [ + 75, + 316, + 468, + 468 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/676730e8cb33e7f21524d0f229c053981161edc42ba4cde84c994b60c10e80c9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LayoutCDFDFDC
Figure 5 (a)85.5270.02
Figure 5 (b)84.9373.57
Figure 5 (c)86.6672.12
Figure 5 (d)87.6074.32
", + "bbox": [ + 169, + 479, + 375, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Study on the numbers of sub-image. We use Swin-B as the baseline for this study to compare the effect of different thumbnail layout schemes on the model's generalization ability. Changing frames to thumbnails involves scaling, so we also investigate the impact of resizing and random cropping pre-processing on model performance. We set up four variants: resizing pre-process with $4 \\times 4$ layout, $3 \\times 3$ layout and $2 \\times 2$ layout; random cropping pre-process with $2 \\times 2$ layout. As shown in Table 6, the model performance degrades sharply when using $4 \\times 4$ layout. This may be due to the small size of each sub-image that the spatial information is not captured well by the model. The result of $3 \\times 3$ layout also slightly decreases. $2 \\times 2$ layout with resizing pre-processing beats $2 \\times 2$ layout with random crop. We also found that TALL-Swin achieves the best performance and the AUC score increases $3.2\\%$ compared to the baseline, suggesting that thumbnails in a $2 \\times 2$ layout are more helpful to TALL-Swin than original frames.", + "bbox": [ + 75, + 613, + 468, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of Sub-image's size. We eliminate the scaling", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/903cc1a1261f9e2e0d900ff2cc3e4a467366129238e62d8387e29832aeee4ea8.jpg", + "table_caption": [ + "Table 5. Effects of different layouts. All models here are trained without mask augmentation." + ], + "table_footnote": [], + "table_body": "
Pre-processLayoutCDFDFDC
None-83.1373.01
Resize4 × 480.1870.45
Resize3 × 383.1872.98
Crop2 × 278.5573.30
Resize2 × 287.6074.32
", + "bbox": [ + 550, + 88, + 834, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "operation for sub-images to allow for more flexible layout settings. However, we've observed that when the number of sub-images grows at their original size, the performance improvements are only slight. Additionally, the computational complexity increases dramatically with the number of frames ( 4.3 times more than the TALL setting), as demonstrated in Table 7. To strike a balance between performance and computational complexity, we reduce the resolution of sub-images in TALL.", + "bbox": [ + 496, + 253, + 893, + 388 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9a72269dcd8e31ee759a5b2ff353d66f785217f6c50acb9689b4286f2c6daf9f.jpg", + "table_caption": [ + "Table 6. Study on the numbers of sub-image. All models here are trained without mask augmentation." + ], + "table_footnote": [], + "table_body": "
Subimage-sizeLayoutFLOPsCDFDFDC
224 × 2243 × 3253G88.6975.98
224 × 2242 × 2185G88.1575.01
112 × 1122 × 247.5G87.6074.32
", + "bbox": [ + 508, + 401, + 879, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Study on absence and order of thumbnails. In this case, we study the impact of missing the last sub-image and the last two sub-images on the model's performance. The first two rows of Table 8 show that all four sub-images contribute to the model performance. Besides, we set the order of the different thumbnails to evaluate the TALL-Swin. We consider three orders: forward, reverse, and random. Forward order performs the best for three different orders. This may be because of the positional encoding of different frames in TALL-Swin.", + "bbox": [ + 496, + 526, + 890, + 676 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/14d54f0cefb5a76bfaef48899b15224d256ae61cbb7e1e63220241a3458bf7b3.jpg", + "table_caption": [ + "Table 7. Effects of Sub-image's size. All models here are trained without mask augmentation." + ], + "table_footnote": [], + "table_body": "
VariantsCDFDFDC
0, 1, 2, -86.4669.51
0, 1, -, -84.2269.09
Random85.8570.30
Reverse86.6572.37
Forward87.6074.32
", + "bbox": [ + 573, + 688, + 810, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablation study of absence and order of thumbnails. All models here are trained without mask augmentation.", + "bbox": [ + 500, + 794, + 890, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of different orders on other backbones. In order to prove that the phenomenon is not incidental, we also conduct experiments on ResNet50 and EfficientNet for three different orders as shown in Table 9. As expected, for-", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22665", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ward order outperforms reverse and random orders both on ResNet50 and EfficientNet, which indicates that TALL can learn the temporal dependency.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1ba2ed4c1d95f8e37238a565bfea4bdf3d1e586b0284d821bc5d896751cba4a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantsCDFDFDC
ResNet50+TALL76.3864.01
Random78.1464.12
Reverse78.5464.87
Forward80.9065.54
EfficientNet+TALL78.1966.81
Random81.0166.13
Reverse81.6666.69
Forward83.3767.15
", + "bbox": [ + 133, + 148, + 405, + 306 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effectiveness of mask strategy. In this work, TALL-Swin is trained on the $\\mathrm{FF}++$ (HQ) dataset without any data enhancement as the baseline except for Multi-scale Crop and Random Horizontal Flip. To validate the effectiveness of the mask strategy, we compare our default baseline with different data augmentation strategies: 1) The Cutout [8] on one sub-image; 2) The Cutout on four sub-images. 3) The combination of Mixup [62] and Cutmix [60] on four sub-images, as shown in Table 10. The performance of a random Cutout [8] on four sub-images is better than on one sub-image. Besides, the mask strategy leads to better performance than the well-known Cutout $(1.46\\%)$ . This supports our hypothesis that strategy encourages models to learn subtle temporal-spatial variations and improves model generalization ability. Further, our augmentation strategy exceeds $1.02\\%$ than the combination of Mixup and Cutmix, demonstrating the augmentation's effectiveness in TALL for video detection.", + "bbox": [ + 75, + 338, + 470, + 609 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/cd3925bde530eb43fec2833a8bf9fba7eb0fc782c4f6f503a13b396cb1d5cdd7.jpg", + "table_caption": [ + "Table 9. Ablation studies of different orders of TALL." + ], + "table_footnote": [], + "table_body": "
AugmentationCountCDFDFDC
None-87.6074.32
Cutout189.0674.07
Cutout489.3375.22
Mixup+CutMix489.7575.33
TALL's mask490.7976.78
", + "bbox": [ + 130, + 619, + 410, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Study on window size. We study the effect of window size on model performance and computational cost. The results are shown in Table 11. Our window expansion for the first three phases will increase the model performance by $1.74\\%$ AUC. The results in the second and third rows show that the first three stages of the window getting the largest would not give a boost to the model. Our analysis of a too-large window may lead to a weakening of the model's ability to learn local information in the sub-image.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b2ed7f0f736ce4b43d50f5e44bb748433161e61c7701b5ca98c3d04f0ce33e59.jpg", + "table_caption": [ + "Table 10. Study of the augmentation strategy in TALL. The count column represents the number of blocks on the thumbnail." + ], + "table_footnote": [], + "table_body": "
Window sizeCDFDFDC
(7,7,7,7)85.6073.32
(14,14,14,7)87.6074.32
(28,28,28,7)86.6574.21
", + "bbox": [ + 581, + 87, + 807, + 152 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 11. Ablation studies of window size. All models here are trained without mask augmentation.", + "bbox": [ + 498, + 156, + 890, + 184 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 208, + 617, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper presents a novel perspective on detecting deepfake videos using TALL. TALL is both simple and effective, enabling joint spatio-temporal modeling without any additional costs. TALL representation reveals normal deepfake patterns with local-global contextual features. We further propose a new baseline for deepfake video detection called TALL-Swin, which efficiently captures the inconsistencies between deepfake video frames. Extensive experiments demonstrate that TALL-Swin achieves promising results for various unseen deepfake types and strong robustness to a wide range of common corruptions.", + "bbox": [ + 496, + 234, + 890, + 401 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 416, + 650, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially funded by National Natural Science Foundation of China under Grants (62276256, U21B2045 and U20A20223) and Beijing Nova Program under Grant Z211100002121108. The authors wish to thank Huaibo Huang and Lijun Sheng in no particular order, for insightful discussions.", + "bbox": [ + 496, + 441, + 890, + 532 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 546, + 596, + 561 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In Proc. WIFS, pages 1-7, 2018. 5", + "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proc. ICCV, pages 6836-6846, 2021. 4, 5, 6", + "[3] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proc. CVPR, pages 4113-4122, 2022. 5", + "[4] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proc. CVPR, pages 6299-6308, 2017. 5", + "[5] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Proc. ECCV, pages 103-120, 2020. 6, 7", + "[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proc. CVPR, pages 1251-1258, 2017. 5, 6, 7", + "[7] deepfakes. Deepfakes. https://github.com/deepfakes/faceswap.2021-11-13.4" + ], + "bbox": [ + 506, + 571, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22666", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[8] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv:1708.04552, 2017. 3, 9", + "[9] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv:2006.07397, 2020. 4", + "[10] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In Proc. CVPR, pages 7865-7874, 2022. 2", + "[11] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Ting Zhang, Weiming Zhang, Nenghai Yu, Dong Chen, Fang Wen, and Baining Guo. Protecting celebrities from deepfake with identity consistency transformer. In Proc. CVPR, pages 9468-9478, 2022. 2, 3", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. ICLR, 2021. 2, 4, 5", + "[13] Jianwei Fei, Yunshu Dai, Peipeng Yu, Tianrun Shen, Zhihua Xia, and Jian Weng. Learning second order local anomaly for general face forgery detection. In Proc. CVPR, pages 20270-20280, 2022. 2", + "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. In Proc. NeurIPS, volume 27, 2014. 1", + "[15] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proc. ACM MM, pages 3473-3481, 2021. 1, 2", + "[16] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, and Lizhuang Ma. Delving into the local: Dynamic inconsistency learning for deepfake video detection. In Proc. AAAI, pages 744-752, 2022. 1, 2, 5", + "[17] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proc. CVPR, 2022. 2, 4, 6, 7", + "[18] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In Proc. CVPR, pages 5039-5049, 2021. 2, 4, 5, 6, 7", + "[19] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Learning spatio-temporal features with 3d residual networks for action recognition. In Proc. ICCV Workshops, pages 3154-3160, 2017. 5", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, pages 770-778, 2016. 5", + "[21] Gee-Sern Hsu, Chun-Hung Tsai, and Hung-Yi Wu. Dual-generator face reenactment. In Proc. CVPR, pages 642-650, 2022. 1" + ], + "bbox": [ + 80, + 92, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Yihao Huang, Felix Juefei-Xu, Qing Guo, Yang Liu, and Geguang Pu. Fakelocator: Robust localization of gan-based face manipulations. IEEE Transactions on Information Forensics and Security, 17:2657-2672, 2022. 1", + "[23] Ge-Peng Ji, Guobao Xiao, Yu-Cheng Chou, Deng-Ping Fan, Kai Zhao, Geng Chen, and Luc Van Gool. Video polyp segmentation: A deep learning perspective. MIR, 19(6):531-549, 2022. 2", + "[24] Ge-Peng Ji, Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Christos Sakaridis, and Luc Van Gool. Masked vision-language transformer in fashion. MIR, 20(3):421-434, 2023. 2", + "[25] Gengyun Jia, Meisong Zheng, Chuanrui Hu, Xin Ma, Yuting Xu, Luoqi Liu, Yafeng Deng, and Ran He. Inconsistency-aware wavelet dual-branch network for face forgery detection. IEEE Transactions on Biometrics, Behavior, and Identity Science, 3(3), 2021. 2", + "[26] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proc. CVPR, pages 2889-2898, 2020. 4, 7", + "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. CVPR, pages 4401-4410, 2019. 1", + "[28] Sohail Ahmed Khan and Hang Dai. Video transformer for deepfake detection with incremental learning. In Proc. ACM MM, pages 1821-1828, 2021. 3", + "[29] Aminollah Khormali and Jiann-Shiun Yuan. Dfdt: An end-to-end deepfake detection framework using vision transformer. Applied Sciences, page 2953, 2022. 3, 5, 6", + "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. 5", + "[31] Jiaming Li, Hongtao Xie, Jiahong Li, Zhongyuan Wang, and Yongdong Zhang. Frequency-aware discriminative feature learning supervised by single-center loss for face forgery detection. In Proc. CVPR, pages 6458-6467, 2021. 2, 5", + "[32] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proc. CVPR, pages 5001-5010, 2020. 2, 5, 6, 7", + "[33] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. In Proc. CVPR Workshops, pages 656-663, 2019. 6", + "[34] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proc. CVPR, pages 3207-3216, 2020. 4", + "[35] Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: Rethinking face forgery detection in frequency domain. In Proc. CVPR, pages 772-781, 2021. 5", + "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. ICCV, pages 10012-10022, 2021. 2, 3, 4, 5", + "[37] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proc. CVPR, pages 8060–8069, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22667", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] MarekKowalski. Faceswap. https://github.com/MarekKowalski/FaceSwap/. 2021-11-13. 4", + "[39] Iacopo Masi, Aditya Killekar, Royston Marian Mascarenhas, Shenoy Pratik Gurudatt, and Wael AbdAlmageed. Two-branch recurrent network for isolating deepfakes in videos. In Proc. ECCV, pages 667–684, 2020. 5", + "[40] Yisroel Mirsky and Wenke Lee. The creation and detection of deepfakes: A survey. ACM CSUR, 54(1):1-41, 2021. 1", + "[41] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In Proc. ICCV, pages 3163-3172, 2021. 5, 6", + "[42] Yuval Nirkin, Lior Wolf, Yosi Keller, and Tal Hassner. Deepfake detection based on discrepancies between faces and their context. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6111-6121, 2021. 2", + "[43] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In Proc. ECCV, pages 86-103, 2020. 2, 5", + "[44] Rameswar Panda Quanfu Fan, Richard Chen. Can an image classifier suffice for action recognition? In ICLR, 2022. 2", + "[45] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. Faceforensics: Learning to detect manipulated facial images. In Proc. ICCV, pages 1-11, 2019. 4", + "[46] Ekraam Sabir, Jiaxin Cheng, Ayush Jaiswal, Wael AbdAlmageed, Iacopo Masi, and Prem Natarajan. Recurrent convolutional strategies for face manipulation detection in videos. In Proc. CVPR Workshops, pages 80-87, 2019. 6, 7", + "[47] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. ICCV, pages 618-626, 2017. 6", + "[48] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In Proc. CVPR, pages 11254-11264, 2022. 1", + "[49] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proc. CVPR, pages 7672-7682, 2022. 1", + "[50] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, pages 6105-6114, 2019. 5", + "[51] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM TOG, 38(4):1-12, 2019. 4", + "[52] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proc. ICCV, pages 2387-2395, 2016. 4", + "[53] Luisa Verdoliva. Media forensics and deepfakes: an overview. IEEE Journal of Selected Topics in Signal Processing, 14(5):910-932, 2020. 1", + "[54] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In Proc. CVPR, pages 14923-14932, 2021. 2" + ], + "bbox": [ + 78, + 92, + 467, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Junke Wang, Zuxuan Wu, Wenhao Ouyang, Xintong Han, Jingjing Chen, Yu-Gang Jiang, and Ser-Nam Li. M2tr: Multi-modal multi-scale transformers for deepfake detection. In Proc. ICMR, pages 615-623, 2022. 2, 3, 5", + "[56] Ping Wang, Kunlin Liu, Wenbo Zhou, Hang Zhou, Honggu Liu, Weiming Zhang, and Nenghai Yu. Adt: Anti-deepfake transformer. In Proc. ICASSP, pages 2899-1903, 2022. 5", + "[57] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proc. CVPR, 2020. 6, 7", + "[58] Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv:2102.11126, 2021. 2", + "[59] Ziming Yang, Jian Liang, Yuting Xu, Xiao-Yu Zhang, and Ran He. Masked relation learning for deepfake detection. IEEE Transactions on Information Forensics and Security, pages 1696-1708, 2023. 2", + "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proc. ICCV, pages 6023-6032, 2019. 9", + "[61] Yu Zeng, Zhe Lin, and Vishal M Patel. Sketched: Mask-free local image manipulation with partial sketches. In Proc. CVPR, pages 5951-5961, 2022. 1", + "[62] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. Proc. ICLR, 2018. 9", + "[63] Yanyi Zhang, Xinyu Li, Chunhui Liu, Bing Shuai, Yi Zhu, Biagio Brattoli, Hao Chen, Ivan Marsic, and Joseph Tighe. Vidtr: Video transformer without convolutions. In Proc. ICCV, pages 13577-13587, 2021. 5, 6", + "[64] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: Interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18:1335-1348, 2023. 5, 6", + "[65] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep-fake detection. In Proc. CVPR, pages 2185–2194, 2021. 2, 5, 6", + "[66] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Self-supervised transformer for deepfake detection. arXiv:2203.01265, 2022. 2, 3", + "[67] Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In Proc. CVPR, pages 15023-15033, 2021. 2", + "[68] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In Proc. ICCV, pages 15044-15054, 2021. 2, 4, 5, 6, 7", + "[69] Xiangyu Zhu, Hao Wang, Hongyan Fei, Zhen Lei, and Stan Z Li. Face forgery detection by 3d decomposition. In Proc. CVPR, pages 2929-2939, 2021. 1", + "[70] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proc. ACM MM, pages 2382-2390, 2020. 5" + ], + "bbox": [ + 501, + 92, + 890, + 895 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22668", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/b759f55b-503f-4164-9793-b1ffe02056fb_model.json b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/b759f55b-503f-4164-9793-b1ffe02056fb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..92f807c6111d728cbab39a7b5e5ed7bae74ded3f --- /dev/null +++ b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/b759f55b-503f-4164-9793-b1ffe02056fb_model.json @@ -0,0 +1,2455 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.201, + 0.131, + 0.773, + 0.154 + ], + "angle": 0, + "content": "TALL:,Thumbsail Layout for Deepfake Video Detection" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.179, + 0.856, + 0.199 + ], + "angle": 0, + "content": "Yuting Xu\\(^{1,3*}\\), Jian Liang\\(^{2,4}\\), Gengyun Jia\\(^{5}\\), Ziming Yang\\(^{1,3}\\), Yanhao Zhang\\(^{6}\\), Ran He\\(^{2,4\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.199, + 0.76, + 0.216 + ], + "angle": 0, + "content": "\\(^{1}\\) Institute of Information Engineering, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.216, + 0.786, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\) CRIPAC & MAIS, Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.234, + 0.786, + 0.252 + ], + "angle": 0, + "content": "\\(^{3}\\) School of Cyber Security, UCAS \\(^{4}\\) School of Artificial Intelligence, UCAS" + }, + { + "type": "list", + "bbox": [ + 0.185, + 0.216, + 0.786, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.252, + 0.857, + 0.269 + ], + "angle": 0, + "content": "\\(^{5}\\) School of Communications and Information Engineering, NJUPT \\(^{6}\\) OPPO Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.273, + 0.782, + 0.287 + ], + "angle": 0, + "content": "yuting.xu@cripac.ia.ac.cn, liangjian92@gmail.com, rhe@nlpr.ia.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.474, + 0.687 + ], + "angle": 0, + "content": "The growing threats of deepfakes to society and cybersecurity have raised enormous public concerns, and increasing efforts have been devoted to this critical topic of deepfake video detection. Existing video methods achieve good performance but are computationally intensive. This paper introduces a simple yet effective strategy named Thumbnail Layout (TALL), which transforms a video clip into a pre-defined layout to realize the preservation of spatial and temporal dependencies. Specifically, consecutive frames are masked in a fixed position in each frame to improve generalization, then resized to sub-images and rearranged into a pre-defined layout as the thumbnail. TALL is model-agnostic and extremely simple by only modifying a few lines of code. Inspired by the success of vision transformers, we incorporate TALL into Swin Transformer, forming an efficient and effective method TALL-Swin. Extensive experiments on intra-dataset and cross-dataset validate the validity and superiority of TALL and SOTA TALL-Swin. TALL-Swin achieves \\(90.79\\%\\) AUC on the challenging cross-dataset task, FaceForensics++ \\(\\rightarrow\\) CelebDF. The code is available at https://github.com/rainy-xu/TALL4Deepfake." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Deepfakes generate and manipulate facial appearances to deceive viewers through generation techniques [61, 48]. With the remarkable success of generative adversarial networks [14, 27], deepfake products have become photorealistic that humans can not distinguish. These deepfake products [21, 49] may be misused for malicious purposes, leading to severe trust issues and security problems, such as financial fraud, identity theft, and celebrity impersonation." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.325, + 0.883, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.548, + 0.895, + 0.619 + ], + "angle": 0, + "content": "Figure 1. The AUC and FLOPs trade-off of different backbones. Image-level backbones with TALL enjoy comparable accuracy-cost trade-offs with the 3DCNN and video transformer family on the unseen Celeb-DF dataset. All models with the same setting are trained on the FF++ (HQ) dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.894, + 0.704 + ], + "angle": 0, + "content": "ation [53, 40]. The rapid development of social media accelerates the abuse of deepfakes. Therefore, it is crucial to develop advanced detection methods to protect the data privacy of individual users." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Most previous image-based methods [22, 69] perform well on intra-dataset, but their generalizability needs to be improved. Recent research has focused on video-based methods to detect deepfake by modeling spatio-temporal dependencies. There are subtle spatio-temporal inconsistencies between frames since the deepfake algorithms are executed frame by frame. The core of video-level approaches for deepfake detection is capturing inconsistencies through temporal modeling. Existing deepfake video detection methods generally follow two directions. Some methods [15, 16] use two-branch networks or modules to learn spatial and temporal information separately and then fuse them. However, these two-branch approaches may frag" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.402, + 0.888 + ], + "angle": 0, + "content": "*This work was done when she was a student in CRIPAC." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.402, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22658" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "ment spatiotemporal cooperation and lead to subtle artifacts being neglected. Others directly use classic temporal models such as LSTM and 3D-CNNs. These methods are computationally intensive. The current rise of transformers for vision task backbones has prompted the emergence of corresponding deepfake detection methods. They are accompanied by significant computational complexity that makes them challenging to deploy and use, despite breakthroughs in performance. To enjoy benefits from both image and video methods, we are curious to see whether it is possible to append information about the temporal dimension to the image dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.517 + ], + "angle": 0, + "content": "This work develops a simple yet effective Thornail Layout (TALL) for deepfake detection by spatio-temporal modeling. TALL is computationally cheap and retains both temporal and spatial information. In detail, we use dense sampling to extract multiple clips in the video and then randomly select four consecutive frames in the video segment. Subsequently, a block is masked at a fixed position in each frame. Finally, the frames are resized as sub-image and sequentially rearranged into a pre-defined layout as a thumbnail, which has the same size as the clip frames. As shown in Figure 1, TALL brings two advantages compared to the previous spatio-temporal modeling methods for deepfake detection: (1) TALL contains local and global contextual deepfake patterns. (2) TALL is a model-agnostic method for spatio-temporal modeling deepfake patterns at zero computation and zero parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.473, + 0.671 + ], + "angle": 0, + "content": "Furthermore, we discover that the better temporal modeling capabilities backbone has, the better performance TALL achieves. Based on the proposed TALL, we complement a baseline for video deepfake detection based on Swin Transformer [36], called TALL-Swin. We validate TALL-Swin on four popular benchmark datasets, including FaceForensics++, Celeb-DF, DFDC, and DeeperForensics. Our method gains a remarkable improvement over the state-of-the-art approaches. The main contributions of our paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.685, + 0.47, + 0.76 + ], + "angle": 0, + "content": "- We provide a new perspective for an efficient strategy for video deepfake detection called Thumbnail Layout (TALL), which incorporates both spatial-temporal dependencies, and allows the model to capture spatial-temporal inconsistencies." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.778, + 0.47, + 0.824 + ], + "angle": 0, + "content": "- We propose a spatio-temporal modeling method called TALL-Swin, which efficiently captures the inconsistencies between deepfake video frames." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the validity of our proposed TALL and TALL-Swin. TALL-Swin outperforms previous methods in both intra-dataset and cross-dataset scenarios." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.685, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.643, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.787, + 0.132 + ], + "angle": 0, + "content": "2.1. Image-Level Deepfake Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.895, + 0.457 + ], + "angle": 0, + "content": "Typically, existing deepfake detection methods fall into two categories: image-level and video-level methods. The image-level methods [25, 13] always exploit the artifacts of deepfake images in the spatial domain, such as discrepancies between local regions [42, 59], grid-like structure in frequency space [10], and differences in global texture statistics [37] that provide specific clues to distinguish deepfakes from the real images. F3Net [43] and FDFL [31] utilize the same pipeline that utilizes frequency-aware features and RGB information to capture the traces in different input spaces separately. RFM [54] and Multi-att [65] propose an attention-guided data augmentation mechanism to guide detectors to discover undetectable deepfake clues. Face X-ray [32] and PCL [67] provide effective ways to outline the boundary of the forged face for detecting deepfakes. ICT [11] exploits an identity extraction module to detect identity inconsistency in the suspect image. Similarly, M2tr [55] detects local inconsistencies within frames at different spatial levels. Generally, image-level methods suffer over-fitting issues when a specific technique manipulates the images, and they ignore temporal information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.783, + 0.483 + ], + "angle": 0, + "content": "2.2. Video-Level Deepfake Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.895, + 0.809 + ], + "angle": 0, + "content": "To improve the generalization of deepfake detectors, many studies generate diversity and generic deepfake data, while other studies capture the temporal incoherence of fake videos as generic clues. Some recent works propose detecting temporal inconsistency using well-designed spatiotemporal neural networks, and others [15, 16] attempt to add modules to image models that capture temporal information. STIL [15] formulates deepfake video detection as a spatial and temporal inconsistency learning process and integrates both spatial and temporal features in a unified 2D CNN framework. FTCN [68] detects temporal-related artifacts instead of spatial artifacts to promote generalization. LipForensics [18] is proposed to learn high-level semantic irregularity in mouth movement in the generated video. RealForensics [17] uses auxiliary data sets during training in exchange for generalization at the cost of higher computational demands. The video-based methods achieve strong generalization but suffer from large computational overhead. To reduce computational costs, we propose TALL which gathers consecutive video frames into thumbnails for learning spatio-temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.88, + 0.833 + ], + "angle": 0, + "content": "2.3. Deepfake Detection with Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recently, ViT [12] has achieved impressive performance in computer vision tasks [24, 23, 44]. Many studies extend the ViT for deepfake detection [66, 58]. These methods achieve better performance compared to CNN-based" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22659" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "models, but also sacrifice computational efficiency. A few works [55, 66] attempt to extend the transformer for deepfake detection due to the advent of the visual transformer (ViT) and the impressive ability to model long-range data, different from two-branch architectures that capture short-range and long-range temporal inconsistencies with a single-branch model. ICT [11] aims to detect identity consistency in deepfake video but may fail in detecting face reenactment and entire face synthesis results. DFLL [28] extract the UV texture map to help the transformer to detect deepfakes, which may disrupt the continuity between video frames. DFTD [29] leverages ViT to consider both global and local information but ignores the problem of excessive model arithmetic requirements. Although the transformer-based approaches achieve promising performance, they are accompanied by significant computational complexity that makes them challenging to deploy and use, and the long-range dependencies may be insufficiently exploited in detection models. Swin Transformer [36] produces a hierarchical feature representation and has linear computational complexity concerning input image size, which is suitable as a general-purpose backbone for various vision tasks. In this paper, we cooperate with Swin Transformer to form our robust and efficient method TALL-Swin." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.467, + 0.172, + 0.483 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.629 + ], + "angle": 0, + "content": "TALL is a deepfake video detection strategy that transforms a video clip into an all-in-one thumbnail without the extra computational overhead. In the following sections, we begin with the motivation of TALL for deepfake detection in Section 3.1. Then we present the technical details of the TALL in Section 3.2. Finally, a generalizable SwINTALL baseline is introduced to explore subtle artifacts in Section 3.3. More details of the layout design are presented in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.636, + 0.2, + 0.65 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.885 + ], + "angle": 0, + "content": "While recent studies have attempted to address noticeable flaws through techniques like slight motion blurring and temporal consistency loss, subtle spatio-temporal artifacts still remain. These artifacts are important for detecting deepfakes, but they introduce two problems: 1) video-based models are less efficient, and 2) analyzing information over long distances may overlook local artifacts, which are critical for deepfake detection. To address these challenges, we propose the TALL strategy, which naturally incorporates temporal information into image-level tasks without disrupting spatial information. This approach enables the image-level model to detect deepfakes in videos. Furthermore, we discovered that TALL provides even greater performance gains when combined with a powerful spatial model, resulting in the TALL-Swin." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In detail, TALL arranges consecutive frames in the" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.088, + 0.71, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.246, + 0.684, + 0.259 + ], + "angle": 0, + "content": "(a) Thumbnail Layout (TALL)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.092, + 0.886, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.246, + 0.86, + 0.259 + ], + "angle": 0, + "content": "(b) Feature Map" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.267, + 0.892, + 0.295 + ], + "angle": 0, + "content": "Figure 2. Illustration of the TALL and shifted window process for computing self-attention in the TALL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.893, + 0.668 + ], + "angle": 0, + "content": "temporal order in a compact \\(2 \\times 2\\) layout, in line with the calculation theory of convolution and shifted window. TALL contains both spatial and temporal information so that model can learn both intra-frame artifacts and interframe inconsistency and obtains comparable performance to video-based methods. Here we use the shifted window to explain TALL's mechanism. As illustrated in Figure 2 (a), the model computes self-attention while accounting for spatial dependencies across sub-images (represented by the solid red box). When the window spans multiple sub-images (represented by the red dash box), the model is able to capture temporal inconsistencies between frames. Moreover, TALL leverages both local and global contexts of deepfake patterns to ensure robust modeling capabilities for short and long-range spatial dependencies. Compared to previous methods, we anticipate that TALL strikes a balance between speed and accuracy, sacrificing a little spatial information while preserving performance. Based on the fact that attention-based models are better at handling contextual features and that the Swin-Transformer uses shifted windows to reduce computation and memory, we further complement TALL-Swin baseline for video deepfake detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.746, + 0.696 + ], + "angle": 0, + "content": "3.2. Thumbnail Layout (TALL)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given a video \\( V \\in \\mathbb{R}^{T \\times C \\times H \\times W} \\), where \\( T \\) is the frame length of the video, \\( C \\) is the number of channels, and \\( H \\times W \\) is the resolution of the frames. Assuming each video contains \\( N \\) clips, we divide a video into \\( N \\) equal segments of length \\( T / N \\) and then sample consecutive \\( t \\) (set to 4 by default) frames from the segments at random locations to form one clip. Then, the thumbnail \\( I \\) is rearranged of sub-images \\( (C \\times \\frac{H}{\\sqrt{t}} \\times \\frac{W}{\\sqrt{t}}) \\) that are resized from the above \\( t \\) frames. To maximize the utility of TALL, we mask the organized \\( N \\) square masks of the thumbnail. It is based on two core designs: 1) The position of the masks is random between different sub-images, which retains the advantages of the Cutout [8] that encourages the network to focus more" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22660" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.079, + 0.091, + 0.466, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode of TALL in a PyTorch-like style." + }, + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.117, + 0.452, + 0.425 + ], + "angle": 0, + "content": "# x: one clip of video (T*C*H*W) \n# T: frame number of clip \n# C: channels; s: mask size \n# d: length of frame included in the thumbnail \n# r: rows of thumbnail \n# x_tall: thumbnail image (224*224) \n#TALL's augmentation strategy \nh = np.random.randint(H) \nw = np.random.randint(W) \n#the mask position is fixed for each frame \nm = np.ones((H, W)) \nh1 = np.clip(h - s // 2, 0, H) \nh2 = np.clip(+s // 2, 0, H) \nw1 = np.clip(w - s // 2, 0, W) \nw2 = np.clip(w + s // 2, 0, W) \nm[h1: h2, w1: w2] = 0 \nm = torch.from_numpy(m) \nm = maskexpand_as(x) \nx = x * m \n#TALL: generation of the thumbnail \nx = x.view(-1, H, W).unsqueeze(0) \nx = x.view((-1, C*d) + x_tall.size() [2:]) \nx = rearrange(x, 'b (th tw c) h w -> b c (th h) (tw w)', th=r, c=C) \nx_tall = interpolate(x, size=H)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.581 + ], + "angle": 0, + "content": "on complementary and less prominent features. 2) We fix the position of the mask within a clip to take advantage of the fact that most deepfake videos are frame-by-frame tampered with, thus forcing the model to detect inconsistencies between adjacent frames of the deepfake videos. We do not allow the mask to appear on the seams of the thumbnail but allow for partial mask inclusion in the thumbnail. The detailed procedure of TALL is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.206, + 0.605 + ], + "angle": 0, + "content": "3.3. TALL-Swin" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To balance efficiency and model performance for spatiotemporal feature learning and to leverage the benefits of attention-based models, we enhanced a baseline deepfake detection model called TALL-Swin by incorporating the Swin Transformer [36]. Given the characteristics of TALL, we slightly modified the window size of Swin-B in TALL-Swin. We first enlarge the window size of the first three stages of the model so that the interaction between frames in the thumbnail becomes more frequent, forcing the model to learn more detailed spatio-temporal dependencies. Next, we set the window size of the last stage to be the same as the feature map size, enabling the window to perform global attention computations while TALL-Swin captures global spatial-temporal dependencies. As a result, the size of the last layer of the feature map became smaller, reducing the window size without introducing any additional computational overhead. Consequently, the window sizes for the four stages of TALL-Swin are [14, 14, 14, 7]. Note that the patch merging process makes TALL-Swin captures a more" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "comprehensive range of dependencies through hierarchical representations, as shown in Figure 2 (b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.892, + 0.228 + ], + "angle": 0, + "content": "Given a video of length \\( T \\), each frame contains \\( N \\) patches, and the window contains \\( P \\) patches. To demonstrate the superiority of TALL-Swin in terms of computational consumption, we show below the computational complexity of the image-level transformer and video-level transformer, including ViT [12], Swin [36], ViViT [2], and TALL-Swin respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.238, + 0.77, + 0.254 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {\\mathrm {V i T}} = 4 T N C ^ {2} + 2 T N ^ {2} C,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.256, + 0.892, + 0.28 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {\\mathrm {S w i n}} = 4 T N C ^ {2} + 2 T P N C, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.274, + 0.783, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {\\mathrm {V i V i T}} = 4 T N C + 2 T ^ {2} N ^ {2} C,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.294, + 0.818, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {\\mathrm {T A L L - S w i n}} = T N C ^ {2} + \\frac {1}{2} T P N C.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.892, + 0.395 + ], + "angle": 0, + "content": "TALL-Swin has the lowest computational complexity compared to image and video-level transformer methods. Subsequent experiments will demonstrate that TALL-Swin maintains performance, albeit at the sacrifice of some spatial information." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.397, + 0.892, + 0.426 + ], + "angle": 0, + "content": "The cross-entropy loss is employed to optimize the TALL-Swin, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.434, + 0.892, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C E} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} y _ {i} \\log \\mathcal {F} \\left(x _ {i}\\right) + (1 - y _ {i}) (\\log (1 - \\mathcal {F} \\left(x _ {i}\\right)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.477, + 0.892, + 0.508 + ], + "angle": 0, + "content": "where \\( x_{i} \\) indicates input clip, \\( y_{i} \\) denotes the label of clip, \\( n \\) is the number of clip, \\( \\mathcal{F} \\) is TALL-Swin." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.521, + 0.633, + 0.538 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.582, + 0.562 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Datasets. Following previous works [18, 17, 68], we evaluate the TALL and TALL-Swin on four widely used datasets. FaceForensics++ [45] is a most-used benchmark on intra-dataset deepfake detection, consisting of 1,000 real videos and 4,000 fake videos in four different manipulations: DeepFake [7], FaceSwap [38], Face2Face [52], and NeuralTextures [51]. Besides, FaceForensics++ contains multiple video qualities, e.g. high quality (HQ), low quality (LQ) and RAW. Celeb-DF (CDF) [34] is a popular benchmark on cross-dataset, which contains 5,693 deepfake videos generated from celebrities. The improved compositing process was used to improve the various visual artifacts presented in the video. Celeb-DF is also suitable for deepfake detection tasks with a reference set. DFDC [9] is a large-scale benchmark developed for Deepfake Detection Challenge. This dataset includes 124k videos from 3,426 paid actors. The existing deepfake detection methods do perform not very well on DFDC due to their sophisticated deepfake techniques. DeeperForensics (DFO) [26] includes 60,000 videos with 17.6 million frames for deepfake detection, whose videos vary in identity, pose, expression, emotion, lighting conditions, and blend shape with high quality." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22661" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.462, + 0.366 + ], + "angle": 0, + "content": "
ModelsTemp.CDFDFDCFLOPsParamsPT
I3D-RGB* [4]✓78.2465.58222.7G25M1K
R3D-50* [19]✓79.6367.73296.6G46M1K
ResNet50* [20]×76.3864.0125.5G21M1K
+TALL✓80.9065.5425.5G21M1K
EffNetB4* [50]×78.1966.818.3G19M1K
+TALL✓83.3767.158.3G19M1K
VTN [41]✓83.2073.50296.6G46M21K
VidTR [63]✓83.3073.30117G93M21K
ViViT* [2]✓86.9674.61628G310M21K
ISTVT [64]✓84.1074.20455.8G--
ViT-B* [12]×82.3372.6455.4G84M21K
+TALL✓86.5874.1055.4G84M21K
Swin-B* [36]×83.1373.0147.5G86M21K
TALL-Swin✓90.7976.7847.5G86M21K
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.37, + 0.47, + 0.482 + ], + "angle": 0, + "content": "Table 1. Performance of different backbones. TALL consistently improves the accuracy over different image-level models. We show the AUC, FLOPs, and number of parameters for each model on the cross-dataset scenario. All models are trained on \\(\\mathrm{FF} + +\\) (HQ). \\(\\checkmark\\) indicates the model enables temporal modeling. \\* indicates our implementation. PT indicates pre-train. 1K and 21K indicate the model pre-trained on ImageNet-1K and 21K respectively. The best results are bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Implementation Details. We use MTCNN to detect face for each frame in the deepfake videos, only extract the maximum area bounding box and add \\(30\\%\\) face crop size from each side as in LipForensics [18]. The ImageNet-21K pretrained Swin-B model is used as our backbone. Excluding ablation experiments, we sample 8 clips using dense sampling, each clip contains 4 frames. The size of the thumbnail is \\(224 \\times 224\\). Following Swin Transformer [36], Adam [30] optimization is used with a learning rate of 1.5e-5 and batch size of 4, using a cosine decay learning rate scheduler and 10 epochs of linear warm-up. We adopt Acc. (accuracy) and AUC (Area Under Receiver Operating Characteristic Curve) as the evaluation metrics for extensive experiments. To ensure a fair comparison, we calculate video-level predictions for the image-based method and average the predictions across the entire video (following previous works [18, 16, 35, 68]). Note that results are directly cited from published papers if we follow the same setting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.295, + 0.803 + ], + "angle": 0, + "content": "4.2. Scaling over Backbones" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To verify our assumption, we adopt several image-level backbones commonly used for deepfake detection for comparison with the video-level backbones. As shown in Table 1 above the double horizontal line, we first compare the accuracy and complexity of the CNN-based video and image backbones. Although I3D [4] and R3D [19] achieve" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "better performance than vanilla ResNet50 [20] and EfficientNet [50], the computation costs are huge, such as R3D-50 with 296G FLOPs. For ResNet and EfficientNet who added TALL, ResNet achieves better AUC both on CDF (76.38 VS 80.93) and DFDC (64.01 VS 65.54) datasets, and EfficientNet achieves \\(5.18\\%\\) better AUC on CDF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.19, + 0.893, + 0.462 + ], + "angle": 0, + "content": "The second section contains the video and image transformers. Compared to video transformers, the image-based ViT and Swin fail to achieve better performance due to the lack of temporal modeling. For example, ViViT achieves \\(86.96\\%\\) AUC on CDF, which is \\(3.6\\%\\) higher than Swin although ViViT with \\(13\\times\\) more computation. By way of contrast, \\(\\mathrm{ViT + TALL}\\) achieves \\(86.58\\%\\) AUC on CDF with 55.4G FLOPs, which is comparable to AUC with ViViT but with low computation. Accordingly, Swin's performance was significantly improved with the addition of TALL without computation increment. On the other hand, TALL boosts higher performance on models with learned long-range dependencies. e.g., ResNet+TALL \\((+4.5\\%)\\) on CDF and \\(+1.5\\%\\) on DFDC) vs. Swin+TALL \\((+7.6\\%)\\) on CDF and \\(+3.6\\%\\) on DFDC). These two section results demonstrate that TALL provides both spatial and temporal information and enables the model to learn spatial and temporal inconsistencies for video deepfake detection." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.49, + 0.889, + 0.83 + ], + "angle": 0, + "content": "
MethodsFF++ (HQ)FF++ (LQ)
Acc.AUCAcc.AUC
MesoNet [1]83.10-70.47-
Xception [6]95.7396.3086.8689.30
Face X-ray [32]-87.35-61.60
Two-branch [39]96.4398.7086.3486.59
Add-Net [70]96.7897.7487.5091.01
F3-Net [43]97.5298.1090.4390.43
FDFL [31]96.6999.3089.0092.40
Multi-Att [65]97.6099.2988.6990.40
RECCE [3]97.0699.3291.0395.02
LipForensics [18]98.8099.7094.2098.10
DFDT [29]98.1899.2692.6794.48
ADT [56]92.0596.3081.4882.52
ST-M2TR [55]-99.42-95.31
VTN [41]98.47-94.02-
VidTR [63]97.42-92.12-
ViViT* [2]92.60-88.02-
ISTVT [64]99.00-96.15-
TALL-Swin98.6599.8792.8294.57
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.498, + 0.834, + 0.892, + 0.876 + ], + "angle": 0, + "content": "Table 2. Intra-dataset evaluations. We report the video-level Acc. \\((\\%)\\) and AUC \\((\\%)\\) on the FF++ dataset. HQ indicates high quality, and LQ indicates low quality." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22662" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.442, + 0.108 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.473, + 0.342 + ], + "angle": 0, + "content": "Intra-dataset evaluations. Following ISTVT [64], we show the results of the \\(\\mathrm{FF}++\\) dataset under both Low Quality (LQ) and High Quality (HQ) videos, and report comparisons against several advanced methods in Table 2. We can observe that advanced video-based transformers have better results than CNN-based methods. Compared to video-based transformer methods, TALL-Swin has comparable performance and lower consumption to the previous video transformer method with HQ settings. However, TALL-Swin gets unsatisfactory results with the LQ setting. The LQ setting is obtained by severely compressing the videos. So the reason for the result may be that TALL scales the frame to a smaller size, causing more spatial information to be lost in the frame. We will investigate the possibility of other designs to further improve performance in the LQ setting." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.355, + 0.462, + 0.643 + ], + "angle": 0, + "content": "
MethodCDFDFDCFShDFoAvg.
Xception [6]73.7070.9072.0084.5075.28
CNN-aug [57]75.6072.1065.7074.4071.95
CNN-GRU [46]69.8068.9080.8074.1073.40
Patch-based [5]69.6065.6057.8081.8068.70
Face X-Ray [32]79.5065.5092.8086.8081.15
Multi-Att [65]75.7068.1066.0077.7071.88
DSP-FWA [33]69.5067.3065.5050.2063.13
LipForensics [18]82.4073.5097.1097.6087.65
FTCN [68]86.9074.0098.8098.8089.63
RealForensics [17]86.9075.9099.7099.3090.45
DFDT [29]88.3076.1097.8096.9089.70
VTN [41]83.2073.5098.7097.7088.30
VidTR [63]83.5073.3098.0097.9088.10
ViViT* [2]86.9674.6199.4199.1990.05
ISTVT [64]84.1074.2099.3098.6089.10
TALL-Swin90.7976.7899.6799.6291.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.647, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Table 3. Generalization to unseen datasets. We report the video-level AUC (\\%) on four unseen datasets: Celeb-DF (CDF), DFDC, FaceShifter (FSh), and DeeperForensics (DFo)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Generalization to unseen datasets. In addition to the intra-dataset comparisons, we also investigate the generalization ability of our method. Adhering to the deepfake video detection cross-dataset protocol [18], we train a model on \\(\\mathrm{FF}++\\) (HQ) then test on Celeb-DF (CDF), DFDC, FaceShifter (FSh), and DeeperForensics (DFo) datasets. As shown in Table 3: (1) Video-based methods generally have better results than image-based methods, which shows that temporal information is helpful for the deepfake video detection task. For example, Lip outperforms Face X-ray's AUC by a wide margin. In addition, most transformer-based models have higher performance than CNN-based models. For the transformer-based models, both achieved" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.212 + ], + "angle": 0, + "content": "an average AUC of \\(88\\%\\), while the best CNN-based video-level models only achieved \\(87\\%\\). (2) TALL-Swin achieves state-of-the-art results on Celeb-DF, DFDC, and Deeper-Forensics datasets, and also beats its competitors on Celeb-DF dataset by a large margin \\((3.8\\%)\\). The results demonstrate that TALL-Swin performs well when encountering unseen datasets with better generalization ability than previous video transformer methods." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.223, + 0.889, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Figure 3. Saliency map visualization of TALL-Swin on different datasets. The first four rows of samples are from the FF++ dataset, and the last four rows are from the unseen datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Analysis of saliency map visualization. We adopt Grad-CAM [47] to visualize where the TALL-Swin is paying its attention to the deepfake faces. In Figure 3, we give the results on intra-dataset and cross-dataset scenarios. All models are trained on \\(\\mathrm{FF} + +\\) (HQ). It can be observed in the first four rows of Figure 3 that TALL-Swin captures method-specific artifacts. Note that the DF transfers the face region from a source video to a target, and the NT only modifies the facial expressions corresponding to the mouth region. TALL-Swin corresponds to focus on the face region and the mouth region. Furthermore, our model traces the more generalized artifacts that are independent of manipulation methods, e.g., blending boundaries (CDF), and abnormal motions in the clip (DFDC, Fsh, Dfo)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22663" + } + ], + [ + { + "type": "image_footnote", + "bbox": [ + 0.179, + 0.093, + 0.82, + 0.106 + ], + "angle": 0, + "content": "Xception Face X-ray ICT Patch-based CNN-GRU LipForensics TALL-Swin" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.114, + 0.297, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.114, + 0.495, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.114, + 0.692, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.114, + 0.89, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.271, + 0.297, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.271, + 0.495, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.271, + 0.692, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.271, + 0.89, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.443, + 0.893, + 0.486 + ], + "angle": 0, + "content": "Figure 4. Robustness to various unseen corruptions. We report the video-level AUC \\((\\%)\\) of our methods under five different levels of seven particular types of corruption. \"Average\" denotes the mean across all corruptions at each severity level. Our TALL-Swin is more robust than previous methods for all corruptions." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.497, + 0.885, + 0.678 + ], + "angle": 0, + "content": "
MethodCleanSaturationContrastBlockNoiseBlurPixelCompressAvg.
Xception (ICCV'19) [6]99.899.398.699.753.860.274.262.178.3
CNN-GRU (CVPRW'19) [46]99.999.098.897.947.971.586.574.582.3
CNN-aug (CVPR'20) [57]99.899.399.195.254.776.591.272.584.1
Patch-based (ECCV'20) [5]99.984.374.299.250.054.456.753.467.5
Face X-ray (CVPR'20) [32]99.897.688.599.149.863.888.655.277.5
LipForensics (ICCV'21) [18]99.999.999.687.473.896.195.695.692.5
FTCN (ICCV'21) [68]99.499.496.797.153.195.898.286.489.5
RealForensics (CVPR'22) [17]99.899.899.698.979.795.398.497.695.6
TALL-Swin w/o mask100.0100.0100.099.883.597.398.497.996.7
TALL-Swin100.0100.0100.0100.085.397.698.598.197.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.681, + 0.893, + 0.711 + ], + "angle": 0, + "content": "Table 4. Average robustness to unseen corruptions. Average Video-level AUC (%) across five intensity levels for each corruption type proposed in DFo [26]. \"Avg\" indicates the mean across all corruptions and all levels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Robustness to unseen perturbations. Deepfake detectors must be robust to common perturbations, given that video propagation on social media causes video compression, noise addition, etc. We also study the performance of robustness to unseen perturbations. Following RealForensics [17], the experiment applies seven unseen perturbations to fake videos at five intensity levels. In Figure 4, we show results of increasing the severity of each corruption. We can observe that other methods degrade dramatically as the perturbations become more severe. TALL-Swin still has a high performance. However, TALL-Swin degrades when" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.887 + ], + "angle": 0, + "content": "the Gaussian noise reaches level five. Table 4 presents the average AUC across all intensity levels for corruption types. We observe that our method is significantly more robust to most perturbations than other methods. The good robustness may be from both the design of TALL and the proposed mask augmentation. The main reason may be the consecutive multi-frame input. We empirically consider that the key to deepfake detection is local inconsistency, the continuous frame design has less redundant information, ensuring that the model finds locally important clues." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22664" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.089, + 0.153, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.172, + 0.147, + 0.185 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.128, + 0.278, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.108, + 0.353, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.108, + 0.428, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.195, + 0.424, + 0.21 + ], + "angle": 0, + "content": "Figure 5. Illustration of different layout designs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.235, + 0.23, + 0.251 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.471, + 0.317 + ], + "angle": 0, + "content": "We perform the ablation study to analyze the effects of each component and hyper-parameter in TALL-Swin. All experiments are trained on \\(\\mathrm{FF}++\\) (HQ) and tested on the CDF and DFDC datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.47, + 0.469 + ], + "angle": 0, + "content": "Effects of different layouts. We train a TALL-Swin model on \\(\\mathrm{FF}++(\\mathrm{HQ})\\) for each layout illustrated in Figure 5, to analyze in which layout of the thumbnails the model learns the strongest generalization of the spatial-temporal dependence of the deepfake patterns. As shown in Table 5, the model with a compact layout like Figure 5 (d) has good generalization ability on the unseen datasets. A compact layout like Figure 5 (d) may help the model to learn the temporal dependence across frames because such a form provides the shortest distance between any two images." + }, + { + "type": "table", + "bbox": [ + 0.17, + 0.481, + 0.376, + 0.568 + ], + "angle": 0, + "content": "
LayoutCDFDFDC
Figure 5 (a)85.5270.02
Figure 5 (b)84.9373.57
Figure 5 (c)86.6672.12
Figure 5 (d)87.6074.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.571, + 0.47, + 0.599 + ], + "angle": 0, + "content": "Table 5. Effects of different layouts. All models here are trained without mask augmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.885 + ], + "angle": 0, + "content": "Study on the numbers of sub-image. We use Swin-B as the baseline for this study to compare the effect of different thumbnail layout schemes on the model's generalization ability. Changing frames to thumbnails involves scaling, so we also investigate the impact of resizing and random cropping pre-processing on model performance. We set up four variants: resizing pre-process with \\(4 \\times 4\\) layout, \\(3 \\times 3\\) layout and \\(2 \\times 2\\) layout; random cropping pre-process with \\(2 \\times 2\\) layout. As shown in Table 6, the model performance degrades sharply when using \\(4 \\times 4\\) layout. This may be due to the small size of each sub-image that the spatial information is not captured well by the model. The result of \\(3 \\times 3\\) layout also slightly decreases. \\(2 \\times 2\\) layout with resizing pre-processing beats \\(2 \\times 2\\) layout with random crop. We also found that TALL-Swin achieves the best performance and the AUC score increases \\(3.2\\%\\) compared to the baseline, suggesting that thumbnails in a \\(2 \\times 2\\) layout are more helpful to TALL-Swin than original frames." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Effects of Sub-image's size. We eliminate the scaling" + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.089, + 0.835, + 0.196 + ], + "angle": 0, + "content": "
Pre-processLayoutCDFDFDC
None-83.1373.01
Resize4 × 480.1870.45
Resize3 × 383.1872.98
Crop2 × 278.5573.30
Resize2 × 287.6074.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.2, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Table 6. Study on the numbers of sub-image. All models here are trained without mask augmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.894, + 0.39 + ], + "angle": 0, + "content": "operation for sub-images to allow for more flexible layout settings. However, we've observed that when the number of sub-images grows at their original size, the performance improvements are only slight. Additionally, the computational complexity increases dramatically with the number of frames ( 4.3 times more than the TALL setting), as demonstrated in Table 7. To strike a balance between performance and computational complexity, we reduce the resolution of sub-images in TALL." + }, + { + "type": "table", + "bbox": [ + 0.509, + 0.402, + 0.88, + 0.48 + ], + "angle": 0, + "content": "
Subimage-sizeLayoutFLOPsCDFDFDC
224 × 2243 × 3253G88.6975.98
224 × 2242 × 2185G88.1575.01
112 × 1122 × 247.5G87.6074.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.483, + 0.892, + 0.512 + ], + "angle": 0, + "content": "Table 7. Effects of Sub-image's size. All models here are trained without mask augmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Study on absence and order of thumbnails. In this case, we study the impact of missing the last sub-image and the last two sub-images on the model's performance. The first two rows of Table 8 show that all four sub-images contribute to the model performance. Besides, we set the order of the different thumbnails to evaluate the TALL-Swin. We consider three orders: forward, reverse, and random. Forward order performs the best for three different orders. This may be because of the positional encoding of different frames in TALL-Swin." + }, + { + "type": "table", + "bbox": [ + 0.574, + 0.689, + 0.812, + 0.791 + ], + "angle": 0, + "content": "
VariantsCDFDFDC
0, 1, 2, -86.4669.51
0, 1, -, -84.2269.09
Random85.8570.30
Reverse86.6572.37
Forward87.6074.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.795, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Table 8. Ablation study of absence and order of thumbnails. All models here are trained without mask augmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effects of different orders on other backbones. In order to prove that the phenomenon is not incidental, we also conduct experiments on ResNet50 and EfficientNet for three different orders as shown in Table 9. As expected, for-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22665" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "ward order outperforms reverse and random orders both on ResNet50 and EfficientNet, which indicates that TALL can learn the temporal dependency." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.149, + 0.406, + 0.308 + ], + "angle": 0, + "content": "
VariantsCDFDFDC
ResNet50+TALL76.3864.01
Random78.1464.12
Reverse78.5464.87
Forward80.9065.54
EfficientNet+TALL78.1966.81
Random81.0166.13
Reverse81.6666.69
Forward83.3767.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.312, + 0.439, + 0.325 + ], + "angle": 0, + "content": "Table 9. Ablation studies of different orders of TALL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.471, + 0.61 + ], + "angle": 0, + "content": "Effectiveness of mask strategy. In this work, TALL-Swin is trained on the \\(\\mathrm{FF}++\\) (HQ) dataset without any data enhancement as the baseline except for Multi-scale Crop and Random Horizontal Flip. To validate the effectiveness of the mask strategy, we compare our default baseline with different data augmentation strategies: 1) The Cutout [8] on one sub-image; 2) The Cutout on four sub-images. 3) The combination of Mixup [62] and Cutmix [60] on four sub-images, as shown in Table 10. The performance of a random Cutout [8] on four sub-images is better than on one sub-image. Besides, the mask strategy leads to better performance than the well-known Cutout \\((1.46\\%)\\). This supports our hypothesis that strategy encourages models to learn subtle temporal-spatial variations and improves model generalization ability. Further, our augmentation strategy exceeds \\(1.02\\%\\) than the combination of Mixup and Cutmix, demonstrating the augmentation's effectiveness in TALL for video detection." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.62, + 0.411, + 0.716 + ], + "angle": 0, + "content": "
AugmentationCountCDFDFDC
None-87.6074.32
Cutout189.0674.07
Cutout489.3375.22
Mixup+CutMix489.7575.33
TALL's mask490.7976.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.72, + 0.468, + 0.747 + ], + "angle": 0, + "content": "Table 10. Study of the augmentation strategy in TALL. The count column represents the number of blocks on the thumbnail." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Study on window size. We study the effect of window size on model performance and computational cost. The results are shown in Table 11. Our window expansion for the first three phases will increase the model performance by \\(1.74\\%\\) AUC. The results in the second and third rows show that the first three stages of the window getting the largest would not give a boost to the model. Our analysis of a too-large window may lead to a weakening of the model's ability to learn local information in the sub-image." + }, + { + "type": "table", + "bbox": [ + 0.582, + 0.088, + 0.808, + 0.153 + ], + "angle": 0, + "content": "
Window sizeCDFDFDC
(7,7,7,7)85.6073.32
(14,14,14,7)87.6074.32
(28,28,28,7)86.6574.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.157, + 0.892, + 0.185 + ], + "angle": 0, + "content": "Table 11. Ablation studies of window size. All models here are trained without mask augmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.209, + 0.619, + 0.225 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.235, + 0.892, + 0.402 + ], + "angle": 0, + "content": "This paper presents a novel perspective on detecting deepfake videos using TALL. TALL is both simple and effective, enabling joint spatio-temporal modeling without any additional costs. TALL representation reveals normal deepfake patterns with local-global contextual features. We further propose a new baseline for deepfake video detection called TALL-Swin, which efficiently captures the inconsistencies between deepfake video frames. Extensive experiments demonstrate that TALL-Swin achieves promising results for various unseen deepfake types and strong robustness to a wide range of common corruptions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.417, + 0.651, + 0.433 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.533 + ], + "angle": 0, + "content": "This work was partially funded by National Natural Science Foundation of China under Grants (62276256, U21B2045 and U20A20223) and Beijing Nova Program under Grant Z211100002121108. The authors wish to thank Huaibo Huang and Lijun Sheng in no particular order, for insightful discussions." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.547, + 0.597, + 0.563 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.572, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In Proc. WIFS, pages 1-7, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.616, + 0.894, + 0.67 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proc. ICCV, pages 6836-6846, 2021. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.673, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[3] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proc. CVPR, pages 4113-4122, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.73, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[4] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proc. CVPR, pages 6299-6308, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.773, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[5] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Proc. ECCV, pages 103-120, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.83, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proc. CVPR, pages 1251-1258, 2017. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[7] deepfakes. Deepfakes. https://github.com/deepfakes/faceswap.2021-11-13.4" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.572, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22666" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[8] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv:1708.04552, 2017. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.137, + 0.468, + 0.192 + ], + "angle": 0, + "content": "[9] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv:2006.07397, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.195, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[10] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In Proc. CVPR, pages 7865-7874, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.252, + 0.468, + 0.321 + ], + "angle": 0, + "content": "[11] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Ting Zhang, Weiming Zhang, Nenghai Yu, Dong Chen, Fang Wen, and Baining Guo. Protecting celebrities from deepfake with identity consistency transformer. In Proc. CVPR, pages 9468-9478, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.324, + 0.468, + 0.407 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. ICLR, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.41, + 0.468, + 0.464 + ], + "angle": 0, + "content": "[13] Jianwei Fei, Yunshu Dai, Peipeng Yu, Tianrun Shen, Zhihua Xia, and Jian Weng. Learning second order local anomaly for general face forgery detection. In Proc. CVPR, pages 20270-20280, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.468, + 0.468, + 0.523 + ], + "angle": 0, + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. In Proc. NeurIPS, volume 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.526, + 0.468, + 0.58 + ], + "angle": 0, + "content": "[15] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proc. ACM MM, pages 3473-3481, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.584, + 0.468, + 0.639 + ], + "angle": 0, + "content": "[16] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, and Lizhuang Ma. Delving into the local: Dynamic inconsistency learning for deepfake video detection. In Proc. AAAI, pages 744-752, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.642, + 0.468, + 0.696 + ], + "angle": 0, + "content": "[17] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proc. CVPR, 2022. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.7, + 0.468, + 0.755 + ], + "angle": 0, + "content": "[18] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In Proc. CVPR, pages 5039-5049, 2021. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.758, + 0.468, + 0.812 + ], + "angle": 0, + "content": "[19] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Learning spatio-temporal features with 3d residual networks for action recognition. In Proc. ICCV Workshops, pages 3154-3160, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.816, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[21] Gee-Sern Hsu, Chun-Hung Tsai, and Hung-Yi Wu. Dual-generator face reenactment. In Proc. CVPR, pages 642-650, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.093, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[22] Yihao Huang, Felix Juefei-Xu, Qing Guo, Yang Liu, and Geguang Pu. Fakelocator: Robust localization of gan-based face manipulations. IEEE Transactions on Information Forensics and Security, 17:2657-2672, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[23] Ge-Peng Ji, Guobao Xiao, Yu-Cheng Chou, Deng-Ping Fan, Kai Zhao, Geng Chen, and Luc Van Gool. Video polyp segmentation: A deep learning perspective. MIR, 19(6):531-549, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[24] Ge-Peng Ji, Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Christos Sakaridis, and Luc Van Gool. Masked vision-language transformer in fashion. MIR, 20(3):421-434, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[25] Gengyun Jia, Meisong Zheng, Chuanrui Hu, Xin Ma, Yuting Xu, Luoqi Liu, Yafeng Deng, and Ran He. Inconsistency-aware wavelet dual-branch network for face forgery detection. IEEE Transactions on Biometrics, Behavior, and Identity Science, 3(3), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[26] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proc. CVPR, pages 2889-2898, 2020. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. CVPR, pages 4401-4410, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[28] Sohail Ahmed Khan and Hang Dai. Video transformer for deepfake detection with incremental learning. In Proc. ACM MM, pages 1821-1828, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[29] Aminollah Khormali and Jiann-Shiun Yuan. Dfdt: An end-to-end deepfake detection framework using vision transformer. Applied Sciences, page 2953, 2022. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[31] Jiaming Li, Hongtao Xie, Jiahong Li, Zhongyuan Wang, and Yongdong Zhang. Frequency-aware discriminative feature learning supervised by single-center loss for face forgery detection. In Proc. CVPR, pages 6458-6467, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[32] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proc. CVPR, pages 5001-5010, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[33] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. In Proc. CVPR Workshops, pages 656-663, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[34] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proc. CVPR, pages 3207-3216, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[35] Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: Rethinking face forgery detection in frequency domain. In Proc. CVPR, pages 772-781, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. ICCV, pages 10012-10022, 2021. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[37] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proc. CVPR, pages 8060–8069, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22667" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "[38] MarekKowalski. Faceswap. https://github.com/MarekKowalski/FaceSwap/. 2021-11-13. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.468, + 0.176 + ], + "angle": 0, + "content": "[39] Iacopo Masi, Aditya Killekar, Royston Marian Mascarenhas, Shenoy Pratik Gurudatt, and Wael AbdAlmageed. Two-branch recurrent network for isolating deepfakes in videos. In Proc. ECCV, pages 667–684, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.468, + 0.204 + ], + "angle": 0, + "content": "[40] Yisroel Mirsky and Wenke Lee. The creation and detection of deepfakes: A survey. ACM CSUR, 54(1):1-41, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.468, + 0.245 + ], + "angle": 0, + "content": "[41] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In Proc. ICCV, pages 3163-3172, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.468, + 0.302 + ], + "angle": 0, + "content": "[42] Yuval Nirkin, Lior Wolf, Yosi Keller, and Tal Hassner. Deepfake detection based on discrepancies between faces and their context. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6111-6121, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[43] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In Proc. ECCV, pages 86-103, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.468, + 0.385 + ], + "angle": 0, + "content": "[44] Rameswar Panda Quanfu Fan, Richard Chen. Can an image classifier suffice for action recognition? In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.468, + 0.442 + ], + "angle": 0, + "content": "[45] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. Faceforensics: Learning to detect manipulated facial images. In Proc. ICCV, pages 1-11, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.468, + 0.497 + ], + "angle": 0, + "content": "[46] Ekraam Sabir, Jiaxin Cheng, Ayush Jaiswal, Wael AbdAlmageed, Iacopo Masi, and Prem Natarajan. Recurrent convolutional strategies for face manipulation detection in videos. In Proc. CVPR Workshops, pages 80-87, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.468, + 0.565 + ], + "angle": 0, + "content": "[47] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. ICCV, pages 618-626, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.468, + 0.622 + ], + "angle": 0, + "content": "[48] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In Proc. CVPR, pages 11254-11264, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.468, + 0.677 + ], + "angle": 0, + "content": "[49] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proc. CVPR, pages 7672-7682, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.468, + 0.719 + ], + "angle": 0, + "content": "[50] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, pages 6105-6114, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.468, + 0.761 + ], + "angle": 0, + "content": "[51] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM TOG, 38(4):1-12, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.468, + 0.817 + ], + "angle": 0, + "content": "[52] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proc. ICCV, pages 2387-2395, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.859 + ], + "angle": 0, + "content": "[53] Luisa Verdoliva. Media forensics and deepfakes: an overview. IEEE Journal of Selected Topics in Signal Processing, 14(5):910-932, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[54] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In Proc. CVPR, pages 14923-14932, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[55] Junke Wang, Zuxuan Wu, Wenhao Ouyang, Xintong Han, Jingjing Chen, Yu-Gang Jiang, and Ser-Nam Li. M2tr: Multi-modal multi-scale transformers for deepfake detection. In Proc. ICMR, pages 615-623, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[56] Ping Wang, Kunlin Liu, Wenbo Zhou, Hang Zhou, Honggu Liu, Weiming Zhang, and Nenghai Yu. Adt: Anti-deepfake transformer. In Proc. ICASSP, pages 2899-1903, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[57] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proc. CVPR, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[58] Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv:2102.11126, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[59] Ziming Yang, Jian Liang, Yuting Xu, Xiao-Yu Zhang, and Ran He. Masked relation learning for deepfake detection. IEEE Transactions on Information Forensics and Security, pages 1696-1708, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.343, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proc. ICCV, pages 6023-6032, 2019. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[61] Yu Zeng, Zhe Lin, and Vishal M Patel. Sketched: Mask-free local image manipulation with partial sketches. In Proc. CVPR, pages 5951-5961, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[62] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. Proc. ICLR, 2018. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.481, + 0.892, + 0.536 + ], + "angle": 0, + "content": "[63] Yanyi Zhang, Xinyu Li, Chunhui Liu, Bing Shuai, Yi Zhu, Biagio Brattoli, Hao Chen, Ivan Marsic, and Joseph Tighe. Vidtr: Video transformer without convolutions. In Proc. ICCV, pages 13577-13587, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.537, + 0.892, + 0.605 + ], + "angle": 0, + "content": "[64] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: Interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18:1335-1348, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.606, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[65] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep-fake detection. In Proc. CVPR, pages 2185–2194, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[66] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Self-supervised transformer for deepfake detection. arXiv:2203.01265, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[67] Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In Proc. CVPR, pages 15023-15033, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[68] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In Proc. ICCV, pages 15044-15054, 2021. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.8, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[69] Xiangyu Zhu, Hao Wang, Hongyan Fei, Zhen Lei, and Stan Z Li. Face forgery detection by 3d decomposition. In Proc. CVPR, pages 2929-2939, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.842, + 0.892, + 0.896 + ], + "angle": 0, + "content": "[70] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proc. ACM MM, pages 2382-2390, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22668" + } + ] +] \ No newline at end of file diff --git a/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/b759f55b-503f-4164-9793-b1ffe02056fb_origin.pdf b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/b759f55b-503f-4164-9793-b1ffe02056fb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e4c93ef4ce4089885657ce5c5884424055d050af --- /dev/null +++ b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/b759f55b-503f-4164-9793-b1ffe02056fb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e690cd539459d4f950096ac984281a761c85e3808ec4b05c1c985e5568939658 +size 3561128 diff --git a/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/full.md b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..868ffa0cd8b13304e7e021b82be9b59045c7a31f --- /dev/null +++ b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/full.md @@ -0,0 +1,350 @@ +# TALL:,Thumbsail Layout for Deepfake Video Detection + +Yuting Xu $^{1,3*}$ , Jian Liang $^{2,4}$ , Gengyun Jia $^{5}$ , Ziming Yang $^{1,3}$ , Yanhao Zhang $^{6}$ , Ran He $^{2,4\dagger}$ + +$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences + +$^{2}$ CRIPAC & MAIS, Institute of Automation, Chinese Academy of Sciences +$^{3}$ School of Cyber Security, UCAS $^{4}$ School of Artificial Intelligence, UCAS + +$^{5}$ School of Communications and Information Engineering, NJUPT $^{6}$ OPPO Research Institute + +yuting.xu@cripac.ia.ac.cn, liangjian92@gmail.com, rhe@nlpr.ia.ac.cn + +# Abstract + +The growing threats of deepfakes to society and cybersecurity have raised enormous public concerns, and increasing efforts have been devoted to this critical topic of deepfake video detection. Existing video methods achieve good performance but are computationally intensive. This paper introduces a simple yet effective strategy named Thumbnail Layout (TALL), which transforms a video clip into a pre-defined layout to realize the preservation of spatial and temporal dependencies. Specifically, consecutive frames are masked in a fixed position in each frame to improve generalization, then resized to sub-images and rearranged into a pre-defined layout as the thumbnail. TALL is model-agnostic and extremely simple by only modifying a few lines of code. Inspired by the success of vision transformers, we incorporate TALL into Swin Transformer, forming an efficient and effective method TALL-Swin. Extensive experiments on intra-dataset and cross-dataset validate the validity and superiority of TALL and SOTA TALL-Swin. TALL-Swin achieves $90.79\%$ AUC on the challenging cross-dataset task, FaceForensics++ $\rightarrow$ CelebDF. The code is available at https://github.com/rainy-xu/TALL4Deepfake. + +# 1. Introduction + +Deepfakes generate and manipulate facial appearances to deceive viewers through generation techniques [61, 48]. With the remarkable success of generative adversarial networks [14, 27], deepfake products have become photorealistic that humans can not distinguish. These deepfake products [21, 49] may be misused for malicious purposes, leading to severe trust issues and security problems, such as financial fraud, identity theft, and celebrity impersonation. + +![](images/a2648b852e6f75e080588e0ff8d14afab55c72aee5ab1b02647d8b2fd55fcd11.jpg) +Figure 1. The AUC and FLOPs trade-off of different backbones. Image-level backbones with TALL enjoy comparable accuracy-cost trade-offs with the 3DCNN and video transformer family on the unseen Celeb-DF dataset. All models with the same setting are trained on the FF++ (HQ) dataset. + +ation [53, 40]. The rapid development of social media accelerates the abuse of deepfakes. Therefore, it is crucial to develop advanced detection methods to protect the data privacy of individual users. + +Most previous image-based methods [22, 69] perform well on intra-dataset, but their generalizability needs to be improved. Recent research has focused on video-based methods to detect deepfake by modeling spatio-temporal dependencies. There are subtle spatio-temporal inconsistencies between frames since the deepfake algorithms are executed frame by frame. The core of video-level approaches for deepfake detection is capturing inconsistencies through temporal modeling. Existing deepfake video detection methods generally follow two directions. Some methods [15, 16] use two-branch networks or modules to learn spatial and temporal information separately and then fuse them. However, these two-branch approaches may frag + +ment spatiotemporal cooperation and lead to subtle artifacts being neglected. Others directly use classic temporal models such as LSTM and 3D-CNNs. These methods are computationally intensive. The current rise of transformers for vision task backbones has prompted the emergence of corresponding deepfake detection methods. They are accompanied by significant computational complexity that makes them challenging to deploy and use, despite breakthroughs in performance. To enjoy benefits from both image and video methods, we are curious to see whether it is possible to append information about the temporal dimension to the image dimension. + +This work develops a simple yet effective Thornail Layout (TALL) for deepfake detection by spatio-temporal modeling. TALL is computationally cheap and retains both temporal and spatial information. In detail, we use dense sampling to extract multiple clips in the video and then randomly select four consecutive frames in the video segment. Subsequently, a block is masked at a fixed position in each frame. Finally, the frames are resized as sub-image and sequentially rearranged into a pre-defined layout as a thumbnail, which has the same size as the clip frames. As shown in Figure 1, TALL brings two advantages compared to the previous spatio-temporal modeling methods for deepfake detection: (1) TALL contains local and global contextual deepfake patterns. (2) TALL is a model-agnostic method for spatio-temporal modeling deepfake patterns at zero computation and zero parameters. + +Furthermore, we discover that the better temporal modeling capabilities backbone has, the better performance TALL achieves. Based on the proposed TALL, we complement a baseline for video deepfake detection based on Swin Transformer [36], called TALL-Swin. We validate TALL-Swin on four popular benchmark datasets, including FaceForensics++, Celeb-DF, DFDC, and DeeperForensics. Our method gains a remarkable improvement over the state-of-the-art approaches. The main contributions of our paper are summarized as follows: + +- We provide a new perspective for an efficient strategy for video deepfake detection called Thumbnail Layout (TALL), which incorporates both spatial-temporal dependencies, and allows the model to capture spatial-temporal inconsistencies. +- We propose a spatio-temporal modeling method called TALL-Swin, which efficiently captures the inconsistencies between deepfake video frames. +- Extensive experiments demonstrate the validity of our proposed TALL and TALL-Swin. TALL-Swin outperforms previous methods in both intra-dataset and cross-dataset scenarios. + +# 2. Related Work + +# 2.1. Image-Level Deepfake Detection + +Typically, existing deepfake detection methods fall into two categories: image-level and video-level methods. The image-level methods [25, 13] always exploit the artifacts of deepfake images in the spatial domain, such as discrepancies between local regions [42, 59], grid-like structure in frequency space [10], and differences in global texture statistics [37] that provide specific clues to distinguish deepfakes from the real images. F3Net [43] and FDFL [31] utilize the same pipeline that utilizes frequency-aware features and RGB information to capture the traces in different input spaces separately. RFM [54] and Multi-att [65] propose an attention-guided data augmentation mechanism to guide detectors to discover undetectable deepfake clues. Face X-ray [32] and PCL [67] provide effective ways to outline the boundary of the forged face for detecting deepfakes. ICT [11] exploits an identity extraction module to detect identity inconsistency in the suspect image. Similarly, M2tr [55] detects local inconsistencies within frames at different spatial levels. Generally, image-level methods suffer over-fitting issues when a specific technique manipulates the images, and they ignore temporal information. + +# 2.2. Video-Level Deepfake Detection + +To improve the generalization of deepfake detectors, many studies generate diversity and generic deepfake data, while other studies capture the temporal incoherence of fake videos as generic clues. Some recent works propose detecting temporal inconsistency using well-designed spatiotemporal neural networks, and others [15, 16] attempt to add modules to image models that capture temporal information. STIL [15] formulates deepfake video detection as a spatial and temporal inconsistency learning process and integrates both spatial and temporal features in a unified 2D CNN framework. FTCN [68] detects temporal-related artifacts instead of spatial artifacts to promote generalization. LipForensics [18] is proposed to learn high-level semantic irregularity in mouth movement in the generated video. RealForensics [17] uses auxiliary data sets during training in exchange for generalization at the cost of higher computational demands. The video-based methods achieve strong generalization but suffer from large computational overhead. To reduce computational costs, we propose TALL which gathers consecutive video frames into thumbnails for learning spatio-temporal consistency. + +# 2.3. Deepfake Detection with Vision Transformer + +Recently, ViT [12] has achieved impressive performance in computer vision tasks [24, 23, 44]. Many studies extend the ViT for deepfake detection [66, 58]. These methods achieve better performance compared to CNN-based + +models, but also sacrifice computational efficiency. A few works [55, 66] attempt to extend the transformer for deepfake detection due to the advent of the visual transformer (ViT) and the impressive ability to model long-range data, different from two-branch architectures that capture short-range and long-range temporal inconsistencies with a single-branch model. ICT [11] aims to detect identity consistency in deepfake video but may fail in detecting face reenactment and entire face synthesis results. DFLL [28] extract the UV texture map to help the transformer to detect deepfakes, which may disrupt the continuity between video frames. DFTD [29] leverages ViT to consider both global and local information but ignores the problem of excessive model arithmetic requirements. Although the transformer-based approaches achieve promising performance, they are accompanied by significant computational complexity that makes them challenging to deploy and use, and the long-range dependencies may be insufficiently exploited in detection models. Swin Transformer [36] produces a hierarchical feature representation and has linear computational complexity concerning input image size, which is suitable as a general-purpose backbone for various vision tasks. In this paper, we cooperate with Swin Transformer to form our robust and efficient method TALL-Swin. + +# 3. Method + +TALL is a deepfake video detection strategy that transforms a video clip into an all-in-one thumbnail without the extra computational overhead. In the following sections, we begin with the motivation of TALL for deepfake detection in Section 3.1. Then we present the technical details of the TALL in Section 3.2. Finally, a generalizable SwINTALL baseline is introduced to explore subtle artifacts in Section 3.3. More details of the layout design are presented in the Appendix. + +# 3.1. Motivation + +While recent studies have attempted to address noticeable flaws through techniques like slight motion blurring and temporal consistency loss, subtle spatio-temporal artifacts still remain. These artifacts are important for detecting deepfakes, but they introduce two problems: 1) video-based models are less efficient, and 2) analyzing information over long distances may overlook local artifacts, which are critical for deepfake detection. To address these challenges, we propose the TALL strategy, which naturally incorporates temporal information into image-level tasks without disrupting spatial information. This approach enables the image-level model to detect deepfakes in videos. Furthermore, we discovered that TALL provides even greater performance gains when combined with a powerful spatial model, resulting in the TALL-Swin. + +In detail, TALL arranges consecutive frames in the + +![](images/9380b024a1cd3eed710e5f8204bc0c4fabb7b49b9666b2f3aca592a19a7c0fda.jpg) +(a) Thumbnail Layout (TALL) +Figure 2. Illustration of the TALL and shifted window process for computing self-attention in the TALL. + +![](images/e79a74cc79cb3363864ed670b85229736b93aab6050a999cedcc01393ba14187.jpg) +(b) Feature Map + +temporal order in a compact $2 \times 2$ layout, in line with the calculation theory of convolution and shifted window. TALL contains both spatial and temporal information so that model can learn both intra-frame artifacts and interframe inconsistency and obtains comparable performance to video-based methods. Here we use the shifted window to explain TALL's mechanism. As illustrated in Figure 2 (a), the model computes self-attention while accounting for spatial dependencies across sub-images (represented by the solid red box). When the window spans multiple sub-images (represented by the red dash box), the model is able to capture temporal inconsistencies between frames. Moreover, TALL leverages both local and global contexts of deepfake patterns to ensure robust modeling capabilities for short and long-range spatial dependencies. Compared to previous methods, we anticipate that TALL strikes a balance between speed and accuracy, sacrificing a little spatial information while preserving performance. Based on the fact that attention-based models are better at handling contextual features and that the Swin-Transformer uses shifted windows to reduce computation and memory, we further complement TALL-Swin baseline for video deepfake detection. + +# 3.2. Thumbnail Layout (TALL) + +Given a video $V \in \mathbb{R}^{T \times C \times H \times W}$ , where $T$ is the frame length of the video, $C$ is the number of channels, and $H \times W$ is the resolution of the frames. Assuming each video contains $N$ clips, we divide a video into $N$ equal segments of length $T / N$ and then sample consecutive $t$ (set to 4 by default) frames from the segments at random locations to form one clip. Then, the thumbnail $I$ is rearranged of sub-images $(C \times \frac{H}{\sqrt{t}} \times \frac{W}{\sqrt{t}})$ that are resized from the above $t$ frames. To maximize the utility of TALL, we mask the organized $N$ square masks of the thumbnail. It is based on two core designs: 1) The position of the masks is random between different sub-images, which retains the advantages of the Cutout [8] that encourages the network to focus more + +Algorithm 1 Pseudocode of TALL in a PyTorch-like style. +# x: one clip of video (T*C*H*W) +# T: frame number of clip +# C: channels; s: mask size +# d: length of frame included in the thumbnail +# r: rows of thumbnail +# x_tall: thumbnail image (224*224) +#TALL's augmentation strategy +h = np.random.randint(H) +w = np.random.randint(W) +#the mask position is fixed for each frame +m = np.ones((H, W)) +h1 = np.clip(h - s // 2, 0, H) +h2 = np.clip(+s // 2, 0, H) +w1 = np.clip(w - s // 2, 0, W) +w2 = np.clip(w + s // 2, 0, W) +m[h1: h2, w1: w2] = 0 +m = torch.from_numpy(m) +m = maskexpand_as(x) +x = x * m +#TALL: generation of the thumbnail +x = x.view(-1, H, W).unsqueeze(0) +x = x.view((-1, C*d) + x_tall.size() [2:]) +x = rearrange(x, 'b (th tw c) h w -> b c (th h) (tw w)', th=r, c=C) +x_tall = interpolate(x, size=H) + +on complementary and less prominent features. 2) We fix the position of the mask within a clip to take advantage of the fact that most deepfake videos are frame-by-frame tampered with, thus forcing the model to detect inconsistencies between adjacent frames of the deepfake videos. We do not allow the mask to appear on the seams of the thumbnail but allow for partial mask inclusion in the thumbnail. The detailed procedure of TALL is summarized in Algorithm 1. + +# 3.3. TALL-Swin + +To balance efficiency and model performance for spatiotemporal feature learning and to leverage the benefits of attention-based models, we enhanced a baseline deepfake detection model called TALL-Swin by incorporating the Swin Transformer [36]. Given the characteristics of TALL, we slightly modified the window size of Swin-B in TALL-Swin. We first enlarge the window size of the first three stages of the model so that the interaction between frames in the thumbnail becomes more frequent, forcing the model to learn more detailed spatio-temporal dependencies. Next, we set the window size of the last stage to be the same as the feature map size, enabling the window to perform global attention computations while TALL-Swin captures global spatial-temporal dependencies. As a result, the size of the last layer of the feature map became smaller, reducing the window size without introducing any additional computational overhead. Consequently, the window sizes for the four stages of TALL-Swin are [14, 14, 14, 7]. Note that the patch merging process makes TALL-Swin captures a more + +comprehensive range of dependencies through hierarchical representations, as shown in Figure 2 (b). + +Given a video of length $T$ , each frame contains $N$ patches, and the window contains $P$ patches. To demonstrate the superiority of TALL-Swin in terms of computational consumption, we show below the computational complexity of the image-level transformer and video-level transformer, including ViT [12], Swin [36], ViViT [2], and TALL-Swin respectively: + +$$ +\Omega_ {\mathrm {V i T}} = 4 T N C ^ {2} + 2 T N ^ {2} C, +$$ + +$$ +\Omega_ {\mathrm {S w i n}} = 4 T N C ^ {2} + 2 T P N C, \tag {1} +$$ + +$$ +\Omega_ {\mathrm {V i V i T}} = 4 T N C + 2 T ^ {2} N ^ {2} C, +$$ + +$$ +\Omega_ {\mathrm {T A L L - S w i n}} = T N C ^ {2} + \frac {1}{2} T P N C. +$$ + +TALL-Swin has the lowest computational complexity compared to image and video-level transformer methods. Subsequent experiments will demonstrate that TALL-Swin maintains performance, albeit at the sacrifice of some spatial information. + +The cross-entropy loss is employed to optimize the TALL-Swin, which is defined as: + +$$ +\mathcal {L} _ {C E} = - \frac {1}{n} \sum_ {i = 1} ^ {n} y _ {i} \log \mathcal {F} \left(x _ {i}\right) + (1 - y _ {i}) (\log (1 - \mathcal {F} \left(x _ {i}\right)), \tag {2} +$$ + +where $x_{i}$ indicates input clip, $y_{i}$ denotes the label of clip, $n$ is the number of clip, $\mathcal{F}$ is TALL-Swin. + +# 4. Experiments + +# 4.1. Setup + +Datasets. Following previous works [18, 17, 68], we evaluate the TALL and TALL-Swin on four widely used datasets. FaceForensics++ [45] is a most-used benchmark on intra-dataset deepfake detection, consisting of 1,000 real videos and 4,000 fake videos in four different manipulations: DeepFake [7], FaceSwap [38], Face2Face [52], and NeuralTextures [51]. Besides, FaceForensics++ contains multiple video qualities, e.g. high quality (HQ), low quality (LQ) and RAW. Celeb-DF (CDF) [34] is a popular benchmark on cross-dataset, which contains 5,693 deepfake videos generated from celebrities. The improved compositing process was used to improve the various visual artifacts presented in the video. Celeb-DF is also suitable for deepfake detection tasks with a reference set. DFDC [9] is a large-scale benchmark developed for Deepfake Detection Challenge. This dataset includes 124k videos from 3,426 paid actors. The existing deepfake detection methods do perform not very well on DFDC due to their sophisticated deepfake techniques. DeeperForensics (DFO) [26] includes 60,000 videos with 17.6 million frames for deepfake detection, whose videos vary in identity, pose, expression, emotion, lighting conditions, and blend shape with high quality. + +
ModelsTemp.CDFDFDCFLOPsParamsPT
I3D-RGB* [4]✓78.2465.58222.7G25M1K
R3D-50* [19]✓79.6367.73296.6G46M1K
ResNet50* [20]×76.3864.0125.5G21M1K
+TALL✓80.9065.5425.5G21M1K
EffNetB4* [50]×78.1966.818.3G19M1K
+TALL✓83.3767.158.3G19M1K
VTN [41]✓83.2073.50296.6G46M21K
VidTR [63]✓83.3073.30117G93M21K
ViViT* [2]✓86.9674.61628G310M21K
ISTVT [64]✓84.1074.20455.8G--
ViT-B* [12]×82.3372.6455.4G84M21K
+TALL✓86.5874.1055.4G84M21K
Swin-B* [36]×83.1373.0147.5G86M21K
TALL-Swin✓90.7976.7847.5G86M21K
+ +Table 1. Performance of different backbones. TALL consistently improves the accuracy over different image-level models. We show the AUC, FLOPs, and number of parameters for each model on the cross-dataset scenario. All models are trained on $\mathrm{FF} + +$ (HQ). $\checkmark$ indicates the model enables temporal modeling. \* indicates our implementation. PT indicates pre-train. 1K and 21K indicate the model pre-trained on ImageNet-1K and 21K respectively. The best results are bold. + +Implementation Details. We use MTCNN to detect face for each frame in the deepfake videos, only extract the maximum area bounding box and add $30\%$ face crop size from each side as in LipForensics [18]. The ImageNet-21K pretrained Swin-B model is used as our backbone. Excluding ablation experiments, we sample 8 clips using dense sampling, each clip contains 4 frames. The size of the thumbnail is $224 \times 224$ . Following Swin Transformer [36], Adam [30] optimization is used with a learning rate of 1.5e-5 and batch size of 4, using a cosine decay learning rate scheduler and 10 epochs of linear warm-up. We adopt Acc. (accuracy) and AUC (Area Under Receiver Operating Characteristic Curve) as the evaluation metrics for extensive experiments. To ensure a fair comparison, we calculate video-level predictions for the image-based method and average the predictions across the entire video (following previous works [18, 16, 35, 68]). Note that results are directly cited from published papers if we follow the same setting. + +# 4.2. Scaling over Backbones + +To verify our assumption, we adopt several image-level backbones commonly used for deepfake detection for comparison with the video-level backbones. As shown in Table 1 above the double horizontal line, we first compare the accuracy and complexity of the CNN-based video and image backbones. Although I3D [4] and R3D [19] achieve + +better performance than vanilla ResNet50 [20] and EfficientNet [50], the computation costs are huge, such as R3D-50 with 296G FLOPs. For ResNet and EfficientNet who added TALL, ResNet achieves better AUC both on CDF (76.38 VS 80.93) and DFDC (64.01 VS 65.54) datasets, and EfficientNet achieves $5.18\%$ better AUC on CDF. + +The second section contains the video and image transformers. Compared to video transformers, the image-based ViT and Swin fail to achieve better performance due to the lack of temporal modeling. For example, ViViT achieves $86.96\%$ AUC on CDF, which is $3.6\%$ higher than Swin although ViViT with $13\times$ more computation. By way of contrast, $\mathrm{ViT + TALL}$ achieves $86.58\%$ AUC on CDF with 55.4G FLOPs, which is comparable to AUC with ViViT but with low computation. Accordingly, Swin's performance was significantly improved with the addition of TALL without computation increment. On the other hand, TALL boosts higher performance on models with learned long-range dependencies. e.g., ResNet+TALL $(+4.5\%)$ on CDF and $+1.5\%$ on DFDC) vs. Swin+TALL $(+7.6\%)$ on CDF and $+3.6\%$ on DFDC). These two section results demonstrate that TALL provides both spatial and temporal information and enables the model to learn spatial and temporal inconsistencies for video deepfake detection. + +
MethodsFF++ (HQ)FF++ (LQ)
Acc.AUCAcc.AUC
MesoNet [1]83.10-70.47-
Xception [6]95.7396.3086.8689.30
Face X-ray [32]-87.35-61.60
Two-branch [39]96.4398.7086.3486.59
Add-Net [70]96.7897.7487.5091.01
F3-Net [43]97.5298.1090.4390.43
FDFL [31]96.6999.3089.0092.40
Multi-Att [65]97.6099.2988.6990.40
RECCE [3]97.0699.3291.0395.02
LipForensics [18]98.8099.7094.2098.10
DFDT [29]98.1899.2692.6794.48
ADT [56]92.0596.3081.4882.52
ST-M2TR [55]-99.42-95.31
VTN [41]98.47-94.02-
VidTR [63]97.42-92.12-
ViViT* [2]92.60-88.02-
ISTVT [64]99.00-96.15-
TALL-Swin98.6599.8792.8294.57
+ +Table 2. Intra-dataset evaluations. We report the video-level Acc. $(\%)$ and AUC $(\%)$ on the FF++ dataset. HQ indicates high quality, and LQ indicates low quality. + +# 4.3. Comparison with State-of-the-art Methods + +Intra-dataset evaluations. Following ISTVT [64], we show the results of the $\mathrm{FF}++$ dataset under both Low Quality (LQ) and High Quality (HQ) videos, and report comparisons against several advanced methods in Table 2. We can observe that advanced video-based transformers have better results than CNN-based methods. Compared to video-based transformer methods, TALL-Swin has comparable performance and lower consumption to the previous video transformer method with HQ settings. However, TALL-Swin gets unsatisfactory results with the LQ setting. The LQ setting is obtained by severely compressing the videos. So the reason for the result may be that TALL scales the frame to a smaller size, causing more spatial information to be lost in the frame. We will investigate the possibility of other designs to further improve performance in the LQ setting. + +
MethodCDFDFDCFShDFoAvg.
Xception [6]73.7070.9072.0084.5075.28
CNN-aug [57]75.6072.1065.7074.4071.95
CNN-GRU [46]69.8068.9080.8074.1073.40
Patch-based [5]69.6065.6057.8081.8068.70
Face X-Ray [32]79.5065.5092.8086.8081.15
Multi-Att [65]75.7068.1066.0077.7071.88
DSP-FWA [33]69.5067.3065.5050.2063.13
LipForensics [18]82.4073.5097.1097.6087.65
FTCN [68]86.9074.0098.8098.8089.63
RealForensics [17]86.9075.9099.7099.3090.45
DFDT [29]88.3076.1097.8096.9089.70
VTN [41]83.2073.5098.7097.7088.30
VidTR [63]83.5073.3098.0097.9088.10
ViViT* [2]86.9674.6199.4199.1990.05
ISTVT [64]84.1074.2099.3098.6089.10
TALL-Swin90.7976.7899.6799.6291.71
+ +Table 3. Generalization to unseen datasets. We report the video-level AUC (\%) on four unseen datasets: Celeb-DF (CDF), DFDC, FaceShifter (FSh), and DeeperForensics (DFo). + +Generalization to unseen datasets. In addition to the intra-dataset comparisons, we also investigate the generalization ability of our method. Adhering to the deepfake video detection cross-dataset protocol [18], we train a model on $\mathrm{FF}++$ (HQ) then test on Celeb-DF (CDF), DFDC, FaceShifter (FSh), and DeeperForensics (DFo) datasets. As shown in Table 3: (1) Video-based methods generally have better results than image-based methods, which shows that temporal information is helpful for the deepfake video detection task. For example, Lip outperforms Face X-ray's AUC by a wide margin. In addition, most transformer-based models have higher performance than CNN-based models. For the transformer-based models, both achieved + +an average AUC of $88\%$ , while the best CNN-based video-level models only achieved $87\%$ . (2) TALL-Swin achieves state-of-the-art results on Celeb-DF, DFDC, and Deeper-Forensics datasets, and also beats its competitors on Celeb-DF dataset by a large margin $(3.8\%)$ . The results demonstrate that TALL-Swin performs well when encountering unseen datasets with better generalization ability than previous video transformer methods. + +![](images/67bba2563ef26452db382cf417d3f1da6d2be31f271f9fa32b71b43dc6e1e2e1.jpg) +Figure 3. Saliency map visualization of TALL-Swin on different datasets. The first four rows of samples are from the FF++ dataset, and the last four rows are from the unseen datasets. + +Analysis of saliency map visualization. We adopt Grad-CAM [47] to visualize where the TALL-Swin is paying its attention to the deepfake faces. In Figure 3, we give the results on intra-dataset and cross-dataset scenarios. All models are trained on $\mathrm{FF} + +$ (HQ). It can be observed in the first four rows of Figure 3 that TALL-Swin captures method-specific artifacts. Note that the DF transfers the face region from a source video to a target, and the NT only modifies the facial expressions corresponding to the mouth region. TALL-Swin corresponds to focus on the face region and the mouth region. Furthermore, our model traces the more generalized artifacts that are independent of manipulation methods, e.g., blending boundaries (CDF), and abnormal motions in the clip (DFDC, Fsh, Dfo). + +![](images/a5b466759f3094b0caecd7262ea43e40088af880ef34f19ca52484688d391b08.jpg) +Xception Face X-ray ICT Patch-based CNN-GRU LipForensics TALL-Swin + +![](images/b9c8c8662d00ed5404ab6de3103706e6d41d22a449160c4a34fd368fdebe3a9e.jpg) + +![](images/a7b257598955b38b8906a145e1afced6ed0b20b2ce4e1c02ecef460c9f66632f.jpg) + +![](images/8f894b31f0b4f348eb464cab653a19b18908a33ca7358a86a0486885792461b6.jpg) + +![](images/1288e83695a1a47eafc8e248fd431a864848075e0e0f2b481c6b972c2ef75ab7.jpg) +Figure 4. Robustness to various unseen corruptions. We report the video-level AUC $(\%)$ of our methods under five different levels of seven particular types of corruption. "Average" denotes the mean across all corruptions at each severity level. Our TALL-Swin is more robust than previous methods for all corruptions. + +![](images/9471f592a87e976ec6610e2702a13b8da8f7c970ae161d7246166f975e8baca1.jpg) + +![](images/b21159c95b3a03adbf4842f84a81ce5140cd6d08b6d0aa5ed8025853777083c5.jpg) + +![](images/e65bbfe61523b63ee11bbbf46dccf1e03b5dee93a39b68ce6d023f8ae9db92d7.jpg) + +
MethodCleanSaturationContrastBlockNoiseBlurPixelCompressAvg.
Xception (ICCV'19) [6]99.899.398.699.753.860.274.262.178.3
CNN-GRU (CVPRW'19) [46]99.999.098.897.947.971.586.574.582.3
CNN-aug (CVPR'20) [57]99.899.399.195.254.776.591.272.584.1
Patch-based (ECCV'20) [5]99.984.374.299.250.054.456.753.467.5
Face X-ray (CVPR'20) [32]99.897.688.599.149.863.888.655.277.5
LipForensics (ICCV'21) [18]99.999.999.687.473.896.195.695.692.5
FTCN (ICCV'21) [68]99.499.496.797.153.195.898.286.489.5
RealForensics (CVPR'22) [17]99.899.899.698.979.795.398.497.695.6
TALL-Swin w/o mask100.0100.0100.099.883.597.398.497.996.7
TALL-Swin100.0100.0100.0100.085.397.698.598.197.1
+ +Table 4. Average robustness to unseen corruptions. Average Video-level AUC (%) across five intensity levels for each corruption type proposed in DFo [26]. "Avg" indicates the mean across all corruptions and all levels. + +Robustness to unseen perturbations. Deepfake detectors must be robust to common perturbations, given that video propagation on social media causes video compression, noise addition, etc. We also study the performance of robustness to unseen perturbations. Following RealForensics [17], the experiment applies seven unseen perturbations to fake videos at five intensity levels. In Figure 4, we show results of increasing the severity of each corruption. We can observe that other methods degrade dramatically as the perturbations become more severe. TALL-Swin still has a high performance. However, TALL-Swin degrades when + +the Gaussian noise reaches level five. Table 4 presents the average AUC across all intensity levels for corruption types. We observe that our method is significantly more robust to most perturbations than other methods. The good robustness may be from both the design of TALL and the proposed mask augmentation. The main reason may be the consecutive multi-frame input. We empirically consider that the key to deepfake detection is local inconsistency, the continuous frame design has less redundant information, ensuring that the model finds locally important clues. + +![](images/c8a499cba746e8873144849dd2f06d2a7bfb91df4a44f8e8d85499bad080da71.jpg) +(a) + +![](images/c3fb8489399ebb8d07696b41d794cdcb57ea4bf936005c3bc7d07cc0e1403dfc.jpg) +Figure 5. Illustration of different layout designs. + +![](images/9dd06e16684826bf5dff944678b7f2b73766ac45bb6fc7836e624b0dd6e9004d.jpg) + +![](images/e6a081a7e563d69a532187f0d2fe7af46d89f58a5e4d20d9cae25f9fd9692e0a.jpg) + +# 4.4. Ablation Study + +We perform the ablation study to analyze the effects of each component and hyper-parameter in TALL-Swin. All experiments are trained on $\mathrm{FF}++$ (HQ) and tested on the CDF and DFDC datasets. + +Effects of different layouts. We train a TALL-Swin model on $\mathrm{FF}++(\mathrm{HQ})$ for each layout illustrated in Figure 5, to analyze in which layout of the thumbnails the model learns the strongest generalization of the spatial-temporal dependence of the deepfake patterns. As shown in Table 5, the model with a compact layout like Figure 5 (d) has good generalization ability on the unseen datasets. A compact layout like Figure 5 (d) may help the model to learn the temporal dependence across frames because such a form provides the shortest distance between any two images. + +
LayoutCDFDFDC
Figure 5 (a)85.5270.02
Figure 5 (b)84.9373.57
Figure 5 (c)86.6672.12
Figure 5 (d)87.6074.32
+ +Study on the numbers of sub-image. We use Swin-B as the baseline for this study to compare the effect of different thumbnail layout schemes on the model's generalization ability. Changing frames to thumbnails involves scaling, so we also investigate the impact of resizing and random cropping pre-processing on model performance. We set up four variants: resizing pre-process with $4 \times 4$ layout, $3 \times 3$ layout and $2 \times 2$ layout; random cropping pre-process with $2 \times 2$ layout. As shown in Table 6, the model performance degrades sharply when using $4 \times 4$ layout. This may be due to the small size of each sub-image that the spatial information is not captured well by the model. The result of $3 \times 3$ layout also slightly decreases. $2 \times 2$ layout with resizing pre-processing beats $2 \times 2$ layout with random crop. We also found that TALL-Swin achieves the best performance and the AUC score increases $3.2\%$ compared to the baseline, suggesting that thumbnails in a $2 \times 2$ layout are more helpful to TALL-Swin than original frames. + +Effects of Sub-image's size. We eliminate the scaling + +Table 5. Effects of different layouts. All models here are trained without mask augmentation. + +
Pre-processLayoutCDFDFDC
None-83.1373.01
Resize4 × 480.1870.45
Resize3 × 383.1872.98
Crop2 × 278.5573.30
Resize2 × 287.6074.32
+ +operation for sub-images to allow for more flexible layout settings. However, we've observed that when the number of sub-images grows at their original size, the performance improvements are only slight. Additionally, the computational complexity increases dramatically with the number of frames ( 4.3 times more than the TALL setting), as demonstrated in Table 7. To strike a balance between performance and computational complexity, we reduce the resolution of sub-images in TALL. + +Table 6. Study on the numbers of sub-image. All models here are trained without mask augmentation. + +
Subimage-sizeLayoutFLOPsCDFDFDC
224 × 2243 × 3253G88.6975.98
224 × 2242 × 2185G88.1575.01
112 × 1122 × 247.5G87.6074.32
+ +Study on absence and order of thumbnails. In this case, we study the impact of missing the last sub-image and the last two sub-images on the model's performance. The first two rows of Table 8 show that all four sub-images contribute to the model performance. Besides, we set the order of the different thumbnails to evaluate the TALL-Swin. We consider three orders: forward, reverse, and random. Forward order performs the best for three different orders. This may be because of the positional encoding of different frames in TALL-Swin. + +Table 7. Effects of Sub-image's size. All models here are trained without mask augmentation. + +
VariantsCDFDFDC
0, 1, 2, -86.4669.51
0, 1, -, -84.2269.09
Random85.8570.30
Reverse86.6572.37
Forward87.6074.32
+ +Table 8. Ablation study of absence and order of thumbnails. All models here are trained without mask augmentation. + +Effects of different orders on other backbones. In order to prove that the phenomenon is not incidental, we also conduct experiments on ResNet50 and EfficientNet for three different orders as shown in Table 9. As expected, for- + +ward order outperforms reverse and random orders both on ResNet50 and EfficientNet, which indicates that TALL can learn the temporal dependency. + +
VariantsCDFDFDC
ResNet50+TALL76.3864.01
Random78.1464.12
Reverse78.5464.87
Forward80.9065.54
EfficientNet+TALL78.1966.81
Random81.0166.13
Reverse81.6666.69
Forward83.3767.15
+ +Effectiveness of mask strategy. In this work, TALL-Swin is trained on the $\mathrm{FF}++$ (HQ) dataset without any data enhancement as the baseline except for Multi-scale Crop and Random Horizontal Flip. To validate the effectiveness of the mask strategy, we compare our default baseline with different data augmentation strategies: 1) The Cutout [8] on one sub-image; 2) The Cutout on four sub-images. 3) The combination of Mixup [62] and Cutmix [60] on four sub-images, as shown in Table 10. The performance of a random Cutout [8] on four sub-images is better than on one sub-image. Besides, the mask strategy leads to better performance than the well-known Cutout $(1.46\%)$ . This supports our hypothesis that strategy encourages models to learn subtle temporal-spatial variations and improves model generalization ability. Further, our augmentation strategy exceeds $1.02\%$ than the combination of Mixup and Cutmix, demonstrating the augmentation's effectiveness in TALL for video detection. + +Table 9. Ablation studies of different orders of TALL. + +
AugmentationCountCDFDFDC
None-87.6074.32
Cutout189.0674.07
Cutout489.3375.22
Mixup+CutMix489.7575.33
TALL's mask490.7976.78
+ +Study on window size. We study the effect of window size on model performance and computational cost. The results are shown in Table 11. Our window expansion for the first three phases will increase the model performance by $1.74\%$ AUC. The results in the second and third rows show that the first three stages of the window getting the largest would not give a boost to the model. Our analysis of a too-large window may lead to a weakening of the model's ability to learn local information in the sub-image. + +Table 10. Study of the augmentation strategy in TALL. The count column represents the number of blocks on the thumbnail. + +
Window sizeCDFDFDC
(7,7,7,7)85.6073.32
(14,14,14,7)87.6074.32
(28,28,28,7)86.6574.21
+ +Table 11. Ablation studies of window size. All models here are trained without mask augmentation. + +# 5. Conclusion + +This paper presents a novel perspective on detecting deepfake videos using TALL. TALL is both simple and effective, enabling joint spatio-temporal modeling without any additional costs. TALL representation reveals normal deepfake patterns with local-global contextual features. We further propose a new baseline for deepfake video detection called TALL-Swin, which efficiently captures the inconsistencies between deepfake video frames. Extensive experiments demonstrate that TALL-Swin achieves promising results for various unseen deepfake types and strong robustness to a wide range of common corruptions. + +# Acknowledgment + +This work was partially funded by National Natural Science Foundation of China under Grants (62276256, U21B2045 and U20A20223) and Beijing Nova Program under Grant Z211100002121108. The authors wish to thank Huaibo Huang and Lijun Sheng in no particular order, for insightful discussions. + +# References + +[1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In Proc. WIFS, pages 1-7, 2018. 5 +[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proc. ICCV, pages 6836-6846, 2021. 4, 5, 6 +[3] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proc. CVPR, pages 4113-4122, 2022. 5 +[4] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proc. CVPR, pages 6299-6308, 2017. 5 +[5] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Proc. ECCV, pages 103-120, 2020. 6, 7 +[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proc. CVPR, pages 1251-1258, 2017. 5, 6, 7 +[7] deepfakes. Deepfakes. https://github.com/deepfakes/faceswap.2021-11-13.4 + +[8] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv:1708.04552, 2017. 3, 9 +[9] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv:2006.07397, 2020. 4 +[10] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In Proc. CVPR, pages 7865-7874, 2022. 2 +[11] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Ting Zhang, Weiming Zhang, Nenghai Yu, Dong Chen, Fang Wen, and Baining Guo. Protecting celebrities from deepfake with identity consistency transformer. In Proc. CVPR, pages 9468-9478, 2022. 2, 3 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. ICLR, 2021. 2, 4, 5 +[13] Jianwei Fei, Yunshu Dai, Peipeng Yu, Tianrun Shen, Zhihua Xia, and Jian Weng. Learning second order local anomaly for general face forgery detection. In Proc. CVPR, pages 20270-20280, 2022. 2 +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. In Proc. NeurIPS, volume 27, 2014. 1 +[15] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proc. ACM MM, pages 3473-3481, 2021. 1, 2 +[16] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, and Lizhuang Ma. Delving into the local: Dynamic inconsistency learning for deepfake video detection. In Proc. AAAI, pages 744-752, 2022. 1, 2, 5 +[17] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proc. CVPR, 2022. 2, 4, 6, 7 +[18] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In Proc. CVPR, pages 5039-5049, 2021. 2, 4, 5, 6, 7 +[19] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Learning spatio-temporal features with 3d residual networks for action recognition. In Proc. ICCV Workshops, pages 3154-3160, 2017. 5 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, pages 770-778, 2016. 5 +[21] Gee-Sern Hsu, Chun-Hung Tsai, and Hung-Yi Wu. Dual-generator face reenactment. In Proc. CVPR, pages 642-650, 2022. 1 + +[22] Yihao Huang, Felix Juefei-Xu, Qing Guo, Yang Liu, and Geguang Pu. Fakelocator: Robust localization of gan-based face manipulations. IEEE Transactions on Information Forensics and Security, 17:2657-2672, 2022. 1 +[23] Ge-Peng Ji, Guobao Xiao, Yu-Cheng Chou, Deng-Ping Fan, Kai Zhao, Geng Chen, and Luc Van Gool. Video polyp segmentation: A deep learning perspective. MIR, 19(6):531-549, 2022. 2 +[24] Ge-Peng Ji, Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Christos Sakaridis, and Luc Van Gool. Masked vision-language transformer in fashion. MIR, 20(3):421-434, 2023. 2 +[25] Gengyun Jia, Meisong Zheng, Chuanrui Hu, Xin Ma, Yuting Xu, Luoqi Liu, Yafeng Deng, and Ran He. Inconsistency-aware wavelet dual-branch network for face forgery detection. IEEE Transactions on Biometrics, Behavior, and Identity Science, 3(3), 2021. 2 +[26] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proc. CVPR, pages 2889-2898, 2020. 4, 7 +[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. CVPR, pages 4401-4410, 2019. 1 +[28] Sohail Ahmed Khan and Hang Dai. Video transformer for deepfake detection with incremental learning. In Proc. ACM MM, pages 1821-1828, 2021. 3 +[29] Aminollah Khormali and Jiann-Shiun Yuan. Dfdt: An end-to-end deepfake detection framework using vision transformer. Applied Sciences, page 2953, 2022. 3, 5, 6 +[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. 5 +[31] Jiaming Li, Hongtao Xie, Jiahong Li, Zhongyuan Wang, and Yongdong Zhang. Frequency-aware discriminative feature learning supervised by single-center loss for face forgery detection. In Proc. CVPR, pages 6458-6467, 2021. 2, 5 +[32] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proc. CVPR, pages 5001-5010, 2020. 2, 5, 6, 7 +[33] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. In Proc. CVPR Workshops, pages 656-663, 2019. 6 +[34] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proc. CVPR, pages 3207-3216, 2020. 4 +[35] Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: Rethinking face forgery detection in frequency domain. In Proc. CVPR, pages 772-781, 2021. 5 +[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. ICCV, pages 10012-10022, 2021. 2, 3, 4, 5 +[37] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proc. CVPR, pages 8060–8069, 2020. 2 + +[38] MarekKowalski. Faceswap. https://github.com/MarekKowalski/FaceSwap/. 2021-11-13. 4 +[39] Iacopo Masi, Aditya Killekar, Royston Marian Mascarenhas, Shenoy Pratik Gurudatt, and Wael AbdAlmageed. Two-branch recurrent network for isolating deepfakes in videos. In Proc. ECCV, pages 667–684, 2020. 5 +[40] Yisroel Mirsky and Wenke Lee. The creation and detection of deepfakes: A survey. ACM CSUR, 54(1):1-41, 2021. 1 +[41] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In Proc. ICCV, pages 3163-3172, 2021. 5, 6 +[42] Yuval Nirkin, Lior Wolf, Yosi Keller, and Tal Hassner. Deepfake detection based on discrepancies between faces and their context. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6111-6121, 2021. 2 +[43] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In Proc. ECCV, pages 86-103, 2020. 2, 5 +[44] Rameswar Panda Quanfu Fan, Richard Chen. Can an image classifier suffice for action recognition? In ICLR, 2022. 2 +[45] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. Faceforensics: Learning to detect manipulated facial images. In Proc. ICCV, pages 1-11, 2019. 4 +[46] Ekraam Sabir, Jiaxin Cheng, Ayush Jaiswal, Wael AbdAlmageed, Iacopo Masi, and Prem Natarajan. Recurrent convolutional strategies for face manipulation detection in videos. In Proc. CVPR Workshops, pages 80-87, 2019. 6, 7 +[47] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. ICCV, pages 618-626, 2017. 6 +[48] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In Proc. CVPR, pages 11254-11264, 2022. 1 +[49] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proc. CVPR, pages 7672-7682, 2022. 1 +[50] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, pages 6105-6114, 2019. 5 +[51] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM TOG, 38(4):1-12, 2019. 4 +[52] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proc. ICCV, pages 2387-2395, 2016. 4 +[53] Luisa Verdoliva. Media forensics and deepfakes: an overview. IEEE Journal of Selected Topics in Signal Processing, 14(5):910-932, 2020. 1 +[54] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In Proc. CVPR, pages 14923-14932, 2021. 2 + +[55] Junke Wang, Zuxuan Wu, Wenhao Ouyang, Xintong Han, Jingjing Chen, Yu-Gang Jiang, and Ser-Nam Li. M2tr: Multi-modal multi-scale transformers for deepfake detection. In Proc. ICMR, pages 615-623, 2022. 2, 3, 5 +[56] Ping Wang, Kunlin Liu, Wenbo Zhou, Hang Zhou, Honggu Liu, Weiming Zhang, and Nenghai Yu. Adt: Anti-deepfake transformer. In Proc. ICASSP, pages 2899-1903, 2022. 5 +[57] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proc. CVPR, 2020. 6, 7 +[58] Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv:2102.11126, 2021. 2 +[59] Ziming Yang, Jian Liang, Yuting Xu, Xiao-Yu Zhang, and Ran He. Masked relation learning for deepfake detection. IEEE Transactions on Information Forensics and Security, pages 1696-1708, 2023. 2 +[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proc. ICCV, pages 6023-6032, 2019. 9 +[61] Yu Zeng, Zhe Lin, and Vishal M Patel. Sketched: Mask-free local image manipulation with partial sketches. In Proc. CVPR, pages 5951-5961, 2022. 1 +[62] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. Proc. ICLR, 2018. 9 +[63] Yanyi Zhang, Xinyu Li, Chunhui Liu, Bing Shuai, Yi Zhu, Biagio Brattoli, Hao Chen, Ivan Marsic, and Joseph Tighe. Vidtr: Video transformer without convolutions. In Proc. ICCV, pages 13577-13587, 2021. 5, 6 +[64] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: Interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18:1335-1348, 2023. 5, 6 +[65] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep-fake detection. In Proc. CVPR, pages 2185–2194, 2021. 2, 5, 6 +[66] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Self-supervised transformer for deepfake detection. arXiv:2203.01265, 2022. 2, 3 +[67] Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In Proc. CVPR, pages 15023-15033, 2021. 2 +[68] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In Proc. ICCV, pages 15044-15054, 2021. 2, 4, 5, 6, 7 +[69] Xiangyu Zhu, Hao Wang, Hongyan Fei, Zhen Lei, and Stan Z Li. Face forgery detection by 3d decomposition. In Proc. CVPR, pages 2929-2939, 2021. 1 +[70] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proc. ACM MM, pages 2382-2390, 2020. 5 \ No newline at end of file diff --git a/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/images.zip b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b7885749192d5418eacb9c954c0bd276c8b8c8a0 --- /dev/null +++ b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96033745342d36b5bb6bb20c682a5e3671dac4d8648b6dd7c0361c62f4e326b4 +size 774750 diff --git a/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/layout.json b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5d00e1433ebe0e239b6d9bf7307228b4b7cd0256 --- /dev/null +++ b/2023/TALL_ Thumbnail Layout for Deepfake Video Detection/layout.json @@ -0,0 +1,9017 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 473, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 473, + 121 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 473, + 121 + ], + "type": "text", + "content": "TALL:,Thumbsail Layout for Deepfake Video Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "spans": [ + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "text", + "content": "Yuting Xu" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "text", + "content": ", Jian Liang" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "text", + "content": ", Gengyun Jia" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "text", + "content": ", Ziming Yang" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "text", + "content": ", Yanhao Zhang" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "text", + "content": ", Ran He" + }, + { + "bbox": [ + 75, + 141, + 523, + 157 + ], + "type": "inline_equation", + "content": "^{2,4\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 157, + 465, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 157, + 465, + 171 + ], + "spans": [ + { + "bbox": [ + 129, + 157, + 465, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 157, + 465, + 171 + ], + "type": "text", + "content": " Institute of Information Engineering, Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 171, + 481, + 199 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 114, + 171, + 481, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 171, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 114, + 171, + 481, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 171, + 481, + 185 + ], + "type": "text", + "content": " CRIPAC & MAIS, Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 185, + 481, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 185, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 113, + 185, + 481, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 185, + 481, + 199 + ], + "type": "text", + "content": " School of Cyber Security, UCAS " + }, + { + "bbox": [ + 113, + 185, + 481, + 199 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 113, + 185, + 481, + 199 + ], + "type": "text", + "content": " School of Artificial Intelligence, UCAS" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 69, + 199, + 524, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 199, + 524, + 213 + ], + "spans": [ + { + "bbox": [ + 69, + 199, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 69, + 199, + 524, + 213 + ], + "type": "text", + "content": " School of Communications and Information Engineering, NJUPT " + }, + { + "bbox": [ + 69, + 199, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 69, + 199, + 524, + 213 + ], + "type": "text", + "content": " OPPO Research Institute" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 216, + 478, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 216, + 478, + 227 + ], + "spans": [ + { + "bbox": [ + 114, + 216, + 478, + 227 + ], + "type": "text", + "content": "yuting.xu@cripac.ia.ac.cn, liangjian92@gmail.com, rhe@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 281, + 290, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 290, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 290, + 544 + ], + "type": "text", + "content": "The growing threats of deepfakes to society and cybersecurity have raised enormous public concerns, and increasing efforts have been devoted to this critical topic of deepfake video detection. Existing video methods achieve good performance but are computationally intensive. This paper introduces a simple yet effective strategy named Thumbnail Layout (TALL), which transforms a video clip into a pre-defined layout to realize the preservation of spatial and temporal dependencies. Specifically, consecutive frames are masked in a fixed position in each frame to improve generalization, then resized to sub-images and rearranged into a pre-defined layout as the thumbnail. TALL is model-agnostic and extremely simple by only modifying a few lines of code. Inspired by the success of vision transformers, we incorporate TALL into Swin Transformer, forming an efficient and effective method TALL-Swin. Extensive experiments on intra-dataset and cross-dataset validate the validity and superiority of TALL and SOTA TALL-Swin. TALL-Swin achieves " + }, + { + "bbox": [ + 46, + 281, + 290, + 544 + ], + "type": "inline_equation", + "content": "90.79\\%" + }, + { + "bbox": [ + 46, + 281, + 290, + 544 + ], + "type": "text", + "content": " AUC on the challenging cross-dataset task, FaceForensics++ " + }, + { + "bbox": [ + 46, + 281, + 290, + 544 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 281, + 290, + 544 + ], + "type": "text", + "content": " CelebDF. The code is available at https://github.com/rainy-xu/TALL4Deepfake." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 588, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 684 + ], + "type": "text", + "content": "Deepfakes generate and manipulate facial appearances to deceive viewers through generation techniques [61, 48]. With the remarkable success of generative adversarial networks [14, 27], deepfake products have become photorealistic that humans can not distinguish. These deepfake products [21, 49] may be misused for malicious purposes, leading to severe trust issues and security problems, such as financial fraud, identity theft, and celebrity impersonation." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 257, + 540, + 429 + ], + "blocks": [ + { + "bbox": [ + 313, + 257, + 540, + 429 + ], + "lines": [ + { + "bbox": [ + 313, + 257, + 540, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 540, + 429 + ], + "type": "image", + "image_path": "a2648b852e6f75e080588e0ff8d14afab55c72aee5ab1b02647d8b2fd55fcd11.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 434, + 547, + 490 + ], + "lines": [ + { + "bbox": [ + 305, + 434, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 547, + 490 + ], + "type": "text", + "content": "Figure 1. The AUC and FLOPs trade-off of different backbones. Image-level backbones with TALL enjoy comparable accuracy-cost trade-offs with the 3DCNN and video transformer family on the unseen Celeb-DF dataset. All models with the same setting are trained on the FF++ (HQ) dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 547, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 557 + ], + "type": "text", + "content": "ation [53, 40]. The rapid development of social media accelerates the abuse of deepfakes. Therefore, it is crucial to develop advanced detection methods to protect the data privacy of individual users." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Most previous image-based methods [22, 69] perform well on intra-dataset, but their generalizability needs to be improved. Recent research has focused on video-based methods to detect deepfake by modeling spatio-temporal dependencies. There are subtle spatio-temporal inconsistencies between frames since the deepfake algorithms are executed frame by frame. The core of video-level approaches for deepfake detection is capturing inconsistencies through temporal modeling. Existing deepfake video detection methods generally follow two directions. Some methods [15, 16] use two-branch networks or modules to learn spatial and temporal information separately and then fuse them. However, these two-branch approaches may frag" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 246, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 246, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 246, + 703 + ], + "type": "text", + "content": "*This work was done when she was a student in CRIPAC." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22658" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "ment spatiotemporal cooperation and lead to subtle artifacts being neglected. Others directly use classic temporal models such as LSTM and 3D-CNNs. These methods are computationally intensive. The current rise of transformers for vision task backbones has prompted the emergence of corresponding deepfake detection methods. They are accompanied by significant computational complexity that makes them challenging to deploy and use, despite breakthroughs in performance. To enjoy benefits from both image and video methods, we are curious to see whether it is possible to append information about the temporal dimension to the image dimension." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 409 + ], + "type": "text", + "content": "This work develops a simple yet effective Thornail Layout (TALL) for deepfake detection by spatio-temporal modeling. TALL is computationally cheap and retains both temporal and spatial information. In detail, we use dense sampling to extract multiple clips in the video and then randomly select four consecutive frames in the video segment. Subsequently, a block is masked at a fixed position in each frame. Finally, the frames are resized as sub-image and sequentially rearranged into a pre-defined layout as a thumbnail, which has the same size as the clip frames. As shown in Figure 1, TALL brings two advantages compared to the previous spatio-temporal modeling methods for deepfake detection: (1) TALL contains local and global contextual deepfake patterns. (2) TALL is a model-agnostic method for spatio-temporal modeling deepfake patterns at zero computation and zero parameters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 410, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 289, + 531 + ], + "type": "text", + "content": "Furthermore, we discover that the better temporal modeling capabilities backbone has, the better performance TALL achieves. Based on the proposed TALL, we complement a baseline for video deepfake detection based on Swin Transformer [36], called TALL-Swin. We validate TALL-Swin on four popular benchmark datasets, including FaceForensics++, Celeb-DF, DFDC, and DeeperForensics. Our method gains a remarkable improvement over the state-of-the-art approaches. The main contributions of our paper are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 542, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 542, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 542, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 59, + 542, + 287, + 601 + ], + "type": "text", + "content": "- We provide a new perspective for an efficient strategy for video deepfake detection called Thumbnail Layout (TALL), which incorporates both spatial-temporal dependencies, and allows the model to capture spatial-temporal inconsistencies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 616, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 616, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 59, + 616, + 287, + 652 + ], + "type": "text", + "content": "- We propose a spatio-temporal modeling method called TALL-Swin, which efficiently captures the inconsistencies between deepfake video frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 665, + 287, + 713 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the validity of our proposed TALL and TALL-Swin. TALL-Swin outperforms previous methods in both intra-dataset and cross-dataset scenarios." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 91, + 481, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 481, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 481, + 104 + ], + "type": "text", + "content": "2.1. Image-Level Deepfake Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 109, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 547, + 361 + ], + "type": "text", + "content": "Typically, existing deepfake detection methods fall into two categories: image-level and video-level methods. The image-level methods [25, 13] always exploit the artifacts of deepfake images in the spatial domain, such as discrepancies between local regions [42, 59], grid-like structure in frequency space [10], and differences in global texture statistics [37] that provide specific clues to distinguish deepfakes from the real images. F3Net [43] and FDFL [31] utilize the same pipeline that utilizes frequency-aware features and RGB information to capture the traces in different input spaces separately. RFM [54] and Multi-att [65] propose an attention-guided data augmentation mechanism to guide detectors to discover undetectable deepfake clues. Face X-ray [32] and PCL [67] provide effective ways to outline the boundary of the forged face for detecting deepfakes. ICT [11] exploits an identity extraction module to detect identity inconsistency in the suspect image. Similarly, M2tr [55] detects local inconsistencies within frames at different spatial levels. Generally, image-level methods suffer over-fitting issues when a specific technique manipulates the images, and they ignore temporal information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 369, + 479, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 479, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 479, + 382 + ], + "type": "text", + "content": "2.2. Video-Level Deepfake Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 387, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 547, + 640 + ], + "type": "text", + "content": "To improve the generalization of deepfake detectors, many studies generate diversity and generic deepfake data, while other studies capture the temporal incoherence of fake videos as generic clues. Some recent works propose detecting temporal inconsistency using well-designed spatiotemporal neural networks, and others [15, 16] attempt to add modules to image models that capture temporal information. STIL [15] formulates deepfake video detection as a spatial and temporal inconsistency learning process and integrates both spatial and temporal features in a unified 2D CNN framework. FTCN [68] detects temporal-related artifacts instead of spatial artifacts to promote generalization. LipForensics [18] is proposed to learn high-level semantic irregularity in mouth movement in the generated video. RealForensics [17] uses auxiliary data sets during training in exchange for generalization at the cost of higher computational demands. The video-based methods achieve strong generalization but suffer from large computational overhead. To reduce computational costs, we propose TALL which gathers consecutive video frames into thumbnails for learning spatio-temporal consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 647, + 538, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 538, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 538, + 659 + ], + "type": "text", + "content": "2.3. Deepfake Detection with Vision Transformer" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Recently, ViT [12] has achieved impressive performance in computer vision tasks [24, 23, 44]. Many studies extend the ViT for deepfake detection [66, 58]. These methods achieve better performance compared to CNN-based" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22659" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "models, but also sacrifice computational efficiency. A few works [55, 66] attempt to extend the transformer for deepfake detection due to the advent of the visual transformer (ViT) and the impressive ability to model long-range data, different from two-branch architectures that capture short-range and long-range temporal inconsistencies with a single-branch model. ICT [11] aims to detect identity consistency in deepfake video but may fail in detecting face reenactment and entire face synthesis results. DFLL [28] extract the UV texture map to help the transformer to detect deepfakes, which may disrupt the continuity between video frames. DFTD [29] leverages ViT to consider both global and local information but ignores the problem of excessive model arithmetic requirements. Although the transformer-based approaches achieve promising performance, they are accompanied by significant computational complexity that makes them challenging to deploy and use, and the long-range dependencies may be insufficiently exploited in detection models. Swin Transformer [36] produces a hierarchical feature representation and has linear computational complexity concerning input image size, which is suitable as a general-purpose backbone for various vision tasks. In this paper, we cooperate with Swin Transformer to form our robust and efficient method TALL-Swin." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 369, + 105, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 105, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 105, + 382 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 389, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 498 + ], + "type": "text", + "content": "TALL is a deepfake video detection strategy that transforms a video clip into an all-in-one thumbnail without the extra computational overhead. In the following sections, we begin with the motivation of TALL for deepfake detection in Section 3.1. Then we present the technical details of the TALL in Section 3.2. Finally, a generalizable SwINTALL baseline is introduced to explore subtle artifacts in Section 3.3. More details of the layout design are presented in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 503, + 122, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 122, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 122, + 514 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 700 + ], + "type": "text", + "content": "While recent studies have attempted to address noticeable flaws through techniques like slight motion blurring and temporal consistency loss, subtle spatio-temporal artifacts still remain. These artifacts are important for detecting deepfakes, but they introduce two problems: 1) video-based models are less efficient, and 2) analyzing information over long distances may overlook local artifacts, which are critical for deepfake detection. To address these challenges, we propose the TALL strategy, which naturally incorporates temporal information into image-level tasks without disrupting spatial information. This approach enables the image-level model to detect deepfakes in videos. Furthermore, we discovered that TALL provides even greater performance gains when combined with a powerful spatial model, resulting in the TALL-Swin." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "In detail, TALL arranges consecutive frames in the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 69, + 434, + 193 + ], + "blocks": [ + { + "bbox": [ + 312, + 69, + 434, + 193 + ], + "lines": [ + { + "bbox": [ + 312, + 69, + 434, + 193 + ], + "spans": [ + { + "bbox": [ + 312, + 69, + 434, + 193 + ], + "type": "image", + "image_path": "9380b024a1cd3eed710e5f8204bc0c4fabb7b49b9666b2f3aca592a19a7c0fda.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 194, + 418, + 205 + ], + "lines": [ + { + "bbox": [ + 319, + 194, + 418, + 205 + ], + "spans": [ + { + "bbox": [ + 319, + 194, + 418, + 205 + ], + "type": "text", + "content": "(a) Thumbnail Layout (TALL)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 211, + 545, + 233 + ], + "lines": [ + { + "bbox": [ + 305, + 211, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 545, + 233 + ], + "type": "text", + "content": "Figure 2. Illustration of the TALL and shifted window process for computing self-attention in the TALL." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 434, + 72, + 542, + 190 + ], + "blocks": [ + { + "bbox": [ + 434, + 72, + 542, + 190 + ], + "lines": [ + { + "bbox": [ + 434, + 72, + 542, + 190 + ], + "spans": [ + { + "bbox": [ + 434, + 72, + 542, + 190 + ], + "type": "image", + "image_path": "e79a74cc79cb3363864ed670b85229736b93aab6050a999cedcc01393ba14187.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 194, + 526, + 205 + ], + "lines": [ + { + "bbox": [ + 473, + 194, + 526, + 205 + ], + "spans": [ + { + "bbox": [ + 473, + 194, + 526, + 205 + ], + "type": "text", + "content": "(b) Feature Map" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 254, + 546, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 546, + 529 + ], + "type": "text", + "content": "temporal order in a compact " + }, + { + "bbox": [ + 304, + 254, + 546, + 529 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 254, + 546, + 529 + ], + "type": "text", + "content": " layout, in line with the calculation theory of convolution and shifted window. TALL contains both spatial and temporal information so that model can learn both intra-frame artifacts and interframe inconsistency and obtains comparable performance to video-based methods. Here we use the shifted window to explain TALL's mechanism. As illustrated in Figure 2 (a), the model computes self-attention while accounting for spatial dependencies across sub-images (represented by the solid red box). When the window spans multiple sub-images (represented by the red dash box), the model is able to capture temporal inconsistencies between frames. Moreover, TALL leverages both local and global contexts of deepfake patterns to ensure robust modeling capabilities for short and long-range spatial dependencies. Compared to previous methods, we anticipate that TALL strikes a balance between speed and accuracy, sacrificing a little spatial information while preserving performance. Based on the fact that attention-based models are better at handling contextual features and that the Swin-Transformer uses shifted windows to reduce computation and memory, we further complement TALL-Swin baseline for video deepfake detection." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 537, + 456, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 456, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 456, + 551 + ], + "type": "text", + "content": "3.2. Thumbnail Layout (TALL)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": "Given a video " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{T \\times C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " is the frame length of the video, " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " is the number of channels, and " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " is the resolution of the frames. Assuming each video contains " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " clips, we divide a video into " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " equal segments of length " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "T / N" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " and then sample consecutive " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " (set to 4 by default) frames from the segments at random locations to form one clip. Then, the thumbnail " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " is rearranged of sub-images " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "(C \\times \\frac{H}{\\sqrt{t}} \\times \\frac{W}{\\sqrt{t}})" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " that are resized from the above " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " frames. To maximize the utility of TALL, we mask the organized " + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 555, + 545, + 713 + ], + "type": "text", + "content": " square masks of the thumbnail. It is based on two core designs: 1) The position of the masks is random between different sub-images, which retains the advantages of the Cutout [8] that encourages the network to focus more" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22660" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 92, + 276, + 336 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 285, + 84 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 285, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 285, + 84 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode of TALL in a PyTorch-like style." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 92, + 276, + 336 + ], + "lines": [ + { + "bbox": [ + 47, + 92, + 276, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 276, + 336 + ], + "type": "text", + "content": "# x: one clip of video (T*C*H*W) \n# T: frame number of clip \n# C: channels; s: mask size \n# d: length of frame included in the thumbnail \n# r: rows of thumbnail \n# x_tall: thumbnail image (224*224) \n#TALL's augmentation strategy \nh = np.random.randint(H) \nw = np.random.randint(W) \n#the mask position is fixed for each frame \nm = np.ones((H, W)) \nh1 = np.clip(h - s // 2, 0, H) \nh2 = np.clip(+s // 2, 0, H) \nw1 = np.clip(w - s // 2, 0, W) \nw2 = np.clip(w + s // 2, 0, W) \nm[h1: h2, w1: w2] = 0 \nm = torch.from_numpy(m) \nm = maskexpand_as(x) \nx = x * m \n#TALL: generation of the thumbnail \nx = x.view(-1, H, W).unsqueeze(0) \nx = x.view((-1, C*d) + x_tall.size() [2:]) \nx = rearrange(x, 'b (th tw c) h w -> b c (th h) (tw w)', th=r, c=C) \nx_tall = interpolate(x, size=H)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 365, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 460 + ], + "type": "text", + "content": "on complementary and less prominent features. 2) We fix the position of the mask within a clip to take advantage of the fact that most deepfake videos are frame-by-frame tampered with, thus forcing the model to detect inconsistencies between adjacent frames of the deepfake videos. We do not allow the mask to appear on the seams of the thumbnail but allow for partial mask inclusion in the thumbnail. The detailed procedure of TALL is summarized in Algorithm 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 468, + 126, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 126, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 126, + 479 + ], + "type": "text", + "content": "3.3. TALL-Swin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": "To balance efficiency and model performance for spatiotemporal feature learning and to leverage the benefits of attention-based models, we enhanced a baseline deepfake detection model called TALL-Swin by incorporating the Swin Transformer [36]. Given the characteristics of TALL, we slightly modified the window size of Swin-B in TALL-Swin. We first enlarge the window size of the first three stages of the model so that the interaction between frames in the thumbnail becomes more frequent, forcing the model to learn more detailed spatio-temporal dependencies. Next, we set the window size of the last stage to be the same as the feature map size, enabling the window to perform global attention computations while TALL-Swin captures global spatial-temporal dependencies. As a result, the size of the last layer of the feature map became smaller, reducing the window size without introducing any additional computational overhead. Consequently, the window sizes for the four stages of TALL-Swin are [14, 14, 14, 7]. Note that the patch merging process makes TALL-Swin captures a more" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "comprehensive range of dependencies through hierarchical representations, as shown in Figure 2 (b)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 95, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 180 + ], + "type": "text", + "content": "Given a video of length " + }, + { + "bbox": [ + 304, + 95, + 545, + 180 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 95, + 545, + 180 + ], + "type": "text", + "content": ", each frame contains " + }, + { + "bbox": [ + 304, + 95, + 545, + 180 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 95, + 545, + 180 + ], + "type": "text", + "content": " patches, and the window contains " + }, + { + "bbox": [ + 304, + 95, + 545, + 180 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 95, + 545, + 180 + ], + "type": "text", + "content": " patches. To demonstrate the superiority of TALL-Swin in terms of computational consumption, we show below the computational complexity of the image-level transformer and video-level transformer, including ViT [12], Swin [36], ViViT [2], and TALL-Swin respectively:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 350, + 188, + 471, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 188, + 471, + 201 + ], + "spans": [ + { + "bbox": [ + 350, + 188, + 471, + 201 + ], + "type": "interline_equation", + "content": "\\Omega_ {\\mathrm {V i T}} = 4 T N C ^ {2} + 2 T N ^ {2} C,", + "image_path": "43ffb9f1d8565af2e90f0a21e6cbb3fb71575b69ea202af3cf09baba75f322ae.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 350, + 202, + 545, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 202, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 350, + 202, + 545, + 221 + ], + "type": "interline_equation", + "content": "\\Omega_ {\\mathrm {S w i n}} = 4 T N C ^ {2} + 2 T P N C, \\tag {1}", + "image_path": "194827f9b86bc6c77cb19ef50b6591e67cdb4d5590c4acbb000cf391ee0a3d49.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 350, + 217, + 479, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 217, + 479, + 230 + ], + "spans": [ + { + "bbox": [ + 350, + 217, + 479, + 230 + ], + "type": "interline_equation", + "content": "\\Omega_ {\\mathrm {V i V i T}} = 4 T N C + 2 T ^ {2} N ^ {2} C,", + "image_path": "951c35a1eb8f1647e04d055f3e3635f37976a8572bd2b5920a9d7677d7b1548f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 232, + 500, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 232, + 500, + 247 + ], + "spans": [ + { + "bbox": [ + 351, + 232, + 500, + 247 + ], + "type": "interline_equation", + "content": "\\Omega_ {\\mathrm {T A L L - S w i n}} = T N C ^ {2} + \\frac {1}{2} T P N C.", + "image_path": "6b0b0e2bb926bb440848687b7203d291f803fa405f9c37b7b3b52db91fac5047.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 254, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 545, + 312 + ], + "type": "text", + "content": "TALL-Swin has the lowest computational complexity compared to image and video-level transformer methods. Subsequent experiments will demonstrate that TALL-Swin maintains performance, albeit at the sacrifice of some spatial information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 314, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 545, + 337 + ], + "type": "text", + "content": "The cross-entropy loss is employed to optimize the TALL-Swin, which is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 343, + 545, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 343, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 310, + 343, + 545, + 371 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C E} = - \\frac {1}{n} \\sum_ {i = 1} ^ {n} y _ {i} \\log \\mathcal {F} \\left(x _ {i}\\right) + (1 - y _ {i}) (\\log (1 - \\mathcal {F} \\left(x _ {i}\\right)), \\tag {2}", + "image_path": "5ec6ab44421013a726430654631cdcbbda5c2b3d80669fb0cbc987f7e06f6e3a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "type": "text", + "content": " indicates input clip, " + }, + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "type": "text", + "content": " denotes the label of clip, " + }, + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "type": "text", + "content": " is the number of clip, " + }, + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 377, + 545, + 402 + ], + "type": "text", + "content": " is TALL-Swin." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 412, + 387, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 387, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 387, + 426 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 432, + 356, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 356, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 356, + 445 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "content": "Datasets. Following previous works [18, 17, 68], we evaluate the TALL and TALL-Swin on four widely used datasets. FaceForensics++ [45] is a most-used benchmark on intra-dataset deepfake detection, consisting of 1,000 real videos and 4,000 fake videos in four different manipulations: DeepFake [7], FaceSwap [38], Face2Face [52], and NeuralTextures [51]. Besides, FaceForensics++ contains multiple video qualities, e.g. high quality (HQ), low quality (LQ) and RAW. Celeb-DF (CDF) [34] is a popular benchmark on cross-dataset, which contains 5,693 deepfake videos generated from celebrities. The improved compositing process was used to improve the various visual artifacts presented in the video. Celeb-DF is also suitable for deepfake detection tasks with a reference set. DFDC [9] is a large-scale benchmark developed for Deepfake Detection Challenge. This dataset includes 124k videos from 3,426 paid actors. The existing deepfake detection methods do perform not very well on DFDC due to their sophisticated deepfake techniques. DeeperForensics (DFO) [26] includes 60,000 videos with 17.6 million frames for deepfake detection, whose videos vary in identity, pose, expression, emotion, lighting conditions, and blend shape with high quality." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22661" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 282, + 289 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 282, + 289 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 282, + 289 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 282, + 289 + ], + "type": "table", + "html": "
ModelsTemp.CDFDFDCFLOPsParamsPT
I3D-RGB* [4]✓78.2465.58222.7G25M1K
R3D-50* [19]✓79.6367.73296.6G46M1K
ResNet50* [20]×76.3864.0125.5G21M1K
+TALL✓80.9065.5425.5G21M1K
EffNetB4* [50]×78.1966.818.3G19M1K
+TALL✓83.3767.158.3G19M1K
VTN [41]✓83.2073.50296.6G46M21K
VidTR [63]✓83.3073.30117G93M21K
ViViT* [2]✓86.9674.61628G310M21K
ISTVT [64]✓84.1074.20455.8G--
ViT-B* [12]×82.3372.6455.4G84M21K
+TALL✓86.5874.1055.4G84M21K
Swin-B* [36]×83.1373.0147.5G86M21K
TALL-Swin✓90.7976.7847.5G86M21K
", + "image_path": "d47850939f4b6a3d246b79ba545e9a3edbbf95adcb1b8dad80bb055de495114d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 293, + 287, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 381 + ], + "type": "text", + "content": "Table 1. Performance of different backbones. TALL consistently improves the accuracy over different image-level models. We show the AUC, FLOPs, and number of parameters for each model on the cross-dataset scenario. All models are trained on " + }, + { + "bbox": [ + 46, + 293, + 287, + 381 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + +" + }, + { + "bbox": [ + 46, + 293, + 287, + 381 + ], + "type": "text", + "content": " (HQ). " + }, + { + "bbox": [ + 46, + 293, + 287, + 381 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 46, + 293, + 287, + 381 + ], + "type": "text", + "content": " indicates the model enables temporal modeling. \\* indicates our implementation. PT indicates pre-train. 1K and 21K indicate the model pre-trained on ImageNet-1K and 21K respectively. The best results are bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "type": "text", + "content": "Implementation Details. We use MTCNN to detect face for each frame in the deepfake videos, only extract the maximum area bounding box and add " + }, + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "type": "text", + "content": " face crop size from each side as in LipForensics [18]. The ImageNet-21K pretrained Swin-B model is used as our backbone. Excluding ablation experiments, we sample 8 clips using dense sampling, each clip contains 4 frames. The size of the thumbnail is " + }, + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 400, + 287, + 616 + ], + "type": "text", + "content": ". Following Swin Transformer [36], Adam [30] optimization is used with a learning rate of 1.5e-5 and batch size of 4, using a cosine decay learning rate scheduler and 10 epochs of linear warm-up. We adopt Acc. (accuracy) and AUC (Area Under Receiver Operating Characteristic Curve) as the evaluation metrics for extensive experiments. To ensure a fair comparison, we calculate video-level predictions for the image-based method and average the predictions across the entire video (following previous works [18, 16, 35, 68]). Note that results are directly cited from published papers if we follow the same setting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 623, + 180, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 180, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 180, + 635 + ], + "type": "text", + "content": "4.2. Scaling over Backbones" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "To verify our assumption, we adopt several image-level backbones commonly used for deepfake detection for comparison with the video-level backbones. As shown in Table 1 above the double horizontal line, we first compare the accuracy and complexity of the CNN-based video and image backbones. Although I3D [4] and R3D [19] achieve" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "better performance than vanilla ResNet50 [20] and EfficientNet [50], the computation costs are huge, such as R3D-50 with 296G FLOPs. For ResNet and EfficientNet who added TALL, ResNet achieves better AUC both on CDF (76.38 VS 80.93) and DFDC (64.01 VS 65.54) datasets, and EfficientNet achieves " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "5.18\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " better AUC on CDF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "content": "The second section contains the video and image transformers. Compared to video transformers, the image-based ViT and Swin fail to achieve better performance due to the lack of temporal modeling. For example, ViViT achieves " + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "inline_equation", + "content": "86.96\\%" + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "content": " AUC on CDF, which is " + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "content": " higher than Swin although ViViT with " + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "inline_equation", + "content": "13\\times" + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "content": " more computation. By way of contrast, " + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "inline_equation", + "content": "\\mathrm{ViT + TALL}" + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "content": " achieves " + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "inline_equation", + "content": "86.58\\%" + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "content": " AUC on CDF with 55.4G FLOPs, which is comparable to AUC with ViViT but with low computation. Accordingly, Swin's performance was significantly improved with the addition of TALL without computation increment. On the other hand, TALL boosts higher performance on models with learned long-range dependencies. e.g., ResNet+TALL " + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "inline_equation", + "content": "(+4.5\\%)" + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "content": " on CDF and " + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "inline_equation", + "content": "+1.5\\%" + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "content": " on DFDC) vs. Swin+TALL " + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "inline_equation", + "content": "(+7.6\\%)" + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "content": " on CDF and " + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "inline_equation", + "content": "+3.6\\%" + }, + { + "bbox": [ + 304, + 150, + 546, + 365 + ], + "type": "text", + "content": " on DFDC). These two section results demonstrate that TALL provides both spatial and temporal information and enables the model to learn spatial and temporal inconsistencies for video deepfake detection." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 388, + 544, + 657 + ], + "blocks": [ + { + "bbox": [ + 307, + 388, + 544, + 657 + ], + "lines": [ + { + "bbox": [ + 307, + 388, + 544, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 544, + 657 + ], + "type": "table", + "html": "
MethodsFF++ (HQ)FF++ (LQ)
Acc.AUCAcc.AUC
MesoNet [1]83.10-70.47-
Xception [6]95.7396.3086.8689.30
Face X-ray [32]-87.35-61.60
Two-branch [39]96.4398.7086.3486.59
Add-Net [70]96.7897.7487.5091.01
F3-Net [43]97.5298.1090.4390.43
FDFL [31]96.6999.3089.0092.40
Multi-Att [65]97.6099.2988.6990.40
RECCE [3]97.0699.3291.0395.02
LipForensics [18]98.8099.7094.2098.10
DFDT [29]98.1899.2692.6794.48
ADT [56]92.0596.3081.4882.52
ST-M2TR [55]-99.42-95.31
VTN [41]98.47-94.02-
VidTR [63]97.42-92.12-
ViViT* [2]92.60-88.02-
ISTVT [64]99.00-96.15-
TALL-Swin98.6599.8792.8294.57
", + "image_path": "63ab115643577cd65dd2d27bef29b6d0783a3ffd068bc9fab1d7656687ad2f2a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 304, + 660, + 545, + 693 + ], + "lines": [ + { + "bbox": [ + 304, + 660, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 660, + 545, + 693 + ], + "type": "text", + "content": "Table 2. Intra-dataset evaluations. We report the video-level Acc. " + }, + { + "bbox": [ + 304, + 660, + 545, + 693 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 304, + 660, + 545, + 693 + ], + "type": "text", + "content": " and AUC " + }, + { + "bbox": [ + 304, + 660, + 545, + 693 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 304, + 660, + 545, + 693 + ], + "type": "text", + "content": " on the FF++ dataset. HQ indicates high quality, and LQ indicates low quality." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22662" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 270, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 270, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 270, + 85 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 289, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 289, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 289, + 270 + ], + "type": "text", + "content": "Intra-dataset evaluations. Following ISTVT [64], we show the results of the " + }, + { + "bbox": [ + 46, + 90, + 289, + 270 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 46, + 90, + 289, + 270 + ], + "type": "text", + "content": " dataset under both Low Quality (LQ) and High Quality (HQ) videos, and report comparisons against several advanced methods in Table 2. We can observe that advanced video-based transformers have better results than CNN-based methods. Compared to video-based transformer methods, TALL-Swin has comparable performance and lower consumption to the previous video transformer method with HQ settings. However, TALL-Swin gets unsatisfactory results with the LQ setting. The LQ setting is obtained by severely compressing the videos. So the reason for the result may be that TALL scales the frame to a smaller size, causing more spatial information to be lost in the frame. We will investigate the possibility of other designs to further improve performance in the LQ setting." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 51, + 281, + 282, + 509 + ], + "blocks": [ + { + "bbox": [ + 51, + 281, + 282, + 509 + ], + "lines": [ + { + "bbox": [ + 51, + 281, + 282, + 509 + ], + "spans": [ + { + "bbox": [ + 51, + 281, + 282, + 509 + ], + "type": "table", + "html": "
MethodCDFDFDCFShDFoAvg.
Xception [6]73.7070.9072.0084.5075.28
CNN-aug [57]75.6072.1065.7074.4071.95
CNN-GRU [46]69.8068.9080.8074.1073.40
Patch-based [5]69.6065.6057.8081.8068.70
Face X-Ray [32]79.5065.5092.8086.8081.15
Multi-Att [65]75.7068.1066.0077.7071.88
DSP-FWA [33]69.5067.3065.5050.2063.13
LipForensics [18]82.4073.5097.1097.6087.65
FTCN [68]86.9074.0098.8098.8089.63
RealForensics [17]86.9075.9099.7099.3090.45
DFDT [29]88.3076.1097.8096.9089.70
VTN [41]83.2073.5098.7097.7088.30
VidTR [63]83.5073.3098.0097.9088.10
ViViT* [2]86.9674.6199.4199.1990.05
ISTVT [64]84.1074.2099.3098.6089.10
TALL-Swin90.7976.7899.6799.6291.71
", + "image_path": "02e03602c7d07f0597c648418a18420c6be6b44a7ba86c65d1742993416db232.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 512, + 287, + 545 + ], + "lines": [ + { + "bbox": [ + 46, + 512, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 287, + 545 + ], + "type": "text", + "content": "Table 3. Generalization to unseen datasets. We report the video-level AUC (\\%) on four unseen datasets: Celeb-DF (CDF), DFDC, FaceShifter (FSh), and DeeperForensics (DFo)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "content": "Generalization to unseen datasets. In addition to the intra-dataset comparisons, we also investigate the generalization ability of our method. Adhering to the deepfake video detection cross-dataset protocol [18], we train a model on " + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "content": " (HQ) then test on Celeb-DF (CDF), DFDC, FaceShifter (FSh), and DeeperForensics (DFo) datasets. As shown in Table 3: (1) Video-based methods generally have better results than image-based methods, which shows that temporal information is helpful for the deepfake video detection task. For example, Lip outperforms Face X-ray's AUC by a wide margin. In addition, most transformer-based models have higher performance than CNN-based models. For the transformer-based models, both achieved" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": "an average AUC of " + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": ", while the best CNN-based video-level models only achieved " + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": ". (2) TALL-Swin achieves state-of-the-art results on Celeb-DF, DFDC, and Deeper-Forensics datasets, and also beats its competitors on Celeb-DF dataset by a large margin " + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "inline_equation", + "content": "(3.8\\%)" + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": ". The results demonstrate that TALL-Swin performs well when encountering unseen datasets with better generalization ability than previous video transformer methods." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 176, + 544, + 497 + ], + "blocks": [ + { + "bbox": [ + 313, + 176, + 544, + 497 + ], + "lines": [ + { + "bbox": [ + 313, + 176, + 544, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 544, + 497 + ], + "type": "image", + "image_path": "67bba2563ef26452db382cf417d3f1da6d2be31f271f9fa32b71b43dc6e1e2e1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 498, + 547, + 533 + ], + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 533 + ], + "type": "text", + "content": "Figure 3. Saliency map visualization of TALL-Swin on different datasets. The first four rows of samples are from the FF++ dataset, and the last four rows are from the unseen datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "Analysis of saliency map visualization. We adopt Grad-CAM [47] to visualize where the TALL-Swin is paying its attention to the deepfake faces. In Figure 3, we give the results on intra-dataset and cross-dataset scenarios. All models are trained on " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + +" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": " (HQ). It can be observed in the first four rows of Figure 3 that TALL-Swin captures method-specific artifacts. Note that the DF transfers the face region from a source video to a target, and the NT only modifies the facial expressions corresponding to the mouth region. TALL-Swin corresponds to focus on the face region and the mouth region. Furthermore, our model traces the more generalized artifacts that are independent of manipulation methods, e.g., blending boundaries (CDF), and abnormal motions in the clip (DFDC, Fsh, Dfo)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22663" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 90, + 181, + 212 + ], + "blocks": [ + { + "bbox": [ + 109, + 73, + 501, + 83 + ], + "lines": [ + { + "bbox": [ + 109, + 73, + 501, + 83 + ], + "spans": [ + { + "bbox": [ + 109, + 73, + 501, + 83 + ], + "type": "text", + "content": "Xception Face X-ray ICT Patch-based CNN-GRU LipForensics TALL-Swin" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 50, + 90, + 181, + 212 + ], + "lines": [ + { + "bbox": [ + 50, + 90, + 181, + 212 + ], + "spans": [ + { + "bbox": [ + 50, + 90, + 181, + 212 + ], + "type": "image", + "image_path": "a5b466759f3094b0caecd7262ea43e40088af880ef34f19ca52484688d391b08.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 182, + 90, + 302, + 213 + ], + "blocks": [ + { + "bbox": [ + 182, + 90, + 302, + 213 + ], + "lines": [ + { + "bbox": [ + 182, + 90, + 302, + 213 + ], + "spans": [ + { + "bbox": [ + 182, + 90, + 302, + 213 + ], + "type": "image", + "image_path": "b9c8c8662d00ed5404ab6de3103706e6d41d22a449160c4a34fd368fdebe3a9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 90, + 423, + 213 + ], + "blocks": [ + { + "bbox": [ + 304, + 90, + 423, + 213 + ], + "lines": [ + { + "bbox": [ + 304, + 90, + 423, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 423, + 213 + ], + "type": "image", + "image_path": "a7b257598955b38b8906a145e1afced6ed0b20b2ce4e1c02ecef460c9f66632f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 425, + 90, + 544, + 213 + ], + "blocks": [ + { + "bbox": [ + 425, + 90, + 544, + 213 + ], + "lines": [ + { + "bbox": [ + 425, + 90, + 544, + 213 + ], + "spans": [ + { + "bbox": [ + 425, + 90, + 544, + 213 + ], + "type": "image", + "image_path": "8f894b31f0b4f348eb464cab653a19b18908a33ca7358a86a0486885792461b6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 214, + 181, + 346 + ], + "blocks": [ + { + "bbox": [ + 49, + 214, + 181, + 346 + ], + "lines": [ + { + "bbox": [ + 49, + 214, + 181, + 346 + ], + "spans": [ + { + "bbox": [ + 49, + 214, + 181, + 346 + ], + "type": "image", + "image_path": "1288e83695a1a47eafc8e248fd431a864848075e0e0f2b481c6b972c2ef75ab7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "type": "text", + "content": "Figure 4. Robustness to various unseen corruptions. We report the video-level AUC " + }, + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 350, + 546, + 384 + ], + "type": "text", + "content": " of our methods under five different levels of seven particular types of corruption. \"Average\" denotes the mean across all corruptions at each severity level. Our TALL-Swin is more robust than previous methods for all corruptions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 182, + 214, + 302, + 346 + ], + "blocks": [ + { + "bbox": [ + 182, + 214, + 302, + 346 + ], + "lines": [ + { + "bbox": [ + 182, + 214, + 302, + 346 + ], + "spans": [ + { + "bbox": [ + 182, + 214, + 302, + 346 + ], + "type": "image", + "image_path": "9471f592a87e976ec6610e2702a13b8da8f7c970ae161d7246166f975e8baca1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 303, + 214, + 423, + 346 + ], + "blocks": [ + { + "bbox": [ + 303, + 214, + 423, + 346 + ], + "lines": [ + { + "bbox": [ + 303, + 214, + 423, + 346 + ], + "spans": [ + { + "bbox": [ + 303, + 214, + 423, + 346 + ], + "type": "image", + "image_path": "b21159c95b3a03adbf4842f84a81ce5140cd6d08b6d0aa5ed8025853777083c5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 425, + 214, + 544, + 346 + ], + "blocks": [ + { + "bbox": [ + 425, + 214, + 544, + 346 + ], + "lines": [ + { + "bbox": [ + 425, + 214, + 544, + 346 + ], + "spans": [ + { + "bbox": [ + 425, + 214, + 544, + 346 + ], + "type": "image", + "image_path": "e65bbfe61523b63ee11bbbf46dccf1e03b5dee93a39b68ce6d023f8ae9db92d7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 52, + 393, + 541, + 536 + ], + "blocks": [ + { + "bbox": [ + 52, + 393, + 541, + 536 + ], + "lines": [ + { + "bbox": [ + 52, + 393, + 541, + 536 + ], + "spans": [ + { + "bbox": [ + 52, + 393, + 541, + 536 + ], + "type": "table", + "html": "
MethodCleanSaturationContrastBlockNoiseBlurPixelCompressAvg.
Xception (ICCV'19) [6]99.899.398.699.753.860.274.262.178.3
CNN-GRU (CVPRW'19) [46]99.999.098.897.947.971.586.574.582.3
CNN-aug (CVPR'20) [57]99.899.399.195.254.776.591.272.584.1
Patch-based (ECCV'20) [5]99.984.374.299.250.054.456.753.467.5
Face X-ray (CVPR'20) [32]99.897.688.599.149.863.888.655.277.5
LipForensics (ICCV'21) [18]99.999.999.687.473.896.195.695.692.5
FTCN (ICCV'21) [68]99.499.496.797.153.195.898.286.489.5
RealForensics (CVPR'22) [17]99.899.899.698.979.795.398.497.695.6
TALL-Swin w/o mask100.0100.0100.099.883.597.398.497.996.7
TALL-Swin100.0100.0100.0100.085.397.698.598.197.1
", + "image_path": "31ce922a2577f7fb9a1a77063a9d3e92be408c7ccc26bbb7bfcf14d115aff431.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 539, + 546, + 563 + ], + "lines": [ + { + "bbox": [ + 46, + 539, + 546, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 546, + 563 + ], + "type": "text", + "content": "Table 4. Average robustness to unseen corruptions. Average Video-level AUC (%) across five intensity levels for each corruption type proposed in DFo [26]. \"Avg\" indicates the mean across all corruptions and all levels." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Robustness to unseen perturbations. Deepfake detectors must be robust to common perturbations, given that video propagation on social media causes video compression, noise addition, etc. We also study the performance of robustness to unseen perturbations. Following RealForensics [17], the experiment applies seven unseen perturbations to fake videos at five intensity levels. In Figure 4, we show results of increasing the severity of each corruption. We can observe that other methods degrade dramatically as the perturbations become more severe. TALL-Swin still has a high performance. However, TALL-Swin degrades when" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 702 + ], + "type": "text", + "content": "the Gaussian noise reaches level five. Table 4 presents the average AUC across all intensity levels for corruption types. We observe that our method is significantly more robust to most perturbations than other methods. The good robustness may be from both the design of TALL and the proposed mask augmentation. The main reason may be the consecutive multi-frame input. We empirically consider that the key to deepfake detection is local inconsistency, the continuous frame design has less redundant information, ensuring that the model finds locally important clues." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22664" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 70, + 93, + 134 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 93, + 134 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 93, + 134 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 93, + 134 + ], + "type": "image", + "image_path": "c8a499cba746e8873144849dd2f06d2a7bfb91df4a44f8e8d85499bad080da71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 136, + 89, + 146 + ], + "lines": [ + { + "bbox": [ + 78, + 136, + 89, + 146 + ], + "spans": [ + { + "bbox": [ + 78, + 136, + 89, + 146 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 105, + 101, + 170, + 132 + ], + "blocks": [ + { + "bbox": [ + 105, + 101, + 170, + 132 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 170, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 170, + 132 + ], + "type": "image", + "image_path": "c3fb8489399ebb8d07696b41d794cdcb57ea4bf936005c3bc7d07cc0e1403dfc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 154, + 259, + 166 + ], + "lines": [ + { + "bbox": [ + 74, + 154, + 259, + 166 + ], + "spans": [ + { + "bbox": [ + 74, + 154, + 259, + 166 + ], + "type": "text", + "content": "Figure 5. Illustration of different layout designs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 182, + 85, + 216, + 132 + ], + "blocks": [ + { + "bbox": [ + 182, + 85, + 216, + 132 + ], + "lines": [ + { + "bbox": [ + 182, + 85, + 216, + 132 + ], + "spans": [ + { + "bbox": [ + 182, + 85, + 216, + 132 + ], + "type": "image", + "image_path": "9dd06e16684826bf5dff944678b7f2b73766ac45bb6fc7836e624b0dd6e9004d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 228, + 85, + 261, + 132 + ], + "blocks": [ + { + "bbox": [ + 228, + 85, + 261, + 132 + ], + "lines": [ + { + "bbox": [ + 228, + 85, + 261, + 132 + ], + "spans": [ + { + "bbox": [ + 228, + 85, + 261, + 132 + ], + "type": "image", + "image_path": "e6a081a7e563d69a532187f0d2fe7af46d89f58a5e4d20d9cae25f9fd9692e0a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 186, + 140, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 140, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 140, + 198 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "text", + "content": "We perform the ablation study to analyze the effects of each component and hyper-parameter in TALL-Swin. All experiments are trained on " + }, + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "text", + "content": " (HQ) and tested on the CDF and DFDC datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 251, + 287, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 371 + ], + "type": "text", + "content": "Effects of different layouts. We train a TALL-Swin model on " + }, + { + "bbox": [ + 46, + 251, + 287, + 371 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++(\\mathrm{HQ})" + }, + { + "bbox": [ + 46, + 251, + 287, + 371 + ], + "type": "text", + "content": " for each layout illustrated in Figure 5, to analyze in which layout of the thumbnails the model learns the strongest generalization of the spatial-temporal dependence of the deepfake patterns. As shown in Table 5, the model with a compact layout like Figure 5 (d) has good generalization ability on the unseen datasets. A compact layout like Figure 5 (d) may help the model to learn the temporal dependence across frames because such a form provides the shortest distance between any two images." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 104, + 380, + 230, + 449 + ], + "blocks": [ + { + "bbox": [ + 104, + 380, + 230, + 449 + ], + "lines": [ + { + "bbox": [ + 104, + 380, + 230, + 449 + ], + "spans": [ + { + "bbox": [ + 104, + 380, + 230, + 449 + ], + "type": "table", + "html": "
LayoutCDFDFDC
Figure 5 (a)85.5270.02
Figure 5 (b)84.9373.57
Figure 5 (c)86.6672.12
Figure 5 (d)87.6074.32
", + "image_path": "676730e8cb33e7f21524d0f229c053981161edc42ba4cde84c994b60c10e80c9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": "Study on the numbers of sub-image. We use Swin-B as the baseline for this study to compare the effect of different thumbnail layout schemes on the model's generalization ability. Changing frames to thumbnails involves scaling, so we also investigate the impact of resizing and random cropping pre-processing on model performance. We set up four variants: resizing pre-process with " + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": " layout, " + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": " layout and " + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": " layout; random cropping pre-process with " + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": " layout. As shown in Table 6, the model performance degrades sharply when using " + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": " layout. This may be due to the small size of each sub-image that the spatial information is not captured well by the model. The result of " + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": " layout also slightly decreases. " + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": " layout with resizing pre-processing beats " + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": " layout with random crop. We also found that TALL-Swin achieves the best performance and the AUC score increases " + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": " compared to the baseline, suggesting that thumbnails in a " + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 486, + 287, + 700 + ], + "type": "text", + "content": " layout are more helpful to TALL-Swin than original frames." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Effects of Sub-image's size. We eliminate the scaling" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 337, + 70, + 511, + 155 + ], + "blocks": [ + { + "bbox": [ + 47, + 452, + 287, + 474 + ], + "lines": [ + { + "bbox": [ + 47, + 452, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 287, + 474 + ], + "type": "text", + "content": "Table 5. Effects of different layouts. All models here are trained without mask augmentation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 70, + 511, + 155 + ], + "lines": [ + { + "bbox": [ + 337, + 70, + 511, + 155 + ], + "spans": [ + { + "bbox": [ + 337, + 70, + 511, + 155 + ], + "type": "table", + "html": "
Pre-processLayoutCDFDFDC
None-83.1373.01
Resize4 × 480.1870.45
Resize3 × 383.1872.98
Crop2 × 278.5573.30
Resize2 × 287.6074.32
", + "image_path": "903cc1a1261f9e2e0d900ff2cc3e4a467366129238e62d8387e29832aeee4ea8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 201, + 547, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 547, + 308 + ], + "type": "text", + "content": "operation for sub-images to allow for more flexible layout settings. However, we've observed that when the number of sub-images grows at their original size, the performance improvements are only slight. Additionally, the computational complexity increases dramatically with the number of frames ( 4.3 times more than the TALL setting), as demonstrated in Table 7. To strike a balance between performance and computational complexity, we reduce the resolution of sub-images in TALL." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 311, + 318, + 538, + 380 + ], + "blocks": [ + { + "bbox": [ + 305, + 158, + 545, + 179 + ], + "lines": [ + { + "bbox": [ + 305, + 158, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 158, + 545, + 179 + ], + "type": "text", + "content": "Table 6. Study on the numbers of sub-image. All models here are trained without mask augmentation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 311, + 318, + 538, + 380 + ], + "lines": [ + { + "bbox": [ + 311, + 318, + 538, + 380 + ], + "spans": [ + { + "bbox": [ + 311, + 318, + 538, + 380 + ], + "type": "table", + "html": "
Subimage-sizeLayoutFLOPsCDFDFDC
224 × 2243 × 3253G88.6975.98
224 × 2242 × 2185G88.1575.01
112 × 1122 × 247.5G87.6074.32
", + "image_path": "9a72269dcd8e31ee759a5b2ff353d66f785217f6c50acb9689b4286f2c6daf9f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 417, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 536 + ], + "type": "text", + "content": "Study on absence and order of thumbnails. In this case, we study the impact of missing the last sub-image and the last two sub-images on the model's performance. The first two rows of Table 8 show that all four sub-images contribute to the model performance. Besides, we set the order of the different thumbnails to evaluate the TALL-Swin. We consider three orders: forward, reverse, and random. Forward order performs the best for three different orders. This may be because of the positional encoding of different frames in TALL-Swin." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 351, + 545, + 496, + 626 + ], + "blocks": [ + { + "bbox": [ + 306, + 382, + 545, + 405 + ], + "lines": [ + { + "bbox": [ + 306, + 382, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 545, + 405 + ], + "type": "text", + "content": "Table 7. Effects of Sub-image's size. All models here are trained without mask augmentation." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 351, + 545, + 496, + 626 + ], + "lines": [ + { + "bbox": [ + 351, + 545, + 496, + 626 + ], + "spans": [ + { + "bbox": [ + 351, + 545, + 496, + 626 + ], + "type": "table", + "html": "
VariantsCDFDFDC
0, 1, 2, -86.4669.51
0, 1, -, -84.2269.09
Random85.8570.30
Reverse86.6572.37
Forward87.6074.32
", + "image_path": "14d54f0cefb5a76bfaef48899b15224d256ae61cbb7e1e63220241a3458bf7b3.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 629, + 545, + 651 + ], + "lines": [ + { + "bbox": [ + 306, + 629, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 629, + 545, + 651 + ], + "type": "text", + "content": "Table 8. Ablation study of absence and order of thumbnails. All models here are trained without mask augmentation." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Effects of different orders on other backbones. In order to prove that the phenomenon is not incidental, we also conduct experiments on ResNet50 and EfficientNet for three different orders as shown in Table 9. As expected, for-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22665" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "ward order outperforms reverse and random orders both on ResNet50 and EfficientNet, which indicates that TALL can learn the temporal dependency." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 118, + 248, + 243 + ], + "blocks": [ + { + "bbox": [ + 82, + 118, + 248, + 243 + ], + "lines": [ + { + "bbox": [ + 82, + 118, + 248, + 243 + ], + "spans": [ + { + "bbox": [ + 82, + 118, + 248, + 243 + ], + "type": "table", + "html": "
VariantsCDFDFDC
ResNet50+TALL76.3864.01
Random78.1464.12
Reverse78.5464.87
Forward80.9065.54
EfficientNet+TALL78.1966.81
Random81.0166.13
Reverse81.6666.69
Forward83.3767.15
", + "image_path": "1ba2ed4c1d95f8e37238a565bfea4bdf3d1e586b0284d821bc5d896751cba4a1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 268, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 288, + 483 + ], + "type": "text", + "content": "Effectiveness of mask strategy. In this work, TALL-Swin is trained on the " + }, + { + "bbox": [ + 46, + 268, + 288, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 46, + 268, + 288, + 483 + ], + "type": "text", + "content": " (HQ) dataset without any data enhancement as the baseline except for Multi-scale Crop and Random Horizontal Flip. To validate the effectiveness of the mask strategy, we compare our default baseline with different data augmentation strategies: 1) The Cutout [8] on one sub-image; 2) The Cutout on four sub-images. 3) The combination of Mixup [62] and Cutmix [60] on four sub-images, as shown in Table 10. The performance of a random Cutout [8] on four sub-images is better than on one sub-image. Besides, the mask strategy leads to better performance than the well-known Cutout " + }, + { + "bbox": [ + 46, + 268, + 288, + 483 + ], + "type": "inline_equation", + "content": "(1.46\\%)" + }, + { + "bbox": [ + 46, + 268, + 288, + 483 + ], + "type": "text", + "content": ". This supports our hypothesis that strategy encourages models to learn subtle temporal-spatial variations and improves model generalization ability. Further, our augmentation strategy exceeds " + }, + { + "bbox": [ + 46, + 268, + 288, + 483 + ], + "type": "inline_equation", + "content": "1.02\\%" + }, + { + "bbox": [ + 46, + 268, + 288, + 483 + ], + "type": "text", + "content": " than the combination of Mixup and Cutmix, demonstrating the augmentation's effectiveness in TALL for video detection." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 80, + 491, + 251, + 567 + ], + "blocks": [ + { + "bbox": [ + 63, + 247, + 268, + 257 + ], + "lines": [ + { + "bbox": [ + 63, + 247, + 268, + 257 + ], + "spans": [ + { + "bbox": [ + 63, + 247, + 268, + 257 + ], + "type": "text", + "content": "Table 9. Ablation studies of different orders of TALL." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 491, + 251, + 567 + ], + "lines": [ + { + "bbox": [ + 80, + 491, + 251, + 567 + ], + "spans": [ + { + "bbox": [ + 80, + 491, + 251, + 567 + ], + "type": "table", + "html": "
AugmentationCountCDFDFDC
None-87.6074.32
Cutout189.0674.07
Cutout489.3375.22
Mixup+CutMix489.7575.33
TALL's mask490.7976.78
", + "image_path": "cd3925bde530eb43fec2833a8bf9fba7eb0fc782c4f6f503a13b396cb1d5cdd7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Study on window size. We study the effect of window size on model performance and computational cost. The results are shown in Table 11. Our window expansion for the first three phases will increase the model performance by " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "1.74\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " AUC. The results in the second and third rows show that the first three stages of the window getting the largest would not give a boost to the model. Our analysis of a too-large window may lead to a weakening of the model's ability to learn local information in the sub-image." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 356, + 69, + 494, + 121 + ], + "blocks": [ + { + "bbox": [ + 47, + 570, + 286, + 591 + ], + "lines": [ + { + "bbox": [ + 47, + 570, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 286, + 591 + ], + "type": "text", + "content": "Table 10. Study of the augmentation strategy in TALL. The count column represents the number of blocks on the thumbnail." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 356, + 69, + 494, + 121 + ], + "lines": [ + { + "bbox": [ + 356, + 69, + 494, + 121 + ], + "spans": [ + { + "bbox": [ + 356, + 69, + 494, + 121 + ], + "type": "table", + "html": "
Window sizeCDFDFDC
(7,7,7,7)85.6073.32
(14,14,14,7)87.6074.32
(28,28,28,7)86.6574.21
", + "image_path": "b2ed7f0f736ce4b43d50f5e44bb748433161e61c7701b5ca98c3d04f0ce33e59.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 124, + 545, + 146 + ], + "lines": [ + { + "bbox": [ + 305, + 124, + 545, + 146 + ], + "spans": [ + { + "bbox": [ + 305, + 124, + 545, + 146 + ], + "type": "text", + "content": "Table 11. Ablation studies of window size. All models here are trained without mask augmentation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 165, + 378, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 165, + 378, + 178 + ], + "spans": [ + { + "bbox": [ + 306, + 165, + 378, + 178 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 186, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 318 + ], + "type": "text", + "content": "This paper presents a novel perspective on detecting deepfake videos using TALL. TALL is both simple and effective, enabling joint spatio-temporal modeling without any additional costs. TALL representation reveals normal deepfake patterns with local-global contextual features. We further propose a new baseline for deepfake video detection called TALL-Swin, which efficiently captures the inconsistencies between deepfake video frames. Extensive experiments demonstrate that TALL-Swin achieves promising results for various unseen deepfake types and strong robustness to a wide range of common corruptions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 330, + 398, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 398, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 398, + 342 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 422 + ], + "type": "text", + "content": "This work was partially funded by National Natural Science Foundation of China under Grants (62276256, U21B2045 and U20A20223) and Beijing Nova Program under Grant Z211100002121108. The authors wish to thank Huaibo Huang and Lijun Sheng in no particular order, for insightful discussions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 433, + 365, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 433, + 365, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 433, + 365, + 445 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 453, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 310, + 453, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 453, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 310, + 453, + 545, + 486 + ], + "type": "text", + "content": "[1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In Proc. WIFS, pages 1-7, 2018. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 487, + 547, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 487, + 547, + 530 + ], + "spans": [ + { + "bbox": [ + 311, + 487, + 547, + 530 + ], + "type": "text", + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proc. ICCV, pages 6836-6846, 2021. 4, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 533, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 533, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 311, + 533, + 545, + 576 + ], + "type": "text", + "content": "[3] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proc. CVPR, pages 4113-4122, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 578, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 578, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 311, + 578, + 545, + 610 + ], + "type": "text", + "content": "[4] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proc. CVPR, pages 6299-6308, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 612, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 612, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 311, + 612, + 545, + 655 + ], + "type": "text", + "content": "[5] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Proc. ECCV, pages 103-120, 2020. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 311, + 657, + 545, + 689 + ], + "type": "text", + "content": "[6] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proc. CVPR, pages 1251-1258, 2017. 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 311, + 691, + 545, + 714 + ], + "type": "text", + "content": "[7] deepfakes. Deepfakes. https://github.com/deepfakes/faceswap.2021-11-13.4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22666" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 73, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 286, + 106 + ], + "type": "text", + "content": "[8] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv:1708.04552, 2017. 3, 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 108, + 286, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 108, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 53, + 108, + 286, + 152 + ], + "type": "text", + "content": "[9] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv:2006.07397, 2020. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 154, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 154, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 154, + 286, + 196 + ], + "type": "text", + "content": "[10] Chengdong Dong, Ajay Kumar, and Eryun Liu. Think twice before detecting gan-generated fake images from their spectral domain imprints. In Proc. CVPR, pages 7865-7874, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 199, + 286, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 199, + 286, + 254 + ], + "spans": [ + { + "bbox": [ + 49, + 199, + 286, + 254 + ], + "type": "text", + "content": "[11] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Ting Zhang, Weiming Zhang, Nenghai Yu, Dong Chen, Fang Wen, and Baining Guo. Protecting celebrities from deepfake with identity consistency transformer. In Proc. CVPR, pages 9468-9478, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 256, + 286, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 256, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 49, + 256, + 286, + 322 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proc. ICLR, 2021. 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 324, + 286, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 324, + 286, + 367 + ], + "spans": [ + { + "bbox": [ + 49, + 324, + 286, + 367 + ], + "type": "text", + "content": "[13] Jianwei Fei, Yunshu Dai, Peipeng Yu, Tianrun Shen, Zhihua Xia, and Jian Weng. Learning second order local anomaly for general face forgery detection. In Proc. CVPR, pages 20270-20280, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 370, + 286, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 370, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 49, + 370, + 286, + 414 + ], + "type": "text", + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. In Proc. NeurIPS, volume 27, 2014. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 416, + 286, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 416, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 49, + 416, + 286, + 459 + ], + "type": "text", + "content": "[15] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proc. ACM MM, pages 3473-3481, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 462, + 286, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 462, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 49, + 462, + 286, + 506 + ], + "type": "text", + "content": "[16] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, and Lizhuang Ma. Delving into the local: Dynamic inconsistency learning for deepfake video detection. In Proc. AAAI, pages 744-752, 2022. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 508, + 286, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 508, + 286, + 551 + ], + "spans": [ + { + "bbox": [ + 49, + 508, + 286, + 551 + ], + "type": "text", + "content": "[17] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proc. CVPR, 2022. 2, 4, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 554, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 554, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 49, + 554, + 286, + 597 + ], + "type": "text", + "content": "[18] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In Proc. CVPR, pages 5039-5049, 2021. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 600, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 600, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 49, + 600, + 286, + 643 + ], + "type": "text", + "content": "[19] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Learning spatio-temporal features with 3d residual networks for action recognition. In Proc. ICCV Workshops, pages 3154-3160, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 646, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 646, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 646, + 286, + 678 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 286, + 712 + ], + "type": "text", + "content": "[21] Gee-Sern Hsu, Chun-Hung Tsai, and Hung-Yi Wu. Dual-generator face reenactment. In Proc. CVPR, pages 642-650, 2022. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[22] Yihao Huang, Felix Juefei-Xu, Qing Guo, Yang Liu, and Geguang Pu. Fakelocator: Robust localization of gan-based face manipulations. IEEE Transactions on Information Forensics and Security, 17:2657-2672, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[23] Ge-Peng Ji, Guobao Xiao, Yu-Cheng Chou, Deng-Ping Fan, Kai Zhao, Geng Chen, and Luc Van Gool. Video polyp segmentation: A deep learning perspective. MIR, 19(6):531-549, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[24] Ge-Peng Ji, Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Christos Sakaridis, and Luc Van Gool. Masked vision-language transformer in fashion. MIR, 20(3):421-434, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "type": "text", + "content": "[25] Gengyun Jia, Meisong Zheng, Chuanrui Hu, Xin Ma, Yuting Xu, Luoqi Liu, Yafeng Deng, and Ran He. Inconsistency-aware wavelet dual-branch network for face forgery detection. IEEE Transactions on Biometrics, Behavior, and Identity Science, 3(3), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 307 + ], + "type": "text", + "content": "[26] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proc. CVPR, pages 2889-2898, 2020. 4, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proc. CVPR, pages 4401-4410, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "text", + "content": "[28] Sohail Ahmed Khan and Hang Dai. Video transformer for deepfake detection with incremental learning. In Proc. ACM MM, pages 1821-1828, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "text", + "content": "[29] Aminollah Khormali and Jiann-Shiun Yuan. Dfdt: An end-to-end deepfake detection framework using vision transformer. Applied Sciences, page 2953, 2022. 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 430 + ], + "type": "text", + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 477 + ], + "type": "text", + "content": "[31] Jiaming Li, Hongtao Xie, Jiahong Li, Zhongyuan Wang, and Yongdong Zhang. Frequency-aware discriminative feature learning supervised by single-center loss for face forgery detection. In Proc. CVPR, pages 6458-6467, 2021. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "text", + "content": "[32] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In Proc. CVPR, pages 5001-5010, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[33] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. In Proc. CVPR Workshops, pages 656-663, 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 557, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 590 + ], + "type": "text", + "content": "[34] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proc. CVPR, pages 3207-3216, 2020. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 591, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 634 + ], + "type": "text", + "content": "[35] Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: Rethinking face forgery detection in frequency domain. In Proc. CVPR, pages 772-781, 2021. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 679 + ], + "type": "text", + "content": "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. ICCV, pages 10012-10022, 2021. 2, 3, 4, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[37] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proc. CVPR, pages 8060–8069, 2020. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22667" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 286, + 94 + ], + "type": "text", + "content": "[38] MarekKowalski. Faceswap. https://github.com/MarekKowalski/FaceSwap/. 2021-11-13. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 286, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 286, + 139 + ], + "type": "text", + "content": "[39] Iacopo Masi, Aditya Killekar, Royston Marian Mascarenhas, Shenoy Pratik Gurudatt, and Wael AbdAlmageed. Two-branch recurrent network for isolating deepfakes in videos. In Proc. ECCV, pages 667–684, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 286, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 286, + 161 + ], + "type": "text", + "content": "[40] Yisroel Mirsky and Wenke Lee. The creation and detection of deepfakes: A survey. ACM CSUR, 54(1):1-41, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 162, + 286, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 286, + 194 + ], + "type": "text", + "content": "[41] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In Proc. ICCV, pages 3163-3172, 2021. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 195, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 286, + 239 + ], + "type": "text", + "content": "[42] Yuval Nirkin, Lior Wolf, Yosi Keller, and Tal Hassner. Deepfake detection based on discrepancies between faces and their context. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6111-6121, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 239, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 286, + 282 + ], + "type": "text", + "content": "[43] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In Proc. ECCV, pages 86-103, 2020. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 284, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 286, + 304 + ], + "type": "text", + "content": "[44] Rameswar Panda Quanfu Fan, Richard Chen. Can an image classifier suffice for action recognition? In ICLR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 305, + 286, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 286, + 350 + ], + "type": "text", + "content": "[45] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. Faceforensics: Learning to detect manipulated facial images. In Proc. ICCV, pages 1-11, 2019. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 350, + 286, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 286, + 393 + ], + "type": "text", + "content": "[46] Ekraam Sabir, Jiaxin Cheng, Ayush Jaiswal, Wael AbdAlmageed, Iacopo Masi, and Prem Natarajan. Recurrent convolutional strategies for face manipulation detection in videos. In Proc. CVPR Workshops, pages 80-87, 2019. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 394, + 286, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 286, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 286, + 447 + ], + "type": "text", + "content": "[47] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proc. ICCV, pages 618-626, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 449, + 286, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 286, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 286, + 492 + ], + "type": "text", + "content": "[48] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In Proc. CVPR, pages 11254-11264, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 493, + 286, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 286, + 536 + ], + "type": "text", + "content": "[49] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proc. CVPR, pages 7672-7682, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 537, + 286, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 286, + 569 + ], + "type": "text", + "content": "[50] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, pages 6105-6114, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 571, + 286, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 286, + 602 + ], + "type": "text", + "content": "[51] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM TOG, 38(4):1-12, 2019. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 604, + 286, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 286, + 647 + ], + "type": "text", + "content": "[52] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proc. ICCV, pages 2387-2395, 2016. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 286, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 680 + ], + "type": "text", + "content": "[53] Luisa Verdoliva. Media forensics and deepfakes: an overview. IEEE Journal of Selected Topics in Signal Processing, 14(5):910-932, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[54] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In Proc. CVPR, pages 14923-14932, 2021. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 709 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[55] Junke Wang, Zuxuan Wu, Wenhao Ouyang, Xintong Han, Jingjing Chen, Yu-Gang Jiang, and Ser-Nam Li. M2tr: Multi-modal multi-scale transformers for deepfake detection. In Proc. ICMR, pages 615-623, 2022. 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[56] Ping Wang, Kunlin Liu, Wenbo Zhou, Hang Zhou, Honggu Liu, Weiming Zhang, and Nenghai Yu. Adt: Anti-deepfake transformer. In Proc. ICASSP, pages 2899-1903, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "text", + "content": "[57] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proc. CVPR, 2020. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 194, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 226 + ], + "type": "text", + "content": "[58] Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv:2102.11126, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "type": "text", + "content": "[59] Ziming Yang, Jian Liang, Yuting Xu, Xiao-Yu Zhang, and Ran He. Masked relation learning for deepfake detection. IEEE Transactions on Information Forensics and Security, pages 1696-1708, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 271, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 271, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 545, + 315 + ], + "type": "text", + "content": "[60] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proc. ICCV, pages 6023-6032, 2019. 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 316, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 347 + ], + "type": "text", + "content": "[61] Yu Zeng, Zhe Lin, and Vishal M Patel. Sketched: Mask-free local image manipulation with partial sketches. In Proc. CVPR, pages 5951-5961, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 348, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 348, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 545, + 380 + ], + "type": "text", + "content": "[62] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond empirical risk minimization. Proc. ICLR, 2018. 9" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 380, + 545, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 380, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 307, + 380, + 545, + 424 + ], + "type": "text", + "content": "[63] Yanyi Zhang, Xinyu Li, Chunhui Liu, Bing Shuai, Yi Zhu, Biagio Brattoli, Hao Chen, Ivan Marsic, and Joseph Tighe. Vidtr: Video transformer without convolutions. In Proc. ICCV, pages 13577-13587, 2021. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 425, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 425, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 307, + 425, + 545, + 479 + ], + "type": "text", + "content": "[64] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: Interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18:1335-1348, 2023. 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "text", + "content": "[65] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep-fake detection. In Proc. CVPR, pages 2185–2194, 2021. 2, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[66] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Weiming Zhang, and Nenghai Yu. Self-supervised transformer for deepfake detection. arXiv:2203.01265, 2022. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "text", + "content": "[67] Tianchen Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In Proc. CVPR, pages 15023-15033, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 632 + ], + "type": "text", + "content": "[68] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In Proc. ICCV, pages 15044-15054, 2021. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 633, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 633, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 633, + 545, + 666 + ], + "type": "text", + "content": "[69] Xiangyu Zhu, Hao Wang, Hongyan Fei, Zhen Lei, and Stan Z Li. Face forgery detection by 3d decomposition. In Proc. CVPR, pages 2929-2939, 2021. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 666, + 545, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 545, + 709 + ], + "type": "text", + "content": "[70] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proc. ACM MM, pages 2382-2390, 2020. 5" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22668" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/f0fe9ecb-f819-46e1-ba9a-de70f0257649_content_list.json b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/f0fe9ecb-f819-46e1-ba9a-de70f0257649_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6f33fdc5c180061328fdb9da8bf352a006967c91 --- /dev/null +++ b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/f0fe9ecb-f819-46e1-ba9a-de70f0257649_content_list.json @@ -0,0 +1,1706 @@ +[ + { + "type": "text", + "text": "TAPIR: Tracking Any Point with per-frame Initialization and temporal Refinement", + "text_level": 1, + "bbox": [ + 125, + 128, + 846, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Carl Doersch*", + "bbox": [ + 114, + 202, + 236, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Yang*", + "bbox": [ + 277, + 203, + 354, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mel Vecerik*†", + "bbox": [ + 392, + 203, + 514, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dilara Gokay*", + "bbox": [ + 553, + 203, + 679, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ankush Gupta*", + "bbox": [ + 718, + 203, + 852, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yusuf Aytar*", + "bbox": [ + 235, + 220, + 346, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joao Carreira*", + "bbox": [ + 387, + 220, + 514, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew Zisserman*†", + "bbox": [ + 555, + 220, + 733, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Google DeepMind", + "bbox": [ + 271, + 257, + 429, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† University College London", + "bbox": [ + 468, + 257, + 697, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{\\ddagger}$ VGG, Department of Engineering Science, University of Oxford", + "bbox": [ + 225, + 275, + 743, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 328, + 313, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a novel model for Tracking Any Point (TAP) that effectively tracks any queried point on any physical surface throughout a video sequence. Our approach employs two stages: (1) a matching stage, which independently locates a suitable candidate point match for the query point on every other frame, and (2) a refinement stage, which updates both the trajectory and query features based on local correlations. The resulting model surpasses all baseline methods by a significant margin on the TAP-Vid benchmark, as demonstrated by an approximate $20\\%$ absolute average Jaccard (AJ) improvement on DAVIS. Our model facilitates fast inference on long and high-resolution video sequences. On a modern GPU, our implementation has the capacity to track points faster than real-time. Given the high-quality trajectories extracted from a large dataset, we demonstrate a proof-of-concept diffusion model which generates trajectories from static images, enabling plausible animations. Visualizations, source code, and pretrained models can be found at https://deepmind-tapir.github.io.", + "bbox": [ + 73, + 359, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 676, + 209, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The problem of point level correspondence —i.e., determining whether two pixels in two different images are projections of the same point on the same physical surface—has long been a fundamental challenge in computer vision, with enormous potential for providing insights about physical properties and 3D shape. We consider its formulation as “Tracking Any Point” (TAP) [12]: given a video and (potentially dense) query points on solid surfaces, an algorithm should reliably output the locations those points correspond to in every other frame where they are visible, and indicate frames where they are not – see Fig. 4 for illustration.", + "bbox": [ + 75, + 703, + 468, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our main contribution is a new model: TAP with per-frame Initialization and temporal Refinement (TAPIR),", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/90016a7bcb80eadb6771cb4c9979ff7cdfcdad16777a4611e5689c693d102ff2.jpg", + "image_caption": [ + "Figure 1. Retrospective evolution of point tracking performance over time on the recent TAP-Vid-Kinetics and TAP-Vid-DAVIS benchmarks, as measured by Average Jaccard (higher is better). In this paper we introduce TAPIR, which significantly improves performance over the state-of-the-art. This unlocks new capabilities, which we demonstrate on motion-based future prediction." + ], + "image_footnote": [], + "bbox": [ + 506, + 330, + 888, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which greatly improves over the state-of-the-art on the recently-proposed TAP-Vid benchmark [12]. There are many challenges to TAP: we must robustly estimate occlusions and recover when points reappear (unlike optical flow and structure-from-motion keypoints), meaning that search must be incorporated; yet when points remain visible for many consecutive frames, it is important to integrate information about appearance and motion across many of those frames in order to optimally predict positions. Furthermore, little real-world ground truth is available to learn from, so supervised-learning algorithms need to learn from synthetic data without overfitting to the data distribution (i.e., sim2real).", + "bbox": [ + 496, + 654, + 893, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are three core design decisions that define TAPIR. The first is to use a coarse-to-fine approach. This approach has been used across many high-precision estimation prob", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10061", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lems in computer vision [7, 28, 32, 38, 41, 58, 69, 73, 78]. For our application, the initial 'coarse' tracking consists of an occlusion-robust matching performed separately on every frame, where tracks are hypothesized using low-resolution features, without enforcing temporal continuity. The 'fine' refinement iteratively uses local, spatio-temporal information at a higher resolution, wherein a neural network can trade-off smoothness of motion with local appearance cues to produce the most likely track. The second design decision is to be fully-convolutional in time: the layers of our neural network consist principally of feature comparisons, spatial convolutions, and temporal convolutions, resulting in a model which efficiently maps onto modern GPU and TPU hardware. The third design decision is that the model should estimate its own uncertainty with regard to its position estimate, in a self-supervised manner. This ensures that low-confidence predictions can be suppressed, which improves benchmark scores. We hypothesize that this may help downstream algorithms (e.g. structure-from-motion) that rely on precision, and can benefit when low-quality matches are removed.", + "bbox": [ + 80, + 90, + 467, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We find that two existing architectures already have some of the pieces we need: TAP-Net [12] and Persistent Independent Particles (PIPs) [19]. Therefore, a key contribution of our work is to effectively combine them while achieving the benefits from both. TAP-Net performs a global search on every frame independently, providing a coarse track that is robust to occlusions. However, it does not make use of the continuity of videos, resulting in jittery, unrealistic tracks. PIPs, meanwhile, gives a recipe for refinement: given an initialization, it searches over a local neighborhood and smooths the track over time. However, PIPs processes videos sequentially in chunks, initializing each chunk with the output from the last. The procedure struggles with occlusion and is difficult to parallelize, resulting in slow processing (i.e., 1 month to evaluate TAP-Vid-Kinetics on 1 GPU). A key contribution of this work is observing the complementarity of these two methods.", + "bbox": [ + 80, + 411, + 467, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1, we find that TAPIR improves over prior works by a large margin, as measured by performance on the TAP-Vid benchmark [12]. On TAP-Vid-DAVIS, TAPIR outperforms TAP-Net by $\\sim 20\\%$ while on the more challenging TAP-Vid-Kinetics, TAPIR outperforms PIPs by $\\sim 20\\%$ , and substantially reduces its inference runtime. To demonstrate the quality of TAPIR trajectories, we showcase a proof-of-concept model trained to generate trajectories given individual images, and find that this model can generate plausible animations from single photographs.", + "bbox": [ + 80, + 670, + 467, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows: 1) We propose a new model for long term point tracking, bridging the gap between TAP-Net and PIPs. 2) We show that the model achieves state-of-the-art results on the challenging TAP-Vid benchmark, with a significant boost on per", + "bbox": [ + 80, + 824, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "formance. 3) We provide an extensive analysis of the architectural decisions that matter for high-performance point tracking. 4) We provide a proof-of-concept of video prediction enabled by TAPIR's high-quality trajectories. Finally, 5) after analyzing components, we separately perform careful tuning of hyperparameters across entire method, in order to develop the best-performing model, which we release at https://www.github.com/deepmind/tapnet for the benefit of the community.", + "bbox": [ + 503, + 90, + 888, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 246, + 638, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physical point correspondence and tracking have been studied in a wide range of settings.", + "bbox": [ + 503, + 273, + 888, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optical Flow focuses on dense motion estimation between image pairs [22, 36]; methods can be divided into classical variational approaches [4, 5] and deep learning approaches [14, 23, 47, 60, 62, 71]. Optical flow operates only on subsequent frames, with no simple method to track across long videos. Groundtruth is hard to obtain [6], so it is typically benchmarked on synthetic scenes [6, 14, 40, 59], though limited real data exists through depth scanners [15].", + "bbox": [ + 503, + 305, + 888, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keypoint Correspondence methods aim at sparse keypoint matching given image pairs, from hand-defined discriminative feature descriptors [3, 34, 35, 54] to deep learning based methods [11, 27, 37, 43]. Though it operates on image pairs, it is arguably \"long term\" as the images may be taken at different times from wide baselines. The goal, however, is typically not to track any point, but to find easily-trackable \"keypoints\" sufficient for reconstruction. They are mainly used in structure-from-motion (SfM) settings and typically ignore occlusion, as SfM can use geometry to filter errors [20, 52, 65, 66]. This line of work is thus often restricted to rigid scenes [39, 75], and is benchmarked accordingly [1, 10, 29, 30, 53, 76].", + "bbox": [ + 503, + 428, + 888, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic Keypoint Tracking is often represented as keypoints or landmarks [26, 45, 56, 61, 64, 70]. Landmarks are often \"joints\", which can have vastly different appearance depending on pose (i.e. viewed from the front or back); thus most methods rely on large supervised datasets [2, 9, 48, 55, 72, 77], although some works track surface meshes [18, 67]. One interesting exception uses sim2real based on motion [13], motivating better surface tracking. Finally, some recent methods discover keypoints on moving objects [24, 25, 63, 74], though these typically require in-domain training on large datasets.", + "bbox": [ + 503, + 626, + 888, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long-term physical point tracking is addressed in a few early works [33, 49, 51, 57, 68], though such hand-engineered methods have fallen out of favor in the deep learning era. Our work instead builds on two recent works, TAP-Net [12] and Persistent Independent Particles (PIPs) [19], which aim to update these prior works to the deep learning era.", + "bbox": [ + 503, + 795, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10062", + "bbox": [ + 482, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/902c79206953071b29eda02102ddbd7a9487ae746f178a91972e615271fa6a5b.jpg", + "image_caption": [ + "Figure 2. TAPIR architecture summary. Our model begins with a global comparison between the query point features and the features for every other frame to compute an initial track estimate, including an uncertainty estimate. Then, we extract features from a local neighborhood (shown in pink) around the initial estimate, and compare these to the query feature at a higher resolution, post-processing the similarities with a temporal depthwise-convolitional network to get an updated position estimate. This updated position is fed back into the next iteration of refinement, repeated for a fixed number of iterations. Note that for simplicity, multi-scale pyramids are not shown." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. TAPIR Model", + "text_level": 1, + "bbox": [ + 76, + 630, + 220, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video and a query point, our goal is to estimate the 2D location $p_t$ that it corresponds to in every other frame $t$ , as well as a 1D probability $o_t$ that the point is occluded and a 1D probability $u_t$ on the uncertainty of the estimated location. To be robust to occlusion, we first match candidate locations in other frames, by comparing the query features with all other features, and post-processing the similarities to arrive at an initial estimate. Then we refine the estimate, based on the local similarities between the query point and the target point. Note that both stages depend mostly on similarities (dot products) between the query point features and features elsewhere (i.e. not on the feature alone); this ensures that we don't overfit to specific features that might only appear in the (synthetic) training data.", + "bbox": [ + 75, + 657, + 468, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 gives an overview of our model. We initialize an estimate of the track by finding the best match within each", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/6dc395be74df4347fbd72725f922e52811f216f82f83d58b1667d713c4f55f3e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TAP-Net [12]PIPs [19]TAPIR
Per-frame Initialization✓✗✓
Fully Convolutional In Time✓✗✓
Temporal Refinement✗✓✓
Multi Scale Feature✗✓✓
Stride-4 Feature✗test-time only✓
Parallel Inference✓✗✓
Uncertainty Estimate✗✗✓
Number of Parameters2.8M28.7M29.3M
", + "bbox": [ + 500, + 88, + 893, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Overview of models. TAPIR combines the merits from both TAP-Net and PIPs and further adds a handful of crucial components to improve performance.", + "bbox": [ + 498, + 210, + 890, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "frame independently, ignoring the temporal nature of the video. This involves a cost volume: for each frame $t$ , we compute the dot product between the query feature $F_{q}$ and all features in the $t$ 'th frame. We post-process the cost volume with a ConvNet, which produces a spatial heatmap, which is then summarized to a point estimate, in a manner broadly similar to TAP-Net. Given the initialization, we then compare the query feature to all features in a small region around an initial track. We feed the comparisons into a neural network, which updates both the query features and the track. We iteratively apply multiple such updates to arrive at a final trajectory estimate. TAPIR accomplishes this with a fully-convolutional architecture, operating on a pyramid of both high-dimensional, low-resolution features and low-dimensional, high-resolution features, in order to maximize efficiency on modern hardware. Finally, we add an estimate of uncertainty with respect to position throughout the architecture in order to suppress low-accuracy predictions. The rest of this section describes TAPIR in detail.", + "bbox": [ + 496, + 263, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Track Initialization", + "text_level": 1, + "bbox": [ + 500, + 560, + 686, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The initial cost volume is computed using a relatively coarse feature map $F \\in \\mathbb{R}^{T \\times H/8 \\times W/8 \\times C}$ , where $T$ is the number of frames, $H$ and $W$ are the image height and width, and $C$ is the number of channels, computed with a standard TSM-ResNet-18 [31] backbone. Features for the query point $F_{q}$ are extracted via bilinear interpolation at the query location, and we perform a dot product between this query feature and all other features in the video.", + "bbox": [ + 496, + 583, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We compute an initial estimate of the position $p_t^0 = (x_t^0, y_t^0)$ and occlusion $o_t^0$ by applying a small ConvNet to the cost volume corresponding to frame $t$ (which is of shape $H / 8 \\times W / 8 \\times 1$ per query). This ConvNet has two outputs: a heatmap of shape $H / 8 \\times W / 8 \\times 1$ for predicting the position, and a single scalar logit for the occlusion estimate, obtained via average pooling followed by projection. The heatmap is converted to a position estimate by a \"spatial soft argmax\", i.e., a softmax across space (to make the heatmap positive and sum to 1). Afterwards, all heatmap values which are too far from the argmax position in this heatmap are set to zero. Then, the positions for each cell of the heatmap are averaged spatially, weighted by the thresholded heatmap", + "bbox": [ + 498, + 704, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10063", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "magnitudes. Thus, the output is typically a location close to the heatmap's maximum; the \"soft argmax\" is differentiable, but the thresholding suppresses spurious matches, and prevents the network from \"averaging\" between multiple matches. This can serve as a good initialization, although the model struggles to localize points to less than a few pixels' accuracy on the original resolution due to the 8-pixel stride of the heatmap.", + "bbox": [ + 75, + 90, + 472, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Position Uncertainty Estimates. A drawback of predicting occlusion and position independently from the cost volume (a choice that is inspired by TAP-Net) is that if a point is visible in the ground truth, it's worse if the algorithm predicts a vastly incorrect location than if it simply incorrectly marks the point as occluded. After all, downstream applications may want to use the track to understand object motion; such downstream pipelines must be robust to occlusion, but may assume that the tracks are otherwise correct. The Average Jaccard metric reflects this: predictions in the wrong location are counted as both a \"false positive\" and a \"false negative.\" This kind of error tends to happen when the algorithm is uncertain about the position, e.g., if there are many potential matches. Therefore, we find it's beneficial for the algorithm to also estimate its own certainty regarding position. We quantify this by making the occlusion pathway output a second logit, estimating the probability $u_{t}^{0}$ that the prediction is likely to be far enough from the ground truth that it isn't useful, even if the model predicts that it's visible. We define \"useful\" as whether the prediction is within a threshold $\\delta$ of the ground truth.", + "bbox": [ + 75, + 237, + 472, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, our loss for the initialization for a given video frame $t$ is $\\mathcal{L}(p_t^0,o_t^0,u_t^0)$ , where $\\mathcal{L}$ is defined as:", + "bbox": [ + 76, + 555, + 470, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} \\left(p _ {t}, o _ {t}, u _ {t}\\right) = \\operatorname {H u b e r} \\left(\\hat {p} _ {t}, p _ {t}\\right) * \\left(1 - \\hat {o} _ {t}\\right) \\\\ + \\operatorname {B C E} \\left(\\hat {o} _ {t}, o _ {t}\\right) \\\\ + \\operatorname {B C E} \\left(\\hat {u} _ {t}, u _ {t}\\right) * \\left(1 - \\hat {o} _ {t}\\right) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 613, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e ,} \\quad \\hat {u} _ {t} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} d (\\hat {p} _ {t}, p _ {t}) > \\delta \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 670, + 379, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\hat{o}_t \\in \\{0,1\\}$ and $\\hat{p} \\in \\mathbb{R}^2$ are the ground truth occlusion and point locations respectively, $d$ is Euclidean distance, $\\delta$ is the distance threshold, Huber is a Huber loss, and BCE is binary cross entropy (both $o_t$ and $u_t$ have sigmoids applied to convert them to probabilities). $\\hat{u}_t$ , the target for the uncertainty estimate $u_t$ , is computed from the ground truth position and the network's prediction: it's 0 if the model's position prediction is close enough to the ground truth (within the threshold $\\delta$ ), and 1 otherwise. At test time, the algorithm should output that the point is visible if it's both predicted as unoccluded and if the model is confident in the prediction. We therefore do a soft combination of", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the two probabilities: the algorithm outputs that the point is visible if $(1 - u_t) * (1 - o_t) > 0.5$ .", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Iterative Refinement", + "text_level": 1, + "bbox": [ + 500, + 133, + 694, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an estimated position, occlusion, and uncertainty for each frame, the goal of each iteration $i$ of our refinement procedure is to compute an update $(\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i})$ , which adjusts the estimate to be closer to the ground truth, integrating information across time. The update is based on a set of \"local score maps\" which capture the query point similarity (i.e. dot products) to the features in the neighborhood of the trajectory. These are computed on a pyramid of different resolutions, so for a given trajectory, they have shape $(H' \\times W' \\times L)$ , where $H' = W' = 7$ , the size of the local neighborhood, and $L$ is the number of levels of the spatial pyramid (different pyramid levels are computed by spatially pooling the feature volume $F$ ). As with the initialization, this set of similarities is post-processed with a network to predict the refined position, occlusion, and uncertainty estimate. Unlike the initialization, however, we include \"local score maps\" for many frames simultaneously as input to the post-processing. We include the current position estimate, the raw query features, and the (flattened) local score maps into a tensor of shape $T \\times (C + K + 4)$ , where $C$ is the number of channels in the query feature, $K = H' \\cdot W' \\cdot L$ is the number of values in the flattened local score map, and 4 extra dimensions for position, occlusion, and uncertainty. The output of this network at the $i$ 'th iteration is a residual $(\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i}, \\Delta F_{q,t,i})$ , which is added to the position, occlusion, uncertainty estimate, and the query feature, respectively. $\\Delta F_{q,t,i}$ is of shape $T \\times C$ ; thus, after the first iteration, slightly different \"query features\" are used on each frame when computing new local score maps.", + "bbox": [ + 496, + 157, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These positions and score maps are fed into a 12-block convolutional network to compute an update $(\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i}, \\Delta F_{q,t,i})$ , where each block consists of a $1 \\times 1$ convolution block and a depthwise convolution block. This architecture is inspired by the MLP-Mixer used for refinement in PIPs: we directly translate the Mixer's cross-channel layers into $1 \\times 1$ convolutions with the same number of channels, and the within-channel operations become similarly within-channel depthwise convolutions. Unlike PIPs, which breaks sequences into 8-frame chunks before running the MLP-Mixer, this convolutional architecture can be run on sequences of any length.", + "bbox": [ + 498, + 598, + 890, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We found that the feature maps used for computing the \"score maps\" are important for achieving high precision. The pyramid levels $l = 1$ through $L - 1$ are computed by spatially average-pooling the raw TSM-ResNet features $F$ at a stride of $8 \\cdot 2^{l - 1}$ , and then computing all dot products between the query feature $F_{q}$ and the pyramid features. For the $t$ 'th frame, we extract a $7 \\times 7$ patch of dot products centered at $p_t$ . At training time, PIPs leaves it here, but at", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10064", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "test time alters the backbone to run at stride 4, introducing a train/test domain gap. We find it is effective to train on stride 4 as well, although this puts pressure on system memory. To alleviate this, we compute a 0-th score map on the 64-channel, stride-4 input convolution map of the TSM-ResNet, and extract a comparable feature for the query from this feature map via bilinear interpolation. Thus, the final set of local score maps has shape $(7 \\cdot 7 \\cdot L)$ .", + "bbox": [ + 75, + 90, + 467, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After $(\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i}, \\Delta F_{q,t,i})$ is obtained from the above architecture, we can iterate the refinement step as many times as desired, re-using the same network parameters. At each iteration, we use the same loss $\\mathcal{L}$ on the output, weighting each iteration the same as the initialization.", + "bbox": [ + 75, + 212, + 467, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion As there is some overlap with prior work, Figure 3 depicts the relationship between TAPIR and other architectures. For the sake of ablations, we also develop a \"simple combination\" model, which is intended to be a simpler combination of a TAP-Net-style initialization with PIPs-style refinements, and follows prior architectural decisions regardless of whether they are optimal. To summarize, the major architectural decisions needed to create the \"simple combination\" of TAP-Net and PIPs (and therefore the contributions of this model) are as follows. First we exploit the complementarity between TAP-Net and PIPs. Second, we remove \"chaining\" from PIPs, replacing that initialization with TAP-Net's initialization. We directly apply PIPs' MLP-Mixer architecture by 'chunking' the input sequence. Note that we adopt TAP-Net's feature network for both initialization and refinement. Finally, we adapt the TAP-Net backbone to extract a multi-scale and high resolution pyramid (follow PIPs stride-4 test-time architecture, but matching TAP-Net's end-to-end training). We apply losses at all layers of the hierarchy (whereas TAP-Net has a single loss, PIPs has no loss on its initialization). The additional contributions of TAPIR are as follows: first, we entirely replace the PIPs MLP-Mixer (which must be applied to fixed-length sequences) with a novel depthwise-convolitional architecture, which has a similar number of parameters but works on sequences of any length. Second, we introduce uncertainty estimate, which are computed at every level of the hierarchy. These are 'self-supervised' in the sense that the targets depend on the model's own output.", + "bbox": [ + 75, + 289, + 468, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Dataset", + "text_level": 1, + "bbox": [ + 76, + 739, + 243, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although TAPIR is designed to be robust to the sim2real gap, we can improve performance by minimizing the gap itself. One subtle difference between the Kubric MOViE [16] dataset and the real world is the lack of panning: although the Kubric camera moves, it is always set to \"look at\" a single point at the center of the workspace. Thus, stationary parts of the scene rotate around the center point of the workspace on the 2D image plane. Models thus tend to fail on real-world videos with panning. Therefore, we mod", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ified the MOVi-E dataset so that the \"look at\" point moves along a random linear trajectory. The start and end points of this trajectory are enforced to be less than 1.5 units off the ground and lie within a radius of 4 units from the workspace center; the trajectory must also pass within 1 unit of the workspace center at some point, to ensure that the camera is usually looking toward the objects in the scene. See Appendix 5.2 for details.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 227, + 633, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate on the TAP-Vid [12] benchmark, a recent large-scale benchmark evaluating the problem of interest. TAP-Vid consists of point track annotations on four different datasets, each with different challenges. DAVIS [44] is a benchmark designed for tracking, and includes complex motion and large changes in object scale. Kinetics [8] contains over 1000 labeled videos, and contains all the complexities of YouTube videos, including difficulties like cuts and camera shake. RGB Stacking is a synthetic dataset of robotics videos with precise point tracks on large textureless regions. Kubric MOVi-E, used for training, contains realistic renderings of objects falling onto a flat surface; it comes with a validation set that we use for comparisons. We conduct training exclusively on the Kubric dataset, and selected our best model primarily by observing DAVIS. We performed no automated evaluation for hyperparameter tuning or model selection on any dataset. We train and evaluate at a resolution of $256 \\times 256$ .", + "bbox": [ + 496, + 253, + 890, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We train for 50,000 steps with a cosine learning rate schedule, with a peak rate of $1 \\times 10^{-3}$ and 1,000 warmup steps. We use 64 TPU-v3 cores, using 4 24-frame videos per core in each step. We use an Adam-W optimizer with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.95$ and weight decay $1 \\times 10^{-2}$ . For most experiments, we use cross-replica batch normalization within the (TSM-)ResNet backbone only, although for our public model we find instance normalization is actually more effective. For most experiments we use $L = 5$ to match PIPs, but in practice we find $L > 3$ has little impact (See Appendix 4.3). Therefore, to save computation, our public model uses $L = 3$ . See Appendix 5.1 for more details.", + "bbox": [ + 496, + 547, + 890, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.Quantitative Comparison", + "text_level": 1, + "bbox": [ + 498, + 755, + 732, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 shows the performance of TAPIR relative to prior works. TAPIR improves by $10.6\\%$ absolute performance over the best model on Kinetics (TAP-Net) and by $19.3\\%$ on DAVIS (PIPs). These improvements are quite significant according to the TAP-Vid metrics: Average Jaccard consists of 5 thresholds, so under perfect occlusion estimation, a $20\\%$ improvement requires improving all points to the next distance threshold (half the distance). It's worth noting", + "bbox": [ + 496, + 779, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10065", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e4b5154c9ea6e729709200e681129490f2cea7954f486846da0cfa18e43b511f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 344, + 243 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6656db070a993c553263d72a54561b74ca215fbbce060ef0b2f849b0166b572b.jpg", + "image_caption": [ + "Figure 3. Comparison of Architectures. Left: the initial TAP-Net and PIPs models: TAP-Net has an independent estimate per frame, whereas PIPs breaks the video into fixed-size chunks and processes chunks sequentially. Middle: a \"simple combination\" of these two architectures, where TAP-Net is used to initialize PIPs-style chunked iterations. Right: Our model, TAPIR, which removes the chunking and adds uncertainty estimate. Note that for simplicity, multi-scale pyramids are not shown." + ], + "image_footnote": [], + "bbox": [ + 80, + 248, + 344, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f43c633f7edbb99052e2883b2306a9ef81cbd330d986f3e608f3e47939be1122.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 88, + 612, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05a05cb07531b542d157902b8e83a28bd058c3647010f2bd4478d233841f48bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 89, + 885, + 381 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fa04f074357fbf5ac5f3e77438d0d06cf116fc6cef83a83a2af6b9cbe88f586e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodKineticsKubricDAVISRGB-Stacking
AJ<δxavgOAAJ<δxavgOAAJ<δxavgOAAJ<δxavgOA
COTR [27]19.038.857.440.160.778.5535.451.380.26.813.579.1
Kubric-VFS-Like [16]40.559.080.051.969.884.633.148.579.457.972.691.9
RAFT [62]34.552.579.741.258.286.430.046.379.644.058.690.4
PIPs [19]35.354.877.459.174.888.642.059.482.137.351.091.6
TAP-Net [12]46.660.985.065.477.793.038.453.182.359.972.890.4
TAPIR (MOVi-E)57.170.087.684.391.895.859.872.387.666.277.493.3
TAPIR (Panning MOVi-E)57.270.187.884.792.195.861.373.688.862.774.691.6
", + "bbox": [ + 81, + 458, + 888, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of TAPIR to prior results on TAP-Vid. $< \\delta_{avg}^{x}$ is the fraction of points unoccluded in the ground truth, where the prediction is within a threshold, ignoring the occlusion estimate; OA is the accuracy of predicting occlusion. AJ is Average Jaccard, which considers both position and occlusion accuracy.", + "bbox": [ + 75, + 607, + 892, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that, like PIPs, TAPIR's performance on DAVIS is slightly higher than on Kinetics, whereas for TAP-Net, Kinetics has higher performance, suggesting that TAPIR's reliance on temporal continuity isn't as useful on Kinetics videos with cuts or camera shake. Note that the bulk of the improvement comes from the improved model rather than the improved dataset (TAPIR (MOVi-E) is trained on the same data as TAP-Net, while 'Panning' is the new dataset). The dataset's contribution is small, mostly improving DAVIS, but we found qualitatively that it makes the most substantial difference on backgrounds when cameras pan. This is not well represented in TAP-Vid's query points, but may matter for real-world applications like stabilization. RGB-stacking, however, is somewhat of an outlier: training on MOVi-E is still optimal (unsurprising as the cameras are", + "bbox": [ + 75, + 667, + 472, + 895 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "static), and there the improvement is only $6.3\\%$ , possibly because TAPIR does less to help with textureless objects, indicating an area for future research. Fig. 4 shows some example predictions on DAVIS. TAPIR corrects many types of errors, including problems with temporal coherence from TAP-Net and problems with occlusion from PIPs. We include further qualitative results on our project webpage, which illustrate how noticeable a $20\\%$ improvement is.", + "bbox": [ + 496, + 667, + 890, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TAP-Vid proposes another evaluation for videos at the resolution of $256 \\times 256$ called 'query-first.' In the typical case (Table 2), the query points are sampled in a 'strided' fashion from the ground-truth tracks: the same trajectory is therefore queried with multiple query points, once every 5 frames. In the 'query first' evaluation, only the first point along each trajectory is used as a query. These points", + "bbox": [ + 496, + 794, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10066", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f909fe9c1e174e514be7f2f8b156ce300e1493268b4d5a6dae45176e10951e02.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Comparison under query first metrics. We see largely the same relative performance trend whether the model is queried in a 'strided' fashion or whether the model is queried with the first frame where the point appeared, although performance is overall lower than the 'strided' evaluation." + ], + "table_body": "
MethodKineticsOADAVISOARGB-Stacking
AJ<δxavgAJ<δxavgAJ<δxavgOA
TAP-Net38.554.480.633.048.678.853.568.186.3
TAP-Net (Panning MOVi-E)39.556.281.436.052.980.150.165.788.3
TAPIR (MOVi-E)49.664.285.255.369.484.456.270.089.3
TAPIR (Panning MOVi-E)49.664.285.056.270.086.555.569.788.0
", + "bbox": [ + 78, + 89, + 467, + 147 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are slightly harder to track than those in the middle of the video, because there are more intervening frames wherein the point's appearance may change. Table 3 shows our results on this approach. We see that the performance broadly follows the results from the 'strided' evaluation: we outperform TAP-Net by a wide margin on real data (10% absolute on Kinetics, 20% on DAVIS), and by a smaller margin on RGB-Stacking (3%). This suggests that TAPIR remains robust even if the output frames are far from the query point.", + "bbox": [ + 76, + 229, + 467, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct a comparison of computational time for model inference on the DAVIS video horsejump-high with a resolution of $256 \\times 256$ . All models are evaluated on a single GPU using 5 runs on average. With Just In Time (JIT) Compilation with JAX, for 50 query points randomly sampled on the first frame, TAP-Net finishes within 0.1 seconds and TAPIR finishes within 0.3 seconds, i.e., roughly 150 frames per second. This is possible because TAPIR, like TAP-Net, maps efficiently to GPU parallel processing. In contrast, PIPs take 34.5 seconds due to a linear increase with respect to the number of points and frames processed. See Appendix 1 for more details.", + "bbox": [ + 76, + 366, + 467, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 559, + 240, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TAPIR vs Simple Combination We analyze the contributions of the main ideas. Our first contribution is to combine PIPs and TAP-Net (Fig. 3, center). Comparing this directly to PIPs is somewhat misleading: running the open-source 'chaining' code on Kinetics requires roughly 1 month on a V100 GPU (roughly 30 minutes to process a 10-second clip). But how crucial is chaining? After all, PIPs MLP-mixers have large spatial receptive fields due to their multi-resolution pyramid. Thus, we trained a naive version of PIPs without chaining: it initializes the entire track to be the query point, and then processes each chunk independently. For a fair comparison, we train using the same dataset and feature extractor, and other hyperparameters as TAPIR, but retain the 8-frame MLP Mixer with the same number of iterations as TAPIR. Table 4 shows that this algorithm (Chain-free PIPs) is surprisingly effective, on par with TAP-Net. Next, we combine the Chain-free PIPs with TAP-Net, replacing the constant initialization with TAP-Net's prediction. The \"simple combination\" (Fig. 3, center) alone improves over PIPs, Chain-free PIPs, and TAP-Net (i.e. $43.3\\%$ vs. $50.0\\%$ on DAVIS). However, the other new", + "bbox": [ + 76, + 583, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a4bfde8e56d762637d91c983d7def8211683d7c4162ce4fde04213fc2ee32c95.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. PIPs, TAP-Net, and a simple combination vs TAPIR. We compare TAPIR (Fig. 3 right) to TAP-Net, PIPs (Fig. 3 left), and the \"simple combination\" (Fig. 3 center). Chain-Free PIPs is our reimplementation of PIPs that removes the extremely slow chaining operation and uses the TAPIR backbone network. All models in this table are trained on Panning MOVi-E except PIPs, which uses its own dataset." + ], + "table_body": "
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubric
PIPs [19]35.342.037.359.1
TAP-Net [12]48.341.159.265.4
Chain-Free PIPs47.243.361.375.3
Simple Combination52.950.062.778.1
TAPIR57.261.362.784.7
", + "bbox": [ + 501, + 89, + 887, + 172 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74360d7d2ea35dec7ea6689570b7a450ea9e1c9977c2930cea0dee86a43c4323.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Model ablation by removing one major component at a time. -No Depthwise Conv uses a MLP-Mixer on a chunked video similar to PIPs. -No Uncertainty Estimate uses the losses directly as described in TAP-Net. -No Higher Res Feature uses only 4 pyramid levels in iterative refinement and has no features at stride 4. -No TAP-Net Initialization uses a constant initialization for iterative refinement. All the components contribute non-trivially to the performance of the model." + ], + "table_body": "
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubic
Full Model57.261.362.784.7
- No Depthwise Conv54.953.861.979.7
- No Uncertainty Estimate54.458.661.583.4
- No Higher Res Feature54.054.059.580.4
- No TAP-Net Initialization54.759.360.384.1
- No Iterative Refinement48.141.660.364.9
", + "bbox": [ + 501, + 279, + 890, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ideas in TAPIR also account for a substantial portion of the performance (i.e. $50.0\\%$ v.s. $61.3\\%$ on DAVIS).", + "bbox": [ + 500, + 494, + 890, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Components Table 5 shows the effect of our architectural decisions. First, we consider two novel architectural elements: the depthwise convolution which eliminates chunking from the PIPs model, and the uncertainty estimate. We see for Kinetics, both methods contribute roughly equally to performance, while for DAVIS, the depthwise conv is more important. A possible interpretation is that DAVIS contains more temporal continuity than Kinetics, so chunking is more harmful. Higher-resolution features are also predictably important to performance, especially on DAVIS. Table 5 also considers a few other model components that are not unique to TAPIR. TAP-Net initialization is important, but the model still gives reasonable performance without it. Interestingly, the full model without refinement (i.e. TAP-Net trained on Panning MOVi-E with the uncertainty estimate) performs similarly to TAP-Net trained on Panning MOVi-E (Table 4, second line); one possible interpretation is that TAP-Net is overfitting to its training data, leading to biased uncertainty estimate.", + "bbox": [ + 500, + 546, + 890, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Iterations of Refinement Finally, Table 6 shows how performance depends on the number of refinement iterations. We use the same number of iterations during training and", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10067", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7bd033911fe7e5d99bd3a9f4e3c8d8c8f3daa046be996530f354b2d454f6ab3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 90, + 380, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1c17f7774e3b29c030a58883aef1cd35b3a8f4087609ad3255d82740418fb5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 90, + 787, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d0c6f2e31b36d134e031dad1dd0cd91ecfe6e88e411874b39b28853b7303e79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 181, + 279, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df8cae52a6cfbbe5b38956f5eb7cae6b377f7fec0bb0463ddbb1a225fe1c6a9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 181, + 482, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b1c7eea26cda9008aebb980d994146f309845bdf1a00324dba7eaac8fe8cad2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 181, + 684, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ad20bbafd95ef327a0496dc830768d8e3f6f4bbbfbae337c4e1171d4c5c0e88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 181, + 887, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1399597778ab6bf89a0ff6c3a268f5afb60f19785cf5f9981df9c1a7a6c092a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 270, + 279, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5cc788129ca4c2062733f833d922e2ab373ae36d45f8861acced953e0b76f27e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 270, + 482, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5060358dbd638861945a6017ef28801c1228c654c8c76260d1fcc410f3049af2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 270, + 684, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77025247b256546591e8dec3d106a5283ce819c8fa945df15eaedb91fe960a31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 270, + 887, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/513d138e95ab57b4046015b9d475203a6f820b6af1354b5883068f8f9589daa6.jpg", + "image_caption": [ + "Figure 4. TAPIR compared to TAP-Net and PIPs. Top shows the query points on the video's first frame. Predictions for two later frames are below. We show predictions (filled circles) relative to ground truth (GT) (ends of the associated segments). x's indicate predictions where the GT is occluded, while empty circles are points visible in GT but predicted occluded (note: models still predict position regardless of occlusion). The majority of the street scene (left) gets occluded by a pedestrian; as a result, PIPs loses many points. TAP-Net fails on the right video, possibly because the textureless clothing is difficult to match without relying on temporal continuity. TAPIR, meanwhile, has far fewer failures." + ], + "image_footnote": [], + "bbox": [ + 81, + 359, + 279, + 445 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9bb190f89efcf11d53c2f28d060278fd150dc7bd2377502bb6bed8a434becf0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 359, + 482, + 445 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a7c33a15f5ac5169a9164dc633849fb12d6ed1520f4b64e736ae90096f625f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 359, + 684, + 445 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ee4ea0b56af29d946205efa1828a7b43082e0cb828e110b428465081e8de8c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 359, + 887, + 445 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7690427e7b56b4901890c670906b68308a525770719d87f6d8b44215e140a8d8.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Evaluation performance against the number of iterative updates. It can be observed that after 4 iterations, the performance no longer improves significantly. This could be attributed to the fact that TAP-Net already provides a good initialization." + ], + "table_body": "
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubric
0 iterations48.141.660.364.9
1 iteration55.755.062.780.1
2 iterations56.758.862.083.1
3 iterations56.860.660.984.1
4 iterations57.261.362.784.7
5 iterations57.861.661.584.3
6 iterations57.260.456.682.9
", + "bbox": [ + 78, + 546, + 468, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "testing for all models. Interestingly, we observed the best performance for TAPIR at 4-5 iterations, whereas PIPs used 6 iterations. Likely a stronger initialization (i.e., TAP-Net) requires fewer iterations to converge. The decrease in performance with more iterations implies that there may be some degenerate behavior, such as oversmoothing, which may be a topic for future research. In this paper, we use 4 iterations for all other experiments.", + "bbox": [ + 75, + 719, + 467, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Further model ablations, including experiments with RNNs and dense convolutions as replacements for our depthwise architecture, the number of feature pyramid layers, and the importance of time-shifting in the features, can", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "be found in Appendix 4. Depthwise convolutions work as well or better than the alternatives while being less expensive. We find little impact of adding more feature pyramid layers as PIPs does, so we remove them from TAPIR. Finally, we find the impact of time-shifting to be minor.", + "bbox": [ + 496, + 549, + 890, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Animating still images with TAPIR", + "text_level": 1, + "bbox": [ + 498, + 643, + 816, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generative modeling of images and videos has experienced an explosion in popularity due to recent striking text-conditioned generation results using diffusion models [17, 21, 42, 46, 50]. Animating still images remains extremely challenging, due to the high computational cost of video modeling and the difficulty of generating realistic motion. Is it possible to improve the generated videos via an explicit representation of surface motion? Here, we develop a proof-of-concept for a diffusion-based system that uses TAPIR tracks to do this.", + "bbox": [ + 496, + 671, + 890, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform video prediction using a pair of diffusion models: one which generates dense trajectories given an input image (the \"trajectory\" model), and a second which takes the input image and the trajectories and generates pixels (the \"pixel\" model). At test time, the outputs from the", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10068", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trajectory model dictate the motion that should be present in the video, ideally ensuring physical plausibility. The pixel model uses the trajectories to attend to the appropriate location in the first frame, in order to produce appearance which is consistent across the video.", + "bbox": [ + 75, + 90, + 470, + 165 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For the trajectory model, we encode the conditioning image with a small ConvNet, which produces a dense grid of features at stride 4. This is used as conditioning for a U-Net that performs diffusion denoising on the trajectories. We find that encoding the trajectories is important: for each frame, we include the raw $x$ and $y$ positions, both in absolute coordinates and coordinates relative to the starting position, and we also include a position encoding of the absolute locations with 32 Fourier features. The U-Net architecture includes self-attention layers, so the Fourier features allow the model to measure whether trajectories are spatially nearby late in an animation. For simplicity and memory efficiency, we stack these values for all frames along the channel axis: thus, the input is a 2D grid of shape $H // 4 \\times W // 4 \\times (T \\cdot (1 + 2 + 2 + 32))$ , (occlusion, relative position, absolute position, fourier encoding), which is processed with a 2D U-Net. The model predicts the clean data given the noisy data, which is the typical \"clean data prediction\" diffusion setup. The pixel model is a more traditional image generation model, which predicts the noise residual given noisy pixels. It takes three inputs, and outputs a single denoised frame. The first input is 3 noisy frames, allowing the model to infer some temporal context. The second is the first frame, warped according to the trajectories from the trajectory model. The third is features for the first frame, computed by a standard ConvNet before, being warped according to the same trajectories. We train on an internal dataset of 1M video clips, processed into 24-frame clips at $256 \\times 256$ resolution. See supplementary for more details.", + "bbox": [ + 75, + 167, + 472, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fig. 5 shows the trajectories produced by our trajectory prediction model on selected images. Our model can make multiple different physically-plausible predictions for motion for each input image, suggesting that it can recognize people and do rudimentary human pose estimation without supervision. Note that this demonstrates two levels of transfer: TAPIR was never trained on humans, while the trajectory-prediction model was trained on video rather than still images like these. It is difficult to judge these motions from a static figure, so we encourage the reader to view videos generated with our pipeline in the supplementary.", + "bbox": [ + 75, + 604, + 470, + 772 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 784, + 194, + 800 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we introduce TAPIR, a novel model for Tracking Any Point (TAP) that adopts a two-stage approach, comprising a matching stage and a refinement stage. It demonstrates a substantial improvement over the state-of-the-art, providing stable predictions and occlusion robustness, and scales to high-resolution videos. We also", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/816c5b75fa56ca10572812a20140fa2670a0ffe37d0150b3d7221821727ac666.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 88, + 694, + 234 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/98ae30fc91e6ce0f9c44cd5ccf47bcb8f4024926cba2ad210adf896b46698a26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 89, + 883, + 233 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ed4d56d95bf060156b51776ed607985448fdf374190505f424b0cf5e46e17b0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 234, + 694, + 378 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6e1cef806f85cced0c84db244635233367330bf56b3b7db5140207acf973696b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 234, + 883, + 378 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/55c5c6769ec5a4c23a01480463c00f80990234af931a9a34862891b7ae756672.jpg", + "image_caption": [ + "Figure 5. Animating still frames. Top row shows the input images. Each following row shows a visualization of a single sample from our trajectory model: the dark purple end is the starting point of the trajectory, while the colors get brighter and yellower with distance. In the left image, one sample shows the hand moving forward in a cutting motion, with a slight incline of the head; in the other, the entire person moves to the left. On the right, the first sample moves one arm in a circle, while the second shows the head turning and the other arm being raised." + ], + "image_footnote": [], + "bbox": [ + 506, + 378, + 694, + 522 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ccba319cec58fa54fa79f2a9956135a88e2862db7e713fd33dea9b61f65a1e48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 378, + 883, + 522 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "show a proof-of-concept of animating still frames, which we hope can serve as a foundation for future graphics research. TAPIR still has some limitations, such as difficulty in precisely perceiving the boundary between background and foreground, and it may still struggle with large appearance changes. Nevertheless, our findings indicate a promising future for TAP in computer vision.", + "bbox": [ + 496, + 664, + 890, + 772 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements We would like to thank Lucas Smaira and Larisa Markeeva for help with datasets and codebases, and Jon Scholz and Todor Davchev for advice on the needs of robotics. We also thank Adam Harley, Viorica Patraucean and Daniel Zoran for helpful discussions, and Mateusz Malinowski for paper comments. Finally, we thank Nando de Freitas for his support.", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10069", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5173-5182, 2017.", + "[2] Marian Stewart Bartlett, Gwen Littlewort, Mark G Frank, Claudia Lainscsek, Ian R Fasel, Javier R Movellan, et al. Automatic recognition of facial actions in spontaneous expressions. J. Multim., 1(6):22-35, 2006.", + "[3] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In Proceedings of European Conference on Computer Vision (ECCV), pages 404-417. Springer, 2006.", + "[4] Thomas Brox, Christoph Bregler, and Jitendra Malik. Large displacement optical flow. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 41-48. IEEE, 2009.", + "[5] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In Proceedings of European Conference on Computer Vision (ECCV), pages 25-36. Springer, 2004.", + "[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Proceedings of European Conference on Computer Vision (ECCV), pages 611-625. Springer, 2012.", + "[7] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4733-4742, 2016.", + "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6299-6308, 2017.", + "[9] Zezhou Cheng, Jong-Chyi Su, and Subhransu Maji. On equivariant and invariant learning of object landmark representations. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 9897-9906, 2021.", + "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5828-5839, 2017.", + "[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 224-236, 2018.", + "[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Proceedings of Neural Information Processing Systems Datasets and Benchmarks Track, 2022.", + "[13] Carl Doersch and Andrew Zisserman. Sim2real transfer learning for 3d human pose estimation: motion to the rescue. Proceedings of Neural Information Processing Systems (NeurIPS), 32, 2019." + ], + "bbox": [ + 78, + 114, + 470, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 2758-2766, 2015.", + "[15] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361. IEEE, 2012.", + "[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: a scalable dataset generator. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022.", + "[17] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10696-10706, 2022.", + "[18] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Denseseposé: Dense human pose estimation in the wild. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7297-7306, 2018.", + "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In Proceedings of European Conference on Computer Vision (ECCV), pages 59-75. Springer, 2022.", + "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003.", + "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022.", + "[22] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185-203, 1981.", + "[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017.", + "[24] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Proceedings of Neural Information Processing Systems (NeurIPS), 31, 2018.", + "[25] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabeled videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8787-8797, 2020.", + "[26] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia" + ], + "bbox": [ + 501, + 92, + 890, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10070", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3192-3199, 2013.", + "[27] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 6207-6217, 2021.", + "[28] Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B Choy, Philip HS Torr, and Manmohan Chandraker. Desire: Distant future prediction in dynamic scenes with interacting agents. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 336-345, 2017.", + "[29] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4521-4530, 2019.", + "[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2041-2050, 2018.", + "[31] Ji Lin, Chuang Gan, Kuan Wang, and Song Han. Tsm: Temporal shift module for efficient and scalable video understanding on edge devices. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020.", + "[32] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6728-6737, 2022.", + "[33] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):978-994, 2010.", + "[34] David G Lowe. Object recognition from local scale-invariant features. In Proceedings of IEEE International Conference on Computer Vision (ICCV), volume 2, pages 1150-1157. IEEE, 1999.", + "[35] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004.", + "[36] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In Proceedings of International Joint Conference on Artificial Intelligence (IJCAI), pages 674-679, 1981.", + "[37] Lucas Manuelli, Yunzhu Li, Pete Florence, and Russ Tedrake. Keypoints into the future: Self-supervised correspondence in model-based reinforcement learning. arXiv preprint arXiv:2009.05085, 2020.", + "[38] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Mantra: Memory augmented networks for multiple trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7143-7152, 2020.", + "[39] David Marr and Tomaso Poggio. A computational theory of human stereo vision. Proceedings of the Royal Society of London. Series B. Biological Sciences, 204(1156):301-328, 1979." + ], + "bbox": [ + 78, + 90, + 470, + 891 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016.", + "[41] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011.", + "[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.", + "[43] Yuki Ono, Eduard Trulls, Pascal Fua, and Kwang Moo Yi. Lf-net: Learning local features from images. Proceedings of Neural Information Processing Systems (NeurIPS), 31, 2018.", + "[44] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017.", + "[45] Deva Ramanan, David A Forsyth, and Andrew Zisserman. Strike a pose: Tracking people by finding stylized poses. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pages 271-278. IEEE, 2005.", + "[46] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "[47] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4161-4170, 2017.", + "[48] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3d reconstruction of interacting hands via collision-aware factorized refinements. In 2021 International Conference on 3D Vision (3DV), pages 432-441. IEEE, 2021.", + "[49] Michael Rubinstein, Ce Liu, and William Freeman. Towards longer long-range motion trajectories. In Proceedings of British Machine Vision Conference, 2012.", + "[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.", + "[51] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International Journal of Computer Vision, 80(1):72-91, 2008.", + "[52] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016.", + "[53] Thomas Schops, Johannes L Schonberger, Silvano Galliani," + ], + "bbox": [ + 501, + 92, + 893, + 892 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3260-3269, 2017.", + "[54] Ishwar K Sethi and Ramesh Jain. Finding trajectories of feature points in a monocular image sequence. IEEE Transactions on Pattern Analysis and Machine Intelligence, (1):56-73, 1987.", + "[55] Jie Shen, Stefanos Zafeiriou, Grigoris G Chrysos, Jean Kossaifi, Georgios Tzimiropoulos, and Maja Pantic. The first facial landmark tracking in-the-wild challenge: Benchmark and results. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, pages 50-58, 2015.", + "[56] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2107-2116, 2017.", + "[57] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. In Proceedings of European Conference on Computer Vision (ECCV), 2004.", + "[58] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6738-6748, 2022.", + "[59] Deqing Sun, Daniel Vlasic, Charles Herrmann, Varun Jampani, Michael Krainin, Huiwen Chang, Ramin Zabih, William T Freeman, and Ce Liu. Autoflow: Learning a better training set for optical flow. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10093-10102, 2021.", + "[60] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8934-8943, 2018.", + "[61] Yi Sun, Yuheng Chen, Xiaogang Wang, and Xiaou Tang. Deep learning face representation by joint identification-verification. Proceedings of Neural Information Processing Systems (NeurIPS), 27, 2014.", + "[62] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of European Conference on Computer Vision (ECCV), pages 402-419. Springer, 2020.", + "[63] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 5916-5925, 2017.", + "[64] Jonathan Tompson, Murphy Stein, Yann Lecun, and Ken Perlin. Real-time continuous pose recovery of human hands using convolutional networks. ACM Transactions on Graphics (ToG), 33(5):1-10, 2014.", + "[65] Philip HS Torr and Andrew Zisserman. Feature based methods for structure and motion estimation. In International" + ], + "bbox": [ + 78, + 90, + 470, + 892 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "workshop on vision algorithms, pages 278-294. Springer, 1999.", + "[66] Sudheendra Vijayanarasimhan, Susanna Ricco, Cordelia Schmid, Rahul Sukthankar, and Katerina Fragkiadaki. Sfmnet: Learning of structure and motion from video. arXiv preprint arXiv:1704.07804, 2017.", + "[67] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of European Conference on Computer Vision (ECCV), pages 601-617, 2018.", + "[68] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3551-3558, 2013.", + "[69] Yue Wang and Justin M Solomon. Prnet: Self-supervised learning for partial-to-partial registration. Proceedings of Neural Information Processing Systems (NeurIPS), 32, 2019.", + "[70] Xuehan Xiong and Fernando De la Torre. Supervised descent method and its applications to face alignment. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 532-539, 2013.", + "[71] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8121-8130, 2022.", + "[72] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 10075-10085, 2021.", + "[73] Zhichao Yin, Trevor Darrell, and Fisher Yu. Hierarchical discrete distribution decomposition for match density estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6044-6053, 2019.", + "[74] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2694-2703, 2018.", + "[75] Zhengyou Zhang. A flexible new technique for camera calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(11):1330-1334, 2000.", + "[76] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018.", + "[77] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2019.", + "[78] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rehmann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew Fitzgibbon, Charles Loop, Christian Theobalt, et al. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (ToG), 33(4):1-12, 2014." + ], + "bbox": [ + 501, + 92, + 890, + 892 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10072", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/f0fe9ecb-f819-46e1-ba9a-de70f0257649_model.json b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/f0fe9ecb-f819-46e1-ba9a-de70f0257649_model.json new file mode 100644 index 0000000000000000000000000000000000000000..78754c9555a20f95278722fe6a0098b719e78bfe --- /dev/null +++ b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/f0fe9ecb-f819-46e1-ba9a-de70f0257649_model.json @@ -0,0 +1,2523 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.13, + 0.847, + 0.175 + ], + "angle": 0, + "content": "TAPIR: Tracking Any Point with per-frame Initialization and temporal Refinement" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.203, + 0.238, + 0.221 + ], + "angle": 0, + "content": "Carl Doersch*" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.204, + 0.355, + 0.222 + ], + "angle": 0, + "content": "Yi Yang*" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.204, + 0.515, + 0.221 + ], + "angle": 0, + "content": "Mel Vecerik*†" + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.204, + 0.68, + 0.222 + ], + "angle": 0, + "content": "Dilara Gokay*" + }, + { + "type": "text", + "bbox": [ + 0.72, + 0.204, + 0.854, + 0.222 + ], + "angle": 0, + "content": "Ankush Gupta*" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.222, + 0.348, + 0.24 + ], + "angle": 0, + "content": "Yusuf Aytar*" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.222, + 0.516, + 0.239 + ], + "angle": 0, + "content": "Joao Carreira*" + }, + { + "type": "text", + "bbox": [ + 0.557, + 0.222, + 0.735, + 0.239 + ], + "angle": 0, + "content": "Andrew Zisserman*†" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.258, + 0.43, + 0.276 + ], + "angle": 0, + "content": "*Google DeepMind" + }, + { + "type": "text", + "bbox": [ + 0.47, + 0.258, + 0.699, + 0.276 + ], + "angle": 0, + "content": "† University College London" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.276, + 0.745, + 0.295 + ], + "angle": 0, + "content": "\\(^{\\ddagger}\\)VGG, Department of Engineering Science, University of Oxford" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.329, + 0.314, + 0.345 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.361, + 0.474, + 0.65 + ], + "angle": 0, + "content": "We present a novel model for Tracking Any Point (TAP) that effectively tracks any queried point on any physical surface throughout a video sequence. Our approach employs two stages: (1) a matching stage, which independently locates a suitable candidate point match for the query point on every other frame, and (2) a refinement stage, which updates both the trajectory and query features based on local correlations. The resulting model surpasses all baseline methods by a significant margin on the TAP-Vid benchmark, as demonstrated by an approximate \\(20\\%\\) absolute average Jaccard (AJ) improvement on DAVIS. Our model facilitates fast inference on long and high-resolution video sequences. On a modern GPU, our implementation has the capacity to track points faster than real-time. Given the high-quality trajectories extracted from a large dataset, we demonstrate a proof-of-concept diffusion model which generates trajectories from static images, enabling plausible animations. Visualizations, source code, and pretrained models can be found at https://deepmind-tapir.github.io." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.678, + 0.21, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.869 + ], + "angle": 0, + "content": "The problem of point level correspondence —i.e., determining whether two pixels in two different images are projections of the same point on the same physical surface—has long been a fundamental challenge in computer vision, with enormous potential for providing insights about physical properties and 3D shape. We consider its formulation as “Tracking Any Point” (TAP) [12]: given a video and (potentially dense) query points on solid surfaces, an algorithm should reliably output the locations those points correspond to in every other frame where they are visible, and indicate frames where they are not – see Fig. 4 for illustration." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Our main contribution is a new model: TAP with per-frame Initialization and temporal Refinement (TAPIR)," + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.332, + 0.889, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.622 + ], + "angle": 0, + "content": "Figure 1. Retrospective evolution of point tracking performance over time on the recent TAP-Vid-Kinetics and TAP-Vid-DAVIS benchmarks, as measured by Average Jaccard (higher is better). In this paper we introduce TAPIR, which significantly improves performance over the state-of-the-art. This unlocks new capabilities, which we demonstrate on motion-based future prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.852 + ], + "angle": 0, + "content": "which greatly improves over the state-of-the-art on the recently-proposed TAP-Vid benchmark [12]. There are many challenges to TAP: we must robustly estimate occlusions and recover when points reappear (unlike optical flow and structure-from-motion keypoints), meaning that search must be incorporated; yet when points remain visible for many consecutive frames, it is important to integrate information about appearance and motion across many of those frames in order to optimally predict positions. Furthermore, little real-world ground truth is available to learn from, so supervised-learning algorithms need to learn from synthetic data without overfitting to the data distribution (i.e., sim2real)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "There are three core design decisions that define TAPIR. The first is to use a coarse-to-fine approach. This approach has been used across many high-precision estimation prob" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10061" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.091, + 0.468, + 0.408 + ], + "angle": 0, + "content": "lems in computer vision [7, 28, 32, 38, 41, 58, 69, 73, 78]. For our application, the initial 'coarse' tracking consists of an occlusion-robust matching performed separately on every frame, where tracks are hypothesized using low-resolution features, without enforcing temporal continuity. The 'fine' refinement iteratively uses local, spatio-temporal information at a higher resolution, wherein a neural network can trade-off smoothness of motion with local appearance cues to produce the most likely track. The second design decision is to be fully-convolutional in time: the layers of our neural network consist principally of feature comparisons, spatial convolutions, and temporal convolutions, resulting in a model which efficiently maps onto modern GPU and TPU hardware. The third design decision is that the model should estimate its own uncertainty with regard to its position estimate, in a self-supervised manner. This ensures that low-confidence predictions can be suppressed, which improves benchmark scores. We hypothesize that this may help downstream algorithms (e.g. structure-from-motion) that rely on precision, and can benefit when low-quality matches are removed." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.412, + 0.468, + 0.668 + ], + "angle": 0, + "content": "We find that two existing architectures already have some of the pieces we need: TAP-Net [12] and Persistent Independent Particles (PIPs) [19]. Therefore, a key contribution of our work is to effectively combine them while achieving the benefits from both. TAP-Net performs a global search on every frame independently, providing a coarse track that is robust to occlusions. However, it does not make use of the continuity of videos, resulting in jittery, unrealistic tracks. PIPs, meanwhile, gives a recipe for refinement: given an initialization, it searches over a local neighborhood and smooths the track over time. However, PIPs processes videos sequentially in chunks, initializing each chunk with the output from the last. The procedure struggles with occlusion and is difficult to parallelize, resulting in slow processing (i.e., 1 month to evaluate TAP-Vid-Kinetics on 1 GPU). A key contribution of this work is observing the complementarity of these two methods." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.671, + 0.468, + 0.822 + ], + "angle": 0, + "content": "As shown in Fig. 1, we find that TAPIR improves over prior works by a large margin, as measured by performance on the TAP-Vid benchmark [12]. On TAP-Vid-DAVIS, TAPIR outperforms TAP-Net by \\(\\sim 20\\%\\) while on the more challenging TAP-Vid-Kinetics, TAPIR outperforms PIPs by \\(\\sim 20\\%\\), and substantially reduces its inference runtime. To demonstrate the quality of TAPIR trajectories, we showcase a proof-of-concept model trained to generate trajectories given individual images, and find that this model can generate plausible animations from single photographs." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "In summary, our contributions are as follows: 1) We propose a new model for long term point tracking, bridging the gap between TAP-Net and PIPs. 2) We show that the model achieves state-of-the-art results on the challenging TAP-Vid benchmark, with a significant boost on per" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.091, + 0.89, + 0.227 + ], + "angle": 0, + "content": "formance. 3) We provide an extensive analysis of the architectural decisions that matter for high-performance point tracking. 4) We provide a proof-of-concept of video prediction enabled by TAPIR's high-quality trajectories. Finally, 5) after analyzing components, we separately perform careful tuning of hyperparameters across entire method, in order to develop the best-performing model, which we release at https://www.github.com/deepmind/tapnet for the benefit of the community." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.247, + 0.64, + 0.263 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.274, + 0.89, + 0.304 + ], + "angle": 0, + "content": "Physical point correspondence and tracking have been studied in a wide range of settings." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.306, + 0.89, + 0.427 + ], + "angle": 0, + "content": "Optical Flow focuses on dense motion estimation between image pairs [22, 36]; methods can be divided into classical variational approaches [4, 5] and deep learning approaches [14, 23, 47, 60, 62, 71]. Optical flow operates only on subsequent frames, with no simple method to track across long videos. Groundtruth is hard to obtain [6], so it is typically benchmarked on synthetic scenes [6, 14, 40, 59], though limited real data exists through depth scanners [15]." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.429, + 0.89, + 0.625 + ], + "angle": 0, + "content": "Keypoint Correspondence methods aim at sparse keypoint matching given image pairs, from hand-defined discriminative feature descriptors [3, 34, 35, 54] to deep learning based methods [11, 27, 37, 43]. Though it operates on image pairs, it is arguably \"long term\" as the images may be taken at different times from wide baselines. The goal, however, is typically not to track any point, but to find easily-trackable \"keypoints\" sufficient for reconstruction. They are mainly used in structure-from-motion (SfM) settings and typically ignore occlusion, as SfM can use geometry to filter errors [20, 52, 65, 66]. This line of work is thus often restricted to rigid scenes [39, 75], and is benchmarked accordingly [1, 10, 29, 30, 53, 76]." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.627, + 0.89, + 0.793 + ], + "angle": 0, + "content": "Semantic Keypoint Tracking is often represented as keypoints or landmarks [26, 45, 56, 61, 64, 70]. Landmarks are often \"joints\", which can have vastly different appearance depending on pose (i.e. viewed from the front or back); thus most methods rely on large supervised datasets [2, 9, 48, 55, 72, 77], although some works track surface meshes [18, 67]. One interesting exception uses sim2real based on motion [13], motivating better surface tracking. Finally, some recent methods discover keypoints on moving objects [24, 25, 63, 74], though these typically require in-domain training on large datasets." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.796, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Long-term physical point tracking is addressed in a few early works [33, 49, 51, 57, 68], though such hand-engineered methods have fallen out of favor in the deep learning era. Our work instead builds on two recent works, TAP-Net [12] and Persistent Independent Particles (PIPs) [19], which aim to update these prior works to the deep learning era." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10062" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.455, + 0.473, + 0.609 + ], + "angle": 0, + "content": "Figure 2. TAPIR architecture summary. Our model begins with a global comparison between the query point features and the features for every other frame to compute an initial track estimate, including an uncertainty estimate. Then, we extract features from a local neighborhood (shown in pink) around the initial estimate, and compare these to the query feature at a higher resolution, post-processing the similarities with a temporal depthwise-convolitional network to get an updated position estimate. This updated position is fed back into the next iteration of refinement, repeated for a fixed number of iterations. Note that for simplicity, multi-scale pyramids are not shown." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.221, + 0.646 + ], + "angle": 0, + "content": "3. TAPIR Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Given a video and a query point, our goal is to estimate the 2D location \\( p_t \\) that it corresponds to in every other frame \\( t \\), as well as a 1D probability \\( o_t \\) that the point is occluded and a 1D probability \\( u_t \\) on the uncertainty of the estimated location. To be robust to occlusion, we first match candidate locations in other frames, by comparing the query features with all other features, and post-processing the similarities to arrive at an initial estimate. Then we refine the estimate, based on the local similarities between the query point and the target point. Note that both stages depend mostly on similarities (dot products) between the query point features and features elsewhere (i.e. not on the feature alone); this ensures that we don't overfit to specific features that might only appear in the (synthetic) training data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Fig. 2 gives an overview of our model. We initialize an estimate of the track by finding the best match within each" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.895, + 0.21 + ], + "angle": 0, + "content": "
TAP-Net [12]PIPs [19]TAPIR
Per-frame Initialization✓✗✓
Fully Convolutional In Time✓✗✓
Temporal Refinement✗✓✓
Multi Scale Feature✗✓✓
Stride-4 Feature✗test-time only✓
Parallel Inference✓✗✓
Uncertainty Estimate✗✗✓
Number of Parameters2.8M28.7M29.3M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.211, + 0.892, + 0.252 + ], + "angle": 0, + "content": "Table 1. Overview of models. TAPIR combines the merits from both TAP-Net and PIPs and further adds a handful of crucial components to improve performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.552 + ], + "angle": 0, + "content": "frame independently, ignoring the temporal nature of the video. This involves a cost volume: for each frame \\( t \\), we compute the dot product between the query feature \\( F_{q} \\) and all features in the \\( t \\)'th frame. We post-process the cost volume with a ConvNet, which produces a spatial heatmap, which is then summarized to a point estimate, in a manner broadly similar to TAP-Net. Given the initialization, we then compare the query feature to all features in a small region around an initial track. We feed the comparisons into a neural network, which updates both the query features and the track. We iteratively apply multiple such updates to arrive at a final trajectory estimate. TAPIR accomplishes this with a fully-convolutional architecture, operating on a pyramid of both high-dimensional, low-resolution features and low-dimensional, high-resolution features, in order to maximize efficiency on modern hardware. Finally, we add an estimate of uncertainty with respect to position throughout the architecture in order to suppress low-accuracy predictions. The rest of this section describes TAPIR in detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.687, + 0.575 + ], + "angle": 0, + "content": "3.1. Track Initialization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.704 + ], + "angle": 0, + "content": "The initial cost volume is computed using a relatively coarse feature map \\( F \\in \\mathbb{R}^{T \\times H/8 \\times W/8 \\times C} \\), where \\( T \\) is the number of frames, \\( H \\) and \\( W \\) are the image height and width, and \\( C \\) is the number of channels, computed with a standard TSM-ResNet-18 [31] backbone. Features for the query point \\( F_{q} \\) are extracted via bilinear interpolation at the query location, and we perform a dot product between this query feature and all other features in the video." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We compute an initial estimate of the position \\( p_t^0 = (x_t^0, y_t^0) \\) and occlusion \\( o_t^0 \\) by applying a small ConvNet to the cost volume corresponding to frame \\( t \\) (which is of shape \\( H / 8 \\times W / 8 \\times 1 \\) per query). This ConvNet has two outputs: a heatmap of shape \\( H / 8 \\times W / 8 \\times 1 \\) for predicting the position, and a single scalar logit for the occlusion estimate, obtained via average pooling followed by projection. The heatmap is converted to a position estimate by a \"spatial soft argmax\", i.e., a softmax across space (to make the heatmap positive and sum to 1). Afterwards, all heatmap values which are too far from the argmax position in this heatmap are set to zero. Then, the positions for each cell of the heatmap are averaged spatially, weighted by the thresholded heatmap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10063" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.214 + ], + "angle": 0, + "content": "magnitudes. Thus, the output is typically a location close to the heatmap's maximum; the \"soft argmax\" is differentiable, but the thresholding suppresses spurious matches, and prevents the network from \"averaging\" between multiple matches. This can serve as a good initialization, although the model struggles to localize points to less than a few pixels' accuracy on the original resolution due to the 8-pixel stride of the heatmap." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.238, + 0.473, + 0.555 + ], + "angle": 0, + "content": "Position Uncertainty Estimates. A drawback of predicting occlusion and position independently from the cost volume (a choice that is inspired by TAP-Net) is that if a point is visible in the ground truth, it's worse if the algorithm predicts a vastly incorrect location than if it simply incorrectly marks the point as occluded. After all, downstream applications may want to use the track to understand object motion; such downstream pipelines must be robust to occlusion, but may assume that the tracks are otherwise correct. The Average Jaccard metric reflects this: predictions in the wrong location are counted as both a \"false positive\" and a \"false negative.\" This kind of error tends to happen when the algorithm is uncertain about the position, e.g., if there are many potential matches. Therefore, we find it's beneficial for the algorithm to also estimate its own certainty regarding position. We quantify this by making the occlusion pathway output a second logit, estimating the probability \\( u_{t}^{0} \\) that the prediction is likely to be far enough from the ground truth that it isn't useful, even if the model predicts that it's visible. We define \"useful\" as whether the prediction is within a threshold \\( \\delta \\) of the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.556, + 0.471, + 0.587 + ], + "angle": 0, + "content": "Therefore, our loss for the initialization for a given video frame \\(t\\) is \\(\\mathcal{L}(p_t^0,o_t^0,u_t^0)\\), where \\(\\mathcal{L}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.614, + 0.469, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} \\left(p _ {t}, o _ {t}, u _ {t}\\right) = \\operatorname {H u b e r} \\left(\\hat {p} _ {t}, p _ {t}\\right) * \\left(1 - \\hat {o} _ {t}\\right) \\\\ + \\operatorname {B C E} \\left(\\hat {o} _ {t}, o _ {t}\\right) \\\\ + \\operatorname {B C E} \\left(\\hat {u} _ {t}, u _ {t}\\right) * \\left(1 - \\hat {o} _ {t}\\right) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.671, + 0.38, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e ,} \\quad \\hat {u} _ {t} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} d (\\hat {p} _ {t}, p _ {t}) > \\delta \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Here, \\(\\hat{o}_t \\in \\{0,1\\}\\) and \\(\\hat{p} \\in \\mathbb{R}^2\\) are the ground truth occlusion and point locations respectively, \\(d\\) is Euclidean distance, \\(\\delta\\) is the distance threshold, Huber is a Huber loss, and BCE is binary cross entropy (both \\(o_t\\) and \\(u_t\\) have sigmoids applied to convert them to probabilities). \\(\\hat{u}_t\\), the target for the uncertainty estimate \\(u_t\\), is computed from the ground truth position and the network's prediction: it's 0 if the model's position prediction is close enough to the ground truth (within the threshold \\(\\delta\\)), and 1 otherwise. At test time, the algorithm should output that the point is visible if it's both predicted as unoccluded and if the model is confident in the prediction. We therefore do a soft combination of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "the two probabilities: the algorithm outputs that the point is visible if \\( (1 - u_t) * (1 - o_t) > 0.5 \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.135, + 0.695, + 0.15 + ], + "angle": 0, + "content": "3.2. Iterative Refinement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.159, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Given an estimated position, occlusion, and uncertainty for each frame, the goal of each iteration \\( i \\) of our refinement procedure is to compute an update \\( (\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i}) \\), which adjusts the estimate to be closer to the ground truth, integrating information across time. The update is based on a set of \"local score maps\" which capture the query point similarity (i.e. dot products) to the features in the neighborhood of the trajectory. These are computed on a pyramid of different resolutions, so for a given trajectory, they have shape \\( (H' \\times W' \\times L) \\), where \\( H' = W' = 7 \\), the size of the local neighborhood, and \\( L \\) is the number of levels of the spatial pyramid (different pyramid levels are computed by spatially pooling the feature volume \\( F \\)). As with the initialization, this set of similarities is post-processed with a network to predict the refined position, occlusion, and uncertainty estimate. Unlike the initialization, however, we include \"local score maps\" for many frames simultaneously as input to the post-processing. We include the current position estimate, the raw query features, and the (flattened) local score maps into a tensor of shape \\( T \\times (C + K + 4) \\), where \\( C \\) is the number of channels in the query feature, \\( K = H' \\cdot W' \\cdot L \\) is the number of values in the flattened local score map, and 4 extra dimensions for position, occlusion, and uncertainty. The output of this network at the \\( i \\)'th iteration is a residual \\( (\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i}, \\Delta F_{q,t,i}) \\), which is added to the position, occlusion, uncertainty estimate, and the query feature, respectively. \\( \\Delta F_{q,t,i} \\) is of shape \\( T \\times C \\); thus, after the first iteration, slightly different \"query features\" are used on each frame when computing new local score maps." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.599, + 0.892, + 0.78 + ], + "angle": 0, + "content": "These positions and score maps are fed into a 12-block convolutional network to compute an update \\((\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i}, \\Delta F_{q,t,i})\\), where each block consists of a \\(1 \\times 1\\) convolution block and a depthwise convolution block. This architecture is inspired by the MLP-Mixer used for refinement in PIPs: we directly translate the Mixer's cross-channel layers into \\(1 \\times 1\\) convolutions with the same number of channels, and the within-channel operations become similarly within-channel depthwise convolutions. Unlike PIPs, which breaks sequences into 8-frame chunks before running the MLP-Mixer, this convolutional architecture can be run on sequences of any length." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We found that the feature maps used for computing the \"score maps\" are important for achieving high precision. The pyramid levels \\( l = 1 \\) through \\( L - 1 \\) are computed by spatially average-pooling the raw TSM-ResNet features \\( F \\) at a stride of \\( 8 \\cdot 2^{l - 1} \\), and then computing all dot products between the query feature \\( F_{q} \\) and the pyramid features. For the \\( t \\)'th frame, we extract a \\( 7 \\times 7 \\) patch of dot products centered at \\( p_t \\). At training time, PIPs leaves it here, but at" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10064" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.212 + ], + "angle": 0, + "content": "test time alters the backbone to run at stride 4, introducing a train/test domain gap. We find it is effective to train on stride 4 as well, although this puts pressure on system memory. To alleviate this, we compute a 0-th score map on the 64-channel, stride-4 input convolution map of the TSM-ResNet, and extract a comparable feature for the query from this feature map via bilinear interpolation. Thus, the final set of local score maps has shape \\((7 \\cdot 7 \\cdot L)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.468, + 0.289 + ], + "angle": 0, + "content": "After \\((\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i}, \\Delta F_{q,t,i})\\) is obtained from the above architecture, we can iterate the refinement step as many times as desired, re-using the same network parameters. At each iteration, we use the same loss \\(\\mathcal{L}\\) on the output, weighting each iteration the same as the initialization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.47, + 0.729 + ], + "angle": 0, + "content": "Discussion As there is some overlap with prior work, Figure 3 depicts the relationship between TAPIR and other architectures. For the sake of ablations, we also develop a \"simple combination\" model, which is intended to be a simpler combination of a TAP-Net-style initialization with PIPs-style refinements, and follows prior architectural decisions regardless of whether they are optimal. To summarize, the major architectural decisions needed to create the \"simple combination\" of TAP-Net and PIPs (and therefore the contributions of this model) are as follows. First we exploit the complementarity between TAP-Net and PIPs. Second, we remove \"chaining\" from PIPs, replacing that initialization with TAP-Net's initialization. We directly apply PIPs' MLP-Mixer architecture by 'chunking' the input sequence. Note that we adopt TAP-Net's feature network for both initialization and refinement. Finally, we adapt the TAP-Net backbone to extract a multi-scale and high resolution pyramid (follow PIPs stride-4 test-time architecture, but matching TAP-Net's end-to-end training). We apply losses at all layers of the hierarchy (whereas TAP-Net has a single loss, PIPs has no loss on its initialization). The additional contributions of TAPIR are as follows: first, we entirely replace the PIPs MLP-Mixer (which must be applied to fixed-length sequences) with a novel depthwise-convolitional architecture, which has a similar number of parameters but works on sequences of any length. Second, we introduce uncertainty estimate, which are computed at every level of the hierarchy. These are 'self-supervised' in the sense that the targets depend on the model's own output." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.245, + 0.758 + ], + "angle": 0, + "content": "3.3. Training Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Although TAPIR is designed to be robust to the sim2real gap, we can improve performance by minimizing the gap itself. One subtle difference between the Kubric MOViE [16] dataset and the real world is the lack of panning: although the Kubric camera moves, it is always set to \"look at\" a single point at the center of the workspace. Thus, stationary parts of the scene rotate around the center point of the workspace on the 2D image plane. Models thus tend to fail on real-world videos with panning. Therefore, we mod" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "ified the MOVi-E dataset so that the \"look at\" point moves along a random linear trajectory. The start and end points of this trajectory are enforced to be less than 1.5 units off the ground and lie within a radius of 4 units from the workspace center; the trajectory must also pass within 1 unit of the workspace center at some point, to ensure that the camera is usually looking toward the objects in the scene. See Appendix 5.2 for details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.228, + 0.634, + 0.246 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.892, + 0.526 + ], + "angle": 0, + "content": "We evaluate on the TAP-Vid [12] benchmark, a recent large-scale benchmark evaluating the problem of interest. TAP-Vid consists of point track annotations on four different datasets, each with different challenges. DAVIS [44] is a benchmark designed for tracking, and includes complex motion and large changes in object scale. Kinetics [8] contains over 1000 labeled videos, and contains all the complexities of YouTube videos, including difficulties like cuts and camera shake. RGB Stacking is a synthetic dataset of robotics videos with precise point tracks on large textureless regions. Kubric MOVi-E, used for training, contains realistic renderings of objects falling onto a flat surface; it comes with a validation set that we use for comparisons. We conduct training exclusively on the Kubric dataset, and selected our best model primarily by observing DAVIS. We performed no automated evaluation for hyperparameter tuning or model selection on any dataset. We train and evaluate at a resolution of \\(256 \\times 256\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Implementation Details We train for 50,000 steps with a cosine learning rate schedule, with a peak rate of \\(1 \\times 10^{-3}\\) and 1,000 warmup steps. We use 64 TPU-v3 cores, using 4 24-frame videos per core in each step. We use an Adam-W optimizer with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.95\\) and weight decay \\(1 \\times 10^{-2}\\). For most experiments, we use cross-replica batch normalization within the (TSM-)ResNet backbone only, although for our public model we find instance normalization is actually more effective. For most experiments we use \\(L = 5\\) to match PIPs, but in practice we find \\(L > 3\\) has little impact (See Appendix 4.3). Therefore, to save computation, our public model uses \\(L = 3\\). See Appendix 5.1 for more details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.756, + 0.733, + 0.773 + ], + "angle": 0, + "content": "4.1.Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Table 2 shows the performance of TAPIR relative to prior works. TAPIR improves by \\(10.6\\%\\) absolute performance over the best model on Kinetics (TAP-Net) and by \\(19.3\\%\\) on DAVIS (PIPs). These improvements are quite significant according to the TAP-Vid metrics: Average Jaccard consists of 5 thresholds, so under perfect occlusion estimation, a \\(20\\%\\) improvement requires improving all points to the next distance threshold (half the distance). It's worth noting" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10065" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.346, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.249, + 0.345, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.089, + 0.614, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.09, + 0.887, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.393, + 0.893, + 0.449 + ], + "angle": 0, + "content": "Figure 3. Comparison of Architectures. Left: the initial TAP-Net and PIPs models: TAP-Net has an independent estimate per frame, whereas PIPs breaks the video into fixed-size chunks and processes chunks sequentially. Middle: a \"simple combination\" of these two architectures, where TAP-Net is used to initialize PIPs-style chunked iterations. Right: Our model, TAPIR, which removes the chunking and adds uncertainty estimate. Note that for simplicity, multi-scale pyramids are not shown." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.459, + 0.89, + 0.607 + ], + "angle": 0, + "content": "
MethodKineticsKubricDAVISRGB-Stacking
AJ<δxavgOAAJ<δxavgOAAJ<δxavgOAAJ<δxavgOA
COTR [27]19.038.857.440.160.778.5535.451.380.26.813.579.1
Kubric-VFS-Like [16]40.559.080.051.969.884.633.148.579.457.972.691.9
RAFT [62]34.552.579.741.258.286.430.046.379.644.058.690.4
PIPs [19]35.354.877.459.174.888.642.059.482.137.351.091.6
TAP-Net [12]46.660.985.065.477.793.038.453.182.359.972.890.4
TAPIR (MOVi-E)57.170.087.684.391.895.859.872.387.666.277.493.3
TAPIR (Panning MOVi-E)57.270.187.884.792.195.861.373.688.862.774.691.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.608, + 0.893, + 0.649 + ], + "angle": 0, + "content": "Table 2. Comparison of TAPIR to prior results on TAP-Vid. \\( < \\delta_{avg}^{x} \\) is the fraction of points unoccluded in the ground truth, where the prediction is within a threshold, ignoring the occlusion estimate; OA is the accuracy of predicting occlusion. AJ is Average Jaccard, which considers both position and occlusion accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.473, + 0.896 + ], + "angle": 0, + "content": "that, like PIPs, TAPIR's performance on DAVIS is slightly higher than on Kinetics, whereas for TAP-Net, Kinetics has higher performance, suggesting that TAPIR's reliance on temporal continuity isn't as useful on Kinetics videos with cuts or camera shake. Note that the bulk of the improvement comes from the improved model rather than the improved dataset (TAPIR (MOVi-E) is trained on the same data as TAP-Net, while 'Panning' is the new dataset). The dataset's contribution is small, mostly improving DAVIS, but we found qualitatively that it makes the most substantial difference on backgrounds when cameras pan. This is not well represented in TAP-Vid's query points, but may matter for real-world applications like stabilization. RGB-stacking, however, is somewhat of an outlier: training on MOVi-E is still optimal (unsurprising as the cameras are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.668, + 0.892, + 0.79 + ], + "angle": 0, + "content": "static), and there the improvement is only \\(6.3\\%\\), possibly because TAPIR does less to help with textureless objects, indicating an area for future research. Fig. 4 shows some example predictions on DAVIS. TAPIR corrects many types of errors, including problems with temporal coherence from TAP-Net and problems with occlusion from PIPs. We include further qualitative results on our project webpage, which illustrate how noticeable a \\(20\\%\\) improvement is." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.903 + ], + "angle": 0, + "content": "TAP-Vid proposes another evaluation for videos at the resolution of \\(256 \\times 256\\) called 'query-first.' In the typical case (Table 2), the query points are sampled in a 'strided' fashion from the ground-truth tracks: the same trajectory is therefore queried with multiple query points, once every 5 frames. In the 'query first' evaluation, only the first point along each trajectory is used as a query. These points" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10066" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.09, + 0.468, + 0.148 + ], + "angle": 0, + "content": "
MethodKineticsOADAVISOARGB-Stacking
AJ<δxavgAJ<δxavgAJ<δxavgOA
TAP-Net38.554.480.633.048.678.853.568.186.3
TAP-Net (Panning MOVi-E)39.556.281.436.052.980.150.165.788.3
TAPIR (MOVi-E)49.664.285.255.369.484.456.270.089.3
TAPIR (Panning MOVi-E)49.664.285.056.270.086.555.569.788.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.078, + 0.149, + 0.468, + 0.216 + ], + "angle": 0, + "content": "Table 3. Comparison under query first metrics. We see largely the same relative performance trend whether the model is queried in a 'strided' fashion or whether the model is queried with the first frame where the point appeared, although performance is overall lower than the 'strided' evaluation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.231, + 0.468, + 0.366 + ], + "angle": 0, + "content": "are slightly harder to track than those in the middle of the video, because there are more intervening frames wherein the point's appearance may change. Table 3 shows our results on this approach. We see that the performance broadly follows the results from the 'strided' evaluation: we outperform TAP-Net by a wide margin on real data (10% absolute on Kinetics, 20% on DAVIS), and by a smaller margin on RGB-Stacking (3%). This suggests that TAPIR remains robust even if the output frames are far from the query point." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.367, + 0.468, + 0.548 + ], + "angle": 0, + "content": "We also conduct a comparison of computational time for model inference on the DAVIS video horsejump-high with a resolution of \\(256 \\times 256\\). All models are evaluated on a single GPU using 5 runs on average. With Just In Time (JIT) Compilation with JAX, for 50 query points randomly sampled on the first frame, TAP-Net finishes within 0.1 seconds and TAPIR finishes within 0.3 seconds, i.e., roughly 150 frames per second. This is possible because TAPIR, like TAP-Net, maps efficiently to GPU parallel processing. In contrast, PIPs take 34.5 seconds due to a linear increase with respect to the number of points and frames processed. See Appendix 1 for more details." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.56, + 0.241, + 0.575 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.584, + 0.468, + 0.9 + ], + "angle": 0, + "content": "TAPIR vs Simple Combination We analyze the contributions of the main ideas. Our first contribution is to combine PIPs and TAP-Net (Fig. 3, center). Comparing this directly to PIPs is somewhat misleading: running the open-source 'chaining' code on Kinetics requires roughly 1 month on a V100 GPU (roughly 30 minutes to process a 10-second clip). But how crucial is chaining? After all, PIPs MLP-mixers have large spatial receptive fields due to their multi-resolution pyramid. Thus, we trained a naive version of PIPs without chaining: it initializes the entire track to be the query point, and then processes each chunk independently. For a fair comparison, we train using the same dataset and feature extractor, and other hyperparameters as TAPIR, but retain the 8-frame MLP Mixer with the same number of iterations as TAPIR. Table 4 shows that this algorithm (Chain-free PIPs) is surprisingly effective, on par with TAP-Net. Next, we combine the Chain-free PIPs with TAP-Net, replacing the constant initialization with TAP-Net's prediction. The \"simple combination\" (Fig. 3, center) alone improves over PIPs, Chain-free PIPs, and TAP-Net (i.e. \\(43.3\\%\\) vs. \\(50.0\\%\\) on DAVIS). However, the other new" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.09, + 0.888, + 0.173 + ], + "angle": 0, + "content": "
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubric
PIPs [19]35.342.037.359.1
TAP-Net [12]48.341.159.265.4
Chain-Free PIPs47.243.361.375.3
Simple Combination52.950.062.778.1
TAPIR57.261.362.784.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.173, + 0.892, + 0.269 + ], + "angle": 0, + "content": "Table 4. PIPs, TAP-Net, and a simple combination vs TAPIR. We compare TAPIR (Fig. 3 right) to TAP-Net, PIPs (Fig. 3 left), and the \"simple combination\" (Fig. 3 center). Chain-Free PIPs is our reimplementation of PIPs that removes the extremely slow chaining operation and uses the TAPIR backbone network. All models in this table are trained on Panning MOVi-E except PIPs, which uses its own dataset." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.28, + 0.892, + 0.369 + ], + "angle": 0, + "content": "
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubic
Full Model57.261.362.784.7
- No Depthwise Conv54.953.861.979.7
- No Uncertainty Estimate54.458.661.583.4
- No Higher Res Feature54.054.059.580.4
- No TAP-Net Initialization54.759.360.384.1
- No Iterative Refinement48.141.660.364.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.5, + 0.371, + 0.892, + 0.48 + ], + "angle": 0, + "content": "Table 5. Model ablation by removing one major component at a time. -No Depthwise Conv uses a MLP-Mixer on a chunked video similar to PIPs. -No Uncertainty Estimate uses the losses directly as described in TAP-Net. -No Higher Res Feature uses only 4 pyramid levels in iterative refinement and has no features at stride 4. -No TAP-Net Initialization uses a constant initialization for iterative refinement. All the components contribute non-trivially to the performance of the model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.496, + 0.891, + 0.527 + ], + "angle": 0, + "content": "ideas in TAPIR also account for a substantial portion of the performance (i.e. \\(50.0\\%\\) v.s. \\(61.3\\%\\) on DAVIS)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.547, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Model Components Table 5 shows the effect of our architectural decisions. First, we consider two novel architectural elements: the depthwise convolution which eliminates chunking from the PIPs model, and the uncertainty estimate. We see for Kinetics, both methods contribute roughly equally to performance, while for DAVIS, the depthwise conv is more important. A possible interpretation is that DAVIS contains more temporal continuity than Kinetics, so chunking is more harmful. Higher-resolution features are also predictably important to performance, especially on DAVIS. Table 5 also considers a few other model components that are not unique to TAPIR. TAP-Net initialization is important, but the model still gives reasonable performance without it. Interestingly, the full model without refinement (i.e. TAP-Net trained on Panning MOVi-E with the uncertainty estimate) performs similarly to TAP-Net trained on Panning MOVi-E (Table 4, second line); one possible interpretation is that TAP-Net is overfitting to its training data, leading to biased uncertainty estimate." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Iterations of Refinement Finally, Table 6 shows how performance depends on the number of refinement iterations. We use the same number of iterations during training and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.092, + 0.382, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.092, + 0.788, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.182, + 0.281, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.182, + 0.483, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.182, + 0.685, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.182, + 0.888, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.271, + 0.281, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.271, + 0.483, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.271, + 0.685, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.271, + 0.888, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.36, + 0.281, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.36, + 0.483, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.36, + 0.685, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.36, + 0.888, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.454, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Figure 4. TAPIR compared to TAP-Net and PIPs. Top shows the query points on the video's first frame. Predictions for two later frames are below. We show predictions (filled circles) relative to ground truth (GT) (ends of the associated segments). x's indicate predictions where the GT is occluded, while empty circles are points visible in GT but predicted occluded (note: models still predict position regardless of occlusion). The majority of the street scene (left) gets occluded by a pedestrian; as a result, PIPs loses many points. TAP-Net fails on the right video, possibly because the textureless clothing is difficult to match without relying on temporal continuity. TAPIR, meanwhile, has far fewer failures." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.547, + 0.47, + 0.656 + ], + "angle": 0, + "content": "
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubric
0 iterations48.141.660.364.9
1 iteration55.755.062.780.1
2 iterations56.758.862.083.1
3 iterations56.860.660.984.1
4 iterations57.261.362.784.7
5 iterations57.861.661.584.3
6 iterations57.260.456.682.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.711 + ], + "angle": 0, + "content": "Table 6. Evaluation performance against the number of iterative updates. It can be observed that after 4 iterations, the performance no longer improves significantly. This could be attributed to the fact that TAP-Net already provides a good initialization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.839 + ], + "angle": 0, + "content": "testing for all models. Interestingly, we observed the best performance for TAPIR at 4-5 iterations, whereas PIPs used 6 iterations. Likely a stronger initialization (i.e., TAP-Net) requires fewer iterations to converge. The decrease in performance with more iterations implies that there may be some degenerate behavior, such as oversmoothing, which may be a topic for future research. In this paper, we use 4 iterations for all other experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Further model ablations, including experiments with RNNs and dense convolutions as replacements for our depthwise architecture, the number of feature pyramid layers, and the importance of time-shifting in the features, can" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.626 + ], + "angle": 0, + "content": "be found in Appendix 4. Depthwise convolutions work as well or better than the alternatives while being less expensive. We find little impact of adding more feature pyramid layers as PIPs does, so we remove them from TAPIR. Finally, we find the impact of time-shifting to be minor." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.645, + 0.818, + 0.663 + ], + "angle": 0, + "content": "5. Animating still images with TAPIR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Generative modeling of images and videos has experienced an explosion in popularity due to recent striking text-conditioned generation results using diffusion models [17, 21, 42, 46, 50]. Animating still images remains extremely challenging, due to the high computational cost of video modeling and the difficulty of generating realistic motion. Is it possible to improve the generated videos via an explicit representation of surface motion? Here, we develop a proof-of-concept for a diffusion-based system that uses TAPIR tracks to do this." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We perform video prediction using a pair of diffusion models: one which generates dense trajectories given an input image (the \"trajectory\" model), and a second which takes the input image and the trajectories and generates pixels (the \"pixel\" model). At test time, the outputs from the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10068" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.166 + ], + "angle": 0, + "content": "trajectory model dictate the motion that should be present in the video, ideally ensuring physical plausibility. The pixel model uses the trajectories to attend to the appropriate location in the first frame, in order to produce appearance which is consistent across the video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.473, + 0.605 + ], + "angle": 0, + "content": "For the trajectory model, we encode the conditioning image with a small ConvNet, which produces a dense grid of features at stride 4. This is used as conditioning for a U-Net that performs diffusion denoising on the trajectories. We find that encoding the trajectories is important: for each frame, we include the raw \\( x \\) and \\( y \\) positions, both in absolute coordinates and coordinates relative to the starting position, and we also include a position encoding of the absolute locations with 32 Fourier features. The U-Net architecture includes self-attention layers, so the Fourier features allow the model to measure whether trajectories are spatially nearby late in an animation. For simplicity and memory efficiency, we stack these values for all frames along the channel axis: thus, the input is a 2D grid of shape \\( H // 4 \\times W // 4 \\times (T \\cdot (1 + 2 + 2 + 32)) \\), (occlusion, relative position, absolute position, fourier encoding), which is processed with a 2D U-Net. The model predicts the clean data given the noisy data, which is the typical \"clean data prediction\" diffusion setup. The pixel model is a more traditional image generation model, which predicts the noise residual given noisy pixels. It takes three inputs, and outputs a single denoised frame. The first input is 3 noisy frames, allowing the model to infer some temporal context. The second is the first frame, warped according to the trajectories from the trajectory model. The third is features for the first frame, computed by a standard ConvNet before, being warped according to the same trajectories. We train on an internal dataset of 1M video clips, processed into 24-frame clips at \\( 256 \\times 256 \\) resolution. See supplementary for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.472, + 0.773 + ], + "angle": 0, + "content": "Fig. 5 shows the trajectories produced by our trajectory prediction model on selected images. Our model can make multiple different physically-plausible predictions for motion for each input image, suggesting that it can recognize people and do rudimentary human pose estimation without supervision. Note that this demonstrates two levels of transfer: TAPIR was never trained on humans, while the trajectory-prediction model was trained on video rather than still images like these. It is difficult to judge these motions from a static figure, so we encourage the reader to view videos generated with our pipeline in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.196, + 0.801 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce TAPIR, a novel model for Tracking Any Point (TAP) that adopts a two-stage approach, comprising a matching stage and a refinement stage. It demonstrates a substantial improvement over the state-of-the-art, providing stable predictions and occlusion robustness, and scales to high-resolution videos. We also" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.696, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.09, + 0.885, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.235, + 0.696, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.235, + 0.885, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.379, + 0.696, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.379, + 0.885, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Figure 5. Animating still frames. Top row shows the input images. Each following row shows a visualization of a single sample from our trajectory model: the dark purple end is the starting point of the trajectory, while the colors get brighter and yellower with distance. In the left image, one sample shows the hand moving forward in a cutting motion, with a slight incline of the head; in the other, the entire person moves to the left. On the right, the first sample moves one arm in a circle, while the second shows the head turning and the other arm being raised." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.773 + ], + "angle": 0, + "content": "show a proof-of-concept of animating still frames, which we hope can serve as a foundation for future graphics research. TAPIR still has some limitations, such as difficulty in precisely perceiving the boundary between background and foreground, and it may still struggle with large appearance changes. Nevertheless, our findings indicate a promising future for TAP in computer vision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements We would like to thank Lucas Smaira and Larisa Markeeva for help with datasets and codebases, and Jon Scholz and Todor Davchev for advice on the needs of robotics. We also thank Adam Harley, Viorica Patraucean and Daniel Zoran for helpful discussions, and Mateusz Malinowski for paper comments. Finally, we thank Nando de Freitas for his support." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10069" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5173-5182, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.184, + 0.471, + 0.238 + ], + "angle": 0, + "content": "[2] Marian Stewart Bartlett, Gwen Littlewort, Mark G Frank, Claudia Lainscsek, Ian R Fasel, Javier R Movellan, et al. Automatic recognition of facial actions in spontaneous expressions. J. Multim., 1(6):22-35, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.238, + 0.471, + 0.292 + ], + "angle": 0, + "content": "[3] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In Proceedings of European Conference on Computer Vision (ECCV), pages 404-417. Springer, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.292, + 0.471, + 0.347 + ], + "angle": 0, + "content": "[4] Thomas Brox, Christoph Bregler, and Jitendra Malik. Large displacement optical flow. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 41-48. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.347, + 0.471, + 0.401 + ], + "angle": 0, + "content": "[5] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In Proceedings of European Conference on Computer Vision (ECCV), pages 25-36. Springer, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Proceedings of European Conference on Computer Vision (ECCV), pages 611-625. Springer, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.455, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[7] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4733-4742, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6299-6308, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.577, + 0.471, + 0.631 + ], + "angle": 0, + "content": "[9] Zezhou Cheng, Jong-Chyi Su, and Subhransu Maji. On equivariant and invariant learning of object landmark representations. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 9897-9906, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5828-5839, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.47, + 0.767 + ], + "angle": 0, + "content": "[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 224-236, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.767, + 0.47, + 0.835 + ], + "angle": 0, + "content": "[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Proceedings of Neural Information Processing Systems Datasets and Benchmarks Track, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.835, + 0.47, + 0.889 + ], + "angle": 0, + "content": "[13] Carl Doersch and Andrew Zisserman. Sim2real transfer learning for 3d human pose estimation: motion to the rescue. Proceedings of Neural Information Processing Systems (NeurIPS), 32, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 2758-2766, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.176, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[15] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361. IEEE, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.243, + 0.892, + 0.408 + ], + "angle": 0, + "content": "[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: a scalable dataset generator. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.408, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[17] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10696-10706, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[18] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Denseseposé: Dense human pose estimation in the wild. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7297-7306, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.53, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In Proceedings of European Conference on Computer Vision (ECCV), pages 59-75. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.584, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[22] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185-203, 1981." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.759, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[24] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Proceedings of Neural Information Processing Systems (NeurIPS), 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.814, + 0.892, + 0.88 + ], + "angle": 0, + "content": "[25] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabeled videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8787-8797, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.88, + 0.892, + 0.895 + ], + "angle": 0, + "content": "[26] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10070" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3192-3199, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.201 + ], + "angle": 0, + "content": "[27] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 6207-6217, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.202, + 0.471, + 0.27 + ], + "angle": 0, + "content": "[28] Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B Choy, Philip HS Torr, and Manmohan Chandraker. Desire: Distant future prediction in dynamic scenes with interacting agents. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 336-345, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.27, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[29] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4521-4530, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.471, + 0.392 + ], + "angle": 0, + "content": "[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2041-2050, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[31] Ji Lin, Chuang Gan, Kuan Wang, and Song Han. Tsm: Temporal shift module for efficient and scalable video understanding on edge devices. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.471, + 0.513 + ], + "angle": 0, + "content": "[32] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6728-6737, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.471, + 0.568 + ], + "angle": 0, + "content": "[33] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):978-994, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.471, + 0.621 + ], + "angle": 0, + "content": "[34] David G Lowe. Object recognition from local scale-invariant features. In Proceedings of IEEE International Conference on Computer Vision (ICCV), volume 2, pages 1150-1157. IEEE, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[35] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.471, + 0.717 + ], + "angle": 0, + "content": "[36] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In Proceedings of International Joint Conference on Artificial Intelligence (IJCAI), pages 674-679, 1981." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[37] Lucas Manuelli, Yunzhu Li, Pete Florence, and Russ Tedrake. Keypoints into the future: Self-supervised correspondence in model-based reinforcement learning. arXiv preprint arXiv:2009.05085, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.771, + 0.471, + 0.839 + ], + "angle": 0, + "content": "[38] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Mantra: Memory augmented networks for multiple trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7143-7152, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.839, + 0.471, + 0.892 + ], + "angle": 0, + "content": "[39] David Marr and Tomaso Poggio. A computational theory of human stereo vision. Proceedings of the Royal Society of London. Series B. Biological Sciences, 204(1156):301-328, 1979." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[40] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.176, + 0.894, + 0.257 + ], + "angle": 0, + "content": "[41] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.257, + 0.894, + 0.324 + ], + "angle": 0, + "content": "[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.325, + 0.894, + 0.366 + ], + "angle": 0, + "content": "[43] Yuki Ono, Eduard Trulls, Pascal Fua, and Kwang Moo Yi. Lf-net: Learning local features from images. Proceedings of Neural Information Processing Systems (NeurIPS), 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.366, + 0.894, + 0.419 + ], + "angle": 0, + "content": "[44] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.893, + 0.487 + ], + "angle": 0, + "content": "[45] Deva Ramanan, David A Forsyth, and Andrew Zisserman. Strike a pose: Tracking people by finding stylized poses. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pages 271-278. IEEE, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.487, + 0.893, + 0.541 + ], + "angle": 0, + "content": "[46] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.894, + 0.595 + ], + "angle": 0, + "content": "[47] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4161-4170, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.893, + 0.662 + ], + "angle": 0, + "content": "[48] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3d reconstruction of interacting hands via collision-aware factorized refinements. In 2021 International Conference on 3D Vision (3DV), pages 432-441. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.894, + 0.704 + ], + "angle": 0, + "content": "[49] Michael Rubinstein, Ce Liu, and William Freeman. Towards longer long-range motion trajectories. In Proceedings of British Machine Vision Conference, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.893, + 0.786 + ], + "angle": 0, + "content": "[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.786, + 0.893, + 0.826 + ], + "angle": 0, + "content": "[51] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International Journal of Computer Vision, 80(1):72-91, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.826, + 0.893, + 0.879 + ], + "angle": 0, + "content": "[52] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.879, + 0.893, + 0.893 + ], + "angle": 0, + "content": "[53] Thomas Schops, Johannes L Schonberger, Silvano Galliani," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10071" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3260-3269, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.161, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[54] Ishwar K Sethi and Ramesh Jain. Finding trajectories of feature points in a monocular image sequence. IEEE Transactions on Pattern Analysis and Machine Intelligence, (1):56-73, 1987." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.215, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[55] Jie Shen, Stefanos Zafeiriou, Grigoris G Chrysos, Jean Kossaifi, Georgios Tzimiropoulos, and Maja Pantic. The first facial landmark tracking in-the-wild challenge: Benchmark and results. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, pages 50-58, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.471, + 0.378 + ], + "angle": 0, + "content": "[56] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2107-2116, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[57] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. In Proceedings of European Conference on Computer Vision (ECCV), 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.501 + ], + "angle": 0, + "content": "[58] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6738-6748, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.501, + 0.469, + 0.583 + ], + "angle": 0, + "content": "[59] Deqing Sun, Daniel Vlasic, Charles Herrmann, Varun Jampani, Michael Krainin, Huiwen Chang, Ramin Zabih, William T Freeman, and Ce Liu. Autoflow: Learning a better training set for optical flow. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10093-10102, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.583, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[60] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8934-8943, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[61] Yi Sun, Yuheng Chen, Xiaogang Wang, and Xiaou Tang. Deep learning face representation by joint identification-verification. Proceedings of Neural Information Processing Systems (NeurIPS), 27, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[62] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of European Conference on Computer Vision (ECCV), pages 402-419. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[63] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 5916-5925, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.813, + 0.469, + 0.867 + ], + "angle": 0, + "content": "[64] Jonathan Tompson, Murphy Stein, Yann Lecun, and Ken Perlin. Real-time continuous pose recovery of human hands using convolutional networks. ACM Transactions on Graphics (ToG), 33(5):1-10, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.868, + 0.469, + 0.893 + ], + "angle": 0, + "content": "[65] Philip HS Torr and Andrew Zisserman. Feature based methods for structure and motion estimation. In International" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "workshop on vision algorithms, pages 278-294. Springer, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[66] Sudheendra Vijayanarasimhan, Susanna Ricco, Cordelia Schmid, Rahul Sukthankar, and Katerina Fragkiadaki. Sfmnet: Learning of structure and motion from video. arXiv preprint arXiv:1704.07804, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.174, + 0.892, + 0.242 + ], + "angle": 0, + "content": "[67] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of European Conference on Computer Vision (ECCV), pages 601-617, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.242, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[68] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3551-3558, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.295, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[69] Yue Wang and Justin M Solomon. Prnet: Self-supervised learning for partial-to-partial registration. Proceedings of Neural Information Processing Systems (NeurIPS), 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[70] Xuehan Xiong and Fernando De la Torre. Supervised descent method and its applications to face alignment. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 532-539, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[71] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8121-8130, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[72] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 10075-10085, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[73] Zhichao Yin, Trevor Darrell, and Fisher Yu. Hierarchical discrete distribution decomposition for match density estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6044-6053, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[74] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2694-2703, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[75] Zhengyou Zhang. A flexible new technique for camera calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(11):1330-1334, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[76] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.81 + ], + "angle": 0, + "content": "[77] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.81, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[78] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rehmann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew Fitzgibbon, Charles Loop, Christian Theobalt, et al. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (ToG), 33(4):1-12, 2014." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10072" + } + ] +] \ No newline at end of file diff --git a/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/f0fe9ecb-f819-46e1-ba9a-de70f0257649_origin.pdf b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/f0fe9ecb-f819-46e1-ba9a-de70f0257649_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6c99bce35018e5f84e66cd8f8346de73f594c46f --- /dev/null +++ b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/f0fe9ecb-f819-46e1-ba9a-de70f0257649_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28d6cf4ee2287a5c3ac390df98dfb3ecbf328a3ad04ce5a20a3b7dfaa312761c +size 4994027 diff --git a/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/full.md b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9cfdf321193858d5f71a03a4556671907a3a4bd1 --- /dev/null +++ b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/full.md @@ -0,0 +1,337 @@ +# TAPIR: Tracking Any Point with per-frame Initialization and temporal Refinement + +Carl Doersch* + +Yi Yang* + +Mel Vecerik*† + +Dilara Gokay* + +Ankush Gupta* + +Yusuf Aytar* + +Joao Carreira* + +Andrew Zisserman*† + +*Google DeepMind + +† University College London + +$^{\ddagger}$ VGG, Department of Engineering Science, University of Oxford + +# Abstract + +We present a novel model for Tracking Any Point (TAP) that effectively tracks any queried point on any physical surface throughout a video sequence. Our approach employs two stages: (1) a matching stage, which independently locates a suitable candidate point match for the query point on every other frame, and (2) a refinement stage, which updates both the trajectory and query features based on local correlations. The resulting model surpasses all baseline methods by a significant margin on the TAP-Vid benchmark, as demonstrated by an approximate $20\%$ absolute average Jaccard (AJ) improvement on DAVIS. Our model facilitates fast inference on long and high-resolution video sequences. On a modern GPU, our implementation has the capacity to track points faster than real-time. Given the high-quality trajectories extracted from a large dataset, we demonstrate a proof-of-concept diffusion model which generates trajectories from static images, enabling plausible animations. Visualizations, source code, and pretrained models can be found at https://deepmind-tapir.github.io. + +# 1. Introduction + +The problem of point level correspondence —i.e., determining whether two pixels in two different images are projections of the same point on the same physical surface—has long been a fundamental challenge in computer vision, with enormous potential for providing insights about physical properties and 3D shape. We consider its formulation as “Tracking Any Point” (TAP) [12]: given a video and (potentially dense) query points on solid surfaces, an algorithm should reliably output the locations those points correspond to in every other frame where they are visible, and indicate frames where they are not – see Fig. 4 for illustration. + +Our main contribution is a new model: TAP with per-frame Initialization and temporal Refinement (TAPIR), + +![](images/90016a7bcb80eadb6771cb4c9979ff7cdfcdad16777a4611e5689c693d102ff2.jpg) +Figure 1. Retrospective evolution of point tracking performance over time on the recent TAP-Vid-Kinetics and TAP-Vid-DAVIS benchmarks, as measured by Average Jaccard (higher is better). In this paper we introduce TAPIR, which significantly improves performance over the state-of-the-art. This unlocks new capabilities, which we demonstrate on motion-based future prediction. + +which greatly improves over the state-of-the-art on the recently-proposed TAP-Vid benchmark [12]. There are many challenges to TAP: we must robustly estimate occlusions and recover when points reappear (unlike optical flow and structure-from-motion keypoints), meaning that search must be incorporated; yet when points remain visible for many consecutive frames, it is important to integrate information about appearance and motion across many of those frames in order to optimally predict positions. Furthermore, little real-world ground truth is available to learn from, so supervised-learning algorithms need to learn from synthetic data without overfitting to the data distribution (i.e., sim2real). + +There are three core design decisions that define TAPIR. The first is to use a coarse-to-fine approach. This approach has been used across many high-precision estimation prob + +lems in computer vision [7, 28, 32, 38, 41, 58, 69, 73, 78]. For our application, the initial 'coarse' tracking consists of an occlusion-robust matching performed separately on every frame, where tracks are hypothesized using low-resolution features, without enforcing temporal continuity. The 'fine' refinement iteratively uses local, spatio-temporal information at a higher resolution, wherein a neural network can trade-off smoothness of motion with local appearance cues to produce the most likely track. The second design decision is to be fully-convolutional in time: the layers of our neural network consist principally of feature comparisons, spatial convolutions, and temporal convolutions, resulting in a model which efficiently maps onto modern GPU and TPU hardware. The third design decision is that the model should estimate its own uncertainty with regard to its position estimate, in a self-supervised manner. This ensures that low-confidence predictions can be suppressed, which improves benchmark scores. We hypothesize that this may help downstream algorithms (e.g. structure-from-motion) that rely on precision, and can benefit when low-quality matches are removed. + +We find that two existing architectures already have some of the pieces we need: TAP-Net [12] and Persistent Independent Particles (PIPs) [19]. Therefore, a key contribution of our work is to effectively combine them while achieving the benefits from both. TAP-Net performs a global search on every frame independently, providing a coarse track that is robust to occlusions. However, it does not make use of the continuity of videos, resulting in jittery, unrealistic tracks. PIPs, meanwhile, gives a recipe for refinement: given an initialization, it searches over a local neighborhood and smooths the track over time. However, PIPs processes videos sequentially in chunks, initializing each chunk with the output from the last. The procedure struggles with occlusion and is difficult to parallelize, resulting in slow processing (i.e., 1 month to evaluate TAP-Vid-Kinetics on 1 GPU). A key contribution of this work is observing the complementarity of these two methods. + +As shown in Fig. 1, we find that TAPIR improves over prior works by a large margin, as measured by performance on the TAP-Vid benchmark [12]. On TAP-Vid-DAVIS, TAPIR outperforms TAP-Net by $\sim 20\%$ while on the more challenging TAP-Vid-Kinetics, TAPIR outperforms PIPs by $\sim 20\%$ , and substantially reduces its inference runtime. To demonstrate the quality of TAPIR trajectories, we showcase a proof-of-concept model trained to generate trajectories given individual images, and find that this model can generate plausible animations from single photographs. + +In summary, our contributions are as follows: 1) We propose a new model for long term point tracking, bridging the gap between TAP-Net and PIPs. 2) We show that the model achieves state-of-the-art results on the challenging TAP-Vid benchmark, with a significant boost on per + +formance. 3) We provide an extensive analysis of the architectural decisions that matter for high-performance point tracking. 4) We provide a proof-of-concept of video prediction enabled by TAPIR's high-quality trajectories. Finally, 5) after analyzing components, we separately perform careful tuning of hyperparameters across entire method, in order to develop the best-performing model, which we release at https://www.github.com/deepmind/tapnet for the benefit of the community. + +# 2. Related Work + +Physical point correspondence and tracking have been studied in a wide range of settings. + +Optical Flow focuses on dense motion estimation between image pairs [22, 36]; methods can be divided into classical variational approaches [4, 5] and deep learning approaches [14, 23, 47, 60, 62, 71]. Optical flow operates only on subsequent frames, with no simple method to track across long videos. Groundtruth is hard to obtain [6], so it is typically benchmarked on synthetic scenes [6, 14, 40, 59], though limited real data exists through depth scanners [15]. + +Keypoint Correspondence methods aim at sparse keypoint matching given image pairs, from hand-defined discriminative feature descriptors [3, 34, 35, 54] to deep learning based methods [11, 27, 37, 43]. Though it operates on image pairs, it is arguably "long term" as the images may be taken at different times from wide baselines. The goal, however, is typically not to track any point, but to find easily-trackable "keypoints" sufficient for reconstruction. They are mainly used in structure-from-motion (SfM) settings and typically ignore occlusion, as SfM can use geometry to filter errors [20, 52, 65, 66]. This line of work is thus often restricted to rigid scenes [39, 75], and is benchmarked accordingly [1, 10, 29, 30, 53, 76]. + +Semantic Keypoint Tracking is often represented as keypoints or landmarks [26, 45, 56, 61, 64, 70]. Landmarks are often "joints", which can have vastly different appearance depending on pose (i.e. viewed from the front or back); thus most methods rely on large supervised datasets [2, 9, 48, 55, 72, 77], although some works track surface meshes [18, 67]. One interesting exception uses sim2real based on motion [13], motivating better surface tracking. Finally, some recent methods discover keypoints on moving objects [24, 25, 63, 74], though these typically require in-domain training on large datasets. + +Long-term physical point tracking is addressed in a few early works [33, 49, 51, 57, 68], though such hand-engineered methods have fallen out of favor in the deep learning era. Our work instead builds on two recent works, TAP-Net [12] and Persistent Independent Particles (PIPs) [19], which aim to update these prior works to the deep learning era. + +![](images/902c79206953071b29eda02102ddbd7a9487ae746f178a91972e615271fa6a5b.jpg) +Figure 2. TAPIR architecture summary. Our model begins with a global comparison between the query point features and the features for every other frame to compute an initial track estimate, including an uncertainty estimate. Then, we extract features from a local neighborhood (shown in pink) around the initial estimate, and compare these to the query feature at a higher resolution, post-processing the similarities with a temporal depthwise-convolitional network to get an updated position estimate. This updated position is fed back into the next iteration of refinement, repeated for a fixed number of iterations. Note that for simplicity, multi-scale pyramids are not shown. + +# 3. TAPIR Model + +Given a video and a query point, our goal is to estimate the 2D location $p_t$ that it corresponds to in every other frame $t$ , as well as a 1D probability $o_t$ that the point is occluded and a 1D probability $u_t$ on the uncertainty of the estimated location. To be robust to occlusion, we first match candidate locations in other frames, by comparing the query features with all other features, and post-processing the similarities to arrive at an initial estimate. Then we refine the estimate, based on the local similarities between the query point and the target point. Note that both stages depend mostly on similarities (dot products) between the query point features and features elsewhere (i.e. not on the feature alone); this ensures that we don't overfit to specific features that might only appear in the (synthetic) training data. + +Fig. 2 gives an overview of our model. We initialize an estimate of the track by finding the best match within each + +
TAP-Net [12]PIPs [19]TAPIR
Per-frame Initialization✓✗✓
Fully Convolutional In Time✓✗✓
Temporal Refinement✗✓✓
Multi Scale Feature✗✓✓
Stride-4 Feature✗test-time only✓
Parallel Inference✓✗✓
Uncertainty Estimate✗✗✓
Number of Parameters2.8M28.7M29.3M
+ +Table 1. Overview of models. TAPIR combines the merits from both TAP-Net and PIPs and further adds a handful of crucial components to improve performance. + +frame independently, ignoring the temporal nature of the video. This involves a cost volume: for each frame $t$ , we compute the dot product between the query feature $F_{q}$ and all features in the $t$ 'th frame. We post-process the cost volume with a ConvNet, which produces a spatial heatmap, which is then summarized to a point estimate, in a manner broadly similar to TAP-Net. Given the initialization, we then compare the query feature to all features in a small region around an initial track. We feed the comparisons into a neural network, which updates both the query features and the track. We iteratively apply multiple such updates to arrive at a final trajectory estimate. TAPIR accomplishes this with a fully-convolutional architecture, operating on a pyramid of both high-dimensional, low-resolution features and low-dimensional, high-resolution features, in order to maximize efficiency on modern hardware. Finally, we add an estimate of uncertainty with respect to position throughout the architecture in order to suppress low-accuracy predictions. The rest of this section describes TAPIR in detail. + +# 3.1. Track Initialization + +The initial cost volume is computed using a relatively coarse feature map $F \in \mathbb{R}^{T \times H/8 \times W/8 \times C}$ , where $T$ is the number of frames, $H$ and $W$ are the image height and width, and $C$ is the number of channels, computed with a standard TSM-ResNet-18 [31] backbone. Features for the query point $F_{q}$ are extracted via bilinear interpolation at the query location, and we perform a dot product between this query feature and all other features in the video. + +We compute an initial estimate of the position $p_t^0 = (x_t^0, y_t^0)$ and occlusion $o_t^0$ by applying a small ConvNet to the cost volume corresponding to frame $t$ (which is of shape $H / 8 \times W / 8 \times 1$ per query). This ConvNet has two outputs: a heatmap of shape $H / 8 \times W / 8 \times 1$ for predicting the position, and a single scalar logit for the occlusion estimate, obtained via average pooling followed by projection. The heatmap is converted to a position estimate by a "spatial soft argmax", i.e., a softmax across space (to make the heatmap positive and sum to 1). Afterwards, all heatmap values which are too far from the argmax position in this heatmap are set to zero. Then, the positions for each cell of the heatmap are averaged spatially, weighted by the thresholded heatmap + +magnitudes. Thus, the output is typically a location close to the heatmap's maximum; the "soft argmax" is differentiable, but the thresholding suppresses spurious matches, and prevents the network from "averaging" between multiple matches. This can serve as a good initialization, although the model struggles to localize points to less than a few pixels' accuracy on the original resolution due to the 8-pixel stride of the heatmap. + +Position Uncertainty Estimates. A drawback of predicting occlusion and position independently from the cost volume (a choice that is inspired by TAP-Net) is that if a point is visible in the ground truth, it's worse if the algorithm predicts a vastly incorrect location than if it simply incorrectly marks the point as occluded. After all, downstream applications may want to use the track to understand object motion; such downstream pipelines must be robust to occlusion, but may assume that the tracks are otherwise correct. The Average Jaccard metric reflects this: predictions in the wrong location are counted as both a "false positive" and a "false negative." This kind of error tends to happen when the algorithm is uncertain about the position, e.g., if there are many potential matches. Therefore, we find it's beneficial for the algorithm to also estimate its own certainty regarding position. We quantify this by making the occlusion pathway output a second logit, estimating the probability $u_{t}^{0}$ that the prediction is likely to be far enough from the ground truth that it isn't useful, even if the model predicts that it's visible. We define "useful" as whether the prediction is within a threshold $\delta$ of the ground truth. + +Therefore, our loss for the initialization for a given video frame $t$ is $\mathcal{L}(p_t^0,o_t^0,u_t^0)$ , where $\mathcal{L}$ is defined as: + +$$ +\begin{array}{l} \mathcal {L} \left(p _ {t}, o _ {t}, u _ {t}\right) = \operatorname {H u b e r} \left(\hat {p} _ {t}, p _ {t}\right) * \left(1 - \hat {o} _ {t}\right) \\ + \operatorname {B C E} \left(\hat {o} _ {t}, o _ {t}\right) \\ + \operatorname {B C E} \left(\hat {u} _ {t}, u _ {t}\right) * \left(1 - \hat {o} _ {t}\right) \tag {1} \\ \end{array} +$$ + +$$ +\text {w h e r e ,} \quad \hat {u} _ {t} = \left\{ \begin{array}{l l} 1 & \text {i f} d (\hat {p} _ {t}, p _ {t}) > \delta \\ 0 & \text {o t h e r w i s e} \end{array} \right. +$$ + +Here, $\hat{o}_t \in \{0,1\}$ and $\hat{p} \in \mathbb{R}^2$ are the ground truth occlusion and point locations respectively, $d$ is Euclidean distance, $\delta$ is the distance threshold, Huber is a Huber loss, and BCE is binary cross entropy (both $o_t$ and $u_t$ have sigmoids applied to convert them to probabilities). $\hat{u}_t$ , the target for the uncertainty estimate $u_t$ , is computed from the ground truth position and the network's prediction: it's 0 if the model's position prediction is close enough to the ground truth (within the threshold $\delta$ ), and 1 otherwise. At test time, the algorithm should output that the point is visible if it's both predicted as unoccluded and if the model is confident in the prediction. We therefore do a soft combination of + +the two probabilities: the algorithm outputs that the point is visible if $(1 - u_t) * (1 - o_t) > 0.5$ . + +# 3.2. Iterative Refinement + +Given an estimated position, occlusion, and uncertainty for each frame, the goal of each iteration $i$ of our refinement procedure is to compute an update $(\Delta p_{t}^{i}, \Delta o_{t}^{i}, \Delta u_{t}^{i})$ , which adjusts the estimate to be closer to the ground truth, integrating information across time. The update is based on a set of "local score maps" which capture the query point similarity (i.e. dot products) to the features in the neighborhood of the trajectory. These are computed on a pyramid of different resolutions, so for a given trajectory, they have shape $(H' \times W' \times L)$ , where $H' = W' = 7$ , the size of the local neighborhood, and $L$ is the number of levels of the spatial pyramid (different pyramid levels are computed by spatially pooling the feature volume $F$ ). As with the initialization, this set of similarities is post-processed with a network to predict the refined position, occlusion, and uncertainty estimate. Unlike the initialization, however, we include "local score maps" for many frames simultaneously as input to the post-processing. We include the current position estimate, the raw query features, and the (flattened) local score maps into a tensor of shape $T \times (C + K + 4)$ , where $C$ is the number of channels in the query feature, $K = H' \cdot W' \cdot L$ is the number of values in the flattened local score map, and 4 extra dimensions for position, occlusion, and uncertainty. The output of this network at the $i$ 'th iteration is a residual $(\Delta p_{t}^{i}, \Delta o_{t}^{i}, \Delta u_{t}^{i}, \Delta F_{q,t,i})$ , which is added to the position, occlusion, uncertainty estimate, and the query feature, respectively. $\Delta F_{q,t,i}$ is of shape $T \times C$ ; thus, after the first iteration, slightly different "query features" are used on each frame when computing new local score maps. + +These positions and score maps are fed into a 12-block convolutional network to compute an update $(\Delta p_{t}^{i}, \Delta o_{t}^{i}, \Delta u_{t}^{i}, \Delta F_{q,t,i})$ , where each block consists of a $1 \times 1$ convolution block and a depthwise convolution block. This architecture is inspired by the MLP-Mixer used for refinement in PIPs: we directly translate the Mixer's cross-channel layers into $1 \times 1$ convolutions with the same number of channels, and the within-channel operations become similarly within-channel depthwise convolutions. Unlike PIPs, which breaks sequences into 8-frame chunks before running the MLP-Mixer, this convolutional architecture can be run on sequences of any length. + +We found that the feature maps used for computing the "score maps" are important for achieving high precision. The pyramid levels $l = 1$ through $L - 1$ are computed by spatially average-pooling the raw TSM-ResNet features $F$ at a stride of $8 \cdot 2^{l - 1}$ , and then computing all dot products between the query feature $F_{q}$ and the pyramid features. For the $t$ 'th frame, we extract a $7 \times 7$ patch of dot products centered at $p_t$ . At training time, PIPs leaves it here, but at + +test time alters the backbone to run at stride 4, introducing a train/test domain gap. We find it is effective to train on stride 4 as well, although this puts pressure on system memory. To alleviate this, we compute a 0-th score map on the 64-channel, stride-4 input convolution map of the TSM-ResNet, and extract a comparable feature for the query from this feature map via bilinear interpolation. Thus, the final set of local score maps has shape $(7 \cdot 7 \cdot L)$ . + +After $(\Delta p_{t}^{i}, \Delta o_{t}^{i}, \Delta u_{t}^{i}, \Delta F_{q,t,i})$ is obtained from the above architecture, we can iterate the refinement step as many times as desired, re-using the same network parameters. At each iteration, we use the same loss $\mathcal{L}$ on the output, weighting each iteration the same as the initialization. + +Discussion As there is some overlap with prior work, Figure 3 depicts the relationship between TAPIR and other architectures. For the sake of ablations, we also develop a "simple combination" model, which is intended to be a simpler combination of a TAP-Net-style initialization with PIPs-style refinements, and follows prior architectural decisions regardless of whether they are optimal. To summarize, the major architectural decisions needed to create the "simple combination" of TAP-Net and PIPs (and therefore the contributions of this model) are as follows. First we exploit the complementarity between TAP-Net and PIPs. Second, we remove "chaining" from PIPs, replacing that initialization with TAP-Net's initialization. We directly apply PIPs' MLP-Mixer architecture by 'chunking' the input sequence. Note that we adopt TAP-Net's feature network for both initialization and refinement. Finally, we adapt the TAP-Net backbone to extract a multi-scale and high resolution pyramid (follow PIPs stride-4 test-time architecture, but matching TAP-Net's end-to-end training). We apply losses at all layers of the hierarchy (whereas TAP-Net has a single loss, PIPs has no loss on its initialization). The additional contributions of TAPIR are as follows: first, we entirely replace the PIPs MLP-Mixer (which must be applied to fixed-length sequences) with a novel depthwise-convolitional architecture, which has a similar number of parameters but works on sequences of any length. Second, we introduce uncertainty estimate, which are computed at every level of the hierarchy. These are 'self-supervised' in the sense that the targets depend on the model's own output. + +# 3.3. Training Dataset + +Although TAPIR is designed to be robust to the sim2real gap, we can improve performance by minimizing the gap itself. One subtle difference between the Kubric MOViE [16] dataset and the real world is the lack of panning: although the Kubric camera moves, it is always set to "look at" a single point at the center of the workspace. Thus, stationary parts of the scene rotate around the center point of the workspace on the 2D image plane. Models thus tend to fail on real-world videos with panning. Therefore, we mod + +ified the MOVi-E dataset so that the "look at" point moves along a random linear trajectory. The start and end points of this trajectory are enforced to be less than 1.5 units off the ground and lie within a radius of 4 units from the workspace center; the trajectory must also pass within 1 unit of the workspace center at some point, to ensure that the camera is usually looking toward the objects in the scene. See Appendix 5.2 for details. + +# 4. Experiments + +We evaluate on the TAP-Vid [12] benchmark, a recent large-scale benchmark evaluating the problem of interest. TAP-Vid consists of point track annotations on four different datasets, each with different challenges. DAVIS [44] is a benchmark designed for tracking, and includes complex motion and large changes in object scale. Kinetics [8] contains over 1000 labeled videos, and contains all the complexities of YouTube videos, including difficulties like cuts and camera shake. RGB Stacking is a synthetic dataset of robotics videos with precise point tracks on large textureless regions. Kubric MOVi-E, used for training, contains realistic renderings of objects falling onto a flat surface; it comes with a validation set that we use for comparisons. We conduct training exclusively on the Kubric dataset, and selected our best model primarily by observing DAVIS. We performed no automated evaluation for hyperparameter tuning or model selection on any dataset. We train and evaluate at a resolution of $256 \times 256$ . + +Implementation Details We train for 50,000 steps with a cosine learning rate schedule, with a peak rate of $1 \times 10^{-3}$ and 1,000 warmup steps. We use 64 TPU-v3 cores, using 4 24-frame videos per core in each step. We use an Adam-W optimizer with $\beta_{1} = 0.9$ and $\beta_{2} = 0.95$ and weight decay $1 \times 10^{-2}$ . For most experiments, we use cross-replica batch normalization within the (TSM-)ResNet backbone only, although for our public model we find instance normalization is actually more effective. For most experiments we use $L = 5$ to match PIPs, but in practice we find $L > 3$ has little impact (See Appendix 4.3). Therefore, to save computation, our public model uses $L = 3$ . See Appendix 5.1 for more details. + +# 4.1.Quantitative Comparison + +Table 2 shows the performance of TAPIR relative to prior works. TAPIR improves by $10.6\%$ absolute performance over the best model on Kinetics (TAP-Net) and by $19.3\%$ on DAVIS (PIPs). These improvements are quite significant according to the TAP-Vid metrics: Average Jaccard consists of 5 thresholds, so under perfect occlusion estimation, a $20\%$ improvement requires improving all points to the next distance threshold (half the distance). It's worth noting + +![](images/e4b5154c9ea6e729709200e681129490f2cea7954f486846da0cfa18e43b511f.jpg) + +![](images/6656db070a993c553263d72a54561b74ca215fbbce060ef0b2f849b0166b572b.jpg) +Figure 3. Comparison of Architectures. Left: the initial TAP-Net and PIPs models: TAP-Net has an independent estimate per frame, whereas PIPs breaks the video into fixed-size chunks and processes chunks sequentially. Middle: a "simple combination" of these two architectures, where TAP-Net is used to initialize PIPs-style chunked iterations. Right: Our model, TAPIR, which removes the chunking and adds uncertainty estimate. Note that for simplicity, multi-scale pyramids are not shown. + +![](images/f43c633f7edbb99052e2883b2306a9ef81cbd330d986f3e608f3e47939be1122.jpg) + +![](images/05a05cb07531b542d157902b8e83a28bd058c3647010f2bd4478d233841f48bf.jpg) + +
MethodKineticsKubricDAVISRGB-Stacking
AJ<δxavgOAAJ<δxavgOAAJ<δxavgOAAJ<δxavgOA
COTR [27]19.038.857.440.160.778.5535.451.380.26.813.579.1
Kubric-VFS-Like [16]40.559.080.051.969.884.633.148.579.457.972.691.9
RAFT [62]34.552.579.741.258.286.430.046.379.644.058.690.4
PIPs [19]35.354.877.459.174.888.642.059.482.137.351.091.6
TAP-Net [12]46.660.985.065.477.793.038.453.182.359.972.890.4
TAPIR (MOVi-E)57.170.087.684.391.895.859.872.387.666.277.493.3
TAPIR (Panning MOVi-E)57.270.187.884.792.195.861.373.688.862.774.691.6
+ +Table 2. Comparison of TAPIR to prior results on TAP-Vid. $< \delta_{avg}^{x}$ is the fraction of points unoccluded in the ground truth, where the prediction is within a threshold, ignoring the occlusion estimate; OA is the accuracy of predicting occlusion. AJ is Average Jaccard, which considers both position and occlusion accuracy. + +that, like PIPs, TAPIR's performance on DAVIS is slightly higher than on Kinetics, whereas for TAP-Net, Kinetics has higher performance, suggesting that TAPIR's reliance on temporal continuity isn't as useful on Kinetics videos with cuts or camera shake. Note that the bulk of the improvement comes from the improved model rather than the improved dataset (TAPIR (MOVi-E) is trained on the same data as TAP-Net, while 'Panning' is the new dataset). The dataset's contribution is small, mostly improving DAVIS, but we found qualitatively that it makes the most substantial difference on backgrounds when cameras pan. This is not well represented in TAP-Vid's query points, but may matter for real-world applications like stabilization. RGB-stacking, however, is somewhat of an outlier: training on MOVi-E is still optimal (unsurprising as the cameras are + +static), and there the improvement is only $6.3\%$ , possibly because TAPIR does less to help with textureless objects, indicating an area for future research. Fig. 4 shows some example predictions on DAVIS. TAPIR corrects many types of errors, including problems with temporal coherence from TAP-Net and problems with occlusion from PIPs. We include further qualitative results on our project webpage, which illustrate how noticeable a $20\%$ improvement is. + +TAP-Vid proposes another evaluation for videos at the resolution of $256 \times 256$ called 'query-first.' In the typical case (Table 2), the query points are sampled in a 'strided' fashion from the ground-truth tracks: the same trajectory is therefore queried with multiple query points, once every 5 frames. In the 'query first' evaluation, only the first point along each trajectory is used as a query. These points + +
MethodKineticsOADAVISOARGB-Stacking
AJ<δxavgAJ<δxavgAJ<δxavgOA
TAP-Net38.554.480.633.048.678.853.568.186.3
TAP-Net (Panning MOVi-E)39.556.281.436.052.980.150.165.788.3
TAPIR (MOVi-E)49.664.285.255.369.484.456.270.089.3
TAPIR (Panning MOVi-E)49.664.285.056.270.086.555.569.788.0
+ +Table 3. Comparison under query first metrics. We see largely the same relative performance trend whether the model is queried in a 'strided' fashion or whether the model is queried with the first frame where the point appeared, although performance is overall lower than the 'strided' evaluation. + +are slightly harder to track than those in the middle of the video, because there are more intervening frames wherein the point's appearance may change. Table 3 shows our results on this approach. We see that the performance broadly follows the results from the 'strided' evaluation: we outperform TAP-Net by a wide margin on real data (10% absolute on Kinetics, 20% on DAVIS), and by a smaller margin on RGB-Stacking (3%). This suggests that TAPIR remains robust even if the output frames are far from the query point. + +We also conduct a comparison of computational time for model inference on the DAVIS video horsejump-high with a resolution of $256 \times 256$ . All models are evaluated on a single GPU using 5 runs on average. With Just In Time (JIT) Compilation with JAX, for 50 query points randomly sampled on the first frame, TAP-Net finishes within 0.1 seconds and TAPIR finishes within 0.3 seconds, i.e., roughly 150 frames per second. This is possible because TAPIR, like TAP-Net, maps efficiently to GPU parallel processing. In contrast, PIPs take 34.5 seconds due to a linear increase with respect to the number of points and frames processed. See Appendix 1 for more details. + +# 4.2. Ablation Studies + +TAPIR vs Simple Combination We analyze the contributions of the main ideas. Our first contribution is to combine PIPs and TAP-Net (Fig. 3, center). Comparing this directly to PIPs is somewhat misleading: running the open-source 'chaining' code on Kinetics requires roughly 1 month on a V100 GPU (roughly 30 minutes to process a 10-second clip). But how crucial is chaining? After all, PIPs MLP-mixers have large spatial receptive fields due to their multi-resolution pyramid. Thus, we trained a naive version of PIPs without chaining: it initializes the entire track to be the query point, and then processes each chunk independently. For a fair comparison, we train using the same dataset and feature extractor, and other hyperparameters as TAPIR, but retain the 8-frame MLP Mixer with the same number of iterations as TAPIR. Table 4 shows that this algorithm (Chain-free PIPs) is surprisingly effective, on par with TAP-Net. Next, we combine the Chain-free PIPs with TAP-Net, replacing the constant initialization with TAP-Net's prediction. The "simple combination" (Fig. 3, center) alone improves over PIPs, Chain-free PIPs, and TAP-Net (i.e. $43.3\%$ vs. $50.0\%$ on DAVIS). However, the other new + +
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubric
PIPs [19]35.342.037.359.1
TAP-Net [12]48.341.159.265.4
Chain-Free PIPs47.243.361.375.3
Simple Combination52.950.062.778.1
TAPIR57.261.362.784.7
+ +Table 4. PIPs, TAP-Net, and a simple combination vs TAPIR. We compare TAPIR (Fig. 3 right) to TAP-Net, PIPs (Fig. 3 left), and the "simple combination" (Fig. 3 center). Chain-Free PIPs is our reimplementation of PIPs that removes the extremely slow chaining operation and uses the TAPIR backbone network. All models in this table are trained on Panning MOVi-E except PIPs, which uses its own dataset. + +
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubic
Full Model57.261.362.784.7
- No Depthwise Conv54.953.861.979.7
- No Uncertainty Estimate54.458.661.583.4
- No Higher Res Feature54.054.059.580.4
- No TAP-Net Initialization54.759.360.384.1
- No Iterative Refinement48.141.660.364.9
+ +Table 5. Model ablation by removing one major component at a time. -No Depthwise Conv uses a MLP-Mixer on a chunked video similar to PIPs. -No Uncertainty Estimate uses the losses directly as described in TAP-Net. -No Higher Res Feature uses only 4 pyramid levels in iterative refinement and has no features at stride 4. -No TAP-Net Initialization uses a constant initialization for iterative refinement. All the components contribute non-trivially to the performance of the model. + +ideas in TAPIR also account for a substantial portion of the performance (i.e. $50.0\%$ v.s. $61.3\%$ on DAVIS). + +Model Components Table 5 shows the effect of our architectural decisions. First, we consider two novel architectural elements: the depthwise convolution which eliminates chunking from the PIPs model, and the uncertainty estimate. We see for Kinetics, both methods contribute roughly equally to performance, while for DAVIS, the depthwise conv is more important. A possible interpretation is that DAVIS contains more temporal continuity than Kinetics, so chunking is more harmful. Higher-resolution features are also predictably important to performance, especially on DAVIS. Table 5 also considers a few other model components that are not unique to TAPIR. TAP-Net initialization is important, but the model still gives reasonable performance without it. Interestingly, the full model without refinement (i.e. TAP-Net trained on Panning MOVi-E with the uncertainty estimate) performs similarly to TAP-Net trained on Panning MOVi-E (Table 4, second line); one possible interpretation is that TAP-Net is overfitting to its training data, leading to biased uncertainty estimate. + +Iterations of Refinement Finally, Table 6 shows how performance depends on the number of refinement iterations. We use the same number of iterations during training and + +![](images/7bd033911fe7e5d99bd3a9f4e3c8d8c8f3daa046be996530f354b2d454f6ab3f.jpg) + +![](images/f1c17f7774e3b29c030a58883aef1cd35b3a8f4087609ad3255d82740418fb5a.jpg) + +![](images/1d0c6f2e31b36d134e031dad1dd0cd91ecfe6e88e411874b39b28853b7303e79.jpg) + +![](images/df8cae52a6cfbbe5b38956f5eb7cae6b377f7fec0bb0463ddbb1a225fe1c6a9c.jpg) + +![](images/4b1c7eea26cda9008aebb980d994146f309845bdf1a00324dba7eaac8fe8cad2.jpg) + +![](images/9ad20bbafd95ef327a0496dc830768d8e3f6f4bbbfbae337c4e1171d4c5c0e88.jpg) + +![](images/1399597778ab6bf89a0ff6c3a268f5afb60f19785cf5f9981df9c1a7a6c092a9.jpg) + +![](images/5cc788129ca4c2062733f833d922e2ab373ae36d45f8861acced953e0b76f27e.jpg) + +![](images/5060358dbd638861945a6017ef28801c1228c654c8c76260d1fcc410f3049af2.jpg) + +![](images/77025247b256546591e8dec3d106a5283ce819c8fa945df15eaedb91fe960a31.jpg) + +![](images/513d138e95ab57b4046015b9d475203a6f820b6af1354b5883068f8f9589daa6.jpg) +Figure 4. TAPIR compared to TAP-Net and PIPs. Top shows the query points on the video's first frame. Predictions for two later frames are below. We show predictions (filled circles) relative to ground truth (GT) (ends of the associated segments). x's indicate predictions where the GT is occluded, while empty circles are points visible in GT but predicted occluded (note: models still predict position regardless of occlusion). The majority of the street scene (left) gets occluded by a pedestrian; as a result, PIPs loses many points. TAP-Net fails on the right video, possibly because the textureless clothing is difficult to match without relying on temporal continuity. TAPIR, meanwhile, has far fewer failures. + +![](images/c9bb190f89efcf11d53c2f28d060278fd150dc7bd2377502bb6bed8a434becf0.jpg) + +![](images/2a7c33a15f5ac5169a9164dc633849fb12d6ed1520f4b64e736ae90096f625f4.jpg) + +![](images/4ee4ea0b56af29d946205efa1828a7b43082e0cb828e110b428465081e8de8c7.jpg) + +
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubric
0 iterations48.141.660.364.9
1 iteration55.755.062.780.1
2 iterations56.758.862.083.1
3 iterations56.860.660.984.1
4 iterations57.261.362.784.7
5 iterations57.861.661.584.3
6 iterations57.260.456.682.9
+ +Table 6. Evaluation performance against the number of iterative updates. It can be observed that after 4 iterations, the performance no longer improves significantly. This could be attributed to the fact that TAP-Net already provides a good initialization. + +testing for all models. Interestingly, we observed the best performance for TAPIR at 4-5 iterations, whereas PIPs used 6 iterations. Likely a stronger initialization (i.e., TAP-Net) requires fewer iterations to converge. The decrease in performance with more iterations implies that there may be some degenerate behavior, such as oversmoothing, which may be a topic for future research. In this paper, we use 4 iterations for all other experiments. + +Further model ablations, including experiments with RNNs and dense convolutions as replacements for our depthwise architecture, the number of feature pyramid layers, and the importance of time-shifting in the features, can + +be found in Appendix 4. Depthwise convolutions work as well or better than the alternatives while being less expensive. We find little impact of adding more feature pyramid layers as PIPs does, so we remove them from TAPIR. Finally, we find the impact of time-shifting to be minor. + +# 5. Animating still images with TAPIR + +Generative modeling of images and videos has experienced an explosion in popularity due to recent striking text-conditioned generation results using diffusion models [17, 21, 42, 46, 50]. Animating still images remains extremely challenging, due to the high computational cost of video modeling and the difficulty of generating realistic motion. Is it possible to improve the generated videos via an explicit representation of surface motion? Here, we develop a proof-of-concept for a diffusion-based system that uses TAPIR tracks to do this. + +We perform video prediction using a pair of diffusion models: one which generates dense trajectories given an input image (the "trajectory" model), and a second which takes the input image and the trajectories and generates pixels (the "pixel" model). At test time, the outputs from the + +trajectory model dictate the motion that should be present in the video, ideally ensuring physical plausibility. The pixel model uses the trajectories to attend to the appropriate location in the first frame, in order to produce appearance which is consistent across the video. + +For the trajectory model, we encode the conditioning image with a small ConvNet, which produces a dense grid of features at stride 4. This is used as conditioning for a U-Net that performs diffusion denoising on the trajectories. We find that encoding the trajectories is important: for each frame, we include the raw $x$ and $y$ positions, both in absolute coordinates and coordinates relative to the starting position, and we also include a position encoding of the absolute locations with 32 Fourier features. The U-Net architecture includes self-attention layers, so the Fourier features allow the model to measure whether trajectories are spatially nearby late in an animation. For simplicity and memory efficiency, we stack these values for all frames along the channel axis: thus, the input is a 2D grid of shape $H // 4 \times W // 4 \times (T \cdot (1 + 2 + 2 + 32))$ , (occlusion, relative position, absolute position, fourier encoding), which is processed with a 2D U-Net. The model predicts the clean data given the noisy data, which is the typical "clean data prediction" diffusion setup. The pixel model is a more traditional image generation model, which predicts the noise residual given noisy pixels. It takes three inputs, and outputs a single denoised frame. The first input is 3 noisy frames, allowing the model to infer some temporal context. The second is the first frame, warped according to the trajectories from the trajectory model. The third is features for the first frame, computed by a standard ConvNet before, being warped according to the same trajectories. We train on an internal dataset of 1M video clips, processed into 24-frame clips at $256 \times 256$ resolution. See supplementary for more details. + +Fig. 5 shows the trajectories produced by our trajectory prediction model on selected images. Our model can make multiple different physically-plausible predictions for motion for each input image, suggesting that it can recognize people and do rudimentary human pose estimation without supervision. Note that this demonstrates two levels of transfer: TAPIR was never trained on humans, while the trajectory-prediction model was trained on video rather than still images like these. It is difficult to judge these motions from a static figure, so we encourage the reader to view videos generated with our pipeline in the supplementary. + +# 6. Conclusion + +In this paper, we introduce TAPIR, a novel model for Tracking Any Point (TAP) that adopts a two-stage approach, comprising a matching stage and a refinement stage. It demonstrates a substantial improvement over the state-of-the-art, providing stable predictions and occlusion robustness, and scales to high-resolution videos. We also + +![](images/816c5b75fa56ca10572812a20140fa2670a0ffe37d0150b3d7221821727ac666.jpg) + +![](images/98ae30fc91e6ce0f9c44cd5ccf47bcb8f4024926cba2ad210adf896b46698a26.jpg) + +![](images/ed4d56d95bf060156b51776ed607985448fdf374190505f424b0cf5e46e17b0b.jpg) + +![](images/6e1cef806f85cced0c84db244635233367330bf56b3b7db5140207acf973696b.jpg) + +![](images/55c5c6769ec5a4c23a01480463c00f80990234af931a9a34862891b7ae756672.jpg) +Figure 5. Animating still frames. Top row shows the input images. Each following row shows a visualization of a single sample from our trajectory model: the dark purple end is the starting point of the trajectory, while the colors get brighter and yellower with distance. In the left image, one sample shows the hand moving forward in a cutting motion, with a slight incline of the head; in the other, the entire person moves to the left. On the right, the first sample moves one arm in a circle, while the second shows the head turning and the other arm being raised. + +![](images/ccba319cec58fa54fa79f2a9956135a88e2862db7e713fd33dea9b61f65a1e48.jpg) + +show a proof-of-concept of animating still frames, which we hope can serve as a foundation for future graphics research. TAPIR still has some limitations, such as difficulty in precisely perceiving the boundary between background and foreground, and it may still struggle with large appearance changes. Nevertheless, our findings indicate a promising future for TAP in computer vision. + +Acknowledgements We would like to thank Lucas Smaira and Larisa Markeeva for help with datasets and codebases, and Jon Scholz and Todor Davchev for advice on the needs of robotics. We also thank Adam Harley, Viorica Patraucean and Daniel Zoran for helpful discussions, and Mateusz Malinowski for paper comments. Finally, we thank Nando de Freitas for his support. + +# References + +[1] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5173-5182, 2017. +[2] Marian Stewart Bartlett, Gwen Littlewort, Mark G Frank, Claudia Lainscsek, Ian R Fasel, Javier R Movellan, et al. Automatic recognition of facial actions in spontaneous expressions. J. Multim., 1(6):22-35, 2006. +[3] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In Proceedings of European Conference on Computer Vision (ECCV), pages 404-417. Springer, 2006. +[4] Thomas Brox, Christoph Bregler, and Jitendra Malik. Large displacement optical flow. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 41-48. IEEE, 2009. +[5] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In Proceedings of European Conference on Computer Vision (ECCV), pages 25-36. Springer, 2004. +[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Proceedings of European Conference on Computer Vision (ECCV), pages 611-625. Springer, 2012. +[7] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4733-4742, 2016. +[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6299-6308, 2017. +[9] Zezhou Cheng, Jong-Chyi Su, and Subhransu Maji. On equivariant and invariant learning of object landmark representations. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 9897-9906, 2021. +[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5828-5839, 2017. +[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 224-236, 2018. +[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Proceedings of Neural Information Processing Systems Datasets and Benchmarks Track, 2022. +[13] Carl Doersch and Andrew Zisserman. Sim2real transfer learning for 3d human pose estimation: motion to the rescue. Proceedings of Neural Information Processing Systems (NeurIPS), 32, 2019. + +[14] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 2758-2766, 2015. +[15] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361. IEEE, 2012. +[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: a scalable dataset generator. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. +[17] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10696-10706, 2022. +[18] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Denseseposé: Dense human pose estimation in the wild. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7297-7306, 2018. +[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In Proceedings of European Conference on Computer Vision (ECCV), pages 59-75. Springer, 2022. +[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. +[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. +[22] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185-203, 1981. +[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017. +[24] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Proceedings of Neural Information Processing Systems (NeurIPS), 31, 2018. +[25] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabeled videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8787-8797, 2020. +[26] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia + +Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3192-3199, 2013. +[27] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 6207-6217, 2021. +[28] Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B Choy, Philip HS Torr, and Manmohan Chandraker. Desire: Distant future prediction in dynamic scenes with interacting agents. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 336-345, 2017. +[29] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4521-4530, 2019. +[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2041-2050, 2018. +[31] Ji Lin, Chuang Gan, Kuan Wang, and Song Han. Tsm: Temporal shift module for efficient and scalable video understanding on edge devices. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. +[32] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6728-6737, 2022. +[33] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):978-994, 2010. +[34] David G Lowe. Object recognition from local scale-invariant features. In Proceedings of IEEE International Conference on Computer Vision (ICCV), volume 2, pages 1150-1157. IEEE, 1999. +[35] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. +[36] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In Proceedings of International Joint Conference on Artificial Intelligence (IJCAI), pages 674-679, 1981. +[37] Lucas Manuelli, Yunzhu Li, Pete Florence, and Russ Tedrake. Keypoints into the future: Self-supervised correspondence in model-based reinforcement learning. arXiv preprint arXiv:2009.05085, 2020. +[38] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Mantra: Memory augmented networks for multiple trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7143-7152, 2020. +[39] David Marr and Tomaso Poggio. A computational theory of human stereo vision. Proceedings of the Royal Society of London. Series B. Biological Sciences, 204(1156):301-328, 1979. + +[40] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. +[41] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. +[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. +[43] Yuki Ono, Eduard Trulls, Pascal Fua, and Kwang Moo Yi. Lf-net: Learning local features from images. Proceedings of Neural Information Processing Systems (NeurIPS), 31, 2018. +[44] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. +[45] Deva Ramanan, David A Forsyth, and Andrew Zisserman. Strike a pose: Tracking people by finding stylized poses. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pages 271-278. IEEE, 2005. +[46] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. +[47] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4161-4170, 2017. +[48] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3d reconstruction of interacting hands via collision-aware factorized refinements. In 2021 International Conference on 3D Vision (3DV), pages 432-441. IEEE, 2021. +[49] Michael Rubinstein, Ce Liu, and William Freeman. Towards longer long-range motion trajectories. In Proceedings of British Machine Vision Conference, 2012. +[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. +[51] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International Journal of Computer Vision, 80(1):72-91, 2008. +[52] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016. +[53] Thomas Schops, Johannes L Schonberger, Silvano Galliani, + +Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3260-3269, 2017. +[54] Ishwar K Sethi and Ramesh Jain. Finding trajectories of feature points in a monocular image sequence. IEEE Transactions on Pattern Analysis and Machine Intelligence, (1):56-73, 1987. +[55] Jie Shen, Stefanos Zafeiriou, Grigoris G Chrysos, Jean Kossaifi, Georgios Tzimiropoulos, and Maja Pantic. The first facial landmark tracking in-the-wild challenge: Benchmark and results. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, pages 50-58, 2015. +[56] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2107-2116, 2017. +[57] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. In Proceedings of European Conference on Computer Vision (ECCV), 2004. +[58] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6738-6748, 2022. +[59] Deqing Sun, Daniel Vlasic, Charles Herrmann, Varun Jampani, Michael Krainin, Huiwen Chang, Ramin Zabih, William T Freeman, and Ce Liu. Autoflow: Learning a better training set for optical flow. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10093-10102, 2021. +[60] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8934-8943, 2018. +[61] Yi Sun, Yuheng Chen, Xiaogang Wang, and Xiaou Tang. Deep learning face representation by joint identification-verification. Proceedings of Neural Information Processing Systems (NeurIPS), 27, 2014. +[62] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of European Conference on Computer Vision (ECCV), pages 402-419. Springer, 2020. +[63] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 5916-5925, 2017. +[64] Jonathan Tompson, Murphy Stein, Yann Lecun, and Ken Perlin. Real-time continuous pose recovery of human hands using convolutional networks. ACM Transactions on Graphics (ToG), 33(5):1-10, 2014. +[65] Philip HS Torr and Andrew Zisserman. Feature based methods for structure and motion estimation. In International + +workshop on vision algorithms, pages 278-294. Springer, 1999. +[66] Sudheendra Vijayanarasimhan, Susanna Ricco, Cordelia Schmid, Rahul Sukthankar, and Katerina Fragkiadaki. Sfmnet: Learning of structure and motion from video. arXiv preprint arXiv:1704.07804, 2017. +[67] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of European Conference on Computer Vision (ECCV), pages 601-617, 2018. +[68] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3551-3558, 2013. +[69] Yue Wang and Justin M Solomon. Prnet: Self-supervised learning for partial-to-partial registration. Proceedings of Neural Information Processing Systems (NeurIPS), 32, 2019. +[70] Xuehan Xiong and Fernando De la Torre. Supervised descent method and its applications to face alignment. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 532-539, 2013. +[71] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8121-8130, 2022. +[72] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 10075-10085, 2021. +[73] Zhichao Yin, Trevor Darrell, and Fisher Yu. Hierarchical discrete distribution decomposition for match density estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6044-6053, 2019. +[74] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2694-2703, 2018. +[75] Zhengyou Zhang. A flexible new technique for camera calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(11):1330-1334, 2000. +[76] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. +[77] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2019. +[78] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rehmann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew Fitzgibbon, Charles Loop, Christian Theobalt, et al. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (ToG), 33(4):1-12, 2014. \ No newline at end of file diff --git a/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/images.zip b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..62bc66c83eb067658f1d7aae5f27bc23d811b0cf --- /dev/null +++ b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a767e675eb3d0993dbc17d15c06f87b15ad5271ca3f37f76cd5e7f683a5e8c3 +size 788567 diff --git a/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/layout.json b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f06443da5c6ea0b00552130ee32ae8cf22723a63 --- /dev/null +++ b/2023/TAPIR_ Tracking Any Point with Per-Frame Initialization and Temporal Refinement/layout.json @@ -0,0 +1,9657 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 518, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 518, + 138 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 518, + 138 + ], + "type": "text", + "content": "TAPIR: Tracking Any Point with per-frame Initialization and temporal Refinement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 160, + 145, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 160, + 145, + 175 + ], + "spans": [ + { + "bbox": [ + 70, + 160, + 145, + 175 + ], + "type": "text", + "content": "Carl Doersch*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 170, + 161, + 217, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 161, + 217, + 175 + ], + "spans": [ + { + "bbox": [ + 170, + 161, + 217, + 175 + ], + "type": "text", + "content": "Yi Yang*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 240, + 161, + 315, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 161, + 315, + 175 + ], + "spans": [ + { + "bbox": [ + 240, + 161, + 315, + 175 + ], + "type": "text", + "content": "Mel Vecerik*†" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 339, + 161, + 416, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 161, + 416, + 175 + ], + "spans": [ + { + "bbox": [ + 339, + 161, + 416, + 175 + ], + "type": "text", + "content": "Dilara Gokay*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 440, + 161, + 522, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 440, + 161, + 522, + 175 + ], + "spans": [ + { + "bbox": [ + 440, + 161, + 522, + 175 + ], + "type": "text", + "content": "Ankush Gupta*" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 144, + 175, + 212, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 175, + 212, + 190 + ], + "spans": [ + { + "bbox": [ + 144, + 175, + 212, + 190 + ], + "type": "text", + "content": "Yusuf Aytar*" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 237, + 175, + 315, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 175, + 315, + 189 + ], + "spans": [ + { + "bbox": [ + 237, + 175, + 315, + 189 + ], + "type": "text", + "content": "Joao Carreira*" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 175, + 449, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 175, + 449, + 189 + ], + "spans": [ + { + "bbox": [ + 340, + 175, + 449, + 189 + ], + "type": "text", + "content": "Andrew Zisserman*†" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 166, + 204, + 263, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 204, + 263, + 218 + ], + "spans": [ + { + "bbox": [ + 166, + 204, + 263, + 218 + ], + "type": "text", + "content": "*Google DeepMind" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 287, + 204, + 427, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 204, + 427, + 218 + ], + "spans": [ + { + "bbox": [ + 287, + 204, + 427, + 218 + ], + "type": "text", + "content": "† University College London" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 218, + 455, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 218, + 455, + 233 + ], + "spans": [ + { + "bbox": [ + 138, + 218, + 455, + 233 + ], + "type": "inline_equation", + "content": "^{\\ddagger}" + }, + { + "bbox": [ + 138, + 218, + 455, + 233 + ], + "type": "text", + "content": "VGG, Department of Engineering Science, University of Oxford" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 260, + 192, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 260, + 192, + 273 + ], + "spans": [ + { + "bbox": [ + 143, + 260, + 192, + 273 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 285, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 285, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 45, + 285, + 290, + 514 + ], + "type": "text", + "content": "We present a novel model for Tracking Any Point (TAP) that effectively tracks any queried point on any physical surface throughout a video sequence. Our approach employs two stages: (1) a matching stage, which independently locates a suitable candidate point match for the query point on every other frame, and (2) a refinement stage, which updates both the trajectory and query features based on local correlations. The resulting model surpasses all baseline methods by a significant margin on the TAP-Vid benchmark, as demonstrated by an approximate " + }, + { + "bbox": [ + 45, + 285, + 290, + 514 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 45, + 285, + 290, + 514 + ], + "type": "text", + "content": " absolute average Jaccard (AJ) improvement on DAVIS. Our model facilitates fast inference on long and high-resolution video sequences. On a modern GPU, our implementation has the capacity to track points faster than real-time. Given the high-quality trajectories extracted from a large dataset, we demonstrate a proof-of-concept diffusion model which generates trajectories from static images, enabling plausible animations. Visualizations, source code, and pretrained models can be found at https://deepmind-tapir.github.io." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 557, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 688 + ], + "type": "text", + "content": "The problem of point level correspondence —i.e., determining whether two pixels in two different images are projections of the same point on the same physical surface—has long been a fundamental challenge in computer vision, with enormous potential for providing insights about physical properties and 3D shape. We consider its formulation as “Tracking Any Point” (TAP) [12]: given a video and (potentially dense) query points on solid surfaces, an algorithm should reliably output the locations those points correspond to in every other frame where they are visible, and indicate frames where they are not – see Fig. 4 for illustration." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Our main contribution is a new model: TAP with per-frame Initialization and temporal Refinement (TAPIR)," + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 310, + 262, + 544, + 411 + ], + "blocks": [ + { + "bbox": [ + 310, + 262, + 544, + 411 + ], + "lines": [ + { + "bbox": [ + 310, + 262, + 544, + 411 + ], + "spans": [ + { + "bbox": [ + 310, + 262, + 544, + 411 + ], + "type": "image", + "image_path": "90016a7bcb80eadb6771cb4c9979ff7cdfcdad16777a4611e5689c693d102ff2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 426, + 547, + 492 + ], + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 492 + ], + "type": "text", + "content": "Figure 1. Retrospective evolution of point tracking performance over time on the recent TAP-Vid-Kinetics and TAP-Vid-DAVIS benchmarks, as measured by Average Jaccard (higher is better). In this paper we introduce TAPIR, which significantly improves performance over the state-of-the-art. This unlocks new capabilities, which we demonstrate on motion-based future prediction." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 518, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 674 + ], + "type": "text", + "content": "which greatly improves over the state-of-the-art on the recently-proposed TAP-Vid benchmark [12]. There are many challenges to TAP: we must robustly estimate occlusions and recover when points reappear (unlike optical flow and structure-from-motion keypoints), meaning that search must be incorporated; yet when points remain visible for many consecutive frames, it is important to integrate information about appearance and motion across many of those frames in order to optimally predict positions. Furthermore, little real-world ground truth is available to learn from, so supervised-learning algorithms need to learn from synthetic data without overfitting to the data distribution (i.e., sim2real)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "There are three core design decisions that define TAPIR. The first is to use a coarse-to-fine approach. This approach has been used across many high-precision estimation prob" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10061" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 323 + ], + "type": "text", + "content": "lems in computer vision [7, 28, 32, 38, 41, 58, 69, 73, 78]. For our application, the initial 'coarse' tracking consists of an occlusion-robust matching performed separately on every frame, where tracks are hypothesized using low-resolution features, without enforcing temporal continuity. The 'fine' refinement iteratively uses local, spatio-temporal information at a higher resolution, wherein a neural network can trade-off smoothness of motion with local appearance cues to produce the most likely track. The second design decision is to be fully-convolutional in time: the layers of our neural network consist principally of feature comparisons, spatial convolutions, and temporal convolutions, resulting in a model which efficiently maps onto modern GPU and TPU hardware. The third design decision is that the model should estimate its own uncertainty with regard to its position estimate, in a self-supervised manner. This ensures that low-confidence predictions can be suppressed, which improves benchmark scores. We hypothesize that this may help downstream algorithms (e.g. structure-from-motion) that rely on precision, and can benefit when low-quality matches are removed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 326, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 326, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 49, + 326, + 286, + 529 + ], + "type": "text", + "content": "We find that two existing architectures already have some of the pieces we need: TAP-Net [12] and Persistent Independent Particles (PIPs) [19]. Therefore, a key contribution of our work is to effectively combine them while achieving the benefits from both. TAP-Net performs a global search on every frame independently, providing a coarse track that is robust to occlusions. However, it does not make use of the continuity of videos, resulting in jittery, unrealistic tracks. PIPs, meanwhile, gives a recipe for refinement: given an initialization, it searches over a local neighborhood and smooths the track over time. However, PIPs processes videos sequentially in chunks, initializing each chunk with the output from the last. The procedure struggles with occlusion and is difficult to parallelize, resulting in slow processing (i.e., 1 month to evaluate TAP-Vid-Kinetics on 1 GPU). A key contribution of this work is observing the complementarity of these two methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 531, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 531, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 49, + 531, + 286, + 651 + ], + "type": "text", + "content": "As shown in Fig. 1, we find that TAPIR improves over prior works by a large margin, as measured by performance on the TAP-Vid benchmark [12]. On TAP-Vid-DAVIS, TAPIR outperforms TAP-Net by " + }, + { + "bbox": [ + 49, + 531, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\sim 20\\%" + }, + { + "bbox": [ + 49, + 531, + 286, + 651 + ], + "type": "text", + "content": " while on the more challenging TAP-Vid-Kinetics, TAPIR outperforms PIPs by " + }, + { + "bbox": [ + 49, + 531, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\sim 20\\%" + }, + { + "bbox": [ + 49, + 531, + 286, + 651 + ], + "type": "text", + "content": ", and substantially reduces its inference runtime. To demonstrate the quality of TAPIR trajectories, we showcase a proof-of-concept model trained to generate trajectories given individual images, and find that this model can generate plausible animations from single photographs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 653, + 286, + 712 + ], + "type": "text", + "content": "In summary, our contributions are as follows: 1) We propose a new model for long term point tracking, bridging the gap between TAP-Net and PIPs. 2) We show that the model achieves state-of-the-art results on the challenging TAP-Vid benchmark, with a significant boost on per" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 72, + 544, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 179 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 179 + ], + "type": "text", + "content": "formance. 3) We provide an extensive analysis of the architectural decisions that matter for high-performance point tracking. 4) We provide a proof-of-concept of video prediction enabled by TAPIR's high-quality trajectories. Finally, 5) after analyzing components, we separately perform careful tuning of hyperparameters across entire method, in order to develop the best-performing model, which we release at https://www.github.com/deepmind/tapnet for the benefit of the community." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 195, + 391, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 391, + 208 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 391, + 208 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 217, + 544, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 544, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 544, + 240 + ], + "type": "text", + "content": "Physical point correspondence and tracking have been studied in a wide range of settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 242, + 544, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 242, + 544, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 242, + 544, + 338 + ], + "type": "text", + "content": "Optical Flow focuses on dense motion estimation between image pairs [22, 36]; methods can be divided into classical variational approaches [4, 5] and deep learning approaches [14, 23, 47, 60, 62, 71]. Optical flow operates only on subsequent frames, with no simple method to track across long videos. Groundtruth is hard to obtain [6], so it is typically benchmarked on synthetic scenes [6, 14, 40, 59], though limited real data exists through depth scanners [15]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 339, + 544, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 339, + 544, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 544, + 495 + ], + "type": "text", + "content": "Keypoint Correspondence methods aim at sparse keypoint matching given image pairs, from hand-defined discriminative feature descriptors [3, 34, 35, 54] to deep learning based methods [11, 27, 37, 43]. Though it operates on image pairs, it is arguably \"long term\" as the images may be taken at different times from wide baselines. The goal, however, is typically not to track any point, but to find easily-trackable \"keypoints\" sufficient for reconstruction. They are mainly used in structure-from-motion (SfM) settings and typically ignore occlusion, as SfM can use geometry to filter errors [20, 52, 65, 66]. This line of work is thus often restricted to rigid scenes [39, 75], and is benchmarked accordingly [1, 10, 29, 30, 53, 76]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 496, + 544, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 496, + 544, + 628 + ], + "spans": [ + { + "bbox": [ + 308, + 496, + 544, + 628 + ], + "type": "text", + "content": "Semantic Keypoint Tracking is often represented as keypoints or landmarks [26, 45, 56, 61, 64, 70]. Landmarks are often \"joints\", which can have vastly different appearance depending on pose (i.e. viewed from the front or back); thus most methods rely on large supervised datasets [2, 9, 48, 55, 72, 77], although some works track surface meshes [18, 67]. One interesting exception uses sim2real based on motion [13], motivating better surface tracking. Finally, some recent methods discover keypoints on moving objects [24, 25, 63, 74], though these typically require in-domain training on large datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 630, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 630, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 630, + 544, + 712 + ], + "type": "text", + "content": "Long-term physical point tracking is addressed in a few early works [33, 49, 51, 57, 68], though such hand-engineered methods have fallen out of favor in the deep learning era. Our work instead builds on two recent works, TAP-Net [12] and Persistent Independent Particles (PIPs) [19], which aim to update these prior works to the deep learning era." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 317, + 757 + ], + "type": "text", + "content": "10062" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 353 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 353 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 353 + ], + "type": "image", + "image_path": "902c79206953071b29eda02102ddbd7a9487ae746f178a91972e615271fa6a5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 360, + 289, + 482 + ], + "lines": [ + { + "bbox": [ + 46, + 360, + 289, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 289, + 482 + ], + "type": "text", + "content": "Figure 2. TAPIR architecture summary. Our model begins with a global comparison between the query point features and the features for every other frame to compute an initial track estimate, including an uncertainty estimate. Then, we extract features from a local neighborhood (shown in pink) around the initial estimate, and compare these to the query feature at a higher resolution, post-processing the similarities with a temporal depthwise-convolitional network to get an updated position estimate. This updated position is fed back into the next iteration of refinement, repeated for a fixed number of iterations. Note that for simplicity, multi-scale pyramids are not shown." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 499, + 135, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 135, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 135, + 511 + ], + "type": "text", + "content": "3. TAPIR Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": "Given a video and a query point, our goal is to estimate the 2D location " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " that it corresponds to in every other frame " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": ", as well as a 1D probability " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " that the point is occluded and a 1D probability " + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 46, + 521, + 287, + 688 + ], + "type": "text", + "content": " on the uncertainty of the estimated location. To be robust to occlusion, we first match candidate locations in other frames, by comparing the query features with all other features, and post-processing the similarities to arrive at an initial estimate. Then we refine the estimate, based on the local similarities between the query point and the target point. Note that both stages depend mostly on similarities (dot products) between the query point features and features elsewhere (i.e. not on the feature alone); this ensures that we don't overfit to specific features that might only appear in the (synthetic) training data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Fig. 2 gives an overview of our model. We initialize an estimate of the track by finding the best match within each" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 547, + 166 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 547, + 166 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 547, + 166 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 547, + 166 + ], + "type": "table", + "html": "
TAP-Net [12]PIPs [19]TAPIR
Per-frame Initialization✓✗✓
Fully Convolutional In Time✓✗✓
Temporal Refinement✗✓✓
Multi Scale Feature✗✓✓
Stride-4 Feature✗test-time only✓
Parallel Inference✓✗✓
Uncertainty Estimate✗✗✓
Number of Parameters2.8M28.7M29.3M
", + "image_path": "6dc395be74df4347fbd72725f922e52811f216f82f83d58b1667d713c4f55f3e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 167, + 545, + 199 + ], + "lines": [ + { + "bbox": [ + 305, + 167, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 545, + 199 + ], + "type": "text", + "content": "Table 1. Overview of models. TAPIR combines the merits from both TAP-Net and PIPs and further adds a handful of crucial components to improve performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 209, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 437 + ], + "type": "text", + "content": "frame independently, ignoring the temporal nature of the video. This involves a cost volume: for each frame " + }, + { + "bbox": [ + 304, + 209, + 545, + 437 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 209, + 545, + 437 + ], + "type": "text", + "content": ", we compute the dot product between the query feature " + }, + { + "bbox": [ + 304, + 209, + 545, + 437 + ], + "type": "inline_equation", + "content": "F_{q}" + }, + { + "bbox": [ + 304, + 209, + 545, + 437 + ], + "type": "text", + "content": " and all features in the " + }, + { + "bbox": [ + 304, + 209, + 545, + 437 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 209, + 545, + 437 + ], + "type": "text", + "content": "'th frame. We post-process the cost volume with a ConvNet, which produces a spatial heatmap, which is then summarized to a point estimate, in a manner broadly similar to TAP-Net. Given the initialization, we then compare the query feature to all features in a small region around an initial track. We feed the comparisons into a neural network, which updates both the query features and the track. We iteratively apply multiple such updates to arrive at a final trajectory estimate. TAPIR accomplishes this with a fully-convolutional architecture, operating on a pyramid of both high-dimensional, low-resolution features and low-dimensional, high-resolution features, in order to maximize efficiency on modern hardware. Finally, we add an estimate of uncertainty with respect to position throughout the architecture in order to suppress low-accuracy predictions. The rest of this section describes TAPIR in detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 444, + 420, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 420, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 420, + 455 + ], + "type": "text", + "content": "3.1. Track Initialization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "content": "The initial cost volume is computed using a relatively coarse feature map " + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{T \\times H/8 \\times W/8 \\times C}" + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "content": " is the number of frames, " + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "content": " are the image height and width, and " + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "content": " is the number of channels, computed with a standard TSM-ResNet-18 [31] backbone. Features for the query point " + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "inline_equation", + "content": "F_{q}" + }, + { + "bbox": [ + 304, + 462, + 545, + 557 + ], + "type": "text", + "content": " are extracted via bilinear interpolation at the query location, and we perform a dot product between this query feature and all other features in the video." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": "We compute an initial estimate of the position " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "p_t^0 = (x_t^0, y_t^0)" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": " and occlusion " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "o_t^0" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": " by applying a small ConvNet to the cost volume corresponding to frame " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": " (which is of shape " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "H / 8 \\times W / 8 \\times 1" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": " per query). This ConvNet has two outputs: a heatmap of shape " + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "H / 8 \\times W / 8 \\times 1" + }, + { + "bbox": [ + 305, + 558, + 546, + 713 + ], + "type": "text", + "content": " for predicting the position, and a single scalar logit for the occlusion estimate, obtained via average pooling followed by projection. The heatmap is converted to a position estimate by a \"spatial soft argmax\", i.e., a softmax across space (to make the heatmap positive and sum to 1). Afterwards, all heatmap values which are too far from the argmax position in this heatmap are set to zero. Then, the positions for each cell of the heatmap are averaged spatially, weighted by the thresholded heatmap" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10063" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": "magnitudes. Thus, the output is typically a location close to the heatmap's maximum; the \"soft argmax\" is differentiable, but the thresholding suppresses spurious matches, and prevents the network from \"averaging\" between multiple matches. This can serve as a good initialization, although the model struggles to localize points to less than a few pixels' accuracy on the original resolution due to the 8-pixel stride of the heatmap." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 188, + 289, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 188, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 289, + 439 + ], + "type": "text", + "content": "Position Uncertainty Estimates. A drawback of predicting occlusion and position independently from the cost volume (a choice that is inspired by TAP-Net) is that if a point is visible in the ground truth, it's worse if the algorithm predicts a vastly incorrect location than if it simply incorrectly marks the point as occluded. After all, downstream applications may want to use the track to understand object motion; such downstream pipelines must be robust to occlusion, but may assume that the tracks are otherwise correct. The Average Jaccard metric reflects this: predictions in the wrong location are counted as both a \"false positive\" and a \"false negative.\" This kind of error tends to happen when the algorithm is uncertain about the position, e.g., if there are many potential matches. Therefore, we find it's beneficial for the algorithm to also estimate its own certainty regarding position. We quantify this by making the occlusion pathway output a second logit, estimating the probability " + }, + { + "bbox": [ + 46, + 188, + 289, + 439 + ], + "type": "inline_equation", + "content": "u_{t}^{0}" + }, + { + "bbox": [ + 46, + 188, + 289, + 439 + ], + "type": "text", + "content": " that the prediction is likely to be far enough from the ground truth that it isn't useful, even if the model predicts that it's visible. We define \"useful\" as whether the prediction is within a threshold " + }, + { + "bbox": [ + 46, + 188, + 289, + 439 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 188, + 289, + 439 + ], + "type": "text", + "content": " of the ground truth." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 440, + 288, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 440, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 288, + 464 + ], + "type": "text", + "content": "Therefore, our loss for the initialization for a given video frame " + }, + { + "bbox": [ + 47, + 440, + 288, + 464 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 440, + 288, + 464 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 47, + 440, + 288, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(p_t^0,o_t^0,u_t^0)" + }, + { + "bbox": [ + 47, + 440, + 288, + 464 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 440, + 288, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 47, + 440, + 288, + 464 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 486, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 486, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 83, + 486, + 287, + 529 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} \\left(p _ {t}, o _ {t}, u _ {t}\\right) = \\operatorname {H u b e r} \\left(\\hat {p} _ {t}, p _ {t}\\right) * \\left(1 - \\hat {o} _ {t}\\right) \\\\ + \\operatorname {B C E} \\left(\\hat {o} _ {t}, o _ {t}\\right) \\\\ + \\operatorname {B C E} \\left(\\hat {u} _ {t}, u _ {t}\\right) * \\left(1 - \\hat {o} _ {t}\\right) \\tag {1} \\\\ \\end{array}", + "image_path": "d6f2ddc4cc6f5b8bd76328c04d89c2e9d35b2d8d2b890202a77f40f732745d79.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 531, + 232, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 531, + 232, + 557 + ], + "spans": [ + { + "bbox": [ + 87, + 531, + 232, + 557 + ], + "type": "interline_equation", + "content": "\\text {w h e r e ,} \\quad \\hat {u} _ {t} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} d (\\hat {p} _ {t}, p _ {t}) > \\delta \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.", + "image_path": "d346c8e85519a1528bc023fccecec68e77ebe3e68a684e10982d51daeaf5e893.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\hat{o}_t \\in \\{0,1\\}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\hat{p} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " are the ground truth occlusion and point locations respectively, " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " is Euclidean distance, " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " is the distance threshold, Huber is a Huber loss, and BCE is binary cross entropy (both " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " have sigmoids applied to convert them to probabilities). " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\hat{u}_t" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ", the target for the uncertainty estimate " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ", is computed from the ground truth position and the network's prediction: it's 0 if the model's position prediction is close enough to the ground truth (within the threshold " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "), and 1 otherwise. At test time, the algorithm should output that the point is visible if it's both predicted as unoccluded and if the model is confident in the prediction. We therefore do a soft combination of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "the two probabilities: the algorithm outputs that the point is visible if " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "(1 - u_t) * (1 - o_t) > 0.5" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 106, + 425, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 106, + 425, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 106, + 425, + 118 + ], + "type": "text", + "content": "3.2. Iterative Refinement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": "Given an estimated position, occlusion, and uncertainty for each frame, the goal of each iteration " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": " of our refinement procedure is to compute an update " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "(\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i})" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": ", which adjusts the estimate to be closer to the ground truth, integrating information across time. The update is based on a set of \"local score maps\" which capture the query point similarity (i.e. dot products) to the features in the neighborhood of the trajectory. These are computed on a pyramid of different resolutions, so for a given trajectory, they have shape " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "(H' \\times W' \\times L)" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "H' = W' = 7" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": ", the size of the local neighborhood, and " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": " is the number of levels of the spatial pyramid (different pyramid levels are computed by spatially pooling the feature volume " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": "). As with the initialization, this set of similarities is post-processed with a network to predict the refined position, occlusion, and uncertainty estimate. Unlike the initialization, however, we include \"local score maps\" for many frames simultaneously as input to the post-processing. We include the current position estimate, the raw query features, and the (flattened) local score maps into a tensor of shape " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "T \\times (C + K + 4)" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": " is the number of channels in the query feature, " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "K = H' \\cdot W' \\cdot L" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": " is the number of values in the flattened local score map, and 4 extra dimensions for position, occlusion, and uncertainty. The output of this network at the " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": "'th iteration is a residual " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "(\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i}, \\Delta F_{q,t,i})" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": ", which is added to the position, occlusion, uncertainty estimate, and the query feature, respectively. " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\Delta F_{q,t,i}" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": " is of shape " + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "inline_equation", + "content": "T \\times C" + }, + { + "bbox": [ + 304, + 125, + 545, + 472 + ], + "type": "text", + "content": "; thus, after the first iteration, slightly different \"query features\" are used on each frame when computing new local score maps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 474, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 545, + 617 + ], + "type": "text", + "content": "These positions and score maps are fed into a 12-block convolutional network to compute an update " + }, + { + "bbox": [ + 305, + 474, + 545, + 617 + ], + "type": "inline_equation", + "content": "(\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i}, \\Delta F_{q,t,i})" + }, + { + "bbox": [ + 305, + 474, + 545, + 617 + ], + "type": "text", + "content": ", where each block consists of a " + }, + { + "bbox": [ + 305, + 474, + 545, + 617 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 305, + 474, + 545, + 617 + ], + "type": "text", + "content": " convolution block and a depthwise convolution block. This architecture is inspired by the MLP-Mixer used for refinement in PIPs: we directly translate the Mixer's cross-channel layers into " + }, + { + "bbox": [ + 305, + 474, + 545, + 617 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 305, + 474, + 545, + 617 + ], + "type": "text", + "content": " convolutions with the same number of channels, and the within-channel operations become similarly within-channel depthwise convolutions. Unlike PIPs, which breaks sequences into 8-frame chunks before running the MLP-Mixer, this convolutional architecture can be run on sequences of any length." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "We found that the feature maps used for computing the \"score maps\" are important for achieving high precision. The pyramid levels " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "L - 1" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " are computed by spatially average-pooling the raw TSM-ResNet features " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " at a stride of " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "8 \\cdot 2^{l - 1}" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": ", and then computing all dot products between the query feature " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "F_{q}" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " and the pyramid features. For the " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "'th frame, we extract a " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " patch of dot products centered at " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": ". At training time, PIPs leaves it here, but at" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10064" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "text", + "content": "test time alters the backbone to run at stride 4, introducing a train/test domain gap. We find it is effective to train on stride 4 as well, although this puts pressure on system memory. To alleviate this, we compute a 0-th score map on the 64-channel, stride-4 input convolution map of the TSM-ResNet, and extract a comparable feature for the query from this feature map via bilinear interpolation. Thus, the final set of local score maps has shape " + }, + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "inline_equation", + "content": "(7 \\cdot 7 \\cdot L)" + }, + { + "bbox": [ + 46, + 72, + 286, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "type": "text", + "content": "After " + }, + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "type": "inline_equation", + "content": "(\\Delta p_{t}^{i}, \\Delta o_{t}^{i}, \\Delta u_{t}^{i}, \\Delta F_{q,t,i})" + }, + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "type": "text", + "content": " is obtained from the above architecture, we can iterate the refinement step as many times as desired, re-using the same network parameters. At each iteration, we use the same loss " + }, + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 168, + 286, + 228 + ], + "type": "text", + "content": " on the output, weighting each iteration the same as the initialization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 229, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 287, + 577 + ], + "type": "text", + "content": "Discussion As there is some overlap with prior work, Figure 3 depicts the relationship between TAPIR and other architectures. For the sake of ablations, we also develop a \"simple combination\" model, which is intended to be a simpler combination of a TAP-Net-style initialization with PIPs-style refinements, and follows prior architectural decisions regardless of whether they are optimal. To summarize, the major architectural decisions needed to create the \"simple combination\" of TAP-Net and PIPs (and therefore the contributions of this model) are as follows. First we exploit the complementarity between TAP-Net and PIPs. Second, we remove \"chaining\" from PIPs, replacing that initialization with TAP-Net's initialization. We directly apply PIPs' MLP-Mixer architecture by 'chunking' the input sequence. Note that we adopt TAP-Net's feature network for both initialization and refinement. Finally, we adapt the TAP-Net backbone to extract a multi-scale and high resolution pyramid (follow PIPs stride-4 test-time architecture, but matching TAP-Net's end-to-end training). We apply losses at all layers of the hierarchy (whereas TAP-Net has a single loss, PIPs has no loss on its initialization). The additional contributions of TAPIR are as follows: first, we entirely replace the PIPs MLP-Mixer (which must be applied to fixed-length sequences) with a novel depthwise-convolitional architecture, which has a similar number of parameters but works on sequences of any length. Second, we introduce uncertainty estimate, which are computed at every level of the hierarchy. These are 'self-supervised' in the sense that the targets depend on the model's own output." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 586, + 149, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 149, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 149, + 600 + ], + "type": "text", + "content": "3.3. Training Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Although TAPIR is designed to be robust to the sim2real gap, we can improve performance by minimizing the gap itself. One subtle difference between the Kubric MOViE [16] dataset and the real world is the lack of panning: although the Kubric camera moves, it is always set to \"look at\" a single point at the center of the workspace. Thus, stationary parts of the scene rotate around the center point of the workspace on the 2D image plane. Models thus tend to fail on real-world videos with panning. Therefore, we mod" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "ified the MOVi-E dataset so that the \"look at\" point moves along a random linear trajectory. The start and end points of this trajectory are enforced to be less than 1.5 units off the ground and lie within a radius of 4 units from the workspace center; the trajectory must also pass within 1 unit of the workspace center at some point, to ensure that the camera is usually looking toward the objects in the scene. See Appendix 5.2 for details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 180, + 388, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 388, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 388, + 194 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 201, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 416 + ], + "type": "text", + "content": "We evaluate on the TAP-Vid [12] benchmark, a recent large-scale benchmark evaluating the problem of interest. TAP-Vid consists of point track annotations on four different datasets, each with different challenges. DAVIS [44] is a benchmark designed for tracking, and includes complex motion and large changes in object scale. Kinetics [8] contains over 1000 labeled videos, and contains all the complexities of YouTube videos, including difficulties like cuts and camera shake. RGB Stacking is a synthetic dataset of robotics videos with precise point tracks on large textureless regions. Kubric MOVi-E, used for training, contains realistic renderings of objects falling onto a flat surface; it comes with a validation set that we use for comparisons. We conduct training exclusively on the Kubric dataset, and selected our best model primarily by observing DAVIS. We performed no automated evaluation for hyperparameter tuning or model selection on any dataset. We train and evaluate at a resolution of " + }, + { + "bbox": [ + 304, + 201, + 545, + 416 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 201, + 545, + 416 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "text", + "content": "Implementation Details We train for 50,000 steps with a cosine learning rate schedule, with a peak rate of " + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "text", + "content": " and 1,000 warmup steps. We use 64 TPU-v3 cores, using 4 24-frame videos per core in each step. We use an Adam-W optimizer with " + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.95" + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "text", + "content": " and weight decay " + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-2}" + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "text", + "content": ". For most experiments, we use cross-replica batch normalization within the (TSM-)ResNet backbone only, although for our public model we find instance normalization is actually more effective. For most experiments we use " + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "L = 5" + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "text", + "content": " to match PIPs, but in practice we find " + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "L > 3" + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "text", + "content": " has little impact (See Appendix 4.3). Therefore, to save computation, our public model uses " + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 304, + 434, + 545, + 589 + ], + "type": "text", + "content": ". See Appendix 5.1 for more details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 598, + 448, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 598, + 448, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 598, + 448, + 612 + ], + "type": "text", + "content": "4.1.Quantitative Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": "Table 2 shows the performance of TAPIR relative to prior works. TAPIR improves by " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " absolute performance over the best model on Kinetics (TAP-Net) and by " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "19.3\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " on DAVIS (PIPs). These improvements are quite significant according to the TAP-Vid metrics: Average Jaccard consists of 5 thresholds, so under perfect occlusion estimation, a " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " improvement requires improving all points to the next distance threshold (half the distance). It's worth noting" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10065" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 211, + 193 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 211, + 193 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 211, + 193 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 211, + 193 + ], + "type": "image", + "image_path": "e4b5154c9ea6e729709200e681129490f2cea7954f486846da0cfa18e43b511f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 197, + 211, + 302 + ], + "blocks": [ + { + "bbox": [ + 49, + 197, + 211, + 302 + ], + "lines": [ + { + "bbox": [ + 49, + 197, + 211, + 302 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 211, + 302 + ], + "type": "image", + "image_path": "6656db070a993c553263d72a54561b74ca215fbbce060ef0b2f849b0166b572b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 311, + 546, + 355 + ], + "lines": [ + { + "bbox": [ + 46, + 311, + 546, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 546, + 355 + ], + "type": "text", + "content": "Figure 3. Comparison of Architectures. Left: the initial TAP-Net and PIPs models: TAP-Net has an independent estimate per frame, whereas PIPs breaks the video into fixed-size chunks and processes chunks sequentially. Middle: a \"simple combination\" of these two architectures, where TAP-Net is used to initialize PIPs-style chunked iterations. Right: Our model, TAPIR, which removes the chunking and adds uncertainty estimate. Note that for simplicity, multi-scale pyramids are not shown." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 212, + 70, + 375, + 302 + ], + "blocks": [ + { + "bbox": [ + 212, + 70, + 375, + 302 + ], + "lines": [ + { + "bbox": [ + 212, + 70, + 375, + 302 + ], + "spans": [ + { + "bbox": [ + 212, + 70, + 375, + 302 + ], + "type": "image", + "image_path": "f43c633f7edbb99052e2883b2306a9ef81cbd330d986f3e608f3e47939be1122.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 379, + 71, + 542, + 302 + ], + "blocks": [ + { + "bbox": [ + 379, + 71, + 542, + 302 + ], + "lines": [ + { + "bbox": [ + 379, + 71, + 542, + 302 + ], + "spans": [ + { + "bbox": [ + 379, + 71, + 542, + 302 + ], + "type": "image", + "image_path": "05a05cb07531b542d157902b8e83a28bd058c3647010f2bd4478d233841f48bf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 363, + 544, + 480 + ], + "blocks": [ + { + "bbox": [ + 50, + 363, + 544, + 480 + ], + "lines": [ + { + "bbox": [ + 50, + 363, + 544, + 480 + ], + "spans": [ + { + "bbox": [ + 50, + 363, + 544, + 480 + ], + "type": "table", + "html": "
MethodKineticsKubricDAVISRGB-Stacking
AJ<δxavgOAAJ<δxavgOAAJ<δxavgOAAJ<δxavgOA
COTR [27]19.038.857.440.160.778.5535.451.380.26.813.579.1
Kubric-VFS-Like [16]40.559.080.051.969.884.633.148.579.457.972.691.9
RAFT [62]34.552.579.741.258.286.430.046.379.644.058.690.4
PIPs [19]35.354.877.459.174.888.642.059.482.137.351.091.6
TAP-Net [12]46.660.985.065.477.793.038.453.182.359.972.890.4
TAPIR (MOVi-E)57.170.087.684.391.895.859.872.387.666.277.493.3
TAPIR (Panning MOVi-E)57.270.187.884.792.195.861.373.688.862.774.691.6
", + "image_path": "fa04f074357fbf5ac5f3e77438d0d06cf116fc6cef83a83a2af6b9cbe88f586e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 481, + 546, + 514 + ], + "lines": [ + { + "bbox": [ + 46, + 481, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 546, + 514 + ], + "type": "text", + "content": "Table 2. Comparison of TAPIR to prior results on TAP-Vid. " + }, + { + "bbox": [ + 46, + 481, + 546, + 514 + ], + "type": "inline_equation", + "content": "< \\delta_{avg}^{x}" + }, + { + "bbox": [ + 46, + 481, + 546, + 514 + ], + "type": "text", + "content": " is the fraction of points unoccluded in the ground truth, where the prediction is within a threshold, ignoring the occlusion estimate; OA is the accuracy of predicting occlusion. AJ is Average Jaccard, which considers both position and occlusion accuracy." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 529, + 289, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 289, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 289, + 709 + ], + "type": "text", + "content": "that, like PIPs, TAPIR's performance on DAVIS is slightly higher than on Kinetics, whereas for TAP-Net, Kinetics has higher performance, suggesting that TAPIR's reliance on temporal continuity isn't as useful on Kinetics videos with cuts or camera shake. Note that the bulk of the improvement comes from the improved model rather than the improved dataset (TAPIR (MOVi-E) is trained on the same data as TAP-Net, while 'Panning' is the new dataset). The dataset's contribution is small, mostly improving DAVIS, but we found qualitatively that it makes the most substantial difference on backgrounds when cameras pan. This is not well represented in TAP-Vid's query points, but may matter for real-world applications like stabilization. RGB-stacking, however, is somewhat of an outlier: training on MOVi-E is still optimal (unsurprising as the cameras are" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "text", + "content": "static), and there the improvement is only " + }, + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "text", + "content": ", possibly because TAPIR does less to help with textureless objects, indicating an area for future research. Fig. 4 shows some example predictions on DAVIS. TAPIR corrects many types of errors, including problems with temporal coherence from TAP-Net and problems with occlusion from PIPs. We include further qualitative results on our project webpage, which illustrate how noticeable a " + }, + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 529, + 545, + 625 + ], + "type": "text", + "content": " improvement is." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": "TAP-Vid proposes another evaluation for videos at the resolution of " + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 629, + 546, + 715 + ], + "type": "text", + "content": " called 'query-first.' In the typical case (Table 2), the query points are sampled in a 'strided' fashion from the ground-truth tracks: the same trajectory is therefore queried with multiple query points, once every 5 frames. In the 'query first' evaluation, only the first point along each trajectory is used as a query. These points" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10066" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 286, + 117 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 286, + 117 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 286, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 286, + 117 + ], + "type": "table", + "html": "
MethodKineticsOADAVISOARGB-Stacking
AJ<δxavgAJ<δxavgAJ<δxavgOA
TAP-Net38.554.480.633.048.678.853.568.186.3
TAP-Net (Panning MOVi-E)39.556.281.436.052.980.150.165.788.3
TAPIR (MOVi-E)49.664.285.255.369.484.456.270.089.3
TAPIR (Panning MOVi-E)49.664.285.056.270.086.555.569.788.0
", + "image_path": "f909fe9c1e174e514be7f2f8b156ce300e1493268b4d5a6dae45176e10951e02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 118, + 286, + 171 + ], + "lines": [ + { + "bbox": [ + 47, + 118, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 286, + 171 + ], + "type": "text", + "content": "Table 3. Comparison under query first metrics. We see largely the same relative performance trend whether the model is queried in a 'strided' fashion or whether the model is queried with the first frame where the point appeared, although performance is overall lower than the 'strided' evaluation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 182, + 286, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 286, + 289 + ], + "type": "text", + "content": "are slightly harder to track than those in the middle of the video, because there are more intervening frames wherein the point's appearance may change. Table 3 shows our results on this approach. We see that the performance broadly follows the results from the 'strided' evaluation: we outperform TAP-Net by a wide margin on real data (10% absolute on Kinetics, 20% on DAVIS), and by a smaller margin on RGB-Stacking (3%). This suggests that TAPIR remains robust even if the output frames are far from the query point." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 290, + 286, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 290, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 290, + 286, + 434 + ], + "type": "text", + "content": "We also conduct a comparison of computational time for model inference on the DAVIS video horsejump-high with a resolution of " + }, + { + "bbox": [ + 47, + 290, + 286, + 434 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 47, + 290, + 286, + 434 + ], + "type": "text", + "content": ". All models are evaluated on a single GPU using 5 runs on average. With Just In Time (JIT) Compilation with JAX, for 50 query points randomly sampled on the first frame, TAP-Net finishes within 0.1 seconds and TAPIR finishes within 0.3 seconds, i.e., roughly 150 frames per second. This is possible because TAPIR, like TAP-Net, maps efficiently to GPU parallel processing. In contrast, PIPs take 34.5 seconds due to a linear increase with respect to the number of points and frames processed. See Appendix 1 for more details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 443, + 147, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 147, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 147, + 455 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 462, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 286, + 712 + ], + "type": "text", + "content": "TAPIR vs Simple Combination We analyze the contributions of the main ideas. Our first contribution is to combine PIPs and TAP-Net (Fig. 3, center). Comparing this directly to PIPs is somewhat misleading: running the open-source 'chaining' code on Kinetics requires roughly 1 month on a V100 GPU (roughly 30 minutes to process a 10-second clip). But how crucial is chaining? After all, PIPs MLP-mixers have large spatial receptive fields due to their multi-resolution pyramid. Thus, we trained a naive version of PIPs without chaining: it initializes the entire track to be the query point, and then processes each chunk independently. For a fair comparison, we train using the same dataset and feature extractor, and other hyperparameters as TAPIR, but retain the 8-frame MLP Mixer with the same number of iterations as TAPIR. Table 4 shows that this algorithm (Chain-free PIPs) is surprisingly effective, on par with TAP-Net. Next, we combine the Chain-free PIPs with TAP-Net, replacing the constant initialization with TAP-Net's prediction. The \"simple combination\" (Fig. 3, center) alone improves over PIPs, Chain-free PIPs, and TAP-Net (i.e. " + }, + { + "bbox": [ + 47, + 462, + 286, + 712 + ], + "type": "inline_equation", + "content": "43.3\\%" + }, + { + "bbox": [ + 47, + 462, + 286, + 712 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 47, + 462, + 286, + 712 + ], + "type": "inline_equation", + "content": "50.0\\%" + }, + { + "bbox": [ + 47, + 462, + 286, + 712 + ], + "type": "text", + "content": " on DAVIS). However, the other new" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 543, + 137 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 543, + 137 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 543, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 543, + 137 + ], + "type": "table", + "html": "
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubric
PIPs [19]35.342.037.359.1
TAP-Net [12]48.341.159.265.4
Chain-Free PIPs47.243.361.375.3
Simple Combination52.950.062.778.1
TAPIR57.261.362.784.7
", + "image_path": "a4bfde8e56d762637d91c983d7def8211683d7c4162ce4fde04213fc2ee32c95.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 137, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 306, + 137, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 137, + 545, + 213 + ], + "type": "text", + "content": "Table 4. PIPs, TAP-Net, and a simple combination vs TAPIR. We compare TAPIR (Fig. 3 right) to TAP-Net, PIPs (Fig. 3 left), and the \"simple combination\" (Fig. 3 center). Chain-Free PIPs is our reimplementation of PIPs that removes the extremely slow chaining operation and uses the TAPIR backbone network. All models in this table are trained on Panning MOVi-E except PIPs, which uses its own dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 221, + 545, + 292 + ], + "blocks": [ + { + "bbox": [ + 307, + 221, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 307, + 221, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 307, + 221, + 545, + 292 + ], + "type": "table", + "html": "
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubic
Full Model57.261.362.784.7
- No Depthwise Conv54.953.861.979.7
- No Uncertainty Estimate54.458.661.583.4
- No Higher Res Feature54.054.059.580.4
- No TAP-Net Initialization54.759.360.384.1
- No Iterative Refinement48.141.660.364.9
", + "image_path": "74360d7d2ea35dec7ea6689570b7a450ea9e1c9977c2930cea0dee86a43c4323.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 306, + 293, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 306, + 293, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 545, + 380 + ], + "type": "text", + "content": "Table 5. Model ablation by removing one major component at a time. -No Depthwise Conv uses a MLP-Mixer on a chunked video similar to PIPs. -No Uncertainty Estimate uses the losses directly as described in TAP-Net. -No Higher Res Feature uses only 4 pyramid levels in iterative refinement and has no features at stride 4. -No TAP-Net Initialization uses a constant initialization for iterative refinement. All the components contribute non-trivially to the performance of the model." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 392, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 545, + 417 + ], + "type": "text", + "content": "ideas in TAPIR also account for a substantial portion of the performance (i.e. " + }, + { + "bbox": [ + 306, + 392, + 545, + 417 + ], + "type": "inline_equation", + "content": "50.0\\%" + }, + { + "bbox": [ + 306, + 392, + 545, + 417 + ], + "type": "text", + "content": " v.s. " + }, + { + "bbox": [ + 306, + 392, + 545, + 417 + ], + "type": "inline_equation", + "content": "61.3\\%" + }, + { + "bbox": [ + 306, + 392, + 545, + 417 + ], + "type": "text", + "content": " on DAVIS)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 433, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 433, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 433, + 545, + 661 + ], + "type": "text", + "content": "Model Components Table 5 shows the effect of our architectural decisions. First, we consider two novel architectural elements: the depthwise convolution which eliminates chunking from the PIPs model, and the uncertainty estimate. We see for Kinetics, both methods contribute roughly equally to performance, while for DAVIS, the depthwise conv is more important. A possible interpretation is that DAVIS contains more temporal continuity than Kinetics, so chunking is more harmful. Higher-resolution features are also predictably important to performance, especially on DAVIS. Table 5 also considers a few other model components that are not unique to TAPIR. TAP-Net initialization is important, but the model still gives reasonable performance without it. Interestingly, the full model without refinement (i.e. TAP-Net trained on Panning MOVi-E with the uncertainty estimate) performs similarly to TAP-Net trained on Panning MOVi-E (Table 4, second line); one possible interpretation is that TAP-Net is overfitting to its training data, leading to biased uncertainty estimate." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "Iterations of Refinement Finally, Table 6 shows how performance depends on the number of refinement iterations. We use the same number of iterations during training and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10067" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 72, + 233, + 141 + ], + "blocks": [ + { + "bbox": [ + 111, + 72, + 233, + 141 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 233, + 141 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 233, + 141 + ], + "type": "image", + "image_path": "7bd033911fe7e5d99bd3a9f4e3c8d8c8f3daa046be996530f354b2d454f6ab3f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 359, + 72, + 482, + 141 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 482, + 141 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 482, + 141 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 482, + 141 + ], + "type": "image", + "image_path": "f1c17f7774e3b29c030a58883aef1cd35b3a8f4087609ad3255d82740418fb5a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 144, + 171, + 212 + ], + "blocks": [ + { + "bbox": [ + 51, + 144, + 171, + 212 + ], + "lines": [ + { + "bbox": [ + 51, + 144, + 171, + 212 + ], + "spans": [ + { + "bbox": [ + 51, + 144, + 171, + 212 + ], + "type": "image", + "image_path": "1d0c6f2e31b36d134e031dad1dd0cd91ecfe6e88e411874b39b28853b7303e79.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 144, + 295, + 213 + ], + "blocks": [ + { + "bbox": [ + 173, + 144, + 295, + 213 + ], + "lines": [ + { + "bbox": [ + 173, + 144, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 173, + 144, + 295, + 213 + ], + "type": "image", + "image_path": "df8cae52a6cfbbe5b38956f5eb7cae6b377f7fec0bb0463ddbb1a225fe1c6a9c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 297, + 144, + 419, + 213 + ], + "blocks": [ + { + "bbox": [ + 297, + 144, + 419, + 213 + ], + "lines": [ + { + "bbox": [ + 297, + 144, + 419, + 213 + ], + "spans": [ + { + "bbox": [ + 297, + 144, + 419, + 213 + ], + "type": "image", + "image_path": "4b1c7eea26cda9008aebb980d994146f309845bdf1a00324dba7eaac8fe8cad2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 421, + 144, + 543, + 213 + ], + "blocks": [ + { + "bbox": [ + 421, + 144, + 543, + 213 + ], + "lines": [ + { + "bbox": [ + 421, + 144, + 543, + 213 + ], + "spans": [ + { + "bbox": [ + 421, + 144, + 543, + 213 + ], + "type": "image", + "image_path": "9ad20bbafd95ef327a0496dc830768d8e3f6f4bbbfbae337c4e1171d4c5c0e88.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 51, + 214, + 171, + 283 + ], + "blocks": [ + { + "bbox": [ + 51, + 214, + 171, + 283 + ], + "lines": [ + { + "bbox": [ + 51, + 214, + 171, + 283 + ], + "spans": [ + { + "bbox": [ + 51, + 214, + 171, + 283 + ], + "type": "image", + "image_path": "1399597778ab6bf89a0ff6c3a268f5afb60f19785cf5f9981df9c1a7a6c092a9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 173, + 214, + 295, + 283 + ], + "blocks": [ + { + "bbox": [ + 173, + 214, + 295, + 283 + ], + "lines": [ + { + "bbox": [ + 173, + 214, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 173, + 214, + 295, + 283 + ], + "type": "image", + "image_path": "5cc788129ca4c2062733f833d922e2ab373ae36d45f8861acced953e0b76f27e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 297, + 214, + 419, + 283 + ], + "blocks": [ + { + "bbox": [ + 297, + 214, + 419, + 283 + ], + "lines": [ + { + "bbox": [ + 297, + 214, + 419, + 283 + ], + "spans": [ + { + "bbox": [ + 297, + 214, + 419, + 283 + ], + "type": "image", + "image_path": "5060358dbd638861945a6017ef28801c1228c654c8c76260d1fcc410f3049af2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 421, + 214, + 543, + 283 + ], + "blocks": [ + { + "bbox": [ + 421, + 214, + 543, + 283 + ], + "lines": [ + { + "bbox": [ + 421, + 214, + 543, + 283 + ], + "spans": [ + { + "bbox": [ + 421, + 214, + 543, + 283 + ], + "type": "image", + "image_path": "77025247b256546591e8dec3d106a5283ce819c8fa945df15eaedb91fe960a31.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 285, + 171, + 353 + ], + "blocks": [ + { + "bbox": [ + 50, + 285, + 171, + 353 + ], + "lines": [ + { + "bbox": [ + 50, + 285, + 171, + 353 + ], + "spans": [ + { + "bbox": [ + 50, + 285, + 171, + 353 + ], + "type": "image", + "image_path": "513d138e95ab57b4046015b9d475203a6f820b6af1354b5883068f8f9589daa6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 359, + 545, + 425 + ], + "lines": [ + { + "bbox": [ + 46, + 359, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 545, + 425 + ], + "type": "text", + "content": "Figure 4. TAPIR compared to TAP-Net and PIPs. Top shows the query points on the video's first frame. Predictions for two later frames are below. We show predictions (filled circles) relative to ground truth (GT) (ends of the associated segments). x's indicate predictions where the GT is occluded, while empty circles are points visible in GT but predicted occluded (note: models still predict position regardless of occlusion). The majority of the street scene (left) gets occluded by a pedestrian; as a result, PIPs loses many points. TAP-Net fails on the right video, possibly because the textureless clothing is difficult to match without relying on temporal continuity. TAPIR, meanwhile, has far fewer failures." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 173, + 285, + 295, + 353 + ], + "blocks": [ + { + "bbox": [ + 173, + 285, + 295, + 353 + ], + "lines": [ + { + "bbox": [ + 173, + 285, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 173, + 285, + 295, + 353 + ], + "type": "image", + "image_path": "c9bb190f89efcf11d53c2f28d060278fd150dc7bd2377502bb6bed8a434becf0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 297, + 285, + 419, + 353 + ], + "blocks": [ + { + "bbox": [ + 297, + 285, + 419, + 353 + ], + "lines": [ + { + "bbox": [ + 297, + 285, + 419, + 353 + ], + "spans": [ + { + "bbox": [ + 297, + 285, + 419, + 353 + ], + "type": "image", + "image_path": "2a7c33a15f5ac5169a9164dc633849fb12d6ed1520f4b64e736ae90096f625f4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 421, + 285, + 543, + 353 + ], + "blocks": [ + { + "bbox": [ + 421, + 285, + 543, + 353 + ], + "lines": [ + { + "bbox": [ + 421, + 285, + 543, + 353 + ], + "spans": [ + { + "bbox": [ + 421, + 285, + 543, + 353 + ], + "type": "image", + "image_path": "4ee4ea0b56af29d946205efa1828a7b43082e0cb828e110b428465081e8de8c7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 48, + 433, + 287, + 519 + ], + "blocks": [ + { + "bbox": [ + 48, + 433, + 287, + 519 + ], + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 519 + ], + "type": "table", + "html": "
Average Jaccard (AJ)KineticsDAVISRGB-StackingKubric
0 iterations48.141.660.364.9
1 iteration55.755.062.780.1
2 iterations56.758.862.083.1
3 iterations56.860.660.984.1
4 iterations57.261.362.784.7
5 iterations57.861.661.584.3
6 iterations57.260.456.682.9
", + "image_path": "7690427e7b56b4901890c670906b68308a525770719d87f6d8b44215e140a8d8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 520, + 287, + 563 + ], + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 563 + ], + "type": "text", + "content": "Table 6. Evaluation performance against the number of iterative updates. It can be observed that after 4 iterations, the performance no longer improves significantly. This could be attributed to the fact that TAP-Net already provides a good initialization." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 570, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 664 + ], + "type": "text", + "content": "testing for all models. Interestingly, we observed the best performance for TAPIR at 4-5 iterations, whereas PIPs used 6 iterations. Likely a stronger initialization (i.e., TAP-Net) requires fewer iterations to converge. The decrease in performance with more iterations implies that there may be some degenerate behavior, such as oversmoothing, which may be a topic for future research. In this paper, we use 4 iterations for all other experiments." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Further model ablations, including experiments with RNNs and dense convolutions as replacements for our depthwise architecture, the number of feature pyramid layers, and the importance of time-shifting in the features, can" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 495 + ], + "type": "text", + "content": "be found in Appendix 4. Depthwise convolutions work as well or better than the alternatives while being less expensive. We find little impact of adding more feature pyramid layers as PIPs does, so we remove them from TAPIR. Finally, we find the impact of time-shifting to be minor." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 510, + 500, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 500, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 500, + 525 + ], + "type": "text", + "content": "5. Animating still images with TAPIR" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 532, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 651 + ], + "type": "text", + "content": "Generative modeling of images and videos has experienced an explosion in popularity due to recent striking text-conditioned generation results using diffusion models [17, 21, 42, 46, 50]. Animating still images remains extremely challenging, due to the high computational cost of video modeling and the difficulty of generating realistic motion. Is it possible to improve the generated videos via an explicit representation of surface motion? Here, we develop a proof-of-concept for a diffusion-based system that uses TAPIR tracks to do this." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "We perform video prediction using a pair of diffusion models: one which generates dense trajectories given an input image (the \"trajectory\" model), and a second which takes the input image and the trajectories and generates pixels (the \"pixel\" model). At test time, the outputs from the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10068" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "content": "trajectory model dictate the motion that should be present in the video, ideally ensuring physical plausibility. The pixel model uses the trajectories to attend to the appropriate location in the first frame, in order to produce appearance which is consistent across the video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "type": "text", + "content": "For the trajectory model, we encode the conditioning image with a small ConvNet, which produces a dense grid of features at stride 4. This is used as conditioning for a U-Net that performs diffusion denoising on the trajectories. We find that encoding the trajectories is important: for each frame, we include the raw " + }, + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "type": "text", + "content": " positions, both in absolute coordinates and coordinates relative to the starting position, and we also include a position encoding of the absolute locations with 32 Fourier features. The U-Net architecture includes self-attention layers, so the Fourier features allow the model to measure whether trajectories are spatially nearby late in an animation. For simplicity and memory efficiency, we stack these values for all frames along the channel axis: thus, the input is a 2D grid of shape " + }, + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "type": "inline_equation", + "content": "H // 4 \\times W // 4 \\times (T \\cdot (1 + 2 + 2 + 32))" + }, + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "type": "text", + "content": ", (occlusion, relative position, absolute position, fourier encoding), which is processed with a 2D U-Net. The model predicts the clean data given the noisy data, which is the typical \"clean data prediction\" diffusion setup. The pixel model is a more traditional image generation model, which predicts the noise residual given noisy pixels. It takes three inputs, and outputs a single denoised frame. The first input is 3 noisy frames, allowing the model to infer some temporal context. The second is the first frame, warped according to the trajectories from the trajectory model. The third is features for the first frame, computed by a standard ConvNet before, being warped according to the same trajectories. We train on an internal dataset of 1M video clips, processed into 24-frame clips at " + }, + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 133, + 289, + 479 + ], + "type": "text", + "content": " resolution. See supplementary for more details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 479, + 288, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 612 + ], + "type": "text", + "content": "Fig. 5 shows the trajectories produced by our trajectory prediction model on selected images. Our model can make multiple different physically-plausible predictions for motion for each input image, suggesting that it can recognize people and do rudimentary human pose estimation without supervision. Note that this demonstrates two levels of transfer: TAPIR was never trained on humans, while the trajectory-prediction model was trained on video rather than still images like these. It is difficult to judge these motions from a static figure, so we encourage the reader to view videos generated with our pipeline in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 621, + 119, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 119, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 119, + 634 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "In this paper, we introduce TAPIR, a novel model for Tracking Any Point (TAP) that adopts a two-stage approach, comprising a matching stage and a refinement stage. It demonstrates a substantial improvement over the state-of-the-art, providing stable predictions and occlusion robustness, and scales to high-resolution videos. We also" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 425, + 186 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 425, + 186 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 425, + 186 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 425, + 186 + ], + "type": "image", + "image_path": "816c5b75fa56ca10572812a20140fa2670a0ffe37d0150b3d7221821727ac666.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 541, + 185 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 541, + 185 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 541, + 185 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 541, + 185 + ], + "type": "image", + "image_path": "98ae30fc91e6ce0f9c44cd5ccf47bcb8f4024926cba2ad210adf896b46698a26.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 186, + 425, + 300 + ], + "blocks": [ + { + "bbox": [ + 310, + 186, + 425, + 300 + ], + "lines": [ + { + "bbox": [ + 310, + 186, + 425, + 300 + ], + "spans": [ + { + "bbox": [ + 310, + 186, + 425, + 300 + ], + "type": "image", + "image_path": "ed4d56d95bf060156b51776ed607985448fdf374190505f424b0cf5e46e17b0b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 186, + 541, + 300 + ], + "blocks": [ + { + "bbox": [ + 427, + 186, + 541, + 300 + ], + "lines": [ + { + "bbox": [ + 427, + 186, + 541, + 300 + ], + "spans": [ + { + "bbox": [ + 427, + 186, + 541, + 300 + ], + "type": "image", + "image_path": "6e1cef806f85cced0c84db244635233367330bf56b3b7db5140207acf973696b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 300, + 425, + 414 + ], + "blocks": [ + { + "bbox": [ + 310, + 300, + 425, + 414 + ], + "lines": [ + { + "bbox": [ + 310, + 300, + 425, + 414 + ], + "spans": [ + { + "bbox": [ + 310, + 300, + 425, + 414 + ], + "type": "image", + "image_path": "55c5c6769ec5a4c23a01480463c00f80990234af931a9a34862891b7ae756672.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 415, + 545, + 515 + ], + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 515 + ], + "type": "text", + "content": "Figure 5. Animating still frames. Top row shows the input images. Each following row shows a visualization of a single sample from our trajectory model: the dark purple end is the starting point of the trajectory, while the colors get brighter and yellower with distance. In the left image, one sample shows the hand moving forward in a cutting motion, with a slight incline of the head; in the other, the entire person moves to the left. On the right, the first sample moves one arm in a circle, while the second shows the head turning and the other arm being raised." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 427, + 300, + 541, + 414 + ], + "blocks": [ + { + "bbox": [ + 427, + 300, + 541, + 414 + ], + "lines": [ + { + "bbox": [ + 427, + 300, + 541, + 414 + ], + "spans": [ + { + "bbox": [ + 427, + 300, + 541, + 414 + ], + "type": "image", + "image_path": "ccba319cec58fa54fa79f2a9956135a88e2862db7e713fd33dea9b61f65a1e48.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 526, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 612 + ], + "type": "text", + "content": "show a proof-of-concept of animating still frames, which we hope can serve as a foundation for future graphics research. TAPIR still has some limitations, such as difficulty in precisely perceiving the boundary between background and foreground, and it may still struggle with large appearance changes. Nevertheless, our findings indicate a promising future for TAP in computer vision." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "Acknowledgements We would like to thank Lucas Smaira and Larisa Markeeva for help with datasets and codebases, and Jon Scholz and Todor Davchev for advice on the needs of robotics. We also thank Adam Harley, Viorica Patraucean and Daniel Zoran for helpful discussions, and Mateusz Malinowski for paper comments. Finally, we thank Nando de Freitas for his support." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10069" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5173-5182, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 145, + 288, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 145, + 288, + 188 + ], + "spans": [ + { + "bbox": [ + 53, + 145, + 288, + 188 + ], + "type": "text", + "content": "[2] Marian Stewart Bartlett, Gwen Littlewort, Mark G Frank, Claudia Lainscsek, Ian R Fasel, Javier R Movellan, et al. Automatic recognition of facial actions in spontaneous expressions. J. Multim., 1(6):22-35, 2006." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 188, + 288, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 188, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 53, + 188, + 288, + 231 + ], + "type": "text", + "content": "[3] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In Proceedings of European Conference on Computer Vision (ECCV), pages 404-417. Springer, 2006." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "type": "text", + "content": "[4] Thomas Brox, Christoph Bregler, and Jitendra Malik. Large displacement optical flow. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 41-48. IEEE, 2009." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "type": "text", + "content": "[5] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In Proceedings of European Conference on Computer Vision (ECCV), pages 25-36. Springer, 2004." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 317, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 288, + 360 + ], + "type": "text", + "content": "[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Proceedings of European Conference on Computer Vision (ECCV), pages 611-625. Springer, 2012." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 360, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 288, + 413 + ], + "type": "text", + "content": "[7] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4733-4742, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 456 + ], + "type": "text", + "content": "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6299-6308, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 456, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 288, + 499 + ], + "type": "text", + "content": "[9] Zezhou Cheng, Jong-Chyi Su, and Subhransu Maji. On equivariant and invariant learning of object landmark representations. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 9897-9906, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 553 + ], + "type": "text", + "content": "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5828-5839, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 553, + 287, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 553, + 287, + 607 + ], + "spans": [ + { + "bbox": [ + 48, + 553, + 287, + 607 + ], + "type": "text", + "content": "[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 224-236, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 607, + 287, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 607, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 48, + 607, + 287, + 661 + ], + "type": "text", + "content": "[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Proceedings of Neural Information Processing Systems Datasets and Benchmarks Track, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 661, + 287, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 661, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 661, + 287, + 704 + ], + "type": "text", + "content": "[13] Carl Doersch and Andrew Zisserman. Sim2real transfer learning for 3d human pose estimation: motion to the rescue. Proceedings of Neural Information Processing Systems (NeurIPS), 32, 2019." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 708 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 2758-2766, 2015." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 139, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 139, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 545, + 191 + ], + "type": "text", + "content": "[15] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361. IEEE, 2012." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 192, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 192, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 192, + 545, + 323 + ], + "type": "text", + "content": "[16] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: a scalable dataset generator. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 323, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 323, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 323, + 545, + 376 + ], + "type": "text", + "content": "[17] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10696-10706, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "text", + "content": "[18] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Denseseposé: Dense human pose estimation in the wild. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7297-7306, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 419, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 419, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 308, + 419, + 545, + 463 + ], + "type": "text", + "content": "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In Proceedings of European Conference on Computer Vision (ECCV), pages 59-75. Springer, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 462, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 462, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 308, + 462, + 545, + 494 + ], + "type": "text", + "content": "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 494, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 545, + 525 + ], + "type": "text", + "content": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 525, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 548 + ], + "type": "text", + "content": "[22] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial Intelligence, 17(1-3):185-203, 1981." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "text", + "content": "[23] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 644 + ], + "type": "text", + "content": "[24] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. Proceedings of Neural Information Processing Systems (NeurIPS), 31, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 644, + 545, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 644, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 308, + 644, + 545, + 696 + ], + "type": "text", + "content": "[25] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Self-supervised learning of interpretable keypoints from unlabeled videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8787-8797, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 696, + 545, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 696, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 308, + 696, + 545, + 708 + ], + "type": "text", + "content": "[26] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10070" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 706 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3192-3199, 2013." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "type": "text", + "content": "[27] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 6207-6217, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 159, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 159, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 288, + 213 + ], + "type": "text", + "content": "[28] Namhoon Lee, Wongun Choi, Paul Vernaza, Christopher B Choy, Philip HS Torr, and Manmohan Chandraker. Desire: Distant future prediction in dynamic scenes with interacting agents. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 336-345, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 213, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 213, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 288, + 267 + ], + "type": "text", + "content": "[29] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4521-4530, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 267, + 288, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 288, + 310 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 288, + 310 + ], + "type": "text", + "content": "[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2041-2050, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 288, + 353 + ], + "type": "text", + "content": "[31] Ji Lin, Chuang Gan, Kuan Wang, and Song Han. Tsm: Temporal shift module for efficient and scalable video understanding on edge devices. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 288, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 288, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 288, + 406 + ], + "type": "text", + "content": "[32] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6728-6737, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 406, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 288, + 449 + ], + "type": "text", + "content": "[33] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):978-994, 2010." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 491 + ], + "type": "text", + "content": "[34] David G Lowe. Object recognition from local scale-invariant features. In Proceedings of IEEE International Conference on Computer Vision (ICCV), volume 2, pages 1150-1157. IEEE, 1999." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 524 + ], + "type": "text", + "content": "[35] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 524, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 288, + 567 + ], + "type": "text", + "content": "[36] Bruce D Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In Proceedings of International Joint Conference on Artificial Intelligence (IJCAI), pages 674-679, 1981." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 609 + ], + "type": "text", + "content": "[37] Lucas Manuelli, Yunzhu Li, Pete Florence, and Russ Tedrake. Keypoints into the future: Self-supervised correspondence in model-based reinforcement learning. arXiv preprint arXiv:2009.05085, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 610, + 288, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 610, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 288, + 664 + ], + "type": "text", + "content": "[38] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. Mantra: Memory augmented networks for multiple trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7143-7152, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 664, + 288, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 664, + 288, + 706 + ], + "spans": [ + { + "bbox": [ + 48, + 664, + 288, + 706 + ], + "type": "text", + "content": "[39] David Marr and Tomaso Poggio. A computational theory of human stereo vision. Proceedings of the Royal Society of London. Series B. Biological Sciences, 204(1156):301-328, 1979." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 707 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "text", + "content": "[40] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 139, + 547, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 139, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 547, + 203 + ], + "type": "text", + "content": "[41] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 203, + 547, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 203, + 547, + 256 + ], + "spans": [ + { + "bbox": [ + 308, + 203, + 547, + 256 + ], + "type": "text", + "content": "[42] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 257, + 547, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 257, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 257, + 547, + 289 + ], + "type": "text", + "content": "[43] Yuki Ono, Eduard Trulls, Pascal Fua, and Kwang Moo Yi. Lf-net: Learning local features from images. Proceedings of Neural Information Processing Systems (NeurIPS), 31, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 289, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 289, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 289, + 547, + 331 + ], + "type": "text", + "content": "[44] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 331, + 546, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 546, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 546, + 385 + ], + "type": "text", + "content": "[45] Deva Ramanan, David A Forsyth, and Andrew Zisserman. Strike a pose: Tracking people by finding stylized poses. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), volume 1, pages 271-278. IEEE, 2005." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 385, + 546, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 385, + 546, + 428 + ], + "spans": [ + { + "bbox": [ + 308, + 385, + 546, + 428 + ], + "type": "text", + "content": "[46] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 428, + 547, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 547, + 471 + ], + "type": "text", + "content": "[47] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4161-4170, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 471, + 546, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 546, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 546, + 524 + ], + "type": "text", + "content": "[48] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3d reconstruction of interacting hands via collision-aware factorized refinements. In 2021 International Conference on 3D Vision (3DV), pages 432-441. IEEE, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 524, + 547, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 547, + 557 + ], + "type": "text", + "content": "[49] Michael Rubinstein, Ce Liu, and William Freeman. Towards longer long-range motion trajectories. In Proceedings of British Machine Vision Conference, 2012." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 546, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 546, + 622 + ], + "type": "text", + "content": "[50] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 622, + 546, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 622, + 546, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 622, + 546, + 654 + ], + "type": "text", + "content": "[51] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International Journal of Computer Vision, 80(1):72-91, 2008." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 654, + 546, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 654, + 546, + 696 + ], + "spans": [ + { + "bbox": [ + 308, + 654, + 546, + 696 + ], + "type": "text", + "content": "[52] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 696, + 546, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 696, + 546, + 707 + ], + "spans": [ + { + "bbox": [ + 308, + 696, + 546, + 707 + ], + "type": "text", + "content": "[53] Thomas Schops, Johannes L Schonberger, Silvano Galliani," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10071" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 707 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "text", + "content": "Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3260-3269, 2017." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 127, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 127, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 287, + 169 + ], + "type": "text", + "content": "[54] Ishwar K Sethi and Ramesh Jain. Finding trajectories of feature points in a monocular image sequence. IEEE Transactions on Pattern Analysis and Machine Intelligence, (1):56-73, 1987." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 170, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 170, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 48, + 170, + 288, + 235 + ], + "type": "text", + "content": "[55] Jie Shen, Stefanos Zafeiriou, Grigoris G Chrysos, Jean Kossaifi, Georgios Tzimiropoulos, and Maja Pantic. The first facial landmark tracking in-the-wild challenge: Benchmark and results. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, pages 50-58, 2015." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 235, + 288, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 288, + 299 + ], + "type": "text", + "content": "[56] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2107-2116, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 300, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 287, + 331 + ], + "type": "text", + "content": "[57] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. In Proceedings of European Conference on Computer Vision (ECCV), 2004." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 396 + ], + "type": "text", + "content": "[58] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6738-6748, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 396, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 287, + 461 + ], + "type": "text", + "content": "[59] Deqing Sun, Daniel Vlasic, Charles Herrmann, Varun Jampani, Michael Krainin, Huiwen Chang, Ramin Zabih, William T Freeman, and Ce Liu. Autoflow: Learning a better training set for optical flow. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10093-10102, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 461, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 461, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 461, + 287, + 514 + ], + "type": "text", + "content": "[60] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8934-8943, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 515, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 557 + ], + "type": "text", + "content": "[61] Yi Sun, Yuheng Chen, Xiaogang Wang, and Xiaou Tang. Deep learning face representation by joint identification-verification. Proceedings of Neural Information Processing Systems (NeurIPS), 27, 2014." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 601 + ], + "type": "text", + "content": "[62] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Proceedings of European Conference on Computer Vision (ECCV), pages 402-419. Springer, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "text", + "content": "[63] James Thewlis, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks by factorized spatial embeddings. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 5916-5925, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 643, + 287, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 643, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 48, + 643, + 287, + 686 + ], + "type": "text", + "content": "[64] Jonathan Tompson, Murphy Stein, Yann Lecun, and Ken Perlin. Real-time continuous pose recovery of human hands using convolutional networks. ACM Transactions on Graphics (ToG), 33(5):1-10, 2014." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 687, + 287, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 687, + 287, + 707 + ], + "spans": [ + { + "bbox": [ + 48, + 687, + 287, + 707 + ], + "type": "text", + "content": "[65] Philip HS Torr and Andrew Zisserman. Feature based methods for structure and motion estimation. In International" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "workshop on vision algorithms, pages 278-294. Springer, 1999." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 137 + ], + "type": "text", + "content": "[66] Sudheendra Vijayanarasimhan, Susanna Ricco, Cordelia Schmid, Rahul Sukthankar, and Katerina Fragkiadaki. Sfmnet: Learning of structure and motion from video. arXiv preprint arXiv:1704.07804, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 137, + 545, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 137, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 545, + 191 + ], + "type": "text", + "content": "[67] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of European Conference on Computer Vision (ECCV), pages 601-617, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 191, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 191, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 545, + 233 + ], + "type": "text", + "content": "[68] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 3551-3558, 2013." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 233, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 233, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 233, + 545, + 266 + ], + "type": "text", + "content": "[69] Yue Wang and Justin M Solomon. Prnet: Self-supervised learning for partial-to-partial registration. Proceedings of Neural Information Processing Systems (NeurIPS), 32, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 266, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 308 + ], + "type": "text", + "content": "[70] Xuehan Xiong and Fernando De la Torre. Supervised descent method and its applications to face alignment. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 532-539, 2013." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 361 + ], + "type": "text", + "content": "[71] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8121-8130, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 361, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 406 + ], + "type": "text", + "content": "[72] Jiarui Xu and Xiaolong Wang. Rethinking self-supervised correspondence learning: A video frame-level similarity perspective. In Proceedings of IEEE International Conference on Computer Vision (ICCV), pages 10075-10085, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 405, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 459 + ], + "type": "text", + "content": "[73] Zhichao Yin, Trevor Darrell, and Fisher Yu. Hierarchical discrete distribution decomposition for match density estimation. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6044-6053, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 459, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 513 + ], + "type": "text", + "content": "[74] Yuting Zhang, Yijie Guo, Yixin Jin, Yijun Luo, Zhiyuan He, and Honglak Lee. Unsupervised discovery of object landmarks as structural representations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2694-2703, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 513, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 545 + ], + "type": "text", + "content": "[75] Zhengyou Zhang. A flexible new technique for camera calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(11):1330-1334, 2000." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 545, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 587 + ], + "type": "text", + "content": "[76] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 587, + 545, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 641 + ], + "type": "text", + "content": "[77] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of IEEE International Conference on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 641, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 641, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 307, + 641, + 545, + 707 + ], + "type": "text", + "content": "[78] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rehmann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew Fitzgibbon, Charles Loop, Christian Theobalt, et al. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (ToG), 33(4):1-12, 2014." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10072" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/ff5641a5-17a3-470a-a519-ccac898db108_content_list.json b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/ff5641a5-17a3-470a-a519-ccac898db108_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ee867abd3c41d0bd3aa2f8dbf6b233023634a5a5 --- /dev/null +++ b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/ff5641a5-17a3-470a-a519-ccac898db108_content_list.json @@ -0,0 +1,1798 @@ +[ + { + "type": "text", + "text": "TARGET: Federated Class-Continual Learning via Exemplar-Free Distillation", + "text_level": 1, + "bbox": [ + 89, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jie Zhang* \nETH Zurich \nzj.jayzhang@gmail.com", + "bbox": [ + 359, + 181, + 611, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Chen & Weiming Zhuang & Lingjuan Lyu† \nSony AI", + "bbox": [ + 292, + 258, + 683, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ChenA.Chen, weiming.zhuang, Lingjuan.Lv}@sony.com", + "bbox": [ + 191, + 295, + 777, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 345, + 313, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper focuses on an under-explored yet important problem: Federated Class-Continual Learning (FCCL), where new classes are dynamically added in federated learning. Existing FCCL works suffer from various limitations, such as requiring additional datasets or storing the private data from previous tasks. In response, we first demonstrate that non-IID data exacerbates catastrophic forgetting issue in FL. Then we propose a novel method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free distTillage), which alleviates catastrophic forgetting in FCCL while preserving client data privacy. Our proposed method leverages the previously trained global model to transfer knowledge of old tasks to the current task at the model level. Moreover, a generator is trained to produce synthetic data to simulate the global distribution of data on each client at the data level. Compared to previous FCCL methods, TARGET does not require any additional datasets or storing real data from previous tasks, which makes it ideal for data-sensitive scenarios.", + "bbox": [ + 75, + 378, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 698, + 209, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) is a privacy-aware learning paradigm that facilitates collaborations among multiple entities (e.g., edge devices or organizations) [38, 20, 28, 66, 17]. Each entity or client in FL retains data locally and transfers only training updates to the central server for aggregation.", + "bbox": [ + 75, + 724, + 468, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional FL studies assume that the data classes and domains are static, but the new classes could emerge and data domains could change over time in reality [64,", + "bbox": [ + 75, + 816, + 468, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "40, 1, 65]. For example, multiple health institutions could use FL to collaborate and train models to identify COVID-19 [56, 9, 2] strains; new COVID-19 strains, however, continue to emerge due to high mutation rate of virus. An intuitive solution to this issue of continuously emerging data classes is training new models from scratch, but this is impractical as it would require significant extra computation cost. Another method is transfer learning from the previously trained model, but this method suffers from catastrophic forgetting [23, 22, 47, 16], degrading performance on the previous classes.", + "bbox": [ + 500, + 348, + 893, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these issues, recent research [36, 11, 3, 19] has introduced the concept of Continual Learning (CL) [61, 42, 51, 49, 48] within the FL framework. These methods, collectively referred to as Federated Continual Learning (FCL), aim to mitigate the problems of catastrophic forgetting in FL. In most FCL scenarios, new classes are dynamically added, which we call Federated Class-Continual Learning (FCCL). FCCL allows local clients to continuously collect new data, and new classes can be added at any time.", + "bbox": [ + 496, + 518, + 893, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, existing FCCL works suffer from various limitations. For example, Ma et al. [36] utilize an unlabeled surrogate dataset to address the catastrophic forgetting problem, which may be difficult to obtain in some data sensitive scenarios. Furthermore, the usage of an unlabeled surrogate dataset may not be ideal for certain types of data, as it may not capture the full complexity of the original data. In CL, exemplar-based methods [44, 33, 54, 13] have achieved leading performance. An exemplar refers to a sample or instance of a previously seen data point that is retained in a memory buffer for future use in the learning process. Dong et al. [11] propose a exemplar-based method that stores historical data to address catastrophic forgetting. However, in many privacy-sensitive scenarios (e.g., hospitals and medical research institutions), users are", + "bbox": [ + 496, + 674, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during internship at Sony AI.", + "bbox": [ + 94, + 875, + 316, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4782", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "not permitted to store data from previous tasks due to privacy and policy concerns and data will not be kept for a long time [21, 53, 34, 50]. In summary, the majority of FCCL methods train the global model with additional datasets or previous task data, which could potentially violate data privacy regulations. This dilemma prompts us to consider the following question:", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Question: How to effectively alleviate the catastrophic forgetting problem in the FCCL without storing the local private data of the client or any additional datasets?", + "bbox": [ + 73, + 198, + 470, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this question, we conduct a systematic analysis and observe that the imbalanced distribution of data among clients in FL exacerbates the catastrophic forgetting problem (see Section 3.3). In order to fix this problem: 1) at the model level, we leverage the previously trained global model to transfer knowledge of the old tasks to the current task. 2) at the data level, we train a generator to produce synthetic data that aims to simulate the global distribution of data on each client. Drawing on these insights, we present a method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillation) that mitigates catastrophic forgetting in FCCL without compromising clients' data privacy.", + "bbox": [ + 75, + 244, + 468, + 440 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be concluded as follows:", + "bbox": [ + 96, + 441, + 406, + 455 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to demonstrate that non-independent and identically distributed (non-IID) data exacerbates catastrophic forgetting issue in FL. Then we propose a novel method called TARGET, which alleviates the catastrophic forgetting in FCCL by leveraging global information.", + "- Compared to previous FCCL methods, TARGET doesn't require extra datasets or data from previous tasks, it can be applied in data sensitive scenarios.", + "- Extensive experiments demonstrate the efficacy of our proposed method. For example, when partitioning the CIFAR-100 dataset into five tasks, our method achieves an accuracy of $36.31\\%$ , which is about $6\\%$ higher than the best baseline method." + ], + "bbox": [ + 94, + 457, + 468, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 684, + 218, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Federated Continual Learning", + "text_level": 1, + "bbox": [ + 76, + 710, + 346, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, deep neural networks (DNNs) have been widely used as a fundamental technology for the development of artificial intelligence, both in established and emerging fields [68, 67, 32, 31, 10]. With the advancement of deep learning, an increasing number of researchers have started to focus on the construction of privacy-preserving deep learning frameworks. FL is a paradigm for collaboratively building a model across multiple clients [24, 39], which is gaining momentum in recent years. However, Federated Continual Learning (FCL), which focused on both FL and CL simultaneously is just emerging and still", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "remains pending further research [52, 43, 58]. Apart from the client-wise catastrophic forgetting, FCL paradigm also poses new challenges such as inter-client interference and communication-efficiency [58]. FedWeIT solves these challenges through decomposing the parameters into three parts, i.e., global parameters, local based parameters, and task-adaptive parameters [58]. Concurrently, Concept-Drift-Aware Federated Averaging (CDA-FedAvg) extends the popular FL algorithm, Federated Averaging (FedAvg), to tackle the CL problem by introducing the concept drift detection and adaptation [14]. FCL with Distillation (CFeD) treats the model trained with the last task as the teacher model to perform a knowledge distillation and proposes a server distillation mechanism to deal with non-i.i.d. issue [37]. Global-Local Forgetting Compensation (GLFC) designs a class-aware gradient compensation loss and a class-semantic relation distillation loss to prevent forgetting, and a proxy server to mitigate the non-independent and identically distributed (non-IID) problem [12]. Additionally, Federated Selective Inter-client Transfer (FedSeIT) applies FCL to NLP through selectively combining model parameters of foreign clients and selecting informative tasks to perform knowledge transfer [5]. These papers illustrate the growing interest in FCL and the need for novel approaches to address its unique challenges. As research in this area continues to advance, we can expect to see more innovative approaches to FCL that further improve its performance, scalability, and privacy preservation.", + "bbox": [ + 496, + 90, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Continual Learning", + "text_level": 1, + "bbox": [ + 500, + 527, + 689, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual Learning has been studied extensively, several training methods have been proposed to address the catastrophic forgetting challenge it presents. Regularization-based approaches: Elastic Weight Consolidation (EWC) selectively penalizes the network parameters that are important for old tasks [23]. Synaptic intelligence (SI) uses a memory buffer to store important network parameters [61]. Incremental moment matching (IMM) modelings the posterior distribution after learning multiple tasks as a mixture of Gaussian models [27]. Stable SGD proposes to carefully design the training regimes such as learning rate decay, batch size, dropout, and optimizer to alleviate forgetting. Replay-based approaches: The methods in this family construct a memory to store the past information which will be presented to the model for reviewing in future tasks. Some store the knowledge of the previous tasks, known as experience replay [35, 45, 49, 46]. iCaRL preserves the most representative samples of each class [44]. Averaged Gradient Episodic Memory (A-GEM) builds an episodic memory of model parameter gradients[8]. Architecture-based: DEN expands the model size [60] and RCL utilizes reinforcement learning [55]. APD divides the model parameters into shared and task-specific parameters to restrict the", + "bbox": [ + 496, + 553, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4783", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model complexity [59].", + "bbox": [ + 76, + 90, + 235, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Exemplar-Free Continual Learning Besides, a promising line of work focuses on data-free Continual Learning. DeepDream perturbs current training samples into images that maximize \"forgetting\" from the previous tasks [41]. DeepInversion proposes a model inversion technique and evaluates its performance in a Continual Learning scenario but found limited success [57]. DFCIL [50] investigates the failure and decomposes the CE-loss into two different losses which guarantee the learning of effective features. These methods enable Continual Learning in scenarios where storing old task data is not feasible or desirable, such as in privacy-sensitive applications.", + "bbox": [ + 75, + 126, + 472, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Catastrophic Forgetting in FCCL", + "text_level": 1, + "bbox": [ + 76, + 321, + 382, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section conducts an in-depth analysis of catastrophic forgetting problem in federated class-continual learning (FCCL). We start by providing a formal definition of the problem. Then, we investigate the forgetting issue in FCCL and discuss potential methods to mitigate it.", + "bbox": [ + 75, + 348, + 468, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 76, + 433, + 261, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Federated Class-Continual Learning (FCCL) focuses on the problem of learning models for new classes over time in FL. An FCCL framework consists of a central server and multiple clients. All clients do not share their raw data with any other client or the central server. Each client learns from a sequence of $n$ tasks, where $k$ -th task contains non-overlapping subsets of classes $C_k \\in C$ , where $C$ is the set of all possible classes. In our privacy-aware scenario, the task stream is presented in an unknown order, and each client can only access its local data from task $k$ during that task's training period, which is no longer accessible thereafter. Note that the models are trained in a distributed manner, where each party has access to only a subset of the classes $C_k$ (i.e. non-IID). We also consider a more challenging and practical setting where the data in each client is heterogeneous. In this paper, we assume the label distribution of data in each client is skewed [29, 63].", + "bbox": [ + 75, + 457, + 472, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Forgetting Issue in FCCL The goal of the global model optimization problem at task $k$ is to minimize the overall classification error on the current set of classes $C_k$ . However, when a new task arises, clients are not able to access data from previous (old) tasks due to privacy concerns and can only update their local model with data from the new task. This often leads to a significant decrease in performance on previous tasks, which is known as catastrophic forgetting [23, 22, 47, 16]. To mitigate catastrophic forgetting in the global model, we aim to minimize the overall classification error on the current set of classes $C_k$ , while", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d92f5b7b0a003af57b65bca690ac044f6a491b05b0137b553fbe616f8cfa6720.jpg", + "image_caption": [ + "(a) Test Accuracy", + "Figure 1: Forgetting under different data partitions. We partitioned the CIFAR-100 dataset into five tasks with 20 classes each, distributed among five clients. \"NIID\" refers to non-IID, where a lower value represents a more imbalanced or skewed distribution of data. Non-IIDness refers to the Dirichlet parameter." + ], + "image_footnote": [], + "bbox": [ + 526, + 98, + 694, + 229 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/db207368bef2e6b47c3e0075a06635eef4acaa8dc916ecf0c50c63b54904634a.jpg", + "image_caption": [ + "(b) Relative Forgetting" + ], + "image_footnote": [], + "bbox": [ + 702, + 98, + 875, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "simultaneously minimizing the changes to the previously learned classes. Formally, the objective function can be written as:", + "bbox": [ + 496, + 391, + 892, + 436 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {k}} \\sum_ {c \\in C _ {k}} \\sum_ {i = 1} ^ {m _ {c}} L \\left(f _ {k} \\left(x _ {i, c}; \\theta_ {k}\\right), c\\right) + \\alpha R \\left(\\theta_ {k}, \\theta_ {k - 1}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 441, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{k}$ is the model parameter at round $k$ , $L$ is a loss function that measures the classification error, $R$ is a regularization term that penalizes changes to the previous model parameters, $m_{c}$ is the number of data in class $c$ , and $\\alpha$ is a hyper-parameter that controls the strength of the regularization. In this formula, $f_{k}(x_{i,c};\\theta_{k})$ represents the classification model that takes as input a data point $x_{i,c}$ associated with class $c$ and outputs a probability distribution over the set of classes in $C_k$ . The regularization term $R$ encourages the new model parameters to be close to the previous model parameters $\\theta_{k - 1}$ , in order to prevent catastrophic forgetting of the previously learned classes.", + "bbox": [ + 496, + 489, + 893, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Heterogeneous Data Exacerbates Forgetting", + "text_level": 1, + "bbox": [ + 498, + 680, + 875, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We argue that the degree of data heterogeneity has a substantial impact on catastrophic forgetting. To verify this, we conduct an experiment on CIFAR-100 dataset [25] with different degrees of data heterogeneity. Inspired by Backward Transfer (BwT) [6], we derive the following formula to measure the severity of forgetting, a popular forgetting measure in CL [4, 7, 8]:", + "bbox": [ + 496, + 703, + 893, + 809 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {k} = \\frac {1}{k - 1} \\sum_ {j = 1} ^ {k - 1} f _ {j} ^ {k}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 816, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_k$ denotes the average forgetting at $k$ -th task and $f_j^k$ quantifies forgetting for the $j$ -th ( $j < k$ ) task after the model", + "bbox": [ + 496, + 869, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4784", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "has been continually trained up to task $k$ . Specifically, for a given data distribution, $f_{j}^{k}$ can be expressed as follows:", + "bbox": [ + 75, + 90, + 468, + 125 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {j} ^ {k} = \\frac {1}{\\left| \\mathcal {C} ^ {j} \\right|} \\sum_ {c \\in \\mathcal {C} ^ {j}} \\max _ {t \\in \\{1, \\dots . N - 1 \\}} \\left(\\mathcal {A} _ {c} ^ {(n)} - \\mathcal {A} _ {c} ^ {(N)}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 133, + 468, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{C}^j$ is a set of classes related to the $j$ -th task, $\\mathcal{A}_c^{(n)}$ is the accuracy on class $c$ at round $t$ , and $\\mathcal{A}_c^{(N)}$ is the final accuracy on class $c$ after learning all tasks. Note that $f_j^k$ captures the average gap between the peak accuracy and the final accuracy for each class of the $j$ -th task after learning the $k$ -th task.", + "bbox": [ + 75, + 183, + 468, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further extend the catastrophic forgetting measurement $\\mathcal{F}$ in Equation 2 to FL under different data partitions and introduce a relative metric $\\mathcal{R}$ to measure forgetting as follows:", + "bbox": [ + 75, + 277, + 468, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {k} = \\frac {\\sum_ {j = 1} ^ {k - 1} f _ {j} ^ {k}}{\\sum_ {j = 1} ^ {k - 1} \\mathcal {A} _ {(j , k)}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 335, + 468, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{A}_{(j,k)}$ is the accuracy on task $j$ after learning task $k$ . For different data partitions, an increased $\\mathcal{R}_k$ indicates a more serious forgetting of previous tasks.", + "bbox": [ + 75, + 381, + 468, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 1 illustrates the impact of catastrophic forgetting under independent and identically distributed (IID) and different levels of non-independent and identically distributed (non-IID) data partitions. In particular, we employ the Dirichlet distribution, which is widely used in FL [29, 28], to simulate the imbalanced label distribution among different clients. Figure 1(a) shows that the accuracy of the model degrades as training proceeds to new tasks in the IID setting. The performance is even worse in the non-IID settings. These results suggest that FCCL faces significant challenge on extreme non-IID settings. We further analyze the forgetting phenomenon in the Figure 1(b). The higher degree of non-IID exacerbates the forgetting phenomenon in FCCL. These empirical studies motivate us to further investigate the catastrophic forgetting issue in FCCL.", + "bbox": [ + 75, + 426, + 468, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Alleviating Forgetting via Global Information", + "text_level": 1, + "bbox": [ + 76, + 662, + 464, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To tackle the issue of catastrophic forgetting in the FCCL, we argue utilizing global information to improve performance. The global information can derive from the global model. We explore this approach and empirically demonstrate that the integration of global information can effectively mitigate catastrophic forgetting.", + "bbox": [ + 75, + 685, + 468, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learn from the Global Model Inspired by LwF [30], we leverage the knowledge from the previously trained global model to the current task via Knowledge Distillation (KD), using only the data from the current task. We conduct experiments on CIFAR100 dataset with 5 continual tasks, where each task contains data of 20 classes. We evaluate the accuracy of the model on each previous task after training on", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1b6e0bfbe75e584dd6f5a8d58ceabdeb600ed022e4c2ebfad2818847ab48772c.jpg", + "table_caption": [ + "Table 1: Test accuracy on previous tasks after training on all tasks. In this context, \"FedAvg\" denotes a naive approach whereby clients learn tasks sequentially." + ], + "table_footnote": [], + "table_body": "
PartitionMethodTask 1Task 2Task 3Task 4Task 5Final
IIDFedAvg0000.0582.616.53
FedLwF6.8 ↑11.5 ↑27.1 ↑44.45 ↑63.2 ↓30.61 ↑
NIID (1)FedAvg000081.6516.33
FedLwF6.65 ↑13.71 ↑29.60 ↑45.41 ↑59.35 ↓30.94 ↑
NIID (0.5)FedAvg0000.1577.3015.49
FedLwF1.6 ↑10.65 ↑27.75 ↑41.3 ↑56.65 ↓27.59 ↑
NIID (0.02)FedAvg0000.159.0011.82
FedLwF0.1 ↑1.7 ↑3.35 ↑28.2 ↑54.0 ↓17.47 ↑
", + "bbox": [ + 503, + 145, + 906, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e28bdbf18d73129be0197101faed2ebb137ecd524954b59c9140dae34e1eae38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 284, + 699, + 349 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b831d8b74780b3c9b36887b41335ac1ed0c8a06b00b2bbadb2079a3033eb7265.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 285, + 875, + 349 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e6b50ad06163db6dcaebff062c27b15c52a8d7263f45bd9804af0b7d99f1ead.jpg", + "image_caption": [ + "Figure 2: Test accuracy by using current task's data, local exemplar and global exemplar." + ], + "image_footnote": [], + "bbox": [ + 514, + 349, + 699, + 422 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/72db22f6aa0020921d7347a583d6aa08ee4aa9f89a707c9dfc872a3af84e567d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 349, + 875, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "all five tasks and report the final accuracy as the average of these results. Table 1 shows that FCCL without knowledge from the global model (i.e., FedAvg [38]) suffers severely from catastrophic forgetting under both IID and non-IID data partitions. Although FedAvg achieves competitive performance on Task 5, it results in $0\\%$ accuracy on previously trained tasks (Task 1, 2, and 3). In contrast, FedLwF achieves better accuracy on Tasks 1, 2, 3, and 4 and final accuracy. The experiments demonstrate that the global model can indeed alleviate the forgetting issue.", + "bbox": [ + 496, + 497, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learn From Global Exemplar In continual learning, an exemplar refers to a sample from a previous task that is stored in a memory buffer for future training. Assuming that data stored in clients are compliant to use for future training, we adopt the idea proposed in iCaRL [44] to save a small proportion of prior training data in memory. The selection of this small proportion of data assumes that the data distribution pertaining the old task is known, but this assumption does not hold in FL because clients are unable to know the data distributions of others due to data privacy concerns. Nevertheless, we employ two exemplar selection methods to illustrate the utility of global data: global exemplar and local exemplar. Global exemplar assumes that the server aggregates a subset of data from clients and distribute these data to clients in training (Note that this method does", + "bbox": [ + 496, + 672, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4785", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/417749987b67f0dcf09c14e375aaadf518cfa018601e6898cb594f773b7544e8.jpg", + "image_caption": [ + "Figure 3: Pipeline of TARGET. We utilize the global model (trained on task $k - 1$ ) to synthesize the data with a global distribution, subsequently employing this synthesized data for training the $k$ -th task." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 477, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "not conform to FL and is only used for comparison). Local exemplar means that each client retains a subset of data from local data in the previous tasks for future training.", + "bbox": [ + 75, + 349, + 468, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2, we show the performance on the current and all previous tasks after learning the current task by using only the new task data, with local exemplar, and with global exemplar. It can be clearly observed that using exemplars can significantly improve the model performance, especially when using global exemplars, which can achieve much higher accuracy than using local exemplars. However, this raises a critical challenge of how to select such exemplars without violating the data privacy.", + "bbox": [ + 75, + 396, + 470, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Our Method: TARGET", + "text_level": 1, + "bbox": [ + 76, + 545, + 302, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 571, + 187, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To utilize the global information without touching on the real exemplars from clients, we present a method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillage), which utilizes global information without storing any real data. A detailed procedure is provided in Algorithm 1. Figure 3 presents an illustration of TARGET, wherein we synthesize data by inverting the global model $\\theta_{k - 1}$ (which was trained on task $k - 1$ ), followed by combining the synthesized data with real data for local model update on task $k$ .", + "bbox": [ + 75, + 594, + 470, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Server Side: Synthesizing Data for Old Tasks", + "text_level": 1, + "bbox": [ + 76, + 755, + 460, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As demonstrated in Figure 2, data with global distributional information is more effective in mitigating the problem of catastrophic forgetting. Therefore, we propose a method of synthesizing data that can model the data distribution of the global model, without the need to preserve any client's privacy data. Specifically, given a global (teacher) model $\\theta_{k-1}$ trained on task $k - 1$ , we first initialize a generator $G$ and a student model $\\theta_S$ . We then repeatedly", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Procedure of TARGET." + ], + "code_body": "Input: $B$ : local minibatch size, $E$ : local epochs, $\\eta$ .. learning rate, synthetic data: $X_{syn} = \\emptyset$ \n1 foreach each task $\\tau = 0,1,\\ldots$ do \n2 Initialize $w_0$ \n3 foreach each round $t = 1,2,\\dots$ do \n4 $S_{t}\\gets$ (random set of m clients) \n5 foreach client $k\\in S_t$ in parallel do \n6 $\\begin{array}{r}\\bigwedge_{t + 1}^{k}\\leftarrow \\text{ClientUpdate}(k,w_{t},\\tau ,X_{syn}) \\end{array}$ \n7 $w_{t + 1}\\gets \\sum_{k = 1}^{m}\\frac{n_k}{n} w_t^k +1$ \n8 $X_{syn} = \\mathsf{DataGeneration}(w_{t + 1})$ \n9 ClientUpdate $(k,w,\\tau ,X_{syn})$ .. \n10 $\\mathcal{B}\\gets$ (split $\\mathcal{P}_k\\cup X_{syn}$ into batches) \n11 set global model $\\mathcal{T}\\gets w$ \n12 foreach each local epoch i from 1 to E do \n13 foreach batch $(b,b_{syn})\\in \\mathcal{B}$ do \n14 $\\ell (w;b) = L_{ce}(w;b)$ if $\\tau \\neq 0$ then \n15 $\\begin{array}{r}\\lfloor \\ell (w;b) + = \\alpha L_{kl}(w,\\mathcal{T};b_{syn})\\rfloor \\end{array}$ \n16 return w to server \n17 DataGeneration(w): \n19 Initialize parameter $\\theta_G,\\theta_S,X_{syn} = \\emptyset$ \n20 foreach round $i = 1,2,\\ldots$ do \n21 Sample noises and labels $\\{\\mathbf{z}_i,\\mathbf{y}_i\\}_{i = 1}^b$ // data generation stage \n22 foreach $j = 1,2,\\ldots$ ,do \n23 Generate $\\{\\hat{\\mathbf{x}}_i\\}_{i = 1}^b$ with $\\{\\mathbf{z}_i\\}_{i = 1}^b$ and $G(\\cdot)$ \n24 Update $\\theta_G$ using Equation 9 \n25 Add a batch of data into $X_{syn}$ \n/ model distillation stage \n26 Update $\\theta_S$ using Equation 10 \n27 return $X_{syn}$", + "bbox": [ + 500, + 109, + 890, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "perform the following two training steps (see line $19\\sim 28$ in Algorithm 1): 1) update the generator by continuously optimizing it to generate data that conforms to the global model distribution; 2) update the student model by distilling knowledge from the teacher model with the synthetic data, hoping that the student model can learn the knowledge of the teacher model sufficiently, which demonstrates the effectiveness of the synthesized data.", + "bbox": [ + 496, + 667, + 890, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Generation First, we utilize $G$ to generate synthetic data from noise $z$ , we need to ensure that the synthetic data $\\hat{x} = G(z)$ is similar to the training dataset. If the synthetic data is similar to the training dataset, their predictions should also be similar. We minimize the cross-entropy (CE) loss on the output of global model $\\theta_{k - 1}(\\hat{x})$ and random la", + "bbox": [ + 496, + 809, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4786", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bels $\\hat{y}$", + "bbox": [ + 76, + 90, + 124, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G} ^ {c e} = C E \\left(\\theta_ {k - 1} (\\hat {x}), \\hat {y}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 104, + 468, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is expected that the synthetic data generated by generator can be classified into a particular class with a high degree of confidence. However, utilizing only the CE loss will cause the generator overfitting to the synthetic data that are far away from the decision boundary (of the global model) [62, 15], thus failing to deliver a good performance. In order to generate samples that are closer to the decision boundary (of the global model) with better transferability, following previous work [62], we introduce a boundary support loss. Additional weight is given to the data on which the global model and the student model diverge in decision making.", + "bbox": [ + 75, + 128, + 468, + 310 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G} ^ {d i v} = - \\omega K L \\left(\\theta_ {k - 1} (\\hat {x}), \\theta_ {S} (\\hat {x})\\right), \\text {a n d} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 321, + 468, + 339 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\omega = \\mathbb {1} (\\arg \\max \\theta_ {k - 1} (\\hat {x}) \\neq \\arg \\max \\theta_ {S} (\\hat {x})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 340, + 468, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $KL$ denotes the Kullback-Leibler (KL) divergence loss, $\\mathbb{1}(a)$ output 1 if $a$ is true and output 0 if $a$ is false. By maximizing the KL divergence loss, the generator can generate more representative data.", + "bbox": [ + 76, + 369, + 468, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motivated by [57], in order to further improve the stability of generator training, we introduce Batch Normalization (BN) loss to make synthetic data conform with the batch normalization statistics.", + "bbox": [ + 76, + 430, + 468, + 489 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G} ^ {b n} = \\sum_ {l} \\left(\\left| \\left| \\mu_ {l} (\\hat {x}) - \\mu_ {l} \\right| \\right| + \\left| \\left| \\sigma_ {l} ^ {2} (\\hat {x}) - \\sigma_ {l} ^ {2} \\right| \\right|\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 500, + 468, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mu_l(\\hat{x})$ and $\\sigma_l^2 (\\hat{x})$ are the batch-wise mean and variance estimate of the $l$ -th BN layer of the generator, $\\mu_{l}$ and $\\sigma_l^2$ are the mean and variance of the $l$ -th BN layer of $f_{S}(\\cdot)$ .", + "bbox": [ + 76, + 542, + 468, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Combining the above losses, we can obtain the loss of the generator as follows,", + "bbox": [ + 76, + 589, + 468, + 619 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G} = \\mathcal {L} _ {G} ^ {c e} + \\lambda_ {1} \\mathcal {L} _ {G} ^ {d i v} + \\lambda_ {2} \\mathcal {L} _ {G} ^ {b n}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 630, + 468, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ is the weight for different loss functions.", + "bbox": [ + 76, + 659, + 459, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Distillation In Equation 7, we introduce a student model to assist in training the generator to produce data with greater diversity. A better student model should lead to a better generator. Therefore, after training the generator for several rounds, we subsequently train the student model using the saved synthesized data and the output of the teacher model, using KL loss for knowledge distillation:", + "bbox": [ + 76, + 694, + 468, + 799 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = K L \\left(\\theta_ {k - 1} (\\hat {x}), \\theta_ {S} (\\hat {x})\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 811, + 468, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this way, we can train a student model with better performance, and then further use it to update the generator. An ideal synthetic dataset should be able to efficiently enable student $\\theta_{S}$ to fully learn the knowledge of teacher model.", + "bbox": [ + 76, + 840, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that when the training of the whole process is over (i.e. the student model can use the synthesized data to obtain high performance), we only retain the synthetic dataset $X_{syn}$ and transfer it to the clients, without saving the generator and student model.", + "bbox": [ + 498, + 90, + 890, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Client Side: Update with Global Information", + "text_level": 1, + "bbox": [ + 498, + 176, + 880, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the client side, we can obtain the data synthesized for the previous task $X_{syn}$ and the real training data of the current task $X_{local}$ , then we train the local model $\\theta_k$ for task $k$ on the two datasets at the same time. We showed in Section 3.3 that the use of global models and global data can alleviate forgetting. Thus we distill the knowledge of global teacher model and global synthetic data by minimizing the following objective function,", + "bbox": [ + 496, + 200, + 890, + 321 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c l i e n t}} = \\underbrace {C E \\left(\\theta_ {k} (x) , y\\right)} _ {\\text {f o r c u r r e n t t a k}} + \\alpha \\cdot \\underbrace {K L \\left(\\theta_ {k - 1} (\\hat {x}) , \\theta_ {k} (\\hat {x})\\right)} _ {\\text {f o r p r e v i o u s t a k s}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 334, + 890, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $(x,y)\\in X_{local}$ and $(\\hat{x})\\in X_{syn}$ . The utilization of the distillation loss facilitates efficient transfer of knowledge from the previous task to the current task model. And $\\alpha$ is a hyper-parameter that controls the strength of the regularization for the previous tasks.", + "bbox": [ + 496, + 382, + 890, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 473, + 632, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 498, + 705, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We experiment on two datasets, namely CIFAR-100 [25], and Tiny-ImageNet [26], to evaluate the performance of our proposed approach. To establish the order of the continual tasks, we adopt the widely used protocols [49, 61, 42, 7]. Specifically, we divide all classes of each dataset equally into multiple tasks by default, i.e. we evenly divide the classes into 5 and 10 tasks to simulate class continual learning scenarios. We employ ResNet18 [18] as the backbone for the classification model.", + "bbox": [ + 496, + 522, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate our approach, we employ the standard continual learning metrics, as used in prior works [23, 22, 47, 16], which include average accuracy across all tasks and a forgetting measure [6] (see Equation 2). For a fair comparison with the baseline class continual learning methods in the FCCL setting, we implement three types of baselines: 1) Finetune, in which each client simply learns tasks in sequence; 2) FedWeIT [58], a regularization-based method in Federated Continual Learning that maximizes the knowledge transfer between clients; 3) Examples of typical continual learning methods that do not store training data for rehearsal, including EWC [27] and LwF [30]. In addition, we compare our method with methods that store real training data of old tasks, such as iCaRL [44]. We implement these traditional continual learning algorithms in the FCCL scenario and name them as FedEWC, FedLwF, and", + "bbox": [ + 496, + 659, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4787", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/935cf74fd0661cff4879794d93af57e1596daecf6b26950ce13e13a43717dd13.jpg", + "table_caption": [ + "Table 2: The Average Accuracy (%) and Forgetting for all learned tasks on CIFAR-100 for various numbers of tasks (5, 10) under both IID and non-IID settings. Results are reported as an average of 3 runs. 'Acc' refers to average accuracy, and 'F' represents the forgetting measure utilized in Equation 2. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
Data partitionIIDNIID (1)NIID (0.5)
TasksT=5T=10T=5T=10T=5T=10
MethodAcc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)
Finetune16.120.787.830.7516.330.778.450.7415.490.747.640.71
FedEWC16.510.718.010.6516.060.688.840.6216.860.668.040.65
FedWeIT28.450.5220.390.4328.560.4919.680.4524.570.5415.450.48
FedLwF30.610.4523.270.3730.940.4221.160.4127.590.4417.980.45
Ours36.310.2224.760.2634.890.2422.850.2633.330.2720.710.29
", + "bbox": [ + 78, + 146, + 893, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c12a0c1865af8eaa726df1e79316f0b03052ac6b238e65dcf404865566e323de.jpg", + "image_caption": [ + "Figure 4: Average accuracy on previous tasks and current task after the model trained on current task." + ], + "image_footnote": [], + "bbox": [ + 84, + 345, + 277, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa93cc02ca34ab091c357fd66b8a5728e8cdeeebc701e9103454e3f65048d25f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 345, + 470, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FedIcaRL. For detailed information on the task configuration, default hyper-parameters and additional experimental results, please refer to the Appendix.", + "bbox": [ + 75, + 550, + 468, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experiments on CIFAR-100", + "text_level": 1, + "bbox": [ + 76, + 604, + 328, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For CIFAR-100 dataset, we conduct experiments on two sets of tasks consisting of 5 and 10 tasks, respectively. We run the experiment on both IID and non-IID scenarios. For non-IID setting, the Dirichlet parameter is set to 0.5 and 1, i.e. $\\mathrm{NIID}(0.5)$ and $\\mathrm{NIID}(1)$ . Table 2 shows the final average accuracy of the FL model trained on all tasks, along with the corresponding forgetting measure for each experiment. It is important to emphasize that an optimal method is characterized by high average accuracy and low forgetting measure.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 indicates that Finetune exhibits the poorest performance when attempting to learn the continuously incoming tasks sequentially, thereby experiencing the issue of catastrophic forgetting. Moreover, we observed that methods based on regularization constraints, such as FedEWC and FedWeIT, were often ineffective in preventing the model from forgetting old tasks due to the lack of available data. On the other hand, we found that distillation-based approaches such as FedLwF and our proposed method were", + "bbox": [ + 75, + 765, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "capable of improving the final average accuracy while simultaneously mitigating the issue of catastrophic forgetting. This is due to the transfer of model knowledge learned from the old task to the new task in a distillation manner, enabling the model to prevent catastrophic forgetting. Obviously, we found that applying our proposed method to generate synthetic datasets for the federated models trained on old tasks and subsequently performing model distillation on these synthetic datasets can lead to a substantial improvement in the average accuracy and reduction in the forgetting measure. This observation underscores the ability of our synthetic data to capture the distribution characteristics of historical task data accurately. For example, in IID setting, when partitioning the CIFAR-100 dataset into five tasks, our method achieves an accuracy of $36.31\\%$ , which is about $6\\%$ higher than the best baseline method FedLwF. It is worth noting that we observed a decrease in the average accuracy of all methods to varying degrees as the number of tasks increased from 5 tasks to 10 tasks due to increased task complexity and the associated forgetting phenomenon. Nonetheless, our proposed method maintains the highest average accuracy and the lowest forgetting measure even under these more demanding conditions.", + "bbox": [ + 496, + 345, + 893, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 illustrates the performance of the models trained by various methods on all previously learned tasks after the completion of each task. Specifically, it shows the average accuracy of the model on both the current and previous tasks after the completion of each task (e.g., after learning the second task, the average accuracy of the model on both the first and second tasks is measured). Based on these curves, it is evident that our proposed model outperforms other competing baseline methods in all incremental tasks, regardless of the number of tasks involved. This finding underscores the effectiveness of our approach in facilitating multiple local clients to learn new classes in a streaming manner while mitigating the forgetting problem.", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4788", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67f20052159e796ecaa4a9ff6da8e20226eb69711fddcd68d8ab91057659124b.jpg", + "image_caption": [ + "Figure 5: The average accuracy $(\\%)$ and Forgetting for all learned tasks on Tiny-ImageNet for 5 tasks under both IID and non-IID settings." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 472, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Experiments on Tiny-ImageNet", + "text_level": 1, + "bbox": [ + 76, + 310, + 357, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also evaluated the performance of our proposed method on the more challenging Tiny-ImageNet dataset and obtained similar results to those observed in the CIFAR-100 experiments. Specifically, in Figure 5, we present the final average accuracy and forgetting measure for all learned tasks in both IID and non-IID settings for the case of 5 tasks. The Dirichlet parameter is set to $\\{0.05, 0.1, 0.5, 1\\}$ . Based on Figure 5, it is evident that our proposed method consistently outperforms FedLwF in terms of average accuracy across all data partitions. Moreover, our method demonstrates a significantly lower forgetting measure than FedLwF under both IID and non-IID settings. Our proposed method achieves an average accuracy that is approximately $3\\%$ higher than that of FedLwF even in the most challenging scenario (i.e. NIID(0.05)). This result highlights the effectiveness of our method in mitigating catastrophic forgetting in the presence of extreme data distributions.", + "bbox": [ + 75, + 334, + 468, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Comparison with Exemplar-based Method", + "text_level": 1, + "bbox": [ + 76, + 604, + 441, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In CL, the most successful approaches to alleviate forgetting require extensive replay of previously seen data, which can be problematic when data legality and privacy concerns exist. Among them, iCaRL [44] is a classic but unrealistic algorithm that relies on the stored exemplars in addition to the network parameters, and it is intuitive that using old task's real data could be beneficial to alleviate forgetting problem. To further understand the performance gap, we compare our proposed TARGET (which uses synthetic data) with iCaRL (which requires storing exemplars from old task's real data), and study the effects of different exemplar memories on the performance of our method and iCaRL in Figure 6. We set the stored exemplar size to $\\{1000, 1500, 2000\\}$ for iCaRL, and $\\{2000, 3000\\}$ for our method on the CIFAR-100 dataset.", + "bbox": [ + 75, + 627, + 468, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The accuracy curve in Figure 6 represents the average accuracy rate measured over all 100 classes learned by the model during the last learning task. While our method", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b360c9d98595787fb5e505121748611600bc34d21883b00e2424542a0a2795f3.jpg", + "image_caption": [ + "Figure 6: Accuracy comparison between real data-based method (iCaRL) and synthetic data-based method (TARGET). We show the result for the last task." + ], + "image_footnote": [], + "bbox": [ + 526, + 97, + 875, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8f9be0c87a631f5edf61211ed2373f74408cb7851152fe0199c4cefef9678c37.jpg", + "table_caption": [ + "Table 3: The effect of $\\alpha$ on the performance of both new and old tasks in CIFAR-100, 2 tasks." + ], + "table_footnote": [], + "table_body": "
α35101525
Old Task (0-49)28.3434.9149.0258.4866.02
New Task (50-99)62.4660.2653.7445.2629.32
Average (0-99)45.4147.5851.3851.8747.76
", + "bbox": [ + 504, + 363, + 887, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "can achieve similar performance to storing 1k real training data by storing 2k synthetic data, it still cannot outperform storing 2k real training data. However, it is worth noting that our method does not require storing any real training data, which can be a significant advantage in scenarios where storing real data is difficult or not allowed due to privacy or legal concerns. Additionally, our method achieves better performance than storing only 1k real training data, which indicates that our synthetic data is effective in mitigating catastrophic forgetting. We observed that when our method stores 3k synthetic data, it achieves better accuracy than when it stores 2k synthetic data. However, surpassing iCaRL in performance with an equal amount of data remains a challenge for our method. How to effectively use fewer synthetic data with more valuable knowledge from previous tasks will be left as a future research direction.", + "bbox": [ + 496, + 470, + 890, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Analysis of Our Method", + "text_level": 1, + "bbox": [ + 500, + 724, + 720, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Trade-off between Backward and Forward Transfer. Continual learning presents a challenge in balancing the trade-off between maintaining high accuracy on old tasks (backwards transfer) and achieving high accuracy on new tasks (forward transfer). In Table 3, we evenly split the CIFAR-100 dataset into two tasks and test the average accuracy of our proposed method under different values of $\\alpha$ . We observe that the trade-off between backward and forward transfer is not always balanced. In order to achieve good backward transfer, a large value of $\\alpha$ should be used", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4789", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eed01bfff92f3aafd5332161b6ad73413ffa3b58b50464ebaa77e341b7be11a9.jpg", + "image_caption": [ + "Figure 7: Accuracy for different size of synthetic dataset." + ], + "image_footnote": [], + "bbox": [ + 86, + 90, + 269, + 229 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b7529b6618ef3d10da9af403483a8a9be090a985a8d10ab89da2872293ab0394.jpg", + "image_caption": [ + "Figure 9: Distillation results for task $\\{1,2,3,4\\}$ when trained on 5 tasks." + ], + "image_footnote": [], + "bbox": [ + 279, + 90, + 452, + 229 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4843a71ba65577c05f88d66f719d9b12961e4fcfea09c2409ecd829015e481d3.jpg", + "image_caption": [ + "Figure 8: Visualization of randomly synthesised data." + ], + "image_footnote": [], + "bbox": [ + 86, + 260, + 460, + 406 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "to prevent the model from forgetting previous tasks and to encourage it to focus more on the old synthetic data. Conversely, to achieve good forward transfer, a small value of $\\alpha$ should be used, allowing the model to learn quickly and effectively from new tasks while still maintaining some knowledge from the old tasks. The experimental results in Table 3 indicate that when the value of $\\alpha$ is between 10-15, the model achieves a good balance between accuracy on new tasks and accuracy on old tasks. We also partition the CIFAR-100 dataset into 5 equal tasks and test our method's performance on all previous tasks after learning each new task (refer to Appendix).", + "bbox": [ + 75, + 444, + 472, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Memory of Synthetic Data. The optimal amount of synthetic data generated for the old task plays a crucial role in determining the final performance of our method. Insufficient synthetic data may not adequately facilitate knowledge transfer from the old tasks, while excessive data generation can lead to increased memory and communication costs. Therefore, determining an appropriate amount of synthetic data that strikes a balance between knowledge transfer and computational cost is critical for the effectiveness and efficiency of our approach. As shown in Figure 7, we test our method on CIFAR-100, which is divided into 5 tasks with different data sizes ranging from 2k to 16k. It can be observed that when the data volume is relatively small, such as 2k and 4k, the performance of the model is poor, especially when the data volume is 2k, the testing curve in task 4 shows a later decline. This is because when the data volume is too small, the model is unable to effectively learn", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/21927968055e9a469543b13752cbbcf6da33269afa682eda7d4b04182252cd35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 92, + 696, + 199 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/decc23f26b9566ccb4f5e2360c30d75fb876661377008b90c6e87f3314be8533.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 92, + 890, + 199 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "knowledge from old tasks. Increasing the data volume to 8k can effectively alleviate the forgetting phenomenon and achieve good performance. However, continuously increasing the data volume to 12k and 16k do not result in significant improvement in the model's performance. It is important to note that the size of the data volume alone does not guarantee the effectiveness of synthetic data in improving machine learning models. The quality and relevance of the synthetic data must also be carefully considered to ensure that it accurately represents the underlying distribution of the real-world data.", + "bbox": [ + 496, + 263, + 893, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visualization on Synthetic Data. To demonstrate the effectiveness of the synthetic data, we present in Figure 8 the visualization results of the synthesized images generated by our method after the model learns the penultimate task on Tiny-ImageNet (CIFAR-100 in Appendix) for 5 tasks. These data have the potential to efficiently enable the student model to approach the performance of the teacher model rapidly. However, it is worth noting that these data exhibit visual dissimilarities from the actual training data.", + "bbox": [ + 496, + 449, + 893, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Analysis on Distillation. Based on the synthetic data, we distill the knowledge of the model trained on old tasks into a student model. The effect of distillation is further demonstrated in Figure 9. It can be observed that even without accessing any private training data from clients, our method can quickly distill the student model on the server to approach the performance of the global model.", + "bbox": [ + 496, + 604, + 893, + 710 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 724, + 619, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In conclusion, this paper introduces a novel method, TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillage), to alleviate the catastrophic forgetting problem in Federated Class-Continual Learning (FCCL). Unlike all the previous methods, our proposed method leverages global knowledge, without requiring any additional datasets or data from previous tasks, making it ideal for privacy-sensitive scenarios. Extensive experimental results demonstrate the effectiveness of our proposed method in comparison to existing FCCL methods.", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4790", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eden Belouadah and Adrian Popescu. Il2m: Class incremental learning with dual memory. In Proceedings of the IEEE/CVF international conference on computer vision, pages 583-592, 2019. 1", + "[2] Cecilia Bjursell. The pandemic as disjuncture: Lifelong learning in a context of fear. International Review of Education, 66(5-6):673-689, 2020. 1", + "[3] Thang D Bui, Cuong V Nguyen, Siddharth Swaroop, and Richard E Turner. Partitioned variational inference: A unified framework encompassing federated and continual learning. arXiv preprint arXiv:1811.11206, 2018. 1", + "[4] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio P. Calmon, and Taesup Moon. Cpr: classifier-projection regularization for continual learning. arXiv preprint arXiv:2006.07326, 2020. 3", + "[5] Yatin Chaudhary, Pranav Rai, Matthias Schubert, Hinrich Schütze, and Pankaj Gupta. Federated continual learning for text classification via selective inter-client transfer. arXiv preprint arXiv:2210.06101, 2022. 2", + "[6] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532-547, 2018. 3, 6", + "[7] Arslan Chaudhry, Naemullah Khan, Puneet Dokania, and Philip Torr. Continual learning in low-rank orthogonal subspaces. Advances in Neural Information Processing Systems, 33:9900-9911, 2020. 3, 6", + "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2, 3", + "[9] Marco Ciotti, Massimo Ciccozzi, Alessandro Terrinoni, Wen-Can Jiang, Cheng-Bin Wang, and Sergio Bernardini. The COVID-19 pandemic. Critical reviews in clinical laboratory sciences, 57(6):365-388, 2020. 1", + "[10] Jiahua Dong, Yang Cong, Gan Sun, Zhen Fang, and Zhengming Ding. Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1(2):1-18, 2021. 2", + "[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1", + "[12] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 2", + "[13] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 86-102. Springer, 2020. 1" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Fernando Estevez Casado, Dylan Lema Pais, Marcos Fernández Criado, Roberto Iglesias Rodríguez, Carlos Vázquez Regueiro, and Senen Barro Ameneiro. Concept drift detection and adaptation for federated and continual learning. 2021. 2", + "[15] Gongfan Fang, Kanya Mo, Xinchao Wang, Jie Song, Shitao Bei, Haofei Zhang, and Mingli Song. Up to 100x faster data-free knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 6597-6604, 2022. 6", + "[16] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013. 1, 3, 6", + "[17] Andrew Hard, Kanishka Rao, Rajiv Mathews, Swaroop Ramaswamy, Françoise Beaufays, Sean Augenstein, Hubert Eichner, Chloe Kiddon, and Daniel Ramage. Federated learning for mobile keyboard prediction. arXiv preprint arXiv:1811.03604, 2018. 1", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[19] Ziyue Jiang, Yi Ren, Ming Lei, and Zhou Zhao. Fed-speech: Federated text-to-speech with continual learning. arXiv preprint arXiv:2110.07216, 2021. 1", + "[20] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and Trends® in Machine Learning, 14(1-2):1-210, 2021. 1", + "[21] Georgios Kaissis, Alexander Ziller, Jonathan Passerat-Palmbach, Theo Ryffel, Dmitrii Usynin, Andrew Trask, Ionésio Lima Jr, Jason Mancuso, Friederike Jungmann, Marc-Matthias Steinborn, et al. End-to-end privacy preserving deep learning on multi-institutional medical imaging. Nature Machine Intelligence, 3(6):473-484, 2021. 2", + "[22] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 1, 3, 6", + "[23] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 1, 2, 3, 6", + "[24] Jakub Konečny, Brendan McMahan, and Daniel Ramage. Federated optimization: Distributed optimization beyond the datacenter. arXiv preprint arXiv:1511.03575, 2015. 2", + "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 3, 6", + "[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6", + "[27] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forget-" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4791", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ting by incremental moment matching. Advances in neural information processing systems, 30, 2017. 2, 6", + "[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1, 4", + "[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. arXiv preprint arXiv:1812.06127, 2018. 3, 4", + "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 4, 6", + "[31] Zexi Li, Qunwei Li, Yi Zhou, Wenliang Zhong, Guannan Zhang, and Chao Wu. Edge-cloud collaborative learning with federated and centralized features. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '23, page 1949-1953, New York, NY, USA, 2023. Association for Computing Machinery. 2", + "[32] Zexi Li, Tao Lin, Xinyi Shang, and Chao Wu. Revisiting weighted aggregation in federated learning with neural networks. In International Conference on Machine Learning. PMLR, 2023. 2", + "[33] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Rmm: Reinforced memory management for class-incremental learning. Advances in Neural Information Processing Systems, 34:3478-3490, 2021. 1", + "[34] Yuang Liu, Wei Zhang, Jun Wang, and Jianyong Wang. Data-free knowledge transfer: A survey. arXiv preprint arXiv:2112.15278, 2021. 2", + "[35] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2", + "[36] Yuhang Ma, Zhongle Xie, Jue Wang, Ke Chen, and Lidan Shou. Continual federated learning based on knowledge distillation. In Lud De Raedt, editor, Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 2182–2188. International Joint Conferences on Artificial Intelligence Organization, 7 2022. Main Track. 1", + "[37] Yuhang Ma, Zhongle Xie, Jue Wang, Ke Chen, and Lidan Shou. Continual federated learning based on knowledge distillation. In Luc De Raedt, editor, Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, pages 2182-2188. ijcai.org, 2022. 2", + "[38] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Aarti Singh and Jerry Zhu, editors, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, volume 54 of Proceedings of Machine Learning Research, pages 1273-1282. PMLR, 20-22 Apr 2017. 1, 4", + "[39] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data." + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 2", + "[40] Sudhanshu Mittal, Silvio Galesso, and Thomas Brox. Essentials for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3513-3522, 2021. 1", + "[41] Alexander Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. 2015. 3", + "[42] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural networks, 113:54-71, 2019. 1, 6", + "[43] Tae Jin Park, Kenichi Kumatani, and Dimitrios Dimitriadis. Tackling dynamics in federated incremental learning with variational embedding rehearsal. arXiv preprint arXiv:2110.09695, 2021. 2", + "[44] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 2, 4, 6, 8", + "[45] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. arXiv preprint arXiv:1810.11910, 2018. 2", + "[46] Amanda Rios and Laurent Itti. Closed-loop gan for continual learning. arXiv preprint arXiv:1811.01146, 2018. 2", + "[47] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 1, 3, 6", + "[48] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International Conference on Machine Learning, pages 4548-4557. PMLR, 2018. 1", + "[49] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 1, 2, 6", + "[50] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9374-9384, 2021. 2, 3", + "[51] Sebastian Thrun. Lifelong learning algorithms. Learning to learn, 8:181-209, 1998. 1", + "[52] Anastasiia Usmanova, François Portet, Philippe Lalanda, and German Vega. A distillation-based approach integrating continual learning and federated learning for pervasive services. arXiv preprint arXiv:2109.04197, 2021. 2", + "[53] Anamaria Vizitiu, Cosmin Loan Nita, Andrei Puiu, Constantin Suciu, and Lucian Mihai Itu. Towards privacy-preserving deep learning based medical imaging applications. In 2019 IEEE international symposium on medical measurements and applications (MeMeA), pages 1-6. IEEE, 2019. 2", + "[54] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In Computer Vision-ECCV 2022: 17th" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4792", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV, pages 398-414. Springer, 2022. 1", + "[55] Ju Xu and Zhanxing Zhu. Reinforced continual learning. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[56] Li Yang, Shasha Liu, Jinyan Liu, Zhixin Zhang, Xiaochun Wan, Bo Huang, Youhai Chen, and Yi Zhang. Covid-19: immunopathogenesis and immunotherapeutics. Signal transduction and targeted therapy, 5(1):128, 2020. 1", + "[57] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715-8724, 2020. 3, 6", + "[58] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086. PMLR, 2021. 2, 6", + "[59] Jaehong Yoon, Saehoon Kim, Eunho Yang, and Sung Ju Hwang. Scalable and order-robust continual learning with additive parameter decomposition. arXiv preprint arXiv:1902.09432, 2019. 3", + "[60] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 2", + "[61] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987-3995. PMLR, 2017. 1, 2, 6", + "[62] Jie Zhang, Chen Chen, Bo Li, Lingjuan Lyu, Shuang Wu, Shouhong Ding, Chunhua Shen, and Chao Wu. Dense: Data-free one-shot federated learning. In Advances in Neural Information Processing Systems. 6", + "[63] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In International Conference on Machine Learning, pages 26311-26329. PMLR, 2022. 3", + "[64] Junting Zhang, Jie Zhang, Shalini Ghosh, Dawei Li, Serafettin Tasci, Larry Heck, Heming Zhang, and C-C Jay Kuo. Class-incremental learning via deep model consolidation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1131-1140, 2020. 1", + "[65] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13208-13217, 2020. 1", + "[66] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1", + "[67] Didi Zhu, Yinchuan Li, Yunfeng Shao, Jianye Hao, Fei Wu, Kun Kuang, Jun Xiao, and Chao Wu. Generalized universal domain adaptation with generative flow networks. arXiv preprint arXiv:2305.04466, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[68] Didi Zhu, Yincuan Li, Junkun Yuan, Zexi Li, Yunfeng Shao, Kun Kuang, and Chao Wu. Universal domain adaptation via compressive attention matching. arXiv preprint arXiv:2304.11862, 2023. 2", + "bbox": [ + 501, + 90, + 890, + 146 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4793", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/ff5641a5-17a3-470a-a519-ccac898db108_model.json b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/ff5641a5-17a3-470a-a519-ccac898db108_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e9e3794b467e269994ae6a7987957625e5b734eb --- /dev/null +++ b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/ff5641a5-17a3-470a-a519-ccac898db108_model.json @@ -0,0 +1,2611 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.882, + 0.153 + ], + "angle": 0, + "content": "TARGET: Federated Class-Continual Learning via Exemplar-Free Distillation" + }, + { + "type": "text", + "bbox": [ + 0.36, + 0.182, + 0.612, + 0.234 + ], + "angle": 0, + "content": "Jie Zhang* \nETH Zurich \nzj.jayzhang@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.259, + 0.684, + 0.295 + ], + "angle": 0, + "content": "Chen Chen & Weiming Zhuang & Lingjuan Lyu† \nSony AI" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.296, + 0.778, + 0.314 + ], + "angle": 0, + "content": "{ChenA.Chen, weiming.zhuang, Lingjuan.Lv}@sony.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.347, + 0.314, + 0.364 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.474, + 0.668 + ], + "angle": 0, + "content": "This paper focuses on an under-explored yet important problem: Federated Class-Continual Learning (FCCL), where new classes are dynamically added in federated learning. Existing FCCL works suffer from various limitations, such as requiring additional datasets or storing the private data from previous tasks. In response, we first demonstrate that non-IID data exacerbates catastrophic forgetting issue in FL. Then we propose a novel method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free distTillage), which alleviates catastrophic forgetting in FCCL while preserving client data privacy. Our proposed method leverages the previously trained global model to transfer knowledge of old tasks to the current task at the model level. Moreover, a generator is trained to produce synthetic data to simulate the global distribution of data on each client at the data level. Compared to previous FCCL methods, TARGET does not require any additional datasets or storing real data from previous tasks, which makes it ideal for data-sensitive scenarios." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.699, + 0.21, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.47, + 0.816 + ], + "angle": 0, + "content": "Federated Learning (FL) is a privacy-aware learning paradigm that facilitates collaborations among multiple entities (e.g., edge devices or organizations) [38, 20, 28, 66, 17]. Each entity or client in FL retains data locally and transfers only training updates to the central server for aggregation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.817, + 0.47, + 0.862 + ], + "angle": 0, + "content": "Conventional FL studies assume that the data classes and domains are static, but the new classes could emerge and data domains could change over time in reality [64," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.349, + 0.895, + 0.514 + ], + "angle": 0, + "content": "40, 1, 65]. For example, multiple health institutions could use FL to collaborate and train models to identify COVID-19 [56, 9, 2] strains; new COVID-19 strains, however, continue to emerge due to high mutation rate of virus. An intuitive solution to this issue of continuously emerging data classes is training new models from scratch, but this is impractical as it would require significant extra computation cost. Another method is transfer learning from the previously trained model, but this method suffers from catastrophic forgetting [23, 22, 47, 16], degrading performance on the previous classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.895, + 0.67 + ], + "angle": 0, + "content": "To address these issues, recent research [36, 11, 3, 19] has introduced the concept of Continual Learning (CL) [61, 42, 51, 49, 48] within the FL framework. These methods, collectively referred to as Federated Continual Learning (FCL), aim to mitigate the problems of catastrophic forgetting in FL. In most FCL scenarios, new classes are dynamically added, which we call Federated Class-Continual Learning (FCCL). FCCL allows local clients to continuously collect new data, and new classes can be added at any time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Unfortunately, existing FCCL works suffer from various limitations. For example, Ma et al. [36] utilize an unlabeled surrogate dataset to address the catastrophic forgetting problem, which may be difficult to obtain in some data sensitive scenarios. Furthermore, the usage of an unlabeled surrogate dataset may not be ideal for certain types of data, as it may not capture the full complexity of the original data. In CL, exemplar-based methods [44, 33, 54, 13] have achieved leading performance. An exemplar refers to a sample or instance of a previously seen data point that is retained in a memory buffer for future use in the learning process. Dong et al. [11] propose a exemplar-based method that stores historical data to address catastrophic forgetting. However, in many privacy-sensitive scenarios (e.g., hospitals and medical research institutions), users are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.318, + 0.888 + ], + "angle": 0, + "content": "*Work done during internship at Sony AI." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.318, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4782" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "not permitted to store data from previous tasks due to privacy and policy concerns and data will not be kept for a long time [21, 53, 34, 50]. In summary, the majority of FCCL methods train the global model with additional datasets or previous task data, which could potentially violate data privacy regulations. This dilemma prompts us to consider the following question:" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.199, + 0.471, + 0.244 + ], + "angle": 0, + "content": "Question: How to effectively alleviate the catastrophic forgetting problem in the FCCL without storing the local private data of the client or any additional datasets?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.47, + 0.441 + ], + "angle": 0, + "content": "To address this question, we conduct a systematic analysis and observe that the imbalanced distribution of data among clients in FL exacerbates the catastrophic forgetting problem (see Section 3.3). In order to fix this problem: 1) at the model level, we leverage the previously trained global model to transfer knowledge of the old tasks to the current task. 2) at the data level, we train a generator to produce synthetic data that aims to simulate the global distribution of data on each client. Drawing on these insights, we present a method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillation) that mitigates catastrophic forgetting in FCCL without compromising clients' data privacy." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.442, + 0.408, + 0.456 + ], + "angle": 0, + "content": "Our contributions can be concluded as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.458, + 0.469, + 0.546 + ], + "angle": 0, + "content": "- We are the first to demonstrate that non-independent and identically distributed (non-IID) data exacerbates catastrophic forgetting issue in FL. Then we propose a novel method called TARGET, which alleviates the catastrophic forgetting in FCCL by leveraging global information." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.548, + 0.468, + 0.593 + ], + "angle": 0, + "content": "- Compared to previous FCCL methods, TARGET doesn't require extra datasets or data from previous tasks, it can be applied in data sensitive scenarios." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.594, + 0.468, + 0.668 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the efficacy of our proposed method. For example, when partitioning the CIFAR-100 dataset into five tasks, our method achieves an accuracy of \\(36.31\\%\\), which is about \\(6\\%\\) higher than the best baseline method." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.458, + 0.469, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.685, + 0.219, + 0.701 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.348, + 0.727 + ], + "angle": 0, + "content": "2.1. Federated Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In recent years, deep neural networks (DNNs) have been widely used as a fundamental technology for the development of artificial intelligence, both in established and emerging fields [68, 67, 32, 31, 10]. With the advancement of deep learning, an increasing number of researchers have started to focus on the construction of privacy-preserving deep learning frameworks. FL is a paradigm for collaboratively building a model across multiple clients [24, 39], which is gaining momentum in recent years. However, Federated Continual Learning (FCL), which focused on both FL and CL simultaneously is just emerging and still" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.515 + ], + "angle": 0, + "content": "remains pending further research [52, 43, 58]. Apart from the client-wise catastrophic forgetting, FCL paradigm also poses new challenges such as inter-client interference and communication-efficiency [58]. FedWeIT solves these challenges through decomposing the parameters into three parts, i.e., global parameters, local based parameters, and task-adaptive parameters [58]. Concurrently, Concept-Drift-Aware Federated Averaging (CDA-FedAvg) extends the popular FL algorithm, Federated Averaging (FedAvg), to tackle the CL problem by introducing the concept drift detection and adaptation [14]. FCL with Distillation (CFeD) treats the model trained with the last task as the teacher model to perform a knowledge distillation and proposes a server distillation mechanism to deal with non-i.i.d. issue [37]. Global-Local Forgetting Compensation (GLFC) designs a class-aware gradient compensation loss and a class-semantic relation distillation loss to prevent forgetting, and a proxy server to mitigate the non-independent and identically distributed (non-IID) problem [12]. Additionally, Federated Selective Inter-client Transfer (FedSeIT) applies FCL to NLP through selectively combining model parameters of foreign clients and selecting informative tasks to perform knowledge transfer [5]. These papers illustrate the growing interest in FCL and the need for novel approaches to address its unique challenges. As research in this area continues to advance, we can expect to see more innovative approaches to FCL that further improve its performance, scalability, and privacy preservation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.691, + 0.545 + ], + "angle": 0, + "content": "2.2. Continual Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Continual Learning has been studied extensively, several training methods have been proposed to address the catastrophic forgetting challenge it presents. Regularization-based approaches: Elastic Weight Consolidation (EWC) selectively penalizes the network parameters that are important for old tasks [23]. Synaptic intelligence (SI) uses a memory buffer to store important network parameters [61]. Incremental moment matching (IMM) modelings the posterior distribution after learning multiple tasks as a mixture of Gaussian models [27]. Stable SGD proposes to carefully design the training regimes such as learning rate decay, batch size, dropout, and optimizer to alleviate forgetting. Replay-based approaches: The methods in this family construct a memory to store the past information which will be presented to the model for reviewing in future tasks. Some store the knowledge of the previous tasks, known as experience replay [35, 45, 49, 46]. iCaRL preserves the most representative samples of each class [44]. Averaged Gradient Episodic Memory (A-GEM) builds an episodic memory of model parameter gradients[8]. Architecture-based: DEN expands the model size [60] and RCL utilizes reinforcement learning [55]. APD divides the model parameters into shared and task-specific parameters to restrict the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.236, + 0.107 + ], + "angle": 0, + "content": "model complexity [59]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.127, + 0.473, + 0.309 + ], + "angle": 0, + "content": "Exemplar-Free Continual Learning Besides, a promising line of work focuses on data-free Continual Learning. DeepDream perturbs current training samples into images that maximize \"forgetting\" from the previous tasks [41]. DeepInversion proposes a model inversion technique and evaluates its performance in a Continual Learning scenario but found limited success [57]. DFCIL [50] investigates the failure and decomposes the CE-loss into two different losses which guarantee the learning of effective features. These methods enable Continual Learning in scenarios where storing old task data is not feasible or desirable, such as in privacy-sensitive applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.323, + 0.383, + 0.34 + ], + "angle": 0, + "content": "3. Catastrophic Forgetting in FCCL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.425 + ], + "angle": 0, + "content": "This section conducts an in-depth analysis of catastrophic forgetting problem in federated class-continual learning (FCCL). We start by providing a formal definition of the problem. Then, we investigate the forgetting issue in FCCL and discuss potential methods to mitigate it." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.434, + 0.262, + 0.45 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.473, + 0.716 + ], + "angle": 0, + "content": "Federated Class-Continual Learning (FCCL) focuses on the problem of learning models for new classes over time in FL. An FCCL framework consists of a central server and multiple clients. All clients do not share their raw data with any other client or the central server. Each client learns from a sequence of \\( n \\) tasks, where \\( k \\)-th task contains non-overlapping subsets of classes \\( C_k \\in C \\), where \\( C \\) is the set of all possible classes. In our privacy-aware scenario, the task stream is presented in an unknown order, and each client can only access its local data from task \\( k \\) during that task's training period, which is no longer accessible thereafter. Note that the models are trained in a distributed manner, where each party has access to only a subset of the classes \\( C_k \\) (i.e. non-IID). We also consider a more challenging and practical setting where the data in each client is heterogeneous. In this paper, we assume the label distribution of data in each client is skewed [29, 63]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Forgetting Issue in FCCL The goal of the global model optimization problem at task \\( k \\) is to minimize the overall classification error on the current set of classes \\( C_k \\). However, when a new task arises, clients are not able to access data from previous (old) tasks due to privacy concerns and can only update their local model with data from the new task. This often leads to a significant decrease in performance on previous tasks, which is known as catastrophic forgetting [23, 22, 47, 16]. To mitigate catastrophic forgetting in the global model, we aim to minimize the overall classification error on the current set of classes \\( C_k \\), while" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.099, + 0.696, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.248, + 0.657, + 0.26 + ], + "angle": 0, + "content": "(a) Test Accuracy" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.099, + 0.877, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.248, + 0.848, + 0.26 + ], + "angle": 0, + "content": "(b) Relative Forgetting" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.276, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Figure 1: Forgetting under different data partitions. We partitioned the CIFAR-100 dataset into five tasks with 20 classes each, distributed among five clients. \"NIID\" refers to non-IID, where a lower value represents a more imbalanced or skewed distribution of data. Non-IIDness refers to the Dirichlet parameter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.893, + 0.437 + ], + "angle": 0, + "content": "simultaneously minimizing the changes to the previously learned classes. Formally, the objective function can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.443, + 0.892, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {k}} \\sum_ {c \\in C _ {k}} \\sum_ {i = 1} ^ {m _ {c}} L \\left(f _ {k} \\left(x _ {i, c}; \\theta_ {k}\\right), c\\right) + \\alpha R \\left(\\theta_ {k}, \\theta_ {k - 1}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.674 + ], + "angle": 0, + "content": "where \\(\\theta_{k}\\) is the model parameter at round \\(k\\), \\(L\\) is a loss function that measures the classification error, \\(R\\) is a regularization term that penalizes changes to the previous model parameters, \\(m_{c}\\) is the number of data in class \\(c\\), and \\(\\alpha\\) is a hyper-parameter that controls the strength of the regularization. In this formula, \\(f_{k}(x_{i,c};\\theta_{k})\\) represents the classification model that takes as input a data point \\(x_{i,c}\\) associated with class \\(c\\) and outputs a probability distribution over the set of classes in \\(C_k\\). The regularization term \\(R\\) encourages the new model parameters to be close to the previous model parameters \\(\\theta_{k - 1}\\), in order to prevent catastrophic forgetting of the previously learned classes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.681, + 0.877, + 0.699 + ], + "angle": 0, + "content": "3.2. Heterogeneous Data Exacerbates Forgetting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.81 + ], + "angle": 0, + "content": "We argue that the degree of data heterogeneity has a substantial impact on catastrophic forgetting. To verify this, we conduct an experiment on CIFAR-100 dataset [25] with different degrees of data heterogeneity. Inspired by Backward Transfer (BwT) [6], we derive the following formula to measure the severity of forgetting, a popular forgetting measure in CL [4, 7, 8]:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.818, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {k} = \\frac {1}{k - 1} \\sum_ {j = 1} ^ {k - 1} f _ {j} ^ {k}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.87, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_k\\) denotes the average forgetting at \\(k\\)-th task and \\(f_j^k\\) quantifies forgetting for the \\(j\\)-th (\\(j < k\\)) task after the model" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4784" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.125 + ], + "angle": 0, + "content": "has been continually trained up to task \\( k \\). Specifically, for a given data distribution, \\( f_{j}^{k} \\) can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.135, + 0.47, + 0.171 + ], + "angle": 0, + "content": "\\[\nf _ {j} ^ {k} = \\frac {1}{\\left| \\mathcal {C} ^ {j} \\right|} \\sum_ {c \\in \\mathcal {C} ^ {j}} \\max _ {t \\in \\{1, \\dots . N - 1 \\}} \\left(\\mathcal {A} _ {c} ^ {(n)} - \\mathcal {A} _ {c} ^ {(N)}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.469, + 0.278 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}^j\\) is a set of classes related to the \\(j\\)-th task, \\(\\mathcal{A}_c^{(n)}\\) is the accuracy on class \\(c\\) at round \\(t\\), and \\(\\mathcal{A}_c^{(N)}\\) is the final accuracy on class \\(c\\) after learning all tasks. Note that \\(f_j^k\\) captures the average gap between the peak accuracy and the final accuracy for each class of the \\(j\\)-th task after learning the \\(k\\)-th task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.469, + 0.336 + ], + "angle": 0, + "content": "We further extend the catastrophic forgetting measurement \\(\\mathcal{F}\\) in Equation 2 to FL under different data partitions and introduce a relative metric \\(\\mathcal{R}\\) to measure forgetting as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.336, + 0.469, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {k} = \\frac {\\sum_ {j = 1} ^ {k - 1} f _ {j} ^ {k}}{\\sum_ {j = 1} ^ {k - 1} \\mathcal {A} _ {(j , k)}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.382, + 0.469, + 0.427 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}_{(j,k)}\\) is the accuracy on task \\(j\\) after learning task \\(k\\). For different data partitions, an increased \\(\\mathcal{R}_k\\) indicates a more serious forgetting of previous tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.469, + 0.653 + ], + "angle": 0, + "content": "Figure 1 illustrates the impact of catastrophic forgetting under independent and identically distributed (IID) and different levels of non-independent and identically distributed (non-IID) data partitions. In particular, we employ the Dirichlet distribution, which is widely used in FL [29, 28], to simulate the imbalanced label distribution among different clients. Figure 1(a) shows that the accuracy of the model degrades as training proceeds to new tasks in the IID setting. The performance is even worse in the non-IID settings. These results suggest that FCCL faces significant challenge on extreme non-IID settings. We further analyze the forgetting phenomenon in the Figure 1(b). The higher degree of non-IID exacerbates the forgetting phenomenon in FCCL. These empirical studies motivate us to further investigate the catastrophic forgetting issue in FCCL." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.663, + 0.465, + 0.68 + ], + "angle": 0, + "content": "3.3. Alleviating Forgetting via Global Information" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.469, + 0.777 + ], + "angle": 0, + "content": "To tackle the issue of catastrophic forgetting in the FCCL, we argue utilizing global information to improve performance. The global information can derive from the global model. We explore this approach and empirically demonstrate that the integration of global information can effectively mitigate catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Learn from the Global Model Inspired by LwF [30], we leverage the knowledge from the previously trained global model to the current task via Knowledge Distillation (KD), using only the data from the current task. We conduct experiments on CIFAR100 dataset with 5 continual tasks, where each task contains data of 20 classes. We evaluate the accuracy of the model on each previous task after training on" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Table 1: Test accuracy on previous tasks after training on all tasks. In this context, \"FedAvg\" denotes a naive approach whereby clients learn tasks sequentially." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.146, + 0.907, + 0.268 + ], + "angle": 0, + "content": "
PartitionMethodTask 1Task 2Task 3Task 4Task 5Final
IIDFedAvg0000.0582.616.53
FedLwF6.8 ↑11.5 ↑27.1 ↑44.45 ↑63.2 ↓30.61 ↑
NIID (1)FedAvg000081.6516.33
FedLwF6.65 ↑13.71 ↑29.60 ↑45.41 ↑59.35 ↓30.94 ↑
NIID (0.5)FedAvg0000.1577.3015.49
FedLwF1.6 ↑10.65 ↑27.75 ↑41.3 ↑56.65 ↓27.59 ↑
NIID (0.02)FedAvg0000.159.0011.82
FedLwF0.1 ↑1.7 ↑3.35 ↑28.2 ↑54.0 ↓17.47 ↑
" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.285, + 0.7, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.286, + 0.877, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.351, + 0.7, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.351, + 0.877, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.439, + 0.892, + 0.47 + ], + "angle": 0, + "content": "Figure 2: Test accuracy by using current task's data, local exemplar and global exemplar." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.65 + ], + "angle": 0, + "content": "all five tasks and report the final accuracy as the average of these results. Table 1 shows that FCCL without knowledge from the global model (i.e., FedAvg [38]) suffers severely from catastrophic forgetting under both IID and non-IID data partitions. Although FedAvg achieves competitive performance on Task 5, it results in \\(0\\%\\) accuracy on previously trained tasks (Task 1, 2, and 3). In contrast, FedLwF achieves better accuracy on Tasks 1, 2, 3, and 4 and final accuracy. The experiments demonstrate that the global model can indeed alleviate the forgetting issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Learn From Global Exemplar In continual learning, an exemplar refers to a sample from a previous task that is stored in a memory buffer for future training. Assuming that data stored in clients are compliant to use for future training, we adopt the idea proposed in iCaRL [44] to save a small proportion of prior training data in memory. The selection of this small proportion of data assumes that the data distribution pertaining the old task is known, but this assumption does not hold in FL because clients are unable to know the data distributions of others due to data privacy concerns. Nevertheless, we employ two exemplar selection methods to illustrate the utility of global data: global exemplar and local exemplar. Global exemplar assumes that the server aggregates a subset of data from clients and distribute these data to clients in training (Note that this method does" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4785" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.478, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.471, + 0.323 + ], + "angle": 0, + "content": "Figure 3: Pipeline of TARGET. We utilize the global model (trained on task \\( k - 1 \\)) to synthesize the data with a global distribution, subsequently employing this synthesized data for training the \\( k \\)-th task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.396 + ], + "angle": 0, + "content": "not conform to FL and is only used for comparison). Local exemplar means that each client retains a subset of data from local data in the previous tasks for future training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.471, + 0.534 + ], + "angle": 0, + "content": "As shown in Figure 2, we show the performance on the current and all previous tasks after learning the current task by using only the new task data, with local exemplar, and with global exemplar. It can be clearly observed that using exemplars can significantly improve the model performance, especially when using global exemplars, which can achieve much higher accuracy than using local exemplars. However, this raises a critical challenge of how to select such exemplars without violating the data privacy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.303, + 0.563 + ], + "angle": 0, + "content": "4. Our Method: TARGET" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.188, + 0.587 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.471, + 0.746 + ], + "angle": 0, + "content": "To utilize the global information without touching on the real exemplars from clients, we present a method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillage), which utilizes global information without storing any real data. A detailed procedure is provided in Algorithm 1. Figure 3 presents an illustration of TARGET, wherein we synthesize data by inverting the global model \\(\\theta_{k - 1}\\) (which was trained on task \\(k - 1\\)), followed by combining the synthesized data with real data for local model update on task \\(k\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.462, + 0.773 + ], + "angle": 0, + "content": "4.2. Server Side: Synthesizing Data for Old Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As demonstrated in Figure 2, data with global distributional information is more effective in mitigating the problem of catastrophic forgetting. Therefore, we propose a method of synthesizing data that can model the data distribution of the global model, without the need to preserve any client's privacy data. Specifically, given a global (teacher) model \\(\\theta_{k-1}\\) trained on task \\(k - 1\\), we first initialize a generator \\(G\\) and a student model \\(\\theta_S\\). We then repeatedly" + }, + { + "type": "code_caption", + "bbox": [ + 0.51, + 0.095, + 0.764, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Procedure of TARGET." + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.111, + 0.892, + 0.641 + ], + "angle": 0, + "content": "Input: \\(B\\) : local minibatch size, \\(E\\) : local epochs, \\(\\eta\\) .. learning rate, synthetic data: \\(X_{syn} = \\emptyset\\) \n1 foreach each task \\(\\tau = 0,1,\\ldots\\) do \n2 Initialize \\(w_0\\) \n3 foreach each round \\(t = 1,2,\\dots\\) do \n4 \\(S_{t}\\gets\\) (random set of m clients) \n5 foreach client \\(k\\in S_t\\) in parallel do \n6 \\(\\begin{array}{r}\\bigwedge_{t + 1}^{k}\\leftarrow \\text{ClientUpdate}(k,w_{t},\\tau ,X_{syn}) \\end{array}\\) \n7 \\(w_{t + 1}\\gets \\sum_{k = 1}^{m}\\frac{n_k}{n} w_t^k +1\\) \n8 \\(X_{syn} = \\mathsf{DataGeneration}(w_{t + 1})\\) \n9 ClientUpdate \\((k,w,\\tau ,X_{syn})\\) .. \n10 \\(\\mathcal{B}\\gets\\) (split \\(\\mathcal{P}_k\\cup X_{syn}\\) into batches) \n11 set global model \\(\\mathcal{T}\\gets w\\) \n12 foreach each local epoch i from 1 to E do \n13 foreach batch \\((b,b_{syn})\\in \\mathcal{B}\\) do \n14 \\(\\ell (w;b) = L_{ce}(w;b)\\) if \\(\\tau \\neq 0\\) then \n15 \\(\\begin{array}{r}\\lfloor \\ell (w;b) + = \\alpha L_{kl}(w,\\mathcal{T};b_{syn})\\rfloor \\end{array}\\) \n16 return w to server \n17 DataGeneration(w): \n19 Initialize parameter \\(\\theta_G,\\theta_S,X_{syn} = \\emptyset\\) \n20 foreach round \\(i = 1,2,\\ldots\\) do \n21 Sample noises and labels \\(\\{\\mathbf{z}_i,\\mathbf{y}_i\\}_{i = 1}^b\\) // data generation stage \n22 foreach \\(j = 1,2,\\ldots\\) ,do \n23 Generate \\(\\{\\hat{\\mathbf{x}}_i\\}_{i = 1}^b\\) with \\(\\{\\mathbf{z}_i\\}_{i = 1}^b\\) and \\(G(\\cdot)\\) \n24 Update \\(\\theta_G\\) using Equation 9 \n25 Add a batch of data into \\(X_{syn}\\) \n/ model distillation stage \n26 Update \\(\\theta_S\\) using Equation 10 \n27 return \\(X_{syn}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.791 + ], + "angle": 0, + "content": "perform the following two training steps (see line \\(19\\sim 28\\) in Algorithm 1): 1) update the generator by continuously optimizing it to generate data that conforms to the global model distribution; 2) update the student model by distilling knowledge from the teacher model with the synthetic data, hoping that the student model can learn the knowledge of the teacher model sufficiently, which demonstrates the effectiveness of the synthesized data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Data Generation First, we utilize \\( G \\) to generate synthetic data from noise \\( z \\), we need to ensure that the synthetic data \\( \\hat{x} = G(z) \\) is similar to the training dataset. If the synthetic data is similar to the training dataset, their predictions should also be similar. We minimize the cross-entropy (CE) loss on the output of global model \\( \\theta_{k - 1}(\\hat{x}) \\) and random la" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4786" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.125, + 0.106 + ], + "angle": 0, + "content": "bels \\(\\hat{y}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.106, + 0.469, + 0.123 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G} ^ {c e} = C E \\left(\\theta_ {k - 1} (\\hat {x}), \\hat {y}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.13, + 0.47, + 0.311 + ], + "angle": 0, + "content": "It is expected that the synthetic data generated by generator can be classified into a particular class with a high degree of confidence. However, utilizing only the CE loss will cause the generator overfitting to the synthetic data that are far away from the decision boundary (of the global model) [62, 15], thus failing to deliver a good performance. In order to generate samples that are closer to the decision boundary (of the global model) with better transferability, following previous work [62], we introduce a boundary support loss. Additional weight is given to the data on which the global model and the student model diverge in decision making." + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.323, + 0.469, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G} ^ {d i v} = - \\omega K L \\left(\\theta_ {k - 1} (\\hat {x}), \\theta_ {S} (\\hat {x})\\right), \\text {a n d} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.342, + 0.469, + 0.359 + ], + "angle": 0, + "content": "\\[\n\\omega = \\mathbb {1} (\\arg \\max \\theta_ {k - 1} (\\hat {x}) \\neq \\arg \\max \\theta_ {S} (\\hat {x})), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.371, + 0.469, + 0.43 + ], + "angle": 0, + "content": "where \\(KL\\) denotes the Kullback-Leibler (KL) divergence loss, \\(\\mathbb{1}(a)\\) output 1 if \\(a\\) is true and output 0 if \\(a\\) is false. By maximizing the KL divergence loss, the generator can generate more representative data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.469, + 0.49 + ], + "angle": 0, + "content": "Motivated by [57], in order to further improve the stability of generator training, we introduce Batch Normalization (BN) loss to make synthetic data conform with the batch normalization statistics." + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.501, + 0.469, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G} ^ {b n} = \\sum_ {l} \\left(\\left| \\left| \\mu_ {l} (\\hat {x}) - \\mu_ {l} \\right| \\right| + \\left| \\left| \\sigma_ {l} ^ {2} (\\hat {x}) - \\sigma_ {l} ^ {2} \\right| \\right|\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.544, + 0.469, + 0.59 + ], + "angle": 0, + "content": "where \\(\\mu_l(\\hat{x})\\) and \\(\\sigma_l^2 (\\hat{x})\\) are the batch-wise mean and variance estimate of the \\(l\\)-th BN layer of the generator, \\(\\mu_{l}\\) and \\(\\sigma_l^2\\) are the mean and variance of the \\(l\\)-th BN layer of \\(f_{S}(\\cdot)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.59, + 0.469, + 0.62 + ], + "angle": 0, + "content": "Combining the above losses, we can obtain the loss of the generator as follows," + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.631, + 0.469, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G} = \\mathcal {L} _ {G} ^ {c e} + \\lambda_ {1} \\mathcal {L} _ {G} ^ {d i v} + \\lambda_ {2} \\mathcal {L} _ {G} ^ {b n}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.46, + 0.675 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) is the weight for different loss functions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.695, + 0.469, + 0.8 + ], + "angle": 0, + "content": "Model Distillation In Equation 7, we introduce a student model to assist in training the generator to produce data with greater diversity. A better student model should lead to a better generator. Therefore, after training the generator for several rounds, we subsequently train the student model using the saved synthesized data and the output of the teacher model, using KL loss for knowledge distillation:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.812, + 0.469, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = K L \\left(\\theta_ {k - 1} (\\hat {x}), \\theta_ {S} (\\hat {x})\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this way, we can train a student model with better performance, and then further use it to update the generator. An ideal synthetic dataset should be able to efficiently enable student \\(\\theta_{S}\\) to fully learn the knowledge of teacher model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "Note that when the training of the whole process is over (i.e. the student model can use the synthesized data to obtain high performance), we only retain the synthetic dataset \\( X_{syn} \\) and transfer it to the clients, without saving the generator and student model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.178, + 0.881, + 0.194 + ], + "angle": 0, + "content": "4.3. Client Side: Update with Global Information" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.202, + 0.892, + 0.322 + ], + "angle": 0, + "content": "On the client side, we can obtain the data synthesized for the previous task \\( X_{syn} \\) and the real training data of the current task \\( X_{local} \\), then we train the local model \\( \\theta_k \\) for task \\( k \\) on the two datasets at the same time. We showed in Section 3.3 that the use of global models and global data can alleviate forgetting. Thus we distill the knowledge of global teacher model and global synthetic data by minimizing the following objective function," + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.335, + 0.892, + 0.371 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c l i e n t}} = \\underbrace {C E \\left(\\theta_ {k} (x) , y\\right)} _ {\\text {f o r c u r r e n t t a k}} + \\alpha \\cdot \\underbrace {K L \\left(\\theta_ {k - 1} (\\hat {x}) , \\theta_ {k} (\\hat {x})\\right)} _ {\\text {f o r p r e v i o u s t a k s}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.459 + ], + "angle": 0, + "content": "where \\((x,y)\\in X_{local}\\) and \\((\\hat{x})\\in X_{syn}\\). The utilization of the distillation loss facilitates efficient transfer of knowledge from the previous task to the current task model. And \\(\\alpha\\) is a hyper-parameter that controls the strength of the regularization for the previous tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.633, + 0.491 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.707, + 0.516 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.659 + ], + "angle": 0, + "content": "We experiment on two datasets, namely CIFAR-100 [25], and Tiny-ImageNet [26], to evaluate the performance of our proposed approach. To establish the order of the continual tasks, we adopt the widely used protocols [49, 61, 42, 7]. Specifically, we divide all classes of each dataset equally into multiple tasks by default, i.e. we evenly divide the classes into 5 and 10 tasks to simulate class continual learning scenarios. We employ ResNet18 [18] as the backbone for the classification model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.9 + ], + "angle": 0, + "content": "To evaluate our approach, we employ the standard continual learning metrics, as used in prior works [23, 22, 47, 16], which include average accuracy across all tasks and a forgetting measure [6] (see Equation 2). For a fair comparison with the baseline class continual learning methods in the FCCL setting, we implement three types of baselines: 1) Finetune, in which each client simply learns tasks in sequence; 2) FedWeIT [58], a regularization-based method in Federated Continual Learning that maximizes the knowledge transfer between clients; 3) Examples of typical continual learning methods that do not store training data for rehearsal, including EWC [27] and LwF [30]. In addition, we compare our method with methods that store real training data of old tasks, such as iCaRL [44]. We implement these traditional continual learning algorithms in the FCCL scenario and name them as FedEWC, FedLwF, and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4787" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.138 + ], + "angle": 0, + "content": "Table 2: The Average Accuracy (%) and Forgetting for all learned tasks on CIFAR-100 for various numbers of tasks (5, 10) under both IID and non-IID settings. Results are reported as an average of 3 runs. 'Acc' refers to average accuracy, and 'F' represents the forgetting measure utilized in Equation 2. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.147, + 0.894, + 0.324 + ], + "angle": 0, + "content": "
Data partitionIIDNIID (1)NIID (0.5)
TasksT=5T=10T=5T=10T=5T=10
MethodAcc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)
Finetune16.120.787.830.7516.330.778.450.7415.490.747.640.71
FedEWC16.510.718.010.6516.060.688.840.6216.860.668.040.65
FedWeIT28.450.5220.390.4328.560.4919.680.4524.570.5415.450.48
FedLwF30.610.4523.270.3730.940.4221.160.4127.590.4417.980.45
Ours36.310.2224.760.2634.890.2422.850.2633.330.2720.710.29
" + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.347, + 0.278, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.346, + 0.471, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.493, + 0.471, + 0.525 + ], + "angle": 0, + "content": "Figure 4: Average accuracy on previous tasks and current task after the model trained on current task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.598 + ], + "angle": 0, + "content": "FedIcaRL. For detailed information on the task configuration, default hyper-parameters and additional experimental results, please refer to the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.33, + 0.623 + ], + "angle": 0, + "content": "5.2. Experiments on CIFAR-100" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.765 + ], + "angle": 0, + "content": "For CIFAR-100 dataset, we conduct experiments on two sets of tasks consisting of 5 and 10 tasks, respectively. We run the experiment on both IID and non-IID scenarios. For non-IID setting, the Dirichlet parameter is set to 0.5 and 1, i.e. \\(\\mathrm{NIID}(0.5)\\) and \\(\\mathrm{NIID}(1)\\). Table 2 shows the final average accuracy of the FL model trained on all tasks, along with the corresponding forgetting measure for each experiment. It is important to emphasize that an optimal method is characterized by high average accuracy and low forgetting measure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Table 2 indicates that Finetune exhibits the poorest performance when attempting to learn the continuously incoming tasks sequentially, thereby experiencing the issue of catastrophic forgetting. Moreover, we observed that methods based on regularization constraints, such as FedEWC and FedWeIT, were often ineffective in preventing the model from forgetting old tasks due to the lack of available data. On the other hand, we found that distillation-based approaches such as FedLwF and our proposed method were" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.895, + 0.695 + ], + "angle": 0, + "content": "capable of improving the final average accuracy while simultaneously mitigating the issue of catastrophic forgetting. This is due to the transfer of model knowledge learned from the old task to the new task in a distillation manner, enabling the model to prevent catastrophic forgetting. Obviously, we found that applying our proposed method to generate synthetic datasets for the federated models trained on old tasks and subsequently performing model distillation on these synthetic datasets can lead to a substantial improvement in the average accuracy and reduction in the forgetting measure. This observation underscores the ability of our synthetic data to capture the distribution characteristics of historical task data accurately. For example, in IID setting, when partitioning the CIFAR-100 dataset into five tasks, our method achieves an accuracy of \\(36.31\\%\\), which is about \\(6\\%\\) higher than the best baseline method FedLwF. It is worth noting that we observed a decrease in the average accuracy of all methods to varying degrees as the number of tasks increased from 5 tasks to 10 tasks due to increased task complexity and the associated forgetting phenomenon. Nonetheless, our proposed method maintains the highest average accuracy and the lowest forgetting measure even under these more demanding conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Figure 4 illustrates the performance of the models trained by various methods on all previously learned tasks after the completion of each task. Specifically, it shows the average accuracy of the model on both the current and previous tasks after the completion of each task (e.g., after learning the second task, the average accuracy of the model on both the first and second tasks is measured). Based on these curves, it is evident that our proposed model outperforms other competing baseline methods in all incremental tasks, regardless of the number of tasks involved. This finding underscores the effectiveness of our approach in facilitating multiple local clients to learn new classes in a streaming manner while mitigating the forgetting problem." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4788" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.473, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.471, + 0.283 + ], + "angle": 0, + "content": "Figure 5: The average accuracy \\((\\%)\\) and Forgetting for all learned tasks on Tiny-ImageNet for 5 tasks under both IID and non-IID settings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.311, + 0.358, + 0.328 + ], + "angle": 0, + "content": "5.3. Experiments on Tiny-ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.47, + 0.593 + ], + "angle": 0, + "content": "We also evaluated the performance of our proposed method on the more challenging Tiny-ImageNet dataset and obtained similar results to those observed in the CIFAR-100 experiments. Specifically, in Figure 5, we present the final average accuracy and forgetting measure for all learned tasks in both IID and non-IID settings for the case of 5 tasks. The Dirichlet parameter is set to \\(\\{0.05, 0.1, 0.5, 1\\}\\). Based on Figure 5, it is evident that our proposed method consistently outperforms FedLwF in terms of average accuracy across all data partitions. Moreover, our method demonstrates a significantly lower forgetting measure than FedLwF under both IID and non-IID settings. Our proposed method achieves an average accuracy that is approximately \\(3\\%\\) higher than that of FedLwF even in the most challenging scenario (i.e. NIID(0.05)). This result highlights the effectiveness of our method in mitigating catastrophic forgetting in the presence of extreme data distributions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.442, + 0.621 + ], + "angle": 0, + "content": "5.4. Comparison with Exemplar-based Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.853 + ], + "angle": 0, + "content": "In CL, the most successful approaches to alleviate forgetting require extensive replay of previously seen data, which can be problematic when data legality and privacy concerns exist. Among them, iCaRL [44] is a classic but unrealistic algorithm that relies on the stored exemplars in addition to the network parameters, and it is intuitive that using old task's real data could be beneficial to alleviate forgetting problem. To further understand the performance gap, we compare our proposed TARGET (which uses synthetic data) with iCaRL (which requires storing exemplars from old task's real data), and study the effects of different exemplar memories on the performance of our method and iCaRL in Figure 6. We set the stored exemplar size to \\(\\{1000, 1500, 2000\\}\\) for iCaRL, and \\(\\{2000, 3000\\}\\) for our method on the CIFAR-100 dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The accuracy curve in Figure 6 represents the average accuracy rate measured over all 100 classes learned by the model during the last learning task. While our method" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.098, + 0.877, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.259, + 0.892, + 0.306 + ], + "angle": 0, + "content": "Figure 6: Accuracy comparison between real data-based method (iCaRL) and synthetic data-based method (TARGET). We show the result for the last task." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.322, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Table 3: The effect of \\( \\alpha \\) on the performance of both new and old tasks in CIFAR-100, 2 tasks." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.364, + 0.888, + 0.446 + ], + "angle": 0, + "content": "
α35101525
Old Task (0-49)28.3434.9149.0258.4866.02
New Task (50-99)62.4660.2653.7445.2629.32
Average (0-99)45.4147.5851.3851.8747.76
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.715 + ], + "angle": 0, + "content": "can achieve similar performance to storing 1k real training data by storing 2k synthetic data, it still cannot outperform storing 2k real training data. However, it is worth noting that our method does not require storing any real training data, which can be a significant advantage in scenarios where storing real data is difficult or not allowed due to privacy or legal concerns. Additionally, our method achieves better performance than storing only 1k real training data, which indicates that our synthetic data is effective in mitigating catastrophic forgetting. We observed that when our method stores 3k synthetic data, it achieves better accuracy than when it stores 2k synthetic data. However, surpassing iCaRL in performance with an equal amount of data remains a challenge for our method. How to effectively use fewer synthetic data with more valuable knowledge from previous tasks will be left as a future research direction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.722, + 0.741 + ], + "angle": 0, + "content": "5.5. Analysis of Our Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Trade-off between Backward and Forward Transfer. Continual learning presents a challenge in balancing the trade-off between maintaining high accuracy on old tasks (backwards transfer) and achieving high accuracy on new tasks (forward transfer). In Table 3, we evenly split the CIFAR-100 dataset into two tasks and test the average accuracy of our proposed method under different values of \\(\\alpha\\). We observe that the trade-off between backward and forward transfer is not always balanced. In order to achieve good backward transfer, a large value of \\(\\alpha\\) should be used" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4789" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.091, + 0.271, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.091, + 0.454, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.232, + 0.462, + 0.246 + ], + "angle": 0, + "content": "Figure 7: Accuracy for different size of synthetic dataset." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.261, + 0.462, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.418, + 0.45, + 0.433 + ], + "angle": 0, + "content": "Figure 8: Visualization of randomly synthesised data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.626 + ], + "angle": 0, + "content": "to prevent the model from forgetting previous tasks and to encourage it to focus more on the old synthetic data. Conversely, to achieve good forward transfer, a small value of \\(\\alpha\\) should be used, allowing the model to learn quickly and effectively from new tasks while still maintaining some knowledge from the old tasks. The experimental results in Table 3 indicate that when the value of \\(\\alpha\\) is between 10-15, the model achieves a good balance between accuracy on new tasks and accuracy on old tasks. We also partition the CIFAR-100 dataset into 5 equal tasks and test our method's performance on all previous tasks after learning each new task (refer to Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Memory of Synthetic Data. The optimal amount of synthetic data generated for the old task plays a crucial role in determining the final performance of our method. Insufficient synthetic data may not adequately facilitate knowledge transfer from the old tasks, while excessive data generation can lead to increased memory and communication costs. Therefore, determining an appropriate amount of synthetic data that strikes a balance between knowledge transfer and computational cost is critical for the effectiveness and efficiency of our approach. As shown in Figure 7, we test our method on CIFAR-100, which is divided into 5 tasks with different data sizes ranging from 2k to 16k. It can be observed that when the data volume is relatively small, such as 2k and 4k, the performance of the model is poor, especially when the data volume is 2k, the testing curve in task 4 shows a later decline. This is because when the data volume is too small, the model is unable to effectively learn" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.093, + 0.697, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.093, + 0.892, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.22, + 0.892, + 0.25 + ], + "angle": 0, + "content": "Figure 9: Distillation results for task \\(\\{1,2,3,4\\}\\) when trained on 5 tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.895, + 0.43 + ], + "angle": 0, + "content": "knowledge from old tasks. Increasing the data volume to 8k can effectively alleviate the forgetting phenomenon and achieve good performance. However, continuously increasing the data volume to 12k and 16k do not result in significant improvement in the model's performance. It is important to note that the size of the data volume alone does not guarantee the effectiveness of synthetic data in improving machine learning models. The quality and relevance of the synthetic data must also be carefully considered to ensure that it accurately represents the underlying distribution of the real-world data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.895, + 0.587 + ], + "angle": 0, + "content": "Visualization on Synthetic Data. To demonstrate the effectiveness of the synthetic data, we present in Figure 8 the visualization results of the synthesized images generated by our method after the model learns the penultimate task on Tiny-ImageNet (CIFAR-100 in Appendix) for 5 tasks. These data have the potential to efficiently enable the student model to approach the performance of the teacher model rapidly. However, it is worth noting that these data exhibit visual dissimilarities from the actual training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.894, + 0.711 + ], + "angle": 0, + "content": "Analysis on Distillation. Based on the synthetic data, we distill the knowledge of the model trained on old tasks into a student model. The effect of distillation is further demonstrated in Figure 9. It can be observed that even without accessing any private training data from clients, our method can quickly distill the student model on the server to approach the performance of the global model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.62, + 0.74 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In conclusion, this paper introduces a novel method, TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillage), to alleviate the catastrophic forgetting problem in Federated Class-Continual Learning (FCCL). Unlike all the previous methods, our proposed method leverages global knowledge, without requiring any additional datasets or data from previous tasks, making it ideal for privacy-sensitive scenarios. Extensive experimental results demonstrate the effectiveness of our proposed method in comparison to existing FCCL methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4790" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.473, + 0.171 + ], + "angle": 0, + "content": "[1] Eden Belouadah and Adrian Popescu. Il2m: Class incremental learning with dual memory. In Proceedings of the IEEE/CVF international conference on computer vision, pages 583-592, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.214 + ], + "angle": 0, + "content": "[2] Cecilia Bjursell. The pandemic as disjuncture: Lifelong learning in a context of fear. International Review of Education, 66(5-6):673-689, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.472, + 0.272 + ], + "angle": 0, + "content": "[3] Thang D Bui, Cuong V Nguyen, Siddharth Swaroop, and Richard E Turner. Partitioned variational inference: A unified framework encompassing federated and continual learning. arXiv preprint arXiv:1811.11206, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.472, + 0.328 + ], + "angle": 0, + "content": "[4] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio P. Calmon, and Taesup Moon. Cpr: classifier-projection regularization for continual learning. arXiv preprint arXiv:2006.07326, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.472, + 0.386 + ], + "angle": 0, + "content": "[5] Yatin Chaudhary, Pranav Rai, Matthias Schubert, Hinrich Schütze, and Pankaj Gupta. Federated continual learning for text classification via selective inter-client transfer. arXiv preprint arXiv:2210.06101, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.389, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[6] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532-547, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[7] Arslan Chaudhry, Naemullah Khan, Puneet Dokania, and Philip Torr. Continual learning in low-rank orthogonal subspaces. Advances in Neural Information Processing Systems, 33:9900-9911, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.472, + 0.559 + ], + "angle": 0, + "content": "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.561, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[9] Marco Ciotti, Massimo Ciccozzi, Alessandro Terrinoni, Wen-Can Jiang, Cheng-Bin Wang, and Sergio Bernardini. The COVID-19 pandemic. Critical reviews in clinical laboratory sciences, 57(6):365-388, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[10] Jiahua Dong, Yang Cong, Gan Sun, Zhen Fang, and Zhengming Ding. Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1(2):1-18, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[12] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 86-102. Springer, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[14] Fernando Estevez Casado, Dylan Lema Pais, Marcos Fernández Criado, Roberto Iglesias Rodríguez, Carlos Vázquez Regueiro, and Senen Barro Ameneiro. Concept drift detection and adaptation for federated and continual learning. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[15] Gongfan Fang, Kanya Mo, Xinchao Wang, Jie Song, Shitao Bei, Haofei Zhang, and Mingli Song. Up to 100x faster data-free knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 6597-6604, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[16] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[17] Andrew Hard, Kanishka Rao, Rajiv Mathews, Swaroop Ramaswamy, Françoise Beaufays, Sean Augenstein, Hubert Eichner, Chloe Kiddon, and Daniel Ramage. Federated learning for mobile keyboard prediction. arXiv preprint arXiv:1811.03604, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[19] Ziyue Jiang, Yi Ren, Ming Lei, and Zhou Zhao. Fed-speech: Federated text-to-speech with continual learning. arXiv preprint arXiv:2110.07216, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[20] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and Trends® in Machine Learning, 14(1-2):1-210, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.546, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[21] Georgios Kaissis, Alexander Ziller, Jonathan Passerat-Palmbach, Theo Ryffel, Dmitrii Usynin, Andrew Trask, Ionésio Lima Jr, Jason Mancuso, Friederike Jungmann, Marc-Matthias Steinborn, et al. End-to-end privacy preserving deep learning on multi-institutional medical imaging. Nature Machine Intelligence, 3(6):473-484, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.631, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[22] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[23] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[24] Jakub Konečny, Brendan McMahan, and Daniel Ramage. Federated optimization: Distributed optimization beyond the datacenter. arXiv preprint arXiv:1511.03575, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.814, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forget-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4791" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "ting by incremental moment matching. Advances in neural information processing systems, 30, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. arXiv preprint arXiv:1812.06127, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.469, + 0.37 + ], + "angle": 0, + "content": "[31] Zexi Li, Qunwei Li, Yi Zhou, Wenliang Zhong, Guannan Zhang, and Chao Wu. Edge-cloud collaborative learning with federated and centralized features. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '23, page 1949-1953, New York, NY, USA, 2023. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.373, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[32] Zexi Li, Tao Lin, Xinyi Shang, and Chao Wu. Revisiting weighted aggregation in federated learning with neural networks. In International Conference on Machine Learning. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[33] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Rmm: Reinforced memory management for class-incremental learning. Advances in Neural Information Processing Systems, 34:3478-3490, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.485, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[34] Yuang Liu, Wei Zhang, Jun Wang, and Jianyong Wang. Data-free knowledge transfer: A survey. arXiv preprint arXiv:2112.15278, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.567 + ], + "angle": 0, + "content": "[35] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[36] Yuhang Ma, Zhongle Xie, Jue Wang, Ke Chen, and Lidan Shou. Continual federated learning based on knowledge distillation. In Lud De Raedt, editor, Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 2182–2188. International Joint Conferences on Artificial Intelligence Organization, 7 2022. Main Track. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[37] Yuhang Ma, Zhongle Xie, Jue Wang, Ke Chen, and Lidan Shou. Continual federated learning based on knowledge distillation. In Luc De Raedt, editor, Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, pages 2182-2188. ijcai.org, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[38] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Aarti Singh and Jerry Zhu, editors, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, volume 54 of Proceedings of Machine Learning Research, pages 1273-1282. PMLR, 20-22 Apr 2017. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[40] Sudhanshu Mittal, Silvio Galesso, and Thomas Brox. Essentials for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3513-3522, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.89, + 0.207 + ], + "angle": 0, + "content": "[41] Alexander Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[42] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural networks, 113:54-71, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[43] Tae Jin Park, Kenichi Kumatani, and Dimitrios Dimitriadis. Tackling dynamics in federated incremental learning with variational embedding rehearsal. arXiv preprint arXiv:2110.09695, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[44] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[45] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. arXiv preprint arXiv:1810.11910, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[46] Amanda Rios and Laurent Itti. Closed-loop gan for continual learning. arXiv preprint arXiv:1811.01146, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[47] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[48] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International Conference on Machine Learning, pages 4548-4557. PMLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[49] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.62, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[50] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9374-9384, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[51] Sebastian Thrun. Lifelong learning algorithms. Learning to learn, 8:181-209, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[52] Anastasiia Usmanova, François Portet, Philippe Lalanda, and German Vega. A distillation-based approach integrating continual learning and federated learning for pervasive services. arXiv preprint arXiv:2109.04197, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[53] Anamaria Vizitiu, Cosmin Loan Nita, Andrei Puiu, Constantin Suciu, and Lucian Mihai Itu. Towards privacy-preserving deep learning based medical imaging applications. In 2019 IEEE international symposium on medical measurements and applications (MeMeA), pages 1-6. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In Computer Vision-ECCV 2022: 17th" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4792" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV, pages 398-414. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.162 + ], + "angle": 0, + "content": "[55] Ju Xu and Zhanxing Zhu. Reinforced continual learning. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[56] Li Yang, Shasha Liu, Jinyan Liu, Zhixin Zhang, Xiaochun Wan, Bo Huang, Youhai Chen, and Yi Zhang. Covid-19: immunopathogenesis and immunotherapeutics. Signal transduction and targeted therapy, 5(1):128, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[57] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715-8724, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[58] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086. PMLR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[59] Jaehong Yoon, Saehoon Kim, Eunho Yang, and Sung Ju Hwang. Scalable and order-robust continual learning with additive parameter decomposition. arXiv preprint arXiv:1902.09432, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[60] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[61] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987-3995. PMLR, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[62] Jie Zhang, Chen Chen, Bo Li, Lingjuan Lyu, Shuang Wu, Shouhong Ding, Chunhua Shen, and Chao Wu. Dense: Data-free one-shot federated learning. In Advances in Neural Information Processing Systems. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[63] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In International Conference on Machine Learning, pages 26311-26329. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[64] Junting Zhang, Jie Zhang, Shalini Ghosh, Dawei Li, Serafettin Tasci, Larry Heck, Heming Zhang, and C-C Jay Kuo. Class-incremental learning via deep model consolidation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1131-1140, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[65] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13208-13217, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[66] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[67] Didi Zhu, Yinchuan Li, Yunfeng Shao, Jianye Hao, Fei Wu, Kun Kuang, Jun Xiao, and Chao Wu. Generalized universal domain adaptation with generative flow networks. arXiv preprint arXiv:2305.04466, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[68] Didi Zhu, Yincuan Li, Junkun Yuan, Zexi Li, Yunfeng Shao, Kun Kuang, and Chao Wu. Universal domain adaptation via compressive attention matching. arXiv preprint arXiv:2304.11862, 2023. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4793" + } + ] +] \ No newline at end of file diff --git a/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/ff5641a5-17a3-470a-a519-ccac898db108_origin.pdf b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/ff5641a5-17a3-470a-a519-ccac898db108_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f4c8eb6a7d1f3b5914a48ec3dcfa76eade5db6e7 --- /dev/null +++ b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/ff5641a5-17a3-470a-a519-ccac898db108_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1f5faef5fabab5b83d583f7983e0bf41254a13a7eacc596beda5a01f40afb05 +size 2231900 diff --git a/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/full.md b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..455e3ec246b95bad9ac6901b3a05366c7dcac4fb --- /dev/null +++ b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/full.md @@ -0,0 +1,396 @@ +# TARGET: Federated Class-Continual Learning via Exemplar-Free Distillation + +Jie Zhang* +ETH Zurich +zj.jayzhang@gmail.com + +Chen Chen & Weiming Zhuang & Lingjuan Lyu† +Sony AI + +{ChenA.Chen, weiming.zhuang, Lingjuan.Lv}@sony.com + +# Abstract + +This paper focuses on an under-explored yet important problem: Federated Class-Continual Learning (FCCL), where new classes are dynamically added in federated learning. Existing FCCL works suffer from various limitations, such as requiring additional datasets or storing the private data from previous tasks. In response, we first demonstrate that non-IID data exacerbates catastrophic forgetting issue in FL. Then we propose a novel method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free distTillage), which alleviates catastrophic forgetting in FCCL while preserving client data privacy. Our proposed method leverages the previously trained global model to transfer knowledge of old tasks to the current task at the model level. Moreover, a generator is trained to produce synthetic data to simulate the global distribution of data on each client at the data level. Compared to previous FCCL methods, TARGET does not require any additional datasets or storing real data from previous tasks, which makes it ideal for data-sensitive scenarios. + +# 1. Introduction + +Federated Learning (FL) is a privacy-aware learning paradigm that facilitates collaborations among multiple entities (e.g., edge devices or organizations) [38, 20, 28, 66, 17]. Each entity or client in FL retains data locally and transfers only training updates to the central server for aggregation. + +Conventional FL studies assume that the data classes and domains are static, but the new classes could emerge and data domains could change over time in reality [64, + +40, 1, 65]. For example, multiple health institutions could use FL to collaborate and train models to identify COVID-19 [56, 9, 2] strains; new COVID-19 strains, however, continue to emerge due to high mutation rate of virus. An intuitive solution to this issue of continuously emerging data classes is training new models from scratch, but this is impractical as it would require significant extra computation cost. Another method is transfer learning from the previously trained model, but this method suffers from catastrophic forgetting [23, 22, 47, 16], degrading performance on the previous classes. + +To address these issues, recent research [36, 11, 3, 19] has introduced the concept of Continual Learning (CL) [61, 42, 51, 49, 48] within the FL framework. These methods, collectively referred to as Federated Continual Learning (FCL), aim to mitigate the problems of catastrophic forgetting in FL. In most FCL scenarios, new classes are dynamically added, which we call Federated Class-Continual Learning (FCCL). FCCL allows local clients to continuously collect new data, and new classes can be added at any time. + +Unfortunately, existing FCCL works suffer from various limitations. For example, Ma et al. [36] utilize an unlabeled surrogate dataset to address the catastrophic forgetting problem, which may be difficult to obtain in some data sensitive scenarios. Furthermore, the usage of an unlabeled surrogate dataset may not be ideal for certain types of data, as it may not capture the full complexity of the original data. In CL, exemplar-based methods [44, 33, 54, 13] have achieved leading performance. An exemplar refers to a sample or instance of a previously seen data point that is retained in a memory buffer for future use in the learning process. Dong et al. [11] propose a exemplar-based method that stores historical data to address catastrophic forgetting. However, in many privacy-sensitive scenarios (e.g., hospitals and medical research institutions), users are + +not permitted to store data from previous tasks due to privacy and policy concerns and data will not be kept for a long time [21, 53, 34, 50]. In summary, the majority of FCCL methods train the global model with additional datasets or previous task data, which could potentially violate data privacy regulations. This dilemma prompts us to consider the following question: + +Question: How to effectively alleviate the catastrophic forgetting problem in the FCCL without storing the local private data of the client or any additional datasets? + +To address this question, we conduct a systematic analysis and observe that the imbalanced distribution of data among clients in FL exacerbates the catastrophic forgetting problem (see Section 3.3). In order to fix this problem: 1) at the model level, we leverage the previously trained global model to transfer knowledge of the old tasks to the current task. 2) at the data level, we train a generator to produce synthetic data that aims to simulate the global distribution of data on each client. Drawing on these insights, we present a method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillation) that mitigates catastrophic forgetting in FCCL without compromising clients' data privacy. + +Our contributions can be concluded as follows: + +- We are the first to demonstrate that non-independent and identically distributed (non-IID) data exacerbates catastrophic forgetting issue in FL. Then we propose a novel method called TARGET, which alleviates the catastrophic forgetting in FCCL by leveraging global information. +- Compared to previous FCCL methods, TARGET doesn't require extra datasets or data from previous tasks, it can be applied in data sensitive scenarios. +- Extensive experiments demonstrate the efficacy of our proposed method. For example, when partitioning the CIFAR-100 dataset into five tasks, our method achieves an accuracy of $36.31\%$ , which is about $6\%$ higher than the best baseline method. + +# 2. Related Work + +# 2.1. Federated Continual Learning + +In recent years, deep neural networks (DNNs) have been widely used as a fundamental technology for the development of artificial intelligence, both in established and emerging fields [68, 67, 32, 31, 10]. With the advancement of deep learning, an increasing number of researchers have started to focus on the construction of privacy-preserving deep learning frameworks. FL is a paradigm for collaboratively building a model across multiple clients [24, 39], which is gaining momentum in recent years. However, Federated Continual Learning (FCL), which focused on both FL and CL simultaneously is just emerging and still + +remains pending further research [52, 43, 58]. Apart from the client-wise catastrophic forgetting, FCL paradigm also poses new challenges such as inter-client interference and communication-efficiency [58]. FedWeIT solves these challenges through decomposing the parameters into three parts, i.e., global parameters, local based parameters, and task-adaptive parameters [58]. Concurrently, Concept-Drift-Aware Federated Averaging (CDA-FedAvg) extends the popular FL algorithm, Federated Averaging (FedAvg), to tackle the CL problem by introducing the concept drift detection and adaptation [14]. FCL with Distillation (CFeD) treats the model trained with the last task as the teacher model to perform a knowledge distillation and proposes a server distillation mechanism to deal with non-i.i.d. issue [37]. Global-Local Forgetting Compensation (GLFC) designs a class-aware gradient compensation loss and a class-semantic relation distillation loss to prevent forgetting, and a proxy server to mitigate the non-independent and identically distributed (non-IID) problem [12]. Additionally, Federated Selective Inter-client Transfer (FedSeIT) applies FCL to NLP through selectively combining model parameters of foreign clients and selecting informative tasks to perform knowledge transfer [5]. These papers illustrate the growing interest in FCL and the need for novel approaches to address its unique challenges. As research in this area continues to advance, we can expect to see more innovative approaches to FCL that further improve its performance, scalability, and privacy preservation. + +# 2.2. Continual Learning + +Continual Learning has been studied extensively, several training methods have been proposed to address the catastrophic forgetting challenge it presents. Regularization-based approaches: Elastic Weight Consolidation (EWC) selectively penalizes the network parameters that are important for old tasks [23]. Synaptic intelligence (SI) uses a memory buffer to store important network parameters [61]. Incremental moment matching (IMM) modelings the posterior distribution after learning multiple tasks as a mixture of Gaussian models [27]. Stable SGD proposes to carefully design the training regimes such as learning rate decay, batch size, dropout, and optimizer to alleviate forgetting. Replay-based approaches: The methods in this family construct a memory to store the past information which will be presented to the model for reviewing in future tasks. Some store the knowledge of the previous tasks, known as experience replay [35, 45, 49, 46]. iCaRL preserves the most representative samples of each class [44]. Averaged Gradient Episodic Memory (A-GEM) builds an episodic memory of model parameter gradients[8]. Architecture-based: DEN expands the model size [60] and RCL utilizes reinforcement learning [55]. APD divides the model parameters into shared and task-specific parameters to restrict the + +model complexity [59]. + +Exemplar-Free Continual Learning Besides, a promising line of work focuses on data-free Continual Learning. DeepDream perturbs current training samples into images that maximize "forgetting" from the previous tasks [41]. DeepInversion proposes a model inversion technique and evaluates its performance in a Continual Learning scenario but found limited success [57]. DFCIL [50] investigates the failure and decomposes the CE-loss into two different losses which guarantee the learning of effective features. These methods enable Continual Learning in scenarios where storing old task data is not feasible or desirable, such as in privacy-sensitive applications. + +# 3. Catastrophic Forgetting in FCCL + +This section conducts an in-depth analysis of catastrophic forgetting problem in federated class-continual learning (FCCL). We start by providing a formal definition of the problem. Then, we investigate the forgetting issue in FCCL and discuss potential methods to mitigate it. + +# 3.1. Problem Definition + +Federated Class-Continual Learning (FCCL) focuses on the problem of learning models for new classes over time in FL. An FCCL framework consists of a central server and multiple clients. All clients do not share their raw data with any other client or the central server. Each client learns from a sequence of $n$ tasks, where $k$ -th task contains non-overlapping subsets of classes $C_k \in C$ , where $C$ is the set of all possible classes. In our privacy-aware scenario, the task stream is presented in an unknown order, and each client can only access its local data from task $k$ during that task's training period, which is no longer accessible thereafter. Note that the models are trained in a distributed manner, where each party has access to only a subset of the classes $C_k$ (i.e. non-IID). We also consider a more challenging and practical setting where the data in each client is heterogeneous. In this paper, we assume the label distribution of data in each client is skewed [29, 63]. + +Forgetting Issue in FCCL The goal of the global model optimization problem at task $k$ is to minimize the overall classification error on the current set of classes $C_k$ . However, when a new task arises, clients are not able to access data from previous (old) tasks due to privacy concerns and can only update their local model with data from the new task. This often leads to a significant decrease in performance on previous tasks, which is known as catastrophic forgetting [23, 22, 47, 16]. To mitigate catastrophic forgetting in the global model, we aim to minimize the overall classification error on the current set of classes $C_k$ , while + +![](images/d92f5b7b0a003af57b65bca690ac044f6a491b05b0137b553fbe616f8cfa6720.jpg) +(a) Test Accuracy +Figure 1: Forgetting under different data partitions. We partitioned the CIFAR-100 dataset into five tasks with 20 classes each, distributed among five clients. "NIID" refers to non-IID, where a lower value represents a more imbalanced or skewed distribution of data. Non-IIDness refers to the Dirichlet parameter. + +![](images/db207368bef2e6b47c3e0075a06635eef4acaa8dc916ecf0c50c63b54904634a.jpg) +(b) Relative Forgetting + +simultaneously minimizing the changes to the previously learned classes. Formally, the objective function can be written as: + +$$ +\min _ {\theta_ {k}} \sum_ {c \in C _ {k}} \sum_ {i = 1} ^ {m _ {c}} L \left(f _ {k} \left(x _ {i, c}; \theta_ {k}\right), c\right) + \alpha R \left(\theta_ {k}, \theta_ {k - 1}\right) \tag {1} +$$ + +where $\theta_{k}$ is the model parameter at round $k$ , $L$ is a loss function that measures the classification error, $R$ is a regularization term that penalizes changes to the previous model parameters, $m_{c}$ is the number of data in class $c$ , and $\alpha$ is a hyper-parameter that controls the strength of the regularization. In this formula, $f_{k}(x_{i,c};\theta_{k})$ represents the classification model that takes as input a data point $x_{i,c}$ associated with class $c$ and outputs a probability distribution over the set of classes in $C_k$ . The regularization term $R$ encourages the new model parameters to be close to the previous model parameters $\theta_{k - 1}$ , in order to prevent catastrophic forgetting of the previously learned classes. + +# 3.2. Heterogeneous Data Exacerbates Forgetting + +We argue that the degree of data heterogeneity has a substantial impact on catastrophic forgetting. To verify this, we conduct an experiment on CIFAR-100 dataset [25] with different degrees of data heterogeneity. Inspired by Backward Transfer (BwT) [6], we derive the following formula to measure the severity of forgetting, a popular forgetting measure in CL [4, 7, 8]: + +$$ +\mathcal {F} _ {k} = \frac {1}{k - 1} \sum_ {j = 1} ^ {k - 1} f _ {j} ^ {k}, \tag {2} +$$ + +where $\mathcal{F}_k$ denotes the average forgetting at $k$ -th task and $f_j^k$ quantifies forgetting for the $j$ -th ( $j < k$ ) task after the model + +has been continually trained up to task $k$ . Specifically, for a given data distribution, $f_{j}^{k}$ can be expressed as follows: + +$$ +f _ {j} ^ {k} = \frac {1}{\left| \mathcal {C} ^ {j} \right|} \sum_ {c \in \mathcal {C} ^ {j}} \max _ {t \in \{1, \dots . N - 1 \}} \left(\mathcal {A} _ {c} ^ {(n)} - \mathcal {A} _ {c} ^ {(N)}\right), \tag {3} +$$ + +where $\mathcal{C}^j$ is a set of classes related to the $j$ -th task, $\mathcal{A}_c^{(n)}$ is the accuracy on class $c$ at round $t$ , and $\mathcal{A}_c^{(N)}$ is the final accuracy on class $c$ after learning all tasks. Note that $f_j^k$ captures the average gap between the peak accuracy and the final accuracy for each class of the $j$ -th task after learning the $k$ -th task. + +We further extend the catastrophic forgetting measurement $\mathcal{F}$ in Equation 2 to FL under different data partitions and introduce a relative metric $\mathcal{R}$ to measure forgetting as follows: + +$$ +\mathcal {R} _ {k} = \frac {\sum_ {j = 1} ^ {k - 1} f _ {j} ^ {k}}{\sum_ {j = 1} ^ {k - 1} \mathcal {A} _ {(j , k)}}, \tag {4} +$$ + +where $\mathcal{A}_{(j,k)}$ is the accuracy on task $j$ after learning task $k$ . For different data partitions, an increased $\mathcal{R}_k$ indicates a more serious forgetting of previous tasks. + +Figure 1 illustrates the impact of catastrophic forgetting under independent and identically distributed (IID) and different levels of non-independent and identically distributed (non-IID) data partitions. In particular, we employ the Dirichlet distribution, which is widely used in FL [29, 28], to simulate the imbalanced label distribution among different clients. Figure 1(a) shows that the accuracy of the model degrades as training proceeds to new tasks in the IID setting. The performance is even worse in the non-IID settings. These results suggest that FCCL faces significant challenge on extreme non-IID settings. We further analyze the forgetting phenomenon in the Figure 1(b). The higher degree of non-IID exacerbates the forgetting phenomenon in FCCL. These empirical studies motivate us to further investigate the catastrophic forgetting issue in FCCL. + +# 3.3. Alleviating Forgetting via Global Information + +To tackle the issue of catastrophic forgetting in the FCCL, we argue utilizing global information to improve performance. The global information can derive from the global model. We explore this approach and empirically demonstrate that the integration of global information can effectively mitigate catastrophic forgetting. + +Learn from the Global Model Inspired by LwF [30], we leverage the knowledge from the previously trained global model to the current task via Knowledge Distillation (KD), using only the data from the current task. We conduct experiments on CIFAR100 dataset with 5 continual tasks, where each task contains data of 20 classes. We evaluate the accuracy of the model on each previous task after training on + +Table 1: Test accuracy on previous tasks after training on all tasks. In this context, "FedAvg" denotes a naive approach whereby clients learn tasks sequentially. + +
PartitionMethodTask 1Task 2Task 3Task 4Task 5Final
IIDFedAvg0000.0582.616.53
FedLwF6.8 ↑11.5 ↑27.1 ↑44.45 ↑63.2 ↓30.61 ↑
NIID (1)FedAvg000081.6516.33
FedLwF6.65 ↑13.71 ↑29.60 ↑45.41 ↑59.35 ↓30.94 ↑
NIID (0.5)FedAvg0000.1577.3015.49
FedLwF1.6 ↑10.65 ↑27.75 ↑41.3 ↑56.65 ↓27.59 ↑
NIID (0.02)FedAvg0000.159.0011.82
FedLwF0.1 ↑1.7 ↑3.35 ↑28.2 ↑54.0 ↓17.47 ↑
+ +![](images/e28bdbf18d73129be0197101faed2ebb137ecd524954b59c9140dae34e1eae38.jpg) + +![](images/b831d8b74780b3c9b36887b41335ac1ed0c8a06b00b2bbadb2079a3033eb7265.jpg) + +![](images/5e6b50ad06163db6dcaebff062c27b15c52a8d7263f45bd9804af0b7d99f1ead.jpg) +Figure 2: Test accuracy by using current task's data, local exemplar and global exemplar. + +![](images/72db22f6aa0020921d7347a583d6aa08ee4aa9f89a707c9dfc872a3af84e567d.jpg) + +all five tasks and report the final accuracy as the average of these results. Table 1 shows that FCCL without knowledge from the global model (i.e., FedAvg [38]) suffers severely from catastrophic forgetting under both IID and non-IID data partitions. Although FedAvg achieves competitive performance on Task 5, it results in $0\%$ accuracy on previously trained tasks (Task 1, 2, and 3). In contrast, FedLwF achieves better accuracy on Tasks 1, 2, 3, and 4 and final accuracy. The experiments demonstrate that the global model can indeed alleviate the forgetting issue. + +Learn From Global Exemplar In continual learning, an exemplar refers to a sample from a previous task that is stored in a memory buffer for future training. Assuming that data stored in clients are compliant to use for future training, we adopt the idea proposed in iCaRL [44] to save a small proportion of prior training data in memory. The selection of this small proportion of data assumes that the data distribution pertaining the old task is known, but this assumption does not hold in FL because clients are unable to know the data distributions of others due to data privacy concerns. Nevertheless, we employ two exemplar selection methods to illustrate the utility of global data: global exemplar and local exemplar. Global exemplar assumes that the server aggregates a subset of data from clients and distribute these data to clients in training (Note that this method does + +![](images/417749987b67f0dcf09c14e375aaadf518cfa018601e6898cb594f773b7544e8.jpg) +Figure 3: Pipeline of TARGET. We utilize the global model (trained on task $k - 1$ ) to synthesize the data with a global distribution, subsequently employing this synthesized data for training the $k$ -th task. + +not conform to FL and is only used for comparison). Local exemplar means that each client retains a subset of data from local data in the previous tasks for future training. + +As shown in Figure 2, we show the performance on the current and all previous tasks after learning the current task by using only the new task data, with local exemplar, and with global exemplar. It can be clearly observed that using exemplars can significantly improve the model performance, especially when using global exemplars, which can achieve much higher accuracy than using local exemplars. However, this raises a critical challenge of how to select such exemplars without violating the data privacy. + +# 4. Our Method: TARGET + +# 4.1. Overview + +To utilize the global information without touching on the real exemplars from clients, we present a method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillage), which utilizes global information without storing any real data. A detailed procedure is provided in Algorithm 1. Figure 3 presents an illustration of TARGET, wherein we synthesize data by inverting the global model $\theta_{k - 1}$ (which was trained on task $k - 1$ ), followed by combining the synthesized data with real data for local model update on task $k$ . + +# 4.2. Server Side: Synthesizing Data for Old Tasks + +As demonstrated in Figure 2, data with global distributional information is more effective in mitigating the problem of catastrophic forgetting. Therefore, we propose a method of synthesizing data that can model the data distribution of the global model, without the need to preserve any client's privacy data. Specifically, given a global (teacher) model $\theta_{k-1}$ trained on task $k - 1$ , we first initialize a generator $G$ and a student model $\theta_S$ . We then repeatedly + +Algorithm 1: Procedure of TARGET. +Input: $B$ : local minibatch size, $E$ : local epochs, $\eta$ .. learning rate, synthetic data: $X_{syn} = \emptyset$ +1 foreach each task $\tau = 0,1,\ldots$ do +2 Initialize $w_0$ +3 foreach each round $t = 1,2,\dots$ do +4 $S_{t}\gets$ (random set of m clients) +5 foreach client $k\in S_t$ in parallel do +6 $\begin{array}{r}\bigwedge_{t + 1}^{k}\leftarrow \text{ClientUpdate}(k,w_{t},\tau ,X_{syn}) \end{array}$ +7 $w_{t + 1}\gets \sum_{k = 1}^{m}\frac{n_k}{n} w_t^k +1$ +8 $X_{syn} = \mathsf{DataGeneration}(w_{t + 1})$ +9 ClientUpdate $(k,w,\tau ,X_{syn})$ .. +10 $\mathcal{B}\gets$ (split $\mathcal{P}_k\cup X_{syn}$ into batches) +11 set global model $\mathcal{T}\gets w$ +12 foreach each local epoch i from 1 to E do +13 foreach batch $(b,b_{syn})\in \mathcal{B}$ do +14 $\ell (w;b) = L_{ce}(w;b)$ if $\tau \neq 0$ then +15 $\begin{array}{r}\lfloor \ell (w;b) + = \alpha L_{kl}(w,\mathcal{T};b_{syn})\rfloor \end{array}$ +16 return w to server +17 DataGeneration(w): +19 Initialize parameter $\theta_G,\theta_S,X_{syn} = \emptyset$ +20 foreach round $i = 1,2,\ldots$ do +21 Sample noises and labels $\{\mathbf{z}_i,\mathbf{y}_i\}_{i = 1}^b$ // data generation stage +22 foreach $j = 1,2,\ldots$ ,do +23 Generate $\{\hat{\mathbf{x}}_i\}_{i = 1}^b$ with $\{\mathbf{z}_i\}_{i = 1}^b$ and $G(\cdot)$ +24 Update $\theta_G$ using Equation 9 +25 Add a batch of data into $X_{syn}$ +/ model distillation stage +26 Update $\theta_S$ using Equation 10 +27 return $X_{syn}$ + +perform the following two training steps (see line $19\sim 28$ in Algorithm 1): 1) update the generator by continuously optimizing it to generate data that conforms to the global model distribution; 2) update the student model by distilling knowledge from the teacher model with the synthetic data, hoping that the student model can learn the knowledge of the teacher model sufficiently, which demonstrates the effectiveness of the synthesized data. + +Data Generation First, we utilize $G$ to generate synthetic data from noise $z$ , we need to ensure that the synthetic data $\hat{x} = G(z)$ is similar to the training dataset. If the synthetic data is similar to the training dataset, their predictions should also be similar. We minimize the cross-entropy (CE) loss on the output of global model $\theta_{k - 1}(\hat{x})$ and random la + +bels $\hat{y}$ + +$$ +\mathcal {L} _ {G} ^ {c e} = C E \left(\theta_ {k - 1} (\hat {x}), \hat {y}\right). \tag {5} +$$ + +It is expected that the synthetic data generated by generator can be classified into a particular class with a high degree of confidence. However, utilizing only the CE loss will cause the generator overfitting to the synthetic data that are far away from the decision boundary (of the global model) [62, 15], thus failing to deliver a good performance. In order to generate samples that are closer to the decision boundary (of the global model) with better transferability, following previous work [62], we introduce a boundary support loss. Additional weight is given to the data on which the global model and the student model diverge in decision making. + +$$ +\mathcal {L} _ {G} ^ {d i v} = - \omega K L \left(\theta_ {k - 1} (\hat {x}), \theta_ {S} (\hat {x})\right), \text {a n d} \tag {6} +$$ + +$$ +\omega = \mathbb {1} (\arg \max \theta_ {k - 1} (\hat {x}) \neq \arg \max \theta_ {S} (\hat {x})), \tag {7} +$$ + +where $KL$ denotes the Kullback-Leibler (KL) divergence loss, $\mathbb{1}(a)$ output 1 if $a$ is true and output 0 if $a$ is false. By maximizing the KL divergence loss, the generator can generate more representative data. + +Motivated by [57], in order to further improve the stability of generator training, we introduce Batch Normalization (BN) loss to make synthetic data conform with the batch normalization statistics. + +$$ +\mathcal {L} _ {G} ^ {b n} = \sum_ {l} \left(\left| \left| \mu_ {l} (\hat {x}) - \mu_ {l} \right| \right| + \left| \left| \sigma_ {l} ^ {2} (\hat {x}) - \sigma_ {l} ^ {2} \right| \right|\right), \tag {8} +$$ + +where $\mu_l(\hat{x})$ and $\sigma_l^2 (\hat{x})$ are the batch-wise mean and variance estimate of the $l$ -th BN layer of the generator, $\mu_{l}$ and $\sigma_l^2$ are the mean and variance of the $l$ -th BN layer of $f_{S}(\cdot)$ . + +Combining the above losses, we can obtain the loss of the generator as follows, + +$$ +\mathcal {L} _ {G} = \mathcal {L} _ {G} ^ {c e} + \lambda_ {1} \mathcal {L} _ {G} ^ {d i v} + \lambda_ {2} \mathcal {L} _ {G} ^ {b n}, \tag {9} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ is the weight for different loss functions. + +Model Distillation In Equation 7, we introduce a student model to assist in training the generator to produce data with greater diversity. A better student model should lead to a better generator. Therefore, after training the generator for several rounds, we subsequently train the student model using the saved synthesized data and the output of the teacher model, using KL loss for knowledge distillation: + +$$ +\mathcal {L} _ {S} = K L \left(\theta_ {k - 1} (\hat {x}), \theta_ {S} (\hat {x})\right). \tag {10} +$$ + +In this way, we can train a student model with better performance, and then further use it to update the generator. An ideal synthetic dataset should be able to efficiently enable student $\theta_{S}$ to fully learn the knowledge of teacher model. + +Note that when the training of the whole process is over (i.e. the student model can use the synthesized data to obtain high performance), we only retain the synthetic dataset $X_{syn}$ and transfer it to the clients, without saving the generator and student model. + +# 4.3. Client Side: Update with Global Information + +On the client side, we can obtain the data synthesized for the previous task $X_{syn}$ and the real training data of the current task $X_{local}$ , then we train the local model $\theta_k$ for task $k$ on the two datasets at the same time. We showed in Section 3.3 that the use of global models and global data can alleviate forgetting. Thus we distill the knowledge of global teacher model and global synthetic data by minimizing the following objective function, + +$$ +\mathcal {L} _ {\text {c l i e n t}} = \underbrace {C E \left(\theta_ {k} (x) , y\right)} _ {\text {f o r c u r r e n t t a k}} + \alpha \cdot \underbrace {K L \left(\theta_ {k - 1} (\hat {x}) , \theta_ {k} (\hat {x})\right)} _ {\text {f o r p r e v i o u s t a k s}}, \tag {11} +$$ + +where $(x,y)\in X_{local}$ and $(\hat{x})\in X_{syn}$ . The utilization of the distillation loss facilitates efficient transfer of knowledge from the previous task to the current task model. And $\alpha$ is a hyper-parameter that controls the strength of the regularization for the previous tasks. + +# 5. Experiments + +# 5.1. Experimental Settings + +We experiment on two datasets, namely CIFAR-100 [25], and Tiny-ImageNet [26], to evaluate the performance of our proposed approach. To establish the order of the continual tasks, we adopt the widely used protocols [49, 61, 42, 7]. Specifically, we divide all classes of each dataset equally into multiple tasks by default, i.e. we evenly divide the classes into 5 and 10 tasks to simulate class continual learning scenarios. We employ ResNet18 [18] as the backbone for the classification model. + +To evaluate our approach, we employ the standard continual learning metrics, as used in prior works [23, 22, 47, 16], which include average accuracy across all tasks and a forgetting measure [6] (see Equation 2). For a fair comparison with the baseline class continual learning methods in the FCCL setting, we implement three types of baselines: 1) Finetune, in which each client simply learns tasks in sequence; 2) FedWeIT [58], a regularization-based method in Federated Continual Learning that maximizes the knowledge transfer between clients; 3) Examples of typical continual learning methods that do not store training data for rehearsal, including EWC [27] and LwF [30]. In addition, we compare our method with methods that store real training data of old tasks, such as iCaRL [44]. We implement these traditional continual learning algorithms in the FCCL scenario and name them as FedEWC, FedLwF, and + +Table 2: The Average Accuracy (%) and Forgetting for all learned tasks on CIFAR-100 for various numbers of tasks (5, 10) under both IID and non-IID settings. Results are reported as an average of 3 runs. 'Acc' refers to average accuracy, and 'F' represents the forgetting measure utilized in Equation 2. The best results are in bold. + +
Data partitionIIDNIID (1)NIID (0.5)
TasksT=5T=10T=5T=10T=5T=10
MethodAcc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)
Finetune16.120.787.830.7516.330.778.450.7415.490.747.640.71
FedEWC16.510.718.010.6516.060.688.840.6216.860.668.040.65
FedWeIT28.450.5220.390.4328.560.4919.680.4524.570.5415.450.48
FedLwF30.610.4523.270.3730.940.4221.160.4127.590.4417.980.45
Ours36.310.2224.760.2634.890.2422.850.2633.330.2720.710.29
+ +![](images/c12a0c1865af8eaa726df1e79316f0b03052ac6b238e65dcf404865566e323de.jpg) +Figure 4: Average accuracy on previous tasks and current task after the model trained on current task. + +![](images/aa93cc02ca34ab091c357fd66b8a5728e8cdeeebc701e9103454e3f65048d25f.jpg) + +FedIcaRL. For detailed information on the task configuration, default hyper-parameters and additional experimental results, please refer to the Appendix. + +# 5.2. Experiments on CIFAR-100 + +For CIFAR-100 dataset, we conduct experiments on two sets of tasks consisting of 5 and 10 tasks, respectively. We run the experiment on both IID and non-IID scenarios. For non-IID setting, the Dirichlet parameter is set to 0.5 and 1, i.e. $\mathrm{NIID}(0.5)$ and $\mathrm{NIID}(1)$ . Table 2 shows the final average accuracy of the FL model trained on all tasks, along with the corresponding forgetting measure for each experiment. It is important to emphasize that an optimal method is characterized by high average accuracy and low forgetting measure. + +Table 2 indicates that Finetune exhibits the poorest performance when attempting to learn the continuously incoming tasks sequentially, thereby experiencing the issue of catastrophic forgetting. Moreover, we observed that methods based on regularization constraints, such as FedEWC and FedWeIT, were often ineffective in preventing the model from forgetting old tasks due to the lack of available data. On the other hand, we found that distillation-based approaches such as FedLwF and our proposed method were + +capable of improving the final average accuracy while simultaneously mitigating the issue of catastrophic forgetting. This is due to the transfer of model knowledge learned from the old task to the new task in a distillation manner, enabling the model to prevent catastrophic forgetting. Obviously, we found that applying our proposed method to generate synthetic datasets for the federated models trained on old tasks and subsequently performing model distillation on these synthetic datasets can lead to a substantial improvement in the average accuracy and reduction in the forgetting measure. This observation underscores the ability of our synthetic data to capture the distribution characteristics of historical task data accurately. For example, in IID setting, when partitioning the CIFAR-100 dataset into five tasks, our method achieves an accuracy of $36.31\%$ , which is about $6\%$ higher than the best baseline method FedLwF. It is worth noting that we observed a decrease in the average accuracy of all methods to varying degrees as the number of tasks increased from 5 tasks to 10 tasks due to increased task complexity and the associated forgetting phenomenon. Nonetheless, our proposed method maintains the highest average accuracy and the lowest forgetting measure even under these more demanding conditions. + +Figure 4 illustrates the performance of the models trained by various methods on all previously learned tasks after the completion of each task. Specifically, it shows the average accuracy of the model on both the current and previous tasks after the completion of each task (e.g., after learning the second task, the average accuracy of the model on both the first and second tasks is measured). Based on these curves, it is evident that our proposed model outperforms other competing baseline methods in all incremental tasks, regardless of the number of tasks involved. This finding underscores the effectiveness of our approach in facilitating multiple local clients to learn new classes in a streaming manner while mitigating the forgetting problem. + +![](images/67f20052159e796ecaa4a9ff6da8e20226eb69711fddcd68d8ab91057659124b.jpg) +Figure 5: The average accuracy $(\%)$ and Forgetting for all learned tasks on Tiny-ImageNet for 5 tasks under both IID and non-IID settings. + +# 5.3. Experiments on Tiny-ImageNet + +We also evaluated the performance of our proposed method on the more challenging Tiny-ImageNet dataset and obtained similar results to those observed in the CIFAR-100 experiments. Specifically, in Figure 5, we present the final average accuracy and forgetting measure for all learned tasks in both IID and non-IID settings for the case of 5 tasks. The Dirichlet parameter is set to $\{0.05, 0.1, 0.5, 1\}$ . Based on Figure 5, it is evident that our proposed method consistently outperforms FedLwF in terms of average accuracy across all data partitions. Moreover, our method demonstrates a significantly lower forgetting measure than FedLwF under both IID and non-IID settings. Our proposed method achieves an average accuracy that is approximately $3\%$ higher than that of FedLwF even in the most challenging scenario (i.e. NIID(0.05)). This result highlights the effectiveness of our method in mitigating catastrophic forgetting in the presence of extreme data distributions. + +# 5.4. Comparison with Exemplar-based Method + +In CL, the most successful approaches to alleviate forgetting require extensive replay of previously seen data, which can be problematic when data legality and privacy concerns exist. Among them, iCaRL [44] is a classic but unrealistic algorithm that relies on the stored exemplars in addition to the network parameters, and it is intuitive that using old task's real data could be beneficial to alleviate forgetting problem. To further understand the performance gap, we compare our proposed TARGET (which uses synthetic data) with iCaRL (which requires storing exemplars from old task's real data), and study the effects of different exemplar memories on the performance of our method and iCaRL in Figure 6. We set the stored exemplar size to $\{1000, 1500, 2000\}$ for iCaRL, and $\{2000, 3000\}$ for our method on the CIFAR-100 dataset. + +The accuracy curve in Figure 6 represents the average accuracy rate measured over all 100 classes learned by the model during the last learning task. While our method + +![](images/b360c9d98595787fb5e505121748611600bc34d21883b00e2424542a0a2795f3.jpg) +Figure 6: Accuracy comparison between real data-based method (iCaRL) and synthetic data-based method (TARGET). We show the result for the last task. + +Table 3: The effect of $\alpha$ on the performance of both new and old tasks in CIFAR-100, 2 tasks. + +
α35101525
Old Task (0-49)28.3434.9149.0258.4866.02
New Task (50-99)62.4660.2653.7445.2629.32
Average (0-99)45.4147.5851.3851.8747.76
+ +can achieve similar performance to storing 1k real training data by storing 2k synthetic data, it still cannot outperform storing 2k real training data. However, it is worth noting that our method does not require storing any real training data, which can be a significant advantage in scenarios where storing real data is difficult or not allowed due to privacy or legal concerns. Additionally, our method achieves better performance than storing only 1k real training data, which indicates that our synthetic data is effective in mitigating catastrophic forgetting. We observed that when our method stores 3k synthetic data, it achieves better accuracy than when it stores 2k synthetic data. However, surpassing iCaRL in performance with an equal amount of data remains a challenge for our method. How to effectively use fewer synthetic data with more valuable knowledge from previous tasks will be left as a future research direction. + +# 5.5. Analysis of Our Method + +Trade-off between Backward and Forward Transfer. Continual learning presents a challenge in balancing the trade-off between maintaining high accuracy on old tasks (backwards transfer) and achieving high accuracy on new tasks (forward transfer). In Table 3, we evenly split the CIFAR-100 dataset into two tasks and test the average accuracy of our proposed method under different values of $\alpha$ . We observe that the trade-off between backward and forward transfer is not always balanced. In order to achieve good backward transfer, a large value of $\alpha$ should be used + +![](images/eed01bfff92f3aafd5332161b6ad73413ffa3b58b50464ebaa77e341b7be11a9.jpg) +Figure 7: Accuracy for different size of synthetic dataset. + +![](images/b7529b6618ef3d10da9af403483a8a9be090a985a8d10ab89da2872293ab0394.jpg) +Figure 9: Distillation results for task $\{1,2,3,4\}$ when trained on 5 tasks. + +![](images/4843a71ba65577c05f88d66f719d9b12961e4fcfea09c2409ecd829015e481d3.jpg) +Figure 8: Visualization of randomly synthesised data. + +to prevent the model from forgetting previous tasks and to encourage it to focus more on the old synthetic data. Conversely, to achieve good forward transfer, a small value of $\alpha$ should be used, allowing the model to learn quickly and effectively from new tasks while still maintaining some knowledge from the old tasks. The experimental results in Table 3 indicate that when the value of $\alpha$ is between 10-15, the model achieves a good balance between accuracy on new tasks and accuracy on old tasks. We also partition the CIFAR-100 dataset into 5 equal tasks and test our method's performance on all previous tasks after learning each new task (refer to Appendix). + +Memory of Synthetic Data. The optimal amount of synthetic data generated for the old task plays a crucial role in determining the final performance of our method. Insufficient synthetic data may not adequately facilitate knowledge transfer from the old tasks, while excessive data generation can lead to increased memory and communication costs. Therefore, determining an appropriate amount of synthetic data that strikes a balance between knowledge transfer and computational cost is critical for the effectiveness and efficiency of our approach. As shown in Figure 7, we test our method on CIFAR-100, which is divided into 5 tasks with different data sizes ranging from 2k to 16k. It can be observed that when the data volume is relatively small, such as 2k and 4k, the performance of the model is poor, especially when the data volume is 2k, the testing curve in task 4 shows a later decline. This is because when the data volume is too small, the model is unable to effectively learn + +![](images/21927968055e9a469543b13752cbbcf6da33269afa682eda7d4b04182252cd35.jpg) + +![](images/decc23f26b9566ccb4f5e2360c30d75fb876661377008b90c6e87f3314be8533.jpg) + +knowledge from old tasks. Increasing the data volume to 8k can effectively alleviate the forgetting phenomenon and achieve good performance. However, continuously increasing the data volume to 12k and 16k do not result in significant improvement in the model's performance. It is important to note that the size of the data volume alone does not guarantee the effectiveness of synthetic data in improving machine learning models. The quality and relevance of the synthetic data must also be carefully considered to ensure that it accurately represents the underlying distribution of the real-world data. + +Visualization on Synthetic Data. To demonstrate the effectiveness of the synthetic data, we present in Figure 8 the visualization results of the synthesized images generated by our method after the model learns the penultimate task on Tiny-ImageNet (CIFAR-100 in Appendix) for 5 tasks. These data have the potential to efficiently enable the student model to approach the performance of the teacher model rapidly. However, it is worth noting that these data exhibit visual dissimilarities from the actual training data. + +Analysis on Distillation. Based on the synthetic data, we distill the knowledge of the model trained on old tasks into a student model. The effect of distillation is further demonstrated in Figure 9. It can be observed that even without accessing any private training data from clients, our method can quickly distill the student model on the server to approach the performance of the global model. + +# 6. Conclusion + +In conclusion, this paper introduces a novel method, TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillage), to alleviate the catastrophic forgetting problem in Federated Class-Continual Learning (FCCL). Unlike all the previous methods, our proposed method leverages global knowledge, without requiring any additional datasets or data from previous tasks, making it ideal for privacy-sensitive scenarios. Extensive experimental results demonstrate the effectiveness of our proposed method in comparison to existing FCCL methods. + +# References + +[1] Eden Belouadah and Adrian Popescu. Il2m: Class incremental learning with dual memory. In Proceedings of the IEEE/CVF international conference on computer vision, pages 583-592, 2019. 1 +[2] Cecilia Bjursell. The pandemic as disjuncture: Lifelong learning in a context of fear. International Review of Education, 66(5-6):673-689, 2020. 1 +[3] Thang D Bui, Cuong V Nguyen, Siddharth Swaroop, and Richard E Turner. Partitioned variational inference: A unified framework encompassing federated and continual learning. arXiv preprint arXiv:1811.11206, 2018. 1 +[4] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio P. Calmon, and Taesup Moon. Cpr: classifier-projection regularization for continual learning. arXiv preprint arXiv:2006.07326, 2020. 3 +[5] Yatin Chaudhary, Pranav Rai, Matthias Schubert, Hinrich Schütze, and Pankaj Gupta. Federated continual learning for text classification via selective inter-client transfer. arXiv preprint arXiv:2210.06101, 2022. 2 +[6] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532-547, 2018. 3, 6 +[7] Arslan Chaudhry, Naemullah Khan, Puneet Dokania, and Philip Torr. Continual learning in low-rank orthogonal subspaces. Advances in Neural Information Processing Systems, 33:9900-9911, 2020. 3, 6 +[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2, 3 +[9] Marco Ciotti, Massimo Ciccozzi, Alessandro Terrinoni, Wen-Can Jiang, Cheng-Bin Wang, and Sergio Bernardini. The COVID-19 pandemic. Critical reviews in clinical laboratory sciences, 57(6):365-388, 2020. 1 +[10] Jiahua Dong, Yang Cong, Gan Sun, Zhen Fang, and Zhengming Ding. Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1(2):1-18, 2021. 2 +[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1 +[12] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 2 +[13] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 86-102. Springer, 2020. 1 + +[14] Fernando Estevez Casado, Dylan Lema Pais, Marcos Fernández Criado, Roberto Iglesias Rodríguez, Carlos Vázquez Regueiro, and Senen Barro Ameneiro. Concept drift detection and adaptation for federated and continual learning. 2021. 2 +[15] Gongfan Fang, Kanya Mo, Xinchao Wang, Jie Song, Shitao Bei, Haofei Zhang, and Mingli Song. Up to 100x faster data-free knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 6597-6604, 2022. 6 +[16] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013. 1, 3, 6 +[17] Andrew Hard, Kanishka Rao, Rajiv Mathews, Swaroop Ramaswamy, Françoise Beaufays, Sean Augenstein, Hubert Eichner, Chloe Kiddon, and Daniel Ramage. Federated learning for mobile keyboard prediction. arXiv preprint arXiv:1811.03604, 2018. 1 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[19] Ziyue Jiang, Yi Ren, Ming Lei, and Zhou Zhao. Fed-speech: Federated text-to-speech with continual learning. arXiv preprint arXiv:2110.07216, 2021. 1 +[20] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and Trends® in Machine Learning, 14(1-2):1-210, 2021. 1 +[21] Georgios Kaissis, Alexander Ziller, Jonathan Passerat-Palmbach, Theo Ryffel, Dmitrii Usynin, Andrew Trask, Ionésio Lima Jr, Jason Mancuso, Friederike Jungmann, Marc-Matthias Steinborn, et al. End-to-end privacy preserving deep learning on multi-institutional medical imaging. Nature Machine Intelligence, 3(6):473-484, 2021. 2 +[22] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 1, 3, 6 +[23] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 1, 2, 3, 6 +[24] Jakub Konečny, Brendan McMahan, and Daniel Ramage. Federated optimization: Distributed optimization beyond the datacenter. arXiv preprint arXiv:1511.03575, 2015. 2 +[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 3, 6 +[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6 +[27] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forget- + +ting by incremental moment matching. Advances in neural information processing systems, 30, 2017. 2, 6 +[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1, 4 +[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. arXiv preprint arXiv:1812.06127, 2018. 3, 4 +[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 4, 6 +[31] Zexi Li, Qunwei Li, Yi Zhou, Wenliang Zhong, Guannan Zhang, and Chao Wu. Edge-cloud collaborative learning with federated and centralized features. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '23, page 1949-1953, New York, NY, USA, 2023. Association for Computing Machinery. 2 +[32] Zexi Li, Tao Lin, Xinyi Shang, and Chao Wu. Revisiting weighted aggregation in federated learning with neural networks. In International Conference on Machine Learning. PMLR, 2023. 2 +[33] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Rmm: Reinforced memory management for class-incremental learning. Advances in Neural Information Processing Systems, 34:3478-3490, 2021. 1 +[34] Yuang Liu, Wei Zhang, Jun Wang, and Jianyong Wang. Data-free knowledge transfer: A survey. arXiv preprint arXiv:2112.15278, 2021. 2 +[35] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2 +[36] Yuhang Ma, Zhongle Xie, Jue Wang, Ke Chen, and Lidan Shou. Continual federated learning based on knowledge distillation. In Lud De Raedt, editor, Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 2182–2188. International Joint Conferences on Artificial Intelligence Organization, 7 2022. Main Track. 1 +[37] Yuhang Ma, Zhongle Xie, Jue Wang, Ke Chen, and Lidan Shou. Continual federated learning based on knowledge distillation. In Luc De Raedt, editor, Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, pages 2182-2188. ijcai.org, 2022. 2 +[38] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Aarti Singh and Jerry Zhu, editors, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, volume 54 of Proceedings of Machine Learning Research, pages 1273-1282. PMLR, 20-22 Apr 2017. 1, 4 +[39] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. + +In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 2 +[40] Sudhanshu Mittal, Silvio Galesso, and Thomas Brox. Essentials for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3513-3522, 2021. 1 +[41] Alexander Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. 2015. 3 +[42] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural networks, 113:54-71, 2019. 1, 6 +[43] Tae Jin Park, Kenichi Kumatani, and Dimitrios Dimitriadis. Tackling dynamics in federated incremental learning with variational embedding rehearsal. arXiv preprint arXiv:2110.09695, 2021. 2 +[44] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 2, 4, 6, 8 +[45] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. arXiv preprint arXiv:1810.11910, 2018. 2 +[46] Amanda Rios and Laurent Itti. Closed-loop gan for continual learning. arXiv preprint arXiv:1811.01146, 2018. 2 +[47] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 1, 3, 6 +[48] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International Conference on Machine Learning, pages 4548-4557. PMLR, 2018. 1 +[49] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 1, 2, 6 +[50] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9374-9384, 2021. 2, 3 +[51] Sebastian Thrun. Lifelong learning algorithms. Learning to learn, 8:181-209, 1998. 1 +[52] Anastasiia Usmanova, François Portet, Philippe Lalanda, and German Vega. A distillation-based approach integrating continual learning and federated learning for pervasive services. arXiv preprint arXiv:2109.04197, 2021. 2 +[53] Anamaria Vizitiu, Cosmin Loan Nita, Andrei Puiu, Constantin Suciu, and Lucian Mihai Itu. Towards privacy-preserving deep learning based medical imaging applications. In 2019 IEEE international symposium on medical measurements and applications (MeMeA), pages 1-6. IEEE, 2019. 2 +[54] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In Computer Vision-ECCV 2022: 17th + +European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV, pages 398-414. Springer, 2022. 1 +[55] Ju Xu and Zhanxing Zhu. Reinforced continual learning. Advances in Neural Information Processing Systems, 31, 2018. 2 +[56] Li Yang, Shasha Liu, Jinyan Liu, Zhixin Zhang, Xiaochun Wan, Bo Huang, Youhai Chen, and Yi Zhang. Covid-19: immunopathogenesis and immunotherapeutics. Signal transduction and targeted therapy, 5(1):128, 2020. 1 +[57] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715-8724, 2020. 3, 6 +[58] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086. PMLR, 2021. 2, 6 +[59] Jaehong Yoon, Saehoon Kim, Eunho Yang, and Sung Ju Hwang. Scalable and order-robust continual learning with additive parameter decomposition. arXiv preprint arXiv:1902.09432, 2019. 3 +[60] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 2 +[61] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987-3995. PMLR, 2017. 1, 2, 6 +[62] Jie Zhang, Chen Chen, Bo Li, Lingjuan Lyu, Shuang Wu, Shouhong Ding, Chunhua Shen, and Chao Wu. Dense: Data-free one-shot federated learning. In Advances in Neural Information Processing Systems. 6 +[63] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In International Conference on Machine Learning, pages 26311-26329. PMLR, 2022. 3 +[64] Junting Zhang, Jie Zhang, Shalini Ghosh, Dawei Li, Serafettin Tasci, Larry Heck, Heming Zhang, and C-C Jay Kuo. Class-incremental learning via deep model consolidation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1131-1140, 2020. 1 +[65] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13208-13217, 2020. 1 +[66] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1 +[67] Didi Zhu, Yinchuan Li, Yunfeng Shao, Jianye Hao, Fei Wu, Kun Kuang, Jun Xiao, and Chao Wu. Generalized universal domain adaptation with generative flow networks. arXiv preprint arXiv:2305.04466, 2023. 2 + +[68] Didi Zhu, Yincuan Li, Junkun Yuan, Zexi Li, Yunfeng Shao, Kun Kuang, and Chao Wu. Universal domain adaptation via compressive attention matching. arXiv preprint arXiv:2304.11862, 2023. 2 \ No newline at end of file diff --git a/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/images.zip b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..43ca0978b97ed2afd6c8e5e5bab3ecaac81feff6 --- /dev/null +++ b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:226a5693f42d594cd25c08ef5cc50067d3455834bdb1d139e10dd49ccd5e3b4c +size 485612 diff --git a/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/layout.json b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2e663aca3cde09c4b5223e4db79c30c30e3d6a99 --- /dev/null +++ b/2023/TARGET_ Federated Class-Continual Learning via Exemplar-Free Distillation/layout.json @@ -0,0 +1,10601 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 539, + 121 + ], + "type": "text", + "content": "TARGET: Federated Class-Continual Learning via Exemplar-Free Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 144, + 374, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 144, + 374, + 185 + ], + "spans": [ + { + "bbox": [ + 220, + 144, + 374, + 185 + ], + "type": "text", + "content": "Jie Zhang* \nETH Zurich \nzj.jayzhang@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 205, + 418, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 205, + 418, + 233 + ], + "spans": [ + { + "bbox": [ + 179, + 205, + 418, + 233 + ], + "type": "text", + "content": "Chen Chen & Weiming Zhuang & Lingjuan Lyu† \nSony AI" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 234, + 476, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 234, + 476, + 248 + ], + "spans": [ + { + "bbox": [ + 117, + 234, + 476, + 248 + ], + "type": "text", + "content": "{ChenA.Chen, weiming.zhuang, Lingjuan.Lv}@sony.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 274, + 192, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 274, + 192, + 288 + ], + "spans": [ + { + "bbox": [ + 143, + 274, + 192, + 288 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 300, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 290, + 529 + ], + "type": "text", + "content": "This paper focuses on an under-explored yet important problem: Federated Class-Continual Learning (FCCL), where new classes are dynamically added in federated learning. Existing FCCL works suffer from various limitations, such as requiring additional datasets or storing the private data from previous tasks. In response, we first demonstrate that non-IID data exacerbates catastrophic forgetting issue in FL. Then we propose a novel method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free distTillage), which alleviates catastrophic forgetting in FCCL while preserving client data privacy. Our proposed method leverages the previously trained global model to transfer knowledge of old tasks to the current task at the model level. Moreover, a generator is trained to produce synthetic data to simulate the global distribution of data on each client at the data level. Compared to previous FCCL methods, TARGET does not require any additional datasets or storing real data from previous tasks, which makes it ideal for data-sensitive scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": "Federated Learning (FL) is a privacy-aware learning paradigm that facilitates collaborations among multiple entities (e.g., edge devices or organizations) [38, 20, 28, 66, 17]. Each entity or client in FL retains data locally and transfers only training updates to the central server for aggregation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 647, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 682 + ], + "type": "text", + "content": "Conventional FL studies assume that the data classes and domains are static, but the new classes could emerge and data domains could change over time in reality [64," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 276, + 547, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 276, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 547, + 407 + ], + "type": "text", + "content": "40, 1, 65]. For example, multiple health institutions could use FL to collaborate and train models to identify COVID-19 [56, 9, 2] strains; new COVID-19 strains, however, continue to emerge due to high mutation rate of virus. An intuitive solution to this issue of continuously emerging data classes is training new models from scratch, but this is impractical as it would require significant extra computation cost. Another method is transfer learning from the previously trained model, but this method suffers from catastrophic forgetting [23, 22, 47, 16], degrading performance on the previous classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 411, + 547, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 530 + ], + "type": "text", + "content": "To address these issues, recent research [36, 11, 3, 19] has introduced the concept of Continual Learning (CL) [61, 42, 51, 49, 48] within the FL framework. These methods, collectively referred to as Federated Continual Learning (FCL), aim to mitigate the problems of catastrophic forgetting in FL. In most FCL scenarios, new classes are dynamically added, which we call Federated Class-Continual Learning (FCCL). FCCL allows local clients to continuously collect new data, and new classes can be added at any time." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 548, + 714 + ], + "type": "text", + "content": "Unfortunately, existing FCCL works suffer from various limitations. For example, Ma et al. [36] utilize an unlabeled surrogate dataset to address the catastrophic forgetting problem, which may be difficult to obtain in some data sensitive scenarios. Furthermore, the usage of an unlabeled surrogate dataset may not be ideal for certain types of data, as it may not capture the full complexity of the original data. In CL, exemplar-based methods [44, 33, 54, 13] have achieved leading performance. An exemplar refers to a sample or instance of a previously seen data point that is retained in a memory buffer for future use in the learning process. Dong et al. [11] propose a exemplar-based method that stores historical data to address catastrophic forgetting. However, in many privacy-sensitive scenarios (e.g., hospitals and medical research institutions), users are" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 194, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 194, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 194, + 703 + ], + "type": "text", + "content": "*Work done during internship at Sony AI." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 136, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4782" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "not permitted to store data from previous tasks due to privacy and policy concerns and data will not be kept for a long time [21, 53, 34, 50]. In summary, the majority of FCCL methods train the global model with additional datasets or previous task data, which could potentially violate data privacy regulations. This dilemma prompts us to consider the following question:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 157, + 288, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 157, + 288, + 193 + ], + "spans": [ + { + "bbox": [ + 45, + 157, + 288, + 193 + ], + "type": "text", + "content": "Question: How to effectively alleviate the catastrophic forgetting problem in the FCCL without storing the local private data of the client or any additional datasets?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 287, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 349 + ], + "type": "text", + "content": "To address this question, we conduct a systematic analysis and observe that the imbalanced distribution of data among clients in FL exacerbates the catastrophic forgetting problem (see Section 3.3). In order to fix this problem: 1) at the model level, we leverage the previously trained global model to transfer knowledge of the old tasks to the current task. 2) at the data level, we train a generator to produce synthetic data that aims to simulate the global distribution of data on each client. Drawing on these insights, we present a method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillation) that mitigates catastrophic forgetting in FCCL without compromising clients' data privacy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 350, + 249, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 350, + 249, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 350, + 249, + 361 + ], + "type": "text", + "content": "Our contributions can be concluded as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 362, + 287, + 529 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 362, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 362, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 58, + 362, + 287, + 432 + ], + "type": "text", + "content": "- We are the first to demonstrate that non-independent and identically distributed (non-IID) data exacerbates catastrophic forgetting issue in FL. Then we propose a novel method called TARGET, which alleviates the catastrophic forgetting in FCCL by leveraging global information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 434, + 286, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 434, + 286, + 469 + ], + "spans": [ + { + "bbox": [ + 58, + 434, + 286, + 469 + ], + "type": "text", + "content": "- Compared to previous FCCL methods, TARGET doesn't require extra datasets or data from previous tasks, it can be applied in data sensitive scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 470, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 470, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 58, + 470, + 286, + 529 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the efficacy of our proposed method. For example, when partitioning the CIFAR-100 dataset into five tasks, our method achieves an accuracy of " + }, + { + "bbox": [ + 58, + 470, + 286, + 529 + ], + "type": "inline_equation", + "content": "36.31\\%" + }, + { + "bbox": [ + 58, + 470, + 286, + 529 + ], + "type": "text", + "content": ", which is about " + }, + { + "bbox": [ + 58, + 470, + 286, + 529 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 58, + 470, + 286, + 529 + ], + "type": "text", + "content": " higher than the best baseline method." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 542, + 134, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 134, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 134, + 555 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 563, + 212, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 212, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 212, + 575 + ], + "type": "text", + "content": "2.1. Federated Continual Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "In recent years, deep neural networks (DNNs) have been widely used as a fundamental technology for the development of artificial intelligence, both in established and emerging fields [68, 67, 32, 31, 10]. With the advancement of deep learning, an increasing number of researchers have started to focus on the construction of privacy-preserving deep learning frameworks. FL is a paradigm for collaboratively building a model across multiple clients [24, 39], which is gaining momentum in recent years. However, Federated Continual Learning (FCL), which focused on both FL and CL simultaneously is just emerging and still" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 407 + ], + "type": "text", + "content": "remains pending further research [52, 43, 58]. Apart from the client-wise catastrophic forgetting, FCL paradigm also poses new challenges such as inter-client interference and communication-efficiency [58]. FedWeIT solves these challenges through decomposing the parameters into three parts, i.e., global parameters, local based parameters, and task-adaptive parameters [58]. Concurrently, Concept-Drift-Aware Federated Averaging (CDA-FedAvg) extends the popular FL algorithm, Federated Averaging (FedAvg), to tackle the CL problem by introducing the concept drift detection and adaptation [14]. FCL with Distillation (CFeD) treats the model trained with the last task as the teacher model to perform a knowledge distillation and proposes a server distillation mechanism to deal with non-i.i.d. issue [37]. Global-Local Forgetting Compensation (GLFC) designs a class-aware gradient compensation loss and a class-semantic relation distillation loss to prevent forgetting, and a proxy server to mitigate the non-independent and identically distributed (non-IID) problem [12]. Additionally, Federated Selective Inter-client Transfer (FedSeIT) applies FCL to NLP through selectively combining model parameters of foreign clients and selecting informative tasks to perform knowledge transfer [5]. These papers illustrate the growing interest in FCL and the need for novel approaches to address its unique challenges. As research in this area continues to advance, we can expect to see more innovative approaches to FCL that further improve its performance, scalability, and privacy preservation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 418, + 422, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 422, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 422, + 431 + ], + "type": "text", + "content": "2.2. Continual Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": "Continual Learning has been studied extensively, several training methods have been proposed to address the catastrophic forgetting challenge it presents. Regularization-based approaches: Elastic Weight Consolidation (EWC) selectively penalizes the network parameters that are important for old tasks [23]. Synaptic intelligence (SI) uses a memory buffer to store important network parameters [61]. Incremental moment matching (IMM) modelings the posterior distribution after learning multiple tasks as a mixture of Gaussian models [27]. Stable SGD proposes to carefully design the training regimes such as learning rate decay, batch size, dropout, and optimizer to alleviate forgetting. Replay-based approaches: The methods in this family construct a memory to store the past information which will be presented to the model for reviewing in future tasks. Some store the knowledge of the previous tasks, known as experience replay [35, 45, 49, 46]. iCaRL preserves the most representative samples of each class [44]. Averaged Gradient Episodic Memory (A-GEM) builds an episodic memory of model parameter gradients[8]. Architecture-based: DEN expands the model size [60] and RCL utilizes reinforcement learning [55]. APD divides the model parameters into shared and task-specific parameters to restrict the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4783" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 144, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 144, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 144, + 84 + ], + "type": "text", + "content": "model complexity [59]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 100, + 289, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 100, + 289, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 100, + 289, + 244 + ], + "type": "text", + "content": "Exemplar-Free Continual Learning Besides, a promising line of work focuses on data-free Continual Learning. DeepDream perturbs current training samples into images that maximize \"forgetting\" from the previous tasks [41]. DeepInversion proposes a model inversion technique and evaluates its performance in a Continual Learning scenario but found limited success [57]. DFCIL [50] investigates the failure and decomposes the CE-loss into two different losses which guarantee the learning of effective features. These methods enable Continual Learning in scenarios where storing old task data is not feasible or desirable, such as in privacy-sensitive applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 255, + 234, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 234, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 234, + 269 + ], + "type": "text", + "content": "3. Catastrophic Forgetting in FCCL" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 276, + 287, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 336 + ], + "type": "text", + "content": "This section conducts an in-depth analysis of catastrophic forgetting problem in federated class-continual learning (FCCL). We start by providing a formal definition of the problem. Then, we investigate the forgetting issue in FCCL and discuss potential methods to mitigate it." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 343, + 160, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 160, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 160, + 356 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "text", + "content": "Federated Class-Continual Learning (FCCL) focuses on the problem of learning models for new classes over time in FL. An FCCL framework consists of a central server and multiple clients. All clients do not share their raw data with any other client or the central server. Each client learns from a sequence of " + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "text", + "content": " tasks, where " + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "text", + "content": "-th task contains non-overlapping subsets of classes " + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "inline_equation", + "content": "C_k \\in C" + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "text", + "content": " is the set of all possible classes. In our privacy-aware scenario, the task stream is presented in an unknown order, and each client can only access its local data from task " + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "text", + "content": " during that task's training period, which is no longer accessible thereafter. Note that the models are trained in a distributed manner, where each party has access to only a subset of the classes " + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 46, + 362, + 289, + 567 + ], + "type": "text", + "content": " (i.e. non-IID). We also consider a more challenging and practical setting where the data in each client is heterogeneous. In this paper, we assume the label distribution of data in each client is skewed [29, 63]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Forgetting Issue in FCCL The goal of the global model optimization problem at task " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is to minimize the overall classification error on the current set of classes " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". However, when a new task arises, clients are not able to access data from previous (old) tasks due to privacy concerns and can only update their local model with data from the new task. This often leads to a significant decrease in performance on previous tasks, which is known as catastrophic forgetting [23, 22, 47, 16]. To mitigate catastrophic forgetting in the global model, we aim to minimize the overall classification error on the current set of classes " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", while" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 78, + 425, + 182 + ], + "blocks": [ + { + "bbox": [ + 322, + 78, + 425, + 182 + ], + "lines": [ + { + "bbox": [ + 322, + 78, + 425, + 182 + ], + "spans": [ + { + "bbox": [ + 322, + 78, + 425, + 182 + ], + "type": "image", + "image_path": "d92f5b7b0a003af57b65bca690ac044f6a491b05b0137b553fbe616f8cfa6720.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 196, + 402, + 205 + ], + "lines": [ + { + "bbox": [ + 342, + 196, + 402, + 205 + ], + "spans": [ + { + "bbox": [ + 342, + 196, + 402, + 205 + ], + "type": "text", + "content": "(a) Test Accuracy" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 218, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 304, + 218, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 547, + 291 + ], + "type": "text", + "content": "Figure 1: Forgetting under different data partitions. We partitioned the CIFAR-100 dataset into five tasks with 20 classes each, distributed among five clients. \"NIID\" refers to non-IID, where a lower value represents a more imbalanced or skewed distribution of data. Non-IIDness refers to the Dirichlet parameter." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 430, + 78, + 536, + 182 + ], + "blocks": [ + { + "bbox": [ + 430, + 78, + 536, + 182 + ], + "lines": [ + { + "bbox": [ + 430, + 78, + 536, + 182 + ], + "spans": [ + { + "bbox": [ + 430, + 78, + 536, + 182 + ], + "type": "image", + "image_path": "db207368bef2e6b47c3e0075a06635eef4acaa8dc916ecf0c50c63b54904634a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 196, + 518, + 205 + ], + "lines": [ + { + "bbox": [ + 443, + 196, + 518, + 205 + ], + "spans": [ + { + "bbox": [ + 443, + 196, + 518, + 205 + ], + "type": "text", + "content": "(b) Relative Forgetting" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 310, + 546, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 546, + 346 + ], + "type": "text", + "content": "simultaneously minimizing the changes to the previously learned classes. Formally, the objective function can be written as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 350, + 545, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 350, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 321, + 350, + 545, + 384 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {k}} \\sum_ {c \\in C _ {k}} \\sum_ {i = 1} ^ {m _ {c}} L \\left(f _ {k} \\left(x _ {i, c}; \\theta_ {k}\\right), c\\right) + \\alpha R \\left(\\theta_ {k}, \\theta_ {k - 1}\\right) \\tag {1}", + "image_path": "a2b775663d5eacc8dfce0e2f2f7b49574e66b10c1a3f6fc0f0bc09d4efe13e2f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\theta_{k}" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " is the model parameter at round " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " is a loss function that measures the classification error, " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " is a regularization term that penalizes changes to the previous model parameters, " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "m_{c}" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " is the number of data in class " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " is a hyper-parameter that controls the strength of the regularization. In this formula, " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "f_{k}(x_{i,c};\\theta_{k})" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " represents the classification model that takes as input a data point " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "x_{i,c}" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " associated with class " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " and outputs a probability distribution over the set of classes in " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": ". The regularization term " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " encourages the new model parameters to be close to the previous model parameters " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\theta_{k - 1}" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": ", in order to prevent catastrophic forgetting of the previously learned classes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 539, + 536, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 539, + 536, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 539, + 536, + 553 + ], + "type": "text", + "content": "3.2. Heterogeneous Data Exacerbates Forgetting" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 641 + ], + "type": "text", + "content": "We argue that the degree of data heterogeneity has a substantial impact on catastrophic forgetting. To verify this, we conduct an experiment on CIFAR-100 dataset [25] with different degrees of data heterogeneity. Inspired by Backward Transfer (BwT) [6], we derive the following formula to measure the severity of forgetting, a popular forgetting measure in CL [4, 7, 8]:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 382, + 647, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 647, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 382, + 647, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {k} = \\frac {1}{k - 1} \\sum_ {j = 1} ^ {k - 1} f _ {j} ^ {k}, \\tag {2}", + "image_path": "9d4a0e4a80c84f57f18a152168aa3ee548295a869334ed1e89b3f3da6a67eda2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_k" + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": " denotes the average forgetting at " + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "-th task and " + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_j^k" + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": " quantifies forgetting for the " + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "-th (" + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "j < k" + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": ") task after the model" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4784" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 99 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 99 + ], + "type": "text", + "content": "has been continually trained up to task " + }, + { + "bbox": [ + 46, + 72, + 287, + 99 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 287, + 99 + ], + "type": "text", + "content": ". Specifically, for a given data distribution, " + }, + { + "bbox": [ + 46, + 72, + 287, + 99 + ], + "type": "inline_equation", + "content": "f_{j}^{k}" + }, + { + "bbox": [ + 46, + 72, + 287, + 99 + ], + "type": "text", + "content": " can be expressed as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 73, + 106, + 287, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 106, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 73, + 106, + 287, + 135 + ], + "type": "interline_equation", + "content": "f _ {j} ^ {k} = \\frac {1}{\\left| \\mathcal {C} ^ {j} \\right|} \\sum_ {c \\in \\mathcal {C} ^ {j}} \\max _ {t \\in \\{1, \\dots . N - 1 \\}} \\left(\\mathcal {A} _ {c} ^ {(n)} - \\mathcal {A} _ {c} ^ {(N)}\\right), \\tag {3}", + "image_path": "f000f3116fb2d7b964cb7aaed81d8399174dcdd728807164df6b4a2fe4ce5a73.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^j" + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": " is a set of classes related to the " + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": "-th task, " + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_c^{(n)}" + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": " is the accuracy on class " + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": " at round " + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_c^{(N)}" + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": " is the final accuracy on class " + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": " after learning all tasks. Note that " + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "inline_equation", + "content": "f_j^k" + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": " captures the average gap between the peak accuracy and the final accuracy for each class of the " + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": "-th task after learning the " + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 145, + 287, + 220 + ], + "type": "text", + "content": "-th task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 220, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 266 + ], + "type": "text", + "content": "We further extend the catastrophic forgetting measurement " + }, + { + "bbox": [ + 46, + 220, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 220, + 287, + 266 + ], + "type": "text", + "content": " in Equation 2 to FL under different data partitions and introduce a relative metric " + }, + { + "bbox": [ + 46, + 220, + 287, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 220, + 287, + 266 + ], + "type": "text", + "content": " to measure forgetting as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 266, + 287, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 266, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 123, + 266, + 287, + 299 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {k} = \\frac {\\sum_ {j = 1} ^ {k - 1} f _ {j} ^ {k}}{\\sum_ {j = 1} ^ {k - 1} \\mathcal {A} _ {(j , k)}}, \\tag {4}", + "image_path": "eedeca5cf6944afc6c642af52de84642a8aff5c2098a034d93092087063794fc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{(j,k)}" + }, + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "type": "text", + "content": " is the accuracy on task " + }, + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "type": "text", + "content": " after learning task " + }, + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "type": "text", + "content": ". For different data partitions, an increased " + }, + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_k" + }, + { + "bbox": [ + 46, + 302, + 287, + 338 + ], + "type": "text", + "content": " indicates a more serious forgetting of previous tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 338, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 517 + ], + "type": "text", + "content": "Figure 1 illustrates the impact of catastrophic forgetting under independent and identically distributed (IID) and different levels of non-independent and identically distributed (non-IID) data partitions. In particular, we employ the Dirichlet distribution, which is widely used in FL [29, 28], to simulate the imbalanced label distribution among different clients. Figure 1(a) shows that the accuracy of the model degrades as training proceeds to new tasks in the IID setting. The performance is even worse in the non-IID settings. These results suggest that FCCL faces significant challenge on extreme non-IID settings. We further analyze the forgetting phenomenon in the Figure 1(b). The higher degree of non-IID exacerbates the forgetting phenomenon in FCCL. These empirical studies motivate us to further investigate the catastrophic forgetting issue in FCCL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 525, + 284, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 284, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 284, + 538 + ], + "type": "text", + "content": "3.3. Alleviating Forgetting via Global Information" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "content": "To tackle the issue of catastrophic forgetting in the FCCL, we argue utilizing global information to improve performance. The global information can derive from the global model. We explore this approach and empirically demonstrate that the integration of global information can effectively mitigate catastrophic forgetting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Learn from the Global Model Inspired by LwF [30], we leverage the knowledge from the previously trained global model to the current task via Knowledge Distillation (KD), using only the data from the current task. We conduct experiments on CIFAR100 dataset with 5 continual tasks, where each task contains data of 20 classes. We evaluate the accuracy of the model on each previous task after training on" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 115, + 555, + 212 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 106 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 106 + ], + "type": "text", + "content": "Table 1: Test accuracy on previous tasks after training on all tasks. In this context, \"FedAvg\" denotes a naive approach whereby clients learn tasks sequentially." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 115, + 555, + 212 + ], + "lines": [ + { + "bbox": [ + 308, + 115, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 308, + 115, + 555, + 212 + ], + "type": "table", + "html": "
PartitionMethodTask 1Task 2Task 3Task 4Task 5Final
IIDFedAvg0000.0582.616.53
FedLwF6.8 ↑11.5 ↑27.1 ↑44.45 ↑63.2 ↓30.61 ↑
NIID (1)FedAvg000081.6516.33
FedLwF6.65 ↑13.71 ↑29.60 ↑45.41 ↑59.35 ↓30.94 ↑
NIID (0.5)FedAvg0000.1577.3015.49
FedLwF1.6 ↑10.65 ↑27.75 ↑41.3 ↑56.65 ↓27.59 ↑
NIID (0.02)FedAvg0000.159.0011.82
FedLwF0.1 ↑1.7 ↑3.35 ↑28.2 ↑54.0 ↓17.47 ↑
", + "image_path": "1b6e0bfbe75e584dd6f5a8d58ceabdeb600ed022e4c2ebfad2818847ab48772c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 225, + 428, + 277 + ], + "blocks": [ + { + "bbox": [ + 315, + 225, + 428, + 277 + ], + "lines": [ + { + "bbox": [ + 315, + 225, + 428, + 277 + ], + "spans": [ + { + "bbox": [ + 315, + 225, + 428, + 277 + ], + "type": "image", + "image_path": "e28bdbf18d73129be0197101faed2ebb137ecd524954b59c9140dae34e1eae38.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 430, + 226, + 536, + 277 + ], + "blocks": [ + { + "bbox": [ + 430, + 226, + 536, + 277 + ], + "lines": [ + { + "bbox": [ + 430, + 226, + 536, + 277 + ], + "spans": [ + { + "bbox": [ + 430, + 226, + 536, + 277 + ], + "type": "image", + "image_path": "b831d8b74780b3c9b36887b41335ac1ed0c8a06b00b2bbadb2079a3033eb7265.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 315, + 277, + 428, + 335 + ], + "blocks": [ + { + "bbox": [ + 315, + 277, + 428, + 335 + ], + "lines": [ + { + "bbox": [ + 315, + 277, + 428, + 335 + ], + "spans": [ + { + "bbox": [ + 315, + 277, + 428, + 335 + ], + "type": "image", + "image_path": "5e6b50ad06163db6dcaebff062c27b15c52a8d7263f45bd9804af0b7d99f1ead.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 347, + 545, + 372 + ], + "lines": [ + { + "bbox": [ + 305, + 347, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 545, + 372 + ], + "type": "text", + "content": "Figure 2: Test accuracy by using current task's data, local exemplar and global exemplar." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 430, + 277, + 536, + 335 + ], + "blocks": [ + { + "bbox": [ + 430, + 277, + 536, + 335 + ], + "lines": [ + { + "bbox": [ + 430, + 277, + 536, + 335 + ], + "spans": [ + { + "bbox": [ + 430, + 277, + 536, + 335 + ], + "type": "image", + "image_path": "72db22f6aa0020921d7347a583d6aa08ee4aa9f89a707c9dfc872a3af84e567d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 394, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 514 + ], + "type": "text", + "content": "all five tasks and report the final accuracy as the average of these results. Table 1 shows that FCCL without knowledge from the global model (i.e., FedAvg [38]) suffers severely from catastrophic forgetting under both IID and non-IID data partitions. Although FedAvg achieves competitive performance on Task 5, it results in " + }, + { + "bbox": [ + 304, + 394, + 545, + 514 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 304, + 394, + 545, + 514 + ], + "type": "text", + "content": " accuracy on previously trained tasks (Task 1, 2, and 3). In contrast, FedLwF achieves better accuracy on Tasks 1, 2, 3, and 4 and final accuracy. The experiments demonstrate that the global model can indeed alleviate the forgetting issue." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": "Learn From Global Exemplar In continual learning, an exemplar refers to a sample from a previous task that is stored in a memory buffer for future training. Assuming that data stored in clients are compliant to use for future training, we adopt the idea proposed in iCaRL [44] to save a small proportion of prior training data in memory. The selection of this small proportion of data assumes that the data distribution pertaining the old task is known, but this assumption does not hold in FL because clients are unable to know the data distributions of others due to data privacy concerns. Nevertheless, we employ two exemplar selection methods to illustrate the utility of global data: global exemplar and local exemplar. Global exemplar assumes that the server aggregates a subset of data from clients and distribute these data to clients in training (Note that this method does" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4785" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 292, + 198 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 292, + 198 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 292, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 292, + 198 + ], + "type": "image", + "image_path": "417749987b67f0dcf09c14e375aaadf518cfa018601e6898cb594f773b7544e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 288, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 288, + 255 + ], + "type": "text", + "content": "Figure 3: Pipeline of TARGET. We utilize the global model (trained on task " + }, + { + "bbox": [ + 46, + 207, + 288, + 255 + ], + "type": "inline_equation", + "content": "k - 1" + }, + { + "bbox": [ + 46, + 207, + 288, + 255 + ], + "type": "text", + "content": ") to synthesize the data with a global distribution, subsequently employing this synthesized data for training the " + }, + { + "bbox": [ + 46, + 207, + 288, + 255 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 207, + 288, + 255 + ], + "type": "text", + "content": "-th task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 287, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 313 + ], + "type": "text", + "content": "not conform to FL and is only used for comparison). Local exemplar means that each client retains a subset of data from local data in the previous tasks for future training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 314, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 288, + 422 + ], + "type": "text", + "content": "As shown in Figure 2, we show the performance on the current and all previous tasks after learning the current task by using only the new task data, with local exemplar, and with global exemplar. It can be clearly observed that using exemplars can significantly improve the model performance, especially when using global exemplars, which can achieve much higher accuracy than using local exemplars. However, this raises a critical challenge of how to select such exemplars without violating the data privacy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 432, + 185, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 185, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 185, + 445 + ], + "type": "text", + "content": "4. Our Method: TARGET" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 115, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 115, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 115, + 464 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 471, + 288, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 288, + 590 + ], + "type": "text", + "content": "To utilize the global information without touching on the real exemplars from clients, we present a method called TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillage), which utilizes global information without storing any real data. A detailed procedure is provided in Algorithm 1. Figure 3 presents an illustration of TARGET, wherein we synthesize data by inverting the global model " + }, + { + "bbox": [ + 46, + 471, + 288, + 590 + ], + "type": "inline_equation", + "content": "\\theta_{k - 1}" + }, + { + "bbox": [ + 46, + 471, + 288, + 590 + ], + "type": "text", + "content": " (which was trained on task " + }, + { + "bbox": [ + 46, + 471, + 288, + 590 + ], + "type": "inline_equation", + "content": "k - 1" + }, + { + "bbox": [ + 46, + 471, + 288, + 590 + ], + "type": "text", + "content": "), followed by combining the synthesized data with real data for local model update on task " + }, + { + "bbox": [ + 46, + 471, + 288, + 590 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 471, + 288, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 598, + 282, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 282, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 282, + 612 + ], + "type": "text", + "content": "4.2. Server Side: Synthesizing Data for Old Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "As demonstrated in Figure 2, data with global distributional information is more effective in mitigating the problem of catastrophic forgetting. Therefore, we propose a method of synthesizing data that can model the data distribution of the global model, without the need to preserve any client's privacy data. Specifically, given a global (teacher) model " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{k-1}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " trained on task " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "k - 1" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ", we first initialize a generator " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " and a student model " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\theta_S" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ". We then repeatedly" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 306, + 87, + 545, + 507 + ], + "blocks": [ + { + "bbox": [ + 312, + 75, + 467, + 87 + ], + "lines": [ + { + "bbox": [ + 312, + 75, + 467, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 75, + 467, + 87 + ], + "type": "text", + "content": "Algorithm 1: Procedure of TARGET." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " : local minibatch size, " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " : local epochs, " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " .. learning rate, synthetic data: " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "X_{syn} = \\emptyset" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " \n1 foreach each task " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\tau = 0,1,\\ldots" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " do \n2 Initialize " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "w_0" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " \n3 foreach each round " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "t = 1,2,\\dots" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " do \n4 " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "S_{t}\\gets" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " (random set of m clients) \n5 foreach client " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "k\\in S_t" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " in parallel do \n6 " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\bigwedge_{t + 1}^{k}\\leftarrow \\text{ClientUpdate}(k,w_{t},\\tau ,X_{syn}) \\end{array}" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " \n7 " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "w_{t + 1}\\gets \\sum_{k = 1}^{m}\\frac{n_k}{n} w_t^k +1" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " \n8 " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "X_{syn} = \\mathsf{DataGeneration}(w_{t + 1})" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " \n9 ClientUpdate " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "(k,w,\\tau ,X_{syn})" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " .. \n10 " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{B}\\gets" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " (split " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k\\cup X_{syn}" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " into batches) \n11 set global model " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\gets w" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " \n12 foreach each local epoch i from 1 to E do \n13 foreach batch " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "(b,b_{syn})\\in \\mathcal{B}" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " do \n14 " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\ell (w;b) = L_{ce}(w;b)" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\tau \\neq 0" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " then \n15 " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\lfloor \\ell (w;b) + = \\alpha L_{kl}(w,\\mathcal{T};b_{syn})\\rfloor \\end{array}" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " \n16 return w to server \n17 DataGeneration(w): \n19 Initialize parameter " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\theta_G,\\theta_S,X_{syn} = \\emptyset" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " \n20 foreach round " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "i = 1,2,\\ldots" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " do \n21 Sample noises and labels " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_i,\\mathbf{y}_i\\}_{i = 1}^b" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " // data generation stage \n22 foreach " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "j = 1,2,\\ldots" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " ,do \n23 Generate " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{x}}_i\\}_{i = 1}^b" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_i\\}_{i = 1}^b" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " \n24 Update " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\theta_G" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " using Equation 9 \n25 Add a batch of data into " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "X_{syn}" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " \n/ model distillation stage \n26 Update " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\theta_S" + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "text", + "content": " using Equation 10 \n27 return " + }, + { + "bbox": [ + 306, + 87, + 545, + 507 + ], + "type": "inline_equation", + "content": "X_{syn}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 529, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 626 + ], + "type": "text", + "content": "perform the following two training steps (see line " + }, + { + "bbox": [ + 304, + 529, + 545, + 626 + ], + "type": "inline_equation", + "content": "19\\sim 28" + }, + { + "bbox": [ + 304, + 529, + 545, + 626 + ], + "type": "text", + "content": " in Algorithm 1): 1) update the generator by continuously optimizing it to generate data that conforms to the global model distribution; 2) update the student model by distilling knowledge from the teacher model with the synthetic data, hoping that the student model can learn the knowledge of the teacher model sufficiently, which demonstrates the effectiveness of the synthesized data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": "Data Generation First, we utilize " + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": " to generate synthetic data from noise " + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": ", we need to ensure that the synthetic data " + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{x} = G(z)" + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": " is similar to the training dataset. If the synthetic data is similar to the training dataset, their predictions should also be similar. We minimize the cross-entropy (CE) loss on the output of global model " + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{k - 1}(\\hat{x})" + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": " and random la" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4786" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 76, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 76, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 76, + 83 + ], + "type": "text", + "content": "bels " + }, + { + "bbox": [ + 47, + 72, + 76, + 83 + ], + "type": "inline_equation", + "content": "\\hat{y}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 83, + 287, + 97 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 83, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 116, + 83, + 287, + 97 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G} ^ {c e} = C E \\left(\\theta_ {k - 1} (\\hat {x}), \\hat {y}\\right). \\tag {5}", + "image_path": "5ce3a90a6c031b980737eb5bca5e3f766827928d5ea472f44ff8faca6791cec8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 102, + 287, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 102, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 102, + 287, + 246 + ], + "type": "text", + "content": "It is expected that the synthetic data generated by generator can be classified into a particular class with a high degree of confidence. However, utilizing only the CE loss will cause the generator overfitting to the synthetic data that are far away from the decision boundary (of the global model) [62, 15], thus failing to deliver a good performance. In order to generate samples that are closer to the decision boundary (of the global model) with better transferability, following previous work [62], we introduce a boundary support loss. Additional weight is given to the data on which the global model and the student model diverge in decision making." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 255, + 287, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 255, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 75, + 255, + 287, + 269 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G} ^ {d i v} = - \\omega K L \\left(\\theta_ {k - 1} (\\hat {x}), \\theta_ {S} (\\hat {x})\\right), \\text {a n d} \\tag {6}", + "image_path": "41a805d36409a43dd97e676ec86aadd00c2265a225de8da0a25eff3580cc1490.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 270, + 287, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 270, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 76, + 270, + 287, + 284 + ], + "type": "interline_equation", + "content": "\\omega = \\mathbb {1} (\\arg \\max \\theta_ {k - 1} (\\hat {x}) \\neq \\arg \\max \\theta_ {S} (\\hat {x})), \\tag {7}", + "image_path": "4fc3329dfa01eeb875f76b883316daf0074e2440eb06e6111b9a8e703a02fb28.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "type": "inline_equation", + "content": "KL" + }, + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "type": "text", + "content": " denotes the Kullback-Leibler (KL) divergence loss, " + }, + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(a)" + }, + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "type": "text", + "content": " output 1 if " + }, + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "type": "text", + "content": " is true and output 0 if " + }, + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 47, + 293, + 287, + 340 + ], + "type": "text", + "content": " is false. By maximizing the KL divergence loss, the generator can generate more representative data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 341, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 287, + 388 + ], + "type": "text", + "content": "Motivated by [57], in order to further improve the stability of generator training, we introduce Batch Normalization (BN) loss to make synthetic data conform with the batch normalization statistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 396, + 287, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 396, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 75, + 396, + 287, + 420 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G} ^ {b n} = \\sum_ {l} \\left(\\left| \\left| \\mu_ {l} (\\hat {x}) - \\mu_ {l} \\right| \\right| + \\left| \\left| \\sigma_ {l} ^ {2} (\\hat {x}) - \\sigma_ {l} ^ {2} \\right| \\right|\\right), \\tag {8}", + "image_path": "930853f6497522982c2ba4513a9dd667b993cc7c9fb8bbe51e1da4f8bf36b4b3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mu_l(\\hat{x})" + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\sigma_l^2 (\\hat{x})" + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "text", + "content": " are the batch-wise mean and variance estimate of the " + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "text", + "content": "-th BN layer of the generator, " + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mu_{l}" + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\sigma_l^2" + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "text", + "content": " are the mean and variance of the " + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "text", + "content": "-th BN layer of " + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "inline_equation", + "content": "f_{S}(\\cdot)" + }, + { + "bbox": [ + 47, + 430, + 287, + 467 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 467, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 287, + 491 + ], + "type": "text", + "content": "Combining the above losses, we can obtain the loss of the generator as follows," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 103, + 499, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 499, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 103, + 499, + 287, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G} = \\mathcal {L} _ {G} ^ {c e} + \\lambda_ {1} \\mathcal {L} _ {G} ^ {d i v} + \\lambda_ {2} \\mathcal {L} _ {G} ^ {b n}, \\tag {9}", + "image_path": "3b086143657b4f1df7cb4a6ae7be02815b30ffaa48fbcf604a738f6989e47f8b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 522, + 281, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 281, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 281, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 522, + 281, + 534 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 47, + 522, + 281, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 522, + 281, + 534 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 47, + 522, + 281, + 534 + ], + "type": "text", + "content": " is the weight for different loss functions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 550, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 287, + 633 + ], + "type": "text", + "content": "Model Distillation In Equation 7, we introduce a student model to assist in training the generator to produce data with greater diversity. A better student model should lead to a better generator. Therefore, after training the generator for several rounds, we subsequently train the student model using the saved synthesized data and the output of the teacher model, using KL loss for knowledge distillation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 108, + 643, + 287, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 643, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 108, + 643, + 287, + 656 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = K L \\left(\\theta_ {k - 1} (\\hat {x}), \\theta_ {S} (\\hat {x})\\right). \\tag {10}", + "image_path": "3a48a03afcb6e5c8a0a997cf2d06b8baf262ebd7f3baad21b4acbc27961e0de1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": "In this way, we can train a student model with better performance, and then further use it to update the generator. An ideal synthetic dataset should be able to efficiently enable student " + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{S}" + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": " to fully learn the knowledge of teacher model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": "Note that when the training of the whole process is over (i.e. the student model can use the synthesized data to obtain high performance), we only retain the synthetic dataset " + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "X_{syn}" + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": " and transfer it to the clients, without saving the generator and student model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 140, + 539, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 539, + 153 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 539, + 153 + ], + "type": "text", + "content": "4.3. Client Side: Update with Global Information" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": "On the client side, we can obtain the data synthesized for the previous task " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "X_{syn}" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": " and the real training data of the current task " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "X_{local}" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": ", then we train the local model " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "\\theta_k" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": " for task " + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 159, + 545, + 255 + ], + "type": "text", + "content": " on the two datasets at the same time. We showed in Section 3.3 that the use of global models and global data can alleviate forgetting. Thus we distill the knowledge of global teacher model and global synthetic data by minimizing the following objective function," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 265, + 545, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 265, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 265, + 545, + 293 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c l i e n t}} = \\underbrace {C E \\left(\\theta_ {k} (x) , y\\right)} _ {\\text {f o r c u r r e n t t a k}} + \\alpha \\cdot \\underbrace {K L \\left(\\theta_ {k - 1} (\\hat {x}) , \\theta_ {k} (\\hat {x})\\right)} _ {\\text {f o r p r e v i o u s t a k s}}, \\tag {11}", + "image_path": "ae8f0c2a0d9684227caba1cb5b3532c3df17d2926fe4ce4c2f0f304e28f4aa56.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "inline_equation", + "content": "(x,y)\\in X_{local}" + }, + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "inline_equation", + "content": "(\\hat{x})\\in X_{syn}" + }, + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "text", + "content": ". The utilization of the distillation loss facilitates efficient transfer of knowledge from the previous task to the current task model. And " + }, + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 303, + 545, + 363 + ], + "type": "text", + "content": " is a hyper-parameter that controls the strength of the regularization for the previous tasks." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 375, + 387, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 387, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 387, + 388 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 395, + 432, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 432, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 432, + 408 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 414, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 521 + ], + "type": "text", + "content": "We experiment on two datasets, namely CIFAR-100 [25], and Tiny-ImageNet [26], to evaluate the performance of our proposed approach. To establish the order of the continual tasks, we adopt the widely used protocols [49, 61, 42, 7]. Specifically, we divide all classes of each dataset equally into multiple tasks by default, i.e. we evenly divide the classes into 5 and 10 tasks to simulate class continual learning scenarios. We employ ResNet18 [18] as the backbone for the classification model." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 522, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 712 + ], + "type": "text", + "content": "To evaluate our approach, we employ the standard continual learning metrics, as used in prior works [23, 22, 47, 16], which include average accuracy across all tasks and a forgetting measure [6] (see Equation 2). For a fair comparison with the baseline class continual learning methods in the FCCL setting, we implement three types of baselines: 1) Finetune, in which each client simply learns tasks in sequence; 2) FedWeIT [58], a regularization-based method in Federated Continual Learning that maximizes the knowledge transfer between clients; 3) Examples of typical continual learning methods that do not store training data for rehearsal, including EWC [27] and LwF [30]. In addition, we compare our method with methods that store real training data of old tasks, such as iCaRL [44]. We implement these traditional continual learning algorithms in the FCCL scenario and name them as FedEWC, FedLwF, and" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4787" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 116, + 547, + 256 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": "Table 2: The Average Accuracy (%) and Forgetting for all learned tasks on CIFAR-100 for various numbers of tasks (5, 10) under both IID and non-IID settings. Results are reported as an average of 3 runs. 'Acc' refers to average accuracy, and 'F' represents the forgetting measure utilized in Equation 2. The best results are in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 116, + 547, + 256 + ], + "lines": [ + { + "bbox": [ + 48, + 116, + 547, + 256 + ], + "spans": [ + { + "bbox": [ + 48, + 116, + 547, + 256 + ], + "type": "table", + "html": "
Data partitionIIDNIID (1)NIID (0.5)
TasksT=5T=10T=5T=10T=5T=10
MethodAcc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)Acc(↑)F(↓)
Finetune16.120.787.830.7516.330.778.450.7415.490.747.640.71
FedEWC16.510.718.010.6516.060.688.840.6216.860.668.040.65
FedWeIT28.450.5220.390.4328.560.4919.680.4524.570.5415.450.48
FedLwF30.610.4523.270.3730.940.4221.160.4127.590.4417.980.45
Ours36.310.2224.760.2634.890.2422.850.2633.330.2720.710.29
", + "image_path": "935cf74fd0661cff4879794d93af57e1596daecf6b26950ce13e13a43717dd13.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 274, + 170, + 377 + ], + "blocks": [ + { + "bbox": [ + 52, + 274, + 170, + 377 + ], + "lines": [ + { + "bbox": [ + 52, + 274, + 170, + 377 + ], + "spans": [ + { + "bbox": [ + 52, + 274, + 170, + 377 + ], + "type": "image", + "image_path": "c12a0c1865af8eaa726df1e79316f0b03052ac6b238e65dcf404865566e323de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 390, + 288, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 390, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 288, + 415 + ], + "type": "text", + "content": "Figure 4: Average accuracy on previous tasks and current task after the model trained on current task." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 274, + 288, + 377 + ], + "blocks": [ + { + "bbox": [ + 171, + 274, + 288, + 377 + ], + "lines": [ + { + "bbox": [ + 171, + 274, + 288, + 377 + ], + "spans": [ + { + "bbox": [ + 171, + 274, + 288, + 377 + ], + "type": "image", + "image_path": "aa93cc02ca34ab091c357fd66b8a5728e8cdeeebc701e9103454e3f65048d25f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 436, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 473 + ], + "type": "text", + "content": "FedIcaRL. For detailed information on the task configuration, default hyper-parameters and additional experimental results, please refer to the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 479, + 201, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 201, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 201, + 493 + ], + "type": "text", + "content": "5.2. Experiments on CIFAR-100" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "For CIFAR-100 dataset, we conduct experiments on two sets of tasks consisting of 5 and 10 tasks, respectively. We run the experiment on both IID and non-IID scenarios. For non-IID setting, the Dirichlet parameter is set to 0.5 and 1, i.e. " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{NIID}(0.5)" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{NIID}(1)" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": ". Table 2 shows the final average accuracy of the FL model trained on all tasks, along with the corresponding forgetting measure for each experiment. It is important to emphasize that an optimal method is characterized by high average accuracy and low forgetting measure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": "Table 2 indicates that Finetune exhibits the poorest performance when attempting to learn the continuously incoming tasks sequentially, thereby experiencing the issue of catastrophic forgetting. Moreover, we observed that methods based on regularization constraints, such as FedEWC and FedWeIT, were often ineffective in preventing the model from forgetting old tasks due to the lack of available data. On the other hand, we found that distillation-based approaches such as FedLwF and our proposed method were" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 274, + 547, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 547, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 547, + 550 + ], + "type": "text", + "content": "capable of improving the final average accuracy while simultaneously mitigating the issue of catastrophic forgetting. This is due to the transfer of model knowledge learned from the old task to the new task in a distillation manner, enabling the model to prevent catastrophic forgetting. Obviously, we found that applying our proposed method to generate synthetic datasets for the federated models trained on old tasks and subsequently performing model distillation on these synthetic datasets can lead to a substantial improvement in the average accuracy and reduction in the forgetting measure. This observation underscores the ability of our synthetic data to capture the distribution characteristics of historical task data accurately. For example, in IID setting, when partitioning the CIFAR-100 dataset into five tasks, our method achieves an accuracy of " + }, + { + "bbox": [ + 304, + 274, + 547, + 550 + ], + "type": "inline_equation", + "content": "36.31\\%" + }, + { + "bbox": [ + 304, + 274, + 547, + 550 + ], + "type": "text", + "content": ", which is about " + }, + { + "bbox": [ + 304, + 274, + 547, + 550 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 304, + 274, + 547, + 550 + ], + "type": "text", + "content": " higher than the best baseline method FedLwF. It is worth noting that we observed a decrease in the average accuracy of all methods to varying degrees as the number of tasks increased from 5 tasks to 10 tasks due to increased task complexity and the associated forgetting phenomenon. Nonetheless, our proposed method maintains the highest average accuracy and the lowest forgetting measure even under these more demanding conditions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Figure 4 illustrates the performance of the models trained by various methods on all previously learned tasks after the completion of each task. Specifically, it shows the average accuracy of the model on both the current and previous tasks after the completion of each task (e.g., after learning the second task, the average accuracy of the model on both the first and second tasks is measured). Based on these curves, it is evident that our proposed model outperforms other competing baseline methods in all incremental tasks, regardless of the number of tasks involved. This finding underscores the effectiveness of our approach in facilitating multiple local clients to learn new classes in a streaming manner while mitigating the forgetting problem." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4788" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 289, + 177 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 289, + 177 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 289, + 177 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 289, + 177 + ], + "type": "image", + "image_path": "67f20052159e796ecaa4a9ff6da8e20226eb69711fddcd68d8ab91057659124b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 187, + 288, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 288, + 224 + ], + "type": "text", + "content": "Figure 5: The average accuracy " + }, + { + "bbox": [ + 46, + 187, + 288, + 224 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 187, + 288, + 224 + ], + "type": "text", + "content": " and Forgetting for all learned tasks on Tiny-ImageNet for 5 tasks under both IID and non-IID settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 246, + 219, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 246, + 219, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 219, + 259 + ], + "type": "text", + "content": "5.3. Experiments on Tiny-ImageNet" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 265, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 469 + ], + "type": "text", + "content": "We also evaluated the performance of our proposed method on the more challenging Tiny-ImageNet dataset and obtained similar results to those observed in the CIFAR-100 experiments. Specifically, in Figure 5, we present the final average accuracy and forgetting measure for all learned tasks in both IID and non-IID settings for the case of 5 tasks. The Dirichlet parameter is set to " + }, + { + "bbox": [ + 46, + 265, + 287, + 469 + ], + "type": "inline_equation", + "content": "\\{0.05, 0.1, 0.5, 1\\}" + }, + { + "bbox": [ + 46, + 265, + 287, + 469 + ], + "type": "text", + "content": ". Based on Figure 5, it is evident that our proposed method consistently outperforms FedLwF in terms of average accuracy across all data partitions. Moreover, our method demonstrates a significantly lower forgetting measure than FedLwF under both IID and non-IID settings. Our proposed method achieves an average accuracy that is approximately " + }, + { + "bbox": [ + 46, + 265, + 287, + 469 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 265, + 287, + 469 + ], + "type": "text", + "content": " higher than that of FedLwF even in the most challenging scenario (i.e. NIID(0.05)). This result highlights the effectiveness of our method in mitigating catastrophic forgetting in the presence of extreme data distributions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 270, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 270, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 270, + 491 + ], + "type": "text", + "content": "5.4. Comparison with Exemplar-based Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "text", + "content": "In CL, the most successful approaches to alleviate forgetting require extensive replay of previously seen data, which can be problematic when data legality and privacy concerns exist. Among them, iCaRL [44] is a classic but unrealistic algorithm that relies on the stored exemplars in addition to the network parameters, and it is intuitive that using old task's real data could be beneficial to alleviate forgetting problem. To further understand the performance gap, we compare our proposed TARGET (which uses synthetic data) with iCaRL (which requires storing exemplars from old task's real data), and study the effects of different exemplar memories on the performance of our method and iCaRL in Figure 6. We set the stored exemplar size to " + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\{1000, 1500, 2000\\}" + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "text", + "content": " for iCaRL, and " + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\{2000, 3000\\}" + }, + { + "bbox": [ + 46, + 497, + 287, + 675 + ], + "type": "text", + "content": " for our method on the CIFAR-100 dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "The accuracy curve in Figure 6 represents the average accuracy rate measured over all 100 classes learned by the model during the last learning task. While our method" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 77, + 536, + 196 + ], + "blocks": [ + { + "bbox": [ + 322, + 77, + 536, + 196 + ], + "lines": [ + { + "bbox": [ + 322, + 77, + 536, + 196 + ], + "spans": [ + { + "bbox": [ + 322, + 77, + 536, + 196 + ], + "type": "image", + "image_path": "b360c9d98595787fb5e505121748611600bc34d21883b00e2424542a0a2795f3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 205, + 545, + 242 + ], + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 242 + ], + "type": "text", + "content": "Figure 6: Accuracy comparison between real data-based method (iCaRL) and synthetic data-based method (TARGET). We show the result for the last task." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 288, + 543, + 353 + ], + "blocks": [ + { + "bbox": [ + 306, + 255, + 545, + 278 + ], + "lines": [ + { + "bbox": [ + 306, + 255, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 306, + 255, + 545, + 278 + ], + "type": "text", + "content": "Table 3: The effect of " + }, + { + "bbox": [ + 306, + 255, + 545, + 278 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 255, + 545, + 278 + ], + "type": "text", + "content": " on the performance of both new and old tasks in CIFAR-100, 2 tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 288, + 543, + 353 + ], + "lines": [ + { + "bbox": [ + 309, + 288, + 543, + 353 + ], + "spans": [ + { + "bbox": [ + 309, + 288, + 543, + 353 + ], + "type": "table", + "html": "
α35101525
Old Task (0-49)28.3434.9149.0258.4866.02
New Task (50-99)62.4660.2653.7445.2629.32
Average (0-99)45.4147.5851.3851.8747.76
", + "image_path": "8f9be0c87a631f5edf61211ed2373f74408cb7851152fe0199c4cefef9678c37.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 373, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 566 + ], + "type": "text", + "content": "can achieve similar performance to storing 1k real training data by storing 2k synthetic data, it still cannot outperform storing 2k real training data. However, it is worth noting that our method does not require storing any real training data, which can be a significant advantage in scenarios where storing real data is difficult or not allowed due to privacy or legal concerns. Additionally, our method achieves better performance than storing only 1k real training data, which indicates that our synthetic data is effective in mitigating catastrophic forgetting. We observed that when our method stores 3k synthetic data, it achieves better accuracy than when it stores 2k synthetic data. However, surpassing iCaRL in performance with an equal amount of data remains a challenge for our method. How to effectively use fewer synthetic data with more valuable knowledge from previous tasks will be left as a future research direction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 574, + 441, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 441, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 441, + 586 + ], + "type": "text", + "content": "5.5. Analysis of Our Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Trade-off between Backward and Forward Transfer. Continual learning presents a challenge in balancing the trade-off between maintaining high accuracy on old tasks (backwards transfer) and achieving high accuracy on new tasks (forward transfer). In Table 3, we evenly split the CIFAR-100 dataset into two tasks and test the average accuracy of our proposed method under different values of " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". We observe that the trade-off between backward and forward transfer is not always balanced. In order to achieve good backward transfer, a large value of " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " should be used" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4789" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 72, + 165, + 182 + ], + "blocks": [ + { + "bbox": [ + 53, + 72, + 165, + 182 + ], + "lines": [ + { + "bbox": [ + 53, + 72, + 165, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 165, + 182 + ], + "type": "image", + "image_path": "eed01bfff92f3aafd5332161b6ad73413ffa3b58b50464ebaa77e341b7be11a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 183, + 282, + 194 + ], + "lines": [ + { + "bbox": [ + 51, + 183, + 282, + 194 + ], + "spans": [ + { + "bbox": [ + 51, + 183, + 282, + 194 + ], + "type": "text", + "content": "Figure 7: Accuracy for different size of synthetic dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 171, + 72, + 277, + 182 + ], + "blocks": [ + { + "bbox": [ + 171, + 72, + 277, + 182 + ], + "lines": [ + { + "bbox": [ + 171, + 72, + 277, + 182 + ], + "spans": [ + { + "bbox": [ + 171, + 72, + 277, + 182 + ], + "type": "image", + "image_path": "b7529b6618ef3d10da9af403483a8a9be090a985a8d10ab89da2872293ab0394.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 174, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 304, + 174, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 545, + 198 + ], + "type": "text", + "content": "Figure 9: Distillation results for task " + }, + { + "bbox": [ + 304, + 174, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\{1,2,3,4\\}" + }, + { + "bbox": [ + 304, + 174, + 545, + 198 + ], + "type": "text", + "content": " when trained on 5 tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 206, + 282, + 322 + ], + "blocks": [ + { + "bbox": [ + 53, + 206, + 282, + 322 + ], + "lines": [ + { + "bbox": [ + 53, + 206, + 282, + 322 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 282, + 322 + ], + "type": "image", + "image_path": "4843a71ba65577c05f88d66f719d9b12961e4fcfea09c2409ecd829015e481d3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 331, + 275, + 342 + ], + "lines": [ + { + "bbox": [ + 58, + 331, + 275, + 342 + ], + "spans": [ + { + "bbox": [ + 58, + 331, + 275, + 342 + ], + "type": "text", + "content": "Figure 8: Visualization of randomly synthesised data." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "text", + "content": "to prevent the model from forgetting previous tasks and to encourage it to focus more on the old synthetic data. Conversely, to achieve good forward transfer, a small value of " + }, + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "text", + "content": " should be used, allowing the model to learn quickly and effectively from new tasks while still maintaining some knowledge from the old tasks. The experimental results in Table 3 indicate that when the value of " + }, + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "text", + "content": " is between 10-15, the model achieves a good balance between accuracy on new tasks and accuracy on old tasks. We also partition the CIFAR-100 dataset into 5 equal tasks and test our method's performance on all previous tasks after learning each new task (refer to Appendix)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Memory of Synthetic Data. The optimal amount of synthetic data generated for the old task plays a crucial role in determining the final performance of our method. Insufficient synthetic data may not adequately facilitate knowledge transfer from the old tasks, while excessive data generation can lead to increased memory and communication costs. Therefore, determining an appropriate amount of synthetic data that strikes a balance between knowledge transfer and computational cost is critical for the effectiveness and efficiency of our approach. As shown in Figure 7, we test our method on CIFAR-100, which is divided into 5 tasks with different data sizes ranging from 2k to 16k. It can be observed that when the data volume is relatively small, such as 2k and 4k, the performance of the model is poor, especially when the data volume is 2k, the testing curve in task 4 shows a later decline. This is because when the data volume is too small, the model is unable to effectively learn" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 426, + 158 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 426, + 158 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 426, + 158 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 426, + 158 + ], + "type": "image", + "image_path": "21927968055e9a469543b13752cbbcf6da33269afa682eda7d4b04182252cd35.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 429, + 73, + 545, + 158 + ], + "blocks": [ + { + "bbox": [ + 429, + 73, + 545, + 158 + ], + "lines": [ + { + "bbox": [ + 429, + 73, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 429, + 73, + 545, + 158 + ], + "type": "image", + "image_path": "decc23f26b9566ccb4f5e2360c30d75fb876661377008b90c6e87f3314be8533.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 209, + 547, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 547, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 547, + 340 + ], + "type": "text", + "content": "knowledge from old tasks. Increasing the data volume to 8k can effectively alleviate the forgetting phenomenon and achieve good performance. However, continuously increasing the data volume to 12k and 16k do not result in significant improvement in the model's performance. It is important to note that the size of the data volume alone does not guarantee the effectiveness of synthetic data in improving machine learning models. The quality and relevance of the synthetic data must also be carefully considered to ensure that it accurately represents the underlying distribution of the real-world data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 356, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 547, + 464 + ], + "type": "text", + "content": "Visualization on Synthetic Data. To demonstrate the effectiveness of the synthetic data, we present in Figure 8 the visualization results of the synthesized images generated by our method after the model learns the penultimate task on Tiny-ImageNet (CIFAR-100 in Appendix) for 5 tasks. These data have the potential to efficiently enable the student model to approach the performance of the teacher model rapidly. However, it is worth noting that these data exhibit visual dissimilarities from the actual training data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 479, + 547, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 563 + ], + "type": "text", + "content": "Analysis on Distillation. Based on the synthetic data, we distill the knowledge of the model trained on old tasks into a student model. The effect of distillation is further demonstrated in Figure 9. It can be observed that even without accessing any private training data from clients, our method can quickly distill the student model on the server to approach the performance of the global model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 574, + 379, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 379, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 379, + 586 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "In conclusion, this paper introduces a novel method, TARGET (federatTed clAss-continual leaRninG via Exemplar-free disTillage), to alleviate the catastrophic forgetting problem in Federated Class-Continual Learning (FCCL). Unlike all the previous methods, our proposed method leverages global knowledge, without requiring any additional datasets or data from previous tasks, making it ideal for privacy-sensitive scenarios. Extensive experimental results demonstrate the effectiveness of our proposed method in comparison to existing FCCL methods." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4790" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 135 + ], + "type": "text", + "content": "[1] Eden Belouadah and Adrian Popescu. Il2m: Class incremental learning with dual memory. In Proceedings of the IEEE/CVF international conference on computer vision, pages 583-592, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 169 + ], + "type": "text", + "content": "[2] Cecilia Bjursell. The pandemic as disjuncture: Lifelong learning in a context of fear. International Review of Education, 66(5-6):673-689, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 215 + ], + "type": "text", + "content": "[3] Thang D Bui, Cuong V Nguyen, Siddharth Swaroop, and Richard E Turner. Partitioned variational inference: A unified framework encompassing federated and continual learning. arXiv preprint arXiv:1811.11206, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 259 + ], + "type": "text", + "content": "[4] Sungmin Cha, Hsiang Hsu, Taebaek Hwang, Flavio P. Calmon, and Taesup Moon. Cpr: classifier-projection regularization for continual learning. arXiv preprint arXiv:2006.07326, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 305 + ], + "type": "text", + "content": "[5] Yatin Chaudhary, Pranav Rai, Matthias Schubert, Hinrich Schütze, and Pankaj Gupta. Federated continual learning for text classification via selective inter-client transfer. arXiv preprint arXiv:2210.06101, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 308, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 288, + 361 + ], + "type": "text", + "content": "[6] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532-547, 2018. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 364, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 288, + 407 + ], + "type": "text", + "content": "[7] Arslan Chaudhry, Naemullah Khan, Puneet Dokania, and Philip Torr. Continual learning in low-rank orthogonal subspaces. Advances in Neural Information Processing Systems, 33:9900-9911, 2020. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 409, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 288, + 442 + ], + "type": "text", + "content": "[8] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 444, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 288, + 487 + ], + "type": "text", + "content": "[9] Marco Ciotti, Massimo Ciccozzi, Alessandro Terrinoni, Wen-Can Jiang, Cheng-Bin Wang, and Sergio Bernardini. The COVID-19 pandemic. Critical reviews in clinical laboratory sciences, 57(6):365-388, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "text", + "content": "[10] Jiahua Dong, Yang Cong, Gan Sun, Zhen Fang, and Zhengming Ding. Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1(2):1-18, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "text", + "content": "[12] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10164-10173, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "type": "text", + "content": "[13] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 86-102. Springer, 2020. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[14] Fernando Estevez Casado, Dylan Lema Pais, Marcos Fernández Criado, Roberto Iglesias Rodríguez, Carlos Vázquez Regueiro, and Senen Barro Ameneiro. Concept drift detection and adaptation for federated and continual learning. 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[15] Gongfan Fang, Kanya Mo, Xinchao Wang, Jie Song, Shitao Bei, Haofei Zhang, and Mingli Song. Up to 100x faster data-free knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 6597-6604, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "text", + "content": "[16] Ian J Goodfellow, Mehdi Mirza, Da Xiao, Aaron Courville, and Yoshua Bengio. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 230, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 284 + ], + "type": "text", + "content": "[17] Andrew Hard, Kanishka Rao, Rajiv Mathews, Swaroop Ramaswamy, Françoise Beaufays, Sean Augenstein, Hubert Eichner, Chloe Kiddon, and Daniel Ramage. Federated learning for mobile keyboard prediction. arXiv preprint arXiv:1811.03604, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 364 + ], + "type": "text", + "content": "[19] Ziyue Jiang, Yi Ren, Ming Lei, and Zhou Zhao. Fed-speech: Federated text-to-speech with continual learning. arXiv preprint arXiv:2110.07216, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "type": "text", + "content": "[20] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and Trends® in Machine Learning, 14(1-2):1-210, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 432, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 432, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 308, + 432, + 545, + 497 + ], + "type": "text", + "content": "[21] Georgios Kaissis, Alexander Ziller, Jonathan Passerat-Palmbach, Theo Ryffel, Dmitrii Usynin, Andrew Trask, Ionésio Lima Jr, Jason Mancuso, Friederike Jungmann, Marc-Matthias Steinborn, et al. End-to-end privacy preserving deep learning on multi-institutional medical imaging. Nature Machine Intelligence, 3(6):473-484, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 499, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 499, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 499, + 545, + 542 + ], + "type": "text", + "content": "[22] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 1, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 544, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 609 + ], + "type": "text", + "content": "[23] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 1, 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 611, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 545, + 643 + ], + "type": "text", + "content": "[24] Jakub Konečny, Brendan McMahan, and Daniel Ramage. Federated optimization: Distributed optimization beyond the datacenter. arXiv preprint arXiv:1511.03575, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 644, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 644, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 644, + 545, + 667 + ], + "type": "text", + "content": "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 689 + ], + "type": "text", + "content": "[26] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Sang-Woo Lee, Jin-Hwa Kim, Jaehyun Jun, Jung-Woo Ha, and Byoung-Tak Zhang. Overcoming catastrophic forget-" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4791" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "text", + "content": "ting by incremental moment matching. Advances in neural information processing systems, 30, 2017. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 138 + ], + "type": "text", + "content": "[28] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[29] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. arXiv preprint arXiv:1812.06127, 2018. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "text", + "content": "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 293 + ], + "type": "text", + "content": "[31] Zexi Li, Qunwei Li, Yi Zhou, Wenliang Zhong, Guannan Zhang, and Chao Wu. Edge-cloud collaborative learning with federated and centralized features. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '23, page 1949-1953, New York, NY, USA, 2023. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 295, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 295, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 295, + 287, + 337 + ], + "type": "text", + "content": "[32] Zexi Li, Tao Lin, Xinyi Shang, and Chao Wu. Revisiting weighted aggregation in federated learning with neural networks. In International Conference on Machine Learning. PMLR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 381 + ], + "type": "text", + "content": "[33] Yaoyao Liu, Bernt Schiele, and Qianru Sun. Rmm: Reinforced memory management for class-incremental learning. Advances in Neural Information Processing Systems, 34:3478-3490, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 384, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 384, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 287, + 415 + ], + "type": "text", + "content": "[34] Yuang Liu, Wei Zhang, Jun Wang, and Jianyong Wang. Data-free knowledge transfer: A survey. arXiv preprint arXiv:2112.15278, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 449 + ], + "type": "text", + "content": "[35] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 525 + ], + "type": "text", + "content": "[36] Yuhang Ma, Zhongle Xie, Jue Wang, Ke Chen, and Lidan Shou. Continual federated learning based on knowledge distillation. In Lud De Raedt, editor, Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 2182–2188. International Joint Conferences on Artificial Intelligence Organization, 7 2022. Main Track. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 592 + ], + "type": "text", + "content": "[37] Yuhang Ma, Zhongle Xie, Jue Wang, Ke Chen, and Lidan Shou. Continual federated learning based on knowledge distillation. In Luc De Raedt, editor, Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, pages 2182-2188. ijcai.org, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 593, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 678 + ], + "type": "text", + "content": "[38] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Aarti Singh and Jerry Zhu, editors, Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, volume 54 of Proceedings of Machine Learning Research, pages 1273-1282. PMLR, 20-22 Apr 2017. 1, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[39] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 544, + 95 + ], + "type": "text", + "content": "In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[40] Sudhanshu Mittal, Silvio Galesso, and Thomas Brox. Essentials for class incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3513-3522, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 544, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 544, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 544, + 163 + ], + "type": "text", + "content": "[41] Alexander Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "text", + "content": "[42] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural networks, 113:54-71, 2019. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "text", + "content": "[43] Tae Jin Park, Kenichi Kumatani, and Dimitrios Dimitriadis. Tackling dynamics in federated incremental learning with variational embedding rehearsal. arXiv preprint arXiv:2110.09695, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "text", + "content": "[44] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "text", + "content": "[45] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. arXiv preprint arXiv:1810.11910, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 354, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 376 + ], + "type": "text", + "content": "[46] Amanda Rios and Laurent Itti. Closed-loop gan for continual learning. arXiv preprint arXiv:1811.01146, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 377, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 409 + ], + "type": "text", + "content": "[47] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 1, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 455 + ], + "type": "text", + "content": "[48] Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. Overcoming catastrophic forgetting with hard attention to the task. In International Conference on Machine Learning, pages 4548-4557. PMLR, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 489 + ], + "type": "text", + "content": "[49] Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "text", + "content": "[50] James Smith, Yen-Chang Hsu, Jonathan Balloch, Yilin Shen, Hongxia Jin, and Zsolt Kira. Always be dreaming: A new approach for data-free class-incremental learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9374-9384, 2021. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 545, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 567 + ], + "type": "text", + "content": "[51] Sebastian Thrun. Lifelong learning algorithms. Learning to learn, 8:181-209, 1998. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "text", + "content": "[52] Anastasiia Usmanova, François Portet, Philippe Lalanda, and German Vega. A distillation-based approach integrating continual learning and federated learning for pervasive services. arXiv preprint arXiv:2109.04197, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 678 + ], + "type": "text", + "content": "[53] Anamaria Vizitiu, Cosmin Loan Nita, Andrei Puiu, Constantin Suciu, and Lucian Mihai Itu. Towards privacy-preserving deep learning based medical imaging applications. In 2019 IEEE international symposium on medical measurements and applications (MeMeA), pages 1-6. IEEE, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[54] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In Computer Vision-ECCV 2022: 17th" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4792" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXV, pages 398-414. Springer, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "text", + "content": "[55] Ju Xu and Zhanxing Zhu. Reinforced continual learning. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "text", + "content": "[56] Li Yang, Shasha Liu, Jinyan Liu, Zhixin Zhang, Xiaochun Wan, Bo Huang, Youhai Chen, and Yi Zhang. Covid-19: immunopathogenesis and immunotherapeutics. Signal transduction and targeted therapy, 5(1):128, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "text", + "content": "[57] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715-8724, 2020. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 296 + ], + "type": "text", + "content": "[58] Jaehong Yoon, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, pages 12073-12086. PMLR, 2021. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 341 + ], + "type": "text", + "content": "[59] Jaehong Yoon, Saehoon Kim, Eunho Yang, and Sung Ju Hwang. Scalable and order-robust continual learning with additive parameter decomposition. arXiv preprint arXiv:1902.09432, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 376 + ], + "type": "text", + "content": "[60] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. arXiv preprint arXiv:1708.01547, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 378, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 420 + ], + "type": "text", + "content": "[61] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987-3995. PMLR, 2017. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[62] Jie Zhang, Chen Chen, Bo Li, Lingjuan Lyu, Shuang Wu, Shouhong Ding, Chunhua Shen, and Chao Wu. Dense: Data-free one-shot federated learning. In Advances in Neural Information Processing Systems. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 520 + ], + "type": "text", + "content": "[63] Jie Zhang, Zhiqi Li, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Chao Wu. Federated learning with label distribution skew via logits calibration. In International Conference on Machine Learning, pages 26311-26329. PMLR, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[64] Junting Zhang, Jie Zhang, Shalini Ghosh, Dawei Li, Serafettin Tasci, Larry Heck, Heming Zhang, and C-C Jay Kuo. Class-incremental learning via deep model consolidation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1131-1140, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "type": "text", + "content": "[65] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu Tao Xia. Maintaining discrimination and fairness in class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13208-13217, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[66] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[67] Didi Zhu, Yinchuan Li, Yunfeng Shao, Jianye Hao, Fei Wu, Kun Kuang, Jun Xiao, and Chao Wu. Generalized universal domain adaptation with generative flow networks. arXiv preprint arXiv:2305.04466, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[68] Didi Zhu, Yincuan Li, Junkun Yuan, Zexi Li, Yunfeng Shao, Kun Kuang, and Chao Wu. Universal domain adaptation via compressive attention matching. arXiv preprint arXiv:2304.11862, 2023. 2" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4793" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_content_list.json b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b57f4b5605c5ab1f74f06bbae75274a07a1aa63f --- /dev/null +++ b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_content_list.json @@ -0,0 +1,1720 @@ +[ + { + "type": "text", + "text": "TCOVIS: Temporally Consistent Online Video Instance Segmentation", + "text_level": 1, + "bbox": [ + 133, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junlong Li Bingyao Yu Yongming Rao Jie Zhou Jiwen Lu*", + "bbox": [ + 233, + 180, + 740, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Automation, Tsinghua University, China", + "bbox": [ + 264, + 199, + 705, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beijing National Research Center for Information Science and Technology, China", + "bbox": [ + 161, + 215, + 807, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, significant progress has been made in video instance segmentation (VIS), with many offline and online methods achieving state-of-the-art performance. While offline methods have the advantage of producing temporally consistent predictions, they are not suitable for real-time scenarios. Conversely, online methods are more practical, but maintaining temporal consistency remains a challenging task. In this paper, we propose a novel online method for video instance segmentation, called TCOVIS, which fully exploits the temporal information in a video clip. The core of our method consists of a global instance assignment strategy and a spatio-temporal enhancement module, which improve the temporal consistency of the features from two aspects. Specifically, we perform global optimal matching between the predictions and ground truth across the whole video clip, and supervise the model with the global optimal objective. We also capture the spatial feature and aggregate it with the semantic feature between frames, thus realizing the spatio-temporal enhancement. We evaluate our method on four widely adopted VIS benchmarks, namely YouTube-VIS 2019/2021/2022 and OVIS, and achieve state-of-the-art performance on all benchmarks without bells-and-whistles. For instance, on YouTube-VIS 2021, TCOVIS achieves 49.5 AP and 61.3 AP with ResNet-50 and Swin-L backbones, respectively. Code is available at https://github.com/jun-long-li/TCOVIS.", + "bbox": [ + 75, + 301, + 472, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 739, + 209, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video instance segmentation (VIS) is a challenging and representative video understanding task recently introduced in [37]. It aims at detecting, segmenting and tracking instances across a video. VIS is attracting increasing attention for various real-world applications such as video editing, video surveillance, augmented reality and autonomous driving. Recently introduced VIS methods can be roughly", + "bbox": [ + 75, + 767, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "categorized into two groups: offline methods and online methods. Offline methods [2, 16, 19, 31, 33, 35] take as input the whole video and perform the segmentation of instance sequence for the whole video at once. Online methods [8, 34, 17, 10, 38], on the contrary, take as input a video frame by frame and generate the pre-frame object instances while associating the frame-wise results across frames. Both offline and online methods have achieved impressing performance on the VIS task.", + "bbox": [ + 496, + 268, + 893, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Offline methods have an inherent advantage in producing temporally consistent predictions, since delicate temporal communication and association mechanisms can be adopted throughout the video [39, 33, 14] to handle the overall temporal information and impose an explicit constraint on the temporal consistency. However, the video-in and video-out offline manner is not suitable for real-time scenarios. Conversely, online methods are more practical and making considerable progress but suffer from temporal inconsistency (as shown in Figure 1), remaining a great challenge.", + "bbox": [ + 496, + 411, + 893, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Online methods rely on specific instance association applied across frames, since only one frame is observed at a time. Existing association techniques can be grouped into two categories, including tracking-by-detection and query propagation-based paradigms. Tracking-by-detection methods [37, 34, 32] generate the per-frame instances independently by existing instance segmentation models [11, 26, 5] and track instances via tracking heads [37] or instance embeddings matching [34, 15]. In this way, the features of different frames are isolated before tracking, which results in temporal inconsistency. Query propagation-based methods [13, 10, 41] are inspired by query-based methods [3, 25] and they propagate the query across frames to decode a unique instance without heuristic matching algorithms. Despite the explicit temporal link of queries, the temporal consistency is impaired by the Local Matching and Propagating (LocPro) scheme, where they first perform local optimal matching between the predictions and ground truth at the beginning of the video, and then propagate the assignment across frames, forcing all features from subsequent frames to follow. The LocPro is not suitable for the holistic optimization across frames and results in temporal inconsistencies.", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1097", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/614e5bdb04624b7cb1a444c6913b3afaf4d077bd360c7bcc669ca7753656b233.jpg", + "image_caption": [ + "Figure 1. Visualization of predictions from the previous online method (the online GenVIS [13]) and our TCOVIS. The previous method generates temporally inconsistent predictions, while our proposed TCOVIS achieves temporal consistency and outperforms the previous method (Best viewed in color)." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 460, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tency. Thus, achieving temporal consistency is challenging for online methods and also not comprehensively investigated by previous online VIS methods.", + "bbox": [ + 75, + 338, + 467, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel online method for video instance segmentation, named TCOVIS, to fully exploit the temporal information within a video. We take as the baseline framework an existing online VIS model (GenVIS [13]) with the query propagation-based instance association. We first introduce the global instance assignment strategy to perform global optimal matching. Different from the previous online methods [13, 41, 10], which obtain per-frame matching cost, assign labels locally on the beginning frame and propagate across frames, we collect the predictions across frames, compute the global matching cost with the video segmentation ground truth and supervise the model with the global instance assignment, encouraging features across the video to be optimized for a global optimal objective. As online methods focus on improving the representative ability of the semantic instance embeddings [38, 34], which is achieved via learning more discriminative semantic embedding in those using heuristic matching [34, 15], or via reviewing the memory of semantic embedding across frames [13] in query propagation-based methods, the spatial features are not comprehensively investigated. We further propose spatio-temporal enhancement module, leveraging the spatial information from the previous frame to enhance the temporal consistency. We perform spatial matting on the pixel embeddings to retrieve the instance-wise spatial features and adopt the cross-attention layer to aggregate the spatial and semantic features across frames, thus realizing the spatio-temporal enhancement. As shown in Figure 1, the previous online method [13] produces temporally inconsistent results, as exemplified by an object abruptly appearing in front of the cat in a mid-frame, while the proposed TCOVIS outperforms the previous one and generates temporally consistent predictions.", + "bbox": [ + 75, + 385, + 468, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate the effectiveness of the proposed method, ex", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "periments are conducted on four widely adopted VIS benchmarks, i.e., YouTube-VIS 2019 [37], YouTube-VIS 2021, YouTube-VIS 2022 and Occluded VIS (OVIS) [24]. Without bells-and-whistles, our proposed method achieves state-of-the-art performance on all benchmarks, outperforming other online methods, e.g., on YouTube-VIS 2021, TCOVIS achieves 49.5 AP and 61.3 AP with ResNet-50 and Swin-L backbones, respectively.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 517, + 210, + 857, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- TCOVIS performs a novel global instance assignment strategy for online video instance segmentation. The model is optimized for the global optimal objective to generate more temporally consistent predictions.", + "- The further proposed spatio-temporal enhancement module captures the spatial feature and aggregates it with the semantic feature between frames, which fully utilizes the spatial information and facilitates the temporal consistency enhancement.", + "- The proposed method achieves state-of-the-art performance on four widely used video instance segmentation benchmarks (YouTube-VIS 2019/2021/2022 and OVIS). Such achievements demonstrate the effectiveness of our proposed method." + ], + "bbox": [ + 517, + 237, + 890, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 482, + 648, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Offline Video Instance Segmentation. Offline methods take as input the whole video and predict instance sequence for all frames at once. Mask propagation and box ensemble techniques are used to improve the predictions and association [1, 2, 19], but they are not end-to-end learnable due to the complex inference process. VisTR [31] extends DETR [42] from the image domain and introduces the transformer [27] to the VIS domain. EfficientVIS [35] and IFC [16] relax the heavy overhead of VisTR via an iterative query-video interaction and memory token communication, respectively. TeViT [39] contains a vision transformer [6] backbone instead of CNN and efficiently builds correspondence between the instance and query. VITA [14] models relationships among instances with the distilled condensed object tokens, without using the dense spatiotemporal backbone features. Offline methods exploit rich temporal knowledge from the whole clip and have the advantage of producing temporally consistent results, however, the offline manner is not suitable for the application in real-time scenarios.", + "bbox": [ + 496, + 507, + 890, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Online Video Instance Segmentation. Instead of processing the entire video before predictions, online methods only leverage the information from the previous frames and segment the video frame-by-frame. The association paradigms of the previous online methods roughly fall into two groups: Tracking-by-detection and Query-propagation.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1098", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most online methods [37, 29, 21, 34] follow the tracking-by-detection paradigm. MaskTrack R-CNN [37] is the baseline method and extends the Mask R-CNN [11] with an extra tracking head for temporal association. CrossVIS [38] proposes a crossover learning scheme to utilize the current contextual information for other frames. VISOLO [9] builds on the image instance segmentation method SOLO [30] and takes advantage of the grid form previous information for memory matching and features aggregation. MinVIS [15] and IDOL [34] make use of the discriminative instance embeddings for matching between frames. With the heuristic matching technique designed for instance association across frames, temporal inconsistency comes from the frame-wise modeling before tracking.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object association with query propagation has been explored in the multi-object tracking (MOT) task [22, 40]. TrackFormer [22] tracks the seen objects of previous frames with a track query subset and detects the newly appeared objects in current frame with an extra object query subset. MOTR [40] extends the paired-frames training scheme to multiple frames for long-range temporal association. The query propagation-based object association is recently introduced to VIS [41, 10, 13]. ROVIS [41] follows TrackFormer [22] to detect and track instances with two subsets of queries. InsPro [10] and GenVIS [13] propagate the queries without heuristics, i.e., handcrafted rules to combine two types of queries, and achieve association across frames. However, previous query propagation-based methods also propagate the local assignment to supervise the model with the local optimal results, which leads to temporal inconsistency across the entire video. We adopt the query propagation framework but introduce the global instance assignment to enhance the temporal consistency.", + "bbox": [ + 75, + 305, + 472, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 614, + 169, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video clip with consecutive image frames, online video instance segmentation methods generate frame-level object instances upon an instance segmentation models [26, 5], utilizing the instance queries propagated from previous frames [35, 13]. We have already discussed that better performance of the online VIS method relies on more temporally consistent instance features among a video. To this end, we propose a novel end-to-end online VIS method TCOVIS (Figure 2), to improve the temporal consistency of instance features via the global instance assignment strategy and spatio-temporal enhancement module. In this section, we first introduce the online VIS pipeline in Section 3.1. Then the details of the proposed global instance assignment strategy and spatio-temporal enhancement module will be described in Section 3.2 and Section 3.3, respectively. Finally, in Section 3.4, we describe the overall loss for training the model end-to-end.", + "bbox": [ + 75, + 643, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Online Video Instance Segmentation", + "text_level": 1, + "bbox": [ + 498, + 90, + 815, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following state-of-the-art VIS methods [15, 4], we adopt the advanced Masked-attention Mask Transformer (Mask2Former [5]) as the image instance segmentation network (Img. SegNet.) in this paper. Assume that the input video clip with T frames is denoted as $x \\in \\mathbb{R}^{T \\times 3 \\times H \\times W}$ . With each frame $H \\times W$ as input, frame-wise activation map is extracted by the backbone and Transformer encoder. Then following the query-based mechanism of DETR [3], $N_{fq}$ object queries of $C$ dimensions are used to parse an input frame, which are called frame object queries $f \\in \\mathbb{R}^{C \\times N_{fq}}$ . Each object in the frame is decoded by the frame object queries from the spatial features through a multiple-level Transformer decoder, and represented as an object embedding of $C$ dimensions. The object embeddings are used for classification and together with the pixel embeddings from the pixel decoder generating the mask for object instances. Class predictions are produced through a linear layer from the object embeddings. Mask embeddings are generated by an MLP linked to the object embeddings, and the model finally segments objects by pixel-wise dot product between per-pixel embeddings $\\mathcal{P} \\in \\mathbb{R}^{C \\times \\frac{H}{S} \\times \\frac{W}{S}}$ and mask embeddings $\\mathcal{M} \\in \\mathbb{R}^{C \\times N_{fq}}$ , where $S$ is the stride of the spatial feature map.", + "bbox": [ + 496, + 112, + 893, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As for online video segmentation, we follow the online scheme of GenVIS [13] which propagates the instance queries from previous frames. To resolve the computational limitation, the framework adopts VITA [14] that regards the frame object queries $f$ as a concise representation of objects in a frame and then feeds them into the Object Encoder $\\mathcal{E}$ for intra-frame relationship. The Object Decoder $\\mathcal{D}$ takes as input $N_v$ video instance queries $q$ and aggregates information from frame object queries. The temporal instance association is implemented through a query-based temporal propagation mechanism, where the output of $\\mathcal{D}$ , instance prototypes denoted as $p$ , are concise representations of instances [13], are not only utilized for classifying and segmenting in current frame at $t$ , but also serve as the instance queries for the next frame at $t + 1$ , i.e.,", + "bbox": [ + 496, + 460, + 893, + 688 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {t + 1} = p ^ {t} = \\mathcal {D} \\left(q ^ {t}, \\mathcal {E} \\left(f ^ {t}\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 694, + 890, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With this propagation mechanism, the model can simply run in an online manner and associate the frame-wise outputs without heuristic matching algorithms.", + "bbox": [ + 496, + 718, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We leave out the Instance Prototype Memory module in [13]. During the training process, we freeze the image instance segmentation model and only train the following modules, to efficiently make use of memory. More implementation details are described in Section 4.2.", + "bbox": [ + 496, + 763, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Global Instance Assignment", + "text_level": 1, + "bbox": [ + 500, + 847, + 751, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the query propagation mechanism discussed above, the associated video instance queries along the temporal", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1099", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0b62bb26eae29dcf53ab8d48dd7ef1e7eee710a0cabdaf22d0115d12b1375694.jpg", + "image_caption": [ + "Figure 2. (a) The overall illustration of our TCOVIS. It supervises the model with a global optimal objective during training and utilizes spatial features via the enhancement module between frames. (b) Details of global instance assignment. Different from the local matching and propagating technique, we conduct global optimal matching and assignment. (c) In the spatio-temporal enhancement module, we perform matting on pixel embeddings according to predicted masks and then aggregate the spatial and semantic features between frames to enhance temporal consistency. Numbers in colors denote the local matching loss and global matching loss (Best viewed in color)." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 872, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dimension $\\{q_{k}^{1:T}\\}_{1:N_{v}}$ , are used to extract the features of a unique instance, i.e. the $k$ -th instance, throughout the whole input video clip. Meanwhile, the associated instance prototypes $\\{p_{k}^{1:T}\\}_{1:N_{v}}$ , i.e. the corresponding outputs from $\\mathcal{D}$ , represent this unique instance within the whole video. For brevity, we define the associated instance prototypes/features as a set of prototypes/features.", + "bbox": [ + 75, + 478, + 470, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previously, online VIS methods that associate frame-level results without heuristic matching algorithms [13, 41, 10] adopt a Local Matching and Propagating (LocPro) technique. This technique matches the predictions and ground truth on a frame level during training. In other words, it conducts one-to-one bipartite matching between ground truth instances and predictions of a frame (what we regard as Local Matching), and propagates the matching from previous frames to subsequent frames. Specifically, given a video clip consisting of $T$ consecutive frames, it first computes the pair-wise matching cost $\\mathcal{L}_{match}^{t}$ between predictions and ground truth instances on the initial frame (the first frame or the frame objects appear):", + "bbox": [ + 75, + 585, + 468, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a t c h}} ^ {t = 1} = \\lambda_ {c l s} \\mathcal {L} _ {c l s} ^ {t = 1} + \\lambda_ {b c e} \\mathcal {L} _ {b c e} ^ {t = 1} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} ^ {t = 1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 786, + 468, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The frame-wise $\\mathcal{L}_{match}^{t}$ is composed of categorical loss and mask loss. The categorical loss adopts the cross entropy loss $\\mathcal{L}_{cls}^{t}$ . The mask loss consists of a binary cross entropy loss $\\mathcal{L}_{bce}^{t}$ and a dice loss [23]. From the cost matrix, it follows DETR [3] and uses Hungarian algorithm [18] for optimal frame-level matching.", + "bbox": [ + 76, + 809, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed in Section 1, Local Matching only obtains optimal matching on the initial frame, and cannot achieve the global optimal matching of the whole video. As illustrated in Figure 2 (b), both dark and light green prototypes attempt to represent the person. With local matching, the ground truth will be assigned to the light one with a smaller local matching cost, however, the dark one performs better from a global perspective and is supposed to be assigned. Training the model with LocPro forces all the predictions of subsequent frames to conform to the initial frame, resulting in temporal inconsistency of the instance features among the video, because inappropriate previous matching brings accumulative error to the model.", + "bbox": [ + 496, + 479, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce Global Instance Assignment (GIA) strategy and expect two functionalities: (1) all frames in the video clip are considered when conducting global matching, and (2) the global assignment encourages the instance features among frames to be optimized for a global optimal objective, both of which serve the temporal consistency of instance features.", + "bbox": [ + 496, + 681, + 892, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, different from previous online VIS methods [8, 34] that conduct frame-level local matching and provide supervision signals to each frame according to the matched pairs, we leave out the halfway matching and supervision. Consecutive input frames of a video clip pass through the VIS model in an online manner and the model generates the predictions of all frames $\\{\\hat{\\mathbf{y}}_k\\}_{1:N_v} =$", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1100", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\{\\hat{y}_k^t\\}_{1:N_v}^{1:T}$ , each of which consists of a category probability $\\hat{c}_k^t$ and a segmentation mask probability $\\hat{m}_k^t$ . To conduct global assignment, we collect predictions of all frames and as well the ground truth video instance segmentation $\\{\\mathbf{y}_k\\}_{1:N_{gt}} = \\{y_k^{1:T}\\}_{1:N_{gt}}$ , including the category label $c_k$ and its binary segmentation masks $\\mathbf{m}_k = m_k^{1:T}$ . Since the ground truth of an instance only has one category label, we first compute the average predicted category probability across a video clip: $\\bar{c}_k = \\sum_{t=1}^T \\hat{c}_k^t / T$ and also collect the masks $\\hat{\\mathbf{m}}_k = \\hat{m}_k^{1:T}$ . The global matching cost is defined as:", + "bbox": [ + 76, + 89, + 472, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {m a t c h}} ^ {\\text {g l o b a l}} = \\lambda_ {c l s} \\mathcal {L} _ {c l s} \\left(c _ {k}, \\bar {c} _ {\\sigma (k)}\\right) + \\lambda_ {b c e} \\mathcal {L} _ {b c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\sigma (k)}\\right) \\\\ + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\sigma (k)}\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 261, + 468, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma \\in \\mathfrak{S}_{N_v}$ is a permutation of $N_{v}$ elements. One-to-one bipartite global matching between $\\{\\hat{\\mathbf{y}}_k\\}_{1:N_v}$ and $\\{\\mathbf{y}_k\\}_{1:N_{gt}}$ is performed to find the global optimal assignment and the objective can be formally described as:", + "bbox": [ + 76, + 313, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\sigma} = \\underset {\\sigma \\in \\mathfrak {S} _ {N _ {v}}} {\\arg \\max } \\sum_ {k = 1} ^ {N _ {g t}} \\mathcal {L} _ {\\text {m a t c h}} ^ {\\text {g l o b a l}} (\\mathbf {y} _ {k}, \\hat {\\mathbf {y}} _ {\\sigma (k)}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 381, + 468, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following prior work [28, 7, 42], we use Hungarian algorithm [18] to search for the global optimal assignment. In contrast to the prior methods that compute matching loss only for $t = 1$ , our method considers the masks and ground truth of the whole clip, computes the matching loss globally, and conducts the global assignment. Finally, given the global optimal assignment, we use the $N_{gt}$ matched video-level predictions to supervise the model with the globally matched instances across the entire video.", + "bbox": [ + 75, + 430, + 468, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the proposed assignment strategy, GIA, the temporal consistency of instance features across the video clip can be effectively enhanced, since we consider all frames as a whole to search for the optimal objective. Specifically, when the set of instance features across the video represents the target instance well in the first few frames, but fails to track it later, this strategy helps to find a more appropriate set to be optimized. As the global matching cost is lower, the selected features fit the target more closely and are more temporally consistent.", + "bbox": [ + 75, + 566, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Spatio-temporal Enhancement", + "text_level": 1, + "bbox": [ + 76, + 726, + 349, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous online video instance segmentation methods focus on improving the representative ability of the semantic instance embeddings [38, 34, 15, 13]. The spatial features are not comprehensively investigated to boost the temporal association for online VIS. Thus, we further introduce Spatio-temporal Enhancement module (STE), leveraging the spatial information from the previous frame to enhance the temporal consistency of the online model.", + "bbox": [ + 75, + 750, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the $t$ -th frame in the video clip, with the framework described in Section 3.1, the mask embedding of", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the $k$ -th instance is generated from the instance prototype through an MLP: $\\mathcal{M}_k^t = MLP(p_k^t)$ , and then the model segments the mask $\\hat{m}_k^t$ by pixel-wise dot product between $\\mathcal{P}_k^t$ and $\\mathcal{M}_k^t$ , which can be formulated as: $\\hat{m}_{k,i,j}^{t} = \\langle \\mathcal{P}_{k,i,j}^{t},\\mathcal{M}_{k}^{t}\\rangle$ , where $i$ and $j$ denote the spatial position of the pixel.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Figure 2 (c), the spatial information of the frame is encoded in the pixel embeddings. To extract instance-wise spatial features, they can be exploited together with the predicted mask. Specifically, we perform spatial matting on pixel embeddings $\\mathcal{P}_k^t$ , similar to image matting, to retrieve the instance-wise pixel embedding according to $\\hat{m}_k^t$ . For each instance, we conduct pixel-wise multiplication between the original pixel embeddings and the binary mask to obtain the retrieved embeddings:", + "bbox": [ + 496, + 184, + 892, + 320 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {k} ^ {t} = \\mathcal {P} _ {k} ^ {t} \\odot \\hat {m} _ {k} ^ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 333, + 890, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes the element-wise multiplication. Many of the retrieved embeddings $\\mathcal{R}_k^t$ are redundant since they describe the same instance, and directly mining the spatial information with them is computationally inefficient. Average pooling is adopted to obtain a concise representation of the spatial features for each instance:", + "bbox": [ + 496, + 364, + 893, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {k} ^ {t} = \\frac {\\sum_ {i , j} \\mathcal {R} _ {k , i , j} ^ {t}}{\\sum_ {i , j} \\mathbb {1} (\\hat {m} _ {k , i , j} ^ {t} = 1)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 468, + 890, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is the indicator function. The concise spatial features $S_{k}^{t}$ are sent to the next frame for temporal association.", + "bbox": [ + 496, + 518, + 890, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Having received from the previous frame the propagated semantic instance queries, i.e., $q^{t + 1} = p^t$ , and the spatial features, we follow the standard multi-head cross-attention layer (MHCA) [27] to incorporate the features from two aspects. The instance prototype $p^t$ is used as the query to decode the spatial features:", + "bbox": [ + 496, + 551, + 890, + 641 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {k} ^ {t + 1} = \\operatorname {M H C A} \\left(p _ {k} ^ {t}, \\mathcal {S} _ {1: N _ {v}} ^ {t}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 654, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q_{k}^{t + 1}$ is the updated instance query now. Notably, the positional embedding is shared by the spatial feature and instance query with regard to the same instance, which helps the model align the instance-wise information between frames. Finally, the updated instance query is fed into the Object Decoder $\\mathcal{D}$ to produce spatio-temporally enhanced features.", + "bbox": [ + 496, + 686, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By performing the proposed enhancement module on the spatial features between frames, we effectively boost the spatio-temporal association of the features. In this way, the temporal consistency of the features is enhanced via the information from the spatial dimension, and the online video instance segmentation model manages to predict more temporally consistent results.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1101", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall Loss", + "text_level": 1, + "bbox": [ + 76, + 90, + 210, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall loss for training with a video clip as input is a linear combination of categorical and mask losses using the one-to-one global optimal assignment $\\hat{\\sigma}$ :", + "bbox": [ + 76, + 113, + 468, + 159 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {o v e r a l l}} = \\lambda_ {c l s} \\mathcal {L} _ {c l s} \\left(c _ {k}, \\bar {c} _ {\\hat {\\sigma} (k)}\\right) + \\lambda_ {b c e} \\mathcal {L} _ {b c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\hat {\\sigma} (k)}\\right) \\\\ + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\tilde {\\sigma} (k)}\\right), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 172, + 468, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{cls}$ is the cross entropy loss, $\\mathcal{L}_{bce}$ is the binary cross entropy loss and $\\mathcal{L}_{dice}$ is the dice loss [23].", + "bbox": [ + 76, + 220, + 468, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 263, + 209, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluated the proposed TCOVIS on four benchmark datasets. Furthermore, we provided in-depth ablation studies of the effectiveness of TCOVIS. Finally, we presented several visualizations of the predictions from our model.", + "bbox": [ + 76, + 289, + 468, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 375, + 179, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated our approach on four VIS datasets: YouTube-VIS 2019 dataset [37], YouTube-VIS 2021 dataset [37], YouTube-VIS 2022 dataset [37] and OVIS dataset [24]. We present a brief description of them:", + "bbox": [ + 76, + 398, + 468, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "YouTube-VIS 2019 & 2021 & 2022: YouTube-VIS 2019 [37] is the first VIS dataset and comprises 40 predefined categories of objects. This dataset included 4,883 unique video instances with 131,000 high-quality manual annotations. We followed the widely utilized training/test set split: 2,238 videos were selected for training, 302 videos were adopted for validation and 343 videos were used for testing. Further, YouTube-VIS 2021 improved the 40-category label set and added 4883 more unique video instances. Then, YouTube-VIS 2022 contained 71 additional long evaluation videos on the top of YouTube-VIS 2021.", + "bbox": [ + 75, + 458, + 468, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OVIS: Occluded video instance segmentation (OVIS) is also a challenging VIS dataset. OVIS included 901 videos in total with 25 semantic categories. This dataset contained 5,223 unique video instances and we followed the widely utilized training/test set split: 607 videos were selected for training, 607 videos were used for validation and 154 videos were adopted for testing.", + "bbox": [ + 75, + 625, + 468, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [37], the video-level average precision (AP) and average recall (AR) were adopted as the evaluation metrics on both YouTube-VIS and OVIS.", + "bbox": [ + 76, + 731, + 468, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 786, + 294, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopted the framework of GenVIS [13] which is built on VITA [14], but left out the similarity loss in VITA and the memory module in GenVIS. With the global assignment, the total loss as well the hyper-parameters were set the same as the video-level loss in VITA, which is a temporally extended loss function [16]. The model was trained", + "bbox": [ + 76, + 809, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with pseudo-videos from COCO images [20] as data augmentation, and with a batch size of 8 video clips of 6 frames. As we froze the backbone and image segmentation model, all experiments were conducted with 8 RTX 2080 Ti GPUs. The method was implemented on detectron2 [36].", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 500, + 176, + 638, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the standard evaluation metrics [37], we compared TCOVIS with state-of-the-art approaches on four VIS benchmarks: YouTube-VIS 2019/2021/2022 and OVIS.", + "bbox": [ + 498, + 200, + 890, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "YouTube-VIS 2019&2021. From Table 1, we can observe that TCOVIS has achieved very competitive performance using both lightweight backbones (ResNet-50) and powerful ones (Swin-L). Moreover, our method can even show better performance than offline methods, such as VITA [14] and offline version GenVIS [13]. On the more difficult YouTube-VIS 2021 dataset, TCOVIS surpasses GenVIS [13] in AP not only with ResNet-50 by 2.4 but also with Swin-L by 1.7.", + "bbox": [ + 496, + 246, + 890, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "YouTube-VIS 2022. As shown in Table 2, TCOVIS performed best in both AP and AR on YouTube-VIS 2022 dataset, which is more challenging than 2019&2021 datasets. Especially with powerful backbone Swin-L, TCOVIS outperformed GenVIS [13] in AP with a huge margin of 4.9, which shows the effectiveness of our method for the complex scenarios.", + "bbox": [ + 496, + 382, + 890, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OVIS. Table 3 presents the comparisons on OVIS dataset, and we can find that TCOVIS also achieved the best 46.7 AP and 19.1 AR with Swin-L backbone. The results demonstrate that TCOVIS can deal with complicated situations where objects are heavily occluded in others. With ResNet-50 backbone, the performance is still the second best compared to the previous state-of-the-art methods.", + "bbox": [ + 496, + 488, + 890, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 604, + 653, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we provided ablation studies and discuss the effects of different settings in the proposed method. The experiments are conducted with a ResNet-50 [12] backbone on YouTube-VIS 2019/2021 [37] valid set.", + "bbox": [ + 496, + 628, + 890, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of the proposed assignment strategy and enhancement module. The ablation studies on the global instance assignment strategy and the spatio-temporal enhancement module are shown in Table 4. As for the assignment strategy, compared to LocPro as the baseline, the model with the proposed global assignment outperformed the baseline model by more than 0.8 on AP, $\\mathrm{AP}_{50}$ and $\\mathrm{AP}_{75}$ on YouTube-VIS 2019, and more than 1.2 on YouTube-VIS 2021. The consistently significant improvement indicates that the strategy with global optimal matching contributes to better overall segmentation, while LocPro only considers the local optimal results. Besides, the baseline forces the posterior features to conform to those at the very beginning leading to temporal inconsistency, since the accumulative", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1102", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a3afe1d7dfc1c1bdcd18c4f331fe65ae6c3ec71853e1b03e3ecf6c38176f6697.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTypeYouTube-VIS 2019YouTube-VIS 2021
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AR_1\\)\\(AR_{10}\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AR_1\\)\\(AR_{10}\\)
ResNet-50EfficientVIS [35]Offline37.959.743.040.346.634.057.537.333.842.5
IFC [16]Offline41.265.144.642.349.635.255.937.732.642.9
Mask2Former-VIS [4]Offline46.468.050.0--40.660.941.8--
\\(TeViT^{\\dagger}\\)[39]Offline46.671.351.644.954.337.961.242.135.144.6
SeqFormer [33]Offline47.469.851.845.554.840.562.443.736.148.1
VITA [14]Offline49.872.654.549.461.045.767.449.540.953.6
\\(GenVIS_{semi-online}\\)[13]Offline51.372.057.849.560.046.367.050.240.653.2
Cross VIS [38]Online36.356.838.935.640.734.254.437.930.438.2
VISOLO [9]Online38.656.343.735.742.536.954.740.230.640.9
MinVIS [15]Online47.469.052.145.755.744.266.048.139.251.7
IDOL [34]Online49.574.052.947.758.743.968.049.638.050.9
\\(GenVIS_{online}\\)[13]Online50.071.554.649.559.747.167.551.541.654.7
TCOVISOnline52.373.557.649.860.249.571.253.841.355.9
Swin-LSeqFormer [33]Offline59.382.166.451.764.451.874.658.242.858.1
Mask2Former-VIS [4]Offline60.484.467.0--52.676.457.2--
VITA [14]Offline63.086.967.956.368.157.580.661.047.762.6
\\(GenVIS_{semi-online}\\)[13]Offline63.885.768.556.368.460.180.966.549.164.7
MinVIS [15]Online61.683.368.654.866.655.376.662.045.960.8
IDOL [34]Online64.387.571.055.669.156.180.863.545.060.1
\\(GenVIS_{online}\\)[13]Online64.084.968.356.169.459.680.965.848.765.0
TCOVISOnline64.186.669.555.869.061.382.968.048.665.1
", + "bbox": [ + 94, + 88, + 867, + 455 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e905c7d1ec9ebf3ac5787ed27e00928460610c74835ab686f6f93ef9d21c8c58.jpg", + "table_caption": [ + "Table 1. Quantitative results on YouTube-VIS 2019 and 2021 validation sets. The results are respectively grouped by method types (Offline or Online) and backbone networks (ResNet-50 and Swin-L). We bold the best performance and underline the second. † denotes using MsgBoxIF [39] backbone which has a similar weight scale with ResNet-50." + ], + "table_footnote": [], + "table_body": "
MethodTypeAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
ResNet-50VITA [14]Offline32.653.939.330.342.6
GenVIS [13]Offline37.258.542.933.240.4
MinVIS [15]Online23.347.919.320.228.0
GenVIS [13]Online37.561.641.532.642.2
TCOVISOnline38.659.441.632.846.7
Swin-LVITA* [14]Offline41.163.044.039.344.3
GenVIS [13]Offline44.369.944.939.948.4
MinVIS* [15]Online33.154.833.729.536.6
GenVIS [13]Online45.169.147.339.848.5
TCOVISOnline51.073.053.541.756.5
", + "bbox": [ + 80, + 518, + 465, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "error impairs the model during the training process. Our proposed assignment strategy encourages the features of all time to fit the global optimal objective, which effectively enhances the temporal consistency.", + "bbox": [ + 76, + 744, + 468, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As for the effectiveness of the spatio-temporal enhancement module, the comparison is also shown in Table 4. The results show that the enhancement module brings improvements of 0.9 AP on YouTube-VIS 2019&2021, compared to using the proposed assignment strategy individually. In particular, the proposed module improved the performance", + "bbox": [ + 76, + 810, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d86aed8c328db85a5ce46cf38bdad612b05bce1891794c08330277e4d3e0f8c1.jpg", + "table_caption": [ + "Table 2. Quantitative results on YouTube-VIS 2022 validation dataset. We bold the highest accuracy and underline the second. $\\star$ : Reproduced by [13]." + ], + "table_footnote": [], + "table_body": "
MethodTypeAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
ResNet-50\\( TeViT^†[39] \\)Offline17.434.915.011.221.8
VITA [14]Offline19.641.217.411.726.0
GenVIS [13]Offline34.559.435.016.638.3
CrossVIS [38]Online14.932.712.110.319.8
VISOLO [9]Online15.331.013.811.121.7
MinVIS [15]Online25.045.524.013.929.7
IDOL [34]Online30.251.330.015.037.5
GenVIS [13]Online35.860.836.216.339.6
TCOVISOnline35.360.736.615.739.5
Swin-LVITA [14]Offline27.751.924.914.933.0
GenVIS [13]Offline45.469.247.818.949.0
MinVIS [15]Online39.461.541.318.143.3
IDOL [34]Online42.665.745.217.949.6
GenVIS [13]Online45.269.148.419.148.6
TCOVISOnline46.770.949.519.150.8
", + "bbox": [ + 503, + 518, + 888, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative results on OVIS validation set. We bold the highest accuracy and underline the second. † denotes using MsgBoxT [39] backbone.", + "bbox": [ + 498, + 744, + 890, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by 1.2 and 1.7 in $\\mathrm{AP}_{50}$ on two datasets, respectively. The performance improvements demonstrate that the proposed module effectively captures the spatial information from the previous frame and aggregates the semantic and spatial features across time. As a result, the delicate spatio-temporal design further enhances the temporal consistency of fea", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1103", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7964e104a624c7656c8dafa917f5b2db2b2e9ac04840fcb9019e39905dee4567.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LocProGIASTEYouTube-VIS 2019YouTube-VIS 2021
AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
✓--50.671.556.147.658.247.468.152.139.853.1
-✓-51.472.357.049.559.948.669.553.640.655.0
-✓✓52.373.557.649.860.249.571.253.841.355.9
", + "bbox": [ + 133, + 88, + 831, + 170 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1959e3d8b7ceee26ef5e77cbaf9acbe21a12a9800c058577c308524c1e08f7ed.jpg", + "table_caption": [ + "Table 4. Ablation study of the method for Global Instance Assignment strategy (GIA) and the Spatio-temporal Enhancement module (STE) on the YouTube-VIS 2019 / 2021 validation sets. LocPro denotes the Local Matching and Propagating technique." + ], + "table_footnote": [], + "table_body": "
ArchitectureAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Cross-Attn.49.571.253.841.355.9
Concat.47.968.652.741.455.1
Self-Cross48.270.051.841.054.1
Dec. Mod.48.569.753.940.654.7
Resp. Pos.48.870.352.541.555.5
Ins. Attn.49.270.653.540.955.1
", + "bbox": [ + 84, + 214, + 454, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study of the module manipulations of the Spatio-temporal Enhancement module on YouTube-VIS 2021.", + "bbox": [ + 75, + 333, + 468, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tures for the online video instance segmentation model.", + "bbox": [ + 75, + 373, + 442, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Incorporating the assignment strategy and the enhancement module, our proposed method gained remarkable performance improvements of 1.6 AP on YouTube-VIS 2019 and 2.1 AP on YouTube-VIS 2021 over the LocPro baseline. In a nutshell, the experimental results indicate that the temporally consistent features learned by our proposed TCOVIS significantly boost the performances on online video instance segmentation.", + "bbox": [ + 75, + 388, + 468, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different manipulations of spatio-temporal enhancement. In Table 5, we compared our proposed spatiotemporal enhancement module with other optional manipulations. Cross-Attn. denotes our proposed manipulation following the standard multi-head cross-attention layer to decode the spatial features with shared positional embedding described in Section 3.3. Concat. indicates that we", + "bbox": [ + 75, + 510, + 468, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "concatenated the corresponding spatial feature and prototype of an instance followed by an MLP to get the updated query. Self-Cross stands for adding an extra self-attention layer ahead of the cross-attention layer. Dec. Mod. is decoder modulation, reviewing the spatial feature for every Transformer decoder layer. Resp .Pos. denotes respective positional embeddings were used for the spatial feature and prototype when we performed cross-attention. Ins. Attn. indicates the instance-wise decoding in cross-attention layer.", + "bbox": [ + 496, + 218, + 893, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown, compared to all its counterparts, our proposed manipulation achieved the best performance. The first three variants are related to the feature aggregation, where $\\text{Concat}$ is too naive to model the spatial and semantic features, while Self-Cross obscures the spatial feature. We inferred the performance decrease of Dec. Mod. comes from the information redundancy when aggregating them in every layer. The last two experiments studied the instance-wise correspondence. In the Resp. Pos. setting, explicit correspondence for spatial and semantic features of the same instance between frames is absent. Ins. Attn. only focuses on the instance itself across time and can slightly enhance the temporal consistency, however, neglecting the spatial features of other instances, the module fails to capture the spatial relationship among instances. The results confirm that the proposed manipulation effectively exploits the spatial information along the temporal dimension to enhance the", + "bbox": [ + 496, + 359, + 893, + 616 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e02a590bf053b004b385f4df18063bc998af74e6a36989d7d8e6857a0c3cebd4.jpg", + "image_caption": [ + "Figure 3. Qualitative results of TCOVIS, compared with VITA [14] and GenVIS [13]. On the left are the predictions on YouTube-VIS 2019 [37] and on the right are on OVIS [24]. Objects displayed in the same color denote the same instance. Our TCOVIS shows more temporally consistent results in these challenging scenes, where there are occlusions of the instance itself or others (Best viewed in color)." + ], + "image_footnote": [], + "bbox": [ + 102, + 636, + 493, + 851 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da32317aa43fe599f2c359f5c3d13a24bbf204d5c8a7ba75e4fba8512b47cdc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 636, + 867, + 851 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1104", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5cd4e3dded4c647a823986ce2103344de4ff575f6242de1dec23ce0ce7ec814d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 122, + 89, + 308, + 172 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1b347761c3d667e652ae9bc45f12f280e35fb98a79414620aadb9107f8918332.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 90, + 493, + 171 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c74aef73dd4df29b67930669ca80865f483e581e23c7211eb9eaa73a791936da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 90, + 679, + 171 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/577032a6dcc42c1025c906b976e3bc80e4d975493c51ec57cc3c7df87a057a31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 90, + 864, + 171 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8e2a2751357596502033264904d4a95c3fbc66562c583f8246492ab9ad64813a.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 122, + 172, + 307, + 253 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9b1384c83a01b1d9d8eda31e6657b274e005e21aff0827d135f2457737b86cd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 172, + 493, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/57ce6d247a3fb0d2e897c3cba32d1861e4c4e2d3fdd702f5642020ecf4234484.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 172, + 679, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/efaf9b8db636caf212f4e04c6d1752d443134b54528dbbb1303617a1b36bf705.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 172, + 864, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/de8517a798c3821fc06cf9feca1b680257d5d64cbb4da75e304319a8e102dac9.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 124, + 253, + 308, + 334 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0203e78a5af85d94493829199498ac92de7e787fc00c2efb26bcf7913ced7953.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 253, + 493, + 334 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1675ef9e6fc7e01c827ec9b67d1817a9cf8ec1d3ac1a9e26bbdd59e7247836c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 253, + 679, + 334 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2b244cfd36a47a45a5006e60a192d09261b8ce75e94d65ef55e5130c4bfbf370.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 253, + 864, + 334 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/db4b9a94343f3c7d4ea8f63b0c91408935f2d0193180c3842369e57082145df4.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 122, + 335, + 308, + 416 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dee460dda09ee30d0615cca21ab36e6242da609cd16d2d5949c405246e06306b.jpg", + "image_caption": [ + "Figure 4. Qualitative results of TCOVIS in four challenging cases on YouTube-VIS 2019 [37] and OVIS [24]: (a) fast movement, (b) heavy occlusion, (c) low resolution, and (d) crowded scene. Objects displayed in the same color denote the same instance. The qualitative results demonstrate the effectiveness and robustness of TCOVIS (Best viewed in color)." + ], + "image_footnote": [], + "bbox": [ + 308, + 335, + 493, + 416 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/72aadeac444e2f6c38758e6e3f9dc4f860c6209b04be8df8a459799051d17ab0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 335, + 679, + 416 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/09056a28bbe9e89a3c8e681e0643adbc5641a6b7c843d60db5b5fc42c08b8567.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 335, + 864, + 416 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "temporal consistency of features.", + "bbox": [ + 76, + 483, + 295, + 498 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5. Qualitative results", + "text_level": 1, + "bbox": [ + 76, + 511, + 253, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Figure 3, we show the qualitative comparisons of the proposed TCOVIS with VITA [14] and GenVIS [13] on YouTube-VIS 2019 and OVIS datasets. In the left scene where there is a zebra with self-occlusion, GenVIS fails to segment its head resulting in fragmented predictions, however, TCOVIS performs temporally consistent segmentation with impressive accuracy. On the right is a difficult case where there are two dogs with similar appearances grappling with each other. VITA incorrectly detects more than two dogs and fails to track the left one, while GenVIS fails to handle the margin of two instances, e.g. the nose and the paw of the left dog. TCOVIS successfully tracks and segments the instances, demonstrating its effectiveness.", + "bbox": [ + 75, + 537, + 468, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Figure 4, we provide more qualitative results of the proposed method in variously challenging cases, which are all chosen from the mentioned benchmarks [37, 24]. As shown in Figure 4 (a), our method successfully tracks the surfer and the surfboard with fast movement. In the second row, we present a difficult case with severe occlusion, in which our method performs admirably by accurately segmenting the claw (at the bottom) of a parrot despite the presence of a wooden stick that partially obstructs its lower body. Figure 4 (c) illustrates a low-resolution case and our method still achieves good performance. In Figure 4 (d),", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "we depict a crowded scene where TCOVIS is capable of handling multiple instances that share similar appearances and exhibit complex interactions. All the qualitative results in the challenging situations demonstrate the effectiveness and robustness of the proposed method.", + "bbox": [ + 498, + 483, + 890, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 580, + 617, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a new online video instance segmentation method, TCOVIS, to fully exploit the temporal information within a video and produce temporally consistent predictions. Based on the query propagation framework, we propose a global instance assignment strategy to perform global optimal matching with the consideration of the entire video and supervise the model with the global optimal objective. We further devise a spatio-temporal enhancement module to capture the spatial feature and aggregate it with the semantic feature between frames. The effectiveness of our method is evaluated with experimental results on YouTube-VIS 2019/2021/2022 and OVIS.", + "bbox": [ + 496, + 609, + 890, + 789 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 811, + 658, + 829 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603.", + "bbox": [ + 498, + 839, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1105", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, pages 158-177, 2020. 2", + "[2] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, pages 9739-9748, 2020. 1, 2", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 3, 4", + "[4] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 3, 7", + "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 1, 3", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2", + "[7] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In ICCV, pages 6910-6919, 2021. 5", + "[8] Yang Fu, Linjie Yang, Ding Liu, Thomas S Huang, and Humphrey Shi. Compfeat: Comprehensive feature aggregation for video instance segmentation. In AAAI, pages 1361-1369, 2021. 1, 4", + "[9] Su Ho Han, Sukjun Hwang, Seoung Wug Oh, Yeonschool Park, Hyunwoo Kim, Min-Jung Kim, and Seon Joo Kim. Visolo: Grid-based space-time aggregation for efficient online video instance segmentation. In CVPR, pages 2896-2905, 2022. 3, 7", + "[10] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. In NeurIPS, 2022. 1, 2, 3, 4", + "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[13] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In CVPR, pages 14623–14632, 2023. 1, 2, 3, 4, 5, 6, 7, 8, 9", + "[14] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. VITA: Video instance segmentation via object token association. In NeurIPS, 2022. 1, 2, 3, 6, 7, 8, 9" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] De-An Huang, Zhiding Yu, and Anima Anandkumar. MinVIS: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 1, 2, 3, 5, 7", + "[16] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 34, 2021. 1, 2, 6, 7", + "[17] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks for multiple object tracking and segmentation. NeurIPS, 34, 2021. 1", + "[18] Harold W Kuhn. The hungarian method for the assignment problem. NRL, 1955. 4, 5", + "[19] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In ICCV, pages 1739-1748, 2021. 1, 2", + "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6", + "[21] Dongfang Liu, Yiming Cui, Wenbo Tan, and Yingjie Chen. Sg-net: Spatial granularity network for one-stage video instance segmentation. In CVPR, pages 9816-9825, 2021. 3", + "[22] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 3", + "[23] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571, 2016. 4, 6", + "[24] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. IJCV, 130(8):2022-2039, 2022. 2, 6, 8, 9", + "[25] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 14454-14463, 2021. 1", + "[26] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, pages 282-298, 2020. 1, 3", + "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2, 5", + "[28] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. 5", + "[29] Tao Wang, Ning Xu, Kean Chen, and Weiyao Lin. End-to-end video instance segmentation via spatial-temporal graph neural networks. In ICCV, pages 10797-10806, 2021. 3", + "[30] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, pages 649-665, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1106", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, pages 8741-8750, 2021. 1, 2", + "[32] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In CVPR, pages 12352-12361, 2021. 1", + "[33] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, pages 553-569, 2022. 1, 7", + "[34] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, pages 588-605, 2022. 1, 2, 3, 4, 5, 7", + "[35] Jialian Wu, Sudhir Yarram, Hui Liang, Tian Lan, Junsong Yuan, Jayan Eledath, and Gerard Medioni. Efficient video instance segmentation via tracklet query and proposal. In CVPR, pages 959-968, 2022. 1, 2, 3, 7", + "[36] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6", + "[37] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, pages 5188-5197, 2019. 1, 2, 3, 6, 8, 9", + "[38] Shusheng Yang, Yuxin Fang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Crossover learning for fast online video instance segmentation. In ICCV, pages 8043-8052, 2021. 1, 2, 3, 5, 7", + "[39] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Wenyu Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In CVPR, pages 2885-2895, 2022. 1, 2, 7", + "[40] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, pages 659-675. Springer, 2022. 3", + "[41] Zitong Zhan, Daniel McKee, and Svetlana Lazebnik. Robust online video instance segmentation with track queries. arXiv preprint arXiv:2211.09108, 2022. 1, 2, 3, 4", + "[42] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable $\\{\\mathrm{detr}\\}$ : Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 5" + ], + "bbox": [ + 78, + 90, + 468, + 700 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1107", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_model.json b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7aceb3ba80f372d00d5461be1d12e826069faec --- /dev/null +++ b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "TCOVIS: Temporally Consistent Online Video Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.181, + 0.741, + 0.2 + ], + "angle": 0, + "content": "Junlong Li Bingyao Yu Yongming Rao Jie Zhou Jiwen Lu*" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.2, + 0.706, + 0.217 + ], + "angle": 0, + "content": "Department of Automation, Tsinghua University, China" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.217, + 0.808, + 0.235 + ], + "angle": 0, + "content": "Beijing National Research Center for Information Science and Technology, China" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.473, + 0.695 + ], + "angle": 0, + "content": "In recent years, significant progress has been made in video instance segmentation (VIS), with many offline and online methods achieving state-of-the-art performance. While offline methods have the advantage of producing temporally consistent predictions, they are not suitable for real-time scenarios. Conversely, online methods are more practical, but maintaining temporal consistency remains a challenging task. In this paper, we propose a novel online method for video instance segmentation, called TCOVIS, which fully exploits the temporal information in a video clip. The core of our method consists of a global instance assignment strategy and a spatio-temporal enhancement module, which improve the temporal consistency of the features from two aspects. Specifically, we perform global optimal matching between the predictions and ground truth across the whole video clip, and supervise the model with the global optimal objective. We also capture the spatial feature and aggregate it with the semantic feature between frames, thus realizing the spatio-temporal enhancement. We evaluate our method on four widely adopted VIS benchmarks, namely YouTube-VIS 2019/2021/2022 and OVIS, and achieve state-of-the-art performance on all benchmarks without bells-and-whistles. For instance, on YouTube-VIS 2021, TCOVIS achieves 49.5 AP and 61.3 AP with ResNet-50 and Swin-L backbones, respectively. Code is available at https://github.com/jun-long-li/TCOVIS." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.741, + 0.21, + 0.758 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.768, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Video instance segmentation (VIS) is a challenging and representative video understanding task recently introduced in [37]. It aims at detecting, segmenting and tracking instances across a video. VIS is attracting increasing attention for various real-world applications such as video editing, video surveillance, augmented reality and autonomous driving. Recently introduced VIS methods can be roughly" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.408 + ], + "angle": 0, + "content": "categorized into two groups: offline methods and online methods. Offline methods [2, 16, 19, 31, 33, 35] take as input the whole video and perform the segmentation of instance sequence for the whole video at once. Online methods [8, 34, 17, 10, 38], on the contrary, take as input a video frame by frame and generate the pre-frame object instances while associating the frame-wise results across frames. Both offline and online methods have achieved impressing performance on the VIS task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.895, + 0.565 + ], + "angle": 0, + "content": "Offline methods have an inherent advantage in producing temporally consistent predictions, since delicate temporal communication and association mechanisms can be adopted throughout the video [39, 33, 14] to handle the overall temporal information and impose an explicit constraint on the temporal consistency. However, the video-in and video-out offline manner is not suitable for real-time scenarios. Conversely, online methods are more practical and making considerable progress but suffer from temporal inconsistency (as shown in Figure 1), remaining a great challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Online methods rely on specific instance association applied across frames, since only one frame is observed at a time. Existing association techniques can be grouped into two categories, including tracking-by-detection and query propagation-based paradigms. Tracking-by-detection methods [37, 34, 32] generate the per-frame instances independently by existing instance segmentation models [11, 26, 5] and track instances via tracking heads [37] or instance embeddings matching [34, 15]. In this way, the features of different frames are isolated before tracking, which results in temporal inconsistency. Query propagation-based methods [13, 10, 41] are inspired by query-based methods [3, 25] and they propagate the query across frames to decode a unique instance without heuristic matching algorithms. Despite the explicit temporal link of queries, the temporal consistency is impaired by the Local Matching and Propagating (LocPro) scheme, where they first perform local optimal matching between the predictions and ground truth at the beginning of the video, and then propagate the assignment across frames, forcing all features from subsequent frames to follow. The LocPro is not suitable for the holistic optimization across frames and results in temporal inconsistencies." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1097" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.462, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.47, + 0.321 + ], + "angle": 0, + "content": "Figure 1. Visualization of predictions from the previous online method (the online GenVIS [13]) and our TCOVIS. The previous method generates temporally inconsistent predictions, while our proposed TCOVIS achieves temporal consistency and outperforms the previous method (Best viewed in color)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.468, + 0.384 + ], + "angle": 0, + "content": "tency. Thus, achieving temporal consistency is challenging for online methods and also not comprehensively investigated by previous online VIS methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.47, + 0.884 + ], + "angle": 0, + "content": "In this paper, we propose a novel online method for video instance segmentation, named TCOVIS, to fully exploit the temporal information within a video. We take as the baseline framework an existing online VIS model (GenVIS [13]) with the query propagation-based instance association. We first introduce the global instance assignment strategy to perform global optimal matching. Different from the previous online methods [13, 41, 10], which obtain per-frame matching cost, assign labels locally on the beginning frame and propagate across frames, we collect the predictions across frames, compute the global matching cost with the video segmentation ground truth and supervise the model with the global instance assignment, encouraging features across the video to be optimized for a global optimal objective. As online methods focus on improving the representative ability of the semantic instance embeddings [38, 34], which is achieved via learning more discriminative semantic embedding in those using heuristic matching [34, 15], or via reviewing the memory of semantic embedding across frames [13] in query propagation-based methods, the spatial features are not comprehensively investigated. We further propose spatio-temporal enhancement module, leveraging the spatial information from the previous frame to enhance the temporal consistency. We perform spatial matting on the pixel embeddings to retrieve the instance-wise spatial features and adopt the cross-attention layer to aggregate the spatial and semantic features across frames, thus realizing the spatio-temporal enhancement. As shown in Figure 1, the previous online method [13] produces temporally inconsistent results, as exemplified by an object abruptly appearing in front of the cat in a mid-frame, while the proposed TCOVIS outperforms the previous one and generates temporally consistent predictions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To validate the effectiveness of the proposed method, ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "periments are conducted on four widely adopted VIS benchmarks, i.e., YouTube-VIS 2019 [37], YouTube-VIS 2021, YouTube-VIS 2022 and Occluded VIS (OVIS) [24]. Without bells-and-whistles, our proposed method achieves state-of-the-art performance on all benchmarks, outperforming other online methods, e.g., on YouTube-VIS 2021, TCOVIS achieves 49.5 AP and 61.3 AP with ResNet-50 and Swin-L backbones, respectively." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.212, + 0.858, + 0.227 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.238, + 0.892, + 0.298 + ], + "angle": 0, + "content": "- TCOVIS performs a novel global instance assignment strategy for online video instance segmentation. The model is optimized for the global optimal objective to generate more temporally consistent predictions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.308, + 0.892, + 0.383 + ], + "angle": 0, + "content": "- The further proposed spatio-temporal enhancement module captures the spatial feature and aggregates it with the semantic feature between frames, which fully utilizes the spatial information and facilitates the temporal consistency enhancement." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.394, + 0.892, + 0.469 + ], + "angle": 0, + "content": "- The proposed method achieves state-of-the-art performance on four widely used video instance segmentation benchmarks (YouTube-VIS 2019/2021/2022 and OVIS). Such achievements demonstrate the effectiveness of our proposed method." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.238, + 0.892, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.483, + 0.65, + 0.499 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Offline Video Instance Segmentation. Offline methods take as input the whole video and predict instance sequence for all frames at once. Mask propagation and box ensemble techniques are used to improve the predictions and association [1, 2, 19], but they are not end-to-end learnable due to the complex inference process. VisTR [31] extends DETR [42] from the image domain and introduces the transformer [27] to the VIS domain. EfficientVIS [35] and IFC [16] relax the heavy overhead of VisTR via an iterative query-video interaction and memory token communication, respectively. TeViT [39] contains a vision transformer [6] backbone instead of CNN and efficiently builds correspondence between the instance and query. VITA [14] models relationships among instances with the distilled condensed object tokens, without using the dense spatiotemporal backbone features. Offline methods exploit rich temporal knowledge from the whole clip and have the advantage of producing temporally consistent results, however, the offline manner is not suitable for the application in real-time scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Online Video Instance Segmentation. Instead of processing the entire video before predictions, online methods only leverage the information from the previous frames and segment the video frame-by-frame. The association paradigms of the previous online methods roughly fall into two groups: Tracking-by-detection and Query-propagation." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1098" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "Most online methods [37, 29, 21, 34] follow the tracking-by-detection paradigm. MaskTrack R-CNN [37] is the baseline method and extends the Mask R-CNN [11] with an extra tracking head for temporal association. CrossVIS [38] proposes a crossover learning scheme to utilize the current contextual information for other frames. VISOLO [9] builds on the image instance segmentation method SOLO [30] and takes advantage of the grid form previous information for memory matching and features aggregation. MinVIS [15] and IDOL [34] make use of the discriminative instance embeddings for matching between frames. With the heuristic matching technique designed for instance association across frames, temporal inconsistency comes from the frame-wise modeling before tracking." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.473, + 0.595 + ], + "angle": 0, + "content": "Object association with query propagation has been explored in the multi-object tracking (MOT) task [22, 40]. TrackFormer [22] tracks the seen objects of previous frames with a track query subset and detects the newly appeared objects in current frame with an extra object query subset. MOTR [40] extends the paired-frames training scheme to multiple frames for long-range temporal association. The query propagation-based object association is recently introduced to VIS [41, 10, 13]. ROVIS [41] follows TrackFormer [22] to detect and track instances with two subsets of queries. InsPro [10] and GenVIS [13] propagate the queries without heuristics, i.e., handcrafted rules to combine two types of queries, and achieve association across frames. However, previous query propagation-based methods also propagate the local assignment to supervise the model with the local optimal results, which leads to temporal inconsistency across the entire video. We adopt the query propagation framework but introduce the global instance assignment to enhance the temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.616, + 0.17, + 0.631 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Given a video clip with consecutive image frames, online video instance segmentation methods generate frame-level object instances upon an instance segmentation models [26, 5], utilizing the instance queries propagated from previous frames [35, 13]. We have already discussed that better performance of the online VIS method relies on more temporally consistent instance features among a video. To this end, we propose a novel end-to-end online VIS method TCOVIS (Figure 2), to improve the temporal consistency of instance features via the global instance assignment strategy and spatio-temporal enhancement module. In this section, we first introduce the online VIS pipeline in Section 3.1. Then the details of the proposed global instance assignment strategy and spatio-temporal enhancement module will be described in Section 3.2 and Section 3.3, respectively. Finally, in Section 3.4, we describe the overall loss for training the model end-to-end." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.816, + 0.108 + ], + "angle": 0, + "content": "3.1. Online Video Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.113, + 0.895, + 0.461 + ], + "angle": 0, + "content": "Following state-of-the-art VIS methods [15, 4], we adopt the advanced Masked-attention Mask Transformer (Mask2Former [5]) as the image instance segmentation network (Img. SegNet.) in this paper. Assume that the input video clip with T frames is denoted as \\( x \\in \\mathbb{R}^{T \\times 3 \\times H \\times W} \\). With each frame \\( H \\times W \\) as input, frame-wise activation map is extracted by the backbone and Transformer encoder. Then following the query-based mechanism of DETR [3], \\( N_{fq} \\) object queries of \\( C \\) dimensions are used to parse an input frame, which are called frame object queries \\( f \\in \\mathbb{R}^{C \\times N_{fq}} \\). Each object in the frame is decoded by the frame object queries from the spatial features through a multiple-level Transformer decoder, and represented as an object embedding of \\( C \\) dimensions. The object embeddings are used for classification and together with the pixel embeddings from the pixel decoder generating the mask for object instances. Class predictions are produced through a linear layer from the object embeddings. Mask embeddings are generated by an MLP linked to the object embeddings, and the model finally segments objects by pixel-wise dot product between per-pixel embeddings \\( \\mathcal{P} \\in \\mathbb{R}^{C \\times \\frac{H}{S} \\times \\frac{W}{S}} \\) and mask embeddings \\( \\mathcal{M} \\in \\mathbb{R}^{C \\times N_{fq}} \\), where \\( S \\) is the stride of the spatial feature map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.895, + 0.689 + ], + "angle": 0, + "content": "As for online video segmentation, we follow the online scheme of GenVIS [13] which propagates the instance queries from previous frames. To resolve the computational limitation, the framework adopts VITA [14] that regards the frame object queries \\( f \\) as a concise representation of objects in a frame and then feeds them into the Object Encoder \\( \\mathcal{E} \\) for intra-frame relationship. The Object Decoder \\( \\mathcal{D} \\) takes as input \\( N_v \\) video instance queries \\( q \\) and aggregates information from frame object queries. The temporal instance association is implemented through a query-based temporal propagation mechanism, where the output of \\( \\mathcal{D} \\), instance prototypes denoted as \\( p \\), are concise representations of instances [13], are not only utilized for classifying and segmenting in current frame at \\( t \\), but also serve as the instance queries for the next frame at \\( t + 1 \\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.695, + 0.892, + 0.713 + ], + "angle": 0, + "content": "\\[\nq ^ {t + 1} = p ^ {t} = \\mathcal {D} \\left(q ^ {t}, \\mathcal {E} \\left(f ^ {t}\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.765 + ], + "angle": 0, + "content": "With this propagation mechanism, the model can simply run in an online manner and associate the frame-wise outputs without heuristic matching algorithms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.84 + ], + "angle": 0, + "content": "We leave out the Instance Prototype Memory module in [13]. During the training process, we freeze the image instance segmentation model and only train the following modules, to efficiently make use of memory. More implementation details are described in Section 4.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.753, + 0.864 + ], + "angle": 0, + "content": "3.2. Global Instance Assignment" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "With the query propagation mechanism discussed above, the associated video instance queries along the temporal" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1099" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.086, + 0.874, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.396, + 0.896, + 0.468 + ], + "angle": 0, + "content": "Figure 2. (a) The overall illustration of our TCOVIS. It supervises the model with a global optimal objective during training and utilizes spatial features via the enhancement module between frames. (b) Details of global instance assignment. Different from the local matching and propagating technique, we conduct global optimal matching and assignment. (c) In the spatio-temporal enhancement module, we perform matting on pixel embeddings according to predicted masks and then aggregate the spatial and semantic features between frames to enhance temporal consistency. Numbers in colors denote the local matching loss and global matching loss (Best viewed in color)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.471, + 0.586 + ], + "angle": 0, + "content": "dimension \\(\\{q_{k}^{1:T}\\}_{1:N_{v}}\\), are used to extract the features of a unique instance, i.e. the \\(k\\)-th instance, throughout the whole input video clip. Meanwhile, the associated instance prototypes \\(\\{p_{k}^{1:T}\\}_{1:N_{v}}\\), i.e. the corresponding outputs from \\(\\mathcal{D}\\), represent this unique instance within the whole video. For brevity, we define the associated instance prototypes/features as a set of prototypes/features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.586, + 0.47, + 0.782 + ], + "angle": 0, + "content": "Previously, online VIS methods that associate frame-level results without heuristic matching algorithms [13, 41, 10] adopt a Local Matching and Propagating (LocPro) technique. This technique matches the predictions and ground truth on a frame level during training. In other words, it conducts one-to-one bipartite matching between ground truth instances and predictions of a frame (what we regard as Local Matching), and propagates the matching from previous frames to subsequent frames. Specifically, given a video clip consisting of \\( T \\) consecutive frames, it first computes the pair-wise matching cost \\( \\mathcal{L}_{match}^{t} \\) between predictions and ground truth instances on the initial frame (the first frame or the frame objects appear):" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.787, + 0.47, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a t c h}} ^ {t = 1} = \\lambda_ {c l s} \\mathcal {L} _ {c l s} ^ {t = 1} + \\lambda_ {b c e} \\mathcal {L} _ {b c e} ^ {t = 1} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} ^ {t = 1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The frame-wise \\(\\mathcal{L}_{match}^{t}\\) is composed of categorical loss and mask loss. The categorical loss adopts the cross entropy loss \\(\\mathcal{L}_{cls}^{t}\\). The mask loss consists of a binary cross entropy loss \\(\\mathcal{L}_{bce}^{t}\\) and a dice loss [23]. From the cost matrix, it follows DETR [3] and uses Hungarian algorithm [18] for optimal frame-level matching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.678 + ], + "angle": 0, + "content": "As discussed in Section 1, Local Matching only obtains optimal matching on the initial frame, and cannot achieve the global optimal matching of the whole video. As illustrated in Figure 2 (b), both dark and light green prototypes attempt to represent the person. With local matching, the ground truth will be assigned to the light one with a smaller local matching cost, however, the dark one performs better from a global perspective and is supposed to be assigned. Training the model with LocPro forces all the predictions of subsequent frames to conform to the initial frame, resulting in temporal inconsistency of the instance features among the video, because inappropriate previous matching brings accumulative error to the model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.893, + 0.789 + ], + "angle": 0, + "content": "We introduce Global Instance Assignment (GIA) strategy and expect two functionalities: (1) all frames in the video clip are considered when conducting global matching, and (2) the global assignment encourages the instance features among frames to be optimized for a global optimal objective, both of which serve the temporal consistency of instance features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "During training, different from previous online VIS methods [8, 34] that conduct frame-level local matching and provide supervision signals to each frame according to the matched pairs, we leave out the halfway matching and supervision. Consecutive input frames of a video clip pass through the VIS model in an online manner and the model generates the predictions of all frames \\(\\{\\hat{\\mathbf{y}}_k\\}_{1:N_v} =\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1100" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.473, + 0.246 + ], + "angle": 0, + "content": "\\(\\{\\hat{y}_k^t\\}_{1:N_v}^{1:T}\\), each of which consists of a category probability \\(\\hat{c}_k^t\\) and a segmentation mask probability \\(\\hat{m}_k^t\\). To conduct global assignment, we collect predictions of all frames and as well the ground truth video instance segmentation \\(\\{\\mathbf{y}_k\\}_{1:N_{gt}} = \\{y_k^{1:T}\\}_{1:N_{gt}}\\), including the category label \\(c_k\\) and its binary segmentation masks \\(\\mathbf{m}_k = m_k^{1:T}\\). Since the ground truth of an instance only has one category label, we first compute the average predicted category probability across a video clip: \\(\\bar{c}_k = \\sum_{t=1}^T \\hat{c}_k^t / T\\) and also collect the masks \\(\\hat{\\mathbf{m}}_k = \\hat{m}_k^{1:T}\\). The global matching cost is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.262, + 0.469, + 0.313 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {m a t c h}} ^ {\\text {g l o b a l}} = \\lambda_ {c l s} \\mathcal {L} _ {c l s} \\left(c _ {k}, \\bar {c} _ {\\sigma (k)}\\right) + \\lambda_ {b c e} \\mathcal {L} _ {b c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\sigma (k)}\\right) \\\\ + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\sigma (k)}\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.314, + 0.47, + 0.375 + ], + "angle": 0, + "content": "where \\(\\sigma \\in \\mathfrak{S}_{N_v}\\) is a permutation of \\(N_{v}\\) elements. One-to-one bipartite global matching between \\(\\{\\hat{\\mathbf{y}}_k\\}_{1:N_v}\\) and \\(\\{\\mathbf{y}_k\\}_{1:N_{gt}}\\) is performed to find the global optimal assignment and the objective can be formally described as:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.382, + 0.47, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\hat {\\sigma} = \\underset {\\sigma \\in \\mathfrak {S} _ {N _ {v}}} {\\arg \\max } \\sum_ {k = 1} ^ {N _ {g t}} \\mathcal {L} _ {\\text {m a t c h}} ^ {\\text {g l o b a l}} (\\mathbf {y} _ {k}, \\hat {\\mathbf {y}} _ {\\sigma (k)}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.567 + ], + "angle": 0, + "content": "Following prior work [28, 7, 42], we use Hungarian algorithm [18] to search for the global optimal assignment. In contrast to the prior methods that compute matching loss only for \\( t = 1 \\), our method considers the masks and ground truth of the whole clip, computes the matching loss globally, and conducts the global assignment. Finally, given the global optimal assignment, we use the \\( N_{gt} \\) matched video-level predictions to supervise the model with the globally matched instances across the entire video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.718 + ], + "angle": 0, + "content": "With the proposed assignment strategy, GIA, the temporal consistency of instance features across the video clip can be effectively enhanced, since we consider all frames as a whole to search for the optimal objective. Specifically, when the set of instance features across the video represents the target instance well in the first few frames, but fails to track it later, this strategy helps to find a more appropriate set to be optimized. As the global matching cost is lower, the selected features fit the target more closely and are more temporally consistent." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.35, + 0.743 + ], + "angle": 0, + "content": "3.3. Spatio-temporal Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Previous online video instance segmentation methods focus on improving the representative ability of the semantic instance embeddings [38, 34, 15, 13]. The spatial features are not comprehensively investigated to boost the temporal association for online VIS. Thus, we further introduce Spatio-temporal Enhancement module (STE), leveraging the spatial information from the previous frame to enhance the temporal consistency of the online model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Given the \\(t\\)-th frame in the video clip, with the framework described in Section 3.1, the mask embedding of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.183 + ], + "angle": 0, + "content": "the \\(k\\)-th instance is generated from the instance prototype through an MLP: \\(\\mathcal{M}_k^t = MLP(p_k^t)\\), and then the model segments the mask \\(\\hat{m}_k^t\\) by pixel-wise dot product between \\(\\mathcal{P}_k^t\\) and \\(\\mathcal{M}_k^t\\), which can be formulated as: \\(\\hat{m}_{k,i,j}^{t} = \\langle \\mathcal{P}_{k,i,j}^{t},\\mathcal{M}_{k}^{t}\\rangle\\), where \\(i\\) and \\(j\\) denote the spatial position of the pixel." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.185, + 0.893, + 0.321 + ], + "angle": 0, + "content": "As illustrated in Figure 2 (c), the spatial information of the frame is encoded in the pixel embeddings. To extract instance-wise spatial features, they can be exploited together with the predicted mask. Specifically, we perform spatial matting on pixel embeddings \\(\\mathcal{P}_k^t\\), similar to image matting, to retrieve the instance-wise pixel embedding according to \\(\\hat{m}_k^t\\). For each instance, we conduct pixel-wise multiplication between the original pixel embeddings and the binary mask to obtain the retrieved embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.334, + 0.892, + 0.352 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {k} ^ {t} = \\mathcal {P} _ {k} ^ {t} \\odot \\hat {m} _ {k} ^ {t}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.894, + 0.456 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes the element-wise multiplication. Many of the retrieved embeddings \\(\\mathcal{R}_k^t\\) are redundant since they describe the same instance, and directly mining the spatial information with them is computationally inefficient. Average pooling is adopted to obtain a concise representation of the spatial features for each instance:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.469, + 0.892, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {k} ^ {t} = \\frac {\\sum_ {i , j} \\mathcal {R} _ {k , i , j} ^ {t}}{\\sum_ {i , j} \\mathbb {1} (\\hat {m} _ {k , i , j} ^ {t} = 1)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.551 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is the indicator function. The concise spatial features \\(S_{k}^{t}\\) are sent to the next frame for temporal association." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.642 + ], + "angle": 0, + "content": "Having received from the previous frame the propagated semantic instance queries, i.e., \\( q^{t + 1} = p^t \\), and the spatial features, we follow the standard multi-head cross-attention layer (MHCA) [27] to incorporate the features from two aspects. The instance prototype \\( p^t \\) is used as the query to decode the spatial features:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.655, + 0.892, + 0.674 + ], + "angle": 0, + "content": "\\[\nq _ {k} ^ {t + 1} = \\operatorname {M H C A} \\left(p _ {k} ^ {t}, \\mathcal {S} _ {1: N _ {v}} ^ {t}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.793 + ], + "angle": 0, + "content": "where \\( q_{k}^{t + 1} \\) is the updated instance query now. Notably, the positional embedding is shared by the spatial feature and instance query with regard to the same instance, which helps the model align the instance-wise information between frames. Finally, the updated instance query is fed into the Object Decoder \\( \\mathcal{D} \\) to produce spatio-temporally enhanced features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "By performing the proposed enhancement module on the spatial features between frames, we effectively boost the spatio-temporal association of the features. In this way, the temporal consistency of the features is enhanced via the information from the spatial dimension, and the online video instance segmentation model manages to predict more temporally consistent results." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1101" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.212, + 0.106 + ], + "angle": 0, + "content": "3.4. Overall Loss" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.114, + 0.47, + 0.16 + ], + "angle": 0, + "content": "The overall loss for training with a video clip as input is a linear combination of categorical and mask losses using the one-to-one global optimal assignment \\(\\hat{\\sigma}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.173, + 0.469, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {o v e r a l l}} = \\lambda_ {c l s} \\mathcal {L} _ {c l s} \\left(c _ {k}, \\bar {c} _ {\\hat {\\sigma} (k)}\\right) + \\lambda_ {b c e} \\mathcal {L} _ {b c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\hat {\\sigma} (k)}\\right) \\\\ + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\tilde {\\sigma} (k)}\\right), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.221, + 0.469, + 0.251 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{cls}\\) is the cross entropy loss, \\(\\mathcal{L}_{bce}\\) is the binary cross entropy loss and \\(\\mathcal{L}_{dice}\\) is the dice loss [23]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.265, + 0.21, + 0.282 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.29, + 0.469, + 0.366 + ], + "angle": 0, + "content": "In this section, we evaluated the proposed TCOVIS on four benchmark datasets. Furthermore, we provided in-depth ablation studies of the effectiveness of TCOVIS. Finally, we presented several visualizations of the predictions from our model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.376, + 0.18, + 0.39 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.399, + 0.469, + 0.459 + ], + "angle": 0, + "content": "We evaluated our approach on four VIS datasets: YouTube-VIS 2019 dataset [37], YouTube-VIS 2021 dataset [37], YouTube-VIS 2022 dataset [37] and OVIS dataset [24]. We present a brief description of them:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.469, + 0.625 + ], + "angle": 0, + "content": "YouTube-VIS 2019 & 2021 & 2022: YouTube-VIS 2019 [37] is the first VIS dataset and comprises 40 predefined categories of objects. This dataset included 4,883 unique video instances with 131,000 high-quality manual annotations. We followed the widely utilized training/test set split: 2,238 videos were selected for training, 302 videos were adopted for validation and 343 videos were used for testing. Further, YouTube-VIS 2021 improved the 40-category label set and added 4883 more unique video instances. Then, YouTube-VIS 2022 contained 71 additional long evaluation videos on the top of YouTube-VIS 2021." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.469, + 0.731 + ], + "angle": 0, + "content": "OVIS: Occluded video instance segmentation (OVIS) is also a challenging VIS dataset. OVIS included 901 videos in total with 25 semantic categories. This dataset contained 5,223 unique video instances and we followed the widely utilized training/test set split: 607 videos were selected for training, 607 videos were used for validation and 154 videos were adopted for testing." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.732, + 0.469, + 0.776 + ], + "angle": 0, + "content": "Following [37], the video-level average precision (AP) and average recall (AR) were adopted as the evaluation metrics on both YouTube-VIS and OVIS." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.295, + 0.803 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We adopted the framework of GenVIS [13] which is built on VITA [14], but left out the similarity loss in VITA and the memory module in GenVIS. With the global assignment, the total loss as well the hyper-parameters were set the same as the video-level loss in VITA, which is a temporally extended loss function [16]. The model was trained" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "with pseudo-videos from COCO images [20] as data augmentation, and with a batch size of 8 video clips of 6 frames. As we froze the backbone and image segmentation model, all experiments were conducted with 8 RTX 2080 Ti GPUs. The method was implemented on detectron2 [36]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.178, + 0.64, + 0.192 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.201, + 0.892, + 0.245 + ], + "angle": 0, + "content": "Following the standard evaluation metrics [37], we compared TCOVIS with state-of-the-art approaches on four VIS benchmarks: YouTube-VIS 2019/2021/2022 and OVIS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.892, + 0.382 + ], + "angle": 0, + "content": "YouTube-VIS 2019&2021. From Table 1, we can observe that TCOVIS has achieved very competitive performance using both lightweight backbones (ResNet-50) and powerful ones (Swin-L). Moreover, our method can even show better performance than offline methods, such as VITA [14] and offline version GenVIS [13]. On the more difficult YouTube-VIS 2021 dataset, TCOVIS surpasses GenVIS [13] in AP not only with ResNet-50 by 2.4 but also with Swin-L by 1.7." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.488 + ], + "angle": 0, + "content": "YouTube-VIS 2022. As shown in Table 2, TCOVIS performed best in both AP and AR on YouTube-VIS 2022 dataset, which is more challenging than 2019&2021 datasets. Especially with powerful backbone Swin-L, TCOVIS outperformed GenVIS [13] in AP with a huge margin of 4.9, which shows the effectiveness of our method for the complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.595 + ], + "angle": 0, + "content": "OVIS. Table 3 presents the comparisons on OVIS dataset, and we can find that TCOVIS also achieved the best 46.7 AP and 19.1 AR with Swin-L backbone. The results demonstrate that TCOVIS can deal with complicated situations where objects are heavily occluded in others. With ResNet-50 backbone, the performance is still the second best compared to the previous state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.655, + 0.621 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.689 + ], + "angle": 0, + "content": "In this section, we provided ablation studies and discuss the effects of different settings in the proposed method. The experiments are conducted with a ResNet-50 [12] backbone on YouTube-VIS 2019/2021 [37] valid set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of the proposed assignment strategy and enhancement module. The ablation studies on the global instance assignment strategy and the spatio-temporal enhancement module are shown in Table 4. As for the assignment strategy, compared to LocPro as the baseline, the model with the proposed global assignment outperformed the baseline model by more than 0.8 on AP, \\(\\mathrm{AP}_{50}\\) and \\(\\mathrm{AP}_{75}\\) on YouTube-VIS 2019, and more than 1.2 on YouTube-VIS 2021. The consistently significant improvement indicates that the strategy with global optimal matching contributes to better overall segmentation, while LocPro only considers the local optimal results. Besides, the baseline forces the posterior features to conform to those at the very beginning leading to temporal inconsistency, since the accumulative" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1102" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.096, + 0.089, + 0.868, + 0.457 + ], + "angle": 0, + "content": "
MethodTypeYouTube-VIS 2019YouTube-VIS 2021
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AR_1\\)\\(AR_{10}\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AR_1\\)\\(AR_{10}\\)
ResNet-50EfficientVIS [35]Offline37.959.743.040.346.634.057.537.333.842.5
IFC [16]Offline41.265.144.642.349.635.255.937.732.642.9
Mask2Former-VIS [4]Offline46.468.050.0--40.660.941.8--
\\(TeViT^{\\dagger}\\)[39]Offline46.671.351.644.954.337.961.242.135.144.6
SeqFormer [33]Offline47.469.851.845.554.840.562.443.736.148.1
VITA [14]Offline49.872.654.549.461.045.767.449.540.953.6
\\(GenVIS_{semi-online}\\)[13]Offline51.372.057.849.560.046.367.050.240.653.2
Cross VIS [38]Online36.356.838.935.640.734.254.437.930.438.2
VISOLO [9]Online38.656.343.735.742.536.954.740.230.640.9
MinVIS [15]Online47.469.052.145.755.744.266.048.139.251.7
IDOL [34]Online49.574.052.947.758.743.968.049.638.050.9
\\(GenVIS_{online}\\)[13]Online50.071.554.649.559.747.167.551.541.654.7
TCOVISOnline52.373.557.649.860.249.571.253.841.355.9
Swin-LSeqFormer [33]Offline59.382.166.451.764.451.874.658.242.858.1
Mask2Former-VIS [4]Offline60.484.467.0--52.676.457.2--
VITA [14]Offline63.086.967.956.368.157.580.661.047.762.6
\\(GenVIS_{semi-online}\\)[13]Offline63.885.768.556.368.460.180.966.549.164.7
MinVIS [15]Online61.683.368.654.866.655.376.662.045.960.8
IDOL [34]Online64.387.571.055.669.156.180.863.545.060.1
\\(GenVIS_{online}\\)[13]Online64.084.968.356.169.459.680.965.848.765.0
TCOVISOnline64.186.669.555.869.061.382.968.048.665.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.463, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Table 1. Quantitative results on YouTube-VIS 2019 and 2021 validation sets. The results are respectively grouped by method types (Offline or Online) and backbone networks (ResNet-50 and Swin-L). We bold the best performance and underline the second. † denotes using MsgBoxIF [39] backbone which has a similar weight scale with ResNet-50." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.519, + 0.466, + 0.677 + ], + "angle": 0, + "content": "
MethodTypeAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
ResNet-50VITA [14]Offline32.653.939.330.342.6
GenVIS [13]Offline37.258.542.933.240.4
MinVIS [15]Online23.347.919.320.228.0
GenVIS [13]Online37.561.641.532.642.2
TCOVISOnline38.659.441.632.846.7
Swin-LVITA* [14]Offline41.163.044.039.344.3
GenVIS [13]Offline44.369.944.939.948.4
MinVIS* [15]Online33.154.833.729.536.6
GenVIS [13]Online45.169.147.339.848.5
TCOVISOnline51.073.053.541.756.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.684, + 0.468, + 0.725 + ], + "angle": 0, + "content": "Table 2. Quantitative results on YouTube-VIS 2022 validation dataset. We bold the highest accuracy and underline the second. \\(\\star\\): Reproduced by [13]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.746, + 0.469, + 0.806 + ], + "angle": 0, + "content": "error impairs the model during the training process. Our proposed assignment strategy encourages the features of all time to fit the global optimal objective, which effectively enhances the temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "As for the effectiveness of the spatio-temporal enhancement module, the comparison is also shown in Table 4. The results show that the enhancement module brings improvements of 0.9 AP on YouTube-VIS 2019&2021, compared to using the proposed assignment strategy individually. In particular, the proposed module improved the performance" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.519, + 0.889, + 0.738 + ], + "angle": 0, + "content": "
MethodTypeAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
ResNet-50\\( TeViT^†[39] \\)Offline17.434.915.011.221.8
VITA [14]Offline19.641.217.411.726.0
GenVIS [13]Offline34.559.435.016.638.3
CrossVIS [38]Online14.932.712.110.319.8
VISOLO [9]Online15.331.013.811.121.7
MinVIS [15]Online25.045.524.013.929.7
IDOL [34]Online30.251.330.015.037.5
GenVIS [13]Online35.860.836.216.339.6
TCOVISOnline35.360.736.615.739.5
Swin-LVITA [14]Offline27.751.924.914.933.0
GenVIS [13]Offline45.469.247.818.949.0
MinVIS [15]Online39.461.541.318.143.3
IDOL [34]Online42.665.745.217.949.6
GenVIS [13]Online45.269.148.419.148.6
TCOVISOnline46.770.949.519.150.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.746, + 0.892, + 0.788 + ], + "angle": 0, + "content": "Table 3. Quantitative results on OVIS validation set. We bold the highest accuracy and underline the second. † denotes using MsgBoxT [39] backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "by 1.2 and 1.7 in \\(\\mathrm{AP}_{50}\\) on two datasets, respectively. The performance improvements demonstrate that the proposed module effectively captures the spatial information from the previous frame and aggregates the semantic and spatial features across time. As a result, the delicate spatio-temporal design further enhances the temporal consistency of fea" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1103" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.135, + 0.089, + 0.832, + 0.171 + ], + "angle": 0, + "content": "
LocProGIASTEYouTube-VIS 2019YouTube-VIS 2021
AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
✓--50.671.556.147.658.247.468.152.139.853.1
-✓-51.472.357.049.559.948.669.553.640.655.0
-✓✓52.373.557.649.860.249.571.253.841.355.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.177, + 0.894, + 0.207 + ], + "angle": 0, + "content": "Table 4. Ablation study of the method for Global Instance Assignment strategy (GIA) and the Spatio-temporal Enhancement module (STE) on the YouTube-VIS 2019 / 2021 validation sets. LocPro denotes the Local Matching and Propagating technique." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.215, + 0.455, + 0.329 + ], + "angle": 0, + "content": "
ArchitectureAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Cross-Attn.49.571.253.841.355.9
Concat.47.968.652.741.455.1
Self-Cross48.270.051.841.054.1
Dec. Mod.48.569.753.940.654.7
Resp. Pos.48.870.352.541.555.5
Ins. Attn.49.270.653.540.955.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.334, + 0.47, + 0.363 + ], + "angle": 0, + "content": "Table 5. Ablation study of the module manipulations of the Spatio-temporal Enhancement module on YouTube-VIS 2021." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.444, + 0.388 + ], + "angle": 0, + "content": "tures for the online video instance segmentation model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.51 + ], + "angle": 0, + "content": "Incorporating the assignment strategy and the enhancement module, our proposed method gained remarkable performance improvements of 1.6 AP on YouTube-VIS 2019 and 2.1 AP on YouTube-VIS 2021 over the LocPro baseline. In a nutshell, the experimental results indicate that the temporally consistent features learned by our proposed TCOVIS significantly boost the performances on online video instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.616 + ], + "angle": 0, + "content": "Different manipulations of spatio-temporal enhancement. In Table 5, we compared our proposed spatiotemporal enhancement module with other optional manipulations. Cross-Attn. denotes our proposed manipulation following the standard multi-head cross-attention layer to decode the spatial features with shared positional embedding described in Section 3.3. Concat. indicates that we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.895, + 0.355 + ], + "angle": 0, + "content": "concatenated the corresponding spatial feature and prototype of an instance followed by an MLP to get the updated query. Self-Cross stands for adding an extra self-attention layer ahead of the cross-attention layer. Dec. Mod. is decoder modulation, reviewing the spatial feature for every Transformer decoder layer. Resp .Pos. denotes respective positional embeddings were used for the spatial feature and prototype when we performed cross-attention. Ins. Attn. indicates the instance-wise decoding in cross-attention layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.36, + 0.895, + 0.617 + ], + "angle": 0, + "content": "As shown, compared to all its counterparts, our proposed manipulation achieved the best performance. The first three variants are related to the feature aggregation, where \\( \\text{Concat} \\) is too naive to model the spatial and semantic features, while Self-Cross obscures the spatial feature. We inferred the performance decrease of Dec. Mod. comes from the information redundancy when aggregating them in every layer. The last two experiments studied the instance-wise correspondence. In the Resp. Pos. setting, explicit correspondence for spatial and semantic features of the same instance between frames is absent. Ins. Attn. only focuses on the instance itself across time and can slightly enhance the temporal consistency, however, neglecting the spatial features of other instances, the module fails to capture the spatial relationship among instances. The results confirm that the proposed manipulation effectively exploits the spatial information along the temporal dimension to enhance the" + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.637, + 0.495, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.637, + 0.869, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.857, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of TCOVIS, compared with VITA [14] and GenVIS [13]. On the left are the predictions on YouTube-VIS 2019 [37] and on the right are on OVIS [24]. Objects displayed in the same color denote the same instance. Our TCOVIS shows more temporally consistent results in these challenging scenes, where there are occlusions of the instance itself or others (Best viewed in color)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1104" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.126, + 0.122, + 0.14 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.09, + 0.309, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.091, + 0.495, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.091, + 0.68, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.091, + 0.866, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.206, + 0.122, + 0.222 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.173, + 0.308, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.173, + 0.494, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.173, + 0.68, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.173, + 0.865, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.289, + 0.122, + 0.304 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.255, + 0.309, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.255, + 0.494, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.255, + 0.68, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.255, + 0.865, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.37, + 0.122, + 0.384 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.336, + 0.309, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.336, + 0.494, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.336, + 0.68, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.336, + 0.866, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.425, + 0.895, + 0.467 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of TCOVIS in four challenging cases on YouTube-VIS 2019 [37] and OVIS [24]: (a) fast movement, (b) heavy occlusion, (c) low resolution, and (d) crowded scene. Objects displayed in the same color denote the same instance. The qualitative results demonstrate the effectiveness and robustness of TCOVIS (Best viewed in color)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.484, + 0.296, + 0.499 + ], + "angle": 0, + "content": "temporal consistency of features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.512, + 0.254, + 0.528 + ], + "angle": 0, + "content": "4.5. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.734 + ], + "angle": 0, + "content": "In Figure 3, we show the qualitative comparisons of the proposed TCOVIS with VITA [14] and GenVIS [13] on YouTube-VIS 2019 and OVIS datasets. In the left scene where there is a zebra with self-occlusion, GenVIS fails to segment its head resulting in fragmented predictions, however, TCOVIS performs temporally consistent segmentation with impressive accuracy. On the right is a difficult case where there are two dogs with similar appearances grappling with each other. VITA incorrectly detects more than two dogs and fails to track the left one, while GenVIS fails to handle the margin of two instances, e.g. the nose and the paw of the left dog. TCOVIS successfully tracks and segments the instances, demonstrating its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In Figure 4, we provide more qualitative results of the proposed method in variously challenging cases, which are all chosen from the mentioned benchmarks [37, 24]. As shown in Figure 4 (a), our method successfully tracks the surfer and the surfboard with fast movement. In the second row, we present a difficult case with severe occlusion, in which our method performs admirably by accurately segmenting the claw (at the bottom) of a parrot despite the presence of a wooden stick that partially obstructs its lower body. Figure 4 (c) illustrates a low-resolution case and our method still achieves good performance. In Figure 4 (d)," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.484, + 0.892, + 0.56 + ], + "angle": 0, + "content": "we depict a crowded scene where TCOVIS is capable of handling multiple instances that share similar appearances and exhibit complex interactions. All the qualitative results in the challenging situations demonstrate the effectiveness and robustness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.619, + 0.597 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.79 + ], + "angle": 0, + "content": "In this paper, we propose a new online video instance segmentation method, TCOVIS, to fully exploit the temporal information within a video and produce temporally consistent predictions. Based on the query propagation framework, we propose a global instance assignment strategy to perform global optimal matching with the consideration of the entire video and supervise the model with the global optimal objective. We further devise a spatio-temporal enhancement module to capture the spatial feature and aggregate it with the semantic feature between frames. The effectiveness of our method is evaluated with experimental results on YouTube-VIS 2019/2021/2022 and OVIS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.813, + 0.66, + 0.83 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1105" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, pages 158-177, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.47, + 0.214 + ], + "angle": 0, + "content": "[2] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, pages 9739-9748, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.47, + 0.271 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.47, + 0.386 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.47 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[7] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In ICCV, pages 6910-6919, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.47, + 0.571 + ], + "angle": 0, + "content": "[8] Yang Fu, Linjie Yang, Ding Liu, Thomas S Huang, and Humphrey Shi. Compfeat: Comprehensive feature aggregation for video instance segmentation. In AAAI, pages 1361-1369, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[9] Su Ho Han, Sukjun Hwang, Seoung Wug Oh, Yeonschool Park, Hyunwoo Kim, Min-Jung Kim, and Seon Joo Kim. Visolo: Grid-based space-time aggregation for efficient online video instance segmentation. In CVPR, pages 2896-2905, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[10] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. In NeurIPS, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[13] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In CVPR, pages 14623–14632, 2023. 1, 2, 3, 4, 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[14] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. VITA: Video instance segmentation via object token association. In NeurIPS, 2022. 1, 2, 3, 6, 7, 8, 9" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] De-An Huang, Zhiding Yu, and Anima Anandkumar. MinVIS: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 34, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[17] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks for multiple object tracking and segmentation. NeurIPS, 34, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[18] Harold W Kuhn. The hungarian method for the assignment problem. NRL, 1955. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[19] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In ICCV, pages 1739-1748, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[21] Dongfang Liu, Yiming Cui, Wenbo Tan, and Yingjie Chen. Sg-net: Spatial granularity network for one-stage video instance segmentation. In CVPR, pages 9816-9825, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[22] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[23] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571, 2016. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[24] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. IJCV, 130(8):2022-2039, 2022. 2, 6, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[25] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 14454-14463, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[26] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, pages 282-298, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[28] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[29] Tao Wang, Ning Xu, Kean Chen, and Weiyao Lin. End-to-end video instance segmentation via spatial-temporal graph neural networks. In ICCV, pages 10797-10806, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, pages 649-665, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1106" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, pages 8741-8750, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.203 + ], + "angle": 0, + "content": "[32] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In CVPR, pages 12352-12361, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.206, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[33] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, pages 553-569, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.468, + 0.302 + ], + "angle": 0, + "content": "[34] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, pages 588-605, 2022. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.36 + ], + "angle": 0, + "content": "[35] Jialian Wu, Sudhir Yarram, Hui Liang, Tian Lan, Junsong Yuan, Jayan Eledath, and Gerard Medioni. Efficient video instance segmentation via tracklet query and proposal. In CVPR, pages 959-968, 2022. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[36] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.468, + 0.444 + ], + "angle": 0, + "content": "[37] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, pages 5188-5197, 2019. 1, 2, 3, 6, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.447, + 0.468, + 0.502 + ], + "angle": 0, + "content": "[38] Shusheng Yang, Yuxin Fang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Crossover learning for fast online video instance segmentation. In ICCV, pages 8043-8052, 2021. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.504, + 0.468, + 0.559 + ], + "angle": 0, + "content": "[39] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Wenyu Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In CVPR, pages 2885-2895, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.561, + 0.468, + 0.616 + ], + "angle": 0, + "content": "[40] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, pages 659-675. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.618, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[41] Zitong Zhan, Daniel McKee, and Svetlana Lazebnik. Robust online video instance segmentation with track queries. arXiv preprint arXiv:2211.09108, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.66, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[42] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable \\(\\{\\mathrm{detr}\\}\\): Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1107" + } + ] +] \ No newline at end of file diff --git a/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_origin.pdf b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4d501c3de7947d1acda3a7c02266c6a8f5122db9 --- /dev/null +++ b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c95fae50f84291a68f586cb116c492e668b1ae618e64d033034eefc2b92c190f +size 7311499 diff --git a/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/full.md b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0a79d728d609bcf16eeb16a183dfd5ba07db0334 --- /dev/null +++ b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/full.md @@ -0,0 +1,326 @@ +# TCOVIS: Temporally Consistent Online Video Instance Segmentation + +Junlong Li Bingyao Yu Yongming Rao Jie Zhou Jiwen Lu* + +Department of Automation, Tsinghua University, China + +Beijing National Research Center for Information Science and Technology, China + +# Abstract + +In recent years, significant progress has been made in video instance segmentation (VIS), with many offline and online methods achieving state-of-the-art performance. While offline methods have the advantage of producing temporally consistent predictions, they are not suitable for real-time scenarios. Conversely, online methods are more practical, but maintaining temporal consistency remains a challenging task. In this paper, we propose a novel online method for video instance segmentation, called TCOVIS, which fully exploits the temporal information in a video clip. The core of our method consists of a global instance assignment strategy and a spatio-temporal enhancement module, which improve the temporal consistency of the features from two aspects. Specifically, we perform global optimal matching between the predictions and ground truth across the whole video clip, and supervise the model with the global optimal objective. We also capture the spatial feature and aggregate it with the semantic feature between frames, thus realizing the spatio-temporal enhancement. We evaluate our method on four widely adopted VIS benchmarks, namely YouTube-VIS 2019/2021/2022 and OVIS, and achieve state-of-the-art performance on all benchmarks without bells-and-whistles. For instance, on YouTube-VIS 2021, TCOVIS achieves 49.5 AP and 61.3 AP with ResNet-50 and Swin-L backbones, respectively. Code is available at https://github.com/jun-long-li/TCOVIS. + +# 1. Introduction + +Video instance segmentation (VIS) is a challenging and representative video understanding task recently introduced in [37]. It aims at detecting, segmenting and tracking instances across a video. VIS is attracting increasing attention for various real-world applications such as video editing, video surveillance, augmented reality and autonomous driving. Recently introduced VIS methods can be roughly + +categorized into two groups: offline methods and online methods. Offline methods [2, 16, 19, 31, 33, 35] take as input the whole video and perform the segmentation of instance sequence for the whole video at once. Online methods [8, 34, 17, 10, 38], on the contrary, take as input a video frame by frame and generate the pre-frame object instances while associating the frame-wise results across frames. Both offline and online methods have achieved impressing performance on the VIS task. + +Offline methods have an inherent advantage in producing temporally consistent predictions, since delicate temporal communication and association mechanisms can be adopted throughout the video [39, 33, 14] to handle the overall temporal information and impose an explicit constraint on the temporal consistency. However, the video-in and video-out offline manner is not suitable for real-time scenarios. Conversely, online methods are more practical and making considerable progress but suffer from temporal inconsistency (as shown in Figure 1), remaining a great challenge. + +Online methods rely on specific instance association applied across frames, since only one frame is observed at a time. Existing association techniques can be grouped into two categories, including tracking-by-detection and query propagation-based paradigms. Tracking-by-detection methods [37, 34, 32] generate the per-frame instances independently by existing instance segmentation models [11, 26, 5] and track instances via tracking heads [37] or instance embeddings matching [34, 15]. In this way, the features of different frames are isolated before tracking, which results in temporal inconsistency. Query propagation-based methods [13, 10, 41] are inspired by query-based methods [3, 25] and they propagate the query across frames to decode a unique instance without heuristic matching algorithms. Despite the explicit temporal link of queries, the temporal consistency is impaired by the Local Matching and Propagating (LocPro) scheme, where they first perform local optimal matching between the predictions and ground truth at the beginning of the video, and then propagate the assignment across frames, forcing all features from subsequent frames to follow. The LocPro is not suitable for the holistic optimization across frames and results in temporal inconsistencies. + +![](images/614e5bdb04624b7cb1a444c6913b3afaf4d077bd360c7bcc669ca7753656b233.jpg) +Figure 1. Visualization of predictions from the previous online method (the online GenVIS [13]) and our TCOVIS. The previous method generates temporally inconsistent predictions, while our proposed TCOVIS achieves temporal consistency and outperforms the previous method (Best viewed in color). + +tency. Thus, achieving temporal consistency is challenging for online methods and also not comprehensively investigated by previous online VIS methods. + +In this paper, we propose a novel online method for video instance segmentation, named TCOVIS, to fully exploit the temporal information within a video. We take as the baseline framework an existing online VIS model (GenVIS [13]) with the query propagation-based instance association. We first introduce the global instance assignment strategy to perform global optimal matching. Different from the previous online methods [13, 41, 10], which obtain per-frame matching cost, assign labels locally on the beginning frame and propagate across frames, we collect the predictions across frames, compute the global matching cost with the video segmentation ground truth and supervise the model with the global instance assignment, encouraging features across the video to be optimized for a global optimal objective. As online methods focus on improving the representative ability of the semantic instance embeddings [38, 34], which is achieved via learning more discriminative semantic embedding in those using heuristic matching [34, 15], or via reviewing the memory of semantic embedding across frames [13] in query propagation-based methods, the spatial features are not comprehensively investigated. We further propose spatio-temporal enhancement module, leveraging the spatial information from the previous frame to enhance the temporal consistency. We perform spatial matting on the pixel embeddings to retrieve the instance-wise spatial features and adopt the cross-attention layer to aggregate the spatial and semantic features across frames, thus realizing the spatio-temporal enhancement. As shown in Figure 1, the previous online method [13] produces temporally inconsistent results, as exemplified by an object abruptly appearing in front of the cat in a mid-frame, while the proposed TCOVIS outperforms the previous one and generates temporally consistent predictions. + +To validate the effectiveness of the proposed method, ex + +periments are conducted on four widely adopted VIS benchmarks, i.e., YouTube-VIS 2019 [37], YouTube-VIS 2021, YouTube-VIS 2022 and Occluded VIS (OVIS) [24]. Without bells-and-whistles, our proposed method achieves state-of-the-art performance on all benchmarks, outperforming other online methods, e.g., on YouTube-VIS 2021, TCOVIS achieves 49.5 AP and 61.3 AP with ResNet-50 and Swin-L backbones, respectively. + +Our main contributions are summarized as follows: + +- TCOVIS performs a novel global instance assignment strategy for online video instance segmentation. The model is optimized for the global optimal objective to generate more temporally consistent predictions. +- The further proposed spatio-temporal enhancement module captures the spatial feature and aggregates it with the semantic feature between frames, which fully utilizes the spatial information and facilitates the temporal consistency enhancement. +- The proposed method achieves state-of-the-art performance on four widely used video instance segmentation benchmarks (YouTube-VIS 2019/2021/2022 and OVIS). Such achievements demonstrate the effectiveness of our proposed method. + +# 2. Related Works + +Offline Video Instance Segmentation. Offline methods take as input the whole video and predict instance sequence for all frames at once. Mask propagation and box ensemble techniques are used to improve the predictions and association [1, 2, 19], but they are not end-to-end learnable due to the complex inference process. VisTR [31] extends DETR [42] from the image domain and introduces the transformer [27] to the VIS domain. EfficientVIS [35] and IFC [16] relax the heavy overhead of VisTR via an iterative query-video interaction and memory token communication, respectively. TeViT [39] contains a vision transformer [6] backbone instead of CNN and efficiently builds correspondence between the instance and query. VITA [14] models relationships among instances with the distilled condensed object tokens, without using the dense spatiotemporal backbone features. Offline methods exploit rich temporal knowledge from the whole clip and have the advantage of producing temporally consistent results, however, the offline manner is not suitable for the application in real-time scenarios. + +Online Video Instance Segmentation. Instead of processing the entire video before predictions, online methods only leverage the information from the previous frames and segment the video frame-by-frame. The association paradigms of the previous online methods roughly fall into two groups: Tracking-by-detection and Query-propagation. + +Most online methods [37, 29, 21, 34] follow the tracking-by-detection paradigm. MaskTrack R-CNN [37] is the baseline method and extends the Mask R-CNN [11] with an extra tracking head for temporal association. CrossVIS [38] proposes a crossover learning scheme to utilize the current contextual information for other frames. VISOLO [9] builds on the image instance segmentation method SOLO [30] and takes advantage of the grid form previous information for memory matching and features aggregation. MinVIS [15] and IDOL [34] make use of the discriminative instance embeddings for matching between frames. With the heuristic matching technique designed for instance association across frames, temporal inconsistency comes from the frame-wise modeling before tracking. + +Object association with query propagation has been explored in the multi-object tracking (MOT) task [22, 40]. TrackFormer [22] tracks the seen objects of previous frames with a track query subset and detects the newly appeared objects in current frame with an extra object query subset. MOTR [40] extends the paired-frames training scheme to multiple frames for long-range temporal association. The query propagation-based object association is recently introduced to VIS [41, 10, 13]. ROVIS [41] follows TrackFormer [22] to detect and track instances with two subsets of queries. InsPro [10] and GenVIS [13] propagate the queries without heuristics, i.e., handcrafted rules to combine two types of queries, and achieve association across frames. However, previous query propagation-based methods also propagate the local assignment to supervise the model with the local optimal results, which leads to temporal inconsistency across the entire video. We adopt the query propagation framework but introduce the global instance assignment to enhance the temporal consistency. + +# 3. Method + +Given a video clip with consecutive image frames, online video instance segmentation methods generate frame-level object instances upon an instance segmentation models [26, 5], utilizing the instance queries propagated from previous frames [35, 13]. We have already discussed that better performance of the online VIS method relies on more temporally consistent instance features among a video. To this end, we propose a novel end-to-end online VIS method TCOVIS (Figure 2), to improve the temporal consistency of instance features via the global instance assignment strategy and spatio-temporal enhancement module. In this section, we first introduce the online VIS pipeline in Section 3.1. Then the details of the proposed global instance assignment strategy and spatio-temporal enhancement module will be described in Section 3.2 and Section 3.3, respectively. Finally, in Section 3.4, we describe the overall loss for training the model end-to-end. + +# 3.1. Online Video Instance Segmentation + +Following state-of-the-art VIS methods [15, 4], we adopt the advanced Masked-attention Mask Transformer (Mask2Former [5]) as the image instance segmentation network (Img. SegNet.) in this paper. Assume that the input video clip with T frames is denoted as $x \in \mathbb{R}^{T \times 3 \times H \times W}$ . With each frame $H \times W$ as input, frame-wise activation map is extracted by the backbone and Transformer encoder. Then following the query-based mechanism of DETR [3], $N_{fq}$ object queries of $C$ dimensions are used to parse an input frame, which are called frame object queries $f \in \mathbb{R}^{C \times N_{fq}}$ . Each object in the frame is decoded by the frame object queries from the spatial features through a multiple-level Transformer decoder, and represented as an object embedding of $C$ dimensions. The object embeddings are used for classification and together with the pixel embeddings from the pixel decoder generating the mask for object instances. Class predictions are produced through a linear layer from the object embeddings. Mask embeddings are generated by an MLP linked to the object embeddings, and the model finally segments objects by pixel-wise dot product between per-pixel embeddings $\mathcal{P} \in \mathbb{R}^{C \times \frac{H}{S} \times \frac{W}{S}}$ and mask embeddings $\mathcal{M} \in \mathbb{R}^{C \times N_{fq}}$ , where $S$ is the stride of the spatial feature map. + +As for online video segmentation, we follow the online scheme of GenVIS [13] which propagates the instance queries from previous frames. To resolve the computational limitation, the framework adopts VITA [14] that regards the frame object queries $f$ as a concise representation of objects in a frame and then feeds them into the Object Encoder $\mathcal{E}$ for intra-frame relationship. The Object Decoder $\mathcal{D}$ takes as input $N_v$ video instance queries $q$ and aggregates information from frame object queries. The temporal instance association is implemented through a query-based temporal propagation mechanism, where the output of $\mathcal{D}$ , instance prototypes denoted as $p$ , are concise representations of instances [13], are not only utilized for classifying and segmenting in current frame at $t$ , but also serve as the instance queries for the next frame at $t + 1$ , i.e., + +$$ +q ^ {t + 1} = p ^ {t} = \mathcal {D} \left(q ^ {t}, \mathcal {E} \left(f ^ {t}\right)\right). \tag {1} +$$ + +With this propagation mechanism, the model can simply run in an online manner and associate the frame-wise outputs without heuristic matching algorithms. + +We leave out the Instance Prototype Memory module in [13]. During the training process, we freeze the image instance segmentation model and only train the following modules, to efficiently make use of memory. More implementation details are described in Section 4.2. + +# 3.2. Global Instance Assignment + +With the query propagation mechanism discussed above, the associated video instance queries along the temporal + +![](images/0b62bb26eae29dcf53ab8d48dd7ef1e7eee710a0cabdaf22d0115d12b1375694.jpg) +Figure 2. (a) The overall illustration of our TCOVIS. It supervises the model with a global optimal objective during training and utilizes spatial features via the enhancement module between frames. (b) Details of global instance assignment. Different from the local matching and propagating technique, we conduct global optimal matching and assignment. (c) In the spatio-temporal enhancement module, we perform matting on pixel embeddings according to predicted masks and then aggregate the spatial and semantic features between frames to enhance temporal consistency. Numbers in colors denote the local matching loss and global matching loss (Best viewed in color). + +dimension $\{q_{k}^{1:T}\}_{1:N_{v}}$ , are used to extract the features of a unique instance, i.e. the $k$ -th instance, throughout the whole input video clip. Meanwhile, the associated instance prototypes $\{p_{k}^{1:T}\}_{1:N_{v}}$ , i.e. the corresponding outputs from $\mathcal{D}$ , represent this unique instance within the whole video. For brevity, we define the associated instance prototypes/features as a set of prototypes/features. + +Previously, online VIS methods that associate frame-level results without heuristic matching algorithms [13, 41, 10] adopt a Local Matching and Propagating (LocPro) technique. This technique matches the predictions and ground truth on a frame level during training. In other words, it conducts one-to-one bipartite matching between ground truth instances and predictions of a frame (what we regard as Local Matching), and propagates the matching from previous frames to subsequent frames. Specifically, given a video clip consisting of $T$ consecutive frames, it first computes the pair-wise matching cost $\mathcal{L}_{match}^{t}$ between predictions and ground truth instances on the initial frame (the first frame or the frame objects appear): + +$$ +\mathcal {L} _ {\text {m a t c h}} ^ {t = 1} = \lambda_ {c l s} \mathcal {L} _ {c l s} ^ {t = 1} + \lambda_ {b c e} \mathcal {L} _ {b c e} ^ {t = 1} + \lambda_ {d i c e} \mathcal {L} _ {d i c e} ^ {t = 1}. \tag {2} +$$ + +The frame-wise $\mathcal{L}_{match}^{t}$ is composed of categorical loss and mask loss. The categorical loss adopts the cross entropy loss $\mathcal{L}_{cls}^{t}$ . The mask loss consists of a binary cross entropy loss $\mathcal{L}_{bce}^{t}$ and a dice loss [23]. From the cost matrix, it follows DETR [3] and uses Hungarian algorithm [18] for optimal frame-level matching. + +As discussed in Section 1, Local Matching only obtains optimal matching on the initial frame, and cannot achieve the global optimal matching of the whole video. As illustrated in Figure 2 (b), both dark and light green prototypes attempt to represent the person. With local matching, the ground truth will be assigned to the light one with a smaller local matching cost, however, the dark one performs better from a global perspective and is supposed to be assigned. Training the model with LocPro forces all the predictions of subsequent frames to conform to the initial frame, resulting in temporal inconsistency of the instance features among the video, because inappropriate previous matching brings accumulative error to the model. + +We introduce Global Instance Assignment (GIA) strategy and expect two functionalities: (1) all frames in the video clip are considered when conducting global matching, and (2) the global assignment encourages the instance features among frames to be optimized for a global optimal objective, both of which serve the temporal consistency of instance features. + +During training, different from previous online VIS methods [8, 34] that conduct frame-level local matching and provide supervision signals to each frame according to the matched pairs, we leave out the halfway matching and supervision. Consecutive input frames of a video clip pass through the VIS model in an online manner and the model generates the predictions of all frames $\{\hat{\mathbf{y}}_k\}_{1:N_v} =$ + +$\{\hat{y}_k^t\}_{1:N_v}^{1:T}$ , each of which consists of a category probability $\hat{c}_k^t$ and a segmentation mask probability $\hat{m}_k^t$ . To conduct global assignment, we collect predictions of all frames and as well the ground truth video instance segmentation $\{\mathbf{y}_k\}_{1:N_{gt}} = \{y_k^{1:T}\}_{1:N_{gt}}$ , including the category label $c_k$ and its binary segmentation masks $\mathbf{m}_k = m_k^{1:T}$ . Since the ground truth of an instance only has one category label, we first compute the average predicted category probability across a video clip: $\bar{c}_k = \sum_{t=1}^T \hat{c}_k^t / T$ and also collect the masks $\hat{\mathbf{m}}_k = \hat{m}_k^{1:T}$ . The global matching cost is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {m a t c h}} ^ {\text {g l o b a l}} = \lambda_ {c l s} \mathcal {L} _ {c l s} \left(c _ {k}, \bar {c} _ {\sigma (k)}\right) + \lambda_ {b c e} \mathcal {L} _ {b c e} \left(\mathbf {m} _ {\mathbf {k}}, \hat {\mathbf {m}} _ {\sigma (k)}\right) \\ + \lambda_ {d i c e} \mathcal {L} _ {d i c e} \left(\mathbf {m} _ {\mathbf {k}}, \hat {\mathbf {m}} _ {\sigma (k)}\right), \tag {3} \\ \end{array} +$$ + +where $\sigma \in \mathfrak{S}_{N_v}$ is a permutation of $N_{v}$ elements. One-to-one bipartite global matching between $\{\hat{\mathbf{y}}_k\}_{1:N_v}$ and $\{\mathbf{y}_k\}_{1:N_{gt}}$ is performed to find the global optimal assignment and the objective can be formally described as: + +$$ +\hat {\sigma} = \underset {\sigma \in \mathfrak {S} _ {N _ {v}}} {\arg \max } \sum_ {k = 1} ^ {N _ {g t}} \mathcal {L} _ {\text {m a t c h}} ^ {\text {g l o b a l}} (\mathbf {y} _ {k}, \hat {\mathbf {y}} _ {\sigma (k)}). \tag {4} +$$ + +Following prior work [28, 7, 42], we use Hungarian algorithm [18] to search for the global optimal assignment. In contrast to the prior methods that compute matching loss only for $t = 1$ , our method considers the masks and ground truth of the whole clip, computes the matching loss globally, and conducts the global assignment. Finally, given the global optimal assignment, we use the $N_{gt}$ matched video-level predictions to supervise the model with the globally matched instances across the entire video. + +With the proposed assignment strategy, GIA, the temporal consistency of instance features across the video clip can be effectively enhanced, since we consider all frames as a whole to search for the optimal objective. Specifically, when the set of instance features across the video represents the target instance well in the first few frames, but fails to track it later, this strategy helps to find a more appropriate set to be optimized. As the global matching cost is lower, the selected features fit the target more closely and are more temporally consistent. + +# 3.3. Spatio-temporal Enhancement + +Previous online video instance segmentation methods focus on improving the representative ability of the semantic instance embeddings [38, 34, 15, 13]. The spatial features are not comprehensively investigated to boost the temporal association for online VIS. Thus, we further introduce Spatio-temporal Enhancement module (STE), leveraging the spatial information from the previous frame to enhance the temporal consistency of the online model. + +Given the $t$ -th frame in the video clip, with the framework described in Section 3.1, the mask embedding of + +the $k$ -th instance is generated from the instance prototype through an MLP: $\mathcal{M}_k^t = MLP(p_k^t)$ , and then the model segments the mask $\hat{m}_k^t$ by pixel-wise dot product between $\mathcal{P}_k^t$ and $\mathcal{M}_k^t$ , which can be formulated as: $\hat{m}_{k,i,j}^{t} = \langle \mathcal{P}_{k,i,j}^{t},\mathcal{M}_{k}^{t}\rangle$ , where $i$ and $j$ denote the spatial position of the pixel. + +As illustrated in Figure 2 (c), the spatial information of the frame is encoded in the pixel embeddings. To extract instance-wise spatial features, they can be exploited together with the predicted mask. Specifically, we perform spatial matting on pixel embeddings $\mathcal{P}_k^t$ , similar to image matting, to retrieve the instance-wise pixel embedding according to $\hat{m}_k^t$ . For each instance, we conduct pixel-wise multiplication between the original pixel embeddings and the binary mask to obtain the retrieved embeddings: + +$$ +\mathcal {R} _ {k} ^ {t} = \mathcal {P} _ {k} ^ {t} \odot \hat {m} _ {k} ^ {t}, \tag {5} +$$ + +where $\odot$ denotes the element-wise multiplication. Many of the retrieved embeddings $\mathcal{R}_k^t$ are redundant since they describe the same instance, and directly mining the spatial information with them is computationally inefficient. Average pooling is adopted to obtain a concise representation of the spatial features for each instance: + +$$ +\mathcal {S} _ {k} ^ {t} = \frac {\sum_ {i , j} \mathcal {R} _ {k , i , j} ^ {t}}{\sum_ {i , j} \mathbb {1} (\hat {m} _ {k , i , j} ^ {t} = 1)}, \tag {6} +$$ + +where $\mathbb{1}(\cdot)$ is the indicator function. The concise spatial features $S_{k}^{t}$ are sent to the next frame for temporal association. + +Having received from the previous frame the propagated semantic instance queries, i.e., $q^{t + 1} = p^t$ , and the spatial features, we follow the standard multi-head cross-attention layer (MHCA) [27] to incorporate the features from two aspects. The instance prototype $p^t$ is used as the query to decode the spatial features: + +$$ +q _ {k} ^ {t + 1} = \operatorname {M H C A} \left(p _ {k} ^ {t}, \mathcal {S} _ {1: N _ {v}} ^ {t}\right), \tag {7} +$$ + +where $q_{k}^{t + 1}$ is the updated instance query now. Notably, the positional embedding is shared by the spatial feature and instance query with regard to the same instance, which helps the model align the instance-wise information between frames. Finally, the updated instance query is fed into the Object Decoder $\mathcal{D}$ to produce spatio-temporally enhanced features. + +By performing the proposed enhancement module on the spatial features between frames, we effectively boost the spatio-temporal association of the features. In this way, the temporal consistency of the features is enhanced via the information from the spatial dimension, and the online video instance segmentation model manages to predict more temporally consistent results. + +# 3.4. Overall Loss + +The overall loss for training with a video clip as input is a linear combination of categorical and mask losses using the one-to-one global optimal assignment $\hat{\sigma}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\text {o v e r a l l}} = \lambda_ {c l s} \mathcal {L} _ {c l s} \left(c _ {k}, \bar {c} _ {\hat {\sigma} (k)}\right) + \lambda_ {b c e} \mathcal {L} _ {b c e} \left(\mathbf {m} _ {\mathbf {k}}, \hat {\mathbf {m}} _ {\hat {\sigma} (k)}\right) \\ + \lambda_ {d i c e} \mathcal {L} _ {d i c e} \left(\mathbf {m} _ {\mathbf {k}}, \hat {\mathbf {m}} _ {\tilde {\sigma} (k)}\right), \tag {8} \\ \end{array} +$$ + +where $\mathcal{L}_{cls}$ is the cross entropy loss, $\mathcal{L}_{bce}$ is the binary cross entropy loss and $\mathcal{L}_{dice}$ is the dice loss [23]. + +# 4. Experiments + +In this section, we evaluated the proposed TCOVIS on four benchmark datasets. Furthermore, we provided in-depth ablation studies of the effectiveness of TCOVIS. Finally, we presented several visualizations of the predictions from our model. + +# 4.1. Datasets + +We evaluated our approach on four VIS datasets: YouTube-VIS 2019 dataset [37], YouTube-VIS 2021 dataset [37], YouTube-VIS 2022 dataset [37] and OVIS dataset [24]. We present a brief description of them: + +YouTube-VIS 2019 & 2021 & 2022: YouTube-VIS 2019 [37] is the first VIS dataset and comprises 40 predefined categories of objects. This dataset included 4,883 unique video instances with 131,000 high-quality manual annotations. We followed the widely utilized training/test set split: 2,238 videos were selected for training, 302 videos were adopted for validation and 343 videos were used for testing. Further, YouTube-VIS 2021 improved the 40-category label set and added 4883 more unique video instances. Then, YouTube-VIS 2022 contained 71 additional long evaluation videos on the top of YouTube-VIS 2021. + +OVIS: Occluded video instance segmentation (OVIS) is also a challenging VIS dataset. OVIS included 901 videos in total with 25 semantic categories. This dataset contained 5,223 unique video instances and we followed the widely utilized training/test set split: 607 videos were selected for training, 607 videos were used for validation and 154 videos were adopted for testing. + +Following [37], the video-level average precision (AP) and average recall (AR) were adopted as the evaluation metrics on both YouTube-VIS and OVIS. + +# 4.2. Implementation Details + +We adopted the framework of GenVIS [13] which is built on VITA [14], but left out the similarity loss in VITA and the memory module in GenVIS. With the global assignment, the total loss as well the hyper-parameters were set the same as the video-level loss in VITA, which is a temporally extended loss function [16]. The model was trained + +with pseudo-videos from COCO images [20] as data augmentation, and with a batch size of 8 video clips of 6 frames. As we froze the backbone and image segmentation model, all experiments were conducted with 8 RTX 2080 Ti GPUs. The method was implemented on detectron2 [36]. + +# 4.3. Main Results + +Following the standard evaluation metrics [37], we compared TCOVIS with state-of-the-art approaches on four VIS benchmarks: YouTube-VIS 2019/2021/2022 and OVIS. + +YouTube-VIS 2019&2021. From Table 1, we can observe that TCOVIS has achieved very competitive performance using both lightweight backbones (ResNet-50) and powerful ones (Swin-L). Moreover, our method can even show better performance than offline methods, such as VITA [14] and offline version GenVIS [13]. On the more difficult YouTube-VIS 2021 dataset, TCOVIS surpasses GenVIS [13] in AP not only with ResNet-50 by 2.4 but also with Swin-L by 1.7. + +YouTube-VIS 2022. As shown in Table 2, TCOVIS performed best in both AP and AR on YouTube-VIS 2022 dataset, which is more challenging than 2019&2021 datasets. Especially with powerful backbone Swin-L, TCOVIS outperformed GenVIS [13] in AP with a huge margin of 4.9, which shows the effectiveness of our method for the complex scenarios. + +OVIS. Table 3 presents the comparisons on OVIS dataset, and we can find that TCOVIS also achieved the best 46.7 AP and 19.1 AR with Swin-L backbone. The results demonstrate that TCOVIS can deal with complicated situations where objects are heavily occluded in others. With ResNet-50 backbone, the performance is still the second best compared to the previous state-of-the-art methods. + +# 4.4. Ablation Study + +In this section, we provided ablation studies and discuss the effects of different settings in the proposed method. The experiments are conducted with a ResNet-50 [12] backbone on YouTube-VIS 2019/2021 [37] valid set. + +Effectiveness of the proposed assignment strategy and enhancement module. The ablation studies on the global instance assignment strategy and the spatio-temporal enhancement module are shown in Table 4. As for the assignment strategy, compared to LocPro as the baseline, the model with the proposed global assignment outperformed the baseline model by more than 0.8 on AP, $\mathrm{AP}_{50}$ and $\mathrm{AP}_{75}$ on YouTube-VIS 2019, and more than 1.2 on YouTube-VIS 2021. The consistently significant improvement indicates that the strategy with global optimal matching contributes to better overall segmentation, while LocPro only considers the local optimal results. Besides, the baseline forces the posterior features to conform to those at the very beginning leading to temporal inconsistency, since the accumulative + +
MethodTypeYouTube-VIS 2019YouTube-VIS 2021
AP\(AP_{50}\)\(AP_{75}\)\(AR_1\)\(AR_{10}\)AP\(AP_{50}\)\(AP_{75}\)\(AR_1\)\(AR_{10}\)
ResNet-50EfficientVIS [35]Offline37.959.743.040.346.634.057.537.333.842.5
IFC [16]Offline41.265.144.642.349.635.255.937.732.642.9
Mask2Former-VIS [4]Offline46.468.050.0--40.660.941.8--
\(TeViT^{\dagger}\)[39]Offline46.671.351.644.954.337.961.242.135.144.6
SeqFormer [33]Offline47.469.851.845.554.840.562.443.736.148.1
VITA [14]Offline49.872.654.549.461.045.767.449.540.953.6
\(GenVIS_{semi-online}\)[13]Offline51.372.057.849.560.046.367.050.240.653.2
Cross VIS [38]Online36.356.838.935.640.734.254.437.930.438.2
VISOLO [9]Online38.656.343.735.742.536.954.740.230.640.9
MinVIS [15]Online47.469.052.145.755.744.266.048.139.251.7
IDOL [34]Online49.574.052.947.758.743.968.049.638.050.9
\(GenVIS_{online}\)[13]Online50.071.554.649.559.747.167.551.541.654.7
TCOVISOnline52.373.557.649.860.249.571.253.841.355.9
Swin-LSeqFormer [33]Offline59.382.166.451.764.451.874.658.242.858.1
Mask2Former-VIS [4]Offline60.484.467.0--52.676.457.2--
VITA [14]Offline63.086.967.956.368.157.580.661.047.762.6
\(GenVIS_{semi-online}\)[13]Offline63.885.768.556.368.460.180.966.549.164.7
MinVIS [15]Online61.683.368.654.866.655.376.662.045.960.8
IDOL [34]Online64.387.571.055.669.156.180.863.545.060.1
\(GenVIS_{online}\)[13]Online64.084.968.356.169.459.680.965.848.765.0
TCOVISOnline64.186.669.555.869.061.382.968.048.665.1
+ +Table 1. Quantitative results on YouTube-VIS 2019 and 2021 validation sets. The results are respectively grouped by method types (Offline or Online) and backbone networks (ResNet-50 and Swin-L). We bold the best performance and underline the second. † denotes using MsgBoxIF [39] backbone which has a similar weight scale with ResNet-50. + +
MethodTypeAP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)\( AR_{10} \)
ResNet-50VITA [14]Offline32.653.939.330.342.6
GenVIS [13]Offline37.258.542.933.240.4
MinVIS [15]Online23.347.919.320.228.0
GenVIS [13]Online37.561.641.532.642.2
TCOVISOnline38.659.441.632.846.7
Swin-LVITA* [14]Offline41.163.044.039.344.3
GenVIS [13]Offline44.369.944.939.948.4
MinVIS* [15]Online33.154.833.729.536.6
GenVIS [13]Online45.169.147.339.848.5
TCOVISOnline51.073.053.541.756.5
+ +error impairs the model during the training process. Our proposed assignment strategy encourages the features of all time to fit the global optimal objective, which effectively enhances the temporal consistency. + +As for the effectiveness of the spatio-temporal enhancement module, the comparison is also shown in Table 4. The results show that the enhancement module brings improvements of 0.9 AP on YouTube-VIS 2019&2021, compared to using the proposed assignment strategy individually. In particular, the proposed module improved the performance + +Table 2. Quantitative results on YouTube-VIS 2022 validation dataset. We bold the highest accuracy and underline the second. $\star$ : Reproduced by [13]. + +
MethodTypeAP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)\( AR_{10} \)
ResNet-50\( TeViT^†[39] \)Offline17.434.915.011.221.8
VITA [14]Offline19.641.217.411.726.0
GenVIS [13]Offline34.559.435.016.638.3
CrossVIS [38]Online14.932.712.110.319.8
VISOLO [9]Online15.331.013.811.121.7
MinVIS [15]Online25.045.524.013.929.7
IDOL [34]Online30.251.330.015.037.5
GenVIS [13]Online35.860.836.216.339.6
TCOVISOnline35.360.736.615.739.5
Swin-LVITA [14]Offline27.751.924.914.933.0
GenVIS [13]Offline45.469.247.818.949.0
MinVIS [15]Online39.461.541.318.143.3
IDOL [34]Online42.665.745.217.949.6
GenVIS [13]Online45.269.148.419.148.6
TCOVISOnline46.770.949.519.150.8
+ +Table 3. Quantitative results on OVIS validation set. We bold the highest accuracy and underline the second. † denotes using MsgBoxT [39] backbone. + +by 1.2 and 1.7 in $\mathrm{AP}_{50}$ on two datasets, respectively. The performance improvements demonstrate that the proposed module effectively captures the spatial information from the previous frame and aggregates the semantic and spatial features across time. As a result, the delicate spatio-temporal design further enhances the temporal consistency of fea + +
LocProGIASTEYouTube-VIS 2019YouTube-VIS 2021
AP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)\( AR_{10} \)AP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)\( AR_{10} \)
✓--50.671.556.147.658.247.468.152.139.853.1
-✓-51.472.357.049.559.948.669.553.640.655.0
-✓✓52.373.557.649.860.249.571.253.841.355.9
+ +Table 4. Ablation study of the method for Global Instance Assignment strategy (GIA) and the Spatio-temporal Enhancement module (STE) on the YouTube-VIS 2019 / 2021 validation sets. LocPro denotes the Local Matching and Propagating technique. + +
ArchitectureAP\( AP_{50} \)\( AP_{75} \)\( AR_1 \)\( AR_{10} \)
Cross-Attn.49.571.253.841.355.9
Concat.47.968.652.741.455.1
Self-Cross48.270.051.841.054.1
Dec. Mod.48.569.753.940.654.7
Resp. Pos.48.870.352.541.555.5
Ins. Attn.49.270.653.540.955.1
+ +Table 5. Ablation study of the module manipulations of the Spatio-temporal Enhancement module on YouTube-VIS 2021. + +tures for the online video instance segmentation model. + +Incorporating the assignment strategy and the enhancement module, our proposed method gained remarkable performance improvements of 1.6 AP on YouTube-VIS 2019 and 2.1 AP on YouTube-VIS 2021 over the LocPro baseline. In a nutshell, the experimental results indicate that the temporally consistent features learned by our proposed TCOVIS significantly boost the performances on online video instance segmentation. + +Different manipulations of spatio-temporal enhancement. In Table 5, we compared our proposed spatiotemporal enhancement module with other optional manipulations. Cross-Attn. denotes our proposed manipulation following the standard multi-head cross-attention layer to decode the spatial features with shared positional embedding described in Section 3.3. Concat. indicates that we + +concatenated the corresponding spatial feature and prototype of an instance followed by an MLP to get the updated query. Self-Cross stands for adding an extra self-attention layer ahead of the cross-attention layer. Dec. Mod. is decoder modulation, reviewing the spatial feature for every Transformer decoder layer. Resp .Pos. denotes respective positional embeddings were used for the spatial feature and prototype when we performed cross-attention. Ins. Attn. indicates the instance-wise decoding in cross-attention layer. + +As shown, compared to all its counterparts, our proposed manipulation achieved the best performance. The first three variants are related to the feature aggregation, where $\text{Concat}$ is too naive to model the spatial and semantic features, while Self-Cross obscures the spatial feature. We inferred the performance decrease of Dec. Mod. comes from the information redundancy when aggregating them in every layer. The last two experiments studied the instance-wise correspondence. In the Resp. Pos. setting, explicit correspondence for spatial and semantic features of the same instance between frames is absent. Ins. Attn. only focuses on the instance itself across time and can slightly enhance the temporal consistency, however, neglecting the spatial features of other instances, the module fails to capture the spatial relationship among instances. The results confirm that the proposed manipulation effectively exploits the spatial information along the temporal dimension to enhance the + +![](images/e02a590bf053b004b385f4df18063bc998af74e6a36989d7d8e6857a0c3cebd4.jpg) +Figure 3. Qualitative results of TCOVIS, compared with VITA [14] and GenVIS [13]. On the left are the predictions on YouTube-VIS 2019 [37] and on the right are on OVIS [24]. Objects displayed in the same color denote the same instance. Our TCOVIS shows more temporally consistent results in these challenging scenes, where there are occlusions of the instance itself or others (Best viewed in color). + +![](images/da32317aa43fe599f2c359f5c3d13a24bbf204d5c8a7ba75e4fba8512b47cdc6.jpg) + +![](images/5cd4e3dded4c647a823986ce2103344de4ff575f6242de1dec23ce0ce7ec814d.jpg) +(a) + +![](images/1b347761c3d667e652ae9bc45f12f280e35fb98a79414620aadb9107f8918332.jpg) + +![](images/c74aef73dd4df29b67930669ca80865f483e581e23c7211eb9eaa73a791936da.jpg) + +![](images/577032a6dcc42c1025c906b976e3bc80e4d975493c51ec57cc3c7df87a057a31.jpg) + +![](images/8e2a2751357596502033264904d4a95c3fbc66562c583f8246492ab9ad64813a.jpg) +(b) + +![](images/9b1384c83a01b1d9d8eda31e6657b274e005e21aff0827d135f2457737b86cd1.jpg) + +![](images/57ce6d247a3fb0d2e897c3cba32d1861e4c4e2d3fdd702f5642020ecf4234484.jpg) + +![](images/efaf9b8db636caf212f4e04c6d1752d443134b54528dbbb1303617a1b36bf705.jpg) + +![](images/de8517a798c3821fc06cf9feca1b680257d5d64cbb4da75e304319a8e102dac9.jpg) +(c) + +![](images/0203e78a5af85d94493829199498ac92de7e787fc00c2efb26bcf7913ced7953.jpg) + +![](images/1675ef9e6fc7e01c827ec9b67d1817a9cf8ec1d3ac1a9e26bbdd59e7247836c5.jpg) + +![](images/2b244cfd36a47a45a5006e60a192d09261b8ce75e94d65ef55e5130c4bfbf370.jpg) + +![](images/db4b9a94343f3c7d4ea8f63b0c91408935f2d0193180c3842369e57082145df4.jpg) +(d) + +![](images/dee460dda09ee30d0615cca21ab36e6242da609cd16d2d5949c405246e06306b.jpg) +Figure 4. Qualitative results of TCOVIS in four challenging cases on YouTube-VIS 2019 [37] and OVIS [24]: (a) fast movement, (b) heavy occlusion, (c) low resolution, and (d) crowded scene. Objects displayed in the same color denote the same instance. The qualitative results demonstrate the effectiveness and robustness of TCOVIS (Best viewed in color). + +![](images/72aadeac444e2f6c38758e6e3f9dc4f860c6209b04be8df8a459799051d17ab0.jpg) + +![](images/09056a28bbe9e89a3c8e681e0643adbc5641a6b7c843d60db5b5fc42c08b8567.jpg) + +temporal consistency of features. + +# 4.5. Qualitative results + +In Figure 3, we show the qualitative comparisons of the proposed TCOVIS with VITA [14] and GenVIS [13] on YouTube-VIS 2019 and OVIS datasets. In the left scene where there is a zebra with self-occlusion, GenVIS fails to segment its head resulting in fragmented predictions, however, TCOVIS performs temporally consistent segmentation with impressive accuracy. On the right is a difficult case where there are two dogs with similar appearances grappling with each other. VITA incorrectly detects more than two dogs and fails to track the left one, while GenVIS fails to handle the margin of two instances, e.g. the nose and the paw of the left dog. TCOVIS successfully tracks and segments the instances, demonstrating its effectiveness. + +In Figure 4, we provide more qualitative results of the proposed method in variously challenging cases, which are all chosen from the mentioned benchmarks [37, 24]. As shown in Figure 4 (a), our method successfully tracks the surfer and the surfboard with fast movement. In the second row, we present a difficult case with severe occlusion, in which our method performs admirably by accurately segmenting the claw (at the bottom) of a parrot despite the presence of a wooden stick that partially obstructs its lower body. Figure 4 (c) illustrates a low-resolution case and our method still achieves good performance. In Figure 4 (d), + +we depict a crowded scene where TCOVIS is capable of handling multiple instances that share similar appearances and exhibit complex interactions. All the qualitative results in the challenging situations demonstrate the effectiveness and robustness of the proposed method. + +# 5. Conclusion + +In this paper, we propose a new online video instance segmentation method, TCOVIS, to fully exploit the temporal information within a video and produce temporally consistent predictions. Based on the query propagation framework, we propose a global instance assignment strategy to perform global optimal matching with the consideration of the entire video and supervise the model with the global optimal objective. We further devise a spatio-temporal enhancement module to capture the spatial feature and aggregate it with the semantic feature between frames. The effectiveness of our method is evaluated with experimental results on YouTube-VIS 2019/2021/2022 and OVIS. + +# Acknowledgement + +This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603. + +# References + +[1] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, pages 158-177, 2020. 2 +[2] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, pages 9739-9748, 2020. 1, 2 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 3, 4 +[4] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 3, 7 +[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 1, 3 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2 +[7] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In ICCV, pages 6910-6919, 2021. 5 +[8] Yang Fu, Linjie Yang, Ding Liu, Thomas S Huang, and Humphrey Shi. Compfeat: Comprehensive feature aggregation for video instance segmentation. In AAAI, pages 1361-1369, 2021. 1, 4 +[9] Su Ho Han, Sukjun Hwang, Seoung Wug Oh, Yeonschool Park, Hyunwoo Kim, Min-Jung Kim, and Seon Joo Kim. Visolo: Grid-based space-time aggregation for efficient online video instance segmentation. In CVPR, pages 2896-2905, 2022. 3, 7 +[10] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. In NeurIPS, 2022. 1, 2, 3, 4 +[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[13] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In CVPR, pages 14623–14632, 2023. 1, 2, 3, 4, 5, 6, 7, 8, 9 +[14] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. VITA: Video instance segmentation via object token association. In NeurIPS, 2022. 1, 2, 3, 6, 7, 8, 9 + +[15] De-An Huang, Zhiding Yu, and Anima Anandkumar. MinVIS: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 1, 2, 3, 5, 7 +[16] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 34, 2021. 1, 2, 6, 7 +[17] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks for multiple object tracking and segmentation. NeurIPS, 34, 2021. 1 +[18] Harold W Kuhn. The hungarian method for the assignment problem. NRL, 1955. 4, 5 +[19] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In ICCV, pages 1739-1748, 2021. 1, 2 +[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6 +[21] Dongfang Liu, Yiming Cui, Wenbo Tan, and Yingjie Chen. Sg-net: Spatial granularity network for one-stage video instance segmentation. In CVPR, pages 9816-9825, 2021. 3 +[22] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 3 +[23] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571, 2016. 4, 6 +[24] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. IJCV, 130(8):2022-2039, 2022. 2, 6, 8, 9 +[25] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 14454-14463, 2021. 1 +[26] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, pages 282-298, 2020. 1, 3 +[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2, 5 +[28] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. 5 +[29] Tao Wang, Ning Xu, Kean Chen, and Weiyao Lin. End-to-end video instance segmentation via spatial-temporal graph neural networks. In ICCV, pages 10797-10806, 2021. 3 +[30] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, pages 649-665, 2020. 3 + +[31] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, pages 8741-8750, 2021. 1, 2 +[32] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In CVPR, pages 12352-12361, 2021. 1 +[33] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, pages 553-569, 2022. 1, 7 +[34] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, pages 588-605, 2022. 1, 2, 3, 4, 5, 7 +[35] Jialian Wu, Sudhir Yarram, Hui Liang, Tian Lan, Junsong Yuan, Jayan Eledath, and Gerard Medioni. Efficient video instance segmentation via tracklet query and proposal. In CVPR, pages 959-968, 2022. 1, 2, 3, 7 +[36] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6 +[37] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, pages 5188-5197, 2019. 1, 2, 3, 6, 8, 9 +[38] Shusheng Yang, Yuxin Fang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Crossover learning for fast online video instance segmentation. In ICCV, pages 8043-8052, 2021. 1, 2, 3, 5, 7 +[39] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Wenyu Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In CVPR, pages 2885-2895, 2022. 1, 2, 7 +[40] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, pages 659-675. Springer, 2022. 3 +[41] Zitong Zhan, Daniel McKee, and Svetlana Lazebnik. Robust online video instance segmentation with track queries. arXiv preprint arXiv:2211.09108, 2022. 1, 2, 3, 4 +[42] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable $\{\mathrm{detr}\}$ : Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 5 \ No newline at end of file diff --git a/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/images.zip b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c250460a83647a16a79098b0c1690a99cc2d6ce4 --- /dev/null +++ b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db408bfab70092d2d39b197e922746be6f00f39406883aa4c27a8f538571904d +size 869319 diff --git a/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/layout.json b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a7dd55740bbdbb4f35de202d460fbc32831be3f3 --- /dev/null +++ b/2023/TCOVIS_ Temporally Consistent Online Video Instance Segmentation/layout.json @@ -0,0 +1,8298 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 512, + 121 + ], + "type": "text", + "content": "TCOVIS: Temporally Consistent Online Video Instance Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 143, + 453, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 143, + 453, + 158 + ], + "spans": [ + { + "bbox": [ + 143, + 143, + 453, + 158 + ], + "type": "text", + "content": "Junlong Li Bingyao Yu Yongming Rao Jie Zhou Jiwen Lu*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 158, + 432, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 158, + 432, + 171 + ], + "spans": [ + { + "bbox": [ + 162, + 158, + 432, + 171 + ], + "type": "text", + "content": "Department of Automation, Tsinghua University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 171, + 494, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 171, + 494, + 186 + ], + "spans": [ + { + "bbox": [ + 99, + 171, + 494, + 186 + ], + "type": "text", + "content": "Beijing National Research Center for Information Science and Technology, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 239, + 289, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 550 + ], + "type": "text", + "content": "In recent years, significant progress has been made in video instance segmentation (VIS), with many offline and online methods achieving state-of-the-art performance. While offline methods have the advantage of producing temporally consistent predictions, they are not suitable for real-time scenarios. Conversely, online methods are more practical, but maintaining temporal consistency remains a challenging task. In this paper, we propose a novel online method for video instance segmentation, called TCOVIS, which fully exploits the temporal information in a video clip. The core of our method consists of a global instance assignment strategy and a spatio-temporal enhancement module, which improve the temporal consistency of the features from two aspects. Specifically, we perform global optimal matching between the predictions and ground truth across the whole video clip, and supervise the model with the global optimal objective. We also capture the spatial feature and aggregate it with the semantic feature between frames, thus realizing the spatio-temporal enhancement. We evaluate our method on four widely adopted VIS benchmarks, namely YouTube-VIS 2019/2021/2022 and OVIS, and achieve state-of-the-art performance on all benchmarks without bells-and-whistles. For instance, on YouTube-VIS 2021, TCOVIS achieves 49.5 AP and 61.3 AP with ResNet-50 and Swin-L backbones, respectively. Code is available at https://github.com/jun-long-li/TCOVIS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 128, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 128, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 128, + 600 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 608, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 608, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 608, + 287, + 693 + ], + "type": "text", + "content": "Video instance segmentation (VIS) is a challenging and representative video understanding task recently introduced in [37]. It aims at detecting, segmenting and tracking instances across a video. VIS is attracting increasing attention for various real-world applications such as video editing, video surveillance, augmented reality and autonomous driving. Recently introduced VIS methods can be roughly" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 323 + ], + "type": "text", + "content": "categorized into two groups: offline methods and online methods. Offline methods [2, 16, 19, 31, 33, 35] take as input the whole video and perform the segmentation of instance sequence for the whole video at once. Online methods [8, 34, 17, 10, 38], on the contrary, take as input a video frame by frame and generate the pre-frame object instances while associating the frame-wise results across frames. Both offline and online methods have achieved impressing performance on the VIS task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 326, + 547, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 447 + ], + "type": "text", + "content": "Offline methods have an inherent advantage in producing temporally consistent predictions, since delicate temporal communication and association mechanisms can be adopted throughout the video [39, 33, 14] to handle the overall temporal information and impose an explicit constraint on the temporal consistency. However, the video-in and video-out offline manner is not suitable for real-time scenarios. Conversely, online methods are more practical and making considerable progress but suffer from temporal inconsistency (as shown in Figure 1), remaining a great challenge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "Online methods rely on specific instance association applied across frames, since only one frame is observed at a time. Existing association techniques can be grouped into two categories, including tracking-by-detection and query propagation-based paradigms. Tracking-by-detection methods [37, 34, 32] generate the per-frame instances independently by existing instance segmentation models [11, 26, 5] and track instances via tracking heads [37] or instance embeddings matching [34, 15]. In this way, the features of different frames are isolated before tracking, which results in temporal inconsistency. Query propagation-based methods [13, 10, 41] are inspired by query-based methods [3, 25] and they propagate the query across frames to decode a unique instance without heuristic matching algorithms. Despite the explicit temporal link of queries, the temporal consistency is impaired by the Local Matching and Propagating (LocPro) scheme, where they first perform local optimal matching between the predictions and ground truth at the beginning of the video, and then propagate the assignment across frames, forcing all features from subsequent frames to follow. The LocPro is not suitable for the holistic optimization across frames and results in temporal inconsistencies." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1097" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 282, + 194 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 282, + 194 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 282, + 194 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 282, + 194 + ], + "type": "image", + "image_path": "614e5bdb04624b7cb1a444c6913b3afaf4d077bd360c7bcc669ca7753656b233.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 287, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 254 + ], + "type": "text", + "content": "Figure 1. Visualization of predictions from the previous online method (the online GenVIS [13]) and our TCOVIS. The previous method generates temporally inconsistent predictions, while our proposed TCOVIS achieves temporal consistency and outperforms the previous method (Best viewed in color)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 268, + 286, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 286, + 304 + ], + "type": "text", + "content": "tency. Thus, achieving temporal consistency is challenging for online methods and also not comprehensively investigated by previous online VIS methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 305, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 700 + ], + "type": "text", + "content": "In this paper, we propose a novel online method for video instance segmentation, named TCOVIS, to fully exploit the temporal information within a video. We take as the baseline framework an existing online VIS model (GenVIS [13]) with the query propagation-based instance association. We first introduce the global instance assignment strategy to perform global optimal matching. Different from the previous online methods [13, 41, 10], which obtain per-frame matching cost, assign labels locally on the beginning frame and propagate across frames, we collect the predictions across frames, compute the global matching cost with the video segmentation ground truth and supervise the model with the global instance assignment, encouraging features across the video to be optimized for a global optimal objective. As online methods focus on improving the representative ability of the semantic instance embeddings [38, 34], which is achieved via learning more discriminative semantic embedding in those using heuristic matching [34, 15], or via reviewing the memory of semantic embedding across frames [13] in query propagation-based methods, the spatial features are not comprehensively investigated. We further propose spatio-temporal enhancement module, leveraging the spatial information from the previous frame to enhance the temporal consistency. We perform spatial matting on the pixel embeddings to retrieve the instance-wise spatial features and adopt the cross-attention layer to aggregate the spatial and semantic features across frames, thus realizing the spatio-temporal enhancement. As shown in Figure 1, the previous online method [13] produces temporally inconsistent results, as exemplified by an object abruptly appearing in front of the cat in a mid-frame, while the proposed TCOVIS outperforms the previous one and generates temporally consistent predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "To validate the effectiveness of the proposed method, ex" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "periments are conducted on four widely adopted VIS benchmarks, i.e., YouTube-VIS 2019 [37], YouTube-VIS 2021, YouTube-VIS 2022 and Occluded VIS (OVIS) [24]. Without bells-and-whistles, our proposed method achieves state-of-the-art performance on all benchmarks, outperforming other online methods, e.g., on YouTube-VIS 2021, TCOVIS achieves 49.5 AP and 61.3 AP with ResNet-50 and Swin-L backbones, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 167, + 525, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 167, + 525, + 179 + ], + "spans": [ + { + "bbox": [ + 317, + 167, + 525, + 179 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 188, + 545, + 371 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 188, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 545, + 236 + ], + "type": "text", + "content": "- TCOVIS performs a novel global instance assignment strategy for online video instance segmentation. The model is optimized for the global optimal objective to generate more temporally consistent predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 243, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 545, + 303 + ], + "type": "text", + "content": "- The further proposed spatio-temporal enhancement module captures the spatial feature and aggregates it with the semantic feature between frames, which fully utilizes the spatial information and facilitates the temporal consistency enhancement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 312, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 312, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 312, + 545, + 371 + ], + "type": "text", + "content": "- The proposed method achieves state-of-the-art performance on four widely used video instance segmentation benchmarks (YouTube-VIS 2019/2021/2022 and OVIS). Such achievements demonstrate the effectiveness of our proposed method." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 382, + 397, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 397, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 397, + 395 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 402, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 640 + ], + "type": "text", + "content": "Offline Video Instance Segmentation. Offline methods take as input the whole video and predict instance sequence for all frames at once. Mask propagation and box ensemble techniques are used to improve the predictions and association [1, 2, 19], but they are not end-to-end learnable due to the complex inference process. VisTR [31] extends DETR [42] from the image domain and introduces the transformer [27] to the VIS domain. EfficientVIS [35] and IFC [16] relax the heavy overhead of VisTR via an iterative query-video interaction and memory token communication, respectively. TeViT [39] contains a vision transformer [6] backbone instead of CNN and efficiently builds correspondence between the instance and query. VITA [14] models relationships among instances with the distilled condensed object tokens, without using the dense spatiotemporal backbone features. Offline methods exploit rich temporal knowledge from the whole clip and have the advantage of producing temporally consistent results, however, the offline manner is not suitable for the application in real-time scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Online Video Instance Segmentation. Instead of processing the entire video before predictions, online methods only leverage the information from the previous frames and segment the video frame-by-frame. The association paradigms of the previous online methods roughly fall into two groups: Tracking-by-detection and Query-propagation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1098" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "Most online methods [37, 29, 21, 34] follow the tracking-by-detection paradigm. MaskTrack R-CNN [37] is the baseline method and extends the Mask R-CNN [11] with an extra tracking head for temporal association. CrossVIS [38] proposes a crossover learning scheme to utilize the current contextual information for other frames. VISOLO [9] builds on the image instance segmentation method SOLO [30] and takes advantage of the grid form previous information for memory matching and features aggregation. MinVIS [15] and IDOL [34] make use of the discriminative instance embeddings for matching between frames. With the heuristic matching technique designed for instance association across frames, temporal inconsistency comes from the frame-wise modeling before tracking." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 289, + 471 + ], + "type": "text", + "content": "Object association with query propagation has been explored in the multi-object tracking (MOT) task [22, 40]. TrackFormer [22] tracks the seen objects of previous frames with a track query subset and detects the newly appeared objects in current frame with an extra object query subset. MOTR [40] extends the paired-frames training scheme to multiple frames for long-range temporal association. The query propagation-based object association is recently introduced to VIS [41, 10, 13]. ROVIS [41] follows TrackFormer [22] to detect and track instances with two subsets of queries. InsPro [10] and GenVIS [13] propagate the queries without heuristics, i.e., handcrafted rules to combine two types of queries, and achieve association across frames. However, previous query propagation-based methods also propagate the local assignment to supervise the model with the local optimal results, which leads to temporal inconsistency across the entire video. We adopt the query propagation framework but introduce the global instance assignment to enhance the temporal consistency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 487, + 104, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 104, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 104, + 499 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 713 + ], + "type": "text", + "content": "Given a video clip with consecutive image frames, online video instance segmentation methods generate frame-level object instances upon an instance segmentation models [26, 5], utilizing the instance queries propagated from previous frames [35, 13]. We have already discussed that better performance of the online VIS method relies on more temporally consistent instance features among a video. To this end, we propose a novel end-to-end online VIS method TCOVIS (Figure 2), to improve the temporal consistency of instance features via the global instance assignment strategy and spatio-temporal enhancement module. In this section, we first introduce the online VIS pipeline in Section 3.1. Then the details of the proposed global instance assignment strategy and spatio-temporal enhancement module will be described in Section 3.2 and Section 3.3, respectively. Finally, in Section 3.4, we describe the overall loss for training the model end-to-end." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 499, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 499, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 499, + 85 + ], + "type": "text", + "content": "3.1. Online Video Instance Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "content": "Following state-of-the-art VIS methods [15, 4], we adopt the advanced Masked-attention Mask Transformer (Mask2Former [5]) as the image instance segmentation network (Img. SegNet.) in this paper. Assume that the input video clip with T frames is denoted as " + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{T \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "content": ". With each frame " + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "content": " as input, frame-wise activation map is extracted by the backbone and Transformer encoder. Then following the query-based mechanism of DETR [3], " + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "inline_equation", + "content": "N_{fq}" + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "content": " object queries of " + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "content": " dimensions are used to parse an input frame, which are called frame object queries " + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{C \\times N_{fq}}" + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "content": ". Each object in the frame is decoded by the frame object queries from the spatial features through a multiple-level Transformer decoder, and represented as an object embedding of " + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "content": " dimensions. The object embeddings are used for classification and together with the pixel embeddings from the pixel decoder generating the mask for object instances. Class predictions are produced through a linear layer from the object embeddings. Mask embeddings are generated by an MLP linked to the object embeddings, and the model finally segments objects by pixel-wise dot product between per-pixel embeddings " + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{C \\times \\frac{H}{S} \\times \\frac{W}{S}}" + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "content": " and mask embeddings " + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{C \\times N_{fq}}" + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 89, + 547, + 365 + ], + "type": "text", + "content": " is the stride of the spatial feature map." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "content": "As for online video segmentation, we follow the online scheme of GenVIS [13] which propagates the instance queries from previous frames. To resolve the computational limitation, the framework adopts VITA [14] that regards the frame object queries " + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "content": " as a concise representation of objects in a frame and then feeds them into the Object Encoder " + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "content": " for intra-frame relationship. The Object Decoder " + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "content": " takes as input " + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "content": " video instance queries " + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "content": " and aggregates information from frame object queries. The temporal instance association is implemented through a query-based temporal propagation mechanism, where the output of " + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "content": ", instance prototypes denoted as " + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "content": ", are concise representations of instances [13], are not only utilized for classifying and segmenting in current frame at " + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "content": ", but also serve as the instance queries for the next frame at " + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 304, + 365, + 547, + 545 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 368, + 550, + 545, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 550, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 368, + 550, + 545, + 564 + ], + "type": "interline_equation", + "content": "q ^ {t + 1} = p ^ {t} = \\mathcal {D} \\left(q ^ {t}, \\mathcal {E} \\left(f ^ {t}\\right)\\right). \\tag {1}", + "image_path": "a451e41ac2cca1d16f6272aa6d8944499bda04af2471f5f0a65f2ceb859a33f1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 569, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 605 + ], + "type": "text", + "content": "With this propagation mechanism, the model can simply run in an online manner and associate the frame-wise outputs without heuristic matching algorithms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 665 + ], + "type": "text", + "content": "We leave out the Instance Prototype Memory module in [13]. During the training process, we freeze the image instance segmentation model and only train the following modules, to efficiently make use of memory. More implementation details are described in Section 4.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 671, + 460, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 460, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 460, + 684 + ], + "type": "text", + "content": "3.2. Global Instance Assignment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "With the query propagation mechanism discussed above, the associated video instance queries along the temporal" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1099" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 534, + 308 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 534, + 308 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 534, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 534, + 308 + ], + "type": "image", + "image_path": "0b62bb26eae29dcf53ab8d48dd7ef1e7eee710a0cabdaf22d0115d12b1375694.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 313, + 548, + 370 + ], + "lines": [ + { + "bbox": [ + 46, + 313, + 548, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 548, + 370 + ], + "type": "text", + "content": "Figure 2. (a) The overall illustration of our TCOVIS. It supervises the model with a global optimal objective during training and utilizes spatial features via the enhancement module between frames. (b) Details of global instance assignment. Different from the local matching and propagating technique, we conduct global optimal matching and assignment. (c) In the spatio-temporal enhancement module, we perform matting on pixel embeddings according to predicted masks and then aggregate the spatial and semantic features between frames to enhance temporal consistency. Numbers in colors denote the local matching loss and global matching loss (Best viewed in color)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "type": "text", + "content": "dimension " + }, + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "type": "inline_equation", + "content": "\\{q_{k}^{1:T}\\}_{1:N_{v}}" + }, + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "type": "text", + "content": ", are used to extract the features of a unique instance, i.e. the " + }, + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "type": "text", + "content": "-th instance, throughout the whole input video clip. Meanwhile, the associated instance prototypes " + }, + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "type": "inline_equation", + "content": "\\{p_{k}^{1:T}\\}_{1:N_{v}}" + }, + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "type": "text", + "content": ", i.e. the corresponding outputs from " + }, + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 379, + 288, + 464 + ], + "type": "text", + "content": ", represent this unique instance within the whole video. For brevity, we define the associated instance prototypes/features as a set of prototypes/features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 464, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 619 + ], + "type": "text", + "content": "Previously, online VIS methods that associate frame-level results without heuristic matching algorithms [13, 41, 10] adopt a Local Matching and Propagating (LocPro) technique. This technique matches the predictions and ground truth on a frame level during training. In other words, it conducts one-to-one bipartite matching between ground truth instances and predictions of a frame (what we regard as Local Matching), and propagates the matching from previous frames to subsequent frames. Specifically, given a video clip consisting of " + }, + { + "bbox": [ + 46, + 464, + 287, + 619 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 464, + 287, + 619 + ], + "type": "text", + "content": " consecutive frames, it first computes the pair-wise matching cost " + }, + { + "bbox": [ + 46, + 464, + 287, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{match}^{t}" + }, + { + "bbox": [ + 46, + 464, + 287, + 619 + ], + "type": "text", + "content": " between predictions and ground truth instances on the initial frame (the first frame or the frame objects appear):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 623, + 287, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 623, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 71, + 623, + 287, + 638 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a t c h}} ^ {t = 1} = \\lambda_ {c l s} \\mathcal {L} _ {c l s} ^ {t = 1} + \\lambda_ {b c e} \\mathcal {L} _ {b c e} ^ {t = 1} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} ^ {t = 1}. \\tag {2}", + "image_path": "795b186bf339c33419799ad332d670d72b7682dcbf2690334ac3e18f750d8df0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": "The frame-wise " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{match}^{t}" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " is composed of categorical loss and mask loss. The categorical loss adopts the cross entropy loss " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}^{t}" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": ". The mask loss consists of a binary cross entropy loss " + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bce}^{t}" + }, + { + "bbox": [ + 47, + 641, + 288, + 714 + ], + "type": "text", + "content": " and a dice loss [23]. From the cost matrix, it follows DETR [3] and uses Hungarian algorithm [18] for optimal frame-level matching." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": "As discussed in Section 1, Local Matching only obtains optimal matching on the initial frame, and cannot achieve the global optimal matching of the whole video. As illustrated in Figure 2 (b), both dark and light green prototypes attempt to represent the person. With local matching, the ground truth will be assigned to the light one with a smaller local matching cost, however, the dark one performs better from a global perspective and is supposed to be assigned. Training the model with LocPro forces all the predictions of subsequent frames to conform to the initial frame, resulting in temporal inconsistency of the instance features among the video, because inappropriate previous matching brings accumulative error to the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 540, + 546, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 546, + 624 + ], + "type": "text", + "content": "We introduce Global Instance Assignment (GIA) strategy and expect two functionalities: (1) all frames in the video clip are considered when conducting global matching, and (2) the global assignment encourages the instance features among frames to be optimized for a global optimal objective, both of which serve the temporal consistency of instance features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "During training, different from previous online VIS methods [8, 34] that conduct frame-level local matching and provide supervision signals to each frame according to the matched pairs, we leave out the halfway matching and supervision. Consecutive input frames of a video clip pass through the VIS model in an online manner and the model generates the predictions of all frames " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{y}}_k\\}_{1:N_v} =" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1100" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\{\\hat{y}_k^t\\}_{1:N_v}^{1:T}" + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "text", + "content": ", each of which consists of a category probability " + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\hat{c}_k^t" + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "text", + "content": " and a segmentation mask probability " + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\hat{m}_k^t" + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "text", + "content": ". To conduct global assignment, we collect predictions of all frames and as well the ground truth video instance segmentation " + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y}_k\\}_{1:N_{gt}} = \\{y_k^{1:T}\\}_{1:N_{gt}}" + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "text", + "content": ", including the category label " + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "text", + "content": " and its binary segmentation masks " + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_k = m_k^{1:T}" + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "text", + "content": ". Since the ground truth of an instance only has one category label, we first compute the average predicted category probability across a video clip: " + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\bar{c}_k = \\sum_{t=1}^T \\hat{c}_k^t / T" + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "text", + "content": " and also collect the masks " + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_k = \\hat{m}_k^{1:T}" + }, + { + "bbox": [ + 47, + 71, + 289, + 194 + ], + "type": "text", + "content": ". The global matching cost is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 207, + 287, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 287, + 247 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {m a t c h}} ^ {\\text {g l o b a l}} = \\lambda_ {c l s} \\mathcal {L} _ {c l s} \\left(c _ {k}, \\bar {c} _ {\\sigma (k)}\\right) + \\lambda_ {b c e} \\mathcal {L} _ {b c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\sigma (k)}\\right) \\\\ + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\sigma (k)}\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "6abffe08f15ae1958966dd8a3d7811e3c193e2983fdd7d3d0490223e2709ea9d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathfrak{S}_{N_v}" + }, + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "type": "text", + "content": " is a permutation of " + }, + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "type": "text", + "content": " elements. One-to-one bipartite global matching between " + }, + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{y}}_k\\}_{1:N_v}" + }, + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y}_k\\}_{1:N_{gt}}" + }, + { + "bbox": [ + 47, + 248, + 287, + 297 + ], + "type": "text", + "content": " is performed to find the global optimal assignment and the objective can be formally described as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 302, + 287, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 302, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 90, + 302, + 287, + 336 + ], + "type": "interline_equation", + "content": "\\hat {\\sigma} = \\underset {\\sigma \\in \\mathfrak {S} _ {N _ {v}}} {\\arg \\max } \\sum_ {k = 1} ^ {N _ {g t}} \\mathcal {L} _ {\\text {m a t c h}} ^ {\\text {g l o b a l}} (\\mathbf {y} _ {k}, \\hat {\\mathbf {y}} _ {\\sigma (k)}). \\tag {4}", + "image_path": "5feb492e9843789295d24e6b42ab229bf5ba2bd339ce962a941c49038d900025.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "text", + "content": "Following prior work [28, 7, 42], we use Hungarian algorithm [18] to search for the global optimal assignment. In contrast to the prior methods that compute matching loss only for " + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "text", + "content": ", our method considers the masks and ground truth of the whole clip, computes the matching loss globally, and conducts the global assignment. Finally, given the global optimal assignment, we use the " + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "inline_equation", + "content": "N_{gt}" + }, + { + "bbox": [ + 46, + 341, + 287, + 449 + ], + "type": "text", + "content": " matched video-level predictions to supervise the model with the globally matched instances across the entire video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": "With the proposed assignment strategy, GIA, the temporal consistency of instance features across the video clip can be effectively enhanced, since we consider all frames as a whole to search for the optimal objective. Specifically, when the set of instance features across the video represents the target instance well in the first few frames, but fails to track it later, this strategy helps to find a more appropriate set to be optimized. As the global matching cost is lower, the selected features fit the target more closely and are more temporally consistent." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 575, + 214, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 214, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 214, + 588 + ], + "type": "text", + "content": "3.3. Spatio-temporal Enhancement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 689 + ], + "type": "text", + "content": "Previous online video instance segmentation methods focus on improving the representative ability of the semantic instance embeddings [38, 34, 15, 13]. The spatial features are not comprehensively investigated to boost the temporal association for online VIS. Thus, we further introduce Spatio-temporal Enhancement module (STE), leveraging the spatial information from the previous frame to enhance the temporal consistency of the online model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Given the " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "-th frame in the video clip, with the framework described in Section 3.1, the mask embedding of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "the " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "-th instance is generated from the instance prototype through an MLP: " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k^t = MLP(p_k^t)" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": ", and then the model segments the mask " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\hat{m}_k^t" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " by pixel-wise dot product between " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k^t" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k^t" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": ", which can be formulated as: " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "\\hat{m}_{k,i,j}^{t} = \\langle \\mathcal{P}_{k,i,j}^{t},\\mathcal{M}_{k}^{t}\\rangle" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": " denote the spatial position of the pixel." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 146, + 546, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 146, + 546, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 146, + 546, + 254 + ], + "type": "text", + "content": "As illustrated in Figure 2 (c), the spatial information of the frame is encoded in the pixel embeddings. To extract instance-wise spatial features, they can be exploited together with the predicted mask. Specifically, we perform spatial matting on pixel embeddings " + }, + { + "bbox": [ + 304, + 146, + 546, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k^t" + }, + { + "bbox": [ + 304, + 146, + 546, + 254 + ], + "type": "text", + "content": ", similar to image matting, to retrieve the instance-wise pixel embedding according to " + }, + { + "bbox": [ + 304, + 146, + 546, + 254 + ], + "type": "inline_equation", + "content": "\\hat{m}_k^t" + }, + { + "bbox": [ + 304, + 146, + 546, + 254 + ], + "type": "text", + "content": ". For each instance, we conduct pixel-wise multiplication between the original pixel embeddings and the binary mask to obtain the retrieved embeddings:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 391, + 264, + 545, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 264, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 391, + 264, + 545, + 278 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {k} ^ {t} = \\mathcal {P} _ {k} ^ {t} \\odot \\hat {m} _ {k} ^ {t}, \\tag {5}", + "image_path": "5bd47d7cafede577dc9f14d3e6d59d9fc7dd3f26e6e09ef78dafbd9b22bcec33.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "type": "text", + "content": " denotes the element-wise multiplication. Many of the retrieved embeddings " + }, + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_k^t" + }, + { + "bbox": [ + 304, + 289, + 547, + 361 + ], + "type": "text", + "content": " are redundant since they describe the same instance, and directly mining the spatial information with them is computationally inefficient. Average pooling is adopted to obtain a concise representation of the spatial features for each instance:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 369, + 371, + 545, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 371, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 369, + 371, + 545, + 402 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {k} ^ {t} = \\frac {\\sum_ {i , j} \\mathcal {R} _ {k , i , j} ^ {t}}{\\sum_ {i , j} \\mathbb {1} (\\hat {m} _ {k , i , j} ^ {t} = 1)}, \\tag {6}", + "image_path": "5d1bf5c5da2ee0629bbad989003dca469ad6863eca34a1968b628912abbea10b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "text", + "content": " is the indicator function. The concise spatial features " + }, + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "inline_equation", + "content": "S_{k}^{t}" + }, + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "text", + "content": " are sent to the next frame for temporal association." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 437, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 508 + ], + "type": "text", + "content": "Having received from the previous frame the propagated semantic instance queries, i.e., " + }, + { + "bbox": [ + 304, + 437, + 545, + 508 + ], + "type": "inline_equation", + "content": "q^{t + 1} = p^t" + }, + { + "bbox": [ + 304, + 437, + 545, + 508 + ], + "type": "text", + "content": ", and the spatial features, we follow the standard multi-head cross-attention layer (MHCA) [27] to incorporate the features from two aspects. The instance prototype " + }, + { + "bbox": [ + 304, + 437, + 545, + 508 + ], + "type": "inline_equation", + "content": "p^t" + }, + { + "bbox": [ + 304, + 437, + 545, + 508 + ], + "type": "text", + "content": " is used as the query to decode the spatial features:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 518, + 545, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 518, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 367, + 518, + 545, + 533 + ], + "type": "interline_equation", + "content": "q _ {k} ^ {t + 1} = \\operatorname {M H C A} \\left(p _ {k} ^ {t}, \\mathcal {S} _ {1: N _ {v}} ^ {t}\\right), \\tag {7}", + "image_path": "1bf4d0858e04ae4da4d5c783a934e85f079cbd9a2da8a61541daa11612dc4dcf.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "inline_equation", + "content": "q_{k}^{t + 1}" + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": " is the updated instance query now. Notably, the positional embedding is shared by the spatial feature and instance query with regard to the same instance, which helps the model align the instance-wise information between frames. Finally, the updated instance query is fed into the Object Decoder " + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": " to produce spatio-temporally enhanced features." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "By performing the proposed enhancement module on the spatial features between frames, we effectively boost the spatio-temporal association of the features. In this way, the temporal consistency of the features is enhanced via the information from the spatial dimension, and the online video instance segmentation model manages to predict more temporally consistent results." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "1101" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 129, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 129, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 129, + 83 + ], + "type": "text", + "content": "3.4. Overall Loss" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 90, + 287, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 90, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 90, + 287, + 126 + ], + "type": "text", + "content": "The overall loss for training with a video clip as input is a linear combination of categorical and mask losses using the one-to-one global optimal assignment " + }, + { + "bbox": [ + 47, + 90, + 287, + 126 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 47, + 90, + 287, + 126 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 287, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 287, + 174 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {o v e r a l l}} = \\lambda_ {c l s} \\mathcal {L} _ {c l s} \\left(c _ {k}, \\bar {c} _ {\\hat {\\sigma} (k)}\\right) + \\lambda_ {b c e} \\mathcal {L} _ {b c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\hat {\\sigma} (k)}\\right) \\\\ + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} \\left(\\mathbf {m} _ {\\mathbf {k}}, \\hat {\\mathbf {m}} _ {\\tilde {\\sigma} (k)}\\right), \\tag {8} \\\\ \\end{array}", + "image_path": "aa01deb0c3576621a6bff0315af8924715252fe3e14d82cfe201d4a6a8de3284.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "content": " is the cross entropy loss, " + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bce}" + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "content": " is the binary cross entropy loss and " + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dice}" + }, + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "content": " is the dice loss [23]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 209, + 128, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 128, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 128, + 223 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 287, + 289 + ], + "type": "text", + "content": "In this section, we evaluated the proposed TCOVIS on four benchmark datasets. Furthermore, we provided in-depth ablation studies of the effectiveness of TCOVIS. Finally, we presented several visualizations of the predictions from our model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 297, + 110, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 110, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 110, + 308 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 316, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 287, + 363 + ], + "type": "text", + "content": "We evaluated our approach on four VIS datasets: YouTube-VIS 2019 dataset [37], YouTube-VIS 2021 dataset [37], YouTube-VIS 2022 dataset [37] and OVIS dataset [24]. We present a brief description of them:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 495 + ], + "type": "text", + "content": "YouTube-VIS 2019 & 2021 & 2022: YouTube-VIS 2019 [37] is the first VIS dataset and comprises 40 predefined categories of objects. This dataset included 4,883 unique video instances with 131,000 high-quality manual annotations. We followed the widely utilized training/test set split: 2,238 videos were selected for training, 302 videos were adopted for validation and 343 videos were used for testing. Further, YouTube-VIS 2021 improved the 40-category label set and added 4883 more unique video instances. Then, YouTube-VIS 2022 contained 71 additional long evaluation videos on the top of YouTube-VIS 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 495, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 578 + ], + "type": "text", + "content": "OVIS: Occluded video instance segmentation (OVIS) is also a challenging VIS dataset. OVIS included 901 videos in total with 25 semantic categories. This dataset contained 5,223 unique video instances and we followed the widely utilized training/test set split: 607 videos were selected for training, 607 videos were used for validation and 154 videos were adopted for testing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 579, + 287, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 287, + 614 + ], + "type": "text", + "content": "Following [37], the video-level average precision (AP) and average recall (AR) were adopted as the evaluation metrics on both YouTube-VIS and OVIS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 623, + 180, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 180, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 180, + 635 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 713 + ], + "type": "text", + "content": "We adopted the framework of GenVIS [13] which is built on VITA [14], but left out the similarity loss in VITA and the memory module in GenVIS. With the global assignment, the total loss as well the hyper-parameters were set the same as the video-level loss in VITA, which is a temporally extended loss function [16]. The model was trained" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "with pseudo-videos from COCO images [20] as data augmentation, and with a batch size of 8 video clips of 6 frames. As we froze the backbone and image segmentation model, all experiments were conducted with 8 RTX 2080 Ti GPUs. The method was implemented on detectron2 [36]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 140, + 391, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 140, + 391, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 140, + 391, + 152 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 159, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 159, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 159, + 545, + 194 + ], + "type": "text", + "content": "Following the standard evaluation metrics [37], we compared TCOVIS with state-of-the-art approaches on four VIS benchmarks: YouTube-VIS 2019/2021/2022 and OVIS." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 195, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 545, + 302 + ], + "type": "text", + "content": "YouTube-VIS 2019&2021. From Table 1, we can observe that TCOVIS has achieved very competitive performance using both lightweight backbones (ResNet-50) and powerful ones (Swin-L). Moreover, our method can even show better performance than offline methods, such as VITA [14] and offline version GenVIS [13]. On the more difficult YouTube-VIS 2021 dataset, TCOVIS surpasses GenVIS [13] in AP not only with ResNet-50 by 2.4 but also with Swin-L by 1.7." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 303, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 386 + ], + "type": "text", + "content": "YouTube-VIS 2022. As shown in Table 2, TCOVIS performed best in both AP and AR on YouTube-VIS 2022 dataset, which is more challenging than 2019&2021 datasets. Especially with powerful backbone Swin-L, TCOVIS outperformed GenVIS [13] in AP with a huge margin of 4.9, which shows the effectiveness of our method for the complex scenarios." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 387, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 471 + ], + "type": "text", + "content": "OVIS. Table 3 presents the comparisons on OVIS dataset, and we can find that TCOVIS also achieved the best 46.7 AP and 19.1 AR with Swin-L backbone. The results demonstrate that TCOVIS can deal with complicated situations where objects are heavily occluded in others. With ResNet-50 backbone, the performance is still the second best compared to the previous state-of-the-art methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 479, + 400, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 400, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 400, + 491 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 545 + ], + "type": "text", + "content": "In this section, we provided ablation studies and discuss the effects of different settings in the proposed method. The experiments are conducted with a ResNet-50 [12] backbone on YouTube-VIS 2019/2021 [37] valid set." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness of the proposed assignment strategy and enhancement module. The ablation studies on the global instance assignment strategy and the spatio-temporal enhancement module are shown in Table 4. As for the assignment strategy, compared to LocPro as the baseline, the model with the proposed global assignment outperformed the baseline model by more than 0.8 on AP, " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " on YouTube-VIS 2019, and more than 1.2 on YouTube-VIS 2021. The consistently significant improvement indicates that the strategy with global optimal matching contributes to better overall segmentation, while LocPro only considers the local optimal results. Besides, the baseline forces the posterior features to conform to those at the very beginning leading to temporal inconsistency, since the accumulative" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1102" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 531, + 361 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 531, + 361 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 531, + 361 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 531, + 361 + ], + "type": "table", + "html": "
MethodTypeYouTube-VIS 2019YouTube-VIS 2021
AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AR_1\\)\\(AR_{10}\\)AP\\(AP_{50}\\)\\(AP_{75}\\)\\(AR_1\\)\\(AR_{10}\\)
ResNet-50EfficientVIS [35]Offline37.959.743.040.346.634.057.537.333.842.5
IFC [16]Offline41.265.144.642.349.635.255.937.732.642.9
Mask2Former-VIS [4]Offline46.468.050.0--40.660.941.8--
\\(TeViT^{\\dagger}\\)[39]Offline46.671.351.644.954.337.961.242.135.144.6
SeqFormer [33]Offline47.469.851.845.554.840.562.443.736.148.1
VITA [14]Offline49.872.654.549.461.045.767.449.540.953.6
\\(GenVIS_{semi-online}\\)[13]Offline51.372.057.849.560.046.367.050.240.653.2
Cross VIS [38]Online36.356.838.935.640.734.254.437.930.438.2
VISOLO [9]Online38.656.343.735.742.536.954.740.230.640.9
MinVIS [15]Online47.469.052.145.755.744.266.048.139.251.7
IDOL [34]Online49.574.052.947.758.743.968.049.638.050.9
\\(GenVIS_{online}\\)[13]Online50.071.554.649.559.747.167.551.541.654.7
TCOVISOnline52.373.557.649.860.249.571.253.841.355.9
Swin-LSeqFormer [33]Offline59.382.166.451.764.451.874.658.242.858.1
Mask2Former-VIS [4]Offline60.484.467.0--52.676.457.2--
VITA [14]Offline63.086.967.956.368.157.580.661.047.762.6
\\(GenVIS_{semi-online}\\)[13]Offline63.885.768.556.368.460.180.966.549.164.7
MinVIS [15]Online61.683.368.654.866.655.376.662.045.960.8
IDOL [34]Online64.387.571.055.669.156.180.863.545.060.1
\\(GenVIS_{online}\\)[13]Online64.084.968.356.169.459.680.965.848.765.0
TCOVISOnline64.186.669.555.869.061.382.968.048.665.1
", + "image_path": "a3afe1d7dfc1c1bdcd18c4f331fe65ae6c3ec71853e1b03e3ecf6c38176f6697.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 411, + 285, + 536 + ], + "blocks": [ + { + "bbox": [ + 47, + 366, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 47, + 366, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 545, + 399 + ], + "type": "text", + "content": "Table 1. Quantitative results on YouTube-VIS 2019 and 2021 validation sets. The results are respectively grouped by method types (Offline or Online) and backbone networks (ResNet-50 and Swin-L). We bold the best performance and underline the second. † denotes using MsgBoxIF [39] backbone which has a similar weight scale with ResNet-50." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 411, + 285, + 536 + ], + "lines": [ + { + "bbox": [ + 49, + 411, + 285, + 536 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 285, + 536 + ], + "type": "table", + "html": "
MethodTypeAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
ResNet-50VITA [14]Offline32.653.939.330.342.6
GenVIS [13]Offline37.258.542.933.240.4
MinVIS [15]Online23.347.919.320.228.0
GenVIS [13]Online37.561.641.532.642.2
TCOVISOnline38.659.441.632.846.7
Swin-LVITA* [14]Offline41.163.044.039.344.3
GenVIS [13]Offline44.369.944.939.948.4
MinVIS* [15]Online33.154.833.729.536.6
GenVIS [13]Online45.169.147.339.848.5
TCOVISOnline51.073.053.541.756.5
", + "image_path": "e905c7d1ec9ebf3ac5787ed27e00928460610c74835ab686f6f93ef9d21c8c58.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 590, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 638 + ], + "type": "text", + "content": "error impairs the model during the training process. Our proposed assignment strategy encourages the features of all time to fit the global optimal objective, which effectively enhances the temporal consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "content": "As for the effectiveness of the spatio-temporal enhancement module, the comparison is also shown in Table 4. The results show that the enhancement module brings improvements of 0.9 AP on YouTube-VIS 2019&2021, compared to using the proposed assignment strategy individually. In particular, the proposed module improved the performance" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 411, + 544, + 584 + ], + "blocks": [ + { + "bbox": [ + 47, + 541, + 286, + 574 + ], + "lines": [ + { + "bbox": [ + 47, + 541, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 286, + 574 + ], + "type": "text", + "content": "Table 2. Quantitative results on YouTube-VIS 2022 validation dataset. We bold the highest accuracy and underline the second. " + }, + { + "bbox": [ + 47, + 541, + 286, + 574 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 47, + 541, + 286, + 574 + ], + "type": "text", + "content": ": Reproduced by [13]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 411, + 544, + 584 + ], + "lines": [ + { + "bbox": [ + 308, + 411, + 544, + 584 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 544, + 584 + ], + "type": "table", + "html": "
MethodTypeAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
ResNet-50\\( TeViT^†[39] \\)Offline17.434.915.011.221.8
VITA [14]Offline19.641.217.411.726.0
GenVIS [13]Offline34.559.435.016.638.3
CrossVIS [38]Online14.932.712.110.319.8
VISOLO [9]Online15.331.013.811.121.7
MinVIS [15]Online25.045.524.013.929.7
IDOL [34]Online30.251.330.015.037.5
GenVIS [13]Online35.860.836.216.339.6
TCOVISOnline35.360.736.615.739.5
Swin-LVITA [14]Offline27.751.924.914.933.0
GenVIS [13]Offline45.469.247.818.949.0
MinVIS [15]Online39.461.541.318.143.3
IDOL [34]Online42.665.745.217.949.6
GenVIS [13]Online45.269.148.419.148.6
TCOVISOnline46.770.949.519.150.8
", + "image_path": "d86aed8c328db85a5ce46cf38bdad612b05bce1891794c08330277e4d3e0f8c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 590, + 545, + 624 + ], + "lines": [ + { + "bbox": [ + 305, + 590, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 590, + 545, + 624 + ], + "type": "text", + "content": "Table 3. Quantitative results on OVIS validation set. We bold the highest accuracy and underline the second. † denotes using MsgBoxT [39] backbone." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "by 1.2 and 1.7 in " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " on two datasets, respectively. The performance improvements demonstrate that the proposed module effectively captures the spatial information from the previous frame and aggregates the semantic and spatial features across time. As a result, the delicate spatio-temporal design further enhances the temporal consistency of fea" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1103" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 509, + 135 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 509, + 135 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 509, + 135 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 509, + 135 + ], + "type": "table", + "html": "
LocProGIASTEYouTube-VIS 2019YouTube-VIS 2021
AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
✓--50.671.556.147.658.247.468.152.139.853.1
-✓-51.472.357.049.559.948.669.553.640.655.0
-✓✓52.373.557.649.860.249.571.253.841.355.9
", + "image_path": "7964e104a624c7656c8dafa917f5b2db2b2e9ac04840fcb9019e39905dee4567.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 170, + 278, + 260 + ], + "blocks": [ + { + "bbox": [ + 46, + 140, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 46, + 140, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 547, + 163 + ], + "type": "text", + "content": "Table 4. Ablation study of the method for Global Instance Assignment strategy (GIA) and the Spatio-temporal Enhancement module (STE) on the YouTube-VIS 2019 / 2021 validation sets. LocPro denotes the Local Matching and Propagating technique." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 170, + 278, + 260 + ], + "lines": [ + { + "bbox": [ + 52, + 170, + 278, + 260 + ], + "spans": [ + { + "bbox": [ + 52, + 170, + 278, + 260 + ], + "type": "table", + "html": "
ArchitectureAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AR_1 \\)\\( AR_{10} \\)
Cross-Attn.49.571.253.841.355.9
Concat.47.968.652.741.455.1
Self-Cross48.270.051.841.054.1
Dec. Mod.48.569.753.940.654.7
Resp. Pos.48.870.352.541.555.5
Ins. Attn.49.270.653.540.955.1
", + "image_path": "1959e3d8b7ceee26ef5e77cbaf9acbe21a12a9800c058577c308524c1e08f7ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 264, + 287, + 287 + ], + "lines": [ + { + "bbox": [ + 46, + 264, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 287, + 287 + ], + "type": "text", + "content": "Table 5. Ablation study of the module manipulations of the Spatio-temporal Enhancement module on YouTube-VIS 2021." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 296, + 271, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 271, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 271, + 307 + ], + "type": "text", + "content": "tures for the online video instance segmentation model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 403 + ], + "type": "text", + "content": "Incorporating the assignment strategy and the enhancement module, our proposed method gained remarkable performance improvements of 1.6 AP on YouTube-VIS 2019 and 2.1 AP on YouTube-VIS 2021 over the LocPro baseline. In a nutshell, the experimental results indicate that the temporally consistent features learned by our proposed TCOVIS significantly boost the performances on online video instance segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 404, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 487 + ], + "type": "text", + "content": "Different manipulations of spatio-temporal enhancement. In Table 5, we compared our proposed spatiotemporal enhancement module with other optional manipulations. Cross-Attn. denotes our proposed manipulation following the standard multi-head cross-attention layer to decode the spatial features with shared positional embedding described in Section 3.3. Concat. indicates that we" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 173, + 547, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 547, + 281 + ], + "type": "text", + "content": "concatenated the corresponding spatial feature and prototype of an instance followed by an MLP to get the updated query. Self-Cross stands for adding an extra self-attention layer ahead of the cross-attention layer. Dec. Mod. is decoder modulation, reviewing the spatial feature for every Transformer decoder layer. Resp .Pos. denotes respective positional embeddings were used for the spatial feature and prototype when we performed cross-attention. Ins. Attn. indicates the instance-wise decoding in cross-attention layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 285, + 547, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 488 + ], + "type": "text", + "content": "As shown, compared to all its counterparts, our proposed manipulation achieved the best performance. The first three variants are related to the feature aggregation, where " + }, + { + "bbox": [ + 304, + 285, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\text{Concat}" + }, + { + "bbox": [ + 304, + 285, + 547, + 488 + ], + "type": "text", + "content": " is too naive to model the spatial and semantic features, while Self-Cross obscures the spatial feature. We inferred the performance decrease of Dec. Mod. comes from the information redundancy when aggregating them in every layer. The last two experiments studied the instance-wise correspondence. In the Resp. Pos. setting, explicit correspondence for spatial and semantic features of the same instance between frames is absent. Ins. Attn. only focuses on the instance itself across time and can slightly enhance the temporal consistency, however, neglecting the spatial features of other instances, the module fails to capture the spatial relationship among instances. The results confirm that the proposed manipulation effectively exploits the spatial information along the temporal dimension to enhance the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 63, + 504, + 302, + 674 + ], + "blocks": [ + { + "bbox": [ + 63, + 504, + 302, + 674 + ], + "lines": [ + { + "bbox": [ + 63, + 504, + 302, + 674 + ], + "spans": [ + { + "bbox": [ + 63, + 504, + 302, + 674 + ], + "type": "image", + "image_path": "e02a590bf053b004b385f4df18063bc998af74e6a36989d7d8e6857a0c3cebd4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 678, + 547, + 712 + ], + "lines": [ + { + "bbox": [ + 46, + 678, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 678, + 547, + 712 + ], + "type": "text", + "content": "Figure 3. Qualitative results of TCOVIS, compared with VITA [14] and GenVIS [13]. On the left are the predictions on YouTube-VIS 2019 [37] and on the right are on OVIS [24]. Objects displayed in the same color denote the same instance. Our TCOVIS shows more temporally consistent results in these challenging scenes, where there are occlusions of the instance itself or others (Best viewed in color)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 305, + 504, + 531, + 674 + ], + "blocks": [ + { + "bbox": [ + 305, + 504, + 531, + 674 + ], + "lines": [ + { + "bbox": [ + 305, + 504, + 531, + 674 + ], + "spans": [ + { + "bbox": [ + 305, + 504, + 531, + 674 + ], + "type": "image", + "image_path": "da32317aa43fe599f2c359f5c3d13a24bbf204d5c8a7ba75e4fba8512b47cdc6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1104" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 189, + 137 + ], + "blocks": [ + { + "bbox": [ + 59, + 99, + 74, + 110 + ], + "lines": [ + { + "bbox": [ + 59, + 99, + 74, + 110 + ], + "spans": [ + { + "bbox": [ + 59, + 99, + 74, + 110 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 71, + 189, + 137 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 189, + 137 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 189, + 137 + ], + "type": "image", + "image_path": "5cd4e3dded4c647a823986ce2103344de4ff575f6242de1dec23ce0ce7ec814d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 189, + 72, + 302, + 136 + ], + "blocks": [ + { + "bbox": [ + 189, + 72, + 302, + 136 + ], + "lines": [ + { + "bbox": [ + 189, + 72, + 302, + 136 + ], + "spans": [ + { + "bbox": [ + 189, + 72, + 302, + 136 + ], + "type": "image", + "image_path": "1b347761c3d667e652ae9bc45f12f280e35fb98a79414620aadb9107f8918332.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 303, + 72, + 416, + 136 + ], + "blocks": [ + { + "bbox": [ + 303, + 72, + 416, + 136 + ], + "lines": [ + { + "bbox": [ + 303, + 72, + 416, + 136 + ], + "spans": [ + { + "bbox": [ + 303, + 72, + 416, + 136 + ], + "type": "image", + "image_path": "c74aef73dd4df29b67930669ca80865f483e581e23c7211eb9eaa73a791936da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 417, + 72, + 529, + 136 + ], + "blocks": [ + { + "bbox": [ + 417, + 72, + 529, + 136 + ], + "lines": [ + { + "bbox": [ + 417, + 72, + 529, + 136 + ], + "spans": [ + { + "bbox": [ + 417, + 72, + 529, + 136 + ], + "type": "image", + "image_path": "577032a6dcc42c1025c906b976e3bc80e4d975493c51ec57cc3c7df87a057a31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 75, + 137, + 188, + 201 + ], + "blocks": [ + { + "bbox": [ + 59, + 163, + 74, + 175 + ], + "lines": [ + { + "bbox": [ + 59, + 163, + 74, + 175 + ], + "spans": [ + { + "bbox": [ + 59, + 163, + 74, + 175 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 137, + 188, + 201 + ], + "lines": [ + { + "bbox": [ + 75, + 137, + 188, + 201 + ], + "spans": [ + { + "bbox": [ + 75, + 137, + 188, + 201 + ], + "type": "image", + "image_path": "8e2a2751357596502033264904d4a95c3fbc66562c583f8246492ab9ad64813a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 189, + 137, + 302, + 200 + ], + "blocks": [ + { + "bbox": [ + 189, + 137, + 302, + 200 + ], + "lines": [ + { + "bbox": [ + 189, + 137, + 302, + 200 + ], + "spans": [ + { + "bbox": [ + 189, + 137, + 302, + 200 + ], + "type": "image", + "image_path": "9b1384c83a01b1d9d8eda31e6657b274e005e21aff0827d135f2457737b86cd1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 302, + 137, + 416, + 200 + ], + "blocks": [ + { + "bbox": [ + 302, + 137, + 416, + 200 + ], + "lines": [ + { + "bbox": [ + 302, + 137, + 416, + 200 + ], + "spans": [ + { + "bbox": [ + 302, + 137, + 416, + 200 + ], + "type": "image", + "image_path": "57ce6d247a3fb0d2e897c3cba32d1861e4c4e2d3fdd702f5642020ecf4234484.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 417, + 137, + 529, + 200 + ], + "blocks": [ + { + "bbox": [ + 417, + 137, + 529, + 200 + ], + "lines": [ + { + "bbox": [ + 417, + 137, + 529, + 200 + ], + "spans": [ + { + "bbox": [ + 417, + 137, + 529, + 200 + ], + "type": "image", + "image_path": "efaf9b8db636caf212f4e04c6d1752d443134b54528dbbb1303617a1b36bf705.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 76, + 201, + 189, + 265 + ], + "blocks": [ + { + "bbox": [ + 60, + 228, + 74, + 240 + ], + "lines": [ + { + "bbox": [ + 60, + 228, + 74, + 240 + ], + "spans": [ + { + "bbox": [ + 60, + 228, + 74, + 240 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 201, + 189, + 265 + ], + "lines": [ + { + "bbox": [ + 76, + 201, + 189, + 265 + ], + "spans": [ + { + "bbox": [ + 76, + 201, + 189, + 265 + ], + "type": "image", + "image_path": "de8517a798c3821fc06cf9feca1b680257d5d64cbb4da75e304319a8e102dac9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 189, + 201, + 302, + 265 + ], + "blocks": [ + { + "bbox": [ + 189, + 201, + 302, + 265 + ], + "lines": [ + { + "bbox": [ + 189, + 201, + 302, + 265 + ], + "spans": [ + { + "bbox": [ + 189, + 201, + 302, + 265 + ], + "type": "image", + "image_path": "0203e78a5af85d94493829199498ac92de7e787fc00c2efb26bcf7913ced7953.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 302, + 201, + 416, + 265 + ], + "blocks": [ + { + "bbox": [ + 302, + 201, + 416, + 265 + ], + "lines": [ + { + "bbox": [ + 302, + 201, + 416, + 265 + ], + "spans": [ + { + "bbox": [ + 302, + 201, + 416, + 265 + ], + "type": "image", + "image_path": "1675ef9e6fc7e01c827ec9b67d1817a9cf8ec1d3ac1a9e26bbdd59e7247836c5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 417, + 201, + 529, + 265 + ], + "blocks": [ + { + "bbox": [ + 417, + 201, + 529, + 265 + ], + "lines": [ + { + "bbox": [ + 417, + 201, + 529, + 265 + ], + "spans": [ + { + "bbox": [ + 417, + 201, + 529, + 265 + ], + "type": "image", + "image_path": "2b244cfd36a47a45a5006e60a192d09261b8ce75e94d65ef55e5130c4bfbf370.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 75, + 266, + 189, + 330 + ], + "blocks": [ + { + "bbox": [ + 59, + 293, + 74, + 304 + ], + "lines": [ + { + "bbox": [ + 59, + 293, + 74, + 304 + ], + "spans": [ + { + "bbox": [ + 59, + 293, + 74, + 304 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 266, + 189, + 330 + ], + "lines": [ + { + "bbox": [ + 75, + 266, + 189, + 330 + ], + "spans": [ + { + "bbox": [ + 75, + 266, + 189, + 330 + ], + "type": "image", + "image_path": "db4b9a94343f3c7d4ea8f63b0c91408935f2d0193180c3842369e57082145df4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 189, + 266, + 302, + 330 + ], + "blocks": [ + { + "bbox": [ + 189, + 266, + 302, + 330 + ], + "lines": [ + { + "bbox": [ + 189, + 266, + 302, + 330 + ], + "spans": [ + { + "bbox": [ + 189, + 266, + 302, + 330 + ], + "type": "image", + "image_path": "dee460dda09ee30d0615cca21ab36e6242da609cd16d2d5949c405246e06306b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 336, + 547, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 336, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 547, + 369 + ], + "type": "text", + "content": "Figure 4. Qualitative results of TCOVIS in four challenging cases on YouTube-VIS 2019 [37] and OVIS [24]: (a) fast movement, (b) heavy occlusion, (c) low resolution, and (d) crowded scene. Objects displayed in the same color denote the same instance. The qualitative results demonstrate the effectiveness and robustness of TCOVIS (Best viewed in color)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 302, + 266, + 416, + 330 + ], + "blocks": [ + { + "bbox": [ + 302, + 266, + 416, + 330 + ], + "lines": [ + { + "bbox": [ + 302, + 266, + 416, + 330 + ], + "spans": [ + { + "bbox": [ + 302, + 266, + 416, + 330 + ], + "type": "image", + "image_path": "72aadeac444e2f6c38758e6e3f9dc4f860c6209b04be8df8a459799051d17ab0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 417, + 266, + 529, + 330 + ], + "blocks": [ + { + "bbox": [ + 417, + 266, + 529, + 330 + ], + "lines": [ + { + "bbox": [ + 417, + 266, + 529, + 330 + ], + "spans": [ + { + "bbox": [ + 417, + 266, + 529, + 330 + ], + "type": "image", + "image_path": "09056a28bbe9e89a3c8e681e0643adbc5641a6b7c843d60db5b5fc42c08b8567.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 383, + 181, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 181, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 181, + 395 + ], + "type": "text", + "content": "temporal consistency of features." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 405, + 155, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 155, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 155, + 418 + ], + "type": "text", + "content": "4.5. Qualitative results" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 426, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 581 + ], + "type": "text", + "content": "In Figure 3, we show the qualitative comparisons of the proposed TCOVIS with VITA [14] and GenVIS [13] on YouTube-VIS 2019 and OVIS datasets. In the left scene where there is a zebra with self-occlusion, GenVIS fails to segment its head resulting in fragmented predictions, however, TCOVIS performs temporally consistent segmentation with impressive accuracy. On the right is a difficult case where there are two dogs with similar appearances grappling with each other. VITA incorrectly detects more than two dogs and fails to track the left one, while GenVIS fails to handle the margin of two instances, e.g. the nose and the paw of the left dog. TCOVIS successfully tracks and segments the instances, demonstrating its effectiveness." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "In Figure 4, we provide more qualitative results of the proposed method in variously challenging cases, which are all chosen from the mentioned benchmarks [37, 24]. As shown in Figure 4 (a), our method successfully tracks the surfer and the surfboard with fast movement. In the second row, we present a difficult case with severe occlusion, in which our method performs admirably by accurately segmenting the claw (at the bottom) of a parrot despite the presence of a wooden stick that partially obstructs its lower body. Figure 4 (c) illustrates a low-resolution case and our method still achieves good performance. In Figure 4 (d)," + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 383, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 383, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 545, + 443 + ], + "type": "text", + "content": "we depict a crowded scene where TCOVIS is capable of handling multiple instances that share similar appearances and exhibit complex interactions. All the qualitative results in the challenging situations demonstrate the effectiveness and robustness of the proposed method." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 483, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 625 + ], + "type": "text", + "content": "In this paper, we propose a new online video instance segmentation method, TCOVIS, to fully exploit the temporal information within a video and produce temporally consistent predictions. Based on the query propagation framework, we propose a global instance assignment strategy to perform global optimal matching with the consideration of the entire video and supervise the model with the global optimal objective. We further devise a spatio-temporal enhancement module to capture the spatial feature and aggregate it with the semantic feature between frames. The effectiveness of our method is evaluated with experimental results on YouTube-VIS 2019/2021/2022 and OVIS." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 643, + 403, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 403, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 403, + 657 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1105" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Ali Athar, Sabarinath Mahadevan, Aljosa Osep, Laura Leal-Taixe, and Bastian Leibe. Stem-seg: Spatio-temporal embeddings for instance segmentation in videos. In ECCV, pages 158-177, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "text", + "content": "[2] Gedas Bertasius and Lorenzo Torresani. Classifying, segmenting, and tracking object instances in video with mask propagation. In CVPR, pages 9739-9748, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 214 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 259 + ], + "type": "text", + "content": "[4] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 305 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 372 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 374, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 287, + 407 + ], + "type": "text", + "content": "[7] Yuxin Fang, Shusheng Yang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Instances as queries. In ICCV, pages 6910-6919, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 409, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 452 + ], + "type": "text", + "content": "[8] Yang Fu, Linjie Yang, Ding Liu, Thomas S Huang, and Humphrey Shi. Compfeat: Comprehensive feature aggregation for video instance segmentation. In AAAI, pages 1361-1369, 2021. 1, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 508 + ], + "type": "text", + "content": "[9] Su Ho Han, Sukjun Hwang, Seoung Wug Oh, Yeonschool Park, Hyunwoo Kim, Min-Jung Kim, and Seon Joo Kim. Visolo: Grid-based space-time aggregation for efficient online video instance segmentation. In CVPR, pages 2896-2905, 2022. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[10] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. In NeurIPS, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 586 + ], + "type": "text", + "content": "[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[13] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A generalized framework for video instance segmentation. In CVPR, pages 14623–14632, 2023. 1, 2, 3, 4, 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[14] Miran Heo, Sukjun Hwang, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. VITA: Video instance segmentation via object token association. In NeurIPS, 2022. 1, 2, 3, 6, 7, 8, 9" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] De-An Huang, Zhiding Yu, and Anima Anandkumar. MinVIS: A minimal video instance segmentation framework without video-based training. In NeurIPS, 2022. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[16] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. NeurIPS, 34, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "text", + "content": "[17] Lei Ke, Xia Li, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Prototypical cross-attention networks for multiple object tracking and segmentation. NeurIPS, 34, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 217 + ], + "type": "text", + "content": "[18] Harold W Kuhn. The hungarian method for the assignment problem. NRL, 1955. 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "text", + "content": "[19] Huaijia Lin, Ruizheng Wu, Shu Liu, Jiangbo Lu, and Jiaya Jia. Video instance segmentation with a propose-reduce paradigm. In ICCV, pages 1739-1748, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "text", + "content": "[21] Dongfang Liu, Yiming Cui, Wenbo Tan, and Yingjie Chen. Sg-net: Spatial granularity network for one-stage video instance segmentation. In CVPR, pages 9816-9825, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 333, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 374 + ], + "type": "text", + "content": "[22] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, pages 8844-8854, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "type": "text", + "content": "[23] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 3DV, pages 565-571, 2016. 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "text", + "content": "[24] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. IJCV, 130(8):2022-2039, 2022. 2, 6, 8, 9" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 468, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 520 + ], + "type": "text", + "content": "[25] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In CVPR, pages 14454-14463, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[26] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, pages 282-298, 2020. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 557, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 599 + ], + "type": "text", + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[28] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[29] Tao Wang, Ning Xu, Kean Chen, and Weiyao Lin. End-to-end video instance segmentation via spatial-temporal graph neural networks. In ICCV, pages 10797-10806, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[30] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, pages 649-665, 2020. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1106" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 555 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In CVPR, pages 8741-8750, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 160 + ], + "type": "text", + "content": "[32] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In CVPR, pages 12352-12361, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 196 + ], + "type": "text", + "content": "[33] Junfeng Wu, Yi Jiang, Song Bai, Wenqing Zhang, and Xiang Bai. Seqformer: Sequential transformer for video instance segmentation. In ECCV, pages 553-569, 2022. 1, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 198, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 286, + 239 + ], + "type": "text", + "content": "[34] Junfeng Wu, Qihao Liu, Yi Jiang, Song Bai, Alan Yuille, and Xiang Bai. In defense of online models for video instance segmentation. In ECCV, pages 588-605, 2022. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 285 + ], + "type": "text", + "content": "[35] Jialian Wu, Sudhir Yarram, Hui Liang, Tian Lan, Junsong Yuan, Jayan Eledath, and Gerard Medioni. Efficient video instance segmentation via tracklet query and proposal. In CVPR, pages 959-968, 2022. 1, 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 286, + 319 + ], + "type": "text", + "content": "[36] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 320, + 286, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 286, + 351 + ], + "type": "text", + "content": "[37] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, pages 5188-5197, 2019. 1, 2, 3, 6, 8, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 354, + 286, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 286, + 397 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 286, + 397 + ], + "type": "text", + "content": "[38] Shusheng Yang, Yuxin Fang, Xinggang Wang, Yu Li, Chen Fang, Ying Shan, Bin Feng, and Wenyu Liu. Crossover learning for fast online video instance segmentation. In ICCV, pages 8043-8052, 2021. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 399, + 286, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 286, + 442 + ], + "type": "text", + "content": "[39] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Wenyu Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In CVPR, pages 2885-2895, 2022. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 444, + 286, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 444, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 444, + 286, + 487 + ], + "type": "text", + "content": "[40] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. In ECCV, pages 659-675. Springer, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 489, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 489, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 49, + 489, + 286, + 520 + ], + "type": "text", + "content": "[41] Zitong Zhan, Daniel McKee, and Svetlana Lazebnik. Robust online video instance segmentation with track queries. arXiv preprint arXiv:2211.09108, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 522, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 522, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 522, + 286, + 555 + ], + "type": "text", + "content": "[42] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable " + }, + { + "bbox": [ + 49, + 522, + 286, + 555 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{detr}\\}" + }, + { + "bbox": [ + 49, + 522, + 286, + 555 + ], + "type": "text", + "content": ": Deformable transformers for end-to-end object detection. In ICLR, 2021. 2, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1107" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_content_list.json b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..221b87c80cbd74df9fabd9b71b299c2b2644ccfc --- /dev/null +++ b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_content_list.json @@ -0,0 +1,1650 @@ +[ + { + "type": "text", + "text": "TEMPO: Efficient Multi-View Pose Estimation, Tracking, and Forecasting", + "text_level": 1, + "bbox": [ + 109, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rohan Choudhury Kris M. Kitani Laszlo A. Jeni Robotics Institute, Carnegie Mellon University", + "bbox": [ + 271, + 181, + 696, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{rchoudhu, kmkitani}@andrew.cmu.edu laszlojeni@cmu.edu", + "bbox": [ + 238, + 219, + 730, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/488d387e29e768b2ba328eb10c29fbb64cdc06680d2eb97cc21bf19ba276dd03.jpg", + "image_caption": [ + "Camera 1" + ], + "image_footnote": [], + "bbox": [ + 94, + 272, + 289, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e90a313eeaf7681e26087e5f2a4c607248cb8e0f6e0d62ad3eac565226b4798a.jpg", + "image_caption": [ + "Camera 2", + "Figure 1: We propose TEMPO: TEMporal POse Estimation, a method for multi-view, multi-person pose estimation, tracking and forecasting. TEMPO uses a recurrent architecture to learn a spatiotemporal representation, significantly improving the pose estimation accuracy while preserving speed at inference time. In each view, we show TEMPO's predicted pose skeletons over 10 frames, colored by tracker identity. Lighter colors correspond to previous frames." + ], + "image_footnote": [], + "bbox": [ + 295, + 272, + 491, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a6f7b555d29676ae93669d76d60f5e2dfe662c93e006896cb1fad21fec23ac3e.jpg", + "image_caption": [ + "Camera 3" + ], + "image_footnote": [], + "bbox": [ + 493, + 272, + 689, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/369d91a9d8e0d389f442d4cefe1ed4a086d1db631dddf1f66bfc23d888e95993.jpg", + "image_caption": [ + "3D Pose" + ], + "image_footnote": [], + "bbox": [ + 705, + 272, + 875, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 505, + 313, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing volumetric methods for predicting 3D human pose estimation are accurate, but computationally expensive and optimized for single time-step prediction. We present TEMPO, an efficient multi-view pose estimation model that learns a robust spatiotemporal representation, improving pose accuracy while also tracking and forecasting human pose. We significantly reduce computation compared to the state-of-the-art by recurrently computing per-person 2D pose features, fusing both spatial and temporal information into a single representation. In doing so, our model is able to use spatiotemporal context to predict more accurate human poses without sacrificing efficiency. We further use this representation to track human poses over time as well as predict future poses. Finally, we demonstrate that our model is able to generalize across datasets without scene-specific fine-tuning. TEMPO achieves $10\\%$ better MPJPE with a $33\\times$ improvement in FPS compared to TesseTrack on the challenging CMU Panoptic Studio dataset. Our code and demos are available at https://rccchoudhury.github.io/tempo2023/.", + "bbox": [ + 73, + 518, + 473, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 843, + 209, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating the pose of several people from multiple overlapping cameras is a crucial vision problem. Volumet-", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ric multi-view methods, which lift 2D image features from each camera view to a feature volume then regress 3D pose, are currently the state of the art [47, 42, 53, 23] in this task. These approaches produce significantly more accurate poses than geometric alternatives, but suffer from two key limitations. First, the most accurate methods use either 3D convolutions [47, 42, 55] or cross-view transformers [49] which are slow and prevent real-time inference. Secondly, most methods are designed for estimating pose at a single timestep and are unable to reason over time, limiting their accuracy and preventing their use for tasks like motion prediction.", + "bbox": [ + 496, + 505, + 893, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose TEMPO, a multi-view TEMporal POse estimation method that addresses both of these issues. TEMPO uses temporal context from previous timesteps to produce smoother and more accurate pose estimates. Our model tracks people over time, predicts future pose and runs efficiently, achieving near real-time performance on existing benchmarks. The key insight behind TEMPO, inspired by work in 3D object detection [31, 20], is that recurrently aggregating spatiotemporal context results in powerful learned representations while being computationally efficient. To do this, we decompose the problem into three stages, illustrated in Figure 2. Given an input RGB video from multiple static, calibrated cameras, at a given timestep $t$ we first detect the locations of each person in the scene by unproject", + "bbox": [ + 496, + 688, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14750", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing image features from each view to a common 3D volume. We then regress 3D bounding boxes centered on each person, and perform tracking by matching the box centers with the detections from the previous timestep $t - 1$ . For each detected person, we compute a spatiotemporal pose representation by recurrently combining features from current and previous timesteps. We then decode the representation into an estimate of the current pose as well as poses at future timesteps. Unlike existing work [47, 53, 42, 55], our method is able to perform temporal tasks like tracking and forecasting without sacrificing efficiency.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method on several pose estimation benchmarks. TEMPO achieves state of the art results on the challenging CMU Panoptic Studio dataset [25] by $10\\%$ , and is competitive on the Campus, Shelf and Human3.6M datasets. We additionally collect our own multiview dataset consisting of highly dynamic scenes, on which TEMPO achieves the best result by a large margin. We show that our model achieves competitive results in pose tracking and evaluate the pose forecasting quality on the CMU Panoptic dataset. Additionally, multi-view pose estimation methods are almost always evaluated on the same dataset they are trained on, leading to results that are specific to certain scenes and camera configurations. We measure our method's ability to generalize across different datasets and find that our method can transfer without additional fine tuning. To summarize, our key contributions are that:", + "bbox": [ + 75, + 257, + 472, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop the most accurate multi-view, multiperson 3D human pose estimation model. Our model uses temporal context to produce smoother and more accurate poses.", + "- Our model runs efficiently with no performance degradation.", + "- Our model tracks and forecasts human pose for every person in the scene.", + "- We evaluate the generalization of our model across multiple datasets and camera configurations." + ], + "bbox": [ + 96, + 523, + 467, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 693, + 218, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Pose Estimation, Tracking and Forecasting Approaches for recovering and tracking 3D human pose are usually limited to monocular video. Such methods for pose estimation [46, 5, 27, 29] and pose tracking [41, 40] are highly efficient, but perform significantly worse than multi-view methods in 3D pose estimation accuracy due to the inherent ambiguity of monocular input.", + "bbox": [ + 75, + 719, + 468, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, methods in human pose forecasting [34] usually predict future motion from ground truth pose histories. Our approach follows [8, 56] and predicts pose directly from a sequence of video frames. Snipper [56] is the closest to our method and uses a spatiotemporal transformer to", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "jointly estimate, track and forecast pose from a monocular video. Our method differs in that it is able to produce highly accurate estimates using multi-view information while running efficiently.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-View Pose Estimation Early work in multi-view human pose estimation was limited to the single-person case, with [3, 30, 37, 1] using pictorial structure models to improve over basic triangulation. More recent approaches [19, 39, 23, 7, 26] improve this result by using advanced deep architectures like 3D CNNs and transformers, and others [9, 10] introduce priors on human shape for additional performance. Our method is most similar to [23], which uses 3D CNNs to regress pose directly from a feature volume. We also follow [23, 26] in analyzing our model's transfer to multiple datasets, extending their qualitative, single-person analysis to a quantitative measurement of performance on several multi-person datasets.", + "bbox": [ + 496, + 154, + 892, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the multi-person setting, early approaches like [3, 12, 13, 52, 6] associate 2D pose estimates from each view, then fuse the matched 2D poses into 3D. Other methods aimed towards multi-person motion capture use Re-ID features [12, 52], 4D graph cuts [54], plane sweep stereo [32], cross-view graph matching [50], or optimize SMPL parameters [14] to produce 3D poses from 2D pose estimates in each view. These methods can generalize across the data sources, but typically have much less accurate predictions compared to volumetric methods. These first unproject learned 2D image features into a 3D volume and regress pose directly from the 3D features with neural networks. Both [23] and [47] use computationally expensive 3D CNNs for the pose estimation step. Follow up work includes Faster VoxelPose [53] which replaces these 3D CNNs with 2D CNNs for a large speedup, and TesseTrack [42], which uses 4D CNNs to reason over multiple timesteps. Our method combines the best of both: we efficiently incorporate spatiotemporal information with only 2D CNNs and a lightweight recurrent network.", + "bbox": [ + 496, + 353, + 893, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Object Detection and Forecasting Historically, 3D object detection and instance segmentation methods for autonomous driving have led development in using multi-view images. One key similarity to our work is the aggregation of 2D image features into a single 3D volume. While [44, 48] use the same bilinear unprojection strategy as our method, several works [38, 17, 31] propose alternatives such as predicting per-pixel depth. Other works also use temporal information for detection for tracking objects through occlusion and spotting hard-to-see objects; [20, 31, 36] concretely demonstrate the benefits of incorporating spatiotemporal context. In particular, FIERY [20] uses temporal information for future instance prediction and BEVFormer [31] efficiently aggregates temporal information with a recurrent architecture, both of which inspired our method. Furthermore, [24, 16] use per-timestep supervision to track", + "bbox": [ + 496, + 657, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14751", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/381410b12e42bd5fdbaedc05dd51fe3e6a1a9ce614aa5acb2dec1525b23ab209.jpg", + "image_caption": [ + "Figure 2: The overall model architecture. We begin by (1) extracting features from each image with the backbone network and unprojecting those features to a 3D volume. In step (2), we use the volume to detect each person in the scene, and (3) associate the detections from the current timestep to the previous one. We then (4) fuse the features from each person with our temporal model and produce a final pose estimate." + ], + "image_footnote": [], + "bbox": [ + 116, + 89, + 872, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pixels through occlusion, an idea which we adapt for reasoning about human pose over multiple frames.", + "bbox": [ + 75, + 474, + 470, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 516, + 168, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method assumes access to calibrated timesynchronized videos from one or more cameras. At training time, we assume access to $T$ sets of $N$ RGB images from different cameras, while at inference time, we have a single set of $N$ images corresponding to the current timestep. In order to enable TEMPO to transfer to new camera configurations and settings, we compute the dimensions of the space and size of the voxel volume directly from the camera matrices. We set the height of the volume to a constant $2\\mathrm{m}$ , while setting the length and width of the volume to be the bounding box of the camera extrinsics from a top-down view, and center the space at the mean of the camera locations.", + "bbox": [ + 75, + 541, + 468, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 744, + 218, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We briefly review the person detection and pose estimation modules used by VoxelPose [47], Tessetrack [42], and Faster VoxelPose [53] that TEMPO builds upon. We refer the reader to the original papers for further detail.", + "bbox": [ + 75, + 768, + 468, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Person Detection", + "text_level": 1, + "bbox": [ + 76, + 845, + 250, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The detection module aims to estimate the location of the root joint as well as tight 3D bounding box for each person", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in the scene. Following previous work, we define the root joint as the mid-hip. At a given time $t$ , the detector module takes as input a set of $N$ images, each corresponding to a different camera view of the same scene at time $t$ . For each image, we extract features with a pretrained backbone, resulting in $N$ feature maps $\\mathbf{F}_1^t, \\mathbf{F}_2^t, \\dots, \\mathbf{F}_N^t$ .", + "bbox": [ + 496, + 474, + 890, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the camera matrices for each view $\\mathbf{C}_1^t,\\mathbf{C}_2^t,\\ldots ,\\mathbf{C}_N^t$ , we use the bilinear sampling procedure from [23, 47, 17]. For a voxel $v\\in V$ with coordinates $\\mathbf{x}$ , we have", + "bbox": [ + 496, + 565, + 890, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nv = \\sum_ {i = 1} ^ {N} \\mathbf {F} _ {i} ^ {t} \\left(\\mathbf {C} _ {i} \\mathbf {x}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 625, + 890, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_i^t (\\mathbf{x})$ is the feature map $\\mathbf{F}_i^t$ 's value at position $\\mathbf{x}$ , obtained by bilinear sampling. We then compute a birds-eye view representation of $V$ by taking the maximum along the $z$ -axis:", + "bbox": [ + 496, + 672, + 890, + 733 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathrm {B E V}} ^ {t} = \\max _ {z} V \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 734, + 890, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use a 2D CNN to produce a 2D heatmap of $\\mathbf{H}^t$ from $\\mathbf{F}_{\\mathrm{BEV}}^t$ of the $(x,y)$ locations of every root joint in the scene. We then sample the top $K$ locations from $\\mathbf{H}^t$ , yielding proposals $(x_1,y_1),(x_2,y_2),\\ldots (x_K,y_K)$ . For each proposal location, we obtain the corresponding feature column $V|_{x,y}$ and apply a 1D CNN to regress a 1D heatmap of the root joint's height, denoted $\\mathbf{H}_k^t$ . We then sample the maximum $z$ coordinate for each $\\mathbf{H}_k^t$ , and combine these to produce a set of detections $D_{t} = \\{(x_{1},y_{1},z_{1}),\\dots (x_{K},y_{k},z_{K})\\}$ . Fi", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14752", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nally, we regress width, length and centerness from $\\mathbf{F}_{\\mathrm{BEV}}^t$ with a multi-headed 2D CNN to produce bounding box predictions for each proposal. The loss function for the detection module has three terms. First, $L_{2D}$ is the distance between the predicted 2D heatmap and the ground truth, given by", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {2 D} = \\sum_ {t = 1} ^ {T} \\sum_ {(x, y))} \\| \\mathbf {H} ^ {t} (x, y) - \\mathbf {H} _ {G T} ^ {t} (x, y) \\| \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 196, + 468, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also compute the loss on the predicted 1D heatmap:", + "bbox": [ + 76, + 248, + 444, + 263 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {1 D} = \\sum_ {t = 1} ^ {T} \\sum_ {k = 1} ^ {K} \\sum_ {z} \\left\\| \\mathbf {H} _ {k} ^ {t} (z) - \\mathbf {H} _ {k, G T} ^ {t} (z) \\right\\| \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 276, + 468, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we include the bounding box regression loss", + "bbox": [ + 76, + 330, + 424, + 345 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {b b o x}} = \\sum_ {t = 1} ^ {T} \\sum_ {(i, j) \\in U} \\| \\mathbf {S} (i, j) - \\mathbf {S} _ {G T} (i, j) \\| _ {1} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 357, + 468, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The total detection loss is the sum of the above terms, with $L_{det} = L_{2D} + L_{1D} + L_{\\mathrm{bbox}}$ .", + "bbox": [ + 76, + 414, + 468, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Instantaneous Pose Estimation", + "text_level": 1, + "bbox": [ + 76, + 467, + 344, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each detection $D$ , we construct a volume of fixed size centered on the detection's center $c_{i}$ , and unproject the backbone features from each camera view into the volume, as in the detection step. As in [53], we mask out all features falling outside the detection's associated bounding box $B_{i}$ , resulting in a feature volume $V_{i}^{t}$ for person $i$ .", + "bbox": [ + 75, + 492, + 468, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 3, we project the feature volume $V_{t}^{i}$ to 2D along each of the $xy, yz$ , and $xz$ planes, resulting in three 2D feature maps, denoted by $\\mathbf{P}_{i,xy}^{t}$ , $\\mathbf{P}_{i,xz}^{t}$ , and $\\mathbf{P}_{i,yz}^{t}$ . The intuition behind this step is that we can predict the 2D position of each joint in each plane, and fuse the predicted 2D positions back together to form a 3D skeleton. Each feature map is passed through a 2D CNN to decode a heatmap of joint likelihood for every person joint, in each of the three planes, and the 2D joint predictions from each plane are fused into 3D with a learned weighting network. We define both the loss for a predicted pose as the mean squared loss between the computed and ground truth 2D heatmaps, as well as the $L_{1}$ loss of the predicted joint locations and ground truth:", + "bbox": [ + 75, + 584, + 468, + 795 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {j o i n t, t} ^ {k} = \\mathrm {M S E} (\\mathbf {P} _ {x y} ^ {t}, \\mathbf {P} _ {x y, G T} ^ {t}) + \\mathrm {M S E} (\\mathbf {P} _ {x z} ^ {t}, \\mathbf {P} _ {x z, G T} ^ {t}) + \\\\ \\operatorname {M S E} \\left(\\mathbf {P} _ {y z} ^ {t}, \\mathbf {P} _ {y z, G T} ^ {t}\\right) + \\sum_ {j = 1} ^ {J} \\left| j _ {i, p r e d} - j _ {i, G T} \\right| \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 806, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with MSE representing the mean squared error.", + "bbox": [ + 76, + 885, + 390, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Person Tracking", + "text_level": 1, + "bbox": [ + 500, + 90, + 663, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now describe how TEMPO uses temporal information. Unlike previous works, TEMPO takes as input a set of $N$ images at time $t$ as well as the person detections $D_{t-1}$ and corresponding 2D pose embeddings $P^{t-1}$ from the previous timestep.", + "bbox": [ + 500, + 116, + 890, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each proposed detection from the previous step consists of a body center $c_{i}$ and a bounding box $B_{i}^{t} = (h_{i}^{t},w_{i}^{t})$ . Given $K$ proposals, we compute a $K\\times K$ cost matrix $\\mathbf{A}$ based on the IoU between $B_{i}^{t}$ and $B_{k}^{t}$ for all detections at time $t$ , resulting in", + "bbox": [ + 498, + 194, + 890, + 270 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} [ i ] [ j ] = \\left\\| c _ {i} - c _ {j} \\right\\| \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 287, + 890, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $c_{i}, c_{j}$ being the associated predicted locations of person $i$ and person $j$ 's root joint.", + "bbox": [ + 498, + 321, + 888, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While VoxelTrack [55] computes cost across every joint in each pose, TEMPO uses the top-down view of the associated bounding box for each person. At inference time, we use the SORT [4] tracker, which is fast and uses a simple Kalman filter with no learned Re-ID mechanism. While [42] uses a learned tracker based on SuperGlue [43], we find that SORT is faster and does not result in degraded performance.", + "bbox": [ + 498, + 353, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Temporal Pose Estimation and Forecasting", + "text_level": 1, + "bbox": [ + 500, + 492, + 864, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After running the tracker, the input to the pose estimation stage is a set of detection proposals $D^{t}$ , the previous timestep's detection proposals $D^{t-1}$ , and the assignment matrix between the detection sets $\\mathbf{A}^{t}$ . We also assume access to pose features $\\mathbf{P}_{i}^{t-1}$ for each person in the previous timestep. Both $D^{t}$ and $D^{t-1}$ have $K$ proposals each.", + "bbox": [ + 498, + 517, + 890, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, $\\mathbf{P}_i^{t - 1}$ and $\\mathbf{P}_i^t$ are centered at $c_{i}^{t - 1}$ and $c_{i}^{t}$ respectively, and thus the pose features from each volume are not in the same coordinate system due to the motion of person $i$ . To fix this, we follow the standard procedure used in temporal birds-eye view prediction [31, 20] and warp the projected features from $\\mathbf{F}_i^{t - 1}$ into the coordinate system of $\\mathbf{P}_i^t$ with a translational warp defined by $c_{i}^{t} - c_{i}^{t - 1}$ .", + "bbox": [ + 498, + 609, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After warping the previous pose features, we run a recurrent network with Spatial Gated Recurrent Units [2] (SpatialGRUs) to produce multiple embeddings: $\\mathbf{F}_i^t$ , representing the current pose, and $\\mathbf{F}_i^{t + 1},\\mathbf{F}_i^{t + 2},\\ldots$ , representing the pose in future frames. While [20] and [31] do not propagate gradients through time and only predict object locations and instance segments at time $T$ , we do backpropagate through time by predicting the pose for each person at every timestep. At training time, we recurrently compute the temporal representation at each timestep $t_0,t_0 + 1,\\dots t_0 + T$ decode a pose for every timestep, and compute losses over all the predicted poses simultaneously. Thus, the final train", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14753", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0677ed1d250ab2f50cccc1e9873d5d000c1a39b0351c3e5cca57fcc2d8fb15df.jpg", + "image_caption": [ + "Figure 3: A closer look at the temporal representation used by our model. Following [53], we first project the feature volume to each of the three planes, and concatenate the projections channel-wise. We pass this feature map through an encoder network. We use this feature encoding as input to the SpatialGRU module, using the spatially warped pose feature from the previous timestep as a hidden state. We use the SpatialGRU module to produce features at the current and future timesteps, which we decode into human poses with the pose decoder network." + ], + "image_footnote": [], + "bbox": [ + 107, + 90, + 880, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing objective is", + "bbox": [ + 76, + 452, + 183, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {p o s e}} = \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {K} L _ {\\text {j o i n t}, t} ^ {i} + L _ {\\text {j o i n t}, t + 1} ^ {i} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 474, + 468, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{joint},t}^{i}$ is the L1 distance between the predictde and ground truth pose at time $t$ Providing supervision to the network at every timestep allows the network to learn a representation that encodes the motion between consecutive frames while enabling temporal smoothness between predictions. As we show in Section 4.5, this technique is crucial to our model's performance.", + "bbox": [ + 75, + 523, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While training, we run the network $T$ times, for each input timestep. However, at inference time, we save the previous embeddings and only receive input images at a single timestep $t$ , significantly reducing the computational burden and allowing our model to use temporal information without sacrificing efficiency.", + "bbox": [ + 75, + 628, + 468, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 746, + 209, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 76, + 771, + 276, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Panoptic Studio The CMU Panoptic Studio dataset [25] is a large multi-view pose estimation dataset with several synchronized camera sequences of multiple interacting subjects. Following prior work [47, 53, 49, 32], we use five HD cameras, specifically cameras 3, 6, 12, 13, and 23. We also use the same training and test split as these works, omitting the sequence 160906_band3 due to data corruption.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human 3.6M The Human3.6M dataset [22, 21] consists of videos of a single subject in an indoor studio with four static cameras. Each video has a professional actor performing a specific action. We follow the training-test split of prior works, using subjects 9 and 11 for validation and the others for training, while omitting corrupted sequences.", + "bbox": [ + 496, + 452, + 890, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Campus and Shelf The Campus and Shelf datasets[3] contain approximately 4000 frames of a single scene. While these datasets are commonly used for benchmarking in previous work, they are missing many annotations. We follow previous work [47, 53, 49] and adopt a synthetic heatmap-based scheme for training.", + "bbox": [ + 496, + 546, + 890, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EgoHumans We include the newly collected EgoHumans multi-view dataset [28]. This benchmark consists of approximately 1 hour of video of up to 5 subjects performing highly dynamic activities, such as playing tag, fencing, or group assembly. It contains videos from up to eight fisheye cameras and includes both egocentric and exocentric camera and pose data.", + "bbox": [ + 496, + 640, + 890, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics For the Panoptic Studio, Human3.6M, and Ego-Humans datasets, we report the mean per-joint position error (MPJPE). We additionally report the Average Precision $(\\mathrm{AP}_K)$ on the Panoptic and EgoHumans datasets and report MPJPE for Human3.6M in line with previous work [37, 35, 23, 42]. On the Shelf and Campus dataset we report the Percentage of Correct Parts (PCP3D). For pose forecasting, we measure the MPJPE between the predicted pose and the ground truth pose 0.33s into the future, matching previous work [56].", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14754", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fe6665ec521cb2e8b9383bab871ec6355b361942fc40521f2ac15ccf3f67aa7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneResolutionAP25↑AP50↑AP100↑AP150↑MPJPE (mm)↓FPS (s)↑
VoxelPose[47]ResNet-50960 × 51283.5998.3399.7699.9117.683.2
Faster VoxelPose [53]ResNet-50960 × 51285.2298.0899.3299.4818.2631.1
PlaneSweepPose [32]ResNet-50960 × 51292.1298.9699.8199.8416.754.3
MvP [32]ResNet-50960 × 51292.2896.697.4597.6915.763.6
OursResNet-50960 × 51289.0199.0899.7699.9314.6829.3
VoxelPose[47]HRNet384 × 38482.4498.5599.7499.9217.632.3
Faster VoxelPose [53]HRNet384 × 38481.6998.3899.6799.8318.7722.4
MvP [49]HRNet384 × 38490.4196.3297.3997.8916.342.8
TesseTrack† [42]HRNet384 × 38486.2498.2999.7299.5016.920.6
OursHRNet384 × 38489.3298.4899.7399.9415.9920.3
", + "bbox": [ + 78, + 89, + 911, + 277 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6b6b1eea07101fc85b84f47d63dbd73911e50ce8c93a27bb5e961b5a50c0f235.jpg", + "table_caption": [ + "Table 1: Pose estimation results on the CMU Panoptic dataset. Our method achieves the best MPJPE and AP while running at speed comparable to Faster VoxelPose. We evaluate our methods at $384 \\times 384$ resolution on the Panoptic dataset, as well as the higher resolution used in other methods. We mark TesseTrack [42] with a $\\dagger$ as their reported results are on a different data split, and the results in this table are from our best reproduction of the method, which is not public." + ], + "table_footnote": [], + "table_body": "
MethodShelfCampus
Actor-1Actor-2Actor-3AverageActor-1Actor2Actor3Average
Belagiannis et al. [3]66.165.083.271.482.072.473.775.8
Ershadi et al. [15]93.375.994.888.094.292.984.690.6
Dong et al.[12]98.894.197.896.997.693.398.096.3
VoxelPose [47]99.394.197.697.097.693.898.896.7
Faster VoxelPose[53]99.496.097.597.696.594.197.996.2
PlaneSweepPose[32]99.396.598.097.998.493.799.097.0
MvP [49]99.395.197.897.498.294.197.496.6
TEMPO (Ours)99.096.398.298.097.795.597.997.3
", + "bbox": [ + 125, + 376, + 844, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: PCP3D accuracy on the Campus and Shelf datasets. We follow the protocol of previous methods and train our backbone on synthetic heatmaps of ground-truth poses. Our method achieves results comparable to the state-of-the-art.", + "bbox": [ + 75, + 573, + 892, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 630, + 294, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [47, 53, 56, 49] we use a ResNet-50 backbone pre-trained on the Panoptic dataset and follow [23] by also using a ResNet-50 backbone pre-trained on Human3.6M for the dataset-specific pose estimation results. We use HR-Net [45] pre-trained on COCO as the model backbone for the generalization experiments, as done in [42] rather than pose estimation backbones that are trained on multi-view datasets. All methods are trained on 8 NVIDIA A100 GPUs with batch size of 2 per GPU. We use Adam with a learning rate of 3e-4, with weight decay of 1e-4 and a linear decay schedule for 10 epochs. We measure FPS using a single A100 GPU, and our code is based off the MMPose [11] library. Additional architectural details are in the supplement.", + "bbox": [ + 75, + 647, + 470, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Pose Estimation Results", + "text_level": 1, + "bbox": [ + 76, + 853, + 297, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first compare the results of our methods other state-of-the-art methods. On the Panoptic Studio dataset, we", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "report results following [53, 47, 49] and use $960 \\times 512$ images and initialize from a ResNet-50 [18, 51] checkpoint pretrained on the Panoptic dataset. We also evaluate our method using an HRNet [45] backbone pretrained on COCO with $384 \\times 384$ images for a fair comparison with TesseTrack. In Table 1 we provide a complete comparison across the state-of-the-art methods, with baselines trained on the same image resolutions and backbones for completeness. TEMPO achieves significantly lower MPJPE across both resolutions and backbones, while running at 29.3 FPS, competitive with Faster VoxelPose. We attribute this performance increase to the smoother and more temporally consistent skeletons our model produces due to its incorporation of temporal context and temporal supervision. We also show in 2 that our model achieves performance competitive with the state of the art on the Campus and Shelf datasets.", + "bbox": [ + 496, + 630, + 893, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14755", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/15873dd9c7c30b7c88d8cdfc7fa331e18f5a4e0323aa5d3ece6a2523d50ac0df.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMOTAIDF1MPJPE
VoxelTrack [55]98.4598.67-
Snipper [56]93.4085.5040.2
TEMPO (Ours)98.4293.6238.5
", + "bbox": [ + 91, + 99, + 352, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(a) Evaluation of tracking and forecasting on the CMU Panoptic dataset. Our method outperforms [56] by using multi-view information and is competitive with VoxelTrack. [55]", + "bbox": [ + 75, + 194, + 367, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/478d7a8f24a18f9705e2390002a6584a0dace38d33eb1b47642c8f2990291a6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training DatasetMPJPE (mm)
PanopticH.6MEgoHumansPanopticH.6MEgoHumans
✓14.6862.96119.8
✓✓36.5625.3108.3
✓✓✓42.832.348.9
", + "bbox": [ + 411, + 88, + 911, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) Evaluating TEMPO's ability to transfer. Unlike previous methods, TEMPO is able to train on multiple datasets and can perform reasonably given enough training data. While our model is able to effectively transfer to Human3.6M, it has difficulty with EgoHumans due likely due to its use of fisheye cameras.", + "bbox": [ + 410, + 191, + 906, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b90e72f1bdc81f6e5f73e22123e355022e51b621dd26a4a3a1b7396b2bd050b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 277, + 230, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2b49a536fcca213c31206216f0d33b90ac0bd065aa7d1e6e7808a5ba16c4132.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 277, + 380, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/225adbddbf0adf933473298923ed665d435772b28e1193ee2e49ef0c9f6a36a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 277, + 529, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddee1d438f47c7ed94bc8d3207f4da99987f22740c281725d1935b31a47d86bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 277, + 679, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60b41b3ccb523c52fd24a94126829f206cebfc4b2c24038bb42fb2347f38dd7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 277, + 818, + 386 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2724d89a55a6ebe7751fb7f267b2e6c6f3beff29bb4332771639fb539e058fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 387, + 230, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8652d7a7a96381970d11f062ceed8a07b322ebc8ad479b7c7ae41f8be3b6604f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 387, + 380, + 491 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7bd6b211d868a076529a179ae772b3c14543c5e0652ca9f1662164b40f9e3bb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 387, + 529, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a3fe3dcaafe6364e73022251465e75e7d324134aae831c42e19c9046e5ede63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 387, + 681, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/76c64d15073a65cf18915a247642f36ebe2521ce1afef6d90d99579fa3d7f9bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 392, + 821, + 493 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64dcd246526b9452aa01e3cd343c6d59b3f19fa5b44b0d24b23f30550c5740d7.jpg", + "image_caption": [ + "Figure 4: Samples of our model's pose estimation performance on the Panoptic Studio, Human3.6M, and EgoHumans datasets. TEMPO predicts accurate poses and tracks them over time." + ], + "image_footnote": [], + "bbox": [ + 83, + 494, + 230, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1c6fdeece528e16cae31aad4301d5f1ea1ed4d513a045197b07275f79fecd9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 494, + 380, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd5b6b064e74e701cd561cb30dfa43fa277fc10f5d6e36643bb4d7a3fcf06df8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 494, + 529, + 598 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d8b13759dedc4460dde312686d3b0643330ff80f8bc5923966e3c5a99195416.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 494, + 679, + 598 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e838f1aaa58fa324e0966d485730bea93e4d61c0c3a5fdbd672f6ed2ef67ec2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 498, + 844, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare pose estimation performance on the Human3.6M and EgoHumans datasets in Table 4. Our results perform comparably to the state-of-the-art. Notably [23] uses ground-truth locations and cropped bounding boxes from input views, while our method is able to match performance despite simultaneously detecting people in the scene. Furthermore, our method significantly outperforms others in the more challenging EgoHumans benchmark, suggesting that temporal context is crucial for handling rapid motion.", + "bbox": [ + 75, + 680, + 472, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Pose Tracking and Forecasting", + "text_level": 1, + "bbox": [ + 76, + 845, + 349, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compared the performance of TEMPO's tracking to VoxelTrack [55] in Table 3a. Our tracker is competitive but", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performs slightly worse, which is expected due to its lack of learned Re-ID features. To our knowledge, pose forecasting has not been attempted for the multi-view case, so we compare against the closest forecasting method, Snipper [56], which estimates future pose from monocular video. Our method takes as input 4 previous timesteps and predicts the pose over the next 3 timesteps, which is 0.33 seconds into the future, matching prior work [56]. Our model achieves state-of-the-art performance, shown in Table 3a. We conducted an additional experiment to measure the performance of our model in transfer on different datasets. The standard practice in monocular 3D pose estimation is to train on and evaluate performance on a combination of multiple datasets. However, the predominant paradigm for", + "bbox": [ + 496, + 680, + 893, + 892 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14756", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e447783c1117f61b2f511ddbddee257e3a6151257d8375faffeb64278e863332.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHuman 3.6MEgoHumans
Martinez et al. [35]57.0-
Pavlakos et al. [37]56.9-
Kadkhodamohammadi et al. [26]49.1-
Ma et al [33]24.4-
VoxelPose [47]19.046.24
Faster VoxelPose [53]19.848.50
MvP [49]18.641.72
Iskakov et al. [23]17.7-
TEMPO (Ours)18.536.74
", + "bbox": [ + 101, + 88, + 446, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "evaluating multi-view pose estimation methods has been to train a single model only on a single dataset and evaluate it on the same dataset. This severely limits the potential of these models to generalize across different settings and thus be deployed in real-world scenarios. Similar to [23], we evaluate the performance of our model trained on multiple combinations of annotated datasets, and report the results for each combination in Table 3b. We run our model with no fine tuning and the same voxel size of $10\\mathrm{cm}^3$ across each dataset. Our method is able to transfer and successfully tracks and predicts pose, but performs noticeably worse, likely due to being trained on a single camera configuration. In particular, the CMU Panoptic training dataset uses 5 cameras, whereas Human3.6M uses 4 and EgoHumans uses 8 fisheye cameras. The model has the most trouble generalizing to EgoHumans, likely due to the significantly larger indoor space and different camera models. We find that TEMPO performs better in transfer after including training data from each dataset, especially on the EgoHumans dataset, suggesting that future works should include diverse multi-view data from different camera configurations and camera models in order to better generalize.", + "bbox": [ + 75, + 434, + 468, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablations", + "text_level": 1, + "bbox": [ + 76, + 777, + 187, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We train ablated models to study the impact of individual components in our method. All our experiments are conducted on the CMU Panoptic dataset, with results shown in table 5. We find that using temporal information slightly helps, but with per-timestep supervision, the model is able to greatly improve. Warping the pose between timesteps further improves the model. We hypothesize that the im", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b866d62bbb3de1b5d9897d85e82664dc34aead5d38481c147b377e617f972d2c.jpg", + "table_caption": [ + "Table 4: Pose Estimation results on the Human3.6M and EgoHumans datasets in MPJPE (mm). Our method is competitive with the state-of-the-art on Human3.6M and surpasses current methods by a significant margin on EgoHumans." + ], + "table_footnote": [], + "table_body": "
MethodTForecastingWarpingPer-t lossMPJPE ↓(mm)
(a)317.83
(b)3✓15.03
(c)3✓✓14.94
(d)3✓✓✓14.68
(e)4✓✓✓14.90
(f)5✓✓✓14.82
", + "bbox": [ + 500, + 88, + 897, + 210 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/29543388f5f1b735d3a6b71bdcb9a88d1fad1789ede8f304863366e502d67bd5.jpg", + "table_caption": [ + "Table 5: Ablation study on various components to our model. The most important component to performance was the per-timestep supervision. Warping the previous feature also improved performance. We observed the forecasting and slightly increasing the length of the input history had no noticeable effect on performance." + ], + "table_footnote": [], + "table_body": "
Cameras12345
MPJPE51.3232.1319.2217.3414.68
", + "bbox": [ + 514, + 335, + 875, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: Ablation on number of cameras. We observe that performance decreases with decreasing number of cameras.", + "bbox": [ + 498, + 386, + 890, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "provement from warping is small due to the relative lack of motion in the Panoptic dataset - the distance between body centers in consecutive timesteps is usually small. We also measure the effect of the history length on performance and found no significant difference. While intuitively, larger history lengths should provide more context, the GPU memory constraints of our method prevent investigating $T > 5$ . We also ablated on the number of cameras on the Panoptic dataset. We found that the MPJPE increases with the number of cameras, matching the findings of [47, 23] and [42].", + "bbox": [ + 496, + 453, + 890, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 631, + 625, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Understanding human behavior from video is a fundamentally temporal problem, requiring accurate and efficient pose estimation algorithms that can reason over time. We presented the first method that satisfies these requirements, achieving state-of-the-art results over existing multi-view pose estimation benchmarks via temporal consistency as a learning objective. Our model is also highly efficient, relying on recurrence to maintain a temporal state while enabling pose tracking and forecasting. TEMPO represents a step closer towards general-purpose human behavior understanding from video.", + "bbox": [ + 496, + 661, + 890, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 854, + 666, + 871 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported partially by Fujitsu.", + "bbox": [ + 519, + 885, + 841, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14757", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sikandar Amin, Mykhaylo Andriluka, Marcus Rohrbach, and Bernt Schiele. Multi-view pictorial structures for 3d human pose estimation. In *Bmvc*, volume 1. Bristol, UK, 2013.", + "[2] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. arXiv preprint arXiv:1511.06432, 2015.", + "[3] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures revisited: Multiple human pose estimation. IEEE Trans. Pattern Anal. Mach. Intell., 38(10):1929-1942, 2016.", + "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016.", + "[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 561-578. Springer, 2016.", + "[6] Lewis Bridgeman, Marco Volino, Jean-Yves Guillemaut, and Adrian Hilton. Multi-person 3d pose estimation and tracking in sports. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019.", + "[7] Simon Bultmann and Sven Behnke. Real-time multi-view 3d human pose estimation using semantic feedback to smart edge sensors. arXiv preprint arXiv:2106.14729, 2021.", + "[8] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 387-404. Springer, 2020.", + "[9] Zhuo Chen, Xu Zhao, and Xiaoyue Wan. Structural triangulation: A closed-form solution to constrained 3d human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 695-711. Springer, 2022.", + "[10] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2262-2271, 2019.", + "[11] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.", + "[12] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[13] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on computer vision and pattern recognition, pages 7792-7801, 2019.", + "[14] Zijian Dong, Jie Song, Xu Chen, Chen Guo, and Otmar Hilliges. Shape-aware multi-person pose estimation from multi-view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11158-11168, 2021.", + "[15] Sara Ershadi-Nasab, Erfan Noury, Shohreh Kasaei, and Esmaeil Sanaei. Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications, 77:15573-15601, 2018.", + "[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 59-75. Springer, 2022.", + "[17] Adam W. Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. Simple-BEV: What really matters for multi-sensor bev perception? In arXiv:2206.07959, 2022.", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016.", + "[19] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7779-7788, 2020.", + "[20] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15273-15282, October 2021.", + "[21] Catalin Ionescu, Fuxin Li, and Cristian Sminchisescu. Latent structured models for human pose estimation. In 2011 International Conference on Computer Vision, pages 2220-2227. IEEE, 2011.", + "[22] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014.", + "[23] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In International Conference on Computer Vision (ICCV), 2019.", + "[24] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European conference on computer vision (ECCV), pages 690-706, 2018.", + "[25] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, 2015.", + "[26] Abdolrahim Kadkhodamohammadi and Nicolas Padoy. A generalizable approach for multi-view 3d human pose regression. Machine Vision and Applications, 32(1):6, 2021." + ], + "bbox": [ + 501, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14758", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. corr abs/1712.06584 (2017). arXiv preprint arXiv:1712.06584, 1(3), 2017.", + "[28] Rawal Khirodkar, Aayush Bansal, Lingni Ma, Richard Newcombe, Minh Vo, and Kris Kitani. Egohumans: An egocentric 3d multi-human benchmark. arXiv preprint arXiv:2305.16487, 2023.", + "[29] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020.", + "[30] Ilya Kostrikov and Juergen Gall. Depth sweep regression forests for estimating 3d human pose from images. In BMVC, volume 1, page 5. Nottingham, UK, 2014.", + "[31] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022.", + "[32] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11886-11895, June 2021.", + "[33] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 424-442. Springer, 2022.", + "[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9489-9497, 2019.", + "[35] Julieta Martinez, Rayat Hossain, Javier Romero, and James J. Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017.", + "[36] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection, 2022.", + "[37] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6988-6997, 2017.", + "[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proceedings of the European Conference on Computer Vision, 2020.", + "[39] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4342-4351, 2019." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people with 3d representations. arXiv preprint arXiv:2111.07868, 2021.", + "[41] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3d appearance, location and pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2740-2749, 2022.", + "[42] N Dinesh Reddy, Laurent Guigues, Leonid Pischulin, Jayan Eledath, and Srinivasa G. Narasimhan. Tessentrack: End-to-end learnable multi-person articulated 3d pose tracking. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021.", + "[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020.", + "[44] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019.", + "[45] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019.", + "[46] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11179-11188, 2021.", + "[47] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision (ECCV), 2020.", + "[48] Hsiao-Yu Fish Tung, Ricson Cheng, and Katerina Fragkiadaki. Learning spatial common sense with geometry-aware recurrent networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2595-2603, 2019.", + "[49] Tao Wang, Jianfeng Zhang, Yujun Cai, Shuicheng Yan, and Jiashi Feng. Direct multi-view multi-person 3d human pose estimation. Advances in Neural Information Processing Systems, 2021.", + "[50] Size Wu, Sheng Jin, Wentao Liu, Lei Bai, Chen Qian, Dong Liu, and Wanli Ouyang. Graph-based 3d multi-person pose estimation using multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11148-11157, 2021.", + "[51] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision (ECCV), 2018.", + "[52] Yan Xu and Kris Kitani. Multi-view multi-person 3d pose estimation with uncalibrated camera networks. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022.", + "[53] Hang Ye, Wentao Zhu, Chunyu Wang, Rujie Wu, and Yizhou Wang. Faster voxelpose: Real-time 3d human pose estimate" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14759", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion by orthographic projection. In European Conference on Computer Vision (ECCV), 2022.", + "[54] Yuxiang Zhang, Liang An, Tao Yu, Xiu Li, Kun Li, and Yebin Liu. 4d association graph for realtime multi-person motion capture using multiple video cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1324-1333, 2020.", + "[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenyu Liu, and Wenjun Zeng. Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2613-2626, 2022.", + "[56] Shihao Zou, Yuanlu Xu, Chao Li, Lingni Ma, Li Cheng, and Minh Vo. Snipper: A spatiotemporal transformer for simultaneous multi-person 3d pose estimation tracking and forecasting on a video snippet, 2022." + ], + "bbox": [ + 78, + 90, + 468, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14760", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_model.json b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7129574f4c3cbd15a370cc187c541fbbd83007b2 --- /dev/null +++ b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "TEMPO: Efficient Multi-View Pose Estimation, Tracking, and Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.182, + 0.697, + 0.218 + ], + "angle": 0, + "content": "Rohan Choudhury Kris M. Kitani Laszlo A. Jeni Robotics Institute, Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.22, + 0.732, + 0.234 + ], + "angle": 0, + "content": "{rchoudhu, kmkitani}@andrew.cmu.edu laszlojeni@cmu.edu" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.273, + 0.29, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.407, + 0.212, + 0.417 + ], + "angle": 0, + "content": "Camera 1" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.273, + 0.492, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.407, + 0.419, + 0.417 + ], + "angle": 0, + "content": "Camera 2" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.273, + 0.691, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.406, + 0.617, + 0.417 + ], + "angle": 0, + "content": "Camera 3" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.273, + 0.876, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.406, + 0.81, + 0.417 + ], + "angle": 0, + "content": "3D Pose" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.431, + 0.895, + 0.493 + ], + "angle": 0, + "content": "Figure 1: We propose TEMPO: TEMporal POse Estimation, a method for multi-view, multi-person pose estimation, tracking and forecasting. TEMPO uses a recurrent architecture to learn a spatiotemporal representation, significantly improving the pose estimation accuracy while preserving speed at inference time. In each view, we show TEMPO's predicted pose skeletons over 10 frames, colored by tracker identity. Lighter colors correspond to previous frames." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.506, + 0.314, + 0.519 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.52, + 0.474, + 0.823 + ], + "angle": 0, + "content": "Existing volumetric methods for predicting 3D human pose estimation are accurate, but computationally expensive and optimized for single time-step prediction. We present TEMPO, an efficient multi-view pose estimation model that learns a robust spatiotemporal representation, improving pose accuracy while also tracking and forecasting human pose. We significantly reduce computation compared to the state-of-the-art by recurrently computing per-person 2D pose features, fusing both spatial and temporal information into a single representation. In doing so, our model is able to use spatiotemporal context to predict more accurate human poses without sacrificing efficiency. We further use this representation to track human poses over time as well as predict future poses. Finally, we demonstrate that our model is able to generalize across datasets without scene-specific fine-tuning. TEMPO achieves \\(10\\%\\) better MPJPE with a \\(33\\times\\) improvement in FPS compared to TesseTrack on the challenging CMU Panoptic Studio dataset. Our code and demos are available at https://rccchoudhury.github.io/tempo2023/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.844, + 0.21, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Estimating the pose of several people from multiple overlapping cameras is a crucial vision problem. Volumet-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.895, + 0.686 + ], + "angle": 0, + "content": "ric multi-view methods, which lift 2D image features from each camera view to a feature volume then regress 3D pose, are currently the state of the art [47, 42, 53, 23] in this task. These approaches produce significantly more accurate poses than geometric alternatives, but suffer from two key limitations. First, the most accurate methods use either 3D convolutions [47, 42, 55] or cross-view transformers [49] which are slow and prevent real-time inference. Secondly, most methods are designed for estimating pose at a single timestep and are unable to reason over time, limiting their accuracy and preventing their use for tasks like motion prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.896, + 0.902 + ], + "angle": 0, + "content": "We propose TEMPO, a multi-view TEMporal POse estimation method that addresses both of these issues. TEMPO uses temporal context from previous timesteps to produce smoother and more accurate pose estimates. Our model tracks people over time, predicts future pose and runs efficiently, achieving near real-time performance on existing benchmarks. The key insight behind TEMPO, inspired by work in 3D object detection [31, 20], is that recurrently aggregating spatiotemporal context results in powerful learned representations while being computationally efficient. To do this, we decompose the problem into three stages, illustrated in Figure 2. Given an input RGB video from multiple static, calibrated cameras, at a given timestep \\(t\\) we first detect the locations of each person in the scene by unproject" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14750" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "ing image features from each view to a common 3D volume. We then regress 3D bounding boxes centered on each person, and perform tracking by matching the box centers with the detections from the previous timestep \\( t - 1 \\). For each detected person, we compute a spatiotemporal pose representation by recurrently combining features from current and previous timesteps. We then decode the representation into an estimate of the current pose as well as poses at future timesteps. Unlike existing work [47, 53, 42, 55], our method is able to perform temporal tasks like tracking and forecasting without sacrificing efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.514 + ], + "angle": 0, + "content": "We evaluate our method on several pose estimation benchmarks. TEMPO achieves state of the art results on the challenging CMU Panoptic Studio dataset [25] by \\(10\\%\\), and is competitive on the Campus, Shelf and Human3.6M datasets. We additionally collect our own multiview dataset consisting of highly dynamic scenes, on which TEMPO achieves the best result by a large margin. We show that our model achieves competitive results in pose tracking and evaluate the pose forecasting quality on the CMU Panoptic dataset. Additionally, multi-view pose estimation methods are almost always evaluated on the same dataset they are trained on, leading to results that are specific to certain scenes and camera configurations. We measure our method's ability to generalize across different datasets and find that our method can transfer without additional fine tuning. To summarize, our key contributions are that:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.524, + 0.468, + 0.584 + ], + "angle": 0, + "content": "- We develop the most accurate multi-view, multiperson 3D human pose estimation model. Our model uses temporal context to produce smoother and more accurate poses." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.587, + 0.468, + 0.615 + ], + "angle": 0, + "content": "- Our model runs efficiently with no performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.619, + 0.468, + 0.648 + ], + "angle": 0, + "content": "- Our model tracks and forecasts human pose for every person in the scene." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.651, + 0.468, + 0.681 + ], + "angle": 0, + "content": "- We evaluate the generalization of our model across multiple datasets and camera configurations." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.524, + 0.468, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.22, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.825 + ], + "angle": 0, + "content": "3D Pose Estimation, Tracking and Forecasting Approaches for recovering and tracking 3D human pose are usually limited to monocular video. Such methods for pose estimation [46, 5, 27, 29] and pose tracking [41, 40] are highly efficient, but perform significantly worse than multi-view methods in 3D pose estimation accuracy due to the inherent ambiguity of monocular input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Furthermore, methods in human pose forecasting [34] usually predict future motion from ground truth pose histories. Our approach follows [8, 56] and predicts pose directly from a sequence of video frames. Snipper [56] is the closest to our method and uses a spatiotemporal transformer to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "jointly estimate, track and forecast pose from a monocular video. Our method differs in that it is able to produce highly accurate estimates using multi-view information while running efficiently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.155, + 0.893, + 0.352 + ], + "angle": 0, + "content": "Multi-View Pose Estimation Early work in multi-view human pose estimation was limited to the single-person case, with [3, 30, 37, 1] using pictorial structure models to improve over basic triangulation. More recent approaches [19, 39, 23, 7, 26] improve this result by using advanced deep architectures like 3D CNNs and transformers, and others [9, 10] introduce priors on human shape for additional performance. Our method is most similar to [23], which uses 3D CNNs to regress pose directly from a feature volume. We also follow [23, 26] in analyzing our model's transfer to multiple datasets, extending their qualitative, single-person analysis to a quantitative measurement of performance on several multi-person datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.895, + 0.656 + ], + "angle": 0, + "content": "In the multi-person setting, early approaches like [3, 12, 13, 52, 6] associate 2D pose estimates from each view, then fuse the matched 2D poses into 3D. Other methods aimed towards multi-person motion capture use Re-ID features [12, 52], 4D graph cuts [54], plane sweep stereo [32], cross-view graph matching [50], or optimize SMPL parameters [14] to produce 3D poses from 2D pose estimates in each view. These methods can generalize across the data sources, but typically have much less accurate predictions compared to volumetric methods. These first unproject learned 2D image features into a 3D volume and regress pose directly from the 3D features with neural networks. Both [23] and [47] use computationally expensive 3D CNNs for the pose estimation step. Follow up work includes Faster VoxelPose [53] which replaces these 3D CNNs with 2D CNNs for a large speedup, and TesseTrack [42], which uses 4D CNNs to reason over multiple timesteps. Our method combines the best of both: we efficiently incorporate spatiotemporal information with only 2D CNNs and a lightweight recurrent network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.901 + ], + "angle": 0, + "content": "3D Object Detection and Forecasting Historically, 3D object detection and instance segmentation methods for autonomous driving have led development in using multi-view images. One key similarity to our work is the aggregation of 2D image features into a single 3D volume. While [44, 48] use the same bilinear unprojection strategy as our method, several works [38, 17, 31] propose alternatives such as predicting per-pixel depth. Other works also use temporal information for detection for tracking objects through occlusion and spotting hard-to-see objects; [20, 31, 36] concretely demonstrate the benefits of incorporating spatiotemporal context. In particular, FIERY [20] uses temporal information for future instance prediction and BEVFormer [31] efficiently aggregates temporal information with a recurrent architecture, both of which inspired our method. Furthermore, [24, 16] use per-timestep supervision to track" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14751" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.873, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.388, + 0.895, + 0.45 + ], + "angle": 0, + "content": "Figure 2: The overall model architecture. We begin by (1) extracting features from each image with the backbone network and unprojecting those features to a 3D volume. In step (2), we use the volume to detect each person in the scene, and (3) associate the detections from the current timestep to the previous one. We then (4) fuse the features from each person with our temporal model and produce a final pose estimate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.471, + 0.506 + ], + "angle": 0, + "content": "pixels through occlusion, an idea which we adapt for reasoning about human pose over multiple frames." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.517, + 0.169, + 0.532 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.737 + ], + "angle": 0, + "content": "Our method assumes access to calibrated timesynchronized videos from one or more cameras. At training time, we assume access to \\( T \\) sets of \\( N \\) RGB images from different cameras, while at inference time, we have a single set of \\( N \\) images corresponding to the current timestep. In order to enable TEMPO to transfer to new camera configurations and settings, we compute the dimensions of the space and size of the voxel volume directly from the camera matrices. We set the height of the volume to a constant \\( 2\\mathrm{m} \\), while setting the length and width of the volume to be the bounding box of the camera extrinsics from a top-down view, and center the space at the mean of the camera locations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.746, + 0.219, + 0.76 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.469, + 0.83 + ], + "angle": 0, + "content": "We briefly review the person detection and pose estimation modules used by VoxelPose [47], Tessetrack [42], and Faster VoxelPose [53] that TEMPO builds upon. We refer the reader to the original papers for further detail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.25, + 0.861 + ], + "angle": 0, + "content": "3.1.1 Person Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The detection module aims to estimate the location of the root joint as well as tight 3D bounding box for each person" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.892, + 0.566 + ], + "angle": 0, + "content": "in the scene. Following previous work, we define the root joint as the mid-hip. At a given time \\( t \\), the detector module takes as input a set of \\( N \\) images, each corresponding to a different camera view of the same scene at time \\( t \\). For each image, we extract features with a pretrained backbone, resulting in \\( N \\) feature maps \\( \\mathbf{F}_1^t, \\mathbf{F}_2^t, \\dots, \\mathbf{F}_N^t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.625 + ], + "angle": 0, + "content": "Given the camera matrices for each view \\(\\mathbf{C}_1^t,\\mathbf{C}_2^t,\\ldots ,\\mathbf{C}_N^t\\) , we use the bilinear sampling procedure from [23, 47, 17]. For a voxel \\(v\\in V\\) with coordinates \\(\\mathbf{x}\\) , we have" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.626, + 0.892, + 0.667 + ], + "angle": 0, + "content": "\\[\nv = \\sum_ {i = 1} ^ {N} \\mathbf {F} _ {i} ^ {t} \\left(\\mathbf {C} _ {i} \\mathbf {x}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.734 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_i^t (\\mathbf{x})\\) is the feature map \\(\\mathbf{F}_i^t\\)'s value at position \\(\\mathbf{x}\\), obtained by bilinear sampling. We then compute a birds-eye view representation of \\(V\\) by taking the maximum along the \\(z\\)-axis:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.735, + 0.892, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathrm {B E V}} ^ {t} = \\max _ {z} V \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We use a 2D CNN to produce a 2D heatmap of \\(\\mathbf{H}^t\\) from \\(\\mathbf{F}_{\\mathrm{BEV}}^t\\) of the \\((x,y)\\) locations of every root joint in the scene. We then sample the top \\(K\\) locations from \\(\\mathbf{H}^t\\), yielding proposals \\((x_1,y_1),(x_2,y_2),\\ldots (x_K,y_K)\\). For each proposal location, we obtain the corresponding feature column \\(V|_{x,y}\\) and apply a 1D CNN to regress a 1D heatmap of the root joint's height, denoted \\(\\mathbf{H}_k^t\\). We then sample the maximum \\(z\\) coordinate for each \\(\\mathbf{H}_k^t\\), and combine these to produce a set of detections \\(D_{t} = \\{(x_{1},y_{1},z_{1}),\\dots (x_{K},y_{k},z_{K})\\}\\). Fi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14752" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.182 + ], + "angle": 0, + "content": "nally, we regress width, length and centerness from \\(\\mathbf{F}_{\\mathrm{BEV}}^t\\) with a multi-headed 2D CNN to produce bounding box predictions for each proposal. The loss function for the detection module has three terms. First, \\(L_{2D}\\) is the distance between the predicted 2D heatmap and the ground truth, given by" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.197, + 0.47, + 0.241 + ], + "angle": 0, + "content": "\\[\nL _ {2 D} = \\sum_ {t = 1} ^ {T} \\sum_ {(x, y))} \\| \\mathbf {H} ^ {t} (x, y) - \\mathbf {H} _ {G T} ^ {t} (x, y) \\| \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.249, + 0.445, + 0.264 + ], + "angle": 0, + "content": "We also compute the loss on the predicted 1D heatmap:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.277, + 0.469, + 0.318 + ], + "angle": 0, + "content": "\\[\nL _ {1 D} = \\sum_ {t = 1} ^ {T} \\sum_ {k = 1} ^ {K} \\sum_ {z} \\left\\| \\mathbf {H} _ {k} ^ {t} (z) - \\mathbf {H} _ {k, G T} ^ {t} (z) \\right\\| \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.331, + 0.426, + 0.346 + ], + "angle": 0, + "content": "Finally, we include the bounding box regression loss" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.358, + 0.469, + 0.402 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {b b o x}} = \\sum_ {t = 1} ^ {T} \\sum_ {(i, j) \\in U} \\| \\mathbf {S} (i, j) - \\mathbf {S} _ {G T} (i, j) \\| _ {1} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.415, + 0.469, + 0.446 + ], + "angle": 0, + "content": "The total detection loss is the sum of the above terms, with \\( L_{det} = L_{2D} + L_{1D} + L_{\\mathrm{bbox}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.468, + 0.345, + 0.482 + ], + "angle": 0, + "content": "3.1.2 Instantaneous Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.469, + 0.584 + ], + "angle": 0, + "content": "For each detection \\( D \\), we construct a volume of fixed size centered on the detection's center \\( c_{i} \\), and unproject the backbone features from each camera view into the volume, as in the detection step. As in [53], we mask out all features falling outside the detection's associated bounding box \\( B_{i} \\), resulting in a feature volume \\( V_{i}^{t} \\) for person \\( i \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.469, + 0.796 + ], + "angle": 0, + "content": "As shown in Figure 3, we project the feature volume \\( V_{t}^{i} \\) to 2D along each of the \\( xy, yz \\), and \\( xz \\) planes, resulting in three 2D feature maps, denoted by \\( \\mathbf{P}_{i,xy}^{t} \\), \\( \\mathbf{P}_{i,xz}^{t} \\), and \\( \\mathbf{P}_{i,yz}^{t} \\). The intuition behind this step is that we can predict the 2D position of each joint in each plane, and fuse the predicted 2D positions back together to form a 3D skeleton. Each feature map is passed through a 2D CNN to decode a heatmap of joint likelihood for every person joint, in each of the three planes, and the 2D joint predictions from each plane are fused into 3D with a learned weighting network. We define both the loss for a predicted pose as the mean squared loss between the computed and ground truth 2D heatmaps, as well as the \\( L_{1} \\) loss of the predicted joint locations and ground truth:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.808, + 0.469, + 0.885 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {j o i n t, t} ^ {k} = \\mathrm {M S E} (\\mathbf {P} _ {x y} ^ {t}, \\mathbf {P} _ {x y, G T} ^ {t}) + \\mathrm {M S E} (\\mathbf {P} _ {x z} ^ {t}, \\mathbf {P} _ {x z, G T} ^ {t}) + \\\\ \\operatorname {M S E} \\left(\\mathbf {P} _ {y z} ^ {t}, \\mathbf {P} _ {y z, G T} ^ {t}\\right) + \\sum_ {j = 1} ^ {J} \\left| j _ {i, p r e d} - j _ {i, G T} \\right| \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.392, + 0.901 + ], + "angle": 0, + "content": "with MSE representing the mean squared error." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.664, + 0.108 + ], + "angle": 0, + "content": "3.2. Person Tracking" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.117, + 0.892, + 0.192 + ], + "angle": 0, + "content": "We now describe how TEMPO uses temporal information. Unlike previous works, TEMPO takes as input a set of \\(N\\) images at time \\(t\\) as well as the person detections \\(D_{t-1}\\) and corresponding 2D pose embeddings \\(P^{t-1}\\) from the previous timestep." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.195, + 0.892, + 0.271 + ], + "angle": 0, + "content": "Each proposed detection from the previous step consists of a body center \\( c_{i} \\) and a bounding box \\( B_{i}^{t} = (h_{i}^{t},w_{i}^{t}) \\). Given \\( K \\) proposals, we compute a \\( K\\times K \\) cost matrix \\( \\mathbf{A} \\) based on the IoU between \\( B_{i}^{t} \\) and \\( B_{k}^{t} \\) for all detections at time \\( t \\), resulting in" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.288, + 0.891, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} [ i ] [ j ] = \\left\\| c _ {i} - c _ {j} \\right\\| \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.322, + 0.89, + 0.352 + ], + "angle": 0, + "content": "with \\( c_{i}, c_{j} \\) being the associated predicted locations of person \\( i \\) and person \\( j \\)'s root joint." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.354, + 0.892, + 0.473 + ], + "angle": 0, + "content": "While VoxelTrack [55] computes cost across every joint in each pose, TEMPO uses the top-down view of the associated bounding box for each person. At inference time, we use the SORT [4] tracker, which is fast and uses a simple Kalman filter with no learned Re-ID mechanism. While [42] uses a learned tracker based on SuperGlue [43], we find that SORT is faster and does not result in degraded performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.493, + 0.866, + 0.509 + ], + "angle": 0, + "content": "3.3. Temporal Pose Estimation and Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.609 + ], + "angle": 0, + "content": "After running the tracker, the input to the pose estimation stage is a set of detection proposals \\( D^{t} \\), the previous timestep's detection proposals \\( D^{t-1} \\), and the assignment matrix between the detection sets \\( \\mathbf{A}^{t} \\). We also assume access to pose features \\( \\mathbf{P}_{i}^{t-1} \\) for each person in the previous timestep. Both \\( D^{t} \\) and \\( D^{t-1} \\) have \\( K \\) proposals each." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.61, + 0.892, + 0.718 + ], + "angle": 0, + "content": "However, \\(\\mathbf{P}_i^{t - 1}\\) and \\(\\mathbf{P}_i^t\\) are centered at \\(c_{i}^{t - 1}\\) and \\(c_{i}^{t}\\) respectively, and thus the pose features from each volume are not in the same coordinate system due to the motion of person \\(i\\). To fix this, we follow the standard procedure used in temporal birds-eye view prediction [31, 20] and warp the projected features from \\(\\mathbf{F}_i^{t - 1}\\) into the coordinate system of \\(\\mathbf{P}_i^t\\) with a translational warp defined by \\(c_{i}^{t} - c_{i}^{t - 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "After warping the previous pose features, we run a recurrent network with Spatial Gated Recurrent Units [2] (SpatialGRUs) to produce multiple embeddings: \\(\\mathbf{F}_i^t\\), representing the current pose, and \\(\\mathbf{F}_i^{t + 1},\\mathbf{F}_i^{t + 2},\\ldots\\), representing the pose in future frames. While [20] and [31] do not propagate gradients through time and only predict object locations and instance segments at time \\(T\\), we do backpropagate through time by predicting the pose for each person at every timestep. At training time, we recurrently compute the temporal representation at each timestep \\(t_0,t_0 + 1,\\dots t_0 + T\\) decode a pose for every timestep, and compute losses over all the predicted poses simultaneously. Thus, the final train" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14753" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.092, + 0.881, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.35, + 0.895, + 0.429 + ], + "angle": 0, + "content": "Figure 3: A closer look at the temporal representation used by our model. Following [53], we first project the feature volume to each of the three planes, and concatenate the projections channel-wise. We pass this feature map through an encoder network. We use this feature encoding as input to the SpatialGRU module, using the spatially warped pose feature from the previous timestep as a hidden state. We use the SpatialGRU module to produce features at the current and future timesteps, which we decode into human poses with the pose decoder network." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.453, + 0.184, + 0.469 + ], + "angle": 0, + "content": "ing objective is" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.475, + 0.47, + 0.517 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {p o s e}} = \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {K} L _ {\\text {j o i n t}, t} ^ {i} + L _ {\\text {j o i n t}, t + 1} ^ {i} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.629 + ], + "angle": 0, + "content": "where \\( L_{\\mathrm{joint},t}^{i} \\) is the L1 distance between the predictde and ground truth pose at time \\( t \\) Providing supervision to the network at every timestep allows the network to learn a representation that encodes the motion between consecutive frames while enabling temporal smoothness between predictions. As we show in Section 4.5, this technique is crucial to our model's performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.722 + ], + "angle": 0, + "content": "While training, we run the network \\( T \\) times, for each input timestep. However, at inference time, we save the previous embeddings and only receive input images at a single timestep \\( t \\), significantly reducing the computational burden and allowing our model to use temporal information without sacrificing efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.747, + 0.21, + 0.765 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.277, + 0.787 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Panoptic Studio The CMU Panoptic Studio dataset [25] is a large multi-view pose estimation dataset with several synchronized camera sequences of multiple interacting subjects. Following prior work [47, 53, 49, 32], we use five HD cameras, specifically cameras 3, 6, 12, 13, and 23. We also use the same training and test split as these works, omitting the sequence 160906_band3 due to data corruption." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Human 3.6M The Human3.6M dataset [22, 21] consists of videos of a single subject in an indoor studio with four static cameras. Each video has a professional actor performing a specific action. We follow the training-test split of prior works, using subjects 9 and 11 for validation and the others for training, while omitting corrupted sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.638 + ], + "angle": 0, + "content": "Campus and Shelf The Campus and Shelf datasets[3] contain approximately 4000 frames of a single scene. While these datasets are commonly used for benchmarking in previous work, they are missing many annotations. We follow previous work [47, 53, 49] and adopt a synthetic heatmap-based scheme for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.892, + 0.747 + ], + "angle": 0, + "content": "EgoHumans We include the newly collected EgoHumans multi-view dataset [28]. This benchmark consists of approximately 1 hour of video of up to 5 subjects performing highly dynamic activities, such as playing tag, fencing, or group assembly. It contains videos from up to eight fisheye cameras and includes both egocentric and exocentric camera and pose data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Metrics For the Panoptic Studio, Human3.6M, and Ego-Humans datasets, we report the mean per-joint position error (MPJPE). We additionally report the Average Precision \\((\\mathrm{AP}_K)\\) on the Panoptic and EgoHumans datasets and report MPJPE for Human3.6M in line with previous work [37, 35, 23, 42]. On the Shelf and Campus dataset we report the Percentage of Correct Parts (PCP3D). For pose forecasting, we measure the MPJPE between the predicted pose and the ground truth pose 0.33s into the future, matching previous work [56]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14754" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.912, + 0.279 + ], + "angle": 0, + "content": "
MethodBackboneResolutionAP25↑AP50↑AP100↑AP150↑MPJPE (mm)↓FPS (s)↑
VoxelPose[47]ResNet-50960 × 51283.5998.3399.7699.9117.683.2
Faster VoxelPose [53]ResNet-50960 × 51285.2298.0899.3299.4818.2631.1
PlaneSweepPose [32]ResNet-50960 × 51292.1298.9699.8199.8416.754.3
MvP [32]ResNet-50960 × 51292.2896.697.4597.6915.763.6
OursResNet-50960 × 51289.0199.0899.7699.9314.6829.3
VoxelPose[47]HRNet384 × 38482.4498.5599.7499.9217.632.3
Faster VoxelPose [53]HRNet384 × 38481.6998.3899.6799.8318.7722.4
MvP [49]HRNet384 × 38490.4196.3297.3997.8916.342.8
TesseTrack† [42]HRNet384 × 38486.2498.2999.7299.5016.920.6
OursHRNet384 × 38489.3298.4899.7399.9415.9920.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.365 + ], + "angle": 0, + "content": "Table 1: Pose estimation results on the CMU Panoptic dataset. Our method achieves the best MPJPE and AP while running at speed comparable to Faster VoxelPose. We evaluate our methods at \\(384 \\times 384\\) resolution on the Panoptic dataset, as well as the higher resolution used in other methods. We mark TesseTrack [42] with a \\(\\dagger\\) as their reported results are on a different data split, and the results in this table are from our best reproduction of the method, which is not public." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.377, + 0.846, + 0.553 + ], + "angle": 0, + "content": "
MethodShelfCampus
Actor-1Actor-2Actor-3AverageActor-1Actor2Actor3Average
Belagiannis et al. [3]66.165.083.271.482.072.473.775.8
Ershadi et al. [15]93.375.994.888.094.292.984.690.6
Dong et al.[12]98.894.197.896.997.693.398.096.3
VoxelPose [47]99.394.197.697.097.693.898.896.7
Faster VoxelPose[53]99.496.097.597.696.594.197.996.2
PlaneSweepPose[32]99.396.598.097.998.493.799.097.0
MvP [49]99.395.197.897.498.294.197.496.6
TEMPO (Ours)99.096.398.298.097.795.597.997.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.574, + 0.893, + 0.606 + ], + "angle": 0, + "content": "Table 2: PCP3D accuracy on the Campus and Shelf datasets. We follow the protocol of previous methods and train our backbone on synthetic heatmaps of ground-truth poses. Our method achieves results comparable to the state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.295, + 0.646 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.471, + 0.845 + ], + "angle": 0, + "content": "Following [47, 53, 56, 49] we use a ResNet-50 backbone pre-trained on the Panoptic dataset and follow [23] by also using a ResNet-50 backbone pre-trained on Human3.6M for the dataset-specific pose estimation results. We use HR-Net [45] pre-trained on COCO as the model backbone for the generalization experiments, as done in [42] rather than pose estimation backbones that are trained on multi-view datasets. All methods are trained on 8 NVIDIA A100 GPUs with batch size of 2 per GPU. We use Adam with a learning rate of 3e-4, with weight decay of 1e-4 and a linear decay schedule for 10 epochs. We measure FPS using a single A100 GPU, and our code is based off the MMPose [11] library. Additional architectural details are in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.854, + 0.298, + 0.869 + ], + "angle": 0, + "content": "4.3. Pose Estimation Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We first compare the results of our methods other state-of-the-art methods. On the Panoptic Studio dataset, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.895, + 0.888 + ], + "angle": 0, + "content": "report results following [53, 47, 49] and use \\(960 \\times 512\\) images and initialize from a ResNet-50 [18, 51] checkpoint pretrained on the Panoptic dataset. We also evaluate our method using an HRNet [45] backbone pretrained on COCO with \\(384 \\times 384\\) images for a fair comparison with TesseTrack. In Table 1 we provide a complete comparison across the state-of-the-art methods, with baselines trained on the same image resolutions and backbones for completeness. TEMPO achieves significantly lower MPJPE across both resolutions and backbones, while running at 29.3 FPS, competitive with Faster VoxelPose. We attribute this performance increase to the smoother and more temporally consistent skeletons our model produces due to its incorporation of temporal context and temporal supervision. We also show in 2 that our model achieves performance competitive with the state of the art on the Campus and Shelf datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "14755" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.1, + 0.353, + 0.177 + ], + "angle": 0, + "content": "
MethodMOTAIDF1MPJPE
VoxelTrack [55]98.4598.67-
Snipper [56]93.4085.5040.2
TEMPO (Ours)98.4293.6238.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.195, + 0.368, + 0.251 + ], + "angle": 0, + "content": "(a) Evaluation of tracking and forecasting on the CMU Panoptic dataset. Our method outperforms [56] by using multi-view information and is competitive with VoxelTrack. [55]" + }, + { + "type": "table", + "bbox": [ + 0.413, + 0.089, + 0.912, + 0.188 + ], + "angle": 0, + "content": "
Training DatasetMPJPE (mm)
PanopticH.6MEgoHumansPanopticH.6MEgoHumans
✓14.6862.96119.8
✓✓36.5625.3108.3
✓✓✓42.832.348.9
" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.193, + 0.907, + 0.25 + ], + "angle": 0, + "content": "(b) Evaluating TEMPO's ability to transfer. Unlike previous methods, TEMPO is able to train on multiple datasets and can perform reasonably given enough training data. While our model is able to effectively transfer to Human3.6M, it has difficulty with EgoHumans due likely due to its use of fisheye cameras." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.279, + 0.232, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.279, + 0.382, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.279, + 0.53, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.279, + 0.68, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.279, + 0.819, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.388, + 0.231, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.388, + 0.382, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.388, + 0.531, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.388, + 0.682, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.393, + 0.822, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.495, + 0.231, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.495, + 0.381, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.495, + 0.531, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.495, + 0.681, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.499, + 0.846, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.623, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Figure 4: Samples of our model's pose estimation performance on the Panoptic Studio, Human3.6M, and EgoHumans datasets. TEMPO predicts accurate poses and tracks them over time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.473, + 0.831 + ], + "angle": 0, + "content": "We compare pose estimation performance on the Human3.6M and EgoHumans datasets in Table 4. Our results perform comparably to the state-of-the-art. Notably [23] uses ground-truth locations and cropped bounding boxes from input views, while our method is able to match performance despite simultaneously detecting people in the scene. Furthermore, our method significantly outperforms others in the more challenging EgoHumans benchmark, suggesting that temporal context is crucial for handling rapid motion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.351, + 0.863 + ], + "angle": 0, + "content": "4.4. Pose Tracking and Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We compared the performance of TEMPO's tracking to VoxelTrack [55] in Table 3a. Our tracker is competitive but" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.895, + 0.893 + ], + "angle": 0, + "content": "performs slightly worse, which is expected due to its lack of learned Re-ID features. To our knowledge, pose forecasting has not been attempted for the multi-view case, so we compare against the closest forecasting method, Snipper [56], which estimates future pose from monocular video. Our method takes as input 4 previous timesteps and predicts the pose over the next 3 timesteps, which is 0.33 seconds into the future, matching prior work [56]. Our model achieves state-of-the-art performance, shown in Table 3a. We conducted an additional experiment to measure the performance of our model in transfer on different datasets. The standard practice in monocular 3D pose estimation is to train on and evaluate performance on a combination of multiple datasets. However, the predominant paradigm for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14756" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.447, + 0.308 + ], + "angle": 0, + "content": "
MethodHuman 3.6MEgoHumans
Martinez et al. [35]57.0-
Pavlakos et al. [37]56.9-
Kadkhodamohammadi et al. [26]49.1-
Ma et al [33]24.4-
VoxelPose [47]19.046.24
Faster VoxelPose [53]19.848.50
MvP [49]18.641.72
Iskakov et al. [23]17.7-
TEMPO (Ours)18.536.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.406 + ], + "angle": 0, + "content": "Table 4: Pose Estimation results on the Human3.6M and EgoHumans datasets in MPJPE (mm). Our method is competitive with the state-of-the-art on Human3.6M and surpasses current methods by a significant margin on EgoHumans." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.768 + ], + "angle": 0, + "content": "evaluating multi-view pose estimation methods has been to train a single model only on a single dataset and evaluate it on the same dataset. This severely limits the potential of these models to generalize across different settings and thus be deployed in real-world scenarios. Similar to [23], we evaluate the performance of our model trained on multiple combinations of annotated datasets, and report the results for each combination in Table 3b. We run our model with no fine tuning and the same voxel size of \\(10\\mathrm{cm}^3\\) across each dataset. Our method is able to transfer and successfully tracks and predicts pose, but performs noticeably worse, likely due to being trained on a single camera configuration. In particular, the CMU Panoptic training dataset uses 5 cameras, whereas Human3.6M uses 4 and EgoHumans uses 8 fisheye cameras. The model has the most trouble generalizing to EgoHumans, likely due to the significantly larger indoor space and different camera models. We find that TEMPO performs better in transfer after including training data from each dataset, especially on the EgoHumans dataset, suggesting that future works should include diverse multi-view data from different camera configurations and camera models in order to better generalize." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.778, + 0.189, + 0.792 + ], + "angle": 0, + "content": "4.5. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We train ablated models to study the impact of individual components in our method. All our experiments are conducted on the CMU Panoptic dataset, with results shown in table 5. We find that using temporal information slightly helps, but with per-timestep supervision, the model is able to greatly improve. Warping the pose between timesteps further improves the model. We hypothesize that the im" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.898, + 0.212 + ], + "angle": 0, + "content": "
MethodTForecastingWarpingPer-t lossMPJPE ↓(mm)
(a)317.83
(b)3✓15.03
(c)3✓✓14.94
(d)3✓✓✓14.68
(e)4✓✓✓14.90
(f)5✓✓✓14.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.222, + 0.892, + 0.313 + ], + "angle": 0, + "content": "Table 5: Ablation study on various components to our model. The most important component to performance was the per-timestep supervision. Warping the previous feature also improved performance. We observed the forecasting and slightly increasing the length of the input history had no noticeable effect on performance." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.336, + 0.877, + 0.376 + ], + "angle": 0, + "content": "
Cameras12345
MPJPE51.3232.1319.2217.3414.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.387, + 0.892, + 0.418 + ], + "angle": 0, + "content": "Table 6: Ablation on number of cameras. We observe that performance decreases with decreasing number of cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.892, + 0.606 + ], + "angle": 0, + "content": "provement from warping is small due to the relative lack of motion in the Panoptic dataset - the distance between body centers in consecutive timesteps is usually small. We also measure the effect of the history length on performance and found no significant difference. While intuitively, larger history lengths should provide more context, the GPU memory constraints of our method prevent investigating \\( T > 5 \\). We also ablated on the number of cameras on the Panoptic dataset. We found that the MPJPE increases with the number of cameras, matching the findings of [47, 23] and [42]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.632, + 0.627, + 0.648 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Understanding human behavior from video is a fundamentally temporal problem, requiring accurate and efficient pose estimation algorithms that can reason over time. We presented the first method that satisfies these requirements, achieving state-of-the-art results over existing multi-view pose estimation benchmarks via temporal consistency as a learning objective. Our model is also highly efficient, relying on recurrence to maintain a temporal state while enabling pose tracking and forecasting. TEMPO represents a step closer towards general-purpose human behavior understanding from video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.856, + 0.667, + 0.872 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.842, + 0.902 + ], + "angle": 0, + "content": "This research was supported partially by Fujitsu." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14757" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.469, + 0.158 + ], + "angle": 0, + "content": "[1] Sikandar Amin, Mykhaylo Andriluka, Marcus Rohrbach, and Bernt Schiele. Multi-view pictorial structures for 3d human pose estimation. In *Bmvc*, volume 1. Bristol, UK, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.469, + 0.214 + ], + "angle": 0, + "content": "[2] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. arXiv preprint arXiv:1511.06432, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[3] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures revisited: Multiple human pose estimation. IEEE Trans. Pattern Anal. Mach. Intell., 38(10):1929-1942, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.332, + 0.469, + 0.428 + ], + "angle": 0, + "content": "[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 561-578. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[6] Lewis Bridgeman, Marco Volino, Jean-Yves Guillemaut, and Adrian Hilton. Multi-person 3d pose estimation and tracking in sports. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[7] Simon Bultmann and Sven Behnke. Real-time multi-view 3d human pose estimation using semantic feedback to smart edge sensors. arXiv preprint arXiv:2106.14729, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[8] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 387-404. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[9] Zhuo Chen, Xu Zhao, and Xiaoyue Wan. Structural triangulation: A closed-form solution to constrained 3d human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 695-711. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[10] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2262-2271, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[11] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[12] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "conference on computer vision and pattern recognition, pages 7792-7801, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[14] Zijian Dong, Jie Song, Xu Chen, Chen Guo, and Otmar Hilliges. Shape-aware multi-person pose estimation from multi-view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11158-11168, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[15] Sara Ershadi-Nasab, Erfan Noury, Shohreh Kasaei, and Esmaeil Sanaei. Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications, 77:15573-15601, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 59-75. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[17] Adam W. Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. Simple-BEV: What really matters for multi-sensor bev perception? In arXiv:2206.07959, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[19] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7779-7788, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[20] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15273-15282, October 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[21] Catalin Ionescu, Fuxin Li, and Cristian Sminchisescu. Latent structured models for human pose estimation. In 2011 International Conference on Computer Vision, pages 2220-2227. IEEE, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[22] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[23] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In International Conference on Computer Vision (ICCV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[24] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European conference on computer vision (ECCV), pages 690-706, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[25] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Abdolrahim Kadkhodamohammadi and Nicolas Padoy. A generalizable approach for multi-view 3d human pose regression. Machine Vision and Applications, 32(1):6, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14758" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. corr abs/1712.06584 (2017). arXiv preprint arXiv:1712.06584, 1(3), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[28] Rawal Khirodkar, Aayush Bansal, Lingni Ma, Richard Newcombe, Minh Vo, and Kris Kitani. Egohumans: An egocentric 3d multi-human benchmark. arXiv preprint arXiv:2305.16487, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.209, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[29] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.469, + 0.323 + ], + "angle": 0, + "content": "[30] Ilya Kostrikov and Juergen Gall. Depth sweep regression forests for estimating 3d human pose from images. In BMVC, volume 1, page 5. Nottingham, UK, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.469, + 0.393 + ], + "angle": 0, + "content": "[31] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[32] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11886-11895, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[33] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 424-442. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9489-9497, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.654 + ], + "angle": 0, + "content": "[35] Julieta Martinez, Rayat Hossain, Javier Romero, and James J. Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[36] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.47, + 0.784 + ], + "angle": 0, + "content": "[37] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6988-6997, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proceedings of the European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4342-4351, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people with 3d representations. arXiv preprint arXiv:2111.07868, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[41] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3d appearance, location and pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2740-2749, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[42] N Dinesh Reddy, Laurent Guigues, Leonid Pischulin, Jayan Eledath, and Srinivasa G. Narasimhan. Tessentrack: End-to-end learnable multi-person articulated 3d pose tracking. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[44] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[45] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[46] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11179-11188, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[47] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[48] Hsiao-Yu Fish Tung, Ricson Cheng, and Katerina Fragkiadaki. Learning spatial common sense with geometry-aware recurrent networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2595-2603, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[49] Tao Wang, Jianfeng Zhang, Yujun Cai, Shuicheng Yan, and Jiashi Feng. Direct multi-view multi-person 3d human pose estimation. Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[50] Size Wu, Sheng Jin, Wentao Liu, Lei Bai, Chen Qian, Dong Liu, and Wanli Ouyang. Graph-based 3d multi-person pose estimation using multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11148-11157, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[51] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision (ECCV), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[52] Yan Xu and Kris Kitani. Multi-view multi-person 3d pose estimation with uncalibrated camera networks. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[53] Hang Ye, Wentao Zhu, Chunyu Wang, Rujie Wu, and Yizhou Wang. Faster voxelpose: Real-time 3d human pose estimate" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14759" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "tion by orthographic projection. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[54] Yuxiang Zhang, Liang An, Tao Yu, Xiu Li, Kun Li, and Yebin Liu. 4d association graph for realtime multi-person motion capture using multiple video cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1324-1333, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenyu Liu, and Wenjun Zeng. Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2613-2626, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[56] Shihao Zou, Yuanlu Xu, Chao Li, Lingni Ma, Li Cheng, and Minh Vo. Snipper: A spatiotemporal transformer for simultaneous multi-person 3d pose estimation tracking and forecasting on a video snippet, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14760" + } + ] +] \ No newline at end of file diff --git a/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_origin.pdf b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90c1f1f126871a92b27ff419ec41e5ac56fc89cd --- /dev/null +++ b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86d56f03e8037ffeb2447a77de808099896fd6f4cba08f87f9fc5d58387508ea +size 2366233 diff --git a/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/full.md b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2dfa9c67c7c31ab78fa6b09658443d36c2a169cb --- /dev/null +++ b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/full.md @@ -0,0 +1,325 @@ +# TEMPO: Efficient Multi-View Pose Estimation, Tracking, and Forecasting + +Rohan Choudhury Kris M. Kitani Laszlo A. Jeni Robotics Institute, Carnegie Mellon University + +{rchoudhu, kmkitani}@andrew.cmu.edu laszlojeni@cmu.edu + +![](images/488d387e29e768b2ba328eb10c29fbb64cdc06680d2eb97cc21bf19ba276dd03.jpg) +Camera 1 + +![](images/e90a313eeaf7681e26087e5f2a4c607248cb8e0f6e0d62ad3eac565226b4798a.jpg) +Camera 2 +Figure 1: We propose TEMPO: TEMporal POse Estimation, a method for multi-view, multi-person pose estimation, tracking and forecasting. TEMPO uses a recurrent architecture to learn a spatiotemporal representation, significantly improving the pose estimation accuracy while preserving speed at inference time. In each view, we show TEMPO's predicted pose skeletons over 10 frames, colored by tracker identity. Lighter colors correspond to previous frames. + +![](images/a6f7b555d29676ae93669d76d60f5e2dfe662c93e006896cb1fad21fec23ac3e.jpg) +Camera 3 + +![](images/369d91a9d8e0d389f442d4cefe1ed4a086d1db631dddf1f66bfc23d888e95993.jpg) +3D Pose + +# Abstract + +Existing volumetric methods for predicting 3D human pose estimation are accurate, but computationally expensive and optimized for single time-step prediction. We present TEMPO, an efficient multi-view pose estimation model that learns a robust spatiotemporal representation, improving pose accuracy while also tracking and forecasting human pose. We significantly reduce computation compared to the state-of-the-art by recurrently computing per-person 2D pose features, fusing both spatial and temporal information into a single representation. In doing so, our model is able to use spatiotemporal context to predict more accurate human poses without sacrificing efficiency. We further use this representation to track human poses over time as well as predict future poses. Finally, we demonstrate that our model is able to generalize across datasets without scene-specific fine-tuning. TEMPO achieves $10\%$ better MPJPE with a $33\times$ improvement in FPS compared to TesseTrack on the challenging CMU Panoptic Studio dataset. Our code and demos are available at https://rccchoudhury.github.io/tempo2023/. + +# 1. Introduction + +Estimating the pose of several people from multiple overlapping cameras is a crucial vision problem. Volumet- + +ric multi-view methods, which lift 2D image features from each camera view to a feature volume then regress 3D pose, are currently the state of the art [47, 42, 53, 23] in this task. These approaches produce significantly more accurate poses than geometric alternatives, but suffer from two key limitations. First, the most accurate methods use either 3D convolutions [47, 42, 55] or cross-view transformers [49] which are slow and prevent real-time inference. Secondly, most methods are designed for estimating pose at a single timestep and are unable to reason over time, limiting their accuracy and preventing their use for tasks like motion prediction. + +We propose TEMPO, a multi-view TEMporal POse estimation method that addresses both of these issues. TEMPO uses temporal context from previous timesteps to produce smoother and more accurate pose estimates. Our model tracks people over time, predicts future pose and runs efficiently, achieving near real-time performance on existing benchmarks. The key insight behind TEMPO, inspired by work in 3D object detection [31, 20], is that recurrently aggregating spatiotemporal context results in powerful learned representations while being computationally efficient. To do this, we decompose the problem into three stages, illustrated in Figure 2. Given an input RGB video from multiple static, calibrated cameras, at a given timestep $t$ we first detect the locations of each person in the scene by unproject + +ing image features from each view to a common 3D volume. We then regress 3D bounding boxes centered on each person, and perform tracking by matching the box centers with the detections from the previous timestep $t - 1$ . For each detected person, we compute a spatiotemporal pose representation by recurrently combining features from current and previous timesteps. We then decode the representation into an estimate of the current pose as well as poses at future timesteps. Unlike existing work [47, 53, 42, 55], our method is able to perform temporal tasks like tracking and forecasting without sacrificing efficiency. + +We evaluate our method on several pose estimation benchmarks. TEMPO achieves state of the art results on the challenging CMU Panoptic Studio dataset [25] by $10\%$ , and is competitive on the Campus, Shelf and Human3.6M datasets. We additionally collect our own multiview dataset consisting of highly dynamic scenes, on which TEMPO achieves the best result by a large margin. We show that our model achieves competitive results in pose tracking and evaluate the pose forecasting quality on the CMU Panoptic dataset. Additionally, multi-view pose estimation methods are almost always evaluated on the same dataset they are trained on, leading to results that are specific to certain scenes and camera configurations. We measure our method's ability to generalize across different datasets and find that our method can transfer without additional fine tuning. To summarize, our key contributions are that: + +- We develop the most accurate multi-view, multiperson 3D human pose estimation model. Our model uses temporal context to produce smoother and more accurate poses. +- Our model runs efficiently with no performance degradation. +- Our model tracks and forecasts human pose for every person in the scene. +- We evaluate the generalization of our model across multiple datasets and camera configurations. + +# 2. Related Work + +3D Pose Estimation, Tracking and Forecasting Approaches for recovering and tracking 3D human pose are usually limited to monocular video. Such methods for pose estimation [46, 5, 27, 29] and pose tracking [41, 40] are highly efficient, but perform significantly worse than multi-view methods in 3D pose estimation accuracy due to the inherent ambiguity of monocular input. + +Furthermore, methods in human pose forecasting [34] usually predict future motion from ground truth pose histories. Our approach follows [8, 56] and predicts pose directly from a sequence of video frames. Snipper [56] is the closest to our method and uses a spatiotemporal transformer to + +jointly estimate, track and forecast pose from a monocular video. Our method differs in that it is able to produce highly accurate estimates using multi-view information while running efficiently. + +Multi-View Pose Estimation Early work in multi-view human pose estimation was limited to the single-person case, with [3, 30, 37, 1] using pictorial structure models to improve over basic triangulation. More recent approaches [19, 39, 23, 7, 26] improve this result by using advanced deep architectures like 3D CNNs and transformers, and others [9, 10] introduce priors on human shape for additional performance. Our method is most similar to [23], which uses 3D CNNs to regress pose directly from a feature volume. We also follow [23, 26] in analyzing our model's transfer to multiple datasets, extending their qualitative, single-person analysis to a quantitative measurement of performance on several multi-person datasets. + +In the multi-person setting, early approaches like [3, 12, 13, 52, 6] associate 2D pose estimates from each view, then fuse the matched 2D poses into 3D. Other methods aimed towards multi-person motion capture use Re-ID features [12, 52], 4D graph cuts [54], plane sweep stereo [32], cross-view graph matching [50], or optimize SMPL parameters [14] to produce 3D poses from 2D pose estimates in each view. These methods can generalize across the data sources, but typically have much less accurate predictions compared to volumetric methods. These first unproject learned 2D image features into a 3D volume and regress pose directly from the 3D features with neural networks. Both [23] and [47] use computationally expensive 3D CNNs for the pose estimation step. Follow up work includes Faster VoxelPose [53] which replaces these 3D CNNs with 2D CNNs for a large speedup, and TesseTrack [42], which uses 4D CNNs to reason over multiple timesteps. Our method combines the best of both: we efficiently incorporate spatiotemporal information with only 2D CNNs and a lightweight recurrent network. + +3D Object Detection and Forecasting Historically, 3D object detection and instance segmentation methods for autonomous driving have led development in using multi-view images. One key similarity to our work is the aggregation of 2D image features into a single 3D volume. While [44, 48] use the same bilinear unprojection strategy as our method, several works [38, 17, 31] propose alternatives such as predicting per-pixel depth. Other works also use temporal information for detection for tracking objects through occlusion and spotting hard-to-see objects; [20, 31, 36] concretely demonstrate the benefits of incorporating spatiotemporal context. In particular, FIERY [20] uses temporal information for future instance prediction and BEVFormer [31] efficiently aggregates temporal information with a recurrent architecture, both of which inspired our method. Furthermore, [24, 16] use per-timestep supervision to track + +![](images/381410b12e42bd5fdbaedc05dd51fe3e6a1a9ce614aa5acb2dec1525b23ab209.jpg) +Figure 2: The overall model architecture. We begin by (1) extracting features from each image with the backbone network and unprojecting those features to a 3D volume. In step (2), we use the volume to detect each person in the scene, and (3) associate the detections from the current timestep to the previous one. We then (4) fuse the features from each person with our temporal model and produce a final pose estimate. + +pixels through occlusion, an idea which we adapt for reasoning about human pose over multiple frames. + +# 3. Method + +Our method assumes access to calibrated timesynchronized videos from one or more cameras. At training time, we assume access to $T$ sets of $N$ RGB images from different cameras, while at inference time, we have a single set of $N$ images corresponding to the current timestep. In order to enable TEMPO to transfer to new camera configurations and settings, we compute the dimensions of the space and size of the voxel volume directly from the camera matrices. We set the height of the volume to a constant $2\mathrm{m}$ , while setting the length and width of the volume to be the bounding box of the camera extrinsics from a top-down view, and center the space at the mean of the camera locations. + +# 3.1. Preliminaries + +We briefly review the person detection and pose estimation modules used by VoxelPose [47], Tessetrack [42], and Faster VoxelPose [53] that TEMPO builds upon. We refer the reader to the original papers for further detail. + +# 3.1.1 Person Detection + +The detection module aims to estimate the location of the root joint as well as tight 3D bounding box for each person + +in the scene. Following previous work, we define the root joint as the mid-hip. At a given time $t$ , the detector module takes as input a set of $N$ images, each corresponding to a different camera view of the same scene at time $t$ . For each image, we extract features with a pretrained backbone, resulting in $N$ feature maps $\mathbf{F}_1^t, \mathbf{F}_2^t, \dots, \mathbf{F}_N^t$ . + +Given the camera matrices for each view $\mathbf{C}_1^t,\mathbf{C}_2^t,\ldots ,\mathbf{C}_N^t$ , we use the bilinear sampling procedure from [23, 47, 17]. For a voxel $v\in V$ with coordinates $\mathbf{x}$ , we have + +$$ +v = \sum_ {i = 1} ^ {N} \mathbf {F} _ {i} ^ {t} \left(\mathbf {C} _ {i} \mathbf {x}\right) \tag {1} +$$ + +where $\mathbf{F}_i^t (\mathbf{x})$ is the feature map $\mathbf{F}_i^t$ 's value at position $\mathbf{x}$ , obtained by bilinear sampling. We then compute a birds-eye view representation of $V$ by taking the maximum along the $z$ -axis: + +$$ +\mathbf {F} _ {\mathrm {B E V}} ^ {t} = \max _ {z} V \tag {2} +$$ + +We use a 2D CNN to produce a 2D heatmap of $\mathbf{H}^t$ from $\mathbf{F}_{\mathrm{BEV}}^t$ of the $(x,y)$ locations of every root joint in the scene. We then sample the top $K$ locations from $\mathbf{H}^t$ , yielding proposals $(x_1,y_1),(x_2,y_2),\ldots (x_K,y_K)$ . For each proposal location, we obtain the corresponding feature column $V|_{x,y}$ and apply a 1D CNN to regress a 1D heatmap of the root joint's height, denoted $\mathbf{H}_k^t$ . We then sample the maximum $z$ coordinate for each $\mathbf{H}_k^t$ , and combine these to produce a set of detections $D_{t} = \{(x_{1},y_{1},z_{1}),\dots (x_{K},y_{k},z_{K})\}$ . Fi + +nally, we regress width, length and centerness from $\mathbf{F}_{\mathrm{BEV}}^t$ with a multi-headed 2D CNN to produce bounding box predictions for each proposal. The loss function for the detection module has three terms. First, $L_{2D}$ is the distance between the predicted 2D heatmap and the ground truth, given by + +$$ +L _ {2 D} = \sum_ {t = 1} ^ {T} \sum_ {(x, y))} \| \mathbf {H} ^ {t} (x, y) - \mathbf {H} _ {G T} ^ {t} (x, y) \| \tag {3} +$$ + +We also compute the loss on the predicted 1D heatmap: + +$$ +L _ {1 D} = \sum_ {t = 1} ^ {T} \sum_ {k = 1} ^ {K} \sum_ {z} \left\| \mathbf {H} _ {k} ^ {t} (z) - \mathbf {H} _ {k, G T} ^ {t} (z) \right\| \tag {4} +$$ + +Finally, we include the bounding box regression loss + +$$ +L _ {\mathrm {b b o x}} = \sum_ {t = 1} ^ {T} \sum_ {(i, j) \in U} \| \mathbf {S} (i, j) - \mathbf {S} _ {G T} (i, j) \| _ {1} \tag {5} +$$ + +The total detection loss is the sum of the above terms, with $L_{det} = L_{2D} + L_{1D} + L_{\mathrm{bbox}}$ . + +# 3.1.2 Instantaneous Pose Estimation + +For each detection $D$ , we construct a volume of fixed size centered on the detection's center $c_{i}$ , and unproject the backbone features from each camera view into the volume, as in the detection step. As in [53], we mask out all features falling outside the detection's associated bounding box $B_{i}$ , resulting in a feature volume $V_{i}^{t}$ for person $i$ . + +As shown in Figure 3, we project the feature volume $V_{t}^{i}$ to 2D along each of the $xy, yz$ , and $xz$ planes, resulting in three 2D feature maps, denoted by $\mathbf{P}_{i,xy}^{t}$ , $\mathbf{P}_{i,xz}^{t}$ , and $\mathbf{P}_{i,yz}^{t}$ . The intuition behind this step is that we can predict the 2D position of each joint in each plane, and fuse the predicted 2D positions back together to form a 3D skeleton. Each feature map is passed through a 2D CNN to decode a heatmap of joint likelihood for every person joint, in each of the three planes, and the 2D joint predictions from each plane are fused into 3D with a learned weighting network. We define both the loss for a predicted pose as the mean squared loss between the computed and ground truth 2D heatmaps, as well as the $L_{1}$ loss of the predicted joint locations and ground truth: + +$$ +\begin{array}{l} L _ {j o i n t, t} ^ {k} = \mathrm {M S E} (\mathbf {P} _ {x y} ^ {t}, \mathbf {P} _ {x y, G T} ^ {t}) + \mathrm {M S E} (\mathbf {P} _ {x z} ^ {t}, \mathbf {P} _ {x z, G T} ^ {t}) + \\ \operatorname {M S E} \left(\mathbf {P} _ {y z} ^ {t}, \mathbf {P} _ {y z, G T} ^ {t}\right) + \sum_ {j = 1} ^ {J} \left| j _ {i, p r e d} - j _ {i, G T} \right| \tag {6} \\ \end{array} +$$ + +with MSE representing the mean squared error. + +# 3.2. Person Tracking + +We now describe how TEMPO uses temporal information. Unlike previous works, TEMPO takes as input a set of $N$ images at time $t$ as well as the person detections $D_{t-1}$ and corresponding 2D pose embeddings $P^{t-1}$ from the previous timestep. + +Each proposed detection from the previous step consists of a body center $c_{i}$ and a bounding box $B_{i}^{t} = (h_{i}^{t},w_{i}^{t})$ . Given $K$ proposals, we compute a $K\times K$ cost matrix $\mathbf{A}$ based on the IoU between $B_{i}^{t}$ and $B_{k}^{t}$ for all detections at time $t$ , resulting in + +$$ +\mathbf {A} [ i ] [ j ] = \left\| c _ {i} - c _ {j} \right\| \tag {7} +$$ + +with $c_{i}, c_{j}$ being the associated predicted locations of person $i$ and person $j$ 's root joint. + +While VoxelTrack [55] computes cost across every joint in each pose, TEMPO uses the top-down view of the associated bounding box for each person. At inference time, we use the SORT [4] tracker, which is fast and uses a simple Kalman filter with no learned Re-ID mechanism. While [42] uses a learned tracker based on SuperGlue [43], we find that SORT is faster and does not result in degraded performance. + +# 3.3. Temporal Pose Estimation and Forecasting + +After running the tracker, the input to the pose estimation stage is a set of detection proposals $D^{t}$ , the previous timestep's detection proposals $D^{t-1}$ , and the assignment matrix between the detection sets $\mathbf{A}^{t}$ . We also assume access to pose features $\mathbf{P}_{i}^{t-1}$ for each person in the previous timestep. Both $D^{t}$ and $D^{t-1}$ have $K$ proposals each. + +However, $\mathbf{P}_i^{t - 1}$ and $\mathbf{P}_i^t$ are centered at $c_{i}^{t - 1}$ and $c_{i}^{t}$ respectively, and thus the pose features from each volume are not in the same coordinate system due to the motion of person $i$ . To fix this, we follow the standard procedure used in temporal birds-eye view prediction [31, 20] and warp the projected features from $\mathbf{F}_i^{t - 1}$ into the coordinate system of $\mathbf{P}_i^t$ with a translational warp defined by $c_{i}^{t} - c_{i}^{t - 1}$ . + +After warping the previous pose features, we run a recurrent network with Spatial Gated Recurrent Units [2] (SpatialGRUs) to produce multiple embeddings: $\mathbf{F}_i^t$ , representing the current pose, and $\mathbf{F}_i^{t + 1},\mathbf{F}_i^{t + 2},\ldots$ , representing the pose in future frames. While [20] and [31] do not propagate gradients through time and only predict object locations and instance segments at time $T$ , we do backpropagate through time by predicting the pose for each person at every timestep. At training time, we recurrently compute the temporal representation at each timestep $t_0,t_0 + 1,\dots t_0 + T$ decode a pose for every timestep, and compute losses over all the predicted poses simultaneously. Thus, the final train + +![](images/0677ed1d250ab2f50cccc1e9873d5d000c1a39b0351c3e5cca57fcc2d8fb15df.jpg) +Figure 3: A closer look at the temporal representation used by our model. Following [53], we first project the feature volume to each of the three planes, and concatenate the projections channel-wise. We pass this feature map through an encoder network. We use this feature encoding as input to the SpatialGRU module, using the spatially warped pose feature from the previous timestep as a hidden state. We use the SpatialGRU module to produce features at the current and future timesteps, which we decode into human poses with the pose decoder network. + +ing objective is + +$$ +L _ {\text {p o s e}} = \sum_ {t = 1} ^ {T} \sum_ {i = 1} ^ {K} L _ {\text {j o i n t}, t} ^ {i} + L _ {\text {j o i n t}, t + 1} ^ {i} \tag {8} +$$ + +where $L_{\mathrm{joint},t}^{i}$ is the L1 distance between the predictde and ground truth pose at time $t$ Providing supervision to the network at every timestep allows the network to learn a representation that encodes the motion between consecutive frames while enabling temporal smoothness between predictions. As we show in Section 4.5, this technique is crucial to our model's performance. + +While training, we run the network $T$ times, for each input timestep. However, at inference time, we save the previous embeddings and only receive input images at a single timestep $t$ , significantly reducing the computational burden and allowing our model to use temporal information without sacrificing efficiency. + +# 4. Experiments + +# 4.1. Datasets and Metrics + +Panoptic Studio The CMU Panoptic Studio dataset [25] is a large multi-view pose estimation dataset with several synchronized camera sequences of multiple interacting subjects. Following prior work [47, 53, 49, 32], we use five HD cameras, specifically cameras 3, 6, 12, 13, and 23. We also use the same training and test split as these works, omitting the sequence 160906_band3 due to data corruption. + +Human 3.6M The Human3.6M dataset [22, 21] consists of videos of a single subject in an indoor studio with four static cameras. Each video has a professional actor performing a specific action. We follow the training-test split of prior works, using subjects 9 and 11 for validation and the others for training, while omitting corrupted sequences. + +Campus and Shelf The Campus and Shelf datasets[3] contain approximately 4000 frames of a single scene. While these datasets are commonly used for benchmarking in previous work, they are missing many annotations. We follow previous work [47, 53, 49] and adopt a synthetic heatmap-based scheme for training. + +EgoHumans We include the newly collected EgoHumans multi-view dataset [28]. This benchmark consists of approximately 1 hour of video of up to 5 subjects performing highly dynamic activities, such as playing tag, fencing, or group assembly. It contains videos from up to eight fisheye cameras and includes both egocentric and exocentric camera and pose data. + +Metrics For the Panoptic Studio, Human3.6M, and Ego-Humans datasets, we report the mean per-joint position error (MPJPE). We additionally report the Average Precision $(\mathrm{AP}_K)$ on the Panoptic and EgoHumans datasets and report MPJPE for Human3.6M in line with previous work [37, 35, 23, 42]. On the Shelf and Campus dataset we report the Percentage of Correct Parts (PCP3D). For pose forecasting, we measure the MPJPE between the predicted pose and the ground truth pose 0.33s into the future, matching previous work [56]. + +
MethodBackboneResolutionAP25↑AP50↑AP100↑AP150↑MPJPE (mm)↓FPS (s)↑
VoxelPose[47]ResNet-50960 × 51283.5998.3399.7699.9117.683.2
Faster VoxelPose [53]ResNet-50960 × 51285.2298.0899.3299.4818.2631.1
PlaneSweepPose [32]ResNet-50960 × 51292.1298.9699.8199.8416.754.3
MvP [32]ResNet-50960 × 51292.2896.697.4597.6915.763.6
OursResNet-50960 × 51289.0199.0899.7699.9314.6829.3
VoxelPose[47]HRNet384 × 38482.4498.5599.7499.9217.632.3
Faster VoxelPose [53]HRNet384 × 38481.6998.3899.6799.8318.7722.4
MvP [49]HRNet384 × 38490.4196.3297.3997.8916.342.8
TesseTrack† [42]HRNet384 × 38486.2498.2999.7299.5016.920.6
OursHRNet384 × 38489.3298.4899.7399.9415.9920.3
+ +Table 1: Pose estimation results on the CMU Panoptic dataset. Our method achieves the best MPJPE and AP while running at speed comparable to Faster VoxelPose. We evaluate our methods at $384 \times 384$ resolution on the Panoptic dataset, as well as the higher resolution used in other methods. We mark TesseTrack [42] with a $\dagger$ as their reported results are on a different data split, and the results in this table are from our best reproduction of the method, which is not public. + +
MethodShelfCampus
Actor-1Actor-2Actor-3AverageActor-1Actor2Actor3Average
Belagiannis et al. [3]66.165.083.271.482.072.473.775.8
Ershadi et al. [15]93.375.994.888.094.292.984.690.6
Dong et al.[12]98.894.197.896.997.693.398.096.3
VoxelPose [47]99.394.197.697.097.693.898.896.7
Faster VoxelPose[53]99.496.097.597.696.594.197.996.2
PlaneSweepPose[32]99.396.598.097.998.493.799.097.0
MvP [49]99.395.197.897.498.294.197.496.6
TEMPO (Ours)99.096.398.298.097.795.597.997.3
+ +Table 2: PCP3D accuracy on the Campus and Shelf datasets. We follow the protocol of previous methods and train our backbone on synthetic heatmaps of ground-truth poses. Our method achieves results comparable to the state-of-the-art. + +# 4.2. Implementation Details + +Following [47, 53, 56, 49] we use a ResNet-50 backbone pre-trained on the Panoptic dataset and follow [23] by also using a ResNet-50 backbone pre-trained on Human3.6M for the dataset-specific pose estimation results. We use HR-Net [45] pre-trained on COCO as the model backbone for the generalization experiments, as done in [42] rather than pose estimation backbones that are trained on multi-view datasets. All methods are trained on 8 NVIDIA A100 GPUs with batch size of 2 per GPU. We use Adam with a learning rate of 3e-4, with weight decay of 1e-4 and a linear decay schedule for 10 epochs. We measure FPS using a single A100 GPU, and our code is based off the MMPose [11] library. Additional architectural details are in the supplement. + +# 4.3. Pose Estimation Results + +We first compare the results of our methods other state-of-the-art methods. On the Panoptic Studio dataset, we + +report results following [53, 47, 49] and use $960 \times 512$ images and initialize from a ResNet-50 [18, 51] checkpoint pretrained on the Panoptic dataset. We also evaluate our method using an HRNet [45] backbone pretrained on COCO with $384 \times 384$ images for a fair comparison with TesseTrack. In Table 1 we provide a complete comparison across the state-of-the-art methods, with baselines trained on the same image resolutions and backbones for completeness. TEMPO achieves significantly lower MPJPE across both resolutions and backbones, while running at 29.3 FPS, competitive with Faster VoxelPose. We attribute this performance increase to the smoother and more temporally consistent skeletons our model produces due to its incorporation of temporal context and temporal supervision. We also show in 2 that our model achieves performance competitive with the state of the art on the Campus and Shelf datasets. + +
MethodMOTAIDF1MPJPE
VoxelTrack [55]98.4598.67-
Snipper [56]93.4085.5040.2
TEMPO (Ours)98.4293.6238.5
+ +(a) Evaluation of tracking and forecasting on the CMU Panoptic dataset. Our method outperforms [56] by using multi-view information and is competitive with VoxelTrack. [55] + +
Training DatasetMPJPE (mm)
PanopticH.6MEgoHumansPanopticH.6MEgoHumans
✓14.6862.96119.8
✓✓36.5625.3108.3
✓✓✓42.832.348.9
+ +(b) Evaluating TEMPO's ability to transfer. Unlike previous methods, TEMPO is able to train on multiple datasets and can perform reasonably given enough training data. While our model is able to effectively transfer to Human3.6M, it has difficulty with EgoHumans due likely due to its use of fisheye cameras. + +![](images/b90e72f1bdc81f6e5f73e22123e355022e51b621dd26a4a3a1b7396b2bd050b4.jpg) + +![](images/a2b49a536fcca213c31206216f0d33b90ac0bd065aa7d1e6e7808a5ba16c4132.jpg) + +![](images/225adbddbf0adf933473298923ed665d435772b28e1193ee2e49ef0c9f6a36a3.jpg) + +![](images/ddee1d438f47c7ed94bc8d3207f4da99987f22740c281725d1935b31a47d86bd.jpg) + +![](images/60b41b3ccb523c52fd24a94126829f206cebfc4b2c24038bb42fb2347f38dd7c.jpg) + +![](images/b2724d89a55a6ebe7751fb7f267b2e6c6f3beff29bb4332771639fb539e058fb.jpg) + +![](images/8652d7a7a96381970d11f062ceed8a07b322ebc8ad479b7c7ae41f8be3b6604f.jpg) + +![](images/7bd6b211d868a076529a179ae772b3c14543c5e0652ca9f1662164b40f9e3bb4.jpg) + +![](images/1a3fe3dcaafe6364e73022251465e75e7d324134aae831c42e19c9046e5ede63.jpg) + +![](images/76c64d15073a65cf18915a247642f36ebe2521ce1afef6d90d99579fa3d7f9bd.jpg) + +![](images/64dcd246526b9452aa01e3cd343c6d59b3f19fa5b44b0d24b23f30550c5740d7.jpg) +Figure 4: Samples of our model's pose estimation performance on the Panoptic Studio, Human3.6M, and EgoHumans datasets. TEMPO predicts accurate poses and tracks them over time. + +![](images/e1c6fdeece528e16cae31aad4301d5f1ea1ed4d513a045197b07275f79fecd9e.jpg) + +![](images/fd5b6b064e74e701cd561cb30dfa43fa277fc10f5d6e36643bb4d7a3fcf06df8.jpg) + +![](images/1d8b13759dedc4460dde312686d3b0643330ff80f8bc5923966e3c5a99195416.jpg) + +![](images/7e838f1aaa58fa324e0966d485730bea93e4d61c0c3a5fdbd672f6ed2ef67ec2.jpg) + +We compare pose estimation performance on the Human3.6M and EgoHumans datasets in Table 4. Our results perform comparably to the state-of-the-art. Notably [23] uses ground-truth locations and cropped bounding boxes from input views, while our method is able to match performance despite simultaneously detecting people in the scene. Furthermore, our method significantly outperforms others in the more challenging EgoHumans benchmark, suggesting that temporal context is crucial for handling rapid motion. + +# 4.4. Pose Tracking and Forecasting + +We compared the performance of TEMPO's tracking to VoxelTrack [55] in Table 3a. Our tracker is competitive but + +performs slightly worse, which is expected due to its lack of learned Re-ID features. To our knowledge, pose forecasting has not been attempted for the multi-view case, so we compare against the closest forecasting method, Snipper [56], which estimates future pose from monocular video. Our method takes as input 4 previous timesteps and predicts the pose over the next 3 timesteps, which is 0.33 seconds into the future, matching prior work [56]. Our model achieves state-of-the-art performance, shown in Table 3a. We conducted an additional experiment to measure the performance of our model in transfer on different datasets. The standard practice in monocular 3D pose estimation is to train on and evaluate performance on a combination of multiple datasets. However, the predominant paradigm for + +
MethodHuman 3.6MEgoHumans
Martinez et al. [35]57.0-
Pavlakos et al. [37]56.9-
Kadkhodamohammadi et al. [26]49.1-
Ma et al [33]24.4-
VoxelPose [47]19.046.24
Faster VoxelPose [53]19.848.50
MvP [49]18.641.72
Iskakov et al. [23]17.7-
TEMPO (Ours)18.536.74
+ +evaluating multi-view pose estimation methods has been to train a single model only on a single dataset and evaluate it on the same dataset. This severely limits the potential of these models to generalize across different settings and thus be deployed in real-world scenarios. Similar to [23], we evaluate the performance of our model trained on multiple combinations of annotated datasets, and report the results for each combination in Table 3b. We run our model with no fine tuning and the same voxel size of $10\mathrm{cm}^3$ across each dataset. Our method is able to transfer and successfully tracks and predicts pose, but performs noticeably worse, likely due to being trained on a single camera configuration. In particular, the CMU Panoptic training dataset uses 5 cameras, whereas Human3.6M uses 4 and EgoHumans uses 8 fisheye cameras. The model has the most trouble generalizing to EgoHumans, likely due to the significantly larger indoor space and different camera models. We find that TEMPO performs better in transfer after including training data from each dataset, especially on the EgoHumans dataset, suggesting that future works should include diverse multi-view data from different camera configurations and camera models in order to better generalize. + +# 4.5. Ablations + +We train ablated models to study the impact of individual components in our method. All our experiments are conducted on the CMU Panoptic dataset, with results shown in table 5. We find that using temporal information slightly helps, but with per-timestep supervision, the model is able to greatly improve. Warping the pose between timesteps further improves the model. We hypothesize that the im + +Table 4: Pose Estimation results on the Human3.6M and EgoHumans datasets in MPJPE (mm). Our method is competitive with the state-of-the-art on Human3.6M and surpasses current methods by a significant margin on EgoHumans. + +
MethodTForecastingWarpingPer-t lossMPJPE ↓(mm)
(a)317.83
(b)3✓15.03
(c)3✓✓14.94
(d)3✓✓✓14.68
(e)4✓✓✓14.90
(f)5✓✓✓14.82
+ +Table 5: Ablation study on various components to our model. The most important component to performance was the per-timestep supervision. Warping the previous feature also improved performance. We observed the forecasting and slightly increasing the length of the input history had no noticeable effect on performance. + +
Cameras12345
MPJPE51.3232.1319.2217.3414.68
+ +Table 6: Ablation on number of cameras. We observe that performance decreases with decreasing number of cameras. + +provement from warping is small due to the relative lack of motion in the Panoptic dataset - the distance between body centers in consecutive timesteps is usually small. We also measure the effect of the history length on performance and found no significant difference. While intuitively, larger history lengths should provide more context, the GPU memory constraints of our method prevent investigating $T > 5$ . We also ablated on the number of cameras on the Panoptic dataset. We found that the MPJPE increases with the number of cameras, matching the findings of [47, 23] and [42]. + +# 5. Conclusions + +Understanding human behavior from video is a fundamentally temporal problem, requiring accurate and efficient pose estimation algorithms that can reason over time. We presented the first method that satisfies these requirements, achieving state-of-the-art results over existing multi-view pose estimation benchmarks via temporal consistency as a learning objective. Our model is also highly efficient, relying on recurrence to maintain a temporal state while enabling pose tracking and forecasting. TEMPO represents a step closer towards general-purpose human behavior understanding from video. + +# Acknowledgements + +This research was supported partially by Fujitsu. + +# References + +[1] Sikandar Amin, Mykhaylo Andriluka, Marcus Rohrbach, and Bernt Schiele. Multi-view pictorial structures for 3d human pose estimation. In *Bmvc*, volume 1. Bristol, UK, 2013. +[2] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. arXiv preprint arXiv:1511.06432, 2015. +[3] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures revisited: Multiple human pose estimation. IEEE Trans. Pattern Anal. Mach. Intell., 38(10):1929-1942, 2016. +[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. +[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 561-578. Springer, 2016. +[6] Lewis Bridgeman, Marco Volino, Jean-Yves Guillemaut, and Adrian Hilton. Multi-person 3d pose estimation and tracking in sports. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. +[7] Simon Bultmann and Sven Behnke. Real-time multi-view 3d human pose estimation using semantic feedback to smart edge sensors. arXiv preprint arXiv:2106.14729, 2021. +[8] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 387-404. Springer, 2020. +[9] Zhuo Chen, Xu Zhao, and Xiaoyue Wan. Structural triangulation: A closed-form solution to constrained 3d human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 695-711. Springer, 2022. +[10] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2262-2271, 2019. +[11] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020. +[12] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[13] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In Proceedings of the IEEE/CVF + +conference on computer vision and pattern recognition, pages 7792-7801, 2019. +[14] Zijian Dong, Jie Song, Xu Chen, Chen Guo, and Otmar Hilliges. Shape-aware multi-person pose estimation from multi-view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11158-11168, 2021. +[15] Sara Ershadi-Nasab, Erfan Noury, Shohreh Kasaei, and Esmaeil Sanaei. Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications, 77:15573-15601, 2018. +[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 59-75. Springer, 2022. +[17] Adam W. Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. Simple-BEV: What really matters for multi-sensor bev perception? In arXiv:2206.07959, 2022. +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. +[19] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7779-7788, 2020. +[20] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15273-15282, October 2021. +[21] Catalin Ionescu, Fuxin Li, and Cristian Sminchisescu. Latent structured models for human pose estimation. In 2011 International Conference on Computer Vision, pages 2220-2227. IEEE, 2011. +[22] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014. +[23] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In International Conference on Computer Vision (ICCV), 2019. +[24] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European conference on computer vision (ECCV), pages 690-706, 2018. +[25] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, 2015. +[26] Abdolrahim Kadkhodamohammadi and Nicolas Padoy. A generalizable approach for multi-view 3d human pose regression. Machine Vision and Applications, 32(1):6, 2021. + +[27] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. corr abs/1712.06584 (2017). arXiv preprint arXiv:1712.06584, 1(3), 2017. +[28] Rawal Khirodkar, Aayush Bansal, Lingni Ma, Richard Newcombe, Minh Vo, and Kris Kitani. Egohumans: An egocentric 3d multi-human benchmark. arXiv preprint arXiv:2305.16487, 2023. +[29] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. +[30] Ilya Kostrikov and Juergen Gall. Depth sweep regression forests for estimating 3d human pose from images. In BMVC, volume 1, page 5. Nottingham, UK, 2014. +[31] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. +[32] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11886-11895, June 2021. +[33] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 424-442. Springer, 2022. +[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9489-9497, 2019. +[35] Julieta Martinez, Rayat Hossain, Javier Romero, and James J. Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. +[36] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection, 2022. +[37] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6988-6997, 2017. +[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proceedings of the European Conference on Computer Vision, 2020. +[39] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4342-4351, 2019. + +[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people with 3d representations. arXiv preprint arXiv:2111.07868, 2021. +[41] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3d appearance, location and pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2740-2749, 2022. +[42] N Dinesh Reddy, Laurent Guigues, Leonid Pischulin, Jayan Eledath, and Srinivasa G. Narasimhan. Tessentrack: End-to-end learnable multi-person articulated 3d pose tracking. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. +[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020. +[44] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. +[45] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. +[46] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11179-11188, 2021. +[47] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision (ECCV), 2020. +[48] Hsiao-Yu Fish Tung, Ricson Cheng, and Katerina Fragkiadaki. Learning spatial common sense with geometry-aware recurrent networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2595-2603, 2019. +[49] Tao Wang, Jianfeng Zhang, Yujun Cai, Shuicheng Yan, and Jiashi Feng. Direct multi-view multi-person 3d human pose estimation. Advances in Neural Information Processing Systems, 2021. +[50] Size Wu, Sheng Jin, Wentao Liu, Lei Bai, Chen Qian, Dong Liu, and Wanli Ouyang. Graph-based 3d multi-person pose estimation using multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11148-11157, 2021. +[51] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision (ECCV), 2018. +[52] Yan Xu and Kris Kitani. Multi-view multi-person 3d pose estimation with uncalibrated camera networks. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022. +[53] Hang Ye, Wentao Zhu, Chunyu Wang, Rujie Wu, and Yizhou Wang. Faster voxelpose: Real-time 3d human pose estimate + +tion by orthographic projection. In European Conference on Computer Vision (ECCV), 2022. +[54] Yuxiang Zhang, Liang An, Tao Yu, Xiu Li, Kun Li, and Yebin Liu. 4d association graph for realtime multi-person motion capture using multiple video cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1324-1333, 2020. +[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenyu Liu, and Wenjun Zeng. Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2613-2626, 2022. +[56] Shihao Zou, Yuanlu Xu, Chao Li, Lingni Ma, Li Cheng, and Minh Vo. Snipper: A spatiotemporal transformer for simultaneous multi-person 3d pose estimation tracking and forecasting on a video snippet, 2022. \ No newline at end of file diff --git a/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/images.zip b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..61972d8207664ff2ec1184913c7f90203fe81663 --- /dev/null +++ b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc41eaa2f8dfb0b9e037677ebac4fc0861bdcb5eeff683858a23b56e4a7be091 +size 637474 diff --git a/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/layout.json b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..052c68bbf59c09fdaed0a852345bee319c2ad5a6 --- /dev/null +++ b/2023/TEMPO_ Efficient Multi-View Pose Estimation, Tracking, and Forecasting/layout.json @@ -0,0 +1,9022 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "text", + "content": "TEMPO: Efficient Multi-View Pose Estimation, Tracking, and Forecasting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 166, + 144, + 426, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 144, + 426, + 172 + ], + "spans": [ + { + "bbox": [ + 166, + 144, + 426, + 172 + ], + "type": "text", + "content": "Rohan Choudhury Kris M. Kitani Laszlo A. Jeni Robotics Institute, Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 174, + 447, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 174, + 447, + 185 + ], + "spans": [ + { + "bbox": [ + 146, + 174, + 447, + 185 + ], + "type": "text", + "content": "{rchoudhu, kmkitani}@andrew.cmu.edu laszlojeni@cmu.edu" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 216, + 177, + 317 + ], + "blocks": [ + { + "bbox": [ + 58, + 216, + 177, + 317 + ], + "lines": [ + { + "bbox": [ + 58, + 216, + 177, + 317 + ], + "spans": [ + { + "bbox": [ + 58, + 216, + 177, + 317 + ], + "type": "image", + "image_path": "488d387e29e768b2ba328eb10c29fbb64cdc06680d2eb97cc21bf19ba276dd03.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 322, + 129, + 330 + ], + "lines": [ + { + "bbox": [ + 99, + 322, + 129, + 330 + ], + "spans": [ + { + "bbox": [ + 99, + 322, + 129, + 330 + ], + "type": "text", + "content": "Camera 1" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 181, + 216, + 301, + 317 + ], + "blocks": [ + { + "bbox": [ + 181, + 216, + 301, + 317 + ], + "lines": [ + { + "bbox": [ + 181, + 216, + 301, + 317 + ], + "spans": [ + { + "bbox": [ + 181, + 216, + 301, + 317 + ], + "type": "image", + "image_path": "e90a313eeaf7681e26087e5f2a4c607248cb8e0f6e0d62ad3eac565226b4798a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 322, + 256, + 330 + ], + "lines": [ + { + "bbox": [ + 224, + 322, + 256, + 330 + ], + "spans": [ + { + "bbox": [ + 224, + 322, + 256, + 330 + ], + "type": "text", + "content": "Camera 2" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 341, + 547, + 390 + ], + "lines": [ + { + "bbox": [ + 45, + 341, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 45, + 341, + 547, + 390 + ], + "type": "text", + "content": "Figure 1: We propose TEMPO: TEMporal POse Estimation, a method for multi-view, multi-person pose estimation, tracking and forecasting. TEMPO uses a recurrent architecture to learn a spatiotemporal representation, significantly improving the pose estimation accuracy while preserving speed at inference time. In each view, we show TEMPO's predicted pose skeletons over 10 frames, colored by tracker identity. Lighter colors correspond to previous frames." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 302, + 216, + 422, + 317 + ], + "blocks": [ + { + "bbox": [ + 302, + 216, + 422, + 317 + ], + "lines": [ + { + "bbox": [ + 302, + 216, + 422, + 317 + ], + "spans": [ + { + "bbox": [ + 302, + 216, + 422, + 317 + ], + "type": "image", + "image_path": "a6f7b555d29676ae93669d76d60f5e2dfe662c93e006896cb1fad21fec23ac3e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 321, + 377, + 330 + ], + "lines": [ + { + "bbox": [ + 347, + 321, + 377, + 330 + ], + "spans": [ + { + "bbox": [ + 347, + 321, + 377, + 330 + ], + "type": "text", + "content": "Camera 3" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 432, + 216, + 536, + 315 + ], + "blocks": [ + { + "bbox": [ + 432, + 216, + 536, + 315 + ], + "lines": [ + { + "bbox": [ + 432, + 216, + 536, + 315 + ], + "spans": [ + { + "bbox": [ + 432, + 216, + 536, + 315 + ], + "type": "image", + "image_path": "369d91a9d8e0d389f442d4cefe1ed4a086d1db631dddf1f66bfc23d888e95993.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 321, + 495, + 330 + ], + "lines": [ + { + "bbox": [ + 468, + 321, + 495, + 330 + ], + "spans": [ + { + "bbox": [ + 468, + 321, + 495, + 330 + ], + "type": "text", + "content": "3D Pose" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 400, + 192, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 400, + 192, + 411 + ], + "spans": [ + { + "bbox": [ + 143, + 400, + 192, + 411 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 411, + 290, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 411, + 290, + 651 + ], + "spans": [ + { + "bbox": [ + 45, + 411, + 290, + 651 + ], + "type": "text", + "content": "Existing volumetric methods for predicting 3D human pose estimation are accurate, but computationally expensive and optimized for single time-step prediction. We present TEMPO, an efficient multi-view pose estimation model that learns a robust spatiotemporal representation, improving pose accuracy while also tracking and forecasting human pose. We significantly reduce computation compared to the state-of-the-art by recurrently computing per-person 2D pose features, fusing both spatial and temporal information into a single representation. In doing so, our model is able to use spatiotemporal context to predict more accurate human poses without sacrificing efficiency. We further use this representation to track human poses over time as well as predict future poses. Finally, we demonstrate that our model is able to generalize across datasets without scene-specific fine-tuning. TEMPO achieves " + }, + { + "bbox": [ + 45, + 411, + 290, + 651 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 45, + 411, + 290, + 651 + ], + "type": "text", + "content": " better MPJPE with a " + }, + { + "bbox": [ + 45, + 411, + 290, + 651 + ], + "type": "inline_equation", + "content": "33\\times" + }, + { + "bbox": [ + 45, + 411, + 290, + 651 + ], + "type": "text", + "content": " improvement in FPS compared to TesseTrack on the challenging CMU Panoptic Studio dataset. Our code and demos are available at https://rccchoudhury.github.io/tempo2023/." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 668, + 128, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 668, + 128, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 668, + 128, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Estimating the pose of several people from multiple overlapping cameras is a crucial vision problem. Volumet-" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 400, + 547, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 543 + ], + "type": "text", + "content": "ric multi-view methods, which lift 2D image features from each camera view to a feature volume then regress 3D pose, are currently the state of the art [47, 42, 53, 23] in this task. These approaches produce significantly more accurate poses than geometric alternatives, but suffer from two key limitations. First, the most accurate methods use either 3D convolutions [47, 42, 55] or cross-view transformers [49] which are slow and prevent real-time inference. Secondly, most methods are designed for estimating pose at a single timestep and are unable to reason over time, limiting their accuracy and preventing their use for tasks like motion prediction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "text", + "content": "We propose TEMPO, a multi-view TEMporal POse estimation method that addresses both of these issues. TEMPO uses temporal context from previous timesteps to produce smoother and more accurate pose estimates. Our model tracks people over time, predicts future pose and runs efficiently, achieving near real-time performance on existing benchmarks. The key insight behind TEMPO, inspired by work in 3D object detection [31, 20], is that recurrently aggregating spatiotemporal context results in powerful learned representations while being computationally efficient. To do this, we decompose the problem into three stages, illustrated in Figure 2. Given an input RGB video from multiple static, calibrated cameras, at a given timestep " + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "text", + "content": " we first detect the locations of each person in the scene by unproject" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14750" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "ing image features from each view to a common 3D volume. We then regress 3D bounding boxes centered on each person, and perform tracking by matching the box centers with the detections from the previous timestep " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ". For each detected person, we compute a spatiotemporal pose representation by recurrently combining features from current and previous timesteps. We then decode the representation into an estimate of the current pose as well as poses at future timesteps. Unlike existing work [47, 53, 42, 55], our method is able to perform temporal tasks like tracking and forecasting without sacrificing efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "type": "text", + "content": "We evaluate our method on several pose estimation benchmarks. TEMPO achieves state of the art results on the challenging CMU Panoptic Studio dataset [25] by " + }, + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 204, + 289, + 407 + ], + "type": "text", + "content": ", and is competitive on the Campus, Shelf and Human3.6M datasets. We additionally collect our own multiview dataset consisting of highly dynamic scenes, on which TEMPO achieves the best result by a large margin. We show that our model achieves competitive results in pose tracking and evaluate the pose forecasting quality on the CMU Panoptic dataset. Additionally, multi-view pose estimation methods are almost always evaluated on the same dataset they are trained on, leading to results that are specific to certain scenes and camera configurations. We measure our method's ability to generalize across different datasets and find that our method can transfer without additional fine tuning. To summarize, our key contributions are that:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 415, + 286, + 539 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 415, + 286, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 415, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 59, + 415, + 286, + 462 + ], + "type": "text", + "content": "- We develop the most accurate multi-view, multiperson 3D human pose estimation model. Our model uses temporal context to produce smoother and more accurate poses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 464, + 286, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 464, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 59, + 464, + 286, + 487 + ], + "type": "text", + "content": "- Our model runs efficiently with no performance degradation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 490, + 286, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 490, + 286, + 513 + ], + "spans": [ + { + "bbox": [ + 59, + 490, + 286, + 513 + ], + "type": "text", + "content": "- Our model tracks and forecasts human pose for every person in the scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 515, + 286, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 515, + 286, + 539 + ], + "spans": [ + { + "bbox": [ + 59, + 515, + 286, + 539 + ], + "type": "text", + "content": "- We evaluate the generalization of our model across multiple datasets and camera configurations." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 549, + 134, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 134, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 134, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": "3D Pose Estimation, Tracking and Forecasting Approaches for recovering and tracking 3D human pose are usually limited to monocular video. Such methods for pose estimation [46, 5, 27, 29] and pose tracking [41, 40] are highly efficient, but perform significantly worse than multi-view methods in 3D pose estimation accuracy due to the inherent ambiguity of monocular input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Furthermore, methods in human pose forecasting [34] usually predict future motion from ground truth pose histories. Our approach follows [8, 56] and predicts pose directly from a sequence of video frames. Snipper [56] is the closest to our method and uses a spatiotemporal transformer to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "jointly estimate, track and forecast pose from a monocular video. Our method differs in that it is able to produce highly accurate estimates using multi-view information while running efficiently." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 122, + 546, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 122, + 546, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 122, + 546, + 278 + ], + "type": "text", + "content": "Multi-View Pose Estimation Early work in multi-view human pose estimation was limited to the single-person case, with [3, 30, 37, 1] using pictorial structure models to improve over basic triangulation. More recent approaches [19, 39, 23, 7, 26] improve this result by using advanced deep architectures like 3D CNNs and transformers, and others [9, 10] introduce priors on human shape for additional performance. Our method is most similar to [23], which uses 3D CNNs to regress pose directly from a feature volume. We also follow [23, 26] in analyzing our model's transfer to multiple datasets, extending their qualitative, single-person analysis to a quantitative measurement of performance on several multi-person datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 280, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 547, + 519 + ], + "type": "text", + "content": "In the multi-person setting, early approaches like [3, 12, 13, 52, 6] associate 2D pose estimates from each view, then fuse the matched 2D poses into 3D. Other methods aimed towards multi-person motion capture use Re-ID features [12, 52], 4D graph cuts [54], plane sweep stereo [32], cross-view graph matching [50], or optimize SMPL parameters [14] to produce 3D poses from 2D pose estimates in each view. These methods can generalize across the data sources, but typically have much less accurate predictions compared to volumetric methods. These first unproject learned 2D image features into a 3D volume and regress pose directly from the 3D features with neural networks. Both [23] and [47] use computationally expensive 3D CNNs for the pose estimation step. Follow up work includes Faster VoxelPose [53] which replaces these 3D CNNs with 2D CNNs for a large speedup, and TesseTrack [42], which uses 4D CNNs to reason over multiple timesteps. Our method combines the best of both: we efficiently incorporate spatiotemporal information with only 2D CNNs and a lightweight recurrent network." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 713 + ], + "type": "text", + "content": "3D Object Detection and Forecasting Historically, 3D object detection and instance segmentation methods for autonomous driving have led development in using multi-view images. One key similarity to our work is the aggregation of 2D image features into a single 3D volume. While [44, 48] use the same bilinear unprojection strategy as our method, several works [38, 17, 31] propose alternatives such as predicting per-pixel depth. Other works also use temporal information for detection for tracking objects through occlusion and spotting hard-to-see objects; [20, 31, 36] concretely demonstrate the benefits of incorporating spatiotemporal context. In particular, FIERY [20] uses temporal information for future instance prediction and BEVFormer [31] efficiently aggregates temporal information with a recurrent architecture, both of which inspired our method. Furthermore, [24, 16] use per-timestep supervision to track" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14751" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 534, + 291 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 534, + 291 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 534, + 291 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 534, + 291 + ], + "type": "image", + "image_path": "381410b12e42bd5fdbaedc05dd51fe3e6a1a9ce614aa5acb2dec1525b23ab209.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 307, + 547, + 356 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 547, + 356 + ], + "type": "text", + "content": "Figure 2: The overall model architecture. We begin by (1) extracting features from each image with the backbone network and unprojecting those features to a 3D volume. In step (2), we use the volume to detect each person in the scene, and (3) associate the detections from the current timestep to the previous one. We then (4) fuse the features from each person with our temporal model and produce a final pose estimate." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 376, + 288, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 400 + ], + "type": "text", + "content": "pixels through occlusion, an idea which we adapt for reasoning about human pose over multiple frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 409, + 103, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 409, + 103, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 409, + 103, + 421 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "type": "text", + "content": "Our method assumes access to calibrated timesynchronized videos from one or more cameras. At training time, we assume access to " + }, + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "type": "text", + "content": " sets of " + }, + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "type": "text", + "content": " RGB images from different cameras, while at inference time, we have a single set of " + }, + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "type": "text", + "content": " images corresponding to the current timestep. In order to enable TEMPO to transfer to new camera configurations and settings, we compute the dimensions of the space and size of the voxel volume directly from the camera matrices. We set the height of the volume to a constant " + }, + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "type": "inline_equation", + "content": "2\\mathrm{m}" + }, + { + "bbox": [ + 46, + 429, + 287, + 583 + ], + "type": "text", + "content": ", while setting the length and width of the volume to be the bounding box of the camera extrinsics from a top-down view, and center the space at the mean of the camera locations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 590, + 134, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 134, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 134, + 601 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": "We briefly review the person detection and pose estimation modules used by VoxelPose [47], Tessetrack [42], and Faster VoxelPose [53] that TEMPO builds upon. We refer the reader to the original papers for further detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 153, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 153, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 153, + 681 + ], + "type": "text", + "content": "3.1.1 Person Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "The detection module aims to estimate the location of the root joint as well as tight 3D bounding box for each person" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "content": "in the scene. Following previous work, we define the root joint as the mid-hip. At a given time " + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "content": ", the detector module takes as input a set of " + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "content": " images, each corresponding to a different camera view of the same scene at time " + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "content": ". For each image, we extract features with a pretrained backbone, resulting in " + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "content": " feature maps " + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_1^t, \\mathbf{F}_2^t, \\dots, \\mathbf{F}_N^t" + }, + { + "bbox": [ + 304, + 376, + 545, + 448 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 448, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 495 + ], + "type": "text", + "content": "Given the camera matrices for each view " + }, + { + "bbox": [ + 304, + 448, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1^t,\\mathbf{C}_2^t,\\ldots ,\\mathbf{C}_N^t" + }, + { + "bbox": [ + 304, + 448, + 545, + 495 + ], + "type": "text", + "content": " , we use the bilinear sampling procedure from [23, 47, 17]. For a voxel " + }, + { + "bbox": [ + 304, + 448, + 545, + 495 + ], + "type": "inline_equation", + "content": "v\\in V" + }, + { + "bbox": [ + 304, + 448, + 545, + 495 + ], + "type": "text", + "content": " with coordinates " + }, + { + "bbox": [ + 304, + 448, + 545, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 448, + 545, + 495 + ], + "type": "text", + "content": " , we have" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 495, + 545, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 495, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 388, + 495, + 545, + 528 + ], + "type": "interline_equation", + "content": "v = \\sum_ {i = 1} ^ {N} \\mathbf {F} _ {i} ^ {t} \\left(\\mathbf {C} _ {i} \\mathbf {x}\\right) \\tag {1}", + "image_path": "a5b246a4f87cf67ac515461ddc74f9388b7dc2aa6d22037a361d8cbb6034762c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^t (\\mathbf{x})" + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "text", + "content": " is the feature map " + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^t" + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "text", + "content": "'s value at position " + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "text", + "content": ", obtained by bilinear sampling. We then compute a birds-eye view representation of " + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "text", + "content": " by taking the maximum along the " + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 533, + 545, + 581 + ], + "type": "text", + "content": "-axis:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 391, + 582, + 545, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 582, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 391, + 582, + 545, + 600 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathrm {B E V}} ^ {t} = \\max _ {z} V \\tag {2}", + "image_path": "b929700dbbb5d8a4ac87b9191154c737a81803dbcc9f9f50579b443f5b79146b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "We use a 2D CNN to produce a 2D heatmap of " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^t" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{BEV}}^t" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " locations of every root joint in the scene. We then sample the top " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " locations from " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^t" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": ", yielding proposals " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "(x_1,y_1),(x_2,y_2),\\ldots (x_K,y_K)" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": ". For each proposal location, we obtain the corresponding feature column " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "V|_{x,y}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " and apply a 1D CNN to regress a 1D heatmap of the root joint's height, denoted " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_k^t" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": ". We then sample the maximum " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " coordinate for each " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_k^t" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": ", and combine these to produce a set of detections " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "D_{t} = \\{(x_{1},y_{1},z_{1}),\\dots (x_{K},y_{k},z_{K})\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": ". Fi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14752" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "nally, we regress width, length and centerness from " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{BEV}}^t" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " with a multi-headed 2D CNN to produce bounding box predictions for each proposal. The loss function for the detection module has three terms. First, " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "L_{2D}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " is the distance between the predicted 2D heatmap and the ground truth, given by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 156, + 287, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 156, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 82, + 156, + 287, + 190 + ], + "type": "interline_equation", + "content": "L _ {2 D} = \\sum_ {t = 1} ^ {T} \\sum_ {(x, y))} \\| \\mathbf {H} ^ {t} (x, y) - \\mathbf {H} _ {G T} ^ {t} (x, y) \\| \\tag {3}", + "image_path": "7b59f4d9a395102a0a8b098ee07eaf4e5866555bd4fe71059cea003f3529cefc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 197, + 272, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 197, + 272, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 197, + 272, + 209 + ], + "type": "text", + "content": "We also compute the loss on the predicted 1D heatmap:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 219, + 287, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 219, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 82, + 219, + 287, + 251 + ], + "type": "interline_equation", + "content": "L _ {1 D} = \\sum_ {t = 1} ^ {T} \\sum_ {k = 1} ^ {K} \\sum_ {z} \\left\\| \\mathbf {H} _ {k} ^ {t} (z) - \\mathbf {H} _ {k, G T} ^ {t} (z) \\right\\| \\tag {4}", + "image_path": "faae5f12bf704b40d71f421d6aab323e4e938d05d1e1bc2b7df79f1f645dcd5a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 262, + 260, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 260, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 260, + 274 + ], + "type": "text", + "content": "Finally, we include the bounding box regression loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 283, + 287, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 283, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 79, + 283, + 287, + 318 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {b b o x}} = \\sum_ {t = 1} ^ {T} \\sum_ {(i, j) \\in U} \\| \\mathbf {S} (i, j) - \\mathbf {S} _ {G T} (i, j) \\| _ {1} \\tag {5}", + "image_path": "efe4ad76591eaa408fd6f2a2852a725cbe6caede21ae6703f44edbb0a30fac2a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 328, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 287, + 353 + ], + "type": "text", + "content": "The total detection loss is the sum of the above terms, with " + }, + { + "bbox": [ + 47, + 328, + 287, + 353 + ], + "type": "inline_equation", + "content": "L_{det} = L_{2D} + L_{1D} + L_{\\mathrm{bbox}}" + }, + { + "bbox": [ + 47, + 328, + 287, + 353 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 370, + 211, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 211, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 211, + 381 + ], + "type": "text", + "content": "3.1.2 Instantaneous Pose Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "text", + "content": "For each detection " + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "text", + "content": ", we construct a volume of fixed size centered on the detection's center " + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "text", + "content": ", and unproject the backbone features from each camera view into the volume, as in the detection step. As in [53], we mask out all features falling outside the detection's associated bounding box " + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "text", + "content": ", resulting in a feature volume " + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "inline_equation", + "content": "V_{i}^{t}" + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "text", + "content": " for person " + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 390, + 287, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "text", + "content": "As shown in Figure 3, we project the feature volume " + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "inline_equation", + "content": "V_{t}^{i}" + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "text", + "content": " to 2D along each of the " + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "inline_equation", + "content": "xy, yz" + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "inline_equation", + "content": "xz" + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "text", + "content": " planes, resulting in three 2D feature maps, denoted by " + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{i,xy}^{t}" + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{i,xz}^{t}" + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{i,yz}^{t}" + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "text", + "content": ". The intuition behind this step is that we can predict the 2D position of each joint in each plane, and fuse the predicted 2D positions back together to form a 3D skeleton. Each feature map is passed through a 2D CNN to decode a heatmap of joint likelihood for every person joint, in each of the three planes, and the 2D joint predictions from each plane are fused into 3D with a learned weighting network. We define both the loss for a predicted pose as the mean squared loss between the computed and ground truth 2D heatmaps, as well as the " + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 463, + 287, + 630 + ], + "type": "text", + "content": " loss of the predicted joint locations and ground truth:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 51, + 639, + 287, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 639, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 51, + 639, + 287, + 700 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {j o i n t, t} ^ {k} = \\mathrm {M S E} (\\mathbf {P} _ {x y} ^ {t}, \\mathbf {P} _ {x y, G T} ^ {t}) + \\mathrm {M S E} (\\mathbf {P} _ {x z} ^ {t}, \\mathbf {P} _ {x z, G T} ^ {t}) + \\\\ \\operatorname {M S E} \\left(\\mathbf {P} _ {y z} ^ {t}, \\mathbf {P} _ {y z, G T} ^ {t}\\right) + \\sum_ {j = 1} ^ {J} \\left| j _ {i, p r e d} - j _ {i, G T} \\right| \\tag {6} \\\\ \\end{array}", + "image_path": "5fd4c253641d29d505ff4f3d00265acc60c9352ea24bb08f090cb75ffecb474c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 701, + 239, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 239, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 239, + 713 + ], + "type": "text", + "content": "with MSE representing the mean squared error." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 406, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 406, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 406, + 85 + ], + "type": "text", + "content": "3.2. Person Tracking" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "type": "text", + "content": "We now describe how TEMPO uses temporal information. Unlike previous works, TEMPO takes as input a set of " + }, + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "type": "text", + "content": " images at time " + }, + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "type": "text", + "content": " as well as the person detections " + }, + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "type": "inline_equation", + "content": "D_{t-1}" + }, + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "type": "text", + "content": " and corresponding 2D pose embeddings " + }, + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "type": "inline_equation", + "content": "P^{t-1}" + }, + { + "bbox": [ + 306, + 92, + 545, + 152 + ], + "type": "text", + "content": " from the previous timestep." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "text", + "content": "Each proposed detection from the previous step consists of a body center " + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "text", + "content": " and a bounding box " + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "B_{i}^{t} = (h_{i}^{t},w_{i}^{t})" + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "text", + "content": " proposals, we compute a " + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "K\\times K" + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "text", + "content": " cost matrix " + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "text", + "content": " based on the IoU between " + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "B_{i}^{t}" + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "B_{k}^{t}" + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "text", + "content": " for all detections at time " + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 154, + 545, + 214 + ], + "type": "text", + "content": ", resulting in" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 384, + 228, + 545, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 228, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 384, + 228, + 545, + 242 + ], + "type": "interline_equation", + "content": "\\mathbf {A} [ i ] [ j ] = \\left\\| c _ {i} - c _ {j} \\right\\| \\tag {7}", + "image_path": "ac24d2a356d29f210f1cfec73cbbad8a39e28caeaacbcef4a627b2c424cbaa05.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 255, + 544, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 544, + 278 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 544, + 278 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 305, + 255, + 544, + 278 + ], + "type": "inline_equation", + "content": "c_{i}, c_{j}" + }, + { + "bbox": [ + 305, + 255, + 544, + 278 + ], + "type": "text", + "content": " being the associated predicted locations of person " + }, + { + "bbox": [ + 305, + 255, + 544, + 278 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 255, + 544, + 278 + ], + "type": "text", + "content": " and person " + }, + { + "bbox": [ + 305, + 255, + 544, + 278 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 255, + 544, + 278 + ], + "type": "text", + "content": "'s root joint." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 280, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 280, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 280, + 545, + 374 + ], + "type": "text", + "content": "While VoxelTrack [55] computes cost across every joint in each pose, TEMPO uses the top-down view of the associated bounding box for each person. At inference time, we use the SORT [4] tracker, which is fast and uses a simple Kalman filter with no learned Re-ID mechanism. While [42] uses a learned tracker based on SuperGlue [43], we find that SORT is faster and does not result in degraded performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 390, + 529, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 390, + 529, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 529, + 403 + ], + "type": "text", + "content": "3.3. Temporal Pose Estimation and Forecasting" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "text", + "content": "After running the tracker, the input to the pose estimation stage is a set of detection proposals " + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "inline_equation", + "content": "D^{t}" + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "text", + "content": ", the previous timestep's detection proposals " + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "inline_equation", + "content": "D^{t-1}" + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "text", + "content": ", and the assignment matrix between the detection sets " + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{t}" + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "text", + "content": ". We also assume access to pose features " + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{i}^{t-1}" + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "text", + "content": " for each person in the previous timestep. Both " + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "inline_equation", + "content": "D^{t}" + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "inline_equation", + "content": "D^{t-1}" + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "text", + "content": " have " + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 410, + 545, + 482 + ], + "type": "text", + "content": " proposals each." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "text", + "content": "However, " + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i^{t - 1}" + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i^t" + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "text", + "content": " are centered at " + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "inline_equation", + "content": "c_{i}^{t - 1}" + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "inline_equation", + "content": "c_{i}^{t}" + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "text", + "content": " respectively, and thus the pose features from each volume are not in the same coordinate system due to the motion of person " + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "text", + "content": ". To fix this, we follow the standard procedure used in temporal birds-eye view prediction [31, 20] and warp the projected features from " + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{t - 1}" + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "text", + "content": " into the coordinate system of " + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i^t" + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "text", + "content": " with a translational warp defined by " + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "inline_equation", + "content": "c_{i}^{t} - c_{i}^{t - 1}" + }, + { + "bbox": [ + 305, + 483, + 545, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "After warping the previous pose features, we run a recurrent network with Spatial Gated Recurrent Units [2] (SpatialGRUs) to produce multiple embeddings: " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^t" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", representing the current pose, and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{t + 1},\\mathbf{F}_i^{t + 2},\\ldots" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", representing the pose in future frames. While [20] and [31] do not propagate gradients through time and only predict object locations and instance segments at time " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", we do backpropagate through time by predicting the pose for each person at every timestep. At training time, we recurrently compute the temporal representation at each timestep " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "t_0,t_0 + 1,\\dots t_0 + T" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " decode a pose for every timestep, and compute losses over all the predicted poses simultaneously. Thus, the final train" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14753" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 72, + 539, + 266 + ], + "blocks": [ + { + "bbox": [ + 66, + 72, + 539, + 266 + ], + "lines": [ + { + "bbox": [ + 66, + 72, + 539, + 266 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 539, + 266 + ], + "type": "image", + "image_path": "0677ed1d250ab2f50cccc1e9873d5d000c1a39b0351c3e5cca57fcc2d8fb15df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 547, + 339 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 547, + 339 + ], + "type": "text", + "content": "Figure 3: A closer look at the temporal representation used by our model. Following [53], we first project the feature volume to each of the three planes, and concatenate the projections channel-wise. We pass this feature map through an encoder network. We use this feature encoding as input to the SpatialGRU module, using the spatially warped pose feature from the previous timestep as a hidden state. We use the SpatialGRU module to produce features at the current and future timesteps, which we decode into human poses with the pose decoder network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 358, + 112, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 112, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 112, + 371 + ], + "type": "text", + "content": "ing objective is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 376, + 287, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 376, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 91, + 376, + 287, + 409 + ], + "type": "interline_equation", + "content": "L _ {\\text {p o s e}} = \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {K} L _ {\\text {j o i n t}, t} ^ {i} + L _ {\\text {j o i n t}, t + 1} ^ {i} \\tag {8}", + "image_path": "f397114b9c757119b648d4f3f96d50fd5d102fa7cc34e37737071068cea8464a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{joint},t}^{i}" + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "text", + "content": " is the L1 distance between the predictde and ground truth pose at time " + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "text", + "content": " Providing supervision to the network at every timestep allows the network to learn a representation that encodes the motion between consecutive frames while enabling temporal smoothness between predictions. As we show in Section 4.5, this technique is crucial to our model's performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 571 + ], + "type": "text", + "content": "While training, we run the network " + }, + { + "bbox": [ + 46, + 498, + 287, + 571 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 498, + 287, + 571 + ], + "type": "text", + "content": " times, for each input timestep. However, at inference time, we save the previous embeddings and only receive input images at a single timestep " + }, + { + "bbox": [ + 46, + 498, + 287, + 571 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 498, + 287, + 571 + ], + "type": "text", + "content": ", significantly reducing the computational burden and allowing our model to use temporal information without sacrificing efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 591, + 128, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 128, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 128, + 605 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 611, + 169, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 169, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 169, + 623 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Panoptic Studio The CMU Panoptic Studio dataset [25] is a large multi-view pose estimation dataset with several synchronized camera sequences of multiple interacting subjects. Following prior work [47, 53, 49, 32], we use five HD cameras, specifically cameras 3, 6, 12, 13, and 23. We also use the same training and test split as these works, omitting the sequence 160906_band3 due to data corruption." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 358, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 431 + ], + "type": "text", + "content": "Human 3.6M The Human3.6M dataset [22, 21] consists of videos of a single subject in an indoor studio with four static cameras. Each video has a professional actor performing a specific action. We follow the training-test split of prior works, using subjects 9 and 11 for validation and the others for training, while omitting corrupted sequences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 433, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 505 + ], + "type": "text", + "content": "Campus and Shelf The Campus and Shelf datasets[3] contain approximately 4000 frames of a single scene. While these datasets are commonly used for benchmarking in previous work, they are missing many annotations. We follow previous work [47, 53, 49] and adopt a synthetic heatmap-based scheme for training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 507, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 545, + 591 + ], + "type": "text", + "content": "EgoHumans We include the newly collected EgoHumans multi-view dataset [28]. This benchmark consists of approximately 1 hour of video of up to 5 subjects performing highly dynamic activities, such as playing tag, fencing, or group assembly. It contains videos from up to eight fisheye cameras and includes both egocentric and exocentric camera and pose data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Metrics For the Panoptic Studio, Human3.6M, and Ego-Humans datasets, we report the mean per-joint position error (MPJPE). We additionally report the Average Precision " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_K)" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": " on the Panoptic and EgoHumans datasets and report MPJPE for Human3.6M in line with previous work [37, 35, 23, 42]. On the Shelf and Campus dataset we report the Percentage of Correct Parts (PCP3D). For pose forecasting, we measure the MPJPE between the predicted pose and the ground truth pose 0.33s into the future, matching previous work [56]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14754" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 558, + 220 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 558, + 220 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 558, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 558, + 220 + ], + "type": "table", + "html": "
MethodBackboneResolutionAP25↑AP50↑AP100↑AP150↑MPJPE (mm)↓FPS (s)↑
VoxelPose[47]ResNet-50960 × 51283.5998.3399.7699.9117.683.2
Faster VoxelPose [53]ResNet-50960 × 51285.2298.0899.3299.4818.2631.1
PlaneSweepPose [32]ResNet-50960 × 51292.1298.9699.8199.8416.754.3
MvP [32]ResNet-50960 × 51292.2896.697.4597.6915.763.6
OursResNet-50960 × 51289.0199.0899.7699.9314.6829.3
VoxelPose[47]HRNet384 × 38482.4498.5599.7499.9217.632.3
Faster VoxelPose [53]HRNet384 × 38481.6998.3899.6799.8318.7722.4
MvP [49]HRNet384 × 38490.4196.3297.3997.8916.342.8
TesseTrack† [42]HRNet384 × 38486.2498.2999.7299.5016.920.6
OursHRNet384 × 38489.3298.4899.7399.9415.9920.3
", + "image_path": "fe6665ec521cb2e8b9383bab871ec6355b361942fc40521f2ac15ccf3f67aa7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 77, + 298, + 517, + 437 + ], + "blocks": [ + { + "bbox": [ + 46, + 239, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 289 + ], + "type": "text", + "content": "Table 1: Pose estimation results on the CMU Panoptic dataset. Our method achieves the best MPJPE and AP while running at speed comparable to Faster VoxelPose. We evaluate our methods at " + }, + { + "bbox": [ + 46, + 239, + 547, + 289 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 46, + 239, + 547, + 289 + ], + "type": "text", + "content": " resolution on the Panoptic dataset, as well as the higher resolution used in other methods. We mark TesseTrack [42] with a " + }, + { + "bbox": [ + 46, + 239, + 547, + 289 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 239, + 547, + 289 + ], + "type": "text", + "content": " as their reported results are on a different data split, and the results in this table are from our best reproduction of the method, which is not public." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 298, + 517, + 437 + ], + "lines": [ + { + "bbox": [ + 77, + 298, + 517, + 437 + ], + "spans": [ + { + "bbox": [ + 77, + 298, + 517, + 437 + ], + "type": "table", + "html": "
MethodShelfCampus
Actor-1Actor-2Actor-3AverageActor-1Actor2Actor3Average
Belagiannis et al. [3]66.165.083.271.482.072.473.775.8
Ershadi et al. [15]93.375.994.888.094.292.984.690.6
Dong et al.[12]98.894.197.896.997.693.398.096.3
VoxelPose [47]99.394.197.697.097.693.898.896.7
Faster VoxelPose[53]99.496.097.597.696.594.197.996.2
PlaneSweepPose[32]99.396.598.097.998.493.799.097.0
MvP [49]99.395.197.897.498.294.197.496.6
TEMPO (Ours)99.096.398.298.097.795.597.997.3
", + "image_path": "6b6b1eea07101fc85b84f47d63dbd73911e50ce8c93a27bb5e961b5a50c0f235.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 454, + 546, + 479 + ], + "lines": [ + { + "bbox": [ + 46, + 454, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 546, + 479 + ], + "type": "text", + "content": "Table 2: PCP3D accuracy on the Campus and Shelf datasets. We follow the protocol of previous methods and train our backbone on synthetic heatmaps of ground-truth poses. Our method achieves results comparable to the state-of-the-art." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 499, + 180, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 180, + 511 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 180, + 511 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 513, + 288, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 288, + 669 + ], + "type": "text", + "content": "Following [47, 53, 56, 49] we use a ResNet-50 backbone pre-trained on the Panoptic dataset and follow [23] by also using a ResNet-50 backbone pre-trained on Human3.6M for the dataset-specific pose estimation results. We use HR-Net [45] pre-trained on COCO as the model backbone for the generalization experiments, as done in [42] rather than pose estimation backbones that are trained on multi-view datasets. All methods are trained on 8 NVIDIA A100 GPUs with batch size of 2 per GPU. We use Adam with a learning rate of 3e-4, with weight decay of 1e-4 and a linear decay schedule for 10 epochs. We measure FPS using a single A100 GPU, and our code is based off the MMPose [11] library. Additional architectural details are in the supplement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 676, + 182, + 688 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 676, + 182, + 688 + ], + "spans": [ + { + "bbox": [ + 47, + 676, + 182, + 688 + ], + "type": "text", + "content": "4.3. Pose Estimation Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "We first compare the results of our methods other state-of-the-art methods. On the Panoptic Studio dataset, we" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 499, + 547, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 547, + 703 + ], + "type": "text", + "content": "report results following [53, 47, 49] and use " + }, + { + "bbox": [ + 304, + 499, + 547, + 703 + ], + "type": "inline_equation", + "content": "960 \\times 512" + }, + { + "bbox": [ + 304, + 499, + 547, + 703 + ], + "type": "text", + "content": " images and initialize from a ResNet-50 [18, 51] checkpoint pretrained on the Panoptic dataset. We also evaluate our method using an HRNet [45] backbone pretrained on COCO with " + }, + { + "bbox": [ + 304, + 499, + 547, + 703 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 304, + 499, + 547, + 703 + ], + "type": "text", + "content": " images for a fair comparison with TesseTrack. In Table 1 we provide a complete comparison across the state-of-the-art methods, with baselines trained on the same image resolutions and backbones for completeness. TEMPO achieves significantly lower MPJPE across both resolutions and backbones, while running at 29.3 FPS, competitive with Faster VoxelPose. We attribute this performance increase to the smoother and more temporally consistent skeletons our model produces due to its incorporation of temporal context and temporal supervision. We also show in 2 that our model achieves performance competitive with the state of the art on the Campus and Shelf datasets." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "14755" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 79, + 216, + 140 + ], + "blocks": [ + { + "bbox": [ + 56, + 79, + 216, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 79, + 216, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 79, + 216, + 140 + ], + "type": "table", + "html": "
MethodMOTAIDF1MPJPE
VoxelTrack [55]98.4598.67-
Snipper [56]93.4085.5040.2
TEMPO (Ours)98.4293.6238.5
", + "image_path": "15873dd9c7c30b7c88d8cdfc7fa331e18f5a4e0323aa5d3ece6a2523d50ac0df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 154, + 225, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 154, + 225, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 225, + 198 + ], + "type": "text", + "content": "(a) Evaluation of tracking and forecasting on the CMU Panoptic dataset. Our method outperforms [56] by using multi-view information and is competitive with VoxelTrack. [55]" + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 252, + 70, + 558, + 148 + ], + "blocks": [ + { + "bbox": [ + 252, + 70, + 558, + 148 + ], + "lines": [ + { + "bbox": [ + 252, + 70, + 558, + 148 + ], + "spans": [ + { + "bbox": [ + 252, + 70, + 558, + 148 + ], + "type": "table", + "html": "
Training DatasetMPJPE (mm)
PanopticH.6MEgoHumansPanopticH.6MEgoHumans
✓14.6862.96119.8
✓✓36.5625.3108.3
✓✓✓42.832.348.9
", + "image_path": "478d7a8f24a18f9705e2390002a6584a0dace38d33eb1b47642c8f2990291a6b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 251, + 152, + 555, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 152, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 251, + 152, + 555, + 198 + ], + "type": "text", + "content": "(b) Evaluating TEMPO's ability to transfer. Unlike previous methods, TEMPO is able to train on multiple datasets and can perform reasonably given enough training data. While our model is able to effectively transfer to Human3.6M, it has difficulty with EgoHumans due likely due to its use of fisheye cameras." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 220, + 141, + 306 + ], + "blocks": [ + { + "bbox": [ + 52, + 220, + 141, + 306 + ], + "lines": [ + { + "bbox": [ + 52, + 220, + 141, + 306 + ], + "spans": [ + { + "bbox": [ + 52, + 220, + 141, + 306 + ], + "type": "image", + "image_path": "b90e72f1bdc81f6e5f73e22123e355022e51b621dd26a4a3a1b7396b2bd050b4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 143, + 220, + 233, + 306 + ], + "blocks": [ + { + "bbox": [ + 143, + 220, + 233, + 306 + ], + "lines": [ + { + "bbox": [ + 143, + 220, + 233, + 306 + ], + "spans": [ + { + "bbox": [ + 143, + 220, + 233, + 306 + ], + "type": "image", + "image_path": "a2b49a536fcca213c31206216f0d33b90ac0bd065aa7d1e6e7808a5ba16c4132.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 235, + 220, + 324, + 306 + ], + "blocks": [ + { + "bbox": [ + 235, + 220, + 324, + 306 + ], + "lines": [ + { + "bbox": [ + 235, + 220, + 324, + 306 + ], + "spans": [ + { + "bbox": [ + 235, + 220, + 324, + 306 + ], + "type": "image", + "image_path": "225adbddbf0adf933473298923ed665d435772b28e1193ee2e49ef0c9f6a36a3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 325, + 220, + 416, + 306 + ], + "blocks": [ + { + "bbox": [ + 325, + 220, + 416, + 306 + ], + "lines": [ + { + "bbox": [ + 325, + 220, + 416, + 306 + ], + "spans": [ + { + "bbox": [ + 325, + 220, + 416, + 306 + ], + "type": "image", + "image_path": "ddee1d438f47c7ed94bc8d3207f4da99987f22740c281725d1935b31a47d86bd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 417, + 220, + 501, + 306 + ], + "blocks": [ + { + "bbox": [ + 417, + 220, + 501, + 306 + ], + "lines": [ + { + "bbox": [ + 417, + 220, + 501, + 306 + ], + "spans": [ + { + "bbox": [ + 417, + 220, + 501, + 306 + ], + "type": "image", + "image_path": "60b41b3ccb523c52fd24a94126829f206cebfc4b2c24038bb42fb2347f38dd7c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 51, + 307, + 141, + 390 + ], + "blocks": [ + { + "bbox": [ + 51, + 307, + 141, + 390 + ], + "lines": [ + { + "bbox": [ + 51, + 307, + 141, + 390 + ], + "spans": [ + { + "bbox": [ + 51, + 307, + 141, + 390 + ], + "type": "image", + "image_path": "b2724d89a55a6ebe7751fb7f267b2e6c6f3beff29bb4332771639fb539e058fb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 143, + 307, + 233, + 389 + ], + "blocks": [ + { + "bbox": [ + 143, + 307, + 233, + 389 + ], + "lines": [ + { + "bbox": [ + 143, + 307, + 233, + 389 + ], + "spans": [ + { + "bbox": [ + 143, + 307, + 233, + 389 + ], + "type": "image", + "image_path": "8652d7a7a96381970d11f062ceed8a07b322ebc8ad479b7c7ae41f8be3b6604f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 235, + 307, + 324, + 390 + ], + "blocks": [ + { + "bbox": [ + 235, + 307, + 324, + 390 + ], + "lines": [ + { + "bbox": [ + 235, + 307, + 324, + 390 + ], + "spans": [ + { + "bbox": [ + 235, + 307, + 324, + 390 + ], + "type": "image", + "image_path": "7bd6b211d868a076529a179ae772b3c14543c5e0652ca9f1662164b40f9e3bb4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 326, + 307, + 417, + 390 + ], + "blocks": [ + { + "bbox": [ + 326, + 307, + 417, + 390 + ], + "lines": [ + { + "bbox": [ + 326, + 307, + 417, + 390 + ], + "spans": [ + { + "bbox": [ + 326, + 307, + 417, + 390 + ], + "type": "image", + "image_path": "1a3fe3dcaafe6364e73022251465e75e7d324134aae831c42e19c9046e5ede63.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 419, + 311, + 503, + 391 + ], + "blocks": [ + { + "bbox": [ + 419, + 311, + 503, + 391 + ], + "lines": [ + { + "bbox": [ + 419, + 311, + 503, + 391 + ], + "spans": [ + { + "bbox": [ + 419, + 311, + 503, + 391 + ], + "type": "image", + "image_path": "76c64d15073a65cf18915a247642f36ebe2521ce1afef6d90d99579fa3d7f9bd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 51, + 392, + 141, + 473 + ], + "blocks": [ + { + "bbox": [ + 51, + 392, + 141, + 473 + ], + "lines": [ + { + "bbox": [ + 51, + 392, + 141, + 473 + ], + "spans": [ + { + "bbox": [ + 51, + 392, + 141, + 473 + ], + "type": "image", + "image_path": "64dcd246526b9452aa01e3cd343c6d59b3f19fa5b44b0d24b23f30550c5740d7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 493, + 545, + 518 + ], + "lines": [ + { + "bbox": [ + 46, + 493, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 545, + 518 + ], + "type": "text", + "content": "Figure 4: Samples of our model's pose estimation performance on the Panoptic Studio, Human3.6M, and EgoHumans datasets. TEMPO predicts accurate poses and tracks them over time." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 143, + 392, + 233, + 473 + ], + "blocks": [ + { + "bbox": [ + 143, + 392, + 233, + 473 + ], + "lines": [ + { + "bbox": [ + 143, + 392, + 233, + 473 + ], + "spans": [ + { + "bbox": [ + 143, + 392, + 233, + 473 + ], + "type": "image", + "image_path": "e1c6fdeece528e16cae31aad4301d5f1ea1ed4d513a045197b07275f79fecd9e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 235, + 392, + 324, + 474 + ], + "blocks": [ + { + "bbox": [ + 235, + 392, + 324, + 474 + ], + "lines": [ + { + "bbox": [ + 235, + 392, + 324, + 474 + ], + "spans": [ + { + "bbox": [ + 235, + 392, + 324, + 474 + ], + "type": "image", + "image_path": "fd5b6b064e74e701cd561cb30dfa43fa277fc10f5d6e36643bb4d7a3fcf06df8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 325, + 392, + 416, + 474 + ], + "blocks": [ + { + "bbox": [ + 325, + 392, + 416, + 474 + ], + "lines": [ + { + "bbox": [ + 325, + 392, + 416, + 474 + ], + "spans": [ + { + "bbox": [ + 325, + 392, + 416, + 474 + ], + "type": "image", + "image_path": "1d8b13759dedc4460dde312686d3b0643330ff80f8bc5923966e3c5a99195416.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 419, + 395, + 517, + 474 + ], + "blocks": [ + { + "bbox": [ + 419, + 395, + 517, + 474 + ], + "lines": [ + { + "bbox": [ + 419, + 395, + 517, + 474 + ], + "spans": [ + { + "bbox": [ + 419, + 395, + 517, + 474 + ], + "type": "image", + "image_path": "7e838f1aaa58fa324e0966d485730bea93e4d61c0c3a5fdbd672f6ed2ef67ec2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 539, + 289, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 289, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 289, + 658 + ], + "type": "text", + "content": "We compare pose estimation performance on the Human3.6M and EgoHumans datasets in Table 4. Our results perform comparably to the state-of-the-art. Notably [23] uses ground-truth locations and cropped bounding boxes from input views, while our method is able to match performance despite simultaneously detecting people in the scene. Furthermore, our method significantly outperforms others in the more challenging EgoHumans benchmark, suggesting that temporal context is crucial for handling rapid motion." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 670, + 214, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 214, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 214, + 683 + ], + "type": "text", + "content": "4.4. Pose Tracking and Forecasting" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "We compared the performance of TEMPO's tracking to VoxelTrack [55] in Table 3a. Our tracker is competitive but" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 539, + 547, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 547, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 547, + 707 + ], + "type": "text", + "content": "performs slightly worse, which is expected due to its lack of learned Re-ID features. To our knowledge, pose forecasting has not been attempted for the multi-view case, so we compare against the closest forecasting method, Snipper [56], which estimates future pose from monocular video. Our method takes as input 4 previous timesteps and predicts the pose over the next 3 timesteps, which is 0.33 seconds into the future, matching prior work [56]. Our model achieves state-of-the-art performance, shown in Table 3a. We conducted an additional experiment to measure the performance of our model in transfer on different datasets. The standard practice in monocular 3D pose estimation is to train on and evaluate performance on a combination of multiple datasets. However, the predominant paradigm for" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14756" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 273, + 243 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 273, + 243 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 273, + 243 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 273, + 243 + ], + "type": "table", + "html": "
MethodHuman 3.6MEgoHumans
Martinez et al. [35]57.0-
Pavlakos et al. [37]56.9-
Kadkhodamohammadi et al. [26]49.1-
Ma et al [33]24.4-
VoxelPose [47]19.046.24
Faster VoxelPose [53]19.848.50
MvP [49]18.641.72
Iskakov et al. [23]17.7-
TEMPO (Ours)18.536.74
", + "image_path": "e447783c1117f61b2f511ddbddee257e3a6151257d8375faffeb64278e863332.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 344, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 608 + ], + "type": "text", + "content": "evaluating multi-view pose estimation methods has been to train a single model only on a single dataset and evaluate it on the same dataset. This severely limits the potential of these models to generalize across different settings and thus be deployed in real-world scenarios. Similar to [23], we evaluate the performance of our model trained on multiple combinations of annotated datasets, and report the results for each combination in Table 3b. We run our model with no fine tuning and the same voxel size of " + }, + { + "bbox": [ + 46, + 344, + 287, + 608 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}^3" + }, + { + "bbox": [ + 46, + 344, + 287, + 608 + ], + "type": "text", + "content": " across each dataset. Our method is able to transfer and successfully tracks and predicts pose, but performs noticeably worse, likely due to being trained on a single camera configuration. In particular, the CMU Panoptic training dataset uses 5 cameras, whereas Human3.6M uses 4 and EgoHumans uses 8 fisheye cameras. The model has the most trouble generalizing to EgoHumans, likely due to the significantly larger indoor space and different camera models. We find that TEMPO performs better in transfer after including training data from each dataset, especially on the EgoHumans dataset, suggesting that future works should include diverse multi-view data from different camera configurations and camera models in order to better generalize." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 616, + 115, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 115, + 627 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 115, + 627 + ], + "type": "text", + "content": "4.5. Ablations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "We train ablated models to study the impact of individual components in our method. All our experiments are conducted on the CMU Panoptic dataset, with results shown in table 5. We find that using temporal information slightly helps, but with per-timestep supervision, the model is able to greatly improve. Warping the pose between timesteps further improves the model. We hypothesize that the im" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 549, + 167 + ], + "blocks": [ + { + "bbox": [ + 46, + 262, + 287, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 321 + ], + "type": "text", + "content": "Table 4: Pose Estimation results on the Human3.6M and EgoHumans datasets in MPJPE (mm). Our method is competitive with the state-of-the-art on Human3.6M and surpasses current methods by a significant margin on EgoHumans." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 70, + 549, + 167 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 549, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 549, + 167 + ], + "type": "table", + "html": "
MethodTForecastingWarpingPer-t lossMPJPE ↓(mm)
(a)317.83
(b)3✓15.03
(c)3✓✓14.94
(d)3✓✓✓14.68
(e)4✓✓✓14.90
(f)5✓✓✓14.82
", + "image_path": "b866d62bbb3de1b5d9897d85e82664dc34aead5d38481c147b377e617f972d2c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 266, + 536, + 297 + ], + "blocks": [ + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "content": "Table 5: Ablation study on various components to our model. The most important component to performance was the per-timestep supervision. Warping the previous feature also improved performance. We observed the forecasting and slightly increasing the length of the input history had no noticeable effect on performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 266, + 536, + 297 + ], + "lines": [ + { + "bbox": [ + 315, + 266, + 536, + 297 + ], + "spans": [ + { + "bbox": [ + 315, + 266, + 536, + 297 + ], + "type": "table", + "html": "
Cameras12345
MPJPE51.3232.1319.2217.3414.68
", + "image_path": "29543388f5f1b735d3a6b71bdcb9a88d1fad1789ede8f304863366e502d67bd5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "text", + "content": "Table 6: Ablation on number of cameras. We observe that performance decreases with decreasing number of cameras." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 359, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 545, + 479 + ], + "type": "text", + "content": "provement from warping is small due to the relative lack of motion in the Panoptic dataset - the distance between body centers in consecutive timesteps is usually small. We also measure the effect of the history length on performance and found no significant difference. While intuitively, larger history lengths should provide more context, the GPU memory constraints of our method prevent investigating " + }, + { + "bbox": [ + 304, + 359, + 545, + 479 + ], + "type": "inline_equation", + "content": "T > 5" + }, + { + "bbox": [ + 304, + 359, + 545, + 479 + ], + "type": "text", + "content": ". We also ablated on the number of cameras on the Panoptic dataset. We found that the MPJPE increases with the number of cameras, matching the findings of [47, 23] and [42]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 500, + 383, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 383, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 383, + 513 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 524, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 524, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 545, + 656 + ], + "type": "text", + "content": "Understanding human behavior from video is a fundamentally temporal problem, requiring accurate and efficient pose estimation algorithms that can reason over time. We presented the first method that satisfies these requirements, achieving state-of-the-art results over existing multi-view pose estimation benchmarks via temporal consistency as a learning objective. Our model is also highly efficient, relying on recurrence to maintain a temporal state while enabling pose tracking and forecasting. TEMPO represents a step closer towards general-purpose human behavior understanding from video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 677, + 408, + 690 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 408, + 690 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 408, + 690 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 701, + 515, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 515, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 515, + 714 + ], + "type": "text", + "content": "This research was supported partially by Fujitsu." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14757" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "text", + "content": "[1] Sikandar Amin, Mykhaylo Andriluka, Marcus Rohrbach, and Bernt Schiele. Multi-view pictorial structures for 3d human pose estimation. In *Bmvc*, volume 1. Bristol, UK, 2013." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 287, + 169 + ], + "type": "text", + "content": "[2] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. arXiv preprint arXiv:1511.06432, 2015." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 215 + ], + "type": "text", + "content": "[3] Vasileios Belagiannis, Sikandar Amin, Mykhaylo Andriluka, Bernt Schiele, Nassir Navab, and Slobodan Ilic. 3d pictorial structures revisited: Multiple human pose estimation. IEEE Trans. Pattern Anal. Mach. Intell., 38(10):1929-1942, 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 261 + ], + "type": "text", + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 338 + ], + "type": "text", + "content": "[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 561-578. Springer, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 396 + ], + "type": "text", + "content": "[6] Lewis Bridgeman, Marco Volino, Jean-Yves Guillemaut, and Adrian Hilton. Multi-person 3d pose estimation and tracking in sports. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "type": "text", + "content": "[7] Simon Bultmann and Sven Behnke. Real-time multi-view 3d human pose estimation using semantic feedback to smart edge sensors. arXiv preprint arXiv:2106.14729, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 496 + ], + "type": "text", + "content": "[8] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 387-404. Springer, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 553 + ], + "type": "text", + "content": "[9] Zhuo Chen, Xu Zhao, and Xiaoyue Wan. Structural triangulation: A closed-form solution to constrained 3d human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 695-711. Springer, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[10] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2262-2271, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 633 + ], + "type": "text", + "content": "[11] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[12] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[13] Junting Dong, Wen Jiang, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose estimation from multiple views. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 545, + 95 + ], + "type": "text", + "content": "conference on computer vision and pattern recognition, pages 7792-7801, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[14] Zijian Dong, Jie Song, Xu Chen, Chen Guo, and Otmar Hilliges. Shape-aware multi-person pose estimation from multi-view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11158-11168, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[15] Sara Ershadi-Nasab, Erfan Noury, Shohreh Kasaei, and Esmaeil Sanaei. Multiple human 3d pose estimation from multiview images. Multimedia Tools and Applications, 77:15573-15601, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 197, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 252 + ], + "type": "text", + "content": "[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 59-75. Springer, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "type": "text", + "content": "[17] Adam W. Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. Simple-BEV: What really matters for multi-sensor bev perception? In arXiv:2206.07959, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 332, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 375 + ], + "type": "text", + "content": "[19] Yihui He, Rui Yan, Katerina Fragkiadaki, and Shouu-I Yu. Epipolar transformers. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7779-7788, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 377, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 443 + ], + "type": "text", + "content": "[20] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15273-15282, October 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "text", + "content": "[21] Catalin Ionescu, Fuxin Li, and Cristian Sminchisescu. Latent structured models for human pose estimation. In 2011 International Conference on Computer Vision, pages 2220-2227. IEEE, 2011." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 489, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 544 + ], + "type": "text", + "content": "[22] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "text", + "content": "[23] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In International Conference on Computer Vision (ICCV), 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 580, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 633 + ], + "type": "text", + "content": "[24] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European conference on computer vision (ECCV), pages 690-706, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[25] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, 2015." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[26] Abdolrahim Kadkhodamohammadi and Nicolas Padoy. A generalizable approach for multi-view 3d human pose regression. Machine Vision and Applications, 32(1):6, 2021." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14758" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. corr abs/1712.06584 (2017). arXiv preprint arXiv:1712.06584, 1(3), 2017." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 163 + ], + "type": "text", + "content": "[28] Rawal Khirodkar, Aayush Bansal, Lingni Ma, Richard Newcombe, Minh Vo, and Kris Kitani. Egohumans: An egocentric 3d multi-human benchmark. arXiv preprint arXiv:2305.16487, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 165, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 165, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 165, + 287, + 219 + ], + "type": "text", + "content": "[29] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 222, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 287, + 255 + ], + "type": "text", + "content": "[30] Ilya Kostrikov and Juergen Gall. Depth sweep regression forests for estimating 3d human pose from images. In BMVC, volume 1, page 5. Nottingham, UK, 2014." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 257, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 257, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 287, + 311 + ], + "type": "text", + "content": "[31] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 314, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 358 + ], + "type": "text", + "content": "[32] Jiahao Lin and Gim Hee Lee. Multi-view multi-person 3d pose estimation with plane sweep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11886-11895, June 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 437 + ], + "type": "text", + "content": "[33] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 424-442. Springer, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 439, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 439, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 439, + 287, + 483 + ], + "type": "text", + "content": "[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9489-9497, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "type": "text", + "content": "[35] Julieta Martinez, Rayat Hossain, Javier Romero, and James J. Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "text", + "content": "[36] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 620 + ], + "type": "text", + "content": "[37] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Harvesting multiple views for marker-less 3d human pose annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6988-6997, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 666 + ], + "type": "text", + "content": "[38] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proceedings of the European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[39] Haibo Qiu, Chunyu Wang, Jingdong Wang, Naiyan Wang, and Wenjun Zeng. Cross view fusion for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4342-4351, 2019." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people with 3d representations. arXiv preprint arXiv:2111.07868, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "type": "text", + "content": "[41] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, and Jitendra Malik. Tracking people by predicting 3d appearance, location and pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2740-2749, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[42] N Dinesh Reddy, Laurent Guigues, Leonid Pischulin, Jayan Eledath, and Srinivasa G. Narasimhan. Tessentrack: End-to-end learnable multi-person articulated 3d pose tracking. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In CVPR, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "type": "text", + "content": "[44] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "text", + "content": "[45] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[46] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11179-11188, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "text", + "content": "[47] Hanyue Tu, Chunyu Wang, and Wenjun Zeng. Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In European Conference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 455, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 509 + ], + "type": "text", + "content": "[48] Hsiao-Yu Fish Tung, Ricson Cheng, and Katerina Fragkiadaki. Learning spatial common sense with geometry-aware recurrent networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2595-2603, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "text", + "content": "[49] Tao Wang, Jianfeng Zhang, Yujun Cai, Shuicheng Yan, and Jiashi Feng. Direct multi-view multi-person 3d human pose estimation. Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "type": "text", + "content": "[50] Size Wu, Sheng Jin, Wentao Liu, Lei Bai, Chen Qian, Dong Liu, and Wanli Ouyang. Graph-based 3d multi-person pose estimation using multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11148-11157, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "text", + "content": "[51] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision (ECCV), 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[52] Yan Xu and Kris Kitani. Multi-view multi-person 3d pose estimation with uncalibrated camera networks. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMVA Press, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[53] Hang Ye, Wentao Zhu, Chunyu Wang, Rujie Wu, and Yizhou Wang. Faster voxelpose: Real-time 3d human pose estimate" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14759" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 251 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "tion by orthographic projection. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[54] Yuxiang Zhang, Liang An, Tao Yu, Xiu Li, Kun Li, and Yebin Liu. 4d association graph for realtime multi-person motion capture using multiple video cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1324-1333, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 205 + ], + "type": "text", + "content": "[55] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenyu Liu, and Wenjun Zeng. Voxeltrack: Multi-person 3d human pose estimation and tracking in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2613-2626, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[56] Shihao Zou, Yuanlu Xu, Chao Li, Lingni Ma, Li Cheng, and Minh Vo. Snipper: A spatiotemporal transformer for simultaneous multi-person 3d pose estimation tracking and forecasting on a video snippet, 2022." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14760" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_content_list.json b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eaff2640cb6ab74ededb1321db618a2475509731 --- /dev/null +++ b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_content_list.json @@ -0,0 +1,1785 @@ +[ + { + "type": "text", + "text": "TF-ICON: Diffusion-Based Training-Free Cross-Domain Image Composition", + "text_level": 1, + "bbox": [ + 96, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shilin Lu1 Yanzhu Liu2 Adams Wai-Kin Kong1", + "bbox": [ + 282, + 180, + 697, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Nanyang Technological University, Singapore $^{2}$ Institute for Infocomm Research (I²R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore", + "bbox": [ + 66, + 199, + 895, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shilin002@e.ntu.edu.sg, liu_yanzhu@i2r.a-star.edu.sg, adamskong@ntu.edu.sg", + "bbox": [ + 163, + 237, + 818, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d26ed3cf3c91eb696cd946fd14db40e6479a9ab39befc57b3e2c7728153d0cec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 253, + 161, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a71505ddcf5c4341f7eab4a1f77c152f11f92b046ef8504485bc601b156a9911.jpg", + "image_caption": [ + "' a professional photograph of an Eiffel tower in the distance, ultra realistic" + ], + "image_footnote": [], + "bbox": [ + 163, + 253, + 321, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/70eea039ca2f245f29c9c5f0094a0da7cc9140d1f8963b3d7573cbbd9e19a7dd.jpg", + "image_caption": [ + "' a professional photograph of a puppy on the beach, ultra realistic'." + ], + "image_footnote": [], + "bbox": [ + 323, + 253, + 482, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ab917b90888d14ede31cb0f51afa86b19cb0afd288ee6917dede397707926f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 253, + 565, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eeb42dd5d159fe37c37357571000d991dd7d7e6278d1fcceff62582039928582.jpg", + "image_caption": [ + "'an oil painting of a hamburger, Van Gogh Style'" + ], + "image_footnote": [], + "bbox": [ + 566, + 253, + 725, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1ad3be380ad420bde06da0ac289b5760085ac199238e925f4c976d68f6f98ac.jpg", + "image_caption": [ + "'an oil painting of a waffle, Van Gogh Style'." + ], + "image_footnote": [], + "bbox": [ + 728, + 253, + 887, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd8f399f2be6462eb7f43715f2e0a73f0383566f77cfa134538a887cd5754a7b.jpg", + "image_caption": [ + "(c) Composition between real & sketchy painting domains", + "Figure 1: Image composition aims to seamlessly blend distinct objects into a specific visual context. Our training-free framework equips attention-based text-driven diffusion models with the capability to achieve this task across various domains (a) photorealism, (b) oil painting, (c) sketching, and (d) cartoon animation, within 20 sampling steps." + ], + "image_footnote": [], + "bbox": [ + 81, + 410, + 161, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3fea9fed12136a363be86ab142e4491af50fd65e07a9f8e735e70fe72f80ea0e.jpg", + "image_caption": [ + "(a) Composition within the photorealism domain", + "' a pencil drawing of a sheep in the sunset" + ], + "image_footnote": [], + "bbox": [ + 163, + 410, + 321, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6843d479466ac00bc61107fd8bd46e9fc315db2b10fa359d57b07e969644329e.jpg", + "image_caption": [ + "' a pencil drawing of a panda in the sunset" + ], + "image_footnote": [], + "bbox": [ + 323, + 410, + 482, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8de6c7ddb49df72fe540feed7935bf14f448769a5ebcaf3cc5c9a27b9d7e9d96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 410, + 565, + 531 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/523e81e026b52099b78832bca53b534810a59a469f9ba6d7e571b6554c898f1d.jpg", + "image_caption": [ + "' a cartoon animation of buildings in the distance'", + "(d) Composition between real & cartoon domains" + ], + "image_footnote": [], + "bbox": [ + 576, + 410, + 725, + 531 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54d85347f8cfb11d4f2f504c98b0e59732ceea6f5a217a40e3ed8e8ed9e067ef.jpg", + "image_caption": [ + "' a cartoon animation of an Eiffel tower in the distance'" + ], + "image_footnote": [], + "bbox": [ + 728, + 410, + 887, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 625, + 313, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-driven diffusion models have exhibited impressive generative capabilities, enabling various image editing tasks. In this paper, we propose TF-ICON, a novel Training-Free Image CompositioN framework that harnesses the power of text-driven diffusion models for cross-domain image-guided composition. This task aims to seamlessly integrate user-provided objects into a specific visual context. Current diffusion-based methods often involve costly instance-based optimization or finetuning of pretrained models on customized datasets, which can potentially undermine their rich prior. In contrast, TF-ICON can leverage off-the-shelf diffusion models to perform cross-domain image-guided composition without requiring additional training, finetuning, or optimization. Moreover, we introduce the exceptional prompt, which contains no information, to facilitate text-driven diffusion models in accurately inverting real images into latent representations,", + "bbox": [ + 73, + 643, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "forming the basis for compositing. Our experiments show that equipping Stable Diffusion with the exceptional prompt outperforms state-of-the-art inversion methods on various datasets (CelebA-HQ, COCO, and ImageNet), and that TF-ICON surpasses prior baselines in versatile visual domains. Code is available at https://github.com/Shilin-LU/TF-ICON", + "bbox": [ + 496, + 627, + 893, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 739, + 632, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image composition task involves incorporating unique objects from different photos to create a harmonious image within a specific visual context, a.k.a. image-guided composition. For instance, consider the scenario where one desires to incorporate a beloved panda into one's favorite artwork, e.g., oil or sketchy painting. The objective is to create a new image where the panda blends seamlessly into the scene without altering the appearance of the panda and the background, just as an artist meticulously crafted this panda", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2294", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for that artwork (See Figure 1). This task is inherently challenging, as it requires maintaining illumination consistency and preserving identifying features. The challenge is further compounded when the photos come from various domains.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While recently large-scale text-to-image models [10, 19, 51, 56, 59, 61, 78] have achieved remarkable success in text-driven image generation, the ambiguity inherent in natural language presents challenges in conveying precise and nuanced visual details, even with highly detailed text prompts. Although this challenge is effectively addressed by enabling personalized concept learning [24, 25, 34, 37, 60], these methods require costly instance-based optimization and are limited in generating concepts with specified backgrounds. Recent studies [67, 77] have shown that diffusion models can achieve image-guided composition by explicitly incorporating additional guiding images. However, these models are retrained from the pretrained diffusion model on tailored datasets, which can damage the rich prior of the model. As a result, these models have limited compositional abilities beyond their training domain and still require significant computational resources.", + "bbox": [ + 75, + 154, + 470, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the wealth of large text-to-image models that have been trained on extensive language-image datasets, we pose a question: how could these models be leveraged for image-guided composition without incurring costly training or finetuning, thereby avoiding damaging the diverse prior? To answer it, we propose the Training-Free Image CompositionN (TF-ICON) framework, which equips attention-based text-to-image diffusion models with the capability to perform image-guided composition without requiring additional training, fine-tuning, extra data, or optimization. To the best of our knowledge, this is the first training-free framework developed for image-guided composition. The framework is compatible with various diffusion model samplers, enabling completion within 20 steps, and harnesses rich semantic knowledge to facilitate image-guided compositions across diverse domains (see Figure 1).", + "bbox": [ + 75, + 414, + 468, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach constitutes an image-guided composition interface through denoising from a reliable starting latent code with the injection of composite self-attention maps. Finding the latent code that allows for reconstructing an input image while maintaining its editability, a.k.a. image inversion, is a challenging yet crucial step for state-of-the-art (SOTA) image editing frameworks involving real images [15, 26, 35, 38, 48, 53, 54, 70]. For diffusion models, while denoising diffusion implicit models (DDIM) inversion [65] has been effective for unconditional diffusion models, it falls short for text-driven diffusion models [26, 50, 70, 71]. To circumvent this, we introduce the exceptional prompt to accurately invert real images into latent codes upon pretrained text-to-image models to serve for further composition generation. The accurate latent codes are composed as the starting noise for the diffusion process. Through the", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gradual injection of composite self-attention maps that are specifically designed to reflect the relations between guiding images, we are able to infuse contextual information from the background into the incorporated objects, which results in harmonious image-guided compositions.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we make the following key contributions:", + "bbox": [ + 517, + 167, + 890, + 183 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We demonstrate the superior performance of high-order diffusion ODE solvers compared to commonly used DDIM inversion for real image inversion.", + "2. We present an exceptional prompt that allows text-driven models to achieve accurate invertibility, laying a solid groundwork for subsequent editing. Experimental results show that it surpasses SOTA inversion methods on three vision datasets.", + "3. We propose the first training-free framework that enables cross-domain image-guided composition for attention-based diffusion models.", + "4. We demonstrate quantitatively and qualitatively that our framework outperforms prior baselines for image-guided composition." + ], + "bbox": [ + 511, + 190, + 890, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 419, + 640, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image composition. Image composition is widely applied to electronic commerce, entertainment, and data augmentation [20, 43] for downstream tasks. It can be broadly categorized into two types: text-guided [5, 6, 11, 22, 45] and image-guided [8, 23, 39, 67, 76, 77, 79]. The former involves composing multiple objects specified by only a text prompt without limiting the appearance of objects, as long as their semantics align with the prompt. Despite the great successes of text-conditioned models, they are often prone to semantic errors [22, 56], especially when the text prompt involves multiple objects. These errors include attribute leakage, attribute interchange, and missing objects, which cause the generated images to critically different from the user's intention [22, 56]. As a result, extensive prompt engineering [74] is often necessary to achieve the desired results. In contrast to text-only guided composition, image-guided composition involves incorporating specific objects and scenarios from user-provided photos, potentially with the aid of a text prompt. However, due to the nature of involving additional real images, it is more challenging particularly when images from different visual domains. Conventionally, image-guided composition is divided into several sub-tasks [52], such as object placement [7, 12, 40, 69, 80], image blending [75, 81], image harmonization [14, 16, 30, 76], and shadow generation [29, 42, 63, 83], each of which is typically addressed by different models and pipelines.", + "bbox": [ + 496, + 446, + 890, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image inversion. Extensive research has been conducted on image inversion for GANs, including latent-based op", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2295", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/43a6bd7a936544f4d728367a28a90fd28694c1e04433d528b991790bc4dbcfeb.jpg", + "image_caption": [ + "(a) Image Inversion" + ], + "image_footnote": [], + "bbox": [ + 80, + 82, + 478, + 165 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7fc4b2d490e4202927ec7ac486839f44305a5ac31372b5187ea219137b06ab2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 83, + 887, + 166 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/707be35a11950f6fa56ef34f8f7b924d43924b6f2a53365461653c62ad4f8c89.jpg", + "image_caption": [ + "(b) Training-Free Image Composition (TF-ICON)", + "Figure 2: The proposed training-free image composition framework. (a) The exceptional prompt is used to invert the main and reference images into noises $\\mathbf{x}_T^{\\mathrm{m}}$ , $\\mathbf{x}_T^{\\mathrm{r}}$ , which are then composed to form the starting point $\\mathbf{x}_T^*$ for the composition process. (b) Three constituents are composed for injecting into the composition process at early timesteps, including self-attention maps from the main and reference image reconstruction processes, as well as cross attention between the main and reference images. For better clarity and readability, the original main and reference images are shown in the pixel space instead of the VAE latent space, and the reference image is presented without resizing and zero-padding." + ], + "image_footnote": [], + "bbox": [ + 83, + 172, + 888, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tization [1, 2, 33], encoders [3, 57, 68], and fine-tuning [4, 58]. For diffusion models, DDIM [65] is a widely used technique for inversion in image editing frameworks. However, in text-driven settings, DDIM leads to significant reconstruction distortion due to the instability resulting from classifier-free guidance (CFG) [18, 28]. Recently, null-text inversion [50] has been proposed to achieve accurate inversion by optimizing the unconditional prediction of the text-to-image model. It demonstrates promising results but requires instance-based optimization. Concurrently, EDICT [71] also achieves near-perfect inversion, albeit doubling the computation time of the diffusion process.", + "bbox": [ + 75, + 545, + 468, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 739, + 202, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion probabilistic models (DPM) [18, 27, 59, 64] are generative models in which an image is generated by progressively denoising from Gaussian noise. The forward diffusion process gradually perturbs data with infinite noise scales, which can be modeled as the solution of a stochastic differential equation (SDE) $\\{\\mathbf{x}_t\\}_{t=0}^T$ . Formally, given a data sample $\\mathbf{x}_0 \\sim p_0 = p_{\\mathrm{data}}$ , random noise is gradually injected, eventually resulting in a sample $\\mathbf{x}_T$ which is typically distributed as a tractable prior $p_T$ without any infer", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mation of $p_0$ , as described by the following SDE [36, 66]:", + "bbox": [ + 500, + 545, + 880, + 561 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} \\mathbf {x} _ {t} = \\mathbf {f} \\left(\\mathbf {x} _ {t}, t\\right) \\mathrm {d} t + g (t) \\mathrm {d} \\mathbf {w} _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 577, + 890, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{w}_t\\in \\mathbb{R}^d$ is the standard Wiener process (a.k.a., Brownian motion), and $\\mathbf{f}(\\cdot ,t)$ and $g(t)$ are commonly designated as the drift and diffusion coefficient, respectively. On the other hand, the reverse diffusion process can be described by the reverse-time SDE from $T$ to 0 [66]:", + "bbox": [ + 496, + 601, + 890, + 676 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} \\mathbf {x} _ {t} = [ \\mathbf {f} (\\mathbf {x} _ {t}, t) - g (t) ^ {2} \\nabla_ {\\mathbf {x}} \\log p _ {t} (\\mathbf {x} _ {t}) ] \\mathrm {d} t + g (t) \\mathrm {d} \\bar {\\mathbf {w}} _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 689, + 890, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{w}}_t$ is the Wiener process in the reverse time. The score function $\\nabla_{\\mathbf{x}}\\log p_t(\\mathbf{x}_t)$ is the only unknown term and can be estimated by a neural network $\\epsilon_{\\theta}(\\mathbf{x}_t,t)$ whose parameter $\\theta$ is optimized by a denoising objective [27, 66].", + "bbox": [ + 496, + 719, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Upon attaining the trained model that predicts the score function accurately, it can be utilized to numerically solve the reverse SDE (Eq. (2)), enabling the generation of samples from a noise distribution. Song et al. [66] outline various methods, including Variance Exploding (VE), Variance Preserving (VP), and sub-VP SDE, for constructing SDEs that perturb the unknown data distribution into a fixed prior. In this work, we leverage the pre-trained text-to-image La", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2296", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tent Diffusion Model (LDM) [59], a.k.a. Stable Diffusion, which applies the VP SDE in the latent space.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 133, + 166, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our objective is to utilize a main (background) image $\\mathbf{I}^{\\mathrm{m}}$ , a reference (foreground) image $\\mathbf{I}^{\\mathrm{r}}$ , a text prompt $\\mathcal{P}$ , and a binary mask $\\mathbf{M}^{\\mathrm{user}}$ which designates the region of interest within the main image, to generate a modified image $\\mathbf{I}^{*}$ . The resultant image $\\mathbf{I}^{*}$ should contain the reference subject with identifying features within the mask, i.e. $id(\\mathbf{I}^{*} \\odot \\mathbf{M}^{\\mathrm{user}}) \\approx id(\\mathbf{I}^{\\mathrm{r}})$ , while concurrently ensuring that the complementing area closely resembles the main image, i.e. $\\mathbf{I}^{*} \\odot (\\mathbf{1} - \\mathbf{M}^{\\mathrm{user}}) \\approx \\mathbf{I}^{\\mathrm{m}} \\odot (\\mathbf{1} - \\mathbf{M}^{\\mathrm{user}})$ . Moreover, it is ideal for the transition between the areas inside and outside the mask to be imperceptible.", + "bbox": [ + 75, + 159, + 468, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a training-free framework that can make use of attention-based pre-trained text-to-image models to perform image-guided composition. To the best of our knowledge, it is the first training-free framework for image-guided composition, which can be accomplished within 20 steps of sampling. The framework is mainly comprised of two steps: image inversion (Section 4.1), and composition generation (Section 4.2), as shown in Figure 2. The full algorithm is presented in Appendix B.2.", + "bbox": [ + 75, + 325, + 470, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Image Inversion with Exceptional Prompt", + "text_level": 1, + "bbox": [ + 76, + 469, + 434, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Achieving precise manipulation of real images often necessitates an accurate inversion process that identifies the corresponding latent representation, which not only provides editability for meaningful manipulation but also accurately reconstructs the input image [26, 68]. For diffusion models, optimal editability is typically characterized by a noise encoding that conforms to the ideal statistical properties of zero-mean, unit-variance Gaussian noise [53].", + "bbox": [ + 75, + 493, + 468, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ODE inversion. Most diffusion frameworks for image editing [15, 26, 35, 38, 53, 70] use DDIM inversion to invert the real image into its latent representation. However, our findings suggest that this may not be the optimal choice for inverting real images. It has been proven that DDIM is a first-order discretization of the associated probability flow ordinary differential equations (ODE) of Eq. (2) [62, 65, 66]:", + "bbox": [ + 75, + 628, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} \\mathbf {x} _ {t} = \\left[ \\mathbf {f} \\left(\\mathbf {x} _ {t}, t\\right) - \\frac {1}{2} g (t) ^ {2} \\nabla_ {\\mathbf {x}} \\log p _ {t} \\left(\\mathbf {x} _ {t}\\right) \\right] \\mathrm {d} t, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 743, + 468, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which can be solved using $\\epsilon_{\\theta}(\\mathbf{x}_t,t)$ and shares the consistent marginal probability distribution $\\{p_t(\\mathbf{x}_t)\\}_{t = 0}^T$ with Eq. (2). Various samplers [32, 44, 46, 47] have been developed for solving the diffusion ODE starting from noise $\\mathbf{x}_T$ to achieve fast sampling (10~20 steps). We offer the insight that utilizing these ODE solvers in turn as encoders starting from the real image $\\mathbf{x}_0$ yields better latent representation $\\mathbf{x}_T$ , compared with those obtained through commonly used", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3f46d3803dea2136fd0ad6e853be951448157d879c981d8fd87998538df68106.jpg", + "image_caption": [ + "(a) Original" + ], + "image_footnote": [], + "bbox": [ + 503, + 90, + 578, + 265 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58d7ffc51efc7a9fb2567a5e35f8c464bafd691b2a196c15014feef48344d7b3.jpg", + "image_caption": [ + "(b) CFG" + ], + "image_footnote": [], + "bbox": [ + 581, + 90, + 656, + 265 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c701696a7df507f13a8f500591db1780aac298926a05d5e765784902f68b9004.jpg", + "image_caption": [ + "(c) Cond." + ], + "image_footnote": [], + "bbox": [ + 658, + 90, + 733, + 265 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea450ef038a402b3765028bf1f78f33eb3abcdebfb9906bc3ab31d749af88288.jpg", + "image_caption": [ + "(d) Uncond." + ], + "image_footnote": [], + "bbox": [ + 733, + 90, + 812, + 265 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7f8820de931ef5f6ccc769ac3ab0caee3950b8dec11c3db166650896a4bb730e.jpg", + "image_caption": [ + "(e) Ours", + "Figure 3: The real image reconstruction results using Stable Diffusion with (b) classifier-free guidance (CFG) output $\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E},\\varnothing)$ ; (c) conditional output $\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E})$ ; (d) unconditional output $\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\varnothing)$ ; and (e) ours. The prompts for (b) and (c) are 'a photo of a singer', 'a photo of a car', and 'an oil painting'. See Appendix A.3 for elaboration." + ], + "image_footnote": [], + "bbox": [ + 813, + 90, + 888, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DDIM. A quantitative analysis is given in Appendix A.1. The enhanced alignment between the forward and backward ODE trajectories in the high-order DPM-Solver++ [47] implies that it is better suited for real image inversion. Thus, this paper employs it for all inversions of diffusion models.", + "bbox": [ + 498, + 381, + 890, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Exceptional prompt. In the unconditional setting $\\epsilon_{\\theta}(\\mathbf{x}_t,t)$ solving the diffusion ODE (Eq. (3)) from 0 to $T$ enables us to obtain the better latent code $\\mathbf{x}_T$ for the real image $\\mathbf{x}_0$ . However, in the text-driven setting $\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E})$ , existing image editing works [26, 50, 70, 71] have shown that the inversion process is prone to significant reconstruction errors, due to the instability induced by CFG [18, 28]:", + "bbox": [ + 498, + 472, + 890, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\mathcal {E}, \\varnothing\\right) = s \\cdot \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\mathcal {E}\\right) + (1 - s) \\cdot \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\varnothing\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 588, + 890, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varnothing = \\psi(\"\")$ and $\\mathcal{E} = \\psi(\\mathcal{P})$ are embeddings of the null and normal prompt, and $s$ is the guidance scale. Our experiments further reveal that even without CFG, both conditional output $\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E})$ and unconditional output $\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\emptyset)$ of text-to-image diffusion models still produce large reconstruction errors, as depicted in Figure 3.", + "bbox": [ + 496, + 613, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve accurate inversion, we present a straightforward yet effective solution, namely exceptional prompt, $\\mathcal{P}_{\\text{exceptional}}$ . Intuitively, any information contained within the input prompt can result in the deviation of the backward ODE trajectories from the forward trajectories. Hence, we remove all information by setting all token numbers to a common value and eliminating positional embeddings for the text prompt, as depicted in Figure 4. Importantly, the exceptional prompt is distinguished from the null prompt by its absence of special tokens, such as [startoftext], [endoftext], and [pad], which still retain information. The exceptional prompt is applied only in image inversion but not in the composition process. The choice of the", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2297", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da46054a4a9d364b4cbf71f21b761c34a32c05fe9ad0cb406a58c716affc4754.jpg", + "image_caption": [ + "Figure 4: The illustration of comparison among (a) Normal Prompt, (b) Null Prompt, and (c) Exceptional Prompt." + ], + "image_footnote": [], + "bbox": [ + 78, + 80, + 472, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "token value does not significantly affect the inversion. The detailed analysis of the exceptional prompt and token value selection is provided in Appendix A.2 and A.4, respectively.", + "bbox": [ + 75, + 253, + 468, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 visually demonstrates the effectiveness of the exceptional prompt. Our results $\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{W})$ with the exceptional prompt embedding $\\mathcal{W} = \\psi (\\mathcal{P}_{\\mathrm{exceptional}})$ are more visually accurate than others. The quantitative experiments are shown in Section 5.1. All the results in Figure 3 are obtained by solving the forward and backward diffusion ODEs using the second-order DPM-Solver++ [47] in 20 steps.", + "bbox": [ + 75, + 299, + 468, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Training-Free Image Composition", + "text_level": 1, + "bbox": [ + 76, + 412, + 375, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Upon equipping the accurate invertibility, image composition can be performed based on it. The composition process consists of two key components: noise incorporation and composite self-attention maps injection.", + "bbox": [ + 75, + 436, + 468, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Noise incorporation. Before inverting images into noises, a simple preprocessing step is necessary for the reference image. Typically, only the foreground in the reference is desired for composition, so the preprocessing step involves using a pretrained segmentation model [84] to remove the background, resizing and repositioning the object to match the user's mask in the main image, and padding it with zeros to ensure it is the same size as the main image (See Appendix B.1 for visual illustration).", + "bbox": [ + 75, + 512, + 468, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the preprocessing is complete, the main and reference images are inverted to corresponding noises $\\mathbf{x}_T^{\\mathrm{m}}$ and $\\mathbf{x}_T^{\\mathrm{r}}$ by solving diffusion ODEs (Eq. (3)) from 0 to $T$ with the exceptional prompt $\\mathcal{P}_{\\mathrm{exceptional}}$ . $\\mathbf{x}_T^{\\mathrm{m}}$ and $\\mathbf{x}_T^{\\mathrm{r}}$ are then merged with standard Gaussian noise $\\mathbf{z}$ to create the starting point $\\mathbf{x}_T^*$ for generating the composition. Formally, the incorporated noise $\\mathbf{x}_T^*$ is calculated by", + "bbox": [ + 75, + 648, + 468, + 755 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {T} ^ {*} = \\mathbf {x} _ {T} ^ {\\mathrm {r}} \\odot \\mathbf {M} ^ {\\text {s e g}} + \\mathbf {x} _ {T} ^ {\\mathrm {m}} \\odot (\\mathbf {1} - \\mathbf {M} ^ {\\text {u s e r}}) + \\mathbf {z} \\odot (\\mathbf {M} ^ {\\text {u s e r}} \\oplus \\mathbf {M} ^ {\\text {s e g}}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 75, + 763, + 468, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , $\\mathbf{M}^{\\mathrm{user}}$ is the user mask, $\\mathbf{M}^{\\mathrm{seg}}$ is the segmentation mask for reference image, and $\\mathbf{M}^{\\mathrm{user}} \\oplus \\mathbf{M}^{\\mathrm{seg}}$ is the XOR of them, which is the transition area. The incorporation of $\\mathbf{z}$ enhances the smoothness of the transition between the regions inside and outside the user mask, by effectively leveraging the prior knowledge of the text-driven diffusion model to inpaint the transition area. Empirically,", + "bbox": [ + 75, + 794, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a62e7ac9c1ef73f3cab714b7f283d8bd8c74a872531768822e4ccc1d95442fbd.jpg", + "image_caption": [ + "Figure 5: A toy example for attention composition." + ], + "image_footnote": [], + "bbox": [ + 558, + 71, + 834, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for cross-domain composition, incorporating the starting point in the noise space usually is more effective, while solely for photorealism, composing in the pixel/latent space and then inverting it as the starting point is more favorable.", + "bbox": [ + 496, + 243, + 890, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Composite self-attention map injection. The incorporated noise $\\mathbf{x}_T^*$ is employed as the starting point for solving the diffusion ODE from $T$ to 0 with a normal prompt $\\mathcal{P}$ to ultimately generate the composition. $\\mathcal{P}$ is intended to assist in inpainting transition areas. However, relying solely on noise incorporation, the pretrained text-to-image model cannot preserve the appearance of the main and reference images effectively, as shown in Figure 7. To tackle this problem, we propose injecting composite self-attention maps in a specially designed manner, as the semantic information is basically retained within the rows and columns of self-attention maps (See Appendix E for visual illustrations).", + "bbox": [ + 496, + 321, + 892, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The composite self-attention map comprises three constituents: two self-attention maps, $\\mathbf{A}_{l,t}^{\\mathrm{m}}$ and $\\mathbf{A}_{l,t}^{\\mathrm{r}}$ , corresponding to the main and reference images, and a cross-attention map, $\\mathbf{A}_{l,t}^{\\mathrm{cross}}$ , calculated between them. The composition way is illustrated in Figure 5. To compose the reference image in the blue regions of Figure 5 (a), its self-attention map $\\mathbf{A}_{l,t}^{\\mathrm{r}}$ should be placed in the corresponding blue regions of Figure 5 (b), since the 5th patch can only attend to the patches 6, 9, and 10 in the context of self-attention. The green regions in Figure 5 (b) should contain the cross-attention map, $\\mathbf{A}_{l,t}^{\\mathrm{cross}}$ , which infuses contextual information from the white regions into the blue regions. If the green regions are preserved as the self-attention of the white regions without replacement, the information stored there only reflects the relation between the original patches, such as the 0th and original 5th patches in the example of index (5,0) or (0,5). This results in a lack of surrounding information being provided to the new 5th patch, such as painting or sketching, which is necessary for seamless object transition to other domains (See ablation in Figure 7).", + "bbox": [ + 496, + 505, + 892, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Essential constituents $\\mathbf{A}_{l,t}^{\\mathrm{m}},\\mathbf{A}_{l,t}^{\\mathrm{r}},\\mathbf{A}_{l,t}^{\\mathrm{cross}}$ are calculated using self-attention modules of the pretrained Stable Diffusion. Typically, a self-attention module at layer $l$ contains three projection matrices $\\mathbf{W}_l^q,\\mathbf{W}_l^k,\\mathbf{W}_l^v$ in the same dimension $\\mathbb{R}^{d\\times d}$ . Denote the spatial features of the main and reference image at timestep $t$ and layer $l$ as $\\mathbf{f}_{l,t}^{\\mathrm{m}}\\in \\mathbb{R}^{(h\\times w)\\times d}$", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2298", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\mathbf{f}_{l,t}^{\\mathrm{r}}\\in \\mathbb{R}^{(h'\\times w')\\times d}$ , respectively, where $h^\\prime \\times w^\\prime$ is the size of the reference image after resizing to match the size of the user mask. The queries, keys, and values for each self-attention module are obtained as:", + "bbox": [ + 75, + 89, + 468, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {l, t} ^ {\\mathrm {m}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {m}} \\mathbf {W} _ {l} ^ {q}, \\quad \\mathbf {k} _ {l, t} ^ {\\mathrm {m}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {m}} \\mathbf {W} _ {l} ^ {k}, \\quad \\mathbf {v} _ {l, t} ^ {\\mathrm {m}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {m}} \\mathbf {W} _ {l} ^ {v}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 157, + 468, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {l, t} ^ {\\mathrm {r}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {r}} \\mathbf {W} _ {l} ^ {q}, \\quad \\mathbf {k} _ {l, t} ^ {\\mathrm {r}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {r}} \\mathbf {W} _ {l} ^ {k}, \\quad \\mathbf {v} _ {l, t} ^ {\\mathrm {r}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {r}} \\mathbf {W} _ {l} ^ {v}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 180, + 468, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{q}_{l,t}^{\\mathrm{m}},\\mathbf{k}_{l,t}^{\\mathrm{m}},\\mathbf{v}_{l,t}^{\\mathrm{m}}\\in \\mathbb{R}^{(h\\times w)\\times d}$ and $\\mathbf{q}_{l,t}^{\\mathrm{r}},\\mathbf{k}_{l,t}^{\\mathrm{r}},\\mathbf{v}_{l,t}^{\\mathrm{r}}\\in$ $\\mathbb{R}^{(h'\\times w')\\times d}$ . Thus, $\\mathbf{A}_{l,t}^{\\mathrm{m}},\\mathbf{A}_{l,t}^{\\mathrm{r}}$ and $\\mathbf{A}_{l,t}^{\\mathrm{cross}}$ are then calculated and composed as $\\mathbf{A}_{l,t}^{*}$ for injection:", + "bbox": [ + 76, + 205, + 468, + 257 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {l, t} ^ {\\mathrm {m}} = \\operatorname {S o f t m a x} \\left(\\mathbf {q} _ {l, t} ^ {\\mathrm {m}} \\cdot \\left(\\mathbf {k} _ {l, t} ^ {\\mathrm {m}}\\right) ^ {\\mathrm {T}} / \\sqrt {d}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 265, + 468, + 292 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {l, t} ^ {\\mathrm {r}} = \\operatorname {S o f t m a x} \\left(\\mathbf {q} _ {l, t} ^ {\\mathrm {r}} \\cdot \\left(\\mathbf {k} _ {l, t} ^ {\\mathrm {r}}\\right) ^ {\\mathrm {T}} / \\sqrt {d}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 294, + 468, + 319 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {l, t} ^ {\\text {c r o s s}} = \\operatorname {S o f t m a x} \\left(\\mathbf {q} _ {l, t} ^ {\\mathrm {m}} \\cdot \\left(\\mathbf {k} _ {l, t} ^ {\\mathrm {r}}\\right) ^ {\\top} / \\sqrt {d}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 321, + 468, + 347 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {l, t} ^ {*} = \\vartheta_ {\\text {c o m p o s e}} \\left(\\mathbf {A} _ {l, t} ^ {\\mathrm {m}}, \\mathbf {A} _ {l, t} ^ {\\mathrm {r}}, \\mathbf {A} _ {l, t} ^ {\\text {c r o s s}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 349, + 468, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{A}_{l,t}^{\\mathrm{m}}\\in \\mathbb{R}^{(h\\times w)\\times (h\\times w)}$ $\\mathbf{A}_{l,t}^{\\mathrm{r}}\\in \\mathbb{R}^{(h^{\\prime}\\times w^{\\prime})\\times (h^{\\prime}\\times w^{\\prime})}$ $\\mathbf{A}_{l,t}^{\\mathrm{cross}}\\in \\mathbb{R}^{(h\\times w)\\times (h^{\\prime}\\times w^{\\prime})}$ , and $\\vartheta_{\\mathrm{compose}}$ is the function to bulid composite self-attention maps $\\mathbf{A}_{l,t}^{*}$ based on patch indices (Figure 5).", + "bbox": [ + 76, + 373, + 468, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a result, three diffusion ODEs are solved simultaneously from $T$ to 0. As depicted in Figure 2 (b), ODEs start from the accurate inverted noises $\\mathbf{x}_T^{\\mathrm{r}},\\mathbf{x}_T^{\\mathrm{m}}$ and the interpolated noise $\\mathbf{x}_T^*$ , respectively. The first two ODEs are solved using the exceptional prompt $\\mathcal{P}_{\\mathrm{exceptional}}$ to progressively reconstruct the main and reference, thus allowing for the precise retention of $\\mathbf{A}_{l,t}^{\\mathrm{m}}$ $\\mathbf{A}_{l,t}^{\\mathrm{r}}$ ,and $\\mathbf{A}_{l,t}^{\\mathrm{cross}}$ at each time step $t$ These attention maps are then composed and injected into the third ODE for generating a natural and cohesive composition with a normal prompt $\\mathcal{P}$", + "bbox": [ + 75, + 440, + 468, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To balance the generation of high-level context and finer details [13, 38], we set a threshold $\\tau_{A}$ to determine the time steps for injecting composite self-attention maps in the early stage $(t\\in [T\\times \\tau_A,T])$ and allow the model to explore ODE trajectories through a normal prompt $\\mathcal{P}$ in the later stage $(t\\in [0,T\\times \\tau_A])$ , guided by the prior of the pretrained model. However, this freedom, without the imposition of attention injection constraints, often results in deviations from the desired background (see Figure 7). Thus, similar to [6], we set an additional threshold, denoted as $\\tau_{B}$ , which regulates the trajectory rectification process. This process entails replacing the regions outside the user mask with the reconstructed main image at various time steps, i.e., $\\hat{\\mathbf{x}}_t^* = \\mathbf{x}_t^*\\odot \\mathbf{M}^{\\mathrm{user}} + \\mathbf{x}_t^{\\mathrm{m}}\\odot (\\mathbf{1} - \\mathbf{M}^{\\mathrm{user}})$ , where $t\\in [T\\times \\tau_B,T]$ . Note that only preserving the background at the final step can lead to noticeable artifacts, as shown in Appendix B.3.", + "bbox": [ + 75, + 590, + 468, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 844, + 207, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section consists of two sets of experiments. The first set assesses the effectiveness of the exceptional prompt", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fe679bdc52de9960f2532ca8a4b5be0e48fdc99893c42458bc76f3ffd1368424.jpg", + "table_caption": [ + "Table 1: The reconstruction comparison on CelebA-HQ." + ], + "table_footnote": [], + "table_body": "
MethodMAE ↓LPIPS ↓SSIM ↑
OptimizationI2S [1]0.0640.1340.872
PTI [58]0.0620.1320.877
EncoderpSp [57]0.0790.1690.793
e4e [68]0.0920.2210.742
ReStyle w/ pSp [3]0.0730.1450.823
ReStyle w/ e4e [3]0.0890.2020.758
HFGI w/ e4e [73]0.0620.1270.877
DiffusionSD w/ CFG0.1340.3400.637
SD w/ Cond.0.1260.3080.654
SD w/ Uncond.0.1260.3040.655
DiffusionCLIP [35]0.0200.0730.914
Ours0.0190.0470.918
Upper BoundVQAE [21]0.0180.0430.919
", + "bbox": [ + 504, + 108, + 890, + 311 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cb366864f96a69e791e9b928388b56eb679cfa9b962c104a86d5735697c7faa0.jpg", + "table_caption": [ + "Table 2: The further reconstruction comparison on COCO and ImageNet. \\*: an upper bound." + ], + "table_footnote": [], + "table_body": "
MethodMSCOCO (5000)ImageNet (3000)
MAE ↓LPIPS ↓SSIM ↑MAE ↓LPIPS ↓SSIM ↑
SD w/ CFG0.1500.4580.5680.1320.4960.575
SD w/ Cond.0.1220.3590.6330.1090.3890.645
SD w/ Uncond.0.1200.3630.6360.1140.4060.635
Ours0.0300.0730.8680.0330.0870.852
VQAE* [21]0.0300.0690.8700.0320.0840.854
", + "bbox": [ + 503, + 358, + 893, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(Section 5.1). The second set evaluates our image composition framework qualitatively and quantitatively (Section 5.2), followed by an ablation study (Section 5.3).", + "bbox": [ + 498, + 468, + 890, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Image Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 523, + 705, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the effectiveness of the exceptional prompt, we compared its performance with SOTA GAN [1, 3, 57, 58, 68, 73] and diffusion [35] inversion methods on the CelebA-HQ [31], following the same setting as described in [35, 73]. Additionally, we conducted experiments on the ImageNet [17] and COCO [41] with Stable Diffusion to further validate our findings. Our results (Tables 1 and 2) show that the exceptional prompt is highly effective in producing reconstructions that closely approximate the upper bound established by the vector quantized autoencoder (VQAE) [21] across all metrics, including MAE, LPIPS [82], and SSIM. The qualitative comparison is shown in Figure 3. All results of Stable Diffusion, including ours, are sampled by the second-order DPM-Solver++ [47]. Experimental settings are detailed in Appendix B.4.", + "bbox": [ + 496, + 547, + 890, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Image Composition Comparisons", + "text_level": 1, + "bbox": [ + 500, + 786, + 792, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test benchmark. As there is currently no benchmark for testing cross-domain image-guided composition as a whole, we developed a test benchmark containing 332 samples. Each sample in the benchmark consists of a main (background) image, a reference (foreground) image, a user mask, and a text prompt. The main images comprise", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2299", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7ee06532bde463cc64858ae73c987269354cc5778dfcf46f8e148bd4dc1e0cfe.jpg", + "image_caption": [ + "'acartoon animation of a cake,acroissant,a piece of bread anda cup of coffee", + "Figure 6: Qualitative comparison with SOTA and concurrent baselines in image-guided composition for sketching, photorealism, painting, and cartoon animation domains. Additional results are available in Appendix H.2." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 888, + 429 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4617d01093fff21c23e7d6f8a2a24735c621ff8da3f87159ccb217464ea019d9.jpg", + "table_caption": [ + "Table 3: Quantitative evaluation results for image composition in the photorealism domain." + ], + "table_footnote": [], + "table_body": "
MethodLPIPS(BG)↓LPIPS(FG)↓CLIP(Image)↑CLIP(Text)↑
SDEdit (0.4) [48]0.350.6280.5627.73
SDEdit (0.6) [48]0.420.6677.6827.98
Blended [5]0.110.7773.2525.19
Paint [77]0.130.7380.2625.92
DIB [81]0.110.6377.5726.84
Ours0.100.6082.8628.11
", + "bbox": [ + 81, + 518, + 470, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "four visual domains: photorealism, pencil sketching, oil painting, and cartoon animation. All reference images are from the photorealism domain as the reference requires segmentation models, which are generally more effective in this domain. Further details are available in Appendix G.", + "bbox": [ + 75, + 626, + 468, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparisons. Our qualitative comparisons are performed across four visual domains, employing SOTA and concurrent baselines that are applicable to image-guided composition, including Deep Image Blending (DIB) [81], DCCF [76], Blended Diffusion [5], Textual Inversion [24], Paint by Example [77], and SDEdit [48] under two different noising levels. As shown in Figure 6, our framework is capable of seamlessly composing objects into various domains while maintaining their identities. In contrast, DIB and DCCF fall short in processing the transition areas, leading to noticeable artifacts. Blended Diffusion's foreground generation and Textual Inversion's", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "background generation relies solely on text prompts, causing deviations from the user's intention. While Paint by Example effectively composes images within its photorealistic training domain, it struggles to adapt to other domains. Additionally, SDEdit with fewer timesteps is suitable for image composition in terms of preserving the identifying features of the reference, but the background is changed. See Appendix H.2 for additional comparisons.", + "bbox": [ + 496, + 486, + 890, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative analysis. The baselines are primarily trained in the photorealism domain, where the objective metrics are more effective; therefore, we focused our quantitative comparison within this domain and relied on user study for comparison in other domains. We assess the same baselines as in the qualitative comparison, with the exception of Textual Inversion [24], which involves instance-based optimization, and DCCF [76], which is used for harmonizing images after copy-and-paste operations. Four metrics are considered: (1) $\\mathrm{LPIPS}_{\\mathrm{(BG)}}$ [82] measures the background consistency, (2) $\\mathrm{LPIPS}_{\\mathrm{(FG)}}$ [82] evaluates the low-level similarity between the edited region and the reference foreground, (3) $\\mathrm{CLIP}_{\\mathrm{(Image)}}$ [55] evaluates the semantic similarity between the edited region and the reference in the CLIP embedding space, and (4) $\\mathrm{CLIP}_{\\mathrm{(Text)}}$ [55] measures the semantic alignment between the text prompt and the resultant image. As presented in Table 3, our method outperforms all baselines. We achieve well preservation", + "bbox": [ + 496, + 628, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2300", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/be758e4f32deb3c30f58edcc2928a4cf0625d3806421da9c95fcf155a5c4e7e1.jpg", + "table_caption": [ + "Table 4: User study: higher score, better ranking. P: photorealism; O: oil painting; S: sketchy painting; C: cartoon." + ], + "table_footnote": [], + "table_body": "
MethodP & PP & OP & SP & CTotal
Blended [5]1.8072.3142.6802.1002.093
SDEdit (0.6) [48]2.3683.0632.4092.7132.485
Paint [77]2.8792.3062.0432.6732.666
DCCF [76]3.8383.2373.2973.4703.583
Ours4.1084.0804.5714.0434.175
", + "bbox": [ + 81, + 125, + 470, + 210 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9d46de8ab927e98e991d5f6840685b0476fc875db14e3a4e5e2f6de49341a38d.jpg", + "table_caption": [ + "Table 5: Ablation study: quantitative comparison of various variants of our framework." + ], + "table_footnote": [], + "table_body": "
ConfigLPIPS(BG) ↓LPIPS(FG) ↓CLIP(Image) ↑CLIP(Text) ↑
Baseline0.340.6575.1329.00
+ Pexceptional0.320.6478.5428.23
+ SA injection0.250.6181.6428.58
+ CA injection0.260.6081.6328.52
+ Background0.100.6082.8628.11
", + "bbox": [ + 81, + 255, + 470, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the background, high object correspondence in both low-level and high-level feature spaces, as well as a high degree of alignment with the text prompt.", + "bbox": [ + 75, + 349, + 468, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User study. We conducted a user study to compare image composition baselines across domains. We recruited 50 participants via Amazon and tasked them with completing 40 ranking questions. Each question consists of 5 options, generated using distinct methods. Details are available in Appendix D. The ranking criteria comprehensively considered foreground preservation, background consistency, seamless composition, and text alignment. The results are listed in Table 4, where the domain information is presented in the format of 'foreground domain & background domain', e.g., photorealism & oil painting. Our method was favored by most participants across different domains.", + "bbox": [ + 75, + 411, + 468, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 604, + 230, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We ablate our key design choices in the following cases: (1) Baseline, where the composition is generated by solving the diffusion ODE from $T$ to 0 using DPM-Solver++ without any injection. The starting point is composed by inverted noises under the normal prompt; (2) The exceptional prompt is applied to obtain accurate inverted noises; (3) The self-attention maps $\\mathbf{A}_{l,t}^{\\mathrm{m}}$ and $\\mathbf{A}_{l,t}^{\\mathrm{r}}$ are composed and then injected; (4) The cross-attention $\\mathbf{A}_{l,t}^{\\mathrm{cross}}$ between the main and reference images is further composed for injection; (5) The background preservation is applied.", + "bbox": [ + 75, + 628, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5 presents the quantitative results, which indicate that the complete algorithm outperforms other variants in all metrics except for $\\mathrm{CLIP}_{\\mathrm{(Text)}}$ . Notably, the baseline achieves the best $\\mathrm{CLIP}_{\\mathrm{(Text)}}$ as it generates compositions solely relying on the normal prompt without any extra constraint. While metrics alone may not reveal the complete effectiveness of cross-attention injection, Figure 7 illustrates that the interactions between the main and reference images", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3af29284b64f3d0ea6355a8598b0fbd45ece9f54c435e9369fcd1dc3969afdb1.jpg", + "image_caption": [ + "Figure 7: Ablation study of different variants of our framework. SA: self-attention. CA: cross-attention." + ], + "image_footnote": [ + "'an oil painting of a burrito, Van Gogh Style" + ], + "bbox": [ + 501, + 90, + 890, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are highly beneficial to both foreground and background in terms of preserving appearance and switching domains. Note that preserving the background at different noise levels affects both background and foreground, which is distinct from preserving solely at the final step (Appendix B.3). Additional ablation results are shown in Appendix H.2.", + "bbox": [ + 496, + 266, + 890, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 498, + 369, + 772, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The primary limitation of our work lies in its inability to generate an object view that critically differs from the given reference. As a result, the choice of reference image may be restricted at times. This is because the model relies on self-attention maps to provide layout and appearance information, which in turn constrains the development of alternative views. While introducing a loose self-attention injection can generate different views, this often compromises the preservation of the object's appearance. To overcome this, further research could explore utilizing personalized concept learning techniques, such as Textual Inversion [24], to encode identity information in the text prompt through special embeddings. Alternatively, utilizing NeRF-relevant techniques [9, 49, 72] can generate other views of the object for a specific scene, but this can require expensive training. Furthermore, due to the fact that our approach relies on Stable Diffusion, it inherits its shortcomings and biases, which may result in producing artifacts in certain scenarios.", + "bbox": [ + 496, + 393, + 890, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 679, + 619, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced a method that leverages the high-order ODE solver with the exceptional prompt to achieve precise inversion of real images, which serves as a foundation for further manipulation. Building upon this, we propose a novel training-free framework, TF-ICON, that enables attention-based text-to-image diffusion models to perform image-guided composition across different domains. Our experimental results demonstrate that our approach outperforms the SOTA baselines for both image inversion and composition. We believe that image composition has the potential to become an essential tool for content creators, offering significant benefits for downstream applications in various industries.", + "bbox": [ + 496, + 704, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2301", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 90, + 236, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported by the National Research Foundation, Singapore under its Strategic Capability Research Centres Funding Initiative. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore.", + "bbox": [ + 75, + 114, + 468, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 220, + 174, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3, 6", + "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 3", + "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 3, 6", + "[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF conference on computer Vision and pattern recognition, pages 18511-18521, 2022. 3", + "[5] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 2, 7, 8", + "[6] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2, 6", + "[7] Samaneh Azadi, Deepak Pathak, Sayna Ebrahimi, and Trevor Darrell. Compositional gan: Learning image-conditional binary composition. International Journal of Computer Vision, 128(10):2570-2585, 2020. 2", + "[8] Andrew Brown, Cheng-Yang Fu, Omkar Parkhi, Tamara L Berg, and Andrea Vedaldi. End-to-end visual editing with a generatively pre-trained artist. arXiv preprint arXiv:2205.01668, 2022. 2", + "[9] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Consistent single-view perpetual view generation with conditional diffusion models. arXiv preprint arXiv:2211.12131, 2022.8", + "[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2", + "[11] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. arXiv preprint arXiv:2301.13826, 2023. 2" + ], + "bbox": [ + 78, + 246, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Bor-Chun Chen and Andrew Kae. Toward realistic image compositing with adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8415-8424, 2019. 2", + "[13] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 6", + "[14] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8394-8403, 2020. 2", + "[15] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2, 4", + "[16] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 2", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3, 4", + "[19] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022. 2", + "[20] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In Proceedings of the IEEE international conference on computer vision, pages 1301-1310, 2017. 2", + "[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 6", + "[22] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 2", + "[23] Oran Gafni and Lior Wolf. Wish you were here: Context-aware human generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7840-7849, 2020. 2", + "[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 7, 8", + "[25] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Designing an encoder" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2302", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 2", + "[26] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 4", + "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3", + "[28] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3, 4", + "[29] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 914-922, 2022. 2", + "[30] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 2", + "[31] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6", + "[32] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 4", + "[33] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3", + "[34] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2", + "[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2, 4, 6", + "[36] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3", + "[37] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022. 2", + "[38] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2, 4, 6", + "[39] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. 2023. 2", + "[40] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative adversarial networks for image compositing. In Proceedings" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9455-9464, 2018. 2", + "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[42] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8139-8148, 2020. 2", + "[43] Liu Liu, Zhenchen Liu, Bo Zhang, Jiangtong Li, Li Niu, Qingyang Liu, and Liqing Zhang. Opa: object placement assessment dataset. arXiv preprint arXiv:2107.01889, 2021. 2", + "[44] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 4", + "[45] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. arXiv preprint arXiv:2206.01714, 2022. 2", + "[46] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 4", + "[47] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 4, 5, 6", + "[48] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 2, 7, 8", + "[49] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 8", + "[50] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 2, 3, 4", + "[51] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[52] Li Niu, Wenyan Cong, Liu Liu, Yan Hong, Bo Zhang, Jing Liang, and Liqing Zhang. Making images real again: A comprehensive survey on deep image composition. arXiv preprint arXiv:2106.14490, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2303", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. arXiv preprint arXiv:2302.03027, 2023. 2, 4", + "[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2", + "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7", + "[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[57] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 3, 6", + "[58] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 3, 6", + "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4", + "[60] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2", + "[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2", + "[62] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022.4", + "[63] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4380-4390, 2021. 2", + "[64] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3", + "[65] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3, 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3, 4", + "[67] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Objectstitch: Generative object compositing. arXiv preprint arXiv:2212.00932, 2022. 2", + "[68] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 3, 4, 6", + "[69] Shashank Tripathi, Siddhartha Chandra, Amit Agrawal, Ambrish Tyagi, James M Rehg, and Visesh Chari. Learning to generate synthetic data via compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 461-470, 2019. 2", + "[70] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 4", + "[71] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446, 2022. 2, 3, 4", + "[72] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022.8", + "[73] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 6", + "[74] Sam Witteveen and Martin Andrews. Investigating prompt engineering in diffusion models. arXiv preprint arXiv:2211.15462, 2022. 2", + "[75] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Gp-gan: Towards realistic high-resolution image blending. In Proceedings of the 27th ACM international conference on multimedia, pages 2487-2495, 2019. 2", + "[76] Ben Xue, Shenghui Ran, Quan Chen, Rongfei Jia, Binqiang Zhao, and Xing Tang. Dccf: Deep comprehensible color filter learning framework for high-resolution image harmonization. In European Conference on Computer Vision, pages 300-316. Springer, 2022. 2, 7, 8", + "[77] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022. 2, 7, 8", + "[78] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2", + "[79] He Zhang, Jianming Zhang, Federico Peruzzi, Zhe Lin, and Vishal M Patel. Deep image compositing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 365-374, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2304", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Lingzhi Zhang, Tarmily Wen, Jie Min, Jiancong Wang, David Han, and Jianbo Shi. Learning object placement by inpainting for compositional data augmentation. In European Conference on Computer Vision, pages 566-581. Springer, 2020. 2", + "[81] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 231-240, 2020. 2, 7", + "[82] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 7", + "[83] Shuyang Zhang, Runze Liang, and Miao Wang. Shadowgan: Shadow synthesis for virtual objects with conditional adversarial networks. Computational Visual Media, 5(1):105-115, 2019. 2", + "[84] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34:10326-10338, 2021. 5" + ], + "bbox": [ + 78, + 90, + 468, + 373 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2305", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_model.json b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6ee6ea19aa2e472c4063755e48f79868533cf691 --- /dev/null +++ b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_model.json @@ -0,0 +1,2897 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.873, + 0.154 + ], + "angle": 0, + "content": "TF-ICON: Diffusion-Based Training-Free Cross-Domain Image Composition" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.181, + 0.699, + 0.2 + ], + "angle": 0, + "content": "Shilin Lu1 Yanzhu Liu2 Adams Wai-Kin Kong1" + }, + { + "type": "text", + "bbox": [ + 0.068, + 0.2, + 0.897, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, Nanyang Technological University, Singapore \\(^{2}\\)Institute for Infocomm Research (I²R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.238, + 0.819, + 0.251 + ], + "angle": 0, + "content": "shilin002@e.ntu.edu.sg, liu_yanzhu@i2r.a-star.edu.sg, adamskong@ntu.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.254, + 0.162, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.254, + 0.323, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.377, + 0.315, + 0.392 + ], + "angle": 0, + "content": "' a professional photograph of an Eiffel tower in the distance, ultra realistic" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.254, + 0.483, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.377, + 0.474, + 0.392 + ], + "angle": 0, + "content": "' a professional photograph of a puppy on the beach, ultra realistic'." + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.254, + 0.566, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.254, + 0.727, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.254, + 0.888, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.377, + 0.72, + 0.392 + ], + "angle": 0, + "content": "'an oil painting of a hamburger, Van Gogh Style'" + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.377, + 0.88, + 0.392 + ], + "angle": 0, + "content": "'an oil painting of a waffle, Van Gogh Style'." + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.395, + 0.452, + 0.406 + ], + "angle": 0, + "content": "(a) Composition within the photorealism domain" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.411, + 0.162, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.165, + 0.411, + 0.322, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.532, + 0.316, + 0.548 + ], + "angle": 0, + "content": "' a pencil drawing of a sheep in the sunset" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.411, + 0.483, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.532, + 0.475, + 0.548 + ], + "angle": 0, + "content": "' a pencil drawing of a panda in the sunset" + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.551, + 0.475, + 0.564 + ], + "angle": 0, + "content": "(c) Composition between real & sketchy painting domains" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.411, + 0.566, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.411, + 0.726, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.533, + 0.722, + 0.548 + ], + "angle": 0, + "content": "' a cartoon animation of buildings in the distance'" + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.411, + 0.888, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.533, + 0.88, + 0.548 + ], + "angle": 0, + "content": "' a cartoon animation of an Eiffel tower in the distance'" + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.551, + 0.86, + 0.563 + ], + "angle": 0, + "content": "(d) Composition between real & cartoon domains" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.57, + 0.893, + 0.616 + ], + "angle": 0, + "content": "Figure 1: Image composition aims to seamlessly blend distinct objects into a specific visual context. Our training-free framework equips attention-based text-driven diffusion models with the capability to achieve this task across various domains (a) photorealism, (b) oil painting, (c) sketching, and (d) cartoon animation, within 20 sampling steps." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.626, + 0.314, + 0.641 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.645, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Text-driven diffusion models have exhibited impressive generative capabilities, enabling various image editing tasks. In this paper, we propose TF-ICON, a novel Training-Free Image CompositioN framework that harnesses the power of text-driven diffusion models for cross-domain image-guided composition. This task aims to seamlessly integrate user-provided objects into a specific visual context. Current diffusion-based methods often involve costly instance-based optimization or finetuning of pretrained models on customized datasets, which can potentially undermine their rich prior. In contrast, TF-ICON can leverage off-the-shelf diffusion models to perform cross-domain image-guided composition without requiring additional training, finetuning, or optimization. Moreover, we introduce the exceptional prompt, which contains no information, to facilitate text-driven diffusion models in accurately inverting real images into latent representations," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.895, + 0.719 + ], + "angle": 0, + "content": "forming the basis for compositing. Our experiments show that equipping Stable Diffusion with the exceptional prompt outperforms state-of-the-art inversion methods on various datasets (CelebA-HQ, COCO, and ImageNet), and that TF-ICON surpasses prior baselines in versatile visual domains. Code is available at https://github.com/Shilin-LU/TF-ICON" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.74, + 0.633, + 0.756 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Image composition task involves incorporating unique objects from different photos to create a harmonious image within a specific visual context, a.k.a. image-guided composition. For instance, consider the scenario where one desires to incorporate a beloved panda into one's favorite artwork, e.g., oil or sketchy painting. The objective is to create a new image where the panda blends seamlessly into the scene without altering the appearance of the panda and the background, just as an artist meticulously crafted this panda" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "2294" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "for that artwork (See Figure 1). This task is inherently challenging, as it requires maintaining illumination consistency and preserving identifying features. The challenge is further compounded when the photos come from various domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.471, + 0.412 + ], + "angle": 0, + "content": "While recently large-scale text-to-image models [10, 19, 51, 56, 59, 61, 78] have achieved remarkable success in text-driven image generation, the ambiguity inherent in natural language presents challenges in conveying precise and nuanced visual details, even with highly detailed text prompts. Although this challenge is effectively addressed by enabling personalized concept learning [24, 25, 34, 37, 60], these methods require costly instance-based optimization and are limited in generating concepts with specified backgrounds. Recent studies [67, 77] have shown that diffusion models can achieve image-guided composition by explicitly incorporating additional guiding images. However, these models are retrained from the pretrained diffusion model on tailored datasets, which can damage the rich prior of the model. As a result, these models have limited compositional abilities beyond their training domain and still require significant computational resources." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Given the wealth of large text-to-image models that have been trained on extensive language-image datasets, we pose a question: how could these models be leveraged for image-guided composition without incurring costly training or finetuning, thereby avoiding damaging the diverse prior? To answer it, we propose the Training-Free Image CompositionN (TF-ICON) framework, which equips attention-based text-to-image diffusion models with the capability to perform image-guided composition without requiring additional training, fine-tuning, extra data, or optimization. To the best of our knowledge, this is the first training-free framework developed for image-guided composition. The framework is compatible with various diffusion model samplers, enabling completion within 20 steps, and harnesses rich semantic knowledge to facilitate image-guided compositions across diverse domains (see Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our approach constitutes an image-guided composition interface through denoising from a reliable starting latent code with the injection of composite self-attention maps. Finding the latent code that allows for reconstructing an input image while maintaining its editability, a.k.a. image inversion, is a challenging yet crucial step for state-of-the-art (SOTA) image editing frameworks involving real images [15, 26, 35, 38, 48, 53, 54, 70]. For diffusion models, while denoising diffusion implicit models (DDIM) inversion [65] has been effective for unconditional diffusion models, it falls short for text-driven diffusion models [26, 50, 70, 71]. To circumvent this, we introduce the exceptional prompt to accurately invert real images into latent codes upon pretrained text-to-image models to serve for further composition generation. The accurate latent codes are composed as the starting noise for the diffusion process. Through the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "gradual injection of composite self-attention maps that are specifically designed to reflect the relations between guiding images, we are able to infuse contextual information from the background into the incorporated objects, which results in harmonious image-guided compositions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.169, + 0.892, + 0.184 + ], + "angle": 0, + "content": "To summarize, we make the following key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.191, + 0.891, + 0.235 + ], + "angle": 0, + "content": "1. We demonstrate the superior performance of high-order diffusion ODE solvers compared to commonly used DDIM inversion for real image inversion." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.236, + 0.892, + 0.31 + ], + "angle": 0, + "content": "2. We present an exceptional prompt that allows text-driven models to achieve accurate invertibility, laying a solid groundwork for subsequent editing. Experimental results show that it surpasses SOTA inversion methods on three vision datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.312, + 0.892, + 0.356 + ], + "angle": 0, + "content": "3. We propose the first training-free framework that enables cross-domain image-guided composition for attention-based diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.358, + 0.892, + 0.402 + ], + "angle": 0, + "content": "4. We demonstrate quantitatively and qualitatively that our framework outperforms prior baselines for image-guided composition." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.191, + 0.892, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.42, + 0.642, + 0.436 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Image composition. Image composition is widely applied to electronic commerce, entertainment, and data augmentation [20, 43] for downstream tasks. It can be broadly categorized into two types: text-guided [5, 6, 11, 22, 45] and image-guided [8, 23, 39, 67, 76, 77, 79]. The former involves composing multiple objects specified by only a text prompt without limiting the appearance of objects, as long as their semantics align with the prompt. Despite the great successes of text-conditioned models, they are often prone to semantic errors [22, 56], especially when the text prompt involves multiple objects. These errors include attribute leakage, attribute interchange, and missing objects, which cause the generated images to critically different from the user's intention [22, 56]. As a result, extensive prompt engineering [74] is often necessary to achieve the desired results. In contrast to text-only guided composition, image-guided composition involves incorporating specific objects and scenarios from user-provided photos, potentially with the aid of a text prompt. However, due to the nature of involving additional real images, it is more challenging particularly when images from different visual domains. Conventionally, image-guided composition is divided into several sub-tasks [52], such as object placement [7, 12, 40, 69, 80], image blending [75, 81], image harmonization [14, 16, 30, 76], and shadow generation [29, 42, 63, 83], each of which is typically addressed by different models and pipelines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image inversion. Extensive research has been conducted on image inversion for GANs, including latent-based op" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2295" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.083, + 0.48, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.084, + 0.888, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.165, + 0.537, + 0.174 + ], + "angle": 0, + "content": "(a) Image Inversion" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.174, + 0.889, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.428, + 0.605, + 0.439 + ], + "angle": 0, + "content": "(b) Training-Free Image Composition (TF-ICON)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.443, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Figure 2: The proposed training-free image composition framework. (a) The exceptional prompt is used to invert the main and reference images into noises \\(\\mathbf{x}_T^{\\mathrm{m}}\\), \\(\\mathbf{x}_T^{\\mathrm{r}}\\), which are then composed to form the starting point \\(\\mathbf{x}_T^*\\) for the composition process. (b) Three constituents are composed for injecting into the composition process at early timesteps, including self-attention maps from the main and reference image reconstruction processes, as well as cross attention between the main and reference images. For better clarity and readability, the original main and reference images are shown in the pixel space instead of the VAE latent space, and the reference image is presented without resizing and zero-padding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.728 + ], + "angle": 0, + "content": "tization [1, 2, 33], encoders [3, 57, 68], and fine-tuning [4, 58]. For diffusion models, DDIM [65] is a widely used technique for inversion in image editing frameworks. However, in text-driven settings, DDIM leads to significant reconstruction distortion due to the instability resulting from classifier-free guidance (CFG) [18, 28]. Recently, null-text inversion [50] has been proposed to achieve accurate inversion by optimizing the unconditional prediction of the text-to-image model. It demonstrates promising results but requires instance-based optimization. Concurrently, EDICT [71] also achieves near-perfect inversion, albeit doubling the computation time of the diffusion process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.204, + 0.758 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Diffusion probabilistic models (DPM) [18, 27, 59, 64] are generative models in which an image is generated by progressively denoising from Gaussian noise. The forward diffusion process gradually perturbs data with infinite noise scales, which can be modeled as the solution of a stochastic differential equation (SDE) \\(\\{\\mathbf{x}_t\\}_{t=0}^T\\). Formally, given a data sample \\(\\mathbf{x}_0 \\sim p_0 = p_{\\mathrm{data}}\\), random noise is gradually injected, eventually resulting in a sample \\(\\mathbf{x}_T\\) which is typically distributed as a tractable prior \\(p_T\\) without any infer" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.546, + 0.882, + 0.562 + ], + "angle": 0, + "content": "mation of \\(p_0\\) , as described by the following SDE [36, 66]:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.578, + 0.891, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} \\mathbf {x} _ {t} = \\mathbf {f} \\left(\\mathbf {x} _ {t}, t\\right) \\mathrm {d} t + g (t) \\mathrm {d} \\mathbf {w} _ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.891, + 0.678 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}_t\\in \\mathbb{R}^d\\) is the standard Wiener process (a.k.a., Brownian motion), and \\(\\mathbf{f}(\\cdot ,t)\\) and \\(g(t)\\) are commonly designated as the drift and diffusion coefficient, respectively. On the other hand, the reverse diffusion process can be described by the reverse-time SDE from \\(T\\) to 0 [66]:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.69, + 0.891, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} \\mathbf {x} _ {t} = [ \\mathbf {f} (\\mathbf {x} _ {t}, t) - g (t) ^ {2} \\nabla_ {\\mathbf {x}} \\log p _ {t} (\\mathbf {x} _ {t}) ] \\mathrm {d} t + g (t) \\mathrm {d} \\bar {\\mathbf {w}} _ {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.891, + 0.78 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{w}}_t\\) is the Wiener process in the reverse time. The score function \\(\\nabla_{\\mathbf{x}}\\log p_t(\\mathbf{x}_t)\\) is the only unknown term and can be estimated by a neural network \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t)\\) whose parameter \\(\\theta\\) is optimized by a denoising objective [27, 66]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Upon attaining the trained model that predicts the score function accurately, it can be utilized to numerically solve the reverse SDE (Eq. (2)), enabling the generation of samples from a noise distribution. Song et al. [66] outline various methods, including Variance Exploding (VE), Variance Preserving (VP), and sub-VP SDE, for constructing SDEs that perturb the unknown data distribution into a fixed prior. In this work, we leverage the pre-trained text-to-image La" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2296" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.122 + ], + "angle": 0, + "content": "tent Diffusion Model (LDM) [59], a.k.a. Stable Diffusion, which applies the VP SDE in the latent space." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.134, + 0.168, + 0.149 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.16, + 0.47, + 0.326 + ], + "angle": 0, + "content": "Our objective is to utilize a main (background) image \\(\\mathbf{I}^{\\mathrm{m}}\\), a reference (foreground) image \\(\\mathbf{I}^{\\mathrm{r}}\\), a text prompt \\(\\mathcal{P}\\), and a binary mask \\(\\mathbf{M}^{\\mathrm{user}}\\) which designates the region of interest within the main image, to generate a modified image \\(\\mathbf{I}^{*}\\). The resultant image \\(\\mathbf{I}^{*}\\) should contain the reference subject with identifying features within the mask, i.e. \\(id(\\mathbf{I}^{*} \\odot \\mathbf{M}^{\\mathrm{user}}) \\approx id(\\mathbf{I}^{\\mathrm{r}})\\), while concurrently ensuring that the complementing area closely resembles the main image, i.e. \\(\\mathbf{I}^{*} \\odot (\\mathbf{1} - \\mathbf{M}^{\\mathrm{user}}) \\approx \\mathbf{I}^{\\mathrm{m}} \\odot (\\mathbf{1} - \\mathbf{M}^{\\mathrm{user}})\\). Moreover, it is ideal for the transition between the areas inside and outside the mask to be imperceptible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.471, + 0.463 + ], + "angle": 0, + "content": "We propose a training-free framework that can make use of attention-based pre-trained text-to-image models to perform image-guided composition. To the best of our knowledge, it is the first training-free framework for image-guided composition, which can be accomplished within 20 steps of sampling. The framework is mainly comprised of two steps: image inversion (Section 4.1), and composition generation (Section 4.2), as shown in Figure 2. The full algorithm is presented in Appendix B.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.436, + 0.487 + ], + "angle": 0, + "content": "4.1. Image Inversion with Exceptional Prompt" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.47, + 0.614 + ], + "angle": 0, + "content": "Achieving precise manipulation of real images often necessitates an accurate inversion process that identifies the corresponding latent representation, which not only provides editability for meaningful manipulation but also accurately reconstructs the input image [26, 68]. For diffusion models, optimal editability is typically characterized by a noise encoding that conforms to the ideal statistical properties of zero-mean, unit-variance Gaussian noise [53]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.736 + ], + "angle": 0, + "content": "ODE inversion. Most diffusion frameworks for image editing [15, 26, 35, 38, 53, 70] use DDIM inversion to invert the real image into its latent representation. However, our findings suggest that this may not be the optimal choice for inverting real images. It has been proven that DDIM is a first-order discretization of the associated probability flow ordinary differential equations (ODE) of Eq. (2) [62, 65, 66]:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.744, + 0.47, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} \\mathbf {x} _ {t} = \\left[ \\mathbf {f} \\left(\\mathbf {x} _ {t}, t\\right) - \\frac {1}{2} g (t) ^ {2} \\nabla_ {\\mathbf {x}} \\log p _ {t} \\left(\\mathbf {x} _ {t}\\right) \\right] \\mathrm {d} t, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "which can be solved using \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t)\\) and shares the consistent marginal probability distribution \\(\\{p_t(\\mathbf{x}_t)\\}_{t = 0}^T\\) with Eq. (2). Various samplers [32, 44, 46, 47] have been developed for solving the diffusion ODE starting from noise \\(\\mathbf{x}_T\\) to achieve fast sampling (10~20 steps). We offer the insight that utilizing these ODE solvers in turn as encoders starting from the real image \\(\\mathbf{x}_0\\) yields better latent representation \\(\\mathbf{x}_T\\), compared with those obtained through commonly used" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.091, + 0.58, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.268, + 0.571, + 0.279 + ], + "angle": 0, + "content": "(a) Original" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.091, + 0.658, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.268, + 0.64, + 0.279 + ], + "angle": 0, + "content": "(b) CFG" + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.091, + 0.735, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.268, + 0.72, + 0.279 + ], + "angle": 0, + "content": "(c) Cond." + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.091, + 0.813, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.745, + 0.268, + 0.804, + 0.279 + ], + "angle": 0, + "content": "(d) Uncond." + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.091, + 0.89, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.268, + 0.872, + 0.279 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.282, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Figure 3: The real image reconstruction results using Stable Diffusion with (b) classifier-free guidance (CFG) output \\(\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E},\\varnothing)\\); (c) conditional output \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E})\\); (d) unconditional output \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\varnothing)\\); and (e) ours. The prompts for (b) and (c) are 'a photo of a singer', 'a photo of a car', and 'an oil painting'. See Appendix A.3 for elaboration." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.459 + ], + "angle": 0, + "content": "DDIM. A quantitative analysis is given in Appendix A.1. The enhanced alignment between the forward and backward ODE trajectories in the high-order DPM-Solver++ [47] implies that it is better suited for real image inversion. Thus, this paper employs it for all inversions of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.473, + 0.892, + 0.58 + ], + "angle": 0, + "content": "Exceptional prompt. In the unconditional setting \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t)\\) solving the diffusion ODE (Eq. (3)) from 0 to \\(T\\) enables us to obtain the better latent code \\(\\mathbf{x}_T\\) for the real image \\(\\mathbf{x}_0\\). However, in the text-driven setting \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E})\\), existing image editing works [26, 50, 70, 71] have shown that the inversion process is prone to significant reconstruction errors, due to the instability induced by CFG [18, 28]:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.589, + 0.892, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\mathcal {E}, \\varnothing\\right) = s \\cdot \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\mathcal {E}\\right) + (1 - s) \\cdot \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\varnothing\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.704 + ], + "angle": 0, + "content": "where \\(\\varnothing = \\psi(\"\")\\) and \\(\\mathcal{E} = \\psi(\\mathcal{P})\\) are embeddings of the null and normal prompt, and \\(s\\) is the guidance scale. Our experiments further reveal that even without CFG, both conditional output \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E})\\) and unconditional output \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\emptyset)\\) of text-to-image diffusion models still produce large reconstruction errors, as depicted in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To achieve accurate inversion, we present a straightforward yet effective solution, namely exceptional prompt, \\( \\mathcal{P}_{\\text{exceptional}} \\). Intuitively, any information contained within the input prompt can result in the deviation of the backward ODE trajectories from the forward trajectories. Hence, we remove all information by setting all token numbers to a common value and eliminating positional embeddings for the text prompt, as depicted in Figure 4. Importantly, the exceptional prompt is distinguished from the null prompt by its absence of special tokens, such as [startoftext], [endoftext], and [pad], which still retain information. The exceptional prompt is applied only in image inversion but not in the composition process. The choice of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2297" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.081, + 0.473, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.245 + ], + "angle": 0, + "content": "Figure 4: The illustration of comparison among (a) Normal Prompt, (b) Null Prompt, and (c) Exceptional Prompt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.469, + 0.3 + ], + "angle": 0, + "content": "token value does not significantly affect the inversion. The detailed analysis of the exceptional prompt and token value selection is provided in Appendix A.2 and A.4, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.47, + 0.406 + ], + "angle": 0, + "content": "Figure 3 visually demonstrates the effectiveness of the exceptional prompt. Our results \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{W})\\) with the exceptional prompt embedding \\(\\mathcal{W} = \\psi (\\mathcal{P}_{\\mathrm{exceptional}})\\) are more visually accurate than others. The quantitative experiments are shown in Section 5.1. All the results in Figure 3 are obtained by solving the forward and backward diffusion ODEs using the second-order DPM-Solver++ [47] in 20 steps." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.414, + 0.377, + 0.43 + ], + "angle": 0, + "content": "4.2. Training-Free Image Composition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.499 + ], + "angle": 0, + "content": "Upon equipping the accurate invertibility, image composition can be performed based on it. The composition process consists of two key components: noise incorporation and composite self-attention maps injection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.469, + 0.648 + ], + "angle": 0, + "content": "Noise incorporation. Before inverting images into noises, a simple preprocessing step is necessary for the reference image. Typically, only the foreground in the reference is desired for composition, so the preprocessing step involves using a pretrained segmentation model [84] to remove the background, resizing and repositioning the object to match the user's mask in the main image, and padding it with zeros to ensure it is the same size as the main image (See Appendix B.1 for visual illustration)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.649, + 0.47, + 0.756 + ], + "angle": 0, + "content": "Once the preprocessing is complete, the main and reference images are inverted to corresponding noises \\(\\mathbf{x}_T^{\\mathrm{m}}\\) and \\(\\mathbf{x}_T^{\\mathrm{r}}\\) by solving diffusion ODEs (Eq. (3)) from 0 to \\(T\\) with the exceptional prompt \\(\\mathcal{P}_{\\mathrm{exceptional}}\\). \\(\\mathbf{x}_T^{\\mathrm{m}}\\) and \\(\\mathbf{x}_T^{\\mathrm{r}}\\) are then merged with standard Gaussian noise \\(\\mathbf{z}\\) to create the starting point \\(\\mathbf{x}_T^*\\) for generating the composition. Formally, the incorporated noise \\(\\mathbf{x}_T^*\\) is calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {T} ^ {*} = \\mathbf {x} _ {T} ^ {\\mathrm {r}} \\odot \\mathbf {M} ^ {\\text {s e g}} + \\mathbf {x} _ {T} ^ {\\mathrm {m}} \\odot (\\mathbf {1} - \\mathbf {M} ^ {\\text {u s e r}}) + \\mathbf {z} \\odot (\\mathbf {M} ^ {\\text {u s e r}} \\oplus \\mathbf {M} ^ {\\text {s e g}}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), \\(\\mathbf{M}^{\\mathrm{user}}\\) is the user mask, \\(\\mathbf{M}^{\\mathrm{seg}}\\) is the segmentation mask for reference image, and \\(\\mathbf{M}^{\\mathrm{user}} \\oplus \\mathbf{M}^{\\mathrm{seg}}\\) is the XOR of them, which is the transition area. The incorporation of \\(\\mathbf{z}\\) enhances the smoothness of the transition between the regions inside and outside the user mask, by effectively leveraging the prior knowledge of the text-driven diffusion model to inpaint the transition area. Empirically," + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.073, + 0.835, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.218, + 0.865, + 0.234 + ], + "angle": 0, + "content": "Figure 5: A toy example for attention composition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.305 + ], + "angle": 0, + "content": "for cross-domain composition, incorporating the starting point in the noise space usually is more effective, while solely for photorealism, composing in the pixel/latent space and then inverting it as the starting point is more favorable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Composite self-attention map injection. The incorporated noise \\(\\mathbf{x}_T^*\\) is employed as the starting point for solving the diffusion ODE from \\(T\\) to 0 with a normal prompt \\(\\mathcal{P}\\) to ultimately generate the composition. \\(\\mathcal{P}\\) is intended to assist in inpainting transition areas. However, relying solely on noise incorporation, the pretrained text-to-image model cannot preserve the appearance of the main and reference images effectively, as shown in Figure 7. To tackle this problem, we propose injecting composite self-attention maps in a specially designed manner, as the semantic information is basically retained within the rows and columns of self-attention maps (See Appendix E for visual illustrations)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.893, + 0.809 + ], + "angle": 0, + "content": "The composite self-attention map comprises three constituents: two self-attention maps, \\(\\mathbf{A}_{l,t}^{\\mathrm{m}}\\) and \\(\\mathbf{A}_{l,t}^{\\mathrm{r}}\\), corresponding to the main and reference images, and a cross-attention map, \\(\\mathbf{A}_{l,t}^{\\mathrm{cross}}\\), calculated between them. The composition way is illustrated in Figure 5. To compose the reference image in the blue regions of Figure 5 (a), its self-attention map \\(\\mathbf{A}_{l,t}^{\\mathrm{r}}\\) should be placed in the corresponding blue regions of Figure 5 (b), since the 5th patch can only attend to the patches 6, 9, and 10 in the context of self-attention. The green regions in Figure 5 (b) should contain the cross-attention map, \\(\\mathbf{A}_{l,t}^{\\mathrm{cross}}\\), which infuses contextual information from the white regions into the blue regions. If the green regions are preserved as the self-attention of the white regions without replacement, the information stored there only reflects the relation between the original patches, such as the 0th and original 5th patches in the example of index (5,0) or (0,5). This results in a lack of surrounding information being provided to the new 5th patch, such as painting or sketching, which is necessary for seamless object transition to other domains (See ablation in Figure 7)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Essential constituents \\(\\mathbf{A}_{l,t}^{\\mathrm{m}},\\mathbf{A}_{l,t}^{\\mathrm{r}},\\mathbf{A}_{l,t}^{\\mathrm{cross}}\\) are calculated using self-attention modules of the pretrained Stable Diffusion. Typically, a self-attention module at layer \\(l\\) contains three projection matrices \\(\\mathbf{W}_l^q,\\mathbf{W}_l^k,\\mathbf{W}_l^v\\) in the same dimension \\(\\mathbb{R}^{d\\times d}\\). Denote the spatial features of the main and reference image at timestep \\(t\\) and layer \\(l\\) as \\(\\mathbf{f}_{l,t}^{\\mathrm{m}}\\in \\mathbb{R}^{(h\\times w)\\times d}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2298" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.47, + 0.152 + ], + "angle": 0, + "content": "and \\(\\mathbf{f}_{l,t}^{\\mathrm{r}}\\in \\mathbb{R}^{(h'\\times w')\\times d}\\), respectively, where \\(h^\\prime \\times w^\\prime\\) is the size of the reference image after resizing to match the size of the user mask. The queries, keys, and values for each self-attention module are obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.159, + 0.469, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {l, t} ^ {\\mathrm {m}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {m}} \\mathbf {W} _ {l} ^ {q}, \\quad \\mathbf {k} _ {l, t} ^ {\\mathrm {m}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {m}} \\mathbf {W} _ {l} ^ {k}, \\quad \\mathbf {v} _ {l, t} ^ {\\mathrm {m}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {m}} \\mathbf {W} _ {l} ^ {v}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.181, + 0.469, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {l, t} ^ {\\mathrm {r}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {r}} \\mathbf {W} _ {l} ^ {q}, \\quad \\mathbf {k} _ {l, t} ^ {\\mathrm {r}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {r}} \\mathbf {W} _ {l} ^ {k}, \\quad \\mathbf {v} _ {l, t} ^ {\\mathrm {r}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {r}} \\mathbf {W} _ {l} ^ {v}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.207, + 0.469, + 0.258 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}_{l,t}^{\\mathrm{m}},\\mathbf{k}_{l,t}^{\\mathrm{m}},\\mathbf{v}_{l,t}^{\\mathrm{m}}\\in \\mathbb{R}^{(h\\times w)\\times d}\\) and \\(\\mathbf{q}_{l,t}^{\\mathrm{r}},\\mathbf{k}_{l,t}^{\\mathrm{r}},\\mathbf{v}_{l,t}^{\\mathrm{r}}\\in\\) \\(\\mathbb{R}^{(h'\\times w')\\times d}\\). Thus, \\(\\mathbf{A}_{l,t}^{\\mathrm{m}},\\mathbf{A}_{l,t}^{\\mathrm{r}}\\) and \\(\\mathbf{A}_{l,t}^{\\mathrm{cross}}\\) are then calculated and composed as \\(\\mathbf{A}_{l,t}^{*}\\) for injection:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.266, + 0.469, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {l, t} ^ {\\mathrm {m}} = \\operatorname {S o f t m a x} \\left(\\mathbf {q} _ {l, t} ^ {\\mathrm {m}} \\cdot \\left(\\mathbf {k} _ {l, t} ^ {\\mathrm {m}}\\right) ^ {\\mathrm {T}} / \\sqrt {d}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.295, + 0.469, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {l, t} ^ {\\mathrm {r}} = \\operatorname {S o f t m a x} \\left(\\mathbf {q} _ {l, t} ^ {\\mathrm {r}} \\cdot \\left(\\mathbf {k} _ {l, t} ^ {\\mathrm {r}}\\right) ^ {\\mathrm {T}} / \\sqrt {d}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.323, + 0.469, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {l, t} ^ {\\text {c r o s s}} = \\operatorname {S o f t m a x} \\left(\\mathbf {q} _ {l, t} ^ {\\mathrm {m}} \\cdot \\left(\\mathbf {k} _ {l, t} ^ {\\mathrm {r}}\\right) ^ {\\top} / \\sqrt {d}\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.35, + 0.469, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {l, t} ^ {*} = \\vartheta_ {\\text {c o m p o s e}} \\left(\\mathbf {A} _ {l, t} ^ {\\mathrm {m}}, \\mathbf {A} _ {l, t} ^ {\\mathrm {r}}, \\mathbf {A} _ {l, t} ^ {\\text {c r o s s}}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.374, + 0.469, + 0.44 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}_{l,t}^{\\mathrm{m}}\\in \\mathbb{R}^{(h\\times w)\\times (h\\times w)}\\) \\(\\mathbf{A}_{l,t}^{\\mathrm{r}}\\in \\mathbb{R}^{(h^{\\prime}\\times w^{\\prime})\\times (h^{\\prime}\\times w^{\\prime})}\\) \\(\\mathbf{A}_{l,t}^{\\mathrm{cross}}\\in \\mathbb{R}^{(h\\times w)\\times (h^{\\prime}\\times w^{\\prime})}\\), and \\(\\vartheta_{\\mathrm{compose}}\\) is the function to bulid composite self-attention maps \\(\\mathbf{A}_{l,t}^{*}\\) based on patch indices (Figure 5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.469, + 0.591 + ], + "angle": 0, + "content": "As a result, three diffusion ODEs are solved simultaneously from \\(T\\) to 0. As depicted in Figure 2 (b), ODEs start from the accurate inverted noises \\(\\mathbf{x}_T^{\\mathrm{r}},\\mathbf{x}_T^{\\mathrm{m}}\\) and the interpolated noise \\(\\mathbf{x}_T^*\\) , respectively. The first two ODEs are solved using the exceptional prompt \\(\\mathcal{P}_{\\mathrm{exceptional}}\\) to progressively reconstruct the main and reference, thus allowing for the precise retention of \\(\\mathbf{A}_{l,t}^{\\mathrm{m}}\\) \\(\\mathbf{A}_{l,t}^{\\mathrm{r}}\\) ,and \\(\\mathbf{A}_{l,t}^{\\mathrm{cross}}\\) at each time step \\(t\\) These attention maps are then composed and injected into the third ODE for generating a natural and cohesive composition with a normal prompt \\(\\mathcal{P}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.469, + 0.833 + ], + "angle": 0, + "content": "To balance the generation of high-level context and finer details [13, 38], we set a threshold \\(\\tau_{A}\\) to determine the time steps for injecting composite self-attention maps in the early stage \\((t\\in [T\\times \\tau_A,T])\\) and allow the model to explore ODE trajectories through a normal prompt \\(\\mathcal{P}\\) in the later stage \\((t\\in [0,T\\times \\tau_A])\\), guided by the prior of the pretrained model. However, this freedom, without the imposition of attention injection constraints, often results in deviations from the desired background (see Figure 7). Thus, similar to [6], we set an additional threshold, denoted as \\(\\tau_{B}\\), which regulates the trajectory rectification process. This process entails replacing the regions outside the user mask with the reconstructed main image at various time steps, i.e., \\(\\hat{\\mathbf{x}}_t^* = \\mathbf{x}_t^*\\odot \\mathbf{M}^{\\mathrm{user}} + \\mathbf{x}_t^{\\mathrm{m}}\\odot (\\mathbf{1} - \\mathbf{M}^{\\mathrm{user}})\\), where \\(t\\in [T\\times \\tau_B,T]\\). Note that only preserving the background at the final step can lead to noticeable artifacts, as shown in Appendix B.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.209, + 0.862 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "This section consists of two sets of experiments. The first set assesses the effectiveness of the exceptional prompt" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.09, + 0.88, + 0.105 + ], + "angle": 0, + "content": "Table 1: The reconstruction comparison on CelebA-HQ." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.109, + 0.892, + 0.312 + ], + "angle": 0, + "content": "
MethodMAE ↓LPIPS ↓SSIM ↑
OptimizationI2S [1]0.0640.1340.872
PTI [58]0.0620.1320.877
EncoderpSp [57]0.0790.1690.793
e4e [68]0.0920.2210.742
ReStyle w/ pSp [3]0.0730.1450.823
ReStyle w/ e4e [3]0.0890.2020.758
HFGI w/ e4e [73]0.0620.1270.877
DiffusionSD w/ CFG0.1340.3400.637
SD w/ Cond.0.1260.3080.654
SD w/ Uncond.0.1260.3040.655
DiffusionCLIP [35]0.0200.0730.914
Ours0.0190.0470.918
Upper BoundVQAE [21]0.0180.0430.919
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.325, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Table 2: The further reconstruction comparison on COCO and ImageNet. \\*: an upper bound." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.359, + 0.894, + 0.455 + ], + "angle": 0, + "content": "
MethodMSCOCO (5000)ImageNet (3000)
MAE ↓LPIPS ↓SSIM ↑MAE ↓LPIPS ↓SSIM ↑
SD w/ CFG0.1500.4580.5680.1320.4960.575
SD w/ Cond.0.1220.3590.6330.1090.3890.645
SD w/ Uncond.0.1200.3630.6360.1140.4060.635
Ours0.0300.0730.8680.0330.0870.852
VQAE* [21]0.0300.0690.8700.0320.0840.854
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.469, + 0.892, + 0.515 + ], + "angle": 0, + "content": "(Section 5.1). The second set evaluates our image composition framework qualitatively and quantitatively (Section 5.2), followed by an ablation study (Section 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.525, + 0.706, + 0.541 + ], + "angle": 0, + "content": "5.1. Image Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.776 + ], + "angle": 0, + "content": "To assess the effectiveness of the exceptional prompt, we compared its performance with SOTA GAN [1, 3, 57, 58, 68, 73] and diffusion [35] inversion methods on the CelebA-HQ [31], following the same setting as described in [35, 73]. Additionally, we conducted experiments on the ImageNet [17] and COCO [41] with Stable Diffusion to further validate our findings. Our results (Tables 1 and 2) show that the exceptional prompt is highly effective in producing reconstructions that closely approximate the upper bound established by the vector quantized autoencoder (VQAE) [21] across all metrics, including MAE, LPIPS [82], and SSIM. The qualitative comparison is shown in Figure 3. All results of Stable Diffusion, including ours, are sampled by the second-order DPM-Solver++ [47]. Experimental settings are detailed in Appendix B.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.794, + 0.803 + ], + "angle": 0, + "content": "5.2. Image Composition Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Test benchmark. As there is currently no benchmark for testing cross-domain image-guided composition as a whole, we developed a test benchmark containing 332 samples. Each sample in the benchmark consists of a main (background) image, a reference (foreground) image, a user mask, and a text prompt. The main images comprise" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2299" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.89, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.431, + 0.653, + 0.441 + ], + "angle": 0, + "content": "'acartoon animation of a cake,acroissant,a piece of bread anda cup of coffee" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.444, + 0.893, + 0.474 + ], + "angle": 0, + "content": "Figure 6: Qualitative comparison with SOTA and concurrent baselines in image-guided composition for sketching, photorealism, painting, and cartoon animation domains. Additional results are available in Appendix H.2." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.485, + 0.47, + 0.515 + ], + "angle": 0, + "content": "Table 3: Quantitative evaluation results for image composition in the photorealism domain." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.52, + 0.472, + 0.609 + ], + "angle": 0, + "content": "
MethodLPIPS(BG)↓LPIPS(FG)↓CLIP(Image)↑CLIP(Text)↑
SDEdit (0.4) [48]0.350.6280.5627.73
SDEdit (0.6) [48]0.420.6677.6827.98
Blended [5]0.110.7773.2525.19
Paint [77]0.130.7380.2625.92
DIB [81]0.110.6377.5726.84
Ours0.100.6082.8628.11
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.704 + ], + "angle": 0, + "content": "four visual domains: photorealism, pencil sketching, oil painting, and cartoon animation. All reference images are from the photorealism domain as the reference requires segmentation models, which are generally more effective in this domain. Further details are available in Appendix G." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Qualitative comparisons. Our qualitative comparisons are performed across four visual domains, employing SOTA and concurrent baselines that are applicable to image-guided composition, including Deep Image Blending (DIB) [81], DCCF [76], Blended Diffusion [5], Textual Inversion [24], Paint by Example [77], and SDEdit [48] under two different noising levels. As shown in Figure 6, our framework is capable of seamlessly composing objects into various domains while maintaining their identities. In contrast, DIB and DCCF fall short in processing the transition areas, leading to noticeable artifacts. Blended Diffusion's foreground generation and Textual Inversion's" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.608 + ], + "angle": 0, + "content": "background generation relies solely on text prompts, causing deviations from the user's intention. While Paint by Example effectively composes images within its photorealistic training domain, it struggles to adapt to other domains. Additionally, SDEdit with fewer timesteps is suitable for image composition in terms of preserving the identifying features of the reference, but the background is changed. See Appendix H.2 for additional comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Quantitative analysis. The baselines are primarily trained in the photorealism domain, where the objective metrics are more effective; therefore, we focused our quantitative comparison within this domain and relied on user study for comparison in other domains. We assess the same baselines as in the qualitative comparison, with the exception of Textual Inversion [24], which involves instance-based optimization, and DCCF [76], which is used for harmonizing images after copy-and-paste operations. Four metrics are considered: (1) \\(\\mathrm{LPIPS}_{\\mathrm{(BG)}}\\) [82] measures the background consistency, (2) \\(\\mathrm{LPIPS}_{\\mathrm{(FG)}}\\) [82] evaluates the low-level similarity between the edited region and the reference foreground, (3) \\(\\mathrm{CLIP}_{\\mathrm{(Image)}}\\) [55] evaluates the semantic similarity between the edited region and the reference in the CLIP embedding space, and (4) \\(\\mathrm{CLIP}_{\\mathrm{(Text)}}\\) [55] measures the semantic alignment between the text prompt and the resultant image. As presented in Table 3, our method outperforms all baselines. We achieve well preservation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2300" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Table 4: User study: higher score, better ranking. P: photorealism; O: oil painting; S: sketchy painting; C: cartoon." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.125, + 0.472, + 0.212 + ], + "angle": 0, + "content": "
MethodP & PP & OP & SP & CTotal
Blended [5]1.8072.3142.6802.1002.093
SDEdit (0.6) [48]2.3683.0632.4092.7132.485
Paint [77]2.8792.3062.0432.6732.666
DCCF [76]3.8383.2373.2973.4703.583
Ours4.1084.0804.5714.0434.175
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.222, + 0.47, + 0.25 + ], + "angle": 0, + "content": "Table 5: Ablation study: quantitative comparison of various variants of our framework." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.256, + 0.471, + 0.337 + ], + "angle": 0, + "content": "
ConfigLPIPS(BG) ↓LPIPS(FG) ↓CLIP(Image) ↑CLIP(Text) ↑
Baseline0.340.6575.1329.00
+ Pexceptional0.320.6478.5428.23
+ SA injection0.250.6181.6428.58
+ CA injection0.260.6081.6328.52
+ Background0.100.6082.8628.11
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.398 + ], + "angle": 0, + "content": "of the background, high object correspondence in both low-level and high-level feature spaces, as well as a high degree of alignment with the text prompt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.593 + ], + "angle": 0, + "content": "User study. We conducted a user study to compare image composition baselines across domains. We recruited 50 participants via Amazon and tasked them with completing 40 ranking questions. Each question consists of 5 options, generated using distinct methods. Details are available in Appendix D. The ranking criteria comprehensively considered foreground preservation, background consistency, seamless composition, and text alignment. The results are listed in Table 4, where the domain information is presented in the format of 'foreground domain & background domain', e.g., photorealism & oil painting. Our method was favored by most participants across different domains." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.231, + 0.621 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.78 + ], + "angle": 0, + "content": "We ablate our key design choices in the following cases: (1) Baseline, where the composition is generated by solving the diffusion ODE from \\(T\\) to 0 using DPM-Solver++ without any injection. The starting point is composed by inverted noises under the normal prompt; (2) The exceptional prompt is applied to obtain accurate inverted noises; (3) The self-attention maps \\(\\mathbf{A}_{l,t}^{\\mathrm{m}}\\) and \\(\\mathbf{A}_{l,t}^{\\mathrm{r}}\\) are composed and then injected; (4) The cross-attention \\(\\mathbf{A}_{l,t}^{\\mathrm{cross}}\\) between the main and reference images is further composed for injection; (5) The background preservation is applied." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Table 5 presents the quantitative results, which indicate that the complete algorithm outperforms other variants in all metrics except for \\(\\mathrm{CLIP}_{\\mathrm{(Text)}}\\). Notably, the baseline achieves the best \\(\\mathrm{CLIP}_{\\mathrm{(Text)}}\\) as it generates compositions solely relying on the normal prompt without any extra constraint. While metrics alone may not reveal the complete effectiveness of cross-attention injection, Figure 7 illustrates that the interactions between the main and reference images" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.892, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.609, + 0.221, + 0.783, + 0.23 + ], + "angle": 0, + "content": "'an oil painting of a burrito, Van Gogh Style" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.23, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Figure 7: Ablation study of different variants of our framework. SA: self-attention. CA: cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.892, + 0.358 + ], + "angle": 0, + "content": "are highly beneficial to both foreground and background in terms of preserving appearance and switching domains. Note that preserving the background at different noise levels affects both background and foreground, which is distinct from preserving solely at the final step (Appendix B.3). Additional ablation results are shown in Appendix H.2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.37, + 0.774, + 0.387 + ], + "angle": 0, + "content": "6. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.668 + ], + "angle": 0, + "content": "The primary limitation of our work lies in its inability to generate an object view that critically differs from the given reference. As a result, the choice of reference image may be restricted at times. This is because the model relies on self-attention maps to provide layout and appearance information, which in turn constrains the development of alternative views. While introducing a loose self-attention injection can generate different views, this often compromises the preservation of the object's appearance. To overcome this, further research could explore utilizing personalized concept learning techniques, such as Textual Inversion [24], to encode identity information in the text prompt through special embeddings. Alternatively, utilizing NeRF-relevant techniques [9, 49, 72] can generate other views of the object for a specific scene, but this can require expensive training. Furthermore, due to the fact that our approach relies on Stable Diffusion, it inherits its shortcomings and biases, which may result in producing artifacts in certain scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.62, + 0.695 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We introduced a method that leverages the high-order ODE solver with the exceptional prompt to achieve precise inversion of real images, which serves as a foundation for further manipulation. Building upon this, we propose a novel training-free framework, TF-ICON, that enables attention-based text-to-image diffusion models to perform image-guided composition across different domains. Our experimental results demonstrate that our approach outperforms the SOTA baselines for both image inversion and composition. We believe that image composition has the potential to become an essential tool for content creators, offering significant benefits for downstream applications in various industries." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2301" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.238, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.47, + 0.207 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore under its Strategic Capability Research Centres Funding Initiative. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.222, + 0.175, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.472, + 0.359 + ], + "angle": 0, + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.361, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.472, + 0.488 + ], + "angle": 0, + "content": "[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF conference on computer Vision and pattern recognition, pages 18511-18521, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.47, + 0.53 + ], + "angle": 0, + "content": "[5] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[6] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.59, + 0.472, + 0.645 + ], + "angle": 0, + "content": "[7] Samaneh Azadi, Deepak Pathak, Sayna Ebrahimi, and Trevor Darrell. Compositional gan: Learning image-conditional binary composition. International Journal of Computer Vision, 128(10):2570-2585, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.647, + 0.47, + 0.701 + ], + "angle": 0, + "content": "[8] Andrew Brown, Cheng-Yang Fu, Omkar Parkhi, Tamara L Berg, and Andrea Vedaldi. End-to-end visual editing with a generatively pre-trained artist. arXiv preprint arXiv:2205.01668, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[9] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Consistent single-view perpetual view generation with conditional diffusion models. arXiv preprint arXiv:2211.12131, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.774, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. arXiv preprint arXiv:2301.13826, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.247, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[12] Bor-Chun Chen and Andrew Kae. Toward realistic image compositing with adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8415-8424, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.895, + 0.219 + ], + "angle": 0, + "content": "[13] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.895, + 0.289 + ], + "angle": 0, + "content": "[14] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8394-8403, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[15] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[16] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[19] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[20] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In Proceedings of the IEEE international conference on computer vision, pages 1301-1310, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[22] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[23] Oran Gafni and Lior Wolf. Wish you were here: Context-aware human generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7840-7849, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Designing an encoder" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2302" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[26] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[28] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[29] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 914-922, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[30] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.428 + ], + "angle": 0, + "content": "[31] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.469 + ], + "angle": 0, + "content": "[32] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.539 + ], + "angle": 0, + "content": "[33] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[34] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.706 + ], + "angle": 0, + "content": "[36] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[37] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[38] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[39] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[40] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative adversarial networks for image compositing. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9455-9464, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[42] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8139-8148, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[43] Liu Liu, Zhenchen Liu, Bo Zhang, Jiangtong Li, Li Niu, Qingyang Liu, and Liqing Zhang. Opa: object placement assessment dataset. arXiv preprint arXiv:2107.01889, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.351, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[44] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.396, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[45] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. arXiv preprint arXiv:2206.01714, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.454, + 0.892, + 0.512 + ], + "angle": 0, + "content": "[46] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[47] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.57, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[48] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[49] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.713, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[50] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[51] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[52] Li Niu, Wenyan Cong, Liu Liu, Yan Hong, Bo Zhang, Jing Liang, and Liqing Zhang. Making images real again: A comprehensive survey on deep image composition. arXiv preprint arXiv:2106.14490, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2303" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[53] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. arXiv preprint arXiv:2302.03027, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.203 + ], + "angle": 0, + "content": "[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.469, + 0.345 + ], + "angle": 0, + "content": "[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[57] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.42, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[58] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.477, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[60] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.605, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[62] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.733, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[63] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4380-4390, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.855 + ], + "angle": 0, + "content": "[64] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[65] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[66] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[67] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Objectstitch: Generative object compositing. arXiv preprint arXiv:2212.00932, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[68] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[69] Shashank Tripathi, Siddhartha Chandra, Amit Agrawal, Ambrish Tyagi, James M Rehg, and Visesh Chari. Learning to generate synthetic data via compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 461-470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.331, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[70] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[71] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.429, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[72] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[73] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[74] Sam Witteveen and Martin Andrews. Investigating prompt engineering in diffusion models. arXiv preprint arXiv:2211.15462, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[75] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Gp-gan: Towards realistic high-resolution image blending. In Proceedings of the 27th ACM international conference on multimedia, pages 2487-2495, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.652, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[76] Ben Xue, Shenghui Ran, Quan Chen, Rongfei Jia, Binqiang Zhao, and Xing Tang. Dccf: Deep comprehensible color filter learning framework for high-resolution image harmonization. In European Conference on Computer Vision, pages 300-316. Springer, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[77] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[78] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[79] He Zhang, Jianming Zhang, Federico Peruzzi, Zhe Lin, and Vishal M Patel. Deep image compositing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 365-374, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2304" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[80] Lingzhi Zhang, Tarmily Wen, Jie Min, Jiancong Wang, David Han, and Jianbo Shi. Learning object placement by inpainting for compositional data augmentation. In European Conference on Computer Vision, pages 566-581. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[81] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 231-240, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[82] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.317 + ], + "angle": 0, + "content": "[83] Shuyang Zhang, Runze Liang, and Miao Wang. Shadowgan: Shadow synthesis for virtual objects with conditional adversarial networks. Computational Visual Media, 5(1):105-115, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[84] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34:10326-10338, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2305" + } + ] +] \ No newline at end of file diff --git a/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_origin.pdf b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..80ada039a7e0663e2f5a12b9c9c635b8404f3f57 --- /dev/null +++ b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80e938b9014cd9f92c6da23a5f028158e7d23053ba50805405bc649fe2b5ccd7 +size 901508 diff --git a/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/full.md b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1f96e722279dba672777ba895fd4a48768a3ca3d --- /dev/null +++ b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/full.md @@ -0,0 +1,390 @@ +# TF-ICON: Diffusion-Based Training-Free Cross-Domain Image Composition + +Shilin Lu1 Yanzhu Liu2 Adams Wai-Kin Kong1 + +$^{1}$ School of Computer Science and Engineering, Nanyang Technological University, Singapore $^{2}$ Institute for Infocomm Research (I²R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore + +shilin002@e.ntu.edu.sg, liu_yanzhu@i2r.a-star.edu.sg, adamskong@ntu.edu.sg + +![](images/d26ed3cf3c91eb696cd946fd14db40e6479a9ab39befc57b3e2c7728153d0cec.jpg) + +![](images/a71505ddcf5c4341f7eab4a1f77c152f11f92b046ef8504485bc601b156a9911.jpg) +' a professional photograph of an Eiffel tower in the distance, ultra realistic + +![](images/70eea039ca2f245f29c9c5f0094a0da7cc9140d1f8963b3d7573cbbd9e19a7dd.jpg) +' a professional photograph of a puppy on the beach, ultra realistic'. + +![](images/7ab917b90888d14ede31cb0f51afa86b19cb0afd288ee6917dede397707926f4.jpg) + +![](images/eeb42dd5d159fe37c37357571000d991dd7d7e6278d1fcceff62582039928582.jpg) +'an oil painting of a hamburger, Van Gogh Style' + +![](images/a1ad3be380ad420bde06da0ac289b5760085ac199238e925f4c976d68f6f98ac.jpg) +'an oil painting of a waffle, Van Gogh Style'. + +![](images/bd8f399f2be6462eb7f43715f2e0a73f0383566f77cfa134538a887cd5754a7b.jpg) +(c) Composition between real & sketchy painting domains +Figure 1: Image composition aims to seamlessly blend distinct objects into a specific visual context. Our training-free framework equips attention-based text-driven diffusion models with the capability to achieve this task across various domains (a) photorealism, (b) oil painting, (c) sketching, and (d) cartoon animation, within 20 sampling steps. + +![](images/3fea9fed12136a363be86ab142e4491af50fd65e07a9f8e735e70fe72f80ea0e.jpg) +(a) Composition within the photorealism domain +' a pencil drawing of a sheep in the sunset + +![](images/6843d479466ac00bc61107fd8bd46e9fc315db2b10fa359d57b07e969644329e.jpg) +' a pencil drawing of a panda in the sunset + +![](images/8de6c7ddb49df72fe540feed7935bf14f448769a5ebcaf3cc5c9a27b9d7e9d96.jpg) + +![](images/523e81e026b52099b78832bca53b534810a59a469f9ba6d7e571b6554c898f1d.jpg) +' a cartoon animation of buildings in the distance' +(d) Composition between real & cartoon domains + +![](images/54d85347f8cfb11d4f2f504c98b0e59732ceea6f5a217a40e3ed8e8ed9e067ef.jpg) +' a cartoon animation of an Eiffel tower in the distance' + +# Abstract + +Text-driven diffusion models have exhibited impressive generative capabilities, enabling various image editing tasks. In this paper, we propose TF-ICON, a novel Training-Free Image CompositioN framework that harnesses the power of text-driven diffusion models for cross-domain image-guided composition. This task aims to seamlessly integrate user-provided objects into a specific visual context. Current diffusion-based methods often involve costly instance-based optimization or finetuning of pretrained models on customized datasets, which can potentially undermine their rich prior. In contrast, TF-ICON can leverage off-the-shelf diffusion models to perform cross-domain image-guided composition without requiring additional training, finetuning, or optimization. Moreover, we introduce the exceptional prompt, which contains no information, to facilitate text-driven diffusion models in accurately inverting real images into latent representations, + +forming the basis for compositing. Our experiments show that equipping Stable Diffusion with the exceptional prompt outperforms state-of-the-art inversion methods on various datasets (CelebA-HQ, COCO, and ImageNet), and that TF-ICON surpasses prior baselines in versatile visual domains. Code is available at https://github.com/Shilin-LU/TF-ICON + +# 1. Introduction + +Image composition task involves incorporating unique objects from different photos to create a harmonious image within a specific visual context, a.k.a. image-guided composition. For instance, consider the scenario where one desires to incorporate a beloved panda into one's favorite artwork, e.g., oil or sketchy painting. The objective is to create a new image where the panda blends seamlessly into the scene without altering the appearance of the panda and the background, just as an artist meticulously crafted this panda + +for that artwork (See Figure 1). This task is inherently challenging, as it requires maintaining illumination consistency and preserving identifying features. The challenge is further compounded when the photos come from various domains. + +While recently large-scale text-to-image models [10, 19, 51, 56, 59, 61, 78] have achieved remarkable success in text-driven image generation, the ambiguity inherent in natural language presents challenges in conveying precise and nuanced visual details, even with highly detailed text prompts. Although this challenge is effectively addressed by enabling personalized concept learning [24, 25, 34, 37, 60], these methods require costly instance-based optimization and are limited in generating concepts with specified backgrounds. Recent studies [67, 77] have shown that diffusion models can achieve image-guided composition by explicitly incorporating additional guiding images. However, these models are retrained from the pretrained diffusion model on tailored datasets, which can damage the rich prior of the model. As a result, these models have limited compositional abilities beyond their training domain and still require significant computational resources. + +Given the wealth of large text-to-image models that have been trained on extensive language-image datasets, we pose a question: how could these models be leveraged for image-guided composition without incurring costly training or finetuning, thereby avoiding damaging the diverse prior? To answer it, we propose the Training-Free Image CompositionN (TF-ICON) framework, which equips attention-based text-to-image diffusion models with the capability to perform image-guided composition without requiring additional training, fine-tuning, extra data, or optimization. To the best of our knowledge, this is the first training-free framework developed for image-guided composition. The framework is compatible with various diffusion model samplers, enabling completion within 20 steps, and harnesses rich semantic knowledge to facilitate image-guided compositions across diverse domains (see Figure 1). + +Our approach constitutes an image-guided composition interface through denoising from a reliable starting latent code with the injection of composite self-attention maps. Finding the latent code that allows for reconstructing an input image while maintaining its editability, a.k.a. image inversion, is a challenging yet crucial step for state-of-the-art (SOTA) image editing frameworks involving real images [15, 26, 35, 38, 48, 53, 54, 70]. For diffusion models, while denoising diffusion implicit models (DDIM) inversion [65] has been effective for unconditional diffusion models, it falls short for text-driven diffusion models [26, 50, 70, 71]. To circumvent this, we introduce the exceptional prompt to accurately invert real images into latent codes upon pretrained text-to-image models to serve for further composition generation. The accurate latent codes are composed as the starting noise for the diffusion process. Through the + +gradual injection of composite self-attention maps that are specifically designed to reflect the relations between guiding images, we are able to infuse contextual information from the background into the incorporated objects, which results in harmonious image-guided compositions. + +To summarize, we make the following key contributions: + +1. We demonstrate the superior performance of high-order diffusion ODE solvers compared to commonly used DDIM inversion for real image inversion. +2. We present an exceptional prompt that allows text-driven models to achieve accurate invertibility, laying a solid groundwork for subsequent editing. Experimental results show that it surpasses SOTA inversion methods on three vision datasets. +3. We propose the first training-free framework that enables cross-domain image-guided composition for attention-based diffusion models. +4. We demonstrate quantitatively and qualitatively that our framework outperforms prior baselines for image-guided composition. + +# 2. Related Work + +Image composition. Image composition is widely applied to electronic commerce, entertainment, and data augmentation [20, 43] for downstream tasks. It can be broadly categorized into two types: text-guided [5, 6, 11, 22, 45] and image-guided [8, 23, 39, 67, 76, 77, 79]. The former involves composing multiple objects specified by only a text prompt without limiting the appearance of objects, as long as their semantics align with the prompt. Despite the great successes of text-conditioned models, they are often prone to semantic errors [22, 56], especially when the text prompt involves multiple objects. These errors include attribute leakage, attribute interchange, and missing objects, which cause the generated images to critically different from the user's intention [22, 56]. As a result, extensive prompt engineering [74] is often necessary to achieve the desired results. In contrast to text-only guided composition, image-guided composition involves incorporating specific objects and scenarios from user-provided photos, potentially with the aid of a text prompt. However, due to the nature of involving additional real images, it is more challenging particularly when images from different visual domains. Conventionally, image-guided composition is divided into several sub-tasks [52], such as object placement [7, 12, 40, 69, 80], image blending [75, 81], image harmonization [14, 16, 30, 76], and shadow generation [29, 42, 63, 83], each of which is typically addressed by different models and pipelines. + +Image inversion. Extensive research has been conducted on image inversion for GANs, including latent-based op + +![](images/43a6bd7a936544f4d728367a28a90fd28694c1e04433d528b991790bc4dbcfeb.jpg) +(a) Image Inversion + +![](images/7fc4b2d490e4202927ec7ac486839f44305a5ac31372b5187ea219137b06ab2f.jpg) + +![](images/707be35a11950f6fa56ef34f8f7b924d43924b6f2a53365461653c62ad4f8c89.jpg) +(b) Training-Free Image Composition (TF-ICON) +Figure 2: The proposed training-free image composition framework. (a) The exceptional prompt is used to invert the main and reference images into noises $\mathbf{x}_T^{\mathrm{m}}$ , $\mathbf{x}_T^{\mathrm{r}}$ , which are then composed to form the starting point $\mathbf{x}_T^*$ for the composition process. (b) Three constituents are composed for injecting into the composition process at early timesteps, including self-attention maps from the main and reference image reconstruction processes, as well as cross attention between the main and reference images. For better clarity and readability, the original main and reference images are shown in the pixel space instead of the VAE latent space, and the reference image is presented without resizing and zero-padding. + +tization [1, 2, 33], encoders [3, 57, 68], and fine-tuning [4, 58]. For diffusion models, DDIM [65] is a widely used technique for inversion in image editing frameworks. However, in text-driven settings, DDIM leads to significant reconstruction distortion due to the instability resulting from classifier-free guidance (CFG) [18, 28]. Recently, null-text inversion [50] has been proposed to achieve accurate inversion by optimizing the unconditional prediction of the text-to-image model. It demonstrates promising results but requires instance-based optimization. Concurrently, EDICT [71] also achieves near-perfect inversion, albeit doubling the computation time of the diffusion process. + +# 3. Preliminary + +Diffusion probabilistic models (DPM) [18, 27, 59, 64] are generative models in which an image is generated by progressively denoising from Gaussian noise. The forward diffusion process gradually perturbs data with infinite noise scales, which can be modeled as the solution of a stochastic differential equation (SDE) $\{\mathbf{x}_t\}_{t=0}^T$ . Formally, given a data sample $\mathbf{x}_0 \sim p_0 = p_{\mathrm{data}}$ , random noise is gradually injected, eventually resulting in a sample $\mathbf{x}_T$ which is typically distributed as a tractable prior $p_T$ without any infer + +mation of $p_0$ , as described by the following SDE [36, 66]: + +$$ +\mathrm {d} \mathbf {x} _ {t} = \mathbf {f} \left(\mathbf {x} _ {t}, t\right) \mathrm {d} t + g (t) \mathrm {d} \mathbf {w} _ {t}, \tag {1} +$$ + +where $\mathbf{w}_t\in \mathbb{R}^d$ is the standard Wiener process (a.k.a., Brownian motion), and $\mathbf{f}(\cdot ,t)$ and $g(t)$ are commonly designated as the drift and diffusion coefficient, respectively. On the other hand, the reverse diffusion process can be described by the reverse-time SDE from $T$ to 0 [66]: + +$$ +\mathrm {d} \mathbf {x} _ {t} = [ \mathbf {f} (\mathbf {x} _ {t}, t) - g (t) ^ {2} \nabla_ {\mathbf {x}} \log p _ {t} (\mathbf {x} _ {t}) ] \mathrm {d} t + g (t) \mathrm {d} \bar {\mathbf {w}} _ {t}, \tag {2} +$$ + +where $\bar{\mathbf{w}}_t$ is the Wiener process in the reverse time. The score function $\nabla_{\mathbf{x}}\log p_t(\mathbf{x}_t)$ is the only unknown term and can be estimated by a neural network $\epsilon_{\theta}(\mathbf{x}_t,t)$ whose parameter $\theta$ is optimized by a denoising objective [27, 66]. + +Upon attaining the trained model that predicts the score function accurately, it can be utilized to numerically solve the reverse SDE (Eq. (2)), enabling the generation of samples from a noise distribution. Song et al. [66] outline various methods, including Variance Exploding (VE), Variance Preserving (VP), and sub-VP SDE, for constructing SDEs that perturb the unknown data distribution into a fixed prior. In this work, we leverage the pre-trained text-to-image La + +tent Diffusion Model (LDM) [59], a.k.a. Stable Diffusion, which applies the VP SDE in the latent space. + +# 4. Method + +Our objective is to utilize a main (background) image $\mathbf{I}^{\mathrm{m}}$ , a reference (foreground) image $\mathbf{I}^{\mathrm{r}}$ , a text prompt $\mathcal{P}$ , and a binary mask $\mathbf{M}^{\mathrm{user}}$ which designates the region of interest within the main image, to generate a modified image $\mathbf{I}^{*}$ . The resultant image $\mathbf{I}^{*}$ should contain the reference subject with identifying features within the mask, i.e. $id(\mathbf{I}^{*} \odot \mathbf{M}^{\mathrm{user}}) \approx id(\mathbf{I}^{\mathrm{r}})$ , while concurrently ensuring that the complementing area closely resembles the main image, i.e. $\mathbf{I}^{*} \odot (\mathbf{1} - \mathbf{M}^{\mathrm{user}}) \approx \mathbf{I}^{\mathrm{m}} \odot (\mathbf{1} - \mathbf{M}^{\mathrm{user}})$ . Moreover, it is ideal for the transition between the areas inside and outside the mask to be imperceptible. + +We propose a training-free framework that can make use of attention-based pre-trained text-to-image models to perform image-guided composition. To the best of our knowledge, it is the first training-free framework for image-guided composition, which can be accomplished within 20 steps of sampling. The framework is mainly comprised of two steps: image inversion (Section 4.1), and composition generation (Section 4.2), as shown in Figure 2. The full algorithm is presented in Appendix B.2. + +# 4.1. Image Inversion with Exceptional Prompt + +Achieving precise manipulation of real images often necessitates an accurate inversion process that identifies the corresponding latent representation, which not only provides editability for meaningful manipulation but also accurately reconstructs the input image [26, 68]. For diffusion models, optimal editability is typically characterized by a noise encoding that conforms to the ideal statistical properties of zero-mean, unit-variance Gaussian noise [53]. + +ODE inversion. Most diffusion frameworks for image editing [15, 26, 35, 38, 53, 70] use DDIM inversion to invert the real image into its latent representation. However, our findings suggest that this may not be the optimal choice for inverting real images. It has been proven that DDIM is a first-order discretization of the associated probability flow ordinary differential equations (ODE) of Eq. (2) [62, 65, 66]: + +$$ +\mathrm {d} \mathbf {x} _ {t} = \left[ \mathbf {f} \left(\mathbf {x} _ {t}, t\right) - \frac {1}{2} g (t) ^ {2} \nabla_ {\mathbf {x}} \log p _ {t} \left(\mathbf {x} _ {t}\right) \right] \mathrm {d} t, \tag {3} +$$ + +which can be solved using $\epsilon_{\theta}(\mathbf{x}_t,t)$ and shares the consistent marginal probability distribution $\{p_t(\mathbf{x}_t)\}_{t = 0}^T$ with Eq. (2). Various samplers [32, 44, 46, 47] have been developed for solving the diffusion ODE starting from noise $\mathbf{x}_T$ to achieve fast sampling (10~20 steps). We offer the insight that utilizing these ODE solvers in turn as encoders starting from the real image $\mathbf{x}_0$ yields better latent representation $\mathbf{x}_T$ , compared with those obtained through commonly used + +![](images/3f46d3803dea2136fd0ad6e853be951448157d879c981d8fd87998538df68106.jpg) +(a) Original + +![](images/58d7ffc51efc7a9fb2567a5e35f8c464bafd691b2a196c15014feef48344d7b3.jpg) +(b) CFG + +![](images/c701696a7df507f13a8f500591db1780aac298926a05d5e765784902f68b9004.jpg) +(c) Cond. + +![](images/ea450ef038a402b3765028bf1f78f33eb3abcdebfb9906bc3ab31d749af88288.jpg) +(d) Uncond. + +![](images/7f8820de931ef5f6ccc769ac3ab0caee3950b8dec11c3db166650896a4bb730e.jpg) +(e) Ours +Figure 3: The real image reconstruction results using Stable Diffusion with (b) classifier-free guidance (CFG) output $\hat{\epsilon}_{\theta}(\mathbf{x}_t,t,\mathcal{E},\varnothing)$ ; (c) conditional output $\epsilon_{\theta}(\mathbf{x}_t,t,\mathcal{E})$ ; (d) unconditional output $\epsilon_{\theta}(\mathbf{x}_t,t,\varnothing)$ ; and (e) ours. The prompts for (b) and (c) are 'a photo of a singer', 'a photo of a car', and 'an oil painting'. See Appendix A.3 for elaboration. + +DDIM. A quantitative analysis is given in Appendix A.1. The enhanced alignment between the forward and backward ODE trajectories in the high-order DPM-Solver++ [47] implies that it is better suited for real image inversion. Thus, this paper employs it for all inversions of diffusion models. + +Exceptional prompt. In the unconditional setting $\epsilon_{\theta}(\mathbf{x}_t,t)$ solving the diffusion ODE (Eq. (3)) from 0 to $T$ enables us to obtain the better latent code $\mathbf{x}_T$ for the real image $\mathbf{x}_0$ . However, in the text-driven setting $\epsilon_{\theta}(\mathbf{x}_t,t,\mathcal{E})$ , existing image editing works [26, 50, 70, 71] have shown that the inversion process is prone to significant reconstruction errors, due to the instability induced by CFG [18, 28]: + +$$ +\hat {\epsilon} _ {\theta} \left(\mathbf {x} _ {t}, t, \mathcal {E}, \varnothing\right) = s \cdot \epsilon_ {\theta} \left(\mathbf {x} _ {t}, t, \mathcal {E}\right) + (1 - s) \cdot \epsilon_ {\theta} \left(\mathbf {x} _ {t}, t, \varnothing\right), \tag {4} +$$ + +where $\varnothing = \psi("")$ and $\mathcal{E} = \psi(\mathcal{P})$ are embeddings of the null and normal prompt, and $s$ is the guidance scale. Our experiments further reveal that even without CFG, both conditional output $\epsilon_{\theta}(\mathbf{x}_t,t,\mathcal{E})$ and unconditional output $\epsilon_{\theta}(\mathbf{x}_t,t,\emptyset)$ of text-to-image diffusion models still produce large reconstruction errors, as depicted in Figure 3. + +To achieve accurate inversion, we present a straightforward yet effective solution, namely exceptional prompt, $\mathcal{P}_{\text{exceptional}}$ . Intuitively, any information contained within the input prompt can result in the deviation of the backward ODE trajectories from the forward trajectories. Hence, we remove all information by setting all token numbers to a common value and eliminating positional embeddings for the text prompt, as depicted in Figure 4. Importantly, the exceptional prompt is distinguished from the null prompt by its absence of special tokens, such as [startoftext], [endoftext], and [pad], which still retain information. The exceptional prompt is applied only in image inversion but not in the composition process. The choice of the + +![](images/da46054a4a9d364b4cbf71f21b761c34a32c05fe9ad0cb406a58c716affc4754.jpg) +Figure 4: The illustration of comparison among (a) Normal Prompt, (b) Null Prompt, and (c) Exceptional Prompt. + +token value does not significantly affect the inversion. The detailed analysis of the exceptional prompt and token value selection is provided in Appendix A.2 and A.4, respectively. + +Figure 3 visually demonstrates the effectiveness of the exceptional prompt. Our results $\epsilon_{\theta}(\mathbf{x}_t,t,\mathcal{W})$ with the exceptional prompt embedding $\mathcal{W} = \psi (\mathcal{P}_{\mathrm{exceptional}})$ are more visually accurate than others. The quantitative experiments are shown in Section 5.1. All the results in Figure 3 are obtained by solving the forward and backward diffusion ODEs using the second-order DPM-Solver++ [47] in 20 steps. + +# 4.2. Training-Free Image Composition + +Upon equipping the accurate invertibility, image composition can be performed based on it. The composition process consists of two key components: noise incorporation and composite self-attention maps injection. + +Noise incorporation. Before inverting images into noises, a simple preprocessing step is necessary for the reference image. Typically, only the foreground in the reference is desired for composition, so the preprocessing step involves using a pretrained segmentation model [84] to remove the background, resizing and repositioning the object to match the user's mask in the main image, and padding it with zeros to ensure it is the same size as the main image (See Appendix B.1 for visual illustration). + +Once the preprocessing is complete, the main and reference images are inverted to corresponding noises $\mathbf{x}_T^{\mathrm{m}}$ and $\mathbf{x}_T^{\mathrm{r}}$ by solving diffusion ODEs (Eq. (3)) from 0 to $T$ with the exceptional prompt $\mathcal{P}_{\mathrm{exceptional}}$ . $\mathbf{x}_T^{\mathrm{m}}$ and $\mathbf{x}_T^{\mathrm{r}}$ are then merged with standard Gaussian noise $\mathbf{z}$ to create the starting point $\mathbf{x}_T^*$ for generating the composition. Formally, the incorporated noise $\mathbf{x}_T^*$ is calculated by + +$$ +\mathbf {x} _ {T} ^ {*} = \mathbf {x} _ {T} ^ {\mathrm {r}} \odot \mathbf {M} ^ {\text {s e g}} + \mathbf {x} _ {T} ^ {\mathrm {m}} \odot (\mathbf {1} - \mathbf {M} ^ {\text {u s e r}}) + \mathbf {z} \odot (\mathbf {M} ^ {\text {u s e r}} \oplus \mathbf {M} ^ {\text {s e g}}), \tag {5} +$$ + +where $\mathbf{z} \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , $\mathbf{M}^{\mathrm{user}}$ is the user mask, $\mathbf{M}^{\mathrm{seg}}$ is the segmentation mask for reference image, and $\mathbf{M}^{\mathrm{user}} \oplus \mathbf{M}^{\mathrm{seg}}$ is the XOR of them, which is the transition area. The incorporation of $\mathbf{z}$ enhances the smoothness of the transition between the regions inside and outside the user mask, by effectively leveraging the prior knowledge of the text-driven diffusion model to inpaint the transition area. Empirically, + +![](images/a62e7ac9c1ef73f3cab714b7f283d8bd8c74a872531768822e4ccc1d95442fbd.jpg) +Figure 5: A toy example for attention composition. + +for cross-domain composition, incorporating the starting point in the noise space usually is more effective, while solely for photorealism, composing in the pixel/latent space and then inverting it as the starting point is more favorable. + +Composite self-attention map injection. The incorporated noise $\mathbf{x}_T^*$ is employed as the starting point for solving the diffusion ODE from $T$ to 0 with a normal prompt $\mathcal{P}$ to ultimately generate the composition. $\mathcal{P}$ is intended to assist in inpainting transition areas. However, relying solely on noise incorporation, the pretrained text-to-image model cannot preserve the appearance of the main and reference images effectively, as shown in Figure 7. To tackle this problem, we propose injecting composite self-attention maps in a specially designed manner, as the semantic information is basically retained within the rows and columns of self-attention maps (See Appendix E for visual illustrations). + +The composite self-attention map comprises three constituents: two self-attention maps, $\mathbf{A}_{l,t}^{\mathrm{m}}$ and $\mathbf{A}_{l,t}^{\mathrm{r}}$ , corresponding to the main and reference images, and a cross-attention map, $\mathbf{A}_{l,t}^{\mathrm{cross}}$ , calculated between them. The composition way is illustrated in Figure 5. To compose the reference image in the blue regions of Figure 5 (a), its self-attention map $\mathbf{A}_{l,t}^{\mathrm{r}}$ should be placed in the corresponding blue regions of Figure 5 (b), since the 5th patch can only attend to the patches 6, 9, and 10 in the context of self-attention. The green regions in Figure 5 (b) should contain the cross-attention map, $\mathbf{A}_{l,t}^{\mathrm{cross}}$ , which infuses contextual information from the white regions into the blue regions. If the green regions are preserved as the self-attention of the white regions without replacement, the information stored there only reflects the relation between the original patches, such as the 0th and original 5th patches in the example of index (5,0) or (0,5). This results in a lack of surrounding information being provided to the new 5th patch, such as painting or sketching, which is necessary for seamless object transition to other domains (See ablation in Figure 7). + +Essential constituents $\mathbf{A}_{l,t}^{\mathrm{m}},\mathbf{A}_{l,t}^{\mathrm{r}},\mathbf{A}_{l,t}^{\mathrm{cross}}$ are calculated using self-attention modules of the pretrained Stable Diffusion. Typically, a self-attention module at layer $l$ contains three projection matrices $\mathbf{W}_l^q,\mathbf{W}_l^k,\mathbf{W}_l^v$ in the same dimension $\mathbb{R}^{d\times d}$ . Denote the spatial features of the main and reference image at timestep $t$ and layer $l$ as $\mathbf{f}_{l,t}^{\mathrm{m}}\in \mathbb{R}^{(h\times w)\times d}$ + +and $\mathbf{f}_{l,t}^{\mathrm{r}}\in \mathbb{R}^{(h'\times w')\times d}$ , respectively, where $h^\prime \times w^\prime$ is the size of the reference image after resizing to match the size of the user mask. The queries, keys, and values for each self-attention module are obtained as: + +$$ +\mathbf {q} _ {l, t} ^ {\mathrm {m}} = \mathbf {f} _ {l, t} ^ {\mathrm {m}} \mathbf {W} _ {l} ^ {q}, \quad \mathbf {k} _ {l, t} ^ {\mathrm {m}} = \mathbf {f} _ {l, t} ^ {\mathrm {m}} \mathbf {W} _ {l} ^ {k}, \quad \mathbf {v} _ {l, t} ^ {\mathrm {m}} = \mathbf {f} _ {l, t} ^ {\mathrm {m}} \mathbf {W} _ {l} ^ {v}, \tag {6} +$$ + +$$ +\mathbf {q} _ {l, t} ^ {\mathrm {r}} = \mathbf {f} _ {l, t} ^ {\mathrm {r}} \mathbf {W} _ {l} ^ {q}, \quad \mathbf {k} _ {l, t} ^ {\mathrm {r}} = \mathbf {f} _ {l, t} ^ {\mathrm {r}} \mathbf {W} _ {l} ^ {k}, \quad \mathbf {v} _ {l, t} ^ {\mathrm {r}} = \mathbf {f} _ {l, t} ^ {\mathrm {r}} \mathbf {W} _ {l} ^ {v}, \tag {7} +$$ + +where $\mathbf{q}_{l,t}^{\mathrm{m}},\mathbf{k}_{l,t}^{\mathrm{m}},\mathbf{v}_{l,t}^{\mathrm{m}}\in \mathbb{R}^{(h\times w)\times d}$ and $\mathbf{q}_{l,t}^{\mathrm{r}},\mathbf{k}_{l,t}^{\mathrm{r}},\mathbf{v}_{l,t}^{\mathrm{r}}\in$ $\mathbb{R}^{(h'\times w')\times d}$ . Thus, $\mathbf{A}_{l,t}^{\mathrm{m}},\mathbf{A}_{l,t}^{\mathrm{r}}$ and $\mathbf{A}_{l,t}^{\mathrm{cross}}$ are then calculated and composed as $\mathbf{A}_{l,t}^{*}$ for injection: + +$$ +\mathbf {A} _ {l, t} ^ {\mathrm {m}} = \operatorname {S o f t m a x} \left(\mathbf {q} _ {l, t} ^ {\mathrm {m}} \cdot \left(\mathbf {k} _ {l, t} ^ {\mathrm {m}}\right) ^ {\mathrm {T}} / \sqrt {d}\right), \tag {8} +$$ + +$$ +\mathbf {A} _ {l, t} ^ {\mathrm {r}} = \operatorname {S o f t m a x} \left(\mathbf {q} _ {l, t} ^ {\mathrm {r}} \cdot \left(\mathbf {k} _ {l, t} ^ {\mathrm {r}}\right) ^ {\mathrm {T}} / \sqrt {d}\right), \tag {9} +$$ + +$$ +\mathbf {A} _ {l, t} ^ {\text {c r o s s}} = \operatorname {S o f t m a x} \left(\mathbf {q} _ {l, t} ^ {\mathrm {m}} \cdot \left(\mathbf {k} _ {l, t} ^ {\mathrm {r}}\right) ^ {\top} / \sqrt {d}\right), \tag {10} +$$ + +$$ +\mathbf {A} _ {l, t} ^ {*} = \vartheta_ {\text {c o m p o s e}} \left(\mathbf {A} _ {l, t} ^ {\mathrm {m}}, \mathbf {A} _ {l, t} ^ {\mathrm {r}}, \mathbf {A} _ {l, t} ^ {\text {c r o s s}}\right), \tag {11} +$$ + +where $\mathbf{A}_{l,t}^{\mathrm{m}}\in \mathbb{R}^{(h\times w)\times (h\times w)}$ $\mathbf{A}_{l,t}^{\mathrm{r}}\in \mathbb{R}^{(h^{\prime}\times w^{\prime})\times (h^{\prime}\times w^{\prime})}$ $\mathbf{A}_{l,t}^{\mathrm{cross}}\in \mathbb{R}^{(h\times w)\times (h^{\prime}\times w^{\prime})}$ , and $\vartheta_{\mathrm{compose}}$ is the function to bulid composite self-attention maps $\mathbf{A}_{l,t}^{*}$ based on patch indices (Figure 5). + +As a result, three diffusion ODEs are solved simultaneously from $T$ to 0. As depicted in Figure 2 (b), ODEs start from the accurate inverted noises $\mathbf{x}_T^{\mathrm{r}},\mathbf{x}_T^{\mathrm{m}}$ and the interpolated noise $\mathbf{x}_T^*$ , respectively. The first two ODEs are solved using the exceptional prompt $\mathcal{P}_{\mathrm{exceptional}}$ to progressively reconstruct the main and reference, thus allowing for the precise retention of $\mathbf{A}_{l,t}^{\mathrm{m}}$ $\mathbf{A}_{l,t}^{\mathrm{r}}$ ,and $\mathbf{A}_{l,t}^{\mathrm{cross}}$ at each time step $t$ These attention maps are then composed and injected into the third ODE for generating a natural and cohesive composition with a normal prompt $\mathcal{P}$ + +To balance the generation of high-level context and finer details [13, 38], we set a threshold $\tau_{A}$ to determine the time steps for injecting composite self-attention maps in the early stage $(t\in [T\times \tau_A,T])$ and allow the model to explore ODE trajectories through a normal prompt $\mathcal{P}$ in the later stage $(t\in [0,T\times \tau_A])$ , guided by the prior of the pretrained model. However, this freedom, without the imposition of attention injection constraints, often results in deviations from the desired background (see Figure 7). Thus, similar to [6], we set an additional threshold, denoted as $\tau_{B}$ , which regulates the trajectory rectification process. This process entails replacing the regions outside the user mask with the reconstructed main image at various time steps, i.e., $\hat{\mathbf{x}}_t^* = \mathbf{x}_t^*\odot \mathbf{M}^{\mathrm{user}} + \mathbf{x}_t^{\mathrm{m}}\odot (\mathbf{1} - \mathbf{M}^{\mathrm{user}})$ , where $t\in [T\times \tau_B,T]$ . Note that only preserving the background at the final step can lead to noticeable artifacts, as shown in Appendix B.3. + +# 5. Experiments + +This section consists of two sets of experiments. The first set assesses the effectiveness of the exceptional prompt + +Table 1: The reconstruction comparison on CelebA-HQ. + +
MethodMAE ↓LPIPS ↓SSIM ↑
OptimizationI2S [1]0.0640.1340.872
PTI [58]0.0620.1320.877
EncoderpSp [57]0.0790.1690.793
e4e [68]0.0920.2210.742
ReStyle w/ pSp [3]0.0730.1450.823
ReStyle w/ e4e [3]0.0890.2020.758
HFGI w/ e4e [73]0.0620.1270.877
DiffusionSD w/ CFG0.1340.3400.637
SD w/ Cond.0.1260.3080.654
SD w/ Uncond.0.1260.3040.655
DiffusionCLIP [35]0.0200.0730.914
Ours0.0190.0470.918
Upper BoundVQAE [21]0.0180.0430.919
+ +Table 2: The further reconstruction comparison on COCO and ImageNet. \*: an upper bound. + +
MethodMSCOCO (5000)ImageNet (3000)
MAE ↓LPIPS ↓SSIM ↑MAE ↓LPIPS ↓SSIM ↑
SD w/ CFG0.1500.4580.5680.1320.4960.575
SD w/ Cond.0.1220.3590.6330.1090.3890.645
SD w/ Uncond.0.1200.3630.6360.1140.4060.635
Ours0.0300.0730.8680.0330.0870.852
VQAE* [21]0.0300.0690.8700.0320.0840.854
+ +(Section 5.1). The second set evaluates our image composition framework qualitatively and quantitatively (Section 5.2), followed by an ablation study (Section 5.3). + +# 5.1. Image Reconstruction + +To assess the effectiveness of the exceptional prompt, we compared its performance with SOTA GAN [1, 3, 57, 58, 68, 73] and diffusion [35] inversion methods on the CelebA-HQ [31], following the same setting as described in [35, 73]. Additionally, we conducted experiments on the ImageNet [17] and COCO [41] with Stable Diffusion to further validate our findings. Our results (Tables 1 and 2) show that the exceptional prompt is highly effective in producing reconstructions that closely approximate the upper bound established by the vector quantized autoencoder (VQAE) [21] across all metrics, including MAE, LPIPS [82], and SSIM. The qualitative comparison is shown in Figure 3. All results of Stable Diffusion, including ours, are sampled by the second-order DPM-Solver++ [47]. Experimental settings are detailed in Appendix B.4. + +# 5.2. Image Composition Comparisons + +Test benchmark. As there is currently no benchmark for testing cross-domain image-guided composition as a whole, we developed a test benchmark containing 332 samples. Each sample in the benchmark consists of a main (background) image, a reference (foreground) image, a user mask, and a text prompt. The main images comprise + +![](images/7ee06532bde463cc64858ae73c987269354cc5778dfcf46f8e148bd4dc1e0cfe.jpg) +'acartoon animation of a cake,acroissant,a piece of bread anda cup of coffee +Figure 6: Qualitative comparison with SOTA and concurrent baselines in image-guided composition for sketching, photorealism, painting, and cartoon animation domains. Additional results are available in Appendix H.2. + +Table 3: Quantitative evaluation results for image composition in the photorealism domain. + +
MethodLPIPS(BG)↓LPIPS(FG)↓CLIP(Image)↑CLIP(Text)↑
SDEdit (0.4) [48]0.350.6280.5627.73
SDEdit (0.6) [48]0.420.6677.6827.98
Blended [5]0.110.7773.2525.19
Paint [77]0.130.7380.2625.92
DIB [81]0.110.6377.5726.84
Ours0.100.6082.8628.11
+ +four visual domains: photorealism, pencil sketching, oil painting, and cartoon animation. All reference images are from the photorealism domain as the reference requires segmentation models, which are generally more effective in this domain. Further details are available in Appendix G. + +Qualitative comparisons. Our qualitative comparisons are performed across four visual domains, employing SOTA and concurrent baselines that are applicable to image-guided composition, including Deep Image Blending (DIB) [81], DCCF [76], Blended Diffusion [5], Textual Inversion [24], Paint by Example [77], and SDEdit [48] under two different noising levels. As shown in Figure 6, our framework is capable of seamlessly composing objects into various domains while maintaining their identities. In contrast, DIB and DCCF fall short in processing the transition areas, leading to noticeable artifacts. Blended Diffusion's foreground generation and Textual Inversion's + +background generation relies solely on text prompts, causing deviations from the user's intention. While Paint by Example effectively composes images within its photorealistic training domain, it struggles to adapt to other domains. Additionally, SDEdit with fewer timesteps is suitable for image composition in terms of preserving the identifying features of the reference, but the background is changed. See Appendix H.2 for additional comparisons. + +Quantitative analysis. The baselines are primarily trained in the photorealism domain, where the objective metrics are more effective; therefore, we focused our quantitative comparison within this domain and relied on user study for comparison in other domains. We assess the same baselines as in the qualitative comparison, with the exception of Textual Inversion [24], which involves instance-based optimization, and DCCF [76], which is used for harmonizing images after copy-and-paste operations. Four metrics are considered: (1) $\mathrm{LPIPS}_{\mathrm{(BG)}}$ [82] measures the background consistency, (2) $\mathrm{LPIPS}_{\mathrm{(FG)}}$ [82] evaluates the low-level similarity between the edited region and the reference foreground, (3) $\mathrm{CLIP}_{\mathrm{(Image)}}$ [55] evaluates the semantic similarity between the edited region and the reference in the CLIP embedding space, and (4) $\mathrm{CLIP}_{\mathrm{(Text)}}$ [55] measures the semantic alignment between the text prompt and the resultant image. As presented in Table 3, our method outperforms all baselines. We achieve well preservation + +Table 4: User study: higher score, better ranking. P: photorealism; O: oil painting; S: sketchy painting; C: cartoon. + +
MethodP & PP & OP & SP & CTotal
Blended [5]1.8072.3142.6802.1002.093
SDEdit (0.6) [48]2.3683.0632.4092.7132.485
Paint [77]2.8792.3062.0432.6732.666
DCCF [76]3.8383.2373.2973.4703.583
Ours4.1084.0804.5714.0434.175
+ +Table 5: Ablation study: quantitative comparison of various variants of our framework. + +
ConfigLPIPS(BG) ↓LPIPS(FG) ↓CLIP(Image) ↑CLIP(Text) ↑
Baseline0.340.6575.1329.00
+ Pexceptional0.320.6478.5428.23
+ SA injection0.250.6181.6428.58
+ CA injection0.260.6081.6328.52
+ Background0.100.6082.8628.11
+ +of the background, high object correspondence in both low-level and high-level feature spaces, as well as a high degree of alignment with the text prompt. + +User study. We conducted a user study to compare image composition baselines across domains. We recruited 50 participants via Amazon and tasked them with completing 40 ranking questions. Each question consists of 5 options, generated using distinct methods. Details are available in Appendix D. The ranking criteria comprehensively considered foreground preservation, background consistency, seamless composition, and text alignment. The results are listed in Table 4, where the domain information is presented in the format of 'foreground domain & background domain', e.g., photorealism & oil painting. Our method was favored by most participants across different domains. + +# 5.3. Ablation Study + +We ablate our key design choices in the following cases: (1) Baseline, where the composition is generated by solving the diffusion ODE from $T$ to 0 using DPM-Solver++ without any injection. The starting point is composed by inverted noises under the normal prompt; (2) The exceptional prompt is applied to obtain accurate inverted noises; (3) The self-attention maps $\mathbf{A}_{l,t}^{\mathrm{m}}$ and $\mathbf{A}_{l,t}^{\mathrm{r}}$ are composed and then injected; (4) The cross-attention $\mathbf{A}_{l,t}^{\mathrm{cross}}$ between the main and reference images is further composed for injection; (5) The background preservation is applied. + +Table 5 presents the quantitative results, which indicate that the complete algorithm outperforms other variants in all metrics except for $\mathrm{CLIP}_{\mathrm{(Text)}}$ . Notably, the baseline achieves the best $\mathrm{CLIP}_{\mathrm{(Text)}}$ as it generates compositions solely relying on the normal prompt without any extra constraint. While metrics alone may not reveal the complete effectiveness of cross-attention injection, Figure 7 illustrates that the interactions between the main and reference images + +Figure 7: Ablation study of different variants of our framework. SA: self-attention. CA: cross-attention. +![](images/3af29284b64f3d0ea6355a8598b0fbd45ece9f54c435e9369fcd1dc3969afdb1.jpg) +'an oil painting of a burrito, Van Gogh Style + +are highly beneficial to both foreground and background in terms of preserving appearance and switching domains. Note that preserving the background at different noise levels affects both background and foreground, which is distinct from preserving solely at the final step (Appendix B.3). Additional ablation results are shown in Appendix H.2. + +# 6. Limitations and Future Work + +The primary limitation of our work lies in its inability to generate an object view that critically differs from the given reference. As a result, the choice of reference image may be restricted at times. This is because the model relies on self-attention maps to provide layout and appearance information, which in turn constrains the development of alternative views. While introducing a loose self-attention injection can generate different views, this often compromises the preservation of the object's appearance. To overcome this, further research could explore utilizing personalized concept learning techniques, such as Textual Inversion [24], to encode identity information in the text prompt through special embeddings. Alternatively, utilizing NeRF-relevant techniques [9, 49, 72] can generate other views of the object for a specific scene, but this can require expensive training. Furthermore, due to the fact that our approach relies on Stable Diffusion, it inherits its shortcomings and biases, which may result in producing artifacts in certain scenarios. + +# 7. Conclusion + +We introduced a method that leverages the high-order ODE solver with the exceptional prompt to achieve precise inversion of real images, which serves as a foundation for further manipulation. Building upon this, we propose a novel training-free framework, TF-ICON, that enables attention-based text-to-image diffusion models to perform image-guided composition across different domains. Our experimental results demonstrate that our approach outperforms the SOTA baselines for both image inversion and composition. We believe that image composition has the potential to become an essential tool for content creators, offering significant benefits for downstream applications in various industries. + +# Acknowledgement + +This research is supported by the National Research Foundation, Singapore under its Strategic Capability Research Centres Funding Initiative. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore. + +# References + +[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3, 6 +[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 3 +[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 3, 6 +[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF conference on computer Vision and pattern recognition, pages 18511-18521, 2022. 3 +[5] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 2, 7, 8 +[6] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2, 6 +[7] Samaneh Azadi, Deepak Pathak, Sayna Ebrahimi, and Trevor Darrell. Compositional gan: Learning image-conditional binary composition. International Journal of Computer Vision, 128(10):2570-2585, 2020. 2 +[8] Andrew Brown, Cheng-Yang Fu, Omkar Parkhi, Tamara L Berg, and Andrea Vedaldi. End-to-end visual editing with a generatively pre-trained artist. arXiv preprint arXiv:2205.01668, 2022. 2 +[9] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Consistent single-view perpetual view generation with conditional diffusion models. arXiv preprint arXiv:2211.12131, 2022.8 +[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2 +[11] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. arXiv preprint arXiv:2301.13826, 2023. 2 + +[12] Bor-Chun Chen and Andrew Kae. Toward realistic image compositing with adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8415-8424, 2019. 2 +[13] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 6 +[14] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8394-8403, 2020. 2 +[15] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2, 4 +[16] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 2 +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3, 4 +[19] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022. 2 +[20] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In Proceedings of the IEEE international conference on computer vision, pages 1301-1310, 2017. 2 +[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 6 +[22] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 2 +[23] Oran Gafni and Lior Wolf. Wish you were here: Context-aware human generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7840-7849, 2020. 2 +[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 7, 8 +[25] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Designing an encoder + +for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 2 +[26] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 4 +[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3 +[28] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3, 4 +[29] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 914-922, 2022. 2 +[30] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 2 +[31] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6 +[32] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 4 +[33] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3 +[34] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2 +[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2, 4, 6 +[36] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3 +[37] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022. 2 +[38] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2, 4, 6 +[39] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. 2023. 2 +[40] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative adversarial networks for image compositing. In Proceedings + +of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9455-9464, 2018. 2 +[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[42] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8139-8148, 2020. 2 +[43] Liu Liu, Zhenchen Liu, Bo Zhang, Jiangtong Li, Li Niu, Qingyang Liu, and Liqing Zhang. Opa: object placement assessment dataset. arXiv preprint arXiv:2107.01889, 2021. 2 +[44] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 4 +[45] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. arXiv preprint arXiv:2206.01714, 2022. 2 +[46] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 4 +[47] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 4, 5, 6 +[48] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 2, 7, 8 +[49] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 8 +[50] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 2, 3, 4 +[51] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[52] Li Niu, Wenyan Cong, Liu Liu, Yan Hong, Bo Zhang, Jing Liang, and Liqing Zhang. Making images real again: A comprehensive survey on deep image composition. arXiv preprint arXiv:2106.14490, 2021. 2 + +[53] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. arXiv preprint arXiv:2302.03027, 2023. 2, 4 +[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2 +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[57] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 3, 6 +[58] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 3, 6 +[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4 +[60] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2 +[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2 +[62] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022.4 +[63] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4380-4390, 2021. 2 +[64] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3 +[65] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3, 4 + +[66] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3, 4 +[67] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Objectstitch: Generative object compositing. arXiv preprint arXiv:2212.00932, 2022. 2 +[68] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 3, 4, 6 +[69] Shashank Tripathi, Siddhartha Chandra, Amit Agrawal, Ambrish Tyagi, James M Rehg, and Visesh Chari. Learning to generate synthetic data via compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 461-470, 2019. 2 +[70] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 4 +[71] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446, 2022. 2, 3, 4 +[72] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022.8 +[73] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 6 +[74] Sam Witteveen and Martin Andrews. Investigating prompt engineering in diffusion models. arXiv preprint arXiv:2211.15462, 2022. 2 +[75] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Gp-gan: Towards realistic high-resolution image blending. In Proceedings of the 27th ACM international conference on multimedia, pages 2487-2495, 2019. 2 +[76] Ben Xue, Shenghui Ran, Quan Chen, Rongfei Jia, Binqiang Zhao, and Xing Tang. Dccf: Deep comprehensible color filter learning framework for high-resolution image harmonization. In European Conference on Computer Vision, pages 300-316. Springer, 2022. 2, 7, 8 +[77] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022. 2, 7, 8 +[78] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2 +[79] He Zhang, Jianming Zhang, Federico Peruzzi, Zhe Lin, and Vishal M Patel. Deep image compositing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 365-374, 2021. 2 + +[80] Lingzhi Zhang, Tarmily Wen, Jie Min, Jiancong Wang, David Han, and Jianbo Shi. Learning object placement by inpainting for compositional data augmentation. In European Conference on Computer Vision, pages 566-581. Springer, 2020. 2 +[81] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 231-240, 2020. 2, 7 +[82] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 7 +[83] Shuyang Zhang, Runze Liang, and Miao Wang. Shadowgan: Shadow synthesis for virtual objects with conditional adversarial networks. Computational Visual Media, 5(1):105-115, 2019. 2 +[84] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34:10326-10338, 2021. 5 \ No newline at end of file diff --git a/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/images.zip b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aee330cde49295672d1262d6deaad799db1317da --- /dev/null +++ b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89e968b35ea0d549ff0a19323dd3abe3db813c0243c1d44efcc8bd8ada9ec93f +size 866459 diff --git a/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/layout.json b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..648fa84067ef421e6a09865671d6116e19c95f4d --- /dev/null +++ b/2023/TF-ICON_ Diffusion-Based Training-Free Cross-Domain Image Composition/layout.json @@ -0,0 +1,11213 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "type": "text", + "content": "TF-ICON: Diffusion-Based Training-Free Cross-Domain Image Composition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 173, + 143, + 427, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 143, + 427, + 158 + ], + "spans": [ + { + "bbox": [ + 173, + 143, + 427, + 158 + ], + "type": "text", + "content": "Shilin Lu1 Yanzhu Liu2 Adams Wai-Kin Kong1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 41, + 158, + 548, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 158, + 548, + 186 + ], + "spans": [ + { + "bbox": [ + 41, + 158, + 548, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 41, + 158, + 548, + 186 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Nanyang Technological University, Singapore " + }, + { + "bbox": [ + 41, + 158, + 548, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 41, + 158, + 548, + 186 + ], + "type": "text", + "content": "Institute for Infocomm Research (I²R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 188, + 501, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 188, + 501, + 198 + ], + "spans": [ + { + "bbox": [ + 100, + 188, + 501, + 198 + ], + "type": "text", + "content": "shilin002@e.ntu.edu.sg, liu_yanzhu@i2r.a-star.edu.sg, adamskong@ntu.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 201, + 99, + 297 + ], + "blocks": [ + { + "bbox": [ + 50, + 201, + 99, + 297 + ], + "lines": [ + { + "bbox": [ + 50, + 201, + 99, + 297 + ], + "spans": [ + { + "bbox": [ + 50, + 201, + 99, + 297 + ], + "type": "image", + "image_path": "d26ed3cf3c91eb696cd946fd14db40e6479a9ab39befc57b3e2c7728153d0cec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 100, + 201, + 197, + 297 + ], + "blocks": [ + { + "bbox": [ + 100, + 201, + 197, + 297 + ], + "lines": [ + { + "bbox": [ + 100, + 201, + 197, + 297 + ], + "spans": [ + { + "bbox": [ + 100, + 201, + 197, + 297 + ], + "type": "image", + "image_path": "a71505ddcf5c4341f7eab4a1f77c152f11f92b046ef8504485bc601b156a9911.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 298, + 192, + 310 + ], + "lines": [ + { + "bbox": [ + 104, + 298, + 192, + 310 + ], + "spans": [ + { + "bbox": [ + 104, + 298, + 192, + 310 + ], + "type": "text", + "content": "' a professional photograph of an Eiffel tower in the distance, ultra realistic" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 198, + 201, + 295, + 297 + ], + "blocks": [ + { + "bbox": [ + 198, + 201, + 295, + 297 + ], + "lines": [ + { + "bbox": [ + 198, + 201, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 198, + 201, + 295, + 297 + ], + "type": "image", + "image_path": "70eea039ca2f245f29c9c5f0094a0da7cc9140d1f8963b3d7573cbbd9e19a7dd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 298, + 290, + 310 + ], + "lines": [ + { + "bbox": [ + 202, + 298, + 290, + 310 + ], + "spans": [ + { + "bbox": [ + 202, + 298, + 290, + 310 + ], + "type": "text", + "content": "' a professional photograph of a puppy on the beach, ultra realistic'." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 298, + 201, + 346, + 297 + ], + "blocks": [ + { + "bbox": [ + 298, + 201, + 346, + 297 + ], + "lines": [ + { + "bbox": [ + 298, + 201, + 346, + 297 + ], + "spans": [ + { + "bbox": [ + 298, + 201, + 346, + 297 + ], + "type": "image", + "image_path": "7ab917b90888d14ede31cb0f51afa86b19cb0afd288ee6917dede397707926f4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 347, + 201, + 444, + 297 + ], + "blocks": [ + { + "bbox": [ + 347, + 201, + 444, + 297 + ], + "lines": [ + { + "bbox": [ + 347, + 201, + 444, + 297 + ], + "spans": [ + { + "bbox": [ + 347, + 201, + 444, + 297 + ], + "type": "image", + "image_path": "eeb42dd5d159fe37c37357571000d991dd7d7e6278d1fcceff62582039928582.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 298, + 440, + 310 + ], + "lines": [ + { + "bbox": [ + 352, + 298, + 440, + 310 + ], + "spans": [ + { + "bbox": [ + 352, + 298, + 440, + 310 + ], + "type": "text", + "content": "'an oil painting of a hamburger, Van Gogh Style'" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 446, + 201, + 543, + 297 + ], + "blocks": [ + { + "bbox": [ + 446, + 201, + 543, + 297 + ], + "lines": [ + { + "bbox": [ + 446, + 201, + 543, + 297 + ], + "spans": [ + { + "bbox": [ + 446, + 201, + 543, + 297 + ], + "type": "image", + "image_path": "a1ad3be380ad420bde06da0ac289b5760085ac199238e925f4c976d68f6f98ac.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 298, + 538, + 310 + ], + "lines": [ + { + "bbox": [ + 451, + 298, + 538, + 310 + ], + "spans": [ + { + "bbox": [ + 451, + 298, + 538, + 310 + ], + "type": "text", + "content": "'an oil painting of a waffle, Van Gogh Style'." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 50, + 325, + 99, + 422 + ], + "blocks": [ + { + "bbox": [ + 50, + 325, + 99, + 422 + ], + "lines": [ + { + "bbox": [ + 50, + 325, + 99, + 422 + ], + "spans": [ + { + "bbox": [ + 50, + 325, + 99, + 422 + ], + "type": "image", + "image_path": "bd8f399f2be6462eb7f43715f2e0a73f0383566f77cfa134538a887cd5754a7b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 436, + 290, + 446 + ], + "lines": [ + { + "bbox": [ + 102, + 436, + 290, + 446 + ], + "spans": [ + { + "bbox": [ + 102, + 436, + 290, + 446 + ], + "type": "text", + "content": "(c) Composition between real & sketchy painting domains" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 451, + 546, + 487 + ], + "lines": [ + { + "bbox": [ + 46, + 451, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 546, + 487 + ], + "type": "text", + "content": "Figure 1: Image composition aims to seamlessly blend distinct objects into a specific visual context. Our training-free framework equips attention-based text-driven diffusion models with the capability to achieve this task across various domains (a) photorealism, (b) oil painting, (c) sketching, and (d) cartoon animation, within 20 sampling steps." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 100, + 325, + 197, + 420 + ], + "blocks": [ + { + "bbox": [ + 119, + 312, + 276, + 321 + ], + "lines": [ + { + "bbox": [ + 119, + 312, + 276, + 321 + ], + "spans": [ + { + "bbox": [ + 119, + 312, + 276, + 321 + ], + "type": "text", + "content": "(a) Composition within the photorealism domain" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 325, + 197, + 420 + ], + "lines": [ + { + "bbox": [ + 100, + 325, + 197, + 420 + ], + "spans": [ + { + "bbox": [ + 100, + 325, + 197, + 420 + ], + "type": "image", + "image_path": "3fea9fed12136a363be86ab142e4491af50fd65e07a9f8e735e70fe72f80ea0e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 421, + 193, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 193, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 193, + 434 + ], + "type": "text", + "content": "' a pencil drawing of a sheep in the sunset" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 198, + 325, + 295, + 420 + ], + "blocks": [ + { + "bbox": [ + 198, + 325, + 295, + 420 + ], + "lines": [ + { + "bbox": [ + 198, + 325, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 198, + 325, + 295, + 420 + ], + "type": "image", + "image_path": "6843d479466ac00bc61107fd8bd46e9fc315db2b10fa359d57b07e969644329e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 421, + 290, + 434 + ], + "lines": [ + { + "bbox": [ + 203, + 421, + 290, + 434 + ], + "spans": [ + { + "bbox": [ + 203, + 421, + 290, + 434 + ], + "type": "text", + "content": "' a pencil drawing of a panda in the sunset" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 298, + 325, + 346, + 421 + ], + "blocks": [ + { + "bbox": [ + 298, + 325, + 346, + 421 + ], + "lines": [ + { + "bbox": [ + 298, + 325, + 346, + 421 + ], + "spans": [ + { + "bbox": [ + 298, + 325, + 346, + 421 + ], + "type": "image", + "image_path": "8de6c7ddb49df72fe540feed7935bf14f448769a5ebcaf3cc5c9a27b9d7e9d96.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 353, + 325, + 444, + 421 + ], + "blocks": [ + { + "bbox": [ + 353, + 325, + 444, + 421 + ], + "lines": [ + { + "bbox": [ + 353, + 325, + 444, + 421 + ], + "spans": [ + { + "bbox": [ + 353, + 325, + 444, + 421 + ], + "type": "image", + "image_path": "523e81e026b52099b78832bca53b534810a59a469f9ba6d7e571b6554c898f1d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 422, + 441, + 434 + ], + "lines": [ + { + "bbox": [ + 354, + 422, + 441, + 434 + ], + "spans": [ + { + "bbox": [ + 354, + 422, + 441, + 434 + ], + "type": "text", + "content": "' a cartoon animation of buildings in the distance'" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 436, + 526, + 445 + ], + "lines": [ + { + "bbox": [ + 365, + 436, + 526, + 445 + ], + "spans": [ + { + "bbox": [ + 365, + 436, + 526, + 445 + ], + "type": "text", + "content": "(d) Composition between real & cartoon domains" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 446, + 325, + 543, + 421 + ], + "blocks": [ + { + "bbox": [ + 446, + 325, + 543, + 421 + ], + "lines": [ + { + "bbox": [ + 446, + 325, + 543, + 421 + ], + "spans": [ + { + "bbox": [ + 446, + 325, + 543, + 421 + ], + "type": "image", + "image_path": "54d85347f8cfb11d4f2f504c98b0e59732ceea6f5a217a40e3ed8e8ed9e067ef.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 422, + 538, + 434 + ], + "lines": [ + { + "bbox": [ + 451, + 422, + 538, + 434 + ], + "spans": [ + { + "bbox": [ + 451, + 422, + 538, + 434 + ], + "type": "text", + "content": "' a cartoon animation of an Eiffel tower in the distance'" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 143, + 495, + 192, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 495, + 192, + 507 + ], + "spans": [ + { + "bbox": [ + 143, + 495, + 192, + 507 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 45, + 510, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 510, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 510, + 290, + 715 + ], + "type": "text", + "content": "Text-driven diffusion models have exhibited impressive generative capabilities, enabling various image editing tasks. In this paper, we propose TF-ICON, a novel Training-Free Image CompositioN framework that harnesses the power of text-driven diffusion models for cross-domain image-guided composition. This task aims to seamlessly integrate user-provided objects into a specific visual context. Current diffusion-based methods often involve costly instance-based optimization or finetuning of pretrained models on customized datasets, which can potentially undermine their rich prior. In contrast, TF-ICON can leverage off-the-shelf diffusion models to perform cross-domain image-guided composition without requiring additional training, finetuning, or optimization. Moreover, we introduce the exceptional prompt, which contains no information, to facilitate text-driven diffusion models in accurately inverting real images into latent representations," + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 497, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 569 + ], + "type": "text", + "content": "forming the basis for compositing. Our experiments show that equipping Stable Diffusion with the exceptional prompt outperforms state-of-the-art inversion methods on various datasets (CelebA-HQ, COCO, and ImageNet), and that TF-ICON surpasses prior baselines in versatile visual domains. Code is available at https://github.com/Shilin-LU/TF-ICON" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 586, + 387, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 387, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 387, + 598 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Image composition task involves incorporating unique objects from different photos to create a harmonious image within a specific visual context, a.k.a. image-guided composition. For instance, consider the scenario where one desires to incorporate a beloved panda into one's favorite artwork, e.g., oil or sketchy painting. The objective is to create a new image where the panda blends seamlessly into the scene without altering the appearance of the panda and the background, just as an artist meticulously crafted this panda" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "2294" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "for that artwork (See Figure 1). This task is inherently challenging, as it requires maintaining illumination consistency and preserving identifying features. The challenge is further compounded when the photos come from various domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 288, + 326 + ], + "type": "text", + "content": "While recently large-scale text-to-image models [10, 19, 51, 56, 59, 61, 78] have achieved remarkable success in text-driven image generation, the ambiguity inherent in natural language presents challenges in conveying precise and nuanced visual details, even with highly detailed text prompts. Although this challenge is effectively addressed by enabling personalized concept learning [24, 25, 34, 37, 60], these methods require costly instance-based optimization and are limited in generating concepts with specified backgrounds. Recent studies [67, 77] have shown that diffusion models can achieve image-guided composition by explicitly incorporating additional guiding images. However, these models are retrained from the pretrained diffusion model on tailored datasets, which can damage the rich prior of the model. As a result, these models have limited compositional abilities beyond their training domain and still require significant computational resources." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 328, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 519 + ], + "type": "text", + "content": "Given the wealth of large text-to-image models that have been trained on extensive language-image datasets, we pose a question: how could these models be leveraged for image-guided composition without incurring costly training or finetuning, thereby avoiding damaging the diverse prior? To answer it, we propose the Training-Free Image CompositionN (TF-ICON) framework, which equips attention-based text-to-image diffusion models with the capability to perform image-guided composition without requiring additional training, fine-tuning, extra data, or optimization. To the best of our knowledge, this is the first training-free framework developed for image-guided composition. The framework is compatible with various diffusion model samplers, enabling completion within 20 steps, and harnesses rich semantic knowledge to facilitate image-guided compositions across diverse domains (see Figure 1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "Our approach constitutes an image-guided composition interface through denoising from a reliable starting latent code with the injection of composite self-attention maps. Finding the latent code that allows for reconstructing an input image while maintaining its editability, a.k.a. image inversion, is a challenging yet crucial step for state-of-the-art (SOTA) image editing frameworks involving real images [15, 26, 35, 38, 48, 53, 54, 70]. For diffusion models, while denoising diffusion implicit models (DDIM) inversion [65] has been effective for unconditional diffusion models, it falls short for text-driven diffusion models [26, 50, 70, 71]. To circumvent this, we introduce the exceptional prompt to accurately invert real images into latent codes upon pretrained text-to-image models to serve for further composition generation. The accurate latent codes are composed as the starting noise for the diffusion process. Through the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "gradual injection of composite self-attention maps that are specifically designed to reflect the relations between guiding images, we are able to infuse contextual information from the background into the incorporated objects, which results in harmonious image-guided compositions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 133, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 133, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 317, + 133, + 545, + 145 + ], + "type": "text", + "content": "To summarize, we make the following key contributions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 151, + 545, + 318 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 151, + 545, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 545, + 186 + ], + "type": "text", + "content": "1. We demonstrate the superior performance of high-order diffusion ODE solvers compared to commonly used DDIM inversion for real image inversion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 186, + 545, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 545, + 245 + ], + "type": "text", + "content": "2. We present an exceptional prompt that allows text-driven models to achieve accurate invertibility, laying a solid groundwork for subsequent editing. Experimental results show that it surpasses SOTA inversion methods on three vision datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 247, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 545, + 281 + ], + "type": "text", + "content": "3. We propose the first training-free framework that enables cross-domain image-guided composition for attention-based diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 283, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 545, + 318 + ], + "type": "text", + "content": "4. We demonstrate quantitatively and qualitatively that our framework outperforms prior baselines for image-guided composition." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 332, + 392, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 332, + 392, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 332, + 392, + 345 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 354, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 677 + ], + "type": "text", + "content": "Image composition. Image composition is widely applied to electronic commerce, entertainment, and data augmentation [20, 43] for downstream tasks. It can be broadly categorized into two types: text-guided [5, 6, 11, 22, 45] and image-guided [8, 23, 39, 67, 76, 77, 79]. The former involves composing multiple objects specified by only a text prompt without limiting the appearance of objects, as long as their semantics align with the prompt. Despite the great successes of text-conditioned models, they are often prone to semantic errors [22, 56], especially when the text prompt involves multiple objects. These errors include attribute leakage, attribute interchange, and missing objects, which cause the generated images to critically different from the user's intention [22, 56]. As a result, extensive prompt engineering [74] is often necessary to achieve the desired results. In contrast to text-only guided composition, image-guided composition involves incorporating specific objects and scenarios from user-provided photos, potentially with the aid of a text prompt. However, due to the nature of involving additional real images, it is more challenging particularly when images from different visual domains. Conventionally, image-guided composition is divided into several sub-tasks [52], such as object placement [7, 12, 40, 69, 80], image blending [75, 81], image harmonization [14, 16, 30, 76], and shadow generation [29, 42, 63, 83], each of which is typically addressed by different models and pipelines." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Image inversion. Extensive research has been conducted on image inversion for GANs, including latent-based op" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2295" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 65, + 293, + 131 + ], + "blocks": [ + { + "bbox": [ + 49, + 65, + 293, + 131 + ], + "lines": [ + { + "bbox": [ + 49, + 65, + 293, + 131 + ], + "spans": [ + { + "bbox": [ + 49, + 65, + 293, + 131 + ], + "type": "image", + "image_path": "43a6bd7a936544f4d728367a28a90fd28694c1e04433d528b991790bc4dbcfeb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 130, + 328, + 137 + ], + "lines": [ + { + "bbox": [ + 272, + 130, + 328, + 137 + ], + "spans": [ + { + "bbox": [ + 272, + 130, + 328, + 137 + ], + "type": "text", + "content": "(a) Image Inversion" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 294, + 66, + 543, + 132 + ], + "blocks": [ + { + "bbox": [ + 294, + 66, + 543, + 132 + ], + "lines": [ + { + "bbox": [ + 294, + 66, + 543, + 132 + ], + "spans": [ + { + "bbox": [ + 294, + 66, + 543, + 132 + ], + "type": "image", + "image_path": "7fc4b2d490e4202927ec7ac486839f44305a5ac31372b5187ea219137b06ab2f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 137, + 544, + 337 + ], + "blocks": [ + { + "bbox": [ + 51, + 137, + 544, + 337 + ], + "lines": [ + { + "bbox": [ + 51, + 137, + 544, + 337 + ], + "spans": [ + { + "bbox": [ + 51, + 137, + 544, + 337 + ], + "type": "image", + "image_path": "707be35a11950f6fa56ef34f8f7b924d43924b6f2a53365461653c62ad4f8c89.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 338, + 370, + 347 + ], + "lines": [ + { + "bbox": [ + 231, + 338, + 370, + 347 + ], + "spans": [ + { + "bbox": [ + 231, + 338, + 370, + 347 + ], + "type": "text", + "content": "(b) Training-Free Image Composition (TF-ICON)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 350, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 545, + 422 + ], + "type": "text", + "content": "Figure 2: The proposed training-free image composition framework. (a) The exceptional prompt is used to invert the main and reference images into noises " + }, + { + "bbox": [ + 46, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 350, + 545, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 350, + 545, + 422 + ], + "type": "text", + "content": ", which are then composed to form the starting point " + }, + { + "bbox": [ + 46, + 350, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^*" + }, + { + "bbox": [ + 46, + 350, + 545, + 422 + ], + "type": "text", + "content": " for the composition process. (b) Three constituents are composed for injecting into the composition process at early timesteps, including self-attention maps from the main and reference image reconstruction processes, as well as cross attention between the main and reference images. For better clarity and readability, the original main and reference images are shown in the pixel space instead of the VAE latent space, and the reference image is presented without resizing and zero-padding." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 432, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 576 + ], + "type": "text", + "content": "tization [1, 2, 33], encoders [3, 57, 68], and fine-tuning [4, 58]. For diffusion models, DDIM [65] is a widely used technique for inversion in image editing frameworks. However, in text-driven settings, DDIM leads to significant reconstruction distortion due to the instability resulting from classifier-free guidance (CFG) [18, 28]. Recently, null-text inversion [50] has been proposed to achieve accurate inversion by optimizing the unconditional prediction of the text-to-image model. It demonstrates promising results but requires instance-based optimization. Concurrently, EDICT [71] also achieves near-perfect inversion, albeit doubling the computation time of the diffusion process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 586, + 124, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 124, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 124, + 600 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Diffusion probabilistic models (DPM) [18, 27, 59, 64] are generative models in which an image is generated by progressively denoising from Gaussian noise. The forward diffusion process gradually perturbs data with infinite noise scales, which can be modeled as the solution of a stochastic differential equation (SDE) " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t\\}_{t=0}^T" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". Formally, given a data sample " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim p_0 = p_{\\mathrm{data}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", random noise is gradually injected, eventually resulting in a sample " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " which is typically distributed as a tractable prior " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "p_T" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " without any infer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 432, + 539, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 539, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 539, + 445 + ], + "type": "text", + "content": "mation of " + }, + { + "bbox": [ + 306, + 432, + 539, + 445 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 306, + 432, + 539, + 445 + ], + "type": "text", + "content": " , as described by the following SDE [36, 66]:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 457, + 545, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 457, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 365, + 457, + 545, + 470 + ], + "type": "interline_equation", + "content": "\\mathrm {d} \\mathbf {x} _ {t} = \\mathbf {f} \\left(\\mathbf {x} _ {t}, t\\right) \\mathrm {d} t + g (t) \\mathrm {d} \\mathbf {w} _ {t}, \\tag {1}", + "image_path": "70a3933e6f78673184a1f0117fdc6139e7048e8c632ce8a8cfdc04e1ae820277.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_t\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "type": "text", + "content": " is the standard Wiener process (a.k.a., Brownian motion), and " + }, + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(\\cdot ,t)" + }, + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "type": "inline_equation", + "content": "g(t)" + }, + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "type": "text", + "content": " are commonly designated as the drift and diffusion coefficient, respectively. On the other hand, the reverse diffusion process can be described by the reverse-time SDE from " + }, + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 476, + 545, + 536 + ], + "type": "text", + "content": " to 0 [66]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 546, + 545, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 546, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 312, + 546, + 545, + 560 + ], + "type": "interline_equation", + "content": "\\mathrm {d} \\mathbf {x} _ {t} = [ \\mathbf {f} (\\mathbf {x} _ {t}, t) - g (t) ^ {2} \\nabla_ {\\mathbf {x}} \\log p _ {t} (\\mathbf {x} _ {t}) ] \\mathrm {d} t + g (t) \\mathrm {d} \\bar {\\mathbf {w}} _ {t}, \\tag {2}", + "image_path": "5da6161847ab72d3a8c758c6353dd5cb9522bd02838f3955bae36e80a909f04a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{w}}_t" + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "content": " is the Wiener process in the reverse time. The score function " + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}}\\log p_t(\\mathbf{x}_t)" + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "content": " is the only unknown term and can be estimated by a neural network " + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "content": " whose parameter " + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "content": " is optimized by a denoising objective [27, 66]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Upon attaining the trained model that predicts the score function accurately, it can be utilized to numerically solve the reverse SDE (Eq. (2)), enabling the generation of samples from a noise distribution. Song et al. [66] outline various methods, including Variance Exploding (VE), Variance Preserving (VP), and sub-VP SDE, for constructing SDEs that perturb the unknown data distribution into a fixed prior. In this work, we leverage the pre-trained text-to-image La" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2296" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "tent Diffusion Model (LDM) [59], a.k.a. Stable Diffusion, which applies the VP SDE in the latent space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 106, + 102, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 102, + 118 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 102, + 118 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": "Our objective is to utilize a main (background) image " + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": ", a reference (foreground) image " + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": ", a text prompt " + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": ", and a binary mask " + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{\\mathrm{user}}" + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": " which designates the region of interest within the main image, to generate a modified image " + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{*}" + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": ". The resultant image " + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{*}" + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": " should contain the reference subject with identifying features within the mask, i.e. " + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "inline_equation", + "content": "id(\\mathbf{I}^{*} \\odot \\mathbf{M}^{\\mathrm{user}}) \\approx id(\\mathbf{I}^{\\mathrm{r}})" + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": ", while concurrently ensuring that the complementing area closely resembles the main image, i.e. " + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{*} \\odot (\\mathbf{1} - \\mathbf{M}^{\\mathrm{user}}) \\approx \\mathbf{I}^{\\mathrm{m}} \\odot (\\mathbf{1} - \\mathbf{M}^{\\mathrm{user}})" + }, + { + "bbox": [ + 46, + 126, + 287, + 258 + ], + "type": "text", + "content": ". Moreover, it is ideal for the transition between the areas inside and outside the mask to be imperceptible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 288, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 288, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 288, + 366 + ], + "type": "text", + "content": "We propose a training-free framework that can make use of attention-based pre-trained text-to-image models to perform image-guided composition. To the best of our knowledge, it is the first training-free framework for image-guided composition, which can be accomplished within 20 steps of sampling. The framework is mainly comprised of two steps: image inversion (Section 4.1), and composition generation (Section 4.2), as shown in Figure 2. The full algorithm is presented in Appendix B.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 372, + 266, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 266, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 266, + 385 + ], + "type": "text", + "content": "4.1. Image Inversion with Exceptional Prompt" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 391, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 287, + 486 + ], + "type": "text", + "content": "Achieving precise manipulation of real images often necessitates an accurate inversion process that identifies the corresponding latent representation, which not only provides editability for meaningful manipulation but also accurately reconstructs the input image [26, 68]. For diffusion models, optimal editability is typically characterized by a noise encoding that conforms to the ideal statistical properties of zero-mean, unit-variance Gaussian noise [53]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 582 + ], + "type": "text", + "content": "ODE inversion. Most diffusion frameworks for image editing [15, 26, 35, 38, 53, 70] use DDIM inversion to invert the real image into its latent representation. However, our findings suggest that this may not be the optimal choice for inverting real images. It has been proven that DDIM is a first-order discretization of the associated probability flow ordinary differential equations (ODE) of Eq. (2) [62, 65, 66]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 589, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 589, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 77, + 589, + 287, + 611 + ], + "type": "interline_equation", + "content": "\\mathrm {d} \\mathbf {x} _ {t} = \\left[ \\mathbf {f} \\left(\\mathbf {x} _ {t}, t\\right) - \\frac {1}{2} g (t) ^ {2} \\nabla_ {\\mathbf {x}} \\log p _ {t} \\left(\\mathbf {x} _ {t}\\right) \\right] \\mathrm {d} t, \\tag {3}", + "image_path": "4d746c10c3fb8b41f304f9bf2126349b5da293eeaccba4c60371a07886e44e7e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "which can be solved using " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and shares the consistent marginal probability distribution " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{p_t(\\mathbf{x}_t)\\}_{t = 0}^T" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " with Eq. (2). Various samplers [32, 44, 46, 47] have been developed for solving the diffusion ODE starting from noise " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " to achieve fast sampling (10~20 steps). We offer the insight that utilizing these ODE solvers in turn as encoders starting from the real image " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " yields better latent representation " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ", compared with those obtained through commonly used" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 72, + 354, + 210 + ], + "blocks": [ + { + "bbox": [ + 308, + 72, + 354, + 210 + ], + "lines": [ + { + "bbox": [ + 308, + 72, + 354, + 210 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 354, + 210 + ], + "type": "image", + "image_path": "3f46d3803dea2136fd0ad6e853be951448157d879c981d8fd87998538df68106.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 212, + 349, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 212, + 349, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 349, + 220 + ], + "type": "text", + "content": "(a) Original" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 356, + 72, + 402, + 210 + ], + "blocks": [ + { + "bbox": [ + 356, + 72, + 402, + 210 + ], + "lines": [ + { + "bbox": [ + 356, + 72, + 402, + 210 + ], + "spans": [ + { + "bbox": [ + 356, + 72, + 402, + 210 + ], + "type": "image", + "image_path": "58d7ffc51efc7a9fb2567a5e35f8c464bafd691b2a196c15014feef48344d7b3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 212, + 391, + 220 + ], + "lines": [ + { + "bbox": [ + 366, + 212, + 391, + 220 + ], + "spans": [ + { + "bbox": [ + 366, + 212, + 391, + 220 + ], + "type": "text", + "content": "(b) CFG" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 449, + 210 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 449, + 210 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 449, + 210 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 449, + 210 + ], + "type": "image", + "image_path": "c701696a7df507f13a8f500591db1780aac298926a05d5e765784902f68b9004.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 212, + 440, + 220 + ], + "lines": [ + { + "bbox": [ + 411, + 212, + 440, + 220 + ], + "spans": [ + { + "bbox": [ + 411, + 212, + 440, + 220 + ], + "type": "text", + "content": "(c) Cond." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 449, + 72, + 497, + 210 + ], + "blocks": [ + { + "bbox": [ + 449, + 72, + 497, + 210 + ], + "lines": [ + { + "bbox": [ + 449, + 72, + 497, + 210 + ], + "spans": [ + { + "bbox": [ + 449, + 72, + 497, + 210 + ], + "type": "image", + "image_path": "ea450ef038a402b3765028bf1f78f33eb3abcdebfb9906bc3ab31d749af88288.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 212, + 492, + 220 + ], + "lines": [ + { + "bbox": [ + 455, + 212, + 492, + 220 + ], + "spans": [ + { + "bbox": [ + 455, + 212, + 492, + 220 + ], + "type": "text", + "content": "(d) Uncond." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 498, + 72, + 544, + 210 + ], + "blocks": [ + { + "bbox": [ + 498, + 72, + 544, + 210 + ], + "lines": [ + { + "bbox": [ + 498, + 72, + 544, + 210 + ], + "spans": [ + { + "bbox": [ + 498, + 72, + 544, + 210 + ], + "type": "image", + "image_path": "7f8820de931ef5f6ccc769ac3ab0caee3950b8dec11c3db166650896a4bb730e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 212, + 533, + 220 + ], + "lines": [ + { + "bbox": [ + 509, + 212, + 533, + 220 + ], + "spans": [ + { + "bbox": [ + 509, + 212, + 533, + 220 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 223, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 305, + 223, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 545, + 294 + ], + "type": "text", + "content": "Figure 3: The real image reconstruction results using Stable Diffusion with (b) classifier-free guidance (CFG) output " + }, + { + "bbox": [ + 305, + 223, + 545, + 294 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E},\\varnothing)" + }, + { + "bbox": [ + 305, + 223, + 545, + 294 + ], + "type": "text", + "content": "; (c) conditional output " + }, + { + "bbox": [ + 305, + 223, + 545, + 294 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E})" + }, + { + "bbox": [ + 305, + 223, + 545, + 294 + ], + "type": "text", + "content": "; (d) unconditional output " + }, + { + "bbox": [ + 305, + 223, + 545, + 294 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\varnothing)" + }, + { + "bbox": [ + 305, + 223, + 545, + 294 + ], + "type": "text", + "content": "; and (e) ours. The prompts for (b) and (c) are 'a photo of a singer', 'a photo of a car', and 'an oil painting'. See Appendix A.3 for elaboration." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 302, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 363 + ], + "type": "text", + "content": "DDIM. A quantitative analysis is given in Appendix A.1. The enhanced alignment between the forward and backward ODE trajectories in the high-order DPM-Solver++ [47] implies that it is better suited for real image inversion. Thus, this paper employs it for all inversions of diffusion models." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "text", + "content": "Exceptional prompt. In the unconditional setting " + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "text", + "content": " solving the diffusion ODE (Eq. (3)) from 0 to " + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "text", + "content": " enables us to obtain the better latent code " + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "text", + "content": " for the real image " + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "text", + "content": ". However, in the text-driven setting " + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E})" + }, + { + "bbox": [ + 305, + 374, + 545, + 459 + ], + "type": "text", + "content": ", existing image editing works [26, 50, 70, 71] have shown that the inversion process is prone to significant reconstruction errors, due to the instability induced by CFG [18, 28]:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 466, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 466, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 310, + 466, + 545, + 479 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\mathcal {E}, \\varnothing\\right) = s \\cdot \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\mathcal {E}\\right) + (1 - s) \\cdot \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t}, t, \\varnothing\\right), \\tag {4}", + "image_path": "028da69bc8e3afc2cbdf425b4bd9a5f3ac94a5eef4b4be3d3879418f8c74ce8e.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\varnothing = \\psi(\"\")" + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\psi(\\mathcal{P})" + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "text", + "content": " are embeddings of the null and normal prompt, and " + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "text", + "content": " is the guidance scale. Our experiments further reveal that even without CFG, both conditional output " + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{E})" + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "text", + "content": " and unconditional output " + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\emptyset)" + }, + { + "bbox": [ + 304, + 486, + 545, + 557 + ], + "type": "text", + "content": " of text-to-image diffusion models still produce large reconstruction errors, as depicted in Figure 3." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "To achieve accurate inversion, we present a straightforward yet effective solution, namely exceptional prompt, " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\text{exceptional}}" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": ". Intuitively, any information contained within the input prompt can result in the deviation of the backward ODE trajectories from the forward trajectories. Hence, we remove all information by setting all token numbers to a common value and eliminating positional embeddings for the text prompt, as depicted in Figure 4. Importantly, the exceptional prompt is distinguished from the null prompt by its absence of special tokens, such as [startoftext], [endoftext], and [pad], which still retain information. The exceptional prompt is applied only in image inversion but not in the composition process. The choice of the" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2297" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 64, + 289, + 167 + ], + "blocks": [ + { + "bbox": [ + 48, + 64, + 289, + 167 + ], + "lines": [ + { + "bbox": [ + 48, + 64, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 48, + 64, + 289, + 167 + ], + "type": "image", + "image_path": "da46054a4a9d364b4cbf71f21b761c34a32c05fe9ad0cb406a58c716affc4754.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 168, + 287, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 194 + ], + "type": "text", + "content": "Figure 4: The illustration of comparison among (a) Normal Prompt, (b) Null Prompt, and (c) Exceptional Prompt." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 287, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 237 + ], + "type": "text", + "content": "token value does not significantly affect the inversion. The detailed analysis of the exceptional prompt and token value selection is provided in Appendix A.2 and A.4, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 237, + 287, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 287, + 321 + ], + "type": "text", + "content": "Figure 3 visually demonstrates the effectiveness of the exceptional prompt. Our results " + }, + { + "bbox": [ + 46, + 237, + 287, + 321 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathcal{W})" + }, + { + "bbox": [ + 46, + 237, + 287, + 321 + ], + "type": "text", + "content": " with the exceptional prompt embedding " + }, + { + "bbox": [ + 46, + 237, + 287, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{W} = \\psi (\\mathcal{P}_{\\mathrm{exceptional}})" + }, + { + "bbox": [ + 46, + 237, + 287, + 321 + ], + "type": "text", + "content": " are more visually accurate than others. The quantitative experiments are shown in Section 5.1. All the results in Figure 3 are obtained by solving the forward and backward diffusion ODEs using the second-order DPM-Solver++ [47] in 20 steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 327, + 230, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 230, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 230, + 340 + ], + "type": "text", + "content": "4.2. Training-Free Image Composition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 346, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 395 + ], + "type": "text", + "content": "Upon equipping the accurate invertibility, image composition can be performed based on it. The composition process consists of two key components: noise incorporation and composite self-attention maps injection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 406, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 287, + 513 + ], + "type": "text", + "content": "Noise incorporation. Before inverting images into noises, a simple preprocessing step is necessary for the reference image. Typically, only the foreground in the reference is desired for composition, so the preprocessing step involves using a pretrained segmentation model [84] to remove the background, resizing and repositioning the object to match the user's mask in the main image, and padding it with zeros to ensure it is the same size as the main image (See Appendix B.1 for visual illustration)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": "Once the preprocessing is complete, the main and reference images are inverted to corresponding noises " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " by solving diffusion ODEs (Eq. (3)) from 0 to " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " with the exceptional prompt " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{exceptional}}" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " are then merged with standard Gaussian noise " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " to create the starting point " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^*" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " for generating the composition. Formally, the incorporated noise " + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^*" + }, + { + "bbox": [ + 46, + 514, + 287, + 598 + ], + "type": "text", + "content": " is calculated by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 630 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {T} ^ {*} = \\mathbf {x} _ {T} ^ {\\mathrm {r}} \\odot \\mathbf {M} ^ {\\text {s e g}} + \\mathbf {x} _ {T} ^ {\\mathrm {m}} \\odot (\\mathbf {1} - \\mathbf {M} ^ {\\text {u s e r}}) + \\mathbf {z} \\odot (\\mathbf {M} ^ {\\text {u s e r}} \\oplus \\mathbf {M} ^ {\\text {s e g}}), \\tag {5}", + "image_path": "0b2a768a3d487a0705087de4291736edaddd9cc4115a2868c97c3046065adbb2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{\\mathrm{user}}" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " is the user mask, " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{\\mathrm{seg}}" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " is the segmentation mask for reference image, and " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{\\mathrm{user}} \\oplus \\mathbf{M}^{\\mathrm{seg}}" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " is the XOR of them, which is the transition area. The incorporation of " + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": " enhances the smoothness of the transition between the regions inside and outside the user mask, by effectively leveraging the prior knowledge of the text-driven diffusion model to inpaint the transition area. Empirically," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 342, + 57, + 511, + 171 + ], + "blocks": [ + { + "bbox": [ + 342, + 57, + 511, + 171 + ], + "lines": [ + { + "bbox": [ + 342, + 57, + 511, + 171 + ], + "spans": [ + { + "bbox": [ + 342, + 57, + 511, + 171 + ], + "type": "image", + "image_path": "a62e7ac9c1ef73f3cab714b7f283d8bd8c74a872531768822e4ccc1d95442fbd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 172, + 529, + 185 + ], + "lines": [ + { + "bbox": [ + 322, + 172, + 529, + 185 + ], + "spans": [ + { + "bbox": [ + 322, + 172, + 529, + 185 + ], + "type": "text", + "content": "Figure 5: A toy example for attention composition." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 193, + 545, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 241 + ], + "type": "text", + "content": "for cross-domain composition, incorporating the starting point in the noise space usually is more effective, while solely for photorealism, composing in the pixel/latent space and then inverting it as the starting point is more favorable." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "type": "text", + "content": "Composite self-attention map injection. The incorporated noise " + }, + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^*" + }, + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "type": "text", + "content": " is employed as the starting point for solving the diffusion ODE from " + }, + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "type": "text", + "content": " to 0 with a normal prompt " + }, + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "type": "text", + "content": " to ultimately generate the composition. " + }, + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 255, + 546, + 398 + ], + "type": "text", + "content": " is intended to assist in inpainting transition areas. However, relying solely on noise incorporation, the pretrained text-to-image model cannot preserve the appearance of the main and reference images effectively, as shown in Figure 7. To tackle this problem, we propose injecting composite self-attention maps in a specially designed manner, as the semantic information is basically retained within the rows and columns of self-attention maps (See Appendix E for visual illustrations)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "text", + "content": "The composite self-attention map comprises three constituents: two self-attention maps, " + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{m}}" + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{r}}" + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "text", + "content": ", corresponding to the main and reference images, and a cross-attention map, " + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{cross}}" + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "text", + "content": ", calculated between them. The composition way is illustrated in Figure 5. To compose the reference image in the blue regions of Figure 5 (a), its self-attention map " + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{r}}" + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "text", + "content": " should be placed in the corresponding blue regions of Figure 5 (b), since the 5th patch can only attend to the patches 6, 9, and 10 in the context of self-attention. The green regions in Figure 5 (b) should contain the cross-attention map, " + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{cross}}" + }, + { + "bbox": [ + 304, + 400, + 546, + 640 + ], + "type": "text", + "content": ", which infuses contextual information from the white regions into the blue regions. If the green regions are preserved as the self-attention of the white regions without replacement, the information stored there only reflects the relation between the original patches, such as the 0th and original 5th patches in the example of index (5,0) or (0,5). This results in a lack of surrounding information being provided to the new 5th patch, such as painting or sketching, which is necessary for seamless object transition to other domains (See ablation in Figure 7)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Essential constituents " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{m}},\\mathbf{A}_{l,t}^{\\mathrm{r}},\\mathbf{A}_{l,t}^{\\mathrm{cross}}" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " are calculated using self-attention modules of the pretrained Stable Diffusion. Typically, a self-attention module at layer " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " contains three projection matrices " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_l^q,\\mathbf{W}_l^k,\\mathbf{W}_l^v" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " in the same dimension " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": ". Denote the spatial features of the main and reference image at timestep " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " and layer " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{l,t}^{\\mathrm{m}}\\in \\mathbb{R}^{(h\\times w)\\times d}" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2298" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 287, + 120 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 71, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{l,t}^{\\mathrm{r}}\\in \\mathbb{R}^{(h'\\times w')\\times d}" + }, + { + "bbox": [ + 46, + 71, + 287, + 120 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 46, + 71, + 287, + 120 + ], + "type": "inline_equation", + "content": "h^\\prime \\times w^\\prime" + }, + { + "bbox": [ + 46, + 71, + 287, + 120 + ], + "type": "text", + "content": " is the size of the reference image after resizing to match the size of the user mask. The queries, keys, and values for each self-attention module are obtained as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 125, + 287, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 125, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 125, + 287, + 140 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {l, t} ^ {\\mathrm {m}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {m}} \\mathbf {W} _ {l} ^ {q}, \\quad \\mathbf {k} _ {l, t} ^ {\\mathrm {m}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {m}} \\mathbf {W} _ {l} ^ {k}, \\quad \\mathbf {v} _ {l, t} ^ {\\mathrm {m}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {m}} \\mathbf {W} _ {l} ^ {v}, \\tag {6}", + "image_path": "15dcd8dd462e6aa2268e66cd2678374781371bf1a03fa472ff13c6b7771bb96c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 143, + 287, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 143, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 70, + 143, + 287, + 158 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {l, t} ^ {\\mathrm {r}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {r}} \\mathbf {W} _ {l} ^ {q}, \\quad \\mathbf {k} _ {l, t} ^ {\\mathrm {r}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {r}} \\mathbf {W} _ {l} ^ {k}, \\quad \\mathbf {v} _ {l, t} ^ {\\mathrm {r}} = \\mathbf {f} _ {l, t} ^ {\\mathrm {r}} \\mathbf {W} _ {l} ^ {v}, \\tag {7}", + "image_path": "7bd2735f980ee833ed5299626159060712ada5bfafc20c7aa8b32854f7f53ebc.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{l,t}^{\\mathrm{m}},\\mathbf{k}_{l,t}^{\\mathrm{m}},\\mathbf{v}_{l,t}^{\\mathrm{m}}\\in \\mathbb{R}^{(h\\times w)\\times d}" + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{l,t}^{\\mathrm{r}},\\mathbf{k}_{l,t}^{\\mathrm{r}},\\mathbf{v}_{l,t}^{\\mathrm{r}}\\in" + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{(h'\\times w')\\times d}" + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{m}},\\mathbf{A}_{l,t}^{\\mathrm{r}}" + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{cross}}" + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "text", + "content": " are then calculated and composed as " + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{*}" + }, + { + "bbox": [ + 47, + 163, + 287, + 204 + ], + "type": "text", + "content": " for injection:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 210, + 287, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 210, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 87, + 210, + 287, + 232 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {l, t} ^ {\\mathrm {m}} = \\operatorname {S o f t m a x} \\left(\\mathbf {q} _ {l, t} ^ {\\mathrm {m}} \\cdot \\left(\\mathbf {k} _ {l, t} ^ {\\mathrm {m}}\\right) ^ {\\mathrm {T}} / \\sqrt {d}\\right), \\tag {8}", + "image_path": "5bf21c8b98d570aa6b9b1b2afccb843df1d88b1c750484922ad60d62aa558bb6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 233, + 287, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 233, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 87, + 233, + 287, + 253 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {l, t} ^ {\\mathrm {r}} = \\operatorname {S o f t m a x} \\left(\\mathbf {q} _ {l, t} ^ {\\mathrm {r}} \\cdot \\left(\\mathbf {k} _ {l, t} ^ {\\mathrm {r}}\\right) ^ {\\mathrm {T}} / \\sqrt {d}\\right), \\tag {9}", + "image_path": "2d979702503c11fd0f25762d3ca508170e8a7d7f5b2f41712d3f9bf621ea4c71.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 255, + 287, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 255, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 87, + 255, + 287, + 275 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {l, t} ^ {\\text {c r o s s}} = \\operatorname {S o f t m a x} \\left(\\mathbf {q} _ {l, t} ^ {\\mathrm {m}} \\cdot \\left(\\mathbf {k} _ {l, t} ^ {\\mathrm {r}}\\right) ^ {\\top} / \\sqrt {d}\\right), \\tag {10}", + "image_path": "e41ca881c28c48cf4046aae6efd4b512f2037669b9349f47d62ad670cb459bf7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 277, + 287, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 277, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 87, + 277, + 287, + 291 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {l, t} ^ {*} = \\vartheta_ {\\text {c o m p o s e}} \\left(\\mathbf {A} _ {l, t} ^ {\\mathrm {m}}, \\mathbf {A} _ {l, t} ^ {\\mathrm {r}}, \\mathbf {A} _ {l, t} ^ {\\text {c r o s s}}\\right), \\tag {11}", + "image_path": "3db9ba8a5214564c288e5e0432fe66df726ccc2df0a9083b24a02afe219a0a2f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{m}}\\in \\mathbb{R}^{(h\\times w)\\times (h\\times w)}" + }, + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{r}}\\in \\mathbb{R}^{(h^{\\prime}\\times w^{\\prime})\\times (h^{\\prime}\\times w^{\\prime})}" + }, + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{cross}}\\in \\mathbb{R}^{(h\\times w)\\times (h^{\\prime}\\times w^{\\prime})}" + }, + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\vartheta_{\\mathrm{compose}}" + }, + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "type": "text", + "content": " is the function to bulid composite self-attention maps " + }, + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{*}" + }, + { + "bbox": [ + 47, + 296, + 287, + 348 + ], + "type": "text", + "content": " based on patch indices (Figure 5)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "text", + "content": "As a result, three diffusion ODEs are solved simultaneously from " + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "text", + "content": " to 0. As depicted in Figure 2 (b), ODEs start from the accurate inverted noises " + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{\\mathrm{r}},\\mathbf{x}_T^{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "text", + "content": " and the interpolated noise " + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^*" + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "text", + "content": " , respectively. The first two ODEs are solved using the exceptional prompt " + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{exceptional}}" + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "text", + "content": " to progressively reconstruct the main and reference, thus allowing for the precise retention of " + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{cross}}" + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "text", + "content": " These attention maps are then composed and injected into the third ODE for generating a natural and cohesive composition with a normal prompt " + }, + { + "bbox": [ + 46, + 349, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "text", + "content": "To balance the generation of high-level context and finer details [13, 38], we set a threshold " + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\tau_{A}" + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "text", + "content": " to determine the time steps for injecting composite self-attention maps in the early stage " + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "inline_equation", + "content": "(t\\in [T\\times \\tau_A,T])" + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "text", + "content": " and allow the model to explore ODE trajectories through a normal prompt " + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "text", + "content": " in the later stage " + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "inline_equation", + "content": "(t\\in [0,T\\times \\tau_A])" + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "text", + "content": ", guided by the prior of the pretrained model. However, this freedom, without the imposition of attention injection constraints, often results in deviations from the desired background (see Figure 7). Thus, similar to [6], we set an additional threshold, denoted as " + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\tau_{B}" + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "text", + "content": ", which regulates the trajectory rectification process. This process entails replacing the regions outside the user mask with the reconstructed main image at various time steps, i.e., " + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_t^* = \\mathbf{x}_t^*\\odot \\mathbf{M}^{\\mathrm{user}} + \\mathbf{x}_t^{\\mathrm{m}}\\odot (\\mathbf{1} - \\mathbf{M}^{\\mathrm{user}})" + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "inline_equation", + "content": "t\\in [T\\times \\tau_B,T]" + }, + { + "bbox": [ + 46, + 468, + 287, + 659 + ], + "type": "text", + "content": ". Note that only preserving the background at the final step can lead to noticeable artifacts, as shown in Appendix B.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 669, + 127, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 127, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 127, + 682 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "This section consists of two sets of experiments. The first set assesses the effectiveness of the exceptional prompt" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 86, + 545, + 247 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 538, + 83 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 538, + 83 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 538, + 83 + ], + "type": "text", + "content": "Table 1: The reconstruction comparison on CelebA-HQ." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 86, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 309, + 86, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 309, + 86, + 545, + 247 + ], + "type": "table", + "html": "
MethodMAE ↓LPIPS ↓SSIM ↑
OptimizationI2S [1]0.0640.1340.872
PTI [58]0.0620.1320.877
EncoderpSp [57]0.0790.1690.793
e4e [68]0.0920.2210.742
ReStyle w/ pSp [3]0.0730.1450.823
ReStyle w/ e4e [3]0.0890.2020.758
HFGI w/ e4e [73]0.0620.1270.877
DiffusionSD w/ CFG0.1340.3400.637
SD w/ Cond.0.1260.3080.654
SD w/ Uncond.0.1260.3040.655
DiffusionCLIP [35]0.0200.0730.914
Ours0.0190.0470.918
Upper BoundVQAE [21]0.0180.0430.919
", + "image_path": "fe679bdc52de9960f2532ca8a4b5be0e48fdc99893c42458bc76f3ffd1368424.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 308, + 284, + 547, + 360 + ], + "blocks": [ + { + "bbox": [ + 306, + 257, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 306, + 257, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 257, + 545, + 281 + ], + "type": "text", + "content": "Table 2: The further reconstruction comparison on COCO and ImageNet. \\*: an upper bound." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 284, + 547, + 360 + ], + "lines": [ + { + "bbox": [ + 308, + 284, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 547, + 360 + ], + "type": "table", + "html": "
MethodMSCOCO (5000)ImageNet (3000)
MAE ↓LPIPS ↓SSIM ↑MAE ↓LPIPS ↓SSIM ↑
SD w/ CFG0.1500.4580.5680.1320.4960.575
SD w/ Cond.0.1220.3590.6330.1090.3890.645
SD w/ Uncond.0.1200.3630.6360.1140.4060.635
Ours0.0300.0730.8680.0330.0870.852
VQAE* [21]0.0300.0690.8700.0320.0840.854
", + "image_path": "cb366864f96a69e791e9b928388b56eb679cfa9b962c104a86d5735697c7faa0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "type": "text", + "content": "(Section 5.1). The second set evaluates our image composition framework qualitatively and quantitatively (Section 5.2), followed by an ablation study (Section 5.3)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 415, + 432, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 432, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 432, + 428 + ], + "type": "text", + "content": "5.1. Image Reconstruction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 434, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 614 + ], + "type": "text", + "content": "To assess the effectiveness of the exceptional prompt, we compared its performance with SOTA GAN [1, 3, 57, 58, 68, 73] and diffusion [35] inversion methods on the CelebA-HQ [31], following the same setting as described in [35, 73]. Additionally, we conducted experiments on the ImageNet [17] and COCO [41] with Stable Diffusion to further validate our findings. Our results (Tables 1 and 2) show that the exceptional prompt is highly effective in producing reconstructions that closely approximate the upper bound established by the vector quantized autoencoder (VQAE) [21] across all metrics, including MAE, LPIPS [82], and SSIM. The qualitative comparison is shown in Figure 3. All results of Stable Diffusion, including ours, are sampled by the second-order DPM-Solver++ [47]. Experimental settings are detailed in Appendix B.4." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 623, + 485, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 485, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 485, + 635 + ], + "type": "text", + "content": "5.2. Image Composition Comparisons" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Test benchmark. As there is currently no benchmark for testing cross-domain image-guided composition as a whole, we developed a test benchmark containing 332 samples. Each sample in the benchmark consists of a main (background) image, a reference (foreground) image, a user mask, and a text prompt. The main images comprise" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2299" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 544, + 340 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 544, + 340 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 544, + 340 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 544, + 340 + ], + "type": "image", + "image_path": "7ee06532bde463cc64858ae73c987269354cc5778dfcf46f8e148bd4dc1e0cfe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 341, + 399, + 349 + ], + "lines": [ + { + "bbox": [ + 194, + 341, + 399, + 349 + ], + "spans": [ + { + "bbox": [ + 194, + 341, + 399, + 349 + ], + "type": "text", + "content": "'acartoon animation of a cake,acroissant,a piece of bread anda cup of coffee" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 351, + 546, + 375 + ], + "lines": [ + { + "bbox": [ + 46, + 351, + 546, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 546, + 375 + ], + "type": "text", + "content": "Figure 6: Qualitative comparison with SOTA and concurrent baselines in image-guided composition for sketching, photorealism, painting, and cartoon animation domains. Additional results are available in Appendix H.2." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 411, + 288, + 482 + ], + "blocks": [ + { + "bbox": [ + 47, + 384, + 287, + 407 + ], + "lines": [ + { + "bbox": [ + 47, + 384, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 287, + 407 + ], + "type": "text", + "content": "Table 3: Quantitative evaluation results for image composition in the photorealism domain." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 411, + 288, + 482 + ], + "lines": [ + { + "bbox": [ + 50, + 411, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 50, + 411, + 288, + 482 + ], + "type": "table", + "html": "
MethodLPIPS(BG)↓LPIPS(FG)↓CLIP(Image)↑CLIP(Text)↑
SDEdit (0.4) [48]0.350.6280.5627.73
SDEdit (0.6) [48]0.420.6677.6827.98
Blended [5]0.110.7773.2525.19
Paint [77]0.130.7380.2625.92
DIB [81]0.110.6377.5726.84
Ours0.100.6082.8628.11
", + "image_path": "4617d01093fff21c23e7d6f8a2a24735c621ff8da3f87159ccb217464ea019d9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 496, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 557 + ], + "type": "text", + "content": "four visual domains: photorealism, pencil sketching, oil painting, and cartoon animation. All reference images are from the photorealism domain as the reference requires segmentation models, which are generally more effective in this domain. Further details are available in Appendix G." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Qualitative comparisons. Our qualitative comparisons are performed across four visual domains, employing SOTA and concurrent baselines that are applicable to image-guided composition, including Deep Image Blending (DIB) [81], DCCF [76], Blended Diffusion [5], Textual Inversion [24], Paint by Example [77], and SDEdit [48] under two different noising levels. As shown in Figure 6, our framework is capable of seamlessly composing objects into various domains while maintaining their identities. In contrast, DIB and DCCF fall short in processing the transition areas, leading to noticeable artifacts. Blended Diffusion's foreground generation and Textual Inversion's" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 481 + ], + "type": "text", + "content": "background generation relies solely on text prompts, causing deviations from the user's intention. While Paint by Example effectively composes images within its photorealistic training domain, it struggles to adapt to other domains. Additionally, SDEdit with fewer timesteps is suitable for image composition in terms of preserving the identifying features of the reference, but the background is changed. See Appendix H.2 for additional comparisons." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": "Quantitative analysis. The baselines are primarily trained in the photorealism domain, where the objective metrics are more effective; therefore, we focused our quantitative comparison within this domain and relied on user study for comparison in other domains. We assess the same baselines as in the qualitative comparison, with the exception of Textual Inversion [24], which involves instance-based optimization, and DCCF [76], which is used for harmonizing images after copy-and-paste operations. Four metrics are considered: (1) " + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS}_{\\mathrm{(BG)}}" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": " [82] measures the background consistency, (2) " + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{LPIPS}_{\\mathrm{(FG)}}" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": " [82] evaluates the low-level similarity between the edited region and the reference foreground, (3) " + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{(Image)}}" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": " [55] evaluates the semantic similarity between the edited region and the reference in the CLIP embedding space, and (4) " + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{(Text)}}" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": " [55] measures the semantic alignment between the text prompt and the resultant image. As presented in Table 3, our method outperforms all baselines. We achieve well preservation" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2300" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 99, + 288, + 167 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 95 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 95 + ], + "type": "text", + "content": "Table 4: User study: higher score, better ranking. P: photorealism; O: oil painting; S: sketchy painting; C: cartoon." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 99, + 288, + 167 + ], + "lines": [ + { + "bbox": [ + 50, + 99, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 50, + 99, + 288, + 167 + ], + "type": "table", + "html": "
MethodP & PP & OP & SP & CTotal
Blended [5]1.8072.3142.6802.1002.093
SDEdit (0.6) [48]2.3683.0632.4092.7132.485
Paint [77]2.8792.3062.0432.6732.666
DCCF [76]3.8383.2373.2973.4703.583
Ours4.1084.0804.5714.0434.175
", + "image_path": "be758e4f32deb3c30f58edcc2928a4cf0625d3806421da9c95fcf155a5c4e7e1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 202, + 288, + 266 + ], + "blocks": [ + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 198 + ], + "type": "text", + "content": "Table 5: Ablation study: quantitative comparison of various variants of our framework." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 202, + 288, + 266 + ], + "lines": [ + { + "bbox": [ + 50, + 202, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 50, + 202, + 288, + 266 + ], + "type": "table", + "html": "
ConfigLPIPS(BG) ↓LPIPS(FG) ↓CLIP(Image) ↑CLIP(Text) ↑
Baseline0.340.6575.1329.00
+ Pexceptional0.320.6478.5428.23
+ SA injection0.250.6181.6428.58
+ CA injection0.260.6081.6328.52
+ Background0.100.6082.8628.11
", + "image_path": "9d46de8ab927e98e991d5f6840685b0476fc875db14e3a4e5e2f6de49341a38d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 277, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 315 + ], + "type": "text", + "content": "of the background, high object correspondence in both low-level and high-level feature spaces, as well as a high degree of alignment with the text prompt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 326, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 469 + ], + "type": "text", + "content": "User study. We conducted a user study to compare image composition baselines across domains. We recruited 50 participants via Amazon and tasked them with completing 40 ranking questions. Each question consists of 5 options, generated using distinct methods. Details are available in Appendix D. The ranking criteria comprehensively considered foreground preservation, background consistency, seamless composition, and text alignment. The results are listed in Table 4, where the domain information is presented in the format of 'foreground domain & background domain', e.g., photorealism & oil painting. Our method was favored by most participants across different domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 479, + 141, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 141, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 141, + 491 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": "We ablate our key design choices in the following cases: (1) Baseline, where the composition is generated by solving the diffusion ODE from " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " to 0 using DPM-Solver++ without any injection. The starting point is composed by inverted noises under the normal prompt; (2) The exceptional prompt is applied to obtain accurate inverted noises; (3) The self-attention maps " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{r}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " are composed and then injected; (4) The cross-attention " + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,t}^{\\mathrm{cross}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 617 + ], + "type": "text", + "content": " between the main and reference images is further composed for injection; (5) The background preservation is applied." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Table 5 presents the quantitative results, which indicate that the complete algorithm outperforms other variants in all metrics except for " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{(Text)}}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ". Notably, the baseline achieves the best " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{(Text)}}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " as it generates compositions solely relying on the normal prompt without any extra constraint. While metrics alone may not reveal the complete effectiveness of cross-attention injection, Figure 7 illustrates that the interactions between the main and reference images" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 545, + 174 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 174 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 174 + ], + "type": "image", + "image_path": "3af29284b64f3d0ea6355a8598b0fbd45ece9f54c435e9369fcd1dc3969afdb1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 175, + 479, + 182 + ], + "lines": [ + { + "bbox": [ + 372, + 175, + 479, + 182 + ], + "spans": [ + { + "bbox": [ + 372, + 175, + 479, + 182 + ], + "type": "text", + "content": "'an oil painting of a burrito, Van Gogh Style" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 306, + 182, + 545, + 204 + ], + "lines": [ + { + "bbox": [ + 306, + 182, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 182, + 545, + 204 + ], + "type": "text", + "content": "Figure 7: Ablation study of different variants of our framework. SA: self-attention. CA: cross-attention." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 211, + 545, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 545, + 283 + ], + "type": "text", + "content": "are highly beneficial to both foreground and background in terms of preserving appearance and switching domains. Note that preserving the background at different noise levels affects both background and foreground, which is distinct from preserving solely at the final step (Appendix B.3). Additional ablation results are shown in Appendix H.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 293, + 473, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 473, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 473, + 306 + ], + "type": "text", + "content": "6. Limitations and Future Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 312, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 529 + ], + "type": "text", + "content": "The primary limitation of our work lies in its inability to generate an object view that critically differs from the given reference. As a result, the choice of reference image may be restricted at times. This is because the model relies on self-attention maps to provide layout and appearance information, which in turn constrains the development of alternative views. While introducing a loose self-attention injection can generate different views, this often compromises the preservation of the object's appearance. To overcome this, further research could explore utilizing personalized concept learning techniques, such as Textual Inversion [24], to encode identity information in the text prompt through special embeddings. Alternatively, utilizing NeRF-relevant techniques [9, 49, 72] can generate other views of the object for a specific scene, but this can require expensive training. Furthermore, due to the fact that our approach relies on Stable Diffusion, it inherits its shortcomings and biases, which may result in producing artifacts in certain scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 538, + 379, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 379, + 550 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 712 + ], + "type": "text", + "content": "We introduced a method that leverages the high-order ODE solver with the exceptional prompt to achieve precise inversion of real images, which serves as a foundation for further manipulation. Building upon this, we propose a novel training-free framework, TF-ICON, that enables attention-based text-to-image diffusion models to perform image-guided composition across different domains. Our experimental results demonstrate that our approach outperforms the SOTA baselines for both image inversion and composition. We believe that image composition has the potential to become an essential tool for content creators, offering significant benefits for downstream applications in various industries." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2301" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 145, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 145, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 145, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 163 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore under its Strategic Capability Research Centres Funding Initiative. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 107, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 107, + 187 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 107, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 195, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 287, + 239 + ], + "type": "text", + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 240, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 288, + 284 + ], + "type": "text", + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 330 + ], + "type": "text", + "content": "[3] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 331, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 288, + 386 + ], + "type": "text", + "content": "[4] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF conference on computer Vision and pattern recognition, pages 18511-18521, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 387, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 287, + 419 + ], + "type": "text", + "content": "[5] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 2, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 421, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 287, + 464 + ], + "type": "text", + "content": "[6] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 467, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 467, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 467, + 288, + 510 + ], + "type": "text", + "content": "[7] Samaneh Azadi, Deepak Pathak, Sayna Ebrahimi, and Trevor Darrell. Compositional gan: Learning image-conditional binary composition. International Journal of Computer Vision, 128(10):2570-2585, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 287, + 555 + ], + "type": "text", + "content": "[8] Andrew Brown, Cheng-Yang Fu, Omkar Parkhi, Tamara L Berg, and Andrea Vedaldi. End-to-end visual editing with a generatively pre-trained artist. arXiv preprint arXiv:2205.01668, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 557, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 287, + 610 + ], + "type": "text", + "content": "[9] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Consistent single-view perpetual view generation with conditional diffusion models. arXiv preprint arXiv:2211.12131, 2022.8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[11] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. arXiv preprint arXiv:2301.13826, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[12] Bor-Chun Chen and Andrew Kae. Toward realistic image compositing with adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8415-8424, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "text", + "content": "[13] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "type": "text", + "content": "[14] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8394-8403, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[15] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 275, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 317 + ], + "type": "text", + "content": "[16] Xiaodong Cun and Chi-Man Pun. Improving the harmony of the composite image by spatial-separated attention module. IEEE Transactions on Image Processing, 29:4759-4771, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 364 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 441 + ], + "type": "text", + "content": "[19] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[20] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In Proceedings of the IEEE international conference on computer vision, pages 1301-1310, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "text", + "content": "[21] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "type": "text", + "content": "[22] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "text", + "content": "[23] Oran Gafni and Lior Wolf. Wish you were here: Context-aware human generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7840-7849, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[24] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Designing an encoder" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2302" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "for fast personalization of text-to-image models. arXiv preprint arXiv:2302.12228, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "text", + "content": "[26] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "text", + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 195 + ], + "type": "text", + "content": "[28] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 239 + ], + "type": "text", + "content": "[29] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 914-922, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 239, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 304 + ], + "type": "text", + "content": "[30] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 338 + ], + "type": "text", + "content": "[31] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 371 + ], + "type": "text", + "content": "[32] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "text", + "content": "[33] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "text", + "content": "[34] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "text", + "content": "[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2426-2435, 2022. 2, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 559 + ], + "type": "text", + "content": "[36] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "text", + "content": "[37] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. arXiv preprint arXiv:2212.04488, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[38] Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. arXiv preprint arXiv:2210.10960, 2022. 2, 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[39] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[40] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative adversarial networks for image compositing. In Proceedings" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9455-9464, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 97, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 163 + ], + "type": "text", + "content": "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 164, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 231 + ], + "type": "text", + "content": "[42] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8139-8148, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 232, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 276 + ], + "type": "text", + "content": "[43] Liu Liu, Zhenchen Liu, Bo Zhang, Jiangtong Li, Li Niu, Qingyang Liu, and Liqing Zhang. Opa: object placement assessment dataset. arXiv preprint arXiv:2107.01889, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 277, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 312 + ], + "type": "text", + "content": "[44] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 313, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 313, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 313, + 545, + 358 + ], + "type": "text", + "content": "[45] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. arXiv preprint arXiv:2206.01714, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 359, + 545, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 359, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 359, + 545, + 405 + ], + "type": "text", + "content": "[46] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 405, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 449 + ], + "type": "text", + "content": "[47] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 4, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 451, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 451, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 451, + 545, + 507 + ], + "type": "text", + "content": "[48] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021. 2, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 508, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 563 + ], + "type": "text", + "content": "[49] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 564, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 564, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 564, + 545, + 609 + ], + "type": "text", + "content": "[50] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 2, 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 611, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 667 + ], + "type": "text", + "content": "[51] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 714 + ], + "type": "text", + "content": "[52] Li Niu, Wenyan Cong, Liu Liu, Yan Hong, Bo Zhang, Jing Liang, and Liqing Zhang. Making images real again: A comprehensive survey on deep image composition. arXiv preprint arXiv:2106.14490, 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2303" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[53] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. arXiv preprint arXiv:2302.03027, 2023. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 160 + ], + "type": "text", + "content": "[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 228 + ], + "type": "text", + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 231, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 287, + 273 + ], + "type": "text", + "content": "[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 276, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 287, + 330 + ], + "type": "text", + "content": "[57] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2287–2296, 2021. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 332, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 332, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 332, + 287, + 374 + ], + "type": "text", + "content": "[58] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 377, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 377, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 49, + 377, + 287, + 431 + ], + "type": "text", + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 434, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 287, + 476 + ], + "type": "text", + "content": "[60] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 479, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 479, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 479, + 287, + 544 + ], + "type": "text", + "content": "[61] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 545, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 287, + 577 + ], + "type": "text", + "content": "[62] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022.4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 580, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 580, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 49, + 580, + 287, + 623 + ], + "type": "text", + "content": "[63] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4380-4390, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 677 + ], + "type": "text", + "content": "[64] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "text", + "content": "[65] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[66] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 160 + ], + "type": "text", + "content": "[67] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Objectstitch: Generative object compositing. arXiv preprint arXiv:2212.00932, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 204 + ], + "type": "text", + "content": "[68] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40(4):1-14, 2021. 3, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "text", + "content": "[69] Shashank Tripathi, Siddhartha Chandra, Amit Agrawal, Ambrish Tyagi, James M Rehg, and Visesh Chari. Learning to generate synthetic data via compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 461-470, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 262, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 303 + ], + "type": "text", + "content": "[70] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. arXiv preprint arXiv:2211.12572, 2022. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 305, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 337 + ], + "type": "text", + "content": "[71] Bram Wallace, Akash Gokul, and Nikhil Naik. Edict: Exact diffusion inversion via coupled transformations. arXiv preprint arXiv:2211.12446, 2022. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 339, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 339, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 545, + 381 + ], + "type": "text", + "content": "[72] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022.8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 384, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 436 + ], + "type": "text", + "content": "[73] Tengfei Wang, Yong Zhang, Yanbo Fan, Jue Wang, and Qifeng Chen. High-fidelity gan inversion for image attribute editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11379-11388, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 438, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 469 + ], + "type": "text", + "content": "[74] Sam Witteveen and Martin Andrews. Investigating prompt engineering in diffusion models. arXiv preprint arXiv:2211.15462, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 514 + ], + "type": "text", + "content": "[75] Huikai Wu, Shuai Zheng, Junge Zhang, and Kaiqi Huang. Gp-gan: Towards realistic high-resolution image blending. In Proceedings of the 27th ACM international conference on multimedia, pages 2487-2495, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 516, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 516, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 516, + 545, + 568 + ], + "type": "text", + "content": "[76] Ben Xue, Shenghui Ran, Quan Chen, Rongfei Jia, Binqiang Zhao, and Xing Tang. Dccf: Deep comprehensible color filter learning framework for high-resolution image harmonization. In European Conference on Computer Vision, pages 300-316. Springer, 2022. 2, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "text", + "content": "[77] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. arXiv preprint arXiv:2211.13227, 2022. 2, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 615, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 668 + ], + "type": "text", + "content": "[78] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[79] He Zhang, Jianming Zhang, Federico Peruzzi, Zhe Lin, and Vishal M Patel. Deep image compositing. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 365-374, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2304" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 296 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[80] Lingzhi Zhang, Tarmily Wen, Jie Min, Jiancong Wang, David Han, and Jianbo Shi. Learning object placement by inpainting for compositional data augmentation. In European Conference on Computer Vision, pages 566-581. Springer, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[81] Lingzhi Zhang, Tarmily Wen, and Jianbo Shi. Deep image blending. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 231-240, 2020. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 206 + ], + "type": "text", + "content": "[82] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[83] Shuyang Zhang, Runze Liang, and Miao Wang. Shadowgan: Shadow synthesis for virtual objects with conditional adversarial networks. Computational Visual Media, 5(1):105-115, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "text", + "content": "[84] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. Advances in Neural Information Processing Systems, 34:10326-10338, 2021. 5" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2305" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file